diff --git a/.gitattributes b/.gitattributes index d9da5316d2ee7d0da0701cdd945562719a3fce0d..ac21194cb9bbff0c778026a00ae195c020283a07 100644 --- a/.gitattributes +++ b/.gitattributes @@ -679,3 +679,67 @@ saved_model/**/* filter=lfs diff=lfs merge=lfs -text 2025/Diffusion-4K_[[:space:]]Ultra-High-Resolution[[:space:]]Image[[:space:]]Synthesis[[:space:]]with[[:space:]]Latent[[:space:]]Diffusion[[:space:]]Models/4935daf0-51e8-442b-8d44-f641c7d877f3_origin.pdf filter=lfs diff=lfs merge=lfs -text 2025/Diffusion-based[[:space:]]Event[[:space:]]Generation[[:space:]]for[[:space:]]High-Quality[[:space:]]Image[[:space:]]Deblurring/d2f268e0-066c-441a-8577-0e87bbf4c0b9_origin.pdf filter=lfs diff=lfs merge=lfs -text 2025/Diffusion-based[[:space:]]Realistic[[:space:]]Listening[[:space:]]Head[[:space:]]Generation[[:space:]]via[[:space:]]Hybrid[[:space:]]Motion[[:space:]]Modeling/54547168-012b-4a96-999b-4f0bfc23f076_origin.pdf filter=lfs diff=lfs merge=lfs -text +2025/DiffusionDrive_[[:space:]]Truncated[[:space:]]Diffusion[[:space:]]Model[[:space:]]for[[:space:]]End-to-End[[:space:]]Autonomous[[:space:]]Driving/186eb20a-812b-4b73-a9f1-a45f01468323_origin.pdf filter=lfs diff=lfs merge=lfs -text +2025/DiffusionSfM_[[:space:]]Predicting[[:space:]]Structure[[:space:]]and[[:space:]]Motion[[:space:]]via[[:space:]]Ray[[:space:]]Origin[[:space:]]and[[:space:]]Endpoint[[:space:]]Diffusion/a0e804a0-3c59-43c5-bb05-b2160d8636bd_origin.pdf filter=lfs diff=lfs merge=lfs -text +2025/Digital[[:space:]]Twin[[:space:]]Catalog_[[:space:]]A[[:space:]]Large-Scale[[:space:]]Photorealistic[[:space:]]3D[[:space:]]Object[[:space:]]Digital[[:space:]]Twin[[:space:]]Dataset/2d29b0d5-5c08-413f-9ad0-63d98b0c6539_origin.pdf filter=lfs diff=lfs merge=lfs -text +2025/Dinomaly_[[:space:]]The[[:space:]]Less[[:space:]]Is[[:space:]]More[[:space:]]Philosophy[[:space:]]in[[:space:]]Multi-Class[[:space:]]Unsupervised[[:space:]]Anomaly[[:space:]]Detection/269a00de-90dc-4a5c-8e52-737661810fb4_origin.pdf filter=lfs diff=lfs merge=lfs -text +2025/DirectTriGS_[[:space:]]Triplane-based[[:space:]]Gaussian[[:space:]]Splatting[[:space:]]Field[[:space:]]Representation[[:space:]]for[[:space:]]3D[[:space:]]Generation/cb304427-02fe-443c-baf8-93b01f1651af_origin.pdf filter=lfs diff=lfs merge=lfs -text +2025/Directional[[:space:]]Label[[:space:]]Diffusion[[:space:]]Model[[:space:]]for[[:space:]]Learning[[:space:]]from[[:space:]]Noisy[[:space:]]Labels/133ce87a-4a21-4fe5-b113-05841acf1ebd_origin.pdf filter=lfs diff=lfs merge=lfs -text +2025/Disco4D_[[:space:]]Disentangled[[:space:]]4D[[:space:]]Human[[:space:]]Generation[[:space:]]and[[:space:]]Animation[[:space:]]from[[:space:]]a[[:space:]]Single[[:space:]]Image/5f039bdf-5dd5-496e-b90f-b0a10473c8cd_origin.pdf filter=lfs diff=lfs merge=lfs -text +2025/DiscoVLA_[[:space:]]Discrepancy[[:space:]]Reduction[[:space:]]in[[:space:]]Vision,[[:space:]]Language,[[:space:]]and[[:space:]]Alignment[[:space:]]for[[:space:]]Parameter-Efficient[[:space:]]Video-Text[[:space:]]Retrieval/9d7a1765-54c3-41c7-9719-3ce7d4a7b6fa_origin.pdf filter=lfs diff=lfs merge=lfs -text +2025/Discovering[[:space:]]Fine-Grained[[:space:]]Visual-Concept[[:space:]]Relations[[:space:]]by[[:space:]]Disentangled[[:space:]]Optimal[[:space:]]Transport[[:space:]]Concept[[:space:]]Bottleneck[[:space:]]Models/353387ac-1b53-4c3c-ac08-91f95db61c80_origin.pdf filter=lfs diff=lfs merge=lfs -text +2025/Discovering[[:space:]]Hidden[[:space:]]Visual[[:space:]]Concepts[[:space:]]Beyond[[:space:]]Linguistic[[:space:]]Input[[:space:]]in[[:space:]]Infant[[:space:]]Learning/50337d5c-a1b4-4b7f-9096-7ccfa12467ab_origin.pdf filter=lfs diff=lfs merge=lfs -text +2025/Discrete[[:space:]]to[[:space:]]Continuous_[[:space:]]Generating[[:space:]]Smooth[[:space:]]Transition[[:space:]]Poses[[:space:]]from[[:space:]]Sign[[:space:]]Language[[:space:]]Observations/4e14bfdb-c7fc-42ce-bc5c-af5b00971815_origin.pdf filter=lfs diff=lfs merge=lfs -text +2025/Disentangled[[:space:]]Pose[[:space:]]and[[:space:]]Appearance[[:space:]]Guidance[[:space:]]for[[:space:]]Multi-Pose[[:space:]]Generation/eaadccfc-d365-4d2d-83e8-71935f857eca_origin.pdf filter=lfs diff=lfs merge=lfs -text +2025/Disentangling[[:space:]]Safe[[:space:]]and[[:space:]]Unsafe[[:space:]]Image[[:space:]]Corruptions[[:space:]]via[[:space:]]Anisotropy[[:space:]]and[[:space:]]Locality/3a52d538-e6a0-40a8-904e-0b81bce10c6a_origin.pdf filter=lfs diff=lfs merge=lfs -text +2025/DiskVPS_[[:space:]]Vanishing[[:space:]]Point[[:space:]]Detector[[:space:]]via[[:space:]]Hough[[:space:]]Transform[[:space:]]in[[:space:]]a[[:space:]]Disk[[:space:]]Region/80e52809-79f5-4a63-9312-c7df737fcd10_origin.pdf filter=lfs diff=lfs merge=lfs -text +2025/Dispider_[[:space:]]Enabling[[:space:]]Video[[:space:]]LLMs[[:space:]]with[[:space:]]Active[[:space:]]Real-Time[[:space:]]Interaction[[:space:]]via[[:space:]]Disentangled[[:space:]]Perception,[[:space:]]Decision,[[:space:]]and[[:space:]]Reaction/509357f7-50b0-4a3e-be02-5d53b2c39c5b_origin.pdf filter=lfs diff=lfs merge=lfs -text +2025/Dissecting[[:space:]]and[[:space:]]Mitigating[[:space:]]Diffusion[[:space:]]Bias[[:space:]]via[[:space:]]Mechanistic[[:space:]]Interpretability/3f305fc0-29c2-45c0-bbbf-2f6d947ae1f1_origin.pdf filter=lfs diff=lfs merge=lfs -text +2025/Distilled[[:space:]]Prompt[[:space:]]Learning[[:space:]]for[[:space:]]Incomplete[[:space:]]Multimodal[[:space:]]Survival[[:space:]]Prediction/5aae416b-ca2c-40f8-8583-f2902bfc3349_origin.pdf filter=lfs diff=lfs merge=lfs -text +2025/Distilling[[:space:]]Long-tailed[[:space:]]Datasets/f09fbc7b-8c7e-4bee-803c-2d67e3d05b6e_origin.pdf filter=lfs diff=lfs merge=lfs -text +2025/Distilling[[:space:]]Monocular[[:space:]]Foundation[[:space:]]Model[[:space:]]for[[:space:]]Fine-grained[[:space:]]Depth[[:space:]]Completion/1663a87e-e3f7-4145-9d6a-27f93678241e_origin.pdf filter=lfs diff=lfs merge=lfs -text +2025/Distilling[[:space:]]Multi-modal[[:space:]]Large[[:space:]]Language[[:space:]]Models[[:space:]]for[[:space:]]Autonomous[[:space:]]Driving/2bea0a6d-f2ee-4f71-9581-707f2552ee73_origin.pdf filter=lfs diff=lfs merge=lfs -text +2025/Distilling[[:space:]]Spatially-Heterogeneous[[:space:]]Distortion[[:space:]]Perception[[:space:]]for[[:space:]]Blind[[:space:]]Image[[:space:]]Quality[[:space:]]Assessment/c7a78868-95dc-4d35-97c7-04bbcfc1063d_origin.pdf filter=lfs diff=lfs merge=lfs -text +2025/Distilling[[:space:]]Spectral[[:space:]]Graph[[:space:]]for[[:space:]]Object-Context[[:space:]]Aware[[:space:]]Open-Vocabulary[[:space:]]Semantic[[:space:]]Segmentation/7ae5aeda-7866-414e-b296-550ab2d843cd_origin.pdf filter=lfs diff=lfs merge=lfs -text +2025/DistinctAD_[[:space:]]Distinctive[[:space:]]Audio[[:space:]]Description[[:space:]]Generation[[:space:]]in[[:space:]]Contexts/d4975861-cea8-4f94-9bb1-e02706824334_origin.pdf filter=lfs diff=lfs merge=lfs -text +2025/Distinguish[[:space:]]Then[[:space:]]Exploit_[[:space:]]Source-free[[:space:]]Open[[:space:]]Set[[:space:]]Domain[[:space:]]Adaptation[[:space:]]via[[:space:]]Weight[[:space:]]Barcode[[:space:]]Estimation[[:space:]]and[[:space:]]Sparse[[:space:]]Label[[:space:]]Assignment/099209ec-4ec9-4836-8000-10ddecb15989_origin.pdf filter=lfs diff=lfs merge=lfs -text +2025/Distraction[[:space:]]is[[:space:]]All[[:space:]]You[[:space:]]Need[[:space:]]for[[:space:]]Multimodal[[:space:]]Large[[:space:]]Language[[:space:]]Model[[:space:]]Jailbreaking/1375bb02-e050-41f2-9b1d-ed16de66c742_origin.pdf filter=lfs diff=lfs merge=lfs -text +2025/Distribution[[:space:]]Prototype[[:space:]]Diffusion[[:space:]]Learning[[:space:]]for[[:space:]]Open-set[[:space:]]Supervised[[:space:]]Anomaly[[:space:]]Detection/5f6492f6-61fa-40e8-b92f-42fc55a187bf_origin.pdf filter=lfs diff=lfs merge=lfs -text +2025/DivPrune_[[:space:]]Diversity-based[[:space:]]Visual[[:space:]]Token[[:space:]]Pruning[[:space:]]for[[:space:]]Large[[:space:]]Multimodal[[:space:]]Models/f2b05006-c955-46f7-a5af-8b7909376348_origin.pdf filter=lfs diff=lfs merge=lfs -text +2025/DiverseFlow_[[:space:]]Sample-Efficient[[:space:]]Diverse[[:space:]]Mode[[:space:]]Coverage[[:space:]]in[[:space:]]Flows/276f7b10-7b9f-439d-a85a-8ad9f72195cf_origin.pdf filter=lfs diff=lfs merge=lfs -text +2025/Divide[[:space:]]and[[:space:]]Conquer_[[:space:]]Heterogeneous[[:space:]]Noise[[:space:]]Integration[[:space:]]for[[:space:]]Diffusion-based[[:space:]]Adversarial[[:space:]]Purification/94f1ab27-85c1-491d-abe8-0116716b932b_origin.pdf filter=lfs diff=lfs merge=lfs -text +2025/Divot_[[:space:]]Diffusion[[:space:]]Powers[[:space:]]Video[[:space:]]Tokenizer[[:space:]]for[[:space:]]Comprehension[[:space:]]and[[:space:]]Generation/dd9cf0f7-0275-41db-a729-40fdeab212b3_origin.pdf filter=lfs diff=lfs merge=lfs -text +2025/DnLUT_[[:space:]]Ultra-Efficient[[:space:]]Color[[:space:]]Image[[:space:]]Denoising[[:space:]]via[[:space:]]Channel-Aware[[:space:]]Lookup[[:space:]]Tables/efb58f94-4255-404d-9c8c-4e560774a0e2_origin.pdf filter=lfs diff=lfs merge=lfs -text +2025/Do[[:space:]]Computer[[:space:]]Vision[[:space:]]Foundation[[:space:]]Models[[:space:]]Learn[[:space:]]the[[:space:]]Low-level[[:space:]]Characteristics[[:space:]]of[[:space:]]the[[:space:]]Human[[:space:]]Visual[[:space:]]System_/4607e0a5-8aa8-470e-b0e4-8cfed36f43c4_origin.pdf filter=lfs diff=lfs merge=lfs -text +2025/Do[[:space:]]ImageNet-trained[[:space:]]Models[[:space:]]Learn[[:space:]]Shortcuts_[[:space:]]The[[:space:]]Impact[[:space:]]of[[:space:]]Frequency[[:space:]]Shortcuts[[:space:]]on[[:space:]]Generalization/f684a2ee-5aa0-48af-b0b3-fb7e06614e21_origin.pdf filter=lfs diff=lfs merge=lfs -text +2025/Do[[:space:]]Visual[[:space:]]Imaginations[[:space:]]Improve[[:space:]]Vision-and-Language[[:space:]]Navigation[[:space:]]Agents_/6684e0af-b484-43ef-8443-50d4bfdefdd1_origin.pdf filter=lfs diff=lfs merge=lfs -text +2025/Do[[:space:]]We[[:space:]]Always[[:space:]]Need[[:space:]]the[[:space:]]Simplicity[[:space:]]Bias_[[:space:]]Looking[[:space:]]for[[:space:]]Optimal[[:space:]]Inductive[[:space:]]Biases[[:space:]]in[[:space:]]the[[:space:]]Wild/f8b7dd25-2878-4e2f-ac24-ae160cb54c72_origin.pdf filter=lfs diff=lfs merge=lfs -text +2025/Do[[:space:]]We[[:space:]]Really[[:space:]]Need[[:space:]]Curated[[:space:]]Malicious[[:space:]]Data[[:space:]]for[[:space:]]Safety[[:space:]]Alignment[[:space:]]in[[:space:]]Multi-modal[[:space:]]Large[[:space:]]Language[[:space:]]Models_/7cc50b70-a1c6-4460-8b07-4c06c5509d66_origin.pdf filter=lfs diff=lfs merge=lfs -text +2025/Do[[:space:]]Your[[:space:]]Best[[:space:]]and[[:space:]]Get[[:space:]]Enough[[:space:]]Rest[[:space:]]for[[:space:]]Continual[[:space:]]Learning/364acbb3-eca1-422d-be23-a76c46789b30_origin.pdf filter=lfs diff=lfs merge=lfs -text +2025/DoF-Gaussian_[[:space:]]Controllable[[:space:]]Depth-of-Field[[:space:]]for[[:space:]]3D[[:space:]]Gaussian[[:space:]]Splatting/1ef9852f-e1e4-4571-8156-b74884315f2d_origin.pdf filter=lfs diff=lfs merge=lfs -text +2025/DocLayLLM_[[:space:]]An[[:space:]]Efficient[[:space:]]Multi-modal[[:space:]]Extension[[:space:]]of[[:space:]]Large[[:space:]]Language[[:space:]]Models[[:space:]]for[[:space:]]Text-rich[[:space:]]Document[[:space:]]Understanding/587c40bc-a1fe-44db-8881-11814790688b_origin.pdf filter=lfs diff=lfs merge=lfs -text +2025/DocSAM_[[:space:]]Unified[[:space:]]Document[[:space:]]Image[[:space:]]Segmentation[[:space:]]via[[:space:]]Query[[:space:]]Decomposition[[:space:]]and[[:space:]]Heterogeneous[[:space:]]Mixed[[:space:]]Learning/cd34a0c4-c9ec-4e5a-a2a8-007f0052f160_origin.pdf filter=lfs diff=lfs merge=lfs -text +2025/DocVLM_[[:space:]]Make[[:space:]]Your[[:space:]]VLM[[:space:]]an[[:space:]]Efficient[[:space:]]Reader/ec872868-b641-4059-95e8-62856f47ad39_origin.pdf filter=lfs diff=lfs merge=lfs -text +2025/Docopilot_[[:space:]]Improving[[:space:]]Multimodal[[:space:]]Models[[:space:]]for[[:space:]]Document-Level[[:space:]]Understanding/2cc56609-3eb1-412a-98df-28e82ae9cd58_origin.pdf filter=lfs diff=lfs merge=lfs -text +2025/Document[[:space:]]Haystacks_[[:space:]]Vision-Language[[:space:]]Reasoning[[:space:]]Over[[:space:]]Piles[[:space:]]of[[:space:]]1000+[[:space:]]Documents/1b0623c9-cc0a-4950-bcf6-716e67ed00b3_origin.pdf filter=lfs diff=lfs merge=lfs -text +2025/Domain[[:space:]]Adaptive[[:space:]]Diabetic[[:space:]]Retinopathy[[:space:]]Grading[[:space:]]with[[:space:]]Model[[:space:]]Absence[[:space:]]and[[:space:]]Flowing[[:space:]]Data/8fc9ee64-90f9-4f33-8fb1-64a5f7b319a1_origin.pdf filter=lfs diff=lfs merge=lfs -text +2025/Domain[[:space:]]Generalization[[:space:]]in[[:space:]]CLIP[[:space:]]via[[:space:]]Learning[[:space:]]with[[:space:]]Diverse[[:space:]]Text[[:space:]]Prompts/51839507-6fb2-45c2-a66c-0644bd7936ed_origin.pdf filter=lfs diff=lfs merge=lfs -text +2025/Don't[[:space:]]Shake[[:space:]]the[[:space:]]Wheel_[[:space:]]Momentum-Aware[[:space:]]Planning[[:space:]]in[[:space:]]End-to-End[[:space:]]Autonomous[[:space:]]Driving/85994b24-3b2b-443a-a1bf-eabbda00b06b_origin.pdf filter=lfs diff=lfs merge=lfs -text +2025/Doppelgangers[[:space:]]and[[:space:]]Adversarial[[:space:]]Vulnerability/85f12467-2b9c-44c2-8a35-40f17a213b17_origin.pdf filter=lfs diff=lfs merge=lfs -text +2025/Doppelgangers++_[[:space:]]Improved[[:space:]]Visual[[:space:]]Disambiguation[[:space:]]with[[:space:]]Geometric[[:space:]]3D[[:space:]]Features/a4009708-9357-4078-b5aa-d3dd7d42ccfe_origin.pdf filter=lfs diff=lfs merge=lfs -text +2025/DoraCycle_[[:space:]]Domain-Oriented[[:space:]]Adaptation[[:space:]]of[[:space:]]Unified[[:space:]]Generative[[:space:]]Model[[:space:]]in[[:space:]]Multimodal[[:space:]]Cycles/fb2a505a-21d1-4afa-86a3-162bb4136d23_origin.pdf filter=lfs diff=lfs merge=lfs -text +2025/Dora_[[:space:]]Sampling[[:space:]]and[[:space:]]Benchmarking[[:space:]]for[[:space:]]3D[[:space:]]Shape[[:space:]]Variational[[:space:]]Auto-Encoders/6b478709-2cef-4587-8775-92fc50ace470_origin.pdf filter=lfs diff=lfs merge=lfs -text +2025/Dr.[[:space:]]Splat_[[:space:]]Directly[[:space:]]Referring[[:space:]]3D[[:space:]]Gaussian[[:space:]]Splatting[[:space:]]via[[:space:]]Direct[[:space:]]Language[[:space:]]Embedding[[:space:]]Registration/04973834-dfc5-4d84-8238-2be21678c244_origin.pdf filter=lfs diff=lfs merge=lfs -text +2025/DrVideo_[[:space:]]Document[[:space:]]Retrieval[[:space:]]Based[[:space:]]Long[[:space:]]Video[[:space:]]Understanding/81160fea-84bd-416d-98f4-7b4c205b85c0_origin.pdf filter=lfs diff=lfs merge=lfs -text +2025/Dragin3D_[[:space:]]Image[[:space:]]Editing[[:space:]]by[[:space:]]Dragging[[:space:]]in[[:space:]]3D[[:space:]]Space/63363afe-acf7-462c-bf4b-9a484c6ae8ae_origin.pdf filter=lfs diff=lfs merge=lfs -text +2025/DreamCache_[[:space:]]Finetuning-Free[[:space:]]Lightweight[[:space:]]Personalized[[:space:]]Image[[:space:]]Generation[[:space:]]via[[:space:]]Feature[[:space:]]Caching/484a2287-cf26-4bf8-9e41-034501dc6452_origin.pdf filter=lfs diff=lfs merge=lfs -text +2025/DreamOmni_[[:space:]]Unified[[:space:]]Image[[:space:]]Generation[[:space:]]and[[:space:]]Editing/ff84d3e1-2391-4334-90b8-2fe54fd66a09_origin.pdf filter=lfs diff=lfs merge=lfs -text +2025/DreamRelation_[[:space:]]Bridging[[:space:]]Customization[[:space:]]and[[:space:]]Relation[[:space:]]Generation/dc0bd82f-ae71-4d19-a1ab-522a355a226a_origin.pdf filter=lfs diff=lfs merge=lfs -text +2025/DreamText_[[:space:]]High[[:space:]]Fidelity[[:space:]]Scene[[:space:]]Text[[:space:]]Synthesis/6bc8ea68-7656-478a-8d11-90857d6ca90e_origin.pdf filter=lfs diff=lfs merge=lfs -text +2025/DreamTrack_[[:space:]]Dreaming[[:space:]]the[[:space:]]Future[[:space:]]for[[:space:]]Multimodal[[:space:]]Visual[[:space:]]Object[[:space:]]Tracking/2e4cde12-368f-4feb-b1b3-a980bb1eb002_origin.pdf filter=lfs diff=lfs merge=lfs -text +2025/DriveDreamer4D_[[:space:]]World[[:space:]]Models[[:space:]]Are[[:space:]]Effective[[:space:]]Data[[:space:]]Machines[[:space:]]for[[:space:]]4D[[:space:]]Driving[[:space:]]Scene[[:space:]]Representation/aa9a31fd-776c-4a2a-a4f7-87c23be589e2_origin.pdf filter=lfs diff=lfs merge=lfs -text +2025/DriveGEN_[[:space:]]Generalized[[:space:]]and[[:space:]]Robust[[:space:]]3D[[:space:]]Detection[[:space:]]in[[:space:]]Driving[[:space:]]via[[:space:]]Controllable[[:space:]]Text-to-Image[[:space:]]Diffusion[[:space:]]Generation/d5fb6699-8753-41df-8e0e-5b82ff7bf5a3_origin.pdf filter=lfs diff=lfs merge=lfs -text +2025/DriveGPT4-V2_[[:space:]]Harnessing[[:space:]]Large[[:space:]]Language[[:space:]]Model[[:space:]]Capabilities[[:space:]]for[[:space:]]Enhanced[[:space:]]Closed-Loop[[:space:]]Autonomous[[:space:]]Driving/31ff1a5c-7171-4c25-a1d0-ee2bacb8057b_origin.pdf filter=lfs diff=lfs merge=lfs -text +2025/DriveScape_[[:space:]]High-Resolution[[:space:]]Driving[[:space:]]Video[[:space:]]Generation[[:space:]]by[[:space:]]Multi-View[[:space:]]Feature[[:space:]]Fusion/6698f995-6b34-4c34-b673-bcdca2c64dfc_origin.pdf filter=lfs diff=lfs merge=lfs -text +2025/Driving[[:space:]]by[[:space:]]the[[:space:]]Rules_[[:space:]]A[[:space:]]Benchmark[[:space:]]for[[:space:]]Integrating[[:space:]]Traffic[[:space:]]Sign[[:space:]]Regulations[[:space:]]into[[:space:]]Vectorized[[:space:]]HD[[:space:]]Map/cbe7a2b6-7e9c-4185-bab6-ca66fed01b85_origin.pdf filter=lfs diff=lfs merge=lfs -text +2025/DrivingSphere_[[:space:]]Building[[:space:]]a[[:space:]]High-fidelity[[:space:]]4D[[:space:]]World[[:space:]]for[[:space:]]Closed-loop[[:space:]]Simulation/e8e4b9c2-6cbc-49b7-a622-765862a5895e_origin.pdf filter=lfs diff=lfs merge=lfs -text diff --git a/2025/DiffusionDrive_ Truncated Diffusion Model for End-to-End Autonomous Driving/186eb20a-812b-4b73-a9f1-a45f01468323_content_list.json b/2025/DiffusionDrive_ Truncated Diffusion Model for End-to-End Autonomous Driving/186eb20a-812b-4b73-a9f1-a45f01468323_content_list.json new file mode 100644 index 0000000000000000000000000000000000000000..77f60f015ccb1fc0a266fc2e936542446b23f67d --- /dev/null +++ b/2025/DiffusionDrive_ Truncated Diffusion Model for End-to-End Autonomous Driving/186eb20a-812b-4b73-a9f1-a45f01468323_content_list.json @@ -0,0 +1,1538 @@ +[ + { + "type": "text", + "text": "DiffusionDrive: Truncated Diffusion Model for End-to-End Autonomous Driving", + "text_level": 1, + "bbox": [ + 91, + 130, + 906, + 154 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Bencheng Liao $^{1,2,\\diamond}$ Shaoyu Chen $^{2,3}$ Haoran Yin $^{3}$ Bo Jiang $^{2,\\diamond}$ Cheng Wang $^{1,2,\\diamond}$ Sixu Yan $^{2}$ Xinbang Zhang $^{3}$ Xiangyu Li $^{3}$ Ying Zhang $^{3}$ Qian Zhang $^{3}$ Xinggang Wang $^{2}$ $^{1}$ Institute of Artificial Intelligence, Huazhong University of Science & Technology \n $^{2}$ School of EIC, Huazhong University of Science & Technology \n $^{3}$ Horizon Robotics \nCode & Model & Demo: hustvl/DiffusionDrive", + "bbox": [ + 99, + 178, + 893, + 292 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Abstract", + "text_level": 1, + "bbox": [ + 246, + 325, + 326, + 342 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Recently, the diffusion model has emerged as a powerful generative technique for robotic policy learning, capable of modeling multi-mode action distributions. Leveraging its capability for end-to-end autonomous driving is a promising direction. However, the numerous denoising steps in the robotic diffusion policy and the more dynamic, open-world nature of traffic scenes pose substantial challenges for generating diverse driving actions at a real-time speed. To address these challenges, we propose a novel truncated diffusion policy that incorporates prior multi-mode anchors and truncates the diffusion schedule, enabling the model to learn denoising from anchored Gaussian distribution to the multi-mode driving action distribution. Additionally, we design an efficient cascade diffusion decoder for enhanced interaction with conditional scene context. The proposed model, DiffusionDrive, demonstrates $10 \\times$ reduction in denoising steps compared to vanilla diffusion policy, delivering superior diversity and quality in just 2 steps. On the planning-oriented NAVSIM dataset, with aligned ResNet-34 backbone, DiffusionDrive achieves 88.1 PDMS without bells and whistles, setting a new record, while running at a real-time speed of 45 FPS on an NVIDIA 4090. Qualitative results on challenging scenarios further confirm that DiffusionDrive can robustly generate diverse plausible driving actions.", + "bbox": [ + 88, + 358, + 485, + 734 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "1. Introduction", + "text_level": 1, + "bbox": [ + 89, + 763, + 220, + 780 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "End-to-end autonomous driving has gained significant attention in recent years due to advancements in perception models (detection [5, 17, 24, 42], tracking [54-56], online mapping [27, 28, 32], etc.), which directly learns the driving policy from the raw sensor inputs. This data-driven ap", + "bbox": [ + 89, + 789, + 483, + 866 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/99b5d280af68662c38a293d2104fb29b5b421a613a42c76a54486e13cbbd3e04.jpg", + "image_caption": [ + "Figure 1. The comparison of different end-to-end paradigms. (a) Single mode regression [8, 16, 20]. (b) Sampling from vocabulary [4, 25]. (c) Vanilla diffusion policy [7, 19]. (d) The proposed truncated diffusion policy." + ], + "image_footnote": [], + "bbox": [ + 514, + 325, + 908, + 501 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "proach offers a scalable and robust alternative to traditional rule-based motion planning, which often struggles to generalize to complex real-world driving settings.", + "bbox": [ + 511, + 561, + 906, + 606 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "To effectively learn from data, mainstream end-to-end planners (e.g., Transfuser [8], UniAD [16], VAD [20]) typically regress a single-mode trajectory from an ego-query as shown in Fig. 1a. However, this paradigm does not account for the inherent uncertainty and multi-mode nature of driving behaviors. Recently, VADv2 [20] introduces a large fixed vocabulary of anchor trajectories (4096 anchors) to discretize the continuous action space and capture a broader range of driving behaviors, and then samples from these anchors based on predicted scores as shown in Fig. 1b. However, this large fixed-vocabulary paradigm is fundamentally constrained by the number and quality of anchor trajectories, often failing in out-of-vocabulary scenarios. Furthermore, managing a large number of anchors presents significant computational challenges for real-time applications. Rather than discretizing the action space, diffusion model [7] has proven to be a powerful genera", + "bbox": [ + 509, + 607, + 908, + 864 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "CVF", + "bbox": [ + 106, + 2, + 181, + 42 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore.", + "bbox": [ + 236, + 0, + 810, + 46 + ], + "page_idx": 0 + }, + { + "type": "page_footnote", + "text": "To distinguish the term \"multimodal\" used to describe input data, we use \"multi-mode\" in this paper to refer to diverse planning decisions.", + "bbox": [ + 511, + 875, + 906, + 900 + ], + "page_idx": 0 + }, + { + "type": "page_footnote", + "text": "$\\diamond$ Intern of Horizon Robotics; $\\boxtimes$ Corresponding author: Xinggang Wang (xgwang@hust.edu.cn).", + "bbox": [ + 89, + 875, + 482, + 900 + ], + "page_idx": 0 + }, + { + "type": "page_number", + "text": "12037", + "bbox": [ + 480, + 944, + 517, + 955 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "tive decision-making policy in the robotics domain, which can directly sample multi-mode physically plausible actions from a Gaussian distribution via iterative denoising process.", + "bbox": [ + 89, + 90, + 480, + 136 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "This inspires us to replicate the success of the diffusion model in the robotics domain to end-to-end autonomous driving. We apply the vanilla robotic diffusion policy to the well-known single-mode-regression method, Transfuser [8], by proposing a variant, $\\text{Transfuser}_{\\text{DP}}$ , which replaces the deterministic MLP regression head with a conditional diffusion model [34]. Though $\\text{Transfuser}_{\\text{DP}}$ improves planning performance, two major issues arise: 1) The numerous 20 denoising steps in the vanilla DDIM diffusion policy introduce heavy computational consumption during inference as shown in Tab. 2, hindering the real-time application for autonomous driving. 2) The trajectories sampled from different Gaussian noises severely overlap with each other, as illustrated in Fig. 2. This underscores the non-trivial challenge of taming the diffusion models for the dynamic and open-world traffic scenes.", + "bbox": [ + 88, + 138, + 482, + 380 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Unlike the vanilla diffusion policy, which samples actions from a random Gaussian noise conditioned on scene context, human drivers adhere to established driving patterns that they dynamically adjust in response to real-time traffic conditions. This insight motivates us to embed these prior driving patterns into the diffusion policy by partitioning the Gaussian distribution into multiple sub-Gaussian distributions centered around prior anchors, referred to as anchored Gaussian distribution. It is implemented by truncating the diffusion schedule to introduce a small portion of Gaussian noise around the prior anchors as shown in Fig. 3. Thanks to the multi-mode distributional expressivity of the diffusion model, the proposed truncated diffusion policy effectively covers the potential action space without requiring a large set of fixed anchors, as VADv2 does. With more reasonable initial noise samples from the anchored Gaussian distribution, we can truncate the denoising process, reducing the required steps from 20 to just 2—a substantial speedup that satisfies the real-time requirements of autonomous driving.", + "bbox": [ + 91, + 383, + 482, + 684 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "To enhance the interaction with conditional scene context, we propose an efficient transformer-based diffusion decoder that interacts not only with structured queries from the perception module but also with Bird's Eye View (BEV) and perspective view (PV) features through a sparse deformable attention mechanism [3, 62]. Additionally, we introduce a cascade mechanism to iteratively refine the trajectory reconstruction within the diffusion decoder at each denoising step.", + "bbox": [ + 89, + 686, + 482, + 823 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "With these innovations, we present DiffusionDrive, a diffusion model for real-time end-to-end autonomous driving. We benchmark our method on the planning-oriented NAVSIM dataset [11] using non-reactive simulation and closed-loop evaluations. Without bells and whis", + "bbox": [ + 89, + 824, + 482, + 900 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "tles, DiffusionDrive achieves 88.1 PDMS on NAVSIM navtest split with the aligned ResNet-34 backbone, significantly outperforming previous state-of-the-art methods. Even compared to the NAVSIM challenge-winning solution Hydra-MDP- $\\nu_{8192}$ -W-EP [25], which follows VADv2 with 8192 anchor trajectories and further incorporates post-processing and additional supervision, DiffusionDrive still outperforms it by 1.6 PDMS through directly learning from human demonstrations and inferring without post-processing, while running at real-time speed of 45 FPS on an NVIDIA 4090. We further validate the superiority of DiffusionDrive on popular nuScenes dataset [2] with open-loop evaluations, DiffusionDrive runs $1.8 \\times$ faster than VAD and outperforms it [20] by $20.8\\%$ lower L2 error and $63.6\\%$ lower collision rate with the same ResNet-50 backbone, demonstrating state-of-the-art planning performance.", + "bbox": [ + 511, + 90, + 903, + 332 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Our contributions can be summarized as follows:", + "bbox": [ + 532, + 333, + 857, + 345 + ], + "page_idx": 1 + }, + { + "type": "list", + "sub_type": "text", + "list_items": [ + "- We firstly introduce the diffusion model to the field of end-to-end autonomous driving and propose a novel truncated diffusion policy to address the issues of mode collapse and heavy computational overhead found in direct adaptation of vanilla diffusion policy to the traffic scene.", + "- We design an efficient transformer-based diffusion decoder that interacts with the conditional information in a cascaded manner for better trajectory reconstruction.", + "- Without bells and whistles, DiffusionDrive significantly outperforms previous state-of-the-art methods, achieving a record-breaking 88.1 PDMS on the NAVSIM navtest split with the same backbone, while maintaining real-time performance at 45 FPS on an NVIDIA 4090.", + "- We qualitatively demonstrate that DiffusionDrive can generate more diverse and plausible trajectories, exhibiting high-quality multi-mode driving actions in various challenging scenarios." + ], + "bbox": [ + 513, + 347, + 903, + 604 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "2. Related Work", + "text_level": 1, + "bbox": [ + 513, + 616, + 653, + 632 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "End-to-end autonomous driving. UniAD [16], as a pioneering work, demonstrates the potential of end-to-end autonomous driving by integrating multiple perception tasks to enhance planning performance. VAD [20] further explores the use of compact vectorized scene representations to improve efficiency. Subsequently, a series of works [6, 8, 12, 23, 26, 43, 45, 58] have adopted the single-trajectory planning paradigm to enhance planning performance further. More recently, VADv2 [4] shifts the paradigm towards multi-mode planning by scoring and sampling from a large fixed vocabulary of anchor trajectories. Hydra-MDP [25] improves the scoring mechanism of VADv2 by introducing extra supervision from a rule-based scorer. SparseDrive [39] explores an alternative BEV-free solution. Unlike existing multi-mode planning approaches, we propose a novel paradigm that leverages powerful generative diffusion models for end-to-end autonomous driving.", + "bbox": [ + 511, + 643, + 906, + 900 + ], + "page_idx": 1 + }, + { + "type": "page_number", + "text": "12038", + "bbox": [ + 480, + 944, + 517, + 955 + ], + "page_idx": 1 + }, + { + "type": "image", + "img_path": "images/eae6cbced2a079f01e0437de2d3082d21ee7f04b1c9e9400b61c9f8776e04382.jpg", + "image_caption": [ + "(b) Top-1's turning left and diverse top-10's lane changing.", + "Figure 2. Qualitative comparison of Transfuser, TransfuserDP and DiffusionDrive on challenging scenes of NAVSIM navtest split. With the same inputs from front cameras and LiDAR, DiffusionDrive achieves the highest planning quality of top-1 scoring trajectory as illustrated in Tab. 2. We render the highlighted diverse trajectories predicted by DiffusionDrive in the front view. (a) and (b) shows that the top-1 scoring trajectory of DiffusionDrive closely matches the ground truth for both going straight and turning left. Additionally, DiffusionDrive's top-10 scoring trajectory demonstrates high-quality lane changing—an ability not observed in multi-mode TransfuserDP and impossible for Transfuser." + ], + "image_footnote": [], + "bbox": [ + 169, + 85, + 826, + 593 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Diffusion model for traffic simulation. Driving diffusion policy has been explored in the traffic simulation by leveraging only abstract perception groundtruth [9, 18, 21, 44]. MotionDiffuser [21] and CTG [60] are pioneering applications of diffusion models for multi-agent motion prediction, using a conditional diffusion model to sample target trajectories from Gaussian noise. $\\mathrm{CTG + + }$ [59] further incorporates a large language model (LLM) for language-driven guidance, improving usability and enabling realistic traffic simulations. Diffusion-ES [48] replaces reward-gradient-guided denoising with evolutionary search. Moving beyond diffusion models limited to traffic simulation with perception groundtruth, our approach unlocks the potential of dif", + "bbox": [ + 89, + 694, + 485, + 891 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "fusion models for real-time, end-to-end autonomous driving through our proposed truncated diffusion policy and efficient diffusion decoder.", + "bbox": [ + 511, + 694, + 906, + 739 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Diffusion model for robotic policy learning. Diffusion policy [7] demonstrates the great potential in robotic policy learning, effectively capturing multi-mode action distributions and high-dimensional action spaces. Diffuser [19] proposes an unconditional diffusion model for trajectory sampling, incorporating techniques such as classifier-free guidance and image inpainting to achieve guided sampling. Subsequently, numerous works have applied diffusion models to various robotic tasks, including stationary manipulation [1, 53], mobile manipulation [47], autonomous nav", + "bbox": [ + 511, + 750, + 908, + 902 + ], + "page_idx": 2 + }, + { + "type": "page_number", + "text": "12039", + "bbox": [ + 480, + 944, + 519, + 957 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Diffusion step: $q(\\tau^i |\\tau^{i - 1}),i\\in [1,T]$", + "bbox": [ + 382, + 90, + 609, + 109 + ], + "page_idx": 3 + }, + { + "type": "image", + "img_path": "images/8cfefc6af824cac1c473f324a6793a77eea43f897f66e0c85228a8fc8fda716a.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 171, + 112, + 826, + 244 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Denoising step: $p_{\\theta}\\big(\\tau^i |\\tau^{i - 1},z\\big),i\\in [1,T]$", + "bbox": [ + 374, + 250, + 625, + 267 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Vanilla Diffusion Policy vs. Truncated Diffusion Policy (ours)", + "text_level": 1, + "bbox": [ + 300, + 276, + 694, + 290 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Diffusion step: $q(\\tau^i |\\tau^{i - 1}),i\\in [1,T_{\\mathrm{trunc}}]$", + "bbox": [ + 210, + 301, + 470, + 316 + ], + "page_idx": 3 + }, + { + "type": "image", + "img_path": "images/908ca90015f3cdef9e8a7a268d4d0e4dffad49b35c5355a18d9a23521945317e.jpg", + "image_caption": [ + "Adding noise to the anchors with truncated diffusion schedule." + ], + "image_footnote": [], + "bbox": [ + 176, + 320, + 478, + 452 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Denosing from anchored Gaussian distribution with fewer steps.", + "bbox": [ + 514, + 303, + 818, + 315 + ], + "page_idx": 3 + }, + { + "type": "image", + "img_path": "images/90462a069ee2704c6c9619343f1e37dcc31b96ef9d735cc9113ee30ae82e944a.jpg", + "image_caption": [ + "Denoising step: $p_{\\theta}(\\tau^i |\\tau^{i - 1},z),i\\in [1,T_{\\mathrm{trunc}}]$", + "Figure 3. Illustration of truncated diffusion policy by comparing with vanilla diffusion policy. We truncate the diffusion process and only add a small portion of Gaussian noise to diffuse the anchor trajectories. Then, we train the diffusion model to reconstruct the ground-truth trajectory from the anchored Gaussian distribution with conditional scene context. During the inference, we also truncate the denoising process by starting from the better samples in the anchored Gaussian distribution than the pure Gaussian noise." + ], + "image_footnote": [], + "bbox": [ + 496, + 315, + 821, + 452 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "igation [37, 51], quadruped locomotion [38], and dexterous manipulation [46]. However, directly applying vanilla diffusion policy to end-to-end autonomous driving poses unique challenges, as it requires real-time efficiency and the generation of plausible multi-mode trajectories in dynamic and open-world traffic scenes. In this work, we propose a novel truncated diffusion policy to address these challenges, introducing concepts that have not yet been explored in the robotics field.", + "bbox": [ + 88, + 542, + 482, + 678 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Diffusion model for image generation. Diffusion models have been extensively adopted for image generation tasks [33, 36, 49, 50, 61]. DDIM [35] enhances DDPM [14] by enabling efficient sampling with significantly fewer steps based on non-Markovian diffusion processes. Flow matching [30, 31] further optimizes the generative process by directly modeling continuous probability flows. TDPM [57] proposes truncated denoising, which initiates the generation process from an implicit intermediate distribution to accelerate sampling. In contrast to these approaches, our method introduces an explicit driving prior within the diffusion policy, effectively guiding the diffusion process toward more accurate and efficient generation tailored specifically for end-to-end autonomous driving.", + "bbox": [ + 89, + 688, + 482, + 900 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "3. Method", + "text_level": 1, + "bbox": [ + 513, + 541, + 604, + 556 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "3.1. Preliminary", + "text_level": 1, + "bbox": [ + 511, + 565, + 643, + 580 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Task formulation. End-to-end autonomous driving takes raw sensor data as input and predicts the future trajectory of the ego-vehicle. The trajectory is represented as a sequence of waypoints $\\tau = \\{(x_{t},y_{t})\\}_{t = 1}^{T_{f}}$ , where $T_{f}$ denotes the planning horizon, and $(x_{t},y_{t})$ is the location of each waypoint at time $t$ in the current ego-vehicle coordinate system.", + "bbox": [ + 511, + 589, + 906, + 681 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Conditional diffusion model. The conditional diffusion model poses a forward diffusion process as gradually adding noise to the data sample, which can be defined as:", + "bbox": [ + 511, + 683, + 905, + 729 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\nq \\left(\\tau^ {i} \\mid \\tau^ {0}\\right) = \\mathcal {N} \\left(\\tau^ {i}; \\sqrt {\\bar {\\alpha} ^ {i}} \\boldsymbol {\\tau} ^ {0}, \\left(1 - \\bar {\\alpha} ^ {i}\\right) \\mathbf {I}\\right), \\tag {1}\n$$\n", + "text_format": "latex", + "bbox": [ + 563, + 739, + 906, + 766 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "where $\\tau^0$ is the clean data sample, and $\\tau^i$ is the data sample with noise at time $i$ (Note: we use superscript $i$ to denote diffusion timestep). The constant $\\bar{\\alpha}^i = \\prod_{s=1}^i \\alpha^s = \\prod_{s=1}^i (1 - \\beta^s)$ and $\\beta^s$ is the noise schedule. We train the reverse process model $f_\\theta(\\tau^i, z, i)$ to predict $\\tau^0$ from $\\tau^i$ with the guidance of conditional information $z$ , where $\\theta$ is the trainable model parameter. During inference, the trained diffusion model $f_\\theta$ progressively refines from the random", + "bbox": [ + 511, + 777, + 906, + 900 + ], + "page_idx": 3 + }, + { + "type": "page_number", + "text": "12040", + "bbox": [ + 480, + 944, + 519, + 955 + ], + "page_idx": 3 + }, + { + "type": "image", + "img_path": "images/39188c906af22969f52df93d3eccb03e17c013632528cf929e933086a49a30e7.jpg", + "image_caption": [ + "Figure 4. Overall architecture of DiffusionDrive. (a) DiffusionDrive can integrate various existing perception modules and sensor inputs. (b) The designed diffusion decoder takes the sampled noisy trajectories from anchored Gaussian distribution as input and progressively denoises them with enhanced interactions with the conditional scene context in a cascade manner to generate the final predictions." + ], + "image_footnote": [], + "bbox": [ + 91, + 87, + 514, + 258 + ], + "page_idx": 4 + }, + { + "type": "image", + "img_path": "images/3501d4a7121777ea39ebdd62ca4e2b3efafa5c0f6dd4428085611e154596d12e.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 519, + 89, + 908, + 261 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "noise $\\tau^T$ sampled in Gaussian distribution to the predicted clean data sample $\\tau^0$ with the guidance of conditional information $z$ , which is defined as:", + "bbox": [ + 89, + 310, + 485, + 357 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\np _ {\\theta} \\left(\\boldsymbol {\\tau} ^ {0} \\mid z\\right) = \\int p \\left(\\boldsymbol {\\tau} ^ {T}\\right) \\prod_ {i = 1} ^ {T} p _ {\\theta} \\left(\\boldsymbol {\\tau} ^ {i - 1} \\mid \\boldsymbol {\\tau} ^ {i}, z\\right) d \\boldsymbol {\\tau} ^ {1: T}. \\tag {2}\n$$\n", + "text_format": "latex", + "bbox": [ + 99, + 367, + 483, + 409 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "3.2. Investigation", + "text_level": 1, + "bbox": [ + 89, + 420, + 227, + 435 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Turn Transfuser [8] into conditional diffusion model. We begin from the representative deterministic end-to-end planner Transfuser [8] and turn it into a generative model $\\mathrm{Transfuser}_{\\mathrm{DP}}$ by simply replacing the regression MLP layers with the conditional diffusion model UNet following vanilla diffusion policy [7]. During the evaluation, we sample a random noise and progressively refine it with 20 steps. Tab. 2 shows that $\\mathrm{Transfuser}_{\\mathrm{DP}}$ achieves better planning quality than deterministic Transfuser.", + "bbox": [ + 89, + 444, + 483, + 580 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Mode collapse. To further investigate the multi-mode property of the vanilla diffusion policy in driving, we sampled 20 random noises from Gaussian distribution and denoised them using 20 steps. As shown in Fig. 2, the different random noises converge to similar trajectories after the denoising process. To quantitatively analyze the phenomenon of mode collapse, we define a mode diversity score $\\mathcal{D}$ based on the mean Intersection over Union (mIoU) between each denoised trajectory and the union of all denoised trajectories:", + "bbox": [ + 89, + 582, + 483, + 731 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {D} = 1 - \\frac {1}{N} \\sum_ {i = 1} ^ {N} \\frac {\\operatorname {A r e a} \\left(\\tau_ {i} \\cap \\bigcup_ {j = 1} ^ {N} \\tau_ {j}\\right)}{\\operatorname {A r e a} \\left(\\tau_ {i} \\cup \\bigcup_ {j = 1} ^ {N} \\tau_ {j}\\right)}, \\tag {3}\n$$\n", + "text_format": "latex", + "bbox": [ + 158, + 729, + 482, + 772 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "where $\\tau_{i}$ represents the $i$ -th denoised trajectory, $N$ is the total number of sampled trajectories and $\\bigcup_{j=1}^{N} \\tau_{j}$ is the union of all denoised trajectories. A higher mIoU indicates less diversity of the denoised trajectories. The quantitative mode diversity results in Tab. 2 further validate the observations presented in Fig. 2.", + "bbox": [ + 89, + 777, + 483, + 868 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Heavy denoising overhead. The DDIM [35] diffusion policy requires 20 denoising steps to transform random noise", + "bbox": [ + 89, + 869, + 483, + 901 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "into a feasible trajectory, which introduces significant computational overhead, reducing the FPS from 60 to 7, as shown in Tab. 2, and making it impractical for real-time online driving applications.", + "bbox": [ + 511, + 311, + 906, + 372 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "3.3. Truncated Diffusion", + "text_level": 1, + "bbox": [ + 511, + 381, + 705, + 395 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Human driving follows fixed patterns, unlike the random noise denoising in vanilla diffusion policy. Motivated by this, we propose a truncated diffusion policy that begins the denoising process from an anchored Gaussian distribution instead of a standard Gaussian distribution. To enable the model to learn to denoise from the anchored Gaussian distribution to the desired driving policy, we further truncate the diffusion schedule during training, adding only a small amount of Gaussian noise to the anchors.", + "bbox": [ + 511, + 402, + 906, + 537 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Training. We first construct the diffusion process by adding Gaussian noise to anchors $\\{\\mathbf{a}_k\\}_{k=1}^{N_{\\text{anchor}}}$ clustered by K-Means on the training set, where $\\mathbf{a}_k = \\{(x_t, y_t)\\}_{t=1}^{T_f}$ . We truncate the diffusion noise schedule to diffuse the anchors to the anchored Gaussian distribution:", + "bbox": [ + 511, + 541, + 906, + 619 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\tau_ {k} ^ {i} = \\sqrt {\\bar {\\alpha} ^ {i}} \\mathbf {a} _ {k} + \\sqrt {1 - \\bar {\\alpha} ^ {i}} \\boldsymbol {\\epsilon}, \\quad \\boldsymbol {\\epsilon} \\sim \\mathcal {N} (0, \\mathbf {I}), \\qquad (4)\n$$\n", + "text_format": "latex", + "bbox": [ + 565, + 628, + 906, + 647 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "where $i \\in [1, T_{\\mathrm{trunc}}]$ and $T_{\\mathrm{trunc}} \\ll T$ is the truncated diffusion steps.", + "bbox": [ + 511, + 657, + 905, + 686 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "During training, the diffusion decoder $f_{\\theta}$ takes as input $N_{\\mathrm{anchor}}$ noisy trajectories $\\{\\tau_k^i\\}_{k=1}^{N_{\\mathrm{anchor}}}$ and predicts classification scores $\\{\\hat{s}_k\\}_{k=1}^{N_{\\mathrm{anchor}}}$ and denoised trajectories $\\{\\hat{\\tau}_k\\}_{k=1}^{N_{\\mathrm{anchor}}}$ :", + "bbox": [ + 511, + 688, + 905, + 738 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\left\\{\\hat {s} _ {k}, \\hat {\\tau} _ {k} \\right\\} _ {k = 1} ^ {N _ {\\text {a n c h o r}}} = f _ {\\theta} \\left(\\left\\{\\tau_ {k} ^ {i} \\right\\} _ {k = 1} ^ {N _ {\\text {a n c h o r}}}, z\\right), \\tag {5}\n$$\n", + "text_format": "latex", + "bbox": [ + 594, + 744, + 906, + 766 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "where $z$ represents the conditional information. We assign the noisy trajectory around the closest anchor to the ground truth trajectory $\\tau_{\\mathrm{gt}}$ as positive sample $(y_k = 1)$ and others as negative samples $(y_k = 0)$ . The training objective combines trajectory reconstruction and classification:", + "bbox": [ + 511, + 775, + 906, + 851 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {L} = \\sum_ {k = 1} ^ {N _ {\\text {a n c h o r}}} \\left[ y _ {k} \\mathcal {L} _ {\\text {r e c}} \\left(\\hat {\\tau} _ {k}, \\tau_ {\\mathrm {g t}}\\right) + \\lambda \\mathrm {B C E} \\left(\\hat {s} _ {k}, y _ {k}\\right) \\right], \\tag {6}\n$$\n", + "text_format": "latex", + "bbox": [ + 562, + 862, + 905, + 902 + ], + "page_idx": 4 + }, + { + "type": "page_number", + "text": "12041", + "bbox": [ + 480, + 944, + 517, + 955 + ], + "page_idx": 4 + }, + { + "type": "table", + "img_path": "images/e45249461eb686dfd45a7e9e800e8405ad0142b839ff26eadf7d5daf32e7d80d.jpg", + "table_caption": [], + "table_footnote": [], + "table_body": "
MethodInputImg. BackboneAnchorNC↑DAC ↑TTC ↑Comf. ↑EP ↑PDMS ↑
UniAD [16]CameraResNet-34 [13]097.891.992.910078.883.4
PARA-Drive [45]CameraResNet-34 [13]097.992.493.099.879.384.0
LTF [8]CameraResNet-34 [13]097.492.892.410079.083.8
Transfuser [8]C & LResNet-34 [13]097.792.892.810079.284.0
DRAMA [52]C & LResNet-34 [13]098.093.194.810080.185.5
VADv2- \\( \\mathcal{V}_{8192} \\) [4]C & LResNet-34 [13]819297.289.191.610076.080.9
Hydra-MDP- \\( \\mathcal{V}_{8192} \\) [25]C & LResNet-34 [13]819297.991.792.910077.683.0
Hydra-MDP- \\( \\mathcal{V}_{8192} \\)-W-EP [25]C & LResNet-34 [13]819298.396.094.610078.786.5
DiffusionDrive (Ours)C & LResNet-34 [13]2098.296.294.710082.288.1
", + "bbox": [ + 93, + 88, + 903, + 262 + ], + "page_idx": 5 + }, + { + "type": "table", + "img_path": "images/90d41fde76a17be1cc5e9e996a65b8c2f21cc99ab9cbe1617e0483aabeced382.jpg", + "table_caption": [ + "Table 1. Comparison on planning-oriented NAVSIM navtest split with closed-loop metrics. “C & L” denotes the use of both camera and LiDAR as sensor inputs. “ $\\mathcal{V}_{8192}$ ” denotes 8192 anchors. “Hydra-MDP- $\\mathcal{V}_{8192}$ -W-EP” is a variant of Hydra-MDP [25], which is further trained to fit the EP evaluation metric with additional supervision from the rule-based evaluator and uses weighted confidence post-processing. DiffusionDrive simply learns from human demonstrations and infers without post-processing. The best and the second best results are denoted by bold and underline." + ], + "table_footnote": [], + "table_body": "
MethodNC↑DAC↑TTC↑Comf.↑EP↑PDMS↑Arch.Plan Module TimeTotal ↓D↑Para.↓FPS↑
Step Time↓Steps ↓
Transfuser97.792.892.810079.284.0MLP0.2ms10.2ms0%56M60
TransfuserDP97.593.792.710079.484.6+0.6UNet6.5ms20130.0ms11%101M7
TransfuserTD97.994.293.910080.285.7+1.7UNet6.9ms213.8ms70%102M27
DiffusionDrive98.296.294.710082.288.1+4.1Dec.3.8ms27.6ms74%60M45
", + "bbox": [ + 94, + 334, + 897, + 438 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Table 2. Roadmap from Transfuser to DiffusionDrive on NAVSIM navtest split. \"TransfuserDP\" denotes Transfuser with vanilla DDIM diffusion policy [7]. \"TransfuserTD\" denotes Transfuser with truncated diffusion policy. \"Step Time\" denotes the runtime of each denoising step. \"FPS\" and runtime are measured on an NVIDIA 4090 GPU. \"D\" denotes the mode diversity score defined in Eq. (3).", + "bbox": [ + 89, + 439, + 906, + 484 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "where $\\lambda$ balances the simple L1 reconstruction loss $\\mathcal{L}_{\\mathrm{rec}}$ and binary cross-entropy (BCE) classification loss.", + "bbox": [ + 89, + 487, + 482, + 517 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Inference. We use a truncated denoising process that starts with noisy trajectories sampled from the anchored Gaussian distribution and progressively denoises them to final predictions. At each denoising timestep, the estimated trajectories from the previous step are passed to the diffusion decoder $f_{\\theta}$ , which predicts classification scores $\\{\\hat{s}_k\\}_{k=1}^{N_{\\mathrm{infer}}}$ and coordinates $\\{\\hat{\\tau}_k\\}_{k=1}^{N_{\\mathrm{infer}}}$ . After obtaining the current timestep's predictions, we apply the DDIM [35] update rule to sample trajectories for the next timestep.", + "bbox": [ + 89, + 518, + 482, + 656 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Inference flexibility. A key advantage of our approach lies in its inference flexibility. While the model is trained with $N_{\\text{anchor}}$ trajectories, the inference process can accommodate an arbitrary number of trajectory samples $N_{\\text{infer}}$ , where $N_{\\text{infer}}$ can be dynamically adjusted based on computational resources or application requirements.", + "bbox": [ + 89, + 657, + 482, + 750 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "3.4. Architecture", + "text_level": 1, + "bbox": [ + 89, + 756, + 225, + 770 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "The overall architecture of our proposed method, DiffusionDrive, is illustrated in Fig. 4. DiffusionDrive can integrate various existing perception modules used in previous end-to-end planners [8, 16, 20, 39] and take different sensor inputs. The designed diffusion decoder is tailored for the complex and challenging driving application, which has enhanced interactions with the conditional scene context.", + "bbox": [ + 89, + 777, + 482, + 883 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Diffusion decoder. Given the set of sampled noisy trajectories", + "bbox": [ + 89, + 885, + 482, + 901 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "ries $\\{\\hat{\\tau}_k\\}_{k=1}^{N_{\\mathrm{infer}}}$ from the anchored Gaussian distribution, we begin by applying deformable spatial cross-attention [29, 42, 62] to interact with Bird's Eye View (BEV) or Perspective View (PV) features based on the trajectory coordinates. Subsequently, cross-attention is performed between the trajectory features and the agent/map queries derived from the perception module, followed by a feed-forward network (FFN). To encode the diffusion timestep information, we utilize a Timestep Modulation layer, which is followed by a Multi-Layer Perceptron (MLP) that predicts the confidence score and the offset relative to the initial noisy trajectory coordinates. The output from this diffusion decoder layer serves as the input for the subsequent cascade diffusion decoder layer. DiffusionDrive further reuses the cascade diffusion decoder to iteratively denoise the trajectory during inference, with parameters shared across the different denoising timesteps. The final trajectory with the highest confidence score is selected as the output.", + "bbox": [ + 511, + 486, + 906, + 758 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "4. Experiment", + "text_level": 1, + "bbox": [ + 511, + 773, + 638, + 792 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "4.1. Dataset", + "text_level": 1, + "bbox": [ + 511, + 800, + 609, + 814 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "NAVSIM. The NAVSIM dataset [11] is a real-world planning-oriented dataset builds upon OpenScene [10], a compact redistribution of nuPlan [22], the largest publicly available annotated driving dataset. NAVSIM leverages eight cameras to achieve a full $360^{\\circ}$ FOV, along with a", + "bbox": [ + 511, + 824, + 906, + 902 + ], + "page_idx": 5 + }, + { + "type": "page_number", + "text": "12042", + "bbox": [ + 480, + 944, + 519, + 957 + ], + "page_idx": 5 + }, + { + "type": "table", + "img_path": "images/5d60719ec93c39d85aaf432bb11752b22d4eff87f1347e8e26a027665583c888.jpg", + "table_caption": [], + "table_footnote": [], + "table_body": "
IDUNet DecoderEgo Query InteractionSpatial Cross-attnAgent/Map Cross-attnCascade DecoderParam.↓NC↑DAC↑Planning MetricPDMS↑
TTC↑Comf.↑EP↑
1✓✓XXX102M97.994.293.910080.285.7
2X✓XXX57M88.783.280.084.843.355.1
3X✓✓XX58M98.295.494.410081.387.1
4X✓X✓X58M97.993.593.810079.885.1
5X✓✓✓X59M98.095.894.410081.787.4
6X✓✓✓✓60M98.296.294.710082.288.1
", + "bbox": [ + 93, + 88, + 903, + 229 + ], + "page_idx": 6 + }, + { + "type": "table", + "img_path": "images/8d14f9f0b1907f2e1c1f24a3c9ffc59f48c791507d86b6b15ede4856c841428c.jpg", + "table_caption": [ + "Table 3. Ablation for design choices. \"Cascade Decoder\" indicates that we stack 2 cascade diffusion decoder layers. ID-1 refers to TransfuserTD in Tab. 2, utilizing conditional UNet and interaction with the ego-query, which Transfuser uses to directly regress the single-mode trajectory." + ], + "table_footnote": [], + "table_body": "
StepsParam.NCDACTTCComf.EPPDMS
160M98.396.094.710082.187.9
260M98.296.294.710082.288.1
360M98.296.394.710092.288.1
", + "bbox": [ + 93, + 277, + 356, + 335 + ], + "page_idx": 6 + }, + { + "type": "table", + "img_path": "images/dfbc91c12c6096eb2909c06a78669315159466720a4e4cce73d6cba83ba5ad34.jpg", + "table_caption": [ + "Table 4. Denoising step number." + ], + "table_footnote": [], + "table_body": "
StagesParam.NCDACTTCComf.EPPDMS
159M98.095.894.410081.787.4
260M98.296.294.710082.288.1
465M98.496.294.910082.488.2
", + "bbox": [ + 367, + 277, + 633, + 335 + ], + "page_idx": 6 + }, + { + "type": "table", + "img_path": "images/f061fbfa1ae90ca4196cf7add05b0aa11c15dc6dfc79c2924f6e7a47949558fb.jpg", + "table_caption": [ + "Table 5. Cascade stages." + ], + "table_footnote": [], + "table_body": "
N_inferParam.NCDACTTCComf.EPPDMS
1060M97.993.593.110080.084.9
2060M98.296.294.710082.288.1
4060M98.596.294.810082.588.2
", + "bbox": [ + 640, + 277, + 900, + 335 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Table 6. Number of sampled noises ${N}_{\\text{infer }}$ .", + "bbox": [ + 638, + 338, + 898, + 351 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "merged LiDAR point cloud derived from five sensors. Annotations are provided at a frequency of $2\\mathrm{Hz}$ and include both HD maps and object bounding boxes. The dataset is designed to emphasize challenging driving scenarios involving dynamic changes in driving intentions, while deliberately excluding trivial situations such as stationary scenes or constant-speed driving.", + "bbox": [ + 88, + 363, + 482, + 469 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "NAVSIM benchmarks planning performance using non-reactive simulations and closed-loop metrics for comprehensive evaluation. In this paper, we employ the proposed PDM score (PDMS) [11], which is a weighted combination of several sub-scores: no at-fault collisions (NC), drivable area compliance (DAC), time-to-collision (TTC), comfort (Comf.), and ego progress (EP).", + "bbox": [ + 88, + 470, + 482, + 575 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "4.2. Implementation Detail", + "text_level": 1, + "bbox": [ + 89, + 590, + 302, + 606 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "We adopt the same perception modules and ResNet-34 backbone [13] as Transfuser for fair comparison. In the diffusion decoder layer, we employ spatial cross-attention to only interact with BEV features following Transfuser's BEV-based setting. We only perform agent cross-attention, since the perception module of Transfuser does not include vectorized map construction. We stack 2 cascade diffusion decoder layers and apply truncated diffusion policy with 20 clustered anchors. The training diffusion schedule is truncated by 50/1000 to diffuse the anchors, while during inference, we use only 2 denoising steps and select the top-1 scoring predicted trajectory for evaluation. The training and inference recipe directly follows Transfuser: We use three cropped and downscaled forward-facing camera images, concatenated as a $1024 \\times 256$ image, and a rasterized BEV LiDAR as input; DiffusionDrive is trained on navtrain split from scratch for 100 epochs with AdamW optimizer on 8 NVIDIA 4090 GPUs with total batch size of 512, setting the learning rate to $6 \\times 10^{-4}$ . No test-time", + "bbox": [ + 91, + 613, + 482, + 900 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "augmentation is applied and the final output for evaluation on navtest split is 8-waypoint trajectory over 4 seconds.", + "bbox": [ + 511, + 364, + 903, + 393 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "4.3. Quantitative Comparison", + "text_level": 1, + "bbox": [ + 511, + 406, + 745, + 422 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Tab. 1 compares DiffusionDrive with state-of-the-art methods on NAVSIM navtest split. With the same ResNet-34 backbone, DiffusionDrive achieves 88.1 PDMS score, demonstrating significant superior performance over the previous learning-based methods. Compared to VADv2, DiffusionDrive surpasses it by 7.2 PDMS while reducing the number of anchors from 8192 to 20, representing a $400 \\times$ reduction. DiffusionDrive also outperforms Hydra-MDP, which follows VADv2's sampling-from-vocabulary paradigm, with a 5.1 PDMS improvement. Even compared to the Hydra-MDP- $\\mathcal{V}_{8192}$ -W-EP, which is a variant of Hydra-MDP [25] by further training to fit the EP evaluation metric with additional supervision and using weighted confidence post-processing, DiffusionDrive still outperforms it by 3.5 EP and 1.6 overall PDMS, relying solely on a straightforward learning-from-human approach without any post-processing. Compared to the Transfuser baseline, where we only differ in the planning module, DiffusionDrive delivers a notable 4.1 PDMS improvement, outperforming it across all sub-scores.", + "bbox": [ + 511, + 428, + 906, + 731 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "4.4. Roadmap", + "text_level": 1, + "bbox": [ + 511, + 742, + 625, + 758 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "In Tab. 2, converting Transfuser $\\mathrm{DP}$ using vanilla diffusion policy improves the PDMS score by 0.6 and the mode diversity score $\\mathcal{D}$ by $11\\%$ . However, it also significantly increases the overhead of the planning module, requiring $20\\times$ more denoising steps and $32\\times$ the step time, resulting in a total $650\\times$ increase in runtime overhead. With the proposed truncated diffusion policy, $\\mathrm{Transfuser}_{\\mathrm{TD}}$ reduces the number of denoising steps from 20 to 2 while improving PDMS by 1.1 and mode diver", + "bbox": [ + 511, + 763, + 906, + 900 + ], + "page_idx": 6 + }, + { + "type": "page_number", + "text": "12043", + "bbox": [ + 480, + 944, + 517, + 955 + ], + "page_idx": 6 + }, + { + "type": "table", + "img_path": "images/fcc3e04c5f71e323a22828beec3a3aea956064444a32df850c3ee01e1054e61d.jpg", + "table_caption": [], + "table_footnote": [], + "table_body": "
MethodInputImg. BackboneL2 (m) ↓Collision Rate (%) ↓FPS ↑
1s2s3sAvg.1s2s3sAvg.
ST-P3 [15]CameraEffNet-b4 [40]1.332.112.902.110.230.621.270.711.6
UniAD [16]CameraResNet-101 [13]0.450.701.040.730.620.580.630.611.8
OccNet [41]CameraResNet-50 [13]1.292.132.992.140.210.591.370.722.6
VAD [20]CameraResNet-50 [13]0.410.701.050.720.070.170.410.224.5
SparseDrive [39]CameraResNet-50 [13]0.290.580.960.610.010.050.180.089.0
DiffusionDrive (Ours)CameraResNet-50 [13]0.270.540.900.570.030.050.160.088.2
", + "bbox": [ + 91, + 88, + 906, + 232 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Table 7. Comparison on nuScenes dataset with open-loop metrics. FPS is measured on a single NVIDIA 4090 GPU following the recipe of SparseDrive [39]. Metric calculation follows ST-P3 [15].", + "bbox": [ + 91, + 233, + 906, + 262 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "sity by $59\\%$ . By further incorporating the proposed diffusion decoder, the final model, DiffusionDrive, reaches 88.1 PDMS and $74\\%$ mode diversity score $\\mathcal{D}$ . Compared to the TransfuserDP, DiffusionDrive shows improvements in 3.5 PDMS and $64\\%$ mode diversity, and a $10\\times$ reduction in denoising steps, resulting in a $6\\times$ speedup in FPS. This enables real-time, high-quality, multi-mode planning.", + "bbox": [ + 89, + 266, + 480, + 375 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "4.5. Ablation Study", + "text_level": 1, + "bbox": [ + 89, + 383, + 243, + 400 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Effect of designs in diffusion decoder. Tab. 3 shows the effectiveness of our design choices in the diffusion decoder. ID-1 is the Transfuser $_{\\text{TD}}$ in the Tab. 2. By comapring ID-6 and ID-1, we can see that the proposed diffusion decoder reduce the $39\\%$ parameters and significantly improves the planning quality by 2.4 PDMS. ID-2 shows severe performance degeneration due to the lack of rich and hierarchical interaction with the environment. By comparing ID-2 and ID-3, we show that spatial cross-attention is vital for accurate planning. ID-5 shows that the proposed cascade mechanism is effective and can further improve the performance.", + "bbox": [ + 89, + 407, + 482, + 574 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Denoising step number. Tab. 4 shows that due to a reasonable start point, DiffusionDrive achieves good planning quality even with only 1 step, and further denoising steps offer quality improvement and inference flexibility for complex environments.", + "bbox": [ + 89, + 575, + 482, + 651 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Cascade stages. Tab. 5 ablates the impact of cascade stage number. Increasing the stage number can improve the planning quality but saturate at the 4 stages and cost more parameters and inference time at each step.", + "bbox": [ + 89, + 652, + 482, + 715 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Number of sampled noises $N_{\\mathrm{infer}}$ . As stated in Sec. 3.3, DiffusionDrive can generate varied trajectories by simply sampling a variable number of noises from anchored Gaussian distribution. Tab. 6 shows that 10 sampled noises can already achieve a decent planning quality. By sampling more noises, DiffusionDrive can cover potential planning action space and lead to improved planning quality.", + "bbox": [ + 89, + 715, + 482, + 823 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "4.6. Qualitative Comparison", + "text_level": 1, + "bbox": [ + 89, + 832, + 312, + 849 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Since the PDMS planning metric calculates based on the top-1 scoring trajectory and our proposed $\\mathcal{D}$ score evaluates mode diversity, these metrics alone cannot fully capture the", + "bbox": [ + 89, + 854, + 482, + 901 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "quality of diverse trajectories. To further validate the quality of multi-mode trajectories, we visualize the planning results of Transfuser, $\\mathrm{Transfuser}_{\\mathrm{DP}}$ and DiffusionDrive on challenging scenarios of NAVSIM navtest split in Fig. 2. The results indicate that the multi-mode trajectories generated by DiffusionDrive are not only diverse but also of high quality. In Fig. 2a, the top-1 scoring trajectory generated by DiffusionDrive closely resembles the ground-truth trajectory, while the highlighted top-10 scoring trajectory surprisingly tries to perform high-quality lane changing. In Fig. 2b, the highlighted top-10 scoring trajectory also performs a lane change, and a neighboring low-scoring trajectory further interacts with surrounding agents to effectively avoid collisions.", + "bbox": [ + 511, + 266, + 906, + 478 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "4.7. Quantitative Comparison on nuScenes dataset", + "text_level": 1, + "bbox": [ + 511, + 488, + 903, + 505 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "The nuScenes dataset is previously popular benchmark for end-to-end planning. Since the major scenarios of nuScenes are simple and trivial situations, we only perform comparison in Tab. 7. We implement DiffusionDrive on top of SparseDrive [39] following its training and inference recipe using open-loop metrics proposed in ST-P3 [15]. We stack 2 cascade diffusion decoder layers and apply the truncated diffusion policy with 18 clustered anchors.", + "bbox": [ + 511, + 511, + 905, + 632 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "As shown in Tab. 7, DiffusionDrive reduces the average L2 error of SparseDrive by $0.04\\mathrm{m}$ , achieving the lowest L2 error and average collision rate. While DiffusionDrive is also efficient and runs $1.8\\times$ faster than VAD with $20.8\\%$ lower L2 error and $63.6\\%$ lower collision rate.", + "bbox": [ + 511, + 633, + 905, + 708 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "5. Conclusion", + "text_level": 1, + "bbox": [ + 511, + 723, + 633, + 739 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "In this work, we propose a novel generative driving decision-making model, DiffusionDrive, for end-to-end autonomous driving by incorporating the proposed truncated diffusion policy and efficient cascade diffusion decoder. DiffusionDrive can denoise a variable number of samples from an anchored Gaussian distribution to generate diverse planning trajectories at real-time speeds. Comprehensive experiments and qualitative comparisons validate the superiority of DiffusionDrive in planning quality, running efficiency, and mode diversity.", + "bbox": [ + 511, + 750, + 906, + 902 + ], + "page_idx": 7 + }, + { + "type": "page_number", + "text": "12044", + "bbox": [ + 480, + 944, + 519, + 957 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Acknowledgement", + "text_level": 1, + "bbox": [ + 91, + 90, + 250, + 107 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "This work was partially supported by the National Natural Science Foundation of China (NSFC) under Grant No. 62276108. We would like to acknowledge Tianheng Cheng for helpful feedback on the draft.", + "bbox": [ + 89, + 114, + 483, + 172 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "References", + "text_level": 1, + "bbox": [ + 91, + 202, + 187, + 217 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[1] Anurag Ajay, Yilun Du, Abhi Gupta, Joshua B. Tenenbaum, Tommi S. Jaakkola, and Pulkit Agrawal. Is conditional generative modeling all you need for decision making? In ICLR, 2023. 3", + "[2] Holger Caesar, Varun Bankiti, Alex H Lang, Sourabh Vora, Venice Erin Liong, Qiang Xu, Anush Krishnan, Yu Pan, Giancarlo Baldan, and Oscar Beijbom. nuscenes: A multimodal dataset for autonomous driving. In CVPR, 2020. 2", + "[3] Shaoyu Chen, Tianheng Cheng, Xinggang Wang, Wenming Meng, Qian Zhang, and Wenyu Liu. Efficient and robust 2d-to-bev representation learning via geometry-guided kernel transformer. CoRR, abs/2206.04584, 2022. 2", + "[4] Shaoyu Chen, Bo Jiang, Hao Gao, Bencheng Liao, Qing Xu, Qian Zhang, Chang Huang, Wenyu Liu, and Xinggang Wang. Vadv2: End-to-end vectorized autonomous driving via probabilistic planning. arXiv preprint arXiv:2402.13243, 2024. 1, 2, 6", + "[5] Shaoyu Chen, Xinggang Wang, Tianheng Cheng, Qian Zhang, Chang Huang, and Wenyu Liu. Polardetr: Polar parametrization for vision-based surround-view 3d detection. Image and Vision Computing, 156:105438, 2025. 1", + "[6] Zhili Chen, Maosheng Ye, Shuangjie Xu, Tongyi Cao, and Qifeng Chen. Ppad: Iterative interactions of prediction and planning for end-to-end autonomous driving. In ECCV, 2024. 2", + "[7] Cheng Chi, Zhenjia Xu, Siyuan Feng, Eric Cousineau, Yilun Du, Benjamin Burchfiel, Russ Tedrake, and Shuran Song. Diffusion policy: Visuomotor policy learning via action diffusion. In RSS, 2023. 1, 3, 5, 6", + "[8] Kashyap Chitta, Aditya Prakash, Bernhard Jaeger, Zehao Yu, Katrin Renz, and Andreas Geiger. Transfuser: Imitation with transformer-based sensor fusion for autonomous driving. TPAMI, 2022. 1, 2, 5, 6", + "[9] Younwoo Choi, Ray Coden Mercurius, Soheil Mohamad Al-izadeh Shabestary, and Amir Rasouli. Dice: Diverse diffusion model with scoring for trajectory prediction. In IV, 2024. 3", + "[10] OpenScene Contributors. Openscene: The largest up-to-date 3d occupancy prediction benchmark in autonomous driving. https://github.com/OpenDriveLab/OpenScene, 2023.6", + "[11] Daniel Dauner, Marcel Hallgarten, Tianyu Li, Xinshuo Weng, Zhiyu Huang, Zetong Yang, Hongyang Li, Igor Gilitschenski, Boris Ivanovic, Marco Pavone, Andreas Geiger, and Kashyap Chitta. Navsim: Data-driven non-reactive autonomous vehicle simulation and benchmarking. In NeurIPS, 2024. 2, 6, 7" + ], + "bbox": [ + 93, + 226, + 483, + 898 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[12] Xunjiang Gu, Guanyu Song, Igor Gilitschenski, Marco Pavone, and Boris Ivanovic. Producing and leveraging online map uncertainty in trajectory prediction. In CVPR, 2024. 2", + "[13] Kaiming He, Xiangyu Zhang, Shaoqing Ren, and Jian Sun. Deep residual learning for image recognition. In CVPR, 2016. 6, 7, 8", + "[14] Jonathan Ho, Ajay Jain, and Pieter Abbeel. Denoising diffusion probabilistic models. Advances in neural information processing systems, 33:6840-6851, 2020. 4", + "[15] Shengchao Hu, Li Chen, Penghao Wu, Hongyang Li, Junchi Yan, and Dacheng Tao. St-p3: End-to-end vision-based autonomous driving via spatial-temporal feature learning. In ECCV, 2022. 8", + "[16] Yihan Hu, Jiazhi Yang, Li Chen, Keyu Li, Chonghao Sima, Xizhou Zhu, Siqi Chai, Senyao Du, Tianwei Lin, Wenhai Wang, et al. Planning-oriented autonomous driving. In CVPR, 2023. 1, 2, 6, 8", + "[17] Junjie Huang, Guan Huang, Zheng Zhu, Yun Ye, and Dalong Du. Bevdet: High-performance multi-camera 3d object detection in bird-eye-view. arXiv preprint arXiv:2112.11790, 2021. 1", + "[18] Zhiyu Huang, Zixu Zhang, Ameya Vaidya, Yuxiao Chen, Chen Lv, and Jaime Fernandez Fisac. Versatile scene-consistent traffic scenario generation as optimization with diffusion. arXiv preprint arXiv:2404.02524, 2024. 3", + "[19] Michael Janner, Yilun Du, Joshua B. Tenenbaum, and Sergey Levine. Planning with diffusion for flexible behavior synthesis. In ICLR, 2022. 1, 3", + "[20] Bo Jiang, Shaoyu Chen, Qing Xu, Bencheng Liao, Jiajie Chen, Helong Zhou, Qian Zhang, Wenyu Liu, Chang Huang, and Xinggang Wang. Vad: Vectorized scene representation for efficient autonomous driving. In ICCV, 2023. 1, 2, 6, 8", + "[21] Chiyu Jiang, Andre Cornman, Cheolho Park, Benjamin Sapp, Yin Zhou, Dragomir Anguelov, et al. Motiondiffuser: Controllable multi-agent motion prediction using diffusion. In CVPR, 2023. 3", + "[22] Napat Karnchanachari, Dimitris Geromichalos, Kok Seang Tan, Nanxiang Li, Christopher Eriksen, Shakiba Yaghoubi, Noushin Mehdipour, Gianmarco Bernasconi, Whye Kit Fong, Yiluan Guo, et al. Towards learning-based planning: The nuplan benchmark for real-world autonomous driving. In ICRA, 2024. 6", + "[23] Yingyan Li, Lue Fan, Jiawei He, Yuqi Wang, Yuntao Chen, Zhaoxiang Zhang, and Tieniu Tan. Enhancing end-to-end autonomous driving with latent world model. arXiv preprint arXiv:2406.08481, 2024. 2", + "[24] Zhiqi Li, Wenhai Wang, Hongyang Li, Enze Xie, Chonghao Sima, Tong Lu, Yu Qiao, and Jifeng Dai. Bevformer: Learning bird's-eye-view representation from multi-camera images via spatiotemporal transformers. In ECCV, 2022. 1", + "[25] Zhenxin Li, Kailin Li, Shihao Wang, Shiyi Lan, Zhiding Yu, Yishen Ji, Zhiqi Li, Ziye Zhu, Jan Kautz, Zuxuan Wu, et al. Hydra-mdp: End-to-end multimodal planning with multi-target hydra-distillation. arXiv preprint arXiv:2406.06978, 2024. 1, 2, 6, 7", + "[26] Zhiqi Li, Zhiding Yu, Shiyi Lan, Jiahan Li, Jan Kautz, Tong Lu, and Jose M Alvarez. Is ego status all you need for open-loop end-to-end autonomous driving? In CVPR, 2024. 2" + ], + "bbox": [ + 516, + 92, + 903, + 900 + ], + "page_idx": 8 + }, + { + "type": "page_number", + "text": "12045", + "bbox": [ + 480, + 944, + 517, + 955 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[27] Benchcheng Liao, Shaoyu Chen, Xinggang Wang, Tianheng Cheng, Qian Zhang, Wenyu Liu, and Chang Huang. MapTR: Structured modeling and learning for online vectorized HD map construction. In ICLR, 2023. 1", + "[28] Benchcheng Liao, Shaoyu Chen, Yunchi Zhang, Bo Jiang, Qian Zhang, Wenyu Liu, Chang Huang, and Xinggang Wang. Maptrv2: An end-to-end framework for online vectorized hd map construction. IJCV, 2024. 1", + "[29] Xuewu Lin, Tianwei Lin, Zixiang Pei, Lichao Huang, and Zhizhong Su. Sparse4d: Multi-view 3d object detection with sparse spatial-temporal fusion. arXiv preprint arXiv:2211.10581, 2022. 6", + "[30] Yaron Lipman, Ricky T. Q. Chen, Heli Ben-Hamu, Maximilian Nickel, and Matthew Le. Flow matching for generative modeling. In The Eleventh International Conference on Learning Representations, 2023. 4", + "[31] Xingchao Liu, Chengyue Gong, and Qiang Liu. Flow straight and fast: Learning to generate and transfer data with rectified flow. arXiv preprint arXiv:2209.03003, 2022. 4", + "[32] Yicheng Liu, Tianyuan Yuan, Yue Wang, Yilun Wang, and Hang Zhao. Vectormapnet: End-to-end vectorized hd map learning. In ICML, 2023. 1", + "[33] Robin Rombach, Andreas Blattmann, Dominik Lorenz, Patrick Esser, and Björn Ommer. High-resolution image synthesis with latent diffusion models. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 10684-10695, 2022. 4", + "[34] Olaf Ronneberger, Philipp Fischer, and Thomas Brox. U-net: Convolutional networks for biomedical image segmentation. In MICCAI, 2015. 2", + "[35] Jiaming Song, Chenlin Meng, and Stefano Ermon. Denoising diffusion implicit models. In ICLR, 2021. 4, 5, 6", + "[36] Yang Song, Jascha Sohl-Dickstein, Diederik P Kingma, Abhishek Kumar, Stefano Ermon, and Ben Poole. Score-based generative modeling through stochastic differential equations. In International Conference on Learning Representations, 2021. 4", + "[37] Ajay Sridhar, Dhruv Shah, Catherine Glossop, and Sergey Levine. Nomad: Goal masked diffusion policies for navigation and exploration. In ICRA, 2024. 4", + "[38] Maria Stamatopoulou, Jianwei Liu, and Dimitrios Kanoulas. Dippest: Diffusion-based path planner for synthesizing trajectories applied on quadruped robots. arXiv preprint arXiv:2405.19232, 2024. 4", + "[39] Wenchao Sun, Xuewu Lin, Yining Shi, Chuang Zhang, Hao ran Wu, and Sifa Zheng. Sparsedrive: End-to-end autonomous driving via sparse scene representation. arXiv preprint arXiv:2405.19620, 2024. 2, 6, 8", + "[40] Mingxing Tan and Quoc Le. Efficientnet: Rethinking model scaling for convolutional neural networks. In ICML, 2019. 8", + "[41] Wenwen Tong, Chonghao Sima, Tai Wang, Li Chen, Silei Wu, Hanming Deng, Yi Gu, Lewei Lu, Ping Luo, Dahua Lin, et al. Scene as occupancy. In ICCV, 2023. 8", + "[42] Yue Wang, Vitor Campagnolo Guizilini, Tianyuan Zhang, Yilun Wang, Hang Zhao, and Justin Solomon. Detr3d: 3d object detection from multi-view images via 3d-to-2d queries. In CoRL, 2022. 1, 6" + ], + "bbox": [ + 91, + 92, + 480, + 898 + ], + "page_idx": 9 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[43] Yuqi Wang, Jiawei He, Lue Fan, Hongxin Li, Yuntao Chen, and Zhaoxiang Zhang. Driving into the future: Multiview visual forecasting and planning with world model for autonomous driving. In CVPR, 2024. 2", + "[44] Yixiao Wang, Chen Tang, Lingfeng Sun, Simone Rossi, Yichen Xie, Chensheng Peng, Thomas Hannagan, Stefano Sabatini, Nicola Poerio, Masayoshi Tomizuka, et al. Optimizing diffusion models for joint trajectory prediction and controllable generation. In ECCV, 2024. 3", + "[45] Xinshuo Weng, Boris Ivanovic, Yan Wang, Yue Wang, and Marco Pavone. Para-drive: Parallelized architecture for real-time autonomous driving. In CVPR, 2024. 2, 6", + "[46] Zehang Weng, Haofei Lu, Danica Kragic, and Jens Lundell. Dexdiffuser: Generating dexterous grasps with diffusion models. arXiv preprint arXiv:2402.02989, 2024. 4", + "[47] Sixu Yan, Zeyu Zhang, Muzhi Han, Zaijin Wang, Qi Xie, Zhitian Li, Zhehan Li, Hangxin Liu, Xinggang Wang, and Song-Chun Zhu. M2diffuser: Diffusion-based trajectory optimization for mobile manipulation in 3d scenes. arXiv preprint arXiv:2410.11402, 2024. 3", + "[48] Brian Yang, Huangyuan Su, Nikolaos Gkanatsios, Tsung-Wei Ke, Ayush Jain, Jeff Schneider, and Katerina Fragkiadaki. Diffusion-es: Gradient-free planning with diffusion for autonomous driving and zero-shot instruction following. In CVPR, 2024. 3", + "[49] Jingfeng Yao and Xinggang Wang. Reconstruction vs. generation: Taming optimization dilemma in latent diffusion models. arXiv preprint arXiv:2501.01423, 2025. 4", + "[50] Jingfeng Yao, Cheng Wang, Wenyu Liu, and Xinggang Wang. Fasteredit: Towards faster diffusion transformers training without architecture modification. arXiv preprint arXiv:2410.10356, 2024. 4", + "[51] Wenhao Yu, Jie Peng, Huanyu Yang, Junrui Zhang, Yifan Duan, Jianmin Ji, and Yanyong Zhang. Ldp: A local diffusion planner for efficient robot navigation and collision avoidance. arXiv preprint arXiv:2407.01950, 2024. 4", + "[52] Chengran Yuan, Zhanqi Zhang, Jiawei Sun, Shuo Sun, Zefan Huang, Christina Dao Wen Lee, Dongen Li, Yuhang Han, Anthony Wong, Keng Peng Tee, et al. Drama: An efficient end-to-end motion planner for autonomous driving with mamba. arXiv preprint arXiv:2408.03601, 2024. 6", + "[53] Yanjie Ze, Gu Zhang, Kangning Zhang, Chenyuan Hu, Muhan Wang, and Huazhe Xu. 3d diffusion policy. In RSS, 2024. 3", + "[54] Fangao Zeng, Bin Dong, Yuang Zhang, Tiancai Wang, Xi-angyu Zhang, and Yichen Wei. Motr: End-to-end multiple-object tracking with transformer. In ECCV, 2022. 1", + "[55] Yifu Zhang, Chunyu Wang, Xinggang Wang, Wenjun Zeng, and Wenyu Liu. Fairmot: On the fairness of detection and re-identification in multiple object tracking. IJCV, 2021.", + "[56] Yifu Zhang, Peize Sun, Yi Jiang, Dongdong Yu, Fucheng Weng, Zehuan Yuan, Ping Luo, Wenyu Liu, and Xinggang Wang. Bytetrack: Multi-object tracking by associating every detection box. In ECCV, 2022. 1", + "[57] Huangjie Zheng, Pengcheng He, Weizhu Chen, and Mingyuan Zhou. Truncated diffusion probabilistic models and diffusion-based adversarial auto-encoders. In The" + ], + "bbox": [ + 516, + 92, + 903, + 898 + ], + "page_idx": 9 + }, + { + "type": "page_number", + "text": "12046", + "bbox": [ + 480, + 945, + 519, + 955 + ], + "page_idx": 9 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "Eleventh International Conference on Learning Representations, 2023. 4", + "[58] Wenzhao Zheng, Ruiqi Song, Xianda Guo, Chenming Zhang, and Long Chen. Genad: Generative end-to-end autonomous driving. In ECCV, 2024. 2", + "[59] Ziyuan Zhong, Davis Rempe, Yuxiao Chen, Boris Ivanovic, Yulong Cao, Danfei Xu, Marco Pavone, and Baishakhi Ray. Language-guided traffic simulation via scene-level diffusion. In CoRL, 2023. 3", + "[60] Ziyuan Zhong, Davis Rempe, Danfei Xu, Yuxiao Chen, Sushant Veer, Tong Che, Baishakhi Ray, and Marco Pavone. Guided conditional diffusion for controllable traffic simulation. In ICRA, 2023. 3", + "[61] Lianghui Zhu, Zilong Huang, Bencheng Liao, Jun Hao Liew, Hanshu Yan, Jiashi Feng, and Xinggang Wang. Dig: Scalable and efficient diffusion models with gated linear attention. arXiv preprint arXiv:2405.18428, 2024. 4", + "[62] Xizhou Zhu, Weijie Su, Lewei Lu, Bin Li, Xiaogang Wang, and Jifeng Dai. Deformable detr: Deformable transformers for end-to-end object detection. In ICLR, 2021. 2, 6" + ], + "bbox": [ + 91, + 90, + 482, + 375 + ], + "page_idx": 10 + }, + { + "type": "page_number", + "text": "12047", + "bbox": [ + 480, + 945, + 517, + 955 + ], + "page_idx": 10 + } +] \ No newline at end of file diff --git a/2025/DiffusionDrive_ Truncated Diffusion Model for End-to-End Autonomous Driving/186eb20a-812b-4b73-a9f1-a45f01468323_model.json b/2025/DiffusionDrive_ Truncated Diffusion Model for End-to-End Autonomous Driving/186eb20a-812b-4b73-a9f1-a45f01468323_model.json new file mode 100644 index 0000000000000000000000000000000000000000..ef20beb10a155300fa82772d208cc84839e9d178 --- /dev/null +++ b/2025/DiffusionDrive_ Truncated Diffusion Model for End-to-End Autonomous Driving/186eb20a-812b-4b73-a9f1-a45f01468323_model.json @@ -0,0 +1,2257 @@ +[ + [ + { + "type": "header", + "bbox": [ + 0.107, + 0.003, + 0.182, + 0.043 + ], + "angle": 0, + "content": "CVF" + }, + { + "type": "header", + "bbox": [ + 0.237, + 0.001, + 0.812, + 0.047 + ], + "angle": 0, + "content": "This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore." + }, + { + "type": "title", + "bbox": [ + 0.092, + 0.131, + 0.908, + 0.155 + ], + "angle": 0, + "content": "DiffusionDrive: Truncated Diffusion Model for End-to-End Autonomous Driving" + }, + { + "type": "text", + "bbox": [ + 0.101, + 0.179, + 0.895, + 0.294 + ], + "angle": 0, + "content": "Bencheng Liao\\(^{1,2,\\diamond}\\) Shaoyu Chen\\(^{2,3}\\) Haoran Yin\\(^{3}\\) Bo Jiang\\(^{2,\\diamond}\\) Cheng Wang\\(^{1,2,\\diamond}\\) Sixu Yan\\(^{2}\\) Xinbang Zhang\\(^{3}\\) Xiangyu Li\\(^{3}\\) Ying Zhang\\(^{3}\\) Qian Zhang\\(^{3}\\) Xinggang Wang\\(^{2}\\) \n\\(^{1}\\) Institute of Artificial Intelligence, Huazhong University of Science & Technology \n\\(^{2}\\) School of EIC, Huazhong University of Science & Technology \n\\(^{3}\\) Horizon Robotics \nCode & Model & Demo: hustvl/DiffusionDrive" + }, + { + "type": "title", + "bbox": [ + 0.248, + 0.327, + 0.327, + 0.343 + ], + "angle": 0, + "content": "Abstract" + }, + { + "type": "text", + "bbox": [ + 0.089, + 0.359, + 0.486, + 0.736 + ], + "angle": 0, + "content": "Recently, the diffusion model has emerged as a powerful generative technique for robotic policy learning, capable of modeling multi-mode action distributions. Leveraging its capability for end-to-end autonomous driving is a promising direction. However, the numerous denoising steps in the robotic diffusion policy and the more dynamic, open-world nature of traffic scenes pose substantial challenges for generating diverse driving actions at a real-time speed. To address these challenges, we propose a novel truncated diffusion policy that incorporates prior multi-mode anchors and truncates the diffusion schedule, enabling the model to learn denoising from anchored Gaussian distribution to the multi-mode driving action distribution. Additionally, we design an efficient cascade diffusion decoder for enhanced interaction with conditional scene context. The proposed model, DiffusionDrive, demonstrates \\(10 \\times\\) reduction in denoising steps compared to vanilla diffusion policy, delivering superior diversity and quality in just 2 steps. On the planning-oriented NAVSIM dataset, with aligned ResNet-34 backbone, DiffusionDrive achieves 88.1 PDMS without bells and whistles, setting a new record, while running at a real-time speed of 45 FPS on an NVIDIA 4090. Qualitative results on challenging scenarios further confirm that DiffusionDrive can robustly generate diverse plausible driving actions." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.765, + 0.222, + 0.781 + ], + "angle": 0, + "content": "1. Introduction" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.79, + 0.484, + 0.867 + ], + "angle": 0, + "content": "End-to-end autonomous driving has gained significant attention in recent years due to advancements in perception models (detection [5, 17, 24, 42], tracking [54-56], online mapping [27, 28, 32], etc.), which directly learns the driving policy from the raw sensor inputs. This data-driven ap" + }, + { + "type": "image", + "bbox": [ + 0.516, + 0.326, + 0.909, + 0.502 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.513, + 0.504, + 0.907, + 0.559 + ], + "angle": 0, + "content": "Figure 1. The comparison of different end-to-end paradigms. (a) Single mode regression [8, 16, 20]. (b) Sampling from vocabulary [4, 25]. (c) Vanilla diffusion policy [7, 19]. (d) The proposed truncated diffusion policy." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.562, + 0.907, + 0.607 + ], + "angle": 0, + "content": "proach offers a scalable and robust alternative to traditional rule-based motion planning, which often struggles to generalize to complex real-world driving settings." + }, + { + "type": "text", + "bbox": [ + 0.511, + 0.608, + 0.909, + 0.865 + ], + "angle": 0, + "content": "To effectively learn from data, mainstream end-to-end planners (e.g., Transfuser [8], UniAD [16], VAD [20]) typically regress a single-mode trajectory from an ego-query as shown in Fig. 1a. However, this paradigm does not account for the inherent uncertainty and multi-mode nature of driving behaviors. Recently, VADv2 [20] introduces a large fixed vocabulary of anchor trajectories (4096 anchors) to discretize the continuous action space and capture a broader range of driving behaviors, and then samples from these anchors based on predicted scores as shown in Fig. 1b. However, this large fixed-vocabulary paradigm is fundamentally constrained by the number and quality of anchor trajectories, often failing in out-of-vocabulary scenarios. Furthermore, managing a large number of anchors presents significant computational challenges for real-time applications. Rather than discretizing the action space, diffusion model [7] has proven to be a powerful genera" + }, + { + "type": "page_footnote", + "bbox": [ + 0.513, + 0.875, + 0.908, + 0.901 + ], + "angle": 0, + "content": "To distinguish the term \"multimodal\" used to describe input data, we use \"multi-mode\" in this paper to refer to diverse planning decisions." + }, + { + "type": "page_footnote", + "bbox": [ + 0.091, + 0.875, + 0.483, + 0.901 + ], + "angle": 0, + "content": "\\(\\diamond\\) Intern of Horizon Robotics; \\(\\boxtimes\\) Corresponding author: Xinggang Wang (xgwang@hust.edu.cn)." + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "12037" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.09, + 0.092, + 0.482, + 0.137 + ], + "angle": 0, + "content": "tive decision-making policy in the robotics domain, which can directly sample multi-mode physically plausible actions from a Gaussian distribution via iterative denoising process." + }, + { + "type": "text", + "bbox": [ + 0.089, + 0.139, + 0.483, + 0.381 + ], + "angle": 0, + "content": "This inspires us to replicate the success of the diffusion model in the robotics domain to end-to-end autonomous driving. We apply the vanilla robotic diffusion policy to the well-known single-mode-regression method, Transfuser [8], by proposing a variant, \\( \\text{Transfuser}_{\\text{DP}} \\), which replaces the deterministic MLP regression head with a conditional diffusion model [34]. Though \\( \\text{Transfuser}_{\\text{DP}} \\) improves planning performance, two major issues arise: 1) The numerous 20 denoising steps in the vanilla DDIM diffusion policy introduce heavy computational consumption during inference as shown in Tab. 2, hindering the real-time application for autonomous driving. 2) The trajectories sampled from different Gaussian noises severely overlap with each other, as illustrated in Fig. 2. This underscores the non-trivial challenge of taming the diffusion models for the dynamic and open-world traffic scenes." + }, + { + "type": "text", + "bbox": [ + 0.093, + 0.384, + 0.483, + 0.685 + ], + "angle": 0, + "content": "Unlike the vanilla diffusion policy, which samples actions from a random Gaussian noise conditioned on scene context, human drivers adhere to established driving patterns that they dynamically adjust in response to real-time traffic conditions. This insight motivates us to embed these prior driving patterns into the diffusion policy by partitioning the Gaussian distribution into multiple sub-Gaussian distributions centered around prior anchors, referred to as anchored Gaussian distribution. It is implemented by truncating the diffusion schedule to introduce a small portion of Gaussian noise around the prior anchors as shown in Fig. 3. Thanks to the multi-mode distributional expressivity of the diffusion model, the proposed truncated diffusion policy effectively covers the potential action space without requiring a large set of fixed anchors, as VADv2 does. With more reasonable initial noise samples from the anchored Gaussian distribution, we can truncate the denoising process, reducing the required steps from 20 to just 2—a substantial speedup that satisfies the real-time requirements of autonomous driving." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.687, + 0.483, + 0.824 + ], + "angle": 0, + "content": "To enhance the interaction with conditional scene context, we propose an efficient transformer-based diffusion decoder that interacts not only with structured queries from the perception module but also with Bird's Eye View (BEV) and perspective view (PV) features through a sparse deformable attention mechanism [3, 62]. Additionally, we introduce a cascade mechanism to iteratively refine the trajectory reconstruction within the diffusion decoder at each denoising step." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.825, + 0.483, + 0.901 + ], + "angle": 0, + "content": "With these innovations, we present DiffusionDrive, a diffusion model for real-time end-to-end autonomous driving. We benchmark our method on the planning-oriented NAVSIM dataset [11] using non-reactive simulation and closed-loop evaluations. Without bells and whis" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.092, + 0.905, + 0.333 + ], + "angle": 0, + "content": "tles, DiffusionDrive achieves 88.1 PDMS on NAVSIM navtest split with the aligned ResNet-34 backbone, significantly outperforming previous state-of-the-art methods. Even compared to the NAVSIM challenge-winning solution Hydra-MDP-\\(\\nu_{8192}\\)-W-EP [25], which follows VADv2 with 8192 anchor trajectories and further incorporates post-processing and additional supervision, DiffusionDrive still outperforms it by 1.6 PDMS through directly learning from human demonstrations and inferring without post-processing, while running at real-time speed of 45 FPS on an NVIDIA 4090. We further validate the superiority of DiffusionDrive on popular nuScenes dataset [2] with open-loop evaluations, DiffusionDrive runs \\(1.8 \\times\\) faster than VAD and outperforms it [20] by \\(20.8\\%\\) lower L2 error and \\(63.6\\%\\) lower collision rate with the same ResNet-50 backbone, demonstrating state-of-the-art planning performance." + }, + { + "type": "text", + "bbox": [ + 0.533, + 0.334, + 0.858, + 0.347 + ], + "angle": 0, + "content": "Our contributions can be summarized as follows:" + }, + { + "type": "text", + "bbox": [ + 0.514, + 0.348, + 0.905, + 0.424 + ], + "angle": 0, + "content": "- We firstly introduce the diffusion model to the field of end-to-end autonomous driving and propose a novel truncated diffusion policy to address the issues of mode collapse and heavy computational overhead found in direct adaptation of vanilla diffusion policy to the traffic scene." + }, + { + "type": "text", + "bbox": [ + 0.515, + 0.424, + 0.905, + 0.469 + ], + "angle": 0, + "content": "- We design an efficient transformer-based diffusion decoder that interacts with the conditional information in a cascaded manner for better trajectory reconstruction." + }, + { + "type": "text", + "bbox": [ + 0.515, + 0.47, + 0.905, + 0.543 + ], + "angle": 0, + "content": "- Without bells and whistles, DiffusionDrive significantly outperforms previous state-of-the-art methods, achieving a record-breaking 88.1 PDMS on the NAVSIM navtest split with the same backbone, while maintaining real-time performance at 45 FPS on an NVIDIA 4090." + }, + { + "type": "text", + "bbox": [ + 0.515, + 0.545, + 0.905, + 0.605 + ], + "angle": 0, + "content": "- We qualitatively demonstrate that DiffusionDrive can generate more diverse and plausible trajectories, exhibiting high-quality multi-mode driving actions in various challenging scenarios." + }, + { + "type": "list", + "bbox": [ + 0.514, + 0.348, + 0.905, + 0.605 + ], + "angle": 0, + "content": null + }, + { + "type": "title", + "bbox": [ + 0.514, + 0.617, + 0.655, + 0.633 + ], + "angle": 0, + "content": "2. Related Work" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.644, + 0.907, + 0.901 + ], + "angle": 0, + "content": "End-to-end autonomous driving. UniAD [16], as a pioneering work, demonstrates the potential of end-to-end autonomous driving by integrating multiple perception tasks to enhance planning performance. VAD [20] further explores the use of compact vectorized scene representations to improve efficiency. Subsequently, a series of works [6, 8, 12, 23, 26, 43, 45, 58] have adopted the single-trajectory planning paradigm to enhance planning performance further. More recently, VADv2 [4] shifts the paradigm towards multi-mode planning by scoring and sampling from a large fixed vocabulary of anchor trajectories. Hydra-MDP [25] improves the scoring mechanism of VADv2 by introducing extra supervision from a rule-based scorer. SparseDrive [39] explores an alternative BEV-free solution. Unlike existing multi-mode planning approaches, we propose a novel paradigm that leverages powerful generative diffusion models for end-to-end autonomous driving." + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "12038" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.171, + 0.086, + 0.827, + 0.594 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.334, + 0.597, + 0.663, + 0.608 + ], + "angle": 0, + "content": "(b) Top-1's turning left and diverse top-10's lane changing." + }, + { + "type": "image_caption", + "bbox": [ + 0.09, + 0.608, + 0.907, + 0.689 + ], + "angle": 0, + "content": "Figure 2. Qualitative comparison of Transfuser, TransfuserDP and DiffusionDrive on challenging scenes of NAVSIM navtest split. With the same inputs from front cameras and LiDAR, DiffusionDrive achieves the highest planning quality of top-1 scoring trajectory as illustrated in Tab. 2. We render the highlighted diverse trajectories predicted by DiffusionDrive in the front view. (a) and (b) shows that the top-1 scoring trajectory of DiffusionDrive closely matches the ground truth for both going straight and turning left. Additionally, DiffusionDrive's top-10 scoring trajectory demonstrates high-quality lane changing—an ability not observed in multi-mode TransfuserDP and impossible for Transfuser." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.695, + 0.486, + 0.892 + ], + "angle": 0, + "content": "Diffusion model for traffic simulation. Driving diffusion policy has been explored in the traffic simulation by leveraging only abstract perception groundtruth [9, 18, 21, 44]. MotionDiffuser [21] and CTG [60] are pioneering applications of diffusion models for multi-agent motion prediction, using a conditional diffusion model to sample target trajectories from Gaussian noise. \\(\\mathrm{CTG + + }\\) [59] further incorporates a large language model (LLM) for language-driven guidance, improving usability and enabling realistic traffic simulations. Diffusion-ES [48] replaces reward-gradient-guided denoising with evolutionary search. Moving beyond diffusion models limited to traffic simulation with perception groundtruth, our approach unlocks the potential of dif" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.695, + 0.907, + 0.74 + ], + "angle": 0, + "content": "fusion models for real-time, end-to-end autonomous driving through our proposed truncated diffusion policy and efficient diffusion decoder." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.75, + 0.909, + 0.903 + ], + "angle": 0, + "content": "Diffusion model for robotic policy learning. Diffusion policy [7] demonstrates the great potential in robotic policy learning, effectively capturing multi-mode action distributions and high-dimensional action spaces. Diffuser [19] proposes an unconditional diffusion model for trajectory sampling, incorporating techniques such as classifier-free guidance and image inpainting to achieve guided sampling. Subsequently, numerous works have applied diffusion models to various robotic tasks, including stationary manipulation [1, 53], mobile manipulation [47], autonomous nav" + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.52, + 0.958 + ], + "angle": 0, + "content": "12039" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.383, + 0.092, + 0.61, + 0.11 + ], + "angle": 0, + "content": "Diffusion step: \\(q(\\tau^i |\\tau^{i - 1}),i\\in [1,T]\\)" + }, + { + "type": "image", + "bbox": [ + 0.172, + 0.113, + 0.827, + 0.245 + ], + "angle": 0, + "content": null + }, + { + "type": "text", + "bbox": [ + 0.375, + 0.251, + 0.626, + 0.268 + ], + "angle": 0, + "content": "Denoising step: \\(p_{\\theta}\\big(\\tau^i |\\tau^{i - 1},z\\big),i\\in [1,T]\\)" + }, + { + "type": "title", + "bbox": [ + 0.302, + 0.277, + 0.696, + 0.291 + ], + "angle": 0, + "content": "Vanilla Diffusion Policy vs. Truncated Diffusion Policy (ours)" + }, + { + "type": "text", + "bbox": [ + 0.212, + 0.302, + 0.471, + 0.318 + ], + "angle": 0, + "content": "Diffusion step: \\(q(\\tau^i |\\tau^{i - 1}),i\\in [1,T_{\\mathrm{trunc}}]\\)" + }, + { + "type": "image", + "bbox": [ + 0.177, + 0.321, + 0.48, + 0.453 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.184, + 0.455, + 0.479, + 0.466 + ], + "angle": 0, + "content": "Adding noise to the anchors with truncated diffusion schedule." + }, + { + "type": "text", + "bbox": [ + 0.516, + 0.304, + 0.82, + 0.316 + ], + "angle": 0, + "content": "Denosing from anchored Gaussian distribution with fewer steps." + }, + { + "type": "image", + "bbox": [ + 0.498, + 0.316, + 0.822, + 0.453 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.525, + 0.458, + 0.808, + 0.474 + ], + "angle": 0, + "content": "Denoising step: \\(p_{\\theta}(\\tau^i |\\tau^{i - 1},z),i\\in [1,T_{\\mathrm{trunc}}]\\)" + }, + { + "type": "image_caption", + "bbox": [ + 0.089, + 0.482, + 0.907, + 0.539 + ], + "angle": 0, + "content": "Figure 3. Illustration of truncated diffusion policy by comparing with vanilla diffusion policy. We truncate the diffusion process and only add a small portion of Gaussian noise to diffuse the anchor trajectories. Then, we train the diffusion model to reconstruct the ground-truth trajectory from the anchored Gaussian distribution with conditional scene context. During the inference, we also truncate the denoising process by starting from the better samples in the anchored Gaussian distribution than the pure Gaussian noise." + }, + { + "type": "text", + "bbox": [ + 0.089, + 0.543, + 0.483, + 0.679 + ], + "angle": 0, + "content": "igation [37, 51], quadruped locomotion [38], and dexterous manipulation [46]. However, directly applying vanilla diffusion policy to end-to-end autonomous driving poses unique challenges, as it requires real-time efficiency and the generation of plausible multi-mode trajectories in dynamic and open-world traffic scenes. In this work, we propose a novel truncated diffusion policy to address these challenges, introducing concepts that have not yet been explored in the robotics field." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.689, + 0.483, + 0.901 + ], + "angle": 0, + "content": "Diffusion model for image generation. Diffusion models have been extensively adopted for image generation tasks [33, 36, 49, 50, 61]. DDIM [35] enhances DDPM [14] by enabling efficient sampling with significantly fewer steps based on non-Markovian diffusion processes. Flow matching [30, 31] further optimizes the generative process by directly modeling continuous probability flows. TDPM [57] proposes truncated denoising, which initiates the generation process from an implicit intermediate distribution to accelerate sampling. In contrast to these approaches, our method introduces an explicit driving prior within the diffusion policy, effectively guiding the diffusion process toward more accurate and efficient generation tailored specifically for end-to-end autonomous driving." + }, + { + "type": "title", + "bbox": [ + 0.514, + 0.542, + 0.605, + 0.557 + ], + "angle": 0, + "content": "3. Method" + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.566, + 0.644, + 0.582 + ], + "angle": 0, + "content": "3.1. Preliminary" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.59, + 0.907, + 0.683 + ], + "angle": 0, + "content": "Task formulation. End-to-end autonomous driving takes raw sensor data as input and predicts the future trajectory of the ego-vehicle. The trajectory is represented as a sequence of waypoints \\(\\tau = \\{(x_{t},y_{t})\\}_{t = 1}^{T_{f}}\\), where \\(T_{f}\\) denotes the planning horizon, and \\((x_{t},y_{t})\\) is the location of each waypoint at time \\(t\\) in the current ego-vehicle coordinate system." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.684, + 0.906, + 0.73 + ], + "angle": 0, + "content": "Conditional diffusion model. The conditional diffusion model poses a forward diffusion process as gradually adding noise to the data sample, which can be defined as:" + }, + { + "type": "equation", + "bbox": [ + 0.565, + 0.741, + 0.907, + 0.767 + ], + "angle": 0, + "content": "\\[\nq \\left(\\tau^ {i} \\mid \\tau^ {0}\\right) = \\mathcal {N} \\left(\\tau^ {i}; \\sqrt {\\bar {\\alpha} ^ {i}} \\boldsymbol {\\tau} ^ {0}, \\left(1 - \\bar {\\alpha} ^ {i}\\right) \\mathbf {I}\\right), \\tag {1}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.778, + 0.907, + 0.901 + ], + "angle": 0, + "content": "where \\(\\tau^0\\) is the clean data sample, and \\(\\tau^i\\) is the data sample with noise at time \\(i\\) (Note: we use superscript \\(i\\) to denote diffusion timestep). The constant \\(\\bar{\\alpha}^i = \\prod_{s=1}^i \\alpha^s = \\prod_{s=1}^i (1 - \\beta^s)\\) and \\(\\beta^s\\) is the noise schedule. We train the reverse process model \\(f_\\theta(\\tau^i, z, i)\\) to predict \\(\\tau^0\\) from \\(\\tau^i\\) with the guidance of conditional information \\(z\\), where \\(\\theta\\) is the trainable model parameter. During inference, the trained diffusion model \\(f_\\theta\\) progressively refines from the random" + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.52, + 0.957 + ], + "angle": 0, + "content": "12040" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.092, + 0.088, + 0.516, + 0.26 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.52, + 0.09, + 0.91, + 0.262 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.09, + 0.265, + 0.908, + 0.308 + ], + "angle": 0, + "content": "Figure 4. Overall architecture of DiffusionDrive. (a) DiffusionDrive can integrate various existing perception modules and sensor inputs. (b) The designed diffusion decoder takes the sampled noisy trajectories from anchored Gaussian distribution as input and progressively denoises them with enhanced interactions with the conditional scene context in a cascade manner to generate the final predictions." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.311, + 0.486, + 0.358 + ], + "angle": 0, + "content": "noise \\(\\tau^T\\) sampled in Gaussian distribution to the predicted clean data sample \\(\\tau^0\\) with the guidance of conditional information \\(z\\), which is defined as:" + }, + { + "type": "equation", + "bbox": [ + 0.101, + 0.368, + 0.484, + 0.41 + ], + "angle": 0, + "content": "\\[\np _ {\\theta} \\left(\\boldsymbol {\\tau} ^ {0} \\mid z\\right) = \\int p \\left(\\boldsymbol {\\tau} ^ {T}\\right) \\prod_ {i = 1} ^ {T} p _ {\\theta} \\left(\\boldsymbol {\\tau} ^ {i - 1} \\mid \\boldsymbol {\\tau} ^ {i}, z\\right) d \\boldsymbol {\\tau} ^ {1: T}. \\tag {2}\n\\]" + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.421, + 0.228, + 0.436 + ], + "angle": 0, + "content": "3.2. Investigation" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.445, + 0.484, + 0.581 + ], + "angle": 0, + "content": "Turn Transfuser [8] into conditional diffusion model. We begin from the representative deterministic end-to-end planner Transfuser [8] and turn it into a generative model \\(\\mathrm{Transfuser}_{\\mathrm{DP}}\\) by simply replacing the regression MLP layers with the conditional diffusion model UNet following vanilla diffusion policy [7]. During the evaluation, we sample a random noise and progressively refine it with 20 steps. Tab. 2 shows that \\(\\mathrm{Transfuser}_{\\mathrm{DP}}\\) achieves better planning quality than deterministic Transfuser." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.583, + 0.484, + 0.732 + ], + "angle": 0, + "content": "Mode collapse. To further investigate the multi-mode property of the vanilla diffusion policy in driving, we sampled 20 random noises from Gaussian distribution and denoised them using 20 steps. As shown in Fig. 2, the different random noises converge to similar trajectories after the denoising process. To quantitatively analyze the phenomenon of mode collapse, we define a mode diversity score \\(\\mathcal{D}\\) based on the mean Intersection over Union (mIoU) between each denoised trajectory and the union of all denoised trajectories:" + }, + { + "type": "equation", + "bbox": [ + 0.16, + 0.731, + 0.483, + 0.773 + ], + "angle": 0, + "content": "\\[\n\\mathcal {D} = 1 - \\frac {1}{N} \\sum_ {i = 1} ^ {N} \\frac {\\operatorname {A r e a} \\left(\\tau_ {i} \\cap \\bigcup_ {j = 1} ^ {N} \\tau_ {j}\\right)}{\\operatorname {A r e a} \\left(\\tau_ {i} \\cup \\bigcup_ {j = 1} ^ {N} \\tau_ {j}\\right)}, \\tag {3}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.778, + 0.484, + 0.869 + ], + "angle": 0, + "content": "where \\(\\tau_{i}\\) represents the \\(i\\)-th denoised trajectory, \\(N\\) is the total number of sampled trajectories and \\(\\bigcup_{j=1}^{N} \\tau_{j}\\) is the union of all denoised trajectories. A higher mIoU indicates less diversity of the denoised trajectories. The quantitative mode diversity results in Tab. 2 further validate the observations presented in Fig. 2." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.871, + 0.484, + 0.902 + ], + "angle": 0, + "content": "Heavy denoising overhead. The DDIM [35] diffusion policy requires 20 denoising steps to transform random noise" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.312, + 0.907, + 0.373 + ], + "angle": 0, + "content": "into a feasible trajectory, which introduces significant computational overhead, reducing the FPS from 60 to 7, as shown in Tab. 2, and making it impractical for real-time online driving applications." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.382, + 0.706, + 0.396 + ], + "angle": 0, + "content": "3.3. Truncated Diffusion" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.403, + 0.907, + 0.539 + ], + "angle": 0, + "content": "Human driving follows fixed patterns, unlike the random noise denoising in vanilla diffusion policy. Motivated by this, we propose a truncated diffusion policy that begins the denoising process from an anchored Gaussian distribution instead of a standard Gaussian distribution. To enable the model to learn to denoise from the anchored Gaussian distribution to the desired driving policy, we further truncate the diffusion schedule during training, adding only a small amount of Gaussian noise to the anchors." + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.542, + 0.907, + 0.62 + ], + "angle": 0, + "content": "Training. We first construct the diffusion process by adding Gaussian noise to anchors \\(\\{\\mathbf{a}_k\\}_{k=1}^{N_{\\text{anchor}}}\\) clustered by K-Means on the training set, where \\(\\mathbf{a}_k = \\{(x_t, y_t)\\}_{t=1}^{T_f}\\). We truncate the diffusion noise schedule to diffuse the anchors to the anchored Gaussian distribution:" + }, + { + "type": "equation", + "bbox": [ + 0.566, + 0.629, + 0.907, + 0.648 + ], + "angle": 0, + "content": "\\[\n\\tau_ {k} ^ {i} = \\sqrt {\\bar {\\alpha} ^ {i}} \\mathbf {a} _ {k} + \\sqrt {1 - \\bar {\\alpha} ^ {i}} \\boldsymbol {\\epsilon}, \\quad \\boldsymbol {\\epsilon} \\sim \\mathcal {N} (0, \\mathbf {I}), \\qquad (4)\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.658, + 0.906, + 0.688 + ], + "angle": 0, + "content": "where \\( i \\in [1, T_{\\mathrm{trunc}}] \\) and \\( T_{\\mathrm{trunc}} \\ll T \\) is the truncated diffusion steps." + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.689, + 0.906, + 0.739 + ], + "angle": 0, + "content": "During training, the diffusion decoder \\( f_{\\theta} \\) takes as input \\( N_{\\mathrm{anchor}} \\) noisy trajectories \\( \\{\\tau_k^i\\}_{k=1}^{N_{\\mathrm{anchor}}} \\) and predicts classification scores \\( \\{\\hat{s}_k\\}_{k=1}^{N_{\\mathrm{anchor}}} \\) and denoised trajectories \\( \\{\\hat{\\tau}_k\\}_{k=1}^{N_{\\mathrm{anchor}}} \\):" + }, + { + "type": "equation", + "bbox": [ + 0.596, + 0.746, + 0.907, + 0.767 + ], + "angle": 0, + "content": "\\[\n\\left\\{\\hat {s} _ {k}, \\hat {\\tau} _ {k} \\right\\} _ {k = 1} ^ {N _ {\\text {a n c h o r}}} = f _ {\\theta} \\left(\\left\\{\\tau_ {k} ^ {i} \\right\\} _ {k = 1} ^ {N _ {\\text {a n c h o r}}}, z\\right), \\tag {5}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.776, + 0.907, + 0.852 + ], + "angle": 0, + "content": "where \\( z \\) represents the conditional information. We assign the noisy trajectory around the closest anchor to the ground truth trajectory \\( \\tau_{\\mathrm{gt}} \\) as positive sample \\( (y_k = 1) \\) and others as negative samples \\( (y_k = 0) \\). The training objective combines trajectory reconstruction and classification:" + }, + { + "type": "equation", + "bbox": [ + 0.563, + 0.863, + 0.906, + 0.904 + ], + "angle": 0, + "content": "\\[\n\\mathcal {L} = \\sum_ {k = 1} ^ {N _ {\\text {a n c h o r}}} \\left[ y _ {k} \\mathcal {L} _ {\\text {r e c}} \\left(\\hat {\\tau} _ {k}, \\tau_ {\\mathrm {g t}}\\right) + \\lambda \\mathrm {B C E} \\left(\\hat {s} _ {k}, y _ {k}\\right) \\right], \\tag {6}\n\\]" + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.518, + 0.957 + ], + "angle": 0, + "content": "12041" + } + ], + [ + { + "type": "table", + "bbox": [ + 0.094, + 0.089, + 0.905, + 0.263 + ], + "angle": 0, + "content": "
MethodInputImg. BackboneAnchorNC↑DAC ↑TTC ↑Comf. ↑EP ↑PDMS ↑
UniAD [16]CameraResNet-34 [13]097.891.992.910078.883.4
PARA-Drive [45]CameraResNet-34 [13]097.992.493.099.879.384.0
LTF [8]CameraResNet-34 [13]097.492.892.410079.083.8
Transfuser [8]C & LResNet-34 [13]097.792.892.810079.284.0
DRAMA [52]C & LResNet-34 [13]098.093.194.810080.185.5
VADv2- \\( \\mathcal{V}_{8192} \\) [4]C & LResNet-34 [13]819297.289.191.610076.080.9
Hydra-MDP- \\( \\mathcal{V}_{8192} \\) [25]C & LResNet-34 [13]819297.991.792.910077.683.0
Hydra-MDP- \\( \\mathcal{V}_{8192} \\)-W-EP [25]C & LResNet-34 [13]819298.396.094.610078.786.5
DiffusionDrive (Ours)C & LResNet-34 [13]2098.296.294.710082.288.1
" + }, + { + "type": "table_caption", + "bbox": [ + 0.09, + 0.264, + 0.907, + 0.334 + ], + "angle": 0, + "content": "Table 1. Comparison on planning-oriented NAVSIM navtest split with closed-loop metrics. “C & L” denotes the use of both camera and LiDAR as sensor inputs. “\\(\\mathcal{V}_{8192}\\)” denotes 8192 anchors. “Hydra-MDP-\\(\\mathcal{V}_{8192}\\)-W-EP” is a variant of Hydra-MDP [25], which is further trained to fit the EP evaluation metric with additional supervision from the rule-based evaluator and uses weighted confidence post-processing. DiffusionDrive simply learns from human demonstrations and infers without post-processing. The best and the second best results are denoted by bold and underline." + }, + { + "type": "table", + "bbox": [ + 0.095, + 0.335, + 0.898, + 0.439 + ], + "angle": 0, + "content": "
MethodNC↑DAC↑TTC↑Comf.↑EP↑PDMS↑Arch.Plan Module TimeTotal ↓D↑Para.↓FPS↑
Step Time↓Steps ↓
Transfuser97.792.892.810079.284.0MLP0.2ms10.2ms0%56M60
TransfuserDP97.593.792.710079.484.6+0.6UNet6.5ms20130.0ms11%101M7
TransfuserTD97.994.293.910080.285.7+1.7UNet6.9ms213.8ms70%102M27
DiffusionDrive98.296.294.710082.288.1+4.1Dec.3.8ms27.6ms74%60M45
" + }, + { + "type": "table_caption", + "bbox": [ + 0.09, + 0.44, + 0.907, + 0.485 + ], + "angle": 0, + "content": "Table 2. Roadmap from Transfuser to DiffusionDrive on NAVSIM navtest split. \"TransfuserDP\" denotes Transfuser with vanilla DDIM diffusion policy [7]. \"TransfuserTD\" denotes Transfuser with truncated diffusion policy. \"Step Time\" denotes the runtime of each denoising step. \"FPS\" and runtime are measured on an NVIDIA 4090 GPU. \"D\" denotes the mode diversity score defined in Eq. (3)." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.488, + 0.483, + 0.518 + ], + "angle": 0, + "content": "where \\(\\lambda\\) balances the simple L1 reconstruction loss \\(\\mathcal{L}_{\\mathrm{rec}}\\) and binary cross-entropy (BCE) classification loss." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.52, + 0.483, + 0.657 + ], + "angle": 0, + "content": "Inference. We use a truncated denoising process that starts with noisy trajectories sampled from the anchored Gaussian distribution and progressively denoises them to final predictions. At each denoising timestep, the estimated trajectories from the previous step are passed to the diffusion decoder \\( f_{\\theta} \\), which predicts classification scores \\( \\{\\hat{s}_k\\}_{k=1}^{N_{\\mathrm{infer}}} \\) and coordinates \\( \\{\\hat{\\tau}_k\\}_{k=1}^{N_{\\mathrm{infer}}} \\). After obtaining the current timestep's predictions, we apply the DDIM [35] update rule to sample trajectories for the next timestep." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.659, + 0.483, + 0.751 + ], + "angle": 0, + "content": "Inference flexibility. A key advantage of our approach lies in its inference flexibility. While the model is trained with \\( N_{\\text{anchor}} \\) trajectories, the inference process can accommodate an arbitrary number of trajectory samples \\( N_{\\text{infer}} \\), where \\( N_{\\text{infer}} \\) can be dynamically adjusted based on computational resources or application requirements." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.757, + 0.226, + 0.771 + ], + "angle": 0, + "content": "3.4. Architecture" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.779, + 0.483, + 0.884 + ], + "angle": 0, + "content": "The overall architecture of our proposed method, DiffusionDrive, is illustrated in Fig. 4. DiffusionDrive can integrate various existing perception modules used in previous end-to-end planners [8, 16, 20, 39] and take different sensor inputs. The designed diffusion decoder is tailored for the complex and challenging driving application, which has enhanced interactions with the conditional scene context." + }, + { + "type": "text", + "bbox": [ + 0.091, + 0.886, + 0.483, + 0.902 + ], + "angle": 0, + "content": "Diffusion decoder. Given the set of sampled noisy trajectories" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.487, + 0.907, + 0.76 + ], + "angle": 0, + "content": "ries \\(\\{\\hat{\\tau}_k\\}_{k=1}^{N_{\\mathrm{infer}}}\\) from the anchored Gaussian distribution, we begin by applying deformable spatial cross-attention [29, 42, 62] to interact with Bird's Eye View (BEV) or Perspective View (PV) features based on the trajectory coordinates. Subsequently, cross-attention is performed between the trajectory features and the agent/map queries derived from the perception module, followed by a feed-forward network (FFN). To encode the diffusion timestep information, we utilize a Timestep Modulation layer, which is followed by a Multi-Layer Perceptron (MLP) that predicts the confidence score and the offset relative to the initial noisy trajectory coordinates. The output from this diffusion decoder layer serves as the input for the subsequent cascade diffusion decoder layer. DiffusionDrive further reuses the cascade diffusion decoder to iteratively denoise the trajectory during inference, with parameters shared across the different denoising timesteps. The final trajectory with the highest confidence score is selected as the output." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.775, + 0.64, + 0.793 + ], + "angle": 0, + "content": "4. Experiment" + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.801, + 0.611, + 0.815 + ], + "angle": 0, + "content": "4.1. Dataset" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.825, + 0.907, + 0.903 + ], + "angle": 0, + "content": "NAVSIM. The NAVSIM dataset [11] is a real-world planning-oriented dataset builds upon OpenScene [10], a compact redistribution of nuPlan [22], the largest publicly available annotated driving dataset. NAVSIM leverages eight cameras to achieve a full \\(360^{\\circ}\\) FOV, along with a" + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.52, + 0.958 + ], + "angle": 0, + "content": "12042" + } + ], + [ + { + "type": "table", + "bbox": [ + 0.094, + 0.089, + 0.905, + 0.23 + ], + "angle": 0, + "content": "
IDUNet DecoderEgo Query InteractionSpatial Cross-attnAgent/Map Cross-attnCascade DecoderParam.↓NC↑DAC↑Planning MetricPDMS↑
TTC↑Comf.↑EP↑
1✓✓XXX102M97.994.293.910080.285.7
2X✓XXX57M88.783.280.084.843.355.1
3X✓✓XX58M98.295.494.410081.387.1
4X✓X✓X58M97.993.593.810079.885.1
5X✓✓✓X59M98.095.894.410081.787.4
6X✓✓✓✓60M98.296.294.710082.288.1
" + }, + { + "type": "table_caption", + "bbox": [ + 0.09, + 0.234, + 0.905, + 0.276 + ], + "angle": 0, + "content": "Table 3. Ablation for design choices. \"Cascade Decoder\" indicates that we stack 2 cascade diffusion decoder layers. ID-1 refers to TransfuserTD in Tab. 2, utilizing conditional UNet and interaction with the ego-query, which Transfuser uses to directly regress the single-mode trajectory." + }, + { + "type": "table", + "bbox": [ + 0.094, + 0.279, + 0.357, + 0.337 + ], + "angle": 0, + "content": "
StepsParam.NCDACTTCComf.EPPDMS
160M98.396.094.710082.187.9
260M98.296.294.710082.288.1
360M98.296.394.710092.288.1
" + }, + { + "type": "table_caption", + "bbox": [ + 0.123, + 0.339, + 0.325, + 0.352 + ], + "angle": 0, + "content": "Table 4. Denoising step number." + }, + { + "type": "table", + "bbox": [ + 0.368, + 0.279, + 0.634, + 0.337 + ], + "angle": 0, + "content": "
StagesParam.NCDACTTCComf.EPPDMS
159M98.095.894.410081.787.4
260M98.296.294.710082.288.1
465M98.496.294.910082.488.2
" + }, + { + "type": "table_caption", + "bbox": [ + 0.42, + 0.339, + 0.573, + 0.352 + ], + "angle": 0, + "content": "Table 5. Cascade stages." + }, + { + "type": "table", + "bbox": [ + 0.642, + 0.279, + 0.901, + 0.337 + ], + "angle": 0, + "content": "
N_inferParam.NCDACTTCComf.EPPDMS
1060M97.993.593.110080.084.9
2060M98.296.294.710082.288.1
4060M98.596.294.810082.588.2
" + }, + { + "type": "table_caption", + "bbox": [ + 0.64, + 0.339, + 0.9, + 0.352 + ], + "angle": 0, + "content": "Table 6. Number of sampled noises \\( {N}_{\\text{infer }} \\) ." + }, + { + "type": "text", + "bbox": [ + 0.089, + 0.364, + 0.483, + 0.47 + ], + "angle": 0, + "content": "merged LiDAR point cloud derived from five sensors. Annotations are provided at a frequency of \\(2\\mathrm{Hz}\\) and include both HD maps and object bounding boxes. The dataset is designed to emphasize challenging driving scenarios involving dynamic changes in driving intentions, while deliberately excluding trivial situations such as stationary scenes or constant-speed driving." + }, + { + "type": "text", + "bbox": [ + 0.089, + 0.471, + 0.483, + 0.577 + ], + "angle": 0, + "content": "NAVSIM benchmarks planning performance using non-reactive simulations and closed-loop metrics for comprehensive evaluation. In this paper, we employ the proposed PDM score (PDMS) [11], which is a weighted combination of several sub-scores: no at-fault collisions (NC), drivable area compliance (DAC), time-to-collision (TTC), comfort (Comf.), and ego progress (EP)." + }, + { + "type": "title", + "bbox": [ + 0.09, + 0.591, + 0.303, + 0.607 + ], + "angle": 0, + "content": "4.2. Implementation Detail" + }, + { + "type": "text", + "bbox": [ + 0.093, + 0.614, + 0.483, + 0.901 + ], + "angle": 0, + "content": "We adopt the same perception modules and ResNet-34 backbone [13] as Transfuser for fair comparison. In the diffusion decoder layer, we employ spatial cross-attention to only interact with BEV features following Transfuser's BEV-based setting. We only perform agent cross-attention, since the perception module of Transfuser does not include vectorized map construction. We stack 2 cascade diffusion decoder layers and apply truncated diffusion policy with 20 clustered anchors. The training diffusion schedule is truncated by 50/1000 to diffuse the anchors, while during inference, we use only 2 denoising steps and select the top-1 scoring predicted trajectory for evaluation. The training and inference recipe directly follows Transfuser: We use three cropped and downscaled forward-facing camera images, concatenated as a \\(1024 \\times 256\\) image, and a rasterized BEV LiDAR as input; DiffusionDrive is trained on navtrain split from scratch for 100 epochs with AdamW optimizer on 8 NVIDIA 4090 GPUs with total batch size of 512, setting the learning rate to \\(6 \\times 10^{-4}\\). No test-time" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.365, + 0.905, + 0.395 + ], + "angle": 0, + "content": "augmentation is applied and the final output for evaluation on navtest split is 8-waypoint trajectory over 4 seconds." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.407, + 0.746, + 0.423 + ], + "angle": 0, + "content": "4.3. Quantitative Comparison" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.429, + 0.907, + 0.732 + ], + "angle": 0, + "content": "Tab. 1 compares DiffusionDrive with state-of-the-art methods on NAVSIM navtest split. With the same ResNet-34 backbone, DiffusionDrive achieves 88.1 PDMS score, demonstrating significant superior performance over the previous learning-based methods. Compared to VADv2, DiffusionDrive surpasses it by 7.2 PDMS while reducing the number of anchors from 8192 to 20, representing a \\(400 \\times\\) reduction. DiffusionDrive also outperforms Hydra-MDP, which follows VADv2's sampling-from-vocabulary paradigm, with a 5.1 PDMS improvement. Even compared to the Hydra-MDP-\\(\\mathcal{V}_{8192}\\)-W-EP, which is a variant of Hydra-MDP [25] by further training to fit the EP evaluation metric with additional supervision and using weighted confidence post-processing, DiffusionDrive still outperforms it by 3.5 EP and 1.6 overall PDMS, relying solely on a straightforward learning-from-human approach without any post-processing. Compared to the Transfuser baseline, where we only differ in the planning module, DiffusionDrive delivers a notable 4.1 PDMS improvement, outperforming it across all sub-scores." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.743, + 0.627, + 0.759 + ], + "angle": 0, + "content": "4.4. Roadmap" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.765, + 0.907, + 0.901 + ], + "angle": 0, + "content": "In Tab. 2, converting Transfuser \\(\\mathrm{DP}\\) using vanilla diffusion policy improves the PDMS score by 0.6 and the mode diversity score \\(\\mathcal{D}\\) by \\(11\\%\\). However, it also significantly increases the overhead of the planning module, requiring \\(20\\times\\) more denoising steps and \\(32\\times\\) the step time, resulting in a total \\(650\\times\\) increase in runtime overhead. With the proposed truncated diffusion policy, \\(\\mathrm{Transfuser}_{\\mathrm{TD}}\\) reduces the number of denoising steps from 20 to 2 while improving PDMS by 1.1 and mode diver" + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "12043" + } + ], + [ + { + "type": "table", + "bbox": [ + 0.093, + 0.089, + 0.907, + 0.233 + ], + "angle": 0, + "content": "
MethodInputImg. BackboneL2 (m) ↓Collision Rate (%) ↓FPS ↑
1s2s3sAvg.1s2s3sAvg.
ST-P3 [15]CameraEffNet-b4 [40]1.332.112.902.110.230.621.270.711.6
UniAD [16]CameraResNet-101 [13]0.450.701.040.730.620.580.630.611.8
OccNet [41]CameraResNet-50 [13]1.292.132.992.140.210.591.370.722.6
VAD [20]CameraResNet-50 [13]0.410.701.050.720.070.170.410.224.5
SparseDrive [39]CameraResNet-50 [13]0.290.580.960.610.010.050.180.089.0
DiffusionDrive (Ours)CameraResNet-50 [13]0.270.540.900.570.030.050.160.088.2
" + }, + { + "type": "table_caption", + "bbox": [ + 0.093, + 0.234, + 0.907, + 0.263 + ], + "angle": 0, + "content": "Table 7. Comparison on nuScenes dataset with open-loop metrics. FPS is measured on a single NVIDIA 4090 GPU following the recipe of SparseDrive [39]. Metric calculation follows ST-P3 [15]." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.267, + 0.482, + 0.375 + ], + "angle": 0, + "content": "sity by \\(59\\%\\). By further incorporating the proposed diffusion decoder, the final model, DiffusionDrive, reaches 88.1 PDMS and \\(74\\%\\) mode diversity score \\(\\mathcal{D}\\). Compared to the TransfuserDP, DiffusionDrive shows improvements in 3.5 PDMS and \\(64\\%\\) mode diversity, and a \\(10\\times\\) reduction in denoising steps, resulting in a \\(6\\times\\) speedup in FPS. This enables real-time, high-quality, multi-mode planning." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.384, + 0.245, + 0.401 + ], + "angle": 0, + "content": "4.5. Ablation Study" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.408, + 0.483, + 0.575 + ], + "angle": 0, + "content": "Effect of designs in diffusion decoder. Tab. 3 shows the effectiveness of our design choices in the diffusion decoder. ID-1 is the Transfuser\\(_{\\text{TD}}\\) in the Tab. 2. By comapring ID-6 and ID-1, we can see that the proposed diffusion decoder reduce the \\(39\\%\\) parameters and significantly improves the planning quality by 2.4 PDMS. ID-2 shows severe performance degeneration due to the lack of rich and hierarchical interaction with the environment. By comparing ID-2 and ID-3, we show that spatial cross-attention is vital for accurate planning. ID-5 shows that the proposed cascade mechanism is effective and can further improve the performance." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.577, + 0.483, + 0.652 + ], + "angle": 0, + "content": "Denoising step number. Tab. 4 shows that due to a reasonable start point, DiffusionDrive achieves good planning quality even with only 1 step, and further denoising steps offer quality improvement and inference flexibility for complex environments." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.654, + 0.483, + 0.716 + ], + "angle": 0, + "content": "Cascade stages. Tab. 5 ablates the impact of cascade stage number. Increasing the stage number can improve the planning quality but saturate at the 4 stages and cost more parameters and inference time at each step." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.717, + 0.483, + 0.824 + ], + "angle": 0, + "content": "Number of sampled noises \\( N_{\\mathrm{infer}} \\). As stated in Sec. 3.3, DiffusionDrive can generate varied trajectories by simply sampling a variable number of noises from anchored Gaussian distribution. Tab. 6 shows that 10 sampled noises can already achieve a decent planning quality. By sampling more noises, DiffusionDrive can cover potential planning action space and lead to improved planning quality." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.833, + 0.313, + 0.85 + ], + "angle": 0, + "content": "4.6. Qualitative Comparison" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.856, + 0.483, + 0.902 + ], + "angle": 0, + "content": "Since the PDMS planning metric calculates based on the top-1 scoring trajectory and our proposed \\(\\mathcal{D}\\) score evaluates mode diversity, these metrics alone cannot fully capture the" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.267, + 0.907, + 0.479 + ], + "angle": 0, + "content": "quality of diverse trajectories. To further validate the quality of multi-mode trajectories, we visualize the planning results of Transfuser, \\(\\mathrm{Transfuser}_{\\mathrm{DP}}\\) and DiffusionDrive on challenging scenarios of NAVSIM navtest split in Fig. 2. The results indicate that the multi-mode trajectories generated by DiffusionDrive are not only diverse but also of high quality. In Fig. 2a, the top-1 scoring trajectory generated by DiffusionDrive closely resembles the ground-truth trajectory, while the highlighted top-10 scoring trajectory surprisingly tries to perform high-quality lane changing. In Fig. 2b, the highlighted top-10 scoring trajectory also performs a lane change, and a neighboring low-scoring trajectory further interacts with surrounding agents to effectively avoid collisions." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.489, + 0.904, + 0.506 + ], + "angle": 0, + "content": "4.7. Quantitative Comparison on nuScenes dataset" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.512, + 0.906, + 0.633 + ], + "angle": 0, + "content": "The nuScenes dataset is previously popular benchmark for end-to-end planning. Since the major scenarios of nuScenes are simple and trivial situations, we only perform comparison in Tab. 7. We implement DiffusionDrive on top of SparseDrive [39] following its training and inference recipe using open-loop metrics proposed in ST-P3 [15]. We stack 2 cascade diffusion decoder layers and apply the truncated diffusion policy with 18 clustered anchors." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.634, + 0.906, + 0.709 + ], + "angle": 0, + "content": "As shown in Tab. 7, DiffusionDrive reduces the average L2 error of SparseDrive by \\(0.04\\mathrm{m}\\), achieving the lowest L2 error and average collision rate. While DiffusionDrive is also efficient and runs \\(1.8\\times\\) faster than VAD with \\(20.8\\%\\) lower L2 error and \\(63.6\\%\\) lower collision rate." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.724, + 0.634, + 0.74 + ], + "angle": 0, + "content": "5. Conclusion" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.75, + 0.907, + 0.903 + ], + "angle": 0, + "content": "In this work, we propose a novel generative driving decision-making model, DiffusionDrive, for end-to-end autonomous driving by incorporating the proposed truncated diffusion policy and efficient cascade diffusion decoder. DiffusionDrive can denoise a variable number of samples from an anchored Gaussian distribution to generate diverse planning trajectories at real-time speeds. Comprehensive experiments and qualitative comparisons validate the superiority of DiffusionDrive in planning quality, running efficiency, and mode diversity." + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.52, + 0.958 + ], + "angle": 0, + "content": "12044" + } + ], + [ + { + "type": "title", + "bbox": [ + 0.092, + 0.091, + 0.251, + 0.108 + ], + "angle": 0, + "content": "Acknowledgement" + }, + { + "type": "text", + "bbox": [ + 0.091, + 0.115, + 0.484, + 0.173 + ], + "angle": 0, + "content": "This work was partially supported by the National Natural Science Foundation of China (NSFC) under Grant No. 62276108. We would like to acknowledge Tianheng Cheng for helpful feedback on the draft." + }, + { + "type": "title", + "bbox": [ + 0.093, + 0.203, + 0.188, + 0.218 + ], + "angle": 0, + "content": "References" + }, + { + "type": "ref_text", + "bbox": [ + 0.101, + 0.227, + 0.484, + 0.282 + ], + "angle": 0, + "content": "[1] Anurag Ajay, Yilun Du, Abhi Gupta, Joshua B. Tenenbaum, Tommi S. Jaakkola, and Pulkit Agrawal. Is conditional generative modeling all you need for decision making? In ICLR, 2023. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.1, + 0.285, + 0.484, + 0.342 + ], + "angle": 0, + "content": "[2] Holger Caesar, Varun Bankiti, Alex H Lang, Sourabh Vora, Venice Erin Liong, Qiang Xu, Anush Krishnan, Yu Pan, Giancarlo Baldan, and Oscar Beijbom. nuscenes: A multimodal dataset for autonomous driving. In CVPR, 2020. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.101, + 0.343, + 0.483, + 0.397 + ], + "angle": 0, + "content": "[3] Shaoyu Chen, Tianheng Cheng, Xinggang Wang, Wenming Meng, Qian Zhang, and Wenyu Liu. Efficient and robust 2d-to-bev representation learning via geometry-guided kernel transformer. CoRR, abs/2206.04584, 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.101, + 0.401, + 0.484, + 0.469 + ], + "angle": 0, + "content": "[4] Shaoyu Chen, Bo Jiang, Hao Gao, Bencheng Liao, Qing Xu, Qian Zhang, Chang Huang, Wenyu Liu, and Xinggang Wang. Vadv2: End-to-end vectorized autonomous driving via probabilistic planning. arXiv preprint arXiv:2402.13243, 2024. 1, 2, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.472, + 0.484, + 0.528 + ], + "angle": 0, + "content": "[5] Shaoyu Chen, Xinggang Wang, Tianheng Cheng, Qian Zhang, Chang Huang, and Wenyu Liu. Polardetr: Polar parametrization for vision-based surround-view 3d detection. Image and Vision Computing, 156:105438, 2025. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.53, + 0.484, + 0.584 + ], + "angle": 0, + "content": "[6] Zhili Chen, Maosheng Ye, Shuangjie Xu, Tongyi Cao, and Qifeng Chen. Ppad: Iterative interactions of prediction and planning for end-to-end autonomous driving. In ECCV, 2024. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.588, + 0.484, + 0.642 + ], + "angle": 0, + "content": "[7] Cheng Chi, Zhenjia Xu, Siyuan Feng, Eric Cousineau, Yilun Du, Benjamin Burchfiel, Russ Tedrake, and Shuran Song. Diffusion policy: Visuomotor policy learning via action diffusion. In RSS, 2023. 1, 3, 5, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.645, + 0.484, + 0.7 + ], + "angle": 0, + "content": "[8] Kashyap Chitta, Aditya Prakash, Bernhard Jaeger, Zehao Yu, Katrin Renz, and Andreas Geiger. Transfuser: Imitation with transformer-based sensor fusion for autonomous driving. TPAMI, 2022. 1, 2, 5, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.703, + 0.484, + 0.757 + ], + "angle": 0, + "content": "[9] Younwoo Choi, Ray Coden Mercurius, Soheil Mohamad Al-izadeh Shabestary, and Amir Rasouli. Dice: Diverse diffusion model with scoring for trajectory prediction. In IV, 2024. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.76, + 0.483, + 0.815 + ], + "angle": 0, + "content": "[10] OpenScene Contributors. Openscene: The largest up-to-date 3d occupancy prediction benchmark in autonomous driving. https://github.com/OpenDriveLab/OpenScene, 2023.6" + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.818, + 0.484, + 0.9 + ], + "angle": 0, + "content": "[11] Daniel Dauner, Marcel Hallgarten, Tianyu Li, Xinshuo Weng, Zhiyu Huang, Zetong Yang, Hongyang Li, Igor Gilitschenski, Boris Ivanovic, Marco Pavone, Andreas Geiger, and Kashyap Chitta. Navsim: Data-driven non-reactive autonomous vehicle simulation and benchmarking. In NeurIPS, 2024. 2, 6, 7" + }, + { + "type": "list", + "bbox": [ + 0.094, + 0.227, + 0.484, + 0.9 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.093, + 0.905, + 0.135 + ], + "angle": 0, + "content": "[12] Xunjiang Gu, Guanyu Song, Igor Gilitschenski, Marco Pavone, and Boris Ivanovic. Producing and leveraging online map uncertainty in trajectory prediction. In CVPR, 2024. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.136, + 0.905, + 0.175 + ], + "angle": 0, + "content": "[13] Kaiming He, Xiangyu Zhang, Shaoqing Ren, and Jian Sun. Deep residual learning for image recognition. In CVPR, 2016. 6, 7, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.178, + 0.905, + 0.218 + ], + "angle": 0, + "content": "[14] Jonathan Ho, Ajay Jain, and Pieter Abbeel. Denoising diffusion probabilistic models. Advances in neural information processing systems, 33:6840-6851, 2020. 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.219, + 0.905, + 0.273 + ], + "angle": 0, + "content": "[15] Shengchao Hu, Li Chen, Penghao Wu, Hongyang Li, Junchi Yan, and Dacheng Tao. St-p3: End-to-end vision-based autonomous driving via spatial-temporal feature learning. In ECCV, 2022. 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.275, + 0.905, + 0.328 + ], + "angle": 0, + "content": "[16] Yihan Hu, Jiazhi Yang, Li Chen, Keyu Li, Chonghao Sima, Xizhou Zhu, Siqi Chai, Senyao Du, Tianwei Lin, Wenhai Wang, et al. Planning-oriented autonomous driving. In CVPR, 2023. 1, 2, 6, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.331, + 0.905, + 0.384 + ], + "angle": 0, + "content": "[17] Junjie Huang, Guan Huang, Zheng Zhu, Yun Ye, and Dalong Du. Bevdet: High-performance multi-camera 3d object detection in bird-eye-view. arXiv preprint arXiv:2112.11790, 2021. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.386, + 0.905, + 0.441 + ], + "angle": 0, + "content": "[18] Zhiyu Huang, Zixu Zhang, Ameya Vaidya, Yuxiao Chen, Chen Lv, and Jaime Fernandez Fisac. Versatile scene-consistent traffic scenario generation as optimization with diffusion. arXiv preprint arXiv:2404.02524, 2024. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.442, + 0.905, + 0.482 + ], + "angle": 0, + "content": "[19] Michael Janner, Yilun Du, Joshua B. Tenenbaum, and Sergey Levine. Planning with diffusion for flexible behavior synthesis. In ICLR, 2022. 1, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.484, + 0.905, + 0.538 + ], + "angle": 0, + "content": "[20] Bo Jiang, Shaoyu Chen, Qing Xu, Bencheng Liao, Jiajie Chen, Helong Zhou, Qian Zhang, Wenyu Liu, Chang Huang, and Xinggang Wang. Vad: Vectorized scene representation for efficient autonomous driving. In ICCV, 2023. 1, 2, 6, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.54, + 0.905, + 0.594 + ], + "angle": 0, + "content": "[21] Chiyu Jiang, Andre Cornman, Cheolho Park, Benjamin Sapp, Yin Zhou, Dragomir Anguelov, et al. Motiondiffuser: Controllable multi-agent motion prediction using diffusion. In CVPR, 2023. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.596, + 0.905, + 0.677 + ], + "angle": 0, + "content": "[22] Napat Karnchanachari, Dimitris Geromichalos, Kok Seang Tan, Nanxiang Li, Christopher Eriksen, Shakiba Yaghoubi, Noushin Mehdipour, Gianmarco Bernasconi, Whye Kit Fong, Yiluan Guo, et al. Towards learning-based planning: The nuplan benchmark for real-world autonomous driving. In ICRA, 2024. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.679, + 0.905, + 0.734 + ], + "angle": 0, + "content": "[23] Yingyan Li, Lue Fan, Jiawei He, Yuqi Wang, Yuntao Chen, Zhaoxiang Zhang, and Tieniu Tan. Enhancing end-to-end autonomous driving with latent world model. arXiv preprint arXiv:2406.08481, 2024. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.735, + 0.905, + 0.789 + ], + "angle": 0, + "content": "[24] Zhiqi Li, Wenhai Wang, Hongyang Li, Enze Xie, Chonghao Sima, Tong Lu, Yu Qiao, and Jifeng Dai. Bevformer: Learning bird's-eye-view representation from multi-camera images via spatiotemporal transformers. In ECCV, 2022. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.79, + 0.905, + 0.858 + ], + "angle": 0, + "content": "[25] Zhenxin Li, Kailin Li, Shihao Wang, Shiyi Lan, Zhiding Yu, Yishen Ji, Zhiqi Li, Ziye Zhu, Jan Kautz, Zuxuan Wu, et al. Hydra-mdp: End-to-end multimodal planning with multi-target hydra-distillation. arXiv preprint arXiv:2406.06978, 2024. 1, 2, 6, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.86, + 0.905, + 0.901 + ], + "angle": 0, + "content": "[26] Zhiqi Li, Zhiding Yu, Shiyi Lan, Jiahan Li, Jan Kautz, Tong Lu, and Jose M Alvarez. Is ego status all you need for open-loop end-to-end autonomous driving? In CVPR, 2024. 2" + }, + { + "type": "list", + "bbox": [ + 0.517, + 0.093, + 0.905, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "12045" + } + ], + [ + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.093, + 0.482, + 0.147 + ], + "angle": 0, + "content": "[27] Benchcheng Liao, Shaoyu Chen, Xinggang Wang, Tianheng Cheng, Qian Zhang, Wenyu Liu, and Chang Huang. MapTR: Structured modeling and learning for online vectorized HD map construction. In ICLR, 2023. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.149, + 0.482, + 0.205 + ], + "angle": 0, + "content": "[28] Benchcheng Liao, Shaoyu Chen, Yunchi Zhang, Bo Jiang, Qian Zhang, Wenyu Liu, Chang Huang, and Xinggang Wang. Maptrv2: An end-to-end framework for online vectorized hd map construction. IJCV, 2024. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.207, + 0.482, + 0.261 + ], + "angle": 0, + "content": "[29] Xuewu Lin, Tianwei Lin, Zixiang Pei, Lichao Huang, and Zhizhong Su. Sparse4d: Multi-view 3d object detection with sparse spatial-temporal fusion. arXiv preprint arXiv:2211.10581, 2022. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.263, + 0.482, + 0.318 + ], + "angle": 0, + "content": "[30] Yaron Lipman, Ricky T. Q. Chen, Heli Ben-Hamu, Maximilian Nickel, and Matthew Le. Flow matching for generative modeling. In The Eleventh International Conference on Learning Representations, 2023. 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.32, + 0.482, + 0.36 + ], + "angle": 0, + "content": "[31] Xingchao Liu, Chengyue Gong, and Qiang Liu. Flow straight and fast: Learning to generate and transfer data with rectified flow. arXiv preprint arXiv:2209.03003, 2022. 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.362, + 0.482, + 0.403 + ], + "angle": 0, + "content": "[32] Yicheng Liu, Tianyuan Yuan, Yue Wang, Yilun Wang, and Hang Zhao. Vectormapnet: End-to-end vectorized hd map learning. In ICML, 2023. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.405, + 0.482, + 0.473 + ], + "angle": 0, + "content": "[33] Robin Rombach, Andreas Blattmann, Dominik Lorenz, Patrick Esser, and Björn Ommer. High-resolution image synthesis with latent diffusion models. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 10684-10695, 2022. 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.476, + 0.482, + 0.516 + ], + "angle": 0, + "content": "[34] Olaf Ronneberger, Philipp Fischer, and Thomas Brox. U-net: Convolutional networks for biomedical image segmentation. In MICCAI, 2015. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.519, + 0.482, + 0.546 + ], + "angle": 0, + "content": "[35] Jiaming Song, Chenlin Meng, and Stefano Ermon. Denoising diffusion implicit models. In ICLR, 2021. 4, 5, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.548, + 0.482, + 0.615 + ], + "angle": 0, + "content": "[36] Yang Song, Jascha Sohl-Dickstein, Diederik P Kingma, Abhishek Kumar, Stefano Ermon, and Ben Poole. Score-based generative modeling through stochastic differential equations. In International Conference on Learning Representations, 2021. 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.618, + 0.482, + 0.658 + ], + "angle": 0, + "content": "[37] Ajay Sridhar, Dhruv Shah, Catherine Glossop, and Sergey Levine. Nomad: Goal masked diffusion policies for navigation and exploration. In ICRA, 2024. 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.661, + 0.482, + 0.715 + ], + "angle": 0, + "content": "[38] Maria Stamatopoulou, Jianwei Liu, and Dimitrios Kanoulas. Dippest: Diffusion-based path planner for synthesizing trajectories applied on quadruped robots. arXiv preprint arXiv:2405.19232, 2024. 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.717, + 0.482, + 0.772 + ], + "angle": 0, + "content": "[39] Wenchao Sun, Xuewu Lin, Yining Shi, Chuang Zhang, Hao ran Wu, and Sifa Zheng. Sparsedrive: End-to-end autonomous driving via sparse scene representation. arXiv preprint arXiv:2405.19620, 2024. 2, 6, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.774, + 0.482, + 0.801 + ], + "angle": 0, + "content": "[40] Mingxing Tan and Quoc Le. Efficientnet: Rethinking model scaling for convolutional neural networks. In ICML, 2019. 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.804, + 0.482, + 0.844 + ], + "angle": 0, + "content": "[41] Wenwen Tong, Chonghao Sima, Tai Wang, Li Chen, Silei Wu, Hanming Deng, Yi Gu, Lewei Lu, Ping Luo, Dahua Lin, et al. Scene as occupancy. In ICCV, 2023. 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.846, + 0.482, + 0.9 + ], + "angle": 0, + "content": "[42] Yue Wang, Vitor Campagnolo Guizilini, Tianyuan Zhang, Yilun Wang, Hang Zhao, and Justin Solomon. Detr3d: 3d object detection from multi-view images via 3d-to-2d queries. In CoRL, 2022. 1, 6" + }, + { + "type": "list", + "bbox": [ + 0.093, + 0.093, + 0.482, + 0.9 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.093, + 0.905, + 0.147 + ], + "angle": 0, + "content": "[43] Yuqi Wang, Jiawei He, Lue Fan, Hongxin Li, Yuntao Chen, and Zhaoxiang Zhang. Driving into the future: Multiview visual forecasting and planning with world model for autonomous driving. In CVPR, 2024. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.149, + 0.905, + 0.219 + ], + "angle": 0, + "content": "[44] Yixiao Wang, Chen Tang, Lingfeng Sun, Simone Rossi, Yichen Xie, Chensheng Peng, Thomas Hannagan, Stefano Sabatini, Nicola Poerio, Masayoshi Tomizuka, et al. Optimizing diffusion models for joint trajectory prediction and controllable generation. In ECCV, 2024. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.221, + 0.905, + 0.261 + ], + "angle": 0, + "content": "[45] Xinshuo Weng, Boris Ivanovic, Yan Wang, Yue Wang, and Marco Pavone. Para-drive: Parallelized architecture for real-time autonomous driving. In CVPR, 2024. 2, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.263, + 0.905, + 0.304 + ], + "angle": 0, + "content": "[46] Zehang Weng, Haofei Lu, Danica Kragic, and Jens Lundell. Dexdiffuser: Generating dexterous grasps with diffusion models. arXiv preprint arXiv:2402.02989, 2024. 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.305, + 0.905, + 0.375 + ], + "angle": 0, + "content": "[47] Sixu Yan, Zeyu Zhang, Muzhi Han, Zaijin Wang, Qi Xie, Zhitian Li, Zhehan Li, Hangxin Liu, Xinggang Wang, and Song-Chun Zhu. M2diffuser: Diffusion-based trajectory optimization for mobile manipulation in 3d scenes. arXiv preprint arXiv:2410.11402, 2024. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.377, + 0.905, + 0.445 + ], + "angle": 0, + "content": "[48] Brian Yang, Huangyuan Su, Nikolaos Gkanatsios, Tsung-Wei Ke, Ayush Jain, Jeff Schneider, and Katerina Fragkiadaki. Diffusion-es: Gradient-free planning with diffusion for autonomous driving and zero-shot instruction following. In CVPR, 2024. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.447, + 0.905, + 0.489 + ], + "angle": 0, + "content": "[49] Jingfeng Yao and Xinggang Wang. Reconstruction vs. generation: Taming optimization dilemma in latent diffusion models. arXiv preprint arXiv:2501.01423, 2025. 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.49, + 0.905, + 0.544 + ], + "angle": 0, + "content": "[50] Jingfeng Yao, Cheng Wang, Wenyu Liu, and Xinggang Wang. Fasteredit: Towards faster diffusion transformers training without architecture modification. arXiv preprint arXiv:2410.10356, 2024. 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.546, + 0.905, + 0.602 + ], + "angle": 0, + "content": "[51] Wenhao Yu, Jie Peng, Huanyu Yang, Junrui Zhang, Yifan Duan, Jianmin Ji, and Yanyong Zhang. Ldp: A local diffusion planner for efficient robot navigation and collision avoidance. arXiv preprint arXiv:2407.01950, 2024. 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.604, + 0.905, + 0.672 + ], + "angle": 0, + "content": "[52] Chengran Yuan, Zhanqi Zhang, Jiawei Sun, Shuo Sun, Zefan Huang, Christina Dao Wen Lee, Dongen Li, Yuhang Han, Anthony Wong, Keng Peng Tee, et al. Drama: An efficient end-to-end motion planner for autonomous driving with mamba. arXiv preprint arXiv:2408.03601, 2024. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.674, + 0.905, + 0.714 + ], + "angle": 0, + "content": "[53] Yanjie Ze, Gu Zhang, Kangning Zhang, Chenyuan Hu, Muhan Wang, and Huazhe Xu. 3d diffusion policy. In RSS, 2024. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.717, + 0.905, + 0.758 + ], + "angle": 0, + "content": "[54] Fangao Zeng, Bin Dong, Yuang Zhang, Tiancai Wang, Xi-angyu Zhang, and Yichen Wei. Motr: End-to-end multiple-object tracking with transformer. In ECCV, 2022. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.76, + 0.905, + 0.801 + ], + "angle": 0, + "content": "[55] Yifu Zhang, Chunyu Wang, Xinggang Wang, Wenjun Zeng, and Wenyu Liu. Fairmot: On the fairness of detection and re-identification in multiple object tracking. IJCV, 2021." + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.803, + 0.905, + 0.857 + ], + "angle": 0, + "content": "[56] Yifu Zhang, Peize Sun, Yi Jiang, Dongdong Yu, Fucheng Weng, Zehuan Yuan, Ping Luo, Wenyu Liu, and Xinggang Wang. Bytetrack: Multi-object tracking by associating every detection box. In ECCV, 2022. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.859, + 0.905, + 0.9 + ], + "angle": 0, + "content": "[57] Huangjie Zheng, Pengcheng He, Weizhu Chen, and Mingyuan Zhou. Truncated diffusion probabilistic models and diffusion-based adversarial auto-encoders. In The" + }, + { + "type": "list", + "bbox": [ + 0.517, + 0.093, + 0.905, + 0.9 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.946, + 0.52, + 0.957 + ], + "angle": 0, + "content": "12046" + } + ], + [ + { + "type": "ref_text", + "bbox": [ + 0.125, + 0.092, + 0.482, + 0.12 + ], + "angle": 0, + "content": "Eleventh International Conference on Learning Representations, 2023. 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.122, + 0.483, + 0.163 + ], + "angle": 0, + "content": "[58] Wenzhao Zheng, Ruiqi Song, Xianda Guo, Chenming Zhang, and Long Chen. Genad: Generative end-to-end autonomous driving. In ECCV, 2024. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.165, + 0.482, + 0.219 + ], + "angle": 0, + "content": "[59] Ziyuan Zhong, Davis Rempe, Yuxiao Chen, Boris Ivanovic, Yulong Cao, Danfei Xu, Marco Pavone, and Baishakhi Ray. Language-guided traffic simulation via scene-level diffusion. In CoRL, 2023. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.221, + 0.482, + 0.276 + ], + "angle": 0, + "content": "[60] Ziyuan Zhong, Davis Rempe, Danfei Xu, Yuxiao Chen, Sushant Veer, Tong Che, Baishakhi Ray, and Marco Pavone. Guided conditional diffusion for controllable traffic simulation. In ICRA, 2023. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.278, + 0.482, + 0.332 + ], + "angle": 0, + "content": "[61] Lianghui Zhu, Zilong Huang, Bencheng Liao, Jun Hao Liew, Hanshu Yan, Jiashi Feng, and Xinggang Wang. Dig: Scalable and efficient diffusion models with gated linear attention. arXiv preprint arXiv:2405.18428, 2024. 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.334, + 0.482, + 0.375 + ], + "angle": 0, + "content": "[62] Xizhou Zhu, Weijie Su, Lewei Lu, Bin Li, Xiaogang Wang, and Jifeng Dai. Deformable detr: Deformable transformers for end-to-end object detection. In ICLR, 2021. 2, 6" + }, + { + "type": "list", + "bbox": [ + 0.093, + 0.092, + 0.483, + 0.375 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.946, + 0.519, + 0.956 + ], + "angle": 0, + "content": "12047" + } + ] +] \ No newline at end of file diff --git a/2025/DiffusionDrive_ Truncated Diffusion Model for End-to-End Autonomous Driving/186eb20a-812b-4b73-a9f1-a45f01468323_origin.pdf b/2025/DiffusionDrive_ Truncated Diffusion Model for End-to-End Autonomous Driving/186eb20a-812b-4b73-a9f1-a45f01468323_origin.pdf new file mode 100644 index 0000000000000000000000000000000000000000..f140fa1b798f1e09b58e8e9eb5c92e2361086b7a --- /dev/null +++ b/2025/DiffusionDrive_ Truncated Diffusion Model for End-to-End Autonomous Driving/186eb20a-812b-4b73-a9f1-a45f01468323_origin.pdf @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:4d23f2f7dfc55c07f768e68df9caabe0154d08d228a22068d95c4fb7cc7e475d +size 1275061 diff --git a/2025/DiffusionDrive_ Truncated Diffusion Model for End-to-End Autonomous Driving/full.md b/2025/DiffusionDrive_ Truncated Diffusion Model for End-to-End Autonomous Driving/full.md new file mode 100644 index 0000000000000000000000000000000000000000..ad3b007b5c1614b6531f20b038ae3b64a30235f0 --- /dev/null +++ b/2025/DiffusionDrive_ Truncated Diffusion Model for End-to-End Autonomous Driving/full.md @@ -0,0 +1,310 @@ +# DiffusionDrive: Truncated Diffusion Model for End-to-End Autonomous Driving + +Bencheng Liao $^{1,2,\diamond}$ Shaoyu Chen $^{2,3}$ Haoran Yin $^{3}$ Bo Jiang $^{2,\diamond}$ Cheng Wang $^{1,2,\diamond}$ Sixu Yan $^{2}$ Xinbang Zhang $^{3}$ Xiangyu Li $^{3}$ Ying Zhang $^{3}$ Qian Zhang $^{3}$ Xinggang Wang $^{2}$ $^{1}$ Institute of Artificial Intelligence, Huazhong University of Science & Technology + $^{2}$ School of EIC, Huazhong University of Science & Technology + $^{3}$ Horizon Robotics +Code & Model & Demo: hustvl/DiffusionDrive + +# Abstract + +Recently, the diffusion model has emerged as a powerful generative technique for robotic policy learning, capable of modeling multi-mode action distributions. Leveraging its capability for end-to-end autonomous driving is a promising direction. However, the numerous denoising steps in the robotic diffusion policy and the more dynamic, open-world nature of traffic scenes pose substantial challenges for generating diverse driving actions at a real-time speed. To address these challenges, we propose a novel truncated diffusion policy that incorporates prior multi-mode anchors and truncates the diffusion schedule, enabling the model to learn denoising from anchored Gaussian distribution to the multi-mode driving action distribution. Additionally, we design an efficient cascade diffusion decoder for enhanced interaction with conditional scene context. The proposed model, DiffusionDrive, demonstrates $10 \times$ reduction in denoising steps compared to vanilla diffusion policy, delivering superior diversity and quality in just 2 steps. On the planning-oriented NAVSIM dataset, with aligned ResNet-34 backbone, DiffusionDrive achieves 88.1 PDMS without bells and whistles, setting a new record, while running at a real-time speed of 45 FPS on an NVIDIA 4090. Qualitative results on challenging scenarios further confirm that DiffusionDrive can robustly generate diverse plausible driving actions. + +# 1. Introduction + +End-to-end autonomous driving has gained significant attention in recent years due to advancements in perception models (detection [5, 17, 24, 42], tracking [54-56], online mapping [27, 28, 32], etc.), which directly learns the driving policy from the raw sensor inputs. This data-driven ap + +![](images/99b5d280af68662c38a293d2104fb29b5b421a613a42c76a54486e13cbbd3e04.jpg) +Figure 1. The comparison of different end-to-end paradigms. (a) Single mode regression [8, 16, 20]. (b) Sampling from vocabulary [4, 25]. (c) Vanilla diffusion policy [7, 19]. (d) The proposed truncated diffusion policy. + +proach offers a scalable and robust alternative to traditional rule-based motion planning, which often struggles to generalize to complex real-world driving settings. + +To effectively learn from data, mainstream end-to-end planners (e.g., Transfuser [8], UniAD [16], VAD [20]) typically regress a single-mode trajectory from an ego-query as shown in Fig. 1a. However, this paradigm does not account for the inherent uncertainty and multi-mode nature of driving behaviors. Recently, VADv2 [20] introduces a large fixed vocabulary of anchor trajectories (4096 anchors) to discretize the continuous action space and capture a broader range of driving behaviors, and then samples from these anchors based on predicted scores as shown in Fig. 1b. However, this large fixed-vocabulary paradigm is fundamentally constrained by the number and quality of anchor trajectories, often failing in out-of-vocabulary scenarios. Furthermore, managing a large number of anchors presents significant computational challenges for real-time applications. Rather than discretizing the action space, diffusion model [7] has proven to be a powerful genera + +tive decision-making policy in the robotics domain, which can directly sample multi-mode physically plausible actions from a Gaussian distribution via iterative denoising process. + +This inspires us to replicate the success of the diffusion model in the robotics domain to end-to-end autonomous driving. We apply the vanilla robotic diffusion policy to the well-known single-mode-regression method, Transfuser [8], by proposing a variant, $\text{Transfuser}_{\text{DP}}$ , which replaces the deterministic MLP regression head with a conditional diffusion model [34]. Though $\text{Transfuser}_{\text{DP}}$ improves planning performance, two major issues arise: 1) The numerous 20 denoising steps in the vanilla DDIM diffusion policy introduce heavy computational consumption during inference as shown in Tab. 2, hindering the real-time application for autonomous driving. 2) The trajectories sampled from different Gaussian noises severely overlap with each other, as illustrated in Fig. 2. This underscores the non-trivial challenge of taming the diffusion models for the dynamic and open-world traffic scenes. + +Unlike the vanilla diffusion policy, which samples actions from a random Gaussian noise conditioned on scene context, human drivers adhere to established driving patterns that they dynamically adjust in response to real-time traffic conditions. This insight motivates us to embed these prior driving patterns into the diffusion policy by partitioning the Gaussian distribution into multiple sub-Gaussian distributions centered around prior anchors, referred to as anchored Gaussian distribution. It is implemented by truncating the diffusion schedule to introduce a small portion of Gaussian noise around the prior anchors as shown in Fig. 3. Thanks to the multi-mode distributional expressivity of the diffusion model, the proposed truncated diffusion policy effectively covers the potential action space without requiring a large set of fixed anchors, as VADv2 does. With more reasonable initial noise samples from the anchored Gaussian distribution, we can truncate the denoising process, reducing the required steps from 20 to just 2—a substantial speedup that satisfies the real-time requirements of autonomous driving. + +To enhance the interaction with conditional scene context, we propose an efficient transformer-based diffusion decoder that interacts not only with structured queries from the perception module but also with Bird's Eye View (BEV) and perspective view (PV) features through a sparse deformable attention mechanism [3, 62]. Additionally, we introduce a cascade mechanism to iteratively refine the trajectory reconstruction within the diffusion decoder at each denoising step. + +With these innovations, we present DiffusionDrive, a diffusion model for real-time end-to-end autonomous driving. We benchmark our method on the planning-oriented NAVSIM dataset [11] using non-reactive simulation and closed-loop evaluations. Without bells and whis + +tles, DiffusionDrive achieves 88.1 PDMS on NAVSIM navtest split with the aligned ResNet-34 backbone, significantly outperforming previous state-of-the-art methods. Even compared to the NAVSIM challenge-winning solution Hydra-MDP- $\nu_{8192}$ -W-EP [25], which follows VADv2 with 8192 anchor trajectories and further incorporates post-processing and additional supervision, DiffusionDrive still outperforms it by 1.6 PDMS through directly learning from human demonstrations and inferring without post-processing, while running at real-time speed of 45 FPS on an NVIDIA 4090. We further validate the superiority of DiffusionDrive on popular nuScenes dataset [2] with open-loop evaluations, DiffusionDrive runs $1.8 \times$ faster than VAD and outperforms it [20] by $20.8\%$ lower L2 error and $63.6\%$ lower collision rate with the same ResNet-50 backbone, demonstrating state-of-the-art planning performance. + +Our contributions can be summarized as follows: + +- We firstly introduce the diffusion model to the field of end-to-end autonomous driving and propose a novel truncated diffusion policy to address the issues of mode collapse and heavy computational overhead found in direct adaptation of vanilla diffusion policy to the traffic scene. +- We design an efficient transformer-based diffusion decoder that interacts with the conditional information in a cascaded manner for better trajectory reconstruction. +- Without bells and whistles, DiffusionDrive significantly outperforms previous state-of-the-art methods, achieving a record-breaking 88.1 PDMS on the NAVSIM navtest split with the same backbone, while maintaining real-time performance at 45 FPS on an NVIDIA 4090. +- We qualitatively demonstrate that DiffusionDrive can generate more diverse and plausible trajectories, exhibiting high-quality multi-mode driving actions in various challenging scenarios. + +# 2. Related Work + +End-to-end autonomous driving. UniAD [16], as a pioneering work, demonstrates the potential of end-to-end autonomous driving by integrating multiple perception tasks to enhance planning performance. VAD [20] further explores the use of compact vectorized scene representations to improve efficiency. Subsequently, a series of works [6, 8, 12, 23, 26, 43, 45, 58] have adopted the single-trajectory planning paradigm to enhance planning performance further. More recently, VADv2 [4] shifts the paradigm towards multi-mode planning by scoring and sampling from a large fixed vocabulary of anchor trajectories. Hydra-MDP [25] improves the scoring mechanism of VADv2 by introducing extra supervision from a rule-based scorer. SparseDrive [39] explores an alternative BEV-free solution. Unlike existing multi-mode planning approaches, we propose a novel paradigm that leverages powerful generative diffusion models for end-to-end autonomous driving. + +![](images/eae6cbced2a079f01e0437de2d3082d21ee7f04b1c9e9400b61c9f8776e04382.jpg) +(b) Top-1's turning left and diverse top-10's lane changing. +Figure 2. Qualitative comparison of Transfuser, TransfuserDP and DiffusionDrive on challenging scenes of NAVSIM navtest split. With the same inputs from front cameras and LiDAR, DiffusionDrive achieves the highest planning quality of top-1 scoring trajectory as illustrated in Tab. 2. We render the highlighted diverse trajectories predicted by DiffusionDrive in the front view. (a) and (b) shows that the top-1 scoring trajectory of DiffusionDrive closely matches the ground truth for both going straight and turning left. Additionally, DiffusionDrive's top-10 scoring trajectory demonstrates high-quality lane changing—an ability not observed in multi-mode TransfuserDP and impossible for Transfuser. + +Diffusion model for traffic simulation. Driving diffusion policy has been explored in the traffic simulation by leveraging only abstract perception groundtruth [9, 18, 21, 44]. MotionDiffuser [21] and CTG [60] are pioneering applications of diffusion models for multi-agent motion prediction, using a conditional diffusion model to sample target trajectories from Gaussian noise. $\mathrm{CTG + + }$ [59] further incorporates a large language model (LLM) for language-driven guidance, improving usability and enabling realistic traffic simulations. Diffusion-ES [48] replaces reward-gradient-guided denoising with evolutionary search. Moving beyond diffusion models limited to traffic simulation with perception groundtruth, our approach unlocks the potential of dif + +fusion models for real-time, end-to-end autonomous driving through our proposed truncated diffusion policy and efficient diffusion decoder. + +Diffusion model for robotic policy learning. Diffusion policy [7] demonstrates the great potential in robotic policy learning, effectively capturing multi-mode action distributions and high-dimensional action spaces. Diffuser [19] proposes an unconditional diffusion model for trajectory sampling, incorporating techniques such as classifier-free guidance and image inpainting to achieve guided sampling. Subsequently, numerous works have applied diffusion models to various robotic tasks, including stationary manipulation [1, 53], mobile manipulation [47], autonomous nav + +Diffusion step: $q(\tau^i |\tau^{i - 1}),i\in [1,T]$ + +![](images/8cfefc6af824cac1c473f324a6793a77eea43f897f66e0c85228a8fc8fda716a.jpg) + +Denoising step: $p_{\theta}\big(\tau^i |\tau^{i - 1},z\big),i\in [1,T]$ + +# Vanilla Diffusion Policy vs. Truncated Diffusion Policy (ours) + +Diffusion step: $q(\tau^i |\tau^{i - 1}),i\in [1,T_{\mathrm{trunc}}]$ + +![](images/908ca90015f3cdef9e8a7a268d4d0e4dffad49b35c5355a18d9a23521945317e.jpg) +Adding noise to the anchors with truncated diffusion schedule. + +Denosing from anchored Gaussian distribution with fewer steps. + +![](images/90462a069ee2704c6c9619343f1e37dcc31b96ef9d735cc9113ee30ae82e944a.jpg) +Denoising step: $p_{\theta}(\tau^i |\tau^{i - 1},z),i\in [1,T_{\mathrm{trunc}}]$ +Figure 3. Illustration of truncated diffusion policy by comparing with vanilla diffusion policy. We truncate the diffusion process and only add a small portion of Gaussian noise to diffuse the anchor trajectories. Then, we train the diffusion model to reconstruct the ground-truth trajectory from the anchored Gaussian distribution with conditional scene context. During the inference, we also truncate the denoising process by starting from the better samples in the anchored Gaussian distribution than the pure Gaussian noise. + +igation [37, 51], quadruped locomotion [38], and dexterous manipulation [46]. However, directly applying vanilla diffusion policy to end-to-end autonomous driving poses unique challenges, as it requires real-time efficiency and the generation of plausible multi-mode trajectories in dynamic and open-world traffic scenes. In this work, we propose a novel truncated diffusion policy to address these challenges, introducing concepts that have not yet been explored in the robotics field. + +Diffusion model for image generation. Diffusion models have been extensively adopted for image generation tasks [33, 36, 49, 50, 61]. DDIM [35] enhances DDPM [14] by enabling efficient sampling with significantly fewer steps based on non-Markovian diffusion processes. Flow matching [30, 31] further optimizes the generative process by directly modeling continuous probability flows. TDPM [57] proposes truncated denoising, which initiates the generation process from an implicit intermediate distribution to accelerate sampling. In contrast to these approaches, our method introduces an explicit driving prior within the diffusion policy, effectively guiding the diffusion process toward more accurate and efficient generation tailored specifically for end-to-end autonomous driving. + +# 3. Method + +# 3.1. Preliminary + +Task formulation. End-to-end autonomous driving takes raw sensor data as input and predicts the future trajectory of the ego-vehicle. The trajectory is represented as a sequence of waypoints $\tau = \{(x_{t},y_{t})\}_{t = 1}^{T_{f}}$ , where $T_{f}$ denotes the planning horizon, and $(x_{t},y_{t})$ is the location of each waypoint at time $t$ in the current ego-vehicle coordinate system. + +Conditional diffusion model. The conditional diffusion model poses a forward diffusion process as gradually adding noise to the data sample, which can be defined as: + +$$ +q \left(\tau^ {i} \mid \tau^ {0}\right) = \mathcal {N} \left(\tau^ {i}; \sqrt {\bar {\alpha} ^ {i}} \boldsymbol {\tau} ^ {0}, \left(1 - \bar {\alpha} ^ {i}\right) \mathbf {I}\right), \tag {1} +$$ + +where $\tau^0$ is the clean data sample, and $\tau^i$ is the data sample with noise at time $i$ (Note: we use superscript $i$ to denote diffusion timestep). The constant $\bar{\alpha}^i = \prod_{s=1}^i \alpha^s = \prod_{s=1}^i (1 - \beta^s)$ and $\beta^s$ is the noise schedule. We train the reverse process model $f_\theta(\tau^i, z, i)$ to predict $\tau^0$ from $\tau^i$ with the guidance of conditional information $z$ , where $\theta$ is the trainable model parameter. During inference, the trained diffusion model $f_\theta$ progressively refines from the random + +![](images/39188c906af22969f52df93d3eccb03e17c013632528cf929e933086a49a30e7.jpg) +Figure 4. Overall architecture of DiffusionDrive. (a) DiffusionDrive can integrate various existing perception modules and sensor inputs. (b) The designed diffusion decoder takes the sampled noisy trajectories from anchored Gaussian distribution as input and progressively denoises them with enhanced interactions with the conditional scene context in a cascade manner to generate the final predictions. + +![](images/3501d4a7121777ea39ebdd62ca4e2b3efafa5c0f6dd4428085611e154596d12e.jpg) + +noise $\tau^T$ sampled in Gaussian distribution to the predicted clean data sample $\tau^0$ with the guidance of conditional information $z$ , which is defined as: + +$$ +p _ {\theta} \left(\boldsymbol {\tau} ^ {0} \mid z\right) = \int p \left(\boldsymbol {\tau} ^ {T}\right) \prod_ {i = 1} ^ {T} p _ {\theta} \left(\boldsymbol {\tau} ^ {i - 1} \mid \boldsymbol {\tau} ^ {i}, z\right) d \boldsymbol {\tau} ^ {1: T}. \tag {2} +$$ + +# 3.2. Investigation + +Turn Transfuser [8] into conditional diffusion model. We begin from the representative deterministic end-to-end planner Transfuser [8] and turn it into a generative model $\mathrm{Transfuser}_{\mathrm{DP}}$ by simply replacing the regression MLP layers with the conditional diffusion model UNet following vanilla diffusion policy [7]. During the evaluation, we sample a random noise and progressively refine it with 20 steps. Tab. 2 shows that $\mathrm{Transfuser}_{\mathrm{DP}}$ achieves better planning quality than deterministic Transfuser. + +Mode collapse. To further investigate the multi-mode property of the vanilla diffusion policy in driving, we sampled 20 random noises from Gaussian distribution and denoised them using 20 steps. As shown in Fig. 2, the different random noises converge to similar trajectories after the denoising process. To quantitatively analyze the phenomenon of mode collapse, we define a mode diversity score $\mathcal{D}$ based on the mean Intersection over Union (mIoU) between each denoised trajectory and the union of all denoised trajectories: + +$$ +\mathcal {D} = 1 - \frac {1}{N} \sum_ {i = 1} ^ {N} \frac {\operatorname {A r e a} \left(\tau_ {i} \cap \bigcup_ {j = 1} ^ {N} \tau_ {j}\right)}{\operatorname {A r e a} \left(\tau_ {i} \cup \bigcup_ {j = 1} ^ {N} \tau_ {j}\right)}, \tag {3} +$$ + +where $\tau_{i}$ represents the $i$ -th denoised trajectory, $N$ is the total number of sampled trajectories and $\bigcup_{j=1}^{N} \tau_{j}$ is the union of all denoised trajectories. A higher mIoU indicates less diversity of the denoised trajectories. The quantitative mode diversity results in Tab. 2 further validate the observations presented in Fig. 2. + +Heavy denoising overhead. The DDIM [35] diffusion policy requires 20 denoising steps to transform random noise + +into a feasible trajectory, which introduces significant computational overhead, reducing the FPS from 60 to 7, as shown in Tab. 2, and making it impractical for real-time online driving applications. + +# 3.3. Truncated Diffusion + +Human driving follows fixed patterns, unlike the random noise denoising in vanilla diffusion policy. Motivated by this, we propose a truncated diffusion policy that begins the denoising process from an anchored Gaussian distribution instead of a standard Gaussian distribution. To enable the model to learn to denoise from the anchored Gaussian distribution to the desired driving policy, we further truncate the diffusion schedule during training, adding only a small amount of Gaussian noise to the anchors. + +Training. We first construct the diffusion process by adding Gaussian noise to anchors $\{\mathbf{a}_k\}_{k=1}^{N_{\text{anchor}}}$ clustered by K-Means on the training set, where $\mathbf{a}_k = \{(x_t, y_t)\}_{t=1}^{T_f}$ . We truncate the diffusion noise schedule to diffuse the anchors to the anchored Gaussian distribution: + +$$ +\tau_ {k} ^ {i} = \sqrt {\bar {\alpha} ^ {i}} \mathbf {a} _ {k} + \sqrt {1 - \bar {\alpha} ^ {i}} \boldsymbol {\epsilon}, \quad \boldsymbol {\epsilon} \sim \mathcal {N} (0, \mathbf {I}), \qquad (4) +$$ + +where $i \in [1, T_{\mathrm{trunc}}]$ and $T_{\mathrm{trunc}} \ll T$ is the truncated diffusion steps. + +During training, the diffusion decoder $f_{\theta}$ takes as input $N_{\mathrm{anchor}}$ noisy trajectories $\{\tau_k^i\}_{k=1}^{N_{\mathrm{anchor}}}$ and predicts classification scores $\{\hat{s}_k\}_{k=1}^{N_{\mathrm{anchor}}}$ and denoised trajectories $\{\hat{\tau}_k\}_{k=1}^{N_{\mathrm{anchor}}}$ : + +$$ +\left\{\hat {s} _ {k}, \hat {\tau} _ {k} \right\} _ {k = 1} ^ {N _ {\text {a n c h o r}}} = f _ {\theta} \left(\left\{\tau_ {k} ^ {i} \right\} _ {k = 1} ^ {N _ {\text {a n c h o r}}}, z\right), \tag {5} +$$ + +where $z$ represents the conditional information. We assign the noisy trajectory around the closest anchor to the ground truth trajectory $\tau_{\mathrm{gt}}$ as positive sample $(y_k = 1)$ and others as negative samples $(y_k = 0)$ . The training objective combines trajectory reconstruction and classification: + +$$ +\mathcal {L} = \sum_ {k = 1} ^ {N _ {\text {a n c h o r}}} \left[ y _ {k} \mathcal {L} _ {\text {r e c}} \left(\hat {\tau} _ {k}, \tau_ {\mathrm {g t}}\right) + \lambda \mathrm {B C E} \left(\hat {s} _ {k}, y _ {k}\right) \right], \tag {6} +$$ + +
MethodInputImg. BackboneAnchorNC↑DAC ↑TTC ↑Comf. ↑EP ↑PDMS ↑
UniAD [16]CameraResNet-34 [13]097.891.992.910078.883.4
PARA-Drive [45]CameraResNet-34 [13]097.992.493.099.879.384.0
LTF [8]CameraResNet-34 [13]097.492.892.410079.083.8
Transfuser [8]C & LResNet-34 [13]097.792.892.810079.284.0
DRAMA [52]C & LResNet-34 [13]098.093.194.810080.185.5
VADv2- \( \mathcal{V}_{8192} \) [4]C & LResNet-34 [13]819297.289.191.610076.080.9
Hydra-MDP- \( \mathcal{V}_{8192} \) [25]C & LResNet-34 [13]819297.991.792.910077.683.0
Hydra-MDP- \( \mathcal{V}_{8192} \)-W-EP [25]C & LResNet-34 [13]819298.396.094.610078.786.5
DiffusionDrive (Ours)C & LResNet-34 [13]2098.296.294.710082.288.1
+ +Table 1. Comparison on planning-oriented NAVSIM navtest split with closed-loop metrics. “C & L” denotes the use of both camera and LiDAR as sensor inputs. “ $\mathcal{V}_{8192}$ ” denotes 8192 anchors. “Hydra-MDP- $\mathcal{V}_{8192}$ -W-EP” is a variant of Hydra-MDP [25], which is further trained to fit the EP evaluation metric with additional supervision from the rule-based evaluator and uses weighted confidence post-processing. DiffusionDrive simply learns from human demonstrations and infers without post-processing. The best and the second best results are denoted by bold and underline. + +
MethodNC↑DAC↑TTC↑Comf.↑EP↑PDMS↑Arch.Plan Module TimeTotal ↓D↑Para.↓FPS↑
Step Time↓Steps ↓
Transfuser97.792.892.810079.284.0MLP0.2ms10.2ms0%56M60
TransfuserDP97.593.792.710079.484.6+0.6UNet6.5ms20130.0ms11%101M7
TransfuserTD97.994.293.910080.285.7+1.7UNet6.9ms213.8ms70%102M27
DiffusionDrive98.296.294.710082.288.1+4.1Dec.3.8ms27.6ms74%60M45
+ +Table 2. Roadmap from Transfuser to DiffusionDrive on NAVSIM navtest split. "TransfuserDP" denotes Transfuser with vanilla DDIM diffusion policy [7]. "TransfuserTD" denotes Transfuser with truncated diffusion policy. "Step Time" denotes the runtime of each denoising step. "FPS" and runtime are measured on an NVIDIA 4090 GPU. "D" denotes the mode diversity score defined in Eq. (3). + +where $\lambda$ balances the simple L1 reconstruction loss $\mathcal{L}_{\mathrm{rec}}$ and binary cross-entropy (BCE) classification loss. + +Inference. We use a truncated denoising process that starts with noisy trajectories sampled from the anchored Gaussian distribution and progressively denoises them to final predictions. At each denoising timestep, the estimated trajectories from the previous step are passed to the diffusion decoder $f_{\theta}$ , which predicts classification scores $\{\hat{s}_k\}_{k=1}^{N_{\mathrm{infer}}}$ and coordinates $\{\hat{\tau}_k\}_{k=1}^{N_{\mathrm{infer}}}$ . After obtaining the current timestep's predictions, we apply the DDIM [35] update rule to sample trajectories for the next timestep. + +Inference flexibility. A key advantage of our approach lies in its inference flexibility. While the model is trained with $N_{\text{anchor}}$ trajectories, the inference process can accommodate an arbitrary number of trajectory samples $N_{\text{infer}}$ , where $N_{\text{infer}}$ can be dynamically adjusted based on computational resources or application requirements. + +# 3.4. Architecture + +The overall architecture of our proposed method, DiffusionDrive, is illustrated in Fig. 4. DiffusionDrive can integrate various existing perception modules used in previous end-to-end planners [8, 16, 20, 39] and take different sensor inputs. The designed diffusion decoder is tailored for the complex and challenging driving application, which has enhanced interactions with the conditional scene context. + +Diffusion decoder. Given the set of sampled noisy trajectories + +ries $\{\hat{\tau}_k\}_{k=1}^{N_{\mathrm{infer}}}$ from the anchored Gaussian distribution, we begin by applying deformable spatial cross-attention [29, 42, 62] to interact with Bird's Eye View (BEV) or Perspective View (PV) features based on the trajectory coordinates. Subsequently, cross-attention is performed between the trajectory features and the agent/map queries derived from the perception module, followed by a feed-forward network (FFN). To encode the diffusion timestep information, we utilize a Timestep Modulation layer, which is followed by a Multi-Layer Perceptron (MLP) that predicts the confidence score and the offset relative to the initial noisy trajectory coordinates. The output from this diffusion decoder layer serves as the input for the subsequent cascade diffusion decoder layer. DiffusionDrive further reuses the cascade diffusion decoder to iteratively denoise the trajectory during inference, with parameters shared across the different denoising timesteps. The final trajectory with the highest confidence score is selected as the output. + +# 4. Experiment + +# 4.1. Dataset + +NAVSIM. The NAVSIM dataset [11] is a real-world planning-oriented dataset builds upon OpenScene [10], a compact redistribution of nuPlan [22], the largest publicly available annotated driving dataset. NAVSIM leverages eight cameras to achieve a full $360^{\circ}$ FOV, along with a + +
IDUNet DecoderEgo Query InteractionSpatial Cross-attnAgent/Map Cross-attnCascade DecoderParam.↓NC↑DAC↑Planning MetricPDMS↑
TTC↑Comf.↑EP↑
1✓✓XXX102M97.994.293.910080.285.7
2X✓XXX57M88.783.280.084.843.355.1
3X✓✓XX58M98.295.494.410081.387.1
4X✓X✓X58M97.993.593.810079.885.1
5X✓✓✓X59M98.095.894.410081.787.4
6X✓✓✓✓60M98.296.294.710082.288.1
+ +Table 3. Ablation for design choices. "Cascade Decoder" indicates that we stack 2 cascade diffusion decoder layers. ID-1 refers to TransfuserTD in Tab. 2, utilizing conditional UNet and interaction with the ego-query, which Transfuser uses to directly regress the single-mode trajectory. + +
StepsParam.NCDACTTCComf.EPPDMS
160M98.396.094.710082.187.9
260M98.296.294.710082.288.1
360M98.296.394.710092.288.1
+ +Table 4. Denoising step number. + +
StagesParam.NCDACTTCComf.EPPDMS
159M98.095.894.410081.787.4
260M98.296.294.710082.288.1
465M98.496.294.910082.488.2
+ +Table 5. Cascade stages. + +
N_inferParam.NCDACTTCComf.EPPDMS
1060M97.993.593.110080.084.9
2060M98.296.294.710082.288.1
4060M98.596.294.810082.588.2
+ +Table 6. Number of sampled noises ${N}_{\text{infer }}$ . + +merged LiDAR point cloud derived from five sensors. Annotations are provided at a frequency of $2\mathrm{Hz}$ and include both HD maps and object bounding boxes. The dataset is designed to emphasize challenging driving scenarios involving dynamic changes in driving intentions, while deliberately excluding trivial situations such as stationary scenes or constant-speed driving. + +NAVSIM benchmarks planning performance using non-reactive simulations and closed-loop metrics for comprehensive evaluation. In this paper, we employ the proposed PDM score (PDMS) [11], which is a weighted combination of several sub-scores: no at-fault collisions (NC), drivable area compliance (DAC), time-to-collision (TTC), comfort (Comf.), and ego progress (EP). + +# 4.2. Implementation Detail + +We adopt the same perception modules and ResNet-34 backbone [13] as Transfuser for fair comparison. In the diffusion decoder layer, we employ spatial cross-attention to only interact with BEV features following Transfuser's BEV-based setting. We only perform agent cross-attention, since the perception module of Transfuser does not include vectorized map construction. We stack 2 cascade diffusion decoder layers and apply truncated diffusion policy with 20 clustered anchors. The training diffusion schedule is truncated by 50/1000 to diffuse the anchors, while during inference, we use only 2 denoising steps and select the top-1 scoring predicted trajectory for evaluation. The training and inference recipe directly follows Transfuser: We use three cropped and downscaled forward-facing camera images, concatenated as a $1024 \times 256$ image, and a rasterized BEV LiDAR as input; DiffusionDrive is trained on navtrain split from scratch for 100 epochs with AdamW optimizer on 8 NVIDIA 4090 GPUs with total batch size of 512, setting the learning rate to $6 \times 10^{-4}$ . No test-time + +augmentation is applied and the final output for evaluation on navtest split is 8-waypoint trajectory over 4 seconds. + +# 4.3. Quantitative Comparison + +Tab. 1 compares DiffusionDrive with state-of-the-art methods on NAVSIM navtest split. With the same ResNet-34 backbone, DiffusionDrive achieves 88.1 PDMS score, demonstrating significant superior performance over the previous learning-based methods. Compared to VADv2, DiffusionDrive surpasses it by 7.2 PDMS while reducing the number of anchors from 8192 to 20, representing a $400 \times$ reduction. DiffusionDrive also outperforms Hydra-MDP, which follows VADv2's sampling-from-vocabulary paradigm, with a 5.1 PDMS improvement. Even compared to the Hydra-MDP- $\mathcal{V}_{8192}$ -W-EP, which is a variant of Hydra-MDP [25] by further training to fit the EP evaluation metric with additional supervision and using weighted confidence post-processing, DiffusionDrive still outperforms it by 3.5 EP and 1.6 overall PDMS, relying solely on a straightforward learning-from-human approach without any post-processing. Compared to the Transfuser baseline, where we only differ in the planning module, DiffusionDrive delivers a notable 4.1 PDMS improvement, outperforming it across all sub-scores. + +# 4.4. Roadmap + +In Tab. 2, converting Transfuser $\mathrm{DP}$ using vanilla diffusion policy improves the PDMS score by 0.6 and the mode diversity score $\mathcal{D}$ by $11\%$ . However, it also significantly increases the overhead of the planning module, requiring $20\times$ more denoising steps and $32\times$ the step time, resulting in a total $650\times$ increase in runtime overhead. With the proposed truncated diffusion policy, $\mathrm{Transfuser}_{\mathrm{TD}}$ reduces the number of denoising steps from 20 to 2 while improving PDMS by 1.1 and mode diver + +
MethodInputImg. BackboneL2 (m) ↓Collision Rate (%) ↓FPS ↑
1s2s3sAvg.1s2s3sAvg.
ST-P3 [15]CameraEffNet-b4 [40]1.332.112.902.110.230.621.270.711.6
UniAD [16]CameraResNet-101 [13]0.450.701.040.730.620.580.630.611.8
OccNet [41]CameraResNet-50 [13]1.292.132.992.140.210.591.370.722.6
VAD [20]CameraResNet-50 [13]0.410.701.050.720.070.170.410.224.5
SparseDrive [39]CameraResNet-50 [13]0.290.580.960.610.010.050.180.089.0
DiffusionDrive (Ours)CameraResNet-50 [13]0.270.540.900.570.030.050.160.088.2
+ +Table 7. Comparison on nuScenes dataset with open-loop metrics. FPS is measured on a single NVIDIA 4090 GPU following the recipe of SparseDrive [39]. Metric calculation follows ST-P3 [15]. + +sity by $59\%$ . By further incorporating the proposed diffusion decoder, the final model, DiffusionDrive, reaches 88.1 PDMS and $74\%$ mode diversity score $\mathcal{D}$ . Compared to the TransfuserDP, DiffusionDrive shows improvements in 3.5 PDMS and $64\%$ mode diversity, and a $10\times$ reduction in denoising steps, resulting in a $6\times$ speedup in FPS. This enables real-time, high-quality, multi-mode planning. + +# 4.5. Ablation Study + +Effect of designs in diffusion decoder. Tab. 3 shows the effectiveness of our design choices in the diffusion decoder. ID-1 is the Transfuser $_{\text{TD}}$ in the Tab. 2. By comapring ID-6 and ID-1, we can see that the proposed diffusion decoder reduce the $39\%$ parameters and significantly improves the planning quality by 2.4 PDMS. ID-2 shows severe performance degeneration due to the lack of rich and hierarchical interaction with the environment. By comparing ID-2 and ID-3, we show that spatial cross-attention is vital for accurate planning. ID-5 shows that the proposed cascade mechanism is effective and can further improve the performance. + +Denoising step number. Tab. 4 shows that due to a reasonable start point, DiffusionDrive achieves good planning quality even with only 1 step, and further denoising steps offer quality improvement and inference flexibility for complex environments. + +Cascade stages. Tab. 5 ablates the impact of cascade stage number. Increasing the stage number can improve the planning quality but saturate at the 4 stages and cost more parameters and inference time at each step. + +Number of sampled noises $N_{\mathrm{infer}}$ . As stated in Sec. 3.3, DiffusionDrive can generate varied trajectories by simply sampling a variable number of noises from anchored Gaussian distribution. Tab. 6 shows that 10 sampled noises can already achieve a decent planning quality. By sampling more noises, DiffusionDrive can cover potential planning action space and lead to improved planning quality. + +# 4.6. Qualitative Comparison + +Since the PDMS planning metric calculates based on the top-1 scoring trajectory and our proposed $\mathcal{D}$ score evaluates mode diversity, these metrics alone cannot fully capture the + +quality of diverse trajectories. To further validate the quality of multi-mode trajectories, we visualize the planning results of Transfuser, $\mathrm{Transfuser}_{\mathrm{DP}}$ and DiffusionDrive on challenging scenarios of NAVSIM navtest split in Fig. 2. The results indicate that the multi-mode trajectories generated by DiffusionDrive are not only diverse but also of high quality. In Fig. 2a, the top-1 scoring trajectory generated by DiffusionDrive closely resembles the ground-truth trajectory, while the highlighted top-10 scoring trajectory surprisingly tries to perform high-quality lane changing. In Fig. 2b, the highlighted top-10 scoring trajectory also performs a lane change, and a neighboring low-scoring trajectory further interacts with surrounding agents to effectively avoid collisions. + +# 4.7. Quantitative Comparison on nuScenes dataset + +The nuScenes dataset is previously popular benchmark for end-to-end planning. Since the major scenarios of nuScenes are simple and trivial situations, we only perform comparison in Tab. 7. We implement DiffusionDrive on top of SparseDrive [39] following its training and inference recipe using open-loop metrics proposed in ST-P3 [15]. We stack 2 cascade diffusion decoder layers and apply the truncated diffusion policy with 18 clustered anchors. + +As shown in Tab. 7, DiffusionDrive reduces the average L2 error of SparseDrive by $0.04\mathrm{m}$ , achieving the lowest L2 error and average collision rate. While DiffusionDrive is also efficient and runs $1.8\times$ faster than VAD with $20.8\%$ lower L2 error and $63.6\%$ lower collision rate. + +# 5. Conclusion + +In this work, we propose a novel generative driving decision-making model, DiffusionDrive, for end-to-end autonomous driving by incorporating the proposed truncated diffusion policy and efficient cascade diffusion decoder. DiffusionDrive can denoise a variable number of samples from an anchored Gaussian distribution to generate diverse planning trajectories at real-time speeds. Comprehensive experiments and qualitative comparisons validate the superiority of DiffusionDrive in planning quality, running efficiency, and mode diversity. + +# Acknowledgement + +This work was partially supported by the National Natural Science Foundation of China (NSFC) under Grant No. 62276108. We would like to acknowledge Tianheng Cheng for helpful feedback on the draft. + +# References + +[1] Anurag Ajay, Yilun Du, Abhi Gupta, Joshua B. Tenenbaum, Tommi S. Jaakkola, and Pulkit Agrawal. Is conditional generative modeling all you need for decision making? In ICLR, 2023. 3 +[2] Holger Caesar, Varun Bankiti, Alex H Lang, Sourabh Vora, Venice Erin Liong, Qiang Xu, Anush Krishnan, Yu Pan, Giancarlo Baldan, and Oscar Beijbom. nuscenes: A multimodal dataset for autonomous driving. In CVPR, 2020. 2 +[3] Shaoyu Chen, Tianheng Cheng, Xinggang Wang, Wenming Meng, Qian Zhang, and Wenyu Liu. Efficient and robust 2d-to-bev representation learning via geometry-guided kernel transformer. CoRR, abs/2206.04584, 2022. 2 +[4] Shaoyu Chen, Bo Jiang, Hao Gao, Bencheng Liao, Qing Xu, Qian Zhang, Chang Huang, Wenyu Liu, and Xinggang Wang. Vadv2: End-to-end vectorized autonomous driving via probabilistic planning. arXiv preprint arXiv:2402.13243, 2024. 1, 2, 6 +[5] Shaoyu Chen, Xinggang Wang, Tianheng Cheng, Qian Zhang, Chang Huang, and Wenyu Liu. Polardetr: Polar parametrization for vision-based surround-view 3d detection. Image and Vision Computing, 156:105438, 2025. 1 +[6] Zhili Chen, Maosheng Ye, Shuangjie Xu, Tongyi Cao, and Qifeng Chen. Ppad: Iterative interactions of prediction and planning for end-to-end autonomous driving. In ECCV, 2024. 2 +[7] Cheng Chi, Zhenjia Xu, Siyuan Feng, Eric Cousineau, Yilun Du, Benjamin Burchfiel, Russ Tedrake, and Shuran Song. Diffusion policy: Visuomotor policy learning via action diffusion. In RSS, 2023. 1, 3, 5, 6 +[8] Kashyap Chitta, Aditya Prakash, Bernhard Jaeger, Zehao Yu, Katrin Renz, and Andreas Geiger. Transfuser: Imitation with transformer-based sensor fusion for autonomous driving. TPAMI, 2022. 1, 2, 5, 6 +[9] Younwoo Choi, Ray Coden Mercurius, Soheil Mohamad Al-izadeh Shabestary, and Amir Rasouli. Dice: Diverse diffusion model with scoring for trajectory prediction. In IV, 2024. 3 +[10] OpenScene Contributors. Openscene: The largest up-to-date 3d occupancy prediction benchmark in autonomous driving. https://github.com/OpenDriveLab/OpenScene, 2023.6 +[11] Daniel Dauner, Marcel Hallgarten, Tianyu Li, Xinshuo Weng, Zhiyu Huang, Zetong Yang, Hongyang Li, Igor Gilitschenski, Boris Ivanovic, Marco Pavone, Andreas Geiger, and Kashyap Chitta. Navsim: Data-driven non-reactive autonomous vehicle simulation and benchmarking. In NeurIPS, 2024. 2, 6, 7 + +[12] Xunjiang Gu, Guanyu Song, Igor Gilitschenski, Marco Pavone, and Boris Ivanovic. Producing and leveraging online map uncertainty in trajectory prediction. In CVPR, 2024. 2 +[13] Kaiming He, Xiangyu Zhang, Shaoqing Ren, and Jian Sun. Deep residual learning for image recognition. In CVPR, 2016. 6, 7, 8 +[14] Jonathan Ho, Ajay Jain, and Pieter Abbeel. Denoising diffusion probabilistic models. Advances in neural information processing systems, 33:6840-6851, 2020. 4 +[15] Shengchao Hu, Li Chen, Penghao Wu, Hongyang Li, Junchi Yan, and Dacheng Tao. St-p3: End-to-end vision-based autonomous driving via spatial-temporal feature learning. In ECCV, 2022. 8 +[16] Yihan Hu, Jiazhi Yang, Li Chen, Keyu Li, Chonghao Sima, Xizhou Zhu, Siqi Chai, Senyao Du, Tianwei Lin, Wenhai Wang, et al. Planning-oriented autonomous driving. In CVPR, 2023. 1, 2, 6, 8 +[17] Junjie Huang, Guan Huang, Zheng Zhu, Yun Ye, and Dalong Du. Bevdet: High-performance multi-camera 3d object detection in bird-eye-view. arXiv preprint arXiv:2112.11790, 2021. 1 +[18] Zhiyu Huang, Zixu Zhang, Ameya Vaidya, Yuxiao Chen, Chen Lv, and Jaime Fernandez Fisac. Versatile scene-consistent traffic scenario generation as optimization with diffusion. arXiv preprint arXiv:2404.02524, 2024. 3 +[19] Michael Janner, Yilun Du, Joshua B. Tenenbaum, and Sergey Levine. Planning with diffusion for flexible behavior synthesis. In ICLR, 2022. 1, 3 +[20] Bo Jiang, Shaoyu Chen, Qing Xu, Bencheng Liao, Jiajie Chen, Helong Zhou, Qian Zhang, Wenyu Liu, Chang Huang, and Xinggang Wang. Vad: Vectorized scene representation for efficient autonomous driving. In ICCV, 2023. 1, 2, 6, 8 +[21] Chiyu Jiang, Andre Cornman, Cheolho Park, Benjamin Sapp, Yin Zhou, Dragomir Anguelov, et al. Motiondiffuser: Controllable multi-agent motion prediction using diffusion. In CVPR, 2023. 3 +[22] Napat Karnchanachari, Dimitris Geromichalos, Kok Seang Tan, Nanxiang Li, Christopher Eriksen, Shakiba Yaghoubi, Noushin Mehdipour, Gianmarco Bernasconi, Whye Kit Fong, Yiluan Guo, et al. Towards learning-based planning: The nuplan benchmark for real-world autonomous driving. In ICRA, 2024. 6 +[23] Yingyan Li, Lue Fan, Jiawei He, Yuqi Wang, Yuntao Chen, Zhaoxiang Zhang, and Tieniu Tan. Enhancing end-to-end autonomous driving with latent world model. arXiv preprint arXiv:2406.08481, 2024. 2 +[24] Zhiqi Li, Wenhai Wang, Hongyang Li, Enze Xie, Chonghao Sima, Tong Lu, Yu Qiao, and Jifeng Dai. Bevformer: Learning bird's-eye-view representation from multi-camera images via spatiotemporal transformers. In ECCV, 2022. 1 +[25] Zhenxin Li, Kailin Li, Shihao Wang, Shiyi Lan, Zhiding Yu, Yishen Ji, Zhiqi Li, Ziye Zhu, Jan Kautz, Zuxuan Wu, et al. Hydra-mdp: End-to-end multimodal planning with multi-target hydra-distillation. arXiv preprint arXiv:2406.06978, 2024. 1, 2, 6, 7 +[26] Zhiqi Li, Zhiding Yu, Shiyi Lan, Jiahan Li, Jan Kautz, Tong Lu, and Jose M Alvarez. Is ego status all you need for open-loop end-to-end autonomous driving? In CVPR, 2024. 2 + +[27] Benchcheng Liao, Shaoyu Chen, Xinggang Wang, Tianheng Cheng, Qian Zhang, Wenyu Liu, and Chang Huang. MapTR: Structured modeling and learning for online vectorized HD map construction. In ICLR, 2023. 1 +[28] Benchcheng Liao, Shaoyu Chen, Yunchi Zhang, Bo Jiang, Qian Zhang, Wenyu Liu, Chang Huang, and Xinggang Wang. Maptrv2: An end-to-end framework for online vectorized hd map construction. IJCV, 2024. 1 +[29] Xuewu Lin, Tianwei Lin, Zixiang Pei, Lichao Huang, and Zhizhong Su. Sparse4d: Multi-view 3d object detection with sparse spatial-temporal fusion. arXiv preprint arXiv:2211.10581, 2022. 6 +[30] Yaron Lipman, Ricky T. Q. Chen, Heli Ben-Hamu, Maximilian Nickel, and Matthew Le. Flow matching for generative modeling. In The Eleventh International Conference on Learning Representations, 2023. 4 +[31] Xingchao Liu, Chengyue Gong, and Qiang Liu. Flow straight and fast: Learning to generate and transfer data with rectified flow. arXiv preprint arXiv:2209.03003, 2022. 4 +[32] Yicheng Liu, Tianyuan Yuan, Yue Wang, Yilun Wang, and Hang Zhao. Vectormapnet: End-to-end vectorized hd map learning. In ICML, 2023. 1 +[33] Robin Rombach, Andreas Blattmann, Dominik Lorenz, Patrick Esser, and Björn Ommer. High-resolution image synthesis with latent diffusion models. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 10684-10695, 2022. 4 +[34] Olaf Ronneberger, Philipp Fischer, and Thomas Brox. U-net: Convolutional networks for biomedical image segmentation. In MICCAI, 2015. 2 +[35] Jiaming Song, Chenlin Meng, and Stefano Ermon. Denoising diffusion implicit models. In ICLR, 2021. 4, 5, 6 +[36] Yang Song, Jascha Sohl-Dickstein, Diederik P Kingma, Abhishek Kumar, Stefano Ermon, and Ben Poole. Score-based generative modeling through stochastic differential equations. In International Conference on Learning Representations, 2021. 4 +[37] Ajay Sridhar, Dhruv Shah, Catherine Glossop, and Sergey Levine. Nomad: Goal masked diffusion policies for navigation and exploration. In ICRA, 2024. 4 +[38] Maria Stamatopoulou, Jianwei Liu, and Dimitrios Kanoulas. Dippest: Diffusion-based path planner for synthesizing trajectories applied on quadruped robots. arXiv preprint arXiv:2405.19232, 2024. 4 +[39] Wenchao Sun, Xuewu Lin, Yining Shi, Chuang Zhang, Hao ran Wu, and Sifa Zheng. Sparsedrive: End-to-end autonomous driving via sparse scene representation. arXiv preprint arXiv:2405.19620, 2024. 2, 6, 8 +[40] Mingxing Tan and Quoc Le. Efficientnet: Rethinking model scaling for convolutional neural networks. In ICML, 2019. 8 +[41] Wenwen Tong, Chonghao Sima, Tai Wang, Li Chen, Silei Wu, Hanming Deng, Yi Gu, Lewei Lu, Ping Luo, Dahua Lin, et al. Scene as occupancy. In ICCV, 2023. 8 +[42] Yue Wang, Vitor Campagnolo Guizilini, Tianyuan Zhang, Yilun Wang, Hang Zhao, and Justin Solomon. Detr3d: 3d object detection from multi-view images via 3d-to-2d queries. In CoRL, 2022. 1, 6 + +[43] Yuqi Wang, Jiawei He, Lue Fan, Hongxin Li, Yuntao Chen, and Zhaoxiang Zhang. Driving into the future: Multiview visual forecasting and planning with world model for autonomous driving. In CVPR, 2024. 2 +[44] Yixiao Wang, Chen Tang, Lingfeng Sun, Simone Rossi, Yichen Xie, Chensheng Peng, Thomas Hannagan, Stefano Sabatini, Nicola Poerio, Masayoshi Tomizuka, et al. Optimizing diffusion models for joint trajectory prediction and controllable generation. In ECCV, 2024. 3 +[45] Xinshuo Weng, Boris Ivanovic, Yan Wang, Yue Wang, and Marco Pavone. Para-drive: Parallelized architecture for real-time autonomous driving. In CVPR, 2024. 2, 6 +[46] Zehang Weng, Haofei Lu, Danica Kragic, and Jens Lundell. Dexdiffuser: Generating dexterous grasps with diffusion models. arXiv preprint arXiv:2402.02989, 2024. 4 +[47] Sixu Yan, Zeyu Zhang, Muzhi Han, Zaijin Wang, Qi Xie, Zhitian Li, Zhehan Li, Hangxin Liu, Xinggang Wang, and Song-Chun Zhu. M2diffuser: Diffusion-based trajectory optimization for mobile manipulation in 3d scenes. arXiv preprint arXiv:2410.11402, 2024. 3 +[48] Brian Yang, Huangyuan Su, Nikolaos Gkanatsios, Tsung-Wei Ke, Ayush Jain, Jeff Schneider, and Katerina Fragkiadaki. Diffusion-es: Gradient-free planning with diffusion for autonomous driving and zero-shot instruction following. In CVPR, 2024. 3 +[49] Jingfeng Yao and Xinggang Wang. Reconstruction vs. generation: Taming optimization dilemma in latent diffusion models. arXiv preprint arXiv:2501.01423, 2025. 4 +[50] Jingfeng Yao, Cheng Wang, Wenyu Liu, and Xinggang Wang. Fasteredit: Towards faster diffusion transformers training without architecture modification. arXiv preprint arXiv:2410.10356, 2024. 4 +[51] Wenhao Yu, Jie Peng, Huanyu Yang, Junrui Zhang, Yifan Duan, Jianmin Ji, and Yanyong Zhang. Ldp: A local diffusion planner for efficient robot navigation and collision avoidance. arXiv preprint arXiv:2407.01950, 2024. 4 +[52] Chengran Yuan, Zhanqi Zhang, Jiawei Sun, Shuo Sun, Zefan Huang, Christina Dao Wen Lee, Dongen Li, Yuhang Han, Anthony Wong, Keng Peng Tee, et al. Drama: An efficient end-to-end motion planner for autonomous driving with mamba. arXiv preprint arXiv:2408.03601, 2024. 6 +[53] Yanjie Ze, Gu Zhang, Kangning Zhang, Chenyuan Hu, Muhan Wang, and Huazhe Xu. 3d diffusion policy. In RSS, 2024. 3 +[54] Fangao Zeng, Bin Dong, Yuang Zhang, Tiancai Wang, Xi-angyu Zhang, and Yichen Wei. Motr: End-to-end multiple-object tracking with transformer. In ECCV, 2022. 1 +[55] Yifu Zhang, Chunyu Wang, Xinggang Wang, Wenjun Zeng, and Wenyu Liu. Fairmot: On the fairness of detection and re-identification in multiple object tracking. IJCV, 2021. +[56] Yifu Zhang, Peize Sun, Yi Jiang, Dongdong Yu, Fucheng Weng, Zehuan Yuan, Ping Luo, Wenyu Liu, and Xinggang Wang. Bytetrack: Multi-object tracking by associating every detection box. In ECCV, 2022. 1 +[57] Huangjie Zheng, Pengcheng He, Weizhu Chen, and Mingyuan Zhou. Truncated diffusion probabilistic models and diffusion-based adversarial auto-encoders. In The + +Eleventh International Conference on Learning Representations, 2023. 4 +[58] Wenzhao Zheng, Ruiqi Song, Xianda Guo, Chenming Zhang, and Long Chen. Genad: Generative end-to-end autonomous driving. In ECCV, 2024. 2 +[59] Ziyuan Zhong, Davis Rempe, Yuxiao Chen, Boris Ivanovic, Yulong Cao, Danfei Xu, Marco Pavone, and Baishakhi Ray. Language-guided traffic simulation via scene-level diffusion. In CoRL, 2023. 3 +[60] Ziyuan Zhong, Davis Rempe, Danfei Xu, Yuxiao Chen, Sushant Veer, Tong Che, Baishakhi Ray, and Marco Pavone. Guided conditional diffusion for controllable traffic simulation. In ICRA, 2023. 3 +[61] Lianghui Zhu, Zilong Huang, Bencheng Liao, Jun Hao Liew, Hanshu Yan, Jiashi Feng, and Xinggang Wang. Dig: Scalable and efficient diffusion models with gated linear attention. arXiv preprint arXiv:2405.18428, 2024. 4 +[62] Xizhou Zhu, Weijie Su, Lewei Lu, Bin Li, Xiaogang Wang, and Jifeng Dai. Deformable detr: Deformable transformers for end-to-end object detection. In ICLR, 2021. 2, 6 \ No newline at end of file diff --git a/2025/DiffusionDrive_ Truncated Diffusion Model for End-to-End Autonomous Driving/images.zip b/2025/DiffusionDrive_ Truncated Diffusion Model for End-to-End Autonomous Driving/images.zip new file mode 100644 index 0000000000000000000000000000000000000000..500466377e1486210b297bf7bbe53100e83957ca --- /dev/null +++ b/2025/DiffusionDrive_ Truncated Diffusion Model for End-to-End Autonomous Driving/images.zip @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:d99560863b92f1c820ade2a22bec8b19e807e1f48ac7540e53ab7977b0f4899d +size 712762 diff --git a/2025/DiffusionDrive_ Truncated Diffusion Model for End-to-End Autonomous Driving/layout.json b/2025/DiffusionDrive_ Truncated Diffusion Model for End-to-End Autonomous Driving/layout.json new file mode 100644 index 0000000000000000000000000000000000000000..b46c072117be9f5819cdda227a2cac6dedd5fcbc --- /dev/null +++ b/2025/DiffusionDrive_ Truncated Diffusion Model for End-to-End Autonomous Driving/layout.json @@ -0,0 +1,8973 @@ +{ + "pdf_info": [ + { + "para_blocks": [ + { + "bbox": [ + 56, + 103, + 555, + 122 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 103, + 555, + 122 + ], + "spans": [ + { + "bbox": [ + 56, + 103, + 555, + 122 + ], + "type": "text", + "content": "DiffusionDrive: Truncated Diffusion Model for End-to-End Autonomous Driving" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 61, + 141, + 547, + 232 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 61, + 141, + 547, + 232 + ], + "spans": [ + { + "bbox": [ + 61, + 141, + 547, + 232 + ], + "type": "text", + "content": "Bencheng Liao" + }, + { + "bbox": [ + 61, + 141, + 547, + 232 + ], + "type": "inline_equation", + "content": "^{1,2,\\diamond}" + }, + { + "bbox": [ + 61, + 141, + 547, + 232 + ], + "type": "text", + "content": " Shaoyu Chen" + }, + { + "bbox": [ + 61, + 141, + 547, + 232 + ], + "type": "inline_equation", + "content": "^{2,3}" + }, + { + "bbox": [ + 61, + 141, + 547, + 232 + ], + "type": "text", + "content": " Haoran Yin" + }, + { + "bbox": [ + 61, + 141, + 547, + 232 + ], + "type": "inline_equation", + "content": "^{3}" + }, + { + "bbox": [ + 61, + 141, + 547, + 232 + ], + "type": "text", + "content": " Bo Jiang" + }, + { + "bbox": [ + 61, + 141, + 547, + 232 + ], + "type": "inline_equation", + "content": "^{2,\\diamond}" + }, + { + "bbox": [ + 61, + 141, + 547, + 232 + ], + "type": "text", + "content": " Cheng Wang" + }, + { + "bbox": [ + 61, + 141, + 547, + 232 + ], + "type": "inline_equation", + "content": "^{1,2,\\diamond}" + }, + { + "bbox": [ + 61, + 141, + 547, + 232 + ], + "type": "text", + "content": " Sixu Yan" + }, + { + "bbox": [ + 61, + 141, + 547, + 232 + ], + "type": "inline_equation", + "content": "^{2}" + }, + { + "bbox": [ + 61, + 141, + 547, + 232 + ], + "type": "text", + "content": " Xinbang Zhang" + }, + { + "bbox": [ + 61, + 141, + 547, + 232 + ], + "type": "inline_equation", + "content": "^{3}" + }, + { + "bbox": [ + 61, + 141, + 547, + 232 + ], + "type": "text", + "content": " Xiangyu Li" + }, + { + "bbox": [ + 61, + 141, + 547, + 232 + ], + "type": "inline_equation", + "content": "^{3}" + }, + { + "bbox": [ + 61, + 141, + 547, + 232 + ], + "type": "text", + "content": " Ying Zhang" + }, + { + "bbox": [ + 61, + 141, + 547, + 232 + ], + "type": "inline_equation", + "content": "^{3}" + }, + { + "bbox": [ + 61, + 141, + 547, + 232 + ], + "type": "text", + "content": " Qian Zhang" + }, + { + "bbox": [ + 61, + 141, + 547, + 232 + ], + "type": "inline_equation", + "content": "^{3}" + }, + { + "bbox": [ + 61, + 141, + 547, + 232 + ], + "type": "text", + "content": " Xinggang Wang" + }, + { + "bbox": [ + 61, + 141, + 547, + 232 + ], + "type": "inline_equation", + "content": "^{2}" + }, + { + "bbox": [ + 61, + 141, + 547, + 232 + ], + "type": "inline_equation", + "content": "^{1}" + }, + { + "bbox": [ + 61, + 141, + 547, + 232 + ], + "type": "text", + "content": " Institute of Artificial Intelligence, Huazhong University of Science & Technology \n" + }, + { + "bbox": [ + 61, + 141, + 547, + 232 + ], + "type": "inline_equation", + "content": "^{2}" + }, + { + "bbox": [ + 61, + 141, + 547, + 232 + ], + "type": "text", + "content": " School of EIC, Huazhong University of Science & Technology \n" + }, + { + "bbox": [ + 61, + 141, + 547, + 232 + ], + "type": "inline_equation", + "content": "^{3}" + }, + { + "bbox": [ + 61, + 141, + 547, + 232 + ], + "type": "text", + "content": " Horizon Robotics \nCode & Model & Demo: hustvl/DiffusionDrive" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 151, + 258, + 200, + 271 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 151, + 258, + 200, + 271 + ], + "spans": [ + { + "bbox": [ + 151, + 258, + 200, + 271 + ], + "type": "text", + "content": "Abstract" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 54, + 284, + 297, + 582 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 54, + 284, + 297, + 582 + ], + "spans": [ + { + "bbox": [ + 54, + 284, + 297, + 582 + ], + "type": "text", + "content": "Recently, the diffusion model has emerged as a powerful generative technique for robotic policy learning, capable of modeling multi-mode action distributions. Leveraging its capability for end-to-end autonomous driving is a promising direction. However, the numerous denoising steps in the robotic diffusion policy and the more dynamic, open-world nature of traffic scenes pose substantial challenges for generating diverse driving actions at a real-time speed. To address these challenges, we propose a novel truncated diffusion policy that incorporates prior multi-mode anchors and truncates the diffusion schedule, enabling the model to learn denoising from anchored Gaussian distribution to the multi-mode driving action distribution. Additionally, we design an efficient cascade diffusion decoder for enhanced interaction with conditional scene context. The proposed model, DiffusionDrive, demonstrates " + }, + { + "bbox": [ + 54, + 284, + 297, + 582 + ], + "type": "inline_equation", + "content": "10 \\times" + }, + { + "bbox": [ + 54, + 284, + 297, + 582 + ], + "type": "text", + "content": " reduction in denoising steps compared to vanilla diffusion policy, delivering superior diversity and quality in just 2 steps. On the planning-oriented NAVSIM dataset, with aligned ResNet-34 backbone, DiffusionDrive achieves 88.1 PDMS without bells and whistles, setting a new record, while running at a real-time speed of 45 FPS on an NVIDIA 4090. Qualitative results on challenging scenarios further confirm that DiffusionDrive can robustly generate diverse plausible driving actions." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 55, + 605, + 135, + 618 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 605, + 135, + 618 + ], + "spans": [ + { + "bbox": [ + 55, + 605, + 135, + 618 + ], + "type": "text", + "content": "1. Introduction" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 55, + 625, + 296, + 686 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 625, + 296, + 686 + ], + "spans": [ + { + "bbox": [ + 55, + 625, + 296, + 686 + ], + "type": "text", + "content": "End-to-end autonomous driving has gained significant attention in recent years due to advancements in perception models (detection [5, 17, 24, 42], tracking [54-56], online mapping [27, 28, 32], etc.), which directly learns the driving policy from the raw sensor inputs. This data-driven ap" + } + ] + } + ], + "index": 7 + }, + { + "type": "image", + "bbox": [ + 315, + 258, + 556, + 397 + ], + "blocks": [ + { + "bbox": [ + 315, + 258, + 556, + 397 + ], + "lines": [ + { + "bbox": [ + 315, + 258, + 556, + 397 + ], + "spans": [ + { + "bbox": [ + 315, + 258, + 556, + 397 + ], + "type": "image", + "image_path": "99b5d280af68662c38a293d2104fb29b5b421a613a42c76a54486e13cbbd3e04.jpg" + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 313, + 399, + 555, + 442 + ], + "lines": [ + { + "bbox": [ + 313, + 399, + 555, + 442 + ], + "spans": [ + { + "bbox": [ + 313, + 399, + 555, + 442 + ], + "type": "text", + "content": "Figure 1. The comparison of different end-to-end paradigms. (a) Single mode regression [8, 16, 20]. (b) Sampling from vocabulary [4, 25]. (c) Vanilla diffusion policy [7, 19]. (d) The proposed truncated diffusion policy." + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "image_caption" + } + ], + "index": 8 + }, + { + "bbox": [ + 313, + 445, + 555, + 480 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 445, + 555, + 480 + ], + "spans": [ + { + "bbox": [ + 313, + 445, + 555, + 480 + ], + "type": "text", + "content": "proach offers a scalable and robust alternative to traditional rule-based motion planning, which often struggles to generalize to complex real-world driving settings." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 312, + 481, + 556, + 685 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 312, + 481, + 556, + 685 + ], + "spans": [ + { + "bbox": [ + 312, + 481, + 556, + 685 + ], + "type": "text", + "content": "To effectively learn from data, mainstream end-to-end planners (e.g., Transfuser [8], UniAD [16], VAD [20]) typically regress a single-mode trajectory from an ego-query as shown in Fig. 1a. However, this paradigm does not account for the inherent uncertainty and multi-mode nature of driving behaviors. Recently, VADv2 [20] introduces a large fixed vocabulary of anchor trajectories (4096 anchors) to discretize the continuous action space and capture a broader range of driving behaviors, and then samples from these anchors based on predicted scores as shown in Fig. 1b. However, this large fixed-vocabulary paradigm is fundamentally constrained by the number and quality of anchor trajectories, often failing in out-of-vocabulary scenarios. Furthermore, managing a large number of anchors presents significant computational challenges for real-time applications. Rather than discretizing the action space, diffusion model [7] has proven to be a powerful genera" + } + ] + } + ], + "index": 11 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "spans": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "type": "text", + "content": "CVF" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 145, + 0, + 496, + 37 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 145, + 0, + 496, + 37 + ], + "spans": [ + { + "bbox": [ + 145, + 0, + 496, + 37 + ], + "type": "text", + "content": "This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 313, + 693, + 555, + 713 + ], + "type": "page_footnote", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 693, + 555, + 713 + ], + "spans": [ + { + "bbox": [ + 313, + 693, + 555, + 713 + ], + "type": "text", + "content": "To distinguish the term \"multimodal\" used to describe input data, we use \"multi-mode\" in this paper to refer to diverse planning decisions." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 55, + 693, + 295, + 713 + ], + "type": "page_footnote", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 693, + 295, + 713 + ], + "spans": [ + { + "bbox": [ + 55, + 693, + 295, + 713 + ], + "type": "inline_equation", + "content": "\\diamond" + }, + { + "bbox": [ + 55, + 693, + 295, + 713 + ], + "type": "text", + "content": " Intern of Horizon Robotics; " + }, + { + "bbox": [ + 55, + 693, + 295, + 713 + ], + "type": "inline_equation", + "content": "\\boxtimes" + }, + { + "bbox": [ + 55, + 693, + 295, + 713 + ], + "type": "text", + "content": " Corresponding author: Xinggang Wang (xgwang@hust.edu.cn)." + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "text", + "content": "12037" + } + ] + } + ], + "index": 14 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 0 + }, + { + "para_blocks": [ + { + "bbox": [ + 55, + 72, + 294, + 108 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 72, + 294, + 108 + ], + "spans": [ + { + "bbox": [ + 55, + 72, + 294, + 108 + ], + "type": "text", + "content": "tive decision-making policy in the robotics domain, which can directly sample multi-mode physically plausible actions from a Gaussian distribution via iterative denoising process." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 54, + 110, + 295, + 301 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 54, + 110, + 295, + 301 + ], + "spans": [ + { + "bbox": [ + 54, + 110, + 295, + 301 + ], + "type": "text", + "content": "This inspires us to replicate the success of the diffusion model in the robotics domain to end-to-end autonomous driving. We apply the vanilla robotic diffusion policy to the well-known single-mode-regression method, Transfuser [8], by proposing a variant, " + }, + { + "bbox": [ + 54, + 110, + 295, + 301 + ], + "type": "inline_equation", + "content": "\\text{Transfuser}_{\\text{DP}}" + }, + { + "bbox": [ + 54, + 110, + 295, + 301 + ], + "type": "text", + "content": ", which replaces the deterministic MLP regression head with a conditional diffusion model [34]. Though " + }, + { + "bbox": [ + 54, + 110, + 295, + 301 + ], + "type": "inline_equation", + "content": "\\text{Transfuser}_{\\text{DP}}" + }, + { + "bbox": [ + 54, + 110, + 295, + 301 + ], + "type": "text", + "content": " improves planning performance, two major issues arise: 1) The numerous 20 denoising steps in the vanilla DDIM diffusion policy introduce heavy computational consumption during inference as shown in Tab. 2, hindering the real-time application for autonomous driving. 2) The trajectories sampled from different Gaussian noises severely overlap with each other, as illustrated in Fig. 2. This underscores the non-trivial challenge of taming the diffusion models for the dynamic and open-world traffic scenes." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 56, + 304, + 295, + 542 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 304, + 295, + 542 + ], + "spans": [ + { + "bbox": [ + 56, + 304, + 295, + 542 + ], + "type": "text", + "content": "Unlike the vanilla diffusion policy, which samples actions from a random Gaussian noise conditioned on scene context, human drivers adhere to established driving patterns that they dynamically adjust in response to real-time traffic conditions. This insight motivates us to embed these prior driving patterns into the diffusion policy by partitioning the Gaussian distribution into multiple sub-Gaussian distributions centered around prior anchors, referred to as anchored Gaussian distribution. It is implemented by truncating the diffusion schedule to introduce a small portion of Gaussian noise around the prior anchors as shown in Fig. 3. Thanks to the multi-mode distributional expressivity of the diffusion model, the proposed truncated diffusion policy effectively covers the potential action space without requiring a large set of fixed anchors, as VADv2 does. With more reasonable initial noise samples from the anchored Gaussian distribution, we can truncate the denoising process, reducing the required steps from 20 to just 2—a substantial speedup that satisfies the real-time requirements of autonomous driving." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 55, + 544, + 295, + 652 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 544, + 295, + 652 + ], + "spans": [ + { + "bbox": [ + 55, + 544, + 295, + 652 + ], + "type": "text", + "content": "To enhance the interaction with conditional scene context, we propose an efficient transformer-based diffusion decoder that interacts not only with structured queries from the perception module but also with Bird's Eye View (BEV) and perspective view (PV) features through a sparse deformable attention mechanism [3, 62]. Additionally, we introduce a cascade mechanism to iteratively refine the trajectory reconstruction within the diffusion decoder at each denoising step." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 55, + 653, + 295, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 653, + 295, + 713 + ], + "spans": [ + { + "bbox": [ + 55, + 653, + 295, + 713 + ], + "type": "text", + "content": "With these innovations, we present DiffusionDrive, a diffusion model for real-time end-to-end autonomous driving. We benchmark our method on the planning-oriented NAVSIM dataset [11] using non-reactive simulation and closed-loop evaluations. Without bells and whis" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 313, + 72, + 553, + 263 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 72, + 553, + 263 + ], + "spans": [ + { + "bbox": [ + 313, + 72, + 553, + 263 + ], + "type": "text", + "content": "tles, DiffusionDrive achieves 88.1 PDMS on NAVSIM navtest split with the aligned ResNet-34 backbone, significantly outperforming previous state-of-the-art methods. Even compared to the NAVSIM challenge-winning solution Hydra-MDP-" + }, + { + "bbox": [ + 313, + 72, + 553, + 263 + ], + "type": "inline_equation", + "content": "\\nu_{8192}" + }, + { + "bbox": [ + 313, + 72, + 553, + 263 + ], + "type": "text", + "content": "-W-EP [25], which follows VADv2 with 8192 anchor trajectories and further incorporates post-processing and additional supervision, DiffusionDrive still outperforms it by 1.6 PDMS through directly learning from human demonstrations and inferring without post-processing, while running at real-time speed of 45 FPS on an NVIDIA 4090. We further validate the superiority of DiffusionDrive on popular nuScenes dataset [2] with open-loop evaluations, DiffusionDrive runs " + }, + { + "bbox": [ + 313, + 72, + 553, + 263 + ], + "type": "inline_equation", + "content": "1.8 \\times" + }, + { + "bbox": [ + 313, + 72, + 553, + 263 + ], + "type": "text", + "content": " faster than VAD and outperforms it [20] by " + }, + { + "bbox": [ + 313, + 72, + 553, + 263 + ], + "type": "inline_equation", + "content": "20.8\\%" + }, + { + "bbox": [ + 313, + 72, + 553, + 263 + ], + "type": "text", + "content": " lower L2 error and " + }, + { + "bbox": [ + 313, + 72, + 553, + 263 + ], + "type": "inline_equation", + "content": "63.6\\%" + }, + { + "bbox": [ + 313, + 72, + 553, + 263 + ], + "type": "text", + "content": " lower collision rate with the same ResNet-50 backbone, demonstrating state-of-the-art planning performance." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 326, + 264, + 525, + 274 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 326, + 264, + 525, + 274 + ], + "spans": [ + { + "bbox": [ + 326, + 264, + 525, + 274 + ], + "type": "text", + "content": "Our contributions can be summarized as follows:" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 314, + 275, + 553, + 479 + ], + "type": "list", + "angle": 0, + "index": 11, + "blocks": [ + { + "bbox": [ + 314, + 275, + 553, + 335 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 314, + 275, + 553, + 335 + ], + "spans": [ + { + "bbox": [ + 314, + 275, + 553, + 335 + ], + "type": "text", + "content": "- We firstly introduce the diffusion model to the field of end-to-end autonomous driving and propose a novel truncated diffusion policy to address the issues of mode collapse and heavy computational overhead found in direct adaptation of vanilla diffusion policy to the traffic scene." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 315, + 335, + 553, + 371 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 315, + 335, + 553, + 371 + ], + "spans": [ + { + "bbox": [ + 315, + 335, + 553, + 371 + ], + "type": "text", + "content": "- We design an efficient transformer-based diffusion decoder that interacts with the conditional information in a cascaded manner for better trajectory reconstruction." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 315, + 372, + 553, + 430 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 315, + 372, + 553, + 430 + ], + "spans": [ + { + "bbox": [ + 315, + 372, + 553, + 430 + ], + "type": "text", + "content": "- Without bells and whistles, DiffusionDrive significantly outperforms previous state-of-the-art methods, achieving a record-breaking 88.1 PDMS on the NAVSIM navtest split with the same backbone, while maintaining real-time performance at 45 FPS on an NVIDIA 4090." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 315, + 431, + 553, + 479 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 315, + 431, + 553, + 479 + ], + "spans": [ + { + "bbox": [ + 315, + 431, + 553, + 479 + ], + "type": "text", + "content": "- We qualitatively demonstrate that DiffusionDrive can generate more diverse and plausible trajectories, exhibiting high-quality multi-mode driving actions in various challenging scenarios." + } + ] + } + ], + "index": 10 + } + ], + "sub_type": "text" + }, + { + "bbox": [ + 314, + 488, + 400, + 501 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 314, + 488, + 400, + 501 + ], + "spans": [ + { + "bbox": [ + 314, + 488, + 400, + 501 + ], + "type": "text", + "content": "2. Related Work" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 313, + 510, + 555, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 510, + 555, + 713 + ], + "spans": [ + { + "bbox": [ + 313, + 510, + 555, + 713 + ], + "type": "text", + "content": "End-to-end autonomous driving. UniAD [16], as a pioneering work, demonstrates the potential of end-to-end autonomous driving by integrating multiple perception tasks to enhance planning performance. VAD [20] further explores the use of compact vectorized scene representations to improve efficiency. Subsequently, a series of works [6, 8, 12, 23, 26, 43, 45, 58] have adopted the single-trajectory planning paradigm to enhance planning performance further. More recently, VADv2 [4] shifts the paradigm towards multi-mode planning by scoring and sampling from a large fixed vocabulary of anchor trajectories. Hydra-MDP [25] improves the scoring mechanism of VADv2 by introducing extra supervision from a rule-based scorer. SparseDrive [39] explores an alternative BEV-free solution. Unlike existing multi-mode planning approaches, we propose a novel paradigm that leverages powerful generative diffusion models for end-to-end autonomous driving." + } + ] + } + ], + "index": 13 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "text", + "content": "12038" + } + ] + } + ], + "index": 14 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 1 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 104, + 68, + 506, + 470 + ], + "blocks": [ + { + "bbox": [ + 104, + 68, + 506, + 470 + ], + "lines": [ + { + "bbox": [ + 104, + 68, + 506, + 470 + ], + "spans": [ + { + "bbox": [ + 104, + 68, + 506, + 470 + ], + "type": "image", + "image_path": "eae6cbced2a079f01e0437de2d3082d21ee7f04b1c9e9400b61c9f8776e04382.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 204, + 472, + 405, + 481 + ], + "lines": [ + { + "bbox": [ + 204, + 472, + 405, + 481 + ], + "spans": [ + { + "bbox": [ + 204, + 472, + 405, + 481 + ], + "type": "text", + "content": "(b) Top-1's turning left and diverse top-10's lane changing." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 55, + 481, + 555, + 545 + ], + "lines": [ + { + "bbox": [ + 55, + 481, + 555, + 545 + ], + "spans": [ + { + "bbox": [ + 55, + 481, + 555, + 545 + ], + "type": "text", + "content": "Figure 2. Qualitative comparison of Transfuser, TransfuserDP and DiffusionDrive on challenging scenes of NAVSIM navtest split. With the same inputs from front cameras and LiDAR, DiffusionDrive achieves the highest planning quality of top-1 scoring trajectory as illustrated in Tab. 2. We render the highlighted diverse trajectories predicted by DiffusionDrive in the front view. (a) and (b) shows that the top-1 scoring trajectory of DiffusionDrive closely matches the ground truth for both going straight and turning left. Additionally, DiffusionDrive's top-10 scoring trajectory demonstrates high-quality lane changing—an ability not observed in multi-mode TransfuserDP and impossible for Transfuser." + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "bbox": [ + 55, + 550, + 297, + 706 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 550, + 297, + 706 + ], + "spans": [ + { + "bbox": [ + 55, + 550, + 297, + 706 + ], + "type": "text", + "content": "Diffusion model for traffic simulation. Driving diffusion policy has been explored in the traffic simulation by leveraging only abstract perception groundtruth [9, 18, 21, 44]. MotionDiffuser [21] and CTG [60] are pioneering applications of diffusion models for multi-agent motion prediction, using a conditional diffusion model to sample target trajectories from Gaussian noise. " + }, + { + "bbox": [ + 55, + 550, + 297, + 706 + ], + "type": "inline_equation", + "content": "\\mathrm{CTG + + }" + }, + { + "bbox": [ + 55, + 550, + 297, + 706 + ], + "type": "text", + "content": " [59] further incorporates a large language model (LLM) for language-driven guidance, improving usability and enabling realistic traffic simulations. Diffusion-ES [48] replaces reward-gradient-guided denoising with evolutionary search. Moving beyond diffusion models limited to traffic simulation with perception groundtruth, our approach unlocks the potential of dif" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 313, + 550, + 555, + 586 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 550, + 555, + 586 + ], + "spans": [ + { + "bbox": [ + 313, + 550, + 555, + 586 + ], + "type": "text", + "content": "fusion models for real-time, end-to-end autonomous driving through our proposed truncated diffusion policy and efficient diffusion decoder." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 313, + 594, + 556, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 594, + 556, + 715 + ], + "spans": [ + { + "bbox": [ + 313, + 594, + 556, + 715 + ], + "type": "text", + "content": "Diffusion model for robotic policy learning. Diffusion policy [7] demonstrates the great potential in robotic policy learning, effectively capturing multi-mode action distributions and high-dimensional action spaces. Diffuser [19] proposes an unconditional diffusion model for trajectory sampling, incorporating techniques such as classifier-free guidance and image inpainting to achieve guided sampling. Subsequently, numerous works have applied diffusion models to various robotic tasks, including stationary manipulation [1, 53], mobile manipulation [47], autonomous nav" + } + ] + } + ], + "index": 5 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 318, + 758 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 318, + 758 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 318, + 758 + ], + "type": "text", + "content": "12039" + } + ] + } + ], + "index": 6 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 2 + }, + { + "para_blocks": [ + { + "bbox": [ + 234, + 72, + 373, + 87 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 234, + 72, + 373, + 87 + ], + "spans": [ + { + "bbox": [ + 234, + 72, + 373, + 87 + ], + "type": "text", + "content": "Diffusion step: " + }, + { + "bbox": [ + 234, + 72, + 373, + 87 + ], + "type": "inline_equation", + "content": "q(\\tau^i |\\tau^{i - 1}),i\\in [1,T]" + } + ] + } + ], + "index": 0 + }, + { + "type": "image", + "bbox": [ + 105, + 89, + 506, + 194 + ], + "blocks": [ + { + "bbox": [ + 105, + 89, + 506, + 194 + ], + "lines": [ + { + "bbox": [ + 105, + 89, + 506, + 194 + ], + "spans": [ + { + "bbox": [ + 105, + 89, + 506, + 194 + ], + "type": "image", + "image_path": "8cfefc6af824cac1c473f324a6793a77eea43f897f66e0c85228a8fc8fda716a.jpg" + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_body" + } + ], + "index": 1 + }, + { + "bbox": [ + 229, + 198, + 383, + 212 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 229, + 198, + 383, + 212 + ], + "spans": [ + { + "bbox": [ + 229, + 198, + 383, + 212 + ], + "type": "text", + "content": "Denoising step: " + }, + { + "bbox": [ + 229, + 198, + 383, + 212 + ], + "type": "inline_equation", + "content": "p_{\\theta}\\big(\\tau^i |\\tau^{i - 1},z\\big),i\\in [1,T]" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 184, + 219, + 425, + 230 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 184, + 219, + 425, + 230 + ], + "spans": [ + { + "bbox": [ + 184, + 219, + 425, + 230 + ], + "type": "text", + "content": "Vanilla Diffusion Policy vs. Truncated Diffusion Policy (ours)" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 129, + 239, + 288, + 251 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 129, + 239, + 288, + 251 + ], + "spans": [ + { + "bbox": [ + 129, + 239, + 288, + 251 + ], + "type": "text", + "content": "Diffusion step: " + }, + { + "bbox": [ + 129, + 239, + 288, + 251 + ], + "type": "inline_equation", + "content": "q(\\tau^i |\\tau^{i - 1}),i\\in [1,T_{\\mathrm{trunc}}]" + } + ] + } + ], + "index": 4 + }, + { + "type": "image", + "bbox": [ + 108, + 254, + 293, + 358 + ], + "blocks": [ + { + "bbox": [ + 108, + 254, + 293, + 358 + ], + "lines": [ + { + "bbox": [ + 108, + 254, + 293, + 358 + ], + "spans": [ + { + "bbox": [ + 108, + 254, + 293, + 358 + ], + "type": "image", + "image_path": "908ca90015f3cdef9e8a7a268d4d0e4dffad49b35c5355a18d9a23521945317e.jpg" + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 112, + 360, + 293, + 369 + ], + "lines": [ + { + "bbox": [ + 112, + 360, + 293, + 369 + ], + "spans": [ + { + "bbox": [ + 112, + 360, + 293, + 369 + ], + "type": "text", + "content": "Adding noise to the anchors with truncated diffusion schedule." + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "image_caption" + } + ], + "index": 5 + }, + { + "bbox": [ + 315, + 240, + 501, + 250 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 315, + 240, + 501, + 250 + ], + "spans": [ + { + "bbox": [ + 315, + 240, + 501, + 250 + ], + "type": "text", + "content": "Denosing from anchored Gaussian distribution with fewer steps." + } + ] + } + ], + "index": 7 + }, + { + "type": "image", + "bbox": [ + 304, + 250, + 503, + 358 + ], + "blocks": [ + { + "bbox": [ + 304, + 250, + 503, + 358 + ], + "lines": [ + { + "bbox": [ + 304, + 250, + 503, + 358 + ], + "spans": [ + { + "bbox": [ + 304, + 250, + 503, + 358 + ], + "type": "image", + "image_path": "90462a069ee2704c6c9619343f1e37dcc31b96ef9d735cc9113ee30ae82e944a.jpg" + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 321, + 362, + 494, + 375 + ], + "lines": [ + { + "bbox": [ + 321, + 362, + 494, + 375 + ], + "spans": [ + { + "bbox": [ + 321, + 362, + 494, + 375 + ], + "type": "text", + "content": "Denoising step: " + }, + { + "bbox": [ + 321, + 362, + 494, + 375 + ], + "type": "inline_equation", + "content": "p_{\\theta}(\\tau^i |\\tau^{i - 1},z),i\\in [1,T_{\\mathrm{trunc}}]" + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 54, + 381, + 555, + 426 + ], + "lines": [ + { + "bbox": [ + 54, + 381, + 555, + 426 + ], + "spans": [ + { + "bbox": [ + 54, + 381, + 555, + 426 + ], + "type": "text", + "content": "Figure 3. Illustration of truncated diffusion policy by comparing with vanilla diffusion policy. We truncate the diffusion process and only add a small portion of Gaussian noise to diffuse the anchor trajectories. Then, we train the diffusion model to reconstruct the ground-truth trajectory from the anchored Gaussian distribution with conditional scene context. During the inference, we also truncate the denoising process by starting from the better samples in the anchored Gaussian distribution than the pure Gaussian noise." + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "image_caption" + } + ], + "index": 8 + }, + { + "bbox": [ + 54, + 430, + 295, + 537 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 54, + 430, + 295, + 537 + ], + "spans": [ + { + "bbox": [ + 54, + 430, + 295, + 537 + ], + "type": "text", + "content": "igation [37, 51], quadruped locomotion [38], and dexterous manipulation [46]. However, directly applying vanilla diffusion policy to end-to-end autonomous driving poses unique challenges, as it requires real-time efficiency and the generation of plausible multi-mode trajectories in dynamic and open-world traffic scenes. In this work, we propose a novel truncated diffusion policy to address these challenges, introducing concepts that have not yet been explored in the robotics field." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 55, + 545, + 295, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 545, + 295, + 713 + ], + "spans": [ + { + "bbox": [ + 55, + 545, + 295, + 713 + ], + "type": "text", + "content": "Diffusion model for image generation. Diffusion models have been extensively adopted for image generation tasks [33, 36, 49, 50, 61]. DDIM [35] enhances DDPM [14] by enabling efficient sampling with significantly fewer steps based on non-Markovian diffusion processes. Flow matching [30, 31] further optimizes the generative process by directly modeling continuous probability flows. TDPM [57] proposes truncated denoising, which initiates the generation process from an implicit intermediate distribution to accelerate sampling. In contrast to these approaches, our method introduces an explicit driving prior within the diffusion policy, effectively guiding the diffusion process toward more accurate and efficient generation tailored specifically for end-to-end autonomous driving." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 314, + 429, + 370, + 441 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 314, + 429, + 370, + 441 + ], + "spans": [ + { + "bbox": [ + 314, + 429, + 370, + 441 + ], + "type": "text", + "content": "3. Method" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 313, + 448, + 394, + 460 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 448, + 394, + 460 + ], + "spans": [ + { + "bbox": [ + 313, + 448, + 394, + 460 + ], + "type": "text", + "content": "3.1. Preliminary" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 313, + 467, + 555, + 540 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 467, + 555, + 540 + ], + "spans": [ + { + "bbox": [ + 313, + 467, + 555, + 540 + ], + "type": "text", + "content": "Task formulation. End-to-end autonomous driving takes raw sensor data as input and predicts the future trajectory of the ego-vehicle. The trajectory is represented as a sequence of waypoints " + }, + { + "bbox": [ + 313, + 467, + 555, + 540 + ], + "type": "inline_equation", + "content": "\\tau = \\{(x_{t},y_{t})\\}_{t = 1}^{T_{f}}" + }, + { + "bbox": [ + 313, + 467, + 555, + 540 + ], + "type": "text", + "content": ", where " + }, + { + "bbox": [ + 313, + 467, + 555, + 540 + ], + "type": "inline_equation", + "content": "T_{f}" + }, + { + "bbox": [ + 313, + 467, + 555, + 540 + ], + "type": "text", + "content": " denotes the planning horizon, and " + }, + { + "bbox": [ + 313, + 467, + 555, + 540 + ], + "type": "inline_equation", + "content": "(x_{t},y_{t})" + }, + { + "bbox": [ + 313, + 467, + 555, + 540 + ], + "type": "text", + "content": " is the location of each waypoint at time " + }, + { + "bbox": [ + 313, + 467, + 555, + 540 + ], + "type": "inline_equation", + "content": "t" + }, + { + "bbox": [ + 313, + 467, + 555, + 540 + ], + "type": "text", + "content": " in the current ego-vehicle coordinate system." + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 313, + 541, + 554, + 578 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 541, + 554, + 578 + ], + "spans": [ + { + "bbox": [ + 313, + 541, + 554, + 578 + ], + "type": "text", + "content": "Conditional diffusion model. The conditional diffusion model poses a forward diffusion process as gradually adding noise to the data sample, which can be defined as:" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 345, + 586, + 555, + 607 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 345, + 586, + 555, + 607 + ], + "spans": [ + { + "bbox": [ + 345, + 586, + 555, + 607 + ], + "type": "interline_equation", + "content": "q \\left(\\tau^ {i} \\mid \\tau^ {0}\\right) = \\mathcal {N} \\left(\\tau^ {i}; \\sqrt {\\bar {\\alpha} ^ {i}} \\boldsymbol {\\tau} ^ {0}, \\left(1 - \\bar {\\alpha} ^ {i}\\right) \\mathbf {I}\\right), \\tag {1}", + "image_path": "388bffc90f3be35885db9476bceffadf6e9164c7fe87869b5fe0718f96c4dbff.jpg" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 313, + 616, + 555, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 616, + 555, + 713 + ], + "spans": [ + { + "bbox": [ + 313, + 616, + 555, + 713 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 313, + 616, + 555, + 713 + ], + "type": "inline_equation", + "content": "\\tau^0" + }, + { + "bbox": [ + 313, + 616, + 555, + 713 + ], + "type": "text", + "content": " is the clean data sample, and " + }, + { + "bbox": [ + 313, + 616, + 555, + 713 + ], + "type": "inline_equation", + "content": "\\tau^i" + }, + { + "bbox": [ + 313, + 616, + 555, + 713 + ], + "type": "text", + "content": " is the data sample with noise at time " + }, + { + "bbox": [ + 313, + 616, + 555, + 713 + ], + "type": "inline_equation", + "content": "i" + }, + { + "bbox": [ + 313, + 616, + 555, + 713 + ], + "type": "text", + "content": " (Note: we use superscript " + }, + { + "bbox": [ + 313, + 616, + 555, + 713 + ], + "type": "inline_equation", + "content": "i" + }, + { + "bbox": [ + 313, + 616, + 555, + 713 + ], + "type": "text", + "content": " to denote diffusion timestep). The constant " + }, + { + "bbox": [ + 313, + 616, + 555, + 713 + ], + "type": "inline_equation", + "content": "\\bar{\\alpha}^i = \\prod_{s=1}^i \\alpha^s = \\prod_{s=1}^i (1 - \\beta^s)" + }, + { + "bbox": [ + 313, + 616, + 555, + 713 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 313, + 616, + 555, + 713 + ], + "type": "inline_equation", + "content": "\\beta^s" + }, + { + "bbox": [ + 313, + 616, + 555, + 713 + ], + "type": "text", + "content": " is the noise schedule. We train the reverse process model " + }, + { + "bbox": [ + 313, + 616, + 555, + 713 + ], + "type": "inline_equation", + "content": "f_\\theta(\\tau^i, z, i)" + }, + { + "bbox": [ + 313, + 616, + 555, + 713 + ], + "type": "text", + "content": " to predict " + }, + { + "bbox": [ + 313, + 616, + 555, + 713 + ], + "type": "inline_equation", + "content": "\\tau^0" + }, + { + "bbox": [ + 313, + 616, + 555, + 713 + ], + "type": "text", + "content": " from " + }, + { + "bbox": [ + 313, + 616, + 555, + 713 + ], + "type": "inline_equation", + "content": "\\tau^i" + }, + { + "bbox": [ + 313, + 616, + 555, + 713 + ], + "type": "text", + "content": " with the guidance of conditional information " + }, + { + "bbox": [ + 313, + 616, + 555, + 713 + ], + "type": "inline_equation", + "content": "z" + }, + { + "bbox": [ + 313, + 616, + 555, + 713 + ], + "type": "text", + "content": ", where " + }, + { + "bbox": [ + 313, + 616, + 555, + 713 + ], + "type": "inline_equation", + "content": "\\theta" + }, + { + "bbox": [ + 313, + 616, + 555, + 713 + ], + "type": "text", + "content": " is the trainable model parameter. During inference, the trained diffusion model " + }, + { + "bbox": [ + 313, + 616, + 555, + 713 + ], + "type": "inline_equation", + "content": "f_\\theta" + }, + { + "bbox": [ + 313, + 616, + 555, + 713 + ], + "type": "text", + "content": " progressively refines from the random" + } + ] + } + ], + "index": 18 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "type": "text", + "content": "12040" + } + ] + } + ], + "index": 19 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 3 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 56, + 69, + 315, + 205 + ], + "blocks": [ + { + "bbox": [ + 56, + 69, + 315, + 205 + ], + "lines": [ + { + "bbox": [ + 56, + 69, + 315, + 205 + ], + "spans": [ + { + "bbox": [ + 56, + 69, + 315, + 205 + ], + "type": "image", + "image_path": "39188c906af22969f52df93d3eccb03e17c013632528cf929e933086a49a30e7.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 55, + 209, + 555, + 243 + ], + "lines": [ + { + "bbox": [ + 55, + 209, + 555, + 243 + ], + "spans": [ + { + "bbox": [ + 55, + 209, + 555, + 243 + ], + "type": "text", + "content": "Figure 4. Overall architecture of DiffusionDrive. (a) DiffusionDrive can integrate various existing perception modules and sensor inputs. (b) The designed diffusion decoder takes the sampled noisy trajectories from anchored Gaussian distribution as input and progressively denoises them with enhanced interactions with the conditional scene context in a cascade manner to generate the final predictions." + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "type": "image", + "bbox": [ + 318, + 71, + 556, + 207 + ], + "blocks": [ + { + "bbox": [ + 318, + 71, + 556, + 207 + ], + "lines": [ + { + "bbox": [ + 318, + 71, + 556, + 207 + ], + "spans": [ + { + "bbox": [ + 318, + 71, + 556, + 207 + ], + "type": "image", + "image_path": "3501d4a7121777ea39ebdd62ca4e2b3efafa5c0f6dd4428085611e154596d12e.jpg" + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_body" + } + ], + "index": 1 + }, + { + "bbox": [ + 55, + 246, + 297, + 283 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 246, + 297, + 283 + ], + "spans": [ + { + "bbox": [ + 55, + 246, + 297, + 283 + ], + "type": "text", + "content": "noise " + }, + { + "bbox": [ + 55, + 246, + 297, + 283 + ], + "type": "inline_equation", + "content": "\\tau^T" + }, + { + "bbox": [ + 55, + 246, + 297, + 283 + ], + "type": "text", + "content": " sampled in Gaussian distribution to the predicted clean data sample " + }, + { + "bbox": [ + 55, + 246, + 297, + 283 + ], + "type": "inline_equation", + "content": "\\tau^0" + }, + { + "bbox": [ + 55, + 246, + 297, + 283 + ], + "type": "text", + "content": " with the guidance of conditional information " + }, + { + "bbox": [ + 55, + 246, + 297, + 283 + ], + "type": "inline_equation", + "content": "z" + }, + { + "bbox": [ + 55, + 246, + 297, + 283 + ], + "type": "text", + "content": ", which is defined as:" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 61, + 291, + 296, + 324 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 61, + 291, + 296, + 324 + ], + "spans": [ + { + "bbox": [ + 61, + 291, + 296, + 324 + ], + "type": "interline_equation", + "content": "p _ {\\theta} \\left(\\boldsymbol {\\tau} ^ {0} \\mid z\\right) = \\int p \\left(\\boldsymbol {\\tau} ^ {T}\\right) \\prod_ {i = 1} ^ {T} p _ {\\theta} \\left(\\boldsymbol {\\tau} ^ {i - 1} \\mid \\boldsymbol {\\tau} ^ {i}, z\\right) d \\boldsymbol {\\tau} ^ {1: T}. \\tag {2}", + "image_path": "7020815be24a1532d82505839ae520da00e2b45f9664128425c96f73f2cfd0ff.jpg" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 55, + 333, + 139, + 345 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 333, + 139, + 345 + ], + "spans": [ + { + "bbox": [ + 55, + 333, + 139, + 345 + ], + "type": "text", + "content": "3.2. Investigation" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 55, + 352, + 296, + 460 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 352, + 296, + 460 + ], + "spans": [ + { + "bbox": [ + 55, + 352, + 296, + 460 + ], + "type": "text", + "content": "Turn Transfuser [8] into conditional diffusion model. We begin from the representative deterministic end-to-end planner Transfuser [8] and turn it into a generative model " + }, + { + "bbox": [ + 55, + 352, + 296, + 460 + ], + "type": "inline_equation", + "content": "\\mathrm{Transfuser}_{\\mathrm{DP}}" + }, + { + "bbox": [ + 55, + 352, + 296, + 460 + ], + "type": "text", + "content": " by simply replacing the regression MLP layers with the conditional diffusion model UNet following vanilla diffusion policy [7]. During the evaluation, we sample a random noise and progressively refine it with 20 steps. Tab. 2 shows that " + }, + { + "bbox": [ + 55, + 352, + 296, + 460 + ], + "type": "inline_equation", + "content": "\\mathrm{Transfuser}_{\\mathrm{DP}}" + }, + { + "bbox": [ + 55, + 352, + 296, + 460 + ], + "type": "text", + "content": " achieves better planning quality than deterministic Transfuser." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 55, + 461, + 296, + 579 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 461, + 296, + 579 + ], + "spans": [ + { + "bbox": [ + 55, + 461, + 296, + 579 + ], + "type": "text", + "content": "Mode collapse. To further investigate the multi-mode property of the vanilla diffusion policy in driving, we sampled 20 random noises from Gaussian distribution and denoised them using 20 steps. As shown in Fig. 2, the different random noises converge to similar trajectories after the denoising process. To quantitatively analyze the phenomenon of mode collapse, we define a mode diversity score " + }, + { + "bbox": [ + 55, + 461, + 296, + 579 + ], + "type": "inline_equation", + "content": "\\mathcal{D}" + }, + { + "bbox": [ + 55, + 461, + 296, + 579 + ], + "type": "text", + "content": " based on the mean Intersection over Union (mIoU) between each denoised trajectory and the union of all denoised trajectories:" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 97, + 578, + 295, + 612 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 97, + 578, + 295, + 612 + ], + "spans": [ + { + "bbox": [ + 97, + 578, + 295, + 612 + ], + "type": "interline_equation", + "content": "\\mathcal {D} = 1 - \\frac {1}{N} \\sum_ {i = 1} ^ {N} \\frac {\\operatorname {A r e a} \\left(\\tau_ {i} \\cap \\bigcup_ {j = 1} ^ {N} \\tau_ {j}\\right)}{\\operatorname {A r e a} \\left(\\tau_ {i} \\cup \\bigcup_ {j = 1} ^ {N} \\tau_ {j}\\right)}, \\tag {3}", + "image_path": "0b6c9549c78f9180e082453a83aaa85ffa234f8e3607124e43f609f1e479a5e2.jpg" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 55, + 616, + 296, + 688 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 616, + 296, + 688 + ], + "spans": [ + { + "bbox": [ + 55, + 616, + 296, + 688 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 55, + 616, + 296, + 688 + ], + "type": "inline_equation", + "content": "\\tau_{i}" + }, + { + "bbox": [ + 55, + 616, + 296, + 688 + ], + "type": "text", + "content": " represents the " + }, + { + "bbox": [ + 55, + 616, + 296, + 688 + ], + "type": "inline_equation", + "content": "i" + }, + { + "bbox": [ + 55, + 616, + 296, + 688 + ], + "type": "text", + "content": "-th denoised trajectory, " + }, + { + "bbox": [ + 55, + 616, + 296, + 688 + ], + "type": "inline_equation", + "content": "N" + }, + { + "bbox": [ + 55, + 616, + 296, + 688 + ], + "type": "text", + "content": " is the total number of sampled trajectories and " + }, + { + "bbox": [ + 55, + 616, + 296, + 688 + ], + "type": "inline_equation", + "content": "\\bigcup_{j=1}^{N} \\tau_{j}" + }, + { + "bbox": [ + 55, + 616, + 296, + 688 + ], + "type": "text", + "content": " is the union of all denoised trajectories. A higher mIoU indicates less diversity of the denoised trajectories. The quantitative mode diversity results in Tab. 2 further validate the observations presented in Fig. 2." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 55, + 689, + 296, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 689, + 296, + 714 + ], + "spans": [ + { + "bbox": [ + 55, + 689, + 296, + 714 + ], + "type": "text", + "content": "Heavy denoising overhead. The DDIM [35] diffusion policy requires 20 denoising steps to transform random noise" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 313, + 247, + 555, + 295 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 247, + 555, + 295 + ], + "spans": [ + { + "bbox": [ + 313, + 247, + 555, + 295 + ], + "type": "text", + "content": "into a feasible trajectory, which introduces significant computational overhead, reducing the FPS from 60 to 7, as shown in Tab. 2, and making it impractical for real-time online driving applications." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 313, + 302, + 432, + 313 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 302, + 432, + 313 + ], + "spans": [ + { + "bbox": [ + 313, + 302, + 432, + 313 + ], + "type": "text", + "content": "3.3. Truncated Diffusion" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 313, + 319, + 555, + 426 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 319, + 555, + 426 + ], + "spans": [ + { + "bbox": [ + 313, + 319, + 555, + 426 + ], + "type": "text", + "content": "Human driving follows fixed patterns, unlike the random noise denoising in vanilla diffusion policy. Motivated by this, we propose a truncated diffusion policy that begins the denoising process from an anchored Gaussian distribution instead of a standard Gaussian distribution. To enable the model to learn to denoise from the anchored Gaussian distribution to the desired driving policy, we further truncate the diffusion schedule during training, adding only a small amount of Gaussian noise to the anchors." + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 313, + 429, + 555, + 491 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 429, + 555, + 491 + ], + "spans": [ + { + "bbox": [ + 313, + 429, + 555, + 491 + ], + "type": "text", + "content": "Training. We first construct the diffusion process by adding Gaussian noise to anchors " + }, + { + "bbox": [ + 313, + 429, + 555, + 491 + ], + "type": "inline_equation", + "content": "\\{\\mathbf{a}_k\\}_{k=1}^{N_{\\text{anchor}}}" + }, + { + "bbox": [ + 313, + 429, + 555, + 491 + ], + "type": "text", + "content": " clustered by K-Means on the training set, where " + }, + { + "bbox": [ + 313, + 429, + 555, + 491 + ], + "type": "inline_equation", + "content": "\\mathbf{a}_k = \\{(x_t, y_t)\\}_{t=1}^{T_f}" + }, + { + "bbox": [ + 313, + 429, + 555, + 491 + ], + "type": "text", + "content": ". We truncate the diffusion noise schedule to diffuse the anchors to the anchored Gaussian distribution:" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 346, + 498, + 555, + 513 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 346, + 498, + 555, + 513 + ], + "spans": [ + { + "bbox": [ + 346, + 498, + 555, + 513 + ], + "type": "interline_equation", + "content": "\\tau_ {k} ^ {i} = \\sqrt {\\bar {\\alpha} ^ {i}} \\mathbf {a} _ {k} + \\sqrt {1 - \\bar {\\alpha} ^ {i}} \\boldsymbol {\\epsilon}, \\quad \\boldsymbol {\\epsilon} \\sim \\mathcal {N} (0, \\mathbf {I}), \\qquad (4)", + "image_path": "7f1c61fb17c42bd8fe8575a653a2d30b286022b1f651ef93103b12612235d647.jpg" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 313, + 521, + 554, + 544 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 521, + 554, + 544 + ], + "spans": [ + { + "bbox": [ + 313, + 521, + 554, + 544 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 313, + 521, + 554, + 544 + ], + "type": "inline_equation", + "content": "i \\in [1, T_{\\mathrm{trunc}}]" + }, + { + "bbox": [ + 313, + 521, + 554, + 544 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 313, + 521, + 554, + 544 + ], + "type": "inline_equation", + "content": "T_{\\mathrm{trunc}} \\ll T" + }, + { + "bbox": [ + 313, + 521, + 554, + 544 + ], + "type": "text", + "content": " is the truncated diffusion steps." + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 313, + 545, + 554, + 585 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 545, + 554, + 585 + ], + "spans": [ + { + "bbox": [ + 313, + 545, + 554, + 585 + ], + "type": "text", + "content": "During training, the diffusion decoder " + }, + { + "bbox": [ + 313, + 545, + 554, + 585 + ], + "type": "inline_equation", + "content": "f_{\\theta}" + }, + { + "bbox": [ + 313, + 545, + 554, + 585 + ], + "type": "text", + "content": " takes as input " + }, + { + "bbox": [ + 313, + 545, + 554, + 585 + ], + "type": "inline_equation", + "content": "N_{\\mathrm{anchor}}" + }, + { + "bbox": [ + 313, + 545, + 554, + 585 + ], + "type": "text", + "content": " noisy trajectories " + }, + { + "bbox": [ + 313, + 545, + 554, + 585 + ], + "type": "inline_equation", + "content": "\\{\\tau_k^i\\}_{k=1}^{N_{\\mathrm{anchor}}}" + }, + { + "bbox": [ + 313, + 545, + 554, + 585 + ], + "type": "text", + "content": " and predicts classification scores " + }, + { + "bbox": [ + 313, + 545, + 554, + 585 + ], + "type": "inline_equation", + "content": "\\{\\hat{s}_k\\}_{k=1}^{N_{\\mathrm{anchor}}}" + }, + { + "bbox": [ + 313, + 545, + 554, + 585 + ], + "type": "text", + "content": " and denoised trajectories " + }, + { + "bbox": [ + 313, + 545, + 554, + 585 + ], + "type": "inline_equation", + "content": "\\{\\hat{\\tau}_k\\}_{k=1}^{N_{\\mathrm{anchor}}}" + }, + { + "bbox": [ + 313, + 545, + 554, + 585 + ], + "type": "text", + "content": ":" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 364, + 590, + 555, + 607 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 364, + 590, + 555, + 607 + ], + "spans": [ + { + "bbox": [ + 364, + 590, + 555, + 607 + ], + "type": "interline_equation", + "content": "\\left\\{\\hat {s} _ {k}, \\hat {\\tau} _ {k} \\right\\} _ {k = 1} ^ {N _ {\\text {a n c h o r}}} = f _ {\\theta} \\left(\\left\\{\\tau_ {k} ^ {i} \\right\\} _ {k = 1} ^ {N _ {\\text {a n c h o r}}}, z\\right), \\tag {5}", + "image_path": "19f37a6810aaff15cd2279259ad8f499d94ca71c3bb6ebc31586b477424a07d4.jpg" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 313, + 614, + 555, + 674 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 614, + 555, + 674 + ], + "spans": [ + { + "bbox": [ + 313, + 614, + 555, + 674 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 313, + 614, + 555, + 674 + ], + "type": "inline_equation", + "content": "z" + }, + { + "bbox": [ + 313, + 614, + 555, + 674 + ], + "type": "text", + "content": " represents the conditional information. We assign the noisy trajectory around the closest anchor to the ground truth trajectory " + }, + { + "bbox": [ + 313, + 614, + 555, + 674 + ], + "type": "inline_equation", + "content": "\\tau_{\\mathrm{gt}}" + }, + { + "bbox": [ + 313, + 614, + 555, + 674 + ], + "type": "text", + "content": " as positive sample " + }, + { + "bbox": [ + 313, + 614, + 555, + 674 + ], + "type": "inline_equation", + "content": "(y_k = 1)" + }, + { + "bbox": [ + 313, + 614, + 555, + 674 + ], + "type": "text", + "content": " and others as negative samples " + }, + { + "bbox": [ + 313, + 614, + 555, + 674 + ], + "type": "inline_equation", + "content": "(y_k = 0)" + }, + { + "bbox": [ + 313, + 614, + 555, + 674 + ], + "type": "text", + "content": ". The training objective combines trajectory reconstruction and classification:" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 344, + 683, + 554, + 715 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 344, + 683, + 554, + 715 + ], + "spans": [ + { + "bbox": [ + 344, + 683, + 554, + 715 + ], + "type": "interline_equation", + "content": "\\mathcal {L} = \\sum_ {k = 1} ^ {N _ {\\text {a n c h o r}}} \\left[ y _ {k} \\mathcal {L} _ {\\text {r e c}} \\left(\\hat {\\tau} _ {k}, \\tau_ {\\mathrm {g t}}\\right) + \\lambda \\mathrm {B C E} \\left(\\hat {s} _ {k}, y _ {k}\\right) \\right], \\tag {6}", + "image_path": "da20312d554176b84059f0864ccc99cf62e77952aa1e7d0c7d4803c0fe0d6bc6.jpg" + } + ] + } + ], + "index": 20 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "text", + "content": "12041" + } + ] + } + ], + "index": 21 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 4 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 57, + 70, + 553, + 208 + ], + "blocks": [ + { + "bbox": [ + 57, + 70, + 553, + 208 + ], + "lines": [ + { + "bbox": [ + 57, + 70, + 553, + 208 + ], + "spans": [ + { + "bbox": [ + 57, + 70, + 553, + 208 + ], + "type": "table", + "html": "
MethodInputImg. BackboneAnchorNC↑DAC ↑TTC ↑Comf. ↑EP ↑PDMS ↑
UniAD [16]CameraResNet-34 [13]097.891.992.910078.883.4
PARA-Drive [45]CameraResNet-34 [13]097.992.493.099.879.384.0
LTF [8]CameraResNet-34 [13]097.492.892.410079.083.8
Transfuser [8]C & LResNet-34 [13]097.792.892.810079.284.0
DRAMA [52]C & LResNet-34 [13]098.093.194.810080.185.5
VADv2- \\( \\mathcal{V}_{8192} \\) [4]C & LResNet-34 [13]819297.289.191.610076.080.9
Hydra-MDP- \\( \\mathcal{V}_{8192} \\) [25]C & LResNet-34 [13]819297.991.792.910077.683.0
Hydra-MDP- \\( \\mathcal{V}_{8192} \\)-W-EP [25]C & LResNet-34 [13]819298.396.094.610078.786.5
DiffusionDrive (Ours)C & LResNet-34 [13]2098.296.294.710082.288.1
", + "image_path": "e45249461eb686dfd45a7e9e800e8405ad0142b839ff26eadf7d5daf32e7d80d.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "table_body" + } + ], + "index": 0 + }, + { + "type": "table", + "bbox": [ + 58, + 265, + 549, + 347 + ], + "blocks": [ + { + "bbox": [ + 55, + 209, + 555, + 264 + ], + "lines": [ + { + "bbox": [ + 55, + 209, + 555, + 264 + ], + "spans": [ + { + "bbox": [ + 55, + 209, + 555, + 264 + ], + "type": "text", + "content": "Table 1. Comparison on planning-oriented NAVSIM navtest split with closed-loop metrics. “C & L” denotes the use of both camera and LiDAR as sensor inputs. “" + }, + { + "bbox": [ + 55, + 209, + 555, + 264 + ], + "type": "inline_equation", + "content": "\\mathcal{V}_{8192}" + }, + { + "bbox": [ + 55, + 209, + 555, + 264 + ], + "type": "text", + "content": "” denotes 8192 anchors. “Hydra-MDP-" + }, + { + "bbox": [ + 55, + 209, + 555, + 264 + ], + "type": "inline_equation", + "content": "\\mathcal{V}_{8192}" + }, + { + "bbox": [ + 55, + 209, + 555, + 264 + ], + "type": "text", + "content": "-W-EP” is a variant of Hydra-MDP [25], which is further trained to fit the EP evaluation metric with additional supervision from the rule-based evaluator and uses weighted confidence post-processing. DiffusionDrive simply learns from human demonstrations and infers without post-processing. The best and the second best results are denoted by bold and underline." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 58, + 265, + 549, + 347 + ], + "lines": [ + { + "bbox": [ + 58, + 265, + 549, + 347 + ], + "spans": [ + { + "bbox": [ + 58, + 265, + 549, + 347 + ], + "type": "table", + "html": "
MethodNC↑DAC↑TTC↑Comf.↑EP↑PDMS↑Arch.Plan Module TimeTotal ↓D↑Para.↓FPS↑
Step Time↓Steps ↓
Transfuser97.792.892.810079.284.0MLP0.2ms10.2ms0%56M60
TransfuserDP97.593.792.710079.484.6+0.6UNet6.5ms20130.0ms11%101M7
TransfuserTD97.994.293.910080.285.7+1.7UNet6.9ms213.8ms70%102M27
DiffusionDrive98.296.294.710082.288.1+4.1Dec.3.8ms27.6ms74%60M45
", + "image_path": "90d41fde76a17be1cc5e9e996a65b8c2f21cc99ab9cbe1617e0483aabeced382.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "table_body" + } + ], + "index": 2 + }, + { + "bbox": [ + 55, + 348, + 555, + 384 + ], + "lines": [ + { + "bbox": [ + 55, + 348, + 555, + 384 + ], + "spans": [ + { + "bbox": [ + 55, + 348, + 555, + 384 + ], + "type": "text", + "content": "Table 2. Roadmap from Transfuser to DiffusionDrive on NAVSIM navtest split. \"TransfuserDP\" denotes Transfuser with vanilla DDIM diffusion policy [7]. \"TransfuserTD\" denotes Transfuser with truncated diffusion policy. \"Step Time\" denotes the runtime of each denoising step. \"FPS\" and runtime are measured on an NVIDIA 4090 GPU. \"D\" denotes the mode diversity score defined in Eq. (3)." + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "text" + }, + { + "bbox": [ + 55, + 386, + 295, + 410 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 386, + 295, + 410 + ], + "spans": [ + { + "bbox": [ + 55, + 386, + 295, + 410 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 55, + 386, + 295, + 410 + ], + "type": "inline_equation", + "content": "\\lambda" + }, + { + "bbox": [ + 55, + 386, + 295, + 410 + ], + "type": "text", + "content": " balances the simple L1 reconstruction loss " + }, + { + "bbox": [ + 55, + 386, + 295, + 410 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{\\mathrm{rec}}" + }, + { + "bbox": [ + 55, + 386, + 295, + 410 + ], + "type": "text", + "content": " and binary cross-entropy (BCE) classification loss." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 55, + 411, + 295, + 520 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 411, + 295, + 520 + ], + "spans": [ + { + "bbox": [ + 55, + 411, + 295, + 520 + ], + "type": "text", + "content": "Inference. We use a truncated denoising process that starts with noisy trajectories sampled from the anchored Gaussian distribution and progressively denoises them to final predictions. At each denoising timestep, the estimated trajectories from the previous step are passed to the diffusion decoder " + }, + { + "bbox": [ + 55, + 411, + 295, + 520 + ], + "type": "inline_equation", + "content": "f_{\\theta}" + }, + { + "bbox": [ + 55, + 411, + 295, + 520 + ], + "type": "text", + "content": ", which predicts classification scores " + }, + { + "bbox": [ + 55, + 411, + 295, + 520 + ], + "type": "inline_equation", + "content": "\\{\\hat{s}_k\\}_{k=1}^{N_{\\mathrm{infer}}}" + }, + { + "bbox": [ + 55, + 411, + 295, + 520 + ], + "type": "text", + "content": " and coordinates " + }, + { + "bbox": [ + 55, + 411, + 295, + 520 + ], + "type": "inline_equation", + "content": "\\{\\hat{\\tau}_k\\}_{k=1}^{N_{\\mathrm{infer}}}" + }, + { + "bbox": [ + 55, + 411, + 295, + 520 + ], + "type": "text", + "content": ". After obtaining the current timestep's predictions, we apply the DDIM [35] update rule to sample trajectories for the next timestep." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 55, + 521, + 295, + 594 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 521, + 295, + 594 + ], + "spans": [ + { + "bbox": [ + 55, + 521, + 295, + 594 + ], + "type": "text", + "content": "Inference flexibility. A key advantage of our approach lies in its inference flexibility. While the model is trained with " + }, + { + "bbox": [ + 55, + 521, + 295, + 594 + ], + "type": "inline_equation", + "content": "N_{\\text{anchor}}" + }, + { + "bbox": [ + 55, + 521, + 295, + 594 + ], + "type": "text", + "content": " trajectories, the inference process can accommodate an arbitrary number of trajectory samples " + }, + { + "bbox": [ + 55, + 521, + 295, + 594 + ], + "type": "inline_equation", + "content": "N_{\\text{infer}}" + }, + { + "bbox": [ + 55, + 521, + 295, + 594 + ], + "type": "text", + "content": ", where " + }, + { + "bbox": [ + 55, + 521, + 295, + 594 + ], + "type": "inline_equation", + "content": "N_{\\text{infer}}" + }, + { + "bbox": [ + 55, + 521, + 295, + 594 + ], + "type": "text", + "content": " can be dynamically adjusted based on computational resources or application requirements." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 55, + 599, + 138, + 610 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 599, + 138, + 610 + ], + "spans": [ + { + "bbox": [ + 55, + 599, + 138, + 610 + ], + "type": "text", + "content": "3.4. Architecture" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 55, + 616, + 295, + 700 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 616, + 295, + 700 + ], + "spans": [ + { + "bbox": [ + 55, + 616, + 295, + 700 + ], + "type": "text", + "content": "The overall architecture of our proposed method, DiffusionDrive, is illustrated in Fig. 4. DiffusionDrive can integrate various existing perception modules used in previous end-to-end planners [8, 16, 20, 39] and take different sensor inputs. The designed diffusion decoder is tailored for the complex and challenging driving application, which has enhanced interactions with the conditional scene context." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 55, + 701, + 295, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 701, + 295, + 714 + ], + "spans": [ + { + "bbox": [ + 55, + 701, + 295, + 714 + ], + "type": "text", + "content": "Diffusion decoder. Given the set of sampled noisy trajectories" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 313, + 385, + 555, + 601 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 385, + 555, + 601 + ], + "spans": [ + { + "bbox": [ + 313, + 385, + 555, + 601 + ], + "type": "text", + "content": "ries " + }, + { + "bbox": [ + 313, + 385, + 555, + 601 + ], + "type": "inline_equation", + "content": "\\{\\hat{\\tau}_k\\}_{k=1}^{N_{\\mathrm{infer}}}" + }, + { + "bbox": [ + 313, + 385, + 555, + 601 + ], + "type": "text", + "content": " from the anchored Gaussian distribution, we begin by applying deformable spatial cross-attention [29, 42, 62] to interact with Bird's Eye View (BEV) or Perspective View (PV) features based on the trajectory coordinates. Subsequently, cross-attention is performed between the trajectory features and the agent/map queries derived from the perception module, followed by a feed-forward network (FFN). To encode the diffusion timestep information, we utilize a Timestep Modulation layer, which is followed by a Multi-Layer Perceptron (MLP) that predicts the confidence score and the offset relative to the initial noisy trajectory coordinates. The output from this diffusion decoder layer serves as the input for the subsequent cascade diffusion decoder layer. DiffusionDrive further reuses the cascade diffusion decoder to iteratively denoise the trajectory during inference, with parameters shared across the different denoising timesteps. The final trajectory with the highest confidence score is selected as the output." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 313, + 613, + 391, + 628 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 613, + 391, + 628 + ], + "spans": [ + { + "bbox": [ + 313, + 613, + 391, + 628 + ], + "type": "text", + "content": "4. Experiment" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 313, + 634, + 373, + 645 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 634, + 373, + 645 + ], + "spans": [ + { + "bbox": [ + 313, + 634, + 373, + 645 + ], + "type": "text", + "content": "4.1. Dataset" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 313, + 653, + 555, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 653, + 555, + 715 + ], + "spans": [ + { + "bbox": [ + 313, + 653, + 555, + 715 + ], + "type": "text", + "content": "NAVSIM. The NAVSIM dataset [11] is a real-world planning-oriented dataset builds upon OpenScene [10], a compact redistribution of nuPlan [22], the largest publicly available annotated driving dataset. NAVSIM leverages eight cameras to achieve a full " + }, + { + "bbox": [ + 313, + 653, + 555, + 715 + ], + "type": "inline_equation", + "content": "360^{\\circ}" + }, + { + "bbox": [ + 313, + 653, + 555, + 715 + ], + "type": "text", + "content": " FOV, along with a" + } + ] + } + ], + "index": 13 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 318, + 758 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 318, + 758 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 318, + 758 + ], + "type": "text", + "content": "12042" + } + ] + } + ], + "index": 14 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 5 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 57, + 70, + 553, + 182 + ], + "blocks": [ + { + "bbox": [ + 57, + 70, + 553, + 182 + ], + "lines": [ + { + "bbox": [ + 57, + 70, + 553, + 182 + ], + "spans": [ + { + "bbox": [ + 57, + 70, + 553, + 182 + ], + "type": "table", + "html": "
IDUNet DecoderEgo Query InteractionSpatial Cross-attnAgent/Map Cross-attnCascade DecoderParam.↓NC↑DAC↑Planning MetricPDMS↑
TTC↑Comf.↑EP↑
1✓✓XXX102M97.994.293.910080.285.7
2X✓XXX57M88.783.280.084.843.355.1
3X✓✓XX58M98.295.494.410081.387.1
4X✓X✓X58M97.993.593.810079.885.1
5X✓✓✓X59M98.095.894.410081.787.4
6X✓✓✓✓60M98.296.294.710082.288.1
", + "image_path": "5d60719ec93c39d85aaf432bb11752b22d4eff87f1347e8e26a027665583c888.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "table_body" + } + ], + "index": 0 + }, + { + "type": "table", + "bbox": [ + 57, + 220, + 218, + 266 + ], + "blocks": [ + { + "bbox": [ + 55, + 185, + 553, + 218 + ], + "lines": [ + { + "bbox": [ + 55, + 185, + 553, + 218 + ], + "spans": [ + { + "bbox": [ + 55, + 185, + 553, + 218 + ], + "type": "text", + "content": "Table 3. Ablation for design choices. \"Cascade Decoder\" indicates that we stack 2 cascade diffusion decoder layers. ID-1 refers to TransfuserTD in Tab. 2, utilizing conditional UNet and interaction with the ego-query, which Transfuser uses to directly regress the single-mode trajectory." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 57, + 220, + 218, + 266 + ], + "lines": [ + { + "bbox": [ + 57, + 220, + 218, + 266 + ], + "spans": [ + { + "bbox": [ + 57, + 220, + 218, + 266 + ], + "type": "table", + "html": "
StepsParam.NCDACTTCComf.EPPDMS
160M98.396.094.710082.187.9
260M98.296.294.710082.288.1
360M98.296.394.710092.288.1
", + "image_path": "8d14f9f0b1907f2e1c1f24a3c9ffc59f48c791507d86b6b15ede4856c841428c.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "table_body" + } + ], + "index": 2 + }, + { + "type": "table", + "bbox": [ + 225, + 220, + 388, + 266 + ], + "blocks": [ + { + "bbox": [ + 75, + 268, + 198, + 278 + ], + "lines": [ + { + "bbox": [ + 75, + 268, + 198, + 278 + ], + "spans": [ + { + "bbox": [ + 75, + 268, + 198, + 278 + ], + "type": "text", + "content": "Table 4. Denoising step number." + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 225, + 220, + 388, + 266 + ], + "lines": [ + { + "bbox": [ + 225, + 220, + 388, + 266 + ], + "spans": [ + { + "bbox": [ + 225, + 220, + 388, + 266 + ], + "type": "table", + "html": "
StagesParam.NCDACTTCComf.EPPDMS
159M98.095.894.410081.787.4
260M98.296.294.710082.288.1
465M98.496.294.910082.488.2
", + "image_path": "dfbc91c12c6096eb2909c06a78669315159466720a4e4cce73d6cba83ba5ad34.jpg" + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "table_body" + } + ], + "index": 4 + }, + { + "type": "table", + "bbox": [ + 392, + 220, + 551, + 266 + ], + "blocks": [ + { + "bbox": [ + 257, + 268, + 350, + 278 + ], + "lines": [ + { + "bbox": [ + 257, + 268, + 350, + 278 + ], + "spans": [ + { + "bbox": [ + 257, + 268, + 350, + 278 + ], + "type": "text", + "content": "Table 5. Cascade stages." + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 392, + 220, + 551, + 266 + ], + "lines": [ + { + "bbox": [ + 392, + 220, + 551, + 266 + ], + "spans": [ + { + "bbox": [ + 392, + 220, + 551, + 266 + ], + "type": "table", + "html": "
N_inferParam.NCDACTTCComf.EPPDMS
1060M97.993.593.110080.084.9
2060M98.296.294.710082.288.1
4060M98.596.294.810082.588.2
", + "image_path": "f061fbfa1ae90ca4196cf7add05b0aa11c15dc6dfc79c2924f6e7a47949558fb.jpg" + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "table_body" + } + ], + "index": 6 + }, + { + "bbox": [ + 391, + 268, + 550, + 278 + ], + "lines": [ + { + "bbox": [ + 391, + 268, + 550, + 278 + ], + "spans": [ + { + "bbox": [ + 391, + 268, + 550, + 278 + ], + "type": "text", + "content": "Table 6. Number of sampled noises " + }, + { + "bbox": [ + 391, + 268, + 550, + 278 + ], + "type": "inline_equation", + "content": "{N}_{\\text{infer }}" + }, + { + "bbox": [ + 391, + 268, + 550, + 278 + ], + "type": "text", + "content": " ." + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "text" + }, + { + "bbox": [ + 54, + 288, + 295, + 372 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 54, + 288, + 295, + 372 + ], + "spans": [ + { + "bbox": [ + 54, + 288, + 295, + 372 + ], + "type": "text", + "content": "merged LiDAR point cloud derived from five sensors. Annotations are provided at a frequency of " + }, + { + "bbox": [ + 54, + 288, + 295, + 372 + ], + "type": "inline_equation", + "content": "2\\mathrm{Hz}" + }, + { + "bbox": [ + 54, + 288, + 295, + 372 + ], + "type": "text", + "content": " and include both HD maps and object bounding boxes. The dataset is designed to emphasize challenging driving scenarios involving dynamic changes in driving intentions, while deliberately excluding trivial situations such as stationary scenes or constant-speed driving." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 54, + 373, + 295, + 456 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 54, + 373, + 295, + 456 + ], + "spans": [ + { + "bbox": [ + 54, + 373, + 295, + 456 + ], + "type": "text", + "content": "NAVSIM benchmarks planning performance using non-reactive simulations and closed-loop metrics for comprehensive evaluation. In this paper, we employ the proposed PDM score (PDMS) [11], which is a weighted combination of several sub-scores: no at-fault collisions (NC), drivable area compliance (DAC), time-to-collision (TTC), comfort (Comf.), and ego progress (EP)." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 55, + 468, + 185, + 480 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 468, + 185, + 480 + ], + "spans": [ + { + "bbox": [ + 55, + 468, + 185, + 480 + ], + "type": "text", + "content": "4.2. Implementation Detail" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 56, + 486, + 295, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 486, + 295, + 713 + ], + "spans": [ + { + "bbox": [ + 56, + 486, + 295, + 713 + ], + "type": "text", + "content": "We adopt the same perception modules and ResNet-34 backbone [13] as Transfuser for fair comparison. In the diffusion decoder layer, we employ spatial cross-attention to only interact with BEV features following Transfuser's BEV-based setting. We only perform agent cross-attention, since the perception module of Transfuser does not include vectorized map construction. We stack 2 cascade diffusion decoder layers and apply truncated diffusion policy with 20 clustered anchors. The training diffusion schedule is truncated by 50/1000 to diffuse the anchors, while during inference, we use only 2 denoising steps and select the top-1 scoring predicted trajectory for evaluation. The training and inference recipe directly follows Transfuser: We use three cropped and downscaled forward-facing camera images, concatenated as a " + }, + { + "bbox": [ + 56, + 486, + 295, + 713 + ], + "type": "inline_equation", + "content": "1024 \\times 256" + }, + { + "bbox": [ + 56, + 486, + 295, + 713 + ], + "type": "text", + "content": " image, and a rasterized BEV LiDAR as input; DiffusionDrive is trained on navtrain split from scratch for 100 epochs with AdamW optimizer on 8 NVIDIA 4090 GPUs with total batch size of 512, setting the learning rate to " + }, + { + "bbox": [ + 56, + 486, + 295, + 713 + ], + "type": "inline_equation", + "content": "6 \\times 10^{-4}" + }, + { + "bbox": [ + 56, + 486, + 295, + 713 + ], + "type": "text", + "content": ". No test-time" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 313, + 289, + 553, + 312 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 289, + 553, + 312 + ], + "spans": [ + { + "bbox": [ + 313, + 289, + 553, + 312 + ], + "type": "text", + "content": "augmentation is applied and the final output for evaluation on navtest split is 8-waypoint trajectory over 4 seconds." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 313, + 322, + 456, + 335 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 322, + 456, + 335 + ], + "spans": [ + { + "bbox": [ + 313, + 322, + 456, + 335 + ], + "type": "text", + "content": "4.3. Quantitative Comparison" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 313, + 339, + 555, + 579 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 339, + 555, + 579 + ], + "spans": [ + { + "bbox": [ + 313, + 339, + 555, + 579 + ], + "type": "text", + "content": "Tab. 1 compares DiffusionDrive with state-of-the-art methods on NAVSIM navtest split. With the same ResNet-34 backbone, DiffusionDrive achieves 88.1 PDMS score, demonstrating significant superior performance over the previous learning-based methods. Compared to VADv2, DiffusionDrive surpasses it by 7.2 PDMS while reducing the number of anchors from 8192 to 20, representing a " + }, + { + "bbox": [ + 313, + 339, + 555, + 579 + ], + "type": "inline_equation", + "content": "400 \\times" + }, + { + "bbox": [ + 313, + 339, + 555, + 579 + ], + "type": "text", + "content": " reduction. DiffusionDrive also outperforms Hydra-MDP, which follows VADv2's sampling-from-vocabulary paradigm, with a 5.1 PDMS improvement. Even compared to the Hydra-MDP-" + }, + { + "bbox": [ + 313, + 339, + 555, + 579 + ], + "type": "inline_equation", + "content": "\\mathcal{V}_{8192}" + }, + { + "bbox": [ + 313, + 339, + 555, + 579 + ], + "type": "text", + "content": "-W-EP, which is a variant of Hydra-MDP [25] by further training to fit the EP evaluation metric with additional supervision and using weighted confidence post-processing, DiffusionDrive still outperforms it by 3.5 EP and 1.6 overall PDMS, relying solely on a straightforward learning-from-human approach without any post-processing. Compared to the Transfuser baseline, where we only differ in the planning module, DiffusionDrive delivers a notable 4.1 PDMS improvement, outperforming it across all sub-scores." + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 313, + 588, + 383, + 601 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 588, + 383, + 601 + ], + "spans": [ + { + "bbox": [ + 313, + 588, + 383, + 601 + ], + "type": "text", + "content": "4.4. Roadmap" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 313, + 605, + 555, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 605, + 555, + 713 + ], + "spans": [ + { + "bbox": [ + 313, + 605, + 555, + 713 + ], + "type": "text", + "content": "In Tab. 2, converting Transfuser " + }, + { + "bbox": [ + 313, + 605, + 555, + 713 + ], + "type": "inline_equation", + "content": "\\mathrm{DP}" + }, + { + "bbox": [ + 313, + 605, + 555, + 713 + ], + "type": "text", + "content": " using vanilla diffusion policy improves the PDMS score by 0.6 and the mode diversity score " + }, + { + "bbox": [ + 313, + 605, + 555, + 713 + ], + "type": "inline_equation", + "content": "\\mathcal{D}" + }, + { + "bbox": [ + 313, + 605, + 555, + 713 + ], + "type": "text", + "content": " by " + }, + { + "bbox": [ + 313, + 605, + 555, + 713 + ], + "type": "inline_equation", + "content": "11\\%" + }, + { + "bbox": [ + 313, + 605, + 555, + 713 + ], + "type": "text", + "content": ". However, it also significantly increases the overhead of the planning module, requiring " + }, + { + "bbox": [ + 313, + 605, + 555, + 713 + ], + "type": "inline_equation", + "content": "20\\times" + }, + { + "bbox": [ + 313, + 605, + 555, + 713 + ], + "type": "text", + "content": " more denoising steps and " + }, + { + "bbox": [ + 313, + 605, + 555, + 713 + ], + "type": "inline_equation", + "content": "32\\times" + }, + { + "bbox": [ + 313, + 605, + 555, + 713 + ], + "type": "text", + "content": " the step time, resulting in a total " + }, + { + "bbox": [ + 313, + 605, + 555, + 713 + ], + "type": "inline_equation", + "content": "650\\times" + }, + { + "bbox": [ + 313, + 605, + 555, + 713 + ], + "type": "text", + "content": " increase in runtime overhead. With the proposed truncated diffusion policy, " + }, + { + "bbox": [ + 313, + 605, + 555, + 713 + ], + "type": "inline_equation", + "content": "\\mathrm{Transfuser}_{\\mathrm{TD}}" + }, + { + "bbox": [ + 313, + 605, + 555, + 713 + ], + "type": "text", + "content": " reduces the number of denoising steps from 20 to 2 while improving PDMS by 1.1 and mode diver" + } + ] + } + ], + "index": 16 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "text", + "content": "12043" + } + ] + } + ], + "index": 17 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 6 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 56, + 70, + 555, + 184 + ], + "blocks": [ + { + "bbox": [ + 56, + 70, + 555, + 184 + ], + "lines": [ + { + "bbox": [ + 56, + 70, + 555, + 184 + ], + "spans": [ + { + "bbox": [ + 56, + 70, + 555, + 184 + ], + "type": "table", + "html": "
MethodInputImg. BackboneL2 (m) ↓Collision Rate (%) ↓FPS ↑
1s2s3sAvg.1s2s3sAvg.
ST-P3 [15]CameraEffNet-b4 [40]1.332.112.902.110.230.621.270.711.6
UniAD [16]CameraResNet-101 [13]0.450.701.040.730.620.580.630.611.8
OccNet [41]CameraResNet-50 [13]1.292.132.992.140.210.591.370.722.6
VAD [20]CameraResNet-50 [13]0.410.701.050.720.070.170.410.224.5
SparseDrive [39]CameraResNet-50 [13]0.290.580.960.610.010.050.180.089.0
DiffusionDrive (Ours)CameraResNet-50 [13]0.270.540.900.570.030.050.160.088.2
", + "image_path": "fcc3e04c5f71e323a22828beec3a3aea956064444a32df850c3ee01e1054e61d.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "table_body" + } + ], + "index": 0 + }, + { + "bbox": [ + 56, + 185, + 555, + 208 + ], + "lines": [ + { + "bbox": [ + 56, + 185, + 555, + 208 + ], + "spans": [ + { + "bbox": [ + 56, + 185, + 555, + 208 + ], + "type": "text", + "content": "Table 7. Comparison on nuScenes dataset with open-loop metrics. FPS is measured on a single NVIDIA 4090 GPU following the recipe of SparseDrive [39]. Metric calculation follows ST-P3 [15]." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "text" + }, + { + "bbox": [ + 55, + 211, + 294, + 297 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 211, + 294, + 297 + ], + "spans": [ + { + "bbox": [ + 55, + 211, + 294, + 297 + ], + "type": "text", + "content": "sity by " + }, + { + "bbox": [ + 55, + 211, + 294, + 297 + ], + "type": "inline_equation", + "content": "59\\%" + }, + { + "bbox": [ + 55, + 211, + 294, + 297 + ], + "type": "text", + "content": ". By further incorporating the proposed diffusion decoder, the final model, DiffusionDrive, reaches 88.1 PDMS and " + }, + { + "bbox": [ + 55, + 211, + 294, + 297 + ], + "type": "inline_equation", + "content": "74\\%" + }, + { + "bbox": [ + 55, + 211, + 294, + 297 + ], + "type": "text", + "content": " mode diversity score " + }, + { + "bbox": [ + 55, + 211, + 294, + 297 + ], + "type": "inline_equation", + "content": "\\mathcal{D}" + }, + { + "bbox": [ + 55, + 211, + 294, + 297 + ], + "type": "text", + "content": ". Compared to the TransfuserDP, DiffusionDrive shows improvements in 3.5 PDMS and " + }, + { + "bbox": [ + 55, + 211, + 294, + 297 + ], + "type": "inline_equation", + "content": "64\\%" + }, + { + "bbox": [ + 55, + 211, + 294, + 297 + ], + "type": "text", + "content": " mode diversity, and a " + }, + { + "bbox": [ + 55, + 211, + 294, + 297 + ], + "type": "inline_equation", + "content": "10\\times" + }, + { + "bbox": [ + 55, + 211, + 294, + 297 + ], + "type": "text", + "content": " reduction in denoising steps, resulting in a " + }, + { + "bbox": [ + 55, + 211, + 294, + 297 + ], + "type": "inline_equation", + "content": "6\\times" + }, + { + "bbox": [ + 55, + 211, + 294, + 297 + ], + "type": "text", + "content": " speedup in FPS. This enables real-time, high-quality, multi-mode planning." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 55, + 304, + 149, + 317 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 304, + 149, + 317 + ], + "spans": [ + { + "bbox": [ + 55, + 304, + 149, + 317 + ], + "type": "text", + "content": "4.5. Ablation Study" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 55, + 323, + 295, + 455 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 323, + 295, + 455 + ], + "spans": [ + { + "bbox": [ + 55, + 323, + 295, + 455 + ], + "type": "text", + "content": "Effect of designs in diffusion decoder. Tab. 3 shows the effectiveness of our design choices in the diffusion decoder. ID-1 is the Transfuser" + }, + { + "bbox": [ + 55, + 323, + 295, + 455 + ], + "type": "inline_equation", + "content": "_{\\text{TD}}" + }, + { + "bbox": [ + 55, + 323, + 295, + 455 + ], + "type": "text", + "content": " in the Tab. 2. By comapring ID-6 and ID-1, we can see that the proposed diffusion decoder reduce the " + }, + { + "bbox": [ + 55, + 323, + 295, + 455 + ], + "type": "inline_equation", + "content": "39\\%" + }, + { + "bbox": [ + 55, + 323, + 295, + 455 + ], + "type": "text", + "content": " parameters and significantly improves the planning quality by 2.4 PDMS. ID-2 shows severe performance degeneration due to the lack of rich and hierarchical interaction with the environment. By comparing ID-2 and ID-3, we show that spatial cross-attention is vital for accurate planning. ID-5 shows that the proposed cascade mechanism is effective and can further improve the performance." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 55, + 456, + 295, + 516 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 456, + 295, + 516 + ], + "spans": [ + { + "bbox": [ + 55, + 456, + 295, + 516 + ], + "type": "text", + "content": "Denoising step number. Tab. 4 shows that due to a reasonable start point, DiffusionDrive achieves good planning quality even with only 1 step, and further denoising steps offer quality improvement and inference flexibility for complex environments." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 55, + 517, + 295, + 567 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 517, + 295, + 567 + ], + "spans": [ + { + "bbox": [ + 55, + 517, + 295, + 567 + ], + "type": "text", + "content": "Cascade stages. Tab. 5 ablates the impact of cascade stage number. Increasing the stage number can improve the planning quality but saturate at the 4 stages and cost more parameters and inference time at each step." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 55, + 567, + 295, + 652 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 567, + 295, + 652 + ], + "spans": [ + { + "bbox": [ + 55, + 567, + 295, + 652 + ], + "type": "text", + "content": "Number of sampled noises " + }, + { + "bbox": [ + 55, + 567, + 295, + 652 + ], + "type": "inline_equation", + "content": "N_{\\mathrm{infer}}" + }, + { + "bbox": [ + 55, + 567, + 295, + 652 + ], + "type": "text", + "content": ". As stated in Sec. 3.3, DiffusionDrive can generate varied trajectories by simply sampling a variable number of noises from anchored Gaussian distribution. Tab. 6 shows that 10 sampled noises can already achieve a decent planning quality. By sampling more noises, DiffusionDrive can cover potential planning action space and lead to improved planning quality." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 55, + 659, + 191, + 673 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 659, + 191, + 673 + ], + "spans": [ + { + "bbox": [ + 55, + 659, + 191, + 673 + ], + "type": "text", + "content": "4.6. Qualitative Comparison" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 55, + 677, + 295, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 677, + 295, + 714 + ], + "spans": [ + { + "bbox": [ + 55, + 677, + 295, + 714 + ], + "type": "text", + "content": "Since the PDMS planning metric calculates based on the top-1 scoring trajectory and our proposed " + }, + { + "bbox": [ + 55, + 677, + 295, + 714 + ], + "type": "inline_equation", + "content": "\\mathcal{D}" + }, + { + "bbox": [ + 55, + 677, + 295, + 714 + ], + "type": "text", + "content": " score evaluates mode diversity, these metrics alone cannot fully capture the" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 313, + 211, + 555, + 379 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 211, + 555, + 379 + ], + "spans": [ + { + "bbox": [ + 313, + 211, + 555, + 379 + ], + "type": "text", + "content": "quality of diverse trajectories. To further validate the quality of multi-mode trajectories, we visualize the planning results of Transfuser, " + }, + { + "bbox": [ + 313, + 211, + 555, + 379 + ], + "type": "inline_equation", + "content": "\\mathrm{Transfuser}_{\\mathrm{DP}}" + }, + { + "bbox": [ + 313, + 211, + 555, + 379 + ], + "type": "text", + "content": " and DiffusionDrive on challenging scenarios of NAVSIM navtest split in Fig. 2. The results indicate that the multi-mode trajectories generated by DiffusionDrive are not only diverse but also of high quality. In Fig. 2a, the top-1 scoring trajectory generated by DiffusionDrive closely resembles the ground-truth trajectory, while the highlighted top-10 scoring trajectory surprisingly tries to perform high-quality lane changing. In Fig. 2b, the highlighted top-10 scoring trajectory also performs a lane change, and a neighboring low-scoring trajectory further interacts with surrounding agents to effectively avoid collisions." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 313, + 387, + 553, + 400 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 387, + 553, + 400 + ], + "spans": [ + { + "bbox": [ + 313, + 387, + 553, + 400 + ], + "type": "text", + "content": "4.7. Quantitative Comparison on nuScenes dataset" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 313, + 405, + 554, + 501 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 405, + 554, + 501 + ], + "spans": [ + { + "bbox": [ + 313, + 405, + 554, + 501 + ], + "type": "text", + "content": "The nuScenes dataset is previously popular benchmark for end-to-end planning. Since the major scenarios of nuScenes are simple and trivial situations, we only perform comparison in Tab. 7. We implement DiffusionDrive on top of SparseDrive [39] following its training and inference recipe using open-loop metrics proposed in ST-P3 [15]. We stack 2 cascade diffusion decoder layers and apply the truncated diffusion policy with 18 clustered anchors." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 313, + 502, + 554, + 561 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 502, + 554, + 561 + ], + "spans": [ + { + "bbox": [ + 313, + 502, + 554, + 561 + ], + "type": "text", + "content": "As shown in Tab. 7, DiffusionDrive reduces the average L2 error of SparseDrive by " + }, + { + "bbox": [ + 313, + 502, + 554, + 561 + ], + "type": "inline_equation", + "content": "0.04\\mathrm{m}" + }, + { + "bbox": [ + 313, + 502, + 554, + 561 + ], + "type": "text", + "content": ", achieving the lowest L2 error and average collision rate. While DiffusionDrive is also efficient and runs " + }, + { + "bbox": [ + 313, + 502, + 554, + 561 + ], + "type": "inline_equation", + "content": "1.8\\times" + }, + { + "bbox": [ + 313, + 502, + 554, + 561 + ], + "type": "text", + "content": " faster than VAD with " + }, + { + "bbox": [ + 313, + 502, + 554, + 561 + ], + "type": "inline_equation", + "content": "20.8\\%" + }, + { + "bbox": [ + 313, + 502, + 554, + 561 + ], + "type": "text", + "content": " lower L2 error and " + }, + { + "bbox": [ + 313, + 502, + 554, + 561 + ], + "type": "inline_equation", + "content": "63.6\\%" + }, + { + "bbox": [ + 313, + 502, + 554, + 561 + ], + "type": "text", + "content": " lower collision rate." + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 313, + 573, + 388, + 586 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 573, + 388, + 586 + ], + "spans": [ + { + "bbox": [ + 313, + 573, + 388, + 586 + ], + "type": "text", + "content": "5. Conclusion" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 313, + 594, + 555, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 594, + 555, + 715 + ], + "spans": [ + { + "bbox": [ + 313, + 594, + 555, + 715 + ], + "type": "text", + "content": "In this work, we propose a novel generative driving decision-making model, DiffusionDrive, for end-to-end autonomous driving by incorporating the proposed truncated diffusion policy and efficient cascade diffusion decoder. DiffusionDrive can denoise a variable number of samples from an anchored Gaussian distribution to generate diverse planning trajectories at real-time speeds. Comprehensive experiments and qualitative comparisons validate the superiority of DiffusionDrive in planning quality, running efficiency, and mode diversity." + } + ] + } + ], + "index": 15 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 318, + 758 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 318, + 758 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 318, + 758 + ], + "type": "text", + "content": "12044" + } + ] + } + ], + "index": 16 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 7 + }, + { + "para_blocks": [ + { + "bbox": [ + 56, + 72, + 153, + 85 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 72, + 153, + 85 + ], + "spans": [ + { + "bbox": [ + 56, + 72, + 153, + 85 + ], + "type": "text", + "content": "Acknowledgement" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 55, + 91, + 296, + 137 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 91, + 296, + 137 + ], + "spans": [ + { + "bbox": [ + 55, + 91, + 296, + 137 + ], + "type": "text", + "content": "This work was partially supported by the National Natural Science Foundation of China (NSFC) under Grant No. 62276108. We would like to acknowledge Tianheng Cheng for helpful feedback on the draft." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 56, + 160, + 115, + 172 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 160, + 115, + 172 + ], + "spans": [ + { + "bbox": [ + 56, + 160, + 115, + 172 + ], + "type": "text", + "content": "References" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 57, + 179, + 296, + 712 + ], + "type": "list", + "angle": 0, + "index": 14, + "blocks": [ + { + "bbox": [ + 61, + 179, + 296, + 223 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 61, + 179, + 296, + 223 + ], + "spans": [ + { + "bbox": [ + 61, + 179, + 296, + 223 + ], + "type": "text", + "content": "[1] Anurag Ajay, Yilun Du, Abhi Gupta, Joshua B. Tenenbaum, Tommi S. Jaakkola, and Pulkit Agrawal. Is conditional generative modeling all you need for decision making? In ICLR, 2023. 3" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 61, + 225, + 296, + 270 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 61, + 225, + 296, + 270 + ], + "spans": [ + { + "bbox": [ + 61, + 225, + 296, + 270 + ], + "type": "text", + "content": "[2] Holger Caesar, Varun Bankiti, Alex H Lang, Sourabh Vora, Venice Erin Liong, Qiang Xu, Anush Krishnan, Yu Pan, Giancarlo Baldan, and Oscar Beijbom. nuscenes: A multimodal dataset for autonomous driving. In CVPR, 2020. 2" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 61, + 271, + 295, + 314 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 61, + 271, + 295, + 314 + ], + "spans": [ + { + "bbox": [ + 61, + 271, + 295, + 314 + ], + "type": "text", + "content": "[3] Shaoyu Chen, Tianheng Cheng, Xinggang Wang, Wenming Meng, Qian Zhang, and Wenyu Liu. Efficient and robust 2d-to-bev representation learning via geometry-guided kernel transformer. CoRR, abs/2206.04584, 2022. 2" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 61, + 317, + 296, + 371 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 61, + 317, + 296, + 371 + ], + "spans": [ + { + "bbox": [ + 61, + 317, + 296, + 371 + ], + "type": "text", + "content": "[4] Shaoyu Chen, Bo Jiang, Hao Gao, Bencheng Liao, Qing Xu, Qian Zhang, Chang Huang, Wenyu Liu, and Xinggang Wang. Vadv2: End-to-end vectorized autonomous driving via probabilistic planning. arXiv preprint arXiv:2402.13243, 2024. 1, 2, 6" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 62, + 373, + 296, + 418 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 373, + 296, + 418 + ], + "spans": [ + { + "bbox": [ + 62, + 373, + 296, + 418 + ], + "type": "text", + "content": "[5] Shaoyu Chen, Xinggang Wang, Tianheng Cheng, Qian Zhang, Chang Huang, and Wenyu Liu. Polardetr: Polar parametrization for vision-based surround-view 3d detection. Image and Vision Computing, 156:105438, 2025. 1" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 62, + 419, + 296, + 462 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 419, + 296, + 462 + ], + "spans": [ + { + "bbox": [ + 62, + 419, + 296, + 462 + ], + "type": "text", + "content": "[6] Zhili Chen, Maosheng Ye, Shuangjie Xu, Tongyi Cao, and Qifeng Chen. Ppad: Iterative interactions of prediction and planning for end-to-end autonomous driving. In ECCV, 2024. 2" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 62, + 465, + 296, + 508 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 465, + 296, + 508 + ], + "spans": [ + { + "bbox": [ + 62, + 465, + 296, + 508 + ], + "type": "text", + "content": "[7] Cheng Chi, Zhenjia Xu, Siyuan Feng, Eric Cousineau, Yilun Du, Benjamin Burchfiel, Russ Tedrake, and Shuran Song. Diffusion policy: Visuomotor policy learning via action diffusion. In RSS, 2023. 1, 3, 5, 6" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 62, + 510, + 296, + 554 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 510, + 296, + 554 + ], + "spans": [ + { + "bbox": [ + 62, + 510, + 296, + 554 + ], + "type": "text", + "content": "[8] Kashyap Chitta, Aditya Prakash, Bernhard Jaeger, Zehao Yu, Katrin Renz, and Andreas Geiger. Transfuser: Imitation with transformer-based sensor fusion for autonomous driving. TPAMI, 2022. 1, 2, 5, 6" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 62, + 556, + 296, + 599 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 556, + 296, + 599 + ], + "spans": [ + { + "bbox": [ + 62, + 556, + 296, + 599 + ], + "type": "text", + "content": "[9] Younwoo Choi, Ray Coden Mercurius, Soheil Mohamad Al-izadeh Shabestary, and Amir Rasouli. Dice: Diverse diffusion model with scoring for trajectory prediction. In IV, 2024. 3" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 57, + 601, + 295, + 645 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 601, + 295, + 645 + ], + "spans": [ + { + "bbox": [ + 57, + 601, + 295, + 645 + ], + "type": "text", + "content": "[10] OpenScene Contributors. Openscene: The largest up-to-date 3d occupancy prediction benchmark in autonomous driving. https://github.com/OpenDriveLab/OpenScene, 2023.6" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 57, + 647, + 296, + 712 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 647, + 296, + 712 + ], + "spans": [ + { + "bbox": [ + 57, + 647, + 296, + 712 + ], + "type": "text", + "content": "[11] Daniel Dauner, Marcel Hallgarten, Tianyu Li, Xinshuo Weng, Zhiyu Huang, Zetong Yang, Hongyang Li, Igor Gilitschenski, Boris Ivanovic, Marco Pavone, Andreas Geiger, and Kashyap Chitta. Navsim: Data-driven non-reactive autonomous vehicle simulation and benchmarking. In NeurIPS, 2024. 2, 6, 7" + } + ] + } + ], + "index": 13 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 316, + 73, + 553, + 713 + ], + "type": "list", + "angle": 0, + "index": 30, + "blocks": [ + { + "bbox": [ + 316, + 73, + 553, + 106 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 73, + 553, + 106 + ], + "spans": [ + { + "bbox": [ + 316, + 73, + 553, + 106 + ], + "type": "text", + "content": "[12] Xunjiang Gu, Guanyu Song, Igor Gilitschenski, Marco Pavone, and Boris Ivanovic. Producing and leveraging online map uncertainty in trajectory prediction. In CVPR, 2024. 2" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 317, + 107, + 553, + 138 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 107, + 553, + 138 + ], + "spans": [ + { + "bbox": [ + 317, + 107, + 553, + 138 + ], + "type": "text", + "content": "[13] Kaiming He, Xiangyu Zhang, Shaoqing Ren, and Jian Sun. Deep residual learning for image recognition. In CVPR, 2016. 6, 7, 8" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 316, + 140, + 553, + 172 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 140, + 553, + 172 + ], + "spans": [ + { + "bbox": [ + 316, + 140, + 553, + 172 + ], + "type": "text", + "content": "[14] Jonathan Ho, Ajay Jain, and Pieter Abbeel. Denoising diffusion probabilistic models. Advances in neural information processing systems, 33:6840-6851, 2020. 4" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 317, + 173, + 553, + 216 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 173, + 553, + 216 + ], + "spans": [ + { + "bbox": [ + 317, + 173, + 553, + 216 + ], + "type": "text", + "content": "[15] Shengchao Hu, Li Chen, Penghao Wu, Hongyang Li, Junchi Yan, and Dacheng Tao. St-p3: End-to-end vision-based autonomous driving via spatial-temporal feature learning. In ECCV, 2022. 8" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 316, + 217, + 553, + 259 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 217, + 553, + 259 + ], + "spans": [ + { + "bbox": [ + 316, + 217, + 553, + 259 + ], + "type": "text", + "content": "[16] Yihan Hu, Jiazhi Yang, Li Chen, Keyu Li, Chonghao Sima, Xizhou Zhu, Siqi Chai, Senyao Du, Tianwei Lin, Wenhai Wang, et al. Planning-oriented autonomous driving. In CVPR, 2023. 1, 2, 6, 8" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 316, + 262, + 553, + 304 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 262, + 553, + 304 + ], + "spans": [ + { + "bbox": [ + 316, + 262, + 553, + 304 + ], + "type": "text", + "content": "[17] Junjie Huang, Guan Huang, Zheng Zhu, Yun Ye, and Dalong Du. Bevdet: High-performance multi-camera 3d object detection in bird-eye-view. arXiv preprint arXiv:2112.11790, 2021. 1" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 316, + 305, + 553, + 349 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 305, + 553, + 349 + ], + "spans": [ + { + "bbox": [ + 316, + 305, + 553, + 349 + ], + "type": "text", + "content": "[18] Zhiyu Huang, Zixu Zhang, Ameya Vaidya, Yuxiao Chen, Chen Lv, and Jaime Fernandez Fisac. Versatile scene-consistent traffic scenario generation as optimization with diffusion. arXiv preprint arXiv:2404.02524, 2024. 3" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 316, + 350, + 553, + 381 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 350, + 553, + 381 + ], + "spans": [ + { + "bbox": [ + 316, + 350, + 553, + 381 + ], + "type": "text", + "content": "[19] Michael Janner, Yilun Du, Joshua B. Tenenbaum, and Sergey Levine. Planning with diffusion for flexible behavior synthesis. In ICLR, 2022. 1, 3" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 316, + 383, + 553, + 426 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 383, + 553, + 426 + ], + "spans": [ + { + "bbox": [ + 316, + 383, + 553, + 426 + ], + "type": "text", + "content": "[20] Bo Jiang, Shaoyu Chen, Qing Xu, Bencheng Liao, Jiajie Chen, Helong Zhou, Qian Zhang, Wenyu Liu, Chang Huang, and Xinggang Wang. Vad: Vectorized scene representation for efficient autonomous driving. In ICCV, 2023. 1, 2, 6, 8" + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 316, + 427, + 553, + 470 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 427, + 553, + 470 + ], + "spans": [ + { + "bbox": [ + 316, + 427, + 553, + 470 + ], + "type": "text", + "content": "[21] Chiyu Jiang, Andre Cornman, Cheolho Park, Benjamin Sapp, Yin Zhou, Dragomir Anguelov, et al. Motiondiffuser: Controllable multi-agent motion prediction using diffusion. In CVPR, 2023. 3" + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 317, + 472, + 553, + 536 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 472, + 553, + 536 + ], + "spans": [ + { + "bbox": [ + 317, + 472, + 553, + 536 + ], + "type": "text", + "content": "[22] Napat Karnchanachari, Dimitris Geromichalos, Kok Seang Tan, Nanxiang Li, Christopher Eriksen, Shakiba Yaghoubi, Noushin Mehdipour, Gianmarco Bernasconi, Whye Kit Fong, Yiluan Guo, et al. Towards learning-based planning: The nuplan benchmark for real-world autonomous driving. In ICRA, 2024. 6" + } + ] + } + ], + "index": 25 + }, + { + "bbox": [ + 316, + 537, + 553, + 581 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 537, + 553, + 581 + ], + "spans": [ + { + "bbox": [ + 316, + 537, + 553, + 581 + ], + "type": "text", + "content": "[23] Yingyan Li, Lue Fan, Jiawei He, Yuqi Wang, Yuntao Chen, Zhaoxiang Zhang, and Tieniu Tan. Enhancing end-to-end autonomous driving with latent world model. arXiv preprint arXiv:2406.08481, 2024. 2" + } + ] + } + ], + "index": 26 + }, + { + "bbox": [ + 316, + 582, + 553, + 624 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 582, + 553, + 624 + ], + "spans": [ + { + "bbox": [ + 316, + 582, + 553, + 624 + ], + "type": "text", + "content": "[24] Zhiqi Li, Wenhai Wang, Hongyang Li, Enze Xie, Chonghao Sima, Tong Lu, Yu Qiao, and Jifeng Dai. Bevformer: Learning bird's-eye-view representation from multi-camera images via spatiotemporal transformers. In ECCV, 2022. 1" + } + ] + } + ], + "index": 27 + }, + { + "bbox": [ + 316, + 625, + 553, + 679 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 625, + 553, + 679 + ], + "spans": [ + { + "bbox": [ + 316, + 625, + 553, + 679 + ], + "type": "text", + "content": "[25] Zhenxin Li, Kailin Li, Shihao Wang, Shiyi Lan, Zhiding Yu, Yishen Ji, Zhiqi Li, Ziye Zhu, Jan Kautz, Zuxuan Wu, et al. Hydra-mdp: End-to-end multimodal planning with multi-target hydra-distillation. arXiv preprint arXiv:2406.06978, 2024. 1, 2, 6, 7" + } + ] + } + ], + "index": 28 + }, + { + "bbox": [ + 316, + 681, + 553, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 681, + 553, + 713 + ], + "spans": [ + { + "bbox": [ + 316, + 681, + 553, + 713 + ], + "type": "text", + "content": "[26] Zhiqi Li, Zhiding Yu, Shiyi Lan, Jiahan Li, Jan Kautz, Tong Lu, and Jose M Alvarez. Is ego status all you need for open-loop end-to-end autonomous driving? In CVPR, 2024. 2" + } + ] + } + ], + "index": 29 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "text", + "content": "12045" + } + ] + } + ], + "index": 31 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 8 + }, + { + "para_blocks": [ + { + "bbox": [ + 56, + 73, + 294, + 712 + ], + "type": "list", + "angle": 0, + "index": 16, + "blocks": [ + { + "bbox": [ + 56, + 73, + 294, + 116 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 73, + 294, + 116 + ], + "spans": [ + { + "bbox": [ + 56, + 73, + 294, + 116 + ], + "type": "text", + "content": "[27] Benchcheng Liao, Shaoyu Chen, Xinggang Wang, Tianheng Cheng, Qian Zhang, Wenyu Liu, and Chang Huang. MapTR: Structured modeling and learning for online vectorized HD map construction. In ICLR, 2023. 1" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 56, + 118, + 294, + 162 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 118, + 294, + 162 + ], + "spans": [ + { + "bbox": [ + 56, + 118, + 294, + 162 + ], + "type": "text", + "content": "[28] Benchcheng Liao, Shaoyu Chen, Yunchi Zhang, Bo Jiang, Qian Zhang, Wenyu Liu, Chang Huang, and Xinggang Wang. Maptrv2: An end-to-end framework for online vectorized hd map construction. IJCV, 2024. 1" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 56, + 163, + 294, + 206 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 163, + 294, + 206 + ], + "spans": [ + { + "bbox": [ + 56, + 163, + 294, + 206 + ], + "type": "text", + "content": "[29] Xuewu Lin, Tianwei Lin, Zixiang Pei, Lichao Huang, and Zhizhong Su. Sparse4d: Multi-view 3d object detection with sparse spatial-temporal fusion. arXiv preprint arXiv:2211.10581, 2022. 6" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 56, + 208, + 294, + 251 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 208, + 294, + 251 + ], + "spans": [ + { + "bbox": [ + 56, + 208, + 294, + 251 + ], + "type": "text", + "content": "[30] Yaron Lipman, Ricky T. Q. Chen, Heli Ben-Hamu, Maximilian Nickel, and Matthew Le. Flow matching for generative modeling. In The Eleventh International Conference on Learning Representations, 2023. 4" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 56, + 253, + 294, + 285 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 253, + 294, + 285 + ], + "spans": [ + { + "bbox": [ + 56, + 253, + 294, + 285 + ], + "type": "text", + "content": "[31] Xingchao Liu, Chengyue Gong, and Qiang Liu. Flow straight and fast: Learning to generate and transfer data with rectified flow. arXiv preprint arXiv:2209.03003, 2022. 4" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 56, + 286, + 294, + 319 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 286, + 294, + 319 + ], + "spans": [ + { + "bbox": [ + 56, + 286, + 294, + 319 + ], + "type": "text", + "content": "[32] Yicheng Liu, Tianyuan Yuan, Yue Wang, Yilun Wang, and Hang Zhao. Vectormapnet: End-to-end vectorized hd map learning. In ICML, 2023. 1" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 56, + 320, + 294, + 374 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 320, + 294, + 374 + ], + "spans": [ + { + "bbox": [ + 56, + 320, + 294, + 374 + ], + "type": "text", + "content": "[33] Robin Rombach, Andreas Blattmann, Dominik Lorenz, Patrick Esser, and Björn Ommer. High-resolution image synthesis with latent diffusion models. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 10684-10695, 2022. 4" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 56, + 376, + 294, + 408 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 376, + 294, + 408 + ], + "spans": [ + { + "bbox": [ + 56, + 376, + 294, + 408 + ], + "type": "text", + "content": "[34] Olaf Ronneberger, Philipp Fischer, and Thomas Brox. U-net: Convolutional networks for biomedical image segmentation. In MICCAI, 2015. 2" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 56, + 411, + 294, + 432 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 411, + 294, + 432 + ], + "spans": [ + { + "bbox": [ + 56, + 411, + 294, + 432 + ], + "type": "text", + "content": "[35] Jiaming Song, Chenlin Meng, and Stefano Ermon. Denoising diffusion implicit models. In ICLR, 2021. 4, 5, 6" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 56, + 434, + 294, + 487 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 434, + 294, + 487 + ], + "spans": [ + { + "bbox": [ + 56, + 434, + 294, + 487 + ], + "type": "text", + "content": "[36] Yang Song, Jascha Sohl-Dickstein, Diederik P Kingma, Abhishek Kumar, Stefano Ermon, and Ben Poole. Score-based generative modeling through stochastic differential equations. In International Conference on Learning Representations, 2021. 4" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 56, + 489, + 294, + 521 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 489, + 294, + 521 + ], + "spans": [ + { + "bbox": [ + 56, + 489, + 294, + 521 + ], + "type": "text", + "content": "[37] Ajay Sridhar, Dhruv Shah, Catherine Glossop, and Sergey Levine. Nomad: Goal masked diffusion policies for navigation and exploration. In ICRA, 2024. 4" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 56, + 523, + 294, + 566 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 523, + 294, + 566 + ], + "spans": [ + { + "bbox": [ + 56, + 523, + 294, + 566 + ], + "type": "text", + "content": "[38] Maria Stamatopoulou, Jianwei Liu, and Dimitrios Kanoulas. Dippest: Diffusion-based path planner for synthesizing trajectories applied on quadruped robots. arXiv preprint arXiv:2405.19232, 2024. 4" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 56, + 567, + 294, + 611 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 567, + 294, + 611 + ], + "spans": [ + { + "bbox": [ + 56, + 567, + 294, + 611 + ], + "type": "text", + "content": "[39] Wenchao Sun, Xuewu Lin, Yining Shi, Chuang Zhang, Hao ran Wu, and Sifa Zheng. Sparsedrive: End-to-end autonomous driving via sparse scene representation. arXiv preprint arXiv:2405.19620, 2024. 2, 6, 8" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 56, + 613, + 294, + 634 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 613, + 294, + 634 + ], + "spans": [ + { + "bbox": [ + 56, + 613, + 294, + 634 + ], + "type": "text", + "content": "[40] Mingxing Tan and Quoc Le. Efficientnet: Rethinking model scaling for convolutional neural networks. In ICML, 2019. 8" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 56, + 636, + 294, + 668 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 636, + 294, + 668 + ], + "spans": [ + { + "bbox": [ + 56, + 636, + 294, + 668 + ], + "type": "text", + "content": "[41] Wenwen Tong, Chonghao Sima, Tai Wang, Li Chen, Silei Wu, Hanming Deng, Yi Gu, Lewei Lu, Ping Luo, Dahua Lin, et al. Scene as occupancy. In ICCV, 2023. 8" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 56, + 670, + 294, + 712 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 670, + 294, + 712 + ], + "spans": [ + { + "bbox": [ + 56, + 670, + 294, + 712 + ], + "type": "text", + "content": "[42] Yue Wang, Vitor Campagnolo Guizilini, Tianyuan Zhang, Yilun Wang, Hang Zhao, and Justin Solomon. Detr3d: 3d object detection from multi-view images via 3d-to-2d queries. In CoRL, 2022. 1, 6" + } + ] + } + ], + "index": 15 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 316, + 73, + 553, + 712 + ], + "type": "list", + "angle": 0, + "index": 32, + "blocks": [ + { + "bbox": [ + 316, + 73, + 553, + 116 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 73, + 553, + 116 + ], + "spans": [ + { + "bbox": [ + 316, + 73, + 553, + 116 + ], + "type": "text", + "content": "[43] Yuqi Wang, Jiawei He, Lue Fan, Hongxin Li, Yuntao Chen, and Zhaoxiang Zhang. Driving into the future: Multiview visual forecasting and planning with world model for autonomous driving. In CVPR, 2024. 2" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 316, + 118, + 553, + 173 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 118, + 553, + 173 + ], + "spans": [ + { + "bbox": [ + 316, + 118, + 553, + 173 + ], + "type": "text", + "content": "[44] Yixiao Wang, Chen Tang, Lingfeng Sun, Simone Rossi, Yichen Xie, Chensheng Peng, Thomas Hannagan, Stefano Sabatini, Nicola Poerio, Masayoshi Tomizuka, et al. Optimizing diffusion models for joint trajectory prediction and controllable generation. In ECCV, 2024. 3" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 316, + 175, + 553, + 206 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 175, + 553, + 206 + ], + "spans": [ + { + "bbox": [ + 316, + 175, + 553, + 206 + ], + "type": "text", + "content": "[45] Xinshuo Weng, Boris Ivanovic, Yan Wang, Yue Wang, and Marco Pavone. Para-drive: Parallelized architecture for real-time autonomous driving. In CVPR, 2024. 2, 6" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 316, + 208, + 553, + 240 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 208, + 553, + 240 + ], + "spans": [ + { + "bbox": [ + 316, + 208, + 553, + 240 + ], + "type": "text", + "content": "[46] Zehang Weng, Haofei Lu, Danica Kragic, and Jens Lundell. Dexdiffuser: Generating dexterous grasps with diffusion models. arXiv preprint arXiv:2402.02989, 2024. 4" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 316, + 241, + 553, + 297 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 241, + 553, + 297 + ], + "spans": [ + { + "bbox": [ + 316, + 241, + 553, + 297 + ], + "type": "text", + "content": "[47] Sixu Yan, Zeyu Zhang, Muzhi Han, Zaijin Wang, Qi Xie, Zhitian Li, Zhehan Li, Hangxin Liu, Xinggang Wang, and Song-Chun Zhu. M2diffuser: Diffusion-based trajectory optimization for mobile manipulation in 3d scenes. arXiv preprint arXiv:2410.11402, 2024. 3" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 316, + 298, + 553, + 352 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 298, + 553, + 352 + ], + "spans": [ + { + "bbox": [ + 316, + 298, + 553, + 352 + ], + "type": "text", + "content": "[48] Brian Yang, Huangyuan Su, Nikolaos Gkanatsios, Tsung-Wei Ke, Ayush Jain, Jeff Schneider, and Katerina Fragkiadaki. Diffusion-es: Gradient-free planning with diffusion for autonomous driving and zero-shot instruction following. In CVPR, 2024. 3" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 316, + 354, + 553, + 387 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 354, + 553, + 387 + ], + "spans": [ + { + "bbox": [ + 316, + 354, + 553, + 387 + ], + "type": "text", + "content": "[49] Jingfeng Yao and Xinggang Wang. Reconstruction vs. generation: Taming optimization dilemma in latent diffusion models. arXiv preprint arXiv:2501.01423, 2025. 4" + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 316, + 388, + 553, + 430 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 388, + 553, + 430 + ], + "spans": [ + { + "bbox": [ + 316, + 388, + 553, + 430 + ], + "type": "text", + "content": "[50] Jingfeng Yao, Cheng Wang, Wenyu Liu, and Xinggang Wang. Fasteredit: Towards faster diffusion transformers training without architecture modification. arXiv preprint arXiv:2410.10356, 2024. 4" + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 316, + 432, + 553, + 476 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 432, + 553, + 476 + ], + "spans": [ + { + "bbox": [ + 316, + 432, + 553, + 476 + ], + "type": "text", + "content": "[51] Wenhao Yu, Jie Peng, Huanyu Yang, Junrui Zhang, Yifan Duan, Jianmin Ji, and Yanyong Zhang. Ldp: A local diffusion planner for efficient robot navigation and collision avoidance. arXiv preprint arXiv:2407.01950, 2024. 4" + } + ] + } + ], + "index": 25 + }, + { + "bbox": [ + 316, + 478, + 553, + 532 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 478, + 553, + 532 + ], + "spans": [ + { + "bbox": [ + 316, + 478, + 553, + 532 + ], + "type": "text", + "content": "[52] Chengran Yuan, Zhanqi Zhang, Jiawei Sun, Shuo Sun, Zefan Huang, Christina Dao Wen Lee, Dongen Li, Yuhang Han, Anthony Wong, Keng Peng Tee, et al. Drama: An efficient end-to-end motion planner for autonomous driving with mamba. arXiv preprint arXiv:2408.03601, 2024. 6" + } + ] + } + ], + "index": 26 + }, + { + "bbox": [ + 316, + 533, + 553, + 565 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 533, + 553, + 565 + ], + "spans": [ + { + "bbox": [ + 316, + 533, + 553, + 565 + ], + "type": "text", + "content": "[53] Yanjie Ze, Gu Zhang, Kangning Zhang, Chenyuan Hu, Muhan Wang, and Huazhe Xu. 3d diffusion policy. In RSS, 2024. 3" + } + ] + } + ], + "index": 27 + }, + { + "bbox": [ + 316, + 567, + 553, + 600 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 567, + 553, + 600 + ], + "spans": [ + { + "bbox": [ + 316, + 567, + 553, + 600 + ], + "type": "text", + "content": "[54] Fangao Zeng, Bin Dong, Yuang Zhang, Tiancai Wang, Xi-angyu Zhang, and Yichen Wei. Motr: End-to-end multiple-object tracking with transformer. In ECCV, 2022. 1" + } + ] + } + ], + "index": 28 + }, + { + "bbox": [ + 316, + 601, + 553, + 634 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 601, + 553, + 634 + ], + "spans": [ + { + "bbox": [ + 316, + 601, + 553, + 634 + ], + "type": "text", + "content": "[55] Yifu Zhang, Chunyu Wang, Xinggang Wang, Wenjun Zeng, and Wenyu Liu. Fairmot: On the fairness of detection and re-identification in multiple object tracking. IJCV, 2021." + } + ] + } + ], + "index": 29 + }, + { + "bbox": [ + 316, + 635, + 553, + 678 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 635, + 553, + 678 + ], + "spans": [ + { + "bbox": [ + 316, + 635, + 553, + 678 + ], + "type": "text", + "content": "[56] Yifu Zhang, Peize Sun, Yi Jiang, Dongdong Yu, Fucheng Weng, Zehuan Yuan, Ping Luo, Wenyu Liu, and Xinggang Wang. Bytetrack: Multi-object tracking by associating every detection box. In ECCV, 2022. 1" + } + ] + } + ], + "index": 30 + }, + { + "bbox": [ + 316, + 680, + 553, + 712 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 680, + 553, + 712 + ], + "spans": [ + { + "bbox": [ + 316, + 680, + 553, + 712 + ], + "type": "text", + "content": "[57] Huangjie Zheng, Pengcheng He, Weizhu Chen, and Mingyuan Zhou. Truncated diffusion probabilistic models and diffusion-based adversarial auto-encoders. In The" + } + ] + } + ], + "index": 31 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 749, + 318, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 749, + 318, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 749, + 318, + 757 + ], + "type": "text", + "content": "12046" + } + ] + } + ], + "index": 33 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 9 + }, + { + "para_blocks": [ + { + "bbox": [ + 56, + 72, + 295, + 297 + ], + "type": "list", + "angle": 0, + "index": 6, + "blocks": [ + { + "bbox": [ + 76, + 72, + 294, + 95 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 76, + 72, + 294, + 95 + ], + "spans": [ + { + "bbox": [ + 76, + 72, + 294, + 95 + ], + "type": "text", + "content": "Eleventh International Conference on Learning Representations, 2023. 4" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 56, + 96, + 295, + 129 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 96, + 295, + 129 + ], + "spans": [ + { + "bbox": [ + 56, + 96, + 295, + 129 + ], + "type": "text", + "content": "[58] Wenzhao Zheng, Ruiqi Song, Xianda Guo, Chenming Zhang, and Long Chen. Genad: Generative end-to-end autonomous driving. In ECCV, 2024. 2" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 56, + 130, + 294, + 173 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 130, + 294, + 173 + ], + "spans": [ + { + "bbox": [ + 56, + 130, + 294, + 173 + ], + "type": "text", + "content": "[59] Ziyuan Zhong, Davis Rempe, Yuxiao Chen, Boris Ivanovic, Yulong Cao, Danfei Xu, Marco Pavone, and Baishakhi Ray. Language-guided traffic simulation via scene-level diffusion. In CoRL, 2023. 3" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 56, + 175, + 294, + 218 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 175, + 294, + 218 + ], + "spans": [ + { + "bbox": [ + 56, + 175, + 294, + 218 + ], + "type": "text", + "content": "[60] Ziyuan Zhong, Davis Rempe, Danfei Xu, Yuxiao Chen, Sushant Veer, Tong Che, Baishakhi Ray, and Marco Pavone. Guided conditional diffusion for controllable traffic simulation. In ICRA, 2023. 3" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 56, + 220, + 294, + 262 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 220, + 294, + 262 + ], + "spans": [ + { + "bbox": [ + 56, + 220, + 294, + 262 + ], + "type": "text", + "content": "[61] Lianghui Zhu, Zilong Huang, Bencheng Liao, Jun Hao Liew, Hanshu Yan, Jiashi Feng, and Xinggang Wang. Dig: Scalable and efficient diffusion models with gated linear attention. arXiv preprint arXiv:2405.18428, 2024. 4" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 56, + 264, + 294, + 297 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 264, + 294, + 297 + ], + "spans": [ + { + "bbox": [ + 56, + 264, + 294, + 297 + ], + "type": "text", + "content": "[62] Xizhou Zhu, Weijie Su, Lewei Lu, Bin Li, Xiaogang Wang, and Jifeng Dai. Deformable detr: Deformable transformers for end-to-end object detection. In ICLR, 2021. 2, 6" + } + ] + } + ], + "index": 5 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 749, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 749, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 749, + 317, + 757 + ], + "type": "text", + "content": "12047" + } + ] + } + ], + "index": 7 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 10 + } + ], + "_backend": "vlm", + "_version_name": "2.6.4" +} \ No newline at end of file diff --git a/2025/DiffusionSfM_ Predicting Structure and Motion via Ray Origin and Endpoint Diffusion/a0e804a0-3c59-43c5-bb05-b2160d8636bd_content_list.json b/2025/DiffusionSfM_ Predicting Structure and Motion via Ray Origin and Endpoint Diffusion/a0e804a0-3c59-43c5-bb05-b2160d8636bd_content_list.json new file mode 100644 index 0000000000000000000000000000000000000000..f18749ad0aa60bfa968bdfc32857b0e71ba444d4 --- /dev/null +++ b/2025/DiffusionSfM_ Predicting Structure and Motion via Ray Origin and Endpoint Diffusion/a0e804a0-3c59-43c5-bb05-b2160d8636bd_content_list.json @@ -0,0 +1,1835 @@ +[ + { + "type": "text", + "text": "DiffusionSfM: Predicting Structure and Motion via Ray Origin and Endpoint Diffusion", + "text_level": 1, + "bbox": [ + 258, + 130, + 738, + 176 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Qitao Zhao, Amy Lin, Jeff Tan, Jason Y. Zhang, Deva Ramanan, Shubham Tulsiani Carnegie Mellon University", + "bbox": [ + 169, + 203, + 826, + 239 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Project page: qitaozhao.github.io/DiffusionSfM", + "bbox": [ + 307, + 239, + 689, + 257 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/ac31d7718dc556e701eba8b776c1f3e30bb00ad6f957b560bb72e53affc2382c.jpg", + "image_caption": [ + "Figure 1. DiffusionSfM. Top: Given a set of multi-view images (left), DiffusionSfM represents scene geometry and cameras (right) as pixel-wise ray origins and endpoints in a global frame. It learns a denoising diffusion model to infer these elements directly from multiview inputs. Unlike traditional Structure-from-Motion (SfM) pipelines, which separate pairwise reasoning and global optimization into two stages, our approach unifies both into a single end-to-end multi-view reasoning framework. Bottom: Example results of inferred scene geometry and cameras for two distinct settings: a real-world outdoor scene (left) and a synthetic indoor scene (right)." + ], + "image_footnote": [], + "bbox": [ + 91, + 268, + 906, + 498 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Abstract", + "text_level": 1, + "bbox": [ + 248, + 597, + 326, + 613 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Current Structure-from-Motion (SfM) methods typically follow a two-stage pipeline, combining learned or geometric pairwise reasoning with a subsequent global optimization step. In contrast, we propose a data-driven multi-view reasoning approach that directly infers 3D scene geometry and camera poses from multi-view images. Our framework, DiffusionSfM, parameterizes scene geometry and cameras as pixel-wise ray origins and endpoints in a global frame and employs a transformer-based denoising diffusion model to predict them from multi-view inputs. To address practical challenges in training diffusion models with missing data and unbounded scene coordinates, we introduce specialized mechanisms that ensure robust learning. We empirically validate DiffusionSfM on both synthetic and real datasets, demonstrating that it outperforms classical and learning-based approaches while naturally modeling uncertainty.", + "bbox": [ + 88, + 642, + 485, + 885 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "1. Introduction", + "text_level": 1, + "bbox": [ + 514, + 597, + 643, + 613 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "The task of recovering structure (geometry) and motion (cameras) from multi-view images has long been a focus of the computer vision community, with typical pipelines [25] performing pairwise correspondence estimation followed by global optimization. While classical methods relied on hand-designed features, matching, and optimization, there has been a recent shift towards incorporating learning-based alternatives [5, 6, 14, 23]. More recently, the widely influential DUSt3R [36] advocates for predicting pairwise 3D pointmaps (instead of only correspondences), demonstrating that this can yield accurate dense geometry and cameras. In order to reconstruct more than two views, DUSt3R (and its variants [10]) still require a global optimization reminiscent of classic bundle adjustment. While these methods, both classical and learning-based, have led to impressive improvements in SfM, the overall approach is largely unchanged – learned or geometric pairwise reasoning followed by global optimization. In this work, we", + "bbox": [ + 511, + 628, + 908, + 902 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "CVF", + "bbox": [ + 106, + 2, + 181, + 42 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore.", + "bbox": [ + 236, + 0, + 810, + 46 + ], + "page_idx": 0 + }, + { + "type": "page_number", + "text": "6317", + "bbox": [ + 482, + 944, + 514, + 955 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "seek to develop an alternative approach that directly predicts both structure and motion, unifying pairwise reasoning and global optimization into a single multi-view framework.", + "bbox": [ + 89, + 90, + 480, + 136 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "We are of course not the first to attempt to find unified alternatives to the two-stage SfM pipeline. In the sparse-view setting where conventional correspondence-based methods struggle, several works employ multi-view architectures to jointly reason across input images. SparsePose [26], RelPose++ [12], and PoseDiffusion [34] all leverage multiview transformers to estimate camera pose for input images, albeit using differing mechanisms such as regression, energy-based modeling, and denoising diffusion. More recently, RayDiffusion [41] argues for a local raymap parameterization of cameras instead of a global extrinsic matrix and shows that existing patch-based transformers can be easily adapted for this task, yielding significantly more accurate pose predictions. Importantly, such methods predict only camera motion and fail to predict scene structure.", + "bbox": [ + 88, + 138, + 482, + 364 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "In this work, we present DiffusionSfM, an end-to-end multi-view model that directly infers dense 3D geometry and cameras from multiple input images. Instead of inferring (depth-agnostic) rays per image patch (as in RayDiffusion [41]) or 3D points per pixel (as in DUSt3R [36]), DiffusionSfM effectively combines both to predict ray origins and endpoints per pixel, directly reporting both scene geometry (endpoints) and generalized cameras (rays). These can readily be converted back to traditional cameras [41]. Compared to RayDiffusion, our model directly predicts structure as well as motion at a finer scale (pixel-wise v.s. patchwise). Compared to DUSt3R, our model directly predicts motion as well as structure but, even more importantly, does so for $N$ views, eliminating the need for memory-intensive global alignment. To model uncertainty, we train a denoising diffusion model but find two key challenges that need to be addressed. First, diffusion models require (noisy) ground truth as input for training, but existing real datasets do not have known endpoints for all pixels due to missing depth in multi-view stereo. Second, the 3D coordinates of endpoints can be potentially unbounded, whereas diffusion models require normalized data. We develop mechanisms to overcome these challenges, leveraging additional \"GT mask conditioning\" as input to inform the model of missing input data and parameterizing 3D points in projective space instead of Euclidean space. We find these strategies allow us to learn accurate predictions for structure and motion.", + "bbox": [ + 91, + 369, + 482, + 776 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "We train and evaluate DiffusionSfM on real-world and synthetic datasets [21, 24, 42] and find that it can infer accurate geometry and cameras for both object-centric and scene-level images (see Fig. 1). In particular, we find that DiffusionSfM yields more accurate camera estimates compared to prior work across these settings while also modeling the underlying uncertainty via the diffusion process. In summary, we show that DiffusionSfM can serve as a unified", + "bbox": [ + 89, + 779, + 482, + 900 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "multi-view reasoning model for 3D geometry and cameras.", + "bbox": [ + 513, + 90, + 901, + 106 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "2. Related Work", + "text_level": 1, + "bbox": [ + 513, + 121, + 653, + 137 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Structure from Motion. Structure-from-Motion (SfM) systems [25] aim to simultaneously recover geometry and cameras given a set of input images. The typical SfM pipeline extracts pixel correspondences from keypoint matching [2, 16], and performs global bundle adjustment (BA) to optimize sparse 3D points and camera parameters by minimizing reprojection errors. Recently, SfM pipelines have been substantially enhanced by replacing classical subcomponents with learning-based methods, such as neural feature descriptors [6, 8], keypoint matching [15, 23, 30], and bundle adjustment [13, 32].", + "bbox": [ + 511, + 147, + 903, + 313 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "More recently, an emerging body of research aims to unify the various SfM subcomponents into an end-to-end neural framework. Notably, ACEZero [3] fits a single neural network to input images and learns pixel-aligned 3D coordinates in a self-supervised manner, while FlowMap [27] predicts per-frame cameras and depth maps using off-the-shelf optical flow as supervision. Though ACEZero and FlowMap are promising attempts to revolutionize SfM pipelines, they both register images incrementally and may suffer under large viewpoint changes. DUSt3R [36] directly regresses 3D pointmaps from image pairs and shows strong generalization ability [11, 21]. Building on DUSt3R, MASt3R [10] introduces a feature head, offering pixel-matching capabilities. MASt3R-SfM [7] is a more scalable SfM pipeline based on MASt3R. While these approaches [7, 10, 36] show impressive performance and robustness under sparse views, they are essentially pair-based, requiring sophisticated global alignment procedures to form a consistent estimate for more than two views.", + "bbox": [ + 511, + 314, + 903, + 599 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Pose Estimation with Global Reasoning. For the task of sparse-view pose estimation, learning-based methods equipped with global reasoning show favorable robustness where traditional SfM methods [25, 28] fail. This line of research includes energy-based [12, 40], regression-based [26], and diffusion-based pose estimators [34, 41]. Among them, diffusion-based methods show a better ability to handle uncertainty [41]. Closest to our work, RayDiffusion [41] leverages a denoising diffusion model [9, 19] with a patch-aligned (depth-unaware) ray representation to predict generic cameras. Our method goes further and pursues a generic representation for both geometry and cameras in the form of ray origins and endpoints for each pixel. In addition to resulting in a richer output, this joint geometry and pose prediction also yields improvements for pose estimation.", + "bbox": [ + 511, + 602, + 903, + 829 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "3. Method", + "text_level": 1, + "bbox": [ + 513, + 844, + 604, + 859 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Given a set of sparse (i.e., 2-8) input images, DiffusionSfM predicts the geometry and cameras of a 3D scene in a global", + "bbox": [ + 511, + 869, + 903, + 900 + ], + "page_idx": 1 + }, + { + "type": "page_number", + "text": "6318", + "bbox": [ + 482, + 945, + 514, + 955 + ], + "page_idx": 1 + }, + { + "type": "image", + "img_path": "images/e313265c3ef039eaa8c98aab188fac7c1ec90a35c2ae39191b416a12c793c328.jpg", + "image_caption": [ + "Figure 2. Method. Given sparse multi-view images as input, DiffusionSfM predicts pixel-wise ray origins and endpoints in a global frame (Sec. 3.1) using a denoising diffusion process (Sec. 3.2). For each image, we compute patch-wise embeddings with DINOv2 [18] and embed noisy ray origins and endpoints into latents using a single downsampling convolutional layer, ensuring alignment with the spatial footprint of the image embeddings. We implement a Diffusion Transformer architecture that predicts clean ray origins and endpoints from noisy samples. A convolutional DPT [20] head outputs full-resolution denoised ray origins and endpoints. To handle incomplete ground truth (GT) during training, we condition the model on a GT mask (Sec. 3.3). At inference, the GT mask is set to all ones, enabling the model to predict origins and endpoints for all pixels. The predicted ray origins and endpoints can be directly visualized in 3D or postprocessed to recover camera extrinsics, intrinsics, and multi-view consistent depth maps." + ], + "image_footnote": [], + "bbox": [ + 94, + 88, + 908, + 215 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "coordinate frame. In Sec. 3.1, we propose to represent 3D scenes as dense pixel-aligned ray origins and endpoints. To predict such scene representations from sparse input images while modeling uncertainty, Sec. 3.2 proposes a denoising diffusion architecture. We then discuss some key practical challenges in training such a model in Sec. 3.3.", + "bbox": [ + 88, + 363, + 483, + 454 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3.1. 3D Scenes as Ray Origins and Endpoints", + "text_level": 1, + "bbox": [ + 89, + 462, + 439, + 478 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Given an input image $\\mathbf{I} \\in \\mathbb{R}^{H \\times W \\times 3}$ with a depth map $\\mathbf{D} \\in \\mathbb{R}^{H \\times W}$ , camera intrinsics $\\mathbf{K} \\in \\mathbb{R}^{3 \\times 3}$ , and world-to-camera extrinsics $\\mathbf{T} \\in \\mathbb{R}^{4 \\times 4}$ (equivalently, rotation $\\mathbf{R} \\in SO(3)$ and translation $\\mathbf{t} \\in \\mathbb{R}^3$ ), each 2D image pixel $\\mathbf{P}_{ij} = [u, v]$ corresponds to a ray that travels from the camera center $\\mathbf{c}$ through the pixel's projected position on the image plane, terminating at the object's surface as specified by the depth map $\\mathbf{D}$ . The endpoint of the ray associated with image pixel $\\mathbf{P}_{ij}$ is given by:", + "bbox": [ + 89, + 482, + 483, + 619 + ], + "page_idx": 2 + }, + { + "type": "equation", + "text": "\n$$\n\\mathbf {E} _ {i j} = \\mathbf {T} ^ {- 1} h (\\mathbf {D} _ {i j} \\cdot \\mathbf {K} ^ {- 1} [ u, v, 1 ] ^ {T}) \\tag {1}\n$$\n", + "text_format": "latex", + "bbox": [ + 166, + 627, + 480, + 646 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "where $h$ maps the 3D point into homogeneous coordinates. The shared ray origin $\\mathbf{O}_{ij}$ for all pixels is equivalent to the camera center $\\mathbf{c}$ , and can be computed as:", + "bbox": [ + 89, + 654, + 482, + 699 + ], + "page_idx": 2 + }, + { + "type": "equation", + "text": "\n$$\n\\mathbf {O} _ {i j} = \\mathbf {c} = h (- \\mathbf {R} ^ {- 1} \\mathbf {t}) \\tag {2}\n$$\n", + "text_format": "latex", + "bbox": [ + 204, + 708, + 480, + 726 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "In summary, we associate each image pixel with a ray origin and endpoint $\\mathbf{S}_{ij} = \\langle \\mathbf{O}_{ij},\\mathbf{E}_{ij}\\rangle$ in world coordinates, describing the location of the observing camera and the observed 3D point on the object surface. Given a bundle of ray origins and endpoints, we can easily extract the corresponding camera pose [41].", + "bbox": [ + 89, + 734, + 482, + 824 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Learning Over-Parameterized Representations. We represent 3D scenes and cameras using distributed ray origins $\\mathbf{O}$ and endpoints $\\mathbf{E}$ rather than global alternatives such as quaternions and translation vectors. This design is inspired by RayDiffusion [41], and it facilitates the use of the", + "bbox": [ + 89, + 825, + 483, + 900 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "distributed deep features learned by state-of-the-art vision backbones, such as DINOv2 [18], which encode image information in a patch-wise manner. Notably, while ray origins $\\mathbf{O}$ should ideally be identical across all pixels, we predict them densely alongside ray endpoints $\\mathbf{E}$ . This encourages ray origins to remain close within the same image, providing implicit regularization during training. Practically, predicting both ray origins and endpoints simultaneously is easy to implement using a single projection head.", + "bbox": [ + 511, + 363, + 906, + 500 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3.2. DiffusionSfM", + "text_level": 1, + "bbox": [ + 511, + 507, + 653, + 522 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "We propose a denoising Diffusion Transformer (DiT) architecture [19] that predicts ray origins and endpoints (Sec. 3.1) via a denoising diffusion process. An overview of DiffusionSfM is given in Fig. 2.", + "bbox": [ + 511, + 529, + 905, + 589 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Diffusion Framework. Given pixel-aligned ray origins and endpoints $S = \\text{stack}(\\{\\mathbf{S}^{(n)}\\}_{n=1}^{N})$ associated with a set of $N$ input images, we apply a forward diffusion process [9, 29] that adds time-dependent Gaussian noise to them. Let $S_{t}$ denote the noisy ray origins and endpoints at timestep $t$ , where $S_{0}$ is the clean sample and $S_{T}$ (at the final diffusion step $T$ ) approximates pure Gaussian noise. The forward diffusion process is defined as:", + "bbox": [ + 511, + 590, + 906, + 710 + ], + "page_idx": 2 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {S} _ {t} = \\sqrt {\\bar {\\alpha} _ {t}} \\mathcal {S} _ {0} + \\sqrt {1 - \\bar {\\alpha} _ {t}} \\epsilon \\tag {3}\n$$\n", + "text_format": "latex", + "bbox": [ + 617, + 720, + 903, + 738 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "where $t \\sim \\mathrm{Uniform}(0, T]$ , $\\epsilon \\sim \\mathcal{N}(0, I)$ , and $\\bar{\\alpha}_t$ follows a pre-defined noise schedule that controls the strength of added noise at each timestep. To perform the reverse diffusion process, which progressively reconstructs the clean sample given noisy observations, we train a diffusion model $f_\\theta$ that takes $\\mathcal{S}_t$ as input and optionally incorporates additional conditioning information $\\mathcal{C}$ . The model is trained using the following loss function (with the \" $x_0$ \" objective):", + "bbox": [ + 511, + 750, + 906, + 869 + ], + "page_idx": 2 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {L} _ {\\text {D i f f u s i o n}} = \\mathbb {E} _ {t, S _ {0}, \\epsilon} \\| S _ {0} - f _ {\\theta} (S _ {t}, t, \\mathcal {C}) \\| ^ {2} \\tag {4}\n$$\n", + "text_format": "latex", + "bbox": [ + 578, + 883, + 903, + 901 + ], + "page_idx": 2 + }, + { + "type": "page_number", + "text": "6319", + "bbox": [ + 480, + 944, + 514, + 955 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Architecture. We implement $f_{\\theta}$ using a DiT [19] architecture conditioned on deep image features $\\mathcal{C} \\in \\mathbb{R}^{N \\times h \\times w \\times c_1}$ from DINOv2 [18], where $h$ and $w$ are the patch resolution and $c_1$ is the embedding dimension. To align pixels to the spatial information learned by DINOv2, we apply a convolutional layer with kernel size and stride equal to ViT patch size: this spatially downsamples the noisy ray origins and endpoints $S_t$ to match the DINOv2 features while increasing their feature dimension:", + "bbox": [ + 89, + 90, + 480, + 226 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {F} = \\operatorname {C o n v} \\left(\\mathcal {S} _ {t}\\right) \\in \\mathbb {R} ^ {N \\times h \\times w \\times c _ {2}} \\tag {5}\n$$\n", + "text_format": "latex", + "bbox": [ + 179, + 234, + 482, + 252 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "The combined DiT input is constructed by concatenating these two feature sets along the channel dimension: $\\mathcal{F} \\oplus \\mathcal{C}$ . Within DiT, patch-wise features attend to others through self-attention [33]. To distinguish between different images and their respective patches, we apply sinusoidal positional encoding [33] based on image and patch indices.", + "bbox": [ + 89, + 260, + 482, + 349 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "While the DiT operates on low-resolution features, our objective is to produce pixel-aligned dense ray origins and endpoints. To achieve this, we employ a DPT (Dense Prediction Transformer) [20] decoder, which takes intermediate feature maps from both DINoV2 and DiT as inputs. The DPT decoder progressively increases the feature resolution through several convolutional layers. The final ray origins and endpoints are decoded from the DPT output using a single linear layer. During inference, we apply the trained model in the reverse diffusion process to iteratively denoise a randomly initialized Gaussian sample.", + "bbox": [ + 89, + 351, + 483, + 517 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "3.3. Practical Training Considerations", + "text_level": 1, + "bbox": [ + 89, + 525, + 387, + 540 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Homogeneous Coordinates for Unbounded Geometry. Real-world scenes often exhibit significant variations in scale, both across different scenes and within a single scene. While normalizing by the average distance of 3D points [36] helps address cross-scene variation, within-scene variations remain. For instance, a background building may be much farther away than foreground elements, potentially resulting in extremely large coordinate values when generating ground-truth ray origins and endpoints. However, neural networks (especially diffusion models) tend to train most effectively when working with bounded inputs and outputs (e.g., between -1 and 1). To stabilize training across the large scale variations present in 3D scene datasets, we propose to represent ray origins and endpoints in homogeneous coordinates. Specifically, given any 3D point, we apply a homogeneous transform from $\\mathbb{R}^3\\to \\mathbb{P}^3$ :", + "bbox": [ + 89, + 547, + 482, + 787 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n(x, y, z) \\rightarrow \\frac {1}{w} (x, y, z, 1) \\tag {6}\n$$\n", + "text_format": "latex", + "bbox": [ + 202, + 796, + 482, + 824 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "where $w$ is an arbitrary scale factor. To encourage bounded coordinates in practice, we choose $w$ such that the homogeneous coordinate is unit-norm:", + "bbox": [ + 89, + 830, + 483, + 875 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\nw := \\sqrt {x ^ {2} + y ^ {2} + z ^ {2} + 1} \\tag {7}\n$$\n", + "text_format": "latex", + "bbox": [ + 197, + 883, + 482, + 901 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Unit normalization allows homogeneous coordinates to serve as a bounded representation for unbounded scene geometry. For example, $(x,y,z,0)$ is a point at infinity in the direction of $(x,y,z)$ . We find that this representation makes large coordinate values more tractable during training.", + "bbox": [ + 511, + 90, + 903, + 167 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Training with Incomplete Ground Truth. Many real-world datasets (e.g., CO3D [21] and MegaDepth [11]), provide only sparse point clouds, leading to incomplete depth information. This presents a significant challenge in diffusion training, as ground-truth (GT) depth values that are used to create clean ray endpoints often contain invalid or missing data. It is highly undesirable for these missing ray endpoints to be interpreted as part of the target distribution. Unlike regression models (e.g., DUSt3R [36]), which do not require (noisy) GT data as input and can simply apply a mask on the loss for supervision, diffusion models must handle incomplete inputs during training.", + "bbox": [ + 511, + 167, + 903, + 348 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "To mitigate this issue, we further apply GT masks $\\mathcal{M} \\in \\mathbb{R}^{N \\times H \\times W}$ to the DiT inputs, where zero values indicate pixels with invalid depth. During training, we multiply DiT inputs with GT masks element-wise, then concatenate along the channel dimension: $S_{t}' = (\\mathcal{M} \\cdot S_{t}) \\oplus \\mathcal{M}$ . Then, we only compute the diffusion loss in Eq. 4 over unmasked pixels. By implementing these strategies, we encourage the model to focus on regions with valid GT values during training. During inference, however, we would like the diffusion process to estimate ray endpoints at all pixels, so we always use GT masks with values set to one.", + "bbox": [ + 511, + 349, + 905, + 515 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Sparse-to-Dense Training. In practice, we find that training the entire model from scratch leads to slow convergence and suboptimal performance. To address this, we propose a sparse-to-dense training approach. First, we train a sparse version of the model, where the DPT decoder is removed, and the output ray origins and endpoints have the same spatial resolution as the DINOv2 features. Unlike Eq. 5, no spatial downsampling is required, so this sparse model uses a single linear layer to embed the noisy ray origins and endpoints. Once the sparse model is trained, we initialize the dense model DiT with the learned weights from the sparse model. This two-stage approach significantly improves performance; see the supplementary material for comparisons.", + "bbox": [ + 511, + 517, + 905, + 714 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "4. Experiments", + "text_level": 1, + "bbox": [ + 511, + 729, + 643, + 747 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "4.1. Experimental Setup", + "text_level": 1, + "bbox": [ + 511, + 756, + 702, + 772 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Datasets. We introduce two model variants, each trained on different datasets. (1) DiffusionSfM-CO3D: Following prior work [12, 41], we train and evaluate our model on the CO3D dataset [21], which consists of turntable video sequences of various object categories. Specifically, we train on 41 object categories and evaluate on both these seen categories and an additional 10 unseen categories to assess generalization. (2) DiffusionSfM: This variant is trained on the", + "bbox": [ + 511, + 779, + 905, + 900 + ], + "page_idx": 3 + }, + { + "type": "page_number", + "text": "6320", + "bbox": [ + 482, + 945, + 514, + 955 + ], + "page_idx": 3 + }, + { + "type": "image", + "img_path": "images/d6e463bb54fb1e9753a107c4e0f00652fabfafb5aeb6a00a0e8c75bcf8a0d463.jpg", + "image_caption": [ + "Input Images /GT Point Clouds" + ], + "image_footnote": [], + "bbox": [ + 122, + 119, + 261, + 172 + ], + "page_idx": 4 + }, + { + "type": "image", + "img_path": "images/dc5f60efbcfe9c7755289675788486c4578c9f213d2b6042429e3e03c31ae41f.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 124, + 172, + 261, + 224 + ], + "page_idx": 4 + }, + { + "type": "image", + "img_path": "images/4a90c02cbf86ede1c3b9f9128f7369efafe2004fac7f244144acbef7b880efc4.jpg", + "image_caption": [ + "DiffusionSfM" + ], + "image_footnote": [], + "bbox": [ + 299, + 128, + 460, + 215 + ], + "page_idx": 4 + }, + { + "type": "image", + "img_path": "images/4dbeed7353c79ac7b9a429bdb597e8c4bf321a4dea76408dad8d2d25a4cb3311.jpg", + "image_caption": [ + "RD + MoGe" + ], + "image_footnote": [], + "bbox": [ + 511, + 128, + 673, + 214 + ], + "page_idx": 4 + }, + { + "type": "image", + "img_path": "images/3c4d4f0156271974c5a1bea2211d4c0da5b8a04dce2f24b68c677ecb90c28e64.jpg", + "image_caption": [ + "DUSt3R" + ], + "image_footnote": [], + "bbox": [ + 718, + 128, + 877, + 214 + ], + "page_idx": 4 + }, + { + "type": "image", + "img_path": "images/08b9d18a8fd24f92149aabec1779e2696ba49c6f961ae07dcdb2a0eb2a56a1d2.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 117, + 232, + 325, + 316 + ], + "page_idx": 4 + }, + { + "type": "image", + "img_path": "images/6d07fe01df1484bab4bf2ec9045876531d07cc800c093177a7bbbbbfac0de292.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 325, + 234, + 519, + 315 + ], + "page_idx": 4 + }, + { + "type": "image", + "img_path": "images/21cddbbf7d3c3a2aa6c186d5790bb68351b7e18e5fec658f160d0f440c0ebba2.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 519, + 232, + 663, + 315 + ], + "page_idx": 4 + }, + { + "type": "image", + "img_path": "images/4467112fd81e570e98956508a4e57480691fb0ba442c38bbf2627e053ef2ac55.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 663, + 232, + 880, + 316 + ], + "page_idx": 4 + }, + { + "type": "image", + "img_path": "images/272dd89e3cebb5972006b1e754848a814b96b262781b027efbac94167d73da4f.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 120, + 329, + 168, + 364 + ], + "page_idx": 4 + }, + { + "type": "image", + "img_path": "images/f9d752bc2a72574ad8da4962e3c75c5038c52407176760b4817e89875dc60734.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 120, + 364, + 168, + 401 + ], + "page_idx": 4 + }, + { + "type": "image", + "img_path": "images/77e3edccd4250d49af8e4e40afe6cba4c9d6793ae40e2556aa588e44db248cf8.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 169, + 330, + 212, + 364 + ], + "page_idx": 4 + }, + { + "type": "image", + "img_path": "images/764b33e09ba83273a035143014a202821327b88eae6f527e802b136b7abbff21.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 214, + 330, + 261, + 364 + ], + "page_idx": 4 + }, + { + "type": "image", + "img_path": "images/abec2e30ad827dfb18d995acedef81d339f0eae11fb77e60e29a29e8e3f6f10d.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 119, + 415, + 315, + 492 + ], + "page_idx": 4 + }, + { + "type": "image", + "img_path": "images/083a0c1897454b692a6d69688d3baf77dee68c4161fa70f086982461b4a28fe5.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 313, + 415, + 517, + 493 + ], + "page_idx": 4 + }, + { + "type": "image", + "img_path": "images/15555cfb9f16f6fdc453589fc0c91b42a2258d1260a60ff53f08904d6c5ffc59.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 521, + 415, + 687, + 491 + ], + "page_idx": 4 + }, + { + "type": "image", + "img_path": "images/d9ba0182fe76e9bcf638276fbe614ef41034a2b339947df162dd809ad1cd3fa0.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 728, + 321, + 880, + 407 + ], + "page_idx": 4 + }, + { + "type": "image", + "img_path": "images/42ead853a68a77308aba6a032385977d1aee9b389b6819fe8f30cd69f3b1bd53.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 691, + 414, + 880, + 493 + ], + "page_idx": 4 + }, + { + "type": "image", + "img_path": "images/3d56dfda9c0cd215776b017ae9736d19c24d055bcab58416313922bee5a94489.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 120, + 508, + 169, + 544 + ], + "page_idx": 4 + }, + { + "type": "image", + "img_path": "images/b1b3bbe97ed1d5d38efbfd8c6a271fd8662c88298f9063cb5d973f46e0ab718e.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 120, + 545, + 168, + 580 + ], + "page_idx": 4 + }, + { + "type": "image", + "img_path": "images/c44111bf71e33372cf04e561079591ede87ef0cb656da94d7fb85cf5cd0f6dba.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 120, + 582, + 168, + 618 + ], + "page_idx": 4 + }, + { + "type": "image", + "img_path": "images/621af54fe4b9ca0aa56cb232e2c29f811110ddb5e15082e17e506a654080d459.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 169, + 508, + 212, + 544 + ], + "page_idx": 4 + }, + { + "type": "image", + "img_path": "images/75308784803a2b539f04991f58dc5590242f78993005180f68724760b4e2d4f9.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 169, + 545, + 212, + 580 + ], + "page_idx": 4 + }, + { + "type": "image", + "img_path": "images/8ff9e2401bc30c177a123e84029058340270d40a3e6b1993cc4d906e2ce533e5.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 169, + 582, + 212, + 618 + ], + "page_idx": 4 + }, + { + "type": "image", + "img_path": "images/64b87ce018823ad58c1359ca2b1ff31994f93d3890ad8717504cbb3578ac1657.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 214, + 508, + 261, + 544 + ], + "page_idx": 4 + }, + { + "type": "image", + "img_path": "images/c919abb7dcae4e0ccf25434d3735e9779fcdd4eeae41a41a769f2dbcea89668c.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 214, + 545, + 261, + 580 + ], + "page_idx": 4 + }, + { + "type": "image", + "img_path": "images/a9ea49154a5cddebf7b1215c95540c58c714ca3a87f5f9090d30a8f5a9a5356c.jpg", + "image_caption": [ + "Figure 3. Qualitative Comparison on Camera Pose Accuracy and Predicted Geometry. For each method, we plot the ground-truth cameras in black and the predicted cameras in other colors. DiffusionSfM demonstrates robust performance even with challenging inputs. Compared to DUSt3R, which sometimes fails to register images in a consistent manner, DiffusionSfM consistently yields a coherent global prediction. Additionally, while we observe that DUSt3R can predict highly precise camera rotations, it often struggles with camera centers (see the backpack example). Input images depicting scenes are out-of-distribution for RayDiffusion, as it is trained on CO3D only." + ], + "image_footnote": [], + "bbox": [ + 114, + 621, + 240, + 723 + ], + "page_idx": 4 + }, + { + "type": "image", + "img_path": "images/1737d179f0ff07dcf799bb9ca29f91d33c878cd8e84667b551b4f902f45aa969.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 302, + 522, + 468, + 601 + ], + "page_idx": 4 + }, + { + "type": "image", + "img_path": "images/b07132582448c16a63a6b68ee425225bc70df6e576c168687035e4f1ffa818fa.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 511, + 522, + 676, + 601 + ], + "page_idx": 4 + }, + { + "type": "image", + "img_path": "images/355211bdc5d80708987da0be50e094a64357cea1d97bc362e30bbd5e8f491890.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 718, + 522, + 883, + 604 + ], + "page_idx": 4 + }, + { + "type": "image", + "img_path": "images/052155bee2775d35833066115b297323f753cce3a22d8fe57f655f5794fe4c81.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 323, + 619, + 446, + 723 + ], + "page_idx": 4 + }, + { + "type": "image", + "img_path": "images/9e2d0801addb462fc02b573b35ec15afd43788406b3ddcbf55c1a5652275a7ab.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 547, + 619, + 661, + 722 + ], + "page_idx": 4 + }, + { + "type": "image", + "img_path": "images/95779d7f8c4d99450912c29e23c3a409997915c022968df709dfd3b6abb7c648.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 745, + 619, + 875, + 723 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "datasets used for DUSt3R [36], excluding Waymo [31] due to its excessively sparse depth maps. The included datasets are Habitat [24], CO3D [21], ScanNet++ [39], ArkitScenes [1], Static Scenes 3D [17], MegaDepth [11], and BlendedMVS [38]. We follow the DUSt3R official repository [37]", + "bbox": [ + 89, + 820, + 480, + 896 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "guidelines to extract image pairs, ensuring reasonable overlap between paired images. To construct multi-view data given these pre-computed pairs, we iteratively sample images using an adjacency matrix, maintaining sufficient overlap with the selected set. Beyond CO3D, this model variant", + "bbox": [ + 511, + 821, + 906, + 896 + ], + "page_idx": 4 + }, + { + "type": "page_number", + "text": "6321", + "bbox": [ + 482, + 945, + 513, + 955 + ], + "page_idx": 4 + }, + { + "type": "image", + "img_path": "images/8113542d26e8415e7fccb43a4ab137fe288e9ddfb3159e473e2445238d6cf63d.jpg", + "image_caption": [ + "Input Images" + ], + "image_footnote": [], + "bbox": [ + 117, + 109, + 248, + 502 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/933b09eb7e3527150f3be3f259b770ea88c020ef8b967f7a83a4d6eafe070ecc.jpg", + "image_caption": [ + "DiffusionSfM" + ], + "image_footnote": [], + "bbox": [ + 258, + 119, + 424, + 204 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/ebb549fcdbcce39636c90442061595ecb9be761d0dac36f498c4259fa3e9299b.jpg", + "image_caption": [ + "RayDiffusion" + ], + "image_footnote": [], + "bbox": [ + 483, + 118, + 643, + 199 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/3d9b95e9ad7374161e5cbf61c41771e2e6b56aedb7c1c0789d6480c724ed3950.jpg", + "image_caption": [ + "DUSt3R" + ], + "image_footnote": [], + "bbox": [ + 702, + 119, + 867, + 203 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/c9e3727a7835cb21302b9686d7fc17c3b5fa3dc25d6a1046592ee7a962cf2cad.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 274, + 229, + 416, + 311 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/3ba083c2b95cf328bae57df050f1ef358611952760041f8e249fa443105d4deb.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 491, + 231, + 660, + 311 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/391548a854ffbf9d9e3b6c9a89f4c7f4a8c2303a99b0f5684284a04001ea7fd6.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 722, + 224, + 879, + 313 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/71982fd3c12d0817f39efa883fc6e901683dd4fcb9958618f0e0e0c14e24d0ec.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 116, + 325, + 248, + 393 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/01f965a7ac2be332782e20e3ebbaef854eed14bc732b6658975749ecb53a0bb9.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 263, + 325, + 419, + 398 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/3a0473971df79c5449389de83b9765f9583258dd642a4923b3125fc4447d8398.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 478, + 315, + 660, + 392 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/fc70de8b11828fc1a07930f61b93d9c100c1903b7fb41fcd5d39ea6065293654.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 702, + 327, + 883, + 404 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/e62c7dc2190fad5f4b7bdae298e3183cc133b3b08212eed26461b84525924ed5.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 116, + 398, + 248, + 467 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/12bc1a5e9df8ccb453a2c73ad7e51bab942f03cf2f0f7966ec16278b98e03951.jpg", + "image_caption": [ + "Figure 4. Additional Qualitative Results on Predicted Camera Poses. DiffusionSfM shows robustness to ambiguous patterns in inputs." + ], + "image_footnote": [], + "bbox": [ + 266, + 407, + 429, + 486 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/43d9c1f6b2fb34cfb80eb04d4a1eace850a5c28b1428f2923e0531d60d3fcbfe.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 483, + 409, + 651, + 484 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/0fb6fb31f2c126f48b00a94419cbada66bd471bb4696337619753970c8eb7ed5.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 707, + 409, + 870, + 484 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "is also evaluated on Habitat and RealEstate10k [42].", + "bbox": [ + 89, + 555, + 434, + 570 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Baselines and Metrics. To evaluate camera pose accuracy in the sparse-view setup, we compare with RayDiffusion [41] and DUSt3R [36], along with previous methods [12, 22, 34]. DUSt3R is trained initially on a mixture of eight datasets, while most other baselines are trained only on CO3D. For a fair and comprehensive comparison, we re-train DUSt3R on the 41-10 split of CO3D, using the authors' official implementation [37] and hyperparameters (referred to as DUSt3R-CO3D). To evaluate camera predictions, we follow prior work [41] and convert predicted rays back to traditional extrinsic matrices and report two pose accuracy metrics: (1) Camera Rotation Accuracy which compares the predicted relative camera rotation between images against ground truth and (2) Camera Center Accuracy which compares predicted camera centers to the ground truth after a similarity alignment. To evaluate the estimated geometry (i.e., ray endpoints), we report Chamfer Distance (CD). Additionally, to explore whether combining an off-the-shelf sparse-view pose estimation method (RayDiffusion) with a monocular depth estimation model (MoGe [35]) is sufficient to infer 3D scene geometry from", + "bbox": [ + 89, + 583, + 483, + 900 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "multiple images, we introduce a new baseline, $\\mathsf{RD} + \\mathsf{MoGe}$ . We include more details in the supplementary. Unless otherwise specified, all evaluations use 2-8 input images.", + "bbox": [ + 511, + 555, + 903, + 602 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "4.2. Evaluation on CO3D", + "text_level": 1, + "bbox": [ + 511, + 618, + 709, + 633 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Camera Pose Accuracy. We present the quantitative results in Tab. 1. For camera rotation accuracy, both versions of DiffusionSfM outperform all baselines, except that DUSt3R achieves slightly higher overall accuracy than DiffusionSfM-CO3D on unseen categories – likely due to its training on more data. For camera center accuracy, our approach consistently outperforms all other methods. We hypothesize that these gains stem from our explicit modeling of camera centers through ray origin prediction. The qualitative results in Fig. 3 illustrate that DiffusionSfM produces robust predictions given challenging inputs, whereas DUSt3R produces inaccurate results. We attribute this improvement to our model's probabilistic modeling capability derived from diffusion, as well as its multi-view reasoning abilities, which together effectively handle these challenging scenarios. While we observe that DUSt3R can predict highly precise camera rotations, it often struggles with cam-", + "bbox": [ + 509, + 643, + 906, + 900 + ], + "page_idx": 5 + }, + { + "type": "page_number", + "text": "6322", + "bbox": [ + 482, + 944, + 514, + 955 + ], + "page_idx": 5 + }, + { + "type": "table", + "img_path": "images/a33cb5d7197aa3e345a5d224916c27764df2b9ba8ce3e1bee23582ecf8870ce7.jpg", + "table_caption": [], + "table_footnote": [], + "table_body": "
# of ImagesRotation Accuracy (↑, @ 15°)Center Accuracy (↑, @ 0.1)
2345678234567
Seen CategoriesCOLMAP [22]30.728.426.526.827.028.130.610034.523.818.915.614.5
PoseDiffusion [34]75.776.476.877.478.078.778.810077.569.765.963.762.8
RelPose++ [12]81.882.884.184.784.985.385.510085.078.074.271.970.3
RayDiffusion [41]91.892.492.692.993.193.393.310094.290.587.886.285.0
DUSt3R-CO3D [36]86.787.988.088.288.688.888.910092.086.883.882.081.1
DUSt3R [36]91.792.793.393.693.894.094.310093.085.781.979.677.8
DiffusionSfM-CO3D93.494.094.594.895.095.295.110095.993.692.291.290.7
DiffusionSfM92.493.594.094.594.795.195.210095.392.991.991.290.8
Unseen CategoriesCOLMAP [22]34.531.831.031.732.735.038.510036.025.520.017.917.6
PoseDiffusion [34]63.264.264.265.766.267.067.710063.650.545.743.041.2
RelPose++ [12]69.871.171.972.873.874.474.910070.658.853.450.447.8
RayDiffusion [41]83.585.686.386.987.287.588.110087.781.177.074.172.4
DUSt3R-CO3D [36]79.881.582.682.783.083.383.710083.677.271.870.068.1
DUSt3R [36]90.892.693.693.693.893.693.410087.979.874.371.769.4
DiffusionSfM-CO3D90.491.292.793.093.193.393.510091.187.785.383.782.7
DiffusionSfM91.592.893.693.994.194.294.410092.487.986.285.384.1
", + "bbox": [ + 91, + 88, + 901, + 361 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "era centers (see the backpack example). Additional qualitative results on camera accuracy are provided in Fig.4.", + "bbox": [ + 88, + 460, + 482, + 491 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Predicted Geometry. To evaluate predicted geometry, we compute Chamfer Distance (CD) and show comparisons against baselines in Tab. 2. We compute CD in two setups (with and without foreground object mask), and find that DiffusionSfM-CO3D performs best without foreground masking. In this setup, the predicted ray endpoints corresponding to the background image pixels tend to have larger coordinate values than foreground ones, and therefore dominate CD. This result indicates that our model provides more accurate predictions for complex image backgrounds. In terms of CD with masking, DUSt3R achieves the best result, while our two model variants outperform DUSt3R-CO3D. See Fig. 3 for visualization. We also notice that naively combining RayDiffusion poses with estimated depths from MoGe does not work well, as the inferred depths are potentially inconsistent across different views, thus resulting in degraded performance.", + "bbox": [ + 91, + 493, + 482, + 750 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "4.3. Evaluation on Scene-Level Datasets", + "text_level": 1, + "bbox": [ + 89, + 768, + 398, + 786 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Beyond the object-centric CO3D dataset [21], we compare DiffusionSfM against DUSt3R on two scene-level datasets: Habitat (in-distribution) [24] and RealEstate10k (out-of-distribution) [42]. The results are presented in Tab. 3. While DiffusionSfM achieves comparable camera rotation accuracy to DUSt3R, it consistently predicts more accurate camera centers, aligning with our findings on CO3D.", + "bbox": [ + 89, + 794, + 482, + 902 + ], + "page_idx": 6 + }, + { + "type": "table", + "img_path": "images/c4ea6f4ebe82bd9ce2a893b5358e6625bbd993f45aeae01820fb06464e3d1fe8.jpg", + "table_caption": [ + "Table 1. Camera Rotation and Center Accuracy on CO3D. On the left, we report the proportion of relative camera rotations within $15^{\\circ}$ of the ground truth. On the right, we report the proportion of camera centers within $10\\%$ of the scene scale, relative to the ground truth. To align the predicted camera centers to ground truth, we apply an optimal similarity transform $(s, \\mathbf{R}, \\mathbf{t})$ , hence the alignment is perfect at $N = 2$ but worsens with more images. DiffusionSfM outperforms all other methods for camera center accuracy, and outperforms all methods trained on equivalent data for rotation accuracy." + ], + "table_footnote": [], + "table_body": "
# of Images2345678
RD*+MoGe [41]0.0590.0640.0710.0620.0630.0610.061
DUSt3R* [36]0.0360.0370.0400.0400.0370.0360.039
DUSt3R [36]0.0210.0230.0240.0240.0250.0250.023
DiffusionSfM*0.0190.0210.0220.0220.0210.0210.021
DiffusionSfM0.0250.0250.0250.0250.0260.0250.028
RD*+MoGe [41]0.0710.0750.0680.0670.0660.0640.064
DUSt3R* [36]0.0380.0360.0360.0360.0340.0330.034
DUSt3R [36]0.0230.0220.0190.0200.0190.0200.020
DiffusionSfM*0.0280.0250.0240.0240.0240.0230.022
DiffusionSfM0.0270.0240.0230.0220.0220.0220.021
", + "bbox": [ + 516, + 457, + 903, + 633 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Table 2. Chamfer Distance $(\\downarrow)$ on CO3D Unseen Categories. Top: CD computed on all scene points. Bottom: CD computed on foreground points only. Models marked with “*” are trained on CO3D only, while those without are trained on multiple datasets. Note that top and bottom values are not directly comparable, as each ground-truth point cloud is individually normalized.", + "bbox": [ + 511, + 643, + 906, + 729 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "4.4. Inference Speed", + "text_level": 1, + "bbox": [ + 511, + 756, + 671, + 773 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Though our method requires iterative diffusion denoising at inference, we can speed this up by performing early stopping. Specifically, we can treat the $x_0$ -prediction from early timesteps as our output instead of iterating over all denoising timesteps. Consistent with observations by Zhang et al. [41], this in fact yields more accurate predictions than the final-step diffusion outputs. As a result we only require 10 denoising diffusion timesteps for inference, taking 1.91", + "bbox": [ + 511, + 779, + 906, + 902 + ], + "page_idx": 6 + }, + { + "type": "page_number", + "text": "6323", + "bbox": [ + 482, + 944, + 514, + 955 + ], + "page_idx": 6 + }, + { + "type": "table", + "img_path": "images/a5b3d4e308ef80db7d45571ba6f502b036de286a0336817e442fa29c3ab997cf.jpg", + "table_caption": [], + "table_footnote": [], + "table_body": "
Hab.DUSt3R97.0/10095.0/97.694.3/95.094.2/93.1
DiffusionSfM92.6/10093.7/99.194.2/98.894.9/98.6
RE10DUSt3R98.1/10097.7/68.797.6/57.997.7/53.3
DiffusionSfM97.8/10097.5/75.697.7/66.297.9/61.9
", + "bbox": [ + 94, + 88, + 478, + 154 + ], + "page_idx": 7 + }, + { + "type": "table", + "img_path": "images/218890669fb73a4f9cd99bd9c950b2301f5873ae02015299bbe805c9375c2644.jpg", + "table_caption": [ + "Table 3. Camera Rotation and Center Accuracy on Two Scene-Level Datasets. Top: Habitat (2-5 views). Bottom: RealEstate10k (2, 4, 6, 8 views). Each grid reports camera rotation accuracy (left, $\\uparrow$ ) and center accuracy (right, $\\uparrow$ ). While DiffusionSfM performs on par with DUSt3R in rotation accuracy, it consistently surpasses DUSt3R in center accuracy." + ], + "table_footnote": [], + "table_body": "
# of Images2345678
DiffusionSfM*0.0190.0210.0220.0220.0210.0210.021
w/o Mask0.0200.0220.0230.0230.0220.0220.024
", + "bbox": [ + 91, + 265, + 482, + 324 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Table 4. Ablation Study on GT Mask Conditioning for CO3D Unseen Categories. We assess the effect of replacing GT mask conditioning with depth interpolation in our CO3D variant (DiffusionSfM*), by reporting the CD for predicted geometry. Incorporating mask conditioning to indicate missing data during training improves geometry quality.", + "bbox": [ + 89, + 334, + 482, + 419 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "seconds on a single A5000 GPU with 8 input images. In contrast, DUSt3R takes 8.73 seconds to run the complete pairwise inference and global alignment procedure. We provide additional analysis in the supplementary material.", + "bbox": [ + 89, + 450, + 483, + 511 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "4.5. Ablation Study", + "text_level": 1, + "bbox": [ + 89, + 526, + 243, + 542 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Homogeneous Coordinates. The use of homogeneous coordinates for ray origins and endpoints is crucial for stable model training. To assess its impact, we replace the proposed homogeneous representation with standard 3D coordinates in $\\mathbb{R}^3$ . Our experiments show that this variant is difficult to train and fails to converge. Further details on this experiment are provided in the supplementary material.", + "bbox": [ + 89, + 550, + 482, + 656 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "GT Mask Conditioning. To evaluate the effectiveness of the proposed GT mask conditioning, we train a baseline model without using this strategy. For missing values in the ground-truth depth maps, we use nearest-neighbor interpolation to fill in invalid pixels. This experiment is conducted on the CO3D dataset, with results presented in Tab. 4. The findings show that removing GT mask conditioning consistently degrades predicted geometry across varying numbers of input views, even when the loss is still masked. While interpolation can effectively fill missing depth within object regions, it often introduces substantial noise in the background (e.g., the sky). This noise negatively impacts diffusion model training, as the entire ray origin and endpoint maps are used as input. We anticipate an even larger performance gap on non-object-centric datasets with more outdoor scenes, such as MegaDepth [11].", + "bbox": [ + 89, + 657, + 482, + 900 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/73e85658c1c87772133020c3736a2822b53d306c91f74b826fc325128df9097e.jpg", + "image_caption": [ + "Figure 5. Multi-modality of DiffusionSfM. We show two distinct samples from DiffusionSfM, starting from the same input images but with different random noise. Sample 1 explains the input images by putting all flowers on the left side, while Sample 2 places one flower on each side (note the difference in the green camera's viewpoint). DiffusionSfM is able to predict multi-modal geometry distributions when the scene layout is ambiguous in the inputs." + ], + "image_footnote": [], + "bbox": [ + 516, + 87, + 908, + 179 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "4.6. Multi-modality from Multiple Sampling", + "text_level": 1, + "bbox": [ + 511, + 311, + 856, + 329 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Diffusion models enable the generation of diverse samples from challenging input images. For instance, in Fig. 5, the vase exhibits symmetric patterns, and we present two distinct predicted endpoints from DiffusionSfM, each offering a different interpretation of the flowers in the image. Compared to regression models such as DUSt3R [8], DiffusionSfM is better suited for handling uncertainty – an inherent aspect of our task – where a sparse set of input images can correspond to multiple plausible 3D scene geometry.", + "bbox": [ + 511, + 334, + 906, + 470 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "5. Discussion", + "text_level": 1, + "bbox": [ + 511, + 483, + 627, + 500 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "We present DiffusionSfM and demonstrate that it recovers accurate predictions of both cameras and geometry from multi-view inputs. Although our results are promising, several challenges and open questions remain.", + "bbox": [ + 511, + 508, + 905, + 569 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Notably, DiffusionSfM employs a pixel-space diffusion model, in contrast to the latent-space models adopted by state-of-the-art T2I generative systems. Operating in pixel space may require greater model capacity, yet our current model remains relatively small - potentially explaining the noisy patterns observed along object boundaries. Learning an expressive latent space for ray origins and endpoints by training a VAE could be a promising direction for future work. In addition, the compute requirement in multi-view transformers scales quadratically with the number of input images: one would require masked attention to deploy systems like ours for a large set of input images.", + "bbox": [ + 511, + 569, + 905, + 750 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Despite these challenges, we believe that our work highlights the potential of a unified approach for multi-view geometry tasks. We envision that our approach can be built upon to train a common system across related geometric tasks, such as SfM (input images with unknown origins and endpoints), registration (some images have known origins and endpoints, whereas others don't), mapping (known rays but unknown endpoints), and view synthesis (unknown pixel values for known rays).", + "bbox": [ + 511, + 750, + 908, + 887 + ], + "page_idx": 7 + }, + { + "type": "page_number", + "text": "6324", + "bbox": [ + 482, + 944, + 514, + 955 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Acknowledgements: We thank the members of the Physical Perception Lab at CMU for their valuable discussions, and extend special thanks to Yanbo Xu for his insights on diffusion models.", + "bbox": [ + 89, + 90, + 480, + 150 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "This work used Bridges-2 [4] at Pittsburgh Supercomputing Center through allocation CIS240166 from the Advanced Cyberinfrastructure Coordination Ecosystem: Services & Support (ACCESS) program, which is supported by National Science Foundation grants #2138259, #2138286, #2138307, #2137603, and #2138296. This work was supported by Intelligence Advanced Research Projects Activity (IARPA) via Department of Interior/Interior Business Center (DOI/IBC) contract number 140D0423C0074. The U.S. Government is authorized to reproduce and distribute reprints for Governmental purposes notwithstanding any copyright annotation thereon. Disclaimer: The views and conclusions contained herein are those of the authors and should not be interpreted as necessarily representing the official policies or endorsements, either expressed or implied, of IARPA, DOI/IBC, or the U.S. Government.", + "bbox": [ + 89, + 152, + 482, + 393 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "References", + "text_level": 1, + "bbox": [ + 91, + 404, + 187, + 419 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[1] Gilad Baruch, Zhuoyuan Chen, Afshin Dehghan, Tal Dimry, Yuri Feigin, Peter Fu, Thomas Gebauer, Brandon Joffe, Daniel Kurz, Arik Schwartz, et al. Arkitsscenes: A diverse real-world dataset for 3d indoor scene understanding using mobile rgb-d data. In NeurIPS D&B, 2021. 5", + "[2] Herbert Bay, Andreas Ess, Tinne Tuytelaars, and Luc Van Gool. Surf: Speeded up robust features. In ECCV, 2006. 2", + "[3] Eric Brachmann, Jamie Wynn, Shuai Chen, Tommaso Cavallari, Aron Monszpart, Daniyar Turmukhambetov, and Victor Adrian Prisacariu. Scene coordinate reconstruction: posing of image collections via incremental learning of a relocalizer. In ECCV, 2024. 2", + "[4] Shawn T Brown, Paola Buitrago, Edward Hanna, Sergiu Sanielevici, Robin Scibek, and Nicholas A Nystrom. Bridges-2: A platform for rapidly-evolving and data intensive research. In Practice and experience in advanced research computing, 2021. 9", + "[5] Ruojin Cai, Joseph Tung, Qianqian Wang, Hadar Averbuch-Elor, Bharath Hariharan, and Noah Snavely. Doppelgangers: Learning to disambiguate images of similar structures. In ICCV, 2023. 1", + "[6] Daniel DeTone, Tomasz Malisiewicz, and Andrew Rabinovich. Superpoint: Self-supervised interest point detection and description. In CVPR, 2018. 1, 2", + "[7] Bardienus Duisterhof, Lojze Zust, Philippe Weinzaepfel, Vincent Leroy, Yohann Cabon, and Jerome Revaud. Mast3r-sfm: a fully-integrated solution for unconstrained structure-from-motion. In 3DV, 2025. 2", + "[8] Johan Edstedt, Qiyu Sun, Georg Bokman, Marten Wadenback, and Michael Felsberg. RoMa: Robust Dense Feature Matching. In CVPR, 2024. 2, 8", + "[9] Jonathan Ho, Ajay Jain, and Pieter Abbeel. Denoising diffusion probabilistic models. In NeurIPS, 2020. 2, 3" + ], + "bbox": [ + 99, + 429, + 482, + 900 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[10] Vincent Leroy, Yohann Cabon, and Jérôme Revaud. Grounding image matching in 3d with mast3r, 2024. ECCV. 1, 2", + "[11] Zhengqi Li and Noah Snavely. Megadepth: Learning single-view depth prediction from internet photos. In CVPR, 2018. 2,4,5,8", + "[12] Amy Lin, Jason Y Zhang, Deva Ramanan, and Shubham Tulsiani. Relpose++: Recovering 6d poses from sparse-view observations. In 3DV, 2024. 2, 4, 6, 7", + "[13] Philipp Lindenberger, Paul-Edouard Sarlin, Viktor Larsson, and Marc Pollefeys. Pixel-perfect structure-from-motion with featuremetric refinement. In ICCV, 2021. 2", + "[14] Philipp Lindenberger, Paul-Edouard Sarlin, Viktor Larsson, and Marc Pollefeys. Pixel-Perfect Structure-from-Motion with Featuremetric Refinement. In ICCV, 2021. 1", + "[15] Philipp Lindenberger, Paul-Edouard Sarlin, and Marc Pollefeys. Lightglue: Local feature matching at light speed. In ICCV, 2023. 2", + "[16] David G Lowe. Distinctive image features from scale-invariant keypoints. In IJCV, 2004. 2", + "[17] Nikolaus Mayer, Eddy Ilg, Philip Hausser, Philipp Fischer, Daniel Cremers, Alexey Dosovitskiy, and Thomas Brox. A large dataset to train convolutional networks for disparity, optical flow, and scene flow estimation. In CVPR, 2016. 5", + "[18] Maxime Oquab, Timothee Darcet, Theo Moutakanni, Huy Vo, Marc Szafraniec, Vasil Khalidov, Pierre Fernandez, Daniel Haziza, Francisco Massa, Alaaeldin El-Nouby, et al. Dinov2: Learning robust visual features without supervision. In TMLR, 2024. 3, 4", + "[19] William Peebles and Saining Xie. Scalable diffusion models with transformers. In ICCV, 2023. 2, 3, 4", + "[20] René Ranftl, Alexey Bochkovskiy, and Vladlen Koltun. Vision transformers for dense prediction. In ICCV, 2021. 3, 4", + "[21] Jeremy Reizenstein, Roman Shapovalov, Philipp Henzler, Luca Sbordone, Patrick Labatut, and David Novotny. Common objects in 3d: Large-scale learning and evaluation of real-life 3d category reconstruction. In ICCV, 2021. 2, 4, 5, 7", + "[22] Paul-Edouard Sarlin, Cesar Cadena, Roland Siegwart, and Marcin Dymczyk. From coarse to fine: Robust hierarchical localization at large scale. In CVPR, 2019. 6, 7", + "[23] Paul-Edouard Sarlin, Daniel DeTone, Tomasz Malisiewicz, and Andrew Rabinovich. Superglue: Learning feature matching with graph neural networks. In CVPR, 2020. 1, 2", + "[24] Manolis Savva, Abhishek Kadian, Oleksandr Maksymets, Yili Zhao, Erik Wijmans, Bhavana Jain, Julian Straub, Jia Liu, Vladlen Koltun, Jitendra Malik, et al. Habitat: A platform for embodied ai research. In ICCV, 2019. 2, 5, 7", + "[25] Johannes L Schonberger and Jan-Michael Frahm. Structure-from-motion revisited. In CVPR, 2016. 1, 2", + "[26] Samarth Sinha, Jason Y Zhang, Andrea Tagliasacchi, Igor Gilitschenski, and David B Lindell. Sparsepose: Sparse-view camera pose regression and refinement. In CVPR, 2023. 2", + "[27] Cameron Smith, David Charatan, Ayush Tewari, and Vincent Sitzmann. Flowmap: High-quality camera poses, intrinsics, and depth via gradient descent. In 3DV, 2025. 2" + ], + "bbox": [ + 516, + 92, + 903, + 900 + ], + "page_idx": 8 + }, + { + "type": "page_number", + "text": "6325", + "bbox": [ + 482, + 944, + 514, + 955 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[28] Noah Snavely, Steven M Seitz, and Richard Szeliski. Photo tourism: exploring photo collections in 3d. In ACM SIGGRAPH, 2006. 2", + "[29] Jiaming Song, Chenlin Meng, and Stefano Ermon. Denoising diffusion implicit models. In ICLR, 2021. 3", + "[30] Jiaming Sun, Zehong Shen, Yuang Wang, Hujun Bao, and Xiaowei Zhou. Loftr: Detector-free local feature matching with transformers. In CVPR, 2021. 2", + "[31] Pei Sun, Henrik Kretzschmar, Xerxes Dotiwalla, Aurelien Chouard, Vijaysai Patnaik, Paul Tsui, James Guo, Yin Zhou, Yuning Chai, Benjamin Caine, et al. Scalability in perception for autonomous driving: Waymo open dataset. In CVPR, 2020. 5", + "[32] Chengzhou Tang and Ping Tan. Ba-net: Dense bundle adjustment network. In ICLR, 2019. 2", + "[33] Ashish Vaswani, Noam Shazeer, Niki Parmar, Jakob Uszkoreit, Llion Jones, Aidan N Gomez, Lukasz Kaiser, and Illia Polosukhin. Attention is all you need. In NeurIPS, 2017. 4", + "[34] Jianyuan Wang, Christian Rupprecht, and David Novotny. Posediffusion: Solving pose estimation via diffusion-aided bundle adjustment. In ICCV, 2023. 2, 6, 7", + "[35] Ruicheng Wang, Sicheng Xu, Cassie Dai, Jianfeng Xiang, Yu Deng, Xin Tong, and Jiaolong Yang. Moge: Unlocking accurate monocular geometry estimation for open-domain images with optimal training supervision. In CVPR, 2025. 6", + "[36] Shuzhe Wang, Vincent Leroy, Yohann Cabon, Boris Chidlovskii, and Jerome Revaud. Dust3r: Geometric 3d vision made easy. In CVPR, 2024. 1, 2, 4, 5, 6, 7", + "[37] Shuzhe Wang, Vincent Leroy, Yohann Cabon, Boris Chidlovskii, and Jerome Revaud. Dust3r: Geometric 3d vision made easy, 2024. https://github.com/naver/dust3r. 5, 6", + "[38] Yao Yao, Zixin Luo, Shiwei Li, Jingyang Zhang, Yufan Ren, Lei Zhou, Tian Fang, and Long Quan. Blendedmvs: A large-scale dataset for generalized multi-view stereo networks. In CVPR, 2020. 5", + "[39] Chandan Yeshwanth, Yueh-Cheng Liu, Matthias Nießner, and Angela Dai. Scannet++: A high-fidelity dataset of 3d indoor scenes. In ICCV, 2023. 5", + "[40] Jason Y Zhang, Deva Ramanan, and Shubham Tulsiani. Relapse: Predicting probabilistic relative rotation for single objects in the wild. In ECCV, 2022. 2", + "[41] Jason Y. Zhang, Amy Lin, Moneish Kumar, Tzu-Hsuan Yang, Deva Ramanan, and Shubham Tulsiani. Cameras as rays: Sparse-view pose estimation via ray diffusion. In ICLR, 2024. 2, 3, 4, 6, 7", + "[42] Tinghui Zhou, Richard Tucker, John Flynn, Graham Fyffe, and Noah Snavely. Stereo magnification: Learning view synthesis using multiplane images. In ACM SIGGRAPH, 2018. 2, 6, 7" + ], + "bbox": [ + 91, + 92, + 482, + 814 + ], + "page_idx": 9 + }, + { + "type": "page_number", + "text": "6326", + "bbox": [ + 482, + 945, + 514, + 955 + ], + "page_idx": 9 + } +] \ No newline at end of file diff --git a/2025/DiffusionSfM_ Predicting Structure and Motion via Ray Origin and Endpoint Diffusion/a0e804a0-3c59-43c5-bb05-b2160d8636bd_model.json b/2025/DiffusionSfM_ Predicting Structure and Motion via Ray Origin and Endpoint Diffusion/a0e804a0-3c59-43c5-bb05-b2160d8636bd_model.json new file mode 100644 index 0000000000000000000000000000000000000000..19bc9323eeca5c963306011e97ef3e37d37eb13f --- /dev/null +++ b/2025/DiffusionSfM_ Predicting Structure and Motion via Ray Origin and Endpoint Diffusion/a0e804a0-3c59-43c5-bb05-b2160d8636bd_model.json @@ -0,0 +1,2266 @@ +[ + [ + { + "type": "header", + "bbox": [ + 0.107, + 0.003, + 0.182, + 0.043 + ], + "angle": 0, + "content": "CVF" + }, + { + "type": "header", + "bbox": [ + 0.237, + 0.0, + 0.812, + 0.047 + ], + "angle": 0, + "content": "This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore." + }, + { + "type": "title", + "bbox": [ + 0.259, + 0.131, + 0.74, + 0.177 + ], + "angle": 0, + "content": "DiffusionSfM: Predicting Structure and Motion via Ray Origin and Endpoint Diffusion" + }, + { + "type": "text", + "bbox": [ + 0.17, + 0.204, + 0.828, + 0.24 + ], + "angle": 0, + "content": "Qitao Zhao, Amy Lin, Jeff Tan, Jason Y. Zhang, Deva Ramanan, Shubham Tulsiani Carnegie Mellon University" + }, + { + "type": "text", + "bbox": [ + 0.308, + 0.24, + 0.69, + 0.258 + ], + "angle": 0, + "content": "Project page: qitaozhao.github.io/DiffusionSfM" + }, + { + "type": "image", + "bbox": [ + 0.092, + 0.27, + 0.907, + 0.499 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.089, + 0.508, + 0.908, + 0.581 + ], + "angle": 0, + "content": "Figure 1. DiffusionSfM. Top: Given a set of multi-view images (left), DiffusionSfM represents scene geometry and cameras (right) as pixel-wise ray origins and endpoints in a global frame. It learns a denoising diffusion model to infer these elements directly from multiview inputs. Unlike traditional Structure-from-Motion (SfM) pipelines, which separate pairwise reasoning and global optimization into two stages, our approach unifies both into a single end-to-end multi-view reasoning framework. Bottom: Example results of inferred scene geometry and cameras for two distinct settings: a real-world outdoor scene (left) and a synthetic indoor scene (right)." + }, + { + "type": "title", + "bbox": [ + 0.249, + 0.598, + 0.327, + 0.614 + ], + "angle": 0, + "content": "Abstract" + }, + { + "type": "text", + "bbox": [ + 0.089, + 0.643, + 0.486, + 0.886 + ], + "angle": 0, + "content": "Current Structure-from-Motion (SfM) methods typically follow a two-stage pipeline, combining learned or geometric pairwise reasoning with a subsequent global optimization step. In contrast, we propose a data-driven multi-view reasoning approach that directly infers 3D scene geometry and camera poses from multi-view images. Our framework, DiffusionSfM, parameterizes scene geometry and cameras as pixel-wise ray origins and endpoints in a global frame and employs a transformer-based denoising diffusion model to predict them from multi-view inputs. To address practical challenges in training diffusion models with missing data and unbounded scene coordinates, we introduce specialized mechanisms that ensure robust learning. We empirically validate DiffusionSfM on both synthetic and real datasets, demonstrating that it outperforms classical and learning-based approaches while naturally modeling uncertainty." + }, + { + "type": "title", + "bbox": [ + 0.515, + 0.598, + 0.645, + 0.614 + ], + "angle": 0, + "content": "1. Introduction" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.63, + 0.909, + 0.904 + ], + "angle": 0, + "content": "The task of recovering structure (geometry) and motion (cameras) from multi-view images has long been a focus of the computer vision community, with typical pipelines [25] performing pairwise correspondence estimation followed by global optimization. While classical methods relied on hand-designed features, matching, and optimization, there has been a recent shift towards incorporating learning-based alternatives [5, 6, 14, 23]. More recently, the widely influential DUSt3R [36] advocates for predicting pairwise 3D pointmaps (instead of only correspondences), demonstrating that this can yield accurate dense geometry and cameras. In order to reconstruct more than two views, DUSt3R (and its variants [10]) still require a global optimization reminiscent of classic bundle adjustment. While these methods, both classical and learning-based, have led to impressive improvements in SfM, the overall approach is largely unchanged – learned or geometric pairwise reasoning followed by global optimization. In this work, we" + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.516, + 0.957 + ], + "angle": 0, + "content": "6317" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.09, + 0.092, + 0.482, + 0.137 + ], + "angle": 0, + "content": "seek to develop an alternative approach that directly predicts both structure and motion, unifying pairwise reasoning and global optimization into a single multi-view framework." + }, + { + "type": "text", + "bbox": [ + 0.089, + 0.14, + 0.483, + 0.366 + ], + "angle": 0, + "content": "We are of course not the first to attempt to find unified alternatives to the two-stage SfM pipeline. In the sparse-view setting where conventional correspondence-based methods struggle, several works employ multi-view architectures to jointly reason across input images. SparsePose [26], RelPose++ [12], and PoseDiffusion [34] all leverage multiview transformers to estimate camera pose for input images, albeit using differing mechanisms such as regression, energy-based modeling, and denoising diffusion. More recently, RayDiffusion [41] argues for a local raymap parameterization of cameras instead of a global extrinsic matrix and shows that existing patch-based transformers can be easily adapted for this task, yielding significantly more accurate pose predictions. Importantly, such methods predict only camera motion and fail to predict scene structure." + }, + { + "type": "text", + "bbox": [ + 0.093, + 0.37, + 0.483, + 0.777 + ], + "angle": 0, + "content": "In this work, we present DiffusionSfM, an end-to-end multi-view model that directly infers dense 3D geometry and cameras from multiple input images. Instead of inferring (depth-agnostic) rays per image patch (as in RayDiffusion [41]) or 3D points per pixel (as in DUSt3R [36]), DiffusionSfM effectively combines both to predict ray origins and endpoints per pixel, directly reporting both scene geometry (endpoints) and generalized cameras (rays). These can readily be converted back to traditional cameras [41]. Compared to RayDiffusion, our model directly predicts structure as well as motion at a finer scale (pixel-wise v.s. patchwise). Compared to DUSt3R, our model directly predicts motion as well as structure but, even more importantly, does so for \\(N\\) views, eliminating the need for memory-intensive global alignment. To model uncertainty, we train a denoising diffusion model but find two key challenges that need to be addressed. First, diffusion models require (noisy) ground truth as input for training, but existing real datasets do not have known endpoints for all pixels due to missing depth in multi-view stereo. Second, the 3D coordinates of endpoints can be potentially unbounded, whereas diffusion models require normalized data. We develop mechanisms to overcome these challenges, leveraging additional \"GT mask conditioning\" as input to inform the model of missing input data and parameterizing 3D points in projective space instead of Euclidean space. We find these strategies allow us to learn accurate predictions for structure and motion." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.78, + 0.483, + 0.901 + ], + "angle": 0, + "content": "We train and evaluate DiffusionSfM on real-world and synthetic datasets [21, 24, 42] and find that it can infer accurate geometry and cameras for both object-centric and scene-level images (see Fig. 1). In particular, we find that DiffusionSfM yields more accurate camera estimates compared to prior work across these settings while also modeling the underlying uncertainty via the diffusion process. In summary, we show that DiffusionSfM can serve as a unified" + }, + { + "type": "text", + "bbox": [ + 0.514, + 0.092, + 0.902, + 0.107 + ], + "angle": 0, + "content": "multi-view reasoning model for 3D geometry and cameras." + }, + { + "type": "title", + "bbox": [ + 0.514, + 0.122, + 0.655, + 0.138 + ], + "angle": 0, + "content": "2. Related Work" + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.148, + 0.905, + 0.314 + ], + "angle": 0, + "content": "Structure from Motion. Structure-from-Motion (SfM) systems [25] aim to simultaneously recover geometry and cameras given a set of input images. The typical SfM pipeline extracts pixel correspondences from keypoint matching [2, 16], and performs global bundle adjustment (BA) to optimize sparse 3D points and camera parameters by minimizing reprojection errors. Recently, SfM pipelines have been substantially enhanced by replacing classical subcomponents with learning-based methods, such as neural feature descriptors [6, 8], keypoint matching [15, 23, 30], and bundle adjustment [13, 32]." + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.315, + 0.905, + 0.6 + ], + "angle": 0, + "content": "More recently, an emerging body of research aims to unify the various SfM subcomponents into an end-to-end neural framework. Notably, ACEZero [3] fits a single neural network to input images and learns pixel-aligned 3D coordinates in a self-supervised manner, while FlowMap [27] predicts per-frame cameras and depth maps using off-the-shelf optical flow as supervision. Though ACEZero and FlowMap are promising attempts to revolutionize SfM pipelines, they both register images incrementally and may suffer under large viewpoint changes. DUSt3R [36] directly regresses 3D pointmaps from image pairs and shows strong generalization ability [11, 21]. Building on DUSt3R, MASt3R [10] introduces a feature head, offering pixel-matching capabilities. MASt3R-SfM [7] is a more scalable SfM pipeline based on MASt3R. While these approaches [7, 10, 36] show impressive performance and robustness under sparse views, they are essentially pair-based, requiring sophisticated global alignment procedures to form a consistent estimate for more than two views." + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.603, + 0.905, + 0.83 + ], + "angle": 0, + "content": "Pose Estimation with Global Reasoning. For the task of sparse-view pose estimation, learning-based methods equipped with global reasoning show favorable robustness where traditional SfM methods [25, 28] fail. This line of research includes energy-based [12, 40], regression-based [26], and diffusion-based pose estimators [34, 41]. Among them, diffusion-based methods show a better ability to handle uncertainty [41]. Closest to our work, RayDiffusion [41] leverages a denoising diffusion model [9, 19] with a patch-aligned (depth-unaware) ray representation to predict generic cameras. Our method goes further and pursues a generic representation for both geometry and cameras in the form of ray origins and endpoints for each pixel. In addition to resulting in a richer output, this joint geometry and pose prediction also yields improvements for pose estimation." + }, + { + "type": "title", + "bbox": [ + 0.514, + 0.845, + 0.605, + 0.86 + ], + "angle": 0, + "content": "3. Method" + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.871, + 0.905, + 0.901 + ], + "angle": 0, + "content": "Given a set of sparse (i.e., 2-8) input images, DiffusionSfM predicts the geometry and cameras of a 3D scene in a global" + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.946, + 0.516, + 0.957 + ], + "angle": 0, + "content": "6318" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.095, + 0.089, + 0.91, + 0.216 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.089, + 0.226, + 0.907, + 0.337 + ], + "angle": 0, + "content": "Figure 2. Method. Given sparse multi-view images as input, DiffusionSfM predicts pixel-wise ray origins and endpoints in a global frame (Sec. 3.1) using a denoising diffusion process (Sec. 3.2). For each image, we compute patch-wise embeddings with DINOv2 [18] and embed noisy ray origins and endpoints into latents using a single downsampling convolutional layer, ensuring alignment with the spatial footprint of the image embeddings. We implement a Diffusion Transformer architecture that predicts clean ray origins and endpoints from noisy samples. A convolutional DPT [20] head outputs full-resolution denoised ray origins and endpoints. To handle incomplete ground truth (GT) during training, we condition the model on a GT mask (Sec. 3.3). At inference, the GT mask is set to all ones, enabling the model to predict origins and endpoints for all pixels. The predicted ray origins and endpoints can be directly visualized in 3D or postprocessed to recover camera extrinsics, intrinsics, and multi-view consistent depth maps." + }, + { + "type": "text", + "bbox": [ + 0.089, + 0.364, + 0.484, + 0.455 + ], + "angle": 0, + "content": "coordinate frame. In Sec. 3.1, we propose to represent 3D scenes as dense pixel-aligned ray origins and endpoints. To predict such scene representations from sparse input images while modeling uncertainty, Sec. 3.2 proposes a denoising diffusion architecture. We then discuss some key practical challenges in training such a model in Sec. 3.3." + }, + { + "type": "title", + "bbox": [ + 0.09, + 0.463, + 0.441, + 0.479 + ], + "angle": 0, + "content": "3.1. 3D Scenes as Ray Origins and Endpoints" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.483, + 0.484, + 0.621 + ], + "angle": 0, + "content": "Given an input image \\(\\mathbf{I} \\in \\mathbb{R}^{H \\times W \\times 3}\\) with a depth map \\(\\mathbf{D} \\in \\mathbb{R}^{H \\times W}\\), camera intrinsics \\(\\mathbf{K} \\in \\mathbb{R}^{3 \\times 3}\\), and world-to-camera extrinsics \\(\\mathbf{T} \\in \\mathbb{R}^{4 \\times 4}\\) (equivalently, rotation \\(\\mathbf{R} \\in SO(3)\\) and translation \\(\\mathbf{t} \\in \\mathbb{R}^3\\)), each 2D image pixel \\(\\mathbf{P}_{ij} = [u, v]\\) corresponds to a ray that travels from the camera center \\(\\mathbf{c}\\) through the pixel's projected position on the image plane, terminating at the object's surface as specified by the depth map \\(\\mathbf{D}\\). The endpoint of the ray associated with image pixel \\(\\mathbf{P}_{ij}\\) is given by:" + }, + { + "type": "equation", + "bbox": [ + 0.167, + 0.628, + 0.482, + 0.647 + ], + "angle": 0, + "content": "\\[\n\\mathbf {E} _ {i j} = \\mathbf {T} ^ {- 1} h (\\mathbf {D} _ {i j} \\cdot \\mathbf {K} ^ {- 1} [ u, v, 1 ] ^ {T}) \\tag {1}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.655, + 0.483, + 0.7 + ], + "angle": 0, + "content": "where \\( h \\) maps the 3D point into homogeneous coordinates. The shared ray origin \\( \\mathbf{O}_{ij} \\) for all pixels is equivalent to the camera center \\( \\mathbf{c} \\), and can be computed as:" + }, + { + "type": "equation", + "bbox": [ + 0.205, + 0.709, + 0.482, + 0.727 + ], + "angle": 0, + "content": "\\[\n\\mathbf {O} _ {i j} = \\mathbf {c} = h (- \\mathbf {R} ^ {- 1} \\mathbf {t}) \\tag {2}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.735, + 0.483, + 0.825 + ], + "angle": 0, + "content": "In summary, we associate each image pixel with a ray origin and endpoint \\(\\mathbf{S}_{ij} = \\langle \\mathbf{O}_{ij},\\mathbf{E}_{ij}\\rangle\\) in world coordinates, describing the location of the observing camera and the observed 3D point on the object surface. Given a bundle of ray origins and endpoints, we can easily extract the corresponding camera pose [41]." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.826, + 0.484, + 0.901 + ], + "angle": 0, + "content": "Learning Over-Parameterized Representations. We represent 3D scenes and cameras using distributed ray origins \\(\\mathbf{O}\\) and endpoints \\(\\mathbf{E}\\) rather than global alternatives such as quaternions and translation vectors. This design is inspired by RayDiffusion [41], and it facilitates the use of the" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.364, + 0.907, + 0.501 + ], + "angle": 0, + "content": "distributed deep features learned by state-of-the-art vision backbones, such as DINOv2 [18], which encode image information in a patch-wise manner. Notably, while ray origins \\(\\mathbf{O}\\) should ideally be identical across all pixels, we predict them densely alongside ray endpoints \\(\\mathbf{E}\\). This encourages ray origins to remain close within the same image, providing implicit regularization during training. Practically, predicting both ray origins and endpoints simultaneously is easy to implement using a single projection head." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.508, + 0.655, + 0.523 + ], + "angle": 0, + "content": "3.2. DiffusionSfM" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.53, + 0.906, + 0.59 + ], + "angle": 0, + "content": "We propose a denoising Diffusion Transformer (DiT) architecture [19] that predicts ray origins and endpoints (Sec. 3.1) via a denoising diffusion process. An overview of DiffusionSfM is given in Fig. 2." + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.591, + 0.907, + 0.711 + ], + "angle": 0, + "content": "Diffusion Framework. Given pixel-aligned ray origins and endpoints \\( S = \\text{stack}(\\{\\mathbf{S}^{(n)}\\}_{n=1}^{N}) \\) associated with a set of \\( N \\) input images, we apply a forward diffusion process [9, 29] that adds time-dependent Gaussian noise to them. Let \\( S_{t} \\) denote the noisy ray origins and endpoints at timestep \\( t \\), where \\( S_{0} \\) is the clean sample and \\( S_{T} \\) (at the final diffusion step \\( T \\)) approximates pure Gaussian noise. The forward diffusion process is defined as:" + }, + { + "type": "equation", + "bbox": [ + 0.619, + 0.722, + 0.905, + 0.739 + ], + "angle": 0, + "content": "\\[\n\\mathcal {S} _ {t} = \\sqrt {\\bar {\\alpha} _ {t}} \\mathcal {S} _ {0} + \\sqrt {1 - \\bar {\\alpha} _ {t}} \\epsilon \\tag {3}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.75, + 0.907, + 0.871 + ], + "angle": 0, + "content": "where \\(t \\sim \\mathrm{Uniform}(0, T]\\), \\(\\epsilon \\sim \\mathcal{N}(0, I)\\), and \\(\\bar{\\alpha}_t\\) follows a pre-defined noise schedule that controls the strength of added noise at each timestep. To perform the reverse diffusion process, which progressively reconstructs the clean sample given noisy observations, we train a diffusion model \\(f_\\theta\\) that takes \\(\\mathcal{S}_t\\) as input and optionally incorporates additional conditioning information \\(\\mathcal{C}\\). The model is trained using the following loss function (with the \" \\(x_0 \\)\" objective):" + }, + { + "type": "equation", + "bbox": [ + 0.579, + 0.885, + 0.905, + 0.902 + ], + "angle": 0, + "content": "\\[\n\\mathcal {L} _ {\\text {D i f f u s i o n}} = \\mathbb {E} _ {t, S _ {0}, \\epsilon} \\| S _ {0} - f _ {\\theta} (S _ {t}, t, \\mathcal {C}) \\| ^ {2} \\tag {4}\n\\]" + }, + { + "type": "page_number", + "bbox": [ + 0.482, + 0.945, + 0.516, + 0.957 + ], + "angle": 0, + "content": "6319" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.09, + 0.092, + 0.482, + 0.227 + ], + "angle": 0, + "content": "Architecture. We implement \\( f_{\\theta} \\) using a DiT [19] architecture conditioned on deep image features \\( \\mathcal{C} \\in \\mathbb{R}^{N \\times h \\times w \\times c_1} \\) from DINOv2 [18], where \\( h \\) and \\( w \\) are the patch resolution and \\( c_1 \\) is the embedding dimension. To align pixels to the spatial information learned by DINOv2, we apply a convolutional layer with kernel size and stride equal to ViT patch size: this spatially downsamples the noisy ray origins and endpoints \\( S_t \\) to match the DINOv2 features while increasing their feature dimension:" + }, + { + "type": "equation", + "bbox": [ + 0.18, + 0.235, + 0.483, + 0.253 + ], + "angle": 0, + "content": "\\[\n\\mathcal {F} = \\operatorname {C o n v} \\left(\\mathcal {S} _ {t}\\right) \\in \\mathbb {R} ^ {N \\times h \\times w \\times c _ {2}} \\tag {5}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.261, + 0.483, + 0.351 + ], + "angle": 0, + "content": "The combined DiT input is constructed by concatenating these two feature sets along the channel dimension: \\(\\mathcal{F} \\oplus \\mathcal{C}\\). Within DiT, patch-wise features attend to others through self-attention [33]. To distinguish between different images and their respective patches, we apply sinusoidal positional encoding [33] based on image and patch indices." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.352, + 0.484, + 0.518 + ], + "angle": 0, + "content": "While the DiT operates on low-resolution features, our objective is to produce pixel-aligned dense ray origins and endpoints. To achieve this, we employ a DPT (Dense Prediction Transformer) [20] decoder, which takes intermediate feature maps from both DINoV2 and DiT as inputs. The DPT decoder progressively increases the feature resolution through several convolutional layers. The final ray origins and endpoints are decoded from the DPT output using a single linear layer. During inference, we apply the trained model in the reverse diffusion process to iteratively denoise a randomly initialized Gaussian sample." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.526, + 0.388, + 0.541 + ], + "angle": 0, + "content": "3.3. Practical Training Considerations" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.548, + 0.483, + 0.789 + ], + "angle": 0, + "content": "Homogeneous Coordinates for Unbounded Geometry. Real-world scenes often exhibit significant variations in scale, both across different scenes and within a single scene. While normalizing by the average distance of 3D points [36] helps address cross-scene variation, within-scene variations remain. For instance, a background building may be much farther away than foreground elements, potentially resulting in extremely large coordinate values when generating ground-truth ray origins and endpoints. However, neural networks (especially diffusion models) tend to train most effectively when working with bounded inputs and outputs (e.g., between -1 and 1). To stabilize training across the large scale variations present in 3D scene datasets, we propose to represent ray origins and endpoints in homogeneous coordinates. Specifically, given any 3D point, we apply a homogeneous transform from \\(\\mathbb{R}^3\\to \\mathbb{P}^3\\):" + }, + { + "type": "equation", + "bbox": [ + 0.203, + 0.797, + 0.483, + 0.825 + ], + "angle": 0, + "content": "\\[\n(x, y, z) \\rightarrow \\frac {1}{w} (x, y, z, 1) \\tag {6}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.832, + 0.484, + 0.876 + ], + "angle": 0, + "content": "where \\( w \\) is an arbitrary scale factor. To encourage bounded coordinates in practice, we choose \\( w \\) such that the homogeneous coordinate is unit-norm:" + }, + { + "type": "equation", + "bbox": [ + 0.199, + 0.884, + 0.483, + 0.902 + ], + "angle": 0, + "content": "\\[\nw := \\sqrt {x ^ {2} + y ^ {2} + z ^ {2} + 1} \\tag {7}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.092, + 0.905, + 0.168 + ], + "angle": 0, + "content": "Unit normalization allows homogeneous coordinates to serve as a bounded representation for unbounded scene geometry. For example, \\((x,y,z,0)\\) is a point at infinity in the direction of \\((x,y,z)\\). We find that this representation makes large coordinate values more tractable during training." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.168, + 0.905, + 0.349 + ], + "angle": 0, + "content": "Training with Incomplete Ground Truth. Many real-world datasets (e.g., CO3D [21] and MegaDepth [11]), provide only sparse point clouds, leading to incomplete depth information. This presents a significant challenge in diffusion training, as ground-truth (GT) depth values that are used to create clean ray endpoints often contain invalid or missing data. It is highly undesirable for these missing ray endpoints to be interpreted as part of the target distribution. Unlike regression models (e.g., DUSt3R [36]), which do not require (noisy) GT data as input and can simply apply a mask on the loss for supervision, diffusion models must handle incomplete inputs during training." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.351, + 0.906, + 0.516 + ], + "angle": 0, + "content": "To mitigate this issue, we further apply GT masks \\(\\mathcal{M} \\in \\mathbb{R}^{N \\times H \\times W}\\) to the DiT inputs, where zero values indicate pixels with invalid depth. During training, we multiply DiT inputs with GT masks element-wise, then concatenate along the channel dimension: \\(S_{t}' = (\\mathcal{M} \\cdot S_{t}) \\oplus \\mathcal{M}\\). Then, we only compute the diffusion loss in Eq. 4 over unmasked pixels. By implementing these strategies, we encourage the model to focus on regions with valid GT values during training. During inference, however, we would like the diffusion process to estimate ray endpoints at all pixels, so we always use GT masks with values set to one." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.518, + 0.906, + 0.715 + ], + "angle": 0, + "content": "Sparse-to-Dense Training. In practice, we find that training the entire model from scratch leads to slow convergence and suboptimal performance. To address this, we propose a sparse-to-dense training approach. First, we train a sparse version of the model, where the DPT decoder is removed, and the output ray origins and endpoints have the same spatial resolution as the DINOv2 features. Unlike Eq. 5, no spatial downsampling is required, so this sparse model uses a single linear layer to embed the noisy ray origins and endpoints. Once the sparse model is trained, we initialize the dense model DiT with the learned weights from the sparse model. This two-stage approach significantly improves performance; see the supplementary material for comparisons." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.731, + 0.645, + 0.748 + ], + "angle": 0, + "content": "4. Experiments" + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.757, + 0.704, + 0.773 + ], + "angle": 0, + "content": "4.1. Experimental Setup" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.78, + 0.906, + 0.901 + ], + "angle": 0, + "content": "Datasets. We introduce two model variants, each trained on different datasets. (1) DiffusionSfM-CO3D: Following prior work [12, 41], we train and evaluate our model on the CO3D dataset [21], which consists of turntable video sequences of various object categories. Specifically, we train on 41 object categories and evaluate on both these seen categories and an additional 10 unseen categories to assess generalization. (2) DiffusionSfM: This variant is trained on the" + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.946, + 0.516, + 0.957 + ], + "angle": 0, + "content": "6320" + } + ], + [ + { + "type": "image_caption", + "bbox": [ + 0.135, + 0.09, + 0.25, + 0.117 + ], + "angle": 0, + "content": "Input Images /GT Point Clouds" + }, + { + "type": "image", + "bbox": [ + 0.124, + 0.12, + 0.262, + 0.173 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.125, + 0.174, + 0.262, + 0.226 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.346, + 0.097, + 0.44, + 0.11 + ], + "angle": 0, + "content": "DiffusionSfM" + }, + { + "type": "image", + "bbox": [ + 0.3, + 0.129, + 0.462, + 0.216 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.561, + 0.097, + 0.646, + 0.11 + ], + "angle": 0, + "content": "RD + MoGe" + }, + { + "type": "image", + "bbox": [ + 0.513, + 0.129, + 0.674, + 0.215 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.783, + 0.097, + 0.842, + 0.11 + ], + "angle": 0, + "content": "DUSt3R" + }, + { + "type": "image", + "bbox": [ + 0.719, + 0.129, + 0.878, + 0.215 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.119, + 0.233, + 0.326, + 0.317 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.326, + 0.235, + 0.52, + 0.316 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.52, + 0.233, + 0.665, + 0.316 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.665, + 0.233, + 0.882, + 0.317 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.122, + 0.33, + 0.169, + 0.366 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.122, + 0.366, + 0.169, + 0.402 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.17, + 0.331, + 0.214, + 0.366 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.215, + 0.331, + 0.263, + 0.366 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.12, + 0.416, + 0.316, + 0.493 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.315, + 0.416, + 0.518, + 0.494 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.522, + 0.416, + 0.688, + 0.492 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.729, + 0.323, + 0.882, + 0.409 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.692, + 0.415, + 0.882, + 0.494 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.122, + 0.509, + 0.17, + 0.545 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.122, + 0.546, + 0.169, + 0.582 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.122, + 0.583, + 0.169, + 0.619 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.17, + 0.509, + 0.214, + 0.545 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.17, + 0.546, + 0.214, + 0.582 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.17, + 0.583, + 0.214, + 0.619 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.215, + 0.509, + 0.263, + 0.545 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.215, + 0.546, + 0.263, + 0.582 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.116, + 0.622, + 0.241, + 0.724 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.303, + 0.523, + 0.47, + 0.602 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.512, + 0.523, + 0.678, + 0.602 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.72, + 0.523, + 0.884, + 0.605 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.325, + 0.62, + 0.447, + 0.724 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.548, + 0.62, + 0.663, + 0.723 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.746, + 0.621, + 0.876, + 0.724 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.09, + 0.738, + 0.907, + 0.807 + ], + "angle": 0, + "content": "Figure 3. Qualitative Comparison on Camera Pose Accuracy and Predicted Geometry. For each method, we plot the ground-truth cameras in black and the predicted cameras in other colors. DiffusionSfM demonstrates robust performance even with challenging inputs. Compared to DUSt3R, which sometimes fails to register images in a consistent manner, DiffusionSfM consistently yields a coherent global prediction. Additionally, while we observe that DUSt3R can predict highly precise camera rotations, it often struggles with camera centers (see the backpack example). Input images depicting scenes are out-of-distribution for RayDiffusion, as it is trained on CO3D only." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.821, + 0.482, + 0.897 + ], + "angle": 0, + "content": "datasets used for DUSt3R [36], excluding Waymo [31] due to its excessively sparse depth maps. The included datasets are Habitat [24], CO3D [21], ScanNet++ [39], ArkitScenes [1], Static Scenes 3D [17], MegaDepth [11], and BlendedMVS [38]. We follow the DUSt3R official repository [37]" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.822, + 0.907, + 0.897 + ], + "angle": 0, + "content": "guidelines to extract image pairs, ensuring reasonable overlap between paired images. To construct multi-view data given these pre-computed pairs, we iteratively sample images using an adjacency matrix, maintaining sufficient overlap with the selected set. Beyond CO3D, this model variant" + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.946, + 0.514, + 0.957 + ], + "angle": 0, + "content": "6321" + } + ], + [ + { + "type": "image_caption", + "bbox": [ + 0.136, + 0.09, + 0.23, + 0.104 + ], + "angle": 0, + "content": "Input Images" + }, + { + "type": "image", + "bbox": [ + 0.118, + 0.111, + 0.249, + 0.503 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.305, + 0.09, + 0.398, + 0.103 + ], + "angle": 0, + "content": "DiffusionSfM" + }, + { + "type": "image", + "bbox": [ + 0.259, + 0.121, + 0.425, + 0.205 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.524, + 0.09, + 0.616, + 0.104 + ], + "angle": 0, + "content": "RayDiffusion" + }, + { + "type": "image", + "bbox": [ + 0.485, + 0.119, + 0.644, + 0.2 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.761, + 0.09, + 0.821, + 0.103 + ], + "angle": 0, + "content": "DUSt3R" + }, + { + "type": "image", + "bbox": [ + 0.703, + 0.121, + 0.868, + 0.204 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.275, + 0.231, + 0.418, + 0.312 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.493, + 0.232, + 0.661, + 0.312 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.723, + 0.226, + 0.88, + 0.314 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.117, + 0.326, + 0.249, + 0.395 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.264, + 0.326, + 0.42, + 0.399 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.479, + 0.316, + 0.661, + 0.393 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.703, + 0.328, + 0.884, + 0.405 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.117, + 0.4, + 0.249, + 0.468 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.267, + 0.409, + 0.43, + 0.487 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.485, + 0.41, + 0.653, + 0.486 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.709, + 0.41, + 0.871, + 0.486 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.091, + 0.515, + 0.904, + 0.53 + ], + "angle": 0, + "content": "Figure 4. Additional Qualitative Results on Predicted Camera Poses. DiffusionSfM shows robustness to ambiguous patterns in inputs." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.556, + 0.435, + 0.571 + ], + "angle": 0, + "content": "is also evaluated on Habitat and RealEstate10k [42]." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.584, + 0.484, + 0.901 + ], + "angle": 0, + "content": "Baselines and Metrics. To evaluate camera pose accuracy in the sparse-view setup, we compare with RayDiffusion [41] and DUSt3R [36], along with previous methods [12, 22, 34]. DUSt3R is trained initially on a mixture of eight datasets, while most other baselines are trained only on CO3D. For a fair and comprehensive comparison, we re-train DUSt3R on the 41-10 split of CO3D, using the authors' official implementation [37] and hyperparameters (referred to as DUSt3R-CO3D). To evaluate camera predictions, we follow prior work [41] and convert predicted rays back to traditional extrinsic matrices and report two pose accuracy metrics: (1) Camera Rotation Accuracy which compares the predicted relative camera rotation between images against ground truth and (2) Camera Center Accuracy which compares predicted camera centers to the ground truth after a similarity alignment. To evaluate the estimated geometry (i.e., ray endpoints), we report Chamfer Distance (CD). Additionally, to explore whether combining an off-the-shelf sparse-view pose estimation method (RayDiffusion) with a monocular depth estimation model (MoGe [35]) is sufficient to infer 3D scene geometry from" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.556, + 0.905, + 0.603 + ], + "angle": 0, + "content": "multiple images, we introduce a new baseline, \\(\\mathsf{RD} + \\mathsf{MoGe}\\). We include more details in the supplementary. Unless otherwise specified, all evaluations use 2-8 input images." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.619, + 0.71, + 0.634 + ], + "angle": 0, + "content": "4.2. Evaluation on CO3D" + }, + { + "type": "text", + "bbox": [ + 0.511, + 0.644, + 0.907, + 0.901 + ], + "angle": 0, + "content": "Camera Pose Accuracy. We present the quantitative results in Tab. 1. For camera rotation accuracy, both versions of DiffusionSfM outperform all baselines, except that DUSt3R achieves slightly higher overall accuracy than DiffusionSfM-CO3D on unseen categories – likely due to its training on more data. For camera center accuracy, our approach consistently outperforms all other methods. We hypothesize that these gains stem from our explicit modeling of camera centers through ray origin prediction. The qualitative results in Fig. 3 illustrate that DiffusionSfM produces robust predictions given challenging inputs, whereas DUSt3R produces inaccurate results. We attribute this improvement to our model's probabilistic modeling capability derived from diffusion, as well as its multi-view reasoning abilities, which together effectively handle these challenging scenarios. While we observe that DUSt3R can predict highly precise camera rotations, it often struggles with cam-" + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.516, + 0.957 + ], + "angle": 0, + "content": "6322" + } + ], + [ + { + "type": "table", + "bbox": [ + 0.093, + 0.089, + 0.902, + 0.362 + ], + "angle": 0, + "content": "
# of ImagesRotation Accuracy (↑, @ 15°)Center Accuracy (↑, @ 0.1)
2345678234567
Seen CategoriesCOLMAP [22]30.728.426.526.827.028.130.610034.523.818.915.614.5
PoseDiffusion [34]75.776.476.877.478.078.778.810077.569.765.963.762.8
RelPose++ [12]81.882.884.184.784.985.385.510085.078.074.271.970.3
RayDiffusion [41]91.892.492.692.993.193.393.310094.290.587.886.285.0
DUSt3R-CO3D [36]86.787.988.088.288.688.888.910092.086.883.882.081.1
DUSt3R [36]91.792.793.393.693.894.094.310093.085.781.979.677.8
DiffusionSfM-CO3D93.494.094.594.895.095.295.110095.993.692.291.290.7
DiffusionSfM92.493.594.094.594.795.195.210095.392.991.991.290.8
Unseen CategoriesCOLMAP [22]34.531.831.031.732.735.038.510036.025.520.017.917.6
PoseDiffusion [34]63.264.264.265.766.267.067.710063.650.545.743.041.2
RelPose++ [12]69.871.171.972.873.874.474.910070.658.853.450.447.8
RayDiffusion [41]83.585.686.386.987.287.588.110087.781.177.074.172.4
DUSt3R-CO3D [36]79.881.582.682.783.083.383.710083.677.271.870.068.1
DUSt3R [36]90.892.693.693.693.893.693.410087.979.874.371.769.4
DiffusionSfM-CO3D90.491.292.793.093.193.393.510091.187.785.383.782.7
DiffusionSfM91.592.893.693.994.194.294.410092.487.986.285.384.1
" + }, + { + "type": "table_caption", + "bbox": [ + 0.09, + 0.371, + 0.907, + 0.442 + ], + "angle": 0, + "content": "Table 1. Camera Rotation and Center Accuracy on CO3D. On the left, we report the proportion of relative camera rotations within \\(15^{\\circ}\\) of the ground truth. On the right, we report the proportion of camera centers within \\(10\\%\\) of the scene scale, relative to the ground truth. To align the predicted camera centers to ground truth, we apply an optimal similarity transform \\((s, \\mathbf{R}, \\mathbf{t})\\), hence the alignment is perfect at \\(N = 2\\) but worsens with more images. DiffusionSfM outperforms all other methods for camera center accuracy, and outperforms all methods trained on equivalent data for rotation accuracy." + }, + { + "type": "text", + "bbox": [ + 0.089, + 0.461, + 0.483, + 0.492 + ], + "angle": 0, + "content": "era centers (see the backpack example). Additional qualitative results on camera accuracy are provided in Fig.4." + }, + { + "type": "text", + "bbox": [ + 0.093, + 0.494, + 0.483, + 0.751 + ], + "angle": 0, + "content": "Predicted Geometry. To evaluate predicted geometry, we compute Chamfer Distance (CD) and show comparisons against baselines in Tab. 2. We compute CD in two setups (with and without foreground object mask), and find that DiffusionSfM-CO3D performs best without foreground masking. In this setup, the predicted ray endpoints corresponding to the background image pixels tend to have larger coordinate values than foreground ones, and therefore dominate CD. This result indicates that our model provides more accurate predictions for complex image backgrounds. In terms of CD with masking, DUSt3R achieves the best result, while our two model variants outperform DUSt3R-CO3D. See Fig. 3 for visualization. We also notice that naively combining RayDiffusion poses with estimated depths from MoGe does not work well, as the inferred depths are potentially inconsistent across different views, thus resulting in degraded performance." + }, + { + "type": "title", + "bbox": [ + 0.09, + 0.77, + 0.4, + 0.787 + ], + "angle": 0, + "content": "4.3. Evaluation on Scene-Level Datasets" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.795, + 0.483, + 0.903 + ], + "angle": 0, + "content": "Beyond the object-centric CO3D dataset [21], we compare DiffusionSfM against DUSt3R on two scene-level datasets: Habitat (in-distribution) [24] and RealEstate10k (out-of-distribution) [42]. The results are presented in Tab. 3. While DiffusionSfM achieves comparable camera rotation accuracy to DUSt3R, it consistently predicts more accurate camera centers, aligning with our findings on CO3D." + }, + { + "type": "table", + "bbox": [ + 0.517, + 0.458, + 0.905, + 0.635 + ], + "angle": 0, + "content": "
# of Images2345678
RD*+MoGe [41]0.0590.0640.0710.0620.0630.0610.061
DUSt3R* [36]0.0360.0370.0400.0400.0370.0360.039
DUSt3R [36]0.0210.0230.0240.0240.0250.0250.023
DiffusionSfM*0.0190.0210.0220.0220.0210.0210.021
DiffusionSfM0.0250.0250.0250.0250.0260.0250.028
RD*+MoGe [41]0.0710.0750.0680.0670.0660.0640.064
DUSt3R* [36]0.0380.0360.0360.0360.0340.0330.034
DUSt3R [36]0.0230.0220.0190.0200.0190.0200.020
DiffusionSfM*0.0280.0250.0240.0240.0240.0230.022
DiffusionSfM0.0270.0240.0230.0220.0220.0220.021
" + }, + { + "type": "table_caption", + "bbox": [ + 0.513, + 0.645, + 0.907, + 0.73 + ], + "angle": 0, + "content": "Table 2. Chamfer Distance \\((\\downarrow)\\) on CO3D Unseen Categories. Top: CD computed on all scene points. Bottom: CD computed on foreground points only. Models marked with “*” are trained on CO3D only, while those without are trained on multiple datasets. Note that top and bottom values are not directly comparable, as each ground-truth point cloud is individually normalized." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.757, + 0.673, + 0.774 + ], + "angle": 0, + "content": "4.4. Inference Speed" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.78, + 0.907, + 0.903 + ], + "angle": 0, + "content": "Though our method requires iterative diffusion denoising at inference, we can speed this up by performing early stopping. Specifically, we can treat the \\( x_0 \\)-prediction from early timesteps as our output instead of iterating over all denoising timesteps. Consistent with observations by Zhang et al. [41], this in fact yields more accurate predictions than the final-step diffusion outputs. As a result we only require 10 denoising diffusion timesteps for inference, taking 1.91" + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.516, + 0.957 + ], + "angle": 0, + "content": "6323" + } + ], + [ + { + "type": "table", + "bbox": [ + 0.095, + 0.089, + 0.48, + 0.155 + ], + "angle": 0, + "content": "
Hab.DUSt3R97.0/10095.0/97.694.3/95.094.2/93.1
DiffusionSfM92.6/10093.7/99.194.2/98.894.9/98.6
RE10DUSt3R98.1/10097.7/68.797.6/57.997.7/53.3
DiffusionSfM97.8/10097.5/75.697.7/66.297.9/61.9
" + }, + { + "type": "table_caption", + "bbox": [ + 0.09, + 0.164, + 0.483, + 0.248 + ], + "angle": 0, + "content": "Table 3. Camera Rotation and Center Accuracy on Two Scene-Level Datasets. Top: Habitat (2-5 views). Bottom: RealEstate10k (2, 4, 6, 8 views). Each grid reports camera rotation accuracy (left, \\(\\uparrow\\)) and center accuracy (right, \\(\\uparrow\\)). While DiffusionSfM performs on par with DUSt3R in rotation accuracy, it consistently surpasses DUSt3R in center accuracy." + }, + { + "type": "table", + "bbox": [ + 0.093, + 0.266, + 0.483, + 0.325 + ], + "angle": 0, + "content": "
# of Images2345678
DiffusionSfM*0.0190.0210.0220.0220.0210.0210.021
w/o Mask0.0200.0220.0230.0230.0220.0220.024
" + }, + { + "type": "table_caption", + "bbox": [ + 0.09, + 0.335, + 0.483, + 0.42 + ], + "angle": 0, + "content": "Table 4. Ablation Study on GT Mask Conditioning for CO3D Unseen Categories. We assess the effect of replacing GT mask conditioning with depth interpolation in our CO3D variant (DiffusionSfM*), by reporting the CD for predicted geometry. Incorporating mask conditioning to indicate missing data during training improves geometry quality." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.451, + 0.484, + 0.512 + ], + "angle": 0, + "content": "seconds on a single A5000 GPU with 8 input images. In contrast, DUSt3R takes 8.73 seconds to run the complete pairwise inference and global alignment procedure. We provide additional analysis in the supplementary material." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.527, + 0.245, + 0.543 + ], + "angle": 0, + "content": "4.5. Ablation Study" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.551, + 0.483, + 0.657 + ], + "angle": 0, + "content": "Homogeneous Coordinates. The use of homogeneous coordinates for ray origins and endpoints is crucial for stable model training. To assess its impact, we replace the proposed homogeneous representation with standard 3D coordinates in \\(\\mathbb{R}^3\\). Our experiments show that this variant is difficult to train and fails to converge. Further details on this experiment are provided in the supplementary material." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.659, + 0.483, + 0.901 + ], + "angle": 0, + "content": "GT Mask Conditioning. To evaluate the effectiveness of the proposed GT mask conditioning, we train a baseline model without using this strategy. For missing values in the ground-truth depth maps, we use nearest-neighbor interpolation to fill in invalid pixels. This experiment is conducted on the CO3D dataset, with results presented in Tab. 4. The findings show that removing GT mask conditioning consistently degrades predicted geometry across varying numbers of input views, even when the loss is still masked. While interpolation can effectively fill missing depth within object regions, it often introduces substantial noise in the background (e.g., the sky). This noise negatively impacts diffusion model training, as the entire ray origin and endpoint maps are used as input. We anticipate an even larger performance gap on non-object-centric datasets with more outdoor scenes, such as MegaDepth [11]." + }, + { + "type": "image", + "bbox": [ + 0.517, + 0.088, + 0.909, + 0.18 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.512, + 0.189, + 0.907, + 0.287 + ], + "angle": 0, + "content": "Figure 5. Multi-modality of DiffusionSfM. We show two distinct samples from DiffusionSfM, starting from the same input images but with different random noise. Sample 1 explains the input images by putting all flowers on the left side, while Sample 2 places one flower on each side (note the difference in the green camera's viewpoint). DiffusionSfM is able to predict multi-modal geometry distributions when the scene layout is ambiguous in the inputs." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.313, + 0.857, + 0.33 + ], + "angle": 0, + "content": "4.6. Multi-modality from Multiple Sampling" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.335, + 0.907, + 0.472 + ], + "angle": 0, + "content": "Diffusion models enable the generation of diverse samples from challenging input images. For instance, in Fig. 5, the vase exhibits symmetric patterns, and we present two distinct predicted endpoints from DiffusionSfM, each offering a different interpretation of the flowers in the image. Compared to regression models such as DUSt3R [8], DiffusionSfM is better suited for handling uncertainty – an inherent aspect of our task – where a sparse set of input images can correspond to multiple plausible 3D scene geometry." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.484, + 0.628, + 0.5 + ], + "angle": 0, + "content": "5. Discussion" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.51, + 0.906, + 0.57 + ], + "angle": 0, + "content": "We present DiffusionSfM and demonstrate that it recovers accurate predictions of both cameras and geometry from multi-view inputs. Although our results are promising, several challenges and open questions remain." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.57, + 0.906, + 0.751 + ], + "angle": 0, + "content": "Notably, DiffusionSfM employs a pixel-space diffusion model, in contrast to the latent-space models adopted by state-of-the-art T2I generative systems. Operating in pixel space may require greater model capacity, yet our current model remains relatively small - potentially explaining the noisy patterns observed along object boundaries. Learning an expressive latent space for ray origins and endpoints by training a VAE could be a promising direction for future work. In addition, the compute requirement in multi-view transformers scales quadratically with the number of input images: one would require masked attention to deploy systems like ours for a large set of input images." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.751, + 0.909, + 0.888 + ], + "angle": 0, + "content": "Despite these challenges, we believe that our work highlights the potential of a unified approach for multi-view geometry tasks. We envision that our approach can be built upon to train a common system across related geometric tasks, such as SfM (input images with unknown origins and endpoints), registration (some images have known origins and endpoints, whereas others don't), mapping (known rays but unknown endpoints), and view synthesis (unknown pixel values for known rays)." + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.516, + 0.957 + ], + "angle": 0, + "content": "6324" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.09, + 0.092, + 0.482, + 0.151 + ], + "angle": 0, + "content": "Acknowledgements: We thank the members of the Physical Perception Lab at CMU for their valuable discussions, and extend special thanks to Yanbo Xu for his insights on diffusion models." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.153, + 0.483, + 0.394 + ], + "angle": 0, + "content": "This work used Bridges-2 [4] at Pittsburgh Supercomputing Center through allocation CIS240166 from the Advanced Cyberinfrastructure Coordination Ecosystem: Services & Support (ACCESS) program, which is supported by National Science Foundation grants #2138259, #2138286, #2138307, #2137603, and #2138296. This work was supported by Intelligence Advanced Research Projects Activity (IARPA) via Department of Interior/Interior Business Center (DOI/IBC) contract number 140D0423C0074. The U.S. Government is authorized to reproduce and distribute reprints for Governmental purposes notwithstanding any copyright annotation thereon. Disclaimer: The views and conclusions contained herein are those of the authors and should not be interpreted as necessarily representing the official policies or endorsements, either expressed or implied, of IARPA, DOI/IBC, or the U.S. Government." + }, + { + "type": "title", + "bbox": [ + 0.093, + 0.405, + 0.188, + 0.42 + ], + "angle": 0, + "content": "References" + }, + { + "type": "ref_text", + "bbox": [ + 0.1, + 0.43, + 0.483, + 0.498 + ], + "angle": 0, + "content": "[1] Gilad Baruch, Zhuoyuan Chen, Afshin Dehghan, Tal Dimry, Yuri Feigin, Peter Fu, Thomas Gebauer, Brandon Joffe, Daniel Kurz, Arik Schwartz, et al. Arkitsscenes: A diverse real-world dataset for 3d indoor scene understanding using mobile rgb-d data. In NeurIPS D&B, 2021. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.101, + 0.5, + 0.482, + 0.539 + ], + "angle": 0, + "content": "[2] Herbert Bay, Andreas Ess, Tinne Tuytelaars, and Luc Van Gool. Surf: Speeded up robust features. In ECCV, 2006. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.541, + 0.483, + 0.609 + ], + "angle": 0, + "content": "[3] Eric Brachmann, Jamie Wynn, Shuai Chen, Tommaso Cavallari, Aron Monszpart, Daniyar Turmukhambetov, and Victor Adrian Prisacariu. Scene coordinate reconstruction: posing of image collections via incremental learning of a relocalizer. In ECCV, 2024. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.611, + 0.482, + 0.679 + ], + "angle": 0, + "content": "[4] Shawn T Brown, Paola Buitrago, Edward Hanna, Sergiu Sanielevici, Robin Scibek, and Nicholas A Nystrom. Bridges-2: A platform for rapidly-evolving and data intensive research. In Practice and experience in advanced research computing, 2021. 9" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.68, + 0.482, + 0.734 + ], + "angle": 0, + "content": "[5] Ruojin Cai, Joseph Tung, Qianqian Wang, Hadar Averbuch-Elor, Bharath Hariharan, and Noah Snavely. Doppelgangers: Learning to disambiguate images of similar structures. In ICCV, 2023. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.736, + 0.482, + 0.775 + ], + "angle": 0, + "content": "[6] Daniel DeTone, Tomasz Malisiewicz, and Andrew Rabinovich. Superpoint: Self-supervised interest point detection and description. In CVPR, 2018. 1, 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.777, + 0.482, + 0.831 + ], + "angle": 0, + "content": "[7] Bardienus Duisterhof, Lojze Zust, Philippe Weinzaepfel, Vincent Leroy, Yohann Cabon, and Jerome Revaud. Mast3r-sfm: a fully-integrated solution for unconstrained structure-from-motion. In 3DV, 2025. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.832, + 0.482, + 0.873 + ], + "angle": 0, + "content": "[8] Johan Edstedt, Qiyu Sun, Georg Bokman, Marten Wadenback, and Michael Felsberg. RoMa: Robust Dense Feature Matching. In CVPR, 2024. 2, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.874, + 0.482, + 0.901 + ], + "angle": 0, + "content": "[9] Jonathan Ho, Ajay Jain, and Pieter Abbeel. Denoising diffusion probabilistic models. In NeurIPS, 2020. 2, 3" + }, + { + "type": "list", + "bbox": [ + 0.1, + 0.43, + 0.483, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.093, + 0.905, + 0.121 + ], + "angle": 0, + "content": "[10] Vincent Leroy, Yohann Cabon, and Jérôme Revaud. Grounding image matching in 3d with mast3r, 2024. ECCV. 1, 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.122, + 0.905, + 0.162 + ], + "angle": 0, + "content": "[11] Zhengqi Li and Noah Snavely. Megadepth: Learning single-view depth prediction from internet photos. In CVPR, 2018. 2,4,5,8" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.163, + 0.905, + 0.203 + ], + "angle": 0, + "content": "[12] Amy Lin, Jason Y Zhang, Deva Ramanan, and Shubham Tulsiani. Relpose++: Recovering 6d poses from sparse-view observations. In 3DV, 2024. 2, 4, 6, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.205, + 0.905, + 0.245 + ], + "angle": 0, + "content": "[13] Philipp Lindenberger, Paul-Edouard Sarlin, Viktor Larsson, and Marc Pollefeys. Pixel-perfect structure-from-motion with featuremetric refinement. In ICCV, 2021. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.247, + 0.905, + 0.287 + ], + "angle": 0, + "content": "[14] Philipp Lindenberger, Paul-Edouard Sarlin, Viktor Larsson, and Marc Pollefeys. Pixel-Perfect Structure-from-Motion with Featuremetric Refinement. In ICCV, 2021. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.288, + 0.905, + 0.328 + ], + "angle": 0, + "content": "[15] Philipp Lindenberger, Paul-Edouard Sarlin, and Marc Pollefeys. Lightglue: Local feature matching at light speed. In ICCV, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.33, + 0.905, + 0.357 + ], + "angle": 0, + "content": "[16] David G Lowe. Distinctive image features from scale-invariant keypoints. In IJCV, 2004. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.359, + 0.905, + 0.413 + ], + "angle": 0, + "content": "[17] Nikolaus Mayer, Eddy Ilg, Philip Hausser, Philipp Fischer, Daniel Cremers, Alexey Dosovitskiy, and Thomas Brox. A large dataset to train convolutional networks for disparity, optical flow, and scene flow estimation. In CVPR, 2016. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.415, + 0.905, + 0.482 + ], + "angle": 0, + "content": "[18] Maxime Oquab, Timothee Darcet, Theo Moutakanni, Huy Vo, Marc Szafraniec, Vasil Khalidov, Pierre Fernandez, Daniel Haziza, Francisco Massa, Alaaeldin El-Nouby, et al. Dinov2: Learning robust visual features without supervision. In TMLR, 2024. 3, 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.484, + 0.905, + 0.51 + ], + "angle": 0, + "content": "[19] William Peebles and Saining Xie. Scalable diffusion models with transformers. In ICCV, 2023. 2, 3, 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.511, + 0.905, + 0.551 + ], + "angle": 0, + "content": "[20] René Ranftl, Alexey Bochkovskiy, and Vladlen Koltun. Vision transformers for dense prediction. In ICCV, 2021. 3, 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.553, + 0.905, + 0.621 + ], + "angle": 0, + "content": "[21] Jeremy Reizenstein, Roman Shapovalov, Philipp Henzler, Luca Sbordone, Patrick Labatut, and David Novotny. Common objects in 3d: Large-scale learning and evaluation of real-life 3d category reconstruction. In ICCV, 2021. 2, 4, 5, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.623, + 0.905, + 0.664 + ], + "angle": 0, + "content": "[22] Paul-Edouard Sarlin, Cesar Cadena, Roland Siegwart, and Marcin Dymczyk. From coarse to fine: Robust hierarchical localization at large scale. In CVPR, 2019. 6, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.665, + 0.905, + 0.718 + ], + "angle": 0, + "content": "[23] Paul-Edouard Sarlin, Daniel DeTone, Tomasz Malisiewicz, and Andrew Rabinovich. Superglue: Learning feature matching with graph neural networks. In CVPR, 2020. 1, 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.72, + 0.905, + 0.775 + ], + "angle": 0, + "content": "[24] Manolis Savva, Abhishek Kadian, Oleksandr Maksymets, Yili Zhao, Erik Wijmans, Bhavana Jain, Julian Straub, Jia Liu, Vladlen Koltun, Jitendra Malik, et al. Habitat: A platform for embodied ai research. In ICCV, 2019. 2, 5, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.777, + 0.905, + 0.802 + ], + "angle": 0, + "content": "[25] Johannes L Schonberger and Jan-Michael Frahm. Structure-from-motion revisited. In CVPR, 2016. 1, 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.804, + 0.905, + 0.857 + ], + "angle": 0, + "content": "[26] Samarth Sinha, Jason Y Zhang, Andrea Tagliasacchi, Igor Gilitschenski, and David B Lindell. Sparsepose: Sparse-view camera pose regression and refinement. In CVPR, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.86, + 0.905, + 0.901 + ], + "angle": 0, + "content": "[27] Cameron Smith, David Charatan, Ayush Tewari, and Vincent Sitzmann. Flowmap: High-quality camera poses, intrinsics, and depth via gradient descent. In 3DV, 2025. 2" + }, + { + "type": "list", + "bbox": [ + 0.517, + 0.093, + 0.905, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.516, + 0.957 + ], + "angle": 0, + "content": "6325" + } + ], + [ + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.093, + 0.482, + 0.134 + ], + "angle": 0, + "content": "[28] Noah Snavely, Steven M Seitz, and Richard Szeliski. Photo tourism: exploring photo collections in 3d. In ACM SIGGRAPH, 2006. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.136, + 0.483, + 0.163 + ], + "angle": 0, + "content": "[29] Jiaming Song, Chenlin Meng, and Stefano Ermon. Denoising diffusion implicit models. In ICLR, 2021. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.165, + 0.482, + 0.205 + ], + "angle": 0, + "content": "[30] Jiaming Sun, Zehong Shen, Yuang Wang, Hujun Bao, and Xiaowei Zhou. Loftr: Detector-free local feature matching with transformers. In CVPR, 2021. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.208, + 0.482, + 0.274 + ], + "angle": 0, + "content": "[31] Pei Sun, Henrik Kretzschmar, Xerxes Dotiwalla, Aurelien Chouard, Vijaysai Patnaik, Paul Tsui, James Guo, Yin Zhou, Yuning Chai, Benjamin Caine, et al. Scalability in perception for autonomous driving: Waymo open dataset. In CVPR, 2020. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.277, + 0.482, + 0.305 + ], + "angle": 0, + "content": "[32] Chengzhou Tang and Ping Tan. Ba-net: Dense bundle adjustment network. In ICLR, 2019. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.307, + 0.482, + 0.348 + ], + "angle": 0, + "content": "[33] Ashish Vaswani, Noam Shazeer, Niki Parmar, Jakob Uszkoreit, Llion Jones, Aidan N Gomez, Lukasz Kaiser, and Illia Polosukhin. Attention is all you need. In NeurIPS, 2017. 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.35, + 0.482, + 0.39 + ], + "angle": 0, + "content": "[34] Jianyuan Wang, Christian Rupprecht, and David Novotny. Posediffusion: Solving pose estimation via diffusion-aided bundle adjustment. In ICCV, 2023. 2, 6, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.392, + 0.482, + 0.459 + ], + "angle": 0, + "content": "[35] Ruicheng Wang, Sicheng Xu, Cassie Dai, Jianfeng Xiang, Yu Deng, Xin Tong, and Jiaolong Yang. Moge: Unlocking accurate monocular geometry estimation for open-domain images with optimal training supervision. In CVPR, 2025. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.463, + 0.482, + 0.504 + ], + "angle": 0, + "content": "[36] Shuzhe Wang, Vincent Leroy, Yohann Cabon, Boris Chidlovskii, and Jerome Revaud. Dust3r: Geometric 3d vision made easy. In CVPR, 2024. 1, 2, 4, 5, 6, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.506, + 0.482, + 0.559 + ], + "angle": 0, + "content": "[37] Shuzhe Wang, Vincent Leroy, Yohann Cabon, Boris Chidlovskii, and Jerome Revaud. Dust3r: Geometric 3d vision made easy, 2024. https://github.com/naver/dust3r. 5, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.562, + 0.482, + 0.616 + ], + "angle": 0, + "content": "[38] Yao Yao, Zixin Luo, Shiwei Li, Jingyang Zhang, Yufan Ren, Lei Zhou, Tian Fang, and Long Quan. Blendedmvs: A large-scale dataset for generalized multi-view stereo networks. In CVPR, 2020. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.619, + 0.482, + 0.659 + ], + "angle": 0, + "content": "[39] Chandan Yeshwanth, Yueh-Cheng Liu, Matthias Nießner, and Angela Dai. Scannet++: A high-fidelity dataset of 3d indoor scenes. In ICCV, 2023. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.662, + 0.482, + 0.702 + ], + "angle": 0, + "content": "[40] Jason Y Zhang, Deva Ramanan, and Shubham Tulsiani. Relapse: Predicting probabilistic relative rotation for single objects in the wild. In ECCV, 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.704, + 0.482, + 0.759 + ], + "angle": 0, + "content": "[41] Jason Y. Zhang, Amy Lin, Moneish Kumar, Tzu-Hsuan Yang, Deva Ramanan, and Shubham Tulsiani. Cameras as rays: Sparse-view pose estimation via ray diffusion. In ICLR, 2024. 2, 3, 4, 6, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.761, + 0.482, + 0.815 + ], + "angle": 0, + "content": "[42] Tinghui Zhou, Richard Tucker, John Flynn, Graham Fyffe, and Noah Snavely. Stereo magnification: Learning view synthesis using multiplane images. In ACM SIGGRAPH, 2018. 2, 6, 7" + }, + { + "type": "list", + "bbox": [ + 0.093, + 0.093, + 0.483, + 0.815 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.946, + 0.516, + 0.956 + ], + "angle": 0, + "content": "6326" + } + ] +] \ No newline at end of file diff --git a/2025/DiffusionSfM_ Predicting Structure and Motion via Ray Origin and Endpoint Diffusion/a0e804a0-3c59-43c5-bb05-b2160d8636bd_origin.pdf b/2025/DiffusionSfM_ Predicting Structure and Motion via Ray Origin and Endpoint Diffusion/a0e804a0-3c59-43c5-bb05-b2160d8636bd_origin.pdf new file mode 100644 index 0000000000000000000000000000000000000000..632349e97db4cfb019488a2902c621885102f7a1 --- /dev/null +++ b/2025/DiffusionSfM_ Predicting Structure and Motion via Ray Origin and Endpoint Diffusion/a0e804a0-3c59-43c5-bb05-b2160d8636bd_origin.pdf @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:88ff6bb8181590e220a5d7f7627b809fa863015cc47ac3591aeab8cad8eaa1d2 +size 7088208 diff --git a/2025/DiffusionSfM_ Predicting Structure and Motion via Ray Origin and Endpoint Diffusion/full.md b/2025/DiffusionSfM_ Predicting Structure and Motion via Ray Origin and Endpoint Diffusion/full.md new file mode 100644 index 0000000000000000000000000000000000000000..ee6445ba518218a952ae70f721801b07b1afa027 --- /dev/null +++ b/2025/DiffusionSfM_ Predicting Structure and Motion via Ray Origin and Endpoint Diffusion/full.md @@ -0,0 +1,339 @@ +# DiffusionSfM: Predicting Structure and Motion via Ray Origin and Endpoint Diffusion + +Qitao Zhao, Amy Lin, Jeff Tan, Jason Y. Zhang, Deva Ramanan, Shubham Tulsiani Carnegie Mellon University + +Project page: qitaozhao.github.io/DiffusionSfM + +![](images/ac31d7718dc556e701eba8b776c1f3e30bb00ad6f957b560bb72e53affc2382c.jpg) +Figure 1. DiffusionSfM. Top: Given a set of multi-view images (left), DiffusionSfM represents scene geometry and cameras (right) as pixel-wise ray origins and endpoints in a global frame. It learns a denoising diffusion model to infer these elements directly from multiview inputs. Unlike traditional Structure-from-Motion (SfM) pipelines, which separate pairwise reasoning and global optimization into two stages, our approach unifies both into a single end-to-end multi-view reasoning framework. Bottom: Example results of inferred scene geometry and cameras for two distinct settings: a real-world outdoor scene (left) and a synthetic indoor scene (right). + +# Abstract + +Current Structure-from-Motion (SfM) methods typically follow a two-stage pipeline, combining learned or geometric pairwise reasoning with a subsequent global optimization step. In contrast, we propose a data-driven multi-view reasoning approach that directly infers 3D scene geometry and camera poses from multi-view images. Our framework, DiffusionSfM, parameterizes scene geometry and cameras as pixel-wise ray origins and endpoints in a global frame and employs a transformer-based denoising diffusion model to predict them from multi-view inputs. To address practical challenges in training diffusion models with missing data and unbounded scene coordinates, we introduce specialized mechanisms that ensure robust learning. We empirically validate DiffusionSfM on both synthetic and real datasets, demonstrating that it outperforms classical and learning-based approaches while naturally modeling uncertainty. + +# 1. Introduction + +The task of recovering structure (geometry) and motion (cameras) from multi-view images has long been a focus of the computer vision community, with typical pipelines [25] performing pairwise correspondence estimation followed by global optimization. While classical methods relied on hand-designed features, matching, and optimization, there has been a recent shift towards incorporating learning-based alternatives [5, 6, 14, 23]. More recently, the widely influential DUSt3R [36] advocates for predicting pairwise 3D pointmaps (instead of only correspondences), demonstrating that this can yield accurate dense geometry and cameras. In order to reconstruct more than two views, DUSt3R (and its variants [10]) still require a global optimization reminiscent of classic bundle adjustment. While these methods, both classical and learning-based, have led to impressive improvements in SfM, the overall approach is largely unchanged – learned or geometric pairwise reasoning followed by global optimization. In this work, we + +seek to develop an alternative approach that directly predicts both structure and motion, unifying pairwise reasoning and global optimization into a single multi-view framework. + +We are of course not the first to attempt to find unified alternatives to the two-stage SfM pipeline. In the sparse-view setting where conventional correspondence-based methods struggle, several works employ multi-view architectures to jointly reason across input images. SparsePose [26], RelPose++ [12], and PoseDiffusion [34] all leverage multiview transformers to estimate camera pose for input images, albeit using differing mechanisms such as regression, energy-based modeling, and denoising diffusion. More recently, RayDiffusion [41] argues for a local raymap parameterization of cameras instead of a global extrinsic matrix and shows that existing patch-based transformers can be easily adapted for this task, yielding significantly more accurate pose predictions. Importantly, such methods predict only camera motion and fail to predict scene structure. + +In this work, we present DiffusionSfM, an end-to-end multi-view model that directly infers dense 3D geometry and cameras from multiple input images. Instead of inferring (depth-agnostic) rays per image patch (as in RayDiffusion [41]) or 3D points per pixel (as in DUSt3R [36]), DiffusionSfM effectively combines both to predict ray origins and endpoints per pixel, directly reporting both scene geometry (endpoints) and generalized cameras (rays). These can readily be converted back to traditional cameras [41]. Compared to RayDiffusion, our model directly predicts structure as well as motion at a finer scale (pixel-wise v.s. patchwise). Compared to DUSt3R, our model directly predicts motion as well as structure but, even more importantly, does so for $N$ views, eliminating the need for memory-intensive global alignment. To model uncertainty, we train a denoising diffusion model but find two key challenges that need to be addressed. First, diffusion models require (noisy) ground truth as input for training, but existing real datasets do not have known endpoints for all pixels due to missing depth in multi-view stereo. Second, the 3D coordinates of endpoints can be potentially unbounded, whereas diffusion models require normalized data. We develop mechanisms to overcome these challenges, leveraging additional "GT mask conditioning" as input to inform the model of missing input data and parameterizing 3D points in projective space instead of Euclidean space. We find these strategies allow us to learn accurate predictions for structure and motion. + +We train and evaluate DiffusionSfM on real-world and synthetic datasets [21, 24, 42] and find that it can infer accurate geometry and cameras for both object-centric and scene-level images (see Fig. 1). In particular, we find that DiffusionSfM yields more accurate camera estimates compared to prior work across these settings while also modeling the underlying uncertainty via the diffusion process. In summary, we show that DiffusionSfM can serve as a unified + +multi-view reasoning model for 3D geometry and cameras. + +# 2. Related Work + +Structure from Motion. Structure-from-Motion (SfM) systems [25] aim to simultaneously recover geometry and cameras given a set of input images. The typical SfM pipeline extracts pixel correspondences from keypoint matching [2, 16], and performs global bundle adjustment (BA) to optimize sparse 3D points and camera parameters by minimizing reprojection errors. Recently, SfM pipelines have been substantially enhanced by replacing classical subcomponents with learning-based methods, such as neural feature descriptors [6, 8], keypoint matching [15, 23, 30], and bundle adjustment [13, 32]. + +More recently, an emerging body of research aims to unify the various SfM subcomponents into an end-to-end neural framework. Notably, ACEZero [3] fits a single neural network to input images and learns pixel-aligned 3D coordinates in a self-supervised manner, while FlowMap [27] predicts per-frame cameras and depth maps using off-the-shelf optical flow as supervision. Though ACEZero and FlowMap are promising attempts to revolutionize SfM pipelines, they both register images incrementally and may suffer under large viewpoint changes. DUSt3R [36] directly regresses 3D pointmaps from image pairs and shows strong generalization ability [11, 21]. Building on DUSt3R, MASt3R [10] introduces a feature head, offering pixel-matching capabilities. MASt3R-SfM [7] is a more scalable SfM pipeline based on MASt3R. While these approaches [7, 10, 36] show impressive performance and robustness under sparse views, they are essentially pair-based, requiring sophisticated global alignment procedures to form a consistent estimate for more than two views. + +Pose Estimation with Global Reasoning. For the task of sparse-view pose estimation, learning-based methods equipped with global reasoning show favorable robustness where traditional SfM methods [25, 28] fail. This line of research includes energy-based [12, 40], regression-based [26], and diffusion-based pose estimators [34, 41]. Among them, diffusion-based methods show a better ability to handle uncertainty [41]. Closest to our work, RayDiffusion [41] leverages a denoising diffusion model [9, 19] with a patch-aligned (depth-unaware) ray representation to predict generic cameras. Our method goes further and pursues a generic representation for both geometry and cameras in the form of ray origins and endpoints for each pixel. In addition to resulting in a richer output, this joint geometry and pose prediction also yields improvements for pose estimation. + +# 3. Method + +Given a set of sparse (i.e., 2-8) input images, DiffusionSfM predicts the geometry and cameras of a 3D scene in a global + +![](images/e313265c3ef039eaa8c98aab188fac7c1ec90a35c2ae39191b416a12c793c328.jpg) +Figure 2. Method. Given sparse multi-view images as input, DiffusionSfM predicts pixel-wise ray origins and endpoints in a global frame (Sec. 3.1) using a denoising diffusion process (Sec. 3.2). For each image, we compute patch-wise embeddings with DINOv2 [18] and embed noisy ray origins and endpoints into latents using a single downsampling convolutional layer, ensuring alignment with the spatial footprint of the image embeddings. We implement a Diffusion Transformer architecture that predicts clean ray origins and endpoints from noisy samples. A convolutional DPT [20] head outputs full-resolution denoised ray origins and endpoints. To handle incomplete ground truth (GT) during training, we condition the model on a GT mask (Sec. 3.3). At inference, the GT mask is set to all ones, enabling the model to predict origins and endpoints for all pixels. The predicted ray origins and endpoints can be directly visualized in 3D or postprocessed to recover camera extrinsics, intrinsics, and multi-view consistent depth maps. + +coordinate frame. In Sec. 3.1, we propose to represent 3D scenes as dense pixel-aligned ray origins and endpoints. To predict such scene representations from sparse input images while modeling uncertainty, Sec. 3.2 proposes a denoising diffusion architecture. We then discuss some key practical challenges in training such a model in Sec. 3.3. + +# 3.1. 3D Scenes as Ray Origins and Endpoints + +Given an input image $\mathbf{I} \in \mathbb{R}^{H \times W \times 3}$ with a depth map $\mathbf{D} \in \mathbb{R}^{H \times W}$ , camera intrinsics $\mathbf{K} \in \mathbb{R}^{3 \times 3}$ , and world-to-camera extrinsics $\mathbf{T} \in \mathbb{R}^{4 \times 4}$ (equivalently, rotation $\mathbf{R} \in SO(3)$ and translation $\mathbf{t} \in \mathbb{R}^3$ ), each 2D image pixel $\mathbf{P}_{ij} = [u, v]$ corresponds to a ray that travels from the camera center $\mathbf{c}$ through the pixel's projected position on the image plane, terminating at the object's surface as specified by the depth map $\mathbf{D}$ . The endpoint of the ray associated with image pixel $\mathbf{P}_{ij}$ is given by: + +$$ +\mathbf {E} _ {i j} = \mathbf {T} ^ {- 1} h (\mathbf {D} _ {i j} \cdot \mathbf {K} ^ {- 1} [ u, v, 1 ] ^ {T}) \tag {1} +$$ + +where $h$ maps the 3D point into homogeneous coordinates. The shared ray origin $\mathbf{O}_{ij}$ for all pixels is equivalent to the camera center $\mathbf{c}$ , and can be computed as: + +$$ +\mathbf {O} _ {i j} = \mathbf {c} = h (- \mathbf {R} ^ {- 1} \mathbf {t}) \tag {2} +$$ + +In summary, we associate each image pixel with a ray origin and endpoint $\mathbf{S}_{ij} = \langle \mathbf{O}_{ij},\mathbf{E}_{ij}\rangle$ in world coordinates, describing the location of the observing camera and the observed 3D point on the object surface. Given a bundle of ray origins and endpoints, we can easily extract the corresponding camera pose [41]. + +Learning Over-Parameterized Representations. We represent 3D scenes and cameras using distributed ray origins $\mathbf{O}$ and endpoints $\mathbf{E}$ rather than global alternatives such as quaternions and translation vectors. This design is inspired by RayDiffusion [41], and it facilitates the use of the + +distributed deep features learned by state-of-the-art vision backbones, such as DINOv2 [18], which encode image information in a patch-wise manner. Notably, while ray origins $\mathbf{O}$ should ideally be identical across all pixels, we predict them densely alongside ray endpoints $\mathbf{E}$ . This encourages ray origins to remain close within the same image, providing implicit regularization during training. Practically, predicting both ray origins and endpoints simultaneously is easy to implement using a single projection head. + +# 3.2. DiffusionSfM + +We propose a denoising Diffusion Transformer (DiT) architecture [19] that predicts ray origins and endpoints (Sec. 3.1) via a denoising diffusion process. An overview of DiffusionSfM is given in Fig. 2. + +Diffusion Framework. Given pixel-aligned ray origins and endpoints $S = \text{stack}(\{\mathbf{S}^{(n)}\}_{n=1}^{N})$ associated with a set of $N$ input images, we apply a forward diffusion process [9, 29] that adds time-dependent Gaussian noise to them. Let $S_{t}$ denote the noisy ray origins and endpoints at timestep $t$ , where $S_{0}$ is the clean sample and $S_{T}$ (at the final diffusion step $T$ ) approximates pure Gaussian noise. The forward diffusion process is defined as: + +$$ +\mathcal {S} _ {t} = \sqrt {\bar {\alpha} _ {t}} \mathcal {S} _ {0} + \sqrt {1 - \bar {\alpha} _ {t}} \epsilon \tag {3} +$$ + +where $t \sim \mathrm{Uniform}(0, T]$ , $\epsilon \sim \mathcal{N}(0, I)$ , and $\bar{\alpha}_t$ follows a pre-defined noise schedule that controls the strength of added noise at each timestep. To perform the reverse diffusion process, which progressively reconstructs the clean sample given noisy observations, we train a diffusion model $f_\theta$ that takes $\mathcal{S}_t$ as input and optionally incorporates additional conditioning information $\mathcal{C}$ . The model is trained using the following loss function (with the " $x_0$ " objective): + +$$ +\mathcal {L} _ {\text {D i f f u s i o n}} = \mathbb {E} _ {t, S _ {0}, \epsilon} \| S _ {0} - f _ {\theta} (S _ {t}, t, \mathcal {C}) \| ^ {2} \tag {4} +$$ + +Architecture. We implement $f_{\theta}$ using a DiT [19] architecture conditioned on deep image features $\mathcal{C} \in \mathbb{R}^{N \times h \times w \times c_1}$ from DINOv2 [18], where $h$ and $w$ are the patch resolution and $c_1$ is the embedding dimension. To align pixels to the spatial information learned by DINOv2, we apply a convolutional layer with kernel size and stride equal to ViT patch size: this spatially downsamples the noisy ray origins and endpoints $S_t$ to match the DINOv2 features while increasing their feature dimension: + +$$ +\mathcal {F} = \operatorname {C o n v} \left(\mathcal {S} _ {t}\right) \in \mathbb {R} ^ {N \times h \times w \times c _ {2}} \tag {5} +$$ + +The combined DiT input is constructed by concatenating these two feature sets along the channel dimension: $\mathcal{F} \oplus \mathcal{C}$ . Within DiT, patch-wise features attend to others through self-attention [33]. To distinguish between different images and their respective patches, we apply sinusoidal positional encoding [33] based on image and patch indices. + +While the DiT operates on low-resolution features, our objective is to produce pixel-aligned dense ray origins and endpoints. To achieve this, we employ a DPT (Dense Prediction Transformer) [20] decoder, which takes intermediate feature maps from both DINoV2 and DiT as inputs. The DPT decoder progressively increases the feature resolution through several convolutional layers. The final ray origins and endpoints are decoded from the DPT output using a single linear layer. During inference, we apply the trained model in the reverse diffusion process to iteratively denoise a randomly initialized Gaussian sample. + +# 3.3. Practical Training Considerations + +Homogeneous Coordinates for Unbounded Geometry. Real-world scenes often exhibit significant variations in scale, both across different scenes and within a single scene. While normalizing by the average distance of 3D points [36] helps address cross-scene variation, within-scene variations remain. For instance, a background building may be much farther away than foreground elements, potentially resulting in extremely large coordinate values when generating ground-truth ray origins and endpoints. However, neural networks (especially diffusion models) tend to train most effectively when working with bounded inputs and outputs (e.g., between -1 and 1). To stabilize training across the large scale variations present in 3D scene datasets, we propose to represent ray origins and endpoints in homogeneous coordinates. Specifically, given any 3D point, we apply a homogeneous transform from $\mathbb{R}^3\to \mathbb{P}^3$ : + +$$ +(x, y, z) \rightarrow \frac {1}{w} (x, y, z, 1) \tag {6} +$$ + +where $w$ is an arbitrary scale factor. To encourage bounded coordinates in practice, we choose $w$ such that the homogeneous coordinate is unit-norm: + +$$ +w := \sqrt {x ^ {2} + y ^ {2} + z ^ {2} + 1} \tag {7} +$$ + +Unit normalization allows homogeneous coordinates to serve as a bounded representation for unbounded scene geometry. For example, $(x,y,z,0)$ is a point at infinity in the direction of $(x,y,z)$ . We find that this representation makes large coordinate values more tractable during training. + +Training with Incomplete Ground Truth. Many real-world datasets (e.g., CO3D [21] and MegaDepth [11]), provide only sparse point clouds, leading to incomplete depth information. This presents a significant challenge in diffusion training, as ground-truth (GT) depth values that are used to create clean ray endpoints often contain invalid or missing data. It is highly undesirable for these missing ray endpoints to be interpreted as part of the target distribution. Unlike regression models (e.g., DUSt3R [36]), which do not require (noisy) GT data as input and can simply apply a mask on the loss for supervision, diffusion models must handle incomplete inputs during training. + +To mitigate this issue, we further apply GT masks $\mathcal{M} \in \mathbb{R}^{N \times H \times W}$ to the DiT inputs, where zero values indicate pixels with invalid depth. During training, we multiply DiT inputs with GT masks element-wise, then concatenate along the channel dimension: $S_{t}' = (\mathcal{M} \cdot S_{t}) \oplus \mathcal{M}$ . Then, we only compute the diffusion loss in Eq. 4 over unmasked pixels. By implementing these strategies, we encourage the model to focus on regions with valid GT values during training. During inference, however, we would like the diffusion process to estimate ray endpoints at all pixels, so we always use GT masks with values set to one. + +Sparse-to-Dense Training. In practice, we find that training the entire model from scratch leads to slow convergence and suboptimal performance. To address this, we propose a sparse-to-dense training approach. First, we train a sparse version of the model, where the DPT decoder is removed, and the output ray origins and endpoints have the same spatial resolution as the DINOv2 features. Unlike Eq. 5, no spatial downsampling is required, so this sparse model uses a single linear layer to embed the noisy ray origins and endpoints. Once the sparse model is trained, we initialize the dense model DiT with the learned weights from the sparse model. This two-stage approach significantly improves performance; see the supplementary material for comparisons. + +# 4. Experiments + +# 4.1. Experimental Setup + +Datasets. We introduce two model variants, each trained on different datasets. (1) DiffusionSfM-CO3D: Following prior work [12, 41], we train and evaluate our model on the CO3D dataset [21], which consists of turntable video sequences of various object categories. Specifically, we train on 41 object categories and evaluate on both these seen categories and an additional 10 unseen categories to assess generalization. (2) DiffusionSfM: This variant is trained on the + +![](images/d6e463bb54fb1e9753a107c4e0f00652fabfafb5aeb6a00a0e8c75bcf8a0d463.jpg) +Input Images /GT Point Clouds + +![](images/dc5f60efbcfe9c7755289675788486c4578c9f213d2b6042429e3e03c31ae41f.jpg) + +![](images/4a90c02cbf86ede1c3b9f9128f7369efafe2004fac7f244144acbef7b880efc4.jpg) +DiffusionSfM + +![](images/4dbeed7353c79ac7b9a429bdb597e8c4bf321a4dea76408dad8d2d25a4cb3311.jpg) +RD + MoGe + +![](images/3c4d4f0156271974c5a1bea2211d4c0da5b8a04dce2f24b68c677ecb90c28e64.jpg) +DUSt3R + +![](images/08b9d18a8fd24f92149aabec1779e2696ba49c6f961ae07dcdb2a0eb2a56a1d2.jpg) + +![](images/6d07fe01df1484bab4bf2ec9045876531d07cc800c093177a7bbbbbfac0de292.jpg) + +![](images/21cddbbf7d3c3a2aa6c186d5790bb68351b7e18e5fec658f160d0f440c0ebba2.jpg) + +![](images/4467112fd81e570e98956508a4e57480691fb0ba442c38bbf2627e053ef2ac55.jpg) + +![](images/272dd89e3cebb5972006b1e754848a814b96b262781b027efbac94167d73da4f.jpg) + +![](images/f9d752bc2a72574ad8da4962e3c75c5038c52407176760b4817e89875dc60734.jpg) + +![](images/77e3edccd4250d49af8e4e40afe6cba4c9d6793ae40e2556aa588e44db248cf8.jpg) + +![](images/764b33e09ba83273a035143014a202821327b88eae6f527e802b136b7abbff21.jpg) + +![](images/abec2e30ad827dfb18d995acedef81d339f0eae11fb77e60e29a29e8e3f6f10d.jpg) + +![](images/083a0c1897454b692a6d69688d3baf77dee68c4161fa70f086982461b4a28fe5.jpg) + +![](images/15555cfb9f16f6fdc453589fc0c91b42a2258d1260a60ff53f08904d6c5ffc59.jpg) + +![](images/d9ba0182fe76e9bcf638276fbe614ef41034a2b339947df162dd809ad1cd3fa0.jpg) + +![](images/42ead853a68a77308aba6a032385977d1aee9b389b6819fe8f30cd69f3b1bd53.jpg) + +![](images/3d56dfda9c0cd215776b017ae9736d19c24d055bcab58416313922bee5a94489.jpg) + +![](images/b1b3bbe97ed1d5d38efbfd8c6a271fd8662c88298f9063cb5d973f46e0ab718e.jpg) + +![](images/c44111bf71e33372cf04e561079591ede87ef0cb656da94d7fb85cf5cd0f6dba.jpg) + +![](images/621af54fe4b9ca0aa56cb232e2c29f811110ddb5e15082e17e506a654080d459.jpg) + +![](images/75308784803a2b539f04991f58dc5590242f78993005180f68724760b4e2d4f9.jpg) + +![](images/8ff9e2401bc30c177a123e84029058340270d40a3e6b1993cc4d906e2ce533e5.jpg) + +![](images/64b87ce018823ad58c1359ca2b1ff31994f93d3890ad8717504cbb3578ac1657.jpg) + +![](images/c919abb7dcae4e0ccf25434d3735e9779fcdd4eeae41a41a769f2dbcea89668c.jpg) + +![](images/a9ea49154a5cddebf7b1215c95540c58c714ca3a87f5f9090d30a8f5a9a5356c.jpg) +Figure 3. Qualitative Comparison on Camera Pose Accuracy and Predicted Geometry. For each method, we plot the ground-truth cameras in black and the predicted cameras in other colors. DiffusionSfM demonstrates robust performance even with challenging inputs. Compared to DUSt3R, which sometimes fails to register images in a consistent manner, DiffusionSfM consistently yields a coherent global prediction. Additionally, while we observe that DUSt3R can predict highly precise camera rotations, it often struggles with camera centers (see the backpack example). Input images depicting scenes are out-of-distribution for RayDiffusion, as it is trained on CO3D only. + +![](images/1737d179f0ff07dcf799bb9ca29f91d33c878cd8e84667b551b4f902f45aa969.jpg) + +![](images/b07132582448c16a63a6b68ee425225bc70df6e576c168687035e4f1ffa818fa.jpg) + +![](images/355211bdc5d80708987da0be50e094a64357cea1d97bc362e30bbd5e8f491890.jpg) + +![](images/052155bee2775d35833066115b297323f753cce3a22d8fe57f655f5794fe4c81.jpg) + +![](images/9e2d0801addb462fc02b573b35ec15afd43788406b3ddcbf55c1a5652275a7ab.jpg) + +![](images/95779d7f8c4d99450912c29e23c3a409997915c022968df709dfd3b6abb7c648.jpg) + +datasets used for DUSt3R [36], excluding Waymo [31] due to its excessively sparse depth maps. The included datasets are Habitat [24], CO3D [21], ScanNet++ [39], ArkitScenes [1], Static Scenes 3D [17], MegaDepth [11], and BlendedMVS [38]. We follow the DUSt3R official repository [37] + +guidelines to extract image pairs, ensuring reasonable overlap between paired images. To construct multi-view data given these pre-computed pairs, we iteratively sample images using an adjacency matrix, maintaining sufficient overlap with the selected set. Beyond CO3D, this model variant + +![](images/8113542d26e8415e7fccb43a4ab137fe288e9ddfb3159e473e2445238d6cf63d.jpg) +Input Images + +![](images/933b09eb7e3527150f3be3f259b770ea88c020ef8b967f7a83a4d6eafe070ecc.jpg) +DiffusionSfM + +![](images/ebb549fcdbcce39636c90442061595ecb9be761d0dac36f498c4259fa3e9299b.jpg) +RayDiffusion + +![](images/3d9b95e9ad7374161e5cbf61c41771e2e6b56aedb7c1c0789d6480c724ed3950.jpg) +DUSt3R + +![](images/c9e3727a7835cb21302b9686d7fc17c3b5fa3dc25d6a1046592ee7a962cf2cad.jpg) + +![](images/3ba083c2b95cf328bae57df050f1ef358611952760041f8e249fa443105d4deb.jpg) + +![](images/391548a854ffbf9d9e3b6c9a89f4c7f4a8c2303a99b0f5684284a04001ea7fd6.jpg) + +![](images/71982fd3c12d0817f39efa883fc6e901683dd4fcb9958618f0e0e0c14e24d0ec.jpg) + +![](images/01f965a7ac2be332782e20e3ebbaef854eed14bc732b6658975749ecb53a0bb9.jpg) + +![](images/3a0473971df79c5449389de83b9765f9583258dd642a4923b3125fc4447d8398.jpg) + +![](images/fc70de8b11828fc1a07930f61b93d9c100c1903b7fb41fcd5d39ea6065293654.jpg) + +![](images/e62c7dc2190fad5f4b7bdae298e3183cc133b3b08212eed26461b84525924ed5.jpg) + +![](images/12bc1a5e9df8ccb453a2c73ad7e51bab942f03cf2f0f7966ec16278b98e03951.jpg) +Figure 4. Additional Qualitative Results on Predicted Camera Poses. DiffusionSfM shows robustness to ambiguous patterns in inputs. + +![](images/43d9c1f6b2fb34cfb80eb04d4a1eace850a5c28b1428f2923e0531d60d3fcbfe.jpg) + +![](images/0fb6fb31f2c126f48b00a94419cbada66bd471bb4696337619753970c8eb7ed5.jpg) + +is also evaluated on Habitat and RealEstate10k [42]. + +Baselines and Metrics. To evaluate camera pose accuracy in the sparse-view setup, we compare with RayDiffusion [41] and DUSt3R [36], along with previous methods [12, 22, 34]. DUSt3R is trained initially on a mixture of eight datasets, while most other baselines are trained only on CO3D. For a fair and comprehensive comparison, we re-train DUSt3R on the 41-10 split of CO3D, using the authors' official implementation [37] and hyperparameters (referred to as DUSt3R-CO3D). To evaluate camera predictions, we follow prior work [41] and convert predicted rays back to traditional extrinsic matrices and report two pose accuracy metrics: (1) Camera Rotation Accuracy which compares the predicted relative camera rotation between images against ground truth and (2) Camera Center Accuracy which compares predicted camera centers to the ground truth after a similarity alignment. To evaluate the estimated geometry (i.e., ray endpoints), we report Chamfer Distance (CD). Additionally, to explore whether combining an off-the-shelf sparse-view pose estimation method (RayDiffusion) with a monocular depth estimation model (MoGe [35]) is sufficient to infer 3D scene geometry from + +multiple images, we introduce a new baseline, $\mathsf{RD} + \mathsf{MoGe}$ . We include more details in the supplementary. Unless otherwise specified, all evaluations use 2-8 input images. + +# 4.2. Evaluation on CO3D + +Camera Pose Accuracy. We present the quantitative results in Tab. 1. For camera rotation accuracy, both versions of DiffusionSfM outperform all baselines, except that DUSt3R achieves slightly higher overall accuracy than DiffusionSfM-CO3D on unseen categories – likely due to its training on more data. For camera center accuracy, our approach consistently outperforms all other methods. We hypothesize that these gains stem from our explicit modeling of camera centers through ray origin prediction. The qualitative results in Fig. 3 illustrate that DiffusionSfM produces robust predictions given challenging inputs, whereas DUSt3R produces inaccurate results. We attribute this improvement to our model's probabilistic modeling capability derived from diffusion, as well as its multi-view reasoning abilities, which together effectively handle these challenging scenarios. While we observe that DUSt3R can predict highly precise camera rotations, it often struggles with cam- + +
# of ImagesRotation Accuracy (↑, @ 15°)Center Accuracy (↑, @ 0.1)
2345678234567
Seen CategoriesCOLMAP [22]30.728.426.526.827.028.130.610034.523.818.915.614.5
PoseDiffusion [34]75.776.476.877.478.078.778.810077.569.765.963.762.8
RelPose++ [12]81.882.884.184.784.985.385.510085.078.074.271.970.3
RayDiffusion [41]91.892.492.692.993.193.393.310094.290.587.886.285.0
DUSt3R-CO3D [36]86.787.988.088.288.688.888.910092.086.883.882.081.1
DUSt3R [36]91.792.793.393.693.894.094.310093.085.781.979.677.8
DiffusionSfM-CO3D93.494.094.594.895.095.295.110095.993.692.291.290.7
DiffusionSfM92.493.594.094.594.795.195.210095.392.991.991.290.8
Unseen CategoriesCOLMAP [22]34.531.831.031.732.735.038.510036.025.520.017.917.6
PoseDiffusion [34]63.264.264.265.766.267.067.710063.650.545.743.041.2
RelPose++ [12]69.871.171.972.873.874.474.910070.658.853.450.447.8
RayDiffusion [41]83.585.686.386.987.287.588.110087.781.177.074.172.4
DUSt3R-CO3D [36]79.881.582.682.783.083.383.710083.677.271.870.068.1
DUSt3R [36]90.892.693.693.693.893.693.410087.979.874.371.769.4
DiffusionSfM-CO3D90.491.292.793.093.193.393.510091.187.785.383.782.7
DiffusionSfM91.592.893.693.994.194.294.410092.487.986.285.384.1
+ +era centers (see the backpack example). Additional qualitative results on camera accuracy are provided in Fig.4. + +Predicted Geometry. To evaluate predicted geometry, we compute Chamfer Distance (CD) and show comparisons against baselines in Tab. 2. We compute CD in two setups (with and without foreground object mask), and find that DiffusionSfM-CO3D performs best without foreground masking. In this setup, the predicted ray endpoints corresponding to the background image pixels tend to have larger coordinate values than foreground ones, and therefore dominate CD. This result indicates that our model provides more accurate predictions for complex image backgrounds. In terms of CD with masking, DUSt3R achieves the best result, while our two model variants outperform DUSt3R-CO3D. See Fig. 3 for visualization. We also notice that naively combining RayDiffusion poses with estimated depths from MoGe does not work well, as the inferred depths are potentially inconsistent across different views, thus resulting in degraded performance. + +# 4.3. Evaluation on Scene-Level Datasets + +Beyond the object-centric CO3D dataset [21], we compare DiffusionSfM against DUSt3R on two scene-level datasets: Habitat (in-distribution) [24] and RealEstate10k (out-of-distribution) [42]. The results are presented in Tab. 3. While DiffusionSfM achieves comparable camera rotation accuracy to DUSt3R, it consistently predicts more accurate camera centers, aligning with our findings on CO3D. + +Table 1. Camera Rotation and Center Accuracy on CO3D. On the left, we report the proportion of relative camera rotations within $15^{\circ}$ of the ground truth. On the right, we report the proportion of camera centers within $10\%$ of the scene scale, relative to the ground truth. To align the predicted camera centers to ground truth, we apply an optimal similarity transform $(s, \mathbf{R}, \mathbf{t})$ , hence the alignment is perfect at $N = 2$ but worsens with more images. DiffusionSfM outperforms all other methods for camera center accuracy, and outperforms all methods trained on equivalent data for rotation accuracy. + +
# of Images2345678
RD*+MoGe [41]0.0590.0640.0710.0620.0630.0610.061
DUSt3R* [36]0.0360.0370.0400.0400.0370.0360.039
DUSt3R [36]0.0210.0230.0240.0240.0250.0250.023
DiffusionSfM*0.0190.0210.0220.0220.0210.0210.021
DiffusionSfM0.0250.0250.0250.0250.0260.0250.028
RD*+MoGe [41]0.0710.0750.0680.0670.0660.0640.064
DUSt3R* [36]0.0380.0360.0360.0360.0340.0330.034
DUSt3R [36]0.0230.0220.0190.0200.0190.0200.020
DiffusionSfM*0.0280.0250.0240.0240.0240.0230.022
DiffusionSfM0.0270.0240.0230.0220.0220.0220.021
+ +Table 2. Chamfer Distance $(\downarrow)$ on CO3D Unseen Categories. Top: CD computed on all scene points. Bottom: CD computed on foreground points only. Models marked with “*” are trained on CO3D only, while those without are trained on multiple datasets. Note that top and bottom values are not directly comparable, as each ground-truth point cloud is individually normalized. + +# 4.4. Inference Speed + +Though our method requires iterative diffusion denoising at inference, we can speed this up by performing early stopping. Specifically, we can treat the $x_0$ -prediction from early timesteps as our output instead of iterating over all denoising timesteps. Consistent with observations by Zhang et al. [41], this in fact yields more accurate predictions than the final-step diffusion outputs. As a result we only require 10 denoising diffusion timesteps for inference, taking 1.91 + +
Hab.DUSt3R97.0/10095.0/97.694.3/95.094.2/93.1
DiffusionSfM92.6/10093.7/99.194.2/98.894.9/98.6
RE10DUSt3R98.1/10097.7/68.797.6/57.997.7/53.3
DiffusionSfM97.8/10097.5/75.697.7/66.297.9/61.9
+ +Table 3. Camera Rotation and Center Accuracy on Two Scene-Level Datasets. Top: Habitat (2-5 views). Bottom: RealEstate10k (2, 4, 6, 8 views). Each grid reports camera rotation accuracy (left, $\uparrow$ ) and center accuracy (right, $\uparrow$ ). While DiffusionSfM performs on par with DUSt3R in rotation accuracy, it consistently surpasses DUSt3R in center accuracy. + +
# of Images2345678
DiffusionSfM*0.0190.0210.0220.0220.0210.0210.021
w/o Mask0.0200.0220.0230.0230.0220.0220.024
+ +Table 4. Ablation Study on GT Mask Conditioning for CO3D Unseen Categories. We assess the effect of replacing GT mask conditioning with depth interpolation in our CO3D variant (DiffusionSfM*), by reporting the CD for predicted geometry. Incorporating mask conditioning to indicate missing data during training improves geometry quality. + +seconds on a single A5000 GPU with 8 input images. In contrast, DUSt3R takes 8.73 seconds to run the complete pairwise inference and global alignment procedure. We provide additional analysis in the supplementary material. + +# 4.5. Ablation Study + +Homogeneous Coordinates. The use of homogeneous coordinates for ray origins and endpoints is crucial for stable model training. To assess its impact, we replace the proposed homogeneous representation with standard 3D coordinates in $\mathbb{R}^3$ . Our experiments show that this variant is difficult to train and fails to converge. Further details on this experiment are provided in the supplementary material. + +GT Mask Conditioning. To evaluate the effectiveness of the proposed GT mask conditioning, we train a baseline model without using this strategy. For missing values in the ground-truth depth maps, we use nearest-neighbor interpolation to fill in invalid pixels. This experiment is conducted on the CO3D dataset, with results presented in Tab. 4. The findings show that removing GT mask conditioning consistently degrades predicted geometry across varying numbers of input views, even when the loss is still masked. While interpolation can effectively fill missing depth within object regions, it often introduces substantial noise in the background (e.g., the sky). This noise negatively impacts diffusion model training, as the entire ray origin and endpoint maps are used as input. We anticipate an even larger performance gap on non-object-centric datasets with more outdoor scenes, such as MegaDepth [11]. + +![](images/73e85658c1c87772133020c3736a2822b53d306c91f74b826fc325128df9097e.jpg) +Figure 5. Multi-modality of DiffusionSfM. We show two distinct samples from DiffusionSfM, starting from the same input images but with different random noise. Sample 1 explains the input images by putting all flowers on the left side, while Sample 2 places one flower on each side (note the difference in the green camera's viewpoint). DiffusionSfM is able to predict multi-modal geometry distributions when the scene layout is ambiguous in the inputs. + +# 4.6. Multi-modality from Multiple Sampling + +Diffusion models enable the generation of diverse samples from challenging input images. For instance, in Fig. 5, the vase exhibits symmetric patterns, and we present two distinct predicted endpoints from DiffusionSfM, each offering a different interpretation of the flowers in the image. Compared to regression models such as DUSt3R [8], DiffusionSfM is better suited for handling uncertainty – an inherent aspect of our task – where a sparse set of input images can correspond to multiple plausible 3D scene geometry. + +# 5. Discussion + +We present DiffusionSfM and demonstrate that it recovers accurate predictions of both cameras and geometry from multi-view inputs. Although our results are promising, several challenges and open questions remain. + +Notably, DiffusionSfM employs a pixel-space diffusion model, in contrast to the latent-space models adopted by state-of-the-art T2I generative systems. Operating in pixel space may require greater model capacity, yet our current model remains relatively small - potentially explaining the noisy patterns observed along object boundaries. Learning an expressive latent space for ray origins and endpoints by training a VAE could be a promising direction for future work. In addition, the compute requirement in multi-view transformers scales quadratically with the number of input images: one would require masked attention to deploy systems like ours for a large set of input images. + +Despite these challenges, we believe that our work highlights the potential of a unified approach for multi-view geometry tasks. We envision that our approach can be built upon to train a common system across related geometric tasks, such as SfM (input images with unknown origins and endpoints), registration (some images have known origins and endpoints, whereas others don't), mapping (known rays but unknown endpoints), and view synthesis (unknown pixel values for known rays). + +Acknowledgements: We thank the members of the Physical Perception Lab at CMU for their valuable discussions, and extend special thanks to Yanbo Xu for his insights on diffusion models. + +This work used Bridges-2 [4] at Pittsburgh Supercomputing Center through allocation CIS240166 from the Advanced Cyberinfrastructure Coordination Ecosystem: Services & Support (ACCESS) program, which is supported by National Science Foundation grants #2138259, #2138286, #2138307, #2137603, and #2138296. This work was supported by Intelligence Advanced Research Projects Activity (IARPA) via Department of Interior/Interior Business Center (DOI/IBC) contract number 140D0423C0074. The U.S. Government is authorized to reproduce and distribute reprints for Governmental purposes notwithstanding any copyright annotation thereon. Disclaimer: The views and conclusions contained herein are those of the authors and should not be interpreted as necessarily representing the official policies or endorsements, either expressed or implied, of IARPA, DOI/IBC, or the U.S. Government. + +# References + +[1] Gilad Baruch, Zhuoyuan Chen, Afshin Dehghan, Tal Dimry, Yuri Feigin, Peter Fu, Thomas Gebauer, Brandon Joffe, Daniel Kurz, Arik Schwartz, et al. Arkitsscenes: A diverse real-world dataset for 3d indoor scene understanding using mobile rgb-d data. In NeurIPS D&B, 2021. 5 +[2] Herbert Bay, Andreas Ess, Tinne Tuytelaars, and Luc Van Gool. Surf: Speeded up robust features. In ECCV, 2006. 2 +[3] Eric Brachmann, Jamie Wynn, Shuai Chen, Tommaso Cavallari, Aron Monszpart, Daniyar Turmukhambetov, and Victor Adrian Prisacariu. Scene coordinate reconstruction: posing of image collections via incremental learning of a relocalizer. In ECCV, 2024. 2 +[4] Shawn T Brown, Paola Buitrago, Edward Hanna, Sergiu Sanielevici, Robin Scibek, and Nicholas A Nystrom. Bridges-2: A platform for rapidly-evolving and data intensive research. In Practice and experience in advanced research computing, 2021. 9 +[5] Ruojin Cai, Joseph Tung, Qianqian Wang, Hadar Averbuch-Elor, Bharath Hariharan, and Noah Snavely. Doppelgangers: Learning to disambiguate images of similar structures. In ICCV, 2023. 1 +[6] Daniel DeTone, Tomasz Malisiewicz, and Andrew Rabinovich. Superpoint: Self-supervised interest point detection and description. In CVPR, 2018. 1, 2 +[7] Bardienus Duisterhof, Lojze Zust, Philippe Weinzaepfel, Vincent Leroy, Yohann Cabon, and Jerome Revaud. Mast3r-sfm: a fully-integrated solution for unconstrained structure-from-motion. In 3DV, 2025. 2 +[8] Johan Edstedt, Qiyu Sun, Georg Bokman, Marten Wadenback, and Michael Felsberg. RoMa: Robust Dense Feature Matching. In CVPR, 2024. 2, 8 +[9] Jonathan Ho, Ajay Jain, and Pieter Abbeel. Denoising diffusion probabilistic models. In NeurIPS, 2020. 2, 3 + +[10] Vincent Leroy, Yohann Cabon, and Jérôme Revaud. Grounding image matching in 3d with mast3r, 2024. ECCV. 1, 2 +[11] Zhengqi Li and Noah Snavely. Megadepth: Learning single-view depth prediction from internet photos. In CVPR, 2018. 2,4,5,8 +[12] Amy Lin, Jason Y Zhang, Deva Ramanan, and Shubham Tulsiani. Relpose++: Recovering 6d poses from sparse-view observations. In 3DV, 2024. 2, 4, 6, 7 +[13] Philipp Lindenberger, Paul-Edouard Sarlin, Viktor Larsson, and Marc Pollefeys. Pixel-perfect structure-from-motion with featuremetric refinement. In ICCV, 2021. 2 +[14] Philipp Lindenberger, Paul-Edouard Sarlin, Viktor Larsson, and Marc Pollefeys. Pixel-Perfect Structure-from-Motion with Featuremetric Refinement. In ICCV, 2021. 1 +[15] Philipp Lindenberger, Paul-Edouard Sarlin, and Marc Pollefeys. Lightglue: Local feature matching at light speed. In ICCV, 2023. 2 +[16] David G Lowe. Distinctive image features from scale-invariant keypoints. In IJCV, 2004. 2 +[17] Nikolaus Mayer, Eddy Ilg, Philip Hausser, Philipp Fischer, Daniel Cremers, Alexey Dosovitskiy, and Thomas Brox. A large dataset to train convolutional networks for disparity, optical flow, and scene flow estimation. In CVPR, 2016. 5 +[18] Maxime Oquab, Timothee Darcet, Theo Moutakanni, Huy Vo, Marc Szafraniec, Vasil Khalidov, Pierre Fernandez, Daniel Haziza, Francisco Massa, Alaaeldin El-Nouby, et al. Dinov2: Learning robust visual features without supervision. In TMLR, 2024. 3, 4 +[19] William Peebles and Saining Xie. Scalable diffusion models with transformers. In ICCV, 2023. 2, 3, 4 +[20] René Ranftl, Alexey Bochkovskiy, and Vladlen Koltun. Vision transformers for dense prediction. In ICCV, 2021. 3, 4 +[21] Jeremy Reizenstein, Roman Shapovalov, Philipp Henzler, Luca Sbordone, Patrick Labatut, and David Novotny. Common objects in 3d: Large-scale learning and evaluation of real-life 3d category reconstruction. In ICCV, 2021. 2, 4, 5, 7 +[22] Paul-Edouard Sarlin, Cesar Cadena, Roland Siegwart, and Marcin Dymczyk. From coarse to fine: Robust hierarchical localization at large scale. In CVPR, 2019. 6, 7 +[23] Paul-Edouard Sarlin, Daniel DeTone, Tomasz Malisiewicz, and Andrew Rabinovich. Superglue: Learning feature matching with graph neural networks. In CVPR, 2020. 1, 2 +[24] Manolis Savva, Abhishek Kadian, Oleksandr Maksymets, Yili Zhao, Erik Wijmans, Bhavana Jain, Julian Straub, Jia Liu, Vladlen Koltun, Jitendra Malik, et al. Habitat: A platform for embodied ai research. In ICCV, 2019. 2, 5, 7 +[25] Johannes L Schonberger and Jan-Michael Frahm. Structure-from-motion revisited. In CVPR, 2016. 1, 2 +[26] Samarth Sinha, Jason Y Zhang, Andrea Tagliasacchi, Igor Gilitschenski, and David B Lindell. Sparsepose: Sparse-view camera pose regression and refinement. In CVPR, 2023. 2 +[27] Cameron Smith, David Charatan, Ayush Tewari, and Vincent Sitzmann. Flowmap: High-quality camera poses, intrinsics, and depth via gradient descent. In 3DV, 2025. 2 + +[28] Noah Snavely, Steven M Seitz, and Richard Szeliski. Photo tourism: exploring photo collections in 3d. In ACM SIGGRAPH, 2006. 2 +[29] Jiaming Song, Chenlin Meng, and Stefano Ermon. Denoising diffusion implicit models. In ICLR, 2021. 3 +[30] Jiaming Sun, Zehong Shen, Yuang Wang, Hujun Bao, and Xiaowei Zhou. Loftr: Detector-free local feature matching with transformers. In CVPR, 2021. 2 +[31] Pei Sun, Henrik Kretzschmar, Xerxes Dotiwalla, Aurelien Chouard, Vijaysai Patnaik, Paul Tsui, James Guo, Yin Zhou, Yuning Chai, Benjamin Caine, et al. Scalability in perception for autonomous driving: Waymo open dataset. In CVPR, 2020. 5 +[32] Chengzhou Tang and Ping Tan. Ba-net: Dense bundle adjustment network. In ICLR, 2019. 2 +[33] Ashish Vaswani, Noam Shazeer, Niki Parmar, Jakob Uszkoreit, Llion Jones, Aidan N Gomez, Lukasz Kaiser, and Illia Polosukhin. Attention is all you need. In NeurIPS, 2017. 4 +[34] Jianyuan Wang, Christian Rupprecht, and David Novotny. Posediffusion: Solving pose estimation via diffusion-aided bundle adjustment. In ICCV, 2023. 2, 6, 7 +[35] Ruicheng Wang, Sicheng Xu, Cassie Dai, Jianfeng Xiang, Yu Deng, Xin Tong, and Jiaolong Yang. Moge: Unlocking accurate monocular geometry estimation for open-domain images with optimal training supervision. In CVPR, 2025. 6 +[36] Shuzhe Wang, Vincent Leroy, Yohann Cabon, Boris Chidlovskii, and Jerome Revaud. Dust3r: Geometric 3d vision made easy. In CVPR, 2024. 1, 2, 4, 5, 6, 7 +[37] Shuzhe Wang, Vincent Leroy, Yohann Cabon, Boris Chidlovskii, and Jerome Revaud. Dust3r: Geometric 3d vision made easy, 2024. https://github.com/naver/dust3r. 5, 6 +[38] Yao Yao, Zixin Luo, Shiwei Li, Jingyang Zhang, Yufan Ren, Lei Zhou, Tian Fang, and Long Quan. Blendedmvs: A large-scale dataset for generalized multi-view stereo networks. In CVPR, 2020. 5 +[39] Chandan Yeshwanth, Yueh-Cheng Liu, Matthias Nießner, and Angela Dai. Scannet++: A high-fidelity dataset of 3d indoor scenes. In ICCV, 2023. 5 +[40] Jason Y Zhang, Deva Ramanan, and Shubham Tulsiani. Relapse: Predicting probabilistic relative rotation for single objects in the wild. In ECCV, 2022. 2 +[41] Jason Y. Zhang, Amy Lin, Moneish Kumar, Tzu-Hsuan Yang, Deva Ramanan, and Shubham Tulsiani. Cameras as rays: Sparse-view pose estimation via ray diffusion. In ICLR, 2024. 2, 3, 4, 6, 7 +[42] Tinghui Zhou, Richard Tucker, John Flynn, Graham Fyffe, and Noah Snavely. Stereo magnification: Learning view synthesis using multiplane images. In ACM SIGGRAPH, 2018. 2, 6, 7 \ No newline at end of file diff --git a/2025/DiffusionSfM_ Predicting Structure and Motion via Ray Origin and Endpoint Diffusion/images.zip b/2025/DiffusionSfM_ Predicting Structure and Motion via Ray Origin and Endpoint Diffusion/images.zip new file mode 100644 index 0000000000000000000000000000000000000000..dda6de230656240ec6052df615a97f8d106e9857 --- /dev/null +++ b/2025/DiffusionSfM_ Predicting Structure and Motion via Ray Origin and Endpoint Diffusion/images.zip @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:41e4c1917e8b864777f3dc7db4d263c84077aaaedce7f5d62d443c84c21c31d8 +size 799706 diff --git a/2025/DiffusionSfM_ Predicting Structure and Motion via Ray Origin and Endpoint Diffusion/layout.json b/2025/DiffusionSfM_ Predicting Structure and Motion via Ray Origin and Endpoint Diffusion/layout.json new file mode 100644 index 0000000000000000000000000000000000000000..95be1111f8c159b1b50b02ec9024c1ecbebc63b6 --- /dev/null +++ b/2025/DiffusionSfM_ Predicting Structure and Motion via Ray Origin and Endpoint Diffusion/layout.json @@ -0,0 +1,8602 @@ +{ + "pdf_info": [ + { + "para_blocks": [ + { + "bbox": [ + 158, + 103, + 452, + 140 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 158, + 103, + 452, + 140 + ], + "spans": [ + { + "bbox": [ + 158, + 103, + 452, + 140 + ], + "type": "text", + "content": "DiffusionSfM: Predicting Structure and Motion via Ray Origin and Endpoint Diffusion" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 104, + 161, + 506, + 190 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 104, + 161, + 506, + 190 + ], + "spans": [ + { + "bbox": [ + 104, + 161, + 506, + 190 + ], + "type": "text", + "content": "Qitao Zhao, Amy Lin, Jeff Tan, Jason Y. Zhang, Deva Ramanan, Shubham Tulsiani Carnegie Mellon University" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 188, + 190, + 422, + 204 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 188, + 190, + 422, + 204 + ], + "spans": [ + { + "bbox": [ + 188, + 190, + 422, + 204 + ], + "type": "text", + "content": "Project page: qitaozhao.github.io/DiffusionSfM" + } + ] + } + ], + "index": 4 + }, + { + "type": "image", + "bbox": [ + 56, + 213, + 555, + 395 + ], + "blocks": [ + { + "bbox": [ + 56, + 213, + 555, + 395 + ], + "lines": [ + { + "bbox": [ + 56, + 213, + 555, + 395 + ], + "spans": [ + { + "bbox": [ + 56, + 213, + 555, + 395 + ], + "type": "image", + "image_path": "ac31d7718dc556e701eba8b776c1f3e30bb00ad6f957b560bb72e53affc2382c.jpg" + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 54, + 402, + 555, + 460 + ], + "lines": [ + { + "bbox": [ + 54, + 402, + 555, + 460 + ], + "spans": [ + { + "bbox": [ + 54, + 402, + 555, + 460 + ], + "type": "text", + "content": "Figure 1. DiffusionSfM. Top: Given a set of multi-view images (left), DiffusionSfM represents scene geometry and cameras (right) as pixel-wise ray origins and endpoints in a global frame. It learns a denoising diffusion model to infer these elements directly from multiview inputs. Unlike traditional Structure-from-Motion (SfM) pipelines, which separate pairwise reasoning and global optimization into two stages, our approach unifies both into a single end-to-end multi-view reasoning framework. Bottom: Example results of inferred scene geometry and cameras for two distinct settings: a real-world outdoor scene (left) and a synthetic indoor scene (right)." + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "image_caption" + } + ], + "index": 5 + }, + { + "bbox": [ + 152, + 473, + 200, + 486 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 152, + 473, + 200, + 486 + ], + "spans": [ + { + "bbox": [ + 152, + 473, + 200, + 486 + ], + "type": "text", + "content": "Abstract" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 54, + 509, + 297, + 701 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 54, + 509, + 297, + 701 + ], + "spans": [ + { + "bbox": [ + 54, + 509, + 297, + 701 + ], + "type": "text", + "content": "Current Structure-from-Motion (SfM) methods typically follow a two-stage pipeline, combining learned or geometric pairwise reasoning with a subsequent global optimization step. In contrast, we propose a data-driven multi-view reasoning approach that directly infers 3D scene geometry and camera poses from multi-view images. Our framework, DiffusionSfM, parameterizes scene geometry and cameras as pixel-wise ray origins and endpoints in a global frame and employs a transformer-based denoising diffusion model to predict them from multi-view inputs. To address practical challenges in training diffusion models with missing data and unbounded scene coordinates, we introduce specialized mechanisms that ensure robust learning. We empirically validate DiffusionSfM on both synthetic and real datasets, demonstrating that it outperforms classical and learning-based approaches while naturally modeling uncertainty." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 315, + 473, + 394, + 486 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 315, + 473, + 394, + 486 + ], + "spans": [ + { + "bbox": [ + 315, + 473, + 394, + 486 + ], + "type": "text", + "content": "1. Introduction" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 313, + 498, + 556, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 498, + 556, + 715 + ], + "spans": [ + { + "bbox": [ + 313, + 498, + 556, + 715 + ], + "type": "text", + "content": "The task of recovering structure (geometry) and motion (cameras) from multi-view images has long been a focus of the computer vision community, with typical pipelines [25] performing pairwise correspondence estimation followed by global optimization. While classical methods relied on hand-designed features, matching, and optimization, there has been a recent shift towards incorporating learning-based alternatives [5, 6, 14, 23]. More recently, the widely influential DUSt3R [36] advocates for predicting pairwise 3D pointmaps (instead of only correspondences), demonstrating that this can yield accurate dense geometry and cameras. In order to reconstruct more than two views, DUSt3R (and its variants [10]) still require a global optimization reminiscent of classic bundle adjustment. While these methods, both classical and learning-based, have led to impressive improvements in SfM, the overall approach is largely unchanged – learned or geometric pairwise reasoning followed by global optimization. In this work, we" + } + ] + } + ], + "index": 10 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "spans": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "type": "text", + "content": "CVF" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 145, + 0, + 496, + 37 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 145, + 0, + 496, + 37 + ], + "spans": [ + { + "bbox": [ + 145, + 0, + 496, + 37 + ], + "type": "text", + "content": "This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "text", + "content": "6317" + } + ] + } + ], + "index": 11 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 0 + }, + { + "para_blocks": [ + { + "bbox": [ + 55, + 72, + 294, + 108 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 72, + 294, + 108 + ], + "spans": [ + { + "bbox": [ + 55, + 72, + 294, + 108 + ], + "type": "text", + "content": "seek to develop an alternative approach that directly predicts both structure and motion, unifying pairwise reasoning and global optimization into a single multi-view framework." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 54, + 110, + 295, + 289 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 54, + 110, + 295, + 289 + ], + "spans": [ + { + "bbox": [ + 54, + 110, + 295, + 289 + ], + "type": "text", + "content": "We are of course not the first to attempt to find unified alternatives to the two-stage SfM pipeline. In the sparse-view setting where conventional correspondence-based methods struggle, several works employ multi-view architectures to jointly reason across input images. SparsePose [26], RelPose++ [12], and PoseDiffusion [34] all leverage multiview transformers to estimate camera pose for input images, albeit using differing mechanisms such as regression, energy-based modeling, and denoising diffusion. More recently, RayDiffusion [41] argues for a local raymap parameterization of cameras instead of a global extrinsic matrix and shows that existing patch-based transformers can be easily adapted for this task, yielding significantly more accurate pose predictions. Importantly, such methods predict only camera motion and fail to predict scene structure." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 56, + 293, + 295, + 615 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 293, + 295, + 615 + ], + "spans": [ + { + "bbox": [ + 56, + 293, + 295, + 615 + ], + "type": "text", + "content": "In this work, we present DiffusionSfM, an end-to-end multi-view model that directly infers dense 3D geometry and cameras from multiple input images. Instead of inferring (depth-agnostic) rays per image patch (as in RayDiffusion [41]) or 3D points per pixel (as in DUSt3R [36]), DiffusionSfM effectively combines both to predict ray origins and endpoints per pixel, directly reporting both scene geometry (endpoints) and generalized cameras (rays). These can readily be converted back to traditional cameras [41]. Compared to RayDiffusion, our model directly predicts structure as well as motion at a finer scale (pixel-wise v.s. patchwise). Compared to DUSt3R, our model directly predicts motion as well as structure but, even more importantly, does so for " + }, + { + "bbox": [ + 56, + 293, + 295, + 615 + ], + "type": "inline_equation", + "content": "N" + }, + { + "bbox": [ + 56, + 293, + 295, + 615 + ], + "type": "text", + "content": " views, eliminating the need for memory-intensive global alignment. To model uncertainty, we train a denoising diffusion model but find two key challenges that need to be addressed. First, diffusion models require (noisy) ground truth as input for training, but existing real datasets do not have known endpoints for all pixels due to missing depth in multi-view stereo. Second, the 3D coordinates of endpoints can be potentially unbounded, whereas diffusion models require normalized data. We develop mechanisms to overcome these challenges, leveraging additional \"GT mask conditioning\" as input to inform the model of missing input data and parameterizing 3D points in projective space instead of Euclidean space. We find these strategies allow us to learn accurate predictions for structure and motion." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 55, + 617, + 295, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 617, + 295, + 713 + ], + "spans": [ + { + "bbox": [ + 55, + 617, + 295, + 713 + ], + "type": "text", + "content": "We train and evaluate DiffusionSfM on real-world and synthetic datasets [21, 24, 42] and find that it can infer accurate geometry and cameras for both object-centric and scene-level images (see Fig. 1). In particular, we find that DiffusionSfM yields more accurate camera estimates compared to prior work across these settings while also modeling the underlying uncertainty via the diffusion process. In summary, we show that DiffusionSfM can serve as a unified" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 314, + 72, + 552, + 84 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 314, + 72, + 552, + 84 + ], + "spans": [ + { + "bbox": [ + 314, + 72, + 552, + 84 + ], + "type": "text", + "content": "multi-view reasoning model for 3D geometry and cameras." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 314, + 96, + 400, + 109 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 314, + 96, + 400, + 109 + ], + "spans": [ + { + "bbox": [ + 314, + 96, + 400, + 109 + ], + "type": "text", + "content": "2. Related Work" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 313, + 117, + 553, + 248 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 117, + 553, + 248 + ], + "spans": [ + { + "bbox": [ + 313, + 117, + 553, + 248 + ], + "type": "text", + "content": "Structure from Motion. Structure-from-Motion (SfM) systems [25] aim to simultaneously recover geometry and cameras given a set of input images. The typical SfM pipeline extracts pixel correspondences from keypoint matching [2, 16], and performs global bundle adjustment (BA) to optimize sparse 3D points and camera parameters by minimizing reprojection errors. Recently, SfM pipelines have been substantially enhanced by replacing classical subcomponents with learning-based methods, such as neural feature descriptors [6, 8], keypoint matching [15, 23, 30], and bundle adjustment [13, 32]." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 313, + 249, + 553, + 475 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 249, + 553, + 475 + ], + "spans": [ + { + "bbox": [ + 313, + 249, + 553, + 475 + ], + "type": "text", + "content": "More recently, an emerging body of research aims to unify the various SfM subcomponents into an end-to-end neural framework. Notably, ACEZero [3] fits a single neural network to input images and learns pixel-aligned 3D coordinates in a self-supervised manner, while FlowMap [27] predicts per-frame cameras and depth maps using off-the-shelf optical flow as supervision. Though ACEZero and FlowMap are promising attempts to revolutionize SfM pipelines, they both register images incrementally and may suffer under large viewpoint changes. DUSt3R [36] directly regresses 3D pointmaps from image pairs and shows strong generalization ability [11, 21]. Building on DUSt3R, MASt3R [10] introduces a feature head, offering pixel-matching capabilities. MASt3R-SfM [7] is a more scalable SfM pipeline based on MASt3R. While these approaches [7, 10, 36] show impressive performance and robustness under sparse views, they are essentially pair-based, requiring sophisticated global alignment procedures to form a consistent estimate for more than two views." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 313, + 477, + 553, + 657 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 477, + 553, + 657 + ], + "spans": [ + { + "bbox": [ + 313, + 477, + 553, + 657 + ], + "type": "text", + "content": "Pose Estimation with Global Reasoning. For the task of sparse-view pose estimation, learning-based methods equipped with global reasoning show favorable robustness where traditional SfM methods [25, 28] fail. This line of research includes energy-based [12, 40], regression-based [26], and diffusion-based pose estimators [34, 41]. Among them, diffusion-based methods show a better ability to handle uncertainty [41]. Closest to our work, RayDiffusion [41] leverages a denoising diffusion model [9, 19] with a patch-aligned (depth-unaware) ray representation to predict generic cameras. Our method goes further and pursues a generic representation for both geometry and cameras in the form of ray origins and endpoints for each pixel. In addition to resulting in a richer output, this joint geometry and pose prediction also yields improvements for pose estimation." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 314, + 669, + 370, + 681 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 314, + 669, + 370, + 681 + ], + "spans": [ + { + "bbox": [ + 314, + 669, + 370, + 681 + ], + "type": "text", + "content": "3. Method" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 313, + 689, + 553, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 689, + 553, + 713 + ], + "spans": [ + { + "bbox": [ + 313, + 689, + 553, + 713 + ], + "type": "text", + "content": "Given a set of sparse (i.e., 2-8) input images, DiffusionSfM predicts the geometry and cameras of a 3D scene in a global" + } + ] + } + ], + "index": 10 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 749, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 749, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 749, + 315, + 757 + ], + "type": "text", + "content": "6318" + } + ] + } + ], + "index": 11 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 1 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 58, + 70, + 556, + 171 + ], + "blocks": [ + { + "bbox": [ + 58, + 70, + 556, + 171 + ], + "lines": [ + { + "bbox": [ + 58, + 70, + 556, + 171 + ], + "spans": [ + { + "bbox": [ + 58, + 70, + 556, + 171 + ], + "type": "image", + "image_path": "e313265c3ef039eaa8c98aab188fac7c1ec90a35c2ae39191b416a12c793c328.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 54, + 178, + 555, + 266 + ], + "lines": [ + { + "bbox": [ + 54, + 178, + 555, + 266 + ], + "spans": [ + { + "bbox": [ + 54, + 178, + 555, + 266 + ], + "type": "text", + "content": "Figure 2. Method. Given sparse multi-view images as input, DiffusionSfM predicts pixel-wise ray origins and endpoints in a global frame (Sec. 3.1) using a denoising diffusion process (Sec. 3.2). For each image, we compute patch-wise embeddings with DINOv2 [18] and embed noisy ray origins and endpoints into latents using a single downsampling convolutional layer, ensuring alignment with the spatial footprint of the image embeddings. We implement a Diffusion Transformer architecture that predicts clean ray origins and endpoints from noisy samples. A convolutional DPT [20] head outputs full-resolution denoised ray origins and endpoints. To handle incomplete ground truth (GT) during training, we condition the model on a GT mask (Sec. 3.3). At inference, the GT mask is set to all ones, enabling the model to predict origins and endpoints for all pixels. The predicted ray origins and endpoints can be directly visualized in 3D or postprocessed to recover camera extrinsics, intrinsics, and multi-view consistent depth maps." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "bbox": [ + 54, + 288, + 296, + 360 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 54, + 288, + 296, + 360 + ], + "spans": [ + { + "bbox": [ + 54, + 288, + 296, + 360 + ], + "type": "text", + "content": "coordinate frame. In Sec. 3.1, we propose to represent 3D scenes as dense pixel-aligned ray origins and endpoints. To predict such scene representations from sparse input images while modeling uncertainty, Sec. 3.2 proposes a denoising diffusion architecture. We then discuss some key practical challenges in training such a model in Sec. 3.3." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 55, + 366, + 269, + 379 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 366, + 269, + 379 + ], + "spans": [ + { + "bbox": [ + 55, + 366, + 269, + 379 + ], + "type": "text", + "content": "3.1. 3D Scenes as Ray Origins and Endpoints" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 55, + 382, + 296, + 491 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 382, + 296, + 491 + ], + "spans": [ + { + "bbox": [ + 55, + 382, + 296, + 491 + ], + "type": "text", + "content": "Given an input image " + }, + { + "bbox": [ + 55, + 382, + 296, + 491 + ], + "type": "inline_equation", + "content": "\\mathbf{I} \\in \\mathbb{R}^{H \\times W \\times 3}" + }, + { + "bbox": [ + 55, + 382, + 296, + 491 + ], + "type": "text", + "content": " with a depth map " + }, + { + "bbox": [ + 55, + 382, + 296, + 491 + ], + "type": "inline_equation", + "content": "\\mathbf{D} \\in \\mathbb{R}^{H \\times W}" + }, + { + "bbox": [ + 55, + 382, + 296, + 491 + ], + "type": "text", + "content": ", camera intrinsics " + }, + { + "bbox": [ + 55, + 382, + 296, + 491 + ], + "type": "inline_equation", + "content": "\\mathbf{K} \\in \\mathbb{R}^{3 \\times 3}" + }, + { + "bbox": [ + 55, + 382, + 296, + 491 + ], + "type": "text", + "content": ", and world-to-camera extrinsics " + }, + { + "bbox": [ + 55, + 382, + 296, + 491 + ], + "type": "inline_equation", + "content": "\\mathbf{T} \\in \\mathbb{R}^{4 \\times 4}" + }, + { + "bbox": [ + 55, + 382, + 296, + 491 + ], + "type": "text", + "content": " (equivalently, rotation " + }, + { + "bbox": [ + 55, + 382, + 296, + 491 + ], + "type": "inline_equation", + "content": "\\mathbf{R} \\in SO(3)" + }, + { + "bbox": [ + 55, + 382, + 296, + 491 + ], + "type": "text", + "content": " and translation " + }, + { + "bbox": [ + 55, + 382, + 296, + 491 + ], + "type": "inline_equation", + "content": "\\mathbf{t} \\in \\mathbb{R}^3" + }, + { + "bbox": [ + 55, + 382, + 296, + 491 + ], + "type": "text", + "content": "), each 2D image pixel " + }, + { + "bbox": [ + 55, + 382, + 296, + 491 + ], + "type": "inline_equation", + "content": "\\mathbf{P}_{ij} = [u, v]" + }, + { + "bbox": [ + 55, + 382, + 296, + 491 + ], + "type": "text", + "content": " corresponds to a ray that travels from the camera center " + }, + { + "bbox": [ + 55, + 382, + 296, + 491 + ], + "type": "inline_equation", + "content": "\\mathbf{c}" + }, + { + "bbox": [ + 55, + 382, + 296, + 491 + ], + "type": "text", + "content": " through the pixel's projected position on the image plane, terminating at the object's surface as specified by the depth map " + }, + { + "bbox": [ + 55, + 382, + 296, + 491 + ], + "type": "inline_equation", + "content": "\\mathbf{D}" + }, + { + "bbox": [ + 55, + 382, + 296, + 491 + ], + "type": "text", + "content": ". The endpoint of the ray associated with image pixel " + }, + { + "bbox": [ + 55, + 382, + 296, + 491 + ], + "type": "inline_equation", + "content": "\\mathbf{P}_{ij}" + }, + { + "bbox": [ + 55, + 382, + 296, + 491 + ], + "type": "text", + "content": " is given by:" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 102, + 497, + 294, + 512 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 102, + 497, + 294, + 512 + ], + "spans": [ + { + "bbox": [ + 102, + 497, + 294, + 512 + ], + "type": "interline_equation", + "content": "\\mathbf {E} _ {i j} = \\mathbf {T} ^ {- 1} h (\\mathbf {D} _ {i j} \\cdot \\mathbf {K} ^ {- 1} [ u, v, 1 ] ^ {T}) \\tag {1}", + "image_path": "3466594519819f9dfcf0eea159bc681d6bc336f8738ce3933a309ea21b2db91a.jpg" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 55, + 518, + 295, + 554 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 518, + 295, + 554 + ], + "spans": [ + { + "bbox": [ + 55, + 518, + 295, + 554 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 55, + 518, + 295, + 554 + ], + "type": "inline_equation", + "content": "h" + }, + { + "bbox": [ + 55, + 518, + 295, + 554 + ], + "type": "text", + "content": " maps the 3D point into homogeneous coordinates. The shared ray origin " + }, + { + "bbox": [ + 55, + 518, + 295, + 554 + ], + "type": "inline_equation", + "content": "\\mathbf{O}_{ij}" + }, + { + "bbox": [ + 55, + 518, + 295, + 554 + ], + "type": "text", + "content": " for all pixels is equivalent to the camera center " + }, + { + "bbox": [ + 55, + 518, + 295, + 554 + ], + "type": "inline_equation", + "content": "\\mathbf{c}" + }, + { + "bbox": [ + 55, + 518, + 295, + 554 + ], + "type": "text", + "content": ", and can be computed as:" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 125, + 561, + 294, + 575 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 125, + 561, + 294, + 575 + ], + "spans": [ + { + "bbox": [ + 125, + 561, + 294, + 575 + ], + "type": "interline_equation", + "content": "\\mathbf {O} _ {i j} = \\mathbf {c} = h (- \\mathbf {R} ^ {- 1} \\mathbf {t}) \\tag {2}", + "image_path": "9e1ce919c04322a8f0014fdbb1d0552ff239fce7b2930e7812a09cfa755f407f.jpg" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 55, + 582, + 295, + 653 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 582, + 295, + 653 + ], + "spans": [ + { + "bbox": [ + 55, + 582, + 295, + 653 + ], + "type": "text", + "content": "In summary, we associate each image pixel with a ray origin and endpoint " + }, + { + "bbox": [ + 55, + 582, + 295, + 653 + ], + "type": "inline_equation", + "content": "\\mathbf{S}_{ij} = \\langle \\mathbf{O}_{ij},\\mathbf{E}_{ij}\\rangle" + }, + { + "bbox": [ + 55, + 582, + 295, + 653 + ], + "type": "text", + "content": " in world coordinates, describing the location of the observing camera and the observed 3D point on the object surface. Given a bundle of ray origins and endpoints, we can easily extract the corresponding camera pose [41]." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 55, + 654, + 296, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 654, + 296, + 713 + ], + "spans": [ + { + "bbox": [ + 55, + 654, + 296, + 713 + ], + "type": "text", + "content": "Learning Over-Parameterized Representations. We represent 3D scenes and cameras using distributed ray origins " + }, + { + "bbox": [ + 55, + 654, + 296, + 713 + ], + "type": "inline_equation", + "content": "\\mathbf{O}" + }, + { + "bbox": [ + 55, + 654, + 296, + 713 + ], + "type": "text", + "content": " and endpoints " + }, + { + "bbox": [ + 55, + 654, + 296, + 713 + ], + "type": "inline_equation", + "content": "\\mathbf{E}" + }, + { + "bbox": [ + 55, + 654, + 296, + 713 + ], + "type": "text", + "content": " rather than global alternatives such as quaternions and translation vectors. This design is inspired by RayDiffusion [41], and it facilitates the use of the" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 313, + 288, + 555, + 396 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 288, + 555, + 396 + ], + "spans": [ + { + "bbox": [ + 313, + 288, + 555, + 396 + ], + "type": "text", + "content": "distributed deep features learned by state-of-the-art vision backbones, such as DINOv2 [18], which encode image information in a patch-wise manner. Notably, while ray origins " + }, + { + "bbox": [ + 313, + 288, + 555, + 396 + ], + "type": "inline_equation", + "content": "\\mathbf{O}" + }, + { + "bbox": [ + 313, + 288, + 555, + 396 + ], + "type": "text", + "content": " should ideally be identical across all pixels, we predict them densely alongside ray endpoints " + }, + { + "bbox": [ + 313, + 288, + 555, + 396 + ], + "type": "inline_equation", + "content": "\\mathbf{E}" + }, + { + "bbox": [ + 313, + 288, + 555, + 396 + ], + "type": "text", + "content": ". This encourages ray origins to remain close within the same image, providing implicit regularization during training. Practically, predicting both ray origins and endpoints simultaneously is easy to implement using a single projection head." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 313, + 402, + 400, + 414 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 402, + 400, + 414 + ], + "spans": [ + { + "bbox": [ + 313, + 402, + 400, + 414 + ], + "type": "text", + "content": "3.2. DiffusionSfM" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 313, + 419, + 554, + 467 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 419, + 554, + 467 + ], + "spans": [ + { + "bbox": [ + 313, + 419, + 554, + 467 + ], + "type": "text", + "content": "We propose a denoising Diffusion Transformer (DiT) architecture [19] that predicts ray origins and endpoints (Sec. 3.1) via a denoising diffusion process. An overview of DiffusionSfM is given in Fig. 2." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 313, + 468, + 555, + 563 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 468, + 555, + 563 + ], + "spans": [ + { + "bbox": [ + 313, + 468, + 555, + 563 + ], + "type": "text", + "content": "Diffusion Framework. Given pixel-aligned ray origins and endpoints " + }, + { + "bbox": [ + 313, + 468, + 555, + 563 + ], + "type": "inline_equation", + "content": "S = \\text{stack}(\\{\\mathbf{S}^{(n)}\\}_{n=1}^{N})" + }, + { + "bbox": [ + 313, + 468, + 555, + 563 + ], + "type": "text", + "content": " associated with a set of " + }, + { + "bbox": [ + 313, + 468, + 555, + 563 + ], + "type": "inline_equation", + "content": "N" + }, + { + "bbox": [ + 313, + 468, + 555, + 563 + ], + "type": "text", + "content": " input images, we apply a forward diffusion process [9, 29] that adds time-dependent Gaussian noise to them. Let " + }, + { + "bbox": [ + 313, + 468, + 555, + 563 + ], + "type": "inline_equation", + "content": "S_{t}" + }, + { + "bbox": [ + 313, + 468, + 555, + 563 + ], + "type": "text", + "content": " denote the noisy ray origins and endpoints at timestep " + }, + { + "bbox": [ + 313, + 468, + 555, + 563 + ], + "type": "inline_equation", + "content": "t" + }, + { + "bbox": [ + 313, + 468, + 555, + 563 + ], + "type": "text", + "content": ", where " + }, + { + "bbox": [ + 313, + 468, + 555, + 563 + ], + "type": "inline_equation", + "content": "S_{0}" + }, + { + "bbox": [ + 313, + 468, + 555, + 563 + ], + "type": "text", + "content": " is the clean sample and " + }, + { + "bbox": [ + 313, + 468, + 555, + 563 + ], + "type": "inline_equation", + "content": "S_{T}" + }, + { + "bbox": [ + 313, + 468, + 555, + 563 + ], + "type": "text", + "content": " (at the final diffusion step " + }, + { + "bbox": [ + 313, + 468, + 555, + 563 + ], + "type": "inline_equation", + "content": "T" + }, + { + "bbox": [ + 313, + 468, + 555, + 563 + ], + "type": "text", + "content": ") approximates pure Gaussian noise. The forward diffusion process is defined as:" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 378, + 571, + 553, + 585 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 378, + 571, + 553, + 585 + ], + "spans": [ + { + "bbox": [ + 378, + 571, + 553, + 585 + ], + "type": "interline_equation", + "content": "\\mathcal {S} _ {t} = \\sqrt {\\bar {\\alpha} _ {t}} \\mathcal {S} _ {0} + \\sqrt {1 - \\bar {\\alpha} _ {t}} \\epsilon \\tag {3}", + "image_path": "66718f1e413aec4c4cd79f699f7bbf32334e032064bf7b39c249fdb4c6510257.jpg" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 313, + 594, + 555, + 689 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 594, + 555, + 689 + ], + "spans": [ + { + "bbox": [ + 313, + 594, + 555, + 689 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 313, + 594, + 555, + 689 + ], + "type": "inline_equation", + "content": "t \\sim \\mathrm{Uniform}(0, T]" + }, + { + "bbox": [ + 313, + 594, + 555, + 689 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 313, + 594, + 555, + 689 + ], + "type": "inline_equation", + "content": "\\epsilon \\sim \\mathcal{N}(0, I)" + }, + { + "bbox": [ + 313, + 594, + 555, + 689 + ], + "type": "text", + "content": ", and " + }, + { + "bbox": [ + 313, + 594, + 555, + 689 + ], + "type": "inline_equation", + "content": "\\bar{\\alpha}_t" + }, + { + "bbox": [ + 313, + 594, + 555, + 689 + ], + "type": "text", + "content": " follows a pre-defined noise schedule that controls the strength of added noise at each timestep. To perform the reverse diffusion process, which progressively reconstructs the clean sample given noisy observations, we train a diffusion model " + }, + { + "bbox": [ + 313, + 594, + 555, + 689 + ], + "type": "inline_equation", + "content": "f_\\theta" + }, + { + "bbox": [ + 313, + 594, + 555, + 689 + ], + "type": "text", + "content": " that takes " + }, + { + "bbox": [ + 313, + 594, + 555, + 689 + ], + "type": "inline_equation", + "content": "\\mathcal{S}_t" + }, + { + "bbox": [ + 313, + 594, + 555, + 689 + ], + "type": "text", + "content": " as input and optionally incorporates additional conditioning information " + }, + { + "bbox": [ + 313, + 594, + 555, + 689 + ], + "type": "inline_equation", + "content": "\\mathcal{C}" + }, + { + "bbox": [ + 313, + 594, + 555, + 689 + ], + "type": "text", + "content": ". The model is trained using the following loss function (with the \" " + }, + { + "bbox": [ + 313, + 594, + 555, + 689 + ], + "type": "inline_equation", + "content": "x_0" + }, + { + "bbox": [ + 313, + 594, + 555, + 689 + ], + "type": "text", + "content": "\" objective):" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 354, + 700, + 553, + 714 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 354, + 700, + 553, + 714 + ], + "spans": [ + { + "bbox": [ + 354, + 700, + 553, + 714 + ], + "type": "interline_equation", + "content": "\\mathcal {L} _ {\\text {D i f f u s i o n}} = \\mathbb {E} _ {t, S _ {0}, \\epsilon} \\| S _ {0} - f _ {\\theta} (S _ {t}, t, \\mathcal {C}) \\| ^ {2} \\tag {4}", + "image_path": "0cc050f6466499f8ebd864bbfa4c3500d89b1dd31ef02f6797963eae989b1c69.jpg" + } + ] + } + ], + "index": 16 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 315, + 757 + ], + "type": "text", + "content": "6319" + } + ] + } + ], + "index": 17 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 2 + }, + { + "para_blocks": [ + { + "bbox": [ + 55, + 72, + 294, + 179 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 72, + 294, + 179 + ], + "spans": [ + { + "bbox": [ + 55, + 72, + 294, + 179 + ], + "type": "text", + "content": "Architecture. We implement " + }, + { + "bbox": [ + 55, + 72, + 294, + 179 + ], + "type": "inline_equation", + "content": "f_{\\theta}" + }, + { + "bbox": [ + 55, + 72, + 294, + 179 + ], + "type": "text", + "content": " using a DiT [19] architecture conditioned on deep image features " + }, + { + "bbox": [ + 55, + 72, + 294, + 179 + ], + "type": "inline_equation", + "content": "\\mathcal{C} \\in \\mathbb{R}^{N \\times h \\times w \\times c_1}" + }, + { + "bbox": [ + 55, + 72, + 294, + 179 + ], + "type": "text", + "content": " from DINOv2 [18], where " + }, + { + "bbox": [ + 55, + 72, + 294, + 179 + ], + "type": "inline_equation", + "content": "h" + }, + { + "bbox": [ + 55, + 72, + 294, + 179 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 55, + 72, + 294, + 179 + ], + "type": "inline_equation", + "content": "w" + }, + { + "bbox": [ + 55, + 72, + 294, + 179 + ], + "type": "text", + "content": " are the patch resolution and " + }, + { + "bbox": [ + 55, + 72, + 294, + 179 + ], + "type": "inline_equation", + "content": "c_1" + }, + { + "bbox": [ + 55, + 72, + 294, + 179 + ], + "type": "text", + "content": " is the embedding dimension. To align pixels to the spatial information learned by DINOv2, we apply a convolutional layer with kernel size and stride equal to ViT patch size: this spatially downsamples the noisy ray origins and endpoints " + }, + { + "bbox": [ + 55, + 72, + 294, + 179 + ], + "type": "inline_equation", + "content": "S_t" + }, + { + "bbox": [ + 55, + 72, + 294, + 179 + ], + "type": "text", + "content": " to match the DINOv2 features while increasing their feature dimension:" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 110, + 186, + 295, + 200 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 110, + 186, + 295, + 200 + ], + "spans": [ + { + "bbox": [ + 110, + 186, + 295, + 200 + ], + "type": "interline_equation", + "content": "\\mathcal {F} = \\operatorname {C o n v} \\left(\\mathcal {S} _ {t}\\right) \\in \\mathbb {R} ^ {N \\times h \\times w \\times c _ {2}} \\tag {5}", + "image_path": "9622ca7a6afa34ea4bfae4e86563229690a837160102dbe3fa20f595833578da.jpg" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 55, + 206, + 295, + 277 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 206, + 295, + 277 + ], + "spans": [ + { + "bbox": [ + 55, + 206, + 295, + 277 + ], + "type": "text", + "content": "The combined DiT input is constructed by concatenating these two feature sets along the channel dimension: " + }, + { + "bbox": [ + 55, + 206, + 295, + 277 + ], + "type": "inline_equation", + "content": "\\mathcal{F} \\oplus \\mathcal{C}" + }, + { + "bbox": [ + 55, + 206, + 295, + 277 + ], + "type": "text", + "content": ". Within DiT, patch-wise features attend to others through self-attention [33]. To distinguish between different images and their respective patches, we apply sinusoidal positional encoding [33] based on image and patch indices." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 55, + 278, + 296, + 410 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 278, + 296, + 410 + ], + "spans": [ + { + "bbox": [ + 55, + 278, + 296, + 410 + ], + "type": "text", + "content": "While the DiT operates on low-resolution features, our objective is to produce pixel-aligned dense ray origins and endpoints. To achieve this, we employ a DPT (Dense Prediction Transformer) [20] decoder, which takes intermediate feature maps from both DINoV2 and DiT as inputs. The DPT decoder progressively increases the feature resolution through several convolutional layers. The final ray origins and endpoints are decoded from the DPT output using a single linear layer. During inference, we apply the trained model in the reverse diffusion process to iteratively denoise a randomly initialized Gaussian sample." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 55, + 416, + 237, + 428 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 416, + 237, + 428 + ], + "spans": [ + { + "bbox": [ + 55, + 416, + 237, + 428 + ], + "type": "text", + "content": "3.3. Practical Training Considerations" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 55, + 434, + 295, + 624 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 434, + 295, + 624 + ], + "spans": [ + { + "bbox": [ + 55, + 434, + 295, + 624 + ], + "type": "text", + "content": "Homogeneous Coordinates for Unbounded Geometry. Real-world scenes often exhibit significant variations in scale, both across different scenes and within a single scene. While normalizing by the average distance of 3D points [36] helps address cross-scene variation, within-scene variations remain. For instance, a background building may be much farther away than foreground elements, potentially resulting in extremely large coordinate values when generating ground-truth ray origins and endpoints. However, neural networks (especially diffusion models) tend to train most effectively when working with bounded inputs and outputs (e.g., between -1 and 1). To stabilize training across the large scale variations present in 3D scene datasets, we propose to represent ray origins and endpoints in homogeneous coordinates. Specifically, given any 3D point, we apply a homogeneous transform from " + }, + { + "bbox": [ + 55, + 434, + 295, + 624 + ], + "type": "inline_equation", + "content": "\\mathbb{R}^3\\to \\mathbb{P}^3" + }, + { + "bbox": [ + 55, + 434, + 295, + 624 + ], + "type": "text", + "content": ":" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 124, + 631, + 295, + 653 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 124, + 631, + 295, + 653 + ], + "spans": [ + { + "bbox": [ + 124, + 631, + 295, + 653 + ], + "type": "interline_equation", + "content": "(x, y, z) \\rightarrow \\frac {1}{w} (x, y, z, 1) \\tag {6}", + "image_path": "1dfba70c4cbb7c37dd03962b52c4974054759c9c678c28c394c8b7c4f03463fd.jpg" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 55, + 658, + 296, + 693 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 658, + 296, + 693 + ], + "spans": [ + { + "bbox": [ + 55, + 658, + 296, + 693 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 55, + 658, + 296, + 693 + ], + "type": "inline_equation", + "content": "w" + }, + { + "bbox": [ + 55, + 658, + 296, + 693 + ], + "type": "text", + "content": " is an arbitrary scale factor. To encourage bounded coordinates in practice, we choose " + }, + { + "bbox": [ + 55, + 658, + 296, + 693 + ], + "type": "inline_equation", + "content": "w" + }, + { + "bbox": [ + 55, + 658, + 296, + 693 + ], + "type": "text", + "content": " such that the homogeneous coordinate is unit-norm:" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 121, + 700, + 295, + 714 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 121, + 700, + 295, + 714 + ], + "spans": [ + { + "bbox": [ + 121, + 700, + 295, + 714 + ], + "type": "interline_equation", + "content": "w := \\sqrt {x ^ {2} + y ^ {2} + z ^ {2} + 1} \\tag {7}", + "image_path": "6357375065d35c7ac6f55dc4bd1bf5308890219438a37d0758b5b0af60af07f4.jpg" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 313, + 72, + 553, + 133 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 72, + 553, + 133 + ], + "spans": [ + { + "bbox": [ + 313, + 72, + 553, + 133 + ], + "type": "text", + "content": "Unit normalization allows homogeneous coordinates to serve as a bounded representation for unbounded scene geometry. For example, " + }, + { + "bbox": [ + 313, + 72, + 553, + 133 + ], + "type": "inline_equation", + "content": "(x,y,z,0)" + }, + { + "bbox": [ + 313, + 72, + 553, + 133 + ], + "type": "text", + "content": " is a point at infinity in the direction of " + }, + { + "bbox": [ + 313, + 72, + 553, + 133 + ], + "type": "inline_equation", + "content": "(x,y,z)" + }, + { + "bbox": [ + 313, + 72, + 553, + 133 + ], + "type": "text", + "content": ". We find that this representation makes large coordinate values more tractable during training." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 313, + 133, + 553, + 276 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 133, + 553, + 276 + ], + "spans": [ + { + "bbox": [ + 313, + 133, + 553, + 276 + ], + "type": "text", + "content": "Training with Incomplete Ground Truth. Many real-world datasets (e.g., CO3D [21] and MegaDepth [11]), provide only sparse point clouds, leading to incomplete depth information. This presents a significant challenge in diffusion training, as ground-truth (GT) depth values that are used to create clean ray endpoints often contain invalid or missing data. It is highly undesirable for these missing ray endpoints to be interpreted as part of the target distribution. Unlike regression models (e.g., DUSt3R [36]), which do not require (noisy) GT data as input and can simply apply a mask on the loss for supervision, diffusion models must handle incomplete inputs during training." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 313, + 277, + 554, + 408 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 277, + 554, + 408 + ], + "spans": [ + { + "bbox": [ + 313, + 277, + 554, + 408 + ], + "type": "text", + "content": "To mitigate this issue, we further apply GT masks " + }, + { + "bbox": [ + 313, + 277, + 554, + 408 + ], + "type": "inline_equation", + "content": "\\mathcal{M} \\in \\mathbb{R}^{N \\times H \\times W}" + }, + { + "bbox": [ + 313, + 277, + 554, + 408 + ], + "type": "text", + "content": " to the DiT inputs, where zero values indicate pixels with invalid depth. During training, we multiply DiT inputs with GT masks element-wise, then concatenate along the channel dimension: " + }, + { + "bbox": [ + 313, + 277, + 554, + 408 + ], + "type": "inline_equation", + "content": "S_{t}' = (\\mathcal{M} \\cdot S_{t}) \\oplus \\mathcal{M}" + }, + { + "bbox": [ + 313, + 277, + 554, + 408 + ], + "type": "text", + "content": ". Then, we only compute the diffusion loss in Eq. 4 over unmasked pixels. By implementing these strategies, we encourage the model to focus on regions with valid GT values during training. During inference, however, we would like the diffusion process to estimate ray endpoints at all pixels, so we always use GT masks with values set to one." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 313, + 410, + 554, + 566 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 410, + 554, + 566 + ], + "spans": [ + { + "bbox": [ + 313, + 410, + 554, + 566 + ], + "type": "text", + "content": "Sparse-to-Dense Training. In practice, we find that training the entire model from scratch leads to slow convergence and suboptimal performance. To address this, we propose a sparse-to-dense training approach. First, we train a sparse version of the model, where the DPT decoder is removed, and the output ray origins and endpoints have the same spatial resolution as the DINOv2 features. Unlike Eq. 5, no spatial downsampling is required, so this sparse model uses a single linear layer to embed the noisy ray origins and endpoints. Once the sparse model is trained, we initialize the dense model DiT with the learned weights from the sparse model. This two-stage approach significantly improves performance; see the supplementary material for comparisons." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 313, + 578, + 394, + 592 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 578, + 394, + 592 + ], + "spans": [ + { + "bbox": [ + 313, + 578, + 394, + 592 + ], + "type": "text", + "content": "4. Experiments" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 313, + 599, + 430, + 612 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 599, + 430, + 612 + ], + "spans": [ + { + "bbox": [ + 313, + 599, + 430, + 612 + ], + "type": "text", + "content": "4.1. Experimental Setup" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 313, + 617, + 554, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 617, + 554, + 713 + ], + "spans": [ + { + "bbox": [ + 313, + 617, + 554, + 713 + ], + "type": "text", + "content": "Datasets. We introduce two model variants, each trained on different datasets. (1) DiffusionSfM-CO3D: Following prior work [12, 41], we train and evaluate our model on the CO3D dataset [21], which consists of turntable video sequences of various object categories. Specifically, we train on 41 object categories and evaluate on both these seen categories and an additional 10 unseen categories to assess generalization. (2) DiffusionSfM: This variant is trained on the" + } + ] + } + ], + "index": 15 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 749, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 749, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 749, + 315, + 757 + ], + "type": "text", + "content": "6320" + } + ] + } + ], + "index": 16 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 3 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 75, + 95, + 160, + 137 + ], + "blocks": [ + { + "bbox": [ + 82, + 71, + 153, + 92 + ], + "lines": [ + { + "bbox": [ + 82, + 71, + 153, + 92 + ], + "spans": [ + { + "bbox": [ + 82, + 71, + 153, + 92 + ], + "type": "text", + "content": "Input Images /GT Point Clouds" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 75, + 95, + 160, + 137 + ], + "lines": [ + { + "bbox": [ + 75, + 95, + 160, + 137 + ], + "spans": [ + { + "bbox": [ + 75, + 95, + 160, + 137 + ], + "type": "image", + "image_path": "d6e463bb54fb1e9753a107c4e0f00652fabfafb5aeb6a00a0e8c75bcf8a0d463.jpg" + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_body" + } + ], + "index": 1 + }, + { + "type": "image", + "bbox": [ + 76, + 137, + 160, + 178 + ], + "blocks": [ + { + "bbox": [ + 76, + 137, + 160, + 178 + ], + "lines": [ + { + "bbox": [ + 76, + 137, + 160, + 178 + ], + "spans": [ + { + "bbox": [ + 76, + 137, + 160, + 178 + ], + "type": "image", + "image_path": "dc5f60efbcfe9c7755289675788486c4578c9f213d2b6042429e3e03c31ae41f.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + } + ], + "index": 2 + }, + { + "type": "image", + "bbox": [ + 183, + 102, + 282, + 171 + ], + "blocks": [ + { + "bbox": [ + 211, + 76, + 269, + 87 + ], + "lines": [ + { + "bbox": [ + 211, + 76, + 269, + 87 + ], + "spans": [ + { + "bbox": [ + 211, + 76, + 269, + 87 + ], + "type": "text", + "content": "DiffusionSfM" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 183, + 102, + 282, + 171 + ], + "lines": [ + { + "bbox": [ + 183, + 102, + 282, + 171 + ], + "spans": [ + { + "bbox": [ + 183, + 102, + 282, + 171 + ], + "type": "image", + "image_path": "4a90c02cbf86ede1c3b9f9128f7369efafe2004fac7f244144acbef7b880efc4.jpg" + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_body" + } + ], + "index": 4 + }, + { + "type": "image", + "bbox": [ + 313, + 102, + 412, + 170 + ], + "blocks": [ + { + "bbox": [ + 343, + 76, + 395, + 87 + ], + "lines": [ + { + "bbox": [ + 343, + 76, + 395, + 87 + ], + "spans": [ + { + "bbox": [ + 343, + 76, + 395, + 87 + ], + "type": "text", + "content": "RD + MoGe" + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 313, + 102, + 412, + 170 + ], + "lines": [ + { + "bbox": [ + 313, + 102, + 412, + 170 + ], + "spans": [ + { + "bbox": [ + 313, + 102, + 412, + 170 + ], + "type": "image", + "image_path": "4dbeed7353c79ac7b9a429bdb597e8c4bf321a4dea76408dad8d2d25a4cb3311.jpg" + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "image_body" + } + ], + "index": 6 + }, + { + "type": "image", + "bbox": [ + 440, + 102, + 537, + 170 + ], + "blocks": [ + { + "bbox": [ + 479, + 76, + 515, + 87 + ], + "lines": [ + { + "bbox": [ + 479, + 76, + 515, + 87 + ], + "spans": [ + { + "bbox": [ + 479, + 76, + 515, + 87 + ], + "type": "text", + "content": "DUSt3R" + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 440, + 102, + 537, + 170 + ], + "lines": [ + { + "bbox": [ + 440, + 102, + 537, + 170 + ], + "spans": [ + { + "bbox": [ + 440, + 102, + 537, + 170 + ], + "type": "image", + "image_path": "3c4d4f0156271974c5a1bea2211d4c0da5b8a04dce2f24b68c677ecb90c28e64.jpg" + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "image_body" + } + ], + "index": 8 + }, + { + "type": "image", + "bbox": [ + 72, + 184, + 199, + 251 + ], + "blocks": [ + { + "bbox": [ + 72, + 184, + 199, + 251 + ], + "lines": [ + { + "bbox": [ + 72, + 184, + 199, + 251 + ], + "spans": [ + { + "bbox": [ + 72, + 184, + 199, + 251 + ], + "type": "image", + "image_path": "08b9d18a8fd24f92149aabec1779e2696ba49c6f961ae07dcdb2a0eb2a56a1d2.jpg" + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "image_body" + } + ], + "index": 9 + }, + { + "type": "image", + "bbox": [ + 199, + 186, + 318, + 250 + ], + "blocks": [ + { + "bbox": [ + 199, + 186, + 318, + 250 + ], + "lines": [ + { + "bbox": [ + 199, + 186, + 318, + 250 + ], + "spans": [ + { + "bbox": [ + 199, + 186, + 318, + 250 + ], + "type": "image", + "image_path": "6d07fe01df1484bab4bf2ec9045876531d07cc800c093177a7bbbbbfac0de292.jpg" + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "image_body" + } + ], + "index": 10 + }, + { + "type": "image", + "bbox": [ + 318, + 184, + 406, + 250 + ], + "blocks": [ + { + "bbox": [ + 318, + 184, + 406, + 250 + ], + "lines": [ + { + "bbox": [ + 318, + 184, + 406, + 250 + ], + "spans": [ + { + "bbox": [ + 318, + 184, + 406, + 250 + ], + "type": "image", + "image_path": "21cddbbf7d3c3a2aa6c186d5790bb68351b7e18e5fec658f160d0f440c0ebba2.jpg" + } + ] + } + ], + "index": 11, + "angle": 0, + "type": "image_body" + } + ], + "index": 11 + }, + { + "type": "image", + "bbox": [ + 406, + 184, + 539, + 251 + ], + "blocks": [ + { + "bbox": [ + 406, + 184, + 539, + 251 + ], + "lines": [ + { + "bbox": [ + 406, + 184, + 539, + 251 + ], + "spans": [ + { + "bbox": [ + 406, + 184, + 539, + 251 + ], + "type": "image", + "image_path": "4467112fd81e570e98956508a4e57480691fb0ba442c38bbf2627e053ef2ac55.jpg" + } + ] + } + ], + "index": 12, + "angle": 0, + "type": "image_body" + } + ], + "index": 12 + }, + { + "type": "image", + "bbox": [ + 74, + 261, + 103, + 289 + ], + "blocks": [ + { + "bbox": [ + 74, + 261, + 103, + 289 + ], + "lines": [ + { + "bbox": [ + 74, + 261, + 103, + 289 + ], + "spans": [ + { + "bbox": [ + 74, + 261, + 103, + 289 + ], + "type": "image", + "image_path": "272dd89e3cebb5972006b1e754848a814b96b262781b027efbac94167d73da4f.jpg" + } + ] + } + ], + "index": 13, + "angle": 0, + "type": "image_body" + } + ], + "index": 13 + }, + { + "type": "image", + "bbox": [ + 74, + 289, + 103, + 318 + ], + "blocks": [ + { + "bbox": [ + 74, + 289, + 103, + 318 + ], + "lines": [ + { + "bbox": [ + 74, + 289, + 103, + 318 + ], + "spans": [ + { + "bbox": [ + 74, + 289, + 103, + 318 + ], + "type": "image", + "image_path": "f9d752bc2a72574ad8da4962e3c75c5038c52407176760b4817e89875dc60734.jpg" + } + ] + } + ], + "index": 14, + "angle": 0, + "type": "image_body" + } + ], + "index": 14 + }, + { + "type": "image", + "bbox": [ + 104, + 262, + 130, + 289 + ], + "blocks": [ + { + "bbox": [ + 104, + 262, + 130, + 289 + ], + "lines": [ + { + "bbox": [ + 104, + 262, + 130, + 289 + ], + "spans": [ + { + "bbox": [ + 104, + 262, + 130, + 289 + ], + "type": "image", + "image_path": "77e3edccd4250d49af8e4e40afe6cba4c9d6793ae40e2556aa588e44db248cf8.jpg" + } + ] + } + ], + "index": 15, + "angle": 0, + "type": "image_body" + } + ], + "index": 15 + }, + { + "type": "image", + "bbox": [ + 131, + 262, + 160, + 289 + ], + "blocks": [ + { + "bbox": [ + 131, + 262, + 160, + 289 + ], + "lines": [ + { + "bbox": [ + 131, + 262, + 160, + 289 + ], + "spans": [ + { + "bbox": [ + 131, + 262, + 160, + 289 + ], + "type": "image", + "image_path": "764b33e09ba83273a035143014a202821327b88eae6f527e802b136b7abbff21.jpg" + } + ] + } + ], + "index": 16, + "angle": 0, + "type": "image_body" + } + ], + "index": 16 + }, + { + "type": "image", + "bbox": [ + 73, + 329, + 193, + 390 + ], + "blocks": [ + { + "bbox": [ + 73, + 329, + 193, + 390 + ], + "lines": [ + { + "bbox": [ + 73, + 329, + 193, + 390 + ], + "spans": [ + { + "bbox": [ + 73, + 329, + 193, + 390 + ], + "type": "image", + "image_path": "abec2e30ad827dfb18d995acedef81d339f0eae11fb77e60e29a29e8e3f6f10d.jpg" + } + ] + } + ], + "index": 17, + "angle": 0, + "type": "image_body" + } + ], + "index": 17 + }, + { + "type": "image", + "bbox": [ + 192, + 329, + 317, + 391 + ], + "blocks": [ + { + "bbox": [ + 192, + 329, + 317, + 391 + ], + "lines": [ + { + "bbox": [ + 192, + 329, + 317, + 391 + ], + "spans": [ + { + "bbox": [ + 192, + 329, + 317, + 391 + ], + "type": "image", + "image_path": "083a0c1897454b692a6d69688d3baf77dee68c4161fa70f086982461b4a28fe5.jpg" + } + ] + } + ], + "index": 18, + "angle": 0, + "type": "image_body" + } + ], + "index": 18 + }, + { + "type": "image", + "bbox": [ + 319, + 329, + 421, + 389 + ], + "blocks": [ + { + "bbox": [ + 319, + 329, + 421, + 389 + ], + "lines": [ + { + "bbox": [ + 319, + 329, + 421, + 389 + ], + "spans": [ + { + "bbox": [ + 319, + 329, + 421, + 389 + ], + "type": "image", + "image_path": "15555cfb9f16f6fdc453589fc0c91b42a2258d1260a60ff53f08904d6c5ffc59.jpg" + } + ] + } + ], + "index": 19, + "angle": 0, + "type": "image_body" + } + ], + "index": 19 + }, + { + "type": "image", + "bbox": [ + 446, + 255, + 539, + 323 + ], + "blocks": [ + { + "bbox": [ + 446, + 255, + 539, + 323 + ], + "lines": [ + { + "bbox": [ + 446, + 255, + 539, + 323 + ], + "spans": [ + { + "bbox": [ + 446, + 255, + 539, + 323 + ], + "type": "image", + "image_path": "d9ba0182fe76e9bcf638276fbe614ef41034a2b339947df162dd809ad1cd3fa0.jpg" + } + ] + } + ], + "index": 20, + "angle": 0, + "type": "image_body" + } + ], + "index": 20 + }, + { + "type": "image", + "bbox": [ + 423, + 328, + 539, + 391 + ], + "blocks": [ + { + "bbox": [ + 423, + 328, + 539, + 391 + ], + "lines": [ + { + "bbox": [ + 423, + 328, + 539, + 391 + ], + "spans": [ + { + "bbox": [ + 423, + 328, + 539, + 391 + ], + "type": "image", + "image_path": "42ead853a68a77308aba6a032385977d1aee9b389b6819fe8f30cd69f3b1bd53.jpg" + } + ] + } + ], + "index": 21, + "angle": 0, + "type": "image_body" + } + ], + "index": 21 + }, + { + "type": "image", + "bbox": [ + 74, + 403, + 104, + 431 + ], + "blocks": [ + { + "bbox": [ + 74, + 403, + 104, + 431 + ], + "lines": [ + { + "bbox": [ + 74, + 403, + 104, + 431 + ], + "spans": [ + { + "bbox": [ + 74, + 403, + 104, + 431 + ], + "type": "image", + "image_path": "3d56dfda9c0cd215776b017ae9736d19c24d055bcab58416313922bee5a94489.jpg" + } + ] + } + ], + "index": 22, + "angle": 0, + "type": "image_body" + } + ], + "index": 22 + }, + { + "type": "image", + "bbox": [ + 74, + 432, + 103, + 460 + ], + "blocks": [ + { + "bbox": [ + 74, + 432, + 103, + 460 + ], + "lines": [ + { + "bbox": [ + 74, + 432, + 103, + 460 + ], + "spans": [ + { + "bbox": [ + 74, + 432, + 103, + 460 + ], + "type": "image", + "image_path": "b1b3bbe97ed1d5d38efbfd8c6a271fd8662c88298f9063cb5d973f46e0ab718e.jpg" + } + ] + } + ], + "index": 23, + "angle": 0, + "type": "image_body" + } + ], + "index": 23 + }, + { + "type": "image", + "bbox": [ + 74, + 461, + 103, + 490 + ], + "blocks": [ + { + "bbox": [ + 74, + 461, + 103, + 490 + ], + "lines": [ + { + "bbox": [ + 74, + 461, + 103, + 490 + ], + "spans": [ + { + "bbox": [ + 74, + 461, + 103, + 490 + ], + "type": "image", + "image_path": "c44111bf71e33372cf04e561079591ede87ef0cb656da94d7fb85cf5cd0f6dba.jpg" + } + ] + } + ], + "index": 24, + "angle": 0, + "type": "image_body" + } + ], + "index": 24 + }, + { + "type": "image", + "bbox": [ + 104, + 403, + 130, + 431 + ], + "blocks": [ + { + "bbox": [ + 104, + 403, + 130, + 431 + ], + "lines": [ + { + "bbox": [ + 104, + 403, + 130, + 431 + ], + "spans": [ + { + "bbox": [ + 104, + 403, + 130, + 431 + ], + "type": "image", + "image_path": "621af54fe4b9ca0aa56cb232e2c29f811110ddb5e15082e17e506a654080d459.jpg" + } + ] + } + ], + "index": 25, + "angle": 0, + "type": "image_body" + } + ], + "index": 25 + }, + { + "type": "image", + "bbox": [ + 104, + 432, + 130, + 460 + ], + "blocks": [ + { + "bbox": [ + 104, + 432, + 130, + 460 + ], + "lines": [ + { + "bbox": [ + 104, + 432, + 130, + 460 + ], + "spans": [ + { + "bbox": [ + 104, + 432, + 130, + 460 + ], + "type": "image", + "image_path": "75308784803a2b539f04991f58dc5590242f78993005180f68724760b4e2d4f9.jpg" + } + ] + } + ], + "index": 26, + "angle": 0, + "type": "image_body" + } + ], + "index": 26 + }, + { + "type": "image", + "bbox": [ + 104, + 461, + 130, + 490 + ], + "blocks": [ + { + "bbox": [ + 104, + 461, + 130, + 490 + ], + "lines": [ + { + "bbox": [ + 104, + 461, + 130, + 490 + ], + "spans": [ + { + "bbox": [ + 104, + 461, + 130, + 490 + ], + "type": "image", + "image_path": "8ff9e2401bc30c177a123e84029058340270d40a3e6b1993cc4d906e2ce533e5.jpg" + } + ] + } + ], + "index": 27, + "angle": 0, + "type": "image_body" + } + ], + "index": 27 + }, + { + "type": "image", + "bbox": [ + 131, + 403, + 160, + 431 + ], + "blocks": [ + { + "bbox": [ + 131, + 403, + 160, + 431 + ], + "lines": [ + { + "bbox": [ + 131, + 403, + 160, + 431 + ], + "spans": [ + { + "bbox": [ + 131, + 403, + 160, + 431 + ], + "type": "image", + "image_path": "64b87ce018823ad58c1359ca2b1ff31994f93d3890ad8717504cbb3578ac1657.jpg" + } + ] + } + ], + "index": 28, + "angle": 0, + "type": "image_body" + } + ], + "index": 28 + }, + { + "type": "image", + "bbox": [ + 131, + 432, + 160, + 460 + ], + "blocks": [ + { + "bbox": [ + 131, + 432, + 160, + 460 + ], + "lines": [ + { + "bbox": [ + 131, + 432, + 160, + 460 + ], + "spans": [ + { + "bbox": [ + 131, + 432, + 160, + 460 + ], + "type": "image", + "image_path": "c919abb7dcae4e0ccf25434d3735e9779fcdd4eeae41a41a769f2dbcea89668c.jpg" + } + ] + } + ], + "index": 29, + "angle": 0, + "type": "image_body" + } + ], + "index": 29 + }, + { + "type": "image", + "bbox": [ + 70, + 492, + 147, + 573 + ], + "blocks": [ + { + "bbox": [ + 70, + 492, + 147, + 573 + ], + "lines": [ + { + "bbox": [ + 70, + 492, + 147, + 573 + ], + "spans": [ + { + "bbox": [ + 70, + 492, + 147, + 573 + ], + "type": "image", + "image_path": "a9ea49154a5cddebf7b1215c95540c58c714ca3a87f5f9090d30a8f5a9a5356c.jpg" + } + ] + } + ], + "index": 30, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 55, + 584, + 555, + 639 + ], + "lines": [ + { + "bbox": [ + 55, + 584, + 555, + 639 + ], + "spans": [ + { + "bbox": [ + 55, + 584, + 555, + 639 + ], + "type": "text", + "content": "Figure 3. Qualitative Comparison on Camera Pose Accuracy and Predicted Geometry. For each method, we plot the ground-truth cameras in black and the predicted cameras in other colors. DiffusionSfM demonstrates robust performance even with challenging inputs. Compared to DUSt3R, which sometimes fails to register images in a consistent manner, DiffusionSfM consistently yields a coherent global prediction. Additionally, while we observe that DUSt3R can predict highly precise camera rotations, it often struggles with camera centers (see the backpack example). Input images depicting scenes are out-of-distribution for RayDiffusion, as it is trained on CO3D only." + } + ] + } + ], + "index": 37, + "angle": 0, + "type": "image_caption" + } + ], + "index": 30 + }, + { + "type": "image", + "bbox": [ + 185, + 414, + 287, + 476 + ], + "blocks": [ + { + "bbox": [ + 185, + 414, + 287, + 476 + ], + "lines": [ + { + "bbox": [ + 185, + 414, + 287, + 476 + ], + "spans": [ + { + "bbox": [ + 185, + 414, + 287, + 476 + ], + "type": "image", + "image_path": "1737d179f0ff07dcf799bb9ca29f91d33c878cd8e84667b551b4f902f45aa969.jpg" + } + ] + } + ], + "index": 31, + "angle": 0, + "type": "image_body" + } + ], + "index": 31 + }, + { + "type": "image", + "bbox": [ + 313, + 414, + 414, + 476 + ], + "blocks": [ + { + "bbox": [ + 313, + 414, + 414, + 476 + ], + "lines": [ + { + "bbox": [ + 313, + 414, + 414, + 476 + ], + "spans": [ + { + "bbox": [ + 313, + 414, + 414, + 476 + ], + "type": "image", + "image_path": "b07132582448c16a63a6b68ee425225bc70df6e576c168687035e4f1ffa818fa.jpg" + } + ] + } + ], + "index": 32, + "angle": 0, + "type": "image_body" + } + ], + "index": 32 + }, + { + "type": "image", + "bbox": [ + 440, + 414, + 541, + 479 + ], + "blocks": [ + { + "bbox": [ + 440, + 414, + 541, + 479 + ], + "lines": [ + { + "bbox": [ + 440, + 414, + 541, + 479 + ], + "spans": [ + { + "bbox": [ + 440, + 414, + 541, + 479 + ], + "type": "image", + "image_path": "355211bdc5d80708987da0be50e094a64357cea1d97bc362e30bbd5e8f491890.jpg" + } + ] + } + ], + "index": 33, + "angle": 0, + "type": "image_body" + } + ], + "index": 33 + }, + { + "type": "image", + "bbox": [ + 198, + 491, + 273, + 573 + ], + "blocks": [ + { + "bbox": [ + 198, + 491, + 273, + 573 + ], + "lines": [ + { + "bbox": [ + 198, + 491, + 273, + 573 + ], + "spans": [ + { + "bbox": [ + 198, + 491, + 273, + 573 + ], + "type": "image", + "image_path": "052155bee2775d35833066115b297323f753cce3a22d8fe57f655f5794fe4c81.jpg" + } + ] + } + ], + "index": 34, + "angle": 0, + "type": "image_body" + } + ], + "index": 34 + }, + { + "type": "image", + "bbox": [ + 335, + 491, + 405, + 572 + ], + "blocks": [ + { + "bbox": [ + 335, + 491, + 405, + 572 + ], + "lines": [ + { + "bbox": [ + 335, + 491, + 405, + 572 + ], + "spans": [ + { + "bbox": [ + 335, + 491, + 405, + 572 + ], + "type": "image", + "image_path": "9e2d0801addb462fc02b573b35ec15afd43788406b3ddcbf55c1a5652275a7ab.jpg" + } + ] + } + ], + "index": 35, + "angle": 0, + "type": "image_body" + } + ], + "index": 35 + }, + { + "type": "image", + "bbox": [ + 456, + 491, + 536, + 573 + ], + "blocks": [ + { + "bbox": [ + 456, + 491, + 536, + 573 + ], + "lines": [ + { + "bbox": [ + 456, + 491, + 536, + 573 + ], + "spans": [ + { + "bbox": [ + 456, + 491, + 536, + 573 + ], + "type": "image", + "image_path": "95779d7f8c4d99450912c29e23c3a409997915c022968df709dfd3b6abb7c648.jpg" + } + ] + } + ], + "index": 36, + "angle": 0, + "type": "image_body" + } + ], + "index": 36 + }, + { + "bbox": [ + 55, + 650, + 294, + 710 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 650, + 294, + 710 + ], + "spans": [ + { + "bbox": [ + 55, + 650, + 294, + 710 + ], + "type": "text", + "content": "datasets used for DUSt3R [36], excluding Waymo [31] due to its excessively sparse depth maps. The included datasets are Habitat [24], CO3D [21], ScanNet++ [39], ArkitScenes [1], Static Scenes 3D [17], MegaDepth [11], and BlendedMVS [38]. We follow the DUSt3R official repository [37]" + } + ] + } + ], + "index": 38 + }, + { + "bbox": [ + 313, + 651, + 555, + 710 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 651, + 555, + 710 + ], + "spans": [ + { + "bbox": [ + 313, + 651, + 555, + 710 + ], + "type": "text", + "content": "guidelines to extract image pairs, ensuring reasonable overlap between paired images. To construct multi-view data given these pre-computed pairs, we iteratively sample images using an adjacency matrix, maintaining sufficient overlap with the selected set. Beyond CO3D, this model variant" + } + ] + } + ], + "index": 39 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 749, + 314, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 749, + 314, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 749, + 314, + 757 + ], + "type": "text", + "content": "6321" + } + ] + } + ], + "index": 40 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 4 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 72, + 87, + 152, + 398 + ], + "blocks": [ + { + "bbox": [ + 83, + 71, + 140, + 82 + ], + "lines": [ + { + "bbox": [ + 83, + 71, + 140, + 82 + ], + "spans": [ + { + "bbox": [ + 83, + 71, + 140, + 82 + ], + "type": "text", + "content": "Input Images" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 72, + 87, + 152, + 398 + ], + "lines": [ + { + "bbox": [ + 72, + 87, + 152, + 398 + ], + "spans": [ + { + "bbox": [ + 72, + 87, + 152, + 398 + ], + "type": "image", + "image_path": "8113542d26e8415e7fccb43a4ab137fe288e9ddfb3159e473e2445238d6cf63d.jpg" + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_body" + } + ], + "index": 1 + }, + { + "type": "image", + "bbox": [ + 158, + 95, + 260, + 162 + ], + "blocks": [ + { + "bbox": [ + 186, + 71, + 243, + 81 + ], + "lines": [ + { + "bbox": [ + 186, + 71, + 243, + 81 + ], + "spans": [ + { + "bbox": [ + 186, + 71, + 243, + 81 + ], + "type": "text", + "content": "DiffusionSfM" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 158, + 95, + 260, + 162 + ], + "lines": [ + { + "bbox": [ + 158, + 95, + 260, + 162 + ], + "spans": [ + { + "bbox": [ + 158, + 95, + 260, + 162 + ], + "type": "image", + "image_path": "933b09eb7e3527150f3be3f259b770ea88c020ef8b967f7a83a4d6eafe070ecc.jpg" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_body" + } + ], + "index": 3 + }, + { + "type": "image", + "bbox": [ + 296, + 94, + 394, + 158 + ], + "blocks": [ + { + "bbox": [ + 320, + 71, + 376, + 82 + ], + "lines": [ + { + "bbox": [ + 320, + 71, + 376, + 82 + ], + "spans": [ + { + "bbox": [ + 320, + 71, + 376, + 82 + ], + "type": "text", + "content": "RayDiffusion" + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 296, + 94, + 394, + 158 + ], + "lines": [ + { + "bbox": [ + 296, + 94, + 394, + 158 + ], + "spans": [ + { + "bbox": [ + 296, + 94, + 394, + 158 + ], + "type": "image", + "image_path": "ebb549fcdbcce39636c90442061595ecb9be761d0dac36f498c4259fa3e9299b.jpg" + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "image_body" + } + ], + "index": 5 + }, + { + "type": "image", + "bbox": [ + 430, + 95, + 531, + 161 + ], + "blocks": [ + { + "bbox": [ + 465, + 71, + 502, + 81 + ], + "lines": [ + { + "bbox": [ + 465, + 71, + 502, + 81 + ], + "spans": [ + { + "bbox": [ + 465, + 71, + 502, + 81 + ], + "type": "text", + "content": "DUSt3R" + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 430, + 95, + 531, + 161 + ], + "lines": [ + { + "bbox": [ + 430, + 95, + 531, + 161 + ], + "spans": [ + { + "bbox": [ + 430, + 95, + 531, + 161 + ], + "type": "image", + "image_path": "3d9b95e9ad7374161e5cbf61c41771e2e6b56aedb7c1c0789d6480c724ed3950.jpg" + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "image_body" + } + ], + "index": 7 + }, + { + "type": "image", + "bbox": [ + 168, + 182, + 255, + 247 + ], + "blocks": [ + { + "bbox": [ + 168, + 182, + 255, + 247 + ], + "lines": [ + { + "bbox": [ + 168, + 182, + 255, + 247 + ], + "spans": [ + { + "bbox": [ + 168, + 182, + 255, + 247 + ], + "type": "image", + "image_path": "c9e3727a7835cb21302b9686d7fc17c3b5fa3dc25d6a1046592ee7a962cf2cad.jpg" + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "image_body" + } + ], + "index": 8 + }, + { + "type": "image", + "bbox": [ + 301, + 183, + 404, + 247 + ], + "blocks": [ + { + "bbox": [ + 301, + 183, + 404, + 247 + ], + "lines": [ + { + "bbox": [ + 301, + 183, + 404, + 247 + ], + "spans": [ + { + "bbox": [ + 301, + 183, + 404, + 247 + ], + "type": "image", + "image_path": "3ba083c2b95cf328bae57df050f1ef358611952760041f8e249fa443105d4deb.jpg" + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "image_body" + } + ], + "index": 9 + }, + { + "type": "image", + "bbox": [ + 442, + 178, + 538, + 248 + ], + "blocks": [ + { + "bbox": [ + 442, + 178, + 538, + 248 + ], + "lines": [ + { + "bbox": [ + 442, + 178, + 538, + 248 + ], + "spans": [ + { + "bbox": [ + 442, + 178, + 538, + 248 + ], + "type": "image", + "image_path": "391548a854ffbf9d9e3b6c9a89f4c7f4a8c2303a99b0f5684284a04001ea7fd6.jpg" + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "image_body" + } + ], + "index": 10 + }, + { + "type": "image", + "bbox": [ + 71, + 258, + 152, + 312 + ], + "blocks": [ + { + "bbox": [ + 71, + 258, + 152, + 312 + ], + "lines": [ + { + "bbox": [ + 71, + 258, + 152, + 312 + ], + "spans": [ + { + "bbox": [ + 71, + 258, + 152, + 312 + ], + "type": "image", + "image_path": "71982fd3c12d0817f39efa883fc6e901683dd4fcb9958618f0e0e0c14e24d0ec.jpg" + } + ] + } + ], + "index": 11, + "angle": 0, + "type": "image_body" + } + ], + "index": 11 + }, + { + "type": "image", + "bbox": [ + 161, + 258, + 257, + 316 + ], + "blocks": [ + { + "bbox": [ + 161, + 258, + 257, + 316 + ], + "lines": [ + { + "bbox": [ + 161, + 258, + 257, + 316 + ], + "spans": [ + { + "bbox": [ + 161, + 258, + 257, + 316 + ], + "type": "image", + "image_path": "01f965a7ac2be332782e20e3ebbaef854eed14bc732b6658975749ecb53a0bb9.jpg" + } + ] + } + ], + "index": 12, + "angle": 0, + "type": "image_body" + } + ], + "index": 12 + }, + { + "type": "image", + "bbox": [ + 293, + 250, + 404, + 311 + ], + "blocks": [ + { + "bbox": [ + 293, + 250, + 404, + 311 + ], + "lines": [ + { + "bbox": [ + 293, + 250, + 404, + 311 + ], + "spans": [ + { + "bbox": [ + 293, + 250, + 404, + 311 + ], + "type": "image", + "image_path": "3a0473971df79c5449389de83b9765f9583258dd642a4923b3125fc4447d8398.jpg" + } + ] + } + ], + "index": 13, + "angle": 0, + "type": "image_body" + } + ], + "index": 13 + }, + { + "type": "image", + "bbox": [ + 430, + 259, + 541, + 320 + ], + "blocks": [ + { + "bbox": [ + 430, + 259, + 541, + 320 + ], + "lines": [ + { + "bbox": [ + 430, + 259, + 541, + 320 + ], + "spans": [ + { + "bbox": [ + 430, + 259, + 541, + 320 + ], + "type": "image", + "image_path": "fc70de8b11828fc1a07930f61b93d9c100c1903b7fb41fcd5d39ea6065293654.jpg" + } + ] + } + ], + "index": 14, + "angle": 0, + "type": "image_body" + } + ], + "index": 14 + }, + { + "type": "image", + "bbox": [ + 71, + 316, + 152, + 370 + ], + "blocks": [ + { + "bbox": [ + 71, + 316, + 152, + 370 + ], + "lines": [ + { + "bbox": [ + 71, + 316, + 152, + 370 + ], + "spans": [ + { + "bbox": [ + 71, + 316, + 152, + 370 + ], + "type": "image", + "image_path": "e62c7dc2190fad5f4b7bdae298e3183cc133b3b08212eed26461b84525924ed5.jpg" + } + ] + } + ], + "index": 15, + "angle": 0, + "type": "image_body" + } + ], + "index": 15 + }, + { + "type": "image", + "bbox": [ + 163, + 323, + 263, + 385 + ], + "blocks": [ + { + "bbox": [ + 163, + 323, + 263, + 385 + ], + "lines": [ + { + "bbox": [ + 163, + 323, + 263, + 385 + ], + "spans": [ + { + "bbox": [ + 163, + 323, + 263, + 385 + ], + "type": "image", + "image_path": "12bc1a5e9df8ccb453a2c73ad7e51bab942f03cf2f0f7966ec16278b98e03951.jpg" + } + ] + } + ], + "index": 16, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 55, + 407, + 553, + 419 + ], + "lines": [ + { + "bbox": [ + 55, + 407, + 553, + 419 + ], + "spans": [ + { + "bbox": [ + 55, + 407, + 553, + 419 + ], + "type": "text", + "content": "Figure 4. Additional Qualitative Results on Predicted Camera Poses. DiffusionSfM shows robustness to ambiguous patterns in inputs." + } + ] + } + ], + "index": 19, + "angle": 0, + "type": "image_caption" + } + ], + "index": 16 + }, + { + "type": "image", + "bbox": [ + 296, + 324, + 399, + 384 + ], + "blocks": [ + { + "bbox": [ + 296, + 324, + 399, + 384 + ], + "lines": [ + { + "bbox": [ + 296, + 324, + 399, + 384 + ], + "spans": [ + { + "bbox": [ + 296, + 324, + 399, + 384 + ], + "type": "image", + "image_path": "43d9c1f6b2fb34cfb80eb04d4a1eace850a5c28b1428f2923e0531d60d3fcbfe.jpg" + } + ] + } + ], + "index": 17, + "angle": 0, + "type": "image_body" + } + ], + "index": 17 + }, + { + "type": "image", + "bbox": [ + 433, + 324, + 533, + 384 + ], + "blocks": [ + { + "bbox": [ + 433, + 324, + 533, + 384 + ], + "lines": [ + { + "bbox": [ + 433, + 324, + 533, + 384 + ], + "spans": [ + { + "bbox": [ + 433, + 324, + 533, + 384 + ], + "type": "image", + "image_path": "0fb6fb31f2c126f48b00a94419cbada66bd471bb4696337619753970c8eb7ed5.jpg" + } + ] + } + ], + "index": 18, + "angle": 0, + "type": "image_body" + } + ], + "index": 18 + }, + { + "bbox": [ + 55, + 440, + 266, + 452 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 440, + 266, + 452 + ], + "spans": [ + { + "bbox": [ + 55, + 440, + 266, + 452 + ], + "type": "text", + "content": "is also evaluated on Habitat and RealEstate10k [42]." + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 55, + 462, + 296, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 462, + 296, + 713 + ], + "spans": [ + { + "bbox": [ + 55, + 462, + 296, + 713 + ], + "type": "text", + "content": "Baselines and Metrics. To evaluate camera pose accuracy in the sparse-view setup, we compare with RayDiffusion [41] and DUSt3R [36], along with previous methods [12, 22, 34]. DUSt3R is trained initially on a mixture of eight datasets, while most other baselines are trained only on CO3D. For a fair and comprehensive comparison, we re-train DUSt3R on the 41-10 split of CO3D, using the authors' official implementation [37] and hyperparameters (referred to as DUSt3R-CO3D). To evaluate camera predictions, we follow prior work [41] and convert predicted rays back to traditional extrinsic matrices and report two pose accuracy metrics: (1) Camera Rotation Accuracy which compares the predicted relative camera rotation between images against ground truth and (2) Camera Center Accuracy which compares predicted camera centers to the ground truth after a similarity alignment. To evaluate the estimated geometry (i.e., ray endpoints), we report Chamfer Distance (CD). Additionally, to explore whether combining an off-the-shelf sparse-view pose estimation method (RayDiffusion) with a monocular depth estimation model (MoGe [35]) is sufficient to infer 3D scene geometry from" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 313, + 440, + 553, + 477 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 440, + 553, + 477 + ], + "spans": [ + { + "bbox": [ + 313, + 440, + 553, + 477 + ], + "type": "text", + "content": "multiple images, we introduce a new baseline, " + }, + { + "bbox": [ + 313, + 440, + 553, + 477 + ], + "type": "inline_equation", + "content": "\\mathsf{RD} + \\mathsf{MoGe}" + }, + { + "bbox": [ + 313, + 440, + 553, + 477 + ], + "type": "text", + "content": ". We include more details in the supplementary. Unless otherwise specified, all evaluations use 2-8 input images." + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 313, + 490, + 434, + 502 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 490, + 434, + 502 + ], + "spans": [ + { + "bbox": [ + 313, + 490, + 434, + 502 + ], + "type": "text", + "content": "4.2. Evaluation on CO3D" + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 312, + 510, + 555, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 312, + 510, + 555, + 713 + ], + "spans": [ + { + "bbox": [ + 312, + 510, + 555, + 713 + ], + "type": "text", + "content": "Camera Pose Accuracy. We present the quantitative results in Tab. 1. For camera rotation accuracy, both versions of DiffusionSfM outperform all baselines, except that DUSt3R achieves slightly higher overall accuracy than DiffusionSfM-CO3D on unseen categories – likely due to its training on more data. For camera center accuracy, our approach consistently outperforms all other methods. We hypothesize that these gains stem from our explicit modeling of camera centers through ray origin prediction. The qualitative results in Fig. 3 illustrate that DiffusionSfM produces robust predictions given challenging inputs, whereas DUSt3R produces inaccurate results. We attribute this improvement to our model's probabilistic modeling capability derived from diffusion, as well as its multi-view reasoning abilities, which together effectively handle these challenging scenarios. While we observe that DUSt3R can predict highly precise camera rotations, it often struggles with cam-" + } + ] + } + ], + "index": 24 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "text", + "content": "6322" + } + ] + } + ], + "index": 25 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 5 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 56, + 70, + 552, + 286 + ], + "blocks": [ + { + "bbox": [ + 56, + 70, + 552, + 286 + ], + "lines": [ + { + "bbox": [ + 56, + 70, + 552, + 286 + ], + "spans": [ + { + "bbox": [ + 56, + 70, + 552, + 286 + ], + "type": "table", + "html": "
# of ImagesRotation Accuracy (↑, @ 15°)Center Accuracy (↑, @ 0.1)
2345678234567
Seen CategoriesCOLMAP [22]30.728.426.526.827.028.130.610034.523.818.915.614.5
PoseDiffusion [34]75.776.476.877.478.078.778.810077.569.765.963.762.8
RelPose++ [12]81.882.884.184.784.985.385.510085.078.074.271.970.3
RayDiffusion [41]91.892.492.692.993.193.393.310094.290.587.886.285.0
DUSt3R-CO3D [36]86.787.988.088.288.688.888.910092.086.883.882.081.1
DUSt3R [36]91.792.793.393.693.894.094.310093.085.781.979.677.8
DiffusionSfM-CO3D93.494.094.594.895.095.295.110095.993.692.291.290.7
DiffusionSfM92.493.594.094.594.795.195.210095.392.991.991.290.8
Unseen CategoriesCOLMAP [22]34.531.831.031.732.735.038.510036.025.520.017.917.6
PoseDiffusion [34]63.264.264.265.766.267.067.710063.650.545.743.041.2
RelPose++ [12]69.871.171.972.873.874.474.910070.658.853.450.447.8
RayDiffusion [41]83.585.686.386.987.287.588.110087.781.177.074.172.4
DUSt3R-CO3D [36]79.881.582.682.783.083.383.710083.677.271.870.068.1
DUSt3R [36]90.892.693.693.693.893.693.410087.979.874.371.769.4
DiffusionSfM-CO3D90.491.292.793.093.193.393.510091.187.785.383.782.7
DiffusionSfM91.592.893.693.994.194.294.410092.487.986.285.384.1
", + "image_path": "a33cb5d7197aa3e345a5d224916c27764df2b9ba8ce3e1bee23582ecf8870ce7.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "table_body" + } + ], + "index": 0 + }, + { + "bbox": [ + 54, + 365, + 295, + 389 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 54, + 365, + 295, + 389 + ], + "spans": [ + { + "bbox": [ + 54, + 365, + 295, + 389 + ], + "type": "text", + "content": "era centers (see the backpack example). Additional qualitative results on camera accuracy are provided in Fig.4." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 56, + 391, + 295, + 594 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 391, + 295, + 594 + ], + "spans": [ + { + "bbox": [ + 56, + 391, + 295, + 594 + ], + "type": "text", + "content": "Predicted Geometry. To evaluate predicted geometry, we compute Chamfer Distance (CD) and show comparisons against baselines in Tab. 2. We compute CD in two setups (with and without foreground object mask), and find that DiffusionSfM-CO3D performs best without foreground masking. In this setup, the predicted ray endpoints corresponding to the background image pixels tend to have larger coordinate values than foreground ones, and therefore dominate CD. This result indicates that our model provides more accurate predictions for complex image backgrounds. In terms of CD with masking, DUSt3R achieves the best result, while our two model variants outperform DUSt3R-CO3D. See Fig. 3 for visualization. We also notice that naively combining RayDiffusion poses with estimated depths from MoGe does not work well, as the inferred depths are potentially inconsistent across different views, thus resulting in degraded performance." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 55, + 609, + 244, + 623 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 609, + 244, + 623 + ], + "spans": [ + { + "bbox": [ + 55, + 609, + 244, + 623 + ], + "type": "text", + "content": "4.3. Evaluation on Scene-Level Datasets" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 55, + 629, + 295, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 629, + 295, + 715 + ], + "spans": [ + { + "bbox": [ + 55, + 629, + 295, + 715 + ], + "type": "text", + "content": "Beyond the object-centric CO3D dataset [21], we compare DiffusionSfM against DUSt3R on two scene-level datasets: Habitat (in-distribution) [24] and RealEstate10k (out-of-distribution) [42]. The results are presented in Tab. 3. While DiffusionSfM achieves comparable camera rotation accuracy to DUSt3R, it consistently predicts more accurate camera centers, aligning with our findings on CO3D." + } + ] + } + ], + "index": 5 + }, + { + "type": "table", + "bbox": [ + 316, + 362, + 553, + 502 + ], + "blocks": [ + { + "bbox": [ + 55, + 293, + 555, + 350 + ], + "lines": [ + { + "bbox": [ + 55, + 293, + 555, + 350 + ], + "spans": [ + { + "bbox": [ + 55, + 293, + 555, + 350 + ], + "type": "text", + "content": "Table 1. Camera Rotation and Center Accuracy on CO3D. On the left, we report the proportion of relative camera rotations within " + }, + { + "bbox": [ + 55, + 293, + 555, + 350 + ], + "type": "inline_equation", + "content": "15^{\\circ}" + }, + { + "bbox": [ + 55, + 293, + 555, + 350 + ], + "type": "text", + "content": " of the ground truth. On the right, we report the proportion of camera centers within " + }, + { + "bbox": [ + 55, + 293, + 555, + 350 + ], + "type": "inline_equation", + "content": "10\\%" + }, + { + "bbox": [ + 55, + 293, + 555, + 350 + ], + "type": "text", + "content": " of the scene scale, relative to the ground truth. To align the predicted camera centers to ground truth, we apply an optimal similarity transform " + }, + { + "bbox": [ + 55, + 293, + 555, + 350 + ], + "type": "inline_equation", + "content": "(s, \\mathbf{R}, \\mathbf{t})" + }, + { + "bbox": [ + 55, + 293, + 555, + 350 + ], + "type": "text", + "content": ", hence the alignment is perfect at " + }, + { + "bbox": [ + 55, + 293, + 555, + 350 + ], + "type": "inline_equation", + "content": "N = 2" + }, + { + "bbox": [ + 55, + 293, + 555, + 350 + ], + "type": "text", + "content": " but worsens with more images. DiffusionSfM outperforms all other methods for camera center accuracy, and outperforms all methods trained on equivalent data for rotation accuracy." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 316, + 362, + 553, + 502 + ], + "lines": [ + { + "bbox": [ + 316, + 362, + 553, + 502 + ], + "spans": [ + { + "bbox": [ + 316, + 362, + 553, + 502 + ], + "type": "table", + "html": "
# of Images2345678
RD*+MoGe [41]0.0590.0640.0710.0620.0630.0610.061
DUSt3R* [36]0.0360.0370.0400.0400.0370.0360.039
DUSt3R [36]0.0210.0230.0240.0240.0250.0250.023
DiffusionSfM*0.0190.0210.0220.0220.0210.0210.021
DiffusionSfM0.0250.0250.0250.0250.0260.0250.028
RD*+MoGe [41]0.0710.0750.0680.0670.0660.0640.064
DUSt3R* [36]0.0380.0360.0360.0360.0340.0330.034
DUSt3R [36]0.0230.0220.0190.0200.0190.0200.020
DiffusionSfM*0.0280.0250.0240.0240.0240.0230.022
DiffusionSfM0.0270.0240.0230.0220.0220.0220.021
", + "image_path": "c4ea6f4ebe82bd9ce2a893b5358e6625bbd993f45aeae01820fb06464e3d1fe8.jpg" + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "table_body" + } + ], + "index": 6 + }, + { + "bbox": [ + 313, + 510, + 555, + 578 + ], + "lines": [ + { + "bbox": [ + 313, + 510, + 555, + 578 + ], + "spans": [ + { + "bbox": [ + 313, + 510, + 555, + 578 + ], + "type": "text", + "content": "Table 2. Chamfer Distance " + }, + { + "bbox": [ + 313, + 510, + 555, + 578 + ], + "type": "inline_equation", + "content": "(\\downarrow)" + }, + { + "bbox": [ + 313, + 510, + 555, + 578 + ], + "type": "text", + "content": " on CO3D Unseen Categories. Top: CD computed on all scene points. Bottom: CD computed on foreground points only. Models marked with “*” are trained on CO3D only, while those without are trained on multiple datasets. Note that top and bottom values are not directly comparable, as each ground-truth point cloud is individually normalized." + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "text" + }, + { + "bbox": [ + 313, + 599, + 411, + 613 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 599, + 411, + 613 + ], + "spans": [ + { + "bbox": [ + 313, + 599, + 411, + 613 + ], + "type": "text", + "content": "4.4. Inference Speed" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 313, + 617, + 555, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 617, + 555, + 715 + ], + "spans": [ + { + "bbox": [ + 313, + 617, + 555, + 715 + ], + "type": "text", + "content": "Though our method requires iterative diffusion denoising at inference, we can speed this up by performing early stopping. Specifically, we can treat the " + }, + { + "bbox": [ + 313, + 617, + 555, + 715 + ], + "type": "inline_equation", + "content": "x_0" + }, + { + "bbox": [ + 313, + 617, + 555, + 715 + ], + "type": "text", + "content": "-prediction from early timesteps as our output instead of iterating over all denoising timesteps. Consistent with observations by Zhang et al. [41], this in fact yields more accurate predictions than the final-step diffusion outputs. As a result we only require 10 denoising diffusion timesteps for inference, taking 1.91" + } + ] + } + ], + "index": 9 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "text", + "content": "6323" + } + ] + } + ], + "index": 10 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 6 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 58, + 70, + 293, + 122 + ], + "blocks": [ + { + "bbox": [ + 58, + 70, + 293, + 122 + ], + "lines": [ + { + "bbox": [ + 58, + 70, + 293, + 122 + ], + "spans": [ + { + "bbox": [ + 58, + 70, + 293, + 122 + ], + "type": "table", + "html": "
Hab.DUSt3R97.0/10095.0/97.694.3/95.094.2/93.1
DiffusionSfM92.6/10093.7/99.194.2/98.894.9/98.6
RE10DUSt3R98.1/10097.7/68.797.6/57.997.7/53.3
DiffusionSfM97.8/10097.5/75.697.7/66.297.9/61.9
", + "image_path": "a5b3d4e308ef80db7d45571ba6f502b036de286a0336817e442fa29c3ab997cf.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "table_body" + } + ], + "index": 0 + }, + { + "type": "table", + "bbox": [ + 56, + 210, + 295, + 257 + ], + "blocks": [ + { + "bbox": [ + 55, + 129, + 295, + 196 + ], + "lines": [ + { + "bbox": [ + 55, + 129, + 295, + 196 + ], + "spans": [ + { + "bbox": [ + 55, + 129, + 295, + 196 + ], + "type": "text", + "content": "Table 3. Camera Rotation and Center Accuracy on Two Scene-Level Datasets. Top: Habitat (2-5 views). Bottom: RealEstate10k (2, 4, 6, 8 views). Each grid reports camera rotation accuracy (left, " + }, + { + "bbox": [ + 55, + 129, + 295, + 196 + ], + "type": "inline_equation", + "content": "\\uparrow" + }, + { + "bbox": [ + 55, + 129, + 295, + 196 + ], + "type": "text", + "content": ") and center accuracy (right, " + }, + { + "bbox": [ + 55, + 129, + 295, + 196 + ], + "type": "inline_equation", + "content": "\\uparrow" + }, + { + "bbox": [ + 55, + 129, + 295, + 196 + ], + "type": "text", + "content": "). While DiffusionSfM performs on par with DUSt3R in rotation accuracy, it consistently surpasses DUSt3R in center accuracy." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 56, + 210, + 295, + 257 + ], + "lines": [ + { + "bbox": [ + 56, + 210, + 295, + 257 + ], + "spans": [ + { + "bbox": [ + 56, + 210, + 295, + 257 + ], + "type": "table", + "html": "
# of Images2345678
DiffusionSfM*0.0190.0210.0220.0220.0210.0210.021
w/o Mask0.0200.0220.0230.0230.0220.0220.024
", + "image_path": "218890669fb73a4f9cd99bd9c950b2301f5873ae02015299bbe805c9375c2644.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "table_body" + } + ], + "index": 2 + }, + { + "bbox": [ + 55, + 265, + 295, + 332 + ], + "lines": [ + { + "bbox": [ + 55, + 265, + 295, + 332 + ], + "spans": [ + { + "bbox": [ + 55, + 265, + 295, + 332 + ], + "type": "text", + "content": "Table 4. Ablation Study on GT Mask Conditioning for CO3D Unseen Categories. We assess the effect of replacing GT mask conditioning with depth interpolation in our CO3D variant (DiffusionSfM*), by reporting the CD for predicted geometry. Incorporating mask conditioning to indicate missing data during training improves geometry quality." + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "text" + }, + { + "bbox": [ + 55, + 357, + 296, + 405 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 357, + 296, + 405 + ], + "spans": [ + { + "bbox": [ + 55, + 357, + 296, + 405 + ], + "type": "text", + "content": "seconds on a single A5000 GPU with 8 input images. In contrast, DUSt3R takes 8.73 seconds to run the complete pairwise inference and global alignment procedure. We provide additional analysis in the supplementary material." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 55, + 417, + 149, + 430 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 417, + 149, + 430 + ], + "spans": [ + { + "bbox": [ + 55, + 417, + 149, + 430 + ], + "type": "text", + "content": "4.5. Ablation Study" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 55, + 436, + 295, + 520 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 436, + 295, + 520 + ], + "spans": [ + { + "bbox": [ + 55, + 436, + 295, + 520 + ], + "type": "text", + "content": "Homogeneous Coordinates. The use of homogeneous coordinates for ray origins and endpoints is crucial for stable model training. To assess its impact, we replace the proposed homogeneous representation with standard 3D coordinates in " + }, + { + "bbox": [ + 55, + 436, + 295, + 520 + ], + "type": "inline_equation", + "content": "\\mathbb{R}^3" + }, + { + "bbox": [ + 55, + 436, + 295, + 520 + ], + "type": "text", + "content": ". Our experiments show that this variant is difficult to train and fails to converge. Further details on this experiment are provided in the supplementary material." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 55, + 521, + 295, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 521, + 295, + 713 + ], + "spans": [ + { + "bbox": [ + 55, + 521, + 295, + 713 + ], + "type": "text", + "content": "GT Mask Conditioning. To evaluate the effectiveness of the proposed GT mask conditioning, we train a baseline model without using this strategy. For missing values in the ground-truth depth maps, we use nearest-neighbor interpolation to fill in invalid pixels. This experiment is conducted on the CO3D dataset, with results presented in Tab. 4. The findings show that removing GT mask conditioning consistently degrades predicted geometry across varying numbers of input views, even when the loss is still masked. While interpolation can effectively fill missing depth within object regions, it often introduces substantial noise in the background (e.g., the sky). This noise negatively impacts diffusion model training, as the entire ray origin and endpoint maps are used as input. We anticipate an even larger performance gap on non-object-centric datasets with more outdoor scenes, such as MegaDepth [11]." + } + ] + } + ], + "index": 7 + }, + { + "type": "image", + "bbox": [ + 316, + 69, + 556, + 142 + ], + "blocks": [ + { + "bbox": [ + 316, + 69, + 556, + 142 + ], + "lines": [ + { + "bbox": [ + 316, + 69, + 556, + 142 + ], + "spans": [ + { + "bbox": [ + 316, + 69, + 556, + 142 + ], + "type": "image", + "image_path": "73e85658c1c87772133020c3736a2822b53d306c91f74b826fc325128df9097e.jpg" + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 313, + 149, + 555, + 227 + ], + "lines": [ + { + "bbox": [ + 313, + 149, + 555, + 227 + ], + "spans": [ + { + "bbox": [ + 313, + 149, + 555, + 227 + ], + "type": "text", + "content": "Figure 5. Multi-modality of DiffusionSfM. We show two distinct samples from DiffusionSfM, starting from the same input images but with different random noise. Sample 1 explains the input images by putting all flowers on the left side, while Sample 2 places one flower on each side (note the difference in the green camera's viewpoint). DiffusionSfM is able to predict multi-modal geometry distributions when the scene layout is ambiguous in the inputs." + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "image_caption" + } + ], + "index": 8 + }, + { + "bbox": [ + 313, + 247, + 524, + 261 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 247, + 524, + 261 + ], + "spans": [ + { + "bbox": [ + 313, + 247, + 524, + 261 + ], + "type": "text", + "content": "4.6. Multi-modality from Multiple Sampling" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 313, + 265, + 555, + 373 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 265, + 555, + 373 + ], + "spans": [ + { + "bbox": [ + 313, + 265, + 555, + 373 + ], + "type": "text", + "content": "Diffusion models enable the generation of diverse samples from challenging input images. For instance, in Fig. 5, the vase exhibits symmetric patterns, and we present two distinct predicted endpoints from DiffusionSfM, each offering a different interpretation of the flowers in the image. Compared to regression models such as DUSt3R [8], DiffusionSfM is better suited for handling uncertainty – an inherent aspect of our task – where a sparse set of input images can correspond to multiple plausible 3D scene geometry." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 313, + 383, + 384, + 396 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 383, + 384, + 396 + ], + "spans": [ + { + "bbox": [ + 313, + 383, + 384, + 396 + ], + "type": "text", + "content": "5. Discussion" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 313, + 403, + 554, + 451 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 403, + 554, + 451 + ], + "spans": [ + { + "bbox": [ + 313, + 403, + 554, + 451 + ], + "type": "text", + "content": "We present DiffusionSfM and demonstrate that it recovers accurate predictions of both cameras and geometry from multi-view inputs. Although our results are promising, several challenges and open questions remain." + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 313, + 451, + 554, + 594 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 451, + 554, + 594 + ], + "spans": [ + { + "bbox": [ + 313, + 451, + 554, + 594 + ], + "type": "text", + "content": "Notably, DiffusionSfM employs a pixel-space diffusion model, in contrast to the latent-space models adopted by state-of-the-art T2I generative systems. Operating in pixel space may require greater model capacity, yet our current model remains relatively small - potentially explaining the noisy patterns observed along object boundaries. Learning an expressive latent space for ray origins and endpoints by training a VAE could be a promising direction for future work. In addition, the compute requirement in multi-view transformers scales quadratically with the number of input images: one would require masked attention to deploy systems like ours for a large set of input images." + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 313, + 594, + 556, + 703 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 594, + 556, + 703 + ], + "spans": [ + { + "bbox": [ + 313, + 594, + 556, + 703 + ], + "type": "text", + "content": "Despite these challenges, we believe that our work highlights the potential of a unified approach for multi-view geometry tasks. We envision that our approach can be built upon to train a common system across related geometric tasks, such as SfM (input images with unknown origins and endpoints), registration (some images have known origins and endpoints, whereas others don't), mapping (known rays but unknown endpoints), and view synthesis (unknown pixel values for known rays)." + } + ] + } + ], + "index": 15 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "text", + "content": "6324" + } + ] + } + ], + "index": 16 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 7 + }, + { + "para_blocks": [ + { + "bbox": [ + 55, + 72, + 294, + 119 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 72, + 294, + 119 + ], + "spans": [ + { + "bbox": [ + 55, + 72, + 294, + 119 + ], + "type": "text", + "content": "Acknowledgements: We thank the members of the Physical Perception Lab at CMU for their valuable discussions, and extend special thanks to Yanbo Xu for his insights on diffusion models." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 55, + 121, + 295, + 312 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 121, + 295, + 312 + ], + "spans": [ + { + "bbox": [ + 55, + 121, + 295, + 312 + ], + "type": "text", + "content": "This work used Bridges-2 [4] at Pittsburgh Supercomputing Center through allocation CIS240166 from the Advanced Cyberinfrastructure Coordination Ecosystem: Services & Support (ACCESS) program, which is supported by National Science Foundation grants #2138259, #2138286, #2138307, #2137603, and #2138296. This work was supported by Intelligence Advanced Research Projects Activity (IARPA) via Department of Interior/Interior Business Center (DOI/IBC) contract number 140D0423C0074. The U.S. Government is authorized to reproduce and distribute reprints for Governmental purposes notwithstanding any copyright annotation thereon. Disclaimer: The views and conclusions contained herein are those of the authors and should not be interpreted as necessarily representing the official policies or endorsements, either expressed or implied, of IARPA, DOI/IBC, or the U.S. Government." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 56, + 320, + 115, + 332 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 320, + 115, + 332 + ], + "spans": [ + { + "bbox": [ + 56, + 320, + 115, + 332 + ], + "type": "text", + "content": "References" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 61, + 340, + 295, + 713 + ], + "type": "list", + "angle": 0, + "index": 12, + "blocks": [ + { + "bbox": [ + 61, + 340, + 295, + 394 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 61, + 340, + 295, + 394 + ], + "spans": [ + { + "bbox": [ + 61, + 340, + 295, + 394 + ], + "type": "text", + "content": "[1] Gilad Baruch, Zhuoyuan Chen, Afshin Dehghan, Tal Dimry, Yuri Feigin, Peter Fu, Thomas Gebauer, Brandon Joffe, Daniel Kurz, Arik Schwartz, et al. Arkitsscenes: A diverse real-world dataset for 3d indoor scene understanding using mobile rgb-d data. In NeurIPS D&B, 2021. 5" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 61, + 396, + 294, + 426 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 61, + 396, + 294, + 426 + ], + "spans": [ + { + "bbox": [ + 61, + 396, + 294, + 426 + ], + "type": "text", + "content": "[2] Herbert Bay, Andreas Ess, Tinne Tuytelaars, and Luc Van Gool. Surf: Speeded up robust features. In ECCV, 2006. 2" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 62, + 428, + 295, + 482 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 428, + 295, + 482 + ], + "spans": [ + { + "bbox": [ + 62, + 428, + 295, + 482 + ], + "type": "text", + "content": "[3] Eric Brachmann, Jamie Wynn, Shuai Chen, Tommaso Cavallari, Aron Monszpart, Daniyar Turmukhambetov, and Victor Adrian Prisacariu. Scene coordinate reconstruction: posing of image collections via incremental learning of a relocalizer. In ECCV, 2024. 2" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 62, + 483, + 294, + 537 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 483, + 294, + 537 + ], + "spans": [ + { + "bbox": [ + 62, + 483, + 294, + 537 + ], + "type": "text", + "content": "[4] Shawn T Brown, Paola Buitrago, Edward Hanna, Sergiu Sanielevici, Robin Scibek, and Nicholas A Nystrom. Bridges-2: A platform for rapidly-evolving and data intensive research. In Practice and experience in advanced research computing, 2021. 9" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 62, + 538, + 294, + 581 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 538, + 294, + 581 + ], + "spans": [ + { + "bbox": [ + 62, + 538, + 294, + 581 + ], + "type": "text", + "content": "[5] Ruojin Cai, Joseph Tung, Qianqian Wang, Hadar Averbuch-Elor, Bharath Hariharan, and Noah Snavely. Doppelgangers: Learning to disambiguate images of similar structures. In ICCV, 2023. 1" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 62, + 582, + 294, + 613 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 582, + 294, + 613 + ], + "spans": [ + { + "bbox": [ + 62, + 582, + 294, + 613 + ], + "type": "text", + "content": "[6] Daniel DeTone, Tomasz Malisiewicz, and Andrew Rabinovich. Superpoint: Self-supervised interest point detection and description. In CVPR, 2018. 1, 2" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 62, + 615, + 294, + 658 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 615, + 294, + 658 + ], + "spans": [ + { + "bbox": [ + 62, + 615, + 294, + 658 + ], + "type": "text", + "content": "[7] Bardienus Duisterhof, Lojze Zust, Philippe Weinzaepfel, Vincent Leroy, Yohann Cabon, and Jerome Revaud. Mast3r-sfm: a fully-integrated solution for unconstrained structure-from-motion. In 3DV, 2025. 2" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 62, + 658, + 294, + 691 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 658, + 294, + 691 + ], + "spans": [ + { + "bbox": [ + 62, + 658, + 294, + 691 + ], + "type": "text", + "content": "[8] Johan Edstedt, Qiyu Sun, Georg Bokman, Marten Wadenback, and Michael Felsberg. RoMa: Robust Dense Feature Matching. In CVPR, 2024. 2, 8" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 62, + 692, + 294, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 692, + 294, + 713 + ], + "spans": [ + { + "bbox": [ + 62, + 692, + 294, + 713 + ], + "type": "text", + "content": "[9] Jonathan Ho, Ajay Jain, and Pieter Abbeel. Denoising diffusion probabilistic models. In NeurIPS, 2020. 2, 3" + } + ] + } + ], + "index": 11 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 316, + 73, + 553, + 713 + ], + "type": "list", + "angle": 0, + "index": 31, + "blocks": [ + { + "bbox": [ + 316, + 73, + 553, + 95 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 73, + 553, + 95 + ], + "spans": [ + { + "bbox": [ + 316, + 73, + 553, + 95 + ], + "type": "text", + "content": "[10] Vincent Leroy, Yohann Cabon, and Jérôme Revaud. Grounding image matching in 3d with mast3r, 2024. ECCV. 1, 2" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 317, + 96, + 553, + 128 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 96, + 553, + 128 + ], + "spans": [ + { + "bbox": [ + 317, + 96, + 553, + 128 + ], + "type": "text", + "content": "[11] Zhengqi Li and Noah Snavely. Megadepth: Learning single-view depth prediction from internet photos. In CVPR, 2018. 2,4,5,8" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 317, + 129, + 553, + 160 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 129, + 553, + 160 + ], + "spans": [ + { + "bbox": [ + 317, + 129, + 553, + 160 + ], + "type": "text", + "content": "[12] Amy Lin, Jason Y Zhang, Deva Ramanan, and Shubham Tulsiani. Relpose++: Recovering 6d poses from sparse-view observations. In 3DV, 2024. 2, 4, 6, 7" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 317, + 162, + 553, + 194 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 162, + 553, + 194 + ], + "spans": [ + { + "bbox": [ + 317, + 162, + 553, + 194 + ], + "type": "text", + "content": "[13] Philipp Lindenberger, Paul-Edouard Sarlin, Viktor Larsson, and Marc Pollefeys. Pixel-perfect structure-from-motion with featuremetric refinement. In ICCV, 2021. 2" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 316, + 195, + 553, + 227 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 195, + 553, + 227 + ], + "spans": [ + { + "bbox": [ + 316, + 195, + 553, + 227 + ], + "type": "text", + "content": "[14] Philipp Lindenberger, Paul-Edouard Sarlin, Viktor Larsson, and Marc Pollefeys. Pixel-Perfect Structure-from-Motion with Featuremetric Refinement. In ICCV, 2021. 1" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 316, + 228, + 553, + 259 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 228, + 553, + 259 + ], + "spans": [ + { + "bbox": [ + 316, + 228, + 553, + 259 + ], + "type": "text", + "content": "[15] Philipp Lindenberger, Paul-Edouard Sarlin, and Marc Pollefeys. Lightglue: Local feature matching at light speed. In ICCV, 2023. 2" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 316, + 261, + 553, + 282 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 261, + 553, + 282 + ], + "spans": [ + { + "bbox": [ + 316, + 261, + 553, + 282 + ], + "type": "text", + "content": "[16] David G Lowe. Distinctive image features from scale-invariant keypoints. In IJCV, 2004. 2" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 316, + 284, + 553, + 327 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 284, + 553, + 327 + ], + "spans": [ + { + "bbox": [ + 316, + 284, + 553, + 327 + ], + "type": "text", + "content": "[17] Nikolaus Mayer, Eddy Ilg, Philip Hausser, Philipp Fischer, Daniel Cremers, Alexey Dosovitskiy, and Thomas Brox. A large dataset to train convolutional networks for disparity, optical flow, and scene flow estimation. In CVPR, 2016. 5" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 316, + 328, + 553, + 381 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 328, + 553, + 381 + ], + "spans": [ + { + "bbox": [ + 316, + 328, + 553, + 381 + ], + "type": "text", + "content": "[18] Maxime Oquab, Timothee Darcet, Theo Moutakanni, Huy Vo, Marc Szafraniec, Vasil Khalidov, Pierre Fernandez, Daniel Haziza, Francisco Massa, Alaaeldin El-Nouby, et al. Dinov2: Learning robust visual features without supervision. In TMLR, 2024. 3, 4" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 316, + 383, + 553, + 403 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 383, + 553, + 403 + ], + "spans": [ + { + "bbox": [ + 316, + 383, + 553, + 403 + ], + "type": "text", + "content": "[19] William Peebles and Saining Xie. Scalable diffusion models with transformers. In ICCV, 2023. 2, 3, 4" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 316, + 404, + 553, + 436 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 404, + 553, + 436 + ], + "spans": [ + { + "bbox": [ + 316, + 404, + 553, + 436 + ], + "type": "text", + "content": "[20] René Ranftl, Alexey Bochkovskiy, and Vladlen Koltun. Vision transformers for dense prediction. In ICCV, 2021. 3, 4" + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 316, + 437, + 553, + 491 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 437, + 553, + 491 + ], + "spans": [ + { + "bbox": [ + 316, + 437, + 553, + 491 + ], + "type": "text", + "content": "[21] Jeremy Reizenstein, Roman Shapovalov, Philipp Henzler, Luca Sbordone, Patrick Labatut, and David Novotny. Common objects in 3d: Large-scale learning and evaluation of real-life 3d category reconstruction. In ICCV, 2021. 2, 4, 5, 7" + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 316, + 493, + 553, + 525 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 493, + 553, + 525 + ], + "spans": [ + { + "bbox": [ + 316, + 493, + 553, + 525 + ], + "type": "text", + "content": "[22] Paul-Edouard Sarlin, Cesar Cadena, Roland Siegwart, and Marcin Dymczyk. From coarse to fine: Robust hierarchical localization at large scale. In CVPR, 2019. 6, 7" + } + ] + } + ], + "index": 25 + }, + { + "bbox": [ + 317, + 526, + 553, + 568 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 526, + 553, + 568 + ], + "spans": [ + { + "bbox": [ + 317, + 526, + 553, + 568 + ], + "type": "text", + "content": "[23] Paul-Edouard Sarlin, Daniel DeTone, Tomasz Malisiewicz, and Andrew Rabinovich. Superglue: Learning feature matching with graph neural networks. In CVPR, 2020. 1, 2" + } + ] + } + ], + "index": 26 + }, + { + "bbox": [ + 317, + 570, + 553, + 613 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 570, + 553, + 613 + ], + "spans": [ + { + "bbox": [ + 317, + 570, + 553, + 613 + ], + "type": "text", + "content": "[24] Manolis Savva, Abhishek Kadian, Oleksandr Maksymets, Yili Zhao, Erik Wijmans, Bhavana Jain, Julian Straub, Jia Liu, Vladlen Koltun, Jitendra Malik, et al. Habitat: A platform for embodied ai research. In ICCV, 2019. 2, 5, 7" + } + ] + } + ], + "index": 27 + }, + { + "bbox": [ + 316, + 615, + 553, + 635 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 615, + 553, + 635 + ], + "spans": [ + { + "bbox": [ + 316, + 615, + 553, + 635 + ], + "type": "text", + "content": "[25] Johannes L Schonberger and Jan-Michael Frahm. Structure-from-motion revisited. In CVPR, 2016. 1, 2" + } + ] + } + ], + "index": 28 + }, + { + "bbox": [ + 317, + 636, + 553, + 678 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 636, + 553, + 678 + ], + "spans": [ + { + "bbox": [ + 317, + 636, + 553, + 678 + ], + "type": "text", + "content": "[26] Samarth Sinha, Jason Y Zhang, Andrea Tagliasacchi, Igor Gilitschenski, and David B Lindell. Sparsepose: Sparse-view camera pose regression and refinement. In CVPR, 2023. 2" + } + ] + } + ], + "index": 29 + }, + { + "bbox": [ + 316, + 681, + 553, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 681, + 553, + 713 + ], + "spans": [ + { + "bbox": [ + 316, + 681, + 553, + 713 + ], + "type": "text", + "content": "[27] Cameron Smith, David Charatan, Ayush Tewari, and Vincent Sitzmann. Flowmap: High-quality camera poses, intrinsics, and depth via gradient descent. In 3DV, 2025. 2" + } + ] + } + ], + "index": 30 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "text", + "content": "6325" + } + ] + } + ], + "index": 32 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 8 + }, + { + "para_blocks": [ + { + "bbox": [ + 56, + 73, + 295, + 645 + ], + "type": "list", + "angle": 0, + "index": 15, + "blocks": [ + { + "bbox": [ + 56, + 73, + 294, + 106 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 73, + 294, + 106 + ], + "spans": [ + { + "bbox": [ + 56, + 73, + 294, + 106 + ], + "type": "text", + "content": "[28] Noah Snavely, Steven M Seitz, and Richard Szeliski. Photo tourism: exploring photo collections in 3d. In ACM SIGGRAPH, 2006. 2" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 56, + 107, + 295, + 129 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 107, + 295, + 129 + ], + "spans": [ + { + "bbox": [ + 56, + 107, + 295, + 129 + ], + "type": "text", + "content": "[29] Jiaming Song, Chenlin Meng, and Stefano Ermon. Denoising diffusion implicit models. In ICLR, 2021. 3" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 56, + 130, + 294, + 162 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 130, + 294, + 162 + ], + "spans": [ + { + "bbox": [ + 56, + 130, + 294, + 162 + ], + "type": "text", + "content": "[30] Jiaming Sun, Zehong Shen, Yuang Wang, Hujun Bao, and Xiaowei Zhou. Loftr: Detector-free local feature matching with transformers. In CVPR, 2021. 2" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 56, + 164, + 294, + 217 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 164, + 294, + 217 + ], + "spans": [ + { + "bbox": [ + 56, + 164, + 294, + 217 + ], + "type": "text", + "content": "[31] Pei Sun, Henrik Kretzschmar, Xerxes Dotiwalla, Aurelien Chouard, Vijaysai Patnaik, Paul Tsui, James Guo, Yin Zhou, Yuning Chai, Benjamin Caine, et al. Scalability in perception for autonomous driving: Waymo open dataset. In CVPR, 2020. 5" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 56, + 219, + 294, + 241 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 219, + 294, + 241 + ], + "spans": [ + { + "bbox": [ + 56, + 219, + 294, + 241 + ], + "type": "text", + "content": "[32] Chengzhou Tang and Ping Tan. Ba-net: Dense bundle adjustment network. In ICLR, 2019. 2" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 56, + 243, + 294, + 275 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 243, + 294, + 275 + ], + "spans": [ + { + "bbox": [ + 56, + 243, + 294, + 275 + ], + "type": "text", + "content": "[33] Ashish Vaswani, Noam Shazeer, Niki Parmar, Jakob Uszkoreit, Llion Jones, Aidan N Gomez, Lukasz Kaiser, and Illia Polosukhin. Attention is all you need. In NeurIPS, 2017. 4" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 56, + 277, + 294, + 308 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 277, + 294, + 308 + ], + "spans": [ + { + "bbox": [ + 56, + 277, + 294, + 308 + ], + "type": "text", + "content": "[34] Jianyuan Wang, Christian Rupprecht, and David Novotny. Posediffusion: Solving pose estimation via diffusion-aided bundle adjustment. In ICCV, 2023. 2, 6, 7" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 56, + 310, + 294, + 363 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 310, + 294, + 363 + ], + "spans": [ + { + "bbox": [ + 56, + 310, + 294, + 363 + ], + "type": "text", + "content": "[35] Ruicheng Wang, Sicheng Xu, Cassie Dai, Jianfeng Xiang, Yu Deng, Xin Tong, and Jiaolong Yang. Moge: Unlocking accurate monocular geometry estimation for open-domain images with optimal training supervision. In CVPR, 2025. 6" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 56, + 366, + 294, + 399 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 366, + 294, + 399 + ], + "spans": [ + { + "bbox": [ + 56, + 366, + 294, + 399 + ], + "type": "text", + "content": "[36] Shuzhe Wang, Vincent Leroy, Yohann Cabon, Boris Chidlovskii, and Jerome Revaud. Dust3r: Geometric 3d vision made easy. In CVPR, 2024. 1, 2, 4, 5, 6, 7" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 56, + 400, + 294, + 442 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 400, + 294, + 442 + ], + "spans": [ + { + "bbox": [ + 56, + 400, + 294, + 442 + ], + "type": "text", + "content": "[37] Shuzhe Wang, Vincent Leroy, Yohann Cabon, Boris Chidlovskii, and Jerome Revaud. Dust3r: Geometric 3d vision made easy, 2024. https://github.com/naver/dust3r. 5, 6" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 56, + 445, + 294, + 487 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 445, + 294, + 487 + ], + "spans": [ + { + "bbox": [ + 56, + 445, + 294, + 487 + ], + "type": "text", + "content": "[38] Yao Yao, Zixin Luo, Shiwei Li, Jingyang Zhang, Yufan Ren, Lei Zhou, Tian Fang, and Long Quan. Blendedmvs: A large-scale dataset for generalized multi-view stereo networks. In CVPR, 2020. 5" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 56, + 490, + 294, + 521 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 490, + 294, + 521 + ], + "spans": [ + { + "bbox": [ + 56, + 490, + 294, + 521 + ], + "type": "text", + "content": "[39] Chandan Yeshwanth, Yueh-Cheng Liu, Matthias Nießner, and Angela Dai. Scannet++: A high-fidelity dataset of 3d indoor scenes. In ICCV, 2023. 5" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 56, + 524, + 294, + 555 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 524, + 294, + 555 + ], + "spans": [ + { + "bbox": [ + 56, + 524, + 294, + 555 + ], + "type": "text", + "content": "[40] Jason Y Zhang, Deva Ramanan, and Shubham Tulsiani. Relapse: Predicting probabilistic relative rotation for single objects in the wild. In ECCV, 2022. 2" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 56, + 557, + 294, + 601 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 557, + 294, + 601 + ], + "spans": [ + { + "bbox": [ + 56, + 557, + 294, + 601 + ], + "type": "text", + "content": "[41] Jason Y. Zhang, Amy Lin, Moneish Kumar, Tzu-Hsuan Yang, Deva Ramanan, and Shubham Tulsiani. Cameras as rays: Sparse-view pose estimation via ray diffusion. In ICLR, 2024. 2, 3, 4, 6, 7" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 56, + 602, + 294, + 645 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 602, + 294, + 645 + ], + "spans": [ + { + "bbox": [ + 56, + 602, + 294, + 645 + ], + "type": "text", + "content": "[42] Tinghui Zhou, Richard Tucker, John Flynn, Graham Fyffe, and Noah Snavely. Stereo magnification: Learning view synthesis using multiplane images. In ACM SIGGRAPH, 2018. 2, 6, 7" + } + ] + } + ], + "index": 14 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 749, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 749, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 749, + 315, + 757 + ], + "type": "text", + "content": "6326" + } + ] + } + ], + "index": 16 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 9 + } + ], + "_backend": "vlm", + "_version_name": "2.6.4" +} \ No newline at end of file diff --git a/2025/Digital Twin Catalog_ A Large-Scale Photorealistic 3D Object Digital Twin Dataset/2d29b0d5-5c08-413f-9ad0-63d98b0c6539_content_list.json b/2025/Digital Twin Catalog_ A Large-Scale Photorealistic 3D Object Digital Twin Dataset/2d29b0d5-5c08-413f-9ad0-63d98b0c6539_content_list.json new file mode 100644 index 0000000000000000000000000000000000000000..b17f0bf5a91ff296490fde21222dd9e5bd0a80bc --- /dev/null +++ b/2025/Digital Twin Catalog_ A Large-Scale Photorealistic 3D Object Digital Twin Dataset/2d29b0d5-5c08-413f-9ad0-63d98b0c6539_content_list.json @@ -0,0 +1,1553 @@ +[ + { + "type": "text", + "text": "Digital Twin Catalog: A Large-Scale Photorealistic 3D Object Digital Twin Dataset", + "text_level": 1, + "bbox": [ + 91, + 130, + 906, + 152 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Zhao Dong $^{1}$ Ka Chen $^{1,*}$ Zhaoyang Lv $^{1,*}$ Hong-Xing Yu $^{2,*}$ Yunzhi Zhang $^{2,*}$ Cheng Zhang $^{1,*}$ Yufeng Zhu $^{1,*}$ Stephen Tian $^{2}$ Zhengqin Li $^{1}$ Georgie Moffatt $^{1}$ Sean Christofferson $^{1}$ James Fort $^{1}$ Xiaqing Pan $^{1}$ Mingfei Yan $^{1}$ Jiajun Wu $^{2}$ Carl Yuheng Ren $^{1}$ Richard Newcombe $^{1}$", + "bbox": [ + 101, + 179, + 893, + 234 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "$^{1}$ Meta Reality Labs Research Stanford University", + "bbox": [ + 287, + 243, + 709, + 263 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/5c6ef2dab8e54265416509670d0062d9671cdfe6f8dc0a4a7cec32c62d6f5d97.jpg", + "image_caption": [ + "(a) Examples of 2,000 digital twins of real-world objects" + ], + "image_footnote": [], + "bbox": [ + 98, + 310, + 433, + 570 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/63ea79faa2c911314cac9024de6bb70ec75275098cc717280376a3c205019268.jpg", + "image_caption": [ + "(b) Shape & PBR Materials" + ], + "image_footnote": [], + "bbox": [ + 462, + 308, + 602, + 570 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/7d21347e865b975546c4908b004218877b7eed45912281de5bc1d261d7ec61c1.jpg", + "image_caption": [ + "(c) DSLR Data", + "Figure 1. The Digital Twin Catalog (DTC) dataset comprises 2,000 digital twins of physical-world objects (a), characterized by millimeter-level geometric accuracy and photorealistic PBR materials (b). DTC includes evaluation data captured using both DSLR cameras and egocentric Aria glasses, featuring captured images with precise foreground object masks and environment lighting for relighting evaluation." + ], + "image_footnote": [], + "bbox": [ + 638, + 297, + 759, + 570 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/8ddbce2402e3d85449771a91d3f6fe0241663684d598440c2308bf4ff77860b6.jpg", + "image_caption": [ + "(d) Egocentric Data" + ], + "image_footnote": [], + "bbox": [ + 774, + 309, + 906, + 566 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Abstract", + "text_level": 1, + "bbox": [ + 246, + 652, + 326, + 667 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "We introduce Digital Twin Catalog (DTC), a new large-scale photorealistic 3D object digital twin dataset. A digital twin of a 3D object is a highly detailed, virtually indistinguishable representation of a physical object, accurately capturing its shape, appearance, physical properties, and other attributes. Recent advances in neural-based 3D reconstruction and inverse rendering have significantly improved the quality of 3D object reconstruction. Despite these advancements, there remains a lack of a large-scale, digital twin quality real-world dataset and benchmark that can quantitatively assess and compare the performance of different reconstruction methods, as well as improve reconstruction quality through", + "bbox": [ + 88, + 686, + 485, + 869 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "training or fine-tuning. Moreover, to democratize 3D digital twin creation, it is essential to integrate creation techniques with next-generation egocentric computing platforms, such as AR glasses. Currently, there is no dataset available to evaluate 3D object reconstruction using egocentric captured images. To address these gaps, the DTC dataset features 2,000 scanned digital twin-quality 3D objects, along with image sequences captured under different lighting conditions using DSLR cameras and egocentric AR glasses. This dataset establishes the first comprehensive real-world evaluation benchmark for 3D digital twin creation tasks, offering a robust foundation for comparing and improving existing reconstruction methods. The DTC dataset is already released at https://www.projectaria.com/datasets/dtc/ and we will also make the baseline evaluations open-source.", + "bbox": [ + 511, + 654, + 908, + 881 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "CVF", + "bbox": [ + 106, + 2, + 181, + 42 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore.", + "bbox": [ + 236, + 0, + 810, + 44 + ], + "page_idx": 0 + }, + { + "type": "page_footnote", + "text": "*Authors contributed equally and are listed in alphabetical order.", + "bbox": [ + 114, + 886, + 457, + 898 + ], + "page_idx": 0 + }, + { + "type": "page_number", + "text": "753", + "bbox": [ + 485, + 945, + 509, + 955 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/787532f76f3c7d05c010f997c86ce2161ca34f8cd6a082de3e89196c9b445026.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 91, + 88, + 251, + 213 + ], + "page_idx": 1 + }, + { + "type": "image", + "img_path": "images/7ac0bd5a69b106213fb119b8eb13f0f3640675ae19d597720bfa6b4fffd7aae1.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 254, + 88, + 415, + 213 + ], + "page_idx": 1 + }, + { + "type": "image", + "img_path": "images/03814b73dbaeeceaca5b864b789e7d1bd77546a4fce1c8fba05868b819ffbd6a.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 416, + 88, + 578, + 213 + ], + "page_idx": 1 + }, + { + "type": "image", + "img_path": "images/162bee86384d06d222c9ba1a0dd1d208552d44bf41cf0d1b69620172067937cf.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 580, + 88, + 741, + 213 + ], + "page_idx": 1 + }, + { + "type": "image", + "img_path": "images/4b26a4302ba3923f69965b63cf545b83d58552ac9cb9e125720cd32dd3681f09.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 741, + 89, + 903, + 213 + ], + "page_idx": 1 + }, + { + "type": "image", + "img_path": "images/c1f9ec98603ce255dfa5b73ba916b26a6b329bd7b9fb068f4a227c67879135ff.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 91, + 215, + 251, + 339 + ], + "page_idx": 1 + }, + { + "type": "image", + "img_path": "images/c3f63b0fd71a76659fed8016ddb3d7ca9f3f5354904b2712fcdf92e3b6356eae.jpg", + "image_caption": [ + "Figure 2. Example DTC models with photorealistic PBR materials." + ], + "image_footnote": [], + "bbox": [ + 254, + 215, + 415, + 339 + ], + "page_idx": 1 + }, + { + "type": "image", + "img_path": "images/c742133f79593415e6754d2c6e6b65a5b75e0dc5457fb9ba9d20bff3338b32a0.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 416, + 215, + 578, + 339 + ], + "page_idx": 1 + }, + { + "type": "image", + "img_path": "images/6d2d671251186c923beaff28ce7b2a88d1983ab8b524b097b52ad10d187569e7.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 580, + 215, + 741, + 339 + ], + "page_idx": 1 + }, + { + "type": "image", + "img_path": "images/33f163d2d32987871165fce759381ea7ecfb7b5057c30b7527a1a50860194a9f.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 741, + 215, + 903, + 339 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "1. Introduction", + "text_level": 1, + "bbox": [ + 91, + 388, + 222, + 404 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "A digital twin of a 3D object is a highly detailed, virtually indistinguishable representation of a physical object, capturing its shape, appearance, physical properties and other attributes with precision. Such a digital twin enables visualization, analysis, and interaction as if it were the real object, supporting simulation, automation, and real-world problem-solving across a wide range of applications in AR/VR [3, 28], spatial/contextual AI [2], and robotics [23, 43]. As fundamental properties of an object, its shape and appearance form the basis for recognizing and interpreting the 3D object, enabling identification, manipulation, and realistic rendering. Recovering these attributes has long been a foundational topic in computer vision and graphics, inspiring extensive research in 3D reconstruction and inverse rendering. Recent breakthroughs in neural-based representation and reconstruction techniques, such as NeRF [52] and 3D Gaussian splattig (3DGS) [37], have significantly elevated the quality of novel view synthesis (NVS) to photorealistic levels. Many subsequent works [35, 62] integrate neural reconstruction with physically-based inverse rendering, enabling relightable appearances. Furthermore, leveraging priors from large reconstruction models (LRMs) [30], high-quality shape and appearance reconstruction can now be achieved with as few as one to four views [41, 69, 82].", + "bbox": [ + 89, + 415, + 485, + 777 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Despite the rapid advancements in 3D object reconstruction, one question remains: does the reconstruction quality truly meet the standard of a digital twin, where virtual representations are indistinguishable from reality? This digital twin standard demands both highly accurate shape matching and photorealistic appearance across different lighting, which present significant acquisition challenges for real-world objects. Existing object-centric datasets for 3D recon", + "bbox": [ + 89, + 780, + 485, + 901 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "struction or inverse rendering have focused on either dataset size [18] or quality of specific aspects [16, 21, 34, 46, 61, 73], often sacrificing comprehensive fidelity and limiting their application scope. This trade-off has led to a lack of datasets that fully satisfy the digital twin criteria, hindering current 3D reconstruction methods from achieving digital twin fidelity. To bridge this gap, we developed the Digital Twin Catalog (DTC) dataset, comprising 2,000 scanned 3D object models (Fig. 1(a)), each with millimeter geometry accuracy and photorealistic PBR materials (Fig. 1(b), Fig. 2).", + "bbox": [ + 511, + 390, + 908, + 542 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "In addition to 3D digital twin models, the DTC dataset includes evaluation data designed to support 3D object reconstruction research. This evaluation data features multiview image sequences with precise foreground object masks and environment lighting information for relighting evaluation. Traditionally, high-quality HDR images captured with modern DSLR cameras have been the standard for 3D reconstruction research. Looking ahead, we encourage the integration of 3D reconstruction research with next-generation human-centric computing platforms, such as egocentric AR glasses, aiming to democratize 3D reconstruction techniques and empower everyone to effortlessly create 3D digital twins. To this end, alongside DSLR-captured evaluation data (Fig. 1(c)), the DTC dataset also provides egocentric evaluation data captured using Project Aria glasses (https://www.projectaria.com) (Fig. 1(d)).", + "bbox": [ + 511, + 547, + 910, + 790 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "The DTC dataset offers extensive opportunities for advancing research in object digital twin creation. We provide a benchmark for state-of-the-art 3D object reconstruction and inverse rendering methods. These benchmarks evaluate performance across novel view synthesis (NVS), shape reconstruction, and relightable appearance reconstruction. We further provide the evaluation of novel view synthesis meth", + "bbox": [ + 511, + 794, + 911, + 902 + ], + "page_idx": 1 + }, + { + "type": "page_number", + "text": "754", + "bbox": [ + 485, + 944, + 509, + 955 + ], + "page_idx": 1 + }, + { + "type": "table", + "img_path": "images/bb306c32194b753c9ba465301f050069d8d24047d747150c927252e682122167.jpg", + "table_caption": [ + "Table 1. Comparison with existing object-centric inverse rendering datasets. *Objaverse [18] consists of both synthetic objects and real scans, only part of which contain PBR materials." + ], + "table_footnote": [], + "table_body": "
Dataset# ObjectsRealScene TypeMulti-viewShapePBR Mat.Relit ImageLightingEgocentric Cap.
ShapeNet-Intrinsics [61]31KXsynthetic✓✓X✓✓X
NeRD Synthetic [11]3Xsynthetic✓✓✓✓✓X
ABO [16]8KXsynthetic✓✓✓✓✓X
MIT Intrasins [25]20✓studio✓XXXXX
DTU-MVS [34]80✓studio✓✓XXXX
Objaverse [18]818K(✓)*studio✓✓(✓)*XXX
DiLiGenT-MV [42]5✓studio✓✓XX✓X
ReNe [65]20✓studio✓XXX✓X
OpenIllumination [46]64✓studio✓X✓✓✓X
GSO [21]1030✓studio✓✓XXXX
Lombardi et al. [47]6✓in-the-wildX✓X✓✓X
NeRD Real [11]4✓in-the-wild✓X✓✓XX
NeROIC [38]3✓in-the-wild✓X✓✓XX
Oxholm et al. [56]4✓in-the-wild✓✓X✓✓X
OmniObject3D [73]6k✓in-the-wild✓✓XXXX
Stanford Orb [39]14✓in-the-wild✓✓✓✓✓X
DTC (ours)2k✓in-the-wild✓✓✓✓✓✓
", + "bbox": [ + 94, + 128, + 901, + 369 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "ods using the egocentric aligned DTC data. Additionally, we explore the dataset's potential in downstream robotics applications by assessing its effectiveness in training robotic policies for pushing and grasping tasks in simulation. These benchmarks and applications provide valuable insights, highlight existing challenges, and uncover promising directions for future research in 3D digital twin creation.", + "bbox": [ + 88, + 396, + 482, + 502 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "2. Related Work", + "text_level": 1, + "bbox": [ + 89, + 521, + 232, + 536 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "We provide a comparison of our DTC dataset to existing object-centric datasets in Table 1. We provide the largest 3D dataset with PBR materials and real world multi-view recordings with digital twin counterparts. We further provide digital twin aligned egocentric recordings, the first of their kind in the egocentric domain. We will discuss the related datasets and methods they can empower as follows.", + "bbox": [ + 88, + 547, + 482, + 654 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3D Digital Twin Datasets Existing 3D digital twin datasets with PBR materials often serve as ground truth for evaluating inverse rendering results. Early efforts [6, 25] provide small-scale intrinsic image of real objects and do not provide shape or PBR material information. Synthetic datasets [11, 16, 45, 61, 70, 72] are widely used for evaluation but do not represent the complexity in a real world environment. For datasets that contain real objects, [18, 34, 42, 47, 56, 65], the reconstruction quality can vary, which leads to imprecise evaluations. Table 1 provides a comparison to the previous work in this domain. Compared to Objaverse[18, 19], which is a collection of existing 3D models with only a small subset containing PBR materials with varying quality, we offer a high quality collection of 3D object data that is also aligned with real world recordings. Compared to OmniObject3D [73], the DTC models", + "bbox": [ + 88, + 657, + 482, + 900 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "provide higher-quality shape and additional PBR materials that are necessary for high quality inverse rendering. We offer the largest quantity of 3D object models compared to all counterparts in various tasks. For real-world evaluation, Stanford-ORB [39] was the prior largest inverse rendering benchmark with in-the-wild lighting. In contrast, we provide more object diversity and higher quality for each object model. The Aria Digital Twin dataset [58] was the first dataset to provide digital twin aligned environments for the scenes and recorded using egocentric device. However, their scene environments are limited and the contained object ground truths inside do not have high quality geometries with PBR materials.", + "bbox": [ + 511, + 396, + 906, + 592 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Object Reconstruction & Inverse Rendering. Using object-centric multi-view images as input, early object reconstruction methods focused on estimating individual object properties, such as shape from shading [5, 9, 83], material acquisition [44, 45, 56, 57, 74], and lighting estimation [67, 78]. Some approaches also aimed to recover reflectance and illumination assuming known object shapes [47, 48]. Inverse rendering, which seeks to invert the rendering equation [36], estimates an image's intrinsic components—geometry, materials, and lighting. The advent of differentiable renderers [14, 15, 49] enabled full-fledged inverse rendering methods to simultaneously recover all these properties for object reconstruction [50].", + "bbox": [ + 511, + 597, + 908, + 792 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Neural volumetric representations such as Neural Radiance Fields (NeRFs) [52] and the like [7, 11-13, 59, 79, 84] encode geometry and appearance as volumetric densities and radiance with a Multi-Layer Perceptron (MLP) network, and render images using the volume rendering equation [51]. 3D-GS [37] introduces 3D Gaussian primitives and rasterization and its following-up variants [31] demonstrates high quality", + "bbox": [ + 511, + 795, + 908, + 902 + ], + "page_idx": 2 + }, + { + "type": "page_number", + "text": "755", + "bbox": [ + 485, + 944, + 509, + 955 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "geometry prediction as well.", + "bbox": [ + 89, + 90, + 279, + 104 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Other surface-based representations [29, 53, 62, 66, 71, 80, 81, 85] extract surfaces as the zero level set, for instance, of a signed distance function (SDF) or an occupancy field [55], allowing them to efficiently model the appearance on the surface with an explicit material model, such as bidirectional reflectance distribution functions (BRDFs). This also enables modeling more complex global illumination effects, such as self-shadows. Most of these methods focus on per-scene optimization and require dense multiple views as input. Recently, researchers have incorporated learning-based models, distilling priors from large training datasets for fast inference on limited test views [8, 10, 33, 45, 61, 70, 72, 86].", + "bbox": [ + 89, + 106, + 483, + 303 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "In this work, we provide the DTC dataset with real-world object recordings that can serve as the benchmark to evaluate object-centric inverse rendering tasks. We evaluate representative baselines from existing work.", + "bbox": [ + 89, + 303, + 483, + 364 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "3. Digital Twin Catalog: A Large-Scale Photorealistic 3D Object Digital Twin Dataset", + "text_level": 1, + "bbox": [ + 89, + 378, + 485, + 415 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "3.1. Dataset Composition", + "text_level": 1, + "bbox": [ + 89, + 422, + 287, + 439 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Our DTC dataset contains: (1) 2,000 scanned 3D object models, featuring millimeter geometric accuracy relative to their physical counterparts, along with a full set of photorealistic PBR material maps (Fig. 1), (2) 100 DSLR-captured evaluation data of 50 objects under different lighting conditions, and (3) 200 egocentric Aria-captured evaluation data of 100 objects with both active and casual observation modes.", + "bbox": [ + 89, + 445, + 483, + 551 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "3.2. Creation of 3D Object Models", + "text_level": 1, + "bbox": [ + 89, + 560, + 356, + 575 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Utilizing the state-of-the-art industrial 3D object scanner [1], we selected 2,000 physical-world objects spanning 40 LVIS [27] categories, carefully chosen to ensure both category diversity and compatibility with the scanner's capabilities. As illustrated in Fig. 3(a) (c), the scanner features a fixed lighting-camera setup within an upper-hemisphere dome, equipped with 8 structured lights for geometry scanning, and 29 spotlights and 29 cameras for material acquisition. During the scanning process, the object is placed on a central holder, and its pose can be adjusted with multi-round scanning to achieve a complete 360-degree scan. For our dataset, each object typically undergoes three pose changes, with a total scanning time of approximately 20 minutes per object.", + "bbox": [ + 89, + 582, + 483, + 794 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "After scanning, a proprietary post-processing pipeline reconstructs both the geometry and the PBR material maps. For 4K-resolution PBR material maps, the post-processing requires approximately 4 hours per object. In terms of quality, the structured-light-based shape reconstruction (Fig. 3 (b)) in the post-processing achieves millimeter-level geometric accuracy. However, the material optimization process", + "bbox": [ + 89, + 795, + 483, + 901 + ], + "page_idx": 3 + }, + { + "type": "image", + "img_path": "images/514e435cb5ecf8ec602ac85b2b43b8a3f2c3a2d8ef0c882b2ac2ddd7d287b30e.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 517, + 89, + 710, + 236 + ], + "page_idx": 3 + }, + { + "type": "image", + "img_path": "images/135bc7d0496f2c9a75be1831c6bd023e92a6f80918f7d1b09193a3d349de2602.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 517, + 236, + 710, + 285 + ], + "page_idx": 3 + }, + { + "type": "image", + "img_path": "images/1e2df084aba61005a1fff73ede9c775b1492db5fa33bf103bb57e655c46b4ffc.jpg", + "image_caption": [ + "Figure 3. 3D object scanner by Covision Media®." + ], + "image_footnote": [], + "bbox": [ + 517, + 285, + 771, + 404 + ], + "page_idx": 3 + }, + { + "type": "image", + "img_path": "images/2221885c36272e45c04b1bb11f68fe0926c7cc4abddbd05d5b0cb6a6dc0834fc.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 710, + 90, + 901, + 237 + ], + "page_idx": 3 + }, + { + "type": "image", + "img_path": "images/b6c70bcdf78249b5ac44914e03c7c6a9ce38c56777016c4c60a6be8291f00057.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 710, + 236, + 901, + 285 + ], + "page_idx": 3 + }, + { + "type": "image", + "img_path": "images/81e1f60c73757d07d21564f4f92073b2c070847f5d4120db339f1d4b88771dbc.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 517, + 444, + 705, + 609 + ], + "page_idx": 3 + }, + { + "type": "image", + "img_path": "images/c7579deec3fdec506ecd75bef2a02066e26fe418be49935994c28f186ed5f5bc.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 707, + 444, + 903, + 609 + ], + "page_idx": 3 + }, + { + "type": "image", + "img_path": "images/ea01e0790fab617bd0e4f3f993dbf25516efb0989fd13c4c1b2cf25eb8b087e6.jpg", + "image_caption": [ + "Figure 4. Rendered DTC models (left) v.s. Photo (Right)." + ], + "image_footnote": [], + "bbox": [ + 517, + 611, + 705, + 772 + ], + "page_idx": 3 + }, + { + "type": "image", + "img_path": "images/2dde835d30df0f67fe2c497f736b8b0b1a2f89ba0109ae86e7edd6a6d6160cc2.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 705, + 611, + 903, + 772 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "performs best for diffuse objects and often struggles with glossy or shiny surfaces. To address this limitation, we hired technical artists to develop a workflow to refine materials for glossy and shiny objects, ensuring that the material quality meets the standards of a digital twin.", + "bbox": [ + 511, + 825, + 905, + 900 + ], + "page_idx": 3 + }, + { + "type": "page_number", + "text": "756", + "bbox": [ + 486, + 945, + 509, + 955 + ], + "page_idx": 3 + }, + { + "type": "image", + "img_path": "images/788b6a091b4172786c2a711155a41ca86b1db107b1def661f06eda6b8ac10fa2.jpg", + "image_caption": [ + "Figure 5. Shape and material (albedo) quality comparison between Stanford-ORB [39] (left) and our DTC (right)." + ], + "image_footnote": [], + "bbox": [ + 93, + 87, + 482, + 224 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "3D Model Accuracy. To validate the material and geometry accuracy of the 3D models, we compared a rendered image of our scanned and processed model with a photograph of the same object taken inside a light box. A virtual light box was meticulously modeled to replicate the light intensity and color temperature of the real light box. The scanned object was then placed in the virtual light box to generate the rendered image. The side-by-side comparison demonstrates a remarkable match between the rendered and real images (Fig. 4).", + "bbox": [ + 88, + 291, + 482, + 441 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Comparison Against Stanford-ORB. We also scanned the objects used in Stanford-ORB [39] to compare the shape and appearance quality. As illustrated in Fig. 5, the Stanford-ORB models exhibit shape artifacts and noisy, lower-quality materials compared to our models.", + "bbox": [ + 89, + 446, + 483, + 521 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "3.3. DSLR Evaluation Data", + "text_level": 1, + "bbox": [ + 89, + 526, + 305, + 541 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Within DTC, we include a DSLR-captured evaluation dataset of 50 objects from Sec. 3.2 captured under two different lighting conditions, resulting in 100 distinct image sequences. For every sequence, we provide (a) approximately 120 HDR and LDR images from different viewing directions, (b) one object pose and (c) per-image camera pose. The two lighting conditions are represented using two environment maps.", + "bbox": [ + 89, + 547, + 482, + 655 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Data Capture. To ensure the DSLR evaluation data quality, we designed and built a DSLR camera rig to automate the capture process (Fig. 6). The rig is designed to rotate the cameras around the centralized object, assuming the environment lighting remains unchanged during the capture. It features a motorized rotary stage with a centrally mounted stationary platform. Attached to the rotary stage is an extrusion frame that forms the gantry arm, supported by a set of castors to bear its weight and enable smooth rotation around the central axis. The extrusion frame is equipped with adjustable camera mounts, allowing DSLR cameras to be positioned flexibly to optimize the capture setup. For our capture process, we utilized three DSLR cameras to perform a 360-degree rotation around the object, capturing images at 9-degree intervals, resulting in 120 photos per object. To ensure precise camera pose estimation, a ChArUco board", + "bbox": [ + 89, + 659, + 483, + 900 + ], + "page_idx": 4 + }, + { + "type": "image", + "img_path": "images/1a6ce4a894ebbfcfb6ec2630ff256afe74443ffa2f522328986bb4e235917f94.jpg", + "image_caption": [ + "Figure 6. DSLR rig for capturing evaluation data." + ], + "image_footnote": [], + "bbox": [ + 516, + 87, + 905, + 233 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "was placed beneath the object during the capture. Example images from this setup are shown in Fig. 6.", + "bbox": [ + 511, + 285, + 906, + 313 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Environment Maps. Following a similar approach in Stanford-ORB [39], we capture the two environment maps using chrome ball images obtained with the capture rig described earlier. With precise camera poses provided by a ChArUco board placed beneath the chrome ball, we first fit a synthetic 3D sphere to the chrome ball by optimizing its 3D position using a geometry-friendly differentiable renderer [40, 75]. Subsequently, using a differentiable Monte Carlo-based renderer [32], we refine the environment map to match the reflection on the chrome ball, employing the single-view light estimation method proposed in [78]. The coordinate system of the environment map is determined by the ChArUco detection.", + "bbox": [ + 511, + 313, + 906, + 507 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Pose Registration for Camera and Object. We first obtain the initial camera poses using the ChArUco board. In most cases, for images captured from the top and middle views, the pose estimates are typically accurate. However, for bottom-view images, inaccuracies arise due to pattern distortion at grazing angles. To mitigate this issue, we refine the camera poses by fitting the rendering of a virtual ChArUco board to the real captured photo using a differentiable renderer [32, 40, 75]. Once the camera poses have been accurately refined, we optimize the object pose by minimizing the mask loss between the rendered mask and the reference one generated by [62]. By providing separate poses for camera and object anchored by ChArUco, the cameras and object are automatically aligned with the optimized environment maps.", + "bbox": [ + 511, + 512, + 908, + 739 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "3.4. Egocentric Evaluation Data", + "text_level": 1, + "bbox": [ + 511, + 750, + 761, + 763 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "We include an evaluation dataset of real world recordings captured by an egocentric device paired with 100 objects from the DTC dataset. We capture the egocentric recordings using the open-sourced Project Aria device [22] and acquire the additional 3D ground truth using its machine perception tool, which includes online device calibration, device trajectory and semi-dense point clouds for each recording. Each recording contains a single 3D object and precisely aligned object poses in the Aria trajectory coordinate frame.", + "bbox": [ + 511, + 763, + 908, + 900 + ], + "page_idx": 4 + }, + { + "type": "page_number", + "text": "757", + "bbox": [ + 486, + 945, + 509, + 955 + ], + "page_idx": 4 + }, + { + "type": "image", + "img_path": "images/dcd62f9fc15c847d5a6c0ee8cc41a039480a020f8dc5ae093f19a24eaae0f820.jpg", + "image_caption": [ + "Figure 7. The workflow to align egocentric video and 3D objects. We acquire the object's semi-dense point cloud from the egocentric recording and the neural reconstructed mesh using Neural-PBIR [62]. Finally, we align the 3D object with the rendered mask from the neural reconstructed mesh using physics-based differentiable rendering (PBDR)." + ], + "image_footnote": [], + "bbox": [ + 158, + 87, + 851, + 390 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Both the object and the trajectory are in metric scale. We can derive additional ground truth for each video from the aligned 3D object properties.", + "bbox": [ + 88, + 469, + 482, + 515 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Data Capture. To feature real world recordings observed from human perspectives that can be representative of 3D object reconstruction for AR/MR devices or robots, we provide two types of recording trajectories for selected objects, termed active and passive respectively, collected by human wearers. The active recording features a complete 360 view of the objects, which is similar to existing object 3D novel view synthesis dataset. The passive recording features casual looks from human wearer, which only contain partial views of the object from certain viewing angles along the trajectory. In dataset creation, we collected the active and passive recordings in the same environment and generate their 3D information in the shared 3D space. This helps reduces potential failures when aligning the object to passive 3D recordings, which are shorter and contains less 3D information. To reduce the effect of noises and motion blur, which are common in egocentric videos in indoor low light environments, we light the capture environment to $3\\mathrm{K}+$ lux illumination and used a fixed low-exposure and gain profile to collect each recording with the appropriate brightness.", + "bbox": [ + 91, + 518, + 483, + 821 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Alignment between Egocentric Video and Object. We provide an illustration of the object alignment to the egocentric video in Fig. 7. Given the images, camera poses and semi-dense point cloud acquired from Project Aria tools, we employ a neural-based mesh reconstruction method [62] to", + "bbox": [ + 88, + 825, + 485, + 902 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "create reference meshes for generating high-quality reference masks and used that to align with the corresponding 3D object mesh. This alignment step serves as an initialization for a more precise pose refinement, which leverages differentiable rendering. For certain objects with symmetric geometry, we observe this process can introduce ambiguities in point cloud registration and subsequent failures. To address such cases, we provide a GUI to manually align and correct the object alignment. Finally, akin to the DSLR camera pose registration phase, we optimize a mask loss over object poses to achieve fine-tuned pose registration.", + "bbox": [ + 511, + 469, + 908, + 636 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "4. Benchmarking and Applications", + "text_level": 1, + "bbox": [ + 511, + 650, + 808, + 667 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "We first use our DLSR and egocentric dataset as a benchmark for existing state-of-the-art methods. For inverse rendering, we design metrics to evaluate the shape and material quality of the recovered 3D object digital twin from three perspectives. For egocentric recording, we evaluate the novel-view synthesis as the initial evaluation. We include additional tasks, e.g. sparse view reconstruction in the supplementary materials for both DSLR and egocentric recordings. Finally we can demonstrate our high quality 3D digital twin models can be beneficial to robotics domain using an application in robotics manipulation.", + "bbox": [ + 511, + 670, + 908, + 837 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "4.1. Application to Inverse Rendering for DSLR", + "text_level": 1, + "bbox": [ + 511, + 847, + 883, + 864 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "The DSLR dataset in DTC provides accurate ground truth, including poses, lighting and 3D models, for inverse rendering", + "bbox": [ + 511, + 869, + 906, + 902 + ], + "page_idx": 5 + }, + { + "type": "page_number", + "text": "758", + "bbox": [ + 486, + 945, + 509, + 955 + ], + "page_idx": 5 + }, + { + "type": "table", + "img_path": "images/66a1b9d8428860a6cacaef98e9368932c4ff416345451293221c6ba11ddb4a8d.jpg", + "table_caption": [ + "Table 2. Benchmark comparison of existing methods on inverse rendering for DSLR. Depth SI-MSE and Shape Chamfer distance $\\times {10}^{-3}$ ." + ], + "table_footnote": [], + "table_body": "
GeometryNovel Scene RelightingNovel View Synthesis
Depth↓Normal↓Shape↓PSNR-H↑PSNR-L↑SSIM↑LPIPS↓PSNR-H↑PSNR-L↑SSIM↑LPIPS↓
Neural-PIL [12]5.710.2525.02N/A28.4235.760.8820.096
PhySG [80]0.310.1611.3127.2832.860.9590.04928.5434.460.9640.045
NVDiffRec [53]0.020.071.6426.9933.270.9510.03728.9534.920.9670.029
NeRD [11]4.550.45108.2026.1032.600.9480.06126.8033.400.8820.102
InvRender [71]0.220.030.7529.5235.980.9610.03731.6437.820.9700.033
NVDiffRecMC [29]0.020.061.3427.7834.550.9520.04231.2738.170.9720.032
", + "bbox": [ + 94, + 114, + 903, + 224 + ], + "page_idx": 6 + }, + { + "type": "table", + "img_path": "images/a234b44032efe9c9f6c7f33fd20e7e0ecc49d74e3d089a7a4236cf3cccd9627d.jpg", + "table_caption": [ + "Table 3. Benchmark on the egocentric aligned recordings." + ], + "table_footnote": [], + "table_body": "
PSNR ↑LPIPS ↓SSIM ↑Depth ↓Normal ↓
3D-GS [37]28.810.0200.98880.17680.3301
2D-GS [31]28.750.0200.98860.17550.2112
", + "bbox": [ + 94, + 272, + 480, + 325 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "tasks and serves as an evaluation suite to benchmark the performance of inverse rendering methods. We select six prior methods for this task and evaluate their performance using the ground truth provided by our dataset. In the following sections, we describe the data splitting strategy, evaluation metrics, and baselines.", + "bbox": [ + 89, + 351, + 483, + 441 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Data Splitting. For benchmarking purposes, we select 15 objects from the DSLR dataset captured under two distinct lighting environments, resulting in a total of 30 image sequences. The selected objects encompass a diverse range of geometric and material properties to ensure a comprehensive evaluation. For each scene, 8 views are selected for testing, while the remaining views are reserved for training.", + "bbox": [ + 89, + 445, + 483, + 551 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Evaluation Metrics. The metrics measure the accuracy of three aspects of baseline performance: geometry estimation, relighting, and novel view synthesis. For geometry estimation, we evaluate the accuracy of predicted depth and normal maps under held-out test views, as well as 3D meshes extracted from baseline methods, compared with the ground truth from our dataset. Relighting metrics evaluate the material decomposition quality of baselines by measuring the accuracy of predicted images under held-out lighting conditions. For view synthesis, we compare the predicted images from viewpoints unseen during training to ground truth captures. We refer to Kuang et al. [39] for metric details.", + "bbox": [ + 89, + 554, + 483, + 736 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Baselines. We include the following baselines: NVDiffRec [53] and NVDiffRecMC [29], with a hybrid shape representation DMTet [60]; InvRender [85] and PhySG [80], which adopt signed distance functions (SDFs) to represent object geometry [76] and utilize implicit neural fields for material decomposition; Neural-PIL [12] and NeRD [11], which use NeRFs [52] as scene representations.", + "bbox": [ + 89, + 739, + 483, + 845 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "4.2. Application to Egocentric Reconstruction", + "text_level": 1, + "bbox": [ + 89, + 854, + 444, + 869 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Our digital twin models, aligned with real world video using the method described in Sec. 3.4, can help obtain accurate", + "bbox": [ + 89, + 869, + 482, + 900 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "ground truth for object-centric images that were previously difficult to acquire. We provide the first evaluation of object-centric novel view synthesis recorded from an egocentric device. We use the projected object shape given the 3D pose of the object and cameras in scene coordinates to acquire the image masks, depth and normal for each object. We selected 15 recordings from the egocentric recording sessions as the evaluation and used the active recordings to benchmark novel view reconstruction. For each recording, we hold out every 8th image view as a testing view.", + "bbox": [ + 511, + 250, + 906, + 400 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "**Evaluations.** We build our baselines based on the gsplats [77] implementation of the 3D Gaussian Splitting (GS) [37] and 2D GS [31], and handle the effect of lens shading from the Project Aria lens [26]. We calculate PSNR, depth and normal based on the observed objects with masks and provide SSIM and LPIPS score on images by masking out the non-object areas as black. Table 3 shows the benchmark results of the baselines. We use the same depth and normal metric in DSLR evaluation. We provide additional qualitative evaluations and analyses on egocentric data towards sparser view settings in the supplementary materials.", + "bbox": [ + 511, + 404, + 908, + 571 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "4.3. Application to Robotic Manipulation", + "text_level": 1, + "bbox": [ + 511, + 582, + 831, + 599 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "High-quality object models have been leveraged in prior work to train real-world robotic agents in scenes represented explicitly [54] or implicitly [64]. These object models have also been shown to facilitate object-centric pose and lighting parameter estimation, enabling model-based planning [63]. In this section, we empirically evaluate the effectiveness of using DTC dataset objects in training robotic policies. Specifically, we consider learning robotic pushing and grasping skills in simulation.", + "bbox": [ + 511, + 604, + 906, + 739 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "First, we sample a subset of 24 cup category objects from the DTC dataset and 24 cup objects from Objaverse-XL [19] ${}^{1}$ . Since not all Objverse-XL objects come with textures, we randomize the colors of those objects uniformly in RGB space. To compute collision meshes for physical simulation, we perform convex decomposition on each object with CoACD [68]. We import these objects along with a UR5e robot equipped with a Robotiq 2F-85 (pushing) or", + "bbox": [ + 511, + 742, + 908, + 863 + ], + "page_idx": 6 + }, + { + "type": "page_footnote", + "text": "1The version of Objaverse-XL used in this work excludes all 3D models sourced from Sketchfab.", + "bbox": [ + 511, + 875, + 906, + 898 + ], + "page_idx": 6 + }, + { + "type": "page_number", + "text": "759", + "bbox": [ + 486, + 945, + 509, + 955 + ], + "page_idx": 6 + }, + { + "type": "table", + "img_path": "images/a6ab799d2042df478496745334f08ef485d78cac29c002571b9ec76578bd80f9.jpg", + "table_caption": [], + "table_footnote": [], + "table_body": "
TaskDTC (ours)Objverse-XL [19]
Pushing @ 2cm36.3% ± 1.5%25.3% ± 6.0%
Pushing @ 3cm43.7% ± 1.2%29.7% ± 6.0%
Pushing @ 5cm47.0% ± 2.6%40.3% ± 5.5%
Grasping42.7% ± 4.7%38.6% ± 11.0%
", + "bbox": [ + 109, + 88, + 467, + 174 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Table 4. Success rate of policies trained on data collected using objects from our DTC dataset and sampled from Objaverse-XL when evaluated on an unseen test object. Errors indicate sample standard deviation over three policy training seeds.", + "bbox": [ + 89, + 184, + 482, + 241 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Robotiq 2F-140 (grasping) gripper into the PyBullet simulator [17] and collect data for each robotic task as described below. After training policies on data from each object set, we evaluate policy performance on a relatively high-quality unseen test object from the StanfordORB dataset [39].", + "bbox": [ + 89, + 280, + 483, + 357 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Pushing. For the pushing task, we collect 5000 trajectories of pseudo-random robotic interaction data for each object set. For each trajectory, a single object from the considered object set and its initial position are randomly selected. Then we train a goal-conditioned neural network policy $\\pi (a|o,o_g)$ where $o$ is an image observation of the current scene and $o_{g}$ is a goal image indicating the desired final object and robot position, $o,o_{g}\\in \\mathbb{R}^{256\\times 256\\times 3}$ , and the action $a\\in \\mathbb{R}^2$ represents a change in the robot's end-effector position in the $x$ and $y$ axes. The $z$ axis end-effector height is held fixed. We sample goals for training via hindsight relabeling [4, 20, 24]. We then perform evaluation on 100 randomly sampled test goals manipulating an unseen test cup from StanfordORB [39].", + "bbox": [ + 88, + 366, + 482, + 561 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Grasping. For grasping, we collect 5000 successful grasp examples for each object set by first placing a single object into the scene, randomizing the object identity and initial position. We then randomly sample candidate grasp poses in a radius around the object's position and simulating their outcomes, rejecting unsuccessful grasps. We train a grasping policy $\\pi(a|o)$ where $o \\in \\mathbb{R}^{256 \\times 256 \\times 3}$ is an image observation of the scene and $a \\in \\mathbb{R}^4$ represents the $x, y, z$ position and $\\theta$ yaw rotation of the robot end-effector at which to attempt the grasp. Again we use 100 test cup object poses.", + "bbox": [ + 89, + 573, + 482, + 724 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Results. We report the results in Table 4 and Fig. 8. We find that across both pushing and grasping tasks, policies trained on DTC dataset objects outperform those trained on Objaverse-XL objects when evaluated on the unseen test object. For pushing, we report performance by defining binary success thresholds based on the final Euclidean distance of the object position to the goal position. Training on DTC objects appears to be especially helpful at enabling policies to make finer adjustments, improving pushing success rates at stricter thresholds. Additional experimental details can be found in the supplementary.", + "bbox": [ + 89, + 734, + 482, + 901 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/3b4d8b3f2caaf2fc9858feb7baa824272d8d29f523c49ef5c6ccb9d47ac20594.jpg", + "image_caption": [ + "Figure 8. Success rates on robotic pushing task when training with our DTC objects and sampled Objaverse-XL objects. Particularly at lower error thresholds, policies trained on DTC objects outperform those trained using Objaverse-XL objects. Shaded bars represent sample standard deviations over policy training random seeds." + ], + "image_footnote": [], + "bbox": [ + 517, + 90, + 901, + 294 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "5. Conclusion", + "text_level": 1, + "bbox": [ + 513, + 383, + 633, + 398 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "We presented a new large scale photorealistic 3D digital twin dataset with the real world recordings that contain its real world counterpart. We provide extensive evaluations of baselines on our DTC dataset serving as new benchmark for inverse rendering and novel view synthesis task. We also demonstrated that high quality digital twin models can be beneficial to applications in robotics domain. We believe our efforts can empower the research community to build and leverage digital twin models for future applications.", + "bbox": [ + 511, + 409, + 906, + 545 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Limitations. Achieving high quality digital twin models currently requires deliberate hardware setup and human efforts in refinement. Solving this challenge without sacrificing quality can significantly further enhance the volume of digital twin models. Our hardware is also limited to objects within a certain size and can not yet recover objects that are deformable, highly specular, or transparent.", + "bbox": [ + 511, + 547, + 908, + 655 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Future work. The existing digital twin model creation in DTC dataset involves lengthy post-processing and may require subjective human refinement, hindering the automation of model generation. However, recent advancements in physics-based differentiable rendering hold promise for enabling faster and more accurate creation of digital twins, especially for material reconstruction. Furthermore, building large-scale digital twins for applications will necessitate efforts to enhance the diversity in object appearance (e.g., transparent objects) and to capture additional attributes, such as physical properties and functionalities.", + "bbox": [ + 511, + 657, + 908, + 824 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "For robotics applications, while deploying manipulation policies learned in simulation to the real world remains generally challenging, we hope that the high-quality digital twin data provided by DTC can serve as a stepping stone towards effective sim-to-real transfer.", + "bbox": [ + 511, + 825, + 908, + 898 + ], + "page_idx": 7 + }, + { + "type": "page_number", + "text": "760", + "bbox": [ + 486, + 945, + 509, + 955 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Acknowledgments. This work is in part supported by NSF CCRI #2120095, RI #2211258, and RI #2338203. YZ is in part supported by the Stanford Interdisciplinary Graduate Fellowship.", + "bbox": [ + 89, + 90, + 485, + 154 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "References", + "text_level": 1, + "bbox": [ + 91, + 167, + 187, + 183 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[1] Covision 3D object scanner. https://www.covisionmedia.ai/en/scanners.4", + "[2] Kosmas Alexopoulos, Nikolaos Nikolakis, and George Chryssolouris. Digital twin-driven supervised machine learning for the development of artificial intelligence applications in manufacturing. International Journal of Computer Integrated Manufacturing, 33(5):429-439, 2020. 2", + "[3] Moayad Aloqaily, Ouns Bouachir, Fakhri Karray, Ismaeel Al Ridhawi, and Abdulmotaleb El Saddik. Integrating digital twin and advanced intelligent technologies to realize the meta-verse. IEEE Consumer Electronics Magazine, 12(6):47-55, 2022. 2", + "[4] Marcin Andrychowicz, Filip Wolski, Alex Ray, Jonas Schneider, Rachel Fong, Peter Welinder, Bob McGrew, Josh Tobin, OpenAI Pieter Abbeel, and Wojciech Zaremba. Hindsight experience replay. In NeurIPS, 2017. 8", + "[5] Jonathan T Barron and Jitendra Malik. Shape, illumination, and reflectance from shading. IEEE TPAMI, 37(8):1670-1687, 2014. 3", + "[6] Sean Bell, Kavita Bala, and Noah Snavely. Intrinsic images in the wild. ACM TOG, 33(4):1-12, 2014. 3", + "[7] Sai Bi, Zexiang Xu, Kalyan Sunkavalli, Miloš Hasan, Yannick Hold-Geoffroy, David Kriegman, and Ravi Ramamoorthi. Deep reflectance volumes: Relightable reconstructions from multi-view photometric images. In ECCV, 2020. 3", + "[8] Sai Bi, Zexiang Xu, Kalyan Sunkavalli, David Kriegman, and Ravi Ramamoorthi. Deep 3D capture: Geometry and reflectance from sparse multi-view images. In CVPR, 2020. 4", + "[9] Martin Bichsel and Alex P Pentland. A simple algorithm for shape from shading. In CVPR, 1992. 3", + "[10] Mark Boss, Varun Jampani, Kihwan Kim, Hendrik P.A. Lensch, and Jan Kautz. Two-shot Spatially-varying BRDF and Shape Estimation. In CVPR, 2020. 4", + "[11] Mark Boss, Raphael Braun, Varun Jampani, Jonathan T. Barron, Ce Liu, and Hendrik P.A. Lensch. NeRD: Neural reflectance decomposition from image collections. In ICCV, 2021. 3, 7", + "[12] Mark Boss, Varun Jampani, Raphael Braun, Ce Liu, Jonathan Barron, and Hendrik Lensch. Neural-PIL: Neural pre-integrated lighting for reflectance decomposition. In NeurIPS, 2021. 7", + "[13] Mark Boss, Andreas Engelhardt, Abhishek Kar, Yuanzhen Li, Deqing Sun, Jonathan Barron, Hendrik Lensch, and Varun Jampani. SAMURAI: Shape and material from unconstrained real-world arbitrary image collections. In NeurIPS, 2022. 3", + "[14] Guangyan Cai, Kai Yan, Zhao Dong, Ioannis Gkioulekas, and Shuang Zhao. Physics-based inverse rendering using combined implicit and explicit geometries. CGF, 41(4):129-138, 2022. 3" + ], + "bbox": [ + 93, + 193, + 485, + 898 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[15] Wenzheng Chen, Joey Litalien, Jun Gao, Zian Wang, Clement Fuji Tsang, Sameh Khamis, Or Litany, and Sanja Fidler. DIBR++: learning to predict lighting and material with a hybrid differentiable renderer. In NeurIPS, 2021. 3", + "[16] Jasmine Collins, Shubham Goel, Kenan Deng, Achleshwar Luthra, Leon Xu, Erhan Gundogdu, Xi Zhang, Tomas F Yago Vicente, Thomas Dideriksen, Himanshu Arora, et al. ABO: Dataset and benchmarks for real-world 3D object understanding. In CVPR, 2022. 2, 3", + "[17] Erwin Coumans and Yunfei Bai. Pybullet, a python module for physics simulation for games, robotics and machine learning. http://pybullet.org, 2016-2019.8", + "[18] Matt Deitke, Dustin Schwenk, Jordi Salvador, Luca Weihs, Oscar Michel, Eli VanderBilt, Ludwig Schmidt, Kiana Ehsani, Aniruddha Kembhavi, and Ali Farhadi. Objverse: A universe of annotated 3D objects. In CVPR, 2023. 2, 3", + "[19] Matt Deitke, Ruoshi Liu, Matthew Wallingford, Huong Ngo, Oscar Michel, Aditya Kusupati, Alan Fan, Christian Laforte, Vikram Voleti, Samir Yitzhak Gadre, et al. Objverse-XL: A universe of $10\\mathrm{m} + 3\\mathrm{D}$ objects. NeurIPS, 2024. 3, 7, 8", + "[20] Yiming Ding, Carlos Florensa, Pieter Abbeel, and Mariano Philipp. Goal-conditioned imitation learning. In NeurIPS, 2019. 8", + "[21] Laura Downs, Anthony Francis, Nate Koenig, Brandon Kinman, Ryan Hickman, Krista Reymann, Thomas B McHugh, and Vincent Vanhoucke. Google scanned objects: A high-quality dataset of 3D scanned household items. In ICRA, 2022. 2, 3", + "[22] Jakob Engel, Kiran Somasundaram, Michael Goesele, Albert Sun, Alexander Gamino, Andrew Turner, Arjang Talatof, Arnie Yuan, Bilal Souti, Brighid Meredith, Cheng Peng, Chris Sweeney, Cole Wilson, Dan Barnes, Daniel DeTone, David Caruso, Derek Valleroy, Dinesh Ginjupalli, Duncan Frost, Edward Miller, Elias Mueggler, Evgeniy Oleinik, Fan Zhang, Guruprasad Somasundaram, Gustavo Solaira, Harry Lanaras, Henry Howard-Jenkins, Huixuan Tang, Hyo Jin Kim, Jaime Rivera, Ji Luo, Jing Dong, Julian Straub, Kevin Bailey, Kevin Eckenhoff, Lingni Ma, Luis Pesqueira, Mark Schwesinger, Maurizio Monge, Nan Yang, Nick Charron, Nikhil Raina, Omkar Parkhi, Peter Borschowa, Pierre Moulon, Prince Gupta, Raul Mur-Artal, Robbie Pennington, Sachin Kulkarni, Sagar Miglani, Santosh Gondi, Saransh Solanki, Sean Diener, Shangyi Cheng, Simon Green, Steve Saarinen, Suvam Patra, Tassos Mourikis, Thomas Whelan, Tripti Singh, Vasileios Balntas, Vijay Baiyya, Wilson Dreewes, Xiaqing Pan, Yang Lou, Yipu Zhao, Yusuf Mansour, Yuyang Zou, Zhaoyang Lv, Zijian Wang, Mingfei Yan, Carl Ren, Renzo De Nardi, and Richard Newcombe. Project Aria: A new tool for egocentric multi-modal ai research, 2023. 5", + "[23] Ruohan Gao, Yen-Yu Chang, Shivani Mall, Li Fei-Fei, and Jiajun Wu. ObjectFolder: A dataset of objects with implicit visual, auditory, and tactile representations. In CoRL, 2021. 2", + "[24] Dibya Ghosh, Abhishek Gupta, Ashwin Reddy, Justin Fu, Coline Devin, Benjamin Eysenbach, and Sergey Levine. Learning to reach goals via iterated supervised learning. In ICML, 2019. 8", + "[25] Roger Grosse, Micah K Johnson, Edward H Adelson, and" + ], + "bbox": [ + 516, + 92, + 906, + 901 + ], + "page_idx": 8 + }, + { + "type": "page_number", + "text": "761", + "bbox": [ + 486, + 945, + 509, + 955 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "William T Freeman. Ground truth dataset and baseline evaluations for intrinsic image algorithms. In ICCV, 2009. 3", + "[26] Qiao Gu, Zhaoyang Lv, Duncan Frost, Simon Green, Julian Straub, and Chris Sweeney. Egolifter: Open-world 3d segmentation for egocentric perception. In ECCV, pages 382-400. Springer, 2024. 7", + "[27] Agrim Gupta, Piotr Dollar, and Ross Girshick. LVIS: A dataset for large vocabulary instance segmentation. In CVPR, 2019. 4", + "[28] Wassim Hamidouche, Lina Bariah, and Mérouane Debbah. Immersive media and massive twinning: Advancing toward the metaverse. IEEE Communications Magazine, 62(7):20-32, 2024. 2", + "[29] Jon Hasselgren, Nikolai Hofmann, and Jacob Munkberg. Shape, light & material decomposition from images using Monte Carlo rendering and denoising. In NeurIPS, 2022. 4, 7", + "[30] Yicong Hong, Kai Zhang, Jiuxiang Gu, Sai Bi, Yang Zhou, Difan Liu, Feng Liu, Kalyan Sunkavalli, Trung Bui, and Hao Tan. LRM: Large reconstruction model for single image to 3D. In The Twelfth International Conference on Learning Representations, 2024. 2", + "[31] Binbin Huang, Zehao Yu, Anpei Chen, Andreas Geiger, and Shenghua Gao. 2D gaussian splatting for geometrically accurate radiance fields. In SIGGRAPH 2024 Conference Papers, 2024. 3, 7", + "[32] Wenzel Jakob, Sébastien Speierer, Nicolas Roussel, Merlin Nimier-David, Delio Vicini, Tizian Zeltner, Baptiste Nicolet, Miguel Crespo, Vincent Leroy, and Ziyi Zhang. Mitsuba 3 renderer, 2022. https://mitsuba-renderer.org. 5", + "[33] Michael Janner, Jiajun Wu, Tejas Kulkarni, Ilker Yildirim, and Joshua B Tenenbaum. Self-supervised intrinsic image decomposition. In NeurIPS, 2017. 4", + "[34] Rasmus Jensen, Anders Dahl, George Vogiatzis, Engin Tola, and Henrik Aanæs. Large scale multi-view stereopsis evaluation. In CVPR, 2014. 2, 3", + "[35] Haian Jin, Isabella Liu, Peijia Xu, Xiaoshuai Zhang, Songfang Han, Sai Bi, Xiaowei Zhou, Zexiang Xu, and Hao Su. TensoIR: Tensorial inverse rendering. In CVPR, 2023. 2", + "[36] James T Kajiya. The rendering equation. In Proceedings of the 13th annual conference on Computer graphics and interactive techniques, pages 143-150, 1986. 3", + "[37] Bernhard Kerbl, Georgios Kopanas, Thomas Leimkuhler, and George Drettakis. 3D gaussian splattering for real-time radiance field rendering. ACM TOG, 2023. 2, 3, 7", + "[38] Zhengfei Kuang, Kyle Olszewski, Mengelei Chai, Zeng Huang, Panos Achlioptas, and Sergey Tulyakov. NeROIC: neural rendering of objects from online image collections. ACM TOG, 41(4):1-12, 2022. 3", + "[39] Zhengfei Kuang, Yunzhi Zhang, Hong-Xing Yu, Samir Agarwala, Elliott Wu, Jiajun Wu, et al. Stanford-orb: a real-world 3D object inverse rendering benchmark. In Advances in Neural Information Processing Systems Datasets and Benchmarks Track, 2023. 3, 5, 7, 8", + "[40] Samuli Laine, Janne Hellsten, Tero Karras, Yeongho Seol, Jaakko Lehtinen, and Timo Aila. Modular primitives for high-performance differentiable rendering. ACM Transactions on Graphics, 39(6), 2020. 5" + ], + "bbox": [ + 91, + 90, + 483, + 900 + ], + "page_idx": 9 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[41] Jiahao Li, Hao Tan, Kai Zhang, Zexiang Xu, Fujun Luan, Yinghao Xu, Yicong Hong, Kalyan Sunkavalli, Greg Shakhnarovich, and Sai Bi. Instant3D: Fast text-to-3D with sparse-view generation and large reconstruction model. In ICLR, 2024. 2", + "[42] Min Li, Zhenglong Zhou, Zhe Wu, Boxin Shi, Changyu Diao, and Ping Tan. Multi-view photometric stereo: A robust solution and benchmark dataset for spatially varying isotropic materials. IEEE TIP, 29:4159-4173, 2020. 3", + "[43] Xuanlin Li, Kyle Hsu, Jiayuan Gu, Karl Pertsch, Oier Mees, Homer Rich Walke, Chuyuan Fu, Ishikaa Lunawat, Isabel Sieh, Sean Kirmani, Sergey Levine, Jiajun Wu, Chelsea Finn, Hao Su, Quan Vuong, and Ted Xiao. Evaluating real-world robot manipulation policies in simulation. arXiv preprint arXiv:2405.05941, 2024. 2", + "[44] Zhengqin Li, Kalyan Sunkavalli, and Manmohan Chandraker. Materials for masses: SVBRDF acquisition with a single mobile phone image. In ECCV, 2018. 3", + "[45] Zhengqin Li, Zexiang Xu, Ravi Ramamoorthi, Kalyan Sunkavalli, and Manmohan Chandraker. Learning to reconstruct shape and spatially-varying reflectance from a single image. ACM TOG, 37(6):1-11, 2018. 3, 4", + "[46] Isabella Liu, Linghao Chen, Ziyang Fu, Liwen Wu, Haian Jin, Zhong Li, Chin Ming Ryan Wong, Yi Xu, Ravi Ramamoorthi, Zexiang Xu, and Hao Su. OpenIllumination: A multi-illumination dataset for inverse rendering evaluation on real objects. NeurIPS, 2023. 2, 3", + "[47] Stephen Lombardi and Ko Nishino. Reflectance and natural illumination from a single image. In ECCV, 2012. 3", + "[48] Stephen Lombardi and Ko Nishino. Reflectance and illumination recovery in the wild. IEEE TPAMI, 2015. 3", + "[49] Fujun Luan, Shuang Zhao, Kavita Bala, and Zhao Dong. Unified shape and SVBRDF recovery using differentiable Monte Carlo rendering. CGF, 40(4):101-113, 2021. 3", + "[50] Stephen Robert Marschner. Inverse Rendering for Computer Graphics. Cornell University, 1998. 3", + "[51] Nelson Max. Optical models for direct volume rendering. IEEE TVCG, 1(2):99-108, 1995. 3", + "[52] Ben Mildenhall, Pratul P Srinivasan, Matthew Tancik, Jonathan T Barron, Ravi Ramamoorthi, and Ren Ng. NeRF: Representing scenes as neural radiance fields for view synthesis. In ECCV, 2020. 2, 3, 7", + "[53] Jacob Munkberg, Jon Hasselgren, Tianchang Shen, Jun Gao, Wenzheng Chen, Alex Evans, Thomas Müller, and Sanja Fidler. Extracting triangular 3D models, materials, and lighting from images. In CVPR, 2022. 4, 7", + "[54] Soroush Nasiriany, Abhiram Maddukuri, Lance Zhang, Adeet Parikh, Aaron Lo, Abhishek Joshi, Ajay Mandlekar, and Yuke Zhu. Robocasa: Large-scale simulation of everyday tasks for generalist robots. In Robotics: Science and Systems (RSS), 2024. 7", + "[55] Michael Niemeyer, Lars Mescheder, Michael Oechsle, and Andreas Geiger. Differentiable volumetric rendering: Learning implicit 3D representations without 3D supervision. In CVPR, 2020. 4", + "[56] Geoffrey Oxholm and Ko Nishino. Multiview shape and reflectance from natural illumination. In CVPR, 2014. 3" + ], + "bbox": [ + 516, + 90, + 905, + 900 + ], + "page_idx": 9 + }, + { + "type": "page_number", + "text": "762", + "bbox": [ + 486, + 945, + 511, + 955 + ], + "page_idx": 9 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[57] Geoffrey Oxholm and Ko Nishino. Shape and reflectance estimation in the wild. IEEE TPAMI, 2015. 3", + "[58] Xiaqing Pan, Nicholas Charron, Yongqian Yang, Scott Peters, Thomas Whelan, Chen Kong, Omkar Parkhi, Richard Newcombe, and Yuheng (Carl) Ren. Aria digital twin: A new benchmark dataset for egocentric 3D machine perception. In Proceedings of the IEEE/CVF International Conference on Computer Vision (ICCV), pages 20133-20143, 2023. 3", + "[59] Viktor Rudnev, Mohamed Elgharib, William Smith, Lingjie Liu, Vladislav Golyanik, and Christian Theobalt. NeRF for outdoor scene relighting. In ECCV, pages 615-631. Springer, 2022. 3", + "[60] Tianchang Shen, Jun Gao, Kangxue Yin, Ming-Yu Liu, and Sanja Fidler. Deep marching tetrahedra: a hybrid representation for high-resolution 3D shape synthesis. In NeurIPS, 2021. 7", + "[61] Jian Shi, Yue Dong, Hao Su, and Stella X Yu. Learning non-Lambertian object intrinsics across ShapeNet categories. In CVPR, 2017. 2, 3, 4", + "[62] Cheng Sun, Guangyan Cai, Zhengqin Li, Kai Yan, Cheng Zhang, Carl Marshall, Jia-Bin Huang, Shuang Zhao, and Zhao Dong. Neural-PBIR reconstruction of shape, material, and illumination. In ICCV, 2023. 2, 4, 5, 6", + "[63] Stephen Tian, Yancheng Cai, Hong-Xing Yu, Sergey Zakharov, Katherine Liu, Adrien Gaidon, Yunzhu Li, and Ji-jun Wu. Multi-object manipulation via object-centric neural scattering functions. In CVPR, 2023. 7", + "[64] Marcel Torne, Anthony Simeonov, Zechu Li, April Chan, Tao Chen, Abhishek Gupta, and Pulkit Agrawal. Reconciling reality through simulation: A real-to-sim-to-real approach for robust manipulation. Arxiv, 2024. 7", + "[65] Marco Toschi, Riccardo De Matteo, Riccardo Spezialetti, Daniele De Gregorio, Luigi Di Stefano, and Samuele Salti. Relight my NeRF: A dataset for novel view synthesis and relighting of real world objects. In CVPR, 2023. 3", + "[66] Peng Wang, Lingjie Liu, Yuan Liu, Christian Theobalt, Taku Komura, and Wenping Wang. NeuS: Learning neural implicit surfaces by volume rendering for multi-view reconstruction. In NeurIPS, 2021. 4", + "[67] Henrique Weber, Donald Prévost, and Jean-François Lalonde. Learning to estimate indoor lighting from 3D objects. In 3DV, 2018. 3", + "[68] Xinyue Wei, Minghua Liu, Zhan Ling, and Hao Su. Approximate convex decomposition for 3D meshes with collision-aware concavity and tree search. ACM TOG, 2022. 7", + "[69] Xinyue Wei, Kai Zhang, Sai Bi, Hao Tan, Fujun Luan, Valentin Deschaintre, Kalyan Sunkavalli, Hao Su, and Zexiang Xu. MeshLRM: Large reconstruction model for high-quality mesh. arXiv preprint arXiv:2404.12385, 2024. 2", + "[70] Felix Wimbauer, Shangzhe Wu, and Christian Rupprecht. De-rendering 3D objects in the wild. In CVPR, 2022. 3, 4", + "[71] Haoqian Wu, Zhipeng Hu, Lincheng Li, Yongqiang Zhang, Changjie Fan, and Xin Yu. NeFII: Inverse rendering for reflectance decomposition with near-field indirect illumination. In CVPR, 2023. 4, 7", + "[72] Shangzhe Wu, Ameesh Makadia, Jiajun Wu, Noah Snavely, Richard Tucker, and Angjoo Kanazawa. De-rendering the world's revolutionary artefacts. In CVPR, 2021. 3, 4" + ], + "bbox": [ + 91, + 90, + 483, + 900 + ], + "page_idx": 10 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[73] Tong Wu, Jiarui Zhang, Xiao Fu, Yuxin Wang, Liang Pan Jiawei Ren, Wayne Wu, Lei Yang, Jiaqi Wang, Chen Qian, Dahua Lin, and Ziwei Liu. OmniObject3D: Large-vocabulary 3D object dataset for realistic perception, reconstruction and generation. In IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), 2023. 2, 3", + "[74] Kohei Yamashita, Yuto Enyo, Shohei Nobuhara, and Ko Nishino. nLMVS-Net: Deep non-lambertian multi-view stereo. In WACV, 2023. 3", + "[75] Kai Yan and Shuang Zhao. PSDR-JIT: Path-space differentiable renderer, 2020. https://github.com/andyyankai/psdr-jit.5", + "[76] Lior Yariv, Yoni Kasten, Dror Moran, Meirav Galun, Matan Atzmon, Basri Ronen, and Yaron Lipman. Multiview neural surface reconstruction by disentangling geometry and appearance. In NeurIPS, 2020. 7", + "[77] Vickie Ye, Ruilong Li, Justin Kerr, Matias Turkulainen, Brent Yi, Zhuoyang Pan, Otto Seiskari, Jianbo Ye, Jeffrey Hu, Matthew Tancik, and Angjoo Kanazawa. gsplat: An open-source library for Gaussian splatting. arXiv preprint arXiv:2409.06765, 2024. 7", + "[78] Hong-Xing Yu, Samir Agarwala, Charles Herrmann, Richard Szeliski, Noah Snavely, Jiajun Wu, and Deqing Sun. Accidental light probes. In CVPR, 2023. 3, 5", + "[79] Hong-Xing Yu, Michelle Guo, Alireza Fathi, Yen-Yu Chang, Eric Ryan Chan, Ruohan Gao, Thomas Funkhouser, and Jia-jun Wu. Learning object-centric neural scattering functions for free-viewpoint relighting and scene composition. TMLR, 2023. 3", + "[80] Kai Zhang, Fujun Luan, Qianqian Wang, Kavita Bala, and Noah Snavely. PhySG: Inverse rendering with spherical gaussians for physics-based material editing and relighting. In CVPR, 2021. 4, 7", + "[81] Kai Zhang, Fujun Luan, Zhengqi Li, and Noah Snavely. IRON: Inverse rendering by optimizing neural SDFs and materials from photometric images. In CVPR, 2022. 4", + "[82] Kai Zhang, Sai Bi, Hao Tan, Yuanbo Xiangli, Nanxuan Zhao, Kalyan Sunkavalli, and Zexiang Xu. GS-LRM: Large reconstruction model for 3D gaussian splatting. arXiv preprint arXiv:2404.19702, 2024. 2", + "[83] Ruo Zhang, Ping-Sing Tsai, James Edwin Cryer, and Mubarak Shah. Shape-from-shading: a survey. IEEE TPAMI, 21(8): 690-706, 1999. 3", + "[84] Xiuming Zhang, Pratul P Srinivasan, Boyang Deng, Paul Debevec, William T Freeman, and Jonathan T Barron. NeR-Factor: Neural factorization of shape and reflectance under an unknown illumination. ACM TOG, 40(6):1-18, 2021. 3", + "[85] Yuanqing Zhang, Jiaming Sun, Xingyi He, Huan Fu, Rongfei Jia, and Xiaowei Zhou. Modeling indirect illumination for inverse rendering. In CVPR, 2022. 4, 7", + "[86] Yunzhi Zhang, Shangzhe Wu, Noah Snavely, and Jiajun Wu. Seeing a rose in five thousand ways. In CVPR, 2023. 4" + ], + "bbox": [ + 516, + 90, + 906, + 828 + ], + "page_idx": 10 + }, + { + "type": "page_number", + "text": "763", + "bbox": [ + 486, + 945, + 511, + 955 + ], + "page_idx": 10 + } +] \ No newline at end of file diff --git a/2025/Digital Twin Catalog_ A Large-Scale Photorealistic 3D Object Digital Twin Dataset/2d29b0d5-5c08-413f-9ad0-63d98b0c6539_model.json b/2025/Digital Twin Catalog_ A Large-Scale Photorealistic 3D Object Digital Twin Dataset/2d29b0d5-5c08-413f-9ad0-63d98b0c6539_model.json new file mode 100644 index 0000000000000000000000000000000000000000..de9d4032fd6d358aefb67d2b96a5bd61efd904d0 --- /dev/null +++ b/2025/Digital Twin Catalog_ A Large-Scale Photorealistic 3D Object Digital Twin Dataset/2d29b0d5-5c08-413f-9ad0-63d98b0c6539_model.json @@ -0,0 +1,2488 @@ +[ + [ + { + "type": "header", + "bbox": [ + 0.107, + 0.003, + 0.182, + 0.043 + ], + "angle": 0, + "content": "CVF" + }, + { + "type": "header", + "bbox": [ + 0.237, + 0.0, + 0.812, + 0.045 + ], + "angle": 0, + "content": "This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore." + }, + { + "type": "title", + "bbox": [ + 0.092, + 0.131, + 0.908, + 0.153 + ], + "angle": 0, + "content": "Digital Twin Catalog: A Large-Scale Photorealistic 3D Object Digital Twin Dataset" + }, + { + "type": "text", + "bbox": [ + 0.102, + 0.18, + 0.895, + 0.235 + ], + "angle": 0, + "content": "Zhao Dong\\(^{1}\\) Ka Chen\\(^{1,*}\\) Zhaoyang Lv\\(^{1,*}\\) Hong-Xing Yu\\(^{2,*}\\) Yunzhi Zhang\\(^{2,*}\\) Cheng Zhang\\(^{1,*}\\) Yufeng Zhu\\(^{1,*}\\) Stephen Tian\\(^{2}\\) Zhengqin Li\\(^{1}\\) Georgie Moffatt\\(^{1}\\) Sean Christofferson\\(^{1}\\) James Fort\\(^{1}\\) Xiaqing Pan\\(^{1}\\) Mingfei Yan\\(^{1}\\) Jiajun Wu\\(^{2}\\) Carl Yuheng Ren\\(^{1}\\) Richard Newcombe\\(^{1}\\)" + }, + { + "type": "text", + "bbox": [ + 0.289, + 0.244, + 0.71, + 0.264 + ], + "angle": 0, + "content": "\\(^{1}\\)Meta Reality Labs Research Stanford University" + }, + { + "type": "image", + "bbox": [ + 0.099, + 0.311, + 0.434, + 0.571 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.093, + 0.578, + 0.44, + 0.592 + ], + "angle": 0, + "content": "(a) Examples of 2,000 digital twins of real-world objects" + }, + { + "type": "image", + "bbox": [ + 0.464, + 0.309, + 0.603, + 0.571 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.451, + 0.578, + 0.615, + 0.592 + ], + "angle": 0, + "content": "(b) Shape & PBR Materials" + }, + { + "type": "image", + "bbox": [ + 0.639, + 0.299, + 0.761, + 0.571 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.657, + 0.578, + 0.746, + 0.592 + ], + "angle": 0, + "content": "(c) DSLR Data" + }, + { + "type": "image", + "bbox": [ + 0.776, + 0.31, + 0.907, + 0.568 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.78, + 0.578, + 0.902, + 0.592 + ], + "angle": 0, + "content": "(d) Egocentric Data" + }, + { + "type": "image_caption", + "bbox": [ + 0.089, + 0.602, + 0.908, + 0.645 + ], + "angle": 0, + "content": "Figure 1. The Digital Twin Catalog (DTC) dataset comprises 2,000 digital twins of physical-world objects (a), characterized by millimeter-level geometric accuracy and photorealistic PBR materials (b). DTC includes evaluation data captured using both DSLR cameras and egocentric Aria glasses, featuring captured images with precise foreground object masks and environment lighting for relighting evaluation." + }, + { + "type": "title", + "bbox": [ + 0.248, + 0.654, + 0.327, + 0.669 + ], + "angle": 0, + "content": "Abstract" + }, + { + "type": "text", + "bbox": [ + 0.089, + 0.687, + 0.486, + 0.87 + ], + "angle": 0, + "content": "We introduce Digital Twin Catalog (DTC), a new large-scale photorealistic 3D object digital twin dataset. A digital twin of a 3D object is a highly detailed, virtually indistinguishable representation of a physical object, accurately capturing its shape, appearance, physical properties, and other attributes. Recent advances in neural-based 3D reconstruction and inverse rendering have significantly improved the quality of 3D object reconstruction. Despite these advancements, there remains a lack of a large-scale, digital twin quality real-world dataset and benchmark that can quantitatively assess and compare the performance of different reconstruction methods, as well as improve reconstruction quality through" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.655, + 0.909, + 0.882 + ], + "angle": 0, + "content": "training or fine-tuning. Moreover, to democratize 3D digital twin creation, it is essential to integrate creation techniques with next-generation egocentric computing platforms, such as AR glasses. Currently, there is no dataset available to evaluate 3D object reconstruction using egocentric captured images. To address these gaps, the DTC dataset features 2,000 scanned digital twin-quality 3D objects, along with image sequences captured under different lighting conditions using DSLR cameras and egocentric AR glasses. This dataset establishes the first comprehensive real-world evaluation benchmark for 3D digital twin creation tasks, offering a robust foundation for comparing and improving existing reconstruction methods. The DTC dataset is already released at https://www.projectaria.com/datasets/dtc/ and we will also make the baseline evaluations open-source." + }, + { + "type": "page_footnote", + "bbox": [ + 0.115, + 0.887, + 0.458, + 0.9 + ], + "angle": 0, + "content": "*Authors contributed equally and are listed in alphabetical order." + }, + { + "type": "page_number", + "bbox": [ + 0.486, + 0.946, + 0.511, + 0.957 + ], + "angle": 0, + "content": "753" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.093, + 0.089, + 0.253, + 0.214 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.255, + 0.089, + 0.416, + 0.214 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.418, + 0.089, + 0.579, + 0.214 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.581, + 0.089, + 0.742, + 0.214 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.743, + 0.09, + 0.905, + 0.214 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.093, + 0.216, + 0.253, + 0.34 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.255, + 0.216, + 0.416, + 0.34 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.418, + 0.216, + 0.579, + 0.34 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.581, + 0.216, + 0.742, + 0.34 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.743, + 0.216, + 0.905, + 0.34 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.298, + 0.35, + 0.698, + 0.365 + ], + "angle": 0, + "content": "Figure 2. Example DTC models with photorealistic PBR materials." + }, + { + "type": "title", + "bbox": [ + 0.092, + 0.39, + 0.223, + 0.405 + ], + "angle": 0, + "content": "1. Introduction" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.416, + 0.486, + 0.779 + ], + "angle": 0, + "content": "A digital twin of a 3D object is a highly detailed, virtually indistinguishable representation of a physical object, capturing its shape, appearance, physical properties and other attributes with precision. Such a digital twin enables visualization, analysis, and interaction as if it were the real object, supporting simulation, automation, and real-world problem-solving across a wide range of applications in AR/VR [3, 28], spatial/contextual AI [2], and robotics [23, 43]. As fundamental properties of an object, its shape and appearance form the basis for recognizing and interpreting the 3D object, enabling identification, manipulation, and realistic rendering. Recovering these attributes has long been a foundational topic in computer vision and graphics, inspiring extensive research in 3D reconstruction and inverse rendering. Recent breakthroughs in neural-based representation and reconstruction techniques, such as NeRF [52] and 3D Gaussian splattig (3DGS) [37], have significantly elevated the quality of novel view synthesis (NVS) to photorealistic levels. Many subsequent works [35, 62] integrate neural reconstruction with physically-based inverse rendering, enabling relightable appearances. Furthermore, leveraging priors from large reconstruction models (LRMs) [30], high-quality shape and appearance reconstruction can now be achieved with as few as one to four views [41, 69, 82]." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.781, + 0.486, + 0.902 + ], + "angle": 0, + "content": "Despite the rapid advancements in 3D object reconstruction, one question remains: does the reconstruction quality truly meet the standard of a digital twin, where virtual representations are indistinguishable from reality? This digital twin standard demands both highly accurate shape matching and photorealistic appearance across different lighting, which present significant acquisition challenges for real-world objects. Existing object-centric datasets for 3D recon" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.391, + 0.909, + 0.543 + ], + "angle": 0, + "content": "struction or inverse rendering have focused on either dataset size [18] or quality of specific aspects [16, 21, 34, 46, 61, 73], often sacrificing comprehensive fidelity and limiting their application scope. This trade-off has led to a lack of datasets that fully satisfy the digital twin criteria, hindering current 3D reconstruction methods from achieving digital twin fidelity. To bridge this gap, we developed the Digital Twin Catalog (DTC) dataset, comprising 2,000 scanned 3D object models (Fig. 1(a)), each with millimeter geometry accuracy and photorealistic PBR materials (Fig. 1(b), Fig. 2)." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.548, + 0.911, + 0.791 + ], + "angle": 0, + "content": "In addition to 3D digital twin models, the DTC dataset includes evaluation data designed to support 3D object reconstruction research. This evaluation data features multiview image sequences with precise foreground object masks and environment lighting information for relighting evaluation. Traditionally, high-quality HDR images captured with modern DSLR cameras have been the standard for 3D reconstruction research. Looking ahead, we encourage the integration of 3D reconstruction research with next-generation human-centric computing platforms, such as egocentric AR glasses, aiming to democratize 3D reconstruction techniques and empower everyone to effortlessly create 3D digital twins. To this end, alongside DSLR-captured evaluation data (Fig. 1(c)), the DTC dataset also provides egocentric evaluation data captured using Project Aria glasses (https://www.projectaria.com) (Fig. 1(d))." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.795, + 0.912, + 0.903 + ], + "angle": 0, + "content": "The DTC dataset offers extensive opportunities for advancing research in object digital twin creation. We provide a benchmark for state-of-the-art 3D object reconstruction and inverse rendering methods. These benchmarks evaluate performance across novel view synthesis (NVS), shape reconstruction, and relightable appearance reconstruction. We further provide the evaluation of novel view synthesis meth" + }, + { + "type": "page_number", + "bbox": [ + 0.486, + 0.945, + 0.511, + 0.956 + ], + "angle": 0, + "content": "754" + } + ], + [ + { + "type": "table_caption", + "bbox": [ + 0.09, + 0.089, + 0.907, + 0.119 + ], + "angle": 0, + "content": "Table 1. Comparison with existing object-centric inverse rendering datasets. *Objaverse [18] consists of both synthetic objects and real scans, only part of which contain PBR materials." + }, + { + "type": "table", + "bbox": [ + 0.095, + 0.13, + 0.903, + 0.371 + ], + "angle": 0, + "content": "
Dataset# ObjectsRealScene TypeMulti-viewShapePBR Mat.Relit ImageLightingEgocentric Cap.
ShapeNet-Intrinsics [61]31KXsynthetic✓✓X✓✓X
NeRD Synthetic [11]3Xsynthetic✓✓✓✓✓X
ABO [16]8KXsynthetic✓✓✓✓✓X
MIT Intrasins [25]20✓studio✓XXXXX
DTU-MVS [34]80✓studio✓✓XXXX
Objaverse [18]818K(✓)*studio✓✓(✓)*XXX
DiLiGenT-MV [42]5✓studio✓✓XX✓X
ReNe [65]20✓studio✓XXX✓X
OpenIllumination [46]64✓studio✓X✓✓✓X
GSO [21]1030✓studio✓✓XXXX
Lombardi et al. [47]6✓in-the-wildX✓X✓✓X
NeRD Real [11]4✓in-the-wild✓X✓✓XX
NeROIC [38]3✓in-the-wild✓X✓✓XX
Oxholm et al. [56]4✓in-the-wild✓✓X✓✓X
OmniObject3D [73]6k✓in-the-wild✓✓XXXX
Stanford Orb [39]14✓in-the-wild✓✓✓✓✓X
DTC (ours)2k✓in-the-wild✓✓✓✓✓✓
" + }, + { + "type": "text", + "bbox": [ + 0.089, + 0.397, + 0.483, + 0.503 + ], + "angle": 0, + "content": "ods using the egocentric aligned DTC data. Additionally, we explore the dataset's potential in downstream robotics applications by assessing its effectiveness in training robotic policies for pushing and grasping tasks in simulation. These benchmarks and applications provide valuable insights, highlight existing challenges, and uncover promising directions for future research in 3D digital twin creation." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.522, + 0.233, + 0.537 + ], + "angle": 0, + "content": "2. Related Work" + }, + { + "type": "text", + "bbox": [ + 0.089, + 0.548, + 0.483, + 0.655 + ], + "angle": 0, + "content": "We provide a comparison of our DTC dataset to existing object-centric datasets in Table 1. We provide the largest 3D dataset with PBR materials and real world multi-view recordings with digital twin counterparts. We further provide digital twin aligned egocentric recordings, the first of their kind in the egocentric domain. We will discuss the related datasets and methods they can empower as follows." + }, + { + "type": "text", + "bbox": [ + 0.089, + 0.659, + 0.483, + 0.901 + ], + "angle": 0, + "content": "3D Digital Twin Datasets Existing 3D digital twin datasets with PBR materials often serve as ground truth for evaluating inverse rendering results. Early efforts [6, 25] provide small-scale intrinsic image of real objects and do not provide shape or PBR material information. Synthetic datasets [11, 16, 45, 61, 70, 72] are widely used for evaluation but do not represent the complexity in a real world environment. For datasets that contain real objects, [18, 34, 42, 47, 56, 65], the reconstruction quality can vary, which leads to imprecise evaluations. Table 1 provides a comparison to the previous work in this domain. Compared to Objaverse[18, 19], which is a collection of existing 3D models with only a small subset containing PBR materials with varying quality, we offer a high quality collection of 3D object data that is also aligned with real world recordings. Compared to OmniObject3D [73], the DTC models" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.397, + 0.908, + 0.593 + ], + "angle": 0, + "content": "provide higher-quality shape and additional PBR materials that are necessary for high quality inverse rendering. We offer the largest quantity of 3D object models compared to all counterparts in various tasks. For real-world evaluation, Stanford-ORB [39] was the prior largest inverse rendering benchmark with in-the-wild lighting. In contrast, we provide more object diversity and higher quality for each object model. The Aria Digital Twin dataset [58] was the first dataset to provide digital twin aligned environments for the scenes and recorded using egocentric device. However, their scene environments are limited and the contained object ground truths inside do not have high quality geometries with PBR materials." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.598, + 0.909, + 0.794 + ], + "angle": 0, + "content": "Object Reconstruction & Inverse Rendering. Using object-centric multi-view images as input, early object reconstruction methods focused on estimating individual object properties, such as shape from shading [5, 9, 83], material acquisition [44, 45, 56, 57, 74], and lighting estimation [67, 78]. Some approaches also aimed to recover reflectance and illumination assuming known object shapes [47, 48]. Inverse rendering, which seeks to invert the rendering equation [36], estimates an image's intrinsic components—geometry, materials, and lighting. The advent of differentiable renderers [14, 15, 49] enabled full-fledged inverse rendering methods to simultaneously recover all these properties for object reconstruction [50]." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.796, + 0.909, + 0.903 + ], + "angle": 0, + "content": "Neural volumetric representations such as Neural Radiance Fields (NeRFs) [52] and the like [7, 11-13, 59, 79, 84] encode geometry and appearance as volumetric densities and radiance with a Multi-Layer Perceptron (MLP) network, and render images using the volume rendering equation [51]. 3D-GS [37] introduces 3D Gaussian primitives and rasterization and its following-up variants [31] demonstrates high quality" + }, + { + "type": "page_number", + "bbox": [ + 0.486, + 0.945, + 0.511, + 0.957 + ], + "angle": 0, + "content": "755" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.091, + 0.092, + 0.281, + 0.106 + ], + "angle": 0, + "content": "geometry prediction as well." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.107, + 0.484, + 0.304 + ], + "angle": 0, + "content": "Other surface-based representations [29, 53, 62, 66, 71, 80, 81, 85] extract surfaces as the zero level set, for instance, of a signed distance function (SDF) or an occupancy field [55], allowing them to efficiently model the appearance on the surface with an explicit material model, such as bidirectional reflectance distribution functions (BRDFs). This also enables modeling more complex global illumination effects, such as self-shadows. Most of these methods focus on per-scene optimization and require dense multiple views as input. Recently, researchers have incorporated learning-based models, distilling priors from large training datasets for fast inference on limited test views [8, 10, 33, 45, 61, 70, 72, 86]." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.304, + 0.484, + 0.365 + ], + "angle": 0, + "content": "In this work, we provide the DTC dataset with real-world object recordings that can serve as the benchmark to evaluate object-centric inverse rendering tasks. We evaluate representative baselines from existing work." + }, + { + "type": "title", + "bbox": [ + 0.09, + 0.38, + 0.486, + 0.416 + ], + "angle": 0, + "content": "3. Digital Twin Catalog: A Large-Scale Photorealistic 3D Object Digital Twin Dataset" + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.424, + 0.288, + 0.44 + ], + "angle": 0, + "content": "3.1. Dataset Composition" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.446, + 0.484, + 0.552 + ], + "angle": 0, + "content": "Our DTC dataset contains: (1) 2,000 scanned 3D object models, featuring millimeter geometric accuracy relative to their physical counterparts, along with a full set of photorealistic PBR material maps (Fig. 1), (2) 100 DSLR-captured evaluation data of 50 objects under different lighting conditions, and (3) 200 egocentric Aria-captured evaluation data of 100 objects with both active and casual observation modes." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.561, + 0.357, + 0.577 + ], + "angle": 0, + "content": "3.2. Creation of 3D Object Models" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.583, + 0.484, + 0.795 + ], + "angle": 0, + "content": "Utilizing the state-of-the-art industrial 3D object scanner [1], we selected 2,000 physical-world objects spanning 40 LVIS [27] categories, carefully chosen to ensure both category diversity and compatibility with the scanner's capabilities. As illustrated in Fig. 3(a) (c), the scanner features a fixed lighting-camera setup within an upper-hemisphere dome, equipped with 8 structured lights for geometry scanning, and 29 spotlights and 29 cameras for material acquisition. During the scanning process, the object is placed on a central holder, and its pose can be adjusted with multi-round scanning to achieve a complete 360-degree scan. For our dataset, each object typically undergoes three pose changes, with a total scanning time of approximately 20 minutes per object." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.796, + 0.484, + 0.902 + ], + "angle": 0, + "content": "After scanning, a proprietary post-processing pipeline reconstructs both the geometry and the PBR material maps. For 4K-resolution PBR material maps, the post-processing requires approximately 4 hours per object. In terms of quality, the structured-light-based shape reconstruction (Fig. 3 (b)) in the post-processing achieves millimeter-level geometric accuracy. However, the material optimization process" + }, + { + "type": "image", + "bbox": [ + 0.518, + 0.09, + 0.712, + 0.237 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.518, + 0.237, + 0.712, + 0.286 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.518, + 0.286, + 0.772, + 0.405 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.712, + 0.091, + 0.903, + 0.238 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.712, + 0.237, + 0.903, + 0.286 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.561, + 0.414, + 0.857, + 0.43 + ], + "angle": 0, + "content": "Figure 3. 3D object scanner by Covision Media®." + }, + { + "type": "image", + "bbox": [ + 0.518, + 0.445, + 0.706, + 0.611 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.708, + 0.445, + 0.905, + 0.611 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.518, + 0.612, + 0.706, + 0.773 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.706, + 0.612, + 0.905, + 0.773 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.538, + 0.784, + 0.88, + 0.798 + ], + "angle": 0, + "content": "Figure 4. Rendered DTC models (left) v.s. Photo (Right)." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.826, + 0.906, + 0.901 + ], + "angle": 0, + "content": "performs best for diffuse objects and often struggles with glossy or shiny surfaces. To address this limitation, we hired technical artists to develop a workflow to refine materials for glossy and shiny objects, ensuring that the material quality meets the standards of a digital twin." + }, + { + "type": "page_number", + "bbox": [ + 0.487, + 0.946, + 0.511, + 0.957 + ], + "angle": 0, + "content": "756" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.094, + 0.088, + 0.483, + 0.226 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.09, + 0.235, + 0.483, + 0.265 + ], + "angle": 0, + "content": "Figure 5. Shape and material (albedo) quality comparison between Stanford-ORB [39] (left) and our DTC (right)." + }, + { + "type": "text", + "bbox": [ + 0.089, + 0.292, + 0.483, + 0.442 + ], + "angle": 0, + "content": "3D Model Accuracy. To validate the material and geometry accuracy of the 3D models, we compared a rendered image of our scanned and processed model with a photograph of the same object taken inside a light box. A virtual light box was meticulously modeled to replicate the light intensity and color temperature of the real light box. The scanned object was then placed in the virtual light box to generate the rendered image. The side-by-side comparison demonstrates a remarkable match between the rendered and real images (Fig. 4)." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.447, + 0.484, + 0.522 + ], + "angle": 0, + "content": "Comparison Against Stanford-ORB. We also scanned the objects used in Stanford-ORB [39] to compare the shape and appearance quality. As illustrated in Fig. 5, the Stanford-ORB models exhibit shape artifacts and noisy, lower-quality materials compared to our models." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.527, + 0.306, + 0.542 + ], + "angle": 0, + "content": "3.3. DSLR Evaluation Data" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.549, + 0.483, + 0.656 + ], + "angle": 0, + "content": "Within DTC, we include a DSLR-captured evaluation dataset of 50 objects from Sec. 3.2 captured under two different lighting conditions, resulting in 100 distinct image sequences. For every sequence, we provide (a) approximately 120 HDR and LDR images from different viewing directions, (b) one object pose and (c) per-image camera pose. The two lighting conditions are represented using two environment maps." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.66, + 0.484, + 0.901 + ], + "angle": 0, + "content": "Data Capture. To ensure the DSLR evaluation data quality, we designed and built a DSLR camera rig to automate the capture process (Fig. 6). The rig is designed to rotate the cameras around the centralized object, assuming the environment lighting remains unchanged during the capture. It features a motorized rotary stage with a centrally mounted stationary platform. Attached to the rotary stage is an extrusion frame that forms the gantry arm, supported by a set of castors to bear its weight and enable smooth rotation around the central axis. The extrusion frame is equipped with adjustable camera mounts, allowing DSLR cameras to be positioned flexibly to optimize the capture setup. For our capture process, we utilized three DSLR cameras to perform a 360-degree rotation around the object, capturing images at 9-degree intervals, resulting in 120 photos per object. To ensure precise camera pose estimation, a ChArUco board" + }, + { + "type": "image", + "bbox": [ + 0.517, + 0.088, + 0.906, + 0.234 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.561, + 0.244, + 0.858, + 0.258 + ], + "angle": 0, + "content": "Figure 6. DSLR rig for capturing evaluation data." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.286, + 0.907, + 0.314 + ], + "angle": 0, + "content": "was placed beneath the object during the capture. Example images from this setup are shown in Fig. 6." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.314, + 0.908, + 0.508 + ], + "angle": 0, + "content": "Environment Maps. Following a similar approach in Stanford-ORB [39], we capture the two environment maps using chrome ball images obtained with the capture rig described earlier. With precise camera poses provided by a ChArUco board placed beneath the chrome ball, we first fit a synthetic 3D sphere to the chrome ball by optimizing its 3D position using a geometry-friendly differentiable renderer [40, 75]. Subsequently, using a differentiable Monte Carlo-based renderer [32], we refine the environment map to match the reflection on the chrome ball, employing the single-view light estimation method proposed in [78]. The coordinate system of the environment map is determined by the ChArUco detection." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.513, + 0.909, + 0.74 + ], + "angle": 0, + "content": "Pose Registration for Camera and Object. We first obtain the initial camera poses using the ChArUco board. In most cases, for images captured from the top and middle views, the pose estimates are typically accurate. However, for bottom-view images, inaccuracies arise due to pattern distortion at grazing angles. To mitigate this issue, we refine the camera poses by fitting the rendering of a virtual ChArUco board to the real captured photo using a differentiable renderer [32, 40, 75]. Once the camera poses have been accurately refined, we optimize the object pose by minimizing the mask loss between the rendered mask and the reference one generated by [62]. By providing separate poses for camera and object anchored by ChArUco, the cameras and object are automatically aligned with the optimized environment maps." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.75, + 0.763, + 0.764 + ], + "angle": 0, + "content": "3.4. Egocentric Evaluation Data" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.765, + 0.909, + 0.901 + ], + "angle": 0, + "content": "We include an evaluation dataset of real world recordings captured by an egocentric device paired with 100 objects from the DTC dataset. We capture the egocentric recordings using the open-sourced Project Aria device [22] and acquire the additional 3D ground truth using its machine perception tool, which includes online device calibration, device trajectory and semi-dense point clouds for each recording. Each recording contains a single 3D object and precisely aligned object poses in the Aria trajectory coordinate frame." + }, + { + "type": "page_number", + "bbox": [ + 0.487, + 0.946, + 0.511, + 0.957 + ], + "angle": 0, + "content": "757" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.159, + 0.088, + 0.852, + 0.391 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.089, + 0.401, + 0.908, + 0.444 + ], + "angle": 0, + "content": "Figure 7. The workflow to align egocentric video and 3D objects. We acquire the object's semi-dense point cloud from the egocentric recording and the neural reconstructed mesh using Neural-PBIR [62]. Finally, we align the 3D object with the rendered mask from the neural reconstructed mesh using physics-based differentiable rendering (PBDR)." + }, + { + "type": "text", + "bbox": [ + 0.089, + 0.47, + 0.483, + 0.516 + ], + "angle": 0, + "content": "Both the object and the trajectory are in metric scale. We can derive additional ground truth for each video from the aligned 3D object properties." + }, + { + "type": "text", + "bbox": [ + 0.093, + 0.519, + 0.485, + 0.822 + ], + "angle": 0, + "content": "Data Capture. To feature real world recordings observed from human perspectives that can be representative of 3D object reconstruction for AR/MR devices or robots, we provide two types of recording trajectories for selected objects, termed active and passive respectively, collected by human wearers. The active recording features a complete 360 view of the objects, which is similar to existing object 3D novel view synthesis dataset. The passive recording features casual looks from human wearer, which only contain partial views of the object from certain viewing angles along the trajectory. In dataset creation, we collected the active and passive recordings in the same environment and generate their 3D information in the shared 3D space. This helps reduces potential failures when aligning the object to passive 3D recordings, which are shorter and contains less 3D information. To reduce the effect of noises and motion blur, which are common in egocentric videos in indoor low light environments, we light the capture environment to \\(3\\mathrm{K}+\\) lux illumination and used a fixed low-exposure and gain profile to collect each recording with the appropriate brightness." + }, + { + "type": "text", + "bbox": [ + 0.089, + 0.826, + 0.486, + 0.903 + ], + "angle": 0, + "content": "Alignment between Egocentric Video and Object. We provide an illustration of the object alignment to the egocentric video in Fig. 7. Given the images, camera poses and semi-dense point cloud acquired from Project Aria tools, we employ a neural-based mesh reconstruction method [62] to" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.47, + 0.909, + 0.637 + ], + "angle": 0, + "content": "create reference meshes for generating high-quality reference masks and used that to align with the corresponding 3D object mesh. This alignment step serves as an initialization for a more precise pose refinement, which leverages differentiable rendering. For certain objects with symmetric geometry, we observe this process can introduce ambiguities in point cloud registration and subsequent failures. To address such cases, we provide a GUI to manually align and correct the object alignment. Finally, akin to the DSLR camera pose registration phase, we optimize a mask loss over object poses to achieve fine-tuned pose registration." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.651, + 0.81, + 0.669 + ], + "angle": 0, + "content": "4. Benchmarking and Applications" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.671, + 0.909, + 0.838 + ], + "angle": 0, + "content": "We first use our DLSR and egocentric dataset as a benchmark for existing state-of-the-art methods. For inverse rendering, we design metrics to evaluate the shape and material quality of the recovered 3D object digital twin from three perspectives. For egocentric recording, we evaluate the novel-view synthesis as the initial evaluation. We include additional tasks, e.g. sparse view reconstruction in the supplementary materials for both DSLR and egocentric recordings. Finally we can demonstrate our high quality 3D digital twin models can be beneficial to robotics domain using an application in robotics manipulation." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.848, + 0.885, + 0.865 + ], + "angle": 0, + "content": "4.1. Application to Inverse Rendering for DSLR" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.871, + 0.908, + 0.903 + ], + "angle": 0, + "content": "The DSLR dataset in DTC provides accurate ground truth, including poses, lighting and 3D models, for inverse rendering" + }, + { + "type": "page_number", + "bbox": [ + 0.487, + 0.946, + 0.511, + 0.957 + ], + "angle": 0, + "content": "758" + } + ], + [ + { + "type": "table_caption", + "bbox": [ + 0.091, + 0.089, + 0.904, + 0.104 + ], + "angle": 0, + "content": "Table 2. Benchmark comparison of existing methods on inverse rendering for DSLR. Depth SI-MSE and Shape Chamfer distance \\( \\times {10}^{-3} \\) ." + }, + { + "type": "table", + "bbox": [ + 0.095, + 0.115, + 0.905, + 0.226 + ], + "angle": 0, + "content": "
GeometryNovel Scene RelightingNovel View Synthesis
Depth↓Normal↓Shape↓PSNR-H↑PSNR-L↑SSIM↑LPIPS↓PSNR-H↑PSNR-L↑SSIM↑LPIPS↓
Neural-PIL [12]5.710.2525.02N/A28.4235.760.8820.096
PhySG [80]0.310.1611.3127.2832.860.9590.04928.5434.460.9640.045
NVDiffRec [53]0.020.071.6426.9933.270.9510.03728.9534.920.9670.029
NeRD [11]4.550.45108.2026.1032.600.9480.06126.8033.400.8820.102
InvRender [71]0.220.030.7529.5235.980.9610.03731.6437.820.9700.033
NVDiffRecMC [29]0.020.061.3427.7834.550.9520.04231.2738.170.9720.032
" + }, + { + "type": "table_caption", + "bbox": [ + 0.115, + 0.248, + 0.459, + 0.263 + ], + "angle": 0, + "content": "Table 3. Benchmark on the egocentric aligned recordings." + }, + { + "type": "table", + "bbox": [ + 0.095, + 0.273, + 0.481, + 0.326 + ], + "angle": 0, + "content": "
PSNR ↑LPIPS ↓SSIM ↑Depth ↓Normal ↓
3D-GS [37]28.810.0200.98880.17680.3301
2D-GS [31]28.750.0200.98860.17550.2112
" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.352, + 0.484, + 0.442 + ], + "angle": 0, + "content": "tasks and serves as an evaluation suite to benchmark the performance of inverse rendering methods. We select six prior methods for this task and evaluate their performance using the ground truth provided by our dataset. In the following sections, we describe the data splitting strategy, evaluation metrics, and baselines." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.446, + 0.484, + 0.552 + ], + "angle": 0, + "content": "Data Splitting. For benchmarking purposes, we select 15 objects from the DSLR dataset captured under two distinct lighting environments, resulting in a total of 30 image sequences. The selected objects encompass a diverse range of geometric and material properties to ensure a comprehensive evaluation. For each scene, 8 views are selected for testing, while the remaining views are reserved for training." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.555, + 0.484, + 0.737 + ], + "angle": 0, + "content": "Evaluation Metrics. The metrics measure the accuracy of three aspects of baseline performance: geometry estimation, relighting, and novel view synthesis. For geometry estimation, we evaluate the accuracy of predicted depth and normal maps under held-out test views, as well as 3D meshes extracted from baseline methods, compared with the ground truth from our dataset. Relighting metrics evaluate the material decomposition quality of baselines by measuring the accuracy of predicted images under held-out lighting conditions. For view synthesis, we compare the predicted images from viewpoints unseen during training to ground truth captures. We refer to Kuang et al. [39] for metric details." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.74, + 0.484, + 0.846 + ], + "angle": 0, + "content": "Baselines. We include the following baselines: NVDiffRec [53] and NVDiffRecMC [29], with a hybrid shape representation DMTet [60]; InvRender [85] and PhySG [80], which adopt signed distance functions (SDFs) to represent object geometry [76] and utilize implicit neural fields for material decomposition; Neural-PIL [12] and NeRD [11], which use NeRFs [52] as scene representations." + }, + { + "type": "title", + "bbox": [ + 0.09, + 0.855, + 0.445, + 0.871 + ], + "angle": 0, + "content": "4.2. Application to Egocentric Reconstruction" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.871, + 0.483, + 0.901 + ], + "angle": 0, + "content": "Our digital twin models, aligned with real world video using the method described in Sec. 3.4, can help obtain accurate" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.25, + 0.907, + 0.401 + ], + "angle": 0, + "content": "ground truth for object-centric images that were previously difficult to acquire. We provide the first evaluation of object-centric novel view synthesis recorded from an egocentric device. We use the projected object shape given the 3D pose of the object and cameras in scene coordinates to acquire the image masks, depth and normal for each object. We selected 15 recordings from the egocentric recording sessions as the evaluation and used the active recordings to benchmark novel view reconstruction. For each recording, we hold out every 8th image view as a testing view." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.405, + 0.909, + 0.572 + ], + "angle": 0, + "content": "**Evaluations.** We build our baselines based on the gsplats [77] implementation of the 3D Gaussian Splitting (GS) [37] and 2D GS [31], and handle the effect of lens shading from the Project Aria lens [26]. We calculate PSNR, depth and normal based on the observed objects with masks and provide SSIM and LPIPS score on images by masking out the non-object areas as black. Table 3 shows the benchmark results of the baselines. We use the same depth and normal metric in DSLR evaluation. We provide additional qualitative evaluations and analyses on egocentric data towards sparser view settings in the supplementary materials." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.583, + 0.833, + 0.6 + ], + "angle": 0, + "content": "4.3. Application to Robotic Manipulation" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.606, + 0.907, + 0.741 + ], + "angle": 0, + "content": "High-quality object models have been leveraged in prior work to train real-world robotic agents in scenes represented explicitly [54] or implicitly [64]. These object models have also been shown to facilitate object-centric pose and lighting parameter estimation, enabling model-based planning [63]. In this section, we empirically evaluate the effectiveness of using DTC dataset objects in training robotic policies. Specifically, we consider learning robotic pushing and grasping skills in simulation." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.743, + 0.909, + 0.864 + ], + "angle": 0, + "content": "First, we sample a subset of 24 cup category objects from the DTC dataset and 24 cup objects from Objaverse-XL [19] \\( {}^{1} \\) . Since not all Objverse-XL objects come with textures, we randomize the colors of those objects uniformly in RGB space. To compute collision meshes for physical simulation, we perform convex decomposition on each object with CoACD [68]. We import these objects along with a UR5e robot equipped with a Robotiq 2F-85 (pushing) or" + }, + { + "type": "page_footnote", + "bbox": [ + 0.513, + 0.875, + 0.907, + 0.9 + ], + "angle": 0, + "content": "1The version of Objaverse-XL used in this work excludes all 3D models sourced from Sketchfab." + }, + { + "type": "page_number", + "bbox": [ + 0.487, + 0.946, + 0.511, + 0.957 + ], + "angle": 0, + "content": "759" + } + ], + [ + { + "type": "table", + "bbox": [ + 0.11, + 0.089, + 0.468, + 0.175 + ], + "angle": 0, + "content": "
TaskDTC (ours)Objverse-XL [19]
Pushing @ 2cm36.3% ± 1.5%25.3% ± 6.0%
Pushing @ 3cm43.7% ± 1.2%29.7% ± 6.0%
Pushing @ 5cm47.0% ± 2.6%40.3% ± 5.5%
Grasping42.7% ± 4.7%38.6% ± 11.0%
" + }, + { + "type": "table_caption", + "bbox": [ + 0.09, + 0.185, + 0.483, + 0.242 + ], + "angle": 0, + "content": "Table 4. Success rate of policies trained on data collected using objects from our DTC dataset and sampled from Objaverse-XL when evaluated on an unseen test object. Errors indicate sample standard deviation over three policy training seeds." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.281, + 0.484, + 0.358 + ], + "angle": 0, + "content": "Robotiq 2F-140 (grasping) gripper into the PyBullet simulator [17] and collect data for each robotic task as described below. After training policies on data from each object set, we evaluate policy performance on a relatively high-quality unseen test object from the StanfordORB dataset [39]." + }, + { + "type": "text", + "bbox": [ + 0.089, + 0.367, + 0.483, + 0.563 + ], + "angle": 0, + "content": "Pushing. For the pushing task, we collect 5000 trajectories of pseudo-random robotic interaction data for each object set. For each trajectory, a single object from the considered object set and its initial position are randomly selected. Then we train a goal-conditioned neural network policy \\(\\pi (a|o,o_g)\\) where \\(o\\) is an image observation of the current scene and \\(o_{g}\\) is a goal image indicating the desired final object and robot position, \\(o,o_{g}\\in \\mathbb{R}^{256\\times 256\\times 3}\\), and the action \\(a\\in \\mathbb{R}^2\\) represents a change in the robot's end-effector position in the \\(x\\) and \\(y\\) axes. The \\(z\\) axis end-effector height is held fixed. We sample goals for training via hindsight relabeling [4, 20, 24]. We then perform evaluation on 100 randomly sampled test goals manipulating an unseen test cup from StanfordORB [39]." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.574, + 0.483, + 0.726 + ], + "angle": 0, + "content": "Grasping. For grasping, we collect 5000 successful grasp examples for each object set by first placing a single object into the scene, randomizing the object identity and initial position. We then randomly sample candidate grasp poses in a radius around the object's position and simulating their outcomes, rejecting unsuccessful grasps. We train a grasping policy \\(\\pi(a|o)\\) where \\(o \\in \\mathbb{R}^{256 \\times 256 \\times 3}\\) is an image observation of the scene and \\(a \\in \\mathbb{R}^4\\) represents the \\(x, y, z\\) position and \\(\\theta\\) yaw rotation of the robot end-effector at which to attempt the grasp. Again we use 100 test cup object poses." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.735, + 0.483, + 0.902 + ], + "angle": 0, + "content": "Results. We report the results in Table 4 and Fig. 8. We find that across both pushing and grasping tasks, policies trained on DTC dataset objects outperform those trained on Objaverse-XL objects when evaluated on the unseen test object. For pushing, we report performance by defining binary success thresholds based on the final Euclidean distance of the object position to the goal position. Training on DTC objects appears to be especially helpful at enabling policies to make finer adjustments, improving pushing success rates at stricter thresholds. Additional experimental details can be found in the supplementary." + }, + { + "type": "image", + "bbox": [ + 0.518, + 0.092, + 0.903, + 0.295 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.513, + 0.307, + 0.907, + 0.378 + ], + "angle": 0, + "content": "Figure 8. Success rates on robotic pushing task when training with our DTC objects and sampled Objaverse-XL objects. Particularly at lower error thresholds, policies trained on DTC objects outperform those trained using Objaverse-XL objects. Shaded bars represent sample standard deviations over policy training random seeds." + }, + { + "type": "title", + "bbox": [ + 0.514, + 0.384, + 0.634, + 0.4 + ], + "angle": 0, + "content": "5. Conclusion" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.41, + 0.907, + 0.546 + ], + "angle": 0, + "content": "We presented a new large scale photorealistic 3D digital twin dataset with the real world recordings that contain its real world counterpart. We provide extensive evaluations of baselines on our DTC dataset serving as new benchmark for inverse rendering and novel view synthesis task. We also demonstrated that high quality digital twin models can be beneficial to applications in robotics domain. We believe our efforts can empower the research community to build and leverage digital twin models for future applications." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.549, + 0.909, + 0.656 + ], + "angle": 0, + "content": "Limitations. Achieving high quality digital twin models currently requires deliberate hardware setup and human efforts in refinement. Solving this challenge without sacrificing quality can significantly further enhance the volume of digital twin models. Our hardware is also limited to objects within a certain size and can not yet recover objects that are deformable, highly specular, or transparent." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.659, + 0.909, + 0.825 + ], + "angle": 0, + "content": "Future work. The existing digital twin model creation in DTC dataset involves lengthy post-processing and may require subjective human refinement, hindering the automation of model generation. However, recent advancements in physics-based differentiable rendering hold promise for enabling faster and more accurate creation of digital twins, especially for material reconstruction. Furthermore, building large-scale digital twins for applications will necessitate efforts to enhance the diversity in object appearance (e.g., transparent objects) and to capture additional attributes, such as physical properties and functionalities." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.826, + 0.909, + 0.9 + ], + "angle": 0, + "content": "For robotics applications, while deploying manipulation policies learned in simulation to the real world remains generally challenging, we hope that the high-quality digital twin data provided by DTC can serve as a stepping stone towards effective sim-to-real transfer." + }, + { + "type": "page_number", + "bbox": [ + 0.487, + 0.946, + 0.511, + 0.957 + ], + "angle": 0, + "content": "760" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.09, + 0.092, + 0.486, + 0.155 + ], + "angle": 0, + "content": "Acknowledgments. This work is in part supported by NSF CCRI #2120095, RI #2211258, and RI #2338203. YZ is in part supported by the Stanford Interdisciplinary Graduate Fellowship." + }, + { + "type": "title", + "bbox": [ + 0.093, + 0.169, + 0.188, + 0.184 + ], + "angle": 0, + "content": "References" + }, + { + "type": "ref_text", + "bbox": [ + 0.101, + 0.194, + 0.486, + 0.221 + ], + "angle": 0, + "content": "[1] Covision 3D object scanner. https://www.covisionmedia.ai/en/scanners.4" + }, + { + "type": "ref_text", + "bbox": [ + 0.101, + 0.224, + 0.484, + 0.293 + ], + "angle": 0, + "content": "[2] Kosmas Alexopoulos, Nikolaos Nikolakis, and George Chryssolouris. Digital twin-driven supervised machine learning for the development of artificial intelligence applications in manufacturing. International Journal of Computer Integrated Manufacturing, 33(5):429-439, 2020. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.296, + 0.484, + 0.363 + ], + "angle": 0, + "content": "[3] Moayad Aloqaily, Ouns Bouachir, Fakhri Karray, Ismaeel Al Ridhawi, and Abdulmotaleb El Saddik. Integrating digital twin and advanced intelligent technologies to realize the meta-verse. IEEE Consumer Electronics Magazine, 12(6):47-55, 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.367, + 0.484, + 0.422 + ], + "angle": 0, + "content": "[4] Marcin Andrychowicz, Filip Wolski, Alex Ray, Jonas Schneider, Rachel Fong, Peter Welinder, Bob McGrew, Josh Tobin, OpenAI Pieter Abbeel, and Wojciech Zaremba. Hindsight experience replay. In NeurIPS, 2017. 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.424, + 0.485, + 0.465 + ], + "angle": 0, + "content": "[5] Jonathan T Barron and Jitendra Malik. Shape, illumination, and reflectance from shading. IEEE TPAMI, 37(8):1670-1687, 2014. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.469, + 0.482, + 0.495 + ], + "angle": 0, + "content": "[6] Sean Bell, Kavita Bala, and Noah Snavely. Intrinsic images in the wild. ACM TOG, 33(4):1-12, 2014. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.498, + 0.484, + 0.553 + ], + "angle": 0, + "content": "[7] Sai Bi, Zexiang Xu, Kalyan Sunkavalli, Miloš Hasan, Yannick Hold-Geoffroy, David Kriegman, and Ravi Ramamoorthi. Deep reflectance volumes: Relightable reconstructions from multi-view photometric images. In ECCV, 2020. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.556, + 0.484, + 0.597 + ], + "angle": 0, + "content": "[8] Sai Bi, Zexiang Xu, Kalyan Sunkavalli, David Kriegman, and Ravi Ramamoorthi. Deep 3D capture: Geometry and reflectance from sparse multi-view images. In CVPR, 2020. 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.6, + 0.482, + 0.627 + ], + "angle": 0, + "content": "[9] Martin Bichsel and Alex P Pentland. A simple algorithm for shape from shading. In CVPR, 1992. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.63, + 0.484, + 0.67 + ], + "angle": 0, + "content": "[10] Mark Boss, Varun Jampani, Kihwan Kim, Hendrik P.A. Lensch, and Jan Kautz. Two-shot Spatially-varying BRDF and Shape Estimation. In CVPR, 2020. 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.673, + 0.484, + 0.727 + ], + "angle": 0, + "content": "[11] Mark Boss, Raphael Braun, Varun Jampani, Jonathan T. Barron, Ce Liu, and Hendrik P.A. Lensch. NeRD: Neural reflectance decomposition from image collections. In ICCV, 2021. 3, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.731, + 0.484, + 0.785 + ], + "angle": 0, + "content": "[12] Mark Boss, Varun Jampani, Raphael Braun, Ce Liu, Jonathan Barron, and Hendrik Lensch. Neural-PIL: Neural pre-integrated lighting for reflectance decomposition. In NeurIPS, 2021. 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.788, + 0.482, + 0.843 + ], + "angle": 0, + "content": "[13] Mark Boss, Andreas Engelhardt, Abhishek Kar, Yuanzhen Li, Deqing Sun, Jonathan Barron, Hendrik Lensch, and Varun Jampani. SAMURAI: Shape and material from unconstrained real-world arbitrary image collections. In NeurIPS, 2022. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.846, + 0.484, + 0.9 + ], + "angle": 0, + "content": "[14] Guangyan Cai, Kai Yan, Zhao Dong, Ioannis Gkioulekas, and Shuang Zhao. Physics-based inverse rendering using combined implicit and explicit geometries. CGF, 41(4):129-138, 2022. 3" + }, + { + "type": "list", + "bbox": [ + 0.094, + 0.194, + 0.486, + 0.9 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.093, + 0.907, + 0.147 + ], + "angle": 0, + "content": "[15] Wenzheng Chen, Joey Litalien, Jun Gao, Zian Wang, Clement Fuji Tsang, Sameh Khamis, Or Litany, and Sanja Fidler. DIBR++: learning to predict lighting and material with a hybrid differentiable renderer. In NeurIPS, 2021. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.151, + 0.907, + 0.219 + ], + "angle": 0, + "content": "[16] Jasmine Collins, Shubham Goel, Kenan Deng, Achleshwar Luthra, Leon Xu, Erhan Gundogdu, Xi Zhang, Tomas F Yago Vicente, Thomas Dideriksen, Himanshu Arora, et al. ABO: Dataset and benchmarks for real-world 3D object understanding. In CVPR, 2022. 2, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.222, + 0.907, + 0.262 + ], + "angle": 0, + "content": "[17] Erwin Coumans and Yunfei Bai. Pybullet, a python module for physics simulation for games, robotics and machine learning. http://pybullet.org, 2016-2019.8" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.265, + 0.907, + 0.32 + ], + "angle": 0, + "content": "[18] Matt Deitke, Dustin Schwenk, Jordi Salvador, Luca Weihs, Oscar Michel, Eli VanderBilt, Ludwig Schmidt, Kiana Ehsani, Aniruddha Kembhavi, and Ali Farhadi. Objverse: A universe of annotated 3D objects. In CVPR, 2023. 2, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.323, + 0.907, + 0.377 + ], + "angle": 0, + "content": "[19] Matt Deitke, Ruoshi Liu, Matthew Wallingford, Huong Ngo, Oscar Michel, Aditya Kusupati, Alan Fan, Christian Laforte, Vikram Voleti, Samir Yitzhak Gadre, et al. Objverse-XL: A universe of \\(10\\mathrm{m} + 3\\mathrm{D}\\) objects. NeurIPS, 2024. 3, 7, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.38, + 0.906, + 0.42 + ], + "angle": 0, + "content": "[20] Yiming Ding, Carlos Florensa, Pieter Abbeel, and Mariano Philipp. Goal-conditioned imitation learning. In NeurIPS, 2019. 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.423, + 0.907, + 0.491 + ], + "angle": 0, + "content": "[21] Laura Downs, Anthony Francis, Nate Koenig, Brandon Kinman, Ryan Hickman, Krista Reymann, Thomas B McHugh, and Vincent Vanhoucke. Google scanned objects: A high-quality dataset of 3D scanned household items. In ICRA, 2022. 2, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.494, + 0.908, + 0.784 + ], + "angle": 0, + "content": "[22] Jakob Engel, Kiran Somasundaram, Michael Goesele, Albert Sun, Alexander Gamino, Andrew Turner, Arjang Talatof, Arnie Yuan, Bilal Souti, Brighid Meredith, Cheng Peng, Chris Sweeney, Cole Wilson, Dan Barnes, Daniel DeTone, David Caruso, Derek Valleroy, Dinesh Ginjupalli, Duncan Frost, Edward Miller, Elias Mueggler, Evgeniy Oleinik, Fan Zhang, Guruprasad Somasundaram, Gustavo Solaira, Harry Lanaras, Henry Howard-Jenkins, Huixuan Tang, Hyo Jin Kim, Jaime Rivera, Ji Luo, Jing Dong, Julian Straub, Kevin Bailey, Kevin Eckenhoff, Lingni Ma, Luis Pesqueira, Mark Schwesinger, Maurizio Monge, Nan Yang, Nick Charron, Nikhil Raina, Omkar Parkhi, Peter Borschowa, Pierre Moulon, Prince Gupta, Raul Mur-Artal, Robbie Pennington, Sachin Kulkarni, Sagar Miglani, Santosh Gondi, Saransh Solanki, Sean Diener, Shangyi Cheng, Simon Green, Steve Saarinen, Suvam Patra, Tassos Mourikis, Thomas Whelan, Tripti Singh, Vasileios Balntas, Vijay Baiyya, Wilson Dreewes, Xiaqing Pan, Yang Lou, Yipu Zhao, Yusuf Mansour, Yuyang Zou, Zhaoyang Lv, Zijian Wang, Mingfei Yan, Carl Ren, Renzo De Nardi, and Richard Newcombe. Project Aria: A new tool for egocentric multi-modal ai research, 2023. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.787, + 0.906, + 0.828 + ], + "angle": 0, + "content": "[23] Ruohan Gao, Yen-Yu Chang, Shivani Mall, Li Fei-Fei, and Jiajun Wu. ObjectFolder: A dataset of objects with implicit visual, auditory, and tactile representations. In CoRL, 2021. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.831, + 0.907, + 0.883 + ], + "angle": 0, + "content": "[24] Dibya Ghosh, Abhishek Gupta, Ashwin Reddy, Justin Fu, Coline Devin, Benjamin Eysenbach, and Sergey Levine. Learning to reach goals via iterated supervised learning. In ICML, 2019. 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.887, + 0.906, + 0.902 + ], + "angle": 0, + "content": "[25] Roger Grosse, Micah K Johnson, Edward H Adelson, and" + }, + { + "type": "list", + "bbox": [ + 0.517, + 0.093, + 0.908, + 0.902 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.487, + 0.946, + 0.511, + 0.957 + ], + "angle": 0, + "content": "761" + } + ], + [ + { + "type": "ref_text", + "bbox": [ + 0.125, + 0.092, + 0.484, + 0.12 + ], + "angle": 0, + "content": "William T Freeman. Ground truth dataset and baseline evaluations for intrinsic image algorithms. In ICCV, 2009. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.122, + 0.484, + 0.177 + ], + "angle": 0, + "content": "[26] Qiao Gu, Zhaoyang Lv, Duncan Frost, Simon Green, Julian Straub, and Chris Sweeney. Egolifter: Open-world 3d segmentation for egocentric perception. In ECCV, pages 382-400. Springer, 2024. 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.179, + 0.484, + 0.219 + ], + "angle": 0, + "content": "[27] Agrim Gupta, Piotr Dollar, and Ross Girshick. LVIS: A dataset for large vocabulary instance segmentation. In CVPR, 2019. 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.221, + 0.484, + 0.275 + ], + "angle": 0, + "content": "[28] Wassim Hamidouche, Lina Bariah, and Mérouane Debbah. Immersive media and massive twinning: Advancing toward the metaverse. IEEE Communications Magazine, 62(7):20-32, 2024. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.278, + 0.484, + 0.32 + ], + "angle": 0, + "content": "[29] Jon Hasselgren, Nikolai Hofmann, and Jacob Munkberg. Shape, light & material decomposition from images using Monte Carlo rendering and denoising. In NeurIPS, 2022. 4, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.321, + 0.484, + 0.389 + ], + "angle": 0, + "content": "[30] Yicong Hong, Kai Zhang, Jiuxiang Gu, Sai Bi, Yang Zhou, Difan Liu, Feng Liu, Kalyan Sunkavalli, Trung Bui, and Hao Tan. LRM: Large reconstruction model for single image to 3D. In The Twelfth International Conference on Learning Representations, 2024. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.391, + 0.484, + 0.445 + ], + "angle": 0, + "content": "[31] Binbin Huang, Zehao Yu, Anpei Chen, Andreas Geiger, and Shenghua Gao. 2D gaussian splatting for geometrically accurate radiance fields. In SIGGRAPH 2024 Conference Papers, 2024. 3, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.448, + 0.484, + 0.504 + ], + "angle": 0, + "content": "[32] Wenzel Jakob, Sébastien Speierer, Nicolas Roussel, Merlin Nimier-David, Delio Vicini, Tizian Zeltner, Baptiste Nicolet, Miguel Crespo, Vincent Leroy, and Ziyi Zhang. Mitsuba 3 renderer, 2022. https://mitsuba-renderer.org. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.505, + 0.484, + 0.546 + ], + "angle": 0, + "content": "[33] Michael Janner, Jiajun Wu, Tejas Kulkarni, Ilker Yildirim, and Joshua B Tenenbaum. Self-supervised intrinsic image decomposition. In NeurIPS, 2017. 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.548, + 0.484, + 0.588 + ], + "angle": 0, + "content": "[34] Rasmus Jensen, Anders Dahl, George Vogiatzis, Engin Tola, and Henrik Aanæs. Large scale multi-view stereopsis evaluation. In CVPR, 2014. 2, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.59, + 0.484, + 0.632 + ], + "angle": 0, + "content": "[35] Haian Jin, Isabella Liu, Peijia Xu, Xiaoshuai Zhang, Songfang Han, Sai Bi, Xiaowei Zhou, Zexiang Xu, and Hao Su. TensoIR: Tensorial inverse rendering. In CVPR, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.633, + 0.484, + 0.675 + ], + "angle": 0, + "content": "[36] James T Kajiya. The rendering equation. In Proceedings of the 13th annual conference on Computer graphics and interactive techniques, pages 143-150, 1986. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.676, + 0.484, + 0.717 + ], + "angle": 0, + "content": "[37] Bernhard Kerbl, Georgios Kopanas, Thomas Leimkuhler, and George Drettakis. 3D gaussian splattering for real-time radiance field rendering. ACM TOG, 2023. 2, 3, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.718, + 0.484, + 0.773 + ], + "angle": 0, + "content": "[38] Zhengfei Kuang, Kyle Olszewski, Mengelei Chai, Zeng Huang, Panos Achlioptas, and Sergey Tulyakov. NeROIC: neural rendering of objects from online image collections. ACM TOG, 41(4):1-12, 2022. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.775, + 0.484, + 0.843 + ], + "angle": 0, + "content": "[39] Zhengfei Kuang, Yunzhi Zhang, Hong-Xing Yu, Samir Agarwala, Elliott Wu, Jiajun Wu, et al. Stanford-orb: a real-world 3D object inverse rendering benchmark. In Advances in Neural Information Processing Systems Datasets and Benchmarks Track, 2023. 3, 5, 7, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.845, + 0.484, + 0.901 + ], + "angle": 0, + "content": "[40] Samuli Laine, Janne Hellsten, Tero Karras, Yeongho Seol, Jaakko Lehtinen, and Timo Aila. Modular primitives for high-performance differentiable rendering. ACM Transactions on Graphics, 39(6), 2020. 5" + }, + { + "type": "list", + "bbox": [ + 0.093, + 0.092, + 0.484, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.092, + 0.906, + 0.161 + ], + "angle": 0, + "content": "[41] Jiahao Li, Hao Tan, Kai Zhang, Zexiang Xu, Fujun Luan, Yinghao Xu, Yicong Hong, Kalyan Sunkavalli, Greg Shakhnarovich, and Sai Bi. Instant3D: Fast text-to-3D with sparse-view generation and large reconstruction model. In ICLR, 2024. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.163, + 0.906, + 0.218 + ], + "angle": 0, + "content": "[42] Min Li, Zhenglong Zhou, Zhe Wu, Boxin Shi, Changyu Diao, and Ping Tan. Multi-view photometric stereo: A robust solution and benchmark dataset for spatially varying isotropic materials. IEEE TIP, 29:4159-4173, 2020. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.22, + 0.906, + 0.302 + ], + "angle": 0, + "content": "[43] Xuanlin Li, Kyle Hsu, Jiayuan Gu, Karl Pertsch, Oier Mees, Homer Rich Walke, Chuyuan Fu, Ishikaa Lunawat, Isabel Sieh, Sean Kirmani, Sergey Levine, Jiajun Wu, Chelsea Finn, Hao Su, Quan Vuong, and Ted Xiao. Evaluating real-world robot manipulation policies in simulation. arXiv preprint arXiv:2405.05941, 2024. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.304, + 0.906, + 0.346 + ], + "angle": 0, + "content": "[44] Zhengqin Li, Kalyan Sunkavalli, and Manmohan Chandraker. Materials for masses: SVBRDF acquisition with a single mobile phone image. In ECCV, 2018. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.347, + 0.906, + 0.402 + ], + "angle": 0, + "content": "[45] Zhengqin Li, Zexiang Xu, Ravi Ramamoorthi, Kalyan Sunkavalli, and Manmohan Chandraker. Learning to reconstruct shape and spatially-varying reflectance from a single image. ACM TOG, 37(6):1-11, 2018. 3, 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.404, + 0.906, + 0.472 + ], + "angle": 0, + "content": "[46] Isabella Liu, Linghao Chen, Ziyang Fu, Liwen Wu, Haian Jin, Zhong Li, Chin Ming Ryan Wong, Yi Xu, Ravi Ramamoorthi, Zexiang Xu, and Hao Su. OpenIllumination: A multi-illumination dataset for inverse rendering evaluation on real objects. NeurIPS, 2023. 2, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.474, + 0.906, + 0.502 + ], + "angle": 0, + "content": "[47] Stephen Lombardi and Ko Nishino. Reflectance and natural illumination from a single image. In ECCV, 2012. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.503, + 0.906, + 0.531 + ], + "angle": 0, + "content": "[48] Stephen Lombardi and Ko Nishino. Reflectance and illumination recovery in the wild. IEEE TPAMI, 2015. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.532, + 0.906, + 0.574 + ], + "angle": 0, + "content": "[49] Fujun Luan, Shuang Zhao, Kavita Bala, and Zhao Dong. Unified shape and SVBRDF recovery using differentiable Monte Carlo rendering. CGF, 40(4):101-113, 2021. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.575, + 0.906, + 0.603 + ], + "angle": 0, + "content": "[50] Stephen Robert Marschner. Inverse Rendering for Computer Graphics. Cornell University, 1998. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.604, + 0.906, + 0.631 + ], + "angle": 0, + "content": "[51] Nelson Max. Optical models for direct volume rendering. IEEE TVCG, 1(2):99-108, 1995. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.633, + 0.906, + 0.688 + ], + "angle": 0, + "content": "[52] Ben Mildenhall, Pratul P Srinivasan, Matthew Tancik, Jonathan T Barron, Ravi Ramamoorthi, and Ren Ng. NeRF: Representing scenes as neural radiance fields for view synthesis. In ECCV, 2020. 2, 3, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.689, + 0.906, + 0.745 + ], + "angle": 0, + "content": "[53] Jacob Munkberg, Jon Hasselgren, Tianchang Shen, Jun Gao, Wenzheng Chen, Alex Evans, Thomas Müller, and Sanja Fidler. Extracting triangular 3D models, materials, and lighting from images. In CVPR, 2022. 4, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.746, + 0.906, + 0.814 + ], + "angle": 0, + "content": "[54] Soroush Nasiriany, Abhiram Maddukuri, Lance Zhang, Adeet Parikh, Aaron Lo, Abhishek Joshi, Ajay Mandlekar, and Yuke Zhu. Robocasa: Large-scale simulation of everyday tasks for generalist robots. In Robotics: Science and Systems (RSS), 2024. 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.816, + 0.906, + 0.871 + ], + "angle": 0, + "content": "[55] Michael Niemeyer, Lars Mescheder, Michael Oechsle, and Andreas Geiger. Differentiable volumetric rendering: Learning implicit 3D representations without 3D supervision. In CVPR, 2020. 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.873, + 0.906, + 0.901 + ], + "angle": 0, + "content": "[56] Geoffrey Oxholm and Ko Nishino. Multiview shape and reflectance from natural illumination. In CVPR, 2014. 3" + }, + { + "type": "list", + "bbox": [ + 0.517, + 0.092, + 0.906, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.487, + 0.946, + 0.512, + 0.957 + ], + "angle": 0, + "content": "762" + } + ], + [ + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.092, + 0.484, + 0.119 + ], + "angle": 0, + "content": "[57] Geoffrey Oxholm and Ko Nishino. Shape and reflectance estimation in the wild. IEEE TPAMI, 2015. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.121, + 0.484, + 0.204 + ], + "angle": 0, + "content": "[58] Xiaqing Pan, Nicholas Charron, Yongqian Yang, Scott Peters, Thomas Whelan, Chen Kong, Omkar Parkhi, Richard Newcombe, and Yuheng (Carl) Ren. Aria digital twin: A new benchmark dataset for egocentric 3D machine perception. In Proceedings of the IEEE/CVF International Conference on Computer Vision (ICCV), pages 20133-20143, 2023. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.205, + 0.484, + 0.258 + ], + "angle": 0, + "content": "[59] Viktor Rudnev, Mohamed Elgharib, William Smith, Lingjie Liu, Vladislav Golyanik, and Christian Theobalt. NeRF for outdoor scene relighting. In ECCV, pages 615-631. Springer, 2022. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.26, + 0.484, + 0.314 + ], + "angle": 0, + "content": "[60] Tianchang Shen, Jun Gao, Kangxue Yin, Ming-Yu Liu, and Sanja Fidler. Deep marching tetrahedra: a hybrid representation for high-resolution 3D shape synthesis. In NeurIPS, 2021. 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.316, + 0.484, + 0.357 + ], + "angle": 0, + "content": "[61] Jian Shi, Yue Dong, Hao Su, and Stella X Yu. Learning non-Lambertian object intrinsics across ShapeNet categories. In CVPR, 2017. 2, 3, 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.358, + 0.484, + 0.412 + ], + "angle": 0, + "content": "[62] Cheng Sun, Guangyan Cai, Zhengqin Li, Kai Yan, Cheng Zhang, Carl Marshall, Jia-Bin Huang, Shuang Zhao, and Zhao Dong. Neural-PBIR reconstruction of shape, material, and illumination. In ICCV, 2023. 2, 4, 5, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.413, + 0.484, + 0.468 + ], + "angle": 0, + "content": "[63] Stephen Tian, Yancheng Cai, Hong-Xing Yu, Sergey Zakharov, Katherine Liu, Adrien Gaidon, Yunzhu Li, and Ji-jun Wu. Multi-object manipulation via object-centric neural scattering functions. In CVPR, 2023. 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.469, + 0.484, + 0.524 + ], + "angle": 0, + "content": "[64] Marcel Torne, Anthony Simeonov, Zechu Li, April Chan, Tao Chen, Abhishek Gupta, and Pulkit Agrawal. Reconciling reality through simulation: A real-to-sim-to-real approach for robust manipulation. Arxiv, 2024. 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.525, + 0.484, + 0.58 + ], + "angle": 0, + "content": "[65] Marco Toschi, Riccardo De Matteo, Riccardo Spezialetti, Daniele De Gregorio, Luigi Di Stefano, and Samuele Salti. Relight my NeRF: A dataset for novel view synthesis and relighting of real world objects. In CVPR, 2023. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.58, + 0.484, + 0.635 + ], + "angle": 0, + "content": "[66] Peng Wang, Lingjie Liu, Yuan Liu, Christian Theobalt, Taku Komura, and Wenping Wang. NeuS: Learning neural implicit surfaces by volume rendering for multi-view reconstruction. In NeurIPS, 2021. 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.636, + 0.484, + 0.676 + ], + "angle": 0, + "content": "[67] Henrique Weber, Donald Prévost, and Jean-François Lalonde. Learning to estimate indoor lighting from 3D objects. In 3DV, 2018. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.678, + 0.484, + 0.718 + ], + "angle": 0, + "content": "[68] Xinyue Wei, Minghua Liu, Zhan Ling, and Hao Su. Approximate convex decomposition for 3D meshes with collision-aware concavity and tree search. ACM TOG, 2022. 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.72, + 0.484, + 0.775 + ], + "angle": 0, + "content": "[69] Xinyue Wei, Kai Zhang, Sai Bi, Hao Tan, Fujun Luan, Valentin Deschaintre, Kalyan Sunkavalli, Hao Su, and Zexiang Xu. MeshLRM: Large reconstruction model for high-quality mesh. arXiv preprint arXiv:2404.12385, 2024. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.776, + 0.484, + 0.803 + ], + "angle": 0, + "content": "[70] Felix Wimbauer, Shangzhe Wu, and Christian Rupprecht. De-rendering 3D objects in the wild. In CVPR, 2022. 3, 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.805, + 0.484, + 0.858 + ], + "angle": 0, + "content": "[71] Haoqian Wu, Zhipeng Hu, Lincheng Li, Yongqiang Zhang, Changjie Fan, and Xin Yu. NeFII: Inverse rendering for reflectance decomposition with near-field indirect illumination. In CVPR, 2023. 4, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.86, + 0.484, + 0.901 + ], + "angle": 0, + "content": "[72] Shangzhe Wu, Ameesh Makadia, Jiajun Wu, Noah Snavely, Richard Tucker, and Angjoo Kanazawa. De-rendering the world's revolutionary artefacts. In CVPR, 2021. 3, 4" + }, + { + "type": "list", + "bbox": [ + 0.093, + 0.092, + 0.484, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.092, + 0.907, + 0.176 + ], + "angle": 0, + "content": "[73] Tong Wu, Jiarui Zhang, Xiao Fu, Yuxin Wang, Liang Pan Jiawei Ren, Wayne Wu, Lei Yang, Jiaqi Wang, Chen Qian, Dahua Lin, and Ziwei Liu. OmniObject3D: Large-vocabulary 3D object dataset for realistic perception, reconstruction and generation. In IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), 2023. 2, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.178, + 0.907, + 0.218 + ], + "angle": 0, + "content": "[74] Kohei Yamashita, Yuto Enyo, Shohei Nobuhara, and Ko Nishino. nLMVS-Net: Deep non-lambertian multi-view stereo. In WACV, 2023. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.22, + 0.907, + 0.26 + ], + "angle": 0, + "content": "[75] Kai Yan and Shuang Zhao. PSDR-JIT: Path-space differentiable renderer, 2020. https://github.com/andyyankai/psdr-jit.5" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.262, + 0.907, + 0.317 + ], + "angle": 0, + "content": "[76] Lior Yariv, Yoni Kasten, Dror Moran, Meirav Galun, Matan Atzmon, Basri Ronen, and Yaron Lipman. Multiview neural surface reconstruction by disentangling geometry and appearance. In NeurIPS, 2020. 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.319, + 0.907, + 0.387 + ], + "angle": 0, + "content": "[77] Vickie Ye, Ruilong Li, Justin Kerr, Matias Turkulainen, Brent Yi, Zhuoyang Pan, Otto Seiskari, Jianbo Ye, Jeffrey Hu, Matthew Tancik, and Angjoo Kanazawa. gsplat: An open-source library for Gaussian splatting. arXiv preprint arXiv:2409.06765, 2024. 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.39, + 0.907, + 0.432 + ], + "angle": 0, + "content": "[78] Hong-Xing Yu, Samir Agarwala, Charles Herrmann, Richard Szeliski, Noah Snavely, Jiajun Wu, and Deqing Sun. Accidental light probes. In CVPR, 2023. 3, 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.433, + 0.907, + 0.5 + ], + "angle": 0, + "content": "[79] Hong-Xing Yu, Michelle Guo, Alireza Fathi, Yen-Yu Chang, Eric Ryan Chan, Ruohan Gao, Thomas Funkhouser, and Jia-jun Wu. Learning object-centric neural scattering functions for free-viewpoint relighting and scene composition. TMLR, 2023. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.503, + 0.907, + 0.557 + ], + "angle": 0, + "content": "[80] Kai Zhang, Fujun Luan, Qianqian Wang, Kavita Bala, and Noah Snavely. PhySG: Inverse rendering with spherical gaussians for physics-based material editing and relighting. In CVPR, 2021. 4, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.559, + 0.907, + 0.601 + ], + "angle": 0, + "content": "[81] Kai Zhang, Fujun Luan, Zhengqi Li, and Noah Snavely. IRON: Inverse rendering by optimizing neural SDFs and materials from photometric images. In CVPR, 2022. 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.602, + 0.907, + 0.657 + ], + "angle": 0, + "content": "[82] Kai Zhang, Sai Bi, Hao Tan, Yuanbo Xiangli, Nanxuan Zhao, Kalyan Sunkavalli, and Zexiang Xu. GS-LRM: Large reconstruction model for 3D gaussian splatting. arXiv preprint arXiv:2404.19702, 2024. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.659, + 0.907, + 0.699 + ], + "angle": 0, + "content": "[83] Ruo Zhang, Ping-Sing Tsai, James Edwin Cryer, and Mubarak Shah. Shape-from-shading: a survey. IEEE TPAMI, 21(8): 690-706, 1999. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.702, + 0.907, + 0.757 + ], + "angle": 0, + "content": "[84] Xiuming Zhang, Pratul P Srinivasan, Boyang Deng, Paul Debevec, William T Freeman, and Jonathan T Barron. NeR-Factor: Neural factorization of shape and reflectance under an unknown illumination. ACM TOG, 40(6):1-18, 2021. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.758, + 0.907, + 0.8 + ], + "angle": 0, + "content": "[85] Yuanqing Zhang, Jiaming Sun, Xingyi He, Huan Fu, Rongfei Jia, and Xiaowei Zhou. Modeling indirect illumination for inverse rendering. In CVPR, 2022. 4, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.801, + 0.907, + 0.829 + ], + "angle": 0, + "content": "[86] Yunzhi Zhang, Shangzhe Wu, Noah Snavely, and Jiajun Wu. Seeing a rose in five thousand ways. In CVPR, 2023. 4" + }, + { + "type": "list", + "bbox": [ + 0.517, + 0.092, + 0.907, + 0.829 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.487, + 0.946, + 0.512, + 0.957 + ], + "angle": 0, + "content": "763" + } + ] +] \ No newline at end of file diff --git a/2025/Digital Twin Catalog_ A Large-Scale Photorealistic 3D Object Digital Twin Dataset/2d29b0d5-5c08-413f-9ad0-63d98b0c6539_origin.pdf b/2025/Digital Twin Catalog_ A Large-Scale Photorealistic 3D Object Digital Twin Dataset/2d29b0d5-5c08-413f-9ad0-63d98b0c6539_origin.pdf new file mode 100644 index 0000000000000000000000000000000000000000..cb0837843ff72cbbad82fd3966d6718c95c9748d --- /dev/null +++ b/2025/Digital Twin Catalog_ A Large-Scale Photorealistic 3D Object Digital Twin Dataset/2d29b0d5-5c08-413f-9ad0-63d98b0c6539_origin.pdf @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:ec8e0357a47f02b6a008f97df39529116e671662b04f8f6db92be6a6c48650da +size 9704564 diff --git a/2025/Digital Twin Catalog_ A Large-Scale Photorealistic 3D Object Digital Twin Dataset/full.md b/2025/Digital Twin Catalog_ A Large-Scale Photorealistic 3D Object Digital Twin Dataset/full.md new file mode 100644 index 0000000000000000000000000000000000000000..43d244e5cbae8ed25133218c42ca8334c4ccec0e --- /dev/null +++ b/2025/Digital Twin Catalog_ A Large-Scale Photorealistic 3D Object Digital Twin Dataset/full.md @@ -0,0 +1,312 @@ +# Digital Twin Catalog: A Large-Scale Photorealistic 3D Object Digital Twin Dataset + +Zhao Dong $^{1}$ Ka Chen $^{1,*}$ Zhaoyang Lv $^{1,*}$ Hong-Xing Yu $^{2,*}$ Yunzhi Zhang $^{2,*}$ Cheng Zhang $^{1,*}$ Yufeng Zhu $^{1,*}$ Stephen Tian $^{2}$ Zhengqin Li $^{1}$ Georgie Moffatt $^{1}$ Sean Christofferson $^{1}$ James Fort $^{1}$ Xiaqing Pan $^{1}$ Mingfei Yan $^{1}$ Jiajun Wu $^{2}$ Carl Yuheng Ren $^{1}$ Richard Newcombe $^{1}$ + +$^{1}$ Meta Reality Labs Research Stanford University + +![](images/5c6ef2dab8e54265416509670d0062d9671cdfe6f8dc0a4a7cec32c62d6f5d97.jpg) +(a) Examples of 2,000 digital twins of real-world objects + +![](images/63ea79faa2c911314cac9024de6bb70ec75275098cc717280376a3c205019268.jpg) +(b) Shape & PBR Materials + +![](images/7d21347e865b975546c4908b004218877b7eed45912281de5bc1d261d7ec61c1.jpg) +(c) DSLR Data +Figure 1. The Digital Twin Catalog (DTC) dataset comprises 2,000 digital twins of physical-world objects (a), characterized by millimeter-level geometric accuracy and photorealistic PBR materials (b). DTC includes evaluation data captured using both DSLR cameras and egocentric Aria glasses, featuring captured images with precise foreground object masks and environment lighting for relighting evaluation. + +![](images/8ddbce2402e3d85449771a91d3f6fe0241663684d598440c2308bf4ff77860b6.jpg) +(d) Egocentric Data + +# Abstract + +We introduce Digital Twin Catalog (DTC), a new large-scale photorealistic 3D object digital twin dataset. A digital twin of a 3D object is a highly detailed, virtually indistinguishable representation of a physical object, accurately capturing its shape, appearance, physical properties, and other attributes. Recent advances in neural-based 3D reconstruction and inverse rendering have significantly improved the quality of 3D object reconstruction. Despite these advancements, there remains a lack of a large-scale, digital twin quality real-world dataset and benchmark that can quantitatively assess and compare the performance of different reconstruction methods, as well as improve reconstruction quality through + +training or fine-tuning. Moreover, to democratize 3D digital twin creation, it is essential to integrate creation techniques with next-generation egocentric computing platforms, such as AR glasses. Currently, there is no dataset available to evaluate 3D object reconstruction using egocentric captured images. To address these gaps, the DTC dataset features 2,000 scanned digital twin-quality 3D objects, along with image sequences captured under different lighting conditions using DSLR cameras and egocentric AR glasses. This dataset establishes the first comprehensive real-world evaluation benchmark for 3D digital twin creation tasks, offering a robust foundation for comparing and improving existing reconstruction methods. The DTC dataset is already released at https://www.projectaria.com/datasets/dtc/ and we will also make the baseline evaluations open-source. + +![](images/787532f76f3c7d05c010f997c86ce2161ca34f8cd6a082de3e89196c9b445026.jpg) + +![](images/7ac0bd5a69b106213fb119b8eb13f0f3640675ae19d597720bfa6b4fffd7aae1.jpg) + +![](images/03814b73dbaeeceaca5b864b789e7d1bd77546a4fce1c8fba05868b819ffbd6a.jpg) + +![](images/162bee86384d06d222c9ba1a0dd1d208552d44bf41cf0d1b69620172067937cf.jpg) + +![](images/4b26a4302ba3923f69965b63cf545b83d58552ac9cb9e125720cd32dd3681f09.jpg) + +![](images/c1f9ec98603ce255dfa5b73ba916b26a6b329bd7b9fb068f4a227c67879135ff.jpg) + +![](images/c3f63b0fd71a76659fed8016ddb3d7ca9f3f5354904b2712fcdf92e3b6356eae.jpg) +Figure 2. Example DTC models with photorealistic PBR materials. + +![](images/c742133f79593415e6754d2c6e6b65a5b75e0dc5457fb9ba9d20bff3338b32a0.jpg) + +![](images/6d2d671251186c923beaff28ce7b2a88d1983ab8b524b097b52ad10d187569e7.jpg) + +![](images/33f163d2d32987871165fce759381ea7ecfb7b5057c30b7527a1a50860194a9f.jpg) + +# 1. Introduction + +A digital twin of a 3D object is a highly detailed, virtually indistinguishable representation of a physical object, capturing its shape, appearance, physical properties and other attributes with precision. Such a digital twin enables visualization, analysis, and interaction as if it were the real object, supporting simulation, automation, and real-world problem-solving across a wide range of applications in AR/VR [3, 28], spatial/contextual AI [2], and robotics [23, 43]. As fundamental properties of an object, its shape and appearance form the basis for recognizing and interpreting the 3D object, enabling identification, manipulation, and realistic rendering. Recovering these attributes has long been a foundational topic in computer vision and graphics, inspiring extensive research in 3D reconstruction and inverse rendering. Recent breakthroughs in neural-based representation and reconstruction techniques, such as NeRF [52] and 3D Gaussian splattig (3DGS) [37], have significantly elevated the quality of novel view synthesis (NVS) to photorealistic levels. Many subsequent works [35, 62] integrate neural reconstruction with physically-based inverse rendering, enabling relightable appearances. Furthermore, leveraging priors from large reconstruction models (LRMs) [30], high-quality shape and appearance reconstruction can now be achieved with as few as one to four views [41, 69, 82]. + +Despite the rapid advancements in 3D object reconstruction, one question remains: does the reconstruction quality truly meet the standard of a digital twin, where virtual representations are indistinguishable from reality? This digital twin standard demands both highly accurate shape matching and photorealistic appearance across different lighting, which present significant acquisition challenges for real-world objects. Existing object-centric datasets for 3D recon + +struction or inverse rendering have focused on either dataset size [18] or quality of specific aspects [16, 21, 34, 46, 61, 73], often sacrificing comprehensive fidelity and limiting their application scope. This trade-off has led to a lack of datasets that fully satisfy the digital twin criteria, hindering current 3D reconstruction methods from achieving digital twin fidelity. To bridge this gap, we developed the Digital Twin Catalog (DTC) dataset, comprising 2,000 scanned 3D object models (Fig. 1(a)), each with millimeter geometry accuracy and photorealistic PBR materials (Fig. 1(b), Fig. 2). + +In addition to 3D digital twin models, the DTC dataset includes evaluation data designed to support 3D object reconstruction research. This evaluation data features multiview image sequences with precise foreground object masks and environment lighting information for relighting evaluation. Traditionally, high-quality HDR images captured with modern DSLR cameras have been the standard for 3D reconstruction research. Looking ahead, we encourage the integration of 3D reconstruction research with next-generation human-centric computing platforms, such as egocentric AR glasses, aiming to democratize 3D reconstruction techniques and empower everyone to effortlessly create 3D digital twins. To this end, alongside DSLR-captured evaluation data (Fig. 1(c)), the DTC dataset also provides egocentric evaluation data captured using Project Aria glasses (https://www.projectaria.com) (Fig. 1(d)). + +The DTC dataset offers extensive opportunities for advancing research in object digital twin creation. We provide a benchmark for state-of-the-art 3D object reconstruction and inverse rendering methods. These benchmarks evaluate performance across novel view synthesis (NVS), shape reconstruction, and relightable appearance reconstruction. We further provide the evaluation of novel view synthesis meth + +Table 1. Comparison with existing object-centric inverse rendering datasets. *Objaverse [18] consists of both synthetic objects and real scans, only part of which contain PBR materials. + +
Dataset# ObjectsRealScene TypeMulti-viewShapePBR Mat.Relit ImageLightingEgocentric Cap.
ShapeNet-Intrinsics [61]31KXsynthetic✓✓X✓✓X
NeRD Synthetic [11]3Xsynthetic✓✓✓✓✓X
ABO [16]8KXsynthetic✓✓✓✓✓X
MIT Intrasins [25]20✓studio✓XXXXX
DTU-MVS [34]80✓studio✓✓XXXX
Objaverse [18]818K(✓)*studio✓✓(✓)*XXX
DiLiGenT-MV [42]5✓studio✓✓XX✓X
ReNe [65]20✓studio✓XXX✓X
OpenIllumination [46]64✓studio✓X✓✓✓X
GSO [21]1030✓studio✓✓XXXX
Lombardi et al. [47]6✓in-the-wildX✓X✓✓X
NeRD Real [11]4✓in-the-wild✓X✓✓XX
NeROIC [38]3✓in-the-wild✓X✓✓XX
Oxholm et al. [56]4✓in-the-wild✓✓X✓✓X
OmniObject3D [73]6k✓in-the-wild✓✓XXXX
Stanford Orb [39]14✓in-the-wild✓✓✓✓✓X
DTC (ours)2k✓in-the-wild✓✓✓✓✓✓
+ +ods using the egocentric aligned DTC data. Additionally, we explore the dataset's potential in downstream robotics applications by assessing its effectiveness in training robotic policies for pushing and grasping tasks in simulation. These benchmarks and applications provide valuable insights, highlight existing challenges, and uncover promising directions for future research in 3D digital twin creation. + +# 2. Related Work + +We provide a comparison of our DTC dataset to existing object-centric datasets in Table 1. We provide the largest 3D dataset with PBR materials and real world multi-view recordings with digital twin counterparts. We further provide digital twin aligned egocentric recordings, the first of their kind in the egocentric domain. We will discuss the related datasets and methods they can empower as follows. + +3D Digital Twin Datasets Existing 3D digital twin datasets with PBR materials often serve as ground truth for evaluating inverse rendering results. Early efforts [6, 25] provide small-scale intrinsic image of real objects and do not provide shape or PBR material information. Synthetic datasets [11, 16, 45, 61, 70, 72] are widely used for evaluation but do not represent the complexity in a real world environment. For datasets that contain real objects, [18, 34, 42, 47, 56, 65], the reconstruction quality can vary, which leads to imprecise evaluations. Table 1 provides a comparison to the previous work in this domain. Compared to Objaverse[18, 19], which is a collection of existing 3D models with only a small subset containing PBR materials with varying quality, we offer a high quality collection of 3D object data that is also aligned with real world recordings. Compared to OmniObject3D [73], the DTC models + +provide higher-quality shape and additional PBR materials that are necessary for high quality inverse rendering. We offer the largest quantity of 3D object models compared to all counterparts in various tasks. For real-world evaluation, Stanford-ORB [39] was the prior largest inverse rendering benchmark with in-the-wild lighting. In contrast, we provide more object diversity and higher quality for each object model. The Aria Digital Twin dataset [58] was the first dataset to provide digital twin aligned environments for the scenes and recorded using egocentric device. However, their scene environments are limited and the contained object ground truths inside do not have high quality geometries with PBR materials. + +Object Reconstruction & Inverse Rendering. Using object-centric multi-view images as input, early object reconstruction methods focused on estimating individual object properties, such as shape from shading [5, 9, 83], material acquisition [44, 45, 56, 57, 74], and lighting estimation [67, 78]. Some approaches also aimed to recover reflectance and illumination assuming known object shapes [47, 48]. Inverse rendering, which seeks to invert the rendering equation [36], estimates an image's intrinsic components—geometry, materials, and lighting. The advent of differentiable renderers [14, 15, 49] enabled full-fledged inverse rendering methods to simultaneously recover all these properties for object reconstruction [50]. + +Neural volumetric representations such as Neural Radiance Fields (NeRFs) [52] and the like [7, 11-13, 59, 79, 84] encode geometry and appearance as volumetric densities and radiance with a Multi-Layer Perceptron (MLP) network, and render images using the volume rendering equation [51]. 3D-GS [37] introduces 3D Gaussian primitives and rasterization and its following-up variants [31] demonstrates high quality + +geometry prediction as well. + +Other surface-based representations [29, 53, 62, 66, 71, 80, 81, 85] extract surfaces as the zero level set, for instance, of a signed distance function (SDF) or an occupancy field [55], allowing them to efficiently model the appearance on the surface with an explicit material model, such as bidirectional reflectance distribution functions (BRDFs). This also enables modeling more complex global illumination effects, such as self-shadows. Most of these methods focus on per-scene optimization and require dense multiple views as input. Recently, researchers have incorporated learning-based models, distilling priors from large training datasets for fast inference on limited test views [8, 10, 33, 45, 61, 70, 72, 86]. + +In this work, we provide the DTC dataset with real-world object recordings that can serve as the benchmark to evaluate object-centric inverse rendering tasks. We evaluate representative baselines from existing work. + +# 3. Digital Twin Catalog: A Large-Scale Photorealistic 3D Object Digital Twin Dataset + +# 3.1. Dataset Composition + +Our DTC dataset contains: (1) 2,000 scanned 3D object models, featuring millimeter geometric accuracy relative to their physical counterparts, along with a full set of photorealistic PBR material maps (Fig. 1), (2) 100 DSLR-captured evaluation data of 50 objects under different lighting conditions, and (3) 200 egocentric Aria-captured evaluation data of 100 objects with both active and casual observation modes. + +# 3.2. Creation of 3D Object Models + +Utilizing the state-of-the-art industrial 3D object scanner [1], we selected 2,000 physical-world objects spanning 40 LVIS [27] categories, carefully chosen to ensure both category diversity and compatibility with the scanner's capabilities. As illustrated in Fig. 3(a) (c), the scanner features a fixed lighting-camera setup within an upper-hemisphere dome, equipped with 8 structured lights for geometry scanning, and 29 spotlights and 29 cameras for material acquisition. During the scanning process, the object is placed on a central holder, and its pose can be adjusted with multi-round scanning to achieve a complete 360-degree scan. For our dataset, each object typically undergoes three pose changes, with a total scanning time of approximately 20 minutes per object. + +After scanning, a proprietary post-processing pipeline reconstructs both the geometry and the PBR material maps. For 4K-resolution PBR material maps, the post-processing requires approximately 4 hours per object. In terms of quality, the structured-light-based shape reconstruction (Fig. 3 (b)) in the post-processing achieves millimeter-level geometric accuracy. However, the material optimization process + +![](images/514e435cb5ecf8ec602ac85b2b43b8a3f2c3a2d8ef0c882b2ac2ddd7d287b30e.jpg) + +![](images/135bc7d0496f2c9a75be1831c6bd023e92a6f80918f7d1b09193a3d349de2602.jpg) + +![](images/1e2df084aba61005a1fff73ede9c775b1492db5fa33bf103bb57e655c46b4ffc.jpg) +Figure 3. 3D object scanner by Covision Media®. + +![](images/2221885c36272e45c04b1bb11f68fe0926c7cc4abddbd05d5b0cb6a6dc0834fc.jpg) + +![](images/b6c70bcdf78249b5ac44914e03c7c6a9ce38c56777016c4c60a6be8291f00057.jpg) + +![](images/81e1f60c73757d07d21564f4f92073b2c070847f5d4120db339f1d4b88771dbc.jpg) + +![](images/c7579deec3fdec506ecd75bef2a02066e26fe418be49935994c28f186ed5f5bc.jpg) + +![](images/ea01e0790fab617bd0e4f3f993dbf25516efb0989fd13c4c1b2cf25eb8b087e6.jpg) +Figure 4. Rendered DTC models (left) v.s. Photo (Right). + +![](images/2dde835d30df0f67fe2c497f736b8b0b1a2f89ba0109ae86e7edd6a6d6160cc2.jpg) + +performs best for diffuse objects and often struggles with glossy or shiny surfaces. To address this limitation, we hired technical artists to develop a workflow to refine materials for glossy and shiny objects, ensuring that the material quality meets the standards of a digital twin. + +![](images/788b6a091b4172786c2a711155a41ca86b1db107b1def661f06eda6b8ac10fa2.jpg) +Figure 5. Shape and material (albedo) quality comparison between Stanford-ORB [39] (left) and our DTC (right). + +3D Model Accuracy. To validate the material and geometry accuracy of the 3D models, we compared a rendered image of our scanned and processed model with a photograph of the same object taken inside a light box. A virtual light box was meticulously modeled to replicate the light intensity and color temperature of the real light box. The scanned object was then placed in the virtual light box to generate the rendered image. The side-by-side comparison demonstrates a remarkable match between the rendered and real images (Fig. 4). + +Comparison Against Stanford-ORB. We also scanned the objects used in Stanford-ORB [39] to compare the shape and appearance quality. As illustrated in Fig. 5, the Stanford-ORB models exhibit shape artifacts and noisy, lower-quality materials compared to our models. + +# 3.3. DSLR Evaluation Data + +Within DTC, we include a DSLR-captured evaluation dataset of 50 objects from Sec. 3.2 captured under two different lighting conditions, resulting in 100 distinct image sequences. For every sequence, we provide (a) approximately 120 HDR and LDR images from different viewing directions, (b) one object pose and (c) per-image camera pose. The two lighting conditions are represented using two environment maps. + +Data Capture. To ensure the DSLR evaluation data quality, we designed and built a DSLR camera rig to automate the capture process (Fig. 6). The rig is designed to rotate the cameras around the centralized object, assuming the environment lighting remains unchanged during the capture. It features a motorized rotary stage with a centrally mounted stationary platform. Attached to the rotary stage is an extrusion frame that forms the gantry arm, supported by a set of castors to bear its weight and enable smooth rotation around the central axis. The extrusion frame is equipped with adjustable camera mounts, allowing DSLR cameras to be positioned flexibly to optimize the capture setup. For our capture process, we utilized three DSLR cameras to perform a 360-degree rotation around the object, capturing images at 9-degree intervals, resulting in 120 photos per object. To ensure precise camera pose estimation, a ChArUco board + +![](images/1a6ce4a894ebbfcfb6ec2630ff256afe74443ffa2f522328986bb4e235917f94.jpg) +Figure 6. DSLR rig for capturing evaluation data. + +was placed beneath the object during the capture. Example images from this setup are shown in Fig. 6. + +Environment Maps. Following a similar approach in Stanford-ORB [39], we capture the two environment maps using chrome ball images obtained with the capture rig described earlier. With precise camera poses provided by a ChArUco board placed beneath the chrome ball, we first fit a synthetic 3D sphere to the chrome ball by optimizing its 3D position using a geometry-friendly differentiable renderer [40, 75]. Subsequently, using a differentiable Monte Carlo-based renderer [32], we refine the environment map to match the reflection on the chrome ball, employing the single-view light estimation method proposed in [78]. The coordinate system of the environment map is determined by the ChArUco detection. + +Pose Registration for Camera and Object. We first obtain the initial camera poses using the ChArUco board. In most cases, for images captured from the top and middle views, the pose estimates are typically accurate. However, for bottom-view images, inaccuracies arise due to pattern distortion at grazing angles. To mitigate this issue, we refine the camera poses by fitting the rendering of a virtual ChArUco board to the real captured photo using a differentiable renderer [32, 40, 75]. Once the camera poses have been accurately refined, we optimize the object pose by minimizing the mask loss between the rendered mask and the reference one generated by [62]. By providing separate poses for camera and object anchored by ChArUco, the cameras and object are automatically aligned with the optimized environment maps. + +# 3.4. Egocentric Evaluation Data + +We include an evaluation dataset of real world recordings captured by an egocentric device paired with 100 objects from the DTC dataset. We capture the egocentric recordings using the open-sourced Project Aria device [22] and acquire the additional 3D ground truth using its machine perception tool, which includes online device calibration, device trajectory and semi-dense point clouds for each recording. Each recording contains a single 3D object and precisely aligned object poses in the Aria trajectory coordinate frame. + +![](images/dcd62f9fc15c847d5a6c0ee8cc41a039480a020f8dc5ae093f19a24eaae0f820.jpg) +Figure 7. The workflow to align egocentric video and 3D objects. We acquire the object's semi-dense point cloud from the egocentric recording and the neural reconstructed mesh using Neural-PBIR [62]. Finally, we align the 3D object with the rendered mask from the neural reconstructed mesh using physics-based differentiable rendering (PBDR). + +Both the object and the trajectory are in metric scale. We can derive additional ground truth for each video from the aligned 3D object properties. + +Data Capture. To feature real world recordings observed from human perspectives that can be representative of 3D object reconstruction for AR/MR devices or robots, we provide two types of recording trajectories for selected objects, termed active and passive respectively, collected by human wearers. The active recording features a complete 360 view of the objects, which is similar to existing object 3D novel view synthesis dataset. The passive recording features casual looks from human wearer, which only contain partial views of the object from certain viewing angles along the trajectory. In dataset creation, we collected the active and passive recordings in the same environment and generate their 3D information in the shared 3D space. This helps reduces potential failures when aligning the object to passive 3D recordings, which are shorter and contains less 3D information. To reduce the effect of noises and motion blur, which are common in egocentric videos in indoor low light environments, we light the capture environment to $3\mathrm{K}+$ lux illumination and used a fixed low-exposure and gain profile to collect each recording with the appropriate brightness. + +Alignment between Egocentric Video and Object. We provide an illustration of the object alignment to the egocentric video in Fig. 7. Given the images, camera poses and semi-dense point cloud acquired from Project Aria tools, we employ a neural-based mesh reconstruction method [62] to + +create reference meshes for generating high-quality reference masks and used that to align with the corresponding 3D object mesh. This alignment step serves as an initialization for a more precise pose refinement, which leverages differentiable rendering. For certain objects with symmetric geometry, we observe this process can introduce ambiguities in point cloud registration and subsequent failures. To address such cases, we provide a GUI to manually align and correct the object alignment. Finally, akin to the DSLR camera pose registration phase, we optimize a mask loss over object poses to achieve fine-tuned pose registration. + +# 4. Benchmarking and Applications + +We first use our DLSR and egocentric dataset as a benchmark for existing state-of-the-art methods. For inverse rendering, we design metrics to evaluate the shape and material quality of the recovered 3D object digital twin from three perspectives. For egocentric recording, we evaluate the novel-view synthesis as the initial evaluation. We include additional tasks, e.g. sparse view reconstruction in the supplementary materials for both DSLR and egocentric recordings. Finally we can demonstrate our high quality 3D digital twin models can be beneficial to robotics domain using an application in robotics manipulation. + +# 4.1. Application to Inverse Rendering for DSLR + +The DSLR dataset in DTC provides accurate ground truth, including poses, lighting and 3D models, for inverse rendering + +Table 2. Benchmark comparison of existing methods on inverse rendering for DSLR. Depth SI-MSE and Shape Chamfer distance $\times {10}^{-3}$ . + +
GeometryNovel Scene RelightingNovel View Synthesis
Depth↓Normal↓Shape↓PSNR-H↑PSNR-L↑SSIM↑LPIPS↓PSNR-H↑PSNR-L↑SSIM↑LPIPS↓
Neural-PIL [12]5.710.2525.02N/A28.4235.760.8820.096
PhySG [80]0.310.1611.3127.2832.860.9590.04928.5434.460.9640.045
NVDiffRec [53]0.020.071.6426.9933.270.9510.03728.9534.920.9670.029
NeRD [11]4.550.45108.2026.1032.600.9480.06126.8033.400.8820.102
InvRender [71]0.220.030.7529.5235.980.9610.03731.6437.820.9700.033
NVDiffRecMC [29]0.020.061.3427.7834.550.9520.04231.2738.170.9720.032
+ +Table 3. Benchmark on the egocentric aligned recordings. + +
PSNR ↑LPIPS ↓SSIM ↑Depth ↓Normal ↓
3D-GS [37]28.810.0200.98880.17680.3301
2D-GS [31]28.750.0200.98860.17550.2112
+ +tasks and serves as an evaluation suite to benchmark the performance of inverse rendering methods. We select six prior methods for this task and evaluate their performance using the ground truth provided by our dataset. In the following sections, we describe the data splitting strategy, evaluation metrics, and baselines. + +Data Splitting. For benchmarking purposes, we select 15 objects from the DSLR dataset captured under two distinct lighting environments, resulting in a total of 30 image sequences. The selected objects encompass a diverse range of geometric and material properties to ensure a comprehensive evaluation. For each scene, 8 views are selected for testing, while the remaining views are reserved for training. + +Evaluation Metrics. The metrics measure the accuracy of three aspects of baseline performance: geometry estimation, relighting, and novel view synthesis. For geometry estimation, we evaluate the accuracy of predicted depth and normal maps under held-out test views, as well as 3D meshes extracted from baseline methods, compared with the ground truth from our dataset. Relighting metrics evaluate the material decomposition quality of baselines by measuring the accuracy of predicted images under held-out lighting conditions. For view synthesis, we compare the predicted images from viewpoints unseen during training to ground truth captures. We refer to Kuang et al. [39] for metric details. + +Baselines. We include the following baselines: NVDiffRec [53] and NVDiffRecMC [29], with a hybrid shape representation DMTet [60]; InvRender [85] and PhySG [80], which adopt signed distance functions (SDFs) to represent object geometry [76] and utilize implicit neural fields for material decomposition; Neural-PIL [12] and NeRD [11], which use NeRFs [52] as scene representations. + +# 4.2. Application to Egocentric Reconstruction + +Our digital twin models, aligned with real world video using the method described in Sec. 3.4, can help obtain accurate + +ground truth for object-centric images that were previously difficult to acquire. We provide the first evaluation of object-centric novel view synthesis recorded from an egocentric device. We use the projected object shape given the 3D pose of the object and cameras in scene coordinates to acquire the image masks, depth and normal for each object. We selected 15 recordings from the egocentric recording sessions as the evaluation and used the active recordings to benchmark novel view reconstruction. For each recording, we hold out every 8th image view as a testing view. + +**Evaluations.** We build our baselines based on the gsplats [77] implementation of the 3D Gaussian Splitting (GS) [37] and 2D GS [31], and handle the effect of lens shading from the Project Aria lens [26]. We calculate PSNR, depth and normal based on the observed objects with masks and provide SSIM and LPIPS score on images by masking out the non-object areas as black. Table 3 shows the benchmark results of the baselines. We use the same depth and normal metric in DSLR evaluation. We provide additional qualitative evaluations and analyses on egocentric data towards sparser view settings in the supplementary materials. + +# 4.3. Application to Robotic Manipulation + +High-quality object models have been leveraged in prior work to train real-world robotic agents in scenes represented explicitly [54] or implicitly [64]. These object models have also been shown to facilitate object-centric pose and lighting parameter estimation, enabling model-based planning [63]. In this section, we empirically evaluate the effectiveness of using DTC dataset objects in training robotic policies. Specifically, we consider learning robotic pushing and grasping skills in simulation. + +First, we sample a subset of 24 cup category objects from the DTC dataset and 24 cup objects from Objaverse-XL [19] ${}^{1}$ . Since not all Objverse-XL objects come with textures, we randomize the colors of those objects uniformly in RGB space. To compute collision meshes for physical simulation, we perform convex decomposition on each object with CoACD [68]. We import these objects along with a UR5e robot equipped with a Robotiq 2F-85 (pushing) or + +
TaskDTC (ours)Objverse-XL [19]
Pushing @ 2cm36.3% ± 1.5%25.3% ± 6.0%
Pushing @ 3cm43.7% ± 1.2%29.7% ± 6.0%
Pushing @ 5cm47.0% ± 2.6%40.3% ± 5.5%
Grasping42.7% ± 4.7%38.6% ± 11.0%
+ +Table 4. Success rate of policies trained on data collected using objects from our DTC dataset and sampled from Objaverse-XL when evaluated on an unseen test object. Errors indicate sample standard deviation over three policy training seeds. + +Robotiq 2F-140 (grasping) gripper into the PyBullet simulator [17] and collect data for each robotic task as described below. After training policies on data from each object set, we evaluate policy performance on a relatively high-quality unseen test object from the StanfordORB dataset [39]. + +Pushing. For the pushing task, we collect 5000 trajectories of pseudo-random robotic interaction data for each object set. For each trajectory, a single object from the considered object set and its initial position are randomly selected. Then we train a goal-conditioned neural network policy $\pi (a|o,o_g)$ where $o$ is an image observation of the current scene and $o_{g}$ is a goal image indicating the desired final object and robot position, $o,o_{g}\in \mathbb{R}^{256\times 256\times 3}$ , and the action $a\in \mathbb{R}^2$ represents a change in the robot's end-effector position in the $x$ and $y$ axes. The $z$ axis end-effector height is held fixed. We sample goals for training via hindsight relabeling [4, 20, 24]. We then perform evaluation on 100 randomly sampled test goals manipulating an unseen test cup from StanfordORB [39]. + +Grasping. For grasping, we collect 5000 successful grasp examples for each object set by first placing a single object into the scene, randomizing the object identity and initial position. We then randomly sample candidate grasp poses in a radius around the object's position and simulating their outcomes, rejecting unsuccessful grasps. We train a grasping policy $\pi(a|o)$ where $o \in \mathbb{R}^{256 \times 256 \times 3}$ is an image observation of the scene and $a \in \mathbb{R}^4$ represents the $x, y, z$ position and $\theta$ yaw rotation of the robot end-effector at which to attempt the grasp. Again we use 100 test cup object poses. + +Results. We report the results in Table 4 and Fig. 8. We find that across both pushing and grasping tasks, policies trained on DTC dataset objects outperform those trained on Objaverse-XL objects when evaluated on the unseen test object. For pushing, we report performance by defining binary success thresholds based on the final Euclidean distance of the object position to the goal position. Training on DTC objects appears to be especially helpful at enabling policies to make finer adjustments, improving pushing success rates at stricter thresholds. Additional experimental details can be found in the supplementary. + +![](images/3b4d8b3f2caaf2fc9858feb7baa824272d8d29f523c49ef5c6ccb9d47ac20594.jpg) +Figure 8. Success rates on robotic pushing task when training with our DTC objects and sampled Objaverse-XL objects. Particularly at lower error thresholds, policies trained on DTC objects outperform those trained using Objaverse-XL objects. Shaded bars represent sample standard deviations over policy training random seeds. + +# 5. Conclusion + +We presented a new large scale photorealistic 3D digital twin dataset with the real world recordings that contain its real world counterpart. We provide extensive evaluations of baselines on our DTC dataset serving as new benchmark for inverse rendering and novel view synthesis task. We also demonstrated that high quality digital twin models can be beneficial to applications in robotics domain. We believe our efforts can empower the research community to build and leverage digital twin models for future applications. + +Limitations. Achieving high quality digital twin models currently requires deliberate hardware setup and human efforts in refinement. Solving this challenge without sacrificing quality can significantly further enhance the volume of digital twin models. Our hardware is also limited to objects within a certain size and can not yet recover objects that are deformable, highly specular, or transparent. + +Future work. The existing digital twin model creation in DTC dataset involves lengthy post-processing and may require subjective human refinement, hindering the automation of model generation. However, recent advancements in physics-based differentiable rendering hold promise for enabling faster and more accurate creation of digital twins, especially for material reconstruction. Furthermore, building large-scale digital twins for applications will necessitate efforts to enhance the diversity in object appearance (e.g., transparent objects) and to capture additional attributes, such as physical properties and functionalities. + +For robotics applications, while deploying manipulation policies learned in simulation to the real world remains generally challenging, we hope that the high-quality digital twin data provided by DTC can serve as a stepping stone towards effective sim-to-real transfer. + +Acknowledgments. This work is in part supported by NSF CCRI #2120095, RI #2211258, and RI #2338203. YZ is in part supported by the Stanford Interdisciplinary Graduate Fellowship. + +# References + +[1] Covision 3D object scanner. https://www.covisionmedia.ai/en/scanners.4 +[2] Kosmas Alexopoulos, Nikolaos Nikolakis, and George Chryssolouris. Digital twin-driven supervised machine learning for the development of artificial intelligence applications in manufacturing. International Journal of Computer Integrated Manufacturing, 33(5):429-439, 2020. 2 +[3] Moayad Aloqaily, Ouns Bouachir, Fakhri Karray, Ismaeel Al Ridhawi, and Abdulmotaleb El Saddik. Integrating digital twin and advanced intelligent technologies to realize the meta-verse. IEEE Consumer Electronics Magazine, 12(6):47-55, 2022. 2 +[4] Marcin Andrychowicz, Filip Wolski, Alex Ray, Jonas Schneider, Rachel Fong, Peter Welinder, Bob McGrew, Josh Tobin, OpenAI Pieter Abbeel, and Wojciech Zaremba. Hindsight experience replay. In NeurIPS, 2017. 8 +[5] Jonathan T Barron and Jitendra Malik. Shape, illumination, and reflectance from shading. IEEE TPAMI, 37(8):1670-1687, 2014. 3 +[6] Sean Bell, Kavita Bala, and Noah Snavely. Intrinsic images in the wild. ACM TOG, 33(4):1-12, 2014. 3 +[7] Sai Bi, Zexiang Xu, Kalyan Sunkavalli, Miloš Hasan, Yannick Hold-Geoffroy, David Kriegman, and Ravi Ramamoorthi. Deep reflectance volumes: Relightable reconstructions from multi-view photometric images. In ECCV, 2020. 3 +[8] Sai Bi, Zexiang Xu, Kalyan Sunkavalli, David Kriegman, and Ravi Ramamoorthi. Deep 3D capture: Geometry and reflectance from sparse multi-view images. In CVPR, 2020. 4 +[9] Martin Bichsel and Alex P Pentland. A simple algorithm for shape from shading. In CVPR, 1992. 3 +[10] Mark Boss, Varun Jampani, Kihwan Kim, Hendrik P.A. Lensch, and Jan Kautz. Two-shot Spatially-varying BRDF and Shape Estimation. In CVPR, 2020. 4 +[11] Mark Boss, Raphael Braun, Varun Jampani, Jonathan T. Barron, Ce Liu, and Hendrik P.A. Lensch. NeRD: Neural reflectance decomposition from image collections. In ICCV, 2021. 3, 7 +[12] Mark Boss, Varun Jampani, Raphael Braun, Ce Liu, Jonathan Barron, and Hendrik Lensch. Neural-PIL: Neural pre-integrated lighting for reflectance decomposition. In NeurIPS, 2021. 7 +[13] Mark Boss, Andreas Engelhardt, Abhishek Kar, Yuanzhen Li, Deqing Sun, Jonathan Barron, Hendrik Lensch, and Varun Jampani. SAMURAI: Shape and material from unconstrained real-world arbitrary image collections. In NeurIPS, 2022. 3 +[14] Guangyan Cai, Kai Yan, Zhao Dong, Ioannis Gkioulekas, and Shuang Zhao. Physics-based inverse rendering using combined implicit and explicit geometries. CGF, 41(4):129-138, 2022. 3 + +[15] Wenzheng Chen, Joey Litalien, Jun Gao, Zian Wang, Clement Fuji Tsang, Sameh Khamis, Or Litany, and Sanja Fidler. DIBR++: learning to predict lighting and material with a hybrid differentiable renderer. In NeurIPS, 2021. 3 +[16] Jasmine Collins, Shubham Goel, Kenan Deng, Achleshwar Luthra, Leon Xu, Erhan Gundogdu, Xi Zhang, Tomas F Yago Vicente, Thomas Dideriksen, Himanshu Arora, et al. ABO: Dataset and benchmarks for real-world 3D object understanding. In CVPR, 2022. 2, 3 +[17] Erwin Coumans and Yunfei Bai. Pybullet, a python module for physics simulation for games, robotics and machine learning. http://pybullet.org, 2016-2019.8 +[18] Matt Deitke, Dustin Schwenk, Jordi Salvador, Luca Weihs, Oscar Michel, Eli VanderBilt, Ludwig Schmidt, Kiana Ehsani, Aniruddha Kembhavi, and Ali Farhadi. Objverse: A universe of annotated 3D objects. In CVPR, 2023. 2, 3 +[19] Matt Deitke, Ruoshi Liu, Matthew Wallingford, Huong Ngo, Oscar Michel, Aditya Kusupati, Alan Fan, Christian Laforte, Vikram Voleti, Samir Yitzhak Gadre, et al. Objverse-XL: A universe of $10\mathrm{m} + 3\mathrm{D}$ objects. NeurIPS, 2024. 3, 7, 8 +[20] Yiming Ding, Carlos Florensa, Pieter Abbeel, and Mariano Philipp. Goal-conditioned imitation learning. In NeurIPS, 2019. 8 +[21] Laura Downs, Anthony Francis, Nate Koenig, Brandon Kinman, Ryan Hickman, Krista Reymann, Thomas B McHugh, and Vincent Vanhoucke. Google scanned objects: A high-quality dataset of 3D scanned household items. In ICRA, 2022. 2, 3 +[22] Jakob Engel, Kiran Somasundaram, Michael Goesele, Albert Sun, Alexander Gamino, Andrew Turner, Arjang Talatof, Arnie Yuan, Bilal Souti, Brighid Meredith, Cheng Peng, Chris Sweeney, Cole Wilson, Dan Barnes, Daniel DeTone, David Caruso, Derek Valleroy, Dinesh Ginjupalli, Duncan Frost, Edward Miller, Elias Mueggler, Evgeniy Oleinik, Fan Zhang, Guruprasad Somasundaram, Gustavo Solaira, Harry Lanaras, Henry Howard-Jenkins, Huixuan Tang, Hyo Jin Kim, Jaime Rivera, Ji Luo, Jing Dong, Julian Straub, Kevin Bailey, Kevin Eckenhoff, Lingni Ma, Luis Pesqueira, Mark Schwesinger, Maurizio Monge, Nan Yang, Nick Charron, Nikhil Raina, Omkar Parkhi, Peter Borschowa, Pierre Moulon, Prince Gupta, Raul Mur-Artal, Robbie Pennington, Sachin Kulkarni, Sagar Miglani, Santosh Gondi, Saransh Solanki, Sean Diener, Shangyi Cheng, Simon Green, Steve Saarinen, Suvam Patra, Tassos Mourikis, Thomas Whelan, Tripti Singh, Vasileios Balntas, Vijay Baiyya, Wilson Dreewes, Xiaqing Pan, Yang Lou, Yipu Zhao, Yusuf Mansour, Yuyang Zou, Zhaoyang Lv, Zijian Wang, Mingfei Yan, Carl Ren, Renzo De Nardi, and Richard Newcombe. Project Aria: A new tool for egocentric multi-modal ai research, 2023. 5 +[23] Ruohan Gao, Yen-Yu Chang, Shivani Mall, Li Fei-Fei, and Jiajun Wu. ObjectFolder: A dataset of objects with implicit visual, auditory, and tactile representations. In CoRL, 2021. 2 +[24] Dibya Ghosh, Abhishek Gupta, Ashwin Reddy, Justin Fu, Coline Devin, Benjamin Eysenbach, and Sergey Levine. Learning to reach goals via iterated supervised learning. In ICML, 2019. 8 +[25] Roger Grosse, Micah K Johnson, Edward H Adelson, and + +William T Freeman. Ground truth dataset and baseline evaluations for intrinsic image algorithms. In ICCV, 2009. 3 +[26] Qiao Gu, Zhaoyang Lv, Duncan Frost, Simon Green, Julian Straub, and Chris Sweeney. Egolifter: Open-world 3d segmentation for egocentric perception. In ECCV, pages 382-400. Springer, 2024. 7 +[27] Agrim Gupta, Piotr Dollar, and Ross Girshick. LVIS: A dataset for large vocabulary instance segmentation. In CVPR, 2019. 4 +[28] Wassim Hamidouche, Lina Bariah, and Mérouane Debbah. Immersive media and massive twinning: Advancing toward the metaverse. IEEE Communications Magazine, 62(7):20-32, 2024. 2 +[29] Jon Hasselgren, Nikolai Hofmann, and Jacob Munkberg. Shape, light & material decomposition from images using Monte Carlo rendering and denoising. In NeurIPS, 2022. 4, 7 +[30] Yicong Hong, Kai Zhang, Jiuxiang Gu, Sai Bi, Yang Zhou, Difan Liu, Feng Liu, Kalyan Sunkavalli, Trung Bui, and Hao Tan. LRM: Large reconstruction model for single image to 3D. In The Twelfth International Conference on Learning Representations, 2024. 2 +[31] Binbin Huang, Zehao Yu, Anpei Chen, Andreas Geiger, and Shenghua Gao. 2D gaussian splatting for geometrically accurate radiance fields. In SIGGRAPH 2024 Conference Papers, 2024. 3, 7 +[32] Wenzel Jakob, Sébastien Speierer, Nicolas Roussel, Merlin Nimier-David, Delio Vicini, Tizian Zeltner, Baptiste Nicolet, Miguel Crespo, Vincent Leroy, and Ziyi Zhang. Mitsuba 3 renderer, 2022. https://mitsuba-renderer.org. 5 +[33] Michael Janner, Jiajun Wu, Tejas Kulkarni, Ilker Yildirim, and Joshua B Tenenbaum. Self-supervised intrinsic image decomposition. In NeurIPS, 2017. 4 +[34] Rasmus Jensen, Anders Dahl, George Vogiatzis, Engin Tola, and Henrik Aanæs. Large scale multi-view stereopsis evaluation. In CVPR, 2014. 2, 3 +[35] Haian Jin, Isabella Liu, Peijia Xu, Xiaoshuai Zhang, Songfang Han, Sai Bi, Xiaowei Zhou, Zexiang Xu, and Hao Su. TensoIR: Tensorial inverse rendering. In CVPR, 2023. 2 +[36] James T Kajiya. The rendering equation. In Proceedings of the 13th annual conference on Computer graphics and interactive techniques, pages 143-150, 1986. 3 +[37] Bernhard Kerbl, Georgios Kopanas, Thomas Leimkuhler, and George Drettakis. 3D gaussian splattering for real-time radiance field rendering. ACM TOG, 2023. 2, 3, 7 +[38] Zhengfei Kuang, Kyle Olszewski, Mengelei Chai, Zeng Huang, Panos Achlioptas, and Sergey Tulyakov. NeROIC: neural rendering of objects from online image collections. ACM TOG, 41(4):1-12, 2022. 3 +[39] Zhengfei Kuang, Yunzhi Zhang, Hong-Xing Yu, Samir Agarwala, Elliott Wu, Jiajun Wu, et al. Stanford-orb: a real-world 3D object inverse rendering benchmark. In Advances in Neural Information Processing Systems Datasets and Benchmarks Track, 2023. 3, 5, 7, 8 +[40] Samuli Laine, Janne Hellsten, Tero Karras, Yeongho Seol, Jaakko Lehtinen, and Timo Aila. Modular primitives for high-performance differentiable rendering. ACM Transactions on Graphics, 39(6), 2020. 5 + +[41] Jiahao Li, Hao Tan, Kai Zhang, Zexiang Xu, Fujun Luan, Yinghao Xu, Yicong Hong, Kalyan Sunkavalli, Greg Shakhnarovich, and Sai Bi. Instant3D: Fast text-to-3D with sparse-view generation and large reconstruction model. In ICLR, 2024. 2 +[42] Min Li, Zhenglong Zhou, Zhe Wu, Boxin Shi, Changyu Diao, and Ping Tan. Multi-view photometric stereo: A robust solution and benchmark dataset for spatially varying isotropic materials. IEEE TIP, 29:4159-4173, 2020. 3 +[43] Xuanlin Li, Kyle Hsu, Jiayuan Gu, Karl Pertsch, Oier Mees, Homer Rich Walke, Chuyuan Fu, Ishikaa Lunawat, Isabel Sieh, Sean Kirmani, Sergey Levine, Jiajun Wu, Chelsea Finn, Hao Su, Quan Vuong, and Ted Xiao. Evaluating real-world robot manipulation policies in simulation. arXiv preprint arXiv:2405.05941, 2024. 2 +[44] Zhengqin Li, Kalyan Sunkavalli, and Manmohan Chandraker. Materials for masses: SVBRDF acquisition with a single mobile phone image. In ECCV, 2018. 3 +[45] Zhengqin Li, Zexiang Xu, Ravi Ramamoorthi, Kalyan Sunkavalli, and Manmohan Chandraker. Learning to reconstruct shape and spatially-varying reflectance from a single image. ACM TOG, 37(6):1-11, 2018. 3, 4 +[46] Isabella Liu, Linghao Chen, Ziyang Fu, Liwen Wu, Haian Jin, Zhong Li, Chin Ming Ryan Wong, Yi Xu, Ravi Ramamoorthi, Zexiang Xu, and Hao Su. OpenIllumination: A multi-illumination dataset for inverse rendering evaluation on real objects. NeurIPS, 2023. 2, 3 +[47] Stephen Lombardi and Ko Nishino. Reflectance and natural illumination from a single image. In ECCV, 2012. 3 +[48] Stephen Lombardi and Ko Nishino. Reflectance and illumination recovery in the wild. IEEE TPAMI, 2015. 3 +[49] Fujun Luan, Shuang Zhao, Kavita Bala, and Zhao Dong. Unified shape and SVBRDF recovery using differentiable Monte Carlo rendering. CGF, 40(4):101-113, 2021. 3 +[50] Stephen Robert Marschner. Inverse Rendering for Computer Graphics. Cornell University, 1998. 3 +[51] Nelson Max. Optical models for direct volume rendering. IEEE TVCG, 1(2):99-108, 1995. 3 +[52] Ben Mildenhall, Pratul P Srinivasan, Matthew Tancik, Jonathan T Barron, Ravi Ramamoorthi, and Ren Ng. NeRF: Representing scenes as neural radiance fields for view synthesis. In ECCV, 2020. 2, 3, 7 +[53] Jacob Munkberg, Jon Hasselgren, Tianchang Shen, Jun Gao, Wenzheng Chen, Alex Evans, Thomas Müller, and Sanja Fidler. Extracting triangular 3D models, materials, and lighting from images. In CVPR, 2022. 4, 7 +[54] Soroush Nasiriany, Abhiram Maddukuri, Lance Zhang, Adeet Parikh, Aaron Lo, Abhishek Joshi, Ajay Mandlekar, and Yuke Zhu. Robocasa: Large-scale simulation of everyday tasks for generalist robots. In Robotics: Science and Systems (RSS), 2024. 7 +[55] Michael Niemeyer, Lars Mescheder, Michael Oechsle, and Andreas Geiger. Differentiable volumetric rendering: Learning implicit 3D representations without 3D supervision. In CVPR, 2020. 4 +[56] Geoffrey Oxholm and Ko Nishino. Multiview shape and reflectance from natural illumination. In CVPR, 2014. 3 + +[57] Geoffrey Oxholm and Ko Nishino. Shape and reflectance estimation in the wild. IEEE TPAMI, 2015. 3 +[58] Xiaqing Pan, Nicholas Charron, Yongqian Yang, Scott Peters, Thomas Whelan, Chen Kong, Omkar Parkhi, Richard Newcombe, and Yuheng (Carl) Ren. Aria digital twin: A new benchmark dataset for egocentric 3D machine perception. In Proceedings of the IEEE/CVF International Conference on Computer Vision (ICCV), pages 20133-20143, 2023. 3 +[59] Viktor Rudnev, Mohamed Elgharib, William Smith, Lingjie Liu, Vladislav Golyanik, and Christian Theobalt. NeRF for outdoor scene relighting. In ECCV, pages 615-631. Springer, 2022. 3 +[60] Tianchang Shen, Jun Gao, Kangxue Yin, Ming-Yu Liu, and Sanja Fidler. Deep marching tetrahedra: a hybrid representation for high-resolution 3D shape synthesis. In NeurIPS, 2021. 7 +[61] Jian Shi, Yue Dong, Hao Su, and Stella X Yu. Learning non-Lambertian object intrinsics across ShapeNet categories. In CVPR, 2017. 2, 3, 4 +[62] Cheng Sun, Guangyan Cai, Zhengqin Li, Kai Yan, Cheng Zhang, Carl Marshall, Jia-Bin Huang, Shuang Zhao, and Zhao Dong. Neural-PBIR reconstruction of shape, material, and illumination. In ICCV, 2023. 2, 4, 5, 6 +[63] Stephen Tian, Yancheng Cai, Hong-Xing Yu, Sergey Zakharov, Katherine Liu, Adrien Gaidon, Yunzhu Li, and Ji-jun Wu. Multi-object manipulation via object-centric neural scattering functions. In CVPR, 2023. 7 +[64] Marcel Torne, Anthony Simeonov, Zechu Li, April Chan, Tao Chen, Abhishek Gupta, and Pulkit Agrawal. Reconciling reality through simulation: A real-to-sim-to-real approach for robust manipulation. Arxiv, 2024. 7 +[65] Marco Toschi, Riccardo De Matteo, Riccardo Spezialetti, Daniele De Gregorio, Luigi Di Stefano, and Samuele Salti. Relight my NeRF: A dataset for novel view synthesis and relighting of real world objects. In CVPR, 2023. 3 +[66] Peng Wang, Lingjie Liu, Yuan Liu, Christian Theobalt, Taku Komura, and Wenping Wang. NeuS: Learning neural implicit surfaces by volume rendering for multi-view reconstruction. In NeurIPS, 2021. 4 +[67] Henrique Weber, Donald Prévost, and Jean-François Lalonde. Learning to estimate indoor lighting from 3D objects. In 3DV, 2018. 3 +[68] Xinyue Wei, Minghua Liu, Zhan Ling, and Hao Su. Approximate convex decomposition for 3D meshes with collision-aware concavity and tree search. ACM TOG, 2022. 7 +[69] Xinyue Wei, Kai Zhang, Sai Bi, Hao Tan, Fujun Luan, Valentin Deschaintre, Kalyan Sunkavalli, Hao Su, and Zexiang Xu. MeshLRM: Large reconstruction model for high-quality mesh. arXiv preprint arXiv:2404.12385, 2024. 2 +[70] Felix Wimbauer, Shangzhe Wu, and Christian Rupprecht. De-rendering 3D objects in the wild. In CVPR, 2022. 3, 4 +[71] Haoqian Wu, Zhipeng Hu, Lincheng Li, Yongqiang Zhang, Changjie Fan, and Xin Yu. NeFII: Inverse rendering for reflectance decomposition with near-field indirect illumination. In CVPR, 2023. 4, 7 +[72] Shangzhe Wu, Ameesh Makadia, Jiajun Wu, Noah Snavely, Richard Tucker, and Angjoo Kanazawa. De-rendering the world's revolutionary artefacts. In CVPR, 2021. 3, 4 + +[73] Tong Wu, Jiarui Zhang, Xiao Fu, Yuxin Wang, Liang Pan Jiawei Ren, Wayne Wu, Lei Yang, Jiaqi Wang, Chen Qian, Dahua Lin, and Ziwei Liu. OmniObject3D: Large-vocabulary 3D object dataset for realistic perception, reconstruction and generation. In IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), 2023. 2, 3 +[74] Kohei Yamashita, Yuto Enyo, Shohei Nobuhara, and Ko Nishino. nLMVS-Net: Deep non-lambertian multi-view stereo. In WACV, 2023. 3 +[75] Kai Yan and Shuang Zhao. PSDR-JIT: Path-space differentiable renderer, 2020. https://github.com/andyyankai/psdr-jit.5 +[76] Lior Yariv, Yoni Kasten, Dror Moran, Meirav Galun, Matan Atzmon, Basri Ronen, and Yaron Lipman. Multiview neural surface reconstruction by disentangling geometry and appearance. In NeurIPS, 2020. 7 +[77] Vickie Ye, Ruilong Li, Justin Kerr, Matias Turkulainen, Brent Yi, Zhuoyang Pan, Otto Seiskari, Jianbo Ye, Jeffrey Hu, Matthew Tancik, and Angjoo Kanazawa. gsplat: An open-source library for Gaussian splatting. arXiv preprint arXiv:2409.06765, 2024. 7 +[78] Hong-Xing Yu, Samir Agarwala, Charles Herrmann, Richard Szeliski, Noah Snavely, Jiajun Wu, and Deqing Sun. Accidental light probes. In CVPR, 2023. 3, 5 +[79] Hong-Xing Yu, Michelle Guo, Alireza Fathi, Yen-Yu Chang, Eric Ryan Chan, Ruohan Gao, Thomas Funkhouser, and Jia-jun Wu. Learning object-centric neural scattering functions for free-viewpoint relighting and scene composition. TMLR, 2023. 3 +[80] Kai Zhang, Fujun Luan, Qianqian Wang, Kavita Bala, and Noah Snavely. PhySG: Inverse rendering with spherical gaussians for physics-based material editing and relighting. In CVPR, 2021. 4, 7 +[81] Kai Zhang, Fujun Luan, Zhengqi Li, and Noah Snavely. IRON: Inverse rendering by optimizing neural SDFs and materials from photometric images. In CVPR, 2022. 4 +[82] Kai Zhang, Sai Bi, Hao Tan, Yuanbo Xiangli, Nanxuan Zhao, Kalyan Sunkavalli, and Zexiang Xu. GS-LRM: Large reconstruction model for 3D gaussian splatting. arXiv preprint arXiv:2404.19702, 2024. 2 +[83] Ruo Zhang, Ping-Sing Tsai, James Edwin Cryer, and Mubarak Shah. Shape-from-shading: a survey. IEEE TPAMI, 21(8): 690-706, 1999. 3 +[84] Xiuming Zhang, Pratul P Srinivasan, Boyang Deng, Paul Debevec, William T Freeman, and Jonathan T Barron. NeR-Factor: Neural factorization of shape and reflectance under an unknown illumination. ACM TOG, 40(6):1-18, 2021. 3 +[85] Yuanqing Zhang, Jiaming Sun, Xingyi He, Huan Fu, Rongfei Jia, and Xiaowei Zhou. Modeling indirect illumination for inverse rendering. In CVPR, 2022. 4, 7 +[86] Yunzhi Zhang, Shangzhe Wu, Noah Snavely, and Jiajun Wu. Seeing a rose in five thousand ways. In CVPR, 2023. 4 \ No newline at end of file diff --git a/2025/Digital Twin Catalog_ A Large-Scale Photorealistic 3D Object Digital Twin Dataset/images.zip b/2025/Digital Twin Catalog_ A Large-Scale Photorealistic 3D Object Digital Twin Dataset/images.zip new file mode 100644 index 0000000000000000000000000000000000000000..e0276441f9e7fce01d3d358b7d040270f5ee6d0e --- /dev/null +++ b/2025/Digital Twin Catalog_ A Large-Scale Photorealistic 3D Object Digital Twin Dataset/images.zip @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:f8919d7206fcd5af8effb8696e227f6594c533c37ceb21802631b135809a95ec +size 592796 diff --git a/2025/Digital Twin Catalog_ A Large-Scale Photorealistic 3D Object Digital Twin Dataset/layout.json b/2025/Digital Twin Catalog_ A Large-Scale Photorealistic 3D Object Digital Twin Dataset/layout.json new file mode 100644 index 0000000000000000000000000000000000000000..69683d9c9c1e00a37d2a4a5d8999b0fa6478a59e --- /dev/null +++ b/2025/Digital Twin Catalog_ A Large-Scale Photorealistic 3D Object Digital Twin Dataset/layout.json @@ -0,0 +1,8461 @@ +{ + "pdf_info": [ + { + "para_blocks": [ + { + "bbox": [ + 56, + 103, + 555, + 121 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 103, + 555, + 121 + ], + "spans": [ + { + "bbox": [ + 56, + 103, + 555, + 121 + ], + "type": "text", + "content": "Digital Twin Catalog: A Large-Scale Photorealistic 3D Object Digital Twin Dataset" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 62, + 142, + 547, + 186 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 142, + 547, + 186 + ], + "spans": [ + { + "bbox": [ + 62, + 142, + 547, + 186 + ], + "type": "text", + "content": "Zhao Dong" + }, + { + "bbox": [ + 62, + 142, + 547, + 186 + ], + "type": "inline_equation", + "content": "^{1}" + }, + { + "bbox": [ + 62, + 142, + 547, + 186 + ], + "type": "text", + "content": " Ka Chen" + }, + { + "bbox": [ + 62, + 142, + 547, + 186 + ], + "type": "inline_equation", + "content": "^{1,*}" + }, + { + "bbox": [ + 62, + 142, + 547, + 186 + ], + "type": "text", + "content": " Zhaoyang Lv" + }, + { + "bbox": [ + 62, + 142, + 547, + 186 + ], + "type": "inline_equation", + "content": "^{1,*}" + }, + { + "bbox": [ + 62, + 142, + 547, + 186 + ], + "type": "text", + "content": " Hong-Xing Yu" + }, + { + "bbox": [ + 62, + 142, + 547, + 186 + ], + "type": "inline_equation", + "content": "^{2,*}" + }, + { + "bbox": [ + 62, + 142, + 547, + 186 + ], + "type": "text", + "content": " Yunzhi Zhang" + }, + { + "bbox": [ + 62, + 142, + 547, + 186 + ], + "type": "inline_equation", + "content": "^{2,*}" + }, + { + "bbox": [ + 62, + 142, + 547, + 186 + ], + "type": "text", + "content": " Cheng Zhang" + }, + { + "bbox": [ + 62, + 142, + 547, + 186 + ], + "type": "inline_equation", + "content": "^{1,*}" + }, + { + "bbox": [ + 62, + 142, + 547, + 186 + ], + "type": "text", + "content": " Yufeng Zhu" + }, + { + "bbox": [ + 62, + 142, + 547, + 186 + ], + "type": "inline_equation", + "content": "^{1,*}" + }, + { + "bbox": [ + 62, + 142, + 547, + 186 + ], + "type": "text", + "content": " Stephen Tian" + }, + { + "bbox": [ + 62, + 142, + 547, + 186 + ], + "type": "inline_equation", + "content": "^{2}" + }, + { + "bbox": [ + 62, + 142, + 547, + 186 + ], + "type": "text", + "content": " Zhengqin Li" + }, + { + "bbox": [ + 62, + 142, + 547, + 186 + ], + "type": "inline_equation", + "content": "^{1}" + }, + { + "bbox": [ + 62, + 142, + 547, + 186 + ], + "type": "text", + "content": " Georgie Moffatt" + }, + { + "bbox": [ + 62, + 142, + 547, + 186 + ], + "type": "inline_equation", + "content": "^{1}" + }, + { + "bbox": [ + 62, + 142, + 547, + 186 + ], + "type": "text", + "content": " Sean Christofferson" + }, + { + "bbox": [ + 62, + 142, + 547, + 186 + ], + "type": "inline_equation", + "content": "^{1}" + }, + { + "bbox": [ + 62, + 142, + 547, + 186 + ], + "type": "text", + "content": " James Fort" + }, + { + "bbox": [ + 62, + 142, + 547, + 186 + ], + "type": "inline_equation", + "content": "^{1}" + }, + { + "bbox": [ + 62, + 142, + 547, + 186 + ], + "type": "text", + "content": " Xiaqing Pan" + }, + { + "bbox": [ + 62, + 142, + 547, + 186 + ], + "type": "inline_equation", + "content": "^{1}" + }, + { + "bbox": [ + 62, + 142, + 547, + 186 + ], + "type": "text", + "content": " Mingfei Yan" + }, + { + "bbox": [ + 62, + 142, + 547, + 186 + ], + "type": "inline_equation", + "content": "^{1}" + }, + { + "bbox": [ + 62, + 142, + 547, + 186 + ], + "type": "text", + "content": " Jiajun Wu" + }, + { + "bbox": [ + 62, + 142, + 547, + 186 + ], + "type": "inline_equation", + "content": "^{2}" + }, + { + "bbox": [ + 62, + 142, + 547, + 186 + ], + "type": "text", + "content": " Carl Yuheng Ren" + }, + { + "bbox": [ + 62, + 142, + 547, + 186 + ], + "type": "inline_equation", + "content": "^{1}" + }, + { + "bbox": [ + 62, + 142, + 547, + 186 + ], + "type": "text", + "content": " Richard Newcombe" + }, + { + "bbox": [ + 62, + 142, + 547, + 186 + ], + "type": "inline_equation", + "content": "^{1}" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 176, + 193, + 434, + 209 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 176, + 193, + 434, + 209 + ], + "spans": [ + { + "bbox": [ + 176, + 193, + 434, + 209 + ], + "type": "inline_equation", + "content": "^{1}" + }, + { + "bbox": [ + 176, + 193, + 434, + 209 + ], + "type": "text", + "content": "Meta Reality Labs Research Stanford University" + } + ] + } + ], + "index": 4 + }, + { + "type": "image", + "bbox": [ + 60, + 246, + 265, + 452 + ], + "blocks": [ + { + "bbox": [ + 60, + 246, + 265, + 452 + ], + "lines": [ + { + "bbox": [ + 60, + 246, + 265, + 452 + ], + "spans": [ + { + "bbox": [ + 60, + 246, + 265, + 452 + ], + "type": "image", + "image_path": "5c6ef2dab8e54265416509670d0062d9671cdfe6f8dc0a4a7cec32c62d6f5d97.jpg" + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 56, + 457, + 269, + 468 + ], + "lines": [ + { + "bbox": [ + 56, + 457, + 269, + 468 + ], + "spans": [ + { + "bbox": [ + 56, + 457, + 269, + 468 + ], + "type": "text", + "content": "(a) Examples of 2,000 digital twins of real-world objects" + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "image_caption" + } + ], + "index": 5 + }, + { + "type": "image", + "bbox": [ + 283, + 244, + 369, + 452 + ], + "blocks": [ + { + "bbox": [ + 283, + 244, + 369, + 452 + ], + "lines": [ + { + "bbox": [ + 283, + 244, + 369, + 452 + ], + "spans": [ + { + "bbox": [ + 283, + 244, + 369, + 452 + ], + "type": "image", + "image_path": "63ea79faa2c911314cac9024de6bb70ec75275098cc717280376a3c205019268.jpg" + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 276, + 457, + 376, + 468 + ], + "lines": [ + { + "bbox": [ + 276, + 457, + 376, + 468 + ], + "spans": [ + { + "bbox": [ + 276, + 457, + 376, + 468 + ], + "type": "text", + "content": "(b) Shape & PBR Materials" + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "image_caption" + } + ], + "index": 7 + }, + { + "type": "image", + "bbox": [ + 391, + 236, + 465, + 452 + ], + "blocks": [ + { + "bbox": [ + 391, + 236, + 465, + 452 + ], + "lines": [ + { + "bbox": [ + 391, + 236, + 465, + 452 + ], + "spans": [ + { + "bbox": [ + 391, + 236, + 465, + 452 + ], + "type": "image", + "image_path": "7d21347e865b975546c4908b004218877b7eed45912281de5bc1d261d7ec61c1.jpg" + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 402, + 457, + 456, + 468 + ], + "lines": [ + { + "bbox": [ + 402, + 457, + 456, + 468 + ], + "spans": [ + { + "bbox": [ + 402, + 457, + 456, + 468 + ], + "type": "text", + "content": "(c) DSLR Data" + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 54, + 476, + 555, + 510 + ], + "lines": [ + { + "bbox": [ + 54, + 476, + 555, + 510 + ], + "spans": [ + { + "bbox": [ + 54, + 476, + 555, + 510 + ], + "type": "text", + "content": "Figure 1. The Digital Twin Catalog (DTC) dataset comprises 2,000 digital twins of physical-world objects (a), characterized by millimeter-level geometric accuracy and photorealistic PBR materials (b). DTC includes evaluation data captured using both DSLR cameras and egocentric Aria glasses, featuring captured images with precise foreground object masks and environment lighting for relighting evaluation." + } + ] + } + ], + "index": 13, + "angle": 0, + "type": "image_caption" + } + ], + "index": 9 + }, + { + "type": "image", + "bbox": [ + 474, + 245, + 555, + 449 + ], + "blocks": [ + { + "bbox": [ + 474, + 245, + 555, + 449 + ], + "lines": [ + { + "bbox": [ + 474, + 245, + 555, + 449 + ], + "spans": [ + { + "bbox": [ + 474, + 245, + 555, + 449 + ], + "type": "image", + "image_path": "8ddbce2402e3d85449771a91d3f6fe0241663684d598440c2308bf4ff77860b6.jpg" + } + ] + } + ], + "index": 11, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 477, + 457, + 552, + 468 + ], + "lines": [ + { + "bbox": [ + 477, + 457, + 552, + 468 + ], + "spans": [ + { + "bbox": [ + 477, + 457, + 552, + 468 + ], + "type": "text", + "content": "(d) Egocentric Data" + } + ] + } + ], + "index": 12, + "angle": 0, + "type": "image_caption" + } + ], + "index": 11 + }, + { + "bbox": [ + 151, + 517, + 200, + 529 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 151, + 517, + 200, + 529 + ], + "spans": [ + { + "bbox": [ + 151, + 517, + 200, + 529 + ], + "type": "text", + "content": "Abstract" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 54, + 544, + 297, + 689 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 54, + 544, + 297, + 689 + ], + "spans": [ + { + "bbox": [ + 54, + 544, + 297, + 689 + ], + "type": "text", + "content": "We introduce Digital Twin Catalog (DTC), a new large-scale photorealistic 3D object digital twin dataset. A digital twin of a 3D object is a highly detailed, virtually indistinguishable representation of a physical object, accurately capturing its shape, appearance, physical properties, and other attributes. Recent advances in neural-based 3D reconstruction and inverse rendering have significantly improved the quality of 3D object reconstruction. Despite these advancements, there remains a lack of a large-scale, digital twin quality real-world dataset and benchmark that can quantitatively assess and compare the performance of different reconstruction methods, as well as improve reconstruction quality through" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 313, + 518, + 556, + 698 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 518, + 556, + 698 + ], + "spans": [ + { + "bbox": [ + 313, + 518, + 556, + 698 + ], + "type": "text", + "content": "training or fine-tuning. Moreover, to democratize 3D digital twin creation, it is essential to integrate creation techniques with next-generation egocentric computing platforms, such as AR glasses. Currently, there is no dataset available to evaluate 3D object reconstruction using egocentric captured images. To address these gaps, the DTC dataset features 2,000 scanned digital twin-quality 3D objects, along with image sequences captured under different lighting conditions using DSLR cameras and egocentric AR glasses. This dataset establishes the first comprehensive real-world evaluation benchmark for 3D digital twin creation tasks, offering a robust foundation for comparing and improving existing reconstruction methods. The DTC dataset is already released at https://www.projectaria.com/datasets/dtc/ and we will also make the baseline evaluations open-source." + } + ] + } + ], + "index": 16 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "spans": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "type": "text", + "content": "CVF" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 145, + 0, + 496, + 35 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 145, + 0, + 496, + 35 + ], + "spans": [ + { + "bbox": [ + 145, + 0, + 496, + 35 + ], + "type": "text", + "content": "This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 70, + 702, + 280, + 712 + ], + "type": "page_footnote", + "angle": 0, + "lines": [ + { + "bbox": [ + 70, + 702, + 280, + 712 + ], + "spans": [ + { + "bbox": [ + 70, + 702, + 280, + 712 + ], + "type": "text", + "content": "*Authors contributed equally and are listed in alphabetical order." + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 297, + 749, + 312, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 297, + 749, + 312, + 757 + ], + "spans": [ + { + "bbox": [ + 297, + 749, + 312, + 757 + ], + "type": "text", + "content": "753" + } + ] + } + ], + "index": 18 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 0 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 56, + 70, + 154, + 169 + ], + "blocks": [ + { + "bbox": [ + 56, + 70, + 154, + 169 + ], + "lines": [ + { + "bbox": [ + 56, + 70, + 154, + 169 + ], + "spans": [ + { + "bbox": [ + 56, + 70, + 154, + 169 + ], + "type": "image", + "image_path": "787532f76f3c7d05c010f997c86ce2161ca34f8cd6a082de3e89196c9b445026.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + } + ], + "index": 0 + }, + { + "type": "image", + "bbox": [ + 156, + 70, + 254, + 169 + ], + "blocks": [ + { + "bbox": [ + 156, + 70, + 254, + 169 + ], + "lines": [ + { + "bbox": [ + 156, + 70, + 254, + 169 + ], + "spans": [ + { + "bbox": [ + 156, + 70, + 254, + 169 + ], + "type": "image", + "image_path": "7ac0bd5a69b106213fb119b8eb13f0f3640675ae19d597720bfa6b4fffd7aae1.jpg" + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_body" + } + ], + "index": 1 + }, + { + "type": "image", + "bbox": [ + 255, + 70, + 354, + 169 + ], + "blocks": [ + { + "bbox": [ + 255, + 70, + 354, + 169 + ], + "lines": [ + { + "bbox": [ + 255, + 70, + 354, + 169 + ], + "spans": [ + { + "bbox": [ + 255, + 70, + 354, + 169 + ], + "type": "image", + "image_path": "03814b73dbaeeceaca5b864b789e7d1bd77546a4fce1c8fba05868b819ffbd6a.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + } + ], + "index": 2 + }, + { + "type": "image", + "bbox": [ + 355, + 70, + 454, + 169 + ], + "blocks": [ + { + "bbox": [ + 355, + 70, + 454, + 169 + ], + "lines": [ + { + "bbox": [ + 355, + 70, + 454, + 169 + ], + "spans": [ + { + "bbox": [ + 355, + 70, + 454, + 169 + ], + "type": "image", + "image_path": "162bee86384d06d222c9ba1a0dd1d208552d44bf41cf0d1b69620172067937cf.jpg" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_body" + } + ], + "index": 3 + }, + { + "type": "image", + "bbox": [ + 454, + 71, + 553, + 169 + ], + "blocks": [ + { + "bbox": [ + 454, + 71, + 553, + 169 + ], + "lines": [ + { + "bbox": [ + 454, + 71, + 553, + 169 + ], + "spans": [ + { + "bbox": [ + 454, + 71, + 553, + 169 + ], + "type": "image", + "image_path": "4b26a4302ba3923f69965b63cf545b83d58552ac9cb9e125720cd32dd3681f09.jpg" + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_body" + } + ], + "index": 4 + }, + { + "type": "image", + "bbox": [ + 56, + 171, + 154, + 269 + ], + "blocks": [ + { + "bbox": [ + 56, + 171, + 154, + 269 + ], + "lines": [ + { + "bbox": [ + 56, + 171, + 154, + 269 + ], + "spans": [ + { + "bbox": [ + 56, + 171, + 154, + 269 + ], + "type": "image", + "image_path": "c1f9ec98603ce255dfa5b73ba916b26a6b329bd7b9fb068f4a227c67879135ff.jpg" + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "image_body" + } + ], + "index": 5 + }, + { + "type": "image", + "bbox": [ + 156, + 171, + 254, + 269 + ], + "blocks": [ + { + "bbox": [ + 156, + 171, + 254, + 269 + ], + "lines": [ + { + "bbox": [ + 156, + 171, + 254, + 269 + ], + "spans": [ + { + "bbox": [ + 156, + 171, + 254, + 269 + ], + "type": "image", + "image_path": "c3f63b0fd71a76659fed8016ddb3d7ca9f3f5354904b2712fcdf92e3b6356eae.jpg" + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 182, + 277, + 427, + 289 + ], + "lines": [ + { + "bbox": [ + 182, + 277, + 427, + 289 + ], + "spans": [ + { + "bbox": [ + 182, + 277, + 427, + 289 + ], + "type": "text", + "content": "Figure 2. Example DTC models with photorealistic PBR materials." + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "image_caption" + } + ], + "index": 6 + }, + { + "type": "image", + "bbox": [ + 255, + 171, + 354, + 269 + ], + "blocks": [ + { + "bbox": [ + 255, + 171, + 354, + 269 + ], + "lines": [ + { + "bbox": [ + 255, + 171, + 354, + 269 + ], + "spans": [ + { + "bbox": [ + 255, + 171, + 354, + 269 + ], + "type": "image", + "image_path": "c742133f79593415e6754d2c6e6b65a5b75e0dc5457fb9ba9d20bff3338b32a0.jpg" + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "image_body" + } + ], + "index": 7 + }, + { + "type": "image", + "bbox": [ + 355, + 171, + 454, + 269 + ], + "blocks": [ + { + "bbox": [ + 355, + 171, + 454, + 269 + ], + "lines": [ + { + "bbox": [ + 355, + 171, + 454, + 269 + ], + "spans": [ + { + "bbox": [ + 355, + 171, + 454, + 269 + ], + "type": "image", + "image_path": "6d2d671251186c923beaff28ce7b2a88d1983ab8b524b097b52ad10d187569e7.jpg" + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "image_body" + } + ], + "index": 8 + }, + { + "type": "image", + "bbox": [ + 454, + 171, + 553, + 269 + ], + "blocks": [ + { + "bbox": [ + 454, + 171, + 553, + 269 + ], + "lines": [ + { + "bbox": [ + 454, + 171, + 553, + 269 + ], + "spans": [ + { + "bbox": [ + 454, + 171, + 553, + 269 + ], + "type": "image", + "image_path": "33f163d2d32987871165fce759381ea7ecfb7b5057c30b7527a1a50860194a9f.jpg" + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "image_body" + } + ], + "index": 9 + }, + { + "bbox": [ + 56, + 308, + 136, + 320 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 308, + 136, + 320 + ], + "spans": [ + { + "bbox": [ + 56, + 308, + 136, + 320 + ], + "type": "text", + "content": "1. Introduction" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 55, + 329, + 297, + 616 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 329, + 297, + 616 + ], + "spans": [ + { + "bbox": [ + 55, + 329, + 297, + 616 + ], + "type": "text", + "content": "A digital twin of a 3D object is a highly detailed, virtually indistinguishable representation of a physical object, capturing its shape, appearance, physical properties and other attributes with precision. Such a digital twin enables visualization, analysis, and interaction as if it were the real object, supporting simulation, automation, and real-world problem-solving across a wide range of applications in AR/VR [3, 28], spatial/contextual AI [2], and robotics [23, 43]. As fundamental properties of an object, its shape and appearance form the basis for recognizing and interpreting the 3D object, enabling identification, manipulation, and realistic rendering. Recovering these attributes has long been a foundational topic in computer vision and graphics, inspiring extensive research in 3D reconstruction and inverse rendering. Recent breakthroughs in neural-based representation and reconstruction techniques, such as NeRF [52] and 3D Gaussian splattig (3DGS) [37], have significantly elevated the quality of novel view synthesis (NVS) to photorealistic levels. Many subsequent works [35, 62] integrate neural reconstruction with physically-based inverse rendering, enabling relightable appearances. Furthermore, leveraging priors from large reconstruction models (LRMs) [30], high-quality shape and appearance reconstruction can now be achieved with as few as one to four views [41, 69, 82]." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 55, + 618, + 297, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 618, + 297, + 714 + ], + "spans": [ + { + "bbox": [ + 55, + 618, + 297, + 714 + ], + "type": "text", + "content": "Despite the rapid advancements in 3D object reconstruction, one question remains: does the reconstruction quality truly meet the standard of a digital twin, where virtual representations are indistinguishable from reality? This digital twin standard demands both highly accurate shape matching and photorealistic appearance across different lighting, which present significant acquisition challenges for real-world objects. Existing object-centric datasets for 3D recon" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 313, + 309, + 556, + 430 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 309, + 556, + 430 + ], + "spans": [ + { + "bbox": [ + 313, + 309, + 556, + 430 + ], + "type": "text", + "content": "struction or inverse rendering have focused on either dataset size [18] or quality of specific aspects [16, 21, 34, 46, 61, 73], often sacrificing comprehensive fidelity and limiting their application scope. This trade-off has led to a lack of datasets that fully satisfy the digital twin criteria, hindering current 3D reconstruction methods from achieving digital twin fidelity. To bridge this gap, we developed the Digital Twin Catalog (DTC) dataset, comprising 2,000 scanned 3D object models (Fig. 1(a)), each with millimeter geometry accuracy and photorealistic PBR materials (Fig. 1(b), Fig. 2)." + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 313, + 434, + 557, + 626 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 434, + 557, + 626 + ], + "spans": [ + { + "bbox": [ + 313, + 434, + 557, + 626 + ], + "type": "text", + "content": "In addition to 3D digital twin models, the DTC dataset includes evaluation data designed to support 3D object reconstruction research. This evaluation data features multiview image sequences with precise foreground object masks and environment lighting information for relighting evaluation. Traditionally, high-quality HDR images captured with modern DSLR cameras have been the standard for 3D reconstruction research. Looking ahead, we encourage the integration of 3D reconstruction research with next-generation human-centric computing platforms, such as egocentric AR glasses, aiming to democratize 3D reconstruction techniques and empower everyone to effortlessly create 3D digital twins. To this end, alongside DSLR-captured evaluation data (Fig. 1(c)), the DTC dataset also provides egocentric evaluation data captured using Project Aria glasses (https://www.projectaria.com) (Fig. 1(d))." + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 313, + 629, + 558, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 629, + 558, + 715 + ], + "spans": [ + { + "bbox": [ + 313, + 629, + 558, + 715 + ], + "type": "text", + "content": "The DTC dataset offers extensive opportunities for advancing research in object digital twin creation. We provide a benchmark for state-of-the-art 3D object reconstruction and inverse rendering methods. These benchmarks evaluate performance across novel view synthesis (NVS), shape reconstruction, and relightable appearance reconstruction. We further provide the evaluation of novel view synthesis meth" + } + ] + } + ], + "index": 16 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 297, + 748, + 312, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 297, + 748, + 312, + 757 + ], + "spans": [ + { + "bbox": [ + 297, + 748, + 312, + 757 + ], + "type": "text", + "content": "754" + } + ] + } + ], + "index": 17 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 1 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 58, + 102, + 552, + 293 + ], + "blocks": [ + { + "bbox": [ + 55, + 70, + 555, + 94 + ], + "lines": [ + { + "bbox": [ + 55, + 70, + 555, + 94 + ], + "spans": [ + { + "bbox": [ + 55, + 70, + 555, + 94 + ], + "type": "text", + "content": "Table 1. Comparison with existing object-centric inverse rendering datasets. *Objaverse [18] consists of both synthetic objects and real scans, only part of which contain PBR materials." + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 58, + 102, + 552, + 293 + ], + "lines": [ + { + "bbox": [ + 58, + 102, + 552, + 293 + ], + "spans": [ + { + "bbox": [ + 58, + 102, + 552, + 293 + ], + "type": "table", + "html": "
Dataset# ObjectsRealScene TypeMulti-viewShapePBR Mat.Relit ImageLightingEgocentric Cap.
ShapeNet-Intrinsics [61]31KXsynthetic✓✓X✓✓X
NeRD Synthetic [11]3Xsynthetic✓✓✓✓✓X
ABO [16]8KXsynthetic✓✓✓✓✓X
MIT Intrasins [25]20✓studio✓XXXXX
DTU-MVS [34]80✓studio✓✓XXXX
Objaverse [18]818K(✓)*studio✓✓(✓)*XXX
DiLiGenT-MV [42]5✓studio✓✓XX✓X
ReNe [65]20✓studio✓XXX✓X
OpenIllumination [46]64✓studio✓X✓✓✓X
GSO [21]1030✓studio✓✓XXXX
Lombardi et al. [47]6✓in-the-wildX✓X✓✓X
NeRD Real [11]4✓in-the-wild✓X✓✓XX
NeROIC [38]3✓in-the-wild✓X✓✓XX
Oxholm et al. [56]4✓in-the-wild✓✓X✓✓X
OmniObject3D [73]6k✓in-the-wild✓✓XXXX
Stanford Orb [39]14✓in-the-wild✓✓✓✓✓X
DTC (ours)2k✓in-the-wild✓✓✓✓✓✓
", + "image_path": "bb306c32194b753c9ba465301f050069d8d24047d747150c927252e682122167.jpg" + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "table_body" + } + ], + "index": 1 + }, + { + "bbox": [ + 54, + 314, + 295, + 398 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 54, + 314, + 295, + 398 + ], + "spans": [ + { + "bbox": [ + 54, + 314, + 295, + 398 + ], + "type": "text", + "content": "ods using the egocentric aligned DTC data. Additionally, we explore the dataset's potential in downstream robotics applications by assessing its effectiveness in training robotic policies for pushing and grasping tasks in simulation. These benchmarks and applications provide valuable insights, highlight existing challenges, and uncover promising directions for future research in 3D digital twin creation." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 55, + 413, + 142, + 425 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 413, + 142, + 425 + ], + "spans": [ + { + "bbox": [ + 55, + 413, + 142, + 425 + ], + "type": "text", + "content": "2. Related Work" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 54, + 434, + 295, + 518 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 54, + 434, + 295, + 518 + ], + "spans": [ + { + "bbox": [ + 54, + 434, + 295, + 518 + ], + "type": "text", + "content": "We provide a comparison of our DTC dataset to existing object-centric datasets in Table 1. We provide the largest 3D dataset with PBR materials and real world multi-view recordings with digital twin counterparts. We further provide digital twin aligned egocentric recordings, the first of their kind in the egocentric domain. We will discuss the related datasets and methods they can empower as follows." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 54, + 521, + 295, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 54, + 521, + 295, + 713 + ], + "spans": [ + { + "bbox": [ + 54, + 521, + 295, + 713 + ], + "type": "text", + "content": "3D Digital Twin Datasets Existing 3D digital twin datasets with PBR materials often serve as ground truth for evaluating inverse rendering results. Early efforts [6, 25] provide small-scale intrinsic image of real objects and do not provide shape or PBR material information. Synthetic datasets [11, 16, 45, 61, 70, 72] are widely used for evaluation but do not represent the complexity in a real world environment. For datasets that contain real objects, [18, 34, 42, 47, 56, 65], the reconstruction quality can vary, which leads to imprecise evaluations. Table 1 provides a comparison to the previous work in this domain. Compared to Objaverse[18, 19], which is a collection of existing 3D models with only a small subset containing PBR materials with varying quality, we offer a high quality collection of 3D object data that is also aligned with real world recordings. Compared to OmniObject3D [73], the DTC models" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 313, + 314, + 555, + 469 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 314, + 555, + 469 + ], + "spans": [ + { + "bbox": [ + 313, + 314, + 555, + 469 + ], + "type": "text", + "content": "provide higher-quality shape and additional PBR materials that are necessary for high quality inverse rendering. We offer the largest quantity of 3D object models compared to all counterparts in various tasks. For real-world evaluation, Stanford-ORB [39] was the prior largest inverse rendering benchmark with in-the-wild lighting. In contrast, we provide more object diversity and higher quality for each object model. The Aria Digital Twin dataset [58] was the first dataset to provide digital twin aligned environments for the scenes and recorded using egocentric device. However, their scene environments are limited and the contained object ground truths inside do not have high quality geometries with PBR materials." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 313, + 473, + 556, + 628 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 473, + 556, + 628 + ], + "spans": [ + { + "bbox": [ + 313, + 473, + 556, + 628 + ], + "type": "text", + "content": "Object Reconstruction & Inverse Rendering. Using object-centric multi-view images as input, early object reconstruction methods focused on estimating individual object properties, such as shape from shading [5, 9, 83], material acquisition [44, 45, 56, 57, 74], and lighting estimation [67, 78]. Some approaches also aimed to recover reflectance and illumination assuming known object shapes [47, 48]. Inverse rendering, which seeks to invert the rendering equation [36], estimates an image's intrinsic components—geometry, materials, and lighting. The advent of differentiable renderers [14, 15, 49] enabled full-fledged inverse rendering methods to simultaneously recover all these properties for object reconstruction [50]." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 313, + 630, + 556, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 630, + 556, + 715 + ], + "spans": [ + { + "bbox": [ + 313, + 630, + 556, + 715 + ], + "type": "text", + "content": "Neural volumetric representations such as Neural Radiance Fields (NeRFs) [52] and the like [7, 11-13, 59, 79, 84] encode geometry and appearance as volumetric densities and radiance with a Multi-Layer Perceptron (MLP) network, and render images using the volume rendering equation [51]. 3D-GS [37] introduces 3D Gaussian primitives and rasterization and its following-up variants [31] demonstrates high quality" + } + ] + } + ], + "index": 8 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 297, + 748, + 312, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 297, + 748, + 312, + 757 + ], + "spans": [ + { + "bbox": [ + 297, + 748, + 312, + 757 + ], + "type": "text", + "content": "755" + } + ] + } + ], + "index": 9 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 2 + }, + { + "para_blocks": [ + { + "bbox": [ + 55, + 72, + 171, + 83 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 72, + 171, + 83 + ], + "spans": [ + { + "bbox": [ + 55, + 72, + 171, + 83 + ], + "type": "text", + "content": "geometry prediction as well." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 55, + 84, + 296, + 240 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 84, + 296, + 240 + ], + "spans": [ + { + "bbox": [ + 55, + 84, + 296, + 240 + ], + "type": "text", + "content": "Other surface-based representations [29, 53, 62, 66, 71, 80, 81, 85] extract surfaces as the zero level set, for instance, of a signed distance function (SDF) or an occupancy field [55], allowing them to efficiently model the appearance on the surface with an explicit material model, such as bidirectional reflectance distribution functions (BRDFs). This also enables modeling more complex global illumination effects, such as self-shadows. Most of these methods focus on per-scene optimization and require dense multiple views as input. Recently, researchers have incorporated learning-based models, distilling priors from large training datasets for fast inference on limited test views [8, 10, 33, 45, 61, 70, 72, 86]." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 55, + 240, + 296, + 289 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 240, + 296, + 289 + ], + "spans": [ + { + "bbox": [ + 55, + 240, + 296, + 289 + ], + "type": "text", + "content": "In this work, we provide the DTC dataset with real-world object recordings that can serve as the benchmark to evaluate object-centric inverse rendering tasks. We evaluate representative baselines from existing work." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 55, + 300, + 297, + 329 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 300, + 297, + 329 + ], + "spans": [ + { + "bbox": [ + 55, + 300, + 297, + 329 + ], + "type": "text", + "content": "3. Digital Twin Catalog: A Large-Scale Photorealistic 3D Object Digital Twin Dataset" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 55, + 335, + 176, + 348 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 335, + 176, + 348 + ], + "spans": [ + { + "bbox": [ + 55, + 335, + 176, + 348 + ], + "type": "text", + "content": "3.1. Dataset Composition" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 55, + 353, + 296, + 437 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 353, + 296, + 437 + ], + "spans": [ + { + "bbox": [ + 55, + 353, + 296, + 437 + ], + "type": "text", + "content": "Our DTC dataset contains: (1) 2,000 scanned 3D object models, featuring millimeter geometric accuracy relative to their physical counterparts, along with a full set of photorealistic PBR material maps (Fig. 1), (2) 100 DSLR-captured evaluation data of 50 objects under different lighting conditions, and (3) 200 egocentric Aria-captured evaluation data of 100 objects with both active and casual observation modes." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 55, + 444, + 218, + 456 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 444, + 218, + 456 + ], + "spans": [ + { + "bbox": [ + 55, + 444, + 218, + 456 + ], + "type": "text", + "content": "3.2. Creation of 3D Object Models" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 55, + 461, + 296, + 629 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 461, + 296, + 629 + ], + "spans": [ + { + "bbox": [ + 55, + 461, + 296, + 629 + ], + "type": "text", + "content": "Utilizing the state-of-the-art industrial 3D object scanner [1], we selected 2,000 physical-world objects spanning 40 LVIS [27] categories, carefully chosen to ensure both category diversity and compatibility with the scanner's capabilities. As illustrated in Fig. 3(a) (c), the scanner features a fixed lighting-camera setup within an upper-hemisphere dome, equipped with 8 structured lights for geometry scanning, and 29 spotlights and 29 cameras for material acquisition. During the scanning process, the object is placed on a central holder, and its pose can be adjusted with multi-round scanning to achieve a complete 360-degree scan. For our dataset, each object typically undergoes three pose changes, with a total scanning time of approximately 20 minutes per object." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 55, + 630, + 296, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 630, + 296, + 714 + ], + "spans": [ + { + "bbox": [ + 55, + 630, + 296, + 714 + ], + "type": "text", + "content": "After scanning, a proprietary post-processing pipeline reconstructs both the geometry and the PBR material maps. For 4K-resolution PBR material maps, the post-processing requires approximately 4 hours per object. In terms of quality, the structured-light-based shape reconstruction (Fig. 3 (b)) in the post-processing achieves millimeter-level geometric accuracy. However, the material optimization process" + } + ] + } + ], + "index": 8 + }, + { + "type": "image", + "bbox": [ + 317, + 71, + 435, + 187 + ], + "blocks": [ + { + "bbox": [ + 317, + 71, + 435, + 187 + ], + "lines": [ + { + "bbox": [ + 317, + 71, + 435, + 187 + ], + "spans": [ + { + "bbox": [ + 317, + 71, + 435, + 187 + ], + "type": "image", + "image_path": "514e435cb5ecf8ec602ac85b2b43b8a3f2c3a2d8ef0c882b2ac2ddd7d287b30e.jpg" + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "image_body" + } + ], + "index": 9 + }, + { + "type": "image", + "bbox": [ + 317, + 187, + 435, + 226 + ], + "blocks": [ + { + "bbox": [ + 317, + 187, + 435, + 226 + ], + "lines": [ + { + "bbox": [ + 317, + 187, + 435, + 226 + ], + "spans": [ + { + "bbox": [ + 317, + 187, + 435, + 226 + ], + "type": "image", + "image_path": "135bc7d0496f2c9a75be1831c6bd023e92a6f80918f7d1b09193a3d349de2602.jpg" + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "image_body" + } + ], + "index": 10 + }, + { + "type": "image", + "bbox": [ + 317, + 226, + 472, + 320 + ], + "blocks": [ + { + "bbox": [ + 317, + 226, + 472, + 320 + ], + "lines": [ + { + "bbox": [ + 317, + 226, + 472, + 320 + ], + "spans": [ + { + "bbox": [ + 317, + 226, + 472, + 320 + ], + "type": "image", + "image_path": "1e2df084aba61005a1fff73ede9c775b1492db5fa33bf103bb57e655c46b4ffc.jpg" + } + ] + } + ], + "index": 11, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 343, + 327, + 524, + 340 + ], + "lines": [ + { + "bbox": [ + 343, + 327, + 524, + 340 + ], + "spans": [ + { + "bbox": [ + 343, + 327, + 524, + 340 + ], + "type": "text", + "content": "Figure 3. 3D object scanner by Covision Media®." + } + ] + } + ], + "index": 14, + "angle": 0, + "type": "image_caption" + } + ], + "index": 11 + }, + { + "type": "image", + "bbox": [ + 435, + 72, + 552, + 188 + ], + "blocks": [ + { + "bbox": [ + 435, + 72, + 552, + 188 + ], + "lines": [ + { + "bbox": [ + 435, + 72, + 552, + 188 + ], + "spans": [ + { + "bbox": [ + 435, + 72, + 552, + 188 + ], + "type": "image", + "image_path": "2221885c36272e45c04b1bb11f68fe0926c7cc4abddbd05d5b0cb6a6dc0834fc.jpg" + } + ] + } + ], + "index": 12, + "angle": 0, + "type": "image_body" + } + ], + "index": 12 + }, + { + "type": "image", + "bbox": [ + 435, + 187, + 552, + 226 + ], + "blocks": [ + { + "bbox": [ + 435, + 187, + 552, + 226 + ], + "lines": [ + { + "bbox": [ + 435, + 187, + 552, + 226 + ], + "spans": [ + { + "bbox": [ + 435, + 187, + 552, + 226 + ], + "type": "image", + "image_path": "b6c70bcdf78249b5ac44914e03c7c6a9ce38c56777016c4c60a6be8291f00057.jpg" + } + ] + } + ], + "index": 13, + "angle": 0, + "type": "image_body" + } + ], + "index": 13 + }, + { + "type": "image", + "bbox": [ + 317, + 352, + 432, + 483 + ], + "blocks": [ + { + "bbox": [ + 317, + 352, + 432, + 483 + ], + "lines": [ + { + "bbox": [ + 317, + 352, + 432, + 483 + ], + "spans": [ + { + "bbox": [ + 317, + 352, + 432, + 483 + ], + "type": "image", + "image_path": "81e1f60c73757d07d21564f4f92073b2c070847f5d4120db339f1d4b88771dbc.jpg" + } + ] + } + ], + "index": 15, + "angle": 0, + "type": "image_body" + } + ], + "index": 15 + }, + { + "type": "image", + "bbox": [ + 433, + 352, + 553, + 483 + ], + "blocks": [ + { + "bbox": [ + 433, + 352, + 553, + 483 + ], + "lines": [ + { + "bbox": [ + 433, + 352, + 553, + 483 + ], + "spans": [ + { + "bbox": [ + 433, + 352, + 553, + 483 + ], + "type": "image", + "image_path": "c7579deec3fdec506ecd75bef2a02066e26fe418be49935994c28f186ed5f5bc.jpg" + } + ] + } + ], + "index": 16, + "angle": 0, + "type": "image_body" + } + ], + "index": 16 + }, + { + "type": "image", + "bbox": [ + 317, + 484, + 432, + 612 + ], + "blocks": [ + { + "bbox": [ + 317, + 484, + 432, + 612 + ], + "lines": [ + { + "bbox": [ + 317, + 484, + 432, + 612 + ], + "spans": [ + { + "bbox": [ + 317, + 484, + 432, + 612 + ], + "type": "image", + "image_path": "ea01e0790fab617bd0e4f3f993dbf25516efb0989fd13c4c1b2cf25eb8b087e6.jpg" + } + ] + } + ], + "index": 17, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 329, + 620, + 538, + 632 + ], + "lines": [ + { + "bbox": [ + 329, + 620, + 538, + 632 + ], + "spans": [ + { + "bbox": [ + 329, + 620, + 538, + 632 + ], + "type": "text", + "content": "Figure 4. Rendered DTC models (left) v.s. Photo (Right)." + } + ] + } + ], + "index": 19, + "angle": 0, + "type": "image_caption" + } + ], + "index": 17 + }, + { + "type": "image", + "bbox": [ + 432, + 484, + 553, + 612 + ], + "blocks": [ + { + "bbox": [ + 432, + 484, + 553, + 612 + ], + "lines": [ + { + "bbox": [ + 432, + 484, + 553, + 612 + ], + "spans": [ + { + "bbox": [ + 432, + 484, + 553, + 612 + ], + "type": "image", + "image_path": "2dde835d30df0f67fe2c497f736b8b0b1a2f89ba0109ae86e7edd6a6d6160cc2.jpg" + } + ] + } + ], + "index": 18, + "angle": 0, + "type": "image_body" + } + ], + "index": 18 + }, + { + "bbox": [ + 313, + 654, + 554, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 654, + 554, + 713 + ], + "spans": [ + { + "bbox": [ + 313, + 654, + 554, + 713 + ], + "type": "text", + "content": "performs best for diffuse objects and often struggles with glossy or shiny surfaces. To address this limitation, we hired technical artists to develop a workflow to refine materials for glossy and shiny objects, ensuring that the material quality meets the standards of a digital twin." + } + ] + } + ], + "index": 20 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 298, + 749, + 312, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 298, + 749, + 312, + 757 + ], + "spans": [ + { + "bbox": [ + 298, + 749, + 312, + 757 + ], + "type": "text", + "content": "756" + } + ] + } + ], + "index": 21 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 3 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 57, + 69, + 295, + 178 + ], + "blocks": [ + { + "bbox": [ + 57, + 69, + 295, + 178 + ], + "lines": [ + { + "bbox": [ + 57, + 69, + 295, + 178 + ], + "spans": [ + { + "bbox": [ + 57, + 69, + 295, + 178 + ], + "type": "image", + "image_path": "788b6a091b4172786c2a711155a41ca86b1db107b1def661f06eda6b8ac10fa2.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 55, + 186, + 295, + 209 + ], + "lines": [ + { + "bbox": [ + 55, + 186, + 295, + 209 + ], + "spans": [ + { + "bbox": [ + 55, + 186, + 295, + 209 + ], + "type": "text", + "content": "Figure 5. Shape and material (albedo) quality comparison between Stanford-ORB [39] (left) and our DTC (right)." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "bbox": [ + 54, + 231, + 295, + 350 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 54, + 231, + 295, + 350 + ], + "spans": [ + { + "bbox": [ + 54, + 231, + 295, + 350 + ], + "type": "text", + "content": "3D Model Accuracy. To validate the material and geometry accuracy of the 3D models, we compared a rendered image of our scanned and processed model with a photograph of the same object taken inside a light box. A virtual light box was meticulously modeled to replicate the light intensity and color temperature of the real light box. The scanned object was then placed in the virtual light box to generate the rendered image. The side-by-side comparison demonstrates a remarkable match between the rendered and real images (Fig. 4)." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 55, + 354, + 296, + 413 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 354, + 296, + 413 + ], + "spans": [ + { + "bbox": [ + 55, + 354, + 296, + 413 + ], + "type": "text", + "content": "Comparison Against Stanford-ORB. We also scanned the objects used in Stanford-ORB [39] to compare the shape and appearance quality. As illustrated in Fig. 5, the Stanford-ORB models exhibit shape artifacts and noisy, lower-quality materials compared to our models." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 55, + 417, + 187, + 429 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 417, + 187, + 429 + ], + "spans": [ + { + "bbox": [ + 55, + 417, + 187, + 429 + ], + "type": "text", + "content": "3.3. DSLR Evaluation Data" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 55, + 434, + 295, + 519 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 434, + 295, + 519 + ], + "spans": [ + { + "bbox": [ + 55, + 434, + 295, + 519 + ], + "type": "text", + "content": "Within DTC, we include a DSLR-captured evaluation dataset of 50 objects from Sec. 3.2 captured under two different lighting conditions, resulting in 100 distinct image sequences. For every sequence, we provide (a) approximately 120 HDR and LDR images from different viewing directions, (b) one object pose and (c) per-image camera pose. The two lighting conditions are represented using two environment maps." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 55, + 522, + 296, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 522, + 296, + 713 + ], + "spans": [ + { + "bbox": [ + 55, + 522, + 296, + 713 + ], + "type": "text", + "content": "Data Capture. To ensure the DSLR evaluation data quality, we designed and built a DSLR camera rig to automate the capture process (Fig. 6). The rig is designed to rotate the cameras around the centralized object, assuming the environment lighting remains unchanged during the capture. It features a motorized rotary stage with a centrally mounted stationary platform. Attached to the rotary stage is an extrusion frame that forms the gantry arm, supported by a set of castors to bear its weight and enable smooth rotation around the central axis. The extrusion frame is equipped with adjustable camera mounts, allowing DSLR cameras to be positioned flexibly to optimize the capture setup. For our capture process, we utilized three DSLR cameras to perform a 360-degree rotation around the object, capturing images at 9-degree intervals, resulting in 120 photos per object. To ensure precise camera pose estimation, a ChArUco board" + } + ] + } + ], + "index": 6 + }, + { + "type": "image", + "bbox": [ + 316, + 69, + 554, + 185 + ], + "blocks": [ + { + "bbox": [ + 316, + 69, + 554, + 185 + ], + "lines": [ + { + "bbox": [ + 316, + 69, + 554, + 185 + ], + "spans": [ + { + "bbox": [ + 316, + 69, + 554, + 185 + ], + "type": "image", + "image_path": "1a6ce4a894ebbfcfb6ec2630ff256afe74443ffa2f522328986bb4e235917f94.jpg" + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 343, + 193, + 525, + 204 + ], + "lines": [ + { + "bbox": [ + 343, + 193, + 525, + 204 + ], + "spans": [ + { + "bbox": [ + 343, + 193, + 525, + 204 + ], + "type": "text", + "content": "Figure 6. DSLR rig for capturing evaluation data." + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "image_caption" + } + ], + "index": 7 + }, + { + "bbox": [ + 313, + 226, + 555, + 248 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 226, + 555, + 248 + ], + "spans": [ + { + "bbox": [ + 313, + 226, + 555, + 248 + ], + "type": "text", + "content": "was placed beneath the object during the capture. Example images from this setup are shown in Fig. 6." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 313, + 248, + 555, + 402 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 248, + 555, + 402 + ], + "spans": [ + { + "bbox": [ + 313, + 248, + 555, + 402 + ], + "type": "text", + "content": "Environment Maps. Following a similar approach in Stanford-ORB [39], we capture the two environment maps using chrome ball images obtained with the capture rig described earlier. With precise camera poses provided by a ChArUco board placed beneath the chrome ball, we first fit a synthetic 3D sphere to the chrome ball by optimizing its 3D position using a geometry-friendly differentiable renderer [40, 75]. Subsequently, using a differentiable Monte Carlo-based renderer [32], we refine the environment map to match the reflection on the chrome ball, employing the single-view light estimation method proposed in [78]. The coordinate system of the environment map is determined by the ChArUco detection." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 313, + 406, + 556, + 586 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 406, + 556, + 586 + ], + "spans": [ + { + "bbox": [ + 313, + 406, + 556, + 586 + ], + "type": "text", + "content": "Pose Registration for Camera and Object. We first obtain the initial camera poses using the ChArUco board. In most cases, for images captured from the top and middle views, the pose estimates are typically accurate. However, for bottom-view images, inaccuracies arise due to pattern distortion at grazing angles. To mitigate this issue, we refine the camera poses by fitting the rendering of a virtual ChArUco board to the real captured photo using a differentiable renderer [32, 40, 75]. Once the camera poses have been accurately refined, we optimize the object pose by minimizing the mask loss between the rendered mask and the reference one generated by [62]. By providing separate poses for camera and object anchored by ChArUco, the cameras and object are automatically aligned with the optimized environment maps." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 313, + 594, + 466, + 605 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 594, + 466, + 605 + ], + "spans": [ + { + "bbox": [ + 313, + 594, + 466, + 605 + ], + "type": "text", + "content": "3.4. Egocentric Evaluation Data" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 313, + 605, + 556, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 605, + 556, + 713 + ], + "spans": [ + { + "bbox": [ + 313, + 605, + 556, + 713 + ], + "type": "text", + "content": "We include an evaluation dataset of real world recordings captured by an egocentric device paired with 100 objects from the DTC dataset. We capture the egocentric recordings using the open-sourced Project Aria device [22] and acquire the additional 3D ground truth using its machine perception tool, which includes online device calibration, device trajectory and semi-dense point clouds for each recording. Each recording contains a single 3D object and precisely aligned object poses in the Aria trajectory coordinate frame." + } + ] + } + ], + "index": 13 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 298, + 749, + 312, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 298, + 749, + 312, + 757 + ], + "spans": [ + { + "bbox": [ + 298, + 749, + 312, + 757 + ], + "type": "text", + "content": "757" + } + ] + } + ], + "index": 14 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 4 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 97, + 69, + 521, + 309 + ], + "blocks": [ + { + "bbox": [ + 97, + 69, + 521, + 309 + ], + "lines": [ + { + "bbox": [ + 97, + 69, + 521, + 309 + ], + "spans": [ + { + "bbox": [ + 97, + 69, + 521, + 309 + ], + "type": "image", + "image_path": "dcd62f9fc15c847d5a6c0ee8cc41a039480a020f8dc5ae093f19a24eaae0f820.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 54, + 317, + 555, + 351 + ], + "lines": [ + { + "bbox": [ + 54, + 317, + 555, + 351 + ], + "spans": [ + { + "bbox": [ + 54, + 317, + 555, + 351 + ], + "type": "text", + "content": "Figure 7. The workflow to align egocentric video and 3D objects. We acquire the object's semi-dense point cloud from the egocentric recording and the neural reconstructed mesh using Neural-PBIR [62]. Finally, we align the 3D object with the rendered mask from the neural reconstructed mesh using physics-based differentiable rendering (PBDR)." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "bbox": [ + 54, + 372, + 295, + 408 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 54, + 372, + 295, + 408 + ], + "spans": [ + { + "bbox": [ + 54, + 372, + 295, + 408 + ], + "type": "text", + "content": "Both the object and the trajectory are in metric scale. We can derive additional ground truth for each video from the aligned 3D object properties." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 56, + 411, + 296, + 651 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 411, + 296, + 651 + ], + "spans": [ + { + "bbox": [ + 56, + 411, + 296, + 651 + ], + "type": "text", + "content": "Data Capture. To feature real world recordings observed from human perspectives that can be representative of 3D object reconstruction for AR/MR devices or robots, we provide two types of recording trajectories for selected objects, termed active and passive respectively, collected by human wearers. The active recording features a complete 360 view of the objects, which is similar to existing object 3D novel view synthesis dataset. The passive recording features casual looks from human wearer, which only contain partial views of the object from certain viewing angles along the trajectory. In dataset creation, we collected the active and passive recordings in the same environment and generate their 3D information in the shared 3D space. This helps reduces potential failures when aligning the object to passive 3D recordings, which are shorter and contains less 3D information. To reduce the effect of noises and motion blur, which are common in egocentric videos in indoor low light environments, we light the capture environment to " + }, + { + "bbox": [ + 56, + 411, + 296, + 651 + ], + "type": "inline_equation", + "content": "3\\mathrm{K}+" + }, + { + "bbox": [ + 56, + 411, + 296, + 651 + ], + "type": "text", + "content": " lux illumination and used a fixed low-exposure and gain profile to collect each recording with the appropriate brightness." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 54, + 654, + 297, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 54, + 654, + 297, + 715 + ], + "spans": [ + { + "bbox": [ + 54, + 654, + 297, + 715 + ], + "type": "text", + "content": "Alignment between Egocentric Video and Object. We provide an illustration of the object alignment to the egocentric video in Fig. 7. Given the images, camera poses and semi-dense point cloud acquired from Project Aria tools, we employ a neural-based mesh reconstruction method [62] to" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 313, + 372, + 556, + 504 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 372, + 556, + 504 + ], + "spans": [ + { + "bbox": [ + 313, + 372, + 556, + 504 + ], + "type": "text", + "content": "create reference meshes for generating high-quality reference masks and used that to align with the corresponding 3D object mesh. This alignment step serves as an initialization for a more precise pose refinement, which leverages differentiable rendering. For certain objects with symmetric geometry, we observe this process can introduce ambiguities in point cloud registration and subsequent failures. To address such cases, we provide a GUI to manually align and correct the object alignment. Finally, akin to the DSLR camera pose registration phase, we optimize a mask loss over object poses to achieve fine-tuned pose registration." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 313, + 515, + 495, + 529 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 515, + 495, + 529 + ], + "spans": [ + { + "bbox": [ + 313, + 515, + 495, + 529 + ], + "type": "text", + "content": "4. Benchmarking and Applications" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 313, + 531, + 556, + 663 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 531, + 556, + 663 + ], + "spans": [ + { + "bbox": [ + 313, + 531, + 556, + 663 + ], + "type": "text", + "content": "We first use our DLSR and egocentric dataset as a benchmark for existing state-of-the-art methods. For inverse rendering, we design metrics to evaluate the shape and material quality of the recovered 3D object digital twin from three perspectives. For egocentric recording, we evaluate the novel-view synthesis as the initial evaluation. We include additional tasks, e.g. sparse view reconstruction in the supplementary materials for both DSLR and egocentric recordings. Finally we can demonstrate our high quality 3D digital twin models can be beneficial to robotics domain using an application in robotics manipulation." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 313, + 671, + 541, + 685 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 671, + 541, + 685 + ], + "spans": [ + { + "bbox": [ + 313, + 671, + 541, + 685 + ], + "type": "text", + "content": "4.1. Application to Inverse Rendering for DSLR" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 313, + 689, + 555, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 689, + 555, + 715 + ], + "spans": [ + { + "bbox": [ + 313, + 689, + 555, + 715 + ], + "type": "text", + "content": "The DSLR dataset in DTC provides accurate ground truth, including poses, lighting and 3D models, for inverse rendering" + } + ] + } + ], + "index": 9 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 298, + 749, + 312, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 298, + 749, + 312, + 757 + ], + "spans": [ + { + "bbox": [ + 298, + 749, + 312, + 757 + ], + "type": "text", + "content": "758" + } + ] + } + ], + "index": 10 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 5 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 58, + 91, + 553, + 178 + ], + "blocks": [ + { + "bbox": [ + 55, + 70, + 553, + 82 + ], + "lines": [ + { + "bbox": [ + 55, + 70, + 553, + 82 + ], + "spans": [ + { + "bbox": [ + 55, + 70, + 553, + 82 + ], + "type": "text", + "content": "Table 2. Benchmark comparison of existing methods on inverse rendering for DSLR. Depth SI-MSE and Shape Chamfer distance " + }, + { + "bbox": [ + 55, + 70, + 553, + 82 + ], + "type": "inline_equation", + "content": "\\times {10}^{-3}" + }, + { + "bbox": [ + 55, + 70, + 553, + 82 + ], + "type": "text", + "content": " ." + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 58, + 91, + 553, + 178 + ], + "lines": [ + { + "bbox": [ + 58, + 91, + 553, + 178 + ], + "spans": [ + { + "bbox": [ + 58, + 91, + 553, + 178 + ], + "type": "table", + "html": "
GeometryNovel Scene RelightingNovel View Synthesis
Depth↓Normal↓Shape↓PSNR-H↑PSNR-L↑SSIM↑LPIPS↓PSNR-H↑PSNR-L↑SSIM↑LPIPS↓
Neural-PIL [12]5.710.2525.02N/A28.4235.760.8820.096
PhySG [80]0.310.1611.3127.2832.860.9590.04928.5434.460.9640.045
NVDiffRec [53]0.020.071.6426.9933.270.9510.03728.9534.920.9670.029
NeRD [11]4.550.45108.2026.1032.600.9480.06126.8033.400.8820.102
InvRender [71]0.220.030.7529.5235.980.9610.03731.6437.820.9700.033
NVDiffRecMC [29]0.020.061.3427.7834.550.9520.04231.2738.170.9720.032
", + "image_path": "66a1b9d8428860a6cacaef98e9368932c4ff416345451293221c6ba11ddb4a8d.jpg" + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "table_body" + } + ], + "index": 1 + }, + { + "type": "table", + "bbox": [ + 58, + 216, + 294, + 258 + ], + "blocks": [ + { + "bbox": [ + 70, + 196, + 280, + 208 + ], + "lines": [ + { + "bbox": [ + 70, + 196, + 280, + 208 + ], + "spans": [ + { + "bbox": [ + 70, + 196, + 280, + 208 + ], + "type": "text", + "content": "Table 3. Benchmark on the egocentric aligned recordings." + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 58, + 216, + 294, + 258 + ], + "lines": [ + { + "bbox": [ + 58, + 216, + 294, + 258 + ], + "spans": [ + { + "bbox": [ + 58, + 216, + 294, + 258 + ], + "type": "table", + "html": "
PSNR ↑LPIPS ↓SSIM ↑Depth ↓Normal ↓
3D-GS [37]28.810.0200.98880.17680.3301
2D-GS [31]28.750.0200.98860.17550.2112
", + "image_path": "a234b44032efe9c9f6c7f33fd20e7e0ecc49d74e3d089a7a4236cf3cccd9627d.jpg" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "table_body" + } + ], + "index": 3 + }, + { + "bbox": [ + 55, + 278, + 296, + 350 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 278, + 296, + 350 + ], + "spans": [ + { + "bbox": [ + 55, + 278, + 296, + 350 + ], + "type": "text", + "content": "tasks and serves as an evaluation suite to benchmark the performance of inverse rendering methods. We select six prior methods for this task and evaluate their performance using the ground truth provided by our dataset. In the following sections, we describe the data splitting strategy, evaluation metrics, and baselines." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 55, + 353, + 296, + 437 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 353, + 296, + 437 + ], + "spans": [ + { + "bbox": [ + 55, + 353, + 296, + 437 + ], + "type": "text", + "content": "Data Splitting. For benchmarking purposes, we select 15 objects from the DSLR dataset captured under two distinct lighting environments, resulting in a total of 30 image sequences. The selected objects encompass a diverse range of geometric and material properties to ensure a comprehensive evaluation. For each scene, 8 views are selected for testing, while the remaining views are reserved for training." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 55, + 439, + 296, + 583 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 439, + 296, + 583 + ], + "spans": [ + { + "bbox": [ + 55, + 439, + 296, + 583 + ], + "type": "text", + "content": "Evaluation Metrics. The metrics measure the accuracy of three aspects of baseline performance: geometry estimation, relighting, and novel view synthesis. For geometry estimation, we evaluate the accuracy of predicted depth and normal maps under held-out test views, as well as 3D meshes extracted from baseline methods, compared with the ground truth from our dataset. Relighting metrics evaluate the material decomposition quality of baselines by measuring the accuracy of predicted images under held-out lighting conditions. For view synthesis, we compare the predicted images from viewpoints unseen during training to ground truth captures. We refer to Kuang et al. [39] for metric details." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 55, + 586, + 296, + 670 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 586, + 296, + 670 + ], + "spans": [ + { + "bbox": [ + 55, + 586, + 296, + 670 + ], + "type": "text", + "content": "Baselines. We include the following baselines: NVDiffRec [53] and NVDiffRecMC [29], with a hybrid shape representation DMTet [60]; InvRender [85] and PhySG [80], which adopt signed distance functions (SDFs) to represent object geometry [76] and utilize implicit neural fields for material decomposition; Neural-PIL [12] and NeRD [11], which use NeRFs [52] as scene representations." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 55, + 677, + 272, + 689 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 677, + 272, + 689 + ], + "spans": [ + { + "bbox": [ + 55, + 677, + 272, + 689 + ], + "type": "text", + "content": "4.2. Application to Egocentric Reconstruction" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 55, + 689, + 295, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 689, + 295, + 713 + ], + "spans": [ + { + "bbox": [ + 55, + 689, + 295, + 713 + ], + "type": "text", + "content": "Our digital twin models, aligned with real world video using the method described in Sec. 3.4, can help obtain accurate" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 313, + 198, + 555, + 317 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 198, + 555, + 317 + ], + "spans": [ + { + "bbox": [ + 313, + 198, + 555, + 317 + ], + "type": "text", + "content": "ground truth for object-centric images that were previously difficult to acquire. We provide the first evaluation of object-centric novel view synthesis recorded from an egocentric device. We use the projected object shape given the 3D pose of the object and cameras in scene coordinates to acquire the image masks, depth and normal for each object. We selected 15 recordings from the egocentric recording sessions as the evaluation and used the active recordings to benchmark novel view reconstruction. For each recording, we hold out every 8th image view as a testing view." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 313, + 320, + 556, + 453 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 320, + 556, + 453 + ], + "spans": [ + { + "bbox": [ + 313, + 320, + 556, + 453 + ], + "type": "text", + "content": "**Evaluations.** We build our baselines based on the gsplats [77] implementation of the 3D Gaussian Splitting (GS) [37] and 2D GS [31], and handle the effect of lens shading from the Project Aria lens [26]. We calculate PSNR, depth and normal based on the observed objects with masks and provide SSIM and LPIPS score on images by masking out the non-object areas as black. Table 3 shows the benchmark results of the baselines. We use the same depth and normal metric in DSLR evaluation. We provide additional qualitative evaluations and analyses on egocentric data towards sparser view settings in the supplementary materials." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 313, + 461, + 509, + 475 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 461, + 509, + 475 + ], + "spans": [ + { + "bbox": [ + 313, + 461, + 509, + 475 + ], + "type": "text", + "content": "4.3. Application to Robotic Manipulation" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 313, + 479, + 555, + 586 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 479, + 555, + 586 + ], + "spans": [ + { + "bbox": [ + 313, + 479, + 555, + 586 + ], + "type": "text", + "content": "High-quality object models have been leveraged in prior work to train real-world robotic agents in scenes represented explicitly [54] or implicitly [64]. These object models have also been shown to facilitate object-centric pose and lighting parameter estimation, enabling model-based planning [63]. In this section, we empirically evaluate the effectiveness of using DTC dataset objects in training robotic policies. Specifically, we consider learning robotic pushing and grasping skills in simulation." + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 313, + 588, + 556, + 684 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 588, + 556, + 684 + ], + "spans": [ + { + "bbox": [ + 313, + 588, + 556, + 684 + ], + "type": "text", + "content": "First, we sample a subset of 24 cup category objects from the DTC dataset and 24 cup objects from Objaverse-XL [19] " + }, + { + "bbox": [ + 313, + 588, + 556, + 684 + ], + "type": "inline_equation", + "content": "{}^{1}" + }, + { + "bbox": [ + 313, + 588, + 556, + 684 + ], + "type": "text", + "content": " . Since not all Objverse-XL objects come with textures, we randomize the colors of those objects uniformly in RGB space. To compute collision meshes for physical simulation, we perform convex decomposition on each object with CoACD [68]. We import these objects along with a UR5e robot equipped with a Robotiq 2F-85 (pushing) or" + } + ] + } + ], + "index": 14 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 313, + 693, + 555, + 712 + ], + "type": "page_footnote", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 693, + 555, + 712 + ], + "spans": [ + { + "bbox": [ + 313, + 693, + 555, + 712 + ], + "type": "text", + "content": "1The version of Objaverse-XL used in this work excludes all 3D models sourced from Sketchfab." + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 298, + 749, + 312, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 298, + 749, + 312, + 757 + ], + "spans": [ + { + "bbox": [ + 298, + 749, + 312, + 757 + ], + "type": "text", + "content": "759" + } + ] + } + ], + "index": 16 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 6 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 67, + 70, + 286, + 138 + ], + "blocks": [ + { + "bbox": [ + 67, + 70, + 286, + 138 + ], + "lines": [ + { + "bbox": [ + 67, + 70, + 286, + 138 + ], + "spans": [ + { + "bbox": [ + 67, + 70, + 286, + 138 + ], + "type": "table", + "html": "
TaskDTC (ours)Objverse-XL [19]
Pushing @ 2cm36.3% ± 1.5%25.3% ± 6.0%
Pushing @ 3cm43.7% ± 1.2%29.7% ± 6.0%
Pushing @ 5cm47.0% ± 2.6%40.3% ± 5.5%
Grasping42.7% ± 4.7%38.6% ± 11.0%
", + "image_path": "a6ab799d2042df478496745334f08ef485d78cac29c002571b9ec76578bd80f9.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "table_body" + } + ], + "index": 0 + }, + { + "bbox": [ + 55, + 146, + 295, + 191 + ], + "lines": [ + { + "bbox": [ + 55, + 146, + 295, + 191 + ], + "spans": [ + { + "bbox": [ + 55, + 146, + 295, + 191 + ], + "type": "text", + "content": "Table 4. Success rate of policies trained on data collected using objects from our DTC dataset and sampled from Objaverse-XL when evaluated on an unseen test object. Errors indicate sample standard deviation over three policy training seeds." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "text" + }, + { + "bbox": [ + 55, + 222, + 296, + 283 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 222, + 296, + 283 + ], + "spans": [ + { + "bbox": [ + 55, + 222, + 296, + 283 + ], + "type": "text", + "content": "Robotiq 2F-140 (grasping) gripper into the PyBullet simulator [17] and collect data for each robotic task as described below. After training policies on data from each object set, we evaluate policy performance on a relatively high-quality unseen test object from the StanfordORB dataset [39]." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 54, + 290, + 295, + 445 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 54, + 290, + 295, + 445 + ], + "spans": [ + { + "bbox": [ + 54, + 290, + 295, + 445 + ], + "type": "text", + "content": "Pushing. For the pushing task, we collect 5000 trajectories of pseudo-random robotic interaction data for each object set. For each trajectory, a single object from the considered object set and its initial position are randomly selected. Then we train a goal-conditioned neural network policy " + }, + { + "bbox": [ + 54, + 290, + 295, + 445 + ], + "type": "inline_equation", + "content": "\\pi (a|o,o_g)" + }, + { + "bbox": [ + 54, + 290, + 295, + 445 + ], + "type": "text", + "content": " where " + }, + { + "bbox": [ + 54, + 290, + 295, + 445 + ], + "type": "inline_equation", + "content": "o" + }, + { + "bbox": [ + 54, + 290, + 295, + 445 + ], + "type": "text", + "content": " is an image observation of the current scene and " + }, + { + "bbox": [ + 54, + 290, + 295, + 445 + ], + "type": "inline_equation", + "content": "o_{g}" + }, + { + "bbox": [ + 54, + 290, + 295, + 445 + ], + "type": "text", + "content": " is a goal image indicating the desired final object and robot position, " + }, + { + "bbox": [ + 54, + 290, + 295, + 445 + ], + "type": "inline_equation", + "content": "o,o_{g}\\in \\mathbb{R}^{256\\times 256\\times 3}" + }, + { + "bbox": [ + 54, + 290, + 295, + 445 + ], + "type": "text", + "content": ", and the action " + }, + { + "bbox": [ + 54, + 290, + 295, + 445 + ], + "type": "inline_equation", + "content": "a\\in \\mathbb{R}^2" + }, + { + "bbox": [ + 54, + 290, + 295, + 445 + ], + "type": "text", + "content": " represents a change in the robot's end-effector position in the " + }, + { + "bbox": [ + 54, + 290, + 295, + 445 + ], + "type": "inline_equation", + "content": "x" + }, + { + "bbox": [ + 54, + 290, + 295, + 445 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 54, + 290, + 295, + 445 + ], + "type": "inline_equation", + "content": "y" + }, + { + "bbox": [ + 54, + 290, + 295, + 445 + ], + "type": "text", + "content": " axes. The " + }, + { + "bbox": [ + 54, + 290, + 295, + 445 + ], + "type": "inline_equation", + "content": "z" + }, + { + "bbox": [ + 54, + 290, + 295, + 445 + ], + "type": "text", + "content": " axis end-effector height is held fixed. We sample goals for training via hindsight relabeling [4, 20, 24]. We then perform evaluation on 100 randomly sampled test goals manipulating an unseen test cup from StanfordORB [39]." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 55, + 454, + 295, + 574 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 454, + 295, + 574 + ], + "spans": [ + { + "bbox": [ + 55, + 454, + 295, + 574 + ], + "type": "text", + "content": "Grasping. For grasping, we collect 5000 successful grasp examples for each object set by first placing a single object into the scene, randomizing the object identity and initial position. We then randomly sample candidate grasp poses in a radius around the object's position and simulating their outcomes, rejecting unsuccessful grasps. We train a grasping policy " + }, + { + "bbox": [ + 55, + 454, + 295, + 574 + ], + "type": "inline_equation", + "content": "\\pi(a|o)" + }, + { + "bbox": [ + 55, + 454, + 295, + 574 + ], + "type": "text", + "content": " where " + }, + { + "bbox": [ + 55, + 454, + 295, + 574 + ], + "type": "inline_equation", + "content": "o \\in \\mathbb{R}^{256 \\times 256 \\times 3}" + }, + { + "bbox": [ + 55, + 454, + 295, + 574 + ], + "type": "text", + "content": " is an image observation of the scene and " + }, + { + "bbox": [ + 55, + 454, + 295, + 574 + ], + "type": "inline_equation", + "content": "a \\in \\mathbb{R}^4" + }, + { + "bbox": [ + 55, + 454, + 295, + 574 + ], + "type": "text", + "content": " represents the " + }, + { + "bbox": [ + 55, + 454, + 295, + 574 + ], + "type": "inline_equation", + "content": "x, y, z" + }, + { + "bbox": [ + 55, + 454, + 295, + 574 + ], + "type": "text", + "content": " position and " + }, + { + "bbox": [ + 55, + 454, + 295, + 574 + ], + "type": "inline_equation", + "content": "\\theta" + }, + { + "bbox": [ + 55, + 454, + 295, + 574 + ], + "type": "text", + "content": " yaw rotation of the robot end-effector at which to attempt the grasp. Again we use 100 test cup object poses." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 55, + 582, + 295, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 582, + 295, + 714 + ], + "spans": [ + { + "bbox": [ + 55, + 582, + 295, + 714 + ], + "type": "text", + "content": "Results. We report the results in Table 4 and Fig. 8. We find that across both pushing and grasping tasks, policies trained on DTC dataset objects outperform those trained on Objaverse-XL objects when evaluated on the unseen test object. For pushing, we report performance by defining binary success thresholds based on the final Euclidean distance of the object position to the goal position. Training on DTC objects appears to be especially helpful at enabling policies to make finer adjustments, improving pushing success rates at stricter thresholds. Additional experimental details can be found in the supplementary." + } + ] + } + ], + "index": 5 + }, + { + "type": "image", + "bbox": [ + 317, + 72, + 552, + 233 + ], + "blocks": [ + { + "bbox": [ + 317, + 72, + 552, + 233 + ], + "lines": [ + { + "bbox": [ + 317, + 72, + 552, + 233 + ], + "spans": [ + { + "bbox": [ + 317, + 72, + 552, + 233 + ], + "type": "image", + "image_path": "3b4d8b3f2caaf2fc9858feb7baa824272d8d29f523c49ef5c6ccb9d47ac20594.jpg" + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 313, + 243, + 555, + 299 + ], + "lines": [ + { + "bbox": [ + 313, + 243, + 555, + 299 + ], + "spans": [ + { + "bbox": [ + 313, + 243, + 555, + 299 + ], + "type": "text", + "content": "Figure 8. Success rates on robotic pushing task when training with our DTC objects and sampled Objaverse-XL objects. Particularly at lower error thresholds, policies trained on DTC objects outperform those trained using Objaverse-XL objects. Shaded bars represent sample standard deviations over policy training random seeds." + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "image_caption" + } + ], + "index": 6 + }, + { + "bbox": [ + 314, + 304, + 388, + 316 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 314, + 304, + 388, + 316 + ], + "spans": [ + { + "bbox": [ + 314, + 304, + 388, + 316 + ], + "type": "text", + "content": "5. Conclusion" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 313, + 324, + 555, + 432 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 324, + 555, + 432 + ], + "spans": [ + { + "bbox": [ + 313, + 324, + 555, + 432 + ], + "type": "text", + "content": "We presented a new large scale photorealistic 3D digital twin dataset with the real world recordings that contain its real world counterpart. We provide extensive evaluations of baselines on our DTC dataset serving as new benchmark for inverse rendering and novel view synthesis task. We also demonstrated that high quality digital twin models can be beneficial to applications in robotics domain. We believe our efforts can empower the research community to build and leverage digital twin models for future applications." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 313, + 434, + 556, + 519 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 434, + 556, + 519 + ], + "spans": [ + { + "bbox": [ + 313, + 434, + 556, + 519 + ], + "type": "text", + "content": "Limitations. Achieving high quality digital twin models currently requires deliberate hardware setup and human efforts in refinement. Solving this challenge without sacrificing quality can significantly further enhance the volume of digital twin models. Our hardware is also limited to objects within a certain size and can not yet recover objects that are deformable, highly specular, or transparent." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 313, + 521, + 556, + 653 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 521, + 556, + 653 + ], + "spans": [ + { + "bbox": [ + 313, + 521, + 556, + 653 + ], + "type": "text", + "content": "Future work. The existing digital twin model creation in DTC dataset involves lengthy post-processing and may require subjective human refinement, hindering the automation of model generation. However, recent advancements in physics-based differentiable rendering hold promise for enabling faster and more accurate creation of digital twins, especially for material reconstruction. Furthermore, building large-scale digital twins for applications will necessitate efforts to enhance the diversity in object appearance (e.g., transparent objects) and to capture additional attributes, such as physical properties and functionalities." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 313, + 654, + 556, + 712 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 654, + 556, + 712 + ], + "spans": [ + { + "bbox": [ + 313, + 654, + 556, + 712 + ], + "type": "text", + "content": "For robotics applications, while deploying manipulation policies learned in simulation to the real world remains generally challenging, we hope that the high-quality digital twin data provided by DTC can serve as a stepping stone towards effective sim-to-real transfer." + } + ] + } + ], + "index": 12 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 298, + 749, + 312, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 298, + 749, + 312, + 757 + ], + "spans": [ + { + "bbox": [ + 298, + 749, + 312, + 757 + ], + "type": "text", + "content": "760" + } + ] + } + ], + "index": 13 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 7 + }, + { + "para_blocks": [ + { + "bbox": [ + 55, + 72, + 297, + 122 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 72, + 297, + 122 + ], + "spans": [ + { + "bbox": [ + 55, + 72, + 297, + 122 + ], + "type": "text", + "content": "Acknowledgments. This work is in part supported by NSF CCRI #2120095, RI #2211258, and RI #2338203. YZ is in part supported by the Stanford Interdisciplinary Graduate Fellowship." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 56, + 133, + 115, + 145 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 133, + 115, + 145 + ], + "spans": [ + { + "bbox": [ + 56, + 133, + 115, + 145 + ], + "type": "text", + "content": "References" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 57, + 153, + 297, + 712 + ], + "type": "list", + "angle": 0, + "index": 16, + "blocks": [ + { + "bbox": [ + 61, + 153, + 297, + 175 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 61, + 153, + 297, + 175 + ], + "spans": [ + { + "bbox": [ + 61, + 153, + 297, + 175 + ], + "type": "text", + "content": "[1] Covision 3D object scanner. https://www.covisionmedia.ai/en/scanners.4" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 61, + 177, + 296, + 232 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 61, + 177, + 296, + 232 + ], + "spans": [ + { + "bbox": [ + 61, + 177, + 296, + 232 + ], + "type": "text", + "content": "[2] Kosmas Alexopoulos, Nikolaos Nikolakis, and George Chryssolouris. Digital twin-driven supervised machine learning for the development of artificial intelligence applications in manufacturing. International Journal of Computer Integrated Manufacturing, 33(5):429-439, 2020. 2" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 62, + 234, + 296, + 287 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 234, + 296, + 287 + ], + "spans": [ + { + "bbox": [ + 62, + 234, + 296, + 287 + ], + "type": "text", + "content": "[3] Moayad Aloqaily, Ouns Bouachir, Fakhri Karray, Ismaeel Al Ridhawi, and Abdulmotaleb El Saddik. Integrating digital twin and advanced intelligent technologies to realize the meta-verse. IEEE Consumer Electronics Magazine, 12(6):47-55, 2022. 2" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 62, + 290, + 296, + 334 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 290, + 296, + 334 + ], + "spans": [ + { + "bbox": [ + 62, + 290, + 296, + 334 + ], + "type": "text", + "content": "[4] Marcin Andrychowicz, Filip Wolski, Alex Ray, Jonas Schneider, Rachel Fong, Peter Welinder, Bob McGrew, Josh Tobin, OpenAI Pieter Abbeel, and Wojciech Zaremba. Hindsight experience replay. In NeurIPS, 2017. 8" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 62, + 335, + 296, + 368 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 335, + 296, + 368 + ], + "spans": [ + { + "bbox": [ + 62, + 335, + 296, + 368 + ], + "type": "text", + "content": "[5] Jonathan T Barron and Jitendra Malik. Shape, illumination, and reflectance from shading. IEEE TPAMI, 37(8):1670-1687, 2014. 3" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 62, + 371, + 294, + 392 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 371, + 294, + 392 + ], + "spans": [ + { + "bbox": [ + 62, + 371, + 294, + 392 + ], + "type": "text", + "content": "[6] Sean Bell, Kavita Bala, and Noah Snavely. Intrinsic images in the wild. ACM TOG, 33(4):1-12, 2014. 3" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 62, + 394, + 296, + 437 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 394, + 296, + 437 + ], + "spans": [ + { + "bbox": [ + 62, + 394, + 296, + 437 + ], + "type": "text", + "content": "[7] Sai Bi, Zexiang Xu, Kalyan Sunkavalli, Miloš Hasan, Yannick Hold-Geoffroy, David Kriegman, and Ravi Ramamoorthi. Deep reflectance volumes: Relightable reconstructions from multi-view photometric images. In ECCV, 2020. 3" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 62, + 440, + 296, + 472 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 440, + 296, + 472 + ], + "spans": [ + { + "bbox": [ + 62, + 440, + 296, + 472 + ], + "type": "text", + "content": "[8] Sai Bi, Zexiang Xu, Kalyan Sunkavalli, David Kriegman, and Ravi Ramamoorthi. Deep 3D capture: Geometry and reflectance from sparse multi-view images. In CVPR, 2020. 4" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 62, + 475, + 294, + 496 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 475, + 294, + 496 + ], + "spans": [ + { + "bbox": [ + 62, + 475, + 294, + 496 + ], + "type": "text", + "content": "[9] Martin Bichsel and Alex P Pentland. A simple algorithm for shape from shading. In CVPR, 1992. 3" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 57, + 498, + 296, + 530 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 498, + 296, + 530 + ], + "spans": [ + { + "bbox": [ + 57, + 498, + 296, + 530 + ], + "type": "text", + "content": "[10] Mark Boss, Varun Jampani, Kihwan Kim, Hendrik P.A. Lensch, and Jan Kautz. Two-shot Spatially-varying BRDF and Shape Estimation. In CVPR, 2020. 4" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 57, + 533, + 296, + 575 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 533, + 296, + 575 + ], + "spans": [ + { + "bbox": [ + 57, + 533, + 296, + 575 + ], + "type": "text", + "content": "[11] Mark Boss, Raphael Braun, Varun Jampani, Jonathan T. Barron, Ce Liu, and Hendrik P.A. Lensch. NeRD: Neural reflectance decomposition from image collections. In ICCV, 2021. 3, 7" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 57, + 578, + 296, + 621 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 578, + 296, + 621 + ], + "spans": [ + { + "bbox": [ + 57, + 578, + 296, + 621 + ], + "type": "text", + "content": "[12] Mark Boss, Varun Jampani, Raphael Braun, Ce Liu, Jonathan Barron, and Hendrik Lensch. Neural-PIL: Neural pre-integrated lighting for reflectance decomposition. In NeurIPS, 2021. 7" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 57, + 624, + 294, + 667 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 624, + 294, + 667 + ], + "spans": [ + { + "bbox": [ + 57, + 624, + 294, + 667 + ], + "type": "text", + "content": "[13] Mark Boss, Andreas Engelhardt, Abhishek Kar, Yuanzhen Li, Deqing Sun, Jonathan Barron, Hendrik Lensch, and Varun Jampani. SAMURAI: Shape and material from unconstrained real-world arbitrary image collections. In NeurIPS, 2022. 3" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 57, + 670, + 296, + 712 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 670, + 296, + 712 + ], + "spans": [ + { + "bbox": [ + 57, + 670, + 296, + 712 + ], + "type": "text", + "content": "[14] Guangyan Cai, Kai Yan, Zhao Dong, Ioannis Gkioulekas, and Shuang Zhao. Physics-based inverse rendering using combined implicit and explicit geometries. CGF, 41(4):129-138, 2022. 3" + } + ] + } + ], + "index": 15 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 316, + 73, + 555, + 714 + ], + "type": "list", + "angle": 0, + "index": 28, + "blocks": [ + { + "bbox": [ + 316, + 73, + 555, + 116 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 73, + 555, + 116 + ], + "spans": [ + { + "bbox": [ + 316, + 73, + 555, + 116 + ], + "type": "text", + "content": "[15] Wenzheng Chen, Joey Litalien, Jun Gao, Zian Wang, Clement Fuji Tsang, Sameh Khamis, Or Litany, and Sanja Fidler. DIBR++: learning to predict lighting and material with a hybrid differentiable renderer. In NeurIPS, 2021. 3" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 317, + 119, + 555, + 173 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 119, + 555, + 173 + ], + "spans": [ + { + "bbox": [ + 317, + 119, + 555, + 173 + ], + "type": "text", + "content": "[16] Jasmine Collins, Shubham Goel, Kenan Deng, Achleshwar Luthra, Leon Xu, Erhan Gundogdu, Xi Zhang, Tomas F Yago Vicente, Thomas Dideriksen, Himanshu Arora, et al. ABO: Dataset and benchmarks for real-world 3D object understanding. In CVPR, 2022. 2, 3" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 316, + 175, + 555, + 207 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 175, + 555, + 207 + ], + "spans": [ + { + "bbox": [ + 316, + 175, + 555, + 207 + ], + "type": "text", + "content": "[17] Erwin Coumans and Yunfei Bai. Pybullet, a python module for physics simulation for games, robotics and machine learning. http://pybullet.org, 2016-2019.8" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 317, + 209, + 555, + 253 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 209, + 555, + 253 + ], + "spans": [ + { + "bbox": [ + 317, + 209, + 555, + 253 + ], + "type": "text", + "content": "[18] Matt Deitke, Dustin Schwenk, Jordi Salvador, Luca Weihs, Oscar Michel, Eli VanderBilt, Ludwig Schmidt, Kiana Ehsani, Aniruddha Kembhavi, and Ali Farhadi. Objverse: A universe of annotated 3D objects. In CVPR, 2023. 2, 3" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 317, + 255, + 555, + 298 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 255, + 555, + 298 + ], + "spans": [ + { + "bbox": [ + 317, + 255, + 555, + 298 + ], + "type": "text", + "content": "[19] Matt Deitke, Ruoshi Liu, Matthew Wallingford, Huong Ngo, Oscar Michel, Aditya Kusupati, Alan Fan, Christian Laforte, Vikram Voleti, Samir Yitzhak Gadre, et al. Objverse-XL: A universe of " + }, + { + "bbox": [ + 317, + 255, + 555, + 298 + ], + "type": "inline_equation", + "content": "10\\mathrm{m} + 3\\mathrm{D}" + }, + { + "bbox": [ + 317, + 255, + 555, + 298 + ], + "type": "text", + "content": " objects. NeurIPS, 2024. 3, 7, 8" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 316, + 300, + 554, + 332 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 300, + 554, + 332 + ], + "spans": [ + { + "bbox": [ + 316, + 300, + 554, + 332 + ], + "type": "text", + "content": "[20] Yiming Ding, Carlos Florensa, Pieter Abbeel, and Mariano Philipp. Goal-conditioned imitation learning. In NeurIPS, 2019. 8" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 316, + 335, + 555, + 388 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 335, + 555, + 388 + ], + "spans": [ + { + "bbox": [ + 316, + 335, + 555, + 388 + ], + "type": "text", + "content": "[21] Laura Downs, Anthony Francis, Nate Koenig, Brandon Kinman, Ryan Hickman, Krista Reymann, Thomas B McHugh, and Vincent Vanhoucke. Google scanned objects: A high-quality dataset of 3D scanned household items. In ICRA, 2022. 2, 3" + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 317, + 391, + 555, + 620 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 391, + 555, + 620 + ], + "spans": [ + { + "bbox": [ + 317, + 391, + 555, + 620 + ], + "type": "text", + "content": "[22] Jakob Engel, Kiran Somasundaram, Michael Goesele, Albert Sun, Alexander Gamino, Andrew Turner, Arjang Talatof, Arnie Yuan, Bilal Souti, Brighid Meredith, Cheng Peng, Chris Sweeney, Cole Wilson, Dan Barnes, Daniel DeTone, David Caruso, Derek Valleroy, Dinesh Ginjupalli, Duncan Frost, Edward Miller, Elias Mueggler, Evgeniy Oleinik, Fan Zhang, Guruprasad Somasundaram, Gustavo Solaira, Harry Lanaras, Henry Howard-Jenkins, Huixuan Tang, Hyo Jin Kim, Jaime Rivera, Ji Luo, Jing Dong, Julian Straub, Kevin Bailey, Kevin Eckenhoff, Lingni Ma, Luis Pesqueira, Mark Schwesinger, Maurizio Monge, Nan Yang, Nick Charron, Nikhil Raina, Omkar Parkhi, Peter Borschowa, Pierre Moulon, Prince Gupta, Raul Mur-Artal, Robbie Pennington, Sachin Kulkarni, Sagar Miglani, Santosh Gondi, Saransh Solanki, Sean Diener, Shangyi Cheng, Simon Green, Steve Saarinen, Suvam Patra, Tassos Mourikis, Thomas Whelan, Tripti Singh, Vasileios Balntas, Vijay Baiyya, Wilson Dreewes, Xiaqing Pan, Yang Lou, Yipu Zhao, Yusuf Mansour, Yuyang Zou, Zhaoyang Lv, Zijian Wang, Mingfei Yan, Carl Ren, Renzo De Nardi, and Richard Newcombe. Project Aria: A new tool for egocentric multi-modal ai research, 2023. 5" + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 316, + 623, + 554, + 655 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 623, + 554, + 655 + ], + "spans": [ + { + "bbox": [ + 316, + 623, + 554, + 655 + ], + "type": "text", + "content": "[23] Ruohan Gao, Yen-Yu Chang, Shivani Mall, Li Fei-Fei, and Jiajun Wu. ObjectFolder: A dataset of objects with implicit visual, auditory, and tactile representations. In CoRL, 2021. 2" + } + ] + } + ], + "index": 25 + }, + { + "bbox": [ + 316, + 658, + 555, + 699 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 658, + 555, + 699 + ], + "spans": [ + { + "bbox": [ + 316, + 658, + 555, + 699 + ], + "type": "text", + "content": "[24] Dibya Ghosh, Abhishek Gupta, Ashwin Reddy, Justin Fu, Coline Devin, Benjamin Eysenbach, and Sergey Levine. Learning to reach goals via iterated supervised learning. In ICML, 2019. 8" + } + ] + } + ], + "index": 26 + }, + { + "bbox": [ + 316, + 702, + 554, + 714 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 702, + 554, + 714 + ], + "spans": [ + { + "bbox": [ + 316, + 702, + 554, + 714 + ], + "type": "text", + "content": "[25] Roger Grosse, Micah K Johnson, Edward H Adelson, and" + } + ] + } + ], + "index": 27 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 298, + 749, + 312, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 298, + 749, + 312, + 757 + ], + "spans": [ + { + "bbox": [ + 298, + 749, + 312, + 757 + ], + "type": "text", + "content": "761" + } + ] + } + ], + "index": 29 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 8 + }, + { + "para_blocks": [ + { + "bbox": [ + 56, + 72, + 296, + 713 + ], + "type": "list", + "angle": 0, + "index": 16, + "blocks": [ + { + "bbox": [ + 76, + 72, + 296, + 95 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 76, + 72, + 296, + 95 + ], + "spans": [ + { + "bbox": [ + 76, + 72, + 296, + 95 + ], + "type": "text", + "content": "William T Freeman. Ground truth dataset and baseline evaluations for intrinsic image algorithms. In ICCV, 2009. 3" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 56, + 96, + 296, + 140 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 96, + 296, + 140 + ], + "spans": [ + { + "bbox": [ + 56, + 96, + 296, + 140 + ], + "type": "text", + "content": "[26] Qiao Gu, Zhaoyang Lv, Duncan Frost, Simon Green, Julian Straub, and Chris Sweeney. Egolifter: Open-world 3d segmentation for egocentric perception. In ECCV, pages 382-400. Springer, 2024. 7" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 56, + 141, + 296, + 173 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 141, + 296, + 173 + ], + "spans": [ + { + "bbox": [ + 56, + 141, + 296, + 173 + ], + "type": "text", + "content": "[27] Agrim Gupta, Piotr Dollar, and Ross Girshick. LVIS: A dataset for large vocabulary instance segmentation. In CVPR, 2019. 4" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 56, + 175, + 296, + 217 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 175, + 296, + 217 + ], + "spans": [ + { + "bbox": [ + 56, + 175, + 296, + 217 + ], + "type": "text", + "content": "[28] Wassim Hamidouche, Lina Bariah, and Mérouane Debbah. Immersive media and massive twinning: Advancing toward the metaverse. IEEE Communications Magazine, 62(7):20-32, 2024. 2" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 56, + 220, + 296, + 253 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 220, + 296, + 253 + ], + "spans": [ + { + "bbox": [ + 56, + 220, + 296, + 253 + ], + "type": "text", + "content": "[29] Jon Hasselgren, Nikolai Hofmann, and Jacob Munkberg. Shape, light & material decomposition from images using Monte Carlo rendering and denoising. In NeurIPS, 2022. 4, 7" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 56, + 254, + 296, + 308 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 254, + 296, + 308 + ], + "spans": [ + { + "bbox": [ + 56, + 254, + 296, + 308 + ], + "type": "text", + "content": "[30] Yicong Hong, Kai Zhang, Jiuxiang Gu, Sai Bi, Yang Zhou, Difan Liu, Feng Liu, Kalyan Sunkavalli, Trung Bui, and Hao Tan. LRM: Large reconstruction model for single image to 3D. In The Twelfth International Conference on Learning Representations, 2024. 2" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 56, + 309, + 296, + 352 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 309, + 296, + 352 + ], + "spans": [ + { + "bbox": [ + 56, + 309, + 296, + 352 + ], + "type": "text", + "content": "[31] Binbin Huang, Zehao Yu, Anpei Chen, Andreas Geiger, and Shenghua Gao. 2D gaussian splatting for geometrically accurate radiance fields. In SIGGRAPH 2024 Conference Papers, 2024. 3, 7" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 56, + 354, + 296, + 399 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 354, + 296, + 399 + ], + "spans": [ + { + "bbox": [ + 56, + 354, + 296, + 399 + ], + "type": "text", + "content": "[32] Wenzel Jakob, Sébastien Speierer, Nicolas Roussel, Merlin Nimier-David, Delio Vicini, Tizian Zeltner, Baptiste Nicolet, Miguel Crespo, Vincent Leroy, and Ziyi Zhang. Mitsuba 3 renderer, 2022. https://mitsuba-renderer.org. 5" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 56, + 399, + 296, + 432 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 399, + 296, + 432 + ], + "spans": [ + { + "bbox": [ + 56, + 399, + 296, + 432 + ], + "type": "text", + "content": "[33] Michael Janner, Jiajun Wu, Tejas Kulkarni, Ilker Yildirim, and Joshua B Tenenbaum. Self-supervised intrinsic image decomposition. In NeurIPS, 2017. 4" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 56, + 434, + 296, + 465 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 434, + 296, + 465 + ], + "spans": [ + { + "bbox": [ + 56, + 434, + 296, + 465 + ], + "type": "text", + "content": "[34] Rasmus Jensen, Anders Dahl, George Vogiatzis, Engin Tola, and Henrik Aanæs. Large scale multi-view stereopsis evaluation. In CVPR, 2014. 2, 3" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 56, + 467, + 296, + 500 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 467, + 296, + 500 + ], + "spans": [ + { + "bbox": [ + 56, + 467, + 296, + 500 + ], + "type": "text", + "content": "[35] Haian Jin, Isabella Liu, Peijia Xu, Xiaoshuai Zhang, Songfang Han, Sai Bi, Xiaowei Zhou, Zexiang Xu, and Hao Su. TensoIR: Tensorial inverse rendering. In CVPR, 2023. 2" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 56, + 501, + 296, + 534 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 501, + 296, + 534 + ], + "spans": [ + { + "bbox": [ + 56, + 501, + 296, + 534 + ], + "type": "text", + "content": "[36] James T Kajiya. The rendering equation. In Proceedings of the 13th annual conference on Computer graphics and interactive techniques, pages 143-150, 1986. 3" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 56, + 535, + 296, + 567 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 535, + 296, + 567 + ], + "spans": [ + { + "bbox": [ + 56, + 535, + 296, + 567 + ], + "type": "text", + "content": "[37] Bernhard Kerbl, Georgios Kopanas, Thomas Leimkuhler, and George Drettakis. 3D gaussian splattering for real-time radiance field rendering. ACM TOG, 2023. 2, 3, 7" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 56, + 568, + 296, + 612 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 568, + 296, + 612 + ], + "spans": [ + { + "bbox": [ + 56, + 568, + 296, + 612 + ], + "type": "text", + "content": "[38] Zhengfei Kuang, Kyle Olszewski, Mengelei Chai, Zeng Huang, Panos Achlioptas, and Sergey Tulyakov. NeROIC: neural rendering of objects from online image collections. ACM TOG, 41(4):1-12, 2022. 3" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 56, + 613, + 296, + 667 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 613, + 296, + 667 + ], + "spans": [ + { + "bbox": [ + 56, + 613, + 296, + 667 + ], + "type": "text", + "content": "[39] Zhengfei Kuang, Yunzhi Zhang, Hong-Xing Yu, Samir Agarwala, Elliott Wu, Jiajun Wu, et al. Stanford-orb: a real-world 3D object inverse rendering benchmark. In Advances in Neural Information Processing Systems Datasets and Benchmarks Track, 2023. 3, 5, 7, 8" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 56, + 669, + 296, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 669, + 296, + 713 + ], + "spans": [ + { + "bbox": [ + 56, + 669, + 296, + 713 + ], + "type": "text", + "content": "[40] Samuli Laine, Janne Hellsten, Tero Karras, Yeongho Seol, Jaakko Lehtinen, and Timo Aila. Modular primitives for high-performance differentiable rendering. ACM Transactions on Graphics, 39(6), 2020. 5" + } + ] + } + ], + "index": 15 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 316, + 72, + 554, + 713 + ], + "type": "list", + "angle": 0, + "index": 33, + "blocks": [ + { + "bbox": [ + 316, + 72, + 554, + 127 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 72, + 554, + 127 + ], + "spans": [ + { + "bbox": [ + 316, + 72, + 554, + 127 + ], + "type": "text", + "content": "[41] Jiahao Li, Hao Tan, Kai Zhang, Zexiang Xu, Fujun Luan, Yinghao Xu, Yicong Hong, Kalyan Sunkavalli, Greg Shakhnarovich, and Sai Bi. Instant3D: Fast text-to-3D with sparse-view generation and large reconstruction model. In ICLR, 2024. 2" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 316, + 129, + 554, + 172 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 129, + 554, + 172 + ], + "spans": [ + { + "bbox": [ + 316, + 129, + 554, + 172 + ], + "type": "text", + "content": "[42] Min Li, Zhenglong Zhou, Zhe Wu, Boxin Shi, Changyu Diao, and Ping Tan. Multi-view photometric stereo: A robust solution and benchmark dataset for spatially varying isotropic materials. IEEE TIP, 29:4159-4173, 2020. 3" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 316, + 174, + 554, + 239 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 174, + 554, + 239 + ], + "spans": [ + { + "bbox": [ + 316, + 174, + 554, + 239 + ], + "type": "text", + "content": "[43] Xuanlin Li, Kyle Hsu, Jiayuan Gu, Karl Pertsch, Oier Mees, Homer Rich Walke, Chuyuan Fu, Ishikaa Lunawat, Isabel Sieh, Sean Kirmani, Sergey Levine, Jiajun Wu, Chelsea Finn, Hao Su, Quan Vuong, and Ted Xiao. Evaluating real-world robot manipulation policies in simulation. arXiv preprint arXiv:2405.05941, 2024. 2" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 316, + 240, + 554, + 274 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 240, + 554, + 274 + ], + "spans": [ + { + "bbox": [ + 316, + 240, + 554, + 274 + ], + "type": "text", + "content": "[44] Zhengqin Li, Kalyan Sunkavalli, and Manmohan Chandraker. Materials for masses: SVBRDF acquisition with a single mobile phone image. In ECCV, 2018. 3" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 316, + 274, + 554, + 318 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 274, + 554, + 318 + ], + "spans": [ + { + "bbox": [ + 316, + 274, + 554, + 318 + ], + "type": "text", + "content": "[45] Zhengqin Li, Zexiang Xu, Ravi Ramamoorthi, Kalyan Sunkavalli, and Manmohan Chandraker. Learning to reconstruct shape and spatially-varying reflectance from a single image. ACM TOG, 37(6):1-11, 2018. 3, 4" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 316, + 319, + 554, + 373 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 319, + 554, + 373 + ], + "spans": [ + { + "bbox": [ + 316, + 319, + 554, + 373 + ], + "type": "text", + "content": "[46] Isabella Liu, Linghao Chen, Ziyang Fu, Liwen Wu, Haian Jin, Zhong Li, Chin Ming Ryan Wong, Yi Xu, Ravi Ramamoorthi, Zexiang Xu, and Hao Su. OpenIllumination: A multi-illumination dataset for inverse rendering evaluation on real objects. NeurIPS, 2023. 2, 3" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 316, + 375, + 554, + 397 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 375, + 554, + 397 + ], + "spans": [ + { + "bbox": [ + 316, + 375, + 554, + 397 + ], + "type": "text", + "content": "[47] Stephen Lombardi and Ko Nishino. Reflectance and natural illumination from a single image. In ECCV, 2012. 3" + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 316, + 398, + 554, + 420 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 398, + 554, + 420 + ], + "spans": [ + { + "bbox": [ + 316, + 398, + 554, + 420 + ], + "type": "text", + "content": "[48] Stephen Lombardi and Ko Nishino. Reflectance and illumination recovery in the wild. IEEE TPAMI, 2015. 3" + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 316, + 421, + 554, + 454 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 421, + 554, + 454 + ], + "spans": [ + { + "bbox": [ + 316, + 421, + 554, + 454 + ], + "type": "text", + "content": "[49] Fujun Luan, Shuang Zhao, Kavita Bala, and Zhao Dong. Unified shape and SVBRDF recovery using differentiable Monte Carlo rendering. CGF, 40(4):101-113, 2021. 3" + } + ] + } + ], + "index": 25 + }, + { + "bbox": [ + 316, + 455, + 554, + 477 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 455, + 554, + 477 + ], + "spans": [ + { + "bbox": [ + 316, + 455, + 554, + 477 + ], + "type": "text", + "content": "[50] Stephen Robert Marschner. Inverse Rendering for Computer Graphics. Cornell University, 1998. 3" + } + ] + } + ], + "index": 26 + }, + { + "bbox": [ + 316, + 478, + 554, + 499 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 478, + 554, + 499 + ], + "spans": [ + { + "bbox": [ + 316, + 478, + 554, + 499 + ], + "type": "text", + "content": "[51] Nelson Max. Optical models for direct volume rendering. IEEE TVCG, 1(2):99-108, 1995. 3" + } + ] + } + ], + "index": 27 + }, + { + "bbox": [ + 316, + 501, + 554, + 544 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 501, + 554, + 544 + ], + "spans": [ + { + "bbox": [ + 316, + 501, + 554, + 544 + ], + "type": "text", + "content": "[52] Ben Mildenhall, Pratul P Srinivasan, Matthew Tancik, Jonathan T Barron, Ravi Ramamoorthi, and Ren Ng. NeRF: Representing scenes as neural radiance fields for view synthesis. In ECCV, 2020. 2, 3, 7" + } + ] + } + ], + "index": 28 + }, + { + "bbox": [ + 316, + 545, + 554, + 590 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 545, + 554, + 590 + ], + "spans": [ + { + "bbox": [ + 316, + 545, + 554, + 590 + ], + "type": "text", + "content": "[53] Jacob Munkberg, Jon Hasselgren, Tianchang Shen, Jun Gao, Wenzheng Chen, Alex Evans, Thomas Müller, and Sanja Fidler. Extracting triangular 3D models, materials, and lighting from images. In CVPR, 2022. 4, 7" + } + ] + } + ], + "index": 29 + }, + { + "bbox": [ + 316, + 590, + 554, + 644 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 590, + 554, + 644 + ], + "spans": [ + { + "bbox": [ + 316, + 590, + 554, + 644 + ], + "type": "text", + "content": "[54] Soroush Nasiriany, Abhiram Maddukuri, Lance Zhang, Adeet Parikh, Aaron Lo, Abhishek Joshi, Ajay Mandlekar, and Yuke Zhu. Robocasa: Large-scale simulation of everyday tasks for generalist robots. In Robotics: Science and Systems (RSS), 2024. 7" + } + ] + } + ], + "index": 30 + }, + { + "bbox": [ + 316, + 646, + 554, + 689 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 646, + 554, + 689 + ], + "spans": [ + { + "bbox": [ + 316, + 646, + 554, + 689 + ], + "type": "text", + "content": "[55] Michael Niemeyer, Lars Mescheder, Michael Oechsle, and Andreas Geiger. Differentiable volumetric rendering: Learning implicit 3D representations without 3D supervision. In CVPR, 2020. 4" + } + ] + } + ], + "index": 31 + }, + { + "bbox": [ + 316, + 691, + 554, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 691, + 554, + 713 + ], + "spans": [ + { + "bbox": [ + 316, + 691, + 554, + 713 + ], + "type": "text", + "content": "[56] Geoffrey Oxholm and Ko Nishino. Multiview shape and reflectance from natural illumination. In CVPR, 2014. 3" + } + ] + } + ], + "index": 32 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 298, + 749, + 313, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 298, + 749, + 313, + 757 + ], + "spans": [ + { + "bbox": [ + 298, + 749, + 313, + 757 + ], + "type": "text", + "content": "762" + } + ] + } + ], + "index": 34 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 9 + }, + { + "para_blocks": [ + { + "bbox": [ + 56, + 72, + 296, + 713 + ], + "type": "list", + "angle": 0, + "index": 16, + "blocks": [ + { + "bbox": [ + 56, + 72, + 296, + 94 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 72, + 296, + 94 + ], + "spans": [ + { + "bbox": [ + 56, + 72, + 296, + 94 + ], + "type": "text", + "content": "[57] Geoffrey Oxholm and Ko Nishino. Shape and reflectance estimation in the wild. IEEE TPAMI, 2015. 3" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 56, + 95, + 296, + 161 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 95, + 296, + 161 + ], + "spans": [ + { + "bbox": [ + 56, + 95, + 296, + 161 + ], + "type": "text", + "content": "[58] Xiaqing Pan, Nicholas Charron, Yongqian Yang, Scott Peters, Thomas Whelan, Chen Kong, Omkar Parkhi, Richard Newcombe, and Yuheng (Carl) Ren. Aria digital twin: A new benchmark dataset for egocentric 3D machine perception. In Proceedings of the IEEE/CVF International Conference on Computer Vision (ICCV), pages 20133-20143, 2023. 3" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 56, + 162, + 296, + 204 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 162, + 296, + 204 + ], + "spans": [ + { + "bbox": [ + 56, + 162, + 296, + 204 + ], + "type": "text", + "content": "[59] Viktor Rudnev, Mohamed Elgharib, William Smith, Lingjie Liu, Vladislav Golyanik, and Christian Theobalt. NeRF for outdoor scene relighting. In ECCV, pages 615-631. Springer, 2022. 3" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 56, + 205, + 296, + 248 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 205, + 296, + 248 + ], + "spans": [ + { + "bbox": [ + 56, + 205, + 296, + 248 + ], + "type": "text", + "content": "[60] Tianchang Shen, Jun Gao, Kangxue Yin, Ming-Yu Liu, and Sanja Fidler. Deep marching tetrahedra: a hybrid representation for high-resolution 3D shape synthesis. In NeurIPS, 2021. 7" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 56, + 250, + 296, + 282 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 250, + 296, + 282 + ], + "spans": [ + { + "bbox": [ + 56, + 250, + 296, + 282 + ], + "type": "text", + "content": "[61] Jian Shi, Yue Dong, Hao Su, and Stella X Yu. Learning non-Lambertian object intrinsics across ShapeNet categories. In CVPR, 2017. 2, 3, 4" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 56, + 283, + 296, + 326 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 283, + 296, + 326 + ], + "spans": [ + { + "bbox": [ + 56, + 283, + 296, + 326 + ], + "type": "text", + "content": "[62] Cheng Sun, Guangyan Cai, Zhengqin Li, Kai Yan, Cheng Zhang, Carl Marshall, Jia-Bin Huang, Shuang Zhao, and Zhao Dong. Neural-PBIR reconstruction of shape, material, and illumination. In ICCV, 2023. 2, 4, 5, 6" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 56, + 327, + 296, + 370 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 327, + 296, + 370 + ], + "spans": [ + { + "bbox": [ + 56, + 327, + 296, + 370 + ], + "type": "text", + "content": "[63] Stephen Tian, Yancheng Cai, Hong-Xing Yu, Sergey Zakharov, Katherine Liu, Adrien Gaidon, Yunzhu Li, and Ji-jun Wu. Multi-object manipulation via object-centric neural scattering functions. In CVPR, 2023. 7" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 56, + 371, + 296, + 415 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 371, + 296, + 415 + ], + "spans": [ + { + "bbox": [ + 56, + 371, + 296, + 415 + ], + "type": "text", + "content": "[64] Marcel Torne, Anthony Simeonov, Zechu Li, April Chan, Tao Chen, Abhishek Gupta, and Pulkit Agrawal. Reconciling reality through simulation: A real-to-sim-to-real approach for robust manipulation. Arxiv, 2024. 7" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 56, + 415, + 296, + 459 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 415, + 296, + 459 + ], + "spans": [ + { + "bbox": [ + 56, + 415, + 296, + 459 + ], + "type": "text", + "content": "[65] Marco Toschi, Riccardo De Matteo, Riccardo Spezialetti, Daniele De Gregorio, Luigi Di Stefano, and Samuele Salti. Relight my NeRF: A dataset for novel view synthesis and relighting of real world objects. In CVPR, 2023. 3" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 56, + 459, + 296, + 502 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 459, + 296, + 502 + ], + "spans": [ + { + "bbox": [ + 56, + 459, + 296, + 502 + ], + "type": "text", + "content": "[66] Peng Wang, Lingjie Liu, Yuan Liu, Christian Theobalt, Taku Komura, and Wenping Wang. NeuS: Learning neural implicit surfaces by volume rendering for multi-view reconstruction. In NeurIPS, 2021. 4" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 56, + 503, + 296, + 535 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 503, + 296, + 535 + ], + "spans": [ + { + "bbox": [ + 56, + 503, + 296, + 535 + ], + "type": "text", + "content": "[67] Henrique Weber, Donald Prévost, and Jean-François Lalonde. Learning to estimate indoor lighting from 3D objects. In 3DV, 2018. 3" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 56, + 536, + 296, + 568 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 536, + 296, + 568 + ], + "spans": [ + { + "bbox": [ + 56, + 536, + 296, + 568 + ], + "type": "text", + "content": "[68] Xinyue Wei, Minghua Liu, Zhan Ling, and Hao Su. Approximate convex decomposition for 3D meshes with collision-aware concavity and tree search. ACM TOG, 2022. 7" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 56, + 570, + 296, + 613 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 570, + 296, + 613 + ], + "spans": [ + { + "bbox": [ + 56, + 570, + 296, + 613 + ], + "type": "text", + "content": "[69] Xinyue Wei, Kai Zhang, Sai Bi, Hao Tan, Fujun Luan, Valentin Deschaintre, Kalyan Sunkavalli, Hao Su, and Zexiang Xu. MeshLRM: Large reconstruction model for high-quality mesh. arXiv preprint arXiv:2404.12385, 2024. 2" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 56, + 614, + 296, + 635 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 614, + 296, + 635 + ], + "spans": [ + { + "bbox": [ + 56, + 614, + 296, + 635 + ], + "type": "text", + "content": "[70] Felix Wimbauer, Shangzhe Wu, and Christian Rupprecht. De-rendering 3D objects in the wild. In CVPR, 2022. 3, 4" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 56, + 637, + 296, + 679 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 637, + 296, + 679 + ], + "spans": [ + { + "bbox": [ + 56, + 637, + 296, + 679 + ], + "type": "text", + "content": "[71] Haoqian Wu, Zhipeng Hu, Lincheng Li, Yongqiang Zhang, Changjie Fan, and Xin Yu. NeFII: Inverse rendering for reflectance decomposition with near-field indirect illumination. In CVPR, 2023. 4, 7" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 56, + 681, + 296, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 681, + 296, + 713 + ], + "spans": [ + { + "bbox": [ + 56, + 681, + 296, + 713 + ], + "type": "text", + "content": "[72] Shangzhe Wu, Ameesh Makadia, Jiajun Wu, Noah Snavely, Richard Tucker, and Angjoo Kanazawa. De-rendering the world's revolutionary artefacts. In CVPR, 2021. 3, 4" + } + ] + } + ], + "index": 15 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 316, + 72, + 555, + 656 + ], + "type": "list", + "angle": 0, + "index": 31, + "blocks": [ + { + "bbox": [ + 316, + 72, + 555, + 139 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 72, + 555, + 139 + ], + "spans": [ + { + "bbox": [ + 316, + 72, + 555, + 139 + ], + "type": "text", + "content": "[73] Tong Wu, Jiarui Zhang, Xiao Fu, Yuxin Wang, Liang Pan Jiawei Ren, Wayne Wu, Lei Yang, Jiaqi Wang, Chen Qian, Dahua Lin, and Ziwei Liu. OmniObject3D: Large-vocabulary 3D object dataset for realistic perception, reconstruction and generation. In IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), 2023. 2, 3" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 316, + 140, + 555, + 172 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 140, + 555, + 172 + ], + "spans": [ + { + "bbox": [ + 316, + 140, + 555, + 172 + ], + "type": "text", + "content": "[74] Kohei Yamashita, Yuto Enyo, Shohei Nobuhara, and Ko Nishino. nLMVS-Net: Deep non-lambertian multi-view stereo. In WACV, 2023. 3" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 316, + 174, + 555, + 205 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 174, + 555, + 205 + ], + "spans": [ + { + "bbox": [ + 316, + 174, + 555, + 205 + ], + "type": "text", + "content": "[75] Kai Yan and Shuang Zhao. PSDR-JIT: Path-space differentiable renderer, 2020. https://github.com/andyyankai/psdr-jit.5" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 316, + 207, + 555, + 251 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 207, + 555, + 251 + ], + "spans": [ + { + "bbox": [ + 316, + 207, + 555, + 251 + ], + "type": "text", + "content": "[76] Lior Yariv, Yoni Kasten, Dror Moran, Meirav Galun, Matan Atzmon, Basri Ronen, and Yaron Lipman. Multiview neural surface reconstruction by disentangling geometry and appearance. In NeurIPS, 2020. 7" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 316, + 252, + 555, + 306 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 252, + 555, + 306 + ], + "spans": [ + { + "bbox": [ + 316, + 252, + 555, + 306 + ], + "type": "text", + "content": "[77] Vickie Ye, Ruilong Li, Justin Kerr, Matias Turkulainen, Brent Yi, Zhuoyang Pan, Otto Seiskari, Jianbo Ye, Jeffrey Hu, Matthew Tancik, and Angjoo Kanazawa. gsplat: An open-source library for Gaussian splatting. arXiv preprint arXiv:2409.06765, 2024. 7" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 316, + 308, + 555, + 342 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 308, + 555, + 342 + ], + "spans": [ + { + "bbox": [ + 316, + 308, + 555, + 342 + ], + "type": "text", + "content": "[78] Hong-Xing Yu, Samir Agarwala, Charles Herrmann, Richard Szeliski, Noah Snavely, Jiajun Wu, and Deqing Sun. Accidental light probes. In CVPR, 2023. 3, 5" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 316, + 342, + 555, + 396 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 342, + 555, + 396 + ], + "spans": [ + { + "bbox": [ + 316, + 342, + 555, + 396 + ], + "type": "text", + "content": "[79] Hong-Xing Yu, Michelle Guo, Alireza Fathi, Yen-Yu Chang, Eric Ryan Chan, Ruohan Gao, Thomas Funkhouser, and Jia-jun Wu. Learning object-centric neural scattering functions for free-viewpoint relighting and scene composition. TMLR, 2023. 3" + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 316, + 398, + 555, + 441 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 398, + 555, + 441 + ], + "spans": [ + { + "bbox": [ + 316, + 398, + 555, + 441 + ], + "type": "text", + "content": "[80] Kai Zhang, Fujun Luan, Qianqian Wang, Kavita Bala, and Noah Snavely. PhySG: Inverse rendering with spherical gaussians for physics-based material editing and relighting. In CVPR, 2021. 4, 7" + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 316, + 442, + 555, + 475 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 442, + 555, + 475 + ], + "spans": [ + { + "bbox": [ + 316, + 442, + 555, + 475 + ], + "type": "text", + "content": "[81] Kai Zhang, Fujun Luan, Zhengqi Li, and Noah Snavely. IRON: Inverse rendering by optimizing neural SDFs and materials from photometric images. In CVPR, 2022. 4" + } + ] + } + ], + "index": 25 + }, + { + "bbox": [ + 316, + 476, + 555, + 520 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 476, + 555, + 520 + ], + "spans": [ + { + "bbox": [ + 316, + 476, + 555, + 520 + ], + "type": "text", + "content": "[82] Kai Zhang, Sai Bi, Hao Tan, Yuanbo Xiangli, Nanxuan Zhao, Kalyan Sunkavalli, and Zexiang Xu. GS-LRM: Large reconstruction model for 3D gaussian splatting. arXiv preprint arXiv:2404.19702, 2024. 2" + } + ] + } + ], + "index": 26 + }, + { + "bbox": [ + 316, + 521, + 555, + 553 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 521, + 555, + 553 + ], + "spans": [ + { + "bbox": [ + 316, + 521, + 555, + 553 + ], + "type": "text", + "content": "[83] Ruo Zhang, Ping-Sing Tsai, James Edwin Cryer, and Mubarak Shah. Shape-from-shading: a survey. IEEE TPAMI, 21(8): 690-706, 1999. 3" + } + ] + } + ], + "index": 27 + }, + { + "bbox": [ + 316, + 555, + 555, + 599 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 555, + 555, + 599 + ], + "spans": [ + { + "bbox": [ + 316, + 555, + 555, + 599 + ], + "type": "text", + "content": "[84] Xiuming Zhang, Pratul P Srinivasan, Boyang Deng, Paul Debevec, William T Freeman, and Jonathan T Barron. NeR-Factor: Neural factorization of shape and reflectance under an unknown illumination. ACM TOG, 40(6):1-18, 2021. 3" + } + ] + } + ], + "index": 28 + }, + { + "bbox": [ + 316, + 600, + 555, + 633 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 600, + 555, + 633 + ], + "spans": [ + { + "bbox": [ + 316, + 600, + 555, + 633 + ], + "type": "text", + "content": "[85] Yuanqing Zhang, Jiaming Sun, Xingyi He, Huan Fu, Rongfei Jia, and Xiaowei Zhou. Modeling indirect illumination for inverse rendering. In CVPR, 2022. 4, 7" + } + ] + } + ], + "index": 29 + }, + { + "bbox": [ + 316, + 634, + 555, + 656 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 634, + 555, + 656 + ], + "spans": [ + { + "bbox": [ + 316, + 634, + 555, + 656 + ], + "type": "text", + "content": "[86] Yunzhi Zhang, Shangzhe Wu, Noah Snavely, and Jiajun Wu. Seeing a rose in five thousand ways. In CVPR, 2023. 4" + } + ] + } + ], + "index": 30 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 298, + 749, + 313, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 298, + 749, + 313, + 757 + ], + "spans": [ + { + "bbox": [ + 298, + 749, + 313, + 757 + ], + "type": "text", + "content": "763" + } + ] + } + ], + "index": 32 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 10 + } + ], + "_backend": "vlm", + "_version_name": "2.6.4" +} \ No newline at end of file diff --git a/2025/Dinomaly_ The Less Is More Philosophy in Multi-Class Unsupervised Anomaly Detection/269a00de-90dc-4a5c-8e52-737661810fb4_content_list.json b/2025/Dinomaly_ The Less Is More Philosophy in Multi-Class Unsupervised Anomaly Detection/269a00de-90dc-4a5c-8e52-737661810fb4_content_list.json new file mode 100644 index 0000000000000000000000000000000000000000..e3725cecfedde1923fd9e0a198600261be93a46a --- /dev/null +++ b/2025/Dinomaly_ The Less Is More Philosophy in Multi-Class Unsupervised Anomaly Detection/269a00de-90dc-4a5c-8e52-737661810fb4_content_list.json @@ -0,0 +1,1621 @@ +[ + { + "type": "text", + "text": "Dinomaly: The Less Is More Philosophy in Multi-Class Unsupervised Anomaly Detection", + "text_level": 1, + "bbox": [ + 267, + 130, + 728, + 176 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Jia Guo $^{1}$ Shuai Lu $^{2}$ Weihang Zhang $^{2\\dagger}$ Fang Chen $^{3}$ Huiqi Li $^{2\\dagger}$ Hongen Liao $^{1,3\\boxtimes}$", + "bbox": [ + 148, + 200, + 846, + 223 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "$^{1}$ School of Biomedical Engineering, Tsinghua University, Beijing, China $^{2}$ Beijing Institute of Technology, Beijing, China", + "bbox": [ + 210, + 238, + 785, + 273 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "$^{3}$ School of Biomedical Engineering, Shanghai Jiao Tong University, Shanghai, China", + "bbox": [ + 163, + 273, + 833, + 292 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "guojia.jeremy@gmail.com lushuaie@163.com zhangweihang@bit.edu.cn chen-fang@sjtu.edu.cn huiqili@bit.edu.cn liao@tsinghua.edu.cn", + "bbox": [ + 209, + 295, + 781, + 325 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Abstract", + "text_level": 1, + "bbox": [ + 248, + 377, + 326, + 393 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Recent studies highlighted a practical setting of unsupervised anomaly detection (UAD) that builds a unified model for multi-class images. Despite various advancements addressing this challenging task, the detection performance under the multi-class setting still lags far behind state-of-the-art class-separated models. Our research aims to bridge this substantial performance gap. In this paper, we present Dinomaly, a minimalist reconstruction-based anomaly detection framework that harnesses pure Transformer architectures without relying on complex designs, additional modules, or specialized tricks. Given this powerful framework consisting of only Attentions and MLPs, we found four simple components that are essential to multi-class anomaly detection: (1) Scalable foundation Transformers that extracts universal and discriminative features, (2) Noisy Bottleneck where pre-existing Dropouts do all the noise injection tricks, (3) Linear Attention that naturally cannot focus, and (4) Loose Reconstruction that does not force layer-to-layer and point-by-point reconstruction. Extensive experiments are conducted across popular anomaly detection benchmarks including MVtec-AD, VisA, Real-IAD, etc. Our proposed Dinomaly achieves impressive image-level AUROC of $99.6\\%$ , $98.7\\%$ , and $89.3\\%$ on the three datasets respectively, which is not only superior to state-of-the-art multi-class UAD methods, but also achieves the most advanced class-separated UAD records. Code is available at: https://github.com/guojiajeremy/Dinomaly", + "bbox": [ + 89, + 415, + 485, + 838 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "1. Introduction", + "text_level": 1, + "bbox": [ + 513, + 361, + 643, + 376 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Unsupervised anomaly detection (UAD) aims to detect abnormal patterns from normal images and further localize the anomalous regions. Because of the diversity of potential anomalies and their scarcity, this task is proposed to model the accessible training sets containing only normal samples as an unsupervised paradigm. UAD has a wide range of applications, e.g., industrial defect detection [3], medical disease screening [13], and video surveillance [37], addressing the difficulty of collecting and labeling all possible anomalies in these scenarios.", + "bbox": [ + 511, + 386, + 906, + 537 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Conventional works on UAD build a separate model for each object category, as shown in Figure 1(a). However, this one-class-one-model setting entails substantial storage overhead for saving models [60], especially when the application scenario necessitates a large number of object classes. For UAD methods, a compact boundary of normal patterns is vital to distinguish anomalies. Once the intra-normal patterns become exceedingly complicated due to various classes, the corresponding distribution becomes challenging to measure, consequently harming the detection performance. Recently, UniAD [60] and successive studies have been proposed to train a unified model for multi-class anomaly detection (MUAD), as shown in Figure 1(b). Under this setting, the \"identity mapping\" that directly copies the input as the output regardless of normal or anomaly harms the performance of conventional methods [60]. This phenomenon is caused by the diversity of multi-class normal patterns that drive the network to generalize on unseen patterns.", + "bbox": [ + 511, + 537, + 908, + 824 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Within two years, a number of methods have been proposed to address MUAD, such as neighbor-masked attention [60], synthetic anomalies [68], vector quantization [36], diffusion model [16, 59], and state space model (Mamba) [17]. However, there is still a non-negligible per", + "bbox": [ + 511, + 825, + 908, + 902 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "CVF", + "bbox": [ + 106, + 2, + 181, + 42 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore.", + "bbox": [ + 236, + 0, + 810, + 46 + ], + "page_idx": 0 + }, + { + "type": "page_footnote", + "text": "Corresponding author. † Former advisors when the project initiated.", + "bbox": [ + 112, + 886, + 482, + 900 + ], + "page_idx": 0 + }, + { + "type": "page_number", + "text": "20405", + "bbox": [ + 478, + 944, + 517, + 957 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/e77de717e7226e870fe1250726262007dd1a4ff99ce23a99dd348328cf5b24f7.jpg", + "image_caption": [ + "(a) Class-Separated UAD Setting" + ], + "image_footnote": [], + "bbox": [ + 114, + 97, + 272, + 189 + ], + "page_idx": 1 + }, + { + "type": "image", + "img_path": "images/21af34168446ad566b0d66a8553a9bddb83a215a657a4e8c6b1775ed7512a181.jpg", + "image_caption": [ + "MVTec-AD (15 classes)" + ], + "image_footnote": [], + "bbox": [ + 287, + 95, + 491, + 219 + ], + "page_idx": 1 + }, + { + "type": "image", + "img_path": "images/752ff9fe62b375b33baafdeedcfd0b9ae29b5ad861e630b336cf9b22f9bea3ab.jpg", + "image_caption": [ + "VisA (12 classes)" + ], + "image_footnote": [], + "bbox": [ + 493, + 95, + 679, + 219 + ], + "page_idx": 1 + }, + { + "type": "image", + "img_path": "images/6de827691005c6a5c2758f19abd6cacf12119cbd3dd32c9c6d1fa1f13c2a2eae.jpg", + "image_caption": [ + "Real-IAD (30 classes)" + ], + "image_footnote": [], + "bbox": [ + 686, + 95, + 870, + 219 + ], + "page_idx": 1 + }, + { + "type": "image", + "img_path": "images/e357a9a888575f41ce9e2b5de7c3db2424bd3b9e759acf39d47a4e7ea73dcc37.jpg", + "image_caption": [ + "(b)", + "Model-Unified", + "Multi-Class UAD Setting" + ], + "image_footnote": [], + "bbox": [ + 114, + 227, + 269, + 311 + ], + "page_idx": 1 + }, + { + "type": "image", + "img_path": "images/bfa686fdd5f1898aaee9f58a6810b419098f70670743e14fdd62616ac3beeb8b.jpg", + "image_caption": [ + "(d) Scaling Properties of Dinomaly on MVTec-AD. $\\dagger$ : default setting." + ], + "image_footnote": [], + "bbox": [ + 295, + 244, + 485, + 325 + ], + "page_idx": 1 + }, + { + "type": "image", + "img_path": "images/7c9079b35c2f53586e2ae811f1f5e9bee8e28950af1f243868cd1825e5748932.jpg", + "image_caption": [ + "(c) Image-Level AUROC on 3 Benchmarks" + ], + "image_footnote": [], + "bbox": [ + 496, + 244, + 681, + 327 + ], + "page_idx": 1 + }, + { + "type": "image", + "img_path": "images/3a384fc4fb010232dc6e2160c574502a492a9d1cb6613d52a8df42eb04d61f13.jpg", + "image_caption": [ + "Figure 1. Setting, benchmarking, and scaling of Dinomaly. (a) Task setting of class-separated UAD. (b) Task setting of MUAD. (c) Comparison with previous SoTA methods on MVtec-AD [3], VisA [70], and Real-IAD [54]. (d) Scaling properties of Dinomaly." + ], + "image_footnote": [], + "bbox": [ + 691, + 246, + 885, + 328 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "formance gap between the state-of-the-art (SoTA) MUAD methods and class-separated UAD methods, restricting the practicability of implementing unified models, as shown in Figure 1(c). In addition, previous methods employ modules and architectures delicately designed, which may not be straightforward, and consequently suffer from limited universality and ease-of-use [18, 36].", + "bbox": [ + 89, + 409, + 483, + 515 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "In this work, we aim to catch up with the performance of class-separated anomaly detection models using a multiclass unified model. We introduce Dinomaly, a minimalist reconstruction-based UAD framework built exclusively by pure Transformer blocks [51], specifically Self-Attentions and Multi-Layer Perceptrons (MLPs). To begin with, we empirically investigate the scaling law of self-supervised pre-trained Vision Transformers (ViT) [12] when serving as the feature encoders for extracting reconstruction objectives. Subsequently, we introduce three straightforward yet crucial elements to address the critical identity mapping phenomenon in MUAD contexts, without increasing complexity or computational burden. First, as an alternative to meticulously designed pseudo anomaly and feature noise, we propose to activate the built-in Dropout in an MLP to prevent the network from restoring both normal and anomalous patterns. Second, we propose to leverage the \"side effect\" of Linear Attention (a computation-efficient counterpart of Softmax Attention) that impedes focus on local regions, thus preventing the forwarding of identical information. Third, previous methods adopt layer-to-layer and region-by-region reconstruction schemes, distilling a decoder that can well mimic the encoder's behavior even for anomalous regions. Therefore, we propose to loosen the reconstruction constraints by grouping multiple layers as a", + "bbox": [ + 89, + 523, + 483, + 901 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "whole and discarding well-reconstructed regions during optimization.", + "bbox": [ + 511, + 409, + 903, + 438 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "To validate the effectiveness of the proposed Dinomaly under MUAD setting, we conduct extensive experiments on a number of widely used benchmarks, including MVTec AD [3] (15 classes), VisA [70] (12 classes), and RealIAD (30 classes). As shown in Figure 1, our base-size Dinomaly achieves unprecedented image-level AUROC of $99.6\\%$ , $98.7\\%$ , and $89.3\\%$ on MVTec AD, VisA, and RealIAD, surpassing previous SoTA methods by a large margin. In addition, scalability is a key feature of Dinomaly. Further scaling up the model size maximizes performance to the fullest level of $99.8\\%$ , $98.9\\%$ , and $90.1\\%$ , respectively; while scaling down parameters and input size can offer efficient solutions in computation-constrained scenarios.", + "bbox": [ + 511, + 439, + 906, + 636 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "2. Related Work", + "text_level": 1, + "bbox": [ + 511, + 648, + 653, + 664 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Multi-Class UAD. UniAD [60] first introduced multi-class anomaly detection, aiming to detect anomalies for different classes using a unified model. In this setting, conventional UAD methods often face the challenge of \"identical shortcuts\", where both anomaly-free and anomaly samples can be effectively recovered during inference [60]. It is believed that this phenomenon is caused by the diversity of multi-class normal patterns that drive the network to generalize on unseen patterns. This contradicts the fundamental assumption of epistemic methods. Many current researches focus on addressing this challenge [14, 31, 36, 59, 60]. UniAD [60] employs a neighbor-masked attention module and a feature-jitter strategy to mitigate these shortcuts. HVQ-Trans [36] proposes a vector quantization (VQ) Transformer model that induces large feature discrepancies", + "bbox": [ + 511, + 672, + 906, + 900 + ], + "page_idx": 1 + }, + { + "type": "page_number", + "text": "20406", + "bbox": [ + 478, + 944, + 517, + 955 + ], + "page_idx": 1 + }, + { + "type": "image", + "img_path": "images/17d3d0fa6c8cb95057e7196ca865d09815a7ac4253dda8c3a5682950e09bd3d4.jpg", + "image_caption": [ + "Figure 2. The framework of Dinomaly, built by simple and pure Transformer building blocks." + ], + "image_footnote": [], + "bbox": [ + 94, + 92, + 908, + 287 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "for anomalies. LafitE [59] utilizes a latent diffusion model and introduces a feature editing strategy to alleviate this issue. DiAD [16] also employs diffusion models to address multi-class UAD settings. OmniAL [68] focuses on anomaly localization in the unified setting, preventing identical reconstruction by using synthesized pseudo anomalies. ReContrast [14] attempted to alleviate the identity mapping by cross-reconstruction between two encoders. ViTAD [5] abstracts a unified feature-reconstruction UAD framework and employ Transformer building blocks. MambaAD [17] explores the recently proposed State Space Model (SSM), Mamba, in the context of multi-class UAD. More related works of UAD are presented in Appendix A.", + "bbox": [ + 88, + 340, + 485, + 541 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3. Method", + "text_level": 1, + "bbox": [ + 89, + 550, + 183, + 566 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3.1. Dinomaly Framework", + "text_level": 1, + "bbox": [ + 89, + 575, + 299, + 590 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "\"What I cannot create, I do not understand\"", + "bbox": [ + 91, + 598, + 387, + 612 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Richer Feynman", + "bbox": [ + 109, + 614, + 256, + 628 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "The ability to recognize anomalies from what we know is an innate human capability, serving as a vital pathway for us to explore the world. Similarly, we construct a reconstruction-based framework that relies on the epistemic characteristic of artificial neural networks. Dinomaly consists of an encoder, a bottleneck, and a reconstruction decoder, as shown in Figure 2. Without loss of generality, a pretrained ViT network [12] with 12 Transformer layers is used as the encoder, extracting informative feature maps with different semantic scales. The bottleneck is a simple MLP (a.k.a. feed-forward network, FFN) that collects the feature representations of the encoder's 8 middle-level layers. The decoder is similar to the encoder, consisting of 8 Transformer layers. During training, the decoder learns to reconstruct the middle-level features of the encoder by maximizing the cosine similarity between feature maps. During inference, the decoder is expected to reconstruct normal regions of feature maps but fails for anomalous regions as it", + "bbox": [ + 89, + 628, + 485, + 901 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "has never seen such samples.", + "bbox": [ + 511, + 342, + 705, + 356 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Foundation Transformers. Foundation models, especially ViTs [12, 33] pre-trained on large-scale datasets, serve as a basis and starting point for specific computer vision tasks. Such networks employ self-supervised learning schemes such as contrastive learning (MoCov3 [6], DINO [4]), masked image modeling (MAE [19], SimMIM [57], BEiT [40]), and their combination (iBOT [69], DINOv2 [39]), producing universal features suitable for image-level visual tasks and pixel-level visual tasks.", + "bbox": [ + 511, + 357, + 906, + 492 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Because of the lack of supervision in UAD, most advanced methods adopt pre-trained networks to extract discriminative features. Recent works [28, 43, 65] have preliminarily discovered the advantage of robust and universal features of self-supervised models over domain-specific ImageNet features in anomaly detection tasks. In this paper, we pioneer the investigation of scaling behaviors in UAD models through a systematic analysis of foundational ViTs, as briefed in Figure 1(d). Our comprehensive evaluation encompasses pre-training strategies (Figure 5), model sizes (Table 4), and input resolutions (Table 5), which are detailed in section 4.4. Considering the balance of detection performance and computational efficiency, we adopt ViT-Base/14 pretrained by DINoV2-Register [7] as the encoder of Dinomaly by default.", + "bbox": [ + 511, + 493, + 908, + 719 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3.2. Noisy Bottleneck.", + "text_level": 1, + "bbox": [ + 511, + 726, + 684, + 742 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "\"Dropout is all you need.\"", + "bbox": [ + 513, + 750, + 692, + 763 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Prior studies [14, 60, 68] attribute the performance degradation of UAD methods trained on diverse multi-class samples to the \"identity mapping\" phenomenon; in this work, we reframe this issue as an \"over-generalization\" problem. Generalization ability is a merit of neural networks, allowing them to perform equally well on unseen test sets. However, generalization is not so wanted in the context of unsupervised anomaly detection that leverages the epistemic nature of neural networks. With the increas", + "bbox": [ + 511, + 765, + 908, + 902 + ], + "page_idx": 2 + }, + { + "type": "page_number", + "text": "20407", + "bbox": [ + 478, + 944, + 519, + 957 + ], + "page_idx": 2 + }, + { + "type": "image", + "img_path": "images/6d4a43b73593602e9e9d33bb1e07ed954c2eb370106d46e6efa33d5071ffbc4f.jpg", + "image_caption": [ + "Figure 3. Softmax Attention vs. Linear Attention. (a) Visualization of attention maps. (b) Attention distribution." + ], + "image_footnote": [], + "bbox": [ + 158, + 92, + 566, + 340 + ], + "page_idx": 3 + }, + { + "type": "image", + "img_path": "images/8d125b0bc5327cd9d9792b71ec97f6b7783286bcb1c5a99b13be4c2f135ca734.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 588, + 116, + 821, + 337 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "ing diversity of images and their patterns due to multi-class UAD settings, the decoder can generalize its reconstruction ability to unseen anomalous samples, resulting in the failure of anomaly detection using reconstruction error.", + "bbox": [ + 88, + 397, + 482, + 457 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "A direct solution for identity mapping is to shift \"reconstruction\" to \"restoration\". Specifically, instead of directly reconstructing the normal images or features given normal inputs, previous works propose to add perturbations as pseudo anomalies on input images [62, 67] or on encoder features [59, 60]; meanwhile, still let the decoder restore anomaly-free images or features, formulating a denoising-like framework. However, such methods employ heuristic and hand-crafted anomaly generation strategies that may not be universal across domains, datasets, and methods. In this work, we turn to leveraging the simple and elegant Dropout techniques. Since its introduction by Hinton et al. [21] in 2014 as a remedy for overfitting, Dropout has become a cornerstone element in neural architectures, including Transformers. In Dinomaly, we employ Dropout to randomly discard neural activations in an MLP bottleneck. Instead of alleviating overfitting, the role of Dropout in Dinomaly can be explained as pseudo feature anomaly that perturb normal representations, analogous to denoising auto-encoders [52, 53]. Without introducing any specific modules, this simple component inherently forces the decoder to restore normal features regardless of whether the test image contains anomalies, in turn, mitigating identical mapping.", + "bbox": [ + 91, + 458, + 483, + 821 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "3.3. Unfocused Linear Attention.", + "text_level": 1, + "bbox": [ + 89, + 830, + 344, + 847 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "\"One man's poison is another man's meat\"", + "bbox": [ + 91, + 854, + 382, + 869 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Softmax Attention is the key mechanism of Transformers, allowing the model to attend to different parts of its", + "bbox": [ + 89, + 869, + 483, + 901 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "input token sequence. Formally, given an input sequence $\\mathbf{X} \\in \\mathbb{R}^{N \\times d}$ with length $N$ , Attention first transforms it into three matrices: the query matrix $\\mathbf{Q} \\in \\mathbb{R}^{N \\times d}$ , the key matrix $\\mathbf{K} \\in \\mathbb{R}^{N \\times d}$ , and the value matrix $\\mathbf{V} \\in \\mathbb{R}^{N \\times d}$ :", + "bbox": [ + 511, + 397, + 906, + 458 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\mathbf {Q} = \\mathbf {X} \\mathbf {W} ^ {Q}, \\mathbf {K} = \\mathbf {X} \\mathbf {W} ^ {K}, \\mathbf {V} = \\mathbf {X} \\mathbf {W} ^ {V}, \\tag {1}\n$$\n", + "text_format": "latex", + "bbox": [ + 568, + 469, + 906, + 488 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "where $\\mathbf{W}^Q, \\mathbf{W}^K, \\mathbf{W}^V \\in \\mathbb{R}^{d \\times d}$ are learnable parameters. By computing the attention map by the query-key similarity, the output of Softmax Attention is given as:", + "bbox": [ + 511, + 500, + 905, + 546 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\operatorname {A t t e n t i o n} (\\mathbf {Q}, \\mathbf {K}, \\mathbf {V}) = \\operatorname {S o f t m a x} (\\mathbf {Q K} ^ {T}) \\mathbf {V}. \\tag {2}\n$$\n", + "text_format": "latex", + "bbox": [ + 563, + 559, + 906, + 578 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Back to MUAD, previous methods [36, 60] suggest adopting Attentions instead of Convolutions because Convolutions can easily learn identical mappings. Nevertheless, both operations are in danger of forming identity mapping by over-concentrating on corresponding input locations for producing the outputs:", + "bbox": [ + 511, + 590, + 906, + 681 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\operatorname {C o n v} \\operatorname {K e r n e l} = \\left[ \\begin{array}{l l l} 0 & 0 & 0 \\\\ 0 & 1 & 0 \\\\ 0 & 0 & 0 \\end{array} \\right], \\quad \\operatorname {A t t n} \\operatorname {M a p} = \\left[ \\begin{array}{l l l} 1 & 0 & 0 \\\\ 0 & 1 & 0 \\\\ 0 & 0 & 1 \\end{array} \\right].\n$$\n", + "text_format": "latex", + "bbox": [ + 511, + 693, + 903, + 743 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Is there any simple solution that prevents Attentions from attending to identical information? In Dinomaly, we turn to leverage the \"unfocusing ability\" of a type of Softmax-free Attention, i.e., Linear Attention. Linear Attention was proposed as a promising alternative to reduce the computation complexity of vanilla Softmax Attention", + "bbox": [ + 511, + 753, + 906, + 845 + ], + "page_idx": 3 + }, + { + "type": "page_footnote", + "text": "1The full form of Attention is Softmax $\\left(\\frac{\\mathbf{Q}\\mathbf{K}^T}{\\sqrt{d}}\\right)\\mathbf{V}$ . The constant denominator is omitted for narrative simplicity. The multi-head mechanism that concatenates multiple Attentions is also omitted.", + "bbox": [ + 511, + 857, + 906, + 900 + ], + "page_idx": 3 + }, + { + "type": "footer", + "text": "20408", + "bbox": [ + 478, + 944, + 519, + 955 + ], + "page_idx": 3 + }, + { + "type": "image", + "img_path": "images/a1dd9fa401691a462a4df3a84a2f8feedcaf54499e447957f60fb2c121c62204.jpg", + "image_caption": [ + "(a) layer-to-layer sparse" + ], + "image_footnote": [], + "bbox": [ + 106, + 89, + 233, + 185 + ], + "page_idx": 4 + }, + { + "type": "image", + "img_path": "images/b866424b103ff548ef2ecbe121963fe58d37f24999027884bb87e659c5bc10ba.jpg", + "image_caption": [ + "(b) layer-to-cat-layer" + ], + "image_footnote": [], + "bbox": [ + 251, + 88, + 356, + 183 + ], + "page_idx": 4 + }, + { + "type": "image", + "img_path": "images/b4bffa389b2c5780fbe925887470053347f3965c3cfab98e72337e6b9155bdfe.jpg", + "image_caption": [ + "(c) layer-to-layer dense" + ], + "image_footnote": [], + "bbox": [ + 395, + 88, + 500, + 183 + ], + "page_idx": 4 + }, + { + "type": "image", + "img_path": "images/2e691d240b522deeace347316f62ed4982cc6efcf42bc8289ce3948972a78386.jpg", + "image_caption": [ + "(d) group-to-group 1 group (Ours)" + ], + "image_footnote": [], + "bbox": [ + 534, + 87, + 635, + 183 + ], + "page_idx": 4 + }, + { + "type": "image", + "img_path": "images/0302c001c32f5d638e646b4c5b0696ae395e2e4637f8de43dad4b5752f96a3ad.jpg", + "image_caption": [ + "(e) group-to-group 2 group (Ours)" + ], + "image_footnote": [], + "bbox": [ + 671, + 88, + 774, + 181 + ], + "page_idx": 4 + }, + { + "type": "image", + "img_path": "images/d3bc1adf7c570b9343648410728143a6c2329294536bf9af7d61d3b9329c354d.jpg", + "image_caption": [ + "encoder feature" + ], + "image_footnote": [], + "bbox": [ + 815, + 95, + 836, + 114 + ], + "page_idx": 4 + }, + { + "type": "image", + "img_path": "images/4767d4991e7ce4e01a71213986c2b2383e2a08849f258d4d882d31d17534b15e.jpg", + "image_caption": [ + "decoder feature" + ], + "image_footnote": [], + "bbox": [ + 816, + 125, + 836, + 138 + ], + "page_idx": 4 + }, + { + "type": "image", + "img_path": "images/f33596ba1387f7c0a6e5d9120039352aced6c9d016bf983caf8f1150221ea210.jpg", + "image_caption": [ + "reconstruct" + ], + "image_footnote": [], + "bbox": [ + 816, + 147, + 836, + 155 + ], + "page_idx": 4 + }, + { + "type": "image", + "img_path": "images/c37867f60176995bbf10bf74c757ea2ab30dd85d912f29e92525a9a0297ad90e.jpg", + "image_caption": [ + "concatenate" + ], + "image_footnote": [], + "bbox": [ + 816, + 162, + 836, + 172 + ], + "page_idx": 4 + }, + { + "type": "image", + "img_path": "images/ac76c81cae97fc845173fcf13a54643878bec7b040506ee9b893173e98d3134b.jpg", + "image_caption": [ + "add", + "Figure 4. Schemes of reconstruction constraint. (a) Layer-to-layer (sparse). (b) Layer-to-cat-layer. (c) Layer-to-layer (dense). (d) Loose group-to-group, 1-group (Ours). (e) Loose group-to-group, 2-group (Ours)." + ], + "image_footnote": [], + "bbox": [ + 816, + 178, + 836, + 186 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "concerning the number of tokens [26]. By substituting Softmax operation with a simple activation function $\\phi (\\cdot)$ (usually $\\phi (x) = \\mathrm{elu}(x) + 1$ ), we can change the computation order from $(\\mathbf{Q}\\mathbf{K}^T)\\mathbf{V}$ to $\\mathbf{Q}(\\mathbf{K}^T\\mathbf{V})$ . Formally, Linear Attention (LA) is given as:", + "bbox": [ + 88, + 273, + 483, + 349 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\operatorname {L A} (\\mathbf {Q}, \\mathbf {K}, \\mathbf {V}) = (\\phi (\\mathbf {Q}) \\phi \\left(\\mathbf {K} ^ {T}\\right)) \\mathbf {V} = \\phi (\\mathbf {Q}) \\left(\\phi \\left(\\mathbf {K} ^ {T}\\right) \\mathbf {V}\\right), \\tag {3}\n$$\n", + "text_format": "latex", + "bbox": [ + 99, + 354, + 483, + 386 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "where the computation complexity is reduced to $\\mathcal{O}(Nd^2)$ from $\\mathcal{O}(N^2 d)$ . The trade-off between complexity and expressiveness is a dilemma. Previous studies [15, 48] attribute Linear Attention's performance degradation on supervised tasks to its incompetence in focusing. Due to the absence of non-linear attention reweighting by Softmax operation, Linear Attention cannot concentrate on important regions related to the query, such as foreground and neighbors. This property, however, is exactly what the reconstruction decoder favors in our contexts.", + "bbox": [ + 88, + 386, + 483, + 537 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "In order to probe how Attentions propagate information, we train two variants of Dinomaly using vanilla Softmax Attention or Linear Attention as the spatial mixer in the decoder and visualize their attention maps. As shown in Figure 3, Softmax Attention tends to focus on the exact region of the query, while Linear Attention spreads its attention across the whole image. This implies that Linear Attention, forced by its incompetence to focus, utilizes more long-range information to restore features at each position, reducing the chance of passing identical information of unseen patterns to the next layer during reconstruction. Of course, employing Linear Attention also benefits from less computation, free of performance drop.", + "bbox": [ + 88, + 537, + 483, + 733 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "3.4. Loose Reconstruction", + "text_level": 1, + "bbox": [ + 89, + 742, + 294, + 756 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "\"The tighter you squeeze, the less you have.\"", + "bbox": [ + 91, + 763, + 388, + 779 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Loose Constraint. Pioneers of feature-reconstruction/distillation UAD methods [10, 46] are inspired by knowledge distillation [20]. Most reconstruction-based methods distill specific encoder layers (e.g. 3 last layers of 3 ResNet stages) by the corresponding decoder layers [10, 46, 65] (Figure 4(a)) or the last decoder layer [58, 60] (Figure 4(b)). Intuitively, with more encoder-decoder feature pairs (Figure 4(c)), UAD model can utilize", + "bbox": [ + 88, + 779, + 483, + 900 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "more information in different layers to discriminate anomalies. However, according to the intuition of knowledge distillation, the student (decoder) can better mimic the behavior of the teacher (encoder) given more layer-to-layer supervision [30], which is harmful for UAD models that detect anomalies by encoder-decoder discrepancy. This phenomenon is also embodied as identity mapping. Thanks to the top-to-bottom consistency of columnar Transformer layers, we propose to loosen the layer-to-layer constraint by adding up all feature maps of interested layers as a whole group, as shown in Figure 4(d). This scheme can be seen as loosening the layer-to-layer correspondence and providing the decoder with more degrees of freedom, so that the decoder is allowed to act much more differently from the encoder when the input pattern is unseen. Because features of shallow layers contain low-level visual characters that are helpful for precise localization, we can further group the features into the low-semantic-level group and high-semantic-level group, as shown in Figure 4(e).", + "bbox": [ + 511, + 273, + 906, + 559 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "**Loose Loss.** Following the above analysis, we also loosen the point-by-point reconstruction loss function by discarding some points in the feature map. Here, we simply borrow the hard-mining global cosine loss [14] that detaches the gradients of well-restored feature points with low cosine distance during training. Let $f_{E}$ and $f_{D}$ denotes (grouped) feature maps of encoder and decoder:", + "bbox": [ + 511, + 560, + 908, + 666 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {L} _ {\\text {g l o b a l} - h m} = \\mathcal {D} _ {\\text {c o s}} \\left(\\mathcal {F} \\left(f _ {E}\\right), \\mathcal {F} \\left(\\hat {f} _ {D}\\right)\\right), \\tag {4}\n$$\n", + "text_format": "latex", + "bbox": [ + 584, + 674, + 903, + 691 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\hat {f} _ {D} (h, w) = \\left\\{ \\begin{array}{l} s g \\left(f _ {D} (h, w)\\right)_{0.1}, \\text {if} \\mathcal {D} _ {\\text {cos}} \\left(f _ {D}, f _ {E}\\right) < k \\% _ {\\text {batch}} \\\\ f _ {D} (h, w), \\text {else} \\end{array} \\right. \\tag{5}\n$$\n", + "text_format": "latex", + "bbox": [ + 514, + 699, + 903, + 739 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {D} _ {\\cos} (a, b) = 1 - \\frac {a ^ {T} \\cdot b}{\\| a \\| \\| b \\|}, \\tag {6}\n$$\n", + "text_format": "latex", + "bbox": [ + 622, + 739, + 903, + 770 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "where $\\mathcal{D}_{cos}$ denotes cosine distance, $\\mathcal{F}(\\cdot)$ denotes flatten operation, $f_{D}(h,w)$ represents the feature point at $(h,w)$ , and $sg(\\cdot)_{0.1}$ denotes shrink the gradient to one-tenth of the original $^2$ . $\\mathcal{D}_{cos}(f_D(h,w),f_E(h,w)) < k\\%_{batch}$ selects $k \\%$ feature points with smaller cosine distance within a batch for gradient shrinking. Total loss is the average $\\mathcal{L}_{global - hm}$ of all encoder-decoder feature pairs.", + "bbox": [ + 511, + 772, + 906, + 878 + ], + "page_idx": 4 + }, + { + "type": "page_footnote", + "text": "${}^{2}$ Complete stop-gradient causes optimization instability occasionally.", + "bbox": [ + 529, + 886, + 897, + 900 + ], + "page_idx": 4 + }, + { + "type": "footer", + "text": "20409", + "bbox": [ + 478, + 944, + 519, + 955 + ], + "page_idx": 4 + }, + { + "type": "table", + "img_path": "images/63abb526d89a5fc1376c69e0a3620b9eeaa44da1e212790f58926cb7f3502cf6.jpg", + "table_caption": [ + "Table 1. Performance under multi-class UAD setting (%). †: method designed for MUAD. Dinomaly†: training schedule is scaled up to 20,000, 20,000, and 100,000 iterations (original: 10,000/10,000/50,000)." + ], + "table_footnote": [], + "table_body": "
DatasetMethodImage-levelPixel-level
AUROCAP\\(F_1\\)-maxAUROCAP\\(F_1\\)-maxAUPRO
MVTec-AD [3]RD4AD [10]94.696.595.296.148.653.891.1
SimpleNet [34]95.398.495.896.945.949.786.5
DeSTSeg [67]89.295.591.693.154.350.964.8
UniAD [60]†96.598.896.296.843.449.590.7
ReContrast [14]†98.399.497.697.160.261.593.2
DiAD [18]†97.299.096.596.852.655.590.7
ViTAD [65]†98.399.497.397.755.358.791.4
MambaAD [17]†98.699.697.897.756.359.293.1
Dinomaly (Ours)99.699.899.098.469.369.294.8
Dinomaly†99.799.899.298.469.369.294.7
VisA [70]RD4AD [10]92.492.489.698.138.042.691.8
SimpleNet [34]87.287.081.896.834.737.881.4
DeSTSeg [67]88.989.085.296.139.643.467.4
UniAD [60]†88.890.885.898.333.739.085.5
ReContrast [14]†95.596.492.098.547.950.691.9
DiAD [18]†86.888.385.196.026.133.075.2
ViTAD [65]†90.591.786.398.236.641.185.1
MambaAD [17]†94.394.589.498.539.444.091.0
Dinomaly (Ours)98.798.996.298.753.255.794.5
Dinomaly†98.999.096.498.853.855.894.5
Real-IAD [54]RD4AD [10]82.479.073.997.325.032.789.6
SimpleNet [34]57.253.461.575.72.86.539.0
DeSTSeg [67]82.379.273.294.637.941.740.6
UniAD [60]†83.080.974.397.321.129.286.7
ReContrast [14]†86.484.277.497.831.638.291.8
DiAD [18]†75.666.469.988.02.97.158.1
ViTAD [65]†82.780.273.797.224.332.384.8
MambaAD [17]†86.384.677.098.533.038.790.5
Dinomaly (Ours)89.386.880.298.842.847.193.9
Dinomaly†89.586.980.498.943.347.494.2
", + "bbox": [ + 133, + 128, + 867, + 539 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "4. Experiments", + "text_level": 1, + "bbox": [ + 89, + 563, + 223, + 580 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "4.1. Experimental Settings", + "text_level": 1, + "bbox": [ + 89, + 588, + 297, + 606 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Datasets. MVtec-AD [3] contains 15 objects (5 texture classes and 10 object classes) with a total of 3,629 normal images as the training set and 1,725 images as the test set (467 normal, 1,258 anomalous). VisA [70] contains 12 objects. Training and test sets are split following the official splitting, resulting in 8,659 normal images in the training set and 2,162 images in the test set (962 normal, 1,200 anomalous). Real-IAD [54] is a large UAD dataset recently released, containing 30 distinct objects. We follow the official splitting that includes all views, resulting in 36,465 normal images in the training set and 114,585 images in the test set (63,256 normal, 51,329 anomalous).", + "bbox": [ + 89, + 612, + 482, + 792 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Metrics. Following prior works [17, 65], we adopt 7 evaluation metrics. Image-level anomaly detection performance is measured by the Area Under the Receiver Operator Curve (AUROC), Average Precision (AP), and $F_{1}$ score under optimal threshold $(F_{1}$ -max). Pixel-level anomaly localization is measured by AUROC, AP, $F_{1}$ -max and the Area Under the Per-Region-Overlap (AUPRO). The results", + "bbox": [ + 89, + 795, + 482, + 901 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "of a dataset is the average of all classes.", + "bbox": [ + 511, + 565, + 777, + 579 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Implementation Details. ViT-Base/14 (patchsize=14) pre-trained by DINOv2-R [7] is used as the encoder by default. The drop rate of Noisy Bottleneck is 0.2 by default and increases to 0.4 on the diverse Real-IAD. Loose constraint with 2 groups is employed, and the anomaly map is given by the mean per-point cosine distance of the 2 groups. The input image is first resized to $448^2$ and then center-cropped to $392^2$ , so the feature map $(28^2)$ is large enough for anomaly localization. StableAdamW optimizer [56] with AMSGrad [41] (more stable than AdamW [35] in training) is utilized with $lr = 2e - 3$ , $\\beta = (0.9,0.999)$ and $wd = 1e - 4$ . The network is trained for 10,000 iterations (steps) on MVTec-AD and VisA, and 50,000 iterations on Real-IAD. Detailed settings are available in Appendix B.", + "bbox": [ + 511, + 580, + 906, + 792 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "4.2. Comparison to Multi-Class UAD SoTAs", + "text_level": 1, + "bbox": [ + 511, + 801, + 854, + 816 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "We compare the proposed Dinomaly with the most advanced UAD and MUAD methods [10, 14, 17, 18, 34, 65, 67]. Experimental results are presented in Table 1, where Dinomaly surpasses compared methods by a large margin on all datasets and all metrics. On the most", + "bbox": [ + 511, + 825, + 905, + 900 + ], + "page_idx": 5 + }, + { + "type": "page_number", + "text": "20410", + "bbox": [ + 478, + 944, + 519, + 957 + ], + "page_idx": 5 + }, + { + "type": "table", + "img_path": "images/cd538305d399bb60a141faeaeca01f08a86dc1a2c840fd10c9d92ee535e6db57.jpg", + "table_caption": [ + "Table 2. Performance under conventional class-separated UAD setting (%). n/a: not available." + ], + "table_footnote": [], + "table_body": "
MethodMVTec-AD [3]VisA [70]Real-IAD [54]
I-AUROCP-AUROCP-AUPROI-AUROCP-AUROCP-AUPROI-AUROCP-AUROCP-AUPRO
Dinomaly (MUAD)99.698.494.898.798.794.589.398.893.9
Dinomaly99.799.995.098.998.995.192.099.195.1
RD4AD [10]98.597.893.996.090.170.987.1n/a93.8
PatchCore [45]99.198.193.594.798.591.889.4n/a91.5
SimpleNet [34]99.698.190.097.198.290.788.5n/a84.6
", + "bbox": [ + 133, + 114, + 864, + 219 + ], + "page_idx": 6 + }, + { + "type": "table", + "img_path": "images/a326e4c362788dd358ee23012ebf6b13e0d43aae440ef876896212edf481cee2.jpg", + "table_caption": [ + "Table 3. Ablations of Dinomaly elements on MVTec-AD (%). NB: Noisy Bottleneck. LA: Linear Attention. LC: Loose Constraint (2 groups). LL: Loose Loss. As MVTec-AD has reached saturation, we also present the results on VisA (Table A4)." + ], + "table_footnote": [], + "table_body": "
NBLALCLLImage-levelPixel-level
AUROCAPF1-maxAUROCAPF1-maxAUPRO
98.4199.0997.4197.1862.9663.8292.95
✓99.0699.5498.3197.6266.2266.7093.71
✓98.5499.2197.6297.2062.9463.7393.09
✓98.3599.0497.4397.1061.0562.7392.60
✓99.0399.4598.1997.6264.1064.9693.34
✓✓99.2799.6298.6397.8567.3667.3394.16
✓✓99.5099.7298.8798.1468.1668.2494.23
✓✓✓99.5299.7398.9298.2068.2568.3494.17
✓✓✓99.5799.7899.0098.2067.9368.2194.50
✓✓✓✓99.6099.7899.0498.3569.2969.1794.79
", + "bbox": [ + 133, + 268, + 864, + 457 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "widely used MVTec-AD, Dinomaly produces image-level performance of 99.6/99.8/99.0 (\\%) and pixel-level performance of 98.4/69.3/69.2/94.8, outperforming previous SoTAs by 1.0/0.2/1.2 and 0.7/9.1/7.7/1.6. This result declares that the image-level performance on the MVTec-AD dataset is nearly saturated under the MUAD setting. On the popular VisA, Dinomaly achieves image-level performance of 98.7/98.9/96.2 and pixel-level performance of 98.7/53.2/55.7/94.5, outperforming previous SoTAs by 3.2/2.5/4.2 and 0.2/5.3/5.1/2.6. On the Real-IAD that contains 30 classes, each with 5 camera views, we produce image-level and pixel-level performance of 89.3/86.8/80.2 and 98.8/42.8/47.1/93.9, outperforming previous SoTAs by 3.0/2.2/3.2 and 0.3/4.9/5.4/3.4, indicating our scalability to extremely complex scenarios. Per-class performances and qualitative visualization are presented in Appendix E and F. We also produce superior results on other popular UAD benchmarks, i.e., MPDD [24], BTAD [38], and UniMedical [66], with I-AUROC of 97.2, 95.4, and 84.9, respectively, as shown in Table A13 in Appendix.", + "bbox": [ + 88, + 481, + 485, + 785 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "4.3. Comparison to Class-Separated UAD SoTAs", + "text_level": 1, + "bbox": [ + 89, + 799, + 467, + 816 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Dinomaly is also compared with class-separated SoTAs, as shown in Table 2. Dinomaly under MUAD setting is comparable to conventional methods [10, 34, 45] that build individual models for each class. On MVTec-AD and VisA, multi-class Dinomaly (first row) is subjected to nearly no", + "bbox": [ + 89, + 824, + 483, + 902 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "performance drop compared to its class-separated counterpart (second row). On the complicated Real-IAD that involves more classes and views, multi-class Dinomaly suffers a moderate performance drop but is still comparable to class-separated SoTAs.", + "bbox": [ + 511, + 481, + 906, + 558 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "4.4. Ablation Study", + "text_level": 1, + "bbox": [ + 511, + 571, + 666, + 588 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Overall Ablation. We conduct experiments to verify the effectiveness of the proposed elements, i.e., Noisy Bottleneck (NB), Linear Attention (LA), Loose Constraint (LC), and Loose Loss (LL). The already-powerful baseline (first row) is Dinomaly with noiseless MLP bottleneck, Softmax Attention, dense layer-to-layer supervision, and global cosine loss. This baseline is very similar to ViTAD [64] and the ViT version of RD4AD [10]. Results on MVtec-AD and VisA are shown in Table 3 and Table A4, respectively. NB and LL can directly contribute to the model performance. LA and LC boost the performance with the presence of NB. The use of LC is not solely beneficial because LC makes the reconstruction too easy without injected noise.", + "bbox": [ + 511, + 595, + 908, + 792 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Model Scalability. Previous works [10, 60, 65] reported that anomaly detection networks do not follow the \"scaling law\". For example, RD4AD [10] found WideResNet50 better than WideResNet101 as the encoder backbone. ViTAD [65] found ViT-Small better than ViT-Base. On the contrary, as shown in Table 4, the performance of the proposed Dinomaly benefits from scaling. Dinomaly equipped with ViT", + "bbox": [ + 511, + 795, + 908, + 902 + ], + "page_idx": 6 + }, + { + "type": "page_number", + "text": "20411", + "bbox": [ + 478, + 944, + 517, + 957 + ], + "page_idx": 6 + }, + { + "type": "table", + "img_path": "images/be8fcebda125d516609706bb55b352c692dfdd1fd3e340fc1f33ba3d0c6c7ba7.jpg", + "table_caption": [ + "Table 4. Scaling of ViT model sizes on MVtec-AD (%). Im/s (Throughoutput, image per second) is measured on NVIDIA RTX3090 with batch size=16. Results on VisA and Real-IAD are shown in Table A3. †:default." + ], + "table_footnote": [], + "table_body": "
Arch.ParamsMACsIm/sImage-levelPixel-level
AUROCAPF1-maxAUROCAPF1-maxAUPRO
ViT-Small37.4M26.3G153.699.2699.6798.7298.0768.2967.7894.36
ViT-Base†148.0M104.7G58.199.6099.7899.0498.3569.2969.1794.79
ViT-Large275.3M413.5G24.299.7799.9299.4598.5470.5370.0495.09
", + "bbox": [ + 133, + 128, + 864, + 218 + ], + "page_idx": 7 + }, + { + "type": "table", + "img_path": "images/5f71ebbf272cad0fdc21a3279e913af6c14f1d181f64cbec2c34a88ad27b3447.jpg", + "table_caption": [ + "Table 5. Scaling input size on MVTec-AD (\\%). †: default. Compared methods yield degradation when increasing input size." + ], + "table_footnote": [], + "table_body": "
MethodInput SizeImage-LevelPixel-Level
RD4AD2562†94.6/96.5/96.196.1/48.6/53.8/91.1
320293.2/96.9/95.695.7/55.1/57.5/91.1
384291.9/96.2/95.094.9/52.1/55.3/90.8
ReContrast2562†98.3/99.4/97.697.1/60.2/61.5/93.2
320298.2/99.2/97.596.8/61.8/62.6/93.3
384295.2/98.0/96.496.5/57.7/59.5/92.6
Dinomaly280299.6/99.8/99.398.2/65.2/66.3/93.6
336299.6/99.8/99.298.3/67.2/67.8/94.2
3922†99.6/99.8/99.098.4/69.3/69.2/94.8
", + "bbox": [ + 102, + 277, + 470, + 436 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Small has already produced state-of-the-art results. ViT-Large further boosts Dinomaly to an unprecedented higher record. This scalability enables users to choose an appropriate model size based on the computational resources available in their specific scenario. A comparison of computational costs with other methods is presented in Table A11. In addition, training schedule can also be scaled up for even better performance without increasing inference costs, as demonstrated in Figure 1 (Dinomaly $\\uparrow$ ).", + "bbox": [ + 88, + 460, + 482, + 598 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Input Scalability. Though it seems unfair to compare Dinomaly with previous works that take smaller images as input, we contend that increasing their input size not only fails to benefit but actively undermines their performance, especially for image-level detection performance, as shown in Table 5. Therefore, we follow the common comparison strategy based on \"optimum vs. optimum\". On the contrary, Dinomaly enjoys scaling input size for anomaly localization, while still producing SoTA performance given smaller images. Details are presented in Table A2 in Appendix.", + "bbox": [ + 88, + 598, + 482, + 750 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "ViT Foundations. We conduct extensive experiments to investigate the impact of diverse pre-trained ViT foundations, including DeiT [50], MAE [19], D-iGPT [44], MOCOV3 [6], DINO [4], iBot [69], DINOv2 [39], and DINOv2-R [7]. As shown in Figure 5, Dinomaly is robust to the choice of backbone. Almost all foundation models can produce SoTA-level results with image-level AUROC higher than $98\\%$ . The only notable exception is MAE, which, without fine-tuning, was reported to be less effective across various unsupervised tasks, e.g., kNN and linear-", + "bbox": [ + 88, + 750, + 482, + 901 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/7534438a9f00e66162f29b3e33941023cda6c2ee6fa9e8f5a27a44c64f62b210.jpg", + "image_caption": [ + "Figure 5. Image-level AUROC of Dinomaly equipped with various ViT foundations, and their linear-probing accuracy on ImageNet. MIM: Masked Image Modeling. CL: Contrastive Learning." + ], + "image_footnote": [], + "bbox": [ + 519, + 239, + 897, + 385 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "probing [39]. The optimal input size varies because the these backbones are pre-trained on different resolutions. Interestingly, we found the anomaly detection performance to be strongly correlated with the accuracy of ImageNet linear-probing (freeze backbone & only tune linear classifier) of the foundation model, suggesting the possibility of further improvement by simply adopting a future foundation model. Detailed results and analysis are presented in Appendix and Table A1.", + "bbox": [ + 511, + 470, + 906, + 607 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Additional experiments and results are detailed in the Appendix C, encompassing evaluations of various pretrained foundations, ablation studies of each components, hyperparameter optimization, and other in-depth analyses.", + "bbox": [ + 511, + 609, + 905, + 672 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "5. Conclusion", + "text_level": 1, + "bbox": [ + 511, + 691, + 633, + 707 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Dinomaly, a minimalistic UAD framework, is proposed to address the under-performed MUAD models in this paper. We present four key elements in Dinomaly, i.e., Foundation Transformer, Noisy MLP Bottleneck, Linear Attention, and Loose Reconstruction, that can boost the performance under the challenging MUAD setting without fancy modules and tricks. Extensive experiments on MVTec AD, VisA, and Real-IAD demonstrate our superiority over previous model-unified multi-class models and even recent class-separated models, indicating the feasibility of implementing a unified model in complicated scenarios free of severe performance degradation.", + "bbox": [ + 511, + 719, + 906, + 900 + ], + "page_idx": 7 + }, + { + "type": "page_number", + "text": "20412", + "bbox": [ + 478, + 944, + 519, + 957 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Acknowledgments", + "text_level": 1, + "bbox": [ + 91, + 90, + 250, + 107 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "References", + "text_level": 1, + "bbox": [ + 91, + 217, + 187, + 233 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[1] Samet Akcay, Amir Atapour-Abarghouei, and Toby P Breckon. Ganomaly: Semi-supervised anomaly detection via adversarial training. In Computer Vision-ACCV 2018: 14th Asian Conference on Computer Vision, Perth, Australia, December 2-6, 2018, Revised Selected Papers, Part III 14, pages 622-637. Springer, 2019. 1", + "[2] Paul Bergmann, Sindy Löwe, Michael Fauser, David Sattlegger, and Carsten Steger. Improving unsupervised defect segmentation by applying structural similarity to autoencoders. arXiv preprint arXiv:1807.02011, 2018. 1", + "[3] Paul Bergmann, Michael Fauser, David Sattlegger, and Carsten Steger. Mvtec ad-a comprehensive real-world dataset for unsupervised anomaly detection. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 9592-9600, 2019. 1, 2, 6, 7", + "[4] Mathilde Caron, Hugo Touvron, Ishan Misra, Hervé Jégou, Julien Mairal, Piotr Bojanowski, and Armand Joulin. Emerging properties in self-supervised vision transformers. In Proceedings of the IEEE/CVF international conference on computer vision, pages 9650-9660, 2021. 3, 8, 2", + "[5] Xinlei Chen and Kaiming He. Exploring simple siamese representation learning. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 15750-15758, 2021. 3", + "[6] Xinlei Chen, Saining Xie, and Kaiming He. An empirical study of training self-supervised vision transformers. In Proceedings of the IEEE/CVF international conference on computer vision, pages 9640-9649, 2021. 3, 8, 2", + "[7] Timothee Darcet, Maxime Oquab, Julien Mairal, and Piotr Bojanowski. Vision transformers need registers. arXiv preprint arXiv:2309.16588, 2023. 3, 6, 8, 1, 2", + "[8] Thomas Defard, Aleksandr Setkov, Angelique Loesch, and Romaric Audigier. Padim: a patch distribution modeling framework for anomaly detection and localization. In International Conference on Pattern Recognition, pages 475-489. Springer, 2021. 1", + "[9] David Dehaene and Pierre Eline. Anomaly localization by modeling perceptual features. arXiv preprint arXiv:2008.05369, 2020. 1", + "[10] Hanqiu Deng and Xingyu Li. Anomaly detection via reverse distillation from one-class embedding. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 9737-9746, 2022. 5, 6, 7, 1, 8, 9, 10", + "[11] Jia Deng, Wei Dong, Richard Socher, Li-Jia Li, Kai Li, and Li Fei-Fei. Imagenet: A large-scale hierarchical image database. In IEEE Conference on Computer Vision and Pattern Recognition, pages 248–255, 2009. 2" + ], + "bbox": [ + 93, + 243, + 483, + 900 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "The authors acknowledge supports from National Natural Science Foundation of China (U22A2051, 82027807), National Key Research and Development Program of China (2022YFC2405200), Tsinghua-Foshan Innovation Special Fund (2021THFS0104), and Institute for Intelligent Healthcare, Tsinghua University (2022ZLB001).", + "[12] Alexey Dosovitskiy, Lucas Beyer, Alexander Kolesnikov, Dirk Weissenborn, Xiaohua Zhai, Thomas Unterthiner, Mostafa Dehghani, Matthias Minderer, Georg Heigold, Sylvain Gelly, et al. An image is worth 16x16 words: Transformers for image recognition at scale. arXiv preprint arXiv:2010.11929, 2020. 2, 3", + "[13] Jia Guo, Shuai Lu, Lize Jia, Weihang Zhang, and Huiqi Li. Encoder-decoder contrast for unsupervised anomaly detection in medical images. IEEE Transactions on Medical Imaging, 2023. 1", + "[14] Jia Guo, Shuai Lu, Lize Jia, Weihang Zhang, and Huiqi Li. Recontrast: Domain-specific anomaly detection via contrastive reconstruction. In Advances in Neural Information Processing Systems (NeurIPS), pages 10721-10740, 2023. 2, 3, 5, 6", + "[15] Dongchen Han, Xuran Pan, Yizeng Han, Shiji Song, and Gao Huang. Flatten transformer: Vision transformer using focused linear attention. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 5961-5971, 2023. 5", + "[16] Haoyang He, Jiangning Zhang, Hongxu Chen, Xuhai Chen, Zhishan Li, Xu Chen, Yabiao Wang, Chengjie Wang, and Lei Xie. Diad: A diffusion-based framework for multi-class anomaly detection. arXiv preprint arXiv:2312.06607, 2023. 1, 3", + "[17] Haoyang He, Yuhu Bai, Jiangning Zhang, Qingdong He, Hongxu Chen, Zhenye Gan, Chengjie Wang, Xiangtai Li, Guanzhong Tian, and Lei Xie. Mambaad: Exploring state space models for multi-class unsupervised anomaly detection. arXiv preprint arXiv:2404.06564, 2024. 1, 3, 6, 7, 8, 10", + "[18] Haoyang He, Jiangning Zhang, Hongxu Chen, Xuhai Chen, Zhishan Li, Xu Chen, Yabiao Wang, Chengjie Wang, and Lei Xie. A diffusion-based framework for multi-class anomaly detection. In Proceedings of the AAAI Conference on Artificial Intelligence, pages 8472-8480, 2024. 2, 6, 7, 8, 9, 10", + "[19] Kaiming He, Xinlei Chen, Saining Xie, Yanghao Li, Piotr Dollar, and Ross Girshick. Masked autoencoders are scalable vision learners. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 16000-16009, 2022. 3, 8, 2, 5", + "[20] Geoffrey Hinton, Oriol Vinyals, and Jeff Dean. Distilling the knowledge in a neural network. arXiv preprint arXiv:1503.02531, 2015. 5", + "[21] Geoffrey E Hinton, Nitish Srivastava, Alex Krizhevsky, Ilya Sutskever, and Ruslan R Salakhutdinov. Improving neural networks by preventing co-adaptation of feature detectors. arXiv preprint arXiv:1207.0580, 2012. 4", + "[22] Chaoqin Huang, Haoyan Guan, Aofan Jiang, Ya Zhang, Michael Spratling, and Yan-Feng Wang. Registration based few-shot anomaly detection. In European Conference on Computer Vision, pages 303-319. Springer, 2022. 7", + "[23] Jongheon Jeong, Yang Zou, Taewan Kim, Dongqing Zhang, Avinash Ravichandran, and Onkar Dabeer. Winclip: Zero/few-shot anomaly classification and segmentation. arXiv preprint arXiv:2303.14814, 2023.7", + "[24] Stepan Jezek, Martin Jonak, Radim Burget, Pavel Dvorak, and Milos Skotak. Deep learning-based defect detection of" + ], + "bbox": [ + 93, + 92, + 905, + 900 + ], + "page_idx": 8 + }, + { + "type": "page_number", + "text": "20413", + "bbox": [ + 478, + 944, + 519, + 955 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "metal parts: evaluating current methods in complex conditions. In 2021 13th International Congress on Ultra Modern Telecommunications and Control Systems and Workshops (ICUMT), pages 66-71. IEEE, 2021. 7, 6", + "[25] Xi Jiang, Jianlin Liu, Jinbao Wang, Qiang Nie, Kai Wu, Yong Liu, Chengjie Wang, and Feng Zheng. Softpatch: Unsupervised anomaly detection with noisy data. Advances in Neural Information Processing Systems, 35:15433-15445, 2022. 6, 7", + "[26] Angelos Katharopoulos, Apoorv Vyas, Nikolaos Pappas, and François Fleuret. Transformers are rnns: Fast autoregressive transformers with linear attention. In International conference on machine learning, pages 5156-5165. PMLR, 2020. 5", + "[27] Sungwook Lee, Seunghyun Lee, and Byung Cheol Song. Cfa: Coupled-hypersphere-based feature adaptation for target-oriented anomaly localization. IEEE Access, 10: 78446-78454, 2022. 1", + "[28] Yujin Lee, Harin Lim, and Hyunsoo Yoon. Selfformaly: Towards task-agnostic unified anomaly detection. arXiv preprint arXiv:2307.12540, 2023. 3", + "[29] Chun-Liang Li, Kihyuk Sohn, Jinsung Yoon, and Tomas Pfister. Cutpaste: Self-supervised learning for anomaly detection and localization. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 9664-9674, 2021. 1", + "[30] Chen Liang, Jiahui Yu, Ming-Hsuan Yang, Matthew Brown, Yin Cui, Tuo Zhao, Boqing Gong, and Tianyi Zhou. Modulewise adaptive distillation for multimodality foundation models. Advances in Neural Information Processing Systems, 36, 2024. 5", + "[31] Jiangqi Liu and Feng Wang. mixed attention auto encoder for multi-class industrial anomaly detection. arXiv preprint arXiv:2309.12700, 2023. 2", + "[32] Wenqian Liu, Runze Li, Meng Zheng, Srikrishna Karanam, Ziyan Wu, Bir Bhanu, Richard J Radke, and Octavia Camps. Towards visually explaining variational autoencoders. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 8642-8651, 2020. 1", + "[33] Ze Liu, Yutong Lin, Yue Cao, Han Hu, Yixuan Wei, Zheng Zhang, Stephen Lin, and Baining Guo. Swin transformer: Hierarchical vision transformer using shifted windows. In IEEE/CVF International Conference on Computer Vision, pages 10012-10022, 2021. 3", + "[34] Zhikang Liu, Yiming Zhou, Yuansheng Xu, and Zilei Wang. Simplenet: A simple network for image anomaly detection and localization. arXiv preprint arXiv:2303.15140, 2023. 6, 7, 1, 8, 9, 10", + "[35] Ilya Loshchilov and Frank Hutter. Decoupled weight decay regularization. arXiv preprint arXiv:1711.05101, 2017. 6", + "[36] Ruiying Lu, YuJie Wu, Long Tian, Dongsheng Wang, Bo Chen, Xiyang Liu, and Ruimin Hu. Hierarchical vector quantized transformer for multi-class unsupervised anomaly detection. arXiv preprint arXiv:2310.14228, 2023. 1, 2, 4", + "[37] Amira Ben Mabrouk and Ezzeddine Zagrouba. Abnormal behavior recognition for intelligent video surveillance systems: A review. Expert Systems with Applications, 91:480-491, 2018. 1" + ], + "bbox": [ + 91, + 90, + 483, + 898 + ], + "page_idx": 9 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[38] Pankaj Mishra, Riccardo Verk, Daniele Fornasier, Claudio Piciarelli, and Gian Luca Foresti. Vt-adl: A vision transformer network for image anomaly detection and localization. In 2021 IEEE 30th International Symposium on Industrial Electronics (ISIE), pages 01–06. IEEE, 2021. 7, 6", + "[39] Maxime Oquab, Timothee Darcet, Theo Moutakanni, Huy Vo, Marc Szafraniec, Vasil Khalidov, Pierre Fernandez, Daniel Haziza, Francisco Massa, Alaaeldin El-Nouby, et al. Dinov2: Learning robust visual features without supervision. arXiv preprint arXiv:2304.07193, 2023. 3, 8, 2", + "[40] Z Peng, L Dong, H Bao, Q Ye, and F Wei. Beit v2: Masked image modeling with vector-quantized visual tokenizers. arXiv preprint arXiv:2208.06366, 2022. 3, 2", + "[41] Sashank J Reddi, Satyen Kale, and Sanjiv Kumar. On the convergence of adam and beyond. arXiv preprint arXiv:1904.09237, 2019. 6, 1", + "[42] Tal Reiss, Niv Cohen, Liron Bergman, and Yedid Hoshen. Panda: Adapting pretrained features for anomaly detection and segmentation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 2806-2814, 2021. 7", + "[43] Tal Reiss, Niv Cohen, Eliahu Horwitz, Ron Abutbul, and Yedid Hoshen. Anomaly detection requires better representations. In European Conference on Computer Vision, pages 56-68. Springer, 2022. 3, 7", + "[44] Sucheng Ren, Zeyu Wang, Hongru Zhu, Junfei Xiao, Alan Yuille, and Cihang Xie. Rejuvenating image-gpt as strong visual representation learners. arXiv preprint arXiv:2312.02147, 2023.8, 2", + "[45] Karsten Roth, Latha Pemula, Joaquin Zepeda, Bernhard Schölkopf, Thomas Brox, and Peter Gehler. Towards total recall in industrial anomaly detection. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 14318-14328, 2022. 7, 1, 3", + "[46] Mohammadreza Salehi, Niousha Sadjadi, Soroosh Baselizadeh, Mohammad H Rohban, and Hamid R Rabiee. Multiresolution knowledge distillation for anomaly detection. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 14902-14912, 2021. 5, 1", + "[47] Thomas Schlegl, Philipp Seebock, Sebastian M Waldstein, Georg Langs, and Ursula Schmidt-Erfurth. f-anogan: Fast unsupervised anomaly detection with generative adversarial networks. Medical image analysis, 54:30-44, 2019. 1", + "[48] Zhuoran Shen, Mingyuan Zhang, Haiyu Zhao, Shuai Yi, and Hongsheng Li. Efficient attention: Attention with linear complexities. In Proceedings of the IEEE/CVF winter conference on applications of computer vision, pages 3531-3539, 2021. 5", + "[49] Shelly Sheynin, Sagie Benaim, and Lior Wolf. A hierarchical transformation-discriminating generative model for few shot anomaly detection. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 8495-8504, 2021. 1", + "[50] Hugo Touvron, Matthieu Cord, Matthijs Douze, Francisco Massa, Alexandre Sablayrolles, and Herv'e J'egou. Training data-efficient image transformers & distillation through attention. arXiv preprint arXiv:2012.12877, 2021. 8, 2" + ], + "bbox": [ + 516, + 90, + 906, + 898 + ], + "page_idx": 9 + }, + { + "type": "page_number", + "text": "20414", + "bbox": [ + 478, + 944, + 519, + 955 + ], + "page_idx": 9 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[51] Ashish Vaswani, Noam Shazeer, Niki Parmar, Jakob Uszkoreit, Llion Jones, Aidan N Gomez, Lukasz Kaiser, and Illia Polosukhin. Attention is all you need. In Advances in Neural Information Processing Systems, pages 5998-6008, 2017. 2", + "[52] Pascal Vincent, Hugo Larochelle, Yoshua Bengio, and Pierre-Antoine Manzagol. Extracting and composing robust features with denoising autoencoders. In Proceedings of the 25th international conference on Machine learning, pages 1096-1103, 2008. 4", + "[53] Pascal Vincent, Hugo Larochelle, Isabelle Lajoie, Yoshua Bengio, and Pierre-Antoine Manzagol. Stacked denoising autoencoders: Learning useful representations in a deep network with a local denoising criterion. In Journal of Machine Learning Research, pages 3371-3408, 2010. 4", + "[54] Chengjie Wang, Wenbing Zhu, Bin-Bin Gao, Zhenye Gan, Jianning Zhang, Zhihao Gu, Shuguang Qian, Mingang Chen, and Lizhuang Ma. Real-iad: A real-world multi-view dataset for benchmarking versatile industrial anomaly detection. arXiv preprint arXiv:2403.12580, 2024. 2, 6, 7, 3", + "[55] Guodong Wang, Shumin Han, Errui Ding, and Di Huang. Student-teacher feature pyramid matching for anomaly detection. In The British Machine Vision Conference (BMVC), 2021. 1", + "[56] Mitchell Wortsman, Tim Dettmers, Luke Zettlemoyer, Ari Morcos, Ali Farhadi, and Ludwig Schmidt. Stable and low-precision training for large-scale vision-language models. Advances in Neural Information Processing Systems, 36: 10271–10298, 2023. 6, 1", + "[57] Zhenda Xie, Zheng Zhang, Yue Cao, Yutong Lin, Jianmin Bao, Zhuliang Yao, Qi Dai, and Han Hu. Simmim: A simple framework for masked image modeling. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 9653-9663, 2022. 3", + "[58] Jie Yang, Yong Shi, and Zhiquan Qi. Dfr: Deep feature reconstruction for unsupervised anomaly segmentation. arXiv preprint arXiv:2012.07122, 2020. 5, 1", + "[59] Haonan Yin, Guanlong Jiao, Qianhui Wu, Borne F Karlsson, Biqing Huang, and Chin Yew Lin. Lafite: Latent diffusion model with feature editing for unsupervised multi-class anomaly detection. arXiv preprint arXiv:2307.08059, 2023.1, 2, 3, 4", + "[60] Zhiyuan You, Lei Cui, Yujun Shen, Kai Yang, Xin Lu, Yu Zheng, and Xinyi Le. A unified model for multi-class anomaly detection. arXiv preprint arXiv:2206.03687, 2022. 1, 2, 3, 4, 5, 6, 7, 8, 9, 10", + "[61] Weihao Yu, Chenyang Si, Pan Zhou, Mi Luo, Yichen Zhou, Jiashi Feng, Shuicheng Yan, and Xinchao Wang. Metaformer baselines for vision. IEEE Transactions on Pattern Analysis and Machine Intelligence, 2023. 4", + "[62] Vitjan Zavrtanik, Matej Kristan, and Danijel Skoc aj. Draem-a discriminatively trained reconstruction embedding for surface anomaly detection. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 8330-8339, 2021. 4, 1", + "[63] Vitjan Zavrtanik, Matej Kristan, and Danijel Skocaj. Reconstruction by inpainting for visual anomaly detection. Pattern Recognition, 112:107706, 2021. 1" + ], + "bbox": [ + 91, + 90, + 480, + 898 + ], + "page_idx": 10 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[64] Dingwen Zhang, Guohai Huang, Qiang Zhang, Jungong Han, Junwei Han, Yizhou Wang, and Yizhou Yu. Exploring task structure for brain tumor segmentation from multimodality mr images. IEEE Transactions on Image Processing, 29:9032-9043, 2020. 7", + "[65] Jiangning Zhang, Xuhai Chen, Yabiao Wang, Chengjie Wang, Yong Liu, Xiangtai Li, Ming-Hsuan Yang, and Dacheng Tao. Exploring plain vit reconstruction for multiclass unsupervised anomaly detection. arXiv preprint arXiv:2312.07495, 2023. 3, 5, 6, 7", + "[66] Jiangning Zhang, Haoyang He, Zhenye Gan, Qingdong He, Yuxuan Cai, Zhucun Xue, Yabiao Wang, Chengjie Wang, Lei Xie, and Yong Liu. Ader: A comprehensive benchmark for multi-class visual anomaly detection. arXiv preprint arXiv:2406.03262, 2024. 7, 2, 6", + "[67] Xuan Zhang, Shiyu Li, Xi Li, Ping Huang, Jiulong Shan, and Ting Chen. Destseg: Segmentation guided denoising student-teacher for anomaly detection. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 3914–3923, 2023. 4, 6, 1, 7, 8, 9, 10", + "[68] Ying Zhao. Omnial: A unified cnn framework for unsupervised anomaly localization. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 3924-3933, 2023. 1, 3", + "[69] Jinghao Zhou, Chen Wei, Huiyu Wang, Wei Shen, Cihang Xie, Alan Yuille, and Tao Kong. ibot: Image bert pre-training with online tokenizer. arXiv preprint arXiv:2111.07832, 2021. 3, 8, 2", + "[70] Yang Zou, Jongheon Jeong, Latha Pemula, Dongqing Zhang, and Onkar Dabeer. Spot-the-difference self-supervised pretraining for anomaly detection and segmentation. In European Conference on Computer Vision, pages 392-408. Springer, 2022. 2, 6, 7, 3" + ], + "bbox": [ + 516, + 90, + 903, + 556 + ], + "page_idx": 10 + }, + { + "type": "page_number", + "text": "20415", + "bbox": [ + 478, + 944, + 517, + 955 + ], + "page_idx": 10 + } +] \ No newline at end of file diff --git a/2025/Dinomaly_ The Less Is More Philosophy in Multi-Class Unsupervised Anomaly Detection/269a00de-90dc-4a5c-8e52-737661810fb4_model.json b/2025/Dinomaly_ The Less Is More Philosophy in Multi-Class Unsupervised Anomaly Detection/269a00de-90dc-4a5c-8e52-737661810fb4_model.json new file mode 100644 index 0000000000000000000000000000000000000000..b2ebf931041467c32abf228e18fe959973efc47c --- /dev/null +++ b/2025/Dinomaly_ The Less Is More Philosophy in Multi-Class Unsupervised Anomaly Detection/269a00de-90dc-4a5c-8e52-737661810fb4_model.json @@ -0,0 +1,2532 @@ +[ + [ + { + "type": "header", + "bbox": [ + 0.107, + 0.003, + 0.182, + 0.043 + ], + "angle": 0, + "content": "CVF" + }, + { + "type": "header", + "bbox": [ + 0.237, + 0.0, + 0.812, + 0.047 + ], + "angle": 0, + "content": "This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore." + }, + { + "type": "title", + "bbox": [ + 0.268, + 0.131, + 0.73, + 0.177 + ], + "angle": 0, + "content": "Dinomaly: The Less Is More Philosophy in Multi-Class Unsupervised Anomaly Detection" + }, + { + "type": "text", + "bbox": [ + 0.149, + 0.202, + 0.848, + 0.224 + ], + "angle": 0, + "content": "Jia Guo\\(^{1}\\) Shuai Lu\\(^{2}\\) Weihang Zhang\\(^{2\\dagger}\\) Fang Chen\\(^{3}\\) Huiqi Li\\(^{2\\dagger}\\) Hongen Liao\\(^{1,3\\boxtimes}\\)" + }, + { + "type": "text", + "bbox": [ + 0.212, + 0.239, + 0.786, + 0.275 + ], + "angle": 0, + "content": "\\(^{1}\\)School of Biomedical Engineering, Tsinghua University, Beijing, China \\(^{2}\\)Beijing Institute of Technology, Beijing, China" + }, + { + "type": "text", + "bbox": [ + 0.164, + 0.275, + 0.834, + 0.293 + ], + "angle": 0, + "content": "\\(^{3}\\)School of Biomedical Engineering, Shanghai Jiao Tong University, Shanghai, China" + }, + { + "type": "text", + "bbox": [ + 0.21, + 0.296, + 0.782, + 0.327 + ], + "angle": 0, + "content": "guojia.jeremy@gmail.com lushuaie@163.com zhangweihang@bit.edu.cn chen-fang@sjtu.edu.cn huiqili@bit.edu.cn liao@tsinghua.edu.cn" + }, + { + "type": "title", + "bbox": [ + 0.249, + 0.378, + 0.327, + 0.394 + ], + "angle": 0, + "content": "Abstract" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.416, + 0.486, + 0.839 + ], + "angle": 0, + "content": "Recent studies highlighted a practical setting of unsupervised anomaly detection (UAD) that builds a unified model for multi-class images. Despite various advancements addressing this challenging task, the detection performance under the multi-class setting still lags far behind state-of-the-art class-separated models. Our research aims to bridge this substantial performance gap. In this paper, we present Dinomaly, a minimalist reconstruction-based anomaly detection framework that harnesses pure Transformer architectures without relying on complex designs, additional modules, or specialized tricks. Given this powerful framework consisting of only Attentions and MLPs, we found four simple components that are essential to multi-class anomaly detection: (1) Scalable foundation Transformers that extracts universal and discriminative features, (2) Noisy Bottleneck where pre-existing Dropouts do all the noise injection tricks, (3) Linear Attention that naturally cannot focus, and (4) Loose Reconstruction that does not force layer-to-layer and point-by-point reconstruction. Extensive experiments are conducted across popular anomaly detection benchmarks including MVtec-AD, VisA, Real-IAD, etc. Our proposed Dinomaly achieves impressive image-level AUROC of \\(99.6\\%\\), \\(98.7\\%\\), and \\(89.3\\%\\) on the three datasets respectively, which is not only superior to state-of-the-art multi-class UAD methods, but also achieves the most advanced class-separated UAD records. Code is available at: https://github.com/guojiajeremy/Dinomaly" + }, + { + "type": "title", + "bbox": [ + 0.514, + 0.362, + 0.645, + 0.377 + ], + "angle": 0, + "content": "1. Introduction" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.387, + 0.907, + 0.538 + ], + "angle": 0, + "content": "Unsupervised anomaly detection (UAD) aims to detect abnormal patterns from normal images and further localize the anomalous regions. Because of the diversity of potential anomalies and their scarcity, this task is proposed to model the accessible training sets containing only normal samples as an unsupervised paradigm. UAD has a wide range of applications, e.g., industrial defect detection [3], medical disease screening [13], and video surveillance [37], addressing the difficulty of collecting and labeling all possible anomalies in these scenarios." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.539, + 0.909, + 0.825 + ], + "angle": 0, + "content": "Conventional works on UAD build a separate model for each object category, as shown in Figure 1(a). However, this one-class-one-model setting entails substantial storage overhead for saving models [60], especially when the application scenario necessitates a large number of object classes. For UAD methods, a compact boundary of normal patterns is vital to distinguish anomalies. Once the intra-normal patterns become exceedingly complicated due to various classes, the corresponding distribution becomes challenging to measure, consequently harming the detection performance. Recently, UniAD [60] and successive studies have been proposed to train a unified model for multi-class anomaly detection (MUAD), as shown in Figure 1(b). Under this setting, the \"identity mapping\" that directly copies the input as the output regardless of normal or anomaly harms the performance of conventional methods [60]. This phenomenon is caused by the diversity of multi-class normal patterns that drive the network to generalize on unseen patterns." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.826, + 0.909, + 0.903 + ], + "angle": 0, + "content": "Within two years, a number of methods have been proposed to address MUAD, such as neighbor-masked attention [60], synthetic anomalies [68], vector quantization [36], diffusion model [16, 59], and state space model (Mamba) [17]. However, there is still a non-negligible per" + }, + { + "type": "page_footnote", + "bbox": [ + 0.114, + 0.887, + 0.483, + 0.901 + ], + "angle": 0, + "content": "Corresponding author. † Former advisors when the project initiated." + }, + { + "type": "page_number", + "bbox": [ + 0.479, + 0.945, + 0.519, + 0.958 + ], + "angle": 0, + "content": "20405" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.116, + 0.098, + 0.273, + 0.19 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.12, + 0.191, + 0.262, + 0.202 + ], + "angle": 0, + "content": "(a) Class-Separated UAD Setting" + }, + { + "type": "image", + "bbox": [ + 0.289, + 0.096, + 0.492, + 0.22 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.349, + 0.22, + 0.454, + 0.23 + ], + "angle": 0, + "content": "MVTec-AD (15 classes)" + }, + { + "type": "image", + "bbox": [ + 0.495, + 0.096, + 0.681, + 0.22 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.554, + 0.221, + 0.631, + 0.23 + ], + "angle": 0, + "content": "VisA (12 classes)" + }, + { + "type": "image", + "bbox": [ + 0.687, + 0.096, + 0.872, + 0.22 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.735, + 0.221, + 0.831, + 0.23 + ], + "angle": 0, + "content": "Real-IAD (30 classes)" + }, + { + "type": "image", + "bbox": [ + 0.116, + 0.228, + 0.27, + 0.313 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.12, + 0.318, + 0.135, + 0.327 + ], + "angle": 0, + "content": "(b)" + }, + { + "type": "image_caption", + "bbox": [ + 0.169, + 0.316, + 0.231, + 0.323 + ], + "angle": 0, + "content": "Model-Unified" + }, + { + "type": "image_caption", + "bbox": [ + 0.148, + 0.324, + 0.252, + 0.333 + ], + "angle": 0, + "content": "Multi-Class UAD Setting" + }, + { + "type": "image", + "bbox": [ + 0.296, + 0.246, + 0.486, + 0.327 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.513, + 0.234, + 0.707, + 0.244 + ], + "angle": 0, + "content": "(c) Image-Level AUROC on 3 Benchmarks" + }, + { + "type": "image", + "bbox": [ + 0.497, + 0.246, + 0.683, + 0.328 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.692, + 0.247, + 0.886, + 0.329 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.456, + 0.332, + 0.766, + 0.344 + ], + "angle": 0, + "content": "(d) Scaling Properties of Dinomaly on MVTec-AD. \\(\\dagger\\) : default setting." + }, + { + "type": "image_caption", + "bbox": [ + 0.09, + 0.355, + 0.907, + 0.385 + ], + "angle": 0, + "content": "Figure 1. Setting, benchmarking, and scaling of Dinomaly. (a) Task setting of class-separated UAD. (b) Task setting of MUAD. (c) Comparison with previous SoTA methods on MVtec-AD [3], VisA [70], and Real-IAD [54]. (d) Scaling properties of Dinomaly." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.41, + 0.484, + 0.516 + ], + "angle": 0, + "content": "formance gap between the state-of-the-art (SoTA) MUAD methods and class-separated UAD methods, restricting the practicability of implementing unified models, as shown in Figure 1(c). In addition, previous methods employ modules and architectures delicately designed, which may not be straightforward, and consequently suffer from limited universality and ease-of-use [18, 36]." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.524, + 0.484, + 0.902 + ], + "angle": 0, + "content": "In this work, we aim to catch up with the performance of class-separated anomaly detection models using a multiclass unified model. We introduce Dinomaly, a minimalist reconstruction-based UAD framework built exclusively by pure Transformer blocks [51], specifically Self-Attentions and Multi-Layer Perceptrons (MLPs). To begin with, we empirically investigate the scaling law of self-supervised pre-trained Vision Transformers (ViT) [12] when serving as the feature encoders for extracting reconstruction objectives. Subsequently, we introduce three straightforward yet crucial elements to address the critical identity mapping phenomenon in MUAD contexts, without increasing complexity or computational burden. First, as an alternative to meticulously designed pseudo anomaly and feature noise, we propose to activate the built-in Dropout in an MLP to prevent the network from restoring both normal and anomalous patterns. Second, we propose to leverage the \"side effect\" of Linear Attention (a computation-efficient counterpart of Softmax Attention) that impedes focus on local regions, thus preventing the forwarding of identical information. Third, previous methods adopt layer-to-layer and region-by-region reconstruction schemes, distilling a decoder that can well mimic the encoder's behavior even for anomalous regions. Therefore, we propose to loosen the reconstruction constraints by grouping multiple layers as a" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.41, + 0.905, + 0.439 + ], + "angle": 0, + "content": "whole and discarding well-reconstructed regions during optimization." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.44, + 0.907, + 0.637 + ], + "angle": 0, + "content": "To validate the effectiveness of the proposed Dinomaly under MUAD setting, we conduct extensive experiments on a number of widely used benchmarks, including MVTec AD [3] (15 classes), VisA [70] (12 classes), and RealIAD (30 classes). As shown in Figure 1, our base-size Dinomaly achieves unprecedented image-level AUROC of \\(99.6\\%\\), \\(98.7\\%\\), and \\(89.3\\%\\) on MVTec AD, VisA, and RealIAD, surpassing previous SoTA methods by a large margin. In addition, scalability is a key feature of Dinomaly. Further scaling up the model size maximizes performance to the fullest level of \\(99.8\\%\\), \\(98.9\\%\\), and \\(90.1\\%\\), respectively; while scaling down parameters and input size can offer efficient solutions in computation-constrained scenarios." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.649, + 0.655, + 0.665 + ], + "angle": 0, + "content": "2. Related Work" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.674, + 0.907, + 0.901 + ], + "angle": 0, + "content": "Multi-Class UAD. UniAD [60] first introduced multi-class anomaly detection, aiming to detect anomalies for different classes using a unified model. In this setting, conventional UAD methods often face the challenge of \"identical shortcuts\", where both anomaly-free and anomaly samples can be effectively recovered during inference [60]. It is believed that this phenomenon is caused by the diversity of multi-class normal patterns that drive the network to generalize on unseen patterns. This contradicts the fundamental assumption of epistemic methods. Many current researches focus on addressing this challenge [14, 31, 36, 59, 60]. UniAD [60] employs a neighbor-masked attention module and a feature-jitter strategy to mitigate these shortcuts. HVQ-Trans [36] proposes a vector quantization (VQ) Transformer model that induces large feature discrepancies" + }, + { + "type": "page_number", + "bbox": [ + 0.48, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "20406" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.095, + 0.093, + 0.91, + 0.288 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.22, + 0.301, + 0.776, + 0.316 + ], + "angle": 0, + "content": "Figure 2. The framework of Dinomaly, built by simple and pure Transformer building blocks." + }, + { + "type": "text", + "bbox": [ + 0.089, + 0.342, + 0.486, + 0.542 + ], + "angle": 0, + "content": "for anomalies. LafitE [59] utilizes a latent diffusion model and introduces a feature editing strategy to alleviate this issue. DiAD [16] also employs diffusion models to address multi-class UAD settings. OmniAL [68] focuses on anomaly localization in the unified setting, preventing identical reconstruction by using synthesized pseudo anomalies. ReContrast [14] attempted to alleviate the identity mapping by cross-reconstruction between two encoders. ViTAD [5] abstracts a unified feature-reconstruction UAD framework and employ Transformer building blocks. MambaAD [17] explores the recently proposed State Space Model (SSM), Mamba, in the context of multi-class UAD. More related works of UAD are presented in Appendix A." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.551, + 0.184, + 0.567 + ], + "angle": 0, + "content": "3. Method" + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.576, + 0.3, + 0.592 + ], + "angle": 0, + "content": "3.1. Dinomaly Framework" + }, + { + "type": "text", + "bbox": [ + 0.093, + 0.599, + 0.388, + 0.613 + ], + "angle": 0, + "content": "\"What I cannot create, I do not understand\"" + }, + { + "type": "text", + "bbox": [ + 0.11, + 0.615, + 0.258, + 0.629 + ], + "angle": 0, + "content": "Richer Feynman" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.63, + 0.486, + 0.902 + ], + "angle": 0, + "content": "The ability to recognize anomalies from what we know is an innate human capability, serving as a vital pathway for us to explore the world. Similarly, we construct a reconstruction-based framework that relies on the epistemic characteristic of artificial neural networks. Dinomaly consists of an encoder, a bottleneck, and a reconstruction decoder, as shown in Figure 2. Without loss of generality, a pretrained ViT network [12] with 12 Transformer layers is used as the encoder, extracting informative feature maps with different semantic scales. The bottleneck is a simple MLP (a.k.a. feed-forward network, FFN) that collects the feature representations of the encoder's 8 middle-level layers. The decoder is similar to the encoder, consisting of 8 Transformer layers. During training, the decoder learns to reconstruct the middle-level features of the encoder by maximizing the cosine similarity between feature maps. During inference, the decoder is expected to reconstruct normal regions of feature maps but fails for anomalous regions as it" + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.343, + 0.707, + 0.357 + ], + "angle": 0, + "content": "has never seen such samples." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.358, + 0.907, + 0.493 + ], + "angle": 0, + "content": "Foundation Transformers. Foundation models, especially ViTs [12, 33] pre-trained on large-scale datasets, serve as a basis and starting point for specific computer vision tasks. Such networks employ self-supervised learning schemes such as contrastive learning (MoCov3 [6], DINO [4]), masked image modeling (MAE [19], SimMIM [57], BEiT [40]), and their combination (iBOT [69], DINOv2 [39]), producing universal features suitable for image-level visual tasks and pixel-level visual tasks." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.494, + 0.909, + 0.72 + ], + "angle": 0, + "content": "Because of the lack of supervision in UAD, most advanced methods adopt pre-trained networks to extract discriminative features. Recent works [28, 43, 65] have preliminarily discovered the advantage of robust and universal features of self-supervised models over domain-specific ImageNet features in anomaly detection tasks. In this paper, we pioneer the investigation of scaling behaviors in UAD models through a systematic analysis of foundational ViTs, as briefed in Figure 1(d). Our comprehensive evaluation encompasses pre-training strategies (Figure 5), model sizes (Table 4), and input resolutions (Table 5), which are detailed in section 4.4. Considering the balance of detection performance and computational efficiency, we adopt ViT-Base/14 pretrained by DINoV2-Register [7] as the encoder of Dinomaly by default." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.727, + 0.686, + 0.743 + ], + "angle": 0, + "content": "3.2. Noisy Bottleneck." + }, + { + "type": "text", + "bbox": [ + 0.514, + 0.75, + 0.694, + 0.765 + ], + "angle": 0, + "content": "\"Dropout is all you need.\"" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.766, + 0.909, + 0.903 + ], + "angle": 0, + "content": "Prior studies [14, 60, 68] attribute the performance degradation of UAD methods trained on diverse multi-class samples to the \"identity mapping\" phenomenon; in this work, we reframe this issue as an \"over-generalization\" problem. Generalization ability is a merit of neural networks, allowing them to perform equally well on unseen test sets. However, generalization is not so wanted in the context of unsupervised anomaly detection that leverages the epistemic nature of neural networks. With the increas" + }, + { + "type": "page_number", + "bbox": [ + 0.479, + 0.945, + 0.521, + 0.958 + ], + "angle": 0, + "content": "20407" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.159, + 0.093, + 0.568, + 0.341 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.589, + 0.117, + 0.822, + 0.338 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.166, + 0.357, + 0.83, + 0.373 + ], + "angle": 0, + "content": "Figure 3. Softmax Attention vs. Linear Attention. (a) Visualization of attention maps. (b) Attention distribution." + }, + { + "type": "text", + "bbox": [ + 0.089, + 0.398, + 0.483, + 0.458 + ], + "angle": 0, + "content": "ing diversity of images and their patterns due to multi-class UAD settings, the decoder can generalize its reconstruction ability to unseen anomalous samples, resulting in the failure of anomaly detection using reconstruction error." + }, + { + "type": "text", + "bbox": [ + 0.092, + 0.459, + 0.485, + 0.822 + ], + "angle": 0, + "content": "A direct solution for identity mapping is to shift \"reconstruction\" to \"restoration\". Specifically, instead of directly reconstructing the normal images or features given normal inputs, previous works propose to add perturbations as pseudo anomalies on input images [62, 67] or on encoder features [59, 60]; meanwhile, still let the decoder restore anomaly-free images or features, formulating a denoising-like framework. However, such methods employ heuristic and hand-crafted anomaly generation strategies that may not be universal across domains, datasets, and methods. In this work, we turn to leveraging the simple and elegant Dropout techniques. Since its introduction by Hinton et al. [21] in 2014 as a remedy for overfitting, Dropout has become a cornerstone element in neural architectures, including Transformers. In Dinomaly, we employ Dropout to randomly discard neural activations in an MLP bottleneck. Instead of alleviating overfitting, the role of Dropout in Dinomaly can be explained as pseudo feature anomaly that perturb normal representations, analogous to denoising auto-encoders [52, 53]. Without introducing any specific modules, this simple component inherently forces the decoder to restore normal features regardless of whether the test image contains anomalies, in turn, mitigating identical mapping." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.832, + 0.345, + 0.848 + ], + "angle": 0, + "content": "3.3. Unfocused Linear Attention." + }, + { + "type": "text", + "bbox": [ + 0.093, + 0.856, + 0.383, + 0.87 + ], + "angle": 0, + "content": "\"One man's poison is another man's meat\"" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.871, + 0.484, + 0.902 + ], + "angle": 0, + "content": "Softmax Attention is the key mechanism of Transformers, allowing the model to attend to different parts of its" + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.398, + 0.907, + 0.459 + ], + "angle": 0, + "content": "input token sequence. Formally, given an input sequence \\(\\mathbf{X} \\in \\mathbb{R}^{N \\times d}\\) with length \\(N\\), Attention first transforms it into three matrices: the query matrix \\(\\mathbf{Q} \\in \\mathbb{R}^{N \\times d}\\), the key matrix \\(\\mathbf{K} \\in \\mathbb{R}^{N \\times d}\\), and the value matrix \\(\\mathbf{V} \\in \\mathbb{R}^{N \\times d}\\):" + }, + { + "type": "equation", + "bbox": [ + 0.569, + 0.47, + 0.907, + 0.489 + ], + "angle": 0, + "content": "\\[\n\\mathbf {Q} = \\mathbf {X} \\mathbf {W} ^ {Q}, \\mathbf {K} = \\mathbf {X} \\mathbf {W} ^ {K}, \\mathbf {V} = \\mathbf {X} \\mathbf {W} ^ {V}, \\tag {1}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.501, + 0.906, + 0.547 + ], + "angle": 0, + "content": "where \\(\\mathbf{W}^Q, \\mathbf{W}^K, \\mathbf{W}^V \\in \\mathbb{R}^{d \\times d}\\) are learnable parameters. By computing the attention map by the query-key similarity, the output of Softmax Attention is given as:" + }, + { + "type": "equation", + "bbox": [ + 0.565, + 0.56, + 0.907, + 0.579 + ], + "angle": 0, + "content": "\\[\n\\operatorname {A t t e n t i o n} (\\mathbf {Q}, \\mathbf {K}, \\mathbf {V}) = \\operatorname {S o f t m a x} (\\mathbf {Q K} ^ {T}) \\mathbf {V}. \\tag {2}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.592, + 0.907, + 0.682 + ], + "angle": 0, + "content": "Back to MUAD, previous methods [36, 60] suggest adopting Attentions instead of Convolutions because Convolutions can easily learn identical mappings. Nevertheless, both operations are in danger of forming identity mapping by over-concentrating on corresponding input locations for producing the outputs:" + }, + { + "type": "equation", + "bbox": [ + 0.513, + 0.694, + 0.904, + 0.744 + ], + "angle": 0, + "content": "\\[\n\\operatorname {C o n v} \\operatorname {K e r n e l} = \\left[ \\begin{array}{l l l} 0 & 0 & 0 \\\\ 0 & 1 & 0 \\\\ 0 & 0 & 0 \\end{array} \\right], \\quad \\operatorname {A t t n} \\operatorname {M a p} = \\left[ \\begin{array}{l l l} 1 & 0 & 0 \\\\ 0 & 1 & 0 \\\\ 0 & 0 & 1 \\end{array} \\right].\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.755, + 0.907, + 0.846 + ], + "angle": 0, + "content": "Is there any simple solution that prevents Attentions from attending to identical information? In Dinomaly, we turn to leverage the \"unfocusing ability\" of a type of Softmax-free Attention, i.e., Linear Attention. Linear Attention was proposed as a promising alternative to reduce the computation complexity of vanilla Softmax Attention" + }, + { + "type": "page_footnote", + "bbox": [ + 0.512, + 0.858, + 0.907, + 0.901 + ], + "angle": 0, + "content": "1The full form of Attention is Softmax \\(\\left(\\frac{\\mathbf{Q}\\mathbf{K}^T}{\\sqrt{d}}\\right)\\mathbf{V}\\). The constant denominator is omitted for narrative simplicity. The multi-head mechanism that concatenates multiple Attentions is also omitted." + }, + { + "type": "footer", + "bbox": [ + 0.479, + 0.945, + 0.52, + 0.957 + ], + "angle": 0, + "content": "20408" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.107, + 0.09, + 0.235, + 0.186 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.128, + 0.187, + 0.231, + 0.197 + ], + "angle": 0, + "content": "(a) layer-to-layer sparse" + }, + { + "type": "image", + "bbox": [ + 0.252, + 0.089, + 0.357, + 0.184 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.259, + 0.186, + 0.352, + 0.197 + ], + "angle": 0, + "content": "(b) layer-to-cat-layer" + }, + { + "type": "image", + "bbox": [ + 0.396, + 0.089, + 0.501, + 0.184 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.397, + 0.186, + 0.499, + 0.197 + ], + "angle": 0, + "content": "(c) layer-to-layer dense" + }, + { + "type": "image", + "bbox": [ + 0.535, + 0.088, + 0.637, + 0.184 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.541, + 0.184, + 0.624, + 0.204 + ], + "angle": 0, + "content": "(d) group-to-group 1 group (Ours)" + }, + { + "type": "image", + "bbox": [ + 0.673, + 0.089, + 0.775, + 0.183 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.679, + 0.184, + 0.761, + 0.204 + ], + "angle": 0, + "content": "(e) group-to-group 2 group (Ours)" + }, + { + "type": "image", + "bbox": [ + 0.816, + 0.097, + 0.838, + 0.116 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.838, + 0.097, + 0.872, + 0.116 + ], + "angle": 0, + "content": "encoder feature" + }, + { + "type": "image", + "bbox": [ + 0.817, + 0.125, + 0.838, + 0.14 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.838, + 0.123, + 0.87, + 0.14 + ], + "angle": 0, + "content": "decoder feature" + }, + { + "type": "image", + "bbox": [ + 0.817, + 0.148, + 0.838, + 0.156 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.838, + 0.147, + 0.889, + 0.156 + ], + "angle": 0, + "content": "reconstruct" + }, + { + "type": "image", + "bbox": [ + 0.817, + 0.163, + 0.838, + 0.173 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.838, + 0.163, + 0.889, + 0.173 + ], + "angle": 0, + "content": "concatenate" + }, + { + "type": "image", + "bbox": [ + 0.817, + 0.179, + 0.838, + 0.188 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.838, + 0.181, + 0.857, + 0.188 + ], + "angle": 0, + "content": "add" + }, + { + "type": "image_caption", + "bbox": [ + 0.089, + 0.219, + 0.907, + 0.248 + ], + "angle": 0, + "content": "Figure 4. Schemes of reconstruction constraint. (a) Layer-to-layer (sparse). (b) Layer-to-cat-layer. (c) Layer-to-layer (dense). (d) Loose group-to-group, 1-group (Ours). (e) Loose group-to-group, 2-group (Ours)." + }, + { + "type": "text", + "bbox": [ + 0.089, + 0.274, + 0.485, + 0.35 + ], + "angle": 0, + "content": "concerning the number of tokens [26]. By substituting Softmax operation with a simple activation function \\(\\phi (\\cdot)\\) (usually \\(\\phi (x) = \\mathrm{elu}(x) + 1\\)), we can change the computation order from \\((\\mathbf{Q}\\mathbf{K}^T)\\mathbf{V}\\) to \\(\\mathbf{Q}(\\mathbf{K}^T\\mathbf{V})\\). Formally, Linear Attention (LA) is given as:" + }, + { + "type": "equation", + "bbox": [ + 0.1, + 0.355, + 0.484, + 0.387 + ], + "angle": 0, + "content": "\\[\n\\operatorname {L A} (\\mathbf {Q}, \\mathbf {K}, \\mathbf {V}) = (\\phi (\\mathbf {Q}) \\phi \\left(\\mathbf {K} ^ {T}\\right)) \\mathbf {V} = \\phi (\\mathbf {Q}) \\left(\\phi \\left(\\mathbf {K} ^ {T}\\right) \\mathbf {V}\\right), \\tag {3}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.089, + 0.387, + 0.484, + 0.538 + ], + "angle": 0, + "content": "where the computation complexity is reduced to \\(\\mathcal{O}(Nd^2)\\) from \\(\\mathcal{O}(N^2 d)\\). The trade-off between complexity and expressiveness is a dilemma. Previous studies [15, 48] attribute Linear Attention's performance degradation on supervised tasks to its incompetence in focusing. Due to the absence of non-linear attention reweighting by Softmax operation, Linear Attention cannot concentrate on important regions related to the query, such as foreground and neighbors. This property, however, is exactly what the reconstruction decoder favors in our contexts." + }, + { + "type": "text", + "bbox": [ + 0.089, + 0.539, + 0.484, + 0.734 + ], + "angle": 0, + "content": "In order to probe how Attentions propagate information, we train two variants of Dinomaly using vanilla Softmax Attention or Linear Attention as the spatial mixer in the decoder and visualize their attention maps. As shown in Figure 3, Softmax Attention tends to focus on the exact region of the query, while Linear Attention spreads its attention across the whole image. This implies that Linear Attention, forced by its incompetence to focus, utilizes more long-range information to restore features at each position, reducing the chance of passing identical information of unseen patterns to the next layer during reconstruction. Of course, employing Linear Attention also benefits from less computation, free of performance drop." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.743, + 0.295, + 0.757 + ], + "angle": 0, + "content": "3.4. Loose Reconstruction" + }, + { + "type": "text", + "bbox": [ + 0.093, + 0.765, + 0.39, + 0.78 + ], + "angle": 0, + "content": "\"The tighter you squeeze, the less you have.\"" + }, + { + "type": "text", + "bbox": [ + 0.089, + 0.78, + 0.484, + 0.901 + ], + "angle": 0, + "content": "Loose Constraint. Pioneers of feature-reconstruction/distillation UAD methods [10, 46] are inspired by knowledge distillation [20]. Most reconstruction-based methods distill specific encoder layers (e.g. 3 last layers of 3 ResNet stages) by the corresponding decoder layers [10, 46, 65] (Figure 4(a)) or the last decoder layer [58, 60] (Figure 4(b)). Intuitively, with more encoder-decoder feature pairs (Figure 4(c)), UAD model can utilize" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.274, + 0.907, + 0.56 + ], + "angle": 0, + "content": "more information in different layers to discriminate anomalies. However, according to the intuition of knowledge distillation, the student (decoder) can better mimic the behavior of the teacher (encoder) given more layer-to-layer supervision [30], which is harmful for UAD models that detect anomalies by encoder-decoder discrepancy. This phenomenon is also embodied as identity mapping. Thanks to the top-to-bottom consistency of columnar Transformer layers, we propose to loosen the layer-to-layer constraint by adding up all feature maps of interested layers as a whole group, as shown in Figure 4(d). This scheme can be seen as loosening the layer-to-layer correspondence and providing the decoder with more degrees of freedom, so that the decoder is allowed to act much more differently from the encoder when the input pattern is unseen. Because features of shallow layers contain low-level visual characters that are helpful for precise localization, we can further group the features into the low-semantic-level group and high-semantic-level group, as shown in Figure 4(e)." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.561, + 0.909, + 0.667 + ], + "angle": 0, + "content": "**Loose Loss.** Following the above analysis, we also loosen the point-by-point reconstruction loss function by discarding some points in the feature map. Here, we simply borrow the hard-mining global cosine loss [14] that detaches the gradients of well-restored feature points with low cosine distance during training. Let \\( f_{E} \\) and \\( f_{D} \\) denotes (grouped) feature maps of encoder and decoder:" + }, + { + "type": "equation", + "bbox": [ + 0.585, + 0.675, + 0.905, + 0.693 + ], + "angle": 0, + "content": "\\[\n\\mathcal {L} _ {\\text {g l o b a l} - h m} = \\mathcal {D} _ {\\text {c o s}} \\left(\\mathcal {F} \\left(f _ {E}\\right), \\mathcal {F} \\left(\\hat {f} _ {D}\\right)\\right), \\tag {4}\n\\]" + }, + { + "type": "equation", + "bbox": [ + 0.515, + 0.7, + 0.905, + 0.741 + ], + "angle": 0, + "content": "\\[\n\\hat {f} _ {D} (h, w) = \\left\\{ \\begin{array}{l} s g \\left(f _ {D} (h, w)\\right)_{0.1}, \\text {if} \\mathcal {D} _ {\\text {cos}} \\left(f _ {D}, f _ {E}\\right) < k \\% _ {\\text {batch}} \\\\ f _ {D} (h, w), \\text {else} \\end{array} \\right. \\tag{5}\n\\]" + }, + { + "type": "equation", + "bbox": [ + 0.623, + 0.741, + 0.905, + 0.771 + ], + "angle": 0, + "content": "\\[\n\\mathcal {D} _ {\\cos} (a, b) = 1 - \\frac {a ^ {T} \\cdot b}{\\| a \\| \\| b \\|}, \\tag {6}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.773, + 0.907, + 0.88 + ], + "angle": 0, + "content": "where \\(\\mathcal{D}_{cos}\\) denotes cosine distance, \\(\\mathcal{F}(\\cdot)\\) denotes flatten operation, \\(f_{D}(h,w)\\) represents the feature point at \\((h,w)\\), and \\(sg(\\cdot)_{0.1}\\) denotes shrink the gradient to one-tenth of the original \\(^2\\). \\(\\mathcal{D}_{cos}(f_D(h,w),f_E(h,w)) < k\\%_{batch}\\) selects \\(k \\%\\) feature points with smaller cosine distance within a batch for gradient shrinking. Total loss is the average \\(\\mathcal{L}_{global - hm}\\) of all encoder-decoder feature pairs." + }, + { + "type": "page_footnote", + "bbox": [ + 0.53, + 0.887, + 0.898, + 0.901 + ], + "angle": 0, + "content": "\\( {}^{2} \\) Complete stop-gradient causes optimization instability occasionally." + }, + { + "type": "footer", + "bbox": [ + 0.479, + 0.945, + 0.52, + 0.957 + ], + "angle": 0, + "content": "20409" + } + ], + [ + { + "type": "table_caption", + "bbox": [ + 0.09, + 0.089, + 0.908, + 0.119 + ], + "angle": 0, + "content": "Table 1. Performance under multi-class UAD setting (%). †: method designed for MUAD. Dinomaly†: training schedule is scaled up to 20,000, 20,000, and 100,000 iterations (original: 10,000/10,000/50,000)." + }, + { + "type": "table", + "bbox": [ + 0.134, + 0.129, + 0.868, + 0.54 + ], + "angle": 0, + "content": "
DatasetMethodImage-levelPixel-level
AUROCAP\\(F_1\\)-maxAUROCAP\\(F_1\\)-maxAUPRO
MVTec-AD [3]RD4AD [10]94.696.595.296.148.653.891.1
SimpleNet [34]95.398.495.896.945.949.786.5
DeSTSeg [67]89.295.591.693.154.350.964.8
UniAD [60]†96.598.896.296.843.449.590.7
ReContrast [14]†98.399.497.697.160.261.593.2
DiAD [18]†97.299.096.596.852.655.590.7
ViTAD [65]†98.399.497.397.755.358.791.4
MambaAD [17]†98.699.697.897.756.359.293.1
Dinomaly (Ours)99.699.899.098.469.369.294.8
Dinomaly†99.799.899.298.469.369.294.7
VisA [70]RD4AD [10]92.492.489.698.138.042.691.8
SimpleNet [34]87.287.081.896.834.737.881.4
DeSTSeg [67]88.989.085.296.139.643.467.4
UniAD [60]†88.890.885.898.333.739.085.5
ReContrast [14]†95.596.492.098.547.950.691.9
DiAD [18]†86.888.385.196.026.133.075.2
ViTAD [65]†90.591.786.398.236.641.185.1
MambaAD [17]†94.394.589.498.539.444.091.0
Dinomaly (Ours)98.798.996.298.753.255.794.5
Dinomaly†98.999.096.498.853.855.894.5
Real-IAD [54]RD4AD [10]82.479.073.997.325.032.789.6
SimpleNet [34]57.253.461.575.72.86.539.0
DeSTSeg [67]82.379.273.294.637.941.740.6
UniAD [60]†83.080.974.397.321.129.286.7
ReContrast [14]†86.484.277.497.831.638.291.8
DiAD [18]†75.666.469.988.02.97.158.1
ViTAD [65]†82.780.273.797.224.332.384.8
MambaAD [17]†86.384.677.098.533.038.790.5
Dinomaly (Ours)89.386.880.298.842.847.193.9
Dinomaly†89.586.980.498.943.347.494.2
" + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.564, + 0.224, + 0.581 + ], + "angle": 0, + "content": "4. Experiments" + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.589, + 0.299, + 0.607 + ], + "angle": 0, + "content": "4.1. Experimental Settings" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.613, + 0.483, + 0.794 + ], + "angle": 0, + "content": "Datasets. MVtec-AD [3] contains 15 objects (5 texture classes and 10 object classes) with a total of 3,629 normal images as the training set and 1,725 images as the test set (467 normal, 1,258 anomalous). VisA [70] contains 12 objects. Training and test sets are split following the official splitting, resulting in 8,659 normal images in the training set and 2,162 images in the test set (962 normal, 1,200 anomalous). Real-IAD [54] is a large UAD dataset recently released, containing 30 distinct objects. We follow the official splitting that includes all views, resulting in 36,465 normal images in the training set and 114,585 images in the test set (63,256 normal, 51,329 anomalous)." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.796, + 0.483, + 0.902 + ], + "angle": 0, + "content": "Metrics. Following prior works [17, 65], we adopt 7 evaluation metrics. Image-level anomaly detection performance is measured by the Area Under the Receiver Operator Curve (AUROC), Average Precision (AP), and \\(F_{1}\\) score under optimal threshold \\((F_{1}\\)-max). Pixel-level anomaly localization is measured by AUROC, AP, \\(F_{1}\\)-max and the Area Under the Per-Region-Overlap (AUPRO). The results" + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.566, + 0.779, + 0.58 + ], + "angle": 0, + "content": "of a dataset is the average of all classes." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.581, + 0.907, + 0.793 + ], + "angle": 0, + "content": "Implementation Details. ViT-Base/14 (patchsize=14) pre-trained by DINOv2-R [7] is used as the encoder by default. The drop rate of Noisy Bottleneck is 0.2 by default and increases to 0.4 on the diverse Real-IAD. Loose constraint with 2 groups is employed, and the anomaly map is given by the mean per-point cosine distance of the 2 groups. The input image is first resized to \\(448^2\\) and then center-cropped to \\(392^2\\), so the feature map \\((28^2)\\) is large enough for anomaly localization. StableAdamW optimizer [56] with AMSGrad [41] (more stable than AdamW [35] in training) is utilized with \\(lr = 2e - 3\\), \\(\\beta = (0.9,0.999)\\) and \\(wd = 1e - 4\\). The network is trained for 10,000 iterations (steps) on MVTec-AD and VisA, and 50,000 iterations on Real-IAD. Detailed settings are available in Appendix B." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.802, + 0.855, + 0.818 + ], + "angle": 0, + "content": "4.2. Comparison to Multi-Class UAD SoTAs" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.826, + 0.906, + 0.901 + ], + "angle": 0, + "content": "We compare the proposed Dinomaly with the most advanced UAD and MUAD methods [10, 14, 17, 18, 34, 65, 67]. Experimental results are presented in Table 1, where Dinomaly surpasses compared methods by a large margin on all datasets and all metrics. On the most" + }, + { + "type": "page_number", + "bbox": [ + 0.48, + 0.945, + 0.521, + 0.958 + ], + "angle": 0, + "content": "20410" + } + ], + [ + { + "type": "table_caption", + "bbox": [ + 0.218, + 0.09, + 0.779, + 0.104 + ], + "angle": 0, + "content": "Table 2. Performance under conventional class-separated UAD setting (%). n/a: not available." + }, + { + "type": "table", + "bbox": [ + 0.134, + 0.115, + 0.865, + 0.22 + ], + "angle": 0, + "content": "
MethodMVTec-AD [3]VisA [70]Real-IAD [54]
I-AUROCP-AUROCP-AUPROI-AUROCP-AUROCP-AUPROI-AUROCP-AUROCP-AUPRO
Dinomaly (MUAD)99.698.494.898.798.794.589.398.893.9
Dinomaly99.799.995.098.998.995.192.099.195.1
RD4AD [10]98.597.893.996.090.170.987.1n/a93.8
PatchCore [45]99.198.193.594.798.591.889.4n/a91.5
SimpleNet [34]99.698.190.097.198.290.788.5n/a84.6
" + }, + { + "type": "table_caption", + "bbox": [ + 0.09, + 0.232, + 0.907, + 0.261 + ], + "angle": 0, + "content": "Table 3. Ablations of Dinomaly elements on MVTec-AD (%). NB: Noisy Bottleneck. LA: Linear Attention. LC: Loose Constraint (2 groups). LL: Loose Loss. As MVTec-AD has reached saturation, we also present the results on VisA (Table A4)." + }, + { + "type": "table", + "bbox": [ + 0.134, + 0.27, + 0.865, + 0.458 + ], + "angle": 0, + "content": "
NBLALCLLImage-levelPixel-level
AUROCAPF1-maxAUROCAPF1-maxAUPRO
98.4199.0997.4197.1862.9663.8292.95
✓99.0699.5498.3197.6266.2266.7093.71
✓98.5499.2197.6297.2062.9463.7393.09
✓98.3599.0497.4397.1061.0562.7392.60
✓99.0399.4598.1997.6264.1064.9693.34
✓✓99.2799.6298.6397.8567.3667.3394.16
✓✓99.5099.7298.8798.1468.1668.2494.23
✓✓✓99.5299.7398.9298.2068.2568.3494.17
✓✓✓99.5799.7899.0098.2067.9368.2194.50
✓✓✓✓99.6099.7899.0498.3569.2969.1794.79
" + }, + { + "type": "text", + "bbox": [ + 0.089, + 0.482, + 0.486, + 0.786 + ], + "angle": 0, + "content": "widely used MVTec-AD, Dinomaly produces image-level performance of 99.6/99.8/99.0 (\\%) and pixel-level performance of 98.4/69.3/69.2/94.8, outperforming previous SoTAs by 1.0/0.2/1.2 and 0.7/9.1/7.7/1.6. This result declares that the image-level performance on the MVTec-AD dataset is nearly saturated under the MUAD setting. On the popular VisA, Dinomaly achieves image-level performance of 98.7/98.9/96.2 and pixel-level performance of 98.7/53.2/55.7/94.5, outperforming previous SoTAs by 3.2/2.5/4.2 and 0.2/5.3/5.1/2.6. On the Real-IAD that contains 30 classes, each with 5 camera views, we produce image-level and pixel-level performance of 89.3/86.8/80.2 and 98.8/42.8/47.1/93.9, outperforming previous SoTAs by 3.0/2.2/3.2 and 0.3/4.9/5.4/3.4, indicating our scalability to extremely complex scenarios. Per-class performances and qualitative visualization are presented in Appendix E and F. We also produce superior results on other popular UAD benchmarks, i.e., MPDD [24], BTAD [38], and UniMedical [66], with I-AUROC of 97.2, 95.4, and 84.9, respectively, as shown in Table A13 in Appendix." + }, + { + "type": "title", + "bbox": [ + 0.09, + 0.8, + 0.468, + 0.817 + ], + "angle": 0, + "content": "4.3. Comparison to Class-Separated UAD SoTAs" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.825, + 0.485, + 0.903 + ], + "angle": 0, + "content": "Dinomaly is also compared with class-separated SoTAs, as shown in Table 2. Dinomaly under MUAD setting is comparable to conventional methods [10, 34, 45] that build individual models for each class. On MVTec-AD and VisA, multi-class Dinomaly (first row) is subjected to nearly no" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.482, + 0.907, + 0.559 + ], + "angle": 0, + "content": "performance drop compared to its class-separated counterpart (second row). On the complicated Real-IAD that involves more classes and views, multi-class Dinomaly suffers a moderate performance drop but is still comparable to class-separated SoTAs." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.572, + 0.667, + 0.589 + ], + "angle": 0, + "content": "4.4. Ablation Study" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.597, + 0.909, + 0.794 + ], + "angle": 0, + "content": "Overall Ablation. We conduct experiments to verify the effectiveness of the proposed elements, i.e., Noisy Bottleneck (NB), Linear Attention (LA), Loose Constraint (LC), and Loose Loss (LL). The already-powerful baseline (first row) is Dinomaly with noiseless MLP bottleneck, Softmax Attention, dense layer-to-layer supervision, and global cosine loss. This baseline is very similar to ViTAD [64] and the ViT version of RD4AD [10]. Results on MVtec-AD and VisA are shown in Table 3 and Table A4, respectively. NB and LL can directly contribute to the model performance. LA and LC boost the performance with the presence of NB. The use of LC is not solely beneficial because LC makes the reconstruction too easy without injected noise." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.796, + 0.909, + 0.903 + ], + "angle": 0, + "content": "Model Scalability. Previous works [10, 60, 65] reported that anomaly detection networks do not follow the \"scaling law\". For example, RD4AD [10] found WideResNet50 better than WideResNet101 as the encoder backbone. ViTAD [65] found ViT-Small better than ViT-Base. On the contrary, as shown in Table 4, the performance of the proposed Dinomaly benefits from scaling. Dinomaly equipped with ViT" + }, + { + "type": "page_number", + "bbox": [ + 0.479, + 0.945, + 0.518, + 0.958 + ], + "angle": 0, + "content": "20411" + } + ], + [ + { + "type": "table_caption", + "bbox": [ + 0.09, + 0.089, + 0.908, + 0.119 + ], + "angle": 0, + "content": "Table 4. Scaling of ViT model sizes on MVtec-AD (%). Im/s (Throughoutput, image per second) is measured on NVIDIA RTX3090 with batch size=16. Results on VisA and Real-IAD are shown in Table A3. †:default." + }, + { + "type": "table", + "bbox": [ + 0.134, + 0.129, + 0.865, + 0.219 + ], + "angle": 0, + "content": "
Arch.ParamsMACsIm/sImage-levelPixel-level
AUROCAPF1-maxAUROCAPF1-maxAUPRO
ViT-Small37.4M26.3G153.699.2699.6798.7298.0768.2967.7894.36
ViT-Base†148.0M104.7G58.199.6099.7899.0498.3569.2969.1794.79
ViT-Large275.3M413.5G24.299.7799.9299.4598.5470.5370.0495.09
" + }, + { + "type": "table_caption", + "bbox": [ + 0.091, + 0.239, + 0.483, + 0.27 + ], + "angle": 0, + "content": "Table 5. Scaling input size on MVTec-AD (\\%). †: default. Compared methods yield degradation when increasing input size." + }, + { + "type": "table", + "bbox": [ + 0.103, + 0.279, + 0.472, + 0.438 + ], + "angle": 0, + "content": "
MethodInput SizeImage-LevelPixel-Level
RD4AD2562†94.6/96.5/96.196.1/48.6/53.8/91.1
320293.2/96.9/95.695.7/55.1/57.5/91.1
384291.9/96.2/95.094.9/52.1/55.3/90.8
ReContrast2562†98.3/99.4/97.697.1/60.2/61.5/93.2
320298.2/99.2/97.596.8/61.8/62.6/93.3
384295.2/98.0/96.496.5/57.7/59.5/92.6
Dinomaly280299.6/99.8/99.398.2/65.2/66.3/93.6
336299.6/99.8/99.298.3/67.2/67.8/94.2
3922†99.6/99.8/99.098.4/69.3/69.2/94.8
" + }, + { + "type": "text", + "bbox": [ + 0.089, + 0.462, + 0.483, + 0.599 + ], + "angle": 0, + "content": "Small has already produced state-of-the-art results. ViT-Large further boosts Dinomaly to an unprecedented higher record. This scalability enables users to choose an appropriate model size based on the computational resources available in their specific scenario. A comparison of computational costs with other methods is presented in Table A11. In addition, training schedule can also be scaled up for even better performance without increasing inference costs, as demonstrated in Figure 1 (Dinomaly\\(\\uparrow\\))." + }, + { + "type": "text", + "bbox": [ + 0.089, + 0.599, + 0.483, + 0.75 + ], + "angle": 0, + "content": "Input Scalability. Though it seems unfair to compare Dinomaly with previous works that take smaller images as input, we contend that increasing their input size not only fails to benefit but actively undermines their performance, especially for image-level detection performance, as shown in Table 5. Therefore, we follow the common comparison strategy based on \"optimum vs. optimum\". On the contrary, Dinomaly enjoys scaling input size for anomaly localization, while still producing SoTA performance given smaller images. Details are presented in Table A2 in Appendix." + }, + { + "type": "text", + "bbox": [ + 0.089, + 0.75, + 0.483, + 0.902 + ], + "angle": 0, + "content": "ViT Foundations. We conduct extensive experiments to investigate the impact of diverse pre-trained ViT foundations, including DeiT [50], MAE [19], D-iGPT [44], MOCOV3 [6], DINO [4], iBot [69], DINOv2 [39], and DINOv2-R [7]. As shown in Figure 5, Dinomaly is robust to the choice of backbone. Almost all foundation models can produce SoTA-level results with image-level AUROC higher than \\(98\\%\\). The only notable exception is MAE, which, without fine-tuning, was reported to be less effective across various unsupervised tasks, e.g., kNN and linear-" + }, + { + "type": "image", + "bbox": [ + 0.521, + 0.241, + 0.898, + 0.386 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.513, + 0.398, + 0.908, + 0.442 + ], + "angle": 0, + "content": "Figure 5. Image-level AUROC of Dinomaly equipped with various ViT foundations, and their linear-probing accuracy on ImageNet. MIM: Masked Image Modeling. CL: Contrastive Learning." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.472, + 0.907, + 0.608 + ], + "angle": 0, + "content": "probing [39]. The optimal input size varies because the these backbones are pre-trained on different resolutions. Interestingly, we found the anomaly detection performance to be strongly correlated with the accuracy of ImageNet linear-probing (freeze backbone & only tune linear classifier) of the foundation model, suggesting the possibility of further improvement by simply adopting a future foundation model. Detailed results and analysis are presented in Appendix and Table A1." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.61, + 0.906, + 0.673 + ], + "angle": 0, + "content": "Additional experiments and results are detailed in the Appendix C, encompassing evaluations of various pretrained foundations, ablation studies of each components, hyperparameter optimization, and other in-depth analyses." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.692, + 0.634, + 0.708 + ], + "angle": 0, + "content": "5. Conclusion" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.72, + 0.907, + 0.901 + ], + "angle": 0, + "content": "Dinomaly, a minimalistic UAD framework, is proposed to address the under-performed MUAD models in this paper. We present four key elements in Dinomaly, i.e., Foundation Transformer, Noisy MLP Bottleneck, Linear Attention, and Loose Reconstruction, that can boost the performance under the challenging MUAD setting without fancy modules and tricks. Extensive experiments on MVTec AD, VisA, and Real-IAD demonstrate our superiority over previous model-unified multi-class models and even recent class-separated models, indicating the feasibility of implementing a unified model in complicated scenarios free of severe performance degradation." + }, + { + "type": "page_number", + "bbox": [ + 0.479, + 0.945, + 0.521, + 0.958 + ], + "angle": 0, + "content": "20412" + } + ], + [ + { + "type": "title", + "bbox": [ + 0.092, + 0.091, + 0.251, + 0.108 + ], + "angle": 0, + "content": "Acknowledgments" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.115, + 0.484, + 0.207 + ], + "angle": 0, + "content": "The authors acknowledge supports from National Natural Science Foundation of China (U22A2051, 82027807), National Key Research and Development Program of China (2022YFC2405200), Tsinghua-Foshan Innovation Special Fund (2021THFS0104), and Institute for Intelligent Healthcare, Tsinghua University (2022ZLB001)." + }, + { + "type": "title", + "bbox": [ + 0.093, + 0.218, + 0.188, + 0.234 + ], + "angle": 0, + "content": "References" + }, + { + "type": "ref_text", + "bbox": [ + 0.101, + 0.244, + 0.484, + 0.327 + ], + "angle": 0, + "content": "[1] Samet Akcay, Amir Atapour-Abarghouei, and Toby P Breckon. Ganomaly: Semi-supervised anomaly detection via adversarial training. In Computer Vision-ACCV 2018: 14th Asian Conference on Computer Vision, Perth, Australia, December 2-6, 2018, Revised Selected Papers, Part III 14, pages 622-637. Springer, 2019. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.328, + 0.483, + 0.383 + ], + "angle": 0, + "content": "[2] Paul Bergmann, Sindy Löwe, Michael Fauser, David Sattlegger, and Carsten Steger. Improving unsupervised defect segmentation by applying structural similarity to autoencoders. arXiv preprint arXiv:1807.02011, 2018. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.384, + 0.483, + 0.453 + ], + "angle": 0, + "content": "[3] Paul Bergmann, Michael Fauser, David Sattlegger, and Carsten Steger. Mvtec ad-a comprehensive real-world dataset for unsupervised anomaly detection. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 9592-9600, 2019. 1, 2, 6, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.454, + 0.482, + 0.523 + ], + "angle": 0, + "content": "[4] Mathilde Caron, Hugo Touvron, Ishan Misra, Hervé Jégou, Julien Mairal, Piotr Bojanowski, and Armand Joulin. Emerging properties in self-supervised vision transformers. In Proceedings of the IEEE/CVF international conference on computer vision, pages 9650-9660, 2021. 3, 8, 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.524, + 0.483, + 0.579 + ], + "angle": 0, + "content": "[5] Xinlei Chen and Kaiming He. Exploring simple siamese representation learning. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 15750-15758, 2021. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.58, + 0.482, + 0.635 + ], + "angle": 0, + "content": "[6] Xinlei Chen, Saining Xie, and Kaiming He. An empirical study of training self-supervised vision transformers. In Proceedings of the IEEE/CVF international conference on computer vision, pages 9640-9649, 2021. 3, 8, 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.636, + 0.482, + 0.678 + ], + "angle": 0, + "content": "[7] Timothee Darcet, Maxime Oquab, Julien Mairal, and Piotr Bojanowski. Vision transformers need registers. arXiv preprint arXiv:2309.16588, 2023. 3, 6, 8, 1, 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.678, + 0.482, + 0.747 + ], + "angle": 0, + "content": "[8] Thomas Defard, Aleksandr Setkov, Angelique Loesch, and Romaric Audigier. Padim: a patch distribution modeling framework for anomaly detection and localization. In International Conference on Pattern Recognition, pages 475-489. Springer, 2021. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.748, + 0.483, + 0.789 + ], + "angle": 0, + "content": "[9] David Dehaene and Pierre Eline. Anomaly localization by modeling perceptual features. arXiv preprint arXiv:2008.05369, 2020. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.096, + 0.79, + 0.484, + 0.845 + ], + "angle": 0, + "content": "[10] Hanqiu Deng and Xingyu Li. Anomaly detection via reverse distillation from one-class embedding. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 9737-9746, 2022. 5, 6, 7, 1, 8, 9, 10" + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.846, + 0.483, + 0.901 + ], + "angle": 0, + "content": "[11] Jia Deng, Wei Dong, Richard Socher, Li-Jia Li, Kai Li, and Li Fei-Fei. Imagenet: A large-scale hierarchical image database. In IEEE Conference on Computer Vision and Pattern Recognition, pages 248–255, 2009. 2" + }, + { + "type": "list", + "bbox": [ + 0.094, + 0.244, + 0.484, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.093, + 0.906, + 0.175 + ], + "angle": 0, + "content": "[12] Alexey Dosovitskiy, Lucas Beyer, Alexander Kolesnikov, Dirk Weissenborn, Xiaohua Zhai, Thomas Unterthiner, Mostafa Dehghani, Matthias Minderer, Georg Heigold, Sylvain Gelly, et al. An image is worth 16x16 words: Transformers for image recognition at scale. arXiv preprint arXiv:2010.11929, 2020. 2, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.177, + 0.905, + 0.232 + ], + "angle": 0, + "content": "[13] Jia Guo, Shuai Lu, Lize Jia, Weihang Zhang, and Huiqi Li. Encoder-decoder contrast for unsupervised anomaly detection in medical images. IEEE Transactions on Medical Imaging, 2023. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.233, + 0.905, + 0.3 + ], + "angle": 0, + "content": "[14] Jia Guo, Shuai Lu, Lize Jia, Weihang Zhang, and Huiqi Li. Recontrast: Domain-specific anomaly detection via contrastive reconstruction. In Advances in Neural Information Processing Systems (NeurIPS), pages 10721-10740, 2023. 2, 3, 5, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.302, + 0.905, + 0.37 + ], + "angle": 0, + "content": "[15] Dongchen Han, Xuran Pan, Yizeng Han, Shiji Song, and Gao Huang. Flatten transformer: Vision transformer using focused linear attention. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 5961-5971, 2023. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.372, + 0.905, + 0.439 + ], + "angle": 0, + "content": "[16] Haoyang He, Jiangning Zhang, Hongxu Chen, Xuhai Chen, Zhishan Li, Xu Chen, Yabiao Wang, Chengjie Wang, and Lei Xie. Diad: A diffusion-based framework for multi-class anomaly detection. arXiv preprint arXiv:2312.06607, 2023. 1, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.442, + 0.905, + 0.522 + ], + "angle": 0, + "content": "[17] Haoyang He, Yuhu Bai, Jiangning Zhang, Qingdong He, Hongxu Chen, Zhenye Gan, Chengjie Wang, Xiangtai Li, Guanzhong Tian, and Lei Xie. Mambaad: Exploring state space models for multi-class unsupervised anomaly detection. arXiv preprint arXiv:2404.06564, 2024. 1, 3, 6, 7, 8, 10" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.525, + 0.905, + 0.594 + ], + "angle": 0, + "content": "[18] Haoyang He, Jiangning Zhang, Hongxu Chen, Xuhai Chen, Zhishan Li, Xu Chen, Yabiao Wang, Chengjie Wang, and Lei Xie. A diffusion-based framework for multi-class anomaly detection. In Proceedings of the AAAI Conference on Artificial Intelligence, pages 8472-8480, 2024. 2, 6, 7, 8, 9, 10" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.596, + 0.905, + 0.662 + ], + "angle": 0, + "content": "[19] Kaiming He, Xinlei Chen, Saining Xie, Yanghao Li, Piotr Dollar, and Ross Girshick. Masked autoencoders are scalable vision learners. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 16000-16009, 2022. 3, 8, 2, 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.664, + 0.905, + 0.705 + ], + "angle": 0, + "content": "[20] Geoffrey Hinton, Oriol Vinyals, and Jeff Dean. Distilling the knowledge in a neural network. arXiv preprint arXiv:1503.02531, 2015. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.707, + 0.905, + 0.761 + ], + "angle": 0, + "content": "[21] Geoffrey E Hinton, Nitish Srivastava, Alex Krizhevsky, Ilya Sutskever, and Ruslan R Salakhutdinov. Improving neural networks by preventing co-adaptation of feature detectors. arXiv preprint arXiv:1207.0580, 2012. 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.762, + 0.905, + 0.817 + ], + "angle": 0, + "content": "[22] Chaoqin Huang, Haoyan Guan, Aofan Jiang, Ya Zhang, Michael Spratling, and Yan-Feng Wang. Registration based few-shot anomaly detection. In European Conference on Computer Vision, pages 303-319. Springer, 2022. 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.818, + 0.905, + 0.873 + ], + "angle": 0, + "content": "[23] Jongheon Jeong, Yang Zou, Taewan Kim, Dongqing Zhang, Avinash Ravichandran, and Onkar Dabeer. Winclip: Zero/few-shot anomaly classification and segmentation. arXiv preprint arXiv:2303.14814, 2023.7" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.874, + 0.905, + 0.901 + ], + "angle": 0, + "content": "[24] Stepan Jezek, Martin Jonak, Radim Burget, Pavel Dvorak, and Milos Skotak. Deep learning-based defect detection of" + }, + { + "type": "list", + "bbox": [ + 0.094, + 0.093, + 0.906, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.479, + 0.945, + 0.521, + 0.957 + ], + "angle": 0, + "content": "20413" + } + ], + [ + { + "type": "ref_text", + "bbox": [ + 0.125, + 0.092, + 0.484, + 0.147 + ], + "angle": 0, + "content": "metal parts: evaluating current methods in complex conditions. In 2021 13th International Congress on Ultra Modern Telecommunications and Control Systems and Workshops (ICUMT), pages 66-71. IEEE, 2021. 7, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.149, + 0.484, + 0.216 + ], + "angle": 0, + "content": "[25] Xi Jiang, Jianlin Liu, Jinbao Wang, Qiang Nie, Kai Wu, Yong Liu, Chengjie Wang, and Feng Zheng. Softpatch: Unsupervised anomaly detection with noisy data. Advances in Neural Information Processing Systems, 35:15433-15445, 2022. 6, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.219, + 0.483, + 0.285 + ], + "angle": 0, + "content": "[26] Angelos Katharopoulos, Apoorv Vyas, Nikolaos Pappas, and François Fleuret. Transformers are rnns: Fast autoregressive transformers with linear attention. In International conference on machine learning, pages 5156-5165. PMLR, 2020. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.288, + 0.483, + 0.342 + ], + "angle": 0, + "content": "[27] Sungwook Lee, Seunghyun Lee, and Byung Cheol Song. Cfa: Coupled-hypersphere-based feature adaptation for target-oriented anomaly localization. IEEE Access, 10: 78446-78454, 2022. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.344, + 0.483, + 0.384 + ], + "angle": 0, + "content": "[28] Yujin Lee, Harin Lim, and Hyunsoo Yoon. Selfformaly: Towards task-agnostic unified anomaly detection. arXiv preprint arXiv:2307.12540, 2023. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.386, + 0.483, + 0.454 + ], + "angle": 0, + "content": "[29] Chun-Liang Li, Kihyuk Sohn, Jinsung Yoon, and Tomas Pfister. Cutpaste: Self-supervised learning for anomaly detection and localization. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 9664-9674, 2021. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.456, + 0.483, + 0.522 + ], + "angle": 0, + "content": "[30] Chen Liang, Jiahui Yu, Ming-Hsuan Yang, Matthew Brown, Yin Cui, Tuo Zhao, Boqing Gong, and Tianyi Zhou. Modulewise adaptive distillation for multimodality foundation models. Advances in Neural Information Processing Systems, 36, 2024. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.525, + 0.483, + 0.565 + ], + "angle": 0, + "content": "[31] Jiangqi Liu and Feng Wang. mixed attention auto encoder for multi-class industrial anomaly detection. arXiv preprint arXiv:2309.12700, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.567, + 0.483, + 0.635 + ], + "angle": 0, + "content": "[32] Wenqian Liu, Runze Li, Meng Zheng, Srikrishna Karanam, Ziyan Wu, Bir Bhanu, Richard J Radke, and Octavia Camps. Towards visually explaining variational autoencoders. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 8642-8651, 2020. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.637, + 0.483, + 0.704 + ], + "angle": 0, + "content": "[33] Ze Liu, Yutong Lin, Yue Cao, Han Hu, Yixuan Wei, Zheng Zhang, Stephen Lin, and Baining Guo. Swin transformer: Hierarchical vision transformer using shifted windows. In IEEE/CVF International Conference on Computer Vision, pages 10012-10022, 2021. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.707, + 0.483, + 0.76 + ], + "angle": 0, + "content": "[34] Zhikang Liu, Yiming Zhou, Yuansheng Xu, and Zilei Wang. Simplenet: A simple network for image anomaly detection and localization. arXiv preprint arXiv:2303.15140, 2023. 6, 7, 1, 8, 9, 10" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.762, + 0.483, + 0.789 + ], + "angle": 0, + "content": "[35] Ilya Loshchilov and Frank Hutter. Decoupled weight decay regularization. arXiv preprint arXiv:1711.05101, 2017. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.79, + 0.483, + 0.844 + ], + "angle": 0, + "content": "[36] Ruiying Lu, YuJie Wu, Long Tian, Dongsheng Wang, Bo Chen, Xiyang Liu, and Ruimin Hu. Hierarchical vector quantized transformer for multi-class unsupervised anomaly detection. arXiv preprint arXiv:2310.14228, 2023. 1, 2, 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.846, + 0.483, + 0.899 + ], + "angle": 0, + "content": "[37] Amira Ben Mabrouk and Ezzeddine Zagrouba. Abnormal behavior recognition for intelligent video surveillance systems: A review. Expert Systems with Applications, 91:480-491, 2018. 1" + }, + { + "type": "list", + "bbox": [ + 0.093, + 0.092, + 0.484, + 0.899 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.092, + 0.905, + 0.161 + ], + "angle": 0, + "content": "[38] Pankaj Mishra, Riccardo Verk, Daniele Fornasier, Claudio Piciarelli, and Gian Luca Foresti. Vt-adl: A vision transformer network for image anomaly detection and localization. In 2021 IEEE 30th International Symposium on Industrial Electronics (ISIE), pages 01–06. IEEE, 2021. 7, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.163, + 0.905, + 0.231 + ], + "angle": 0, + "content": "[39] Maxime Oquab, Timothee Darcet, Theo Moutakanni, Huy Vo, Marc Szafraniec, Vasil Khalidov, Pierre Fernandez, Daniel Haziza, Francisco Massa, Alaaeldin El-Nouby, et al. Dinov2: Learning robust visual features without supervision. arXiv preprint arXiv:2304.07193, 2023. 3, 8, 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.233, + 0.905, + 0.273 + ], + "angle": 0, + "content": "[40] Z Peng, L Dong, H Bao, Q Ye, and F Wei. Beit v2: Masked image modeling with vector-quantized visual tokenizers. arXiv preprint arXiv:2208.06366, 2022. 3, 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.274, + 0.905, + 0.314 + ], + "angle": 0, + "content": "[41] Sashank J Reddi, Satyen Kale, and Sanjiv Kumar. On the convergence of adam and beyond. arXiv preprint arXiv:1904.09237, 2019. 6, 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.316, + 0.905, + 0.383 + ], + "angle": 0, + "content": "[42] Tal Reiss, Niv Cohen, Liron Bergman, and Yedid Hoshen. Panda: Adapting pretrained features for anomaly detection and segmentation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 2806-2814, 2021. 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.385, + 0.905, + 0.44 + ], + "angle": 0, + "content": "[43] Tal Reiss, Niv Cohen, Eliahu Horwitz, Ron Abutbul, and Yedid Hoshen. Anomaly detection requires better representations. In European Conference on Computer Vision, pages 56-68. Springer, 2022. 3, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.442, + 0.905, + 0.495 + ], + "angle": 0, + "content": "[44] Sucheng Ren, Zeyu Wang, Hongru Zhu, Junfei Xiao, Alan Yuille, and Cihang Xie. Rejuvenating image-gpt as strong visual representation learners. arXiv preprint arXiv:2312.02147, 2023.8, 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.497, + 0.907, + 0.566 + ], + "angle": 0, + "content": "[45] Karsten Roth, Latha Pemula, Joaquin Zepeda, Bernhard Schölkopf, Thomas Brox, and Peter Gehler. Towards total recall in industrial anomaly detection. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 14318-14328, 2022. 7, 1, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.567, + 0.905, + 0.648 + ], + "angle": 0, + "content": "[46] Mohammadreza Salehi, Niousha Sadjadi, Soroosh Baselizadeh, Mohammad H Rohban, and Hamid R Rabiee. Multiresolution knowledge distillation for anomaly detection. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 14902-14912, 2021. 5, 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.65, + 0.905, + 0.705 + ], + "angle": 0, + "content": "[47] Thomas Schlegl, Philipp Seebock, Sebastian M Waldstein, Georg Langs, and Ursula Schmidt-Erfurth. f-anogan: Fast unsupervised anomaly detection with generative adversarial networks. Medical image analysis, 54:30-44, 2019. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.707, + 0.905, + 0.773 + ], + "angle": 0, + "content": "[48] Zhuoran Shen, Mingyuan Zhang, Haiyu Zhao, Shuai Yi, and Hongsheng Li. Efficient attention: Attention with linear complexities. In Proceedings of the IEEE/CVF winter conference on applications of computer vision, pages 3531-3539, 2021. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.776, + 0.905, + 0.843 + ], + "angle": 0, + "content": "[49] Shelly Sheynin, Sagie Benaim, and Lior Wolf. A hierarchical transformation-discriminating generative model for few shot anomaly detection. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 8495-8504, 2021. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.845, + 0.905, + 0.9 + ], + "angle": 0, + "content": "[50] Hugo Touvron, Matthieu Cord, Matthijs Douze, Francisco Massa, Alexandre Sablayrolles, and Herv'e J'egou. Training data-efficient image transformers & distillation through attention. arXiv preprint arXiv:2012.12877, 2021. 8, 2" + }, + { + "type": "list", + "bbox": [ + 0.517, + 0.092, + 0.907, + 0.9 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.48, + 0.945, + 0.521, + 0.957 + ], + "angle": 0, + "content": "20414" + } + ], + [ + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.092, + 0.482, + 0.148 + ], + "angle": 0, + "content": "[51] Ashish Vaswani, Noam Shazeer, Niki Parmar, Jakob Uszkoreit, Llion Jones, Aidan N Gomez, Lukasz Kaiser, and Illia Polosukhin. Attention is all you need. In Advances in Neural Information Processing Systems, pages 5998-6008, 2017. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.149, + 0.482, + 0.219 + ], + "angle": 0, + "content": "[52] Pascal Vincent, Hugo Larochelle, Yoshua Bengio, and Pierre-Antoine Manzagol. Extracting and composing robust features with denoising autoencoders. In Proceedings of the 25th international conference on Machine learning, pages 1096-1103, 2008. 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.221, + 0.482, + 0.29 + ], + "angle": 0, + "content": "[53] Pascal Vincent, Hugo Larochelle, Isabelle Lajoie, Yoshua Bengio, and Pierre-Antoine Manzagol. Stacked denoising autoencoders: Learning useful representations in a deep network with a local denoising criterion. In Journal of Machine Learning Research, pages 3371-3408, 2010. 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.291, + 0.482, + 0.36 + ], + "angle": 0, + "content": "[54] Chengjie Wang, Wenbing Zhu, Bin-Bin Gao, Zhenye Gan, Jianning Zhang, Zhihao Gu, Shuguang Qian, Mingang Chen, and Lizhuang Ma. Real-iad: A real-world multi-view dataset for benchmarking versatile industrial anomaly detection. arXiv preprint arXiv:2403.12580, 2024. 2, 6, 7, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.362, + 0.482, + 0.417 + ], + "angle": 0, + "content": "[55] Guodong Wang, Shumin Han, Errui Ding, and Di Huang. Student-teacher feature pyramid matching for anomaly detection. In The British Machine Vision Conference (BMVC), 2021. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.419, + 0.482, + 0.487 + ], + "angle": 0, + "content": "[56] Mitchell Wortsman, Tim Dettmers, Luke Zettlemoyer, Ari Morcos, Ali Farhadi, and Ludwig Schmidt. Stable and low-precision training for large-scale vision-language models. Advances in Neural Information Processing Systems, 36: 10271–10298, 2023. 6, 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.489, + 0.482, + 0.559 + ], + "angle": 0, + "content": "[57] Zhenda Xie, Zheng Zhang, Yue Cao, Yutong Lin, Jianmin Bao, Zhuliang Yao, Qi Dai, and Han Hu. Simmim: A simple framework for masked image modeling. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 9653-9663, 2022. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.561, + 0.482, + 0.602 + ], + "angle": 0, + "content": "[58] Jie Yang, Yong Shi, and Zhiquan Qi. Dfr: Deep feature reconstruction for unsupervised anomaly segmentation. arXiv preprint arXiv:2012.07122, 2020. 5, 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.604, + 0.482, + 0.671 + ], + "angle": 0, + "content": "[59] Haonan Yin, Guanlong Jiao, Qianhui Wu, Borne F Karlsson, Biqing Huang, and Chin Yew Lin. Lafite: Latent diffusion model with feature editing for unsupervised multi-class anomaly detection. arXiv preprint arXiv:2307.08059, 2023.1, 2, 3, 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.675, + 0.482, + 0.729 + ], + "angle": 0, + "content": "[60] Zhiyuan You, Lei Cui, Yujun Shen, Kai Yang, Xin Lu, Yu Zheng, and Xinyi Le. A unified model for multi-class anomaly detection. arXiv preprint arXiv:2206.03687, 2022. 1, 2, 3, 4, 5, 6, 7, 8, 9, 10" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.732, + 0.482, + 0.787 + ], + "angle": 0, + "content": "[61] Weihao Yu, Chenyang Si, Pan Zhou, Mi Luo, Yichen Zhou, Jiashi Feng, Shuicheng Yan, and Xinchao Wang. Metaformer baselines for vision. IEEE Transactions on Pattern Analysis and Machine Intelligence, 2023. 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.789, + 0.482, + 0.856 + ], + "angle": 0, + "content": "[62] Vitjan Zavrtanik, Matej Kristan, and Danijel Skoc aj. Draem-a discriminatively trained reconstruction embedding for surface anomaly detection. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 8330-8339, 2021. 4, 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.859, + 0.482, + 0.9 + ], + "angle": 0, + "content": "[63] Vitjan Zavrtanik, Matej Kristan, and Danijel Skocaj. Reconstruction by inpainting for visual anomaly detection. Pattern Recognition, 112:107706, 2021. 1" + }, + { + "type": "list", + "bbox": [ + 0.093, + 0.092, + 0.482, + 0.9 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.092, + 0.905, + 0.161 + ], + "angle": 0, + "content": "[64] Dingwen Zhang, Guohai Huang, Qiang Zhang, Jungong Han, Junwei Han, Yizhou Wang, and Yizhou Yu. Exploring task structure for brain tumor segmentation from multimodality mr images. IEEE Transactions on Image Processing, 29:9032-9043, 2020. 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.164, + 0.905, + 0.232 + ], + "angle": 0, + "content": "[65] Jiangning Zhang, Xuhai Chen, Yabiao Wang, Chengjie Wang, Yong Liu, Xiangtai Li, Ming-Hsuan Yang, and Dacheng Tao. Exploring plain vit reconstruction for multiclass unsupervised anomaly detection. arXiv preprint arXiv:2312.07495, 2023. 3, 5, 6, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.234, + 0.905, + 0.302 + ], + "angle": 0, + "content": "[66] Jiangning Zhang, Haoyang He, Zhenye Gan, Qingdong He, Yuxuan Cai, Zhucun Xue, Yabiao Wang, Chengjie Wang, Lei Xie, and Yong Liu. Ader: A comprehensive benchmark for multi-class visual anomaly detection. arXiv preprint arXiv:2406.03262, 2024. 7, 2, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.304, + 0.905, + 0.373 + ], + "angle": 0, + "content": "[67] Xuan Zhang, Shiyu Li, Xi Li, Ping Huang, Jiulong Shan, and Ting Chen. Destseg: Segmentation guided denoising student-teacher for anomaly detection. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 3914–3923, 2023. 4, 6, 1, 7, 8, 9, 10" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.375, + 0.905, + 0.43 + ], + "angle": 0, + "content": "[68] Ying Zhao. Omnial: A unified cnn framework for unsupervised anomaly localization. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 3924-3933, 2023. 1, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.432, + 0.905, + 0.485 + ], + "angle": 0, + "content": "[69] Jinghao Zhou, Chen Wei, Huiyu Wang, Wei Shen, Cihang Xie, Alan Yuille, and Tao Kong. ibot: Image bert pre-training with online tokenizer. arXiv preprint arXiv:2111.07832, 2021. 3, 8, 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.488, + 0.905, + 0.557 + ], + "angle": 0, + "content": "[70] Yang Zou, Jongheon Jeong, Latha Pemula, Dongqing Zhang, and Onkar Dabeer. Spot-the-difference self-supervised pretraining for anomaly detection and segmentation. In European Conference on Computer Vision, pages 392-408. Springer, 2022. 2, 6, 7, 3" + }, + { + "type": "list", + "bbox": [ + 0.517, + 0.092, + 0.905, + 0.557 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.48, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "20415" + } + ] +] \ No newline at end of file diff --git a/2025/Dinomaly_ The Less Is More Philosophy in Multi-Class Unsupervised Anomaly Detection/269a00de-90dc-4a5c-8e52-737661810fb4_origin.pdf b/2025/Dinomaly_ The Less Is More Philosophy in Multi-Class Unsupervised Anomaly Detection/269a00de-90dc-4a5c-8e52-737661810fb4_origin.pdf new file mode 100644 index 0000000000000000000000000000000000000000..2248cf3bdebc9bc1a6d38b930b9a9c48edb990cb --- /dev/null +++ b/2025/Dinomaly_ The Less Is More Philosophy in Multi-Class Unsupervised Anomaly Detection/269a00de-90dc-4a5c-8e52-737661810fb4_origin.pdf @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:52c736ae4dc771d06bb3d05477763fdaf97532e01f9cf1933d490f18ee6a25b8 +size 3988662 diff --git a/2025/Dinomaly_ The Less Is More Philosophy in Multi-Class Unsupervised Anomaly Detection/full.md b/2025/Dinomaly_ The Less Is More Philosophy in Multi-Class Unsupervised Anomaly Detection/full.md new file mode 100644 index 0000000000000000000000000000000000000000..2ad45b3999ca140a12aea3c0b2ec723fde3a9941 --- /dev/null +++ b/2025/Dinomaly_ The Less Is More Philosophy in Multi-Class Unsupervised Anomaly Detection/full.md @@ -0,0 +1,335 @@ +# Dinomaly: The Less Is More Philosophy in Multi-Class Unsupervised Anomaly Detection + +Jia Guo $^{1}$ Shuai Lu $^{2}$ Weihang Zhang $^{2\dagger}$ Fang Chen $^{3}$ Huiqi Li $^{2\dagger}$ Hongen Liao $^{1,3\boxtimes}$ + +$^{1}$ School of Biomedical Engineering, Tsinghua University, Beijing, China $^{2}$ Beijing Institute of Technology, Beijing, China + +$^{3}$ School of Biomedical Engineering, Shanghai Jiao Tong University, Shanghai, China + +guojia.jeremy@gmail.com lushuaie@163.com zhangweihang@bit.edu.cn chen-fang@sjtu.edu.cn huiqili@bit.edu.cn liao@tsinghua.edu.cn + +# Abstract + +Recent studies highlighted a practical setting of unsupervised anomaly detection (UAD) that builds a unified model for multi-class images. Despite various advancements addressing this challenging task, the detection performance under the multi-class setting still lags far behind state-of-the-art class-separated models. Our research aims to bridge this substantial performance gap. In this paper, we present Dinomaly, a minimalist reconstruction-based anomaly detection framework that harnesses pure Transformer architectures without relying on complex designs, additional modules, or specialized tricks. Given this powerful framework consisting of only Attentions and MLPs, we found four simple components that are essential to multi-class anomaly detection: (1) Scalable foundation Transformers that extracts universal and discriminative features, (2) Noisy Bottleneck where pre-existing Dropouts do all the noise injection tricks, (3) Linear Attention that naturally cannot focus, and (4) Loose Reconstruction that does not force layer-to-layer and point-by-point reconstruction. Extensive experiments are conducted across popular anomaly detection benchmarks including MVtec-AD, VisA, Real-IAD, etc. Our proposed Dinomaly achieves impressive image-level AUROC of $99.6\%$ , $98.7\%$ , and $89.3\%$ on the three datasets respectively, which is not only superior to state-of-the-art multi-class UAD methods, but also achieves the most advanced class-separated UAD records. Code is available at: https://github.com/guojiajeremy/Dinomaly + +# 1. Introduction + +Unsupervised anomaly detection (UAD) aims to detect abnormal patterns from normal images and further localize the anomalous regions. Because of the diversity of potential anomalies and their scarcity, this task is proposed to model the accessible training sets containing only normal samples as an unsupervised paradigm. UAD has a wide range of applications, e.g., industrial defect detection [3], medical disease screening [13], and video surveillance [37], addressing the difficulty of collecting and labeling all possible anomalies in these scenarios. + +Conventional works on UAD build a separate model for each object category, as shown in Figure 1(a). However, this one-class-one-model setting entails substantial storage overhead for saving models [60], especially when the application scenario necessitates a large number of object classes. For UAD methods, a compact boundary of normal patterns is vital to distinguish anomalies. Once the intra-normal patterns become exceedingly complicated due to various classes, the corresponding distribution becomes challenging to measure, consequently harming the detection performance. Recently, UniAD [60] and successive studies have been proposed to train a unified model for multi-class anomaly detection (MUAD), as shown in Figure 1(b). Under this setting, the "identity mapping" that directly copies the input as the output regardless of normal or anomaly harms the performance of conventional methods [60]. This phenomenon is caused by the diversity of multi-class normal patterns that drive the network to generalize on unseen patterns. + +Within two years, a number of methods have been proposed to address MUAD, such as neighbor-masked attention [60], synthetic anomalies [68], vector quantization [36], diffusion model [16, 59], and state space model (Mamba) [17]. However, there is still a non-negligible per + +![](images/e77de717e7226e870fe1250726262007dd1a4ff99ce23a99dd348328cf5b24f7.jpg) +(a) Class-Separated UAD Setting + +![](images/21af34168446ad566b0d66a8553a9bddb83a215a657a4e8c6b1775ed7512a181.jpg) +MVTec-AD (15 classes) + +![](images/752ff9fe62b375b33baafdeedcfd0b9ae29b5ad861e630b336cf9b22f9bea3ab.jpg) +VisA (12 classes) + +![](images/6de827691005c6a5c2758f19abd6cacf12119cbd3dd32c9c6d1fa1f13c2a2eae.jpg) +Real-IAD (30 classes) + +![](images/e357a9a888575f41ce9e2b5de7c3db2424bd3b9e759acf39d47a4e7ea73dcc37.jpg) +(b) +Model-Unified +Multi-Class UAD Setting + +![](images/bfa686fdd5f1898aaee9f58a6810b419098f70670743e14fdd62616ac3beeb8b.jpg) +(d) Scaling Properties of Dinomaly on MVTec-AD. $\dagger$ : default setting. + +![](images/7c9079b35c2f53586e2ae811f1f5e9bee8e28950af1f243868cd1825e5748932.jpg) +(c) Image-Level AUROC on 3 Benchmarks + +![](images/3a384fc4fb010232dc6e2160c574502a492a9d1cb6613d52a8df42eb04d61f13.jpg) +Figure 1. Setting, benchmarking, and scaling of Dinomaly. (a) Task setting of class-separated UAD. (b) Task setting of MUAD. (c) Comparison with previous SoTA methods on MVtec-AD [3], VisA [70], and Real-IAD [54]. (d) Scaling properties of Dinomaly. + +formance gap between the state-of-the-art (SoTA) MUAD methods and class-separated UAD methods, restricting the practicability of implementing unified models, as shown in Figure 1(c). In addition, previous methods employ modules and architectures delicately designed, which may not be straightforward, and consequently suffer from limited universality and ease-of-use [18, 36]. + +In this work, we aim to catch up with the performance of class-separated anomaly detection models using a multiclass unified model. We introduce Dinomaly, a minimalist reconstruction-based UAD framework built exclusively by pure Transformer blocks [51], specifically Self-Attentions and Multi-Layer Perceptrons (MLPs). To begin with, we empirically investigate the scaling law of self-supervised pre-trained Vision Transformers (ViT) [12] when serving as the feature encoders for extracting reconstruction objectives. Subsequently, we introduce three straightforward yet crucial elements to address the critical identity mapping phenomenon in MUAD contexts, without increasing complexity or computational burden. First, as an alternative to meticulously designed pseudo anomaly and feature noise, we propose to activate the built-in Dropout in an MLP to prevent the network from restoring both normal and anomalous patterns. Second, we propose to leverage the "side effect" of Linear Attention (a computation-efficient counterpart of Softmax Attention) that impedes focus on local regions, thus preventing the forwarding of identical information. Third, previous methods adopt layer-to-layer and region-by-region reconstruction schemes, distilling a decoder that can well mimic the encoder's behavior even for anomalous regions. Therefore, we propose to loosen the reconstruction constraints by grouping multiple layers as a + +whole and discarding well-reconstructed regions during optimization. + +To validate the effectiveness of the proposed Dinomaly under MUAD setting, we conduct extensive experiments on a number of widely used benchmarks, including MVTec AD [3] (15 classes), VisA [70] (12 classes), and RealIAD (30 classes). As shown in Figure 1, our base-size Dinomaly achieves unprecedented image-level AUROC of $99.6\%$ , $98.7\%$ , and $89.3\%$ on MVTec AD, VisA, and RealIAD, surpassing previous SoTA methods by a large margin. In addition, scalability is a key feature of Dinomaly. Further scaling up the model size maximizes performance to the fullest level of $99.8\%$ , $98.9\%$ , and $90.1\%$ , respectively; while scaling down parameters and input size can offer efficient solutions in computation-constrained scenarios. + +# 2. Related Work + +Multi-Class UAD. UniAD [60] first introduced multi-class anomaly detection, aiming to detect anomalies for different classes using a unified model. In this setting, conventional UAD methods often face the challenge of "identical shortcuts", where both anomaly-free and anomaly samples can be effectively recovered during inference [60]. It is believed that this phenomenon is caused by the diversity of multi-class normal patterns that drive the network to generalize on unseen patterns. This contradicts the fundamental assumption of epistemic methods. Many current researches focus on addressing this challenge [14, 31, 36, 59, 60]. UniAD [60] employs a neighbor-masked attention module and a feature-jitter strategy to mitigate these shortcuts. HVQ-Trans [36] proposes a vector quantization (VQ) Transformer model that induces large feature discrepancies + +![](images/17d3d0fa6c8cb95057e7196ca865d09815a7ac4253dda8c3a5682950e09bd3d4.jpg) +Figure 2. The framework of Dinomaly, built by simple and pure Transformer building blocks. + +for anomalies. LafitE [59] utilizes a latent diffusion model and introduces a feature editing strategy to alleviate this issue. DiAD [16] also employs diffusion models to address multi-class UAD settings. OmniAL [68] focuses on anomaly localization in the unified setting, preventing identical reconstruction by using synthesized pseudo anomalies. ReContrast [14] attempted to alleviate the identity mapping by cross-reconstruction between two encoders. ViTAD [5] abstracts a unified feature-reconstruction UAD framework and employ Transformer building blocks. MambaAD [17] explores the recently proposed State Space Model (SSM), Mamba, in the context of multi-class UAD. More related works of UAD are presented in Appendix A. + +# 3. Method + +# 3.1. Dinomaly Framework + +"What I cannot create, I do not understand" + +Richer Feynman + +The ability to recognize anomalies from what we know is an innate human capability, serving as a vital pathway for us to explore the world. Similarly, we construct a reconstruction-based framework that relies on the epistemic characteristic of artificial neural networks. Dinomaly consists of an encoder, a bottleneck, and a reconstruction decoder, as shown in Figure 2. Without loss of generality, a pretrained ViT network [12] with 12 Transformer layers is used as the encoder, extracting informative feature maps with different semantic scales. The bottleneck is a simple MLP (a.k.a. feed-forward network, FFN) that collects the feature representations of the encoder's 8 middle-level layers. The decoder is similar to the encoder, consisting of 8 Transformer layers. During training, the decoder learns to reconstruct the middle-level features of the encoder by maximizing the cosine similarity between feature maps. During inference, the decoder is expected to reconstruct normal regions of feature maps but fails for anomalous regions as it + +has never seen such samples. + +Foundation Transformers. Foundation models, especially ViTs [12, 33] pre-trained on large-scale datasets, serve as a basis and starting point for specific computer vision tasks. Such networks employ self-supervised learning schemes such as contrastive learning (MoCov3 [6], DINO [4]), masked image modeling (MAE [19], SimMIM [57], BEiT [40]), and their combination (iBOT [69], DINOv2 [39]), producing universal features suitable for image-level visual tasks and pixel-level visual tasks. + +Because of the lack of supervision in UAD, most advanced methods adopt pre-trained networks to extract discriminative features. Recent works [28, 43, 65] have preliminarily discovered the advantage of robust and universal features of self-supervised models over domain-specific ImageNet features in anomaly detection tasks. In this paper, we pioneer the investigation of scaling behaviors in UAD models through a systematic analysis of foundational ViTs, as briefed in Figure 1(d). Our comprehensive evaluation encompasses pre-training strategies (Figure 5), model sizes (Table 4), and input resolutions (Table 5), which are detailed in section 4.4. Considering the balance of detection performance and computational efficiency, we adopt ViT-Base/14 pretrained by DINoV2-Register [7] as the encoder of Dinomaly by default. + +# 3.2. Noisy Bottleneck. + +"Dropout is all you need." + +Prior studies [14, 60, 68] attribute the performance degradation of UAD methods trained on diverse multi-class samples to the "identity mapping" phenomenon; in this work, we reframe this issue as an "over-generalization" problem. Generalization ability is a merit of neural networks, allowing them to perform equally well on unseen test sets. However, generalization is not so wanted in the context of unsupervised anomaly detection that leverages the epistemic nature of neural networks. With the increas + +![](images/6d4a43b73593602e9e9d33bb1e07ed954c2eb370106d46e6efa33d5071ffbc4f.jpg) +Figure 3. Softmax Attention vs. Linear Attention. (a) Visualization of attention maps. (b) Attention distribution. + +![](images/8d125b0bc5327cd9d9792b71ec97f6b7783286bcb1c5a99b13be4c2f135ca734.jpg) + +ing diversity of images and their patterns due to multi-class UAD settings, the decoder can generalize its reconstruction ability to unseen anomalous samples, resulting in the failure of anomaly detection using reconstruction error. + +A direct solution for identity mapping is to shift "reconstruction" to "restoration". Specifically, instead of directly reconstructing the normal images or features given normal inputs, previous works propose to add perturbations as pseudo anomalies on input images [62, 67] or on encoder features [59, 60]; meanwhile, still let the decoder restore anomaly-free images or features, formulating a denoising-like framework. However, such methods employ heuristic and hand-crafted anomaly generation strategies that may not be universal across domains, datasets, and methods. In this work, we turn to leveraging the simple and elegant Dropout techniques. Since its introduction by Hinton et al. [21] in 2014 as a remedy for overfitting, Dropout has become a cornerstone element in neural architectures, including Transformers. In Dinomaly, we employ Dropout to randomly discard neural activations in an MLP bottleneck. Instead of alleviating overfitting, the role of Dropout in Dinomaly can be explained as pseudo feature anomaly that perturb normal representations, analogous to denoising auto-encoders [52, 53]. Without introducing any specific modules, this simple component inherently forces the decoder to restore normal features regardless of whether the test image contains anomalies, in turn, mitigating identical mapping. + +# 3.3. Unfocused Linear Attention. + +"One man's poison is another man's meat" + +Softmax Attention is the key mechanism of Transformers, allowing the model to attend to different parts of its + +input token sequence. Formally, given an input sequence $\mathbf{X} \in \mathbb{R}^{N \times d}$ with length $N$ , Attention first transforms it into three matrices: the query matrix $\mathbf{Q} \in \mathbb{R}^{N \times d}$ , the key matrix $\mathbf{K} \in \mathbb{R}^{N \times d}$ , and the value matrix $\mathbf{V} \in \mathbb{R}^{N \times d}$ : + +$$ +\mathbf {Q} = \mathbf {X} \mathbf {W} ^ {Q}, \mathbf {K} = \mathbf {X} \mathbf {W} ^ {K}, \mathbf {V} = \mathbf {X} \mathbf {W} ^ {V}, \tag {1} +$$ + +where $\mathbf{W}^Q, \mathbf{W}^K, \mathbf{W}^V \in \mathbb{R}^{d \times d}$ are learnable parameters. By computing the attention map by the query-key similarity, the output of Softmax Attention is given as: + +$$ +\operatorname {A t t e n t i o n} (\mathbf {Q}, \mathbf {K}, \mathbf {V}) = \operatorname {S o f t m a x} (\mathbf {Q K} ^ {T}) \mathbf {V}. \tag {2} +$$ + +Back to MUAD, previous methods [36, 60] suggest adopting Attentions instead of Convolutions because Convolutions can easily learn identical mappings. Nevertheless, both operations are in danger of forming identity mapping by over-concentrating on corresponding input locations for producing the outputs: + +$$ +\operatorname {C o n v} \operatorname {K e r n e l} = \left[ \begin{array}{l l l} 0 & 0 & 0 \\ 0 & 1 & 0 \\ 0 & 0 & 0 \end{array} \right], \quad \operatorname {A t t n} \operatorname {M a p} = \left[ \begin{array}{l l l} 1 & 0 & 0 \\ 0 & 1 & 0 \\ 0 & 0 & 1 \end{array} \right]. +$$ + +Is there any simple solution that prevents Attentions from attending to identical information? In Dinomaly, we turn to leverage the "unfocusing ability" of a type of Softmax-free Attention, i.e., Linear Attention. Linear Attention was proposed as a promising alternative to reduce the computation complexity of vanilla Softmax Attention + +![](images/a1dd9fa401691a462a4df3a84a2f8feedcaf54499e447957f60fb2c121c62204.jpg) +(a) layer-to-layer sparse + +![](images/b866424b103ff548ef2ecbe121963fe58d37f24999027884bb87e659c5bc10ba.jpg) +(b) layer-to-cat-layer + +![](images/b4bffa389b2c5780fbe925887470053347f3965c3cfab98e72337e6b9155bdfe.jpg) +(c) layer-to-layer dense + +![](images/2e691d240b522deeace347316f62ed4982cc6efcf42bc8289ce3948972a78386.jpg) +(d) group-to-group 1 group (Ours) + +![](images/0302c001c32f5d638e646b4c5b0696ae395e2e4637f8de43dad4b5752f96a3ad.jpg) +(e) group-to-group 2 group (Ours) + +![](images/d3bc1adf7c570b9343648410728143a6c2329294536bf9af7d61d3b9329c354d.jpg) +encoder feature + +![](images/4767d4991e7ce4e01a71213986c2b2383e2a08849f258d4d882d31d17534b15e.jpg) +decoder feature + +![](images/f33596ba1387f7c0a6e5d9120039352aced6c9d016bf983caf8f1150221ea210.jpg) +reconstruct + +![](images/c37867f60176995bbf10bf74c757ea2ab30dd85d912f29e92525a9a0297ad90e.jpg) +concatenate + +![](images/ac76c81cae97fc845173fcf13a54643878bec7b040506ee9b893173e98d3134b.jpg) +add +Figure 4. Schemes of reconstruction constraint. (a) Layer-to-layer (sparse). (b) Layer-to-cat-layer. (c) Layer-to-layer (dense). (d) Loose group-to-group, 1-group (Ours). (e) Loose group-to-group, 2-group (Ours). + +concerning the number of tokens [26]. By substituting Softmax operation with a simple activation function $\phi (\cdot)$ (usually $\phi (x) = \mathrm{elu}(x) + 1$ ), we can change the computation order from $(\mathbf{Q}\mathbf{K}^T)\mathbf{V}$ to $\mathbf{Q}(\mathbf{K}^T\mathbf{V})$ . Formally, Linear Attention (LA) is given as: + +$$ +\operatorname {L A} (\mathbf {Q}, \mathbf {K}, \mathbf {V}) = (\phi (\mathbf {Q}) \phi \left(\mathbf {K} ^ {T}\right)) \mathbf {V} = \phi (\mathbf {Q}) \left(\phi \left(\mathbf {K} ^ {T}\right) \mathbf {V}\right), \tag {3} +$$ + +where the computation complexity is reduced to $\mathcal{O}(Nd^2)$ from $\mathcal{O}(N^2 d)$ . The trade-off between complexity and expressiveness is a dilemma. Previous studies [15, 48] attribute Linear Attention's performance degradation on supervised tasks to its incompetence in focusing. Due to the absence of non-linear attention reweighting by Softmax operation, Linear Attention cannot concentrate on important regions related to the query, such as foreground and neighbors. This property, however, is exactly what the reconstruction decoder favors in our contexts. + +In order to probe how Attentions propagate information, we train two variants of Dinomaly using vanilla Softmax Attention or Linear Attention as the spatial mixer in the decoder and visualize their attention maps. As shown in Figure 3, Softmax Attention tends to focus on the exact region of the query, while Linear Attention spreads its attention across the whole image. This implies that Linear Attention, forced by its incompetence to focus, utilizes more long-range information to restore features at each position, reducing the chance of passing identical information of unseen patterns to the next layer during reconstruction. Of course, employing Linear Attention also benefits from less computation, free of performance drop. + +# 3.4. Loose Reconstruction + +"The tighter you squeeze, the less you have." + +Loose Constraint. Pioneers of feature-reconstruction/distillation UAD methods [10, 46] are inspired by knowledge distillation [20]. Most reconstruction-based methods distill specific encoder layers (e.g. 3 last layers of 3 ResNet stages) by the corresponding decoder layers [10, 46, 65] (Figure 4(a)) or the last decoder layer [58, 60] (Figure 4(b)). Intuitively, with more encoder-decoder feature pairs (Figure 4(c)), UAD model can utilize + +more information in different layers to discriminate anomalies. However, according to the intuition of knowledge distillation, the student (decoder) can better mimic the behavior of the teacher (encoder) given more layer-to-layer supervision [30], which is harmful for UAD models that detect anomalies by encoder-decoder discrepancy. This phenomenon is also embodied as identity mapping. Thanks to the top-to-bottom consistency of columnar Transformer layers, we propose to loosen the layer-to-layer constraint by adding up all feature maps of interested layers as a whole group, as shown in Figure 4(d). This scheme can be seen as loosening the layer-to-layer correspondence and providing the decoder with more degrees of freedom, so that the decoder is allowed to act much more differently from the encoder when the input pattern is unseen. Because features of shallow layers contain low-level visual characters that are helpful for precise localization, we can further group the features into the low-semantic-level group and high-semantic-level group, as shown in Figure 4(e). + +**Loose Loss.** Following the above analysis, we also loosen the point-by-point reconstruction loss function by discarding some points in the feature map. Here, we simply borrow the hard-mining global cosine loss [14] that detaches the gradients of well-restored feature points with low cosine distance during training. Let $f_{E}$ and $f_{D}$ denotes (grouped) feature maps of encoder and decoder: + +$$ +\mathcal {L} _ {\text {g l o b a l} - h m} = \mathcal {D} _ {\text {c o s}} \left(\mathcal {F} \left(f _ {E}\right), \mathcal {F} \left(\hat {f} _ {D}\right)\right), \tag {4} +$$ + +$$ +\hat {f} _ {D} (h, w) = \left\{ \begin{array}{l} s g \left(f _ {D} (h, w)\right)_{0.1}, \text {if} \mathcal {D} _ {\text {cos}} \left(f _ {D}, f _ {E}\right) < k \% _ {\text {batch}} \\ f _ {D} (h, w), \text {else} \end{array} \right. \tag{5} +$$ + +$$ +\mathcal {D} _ {\cos} (a, b) = 1 - \frac {a ^ {T} \cdot b}{\| a \| \| b \|}, \tag {6} +$$ + +where $\mathcal{D}_{cos}$ denotes cosine distance, $\mathcal{F}(\cdot)$ denotes flatten operation, $f_{D}(h,w)$ represents the feature point at $(h,w)$ , and $sg(\cdot)_{0.1}$ denotes shrink the gradient to one-tenth of the original $^2$ . $\mathcal{D}_{cos}(f_D(h,w),f_E(h,w)) < k\%_{batch}$ selects $k \%$ feature points with smaller cosine distance within a batch for gradient shrinking. Total loss is the average $\mathcal{L}_{global - hm}$ of all encoder-decoder feature pairs. + +Table 1. Performance under multi-class UAD setting (%). †: method designed for MUAD. Dinomaly†: training schedule is scaled up to 20,000, 20,000, and 100,000 iterations (original: 10,000/10,000/50,000). + +
DatasetMethodImage-levelPixel-level
AUROCAP\(F_1\)-maxAUROCAP\(F_1\)-maxAUPRO
MVTec-AD [3]RD4AD [10]94.696.595.296.148.653.891.1
SimpleNet [34]95.398.495.896.945.949.786.5
DeSTSeg [67]89.295.591.693.154.350.964.8
UniAD [60]†96.598.896.296.843.449.590.7
ReContrast [14]†98.399.497.697.160.261.593.2
DiAD [18]†97.299.096.596.852.655.590.7
ViTAD [65]†98.399.497.397.755.358.791.4
MambaAD [17]†98.699.697.897.756.359.293.1
Dinomaly (Ours)99.699.899.098.469.369.294.8
Dinomaly†99.799.899.298.469.369.294.7
VisA [70]RD4AD [10]92.492.489.698.138.042.691.8
SimpleNet [34]87.287.081.896.834.737.881.4
DeSTSeg [67]88.989.085.296.139.643.467.4
UniAD [60]†88.890.885.898.333.739.085.5
ReContrast [14]†95.596.492.098.547.950.691.9
DiAD [18]†86.888.385.196.026.133.075.2
ViTAD [65]†90.591.786.398.236.641.185.1
MambaAD [17]†94.394.589.498.539.444.091.0
Dinomaly (Ours)98.798.996.298.753.255.794.5
Dinomaly†98.999.096.498.853.855.894.5
Real-IAD [54]RD4AD [10]82.479.073.997.325.032.789.6
SimpleNet [34]57.253.461.575.72.86.539.0
DeSTSeg [67]82.379.273.294.637.941.740.6
UniAD [60]†83.080.974.397.321.129.286.7
ReContrast [14]†86.484.277.497.831.638.291.8
DiAD [18]†75.666.469.988.02.97.158.1
ViTAD [65]†82.780.273.797.224.332.384.8
MambaAD [17]†86.384.677.098.533.038.790.5
Dinomaly (Ours)89.386.880.298.842.847.193.9
Dinomaly†89.586.980.498.943.347.494.2
+ +# 4. Experiments + +# 4.1. Experimental Settings + +Datasets. MVtec-AD [3] contains 15 objects (5 texture classes and 10 object classes) with a total of 3,629 normal images as the training set and 1,725 images as the test set (467 normal, 1,258 anomalous). VisA [70] contains 12 objects. Training and test sets are split following the official splitting, resulting in 8,659 normal images in the training set and 2,162 images in the test set (962 normal, 1,200 anomalous). Real-IAD [54] is a large UAD dataset recently released, containing 30 distinct objects. We follow the official splitting that includes all views, resulting in 36,465 normal images in the training set and 114,585 images in the test set (63,256 normal, 51,329 anomalous). + +Metrics. Following prior works [17, 65], we adopt 7 evaluation metrics. Image-level anomaly detection performance is measured by the Area Under the Receiver Operator Curve (AUROC), Average Precision (AP), and $F_{1}$ score under optimal threshold $(F_{1}$ -max). Pixel-level anomaly localization is measured by AUROC, AP, $F_{1}$ -max and the Area Under the Per-Region-Overlap (AUPRO). The results + +of a dataset is the average of all classes. + +Implementation Details. ViT-Base/14 (patchsize=14) pre-trained by DINOv2-R [7] is used as the encoder by default. The drop rate of Noisy Bottleneck is 0.2 by default and increases to 0.4 on the diverse Real-IAD. Loose constraint with 2 groups is employed, and the anomaly map is given by the mean per-point cosine distance of the 2 groups. The input image is first resized to $448^2$ and then center-cropped to $392^2$ , so the feature map $(28^2)$ is large enough for anomaly localization. StableAdamW optimizer [56] with AMSGrad [41] (more stable than AdamW [35] in training) is utilized with $lr = 2e - 3$ , $\beta = (0.9,0.999)$ and $wd = 1e - 4$ . The network is trained for 10,000 iterations (steps) on MVTec-AD and VisA, and 50,000 iterations on Real-IAD. Detailed settings are available in Appendix B. + +# 4.2. Comparison to Multi-Class UAD SoTAs + +We compare the proposed Dinomaly with the most advanced UAD and MUAD methods [10, 14, 17, 18, 34, 65, 67]. Experimental results are presented in Table 1, where Dinomaly surpasses compared methods by a large margin on all datasets and all metrics. On the most + +Table 2. Performance under conventional class-separated UAD setting (%). n/a: not available. + +
MethodMVTec-AD [3]VisA [70]Real-IAD [54]
I-AUROCP-AUROCP-AUPROI-AUROCP-AUROCP-AUPROI-AUROCP-AUROCP-AUPRO
Dinomaly (MUAD)99.698.494.898.798.794.589.398.893.9
Dinomaly99.799.995.098.998.995.192.099.195.1
RD4AD [10]98.597.893.996.090.170.987.1n/a93.8
PatchCore [45]99.198.193.594.798.591.889.4n/a91.5
SimpleNet [34]99.698.190.097.198.290.788.5n/a84.6
+ +Table 3. Ablations of Dinomaly elements on MVTec-AD (%). NB: Noisy Bottleneck. LA: Linear Attention. LC: Loose Constraint (2 groups). LL: Loose Loss. As MVTec-AD has reached saturation, we also present the results on VisA (Table A4). + +
NBLALCLLImage-levelPixel-level
AUROCAPF1-maxAUROCAPF1-maxAUPRO
98.4199.0997.4197.1862.9663.8292.95
✓99.0699.5498.3197.6266.2266.7093.71
✓98.5499.2197.6297.2062.9463.7393.09
✓98.3599.0497.4397.1061.0562.7392.60
✓99.0399.4598.1997.6264.1064.9693.34
✓✓99.2799.6298.6397.8567.3667.3394.16
✓✓99.5099.7298.8798.1468.1668.2494.23
✓✓✓99.5299.7398.9298.2068.2568.3494.17
✓✓✓99.5799.7899.0098.2067.9368.2194.50
✓✓✓✓99.6099.7899.0498.3569.2969.1794.79
+ +widely used MVTec-AD, Dinomaly produces image-level performance of 99.6/99.8/99.0 (\%) and pixel-level performance of 98.4/69.3/69.2/94.8, outperforming previous SoTAs by 1.0/0.2/1.2 and 0.7/9.1/7.7/1.6. This result declares that the image-level performance on the MVTec-AD dataset is nearly saturated under the MUAD setting. On the popular VisA, Dinomaly achieves image-level performance of 98.7/98.9/96.2 and pixel-level performance of 98.7/53.2/55.7/94.5, outperforming previous SoTAs by 3.2/2.5/4.2 and 0.2/5.3/5.1/2.6. On the Real-IAD that contains 30 classes, each with 5 camera views, we produce image-level and pixel-level performance of 89.3/86.8/80.2 and 98.8/42.8/47.1/93.9, outperforming previous SoTAs by 3.0/2.2/3.2 and 0.3/4.9/5.4/3.4, indicating our scalability to extremely complex scenarios. Per-class performances and qualitative visualization are presented in Appendix E and F. We also produce superior results on other popular UAD benchmarks, i.e., MPDD [24], BTAD [38], and UniMedical [66], with I-AUROC of 97.2, 95.4, and 84.9, respectively, as shown in Table A13 in Appendix. + +# 4.3. Comparison to Class-Separated UAD SoTAs + +Dinomaly is also compared with class-separated SoTAs, as shown in Table 2. Dinomaly under MUAD setting is comparable to conventional methods [10, 34, 45] that build individual models for each class. On MVTec-AD and VisA, multi-class Dinomaly (first row) is subjected to nearly no + +performance drop compared to its class-separated counterpart (second row). On the complicated Real-IAD that involves more classes and views, multi-class Dinomaly suffers a moderate performance drop but is still comparable to class-separated SoTAs. + +# 4.4. Ablation Study + +Overall Ablation. We conduct experiments to verify the effectiveness of the proposed elements, i.e., Noisy Bottleneck (NB), Linear Attention (LA), Loose Constraint (LC), and Loose Loss (LL). The already-powerful baseline (first row) is Dinomaly with noiseless MLP bottleneck, Softmax Attention, dense layer-to-layer supervision, and global cosine loss. This baseline is very similar to ViTAD [64] and the ViT version of RD4AD [10]. Results on MVtec-AD and VisA are shown in Table 3 and Table A4, respectively. NB and LL can directly contribute to the model performance. LA and LC boost the performance with the presence of NB. The use of LC is not solely beneficial because LC makes the reconstruction too easy without injected noise. + +Model Scalability. Previous works [10, 60, 65] reported that anomaly detection networks do not follow the "scaling law". For example, RD4AD [10] found WideResNet50 better than WideResNet101 as the encoder backbone. ViTAD [65] found ViT-Small better than ViT-Base. On the contrary, as shown in Table 4, the performance of the proposed Dinomaly benefits from scaling. Dinomaly equipped with ViT + +Table 4. Scaling of ViT model sizes on MVtec-AD (%). Im/s (Throughoutput, image per second) is measured on NVIDIA RTX3090 with batch size=16. Results on VisA and Real-IAD are shown in Table A3. †:default. + +
Arch.ParamsMACsIm/sImage-levelPixel-level
AUROCAPF1-maxAUROCAPF1-maxAUPRO
ViT-Small37.4M26.3G153.699.2699.6798.7298.0768.2967.7894.36
ViT-Base†148.0M104.7G58.199.6099.7899.0498.3569.2969.1794.79
ViT-Large275.3M413.5G24.299.7799.9299.4598.5470.5370.0495.09
+ +Table 5. Scaling input size on MVTec-AD (\%). †: default. Compared methods yield degradation when increasing input size. + +
MethodInput SizeImage-LevelPixel-Level
RD4AD2562†94.6/96.5/96.196.1/48.6/53.8/91.1
320293.2/96.9/95.695.7/55.1/57.5/91.1
384291.9/96.2/95.094.9/52.1/55.3/90.8
ReContrast2562†98.3/99.4/97.697.1/60.2/61.5/93.2
320298.2/99.2/97.596.8/61.8/62.6/93.3
384295.2/98.0/96.496.5/57.7/59.5/92.6
Dinomaly280299.6/99.8/99.398.2/65.2/66.3/93.6
336299.6/99.8/99.298.3/67.2/67.8/94.2
3922†99.6/99.8/99.098.4/69.3/69.2/94.8
+ +Small has already produced state-of-the-art results. ViT-Large further boosts Dinomaly to an unprecedented higher record. This scalability enables users to choose an appropriate model size based on the computational resources available in their specific scenario. A comparison of computational costs with other methods is presented in Table A11. In addition, training schedule can also be scaled up for even better performance without increasing inference costs, as demonstrated in Figure 1 (Dinomaly $\uparrow$ ). + +Input Scalability. Though it seems unfair to compare Dinomaly with previous works that take smaller images as input, we contend that increasing their input size not only fails to benefit but actively undermines their performance, especially for image-level detection performance, as shown in Table 5. Therefore, we follow the common comparison strategy based on "optimum vs. optimum". On the contrary, Dinomaly enjoys scaling input size for anomaly localization, while still producing SoTA performance given smaller images. Details are presented in Table A2 in Appendix. + +ViT Foundations. We conduct extensive experiments to investigate the impact of diverse pre-trained ViT foundations, including DeiT [50], MAE [19], D-iGPT [44], MOCOV3 [6], DINO [4], iBot [69], DINOv2 [39], and DINOv2-R [7]. As shown in Figure 5, Dinomaly is robust to the choice of backbone. Almost all foundation models can produce SoTA-level results with image-level AUROC higher than $98\%$ . The only notable exception is MAE, which, without fine-tuning, was reported to be less effective across various unsupervised tasks, e.g., kNN and linear- + +![](images/7534438a9f00e66162f29b3e33941023cda6c2ee6fa9e8f5a27a44c64f62b210.jpg) +Figure 5. Image-level AUROC of Dinomaly equipped with various ViT foundations, and their linear-probing accuracy on ImageNet. MIM: Masked Image Modeling. CL: Contrastive Learning. + +probing [39]. The optimal input size varies because the these backbones are pre-trained on different resolutions. Interestingly, we found the anomaly detection performance to be strongly correlated with the accuracy of ImageNet linear-probing (freeze backbone & only tune linear classifier) of the foundation model, suggesting the possibility of further improvement by simply adopting a future foundation model. Detailed results and analysis are presented in Appendix and Table A1. + +Additional experiments and results are detailed in the Appendix C, encompassing evaluations of various pretrained foundations, ablation studies of each components, hyperparameter optimization, and other in-depth analyses. + +# 5. Conclusion + +Dinomaly, a minimalistic UAD framework, is proposed to address the under-performed MUAD models in this paper. We present four key elements in Dinomaly, i.e., Foundation Transformer, Noisy MLP Bottleneck, Linear Attention, and Loose Reconstruction, that can boost the performance under the challenging MUAD setting without fancy modules and tricks. Extensive experiments on MVTec AD, VisA, and Real-IAD demonstrate our superiority over previous model-unified multi-class models and even recent class-separated models, indicating the feasibility of implementing a unified model in complicated scenarios free of severe performance degradation. + +# Acknowledgments + +# References + +[1] Samet Akcay, Amir Atapour-Abarghouei, and Toby P Breckon. Ganomaly: Semi-supervised anomaly detection via adversarial training. In Computer Vision-ACCV 2018: 14th Asian Conference on Computer Vision, Perth, Australia, December 2-6, 2018, Revised Selected Papers, Part III 14, pages 622-637. Springer, 2019. 1 +[2] Paul Bergmann, Sindy Löwe, Michael Fauser, David Sattlegger, and Carsten Steger. Improving unsupervised defect segmentation by applying structural similarity to autoencoders. arXiv preprint arXiv:1807.02011, 2018. 1 +[3] Paul Bergmann, Michael Fauser, David Sattlegger, and Carsten Steger. Mvtec ad-a comprehensive real-world dataset for unsupervised anomaly detection. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 9592-9600, 2019. 1, 2, 6, 7 +[4] Mathilde Caron, Hugo Touvron, Ishan Misra, Hervé Jégou, Julien Mairal, Piotr Bojanowski, and Armand Joulin. Emerging properties in self-supervised vision transformers. In Proceedings of the IEEE/CVF international conference on computer vision, pages 9650-9660, 2021. 3, 8, 2 +[5] Xinlei Chen and Kaiming He. Exploring simple siamese representation learning. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 15750-15758, 2021. 3 +[6] Xinlei Chen, Saining Xie, and Kaiming He. An empirical study of training self-supervised vision transformers. In Proceedings of the IEEE/CVF international conference on computer vision, pages 9640-9649, 2021. 3, 8, 2 +[7] Timothee Darcet, Maxime Oquab, Julien Mairal, and Piotr Bojanowski. Vision transformers need registers. arXiv preprint arXiv:2309.16588, 2023. 3, 6, 8, 1, 2 +[8] Thomas Defard, Aleksandr Setkov, Angelique Loesch, and Romaric Audigier. Padim: a patch distribution modeling framework for anomaly detection and localization. In International Conference on Pattern Recognition, pages 475-489. Springer, 2021. 1 +[9] David Dehaene and Pierre Eline. Anomaly localization by modeling perceptual features. arXiv preprint arXiv:2008.05369, 2020. 1 +[10] Hanqiu Deng and Xingyu Li. Anomaly detection via reverse distillation from one-class embedding. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 9737-9746, 2022. 5, 6, 7, 1, 8, 9, 10 +[11] Jia Deng, Wei Dong, Richard Socher, Li-Jia Li, Kai Li, and Li Fei-Fei. Imagenet: A large-scale hierarchical image database. In IEEE Conference on Computer Vision and Pattern Recognition, pages 248–255, 2009. 2 + +The authors acknowledge supports from National Natural Science Foundation of China (U22A2051, 82027807), National Key Research and Development Program of China (2022YFC2405200), Tsinghua-Foshan Innovation Special Fund (2021THFS0104), and Institute for Intelligent Healthcare, Tsinghua University (2022ZLB001). +[12] Alexey Dosovitskiy, Lucas Beyer, Alexander Kolesnikov, Dirk Weissenborn, Xiaohua Zhai, Thomas Unterthiner, Mostafa Dehghani, Matthias Minderer, Georg Heigold, Sylvain Gelly, et al. An image is worth 16x16 words: Transformers for image recognition at scale. arXiv preprint arXiv:2010.11929, 2020. 2, 3 +[13] Jia Guo, Shuai Lu, Lize Jia, Weihang Zhang, and Huiqi Li. Encoder-decoder contrast for unsupervised anomaly detection in medical images. IEEE Transactions on Medical Imaging, 2023. 1 +[14] Jia Guo, Shuai Lu, Lize Jia, Weihang Zhang, and Huiqi Li. Recontrast: Domain-specific anomaly detection via contrastive reconstruction. In Advances in Neural Information Processing Systems (NeurIPS), pages 10721-10740, 2023. 2, 3, 5, 6 +[15] Dongchen Han, Xuran Pan, Yizeng Han, Shiji Song, and Gao Huang. Flatten transformer: Vision transformer using focused linear attention. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 5961-5971, 2023. 5 +[16] Haoyang He, Jiangning Zhang, Hongxu Chen, Xuhai Chen, Zhishan Li, Xu Chen, Yabiao Wang, Chengjie Wang, and Lei Xie. Diad: A diffusion-based framework for multi-class anomaly detection. arXiv preprint arXiv:2312.06607, 2023. 1, 3 +[17] Haoyang He, Yuhu Bai, Jiangning Zhang, Qingdong He, Hongxu Chen, Zhenye Gan, Chengjie Wang, Xiangtai Li, Guanzhong Tian, and Lei Xie. Mambaad: Exploring state space models for multi-class unsupervised anomaly detection. arXiv preprint arXiv:2404.06564, 2024. 1, 3, 6, 7, 8, 10 +[18] Haoyang He, Jiangning Zhang, Hongxu Chen, Xuhai Chen, Zhishan Li, Xu Chen, Yabiao Wang, Chengjie Wang, and Lei Xie. A diffusion-based framework for multi-class anomaly detection. In Proceedings of the AAAI Conference on Artificial Intelligence, pages 8472-8480, 2024. 2, 6, 7, 8, 9, 10 +[19] Kaiming He, Xinlei Chen, Saining Xie, Yanghao Li, Piotr Dollar, and Ross Girshick. Masked autoencoders are scalable vision learners. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 16000-16009, 2022. 3, 8, 2, 5 +[20] Geoffrey Hinton, Oriol Vinyals, and Jeff Dean. Distilling the knowledge in a neural network. arXiv preprint arXiv:1503.02531, 2015. 5 +[21] Geoffrey E Hinton, Nitish Srivastava, Alex Krizhevsky, Ilya Sutskever, and Ruslan R Salakhutdinov. Improving neural networks by preventing co-adaptation of feature detectors. arXiv preprint arXiv:1207.0580, 2012. 4 +[22] Chaoqin Huang, Haoyan Guan, Aofan Jiang, Ya Zhang, Michael Spratling, and Yan-Feng Wang. Registration based few-shot anomaly detection. In European Conference on Computer Vision, pages 303-319. Springer, 2022. 7 +[23] Jongheon Jeong, Yang Zou, Taewan Kim, Dongqing Zhang, Avinash Ravichandran, and Onkar Dabeer. Winclip: Zero/few-shot anomaly classification and segmentation. arXiv preprint arXiv:2303.14814, 2023.7 +[24] Stepan Jezek, Martin Jonak, Radim Burget, Pavel Dvorak, and Milos Skotak. Deep learning-based defect detection of + +metal parts: evaluating current methods in complex conditions. In 2021 13th International Congress on Ultra Modern Telecommunications and Control Systems and Workshops (ICUMT), pages 66-71. IEEE, 2021. 7, 6 +[25] Xi Jiang, Jianlin Liu, Jinbao Wang, Qiang Nie, Kai Wu, Yong Liu, Chengjie Wang, and Feng Zheng. Softpatch: Unsupervised anomaly detection with noisy data. Advances in Neural Information Processing Systems, 35:15433-15445, 2022. 6, 7 +[26] Angelos Katharopoulos, Apoorv Vyas, Nikolaos Pappas, and François Fleuret. Transformers are rnns: Fast autoregressive transformers with linear attention. In International conference on machine learning, pages 5156-5165. PMLR, 2020. 5 +[27] Sungwook Lee, Seunghyun Lee, and Byung Cheol Song. Cfa: Coupled-hypersphere-based feature adaptation for target-oriented anomaly localization. IEEE Access, 10: 78446-78454, 2022. 1 +[28] Yujin Lee, Harin Lim, and Hyunsoo Yoon. Selfformaly: Towards task-agnostic unified anomaly detection. arXiv preprint arXiv:2307.12540, 2023. 3 +[29] Chun-Liang Li, Kihyuk Sohn, Jinsung Yoon, and Tomas Pfister. Cutpaste: Self-supervised learning for anomaly detection and localization. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 9664-9674, 2021. 1 +[30] Chen Liang, Jiahui Yu, Ming-Hsuan Yang, Matthew Brown, Yin Cui, Tuo Zhao, Boqing Gong, and Tianyi Zhou. Modulewise adaptive distillation for multimodality foundation models. Advances in Neural Information Processing Systems, 36, 2024. 5 +[31] Jiangqi Liu and Feng Wang. mixed attention auto encoder for multi-class industrial anomaly detection. arXiv preprint arXiv:2309.12700, 2023. 2 +[32] Wenqian Liu, Runze Li, Meng Zheng, Srikrishna Karanam, Ziyan Wu, Bir Bhanu, Richard J Radke, and Octavia Camps. Towards visually explaining variational autoencoders. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 8642-8651, 2020. 1 +[33] Ze Liu, Yutong Lin, Yue Cao, Han Hu, Yixuan Wei, Zheng Zhang, Stephen Lin, and Baining Guo. Swin transformer: Hierarchical vision transformer using shifted windows. In IEEE/CVF International Conference on Computer Vision, pages 10012-10022, 2021. 3 +[34] Zhikang Liu, Yiming Zhou, Yuansheng Xu, and Zilei Wang. Simplenet: A simple network for image anomaly detection and localization. arXiv preprint arXiv:2303.15140, 2023. 6, 7, 1, 8, 9, 10 +[35] Ilya Loshchilov and Frank Hutter. Decoupled weight decay regularization. arXiv preprint arXiv:1711.05101, 2017. 6 +[36] Ruiying Lu, YuJie Wu, Long Tian, Dongsheng Wang, Bo Chen, Xiyang Liu, and Ruimin Hu. Hierarchical vector quantized transformer for multi-class unsupervised anomaly detection. arXiv preprint arXiv:2310.14228, 2023. 1, 2, 4 +[37] Amira Ben Mabrouk and Ezzeddine Zagrouba. Abnormal behavior recognition for intelligent video surveillance systems: A review. Expert Systems with Applications, 91:480-491, 2018. 1 + +[38] Pankaj Mishra, Riccardo Verk, Daniele Fornasier, Claudio Piciarelli, and Gian Luca Foresti. Vt-adl: A vision transformer network for image anomaly detection and localization. In 2021 IEEE 30th International Symposium on Industrial Electronics (ISIE), pages 01–06. IEEE, 2021. 7, 6 +[39] Maxime Oquab, Timothee Darcet, Theo Moutakanni, Huy Vo, Marc Szafraniec, Vasil Khalidov, Pierre Fernandez, Daniel Haziza, Francisco Massa, Alaaeldin El-Nouby, et al. Dinov2: Learning robust visual features without supervision. arXiv preprint arXiv:2304.07193, 2023. 3, 8, 2 +[40] Z Peng, L Dong, H Bao, Q Ye, and F Wei. Beit v2: Masked image modeling with vector-quantized visual tokenizers. arXiv preprint arXiv:2208.06366, 2022. 3, 2 +[41] Sashank J Reddi, Satyen Kale, and Sanjiv Kumar. On the convergence of adam and beyond. arXiv preprint arXiv:1904.09237, 2019. 6, 1 +[42] Tal Reiss, Niv Cohen, Liron Bergman, and Yedid Hoshen. Panda: Adapting pretrained features for anomaly detection and segmentation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 2806-2814, 2021. 7 +[43] Tal Reiss, Niv Cohen, Eliahu Horwitz, Ron Abutbul, and Yedid Hoshen. Anomaly detection requires better representations. In European Conference on Computer Vision, pages 56-68. Springer, 2022. 3, 7 +[44] Sucheng Ren, Zeyu Wang, Hongru Zhu, Junfei Xiao, Alan Yuille, and Cihang Xie. Rejuvenating image-gpt as strong visual representation learners. arXiv preprint arXiv:2312.02147, 2023.8, 2 +[45] Karsten Roth, Latha Pemula, Joaquin Zepeda, Bernhard Schölkopf, Thomas Brox, and Peter Gehler. Towards total recall in industrial anomaly detection. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 14318-14328, 2022. 7, 1, 3 +[46] Mohammadreza Salehi, Niousha Sadjadi, Soroosh Baselizadeh, Mohammad H Rohban, and Hamid R Rabiee. Multiresolution knowledge distillation for anomaly detection. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 14902-14912, 2021. 5, 1 +[47] Thomas Schlegl, Philipp Seebock, Sebastian M Waldstein, Georg Langs, and Ursula Schmidt-Erfurth. f-anogan: Fast unsupervised anomaly detection with generative adversarial networks. Medical image analysis, 54:30-44, 2019. 1 +[48] Zhuoran Shen, Mingyuan Zhang, Haiyu Zhao, Shuai Yi, and Hongsheng Li. Efficient attention: Attention with linear complexities. In Proceedings of the IEEE/CVF winter conference on applications of computer vision, pages 3531-3539, 2021. 5 +[49] Shelly Sheynin, Sagie Benaim, and Lior Wolf. A hierarchical transformation-discriminating generative model for few shot anomaly detection. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 8495-8504, 2021. 1 +[50] Hugo Touvron, Matthieu Cord, Matthijs Douze, Francisco Massa, Alexandre Sablayrolles, and Herv'e J'egou. Training data-efficient image transformers & distillation through attention. arXiv preprint arXiv:2012.12877, 2021. 8, 2 + +[51] Ashish Vaswani, Noam Shazeer, Niki Parmar, Jakob Uszkoreit, Llion Jones, Aidan N Gomez, Lukasz Kaiser, and Illia Polosukhin. Attention is all you need. In Advances in Neural Information Processing Systems, pages 5998-6008, 2017. 2 +[52] Pascal Vincent, Hugo Larochelle, Yoshua Bengio, and Pierre-Antoine Manzagol. Extracting and composing robust features with denoising autoencoders. In Proceedings of the 25th international conference on Machine learning, pages 1096-1103, 2008. 4 +[53] Pascal Vincent, Hugo Larochelle, Isabelle Lajoie, Yoshua Bengio, and Pierre-Antoine Manzagol. Stacked denoising autoencoders: Learning useful representations in a deep network with a local denoising criterion. In Journal of Machine Learning Research, pages 3371-3408, 2010. 4 +[54] Chengjie Wang, Wenbing Zhu, Bin-Bin Gao, Zhenye Gan, Jianning Zhang, Zhihao Gu, Shuguang Qian, Mingang Chen, and Lizhuang Ma. Real-iad: A real-world multi-view dataset for benchmarking versatile industrial anomaly detection. arXiv preprint arXiv:2403.12580, 2024. 2, 6, 7, 3 +[55] Guodong Wang, Shumin Han, Errui Ding, and Di Huang. Student-teacher feature pyramid matching for anomaly detection. In The British Machine Vision Conference (BMVC), 2021. 1 +[56] Mitchell Wortsman, Tim Dettmers, Luke Zettlemoyer, Ari Morcos, Ali Farhadi, and Ludwig Schmidt. Stable and low-precision training for large-scale vision-language models. Advances in Neural Information Processing Systems, 36: 10271–10298, 2023. 6, 1 +[57] Zhenda Xie, Zheng Zhang, Yue Cao, Yutong Lin, Jianmin Bao, Zhuliang Yao, Qi Dai, and Han Hu. Simmim: A simple framework for masked image modeling. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 9653-9663, 2022. 3 +[58] Jie Yang, Yong Shi, and Zhiquan Qi. Dfr: Deep feature reconstruction for unsupervised anomaly segmentation. arXiv preprint arXiv:2012.07122, 2020. 5, 1 +[59] Haonan Yin, Guanlong Jiao, Qianhui Wu, Borne F Karlsson, Biqing Huang, and Chin Yew Lin. Lafite: Latent diffusion model with feature editing for unsupervised multi-class anomaly detection. arXiv preprint arXiv:2307.08059, 2023.1, 2, 3, 4 +[60] Zhiyuan You, Lei Cui, Yujun Shen, Kai Yang, Xin Lu, Yu Zheng, and Xinyi Le. A unified model for multi-class anomaly detection. arXiv preprint arXiv:2206.03687, 2022. 1, 2, 3, 4, 5, 6, 7, 8, 9, 10 +[61] Weihao Yu, Chenyang Si, Pan Zhou, Mi Luo, Yichen Zhou, Jiashi Feng, Shuicheng Yan, and Xinchao Wang. Metaformer baselines for vision. IEEE Transactions on Pattern Analysis and Machine Intelligence, 2023. 4 +[62] Vitjan Zavrtanik, Matej Kristan, and Danijel Skoc aj. Draem-a discriminatively trained reconstruction embedding for surface anomaly detection. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 8330-8339, 2021. 4, 1 +[63] Vitjan Zavrtanik, Matej Kristan, and Danijel Skocaj. Reconstruction by inpainting for visual anomaly detection. Pattern Recognition, 112:107706, 2021. 1 + +[64] Dingwen Zhang, Guohai Huang, Qiang Zhang, Jungong Han, Junwei Han, Yizhou Wang, and Yizhou Yu. Exploring task structure for brain tumor segmentation from multimodality mr images. IEEE Transactions on Image Processing, 29:9032-9043, 2020. 7 +[65] Jiangning Zhang, Xuhai Chen, Yabiao Wang, Chengjie Wang, Yong Liu, Xiangtai Li, Ming-Hsuan Yang, and Dacheng Tao. Exploring plain vit reconstruction for multiclass unsupervised anomaly detection. arXiv preprint arXiv:2312.07495, 2023. 3, 5, 6, 7 +[66] Jiangning Zhang, Haoyang He, Zhenye Gan, Qingdong He, Yuxuan Cai, Zhucun Xue, Yabiao Wang, Chengjie Wang, Lei Xie, and Yong Liu. Ader: A comprehensive benchmark for multi-class visual anomaly detection. arXiv preprint arXiv:2406.03262, 2024. 7, 2, 6 +[67] Xuan Zhang, Shiyu Li, Xi Li, Ping Huang, Jiulong Shan, and Ting Chen. Destseg: Segmentation guided denoising student-teacher for anomaly detection. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 3914–3923, 2023. 4, 6, 1, 7, 8, 9, 10 +[68] Ying Zhao. Omnial: A unified cnn framework for unsupervised anomaly localization. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 3924-3933, 2023. 1, 3 +[69] Jinghao Zhou, Chen Wei, Huiyu Wang, Wei Shen, Cihang Xie, Alan Yuille, and Tao Kong. ibot: Image bert pre-training with online tokenizer. arXiv preprint arXiv:2111.07832, 2021. 3, 8, 2 +[70] Yang Zou, Jongheon Jeong, Latha Pemula, Dongqing Zhang, and Onkar Dabeer. Spot-the-difference self-supervised pretraining for anomaly detection and segmentation. In European Conference on Computer Vision, pages 392-408. Springer, 2022. 2, 6, 7, 3 \ No newline at end of file diff --git a/2025/Dinomaly_ The Less Is More Philosophy in Multi-Class Unsupervised Anomaly Detection/images.zip b/2025/Dinomaly_ The Less Is More Philosophy in Multi-Class Unsupervised Anomaly Detection/images.zip new file mode 100644 index 0000000000000000000000000000000000000000..f7cd571d4f0040343953636efee8fbc2adb98d43 --- /dev/null +++ b/2025/Dinomaly_ The Less Is More Philosophy in Multi-Class Unsupervised Anomaly Detection/images.zip @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:274491dfb9d6ea1b017e32472658031dacc839a26a74ed0fd02572ca1279824c +size 651967 diff --git a/2025/Dinomaly_ The Less Is More Philosophy in Multi-Class Unsupervised Anomaly Detection/layout.json b/2025/Dinomaly_ The Less Is More Philosophy in Multi-Class Unsupervised Anomaly Detection/layout.json new file mode 100644 index 0000000000000000000000000000000000000000..5a083398d320f9bd3ca343d1d99ea62c0e8a4477 --- /dev/null +++ b/2025/Dinomaly_ The Less Is More Philosophy in Multi-Class Unsupervised Anomaly Detection/layout.json @@ -0,0 +1,8933 @@ +{ + "pdf_info": [ + { + "para_blocks": [ + { + "bbox": [ + 164, + 103, + 446, + 140 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 164, + 103, + 446, + 140 + ], + "spans": [ + { + "bbox": [ + 164, + 103, + 446, + 140 + ], + "type": "text", + "content": "Dinomaly: The Less Is More Philosophy in Multi-Class Unsupervised Anomaly Detection" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 91, + 159, + 518, + 177 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 91, + 159, + 518, + 177 + ], + "spans": [ + { + "bbox": [ + 91, + 159, + 518, + 177 + ], + "type": "text", + "content": "Jia Guo" + }, + { + "bbox": [ + 91, + 159, + 518, + 177 + ], + "type": "inline_equation", + "content": "^{1}" + }, + { + "bbox": [ + 91, + 159, + 518, + 177 + ], + "type": "text", + "content": " Shuai Lu" + }, + { + "bbox": [ + 91, + 159, + 518, + 177 + ], + "type": "inline_equation", + "content": "^{2}" + }, + { + "bbox": [ + 91, + 159, + 518, + 177 + ], + "type": "text", + "content": " Weihang Zhang" + }, + { + "bbox": [ + 91, + 159, + 518, + 177 + ], + "type": "inline_equation", + "content": "^{2\\dagger}" + }, + { + "bbox": [ + 91, + 159, + 518, + 177 + ], + "type": "text", + "content": " Fang Chen" + }, + { + "bbox": [ + 91, + 159, + 518, + 177 + ], + "type": "inline_equation", + "content": "^{3}" + }, + { + "bbox": [ + 91, + 159, + 518, + 177 + ], + "type": "text", + "content": " Huiqi Li" + }, + { + "bbox": [ + 91, + 159, + 518, + 177 + ], + "type": "inline_equation", + "content": "^{2\\dagger}" + }, + { + "bbox": [ + 91, + 159, + 518, + 177 + ], + "type": "text", + "content": " Hongen Liao" + }, + { + "bbox": [ + 91, + 159, + 518, + 177 + ], + "type": "inline_equation", + "content": "^{1,3\\boxtimes}" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 129, + 189, + 481, + 217 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 129, + 189, + 481, + 217 + ], + "spans": [ + { + "bbox": [ + 129, + 189, + 481, + 217 + ], + "type": "inline_equation", + "content": "^{1}" + }, + { + "bbox": [ + 129, + 189, + 481, + 217 + ], + "type": "text", + "content": "School of Biomedical Engineering, Tsinghua University, Beijing, China " + }, + { + "bbox": [ + 129, + 189, + 481, + 217 + ], + "type": "inline_equation", + "content": "^{2}" + }, + { + "bbox": [ + 129, + 189, + 481, + 217 + ], + "type": "text", + "content": "Beijing Institute of Technology, Beijing, China" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 100, + 217, + 510, + 232 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 100, + 217, + 510, + 232 + ], + "spans": [ + { + "bbox": [ + 100, + 217, + 510, + 232 + ], + "type": "inline_equation", + "content": "^{3}" + }, + { + "bbox": [ + 100, + 217, + 510, + 232 + ], + "type": "text", + "content": "School of Biomedical Engineering, Shanghai Jiao Tong University, Shanghai, China" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 128, + 234, + 478, + 258 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 128, + 234, + 478, + 258 + ], + "spans": [ + { + "bbox": [ + 128, + 234, + 478, + 258 + ], + "type": "text", + "content": "guojia.jeremy@gmail.com lushuaie@163.com zhangweihang@bit.edu.cn chen-fang@sjtu.edu.cn huiqili@bit.edu.cn liao@tsinghua.edu.cn" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 152, + 299, + 200, + 312 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 152, + 299, + 200, + 312 + ], + "spans": [ + { + "bbox": [ + 152, + 299, + 200, + 312 + ], + "type": "text", + "content": "Abstract" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 55, + 329, + 297, + 664 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 329, + 297, + 664 + ], + "spans": [ + { + "bbox": [ + 55, + 329, + 297, + 664 + ], + "type": "text", + "content": "Recent studies highlighted a practical setting of unsupervised anomaly detection (UAD) that builds a unified model for multi-class images. Despite various advancements addressing this challenging task, the detection performance under the multi-class setting still lags far behind state-of-the-art class-separated models. Our research aims to bridge this substantial performance gap. In this paper, we present Dinomaly, a minimalist reconstruction-based anomaly detection framework that harnesses pure Transformer architectures without relying on complex designs, additional modules, or specialized tricks. Given this powerful framework consisting of only Attentions and MLPs, we found four simple components that are essential to multi-class anomaly detection: (1) Scalable foundation Transformers that extracts universal and discriminative features, (2) Noisy Bottleneck where pre-existing Dropouts do all the noise injection tricks, (3) Linear Attention that naturally cannot focus, and (4) Loose Reconstruction that does not force layer-to-layer and point-by-point reconstruction. Extensive experiments are conducted across popular anomaly detection benchmarks including MVtec-AD, VisA, Real-IAD, etc. Our proposed Dinomaly achieves impressive image-level AUROC of " + }, + { + "bbox": [ + 55, + 329, + 297, + 664 + ], + "type": "inline_equation", + "content": "99.6\\%" + }, + { + "bbox": [ + 55, + 329, + 297, + 664 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 55, + 329, + 297, + 664 + ], + "type": "inline_equation", + "content": "98.7\\%" + }, + { + "bbox": [ + 55, + 329, + 297, + 664 + ], + "type": "text", + "content": ", and " + }, + { + "bbox": [ + 55, + 329, + 297, + 664 + ], + "type": "inline_equation", + "content": "89.3\\%" + }, + { + "bbox": [ + 55, + 329, + 297, + 664 + ], + "type": "text", + "content": " on the three datasets respectively, which is not only superior to state-of-the-art multi-class UAD methods, but also achieves the most advanced class-separated UAD records. Code is available at: https://github.com/guojiajeremy/Dinomaly" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 314, + 286, + 394, + 298 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 314, + 286, + 394, + 298 + ], + "spans": [ + { + "bbox": [ + 314, + 286, + 394, + 298 + ], + "type": "text", + "content": "1. Introduction" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 313, + 306, + 555, + 426 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 306, + 555, + 426 + ], + "spans": [ + { + "bbox": [ + 313, + 306, + 555, + 426 + ], + "type": "text", + "content": "Unsupervised anomaly detection (UAD) aims to detect abnormal patterns from normal images and further localize the anomalous regions. Because of the diversity of potential anomalies and their scarcity, this task is proposed to model the accessible training sets containing only normal samples as an unsupervised paradigm. UAD has a wide range of applications, e.g., industrial defect detection [3], medical disease screening [13], and video surveillance [37], addressing the difficulty of collecting and labeling all possible anomalies in these scenarios." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 313, + 426, + 556, + 653 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 426, + 556, + 653 + ], + "spans": [ + { + "bbox": [ + 313, + 426, + 556, + 653 + ], + "type": "text", + "content": "Conventional works on UAD build a separate model for each object category, as shown in Figure 1(a). However, this one-class-one-model setting entails substantial storage overhead for saving models [60], especially when the application scenario necessitates a large number of object classes. For UAD methods, a compact boundary of normal patterns is vital to distinguish anomalies. Once the intra-normal patterns become exceedingly complicated due to various classes, the corresponding distribution becomes challenging to measure, consequently harming the detection performance. Recently, UniAD [60] and successive studies have been proposed to train a unified model for multi-class anomaly detection (MUAD), as shown in Figure 1(b). Under this setting, the \"identity mapping\" that directly copies the input as the output regardless of normal or anomaly harms the performance of conventional methods [60]. This phenomenon is caused by the diversity of multi-class normal patterns that drive the network to generalize on unseen patterns." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 313, + 654, + 556, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 654, + 556, + 715 + ], + "spans": [ + { + "bbox": [ + 313, + 654, + 556, + 715 + ], + "type": "text", + "content": "Within two years, a number of methods have been proposed to address MUAD, such as neighbor-masked attention [60], synthetic anomalies [68], vector quantization [36], diffusion model [16, 59], and state space model (Mamba) [17]. However, there is still a non-negligible per" + } + ] + } + ], + "index": 12 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "spans": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "type": "text", + "content": "CVF" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 145, + 0, + 496, + 37 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 145, + 0, + 496, + 37 + ], + "spans": [ + { + "bbox": [ + 145, + 0, + 496, + 37 + ], + "type": "text", + "content": "This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 69, + 702, + 295, + 713 + ], + "type": "page_footnote", + "angle": 0, + "lines": [ + { + "bbox": [ + 69, + 702, + 295, + 713 + ], + "spans": [ + { + "bbox": [ + 69, + 702, + 295, + 713 + ], + "type": "text", + "content": "Corresponding author. † Former advisors when the project initiated." + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 293, + 748, + 317, + 758 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 317, + 758 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 317, + 758 + ], + "type": "text", + "content": "20405" + } + ] + } + ], + "index": 14 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 0 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 70, + 77, + 167, + 150 + ], + "blocks": [ + { + "bbox": [ + 70, + 77, + 167, + 150 + ], + "lines": [ + { + "bbox": [ + 70, + 77, + 167, + 150 + ], + "spans": [ + { + "bbox": [ + 70, + 77, + 167, + 150 + ], + "type": "image", + "image_path": "e77de717e7226e870fe1250726262007dd1a4ff99ce23a99dd348328cf5b24f7.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 73, + 151, + 160, + 159 + ], + "lines": [ + { + "bbox": [ + 73, + 151, + 160, + 159 + ], + "spans": [ + { + "bbox": [ + 73, + 151, + 160, + 159 + ], + "type": "text", + "content": "(a) Class-Separated UAD Setting" + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "type": "image", + "bbox": [ + 176, + 76, + 301, + 174 + ], + "blocks": [ + { + "bbox": [ + 176, + 76, + 301, + 174 + ], + "lines": [ + { + "bbox": [ + 176, + 76, + 301, + 174 + ], + "spans": [ + { + "bbox": [ + 176, + 76, + 301, + 174 + ], + "type": "image", + "image_path": "21af34168446ad566b0d66a8553a9bddb83a215a657a4e8c6b1775ed7512a181.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 213, + 174, + 277, + 182 + ], + "lines": [ + { + "bbox": [ + 213, + 174, + 277, + 182 + ], + "spans": [ + { + "bbox": [ + 213, + 174, + 277, + 182 + ], + "type": "text", + "content": "MVTec-AD (15 classes)" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_caption" + } + ], + "index": 2 + }, + { + "type": "image", + "bbox": [ + 302, + 76, + 416, + 174 + ], + "blocks": [ + { + "bbox": [ + 302, + 76, + 416, + 174 + ], + "lines": [ + { + "bbox": [ + 302, + 76, + 416, + 174 + ], + "spans": [ + { + "bbox": [ + 302, + 76, + 416, + 174 + ], + "type": "image", + "image_path": "752ff9fe62b375b33baafdeedcfd0b9ae29b5ad861e630b336cf9b22f9bea3ab.jpg" + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 339, + 175, + 386, + 182 + ], + "lines": [ + { + "bbox": [ + 339, + 175, + 386, + 182 + ], + "spans": [ + { + "bbox": [ + 339, + 175, + 386, + 182 + ], + "type": "text", + "content": "VisA (12 classes)" + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "image_caption" + } + ], + "index": 4 + }, + { + "type": "image", + "bbox": [ + 420, + 76, + 533, + 174 + ], + "blocks": [ + { + "bbox": [ + 420, + 76, + 533, + 174 + ], + "lines": [ + { + "bbox": [ + 420, + 76, + 533, + 174 + ], + "spans": [ + { + "bbox": [ + 420, + 76, + 533, + 174 + ], + "type": "image", + "image_path": "6de827691005c6a5c2758f19abd6cacf12119cbd3dd32c9c6d1fa1f13c2a2eae.jpg" + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 449, + 175, + 508, + 182 + ], + "lines": [ + { + "bbox": [ + 449, + 175, + 508, + 182 + ], + "spans": [ + { + "bbox": [ + 449, + 175, + 508, + 182 + ], + "type": "text", + "content": "Real-IAD (30 classes)" + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "image_caption" + } + ], + "index": 6 + }, + { + "type": "image", + "bbox": [ + 70, + 180, + 165, + 247 + ], + "blocks": [ + { + "bbox": [ + 70, + 180, + 165, + 247 + ], + "lines": [ + { + "bbox": [ + 70, + 180, + 165, + 247 + ], + "spans": [ + { + "bbox": [ + 70, + 180, + 165, + 247 + ], + "type": "image", + "image_path": "e357a9a888575f41ce9e2b5de7c3db2424bd3b9e759acf39d47a4e7ea73dcc37.jpg" + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 73, + 251, + 82, + 258 + ], + "lines": [ + { + "bbox": [ + 73, + 251, + 82, + 258 + ], + "spans": [ + { + "bbox": [ + 73, + 251, + 82, + 258 + ], + "type": "text", + "content": "(b)" + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 103, + 250, + 141, + 255 + ], + "lines": [ + { + "bbox": [ + 103, + 250, + 141, + 255 + ], + "spans": [ + { + "bbox": [ + 103, + 250, + 141, + 255 + ], + "type": "text", + "content": "Model-Unified" + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 90, + 256, + 154, + 263 + ], + "lines": [ + { + "bbox": [ + 90, + 256, + 154, + 263 + ], + "spans": [ + { + "bbox": [ + 90, + 256, + 154, + 263 + ], + "type": "text", + "content": "Multi-Class UAD Setting" + } + ] + } + ], + "index": 11, + "angle": 0, + "type": "image_caption" + } + ], + "index": 8 + }, + { + "type": "image", + "bbox": [ + 181, + 194, + 297, + 258 + ], + "blocks": [ + { + "bbox": [ + 181, + 194, + 297, + 258 + ], + "lines": [ + { + "bbox": [ + 181, + 194, + 297, + 258 + ], + "spans": [ + { + "bbox": [ + 181, + 194, + 297, + 258 + ], + "type": "image", + "image_path": "bfa686fdd5f1898aaee9f58a6810b419098f70670743e14fdd62616ac3beeb8b.jpg" + } + ] + } + ], + "index": 12, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 279, + 262, + 468, + 272 + ], + "lines": [ + { + "bbox": [ + 279, + 262, + 468, + 272 + ], + "spans": [ + { + "bbox": [ + 279, + 262, + 468, + 272 + ], + "type": "text", + "content": "(d) Scaling Properties of Dinomaly on MVTec-AD. " + }, + { + "bbox": [ + 279, + 262, + 468, + 272 + ], + "type": "inline_equation", + "content": "\\dagger" + }, + { + "bbox": [ + 279, + 262, + 468, + 272 + ], + "type": "text", + "content": " : default setting." + } + ] + } + ], + "index": 16, + "angle": 0, + "type": "image_caption" + } + ], + "index": 12 + }, + { + "type": "image", + "bbox": [ + 304, + 194, + 417, + 259 + ], + "blocks": [ + { + "bbox": [ + 313, + 185, + 432, + 193 + ], + "lines": [ + { + "bbox": [ + 313, + 185, + 432, + 193 + ], + "spans": [ + { + "bbox": [ + 313, + 185, + 432, + 193 + ], + "type": "text", + "content": "(c) Image-Level AUROC on 3 Benchmarks" + } + ] + } + ], + "index": 13, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 304, + 194, + 417, + 259 + ], + "lines": [ + { + "bbox": [ + 304, + 194, + 417, + 259 + ], + "spans": [ + { + "bbox": [ + 304, + 194, + 417, + 259 + ], + "type": "image", + "image_path": "7c9079b35c2f53586e2ae811f1f5e9bee8e28950af1f243868cd1825e5748932.jpg" + } + ] + } + ], + "index": 14, + "angle": 0, + "type": "image_body" + } + ], + "index": 14 + }, + { + "type": "image", + "bbox": [ + 423, + 195, + 542, + 260 + ], + "blocks": [ + { + "bbox": [ + 423, + 195, + 542, + 260 + ], + "lines": [ + { + "bbox": [ + 423, + 195, + 542, + 260 + ], + "spans": [ + { + "bbox": [ + 423, + 195, + 542, + 260 + ], + "type": "image", + "image_path": "3a384fc4fb010232dc6e2160c574502a492a9d1cb6613d52a8df42eb04d61f13.jpg" + } + ] + } + ], + "index": 15, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 55, + 281, + 555, + 304 + ], + "lines": [ + { + "bbox": [ + 55, + 281, + 555, + 304 + ], + "spans": [ + { + "bbox": [ + 55, + 281, + 555, + 304 + ], + "type": "text", + "content": "Figure 1. Setting, benchmarking, and scaling of Dinomaly. (a) Task setting of class-separated UAD. (b) Task setting of MUAD. (c) Comparison with previous SoTA methods on MVtec-AD [3], VisA [70], and Real-IAD [54]. (d) Scaling properties of Dinomaly." + } + ] + } + ], + "index": 17, + "angle": 0, + "type": "image_caption" + } + ], + "index": 15 + }, + { + "bbox": [ + 55, + 324, + 296, + 408 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 324, + 296, + 408 + ], + "spans": [ + { + "bbox": [ + 55, + 324, + 296, + 408 + ], + "type": "text", + "content": "formance gap between the state-of-the-art (SoTA) MUAD methods and class-separated UAD methods, restricting the practicability of implementing unified models, as shown in Figure 1(c). In addition, previous methods employ modules and architectures delicately designed, which may not be straightforward, and consequently suffer from limited universality and ease-of-use [18, 36]." + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 55, + 415, + 296, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 415, + 296, + 714 + ], + "spans": [ + { + "bbox": [ + 55, + 415, + 296, + 714 + ], + "type": "text", + "content": "In this work, we aim to catch up with the performance of class-separated anomaly detection models using a multiclass unified model. We introduce Dinomaly, a minimalist reconstruction-based UAD framework built exclusively by pure Transformer blocks [51], specifically Self-Attentions and Multi-Layer Perceptrons (MLPs). To begin with, we empirically investigate the scaling law of self-supervised pre-trained Vision Transformers (ViT) [12] when serving as the feature encoders for extracting reconstruction objectives. Subsequently, we introduce three straightforward yet crucial elements to address the critical identity mapping phenomenon in MUAD contexts, without increasing complexity or computational burden. First, as an alternative to meticulously designed pseudo anomaly and feature noise, we propose to activate the built-in Dropout in an MLP to prevent the network from restoring both normal and anomalous patterns. Second, we propose to leverage the \"side effect\" of Linear Attention (a computation-efficient counterpart of Softmax Attention) that impedes focus on local regions, thus preventing the forwarding of identical information. Third, previous methods adopt layer-to-layer and region-by-region reconstruction schemes, distilling a decoder that can well mimic the encoder's behavior even for anomalous regions. Therefore, we propose to loosen the reconstruction constraints by grouping multiple layers as a" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 313, + 324, + 553, + 347 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 324, + 553, + 347 + ], + "spans": [ + { + "bbox": [ + 313, + 324, + 553, + 347 + ], + "type": "text", + "content": "whole and discarding well-reconstructed regions during optimization." + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 313, + 348, + 555, + 504 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 348, + 555, + 504 + ], + "spans": [ + { + "bbox": [ + 313, + 348, + 555, + 504 + ], + "type": "text", + "content": "To validate the effectiveness of the proposed Dinomaly under MUAD setting, we conduct extensive experiments on a number of widely used benchmarks, including MVTec AD [3] (15 classes), VisA [70] (12 classes), and RealIAD (30 classes). As shown in Figure 1, our base-size Dinomaly achieves unprecedented image-level AUROC of " + }, + { + "bbox": [ + 313, + 348, + 555, + 504 + ], + "type": "inline_equation", + "content": "99.6\\%" + }, + { + "bbox": [ + 313, + 348, + 555, + 504 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 313, + 348, + 555, + 504 + ], + "type": "inline_equation", + "content": "98.7\\%" + }, + { + "bbox": [ + 313, + 348, + 555, + 504 + ], + "type": "text", + "content": ", and " + }, + { + "bbox": [ + 313, + 348, + 555, + 504 + ], + "type": "inline_equation", + "content": "89.3\\%" + }, + { + "bbox": [ + 313, + 348, + 555, + 504 + ], + "type": "text", + "content": " on MVTec AD, VisA, and RealIAD, surpassing previous SoTA methods by a large margin. In addition, scalability is a key feature of Dinomaly. Further scaling up the model size maximizes performance to the fullest level of " + }, + { + "bbox": [ + 313, + 348, + 555, + 504 + ], + "type": "inline_equation", + "content": "99.8\\%" + }, + { + "bbox": [ + 313, + 348, + 555, + 504 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 313, + 348, + 555, + 504 + ], + "type": "inline_equation", + "content": "98.9\\%" + }, + { + "bbox": [ + 313, + 348, + 555, + 504 + ], + "type": "text", + "content": ", and " + }, + { + "bbox": [ + 313, + 348, + 555, + 504 + ], + "type": "inline_equation", + "content": "90.1\\%" + }, + { + "bbox": [ + 313, + 348, + 555, + 504 + ], + "type": "text", + "content": ", respectively; while scaling down parameters and input size can offer efficient solutions in computation-constrained scenarios." + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 313, + 514, + 400, + 526 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 514, + 400, + 526 + ], + "spans": [ + { + "bbox": [ + 313, + 514, + 400, + 526 + ], + "type": "text", + "content": "2. Related Work" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 313, + 533, + 555, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 533, + 555, + 713 + ], + "spans": [ + { + "bbox": [ + 313, + 533, + 555, + 713 + ], + "type": "text", + "content": "Multi-Class UAD. UniAD [60] first introduced multi-class anomaly detection, aiming to detect anomalies for different classes using a unified model. In this setting, conventional UAD methods often face the challenge of \"identical shortcuts\", where both anomaly-free and anomaly samples can be effectively recovered during inference [60]. It is believed that this phenomenon is caused by the diversity of multi-class normal patterns that drive the network to generalize on unseen patterns. This contradicts the fundamental assumption of epistemic methods. Many current researches focus on addressing this challenge [14, 31, 36, 59, 60]. UniAD [60] employs a neighbor-masked attention module and a feature-jitter strategy to mitigate these shortcuts. HVQ-Trans [36] proposes a vector quantization (VQ) Transformer model that induces large feature discrepancies" + } + ] + } + ], + "index": 23 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "text", + "content": "20406" + } + ] + } + ], + "index": 24 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 1 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 58, + 73, + 556, + 228 + ], + "blocks": [ + { + "bbox": [ + 58, + 73, + 556, + 228 + ], + "lines": [ + { + "bbox": [ + 58, + 73, + 556, + 228 + ], + "spans": [ + { + "bbox": [ + 58, + 73, + 556, + 228 + ], + "type": "image", + "image_path": "17d3d0fa6c8cb95057e7196ca865d09815a7ac4253dda8c3a5682950e09bd3d4.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 134, + 238, + 474, + 250 + ], + "lines": [ + { + "bbox": [ + 134, + 238, + 474, + 250 + ], + "spans": [ + { + "bbox": [ + 134, + 238, + 474, + 250 + ], + "type": "text", + "content": "Figure 2. The framework of Dinomaly, built by simple and pure Transformer building blocks." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "bbox": [ + 54, + 270, + 297, + 429 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 54, + 270, + 297, + 429 + ], + "spans": [ + { + "bbox": [ + 54, + 270, + 297, + 429 + ], + "type": "text", + "content": "for anomalies. LafitE [59] utilizes a latent diffusion model and introduces a feature editing strategy to alleviate this issue. DiAD [16] also employs diffusion models to address multi-class UAD settings. OmniAL [68] focuses on anomaly localization in the unified setting, preventing identical reconstruction by using synthesized pseudo anomalies. ReContrast [14] attempted to alleviate the identity mapping by cross-reconstruction between two encoders. ViTAD [5] abstracts a unified feature-reconstruction UAD framework and employ Transformer building blocks. MambaAD [17] explores the recently proposed State Space Model (SSM), Mamba, in the context of multi-class UAD. More related works of UAD are presented in Appendix A." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 55, + 436, + 112, + 449 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 436, + 112, + 449 + ], + "spans": [ + { + "bbox": [ + 55, + 436, + 112, + 449 + ], + "type": "text", + "content": "3. Method" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 55, + 456, + 183, + 468 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 456, + 183, + 468 + ], + "spans": [ + { + "bbox": [ + 55, + 456, + 183, + 468 + ], + "type": "text", + "content": "3.1. Dinomaly Framework" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 56, + 474, + 237, + 485 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 474, + 237, + 485 + ], + "spans": [ + { + "bbox": [ + 56, + 474, + 237, + 485 + ], + "type": "text", + "content": "\"What I cannot create, I do not understand\"" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 67, + 487, + 157, + 498 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 67, + 487, + 157, + 498 + ], + "spans": [ + { + "bbox": [ + 67, + 487, + 157, + 498 + ], + "type": "text", + "content": "Richer Feynman" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 55, + 498, + 297, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 498, + 297, + 714 + ], + "spans": [ + { + "bbox": [ + 55, + 498, + 297, + 714 + ], + "type": "text", + "content": "The ability to recognize anomalies from what we know is an innate human capability, serving as a vital pathway for us to explore the world. Similarly, we construct a reconstruction-based framework that relies on the epistemic characteristic of artificial neural networks. Dinomaly consists of an encoder, a bottleneck, and a reconstruction decoder, as shown in Figure 2. Without loss of generality, a pretrained ViT network [12] with 12 Transformer layers is used as the encoder, extracting informative feature maps with different semantic scales. The bottleneck is a simple MLP (a.k.a. feed-forward network, FFN) that collects the feature representations of the encoder's 8 middle-level layers. The decoder is similar to the encoder, consisting of 8 Transformer layers. During training, the decoder learns to reconstruct the middle-level features of the encoder by maximizing the cosine similarity between feature maps. During inference, the decoder is expected to reconstruct normal regions of feature maps but fails for anomalous regions as it" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 313, + 271, + 432, + 282 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 271, + 432, + 282 + ], + "spans": [ + { + "bbox": [ + 313, + 271, + 432, + 282 + ], + "type": "text", + "content": "has never seen such samples." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 313, + 283, + 555, + 390 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 283, + 555, + 390 + ], + "spans": [ + { + "bbox": [ + 313, + 283, + 555, + 390 + ], + "type": "text", + "content": "Foundation Transformers. Foundation models, especially ViTs [12, 33] pre-trained on large-scale datasets, serve as a basis and starting point for specific computer vision tasks. Such networks employ self-supervised learning schemes such as contrastive learning (MoCov3 [6], DINO [4]), masked image modeling (MAE [19], SimMIM [57], BEiT [40]), and their combination (iBOT [69], DINOv2 [39]), producing universal features suitable for image-level visual tasks and pixel-level visual tasks." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 313, + 391, + 556, + 570 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 391, + 556, + 570 + ], + "spans": [ + { + "bbox": [ + 313, + 391, + 556, + 570 + ], + "type": "text", + "content": "Because of the lack of supervision in UAD, most advanced methods adopt pre-trained networks to extract discriminative features. Recent works [28, 43, 65] have preliminarily discovered the advantage of robust and universal features of self-supervised models over domain-specific ImageNet features in anomaly detection tasks. In this paper, we pioneer the investigation of scaling behaviors in UAD models through a systematic analysis of foundational ViTs, as briefed in Figure 1(d). Our comprehensive evaluation encompasses pre-training strategies (Figure 5), model sizes (Table 4), and input resolutions (Table 5), which are detailed in section 4.4. Considering the balance of detection performance and computational efficiency, we adopt ViT-Base/14 pretrained by DINoV2-Register [7] as the encoder of Dinomaly by default." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 313, + 575, + 419, + 588 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 575, + 419, + 588 + ], + "spans": [ + { + "bbox": [ + 313, + 575, + 419, + 588 + ], + "type": "text", + "content": "3.2. Noisy Bottleneck." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 314, + 594, + 424, + 605 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 314, + 594, + 424, + 605 + ], + "spans": [ + { + "bbox": [ + 314, + 594, + 424, + 605 + ], + "type": "text", + "content": "\"Dropout is all you need.\"" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 313, + 606, + 556, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 606, + 556, + 715 + ], + "spans": [ + { + "bbox": [ + 313, + 606, + 556, + 715 + ], + "type": "text", + "content": "Prior studies [14, 60, 68] attribute the performance degradation of UAD methods trained on diverse multi-class samples to the \"identity mapping\" phenomenon; in this work, we reframe this issue as an \"over-generalization\" problem. Generalization ability is a merit of neural networks, allowing them to perform equally well on unseen test sets. However, generalization is not so wanted in the context of unsupervised anomaly detection that leverages the epistemic nature of neural networks. With the increas" + } + ] + } + ], + "index": 13 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "type": "text", + "content": "20407" + } + ] + } + ], + "index": 14 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 2 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 97, + 73, + 347, + 270 + ], + "blocks": [ + { + "bbox": [ + 97, + 73, + 347, + 270 + ], + "lines": [ + { + "bbox": [ + 97, + 73, + 347, + 270 + ], + "spans": [ + { + "bbox": [ + 97, + 73, + 347, + 270 + ], + "type": "image", + "image_path": "6d4a43b73593602e9e9d33bb1e07ed954c2eb370106d46e6efa33d5071ffbc4f.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 101, + 282, + 507, + 295 + ], + "lines": [ + { + "bbox": [ + 101, + 282, + 507, + 295 + ], + "spans": [ + { + "bbox": [ + 101, + 282, + 507, + 295 + ], + "type": "text", + "content": "Figure 3. Softmax Attention vs. Linear Attention. (a) Visualization of attention maps. (b) Attention distribution." + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "type": "image", + "bbox": [ + 360, + 92, + 503, + 267 + ], + "blocks": [ + { + "bbox": [ + 360, + 92, + 503, + 267 + ], + "lines": [ + { + "bbox": [ + 360, + 92, + 503, + 267 + ], + "spans": [ + { + "bbox": [ + 360, + 92, + 503, + 267 + ], + "type": "image", + "image_path": "8d125b0bc5327cd9d9792b71ec97f6b7783286bcb1c5a99b13be4c2f135ca734.jpg" + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_body" + } + ], + "index": 1 + }, + { + "bbox": [ + 54, + 315, + 295, + 362 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 54, + 315, + 295, + 362 + ], + "spans": [ + { + "bbox": [ + 54, + 315, + 295, + 362 + ], + "type": "text", + "content": "ing diversity of images and their patterns due to multi-class UAD settings, the decoder can generalize its reconstruction ability to unseen anomalous samples, resulting in the failure of anomaly detection using reconstruction error." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 56, + 363, + 296, + 651 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 363, + 296, + 651 + ], + "spans": [ + { + "bbox": [ + 56, + 363, + 296, + 651 + ], + "type": "text", + "content": "A direct solution for identity mapping is to shift \"reconstruction\" to \"restoration\". Specifically, instead of directly reconstructing the normal images or features given normal inputs, previous works propose to add perturbations as pseudo anomalies on input images [62, 67] or on encoder features [59, 60]; meanwhile, still let the decoder restore anomaly-free images or features, formulating a denoising-like framework. However, such methods employ heuristic and hand-crafted anomaly generation strategies that may not be universal across domains, datasets, and methods. In this work, we turn to leveraging the simple and elegant Dropout techniques. Since its introduction by Hinton et al. [21] in 2014 as a remedy for overfitting, Dropout has become a cornerstone element in neural architectures, including Transformers. In Dinomaly, we employ Dropout to randomly discard neural activations in an MLP bottleneck. Instead of alleviating overfitting, the role of Dropout in Dinomaly can be explained as pseudo feature anomaly that perturb normal representations, analogous to denoising auto-encoders [52, 53]. Without introducing any specific modules, this simple component inherently forces the decoder to restore normal features regardless of whether the test image contains anomalies, in turn, mitigating identical mapping." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 55, + 658, + 211, + 671 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 658, + 211, + 671 + ], + "spans": [ + { + "bbox": [ + 55, + 658, + 211, + 671 + ], + "type": "text", + "content": "3.3. Unfocused Linear Attention." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 56, + 677, + 234, + 689 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 677, + 234, + 689 + ], + "spans": [ + { + "bbox": [ + 56, + 677, + 234, + 689 + ], + "type": "text", + "content": "\"One man's poison is another man's meat\"" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 55, + 689, + 296, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 689, + 296, + 714 + ], + "spans": [ + { + "bbox": [ + 55, + 689, + 296, + 714 + ], + "type": "text", + "content": "Softmax Attention is the key mechanism of Transformers, allowing the model to attend to different parts of its" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 313, + 315, + 555, + 363 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 315, + 555, + 363 + ], + "spans": [ + { + "bbox": [ + 313, + 315, + 555, + 363 + ], + "type": "text", + "content": "input token sequence. Formally, given an input sequence " + }, + { + "bbox": [ + 313, + 315, + 555, + 363 + ], + "type": "inline_equation", + "content": "\\mathbf{X} \\in \\mathbb{R}^{N \\times d}" + }, + { + "bbox": [ + 313, + 315, + 555, + 363 + ], + "type": "text", + "content": " with length " + }, + { + "bbox": [ + 313, + 315, + 555, + 363 + ], + "type": "inline_equation", + "content": "N" + }, + { + "bbox": [ + 313, + 315, + 555, + 363 + ], + "type": "text", + "content": ", Attention first transforms it into three matrices: the query matrix " + }, + { + "bbox": [ + 313, + 315, + 555, + 363 + ], + "type": "inline_equation", + "content": "\\mathbf{Q} \\in \\mathbb{R}^{N \\times d}" + }, + { + "bbox": [ + 313, + 315, + 555, + 363 + ], + "type": "text", + "content": ", the key matrix " + }, + { + "bbox": [ + 313, + 315, + 555, + 363 + ], + "type": "inline_equation", + "content": "\\mathbf{K} \\in \\mathbb{R}^{N \\times d}" + }, + { + "bbox": [ + 313, + 315, + 555, + 363 + ], + "type": "text", + "content": ", and the value matrix " + }, + { + "bbox": [ + 313, + 315, + 555, + 363 + ], + "type": "inline_equation", + "content": "\\mathbf{V} \\in \\mathbb{R}^{N \\times d}" + }, + { + "bbox": [ + 313, + 315, + 555, + 363 + ], + "type": "text", + "content": ":" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 348, + 372, + 555, + 387 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 348, + 372, + 555, + 387 + ], + "spans": [ + { + "bbox": [ + 348, + 372, + 555, + 387 + ], + "type": "interline_equation", + "content": "\\mathbf {Q} = \\mathbf {X} \\mathbf {W} ^ {Q}, \\mathbf {K} = \\mathbf {X} \\mathbf {W} ^ {K}, \\mathbf {V} = \\mathbf {X} \\mathbf {W} ^ {V}, \\tag {1}", + "image_path": "a658d99776cdfa23ba40e672a5dfae4de2a31d40e961638b8e47e4a2acadcfb9.jpg" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 313, + 396, + 554, + 433 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 396, + 554, + 433 + ], + "spans": [ + { + "bbox": [ + 313, + 396, + 554, + 433 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 313, + 396, + 554, + 433 + ], + "type": "inline_equation", + "content": "\\mathbf{W}^Q, \\mathbf{W}^K, \\mathbf{W}^V \\in \\mathbb{R}^{d \\times d}" + }, + { + "bbox": [ + 313, + 396, + 554, + 433 + ], + "type": "text", + "content": " are learnable parameters. By computing the attention map by the query-key similarity, the output of Softmax Attention is given as:" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 345, + 443, + 555, + 458 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 345, + 443, + 555, + 458 + ], + "spans": [ + { + "bbox": [ + 345, + 443, + 555, + 458 + ], + "type": "interline_equation", + "content": "\\operatorname {A t t e n t i o n} (\\mathbf {Q}, \\mathbf {K}, \\mathbf {V}) = \\operatorname {S o f t m a x} (\\mathbf {Q K} ^ {T}) \\mathbf {V}. \\tag {2}", + "image_path": "583c7cd561209d952790d45eddfea2bd1258715f673215b4f3a76f3d0078a7d0.jpg" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 313, + 468, + 555, + 540 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 468, + 555, + 540 + ], + "spans": [ + { + "bbox": [ + 313, + 468, + 555, + 540 + ], + "type": "text", + "content": "Back to MUAD, previous methods [36, 60] suggest adopting Attentions instead of Convolutions because Convolutions can easily learn identical mappings. Nevertheless, both operations are in danger of forming identity mapping by over-concentrating on corresponding input locations for producing the outputs:" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 313, + 549, + 553, + 589 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 549, + 553, + 589 + ], + "spans": [ + { + "bbox": [ + 313, + 549, + 553, + 589 + ], + "type": "interline_equation", + "content": "\\operatorname {C o n v} \\operatorname {K e r n e l} = \\left[ \\begin{array}{l l l} 0 & 0 & 0 \\\\ 0 & 1 & 0 \\\\ 0 & 0 & 0 \\end{array} \\right], \\quad \\operatorname {A t t n} \\operatorname {M a p} = \\left[ \\begin{array}{l l l} 1 & 0 & 0 \\\\ 0 & 1 & 0 \\\\ 0 & 0 & 1 \\end{array} \\right].", + "image_path": "0d5a2fd704ede7a450586e1c9bd25ca8ac78c3470b5739bf5e50d53dde875676.jpg" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 313, + 597, + 555, + 670 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 597, + 555, + 670 + ], + "spans": [ + { + "bbox": [ + 313, + 597, + 555, + 670 + ], + "type": "text", + "content": "Is there any simple solution that prevents Attentions from attending to identical information? In Dinomaly, we turn to leverage the \"unfocusing ability\" of a type of Softmax-free Attention, i.e., Linear Attention. Linear Attention was proposed as a promising alternative to reduce the computation complexity of vanilla Softmax Attention" + } + ] + } + ], + "index": 14 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 313, + 679, + 555, + 713 + ], + "type": "page_footnote", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 679, + 555, + 713 + ], + "spans": [ + { + "bbox": [ + 313, + 679, + 555, + 713 + ], + "type": "text", + "content": "1The full form of Attention is Softmax " + }, + { + "bbox": [ + 313, + 679, + 555, + 713 + ], + "type": "inline_equation", + "content": "\\left(\\frac{\\mathbf{Q}\\mathbf{K}^T}{\\sqrt{d}}\\right)\\mathbf{V}" + }, + { + "bbox": [ + 313, + 679, + 555, + 713 + ], + "type": "text", + "content": ". The constant denominator is omitted for narrative simplicity. The multi-head mechanism that concatenates multiple Attentions is also omitted." + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 293, + 748, + 318, + 757 + ], + "type": "footer", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 318, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 318, + 757 + ], + "type": "text", + "content": "20408" + } + ] + } + ], + "index": 16 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 3 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 65, + 71, + 143, + 147 + ], + "blocks": [ + { + "bbox": [ + 65, + 71, + 143, + 147 + ], + "lines": [ + { + "bbox": [ + 65, + 71, + 143, + 147 + ], + "spans": [ + { + "bbox": [ + 65, + 71, + 143, + 147 + ], + "type": "image", + "image_path": "a1dd9fa401691a462a4df3a84a2f8feedcaf54499e447957f60fb2c121c62204.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 78, + 148, + 141, + 156 + ], + "lines": [ + { + "bbox": [ + 78, + 148, + 141, + 156 + ], + "spans": [ + { + "bbox": [ + 78, + 148, + 141, + 156 + ], + "type": "text", + "content": "(a) layer-to-layer sparse" + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "type": "image", + "bbox": [ + 154, + 70, + 218, + 145 + ], + "blocks": [ + { + "bbox": [ + 154, + 70, + 218, + 145 + ], + "lines": [ + { + "bbox": [ + 154, + 70, + 218, + 145 + ], + "spans": [ + { + "bbox": [ + 154, + 70, + 218, + 145 + ], + "type": "image", + "image_path": "b866424b103ff548ef2ecbe121963fe58d37f24999027884bb87e659c5bc10ba.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 158, + 147, + 215, + 156 + ], + "lines": [ + { + "bbox": [ + 158, + 147, + 215, + 156 + ], + "spans": [ + { + "bbox": [ + 158, + 147, + 215, + 156 + ], + "type": "text", + "content": "(b) layer-to-cat-layer" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_caption" + } + ], + "index": 2 + }, + { + "type": "image", + "bbox": [ + 242, + 70, + 306, + 145 + ], + "blocks": [ + { + "bbox": [ + 242, + 70, + 306, + 145 + ], + "lines": [ + { + "bbox": [ + 242, + 70, + 306, + 145 + ], + "spans": [ + { + "bbox": [ + 242, + 70, + 306, + 145 + ], + "type": "image", + "image_path": "b4bffa389b2c5780fbe925887470053347f3965c3cfab98e72337e6b9155bdfe.jpg" + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 242, + 147, + 305, + 156 + ], + "lines": [ + { + "bbox": [ + 242, + 147, + 305, + 156 + ], + "spans": [ + { + "bbox": [ + 242, + 147, + 305, + 156 + ], + "type": "text", + "content": "(c) layer-to-layer dense" + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "image_caption" + } + ], + "index": 4 + }, + { + "type": "image", + "bbox": [ + 327, + 69, + 389, + 145 + ], + "blocks": [ + { + "bbox": [ + 327, + 69, + 389, + 145 + ], + "lines": [ + { + "bbox": [ + 327, + 69, + 389, + 145 + ], + "spans": [ + { + "bbox": [ + 327, + 69, + 389, + 145 + ], + "type": "image", + "image_path": "2e691d240b522deeace347316f62ed4982cc6efcf42bc8289ce3948972a78386.jpg" + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 331, + 145, + 381, + 161 + ], + "lines": [ + { + "bbox": [ + 331, + 145, + 381, + 161 + ], + "spans": [ + { + "bbox": [ + 331, + 145, + 381, + 161 + ], + "type": "text", + "content": "(d) group-to-group 1 group (Ours)" + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "image_caption" + } + ], + "index": 6 + }, + { + "type": "image", + "bbox": [ + 411, + 70, + 474, + 144 + ], + "blocks": [ + { + "bbox": [ + 411, + 70, + 474, + 144 + ], + "lines": [ + { + "bbox": [ + 411, + 70, + 474, + 144 + ], + "spans": [ + { + "bbox": [ + 411, + 70, + 474, + 144 + ], + "type": "image", + "image_path": "0302c001c32f5d638e646b4c5b0696ae395e2e4637f8de43dad4b5752f96a3ad.jpg" + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 415, + 145, + 465, + 161 + ], + "lines": [ + { + "bbox": [ + 415, + 145, + 465, + 161 + ], + "spans": [ + { + "bbox": [ + 415, + 145, + 465, + 161 + ], + "type": "text", + "content": "(e) group-to-group 2 group (Ours)" + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "image_caption" + } + ], + "index": 8 + }, + { + "type": "image", + "bbox": [ + 499, + 76, + 512, + 91 + ], + "blocks": [ + { + "bbox": [ + 499, + 76, + 512, + 91 + ], + "lines": [ + { + "bbox": [ + 499, + 76, + 512, + 91 + ], + "spans": [ + { + "bbox": [ + 499, + 76, + 512, + 91 + ], + "type": "image", + "image_path": "d3bc1adf7c570b9343648410728143a6c2329294536bf9af7d61d3b9329c354d.jpg" + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 512, + 76, + 533, + 91 + ], + "lines": [ + { + "bbox": [ + 512, + 76, + 533, + 91 + ], + "spans": [ + { + "bbox": [ + 512, + 76, + 533, + 91 + ], + "type": "text", + "content": "encoder feature" + } + ] + } + ], + "index": 11, + "angle": 0, + "type": "image_caption" + } + ], + "index": 10 + }, + { + "type": "image", + "bbox": [ + 500, + 99, + 512, + 110 + ], + "blocks": [ + { + "bbox": [ + 500, + 99, + 512, + 110 + ], + "lines": [ + { + "bbox": [ + 500, + 99, + 512, + 110 + ], + "spans": [ + { + "bbox": [ + 500, + 99, + 512, + 110 + ], + "type": "image", + "image_path": "4767d4991e7ce4e01a71213986c2b2383e2a08849f258d4d882d31d17534b15e.jpg" + } + ] + } + ], + "index": 12, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 512, + 97, + 532, + 110 + ], + "lines": [ + { + "bbox": [ + 512, + 97, + 532, + 110 + ], + "spans": [ + { + "bbox": [ + 512, + 97, + 532, + 110 + ], + "type": "text", + "content": "decoder feature" + } + ] + } + ], + "index": 13, + "angle": 0, + "type": "image_caption" + } + ], + "index": 12 + }, + { + "type": "image", + "bbox": [ + 500, + 117, + 512, + 123 + ], + "blocks": [ + { + "bbox": [ + 500, + 117, + 512, + 123 + ], + "lines": [ + { + "bbox": [ + 500, + 117, + 512, + 123 + ], + "spans": [ + { + "bbox": [ + 500, + 117, + 512, + 123 + ], + "type": "image", + "image_path": "f33596ba1387f7c0a6e5d9120039352aced6c9d016bf983caf8f1150221ea210.jpg" + } + ] + } + ], + "index": 14, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 512, + 116, + 544, + 123 + ], + "lines": [ + { + "bbox": [ + 512, + 116, + 544, + 123 + ], + "spans": [ + { + "bbox": [ + 512, + 116, + 544, + 123 + ], + "type": "text", + "content": "reconstruct" + } + ] + } + ], + "index": 15, + "angle": 0, + "type": "image_caption" + } + ], + "index": 14 + }, + { + "type": "image", + "bbox": [ + 500, + 129, + 512, + 137 + ], + "blocks": [ + { + "bbox": [ + 500, + 129, + 512, + 137 + ], + "lines": [ + { + "bbox": [ + 500, + 129, + 512, + 137 + ], + "spans": [ + { + "bbox": [ + 500, + 129, + 512, + 137 + ], + "type": "image", + "image_path": "c37867f60176995bbf10bf74c757ea2ab30dd85d912f29e92525a9a0297ad90e.jpg" + } + ] + } + ], + "index": 16, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 512, + 129, + 544, + 137 + ], + "lines": [ + { + "bbox": [ + 512, + 129, + 544, + 137 + ], + "spans": [ + { + "bbox": [ + 512, + 129, + 544, + 137 + ], + "type": "text", + "content": "concatenate" + } + ] + } + ], + "index": 17, + "angle": 0, + "type": "image_caption" + } + ], + "index": 16 + }, + { + "type": "image", + "bbox": [ + 500, + 141, + 512, + 148 + ], + "blocks": [ + { + "bbox": [ + 500, + 141, + 512, + 148 + ], + "lines": [ + { + "bbox": [ + 500, + 141, + 512, + 148 + ], + "spans": [ + { + "bbox": [ + 500, + 141, + 512, + 148 + ], + "type": "image", + "image_path": "ac76c81cae97fc845173fcf13a54643878bec7b040506ee9b893173e98d3134b.jpg" + } + ] + } + ], + "index": 18, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 512, + 143, + 524, + 148 + ], + "lines": [ + { + "bbox": [ + 512, + 143, + 524, + 148 + ], + "spans": [ + { + "bbox": [ + 512, + 143, + 524, + 148 + ], + "type": "text", + "content": "add" + } + ] + } + ], + "index": 19, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 54, + 173, + 555, + 196 + ], + "lines": [ + { + "bbox": [ + 54, + 173, + 555, + 196 + ], + "spans": [ + { + "bbox": [ + 54, + 173, + 555, + 196 + ], + "type": "text", + "content": "Figure 4. Schemes of reconstruction constraint. (a) Layer-to-layer (sparse). (b) Layer-to-cat-layer. (c) Layer-to-layer (dense). (d) Loose group-to-group, 1-group (Ours). (e) Loose group-to-group, 2-group (Ours)." + } + ] + } + ], + "index": 20, + "angle": 0, + "type": "image_caption" + } + ], + "index": 18 + }, + { + "bbox": [ + 54, + 217, + 296, + 277 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 54, + 217, + 296, + 277 + ], + "spans": [ + { + "bbox": [ + 54, + 217, + 296, + 277 + ], + "type": "text", + "content": "concerning the number of tokens [26]. By substituting Softmax operation with a simple activation function " + }, + { + "bbox": [ + 54, + 217, + 296, + 277 + ], + "type": "inline_equation", + "content": "\\phi (\\cdot)" + }, + { + "bbox": [ + 54, + 217, + 296, + 277 + ], + "type": "text", + "content": " (usually " + }, + { + "bbox": [ + 54, + 217, + 296, + 277 + ], + "type": "inline_equation", + "content": "\\phi (x) = \\mathrm{elu}(x) + 1" + }, + { + "bbox": [ + 54, + 217, + 296, + 277 + ], + "type": "text", + "content": "), we can change the computation order from " + }, + { + "bbox": [ + 54, + 217, + 296, + 277 + ], + "type": "inline_equation", + "content": "(\\mathbf{Q}\\mathbf{K}^T)\\mathbf{V}" + }, + { + "bbox": [ + 54, + 217, + 296, + 277 + ], + "type": "text", + "content": " to " + }, + { + "bbox": [ + 54, + 217, + 296, + 277 + ], + "type": "inline_equation", + "content": "\\mathbf{Q}(\\mathbf{K}^T\\mathbf{V})" + }, + { + "bbox": [ + 54, + 217, + 296, + 277 + ], + "type": "text", + "content": ". Formally, Linear Attention (LA) is given as:" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 61, + 281, + 296, + 306 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 61, + 281, + 296, + 306 + ], + "spans": [ + { + "bbox": [ + 61, + 281, + 296, + 306 + ], + "type": "interline_equation", + "content": "\\operatorname {L A} (\\mathbf {Q}, \\mathbf {K}, \\mathbf {V}) = (\\phi (\\mathbf {Q}) \\phi \\left(\\mathbf {K} ^ {T}\\right)) \\mathbf {V} = \\phi (\\mathbf {Q}) \\left(\\phi \\left(\\mathbf {K} ^ {T}\\right) \\mathbf {V}\\right), \\tag {3}", + "image_path": "d9524ebd348f8b899c3f1270c8df519e23bd53f56f5896c9df614192a2c1fdc7.jpg" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 54, + 306, + 296, + 426 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 54, + 306, + 296, + 426 + ], + "spans": [ + { + "bbox": [ + 54, + 306, + 296, + 426 + ], + "type": "text", + "content": "where the computation complexity is reduced to " + }, + { + "bbox": [ + 54, + 306, + 296, + 426 + ], + "type": "inline_equation", + "content": "\\mathcal{O}(Nd^2)" + }, + { + "bbox": [ + 54, + 306, + 296, + 426 + ], + "type": "text", + "content": " from " + }, + { + "bbox": [ + 54, + 306, + 296, + 426 + ], + "type": "inline_equation", + "content": "\\mathcal{O}(N^2 d)" + }, + { + "bbox": [ + 54, + 306, + 296, + 426 + ], + "type": "text", + "content": ". The trade-off between complexity and expressiveness is a dilemma. Previous studies [15, 48] attribute Linear Attention's performance degradation on supervised tasks to its incompetence in focusing. Due to the absence of non-linear attention reweighting by Softmax operation, Linear Attention cannot concentrate on important regions related to the query, such as foreground and neighbors. This property, however, is exactly what the reconstruction decoder favors in our contexts." + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 54, + 426, + 296, + 581 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 54, + 426, + 296, + 581 + ], + "spans": [ + { + "bbox": [ + 54, + 426, + 296, + 581 + ], + "type": "text", + "content": "In order to probe how Attentions propagate information, we train two variants of Dinomaly using vanilla Softmax Attention or Linear Attention as the spatial mixer in the decoder and visualize their attention maps. As shown in Figure 3, Softmax Attention tends to focus on the exact region of the query, while Linear Attention spreads its attention across the whole image. This implies that Linear Attention, forced by its incompetence to focus, utilizes more long-range information to restore features at each position, reducing the chance of passing identical information of unseen patterns to the next layer during reconstruction. Of course, employing Linear Attention also benefits from less computation, free of performance drop." + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 55, + 588, + 180, + 599 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 588, + 180, + 599 + ], + "spans": [ + { + "bbox": [ + 55, + 588, + 180, + 599 + ], + "type": "text", + "content": "3.4. Loose Reconstruction" + } + ] + } + ], + "index": 25 + }, + { + "bbox": [ + 56, + 605, + 238, + 617 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 605, + 238, + 617 + ], + "spans": [ + { + "bbox": [ + 56, + 605, + 238, + 617 + ], + "type": "text", + "content": "\"The tighter you squeeze, the less you have.\"" + } + ] + } + ], + "index": 26 + }, + { + "bbox": [ + 54, + 617, + 296, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 54, + 617, + 296, + 713 + ], + "spans": [ + { + "bbox": [ + 54, + 617, + 296, + 713 + ], + "type": "text", + "content": "Loose Constraint. Pioneers of feature-reconstruction/distillation UAD methods [10, 46] are inspired by knowledge distillation [20]. Most reconstruction-based methods distill specific encoder layers (e.g. 3 last layers of 3 ResNet stages) by the corresponding decoder layers [10, 46, 65] (Figure 4(a)) or the last decoder layer [58, 60] (Figure 4(b)). Intuitively, with more encoder-decoder feature pairs (Figure 4(c)), UAD model can utilize" + } + ] + } + ], + "index": 27 + }, + { + "bbox": [ + 313, + 217, + 555, + 443 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 217, + 555, + 443 + ], + "spans": [ + { + "bbox": [ + 313, + 217, + 555, + 443 + ], + "type": "text", + "content": "more information in different layers to discriminate anomalies. However, according to the intuition of knowledge distillation, the student (decoder) can better mimic the behavior of the teacher (encoder) given more layer-to-layer supervision [30], which is harmful for UAD models that detect anomalies by encoder-decoder discrepancy. This phenomenon is also embodied as identity mapping. Thanks to the top-to-bottom consistency of columnar Transformer layers, we propose to loosen the layer-to-layer constraint by adding up all feature maps of interested layers as a whole group, as shown in Figure 4(d). This scheme can be seen as loosening the layer-to-layer correspondence and providing the decoder with more degrees of freedom, so that the decoder is allowed to act much more differently from the encoder when the input pattern is unseen. Because features of shallow layers contain low-level visual characters that are helpful for precise localization, we can further group the features into the low-semantic-level group and high-semantic-level group, as shown in Figure 4(e)." + } + ] + } + ], + "index": 28 + }, + { + "bbox": [ + 313, + 444, + 556, + 528 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 444, + 556, + 528 + ], + "spans": [ + { + "bbox": [ + 313, + 444, + 556, + 528 + ], + "type": "text", + "content": "**Loose Loss.** Following the above analysis, we also loosen the point-by-point reconstruction loss function by discarding some points in the feature map. Here, we simply borrow the hard-mining global cosine loss [14] that detaches the gradients of well-restored feature points with low cosine distance during training. Let " + }, + { + "bbox": [ + 313, + 444, + 556, + 528 + ], + "type": "inline_equation", + "content": "f_{E}" + }, + { + "bbox": [ + 313, + 444, + 556, + 528 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 313, + 444, + 556, + 528 + ], + "type": "inline_equation", + "content": "f_{D}" + }, + { + "bbox": [ + 313, + 444, + 556, + 528 + ], + "type": "text", + "content": " denotes (grouped) feature maps of encoder and decoder:" + } + ] + } + ], + "index": 29 + }, + { + "bbox": [ + 358, + 534, + 553, + 548 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 358, + 534, + 553, + 548 + ], + "spans": [ + { + "bbox": [ + 358, + 534, + 553, + 548 + ], + "type": "interline_equation", + "content": "\\mathcal {L} _ {\\text {g l o b a l} - h m} = \\mathcal {D} _ {\\text {c o s}} \\left(\\mathcal {F} \\left(f _ {E}\\right), \\mathcal {F} \\left(\\hat {f} _ {D}\\right)\\right), \\tag {4}", + "image_path": "00f9a16c161cdd9f823643e1e8f0afeaa93565cad869b1c8e21b097292c4de06.jpg" + } + ] + } + ], + "index": 30 + }, + { + "bbox": [ + 315, + 554, + 553, + 586 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 315, + 554, + 553, + 586 + ], + "spans": [ + { + "bbox": [ + 315, + 554, + 553, + 586 + ], + "type": "interline_equation", + "content": "\\hat {f} _ {D} (h, w) = \\left\\{ \\begin{array}{l} s g \\left(f _ {D} (h, w)\\right)_{0.1}, \\text {if} \\mathcal {D} _ {\\text {cos}} \\left(f _ {D}, f _ {E}\\right) < k \\% _ {\\text {batch}} \\\\ f _ {D} (h, w), \\text {else} \\end{array} \\right. \\tag{5}", + "image_path": "d632127e43dddf6b738e825e4a58d5e9a2e002d982a30fa4f1718c681ac6b9fb.jpg" + } + ] + } + ], + "index": 31 + }, + { + "bbox": [ + 381, + 586, + 553, + 610 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 381, + 586, + 553, + 610 + ], + "spans": [ + { + "bbox": [ + 381, + 586, + 553, + 610 + ], + "type": "interline_equation", + "content": "\\mathcal {D} _ {\\cos} (a, b) = 1 - \\frac {a ^ {T} \\cdot b}{\\| a \\| \\| b \\|}, \\tag {6}", + "image_path": "0c24c9c3f09b517675a42f6d020924fe411315e04a63f4c928dd2059cb7a48bf.jpg" + } + ] + } + ], + "index": 32 + }, + { + "bbox": [ + 313, + 612, + 555, + 696 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 612, + 555, + 696 + ], + "spans": [ + { + "bbox": [ + 313, + 612, + 555, + 696 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 313, + 612, + 555, + 696 + ], + "type": "inline_equation", + "content": "\\mathcal{D}_{cos}" + }, + { + "bbox": [ + 313, + 612, + 555, + 696 + ], + "type": "text", + "content": " denotes cosine distance, " + }, + { + "bbox": [ + 313, + 612, + 555, + 696 + ], + "type": "inline_equation", + "content": "\\mathcal{F}(\\cdot)" + }, + { + "bbox": [ + 313, + 612, + 555, + 696 + ], + "type": "text", + "content": " denotes flatten operation, " + }, + { + "bbox": [ + 313, + 612, + 555, + 696 + ], + "type": "inline_equation", + "content": "f_{D}(h,w)" + }, + { + "bbox": [ + 313, + 612, + 555, + 696 + ], + "type": "text", + "content": " represents the feature point at " + }, + { + "bbox": [ + 313, + 612, + 555, + 696 + ], + "type": "inline_equation", + "content": "(h,w)" + }, + { + "bbox": [ + 313, + 612, + 555, + 696 + ], + "type": "text", + "content": ", and " + }, + { + "bbox": [ + 313, + 612, + 555, + 696 + ], + "type": "inline_equation", + "content": "sg(\\cdot)_{0.1}" + }, + { + "bbox": [ + 313, + 612, + 555, + 696 + ], + "type": "text", + "content": " denotes shrink the gradient to one-tenth of the original " + }, + { + "bbox": [ + 313, + 612, + 555, + 696 + ], + "type": "inline_equation", + "content": "^2" + }, + { + "bbox": [ + 313, + 612, + 555, + 696 + ], + "type": "text", + "content": ". " + }, + { + "bbox": [ + 313, + 612, + 555, + 696 + ], + "type": "inline_equation", + "content": "\\mathcal{D}_{cos}(f_D(h,w),f_E(h,w)) < k\\%_{batch}" + }, + { + "bbox": [ + 313, + 612, + 555, + 696 + ], + "type": "text", + "content": " selects " + }, + { + "bbox": [ + 313, + 612, + 555, + 696 + ], + "type": "inline_equation", + "content": "k \\%" + }, + { + "bbox": [ + 313, + 612, + 555, + 696 + ], + "type": "text", + "content": " feature points with smaller cosine distance within a batch for gradient shrinking. Total loss is the average " + }, + { + "bbox": [ + 313, + 612, + 555, + 696 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{global - hm}" + }, + { + "bbox": [ + 313, + 612, + 555, + 696 + ], + "type": "text", + "content": " of all encoder-decoder feature pairs." + } + ] + } + ], + "index": 33 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 324, + 702, + 549, + 713 + ], + "type": "page_footnote", + "angle": 0, + "lines": [ + { + "bbox": [ + 324, + 702, + 549, + 713 + ], + "spans": [ + { + "bbox": [ + 324, + 702, + 549, + 713 + ], + "type": "inline_equation", + "content": "{}^{2}" + }, + { + "bbox": [ + 324, + 702, + 549, + 713 + ], + "type": "text", + "content": " Complete stop-gradient causes optimization instability occasionally." + } + ] + } + ], + "index": 34 + }, + { + "bbox": [ + 293, + 748, + 318, + 757 + ], + "type": "footer", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 318, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 318, + 757 + ], + "type": "text", + "content": "20409" + } + ] + } + ], + "index": 35 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 4 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 82, + 102, + 531, + 427 + ], + "blocks": [ + { + "bbox": [ + 55, + 70, + 555, + 94 + ], + "lines": [ + { + "bbox": [ + 55, + 70, + 555, + 94 + ], + "spans": [ + { + "bbox": [ + 55, + 70, + 555, + 94 + ], + "type": "text", + "content": "Table 1. Performance under multi-class UAD setting (%). †: method designed for MUAD. Dinomaly†: training schedule is scaled up to 20,000, 20,000, and 100,000 iterations (original: 10,000/10,000/50,000)." + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 82, + 102, + 531, + 427 + ], + "lines": [ + { + "bbox": [ + 82, + 102, + 531, + 427 + ], + "spans": [ + { + "bbox": [ + 82, + 102, + 531, + 427 + ], + "type": "table", + "html": "
DatasetMethodImage-levelPixel-level
AUROCAP\\(F_1\\)-maxAUROCAP\\(F_1\\)-maxAUPRO
MVTec-AD [3]RD4AD [10]94.696.595.296.148.653.891.1
SimpleNet [34]95.398.495.896.945.949.786.5
DeSTSeg [67]89.295.591.693.154.350.964.8
UniAD [60]†96.598.896.296.843.449.590.7
ReContrast [14]†98.399.497.697.160.261.593.2
DiAD [18]†97.299.096.596.852.655.590.7
ViTAD [65]†98.399.497.397.755.358.791.4
MambaAD [17]†98.699.697.897.756.359.293.1
Dinomaly (Ours)99.699.899.098.469.369.294.8
Dinomaly†99.799.899.298.469.369.294.7
VisA [70]RD4AD [10]92.492.489.698.138.042.691.8
SimpleNet [34]87.287.081.896.834.737.881.4
DeSTSeg [67]88.989.085.296.139.643.467.4
UniAD [60]†88.890.885.898.333.739.085.5
ReContrast [14]†95.596.492.098.547.950.691.9
DiAD [18]†86.888.385.196.026.133.075.2
ViTAD [65]†90.591.786.398.236.641.185.1
MambaAD [17]†94.394.589.498.539.444.091.0
Dinomaly (Ours)98.798.996.298.753.255.794.5
Dinomaly†98.999.096.498.853.855.894.5
Real-IAD [54]RD4AD [10]82.479.073.997.325.032.789.6
SimpleNet [34]57.253.461.575.72.86.539.0
DeSTSeg [67]82.379.273.294.637.941.740.6
UniAD [60]†83.080.974.397.321.129.286.7
ReContrast [14]†86.484.277.497.831.638.291.8
DiAD [18]†75.666.469.988.02.97.158.1
ViTAD [65]†82.780.273.797.224.332.384.8
MambaAD [17]†86.384.677.098.533.038.790.5
Dinomaly (Ours)89.386.880.298.842.847.193.9
Dinomaly†89.586.980.498.943.347.494.2
", + "image_path": "63abb526d89a5fc1376c69e0a3620b9eeaa44da1e212790f58926cb7f3502cf6.jpg" + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "table_body" + } + ], + "index": 1 + }, + { + "bbox": [ + 55, + 446, + 137, + 460 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 446, + 137, + 460 + ], + "spans": [ + { + "bbox": [ + 55, + 446, + 137, + 460 + ], + "type": "text", + "content": "4. Experiments" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 55, + 466, + 182, + 480 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 466, + 182, + 480 + ], + "spans": [ + { + "bbox": [ + 55, + 466, + 182, + 480 + ], + "type": "text", + "content": "4.1. Experimental Settings" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 55, + 485, + 295, + 628 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 485, + 295, + 628 + ], + "spans": [ + { + "bbox": [ + 55, + 485, + 295, + 628 + ], + "type": "text", + "content": "Datasets. MVtec-AD [3] contains 15 objects (5 texture classes and 10 object classes) with a total of 3,629 normal images as the training set and 1,725 images as the test set (467 normal, 1,258 anomalous). VisA [70] contains 12 objects. Training and test sets are split following the official splitting, resulting in 8,659 normal images in the training set and 2,162 images in the test set (962 normal, 1,200 anomalous). Real-IAD [54] is a large UAD dataset recently released, containing 30 distinct objects. We follow the official splitting that includes all views, resulting in 36,465 normal images in the training set and 114,585 images in the test set (63,256 normal, 51,329 anomalous)." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 55, + 630, + 295, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 630, + 295, + 714 + ], + "spans": [ + { + "bbox": [ + 55, + 630, + 295, + 714 + ], + "type": "text", + "content": "Metrics. Following prior works [17, 65], we adopt 7 evaluation metrics. Image-level anomaly detection performance is measured by the Area Under the Receiver Operator Curve (AUROC), Average Precision (AP), and " + }, + { + "bbox": [ + 55, + 630, + 295, + 714 + ], + "type": "inline_equation", + "content": "F_{1}" + }, + { + "bbox": [ + 55, + 630, + 295, + 714 + ], + "type": "text", + "content": " score under optimal threshold " + }, + { + "bbox": [ + 55, + 630, + 295, + 714 + ], + "type": "inline_equation", + "content": "(F_{1}" + }, + { + "bbox": [ + 55, + 630, + 295, + 714 + ], + "type": "text", + "content": "-max). Pixel-level anomaly localization is measured by AUROC, AP, " + }, + { + "bbox": [ + 55, + 630, + 295, + 714 + ], + "type": "inline_equation", + "content": "F_{1}" + }, + { + "bbox": [ + 55, + 630, + 295, + 714 + ], + "type": "text", + "content": "-max and the Area Under the Per-Region-Overlap (AUPRO). The results" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 313, + 448, + 476, + 459 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 448, + 476, + 459 + ], + "spans": [ + { + "bbox": [ + 313, + 448, + 476, + 459 + ], + "type": "text", + "content": "of a dataset is the average of all classes." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 313, + 460, + 555, + 628 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 460, + 555, + 628 + ], + "spans": [ + { + "bbox": [ + 313, + 460, + 555, + 628 + ], + "type": "text", + "content": "Implementation Details. ViT-Base/14 (patchsize=14) pre-trained by DINOv2-R [7] is used as the encoder by default. The drop rate of Noisy Bottleneck is 0.2 by default and increases to 0.4 on the diverse Real-IAD. Loose constraint with 2 groups is employed, and the anomaly map is given by the mean per-point cosine distance of the 2 groups. The input image is first resized to " + }, + { + "bbox": [ + 313, + 460, + 555, + 628 + ], + "type": "inline_equation", + "content": "448^2" + }, + { + "bbox": [ + 313, + 460, + 555, + 628 + ], + "type": "text", + "content": " and then center-cropped to " + }, + { + "bbox": [ + 313, + 460, + 555, + 628 + ], + "type": "inline_equation", + "content": "392^2" + }, + { + "bbox": [ + 313, + 460, + 555, + 628 + ], + "type": "text", + "content": ", so the feature map " + }, + { + "bbox": [ + 313, + 460, + 555, + 628 + ], + "type": "inline_equation", + "content": "(28^2)" + }, + { + "bbox": [ + 313, + 460, + 555, + 628 + ], + "type": "text", + "content": " is large enough for anomaly localization. StableAdamW optimizer [56] with AMSGrad [41] (more stable than AdamW [35] in training) is utilized with " + }, + { + "bbox": [ + 313, + 460, + 555, + 628 + ], + "type": "inline_equation", + "content": "lr = 2e - 3" + }, + { + "bbox": [ + 313, + 460, + 555, + 628 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 313, + 460, + 555, + 628 + ], + "type": "inline_equation", + "content": "\\beta = (0.9,0.999)" + }, + { + "bbox": [ + 313, + 460, + 555, + 628 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 313, + 460, + 555, + 628 + ], + "type": "inline_equation", + "content": "wd = 1e - 4" + }, + { + "bbox": [ + 313, + 460, + 555, + 628 + ], + "type": "text", + "content": ". The network is trained for 10,000 iterations (steps) on MVTec-AD and VisA, and 50,000 iterations on Real-IAD. Detailed settings are available in Appendix B." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 313, + 635, + 523, + 647 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 635, + 523, + 647 + ], + "spans": [ + { + "bbox": [ + 313, + 635, + 523, + 647 + ], + "type": "text", + "content": "4.2. Comparison to Multi-Class UAD SoTAs" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 313, + 654, + 554, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 654, + 554, + 713 + ], + "spans": [ + { + "bbox": [ + 313, + 654, + 554, + 713 + ], + "type": "text", + "content": "We compare the proposed Dinomaly with the most advanced UAD and MUAD methods [10, 14, 17, 18, 34, 65, 67]. Experimental results are presented in Table 1, where Dinomaly surpasses compared methods by a large margin on all datasets and all metrics. On the most" + } + ] + } + ], + "index": 9 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "type": "text", + "content": "20410" + } + ] + } + ], + "index": 10 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 5 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 82, + 91, + 529, + 174 + ], + "blocks": [ + { + "bbox": [ + 133, + 71, + 476, + 82 + ], + "lines": [ + { + "bbox": [ + 133, + 71, + 476, + 82 + ], + "spans": [ + { + "bbox": [ + 133, + 71, + 476, + 82 + ], + "type": "text", + "content": "Table 2. Performance under conventional class-separated UAD setting (%). n/a: not available." + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 82, + 91, + 529, + 174 + ], + "lines": [ + { + "bbox": [ + 82, + 91, + 529, + 174 + ], + "spans": [ + { + "bbox": [ + 82, + 91, + 529, + 174 + ], + "type": "table", + "html": "
MethodMVTec-AD [3]VisA [70]Real-IAD [54]
I-AUROCP-AUROCP-AUPROI-AUROCP-AUROCP-AUPROI-AUROCP-AUROCP-AUPRO
Dinomaly (MUAD)99.698.494.898.798.794.589.398.893.9
Dinomaly99.799.995.098.998.995.192.099.195.1
RD4AD [10]98.597.893.996.090.170.987.1n/a93.8
PatchCore [45]99.198.193.594.798.591.889.4n/a91.5
SimpleNet [34]99.698.190.097.198.290.788.5n/a84.6
", + "image_path": "cd538305d399bb60a141faeaeca01f08a86dc1a2c840fd10c9d92ee535e6db57.jpg" + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "table_body" + } + ], + "index": 1 + }, + { + "type": "table", + "bbox": [ + 82, + 213, + 529, + 362 + ], + "blocks": [ + { + "bbox": [ + 55, + 183, + 555, + 206 + ], + "lines": [ + { + "bbox": [ + 55, + 183, + 555, + 206 + ], + "spans": [ + { + "bbox": [ + 55, + 183, + 555, + 206 + ], + "type": "text", + "content": "Table 3. Ablations of Dinomaly elements on MVTec-AD (%). NB: Noisy Bottleneck. LA: Linear Attention. LC: Loose Constraint (2 groups). LL: Loose Loss. As MVTec-AD has reached saturation, we also present the results on VisA (Table A4)." + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 82, + 213, + 529, + 362 + ], + "lines": [ + { + "bbox": [ + 82, + 213, + 529, + 362 + ], + "spans": [ + { + "bbox": [ + 82, + 213, + 529, + 362 + ], + "type": "table", + "html": "
NBLALCLLImage-levelPixel-level
AUROCAPF1-maxAUROCAPF1-maxAUPRO
98.4199.0997.4197.1862.9663.8292.95
✓99.0699.5498.3197.6266.2266.7093.71
✓98.5499.2197.6297.2062.9463.7393.09
✓98.3599.0497.4397.1061.0562.7392.60
✓99.0399.4598.1997.6264.1064.9693.34
✓✓99.2799.6298.6397.8567.3667.3394.16
✓✓99.5099.7298.8798.1468.1668.2494.23
✓✓✓99.5299.7398.9298.2068.2568.3494.17
✓✓✓99.5799.7899.0098.2067.9368.2194.50
✓✓✓✓99.6099.7899.0498.3569.2969.1794.79
", + "image_path": "a326e4c362788dd358ee23012ebf6b13e0d43aae440ef876896212edf481cee2.jpg" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "table_body" + } + ], + "index": 3 + }, + { + "bbox": [ + 54, + 381, + 297, + 622 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 54, + 381, + 297, + 622 + ], + "spans": [ + { + "bbox": [ + 54, + 381, + 297, + 622 + ], + "type": "text", + "content": "widely used MVTec-AD, Dinomaly produces image-level performance of 99.6/99.8/99.0 (\\%) and pixel-level performance of 98.4/69.3/69.2/94.8, outperforming previous SoTAs by 1.0/0.2/1.2 and 0.7/9.1/7.7/1.6. This result declares that the image-level performance on the MVTec-AD dataset is nearly saturated under the MUAD setting. On the popular VisA, Dinomaly achieves image-level performance of 98.7/98.9/96.2 and pixel-level performance of 98.7/53.2/55.7/94.5, outperforming previous SoTAs by 3.2/2.5/4.2 and 0.2/5.3/5.1/2.6. On the Real-IAD that contains 30 classes, each with 5 camera views, we produce image-level and pixel-level performance of 89.3/86.8/80.2 and 98.8/42.8/47.1/93.9, outperforming previous SoTAs by 3.0/2.2/3.2 and 0.3/4.9/5.4/3.4, indicating our scalability to extremely complex scenarios. Per-class performances and qualitative visualization are presented in Appendix E and F. We also produce superior results on other popular UAD benchmarks, i.e., MPDD [24], BTAD [38], and UniMedical [66], with I-AUROC of 97.2, 95.4, and 84.9, respectively, as shown in Table A13 in Appendix." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 55, + 633, + 286, + 647 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 633, + 286, + 647 + ], + "spans": [ + { + "bbox": [ + 55, + 633, + 286, + 647 + ], + "type": "text", + "content": "4.3. Comparison to Class-Separated UAD SoTAs" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 55, + 653, + 296, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 653, + 296, + 715 + ], + "spans": [ + { + "bbox": [ + 55, + 653, + 296, + 715 + ], + "type": "text", + "content": "Dinomaly is also compared with class-separated SoTAs, as shown in Table 2. Dinomaly under MUAD setting is comparable to conventional methods [10, 34, 45] that build individual models for each class. On MVTec-AD and VisA, multi-class Dinomaly (first row) is subjected to nearly no" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 313, + 381, + 555, + 442 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 381, + 555, + 442 + ], + "spans": [ + { + "bbox": [ + 313, + 381, + 555, + 442 + ], + "type": "text", + "content": "performance drop compared to its class-separated counterpart (second row). On the complicated Real-IAD that involves more classes and views, multi-class Dinomaly suffers a moderate performance drop but is still comparable to class-separated SoTAs." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 313, + 453, + 408, + 466 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 453, + 408, + 466 + ], + "spans": [ + { + "bbox": [ + 313, + 453, + 408, + 466 + ], + "type": "text", + "content": "4.4. Ablation Study" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 313, + 472, + 556, + 628 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 472, + 556, + 628 + ], + "spans": [ + { + "bbox": [ + 313, + 472, + 556, + 628 + ], + "type": "text", + "content": "Overall Ablation. We conduct experiments to verify the effectiveness of the proposed elements, i.e., Noisy Bottleneck (NB), Linear Attention (LA), Loose Constraint (LC), and Loose Loss (LL). The already-powerful baseline (first row) is Dinomaly with noiseless MLP bottleneck, Softmax Attention, dense layer-to-layer supervision, and global cosine loss. This baseline is very similar to ViTAD [64] and the ViT version of RD4AD [10]. Results on MVtec-AD and VisA are shown in Table 3 and Table A4, respectively. NB and LL can directly contribute to the model performance. LA and LC boost the performance with the presence of NB. The use of LC is not solely beneficial because LC makes the reconstruction too easy without injected noise." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 313, + 630, + 556, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 630, + 556, + 715 + ], + "spans": [ + { + "bbox": [ + 313, + 630, + 556, + 715 + ], + "type": "text", + "content": "Model Scalability. Previous works [10, 60, 65] reported that anomaly detection networks do not follow the \"scaling law\". For example, RD4AD [10] found WideResNet50 better than WideResNet101 as the encoder backbone. ViTAD [65] found ViT-Small better than ViT-Base. On the contrary, as shown in Table 4, the performance of the proposed Dinomaly benefits from scaling. Dinomaly equipped with ViT" + } + ] + } + ], + "index": 10 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 317, + 758 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 317, + 758 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 317, + 758 + ], + "type": "text", + "content": "20411" + } + ] + } + ], + "index": 11 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 6 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 82, + 102, + 529, + 173 + ], + "blocks": [ + { + "bbox": [ + 55, + 70, + 555, + 94 + ], + "lines": [ + { + "bbox": [ + 55, + 70, + 555, + 94 + ], + "spans": [ + { + "bbox": [ + 55, + 70, + 555, + 94 + ], + "type": "text", + "content": "Table 4. Scaling of ViT model sizes on MVtec-AD (%). Im/s (Throughoutput, image per second) is measured on NVIDIA RTX3090 with batch size=16. Results on VisA and Real-IAD are shown in Table A3. †:default." + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 82, + 102, + 529, + 173 + ], + "lines": [ + { + "bbox": [ + 82, + 102, + 529, + 173 + ], + "spans": [ + { + "bbox": [ + 82, + 102, + 529, + 173 + ], + "type": "table", + "html": "
Arch.ParamsMACsIm/sImage-levelPixel-level
AUROCAPF1-maxAUROCAPF1-maxAUPRO
ViT-Small37.4M26.3G153.699.2699.6798.7298.0768.2967.7894.36
ViT-Base†148.0M104.7G58.199.6099.7899.0498.3569.2969.1794.79
ViT-Large275.3M413.5G24.299.7799.9299.4598.5470.5370.0495.09
", + "image_path": "be8fcebda125d516609706bb55b352c692dfdd1fd3e340fc1f33ba3d0c6c7ba7.jpg" + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "table_body" + } + ], + "index": 1 + }, + { + "type": "table", + "bbox": [ + 63, + 220, + 288, + 346 + ], + "blocks": [ + { + "bbox": [ + 55, + 189, + 295, + 213 + ], + "lines": [ + { + "bbox": [ + 55, + 189, + 295, + 213 + ], + "spans": [ + { + "bbox": [ + 55, + 189, + 295, + 213 + ], + "type": "text", + "content": "Table 5. Scaling input size on MVTec-AD (\\%). †: default. Compared methods yield degradation when increasing input size." + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 63, + 220, + 288, + 346 + ], + "lines": [ + { + "bbox": [ + 63, + 220, + 288, + 346 + ], + "spans": [ + { + "bbox": [ + 63, + 220, + 288, + 346 + ], + "type": "table", + "html": "
MethodInput SizeImage-LevelPixel-Level
RD4AD2562†94.6/96.5/96.196.1/48.6/53.8/91.1
320293.2/96.9/95.695.7/55.1/57.5/91.1
384291.9/96.2/95.094.9/52.1/55.3/90.8
ReContrast2562†98.3/99.4/97.697.1/60.2/61.5/93.2
320298.2/99.2/97.596.8/61.8/62.6/93.3
384295.2/98.0/96.496.5/57.7/59.5/92.6
Dinomaly280299.6/99.8/99.398.2/65.2/66.3/93.6
336299.6/99.8/99.298.3/67.2/67.8/94.2
3922†99.6/99.8/99.098.4/69.3/69.2/94.8
", + "image_path": "5f71ebbf272cad0fdc21a3279e913af6c14f1d181f64cbec2c34a88ad27b3447.jpg" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "table_body" + } + ], + "index": 3 + }, + { + "bbox": [ + 54, + 365, + 295, + 474 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 54, + 365, + 295, + 474 + ], + "spans": [ + { + "bbox": [ + 54, + 365, + 295, + 474 + ], + "type": "text", + "content": "Small has already produced state-of-the-art results. ViT-Large further boosts Dinomaly to an unprecedented higher record. This scalability enables users to choose an appropriate model size based on the computational resources available in their specific scenario. A comparison of computational costs with other methods is presented in Table A11. In addition, training schedule can also be scaled up for even better performance without increasing inference costs, as demonstrated in Figure 1 (Dinomaly" + }, + { + "bbox": [ + 54, + 365, + 295, + 474 + ], + "type": "inline_equation", + "content": "\\uparrow" + }, + { + "bbox": [ + 54, + 365, + 295, + 474 + ], + "type": "text", + "content": ")." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 54, + 474, + 295, + 594 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 54, + 474, + 295, + 594 + ], + "spans": [ + { + "bbox": [ + 54, + 474, + 295, + 594 + ], + "type": "text", + "content": "Input Scalability. Though it seems unfair to compare Dinomaly with previous works that take smaller images as input, we contend that increasing their input size not only fails to benefit but actively undermines their performance, especially for image-level detection performance, as shown in Table 5. Therefore, we follow the common comparison strategy based on \"optimum vs. optimum\". On the contrary, Dinomaly enjoys scaling input size for anomaly localization, while still producing SoTA performance given smaller images. Details are presented in Table A2 in Appendix." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 54, + 594, + 295, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 54, + 594, + 295, + 714 + ], + "spans": [ + { + "bbox": [ + 54, + 594, + 295, + 714 + ], + "type": "text", + "content": "ViT Foundations. We conduct extensive experiments to investigate the impact of diverse pre-trained ViT foundations, including DeiT [50], MAE [19], D-iGPT [44], MOCOV3 [6], DINO [4], iBot [69], DINOv2 [39], and DINOv2-R [7]. As shown in Figure 5, Dinomaly is robust to the choice of backbone. Almost all foundation models can produce SoTA-level results with image-level AUROC higher than " + }, + { + "bbox": [ + 54, + 594, + 295, + 714 + ], + "type": "inline_equation", + "content": "98\\%" + }, + { + "bbox": [ + 54, + 594, + 295, + 714 + ], + "type": "text", + "content": ". The only notable exception is MAE, which, without fine-tuning, was reported to be less effective across various unsupervised tasks, e.g., kNN and linear-" + } + ] + } + ], + "index": 6 + }, + { + "type": "image", + "bbox": [ + 318, + 190, + 549, + 305 + ], + "blocks": [ + { + "bbox": [ + 318, + 190, + 549, + 305 + ], + "lines": [ + { + "bbox": [ + 318, + 190, + 549, + 305 + ], + "spans": [ + { + "bbox": [ + 318, + 190, + 549, + 305 + ], + "type": "image", + "image_path": "7534438a9f00e66162f29b3e33941023cda6c2ee6fa9e8f5a27a44c64f62b210.jpg" + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 313, + 315, + 555, + 350 + ], + "lines": [ + { + "bbox": [ + 313, + 315, + 555, + 350 + ], + "spans": [ + { + "bbox": [ + 313, + 315, + 555, + 350 + ], + "type": "text", + "content": "Figure 5. Image-level AUROC of Dinomaly equipped with various ViT foundations, and their linear-probing accuracy on ImageNet. MIM: Masked Image Modeling. CL: Contrastive Learning." + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "image_caption" + } + ], + "index": 7 + }, + { + "bbox": [ + 313, + 373, + 555, + 481 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 373, + 555, + 481 + ], + "spans": [ + { + "bbox": [ + 313, + 373, + 555, + 481 + ], + "type": "text", + "content": "probing [39]. The optimal input size varies because the these backbones are pre-trained on different resolutions. Interestingly, we found the anomaly detection performance to be strongly correlated with the accuracy of ImageNet linear-probing (freeze backbone & only tune linear classifier) of the foundation model, suggesting the possibility of further improvement by simply adopting a future foundation model. Detailed results and analysis are presented in Appendix and Table A1." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 313, + 483, + 554, + 533 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 483, + 554, + 533 + ], + "spans": [ + { + "bbox": [ + 313, + 483, + 554, + 533 + ], + "type": "text", + "content": "Additional experiments and results are detailed in the Appendix C, encompassing evaluations of various pretrained foundations, ablation studies of each components, hyperparameter optimization, and other in-depth analyses." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 313, + 548, + 388, + 560 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 548, + 388, + 560 + ], + "spans": [ + { + "bbox": [ + 313, + 548, + 388, + 560 + ], + "type": "text", + "content": "5. Conclusion" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 313, + 570, + 555, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 570, + 555, + 713 + ], + "spans": [ + { + "bbox": [ + 313, + 570, + 555, + 713 + ], + "type": "text", + "content": "Dinomaly, a minimalistic UAD framework, is proposed to address the under-performed MUAD models in this paper. We present four key elements in Dinomaly, i.e., Foundation Transformer, Noisy MLP Bottleneck, Linear Attention, and Loose Reconstruction, that can boost the performance under the challenging MUAD setting without fancy modules and tricks. Extensive experiments on MVTec AD, VisA, and Real-IAD demonstrate our superiority over previous model-unified multi-class models and even recent class-separated models, indicating the feasibility of implementing a unified model in complicated scenarios free of severe performance degradation." + } + ] + } + ], + "index": 12 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "type": "text", + "content": "20412" + } + ] + } + ], + "index": 13 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 7 + }, + { + "para_blocks": [ + { + "bbox": [ + 56, + 72, + 153, + 85 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 72, + 153, + 85 + ], + "spans": [ + { + "bbox": [ + 56, + 72, + 153, + 85 + ], + "type": "text", + "content": "Acknowledgments" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 56, + 172, + 115, + 185 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 172, + 115, + 185 + ], + "spans": [ + { + "bbox": [ + 56, + 172, + 115, + 185 + ], + "type": "text", + "content": "References" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 57, + 193, + 296, + 713 + ], + "type": "list", + "angle": 0, + "index": 14, + "blocks": [ + { + "bbox": [ + 61, + 193, + 296, + 258 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 61, + 193, + 296, + 258 + ], + "spans": [ + { + "bbox": [ + 61, + 193, + 296, + 258 + ], + "type": "text", + "content": "[1] Samet Akcay, Amir Atapour-Abarghouei, and Toby P Breckon. Ganomaly: Semi-supervised anomaly detection via adversarial training. In Computer Vision-ACCV 2018: 14th Asian Conference on Computer Vision, Perth, Australia, December 2-6, 2018, Revised Selected Papers, Part III 14, pages 622-637. Springer, 2019. 1" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 62, + 259, + 295, + 303 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 259, + 295, + 303 + ], + "spans": [ + { + "bbox": [ + 62, + 259, + 295, + 303 + ], + "type": "text", + "content": "[2] Paul Bergmann, Sindy Löwe, Michael Fauser, David Sattlegger, and Carsten Steger. Improving unsupervised defect segmentation by applying structural similarity to autoencoders. arXiv preprint arXiv:1807.02011, 2018. 1" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 62, + 304, + 295, + 358 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 304, + 295, + 358 + ], + "spans": [ + { + "bbox": [ + 62, + 304, + 295, + 358 + ], + "type": "text", + "content": "[3] Paul Bergmann, Michael Fauser, David Sattlegger, and Carsten Steger. Mvtec ad-a comprehensive real-world dataset for unsupervised anomaly detection. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 9592-9600, 2019. 1, 2, 6, 7" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 62, + 359, + 294, + 414 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 359, + 294, + 414 + ], + "spans": [ + { + "bbox": [ + 62, + 359, + 294, + 414 + ], + "type": "text", + "content": "[4] Mathilde Caron, Hugo Touvron, Ishan Misra, Hervé Jégou, Julien Mairal, Piotr Bojanowski, and Armand Joulin. Emerging properties in self-supervised vision transformers. In Proceedings of the IEEE/CVF international conference on computer vision, pages 9650-9660, 2021. 3, 8, 2" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 62, + 415, + 295, + 458 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 415, + 295, + 458 + ], + "spans": [ + { + "bbox": [ + 62, + 415, + 295, + 458 + ], + "type": "text", + "content": "[5] Xinlei Chen and Kaiming He. Exploring simple siamese representation learning. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 15750-15758, 2021. 3" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 62, + 459, + 294, + 502 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 459, + 294, + 502 + ], + "spans": [ + { + "bbox": [ + 62, + 459, + 294, + 502 + ], + "type": "text", + "content": "[6] Xinlei Chen, Saining Xie, and Kaiming He. An empirical study of training self-supervised vision transformers. In Proceedings of the IEEE/CVF international conference on computer vision, pages 9640-9649, 2021. 3, 8, 2" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 62, + 503, + 294, + 536 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 503, + 294, + 536 + ], + "spans": [ + { + "bbox": [ + 62, + 503, + 294, + 536 + ], + "type": "text", + "content": "[7] Timothee Darcet, Maxime Oquab, Julien Mairal, and Piotr Bojanowski. Vision transformers need registers. arXiv preprint arXiv:2309.16588, 2023. 3, 6, 8, 1, 2" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 62, + 536, + 294, + 591 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 536, + 294, + 591 + ], + "spans": [ + { + "bbox": [ + 62, + 536, + 294, + 591 + ], + "type": "text", + "content": "[8] Thomas Defard, Aleksandr Setkov, Angelique Loesch, and Romaric Audigier. Padim: a patch distribution modeling framework for anomaly detection and localization. In International Conference on Pattern Recognition, pages 475-489. Springer, 2021. 1" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 62, + 592, + 295, + 624 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 592, + 295, + 624 + ], + "spans": [ + { + "bbox": [ + 62, + 592, + 295, + 624 + ], + "type": "text", + "content": "[9] David Dehaene and Pierre Eline. Anomaly localization by modeling perceptual features. arXiv preprint arXiv:2008.05369, 2020. 1" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 58, + 625, + 296, + 669 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 58, + 625, + 296, + 669 + ], + "spans": [ + { + "bbox": [ + 58, + 625, + 296, + 669 + ], + "type": "text", + "content": "[10] Hanqiu Deng and Xingyu Li. Anomaly detection via reverse distillation from one-class embedding. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 9737-9746, 2022. 5, 6, 7, 1, 8, 9, 10" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 57, + 670, + 295, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 670, + 295, + 713 + ], + "spans": [ + { + "bbox": [ + 57, + 670, + 295, + 713 + ], + "type": "text", + "content": "[11] Jia Deng, Wei Dong, Richard Socher, Li-Jia Li, Kai Li, and Li Fei-Fei. Imagenet: A large-scale hierarchical image database. In IEEE Conference on Computer Vision and Pattern Recognition, pages 248–255, 2009. 2" + } + ] + } + ], + "index": 13 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 57, + 73, + 554, + 713 + ], + "type": "list", + "angle": 0, + "index": 28, + "blocks": [ + { + "bbox": [ + 55, + 91, + 296, + 163 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 91, + 296, + 163 + ], + "spans": [ + { + "bbox": [ + 55, + 91, + 296, + 163 + ], + "type": "text", + "content": "The authors acknowledge supports from National Natural Science Foundation of China (U22A2051, 82027807), National Key Research and Development Program of China (2022YFC2405200), Tsinghua-Foshan Innovation Special Fund (2021THFS0104), and Institute for Intelligent Healthcare, Tsinghua University (2022ZLB001)." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 316, + 73, + 554, + 138 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 73, + 554, + 138 + ], + "spans": [ + { + "bbox": [ + 316, + 73, + 554, + 138 + ], + "type": "text", + "content": "[12] Alexey Dosovitskiy, Lucas Beyer, Alexander Kolesnikov, Dirk Weissenborn, Xiaohua Zhai, Thomas Unterthiner, Mostafa Dehghani, Matthias Minderer, Georg Heigold, Sylvain Gelly, et al. An image is worth 16x16 words: Transformers for image recognition at scale. arXiv preprint arXiv:2010.11929, 2020. 2, 3" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 316, + 140, + 553, + 183 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 140, + 553, + 183 + ], + "spans": [ + { + "bbox": [ + 316, + 140, + 553, + 183 + ], + "type": "text", + "content": "[13] Jia Guo, Shuai Lu, Lize Jia, Weihang Zhang, and Huiqi Li. Encoder-decoder contrast for unsupervised anomaly detection in medical images. IEEE Transactions on Medical Imaging, 2023. 1" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 317, + 184, + 553, + 237 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 184, + 553, + 237 + ], + "spans": [ + { + "bbox": [ + 317, + 184, + 553, + 237 + ], + "type": "text", + "content": "[14] Jia Guo, Shuai Lu, Lize Jia, Weihang Zhang, and Huiqi Li. Recontrast: Domain-specific anomaly detection via contrastive reconstruction. In Advances in Neural Information Processing Systems (NeurIPS), pages 10721-10740, 2023. 2, 3, 5, 6" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 316, + 239, + 553, + 293 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 239, + 553, + 293 + ], + "spans": [ + { + "bbox": [ + 316, + 239, + 553, + 293 + ], + "type": "text", + "content": "[15] Dongchen Han, Xuran Pan, Yizeng Han, Shiji Song, and Gao Huang. Flatten transformer: Vision transformer using focused linear attention. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 5961-5971, 2023. 5" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 316, + 294, + 553, + 347 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 294, + 553, + 347 + ], + "spans": [ + { + "bbox": [ + 316, + 294, + 553, + 347 + ], + "type": "text", + "content": "[16] Haoyang He, Jiangning Zhang, Hongxu Chen, Xuhai Chen, Zhishan Li, Xu Chen, Yabiao Wang, Chengjie Wang, and Lei Xie. Diad: A diffusion-based framework for multi-class anomaly detection. arXiv preprint arXiv:2312.06607, 2023. 1, 3" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 317, + 350, + 553, + 413 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 350, + 553, + 413 + ], + "spans": [ + { + "bbox": [ + 317, + 350, + 553, + 413 + ], + "type": "text", + "content": "[17] Haoyang He, Yuhu Bai, Jiangning Zhang, Qingdong He, Hongxu Chen, Zhenye Gan, Chengjie Wang, Xiangtai Li, Guanzhong Tian, and Lei Xie. Mambaad: Exploring state space models for multi-class unsupervised anomaly detection. arXiv preprint arXiv:2404.06564, 2024. 1, 3, 6, 7, 8, 10" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 316, + 415, + 553, + 470 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 415, + 553, + 470 + ], + "spans": [ + { + "bbox": [ + 316, + 415, + 553, + 470 + ], + "type": "text", + "content": "[18] Haoyang He, Jiangning Zhang, Hongxu Chen, Xuhai Chen, Zhishan Li, Xu Chen, Yabiao Wang, Chengjie Wang, and Lei Xie. A diffusion-based framework for multi-class anomaly detection. In Proceedings of the AAAI Conference on Artificial Intelligence, pages 8472-8480, 2024. 2, 6, 7, 8, 9, 10" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 317, + 472, + 553, + 524 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 472, + 553, + 524 + ], + "spans": [ + { + "bbox": [ + 317, + 472, + 553, + 524 + ], + "type": "text", + "content": "[19] Kaiming He, Xinlei Chen, Saining Xie, Yanghao Li, Piotr Dollar, and Ross Girshick. Masked autoencoders are scalable vision learners. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 16000-16009, 2022. 3, 8, 2, 5" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 316, + 525, + 553, + 558 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 525, + 553, + 558 + ], + "spans": [ + { + "bbox": [ + 316, + 525, + 553, + 558 + ], + "type": "text", + "content": "[20] Geoffrey Hinton, Oriol Vinyals, and Jeff Dean. Distilling the knowledge in a neural network. arXiv preprint arXiv:1503.02531, 2015. 5" + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 317, + 559, + 553, + 602 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 559, + 553, + 602 + ], + "spans": [ + { + "bbox": [ + 317, + 559, + 553, + 602 + ], + "type": "text", + "content": "[21] Geoffrey E Hinton, Nitish Srivastava, Alex Krizhevsky, Ilya Sutskever, and Ruslan R Salakhutdinov. Improving neural networks by preventing co-adaptation of feature detectors. arXiv preprint arXiv:1207.0580, 2012. 4" + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 316, + 603, + 553, + 647 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 603, + 553, + 647 + ], + "spans": [ + { + "bbox": [ + 316, + 603, + 553, + 647 + ], + "type": "text", + "content": "[22] Chaoqin Huang, Haoyan Guan, Aofan Jiang, Ya Zhang, Michael Spratling, and Yan-Feng Wang. Registration based few-shot anomaly detection. In European Conference on Computer Vision, pages 303-319. Springer, 2022. 7" + } + ] + } + ], + "index": 25 + }, + { + "bbox": [ + 316, + 647, + 553, + 691 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 647, + 553, + 691 + ], + "spans": [ + { + "bbox": [ + 316, + 647, + 553, + 691 + ], + "type": "text", + "content": "[23] Jongheon Jeong, Yang Zou, Taewan Kim, Dongqing Zhang, Avinash Ravichandran, and Onkar Dabeer. Winclip: Zero/few-shot anomaly classification and segmentation. arXiv preprint arXiv:2303.14814, 2023.7" + } + ] + } + ], + "index": 26 + }, + { + "bbox": [ + 316, + 692, + 553, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 692, + 553, + 713 + ], + "spans": [ + { + "bbox": [ + 316, + 692, + 553, + 713 + ], + "type": "text", + "content": "[24] Stepan Jezek, Martin Jonak, Radim Burget, Pavel Dvorak, and Milos Skotak. Deep learning-based defect detection of" + } + ] + } + ], + "index": 27 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 318, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 318, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 318, + 757 + ], + "type": "text", + "content": "20413" + } + ] + } + ], + "index": 29 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 8 + }, + { + "para_blocks": [ + { + "bbox": [ + 56, + 72, + 296, + 712 + ], + "type": "list", + "angle": 0, + "index": 14, + "blocks": [ + { + "bbox": [ + 76, + 72, + 296, + 116 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 76, + 72, + 296, + 116 + ], + "spans": [ + { + "bbox": [ + 76, + 72, + 296, + 116 + ], + "type": "text", + "content": "metal parts: evaluating current methods in complex conditions. In 2021 13th International Congress on Ultra Modern Telecommunications and Control Systems and Workshops (ICUMT), pages 66-71. IEEE, 2021. 7, 6" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 56, + 118, + 296, + 171 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 118, + 296, + 171 + ], + "spans": [ + { + "bbox": [ + 56, + 118, + 296, + 171 + ], + "type": "text", + "content": "[25] Xi Jiang, Jianlin Liu, Jinbao Wang, Qiang Nie, Kai Wu, Yong Liu, Chengjie Wang, and Feng Zheng. Softpatch: Unsupervised anomaly detection with noisy data. Advances in Neural Information Processing Systems, 35:15433-15445, 2022. 6, 7" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 56, + 173, + 295, + 225 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 173, + 295, + 225 + ], + "spans": [ + { + "bbox": [ + 56, + 173, + 295, + 225 + ], + "type": "text", + "content": "[26] Angelos Katharopoulos, Apoorv Vyas, Nikolaos Pappas, and François Fleuret. Transformers are rnns: Fast autoregressive transformers with linear attention. In International conference on machine learning, pages 5156-5165. PMLR, 2020. 5" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 56, + 228, + 295, + 270 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 228, + 295, + 270 + ], + "spans": [ + { + "bbox": [ + 56, + 228, + 295, + 270 + ], + "type": "text", + "content": "[27] Sungwook Lee, Seunghyun Lee, and Byung Cheol Song. Cfa: Coupled-hypersphere-based feature adaptation for target-oriented anomaly localization. IEEE Access, 10: 78446-78454, 2022. 1" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 56, + 272, + 295, + 304 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 272, + 295, + 304 + ], + "spans": [ + { + "bbox": [ + 56, + 272, + 295, + 304 + ], + "type": "text", + "content": "[28] Yujin Lee, Harin Lim, and Hyunsoo Yoon. Selfformaly: Towards task-agnostic unified anomaly detection. arXiv preprint arXiv:2307.12540, 2023. 3" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 56, + 305, + 295, + 359 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 305, + 295, + 359 + ], + "spans": [ + { + "bbox": [ + 56, + 305, + 295, + 359 + ], + "type": "text", + "content": "[29] Chun-Liang Li, Kihyuk Sohn, Jinsung Yoon, and Tomas Pfister. Cutpaste: Self-supervised learning for anomaly detection and localization. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 9664-9674, 2021. 1" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 56, + 361, + 295, + 413 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 361, + 295, + 413 + ], + "spans": [ + { + "bbox": [ + 56, + 361, + 295, + 413 + ], + "type": "text", + "content": "[30] Chen Liang, Jiahui Yu, Ming-Hsuan Yang, Matthew Brown, Yin Cui, Tuo Zhao, Boqing Gong, and Tianyi Zhou. Modulewise adaptive distillation for multimodality foundation models. Advances in Neural Information Processing Systems, 36, 2024. 5" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 56, + 415, + 295, + 447 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 415, + 295, + 447 + ], + "spans": [ + { + "bbox": [ + 56, + 415, + 295, + 447 + ], + "type": "text", + "content": "[31] Jiangqi Liu and Feng Wang. mixed attention auto encoder for multi-class industrial anomaly detection. arXiv preprint arXiv:2309.12700, 2023. 2" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 56, + 449, + 295, + 502 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 449, + 295, + 502 + ], + "spans": [ + { + "bbox": [ + 56, + 449, + 295, + 502 + ], + "type": "text", + "content": "[32] Wenqian Liu, Runze Li, Meng Zheng, Srikrishna Karanam, Ziyan Wu, Bir Bhanu, Richard J Radke, and Octavia Camps. Towards visually explaining variational autoencoders. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 8642-8651, 2020. 1" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 56, + 504, + 295, + 557 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 504, + 295, + 557 + ], + "spans": [ + { + "bbox": [ + 56, + 504, + 295, + 557 + ], + "type": "text", + "content": "[33] Ze Liu, Yutong Lin, Yue Cao, Han Hu, Yixuan Wei, Zheng Zhang, Stephen Lin, and Baining Guo. Swin transformer: Hierarchical vision transformer using shifted windows. In IEEE/CVF International Conference on Computer Vision, pages 10012-10022, 2021. 3" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 56, + 559, + 295, + 601 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 559, + 295, + 601 + ], + "spans": [ + { + "bbox": [ + 56, + 559, + 295, + 601 + ], + "type": "text", + "content": "[34] Zhikang Liu, Yiming Zhou, Yuansheng Xu, and Zilei Wang. Simplenet: A simple network for image anomaly detection and localization. arXiv preprint arXiv:2303.15140, 2023. 6, 7, 1, 8, 9, 10" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 56, + 603, + 295, + 624 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 603, + 295, + 624 + ], + "spans": [ + { + "bbox": [ + 56, + 603, + 295, + 624 + ], + "type": "text", + "content": "[35] Ilya Loshchilov and Frank Hutter. Decoupled weight decay regularization. arXiv preprint arXiv:1711.05101, 2017. 6" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 56, + 625, + 295, + 668 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 625, + 295, + 668 + ], + "spans": [ + { + "bbox": [ + 56, + 625, + 295, + 668 + ], + "type": "text", + "content": "[36] Ruiying Lu, YuJie Wu, Long Tian, Dongsheng Wang, Bo Chen, Xiyang Liu, and Ruimin Hu. Hierarchical vector quantized transformer for multi-class unsupervised anomaly detection. arXiv preprint arXiv:2310.14228, 2023. 1, 2, 4" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 56, + 670, + 295, + 712 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 670, + 295, + 712 + ], + "spans": [ + { + "bbox": [ + 56, + 670, + 295, + 712 + ], + "type": "text", + "content": "[37] Amira Ben Mabrouk and Ezzeddine Zagrouba. Abnormal behavior recognition for intelligent video surveillance systems: A review. Expert Systems with Applications, 91:480-491, 2018. 1" + } + ] + } + ], + "index": 13 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 316, + 72, + 555, + 712 + ], + "type": "list", + "angle": 0, + "index": 28, + "blocks": [ + { + "bbox": [ + 316, + 72, + 553, + 127 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 72, + 553, + 127 + ], + "spans": [ + { + "bbox": [ + 316, + 72, + 553, + 127 + ], + "type": "text", + "content": "[38] Pankaj Mishra, Riccardo Verk, Daniele Fornasier, Claudio Piciarelli, and Gian Luca Foresti. Vt-adl: A vision transformer network for image anomaly detection and localization. In 2021 IEEE 30th International Symposium on Industrial Electronics (ISIE), pages 01–06. IEEE, 2021. 7, 6" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 316, + 129, + 553, + 182 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 129, + 553, + 182 + ], + "spans": [ + { + "bbox": [ + 316, + 129, + 553, + 182 + ], + "type": "text", + "content": "[39] Maxime Oquab, Timothee Darcet, Theo Moutakanni, Huy Vo, Marc Szafraniec, Vasil Khalidov, Pierre Fernandez, Daniel Haziza, Francisco Massa, Alaaeldin El-Nouby, et al. Dinov2: Learning robust visual features without supervision. arXiv preprint arXiv:2304.07193, 2023. 3, 8, 2" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 316, + 184, + 553, + 216 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 184, + 553, + 216 + ], + "spans": [ + { + "bbox": [ + 316, + 184, + 553, + 216 + ], + "type": "text", + "content": "[40] Z Peng, L Dong, H Bao, Q Ye, and F Wei. Beit v2: Masked image modeling with vector-quantized visual tokenizers. arXiv preprint arXiv:2208.06366, 2022. 3, 2" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 316, + 217, + 553, + 248 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 217, + 553, + 248 + ], + "spans": [ + { + "bbox": [ + 316, + 217, + 553, + 248 + ], + "type": "text", + "content": "[41] Sashank J Reddi, Satyen Kale, and Sanjiv Kumar. On the convergence of adam and beyond. arXiv preprint arXiv:1904.09237, 2019. 6, 1" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 316, + 250, + 553, + 303 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 250, + 553, + 303 + ], + "spans": [ + { + "bbox": [ + 316, + 250, + 553, + 303 + ], + "type": "text", + "content": "[42] Tal Reiss, Niv Cohen, Liron Bergman, and Yedid Hoshen. Panda: Adapting pretrained features for anomaly detection and segmentation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 2806-2814, 2021. 7" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 316, + 304, + 553, + 348 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 304, + 553, + 348 + ], + "spans": [ + { + "bbox": [ + 316, + 304, + 553, + 348 + ], + "type": "text", + "content": "[43] Tal Reiss, Niv Cohen, Eliahu Horwitz, Ron Abutbul, and Yedid Hoshen. Anomaly detection requires better representations. In European Conference on Computer Vision, pages 56-68. Springer, 2022. 3, 7" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 316, + 350, + 553, + 392 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 350, + 553, + 392 + ], + "spans": [ + { + "bbox": [ + 316, + 350, + 553, + 392 + ], + "type": "text", + "content": "[44] Sucheng Ren, Zeyu Wang, Hongru Zhu, Junfei Xiao, Alan Yuille, and Cihang Xie. Rejuvenating image-gpt as strong visual representation learners. arXiv preprint arXiv:2312.02147, 2023.8, 2" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 316, + 393, + 555, + 448 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 393, + 555, + 448 + ], + "spans": [ + { + "bbox": [ + 316, + 393, + 555, + 448 + ], + "type": "text", + "content": "[45] Karsten Roth, Latha Pemula, Joaquin Zepeda, Bernhard Schölkopf, Thomas Brox, and Peter Gehler. Towards total recall in industrial anomaly detection. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 14318-14328, 2022. 7, 1, 3" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 316, + 449, + 553, + 513 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 449, + 553, + 513 + ], + "spans": [ + { + "bbox": [ + 316, + 449, + 553, + 513 + ], + "type": "text", + "content": "[46] Mohammadreza Salehi, Niousha Sadjadi, Soroosh Baselizadeh, Mohammad H Rohban, and Hamid R Rabiee. Multiresolution knowledge distillation for anomaly detection. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 14902-14912, 2021. 5, 1" + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 316, + 514, + 553, + 558 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 514, + 553, + 558 + ], + "spans": [ + { + "bbox": [ + 316, + 514, + 553, + 558 + ], + "type": "text", + "content": "[47] Thomas Schlegl, Philipp Seebock, Sebastian M Waldstein, Georg Langs, and Ursula Schmidt-Erfurth. f-anogan: Fast unsupervised anomaly detection with generative adversarial networks. Medical image analysis, 54:30-44, 2019. 1" + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 316, + 559, + 553, + 612 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 559, + 553, + 612 + ], + "spans": [ + { + "bbox": [ + 316, + 559, + 553, + 612 + ], + "type": "text", + "content": "[48] Zhuoran Shen, Mingyuan Zhang, Haiyu Zhao, Shuai Yi, and Hongsheng Li. Efficient attention: Attention with linear complexities. In Proceedings of the IEEE/CVF winter conference on applications of computer vision, pages 3531-3539, 2021. 5" + } + ] + } + ], + "index": 25 + }, + { + "bbox": [ + 316, + 614, + 553, + 667 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 614, + 553, + 667 + ], + "spans": [ + { + "bbox": [ + 316, + 614, + 553, + 667 + ], + "type": "text", + "content": "[49] Shelly Sheynin, Sagie Benaim, and Lior Wolf. A hierarchical transformation-discriminating generative model for few shot anomaly detection. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 8495-8504, 2021. 1" + } + ] + } + ], + "index": 26 + }, + { + "bbox": [ + 316, + 669, + 553, + 712 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 669, + 553, + 712 + ], + "spans": [ + { + "bbox": [ + 316, + 669, + 553, + 712 + ], + "type": "text", + "content": "[50] Hugo Touvron, Matthieu Cord, Matthijs Douze, Francisco Massa, Alexandre Sablayrolles, and Herv'e J'egou. Training data-efficient image transformers & distillation through attention. arXiv preprint arXiv:2012.12877, 2021. 8, 2" + } + ] + } + ], + "index": 27 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 318, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 318, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 318, + 757 + ], + "type": "text", + "content": "20414" + } + ] + } + ], + "index": 29 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 9 + }, + { + "para_blocks": [ + { + "bbox": [ + 56, + 72, + 294, + 712 + ], + "type": "list", + "angle": 0, + "index": 13, + "blocks": [ + { + "bbox": [ + 56, + 72, + 294, + 117 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 72, + 294, + 117 + ], + "spans": [ + { + "bbox": [ + 56, + 72, + 294, + 117 + ], + "type": "text", + "content": "[51] Ashish Vaswani, Noam Shazeer, Niki Parmar, Jakob Uszkoreit, Llion Jones, Aidan N Gomez, Lukasz Kaiser, and Illia Polosukhin. Attention is all you need. In Advances in Neural Information Processing Systems, pages 5998-6008, 2017. 2" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 56, + 118, + 294, + 173 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 118, + 294, + 173 + ], + "spans": [ + { + "bbox": [ + 56, + 118, + 294, + 173 + ], + "type": "text", + "content": "[52] Pascal Vincent, Hugo Larochelle, Yoshua Bengio, and Pierre-Antoine Manzagol. Extracting and composing robust features with denoising autoencoders. In Proceedings of the 25th international conference on Machine learning, pages 1096-1103, 2008. 4" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 56, + 175, + 294, + 229 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 175, + 294, + 229 + ], + "spans": [ + { + "bbox": [ + 56, + 175, + 294, + 229 + ], + "type": "text", + "content": "[53] Pascal Vincent, Hugo Larochelle, Isabelle Lajoie, Yoshua Bengio, and Pierre-Antoine Manzagol. Stacked denoising autoencoders: Learning useful representations in a deep network with a local denoising criterion. In Journal of Machine Learning Research, pages 3371-3408, 2010. 4" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 56, + 230, + 294, + 285 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 230, + 294, + 285 + ], + "spans": [ + { + "bbox": [ + 56, + 230, + 294, + 285 + ], + "type": "text", + "content": "[54] Chengjie Wang, Wenbing Zhu, Bin-Bin Gao, Zhenye Gan, Jianning Zhang, Zhihao Gu, Shuguang Qian, Mingang Chen, and Lizhuang Ma. Real-iad: A real-world multi-view dataset for benchmarking versatile industrial anomaly detection. arXiv preprint arXiv:2403.12580, 2024. 2, 6, 7, 3" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 56, + 286, + 294, + 330 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 286, + 294, + 330 + ], + "spans": [ + { + "bbox": [ + 56, + 286, + 294, + 330 + ], + "type": "text", + "content": "[55] Guodong Wang, Shumin Han, Errui Ding, and Di Huang. Student-teacher feature pyramid matching for anomaly detection. In The British Machine Vision Conference (BMVC), 2021. 1" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 56, + 331, + 294, + 385 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 331, + 294, + 385 + ], + "spans": [ + { + "bbox": [ + 56, + 331, + 294, + 385 + ], + "type": "text", + "content": "[56] Mitchell Wortsman, Tim Dettmers, Luke Zettlemoyer, Ari Morcos, Ali Farhadi, and Ludwig Schmidt. Stable and low-precision training for large-scale vision-language models. Advances in Neural Information Processing Systems, 36: 10271–10298, 2023. 6, 1" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 56, + 387, + 294, + 442 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 387, + 294, + 442 + ], + "spans": [ + { + "bbox": [ + 56, + 387, + 294, + 442 + ], + "type": "text", + "content": "[57] Zhenda Xie, Zheng Zhang, Yue Cao, Yutong Lin, Jianmin Bao, Zhuliang Yao, Qi Dai, and Han Hu. Simmim: A simple framework for masked image modeling. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 9653-9663, 2022. 3" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 56, + 444, + 294, + 476 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 444, + 294, + 476 + ], + "spans": [ + { + "bbox": [ + 56, + 444, + 294, + 476 + ], + "type": "text", + "content": "[58] Jie Yang, Yong Shi, and Zhiquan Qi. Dfr: Deep feature reconstruction for unsupervised anomaly segmentation. arXiv preprint arXiv:2012.07122, 2020. 5, 1" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 56, + 478, + 294, + 531 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 478, + 294, + 531 + ], + "spans": [ + { + "bbox": [ + 56, + 478, + 294, + 531 + ], + "type": "text", + "content": "[59] Haonan Yin, Guanlong Jiao, Qianhui Wu, Borne F Karlsson, Biqing Huang, and Chin Yew Lin. Lafite: Latent diffusion model with feature editing for unsupervised multi-class anomaly detection. arXiv preprint arXiv:2307.08059, 2023.1, 2, 3, 4" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 56, + 534, + 294, + 577 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 534, + 294, + 577 + ], + "spans": [ + { + "bbox": [ + 56, + 534, + 294, + 577 + ], + "type": "text", + "content": "[60] Zhiyuan You, Lei Cui, Yujun Shen, Kai Yang, Xin Lu, Yu Zheng, and Xinyi Le. A unified model for multi-class anomaly detection. arXiv preprint arXiv:2206.03687, 2022. 1, 2, 3, 4, 5, 6, 7, 8, 9, 10" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 56, + 579, + 294, + 623 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 579, + 294, + 623 + ], + "spans": [ + { + "bbox": [ + 56, + 579, + 294, + 623 + ], + "type": "text", + "content": "[61] Weihao Yu, Chenyang Si, Pan Zhou, Mi Luo, Yichen Zhou, Jiashi Feng, Shuicheng Yan, and Xinchao Wang. Metaformer baselines for vision. IEEE Transactions on Pattern Analysis and Machine Intelligence, 2023. 4" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 56, + 624, + 294, + 677 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 624, + 294, + 677 + ], + "spans": [ + { + "bbox": [ + 56, + 624, + 294, + 677 + ], + "type": "text", + "content": "[62] Vitjan Zavrtanik, Matej Kristan, and Danijel Skoc aj. Draem-a discriminatively trained reconstruction embedding for surface anomaly detection. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 8330-8339, 2021. 4, 1" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 56, + 680, + 294, + 712 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 680, + 294, + 712 + ], + "spans": [ + { + "bbox": [ + 56, + 680, + 294, + 712 + ], + "type": "text", + "content": "[63] Vitjan Zavrtanik, Matej Kristan, and Danijel Skocaj. Reconstruction by inpainting for visual anomaly detection. Pattern Recognition, 112:107706, 2021. 1" + } + ] + } + ], + "index": 12 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 316, + 72, + 553, + 441 + ], + "type": "list", + "angle": 0, + "index": 21, + "blocks": [ + { + "bbox": [ + 316, + 72, + 553, + 127 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 72, + 553, + 127 + ], + "spans": [ + { + "bbox": [ + 316, + 72, + 553, + 127 + ], + "type": "text", + "content": "[64] Dingwen Zhang, Guohai Huang, Qiang Zhang, Jungong Han, Junwei Han, Yizhou Wang, and Yizhou Yu. Exploring task structure for brain tumor segmentation from multimodality mr images. IEEE Transactions on Image Processing, 29:9032-9043, 2020. 7" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 316, + 129, + 553, + 183 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 129, + 553, + 183 + ], + "spans": [ + { + "bbox": [ + 316, + 129, + 553, + 183 + ], + "type": "text", + "content": "[65] Jiangning Zhang, Xuhai Chen, Yabiao Wang, Chengjie Wang, Yong Liu, Xiangtai Li, Ming-Hsuan Yang, and Dacheng Tao. Exploring plain vit reconstruction for multiclass unsupervised anomaly detection. arXiv preprint arXiv:2312.07495, 2023. 3, 5, 6, 7" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 316, + 185, + 553, + 239 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 185, + 553, + 239 + ], + "spans": [ + { + "bbox": [ + 316, + 185, + 553, + 239 + ], + "type": "text", + "content": "[66] Jiangning Zhang, Haoyang He, Zhenye Gan, Qingdong He, Yuxuan Cai, Zhucun Xue, Yabiao Wang, Chengjie Wang, Lei Xie, and Yong Liu. Ader: A comprehensive benchmark for multi-class visual anomaly detection. arXiv preprint arXiv:2406.03262, 2024. 7, 2, 6" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 316, + 240, + 553, + 295 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 240, + 553, + 295 + ], + "spans": [ + { + "bbox": [ + 316, + 240, + 553, + 295 + ], + "type": "text", + "content": "[67] Xuan Zhang, Shiyu Li, Xi Li, Ping Huang, Jiulong Shan, and Ting Chen. Destseg: Segmentation guided denoising student-teacher for anomaly detection. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 3914–3923, 2023. 4, 6, 1, 7, 8, 9, 10" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 316, + 297, + 553, + 340 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 297, + 553, + 340 + ], + "spans": [ + { + "bbox": [ + 316, + 297, + 553, + 340 + ], + "type": "text", + "content": "[68] Ying Zhao. Omnial: A unified cnn framework for unsupervised anomaly localization. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 3924-3933, 2023. 1, 3" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 316, + 342, + 553, + 384 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 342, + 553, + 384 + ], + "spans": [ + { + "bbox": [ + 316, + 342, + 553, + 384 + ], + "type": "text", + "content": "[69] Jinghao Zhou, Chen Wei, Huiyu Wang, Wei Shen, Cihang Xie, Alan Yuille, and Tao Kong. ibot: Image bert pre-training with online tokenizer. arXiv preprint arXiv:2111.07832, 2021. 3, 8, 2" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 316, + 386, + 553, + 441 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 386, + 553, + 441 + ], + "spans": [ + { + "bbox": [ + 316, + 386, + 553, + 441 + ], + "type": "text", + "content": "[70] Yang Zou, Jongheon Jeong, Latha Pemula, Dongqing Zhang, and Onkar Dabeer. Spot-the-difference self-supervised pretraining for anomaly detection and segmentation. In European Conference on Computer Vision, pages 392-408. Springer, 2022. 2, 6, 7, 3" + } + ] + } + ], + "index": 20 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "text", + "content": "20415" + } + ] + } + ], + "index": 22 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 10 + } + ], + "_backend": "vlm", + "_version_name": "2.6.4" +} \ No newline at end of file diff --git a/2025/DirectTriGS_ Triplane-based Gaussian Splatting Field Representation for 3D Generation/cb304427-02fe-443c-baf8-93b01f1651af_content_list.json b/2025/DirectTriGS_ Triplane-based Gaussian Splatting Field Representation for 3D Generation/cb304427-02fe-443c-baf8-93b01f1651af_content_list.json new file mode 100644 index 0000000000000000000000000000000000000000..be90c85b2e13105e3d4f90df0a93a01ef6cf19f5 --- /dev/null +++ b/2025/DirectTriGS_ Triplane-based Gaussian Splatting Field Representation for 3D Generation/cb304427-02fe-443c-baf8-93b01f1651af_content_list.json @@ -0,0 +1,1665 @@ +[ + { + "type": "text", + "text": "DirectTriGS: Triplane-based Gaussian Splitting Field Representation for 3D Generation", + "text_level": 1, + "bbox": [ + 112, + 128, + 883, + 174 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Xiaoliang Ju, Hongsheng Li $^{1,2}$ $^{1}$ CUHK MMLab $^{2}$ CPII under InnoHK", + "bbox": [ + 328, + 202, + 669, + 237 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "akira@link.cuhk.edu.hk", + "bbox": [ + 397, + 242, + 594, + 255 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Abstract", + "text_level": 1, + "bbox": [ + 246, + 291, + 326, + 306 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "We present DirectTriGS, a novel framework designed for 3D object generation with Gaussian Splatting (GS). GS-based rendering for 3D content has gained considerable attention recently. However, there has been limited exploration in directly generating 3D Gaussians compared to traditional generative modeling approaches. The main challenge lies in the complex data structure of GS represented by discrete point clouds with multiple channels. To overcome this challenge, we propose employing the triplane representation, which allows us to represent Gaussian Splatting as an image-like continuous field. This representation effectively encodes both the geometry and texture information, enabling smooth transformation back to Gaussian point clouds and rendering into images by a TriRenderer, with only 2D supervisions. The proposed TriRenderer is fully differentiable, so that the rendering loss can supervise both texture and geometry encoding. Furthermore, the triplane representation can be compressed using a Variational Autoencoder (VAE), which can subsequently be utilized in latent diffusion to generate 3D objects. The experiments demonstrate that the proposed generation framework can produce high-quality 3D object geometry and rendering results in the text-to-3D task.", + "bbox": [ + 89, + 324, + 485, + 672 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "1. Introduction", + "text_level": 1, + "bbox": [ + 89, + 707, + 220, + 722 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Neural rendering has grown to a focal point in rendering techniques in recent years, as it achieves more realistic rendering effects by leveraging the great expressiveness of neural network. The representatives are neural radiance field (NeRF) [32] and the newly emerged Gaussian Splatting (GS) [21]. However, when applied to the field of 3D generation, the slow rendering and training speed become a strong limitation for NeRF. While GS is more flexible with greater rendering efficiency and editability, few works have addressed the challenge of direct GS generation in 3D due to its complex data structure. In this paper, we propose a", + "bbox": [ + 89, + 734, + 483, + 902 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "novel approach called DirectTriGS, which encodes GS as triplane representation, and introduce its corresponding renderer, TriRenderer. Subsequently, we apply latent diffusion on the Triplane representation to generate high-quality GS objects.", + "bbox": [ + 511, + 292, + 906, + 367 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Gaussian splatting uses multi-channel point cloud of \"splats\" to describe the 3D contents. With a differentiable splats rasterizer, GS has fast rendering speed. However, the sparsity, multiple channels, and uneven distributed density of 3D GS bring great difficulty for 3D generation. We propose to encode GS contents as multi-channel triplane representations, which have been shown to possess favorable properties for representing 3D geometry or NeRF, as demonstrated in previous works Shue et al. [40], Wu et al. [47]. In our work, we leverage the triplane representation to encode both the geometry information and other GS attributes in two separate groups of channels. By training on a dataset of 3D objects, we obtain a shared TriRenderer that is capable of decoding different triplane representations to GS and then render it to image. TriRenderer is fully differentiable, enabling the use of only 2D rendering loss to supervise both the texture and geometry of 3D GS. The advantage of using Triplane to represent GS are two folds. First, it leads to high memory efficiency compared with dense voxels and is expressive enough for generating various 3D GS. Second, Triplane is more compatible with the convolution-based encoders compared with the original sparse GS point cloud, which require specifically designed networks on processing sparse point clouds.", + "bbox": [ + 511, + 369, + 908, + 733 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "We follow stable diffusion [36] to train DirectTriGS on the proposed triplane representation. Specifically, a VAE is designed to further convert the Triplanes into latent code. We use two separate decoders to decouple the decoding of geometry and GS attributes. Second, we roll out the triplane latent to an expanded multi-channel image, then exploit staged latent diffusion to do generation. Two-stage diffusion is employed to generate the geometry and the corresponding GS appearance. Finally, score distillation sampling (SDS) [35] is adopted as a optional post-processing to refine or restyle the generated 3D objects.", + "bbox": [ + 511, + 734, + 910, + 902 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "CVF", + "bbox": [ + 106, + 2, + 181, + 42 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore.", + "bbox": [ + 236, + 0, + 810, + 46 + ], + "page_idx": 0 + }, + { + "type": "page_number", + "text": "16229", + "bbox": [ + 480, + 944, + 519, + 957 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Our contribution can summarize as follows. 1) We propose a Triplane representation for direct 3D GS generation, which is memory efficient to derive GS point clouds with vivid rendering. 2) We design a fully differentiable TriRenderer to enable the end-to-end training of triplane representations, with only 2D supervisions. 3) We develop a Triplane-based GS generation framework DirectTriGS that incorporates a specially designed VAE, latent diffusion module, and a SDS based refiner. 4) Experiments demonstrate our method produces competitive performance in both 3D geometry and multi-view rendering quality in text-to-3D tasks.", + "bbox": [ + 89, + 90, + 480, + 271 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "2. Related Works", + "text_level": 1, + "bbox": [ + 89, + 286, + 238, + 301 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "3D Object Generation. 1) Multi-view images-to-3D. The first route generates multi-view colored 3D image and then reconstruct the 3D shape and project image to texture, or use score distillation sampling to build 3D object from images produced by 2D diffusion, such as [6, 13, 27, 35, 39, 41]. Since these approaches are actually 2D generator, it is difficult to maintain 3D consistency, which may lead to Janus problem. 2) Image-to-3D. The second route attempt to retrieve 3D geometry from a single input image, such as [25, 28, 53]. With the certain one image as input, the users may be more concerned with the reconstruction quality. 3)Text-to-3D. The third group mainly operate in 3D, which can be clustered by different 3D representation. For example, [10, 11] are able to generate high-quality 3D textured meshes using differentiable rendering; [34, 46] generate colored pointcloud; [1, 14, 23, 31] generate NeRF volumes; [18, 47] generate SDF volumes.", + "bbox": [ + 89, + 311, + 482, + 566 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Remark. The input conditions of these three generation methods are gradually becoming abstract, corresponding to increasing difficulty of the generation task, commonly resulting in a gradual decline in the generation quality.", + "bbox": [ + 89, + 568, + 482, + 628 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Gaussian Splatting. Gaussian Splatting [21] exploits multi-channel pointcloud of splats to describe the 3D content, which allows real-time rendering of photorealistic scenes. Recently, there emerges so many research focusing on its reconstruction quality enhancement and algorithm optimization [5, 8, 15, 29, 51]. Another related active area includes 4D dynamic GS modeling [3, 44, 49], scene editing [2, 9] and its applications like SLAM [30, 48].", + "bbox": [ + 89, + 628, + 482, + 750 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "As for GS generation, some works exploit multi-view image based route such as [6, 41], where the image generator output image from required views to enhance the GS rendering in a recursive manner. Rare works explore the GS generation directly in 3D space, such as [53] generates a point cloud using a Transformer structure with 2D image as input, and build a mapping of image to tri-plane encoding of GS attribute; GaussianCube [52] proposes to use optimal transport to model the 3D GS for text-to-3D generation; [16] combine explicit anchor points with implicit", + "bbox": [ + 89, + 750, + 482, + 901 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "triplane to directly generate GS points.", + "bbox": [ + 511, + 90, + 771, + 107 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "3. Revisiting Gaussian Splitting", + "text_level": 1, + "bbox": [ + 511, + 118, + 785, + 136 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Gaussian Splatting [21] uses a set of Gaussian points to describe a 3D object. The Gaussian points are defined by a full 3D covariance matrix $\\pmb{\\Sigma}$ in a world space as $G(\\mathbf{x}) = e^{-\\frac{1}{2}\\mathbf{x}^T\\pmb{\\Sigma}^{-1}\\mathbf{x}}$ , centered at point (mean) $\\mathbf{p} \\in \\mathbb{R}^3$ . Each point is with a opacity scalar $\\alpha \\in [0,1]$ for blending and a series of Spherical Harmonics (SH) coefficients to correctly capture the view-dependent appearance of the scene. The number of SH coefficients are $3 \\times (n + 1)^2$ , where $n$ denotes the SH order, and higher $n$ corresponds to more accurate view-dependent appearance.", + "bbox": [ + 511, + 143, + 903, + 297 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "During rendering, the Gaussians are project to 2D given a viewing transformation $\\mathbf{W}$ . The covariance is transformed as $\\boldsymbol{\\Sigma}^{\\prime} = \\mathbf{J}\\mathbf{W}\\boldsymbol{\\Sigma}\\mathbf{W}^{T}\\mathbf{J}^{T}$ , where $\\mathbf{J}$ is the Jacobian of the affine approximation of the projective transformation. To ensure $\\boldsymbol{\\Sigma}$ to be semi-positive in the whole training process, $\\boldsymbol{\\Sigma}$ can be represented by a Cholesky decomposition as $\\boldsymbol{\\Sigma} = \\mathbf{R}\\mathbf{S}\\mathbf{S}^{T}\\mathbf{R}^{T}$ , represented by a tuple $s = (s_{1}, s_{2}, s_{3})$ for scaling and a unnormalized quaternion $\\mathbf{q} \\in \\mathbb{R}^{4}$ for rotation. In summary, each Gaussian point are a union set of position $\\mathbf{p}$ and its GS attributes:", + "bbox": [ + 511, + 297, + 905, + 448 + ], + "page_idx": 1 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {G} _ {\\mathbf {p}} := \\{s, \\mathbf {q}, \\mathrm {S H}, \\alpha \\}. \\tag {1}\n$$\n", + "text_format": "latex", + "bbox": [ + 637, + 458, + 903, + 474 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "GS pointcloud can be efficiently rendered by a rasterization-based splatting renderer [21], which is fully differentiable.", + "bbox": [ + 511, + 484, + 903, + 515 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "4. Methodology", + "text_level": 1, + "bbox": [ + 511, + 527, + 648, + 545 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Our motivation is to directly generate 3D GS with fast speed by utilizing the proposed triplane-based GS representation. While 2D-lifting methods like SDS-based approaches yield intricate results, they are time-intensive and susceptible to Janus problem. Direct generation, notably with fast sampling, can complete 3D generation in under a minute, contrasting with SDS methods that take over 10 minutes or longer, such as [4][43], posing challenges for users. Therefore, direct 3D generation is more practical, with 2D-lifting reserved for enhancing texture details when needed.", + "bbox": [ + 511, + 551, + 905, + 703 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "The procedures of our generation framework can be divide into 4 stages as shown in Fig.1 (a). Firstly, we encode the 3D objects into triplanes using the TriRenderer. The triplane encoding and TriRenderer training can be concurrently executed with only the multi-view RGBA image and camera poses as supervision. Secondly, a VAE is trained to compress the triplanes to latent space, enabling effective capture of high-level information for subsequent modules. Thirdly, a diffusion model is trained on the triplane latent code. During inference, the diffusion model produces the latent code, which can be decoded back to its corresponding triplane by the VAE decoder. Subsequently, the TriRenderer decodes the triplane into the standard GS point", + "bbox": [ + 511, + 704, + 906, + 900 + ], + "page_idx": 1 + }, + { + "type": "page_number", + "text": "16230", + "bbox": [ + 480, + 944, + 517, + 955 + ], + "page_idx": 1 + }, + { + "type": "image", + "img_path": "images/0f6b3129f9dafeb1ea97a8c07c5e1e6f91700926f675d7de8e63579c4810257d.jpg", + "image_caption": [ + "(a) The overall procedure of our 3D content generation framework." + ], + "image_footnote": [], + "bbox": [ + 135, + 89, + 307, + 142 + ], + "page_idx": 2 + }, + { + "type": "image", + "img_path": "images/e89ee7f07caa7332f781434de1779f2cd6006aa281d39e6a1c03fab792483ba9.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 321, + 90, + 491, + 142 + ], + "page_idx": 2 + }, + { + "type": "image", + "img_path": "images/ac0245851813463f977695cdcb7376eb06ba9d71159ac02010405b1d656ef03f.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 504, + 90, + 676, + 142 + ], + "page_idx": 2 + }, + { + "type": "image", + "img_path": "images/698df0cca85d4e2868585e5d2158dea4c7d18ef59fe1938a7f12036db508cf97.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 689, + 90, + 862, + 142 + ], + "page_idx": 2 + }, + { + "type": "image", + "img_path": "images/a8f4465db2ff41f986e7112ab90158b5fd05a67645196976b27a8d94f6dc0e1c.jpg", + "image_caption": [ + "Figure 1. The overview of the proposed 3D GS generation framework and the core component TriRenderer. We use triplane to represent 3D objects, and design a TriRenderer to decode triplane to GS pointcloud and render multi-view images." + ], + "image_footnote": [ + "(b) The forward pass of the TriRenderer pipeline: rendering a Triolane to multi-view images." + ], + "bbox": [ + 133, + 171, + 862, + 344 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "cloud. Finally, an SDS-based refiner can be utilized as an optional post-processing step to enhance or restyle the generated 3D object. Leveraging the pretrained 2D diffusion model within the SDS, the generated GS can be further improved to achieve a more detailed appearance.", + "bbox": [ + 89, + 425, + 483, + 500 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "4.1. Triplane-based GS Field", + "text_level": 1, + "bbox": [ + 89, + 511, + 313, + 527 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "We employ triplanes to convert the discrete multi-channel GS points into a continuous field, tailored for more efficient encoding in the subsequent generative model. Triplanes are 3 planes formed by every 2 axes of $x,y,z$ axes in 3D space, where every 3D point $\\mathbf{p}$ can query these planes to get corresponding features $F_{\\mathbf{p}} = (F_{xy},F_{xz},F_{yz})$ using orthogonal projection. In real implementation, triplane is a tensor with the size of $3\\times H\\times H\\times C$ , where $H$ denote the resolution along every axis and $C$ is the feature channel. The query of point with arbitrary continuous coordinate should be a bi-linear interpolation in the triplane grids. Considering GS is a set of points with multiple channels of attributes, it is natural to use triplane for GS encoding. If the triplane feature and GS attributes of a point can be transformed to each other, the sparse GS pointcloud of every object can be represented as continuous triplane GS field for further encoding. We use separate channels to encode 3D geometry and other GS attributes as described in Eq. 1, which is an experimental setting for better convergence in the training.", + "bbox": [ + 89, + 534, + 483, + 821 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "4.2. TriRenderer", + "text_level": 1, + "bbox": [ + 89, + 832, + 223, + 847 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "The TriRenderer serves as the crucial component for converting a 3D object into a triplane-based GS field. It acts as a fully differentiable bridge connecting the GS field with", + "bbox": [ + 89, + 854, + 483, + 900 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "rendering, enabling the optimization from 2D images to geometry and appearance encoding.", + "bbox": [ + 511, + 425, + 903, + 455 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "As depicted in Fig. 1 (b), the TriRenderer comprises a geometry branch and a GS attribute branch, each equipped with independent decoders to decode the triplane. The geometry branch is responsible for retrieve the geometry as triangular mesh from triplane, with a surface sampler for GS pointcloud sampling. Then GS attribute branch uses the obtained pointcloud to query the triplane to obtain the GS attribute corresponding to each point. In this way, the GS pointcloud in the original format is retrieved, and can be rendered using the original GS renderer. It is worth mentioning that all triplanes of different object share a common TriRenderer, which ensures the features on different triplanes subject to a similar distribution.", + "bbox": [ + 511, + 457, + 906, + 654 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Geometry Branch. We use signed distance function (SDF) to represent the geometry, and allow every vertex deform as [37], so that the geometry branch decodes $F_{p}$ to a SDF volume and its vertex deformation. By query all grid coordinates of the designated resolution $L \\times L \\times L$ , the deformable SDF volume is reconstructed. Then we exploit a differentiable Marching Cube algorithm FlexiCubes [38] to extract triangular mesh from the SDF volume.", + "bbox": [ + 511, + 656, + 905, + 777 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Surface Sampling. Considering that GS pointcloud is generally gathered on the surface of objects, we randomly sample GS points on the faces of triangular mesh with barycentric coordinates. We expect that every splat is flat and with a normal consistent with its source triangular face. Assume a face is formed by vertices $< v_{a}, v_{b}, v_{c} >$ , and vector $\\mathbf{v}_{1} = (v_{a} - v_{b})$ , $\\mathbf{v}_{2} = (v_{b} - v_{c})$ , the face normal can be calculated as $\\mathbf{n} = \\mathbf{v}_{1} \\times \\mathbf{v}_{2}$ . Then the rotation matrix of every", + "bbox": [ + 511, + 779, + 908, + 901 + ], + "page_idx": 2 + }, + { + "type": "page_number", + "text": "16231", + "bbox": [ + 480, + 944, + 517, + 955 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "splat drawn from this face can be formulated as", + "bbox": [ + 89, + 90, + 406, + 104 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\mathbf {R} = \\left[ \\mathbf {v} _ {1}, \\mathbf {n} \\times \\mathbf {v} _ {1}, \\mathbf {n} \\right], \\tag {2}\n$$\n", + "text_format": "latex", + "bbox": [ + 212, + 116, + 482, + 133 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "which can be transform to $\\mathbf{q}$ in Eq.1 by standard matrix-to-quaternion algorithm. To make the splats flat, $s_3$ in Eq.1 is fixed to an infinite small value.", + "bbox": [ + 89, + 143, + 482, + 188 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "To further reduce the computational load in the subsequent GS rendering process, only faces oriented toward the camera are sampled. These face indices can be obtained through fast rasterization of the mesh faces. Note this rasterization is not required to be differentiable, and we use the library Nvdiffrast [22] for implementation.", + "bbox": [ + 89, + 189, + 483, + 280 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "GS Attribute Branch. By querying the triplane GS channels with the sampled pointcloud, we can obtain the GS feature and use the GS attribute branch to transform them to the rest GS attributes. Then the GS renderer can render GS pointcloud to multi-view images. Since different GS attributes have different numerical scales and distributions, we customize individual headers to decode them respectively, which is similar to [53].", + "bbox": [ + 89, + 280, + 483, + 400 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Training. The training consists of 2 stages. In the first stage, a small batch of data is used to train the triplanes and the TriRenderer together. The second stage involves the whole dataset to train all triplanes with the parameters of TriRenderer frozen. In this way, we can handle large dataset by distributed training in the second stage, and the same TriRenderer can be shared by all triplanes. There is no need to pre-train the GS pointcloud using original GS training as [21].", + "bbox": [ + 89, + 401, + 483, + 537 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "The only required supervision data is $N$ multi-view images $\\{I_i\\}_{i=1}^N$ with camera poses, and the training losses consist of the rendering loss $L_{\\mathrm{render}}$ and the geometric regularization $L_{\\mathrm{geo}}$ . The rendering loss is a weighted sum as", + "bbox": [ + 89, + 537, + 483, + 598 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\nL _ {\\text {r e n d e r}} = w _ {1} L _ {\\text {a l p h a}} + w _ {2} L _ {\\text {r g b}} + w _ {3} L _ {\\text {p i p s}}, \\tag {3}\n$$\n", + "text_format": "latex", + "bbox": [ + 143, + 612, + 482, + 628 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "where $L_{\\mathrm{alpha}}$ is a $L_{1}$ loss on the alpha map, namely the silhouette image loss; $L_{\\mathrm{rgb}}$ is a combination of $L_{1}$ pixel loss and SSIM loss [42] between rendered image $\\hat{I}_i$ and ground truth $I_i$ :", + "bbox": [ + 89, + 633, + 483, + 696 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\nL _ {\\mathrm {r g b}} = \\sum_ {i = 1} ^ {N} (1 - \\beta) \\| I _ {i} - \\hat {I} _ {i} \\| _ {1} + \\beta \\operatorname {S S I M} \\left(I _ {i}, \\hat {I} _ {i}\\right), \\tag {4}\n$$\n", + "text_format": "latex", + "bbox": [ + 117, + 705, + 482, + 748 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "and $L_{\\mathrm{pips}}$ is the perceptual loss [17]. $w_{1}, w_{2}, w_{3}$ and $\\beta$ are all weighting factors determined by experiments.", + "bbox": [ + 89, + 756, + 482, + 786 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "The geometric regularization loss on the SDF volume $V$ is defined as", + "bbox": [ + 89, + 786, + 483, + 816 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\nL _ {\\text {g e o}} = \\gamma_ {1} L _ {\\text {d e v}} + \\gamma_ {2} L _ {\\text {w e i g h t i n g}} + \\gamma_ {3} L _ {\\text {C E}} + \\gamma_ {4} L _ {\\text {s i g n}}, \\tag {5}\n$$\n", + "text_format": "latex", + "bbox": [ + 102, + 828, + 482, + 845 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "where the $L_{\\mathrm{dev}}$ and $L_{\\mathrm{weighting}}$ are defined by Flex-cubes [38] to regularize the extracted connectivity and the weighting scale of SDF vertices. To penalize the sign", + "bbox": [ + 89, + 854, + 483, + 901 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "changes on all grid edges, we follow [33] to define $L_{\\mathrm{CE}}$ as", + "bbox": [ + 511, + 90, + 903, + 119 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\nL _ {\\mathrm {C E}} = \\sum_ {(s _ {a}, s _ {b}) \\in \\epsilon} \\operatorname {C E} \\left(\\sigma \\left(s _ {a}\\right), \\operatorname {s i g n} \\left(s _ {b}\\right)\\right), \\tag {6}\n$$\n", + "text_format": "latex", + "bbox": [ + 581, + 119, + 903, + 152 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "where CE denotes cross entropy; $\\epsilon$ is the set of all edges connecting vertices with different signs; and $\\sigma$ is the sigmoid function. Finally, $L_{\\mathrm{sign}}$ is designed to prevent the SDF volume from being trapped in an empty shape, i.e. a fully positive or fully negative SDF volume,", + "bbox": [ + 511, + 155, + 906, + 231 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\nL _ {\\text {s i g n}} = \\delta (V) M (| V |), \\tag {7}\n$$\n", + "text_format": "latex", + "bbox": [ + 633, + 239, + 903, + 257 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "where $\\delta (V) = 1$ if $V$ is empty, otherwise $\\delta (V) = 0$ . $M(|V|)$ denotes the mean of the absolute value of $V$ .", + "bbox": [ + 511, + 263, + 903, + 295 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "4.3. Triplane VAE", + "text_level": 1, + "bbox": [ + 511, + 301, + 656, + 318 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "We employ an UNet-like structure of VAE to compress the triplane to latent space. The training pipeline is demonstrated in Fig. 2 (a). Considering that the 3 planes of triplanes encodes features from 3 orthogonal view direction, these 3 planes should be homogeneous data, so that We reshape the triplanes of batchsize $B$ to a new batch as $B \\times 3 \\times H \\times H \\times C$ to $3B \\times H \\times H \\times C$ for the VAE training. We use decoupled encoders and decoders for geometry and GS channels of triplanes, for experiments prove that a mixed encoding may lead to blur of the rendering results. After the triplane is reconstructed, the TriRenderer trained in Section.4.1 can be used to retrieve the GS pointclouds and render it to images.", + "bbox": [ + 511, + 324, + 906, + 520 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "In the training process, same loss functions as Eq. 3 and Eq. 5 are used, along with an $L_{1}$ loss between the input and reconstructed triplanes, denoted as $L_{\\mathrm{Tri}}$ . Additionally, a Kullback-Leibler divergence loss $L_{\\mathrm{KL}}$ is included to ensure that the latent variables do not deviate significantly from a normal distribution. The total losses is summarized as", + "bbox": [ + 511, + 520, + 905, + 609 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\nL _ {\\mathrm {V A E}} = L _ {\\mathrm {T r i}} + L _ {\\mathrm {G e o}} + L _ {\\mathrm {R e n d e r}} + \\gamma L _ {\\mathrm {K L}}, \\tag {8}\n$$\n", + "text_format": "latex", + "bbox": [ + 566, + 619, + 903, + 636 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "where $\\gamma$ is a small weight.", + "bbox": [ + 511, + 643, + 691, + 660 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "4.4. Two-stage LDM for Text-to-3D Generation", + "text_level": 1, + "bbox": [ + 511, + 667, + 877, + 683 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Taking into account the significant relevance of texture appearance to the underlying geometry, we proposed to utilize staged diffusion to generate geometry and GS attribute successively as shown in Fig. 2 (b). It is easy to implement it because the latent codes for geometry and other GS attributes are totally decoupled by VAE, as mentioned earlier. In this way, the geometry code can be a new condition for the second stage.", + "bbox": [ + 511, + 689, + 905, + 809 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "We follow DDPM [12, 20] to implement the text-conditioned LDM. To better capture the relation between different planes, we roll out the triplane latent of $B \\times 3 \\times h \\times h \\times c$ to a image-like $B \\times 3h \\times h \\times c$ as the input of diffusion model, and the generated result will be transformed back to the original shape for TriRenderer decoding.", + "bbox": [ + 511, + 810, + 905, + 901 + ], + "page_idx": 3 + }, + { + "type": "page_number", + "text": "16232", + "bbox": [ + 480, + 944, + 517, + 955 + ], + "page_idx": 3 + }, + { + "type": "image", + "img_path": "images/d4d455d7b936a86017daee34945deaf08686bcf6beb9a8f6ab8da1af198305b4.jpg", + "image_caption": [ + "Figure 2. Triplane VAE and 2-stage diffusion." + ], + "image_footnote": [], + "bbox": [ + 138, + 87, + 857, + 220 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "4.5. SDS-based Texture Booster", + "text_level": 1, + "bbox": [ + 89, + 275, + 336, + 290 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "To refine or restyle the generated GS pointcloud, we utilize a pretrained 2D diffusion model for the SDS training process [35] as shown in Fig. 3. As the generated 3D GS is constrained by the mesh reconstructed from SDF, detailed in Section 4.2, we maintain the GS splats' adherence as previously established. However, we enable the mesh vertices to shift within a limited radius to enhance the stability of the SDS process. Additionally, throughout the entire SDS procedure, there is no need for densification or pruning operations. This ensures that the number of points can be controlled to align with the number of mesh faces.", + "bbox": [ + 89, + 301, + 483, + 468 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "With the image rendered from the initial GS pointcloud, the 2D diffusion model can produce images with better 2D/3D consistence in different views, so that the optimization process can converge rapidly without Janus problem.", + "bbox": [ + 89, + 473, + 483, + 535 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "5. Experiment", + "text_level": 1, + "bbox": [ + 89, + 563, + 215, + 579 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "The experiments are conducted step-by-step according to the generation pipeline proposed as Fig.1 (a). We use OmniObject3D [45] dataset for toy experiments, and Objiverse [7] dataset for main evaluation. First, we sample 200K object to train their co顺应ing triplanes and TriNDERer. Then we train the VAE and diffusion model for 3DS GS generation. Finally, we exploit SDS to boost the texture quality for unsatisfactory generated objects. The procedure of data pre-processing and the implementation details are included in the supplementary materials.", + "bbox": [ + 89, + 593, + 482, + 744 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "The experimental results primarily consist of the following components: 1) For examining the proposed triplane modeling for GS generation, we focus on the text-to-3D task and showcase the direct GS generation outcomes qualitatively and quantitatively. These results are compared with Shap-E [19], Direct3D [26], and the most recent Gaussian-Cube [52]. 2) In comparison to 2D-lifting GS generation methods, we present the performance of SDS refinement on unsatisfying samples, with GsGEN [4] and GaussianDreamer [50] as baselines.", + "bbox": [ + 89, + 750, + 483, + 900 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "5.1. Triplane Fitting and VAE Reconstruction", + "text_level": 1, + "bbox": [ + 511, + 275, + 867, + 292 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Triplane Fitting. We adopt a resolution of $3 \\times 128 \\times 128 \\times 16$ for all triplane fitting, with a resolution of $64 \\times 64 \\times 64$ for the deformable SDF grids. We randomly sample 1000 objects for the shared TriRenderer training, and train all residual objects distributively. Every triplane is initialized with Gaussian noise before training. Under such settings, every triplane takes less than 30 seconds for geometry and GS appearance reconstruction.", + "bbox": [ + 511, + 297, + 903, + 417 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "VAE Reconstruction. We use a down-sample factor of 4 to compress the triplanes to latent space.", + "bbox": [ + 511, + 417, + 903, + 448 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "We visualize the triplane and the VAE reconstruction results of several data samples for a simple check, please refer to the supplementary materials.", + "bbox": [ + 511, + 449, + 903, + 494 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "5.2. Direct Text-to-3D GS Generation by LDM", + "text_level": 1, + "bbox": [ + 511, + 501, + 872, + 517 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "In this section, we present the results of our 3D GS object generation without the SDS refinement, both quantitatively and qualitatively. Due to the current scarcity of work targeting direct 3DGS generation, we selected two state-of-the-art text-to-3D works based on other 3D representations(NeRF) Shap-E [19] and Direct3D [26] for comparison. For fairness, these selected methods also perform direct generation in the 3D domain without any 2D prior knowledge or refinement based on 2D diffusion. Additionally, for the limited space, we provide more generation results and comparison with other methods such as GaussianCube [52] in the supplementary materials.", + "bbox": [ + 511, + 523, + 905, + 703 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Qualitative Results. The generation results from different methods are listed in Fig. 4. Every generated sample is rendered from different views accompanied by the provided text prompt captioned beneath the images. Our method showcase enhanced proficiency in both geometry and rendering quality, resulting in sharper and clearer outputs, which can be further verified in the subsequent quantitative evaluation. For more generated samples, please check the supplementary materials.", + "bbox": [ + 511, + 704, + 905, + 839 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Quantitative Results. We use CLIP score to evaluate the text-to-3d consistency, and an user study is conducted to evaluate the generation results from various aspect such as geometry, texture, realistic rendering and the consistency", + "bbox": [ + 511, + 840, + 903, + 901 + ], + "page_idx": 4 + }, + { + "type": "page_number", + "text": "16233", + "bbox": [ + 480, + 944, + 517, + 955 + ], + "page_idx": 4 + }, + { + "type": "image", + "img_path": "images/1c5fae7808bcece9143adcb3cb11fbe33f733882894530679f672ec3c79ce3e9.jpg", + "image_caption": [ + "Figure 3. Score distillation process(SDS) for optional texture refinement." + ], + "image_footnote": [], + "bbox": [ + 153, + 90, + 823, + 143 + ], + "page_idx": 5 + }, + { + "type": "table", + "img_path": "images/1593ef762653557dfd10d55a85de4bd7e62431342865575229bc4225a6f20f4a.jpg", + "table_caption": [], + "table_footnote": [], + "table_body": "
Geometry↑Texture↑Realism↑Consistency↑Overall↑
Shap-E3.0803.0192.9473.1213.042
Direct3D3.2223.1503.1183.2423.183
Ours3.4563.3833.3323.5203.423
", + "bbox": [ + 94, + 185, + 478, + 244 + ], + "page_idx": 5 + }, + { + "type": "table", + "img_path": "images/52b2bbc7f93175f53ef3335d446e33c6d9841e53528f72eb0feaaeb47a7e98f0.jpg", + "table_caption": [ + "Table 1. User study on generated 3D objects (Our results are raw outputs from LDM without SDS refinement)." + ], + "table_footnote": [], + "table_body": "
Shap-EDirect3DOurs
openai:ViT-L-140.23980.21520.2456
openai:ViT-L-14-3360.24260.22200.2462
", + "bbox": [ + 106, + 304, + 468, + 364 + ], + "page_idx": 5 + }, + { + "type": "table", + "img_path": "images/43dcfffad70634b65261c29f50cd9cba3bcace2e9f15b04b62c0598dbac6ce5c.jpg", + "table_caption": [ + "Table 2. CLIP score(↑) for evaluation of similarity between rendered images and given text prompt((Our results are raw outputs from LDM without SDS refinement))." + ], + "table_footnote": [], + "table_body": "
More Realistic (%)More Detailed (%)Overall (%)
GaussianDreamer35.050.042.5
GSGEN17.55.011.3
Ours47.545.046.3
", + "bbox": [ + 104, + 439, + 470, + 494 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "with given prompt. 49 users participated in the user study to score the over 50 3D samples from 1 to 5 points, and the average results are shown in Tab. 1. As for the CLIP score, the open-source repository t2vmetrics [24] is used to calculate the CLIP score on two versions of ViT models, and the results are demonstrated in Tab. 2. Both the CLIP score and the user study indicate that the proposed method produces better performance.", + "bbox": [ + 89, + 530, + 483, + 650 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "5.3. Optional Texture Boosting by SDS", + "text_level": 1, + "bbox": [ + 89, + 672, + 390, + 690 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "To compare with other SDS-based 3D GS generation methods, we implement a version of SDS for our mesh-binding GS representation as described in Section 4.5. GsGEN [4] and GaussianDreamer [50] are selected as our baselines. The generated samples, as illustrated in Fig. 5, highlight that with the SDS refiner, our DirectTriGS can deliver competitive outcomes comparable to the state-of-the-art 2D-lifting methods.", + "bbox": [ + 89, + 699, + 482, + 819 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Furthermore, we conducted a user study for quantitative evaluation. 40 users were tasked with ranking 15 samples from different methods based on two criteria: realism and level of detail. The results indicate that our approach marginally outperforms the other baseline methods.", + "bbox": [ + 89, + 825, + 483, + 902 + ], + "page_idx": 5 + }, + { + "type": "table", + "img_path": "images/fd92d3b8f0105832a7be991c6cdea016b3e28d2de73d32e81c47b53873ff58b0.jpg", + "table_caption": [ + "Table 3. User study on generated 3D objects (with SDS)." + ], + "table_footnote": [], + "table_body": "
TFLOPsParas.(M)GPU Mem.(GB)Running Time (Sec.)
Generation0.0082356.652.9416.2
", + "bbox": [ + 517, + 185, + 903, + 217 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Table 4. Generation inference efficiency, with the batch size is 1.", + "bbox": [ + 514, + 226, + 900, + 239 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "5.4. Ablation Study", + "text_level": 1, + "bbox": [ + 511, + 248, + 666, + 265 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "3D Diffusion for voxel-based Gaussian Splitting. We attempt to do GS attribute generation conditioned on given voxel occupancy via 3D diffusion model. However, even a toy experiment of over-fitting one single object failed, which may attributes to the complex multiple channels with different distributions especially the non-Euclidean ones such as quaternion. As shown in Fig.6, the diffusion model learns to generate color but fails to generate splat scaling, opacity and orientation.", + "bbox": [ + 511, + 270, + 906, + 406 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Triplane Diffusion without VAE. Since VAE inherently involves a reconstruction loss, we attempt to use direct diffusion on the triplane space for generation. The experiment results show that such direct diffusion may cause serious noise on decoded geometry. Randomly generated samples are visualized in Fig.7. A possible reason is that the multichannels of triplane contain considerable noisy redundancy, which is difficult for the LDM to filter.", + "bbox": [ + 511, + 406, + 905, + 526 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Inference Efficiency. The inference efficiency is listed in Table 4. This experiment is conducted on the platform equipped with RTX3090 GPU with 24GB memory.", + "bbox": [ + 511, + 527, + 905, + 574 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "6. Conclusion & Acknowledgment", + "text_level": 1, + "bbox": [ + 511, + 585, + 803, + 604 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "We have presented a novel framework DirectTriGS for Gaussian Splatting Field generation. DirectTriGS mainly consists of 3 parts: 1) a light-weight triplane representation for 3D object with the format of Gaussian Splatting; 2) a fully differentiable TriRenderer which can decode triplane to original GS and render it to multi-view images; 3) the triplane VAE and staged diffusion model for the whole generation process. By utilizing our DirectTriGS, the intricate GS data can be generated directly and efficiently. 4) Additionally, we incorporate a SDS refiner to further improve the texture and details of generated objects.", + "bbox": [ + 511, + 611, + 906, + 777 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "This project is funded in part by National Key R&D Program of China Project 2022ZD0161100, by the Centre for Perceptual and Interactive Intelligence (CPII) Ltd under the Innovation and Technology Commission (ITC)'s InnoHK, by General Research Fund of Hong Kong RGC Project 14204021, by NSFC-RGC Joint Fund Project N_CUHK498/24. Hongsheng Li is a PI of CPII under the InnoHK.", + "bbox": [ + 511, + 777, + 908, + 898 + ], + "page_idx": 5 + }, + { + "type": "page_number", + "text": "16234", + "bbox": [ + 480, + 944, + 517, + 955 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/e02b0b9e340b51003f4a0af94ca505b938ab38185310a1aeae72930ffb0077bf.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 109, + 90, + 339, + 203 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/a8d52a97ba35677a7fd3927da59bc8ef5a17462e5f80d1462158dcac99b07aee.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 357, + 90, + 620, + 208 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/c451f8c67b51d102c44bd5f968217768c2cc62129beff2a09320df123d8df510.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 640, + 90, + 880, + 203 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/ecdf04e07c1cb4999e734e4316aba7b2d96e2e51b03be480578dcba990b57259.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 133, + 205, + 339, + 320 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/2ceb55e32e7787145e7f5b7b418411f8833e84c28a3f12e7ba41dfae5427d41f.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 346, + 210, + 653, + 320 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/c93a0fea47463fb7388dd9b94157b2696942a779c3ac022e0750f2e3979a18f9.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 663, + 207, + 864, + 314 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/939ef3c413b258a30bc7602ce90f5916e697e334b760bbc40b742c461fa27860.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 125, + 333, + 346, + 415 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/0db8dc473bcd8b9a0ec8866a96e32cf4eb5c73d79cb8f8e5cc1486d0b098f5af.jpg", + "image_caption": [ + "(a) Comparison with Direct3D." + ], + "image_footnote": [], + "bbox": [ + 359, + 321, + 643, + 421 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/3bbb6ee05be3c129040b8c07a2869f056326c0048e52df83b029f2edbef790fa.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 668, + 321, + 867, + 417 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/7e68078254c1997822946878565641ef1c291448b2668b07fbb52580f39748ae.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 109, + 462, + 354, + 583 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/9a8f5f4e995cc14f49033c609e54d995cee059d85cbd38496bf0708bf7e03383.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 370, + 463, + 611, + 579 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/67d5c93c197f13c32b1ede71b73b93751838c6c0d5ccf21da19b76caa3ce5c98.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 625, + 462, + 887, + 583 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/ef41a3cd14b7ec0b934bd4833be2c8df5f63f92d873c950331dddcc120b9a90e.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 114, + 587, + 349, + 809 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/1c07367dad90481bf51654c2997d2915314471b3926331b615d5c02e577bc444.jpg", + "image_caption": [ + "(b) Comparison with Shap-E." + ], + "image_footnote": [], + "bbox": [ + 351, + 587, + 624, + 809 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/64b850aa9cfd764ef1cd388e358115550c22469c903a819a75a3b9afb6f2b7ee.jpg", + "image_caption": [ + "Figure 4. Comparison of generated 3D objects (Raw results without any refinement such as SDS)." + ], + "image_footnote": [], + "bbox": [ + 643, + 588, + 880, + 811 + ], + "page_idx": 6 + }, + { + "type": "page_number", + "text": "16235", + "bbox": [ + 480, + 945, + 517, + 955 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/f4e11a8500518460496587dd8c168e83c9614b6ac6d4ada9d2df5d5f61050a85.jpg", + "image_caption": [ + "Figure 5. Refinement of unsatisfactory results by SDS, compared with pure SDS-based methods GsGEN and GaussianDreamer." + ], + "image_footnote": [], + "bbox": [ + 161, + 87, + 823, + 683 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/98d34b585d9c24f1d501d70cd8477cd25c8c9773309fff17b241ddca3c0e9f35.jpg", + "image_caption": [ + "Figure 6. Failure of GS attribute generation." + ], + "image_footnote": [], + "bbox": [ + 184, + 734, + 328, + 829 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/95a0e5c3a1fbb9d463967d6d26657d1dad5d10e760be06d36e579c5a787388b9.jpg", + "image_caption": [ + "Figure 7. Noisy results of direct diffusion on triplane. The condition prompt are \"train\", \"watermelon\" and \"box\"." + ], + "image_footnote": [], + "bbox": [ + 364, + 736, + 828, + 827 + ], + "page_idx": 7 + }, + { + "type": "page_number", + "text": "16236", + "bbox": [ + 480, + 945, + 519, + 955 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "References", + "text_level": 1, + "bbox": [ + 91, + 89, + 187, + 104 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[1] Hansheng Chen, Jiatao Gu, Anpei Chen, Wei Tian, Zhuowen Tu, Lingjie Liu, and Hao Su. Single-stage diffusion nef: A unified approach to 3d generation and reconstruction. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 2416-2425, 2023. 2", + "[2] Yiwen Chen, Zilong Chen, Chi Zhang, Feng Wang, Xiaofeng Yang, Yikai Wang, Zhongang Cai, Lei Yang, Huaping Liu, and Guosheng Lin. Gaussianeditor: Swift and controllable 3d editing with gaussian splatting. arXiv preprint arXiv:2311.14521, 2023. 2", + "[3] Yurui Chen, Chun Gu, Junzhe Jiang, Xiatian Zhu, and Li Zhang. Periodic vibration gaussian: Dynamic urban scene reconstruction and real-time rendering. arXiv preprint arXiv:2311.18561, 2023. 2", + "[4] Zilong Chen, Feng Wang, Yikai Wang, and Huaping Liu. Text-to-3d using gaussian splatting. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 21401-21412, 2024. 2, 5, 6", + "[5] Kai Cheng, Xiaoxiao Long, Kaizhi Yang, Yao Yao, Wei Yin, Yuexin Ma, Wenping Wang, and Xuejin Chen. Gaussianpro: 3d gaussian splatting with progressive propagation. arXiv preprint arXiv:2402.14650, 2024. 2", + "[6] Jaeyoung Chung, Suyoung Lee, Hyeongjin Nam, Jaerin Lee, and Kyoung Mu Lee. Luciddreamer: Domain-free generation of 3d gaussian splattering scenes. arXiv preprint arXiv:2311.13384, 2023. 2", + "[7] Matt Deitke, Dustin Schwenk, Jordi Salvador, Luca Weihs, Oscar Michel, Eli VanderBilt, Ludwig Schmidt, Kiana Ehsani, Aniruddha Kembhavi, and Ali Farhadi. Objverse: A universe of annotated 3d objects. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 13142-13153, 2023. 5", + "[8] Zhiwen Fan, Kevin Wang, Kairun Wen, Zehao Zhu, Dejia Xu, and Zhangyang Wang. Lightgaussian: Unbounded 3d gaussian compression with 15x reduction and $200+$ fps. arXiv preprint arXiv:2311.17245, 2023. 2", + "[9] Jiemin Fang, Junjie Wang, Xiaopeng Zhang, Lingxi Xie, and Qi Tian. Gaussianeditor: Editing 3d gaussians delicately with text instructions. arXiv preprint arXiv:2311.16037, 2023. 2", + "[10] Jun Gao, Tianchang Shen, Zian Wang, Wenzheng Chen, Kangxue Yin, Daiqing Li, Or Litany, Zan Gojcic, and Sanja Fidler. Get3d: A generative model of high quality 3d textured shapes learned from images. Advances In Neural Information Processing Systems, 35:31841-31854, 2022. 2", + "[11] Anchit Gupta, Wenhan Xiong, Yixin Nie, Ian Jones, and Barlas Oğuz. 3dgen: Triplane latent diffusion for textured mesh generation. arXiv preprint arXiv:2303.05371, 2023. 2", + "[12] Jonathan Ho, Ajay Jain, and Pieter Abbeel. Denoising diffusion probabilistic models. Advances in neural information processing systems, 33:6840-6851, 2020. 4", + "[13] Lukas Hollein, Ang Cao, Andrew Owens, Justin Johnson, and Matthias Nießner. Text2room: Extracting textured 3d meshes from 2d text-to-image models. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 7909-7920, 2023. 2" + ], + "bbox": [ + 93, + 114, + 483, + 900 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[14] Fangzhou Hong, Jiaxiang Tang, Ziang Cao, Min Shi, Tong Wu, Zhaoxi Chen, Shuai Yang, Tengfei Wang, Liang Pan, Dahua Lin, et al. 3dtopia: Large text-to-3d generation model with hybrid diffusion priors. arXiv preprint arXiv:2403.02234, 2024. 2", + "[15] Binbin Huang, Zehao Yu, Anpei Chen, Andreas Geiger, and Shenghua Gao. 2d gaussian splatting for geometrically accurate radiance fields. In ACM SIGGRAPH 2024 Conference Papers, pages 1-11, 2024. 2", + "[16] Lutao Jiang and Lin Wang. Brightdreamer: Generic 3d gaussian generative framework for fast text-to-3d synthesis. arXiv preprint arXiv:2403.11273, 2024. 2", + "[17] Justin Johnson, Alexandre Alahi, and Li Fei-Fei. Perceptual losses for real-time style transfer and super-resolution. In Computer Vision-ECCV 2016: 14th European Conference, Amsterdam, The Netherlands, October 11-14, 2016, Proceedings, Part II 14, pages 694-711. Springer, 2016. 4", + "[18] Xiaoliang Ju, Zhaoyang Huang, Yijin Li, Guofeng Zhang, Yu Qiao, and Hongsheng Li. Diffindscene: Diffusion-based high-quality 3d indoor scene generation, 2023. 2", + "[19] Heewoo Jun and Alex Nichol. Shap-e: Generating conditional 3d implicit functions. arXiv preprint arXiv:2305.02463, 2023. 5", + "[20] Tero Karras, Miika Aittala, Timo Aila, and Samuli Laine. Elucidating the design space of diffusion-based generative models. Advances in neural information processing systems, 35:26565-26577, 2022. 4", + "[21] Bernhard Kerbl, Georgios Kopanas, Thomas Leimkuhler, and George Drettakis. 3d gaussian splatting for real-time radiance field rendering. ACM Transactions on Graphics, 42 (4):1-14, 2023. 1, 2, 4", + "[22] Samuli Laine, Janne Hellsten, Tero Karras, Yeongho Seol, Jaakko Lehtinen, and Timo Aila. Modular primitives for high-performance differentiable rendering. ACM Transactions on Graphics, 39(6), 2020. 4", + "[23] Yushi Lan, Fangzhou Hong, Shuai Yang, Shangchen Zhou, Xuyi Meng, Bo Dai, Xingang Pan, and Chen Change Loy. Ln3diff: Scalable latent neural fields diffusion for speedy 3d generation. In European Conference on Computer Vision, pages 112-130. Springer, 2025. 2", + "[24] Zhiqiu Lin, Deepak Pathak, Baiqi Li, Jiayao Li, Xide Xia, Graham Neubig, Pengchuan Zhang, and Deva Ramanan. Evaluating text-to-visual generation with image-to-text generation. arXiv preprint arXiv:2404.01291, 2024. 6", + "[25] Minghua Liu, Chao Xu, Haian Jin, Linghao Chen, Mukund Varma T, Zexiang Xu, and Hao Su. One-2-3-45: Any single image to 3d mesh in 45 seconds without per-shape optimization. Advances in Neural Information Processing Systems, 36, 2024. 2", + "[26] Qihao Liu, Yi Zhang, Song Bai, Adam Kortylewski, and Alan Yuille. Direct-3d: Learning direct text-to-3d generation on massive noisy 3d data. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 6881–6891, 2024. 5", + "[27] Ruoshi Liu, Rundi Wu, Basile Van Hoorick, Pavel Tokmakov, Sergey Zakharov, and Carl Vondrick. Zero-1-to-3: Zero-shot one image to 3d object. In Proceedings of the" + ], + "bbox": [ + 516, + 92, + 905, + 900 + ], + "page_idx": 8 + }, + { + "type": "page_number", + "text": "16237", + "bbox": [ + 480, + 944, + 517, + 955 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "IEEE/CVF International Conference on Computer Vision, pages 9298-9309, 2023. 2", + "[28] Xiaoxiao Long, Yuan-Chen Guo, Cheng Lin, Yuan Liu, Zhiyang Dou, Lingjie Liu, Yuexin Ma, Song-Hai Zhang, Marc Habermann, Christian Theobalt, et al. Wonder3d: Single image to 3d using cross-domain diffusion. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 9970-9980, 2024. 2", + "[29] Tao Lu, Mulin Yu, Linning Xu, Yuanbo Xiangli, Limin Wang, Dahua Lin, and Bo Dai. Scaffold-gs: Structured 3d gaussians for view-adaptive rendering. arXiv preprint arXiv:2312.00109, 2023. 2", + "[30] Hidenobu Matsuki, Riku Murai, Paul HJ Kelly, and Andrew J Davison. Gaussian splattering slam. arXiv preprint arXiv:2312.06741, 2023. 2", + "[31] Gal Metzer, Elad Richardson, Or Patashnik, Raja Giryes, and Daniel Cohen-Or. Latent-nerf for shape-guided generation of 3d shapes and textures. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 12663–12673, 2023. 2", + "[32] Ben Mildenhall, Pratul P Srinivasan, Matthew Tancik, Jonathan T Barron, Ravi Ramamoorthi, and Ren Ng. Nerf: Representing scenes as neural radiance fields for view synthesis. Communications of the ACM, 65(1):99-106, 2021. 1", + "[33] Jacob Munkberg, Jon Hasselgren, Tianchang Shen, Jun Gao, Wenzheng Chen, Alex Evans, Thomas Müller, and Sanja Fidler. Extracting triangular 3d models, materials, and lighting from images. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 8280-8290, 2022. 4", + "[34] Alex Nichol, Heewoo Jun, Prafulla Dhariwal, Pamela Mishkin, and Mark Chen. Point-e: A system for generating 3d point clouds from complex prompts. arXiv preprint arXiv:2212.08751, 2022. 2", + "[35] Ben Poole, Ajay Jain, Jonathan T Barron, and Ben Mildenhall. Dreamfusion: Text-to-3d using 2d diffusion. arXiv preprint arXiv:2209.14988, 2022. 1, 2, 5", + "[36] Robin Rombach, Andreas Blattmann, Dominik Lorenz, Patrick Esser, and Björn Ommer. High-resolution image synthesis with latent diffusion models. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 10684-10695, 2022. 1", + "[37] Tianchang Shen, Jun Gao, Kangxue Yin, Ming-Yu Liu, and Sanja Fidler. Deep marching tetrahedra: a hybrid representation for high-resolution 3d shape synthesis. In Advances in Neural Information Processing Systems (NeurIPS), 2021. 3", + "[38] Tianchang Shen, Jacob Munkberg, Jon Hasselgren, Kangxue Yin, Zian Wang, Wenzheng Chen, Zan Gojcic, Sanja Fidler, Nicholas Sharp, and Jun Gao. Flexible isosurface extraction for gradient-based mesh optimization. ACM Trans. Graph., 42(4), 2023. 3, 4", + "[39] Yichun Shi, Peng Wang, Jianglong Ye, Mai Long, Kejie Li, and Xiao Yang. Mvdream: Multi-view diffusion for 3d generation. arXiv preprint arXiv:2308.16512, 2023. 2", + "[40] J Ryan Shue, Eric Ryan Chan, Ryan Po, Zachary Ankner, Jiajun Wu, and Gordon Wetzstein. 3d neural field generation" + ], + "bbox": [ + 91, + 90, + 482, + 900 + ], + "page_idx": 9 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "using triplane diffusion. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 20875-20886, 2023. 1", + "[41] Jiaxiang Tang, Jiawei Ren, Hang Zhou, Ziwei Liu, and Gang Zeng. Dreamgaussian: Generative gaussian splatting for efficient 3d content creation. arXiv preprint arXiv:2309.16653, 2023. 2", + "[42] Zhou Wang, Alan C Bovik, Hamid R Sheikh, and Eero P Simoncelli. Image quality assessment: from error visibility to structural similarity. IEEE transactions on image processing, 13(4):600-612, 2004. 4", + "[43] Zhengyi Wang, Cheng Lu, Yikai Wang, Fan Bao, Chongxuan Li, Hang Su, and Jun Zhu. Prolificdreamer: High-fidelity and diverse text-to-3d generation with variational score distillation. Advances in Neural Information Processing Systems, 36, 2024. 2", + "[44] Guanjun Wu, Taoran Yi, Jiemin Fang, Lingxi Xie, Xiaopeng Zhang, Wei Wei, Wenyu Liu, Qi Tian, and Xinggang Wang. 4d gaussian splatting for real-time dynamic scene rendering. arXiv preprint arXiv:2310.08528, 2023. 2", + "[45] Tong Wu, Jiarui Zhang, Xiao Fu, Yuxin Wang, Liang Pan Jiawei Ren, Wayne Wu, Lei Yang, Jiaqi Wang, Chen Qian, Dahua Lin, and Ziwei Liu. Omniobject3d: Large-vocabulary 3d object dataset for realistic perception, reconstruction and generation. In IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), 2023. 5", + "[46] Zijie Wu, Yaonan Wang, Mingtao Feng, He Xie, and Ajmal Mian. Sketch and text guided diffusion model for colored point cloud generation. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 8929-8939, 2023. 2", + "[47] Zhennan Wu, Yang Li, Han Yan, Taizhang Shang, Weixuan Sun, Senbo Wang, Ruikai Cui, Weizhe Liu, Hiroyuki Sato, Hongdong Li, et al. Blockfusion: Expandable 3d scene generation using latent tri-plane extrapolation. arXiv preprint arXiv:2401.17053, 2024. 1, 2", + "[48] Chi Yan, Delin Qu, Dong Wang, Dan Xu, Zhigang Wang, Bin Zhao, and Xuelong Li. Gs-slam: Dense visual slam with 3d gaussian splatting. arXiv preprint arXiv:2311.11700, 2023. 2", + "[49] Zeyu Yang, Hongye Yang, Zijie Pan, Xiatian Zhu, and Li Zhang. Real-time photorealistic dynamic scene representation and rendering with 4d gaussian splatting. arXiv preprint arXiv:2310.10642, 2023. 2", + "[50] Taoran Yi, Jiemin Fang, Junjie Wang, Guanjun Wu, Lingxi Xie, Xiaopeng Zhang, Wenyu Liu, Qi Tian, and Xinggang Wang. Gaussiandreamer: Fast generation from text to 3d gaussians by bridging 2d and 3d diffusion models. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 6796-6807, 2024. 5, 6", + "[51] Zehao Yu, Anpei Chen, Binbin Huang, Torsten Sattler, and Andreas Geiger. Mip-splatting: Alias-free 3d gaussian splattering. arXiv preprint arXiv:2311.16493, 2023. 2", + "[52] Bowen Zhang, Yiji Cheng, Jiaolong Yang, Chunyu Wang, Feng Zhao, Yansong Tang, Dong Chen, and Baining Guo. Gaussiancube: Structuring gaussian splatting using optimal transport for 3d generative modeling. arXiv preprint arXiv:2403.19655, 2024. 2, 5" + ], + "bbox": [ + 516, + 92, + 903, + 898 + ], + "page_idx": 9 + }, + { + "type": "page_number", + "text": "16238", + "bbox": [ + 480, + 944, + 519, + 955 + ], + "page_idx": 9 + }, + { + "type": "ref_text", + "text": "[53] Zi-Xin Zou, Zhipeng Yu, Yuan-Chen Guo, Yangguang Li, Ding Liang, Yan-Pei Cao, and Song-Hai Zhang. Triplane meets gaussian splatting: Fast and generalizable single-view 3d reconstruction with transformers. arXiv preprint arXiv:2312.09147, 2023. 2, 4", + "bbox": [ + 91, + 90, + 486, + 161 + ], + "page_idx": 10 + }, + { + "type": "page_number", + "text": "16239", + "bbox": [ + 480, + 945, + 519, + 955 + ], + "page_idx": 10 + } +] \ No newline at end of file diff --git a/2025/DirectTriGS_ Triplane-based Gaussian Splatting Field Representation for 3D Generation/cb304427-02fe-443c-baf8-93b01f1651af_model.json b/2025/DirectTriGS_ Triplane-based Gaussian Splatting Field Representation for 3D Generation/cb304427-02fe-443c-baf8-93b01f1651af_model.json new file mode 100644 index 0000000000000000000000000000000000000000..79363d59b6a0cbe7b94500af603fb20dbf986a5a --- /dev/null +++ b/2025/DirectTriGS_ Triplane-based Gaussian Splatting Field Representation for 3D Generation/cb304427-02fe-443c-baf8-93b01f1651af_model.json @@ -0,0 +1,2257 @@ +[ + [ + { + "type": "header", + "bbox": [ + 0.107, + 0.003, + 0.182, + 0.043 + ], + "angle": 0, + "content": "CVF" + }, + { + "type": "header", + "bbox": [ + 0.237, + 0.001, + 0.812, + 0.047 + ], + "angle": 0, + "content": "This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore." + }, + { + "type": "title", + "bbox": [ + 0.113, + 0.13, + 0.885, + 0.175 + ], + "angle": 0, + "content": "DirectTriGS: Triplane-based Gaussian Splitting Field Representation for 3D Generation" + }, + { + "type": "text", + "bbox": [ + 0.329, + 0.203, + 0.67, + 0.238 + ], + "angle": 0, + "content": "Xiaoliang Ju, Hongsheng Li\\(^{1,2}\\) \n\\(^{1}\\) CUHK MMLab \\(^{2}\\) CPII under InnoHK" + }, + { + "type": "text", + "bbox": [ + 0.398, + 0.243, + 0.595, + 0.256 + ], + "angle": 0, + "content": "akira@link.cuhk.edu.hk" + }, + { + "type": "title", + "bbox": [ + 0.248, + 0.292, + 0.327, + 0.308 + ], + "angle": 0, + "content": "Abstract" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.325, + 0.486, + 0.673 + ], + "angle": 0, + "content": "We present DirectTriGS, a novel framework designed for 3D object generation with Gaussian Splatting (GS). GS-based rendering for 3D content has gained considerable attention recently. However, there has been limited exploration in directly generating 3D Gaussians compared to traditional generative modeling approaches. The main challenge lies in the complex data structure of GS represented by discrete point clouds with multiple channels. To overcome this challenge, we propose employing the triplane representation, which allows us to represent Gaussian Splatting as an image-like continuous field. This representation effectively encodes both the geometry and texture information, enabling smooth transformation back to Gaussian point clouds and rendering into images by a TriRenderer, with only 2D supervisions. The proposed TriRenderer is fully differentiable, so that the rendering loss can supervise both texture and geometry encoding. Furthermore, the triplane representation can be compressed using a Variational Autoencoder (VAE), which can subsequently be utilized in latent diffusion to generate 3D objects. The experiments demonstrate that the proposed generation framework can produce high-quality 3D object geometry and rendering results in the text-to-3D task." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.708, + 0.222, + 0.723 + ], + "angle": 0, + "content": "1. Introduction" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.735, + 0.485, + 0.903 + ], + "angle": 0, + "content": "Neural rendering has grown to a focal point in rendering techniques in recent years, as it achieves more realistic rendering effects by leveraging the great expressiveness of neural network. The representatives are neural radiance field (NeRF) [32] and the newly emerged Gaussian Splatting (GS) [21]. However, when applied to the field of 3D generation, the slow rendering and training speed become a strong limitation for NeRF. While GS is more flexible with greater rendering efficiency and editability, few works have addressed the challenge of direct GS generation in 3D due to its complex data structure. In this paper, we propose a" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.293, + 0.907, + 0.368 + ], + "angle": 0, + "content": "novel approach called DirectTriGS, which encodes GS as triplane representation, and introduce its corresponding renderer, TriRenderer. Subsequently, we apply latent diffusion on the Triplane representation to generate high-quality GS objects." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.371, + 0.909, + 0.734 + ], + "angle": 0, + "content": "Gaussian splatting uses multi-channel point cloud of \"splats\" to describe the 3D contents. With a differentiable splats rasterizer, GS has fast rendering speed. However, the sparsity, multiple channels, and uneven distributed density of 3D GS bring great difficulty for 3D generation. We propose to encode GS contents as multi-channel triplane representations, which have been shown to possess favorable properties for representing 3D geometry or NeRF, as demonstrated in previous works Shue et al. [40], Wu et al. [47]. In our work, we leverage the triplane representation to encode both the geometry information and other GS attributes in two separate groups of channels. By training on a dataset of 3D objects, we obtain a shared TriRenderer that is capable of decoding different triplane representations to GS and then render it to image. TriRenderer is fully differentiable, enabling the use of only 2D rendering loss to supervise both the texture and geometry of 3D GS. The advantage of using Triplane to represent GS are two folds. First, it leads to high memory efficiency compared with dense voxels and is expressive enough for generating various 3D GS. Second, Triplane is more compatible with the convolution-based encoders compared with the original sparse GS point cloud, which require specifically designed networks on processing sparse point clouds." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.735, + 0.911, + 0.903 + ], + "angle": 0, + "content": "We follow stable diffusion [36] to train DirectTriGS on the proposed triplane representation. Specifically, a VAE is designed to further convert the Triplanes into latent code. We use two separate decoders to decouple the decoding of geometry and GS attributes. Second, we roll out the triplane latent to an expanded multi-channel image, then exploit staged latent diffusion to do generation. Two-stage diffusion is employed to generate the geometry and the corresponding GS appearance. Finally, score distillation sampling (SDS) [35] is adopted as a optional post-processing to refine or restyle the generated 3D objects." + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.52, + 0.958 + ], + "angle": 0, + "content": "16229" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.09, + 0.092, + 0.482, + 0.272 + ], + "angle": 0, + "content": "Our contribution can summarize as follows. 1) We propose a Triplane representation for direct 3D GS generation, which is memory efficient to derive GS point clouds with vivid rendering. 2) We design a fully differentiable TriRenderer to enable the end-to-end training of triplane representations, with only 2D supervisions. 3) We develop a Triplane-based GS generation framework DirectTriGS that incorporates a specially designed VAE, latent diffusion module, and a SDS based refiner. 4) Experiments demonstrate our method produces competitive performance in both 3D geometry and multi-view rendering quality in text-to-3D tasks." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.287, + 0.24, + 0.303 + ], + "angle": 0, + "content": "2. Related Works" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.312, + 0.483, + 0.568 + ], + "angle": 0, + "content": "3D Object Generation. 1) Multi-view images-to-3D. The first route generates multi-view colored 3D image and then reconstruct the 3D shape and project image to texture, or use score distillation sampling to build 3D object from images produced by 2D diffusion, such as [6, 13, 27, 35, 39, 41]. Since these approaches are actually 2D generator, it is difficult to maintain 3D consistency, which may lead to Janus problem. 2) Image-to-3D. The second route attempt to retrieve 3D geometry from a single input image, such as [25, 28, 53]. With the certain one image as input, the users may be more concerned with the reconstruction quality. 3)Text-to-3D. The third group mainly operate in 3D, which can be clustered by different 3D representation. For example, [10, 11] are able to generate high-quality 3D textured meshes using differentiable rendering; [34, 46] generate colored pointcloud; [1, 14, 23, 31] generate NeRF volumes; [18, 47] generate SDF volumes." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.569, + 0.483, + 0.629 + ], + "angle": 0, + "content": "Remark. The input conditions of these three generation methods are gradually becoming abstract, corresponding to increasing difficulty of the generation task, commonly resulting in a gradual decline in the generation quality." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.63, + 0.483, + 0.75 + ], + "angle": 0, + "content": "Gaussian Splatting. Gaussian Splatting [21] exploits multi-channel pointcloud of splats to describe the 3D content, which allows real-time rendering of photorealistic scenes. Recently, there emerges so many research focusing on its reconstruction quality enhancement and algorithm optimization [5, 8, 15, 29, 51]. Another related active area includes 4D dynamic GS modeling [3, 44, 49], scene editing [2, 9] and its applications like SLAM [30, 48]." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.751, + 0.483, + 0.902 + ], + "angle": 0, + "content": "As for GS generation, some works exploit multi-view image based route such as [6, 41], where the image generator output image from required views to enhance the GS rendering in a recursive manner. Rare works explore the GS generation directly in 3D space, such as [53] generates a point cloud using a Transformer structure with 2D image as input, and build a mapping of image to tri-plane encoding of GS attribute; GaussianCube [52] proposes to use optimal transport to model the 3D GS for text-to-3D generation; [16] combine explicit anchor points with implicit" + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.092, + 0.772, + 0.108 + ], + "angle": 0, + "content": "triplane to directly generate GS points." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.119, + 0.787, + 0.137 + ], + "angle": 0, + "content": "3. Revisiting Gaussian Splitting" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.145, + 0.905, + 0.298 + ], + "angle": 0, + "content": "Gaussian Splatting [21] uses a set of Gaussian points to describe a 3D object. The Gaussian points are defined by a full 3D covariance matrix \\(\\pmb{\\Sigma}\\) in a world space as \\(G(\\mathbf{x}) = e^{-\\frac{1}{2}\\mathbf{x}^T\\pmb{\\Sigma}^{-1}\\mathbf{x}}\\), centered at point (mean) \\(\\mathbf{p} \\in \\mathbb{R}^3\\). Each point is with a opacity scalar \\(\\alpha \\in [0,1]\\) for blending and a series of Spherical Harmonics (SH) coefficients to correctly capture the view-dependent appearance of the scene. The number of SH coefficients are \\(3 \\times (n + 1)^2\\), where \\(n\\) denotes the SH order, and higher \\(n\\) corresponds to more accurate view-dependent appearance." + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.298, + 0.906, + 0.449 + ], + "angle": 0, + "content": "During rendering, the Gaussians are project to 2D given a viewing transformation \\(\\mathbf{W}\\). The covariance is transformed as \\(\\boldsymbol{\\Sigma}^{\\prime} = \\mathbf{J}\\mathbf{W}\\boldsymbol{\\Sigma}\\mathbf{W}^{T}\\mathbf{J}^{T}\\), where \\(\\mathbf{J}\\) is the Jacobian of the affine approximation of the projective transformation. To ensure \\(\\boldsymbol{\\Sigma}\\) to be semi-positive in the whole training process, \\(\\boldsymbol{\\Sigma}\\) can be represented by a Cholesky decomposition as \\(\\boldsymbol{\\Sigma} = \\mathbf{R}\\mathbf{S}\\mathbf{S}^{T}\\mathbf{R}^{T}\\), represented by a tuple \\(s = (s_{1}, s_{2}, s_{3})\\) for scaling and a unnormalized quaternion \\(\\mathbf{q} \\in \\mathbb{R}^{4}\\) for rotation. In summary, each Gaussian point are a union set of position \\(\\mathbf{p}\\) and its GS attributes:" + }, + { + "type": "equation", + "bbox": [ + 0.638, + 0.459, + 0.905, + 0.476 + ], + "angle": 0, + "content": "\\[\n\\mathcal {G} _ {\\mathbf {p}} := \\{s, \\mathbf {q}, \\mathrm {S H}, \\alpha \\}. \\tag {1}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.485, + 0.904, + 0.516 + ], + "angle": 0, + "content": "GS pointcloud can be efficiently rendered by a rasterization-based splatting renderer [21], which is fully differentiable." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.529, + 0.649, + 0.546 + ], + "angle": 0, + "content": "4. Methodology" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.553, + 0.906, + 0.704 + ], + "angle": 0, + "content": "Our motivation is to directly generate 3D GS with fast speed by utilizing the proposed triplane-based GS representation. While 2D-lifting methods like SDS-based approaches yield intricate results, they are time-intensive and susceptible to Janus problem. Direct generation, notably with fast sampling, can complete 3D generation in under a minute, contrasting with SDS methods that take over 10 minutes or longer, such as [4][43], posing challenges for users. Therefore, direct 3D generation is more practical, with 2D-lifting reserved for enhancing texture details when needed." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.705, + 0.907, + 0.901 + ], + "angle": 0, + "content": "The procedures of our generation framework can be divide into 4 stages as shown in Fig.1 (a). Firstly, we encode the 3D objects into triplanes using the TriRenderer. The triplane encoding and TriRenderer training can be concurrently executed with only the multi-view RGBA image and camera poses as supervision. Secondly, a VAE is trained to compress the triplanes to latent space, enabling effective capture of high-level information for subsequent modules. Thirdly, a diffusion model is trained on the triplane latent code. During inference, the diffusion model produces the latent code, which can be decoded back to its corresponding triplane by the VAE decoder. Subsequently, the TriRenderer decodes the triplane into the standard GS point" + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "16230" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.137, + 0.09, + 0.308, + 0.143 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.322, + 0.091, + 0.493, + 0.143 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.506, + 0.091, + 0.678, + 0.143 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.691, + 0.092, + 0.863, + 0.143 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.304, + 0.149, + 0.66, + 0.163 + ], + "angle": 0, + "content": "(a) The overall procedure of our 3D content generation framework." + }, + { + "type": "image", + "bbox": [ + 0.135, + 0.172, + 0.863, + 0.345 + ], + "angle": 0, + "content": null + }, + { + "type": "image_footnote", + "bbox": [ + 0.262, + 0.349, + 0.747, + 0.361 + ], + "angle": 0, + "content": "(b) The forward pass of the TriRenderer pipeline: rendering a Triolane to multi-view images." + }, + { + "type": "image_caption", + "bbox": [ + 0.09, + 0.371, + 0.908, + 0.399 + ], + "angle": 0, + "content": "Figure 1. The overview of the proposed 3D GS generation framework and the core component TriRenderer. We use triplane to represent 3D objects, and design a TriRenderer to decode triplane to GS pointcloud and render multi-view images." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.426, + 0.484, + 0.501 + ], + "angle": 0, + "content": "cloud. Finally, an SDS-based refiner can be utilized as an optional post-processing step to enhance or restyle the generated 3D object. Leveraging the pretrained 2D diffusion model within the SDS, the generated GS can be further improved to achieve a more detailed appearance." + }, + { + "type": "title", + "bbox": [ + 0.09, + 0.512, + 0.315, + 0.528 + ], + "angle": 0, + "content": "4.1. Triplane-based GS Field" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.535, + 0.484, + 0.822 + ], + "angle": 0, + "content": "We employ triplanes to convert the discrete multi-channel GS points into a continuous field, tailored for more efficient encoding in the subsequent generative model. Triplanes are 3 planes formed by every 2 axes of \\(x,y,z\\) axes in 3D space, where every 3D point \\(\\mathbf{p}\\) can query these planes to get corresponding features \\(F_{\\mathbf{p}} = (F_{xy},F_{xz},F_{yz})\\) using orthogonal projection. In real implementation, triplane is a tensor with the size of \\(3\\times H\\times H\\times C\\), where \\(H\\) denote the resolution along every axis and \\(C\\) is the feature channel. The query of point with arbitrary continuous coordinate should be a bi-linear interpolation in the triplane grids. Considering GS is a set of points with multiple channels of attributes, it is natural to use triplane for GS encoding. If the triplane feature and GS attributes of a point can be transformed to each other, the sparse GS pointcloud of every object can be represented as continuous triplane GS field for further encoding. We use separate channels to encode 3D geometry and other GS attributes as described in Eq. 1, which is an experimental setting for better convergence in the training." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.833, + 0.224, + 0.848 + ], + "angle": 0, + "content": "4.2. TriRenderer" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.856, + 0.484, + 0.901 + ], + "angle": 0, + "content": "The TriRenderer serves as the crucial component for converting a 3D object into a triplane-based GS field. It acts as a fully differentiable bridge connecting the GS field with" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.426, + 0.905, + 0.456 + ], + "angle": 0, + "content": "rendering, enabling the optimization from 2D images to geometry and appearance encoding." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.458, + 0.907, + 0.655 + ], + "angle": 0, + "content": "As depicted in Fig. 1 (b), the TriRenderer comprises a geometry branch and a GS attribute branch, each equipped with independent decoders to decode the triplane. The geometry branch is responsible for retrieve the geometry as triangular mesh from triplane, with a surface sampler for GS pointcloud sampling. Then GS attribute branch uses the obtained pointcloud to query the triplane to obtain the GS attribute corresponding to each point. In this way, the GS pointcloud in the original format is retrieved, and can be rendered using the original GS renderer. It is worth mentioning that all triplanes of different object share a common TriRenderer, which ensures the features on different triplanes subject to a similar distribution." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.657, + 0.906, + 0.779 + ], + "angle": 0, + "content": "Geometry Branch. We use signed distance function (SDF) to represent the geometry, and allow every vertex deform as [37], so that the geometry branch decodes \\( F_{p} \\) to a SDF volume and its vertex deformation. By query all grid coordinates of the designated resolution \\( L \\times L \\times L \\), the deformable SDF volume is reconstructed. Then we exploit a differentiable Marching Cube algorithm FlexiCubes [38] to extract triangular mesh from the SDF volume." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.78, + 0.909, + 0.902 + ], + "angle": 0, + "content": "Surface Sampling. Considering that GS pointcloud is generally gathered on the surface of objects, we randomly sample GS points on the faces of triangular mesh with barycentric coordinates. We expect that every splat is flat and with a normal consistent with its source triangular face. Assume a face is formed by vertices \\( < v_{a}, v_{b}, v_{c} > \\), and vector \\( \\mathbf{v}_{1} = (v_{a} - v_{b}) \\), \\( \\mathbf{v}_{2} = (v_{b} - v_{c}) \\), the face normal can be calculated as \\( \\mathbf{n} = \\mathbf{v}_{1} \\times \\mathbf{v}_{2} \\). Then the rotation matrix of every" + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.518, + 0.957 + ], + "angle": 0, + "content": "16231" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.091, + 0.092, + 0.407, + 0.106 + ], + "angle": 0, + "content": "splat drawn from this face can be formulated as" + }, + { + "type": "equation", + "bbox": [ + 0.213, + 0.117, + 0.483, + 0.134 + ], + "angle": 0, + "content": "\\[\n\\mathbf {R} = \\left[ \\mathbf {v} _ {1}, \\mathbf {n} \\times \\mathbf {v} _ {1}, \\mathbf {n} \\right], \\tag {2}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.145, + 0.483, + 0.189 + ], + "angle": 0, + "content": "which can be transform to \\(\\mathbf{q}\\) in Eq.1 by standard matrix-to-quaternion algorithm. To make the splats flat, \\(s_3\\) in Eq.1 is fixed to an infinite small value." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.19, + 0.484, + 0.281 + ], + "angle": 0, + "content": "To further reduce the computational load in the subsequent GS rendering process, only faces oriented toward the camera are sampled. These face indices can be obtained through fast rasterization of the mesh faces. Note this rasterization is not required to be differentiable, and we use the library Nvdiffrast [22] for implementation." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.281, + 0.484, + 0.401 + ], + "angle": 0, + "content": "GS Attribute Branch. By querying the triplane GS channels with the sampled pointcloud, we can obtain the GS feature and use the GS attribute branch to transform them to the rest GS attributes. Then the GS renderer can render GS pointcloud to multi-view images. Since different GS attributes have different numerical scales and distributions, we customize individual headers to decode them respectively, which is similar to [53]." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.402, + 0.484, + 0.538 + ], + "angle": 0, + "content": "Training. The training consists of 2 stages. In the first stage, a small batch of data is used to train the triplanes and the TriRenderer together. The second stage involves the whole dataset to train all triplanes with the parameters of TriRenderer frozen. In this way, we can handle large dataset by distributed training in the second stage, and the same TriRenderer can be shared by all triplanes. There is no need to pre-train the GS pointcloud using original GS training as [21]." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.538, + 0.484, + 0.599 + ], + "angle": 0, + "content": "The only required supervision data is \\(N\\) multi-view images \\(\\{I_i\\}_{i=1}^N\\) with camera poses, and the training losses consist of the rendering loss \\(L_{\\mathrm{render}}\\) and the geometric regularization \\(L_{\\mathrm{geo}}\\). The rendering loss is a weighted sum as" + }, + { + "type": "equation", + "bbox": [ + 0.144, + 0.613, + 0.483, + 0.629 + ], + "angle": 0, + "content": "\\[\nL _ {\\text {r e n d e r}} = w _ {1} L _ {\\text {a l p h a}} + w _ {2} L _ {\\text {r g b}} + w _ {3} L _ {\\text {p i p s}}, \\tag {3}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.635, + 0.484, + 0.697 + ], + "angle": 0, + "content": "where \\( L_{\\mathrm{alpha}} \\) is a \\( L_{1} \\) loss on the alpha map, namely the silhouette image loss; \\( L_{\\mathrm{rgb}} \\) is a combination of \\( L_{1} \\) pixel loss and SSIM loss [42] between rendered image \\( \\hat{I}_i \\) and ground truth \\( I_i \\):" + }, + { + "type": "equation", + "bbox": [ + 0.118, + 0.706, + 0.483, + 0.749 + ], + "angle": 0, + "content": "\\[\nL _ {\\mathrm {r g b}} = \\sum_ {i = 1} ^ {N} (1 - \\beta) \\| I _ {i} - \\hat {I} _ {i} \\| _ {1} + \\beta \\operatorname {S S I M} \\left(I _ {i}, \\hat {I} _ {i}\\right), \\tag {4}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.757, + 0.483, + 0.787 + ], + "angle": 0, + "content": "and \\(L_{\\mathrm{pips}}\\) is the perceptual loss [17]. \\(w_{1}, w_{2}, w_{3}\\) and \\(\\beta\\) are all weighting factors determined by experiments." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.787, + 0.484, + 0.817 + ], + "angle": 0, + "content": "The geometric regularization loss on the SDF volume \\(V\\) is defined as" + }, + { + "type": "equation", + "bbox": [ + 0.104, + 0.829, + 0.483, + 0.846 + ], + "angle": 0, + "content": "\\[\nL _ {\\text {g e o}} = \\gamma_ {1} L _ {\\text {d e v}} + \\gamma_ {2} L _ {\\text {w e i g h t i n g}} + \\gamma_ {3} L _ {\\text {C E}} + \\gamma_ {4} L _ {\\text {s i g n}}, \\tag {5}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.856, + 0.484, + 0.902 + ], + "angle": 0, + "content": "where the \\(L_{\\mathrm{dev}}\\) and \\(L_{\\mathrm{weighting}}\\) are defined by Flex-cubes [38] to regularize the extracted connectivity and the weighting scale of SDF vertices. To penalize the sign" + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.092, + 0.905, + 0.12 + ], + "angle": 0, + "content": "changes on all grid edges, we follow [33] to define \\(L_{\\mathrm{CE}}\\) as" + }, + { + "type": "equation", + "bbox": [ + 0.583, + 0.12, + 0.905, + 0.153 + ], + "angle": 0, + "content": "\\[\nL _ {\\mathrm {C E}} = \\sum_ {(s _ {a}, s _ {b}) \\in \\epsilon} \\operatorname {C E} \\left(\\sigma \\left(s _ {a}\\right), \\operatorname {s i g n} \\left(s _ {b}\\right)\\right), \\tag {6}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.156, + 0.907, + 0.232 + ], + "angle": 0, + "content": "where CE denotes cross entropy; \\(\\epsilon\\) is the set of all edges connecting vertices with different signs; and \\(\\sigma\\) is the sigmoid function. Finally, \\(L_{\\mathrm{sign}}\\) is designed to prevent the SDF volume from being trapped in an empty shape, i.e. a fully positive or fully negative SDF volume," + }, + { + "type": "equation", + "bbox": [ + 0.634, + 0.24, + 0.905, + 0.258 + ], + "angle": 0, + "content": "\\[\nL _ {\\text {s i g n}} = \\delta (V) M (| V |), \\tag {7}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.264, + 0.905, + 0.296 + ], + "angle": 0, + "content": "where \\(\\delta (V) = 1\\) if \\(V\\) is empty, otherwise \\(\\delta (V) = 0\\). \\(M(|V|)\\) denotes the mean of the absolute value of \\(V\\)." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.303, + 0.657, + 0.319 + ], + "angle": 0, + "content": "4.3. Triplane VAE" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.325, + 0.907, + 0.521 + ], + "angle": 0, + "content": "We employ an UNet-like structure of VAE to compress the triplane to latent space. The training pipeline is demonstrated in Fig. 2 (a). Considering that the 3 planes of triplanes encodes features from 3 orthogonal view direction, these 3 planes should be homogeneous data, so that We reshape the triplanes of batchsize \\( B \\) to a new batch as \\( B \\times 3 \\times H \\times H \\times C \\) to \\( 3B \\times H \\times H \\times C \\) for the VAE training. We use decoupled encoders and decoders for geometry and GS channels of triplanes, for experiments prove that a mixed encoding may lead to blur of the rendering results. After the triplane is reconstructed, the TriRenderer trained in Section.4.1 can be used to retrieve the GS pointclouds and render it to images." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.521, + 0.906, + 0.611 + ], + "angle": 0, + "content": "In the training process, same loss functions as Eq. 3 and Eq. 5 are used, along with an \\( L_{1} \\) loss between the input and reconstructed triplanes, denoted as \\( L_{\\mathrm{Tri}} \\). Additionally, a Kullback-Leibler divergence loss \\( L_{\\mathrm{KL}} \\) is included to ensure that the latent variables do not deviate significantly from a normal distribution. The total losses is summarized as" + }, + { + "type": "equation", + "bbox": [ + 0.568, + 0.621, + 0.905, + 0.637 + ], + "angle": 0, + "content": "\\[\nL _ {\\mathrm {V A E}} = L _ {\\mathrm {T r i}} + L _ {\\mathrm {G e o}} + L _ {\\mathrm {R e n d e r}} + \\gamma L _ {\\mathrm {K L}}, \\tag {8}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.645, + 0.692, + 0.661 + ], + "angle": 0, + "content": "where \\(\\gamma\\) is a small weight." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.668, + 0.879, + 0.684 + ], + "angle": 0, + "content": "4.4. Two-stage LDM for Text-to-3D Generation" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.69, + 0.906, + 0.81 + ], + "angle": 0, + "content": "Taking into account the significant relevance of texture appearance to the underlying geometry, we proposed to utilize staged diffusion to generate geometry and GS attribute successively as shown in Fig. 2 (b). It is easy to implement it because the latent codes for geometry and other GS attributes are totally decoupled by VAE, as mentioned earlier. In this way, the geometry code can be a new condition for the second stage." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.811, + 0.906, + 0.902 + ], + "angle": 0, + "content": "We follow DDPM [12, 20] to implement the text-conditioned LDM. To better capture the relation between different planes, we roll out the triplane latent of \\( B \\times 3 \\times h \\times h \\times c \\) to a image-like \\( B \\times 3h \\times h \\times c \\) as the input of diffusion model, and the generated result will be transformed back to the original shape for TriRenderer decoding." + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "16232" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.139, + 0.088, + 0.858, + 0.222 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.36, + 0.235, + 0.637, + 0.25 + ], + "angle": 0, + "content": "Figure 2. Triplane VAE and 2-stage diffusion." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.276, + 0.338, + 0.291 + ], + "angle": 0, + "content": "4.5. SDS-based Texture Booster" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.303, + 0.484, + 0.469 + ], + "angle": 0, + "content": "To refine or restyle the generated GS pointcloud, we utilize a pretrained 2D diffusion model for the SDS training process [35] as shown in Fig. 3. As the generated 3D GS is constrained by the mesh reconstructed from SDF, detailed in Section 4.2, we maintain the GS splats' adherence as previously established. However, we enable the mesh vertices to shift within a limited radius to enhance the stability of the SDS process. Additionally, throughout the entire SDS procedure, there is no need for densification or pruning operations. This ensures that the number of points can be controlled to align with the number of mesh faces." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.474, + 0.484, + 0.536 + ], + "angle": 0, + "content": "With the image rendered from the initial GS pointcloud, the 2D diffusion model can produce images with better 2D/3D consistence in different views, so that the optimization process can converge rapidly without Janus problem." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.564, + 0.216, + 0.58 + ], + "angle": 0, + "content": "5. Experiment" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.594, + 0.483, + 0.746 + ], + "angle": 0, + "content": "The experiments are conducted step-by-step according to the generation pipeline proposed as Fig.1 (a). We use OmniObject3D [45] dataset for toy experiments, and Objiverse [7] dataset for main evaluation. First, we sample 200K object to train their co顺应ing triplanes and TriNDERer. Then we train the VAE and diffusion model for 3DS GS generation. Finally, we exploit SDS to boost the texture quality for unsatisfactory generated objects. The procedure of data pre-processing and the implementation details are included in the supplementary materials." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.75, + 0.484, + 0.901 + ], + "angle": 0, + "content": "The experimental results primarily consist of the following components: 1) For examining the proposed triplane modeling for GS generation, we focus on the text-to-3D task and showcase the direct GS generation outcomes qualitatively and quantitatively. These results are compared with Shap-E [19], Direct3D [26], and the most recent Gaussian-Cube [52]. 2) In comparison to 2D-lifting GS generation methods, we present the performance of SDS refinement on unsatisfying samples, with GsGEN [4] and GaussianDreamer [50] as baselines." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.276, + 0.868, + 0.293 + ], + "angle": 0, + "content": "5.1. Triplane Fitting and VAE Reconstruction" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.298, + 0.905, + 0.418 + ], + "angle": 0, + "content": "Triplane Fitting. We adopt a resolution of \\(3 \\times 128 \\times 128 \\times 16\\) for all triplane fitting, with a resolution of \\(64 \\times 64 \\times 64\\) for the deformable SDF grids. We randomly sample 1000 objects for the shared TriRenderer training, and train all residual objects distributively. Every triplane is initialized with Gaussian noise before training. Under such settings, every triplane takes less than 30 seconds for geometry and GS appearance reconstruction." + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.419, + 0.905, + 0.449 + ], + "angle": 0, + "content": "VAE Reconstruction. We use a down-sample factor of 4 to compress the triplanes to latent space." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.45, + 0.905, + 0.495 + ], + "angle": 0, + "content": "We visualize the triplane and the VAE reconstruction results of several data samples for a simple check, please refer to the supplementary materials." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.502, + 0.874, + 0.518 + ], + "angle": 0, + "content": "5.2. Direct Text-to-3D GS Generation by LDM" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.524, + 0.906, + 0.704 + ], + "angle": 0, + "content": "In this section, we present the results of our 3D GS object generation without the SDS refinement, both quantitatively and qualitatively. Due to the current scarcity of work targeting direct 3DGS generation, we selected two state-of-the-art text-to-3D works based on other 3D representations(NeRF) Shap-E [19] and Direct3D [26] for comparison. For fairness, these selected methods also perform direct generation in the 3D domain without any 2D prior knowledge or refinement based on 2D diffusion. Additionally, for the limited space, we provide more generation results and comparison with other methods such as GaussianCube [52] in the supplementary materials." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.705, + 0.906, + 0.84 + ], + "angle": 0, + "content": "Qualitative Results. The generation results from different methods are listed in Fig. 4. Every generated sample is rendered from different views accompanied by the provided text prompt captioned beneath the images. Our method showcase enhanced proficiency in both geometry and rendering quality, resulting in sharper and clearer outputs, which can be further verified in the subsequent quantitative evaluation. For more generated samples, please check the supplementary materials." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.841, + 0.905, + 0.902 + ], + "angle": 0, + "content": "Quantitative Results. We use CLIP score to evaluate the text-to-3d consistency, and an user study is conducted to evaluate the generation results from various aspect such as geometry, texture, realistic rendering and the consistency" + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "16233" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.154, + 0.091, + 0.825, + 0.144 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.281, + 0.155, + 0.715, + 0.171 + ], + "angle": 0, + "content": "Figure 3. Score distillation process(SDS) for optional texture refinement." + }, + { + "type": "table", + "bbox": [ + 0.095, + 0.186, + 0.48, + 0.245 + ], + "angle": 0, + "content": "
Geometry↑Texture↑Realism↑Consistency↑Overall↑
Shap-E3.0803.0192.9473.1213.042
Direct3D3.2223.1503.1183.2423.183
Ours3.4563.3833.3323.5203.423
" + }, + { + "type": "table_caption", + "bbox": [ + 0.09, + 0.255, + 0.483, + 0.283 + ], + "angle": 0, + "content": "Table 1. User study on generated 3D objects (Our results are raw outputs from LDM without SDS refinement)." + }, + { + "type": "table", + "bbox": [ + 0.107, + 0.305, + 0.47, + 0.366 + ], + "angle": 0, + "content": "
Shap-EDirect3DOurs
openai:ViT-L-140.23980.21520.2456
openai:ViT-L-14-3360.24260.22200.2462
" + }, + { + "type": "table_caption", + "bbox": [ + 0.09, + 0.376, + 0.484, + 0.418 + ], + "angle": 0, + "content": "Table 2. CLIP score(↑) for evaluation of similarity between rendered images and given text prompt((Our results are raw outputs from LDM without SDS refinement))." + }, + { + "type": "table", + "bbox": [ + 0.105, + 0.44, + 0.472, + 0.496 + ], + "angle": 0, + "content": "
More Realistic (%)More Detailed (%)Overall (%)
GaussianDreamer35.050.042.5
GSGEN17.55.011.3
Ours47.545.046.3
" + }, + { + "type": "table_caption", + "bbox": [ + 0.117, + 0.506, + 0.456, + 0.521 + ], + "angle": 0, + "content": "Table 3. User study on generated 3D objects (with SDS)." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.531, + 0.484, + 0.651 + ], + "angle": 0, + "content": "with given prompt. 49 users participated in the user study to score the over 50 3D samples from 1 to 5 points, and the average results are shown in Tab. 1. As for the CLIP score, the open-source repository t2vmetrics [24] is used to calculate the CLIP score on two versions of ViT models, and the results are demonstrated in Tab. 2. Both the CLIP score and the user study indicate that the proposed method produces better performance." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.673, + 0.391, + 0.691 + ], + "angle": 0, + "content": "5.3. Optional Texture Boosting by SDS" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.7, + 0.483, + 0.82 + ], + "angle": 0, + "content": "To compare with other SDS-based 3D GS generation methods, we implement a version of SDS for our mesh-binding GS representation as described in Section 4.5. GsGEN [4] and GaussianDreamer [50] are selected as our baselines. The generated samples, as illustrated in Fig. 5, highlight that with the SDS refiner, our DirectTriGS can deliver competitive outcomes comparable to the state-of-the-art 2D-lifting methods." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.826, + 0.484, + 0.903 + ], + "angle": 0, + "content": "Furthermore, we conducted a user study for quantitative evaluation. 40 users were tasked with ranking 15 samples from different methods based on two criteria: realism and level of detail. The results indicate that our approach marginally outperforms the other baseline methods." + }, + { + "type": "table", + "bbox": [ + 0.518, + 0.186, + 0.905, + 0.218 + ], + "angle": 0, + "content": "
TFLOPsParas.(M)GPU Mem.(GB)Running Time (Sec.)
Generation0.0082356.652.9416.2
" + }, + { + "type": "table_caption", + "bbox": [ + 0.516, + 0.227, + 0.901, + 0.241 + ], + "angle": 0, + "content": "Table 4. Generation inference efficiency, with the batch size is 1." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.249, + 0.667, + 0.266 + ], + "angle": 0, + "content": "5.4. Ablation Study" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.271, + 0.907, + 0.407 + ], + "angle": 0, + "content": "3D Diffusion for voxel-based Gaussian Splitting. We attempt to do GS attribute generation conditioned on given voxel occupancy via 3D diffusion model. However, even a toy experiment of over-fitting one single object failed, which may attributes to the complex multiple channels with different distributions especially the non-Euclidean ones such as quaternion. As shown in Fig.6, the diffusion model learns to generate color but fails to generate splat scaling, opacity and orientation." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.407, + 0.906, + 0.527 + ], + "angle": 0, + "content": "Triplane Diffusion without VAE. Since VAE inherently involves a reconstruction loss, we attempt to use direct diffusion on the triplane space for generation. The experiment results show that such direct diffusion may cause serious noise on decoded geometry. Randomly generated samples are visualized in Fig.7. A possible reason is that the multichannels of triplane contain considerable noisy redundancy, which is difficult for the LDM to filter." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.528, + 0.906, + 0.575 + ], + "angle": 0, + "content": "Inference Efficiency. The inference efficiency is listed in Table 4. This experiment is conducted on the platform equipped with RTX3090 GPU with 24GB memory." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.587, + 0.805, + 0.605 + ], + "angle": 0, + "content": "6. Conclusion & Acknowledgment" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.612, + 0.907, + 0.779 + ], + "angle": 0, + "content": "We have presented a novel framework DirectTriGS for Gaussian Splatting Field generation. DirectTriGS mainly consists of 3 parts: 1) a light-weight triplane representation for 3D object with the format of Gaussian Splatting; 2) a fully differentiable TriRenderer which can decode triplane to original GS and render it to multi-view images; 3) the triplane VAE and staged diffusion model for the whole generation process. By utilizing our DirectTriGS, the intricate GS data can be generated directly and efficiently. 4) Additionally, we incorporate a SDS refiner to further improve the texture and details of generated objects." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.779, + 0.909, + 0.899 + ], + "angle": 0, + "content": "This project is funded in part by National Key R&D Program of China Project 2022ZD0161100, by the Centre for Perceptual and Interactive Intelligence (CPII) Ltd under the Innovation and Technology Commission (ITC)'s InnoHK, by General Research Fund of Hong Kong RGC Project 14204021, by NSFC-RGC Joint Fund Project N_CUHK498/24. Hongsheng Li is a PI of CPII under the InnoHK." + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "16234" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.111, + 0.092, + 0.34, + 0.204 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.359, + 0.092, + 0.622, + 0.209 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.642, + 0.092, + 0.882, + 0.204 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.134, + 0.207, + 0.34, + 0.321 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.347, + 0.212, + 0.655, + 0.321 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.665, + 0.208, + 0.865, + 0.315 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.127, + 0.334, + 0.348, + 0.416 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.361, + 0.323, + 0.644, + 0.422 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.669, + 0.322, + 0.868, + 0.419 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.407, + 0.427, + 0.589, + 0.44 + ], + "angle": 0, + "content": "(a) Comparison with Direct3D." + }, + { + "type": "image", + "bbox": [ + 0.111, + 0.463, + 0.355, + 0.584 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.371, + 0.464, + 0.612, + 0.58 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.626, + 0.463, + 0.888, + 0.584 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.116, + 0.588, + 0.351, + 0.81 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.352, + 0.588, + 0.625, + 0.81 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.644, + 0.589, + 0.882, + 0.813 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.407, + 0.82, + 0.581, + 0.834 + ], + "angle": 0, + "content": "(b) Comparison with Shap-E." + }, + { + "type": "image_caption", + "bbox": [ + 0.2, + 0.868, + 0.798, + 0.882 + ], + "angle": 0, + "content": "Figure 4. Comparison of generated 3D objects (Raw results without any refinement such as SDS)." + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.946, + 0.519, + 0.957 + ], + "angle": 0, + "content": "16235" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.162, + 0.088, + 0.824, + 0.684 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.119, + 0.692, + 0.879, + 0.708 + ], + "angle": 0, + "content": "Figure 5. Refinement of unsatisfactory results by SDS, compared with pure SDS-based methods GsGEN and GaussianDreamer." + }, + { + "type": "image", + "bbox": [ + 0.186, + 0.735, + 0.33, + 0.83 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.167, + 0.845, + 0.349, + 0.873 + ], + "angle": 0, + "content": "Figure 6. Failure of GS attribute generation." + }, + { + "type": "image", + "bbox": [ + 0.366, + 0.737, + 0.83, + 0.828 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.364, + 0.841, + 0.83, + 0.868 + ], + "angle": 0, + "content": "Figure 7. Noisy results of direct diffusion on triplane. The condition prompt are \"train\", \"watermelon\" and \"box\"." + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.946, + 0.52, + 0.957 + ], + "angle": 0, + "content": "16236" + } + ], + [ + { + "type": "title", + "bbox": [ + 0.093, + 0.09, + 0.188, + 0.106 + ], + "angle": 0, + "content": "References" + }, + { + "type": "ref_text", + "bbox": [ + 0.101, + 0.115, + 0.484, + 0.184 + ], + "angle": 0, + "content": "[1] Hansheng Chen, Jiatao Gu, Anpei Chen, Wei Tian, Zhuowen Tu, Lingjie Liu, and Hao Su. Single-stage diffusion nef: A unified approach to 3d generation and reconstruction. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 2416-2425, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.101, + 0.186, + 0.484, + 0.254 + ], + "angle": 0, + "content": "[2] Yiwen Chen, Zilong Chen, Chi Zhang, Feng Wang, Xiaofeng Yang, Yikai Wang, Zhongang Cai, Lei Yang, Huaping Liu, and Guosheng Lin. Gaussianeditor: Swift and controllable 3d editing with gaussian splatting. arXiv preprint arXiv:2311.14521, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.256, + 0.483, + 0.31 + ], + "angle": 0, + "content": "[3] Yurui Chen, Chun Gu, Junzhe Jiang, Xiatian Zhu, and Li Zhang. Periodic vibration gaussian: Dynamic urban scene reconstruction and real-time rendering. arXiv preprint arXiv:2311.18561, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.312, + 0.484, + 0.367 + ], + "angle": 0, + "content": "[4] Zilong Chen, Feng Wang, Yikai Wang, and Huaping Liu. Text-to-3d using gaussian splatting. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 21401-21412, 2024. 2, 5, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.368, + 0.483, + 0.422 + ], + "angle": 0, + "content": "[5] Kai Cheng, Xiaoxiao Long, Kaizhi Yang, Yao Yao, Wei Yin, Yuexin Ma, Wenping Wang, and Xuejin Chen. Gaussianpro: 3d gaussian splatting with progressive propagation. arXiv preprint arXiv:2402.14650, 2024. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.424, + 0.483, + 0.478 + ], + "angle": 0, + "content": "[6] Jaeyoung Chung, Suyoung Lee, Hyeongjin Nam, Jaerin Lee, and Kyoung Mu Lee. Luciddreamer: Domain-free generation of 3d gaussian splattering scenes. arXiv preprint arXiv:2311.13384, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.481, + 0.484, + 0.564 + ], + "angle": 0, + "content": "[7] Matt Deitke, Dustin Schwenk, Jordi Salvador, Luca Weihs, Oscar Michel, Eli VanderBilt, Ludwig Schmidt, Kiana Ehsani, Aniruddha Kembhavi, and Ali Farhadi. Objverse: A universe of annotated 3d objects. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 13142-13153, 2023. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.565, + 0.483, + 0.619 + ], + "angle": 0, + "content": "[8] Zhiwen Fan, Kevin Wang, Kairun Wen, Zehao Zhu, Dejia Xu, and Zhangyang Wang. Lightgaussian: Unbounded 3d gaussian compression with 15x reduction and \\(200+\\) fps. arXiv preprint arXiv:2311.17245, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.621, + 0.483, + 0.674 + ], + "angle": 0, + "content": "[9] Jiemin Fang, Junjie Wang, Xiaopeng Zhang, Lingxi Xie, and Qi Tian. Gaussianeditor: Editing 3d gaussians delicately with text instructions. arXiv preprint arXiv:2311.16037, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.677, + 0.483, + 0.746 + ], + "angle": 0, + "content": "[10] Jun Gao, Tianchang Shen, Zian Wang, Wenzheng Chen, Kangxue Yin, Daiqing Li, Or Litany, Zan Gojcic, and Sanja Fidler. Get3d: A generative model of high quality 3d textured shapes learned from images. Advances In Neural Information Processing Systems, 35:31841-31854, 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.747, + 0.483, + 0.789 + ], + "angle": 0, + "content": "[11] Anchit Gupta, Wenhan Xiong, Yixin Nie, Ian Jones, and Barlas Oğuz. 3dgen: Triplane latent diffusion for textured mesh generation. arXiv preprint arXiv:2303.05371, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.79, + 0.483, + 0.831 + ], + "angle": 0, + "content": "[12] Jonathan Ho, Ajay Jain, and Pieter Abbeel. Denoising diffusion probabilistic models. Advances in neural information processing systems, 33:6840-6851, 2020. 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.832, + 0.484, + 0.901 + ], + "angle": 0, + "content": "[13] Lukas Hollein, Ang Cao, Andrew Owens, Justin Johnson, and Matthias Nießner. Text2room: Extracting textured 3d meshes from 2d text-to-image models. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 7909-7920, 2023. 2" + }, + { + "type": "list", + "bbox": [ + 0.094, + 0.115, + 0.484, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.093, + 0.906, + 0.161 + ], + "angle": 0, + "content": "[14] Fangzhou Hong, Jiaxiang Tang, Ziang Cao, Min Shi, Tong Wu, Zhaoxi Chen, Shuai Yang, Tengfei Wang, Liang Pan, Dahua Lin, et al. 3dtopia: Large text-to-3d generation model with hybrid diffusion priors. arXiv preprint arXiv:2403.02234, 2024. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.164, + 0.906, + 0.219 + ], + "angle": 0, + "content": "[15] Binbin Huang, Zehao Yu, Anpei Chen, Andreas Geiger, and Shenghua Gao. 2d gaussian splatting for geometrically accurate radiance fields. In ACM SIGGRAPH 2024 Conference Papers, pages 1-11, 2024. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.221, + 0.905, + 0.262 + ], + "angle": 0, + "content": "[16] Lutao Jiang and Lin Wang. Brightdreamer: Generic 3d gaussian generative framework for fast text-to-3d synthesis. arXiv preprint arXiv:2403.11273, 2024. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.264, + 0.905, + 0.332 + ], + "angle": 0, + "content": "[17] Justin Johnson, Alexandre Alahi, and Li Fei-Fei. Perceptual losses for real-time style transfer and super-resolution. In Computer Vision-ECCV 2016: 14th European Conference, Amsterdam, The Netherlands, October 11-14, 2016, Proceedings, Part II 14, pages 694-711. Springer, 2016. 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.334, + 0.905, + 0.375 + ], + "angle": 0, + "content": "[18] Xiaoliang Ju, Zhaoyang Huang, Yijin Li, Guofeng Zhang, Yu Qiao, and Hongsheng Li. Diffindscene: Diffusion-based high-quality 3d indoor scene generation, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.378, + 0.905, + 0.417 + ], + "angle": 0, + "content": "[19] Heewoo Jun and Alex Nichol. Shap-e: Generating conditional 3d implicit functions. arXiv preprint arXiv:2305.02463, 2023. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.42, + 0.905, + 0.474 + ], + "angle": 0, + "content": "[20] Tero Karras, Miika Aittala, Timo Aila, and Samuli Laine. Elucidating the design space of diffusion-based generative models. Advances in neural information processing systems, 35:26565-26577, 2022. 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.477, + 0.905, + 0.532 + ], + "angle": 0, + "content": "[21] Bernhard Kerbl, Georgios Kopanas, Thomas Leimkuhler, and George Drettakis. 3d gaussian splatting for real-time radiance field rendering. ACM Transactions on Graphics, 42 (4):1-14, 2023. 1, 2, 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.534, + 0.905, + 0.589 + ], + "angle": 0, + "content": "[22] Samuli Laine, Janne Hellsten, Tero Karras, Yeongho Seol, Jaakko Lehtinen, and Timo Aila. Modular primitives for high-performance differentiable rendering. ACM Transactions on Graphics, 39(6), 2020. 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.591, + 0.905, + 0.659 + ], + "angle": 0, + "content": "[23] Yushi Lan, Fangzhou Hong, Shuai Yang, Shangchen Zhou, Xuyi Meng, Bo Dai, Xingang Pan, and Chen Change Loy. Ln3diff: Scalable latent neural fields diffusion for speedy 3d generation. In European Conference on Computer Vision, pages 112-130. Springer, 2025. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.661, + 0.905, + 0.716 + ], + "angle": 0, + "content": "[24] Zhiqiu Lin, Deepak Pathak, Baiqi Li, Jiayao Li, Xide Xia, Graham Neubig, Pengchuan Zhang, and Deva Ramanan. Evaluating text-to-visual generation with image-to-text generation. arXiv preprint arXiv:2404.01291, 2024. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.718, + 0.905, + 0.786 + ], + "angle": 0, + "content": "[25] Minghua Liu, Chao Xu, Haian Jin, Linghao Chen, Mukund Varma T, Zexiang Xu, and Hao Su. One-2-3-45: Any single image to 3d mesh in 45 seconds without per-shape optimization. Advances in Neural Information Processing Systems, 36, 2024. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.789, + 0.905, + 0.858 + ], + "angle": 0, + "content": "[26] Qihao Liu, Yi Zhang, Song Bai, Adam Kortylewski, and Alan Yuille. Direct-3d: Learning direct text-to-3d generation on massive noisy 3d data. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 6881–6891, 2024. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.86, + 0.905, + 0.901 + ], + "angle": 0, + "content": "[27] Ruoshi Liu, Rundi Wu, Basile Van Hoorick, Pavel Tokmakov, Sergey Zakharov, and Carl Vondrick. Zero-1-to-3: Zero-shot one image to 3d object. In Proceedings of the" + }, + { + "type": "list", + "bbox": [ + 0.517, + 0.093, + 0.906, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "16237" + } + ], + [ + { + "type": "ref_text", + "bbox": [ + 0.126, + 0.092, + 0.482, + 0.12 + ], + "angle": 0, + "content": "IEEE/CVF International Conference on Computer Vision, pages 9298-9309, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.122, + 0.483, + 0.205 + ], + "angle": 0, + "content": "[28] Xiaoxiao Long, Yuan-Chen Guo, Cheng Lin, Yuan Liu, Zhiyang Dou, Lingjie Liu, Yuexin Ma, Song-Hai Zhang, Marc Habermann, Christian Theobalt, et al. Wonder3d: Single image to 3d using cross-domain diffusion. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 9970-9980, 2024. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.207, + 0.483, + 0.261 + ], + "angle": 0, + "content": "[29] Tao Lu, Mulin Yu, Linning Xu, Yuanbo Xiangli, Limin Wang, Dahua Lin, and Bo Dai. Scaffold-gs: Structured 3d gaussians for view-adaptive rendering. arXiv preprint arXiv:2312.00109, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.264, + 0.482, + 0.304 + ], + "angle": 0, + "content": "[30] Hidenobu Matsuki, Riku Murai, Paul HJ Kelly, and Andrew J Davison. Gaussian splattering slam. arXiv preprint arXiv:2312.06741, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.307, + 0.483, + 0.375 + ], + "angle": 0, + "content": "[31] Gal Metzer, Elad Richardson, Or Patashnik, Raja Giryes, and Daniel Cohen-Or. Latent-nerf for shape-guided generation of 3d shapes and textures. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 12663–12673, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.378, + 0.482, + 0.445 + ], + "angle": 0, + "content": "[32] Ben Mildenhall, Pratul P Srinivasan, Matthew Tancik, Jonathan T Barron, Ravi Ramamoorthi, and Ren Ng. Nerf: Representing scenes as neural radiance fields for view synthesis. Communications of the ACM, 65(1):99-106, 2021. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.448, + 0.483, + 0.53 + ], + "angle": 0, + "content": "[33] Jacob Munkberg, Jon Hasselgren, Tianchang Shen, Jun Gao, Wenzheng Chen, Alex Evans, Thomas Müller, and Sanja Fidler. Extracting triangular 3d models, materials, and lighting from images. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 8280-8290, 2022. 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.533, + 0.483, + 0.587 + ], + "angle": 0, + "content": "[34] Alex Nichol, Heewoo Jun, Prafulla Dhariwal, Pamela Mishkin, and Mark Chen. Point-e: A system for generating 3d point clouds from complex prompts. arXiv preprint arXiv:2212.08751, 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.59, + 0.483, + 0.63 + ], + "angle": 0, + "content": "[35] Ben Poole, Ajay Jain, Jonathan T Barron, and Ben Mildenhall. Dreamfusion: Text-to-3d using 2d diffusion. arXiv preprint arXiv:2209.14988, 2022. 1, 2, 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.633, + 0.483, + 0.701 + ], + "angle": 0, + "content": "[36] Robin Rombach, Andreas Blattmann, Dominik Lorenz, Patrick Esser, and Björn Ommer. High-resolution image synthesis with latent diffusion models. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 10684-10695, 2022. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.704, + 0.483, + 0.758 + ], + "angle": 0, + "content": "[37] Tianchang Shen, Jun Gao, Kangxue Yin, Ming-Yu Liu, and Sanja Fidler. Deep marching tetrahedra: a hybrid representation for high-resolution 3d shape synthesis. In Advances in Neural Information Processing Systems (NeurIPS), 2021. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.761, + 0.483, + 0.828 + ], + "angle": 0, + "content": "[38] Tianchang Shen, Jacob Munkberg, Jon Hasselgren, Kangxue Yin, Zian Wang, Wenzheng Chen, Zan Gojcic, Sanja Fidler, Nicholas Sharp, and Jun Gao. Flexible isosurface extraction for gradient-based mesh optimization. ACM Trans. Graph., 42(4), 2023. 3, 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.831, + 0.483, + 0.871 + ], + "angle": 0, + "content": "[39] Yichun Shi, Peng Wang, Jianglong Ye, Mai Long, Kejie Li, and Xiao Yang. Mvdream: Multi-view diffusion for 3d generation. arXiv preprint arXiv:2308.16512, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.874, + 0.483, + 0.901 + ], + "angle": 0, + "content": "[40] J Ryan Shue, Eric Ryan Chan, Ryan Po, Zachary Ankner, Jiajun Wu, and Gordon Wetzstein. 3d neural field generation" + }, + { + "type": "list", + "bbox": [ + 0.093, + 0.092, + 0.483, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.549, + 0.093, + 0.905, + 0.134 + ], + "angle": 0, + "content": "using triplane diffusion. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 20875-20886, 2023. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.136, + 0.905, + 0.188 + ], + "angle": 0, + "content": "[41] Jiaxiang Tang, Jiawei Ren, Hang Zhou, Ziwei Liu, and Gang Zeng. Dreamgaussian: Generative gaussian splatting for efficient 3d content creation. arXiv preprint arXiv:2309.16653, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.191, + 0.905, + 0.244 + ], + "angle": 0, + "content": "[42] Zhou Wang, Alan C Bovik, Hamid R Sheikh, and Eero P Simoncelli. Image quality assessment: from error visibility to structural similarity. IEEE transactions on image processing, 13(4):600-612, 2004. 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.247, + 0.905, + 0.313 + ], + "angle": 0, + "content": "[43] Zhengyi Wang, Cheng Lu, Yikai Wang, Fan Bao, Chongxuan Li, Hang Su, and Jun Zhu. Prolificdreamer: High-fidelity and diverse text-to-3d generation with variational score distillation. Advances in Neural Information Processing Systems, 36, 2024. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.317, + 0.905, + 0.371 + ], + "angle": 0, + "content": "[44] Guanjun Wu, Taoran Yi, Jiemin Fang, Lingxi Xie, Xiaopeng Zhang, Wei Wei, Wenyu Liu, Qi Tian, and Xinggang Wang. 4d gaussian splatting for real-time dynamic scene rendering. arXiv preprint arXiv:2310.08528, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.373, + 0.905, + 0.455 + ], + "angle": 0, + "content": "[45] Tong Wu, Jiarui Zhang, Xiao Fu, Yuxin Wang, Liang Pan Jiawei Ren, Wayne Wu, Lei Yang, Jiaqi Wang, Chen Qian, Dahua Lin, and Ziwei Liu. Omniobject3d: Large-vocabulary 3d object dataset for realistic perception, reconstruction and generation. In IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), 2023. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.457, + 0.905, + 0.523 + ], + "angle": 0, + "content": "[46] Zijie Wu, Yaonan Wang, Mingtao Feng, He Xie, and Ajmal Mian. Sketch and text guided diffusion model for colored point cloud generation. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 8929-8939, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.526, + 0.905, + 0.593 + ], + "angle": 0, + "content": "[47] Zhennan Wu, Yang Li, Han Yan, Taizhang Shang, Weixuan Sun, Senbo Wang, Ruikai Cui, Weizhe Liu, Hiroyuki Sato, Hongdong Li, et al. Blockfusion: Expandable 3d scene generation using latent tri-plane extrapolation. arXiv preprint arXiv:2401.17053, 2024. 1, 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.596, + 0.905, + 0.649 + ], + "angle": 0, + "content": "[48] Chi Yan, Delin Qu, Dong Wang, Dan Xu, Zhigang Wang, Bin Zhao, and Xuelong Li. Gs-slam: Dense visual slam with 3d gaussian splatting. arXiv preprint arXiv:2311.11700, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.651, + 0.905, + 0.704 + ], + "angle": 0, + "content": "[49] Zeyu Yang, Hongye Yang, Zijie Pan, Xiatian Zhu, and Li Zhang. Real-time photorealistic dynamic scene representation and rendering with 4d gaussian splatting. arXiv preprint arXiv:2310.10642, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.707, + 0.905, + 0.788 + ], + "angle": 0, + "content": "[50] Taoran Yi, Jiemin Fang, Junjie Wang, Guanjun Wu, Lingxi Xie, Xiaopeng Zhang, Wenyu Liu, Qi Tian, and Xinggang Wang. Gaussiandreamer: Fast generation from text to 3d gaussians by bridging 2d and 3d diffusion models. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 6796-6807, 2024. 5, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.79, + 0.905, + 0.831 + ], + "angle": 0, + "content": "[51] Zehao Yu, Anpei Chen, Binbin Huang, Torsten Sattler, and Andreas Geiger. Mip-splatting: Alias-free 3d gaussian splattering. arXiv preprint arXiv:2311.16493, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.833, + 0.905, + 0.9 + ], + "angle": 0, + "content": "[52] Bowen Zhang, Yiji Cheng, Jiaolong Yang, Chunyu Wang, Feng Zhao, Yansong Tang, Dong Chen, and Baining Guo. Gaussiancube: Structuring gaussian splatting using optimal transport for 3d generative modeling. arXiv preprint arXiv:2403.19655, 2024. 2, 5" + }, + { + "type": "list", + "bbox": [ + 0.517, + 0.093, + 0.905, + 0.9 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.52, + 0.957 + ], + "angle": 0, + "content": "16238" + } + ], + [ + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.092, + 0.487, + 0.162 + ], + "angle": 0, + "content": "[53] Zi-Xin Zou, Zhipeng Yu, Yuan-Chen Guo, Yangguang Li, Ding Liang, Yan-Pei Cao, and Song-Hai Zhang. Triplane meets gaussian splatting: Fast and generalizable single-view 3d reconstruction with transformers. arXiv preprint arXiv:2312.09147, 2023. 2, 4" + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.946, + 0.52, + 0.957 + ], + "angle": 0, + "content": "16239" + } + ] +] \ No newline at end of file diff --git a/2025/DirectTriGS_ Triplane-based Gaussian Splatting Field Representation for 3D Generation/cb304427-02fe-443c-baf8-93b01f1651af_origin.pdf b/2025/DirectTriGS_ Triplane-based Gaussian Splatting Field Representation for 3D Generation/cb304427-02fe-443c-baf8-93b01f1651af_origin.pdf new file mode 100644 index 0000000000000000000000000000000000000000..94b3d3bd534ff7a0c9b2f2a6da268e8d25e99648 --- /dev/null +++ b/2025/DirectTriGS_ Triplane-based Gaussian Splatting Field Representation for 3D Generation/cb304427-02fe-443c-baf8-93b01f1651af_origin.pdf @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:eec20267234b13784a07c778eac3d770a8c95e8835f1963182bcc900c1740e77 +size 4032146 diff --git a/2025/DirectTriGS_ Triplane-based Gaussian Splatting Field Representation for 3D Generation/full.md b/2025/DirectTriGS_ Triplane-based Gaussian Splatting Field Representation for 3D Generation/full.md new file mode 100644 index 0000000000000000000000000000000000000000..0888278b65ec9a35956ce07a6237b02a290d9eef --- /dev/null +++ b/2025/DirectTriGS_ Triplane-based Gaussian Splatting Field Representation for 3D Generation/full.md @@ -0,0 +1,326 @@ +# DirectTriGS: Triplane-based Gaussian Splitting Field Representation for 3D Generation + +Xiaoliang Ju, Hongsheng Li $^{1,2}$ $^{1}$ CUHK MMLab $^{2}$ CPII under InnoHK + +akira@link.cuhk.edu.hk + +# Abstract + +We present DirectTriGS, a novel framework designed for 3D object generation with Gaussian Splatting (GS). GS-based rendering for 3D content has gained considerable attention recently. However, there has been limited exploration in directly generating 3D Gaussians compared to traditional generative modeling approaches. The main challenge lies in the complex data structure of GS represented by discrete point clouds with multiple channels. To overcome this challenge, we propose employing the triplane representation, which allows us to represent Gaussian Splatting as an image-like continuous field. This representation effectively encodes both the geometry and texture information, enabling smooth transformation back to Gaussian point clouds and rendering into images by a TriRenderer, with only 2D supervisions. The proposed TriRenderer is fully differentiable, so that the rendering loss can supervise both texture and geometry encoding. Furthermore, the triplane representation can be compressed using a Variational Autoencoder (VAE), which can subsequently be utilized in latent diffusion to generate 3D objects. The experiments demonstrate that the proposed generation framework can produce high-quality 3D object geometry and rendering results in the text-to-3D task. + +# 1. Introduction + +Neural rendering has grown to a focal point in rendering techniques in recent years, as it achieves more realistic rendering effects by leveraging the great expressiveness of neural network. The representatives are neural radiance field (NeRF) [32] and the newly emerged Gaussian Splatting (GS) [21]. However, when applied to the field of 3D generation, the slow rendering and training speed become a strong limitation for NeRF. While GS is more flexible with greater rendering efficiency and editability, few works have addressed the challenge of direct GS generation in 3D due to its complex data structure. In this paper, we propose a + +novel approach called DirectTriGS, which encodes GS as triplane representation, and introduce its corresponding renderer, TriRenderer. Subsequently, we apply latent diffusion on the Triplane representation to generate high-quality GS objects. + +Gaussian splatting uses multi-channel point cloud of "splats" to describe the 3D contents. With a differentiable splats rasterizer, GS has fast rendering speed. However, the sparsity, multiple channels, and uneven distributed density of 3D GS bring great difficulty for 3D generation. We propose to encode GS contents as multi-channel triplane representations, which have been shown to possess favorable properties for representing 3D geometry or NeRF, as demonstrated in previous works Shue et al. [40], Wu et al. [47]. In our work, we leverage the triplane representation to encode both the geometry information and other GS attributes in two separate groups of channels. By training on a dataset of 3D objects, we obtain a shared TriRenderer that is capable of decoding different triplane representations to GS and then render it to image. TriRenderer is fully differentiable, enabling the use of only 2D rendering loss to supervise both the texture and geometry of 3D GS. The advantage of using Triplane to represent GS are two folds. First, it leads to high memory efficiency compared with dense voxels and is expressive enough for generating various 3D GS. Second, Triplane is more compatible with the convolution-based encoders compared with the original sparse GS point cloud, which require specifically designed networks on processing sparse point clouds. + +We follow stable diffusion [36] to train DirectTriGS on the proposed triplane representation. Specifically, a VAE is designed to further convert the Triplanes into latent code. We use two separate decoders to decouple the decoding of geometry and GS attributes. Second, we roll out the triplane latent to an expanded multi-channel image, then exploit staged latent diffusion to do generation. Two-stage diffusion is employed to generate the geometry and the corresponding GS appearance. Finally, score distillation sampling (SDS) [35] is adopted as a optional post-processing to refine or restyle the generated 3D objects. + +Our contribution can summarize as follows. 1) We propose a Triplane representation for direct 3D GS generation, which is memory efficient to derive GS point clouds with vivid rendering. 2) We design a fully differentiable TriRenderer to enable the end-to-end training of triplane representations, with only 2D supervisions. 3) We develop a Triplane-based GS generation framework DirectTriGS that incorporates a specially designed VAE, latent diffusion module, and a SDS based refiner. 4) Experiments demonstrate our method produces competitive performance in both 3D geometry and multi-view rendering quality in text-to-3D tasks. + +# 2. Related Works + +3D Object Generation. 1) Multi-view images-to-3D. The first route generates multi-view colored 3D image and then reconstruct the 3D shape and project image to texture, or use score distillation sampling to build 3D object from images produced by 2D diffusion, such as [6, 13, 27, 35, 39, 41]. Since these approaches are actually 2D generator, it is difficult to maintain 3D consistency, which may lead to Janus problem. 2) Image-to-3D. The second route attempt to retrieve 3D geometry from a single input image, such as [25, 28, 53]. With the certain one image as input, the users may be more concerned with the reconstruction quality. 3)Text-to-3D. The third group mainly operate in 3D, which can be clustered by different 3D representation. For example, [10, 11] are able to generate high-quality 3D textured meshes using differentiable rendering; [34, 46] generate colored pointcloud; [1, 14, 23, 31] generate NeRF volumes; [18, 47] generate SDF volumes. + +Remark. The input conditions of these three generation methods are gradually becoming abstract, corresponding to increasing difficulty of the generation task, commonly resulting in a gradual decline in the generation quality. + +Gaussian Splatting. Gaussian Splatting [21] exploits multi-channel pointcloud of splats to describe the 3D content, which allows real-time rendering of photorealistic scenes. Recently, there emerges so many research focusing on its reconstruction quality enhancement and algorithm optimization [5, 8, 15, 29, 51]. Another related active area includes 4D dynamic GS modeling [3, 44, 49], scene editing [2, 9] and its applications like SLAM [30, 48]. + +As for GS generation, some works exploit multi-view image based route such as [6, 41], where the image generator output image from required views to enhance the GS rendering in a recursive manner. Rare works explore the GS generation directly in 3D space, such as [53] generates a point cloud using a Transformer structure with 2D image as input, and build a mapping of image to tri-plane encoding of GS attribute; GaussianCube [52] proposes to use optimal transport to model the 3D GS for text-to-3D generation; [16] combine explicit anchor points with implicit + +triplane to directly generate GS points. + +# 3. Revisiting Gaussian Splitting + +Gaussian Splatting [21] uses a set of Gaussian points to describe a 3D object. The Gaussian points are defined by a full 3D covariance matrix $\pmb{\Sigma}$ in a world space as $G(\mathbf{x}) = e^{-\frac{1}{2}\mathbf{x}^T\pmb{\Sigma}^{-1}\mathbf{x}}$ , centered at point (mean) $\mathbf{p} \in \mathbb{R}^3$ . Each point is with a opacity scalar $\alpha \in [0,1]$ for blending and a series of Spherical Harmonics (SH) coefficients to correctly capture the view-dependent appearance of the scene. The number of SH coefficients are $3 \times (n + 1)^2$ , where $n$ denotes the SH order, and higher $n$ corresponds to more accurate view-dependent appearance. + +During rendering, the Gaussians are project to 2D given a viewing transformation $\mathbf{W}$ . The covariance is transformed as $\boldsymbol{\Sigma}^{\prime} = \mathbf{J}\mathbf{W}\boldsymbol{\Sigma}\mathbf{W}^{T}\mathbf{J}^{T}$ , where $\mathbf{J}$ is the Jacobian of the affine approximation of the projective transformation. To ensure $\boldsymbol{\Sigma}$ to be semi-positive in the whole training process, $\boldsymbol{\Sigma}$ can be represented by a Cholesky decomposition as $\boldsymbol{\Sigma} = \mathbf{R}\mathbf{S}\mathbf{S}^{T}\mathbf{R}^{T}$ , represented by a tuple $s = (s_{1}, s_{2}, s_{3})$ for scaling and a unnormalized quaternion $\mathbf{q} \in \mathbb{R}^{4}$ for rotation. In summary, each Gaussian point are a union set of position $\mathbf{p}$ and its GS attributes: + +$$ +\mathcal {G} _ {\mathbf {p}} := \{s, \mathbf {q}, \mathrm {S H}, \alpha \}. \tag {1} +$$ + +GS pointcloud can be efficiently rendered by a rasterization-based splatting renderer [21], which is fully differentiable. + +# 4. Methodology + +Our motivation is to directly generate 3D GS with fast speed by utilizing the proposed triplane-based GS representation. While 2D-lifting methods like SDS-based approaches yield intricate results, they are time-intensive and susceptible to Janus problem. Direct generation, notably with fast sampling, can complete 3D generation in under a minute, contrasting with SDS methods that take over 10 minutes or longer, such as [4][43], posing challenges for users. Therefore, direct 3D generation is more practical, with 2D-lifting reserved for enhancing texture details when needed. + +The procedures of our generation framework can be divide into 4 stages as shown in Fig.1 (a). Firstly, we encode the 3D objects into triplanes using the TriRenderer. The triplane encoding and TriRenderer training can be concurrently executed with only the multi-view RGBA image and camera poses as supervision. Secondly, a VAE is trained to compress the triplanes to latent space, enabling effective capture of high-level information for subsequent modules. Thirdly, a diffusion model is trained on the triplane latent code. During inference, the diffusion model produces the latent code, which can be decoded back to its corresponding triplane by the VAE decoder. Subsequently, the TriRenderer decodes the triplane into the standard GS point + +![](images/0f6b3129f9dafeb1ea97a8c07c5e1e6f91700926f675d7de8e63579c4810257d.jpg) +(a) The overall procedure of our 3D content generation framework. + +![](images/e89ee7f07caa7332f781434de1779f2cd6006aa281d39e6a1c03fab792483ba9.jpg) + +![](images/ac0245851813463f977695cdcb7376eb06ba9d71159ac02010405b1d656ef03f.jpg) + +![](images/698df0cca85d4e2868585e5d2158dea4c7d18ef59fe1938a7f12036db508cf97.jpg) + +Figure 1. The overview of the proposed 3D GS generation framework and the core component TriRenderer. We use triplane to represent 3D objects, and design a TriRenderer to decode triplane to GS pointcloud and render multi-view images. +![](images/a8f4465db2ff41f986e7112ab90158b5fd05a67645196976b27a8d94f6dc0e1c.jpg) +(b) The forward pass of the TriRenderer pipeline: rendering a Triolane to multi-view images. + +cloud. Finally, an SDS-based refiner can be utilized as an optional post-processing step to enhance or restyle the generated 3D object. Leveraging the pretrained 2D diffusion model within the SDS, the generated GS can be further improved to achieve a more detailed appearance. + +# 4.1. Triplane-based GS Field + +We employ triplanes to convert the discrete multi-channel GS points into a continuous field, tailored for more efficient encoding in the subsequent generative model. Triplanes are 3 planes formed by every 2 axes of $x,y,z$ axes in 3D space, where every 3D point $\mathbf{p}$ can query these planes to get corresponding features $F_{\mathbf{p}} = (F_{xy},F_{xz},F_{yz})$ using orthogonal projection. In real implementation, triplane is a tensor with the size of $3\times H\times H\times C$ , where $H$ denote the resolution along every axis and $C$ is the feature channel. The query of point with arbitrary continuous coordinate should be a bi-linear interpolation in the triplane grids. Considering GS is a set of points with multiple channels of attributes, it is natural to use triplane for GS encoding. If the triplane feature and GS attributes of a point can be transformed to each other, the sparse GS pointcloud of every object can be represented as continuous triplane GS field for further encoding. We use separate channels to encode 3D geometry and other GS attributes as described in Eq. 1, which is an experimental setting for better convergence in the training. + +# 4.2. TriRenderer + +The TriRenderer serves as the crucial component for converting a 3D object into a triplane-based GS field. It acts as a fully differentiable bridge connecting the GS field with + +rendering, enabling the optimization from 2D images to geometry and appearance encoding. + +As depicted in Fig. 1 (b), the TriRenderer comprises a geometry branch and a GS attribute branch, each equipped with independent decoders to decode the triplane. The geometry branch is responsible for retrieve the geometry as triangular mesh from triplane, with a surface sampler for GS pointcloud sampling. Then GS attribute branch uses the obtained pointcloud to query the triplane to obtain the GS attribute corresponding to each point. In this way, the GS pointcloud in the original format is retrieved, and can be rendered using the original GS renderer. It is worth mentioning that all triplanes of different object share a common TriRenderer, which ensures the features on different triplanes subject to a similar distribution. + +Geometry Branch. We use signed distance function (SDF) to represent the geometry, and allow every vertex deform as [37], so that the geometry branch decodes $F_{p}$ to a SDF volume and its vertex deformation. By query all grid coordinates of the designated resolution $L \times L \times L$ , the deformable SDF volume is reconstructed. Then we exploit a differentiable Marching Cube algorithm FlexiCubes [38] to extract triangular mesh from the SDF volume. + +Surface Sampling. Considering that GS pointcloud is generally gathered on the surface of objects, we randomly sample GS points on the faces of triangular mesh with barycentric coordinates. We expect that every splat is flat and with a normal consistent with its source triangular face. Assume a face is formed by vertices $< v_{a}, v_{b}, v_{c} >$ , and vector $\mathbf{v}_{1} = (v_{a} - v_{b})$ , $\mathbf{v}_{2} = (v_{b} - v_{c})$ , the face normal can be calculated as $\mathbf{n} = \mathbf{v}_{1} \times \mathbf{v}_{2}$ . Then the rotation matrix of every + +splat drawn from this face can be formulated as + +$$ +\mathbf {R} = \left[ \mathbf {v} _ {1}, \mathbf {n} \times \mathbf {v} _ {1}, \mathbf {n} \right], \tag {2} +$$ + +which can be transform to $\mathbf{q}$ in Eq.1 by standard matrix-to-quaternion algorithm. To make the splats flat, $s_3$ in Eq.1 is fixed to an infinite small value. + +To further reduce the computational load in the subsequent GS rendering process, only faces oriented toward the camera are sampled. These face indices can be obtained through fast rasterization of the mesh faces. Note this rasterization is not required to be differentiable, and we use the library Nvdiffrast [22] for implementation. + +GS Attribute Branch. By querying the triplane GS channels with the sampled pointcloud, we can obtain the GS feature and use the GS attribute branch to transform them to the rest GS attributes. Then the GS renderer can render GS pointcloud to multi-view images. Since different GS attributes have different numerical scales and distributions, we customize individual headers to decode them respectively, which is similar to [53]. + +Training. The training consists of 2 stages. In the first stage, a small batch of data is used to train the triplanes and the TriRenderer together. The second stage involves the whole dataset to train all triplanes with the parameters of TriRenderer frozen. In this way, we can handle large dataset by distributed training in the second stage, and the same TriRenderer can be shared by all triplanes. There is no need to pre-train the GS pointcloud using original GS training as [21]. + +The only required supervision data is $N$ multi-view images $\{I_i\}_{i=1}^N$ with camera poses, and the training losses consist of the rendering loss $L_{\mathrm{render}}$ and the geometric regularization $L_{\mathrm{geo}}$ . The rendering loss is a weighted sum as + +$$ +L _ {\text {r e n d e r}} = w _ {1} L _ {\text {a l p h a}} + w _ {2} L _ {\text {r g b}} + w _ {3} L _ {\text {p i p s}}, \tag {3} +$$ + +where $L_{\mathrm{alpha}}$ is a $L_{1}$ loss on the alpha map, namely the silhouette image loss; $L_{\mathrm{rgb}}$ is a combination of $L_{1}$ pixel loss and SSIM loss [42] between rendered image $\hat{I}_i$ and ground truth $I_i$ : + +$$ +L _ {\mathrm {r g b}} = \sum_ {i = 1} ^ {N} (1 - \beta) \| I _ {i} - \hat {I} _ {i} \| _ {1} + \beta \operatorname {S S I M} \left(I _ {i}, \hat {I} _ {i}\right), \tag {4} +$$ + +and $L_{\mathrm{pips}}$ is the perceptual loss [17]. $w_{1}, w_{2}, w_{3}$ and $\beta$ are all weighting factors determined by experiments. + +The geometric regularization loss on the SDF volume $V$ is defined as + +$$ +L _ {\text {g e o}} = \gamma_ {1} L _ {\text {d e v}} + \gamma_ {2} L _ {\text {w e i g h t i n g}} + \gamma_ {3} L _ {\text {C E}} + \gamma_ {4} L _ {\text {s i g n}}, \tag {5} +$$ + +where the $L_{\mathrm{dev}}$ and $L_{\mathrm{weighting}}$ are defined by Flex-cubes [38] to regularize the extracted connectivity and the weighting scale of SDF vertices. To penalize the sign + +changes on all grid edges, we follow [33] to define $L_{\mathrm{CE}}$ as + +$$ +L _ {\mathrm {C E}} = \sum_ {(s _ {a}, s _ {b}) \in \epsilon} \operatorname {C E} \left(\sigma \left(s _ {a}\right), \operatorname {s i g n} \left(s _ {b}\right)\right), \tag {6} +$$ + +where CE denotes cross entropy; $\epsilon$ is the set of all edges connecting vertices with different signs; and $\sigma$ is the sigmoid function. Finally, $L_{\mathrm{sign}}$ is designed to prevent the SDF volume from being trapped in an empty shape, i.e. a fully positive or fully negative SDF volume, + +$$ +L _ {\text {s i g n}} = \delta (V) M (| V |), \tag {7} +$$ + +where $\delta (V) = 1$ if $V$ is empty, otherwise $\delta (V) = 0$ . $M(|V|)$ denotes the mean of the absolute value of $V$ . + +# 4.3. Triplane VAE + +We employ an UNet-like structure of VAE to compress the triplane to latent space. The training pipeline is demonstrated in Fig. 2 (a). Considering that the 3 planes of triplanes encodes features from 3 orthogonal view direction, these 3 planes should be homogeneous data, so that We reshape the triplanes of batchsize $B$ to a new batch as $B \times 3 \times H \times H \times C$ to $3B \times H \times H \times C$ for the VAE training. We use decoupled encoders and decoders for geometry and GS channels of triplanes, for experiments prove that a mixed encoding may lead to blur of the rendering results. After the triplane is reconstructed, the TriRenderer trained in Section.4.1 can be used to retrieve the GS pointclouds and render it to images. + +In the training process, same loss functions as Eq. 3 and Eq. 5 are used, along with an $L_{1}$ loss between the input and reconstructed triplanes, denoted as $L_{\mathrm{Tri}}$ . Additionally, a Kullback-Leibler divergence loss $L_{\mathrm{KL}}$ is included to ensure that the latent variables do not deviate significantly from a normal distribution. The total losses is summarized as + +$$ +L _ {\mathrm {V A E}} = L _ {\mathrm {T r i}} + L _ {\mathrm {G e o}} + L _ {\mathrm {R e n d e r}} + \gamma L _ {\mathrm {K L}}, \tag {8} +$$ + +where $\gamma$ is a small weight. + +# 4.4. Two-stage LDM for Text-to-3D Generation + +Taking into account the significant relevance of texture appearance to the underlying geometry, we proposed to utilize staged diffusion to generate geometry and GS attribute successively as shown in Fig. 2 (b). It is easy to implement it because the latent codes for geometry and other GS attributes are totally decoupled by VAE, as mentioned earlier. In this way, the geometry code can be a new condition for the second stage. + +We follow DDPM [12, 20] to implement the text-conditioned LDM. To better capture the relation between different planes, we roll out the triplane latent of $B \times 3 \times h \times h \times c$ to a image-like $B \times 3h \times h \times c$ as the input of diffusion model, and the generated result will be transformed back to the original shape for TriRenderer decoding. + +![](images/d4d455d7b936a86017daee34945deaf08686bcf6beb9a8f6ab8da1af198305b4.jpg) +Figure 2. Triplane VAE and 2-stage diffusion. + +# 4.5. SDS-based Texture Booster + +To refine or restyle the generated GS pointcloud, we utilize a pretrained 2D diffusion model for the SDS training process [35] as shown in Fig. 3. As the generated 3D GS is constrained by the mesh reconstructed from SDF, detailed in Section 4.2, we maintain the GS splats' adherence as previously established. However, we enable the mesh vertices to shift within a limited radius to enhance the stability of the SDS process. Additionally, throughout the entire SDS procedure, there is no need for densification or pruning operations. This ensures that the number of points can be controlled to align with the number of mesh faces. + +With the image rendered from the initial GS pointcloud, the 2D diffusion model can produce images with better 2D/3D consistence in different views, so that the optimization process can converge rapidly without Janus problem. + +# 5. Experiment + +The experiments are conducted step-by-step according to the generation pipeline proposed as Fig.1 (a). We use OmniObject3D [45] dataset for toy experiments, and Objiverse [7] dataset for main evaluation. First, we sample 200K object to train their co顺应ing triplanes and TriNDERer. Then we train the VAE and diffusion model for 3DS GS generation. Finally, we exploit SDS to boost the texture quality for unsatisfactory generated objects. The procedure of data pre-processing and the implementation details are included in the supplementary materials. + +The experimental results primarily consist of the following components: 1) For examining the proposed triplane modeling for GS generation, we focus on the text-to-3D task and showcase the direct GS generation outcomes qualitatively and quantitatively. These results are compared with Shap-E [19], Direct3D [26], and the most recent Gaussian-Cube [52]. 2) In comparison to 2D-lifting GS generation methods, we present the performance of SDS refinement on unsatisfying samples, with GsGEN [4] and GaussianDreamer [50] as baselines. + +# 5.1. Triplane Fitting and VAE Reconstruction + +Triplane Fitting. We adopt a resolution of $3 \times 128 \times 128 \times 16$ for all triplane fitting, with a resolution of $64 \times 64 \times 64$ for the deformable SDF grids. We randomly sample 1000 objects for the shared TriRenderer training, and train all residual objects distributively. Every triplane is initialized with Gaussian noise before training. Under such settings, every triplane takes less than 30 seconds for geometry and GS appearance reconstruction. + +VAE Reconstruction. We use a down-sample factor of 4 to compress the triplanes to latent space. + +We visualize the triplane and the VAE reconstruction results of several data samples for a simple check, please refer to the supplementary materials. + +# 5.2. Direct Text-to-3D GS Generation by LDM + +In this section, we present the results of our 3D GS object generation without the SDS refinement, both quantitatively and qualitatively. Due to the current scarcity of work targeting direct 3DGS generation, we selected two state-of-the-art text-to-3D works based on other 3D representations(NeRF) Shap-E [19] and Direct3D [26] for comparison. For fairness, these selected methods also perform direct generation in the 3D domain without any 2D prior knowledge or refinement based on 2D diffusion. Additionally, for the limited space, we provide more generation results and comparison with other methods such as GaussianCube [52] in the supplementary materials. + +Qualitative Results. The generation results from different methods are listed in Fig. 4. Every generated sample is rendered from different views accompanied by the provided text prompt captioned beneath the images. Our method showcase enhanced proficiency in both geometry and rendering quality, resulting in sharper and clearer outputs, which can be further verified in the subsequent quantitative evaluation. For more generated samples, please check the supplementary materials. + +Quantitative Results. We use CLIP score to evaluate the text-to-3d consistency, and an user study is conducted to evaluate the generation results from various aspect such as geometry, texture, realistic rendering and the consistency + +![](images/1c5fae7808bcece9143adcb3cb11fbe33f733882894530679f672ec3c79ce3e9.jpg) +Figure 3. Score distillation process(SDS) for optional texture refinement. + +
Geometry↑Texture↑Realism↑Consistency↑Overall↑
Shap-E3.0803.0192.9473.1213.042
Direct3D3.2223.1503.1183.2423.183
Ours3.4563.3833.3323.5203.423
+ +Table 1. User study on generated 3D objects (Our results are raw outputs from LDM without SDS refinement). + +
Shap-EDirect3DOurs
openai:ViT-L-140.23980.21520.2456
openai:ViT-L-14-3360.24260.22200.2462
+ +Table 2. CLIP score(↑) for evaluation of similarity between rendered images and given text prompt((Our results are raw outputs from LDM without SDS refinement)). + +
More Realistic (%)More Detailed (%)Overall (%)
GaussianDreamer35.050.042.5
GSGEN17.55.011.3
Ours47.545.046.3
+ +with given prompt. 49 users participated in the user study to score the over 50 3D samples from 1 to 5 points, and the average results are shown in Tab. 1. As for the CLIP score, the open-source repository t2vmetrics [24] is used to calculate the CLIP score on two versions of ViT models, and the results are demonstrated in Tab. 2. Both the CLIP score and the user study indicate that the proposed method produces better performance. + +# 5.3. Optional Texture Boosting by SDS + +To compare with other SDS-based 3D GS generation methods, we implement a version of SDS for our mesh-binding GS representation as described in Section 4.5. GsGEN [4] and GaussianDreamer [50] are selected as our baselines. The generated samples, as illustrated in Fig. 5, highlight that with the SDS refiner, our DirectTriGS can deliver competitive outcomes comparable to the state-of-the-art 2D-lifting methods. + +Furthermore, we conducted a user study for quantitative evaluation. 40 users were tasked with ranking 15 samples from different methods based on two criteria: realism and level of detail. The results indicate that our approach marginally outperforms the other baseline methods. + +Table 3. User study on generated 3D objects (with SDS). + +
TFLOPsParas.(M)GPU Mem.(GB)Running Time (Sec.)
Generation0.0082356.652.9416.2
+ +Table 4. Generation inference efficiency, with the batch size is 1. + +# 5.4. Ablation Study + +3D Diffusion for voxel-based Gaussian Splitting. We attempt to do GS attribute generation conditioned on given voxel occupancy via 3D diffusion model. However, even a toy experiment of over-fitting one single object failed, which may attributes to the complex multiple channels with different distributions especially the non-Euclidean ones such as quaternion. As shown in Fig.6, the diffusion model learns to generate color but fails to generate splat scaling, opacity and orientation. + +Triplane Diffusion without VAE. Since VAE inherently involves a reconstruction loss, we attempt to use direct diffusion on the triplane space for generation. The experiment results show that such direct diffusion may cause serious noise on decoded geometry. Randomly generated samples are visualized in Fig.7. A possible reason is that the multichannels of triplane contain considerable noisy redundancy, which is difficult for the LDM to filter. + +Inference Efficiency. The inference efficiency is listed in Table 4. This experiment is conducted on the platform equipped with RTX3090 GPU with 24GB memory. + +# 6. Conclusion & Acknowledgment + +We have presented a novel framework DirectTriGS for Gaussian Splatting Field generation. DirectTriGS mainly consists of 3 parts: 1) a light-weight triplane representation for 3D object with the format of Gaussian Splatting; 2) a fully differentiable TriRenderer which can decode triplane to original GS and render it to multi-view images; 3) the triplane VAE and staged diffusion model for the whole generation process. By utilizing our DirectTriGS, the intricate GS data can be generated directly and efficiently. 4) Additionally, we incorporate a SDS refiner to further improve the texture and details of generated objects. + +This project is funded in part by National Key R&D Program of China Project 2022ZD0161100, by the Centre for Perceptual and Interactive Intelligence (CPII) Ltd under the Innovation and Technology Commission (ITC)'s InnoHK, by General Research Fund of Hong Kong RGC Project 14204021, by NSFC-RGC Joint Fund Project N_CUHK498/24. Hongsheng Li is a PI of CPII under the InnoHK. + +![](images/e02b0b9e340b51003f4a0af94ca505b938ab38185310a1aeae72930ffb0077bf.jpg) + +![](images/a8d52a97ba35677a7fd3927da59bc8ef5a17462e5f80d1462158dcac99b07aee.jpg) + +![](images/c451f8c67b51d102c44bd5f968217768c2cc62129beff2a09320df123d8df510.jpg) + +![](images/ecdf04e07c1cb4999e734e4316aba7b2d96e2e51b03be480578dcba990b57259.jpg) + +![](images/2ceb55e32e7787145e7f5b7b418411f8833e84c28a3f12e7ba41dfae5427d41f.jpg) + +![](images/c93a0fea47463fb7388dd9b94157b2696942a779c3ac022e0750f2e3979a18f9.jpg) + +![](images/939ef3c413b258a30bc7602ce90f5916e697e334b760bbc40b742c461fa27860.jpg) + +![](images/0db8dc473bcd8b9a0ec8866a96e32cf4eb5c73d79cb8f8e5cc1486d0b098f5af.jpg) +(a) Comparison with Direct3D. + +![](images/3bbb6ee05be3c129040b8c07a2869f056326c0048e52df83b029f2edbef790fa.jpg) + +![](images/7e68078254c1997822946878565641ef1c291448b2668b07fbb52580f39748ae.jpg) + +![](images/9a8f5f4e995cc14f49033c609e54d995cee059d85cbd38496bf0708bf7e03383.jpg) + +![](images/67d5c93c197f13c32b1ede71b73b93751838c6c0d5ccf21da19b76caa3ce5c98.jpg) + +![](images/ef41a3cd14b7ec0b934bd4833be2c8df5f63f92d873c950331dddcc120b9a90e.jpg) + +![](images/1c07367dad90481bf51654c2997d2915314471b3926331b615d5c02e577bc444.jpg) +(b) Comparison with Shap-E. + +![](images/64b850aa9cfd764ef1cd388e358115550c22469c903a819a75a3b9afb6f2b7ee.jpg) +Figure 4. Comparison of generated 3D objects (Raw results without any refinement such as SDS). + +![](images/f4e11a8500518460496587dd8c168e83c9614b6ac6d4ada9d2df5d5f61050a85.jpg) +Figure 5. Refinement of unsatisfactory results by SDS, compared with pure SDS-based methods GsGEN and GaussianDreamer. + +![](images/98d34b585d9c24f1d501d70cd8477cd25c8c9773309fff17b241ddca3c0e9f35.jpg) +Figure 6. Failure of GS attribute generation. + +![](images/95a0e5c3a1fbb9d463967d6d26657d1dad5d10e760be06d36e579c5a787388b9.jpg) +Figure 7. Noisy results of direct diffusion on triplane. The condition prompt are "train", "watermelon" and "box". + +# References + +[1] Hansheng Chen, Jiatao Gu, Anpei Chen, Wei Tian, Zhuowen Tu, Lingjie Liu, and Hao Su. Single-stage diffusion nef: A unified approach to 3d generation and reconstruction. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 2416-2425, 2023. 2 +[2] Yiwen Chen, Zilong Chen, Chi Zhang, Feng Wang, Xiaofeng Yang, Yikai Wang, Zhongang Cai, Lei Yang, Huaping Liu, and Guosheng Lin. Gaussianeditor: Swift and controllable 3d editing with gaussian splatting. arXiv preprint arXiv:2311.14521, 2023. 2 +[3] Yurui Chen, Chun Gu, Junzhe Jiang, Xiatian Zhu, and Li Zhang. Periodic vibration gaussian: Dynamic urban scene reconstruction and real-time rendering. arXiv preprint arXiv:2311.18561, 2023. 2 +[4] Zilong Chen, Feng Wang, Yikai Wang, and Huaping Liu. Text-to-3d using gaussian splatting. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 21401-21412, 2024. 2, 5, 6 +[5] Kai Cheng, Xiaoxiao Long, Kaizhi Yang, Yao Yao, Wei Yin, Yuexin Ma, Wenping Wang, and Xuejin Chen. Gaussianpro: 3d gaussian splatting with progressive propagation. arXiv preprint arXiv:2402.14650, 2024. 2 +[6] Jaeyoung Chung, Suyoung Lee, Hyeongjin Nam, Jaerin Lee, and Kyoung Mu Lee. Luciddreamer: Domain-free generation of 3d gaussian splattering scenes. arXiv preprint arXiv:2311.13384, 2023. 2 +[7] Matt Deitke, Dustin Schwenk, Jordi Salvador, Luca Weihs, Oscar Michel, Eli VanderBilt, Ludwig Schmidt, Kiana Ehsani, Aniruddha Kembhavi, and Ali Farhadi. Objverse: A universe of annotated 3d objects. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 13142-13153, 2023. 5 +[8] Zhiwen Fan, Kevin Wang, Kairun Wen, Zehao Zhu, Dejia Xu, and Zhangyang Wang. Lightgaussian: Unbounded 3d gaussian compression with 15x reduction and $200+$ fps. arXiv preprint arXiv:2311.17245, 2023. 2 +[9] Jiemin Fang, Junjie Wang, Xiaopeng Zhang, Lingxi Xie, and Qi Tian. Gaussianeditor: Editing 3d gaussians delicately with text instructions. arXiv preprint arXiv:2311.16037, 2023. 2 +[10] Jun Gao, Tianchang Shen, Zian Wang, Wenzheng Chen, Kangxue Yin, Daiqing Li, Or Litany, Zan Gojcic, and Sanja Fidler. Get3d: A generative model of high quality 3d textured shapes learned from images. Advances In Neural Information Processing Systems, 35:31841-31854, 2022. 2 +[11] Anchit Gupta, Wenhan Xiong, Yixin Nie, Ian Jones, and Barlas Oğuz. 3dgen: Triplane latent diffusion for textured mesh generation. arXiv preprint arXiv:2303.05371, 2023. 2 +[12] Jonathan Ho, Ajay Jain, and Pieter Abbeel. Denoising diffusion probabilistic models. Advances in neural information processing systems, 33:6840-6851, 2020. 4 +[13] Lukas Hollein, Ang Cao, Andrew Owens, Justin Johnson, and Matthias Nießner. Text2room: Extracting textured 3d meshes from 2d text-to-image models. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 7909-7920, 2023. 2 + +[14] Fangzhou Hong, Jiaxiang Tang, Ziang Cao, Min Shi, Tong Wu, Zhaoxi Chen, Shuai Yang, Tengfei Wang, Liang Pan, Dahua Lin, et al. 3dtopia: Large text-to-3d generation model with hybrid diffusion priors. arXiv preprint arXiv:2403.02234, 2024. 2 +[15] Binbin Huang, Zehao Yu, Anpei Chen, Andreas Geiger, and Shenghua Gao. 2d gaussian splatting for geometrically accurate radiance fields. In ACM SIGGRAPH 2024 Conference Papers, pages 1-11, 2024. 2 +[16] Lutao Jiang and Lin Wang. Brightdreamer: Generic 3d gaussian generative framework for fast text-to-3d synthesis. arXiv preprint arXiv:2403.11273, 2024. 2 +[17] Justin Johnson, Alexandre Alahi, and Li Fei-Fei. Perceptual losses for real-time style transfer and super-resolution. In Computer Vision-ECCV 2016: 14th European Conference, Amsterdam, The Netherlands, October 11-14, 2016, Proceedings, Part II 14, pages 694-711. Springer, 2016. 4 +[18] Xiaoliang Ju, Zhaoyang Huang, Yijin Li, Guofeng Zhang, Yu Qiao, and Hongsheng Li. Diffindscene: Diffusion-based high-quality 3d indoor scene generation, 2023. 2 +[19] Heewoo Jun and Alex Nichol. Shap-e: Generating conditional 3d implicit functions. arXiv preprint arXiv:2305.02463, 2023. 5 +[20] Tero Karras, Miika Aittala, Timo Aila, and Samuli Laine. Elucidating the design space of diffusion-based generative models. Advances in neural information processing systems, 35:26565-26577, 2022. 4 +[21] Bernhard Kerbl, Georgios Kopanas, Thomas Leimkuhler, and George Drettakis. 3d gaussian splatting for real-time radiance field rendering. ACM Transactions on Graphics, 42 (4):1-14, 2023. 1, 2, 4 +[22] Samuli Laine, Janne Hellsten, Tero Karras, Yeongho Seol, Jaakko Lehtinen, and Timo Aila. Modular primitives for high-performance differentiable rendering. ACM Transactions on Graphics, 39(6), 2020. 4 +[23] Yushi Lan, Fangzhou Hong, Shuai Yang, Shangchen Zhou, Xuyi Meng, Bo Dai, Xingang Pan, and Chen Change Loy. Ln3diff: Scalable latent neural fields diffusion for speedy 3d generation. In European Conference on Computer Vision, pages 112-130. Springer, 2025. 2 +[24] Zhiqiu Lin, Deepak Pathak, Baiqi Li, Jiayao Li, Xide Xia, Graham Neubig, Pengchuan Zhang, and Deva Ramanan. Evaluating text-to-visual generation with image-to-text generation. arXiv preprint arXiv:2404.01291, 2024. 6 +[25] Minghua Liu, Chao Xu, Haian Jin, Linghao Chen, Mukund Varma T, Zexiang Xu, and Hao Su. One-2-3-45: Any single image to 3d mesh in 45 seconds without per-shape optimization. Advances in Neural Information Processing Systems, 36, 2024. 2 +[26] Qihao Liu, Yi Zhang, Song Bai, Adam Kortylewski, and Alan Yuille. Direct-3d: Learning direct text-to-3d generation on massive noisy 3d data. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 6881–6891, 2024. 5 +[27] Ruoshi Liu, Rundi Wu, Basile Van Hoorick, Pavel Tokmakov, Sergey Zakharov, and Carl Vondrick. Zero-1-to-3: Zero-shot one image to 3d object. In Proceedings of the + +IEEE/CVF International Conference on Computer Vision, pages 9298-9309, 2023. 2 +[28] Xiaoxiao Long, Yuan-Chen Guo, Cheng Lin, Yuan Liu, Zhiyang Dou, Lingjie Liu, Yuexin Ma, Song-Hai Zhang, Marc Habermann, Christian Theobalt, et al. Wonder3d: Single image to 3d using cross-domain diffusion. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 9970-9980, 2024. 2 +[29] Tao Lu, Mulin Yu, Linning Xu, Yuanbo Xiangli, Limin Wang, Dahua Lin, and Bo Dai. Scaffold-gs: Structured 3d gaussians for view-adaptive rendering. arXiv preprint arXiv:2312.00109, 2023. 2 +[30] Hidenobu Matsuki, Riku Murai, Paul HJ Kelly, and Andrew J Davison. Gaussian splattering slam. arXiv preprint arXiv:2312.06741, 2023. 2 +[31] Gal Metzer, Elad Richardson, Or Patashnik, Raja Giryes, and Daniel Cohen-Or. Latent-nerf for shape-guided generation of 3d shapes and textures. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 12663–12673, 2023. 2 +[32] Ben Mildenhall, Pratul P Srinivasan, Matthew Tancik, Jonathan T Barron, Ravi Ramamoorthi, and Ren Ng. Nerf: Representing scenes as neural radiance fields for view synthesis. Communications of the ACM, 65(1):99-106, 2021. 1 +[33] Jacob Munkberg, Jon Hasselgren, Tianchang Shen, Jun Gao, Wenzheng Chen, Alex Evans, Thomas Müller, and Sanja Fidler. Extracting triangular 3d models, materials, and lighting from images. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 8280-8290, 2022. 4 +[34] Alex Nichol, Heewoo Jun, Prafulla Dhariwal, Pamela Mishkin, and Mark Chen. Point-e: A system for generating 3d point clouds from complex prompts. arXiv preprint arXiv:2212.08751, 2022. 2 +[35] Ben Poole, Ajay Jain, Jonathan T Barron, and Ben Mildenhall. Dreamfusion: Text-to-3d using 2d diffusion. arXiv preprint arXiv:2209.14988, 2022. 1, 2, 5 +[36] Robin Rombach, Andreas Blattmann, Dominik Lorenz, Patrick Esser, and Björn Ommer. High-resolution image synthesis with latent diffusion models. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 10684-10695, 2022. 1 +[37] Tianchang Shen, Jun Gao, Kangxue Yin, Ming-Yu Liu, and Sanja Fidler. Deep marching tetrahedra: a hybrid representation for high-resolution 3d shape synthesis. In Advances in Neural Information Processing Systems (NeurIPS), 2021. 3 +[38] Tianchang Shen, Jacob Munkberg, Jon Hasselgren, Kangxue Yin, Zian Wang, Wenzheng Chen, Zan Gojcic, Sanja Fidler, Nicholas Sharp, and Jun Gao. Flexible isosurface extraction for gradient-based mesh optimization. ACM Trans. Graph., 42(4), 2023. 3, 4 +[39] Yichun Shi, Peng Wang, Jianglong Ye, Mai Long, Kejie Li, and Xiao Yang. Mvdream: Multi-view diffusion for 3d generation. arXiv preprint arXiv:2308.16512, 2023. 2 +[40] J Ryan Shue, Eric Ryan Chan, Ryan Po, Zachary Ankner, Jiajun Wu, and Gordon Wetzstein. 3d neural field generation + +using triplane diffusion. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 20875-20886, 2023. 1 +[41] Jiaxiang Tang, Jiawei Ren, Hang Zhou, Ziwei Liu, and Gang Zeng. Dreamgaussian: Generative gaussian splatting for efficient 3d content creation. arXiv preprint arXiv:2309.16653, 2023. 2 +[42] Zhou Wang, Alan C Bovik, Hamid R Sheikh, and Eero P Simoncelli. Image quality assessment: from error visibility to structural similarity. IEEE transactions on image processing, 13(4):600-612, 2004. 4 +[43] Zhengyi Wang, Cheng Lu, Yikai Wang, Fan Bao, Chongxuan Li, Hang Su, and Jun Zhu. Prolificdreamer: High-fidelity and diverse text-to-3d generation with variational score distillation. Advances in Neural Information Processing Systems, 36, 2024. 2 +[44] Guanjun Wu, Taoran Yi, Jiemin Fang, Lingxi Xie, Xiaopeng Zhang, Wei Wei, Wenyu Liu, Qi Tian, and Xinggang Wang. 4d gaussian splatting for real-time dynamic scene rendering. arXiv preprint arXiv:2310.08528, 2023. 2 +[45] Tong Wu, Jiarui Zhang, Xiao Fu, Yuxin Wang, Liang Pan Jiawei Ren, Wayne Wu, Lei Yang, Jiaqi Wang, Chen Qian, Dahua Lin, and Ziwei Liu. Omniobject3d: Large-vocabulary 3d object dataset for realistic perception, reconstruction and generation. In IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), 2023. 5 +[46] Zijie Wu, Yaonan Wang, Mingtao Feng, He Xie, and Ajmal Mian. Sketch and text guided diffusion model for colored point cloud generation. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 8929-8939, 2023. 2 +[47] Zhennan Wu, Yang Li, Han Yan, Taizhang Shang, Weixuan Sun, Senbo Wang, Ruikai Cui, Weizhe Liu, Hiroyuki Sato, Hongdong Li, et al. Blockfusion: Expandable 3d scene generation using latent tri-plane extrapolation. arXiv preprint arXiv:2401.17053, 2024. 1, 2 +[48] Chi Yan, Delin Qu, Dong Wang, Dan Xu, Zhigang Wang, Bin Zhao, and Xuelong Li. Gs-slam: Dense visual slam with 3d gaussian splatting. arXiv preprint arXiv:2311.11700, 2023. 2 +[49] Zeyu Yang, Hongye Yang, Zijie Pan, Xiatian Zhu, and Li Zhang. Real-time photorealistic dynamic scene representation and rendering with 4d gaussian splatting. arXiv preprint arXiv:2310.10642, 2023. 2 +[50] Taoran Yi, Jiemin Fang, Junjie Wang, Guanjun Wu, Lingxi Xie, Xiaopeng Zhang, Wenyu Liu, Qi Tian, and Xinggang Wang. Gaussiandreamer: Fast generation from text to 3d gaussians by bridging 2d and 3d diffusion models. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 6796-6807, 2024. 5, 6 +[51] Zehao Yu, Anpei Chen, Binbin Huang, Torsten Sattler, and Andreas Geiger. Mip-splatting: Alias-free 3d gaussian splattering. arXiv preprint arXiv:2311.16493, 2023. 2 +[52] Bowen Zhang, Yiji Cheng, Jiaolong Yang, Chunyu Wang, Feng Zhao, Yansong Tang, Dong Chen, and Baining Guo. Gaussiancube: Structuring gaussian splatting using optimal transport for 3d generative modeling. arXiv preprint arXiv:2403.19655, 2024. 2, 5 + +[53] Zi-Xin Zou, Zhipeng Yu, Yuan-Chen Guo, Yangguang Li, Ding Liang, Yan-Pei Cao, and Song-Hai Zhang. Triplane meets gaussian splatting: Fast and generalizable single-view 3d reconstruction with transformers. arXiv preprint arXiv:2312.09147, 2023. 2, 4 \ No newline at end of file diff --git a/2025/DirectTriGS_ Triplane-based Gaussian Splatting Field Representation for 3D Generation/images.zip b/2025/DirectTriGS_ Triplane-based Gaussian Splatting Field Representation for 3D Generation/images.zip new file mode 100644 index 0000000000000000000000000000000000000000..b03f9a099757f1f0435561173301ea3a73cdca63 --- /dev/null +++ b/2025/DirectTriGS_ Triplane-based Gaussian Splatting Field Representation for 3D Generation/images.zip @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:43b7b8c09ee7e9b69e7e40dfbb32e4e779b9ba99ae415b72f20e0d06c4850b85 +size 683907 diff --git a/2025/DirectTriGS_ Triplane-based Gaussian Splatting Field Representation for 3D Generation/layout.json b/2025/DirectTriGS_ Triplane-based Gaussian Splatting Field Representation for 3D Generation/layout.json new file mode 100644 index 0000000000000000000000000000000000000000..03da4fa158a684c6ea3983ff02016513d57fcf19 --- /dev/null +++ b/2025/DirectTriGS_ Triplane-based Gaussian Splatting Field Representation for 3D Generation/layout.json @@ -0,0 +1,8500 @@ +{ + "pdf_info": [ + { + "para_blocks": [ + { + "bbox": [ + 69, + 102, + 541, + 138 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 69, + 102, + 541, + 138 + ], + "spans": [ + { + "bbox": [ + 69, + 102, + 541, + 138 + ], + "type": "text", + "content": "DirectTriGS: Triplane-based Gaussian Splitting Field Representation for 3D Generation" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 201, + 160, + 410, + 188 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 201, + 160, + 410, + 188 + ], + "spans": [ + { + "bbox": [ + 201, + 160, + 410, + 188 + ], + "type": "text", + "content": "Xiaoliang Ju, Hongsheng Li" + }, + { + "bbox": [ + 201, + 160, + 410, + 188 + ], + "type": "inline_equation", + "content": "^{1,2}" + }, + { + "bbox": [ + 201, + 160, + 410, + 188 + ], + "type": "inline_equation", + "content": "^{1}" + }, + { + "bbox": [ + 201, + 160, + 410, + 188 + ], + "type": "text", + "content": " CUHK MMLab " + }, + { + "bbox": [ + 201, + 160, + 410, + 188 + ], + "type": "inline_equation", + "content": "^{2}" + }, + { + "bbox": [ + 201, + 160, + 410, + 188 + ], + "type": "text", + "content": " CPII under InnoHK" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 243, + 192, + 364, + 202 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 243, + 192, + 364, + 202 + ], + "spans": [ + { + "bbox": [ + 243, + 192, + 364, + 202 + ], + "type": "text", + "content": "akira@link.cuhk.edu.hk" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 151, + 231, + 200, + 243 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 151, + 231, + 200, + 243 + ], + "spans": [ + { + "bbox": [ + 151, + 231, + 200, + 243 + ], + "type": "text", + "content": "Abstract" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 55, + 257, + 297, + 533 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 257, + 297, + 533 + ], + "spans": [ + { + "bbox": [ + 55, + 257, + 297, + 533 + ], + "type": "text", + "content": "We present DirectTriGS, a novel framework designed for 3D object generation with Gaussian Splatting (GS). GS-based rendering for 3D content has gained considerable attention recently. However, there has been limited exploration in directly generating 3D Gaussians compared to traditional generative modeling approaches. The main challenge lies in the complex data structure of GS represented by discrete point clouds with multiple channels. To overcome this challenge, we propose employing the triplane representation, which allows us to represent Gaussian Splatting as an image-like continuous field. This representation effectively encodes both the geometry and texture information, enabling smooth transformation back to Gaussian point clouds and rendering into images by a TriRenderer, with only 2D supervisions. The proposed TriRenderer is fully differentiable, so that the rendering loss can supervise both texture and geometry encoding. Furthermore, the triplane representation can be compressed using a Variational Autoencoder (VAE), which can subsequently be utilized in latent diffusion to generate 3D objects. The experiments demonstrate that the proposed generation framework can produce high-quality 3D object geometry and rendering results in the text-to-3D task." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 55, + 560, + 135, + 572 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 560, + 135, + 572 + ], + "spans": [ + { + "bbox": [ + 55, + 560, + 135, + 572 + ], + "type": "text", + "content": "1. Introduction" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 55, + 582, + 296, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 582, + 296, + 715 + ], + "spans": [ + { + "bbox": [ + 55, + 582, + 296, + 715 + ], + "type": "text", + "content": "Neural rendering has grown to a focal point in rendering techniques in recent years, as it achieves more realistic rendering effects by leveraging the great expressiveness of neural network. The representatives are neural radiance field (NeRF) [32] and the newly emerged Gaussian Splatting (GS) [21]. However, when applied to the field of 3D generation, the slow rendering and training speed become a strong limitation for NeRF. While GS is more flexible with greater rendering efficiency and editability, few works have addressed the challenge of direct GS generation in 3D due to its complex data structure. In this paper, we propose a" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 313, + 232, + 555, + 291 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 232, + 555, + 291 + ], + "spans": [ + { + "bbox": [ + 313, + 232, + 555, + 291 + ], + "type": "text", + "content": "novel approach called DirectTriGS, which encodes GS as triplane representation, and introduce its corresponding renderer, TriRenderer. Subsequently, we apply latent diffusion on the Triplane representation to generate high-quality GS objects." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 313, + 293, + 556, + 581 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 293, + 556, + 581 + ], + "spans": [ + { + "bbox": [ + 313, + 293, + 556, + 581 + ], + "type": "text", + "content": "Gaussian splatting uses multi-channel point cloud of \"splats\" to describe the 3D contents. With a differentiable splats rasterizer, GS has fast rendering speed. However, the sparsity, multiple channels, and uneven distributed density of 3D GS bring great difficulty for 3D generation. We propose to encode GS contents as multi-channel triplane representations, which have been shown to possess favorable properties for representing 3D geometry or NeRF, as demonstrated in previous works Shue et al. [40], Wu et al. [47]. In our work, we leverage the triplane representation to encode both the geometry information and other GS attributes in two separate groups of channels. By training on a dataset of 3D objects, we obtain a shared TriRenderer that is capable of decoding different triplane representations to GS and then render it to image. TriRenderer is fully differentiable, enabling the use of only 2D rendering loss to supervise both the texture and geometry of 3D GS. The advantage of using Triplane to represent GS are two folds. First, it leads to high memory efficiency compared with dense voxels and is expressive enough for generating various 3D GS. Second, Triplane is more compatible with the convolution-based encoders compared with the original sparse GS point cloud, which require specifically designed networks on processing sparse point clouds." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 313, + 582, + 557, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 582, + 557, + 715 + ], + "spans": [ + { + "bbox": [ + 313, + 582, + 557, + 715 + ], + "type": "text", + "content": "We follow stable diffusion [36] to train DirectTriGS on the proposed triplane representation. Specifically, a VAE is designed to further convert the Triplanes into latent code. We use two separate decoders to decouple the decoding of geometry and GS attributes. Second, we roll out the triplane latent to an expanded multi-channel image, then exploit staged latent diffusion to do generation. Two-stage diffusion is employed to generate the geometry and the corresponding GS appearance. Finally, score distillation sampling (SDS) [35] is adopted as a optional post-processing to refine or restyle the generated 3D objects." + } + ] + } + ], + "index": 11 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "spans": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "type": "text", + "content": "CVF" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 145, + 0, + 496, + 37 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 145, + 0, + 496, + 37 + ], + "spans": [ + { + "bbox": [ + 145, + 0, + 496, + 37 + ], + "type": "text", + "content": "This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 294, + 748, + 318, + 758 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 318, + 758 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 318, + 758 + ], + "type": "text", + "content": "16229" + } + ] + } + ], + "index": 12 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 0 + }, + { + "para_blocks": [ + { + "bbox": [ + 55, + 72, + 294, + 215 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 72, + 294, + 215 + ], + "spans": [ + { + "bbox": [ + 55, + 72, + 294, + 215 + ], + "type": "text", + "content": "Our contribution can summarize as follows. 1) We propose a Triplane representation for direct 3D GS generation, which is memory efficient to derive GS point clouds with vivid rendering. 2) We design a fully differentiable TriRenderer to enable the end-to-end training of triplane representations, with only 2D supervisions. 3) We develop a Triplane-based GS generation framework DirectTriGS that incorporates a specially designed VAE, latent diffusion module, and a SDS based refiner. 4) Experiments demonstrate our method produces competitive performance in both 3D geometry and multi-view rendering quality in text-to-3D tasks." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 55, + 227, + 146, + 239 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 227, + 146, + 239 + ], + "spans": [ + { + "bbox": [ + 55, + 227, + 146, + 239 + ], + "type": "text", + "content": "2. Related Works" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 55, + 247, + 295, + 449 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 247, + 295, + 449 + ], + "spans": [ + { + "bbox": [ + 55, + 247, + 295, + 449 + ], + "type": "text", + "content": "3D Object Generation. 1) Multi-view images-to-3D. The first route generates multi-view colored 3D image and then reconstruct the 3D shape and project image to texture, or use score distillation sampling to build 3D object from images produced by 2D diffusion, such as [6, 13, 27, 35, 39, 41]. Since these approaches are actually 2D generator, it is difficult to maintain 3D consistency, which may lead to Janus problem. 2) Image-to-3D. The second route attempt to retrieve 3D geometry from a single input image, such as [25, 28, 53]. With the certain one image as input, the users may be more concerned with the reconstruction quality. 3)Text-to-3D. The third group mainly operate in 3D, which can be clustered by different 3D representation. For example, [10, 11] are able to generate high-quality 3D textured meshes using differentiable rendering; [34, 46] generate colored pointcloud; [1, 14, 23, 31] generate NeRF volumes; [18, 47] generate SDF volumes." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 55, + 450, + 295, + 498 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 450, + 295, + 498 + ], + "spans": [ + { + "bbox": [ + 55, + 450, + 295, + 498 + ], + "type": "text", + "content": "Remark. The input conditions of these three generation methods are gradually becoming abstract, corresponding to increasing difficulty of the generation task, commonly resulting in a gradual decline in the generation quality." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 55, + 498, + 295, + 594 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 498, + 295, + 594 + ], + "spans": [ + { + "bbox": [ + 55, + 498, + 295, + 594 + ], + "type": "text", + "content": "Gaussian Splatting. Gaussian Splatting [21] exploits multi-channel pointcloud of splats to describe the 3D content, which allows real-time rendering of photorealistic scenes. Recently, there emerges so many research focusing on its reconstruction quality enhancement and algorithm optimization [5, 8, 15, 29, 51]. Another related active area includes 4D dynamic GS modeling [3, 44, 49], scene editing [2, 9] and its applications like SLAM [30, 48]." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 55, + 594, + 295, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 594, + 295, + 714 + ], + "spans": [ + { + "bbox": [ + 55, + 594, + 295, + 714 + ], + "type": "text", + "content": "As for GS generation, some works exploit multi-view image based route such as [6, 41], where the image generator output image from required views to enhance the GS rendering in a recursive manner. Rare works explore the GS generation directly in 3D space, such as [53] generates a point cloud using a Transformer structure with 2D image as input, and build a mapping of image to tri-plane encoding of GS attribute; GaussianCube [52] proposes to use optimal transport to model the 3D GS for text-to-3D generation; [16] combine explicit anchor points with implicit" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 313, + 72, + 472, + 85 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 72, + 472, + 85 + ], + "spans": [ + { + "bbox": [ + 313, + 72, + 472, + 85 + ], + "type": "text", + "content": "triplane to directly generate GS points." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 313, + 94, + 481, + 108 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 94, + 481, + 108 + ], + "spans": [ + { + "bbox": [ + 313, + 94, + 481, + 108 + ], + "type": "text", + "content": "3. Revisiting Gaussian Splitting" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 313, + 114, + 553, + 236 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 114, + 553, + 236 + ], + "spans": [ + { + "bbox": [ + 313, + 114, + 553, + 236 + ], + "type": "text", + "content": "Gaussian Splatting [21] uses a set of Gaussian points to describe a 3D object. The Gaussian points are defined by a full 3D covariance matrix " + }, + { + "bbox": [ + 313, + 114, + 553, + 236 + ], + "type": "inline_equation", + "content": "\\pmb{\\Sigma}" + }, + { + "bbox": [ + 313, + 114, + 553, + 236 + ], + "type": "text", + "content": " in a world space as " + }, + { + "bbox": [ + 313, + 114, + 553, + 236 + ], + "type": "inline_equation", + "content": "G(\\mathbf{x}) = e^{-\\frac{1}{2}\\mathbf{x}^T\\pmb{\\Sigma}^{-1}\\mathbf{x}}" + }, + { + "bbox": [ + 313, + 114, + 553, + 236 + ], + "type": "text", + "content": ", centered at point (mean) " + }, + { + "bbox": [ + 313, + 114, + 553, + 236 + ], + "type": "inline_equation", + "content": "\\mathbf{p} \\in \\mathbb{R}^3" + }, + { + "bbox": [ + 313, + 114, + 553, + 236 + ], + "type": "text", + "content": ". Each point is with a opacity scalar " + }, + { + "bbox": [ + 313, + 114, + 553, + 236 + ], + "type": "inline_equation", + "content": "\\alpha \\in [0,1]" + }, + { + "bbox": [ + 313, + 114, + 553, + 236 + ], + "type": "text", + "content": " for blending and a series of Spherical Harmonics (SH) coefficients to correctly capture the view-dependent appearance of the scene. The number of SH coefficients are " + }, + { + "bbox": [ + 313, + 114, + 553, + 236 + ], + "type": "inline_equation", + "content": "3 \\times (n + 1)^2" + }, + { + "bbox": [ + 313, + 114, + 553, + 236 + ], + "type": "text", + "content": ", where " + }, + { + "bbox": [ + 313, + 114, + 553, + 236 + ], + "type": "inline_equation", + "content": "n" + }, + { + "bbox": [ + 313, + 114, + 553, + 236 + ], + "type": "text", + "content": " denotes the SH order, and higher " + }, + { + "bbox": [ + 313, + 114, + 553, + 236 + ], + "type": "inline_equation", + "content": "n" + }, + { + "bbox": [ + 313, + 114, + 553, + 236 + ], + "type": "text", + "content": " corresponds to more accurate view-dependent appearance." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 313, + 236, + 554, + 355 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 236, + 554, + 355 + ], + "spans": [ + { + "bbox": [ + 313, + 236, + 554, + 355 + ], + "type": "text", + "content": "During rendering, the Gaussians are project to 2D given a viewing transformation " + }, + { + "bbox": [ + 313, + 236, + 554, + 355 + ], + "type": "inline_equation", + "content": "\\mathbf{W}" + }, + { + "bbox": [ + 313, + 236, + 554, + 355 + ], + "type": "text", + "content": ". The covariance is transformed as " + }, + { + "bbox": [ + 313, + 236, + 554, + 355 + ], + "type": "inline_equation", + "content": "\\boldsymbol{\\Sigma}^{\\prime} = \\mathbf{J}\\mathbf{W}\\boldsymbol{\\Sigma}\\mathbf{W}^{T}\\mathbf{J}^{T}" + }, + { + "bbox": [ + 313, + 236, + 554, + 355 + ], + "type": "text", + "content": ", where " + }, + { + "bbox": [ + 313, + 236, + 554, + 355 + ], + "type": "inline_equation", + "content": "\\mathbf{J}" + }, + { + "bbox": [ + 313, + 236, + 554, + 355 + ], + "type": "text", + "content": " is the Jacobian of the affine approximation of the projective transformation. To ensure " + }, + { + "bbox": [ + 313, + 236, + 554, + 355 + ], + "type": "inline_equation", + "content": "\\boldsymbol{\\Sigma}" + }, + { + "bbox": [ + 313, + 236, + 554, + 355 + ], + "type": "text", + "content": " to be semi-positive in the whole training process, " + }, + { + "bbox": [ + 313, + 236, + 554, + 355 + ], + "type": "inline_equation", + "content": "\\boldsymbol{\\Sigma}" + }, + { + "bbox": [ + 313, + 236, + 554, + 355 + ], + "type": "text", + "content": " can be represented by a Cholesky decomposition as " + }, + { + "bbox": [ + 313, + 236, + 554, + 355 + ], + "type": "inline_equation", + "content": "\\boldsymbol{\\Sigma} = \\mathbf{R}\\mathbf{S}\\mathbf{S}^{T}\\mathbf{R}^{T}" + }, + { + "bbox": [ + 313, + 236, + 554, + 355 + ], + "type": "text", + "content": ", represented by a tuple " + }, + { + "bbox": [ + 313, + 236, + 554, + 355 + ], + "type": "inline_equation", + "content": "s = (s_{1}, s_{2}, s_{3})" + }, + { + "bbox": [ + 313, + 236, + 554, + 355 + ], + "type": "text", + "content": " for scaling and a unnormalized quaternion " + }, + { + "bbox": [ + 313, + 236, + 554, + 355 + ], + "type": "inline_equation", + "content": "\\mathbf{q} \\in \\mathbb{R}^{4}" + }, + { + "bbox": [ + 313, + 236, + 554, + 355 + ], + "type": "text", + "content": " for rotation. In summary, each Gaussian point are a union set of position " + }, + { + "bbox": [ + 313, + 236, + 554, + 355 + ], + "type": "inline_equation", + "content": "\\mathbf{p}" + }, + { + "bbox": [ + 313, + 236, + 554, + 355 + ], + "type": "text", + "content": " and its GS attributes:" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 390, + 363, + 553, + 376 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 390, + 363, + 553, + 376 + ], + "spans": [ + { + "bbox": [ + 390, + 363, + 553, + 376 + ], + "type": "interline_equation", + "content": "\\mathcal {G} _ {\\mathbf {p}} := \\{s, \\mathbf {q}, \\mathrm {S H}, \\alpha \\}. \\tag {1}", + "image_path": "e197883ddbd6e9dab2b87e5cf0db41622e31205d580fccf2cbb2505cacce3882.jpg" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 313, + 384, + 553, + 408 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 384, + 553, + 408 + ], + "spans": [ + { + "bbox": [ + 313, + 384, + 553, + 408 + ], + "type": "text", + "content": "GS pointcloud can be efficiently rendered by a rasterization-based splatting renderer [21], which is fully differentiable." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 313, + 418, + 397, + 432 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 418, + 397, + 432 + ], + "spans": [ + { + "bbox": [ + 313, + 418, + 397, + 432 + ], + "type": "text", + "content": "4. Methodology" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 313, + 437, + 554, + 557 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 437, + 554, + 557 + ], + "spans": [ + { + "bbox": [ + 313, + 437, + 554, + 557 + ], + "type": "text", + "content": "Our motivation is to directly generate 3D GS with fast speed by utilizing the proposed triplane-based GS representation. While 2D-lifting methods like SDS-based approaches yield intricate results, they are time-intensive and susceptible to Janus problem. Direct generation, notably with fast sampling, can complete 3D generation in under a minute, contrasting with SDS methods that take over 10 minutes or longer, such as [4][43], posing challenges for users. Therefore, direct 3D generation is more practical, with 2D-lifting reserved for enhancing texture details when needed." + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 313, + 558, + 555, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 558, + 555, + 713 + ], + "spans": [ + { + "bbox": [ + 313, + 558, + 555, + 713 + ], + "type": "text", + "content": "The procedures of our generation framework can be divide into 4 stages as shown in Fig.1 (a). Firstly, we encode the 3D objects into triplanes using the TriRenderer. The triplane encoding and TriRenderer training can be concurrently executed with only the multi-view RGBA image and camera poses as supervision. Secondly, a VAE is trained to compress the triplanes to latent space, enabling effective capture of high-level information for subsequent modules. Thirdly, a diffusion model is trained on the triplane latent code. During inference, the diffusion model produces the latent code, which can be decoded back to its corresponding triplane by the VAE decoder. Subsequently, the TriRenderer decodes the triplane into the standard GS point" + } + ] + } + ], + "index": 14 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "text", + "content": "16230" + } + ] + } + ], + "index": 15 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 1 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 83, + 71, + 188, + 113 + ], + "blocks": [ + { + "bbox": [ + 83, + 71, + 188, + 113 + ], + "lines": [ + { + "bbox": [ + 83, + 71, + 188, + 113 + ], + "spans": [ + { + "bbox": [ + 83, + 71, + 188, + 113 + ], + "type": "image", + "image_path": "0f6b3129f9dafeb1ea97a8c07c5e1e6f91700926f675d7de8e63579c4810257d.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 186, + 118, + 403, + 129 + ], + "lines": [ + { + "bbox": [ + 186, + 118, + 403, + 129 + ], + "spans": [ + { + "bbox": [ + 186, + 118, + 403, + 129 + ], + "type": "text", + "content": "(a) The overall procedure of our 3D content generation framework." + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "type": "image", + "bbox": [ + 197, + 72, + 301, + 113 + ], + "blocks": [ + { + "bbox": [ + 197, + 72, + 301, + 113 + ], + "lines": [ + { + "bbox": [ + 197, + 72, + 301, + 113 + ], + "spans": [ + { + "bbox": [ + 197, + 72, + 301, + 113 + ], + "type": "image", + "image_path": "e89ee7f07caa7332f781434de1779f2cd6006aa281d39e6a1c03fab792483ba9.jpg" + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_body" + } + ], + "index": 1 + }, + { + "type": "image", + "bbox": [ + 309, + 72, + 414, + 113 + ], + "blocks": [ + { + "bbox": [ + 309, + 72, + 414, + 113 + ], + "lines": [ + { + "bbox": [ + 309, + 72, + 414, + 113 + ], + "spans": [ + { + "bbox": [ + 309, + 72, + 414, + 113 + ], + "type": "image", + "image_path": "ac0245851813463f977695cdcb7376eb06ba9d71159ac02010405b1d656ef03f.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + } + ], + "index": 2 + }, + { + "type": "image", + "bbox": [ + 422, + 72, + 528, + 113 + ], + "blocks": [ + { + "bbox": [ + 422, + 72, + 528, + 113 + ], + "lines": [ + { + "bbox": [ + 422, + 72, + 528, + 113 + ], + "spans": [ + { + "bbox": [ + 422, + 72, + 528, + 113 + ], + "type": "image", + "image_path": "698df0cca85d4e2868585e5d2158dea4c7d18ef59fe1938a7f12036db508cf97.jpg" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_body" + } + ], + "index": 3 + }, + { + "type": "image", + "bbox": [ + 82, + 136, + 528, + 273 + ], + "blocks": [ + { + "bbox": [ + 82, + 136, + 528, + 273 + ], + "lines": [ + { + "bbox": [ + 82, + 136, + 528, + 273 + ], + "spans": [ + { + "bbox": [ + 82, + 136, + 528, + 273 + ], + "type": "image", + "image_path": "a8f4465db2ff41f986e7112ab90158b5fd05a67645196976b27a8d94f6dc0e1c.jpg" + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 160, + 276, + 457, + 285 + ], + "lines": [ + { + "bbox": [ + 160, + 276, + 457, + 285 + ], + "spans": [ + { + "bbox": [ + 160, + 276, + 457, + 285 + ], + "type": "text", + "content": "(b) The forward pass of the TriRenderer pipeline: rendering a Triolane to multi-view images." + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "image_footnote" + }, + { + "bbox": [ + 55, + 293, + 555, + 316 + ], + "lines": [ + { + "bbox": [ + 55, + 293, + 555, + 316 + ], + "spans": [ + { + "bbox": [ + 55, + 293, + 555, + 316 + ], + "type": "text", + "content": "Figure 1. The overview of the proposed 3D GS generation framework and the core component TriRenderer. We use triplane to represent 3D objects, and design a TriRenderer to decode triplane to GS pointcloud and render multi-view images." + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "image_caption" + } + ], + "index": 5 + }, + { + "bbox": [ + 55, + 337, + 296, + 396 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 337, + 296, + 396 + ], + "spans": [ + { + "bbox": [ + 55, + 337, + 296, + 396 + ], + "type": "text", + "content": "cloud. Finally, an SDS-based refiner can be utilized as an optional post-processing step to enhance or restyle the generated 3D object. Leveraging the pretrained 2D diffusion model within the SDS, the generated GS can be further improved to achieve a more detailed appearance." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 55, + 405, + 192, + 418 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 405, + 192, + 418 + ], + "spans": [ + { + "bbox": [ + 55, + 405, + 192, + 418 + ], + "type": "text", + "content": "4.1. Triplane-based GS Field" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 55, + 423, + 296, + 651 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 423, + 296, + 651 + ], + "spans": [ + { + "bbox": [ + 55, + 423, + 296, + 651 + ], + "type": "text", + "content": "We employ triplanes to convert the discrete multi-channel GS points into a continuous field, tailored for more efficient encoding in the subsequent generative model. Triplanes are 3 planes formed by every 2 axes of " + }, + { + "bbox": [ + 55, + 423, + 296, + 651 + ], + "type": "inline_equation", + "content": "x,y,z" + }, + { + "bbox": [ + 55, + 423, + 296, + 651 + ], + "type": "text", + "content": " axes in 3D space, where every 3D point " + }, + { + "bbox": [ + 55, + 423, + 296, + 651 + ], + "type": "inline_equation", + "content": "\\mathbf{p}" + }, + { + "bbox": [ + 55, + 423, + 296, + 651 + ], + "type": "text", + "content": " can query these planes to get corresponding features " + }, + { + "bbox": [ + 55, + 423, + 296, + 651 + ], + "type": "inline_equation", + "content": "F_{\\mathbf{p}} = (F_{xy},F_{xz},F_{yz})" + }, + { + "bbox": [ + 55, + 423, + 296, + 651 + ], + "type": "text", + "content": " using orthogonal projection. In real implementation, triplane is a tensor with the size of " + }, + { + "bbox": [ + 55, + 423, + 296, + 651 + ], + "type": "inline_equation", + "content": "3\\times H\\times H\\times C" + }, + { + "bbox": [ + 55, + 423, + 296, + 651 + ], + "type": "text", + "content": ", where " + }, + { + "bbox": [ + 55, + 423, + 296, + 651 + ], + "type": "inline_equation", + "content": "H" + }, + { + "bbox": [ + 55, + 423, + 296, + 651 + ], + "type": "text", + "content": " denote the resolution along every axis and " + }, + { + "bbox": [ + 55, + 423, + 296, + 651 + ], + "type": "inline_equation", + "content": "C" + }, + { + "bbox": [ + 55, + 423, + 296, + 651 + ], + "type": "text", + "content": " is the feature channel. The query of point with arbitrary continuous coordinate should be a bi-linear interpolation in the triplane grids. Considering GS is a set of points with multiple channels of attributes, it is natural to use triplane for GS encoding. If the triplane feature and GS attributes of a point can be transformed to each other, the sparse GS pointcloud of every object can be represented as continuous triplane GS field for further encoding. We use separate channels to encode 3D geometry and other GS attributes as described in Eq. 1, which is an experimental setting for better convergence in the training." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 55, + 659, + 137, + 671 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 659, + 137, + 671 + ], + "spans": [ + { + "bbox": [ + 55, + 659, + 137, + 671 + ], + "type": "text", + "content": "4.2. TriRenderer" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 55, + 677, + 296, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 677, + 296, + 713 + ], + "spans": [ + { + "bbox": [ + 55, + 677, + 296, + 713 + ], + "type": "text", + "content": "The TriRenderer serves as the crucial component for converting a 3D object into a triplane-based GS field. It acts as a fully differentiable bridge connecting the GS field with" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 313, + 337, + 553, + 361 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 337, + 553, + 361 + ], + "spans": [ + { + "bbox": [ + 313, + 337, + 553, + 361 + ], + "type": "text", + "content": "rendering, enabling the optimization from 2D images to geometry and appearance encoding." + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 313, + 362, + 555, + 518 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 362, + 555, + 518 + ], + "spans": [ + { + "bbox": [ + 313, + 362, + 555, + 518 + ], + "type": "text", + "content": "As depicted in Fig. 1 (b), the TriRenderer comprises a geometry branch and a GS attribute branch, each equipped with independent decoders to decode the triplane. The geometry branch is responsible for retrieve the geometry as triangular mesh from triplane, with a surface sampler for GS pointcloud sampling. Then GS attribute branch uses the obtained pointcloud to query the triplane to obtain the GS attribute corresponding to each point. In this way, the GS pointcloud in the original format is retrieved, and can be rendered using the original GS renderer. It is worth mentioning that all triplanes of different object share a common TriRenderer, which ensures the features on different triplanes subject to a similar distribution." + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 313, + 520, + 554, + 616 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 520, + 554, + 616 + ], + "spans": [ + { + "bbox": [ + 313, + 520, + 554, + 616 + ], + "type": "text", + "content": "Geometry Branch. We use signed distance function (SDF) to represent the geometry, and allow every vertex deform as [37], so that the geometry branch decodes " + }, + { + "bbox": [ + 313, + 520, + 554, + 616 + ], + "type": "inline_equation", + "content": "F_{p}" + }, + { + "bbox": [ + 313, + 520, + 554, + 616 + ], + "type": "text", + "content": " to a SDF volume and its vertex deformation. By query all grid coordinates of the designated resolution " + }, + { + "bbox": [ + 313, + 520, + 554, + 616 + ], + "type": "inline_equation", + "content": "L \\times L \\times L" + }, + { + "bbox": [ + 313, + 520, + 554, + 616 + ], + "type": "text", + "content": ", the deformable SDF volume is reconstructed. Then we exploit a differentiable Marching Cube algorithm FlexiCubes [38] to extract triangular mesh from the SDF volume." + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 313, + 617, + 556, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 617, + 556, + 714 + ], + "spans": [ + { + "bbox": [ + 313, + 617, + 556, + 714 + ], + "type": "text", + "content": "Surface Sampling. Considering that GS pointcloud is generally gathered on the surface of objects, we randomly sample GS points on the faces of triangular mesh with barycentric coordinates. We expect that every splat is flat and with a normal consistent with its source triangular face. Assume a face is formed by vertices " + }, + { + "bbox": [ + 313, + 617, + 556, + 714 + ], + "type": "inline_equation", + "content": "< v_{a}, v_{b}, v_{c} >" + }, + { + "bbox": [ + 313, + 617, + 556, + 714 + ], + "type": "text", + "content": ", and vector " + }, + { + "bbox": [ + 313, + 617, + 556, + 714 + ], + "type": "inline_equation", + "content": "\\mathbf{v}_{1} = (v_{a} - v_{b})" + }, + { + "bbox": [ + 313, + 617, + 556, + 714 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 313, + 617, + 556, + 714 + ], + "type": "inline_equation", + "content": "\\mathbf{v}_{2} = (v_{b} - v_{c})" + }, + { + "bbox": [ + 313, + 617, + 556, + 714 + ], + "type": "text", + "content": ", the face normal can be calculated as " + }, + { + "bbox": [ + 313, + 617, + 556, + 714 + ], + "type": "inline_equation", + "content": "\\mathbf{n} = \\mathbf{v}_{1} \\times \\mathbf{v}_{2}" + }, + { + "bbox": [ + 313, + 617, + 556, + 714 + ], + "type": "text", + "content": ". Then the rotation matrix of every" + } + ] + } + ], + "index": 16 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "text", + "content": "16231" + } + ] + } + ], + "index": 17 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 2 + }, + { + "para_blocks": [ + { + "bbox": [ + 55, + 72, + 249, + 83 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 72, + 249, + 83 + ], + "spans": [ + { + "bbox": [ + 55, + 72, + 249, + 83 + ], + "type": "text", + "content": "splat drawn from this face can be formulated as" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 130, + 92, + 295, + 106 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 92, + 295, + 106 + ], + "spans": [ + { + "bbox": [ + 130, + 92, + 295, + 106 + ], + "type": "interline_equation", + "content": "\\mathbf {R} = \\left[ \\mathbf {v} _ {1}, \\mathbf {n} \\times \\mathbf {v} _ {1}, \\mathbf {n} \\right], \\tag {2}", + "image_path": "7d5bdbd456a682db203f73c09070d6b9b03d36f95ec3ebd49f6fbf06e2f3656a.jpg" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 55, + 114, + 295, + 149 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 114, + 295, + 149 + ], + "spans": [ + { + "bbox": [ + 55, + 114, + 295, + 149 + ], + "type": "text", + "content": "which can be transform to " + }, + { + "bbox": [ + 55, + 114, + 295, + 149 + ], + "type": "inline_equation", + "content": "\\mathbf{q}" + }, + { + "bbox": [ + 55, + 114, + 295, + 149 + ], + "type": "text", + "content": " in Eq.1 by standard matrix-to-quaternion algorithm. To make the splats flat, " + }, + { + "bbox": [ + 55, + 114, + 295, + 149 + ], + "type": "inline_equation", + "content": "s_3" + }, + { + "bbox": [ + 55, + 114, + 295, + 149 + ], + "type": "text", + "content": " in Eq.1 is fixed to an infinite small value." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 55, + 150, + 296, + 222 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 150, + 296, + 222 + ], + "spans": [ + { + "bbox": [ + 55, + 150, + 296, + 222 + ], + "type": "text", + "content": "To further reduce the computational load in the subsequent GS rendering process, only faces oriented toward the camera are sampled. These face indices can be obtained through fast rasterization of the mesh faces. Note this rasterization is not required to be differentiable, and we use the library Nvdiffrast [22] for implementation." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 55, + 222, + 296, + 317 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 222, + 296, + 317 + ], + "spans": [ + { + "bbox": [ + 55, + 222, + 296, + 317 + ], + "type": "text", + "content": "GS Attribute Branch. By querying the triplane GS channels with the sampled pointcloud, we can obtain the GS feature and use the GS attribute branch to transform them to the rest GS attributes. Then the GS renderer can render GS pointcloud to multi-view images. Since different GS attributes have different numerical scales and distributions, we customize individual headers to decode them respectively, which is similar to [53]." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 55, + 318, + 296, + 426 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 318, + 296, + 426 + ], + "spans": [ + { + "bbox": [ + 55, + 318, + 296, + 426 + ], + "type": "text", + "content": "Training. The training consists of 2 stages. In the first stage, a small batch of data is used to train the triplanes and the TriRenderer together. The second stage involves the whole dataset to train all triplanes with the parameters of TriRenderer frozen. In this way, we can handle large dataset by distributed training in the second stage, and the same TriRenderer can be shared by all triplanes. There is no need to pre-train the GS pointcloud using original GS training as [21]." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 55, + 426, + 296, + 474 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 426, + 296, + 474 + ], + "spans": [ + { + "bbox": [ + 55, + 426, + 296, + 474 + ], + "type": "text", + "content": "The only required supervision data is " + }, + { + "bbox": [ + 55, + 426, + 296, + 474 + ], + "type": "inline_equation", + "content": "N" + }, + { + "bbox": [ + 55, + 426, + 296, + 474 + ], + "type": "text", + "content": " multi-view images " + }, + { + "bbox": [ + 55, + 426, + 296, + 474 + ], + "type": "inline_equation", + "content": "\\{I_i\\}_{i=1}^N" + }, + { + "bbox": [ + 55, + 426, + 296, + 474 + ], + "type": "text", + "content": " with camera poses, and the training losses consist of the rendering loss " + }, + { + "bbox": [ + 55, + 426, + 296, + 474 + ], + "type": "inline_equation", + "content": "L_{\\mathrm{render}}" + }, + { + "bbox": [ + 55, + 426, + 296, + 474 + ], + "type": "text", + "content": " and the geometric regularization " + }, + { + "bbox": [ + 55, + 426, + 296, + 474 + ], + "type": "inline_equation", + "content": "L_{\\mathrm{geo}}" + }, + { + "bbox": [ + 55, + 426, + 296, + 474 + ], + "type": "text", + "content": ". The rendering loss is a weighted sum as" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 88, + 485, + 295, + 498 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 88, + 485, + 295, + 498 + ], + "spans": [ + { + "bbox": [ + 88, + 485, + 295, + 498 + ], + "type": "interline_equation", + "content": "L _ {\\text {r e n d e r}} = w _ {1} L _ {\\text {a l p h a}} + w _ {2} L _ {\\text {r g b}} + w _ {3} L _ {\\text {p i p s}}, \\tag {3}", + "image_path": "1a44aaa96cecc3494f497deb5daa01fb01f62102a936f2e2155210214bc9dd85.jpg" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 55, + 502, + 296, + 552 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 502, + 296, + 552 + ], + "spans": [ + { + "bbox": [ + 55, + 502, + 296, + 552 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 55, + 502, + 296, + 552 + ], + "type": "inline_equation", + "content": "L_{\\mathrm{alpha}}" + }, + { + "bbox": [ + 55, + 502, + 296, + 552 + ], + "type": "text", + "content": " is a " + }, + { + "bbox": [ + 55, + 502, + 296, + 552 + ], + "type": "inline_equation", + "content": "L_{1}" + }, + { + "bbox": [ + 55, + 502, + 296, + 552 + ], + "type": "text", + "content": " loss on the alpha map, namely the silhouette image loss; " + }, + { + "bbox": [ + 55, + 502, + 296, + 552 + ], + "type": "inline_equation", + "content": "L_{\\mathrm{rgb}}" + }, + { + "bbox": [ + 55, + 502, + 296, + 552 + ], + "type": "text", + "content": " is a combination of " + }, + { + "bbox": [ + 55, + 502, + 296, + 552 + ], + "type": "inline_equation", + "content": "L_{1}" + }, + { + "bbox": [ + 55, + 502, + 296, + 552 + ], + "type": "text", + "content": " pixel loss and SSIM loss [42] between rendered image " + }, + { + "bbox": [ + 55, + 502, + 296, + 552 + ], + "type": "inline_equation", + "content": "\\hat{I}_i" + }, + { + "bbox": [ + 55, + 502, + 296, + 552 + ], + "type": "text", + "content": " and ground truth " + }, + { + "bbox": [ + 55, + 502, + 296, + 552 + ], + "type": "inline_equation", + "content": "I_i" + }, + { + "bbox": [ + 55, + 502, + 296, + 552 + ], + "type": "text", + "content": ":" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 72, + 559, + 295, + 593 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 72, + 559, + 295, + 593 + ], + "spans": [ + { + "bbox": [ + 72, + 559, + 295, + 593 + ], + "type": "interline_equation", + "content": "L _ {\\mathrm {r g b}} = \\sum_ {i = 1} ^ {N} (1 - \\beta) \\| I _ {i} - \\hat {I} _ {i} \\| _ {1} + \\beta \\operatorname {S S I M} \\left(I _ {i}, \\hat {I} _ {i}\\right), \\tag {4}", + "image_path": "79dd72220aed0d619f508baef8f4962d6379962212fdfb6f8a8003e8e528a16d.jpg" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 55, + 599, + 295, + 623 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 599, + 295, + 623 + ], + "spans": [ + { + "bbox": [ + 55, + 599, + 295, + 623 + ], + "type": "text", + "content": "and " + }, + { + "bbox": [ + 55, + 599, + 295, + 623 + ], + "type": "inline_equation", + "content": "L_{\\mathrm{pips}}" + }, + { + "bbox": [ + 55, + 599, + 295, + 623 + ], + "type": "text", + "content": " is the perceptual loss [17]. " + }, + { + "bbox": [ + 55, + 599, + 295, + 623 + ], + "type": "inline_equation", + "content": "w_{1}, w_{2}, w_{3}" + }, + { + "bbox": [ + 55, + 599, + 295, + 623 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 55, + 599, + 295, + 623 + ], + "type": "inline_equation", + "content": "\\beta" + }, + { + "bbox": [ + 55, + 599, + 295, + 623 + ], + "type": "text", + "content": " are all weighting factors determined by experiments." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 55, + 623, + 296, + 647 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 623, + 296, + 647 + ], + "spans": [ + { + "bbox": [ + 55, + 623, + 296, + 647 + ], + "type": "text", + "content": "The geometric regularization loss on the SDF volume " + }, + { + "bbox": [ + 55, + 623, + 296, + 647 + ], + "type": "inline_equation", + "content": "V" + }, + { + "bbox": [ + 55, + 623, + 296, + 647 + ], + "type": "text", + "content": " is defined as" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 63, + 656, + 295, + 670 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 63, + 656, + 295, + 670 + ], + "spans": [ + { + "bbox": [ + 63, + 656, + 295, + 670 + ], + "type": "interline_equation", + "content": "L _ {\\text {g e o}} = \\gamma_ {1} L _ {\\text {d e v}} + \\gamma_ {2} L _ {\\text {w e i g h t i n g}} + \\gamma_ {3} L _ {\\text {C E}} + \\gamma_ {4} L _ {\\text {s i g n}}, \\tag {5}", + "image_path": "7cee9e27704a4bf9b9c3b29fe25be01ebae8be02c226409d9e89bc20ce031c0c.jpg" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 55, + 677, + 296, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 677, + 296, + 714 + ], + "spans": [ + { + "bbox": [ + 55, + 677, + 296, + 714 + ], + "type": "text", + "content": "where the " + }, + { + "bbox": [ + 55, + 677, + 296, + 714 + ], + "type": "inline_equation", + "content": "L_{\\mathrm{dev}}" + }, + { + "bbox": [ + 55, + 677, + 296, + 714 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 55, + 677, + 296, + 714 + ], + "type": "inline_equation", + "content": "L_{\\mathrm{weighting}}" + }, + { + "bbox": [ + 55, + 677, + 296, + 714 + ], + "type": "text", + "content": " are defined by Flex-cubes [38] to regularize the extracted connectivity and the weighting scale of SDF vertices. To penalize the sign" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 313, + 72, + 553, + 95 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 72, + 553, + 95 + ], + "spans": [ + { + "bbox": [ + 313, + 72, + 553, + 95 + ], + "type": "text", + "content": "changes on all grid edges, we follow [33] to define " + }, + { + "bbox": [ + 313, + 72, + 553, + 95 + ], + "type": "inline_equation", + "content": "L_{\\mathrm{CE}}" + }, + { + "bbox": [ + 313, + 72, + 553, + 95 + ], + "type": "text", + "content": " as" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 356, + 95, + 553, + 121 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 356, + 95, + 553, + 121 + ], + "spans": [ + { + "bbox": [ + 356, + 95, + 553, + 121 + ], + "type": "interline_equation", + "content": "L _ {\\mathrm {C E}} = \\sum_ {(s _ {a}, s _ {b}) \\in \\epsilon} \\operatorname {C E} \\left(\\sigma \\left(s _ {a}\\right), \\operatorname {s i g n} \\left(s _ {b}\\right)\\right), \\tag {6}", + "image_path": "07612b9e9de2e45c2ea5eb7b8815b849fb4cc5264cc63620354c2704d481df7a.jpg" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 313, + 123, + 555, + 183 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 123, + 555, + 183 + ], + "spans": [ + { + "bbox": [ + 313, + 123, + 555, + 183 + ], + "type": "text", + "content": "where CE denotes cross entropy; " + }, + { + "bbox": [ + 313, + 123, + 555, + 183 + ], + "type": "inline_equation", + "content": "\\epsilon" + }, + { + "bbox": [ + 313, + 123, + 555, + 183 + ], + "type": "text", + "content": " is the set of all edges connecting vertices with different signs; and " + }, + { + "bbox": [ + 313, + 123, + 555, + 183 + ], + "type": "inline_equation", + "content": "\\sigma" + }, + { + "bbox": [ + 313, + 123, + 555, + 183 + ], + "type": "text", + "content": " is the sigmoid function. Finally, " + }, + { + "bbox": [ + 313, + 123, + 555, + 183 + ], + "type": "inline_equation", + "content": "L_{\\mathrm{sign}}" + }, + { + "bbox": [ + 313, + 123, + 555, + 183 + ], + "type": "text", + "content": " is designed to prevent the SDF volume from being trapped in an empty shape, i.e. a fully positive or fully negative SDF volume," + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 388, + 190, + 553, + 204 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 388, + 190, + 553, + 204 + ], + "spans": [ + { + "bbox": [ + 388, + 190, + 553, + 204 + ], + "type": "interline_equation", + "content": "L _ {\\text {s i g n}} = \\delta (V) M (| V |), \\tag {7}", + "image_path": "0137b04732afcea82f48d7fc628d26a335fac3e672cb0de795ec32c2b2672c31.jpg" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 313, + 209, + 553, + 234 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 209, + 553, + 234 + ], + "spans": [ + { + "bbox": [ + 313, + 209, + 553, + 234 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 313, + 209, + 553, + 234 + ], + "type": "inline_equation", + "content": "\\delta (V) = 1" + }, + { + "bbox": [ + 313, + 209, + 553, + 234 + ], + "type": "text", + "content": " if " + }, + { + "bbox": [ + 313, + 209, + 553, + 234 + ], + "type": "inline_equation", + "content": "V" + }, + { + "bbox": [ + 313, + 209, + 553, + 234 + ], + "type": "text", + "content": " is empty, otherwise " + }, + { + "bbox": [ + 313, + 209, + 553, + 234 + ], + "type": "inline_equation", + "content": "\\delta (V) = 0" + }, + { + "bbox": [ + 313, + 209, + 553, + 234 + ], + "type": "text", + "content": ". " + }, + { + "bbox": [ + 313, + 209, + 553, + 234 + ], + "type": "inline_equation", + "content": "M(|V|)" + }, + { + "bbox": [ + 313, + 209, + 553, + 234 + ], + "type": "text", + "content": " denotes the mean of the absolute value of " + }, + { + "bbox": [ + 313, + 209, + 553, + 234 + ], + "type": "inline_equation", + "content": "V" + }, + { + "bbox": [ + 313, + 209, + 553, + 234 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 313, + 239, + 402, + 252 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 239, + 402, + 252 + ], + "spans": [ + { + "bbox": [ + 313, + 239, + 402, + 252 + ], + "type": "text", + "content": "4.3. Triplane VAE" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 313, + 257, + 555, + 412 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 257, + 555, + 412 + ], + "spans": [ + { + "bbox": [ + 313, + 257, + 555, + 412 + ], + "type": "text", + "content": "We employ an UNet-like structure of VAE to compress the triplane to latent space. The training pipeline is demonstrated in Fig. 2 (a). Considering that the 3 planes of triplanes encodes features from 3 orthogonal view direction, these 3 planes should be homogeneous data, so that We reshape the triplanes of batchsize " + }, + { + "bbox": [ + 313, + 257, + 555, + 412 + ], + "type": "inline_equation", + "content": "B" + }, + { + "bbox": [ + 313, + 257, + 555, + 412 + ], + "type": "text", + "content": " to a new batch as " + }, + { + "bbox": [ + 313, + 257, + 555, + 412 + ], + "type": "inline_equation", + "content": "B \\times 3 \\times H \\times H \\times C" + }, + { + "bbox": [ + 313, + 257, + 555, + 412 + ], + "type": "text", + "content": " to " + }, + { + "bbox": [ + 313, + 257, + 555, + 412 + ], + "type": "inline_equation", + "content": "3B \\times H \\times H \\times C" + }, + { + "bbox": [ + 313, + 257, + 555, + 412 + ], + "type": "text", + "content": " for the VAE training. We use decoupled encoders and decoders for geometry and GS channels of triplanes, for experiments prove that a mixed encoding may lead to blur of the rendering results. After the triplane is reconstructed, the TriRenderer trained in Section.4.1 can be used to retrieve the GS pointclouds and render it to images." + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 313, + 412, + 554, + 483 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 412, + 554, + 483 + ], + "spans": [ + { + "bbox": [ + 313, + 412, + 554, + 483 + ], + "type": "text", + "content": "In the training process, same loss functions as Eq. 3 and Eq. 5 are used, along with an " + }, + { + "bbox": [ + 313, + 412, + 554, + 483 + ], + "type": "inline_equation", + "content": "L_{1}" + }, + { + "bbox": [ + 313, + 412, + 554, + 483 + ], + "type": "text", + "content": " loss between the input and reconstructed triplanes, denoted as " + }, + { + "bbox": [ + 313, + 412, + 554, + 483 + ], + "type": "inline_equation", + "content": "L_{\\mathrm{Tri}}" + }, + { + "bbox": [ + 313, + 412, + 554, + 483 + ], + "type": "text", + "content": ". Additionally, a Kullback-Leibler divergence loss " + }, + { + "bbox": [ + 313, + 412, + 554, + 483 + ], + "type": "inline_equation", + "content": "L_{\\mathrm{KL}}" + }, + { + "bbox": [ + 313, + 412, + 554, + 483 + ], + "type": "text", + "content": " is included to ensure that the latent variables do not deviate significantly from a normal distribution. The total losses is summarized as" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 347, + 491, + 553, + 504 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 347, + 491, + 553, + 504 + ], + "spans": [ + { + "bbox": [ + 347, + 491, + 553, + 504 + ], + "type": "interline_equation", + "content": "L _ {\\mathrm {V A E}} = L _ {\\mathrm {T r i}} + L _ {\\mathrm {G e o}} + L _ {\\mathrm {R e n d e r}} + \\gamma L _ {\\mathrm {K L}}, \\tag {8}", + "image_path": "affb193e7177be819d548de8da1eb59d6cdc18d5c43ba70cee19ef214053781b.jpg" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 313, + 510, + 423, + 523 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 510, + 423, + 523 + ], + "spans": [ + { + "bbox": [ + 313, + 510, + 423, + 523 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 313, + 510, + 423, + 523 + ], + "type": "inline_equation", + "content": "\\gamma" + }, + { + "bbox": [ + 313, + 510, + 423, + 523 + ], + "type": "text", + "content": " is a small weight." + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 313, + 529, + 537, + 541 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 529, + 537, + 541 + ], + "spans": [ + { + "bbox": [ + 313, + 529, + 537, + 541 + ], + "type": "text", + "content": "4.4. Two-stage LDM for Text-to-3D Generation" + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 313, + 546, + 554, + 641 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 546, + 554, + 641 + ], + "spans": [ + { + "bbox": [ + 313, + 546, + 554, + 641 + ], + "type": "text", + "content": "Taking into account the significant relevance of texture appearance to the underlying geometry, we proposed to utilize staged diffusion to generate geometry and GS attribute successively as shown in Fig. 2 (b). It is easy to implement it because the latent codes for geometry and other GS attributes are totally decoupled by VAE, as mentioned earlier. In this way, the geometry code can be a new condition for the second stage." + } + ] + } + ], + "index": 25 + }, + { + "bbox": [ + 313, + 642, + 554, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 642, + 554, + 714 + ], + "spans": [ + { + "bbox": [ + 313, + 642, + 554, + 714 + ], + "type": "text", + "content": "We follow DDPM [12, 20] to implement the text-conditioned LDM. To better capture the relation between different planes, we roll out the triplane latent of " + }, + { + "bbox": [ + 313, + 642, + 554, + 714 + ], + "type": "inline_equation", + "content": "B \\times 3 \\times h \\times h \\times c" + }, + { + "bbox": [ + 313, + 642, + 554, + 714 + ], + "type": "text", + "content": " to a image-like " + }, + { + "bbox": [ + 313, + 642, + 554, + 714 + ], + "type": "inline_equation", + "content": "B \\times 3h \\times h \\times c" + }, + { + "bbox": [ + 313, + 642, + 554, + 714 + ], + "type": "text", + "content": " as the input of diffusion model, and the generated result will be transformed back to the original shape for TriRenderer decoding." + } + ] + } + ], + "index": 26 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "text", + "content": "16232" + } + ] + } + ], + "index": 27 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 3 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 85, + 69, + 525, + 175 + ], + "blocks": [ + { + "bbox": [ + 85, + 69, + 525, + 175 + ], + "lines": [ + { + "bbox": [ + 85, + 69, + 525, + 175 + ], + "spans": [ + { + "bbox": [ + 85, + 69, + 525, + 175 + ], + "type": "image", + "image_path": "d4d455d7b936a86017daee34945deaf08686bcf6beb9a8f6ab8da1af198305b4.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 220, + 186, + 389, + 198 + ], + "lines": [ + { + "bbox": [ + 220, + 186, + 389, + 198 + ], + "spans": [ + { + "bbox": [ + 220, + 186, + 389, + 198 + ], + "type": "text", + "content": "Figure 2. Triplane VAE and 2-stage diffusion." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "bbox": [ + 55, + 218, + 206, + 230 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 218, + 206, + 230 + ], + "spans": [ + { + "bbox": [ + 55, + 218, + 206, + 230 + ], + "type": "text", + "content": "4.5. SDS-based Texture Booster" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 55, + 239, + 296, + 371 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 239, + 296, + 371 + ], + "spans": [ + { + "bbox": [ + 55, + 239, + 296, + 371 + ], + "type": "text", + "content": "To refine or restyle the generated GS pointcloud, we utilize a pretrained 2D diffusion model for the SDS training process [35] as shown in Fig. 3. As the generated 3D GS is constrained by the mesh reconstructed from SDF, detailed in Section 4.2, we maintain the GS splats' adherence as previously established. However, we enable the mesh vertices to shift within a limited radius to enhance the stability of the SDS process. Additionally, throughout the entire SDS procedure, there is no need for densification or pruning operations. This ensures that the number of points can be controlled to align with the number of mesh faces." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 55, + 375, + 296, + 424 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 375, + 296, + 424 + ], + "spans": [ + { + "bbox": [ + 55, + 375, + 296, + 424 + ], + "type": "text", + "content": "With the image rendered from the initial GS pointcloud, the 2D diffusion model can produce images with better 2D/3D consistence in different views, so that the optimization process can converge rapidly without Janus problem." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 55, + 446, + 132, + 459 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 446, + 132, + 459 + ], + "spans": [ + { + "bbox": [ + 55, + 446, + 132, + 459 + ], + "type": "text", + "content": "5. Experiment" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 55, + 470, + 295, + 590 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 470, + 295, + 590 + ], + "spans": [ + { + "bbox": [ + 55, + 470, + 295, + 590 + ], + "type": "text", + "content": "The experiments are conducted step-by-step according to the generation pipeline proposed as Fig.1 (a). We use OmniObject3D [45] dataset for toy experiments, and Objiverse [7] dataset for main evaluation. First, we sample 200K object to train their co顺应ing triplanes and TriNDERer. Then we train the VAE and diffusion model for 3DS GS generation. Finally, we exploit SDS to boost the texture quality for unsatisfactory generated objects. The procedure of data pre-processing and the implementation details are included in the supplementary materials." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 55, + 594, + 296, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 594, + 296, + 713 + ], + "spans": [ + { + "bbox": [ + 55, + 594, + 296, + 713 + ], + "type": "text", + "content": "The experimental results primarily consist of the following components: 1) For examining the proposed triplane modeling for GS generation, we focus on the text-to-3D task and showcase the direct GS generation outcomes qualitatively and quantitatively. These results are compared with Shap-E [19], Direct3D [26], and the most recent Gaussian-Cube [52]. 2) In comparison to 2D-lifting GS generation methods, we present the performance of SDS refinement on unsatisfying samples, with GsGEN [4] and GaussianDreamer [50] as baselines." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 313, + 218, + 531, + 232 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 218, + 531, + 232 + ], + "spans": [ + { + "bbox": [ + 313, + 218, + 531, + 232 + ], + "type": "text", + "content": "5.1. Triplane Fitting and VAE Reconstruction" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 313, + 236, + 553, + 331 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 236, + 553, + 331 + ], + "spans": [ + { + "bbox": [ + 313, + 236, + 553, + 331 + ], + "type": "text", + "content": "Triplane Fitting. We adopt a resolution of " + }, + { + "bbox": [ + 313, + 236, + 553, + 331 + ], + "type": "inline_equation", + "content": "3 \\times 128 \\times 128 \\times 16" + }, + { + "bbox": [ + 313, + 236, + 553, + 331 + ], + "type": "text", + "content": " for all triplane fitting, with a resolution of " + }, + { + "bbox": [ + 313, + 236, + 553, + 331 + ], + "type": "inline_equation", + "content": "64 \\times 64 \\times 64" + }, + { + "bbox": [ + 313, + 236, + 553, + 331 + ], + "type": "text", + "content": " for the deformable SDF grids. We randomly sample 1000 objects for the shared TriRenderer training, and train all residual objects distributively. Every triplane is initialized with Gaussian noise before training. Under such settings, every triplane takes less than 30 seconds for geometry and GS appearance reconstruction." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 313, + 331, + 553, + 355 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 331, + 553, + 355 + ], + "spans": [ + { + "bbox": [ + 313, + 331, + 553, + 355 + ], + "type": "text", + "content": "VAE Reconstruction. We use a down-sample factor of 4 to compress the triplanes to latent space." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 313, + 356, + 553, + 392 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 356, + 553, + 392 + ], + "spans": [ + { + "bbox": [ + 313, + 356, + 553, + 392 + ], + "type": "text", + "content": "We visualize the triplane and the VAE reconstruction results of several data samples for a simple check, please refer to the supplementary materials." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 313, + 397, + 534, + 410 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 397, + 534, + 410 + ], + "spans": [ + { + "bbox": [ + 313, + 397, + 534, + 410 + ], + "type": "text", + "content": "5.2. Direct Text-to-3D GS Generation by LDM" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 313, + 415, + 554, + 557 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 415, + 554, + 557 + ], + "spans": [ + { + "bbox": [ + 313, + 415, + 554, + 557 + ], + "type": "text", + "content": "In this section, we present the results of our 3D GS object generation without the SDS refinement, both quantitatively and qualitatively. Due to the current scarcity of work targeting direct 3DGS generation, we selected two state-of-the-art text-to-3D works based on other 3D representations(NeRF) Shap-E [19] and Direct3D [26] for comparison. For fairness, these selected methods also perform direct generation in the 3D domain without any 2D prior knowledge or refinement based on 2D diffusion. Additionally, for the limited space, we provide more generation results and comparison with other methods such as GaussianCube [52] in the supplementary materials." + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 313, + 558, + 554, + 665 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 558, + 554, + 665 + ], + "spans": [ + { + "bbox": [ + 313, + 558, + 554, + 665 + ], + "type": "text", + "content": "Qualitative Results. The generation results from different methods are listed in Fig. 4. Every generated sample is rendered from different views accompanied by the provided text prompt captioned beneath the images. Our method showcase enhanced proficiency in both geometry and rendering quality, resulting in sharper and clearer outputs, which can be further verified in the subsequent quantitative evaluation. For more generated samples, please check the supplementary materials." + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 313, + 666, + 553, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 666, + 553, + 714 + ], + "spans": [ + { + "bbox": [ + 313, + 666, + 553, + 714 + ], + "type": "text", + "content": "Quantitative Results. We use CLIP score to evaluate the text-to-3d consistency, and an user study is conducted to evaluate the generation results from various aspect such as geometry, texture, realistic rendering and the consistency" + } + ] + } + ], + "index": 15 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "text", + "content": "16233" + } + ] + } + ], + "index": 16 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 4 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 94, + 72, + 504, + 114 + ], + "blocks": [ + { + "bbox": [ + 94, + 72, + 504, + 114 + ], + "lines": [ + { + "bbox": [ + 94, + 72, + 504, + 114 + ], + "spans": [ + { + "bbox": [ + 94, + 72, + 504, + 114 + ], + "type": "image", + "image_path": "1c5fae7808bcece9143adcb3cb11fbe33f733882894530679f672ec3c79ce3e9.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 171, + 122, + 437, + 135 + ], + "lines": [ + { + "bbox": [ + 171, + 122, + 437, + 135 + ], + "spans": [ + { + "bbox": [ + 171, + 122, + 437, + 135 + ], + "type": "text", + "content": "Figure 3. Score distillation process(SDS) for optional texture refinement." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "type": "table", + "bbox": [ + 58, + 147, + 293, + 194 + ], + "blocks": [ + { + "bbox": [ + 58, + 147, + 293, + 194 + ], + "lines": [ + { + "bbox": [ + 58, + 147, + 293, + 194 + ], + "spans": [ + { + "bbox": [ + 58, + 147, + 293, + 194 + ], + "type": "table", + "html": "
Geometry↑Texture↑Realism↑Consistency↑Overall↑
Shap-E3.0803.0192.9473.1213.042
Direct3D3.2223.1503.1183.2423.183
Ours3.4563.3833.3323.5203.423
", + "image_path": "1593ef762653557dfd10d55a85de4bd7e62431342865575229bc4225a6f20f4a.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "table_body" + } + ], + "index": 2 + }, + { + "type": "table", + "bbox": [ + 65, + 241, + 287, + 289 + ], + "blocks": [ + { + "bbox": [ + 55, + 201, + 295, + 224 + ], + "lines": [ + { + "bbox": [ + 55, + 201, + 295, + 224 + ], + "spans": [ + { + "bbox": [ + 55, + 201, + 295, + 224 + ], + "type": "text", + "content": "Table 1. User study on generated 3D objects (Our results are raw outputs from LDM without SDS refinement)." + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 65, + 241, + 287, + 289 + ], + "lines": [ + { + "bbox": [ + 65, + 241, + 287, + 289 + ], + "spans": [ + { + "bbox": [ + 65, + 241, + 287, + 289 + ], + "type": "table", + "html": "
Shap-EDirect3DOurs
openai:ViT-L-140.23980.21520.2456
openai:ViT-L-14-3360.24260.22200.2462
", + "image_path": "52b2bbc7f93175f53ef3335d446e33c6d9841e53528f72eb0feaaeb47a7e98f0.jpg" + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "table_body" + } + ], + "index": 4 + }, + { + "type": "table", + "bbox": [ + 64, + 348, + 288, + 392 + ], + "blocks": [ + { + "bbox": [ + 55, + 297, + 296, + 331 + ], + "lines": [ + { + "bbox": [ + 55, + 297, + 296, + 331 + ], + "spans": [ + { + "bbox": [ + 55, + 297, + 296, + 331 + ], + "type": "text", + "content": "Table 2. CLIP score(↑) for evaluation of similarity between rendered images and given text prompt((Our results are raw outputs from LDM without SDS refinement))." + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 64, + 348, + 288, + 392 + ], + "lines": [ + { + "bbox": [ + 64, + 348, + 288, + 392 + ], + "spans": [ + { + "bbox": [ + 64, + 348, + 288, + 392 + ], + "type": "table", + "html": "
More Realistic (%)More Detailed (%)Overall (%)
GaussianDreamer35.050.042.5
GSGEN17.55.011.3
Ours47.545.046.3
", + "image_path": "43dcfffad70634b65261c29f50cd9cba3bcace2e9f15b04b62c0598dbac6ce5c.jpg" + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "table_body" + } + ], + "index": 6 + }, + { + "bbox": [ + 55, + 420, + 296, + 515 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 420, + 296, + 515 + ], + "spans": [ + { + "bbox": [ + 55, + 420, + 296, + 515 + ], + "type": "text", + "content": "with given prompt. 49 users participated in the user study to score the over 50 3D samples from 1 to 5 points, and the average results are shown in Tab. 1. As for the CLIP score, the open-source repository t2vmetrics [24] is used to calculate the CLIP score on two versions of ViT models, and the results are demonstrated in Tab. 2. Both the CLIP score and the user study indicate that the proposed method produces better performance." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 55, + 533, + 239, + 547 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 533, + 239, + 547 + ], + "spans": [ + { + "bbox": [ + 55, + 533, + 239, + 547 + ], + "type": "text", + "content": "5.3. Optional Texture Boosting by SDS" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 55, + 554, + 295, + 649 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 554, + 295, + 649 + ], + "spans": [ + { + "bbox": [ + 55, + 554, + 295, + 649 + ], + "type": "text", + "content": "To compare with other SDS-based 3D GS generation methods, we implement a version of SDS for our mesh-binding GS representation as described in Section 4.5. GsGEN [4] and GaussianDreamer [50] are selected as our baselines. The generated samples, as illustrated in Fig. 5, highlight that with the SDS refiner, our DirectTriGS can deliver competitive outcomes comparable to the state-of-the-art 2D-lifting methods." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 55, + 654, + 296, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 654, + 296, + 715 + ], + "spans": [ + { + "bbox": [ + 55, + 654, + 296, + 715 + ], + "type": "text", + "content": "Furthermore, we conducted a user study for quantitative evaluation. 40 users were tasked with ranking 15 samples from different methods based on two criteria: realism and level of detail. The results indicate that our approach marginally outperforms the other baseline methods." + } + ] + } + ], + "index": 11 + }, + { + "type": "table", + "bbox": [ + 317, + 147, + 553, + 172 + ], + "blocks": [ + { + "bbox": [ + 71, + 400, + 279, + 412 + ], + "lines": [ + { + "bbox": [ + 71, + 400, + 279, + 412 + ], + "spans": [ + { + "bbox": [ + 71, + 400, + 279, + 412 + ], + "type": "text", + "content": "Table 3. User study on generated 3D objects (with SDS)." + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 317, + 147, + 553, + 172 + ], + "lines": [ + { + "bbox": [ + 317, + 147, + 553, + 172 + ], + "spans": [ + { + "bbox": [ + 317, + 147, + 553, + 172 + ], + "type": "table", + "html": "
TFLOPsParas.(M)GPU Mem.(GB)Running Time (Sec.)
Generation0.0082356.652.9416.2
", + "image_path": "fd92d3b8f0105832a7be991c6cdea016b3e28d2de73d32e81c47b53873ff58b0.jpg" + } + ] + } + ], + "index": 12, + "angle": 0, + "type": "table_body" + } + ], + "index": 12 + }, + { + "bbox": [ + 315, + 179, + 551, + 190 + ], + "lines": [ + { + "bbox": [ + 315, + 179, + 551, + 190 + ], + "spans": [ + { + "bbox": [ + 315, + 179, + 551, + 190 + ], + "type": "text", + "content": "Table 4. Generation inference efficiency, with the batch size is 1." + } + ] + } + ], + "index": 13, + "angle": 0, + "type": "text" + }, + { + "bbox": [ + 313, + 197, + 408, + 210 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 197, + 408, + 210 + ], + "spans": [ + { + "bbox": [ + 313, + 197, + 408, + 210 + ], + "type": "text", + "content": "5.4. Ablation Study" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 313, + 214, + 555, + 322 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 214, + 555, + 322 + ], + "spans": [ + { + "bbox": [ + 313, + 214, + 555, + 322 + ], + "type": "text", + "content": "3D Diffusion for voxel-based Gaussian Splitting. We attempt to do GS attribute generation conditioned on given voxel occupancy via 3D diffusion model. However, even a toy experiment of over-fitting one single object failed, which may attributes to the complex multiple channels with different distributions especially the non-Euclidean ones such as quaternion. As shown in Fig.6, the diffusion model learns to generate color but fails to generate splat scaling, opacity and orientation." + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 313, + 322, + 554, + 417 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 322, + 554, + 417 + ], + "spans": [ + { + "bbox": [ + 313, + 322, + 554, + 417 + ], + "type": "text", + "content": "Triplane Diffusion without VAE. Since VAE inherently involves a reconstruction loss, we attempt to use direct diffusion on the triplane space for generation. The experiment results show that such direct diffusion may cause serious noise on decoded geometry. Randomly generated samples are visualized in Fig.7. A possible reason is that the multichannels of triplane contain considerable noisy redundancy, which is difficult for the LDM to filter." + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 313, + 418, + 554, + 455 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 418, + 554, + 455 + ], + "spans": [ + { + "bbox": [ + 313, + 418, + 554, + 455 + ], + "type": "text", + "content": "Inference Efficiency. The inference efficiency is listed in Table 4. This experiment is conducted on the platform equipped with RTX3090 GPU with 24GB memory." + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 313, + 464, + 492, + 479 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 464, + 492, + 479 + ], + "spans": [ + { + "bbox": [ + 313, + 464, + 492, + 479 + ], + "type": "text", + "content": "6. Conclusion & Acknowledgment" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 313, + 484, + 555, + 616 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 484, + 555, + 616 + ], + "spans": [ + { + "bbox": [ + 313, + 484, + 555, + 616 + ], + "type": "text", + "content": "We have presented a novel framework DirectTriGS for Gaussian Splatting Field generation. DirectTriGS mainly consists of 3 parts: 1) a light-weight triplane representation for 3D object with the format of Gaussian Splatting; 2) a fully differentiable TriRenderer which can decode triplane to original GS and render it to multi-view images; 3) the triplane VAE and staged diffusion model for the whole generation process. By utilizing our DirectTriGS, the intricate GS data can be generated directly and efficiently. 4) Additionally, we incorporate a SDS refiner to further improve the texture and details of generated objects." + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 313, + 616, + 556, + 712 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 616, + 556, + 712 + ], + "spans": [ + { + "bbox": [ + 313, + 616, + 556, + 712 + ], + "type": "text", + "content": "This project is funded in part by National Key R&D Program of China Project 2022ZD0161100, by the Centre for Perceptual and Interactive Intelligence (CPII) Ltd under the Innovation and Technology Commission (ITC)'s InnoHK, by General Research Fund of Hong Kong RGC Project 14204021, by NSFC-RGC Joint Fund Project N_CUHK498/24. Hongsheng Li is a PI of CPII under the InnoHK." + } + ] + } + ], + "index": 20 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "text", + "content": "16234" + } + ] + } + ], + "index": 21 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 5 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 67, + 72, + 208, + 161 + ], + "blocks": [ + { + "bbox": [ + 67, + 72, + 208, + 161 + ], + "lines": [ + { + "bbox": [ + 67, + 72, + 208, + 161 + ], + "spans": [ + { + "bbox": [ + 67, + 72, + 208, + 161 + ], + "type": "image", + "image_path": "e02b0b9e340b51003f4a0af94ca505b938ab38185310a1aeae72930ffb0077bf.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + } + ], + "index": 0 + }, + { + "type": "image", + "bbox": [ + 219, + 72, + 380, + 165 + ], + "blocks": [ + { + "bbox": [ + 219, + 72, + 380, + 165 + ], + "lines": [ + { + "bbox": [ + 219, + 72, + 380, + 165 + ], + "spans": [ + { + "bbox": [ + 219, + 72, + 380, + 165 + ], + "type": "image", + "image_path": "a8d52a97ba35677a7fd3927da59bc8ef5a17462e5f80d1462158dcac99b07aee.jpg" + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_body" + } + ], + "index": 1 + }, + { + "type": "image", + "bbox": [ + 392, + 72, + 539, + 161 + ], + "blocks": [ + { + "bbox": [ + 392, + 72, + 539, + 161 + ], + "lines": [ + { + "bbox": [ + 392, + 72, + 539, + 161 + ], + "spans": [ + { + "bbox": [ + 392, + 72, + 539, + 161 + ], + "type": "image", + "image_path": "c451f8c67b51d102c44bd5f968217768c2cc62129beff2a09320df123d8df510.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + } + ], + "index": 2 + }, + { + "type": "image", + "bbox": [ + 82, + 163, + 208, + 254 + ], + "blocks": [ + { + "bbox": [ + 82, + 163, + 208, + 254 + ], + "lines": [ + { + "bbox": [ + 82, + 163, + 208, + 254 + ], + "spans": [ + { + "bbox": [ + 82, + 163, + 208, + 254 + ], + "type": "image", + "image_path": "ecdf04e07c1cb4999e734e4316aba7b2d96e2e51b03be480578dcba990b57259.jpg" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_body" + } + ], + "index": 3 + }, + { + "type": "image", + "bbox": [ + 212, + 167, + 400, + 254 + ], + "blocks": [ + { + "bbox": [ + 212, + 167, + 400, + 254 + ], + "lines": [ + { + "bbox": [ + 212, + 167, + 400, + 254 + ], + "spans": [ + { + "bbox": [ + 212, + 167, + 400, + 254 + ], + "type": "image", + "image_path": "2ceb55e32e7787145e7f5b7b418411f8833e84c28a3f12e7ba41dfae5427d41f.jpg" + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_body" + } + ], + "index": 4 + }, + { + "type": "image", + "bbox": [ + 406, + 164, + 529, + 249 + ], + "blocks": [ + { + "bbox": [ + 406, + 164, + 529, + 249 + ], + "lines": [ + { + "bbox": [ + 406, + 164, + 529, + 249 + ], + "spans": [ + { + "bbox": [ + 406, + 164, + 529, + 249 + ], + "type": "image", + "image_path": "c93a0fea47463fb7388dd9b94157b2696942a779c3ac022e0750f2e3979a18f9.jpg" + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "image_body" + } + ], + "index": 5 + }, + { + "type": "image", + "bbox": [ + 77, + 264, + 212, + 329 + ], + "blocks": [ + { + "bbox": [ + 77, + 264, + 212, + 329 + ], + "lines": [ + { + "bbox": [ + 77, + 264, + 212, + 329 + ], + "spans": [ + { + "bbox": [ + 77, + 264, + 212, + 329 + ], + "type": "image", + "image_path": "939ef3c413b258a30bc7602ce90f5916e697e334b760bbc40b742c461fa27860.jpg" + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "image_body" + } + ], + "index": 6 + }, + { + "type": "image", + "bbox": [ + 220, + 255, + 394, + 334 + ], + "blocks": [ + { + "bbox": [ + 220, + 255, + 394, + 334 + ], + "lines": [ + { + "bbox": [ + 220, + 255, + 394, + 334 + ], + "spans": [ + { + "bbox": [ + 220, + 255, + 394, + 334 + ], + "type": "image", + "image_path": "0db8dc473bcd8b9a0ec8866a96e32cf4eb5c73d79cb8f8e5cc1486d0b098f5af.jpg" + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 249, + 338, + 360, + 348 + ], + "lines": [ + { + "bbox": [ + 249, + 338, + 360, + 348 + ], + "spans": [ + { + "bbox": [ + 249, + 338, + 360, + 348 + ], + "type": "text", + "content": "(a) Comparison with Direct3D." + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "image_caption" + } + ], + "index": 7 + }, + { + "type": "image", + "bbox": [ + 409, + 255, + 531, + 331 + ], + "blocks": [ + { + "bbox": [ + 409, + 255, + 531, + 331 + ], + "lines": [ + { + "bbox": [ + 409, + 255, + 531, + 331 + ], + "spans": [ + { + "bbox": [ + 409, + 255, + 531, + 331 + ], + "type": "image", + "image_path": "3bbb6ee05be3c129040b8c07a2869f056326c0048e52df83b029f2edbef790fa.jpg" + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "image_body" + } + ], + "index": 8 + }, + { + "type": "image", + "bbox": [ + 67, + 366, + 217, + 462 + ], + "blocks": [ + { + "bbox": [ + 67, + 366, + 217, + 462 + ], + "lines": [ + { + "bbox": [ + 67, + 366, + 217, + 462 + ], + "spans": [ + { + "bbox": [ + 67, + 366, + 217, + 462 + ], + "type": "image", + "image_path": "7e68078254c1997822946878565641ef1c291448b2668b07fbb52580f39748ae.jpg" + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "image_body" + } + ], + "index": 10 + }, + { + "type": "image", + "bbox": [ + 227, + 367, + 374, + 459 + ], + "blocks": [ + { + "bbox": [ + 227, + 367, + 374, + 459 + ], + "lines": [ + { + "bbox": [ + 227, + 367, + 374, + 459 + ], + "spans": [ + { + "bbox": [ + 227, + 367, + 374, + 459 + ], + "type": "image", + "image_path": "9a8f5f4e995cc14f49033c609e54d995cee059d85cbd38496bf0708bf7e03383.jpg" + } + ] + } + ], + "index": 11, + "angle": 0, + "type": "image_body" + } + ], + "index": 11 + }, + { + "type": "image", + "bbox": [ + 383, + 366, + 543, + 462 + ], + "blocks": [ + { + "bbox": [ + 383, + 366, + 543, + 462 + ], + "lines": [ + { + "bbox": [ + 383, + 366, + 543, + 462 + ], + "spans": [ + { + "bbox": [ + 383, + 366, + 543, + 462 + ], + "type": "image", + "image_path": "67d5c93c197f13c32b1ede71b73b93751838c6c0d5ccf21da19b76caa3ce5c98.jpg" + } + ] + } + ], + "index": 12, + "angle": 0, + "type": "image_body" + } + ], + "index": 12 + }, + { + "type": "image", + "bbox": [ + 70, + 465, + 214, + 641 + ], + "blocks": [ + { + "bbox": [ + 70, + 465, + 214, + 641 + ], + "lines": [ + { + "bbox": [ + 70, + 465, + 214, + 641 + ], + "spans": [ + { + "bbox": [ + 70, + 465, + 214, + 641 + ], + "type": "image", + "image_path": "ef41a3cd14b7ec0b934bd4833be2c8df5f63f92d873c950331dddcc120b9a90e.jpg" + } + ] + } + ], + "index": 13, + "angle": 0, + "type": "image_body" + } + ], + "index": 13 + }, + { + "type": "image", + "bbox": [ + 215, + 465, + 382, + 641 + ], + "blocks": [ + { + "bbox": [ + 215, + 465, + 382, + 641 + ], + "lines": [ + { + "bbox": [ + 215, + 465, + 382, + 641 + ], + "spans": [ + { + "bbox": [ + 215, + 465, + 382, + 641 + ], + "type": "image", + "image_path": "1c07367dad90481bf51654c2997d2915314471b3926331b615d5c02e577bc444.jpg" + } + ] + } + ], + "index": 14, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 249, + 649, + 355, + 660 + ], + "lines": [ + { + "bbox": [ + 249, + 649, + 355, + 660 + ], + "spans": [ + { + "bbox": [ + 249, + 649, + 355, + 660 + ], + "type": "text", + "content": "(b) Comparison with Shap-E." + } + ] + } + ], + "index": 16, + "angle": 0, + "type": "image_caption" + } + ], + "index": 14 + }, + { + "type": "image", + "bbox": [ + 394, + 466, + 539, + 643 + ], + "blocks": [ + { + "bbox": [ + 394, + 466, + 539, + 643 + ], + "lines": [ + { + "bbox": [ + 394, + 466, + 539, + 643 + ], + "spans": [ + { + "bbox": [ + 394, + 466, + 539, + 643 + ], + "type": "image", + "image_path": "64b850aa9cfd764ef1cd388e358115550c22469c903a819a75a3b9afb6f2b7ee.jpg" + } + ] + } + ], + "index": 15, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 122, + 687, + 488, + 698 + ], + "lines": [ + { + "bbox": [ + 122, + 687, + 488, + 698 + ], + "spans": [ + { + "bbox": [ + 122, + 687, + 488, + 698 + ], + "type": "text", + "content": "Figure 4. Comparison of generated 3D objects (Raw results without any refinement such as SDS)." + } + ] + } + ], + "index": 17, + "angle": 0, + "type": "image_caption" + } + ], + "index": 15 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 749, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 749, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 749, + 317, + 757 + ], + "type": "text", + "content": "16235" + } + ] + } + ], + "index": 18 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 6 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 99, + 69, + 504, + 541 + ], + "blocks": [ + { + "bbox": [ + 99, + 69, + 504, + 541 + ], + "lines": [ + { + "bbox": [ + 99, + 69, + 504, + 541 + ], + "spans": [ + { + "bbox": [ + 99, + 69, + 504, + 541 + ], + "type": "image", + "image_path": "f4e11a8500518460496587dd8c168e83c9614b6ac6d4ada9d2df5d5f61050a85.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 72, + 548, + 537, + 560 + ], + "lines": [ + { + "bbox": [ + 72, + 548, + 537, + 560 + ], + "spans": [ + { + "bbox": [ + 72, + 548, + 537, + 560 + ], + "type": "text", + "content": "Figure 5. Refinement of unsatisfactory results by SDS, compared with pure SDS-based methods GsGEN and GaussianDreamer." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "type": "image", + "bbox": [ + 113, + 582, + 201, + 657 + ], + "blocks": [ + { + "bbox": [ + 113, + 582, + 201, + 657 + ], + "lines": [ + { + "bbox": [ + 113, + 582, + 201, + 657 + ], + "spans": [ + { + "bbox": [ + 113, + 582, + 201, + 657 + ], + "type": "image", + "image_path": "98d34b585d9c24f1d501d70cd8477cd25c8c9773309fff17b241ddca3c0e9f35.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 102, + 669, + 213, + 691 + ], + "lines": [ + { + "bbox": [ + 102, + 669, + 213, + 691 + ], + "spans": [ + { + "bbox": [ + 102, + 669, + 213, + 691 + ], + "type": "text", + "content": "Figure 6. Failure of GS attribute generation." + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_caption" + } + ], + "index": 2 + }, + { + "type": "image", + "bbox": [ + 223, + 583, + 507, + 655 + ], + "blocks": [ + { + "bbox": [ + 223, + 583, + 507, + 655 + ], + "lines": [ + { + "bbox": [ + 223, + 583, + 507, + 655 + ], + "spans": [ + { + "bbox": [ + 223, + 583, + 507, + 655 + ], + "type": "image", + "image_path": "95a0e5c3a1fbb9d463967d6d26657d1dad5d10e760be06d36e579c5a787388b9.jpg" + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 222, + 666, + 507, + 687 + ], + "lines": [ + { + "bbox": [ + 222, + 666, + 507, + 687 + ], + "spans": [ + { + "bbox": [ + 222, + 666, + 507, + 687 + ], + "type": "text", + "content": "Figure 7. Noisy results of direct diffusion on triplane. The condition prompt are \"train\", \"watermelon\" and \"box\"." + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "image_caption" + } + ], + "index": 4 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 749, + 318, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 749, + 318, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 749, + 318, + 757 + ], + "type": "text", + "content": "16236" + } + ] + } + ], + "index": 6 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 7 + }, + { + "para_blocks": [ + { + "bbox": [ + 56, + 71, + 115, + 83 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 71, + 115, + 83 + ], + "spans": [ + { + "bbox": [ + 56, + 71, + 115, + 83 + ], + "type": "text", + "content": "References" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 57, + 91, + 296, + 713 + ], + "type": "list", + "angle": 0, + "index": 14, + "blocks": [ + { + "bbox": [ + 61, + 91, + 296, + 145 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 61, + 91, + 296, + 145 + ], + "spans": [ + { + "bbox": [ + 61, + 91, + 296, + 145 + ], + "type": "text", + "content": "[1] Hansheng Chen, Jiatao Gu, Anpei Chen, Wei Tian, Zhuowen Tu, Lingjie Liu, and Hao Su. Single-stage diffusion nef: A unified approach to 3d generation and reconstruction. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 2416-2425, 2023. 2" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 61, + 147, + 296, + 201 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 61, + 147, + 296, + 201 + ], + "spans": [ + { + "bbox": [ + 61, + 147, + 296, + 201 + ], + "type": "text", + "content": "[2] Yiwen Chen, Zilong Chen, Chi Zhang, Feng Wang, Xiaofeng Yang, Yikai Wang, Zhongang Cai, Lei Yang, Huaping Liu, and Guosheng Lin. Gaussianeditor: Swift and controllable 3d editing with gaussian splatting. arXiv preprint arXiv:2311.14521, 2023. 2" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 62, + 202, + 295, + 245 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 202, + 295, + 245 + ], + "spans": [ + { + "bbox": [ + 62, + 202, + 295, + 245 + ], + "type": "text", + "content": "[3] Yurui Chen, Chun Gu, Junzhe Jiang, Xiatian Zhu, and Li Zhang. Periodic vibration gaussian: Dynamic urban scene reconstruction and real-time rendering. arXiv preprint arXiv:2311.18561, 2023. 2" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 62, + 247, + 296, + 290 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 247, + 296, + 290 + ], + "spans": [ + { + "bbox": [ + 62, + 247, + 296, + 290 + ], + "type": "text", + "content": "[4] Zilong Chen, Feng Wang, Yikai Wang, and Huaping Liu. Text-to-3d using gaussian splatting. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 21401-21412, 2024. 2, 5, 6" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 62, + 291, + 295, + 334 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 291, + 295, + 334 + ], + "spans": [ + { + "bbox": [ + 62, + 291, + 295, + 334 + ], + "type": "text", + "content": "[5] Kai Cheng, Xiaoxiao Long, Kaizhi Yang, Yao Yao, Wei Yin, Yuexin Ma, Wenping Wang, and Xuejin Chen. Gaussianpro: 3d gaussian splatting with progressive propagation. arXiv preprint arXiv:2402.14650, 2024. 2" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 62, + 335, + 295, + 378 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 335, + 295, + 378 + ], + "spans": [ + { + "bbox": [ + 62, + 335, + 295, + 378 + ], + "type": "text", + "content": "[6] Jaeyoung Chung, Suyoung Lee, Hyeongjin Nam, Jaerin Lee, and Kyoung Mu Lee. Luciddreamer: Domain-free generation of 3d gaussian splattering scenes. arXiv preprint arXiv:2311.13384, 2023. 2" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 62, + 380, + 296, + 446 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 380, + 296, + 446 + ], + "spans": [ + { + "bbox": [ + 62, + 380, + 296, + 446 + ], + "type": "text", + "content": "[7] Matt Deitke, Dustin Schwenk, Jordi Salvador, Luca Weihs, Oscar Michel, Eli VanderBilt, Ludwig Schmidt, Kiana Ehsani, Aniruddha Kembhavi, and Ali Farhadi. Objverse: A universe of annotated 3d objects. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 13142-13153, 2023. 5" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 62, + 447, + 295, + 490 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 447, + 295, + 490 + ], + "spans": [ + { + "bbox": [ + 62, + 447, + 295, + 490 + ], + "type": "text", + "content": "[8] Zhiwen Fan, Kevin Wang, Kairun Wen, Zehao Zhu, Dejia Xu, and Zhangyang Wang. Lightgaussian: Unbounded 3d gaussian compression with 15x reduction and " + }, + { + "bbox": [ + 62, + 447, + 295, + 490 + ], + "type": "inline_equation", + "content": "200+" + }, + { + "bbox": [ + 62, + 447, + 295, + 490 + ], + "type": "text", + "content": " fps. arXiv preprint arXiv:2311.17245, 2023. 2" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 62, + 491, + 295, + 533 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 491, + 295, + 533 + ], + "spans": [ + { + "bbox": [ + 62, + 491, + 295, + 533 + ], + "type": "text", + "content": "[9] Jiemin Fang, Junjie Wang, Xiaopeng Zhang, Lingxi Xie, and Qi Tian. Gaussianeditor: Editing 3d gaussians delicately with text instructions. arXiv preprint arXiv:2311.16037, 2023. 2" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 57, + 536, + 295, + 590 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 536, + 295, + 590 + ], + "spans": [ + { + "bbox": [ + 57, + 536, + 295, + 590 + ], + "type": "text", + "content": "[10] Jun Gao, Tianchang Shen, Zian Wang, Wenzheng Chen, Kangxue Yin, Daiqing Li, Or Litany, Zan Gojcic, and Sanja Fidler. Get3d: A generative model of high quality 3d textured shapes learned from images. Advances In Neural Information Processing Systems, 35:31841-31854, 2022. 2" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 57, + 591, + 295, + 624 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 591, + 295, + 624 + ], + "spans": [ + { + "bbox": [ + 57, + 591, + 295, + 624 + ], + "type": "text", + "content": "[11] Anchit Gupta, Wenhan Xiong, Yixin Nie, Ian Jones, and Barlas Oğuz. 3dgen: Triplane latent diffusion for textured mesh generation. arXiv preprint arXiv:2303.05371, 2023. 2" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 57, + 625, + 295, + 658 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 625, + 295, + 658 + ], + "spans": [ + { + "bbox": [ + 57, + 625, + 295, + 658 + ], + "type": "text", + "content": "[12] Jonathan Ho, Ajay Jain, and Pieter Abbeel. Denoising diffusion probabilistic models. Advances in neural information processing systems, 33:6840-6851, 2020. 4" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 57, + 658, + 296, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 658, + 296, + 713 + ], + "spans": [ + { + "bbox": [ + 57, + 658, + 296, + 713 + ], + "type": "text", + "content": "[13] Lukas Hollein, Ang Cao, Andrew Owens, Justin Johnson, and Matthias Nießner. Text2room: Extracting textured 3d meshes from 2d text-to-image models. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 7909-7920, 2023. 2" + } + ] + } + ], + "index": 13 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 316, + 73, + 554, + 713 + ], + "type": "list", + "angle": 0, + "index": 29, + "blocks": [ + { + "bbox": [ + 316, + 73, + 554, + 127 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 73, + 554, + 127 + ], + "spans": [ + { + "bbox": [ + 316, + 73, + 554, + 127 + ], + "type": "text", + "content": "[14] Fangzhou Hong, Jiaxiang Tang, Ziang Cao, Min Shi, Tong Wu, Zhaoxi Chen, Shuai Yang, Tengfei Wang, Liang Pan, Dahua Lin, et al. 3dtopia: Large text-to-3d generation model with hybrid diffusion priors. arXiv preprint arXiv:2403.02234, 2024. 2" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 317, + 129, + 554, + 173 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 129, + 554, + 173 + ], + "spans": [ + { + "bbox": [ + 317, + 129, + 554, + 173 + ], + "type": "text", + "content": "[15] Binbin Huang, Zehao Yu, Anpei Chen, Andreas Geiger, and Shenghua Gao. 2d gaussian splatting for geometrically accurate radiance fields. In ACM SIGGRAPH 2024 Conference Papers, pages 1-11, 2024. 2" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 316, + 175, + 553, + 207 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 175, + 553, + 207 + ], + "spans": [ + { + "bbox": [ + 316, + 175, + 553, + 207 + ], + "type": "text", + "content": "[16] Lutao Jiang and Lin Wang. Brightdreamer: Generic 3d gaussian generative framework for fast text-to-3d synthesis. arXiv preprint arXiv:2403.11273, 2024. 2" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 317, + 209, + 553, + 262 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 209, + 553, + 262 + ], + "spans": [ + { + "bbox": [ + 317, + 209, + 553, + 262 + ], + "type": "text", + "content": "[17] Justin Johnson, Alexandre Alahi, and Li Fei-Fei. Perceptual losses for real-time style transfer and super-resolution. In Computer Vision-ECCV 2016: 14th European Conference, Amsterdam, The Netherlands, October 11-14, 2016, Proceedings, Part II 14, pages 694-711. Springer, 2016. 4" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 316, + 264, + 553, + 297 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 264, + 553, + 297 + ], + "spans": [ + { + "bbox": [ + 316, + 264, + 553, + 297 + ], + "type": "text", + "content": "[18] Xiaoliang Ju, Zhaoyang Huang, Yijin Li, Guofeng Zhang, Yu Qiao, and Hongsheng Li. Diffindscene: Diffusion-based high-quality 3d indoor scene generation, 2023. 2" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 316, + 299, + 553, + 330 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 299, + 553, + 330 + ], + "spans": [ + { + "bbox": [ + 316, + 299, + 553, + 330 + ], + "type": "text", + "content": "[19] Heewoo Jun and Alex Nichol. Shap-e: Generating conditional 3d implicit functions. arXiv preprint arXiv:2305.02463, 2023. 5" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 316, + 332, + 553, + 375 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 332, + 553, + 375 + ], + "spans": [ + { + "bbox": [ + 316, + 332, + 553, + 375 + ], + "type": "text", + "content": "[20] Tero Karras, Miika Aittala, Timo Aila, and Samuli Laine. Elucidating the design space of diffusion-based generative models. Advances in neural information processing systems, 35:26565-26577, 2022. 4" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 316, + 377, + 553, + 421 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 377, + 553, + 421 + ], + "spans": [ + { + "bbox": [ + 316, + 377, + 553, + 421 + ], + "type": "text", + "content": "[21] Bernhard Kerbl, Georgios Kopanas, Thomas Leimkuhler, and George Drettakis. 3d gaussian splatting for real-time radiance field rendering. ACM Transactions on Graphics, 42 (4):1-14, 2023. 1, 2, 4" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 316, + 422, + 553, + 466 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 422, + 553, + 466 + ], + "spans": [ + { + "bbox": [ + 316, + 422, + 553, + 466 + ], + "type": "text", + "content": "[22] Samuli Laine, Janne Hellsten, Tero Karras, Yeongho Seol, Jaakko Lehtinen, and Timo Aila. Modular primitives for high-performance differentiable rendering. ACM Transactions on Graphics, 39(6), 2020. 4" + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 317, + 468, + 553, + 521 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 468, + 553, + 521 + ], + "spans": [ + { + "bbox": [ + 317, + 468, + 553, + 521 + ], + "type": "text", + "content": "[23] Yushi Lan, Fangzhou Hong, Shuai Yang, Shangchen Zhou, Xuyi Meng, Bo Dai, Xingang Pan, and Chen Change Loy. Ln3diff: Scalable latent neural fields diffusion for speedy 3d generation. In European Conference on Computer Vision, pages 112-130. Springer, 2025. 2" + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 316, + 523, + 553, + 567 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 523, + 553, + 567 + ], + "spans": [ + { + "bbox": [ + 316, + 523, + 553, + 567 + ], + "type": "text", + "content": "[24] Zhiqiu Lin, Deepak Pathak, Baiqi Li, Jiayao Li, Xide Xia, Graham Neubig, Pengchuan Zhang, and Deva Ramanan. Evaluating text-to-visual generation with image-to-text generation. arXiv preprint arXiv:2404.01291, 2024. 6" + } + ] + } + ], + "index": 25 + }, + { + "bbox": [ + 317, + 568, + 553, + 622 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 568, + 553, + 622 + ], + "spans": [ + { + "bbox": [ + 317, + 568, + 553, + 622 + ], + "type": "text", + "content": "[25] Minghua Liu, Chao Xu, Haian Jin, Linghao Chen, Mukund Varma T, Zexiang Xu, and Hao Su. One-2-3-45: Any single image to 3d mesh in 45 seconds without per-shape optimization. Advances in Neural Information Processing Systems, 36, 2024. 2" + } + ] + } + ], + "index": 26 + }, + { + "bbox": [ + 316, + 624, + 553, + 679 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 624, + 553, + 679 + ], + "spans": [ + { + "bbox": [ + 316, + 624, + 553, + 679 + ], + "type": "text", + "content": "[26] Qihao Liu, Yi Zhang, Song Bai, Adam Kortylewski, and Alan Yuille. Direct-3d: Learning direct text-to-3d generation on massive noisy 3d data. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 6881–6891, 2024. 5" + } + ] + } + ], + "index": 27 + }, + { + "bbox": [ + 316, + 681, + 553, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 681, + 553, + 713 + ], + "spans": [ + { + "bbox": [ + 316, + 681, + 553, + 713 + ], + "type": "text", + "content": "[27] Ruoshi Liu, Rundi Wu, Basile Van Hoorick, Pavel Tokmakov, Sergey Zakharov, and Carl Vondrick. Zero-1-to-3: Zero-shot one image to 3d object. In Proceedings of the" + } + ] + } + ], + "index": 28 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "text", + "content": "16237" + } + ] + } + ], + "index": 30 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 8 + }, + { + "para_blocks": [ + { + "bbox": [ + 56, + 72, + 295, + 713 + ], + "type": "list", + "angle": 0, + "index": 14, + "blocks": [ + { + "bbox": [ + 77, + 72, + 294, + 95 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 77, + 72, + 294, + 95 + ], + "spans": [ + { + "bbox": [ + 77, + 72, + 294, + 95 + ], + "type": "text", + "content": "IEEE/CVF International Conference on Computer Vision, pages 9298-9309, 2023. 2" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 56, + 96, + 295, + 162 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 96, + 295, + 162 + ], + "spans": [ + { + "bbox": [ + 56, + 96, + 295, + 162 + ], + "type": "text", + "content": "[28] Xiaoxiao Long, Yuan-Chen Guo, Cheng Lin, Yuan Liu, Zhiyang Dou, Lingjie Liu, Yuexin Ma, Song-Hai Zhang, Marc Habermann, Christian Theobalt, et al. Wonder3d: Single image to 3d using cross-domain diffusion. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 9970-9980, 2024. 2" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 56, + 163, + 295, + 206 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 163, + 295, + 206 + ], + "spans": [ + { + "bbox": [ + 56, + 163, + 295, + 206 + ], + "type": "text", + "content": "[29] Tao Lu, Mulin Yu, Linning Xu, Yuanbo Xiangli, Limin Wang, Dahua Lin, and Bo Dai. Scaffold-gs: Structured 3d gaussians for view-adaptive rendering. arXiv preprint arXiv:2312.00109, 2023. 2" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 56, + 209, + 294, + 240 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 209, + 294, + 240 + ], + "spans": [ + { + "bbox": [ + 56, + 209, + 294, + 240 + ], + "type": "text", + "content": "[30] Hidenobu Matsuki, Riku Murai, Paul HJ Kelly, and Andrew J Davison. Gaussian splattering slam. arXiv preprint arXiv:2312.06741, 2023. 2" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 56, + 243, + 295, + 297 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 243, + 295, + 297 + ], + "spans": [ + { + "bbox": [ + 56, + 243, + 295, + 297 + ], + "type": "text", + "content": "[31] Gal Metzer, Elad Richardson, Or Patashnik, Raja Giryes, and Daniel Cohen-Or. Latent-nerf for shape-guided generation of 3d shapes and textures. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 12663–12673, 2023. 2" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 56, + 299, + 294, + 352 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 299, + 294, + 352 + ], + "spans": [ + { + "bbox": [ + 56, + 299, + 294, + 352 + ], + "type": "text", + "content": "[32] Ben Mildenhall, Pratul P Srinivasan, Matthew Tancik, Jonathan T Barron, Ravi Ramamoorthi, and Ren Ng. Nerf: Representing scenes as neural radiance fields for view synthesis. Communications of the ACM, 65(1):99-106, 2021. 1" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 56, + 354, + 295, + 419 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 354, + 295, + 419 + ], + "spans": [ + { + "bbox": [ + 56, + 354, + 295, + 419 + ], + "type": "text", + "content": "[33] Jacob Munkberg, Jon Hasselgren, Tianchang Shen, Jun Gao, Wenzheng Chen, Alex Evans, Thomas Müller, and Sanja Fidler. Extracting triangular 3d models, materials, and lighting from images. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 8280-8290, 2022. 4" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 56, + 422, + 295, + 464 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 422, + 295, + 464 + ], + "spans": [ + { + "bbox": [ + 56, + 422, + 295, + 464 + ], + "type": "text", + "content": "[34] Alex Nichol, Heewoo Jun, Prafulla Dhariwal, Pamela Mishkin, and Mark Chen. Point-e: A system for generating 3d point clouds from complex prompts. arXiv preprint arXiv:2212.08751, 2022. 2" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 56, + 467, + 295, + 498 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 467, + 295, + 498 + ], + "spans": [ + { + "bbox": [ + 56, + 467, + 295, + 498 + ], + "type": "text", + "content": "[35] Ben Poole, Ajay Jain, Jonathan T Barron, and Ben Mildenhall. Dreamfusion: Text-to-3d using 2d diffusion. arXiv preprint arXiv:2209.14988, 2022. 1, 2, 5" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 56, + 501, + 295, + 555 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 501, + 295, + 555 + ], + "spans": [ + { + "bbox": [ + 56, + 501, + 295, + 555 + ], + "type": "text", + "content": "[36] Robin Rombach, Andreas Blattmann, Dominik Lorenz, Patrick Esser, and Björn Ommer. High-resolution image synthesis with latent diffusion models. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 10684-10695, 2022. 1" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 56, + 557, + 295, + 600 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 557, + 295, + 600 + ], + "spans": [ + { + "bbox": [ + 56, + 557, + 295, + 600 + ], + "type": "text", + "content": "[37] Tianchang Shen, Jun Gao, Kangxue Yin, Ming-Yu Liu, and Sanja Fidler. Deep marching tetrahedra: a hybrid representation for high-resolution 3d shape synthesis. In Advances in Neural Information Processing Systems (NeurIPS), 2021. 3" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 56, + 602, + 295, + 655 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 602, + 295, + 655 + ], + "spans": [ + { + "bbox": [ + 56, + 602, + 295, + 655 + ], + "type": "text", + "content": "[38] Tianchang Shen, Jacob Munkberg, Jon Hasselgren, Kangxue Yin, Zian Wang, Wenzheng Chen, Zan Gojcic, Sanja Fidler, Nicholas Sharp, and Jun Gao. Flexible isosurface extraction for gradient-based mesh optimization. ACM Trans. Graph., 42(4), 2023. 3, 4" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 56, + 658, + 295, + 689 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 658, + 295, + 689 + ], + "spans": [ + { + "bbox": [ + 56, + 658, + 295, + 689 + ], + "type": "text", + "content": "[39] Yichun Shi, Peng Wang, Jianglong Ye, Mai Long, Kejie Li, and Xiao Yang. Mvdream: Multi-view diffusion for 3d generation. arXiv preprint arXiv:2308.16512, 2023. 2" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 56, + 692, + 295, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 692, + 295, + 713 + ], + "spans": [ + { + "bbox": [ + 56, + 692, + 295, + 713 + ], + "type": "text", + "content": "[40] J Ryan Shue, Eric Ryan Chan, Ryan Po, Zachary Ankner, Jiajun Wu, and Gordon Wetzstein. 3d neural field generation" + } + ] + } + ], + "index": 13 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 316, + 73, + 553, + 712 + ], + "type": "list", + "angle": 0, + "index": 28, + "blocks": [ + { + "bbox": [ + 335, + 73, + 553, + 106 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 335, + 73, + 553, + 106 + ], + "spans": [ + { + "bbox": [ + 335, + 73, + 553, + 106 + ], + "type": "text", + "content": "using triplane diffusion. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 20875-20886, 2023. 1" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 316, + 107, + 553, + 148 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 107, + 553, + 148 + ], + "spans": [ + { + "bbox": [ + 316, + 107, + 553, + 148 + ], + "type": "text", + "content": "[41] Jiaxiang Tang, Jiawei Ren, Hang Zhou, Ziwei Liu, and Gang Zeng. Dreamgaussian: Generative gaussian splatting for efficient 3d content creation. arXiv preprint arXiv:2309.16653, 2023. 2" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 316, + 151, + 553, + 193 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 151, + 553, + 193 + ], + "spans": [ + { + "bbox": [ + 316, + 151, + 553, + 193 + ], + "type": "text", + "content": "[42] Zhou Wang, Alan C Bovik, Hamid R Sheikh, and Eero P Simoncelli. Image quality assessment: from error visibility to structural similarity. IEEE transactions on image processing, 13(4):600-612, 2004. 4" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 316, + 195, + 553, + 247 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 195, + 553, + 247 + ], + "spans": [ + { + "bbox": [ + 316, + 195, + 553, + 247 + ], + "type": "text", + "content": "[43] Zhengyi Wang, Cheng Lu, Yikai Wang, Fan Bao, Chongxuan Li, Hang Su, and Jun Zhu. Prolificdreamer: High-fidelity and diverse text-to-3d generation with variational score distillation. Advances in Neural Information Processing Systems, 36, 2024. 2" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 316, + 251, + 553, + 293 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 251, + 553, + 293 + ], + "spans": [ + { + "bbox": [ + 316, + 251, + 553, + 293 + ], + "type": "text", + "content": "[44] Guanjun Wu, Taoran Yi, Jiemin Fang, Lingxi Xie, Xiaopeng Zhang, Wei Wei, Wenyu Liu, Qi Tian, and Xinggang Wang. 4d gaussian splatting for real-time dynamic scene rendering. arXiv preprint arXiv:2310.08528, 2023. 2" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 316, + 295, + 553, + 360 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 295, + 553, + 360 + ], + "spans": [ + { + "bbox": [ + 316, + 295, + 553, + 360 + ], + "type": "text", + "content": "[45] Tong Wu, Jiarui Zhang, Xiao Fu, Yuxin Wang, Liang Pan Jiawei Ren, Wayne Wu, Lei Yang, Jiaqi Wang, Chen Qian, Dahua Lin, and Ziwei Liu. Omniobject3d: Large-vocabulary 3d object dataset for realistic perception, reconstruction and generation. In IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), 2023. 5" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 316, + 361, + 553, + 414 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 361, + 553, + 414 + ], + "spans": [ + { + "bbox": [ + 316, + 361, + 553, + 414 + ], + "type": "text", + "content": "[46] Zijie Wu, Yaonan Wang, Mingtao Feng, He Xie, and Ajmal Mian. Sketch and text guided diffusion model for colored point cloud generation. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 8929-8939, 2023. 2" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 316, + 416, + 553, + 469 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 416, + 553, + 469 + ], + "spans": [ + { + "bbox": [ + 316, + 416, + 553, + 469 + ], + "type": "text", + "content": "[47] Zhennan Wu, Yang Li, Han Yan, Taizhang Shang, Weixuan Sun, Senbo Wang, Ruikai Cui, Weizhe Liu, Hiroyuki Sato, Hongdong Li, et al. Blockfusion: Expandable 3d scene generation using latent tri-plane extrapolation. arXiv preprint arXiv:2401.17053, 2024. 1, 2" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 316, + 472, + 553, + 514 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 472, + 553, + 514 + ], + "spans": [ + { + "bbox": [ + 316, + 472, + 553, + 514 + ], + "type": "text", + "content": "[48] Chi Yan, Delin Qu, Dong Wang, Dan Xu, Zhigang Wang, Bin Zhao, and Xuelong Li. Gs-slam: Dense visual slam with 3d gaussian splatting. arXiv preprint arXiv:2311.11700, 2023. 2" + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 316, + 515, + 553, + 557 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 515, + 553, + 557 + ], + "spans": [ + { + "bbox": [ + 316, + 515, + 553, + 557 + ], + "type": "text", + "content": "[49] Zeyu Yang, Hongye Yang, Zijie Pan, Xiatian Zhu, and Li Zhang. Real-time photorealistic dynamic scene representation and rendering with 4d gaussian splatting. arXiv preprint arXiv:2310.10642, 2023. 2" + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 316, + 559, + 553, + 624 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 559, + 553, + 624 + ], + "spans": [ + { + "bbox": [ + 316, + 559, + 553, + 624 + ], + "type": "text", + "content": "[50] Taoran Yi, Jiemin Fang, Junjie Wang, Guanjun Wu, Lingxi Xie, Xiaopeng Zhang, Wenyu Liu, Qi Tian, and Xinggang Wang. Gaussiandreamer: Fast generation from text to 3d gaussians by bridging 2d and 3d diffusion models. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 6796-6807, 2024. 5, 6" + } + ] + } + ], + "index": 25 + }, + { + "bbox": [ + 316, + 625, + 553, + 658 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 625, + 553, + 658 + ], + "spans": [ + { + "bbox": [ + 316, + 625, + 553, + 658 + ], + "type": "text", + "content": "[51] Zehao Yu, Anpei Chen, Binbin Huang, Torsten Sattler, and Andreas Geiger. Mip-splatting: Alias-free 3d gaussian splattering. arXiv preprint arXiv:2311.16493, 2023. 2" + } + ] + } + ], + "index": 26 + }, + { + "bbox": [ + 316, + 659, + 553, + 712 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 659, + 553, + 712 + ], + "spans": [ + { + "bbox": [ + 316, + 659, + 553, + 712 + ], + "type": "text", + "content": "[52] Bowen Zhang, Yiji Cheng, Jiaolong Yang, Chunyu Wang, Feng Zhao, Yansong Tang, Dong Chen, and Baining Guo. Gaussiancube: Structuring gaussian splatting using optimal transport for 3d generative modeling. arXiv preprint arXiv:2403.19655, 2024. 2, 5" + } + ] + } + ], + "index": 27 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "type": "text", + "content": "16238" + } + ] + } + ], + "index": 29 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 9 + }, + { + "para_blocks": [ + { + "bbox": [ + 56, + 72, + 298, + 128 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 72, + 298, + 128 + ], + "spans": [ + { + "bbox": [ + 56, + 72, + 298, + 128 + ], + "type": "text", + "content": "[53] Zi-Xin Zou, Zhipeng Yu, Yuan-Chen Guo, Yangguang Li, Ding Liang, Yan-Pei Cao, and Song-Hai Zhang. Triplane meets gaussian splatting: Fast and generalizable single-view 3d reconstruction with transformers. arXiv preprint arXiv:2312.09147, 2023. 2, 4" + } + ] + } + ], + "index": 0 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 749, + 318, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 749, + 318, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 749, + 318, + 757 + ], + "type": "text", + "content": "16239" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 10 + } + ], + "_backend": "vlm", + "_version_name": "2.6.4" +} \ No newline at end of file diff --git a/2025/Directional Label Diffusion Model for Learning from Noisy Labels/133ce87a-4a21-4fe5-b113-05841acf1ebd_content_list.json b/2025/Directional Label Diffusion Model for Learning from Noisy Labels/133ce87a-4a21-4fe5-b113-05841acf1ebd_content_list.json new file mode 100644 index 0000000000000000000000000000000000000000..f56746f95f913e9206d5e4d41794e18ba887b140 --- /dev/null +++ b/2025/Directional Label Diffusion Model for Learning from Noisy Labels/133ce87a-4a21-4fe5-b113-05841acf1ebd_content_list.json @@ -0,0 +1,1545 @@ +[ + { + "type": "text", + "text": "Directional Label Diffusion Model for Learning from Noisy Labels", + "text_level": 1, + "bbox": [ + 163, + 130, + 833, + 152 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Senyu Hou $^{1}$ , Gaoxia Jiang $^{1}$ , Jia Zhang $^{1}$ , Shangrong Yang $^{1}$ , Husheng Guo $^{1}$ , Yaqing Guo $^{1}$ , Wenjian Wang $^{2*}$", + "bbox": [ + 200, + 179, + 795, + 215 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "$^{1}$ School of Computer and Information Technology, Shanxi University, China \n $^{2}$ Key Laboratory of Computational Intelligence and Chinese Information Processing Ministry of Education, Shanxi University, China", + "bbox": [ + 165, + 215, + 828, + 268 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "{202312407005, 202412407035}@email.sxu.edu.cn", + "bbox": [ + 299, + 271, + 697, + 287 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "{jianggaoxia, sryang, guohusheng, yqguo, wjwang}@sxu.edu.cn", + "bbox": [ + 236, + 289, + 756, + 304 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Abstract", + "text_level": 1, + "bbox": [ + 246, + 339, + 326, + 354 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "In image classification, the label quality of training data critically influences model generalization, especially for deep neural networks (DNNs). Traditionally, learning from noisy labels (LNL) can improve the generalization of DNNs through complex architectures or a series of robust techniques, but its performance improvement is limited by the discriminative paradigm. Unlike traditional ways, we resolve the LNL problems from the perspective of robust label generation, based on diffusion models within the generative paradigm. To expand the diffusion model into a robust classifier that explicitly accommodates more noise knowledge, we propose a Directional Label Diffusion (DLD) model. It disentangles the diffusion process into two paths, i.e., directional diffusion and random diffusion. Specifically, directional diffusion simulates the corruption of true labels into a directed noise distribution, prioritizing the removal of likely noise, whereas random diffusion introduces inherent randomness to support label recovery. This architecture enable DLD to gradually infer labels from an initial random state, interpretably diverging from the specified noise distribution. To adapt the model to diverse noisy environments, we design a low-cost label precorrection method that automatically supplies more accurate label information to the diffusion model, without requiring manual intervention or additional iterations. Our approach outperforms state-of-the-art methods on both simulated and real-world noisy datasets. Code is available at https://github.com/SenyuHou/DLD.", + "bbox": [ + 89, + 371, + 485, + 794 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "1. Introduction", + "text_level": 1, + "bbox": [ + 89, + 805, + 220, + 821 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Deep neural networks (DNNs) have achieved notable success across key areas of machine learning, driven by the growth of datasets and the increase in model parameters.", + "bbox": [ + 89, + 828, + 483, + 875 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "These models are primarily trained using supervised learning, which relies heavily on large volumes of labeled data [5]. However, real-world data often suffer from low-quality labels due to subjective errors like indistinguishable samples or objective errors such as incorrect label recording, making noisy labels a widespread issue [13, 55]. Studies have shown that the generalization capability of deep neural networks is adversely affected when trained with noisy labels [1, 59]. Consequently, it is crucial to develop methods for learning from noisy labels (LNL) on large-scale but low-quality datasets.", + "bbox": [ + 511, + 340, + 905, + 505 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Due to the varied distribution of label noise, historical research has classified it into class-conditional noise (CCN) and instance-dependent noise (IDN) [57]. Recent studies indicate that label noise in real datasets predominantly aligns with the IDN assumption, wherein noise is unevenly distributed and more likely in samples with similar features [2, 6, 10–12, 51]. To tackle these complex noise patterns, mainstream LNL methods employ strategies combining filtering and training approaches. Some studies demonstrate a phenomenon known as the memorization effect [1, 3, 29, 56], i.e., DNNs first learn from simple data before fitting to noisy samples. Based on this observation, sample selection strategies utilize confidence levels or prediction loss during training intervals to identify noisy samples [17, 24, 28, 58]. However, this strategy is inherently risky, as the timing of noise memorization varies across datasets and noise distributions, complicating the selection criteria [6, 39]. Consequently, mainline LNL methods can often be integrated with training strategies to enhance model robustness, employing techniques like label filtering [17, 36, 50, 58], label correction [47, 60, 64], robust loss [16, 48, 61], and semi-supervised learning [28]. Among these techniques, semi-supervised learning, by combining multiple approaches [34, 37, 49], often outperforms other LNL methods but is still constrained by the discriminative framework. Although some researches attempt to combine", + "bbox": [ + 511, + 508, + 906, + 901 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "CVF", + "bbox": [ + 106, + 2, + 181, + 42 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore.", + "bbox": [ + 236, + 0, + 810, + 46 + ], + "page_idx": 0 + }, + { + "type": "page_footnote", + "text": "*Corresponding author.", + "bbox": [ + 109, + 887, + 236, + 898 + ], + "page_idx": 0 + }, + { + "type": "page_number", + "text": "25738", + "bbox": [ + 478, + 945, + 519, + 955 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "variational autoencoders (VAE) [25] with DivideMix [28] to address label noise [15], most methods remain focused on discriminative models.", + "bbox": [ + 89, + 90, + 480, + 136 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "In this paper, we shift away from the traditional LNL perspective, proposing robust classification models using label diffusion models that focus on the generative modeling and deeply explores the generative relationship between features and labels. Under this paradigm, labels are generated by gradually refining an initial random guess. The uncertainty modeling in diffusion models inherently meets some robustness requirements, as random label generation can offset some noise interference [4, 18]. However, this type of diffusion process lacks information about noisy labels. Intrinsic uncertainty alone is insufficient to handle complex noisy environments. Inspired by the residual diffusion denoising models (RDDM) [32], we explore a novel Directional Label Diffusion (DLD) model that incorporates noisy label information into the diffusion process. In DLD, we decouple the diffusion process into two channels: directional diffusion and random diffusion (see Figure 1). The directional diffusion guides the initial label towards a more probable noise direction (e.g., a green automobile may be misclassified as a frog or other classes, as indicated by the probability distribution in the third column in Figure 1), while random diffusion retains the original diffusion mechanism. This approach injects information about which categories should not be generated into the label generation paradigm, enhancing the interpretability of the diffusion model's robust inference. Unlike conventional label diffusion, our diffusion process concludes with an endpoint that includes the potential noise distribution of the sample (i.e., $\\mathbf{y}_T = \\mathbf{y}_n + \\epsilon$ ) and employs two independent coefficient schedules to control the diffusion speed of both channels.", + "bbox": [ + 91, + 141, + 483, + 593 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "One challenge is that only noisy labels are available in practice for training, so neither $\\mathbf{y}_0$ nor $\\mathbf{y}_n$ can be obtained directly. Few researchers employed neighborhood detection techniques like KNN to re-estimate $\\mathbf{y}_0$ [4]. However, under the IDN assumption, neighborhood labels of noisy samples often contain similar errors, making naive label estimation methods unreliable and potentially causing $\\mathbf{y}_0$ to deviate from the ground truth (see Figure 3). To address these issues, we further propose a straightforward and efficient label pre-correction method that automatically provides diverse label information for diffusion model training. This framework employs feature extractors to identify the neighborhood label distribution of samples. Note that these extractors are either trained with self-supervised methods [7, 19, 65], which require few additional training iterations, or use pre-trained models without further training, thus remaining unaffected by label noise. Additionally, this pre-correction strategy allows the seamless integration of pre-trained large-scale models such as Sim-CLR [7] and vision transformer (ViT) [40] into our framework to further", + "bbox": [ + 89, + 598, + 483, + 900 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "boost the performance. The main contributions of this work are summarized as follows:", + "bbox": [ + 511, + 90, + 903, + 119 + ], + "page_idx": 1 + }, + { + "type": "list", + "sub_type": "text", + "list_items": [ + "- We propose a novel Directional Label Diffusion (DLD) model that explicitly introduces interpretable noise information through a dual diffusion architecture, enabling robust learning of label generation patterns from noisy data.", + "- We design a fast label pre-correction method that automatically provides more accurate label information for the diffusion process and richer feature conditions for the generation process. By leveraging the consistency principle, this method pre-estimates the label distribution in a single iteration without incurring additional costs.", + "- We refine the paradigm for introducing feature conditions with rigorous theoretical deduction and provide complementary training strategies to enhance robustness further.", + "- We experimentally show that our model surpasses the state-of-the-art (SOTA) methods on both simulated and real-world noisy datasets, e.g., $6\\% \\sim 21\\%$ accuracy improvement on the CIFAR-100 dataset with IDN." + ], + "bbox": [ + 513, + 121, + 903, + 376 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "2. Related Work", + "text_level": 1, + "bbox": [ + 513, + 388, + 653, + 402 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "2.1. Learning from Noisy Labels", + "text_level": 1, + "bbox": [ + 513, + 409, + 764, + 425 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Screening and training strategies are central to LNL methods, aiming to distinguish noisy from clean samples and to optimize model learning under noisy conditions. Screening strategies, like Co-teaching [17], select low-loss samples as clean, while DivideMix [28] uses a Gaussian mixture model to automatically identify clean samples during training. Confidence-based methods like CC [62] and DISC [31] classify samples with unreliable predictions as noisy, whereas neighbor consistency regularization (NCR) [22] promotes learning from samples with high neighbor consistency. Our pre-correction method is based on similarity constraints across multi-view label distributions and leverages a feature extractor to provide high-quality information to the diffusion model without requiring additional iterative training. For training, models such as SEAL [6] iteratively correct noisy labels, but this often risks worsening outcomes [23]. Robust loss functions, like mean absolute error (MAE) [16] and generalized cross entropy (GCE) [61], enhance resilience to noisy labels, while DAC [45] and DMI [54] offer advanced mutual information-based losses. Approaches like DivideMix [28] treat LNL as semi-supervised learning, supported by methods like CC [62], C2D [63], and ELR [33] to refine the process. Without complex iterations or loss design, our method utilizes a lightweight diffusion network and a simple L2 loss to quickly learn the feature-label generation relationship.", + "bbox": [ + 511, + 431, + 906, + 824 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "2.2. Label Denoising Diffusion Models", + "text_level": 1, + "bbox": [ + 513, + 833, + 808, + 849 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Diffusion models were initially designed for image generation tasks and can be understood through the lenses of score matching and dynamics. Ho et al. [21] introduced a", + "bbox": [ + 511, + 854, + 906, + 900 + ], + "page_idx": 1 + }, + { + "type": "page_number", + "text": "25739", + "bbox": [ + 478, + 944, + 517, + 955 + ], + "page_idx": 1 + }, + { + "type": "image", + "img_path": "images/c07b814ed34be5d0f57d0b65e23ec6f71e12a29dad6bf275b3b5b1a6c5210b02.jpg", + "image_caption": [ + "Figure 1. The proposed Directional Label Diffusion (DLD) model includes diagrams illustrating both the forward and reverse processes. In the forward process, diffusion is decoupled into two channels: random diffusion and directional diffusion. Random diffusion represents the transformation of $\\mathbf{y}_0$ into random Gaussian noise $\\epsilon$ , while directional diffusion describes the transformation of $\\mathbf{y}_0$ into a specific sample noise label distribution $\\mathbf{y}_n$ . Ultimately, $\\mathbf{y}_0$ diffuses to the endpoint $\\mathbf{y}_T = \\mathbf{y}_0 + \\epsilon$ . Notably, the features of the samples serve as control conditions, guiding the generation of labels only during the reverse process." + ], + "image_footnote": [], + "bbox": [ + 133, + 88, + 867, + 300 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "more classical interpretative framework known as denoising diffusion probabilistic models (DDPM), which have been widely applied across diverse domains. RDDM expands the denoising-based diffusion models, initially uninterpretable for image restoration, into a unified and interpretable model for both image generation and restoration by introducing residuals. Classification and regression diffusion models (CARD) [18] reconstruct the classification as a conditional generative task, generating labels from dimensionally reduced image features via a pre-trained encoder. Building on CARD, label-retrieval-augmented (LRA) diffusion model [4] re-estimates the generation targets through neighbor label retrieval, improving label accuracy. Our model disentangles the diffusion process into directional and random diffusion channels, enhancing diffusion quality through precorrection, which leads to train a more robust label diffusion model.", + "bbox": [ + 91, + 401, + 483, + 656 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3. Preliminary", + "text_level": 1, + "bbox": [ + 89, + 665, + 217, + 683 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "In this section, we primarily introduce the label denoising diffusion models (LDDM), which is constructed based on the theoretical framework of the DDPM [21]. We start by defining the data distribution $\\mathbf{y}_0\\sim q(\\mathbf{y}_0)$ and a Markovian forward (diffusion) process $q$ , which progressively adds Gaussian noise $\\epsilon \\sim \\mathcal{N}(0,\\mathbf{I})$ to the labels based on diffusion strength $\\beta_{t}$ , diffusing label $\\mathbf{y}_0$ to $\\mathbf{y}_T$ ..", + "bbox": [ + 89, + 686, + 482, + 794 + ], + "page_idx": 2 + }, + { + "type": "equation", + "text": "\n$$\nq \\left(\\mathbf {y} _ {t} \\mid \\mathbf {y} _ {t - 1}\\right) = \\mathcal {N} \\left(\\mathbf {y} _ {t}; \\sqrt {1 - \\beta_ {t}} \\mathbf {y} _ {t - 1}, \\beta_ {t} \\mathbf {I}\\right), \\tag {1}\n$$\n", + "text_format": "latex", + "bbox": [ + 135, + 797, + 482, + 824 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Under Bayes' theorem, one finds that the posterior distribution of the forward process can be represented as a Gaussian distribution with mean $\\tilde{\\mu}_t(\\mathbf{y}_t,\\mathbf{y}_0)$ and variance $\\tilde{\\beta}_t$ :", + "bbox": [ + 89, + 827, + 483, + 875 + ], + "page_idx": 2 + }, + { + "type": "equation", + "text": "\n$$\nq \\left(\\mathbf {y} _ {t - 1} \\mid \\mathbf {y} _ {t}, \\mathbf {y} _ {0}\\right) = \\mathcal {N} \\left(\\mathbf {y} _ {t - 1}; \\tilde {\\mu} _ {t} \\left(\\mathbf {y} _ {t}, \\mathbf {y} _ {0}\\right), \\tilde {\\beta} _ {t} \\mathbf {I}\\right), \\tag {2}\n$$\n", + "text_format": "latex", + "bbox": [ + 114, + 878, + 482, + 905 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "where $\\tilde{\\mu}_t(\\mathbf{y}_t,\\mathbf{y}_0) = \\frac{\\sqrt{\\bar{\\alpha}_{t - 1}}\\beta_t}{1 - \\bar{\\alpha}_t}\\mathbf{y}_0 + \\frac{\\sqrt{\\alpha_t(1 - \\bar{\\alpha}_{t - 1})}}{1 - \\bar{\\alpha}_t}\\mathbf{y}_t$ and $\\tilde{\\beta}_t = \\frac{1 - \\bar{\\alpha}_{t - 1}}{1 - \\bar{\\alpha}_t}\\beta_t$ , In the reverse process, we naturally employ a neural network to perform Markovian inference from $q(\\mathbf{y}_T)$ , fitting the posterior distribution $q(\\mathbf{y}_{t - 1}|\\mathbf{y}_t)$ through to $\\mathbf{y}_0$ . Since the variance of the posterior Gaussian distribution depends only on the diffusion strength corresponding to time $t$ , the neural network focuses primarily on predicting and learning the mean:", + "bbox": [ + 511, + 398, + 906, + 526 + ], + "page_idx": 2 + }, + { + "type": "equation", + "text": "\n$$\np _ {\\theta} \\left(\\mathbf {y} _ {t - 1} \\mid \\mathbf {y} _ {t}\\right) := \\mathcal {N} \\left(\\mathbf {y} _ {t - 1}; \\mu_ {\\theta} \\left(\\mathbf {y} _ {t}, t\\right), \\tilde {\\beta} _ {t} \\mathbf {I}\\right). \\tag {3}\n$$\n", + "text_format": "latex", + "bbox": [ + 555, + 537, + 906, + 564 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "To train this model such that $p_{\\theta}(\\mathbf{y}_0)$ closely approximates $q(\\mathbf{y}_0)$ , we can optimize the variational lower bound (VLB), however, research [7] found that directly predicting the mean was not effective, so they reparameterized the estimation target, constraining the neural network to estimate the actual noise of the diffusion process and simplified the loss function:", + "bbox": [ + 511, + 573, + 905, + 678 + ], + "page_idx": 2 + }, + { + "type": "equation", + "text": "\n$$\nL _ {\\text {s i m p l e}} := E _ {t \\sim [ 1, T ], \\mathbf {y} _ {0} \\sim q (\\mathbf {y} _ {0}), \\epsilon \\sim \\mathcal {N} (0, \\mathbf {I})} \\left[ \\| \\epsilon - \\epsilon_ {\\theta} (\\mathbf {y} _ {t}, t) \\| ^ {2} \\right]. \\tag {4}\n$$\n", + "text_format": "latex", + "bbox": [ + 514, + 686, + 903, + 724 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "During inference, we utilize the label noise estimated by the model, restoring the mean of the label at time $t - 1$ to retrieve its distribution form, and then gradually performing random sampling until generating the label $\\mathbf{y}_0$ :", + "bbox": [ + 511, + 724, + 905, + 786 + ], + "page_idx": 2 + }, + { + "type": "equation", + "text": "\n$$\n\\mathbf {y} _ {t - 1} = \\frac {1}{\\sqrt {\\alpha_ {t}}} \\left(\\mathbf {y} _ {t} - \\frac {1 - \\alpha_ {t}}{\\sqrt {1 - \\bar {\\alpha} _ {t}}} \\epsilon_ {\\theta} (\\mathbf {y} _ {t}, t)\\right) + \\sqrt {\\tilde {\\beta} _ {t}} \\mathbf {z}, \\tag {5}\n$$\n", + "text_format": "latex", + "bbox": [ + 527, + 795, + 906, + 830 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "where $\\mathbf{z} \\sim \\mathcal{N}(0, \\mathbf{I})$ . Unfortunately, this label generation process lacks any conditional information, resulting in random label generation for images with identical features, which fails to achieve classification. To address this, we", + "bbox": [ + 511, + 839, + 906, + 900 + ], + "page_idx": 2 + }, + { + "type": "page_number", + "text": "25740", + "bbox": [ + 478, + 944, + 519, + 957 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "analyze the foundational paradigm of conditional label diffusion models in Appendix A and simplify the method of feature conditioning.", + "bbox": [ + 89, + 90, + 485, + 137 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "4. Directional Label Diffusion Model", + "text_level": 1, + "bbox": [ + 89, + 148, + 403, + 165 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Based on LDDM, our goal is to develop a directional diffusion process on labels to introduce more knowledge about noise. We modify the representation of $\\mathbf{y}_T = \\epsilon$ in LDDM to $\\mathbf{y}_T = \\mathbf{y}_n + \\epsilon$ in our DLD, where $\\mathbf{y}_n$ is an estimation of noisy label distribution. The improved forward process is no longer a simple random noise diffusion but is disentangled into two paths of directional diffusion and random diffusion, as shown in Figure 1.", + "bbox": [ + 89, + 170, + 483, + 291 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "In the following subsections, we first outline the overall framework of our model in Section 4.1. Next, in Section 4.2, we present the foundational theory of the DLD diffusion process and analyze the method of conditional control. In Section 4.3, we describe the reverse process and training method, including optimizations for introducing feature conditions and an adapted accelerated inference method. Finally, in Section 4.4, we describe the rapid label pre-correction method within the DLD framework.", + "bbox": [ + 89, + 292, + 483, + 428 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "4.1. Model Overview", + "text_level": 1, + "bbox": [ + 89, + 439, + 254, + 454 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "We propose the Directional Label Diffusion (DLD) framework, with the overall process illustrated in Figure 2. The framework contains two main stages: label pre-correction (Stage 1) and the directional label diffusion model (Stage 2). In stage 1, we first use two different data augmentation methods to provide two views for each sample, then use a pre-trained feature extractor $f_{ext}$ to observe the neighborhood label distribution $p_w$ and $p_s$ of the same sample in two feature spaces. Based on similarity constraints and the principle of neighborhood consistency, we divide the dataset into clean, noisy, and hard sets. Then, the label precorrector and noise distribution estimator prepare $\\mathbf{y}_0$ and $\\mathbf{y}_n$ based on the sample partitioning results, using $\\mathbf{y}_0$ as the initial diffusion label and $\\mathbf{y}_n$ as part of the diffusion endpoint for the next stage of model training. Otherwise, we introduce both augmented image features and high-quality features extracted by $f_{ext}$ into the training process to jointly guide the learning of the robust diffusion model.", + "bbox": [ + 89, + 460, + 483, + 734 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "4.2. Forward Process of DLD", + "text_level": 1, + "bbox": [ + 89, + 744, + 318, + 758 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "To model the random diffusion of labels and the addition of directional noise information, we define the single forward (diffusion) process step in our DLD as follows:", + "bbox": [ + 89, + 767, + 483, + 813 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\mathbf {y} _ {t} = \\mathbf {y} _ {t - 1} + \\alpha_ {t} \\mathbf {y} _ {d} + \\beta_ {t} \\epsilon , \\tag {6}\n$$\n", + "text_format": "latex", + "bbox": [ + 197, + 827, + 482, + 842 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "where $\\mathbf{y}_d$ is the deviation between $\\mathbf{y}_n$ and $y_0$ (i.e., $\\mathbf{y}_d = \\mathbf{y}_n - \\mathbf{y}_0$ ), representing the directional shift contained in the diffusion from $\\mathbf{y}_{t-1}$ to $\\mathbf{y}_t$ . Two independent coefficients, $\\alpha_t$", + "bbox": [ + 89, + 854, + 483, + 901 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "and $\\beta_{t}$ control the degrees of directional and random diffusion, respectively. Based on the additivity of Gaussian noise, we derive the jump diffusion formula from $\\mathbf{y}_0$ to $\\mathbf{y}_t$ without requiring single-step continuous diffusion (see Appendix B for the proof):", + "bbox": [ + 511, + 90, + 905, + 167 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\mathbf {y} _ {t} = \\mathbf {y} _ {0} + \\bar {\\alpha} _ {t} \\mathbf {y} _ {d} + \\bar {\\beta} _ {t} \\epsilon , \\tag {7}\n$$\n", + "text_format": "latex", + "bbox": [ + 627, + 176, + 906, + 195 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "where $\\bar{\\alpha}_t = \\sum_{i=1}^t \\alpha_i$ and $\\bar{\\beta}_t = \\sqrt{\\sum_{i=1}^t \\beta_i^2}$ . When $t = T$ , we hope $\\mathbf{y}_T = \\mathbf{y}_n + \\epsilon = \\mathbf{y}_0 + \\mathbf{y}_d + \\epsilon$ , so we usually set $\\bar{\\alpha}_T = \\bar{\\beta}_T = 1$ . This implies that the two coefficients $\\alpha_t$ and $\\beta_t$ are not arbitrary transformations; rather, they should satisfy a specific sum constraint. Then, we redefine the conditional Markov diffusion process similar to the directional label forward distribution $q$ without condition (detailed theoretical framework in Appendix C), assuming that when conditioned on features $\\mathbf{x}$ , the prior distribution of $\\hat{q}$ is the same as $q$ :", + "bbox": [ + 511, + 207, + 906, + 364 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\begin{array}{l} \\hat {q} \\left(\\mathbf {y} _ {t} \\mid \\mathbf {y} _ {t - 1}, \\mathbf {y} _ {d}, \\mathbf {x}\\right) := q \\left(\\mathbf {y} _ {t} \\mid \\mathbf {y} _ {t - 1}, \\mathbf {y} _ {d}\\right) \\tag {8} \\\\ = \\mathcal {N} \\left(\\mathbf {y} _ {t}; \\mathbf {y} _ {t - 1} + \\alpha_ {t} \\mathbf {y} _ {d}, \\beta_ {t} ^ {2} \\mathbf {I}\\right), \\\\ \\end{array}\n$$\n", + "text_format": "latex", + "bbox": [ + 535, + 375, + 903, + 412 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\begin{array}{l} \\hat {q} \\left(\\mathbf {y} _ {1: T} \\mid \\mathbf {y} _ {0}, \\mathbf {y} _ {d}, \\mathbf {x}\\right) := \\prod_ {t = 1} ^ {T} \\hat {q} \\left(\\mathbf {y} _ {t} \\mid \\mathbf {y} _ {t - 1}, \\mathbf {y} _ {d}, \\mathbf {x}\\right) \\tag {9} \\\\ = \\mathcal {N} \\left(\\mathbf {y} _ {t}; \\mathbf {y} _ {0} + \\bar {\\alpha} _ {t} \\mathbf {y} _ {d}, \\bar {\\beta} _ {t} ^ {2} \\mathbf {I}\\right), \\\\ \\end{array}\n$$\n", + "text_format": "latex", + "bbox": [ + 539, + 438, + 905, + 500 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "which means that the diffusion process of the labels is not influenced by any image features.", + "bbox": [ + 511, + 508, + 905, + 539 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "In Appendix C, using Bayes' rule and Taylor expansion, we derive the conditional posterior distribution during the forward process:", + "bbox": [ + 511, + 540, + 905, + 584 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\hat {q} \\left(\\mathbf {y} _ {t - 1} \\mid \\mathbf {y} _ {t}, \\mathbf {y} _ {0}, \\mathbf {y} _ {d}, \\mathbf {x}\\right) = \\mathcal {N} \\left(\\mathbf {y} _ {t - 1}; \\hat {\\mu} _ {t}, \\hat {\\sigma} _ {t} \\mathbf {I}\\right), \\tag {10}\n$$\n", + "text_format": "latex", + "bbox": [ + 544, + 598, + 903, + 616 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "where $\\hat{\\mu}_t = y_t - \\alpha_t y_d - \\frac{\\beta_t^2}{\\beta_t} \\epsilon + \\hat{\\sigma}_t \\nabla_{y_{t-1}} \\log \\hat{q}(\\mathbf{x} | \\mathbf{y}_{t-1}, \\mathbf{y}_d)$ and $\\hat{\\sigma}_t = \\frac{\\beta_t^2 \\bar{\\beta}_{t-1}^2}{\\beta_t^2}$ . The last term in $\\hat{\\mu}_t$ can be considered a deviation to the unconditional distribution mean, guided by the first derivative of $\\hat{q}(\\mathbf{x} | \\mathbf{y}_{t-1}, \\mathbf{y}_d)$ , indicating that the decoding information from $\\mathbf{y}_{t-1}$ to $\\mathbf{x}$ should be introduced in training.", + "bbox": [ + 511, + 627, + 905, + 733 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "4.3. Reverse and Training Process of DLD", + "text_level": 1, + "bbox": [ + 511, + 741, + 839, + 758 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Reverse Process and Training. In the reverse (generation) process, as indicated by Eq. 6, we need to simultaneously predict $\\mathbf{y}_d$ and $\\epsilon$ to progressively recover $\\mathbf{y}_T$ to $\\mathbf{y}_0$ . Therefore, we can train two independent networks to estimate $\\mathbf{y}_d^{\\theta_1}$ and $\\epsilon_{\\theta_2}$ , respectively, to compute the reverse posterior distribution. Furthermore, we directly use feature information as input for two networks:", + "bbox": [ + 511, + 763, + 905, + 869 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\np _ {\\theta} \\left(\\mathbf {y} _ {t - 1} \\mid \\mathbf {y} _ {t}, \\mathbf {x}\\right) := \\mathcal {N} \\left(\\mathbf {y} _ {t - 1}; \\mu_ {\\theta}, \\hat {\\sigma} _ {t} ^ {2} \\mathbf {I}\\right). \\tag {11}\n$$\n", + "text_format": "latex", + "bbox": [ + 571, + 881, + 903, + 900 + ], + "page_idx": 3 + }, + { + "type": "page_number", + "text": "25741", + "bbox": [ + 478, + 944, + 517, + 955 + ], + "page_idx": 3 + }, + { + "type": "image", + "img_path": "images/e97e2bb5b87d4508df05e432f7987aa9cc5744c0ac32100494bd5f996e06b6d5.jpg", + "image_caption": [ + "Figure 2. Overview of the DLD framework. During Stage 1, the noisy dataset is quickly divided into three subsets. Then, we estimate $\\mathbf{y}_0$ and $\\mathbf{y}_n$ for each samples based on Eq. 14 and Eq. 15. In Stage 2, the label information and augmented features from Stage 1 are jointly used to train a directional diffusion model (blue bars represent random diffusion and gray bars represent directional diffusion)." + ], + "image_footnote": [], + "bbox": [ + 169, + 92, + 823, + 375 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Based on the analysis of the forward process, we adopt a straightforward approach which naturally integrates the decoding process into the diffusion model architecture by a fixed decoding layer $L_{1}$ to map $\\mathbf{y}$ into the feature space of $\\mathbf{x}$ . Detailed information about the specific diffusion model network structure is shown in Appendix Figure C.1.", + "bbox": [ + 88, + 454, + 483, + 545 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "By optimizing the VLB and reparameterization [21], we derive the following simplified loss function for training:", + "bbox": [ + 89, + 546, + 483, + 575 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {L} _ {d} = \\left\\| \\mathbf {y} _ {d} - \\mathbf {y} _ {d} ^ {\\theta_ {1}} \\left(\\mathbf {y} _ {t}, \\mathbf {y} _ {n}, \\mathbf {x}, t\\right) \\right\\| ^ {2}, \\tag {12}\n$$\n", + "text_format": "latex", + "bbox": [ + 169, + 587, + 482, + 609 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {L} _ {\\epsilon} = \\left\\| \\epsilon - \\epsilon_ {\\theta_ {2}} \\left(\\mathbf {y} _ {t}, \\mathbf {y} _ {n}, \\mathbf {x}, t\\right) \\right\\| ^ {2}, \\tag {13}\n$$\n", + "text_format": "latex", + "bbox": [ + 176, + 633, + 482, + 652 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "where $\\epsilon \\sim \\mathcal{N}(0,\\mathbf{I})$ , and the training input $\\mathbf{y}_t$ is synthesized using Eq. 7. Algorithm 1 describes the training procedure.", + "bbox": [ + 89, + 657, + 482, + 688 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Inference. Since the classification diffusion model is deterministic in inference and the dimensionality of the label vector is much lower than that of images, we mimic the inference process of DDIM [44] to generate labels in as few steps as possible without affecting the generated categories. We can gradually sample labels step-by-step following $\\mathbf{y}_{t-1} = \\mathbf{y}_t - (\\bar{\\alpha}_t - \\bar{\\alpha}_{t-1})\\mathbf{y}_d^{\\theta_1} - (\\bar{\\beta}_t - \\bar{\\beta}_{t-1})\\epsilon_{\\theta_2}$ , where the transition from $t-1$ to $t$ is discrete rather than continuous in time. Appendix D provides a detailed explanation of the de-Markovization process, the accelerated sampling algorithm and the visual demonstration of the sampling process. In experiments, we use 5 steps for DLD's inference, which is on the same order of magnitude as the speed of traditional classification models.", + "bbox": [ + 89, + 688, + 483, + 900 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "4.4. Label Pre-correction Method", + "text_level": 1, + "bbox": [ + 511, + 454, + 774, + 469 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Although diffusion models are powerful, under the guidance of incorrect $\\mathbf{y}_0$ and $\\mathbf{y}_n$ , they can easily fit the wrong conditional distributions. Therefore, we propose a one-step label pre-correction method that automatically provides accurate label information and feature conditions for the DLD model, requiring no additional parameter training, making it simple, straightforward, and efficient.", + "bbox": [ + 511, + 479, + 906, + 584 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Label Pre-correction for $\\mathbf{y}_0$ . Inspired by the concepts of multi-view[42, 53] and neighbor consistency regularization [22], our primary assumption is that the same sample should have similar neighborhood label distributions in different feature spaces. We define the original noisy dataset as $\\mathcal{D} = \\{(x_i,\\tilde{y}_i)\\mid x_i\\in \\mathbb{R}^d,\\tilde{y}_i\\in \\mathbb{R}^m,i = 1,\\ldots ,n\\}$ . To obtain different feature spaces by a shared feature extractor $f_{ext}$ , we first apply two different data augmentation methods and map them into $\\mathbb{R}^w$ and $\\mathbb{R}^s$ , with the augmented datasets respectively denoted as $\\mathcal{D}^w$ and $\\mathcal{D}^s$ . Then, we use Euclidean or Cosine similarity to obtain the nearest $K$ neighborhood labels $y_{k = 1}^{K}$ and feature space distances $v_{k = 1}^{K}$ then we can get two weighted label distributions $p_w$ and $p_s$ by $p(y_{i}) = \\left(\\sum_{k = 1}^{K}w_{k}\\cdot \\mathbf{1}(y_{k} = y_{i})\\right)\\cdot \\left(\\sum_{k = 1}^{K}w_{k}\\right)^{-1}$ where $w_{k} = \\frac{1}{v_{k} + \\delta}$ , where $\\delta$ is a small constant to prevent division by zero. Next, we calculate the divergence between the two distributions using the Kullback-Leibler divergence and a via Gaussian mixture model (GMM) [38] to re-partitioning three types of samples: Samples with high means are reassigned to the hard set $\\mathcal{D}_H$ , which represents", + "bbox": [ + 511, + 587, + 908, + 902 + ], + "page_idx": 4 + }, + { + "type": "page_number", + "text": "25742", + "bbox": [ + 478, + 944, + 519, + 957 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "samples whose correctness is hard to determine and typically lie at the boundaries of classes. Samples with low means indicate that similar neighborhood distributions under both views. Using the principle of neighborhood consistency, we compare $\\text{argmax}(p_{ws})$ with the original label $\\tilde{y}$ , where $p_{ws} = \\frac{p_w + p_s}{2}$ . If the class with the highest frequency in the average neighborhood distribution matches the original label, then the sample is assigned to the clean set $\\mathcal{D}_C$ ; otherwise, it is assigned to the noisy set $\\mathcal{D}_N$ .", + "bbox": [ + 89, + 90, + 480, + 226 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Based on the repartitioned sample subsets and the original noisy labels $\\tilde{y}$ , we define a label pre-corrector, which functions similarly to a transition matrix. It can automatically perform label flipping or refinement:", + "bbox": [ + 89, + 227, + 480, + 287 + ], + "page_idx": 5 + }, + { + "type": "equation", + "text": "\n$$\nP \\left(y _ {0} \\mid \\tilde {y}, x\\right) = \\left\\{ \\begin{array}{l l} \\tilde {y} & \\text {i f} x \\in \\mathcal {D} _ {C}, \\\\ \\operatorname {a r g m a x} \\left(p _ {w s}\\right) & \\text {i f} x \\in \\mathcal {D} _ {N}, \\\\ p _ {w s} & \\text {i f} x \\in \\mathcal {D} _ {H}. \\end{array} \\right. \\tag {14}\n$$\n", + "text_format": "latex", + "bbox": [ + 109, + 297, + 482, + 347 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "According to Eq. 14, for samples with deterministic neighborhood label distributions, we use hard labels to minimize the interference of redundant information on $y_0$ . For samples with more contentious neighborhood label distributions, we adopt a soft strategy for label refinement. Instead of point estimation, we estimate through a label vector, retaining most of the neighborhood information. This strategy enhances the accuracy of the $y_0$ while preserving as much robust information as possible.", + "bbox": [ + 89, + 357, + 482, + 492 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Estimate Noise distribution for $\\mathbf{y}_n$ . Similar to obtaining $\\mathbf{y}_0$ , we use the subset of divided samples to estimate the noise distribution of the samples. According to Eq. 15, when a sample is identified as clean, the directed diffusion term is used to gradually fade out $\\mathbf{y}_0$ ( $\\mathbf{y}_d = -\\mathbf{y}_0$ ), allowing the model to learn to recover a label closer to the ground truth from an uninformed $\\epsilon$ . When the sample contains noise, the endpoint of diffusion should include a clear representation of noise, enabling the model to strictly avoid recovering such noisy vector values during reverse generation. For uncertain samples, we simulate potential noise distributions by retaining the difference in label distributions between two views. The greater the difference in the distribution values for a particular class in the two views, the more likely the sample is incorrectly labeled as that class.", + "bbox": [ + 89, + 494, + 482, + 720 + ], + "page_idx": 5 + }, + { + "type": "equation", + "text": "\n$$\nP \\left(y _ {n} \\mid \\tilde {y}, x\\right) = \\left\\{ \\begin{array}{l l} 0 & \\text {i f} x \\in \\mathcal {D} _ {C}, \\\\ \\tilde {y} & \\text {i f} x \\in \\mathcal {D} _ {N}, \\\\ \\frac {\\left| p _ {w} - p _ {s} \\right|}{\\sum \\left| p _ {w} - p _ {s} \\right|} & \\text {i f} x \\in \\mathcal {D} _ {H}. \\end{array} \\right. \\tag {15}\n$$\n", + "text_format": "latex", + "bbox": [ + 122, + 732, + 482, + 785 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "The complete training process is outlined in Algorithm 1, consisting of the following phases: Steps $1 \\sim 3$ for sample re-partitioning, Steps $4 \\sim 9$ for diffusion model training, with Step 6 dedicated to $\\mathbf{y}_0$ pre-correction and $\\mathbf{y}_n$ estimation. Notably, in Step 5, instead of sampling features from the original noisy set $\\mathcal{D}$ , we use a mixture of sample features from $\\mathcal{D}^w$ and $\\mathcal{D}^s$ as the training input for the diffusion", + "bbox": [ + 89, + 794, + 482, + 900 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Algorithm 1 DLD Training", + "text_level": 1, + "bbox": [ + 514, + 90, + 700, + 106 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Input: Noisy training set $\\mathcal{D} = \\{\\mathbf{X},\\tilde{\\mathbf{Y}}\\}$ , feature extractor $f_{\\mathrm{ext}}$ , two network parameters $\\theta_{1}$ and $\\theta_{2}$", + "bbox": [ + 513, + 109, + 903, + 140 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Output: $\\theta_{1}$ and $\\theta_{2}$", + "bbox": [ + 514, + 141, + 640, + 155 + ], + "page_idx": 5 + }, + { + "type": "list", + "sub_type": "text", + "list_items": [ + "1: Obtain two augmented features $\\mathcal{D}^w$ and $\\mathcal{D}^s$ using the feature extractor $f_{\\mathrm{ext}}$ .", + "2: Calculate K-nearest label distributions $p_w$ and $p_s$ .", + "3: Re-partitioning samples into $\\mathcal{D}_C, \\mathcal{D}_N$ and $\\mathcal{D}_H$ .", + "4: while not converged do", + "5: Sample data $(\\mathbf{x}_{ws},\\tilde{\\mathbf{y}})\\sim \\{\\mathcal{D}^w,\\mathcal{D}^s\\}$ ; time slice $t\\sim$ $\\{1,\\dots ,T\\}$ ; and noise $\\epsilon \\sim \\mathcal{N}(0,\\mathbf{I})$", + "6: Pre-estimate $\\mathbf{y}_0$ and $\\mathbf{y}_n$ based on Eq. 14 and Eq. 15, respectively.", + "7: Accumulate losses through Eq. 12 and Eq. 13.", + "8: Take gradient descent step on the losses to update network parameters.", + "9: end while" + ], + "bbox": [ + 521, + 157, + 903, + 353 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "model, where $\\mathbf{x}_{ws} = (1 - \\gamma)\\mathbf{x}_w + \\gamma \\mathbf{x}_s$ , with $\\gamma = \\frac{|\\mathcal{D}_H|}{|\\mathcal{D}|}$ representing the proportion of uncertain samples, which means that the more difficult the sample classification, the richer the feature conditions the model automatically introduces. This reflects an implicit regularization method that mitigates the issue of the diffusion model being overly confident when still learning incorrect information.", + "bbox": [ + 511, + 380, + 906, + 489 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "5. Experimental Results and Analysis", + "text_level": 1, + "bbox": [ + 511, + 500, + 830, + 517 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "In this section, we first evaluate the DLD's effectiveness on CIFAR datasets with varying levels of IDN in Section 5.1. Then, We conduct ablation studies in Section 5.2 to analyze the contributions of the pre-trained feature extractor, the label pre-correction technology, and other components. In Section 5.3, we compare the DLD to the SOTA methods across four real-world noisy datasets. Appendix E provides more details about our experiments.", + "bbox": [ + 511, + 521, + 906, + 643 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "5.1. Results on Simulated Noisy Datasets", + "text_level": 1, + "bbox": [ + 511, + 652, + 828, + 667 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "We conduct experiments on the CIFAR-10 and CIFAR-100 datasets [27] with simulated IDN [6]. Details of the IDN generation algorithm and examples are provided in Appendix E.1. In our experiments, we generate noise levels of $20\\%$ , $40\\%$ , and $60\\%$ . To obtain the feature space for label pre-correction, we provide multiple configurations of feature extractors, including DLD+ResNet, DLD+SimCLR and DLD+ViT (see Appendix F.2 for detailed settings).", + "bbox": [ + 511, + 672, + 906, + 795 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "As shown in Table 1, the DLD+ResNet configuration outperforms the baseline across most noise levels on CIFAR, especially in high-noise environments. We analyze that the main factor affecting its performance in low noise environment is the introduction of too much robust information, which leads to the interference of the model. When using the DLD+ViT configuration, its performance surpasses", + "bbox": [ + 511, + 795, + 906, + 900 + ], + "page_idx": 5 + }, + { + "type": "page_number", + "text": "25743", + "bbox": [ + 478, + 944, + 517, + 955 + ], + "page_idx": 5 + }, + { + "type": "table", + "img_path": "images/1bf87349bc6108a20b0cc43e23ffbb4c475582bd095e76175d2b195376dc57b4.jpg", + "table_caption": [ + "Table 1. Classification accuracies (%) of CIFAR datasets under different levels of IDN." + ], + "table_footnote": [], + "table_body": "
MethodCIFAR-10CIFAR-100
20%40%60%20%40%60%
Co-teaching [17]87.28±0.2078.72±0.4762.51±1.9869.26±0.3959.53±0.5746.80±0.51
GCE [61]86.44±0.2376.71±0.3960.76±3.0868.24±0.2755.76±1.1846.75±2.46
DAC [45]86.16±0.1374.80±0.3259.97±0.9068.70±0.7855.71±1.0146.29±3.35
DMI [54]86.57±0.1677.81±0.8569.94±1.3168.71±1.2157.99±0.9250.96±2.08
SEAL [6]87.79±0.0982.98±0.0570.07±1.5470.61±0.1961.18±0.3551.86±1.32
NCR [22]87.74±0.1583.03±0.2169.44±1.0170.68±0.3761.93±0.5449.13±2.77
CC+DivideMix [62]88.36±0.1184.18±0.4071.52±1.7370.09±0.1964.35±0.3553.20±0.66
DISC [31]90.68±0.1085.61±0.3273.59±1.1371.49±0.1364.46±0.2953.62±0.51
LRA-diffusion+ResNet [4]89.94±0.1586.73±0.3774.66±1.0272.35±0.1465.55±0.3054.43±0.53
DLD+ResNet (Ours)90.43±0.1287.17±0.2976.07±1.1072.60±0.1566.59±0.3155.72±0.24
EPL+ViT [26]95.95±0.1793.01±0.3376.16±1.1872.24±0.1267.20±0.2055.59±0.13
LRA-diffusion+ViT [4]96.65±0.1893.68±0.3478.12±1.1773.59±0.1367.67±0.2157.21±0.12
DLD+ViT (Ours)96.98±0.1296.49±0.0783.14±0.8178.62±0.1176.03±0.1469.57±0.17
", + "bbox": [ + 169, + 107, + 834, + 345 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "EPL+ViT and LRA-diffusion+ViT. This implies that DLD more effectively utilizes the provided prior knowledge without requiring additional training conditions. Across various datasets, DLD's superiority is more pronounced on CIFAR-100, indicating that the more complex IDN on CIFAR-100 severely impacts the performance of other methods, whereas our model maintains robust learning.", + "bbox": [ + 89, + 368, + 480, + 474 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Additionally, we conducted experiments on the simulated CCN dataset. The results (see Appendix E.2) show that CCN is easier to handle than IDN, which aligns with our expectations. Compared to other methods, DLD still achieves the highest classification accuracy. Due to the lightweight architecture of the diffusion model and the single-step pre-correction process, DLD outperforms traditional LNL methods without incurring significant computational overhead. A detailed analysis of training time overhead and model parameters is provided in Appendix E.3.", + "bbox": [ + 88, + 474, + 482, + 628 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "5.2. Ablation Studies", + "text_level": 1, + "bbox": [ + 89, + 636, + 254, + 651 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Power of label pre-correction. We validate the effectiveness of label pre-correction by freezing the diffusion model. As shown in Figure 3, our pre-correction method consistently outperforms in various noisy environments, while Figure 3b demonstrates that the LRA estimation method disrupts originally correct labels. Then, we unfroze the diffusion model to verify that DLD does not fully depend on the pre-correction method. Results show that DLD further enhances generalization performance based on pre-correction, indicating that the two methods complement each other effectively. Additionally, we design more comprehensive cross-ablation experiments, e.g., replacing the LRA technique in LRA-diffusion with the pre-correction module to demonstrate the necessity of directional diffusion. See Appendix F.1 for more details.", + "bbox": [ + 89, + 657, + 482, + 883 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Contribution of different pre-trained features and other", + "bbox": [ + 89, + 885, + 482, + 900 + ], + "page_idx": 6 + }, + { + "type": "table", + "img_path": "images/d8aeea1dae7b845f36b039d219ba80f9232e176804ce280ca4353596adad6cbf.jpg", + "table_caption": [ + "Table 2. Classification accuracies (%) on Animal-10N." + ], + "table_footnote": [], + "table_body": "
MethodAccuracy
CE [35]80.6
GCE [61]81.5
NCT [8]84.1
DISC [31]87.3
SSR [14]88.5
LRA-diffusion [4]88.6
J-ViT [9]89.0
DLD (Ours)89.4
", + "bbox": [ + 550, + 392, + 864, + 534 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "training components. We explore the impact of various feature extractors, including ResNet34 [20], SimCLR [7] and ViT-L/14 [40], and conduct ablation studies comparing our method with CARD and LRA-diffusion. Here, DISC [31] served as a baseline that does not require any pretrained models. Detailed results and analyses are provided in Appendix F.2. Our findings reveal that DLD outperforms other diffusion models when combined with different pretrained models, achieving optimal performance when integrated with ViT-L/14. This underscores ViT-L/14's effectiveness in enhancing label pre-correction and facilitating diffusion model learning. Furthermore, Appendix F.3 details additional ablation studies on various hyperparameter settings. We validate the optimal training strategy for the DLD method and recommend this combination.", + "bbox": [ + 511, + 561, + 906, + 787 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "5.3. Results on Real-world Noisy Datasets", + "text_level": 1, + "bbox": [ + 511, + 801, + 836, + 816 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "We select four real-world noisy datasets for our experiments to demonstrate the superior performance of DLD: Animal10N [43], Clothing1M [52], WebVision [30], and ImageNet ILSVRC2012 [41]. Comprehensive descriptions of these datasets and detailed experimental setups can be", + "bbox": [ + 511, + 824, + 905, + 900 + ], + "page_idx": 6 + }, + { + "type": "page_number", + "text": "25744", + "bbox": [ + 478, + 944, + 517, + 955 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/583fbcb0c2d28e4097d41b09173dd3eece8bb464b0de20c1edf1f9cfafa91618.jpg", + "image_caption": [ + "(a) Experimental results on the CIFAR-10 dataset" + ], + "image_footnote": [], + "bbox": [ + 130, + 88, + 496, + 286 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/7e01a5932844ab8d099133ec3ebdb8ff0faa62526609a98a5f2823c22774ac80.jpg", + "image_caption": [ + "(b) Experimental results on the CIFAR-100 dataset", + "Figure 3. Comparison of different label correction strategies' accuracies, where Baseline represents the proportion of clean labels." + ], + "image_footnote": [], + "bbox": [ + 500, + 89, + 864, + 287 + ], + "page_idx": 7 + }, + { + "type": "table", + "img_path": "images/c70e336627993e4cee09268457f476cc37ad7f5bade2f20f791dcfa59cb762f8.jpg", + "table_caption": [ + "Table 3. Classification accuracies (%) on Clothing1M." + ], + "table_footnote": [], + "table_body": "
MethodAccuracy
Co-teaching [17]69.21
DMI [54]72.46
DivideMix [28]74.76
CC [62]75.40
DISC [31]74.79
EPL [26]75.21
LRA-diffusion [4]74.46
SANM [46]75.63
DLD (Ours)75.69
", + "bbox": [ + 129, + 375, + 442, + 532 + ], + "page_idx": 7 + }, + { + "type": "table", + "img_path": "images/b7a33f3397a47f0357f5c83b4ba3b2223260844b524a52c2b4366ba6f881fcbb.jpg", + "table_caption": [ + "Table 4. Classification accuracies (%) on WebVision and ILSVRC2012 datasets." + ], + "table_footnote": [], + "table_body": "
MethodWebVisionILSVRC12
Co-teaching [17]63.5861.48
DivideMix [28]77.3275.20
ELR [33]77.7870.29
EPL [26]78.7776.51
C2D [63]79.4278.57
CC [62]79.3676.08
DISC [31]80.2877.44
LRA-diffusion [4]84.1682.56
DLD (Ours)84.5183.74
", + "bbox": [ + 129, + 584, + 442, + 739 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "found in Appendix E.4 and Appendix E.5, respectively. As demonstrated in Table 2, Table 3, and Table 4, DLD outperforms all SOTA methods across these datasets. Note that DLD's relative classification performance improved by $7.26\\%$ and $0.44\\%$ over J-ViT [9] and EPL [26], respectively, both employing the ViT architecture. Compared to LRA-diffusion, which also employs ViT and label diffusion models, DLD achieves varying degrees of performance improvement, demonstrating a more effective and reason-", + "bbox": [ + 89, + 763, + 482, + 900 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "able use of prior knowledge and diffusion models. Besides, while ViT-L/14 improved performance on Clothing1M, Appendix G shows that other extractors like CC can be even more effective, highlighting our method's versatility with different LNL methods. Besides, DLD's performance advantage partly stems from the pre-correction method. In future work, we aim to integrate discriminative and generative paradigms, using diffusion models to refine pre-correction, generate label information, and enhance representational power and multimodal generalization. Appendix H discusses additional limitations and future directions.", + "bbox": [ + 511, + 352, + 906, + 517 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "6. Conclusion", + "text_level": 1, + "bbox": [ + 511, + 540, + 633, + 556 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "This paper proposes DLD, a diffusion model specifically designed to learn effectively from noisy labels for robust classification. Our main idea is to decouple the label diffusion process into two channels: directional diffusion and random diffusion. This approach explicitly introduces noisy label information into the diffusion model for the first time, enabling the model to interpretably mitigate the generation of incorrect labels, thereby enhancing its ability to generalize from noisy to clean data. By incorporating our low-cost label pre-correction method, DLD can automatically select label diffusion information for each sample in diverse noisy environments, without requiring additional adjustments or iterations. Additionally, our model synergizes with various pre-trained large-scale models, like ViT, leveraging their prior knowledge more effectively than other methods. Experiments on simulated and real-world noisy datasets show that DLD surpasses conventional LNL methods in both accuracy and efficiency. By framing LNL within a generative model, our method bridges LNL with other learning frameworks. We believe this work will reduce the workload for practitioners handling large-scale, low-quality labeled data, benefiting the development of the image recognition field.", + "bbox": [ + 511, + 568, + 906, + 900 + ], + "page_idx": 7 + }, + { + "type": "page_number", + "text": "25745", + "bbox": [ + 478, + 944, + 517, + 955 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Acknowledgements", + "text_level": 1, + "bbox": [ + 91, + 90, + 258, + 107 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "This work was supported in part by the National Natural Science Foundation of China under Grant U21A20513, Grant 62476157, Grant 62276161, Grant 62276157 and Grant 62406179, and in part by the Key Research and Development Program of Shanxi Province under Grant 202202020101003 and the Fundamental Research Program of Shanxi Province under Grant 202403021222026.", + "bbox": [ + 89, + 114, + 483, + 220 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "References", + "text_level": 1, + "bbox": [ + 91, + 237, + 187, + 252 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[1] Devansh Arpit, Stanisław Jastrzebski, Nicolas Ballas, David Krueger, Emmanuel Bengio, Maxinder S Kanwal, Tegan Maharaj, Asja Fischer, Aaron Courville, Yoshua Bengio, et al. A closer look at memorization in deep networks. In International conference on machine learning, pages 233-242. PMLR, 2017. 1", + "[2] Antonin Berthon, Bo Han, Gang Niu, Tongliang Liu, and Masashi Sugiyama. Confidence scores make instance-dependent label-noise learning possible. In International conference on machine learning, pages 825-836. PMLR, 2021. 1", + "[3] Stella Biderman, Usvsn Prashanth, Lintang Sutawika, Hailey Schoelkopf, Quentin Anthony, Shivanshu Purohit, and Edward Raff. Emergent and predictable memorization in large language models. Advances in Neural Information Processing Systems, 36:28072-28090, 2023. 1", + "[4] Jian Chen, Ruiyi Zhang, Tong Yu, Rohan Sharma, Zhiqiang Xu, Tong Sun, and Changyou Chen. Label-retrieval augmented diffusion models for learning from noisy labels. Advances in Neural Information Processing Systems, 36, 2024. 2, 3, 7, 8", + "[5] Pengfei Chen, Ben Ben Liao, Guangyong Chen, and Shengyu Zhang. Understanding and utilizing deep neural networks trained with noisy labels. In International conference on machine learning, pages 1062-1070. PMLR, 2019. 1", + "[6] Pengfei Chen, Junjie Ye, Guangyong Chen, Jingwei Zhao, and Pheng-Ann Heng. Beyond class-conditional assumption: A primary attempt to combat instance-dependent label noise. In Proceedings of the AAAI Conference on Artificial Intelligence, pages 11442–11450, 2021. 1, 2, 6, 7", + "[7] Ting Chen, Simon Kornblith, Mohammad Norouzi, and Geoffrey Hinton. A simple framework for contrastive learning of visual representations. In International conference on machine learning, pages 1597-1607. PMLR, 2020. 2, 3, 7", + "[8] Yingyi Chen, Shell Xu Hu, Xi Shen, Chunrong Ai, and Johan AK Suykens. Compressing features for learning with noisy labels. IEEE Transactions on Neural Networks and Learning Systems, 35(2):2124-2138, 2022. 7", + "[9] Yingyi Chen, Xi Shen, Yahui Liu, Qinghua Tao, and Johan AK Suykens. Jigsaw-vit: Learning jigsaw puzzles in vision transformer. Pattern Recognition Letters, 166:53-60, 2023. 7, 8", + "[10] De Cheng, Tongliang Liu, Yixiong Ning, Nannan Wang, Bo Han, Gang Niu, Xinbo Gao, and Masashi Sugiyama." + ], + "bbox": [ + 93, + 261, + 483, + 900 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "Instance-dependent label-noise learning with manifold-regularized transition matrix estimation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 16630-16639, 2022. 1", + "[11] Hao Cheng, Zhaowei Zhu, Xingyu Li, Yifei Gong, Xing Sun, and Yang Liu. Learning with instance-dependent label noise: A sample sieve approach. arXiv preprint arXiv:2010.02347, 2020.", + "[12] Jiacheng Cheng, Tongliang Liu, Kotagiri Ramamohanarao, and Dacheng Tao. Learning with bounded instance and label-dependent label noise. In International conference on machine learning, pages 1789-1799. PMLR, 2020. 1", + "[13] C Fabian Benitez-Quiroz, Ramprakash Srinivasan, and Aleix M Martinez. Emotionet: An accurate, real-time algorithm for the automatic annotation of a million facial expressions in the wild. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 5562–5570, 2016. 1", + "[14] Chen Feng, Georgios Tzimiropoulos, and Ioannis Patras. Ssr: An efficient and robust framework for learning with unknown label noise. arXiv preprint arXiv:2111.11288, 2021. 7", + "[15] Arpit Garg, Cuong Nguyen, Rafael Felix, Thanh-Toan Do, and Gustavo Carneiro. Instance-dependent noisy label learning via graphical modelling. In Proceedings of the IEEE/CVF winter conference on applications of computer vision, pages 2288-2298, 2023. 2", + "[16] Aritra Ghosh, Himanshu Kumar, and P Shanti Sastry. Robust loss functions under label noise for deep neural networks. In Proceedings of the AAAI conference on artificial intelligence, 2017. 1, 2", + "[17] Bo Han, Quanming Yao, Xingrui Yu, Gang Niu, Miao Xu, Weihua Hu, Ivor Tsang, and Masashi Sugiyama. Co-teaching: Robust training of deep neural networks with extremely noisy labels. Advances in neural information processing systems, 31, 2018. 1, 2, 7, 8", + "[18] Xizewen Han, Huangjie Zheng, and Mingyuan Zhou. Card: Classification and regression diffusion models. Advances in Neural Information Processing Systems, 35:18100-18115, 2022. 2, 3", + "[19] Kaiming He, Xiangyu Zhang, Shaoqing Ren, and Jian Sun. Delving deep into rectifiers: Surpassing human-level performance on imagenet classification. In Proceedings of the IEEE international conference on computer vision, pages 1026-1034, 2015. 2", + "[20] Kaiming He, Xiangyu Zhang, Shaoqing Ren, and Jian Sun. Deep residual learning for image recognition. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 770-778, 2016. 7", + "[21] Jonathan Ho, Ajay Jain, and Pieter Abbeel. Denoising diffusion probabilistic models. Advances in neural information processing systems, 33:6840-6851, 2020. 2, 3, 5", + "[22] Ahmet Iscen, Jack Valmadre, Anurag Arnab, and Cordelia Schmid. Learning with neighbor consistency for noisy labels. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 4672-4681, 2022. 2, 5, 7" + ], + "bbox": [ + 516, + 92, + 906, + 898 + ], + "page_idx": 8 + }, + { + "type": "page_number", + "text": "25746", + "bbox": [ + 478, + 944, + 519, + 955 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[23] Gaoxia Jiang, Jia Zhang, Xuefei Bai, Wenjian Wang, and Deyu Meng. Which is more effective in label noise cleaning, correction or filtering? In Proceedings of the AAAI Conference on Artificial Intelligence, pages 12866-12873, 2024. 2", + "[24] Lu Jiang, Zhengyuan Zhou, Thomas Leung, Li-Jia Li, and Li Fei-Fei. Mentornet: Learning data-driven curriculum for very deep neural networks on corrupted labels. In International conference on machine learning, pages 2304–2313. PMLR, 2018. 1", + "[25] Diederik P Kingma. Auto-encoding variational bayes. arXiv preprint arXiv:1312.6114, 2013. 2", + "[26] Jongwoo Ko, Sumyeong Ahn, and Se-Young Yun. Efficient utilization of pre-trained model for learning with noisy labels. In ICLR 2023 Workshop on Pitfalls of limited data and computation for Trustworthy ML, 2023. 7, 8", + "[27] Alex Krizhevsky. Learning multiple layers of features from tiny images. Master's thesis, University of Toronto, 2009. 6", + "[28] Junnan Li, Richard Socher, and Steven CH Hoi. Dividemix: Learning with noisy labels as semi-supervised learning. arXiv preprint arXiv:2002.07394, 2020. 1, 2, 8", + "[29] Mingchen Li, Mahdi Soltanolkotabi, and Samet Oymak. Gradient descent with early stopping is provably robust to label noise for overparameterized neural networks. In International conference on artificial intelligence and statistics, pages 4313-4324. PMLR, 2020. 1", + "[30] Wen Li, Limin Wang, Wei Li, Eirikur Agustsson, and Luc Van Gool. Webvision database: Visual learning and understanding from web data. arXiv preprint arXiv:1708.02862, 2017. 7", + "[31] Yifan Li, Hu Han, Shiguang Shan, and Xilin Chen. Disc: Learning from noisy labels via dynamic instance-specific selection and correction. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 24070-24079, 2023. 2, 7, 8", + "[32] Jiawei Liu, Qiang Wang, Huijie Fan, Yinong Wang, Yandong Tang, and Liangqiong Qu. Residual denoising diffusion models. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 2773-2783, 2024. 2", + "[33] Sheng Liu, Jonathan Niles-Weed, Narges Razavian, and Carlos Fernandez-Granda. Early-learning regularization prevents memorization of noisy labels. Advances in neural information processing systems, 33:20331-20342, 2020. 2, 8", + "[34] Yangdi Lu, Yang Bo, and Wenbo He. Co-matching: Combating noisy labels by augmentation anchoring. arXiv preprint arXiv:2103.12814, 2021. 1", + "[35] Anqi Mao, Mehryar Mohri, and Yutao Zhong. Cross-entropy loss functions: Theoretical analysis and applications. In International conference on Machine learning, pages 23803–23828. PMLR, 2023. 7", + "[36] Duc Tam Nguyen, Chaithanya Kumar Mummadi, Thi Phuong Nhung Ngo, Thi Hoai Phuong Nguyen, Laura Beggel, and Thomas Brox. Self: Learning to filter noisy labels with self-ensembling. arXiv preprint arXiv:1910.01842, 2019. 1" + ], + "bbox": [ + 91, + 90, + 482, + 898 + ], + "page_idx": 9 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[37] Kento Nishi, Yi Ding, Alex Rich, and Tobias Hollerer. Augmentation strategies for learning with noisy labels. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 8022-8031, 2021. 1", + "[38] Haim Permuter, Joseph Francos, and Ian Jermyn. A study of gaussian mixture models of color and texture features for image classification and segmentation. Pattern recognition, 39(4):695-706, 2006. 5", + "[39] Lutz Prechelt. Early stopping-but when? In Neural Networks: Tricks of the trade, pages 55-69. Springer, 2002. 1", + "[40] Alec Radford, Jong Wook Kim, Chris Hallacy, Aditya Ramesh, Gabriel Goh, Sandhini Agarwal, Girish Sastry, Amanda Askell, Pamela Mishkin, Jack Clark, et al. Learning transferable visual models from natural language supervision. In International conference on machine learning, pages 8748-8763. PMLR, 2021. 2, 7", + "[41] Olga Russakovsky, Jia Deng, Hao Su, Jonathan Krause, Sanjeev Satheesh, Sean Ma, Zhiheng Huang, Andrej Karpathy, Aditya Khosla, Michael Bernstein, et al. Imagenet large scale visual recognition challenge. International journal of computer vision, 115:211-252, 2015. 7", + "[42] Kihyuk Sohn, David Berthelot, Nicholas Carlini, Zizhao Zhang, Han Zhang, Colin A Raffel, Ekin Dogus Cubuk, Alexey Kurakin, and Chun-Liang Li. Fixmatch: Simplifying semi-supervised learning with consistency and confidence. Advances in neural information processing systems, 33:596-608, 2020. 5", + "[43] Hwanjun Song, Minseok Kim, and Jae-Gil Lee. Selfie: Refurbishing unclean samples for robust deep learning. In International conference on machine learning, pages 5907-5915. PMLR, 2019. 7", + "[44] Jiaming Song, Chenlin Meng, and Stefano Ermon. Denoising diffusion implicit models. arXiv preprint arXiv:2010.02502, 2020. 5", + "[45] Sunil Thulasidasan, Tanmoy Bhattacharya, Jeff Bilmes, Gopinath Chennupati, and Jamal Mohd-Yusof. Combating label noise in deep learning using abstention. arXiv preprint arXiv:1905.10964, 2019. 2, 7", + "[46] Yuanpeng Tu, Boshen Zhang, Yuxi Li, Liang Liu, Jian Li, Jiangning Zhang, Yabiao Wang, Chengjie Wang, and Cai Rong Zhao. Learning with noisy labels via self-supervised adversarial noisy masking. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 16186-16195, 2023. 8", + "[47] Yisen Wang, Weiyang Liu, Xingjun Ma, James Bailey, Hongyuan Zha, Le Song, and Shu-Tao Xia. Iterative learning with open-set noisy labels. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 8688-8696, 2018. 1", + "[48] Yisen Wang, Xingjun Ma, Zaiyi Chen, Yuan Luo, Jinfeng Yi, and James Bailey. Symmetric cross entropy for robust learning with noisy labels. In Proceedings of the IEEE/CVF international computer vision and pattern recognition, pages 322-330, 2019. 1", + "[49] Hongxin Wei, Lei Feng, Xiangyu Chen, and Bo An. Combating noisy labels by agreement: A joint training method with co-regularization. In Proceedings of the IEEE/CVF con" + ], + "bbox": [ + 516, + 92, + 903, + 900 + ], + "page_idx": 9 + }, + { + "type": "page_number", + "text": "25747", + "bbox": [ + 478, + 944, + 517, + 955 + ], + "page_idx": 9 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "ference on computer vision and pattern recognition, pages 13726-13735, 2020. 1", + "[50] Pengxiang Wu, Songzhu Zheng, Mayank Goswami, Dimitris Metaxas, and Chao Chen. A topological filter for learning with label noise. Advances in neural information processing systems, 33:21382-21393, 2020. 1", + "[51] Xiaobo Xia, Tongliang Liu, Bo Han, Nannan Wang, Mingming Gong, Haifeng Liu, Gang Niu, Dacheng Tao, and Masashi Sugiyama. Part-dependent label noise: Towards instance-dependent label noise. Advances in Neural Information Processing Systems, 33:7597-7610, 2020. 1", + "[52] Tong Xiao, Tian Xia, Yi Yang, Chang Huang, and Xiaogang Wang. Learning from massive noisy labeled data for image classification. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 2691-2699, 2015. 7", + "[53] Cai Xu, Yilin Zhang, Ziyu Guan, and Wei Zhao. Trusted multi-view learning with label noise. arXiv preprint arXiv:2404.11944, 2024. 5", + "[54] Yilun Xu, Peng Cao, Yuqing Kong, and Yizhou Wang. L.dmi: A novel information-theoretic loss function for training deep nets robust to label noise. Advances in neural information processing systems, 32, 2019. 2, 7, 8", + "[55] Yan Yan, Rómer Rosales, Glenn Fung, Ramanathan Subramanian, and Jennifer Dy. Learning from multiple annotators with varying expertise. Machine learning, 95:291-327, 2014. 1", + "[56] Quanming Yao, Hansi Yang, Bo Han, Gang Niu, and James Tin-Yau Kwok. Searching to exploit memorization effect in learning with noisy labels. In International Conference on Machine Learning, pages 10789-10798. PMLR, 2020. 1", + "[57] Yu Yao, Tongliang Liu, Mingming Gong, Bo Han, Gang Niu, and Kun Zhang. Instance-dependent label-noise learning under a structural causal model. Advances in Neural Information Processing Systems, 34:4409-4420, 2021. 1", + "[58] Xingrui Yu, Bo Han, Jiangchao Yao, Gang Niu, Ivor Tsang, and Masashi Sugiyama. How does disagreement help generalization against label corruption? In International conference on machine learning, pages 7164-7173. PMLR, 2019. 1", + "[59] Chiyuan Zhang, Samy Bengio, Moritz Hardt, Benjamin Recht, and Oriol Vinyals. Understanding deep learning (still) requires rethinking generalization. Communications of the ACM, 64(3):107-115, 2021. 1", + "[60] Yikai Zhang, Songzhu Zheng, Pengxiang Wu, Mayank Goswami, and Chao Chen. Learning with feature-dependent label noise: A progressive approach. arXiv preprint arXiv:2103.07756, 2021. 1", + "[61] Zhilu Zhang and Mert Sabuncu. Generalized cross entropy loss for training deep neural networks with noisy labels. Advances in neural information processing systems, 31, 2018, 1, 2, 7", + "[62] Ganlong Zhao, Guanbin Li, Yipeng Qin, Feng Liu, and Yizhou Yu. Centrality and consistency: two-stage clean samples identification for learning with instance-dependent noisy labels. In European Conference on Computer Vision, pages 21-37. Springer, 2022. 2, 7, 8" + ], + "bbox": [ + 91, + 92, + 482, + 900 + ], + "page_idx": 10 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[63] Evgenii Zheltonozhskii, Chaim Baskin, Avi Mendelson, Alex M Bronstein, and Or Litany. Contrast to divide: Self-supervised pre-training for learning with noisy labels. In Proceedings of the IEEE/CVF Winter Conference on Applications of Computer Vision, pages 1657-1667, 2022. 2, 8", + "[64] Songzhu Zheng, Pengxiang Wu, Aman Goswami, Mayank Goswami, Dimitris Metaxas, and Chao Chen. Error-bounded correction of noisy labels. In International Conference on Machine Learning, pages 11447-11457. PMLR, 2020. 1", + "[65] Yuanyi Zhong, Haoran Tang, Junkun Chen, Jian Peng, and Yu-Xiong Wang. Is self-supervised learning more robust than supervised learning? arXiv preprint arXiv:2206.05259, 2022. 2" + ], + "bbox": [ + 516, + 92, + 903, + 273 + ], + "page_idx": 10 + }, + { + "type": "page_number", + "text": "25748", + "bbox": [ + 478, + 945, + 517, + 955 + ], + "page_idx": 10 + } +] \ No newline at end of file diff --git a/2025/Directional Label Diffusion Model for Learning from Noisy Labels/133ce87a-4a21-4fe5-b113-05841acf1ebd_model.json b/2025/Directional Label Diffusion Model for Learning from Noisy Labels/133ce87a-4a21-4fe5-b113-05841acf1ebd_model.json new file mode 100644 index 0000000000000000000000000000000000000000..da58a71b4a6b90b2f4275f43f8f57545fb8a947b --- /dev/null +++ b/2025/Directional Label Diffusion Model for Learning from Noisy Labels/133ce87a-4a21-4fe5-b113-05841acf1ebd_model.json @@ -0,0 +1,2378 @@ +[ + [ + { + "type": "header", + "bbox": [ + 0.107, + 0.003, + 0.182, + 0.043 + ], + "angle": 0, + "content": "CVF" + }, + { + "type": "header", + "bbox": [ + 0.237, + 0.001, + 0.812, + 0.047 + ], + "angle": 0, + "content": "This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore." + }, + { + "type": "title", + "bbox": [ + 0.165, + 0.131, + 0.834, + 0.153 + ], + "angle": 0, + "content": "Directional Label Diffusion Model for Learning from Noisy Labels" + }, + { + "type": "text", + "bbox": [ + 0.202, + 0.18, + 0.796, + 0.217 + ], + "angle": 0, + "content": "Senyu Hou\\(^{1}\\), Gaoxia Jiang\\(^{1}\\), Jia Zhang\\(^{1}\\), Shangrong Yang\\(^{1}\\), Husheng Guo\\(^{1}\\), Yaqing Guo\\(^{1}\\), Wenjian Wang\\(^{2*}\\)" + }, + { + "type": "text", + "bbox": [ + 0.166, + 0.217, + 0.83, + 0.27 + ], + "angle": 0, + "content": "\\(^{1}\\)School of Computer and Information Technology, Shanxi University, China \n\\(^{2}\\)Key Laboratory of Computational Intelligence and Chinese Information Processing Ministry of Education, Shanxi University, China" + }, + { + "type": "text", + "bbox": [ + 0.3, + 0.272, + 0.699, + 0.288 + ], + "angle": 0, + "content": "{202312407005, 202412407035}@email.sxu.edu.cn" + }, + { + "type": "text", + "bbox": [ + 0.237, + 0.29, + 0.758, + 0.305 + ], + "angle": 0, + "content": "{jianggaoxia, sryang, guohusheng, yqguo, wjwang}@sxu.edu.cn" + }, + { + "type": "title", + "bbox": [ + 0.248, + 0.34, + 0.327, + 0.356 + ], + "angle": 0, + "content": "Abstract" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.372, + 0.486, + 0.795 + ], + "angle": 0, + "content": "In image classification, the label quality of training data critically influences model generalization, especially for deep neural networks (DNNs). Traditionally, learning from noisy labels (LNL) can improve the generalization of DNNs through complex architectures or a series of robust techniques, but its performance improvement is limited by the discriminative paradigm. Unlike traditional ways, we resolve the LNL problems from the perspective of robust label generation, based on diffusion models within the generative paradigm. To expand the diffusion model into a robust classifier that explicitly accommodates more noise knowledge, we propose a Directional Label Diffusion (DLD) model. It disentangles the diffusion process into two paths, i.e., directional diffusion and random diffusion. Specifically, directional diffusion simulates the corruption of true labels into a directed noise distribution, prioritizing the removal of likely noise, whereas random diffusion introduces inherent randomness to support label recovery. This architecture enable DLD to gradually infer labels from an initial random state, interpretably diverging from the specified noise distribution. To adapt the model to diverse noisy environments, we design a low-cost label precorrection method that automatically supplies more accurate label information to the diffusion model, without requiring manual intervention or additional iterations. Our approach outperforms state-of-the-art methods on both simulated and real-world noisy datasets. Code is available at https://github.com/SenyuHou/DLD." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.806, + 0.222, + 0.822 + ], + "angle": 0, + "content": "1. Introduction" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.829, + 0.484, + 0.875 + ], + "angle": 0, + "content": "Deep neural networks (DNNs) have achieved notable success across key areas of machine learning, driven by the growth of datasets and the increase in model parameters." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.341, + 0.906, + 0.506 + ], + "angle": 0, + "content": "These models are primarily trained using supervised learning, which relies heavily on large volumes of labeled data [5]. However, real-world data often suffer from low-quality labels due to subjective errors like indistinguishable samples or objective errors such as incorrect label recording, making noisy labels a widespread issue [13, 55]. Studies have shown that the generalization capability of deep neural networks is adversely affected when trained with noisy labels [1, 59]. Consequently, it is crucial to develop methods for learning from noisy labels (LNL) on large-scale but low-quality datasets." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.509, + 0.908, + 0.902 + ], + "angle": 0, + "content": "Due to the varied distribution of label noise, historical research has classified it into class-conditional noise (CCN) and instance-dependent noise (IDN) [57]. Recent studies indicate that label noise in real datasets predominantly aligns with the IDN assumption, wherein noise is unevenly distributed and more likely in samples with similar features [2, 6, 10–12, 51]. To tackle these complex noise patterns, mainstream LNL methods employ strategies combining filtering and training approaches. Some studies demonstrate a phenomenon known as the memorization effect [1, 3, 29, 56], i.e., DNNs first learn from simple data before fitting to noisy samples. Based on this observation, sample selection strategies utilize confidence levels or prediction loss during training intervals to identify noisy samples [17, 24, 28, 58]. However, this strategy is inherently risky, as the timing of noise memorization varies across datasets and noise distributions, complicating the selection criteria [6, 39]. Consequently, mainline LNL methods can often be integrated with training strategies to enhance model robustness, employing techniques like label filtering [17, 36, 50, 58], label correction [47, 60, 64], robust loss [16, 48, 61], and semi-supervised learning [28]. Among these techniques, semi-supervised learning, by combining multiple approaches [34, 37, 49], often outperforms other LNL methods but is still constrained by the discriminative framework. Although some researches attempt to combine" + }, + { + "type": "page_footnote", + "bbox": [ + 0.11, + 0.888, + 0.237, + 0.9 + ], + "angle": 0, + "content": "*Corresponding author." + }, + { + "type": "page_number", + "bbox": [ + 0.479, + 0.946, + 0.521, + 0.957 + ], + "angle": 0, + "content": "25738" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.09, + 0.092, + 0.482, + 0.137 + ], + "angle": 0, + "content": "variational autoencoders (VAE) [25] with DivideMix [28] to address label noise [15], most methods remain focused on discriminative models." + }, + { + "type": "text", + "bbox": [ + 0.093, + 0.142, + 0.485, + 0.594 + ], + "angle": 0, + "content": "In this paper, we shift away from the traditional LNL perspective, proposing robust classification models using label diffusion models that focus on the generative modeling and deeply explores the generative relationship between features and labels. Under this paradigm, labels are generated by gradually refining an initial random guess. The uncertainty modeling in diffusion models inherently meets some robustness requirements, as random label generation can offset some noise interference [4, 18]. However, this type of diffusion process lacks information about noisy labels. Intrinsic uncertainty alone is insufficient to handle complex noisy environments. Inspired by the residual diffusion denoising models (RDDM) [32], we explore a novel Directional Label Diffusion (DLD) model that incorporates noisy label information into the diffusion process. In DLD, we decouple the diffusion process into two channels: directional diffusion and random diffusion (see Figure 1). The directional diffusion guides the initial label towards a more probable noise direction (e.g., a green automobile may be misclassified as a frog or other classes, as indicated by the probability distribution in the third column in Figure 1), while random diffusion retains the original diffusion mechanism. This approach injects information about which categories should not be generated into the label generation paradigm, enhancing the interpretability of the diffusion model's robust inference. Unlike conventional label diffusion, our diffusion process concludes with an endpoint that includes the potential noise distribution of the sample (i.e., \\(\\mathbf{y}_T = \\mathbf{y}_n + \\epsilon\\)) and employs two independent coefficient schedules to control the diffusion speed of both channels." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.599, + 0.484, + 0.901 + ], + "angle": 0, + "content": "One challenge is that only noisy labels are available in practice for training, so neither \\(\\mathbf{y}_0\\) nor \\(\\mathbf{y}_n\\) can be obtained directly. Few researchers employed neighborhood detection techniques like KNN to re-estimate \\(\\mathbf{y}_0\\) [4]. However, under the IDN assumption, neighborhood labels of noisy samples often contain similar errors, making naive label estimation methods unreliable and potentially causing \\(\\mathbf{y}_0\\) to deviate from the ground truth (see Figure 3). To address these issues, we further propose a straightforward and efficient label pre-correction method that automatically provides diverse label information for diffusion model training. This framework employs feature extractors to identify the neighborhood label distribution of samples. Note that these extractors are either trained with self-supervised methods [7, 19, 65], which require few additional training iterations, or use pre-trained models without further training, thus remaining unaffected by label noise. Additionally, this pre-correction strategy allows the seamless integration of pre-trained large-scale models such as Sim-CLR [7] and vision transformer (ViT) [40] into our framework to further" + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.092, + 0.905, + 0.121 + ], + "angle": 0, + "content": "boost the performance. The main contributions of this work are summarized as follows:" + }, + { + "type": "text", + "bbox": [ + 0.514, + 0.122, + 0.905, + 0.182 + ], + "angle": 0, + "content": "- We propose a novel Directional Label Diffusion (DLD) model that explicitly introduces interpretable noise information through a dual diffusion architecture, enabling robust learning of label generation patterns from noisy data." + }, + { + "type": "text", + "bbox": [ + 0.515, + 0.183, + 0.905, + 0.272 + ], + "angle": 0, + "content": "- We design a fast label pre-correction method that automatically provides more accurate label information for the diffusion process and richer feature conditions for the generation process. By leveraging the consistency principle, this method pre-estimates the label distribution in a single iteration without incurring additional costs." + }, + { + "type": "text", + "bbox": [ + 0.515, + 0.273, + 0.905, + 0.317 + ], + "angle": 0, + "content": "- We refine the paradigm for introducing feature conditions with rigorous theoretical deduction and provide complementary training strategies to enhance robustness further." + }, + { + "type": "text", + "bbox": [ + 0.515, + 0.318, + 0.905, + 0.377 + ], + "angle": 0, + "content": "- We experimentally show that our model surpasses the state-of-the-art (SOTA) methods on both simulated and real-world noisy datasets, e.g., \\(6\\% \\sim 21\\%\\) accuracy improvement on the CIFAR-100 dataset with IDN." + }, + { + "type": "list", + "bbox": [ + 0.514, + 0.122, + 0.905, + 0.377 + ], + "angle": 0, + "content": null + }, + { + "type": "title", + "bbox": [ + 0.514, + 0.389, + 0.655, + 0.403 + ], + "angle": 0, + "content": "2. Related Work" + }, + { + "type": "title", + "bbox": [ + 0.514, + 0.41, + 0.766, + 0.426 + ], + "angle": 0, + "content": "2.1. Learning from Noisy Labels" + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.432, + 0.907, + 0.825 + ], + "angle": 0, + "content": "Screening and training strategies are central to LNL methods, aiming to distinguish noisy from clean samples and to optimize model learning under noisy conditions. Screening strategies, like Co-teaching [17], select low-loss samples as clean, while DivideMix [28] uses a Gaussian mixture model to automatically identify clean samples during training. Confidence-based methods like CC [62] and DISC [31] classify samples with unreliable predictions as noisy, whereas neighbor consistency regularization (NCR) [22] promotes learning from samples with high neighbor consistency. Our pre-correction method is based on similarity constraints across multi-view label distributions and leverages a feature extractor to provide high-quality information to the diffusion model without requiring additional iterative training. For training, models such as SEAL [6] iteratively correct noisy labels, but this often risks worsening outcomes [23]. Robust loss functions, like mean absolute error (MAE) [16] and generalized cross entropy (GCE) [61], enhance resilience to noisy labels, while DAC [45] and DMI [54] offer advanced mutual information-based losses. Approaches like DivideMix [28] treat LNL as semi-supervised learning, supported by methods like CC [62], C2D [63], and ELR [33] to refine the process. Without complex iterations or loss design, our method utilizes a lightweight diffusion network and a simple L2 loss to quickly learn the feature-label generation relationship." + }, + { + "type": "title", + "bbox": [ + 0.514, + 0.834, + 0.81, + 0.85 + ], + "angle": 0, + "content": "2.2. Label Denoising Diffusion Models" + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.856, + 0.908, + 0.901 + ], + "angle": 0, + "content": "Diffusion models were initially designed for image generation tasks and can be understood through the lenses of score matching and dynamics. Ho et al. [21] introduced a" + }, + { + "type": "page_number", + "bbox": [ + 0.48, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "25739" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.134, + 0.089, + 0.868, + 0.301 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.089, + 0.306, + 0.908, + 0.377 + ], + "angle": 0, + "content": "Figure 1. The proposed Directional Label Diffusion (DLD) model includes diagrams illustrating both the forward and reverse processes. In the forward process, diffusion is decoupled into two channels: random diffusion and directional diffusion. Random diffusion represents the transformation of \\(\\mathbf{y}_0\\) into random Gaussian noise \\(\\epsilon\\), while directional diffusion describes the transformation of \\(\\mathbf{y}_0\\) into a specific sample noise label distribution \\(\\mathbf{y}_n\\). Ultimately, \\(\\mathbf{y}_0\\) diffuses to the endpoint \\(\\mathbf{y}_T = \\mathbf{y}_0 + \\epsilon\\). Notably, the features of the samples serve as control conditions, guiding the generation of labels only during the reverse process." + }, + { + "type": "text", + "bbox": [ + 0.093, + 0.402, + 0.485, + 0.657 + ], + "angle": 0, + "content": "more classical interpretative framework known as denoising diffusion probabilistic models (DDPM), which have been widely applied across diverse domains. RDDM expands the denoising-based diffusion models, initially uninterpretable for image restoration, into a unified and interpretable model for both image generation and restoration by introducing residuals. Classification and regression diffusion models (CARD) [18] reconstruct the classification as a conditional generative task, generating labels from dimensionally reduced image features via a pre-trained encoder. Building on CARD, label-retrieval-augmented (LRA) diffusion model [4] re-estimates the generation targets through neighbor label retrieval, improving label accuracy. Our model disentangles the diffusion process into directional and random diffusion channels, enhancing diffusion quality through precorrection, which leads to train a more robust label diffusion model." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.666, + 0.218, + 0.684 + ], + "angle": 0, + "content": "3. Preliminary" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.688, + 0.483, + 0.795 + ], + "angle": 0, + "content": "In this section, we primarily introduce the label denoising diffusion models (LDDM), which is constructed based on the theoretical framework of the DDPM [21]. We start by defining the data distribution \\(\\mathbf{y}_0\\sim q(\\mathbf{y}_0)\\) and a Markovian forward (diffusion) process \\(q\\) , which progressively adds Gaussian noise \\(\\epsilon \\sim \\mathcal{N}(0,\\mathbf{I})\\) to the labels based on diffusion strength \\(\\beta_{t}\\) , diffusing label \\(\\mathbf{y}_0\\) to \\(\\mathbf{y}_T\\) .." + }, + { + "type": "equation", + "bbox": [ + 0.136, + 0.799, + 0.483, + 0.825 + ], + "angle": 0, + "content": "\\[\nq \\left(\\mathbf {y} _ {t} \\mid \\mathbf {y} _ {t - 1}\\right) = \\mathcal {N} \\left(\\mathbf {y} _ {t}; \\sqrt {1 - \\beta_ {t}} \\mathbf {y} _ {t - 1}, \\beta_ {t} \\mathbf {I}\\right), \\tag {1}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.828, + 0.484, + 0.876 + ], + "angle": 0, + "content": "Under Bayes' theorem, one finds that the posterior distribution of the forward process can be represented as a Gaussian distribution with mean \\(\\tilde{\\mu}_t(\\mathbf{y}_t,\\mathbf{y}_0)\\) and variance \\(\\tilde{\\beta}_t\\) :" + }, + { + "type": "equation", + "bbox": [ + 0.115, + 0.879, + 0.483, + 0.906 + ], + "angle": 0, + "content": "\\[\nq \\left(\\mathbf {y} _ {t - 1} \\mid \\mathbf {y} _ {t}, \\mathbf {y} _ {0}\\right) = \\mathcal {N} \\left(\\mathbf {y} _ {t - 1}; \\tilde {\\mu} _ {t} \\left(\\mathbf {y} _ {t}, \\mathbf {y} _ {0}\\right), \\tilde {\\beta} _ {t} \\mathbf {I}\\right), \\tag {2}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.399, + 0.907, + 0.527 + ], + "angle": 0, + "content": "where \\(\\tilde{\\mu}_t(\\mathbf{y}_t,\\mathbf{y}_0) = \\frac{\\sqrt{\\bar{\\alpha}_{t - 1}}\\beta_t}{1 - \\bar{\\alpha}_t}\\mathbf{y}_0 + \\frac{\\sqrt{\\alpha_t(1 - \\bar{\\alpha}_{t - 1})}}{1 - \\bar{\\alpha}_t}\\mathbf{y}_t\\) and \\(\\tilde{\\beta}_t = \\frac{1 - \\bar{\\alpha}_{t - 1}}{1 - \\bar{\\alpha}_t}\\beta_t\\), In the reverse process, we naturally employ a neural network to perform Markovian inference from \\(q(\\mathbf{y}_T)\\), fitting the posterior distribution \\(q(\\mathbf{y}_{t - 1}|\\mathbf{y}_t)\\) through to \\(\\mathbf{y}_0\\). Since the variance of the posterior Gaussian distribution depends only on the diffusion strength corresponding to time \\(t\\), the neural network focuses primarily on predicting and learning the mean:" + }, + { + "type": "equation", + "bbox": [ + 0.557, + 0.538, + 0.907, + 0.565 + ], + "angle": 0, + "content": "\\[\np _ {\\theta} \\left(\\mathbf {y} _ {t - 1} \\mid \\mathbf {y} _ {t}\\right) := \\mathcal {N} \\left(\\mathbf {y} _ {t - 1}; \\mu_ {\\theta} \\left(\\mathbf {y} _ {t}, t\\right), \\tilde {\\beta} _ {t} \\mathbf {I}\\right). \\tag {3}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.574, + 0.906, + 0.679 + ], + "angle": 0, + "content": "To train this model such that \\( p_{\\theta}(\\mathbf{y}_0) \\) closely approximates \\( q(\\mathbf{y}_0) \\), we can optimize the variational lower bound (VLB), however, research [7] found that directly predicting the mean was not effective, so they reparameterized the estimation target, constraining the neural network to estimate the actual noise of the diffusion process and simplified the loss function:" + }, + { + "type": "equation", + "bbox": [ + 0.515, + 0.687, + 0.905, + 0.725 + ], + "angle": 0, + "content": "\\[\nL _ {\\text {s i m p l e}} := E _ {t \\sim [ 1, T ], \\mathbf {y} _ {0} \\sim q (\\mathbf {y} _ {0}), \\epsilon \\sim \\mathcal {N} (0, \\mathbf {I})} \\left[ \\| \\epsilon - \\epsilon_ {\\theta} (\\mathbf {y} _ {t}, t) \\| ^ {2} \\right]. \\tag {4}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.725, + 0.906, + 0.787 + ], + "angle": 0, + "content": "During inference, we utilize the label noise estimated by the model, restoring the mean of the label at time \\( t - 1 \\) to retrieve its distribution form, and then gradually performing random sampling until generating the label \\( \\mathbf{y}_0 \\):" + }, + { + "type": "equation", + "bbox": [ + 0.529, + 0.796, + 0.907, + 0.831 + ], + "angle": 0, + "content": "\\[\n\\mathbf {y} _ {t - 1} = \\frac {1}{\\sqrt {\\alpha_ {t}}} \\left(\\mathbf {y} _ {t} - \\frac {1 - \\alpha_ {t}}{\\sqrt {1 - \\bar {\\alpha} _ {t}}} \\epsilon_ {\\theta} (\\mathbf {y} _ {t}, t)\\right) + \\sqrt {\\tilde {\\beta} _ {t}} \\mathbf {z}, \\tag {5}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.84, + 0.907, + 0.901 + ], + "angle": 0, + "content": "where \\(\\mathbf{z} \\sim \\mathcal{N}(0, \\mathbf{I})\\). Unfortunately, this label generation process lacks any conditional information, resulting in random label generation for images with identical features, which fails to achieve classification. To address this, we" + }, + { + "type": "page_number", + "bbox": [ + 0.479, + 0.945, + 0.521, + 0.958 + ], + "angle": 0, + "content": "25740" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.09, + 0.092, + 0.486, + 0.138 + ], + "angle": 0, + "content": "analyze the foundational paradigm of conditional label diffusion models in Appendix A and simplify the method of feature conditioning." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.149, + 0.405, + 0.166 + ], + "angle": 0, + "content": "4. Directional Label Diffusion Model" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.171, + 0.484, + 0.292 + ], + "angle": 0, + "content": "Based on LDDM, our goal is to develop a directional diffusion process on labels to introduce more knowledge about noise. We modify the representation of \\(\\mathbf{y}_T = \\epsilon\\) in LDDM to \\(\\mathbf{y}_T = \\mathbf{y}_n + \\epsilon\\) in our DLD, where \\(\\mathbf{y}_n\\) is an estimation of noisy label distribution. The improved forward process is no longer a simple random noise diffusion but is disentangled into two paths of directional diffusion and random diffusion, as shown in Figure 1." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.293, + 0.484, + 0.429 + ], + "angle": 0, + "content": "In the following subsections, we first outline the overall framework of our model in Section 4.1. Next, in Section 4.2, we present the foundational theory of the DLD diffusion process and analyze the method of conditional control. In Section 4.3, we describe the reverse process and training method, including optimizations for introducing feature conditions and an adapted accelerated inference method. Finally, in Section 4.4, we describe the rapid label pre-correction method within the DLD framework." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.44, + 0.255, + 0.455 + ], + "angle": 0, + "content": "4.1. Model Overview" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.462, + 0.485, + 0.735 + ], + "angle": 0, + "content": "We propose the Directional Label Diffusion (DLD) framework, with the overall process illustrated in Figure 2. The framework contains two main stages: label pre-correction (Stage 1) and the directional label diffusion model (Stage 2). In stage 1, we first use two different data augmentation methods to provide two views for each sample, then use a pre-trained feature extractor \\( f_{ext} \\) to observe the neighborhood label distribution \\( p_w \\) and \\( p_s \\) of the same sample in two feature spaces. Based on similarity constraints and the principle of neighborhood consistency, we divide the dataset into clean, noisy, and hard sets. Then, the label precorrector and noise distribution estimator prepare \\( \\mathbf{y}_0 \\) and \\( \\mathbf{y}_n \\) based on the sample partitioning results, using \\( \\mathbf{y}_0 \\) as the initial diffusion label and \\( \\mathbf{y}_n \\) as part of the diffusion endpoint for the next stage of model training. Otherwise, we introduce both augmented image features and high-quality features extracted by \\( f_{ext} \\) into the training process to jointly guide the learning of the robust diffusion model." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.746, + 0.32, + 0.76 + ], + "angle": 0, + "content": "4.2. Forward Process of DLD" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.768, + 0.485, + 0.814 + ], + "angle": 0, + "content": "To model the random diffusion of labels and the addition of directional noise information, we define the single forward (diffusion) process step in our DLD as follows:" + }, + { + "type": "equation", + "bbox": [ + 0.199, + 0.828, + 0.483, + 0.843 + ], + "angle": 0, + "content": "\\[\n\\mathbf {y} _ {t} = \\mathbf {y} _ {t - 1} + \\alpha_ {t} \\mathbf {y} _ {d} + \\beta_ {t} \\epsilon , \\tag {6}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.856, + 0.484, + 0.902 + ], + "angle": 0, + "content": "where \\(\\mathbf{y}_d\\) is the deviation between \\(\\mathbf{y}_n\\) and \\(y_0\\) (i.e., \\(\\mathbf{y}_d = \\mathbf{y}_n - \\mathbf{y}_0\\)), representing the directional shift contained in the diffusion from \\(\\mathbf{y}_{t-1}\\) to \\(\\mathbf{y}_t\\). Two independent coefficients, \\(\\alpha_t\\)" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.092, + 0.906, + 0.168 + ], + "angle": 0, + "content": "and \\(\\beta_{t}\\) control the degrees of directional and random diffusion, respectively. Based on the additivity of Gaussian noise, we derive the jump diffusion formula from \\(\\mathbf{y}_0\\) to \\(\\mathbf{y}_t\\) without requiring single-step continuous diffusion (see Appendix B for the proof):" + }, + { + "type": "equation", + "bbox": [ + 0.629, + 0.178, + 0.907, + 0.196 + ], + "angle": 0, + "content": "\\[\n\\mathbf {y} _ {t} = \\mathbf {y} _ {0} + \\bar {\\alpha} _ {t} \\mathbf {y} _ {d} + \\bar {\\beta} _ {t} \\epsilon , \\tag {7}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.208, + 0.907, + 0.365 + ], + "angle": 0, + "content": "where \\(\\bar{\\alpha}_t = \\sum_{i=1}^t \\alpha_i\\) and \\(\\bar{\\beta}_t = \\sqrt{\\sum_{i=1}^t \\beta_i^2}\\). When \\(t = T\\), we hope \\(\\mathbf{y}_T = \\mathbf{y}_n + \\epsilon = \\mathbf{y}_0 + \\mathbf{y}_d + \\epsilon\\), so we usually set \\(\\bar{\\alpha}_T = \\bar{\\beta}_T = 1\\). This implies that the two coefficients \\(\\alpha_t\\) and \\(\\beta_t\\) are not arbitrary transformations; rather, they should satisfy a specific sum constraint. Then, we redefine the conditional Markov diffusion process similar to the directional label forward distribution \\(q\\) without condition (detailed theoretical framework in Appendix C), assuming that when conditioned on features \\(\\mathbf{x}\\), the prior distribution of \\(\\hat{q}\\) is the same as \\(q\\):" + }, + { + "type": "equation", + "bbox": [ + 0.537, + 0.375, + 0.905, + 0.414 + ], + "angle": 0, + "content": "\\[\n\\begin{array}{l} \\hat {q} \\left(\\mathbf {y} _ {t} \\mid \\mathbf {y} _ {t - 1}, \\mathbf {y} _ {d}, \\mathbf {x}\\right) := q \\left(\\mathbf {y} _ {t} \\mid \\mathbf {y} _ {t - 1}, \\mathbf {y} _ {d}\\right) \\tag {8} \\\\ = \\mathcal {N} \\left(\\mathbf {y} _ {t}; \\mathbf {y} _ {t - 1} + \\alpha_ {t} \\mathbf {y} _ {d}, \\beta_ {t} ^ {2} \\mathbf {I}\\right), \\\\ \\end{array}\n\\]" + }, + { + "type": "equation", + "bbox": [ + 0.54, + 0.439, + 0.906, + 0.501 + ], + "angle": 0, + "content": "\\[\n\\begin{array}{l} \\hat {q} \\left(\\mathbf {y} _ {1: T} \\mid \\mathbf {y} _ {0}, \\mathbf {y} _ {d}, \\mathbf {x}\\right) := \\prod_ {t = 1} ^ {T} \\hat {q} \\left(\\mathbf {y} _ {t} \\mid \\mathbf {y} _ {t - 1}, \\mathbf {y} _ {d}, \\mathbf {x}\\right) \\tag {9} \\\\ = \\mathcal {N} \\left(\\mathbf {y} _ {t}; \\mathbf {y} _ {0} + \\bar {\\alpha} _ {t} \\mathbf {y} _ {d}, \\bar {\\beta} _ {t} ^ {2} \\mathbf {I}\\right), \\\\ \\end{array}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.51, + 0.906, + 0.54 + ], + "angle": 0, + "content": "which means that the diffusion process of the labels is not influenced by any image features." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.541, + 0.906, + 0.585 + ], + "angle": 0, + "content": "In Appendix C, using Bayes' rule and Taylor expansion, we derive the conditional posterior distribution during the forward process:" + }, + { + "type": "equation", + "bbox": [ + 0.545, + 0.599, + 0.905, + 0.617 + ], + "angle": 0, + "content": "\\[\n\\hat {q} \\left(\\mathbf {y} _ {t - 1} \\mid \\mathbf {y} _ {t}, \\mathbf {y} _ {0}, \\mathbf {y} _ {d}, \\mathbf {x}\\right) = \\mathcal {N} \\left(\\mathbf {y} _ {t - 1}; \\hat {\\mu} _ {t}, \\hat {\\sigma} _ {t} \\mathbf {I}\\right), \\tag {10}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.628, + 0.906, + 0.734 + ], + "angle": 0, + "content": "where \\(\\hat{\\mu}_t = y_t - \\alpha_t y_d - \\frac{\\beta_t^2}{\\beta_t} \\epsilon + \\hat{\\sigma}_t \\nabla_{y_{t-1}} \\log \\hat{q}(\\mathbf{x} | \\mathbf{y}_{t-1}, \\mathbf{y}_d)\\) and \\(\\hat{\\sigma}_t = \\frac{\\beta_t^2 \\bar{\\beta}_{t-1}^2}{\\beta_t^2}\\). The last term in \\(\\hat{\\mu}_t\\) can be considered a deviation to the unconditional distribution mean, guided by the first derivative of \\(\\hat{q}(\\mathbf{x} | \\mathbf{y}_{t-1}, \\mathbf{y}_d)\\), indicating that the decoding information from \\(\\mathbf{y}_{t-1}\\) to \\(\\mathbf{x}\\) should be introduced in training." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.742, + 0.84, + 0.759 + ], + "angle": 0, + "content": "4.3. Reverse and Training Process of DLD" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.764, + 0.906, + 0.87 + ], + "angle": 0, + "content": "Reverse Process and Training. In the reverse (generation) process, as indicated by Eq. 6, we need to simultaneously predict \\(\\mathbf{y}_d\\) and \\(\\epsilon\\) to progressively recover \\(\\mathbf{y}_T\\) to \\(\\mathbf{y}_0\\). Therefore, we can train two independent networks to estimate \\(\\mathbf{y}_d^{\\theta_1}\\) and \\(\\epsilon_{\\theta_2}\\), respectively, to compute the reverse posterior distribution. Furthermore, we directly use feature information as input for two networks:" + }, + { + "type": "equation", + "bbox": [ + 0.573, + 0.882, + 0.905, + 0.901 + ], + "angle": 0, + "content": "\\[\np _ {\\theta} \\left(\\mathbf {y} _ {t - 1} \\mid \\mathbf {y} _ {t}, \\mathbf {x}\\right) := \\mathcal {N} \\left(\\mathbf {y} _ {t - 1}; \\mu_ {\\theta}, \\hat {\\sigma} _ {t} ^ {2} \\mathbf {I}\\right). \\tag {11}\n\\]" + }, + { + "type": "page_number", + "bbox": [ + 0.479, + 0.945, + 0.518, + 0.957 + ], + "angle": 0, + "content": "25741" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.171, + 0.093, + 0.825, + 0.375 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.089, + 0.387, + 0.908, + 0.43 + ], + "angle": 0, + "content": "Figure 2. Overview of the DLD framework. During Stage 1, the noisy dataset is quickly divided into three subsets. Then, we estimate \\(\\mathbf{y}_0\\) and \\(\\mathbf{y}_n\\) for each samples based on Eq. 14 and Eq. 15. In Stage 2, the label information and augmented features from Stage 1 are jointly used to train a directional diffusion model (blue bars represent random diffusion and gray bars represent directional diffusion)." + }, + { + "type": "text", + "bbox": [ + 0.089, + 0.455, + 0.484, + 0.546 + ], + "angle": 0, + "content": "Based on the analysis of the forward process, we adopt a straightforward approach which naturally integrates the decoding process into the diffusion model architecture by a fixed decoding layer \\( L_{1} \\) to map \\( \\mathbf{y} \\) into the feature space of \\( \\mathbf{x} \\). Detailed information about the specific diffusion model network structure is shown in Appendix Figure C.1." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.547, + 0.484, + 0.577 + ], + "angle": 0, + "content": "By optimizing the VLB and reparameterization [21], we derive the following simplified loss function for training:" + }, + { + "type": "equation", + "bbox": [ + 0.17, + 0.588, + 0.483, + 0.61 + ], + "angle": 0, + "content": "\\[\n\\mathcal {L} _ {d} = \\left\\| \\mathbf {y} _ {d} - \\mathbf {y} _ {d} ^ {\\theta_ {1}} \\left(\\mathbf {y} _ {t}, \\mathbf {y} _ {n}, \\mathbf {x}, t\\right) \\right\\| ^ {2}, \\tag {12}\n\\]" + }, + { + "type": "equation", + "bbox": [ + 0.178, + 0.634, + 0.483, + 0.653 + ], + "angle": 0, + "content": "\\[\n\\mathcal {L} _ {\\epsilon} = \\left\\| \\epsilon - \\epsilon_ {\\theta_ {2}} \\left(\\mathbf {y} _ {t}, \\mathbf {y} _ {n}, \\mathbf {x}, t\\right) \\right\\| ^ {2}, \\tag {13}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.659, + 0.483, + 0.689 + ], + "angle": 0, + "content": "where \\(\\epsilon \\sim \\mathcal{N}(0,\\mathbf{I})\\), and the training input \\(\\mathbf{y}_t\\) is synthesized using Eq. 7. Algorithm 1 describes the training procedure." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.689, + 0.484, + 0.901 + ], + "angle": 0, + "content": "Inference. Since the classification diffusion model is deterministic in inference and the dimensionality of the label vector is much lower than that of images, we mimic the inference process of DDIM [44] to generate labels in as few steps as possible without affecting the generated categories. We can gradually sample labels step-by-step following \\(\\mathbf{y}_{t-1} = \\mathbf{y}_t - (\\bar{\\alpha}_t - \\bar{\\alpha}_{t-1})\\mathbf{y}_d^{\\theta_1} - (\\bar{\\beta}_t - \\bar{\\beta}_{t-1})\\epsilon_{\\theta_2}\\), where the transition from \\(t-1\\) to \\(t\\) is discrete rather than continuous in time. Appendix D provides a detailed explanation of the de-Markovization process, the accelerated sampling algorithm and the visual demonstration of the sampling process. In experiments, we use 5 steps for DLD's inference, which is on the same order of magnitude as the speed of traditional classification models." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.455, + 0.776, + 0.47 + ], + "angle": 0, + "content": "4.4. Label Pre-correction Method" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.48, + 0.907, + 0.585 + ], + "angle": 0, + "content": "Although diffusion models are powerful, under the guidance of incorrect \\(\\mathbf{y}_0\\) and \\(\\mathbf{y}_n\\), they can easily fit the wrong conditional distributions. Therefore, we propose a one-step label pre-correction method that automatically provides accurate label information and feature conditions for the DLD model, requiring no additional parameter training, making it simple, straightforward, and efficient." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.588, + 0.909, + 0.903 + ], + "angle": 0, + "content": "Label Pre-correction for \\(\\mathbf{y}_0\\). Inspired by the concepts of multi-view[42, 53] and neighbor consistency regularization [22], our primary assumption is that the same sample should have similar neighborhood label distributions in different feature spaces. We define the original noisy dataset as \\(\\mathcal{D} = \\{(x_i,\\tilde{y}_i)\\mid x_i\\in \\mathbb{R}^d,\\tilde{y}_i\\in \\mathbb{R}^m,i = 1,\\ldots ,n\\}\\). To obtain different feature spaces by a shared feature extractor \\(f_{ext}\\), we first apply two different data augmentation methods and map them into \\(\\mathbb{R}^w\\) and \\(\\mathbb{R}^s\\), with the augmented datasets respectively denoted as \\(\\mathcal{D}^w\\) and \\(\\mathcal{D}^s\\). Then, we use Euclidean or Cosine similarity to obtain the nearest \\(K\\) neighborhood labels \\(y_{k = 1}^{K}\\) and feature space distances \\(v_{k = 1}^{K}\\) then we can get two weighted label distributions \\(p_w\\) and \\(p_s\\) by \\(p(y_{i}) = \\left(\\sum_{k = 1}^{K}w_{k}\\cdot \\mathbf{1}(y_{k} = y_{i})\\right)\\cdot \\left(\\sum_{k = 1}^{K}w_{k}\\right)^{-1}\\) where \\(w_{k} = \\frac{1}{v_{k} + \\delta}\\), where \\(\\delta\\) is a small constant to prevent division by zero. Next, we calculate the divergence between the two distributions using the Kullback-Leibler divergence and a via Gaussian mixture model (GMM) [38] to re-partitioning three types of samples: Samples with high means are reassigned to the hard set \\(\\mathcal{D}_H\\), which represents" + }, + { + "type": "page_number", + "bbox": [ + 0.479, + 0.945, + 0.521, + 0.958 + ], + "angle": 0, + "content": "25742" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.09, + 0.092, + 0.482, + 0.227 + ], + "angle": 0, + "content": "samples whose correctness is hard to determine and typically lie at the boundaries of classes. Samples with low means indicate that similar neighborhood distributions under both views. Using the principle of neighborhood consistency, we compare \\( \\text{argmax}(p_{ws}) \\) with the original label \\( \\tilde{y} \\), where \\( p_{ws} = \\frac{p_w + p_s}{2} \\). If the class with the highest frequency in the average neighborhood distribution matches the original label, then the sample is assigned to the clean set \\( \\mathcal{D}_C \\); otherwise, it is assigned to the noisy set \\( \\mathcal{D}_N \\)." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.228, + 0.482, + 0.289 + ], + "angle": 0, + "content": "Based on the repartitioned sample subsets and the original noisy labels \\(\\tilde{y}\\), we define a label pre-corrector, which functions similarly to a transition matrix. It can automatically perform label flipping or refinement:" + }, + { + "type": "equation", + "bbox": [ + 0.11, + 0.299, + 0.483, + 0.348 + ], + "angle": 0, + "content": "\\[\nP \\left(y _ {0} \\mid \\tilde {y}, x\\right) = \\left\\{ \\begin{array}{l l} \\tilde {y} & \\text {i f} x \\in \\mathcal {D} _ {C}, \\\\ \\operatorname {a r g m a x} \\left(p _ {w s}\\right) & \\text {i f} x \\in \\mathcal {D} _ {N}, \\\\ p _ {w s} & \\text {i f} x \\in \\mathcal {D} _ {H}. \\end{array} \\right. \\tag {14}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.358, + 0.483, + 0.493 + ], + "angle": 0, + "content": "According to Eq. 14, for samples with deterministic neighborhood label distributions, we use hard labels to minimize the interference of redundant information on \\( y_0 \\). For samples with more contentious neighborhood label distributions, we adopt a soft strategy for label refinement. Instead of point estimation, we estimate through a label vector, retaining most of the neighborhood information. This strategy enhances the accuracy of the \\( y_0 \\) while preserving as much robust information as possible." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.495, + 0.483, + 0.722 + ], + "angle": 0, + "content": "Estimate Noise distribution for \\(\\mathbf{y}_n\\). Similar to obtaining \\(\\mathbf{y}_0\\), we use the subset of divided samples to estimate the noise distribution of the samples. According to Eq. 15, when a sample is identified as clean, the directed diffusion term is used to gradually fade out \\(\\mathbf{y}_0\\) (\\(\\mathbf{y}_d = -\\mathbf{y}_0\\)), allowing the model to learn to recover a label closer to the ground truth from an uninformed \\(\\epsilon\\). When the sample contains noise, the endpoint of diffusion should include a clear representation of noise, enabling the model to strictly avoid recovering such noisy vector values during reverse generation. For uncertain samples, we simulate potential noise distributions by retaining the difference in label distributions between two views. The greater the difference in the distribution values for a particular class in the two views, the more likely the sample is incorrectly labeled as that class." + }, + { + "type": "equation", + "bbox": [ + 0.124, + 0.733, + 0.483, + 0.786 + ], + "angle": 0, + "content": "\\[\nP \\left(y _ {n} \\mid \\tilde {y}, x\\right) = \\left\\{ \\begin{array}{l l} 0 & \\text {i f} x \\in \\mathcal {D} _ {C}, \\\\ \\tilde {y} & \\text {i f} x \\in \\mathcal {D} _ {N}, \\\\ \\frac {\\left| p _ {w} - p _ {s} \\right|}{\\sum \\left| p _ {w} - p _ {s} \\right|} & \\text {i f} x \\in \\mathcal {D} _ {H}. \\end{array} \\right. \\tag {15}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.795, + 0.483, + 0.901 + ], + "angle": 0, + "content": "The complete training process is outlined in Algorithm 1, consisting of the following phases: Steps \\(1 \\sim 3\\) for sample re-partitioning, Steps \\(4 \\sim 9\\) for diffusion model training, with Step 6 dedicated to \\(\\mathbf{y}_0\\) pre-correction and \\(\\mathbf{y}_n\\) estimation. Notably, in Step 5, instead of sampling features from the original noisy set \\(\\mathcal{D}\\), we use a mixture of sample features from \\(\\mathcal{D}^w\\) and \\(\\mathcal{D}^s\\) as the training input for the diffusion" + }, + { + "type": "title", + "bbox": [ + 0.516, + 0.092, + 0.702, + 0.107 + ], + "angle": 0, + "content": "Algorithm 1 DLD Training" + }, + { + "type": "text", + "bbox": [ + 0.514, + 0.111, + 0.905, + 0.141 + ], + "angle": 0, + "content": "Input: Noisy training set \\(\\mathcal{D} = \\{\\mathbf{X},\\tilde{\\mathbf{Y}}\\}\\), feature extractor \\(f_{\\mathrm{ext}}\\), two network parameters \\(\\theta_{1}\\) and \\(\\theta_{2}\\)" + }, + { + "type": "text", + "bbox": [ + 0.516, + 0.142, + 0.641, + 0.156 + ], + "angle": 0, + "content": "Output: \\(\\theta_{1}\\) and \\(\\theta_{2}\\)" + }, + { + "type": "text", + "bbox": [ + 0.522, + 0.159, + 0.905, + 0.189 + ], + "angle": 0, + "content": "1: Obtain two augmented features \\(\\mathcal{D}^w\\) and \\(\\mathcal{D}^s\\) using the feature extractor \\(f_{\\mathrm{ext}}\\)." + }, + { + "type": "text", + "bbox": [ + 0.522, + 0.19, + 0.871, + 0.204 + ], + "angle": 0, + "content": "2: Calculate K-nearest label distributions \\( p_w \\) and \\( p_s \\)." + }, + { + "type": "text", + "bbox": [ + 0.522, + 0.205, + 0.853, + 0.219 + ], + "angle": 0, + "content": "3: Re-partitioning samples into \\(\\mathcal{D}_C, \\mathcal{D}_N\\) and \\(\\mathcal{D}_H\\)." + }, + { + "type": "text", + "bbox": [ + 0.522, + 0.22, + 0.702, + 0.234 + ], + "angle": 0, + "content": "4: while not converged do" + }, + { + "type": "text", + "bbox": [ + 0.522, + 0.235, + 0.905, + 0.265 + ], + "angle": 0, + "content": "5: Sample data \\((\\mathbf{x}_{ws},\\tilde{\\mathbf{y}})\\sim \\{\\mathcal{D}^w,\\mathcal{D}^s\\}\\) ; time slice \\(t\\sim\\) \\(\\{1,\\dots ,T\\}\\) ; and noise \\(\\epsilon \\sim \\mathcal{N}(0,\\mathbf{I})\\)" + }, + { + "type": "text", + "bbox": [ + 0.522, + 0.266, + 0.905, + 0.295 + ], + "angle": 0, + "content": "6: Pre-estimate \\(\\mathbf{y}_0\\) and \\(\\mathbf{y}_n\\) based on Eq. 14 and Eq. 15, respectively." + }, + { + "type": "text", + "bbox": [ + 0.522, + 0.296, + 0.866, + 0.31 + ], + "angle": 0, + "content": "7: Accumulate losses through Eq. 12 and Eq. 13." + }, + { + "type": "text", + "bbox": [ + 0.522, + 0.311, + 0.905, + 0.34 + ], + "angle": 0, + "content": "8: Take gradient descent step on the losses to update network parameters." + }, + { + "type": "text", + "bbox": [ + 0.522, + 0.341, + 0.614, + 0.354 + ], + "angle": 0, + "content": "9: end while" + }, + { + "type": "list", + "bbox": [ + 0.522, + 0.159, + 0.905, + 0.354 + ], + "angle": 0, + "content": null + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.381, + 0.907, + 0.491 + ], + "angle": 0, + "content": "model, where \\(\\mathbf{x}_{ws} = (1 - \\gamma)\\mathbf{x}_w + \\gamma \\mathbf{x}_s\\), with \\(\\gamma = \\frac{|\\mathcal{D}_H|}{|\\mathcal{D}|}\\) representing the proportion of uncertain samples, which means that the more difficult the sample classification, the richer the feature conditions the model automatically introduces. This reflects an implicit regularization method that mitigates the issue of the diffusion model being overly confident when still learning incorrect information." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.501, + 0.831, + 0.518 + ], + "angle": 0, + "content": "5. Experimental Results and Analysis" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.522, + 0.907, + 0.644 + ], + "angle": 0, + "content": "In this section, we first evaluate the DLD's effectiveness on CIFAR datasets with varying levels of IDN in Section 5.1. Then, We conduct ablation studies in Section 5.2 to analyze the contributions of the pre-trained feature extractor, the label pre-correction technology, and other components. In Section 5.3, we compare the DLD to the SOTA methods across four real-world noisy datasets. Appendix E provides more details about our experiments." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.653, + 0.829, + 0.669 + ], + "angle": 0, + "content": "5.1. Results on Simulated Noisy Datasets" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.674, + 0.907, + 0.796 + ], + "angle": 0, + "content": "We conduct experiments on the CIFAR-10 and CIFAR-100 datasets [27] with simulated IDN [6]. Details of the IDN generation algorithm and examples are provided in Appendix E.1. In our experiments, we generate noise levels of \\(20\\%\\), \\(40\\%\\), and \\(60\\%\\). To obtain the feature space for label pre-correction, we provide multiple configurations of feature extractors, including DLD+ResNet, DLD+SimCLR and DLD+ViT (see Appendix F.2 for detailed settings)." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.796, + 0.907, + 0.901 + ], + "angle": 0, + "content": "As shown in Table 1, the DLD+ResNet configuration outperforms the baseline across most noise levels on CIFAR, especially in high-noise environments. We analyze that the main factor affecting its performance in low noise environment is the introduction of too much robust information, which leads to the interference of the model. When using the DLD+ViT configuration, its performance surpasses" + }, + { + "type": "page_number", + "bbox": [ + 0.479, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "25743" + } + ], + [ + { + "type": "table_caption", + "bbox": [ + 0.24, + 0.089, + 0.756, + 0.103 + ], + "angle": 0, + "content": "Table 1. Classification accuracies (%) of CIFAR datasets under different levels of IDN." + }, + { + "type": "table", + "bbox": [ + 0.171, + 0.108, + 0.835, + 0.346 + ], + "angle": 0, + "content": "
MethodCIFAR-10CIFAR-100
20%40%60%20%40%60%
Co-teaching [17]87.28±0.2078.72±0.4762.51±1.9869.26±0.3959.53±0.5746.80±0.51
GCE [61]86.44±0.2376.71±0.3960.76±3.0868.24±0.2755.76±1.1846.75±2.46
DAC [45]86.16±0.1374.80±0.3259.97±0.9068.70±0.7855.71±1.0146.29±3.35
DMI [54]86.57±0.1677.81±0.8569.94±1.3168.71±1.2157.99±0.9250.96±2.08
SEAL [6]87.79±0.0982.98±0.0570.07±1.5470.61±0.1961.18±0.3551.86±1.32
NCR [22]87.74±0.1583.03±0.2169.44±1.0170.68±0.3761.93±0.5449.13±2.77
CC+DivideMix [62]88.36±0.1184.18±0.4071.52±1.7370.09±0.1964.35±0.3553.20±0.66
DISC [31]90.68±0.1085.61±0.3273.59±1.1371.49±0.1364.46±0.2953.62±0.51
LRA-diffusion+ResNet [4]89.94±0.1586.73±0.3774.66±1.0272.35±0.1465.55±0.3054.43±0.53
DLD+ResNet (Ours)90.43±0.1287.17±0.2976.07±1.1072.60±0.1566.59±0.3155.72±0.24
EPL+ViT [26]95.95±0.1793.01±0.3376.16±1.1872.24±0.1267.20±0.2055.59±0.13
LRA-diffusion+ViT [4]96.65±0.1893.68±0.3478.12±1.1773.59±0.1367.67±0.2157.21±0.12
DLD+ViT (Ours)96.98±0.1296.49±0.0783.14±0.8178.62±0.1176.03±0.1469.57±0.17
" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.369, + 0.482, + 0.476 + ], + "angle": 0, + "content": "EPL+ViT and LRA-diffusion+ViT. This implies that DLD more effectively utilizes the provided prior knowledge without requiring additional training conditions. Across various datasets, DLD's superiority is more pronounced on CIFAR-100, indicating that the more complex IDN on CIFAR-100 severely impacts the performance of other methods, whereas our model maintains robust learning." + }, + { + "type": "text", + "bbox": [ + 0.089, + 0.476, + 0.483, + 0.629 + ], + "angle": 0, + "content": "Additionally, we conducted experiments on the simulated CCN dataset. The results (see Appendix E.2) show that CCN is easier to handle than IDN, which aligns with our expectations. Compared to other methods, DLD still achieves the highest classification accuracy. Due to the lightweight architecture of the diffusion model and the single-step pre-correction process, DLD outperforms traditional LNL methods without incurring significant computational overhead. A detailed analysis of training time overhead and model parameters is provided in Appendix E.3." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.637, + 0.255, + 0.652 + ], + "angle": 0, + "content": "5.2. Ablation Studies" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.659, + 0.483, + 0.885 + ], + "angle": 0, + "content": "Power of label pre-correction. We validate the effectiveness of label pre-correction by freezing the diffusion model. As shown in Figure 3, our pre-correction method consistently outperforms in various noisy environments, while Figure 3b demonstrates that the LRA estimation method disrupts originally correct labels. Then, we unfroze the diffusion model to verify that DLD does not fully depend on the pre-correction method. Results show that DLD further enhances generalization performance based on pre-correction, indicating that the two methods complement each other effectively. Additionally, we design more comprehensive cross-ablation experiments, e.g., replacing the LRA technique in LRA-diffusion with the pre-correction module to demonstrate the necessity of directional diffusion. See Appendix F.1 for more details." + }, + { + "type": "text", + "bbox": [ + 0.091, + 0.886, + 0.483, + 0.901 + ], + "angle": 0, + "content": "Contribution of different pre-trained features and other" + }, + { + "type": "table_caption", + "bbox": [ + 0.547, + 0.368, + 0.872, + 0.382 + ], + "angle": 0, + "content": "Table 2. Classification accuracies (%) on Animal-10N." + }, + { + "type": "table", + "bbox": [ + 0.552, + 0.393, + 0.866, + 0.535 + ], + "angle": 0, + "content": "
MethodAccuracy
CE [35]80.6
GCE [61]81.5
NCT [8]84.1
DISC [31]87.3
SSR [14]88.5
LRA-diffusion [4]88.6
J-ViT [9]89.0
DLD (Ours)89.4
" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.563, + 0.907, + 0.789 + ], + "angle": 0, + "content": "training components. We explore the impact of various feature extractors, including ResNet34 [20], SimCLR [7] and ViT-L/14 [40], and conduct ablation studies comparing our method with CARD and LRA-diffusion. Here, DISC [31] served as a baseline that does not require any pretrained models. Detailed results and analyses are provided in Appendix F.2. Our findings reveal that DLD outperforms other diffusion models when combined with different pretrained models, achieving optimal performance when integrated with ViT-L/14. This underscores ViT-L/14's effectiveness in enhancing label pre-correction and facilitating diffusion model learning. Furthermore, Appendix F.3 details additional ablation studies on various hyperparameter settings. We validate the optimal training strategy for the DLD method and recommend this combination." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.802, + 0.837, + 0.818 + ], + "angle": 0, + "content": "5.3. Results on Real-world Noisy Datasets" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.825, + 0.906, + 0.901 + ], + "angle": 0, + "content": "We select four real-world noisy datasets for our experiments to demonstrate the superior performance of DLD: Animal10N [43], Clothing1M [52], WebVision [30], and ImageNet ILSVRC2012 [41]. Comprehensive descriptions of these datasets and detailed experimental setups can be" + }, + { + "type": "page_number", + "bbox": [ + 0.48, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "25744" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.132, + 0.089, + 0.497, + 0.287 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.182, + 0.291, + 0.443, + 0.304 + ], + "angle": 0, + "content": "(a) Experimental results on the CIFAR-10 dataset" + }, + { + "type": "image", + "bbox": [ + 0.501, + 0.09, + 0.865, + 0.288 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.547, + 0.291, + 0.816, + 0.304 + ], + "angle": 0, + "content": "(b) Experimental results on the CIFAR-100 dataset" + }, + { + "type": "image_caption", + "bbox": [ + 0.114, + 0.312, + 0.88, + 0.327 + ], + "angle": 0, + "content": "Figure 3. Comparison of different label correction strategies' accuracies, where Baseline represents the proportion of clean labels." + }, + { + "type": "table_caption", + "bbox": [ + 0.126, + 0.351, + 0.447, + 0.366 + ], + "angle": 0, + "content": "Table 3. Classification accuracies (%) on Clothing1M." + }, + { + "type": "table", + "bbox": [ + 0.13, + 0.376, + 0.443, + 0.533 + ], + "angle": 0, + "content": "
MethodAccuracy
Co-teaching [17]69.21
DMI [54]72.46
DivideMix [28]74.76
CC [62]75.40
DISC [31]74.79
EPL [26]75.21
LRA-diffusion [4]74.46
SANM [46]75.63
DLD (Ours)75.69
" + }, + { + "type": "table_caption", + "bbox": [ + 0.09, + 0.545, + 0.483, + 0.572 + ], + "angle": 0, + "content": "Table 4. Classification accuracies (%) on WebVision and ILSVRC2012 datasets." + }, + { + "type": "table", + "bbox": [ + 0.13, + 0.585, + 0.443, + 0.74 + ], + "angle": 0, + "content": "
MethodWebVisionILSVRC12
Co-teaching [17]63.5861.48
DivideMix [28]77.3275.20
ELR [33]77.7870.29
EPL [26]78.7776.51
C2D [63]79.4278.57
CC [62]79.3676.08
DISC [31]80.2877.44
LRA-diffusion [4]84.1682.56
DLD (Ours)84.5183.74
" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.765, + 0.483, + 0.901 + ], + "angle": 0, + "content": "found in Appendix E.4 and Appendix E.5, respectively. As demonstrated in Table 2, Table 3, and Table 4, DLD outperforms all SOTA methods across these datasets. Note that DLD's relative classification performance improved by \\(7.26\\%\\) and \\(0.44\\%\\) over J-ViT [9] and EPL [26], respectively, both employing the ViT architecture. Compared to LRA-diffusion, which also employs ViT and label diffusion models, DLD achieves varying degrees of performance improvement, demonstrating a more effective and reason-" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.353, + 0.907, + 0.518 + ], + "angle": 0, + "content": "able use of prior knowledge and diffusion models. Besides, while ViT-L/14 improved performance on Clothing1M, Appendix G shows that other extractors like CC can be even more effective, highlighting our method's versatility with different LNL methods. Besides, DLD's performance advantage partly stems from the pre-correction method. In future work, we aim to integrate discriminative and generative paradigms, using diffusion models to refine pre-correction, generate label information, and enhance representational power and multimodal generalization. Appendix H discusses additional limitations and future directions." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.541, + 0.634, + 0.557 + ], + "angle": 0, + "content": "6. Conclusion" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.569, + 0.907, + 0.901 + ], + "angle": 0, + "content": "This paper proposes DLD, a diffusion model specifically designed to learn effectively from noisy labels for robust classification. Our main idea is to decouple the label diffusion process into two channels: directional diffusion and random diffusion. This approach explicitly introduces noisy label information into the diffusion model for the first time, enabling the model to interpretably mitigate the generation of incorrect labels, thereby enhancing its ability to generalize from noisy to clean data. By incorporating our low-cost label pre-correction method, DLD can automatically select label diffusion information for each sample in diverse noisy environments, without requiring additional adjustments or iterations. Additionally, our model synergizes with various pre-trained large-scale models, like ViT, leveraging their prior knowledge more effectively than other methods. Experiments on simulated and real-world noisy datasets show that DLD surpasses conventional LNL methods in both accuracy and efficiency. By framing LNL within a generative model, our method bridges LNL with other learning frameworks. We believe this work will reduce the workload for practitioners handling large-scale, low-quality labeled data, benefiting the development of the image recognition field." + }, + { + "type": "page_number", + "bbox": [ + 0.48, + 0.945, + 0.518, + 0.957 + ], + "angle": 0, + "content": "25745" + } + ], + [ + { + "type": "title", + "bbox": [ + 0.092, + 0.091, + 0.259, + 0.108 + ], + "angle": 0, + "content": "Acknowledgements" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.116, + 0.484, + 0.222 + ], + "angle": 0, + "content": "This work was supported in part by the National Natural Science Foundation of China under Grant U21A20513, Grant 62476157, Grant 62276161, Grant 62276157 and Grant 62406179, and in part by the Key Research and Development Program of Shanxi Province under Grant 202202020101003 and the Fundamental Research Program of Shanxi Province under Grant 202403021222026." + }, + { + "type": "title", + "bbox": [ + 0.093, + 0.238, + 0.188, + 0.253 + ], + "angle": 0, + "content": "References" + }, + { + "type": "ref_text", + "bbox": [ + 0.101, + 0.262, + 0.484, + 0.345 + ], + "angle": 0, + "content": "[1] Devansh Arpit, Stanisław Jastrzebski, Nicolas Ballas, David Krueger, Emmanuel Bengio, Maxinder S Kanwal, Tegan Maharaj, Asja Fischer, Aaron Courville, Yoshua Bengio, et al. A closer look at memorization in deep networks. In International conference on machine learning, pages 233-242. PMLR, 2017. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.101, + 0.347, + 0.484, + 0.415 + ], + "angle": 0, + "content": "[2] Antonin Berthon, Bo Han, Gang Niu, Tongliang Liu, and Masashi Sugiyama. Confidence scores make instance-dependent label-noise learning possible. In International conference on machine learning, pages 825-836. PMLR, 2021. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.418, + 0.483, + 0.487 + ], + "angle": 0, + "content": "[3] Stella Biderman, Usvsn Prashanth, Lintang Sutawika, Hailey Schoelkopf, Quentin Anthony, Shivanshu Purohit, and Edward Raff. Emergent and predictable memorization in large language models. Advances in Neural Information Processing Systems, 36:28072-28090, 2023. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.489, + 0.483, + 0.558 + ], + "angle": 0, + "content": "[4] Jian Chen, Ruiyi Zhang, Tong Yu, Rohan Sharma, Zhiqiang Xu, Tong Sun, and Changyou Chen. Label-retrieval augmented diffusion models for learning from noisy labels. Advances in Neural Information Processing Systems, 36, 2024. 2, 3, 7, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.56, + 0.482, + 0.628 + ], + "angle": 0, + "content": "[5] Pengfei Chen, Ben Ben Liao, Guangyong Chen, and Shengyu Zhang. Understanding and utilizing deep neural networks trained with noisy labels. In International conference on machine learning, pages 1062-1070. PMLR, 2019. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.631, + 0.482, + 0.7 + ], + "angle": 0, + "content": "[6] Pengfei Chen, Junjie Ye, Guangyong Chen, Jingwei Zhao, and Pheng-Ann Heng. Beyond class-conditional assumption: A primary attempt to combat instance-dependent label noise. In Proceedings of the AAAI Conference on Artificial Intelligence, pages 11442–11450, 2021. 1, 2, 6, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.702, + 0.482, + 0.758 + ], + "angle": 0, + "content": "[7] Ting Chen, Simon Kornblith, Mohammad Norouzi, and Geoffrey Hinton. A simple framework for contrastive learning of visual representations. In International conference on machine learning, pages 1597-1607. PMLR, 2020. 2, 3, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.76, + 0.483, + 0.814 + ], + "angle": 0, + "content": "[8] Yingyi Chen, Shell Xu Hu, Xi Shen, Chunrong Ai, and Johan AK Suykens. Compressing features for learning with noisy labels. IEEE Transactions on Neural Networks and Learning Systems, 35(2):2124-2138, 2022. 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.817, + 0.483, + 0.87 + ], + "angle": 0, + "content": "[9] Yingyi Chen, Xi Shen, Yahui Liu, Qinghua Tao, and Johan AK Suykens. Jigsaw-vit: Learning jigsaw puzzles in vision transformer. Pattern Recognition Letters, 166:53-60, 2023. 7, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.873, + 0.482, + 0.901 + ], + "angle": 0, + "content": "[10] De Cheng, Tongliang Liu, Yixiong Ning, Nannan Wang, Bo Han, Gang Niu, Xinbo Gao, and Masashi Sugiyama." + }, + { + "type": "list", + "bbox": [ + 0.094, + 0.262, + 0.484, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.545, + 0.093, + 0.907, + 0.148 + ], + "angle": 0, + "content": "Instance-dependent label-noise learning with manifold-regularized transition matrix estimation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 16630-16639, 2022. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.151, + 0.906, + 0.204 + ], + "angle": 0, + "content": "[11] Hao Cheng, Zhaowei Zhu, Xingyu Li, Yifei Gong, Xing Sun, and Yang Liu. Learning with instance-dependent label noise: A sample sieve approach. arXiv preprint arXiv:2010.02347, 2020." + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.207, + 0.905, + 0.262 + ], + "angle": 0, + "content": "[12] Jiacheng Cheng, Tongliang Liu, Kotagiri Ramamohanarao, and Dacheng Tao. Learning with bounded instance and label-dependent label noise. In International conference on machine learning, pages 1789-1799. PMLR, 2020. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.264, + 0.906, + 0.345 + ], + "angle": 0, + "content": "[13] C Fabian Benitez-Quiroz, Ramprakash Srinivasan, and Aleix M Martinez. Emotionet: An accurate, real-time algorithm for the automatic annotation of a million facial expressions in the wild. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 5562–5570, 2016. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.349, + 0.905, + 0.401 + ], + "angle": 0, + "content": "[14] Chen Feng, Georgios Tzimiropoulos, and Ioannis Patras. Ssr: An efficient and robust framework for learning with unknown label noise. arXiv preprint arXiv:2111.11288, 2021. 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.406, + 0.906, + 0.474 + ], + "angle": 0, + "content": "[15] Arpit Garg, Cuong Nguyen, Rafael Felix, Thanh-Toan Do, and Gustavo Carneiro. Instance-dependent noisy label learning via graphical modelling. In Proceedings of the IEEE/CVF winter conference on applications of computer vision, pages 2288-2298, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.477, + 0.905, + 0.531 + ], + "angle": 0, + "content": "[16] Aritra Ghosh, Himanshu Kumar, and P Shanti Sastry. Robust loss functions under label noise for deep neural networks. In Proceedings of the AAAI conference on artificial intelligence, 2017. 1, 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.533, + 0.905, + 0.602 + ], + "angle": 0, + "content": "[17] Bo Han, Quanming Yao, Xingrui Yu, Gang Niu, Miao Xu, Weihua Hu, Ivor Tsang, and Masashi Sugiyama. Co-teaching: Robust training of deep neural networks with extremely noisy labels. Advances in neural information processing systems, 31, 2018. 1, 2, 7, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.604, + 0.905, + 0.658 + ], + "angle": 0, + "content": "[18] Xizewen Han, Huangjie Zheng, and Mingyuan Zhou. Card: Classification and regression diffusion models. Advances in Neural Information Processing Systems, 35:18100-18115, 2022. 2, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.661, + 0.905, + 0.729 + ], + "angle": 0, + "content": "[19] Kaiming He, Xiangyu Zhang, Shaoqing Ren, and Jian Sun. Delving deep into rectifiers: Surpassing human-level performance on imagenet classification. In Proceedings of the IEEE international conference on computer vision, pages 1026-1034, 2015. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.732, + 0.905, + 0.787 + ], + "angle": 0, + "content": "[20] Kaiming He, Xiangyu Zhang, Shaoqing Ren, and Jian Sun. Deep residual learning for image recognition. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 770-778, 2016. 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.789, + 0.905, + 0.83 + ], + "angle": 0, + "content": "[21] Jonathan Ho, Ajay Jain, and Pieter Abbeel. Denoising diffusion probabilistic models. Advances in neural information processing systems, 33:6840-6851, 2020. 2, 3, 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.832, + 0.905, + 0.9 + ], + "angle": 0, + "content": "[22] Ahmet Iscen, Jack Valmadre, Anurag Arnab, and Cordelia Schmid. Learning with neighbor consistency for noisy labels. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 4672-4681, 2022. 2, 5, 7" + }, + { + "type": "list", + "bbox": [ + 0.517, + 0.093, + 0.907, + 0.9 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.48, + 0.945, + 0.521, + 0.957 + ], + "angle": 0, + "content": "25746" + } + ], + [ + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.092, + 0.482, + 0.16 + ], + "angle": 0, + "content": "[23] Gaoxia Jiang, Jia Zhang, Xuefei Bai, Wenjian Wang, and Deyu Meng. Which is more effective in label noise cleaning, correction or filtering? In Proceedings of the AAAI Conference on Artificial Intelligence, pages 12866-12873, 2024. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.165, + 0.482, + 0.233 + ], + "angle": 0, + "content": "[24] Lu Jiang, Zhengyuan Zhou, Thomas Leung, Li-Jia Li, and Li Fei-Fei. Mentornet: Learning data-driven curriculum for very deep neural networks on corrupted labels. In International conference on machine learning, pages 2304–2313. PMLR, 2018. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.236, + 0.483, + 0.264 + ], + "angle": 0, + "content": "[25] Diederik P Kingma. Auto-encoding variational bayes. arXiv preprint arXiv:1312.6114, 2013. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.267, + 0.483, + 0.322 + ], + "angle": 0, + "content": "[26] Jongwoo Ko, Sumyeong Ahn, and Se-Young Yun. Efficient utilization of pre-trained model for learning with noisy labels. In ICLR 2023 Workshop on Pitfalls of limited data and computation for Trustworthy ML, 2023. 7, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.325, + 0.482, + 0.353 + ], + "angle": 0, + "content": "[27] Alex Krizhevsky. Learning multiple layers of features from tiny images. Master's thesis, University of Toronto, 2009. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.355, + 0.482, + 0.396 + ], + "angle": 0, + "content": "[28] Junnan Li, Richard Socher, and Steven CH Hoi. Dividemix: Learning with noisy labels as semi-supervised learning. arXiv preprint arXiv:2002.07394, 2020. 1, 2, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.399, + 0.482, + 0.468 + ], + "angle": 0, + "content": "[29] Mingchen Li, Mahdi Soltanolkotabi, and Samet Oymak. Gradient descent with early stopping is provably robust to label noise for overparameterized neural networks. In International conference on artificial intelligence and statistics, pages 4313-4324. PMLR, 2020. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.471, + 0.482, + 0.525 + ], + "angle": 0, + "content": "[30] Wen Li, Limin Wang, Wei Li, Eirikur Agustsson, and Luc Van Gool. Webvision database: Visual learning and understanding from web data. arXiv preprint arXiv:1708.02862, 2017. 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.528, + 0.482, + 0.598 + ], + "angle": 0, + "content": "[31] Yifan Li, Hu Han, Shiguang Shan, and Xilin Chen. Disc: Learning from noisy labels via dynamic instance-specific selection and correction. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 24070-24079, 2023. 2, 7, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.601, + 0.482, + 0.668 + ], + "angle": 0, + "content": "[32] Jiawei Liu, Qiang Wang, Huijie Fan, Yinong Wang, Yandong Tang, and Liangqiong Qu. Residual denoising diffusion models. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 2773-2783, 2024. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.672, + 0.482, + 0.728 + ], + "angle": 0, + "content": "[33] Sheng Liu, Jonathan Niles-Weed, Narges Razavian, and Carlos Fernandez-Granda. Early-learning regularization prevents memorization of noisy labels. Advances in neural information processing systems, 33:20331-20342, 2020. 2, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.73, + 0.482, + 0.771 + ], + "angle": 0, + "content": "[34] Yangdi Lu, Yang Bo, and Wenbo He. Co-matching: Combating noisy labels by augmentation anchoring. arXiv preprint arXiv:2103.12814, 2021. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.774, + 0.482, + 0.828 + ], + "angle": 0, + "content": "[35] Anqi Mao, Mehryar Mohri, and Yutao Zhong. Cross-entropy loss functions: Theoretical analysis and applications. In International conference on Machine learning, pages 23803–23828. PMLR, 2023. 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.832, + 0.482, + 0.899 + ], + "angle": 0, + "content": "[36] Duc Tam Nguyen, Chaithanya Kumar Mummadi, Thi Phuong Nhung Ngo, Thi Hoai Phuong Nguyen, Laura Beggel, and Thomas Brox. Self: Learning to filter noisy labels with self-ensembling. arXiv preprint arXiv:1910.01842, 2019. 1" + }, + { + "type": "list", + "bbox": [ + 0.093, + 0.092, + 0.483, + 0.899 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.093, + 0.905, + 0.148 + ], + "angle": 0, + "content": "[37] Kento Nishi, Yi Ding, Alex Rich, and Tobias Hollerer. Augmentation strategies for learning with noisy labels. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 8022-8031, 2021. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.151, + 0.905, + 0.204 + ], + "angle": 0, + "content": "[38] Haim Permuter, Joseph Francos, and Ian Jermyn. A study of gaussian mixture models of color and texture features for image classification and segmentation. Pattern recognition, 39(4):695-706, 2006. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.207, + 0.905, + 0.235 + ], + "angle": 0, + "content": "[39] Lutz Prechelt. Early stopping-but when? In Neural Networks: Tricks of the trade, pages 55-69. Springer, 2002. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.236, + 0.905, + 0.318 + ], + "angle": 0, + "content": "[40] Alec Radford, Jong Wook Kim, Chris Hallacy, Aditya Ramesh, Gabriel Goh, Sandhini Agarwal, Girish Sastry, Amanda Askell, Pamela Mishkin, Jack Clark, et al. Learning transferable visual models from natural language supervision. In International conference on machine learning, pages 8748-8763. PMLR, 2021. 2, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.321, + 0.905, + 0.389 + ], + "angle": 0, + "content": "[41] Olga Russakovsky, Jia Deng, Hao Su, Jonathan Krause, Sanjeev Satheesh, Sean Ma, Zhiheng Huang, Andrej Karpathy, Aditya Khosla, Michael Bernstein, et al. Imagenet large scale visual recognition challenge. International journal of computer vision, 115:211-252, 2015. 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.392, + 0.905, + 0.473 + ], + "angle": 0, + "content": "[42] Kihyuk Sohn, David Berthelot, Nicholas Carlini, Zizhao Zhang, Han Zhang, Colin A Raffel, Ekin Dogus Cubuk, Alexey Kurakin, and Chun-Liang Li. Fixmatch: Simplifying semi-supervised learning with consistency and confidence. Advances in neural information processing systems, 33:596-608, 2020. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.476, + 0.905, + 0.531 + ], + "angle": 0, + "content": "[43] Hwanjun Song, Minseok Kim, and Jae-Gil Lee. Selfie: Refurbishing unclean samples for robust deep learning. In International conference on machine learning, pages 5907-5915. PMLR, 2019. 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.533, + 0.905, + 0.573 + ], + "angle": 0, + "content": "[44] Jiaming Song, Chenlin Meng, and Stefano Ermon. Denoising diffusion implicit models. arXiv preprint arXiv:2010.02502, 2020. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.576, + 0.905, + 0.631 + ], + "angle": 0, + "content": "[45] Sunil Thulasidasan, Tanmoy Bhattacharya, Jeff Bilmes, Gopinath Chennupati, and Jamal Mohd-Yusof. Combating label noise in deep learning using abstention. arXiv preprint arXiv:1905.10964, 2019. 2, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.633, + 0.905, + 0.716 + ], + "angle": 0, + "content": "[46] Yuanpeng Tu, Boshen Zhang, Yuxi Li, Liang Liu, Jian Li, Jiangning Zhang, Yabiao Wang, Chengjie Wang, and Cai Rong Zhao. Learning with noisy labels via self-supervised adversarial noisy masking. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 16186-16195, 2023. 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.718, + 0.905, + 0.786 + ], + "angle": 0, + "content": "[47] Yisen Wang, Weiyang Liu, Xingjun Ma, James Bailey, Hongyuan Zha, Le Song, and Shu-Tao Xia. Iterative learning with open-set noisy labels. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 8688-8696, 2018. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.789, + 0.905, + 0.856 + ], + "angle": 0, + "content": "[48] Yisen Wang, Xingjun Ma, Zaiyi Chen, Yuan Luo, Jinfeng Yi, and James Bailey. Symmetric cross entropy for robust learning with noisy labels. In Proceedings of the IEEE/CVF international computer vision and pattern recognition, pages 322-330, 2019. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.859, + 0.905, + 0.901 + ], + "angle": 0, + "content": "[49] Hongxin Wei, Lei Feng, Xiangyu Chen, and Bo An. Combating noisy labels by agreement: A joint training method with co-regularization. In Proceedings of the IEEE/CVF con" + }, + { + "type": "list", + "bbox": [ + 0.517, + 0.093, + 0.905, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.48, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "25747" + } + ], + [ + { + "type": "ref_text", + "bbox": [ + 0.123, + 0.093, + 0.482, + 0.119 + ], + "angle": 0, + "content": "ference on computer vision and pattern recognition, pages 13726-13735, 2020. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.122, + 0.483, + 0.177 + ], + "angle": 0, + "content": "[50] Pengxiang Wu, Songzhu Zheng, Mayank Goswami, Dimitris Metaxas, and Chao Chen. A topological filter for learning with label noise. Advances in neural information processing systems, 33:21382-21393, 2020. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.179, + 0.482, + 0.248 + ], + "angle": 0, + "content": "[51] Xiaobo Xia, Tongliang Liu, Bo Han, Nannan Wang, Mingming Gong, Haifeng Liu, Gang Niu, Dacheng Tao, and Masashi Sugiyama. Part-dependent label noise: Towards instance-dependent label noise. Advances in Neural Information Processing Systems, 33:7597-7610, 2020. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.25, + 0.482, + 0.317 + ], + "angle": 0, + "content": "[52] Tong Xiao, Tian Xia, Yi Yang, Chang Huang, and Xiaogang Wang. Learning from massive noisy labeled data for image classification. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 2691-2699, 2015. 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.32, + 0.482, + 0.36 + ], + "angle": 0, + "content": "[53] Cai Xu, Yilin Zhang, Ziyu Guan, and Wei Zhao. Trusted multi-view learning with label noise. arXiv preprint arXiv:2404.11944, 2024. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.363, + 0.482, + 0.418 + ], + "angle": 0, + "content": "[54] Yilun Xu, Peng Cao, Yuqing Kong, and Yizhou Wang. L.dmi: A novel information-theoretic loss function for training deep nets robust to label noise. Advances in neural information processing systems, 32, 2019. 2, 7, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.42, + 0.482, + 0.473 + ], + "angle": 0, + "content": "[55] Yan Yan, Rómer Rosales, Glenn Fung, Ramanathan Subramanian, and Jennifer Dy. Learning from multiple annotators with varying expertise. Machine learning, 95:291-327, 2014. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.478, + 0.482, + 0.533 + ], + "angle": 0, + "content": "[56] Quanming Yao, Hansi Yang, Bo Han, Gang Niu, and James Tin-Yau Kwok. Searching to exploit memorization effect in learning with noisy labels. In International Conference on Machine Learning, pages 10789-10798. PMLR, 2020. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.535, + 0.482, + 0.588 + ], + "angle": 0, + "content": "[57] Yu Yao, Tongliang Liu, Mingming Gong, Bo Han, Gang Niu, and Kun Zhang. Instance-dependent label-noise learning under a structural causal model. Advances in Neural Information Processing Systems, 34:4409-4420, 2021. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.591, + 0.482, + 0.657 + ], + "angle": 0, + "content": "[58] Xingrui Yu, Bo Han, Jiangchao Yao, Gang Niu, Ivor Tsang, and Masashi Sugiyama. How does disagreement help generalization against label corruption? In International conference on machine learning, pages 7164-7173. PMLR, 2019. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.661, + 0.482, + 0.716 + ], + "angle": 0, + "content": "[59] Chiyuan Zhang, Samy Bengio, Moritz Hardt, Benjamin Recht, and Oriol Vinyals. Understanding deep learning (still) requires rethinking generalization. Communications of the ACM, 64(3):107-115, 2021. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.718, + 0.482, + 0.772 + ], + "angle": 0, + "content": "[60] Yikai Zhang, Songzhu Zheng, Pengxiang Wu, Mayank Goswami, and Chao Chen. Learning with feature-dependent label noise: A progressive approach. arXiv preprint arXiv:2103.07756, 2021. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.775, + 0.482, + 0.829 + ], + "angle": 0, + "content": "[61] Zhilu Zhang and Mert Sabuncu. Generalized cross entropy loss for training deep neural networks with noisy labels. Advances in neural information processing systems, 31, 2018, 1, 2, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.832, + 0.482, + 0.901 + ], + "angle": 0, + "content": "[62] Ganlong Zhao, Guanbin Li, Yipeng Qin, Feng Liu, and Yizhou Yu. Centrality and consistency: two-stage clean samples identification for learning with instance-dependent noisy labels. In European Conference on Computer Vision, pages 21-37. Springer, 2022. 2, 7, 8" + }, + { + "type": "list", + "bbox": [ + 0.093, + 0.093, + 0.483, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.093, + 0.905, + 0.162 + ], + "angle": 0, + "content": "[63] Evgenii Zheltonozhskii, Chaim Baskin, Avi Mendelson, Alex M Bronstein, and Or Litany. Contrast to divide: Self-supervised pre-training for learning with noisy labels. In Proceedings of the IEEE/CVF Winter Conference on Applications of Computer Vision, pages 1657-1667, 2022. 2, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.164, + 0.905, + 0.218 + ], + "angle": 0, + "content": "[64] Songzhu Zheng, Pengxiang Wu, Aman Goswami, Mayank Goswami, Dimitris Metaxas, and Chao Chen. Error-bounded correction of noisy labels. In International Conference on Machine Learning, pages 11447-11457. PMLR, 2020. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.22, + 0.905, + 0.274 + ], + "angle": 0, + "content": "[65] Yuanyi Zhong, Haoran Tang, Junkun Chen, Jian Peng, and Yu-Xiong Wang. Is self-supervised learning more robust than supervised learning? arXiv preprint arXiv:2206.05259, 2022. 2" + }, + { + "type": "list", + "bbox": [ + 0.517, + 0.093, + 0.905, + 0.274 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.48, + 0.946, + 0.519, + 0.957 + ], + "angle": 0, + "content": "25748" + } + ] +] \ No newline at end of file diff --git a/2025/Directional Label Diffusion Model for Learning from Noisy Labels/133ce87a-4a21-4fe5-b113-05841acf1ebd_origin.pdf b/2025/Directional Label Diffusion Model for Learning from Noisy Labels/133ce87a-4a21-4fe5-b113-05841acf1ebd_origin.pdf new file mode 100644 index 0000000000000000000000000000000000000000..24a4910fbfc6bb8202d72e4bedd087aca3d4748f --- /dev/null +++ b/2025/Directional Label Diffusion Model for Learning from Noisy Labels/133ce87a-4a21-4fe5-b113-05841acf1ebd_origin.pdf @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:287734477c7243ca89d8b9959559226afb41092b88799f6b5d581ab9fefca2d4 +size 1302746 diff --git a/2025/Directional Label Diffusion Model for Learning from Noisy Labels/full.md b/2025/Directional Label Diffusion Model for Learning from Noisy Labels/full.md new file mode 100644 index 0000000000000000000000000000000000000000..c9d6112e323f4833867b4289fa93fd7473b0180e --- /dev/null +++ b/2025/Directional Label Diffusion Model for Learning from Noisy Labels/full.md @@ -0,0 +1,337 @@ +# Directional Label Diffusion Model for Learning from Noisy Labels + +Senyu Hou $^{1}$ , Gaoxia Jiang $^{1}$ , Jia Zhang $^{1}$ , Shangrong Yang $^{1}$ , Husheng Guo $^{1}$ , Yaqing Guo $^{1}$ , Wenjian Wang $^{2*}$ + +$^{1}$ School of Computer and Information Technology, Shanxi University, China + $^{2}$ Key Laboratory of Computational Intelligence and Chinese Information Processing Ministry of Education, Shanxi University, China + +{202312407005, 202412407035}@email.sxu.edu.cn + +{jianggaoxia, sryang, guohusheng, yqguo, wjwang}@sxu.edu.cn + +# Abstract + +In image classification, the label quality of training data critically influences model generalization, especially for deep neural networks (DNNs). Traditionally, learning from noisy labels (LNL) can improve the generalization of DNNs through complex architectures or a series of robust techniques, but its performance improvement is limited by the discriminative paradigm. Unlike traditional ways, we resolve the LNL problems from the perspective of robust label generation, based on diffusion models within the generative paradigm. To expand the diffusion model into a robust classifier that explicitly accommodates more noise knowledge, we propose a Directional Label Diffusion (DLD) model. It disentangles the diffusion process into two paths, i.e., directional diffusion and random diffusion. Specifically, directional diffusion simulates the corruption of true labels into a directed noise distribution, prioritizing the removal of likely noise, whereas random diffusion introduces inherent randomness to support label recovery. This architecture enable DLD to gradually infer labels from an initial random state, interpretably diverging from the specified noise distribution. To adapt the model to diverse noisy environments, we design a low-cost label precorrection method that automatically supplies more accurate label information to the diffusion model, without requiring manual intervention or additional iterations. Our approach outperforms state-of-the-art methods on both simulated and real-world noisy datasets. Code is available at https://github.com/SenyuHou/DLD. + +# 1. Introduction + +Deep neural networks (DNNs) have achieved notable success across key areas of machine learning, driven by the growth of datasets and the increase in model parameters. + +These models are primarily trained using supervised learning, which relies heavily on large volumes of labeled data [5]. However, real-world data often suffer from low-quality labels due to subjective errors like indistinguishable samples or objective errors such as incorrect label recording, making noisy labels a widespread issue [13, 55]. Studies have shown that the generalization capability of deep neural networks is adversely affected when trained with noisy labels [1, 59]. Consequently, it is crucial to develop methods for learning from noisy labels (LNL) on large-scale but low-quality datasets. + +Due to the varied distribution of label noise, historical research has classified it into class-conditional noise (CCN) and instance-dependent noise (IDN) [57]. Recent studies indicate that label noise in real datasets predominantly aligns with the IDN assumption, wherein noise is unevenly distributed and more likely in samples with similar features [2, 6, 10–12, 51]. To tackle these complex noise patterns, mainstream LNL methods employ strategies combining filtering and training approaches. Some studies demonstrate a phenomenon known as the memorization effect [1, 3, 29, 56], i.e., DNNs first learn from simple data before fitting to noisy samples. Based on this observation, sample selection strategies utilize confidence levels or prediction loss during training intervals to identify noisy samples [17, 24, 28, 58]. However, this strategy is inherently risky, as the timing of noise memorization varies across datasets and noise distributions, complicating the selection criteria [6, 39]. Consequently, mainline LNL methods can often be integrated with training strategies to enhance model robustness, employing techniques like label filtering [17, 36, 50, 58], label correction [47, 60, 64], robust loss [16, 48, 61], and semi-supervised learning [28]. Among these techniques, semi-supervised learning, by combining multiple approaches [34, 37, 49], often outperforms other LNL methods but is still constrained by the discriminative framework. Although some researches attempt to combine + +variational autoencoders (VAE) [25] with DivideMix [28] to address label noise [15], most methods remain focused on discriminative models. + +In this paper, we shift away from the traditional LNL perspective, proposing robust classification models using label diffusion models that focus on the generative modeling and deeply explores the generative relationship between features and labels. Under this paradigm, labels are generated by gradually refining an initial random guess. The uncertainty modeling in diffusion models inherently meets some robustness requirements, as random label generation can offset some noise interference [4, 18]. However, this type of diffusion process lacks information about noisy labels. Intrinsic uncertainty alone is insufficient to handle complex noisy environments. Inspired by the residual diffusion denoising models (RDDM) [32], we explore a novel Directional Label Diffusion (DLD) model that incorporates noisy label information into the diffusion process. In DLD, we decouple the diffusion process into two channels: directional diffusion and random diffusion (see Figure 1). The directional diffusion guides the initial label towards a more probable noise direction (e.g., a green automobile may be misclassified as a frog or other classes, as indicated by the probability distribution in the third column in Figure 1), while random diffusion retains the original diffusion mechanism. This approach injects information about which categories should not be generated into the label generation paradigm, enhancing the interpretability of the diffusion model's robust inference. Unlike conventional label diffusion, our diffusion process concludes with an endpoint that includes the potential noise distribution of the sample (i.e., $\mathbf{y}_T = \mathbf{y}_n + \epsilon$ ) and employs two independent coefficient schedules to control the diffusion speed of both channels. + +One challenge is that only noisy labels are available in practice for training, so neither $\mathbf{y}_0$ nor $\mathbf{y}_n$ can be obtained directly. Few researchers employed neighborhood detection techniques like KNN to re-estimate $\mathbf{y}_0$ [4]. However, under the IDN assumption, neighborhood labels of noisy samples often contain similar errors, making naive label estimation methods unreliable and potentially causing $\mathbf{y}_0$ to deviate from the ground truth (see Figure 3). To address these issues, we further propose a straightforward and efficient label pre-correction method that automatically provides diverse label information for diffusion model training. This framework employs feature extractors to identify the neighborhood label distribution of samples. Note that these extractors are either trained with self-supervised methods [7, 19, 65], which require few additional training iterations, or use pre-trained models without further training, thus remaining unaffected by label noise. Additionally, this pre-correction strategy allows the seamless integration of pre-trained large-scale models such as Sim-CLR [7] and vision transformer (ViT) [40] into our framework to further + +boost the performance. The main contributions of this work are summarized as follows: + +- We propose a novel Directional Label Diffusion (DLD) model that explicitly introduces interpretable noise information through a dual diffusion architecture, enabling robust learning of label generation patterns from noisy data. +- We design a fast label pre-correction method that automatically provides more accurate label information for the diffusion process and richer feature conditions for the generation process. By leveraging the consistency principle, this method pre-estimates the label distribution in a single iteration without incurring additional costs. +- We refine the paradigm for introducing feature conditions with rigorous theoretical deduction and provide complementary training strategies to enhance robustness further. +- We experimentally show that our model surpasses the state-of-the-art (SOTA) methods on both simulated and real-world noisy datasets, e.g., $6\% \sim 21\%$ accuracy improvement on the CIFAR-100 dataset with IDN. + +# 2. Related Work + +# 2.1. Learning from Noisy Labels + +Screening and training strategies are central to LNL methods, aiming to distinguish noisy from clean samples and to optimize model learning under noisy conditions. Screening strategies, like Co-teaching [17], select low-loss samples as clean, while DivideMix [28] uses a Gaussian mixture model to automatically identify clean samples during training. Confidence-based methods like CC [62] and DISC [31] classify samples with unreliable predictions as noisy, whereas neighbor consistency regularization (NCR) [22] promotes learning from samples with high neighbor consistency. Our pre-correction method is based on similarity constraints across multi-view label distributions and leverages a feature extractor to provide high-quality information to the diffusion model without requiring additional iterative training. For training, models such as SEAL [6] iteratively correct noisy labels, but this often risks worsening outcomes [23]. Robust loss functions, like mean absolute error (MAE) [16] and generalized cross entropy (GCE) [61], enhance resilience to noisy labels, while DAC [45] and DMI [54] offer advanced mutual information-based losses. Approaches like DivideMix [28] treat LNL as semi-supervised learning, supported by methods like CC [62], C2D [63], and ELR [33] to refine the process. Without complex iterations or loss design, our method utilizes a lightweight diffusion network and a simple L2 loss to quickly learn the feature-label generation relationship. + +# 2.2. Label Denoising Diffusion Models + +Diffusion models were initially designed for image generation tasks and can be understood through the lenses of score matching and dynamics. Ho et al. [21] introduced a + +![](images/c07b814ed34be5d0f57d0b65e23ec6f71e12a29dad6bf275b3b5b1a6c5210b02.jpg) +Figure 1. The proposed Directional Label Diffusion (DLD) model includes diagrams illustrating both the forward and reverse processes. In the forward process, diffusion is decoupled into two channels: random diffusion and directional diffusion. Random diffusion represents the transformation of $\mathbf{y}_0$ into random Gaussian noise $\epsilon$ , while directional diffusion describes the transformation of $\mathbf{y}_0$ into a specific sample noise label distribution $\mathbf{y}_n$ . Ultimately, $\mathbf{y}_0$ diffuses to the endpoint $\mathbf{y}_T = \mathbf{y}_0 + \epsilon$ . Notably, the features of the samples serve as control conditions, guiding the generation of labels only during the reverse process. + +more classical interpretative framework known as denoising diffusion probabilistic models (DDPM), which have been widely applied across diverse domains. RDDM expands the denoising-based diffusion models, initially uninterpretable for image restoration, into a unified and interpretable model for both image generation and restoration by introducing residuals. Classification and regression diffusion models (CARD) [18] reconstruct the classification as a conditional generative task, generating labels from dimensionally reduced image features via a pre-trained encoder. Building on CARD, label-retrieval-augmented (LRA) diffusion model [4] re-estimates the generation targets through neighbor label retrieval, improving label accuracy. Our model disentangles the diffusion process into directional and random diffusion channels, enhancing diffusion quality through precorrection, which leads to train a more robust label diffusion model. + +# 3. Preliminary + +In this section, we primarily introduce the label denoising diffusion models (LDDM), which is constructed based on the theoretical framework of the DDPM [21]. We start by defining the data distribution $\mathbf{y}_0\sim q(\mathbf{y}_0)$ and a Markovian forward (diffusion) process $q$ , which progressively adds Gaussian noise $\epsilon \sim \mathcal{N}(0,\mathbf{I})$ to the labels based on diffusion strength $\beta_{t}$ , diffusing label $\mathbf{y}_0$ to $\mathbf{y}_T$ .. + +$$ +q \left(\mathbf {y} _ {t} \mid \mathbf {y} _ {t - 1}\right) = \mathcal {N} \left(\mathbf {y} _ {t}; \sqrt {1 - \beta_ {t}} \mathbf {y} _ {t - 1}, \beta_ {t} \mathbf {I}\right), \tag {1} +$$ + +Under Bayes' theorem, one finds that the posterior distribution of the forward process can be represented as a Gaussian distribution with mean $\tilde{\mu}_t(\mathbf{y}_t,\mathbf{y}_0)$ and variance $\tilde{\beta}_t$ : + +$$ +q \left(\mathbf {y} _ {t - 1} \mid \mathbf {y} _ {t}, \mathbf {y} _ {0}\right) = \mathcal {N} \left(\mathbf {y} _ {t - 1}; \tilde {\mu} _ {t} \left(\mathbf {y} _ {t}, \mathbf {y} _ {0}\right), \tilde {\beta} _ {t} \mathbf {I}\right), \tag {2} +$$ + +where $\tilde{\mu}_t(\mathbf{y}_t,\mathbf{y}_0) = \frac{\sqrt{\bar{\alpha}_{t - 1}}\beta_t}{1 - \bar{\alpha}_t}\mathbf{y}_0 + \frac{\sqrt{\alpha_t(1 - \bar{\alpha}_{t - 1})}}{1 - \bar{\alpha}_t}\mathbf{y}_t$ and $\tilde{\beta}_t = \frac{1 - \bar{\alpha}_{t - 1}}{1 - \bar{\alpha}_t}\beta_t$ , In the reverse process, we naturally employ a neural network to perform Markovian inference from $q(\mathbf{y}_T)$ , fitting the posterior distribution $q(\mathbf{y}_{t - 1}|\mathbf{y}_t)$ through to $\mathbf{y}_0$ . Since the variance of the posterior Gaussian distribution depends only on the diffusion strength corresponding to time $t$ , the neural network focuses primarily on predicting and learning the mean: + +$$ +p _ {\theta} \left(\mathbf {y} _ {t - 1} \mid \mathbf {y} _ {t}\right) := \mathcal {N} \left(\mathbf {y} _ {t - 1}; \mu_ {\theta} \left(\mathbf {y} _ {t}, t\right), \tilde {\beta} _ {t} \mathbf {I}\right). \tag {3} +$$ + +To train this model such that $p_{\theta}(\mathbf{y}_0)$ closely approximates $q(\mathbf{y}_0)$ , we can optimize the variational lower bound (VLB), however, research [7] found that directly predicting the mean was not effective, so they reparameterized the estimation target, constraining the neural network to estimate the actual noise of the diffusion process and simplified the loss function: + +$$ +L _ {\text {s i m p l e}} := E _ {t \sim [ 1, T ], \mathbf {y} _ {0} \sim q (\mathbf {y} _ {0}), \epsilon \sim \mathcal {N} (0, \mathbf {I})} \left[ \| \epsilon - \epsilon_ {\theta} (\mathbf {y} _ {t}, t) \| ^ {2} \right]. \tag {4} +$$ + +During inference, we utilize the label noise estimated by the model, restoring the mean of the label at time $t - 1$ to retrieve its distribution form, and then gradually performing random sampling until generating the label $\mathbf{y}_0$ : + +$$ +\mathbf {y} _ {t - 1} = \frac {1}{\sqrt {\alpha_ {t}}} \left(\mathbf {y} _ {t} - \frac {1 - \alpha_ {t}}{\sqrt {1 - \bar {\alpha} _ {t}}} \epsilon_ {\theta} (\mathbf {y} _ {t}, t)\right) + \sqrt {\tilde {\beta} _ {t}} \mathbf {z}, \tag {5} +$$ + +where $\mathbf{z} \sim \mathcal{N}(0, \mathbf{I})$ . Unfortunately, this label generation process lacks any conditional information, resulting in random label generation for images with identical features, which fails to achieve classification. To address this, we + +analyze the foundational paradigm of conditional label diffusion models in Appendix A and simplify the method of feature conditioning. + +# 4. Directional Label Diffusion Model + +Based on LDDM, our goal is to develop a directional diffusion process on labels to introduce more knowledge about noise. We modify the representation of $\mathbf{y}_T = \epsilon$ in LDDM to $\mathbf{y}_T = \mathbf{y}_n + \epsilon$ in our DLD, where $\mathbf{y}_n$ is an estimation of noisy label distribution. The improved forward process is no longer a simple random noise diffusion but is disentangled into two paths of directional diffusion and random diffusion, as shown in Figure 1. + +In the following subsections, we first outline the overall framework of our model in Section 4.1. Next, in Section 4.2, we present the foundational theory of the DLD diffusion process and analyze the method of conditional control. In Section 4.3, we describe the reverse process and training method, including optimizations for introducing feature conditions and an adapted accelerated inference method. Finally, in Section 4.4, we describe the rapid label pre-correction method within the DLD framework. + +# 4.1. Model Overview + +We propose the Directional Label Diffusion (DLD) framework, with the overall process illustrated in Figure 2. The framework contains two main stages: label pre-correction (Stage 1) and the directional label diffusion model (Stage 2). In stage 1, we first use two different data augmentation methods to provide two views for each sample, then use a pre-trained feature extractor $f_{ext}$ to observe the neighborhood label distribution $p_w$ and $p_s$ of the same sample in two feature spaces. Based on similarity constraints and the principle of neighborhood consistency, we divide the dataset into clean, noisy, and hard sets. Then, the label precorrector and noise distribution estimator prepare $\mathbf{y}_0$ and $\mathbf{y}_n$ based on the sample partitioning results, using $\mathbf{y}_0$ as the initial diffusion label and $\mathbf{y}_n$ as part of the diffusion endpoint for the next stage of model training. Otherwise, we introduce both augmented image features and high-quality features extracted by $f_{ext}$ into the training process to jointly guide the learning of the robust diffusion model. + +# 4.2. Forward Process of DLD + +To model the random diffusion of labels and the addition of directional noise information, we define the single forward (diffusion) process step in our DLD as follows: + +$$ +\mathbf {y} _ {t} = \mathbf {y} _ {t - 1} + \alpha_ {t} \mathbf {y} _ {d} + \beta_ {t} \epsilon , \tag {6} +$$ + +where $\mathbf{y}_d$ is the deviation between $\mathbf{y}_n$ and $y_0$ (i.e., $\mathbf{y}_d = \mathbf{y}_n - \mathbf{y}_0$ ), representing the directional shift contained in the diffusion from $\mathbf{y}_{t-1}$ to $\mathbf{y}_t$ . Two independent coefficients, $\alpha_t$ + +and $\beta_{t}$ control the degrees of directional and random diffusion, respectively. Based on the additivity of Gaussian noise, we derive the jump diffusion formula from $\mathbf{y}_0$ to $\mathbf{y}_t$ without requiring single-step continuous diffusion (see Appendix B for the proof): + +$$ +\mathbf {y} _ {t} = \mathbf {y} _ {0} + \bar {\alpha} _ {t} \mathbf {y} _ {d} + \bar {\beta} _ {t} \epsilon , \tag {7} +$$ + +where $\bar{\alpha}_t = \sum_{i=1}^t \alpha_i$ and $\bar{\beta}_t = \sqrt{\sum_{i=1}^t \beta_i^2}$ . When $t = T$ , we hope $\mathbf{y}_T = \mathbf{y}_n + \epsilon = \mathbf{y}_0 + \mathbf{y}_d + \epsilon$ , so we usually set $\bar{\alpha}_T = \bar{\beta}_T = 1$ . This implies that the two coefficients $\alpha_t$ and $\beta_t$ are not arbitrary transformations; rather, they should satisfy a specific sum constraint. Then, we redefine the conditional Markov diffusion process similar to the directional label forward distribution $q$ without condition (detailed theoretical framework in Appendix C), assuming that when conditioned on features $\mathbf{x}$ , the prior distribution of $\hat{q}$ is the same as $q$ : + +$$ +\begin{array}{l} \hat {q} \left(\mathbf {y} _ {t} \mid \mathbf {y} _ {t - 1}, \mathbf {y} _ {d}, \mathbf {x}\right) := q \left(\mathbf {y} _ {t} \mid \mathbf {y} _ {t - 1}, \mathbf {y} _ {d}\right) \tag {8} \\ = \mathcal {N} \left(\mathbf {y} _ {t}; \mathbf {y} _ {t - 1} + \alpha_ {t} \mathbf {y} _ {d}, \beta_ {t} ^ {2} \mathbf {I}\right), \\ \end{array} +$$ + +$$ +\begin{array}{l} \hat {q} \left(\mathbf {y} _ {1: T} \mid \mathbf {y} _ {0}, \mathbf {y} _ {d}, \mathbf {x}\right) := \prod_ {t = 1} ^ {T} \hat {q} \left(\mathbf {y} _ {t} \mid \mathbf {y} _ {t - 1}, \mathbf {y} _ {d}, \mathbf {x}\right) \tag {9} \\ = \mathcal {N} \left(\mathbf {y} _ {t}; \mathbf {y} _ {0} + \bar {\alpha} _ {t} \mathbf {y} _ {d}, \bar {\beta} _ {t} ^ {2} \mathbf {I}\right), \\ \end{array} +$$ + +which means that the diffusion process of the labels is not influenced by any image features. + +In Appendix C, using Bayes' rule and Taylor expansion, we derive the conditional posterior distribution during the forward process: + +$$ +\hat {q} \left(\mathbf {y} _ {t - 1} \mid \mathbf {y} _ {t}, \mathbf {y} _ {0}, \mathbf {y} _ {d}, \mathbf {x}\right) = \mathcal {N} \left(\mathbf {y} _ {t - 1}; \hat {\mu} _ {t}, \hat {\sigma} _ {t} \mathbf {I}\right), \tag {10} +$$ + +where $\hat{\mu}_t = y_t - \alpha_t y_d - \frac{\beta_t^2}{\beta_t} \epsilon + \hat{\sigma}_t \nabla_{y_{t-1}} \log \hat{q}(\mathbf{x} | \mathbf{y}_{t-1}, \mathbf{y}_d)$ and $\hat{\sigma}_t = \frac{\beta_t^2 \bar{\beta}_{t-1}^2}{\beta_t^2}$ . The last term in $\hat{\mu}_t$ can be considered a deviation to the unconditional distribution mean, guided by the first derivative of $\hat{q}(\mathbf{x} | \mathbf{y}_{t-1}, \mathbf{y}_d)$ , indicating that the decoding information from $\mathbf{y}_{t-1}$ to $\mathbf{x}$ should be introduced in training. + +# 4.3. Reverse and Training Process of DLD + +Reverse Process and Training. In the reverse (generation) process, as indicated by Eq. 6, we need to simultaneously predict $\mathbf{y}_d$ and $\epsilon$ to progressively recover $\mathbf{y}_T$ to $\mathbf{y}_0$ . Therefore, we can train two independent networks to estimate $\mathbf{y}_d^{\theta_1}$ and $\epsilon_{\theta_2}$ , respectively, to compute the reverse posterior distribution. Furthermore, we directly use feature information as input for two networks: + +$$ +p _ {\theta} \left(\mathbf {y} _ {t - 1} \mid \mathbf {y} _ {t}, \mathbf {x}\right) := \mathcal {N} \left(\mathbf {y} _ {t - 1}; \mu_ {\theta}, \hat {\sigma} _ {t} ^ {2} \mathbf {I}\right). \tag {11} +$$ + +![](images/e97e2bb5b87d4508df05e432f7987aa9cc5744c0ac32100494bd5f996e06b6d5.jpg) +Figure 2. Overview of the DLD framework. During Stage 1, the noisy dataset is quickly divided into three subsets. Then, we estimate $\mathbf{y}_0$ and $\mathbf{y}_n$ for each samples based on Eq. 14 and Eq. 15. In Stage 2, the label information and augmented features from Stage 1 are jointly used to train a directional diffusion model (blue bars represent random diffusion and gray bars represent directional diffusion). + +Based on the analysis of the forward process, we adopt a straightforward approach which naturally integrates the decoding process into the diffusion model architecture by a fixed decoding layer $L_{1}$ to map $\mathbf{y}$ into the feature space of $\mathbf{x}$ . Detailed information about the specific diffusion model network structure is shown in Appendix Figure C.1. + +By optimizing the VLB and reparameterization [21], we derive the following simplified loss function for training: + +$$ +\mathcal {L} _ {d} = \left\| \mathbf {y} _ {d} - \mathbf {y} _ {d} ^ {\theta_ {1}} \left(\mathbf {y} _ {t}, \mathbf {y} _ {n}, \mathbf {x}, t\right) \right\| ^ {2}, \tag {12} +$$ + +$$ +\mathcal {L} _ {\epsilon} = \left\| \epsilon - \epsilon_ {\theta_ {2}} \left(\mathbf {y} _ {t}, \mathbf {y} _ {n}, \mathbf {x}, t\right) \right\| ^ {2}, \tag {13} +$$ + +where $\epsilon \sim \mathcal{N}(0,\mathbf{I})$ , and the training input $\mathbf{y}_t$ is synthesized using Eq. 7. Algorithm 1 describes the training procedure. + +Inference. Since the classification diffusion model is deterministic in inference and the dimensionality of the label vector is much lower than that of images, we mimic the inference process of DDIM [44] to generate labels in as few steps as possible without affecting the generated categories. We can gradually sample labels step-by-step following $\mathbf{y}_{t-1} = \mathbf{y}_t - (\bar{\alpha}_t - \bar{\alpha}_{t-1})\mathbf{y}_d^{\theta_1} - (\bar{\beta}_t - \bar{\beta}_{t-1})\epsilon_{\theta_2}$ , where the transition from $t-1$ to $t$ is discrete rather than continuous in time. Appendix D provides a detailed explanation of the de-Markovization process, the accelerated sampling algorithm and the visual demonstration of the sampling process. In experiments, we use 5 steps for DLD's inference, which is on the same order of magnitude as the speed of traditional classification models. + +# 4.4. Label Pre-correction Method + +Although diffusion models are powerful, under the guidance of incorrect $\mathbf{y}_0$ and $\mathbf{y}_n$ , they can easily fit the wrong conditional distributions. Therefore, we propose a one-step label pre-correction method that automatically provides accurate label information and feature conditions for the DLD model, requiring no additional parameter training, making it simple, straightforward, and efficient. + +Label Pre-correction for $\mathbf{y}_0$ . Inspired by the concepts of multi-view[42, 53] and neighbor consistency regularization [22], our primary assumption is that the same sample should have similar neighborhood label distributions in different feature spaces. We define the original noisy dataset as $\mathcal{D} = \{(x_i,\tilde{y}_i)\mid x_i\in \mathbb{R}^d,\tilde{y}_i\in \mathbb{R}^m,i = 1,\ldots ,n\}$ . To obtain different feature spaces by a shared feature extractor $f_{ext}$ , we first apply two different data augmentation methods and map them into $\mathbb{R}^w$ and $\mathbb{R}^s$ , with the augmented datasets respectively denoted as $\mathcal{D}^w$ and $\mathcal{D}^s$ . Then, we use Euclidean or Cosine similarity to obtain the nearest $K$ neighborhood labels $y_{k = 1}^{K}$ and feature space distances $v_{k = 1}^{K}$ then we can get two weighted label distributions $p_w$ and $p_s$ by $p(y_{i}) = \left(\sum_{k = 1}^{K}w_{k}\cdot \mathbf{1}(y_{k} = y_{i})\right)\cdot \left(\sum_{k = 1}^{K}w_{k}\right)^{-1}$ where $w_{k} = \frac{1}{v_{k} + \delta}$ , where $\delta$ is a small constant to prevent division by zero. Next, we calculate the divergence between the two distributions using the Kullback-Leibler divergence and a via Gaussian mixture model (GMM) [38] to re-partitioning three types of samples: Samples with high means are reassigned to the hard set $\mathcal{D}_H$ , which represents + +samples whose correctness is hard to determine and typically lie at the boundaries of classes. Samples with low means indicate that similar neighborhood distributions under both views. Using the principle of neighborhood consistency, we compare $\text{argmax}(p_{ws})$ with the original label $\tilde{y}$ , where $p_{ws} = \frac{p_w + p_s}{2}$ . If the class with the highest frequency in the average neighborhood distribution matches the original label, then the sample is assigned to the clean set $\mathcal{D}_C$ ; otherwise, it is assigned to the noisy set $\mathcal{D}_N$ . + +Based on the repartitioned sample subsets and the original noisy labels $\tilde{y}$ , we define a label pre-corrector, which functions similarly to a transition matrix. It can automatically perform label flipping or refinement: + +$$ +P \left(y _ {0} \mid \tilde {y}, x\right) = \left\{ \begin{array}{l l} \tilde {y} & \text {i f} x \in \mathcal {D} _ {C}, \\ \operatorname {a r g m a x} \left(p _ {w s}\right) & \text {i f} x \in \mathcal {D} _ {N}, \\ p _ {w s} & \text {i f} x \in \mathcal {D} _ {H}. \end{array} \right. \tag {14} +$$ + +According to Eq. 14, for samples with deterministic neighborhood label distributions, we use hard labels to minimize the interference of redundant information on $y_0$ . For samples with more contentious neighborhood label distributions, we adopt a soft strategy for label refinement. Instead of point estimation, we estimate through a label vector, retaining most of the neighborhood information. This strategy enhances the accuracy of the $y_0$ while preserving as much robust information as possible. + +Estimate Noise distribution for $\mathbf{y}_n$ . Similar to obtaining $\mathbf{y}_0$ , we use the subset of divided samples to estimate the noise distribution of the samples. According to Eq. 15, when a sample is identified as clean, the directed diffusion term is used to gradually fade out $\mathbf{y}_0$ ( $\mathbf{y}_d = -\mathbf{y}_0$ ), allowing the model to learn to recover a label closer to the ground truth from an uninformed $\epsilon$ . When the sample contains noise, the endpoint of diffusion should include a clear representation of noise, enabling the model to strictly avoid recovering such noisy vector values during reverse generation. For uncertain samples, we simulate potential noise distributions by retaining the difference in label distributions between two views. The greater the difference in the distribution values for a particular class in the two views, the more likely the sample is incorrectly labeled as that class. + +$$ +P \left(y _ {n} \mid \tilde {y}, x\right) = \left\{ \begin{array}{l l} 0 & \text {i f} x \in \mathcal {D} _ {C}, \\ \tilde {y} & \text {i f} x \in \mathcal {D} _ {N}, \\ \frac {\left| p _ {w} - p _ {s} \right|}{\sum \left| p _ {w} - p _ {s} \right|} & \text {i f} x \in \mathcal {D} _ {H}. \end{array} \right. \tag {15} +$$ + +The complete training process is outlined in Algorithm 1, consisting of the following phases: Steps $1 \sim 3$ for sample re-partitioning, Steps $4 \sim 9$ for diffusion model training, with Step 6 dedicated to $\mathbf{y}_0$ pre-correction and $\mathbf{y}_n$ estimation. Notably, in Step 5, instead of sampling features from the original noisy set $\mathcal{D}$ , we use a mixture of sample features from $\mathcal{D}^w$ and $\mathcal{D}^s$ as the training input for the diffusion + +# Algorithm 1 DLD Training + +Input: Noisy training set $\mathcal{D} = \{\mathbf{X},\tilde{\mathbf{Y}}\}$ , feature extractor $f_{\mathrm{ext}}$ , two network parameters $\theta_{1}$ and $\theta_{2}$ + +Output: $\theta_{1}$ and $\theta_{2}$ + +1: Obtain two augmented features $\mathcal{D}^w$ and $\mathcal{D}^s$ using the feature extractor $f_{\mathrm{ext}}$ . +2: Calculate K-nearest label distributions $p_w$ and $p_s$ . +3: Re-partitioning samples into $\mathcal{D}_C, \mathcal{D}_N$ and $\mathcal{D}_H$ . +4: while not converged do +5: Sample data $(\mathbf{x}_{ws},\tilde{\mathbf{y}})\sim \{\mathcal{D}^w,\mathcal{D}^s\}$ ; time slice $t\sim$ $\{1,\dots ,T\}$ ; and noise $\epsilon \sim \mathcal{N}(0,\mathbf{I})$ +6: Pre-estimate $\mathbf{y}_0$ and $\mathbf{y}_n$ based on Eq. 14 and Eq. 15, respectively. +7: Accumulate losses through Eq. 12 and Eq. 13. +8: Take gradient descent step on the losses to update network parameters. +9: end while + +model, where $\mathbf{x}_{ws} = (1 - \gamma)\mathbf{x}_w + \gamma \mathbf{x}_s$ , with $\gamma = \frac{|\mathcal{D}_H|}{|\mathcal{D}|}$ representing the proportion of uncertain samples, which means that the more difficult the sample classification, the richer the feature conditions the model automatically introduces. This reflects an implicit regularization method that mitigates the issue of the diffusion model being overly confident when still learning incorrect information. + +# 5. Experimental Results and Analysis + +In this section, we first evaluate the DLD's effectiveness on CIFAR datasets with varying levels of IDN in Section 5.1. Then, We conduct ablation studies in Section 5.2 to analyze the contributions of the pre-trained feature extractor, the label pre-correction technology, and other components. In Section 5.3, we compare the DLD to the SOTA methods across four real-world noisy datasets. Appendix E provides more details about our experiments. + +# 5.1. Results on Simulated Noisy Datasets + +We conduct experiments on the CIFAR-10 and CIFAR-100 datasets [27] with simulated IDN [6]. Details of the IDN generation algorithm and examples are provided in Appendix E.1. In our experiments, we generate noise levels of $20\%$ , $40\%$ , and $60\%$ . To obtain the feature space for label pre-correction, we provide multiple configurations of feature extractors, including DLD+ResNet, DLD+SimCLR and DLD+ViT (see Appendix F.2 for detailed settings). + +As shown in Table 1, the DLD+ResNet configuration outperforms the baseline across most noise levels on CIFAR, especially in high-noise environments. We analyze that the main factor affecting its performance in low noise environment is the introduction of too much robust information, which leads to the interference of the model. When using the DLD+ViT configuration, its performance surpasses + +Table 1. Classification accuracies (%) of CIFAR datasets under different levels of IDN. + +
MethodCIFAR-10CIFAR-100
20%40%60%20%40%60%
Co-teaching [17]87.28±0.2078.72±0.4762.51±1.9869.26±0.3959.53±0.5746.80±0.51
GCE [61]86.44±0.2376.71±0.3960.76±3.0868.24±0.2755.76±1.1846.75±2.46
DAC [45]86.16±0.1374.80±0.3259.97±0.9068.70±0.7855.71±1.0146.29±3.35
DMI [54]86.57±0.1677.81±0.8569.94±1.3168.71±1.2157.99±0.9250.96±2.08
SEAL [6]87.79±0.0982.98±0.0570.07±1.5470.61±0.1961.18±0.3551.86±1.32
NCR [22]87.74±0.1583.03±0.2169.44±1.0170.68±0.3761.93±0.5449.13±2.77
CC+DivideMix [62]88.36±0.1184.18±0.4071.52±1.7370.09±0.1964.35±0.3553.20±0.66
DISC [31]90.68±0.1085.61±0.3273.59±1.1371.49±0.1364.46±0.2953.62±0.51
LRA-diffusion+ResNet [4]89.94±0.1586.73±0.3774.66±1.0272.35±0.1465.55±0.3054.43±0.53
DLD+ResNet (Ours)90.43±0.1287.17±0.2976.07±1.1072.60±0.1566.59±0.3155.72±0.24
EPL+ViT [26]95.95±0.1793.01±0.3376.16±1.1872.24±0.1267.20±0.2055.59±0.13
LRA-diffusion+ViT [4]96.65±0.1893.68±0.3478.12±1.1773.59±0.1367.67±0.2157.21±0.12
DLD+ViT (Ours)96.98±0.1296.49±0.0783.14±0.8178.62±0.1176.03±0.1469.57±0.17
+ +EPL+ViT and LRA-diffusion+ViT. This implies that DLD more effectively utilizes the provided prior knowledge without requiring additional training conditions. Across various datasets, DLD's superiority is more pronounced on CIFAR-100, indicating that the more complex IDN on CIFAR-100 severely impacts the performance of other methods, whereas our model maintains robust learning. + +Additionally, we conducted experiments on the simulated CCN dataset. The results (see Appendix E.2) show that CCN is easier to handle than IDN, which aligns with our expectations. Compared to other methods, DLD still achieves the highest classification accuracy. Due to the lightweight architecture of the diffusion model and the single-step pre-correction process, DLD outperforms traditional LNL methods without incurring significant computational overhead. A detailed analysis of training time overhead and model parameters is provided in Appendix E.3. + +# 5.2. Ablation Studies + +Power of label pre-correction. We validate the effectiveness of label pre-correction by freezing the diffusion model. As shown in Figure 3, our pre-correction method consistently outperforms in various noisy environments, while Figure 3b demonstrates that the LRA estimation method disrupts originally correct labels. Then, we unfroze the diffusion model to verify that DLD does not fully depend on the pre-correction method. Results show that DLD further enhances generalization performance based on pre-correction, indicating that the two methods complement each other effectively. Additionally, we design more comprehensive cross-ablation experiments, e.g., replacing the LRA technique in LRA-diffusion with the pre-correction module to demonstrate the necessity of directional diffusion. See Appendix F.1 for more details. + +Contribution of different pre-trained features and other + +Table 2. Classification accuracies (%) on Animal-10N. + +
MethodAccuracy
CE [35]80.6
GCE [61]81.5
NCT [8]84.1
DISC [31]87.3
SSR [14]88.5
LRA-diffusion [4]88.6
J-ViT [9]89.0
DLD (Ours)89.4
+ +training components. We explore the impact of various feature extractors, including ResNet34 [20], SimCLR [7] and ViT-L/14 [40], and conduct ablation studies comparing our method with CARD and LRA-diffusion. Here, DISC [31] served as a baseline that does not require any pretrained models. Detailed results and analyses are provided in Appendix F.2. Our findings reveal that DLD outperforms other diffusion models when combined with different pretrained models, achieving optimal performance when integrated with ViT-L/14. This underscores ViT-L/14's effectiveness in enhancing label pre-correction and facilitating diffusion model learning. Furthermore, Appendix F.3 details additional ablation studies on various hyperparameter settings. We validate the optimal training strategy for the DLD method and recommend this combination. + +# 5.3. Results on Real-world Noisy Datasets + +We select four real-world noisy datasets for our experiments to demonstrate the superior performance of DLD: Animal10N [43], Clothing1M [52], WebVision [30], and ImageNet ILSVRC2012 [41]. Comprehensive descriptions of these datasets and detailed experimental setups can be + +![](images/583fbcb0c2d28e4097d41b09173dd3eece8bb464b0de20c1edf1f9cfafa91618.jpg) +(a) Experimental results on the CIFAR-10 dataset + +![](images/7e01a5932844ab8d099133ec3ebdb8ff0faa62526609a98a5f2823c22774ac80.jpg) +(b) Experimental results on the CIFAR-100 dataset +Figure 3. Comparison of different label correction strategies' accuracies, where Baseline represents the proportion of clean labels. + +Table 3. Classification accuracies (%) on Clothing1M. + +
MethodAccuracy
Co-teaching [17]69.21
DMI [54]72.46
DivideMix [28]74.76
CC [62]75.40
DISC [31]74.79
EPL [26]75.21
LRA-diffusion [4]74.46
SANM [46]75.63
DLD (Ours)75.69
+ +Table 4. Classification accuracies (%) on WebVision and ILSVRC2012 datasets. + +
MethodWebVisionILSVRC12
Co-teaching [17]63.5861.48
DivideMix [28]77.3275.20
ELR [33]77.7870.29
EPL [26]78.7776.51
C2D [63]79.4278.57
CC [62]79.3676.08
DISC [31]80.2877.44
LRA-diffusion [4]84.1682.56
DLD (Ours)84.5183.74
+ +found in Appendix E.4 and Appendix E.5, respectively. As demonstrated in Table 2, Table 3, and Table 4, DLD outperforms all SOTA methods across these datasets. Note that DLD's relative classification performance improved by $7.26\%$ and $0.44\%$ over J-ViT [9] and EPL [26], respectively, both employing the ViT architecture. Compared to LRA-diffusion, which also employs ViT and label diffusion models, DLD achieves varying degrees of performance improvement, demonstrating a more effective and reason- + +able use of prior knowledge and diffusion models. Besides, while ViT-L/14 improved performance on Clothing1M, Appendix G shows that other extractors like CC can be even more effective, highlighting our method's versatility with different LNL methods. Besides, DLD's performance advantage partly stems from the pre-correction method. In future work, we aim to integrate discriminative and generative paradigms, using diffusion models to refine pre-correction, generate label information, and enhance representational power and multimodal generalization. Appendix H discusses additional limitations and future directions. + +# 6. Conclusion + +This paper proposes DLD, a diffusion model specifically designed to learn effectively from noisy labels for robust classification. Our main idea is to decouple the label diffusion process into two channels: directional diffusion and random diffusion. This approach explicitly introduces noisy label information into the diffusion model for the first time, enabling the model to interpretably mitigate the generation of incorrect labels, thereby enhancing its ability to generalize from noisy to clean data. By incorporating our low-cost label pre-correction method, DLD can automatically select label diffusion information for each sample in diverse noisy environments, without requiring additional adjustments or iterations. Additionally, our model synergizes with various pre-trained large-scale models, like ViT, leveraging their prior knowledge more effectively than other methods. Experiments on simulated and real-world noisy datasets show that DLD surpasses conventional LNL methods in both accuracy and efficiency. By framing LNL within a generative model, our method bridges LNL with other learning frameworks. We believe this work will reduce the workload for practitioners handling large-scale, low-quality labeled data, benefiting the development of the image recognition field. + +# Acknowledgements + +This work was supported in part by the National Natural Science Foundation of China under Grant U21A20513, Grant 62476157, Grant 62276161, Grant 62276157 and Grant 62406179, and in part by the Key Research and Development Program of Shanxi Province under Grant 202202020101003 and the Fundamental Research Program of Shanxi Province under Grant 202403021222026. + +# References + +[1] Devansh Arpit, Stanisław Jastrzebski, Nicolas Ballas, David Krueger, Emmanuel Bengio, Maxinder S Kanwal, Tegan Maharaj, Asja Fischer, Aaron Courville, Yoshua Bengio, et al. A closer look at memorization in deep networks. In International conference on machine learning, pages 233-242. PMLR, 2017. 1 +[2] Antonin Berthon, Bo Han, Gang Niu, Tongliang Liu, and Masashi Sugiyama. Confidence scores make instance-dependent label-noise learning possible. In International conference on machine learning, pages 825-836. PMLR, 2021. 1 +[3] Stella Biderman, Usvsn Prashanth, Lintang Sutawika, Hailey Schoelkopf, Quentin Anthony, Shivanshu Purohit, and Edward Raff. Emergent and predictable memorization in large language models. Advances in Neural Information Processing Systems, 36:28072-28090, 2023. 1 +[4] Jian Chen, Ruiyi Zhang, Tong Yu, Rohan Sharma, Zhiqiang Xu, Tong Sun, and Changyou Chen. Label-retrieval augmented diffusion models for learning from noisy labels. Advances in Neural Information Processing Systems, 36, 2024. 2, 3, 7, 8 +[5] Pengfei Chen, Ben Ben Liao, Guangyong Chen, and Shengyu Zhang. Understanding and utilizing deep neural networks trained with noisy labels. In International conference on machine learning, pages 1062-1070. PMLR, 2019. 1 +[6] Pengfei Chen, Junjie Ye, Guangyong Chen, Jingwei Zhao, and Pheng-Ann Heng. Beyond class-conditional assumption: A primary attempt to combat instance-dependent label noise. In Proceedings of the AAAI Conference on Artificial Intelligence, pages 11442–11450, 2021. 1, 2, 6, 7 +[7] Ting Chen, Simon Kornblith, Mohammad Norouzi, and Geoffrey Hinton. A simple framework for contrastive learning of visual representations. In International conference on machine learning, pages 1597-1607. PMLR, 2020. 2, 3, 7 +[8] Yingyi Chen, Shell Xu Hu, Xi Shen, Chunrong Ai, and Johan AK Suykens. Compressing features for learning with noisy labels. IEEE Transactions on Neural Networks and Learning Systems, 35(2):2124-2138, 2022. 7 +[9] Yingyi Chen, Xi Shen, Yahui Liu, Qinghua Tao, and Johan AK Suykens. Jigsaw-vit: Learning jigsaw puzzles in vision transformer. Pattern Recognition Letters, 166:53-60, 2023. 7, 8 +[10] De Cheng, Tongliang Liu, Yixiong Ning, Nannan Wang, Bo Han, Gang Niu, Xinbo Gao, and Masashi Sugiyama. + +Instance-dependent label-noise learning with manifold-regularized transition matrix estimation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 16630-16639, 2022. 1 +[11] Hao Cheng, Zhaowei Zhu, Xingyu Li, Yifei Gong, Xing Sun, and Yang Liu. Learning with instance-dependent label noise: A sample sieve approach. arXiv preprint arXiv:2010.02347, 2020. +[12] Jiacheng Cheng, Tongliang Liu, Kotagiri Ramamohanarao, and Dacheng Tao. Learning with bounded instance and label-dependent label noise. In International conference on machine learning, pages 1789-1799. PMLR, 2020. 1 +[13] C Fabian Benitez-Quiroz, Ramprakash Srinivasan, and Aleix M Martinez. Emotionet: An accurate, real-time algorithm for the automatic annotation of a million facial expressions in the wild. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 5562–5570, 2016. 1 +[14] Chen Feng, Georgios Tzimiropoulos, and Ioannis Patras. Ssr: An efficient and robust framework for learning with unknown label noise. arXiv preprint arXiv:2111.11288, 2021. 7 +[15] Arpit Garg, Cuong Nguyen, Rafael Felix, Thanh-Toan Do, and Gustavo Carneiro. Instance-dependent noisy label learning via graphical modelling. In Proceedings of the IEEE/CVF winter conference on applications of computer vision, pages 2288-2298, 2023. 2 +[16] Aritra Ghosh, Himanshu Kumar, and P Shanti Sastry. Robust loss functions under label noise for deep neural networks. In Proceedings of the AAAI conference on artificial intelligence, 2017. 1, 2 +[17] Bo Han, Quanming Yao, Xingrui Yu, Gang Niu, Miao Xu, Weihua Hu, Ivor Tsang, and Masashi Sugiyama. Co-teaching: Robust training of deep neural networks with extremely noisy labels. Advances in neural information processing systems, 31, 2018. 1, 2, 7, 8 +[18] Xizewen Han, Huangjie Zheng, and Mingyuan Zhou. Card: Classification and regression diffusion models. Advances in Neural Information Processing Systems, 35:18100-18115, 2022. 2, 3 +[19] Kaiming He, Xiangyu Zhang, Shaoqing Ren, and Jian Sun. Delving deep into rectifiers: Surpassing human-level performance on imagenet classification. In Proceedings of the IEEE international conference on computer vision, pages 1026-1034, 2015. 2 +[20] Kaiming He, Xiangyu Zhang, Shaoqing Ren, and Jian Sun. Deep residual learning for image recognition. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 770-778, 2016. 7 +[21] Jonathan Ho, Ajay Jain, and Pieter Abbeel. Denoising diffusion probabilistic models. Advances in neural information processing systems, 33:6840-6851, 2020. 2, 3, 5 +[22] Ahmet Iscen, Jack Valmadre, Anurag Arnab, and Cordelia Schmid. Learning with neighbor consistency for noisy labels. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 4672-4681, 2022. 2, 5, 7 + +[23] Gaoxia Jiang, Jia Zhang, Xuefei Bai, Wenjian Wang, and Deyu Meng. Which is more effective in label noise cleaning, correction or filtering? In Proceedings of the AAAI Conference on Artificial Intelligence, pages 12866-12873, 2024. 2 +[24] Lu Jiang, Zhengyuan Zhou, Thomas Leung, Li-Jia Li, and Li Fei-Fei. Mentornet: Learning data-driven curriculum for very deep neural networks on corrupted labels. In International conference on machine learning, pages 2304–2313. PMLR, 2018. 1 +[25] Diederik P Kingma. Auto-encoding variational bayes. arXiv preprint arXiv:1312.6114, 2013. 2 +[26] Jongwoo Ko, Sumyeong Ahn, and Se-Young Yun. Efficient utilization of pre-trained model for learning with noisy labels. In ICLR 2023 Workshop on Pitfalls of limited data and computation for Trustworthy ML, 2023. 7, 8 +[27] Alex Krizhevsky. Learning multiple layers of features from tiny images. Master's thesis, University of Toronto, 2009. 6 +[28] Junnan Li, Richard Socher, and Steven CH Hoi. Dividemix: Learning with noisy labels as semi-supervised learning. arXiv preprint arXiv:2002.07394, 2020. 1, 2, 8 +[29] Mingchen Li, Mahdi Soltanolkotabi, and Samet Oymak. Gradient descent with early stopping is provably robust to label noise for overparameterized neural networks. In International conference on artificial intelligence and statistics, pages 4313-4324. PMLR, 2020. 1 +[30] Wen Li, Limin Wang, Wei Li, Eirikur Agustsson, and Luc Van Gool. Webvision database: Visual learning and understanding from web data. arXiv preprint arXiv:1708.02862, 2017. 7 +[31] Yifan Li, Hu Han, Shiguang Shan, and Xilin Chen. Disc: Learning from noisy labels via dynamic instance-specific selection and correction. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 24070-24079, 2023. 2, 7, 8 +[32] Jiawei Liu, Qiang Wang, Huijie Fan, Yinong Wang, Yandong Tang, and Liangqiong Qu. Residual denoising diffusion models. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 2773-2783, 2024. 2 +[33] Sheng Liu, Jonathan Niles-Weed, Narges Razavian, and Carlos Fernandez-Granda. Early-learning regularization prevents memorization of noisy labels. Advances in neural information processing systems, 33:20331-20342, 2020. 2, 8 +[34] Yangdi Lu, Yang Bo, and Wenbo He. Co-matching: Combating noisy labels by augmentation anchoring. arXiv preprint arXiv:2103.12814, 2021. 1 +[35] Anqi Mao, Mehryar Mohri, and Yutao Zhong. Cross-entropy loss functions: Theoretical analysis and applications. In International conference on Machine learning, pages 23803–23828. PMLR, 2023. 7 +[36] Duc Tam Nguyen, Chaithanya Kumar Mummadi, Thi Phuong Nhung Ngo, Thi Hoai Phuong Nguyen, Laura Beggel, and Thomas Brox. Self: Learning to filter noisy labels with self-ensembling. arXiv preprint arXiv:1910.01842, 2019. 1 + +[37] Kento Nishi, Yi Ding, Alex Rich, and Tobias Hollerer. Augmentation strategies for learning with noisy labels. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 8022-8031, 2021. 1 +[38] Haim Permuter, Joseph Francos, and Ian Jermyn. A study of gaussian mixture models of color and texture features for image classification and segmentation. Pattern recognition, 39(4):695-706, 2006. 5 +[39] Lutz Prechelt. Early stopping-but when? In Neural Networks: Tricks of the trade, pages 55-69. Springer, 2002. 1 +[40] Alec Radford, Jong Wook Kim, Chris Hallacy, Aditya Ramesh, Gabriel Goh, Sandhini Agarwal, Girish Sastry, Amanda Askell, Pamela Mishkin, Jack Clark, et al. Learning transferable visual models from natural language supervision. In International conference on machine learning, pages 8748-8763. PMLR, 2021. 2, 7 +[41] Olga Russakovsky, Jia Deng, Hao Su, Jonathan Krause, Sanjeev Satheesh, Sean Ma, Zhiheng Huang, Andrej Karpathy, Aditya Khosla, Michael Bernstein, et al. Imagenet large scale visual recognition challenge. International journal of computer vision, 115:211-252, 2015. 7 +[42] Kihyuk Sohn, David Berthelot, Nicholas Carlini, Zizhao Zhang, Han Zhang, Colin A Raffel, Ekin Dogus Cubuk, Alexey Kurakin, and Chun-Liang Li. Fixmatch: Simplifying semi-supervised learning with consistency and confidence. Advances in neural information processing systems, 33:596-608, 2020. 5 +[43] Hwanjun Song, Minseok Kim, and Jae-Gil Lee. Selfie: Refurbishing unclean samples for robust deep learning. In International conference on machine learning, pages 5907-5915. PMLR, 2019. 7 +[44] Jiaming Song, Chenlin Meng, and Stefano Ermon. Denoising diffusion implicit models. arXiv preprint arXiv:2010.02502, 2020. 5 +[45] Sunil Thulasidasan, Tanmoy Bhattacharya, Jeff Bilmes, Gopinath Chennupati, and Jamal Mohd-Yusof. Combating label noise in deep learning using abstention. arXiv preprint arXiv:1905.10964, 2019. 2, 7 +[46] Yuanpeng Tu, Boshen Zhang, Yuxi Li, Liang Liu, Jian Li, Jiangning Zhang, Yabiao Wang, Chengjie Wang, and Cai Rong Zhao. Learning with noisy labels via self-supervised adversarial noisy masking. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 16186-16195, 2023. 8 +[47] Yisen Wang, Weiyang Liu, Xingjun Ma, James Bailey, Hongyuan Zha, Le Song, and Shu-Tao Xia. Iterative learning with open-set noisy labels. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 8688-8696, 2018. 1 +[48] Yisen Wang, Xingjun Ma, Zaiyi Chen, Yuan Luo, Jinfeng Yi, and James Bailey. Symmetric cross entropy for robust learning with noisy labels. In Proceedings of the IEEE/CVF international computer vision and pattern recognition, pages 322-330, 2019. 1 +[49] Hongxin Wei, Lei Feng, Xiangyu Chen, and Bo An. Combating noisy labels by agreement: A joint training method with co-regularization. In Proceedings of the IEEE/CVF con + +ference on computer vision and pattern recognition, pages 13726-13735, 2020. 1 +[50] Pengxiang Wu, Songzhu Zheng, Mayank Goswami, Dimitris Metaxas, and Chao Chen. A topological filter for learning with label noise. Advances in neural information processing systems, 33:21382-21393, 2020. 1 +[51] Xiaobo Xia, Tongliang Liu, Bo Han, Nannan Wang, Mingming Gong, Haifeng Liu, Gang Niu, Dacheng Tao, and Masashi Sugiyama. Part-dependent label noise: Towards instance-dependent label noise. Advances in Neural Information Processing Systems, 33:7597-7610, 2020. 1 +[52] Tong Xiao, Tian Xia, Yi Yang, Chang Huang, and Xiaogang Wang. Learning from massive noisy labeled data for image classification. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 2691-2699, 2015. 7 +[53] Cai Xu, Yilin Zhang, Ziyu Guan, and Wei Zhao. Trusted multi-view learning with label noise. arXiv preprint arXiv:2404.11944, 2024. 5 +[54] Yilun Xu, Peng Cao, Yuqing Kong, and Yizhou Wang. L.dmi: A novel information-theoretic loss function for training deep nets robust to label noise. Advances in neural information processing systems, 32, 2019. 2, 7, 8 +[55] Yan Yan, Rómer Rosales, Glenn Fung, Ramanathan Subramanian, and Jennifer Dy. Learning from multiple annotators with varying expertise. Machine learning, 95:291-327, 2014. 1 +[56] Quanming Yao, Hansi Yang, Bo Han, Gang Niu, and James Tin-Yau Kwok. Searching to exploit memorization effect in learning with noisy labels. In International Conference on Machine Learning, pages 10789-10798. PMLR, 2020. 1 +[57] Yu Yao, Tongliang Liu, Mingming Gong, Bo Han, Gang Niu, and Kun Zhang. Instance-dependent label-noise learning under a structural causal model. Advances in Neural Information Processing Systems, 34:4409-4420, 2021. 1 +[58] Xingrui Yu, Bo Han, Jiangchao Yao, Gang Niu, Ivor Tsang, and Masashi Sugiyama. How does disagreement help generalization against label corruption? In International conference on machine learning, pages 7164-7173. PMLR, 2019. 1 +[59] Chiyuan Zhang, Samy Bengio, Moritz Hardt, Benjamin Recht, and Oriol Vinyals. Understanding deep learning (still) requires rethinking generalization. Communications of the ACM, 64(3):107-115, 2021. 1 +[60] Yikai Zhang, Songzhu Zheng, Pengxiang Wu, Mayank Goswami, and Chao Chen. Learning with feature-dependent label noise: A progressive approach. arXiv preprint arXiv:2103.07756, 2021. 1 +[61] Zhilu Zhang and Mert Sabuncu. Generalized cross entropy loss for training deep neural networks with noisy labels. Advances in neural information processing systems, 31, 2018, 1, 2, 7 +[62] Ganlong Zhao, Guanbin Li, Yipeng Qin, Feng Liu, and Yizhou Yu. Centrality and consistency: two-stage clean samples identification for learning with instance-dependent noisy labels. In European Conference on Computer Vision, pages 21-37. Springer, 2022. 2, 7, 8 + +[63] Evgenii Zheltonozhskii, Chaim Baskin, Avi Mendelson, Alex M Bronstein, and Or Litany. Contrast to divide: Self-supervised pre-training for learning with noisy labels. In Proceedings of the IEEE/CVF Winter Conference on Applications of Computer Vision, pages 1657-1667, 2022. 2, 8 +[64] Songzhu Zheng, Pengxiang Wu, Aman Goswami, Mayank Goswami, Dimitris Metaxas, and Chao Chen. Error-bounded correction of noisy labels. In International Conference on Machine Learning, pages 11447-11457. PMLR, 2020. 1 +[65] Yuanyi Zhong, Haoran Tang, Junkun Chen, Jian Peng, and Yu-Xiong Wang. Is self-supervised learning more robust than supervised learning? arXiv preprint arXiv:2206.05259, 2022. 2 \ No newline at end of file diff --git a/2025/Directional Label Diffusion Model for Learning from Noisy Labels/images.zip b/2025/Directional Label Diffusion Model for Learning from Noisy Labels/images.zip new file mode 100644 index 0000000000000000000000000000000000000000..b2c01275952a4493695ee84f7b1ffcf4281d7b65 --- /dev/null +++ b/2025/Directional Label Diffusion Model for Learning from Noisy Labels/images.zip @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:3b1290c664dac6450bb75001fe2d99358078759c842233bb009f52e2d295fb0f +size 488664 diff --git a/2025/Directional Label Diffusion Model for Learning from Noisy Labels/layout.json b/2025/Directional Label Diffusion Model for Learning from Noisy Labels/layout.json new file mode 100644 index 0000000000000000000000000000000000000000..46f19127c903569a741f039ef30b076e64267a32 --- /dev/null +++ b/2025/Directional Label Diffusion Model for Learning from Noisy Labels/layout.json @@ -0,0 +1,10242 @@ +{ + "pdf_info": [ + { + "para_blocks": [ + { + "bbox": [ + 100, + 103, + 510, + 121 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 100, + 103, + 510, + 121 + ], + "spans": [ + { + "bbox": [ + 100, + 103, + 510, + 121 + ], + "type": "text", + "content": "Directional Label Diffusion Model for Learning from Noisy Labels" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 123, + 142, + 487, + 171 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 123, + 142, + 487, + 171 + ], + "spans": [ + { + "bbox": [ + 123, + 142, + 487, + 171 + ], + "type": "text", + "content": "Senyu Hou" + }, + { + "bbox": [ + 123, + 142, + 487, + 171 + ], + "type": "inline_equation", + "content": "^{1}" + }, + { + "bbox": [ + 123, + 142, + 487, + 171 + ], + "type": "text", + "content": ", Gaoxia Jiang" + }, + { + "bbox": [ + 123, + 142, + 487, + 171 + ], + "type": "inline_equation", + "content": "^{1}" + }, + { + "bbox": [ + 123, + 142, + 487, + 171 + ], + "type": "text", + "content": ", Jia Zhang" + }, + { + "bbox": [ + 123, + 142, + 487, + 171 + ], + "type": "inline_equation", + "content": "^{1}" + }, + { + "bbox": [ + 123, + 142, + 487, + 171 + ], + "type": "text", + "content": ", Shangrong Yang" + }, + { + "bbox": [ + 123, + 142, + 487, + 171 + ], + "type": "inline_equation", + "content": "^{1}" + }, + { + "bbox": [ + 123, + 142, + 487, + 171 + ], + "type": "text", + "content": ", Husheng Guo" + }, + { + "bbox": [ + 123, + 142, + 487, + 171 + ], + "type": "inline_equation", + "content": "^{1}" + }, + { + "bbox": [ + 123, + 142, + 487, + 171 + ], + "type": "text", + "content": ", Yaqing Guo" + }, + { + "bbox": [ + 123, + 142, + 487, + 171 + ], + "type": "inline_equation", + "content": "^{1}" + }, + { + "bbox": [ + 123, + 142, + 487, + 171 + ], + "type": "text", + "content": ", Wenjian Wang" + }, + { + "bbox": [ + 123, + 142, + 487, + 171 + ], + "type": "inline_equation", + "content": "^{2*}" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 101, + 171, + 507, + 213 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 101, + 171, + 507, + 213 + ], + "spans": [ + { + "bbox": [ + 101, + 171, + 507, + 213 + ], + "type": "inline_equation", + "content": "^{1}" + }, + { + "bbox": [ + 101, + 171, + 507, + 213 + ], + "type": "text", + "content": "School of Computer and Information Technology, Shanxi University, China \n" + }, + { + "bbox": [ + 101, + 171, + 507, + 213 + ], + "type": "inline_equation", + "content": "^{2}" + }, + { + "bbox": [ + 101, + 171, + 507, + 213 + ], + "type": "text", + "content": "Key Laboratory of Computational Intelligence and Chinese Information Processing Ministry of Education, Shanxi University, China" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 183, + 215, + 427, + 228 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 183, + 215, + 427, + 228 + ], + "spans": [ + { + "bbox": [ + 183, + 215, + 427, + 228 + ], + "type": "text", + "content": "{202312407005, 202412407035}@email.sxu.edu.cn" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 145, + 229, + 463, + 241 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 145, + 229, + 463, + 241 + ], + "spans": [ + { + "bbox": [ + 145, + 229, + 463, + 241 + ], + "type": "text", + "content": "{jianggaoxia, sryang, guohusheng, yqguo, wjwang}@sxu.edu.cn" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 151, + 269, + 200, + 281 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 151, + 269, + 200, + 281 + ], + "spans": [ + { + "bbox": [ + 151, + 269, + 200, + 281 + ], + "type": "text", + "content": "Abstract" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 55, + 294, + 297, + 629 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 294, + 297, + 629 + ], + "spans": [ + { + "bbox": [ + 55, + 294, + 297, + 629 + ], + "type": "text", + "content": "In image classification, the label quality of training data critically influences model generalization, especially for deep neural networks (DNNs). Traditionally, learning from noisy labels (LNL) can improve the generalization of DNNs through complex architectures or a series of robust techniques, but its performance improvement is limited by the discriminative paradigm. Unlike traditional ways, we resolve the LNL problems from the perspective of robust label generation, based on diffusion models within the generative paradigm. To expand the diffusion model into a robust classifier that explicitly accommodates more noise knowledge, we propose a Directional Label Diffusion (DLD) model. It disentangles the diffusion process into two paths, i.e., directional diffusion and random diffusion. Specifically, directional diffusion simulates the corruption of true labels into a directed noise distribution, prioritizing the removal of likely noise, whereas random diffusion introduces inherent randomness to support label recovery. This architecture enable DLD to gradually infer labels from an initial random state, interpretably diverging from the specified noise distribution. To adapt the model to diverse noisy environments, we design a low-cost label precorrection method that automatically supplies more accurate label information to the diffusion model, without requiring manual intervention or additional iterations. Our approach outperforms state-of-the-art methods on both simulated and real-world noisy datasets. Code is available at https://github.com/SenyuHou/DLD." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 55, + 638, + 135, + 651 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 638, + 135, + 651 + ], + "spans": [ + { + "bbox": [ + 55, + 638, + 135, + 651 + ], + "type": "text", + "content": "1. Introduction" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 55, + 656, + 296, + 693 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 656, + 296, + 693 + ], + "spans": [ + { + "bbox": [ + 55, + 656, + 296, + 693 + ], + "type": "text", + "content": "Deep neural networks (DNNs) have achieved notable success across key areas of machine learning, driven by the growth of datasets and the increase in model parameters." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 313, + 270, + 554, + 400 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 270, + 554, + 400 + ], + "spans": [ + { + "bbox": [ + 313, + 270, + 554, + 400 + ], + "type": "text", + "content": "These models are primarily trained using supervised learning, which relies heavily on large volumes of labeled data [5]. However, real-world data often suffer from low-quality labels due to subjective errors like indistinguishable samples or objective errors such as incorrect label recording, making noisy labels a widespread issue [13, 55]. Studies have shown that the generalization capability of deep neural networks is adversely affected when trained with noisy labels [1, 59]. Consequently, it is crucial to develop methods for learning from noisy labels (LNL) on large-scale but low-quality datasets." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 313, + 403, + 555, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 403, + 555, + 714 + ], + "spans": [ + { + "bbox": [ + 313, + 403, + 555, + 714 + ], + "type": "text", + "content": "Due to the varied distribution of label noise, historical research has classified it into class-conditional noise (CCN) and instance-dependent noise (IDN) [57]. Recent studies indicate that label noise in real datasets predominantly aligns with the IDN assumption, wherein noise is unevenly distributed and more likely in samples with similar features [2, 6, 10–12, 51]. To tackle these complex noise patterns, mainstream LNL methods employ strategies combining filtering and training approaches. Some studies demonstrate a phenomenon known as the memorization effect [1, 3, 29, 56], i.e., DNNs first learn from simple data before fitting to noisy samples. Based on this observation, sample selection strategies utilize confidence levels or prediction loss during training intervals to identify noisy samples [17, 24, 28, 58]. However, this strategy is inherently risky, as the timing of noise memorization varies across datasets and noise distributions, complicating the selection criteria [6, 39]. Consequently, mainline LNL methods can often be integrated with training strategies to enhance model robustness, employing techniques like label filtering [17, 36, 50, 58], label correction [47, 60, 64], robust loss [16, 48, 61], and semi-supervised learning [28]. Among these techniques, semi-supervised learning, by combining multiple approaches [34, 37, 49], often outperforms other LNL methods but is still constrained by the discriminative framework. Although some researches attempt to combine" + } + ] + } + ], + "index": 12 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "spans": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "type": "text", + "content": "CVF" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 145, + 0, + 496, + 37 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 145, + 0, + 496, + 37 + ], + "spans": [ + { + "bbox": [ + 145, + 0, + 496, + 37 + ], + "type": "text", + "content": "This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 67, + 703, + 145, + 712 + ], + "type": "page_footnote", + "angle": 0, + "lines": [ + { + "bbox": [ + 67, + 703, + 145, + 712 + ], + "spans": [ + { + "bbox": [ + 67, + 703, + 145, + 712 + ], + "type": "text", + "content": "*Corresponding author." + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 293, + 749, + 318, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 749, + 318, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 749, + 318, + 757 + ], + "type": "text", + "content": "25738" + } + ] + } + ], + "index": 14 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 0 + }, + { + "para_blocks": [ + { + "bbox": [ + 55, + 72, + 294, + 108 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 72, + 294, + 108 + ], + "spans": [ + { + "bbox": [ + 55, + 72, + 294, + 108 + ], + "type": "text", + "content": "variational autoencoders (VAE) [25] with DivideMix [28] to address label noise [15], most methods remain focused on discriminative models." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 56, + 112, + 296, + 470 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 112, + 296, + 470 + ], + "spans": [ + { + "bbox": [ + 56, + 112, + 296, + 470 + ], + "type": "text", + "content": "In this paper, we shift away from the traditional LNL perspective, proposing robust classification models using label diffusion models that focus on the generative modeling and deeply explores the generative relationship between features and labels. Under this paradigm, labels are generated by gradually refining an initial random guess. The uncertainty modeling in diffusion models inherently meets some robustness requirements, as random label generation can offset some noise interference [4, 18]. However, this type of diffusion process lacks information about noisy labels. Intrinsic uncertainty alone is insufficient to handle complex noisy environments. Inspired by the residual diffusion denoising models (RDDM) [32], we explore a novel Directional Label Diffusion (DLD) model that incorporates noisy label information into the diffusion process. In DLD, we decouple the diffusion process into two channels: directional diffusion and random diffusion (see Figure 1). The directional diffusion guides the initial label towards a more probable noise direction (e.g., a green automobile may be misclassified as a frog or other classes, as indicated by the probability distribution in the third column in Figure 1), while random diffusion retains the original diffusion mechanism. This approach injects information about which categories should not be generated into the label generation paradigm, enhancing the interpretability of the diffusion model's robust inference. Unlike conventional label diffusion, our diffusion process concludes with an endpoint that includes the potential noise distribution of the sample (i.e., " + }, + { + "bbox": [ + 56, + 112, + 296, + 470 + ], + "type": "inline_equation", + "content": "\\mathbf{y}_T = \\mathbf{y}_n + \\epsilon" + }, + { + "bbox": [ + 56, + 112, + 296, + 470 + ], + "type": "text", + "content": ") and employs two independent coefficient schedules to control the diffusion speed of both channels." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 55, + 474, + 296, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 474, + 296, + 713 + ], + "spans": [ + { + "bbox": [ + 55, + 474, + 296, + 713 + ], + "type": "text", + "content": "One challenge is that only noisy labels are available in practice for training, so neither " + }, + { + "bbox": [ + 55, + 474, + 296, + 713 + ], + "type": "inline_equation", + "content": "\\mathbf{y}_0" + }, + { + "bbox": [ + 55, + 474, + 296, + 713 + ], + "type": "text", + "content": " nor " + }, + { + "bbox": [ + 55, + 474, + 296, + 713 + ], + "type": "inline_equation", + "content": "\\mathbf{y}_n" + }, + { + "bbox": [ + 55, + 474, + 296, + 713 + ], + "type": "text", + "content": " can be obtained directly. Few researchers employed neighborhood detection techniques like KNN to re-estimate " + }, + { + "bbox": [ + 55, + 474, + 296, + 713 + ], + "type": "inline_equation", + "content": "\\mathbf{y}_0" + }, + { + "bbox": [ + 55, + 474, + 296, + 713 + ], + "type": "text", + "content": " [4]. However, under the IDN assumption, neighborhood labels of noisy samples often contain similar errors, making naive label estimation methods unreliable and potentially causing " + }, + { + "bbox": [ + 55, + 474, + 296, + 713 + ], + "type": "inline_equation", + "content": "\\mathbf{y}_0" + }, + { + "bbox": [ + 55, + 474, + 296, + 713 + ], + "type": "text", + "content": " to deviate from the ground truth (see Figure 3). To address these issues, we further propose a straightforward and efficient label pre-correction method that automatically provides diverse label information for diffusion model training. This framework employs feature extractors to identify the neighborhood label distribution of samples. Note that these extractors are either trained with self-supervised methods [7, 19, 65], which require few additional training iterations, or use pre-trained models without further training, thus remaining unaffected by label noise. Additionally, this pre-correction strategy allows the seamless integration of pre-trained large-scale models such as Sim-CLR [7] and vision transformer (ViT) [40] into our framework to further" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 313, + 72, + 553, + 95 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 72, + 553, + 95 + ], + "spans": [ + { + "bbox": [ + 313, + 72, + 553, + 95 + ], + "type": "text", + "content": "boost the performance. The main contributions of this work are summarized as follows:" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 314, + 96, + 553, + 298 + ], + "type": "list", + "angle": 0, + "index": 8, + "blocks": [ + { + "bbox": [ + 314, + 96, + 553, + 144 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 314, + 96, + 553, + 144 + ], + "spans": [ + { + "bbox": [ + 314, + 96, + 553, + 144 + ], + "type": "text", + "content": "- We propose a novel Directional Label Diffusion (DLD) model that explicitly introduces interpretable noise information through a dual diffusion architecture, enabling robust learning of label generation patterns from noisy data." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 315, + 144, + 553, + 215 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 315, + 144, + 553, + 215 + ], + "spans": [ + { + "bbox": [ + 315, + 144, + 553, + 215 + ], + "type": "text", + "content": "- We design a fast label pre-correction method that automatically provides more accurate label information for the diffusion process and richer feature conditions for the generation process. By leveraging the consistency principle, this method pre-estimates the label distribution in a single iteration without incurring additional costs." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 315, + 216, + 553, + 251 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 315, + 216, + 553, + 251 + ], + "spans": [ + { + "bbox": [ + 315, + 216, + 553, + 251 + ], + "type": "text", + "content": "- We refine the paradigm for introducing feature conditions with rigorous theoretical deduction and provide complementary training strategies to enhance robustness further." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 315, + 251, + 553, + 298 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 315, + 251, + 553, + 298 + ], + "spans": [ + { + "bbox": [ + 315, + 251, + 553, + 298 + ], + "type": "text", + "content": "- We experimentally show that our model surpasses the state-of-the-art (SOTA) methods on both simulated and real-world noisy datasets, e.g., " + }, + { + "bbox": [ + 315, + 251, + 553, + 298 + ], + "type": "inline_equation", + "content": "6\\% \\sim 21\\%" + }, + { + "bbox": [ + 315, + 251, + 553, + 298 + ], + "type": "text", + "content": " accuracy improvement on the CIFAR-100 dataset with IDN." + } + ] + } + ], + "index": 7 + } + ], + "sub_type": "text" + }, + { + "bbox": [ + 314, + 308, + 400, + 319 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 314, + 308, + 400, + 319 + ], + "spans": [ + { + "bbox": [ + 314, + 308, + 400, + 319 + ], + "type": "text", + "content": "2. Related Work" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 314, + 324, + 468, + 337 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 314, + 324, + 468, + 337 + ], + "spans": [ + { + "bbox": [ + 314, + 324, + 468, + 337 + ], + "type": "text", + "content": "2.1. Learning from Noisy Labels" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 313, + 342, + 555, + 653 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 342, + 555, + 653 + ], + "spans": [ + { + "bbox": [ + 313, + 342, + 555, + 653 + ], + "type": "text", + "content": "Screening and training strategies are central to LNL methods, aiming to distinguish noisy from clean samples and to optimize model learning under noisy conditions. Screening strategies, like Co-teaching [17], select low-loss samples as clean, while DivideMix [28] uses a Gaussian mixture model to automatically identify clean samples during training. Confidence-based methods like CC [62] and DISC [31] classify samples with unreliable predictions as noisy, whereas neighbor consistency regularization (NCR) [22] promotes learning from samples with high neighbor consistency. Our pre-correction method is based on similarity constraints across multi-view label distributions and leverages a feature extractor to provide high-quality information to the diffusion model without requiring additional iterative training. For training, models such as SEAL [6] iteratively correct noisy labels, but this often risks worsening outcomes [23]. Robust loss functions, like mean absolute error (MAE) [16] and generalized cross entropy (GCE) [61], enhance resilience to noisy labels, while DAC [45] and DMI [54] offer advanced mutual information-based losses. Approaches like DivideMix [28] treat LNL as semi-supervised learning, supported by methods like CC [62], C2D [63], and ELR [33] to refine the process. Without complex iterations or loss design, our method utilizes a lightweight diffusion network and a simple L2 loss to quickly learn the feature-label generation relationship." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 314, + 660, + 495, + 673 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 314, + 660, + 495, + 673 + ], + "spans": [ + { + "bbox": [ + 314, + 660, + 495, + 673 + ], + "type": "text", + "content": "2.2. Label Denoising Diffusion Models" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 313, + 677, + 555, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 677, + 555, + 713 + ], + "spans": [ + { + "bbox": [ + 313, + 677, + 555, + 713 + ], + "type": "text", + "content": "Diffusion models were initially designed for image generation tasks and can be understood through the lenses of score matching and dynamics. Ho et al. [21] introduced a" + } + ] + } + ], + "index": 13 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "text", + "content": "25739" + } + ] + } + ], + "index": 14 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 1 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 82, + 70, + 531, + 238 + ], + "blocks": [ + { + "bbox": [ + 82, + 70, + 531, + 238 + ], + "lines": [ + { + "bbox": [ + 82, + 70, + 531, + 238 + ], + "spans": [ + { + "bbox": [ + 82, + 70, + 531, + 238 + ], + "type": "image", + "image_path": "c07b814ed34be5d0f57d0b65e23ec6f71e12a29dad6bf275b3b5b1a6c5210b02.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 54, + 242, + 555, + 298 + ], + "lines": [ + { + "bbox": [ + 54, + 242, + 555, + 298 + ], + "spans": [ + { + "bbox": [ + 54, + 242, + 555, + 298 + ], + "type": "text", + "content": "Figure 1. The proposed Directional Label Diffusion (DLD) model includes diagrams illustrating both the forward and reverse processes. In the forward process, diffusion is decoupled into two channels: random diffusion and directional diffusion. Random diffusion represents the transformation of " + }, + { + "bbox": [ + 54, + 242, + 555, + 298 + ], + "type": "inline_equation", + "content": "\\mathbf{y}_0" + }, + { + "bbox": [ + 54, + 242, + 555, + 298 + ], + "type": "text", + "content": " into random Gaussian noise " + }, + { + "bbox": [ + 54, + 242, + 555, + 298 + ], + "type": "inline_equation", + "content": "\\epsilon" + }, + { + "bbox": [ + 54, + 242, + 555, + 298 + ], + "type": "text", + "content": ", while directional diffusion describes the transformation of " + }, + { + "bbox": [ + 54, + 242, + 555, + 298 + ], + "type": "inline_equation", + "content": "\\mathbf{y}_0" + }, + { + "bbox": [ + 54, + 242, + 555, + 298 + ], + "type": "text", + "content": " into a specific sample noise label distribution " + }, + { + "bbox": [ + 54, + 242, + 555, + 298 + ], + "type": "inline_equation", + "content": "\\mathbf{y}_n" + }, + { + "bbox": [ + 54, + 242, + 555, + 298 + ], + "type": "text", + "content": ". Ultimately, " + }, + { + "bbox": [ + 54, + 242, + 555, + 298 + ], + "type": "inline_equation", + "content": "\\mathbf{y}_0" + }, + { + "bbox": [ + 54, + 242, + 555, + 298 + ], + "type": "text", + "content": " diffuses to the endpoint " + }, + { + "bbox": [ + 54, + 242, + 555, + 298 + ], + "type": "inline_equation", + "content": "\\mathbf{y}_T = \\mathbf{y}_0 + \\epsilon" + }, + { + "bbox": [ + 54, + 242, + 555, + 298 + ], + "type": "text", + "content": ". Notably, the features of the samples serve as control conditions, guiding the generation of labels only during the reverse process." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "bbox": [ + 56, + 318, + 296, + 520 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 318, + 296, + 520 + ], + "spans": [ + { + "bbox": [ + 56, + 318, + 296, + 520 + ], + "type": "text", + "content": "more classical interpretative framework known as denoising diffusion probabilistic models (DDPM), which have been widely applied across diverse domains. RDDM expands the denoising-based diffusion models, initially uninterpretable for image restoration, into a unified and interpretable model for both image generation and restoration by introducing residuals. Classification and regression diffusion models (CARD) [18] reconstruct the classification as a conditional generative task, generating labels from dimensionally reduced image features via a pre-trained encoder. Building on CARD, label-retrieval-augmented (LRA) diffusion model [4] re-estimates the generation targets through neighbor label retrieval, improving label accuracy. Our model disentangles the diffusion process into directional and random diffusion channels, enhancing diffusion quality through precorrection, which leads to train a more robust label diffusion model." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 55, + 527, + 133, + 541 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 527, + 133, + 541 + ], + "spans": [ + { + "bbox": [ + 55, + 527, + 133, + 541 + ], + "type": "text", + "content": "3. Preliminary" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 55, + 544, + 295, + 629 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 544, + 295, + 629 + ], + "spans": [ + { + "bbox": [ + 55, + 544, + 295, + 629 + ], + "type": "text", + "content": "In this section, we primarily introduce the label denoising diffusion models (LDDM), which is constructed based on the theoretical framework of the DDPM [21]. We start by defining the data distribution " + }, + { + "bbox": [ + 55, + 544, + 295, + 629 + ], + "type": "inline_equation", + "content": "\\mathbf{y}_0\\sim q(\\mathbf{y}_0)" + }, + { + "bbox": [ + 55, + 544, + 295, + 629 + ], + "type": "text", + "content": " and a Markovian forward (diffusion) process " + }, + { + "bbox": [ + 55, + 544, + 295, + 629 + ], + "type": "inline_equation", + "content": "q" + }, + { + "bbox": [ + 55, + 544, + 295, + 629 + ], + "type": "text", + "content": " , which progressively adds Gaussian noise " + }, + { + "bbox": [ + 55, + 544, + 295, + 629 + ], + "type": "inline_equation", + "content": "\\epsilon \\sim \\mathcal{N}(0,\\mathbf{I})" + }, + { + "bbox": [ + 55, + 544, + 295, + 629 + ], + "type": "text", + "content": " to the labels based on diffusion strength " + }, + { + "bbox": [ + 55, + 544, + 295, + 629 + ], + "type": "inline_equation", + "content": "\\beta_{t}" + }, + { + "bbox": [ + 55, + 544, + 295, + 629 + ], + "type": "text", + "content": " , diffusing label " + }, + { + "bbox": [ + 55, + 544, + 295, + 629 + ], + "type": "inline_equation", + "content": "\\mathbf{y}_0" + }, + { + "bbox": [ + 55, + 544, + 295, + 629 + ], + "type": "text", + "content": " to " + }, + { + "bbox": [ + 55, + 544, + 295, + 629 + ], + "type": "inline_equation", + "content": "\\mathbf{y}_T" + }, + { + "bbox": [ + 55, + 544, + 295, + 629 + ], + "type": "text", + "content": " .." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 83, + 632, + 295, + 653 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 83, + 632, + 295, + 653 + ], + "spans": [ + { + "bbox": [ + 83, + 632, + 295, + 653 + ], + "type": "interline_equation", + "content": "q \\left(\\mathbf {y} _ {t} \\mid \\mathbf {y} _ {t - 1}\\right) = \\mathcal {N} \\left(\\mathbf {y} _ {t}; \\sqrt {1 - \\beta_ {t}} \\mathbf {y} _ {t - 1}, \\beta_ {t} \\mathbf {I}\\right), \\tag {1}", + "image_path": "64b10567c0e29611aad7f8b893cf8dffa2524779ecf3bb6781a637f9ec1bfaf9.jpg" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 55, + 655, + 296, + 693 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 655, + 296, + 693 + ], + "spans": [ + { + "bbox": [ + 55, + 655, + 296, + 693 + ], + "type": "text", + "content": "Under Bayes' theorem, one finds that the posterior distribution of the forward process can be represented as a Gaussian distribution with mean " + }, + { + "bbox": [ + 55, + 655, + 296, + 693 + ], + "type": "inline_equation", + "content": "\\tilde{\\mu}_t(\\mathbf{y}_t,\\mathbf{y}_0)" + }, + { + "bbox": [ + 55, + 655, + 296, + 693 + ], + "type": "text", + "content": " and variance " + }, + { + "bbox": [ + 55, + 655, + 296, + 693 + ], + "type": "inline_equation", + "content": "\\tilde{\\beta}_t" + }, + { + "bbox": [ + 55, + 655, + 296, + 693 + ], + "type": "text", + "content": " :" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 70, + 696, + 295, + 717 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 70, + 696, + 295, + 717 + ], + "spans": [ + { + "bbox": [ + 70, + 696, + 295, + 717 + ], + "type": "interline_equation", + "content": "q \\left(\\mathbf {y} _ {t - 1} \\mid \\mathbf {y} _ {t}, \\mathbf {y} _ {0}\\right) = \\mathcal {N} \\left(\\mathbf {y} _ {t - 1}; \\tilde {\\mu} _ {t} \\left(\\mathbf {y} _ {t}, \\mathbf {y} _ {0}\\right), \\tilde {\\beta} _ {t} \\mathbf {I}\\right), \\tag {2}", + "image_path": "edf63fa15ef20157a3fd24203f4f58eec86984bcde0bce6979d8a44d806960a2.jpg" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 313, + 316, + 555, + 417 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 316, + 555, + 417 + ], + "spans": [ + { + "bbox": [ + 313, + 316, + 555, + 417 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 313, + 316, + 555, + 417 + ], + "type": "inline_equation", + "content": "\\tilde{\\mu}_t(\\mathbf{y}_t,\\mathbf{y}_0) = \\frac{\\sqrt{\\bar{\\alpha}_{t - 1}}\\beta_t}{1 - \\bar{\\alpha}_t}\\mathbf{y}_0 + \\frac{\\sqrt{\\alpha_t(1 - \\bar{\\alpha}_{t - 1})}}{1 - \\bar{\\alpha}_t}\\mathbf{y}_t" + }, + { + "bbox": [ + 313, + 316, + 555, + 417 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 313, + 316, + 555, + 417 + ], + "type": "inline_equation", + "content": "\\tilde{\\beta}_t = \\frac{1 - \\bar{\\alpha}_{t - 1}}{1 - \\bar{\\alpha}_t}\\beta_t" + }, + { + "bbox": [ + 313, + 316, + 555, + 417 + ], + "type": "text", + "content": ", In the reverse process, we naturally employ a neural network to perform Markovian inference from " + }, + { + "bbox": [ + 313, + 316, + 555, + 417 + ], + "type": "inline_equation", + "content": "q(\\mathbf{y}_T)" + }, + { + "bbox": [ + 313, + 316, + 555, + 417 + ], + "type": "text", + "content": ", fitting the posterior distribution " + }, + { + "bbox": [ + 313, + 316, + 555, + 417 + ], + "type": "inline_equation", + "content": "q(\\mathbf{y}_{t - 1}|\\mathbf{y}_t)" + }, + { + "bbox": [ + 313, + 316, + 555, + 417 + ], + "type": "text", + "content": " through to " + }, + { + "bbox": [ + 313, + 316, + 555, + 417 + ], + "type": "inline_equation", + "content": "\\mathbf{y}_0" + }, + { + "bbox": [ + 313, + 316, + 555, + 417 + ], + "type": "text", + "content": ". Since the variance of the posterior Gaussian distribution depends only on the diffusion strength corresponding to time " + }, + { + "bbox": [ + 313, + 316, + 555, + 417 + ], + "type": "inline_equation", + "content": "t" + }, + { + "bbox": [ + 313, + 316, + 555, + 417 + ], + "type": "text", + "content": ", the neural network focuses primarily on predicting and learning the mean:" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 340, + 426, + 555, + 447 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 340, + 426, + 555, + 447 + ], + "spans": [ + { + "bbox": [ + 340, + 426, + 555, + 447 + ], + "type": "interline_equation", + "content": "p _ {\\theta} \\left(\\mathbf {y} _ {t - 1} \\mid \\mathbf {y} _ {t}\\right) := \\mathcal {N} \\left(\\mathbf {y} _ {t - 1}; \\mu_ {\\theta} \\left(\\mathbf {y} _ {t}, t\\right), \\tilde {\\beta} _ {t} \\mathbf {I}\\right). \\tag {3}", + "image_path": "9ac913f93476ecaf669bcf363091a73a4121dfcef51ace42f2cd758a35c6b5ad.jpg" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 313, + 454, + 554, + 537 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 454, + 554, + 537 + ], + "spans": [ + { + "bbox": [ + 313, + 454, + 554, + 537 + ], + "type": "text", + "content": "To train this model such that " + }, + { + "bbox": [ + 313, + 454, + 554, + 537 + ], + "type": "inline_equation", + "content": "p_{\\theta}(\\mathbf{y}_0)" + }, + { + "bbox": [ + 313, + 454, + 554, + 537 + ], + "type": "text", + "content": " closely approximates " + }, + { + "bbox": [ + 313, + 454, + 554, + 537 + ], + "type": "inline_equation", + "content": "q(\\mathbf{y}_0)" + }, + { + "bbox": [ + 313, + 454, + 554, + 537 + ], + "type": "text", + "content": ", we can optimize the variational lower bound (VLB), however, research [7] found that directly predicting the mean was not effective, so they reparameterized the estimation target, constraining the neural network to estimate the actual noise of the diffusion process and simplified the loss function:" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 315, + 544, + 553, + 574 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 315, + 544, + 553, + 574 + ], + "spans": [ + { + "bbox": [ + 315, + 544, + 553, + 574 + ], + "type": "interline_equation", + "content": "L _ {\\text {s i m p l e}} := E _ {t \\sim [ 1, T ], \\mathbf {y} _ {0} \\sim q (\\mathbf {y} _ {0}), \\epsilon \\sim \\mathcal {N} (0, \\mathbf {I})} \\left[ \\| \\epsilon - \\epsilon_ {\\theta} (\\mathbf {y} _ {t}, t) \\| ^ {2} \\right]. \\tag {4}", + "image_path": "0a4cd8d06da166a4aa9435998621a30c3270e40f3815acfb8c0e2e51e3d3a7ca.jpg" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 313, + 574, + 554, + 623 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 574, + 554, + 623 + ], + "spans": [ + { + "bbox": [ + 313, + 574, + 554, + 623 + ], + "type": "text", + "content": "During inference, we utilize the label noise estimated by the model, restoring the mean of the label at time " + }, + { + "bbox": [ + 313, + 574, + 554, + 623 + ], + "type": "inline_equation", + "content": "t - 1" + }, + { + "bbox": [ + 313, + 574, + 554, + 623 + ], + "type": "text", + "content": " to retrieve its distribution form, and then gradually performing random sampling until generating the label " + }, + { + "bbox": [ + 313, + 574, + 554, + 623 + ], + "type": "inline_equation", + "content": "\\mathbf{y}_0" + }, + { + "bbox": [ + 313, + 574, + 554, + 623 + ], + "type": "text", + "content": ":" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 323, + 630, + 555, + 658 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 323, + 630, + 555, + 658 + ], + "spans": [ + { + "bbox": [ + 323, + 630, + 555, + 658 + ], + "type": "interline_equation", + "content": "\\mathbf {y} _ {t - 1} = \\frac {1}{\\sqrt {\\alpha_ {t}}} \\left(\\mathbf {y} _ {t} - \\frac {1 - \\alpha_ {t}}{\\sqrt {1 - \\bar {\\alpha} _ {t}}} \\epsilon_ {\\theta} (\\mathbf {y} _ {t}, t)\\right) + \\sqrt {\\tilde {\\beta} _ {t}} \\mathbf {z}, \\tag {5}", + "image_path": "19007f4685932d76f04e2689045558c74405f649227e5f8cb3b96423965df6d8.jpg" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 313, + 665, + 555, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 665, + 555, + 713 + ], + "spans": [ + { + "bbox": [ + 313, + 665, + 555, + 713 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 313, + 665, + 555, + 713 + ], + "type": "inline_equation", + "content": "\\mathbf{z} \\sim \\mathcal{N}(0, \\mathbf{I})" + }, + { + "bbox": [ + 313, + 665, + 555, + 713 + ], + "type": "text", + "content": ". Unfortunately, this label generation process lacks any conditional information, resulting in random label generation for images with identical features, which fails to achieve classification. To address this, we" + } + ] + } + ], + "index": 14 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "type": "text", + "content": "25740" + } + ] + } + ], + "index": 15 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 2 + }, + { + "para_blocks": [ + { + "bbox": [ + 55, + 72, + 297, + 109 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 72, + 297, + 109 + ], + "spans": [ + { + "bbox": [ + 55, + 72, + 297, + 109 + ], + "type": "text", + "content": "analyze the foundational paradigm of conditional label diffusion models in Appendix A and simplify the method of feature conditioning." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 55, + 118, + 247, + 131 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 118, + 247, + 131 + ], + "spans": [ + { + "bbox": [ + 55, + 118, + 247, + 131 + ], + "type": "text", + "content": "4. Directional Label Diffusion Model" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 55, + 135, + 296, + 231 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 135, + 296, + 231 + ], + "spans": [ + { + "bbox": [ + 55, + 135, + 296, + 231 + ], + "type": "text", + "content": "Based on LDDM, our goal is to develop a directional diffusion process on labels to introduce more knowledge about noise. We modify the representation of " + }, + { + "bbox": [ + 55, + 135, + 296, + 231 + ], + "type": "inline_equation", + "content": "\\mathbf{y}_T = \\epsilon" + }, + { + "bbox": [ + 55, + 135, + 296, + 231 + ], + "type": "text", + "content": " in LDDM to " + }, + { + "bbox": [ + 55, + 135, + 296, + 231 + ], + "type": "inline_equation", + "content": "\\mathbf{y}_T = \\mathbf{y}_n + \\epsilon" + }, + { + "bbox": [ + 55, + 135, + 296, + 231 + ], + "type": "text", + "content": " in our DLD, where " + }, + { + "bbox": [ + 55, + 135, + 296, + 231 + ], + "type": "inline_equation", + "content": "\\mathbf{y}_n" + }, + { + "bbox": [ + 55, + 135, + 296, + 231 + ], + "type": "text", + "content": " is an estimation of noisy label distribution. The improved forward process is no longer a simple random noise diffusion but is disentangled into two paths of directional diffusion and random diffusion, as shown in Figure 1." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 55, + 232, + 296, + 339 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 232, + 296, + 339 + ], + "spans": [ + { + "bbox": [ + 55, + 232, + 296, + 339 + ], + "type": "text", + "content": "In the following subsections, we first outline the overall framework of our model in Section 4.1. Next, in Section 4.2, we present the foundational theory of the DLD diffusion process and analyze the method of conditional control. In Section 4.3, we describe the reverse process and training method, including optimizations for introducing feature conditions and an adapted accelerated inference method. Finally, in Section 4.4, we describe the rapid label pre-correction method within the DLD framework." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 55, + 348, + 156, + 360 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 348, + 156, + 360 + ], + "spans": [ + { + "bbox": [ + 55, + 348, + 156, + 360 + ], + "type": "text", + "content": "4.1. Model Overview" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 55, + 365, + 296, + 582 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 365, + 296, + 582 + ], + "spans": [ + { + "bbox": [ + 55, + 365, + 296, + 582 + ], + "type": "text", + "content": "We propose the Directional Label Diffusion (DLD) framework, with the overall process illustrated in Figure 2. The framework contains two main stages: label pre-correction (Stage 1) and the directional label diffusion model (Stage 2). In stage 1, we first use two different data augmentation methods to provide two views for each sample, then use a pre-trained feature extractor " + }, + { + "bbox": [ + 55, + 365, + 296, + 582 + ], + "type": "inline_equation", + "content": "f_{ext}" + }, + { + "bbox": [ + 55, + 365, + 296, + 582 + ], + "type": "text", + "content": " to observe the neighborhood label distribution " + }, + { + "bbox": [ + 55, + 365, + 296, + 582 + ], + "type": "inline_equation", + "content": "p_w" + }, + { + "bbox": [ + 55, + 365, + 296, + 582 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 55, + 365, + 296, + 582 + ], + "type": "inline_equation", + "content": "p_s" + }, + { + "bbox": [ + 55, + 365, + 296, + 582 + ], + "type": "text", + "content": " of the same sample in two feature spaces. Based on similarity constraints and the principle of neighborhood consistency, we divide the dataset into clean, noisy, and hard sets. Then, the label precorrector and noise distribution estimator prepare " + }, + { + "bbox": [ + 55, + 365, + 296, + 582 + ], + "type": "inline_equation", + "content": "\\mathbf{y}_0" + }, + { + "bbox": [ + 55, + 365, + 296, + 582 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 55, + 365, + 296, + 582 + ], + "type": "inline_equation", + "content": "\\mathbf{y}_n" + }, + { + "bbox": [ + 55, + 365, + 296, + 582 + ], + "type": "text", + "content": " based on the sample partitioning results, using " + }, + { + "bbox": [ + 55, + 365, + 296, + 582 + ], + "type": "inline_equation", + "content": "\\mathbf{y}_0" + }, + { + "bbox": [ + 55, + 365, + 296, + 582 + ], + "type": "text", + "content": " as the initial diffusion label and " + }, + { + "bbox": [ + 55, + 365, + 296, + 582 + ], + "type": "inline_equation", + "content": "\\mathbf{y}_n" + }, + { + "bbox": [ + 55, + 365, + 296, + 582 + ], + "type": "text", + "content": " as part of the diffusion endpoint for the next stage of model training. Otherwise, we introduce both augmented image features and high-quality features extracted by " + }, + { + "bbox": [ + 55, + 365, + 296, + 582 + ], + "type": "inline_equation", + "content": "f_{ext}" + }, + { + "bbox": [ + 55, + 365, + 296, + 582 + ], + "type": "text", + "content": " into the training process to jointly guide the learning of the robust diffusion model." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 55, + 590, + 195, + 601 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 590, + 195, + 601 + ], + "spans": [ + { + "bbox": [ + 55, + 590, + 195, + 601 + ], + "type": "text", + "content": "4.2. Forward Process of DLD" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 55, + 608, + 296, + 644 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 608, + 296, + 644 + ], + "spans": [ + { + "bbox": [ + 55, + 608, + 296, + 644 + ], + "type": "text", + "content": "To model the random diffusion of labels and the addition of directional noise information, we define the single forward (diffusion) process step in our DLD as follows:" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 121, + 655, + 295, + 667 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 121, + 655, + 295, + 667 + ], + "spans": [ + { + "bbox": [ + 121, + 655, + 295, + 667 + ], + "type": "interline_equation", + "content": "\\mathbf {y} _ {t} = \\mathbf {y} _ {t - 1} + \\alpha_ {t} \\mathbf {y} _ {d} + \\beta_ {t} \\epsilon , \\tag {6}", + "image_path": "51b6a38ec320ebe2463414eb7e5d3ceaab82129a49925690a4a2d72b42991bfa.jpg" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 55, + 677, + 296, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 677, + 296, + 714 + ], + "spans": [ + { + "bbox": [ + 55, + 677, + 296, + 714 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 55, + 677, + 296, + 714 + ], + "type": "inline_equation", + "content": "\\mathbf{y}_d" + }, + { + "bbox": [ + 55, + 677, + 296, + 714 + ], + "type": "text", + "content": " is the deviation between " + }, + { + "bbox": [ + 55, + 677, + 296, + 714 + ], + "type": "inline_equation", + "content": "\\mathbf{y}_n" + }, + { + "bbox": [ + 55, + 677, + 296, + 714 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 55, + 677, + 296, + 714 + ], + "type": "inline_equation", + "content": "y_0" + }, + { + "bbox": [ + 55, + 677, + 296, + 714 + ], + "type": "text", + "content": " (i.e., " + }, + { + "bbox": [ + 55, + 677, + 296, + 714 + ], + "type": "inline_equation", + "content": "\\mathbf{y}_d = \\mathbf{y}_n - \\mathbf{y}_0" + }, + { + "bbox": [ + 55, + 677, + 296, + 714 + ], + "type": "text", + "content": "), representing the directional shift contained in the diffusion from " + }, + { + "bbox": [ + 55, + 677, + 296, + 714 + ], + "type": "inline_equation", + "content": "\\mathbf{y}_{t-1}" + }, + { + "bbox": [ + 55, + 677, + 296, + 714 + ], + "type": "text", + "content": " to " + }, + { + "bbox": [ + 55, + 677, + 296, + 714 + ], + "type": "inline_equation", + "content": "\\mathbf{y}_t" + }, + { + "bbox": [ + 55, + 677, + 296, + 714 + ], + "type": "text", + "content": ". Two independent coefficients, " + }, + { + "bbox": [ + 55, + 677, + 296, + 714 + ], + "type": "inline_equation", + "content": "\\alpha_t" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 313, + 72, + 554, + 133 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 72, + 554, + 133 + ], + "spans": [ + { + "bbox": [ + 313, + 72, + 554, + 133 + ], + "type": "text", + "content": "and " + }, + { + "bbox": [ + 313, + 72, + 554, + 133 + ], + "type": "inline_equation", + "content": "\\beta_{t}" + }, + { + "bbox": [ + 313, + 72, + 554, + 133 + ], + "type": "text", + "content": " control the degrees of directional and random diffusion, respectively. Based on the additivity of Gaussian noise, we derive the jump diffusion formula from " + }, + { + "bbox": [ + 313, + 72, + 554, + 133 + ], + "type": "inline_equation", + "content": "\\mathbf{y}_0" + }, + { + "bbox": [ + 313, + 72, + 554, + 133 + ], + "type": "text", + "content": " to " + }, + { + "bbox": [ + 313, + 72, + 554, + 133 + ], + "type": "inline_equation", + "content": "\\mathbf{y}_t" + }, + { + "bbox": [ + 313, + 72, + 554, + 133 + ], + "type": "text", + "content": " without requiring single-step continuous diffusion (see Appendix B for the proof):" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 384, + 140, + 555, + 155 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 384, + 140, + 555, + 155 + ], + "spans": [ + { + "bbox": [ + 384, + 140, + 555, + 155 + ], + "type": "interline_equation", + "content": "\\mathbf {y} _ {t} = \\mathbf {y} _ {0} + \\bar {\\alpha} _ {t} \\mathbf {y} _ {d} + \\bar {\\beta} _ {t} \\epsilon , \\tag {7}", + "image_path": "78f27ee2878a35be1d20032f92f136378f15228bb3b1f86cfc67a41895894d44.jpg" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 313, + 164, + 555, + 289 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 164, + 555, + 289 + ], + "spans": [ + { + "bbox": [ + 313, + 164, + 555, + 289 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 313, + 164, + 555, + 289 + ], + "type": "inline_equation", + "content": "\\bar{\\alpha}_t = \\sum_{i=1}^t \\alpha_i" + }, + { + "bbox": [ + 313, + 164, + 555, + 289 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 313, + 164, + 555, + 289 + ], + "type": "inline_equation", + "content": "\\bar{\\beta}_t = \\sqrt{\\sum_{i=1}^t \\beta_i^2}" + }, + { + "bbox": [ + 313, + 164, + 555, + 289 + ], + "type": "text", + "content": ". When " + }, + { + "bbox": [ + 313, + 164, + 555, + 289 + ], + "type": "inline_equation", + "content": "t = T" + }, + { + "bbox": [ + 313, + 164, + 555, + 289 + ], + "type": "text", + "content": ", we hope " + }, + { + "bbox": [ + 313, + 164, + 555, + 289 + ], + "type": "inline_equation", + "content": "\\mathbf{y}_T = \\mathbf{y}_n + \\epsilon = \\mathbf{y}_0 + \\mathbf{y}_d + \\epsilon" + }, + { + "bbox": [ + 313, + 164, + 555, + 289 + ], + "type": "text", + "content": ", so we usually set " + }, + { + "bbox": [ + 313, + 164, + 555, + 289 + ], + "type": "inline_equation", + "content": "\\bar{\\alpha}_T = \\bar{\\beta}_T = 1" + }, + { + "bbox": [ + 313, + 164, + 555, + 289 + ], + "type": "text", + "content": ". This implies that the two coefficients " + }, + { + "bbox": [ + 313, + 164, + 555, + 289 + ], + "type": "inline_equation", + "content": "\\alpha_t" + }, + { + "bbox": [ + 313, + 164, + 555, + 289 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 313, + 164, + 555, + 289 + ], + "type": "inline_equation", + "content": "\\beta_t" + }, + { + "bbox": [ + 313, + 164, + 555, + 289 + ], + "type": "text", + "content": " are not arbitrary transformations; rather, they should satisfy a specific sum constraint. Then, we redefine the conditional Markov diffusion process similar to the directional label forward distribution " + }, + { + "bbox": [ + 313, + 164, + 555, + 289 + ], + "type": "inline_equation", + "content": "q" + }, + { + "bbox": [ + 313, + 164, + 555, + 289 + ], + "type": "text", + "content": " without condition (detailed theoretical framework in Appendix C), assuming that when conditioned on features " + }, + { + "bbox": [ + 313, + 164, + 555, + 289 + ], + "type": "inline_equation", + "content": "\\mathbf{x}" + }, + { + "bbox": [ + 313, + 164, + 555, + 289 + ], + "type": "text", + "content": ", the prior distribution of " + }, + { + "bbox": [ + 313, + 164, + 555, + 289 + ], + "type": "inline_equation", + "content": "\\hat{q}" + }, + { + "bbox": [ + 313, + 164, + 555, + 289 + ], + "type": "text", + "content": " is the same as " + }, + { + "bbox": [ + 313, + 164, + 555, + 289 + ], + "type": "inline_equation", + "content": "q" + }, + { + "bbox": [ + 313, + 164, + 555, + 289 + ], + "type": "text", + "content": ":" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 328, + 297, + 553, + 327 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 328, + 297, + 553, + 327 + ], + "spans": [ + { + "bbox": [ + 328, + 297, + 553, + 327 + ], + "type": "interline_equation", + "content": "\\begin{array}{l} \\hat {q} \\left(\\mathbf {y} _ {t} \\mid \\mathbf {y} _ {t - 1}, \\mathbf {y} _ {d}, \\mathbf {x}\\right) := q \\left(\\mathbf {y} _ {t} \\mid \\mathbf {y} _ {t - 1}, \\mathbf {y} _ {d}\\right) \\tag {8} \\\\ = \\mathcal {N} \\left(\\mathbf {y} _ {t}; \\mathbf {y} _ {t - 1} + \\alpha_ {t} \\mathbf {y} _ {d}, \\beta_ {t} ^ {2} \\mathbf {I}\\right), \\\\ \\end{array}", + "image_path": "cbce3d1914818ea6bc58ef9f7a2aefbde0e06a9c2cae4a479f5e72e8a97aadd2.jpg" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 330, + 347, + 554, + 396 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 330, + 347, + 554, + 396 + ], + "spans": [ + { + "bbox": [ + 330, + 347, + 554, + 396 + ], + "type": "interline_equation", + "content": "\\begin{array}{l} \\hat {q} \\left(\\mathbf {y} _ {1: T} \\mid \\mathbf {y} _ {0}, \\mathbf {y} _ {d}, \\mathbf {x}\\right) := \\prod_ {t = 1} ^ {T} \\hat {q} \\left(\\mathbf {y} _ {t} \\mid \\mathbf {y} _ {t - 1}, \\mathbf {y} _ {d}, \\mathbf {x}\\right) \\tag {9} \\\\ = \\mathcal {N} \\left(\\mathbf {y} _ {t}; \\mathbf {y} _ {0} + \\bar {\\alpha} _ {t} \\mathbf {y} _ {d}, \\bar {\\beta} _ {t} ^ {2} \\mathbf {I}\\right), \\\\ \\end{array}", + "image_path": "bd61c801427c3058db2cc281bbdb6fbe2e16faa42fc070ff8db6a99d82029484.jpg" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 313, + 403, + 554, + 427 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 403, + 554, + 427 + ], + "spans": [ + { + "bbox": [ + 313, + 403, + 554, + 427 + ], + "type": "text", + "content": "which means that the diffusion process of the labels is not influenced by any image features." + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 313, + 428, + 554, + 463 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 428, + 554, + 463 + ], + "spans": [ + { + "bbox": [ + 313, + 428, + 554, + 463 + ], + "type": "text", + "content": "In Appendix C, using Bayes' rule and Taylor expansion, we derive the conditional posterior distribution during the forward process:" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 333, + 474, + 553, + 488 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 333, + 474, + 553, + 488 + ], + "spans": [ + { + "bbox": [ + 333, + 474, + 553, + 488 + ], + "type": "interline_equation", + "content": "\\hat {q} \\left(\\mathbf {y} _ {t - 1} \\mid \\mathbf {y} _ {t}, \\mathbf {y} _ {0}, \\mathbf {y} _ {d}, \\mathbf {x}\\right) = \\mathcal {N} \\left(\\mathbf {y} _ {t - 1}; \\hat {\\mu} _ {t}, \\hat {\\sigma} _ {t} \\mathbf {I}\\right), \\tag {10}", + "image_path": "9c5171cfbc564e53876f9c301f76ad75baef3c1b2a41dd3b8c846b077421a2a6.jpg" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 313, + 497, + 554, + 581 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 497, + 554, + 581 + ], + "spans": [ + { + "bbox": [ + 313, + 497, + 554, + 581 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 313, + 497, + 554, + 581 + ], + "type": "inline_equation", + "content": "\\hat{\\mu}_t = y_t - \\alpha_t y_d - \\frac{\\beta_t^2}{\\beta_t} \\epsilon + \\hat{\\sigma}_t \\nabla_{y_{t-1}} \\log \\hat{q}(\\mathbf{x} | \\mathbf{y}_{t-1}, \\mathbf{y}_d)" + }, + { + "bbox": [ + 313, + 497, + 554, + 581 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 313, + 497, + 554, + 581 + ], + "type": "inline_equation", + "content": "\\hat{\\sigma}_t = \\frac{\\beta_t^2 \\bar{\\beta}_{t-1}^2}{\\beta_t^2}" + }, + { + "bbox": [ + 313, + 497, + 554, + 581 + ], + "type": "text", + "content": ". The last term in " + }, + { + "bbox": [ + 313, + 497, + 554, + 581 + ], + "type": "inline_equation", + "content": "\\hat{\\mu}_t" + }, + { + "bbox": [ + 313, + 497, + 554, + 581 + ], + "type": "text", + "content": " can be considered a deviation to the unconditional distribution mean, guided by the first derivative of " + }, + { + "bbox": [ + 313, + 497, + 554, + 581 + ], + "type": "inline_equation", + "content": "\\hat{q}(\\mathbf{x} | \\mathbf{y}_{t-1}, \\mathbf{y}_d)" + }, + { + "bbox": [ + 313, + 497, + 554, + 581 + ], + "type": "text", + "content": ", indicating that the decoding information from " + }, + { + "bbox": [ + 313, + 497, + 554, + 581 + ], + "type": "inline_equation", + "content": "\\mathbf{y}_{t-1}" + }, + { + "bbox": [ + 313, + 497, + 554, + 581 + ], + "type": "text", + "content": " to " + }, + { + "bbox": [ + 313, + 497, + 554, + 581 + ], + "type": "inline_equation", + "content": "\\mathbf{x}" + }, + { + "bbox": [ + 313, + 497, + 554, + 581 + ], + "type": "text", + "content": " should be introduced in training." + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 313, + 587, + 514, + 601 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 587, + 514, + 601 + ], + "spans": [ + { + "bbox": [ + 313, + 587, + 514, + 601 + ], + "type": "text", + "content": "4.3. Reverse and Training Process of DLD" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 313, + 605, + 554, + 689 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 605, + 554, + 689 + ], + "spans": [ + { + "bbox": [ + 313, + 605, + 554, + 689 + ], + "type": "text", + "content": "Reverse Process and Training. In the reverse (generation) process, as indicated by Eq. 6, we need to simultaneously predict " + }, + { + "bbox": [ + 313, + 605, + 554, + 689 + ], + "type": "inline_equation", + "content": "\\mathbf{y}_d" + }, + { + "bbox": [ + 313, + 605, + 554, + 689 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 313, + 605, + 554, + 689 + ], + "type": "inline_equation", + "content": "\\epsilon" + }, + { + "bbox": [ + 313, + 605, + 554, + 689 + ], + "type": "text", + "content": " to progressively recover " + }, + { + "bbox": [ + 313, + 605, + 554, + 689 + ], + "type": "inline_equation", + "content": "\\mathbf{y}_T" + }, + { + "bbox": [ + 313, + 605, + 554, + 689 + ], + "type": "text", + "content": " to " + }, + { + "bbox": [ + 313, + 605, + 554, + 689 + ], + "type": "inline_equation", + "content": "\\mathbf{y}_0" + }, + { + "bbox": [ + 313, + 605, + 554, + 689 + ], + "type": "text", + "content": ". Therefore, we can train two independent networks to estimate " + }, + { + "bbox": [ + 313, + 605, + 554, + 689 + ], + "type": "inline_equation", + "content": "\\mathbf{y}_d^{\\theta_1}" + }, + { + "bbox": [ + 313, + 605, + 554, + 689 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 313, + 605, + 554, + 689 + ], + "type": "inline_equation", + "content": "\\epsilon_{\\theta_2}" + }, + { + "bbox": [ + 313, + 605, + 554, + 689 + ], + "type": "text", + "content": ", respectively, to compute the reverse posterior distribution. Furthermore, we directly use feature information as input for two networks:" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 350, + 698, + 553, + 713 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 350, + 698, + 553, + 713 + ], + "spans": [ + { + "bbox": [ + 350, + 698, + 553, + 713 + ], + "type": "interline_equation", + "content": "p _ {\\theta} \\left(\\mathbf {y} _ {t - 1} \\mid \\mathbf {y} _ {t}, \\mathbf {x}\\right) := \\mathcal {N} \\left(\\mathbf {y} _ {t - 1}; \\mu_ {\\theta}, \\hat {\\sigma} _ {t} ^ {2} \\mathbf {I}\\right). \\tag {11}", + "image_path": "2adf7c81412af28665b7477808c5bab6c6151ebd718002989c67a9d14e7e6552.jpg" + } + ] + } + ], + "index": 21 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "text", + "content": "25741" + } + ] + } + ], + "index": 22 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 3 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 104, + 73, + 504, + 297 + ], + "blocks": [ + { + "bbox": [ + 104, + 73, + 504, + 297 + ], + "lines": [ + { + "bbox": [ + 104, + 73, + 504, + 297 + ], + "spans": [ + { + "bbox": [ + 104, + 73, + 504, + 297 + ], + "type": "image", + "image_path": "e97e2bb5b87d4508df05e432f7987aa9cc5744c0ac32100494bd5f996e06b6d5.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 54, + 306, + 555, + 340 + ], + "lines": [ + { + "bbox": [ + 54, + 306, + 555, + 340 + ], + "spans": [ + { + "bbox": [ + 54, + 306, + 555, + 340 + ], + "type": "text", + "content": "Figure 2. Overview of the DLD framework. During Stage 1, the noisy dataset is quickly divided into three subsets. Then, we estimate " + }, + { + "bbox": [ + 54, + 306, + 555, + 340 + ], + "type": "inline_equation", + "content": "\\mathbf{y}_0" + }, + { + "bbox": [ + 54, + 306, + 555, + 340 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 54, + 306, + 555, + 340 + ], + "type": "inline_equation", + "content": "\\mathbf{y}_n" + }, + { + "bbox": [ + 54, + 306, + 555, + 340 + ], + "type": "text", + "content": " for each samples based on Eq. 14 and Eq. 15. In Stage 2, the label information and augmented features from Stage 1 are jointly used to train a directional diffusion model (blue bars represent random diffusion and gray bars represent directional diffusion)." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "bbox": [ + 54, + 360, + 296, + 432 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 54, + 360, + 296, + 432 + ], + "spans": [ + { + "bbox": [ + 54, + 360, + 296, + 432 + ], + "type": "text", + "content": "Based on the analysis of the forward process, we adopt a straightforward approach which naturally integrates the decoding process into the diffusion model architecture by a fixed decoding layer " + }, + { + "bbox": [ + 54, + 360, + 296, + 432 + ], + "type": "inline_equation", + "content": "L_{1}" + }, + { + "bbox": [ + 54, + 360, + 296, + 432 + ], + "type": "text", + "content": " to map " + }, + { + "bbox": [ + 54, + 360, + 296, + 432 + ], + "type": "inline_equation", + "content": "\\mathbf{y}" + }, + { + "bbox": [ + 54, + 360, + 296, + 432 + ], + "type": "text", + "content": " into the feature space of " + }, + { + "bbox": [ + 54, + 360, + 296, + 432 + ], + "type": "inline_equation", + "content": "\\mathbf{x}" + }, + { + "bbox": [ + 54, + 360, + 296, + 432 + ], + "type": "text", + "content": ". Detailed information about the specific diffusion model network structure is shown in Appendix Figure C.1." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 55, + 433, + 296, + 456 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 433, + 296, + 456 + ], + "spans": [ + { + "bbox": [ + 55, + 433, + 296, + 456 + ], + "type": "text", + "content": "By optimizing the VLB and reparameterization [21], we derive the following simplified loss function for training:" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 104, + 465, + 295, + 483 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 104, + 465, + 295, + 483 + ], + "spans": [ + { + "bbox": [ + 104, + 465, + 295, + 483 + ], + "type": "interline_equation", + "content": "\\mathcal {L} _ {d} = \\left\\| \\mathbf {y} _ {d} - \\mathbf {y} _ {d} ^ {\\theta_ {1}} \\left(\\mathbf {y} _ {t}, \\mathbf {y} _ {n}, \\mathbf {x}, t\\right) \\right\\| ^ {2}, \\tag {12}", + "image_path": "4268f15a5192342a4b7bbc1d6afe8db12019dd3a41b6f9ac31fd34dcf7d32e44.jpg" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 108, + 502, + 295, + 517 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 108, + 502, + 295, + 517 + ], + "spans": [ + { + "bbox": [ + 108, + 502, + 295, + 517 + ], + "type": "interline_equation", + "content": "\\mathcal {L} _ {\\epsilon} = \\left\\| \\epsilon - \\epsilon_ {\\theta_ {2}} \\left(\\mathbf {y} _ {t}, \\mathbf {y} _ {n}, \\mathbf {x}, t\\right) \\right\\| ^ {2}, \\tag {13}", + "image_path": "768cede62a62945671884d411769a5d840230656c086d1478fe7f80f3598c04b.jpg" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 55, + 521, + 295, + 545 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 521, + 295, + 545 + ], + "spans": [ + { + "bbox": [ + 55, + 521, + 295, + 545 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 55, + 521, + 295, + 545 + ], + "type": "inline_equation", + "content": "\\epsilon \\sim \\mathcal{N}(0,\\mathbf{I})" + }, + { + "bbox": [ + 55, + 521, + 295, + 545 + ], + "type": "text", + "content": ", and the training input " + }, + { + "bbox": [ + 55, + 521, + 295, + 545 + ], + "type": "inline_equation", + "content": "\\mathbf{y}_t" + }, + { + "bbox": [ + 55, + 521, + 295, + 545 + ], + "type": "text", + "content": " is synthesized using Eq. 7. Algorithm 1 describes the training procedure." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 55, + 545, + 296, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 545, + 296, + 713 + ], + "spans": [ + { + "bbox": [ + 55, + 545, + 296, + 713 + ], + "type": "text", + "content": "Inference. Since the classification diffusion model is deterministic in inference and the dimensionality of the label vector is much lower than that of images, we mimic the inference process of DDIM [44] to generate labels in as few steps as possible without affecting the generated categories. We can gradually sample labels step-by-step following " + }, + { + "bbox": [ + 55, + 545, + 296, + 713 + ], + "type": "inline_equation", + "content": "\\mathbf{y}_{t-1} = \\mathbf{y}_t - (\\bar{\\alpha}_t - \\bar{\\alpha}_{t-1})\\mathbf{y}_d^{\\theta_1} - (\\bar{\\beta}_t - \\bar{\\beta}_{t-1})\\epsilon_{\\theta_2}" + }, + { + "bbox": [ + 55, + 545, + 296, + 713 + ], + "type": "text", + "content": ", where the transition from " + }, + { + "bbox": [ + 55, + 545, + 296, + 713 + ], + "type": "inline_equation", + "content": "t-1" + }, + { + "bbox": [ + 55, + 545, + 296, + 713 + ], + "type": "text", + "content": " to " + }, + { + "bbox": [ + 55, + 545, + 296, + 713 + ], + "type": "inline_equation", + "content": "t" + }, + { + "bbox": [ + 55, + 545, + 296, + 713 + ], + "type": "text", + "content": " is discrete rather than continuous in time. Appendix D provides a detailed explanation of the de-Markovization process, the accelerated sampling algorithm and the visual demonstration of the sampling process. In experiments, we use 5 steps for DLD's inference, which is on the same order of magnitude as the speed of traditional classification models." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 313, + 360, + 474, + 372 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 360, + 474, + 372 + ], + "spans": [ + { + "bbox": [ + 313, + 360, + 474, + 372 + ], + "type": "text", + "content": "4.4. Label Pre-correction Method" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 313, + 380, + 555, + 463 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 380, + 555, + 463 + ], + "spans": [ + { + "bbox": [ + 313, + 380, + 555, + 463 + ], + "type": "text", + "content": "Although diffusion models are powerful, under the guidance of incorrect " + }, + { + "bbox": [ + 313, + 380, + 555, + 463 + ], + "type": "inline_equation", + "content": "\\mathbf{y}_0" + }, + { + "bbox": [ + 313, + 380, + 555, + 463 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 313, + 380, + 555, + 463 + ], + "type": "inline_equation", + "content": "\\mathbf{y}_n" + }, + { + "bbox": [ + 313, + 380, + 555, + 463 + ], + "type": "text", + "content": ", they can easily fit the wrong conditional distributions. Therefore, we propose a one-step label pre-correction method that automatically provides accurate label information and feature conditions for the DLD model, requiring no additional parameter training, making it simple, straightforward, and efficient." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 313, + 465, + 556, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 465, + 556, + 715 + ], + "spans": [ + { + "bbox": [ + 313, + 465, + 556, + 715 + ], + "type": "text", + "content": "Label Pre-correction for " + }, + { + "bbox": [ + 313, + 465, + 556, + 715 + ], + "type": "inline_equation", + "content": "\\mathbf{y}_0" + }, + { + "bbox": [ + 313, + 465, + 556, + 715 + ], + "type": "text", + "content": ". Inspired by the concepts of multi-view[42, 53] and neighbor consistency regularization [22], our primary assumption is that the same sample should have similar neighborhood label distributions in different feature spaces. We define the original noisy dataset as " + }, + { + "bbox": [ + 313, + 465, + 556, + 715 + ], + "type": "inline_equation", + "content": "\\mathcal{D} = \\{(x_i,\\tilde{y}_i)\\mid x_i\\in \\mathbb{R}^d,\\tilde{y}_i\\in \\mathbb{R}^m,i = 1,\\ldots ,n\\}" + }, + { + "bbox": [ + 313, + 465, + 556, + 715 + ], + "type": "text", + "content": ". To obtain different feature spaces by a shared feature extractor " + }, + { + "bbox": [ + 313, + 465, + 556, + 715 + ], + "type": "inline_equation", + "content": "f_{ext}" + }, + { + "bbox": [ + 313, + 465, + 556, + 715 + ], + "type": "text", + "content": ", we first apply two different data augmentation methods and map them into " + }, + { + "bbox": [ + 313, + 465, + 556, + 715 + ], + "type": "inline_equation", + "content": "\\mathbb{R}^w" + }, + { + "bbox": [ + 313, + 465, + 556, + 715 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 313, + 465, + 556, + 715 + ], + "type": "inline_equation", + "content": "\\mathbb{R}^s" + }, + { + "bbox": [ + 313, + 465, + 556, + 715 + ], + "type": "text", + "content": ", with the augmented datasets respectively denoted as " + }, + { + "bbox": [ + 313, + 465, + 556, + 715 + ], + "type": "inline_equation", + "content": "\\mathcal{D}^w" + }, + { + "bbox": [ + 313, + 465, + 556, + 715 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 313, + 465, + 556, + 715 + ], + "type": "inline_equation", + "content": "\\mathcal{D}^s" + }, + { + "bbox": [ + 313, + 465, + 556, + 715 + ], + "type": "text", + "content": ". Then, we use Euclidean or Cosine similarity to obtain the nearest " + }, + { + "bbox": [ + 313, + 465, + 556, + 715 + ], + "type": "inline_equation", + "content": "K" + }, + { + "bbox": [ + 313, + 465, + 556, + 715 + ], + "type": "text", + "content": " neighborhood labels " + }, + { + "bbox": [ + 313, + 465, + 556, + 715 + ], + "type": "inline_equation", + "content": "y_{k = 1}^{K}" + }, + { + "bbox": [ + 313, + 465, + 556, + 715 + ], + "type": "text", + "content": " and feature space distances " + }, + { + "bbox": [ + 313, + 465, + 556, + 715 + ], + "type": "inline_equation", + "content": "v_{k = 1}^{K}" + }, + { + "bbox": [ + 313, + 465, + 556, + 715 + ], + "type": "text", + "content": " then we can get two weighted label distributions " + }, + { + "bbox": [ + 313, + 465, + 556, + 715 + ], + "type": "inline_equation", + "content": "p_w" + }, + { + "bbox": [ + 313, + 465, + 556, + 715 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 313, + 465, + 556, + 715 + ], + "type": "inline_equation", + "content": "p_s" + }, + { + "bbox": [ + 313, + 465, + 556, + 715 + ], + "type": "text", + "content": " by " + }, + { + "bbox": [ + 313, + 465, + 556, + 715 + ], + "type": "inline_equation", + "content": "p(y_{i}) = \\left(\\sum_{k = 1}^{K}w_{k}\\cdot \\mathbf{1}(y_{k} = y_{i})\\right)\\cdot \\left(\\sum_{k = 1}^{K}w_{k}\\right)^{-1}" + }, + { + "bbox": [ + 313, + 465, + 556, + 715 + ], + "type": "text", + "content": " where " + }, + { + "bbox": [ + 313, + 465, + 556, + 715 + ], + "type": "inline_equation", + "content": "w_{k} = \\frac{1}{v_{k} + \\delta}" + }, + { + "bbox": [ + 313, + 465, + 556, + 715 + ], + "type": "text", + "content": ", where " + }, + { + "bbox": [ + 313, + 465, + 556, + 715 + ], + "type": "inline_equation", + "content": "\\delta" + }, + { + "bbox": [ + 313, + 465, + 556, + 715 + ], + "type": "text", + "content": " is a small constant to prevent division by zero. Next, we calculate the divergence between the two distributions using the Kullback-Leibler divergence and a via Gaussian mixture model (GMM) [38] to re-partitioning three types of samples: Samples with high means are reassigned to the hard set " + }, + { + "bbox": [ + 313, + 465, + 556, + 715 + ], + "type": "inline_equation", + "content": "\\mathcal{D}_H" + }, + { + "bbox": [ + 313, + 465, + 556, + 715 + ], + "type": "text", + "content": ", which represents" + } + ] + } + ], + "index": 10 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "type": "text", + "content": "25742" + } + ] + } + ], + "index": 11 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 4 + }, + { + "para_blocks": [ + { + "bbox": [ + 55, + 72, + 294, + 179 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 72, + 294, + 179 + ], + "spans": [ + { + "bbox": [ + 55, + 72, + 294, + 179 + ], + "type": "text", + "content": "samples whose correctness is hard to determine and typically lie at the boundaries of classes. Samples with low means indicate that similar neighborhood distributions under both views. Using the principle of neighborhood consistency, we compare " + }, + { + "bbox": [ + 55, + 72, + 294, + 179 + ], + "type": "inline_equation", + "content": "\\text{argmax}(p_{ws})" + }, + { + "bbox": [ + 55, + 72, + 294, + 179 + ], + "type": "text", + "content": " with the original label " + }, + { + "bbox": [ + 55, + 72, + 294, + 179 + ], + "type": "inline_equation", + "content": "\\tilde{y}" + }, + { + "bbox": [ + 55, + 72, + 294, + 179 + ], + "type": "text", + "content": ", where " + }, + { + "bbox": [ + 55, + 72, + 294, + 179 + ], + "type": "inline_equation", + "content": "p_{ws} = \\frac{p_w + p_s}{2}" + }, + { + "bbox": [ + 55, + 72, + 294, + 179 + ], + "type": "text", + "content": ". If the class with the highest frequency in the average neighborhood distribution matches the original label, then the sample is assigned to the clean set " + }, + { + "bbox": [ + 55, + 72, + 294, + 179 + ], + "type": "inline_equation", + "content": "\\mathcal{D}_C" + }, + { + "bbox": [ + 55, + 72, + 294, + 179 + ], + "type": "text", + "content": "; otherwise, it is assigned to the noisy set " + }, + { + "bbox": [ + 55, + 72, + 294, + 179 + ], + "type": "inline_equation", + "content": "\\mathcal{D}_N" + }, + { + "bbox": [ + 55, + 72, + 294, + 179 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 55, + 180, + 294, + 228 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 180, + 294, + 228 + ], + "spans": [ + { + "bbox": [ + 55, + 180, + 294, + 228 + ], + "type": "text", + "content": "Based on the repartitioned sample subsets and the original noisy labels " + }, + { + "bbox": [ + 55, + 180, + 294, + 228 + ], + "type": "inline_equation", + "content": "\\tilde{y}" + }, + { + "bbox": [ + 55, + 180, + 294, + 228 + ], + "type": "text", + "content": ", we define a label pre-corrector, which functions similarly to a transition matrix. It can automatically perform label flipping or refinement:" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 67, + 236, + 295, + 275 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 67, + 236, + 295, + 275 + ], + "spans": [ + { + "bbox": [ + 67, + 236, + 295, + 275 + ], + "type": "interline_equation", + "content": "P \\left(y _ {0} \\mid \\tilde {y}, x\\right) = \\left\\{ \\begin{array}{l l} \\tilde {y} & \\text {i f} x \\in \\mathcal {D} _ {C}, \\\\ \\operatorname {a r g m a x} \\left(p _ {w s}\\right) & \\text {i f} x \\in \\mathcal {D} _ {N}, \\\\ p _ {w s} & \\text {i f} x \\in \\mathcal {D} _ {H}. \\end{array} \\right. \\tag {14}", + "image_path": "987608727c160cba36cfdc97bcd27afcdc5cff0919696fe58e3d4f9b8056224e.jpg" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 55, + 283, + 295, + 390 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 283, + 295, + 390 + ], + "spans": [ + { + "bbox": [ + 55, + 283, + 295, + 390 + ], + "type": "text", + "content": "According to Eq. 14, for samples with deterministic neighborhood label distributions, we use hard labels to minimize the interference of redundant information on " + }, + { + "bbox": [ + 55, + 283, + 295, + 390 + ], + "type": "inline_equation", + "content": "y_0" + }, + { + "bbox": [ + 55, + 283, + 295, + 390 + ], + "type": "text", + "content": ". For samples with more contentious neighborhood label distributions, we adopt a soft strategy for label refinement. Instead of point estimation, we estimate through a label vector, retaining most of the neighborhood information. This strategy enhances the accuracy of the " + }, + { + "bbox": [ + 55, + 283, + 295, + 390 + ], + "type": "inline_equation", + "content": "y_0" + }, + { + "bbox": [ + 55, + 283, + 295, + 390 + ], + "type": "text", + "content": " while preserving as much robust information as possible." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 55, + 392, + 295, + 571 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 392, + 295, + 571 + ], + "spans": [ + { + "bbox": [ + 55, + 392, + 295, + 571 + ], + "type": "text", + "content": "Estimate Noise distribution for " + }, + { + "bbox": [ + 55, + 392, + 295, + 571 + ], + "type": "inline_equation", + "content": "\\mathbf{y}_n" + }, + { + "bbox": [ + 55, + 392, + 295, + 571 + ], + "type": "text", + "content": ". Similar to obtaining " + }, + { + "bbox": [ + 55, + 392, + 295, + 571 + ], + "type": "inline_equation", + "content": "\\mathbf{y}_0" + }, + { + "bbox": [ + 55, + 392, + 295, + 571 + ], + "type": "text", + "content": ", we use the subset of divided samples to estimate the noise distribution of the samples. According to Eq. 15, when a sample is identified as clean, the directed diffusion term is used to gradually fade out " + }, + { + "bbox": [ + 55, + 392, + 295, + 571 + ], + "type": "inline_equation", + "content": "\\mathbf{y}_0" + }, + { + "bbox": [ + 55, + 392, + 295, + 571 + ], + "type": "text", + "content": " (" + }, + { + "bbox": [ + 55, + 392, + 295, + 571 + ], + "type": "inline_equation", + "content": "\\mathbf{y}_d = -\\mathbf{y}_0" + }, + { + "bbox": [ + 55, + 392, + 295, + 571 + ], + "type": "text", + "content": "), allowing the model to learn to recover a label closer to the ground truth from an uninformed " + }, + { + "bbox": [ + 55, + 392, + 295, + 571 + ], + "type": "inline_equation", + "content": "\\epsilon" + }, + { + "bbox": [ + 55, + 392, + 295, + 571 + ], + "type": "text", + "content": ". When the sample contains noise, the endpoint of diffusion should include a clear representation of noise, enabling the model to strictly avoid recovering such noisy vector values during reverse generation. For uncertain samples, we simulate potential noise distributions by retaining the difference in label distributions between two views. The greater the difference in the distribution values for a particular class in the two views, the more likely the sample is incorrectly labeled as that class." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 75, + 580, + 295, + 622 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 75, + 580, + 295, + 622 + ], + "spans": [ + { + "bbox": [ + 75, + 580, + 295, + 622 + ], + "type": "interline_equation", + "content": "P \\left(y _ {n} \\mid \\tilde {y}, x\\right) = \\left\\{ \\begin{array}{l l} 0 & \\text {i f} x \\in \\mathcal {D} _ {C}, \\\\ \\tilde {y} & \\text {i f} x \\in \\mathcal {D} _ {N}, \\\\ \\frac {\\left| p _ {w} - p _ {s} \\right|}{\\sum \\left| p _ {w} - p _ {s} \\right|} & \\text {i f} x \\in \\mathcal {D} _ {H}. \\end{array} \\right. \\tag {15}", + "image_path": "c71d59e4a75ee427b756b978bb7c5476ac84d0e90491be4e3ab1f4ccd87f9e0f.jpg" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 55, + 629, + 295, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 629, + 295, + 713 + ], + "spans": [ + { + "bbox": [ + 55, + 629, + 295, + 713 + ], + "type": "text", + "content": "The complete training process is outlined in Algorithm 1, consisting of the following phases: Steps " + }, + { + "bbox": [ + 55, + 629, + 295, + 713 + ], + "type": "inline_equation", + "content": "1 \\sim 3" + }, + { + "bbox": [ + 55, + 629, + 295, + 713 + ], + "type": "text", + "content": " for sample re-partitioning, Steps " + }, + { + "bbox": [ + 55, + 629, + 295, + 713 + ], + "type": "inline_equation", + "content": "4 \\sim 9" + }, + { + "bbox": [ + 55, + 629, + 295, + 713 + ], + "type": "text", + "content": " for diffusion model training, with Step 6 dedicated to " + }, + { + "bbox": [ + 55, + 629, + 295, + 713 + ], + "type": "inline_equation", + "content": "\\mathbf{y}_0" + }, + { + "bbox": [ + 55, + 629, + 295, + 713 + ], + "type": "text", + "content": " pre-correction and " + }, + { + "bbox": [ + 55, + 629, + 295, + 713 + ], + "type": "inline_equation", + "content": "\\mathbf{y}_n" + }, + { + "bbox": [ + 55, + 629, + 295, + 713 + ], + "type": "text", + "content": " estimation. Notably, in Step 5, instead of sampling features from the original noisy set " + }, + { + "bbox": [ + 55, + 629, + 295, + 713 + ], + "type": "inline_equation", + "content": "\\mathcal{D}" + }, + { + "bbox": [ + 55, + 629, + 295, + 713 + ], + "type": "text", + "content": ", we use a mixture of sample features from " + }, + { + "bbox": [ + 55, + 629, + 295, + 713 + ], + "type": "inline_equation", + "content": "\\mathcal{D}^w" + }, + { + "bbox": [ + 55, + 629, + 295, + 713 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 55, + 629, + 295, + 713 + ], + "type": "inline_equation", + "content": "\\mathcal{D}^s" + }, + { + "bbox": [ + 55, + 629, + 295, + 713 + ], + "type": "text", + "content": " as the training input for the diffusion" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 315, + 72, + 429, + 84 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 315, + 72, + 429, + 84 + ], + "spans": [ + { + "bbox": [ + 315, + 72, + 429, + 84 + ], + "type": "text", + "content": "Algorithm 1 DLD Training" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 314, + 87, + 553, + 111 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 314, + 87, + 553, + 111 + ], + "spans": [ + { + "bbox": [ + 314, + 87, + 553, + 111 + ], + "type": "text", + "content": "Input: Noisy training set " + }, + { + "bbox": [ + 314, + 87, + 553, + 111 + ], + "type": "inline_equation", + "content": "\\mathcal{D} = \\{\\mathbf{X},\\tilde{\\mathbf{Y}}\\}" + }, + { + "bbox": [ + 314, + 87, + 553, + 111 + ], + "type": "text", + "content": ", feature extractor " + }, + { + "bbox": [ + 314, + 87, + 553, + 111 + ], + "type": "inline_equation", + "content": "f_{\\mathrm{ext}}" + }, + { + "bbox": [ + 314, + 87, + 553, + 111 + ], + "type": "text", + "content": ", two network parameters " + }, + { + "bbox": [ + 314, + 87, + 553, + 111 + ], + "type": "inline_equation", + "content": "\\theta_{1}" + }, + { + "bbox": [ + 314, + 87, + 553, + 111 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 314, + 87, + 553, + 111 + ], + "type": "inline_equation", + "content": "\\theta_{2}" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 315, + 112, + 392, + 123 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 315, + 112, + 392, + 123 + ], + "spans": [ + { + "bbox": [ + 315, + 112, + 392, + 123 + ], + "type": "text", + "content": "Output: " + }, + { + "bbox": [ + 315, + 112, + 392, + 123 + ], + "type": "inline_equation", + "content": "\\theta_{1}" + }, + { + "bbox": [ + 315, + 112, + 392, + 123 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 315, + 112, + 392, + 123 + ], + "type": "inline_equation", + "content": "\\theta_{2}" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 319, + 125, + 553, + 280 + ], + "type": "list", + "angle": 0, + "index": 19, + "blocks": [ + { + "bbox": [ + 319, + 125, + 553, + 149 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 319, + 125, + 553, + 149 + ], + "spans": [ + { + "bbox": [ + 319, + 125, + 553, + 149 + ], + "type": "text", + "content": "1: Obtain two augmented features " + }, + { + "bbox": [ + 319, + 125, + 553, + 149 + ], + "type": "inline_equation", + "content": "\\mathcal{D}^w" + }, + { + "bbox": [ + 319, + 125, + 553, + 149 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 319, + 125, + 553, + 149 + ], + "type": "inline_equation", + "content": "\\mathcal{D}^s" + }, + { + "bbox": [ + 319, + 125, + 553, + 149 + ], + "type": "text", + "content": " using the feature extractor " + }, + { + "bbox": [ + 319, + 125, + 553, + 149 + ], + "type": "inline_equation", + "content": "f_{\\mathrm{ext}}" + }, + { + "bbox": [ + 319, + 125, + 553, + 149 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 319, + 150, + 533, + 161 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 319, + 150, + 533, + 161 + ], + "spans": [ + { + "bbox": [ + 319, + 150, + 533, + 161 + ], + "type": "text", + "content": "2: Calculate K-nearest label distributions " + }, + { + "bbox": [ + 319, + 150, + 533, + 161 + ], + "type": "inline_equation", + "content": "p_w" + }, + { + "bbox": [ + 319, + 150, + 533, + 161 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 319, + 150, + 533, + 161 + ], + "type": "inline_equation", + "content": "p_s" + }, + { + "bbox": [ + 319, + 150, + 533, + 161 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 319, + 162, + 522, + 173 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 319, + 162, + 522, + 173 + ], + "spans": [ + { + "bbox": [ + 319, + 162, + 522, + 173 + ], + "type": "text", + "content": "3: Re-partitioning samples into " + }, + { + "bbox": [ + 319, + 162, + 522, + 173 + ], + "type": "inline_equation", + "content": "\\mathcal{D}_C, \\mathcal{D}_N" + }, + { + "bbox": [ + 319, + 162, + 522, + 173 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 319, + 162, + 522, + 173 + ], + "type": "inline_equation", + "content": "\\mathcal{D}_H" + }, + { + "bbox": [ + 319, + 162, + 522, + 173 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 319, + 174, + 429, + 185 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 319, + 174, + 429, + 185 + ], + "spans": [ + { + "bbox": [ + 319, + 174, + 429, + 185 + ], + "type": "text", + "content": "4: while not converged do" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 319, + 186, + 553, + 209 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 319, + 186, + 553, + 209 + ], + "spans": [ + { + "bbox": [ + 319, + 186, + 553, + 209 + ], + "type": "text", + "content": "5: Sample data " + }, + { + "bbox": [ + 319, + 186, + 553, + 209 + ], + "type": "inline_equation", + "content": "(\\mathbf{x}_{ws},\\tilde{\\mathbf{y}})\\sim \\{\\mathcal{D}^w,\\mathcal{D}^s\\}" + }, + { + "bbox": [ + 319, + 186, + 553, + 209 + ], + "type": "text", + "content": " ; time slice " + }, + { + "bbox": [ + 319, + 186, + 553, + 209 + ], + "type": "inline_equation", + "content": "t\\sim" + }, + { + "bbox": [ + 319, + 186, + 553, + 209 + ], + "type": "inline_equation", + "content": "\\{1,\\dots ,T\\}" + }, + { + "bbox": [ + 319, + 186, + 553, + 209 + ], + "type": "text", + "content": " ; and noise " + }, + { + "bbox": [ + 319, + 186, + 553, + 209 + ], + "type": "inline_equation", + "content": "\\epsilon \\sim \\mathcal{N}(0,\\mathbf{I})" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 319, + 210, + 553, + 233 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 319, + 210, + 553, + 233 + ], + "spans": [ + { + "bbox": [ + 319, + 210, + 553, + 233 + ], + "type": "text", + "content": "6: Pre-estimate " + }, + { + "bbox": [ + 319, + 210, + 553, + 233 + ], + "type": "inline_equation", + "content": "\\mathbf{y}_0" + }, + { + "bbox": [ + 319, + 210, + 553, + 233 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 319, + 210, + 553, + 233 + ], + "type": "inline_equation", + "content": "\\mathbf{y}_n" + }, + { + "bbox": [ + 319, + 210, + 553, + 233 + ], + "type": "text", + "content": " based on Eq. 14 and Eq. 15, respectively." + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 319, + 234, + 529, + 245 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 319, + 234, + 529, + 245 + ], + "spans": [ + { + "bbox": [ + 319, + 234, + 529, + 245 + ], + "type": "text", + "content": "7: Accumulate losses through Eq. 12 and Eq. 13." + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 319, + 246, + 553, + 269 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 319, + 246, + 553, + 269 + ], + "spans": [ + { + "bbox": [ + 319, + 246, + 553, + 269 + ], + "type": "text", + "content": "8: Take gradient descent step on the losses to update network parameters." + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 319, + 270, + 375, + 280 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 319, + 270, + 375, + 280 + ], + "spans": [ + { + "bbox": [ + 319, + 270, + 375, + 280 + ], + "type": "text", + "content": "9: end while" + } + ] + } + ], + "index": 18 + } + ], + "sub_type": "text" + }, + { + "bbox": [ + 313, + 301, + 555, + 388 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 301, + 555, + 388 + ], + "spans": [ + { + "bbox": [ + 313, + 301, + 555, + 388 + ], + "type": "text", + "content": "model, where " + }, + { + "bbox": [ + 313, + 301, + 555, + 388 + ], + "type": "inline_equation", + "content": "\\mathbf{x}_{ws} = (1 - \\gamma)\\mathbf{x}_w + \\gamma \\mathbf{x}_s" + }, + { + "bbox": [ + 313, + 301, + 555, + 388 + ], + "type": "text", + "content": ", with " + }, + { + "bbox": [ + 313, + 301, + 555, + 388 + ], + "type": "inline_equation", + "content": "\\gamma = \\frac{|\\mathcal{D}_H|}{|\\mathcal{D}|}" + }, + { + "bbox": [ + 313, + 301, + 555, + 388 + ], + "type": "text", + "content": " representing the proportion of uncertain samples, which means that the more difficult the sample classification, the richer the feature conditions the model automatically introduces. This reflects an implicit regularization method that mitigates the issue of the diffusion model being overly confident when still learning incorrect information." + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 313, + 396, + 508, + 410 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 396, + 508, + 410 + ], + "spans": [ + { + "bbox": [ + 313, + 396, + 508, + 410 + ], + "type": "text", + "content": "5. Experimental Results and Analysis" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 313, + 413, + 555, + 510 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 413, + 555, + 510 + ], + "spans": [ + { + "bbox": [ + 313, + 413, + 555, + 510 + ], + "type": "text", + "content": "In this section, we first evaluate the DLD's effectiveness on CIFAR datasets with varying levels of IDN in Section 5.1. Then, We conduct ablation studies in Section 5.2 to analyze the contributions of the pre-trained feature extractor, the label pre-correction technology, and other components. In Section 5.3, we compare the DLD to the SOTA methods across four real-world noisy datasets. Appendix E provides more details about our experiments." + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 313, + 517, + 507, + 529 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 517, + 507, + 529 + ], + "spans": [ + { + "bbox": [ + 313, + 517, + 507, + 529 + ], + "type": "text", + "content": "5.1. Results on Simulated Noisy Datasets" + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 313, + 533, + 555, + 630 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 533, + 555, + 630 + ], + "spans": [ + { + "bbox": [ + 313, + 533, + 555, + 630 + ], + "type": "text", + "content": "We conduct experiments on the CIFAR-10 and CIFAR-100 datasets [27] with simulated IDN [6]. Details of the IDN generation algorithm and examples are provided in Appendix E.1. In our experiments, we generate noise levels of " + }, + { + "bbox": [ + 313, + 533, + 555, + 630 + ], + "type": "inline_equation", + "content": "20\\%" + }, + { + "bbox": [ + 313, + 533, + 555, + 630 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 313, + 533, + 555, + 630 + ], + "type": "inline_equation", + "content": "40\\%" + }, + { + "bbox": [ + 313, + 533, + 555, + 630 + ], + "type": "text", + "content": ", and " + }, + { + "bbox": [ + 313, + 533, + 555, + 630 + ], + "type": "inline_equation", + "content": "60\\%" + }, + { + "bbox": [ + 313, + 533, + 555, + 630 + ], + "type": "text", + "content": ". To obtain the feature space for label pre-correction, we provide multiple configurations of feature extractors, including DLD+ResNet, DLD+SimCLR and DLD+ViT (see Appendix F.2 for detailed settings)." + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 313, + 630, + 555, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 630, + 555, + 713 + ], + "spans": [ + { + "bbox": [ + 313, + 630, + 555, + 713 + ], + "type": "text", + "content": "As shown in Table 1, the DLD+ResNet configuration outperforms the baseline across most noise levels on CIFAR, especially in high-noise environments. We analyze that the main factor affecting its performance in low noise environment is the introduction of too much robust information, which leads to the interference of the model. When using the DLD+ViT configuration, its performance surpasses" + } + ] + } + ], + "index": 25 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "text", + "content": "25743" + } + ] + } + ], + "index": 26 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 5 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 104, + 85, + 511, + 274 + ], + "blocks": [ + { + "bbox": [ + 146, + 70, + 462, + 81 + ], + "lines": [ + { + "bbox": [ + 146, + 70, + 462, + 81 + ], + "spans": [ + { + "bbox": [ + 146, + 70, + 462, + 81 + ], + "type": "text", + "content": "Table 1. Classification accuracies (%) of CIFAR datasets under different levels of IDN." + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 104, + 85, + 511, + 274 + ], + "lines": [ + { + "bbox": [ + 104, + 85, + 511, + 274 + ], + "spans": [ + { + "bbox": [ + 104, + 85, + 511, + 274 + ], + "type": "table", + "html": "
MethodCIFAR-10CIFAR-100
20%40%60%20%40%60%
Co-teaching [17]87.28±0.2078.72±0.4762.51±1.9869.26±0.3959.53±0.5746.80±0.51
GCE [61]86.44±0.2376.71±0.3960.76±3.0868.24±0.2755.76±1.1846.75±2.46
DAC [45]86.16±0.1374.80±0.3259.97±0.9068.70±0.7855.71±1.0146.29±3.35
DMI [54]86.57±0.1677.81±0.8569.94±1.3168.71±1.2157.99±0.9250.96±2.08
SEAL [6]87.79±0.0982.98±0.0570.07±1.5470.61±0.1961.18±0.3551.86±1.32
NCR [22]87.74±0.1583.03±0.2169.44±1.0170.68±0.3761.93±0.5449.13±2.77
CC+DivideMix [62]88.36±0.1184.18±0.4071.52±1.7370.09±0.1964.35±0.3553.20±0.66
DISC [31]90.68±0.1085.61±0.3273.59±1.1371.49±0.1364.46±0.2953.62±0.51
LRA-diffusion+ResNet [4]89.94±0.1586.73±0.3774.66±1.0272.35±0.1465.55±0.3054.43±0.53
DLD+ResNet (Ours)90.43±0.1287.17±0.2976.07±1.1072.60±0.1566.59±0.3155.72±0.24
EPL+ViT [26]95.95±0.1793.01±0.3376.16±1.1872.24±0.1267.20±0.2055.59±0.13
LRA-diffusion+ViT [4]96.65±0.1893.68±0.3478.12±1.1773.59±0.1367.67±0.2157.21±0.12
DLD+ViT (Ours)96.98±0.1296.49±0.0783.14±0.8178.62±0.1176.03±0.1469.57±0.17
", + "image_path": "1bf87349bc6108a20b0cc43e23ffbb4c475582bd095e76175d2b195376dc57b4.jpg" + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "table_body" + } + ], + "index": 1 + }, + { + "bbox": [ + 55, + 292, + 294, + 376 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 292, + 294, + 376 + ], + "spans": [ + { + "bbox": [ + 55, + 292, + 294, + 376 + ], + "type": "text", + "content": "EPL+ViT and LRA-diffusion+ViT. This implies that DLD more effectively utilizes the provided prior knowledge without requiring additional training conditions. Across various datasets, DLD's superiority is more pronounced on CIFAR-100, indicating that the more complex IDN on CIFAR-100 severely impacts the performance of other methods, whereas our model maintains robust learning." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 54, + 376, + 295, + 498 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 54, + 376, + 295, + 498 + ], + "spans": [ + { + "bbox": [ + 54, + 376, + 295, + 498 + ], + "type": "text", + "content": "Additionally, we conducted experiments on the simulated CCN dataset. The results (see Appendix E.2) show that CCN is easier to handle than IDN, which aligns with our expectations. Compared to other methods, DLD still achieves the highest classification accuracy. Due to the lightweight architecture of the diffusion model and the single-step pre-correction process, DLD outperforms traditional LNL methods without incurring significant computational overhead. A detailed analysis of training time overhead and model parameters is provided in Appendix E.3." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 55, + 504, + 156, + 516 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 504, + 156, + 516 + ], + "spans": [ + { + "bbox": [ + 55, + 504, + 156, + 516 + ], + "type": "text", + "content": "5.2. Ablation Studies" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 55, + 521, + 295, + 700 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 521, + 295, + 700 + ], + "spans": [ + { + "bbox": [ + 55, + 521, + 295, + 700 + ], + "type": "text", + "content": "Power of label pre-correction. We validate the effectiveness of label pre-correction by freezing the diffusion model. As shown in Figure 3, our pre-correction method consistently outperforms in various noisy environments, while Figure 3b demonstrates that the LRA estimation method disrupts originally correct labels. Then, we unfroze the diffusion model to verify that DLD does not fully depend on the pre-correction method. Results show that DLD further enhances generalization performance based on pre-correction, indicating that the two methods complement each other effectively. Additionally, we design more comprehensive cross-ablation experiments, e.g., replacing the LRA technique in LRA-diffusion with the pre-correction module to demonstrate the necessity of directional diffusion. See Appendix F.1 for more details." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 55, + 701, + 295, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 701, + 295, + 713 + ], + "spans": [ + { + "bbox": [ + 55, + 701, + 295, + 713 + ], + "type": "text", + "content": "Contribution of different pre-trained features and other" + } + ] + } + ], + "index": 6 + }, + { + "type": "table", + "bbox": [ + 337, + 311, + 529, + 423 + ], + "blocks": [ + { + "bbox": [ + 334, + 291, + 533, + 302 + ], + "lines": [ + { + "bbox": [ + 334, + 291, + 533, + 302 + ], + "spans": [ + { + "bbox": [ + 334, + 291, + 533, + 302 + ], + "type": "text", + "content": "Table 2. Classification accuracies (%) on Animal-10N." + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 337, + 311, + 529, + 423 + ], + "lines": [ + { + "bbox": [ + 337, + 311, + 529, + 423 + ], + "spans": [ + { + "bbox": [ + 337, + 311, + 529, + 423 + ], + "type": "table", + "html": "
MethodAccuracy
CE [35]80.6
GCE [61]81.5
NCT [8]84.1
DISC [31]87.3
SSR [14]88.5
LRA-diffusion [4]88.6
J-ViT [9]89.0
DLD (Ours)89.4
", + "image_path": "d8aeea1dae7b845f36b039d219ba80f9232e176804ce280ca4353596adad6cbf.jpg" + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "table_body" + } + ], + "index": 8 + }, + { + "bbox": [ + 313, + 445, + 555, + 624 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 445, + 555, + 624 + ], + "spans": [ + { + "bbox": [ + 313, + 445, + 555, + 624 + ], + "type": "text", + "content": "training components. We explore the impact of various feature extractors, including ResNet34 [20], SimCLR [7] and ViT-L/14 [40], and conduct ablation studies comparing our method with CARD and LRA-diffusion. Here, DISC [31] served as a baseline that does not require any pretrained models. Detailed results and analyses are provided in Appendix F.2. Our findings reveal that DLD outperforms other diffusion models when combined with different pretrained models, achieving optimal performance when integrated with ViT-L/14. This underscores ViT-L/14's effectiveness in enhancing label pre-correction and facilitating diffusion model learning. Furthermore, Appendix F.3 details additional ablation studies on various hyperparameter settings. We validate the optimal training strategy for the DLD method and recommend this combination." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 313, + 635, + 512, + 647 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 635, + 512, + 647 + ], + "spans": [ + { + "bbox": [ + 313, + 635, + 512, + 647 + ], + "type": "text", + "content": "5.3. Results on Real-world Noisy Datasets" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 313, + 653, + 554, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 653, + 554, + 713 + ], + "spans": [ + { + "bbox": [ + 313, + 653, + 554, + 713 + ], + "type": "text", + "content": "We select four real-world noisy datasets for our experiments to demonstrate the superior performance of DLD: Animal10N [43], Clothing1M [52], WebVision [30], and ImageNet ILSVRC2012 [41]. Comprehensive descriptions of these datasets and detailed experimental setups can be" + } + ] + } + ], + "index": 11 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "text", + "content": "25744" + } + ] + } + ], + "index": 12 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 6 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 80, + 70, + 304, + 227 + ], + "blocks": [ + { + "bbox": [ + 80, + 70, + 304, + 227 + ], + "lines": [ + { + "bbox": [ + 80, + 70, + 304, + 227 + ], + "spans": [ + { + "bbox": [ + 80, + 70, + 304, + 227 + ], + "type": "image", + "image_path": "583fbcb0c2d28e4097d41b09173dd3eece8bb464b0de20c1edf1f9cfafa91618.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 111, + 230, + 271, + 240 + ], + "lines": [ + { + "bbox": [ + 111, + 230, + 271, + 240 + ], + "spans": [ + { + "bbox": [ + 111, + 230, + 271, + 240 + ], + "type": "text", + "content": "(a) Experimental results on the CIFAR-10 dataset" + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "type": "image", + "bbox": [ + 306, + 71, + 529, + 228 + ], + "blocks": [ + { + "bbox": [ + 306, + 71, + 529, + 228 + ], + "lines": [ + { + "bbox": [ + 306, + 71, + 529, + 228 + ], + "spans": [ + { + "bbox": [ + 306, + 71, + 529, + 228 + ], + "type": "image", + "image_path": "7e01a5932844ab8d099133ec3ebdb8ff0faa62526609a98a5f2823c22774ac80.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 334, + 230, + 499, + 240 + ], + "lines": [ + { + "bbox": [ + 334, + 230, + 499, + 240 + ], + "spans": [ + { + "bbox": [ + 334, + 230, + 499, + 240 + ], + "type": "text", + "content": "(b) Experimental results on the CIFAR-100 dataset" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 69, + 247, + 538, + 258 + ], + "lines": [ + { + "bbox": [ + 69, + 247, + 538, + 258 + ], + "spans": [ + { + "bbox": [ + 69, + 247, + 538, + 258 + ], + "type": "text", + "content": "Figure 3. Comparison of different label correction strategies' accuracies, where Baseline represents the proportion of clean labels." + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_caption" + } + ], + "index": 2 + }, + { + "type": "table", + "bbox": [ + 79, + 297, + 271, + 422 + ], + "blocks": [ + { + "bbox": [ + 77, + 277, + 273, + 289 + ], + "lines": [ + { + "bbox": [ + 77, + 277, + 273, + 289 + ], + "spans": [ + { + "bbox": [ + 77, + 277, + 273, + 289 + ], + "type": "text", + "content": "Table 3. Classification accuracies (%) on Clothing1M." + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 79, + 297, + 271, + 422 + ], + "lines": [ + { + "bbox": [ + 79, + 297, + 271, + 422 + ], + "spans": [ + { + "bbox": [ + 79, + 297, + 271, + 422 + ], + "type": "table", + "html": "
MethodAccuracy
Co-teaching [17]69.21
DMI [54]72.46
DivideMix [28]74.76
CC [62]75.40
DISC [31]74.79
EPL [26]75.21
LRA-diffusion [4]74.46
SANM [46]75.63
DLD (Ours)75.69
", + "image_path": "c70e336627993e4cee09268457f476cc37ad7f5bade2f20f791dcfa59cb762f8.jpg" + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "table_body" + } + ], + "index": 6 + }, + { + "type": "table", + "bbox": [ + 79, + 463, + 271, + 586 + ], + "blocks": [ + { + "bbox": [ + 55, + 431, + 295, + 453 + ], + "lines": [ + { + "bbox": [ + 55, + 431, + 295, + 453 + ], + "spans": [ + { + "bbox": [ + 55, + 431, + 295, + 453 + ], + "type": "text", + "content": "Table 4. Classification accuracies (%) on WebVision and ILSVRC2012 datasets." + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 79, + 463, + 271, + 586 + ], + "lines": [ + { + "bbox": [ + 79, + 463, + 271, + 586 + ], + "spans": [ + { + "bbox": [ + 79, + 463, + 271, + 586 + ], + "type": "table", + "html": "
MethodWebVisionILSVRC12
Co-teaching [17]63.5861.48
DivideMix [28]77.3275.20
ELR [33]77.7870.29
EPL [26]78.7776.51
C2D [63]79.4278.57
CC [62]79.3676.08
DISC [31]80.2877.44
LRA-diffusion [4]84.1682.56
DLD (Ours)84.5183.74
", + "image_path": "b7a33f3397a47f0357f5c83b4ba3b2223260844b524a52c2b4366ba6f881fcbb.jpg" + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "table_body" + } + ], + "index": 8 + }, + { + "bbox": [ + 55, + 605, + 295, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 605, + 295, + 713 + ], + "spans": [ + { + "bbox": [ + 55, + 605, + 295, + 713 + ], + "type": "text", + "content": "found in Appendix E.4 and Appendix E.5, respectively. As demonstrated in Table 2, Table 3, and Table 4, DLD outperforms all SOTA methods across these datasets. Note that DLD's relative classification performance improved by " + }, + { + "bbox": [ + 55, + 605, + 295, + 713 + ], + "type": "inline_equation", + "content": "7.26\\%" + }, + { + "bbox": [ + 55, + 605, + 295, + 713 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 55, + 605, + 295, + 713 + ], + "type": "inline_equation", + "content": "0.44\\%" + }, + { + "bbox": [ + 55, + 605, + 295, + 713 + ], + "type": "text", + "content": " over J-ViT [9] and EPL [26], respectively, both employing the ViT architecture. Compared to LRA-diffusion, which also employs ViT and label diffusion models, DLD achieves varying degrees of performance improvement, demonstrating a more effective and reason-" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 313, + 279, + 555, + 410 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 279, + 555, + 410 + ], + "spans": [ + { + "bbox": [ + 313, + 279, + 555, + 410 + ], + "type": "text", + "content": "able use of prior knowledge and diffusion models. Besides, while ViT-L/14 improved performance on Clothing1M, Appendix G shows that other extractors like CC can be even more effective, highlighting our method's versatility with different LNL methods. Besides, DLD's performance advantage partly stems from the pre-correction method. In future work, we aim to integrate discriminative and generative paradigms, using diffusion models to refine pre-correction, generate label information, and enhance representational power and multimodal generalization. Appendix H discusses additional limitations and future directions." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 313, + 428, + 388, + 441 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 428, + 388, + 441 + ], + "spans": [ + { + "bbox": [ + 313, + 428, + 388, + 441 + ], + "type": "text", + "content": "6. Conclusion" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 313, + 450, + 555, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 450, + 555, + 713 + ], + "spans": [ + { + "bbox": [ + 313, + 450, + 555, + 713 + ], + "type": "text", + "content": "This paper proposes DLD, a diffusion model specifically designed to learn effectively from noisy labels for robust classification. Our main idea is to decouple the label diffusion process into two channels: directional diffusion and random diffusion. This approach explicitly introduces noisy label information into the diffusion model for the first time, enabling the model to interpretably mitigate the generation of incorrect labels, thereby enhancing its ability to generalize from noisy to clean data. By incorporating our low-cost label pre-correction method, DLD can automatically select label diffusion information for each sample in diverse noisy environments, without requiring additional adjustments or iterations. Additionally, our model synergizes with various pre-trained large-scale models, like ViT, leveraging their prior knowledge more effectively than other methods. Experiments on simulated and real-world noisy datasets show that DLD surpasses conventional LNL methods in both accuracy and efficiency. By framing LNL within a generative model, our method bridges LNL with other learning frameworks. We believe this work will reduce the workload for practitioners handling large-scale, low-quality labeled data, benefiting the development of the image recognition field." + } + ] + } + ], + "index": 12 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "text", + "content": "25745" + } + ] + } + ], + "index": 13 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 7 + }, + { + "para_blocks": [ + { + "bbox": [ + 56, + 72, + 158, + 85 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 72, + 158, + 85 + ], + "spans": [ + { + "bbox": [ + 56, + 72, + 158, + 85 + ], + "type": "text", + "content": "Acknowledgements" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 55, + 91, + 296, + 175 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 91, + 296, + 175 + ], + "spans": [ + { + "bbox": [ + 55, + 91, + 296, + 175 + ], + "type": "text", + "content": "This work was supported in part by the National Natural Science Foundation of China under Grant U21A20513, Grant 62476157, Grant 62276161, Grant 62276157 and Grant 62406179, and in part by the Key Research and Development Program of Shanxi Province under Grant 202202020101003 and the Fundamental Research Program of Shanxi Province under Grant 202403021222026." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 56, + 188, + 115, + 200 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 188, + 115, + 200 + ], + "spans": [ + { + "bbox": [ + 56, + 188, + 115, + 200 + ], + "type": "text", + "content": "References" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 57, + 207, + 296, + 713 + ], + "type": "list", + "angle": 0, + "index": 13, + "blocks": [ + { + "bbox": [ + 61, + 207, + 296, + 273 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 61, + 207, + 296, + 273 + ], + "spans": [ + { + "bbox": [ + 61, + 207, + 296, + 273 + ], + "type": "text", + "content": "[1] Devansh Arpit, Stanisław Jastrzebski, Nicolas Ballas, David Krueger, Emmanuel Bengio, Maxinder S Kanwal, Tegan Maharaj, Asja Fischer, Aaron Courville, Yoshua Bengio, et al. A closer look at memorization in deep networks. In International conference on machine learning, pages 233-242. PMLR, 2017. 1" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 61, + 274, + 296, + 328 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 61, + 274, + 296, + 328 + ], + "spans": [ + { + "bbox": [ + 61, + 274, + 296, + 328 + ], + "type": "text", + "content": "[2] Antonin Berthon, Bo Han, Gang Niu, Tongliang Liu, and Masashi Sugiyama. Confidence scores make instance-dependent label-noise learning possible. In International conference on machine learning, pages 825-836. PMLR, 2021. 1" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 62, + 331, + 295, + 385 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 331, + 295, + 385 + ], + "spans": [ + { + "bbox": [ + 62, + 331, + 295, + 385 + ], + "type": "text", + "content": "[3] Stella Biderman, Usvsn Prashanth, Lintang Sutawika, Hailey Schoelkopf, Quentin Anthony, Shivanshu Purohit, and Edward Raff. Emergent and predictable memorization in large language models. Advances in Neural Information Processing Systems, 36:28072-28090, 2023. 1" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 62, + 387, + 295, + 441 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 387, + 295, + 441 + ], + "spans": [ + { + "bbox": [ + 62, + 387, + 295, + 441 + ], + "type": "text", + "content": "[4] Jian Chen, Ruiyi Zhang, Tong Yu, Rohan Sharma, Zhiqiang Xu, Tong Sun, and Changyou Chen. Label-retrieval augmented diffusion models for learning from noisy labels. Advances in Neural Information Processing Systems, 36, 2024. 2, 3, 7, 8" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 62, + 443, + 294, + 497 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 443, + 294, + 497 + ], + "spans": [ + { + "bbox": [ + 62, + 443, + 294, + 497 + ], + "type": "text", + "content": "[5] Pengfei Chen, Ben Ben Liao, Guangyong Chen, and Shengyu Zhang. Understanding and utilizing deep neural networks trained with noisy labels. In International conference on machine learning, pages 1062-1070. PMLR, 2019. 1" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 62, + 499, + 294, + 554 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 499, + 294, + 554 + ], + "spans": [ + { + "bbox": [ + 62, + 499, + 294, + 554 + ], + "type": "text", + "content": "[6] Pengfei Chen, Junjie Ye, Guangyong Chen, Jingwei Zhao, and Pheng-Ann Heng. Beyond class-conditional assumption: A primary attempt to combat instance-dependent label noise. In Proceedings of the AAAI Conference on Artificial Intelligence, pages 11442–11450, 2021. 1, 2, 6, 7" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 62, + 555, + 294, + 600 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 555, + 294, + 600 + ], + "spans": [ + { + "bbox": [ + 62, + 555, + 294, + 600 + ], + "type": "text", + "content": "[7] Ting Chen, Simon Kornblith, Mohammad Norouzi, and Geoffrey Hinton. A simple framework for contrastive learning of visual representations. In International conference on machine learning, pages 1597-1607. PMLR, 2020. 2, 3, 7" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 62, + 601, + 295, + 644 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 601, + 295, + 644 + ], + "spans": [ + { + "bbox": [ + 62, + 601, + 295, + 644 + ], + "type": "text", + "content": "[8] Yingyi Chen, Shell Xu Hu, Xi Shen, Chunrong Ai, and Johan AK Suykens. Compressing features for learning with noisy labels. IEEE Transactions on Neural Networks and Learning Systems, 35(2):2124-2138, 2022. 7" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 62, + 647, + 295, + 689 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 647, + 295, + 689 + ], + "spans": [ + { + "bbox": [ + 62, + 647, + 295, + 689 + ], + "type": "text", + "content": "[9] Yingyi Chen, Xi Shen, Yahui Liu, Qinghua Tao, and Johan AK Suykens. Jigsaw-vit: Learning jigsaw puzzles in vision transformer. Pattern Recognition Letters, 166:53-60, 2023. 7, 8" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 57, + 691, + 294, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 691, + 294, + 713 + ], + "spans": [ + { + "bbox": [ + 57, + 691, + 294, + 713 + ], + "type": "text", + "content": "[10] De Cheng, Tongliang Liu, Yixiong Ning, Nannan Wang, Bo Han, Gang Niu, Xinbo Gao, and Masashi Sugiyama." + } + ] + } + ], + "index": 12 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 316, + 73, + 555, + 712 + ], + "type": "list", + "angle": 0, + "index": 27, + "blocks": [ + { + "bbox": [ + 333, + 73, + 555, + 117 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 333, + 73, + 555, + 117 + ], + "spans": [ + { + "bbox": [ + 333, + 73, + 555, + 117 + ], + "type": "text", + "content": "Instance-dependent label-noise learning with manifold-regularized transition matrix estimation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 16630-16639, 2022. 1" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 316, + 119, + 554, + 161 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 119, + 554, + 161 + ], + "spans": [ + { + "bbox": [ + 316, + 119, + 554, + 161 + ], + "type": "text", + "content": "[11] Hao Cheng, Zhaowei Zhu, Xingyu Li, Yifei Gong, Xing Sun, and Yang Liu. Learning with instance-dependent label noise: A sample sieve approach. arXiv preprint arXiv:2010.02347, 2020." + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 317, + 163, + 553, + 207 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 163, + 553, + 207 + ], + "spans": [ + { + "bbox": [ + 317, + 163, + 553, + 207 + ], + "type": "text", + "content": "[12] Jiacheng Cheng, Tongliang Liu, Kotagiri Ramamohanarao, and Dacheng Tao. Learning with bounded instance and label-dependent label noise. In International conference on machine learning, pages 1789-1799. PMLR, 2020. 1" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 317, + 209, + 554, + 273 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 209, + 554, + 273 + ], + "spans": [ + { + "bbox": [ + 317, + 209, + 554, + 273 + ], + "type": "text", + "content": "[13] C Fabian Benitez-Quiroz, Ramprakash Srinivasan, and Aleix M Martinez. Emotionet: An accurate, real-time algorithm for the automatic annotation of a million facial expressions in the wild. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 5562–5570, 2016. 1" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 316, + 276, + 553, + 317 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 276, + 553, + 317 + ], + "spans": [ + { + "bbox": [ + 316, + 276, + 553, + 317 + ], + "type": "text", + "content": "[14] Chen Feng, Georgios Tzimiropoulos, and Ioannis Patras. Ssr: An efficient and robust framework for learning with unknown label noise. arXiv preprint arXiv:2111.11288, 2021. 7" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 317, + 321, + 554, + 375 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 321, + 554, + 375 + ], + "spans": [ + { + "bbox": [ + 317, + 321, + 554, + 375 + ], + "type": "text", + "content": "[15] Arpit Garg, Cuong Nguyen, Rafael Felix, Thanh-Toan Do, and Gustavo Carneiro. Instance-dependent noisy label learning via graphical modelling. In Proceedings of the IEEE/CVF winter conference on applications of computer vision, pages 2288-2298, 2023. 2" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 317, + 377, + 553, + 420 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 377, + 553, + 420 + ], + "spans": [ + { + "bbox": [ + 317, + 377, + 553, + 420 + ], + "type": "text", + "content": "[16] Aritra Ghosh, Himanshu Kumar, and P Shanti Sastry. Robust loss functions under label noise for deep neural networks. In Proceedings of the AAAI conference on artificial intelligence, 2017. 1, 2" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 317, + 422, + 553, + 476 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 422, + 553, + 476 + ], + "spans": [ + { + "bbox": [ + 317, + 422, + 553, + 476 + ], + "type": "text", + "content": "[17] Bo Han, Quanming Yao, Xingrui Yu, Gang Niu, Miao Xu, Weihua Hu, Ivor Tsang, and Masashi Sugiyama. Co-teaching: Robust training of deep neural networks with extremely noisy labels. Advances in neural information processing systems, 31, 2018. 1, 2, 7, 8" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 317, + 478, + 553, + 521 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 478, + 553, + 521 + ], + "spans": [ + { + "bbox": [ + 317, + 478, + 553, + 521 + ], + "type": "text", + "content": "[18] Xizewen Han, Huangjie Zheng, and Mingyuan Zhou. Card: Classification and regression diffusion models. Advances in Neural Information Processing Systems, 35:18100-18115, 2022. 2, 3" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 317, + 523, + 553, + 577 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 523, + 553, + 577 + ], + "spans": [ + { + "bbox": [ + 317, + 523, + 553, + 577 + ], + "type": "text", + "content": "[19] Kaiming He, Xiangyu Zhang, Shaoqing Ren, and Jian Sun. Delving deep into rectifiers: Surpassing human-level performance on imagenet classification. In Proceedings of the IEEE international conference on computer vision, pages 1026-1034, 2015. 2" + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 317, + 579, + 553, + 623 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 579, + 553, + 623 + ], + "spans": [ + { + "bbox": [ + 317, + 579, + 553, + 623 + ], + "type": "text", + "content": "[20] Kaiming He, Xiangyu Zhang, Shaoqing Ren, and Jian Sun. Deep residual learning for image recognition. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 770-778, 2016. 7" + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 316, + 624, + 553, + 657 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 624, + 553, + 657 + ], + "spans": [ + { + "bbox": [ + 316, + 624, + 553, + 657 + ], + "type": "text", + "content": "[21] Jonathan Ho, Ajay Jain, and Pieter Abbeel. Denoising diffusion probabilistic models. Advances in neural information processing systems, 33:6840-6851, 2020. 2, 3, 5" + } + ] + } + ], + "index": 25 + }, + { + "bbox": [ + 317, + 658, + 553, + 712 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 658, + 553, + 712 + ], + "spans": [ + { + "bbox": [ + 317, + 658, + 553, + 712 + ], + "type": "text", + "content": "[22] Ahmet Iscen, Jack Valmadre, Anurag Arnab, and Cordelia Schmid. Learning with neighbor consistency for noisy labels. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 4672-4681, 2022. 2, 5, 7" + } + ] + } + ], + "index": 26 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 318, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 318, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 318, + 757 + ], + "type": "text", + "content": "25746" + } + ] + } + ], + "index": 28 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 8 + }, + { + "para_blocks": [ + { + "bbox": [ + 56, + 72, + 295, + 712 + ], + "type": "list", + "angle": 0, + "index": 14, + "blocks": [ + { + "bbox": [ + 56, + 72, + 294, + 126 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 72, + 294, + 126 + ], + "spans": [ + { + "bbox": [ + 56, + 72, + 294, + 126 + ], + "type": "text", + "content": "[23] Gaoxia Jiang, Jia Zhang, Xuefei Bai, Wenjian Wang, and Deyu Meng. Which is more effective in label noise cleaning, correction or filtering? In Proceedings of the AAAI Conference on Artificial Intelligence, pages 12866-12873, 2024. 2" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 56, + 130, + 294, + 184 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 130, + 294, + 184 + ], + "spans": [ + { + "bbox": [ + 56, + 130, + 294, + 184 + ], + "type": "text", + "content": "[24] Lu Jiang, Zhengyuan Zhou, Thomas Leung, Li-Jia Li, and Li Fei-Fei. Mentornet: Learning data-driven curriculum for very deep neural networks on corrupted labels. In International conference on machine learning, pages 2304–2313. PMLR, 2018. 1" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 57, + 186, + 295, + 209 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 186, + 295, + 209 + ], + "spans": [ + { + "bbox": [ + 57, + 186, + 295, + 209 + ], + "type": "text", + "content": "[25] Diederik P Kingma. Auto-encoding variational bayes. arXiv preprint arXiv:1312.6114, 2013. 2" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 57, + 211, + 295, + 255 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 211, + 295, + 255 + ], + "spans": [ + { + "bbox": [ + 57, + 211, + 295, + 255 + ], + "type": "text", + "content": "[26] Jongwoo Ko, Sumyeong Ahn, and Se-Young Yun. Efficient utilization of pre-trained model for learning with noisy labels. In ICLR 2023 Workshop on Pitfalls of limited data and computation for Trustworthy ML, 2023. 7, 8" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 57, + 257, + 294, + 279 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 257, + 294, + 279 + ], + "spans": [ + { + "bbox": [ + 57, + 257, + 294, + 279 + ], + "type": "text", + "content": "[27] Alex Krizhevsky. Learning multiple layers of features from tiny images. Master's thesis, University of Toronto, 2009. 6" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 57, + 281, + 294, + 313 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 281, + 294, + 313 + ], + "spans": [ + { + "bbox": [ + 57, + 281, + 294, + 313 + ], + "type": "text", + "content": "[28] Junnan Li, Richard Socher, and Steven CH Hoi. Dividemix: Learning with noisy labels as semi-supervised learning. arXiv preprint arXiv:2002.07394, 2020. 1, 2, 8" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 57, + 316, + 294, + 370 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 316, + 294, + 370 + ], + "spans": [ + { + "bbox": [ + 57, + 316, + 294, + 370 + ], + "type": "text", + "content": "[29] Mingchen Li, Mahdi Soltanolkotabi, and Samet Oymak. Gradient descent with early stopping is provably robust to label noise for overparameterized neural networks. In International conference on artificial intelligence and statistics, pages 4313-4324. PMLR, 2020. 1" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 57, + 373, + 294, + 415 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 373, + 294, + 415 + ], + "spans": [ + { + "bbox": [ + 57, + 373, + 294, + 415 + ], + "type": "text", + "content": "[30] Wen Li, Limin Wang, Wei Li, Eirikur Agustsson, and Luc Van Gool. Webvision database: Visual learning and understanding from web data. arXiv preprint arXiv:1708.02862, 2017. 7" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 57, + 418, + 294, + 473 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 418, + 294, + 473 + ], + "spans": [ + { + "bbox": [ + 57, + 418, + 294, + 473 + ], + "type": "text", + "content": "[31] Yifan Li, Hu Han, Shiguang Shan, and Xilin Chen. Disc: Learning from noisy labels via dynamic instance-specific selection and correction. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 24070-24079, 2023. 2, 7, 8" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 57, + 475, + 294, + 529 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 475, + 294, + 529 + ], + "spans": [ + { + "bbox": [ + 57, + 475, + 294, + 529 + ], + "type": "text", + "content": "[32] Jiawei Liu, Qiang Wang, Huijie Fan, Yinong Wang, Yandong Tang, and Liangqiong Qu. Residual denoising diffusion models. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 2773-2783, 2024. 2" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 57, + 532, + 294, + 576 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 532, + 294, + 576 + ], + "spans": [ + { + "bbox": [ + 57, + 532, + 294, + 576 + ], + "type": "text", + "content": "[33] Sheng Liu, Jonathan Niles-Weed, Narges Razavian, and Carlos Fernandez-Granda. Early-learning regularization prevents memorization of noisy labels. Advances in neural information processing systems, 33:20331-20342, 2020. 2, 8" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 57, + 578, + 294, + 610 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 578, + 294, + 610 + ], + "spans": [ + { + "bbox": [ + 57, + 578, + 294, + 610 + ], + "type": "text", + "content": "[34] Yangdi Lu, Yang Bo, and Wenbo He. Co-matching: Combating noisy labels by augmentation anchoring. arXiv preprint arXiv:2103.12814, 2021. 1" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 57, + 613, + 294, + 655 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 613, + 294, + 655 + ], + "spans": [ + { + "bbox": [ + 57, + 613, + 294, + 655 + ], + "type": "text", + "content": "[35] Anqi Mao, Mehryar Mohri, and Yutao Zhong. Cross-entropy loss functions: Theoretical analysis and applications. In International conference on Machine learning, pages 23803–23828. PMLR, 2023. 7" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 57, + 658, + 294, + 712 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 658, + 294, + 712 + ], + "spans": [ + { + "bbox": [ + 57, + 658, + 294, + 712 + ], + "type": "text", + "content": "[36] Duc Tam Nguyen, Chaithanya Kumar Mummadi, Thi Phuong Nhung Ngo, Thi Hoai Phuong Nguyen, Laura Beggel, and Thomas Brox. Self: Learning to filter noisy labels with self-ensembling. arXiv preprint arXiv:1910.01842, 2019. 1" + } + ] + } + ], + "index": 13 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 316, + 73, + 553, + 713 + ], + "type": "list", + "angle": 0, + "index": 28, + "blocks": [ + { + "bbox": [ + 316, + 73, + 553, + 117 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 73, + 553, + 117 + ], + "spans": [ + { + "bbox": [ + 316, + 73, + 553, + 117 + ], + "type": "text", + "content": "[37] Kento Nishi, Yi Ding, Alex Rich, and Tobias Hollerer. Augmentation strategies for learning with noisy labels. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 8022-8031, 2021. 1" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 317, + 119, + 553, + 161 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 119, + 553, + 161 + ], + "spans": [ + { + "bbox": [ + 317, + 119, + 553, + 161 + ], + "type": "text", + "content": "[38] Haim Permuter, Joseph Francos, and Ian Jermyn. A study of gaussian mixture models of color and texture features for image classification and segmentation. Pattern recognition, 39(4):695-706, 2006. 5" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 317, + 163, + 553, + 186 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 163, + 553, + 186 + ], + "spans": [ + { + "bbox": [ + 317, + 163, + 553, + 186 + ], + "type": "text", + "content": "[39] Lutz Prechelt. Early stopping-but when? In Neural Networks: Tricks of the trade, pages 55-69. Springer, 2002. 1" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 317, + 186, + 553, + 251 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 186, + 553, + 251 + ], + "spans": [ + { + "bbox": [ + 317, + 186, + 553, + 251 + ], + "type": "text", + "content": "[40] Alec Radford, Jong Wook Kim, Chris Hallacy, Aditya Ramesh, Gabriel Goh, Sandhini Agarwal, Girish Sastry, Amanda Askell, Pamela Mishkin, Jack Clark, et al. Learning transferable visual models from natural language supervision. In International conference on machine learning, pages 8748-8763. PMLR, 2021. 2, 7" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 317, + 254, + 553, + 308 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 254, + 553, + 308 + ], + "spans": [ + { + "bbox": [ + 317, + 254, + 553, + 308 + ], + "type": "text", + "content": "[41] Olga Russakovsky, Jia Deng, Hao Su, Jonathan Krause, Sanjeev Satheesh, Sean Ma, Zhiheng Huang, Andrej Karpathy, Aditya Khosla, Michael Bernstein, et al. Imagenet large scale visual recognition challenge. International journal of computer vision, 115:211-252, 2015. 7" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 317, + 310, + 553, + 374 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 310, + 553, + 374 + ], + "spans": [ + { + "bbox": [ + 317, + 310, + 553, + 374 + ], + "type": "text", + "content": "[42] Kihyuk Sohn, David Berthelot, Nicholas Carlini, Zizhao Zhang, Han Zhang, Colin A Raffel, Ekin Dogus Cubuk, Alexey Kurakin, and Chun-Liang Li. Fixmatch: Simplifying semi-supervised learning with consistency and confidence. Advances in neural information processing systems, 33:596-608, 2020. 5" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 317, + 376, + 553, + 420 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 376, + 553, + 420 + ], + "spans": [ + { + "bbox": [ + 317, + 376, + 553, + 420 + ], + "type": "text", + "content": "[43] Hwanjun Song, Minseok Kim, and Jae-Gil Lee. Selfie: Refurbishing unclean samples for robust deep learning. In International conference on machine learning, pages 5907-5915. PMLR, 2019. 7" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 317, + 422, + 553, + 453 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 422, + 553, + 453 + ], + "spans": [ + { + "bbox": [ + 317, + 422, + 553, + 453 + ], + "type": "text", + "content": "[44] Jiaming Song, Chenlin Meng, and Stefano Ermon. Denoising diffusion implicit models. arXiv preprint arXiv:2010.02502, 2020. 5" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 317, + 456, + 553, + 499 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 456, + 553, + 499 + ], + "spans": [ + { + "bbox": [ + 317, + 456, + 553, + 499 + ], + "type": "text", + "content": "[45] Sunil Thulasidasan, Tanmoy Bhattacharya, Jeff Bilmes, Gopinath Chennupati, and Jamal Mohd-Yusof. Combating label noise in deep learning using abstention. arXiv preprint arXiv:1905.10964, 2019. 2, 7" + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 317, + 501, + 553, + 567 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 501, + 553, + 567 + ], + "spans": [ + { + "bbox": [ + 317, + 501, + 553, + 567 + ], + "type": "text", + "content": "[46] Yuanpeng Tu, Boshen Zhang, Yuxi Li, Liang Liu, Jian Li, Jiangning Zhang, Yabiao Wang, Chengjie Wang, and Cai Rong Zhao. Learning with noisy labels via self-supervised adversarial noisy masking. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 16186-16195, 2023. 8" + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 317, + 568, + 553, + 622 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 568, + 553, + 622 + ], + "spans": [ + { + "bbox": [ + 317, + 568, + 553, + 622 + ], + "type": "text", + "content": "[47] Yisen Wang, Weiyang Liu, Xingjun Ma, James Bailey, Hongyuan Zha, Le Song, and Shu-Tao Xia. Iterative learning with open-set noisy labels. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 8688-8696, 2018. 1" + } + ] + } + ], + "index": 25 + }, + { + "bbox": [ + 317, + 624, + 553, + 677 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 624, + 553, + 677 + ], + "spans": [ + { + "bbox": [ + 317, + 624, + 553, + 677 + ], + "type": "text", + "content": "[48] Yisen Wang, Xingjun Ma, Zaiyi Chen, Yuan Luo, Jinfeng Yi, and James Bailey. Symmetric cross entropy for robust learning with noisy labels. In Proceedings of the IEEE/CVF international computer vision and pattern recognition, pages 322-330, 2019. 1" + } + ] + } + ], + "index": 26 + }, + { + "bbox": [ + 317, + 680, + 553, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 680, + 553, + 713 + ], + "spans": [ + { + "bbox": [ + 317, + 680, + 553, + 713 + ], + "type": "text", + "content": "[49] Hongxin Wei, Lei Feng, Xiangyu Chen, and Bo An. Combating noisy labels by agreement: A joint training method with co-regularization. In Proceedings of the IEEE/CVF con" + } + ] + } + ], + "index": 27 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "text", + "content": "25747" + } + ] + } + ], + "index": 29 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 9 + }, + { + "para_blocks": [ + { + "bbox": [ + 56, + 73, + 295, + 713 + ], + "type": "list", + "angle": 0, + "index": 14, + "blocks": [ + { + "bbox": [ + 75, + 73, + 294, + 94 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 75, + 73, + 294, + 94 + ], + "spans": [ + { + "bbox": [ + 75, + 73, + 294, + 94 + ], + "type": "text", + "content": "ference on computer vision and pattern recognition, pages 13726-13735, 2020. 1" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 56, + 96, + 295, + 140 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 96, + 295, + 140 + ], + "spans": [ + { + "bbox": [ + 56, + 96, + 295, + 140 + ], + "type": "text", + "content": "[50] Pengxiang Wu, Songzhu Zheng, Mayank Goswami, Dimitris Metaxas, and Chao Chen. A topological filter for learning with label noise. Advances in neural information processing systems, 33:21382-21393, 2020. 1" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 56, + 141, + 294, + 196 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 141, + 294, + 196 + ], + "spans": [ + { + "bbox": [ + 56, + 141, + 294, + 196 + ], + "type": "text", + "content": "[51] Xiaobo Xia, Tongliang Liu, Bo Han, Nannan Wang, Mingming Gong, Haifeng Liu, Gang Niu, Dacheng Tao, and Masashi Sugiyama. Part-dependent label noise: Towards instance-dependent label noise. Advances in Neural Information Processing Systems, 33:7597-7610, 2020. 1" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 56, + 198, + 294, + 251 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 198, + 294, + 251 + ], + "spans": [ + { + "bbox": [ + 56, + 198, + 294, + 251 + ], + "type": "text", + "content": "[52] Tong Xiao, Tian Xia, Yi Yang, Chang Huang, and Xiaogang Wang. Learning from massive noisy labeled data for image classification. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 2691-2699, 2015. 7" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 56, + 253, + 294, + 285 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 253, + 294, + 285 + ], + "spans": [ + { + "bbox": [ + 56, + 253, + 294, + 285 + ], + "type": "text", + "content": "[53] Cai Xu, Yilin Zhang, Ziyu Guan, and Wei Zhao. Trusted multi-view learning with label noise. arXiv preprint arXiv:2404.11944, 2024. 5" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 56, + 287, + 294, + 331 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 287, + 294, + 331 + ], + "spans": [ + { + "bbox": [ + 56, + 287, + 294, + 331 + ], + "type": "text", + "content": "[54] Yilun Xu, Peng Cao, Yuqing Kong, and Yizhou Wang. L.dmi: A novel information-theoretic loss function for training deep nets robust to label noise. Advances in neural information processing systems, 32, 2019. 2, 7, 8" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 56, + 332, + 294, + 374 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 332, + 294, + 374 + ], + "spans": [ + { + "bbox": [ + 56, + 332, + 294, + 374 + ], + "type": "text", + "content": "[55] Yan Yan, Rómer Rosales, Glenn Fung, Ramanathan Subramanian, and Jennifer Dy. Learning from multiple annotators with varying expertise. Machine learning, 95:291-327, 2014. 1" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 56, + 378, + 294, + 422 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 378, + 294, + 422 + ], + "spans": [ + { + "bbox": [ + 56, + 378, + 294, + 422 + ], + "type": "text", + "content": "[56] Quanming Yao, Hansi Yang, Bo Han, Gang Niu, and James Tin-Yau Kwok. Searching to exploit memorization effect in learning with noisy labels. In International Conference on Machine Learning, pages 10789-10798. PMLR, 2020. 1" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 56, + 423, + 294, + 465 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 423, + 294, + 465 + ], + "spans": [ + { + "bbox": [ + 56, + 423, + 294, + 465 + ], + "type": "text", + "content": "[57] Yu Yao, Tongliang Liu, Mingming Gong, Bo Han, Gang Niu, and Kun Zhang. Instance-dependent label-noise learning under a structural causal model. Advances in Neural Information Processing Systems, 34:4409-4420, 2021. 1" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 56, + 468, + 294, + 520 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 468, + 294, + 520 + ], + "spans": [ + { + "bbox": [ + 56, + 468, + 294, + 520 + ], + "type": "text", + "content": "[58] Xingrui Yu, Bo Han, Jiangchao Yao, Gang Niu, Ivor Tsang, and Masashi Sugiyama. How does disagreement help generalization against label corruption? In International conference on machine learning, pages 7164-7173. PMLR, 2019. 1" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 56, + 523, + 294, + 567 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 523, + 294, + 567 + ], + "spans": [ + { + "bbox": [ + 56, + 523, + 294, + 567 + ], + "type": "text", + "content": "[59] Chiyuan Zhang, Samy Bengio, Moritz Hardt, Benjamin Recht, and Oriol Vinyals. Understanding deep learning (still) requires rethinking generalization. Communications of the ACM, 64(3):107-115, 2021. 1" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 56, + 568, + 294, + 611 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 568, + 294, + 611 + ], + "spans": [ + { + "bbox": [ + 56, + 568, + 294, + 611 + ], + "type": "text", + "content": "[60] Yikai Zhang, Songzhu Zheng, Pengxiang Wu, Mayank Goswami, and Chao Chen. Learning with feature-dependent label noise: A progressive approach. arXiv preprint arXiv:2103.07756, 2021. 1" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 56, + 613, + 294, + 656 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 613, + 294, + 656 + ], + "spans": [ + { + "bbox": [ + 56, + 613, + 294, + 656 + ], + "type": "text", + "content": "[61] Zhilu Zhang and Mert Sabuncu. Generalized cross entropy loss for training deep neural networks with noisy labels. Advances in neural information processing systems, 31, 2018, 1, 2, 7" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 56, + 658, + 294, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 658, + 294, + 713 + ], + "spans": [ + { + "bbox": [ + 56, + 658, + 294, + 713 + ], + "type": "text", + "content": "[62] Ganlong Zhao, Guanbin Li, Yipeng Qin, Feng Liu, and Yizhou Yu. Centrality and consistency: two-stage clean samples identification for learning with instance-dependent noisy labels. In European Conference on Computer Vision, pages 21-37. Springer, 2022. 2, 7, 8" + } + ] + } + ], + "index": 13 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 316, + 73, + 553, + 217 + ], + "type": "list", + "angle": 0, + "index": 18, + "blocks": [ + { + "bbox": [ + 316, + 73, + 553, + 128 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 73, + 553, + 128 + ], + "spans": [ + { + "bbox": [ + 316, + 73, + 553, + 128 + ], + "type": "text", + "content": "[63] Evgenii Zheltonozhskii, Chaim Baskin, Avi Mendelson, Alex M Bronstein, and Or Litany. Contrast to divide: Self-supervised pre-training for learning with noisy labels. In Proceedings of the IEEE/CVF Winter Conference on Applications of Computer Vision, pages 1657-1667, 2022. 2, 8" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 316, + 129, + 553, + 172 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 129, + 553, + 172 + ], + "spans": [ + { + "bbox": [ + 316, + 129, + 553, + 172 + ], + "type": "text", + "content": "[64] Songzhu Zheng, Pengxiang Wu, Aman Goswami, Mayank Goswami, Dimitris Metaxas, and Chao Chen. Error-bounded correction of noisy labels. In International Conference on Machine Learning, pages 11447-11457. PMLR, 2020. 1" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 316, + 174, + 553, + 217 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 174, + 553, + 217 + ], + "spans": [ + { + "bbox": [ + 316, + 174, + 553, + 217 + ], + "type": "text", + "content": "[65] Yuanyi Zhong, Haoran Tang, Junkun Chen, Jian Peng, and Yu-Xiong Wang. Is self-supervised learning more robust than supervised learning? arXiv preprint arXiv:2206.05259, 2022. 2" + } + ] + } + ], + "index": 17 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 749, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 749, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 749, + 317, + 757 + ], + "type": "text", + "content": "25748" + } + ] + } + ], + "index": 19 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 10 + } + ], + "_backend": "vlm", + "_version_name": "2.6.4" +} \ No newline at end of file diff --git a/2025/Disco4D_ Disentangled 4D Human Generation and Animation from a Single Image/5f039bdf-5dd5-496e-b90f-b0a10473c8cd_content_list.json b/2025/Disco4D_ Disentangled 4D Human Generation and Animation from a Single Image/5f039bdf-5dd5-496e-b90f-b0a10473c8cd_content_list.json new file mode 100644 index 0000000000000000000000000000000000000000..abde37bf221d0019987468711f25a3b39bbba56d --- /dev/null +++ b/2025/Disco4D_ Disentangled 4D Human Generation and Animation from a Single Image/5f039bdf-5dd5-496e-b90f-b0a10473c8cd_content_list.json @@ -0,0 +1,1512 @@ +[ + { + "type": "text", + "text": "Disco4D: Disentangled 4D Human Generation and Animation from a Single Image", + "text_level": 1, + "bbox": [ + 91, + 130, + 906, + 154 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Hui En Pang $^{1}$ , Shuai Liu $^{3}$ , Zhongang Cai $^{1,2,3}$ , Lei Yang $^{2,3}$ , Tianwei Zhang $^{1}$ , Ziwei Liu $^{1}$", + "bbox": [ + 236, + 178, + 753, + 215 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "$^{1}$ S-Lab, Nanyang Technological University $^{2}$ SenseTime Research $^{3}$ Shanghai AI Laboratory", + "bbox": [ + 116, + 215, + 874, + 234 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/95e2c828af253cd8c94fd8208bcc383a17ce69683a910e6599a18759da9fab83.jpg", + "image_caption": [ + "Figure 1. Disco4D is a novel Gaussian Splatting framework for 4D disentangled human generation from a single image. Clothing and assets are modeled as Gaussians on top of the SMPL-X body, enabling dynamic animation and flexible editing." + ], + "image_footnote": [], + "bbox": [ + 102, + 252, + 893, + 393 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Abstract", + "text_level": 1, + "bbox": [ + 248, + 430, + 326, + 445 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "We present Disco4D, a novel Gaussian Splatting framework for 4D human generation and animation from a single image. Different from existing methods, Disco4D distinctively disentangles clothings (with Gaussian models) from the human body (with SMPL-X model), significantly enhancing the generation details and flexibility. Specifically, 1) Disco4D learns to efficiently fit the clothing Gaussians over the SMPL-X Gaussians. 2) Next, Disco4D adopts diffusion models to enhance the 3D generation process, e.g., modeling occluded parts not visible in the input image. 3) Finally, Disco4D learns an identity encoding for each clothing Gaussian to facilitate the separation and extraction of clothing assets. Furthermore, Disco4D naturally supports 4D human animation with vivid dynamics. Extensive experiments demonstrate the superiority of Disco4D on 4D human generation and animation tasks. Our code is available at https://github.com/disco-4d/Disco4D.", + "bbox": [ + 88, + 463, + 485, + 720 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "1. Introduction", + "text_level": 1, + "bbox": [ + 89, + 753, + 220, + 768 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "The development of high-fidelity 3D digital humans is increasingly important across a variety of augmented and virtual reality applications. To streamline the creation of these digital avatars from easily accessible in-the-wild images, a multitude of research efforts have been made on reconstructing 3D clothed human models from a single image [2, 33, 40-42, 81, 82, 103, 104, 104, 117]. These works predominantly focus on the simultaneous reconstruction of", + "bbox": [ + 88, + 779, + 485, + 901 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "the human body and clothing. Unfortunately, these works have inherent limitations, and integrating them into applications that require virtual try-on or avatar customization poses significant challenges. This is primarily because the models are rendered as single-layer, non-animatable meshes where distinct attributes (e.g., hair, clothing, accessories) are merged into one continuous surface, with underlying layers completely obscured and self-contact areas inseparably connected. Such limitation complicates the re-animation and dynamic customization tasks. Existing works that perform layered reconstruction [25, 26] rely on self-rotating video inputs with extensive frames and viewpoints and involve substantial processing times.", + "bbox": [ + 511, + 430, + 908, + 628 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "To address these issues, we propose Disco4D, a novel 4D clothed human reconstruction method that distinctly separates the human body from clothing elements from a single image. It supports human animation as the 4th dimension, which cannot be realized by prior static 3D reconstruction works [2, 33, 42, 81, 82, 103, 104, 117]. To achieve this, it employs the SMPL-X [71] parametric model to represent the human body, capitalizing on its efficacy in capturing body structure and kinematics. Conversely, clothing, along with dynamic and variable elements such as hair and accessories, is represented using Gaussian models, which are able to model the large variability in clothing. By binding Gaussians to a SMPL-X model and fixing it during the training phase, Disco4D ensures the integrity of the body while focusing the learning process on the appearance aspects. To model occluded portions not visible in the input image, diffusion models are used to enhance the 3D generation process. Moreover, Disco4D includes an identity grouping mecha", + "bbox": [ + 511, + 628, + 910, + 901 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "CVF", + "bbox": [ + 106, + 2, + 181, + 42 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore.", + "bbox": [ + 236, + 0, + 810, + 46 + ], + "page_idx": 0 + }, + { + "type": "page_number", + "text": "26331", + "bbox": [ + 478, + 944, + 517, + 955 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "nism for the Gaussians, which is instrumental in maintaining the separability and individuality of each clothing asset.", + "bbox": [ + 89, + 90, + 482, + 119 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "The independent reconstruction of clothing and body offers several advantages. (1) Enhanced reconstruction fidelity. The SMPL-X body serves as a stable anchor for the clothing to conform to. By isolating the focus to learn clothing Gaussians, we achieve a more refined geometry and intricate detailing in the clothed model. (2) Fine-grained categorization and extraction of clothing items. Disco4D is able to separate clothing Gaussians into their respective categories, which is crucial for the recovery and utilization of individual clothing assets. (3) Extensive editing capabilities. Disco4D supports different editing functions, including the removal of specific items, inpainting (altering color or material), and other modifications. Such rich editing options allow for precise adjustments to individual assets without inadvertently affecting adjacent elements. This level of control is particularly beneficial in applications requiring detailed customization, such as virtual fashion design and digital content creation. (4) Improved animation capabilities. The body Gaussians adhere to the deformations dictated by the SMPL-X model, while clothing Gaussians conform to the underlying body movements but also exhibit behaviors true to their material characteristics. The disentangled deformation allows for nuanced adjustments to clothing behavior in response to complex body movements, thereby elevating the quality of clothed human animation.", + "bbox": [ + 91, + 122, + 485, + 498 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "2. Related Works", + "text_level": 1, + "bbox": [ + 89, + 511, + 241, + 527 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Table 1 summarizes the relevant 3D/4D generation methods. We describe their details below.", + "bbox": [ + 89, + 537, + 485, + 566 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "2.1.3D Generation", + "text_level": 1, + "bbox": [ + 89, + 575, + 241, + 590 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Single-image 3D Generation. Single-image reconstruction leverages advanced methods [45, 67] to generate 3D assets in the form of 3D point clouds or NeRF [65] from one image. While earlier efforts using auto-encoders focused on synthetic objects [12, 14, 21, 88, 93, 105], newer approaches treat the task as conditional generation, employing diffusion models [35] for 3D generation from both image and text [19, 35, 60, 64, 74, 76, 79, 90]. One-2-3-45 [59] uses 2D diffusion models [60, 87] to generate multiview images for reconstruction, while LRM [36] adopts transformer-based architecture to scale up the task on large datasets [19, 112]. Gaussian-based methods [47], particularly DreamGaussian [89] and LGM [91], offer efficient, high-resolution 3D model generation from text or images. Recently, video diffusion models have attracted significant attention due to their remarkable ability to generate intricate scenes and complex dynamics with great spatio-temporal consistency [4, 7-9, 31, 56, 116]. They are employed to generate consistent multi-view images, and then reconstruct underlying 3D assets with high quality [15].", + "bbox": [ + 89, + 598, + 485, + 901 + ], + "page_idx": 1 + }, + { + "type": "table", + "img_path": "images/2963fa4e815a0db6e0cb0f849e4b7dc57e671763a5eef8ce332a45006557f049.jpg", + "table_caption": [ + "Table 1. 3D/4D generation methods from a single image." + ], + "table_footnote": [], + "table_body": "
MethodTypeLayeredAnimatable
LGM [91]GeneralXX
PiFU [81]Human-centricXX
DreamFusion [74]GeneralXX
DreamGaussian [89]GeneralXX
PiFU [81]Human-centricXX
D-IF [107]Human-centricXX
HiLo [108]Human-centricXX
ECON [104]Human-centricXX
SHERF [40]Human-centricX✓
Disco4DHuman-centric✓✓
", + "bbox": [ + 565, + 109, + 854, + 231 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Single-image human-centric 3D Generation. Significant research efforts have been made for 3D human reconstruction, which can be classified into the following categories. (1) Explicit-shape-based methods rely on Human Mesh Recovery (HMR) using parametric models like SMPL [62] and SMPL-X [71] to generate 3D body meshes [16, 17, 22, 24, 44, 46, 48, 49, 51, 66, 80, 118]. To account for 3D garments, several approaches incorporate offsets [99, 119] or templates, utilize deformable garment templates [6, 43], or employ non-parametric forms for clothed figures [27, 104, 115]. Despite their advancements, they face limitations in handling complex outfit variations and loose clothing due to inherent topological constraints. (2) Implicit-function-based methods utilize implicit representations like occupancy or distance fields for modeling clothed humans with complex geometries, such as loose garments. Techniques range from end-to-end regression of free-form implicit surfaces [2, 81, 82] to use of geometric priors [33, 42, 103, 104, 117] and implicit shape completion [104]. Notable works such as PIFu [81], ARCH(++) [33, 42], and PaMIR [117] can extract textured models from images, but struggle with depth ambiguities and texture inconsistencies. (3) NeRF-based methods incorporate model-based priors (i.e., SMPL-X) for accurate human reconstruction. Efforts like SHERF [40] and ELICIT [41] improve the reconstruction coherence by addressing 2D observation incompleteness leveraging appearance priors. Most of these 3D clothed human reconstruction and animation works [2, 33, 42, 81, 82, 103, 104, 117] require training on human-specific datasets, which brings another limitation on the availability of such datasets.", + "bbox": [ + 511, + 246, + 908, + 715 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "3D Clothing Modeling. Reconstructing clothing from images and videos as a separate layer over the human body poses significant challenges due to the diversity of clothing topologies. Previous efforts relied on either template meshes or implicit surface models, and required extensive, high-quality 3D data from simulations [5, 70, 83, 95] or tailored template meshes [13, 32, 73, 100]. New methods were developed [34, 43] for multi-clothing models and versatile template meshes, respectively, facilitating diverse clothing topology encoding. However, these techniques typically fall short in capturing the clothing texture and appearance. The reliance on predefined clothing style templates further con", + "bbox": [ + 511, + 719, + 910, + 901 + ], + "page_idx": 1 + }, + { + "type": "page_number", + "text": "26332", + "bbox": [ + 478, + 944, + 519, + 957 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "strains their ability to handle real-world clothing variations. Corona et al. [18] addressed these shortcomings by representing clothing layers with deep unsigned distance functions and an auto-decoder for style and cut differentiation, though this often produces overly-smooth reconstructions [18]. On the other hand, SCARF [25] and DELTA [26] significantly enhance the visual fidelity by applying NeRF to clothing layers, but require self-rotating video inputs and considerable processing times.", + "bbox": [ + 89, + 90, + 485, + 227 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "2.2. 4D Animation", + "text_level": 1, + "bbox": [ + 89, + 246, + 235, + 260 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "4D Animation. This task aims at capturing dynamic 3D scenes over time. Two primary approaches have emerged: modeling 4D scenes by adding time dimension $t$ or latent codes to spatial coordinates [28, 57, 98]; combining deformation fields with static 3D scenes [20, 58, 68, 69, 75, 92, 114]. Recent efforts in explicit or hybrid representations, like planar decomposition [11, 84, 85], hash representations [94], and other innovative methods [1, 23, 29], have improved reconstruction speed and quality. Gaussian Splatting, especially, stands out for balancing efficiency with quality, with dynamic 3D Gaussians [63] and 4D Gaussian Splatting [97, 110] techniques introducing time-dependent deformations to enhance reconstructions. Notably, DreamGaussian4D [78] stands out by minimizing the optimization time while achieving high-quality 4D reconstructions.", + "bbox": [ + 89, + 271, + 483, + 498 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Human-centric 4D Animation. Recent works leverage Gaussian-based methods [38, 50, 54, 55, 61, 77, 113] for 4D human reconstruction, requiring extensive frame sequences (50-100 frames) and/or multiple viewpoints. Currently there has not been any work on 4D layered human generation and animation from a single image or a video with few images, which will be achieved in this paper.", + "bbox": [ + 89, + 501, + 483, + 608 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3. Methodology", + "text_level": 1, + "bbox": [ + 89, + 630, + 225, + 648 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3.1. Preliminaries", + "text_level": 1, + "bbox": [ + 89, + 657, + 230, + 672 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3D Gaussian Splatting employs explicit 3D Gaussian points as its primary rendering entities. A 3D Gaussian point is defined as a function $G(x) = e^{-\\frac{1}{2}(x - \\mu)^T\\Sigma^{-1}(x - \\mu)}$ , where $\\mu$ and $\\Sigma$ are the spatial mean and covariance matrix, respectively. Each Gaussian is also associated with its own rotation $r$ , scaling $s$ , opacity $\\alpha$ , a view-dependent color $c$ represented by spherical harmonic coefficients $f$ .", + "bbox": [ + 89, + 684, + 483, + 792 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "SMPL-X parameterization [71] is an extension of the SMPL body model [62] with face and hand, designed to capture a more accurate representation of intricate body movements. SMPL-X is defined as a function $M(\\beta, \\theta, \\psi) : \\mathbb{R}^{|\\beta| \\times |\\theta| \\times |\\psi|} \\to \\mathbb{R}^{3N}$ , parametrized by the pose $\\theta \\in \\mathbb{R}^{3J}$ (where $J$ denotes the number of full body joints), body shape $\\beta \\in \\mathbb{R}^{|\\beta|}$ and facial expression $\\psi \\in \\mathbb{R}^{|\\psi|}$ .", + "bbox": [ + 89, + 795, + 483, + 901 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3.2. Overview", + "text_level": 1, + "bbox": [ + 513, + 90, + 624, + 104 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Given a single image, Disco4D generates animatable 3D clothed human avatars in a bottom-up manner, facilitating natural separability. Our generated 3D clothed avatars, denoted as $S_{human}$ , are represented as the concatenation of $S_{body}$ and $S_{cloth}$ . Inspired by prior works [89, 91], $S$ capitalizes on Gaussian representations:", + "bbox": [ + 511, + 112, + 906, + 203 + ], + "page_idx": 2 + }, + { + "type": "equation", + "text": "\n$$\nS = G (\\mu , r, s, \\alpha , c, e), \\tag {1}\n$$\n", + "text_format": "latex", + "bbox": [ + 632, + 215, + 906, + 233 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "where $\\mu, r, s, \\alpha, c$ and $e$ denote positions, rotation, scaling, opacity, spherical harmonics coefficients and identity encoding, respectively. Different from traditional Gaussian representations, we add identity encoding $e$ to associate each Gaussian with its clothing category.", + "bbox": [ + 511, + 244, + 906, + 320 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Figure 2 depicts our framework. We start by generating colored SMPL-X Gaussians representing the body beneath clothing (Sec. 3.3). We obtain a visual hull for canonicalization and refine Gaussian predictions to align and envelop the SMPL-X mesh (Sec. 3.4). Next, we iteratively optimize canonical clothing Gaussians external to the SMPL-X mesh (Sec. 3.5). Lastly, we showcase the animation and editing of generated clothed avatars (Sec. 3.6). Notably, we leverage diffusion models to refine textures during 3D generation (Sec. 3.5) and extrapolate unseen views during 4D animation (Sec. 3.6).", + "bbox": [ + 511, + 321, + 908, + 487 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3.3. SMPL-X Gaussians", + "text_level": 1, + "bbox": [ + 513, + 497, + 700, + 511 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Given an image, we first estimate coarse SMPL-X parameters with an off-the-shelf model [10], and then refine coarse predictions by fitting on 2D keypoints and clothing segmentation masks [72], obtaining pixel-aligned SMPL-X parameters $(\\beta, \\theta, \\psi)$ .", + "bbox": [ + 511, + 518, + 906, + 594 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Mesh Binding. To convert the SMPL-X [71] mesh $M(\\beta, \\theta, \\psi)$ into Gaussians $S_{body}$ for rendering, flat 3D Gaussians are bound to each mesh triangle, similar to SuGaR [30]. Gaussian means $\\mu_{body}$ are computed using predefined barycentric coordinates, while Gaussian rotations $r_{body}$ derive from surface normals. The initial scaling $s_{body}$ ensures dense mesh coverage, with the last axis set to 0.1 for a uniformly thin surface. For color representation beneath clothing, opacity $\\alpha_{body}$ is set to 1.0, with spherical harmonics $c_{body}$ optimized for each Gaussian. Visible skin color is supervised, while occluded skin color aligns with visible regions. A fixed label $e_{body}$ is assigned for rendering, remaining unchanged during training. When optimizing clothing Gaussians $S_{cloth}$ , SMPL-X Gaussians $S_{body}$ parameters stay fixed, preserving the body structure while allowing flexible learning for clothing.", + "bbox": [ + 511, + 595, + 908, + 838 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3.4. Initialization of Clothing Gaussians", + "text_level": 1, + "bbox": [ + 511, + 847, + 821, + 864 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Cloth styles are diverse, making proper initialization crucial for effective clothing modeling. In synchronization with", + "bbox": [ + 511, + 869, + 906, + 900 + ], + "page_idx": 2 + }, + { + "type": "page_number", + "text": "26333", + "bbox": [ + 478, + 944, + 517, + 955 + ], + "page_idx": 2 + }, + { + "type": "image", + "img_path": "images/a42620725e9459866868a1c7ce133c3e2b8b35f0d94e5e20c14ece6d76cafb14.jpg", + "image_caption": [ + "Figure 2. Framework Overview of Disco4D. (a) 3D Generation utilizes a single image to obtain disentangled body and clothing Gaussians. Body, face and hand poses are refined to be pixel-aligned. For faster initialization, clothing Gaussians and visual hull are obtained with Gaussian Reconstruction Models. These clothing Gaussians are embedded to SMPL-X mesh and adopt the local coordinate system of the triangle. Subsequently, the iterative optimization process (pruning, identity encoding and densifying) separates the body and garments. The learned identity encodings guide the densification of the clothing Gaussians. (b) 4D Animation is achieved by either direct driving of SMPL-X poses or leveraging video to learn extra clothing deformation (refer to Figure 3 for more details). Various (c) 3D/4D Editing operations can be performed with our disentangled representation." + ], + "image_footnote": [], + "bbox": [ + 117, + 85, + 880, + 292 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "estimating SMPL-X, we first employ the Video Diffusion Model [7] to estimate multi-view images. Subsequently, we leverage Gaussian Reconstruction Models [91] to obtain initial 3D Gaussians and their corresponding visual hull. Yet, the reconstructed 3D outputs often suffer from geometric inaccuracies, such as incorrect poses due to pose ambiguity or missing limbs. To address this, we refine the coarse visual hull to ensure it accurately aligns with and overlays the SMPL-X mesh and encapsulates a good geometry for the clothed figure. With SMPL-X aligned visual hull, we derive the refined Gaussians by adopting properties from their nearest neighbors. The refined visual hull and Gaussians are then canonicalized for the optimization phase.", + "bbox": [ + 88, + 400, + 485, + 595 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Mesh embedding. Each 3D clothing Gaussian is embedded on a triangle of the canonical mesh, defining its position in both canonical and posed spaces. The mean vertex position $\\mathbf{O}$ serves as the origin of the local coordinate system, with the Gaussian positioned by an offset vector $\\mathbf{v} = \\sigma \\mathbf{i} + \\beta \\mathbf{j} + \\gamma \\mathbf{k}$ , where $\\sigma$ , $\\beta$ , and $\\gamma$ are the components of the displacement vector along the tangent $\\mathbf{i}$ , bitangent $\\mathbf{j}$ , and normal $\\mathbf{k}$ . Unlike SplattingAvatar [86], which displaces Gaussians along the normal, our approach allows embedding to the most suitable triangle rather than the nearest one. For example, hair Gaussians are tagged to head faces instead of the nearest face for reposing [39, 86] (Figure 9 in Appendix). In animation, the Gaussian rotates with its embedded triangle face $(\\delta r)$ , while scaling $(\\delta s)$ is adjusted dynamically based on changes in edge lengths. During optimization, Gaussian and embedding parameters $(\\mathbf{O}, \\mathbf{v}, \\delta r, \\text{and } \\delta s)$ are jointly updated.", + "bbox": [ + 89, + 597, + 485, + 839 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "3.5. Optimization of Separable Gaussians", + "text_level": 1, + "bbox": [ + 89, + 848, + 413, + 864 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "With the SMPL-X Gaussian and initialized clothing Gaussian, we aim to optimize canonical clothing Gaussians $S_{cloth}$", + "bbox": [ + 89, + 869, + 483, + 902 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "outside the SMPL-X mesh. This involves three steps: 1) we use Signed Distance Function (SDF) loss and pruning to discourage and remove Gaussians that reside within the body; 2) we introduce identity encoding $e$ to attach a clothing label for each clothing Gaussian, by lifting multi-view 2D segmentations of the target object onto the 3D Gaussians; and 3) guided by $e_{body}$ and $e_{cloth}$ , we selectively densify only the relevant clothing points while ignoring body points. Once the disentangled clothing is obtained, we use SDS loss to in-paint high-resolution texture from the reference image to individual clothing Gaussians, thereby enriching the details of unseen regions.", + "bbox": [ + 511, + 400, + 906, + 582 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "SDF Loss and Pruning. In reality, the clothing is always external to the body. During refinement, we ensure that the clothing Gaussians are positioned externally to the SMPL-X mesh by applying the SDF loss and a pruning strategy. Specifically, the SDF loss $\\mathcal{L}_{sdf}$ penalizes any new densified Gaussians that intrude into the space of the SMPL-X mesh, ensuring that the clothing Gaussians consistently remain outside the body's surface. Pruning is applied at fixed intervals to reinforce this separation, and systematically remove any Gaussians located within the SDF of the SMPL-X mesh.", + "bbox": [ + 511, + 580, + 908, + 733 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Identity encoding. To associate each Gaussian to its clothing category, we introduce Identity Encoding $(e)$ , a learnable and compact vector of length 15, representing clothing categories from SegFormer [101] segmentation masks1. During training, the encodings are rendered into 2D segmentation masks in a differentiable manner following [111]. For classification, we apply a softmax to the rendered features $E_{\\mathrm{id}}$ and use cross-entropy loss $\\mathcal{L}_{2d}$ for $(K + 1)$ -category classifica", + "bbox": [ + 511, + 732, + 908, + 854 + ], + "page_idx": 3 + }, + { + "type": "page_footnote", + "text": "1Categories: 0: \"Background\", 1: \"Hat\", 2: \"Hair\", 3: \"Sunglasses\", 4: \"Upper-clothes\", 5: \"Skirt\", 6: \"Pants\", 7: \"Dress\", 8: \"Belt\", 9: \"Left-shoe\", 10: \"Right-shoe\", 11: \"Face\", 12: \"Skin\", 13: \"Bag\", 14: \"Scarf\"", + "bbox": [ + 511, + 862, + 906, + 902 + ], + "page_idx": 3 + }, + { + "type": "page_number", + "text": "26334", + "bbox": [ + 478, + 944, + 519, + 957 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "tion. An unsupervised 3D regularization loss $\\mathcal{L}_{3d}$ promotes spatial consistency among the top $k$ -nearest 3D Gaussians' Identity Encodings. Consequently, the overall identity loss is $\\mathcal{L}_{id} = \\mathcal{L}_{2d} + \\mathcal{L}_{3d}$ . Refer to Appendix 7.2 for more details. Densification of clothing Gaussians. To learn clothing more efficiently, we perform sampling for categorical Gaussians that belong to the same clothing category and embedding. We find the $k$ -nearest Gaussian points for the resampled points and inherit their Gaussian properties (scaling, rotation, opacity, SH properties). By selectively densifying clothing Gaussians, we only add necessary Gaussians while ignoring body Gaussians.", + "bbox": [ + 89, + 90, + 483, + 272 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Anisotropy. To prevent overly-skinny kernels that point outward from the object surface under large deformations, we enforce the anisotropy of Gaussian kernels following [102]. During optimization, we employ $\\mathcal{L}_{\\mathrm{ani}} = \\frac{1}{|P|}\\sum_{p\\in P}\\max \\left(\\frac{\\max(s_p)}{\\min(s_p)},\\tau\\right) - \\tau$ , where $s_p$ is the scalings of 3D Gaussians. This loss constrains the ratio between the major and minor axis lengths below threshold $\\tau$ .", + "bbox": [ + 89, + 273, + 483, + 383 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Total loss. To inpaint occluded textures, we use the $\\mathcal{L}_{SDS}$ loss on the Gaussians in the canonical pose after optimizing the front view for 500 steps. Combined with the conventional 3D Gaussian Loss $\\mathcal{L}_{ori}$ on image rendering, the total loss $L$ for end-to-end optimization of clothing Gaussians via network $C_N$ is:", + "bbox": [ + 89, + 385, + 483, + 474 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {L} = \\lambda_ {\\text {o r i}} \\mathcal {L} _ {\\text {o r i}} + \\lambda_ {\\text {i d}} \\mathcal {L} _ {\\text {i d}} + \\lambda_ {\\text {a n i}} \\mathcal {L} _ {\\text {a n i}} + \\lambda_ {\\text {s d f}} \\mathcal {L} _ {\\text {s d f}} + \\lambda_ {\\text {S D S}} \\mathcal {L} _ {\\text {S D S}} \\tag {2}\n$$\n", + "text_format": "latex", + "bbox": [ + 94, + 483, + 483, + 513 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "3.6. 4D Human Animation and Editing", + "text_level": 1, + "bbox": [ + 89, + 517, + 393, + 534 + ], + "page_idx": 4 + }, + { + "type": "image", + "img_path": "images/0f4a91071a9f36466d5bc8dae4a308e65ce2cab82b34981bcabd105fe8f64912.jpg", + "image_caption": [ + "a) SMPL-X Pose Driven Animation" + ], + "image_footnote": [], + "bbox": [ + 117, + 554, + 441, + 608 + ], + "page_idx": 4 + }, + { + "type": "image", + "img_path": "images/18159098878650bcb17563218212498802af2f353eefda2c9c7cb61d3d527ddb.jpg", + "image_caption": [ + "b) Video-to-4D Reconstruction", + "Figure 3. 4D animation is achieved by (a) driving SMPL-X poses or (b) using video to learn additional clothing deformations. From the first frame, a static 3D disentangled GS model is generated. Pose transformations deform body and clothing Gaussians, and a deformation network is optimized to capture additional clothing deformations over time." + ], + "image_footnote": [], + "bbox": [ + 109, + 623, + 467, + 820 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Disco4D's disentangled representation naturally supports animation and editing. The canonical Gaussians $S_{body}$ and $S_{cloth}$ enable separate deformations for clothing and body, ensuring realistic animation. Besides, individual clothing categories can be easily edited using image or text prompts. The learned clothing can be transferred to different body shapes and poses, for versatile customization.", + "bbox": [ + 511, + 90, + 906, + 196 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Animating Gaussians. As shown in Figure 2, Disco4D enables animation of the canonical human Gaussian via two methods. Firstly, Gaussians can be directly driven using 3D SMPL-X sequences obtained from a motion database or estimated from 2D videos. Secondly, Disco4D enhances the model by learning detailed clothing dynamics from monocular videos. This disentanglement enables the focused modeling of clothing dynamics without altering the underlying human representation.", + "bbox": [ + 511, + 198, + 908, + 335 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "To extend static 3D Gaussians into dynamic 4D Gaussians, a deformation network is trained to predict changes in position, rotation, and scale of the reposed clothing Gaussians based on a timestamp, as described in DreamGaussian4D [78]. Unlike [78], which learns deformations for all Gaussians, Disco4D models body Gaussians using the SMPL-X mesh, while clothing Gaussians employ posed transformations and learned deformations. The transformation is defined as $S'' = D_N(S', t)$ where $D_N$ is the deformation network, $S'$ is the spatial descriptions of the reposed 3D clothing Gaussian, $t$ is the timestamp, and $S''$ is the spatial descriptions of the deformed and reposed 3D clothing Gaussians. Following [78], the deformation model is initialized to predict zero deformation at the start of training to avoid divergence between dynamic and static models. The weights and biases of the final prediction heads are initialized to zero, and skip connections are introduced to enable gradient backpropagation.", + "bbox": [ + 511, + 335, + 908, + 609 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "To optimize the deformation field using the reference view video, we minimize the reconstruction loss $\\mathcal{L}_{Ref}$ between the rendered image and video frame at each timestep. To propagate the motion from the reference view to the entire 3D model, we leverage Zero-1-to-3-XL [19] to predict the deformation of the unseen part to calculate $\\mathcal{L}_{SDS}$ . Despite per-frame predictions of image diffusion models, the fixed color and opacity of static 3D Gaussians help preserve temporal consistency.", + "bbox": [ + 511, + 611, + 908, + 747 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Editing Clothing Gaussians. We extract the Gaussians corresponding to the specific category and edit them. This allows fine-grained editing and ensures that other Gaussians are not affected. Instead of fine-tuning all 3D Gaussians, we freeze the properties for most of the well-trained Gaussians and only adjust a small part of 3D Gaussians relevant to the target categories. For 3D object removal, we simply delete the 3D Gaussians of the editing target. For 3D object colorization by in-painting or text guidance, we reinitialize the color and tune the color (SH) parameters of the correspond-", + "bbox": [ + 511, + 750, + 910, + 901 + ], + "page_idx": 4 + }, + { + "type": "page_number", + "text": "26335", + "bbox": [ + 478, + 944, + 517, + 955 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "ing Gaussian group, while fixing the 3D positions and other properties to preserve the learned 3D geometry.", + "bbox": [ + 89, + 90, + 485, + 122 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "4. Experiments", + "text_level": 1, + "bbox": [ + 89, + 133, + 225, + 152 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Our detailed implementation and experiment setup can be found in Appendix 7.3.", + "bbox": [ + 89, + 159, + 482, + 191 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "4.1.3D Generation", + "text_level": 1, + "bbox": [ + 89, + 198, + 241, + 213 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Generation and Disentanglement. Our generation and disentanglement results are presented in Figure 4 and Table 2. We assessed the disentanglement quality using the Synbody [109] and CloSe [3] datasets, rendering 30 and 110 clothed human meshes respectively from four angles and evaluating CLIP-similarity, PSNR, SSIM, and LPIPS for various poses and views within the CloSe dataset. disco4D leverages diffusion models without requiring training on human specific datasets. Therefore, we compare it with DreamGaussian [89] and LGM [91] which reconstruct 3D objects from diffusion models. Additionally, we conducted comparisons with SHERF, a human-centric baseline for evaluating novel poses and views. Figure 4 shows disco4D has higher fidelity and better geometry for body parts such as face and limbs due to the representation using SMPL-X Gaussians. It outperforms DreamGaussian and SHERF on SynBody and CloSe benchmarks. disco4D performs worse than LGM on novel views, likely due to its optimization of Gaussians in canonical space for pose generalization, compromising view-specific detail. Editing. We can edit specific clothing appearance given an image or text prompt, repose the person and transfer person characteristics. The disentanglement allows fine-grained editing and modification of individual assets without affecting other assets, and stacking multiple edits (Figure 4). User study. We conducted a user study to evaluate the generative quality of our image-to-3D Gaussians reconstruction on random in-the-wild images from SHHQ, detailed in Table 3. This study focuses on reference view consistency and overall generation quality, crucial aspects in image reconstruction tasks. We rendered 360-degree rotation videos for 25 images generated by DreamGaussian, LGM, and disco4D. We invited 43 volunteers to rate $24\\sim 27$ mixed samples from these methods on image consistency and overall model quality, yielding 1080 valid scores. As shown in Table 3, disco4D was preferred, demonstrating better alignment with the original image content and superior overall quality.", + "bbox": [ + 89, + 220, + 485, + 765 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "4.2. 4D Animation", + "text_level": 1, + "bbox": [ + 89, + 773, + 236, + 787 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Pose-Driven Animation. disco4D generates canonical Gaussians that can be animated with any pose sequence. Figure 12 in the Appendix demonstrates our animation capabilities and compares them with current SOTA 2D animation methods. Using identical inputs—a single frame and pose sequence—our approach more effectively preserves the body shape and fine details such as facial features and clothing. It", + "bbox": [ + 89, + 794, + 483, + 901 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "surpasses Animate-Anyone [37] and Magic-Animate [106] in accurately modeling fine-grained body parts like hands and faces, and exhibits greater consistency compared to CHAMP [120]. The disentanglement feature of Disco4D further allows for direct manipulation of Clothing Gaussians, as shown in Figure 6.", + "bbox": [ + 511, + 90, + 906, + 181 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "4D Reconstruction. For the 4D-Dress Dataset [96], we evaluated 8 sequences, assessing CLIP similarity scores against ground-truth meshes and disentangled assets, along with novel view performance (PSNR, SSIM, LPIPS) from four viewpoints. Table 4 summarizes our quantitative results, benchmarking Disco4D against existing video-to-4D general GS approaches, such as DreamGaussian4D [78], as well as human-centric GS methods, including MonoHuman [113], GART [54], and GaussianAvatar [38]. We evaluate on monocular videos comprising 14 frames, captured from a limited front-view perspective, without full-body visibility across frames.", + "bbox": [ + 511, + 184, + 908, + 364 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Disco4D outperforms MonoHuman [113], GART [54], and GaussianAvatar [38] (Table 4) as these methods reconstruct using known video information, unable to model unseen regions. Consequently, these methods cannot accurately model back views from front-facing videos, leading to artifacts in other perspectives and canonical space (see Figure 5). In contrast, Disco4D first performs reconstruction and subsequently incorporates details, such as clothing deformation, from the input frames, enabling consistent reconstruction even in unseen viewpoints.", + "bbox": [ + 511, + 367, + 908, + 518 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "While DreamGaussian4D [78] is capable of modeling back-view information, the details remain coarse. Our results demonstrate that initializing with our model from the first frame (DreamGaussian4D Disco4D-init) significantly outperforms other initialization methods (DreamGaussian4DLGM init, DreamGaussian init) in both fidelity and geometry (Table 4). Nevertheless, without incorporating human priors, DreamGaussian4D [78] still faces challenges, such as missing limbs and difficulty modeling fine details like facial features (see Figure 13 in Appendix).", + "bbox": [ + 511, + 520, + 908, + 672 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Reposing our canonical avatar enables us to align the body and assets accurately with the inferred postures from the source video, yielding high-quality reconstruction of faces, hands, and garments. Our reposed method surpasses DreamGaussian4D in geometry and fidelity by incorporating human priors. However, reposing alone cannot capture clothing dynamics. To address this, our disentangled approach models clothing deformations on the reposed Gaussians, guided by a diffusion model. As demonstrated in Figure 13 and Table 4, this process enhances the accuracy of clothing resemblance to the ground truth. The combination of asset repositioning and learned deformations improves modeling quality, with repositioning handling pose-driven changes and learned deformations simulating dynamic asset movements as observed in the driving video.", + "bbox": [ + 511, + 674, + 910, + 900 + ], + "page_idx": 5 + }, + { + "type": "page_number", + "text": "26336", + "bbox": [ + 478, + 945, + 519, + 955 + ], + "page_idx": 5 + }, + { + "type": "table", + "img_path": "images/f8a529b962d52f736868671611637835ae46b599aa08d4ce47c5180fad51cd45.jpg", + "table_caption": [ + "Table 2. CLIP-embedding loss for generated humans and segmented assets, and performance (PSNR, SSIM, LPIPS) comparisons for novel poses and views on the Synbody and CloSe datasets across DreamGaussian, LGM, SHERF, and Disco4D." + ], + "table_footnote": [], + "table_body": "
MethodSynBodyCloSe
CLIPNovel ViewCLIPNovel ViewNovel Pose
All ↑Pants ↑Shirt ↑Shoes ↑PSNR ↑SSIM ↑LPIPS ↓All ↑Pants ↑Shirt ↑Shoes ↑PSNR ↑SSIM ↑LPIPS ↓PSNR ↑SSIM ↑LPIPS ↓
DreamGaussian0.7510.7150.7100.74913.1180.8830.2290.7340.6930.6740.76720.080.9390.089---
LGM0.8070.7240.7470.76012.8840.8760.2280.8290.7270.7120.77820.500.9390.077---
SHERF0.7660.6490.6360.71415.1890.8520.1890.7770.7850.7290.80118.960.9120.08315.540.8440.165
Disco4D0.8510.7840.7530.80115.6910.8480.1850.8560.8580.8100.84220.100.9180.08117.960.8510.136
", + "bbox": [ + 91, + 117, + 903, + 202 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/d4c087aa346e36287966f162e97293bec9476662da2b7b9f9bf420bcef256338.jpg", + "image_caption": [ + "Figure 4. Qualitative comparison of image generation across DreamGaussian, LGM, SHERF, and Disco4D." + ], + "image_footnote": [], + "bbox": [ + 101, + 214, + 890, + 467 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/57fd506cb5851f16f4b2827cb1b859baacff518a16203678fe9f75a94c0c5543.jpg", + "image_caption": [ + "Figure 5. Qualitative comparison of 4D generation between DreamGaussian4D, MonoHuman, GART, GaussianAvatar, and Disco4D." + ], + "image_footnote": [], + "bbox": [ + 112, + 491, + 880, + 861 + ], + "page_idx": 6 + }, + { + "type": "page_number", + "text": "26337", + "bbox": [ + 478, + 945, + 519, + 955 + ], + "page_idx": 6 + }, + { + "type": "table", + "img_path": "images/4b2da948bac6ce9f771e239d3de5f968ccd804fd06c3345ed3cba2c13270a4c0.jpg", + "table_caption": [ + "Table 3. User study rates quality of generated 3D Gaussians from 1-5. The higher the better." + ], + "table_footnote": [], + "table_body": "
MetricImage Consistency ↑Overall Quality ↑
DreamGaussian2.0171.852
LGM2.3382.017
Disco4D3.1423.037
", + "bbox": [ + 151, + 117, + 419, + 172 + ], + "page_idx": 7 + }, + { + "type": "table", + "img_path": "images/dbfe3cda0184f66b403e6885cc745a30396ef77df4d91d224cddfc6d86550869.jpg", + "table_caption": [ + "Table 4. CLIP-embedding loss for generated humans and segmented assets, and performance (PSNR, SSIM, LPIPS) comparison on 4D-Dress across various video-to-4D methods." + ], + "table_footnote": [], + "table_body": "
All ↑Assets ↑PSNR ↑SSIM ↑LPIPS ↓
DreamGaussian4D0.7840.76920.540.930.080
MonoHuman0.7620.74320.220.920.086
GART0.8000.77218.810.920.086
GaussianAvatar0.8220.76820.010.930.069
DreamGaussian4D (LGM init)0.8090.79519.160.930.086
DreamGaussian4D (Disco4D init)0.8700.84921.020.930.065
Disco4D (reposed)0.8530.77423.940.950.049
Disco4D (reposed)+learned deformations0.9000.86525.460.960.035
", + "bbox": [ + 91, + 227, + 483, + 325 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/465ac7c73588c24ace8492ee9eb936cb4a42e2d7e2432f4d7178a055f7c44cce.jpg", + "image_caption": [ + "Figure 6. First frame Editing and Animation. Betas Editing, Recoloring (Text/Image-guided), Composition (Removal, Swap)." + ], + "image_footnote": [], + "bbox": [ + 94, + 327, + 478, + 501 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "4D Editing. For a normal pipeline in character animation, editing the person in the video requires high consistency throughout all frames. For pose-driven animation methods, first frame editing and generation is required. Our method directly edits the Gaussians, which is more straightforward, fine-grained and consistent. This is seen from Figure 6.", + "bbox": [ + 89, + 547, + 483, + 640 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "4.3. Ablation Studies", + "text_level": 1, + "bbox": [ + 89, + 650, + 254, + 666 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Initialization of Clothing Gaussians. This process is crucial for high fidelity reconstruction. As shown in Figure 8 in the Appendix, we evaluate different strategies, including random, surface, and hull-based initialization. Hull-based initialization significantly enhances the model accuracy and realism over other methods. Initialization directly on the SMPL-X surface often leads to inaccurate geometries, particularly with complex or loose garments, creating elongated, thin Gaussians and visual artifacts. In contrast, hull-based initialization captures garment details more effectively and maintains pose consistency, closely aligning with the true geometry of the clothed body.", + "bbox": [ + 89, + 672, + 482, + 853 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Geometry of Clothing Gaussians. Figure 14 in the Appendix highlights the differences in clothing geometry between DreamGaussian [89], LGM [91] and Disco4D. In", + "bbox": [ + 89, + 854, + 483, + 900 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "DreamGaussian, all points are confined within the body geometry, whereas in LGM, about half of the points extend beyond the SMPL-X body. Removing internal points leaves sparse, translucent representations for clothing. This sparsity suggests reliance on internal points for visual representation, failing to accurately depict the object's geometry where appearance should primarily originate from surface points. Often, clothing Gaussian points are incorrectly positioned inside the body's hull rather than on the surface. To better represent clothing geometry, Disco4D positions all clothing Gaussians externally to the SMPL-X body mesh, accurately reflecting the garment's actual physical characteristics.", + "bbox": [ + 511, + 90, + 906, + 272 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Clothing editing. Figure 14 shows our editing results with the prompt \"Color the top pink\". disco4D allows for precise editing of the targeted clothing without affecting other areas.", + "bbox": [ + 511, + 273, + 906, + 319 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "5. Discussion", + "text_level": 1, + "bbox": [ + 513, + 334, + 625, + 349 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Despite achieving impressive results, some failure cases still exist, as shown in Figure 7 in the Appendix. disco4D relies on robust and pixel-aligned SMPL-X estimation, which is still an unsolved problem. It occasionally fails for poor visual hull initialization. The extraction of mesh assets from clothing Gaussians using Local Density Query, as per DreamGaussian [89], currently loses fine-grained details. Enhancing the detail level of geometry derived from clothing Gaussians could bolster the utility of reconstructed assets in animation and simulation applications. Furthermore, the initialized visual hulls obtained from multi-view SMPL-X guided images are often of suboptimal quality and suffer from poor side and back views, necessitating refinement. Improving pose guidance models to achieve more accurate visual hulls could alleviate the need for extensive refinement. In addition, future works could look into modeling multi-layered clothing and reconstructing the occluded clothing. disco4D has many positive applications, but it also has the potential to facilitate deepfake avatars and raise IP concerns. Regulations should be built to address these issues alongside its benefits in the entertainment industry.", + "bbox": [ + 511, + 359, + 908, + 678 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "6. Conclusion", + "text_level": 1, + "bbox": [ + 513, + 693, + 633, + 709 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "We propose Disco4D, a novel approach for the generation of 3D animatable clothed human Gaussians from a single image, emphasizing high-fidelity detail and separation of assets. We manage to compositionally generate separate components, such as haircut, accessories, and decoupled outfits. Our core insight is the fixing of SMPL-X Gaussians, fitting segmented Gaussians over SMPL-X Gaussians, and application of diffusion models to enhance 3D reconstruction, including modeling occluded parts not visible in the input image. Its capability to separate assets offers significant advantages, including localized, fine-grained editing of individual assets and enhanced animatability.", + "bbox": [ + 511, + 719, + 908, + 900 + ], + "page_idx": 7 + }, + { + "type": "page_number", + "text": "26338", + "bbox": [ + 478, + 945, + 517, + 955 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Acknowledgements", + "text_level": 1, + "bbox": [ + 91, + 90, + 258, + 107 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "We sincerely thank the anonymous reviewers for their valuable comments on this paper. This research/project is supported by the National Research Foundation, Singapore under its AI Singapore Programme(AISG Award No: AISG3-PhD-2023-08-049T). This study is supported by the Ministry of Education, Singapore, under its MOE AcRF Tier 2 (MOET2EP20221-0012, MOE-T2EP20223-0002), and under the RIE2020 Industry Alignment Fund - Industry Collaboration Projects (IAF-ICP) Funding Initiative, as well as cash and in-kind contribution from the industry partner(s).", + "bbox": [ + 89, + 116, + 485, + 268 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "References", + "text_level": 1, + "bbox": [ + 91, + 284, + 187, + 300 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[1] Jad Abou-Chakra, Feras Dayoub, and Niko Sunderhauf. Particlenerf: Particle based encoding for online neural radiance fields. arXiv preprint arXiv:2211.04041, 2022.3", + "[2] Thiemo Alldieck, Mihai Zanfir, and Cristian Sminchisescu. Photorealistic Monocular 3D Reconstruction of Humans Wearing Clothing. Proceedings of the IEEE Computer Society Conference on Computer Vision and Pattern Recognition, 2022-June:1496-1505, 2022. 1, 2", + "[3] Dimitrije Antić, Garvita Tiwari, Batuhan Ozcomlekci, Riccardo Marin, and Gerard Pons-Moll. CloSe: A 3D clothing segmentation dataset and model. In International Conference on 3D Vision (3DV), 2024. 6", + "[4] Omer Bar-Tal, Hila Chefer, Omer Tov, Charles Herrmann, Roni Paiss, Shiran Zada, Ariel Ephrat, Junhwa Hur, Guanghui Liu, Amit Raj, Yanzhen Li, Michael Rubinstein, Tomer Michaeli, Oliver Wang, Deqing Sun, Tali Dekel, and Inbar Mosseri. Lumiere: A spacetime diffusion model for video generation, 2024. 2", + "[5] Hugo Bertiche, Meysam Madadi, and Sergio Escalera. Cloth3d: clothed 3d humans. In European Conference on Computer Vision, pages 344-359. Springer, 2020. 2", + "[6] Bharat Lal Bhatnagar, Garvita Tiwari, Christian Theobalt, and Gerard Pons-Moll. Multi-garment net: Learning to dress 3d people from images. In IEEE International Conference on Computer Vision (ICCV). IEEE, 2019. 2", + "[7] Andreas Blattmann, Tim Dockhorn, Sumith Kulal, Daniel Mendelevitch, Maciej Kilian, Dominik Lorenz, Yam Levi, Zion English, Vikram Voleti, Adam Letts, Varun Jampani, and Robin Rombach. Stable video diffusion: Scaling latent video diffusion models to large datasets, 2023. 2, 4", + "[8] Andreas Blattmann, Robin Rombach, Huan Ling, Tim Dockhorn, Seung Wook Kim, Sanja Fidler, and" + ], + "bbox": [ + 107, + 309, + 486, + 901 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "Karsten Kreis. Align your latents: High-resolution video synthesis with latent diffusion models, 2023.", + "[9] Tim Brooks, Bill Peebles, Connor Holmes, Will DePue, Yufei Guo, Li Jing, David Schnurr, Joe Taylor, Troy Luhman, Eric Luhman, Clarence Ng, Ricky Wang, and Aditya Ramesh. Video generation models as world simulators. 2024. 2", + "[10] Zhongang Cai, Wanqi Yin, Ailing Zeng, Chen Wei, Qingping Sun, Wang Yanjun, Hui En Pang, Haiyi Mei, Mingyuan Zhang, Lei Zhang, et al. Smpler-x: Scaling up expressive human pose and shape estimation. Advances in Neural Information Processing Systems, 36, 2024. 3", + "[11] Ang Cao and Justin Johnson. Hexplane: A fast representation for dynamic scenes. CVPR, 2023. 3", + "[12] Angel X. Chang, Thomas Funkhouser, Leonidas Guibas, Pat Hanrahan, Qixing Huang, Zimo Li, Silvio Savarese, Manolis Savva, Shuran Song, Hao Su, Jianxiong Xiao, Li Yi, and Fisher Yu. ShapeNet: An Information-Rich 3D Model Repository. Technical Report arXiv:1512.03012 [cs.GR], Stanford University — Princeton University — Toyota Technological Institute at Chicago, 2015. 2", + "[13] Xin Chen, Anqi Pang, Yang Wei, Wang Peihao, Lan Xu, and Jingyi Yu. Tightcap: 3d human shape capture with clothing tightness field. ACM Transactions on Graphics (Presented at ACM SIGGRAPH), 2021. 2", + "[14] Zhiqin Chen, Andrea Tagliasacchi, and Hao Zhang. BSP-net: Generating compact meshes via binary space partitioning. Proceedings of IEEE Conference on Computer Vision and Pattern Recognition (CVPR), 2020. 2", + "[15] Zilong Chen, Yikai Wang, Feng Wang, Zhengyi Wang, and Huaping Liu. V3d: Video diffusion models are effective 3d generators, 2024. 2", + "[16] Hongsuk Choi, Gyeongsik Moon, and Kyoung Mu Lee. Pose2Mesh: Graph Convolutional Network for 3D Human Pose and Mesh Recovery from a 2D Human Pose. Lecture Notes in Computer Science (including subseries Lecture Notes in Artificial Intelligence and Lecture Notes in Bioinformatics), 12352 LNCS: 769-787, 2020. 2", + "[17] Vasileios Choutas, Georgios Pavlakos, Timo Bolkart, Dimitrios Tzionas, and Michael J. Black. Monocular expressive body regression through body-driven attention. In European Conference on Computer Vision (ECCV), 2020. 2", + "[18] Enric Corona, Albert Pumarola, Guillem Alenyà, Gerard Pons-Moll, and Francesc Moreno-Noguer. Simplicit: Topology-aware generative model for clothed people. In CVPR, 2021. 3", + "[19] Matt Deitke, Ruoshi Liu, Matthew Wallingford, Huong Ngo, Oscar Michel, Aditya Kusupati, Alan" + ], + "bbox": [ + 522, + 90, + 906, + 901 + ], + "page_idx": 8 + }, + { + "type": "page_number", + "text": "26339", + "bbox": [ + 478, + 944, + 519, + 955 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "Fan, Christian Laforte, Vikram Voleti, Samir Yitzhak Gadre, Eli VanderBilt, Aniruddha Kembhavi, Carl Vondrick, Georgia Gkioxari, Kiana Ehsani, Ludwig Schmidt, and Ali Farhadi. Objverse-xl: A universe of $10\\mathrm{m} + 3\\mathrm{d}$ objects. arXiv preprint arXiv:2307.05663, 2023. 2, 5", + "[20] Yilun Du, Yinan Zhang, Hong-Xing Yu, Joshua B. Tenenbaum, and Jiajun Wu. Neural radiance flow for 4d view synthesis and video processing. In Proceedings of the IEEE/CVF International Conference on Computer Vision, 2021. 3", + "[21] Shivam Duggal and Deepak Pathak. Topologically-aware deformation fields for single-view 3d reconstruction. CVPR, 2022. 2", + "[22] Sai Kumar Dwivedi, Nikos Athanasiou, Muhammed Kocabas, and Michael J. Black. Learning to Regress Bodies from Images using Differentiable Semantic Rendering. In IEEE/CVF International Conference on Computer Vision (ICCV), pages 11230-11239, 2021. 2", + "[23] Jiemin Fang, Taoran Yi, Xinggang Wang, Lingxi Xie, Xiaopeng Zhang, Wenyu Liu, Matthias Nießner, and Qi Tian. Fast dynamic radiance fields with time-aware neural voxels. In SIGGRAPH Asia 2022 Conference Papers. ACM, 2022. 3", + "[24] Yao Feng, Vasileios Choutas, Timo Bolkart, Dimitrios Tzionas, and Michael J. Black. Collaborative regression of expressive bodies using moderation. In International Conference on 3D Vision (3DV), 2021.", + "[25] Yao Feng, Jinlong Yang, Marc Pollefeys, Michael J. Black, and Timo Bolkart. Capturing and animation of body and clothing from monocular video. In SIGGRAPH Asia 2022 Conference Papers, 2022. 1, 3", + "[26] Yao Feng, Weiyang Liu, Timo Bolkart, Jinlong Yang, Marc Pollefeys, and Michael J. Black. Learning disentangled avatars with hybrid 3d representations. arXiv, 2023. 1, 3", + "[27] Valentin Gabeur, Jean-Sebastien Franco, Xavier Martin, Cordelia Schmid, and Gregory Rogez. Moulding humans: Non-parametric 3d human shape estimation from single images, 2019. 2", + "[28] Chen Gao, Ayush Saraf, Johannes Kopf, and Jia-Bin Huang. Dynamic view synthesis from dynamic monocular video. In Proceedings of the IEEE International Conference on Computer Vision, 2021. 3", + "[29] Shanyan Guan, Huayu Deng, Yunbo Wang, and Xiaokang Yang. Neurofluid: Fluid dynamics grounding with particle-driven neural radiance fields, 2022. 3", + "[30] Antoine Guédon and Vincent Lepetit. Sugar: Surface-aligned gaussian splatting for efficient 3d mesh reconstruction and high-quality mesh rendering. arXiv preprint arXiv:2311.12775, 2023. 3" + ], + "bbox": [ + 101, + 90, + 483, + 900 + ], + "page_idx": 9 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[31] Yuwei Guo, Ceyuan Yang, Anyi Rao, Zhengyang Liang, Yaohui Wang, Yu Qiao, Maneesh Agrawala, Dahua Lin, and Bo Dai. Animatediff: Animate your personalized text-to-image diffusion models without specific tuning, 2023. 2", + "[32] Oshri Halimi, Fabian Prada, Tuur Stuyck, Donglai Xiang, Timur Bagautdinov, He Wen, Ron Kimmel, Takaaki Shiratori, Chenglei Wu, and Yaser Sheikh. Garment avatars: Realistic cloth driving using pattern registration, 2022. 2", + "[33] Tong He, Yuanlu Xu, Shunsuke Saito, Stefano Soatto, and Tony Tung. $\\mathsf{ARCH} + +$ : Animation-Ready Clothed Human Reconstruction Revisited. Proceedings of the IEEE International Conference on Computer Vision, pages 11026–11036, 2021. 1, 2", + "[34] Zhu Heming, Cao Yu, Jin Hang, Chen Weikai, Du Dong, Wang Zhangye, Cui Shuguang, and Han Xiaoguang. Deep fashion3d: A dataset and benchmark for 3d garment reconstruction from single images. In Computer Vision - ECCV 2020, pages 512-530. Springer International Publishing, 2020. 2", + "[35] Jonathan Ho, Ajay Jain, and Pieter Abbeel. Denoising diffusion probabilistic models. arXiv preprint arxiv:2006.11239, 2020. 2", + "[36] Yicong Hong, Kai Zhang, Jiaxiang Gu, Sai Bi, Yang Zhou, Difan Liu, Feng Liu, Kalyan Sunkavalli, Trung Bui, and Hao Tan. Lrm: Large reconstruction model for single image to 3d. arXiv preprint arXiv:2311.04400, 2023. 2", + "[37] Li Hu, Xin Gao, Peng Zhang, Ke Sun, Bang Zhang, and Liefeng Bo. Animate anyone: Consistent and controllable image-to-video synthesis for character animation. arXiv preprint arXiv:2311.17117, 2023. 6", + "[38] Liangxiao Hu, Hongwen Zhang, Yuxiang Zhang, Boyao Zhou, Boning Liu, Shengping Zhang, and Liqiang Nie. Gaussianavatar: Towards realistic human avatar modeling from a single video via animatable 3d gaussians. In IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), 2024. 3, 6", + "[39] Shoukang Hu and Ziwei Liu. Gauhuman: Articulated gaussian splatting from monocular human videos. arXiv preprint arXiv:, 2023. 4", + "[40] Shoukang Hu, Fangzhou Hong, Liang Pan, Haiyi Mei, Lei Yang, and Ziwei Liu. Sherf: Generalizable human nerf from a single image. arXiv preprint arXiv:2303.12791, 2023. 1, 2", + "[41] Yangyi Huang, Hongwei Yi, Weiyang Liu, Haofan Wang, Boxi Wu, Wenxiao Wang, Binbin Lin, Debing Zhang, and Deng Cai. One-shot implicit animatable avatars with model-based priors. In IEEE Conference on Computer Vision (ICCV), 2023. 2" + ], + "bbox": [ + 524, + 90, + 905, + 900 + ], + "page_idx": 9 + }, + { + "type": "page_number", + "text": "26340", + "bbox": [ + 478, + 944, + 519, + 955 + ], + "page_idx": 9 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[42] Zeng Huang, Yuanlu Xu, Christoph Lassner, Hao Li, and Tony Tung. ARCH: Animatable Reconstruction of Clothed Humans. Proceedings of the IEEE Computer Society Conference on Computer Vision and Pattern Recognition, pages 3090-3099, 2020. 1, 2", + "[43] Boyi Jiang, Juyong Zhang, Yang Hong, Jinhao Luo, Ligang Liu, and Hujun Bao. BCnet: Learning body and cloth shape from a single image. In European Conference on Computer Vision. Springer, 2020. 2", + "[44] Hanbyul Joo, Natalia Neverova, and Andrea Vedaldi. Exemplar Fine-Tuning for 3D Human Model Fitting Towards In-the-Wild 3D Human Pose Estimation. Proceedings - 2021 International Conference on 3D Vision, 3DV 2021, pages 42-52, 2021. 2", + "[45] Heewoo Jun and Alex Nichol. Shap-e: Generating conditional 3d implicit functions, 2023. 2", + "[46] Angjoo Kanazawa, Michael J Black, David W Jacobs, and Jitendra Malik. End-to-End Recovery of Human Shape and Pose. In Proceedings of the IEEE Computer Society Conference on Computer Vision and Pattern Recognition, pages 7122–7131, 2018. 2", + "[47] Bernhard Kerbl, Georgios Kopanas, Thomas Leimkuhler, and George Drettakis. 3d gaussian splatting for real-time radiance field rendering. ACM Transactions on Graphics, 42(4), 2023. 2, 15", + "[48] Muhammed Kocabas, Chun-Hao P. Huang, Otmar Hilliges, and Michael J. Black. PARE: Part attention regressor for 3D human body estimation. In Proc. International Conference on Computer Vision (ICCV), pages 11127-11137, 2021. 2", + "[49] Muhammed Kocabas, Chun-Hao P. Huang, Joachim Tesch, Lea Müller, Otmar Hilliges, and Michael J. Black. SPEC: Seeing people in the wild with an estimated camera. In Proc. International Conference on Computer Vision (ICCV), pages 11035-11045, 2021. 2", + "[50] Muhammed Kocabas, Rick Chang, James Gabriel, Oncel Tuzel, and Anurag Ranjan. Hugs: Human gaussian splats, 2023. 3", + "[51] Nikos Kolotouros, Georgios Pavlakos, Michael Black, and Kostas Daniilidis. Learning to reconstruct 3D human pose and shape via model-fitting in the loop. Proceedings of the IEEE International Conference on Computer Vision, 2019-Octob:2252-2261, 2019. 2", + "[52] Georgios Kopanas, Julien Philip, Thomas Leimkuhler, and George Drettakis. Point-based neural rendering with per-view optimization. Computer Graphics Forum (Proceedings of the Eurographics Symposium on Rendering), 40(4), 2021. 15", + "[53] Georgios Kopanas, Thomas Leimkuhler, Gilles Rainer, Clément Jambon, and George Drettakis. Neural point catauastics for novel-view synthesis of re" + ], + "bbox": [ + 99, + 90, + 485, + 900 + ], + "page_idx": 10 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "flections. ACM Transactions on Graphics, 41(6): Article-201, 2022. 15", + "[54] Jiahui Lei, Yufu Wang, Georgios Pavlakos, Lingjie Liu, and Kostas Daniilidis. Gart: Gaussian articulated template models, 2023. 3, 6", + "[55] Mengtian Li, Shengxiang Yao, Zhifeng Xie, and Keyu Chen. Gaussianbody:Clothed human reconstruction via 3d gaussian splatting,2024.3", + "[56] Xuanyi Li, Daquan Zhou, Chenxu Zhang, Shaodong Wei, Qibin Hou, and Ming-Ming Cheng. Sora generates videos with stunning geometrical consistency. arXiv preprint arXiv: 2402.17403, 2024. 2", + "[57] Zhengqi Li, Simon Niklaus, Noah Snavely, and Oliver Wang. Neural scene flow fields for space-time view synthesis of dynamic scenes. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), 2021. 3", + "[58] Zhengqi Li, Qianqian Wang, Forrester Cole, Richard Tucker, and Noah Snavely. Dynibar: Neural dynamic image-based rendering. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), 2023. 3", + "[59] Minghua Liu, Chao Xu, Haian Jin, Linghao Chen, Zexiang Xu, Hao Su, et al. One-2-3-45: Any single image to 3d mesh in 45 seconds without per-shape optimization. arXiv preprint arXiv:2306.16928, 2023. 2", + "[60] Ruoshi Liu, Rundi Wu, Basile Van Hoorick, Pavel Tokmakov, Sergey Zakharov, and Carl Vondrick. Zero-1-to-3: Zero-shot one image to 3d object, 2023. 2", + "[61] Yang Liu, Xiang Huang, Minghan Qin, Qinwei Lin, and Haoqian Wang. Animatable 3d gaussian: Fast and high-quality reconstruction of multiple human avatars. arXiv preprint arXiv:2311.16482, 2023. 3", + "[62] Matthew Loper, Naureen Mahmood, Javier Romero, Gerard Pons-Moll, and Michael J. Black. SMPL: A skinned multi-person linear model. ACM Trans. Graphics (Proc. SIGGRAPH Asia), 34(6):248:1-248:16, 2015. 2, 3, 15", + "[63] Jonathon Luiten, Georgios Kopanas, Bastian Leibe, and Deva Ramanan. Dynamic 3d gaussians: Tracking by persistent dynamic view synthesis. In 3DV, 2024. 3", + "[64] Luke Melas-Kyriazi, Iro Laina, Christian Rupprecht, and Andrea Vedaldi. Realfusion: 360deg reconstruction of any object from a single image. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 8446-8455, 2023. 2", + "[65] Ben Mildenhall, Pratul P. Srinivasan, Matthew Tancik, Jonathan T. Barron, Ravi Ramamoorthi, and Ren Ng. Nerf: Representing scenes as neural radiance fields for view synthesis. In ECCV, 2020. 2" + ], + "bbox": [ + 522, + 90, + 906, + 900 + ], + "page_idx": 10 + }, + { + "type": "page_number", + "text": "26341", + "bbox": [ + 478, + 944, + 517, + 955 + ], + "page_idx": 10 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[66] Gyeongsik Moon, Hongsuk Choi, and Kyoung Mu Lee. Accurate 3d hand pose estimation for whole-body 3d human mesh estimation. In Computer Vision and Pattern Recognition Workshop (CVPRW), 2022. 2", + "[67] Alex Nichol, Heewoo Jun, Prafulla Dhariwal, Pamela Mishkin, and Mark Chen. Point-e: A system for generating 3d point clouds from complex prompts, 2022. 2", + "[68] Keunhong Park, Utkarsh Sinha, Jonathan T. Barron, Sofien Bouaziz, Dan B Goldman, Steven M. Seitz, and Ricardo Martin-Brualla. Nerfies: Deformable neural radiance fields. ICCV, 2021. 3", + "[69] Keunhong Park, Utkarsh Sinha, Peter Hedman, Jonathan T. Barron, Sofien Bouaziz, Dan B Goldman, Ricardo Martin-Brualla, and Steven M. Seitz. Hypernerf: A higher-dimensional representation for topologically varying neural radiance fields. ACM Trans. Graph., 40(6), 2021. 3", + "[70] Chaitanya Patel, Zhouyingcheng Liao, and Gerard Pons-Moll. Tailornet: Predicting clothing in 3d as a function of human pose, shape and garment style. In IEEE Conference on Computer Vision and Pattern Recognition (CVPR). IEEE, 2020. 2", + "[71] Georgios Pavlakos, Vasileios Choutas, Nima Ghorbani, Timo Bolkart, Ahmed A. Osman, Dmitrios Tzionas, and Michael J. Black. Expressive body capture: 3D hands, face, and body from a single image. In Proceedings of the IEEE Computer Society Conference on Computer Vision and Pattern Recognition, pages 10967-10977, 2019. 1, 2, 3, 15", + "[72] Georgios Pavlakos, Vasileios Choutas, Nima Ghorbani, Timo Bolkart, Ahmed AA Osman, Dimitrios Tzionas, and Michael J Black. Expressive body capture: 3d hands, face, and body from a single image. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 10975-10985, 2019. 3", + "[73] Gerard Pons-Moll, Sergi Pujades, Sonny Hu, and Michael Black. Clothcap: Seamless 4d clothing capture and retargeting. ACM Transactions on Graphics, (Proc. SIGGRAPH), 36(4), 2017. Two first authors contributed equally. 2", + "[74] Ben Poole, Ajay Jain, Jonathan T. Barron, and Ben Mildenhall. Dreamfusion: Text-to-3d using 2d diffusion. arXiv, 2022. 2", + "[75] Albert Pumarola, Enric Corona, Gerard Pons-Moll, and Francesc Moreno-Noguer. D-NeRF: Neural Radiance Fields for Dynamic Scenes. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, 2020. 3", + "[76] Guocheng Qian, Jinjie Mai, Abdullah Hamdi, Jian Ren, Aliaksandr Siarohin, Bing Li, Hsin-Ying Lee," + ], + "bbox": [ + 101, + 90, + 485, + 902 + ], + "page_idx": 11 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "Ivan Skorokhodov, Peter Wonka, Sergey Tulyakov, and Bernard Ghanem. Magic123: One image to high-quality 3d object generation using both 2d and 3d diffusion priors. In The Twelfth International Conference on Learning Representations (ICLR), 2024. 2", + "[77] Zhiyin Qian, Shaofei Wang, Marko Mihajlovic, Andreas Geiger, and Siyu Tang. 3dgs- avatar: Animatable avatars via deformable 3d gaussian splatting. 2024. 3", + "[78] Jiawei Ren, Liang Pan, Jiaxiang Tang, Chi Zhang, Ang Cao, Gang Zeng, and Ziwei Liu. Dreamgaussian4d: Generative 4d gaussian splatting. arXiv preprint arXiv:2312.17142, 2023. 3, 5, 6", + "[79] Robin Rombach, Andreas Blattmann, Dominik Lorenz, Patrick Esser, and Björn Ommer. High-resolution image synthesis with latent diffusion models, 2021. 2", + "[80] Yu Rong, Takaaki Shiratori, and Hanbyul Joo. FrankMocap: A Monocular 3D Whole-Body Pose Estimation System via Regression and Integration. In Proceedings of the IEEE International Conference on Computer Vision, pages 1749–1759, 2021. 2", + "[81] Shunsuke Saito, Zeng Huang, Ryota Natsume, Shigeo Morishima, Angjoo Kanazawa, and Hao Li. Pifu: Pixel-aligned implicit function for high-resolution clothed human digitization. In The IEEE International Conference on Computer Vision (ICCV), 2019. 1, 2", + "[82] Shunsuke Saito, Tomas Simon, Jason Saragih, and Hanbyul Joo. Pifuhd: Multi-level pixel-aligned implicit function for high-resolution 3d human digitization. In CVPR, 2020. 1, 2", + "[83] Igor Santesteban, Miguel A. Otaduy, and Dan Casas. Learning-Based Animation of Clothing for Virtual Try-On. Computer Graphics Forum (Proc. Eurographics), 2019. 2", + "[84] Sara Fridovich-Keil and Giacomo Meanti, Frederik Rahbæk Warburg, Benjamin Recht, and Angjoo Kanazawa. K-planes: Explicit radiance fields in space, time, and appearance. In CVPR, 2023. 3", + "[85] Ruizhi Shao, Zerong Zheng, Hanzhang Tu, Boning Liu, Hongwen Zhang, and Yebin Liu. Tensor4d: Efficient neural 4d decomposition for high-fidelity dynamic reconstruction and rendering. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition, 2023. 3", + "[86] Zhijing Shao, Zhaolong Wang, Zhuang Li, Duotun Wang, Xiangru Lin, Yu Zhang, Mingming Fan, and Zeyu Wang. SplattingAvatar: Realistic Real-Time Human Avatars with Mesh-Embedded Gaussian Splatting. In Computer Vision and Pattern Recognition (CVPR), 2024. 4" + ], + "bbox": [ + 522, + 90, + 906, + 900 + ], + "page_idx": 11 + }, + { + "type": "page_number", + "text": "26342", + "bbox": [ + 478, + 944, + 519, + 955 + ], + "page_idx": 11 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[87] Ruoxi Shi, Hansheng Chen, Zhuoyang Zhang, Minghua Liu, Chao Xu, Xinyue Wei, Linghao Chen, Chong Zeng, and Hao Su. Zero123++: a single image to consistent multi-view diffusion base model, 2023. 2", + "[88] Stanislaw Szymanowicz, Christian Rupprecht, and Andrea Vedaldi. Splatter image: Ultra-fast single-view 3d reconstruction, 2023. 2", + "[89] Jiaxiang Tang, Jiawei Ren, Hang Zhou, Ziwei Liu, and Gang Zeng. Dreamgaussian: Generative gaussian splatting for efficient 3d content creation. arXiv preprint arXiv:2309.16653, 2023. 2, 3, 6, 8", + "[90] Junshu Tang, Tengfei Wang, Bo Zhang, Ting Zhang, Ran Yi, Lizhuang Ma, and Dong Chen. Make-it-3d: High-fidelity 3d creation from a single image with diffusion prior. In Proceedings of the IEEE/CVF International Conference on Computer Vision (ICCV), pages 22819–22829, 2023. 2", + "[91] Jiaxiang Tang, Zhaoxi Chen, Xiaokang Chen, Tengfei Wang, Gang Zeng, and Ziwei Liu. Lgm: Large multi-view gaussian model for high-resolution 3d content creation. arXiv preprint arXiv:2402.05054, 2024. 2, 3, 4, 6, 8", + "[92] Edgar Tretschk, Ayush Tewari, Vladislav Golyanik, Michael Zollhöfer, Christoph Lassner, and Christian Theobalt. Non-rigid neural radiance fields: Reconstruction and novel view synthesis of a dynamic scene from monocular video. In IEEE International Conference on Computer Vision (ICCV). IEEE, 2021. 3", + "[93] Alex Trevithick and Bo Yang. Grf: Learning a general radiance field for 3d scene representation and rendering. In arXiv:2010.04595, 2020. 2", + "[94] Haithem Turki, Jason Y. Zhang, Francesco Ferroni, and Deva Ramanan. Suits: Scalable urban dynamic scenes, 2023. 3", + "[95] Raquel Vidaurre, Igor Santesteban, Elena Garces, and Dan Casas. Fully Convolutional Graph Neural Networks for Parametric Virtual Try-On. Computer Graphics Forum (Proc. SCA), 2020. 2", + "[96] Wenbo Wang, Hsuan-I Ho, Chen Guo, Boxiang Rong, Artur Grigorev, Jie Song, Juan Jose Zarate, and Otmar Hilliges. 4d-dress: A 4d dataset of real-world human clothing with semantic annotations. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition (CVPR), 2024. 6, 16", + "[97] Guanjun Wu, Taoran Yi, Jiemin Fang, Lingxi Xie, Xiaopeng Zhang, Wei Wei, Wenyu Liu, Qi Tian, and Wang Xinggang. 4d gaussian splatting for real-time dynamic scene rendering. arXiv preprint arXiv:2310.08528, 2023. 3", + "[98] Wenqi Xian, Jia-Bin Huang, Johannes Kopf, and Changil Kim. Space-time neural irradiance fields for free-viewpoint video. In Proceedings of the" + ], + "bbox": [ + 99, + 90, + 485, + 900 + ], + "page_idx": 12 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 9421-9431, 2021. 3", + "[99] Donglai Xiang, Fabian Prada, Chenglei Wu, and Jessica Hodgins. Monoclothcap: Towards temporally coherent clothing capture from monocular rgb video. In Proceedings of International Conference on 3D Vision (3DV '20), pages 322 - 332, 2020. 2", + "[100] Donglai Xiang, Fabian Prada, Timur Bagautdinov, Weipeng Xu, Yuan Dong, He Wen, Jessica Hodgins, and Chenglei Wu. Modeling clothing as a separate layer for an animatable human avatar. ACM Transactions on Graphics, 40(6):1-15, 2021. 2", + "[101] Enze Xie, Wenhai Wang, Zhiding Yu, Anima Anandkumar, Jose M. Alvarez, and Ping Luo. Segformer: Simple and efficient design for semantic segmentation with transformers. CoRR, abs/2105.15203, 2021. 4", + "[102] Tianyi Xie, Zeshun Zong, Yuxing Qiu, Xuan Li, Yutao Feng, Yin Yang, and Chenfanfu Jiang. Physgaussian: Physics-integrated 3d gaussians for generative dynamics. arXiv preprint arXiv:2311.12198, 2023. 5", + "[103] Yuliang Xiu, Jinlong Yang, Dimitrios Tzionas, and Michael J. Black. ICON: Implicit Clothed humans Obtained from Normals. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 13296-13306, 2022. 1, 2", + "[104] Yuliang Xiu, Jinlong Yang, Xu Cao, Dimitrios Tzionas, and Michael J. Black. ECON: Explicit Clothed humans Optimized via Normal integration. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), 2023. 1, 2", + "[105] Qiangeng Xu, Weiyue Wang, Duygu Ceylan, Radomir Mech, and Ulrich Neumann. Disn: Deep implicit surface network for high-quality single-view 3d reconstruction. In Advances in Neural Information Processing Systems 32, pages 492-502. 2019. 2", + "[106] Zhongcong Xu, Jianfeng Zhang, Jun Hao Liew, Hanshu Yan, Jia-Wei Liu, Chenxu Zhang, Jiashi Feng, and Mike Zheng Shou. Magicanimate: Temporally consistent human image animation using diffusion model. 2023. 6", + "[107] Xueting Yang, Yihao Luo, Yuliang Xiu, Wei Wang, Hao Xu, and Zhaoxin Fan. D-IF: Uncertainty-aware Human Digitization via Implicit Distribution Field. In Proceedings of the IEEE/CVF international conference on computer vision, 2023. 2", + "[108] Yifan Yang, Dong Liu, Shuhai Zhang, Zeshuai Deng, Zixiong Huang, and Mingkui Tan. Hilo: Detailed and robust 3d clothed human reconstruction with high- and low-frequency information of parametric models. In Proceedings of the IEEE/CVF Conference" + ], + "bbox": [ + 516, + 90, + 906, + 900 + ], + "page_idx": 12 + }, + { + "type": "page_number", + "text": "26343", + "bbox": [ + 478, + 944, + 517, + 955 + ], + "page_idx": 12 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "on Computer Vision and Pattern Recognition, pages 10671-10681, 2024. 2", + "[109] Zhitao Yang, Zhongang Cai, Haiyi Mei, Shuai Liu, Zhaoxi Chen, Weiye Xiao, Yukun Wei, Zhongfei Qing, Chen Wei, Bo Dai, Wayne Wu, Chen Qian, Dahua Lin, Ziwei Liu, and Lei Yang. Synbody: Synthetic dataset with layered human models for 3d human perception and modeling. In Proceedings of the IEEE/CVF International Conference on Computer Vision (ICCV), pages 20282-20292, 2023. 6", + "[110] Ziyi Yang, Xinyu Gao, Wen Zhou, Shaohui Jiao, Yuqing Zhang, and Xiaogang Jin. Deformable 3d gaussians for high-fidelity monocular dynamic scene reconstruction. arXiv preprint arXiv:2309.13101, 2023. 3", + "[111] Mingqiao Ye, Martin Danelljan, Fisher Yu, and Lei Ke. Gaussian grouping: Segment and edit anything in 3d scenes. arXiv preprint arXiv:2312.00732, 2023. 4, 15", + "[112] Xianggang Yu, Mutian Xu, Yidan Zhang, Haolin Liu, Chongjie Ye, Yushuang Wu, Zizheng Yan, Tianyou Liang, Guanying Chen, Shuguang Cui, and Xiaoguang Han. Mvimgnet: A large-scale dataset of multi-view images. In CVPR, 2023. 2", + "[113] Zhengming Yu, Wei Cheng, Xian Liu, Wayne Wu, and Kwan-Yee Lin. Monohuman: Animatable human neural field from monocular video. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 16943-16953, 2023. 3, 6", + "[114] Wentao Yuan, Zhaoyang Lv, Tanner Schmidt, and Steven Lovegrove. Star: Self-supervised tracking and reconstruction of rigid objects in motion with neural rendering. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 13144-13152, 2021. 3", + "[115] Ilya Zakharkin, Kirill Mazur, Artur Grigorev, and Victor Lempitsky. Point-based modeling of human clothing. In Proceedings of the IEEE/CVF International Conference on Computer Vision (ICCV), pages 14718-14727, 2021. 2", + "[116] Yan Zeng, Guoqiang Wei, Jiani Zheng, Jiaxin Zou, Yang Wei, Yuchen Zhang, and Hang Li. Make pixels dance: High-dynamic video generation. arXiv:2311.10982, 2023. 2", + "[117] Zheng Zerong, Yu Tao, Liu Yebin, and Dai Qionghai. Pamir: Parametric model-conditioned implicit representation for image-based human reconstruction, 2021. 1, 2", + "[118] Hongwen Zhang, Yating Tian, Yuxiang Zhang, Mengcheng Li, Liang An, Zhenan Sun, and Yebin Liu. Pymaf-x: Towards well-aligned full-body model regression from monocular images. IEEE Transac-" + ], + "bbox": [ + 91, + 90, + 483, + 900 + ], + "page_idx": 13 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "tions on Pattern Analysis and Machine Intelligence, 2023. 2", + "[119] Hao Zhu, Xinxin Zuo, Sen Wang, Xun Cao, and Ruigang Yang. Detailed human shape estimation from a single image by hierarchical mesh deformation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 4491-4500, 2019. 2", + "[120] Shenhao Zhu, Junming Leo Chen, Zuozhuo Dai, Yinghui Xu, Xun Cao, Yao Yao, Hao Zhu, and Siyu Zhu. Champ: Controllable and consistent human image animation with 3d parametric guidance, 2024. 6" + ], + "bbox": [ + 516, + 90, + 906, + 287 + ], + "page_idx": 13 + }, + { + "type": "page_number", + "text": "26344", + "bbox": [ + 478, + 945, + 517, + 955 + ], + "page_idx": 13 + } +] \ No newline at end of file diff --git a/2025/Disco4D_ Disentangled 4D Human Generation and Animation from a Single Image/5f039bdf-5dd5-496e-b90f-b0a10473c8cd_model.json b/2025/Disco4D_ Disentangled 4D Human Generation and Animation from a Single Image/5f039bdf-5dd5-496e-b90f-b0a10473c8cd_model.json new file mode 100644 index 0000000000000000000000000000000000000000..1e1576037ecdb4c3bf26a5b30fcef9cc28cf1dc0 --- /dev/null +++ b/2025/Disco4D_ Disentangled 4D Human Generation and Animation from a Single Image/5f039bdf-5dd5-496e-b90f-b0a10473c8cd_model.json @@ -0,0 +1,2846 @@ +[ + [ + { + "type": "header", + "bbox": [ + 0.107, + 0.003, + 0.182, + 0.043 + ], + "angle": 0, + "content": "CVF" + }, + { + "type": "header", + "bbox": [ + 0.237, + 0.001, + 0.812, + 0.047 + ], + "angle": 0, + "content": "This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore." + }, + { + "type": "title", + "bbox": [ + 0.092, + 0.131, + 0.908, + 0.155 + ], + "angle": 0, + "content": "Disco4D: Disentangled 4D Human Generation and Animation from a Single Image" + }, + { + "type": "text", + "bbox": [ + 0.238, + 0.179, + 0.754, + 0.217 + ], + "angle": 0, + "content": "Hui En Pang\\(^{1}\\), Shuai Liu\\(^{3}\\), Zhongang Cai\\(^{1,2,3}\\), Lei Yang\\(^{2,3}\\), Tianwei Zhang\\(^{1}\\), Ziwei Liu\\(^{1}\\)" + }, + { + "type": "text", + "bbox": [ + 0.117, + 0.217, + 0.875, + 0.235 + ], + "angle": 0, + "content": "\\(^{1}\\)S-Lab, Nanyang Technological University \\(^{2}\\)SenseTime Research \\(^{3}\\)Shanghai AI Laboratory" + }, + { + "type": "image", + "bbox": [ + 0.104, + 0.253, + 0.895, + 0.394 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.09, + 0.396, + 0.908, + 0.425 + ], + "angle": 0, + "content": "Figure 1. Disco4D is a novel Gaussian Splatting framework for 4D disentangled human generation from a single image. Clothing and assets are modeled as Gaussians on top of the SMPL-X body, enabling dynamic animation and flexible editing." + }, + { + "type": "title", + "bbox": [ + 0.249, + 0.431, + 0.327, + 0.446 + ], + "angle": 0, + "content": "Abstract" + }, + { + "type": "text", + "bbox": [ + 0.089, + 0.464, + 0.486, + 0.722 + ], + "angle": 0, + "content": "We present Disco4D, a novel Gaussian Splatting framework for 4D human generation and animation from a single image. Different from existing methods, Disco4D distinctively disentangles clothings (with Gaussian models) from the human body (with SMPL-X model), significantly enhancing the generation details and flexibility. Specifically, 1) Disco4D learns to efficiently fit the clothing Gaussians over the SMPL-X Gaussians. 2) Next, Disco4D adopts diffusion models to enhance the 3D generation process, e.g., modeling occluded parts not visible in the input image. 3) Finally, Disco4D learns an identity encoding for each clothing Gaussian to facilitate the separation and extraction of clothing assets. Furthermore, Disco4D naturally supports 4D human animation with vivid dynamics. Extensive experiments demonstrate the superiority of Disco4D on 4D human generation and animation tasks. Our code is available at https://github.com/disco-4d/Disco4D." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.754, + 0.222, + 0.769 + ], + "angle": 0, + "content": "1. Introduction" + }, + { + "type": "text", + "bbox": [ + 0.089, + 0.78, + 0.486, + 0.902 + ], + "angle": 0, + "content": "The development of high-fidelity 3D digital humans is increasingly important across a variety of augmented and virtual reality applications. To streamline the creation of these digital avatars from easily accessible in-the-wild images, a multitude of research efforts have been made on reconstructing 3D clothed human models from a single image [2, 33, 40-42, 81, 82, 103, 104, 104, 117]. These works predominantly focus on the simultaneous reconstruction of" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.431, + 0.909, + 0.629 + ], + "angle": 0, + "content": "the human body and clothing. Unfortunately, these works have inherent limitations, and integrating them into applications that require virtual try-on or avatar customization poses significant challenges. This is primarily because the models are rendered as single-layer, non-animatable meshes where distinct attributes (e.g., hair, clothing, accessories) are merged into one continuous surface, with underlying layers completely obscured and self-contact areas inseparably connected. Such limitation complicates the re-animation and dynamic customization tasks. Existing works that perform layered reconstruction [25, 26] rely on self-rotating video inputs with extensive frames and viewpoints and involve substantial processing times." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.63, + 0.911, + 0.902 + ], + "angle": 0, + "content": "To address these issues, we propose Disco4D, a novel 4D clothed human reconstruction method that distinctly separates the human body from clothing elements from a single image. It supports human animation as the 4th dimension, which cannot be realized by prior static 3D reconstruction works [2, 33, 42, 81, 82, 103, 104, 117]. To achieve this, it employs the SMPL-X [71] parametric model to represent the human body, capitalizing on its efficacy in capturing body structure and kinematics. Conversely, clothing, along with dynamic and variable elements such as hair and accessories, is represented using Gaussian models, which are able to model the large variability in clothing. By binding Gaussians to a SMPL-X model and fixing it during the training phase, Disco4D ensures the integrity of the body while focusing the learning process on the appearance aspects. To model occluded portions not visible in the input image, diffusion models are used to enhance the 3D generation process. Moreover, Disco4D includes an identity grouping mecha" + }, + { + "type": "page_number", + "bbox": [ + 0.479, + 0.945, + 0.518, + 0.957 + ], + "angle": 0, + "content": "26331" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.09, + 0.092, + 0.483, + 0.121 + ], + "angle": 0, + "content": "nism for the Gaussians, which is instrumental in maintaining the separability and individuality of each clothing asset." + }, + { + "type": "text", + "bbox": [ + 0.093, + 0.123, + 0.486, + 0.499 + ], + "angle": 0, + "content": "The independent reconstruction of clothing and body offers several advantages. (1) Enhanced reconstruction fidelity. The SMPL-X body serves as a stable anchor for the clothing to conform to. By isolating the focus to learn clothing Gaussians, we achieve a more refined geometry and intricate detailing in the clothed model. (2) Fine-grained categorization and extraction of clothing items. Disco4D is able to separate clothing Gaussians into their respective categories, which is crucial for the recovery and utilization of individual clothing assets. (3) Extensive editing capabilities. Disco4D supports different editing functions, including the removal of specific items, inpainting (altering color or material), and other modifications. Such rich editing options allow for precise adjustments to individual assets without inadvertently affecting adjacent elements. This level of control is particularly beneficial in applications requiring detailed customization, such as virtual fashion design and digital content creation. (4) Improved animation capabilities. The body Gaussians adhere to the deformations dictated by the SMPL-X model, while clothing Gaussians conform to the underlying body movements but also exhibit behaviors true to their material characteristics. The disentangled deformation allows for nuanced adjustments to clothing behavior in response to complex body movements, thereby elevating the quality of clothed human animation." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.512, + 0.242, + 0.529 + ], + "angle": 0, + "content": "2. Related Works" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.538, + 0.486, + 0.568 + ], + "angle": 0, + "content": "Table 1 summarizes the relevant 3D/4D generation methods. We describe their details below." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.577, + 0.242, + 0.592 + ], + "angle": 0, + "content": "2.1.3D Generation" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.599, + 0.486, + 0.902 + ], + "angle": 0, + "content": "Single-image 3D Generation. Single-image reconstruction leverages advanced methods [45, 67] to generate 3D assets in the form of 3D point clouds or NeRF [65] from one image. While earlier efforts using auto-encoders focused on synthetic objects [12, 14, 21, 88, 93, 105], newer approaches treat the task as conditional generation, employing diffusion models [35] for 3D generation from both image and text [19, 35, 60, 64, 74, 76, 79, 90]. One-2-3-45 [59] uses 2D diffusion models [60, 87] to generate multiview images for reconstruction, while LRM [36] adopts transformer-based architecture to scale up the task on large datasets [19, 112]. Gaussian-based methods [47], particularly DreamGaussian [89] and LGM [91], offer efficient, high-resolution 3D model generation from text or images. Recently, video diffusion models have attracted significant attention due to their remarkable ability to generate intricate scenes and complex dynamics with great spatio-temporal consistency [4, 7-9, 31, 56, 116]. They are employed to generate consistent multi-view images, and then reconstruct underlying 3D assets with high quality [15]." + }, + { + "type": "table_caption", + "bbox": [ + 0.534, + 0.09, + 0.886, + 0.105 + ], + "angle": 0, + "content": "Table 1. 3D/4D generation methods from a single image." + }, + { + "type": "table", + "bbox": [ + 0.566, + 0.11, + 0.855, + 0.232 + ], + "angle": 0, + "content": "
MethodTypeLayeredAnimatable
LGM [91]GeneralXX
PiFU [81]Human-centricXX
DreamFusion [74]GeneralXX
DreamGaussian [89]GeneralXX
PiFU [81]Human-centricXX
D-IF [107]Human-centricXX
HiLo [108]Human-centricXX
ECON [104]Human-centricXX
SHERF [40]Human-centricX✓
Disco4DHuman-centric✓✓
" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.247, + 0.909, + 0.716 + ], + "angle": 0, + "content": "Single-image human-centric 3D Generation. Significant research efforts have been made for 3D human reconstruction, which can be classified into the following categories. (1) Explicit-shape-based methods rely on Human Mesh Recovery (HMR) using parametric models like SMPL [62] and SMPL-X [71] to generate 3D body meshes [16, 17, 22, 24, 44, 46, 48, 49, 51, 66, 80, 118]. To account for 3D garments, several approaches incorporate offsets [99, 119] or templates, utilize deformable garment templates [6, 43], or employ non-parametric forms for clothed figures [27, 104, 115]. Despite their advancements, they face limitations in handling complex outfit variations and loose clothing due to inherent topological constraints. (2) Implicit-function-based methods utilize implicit representations like occupancy or distance fields for modeling clothed humans with complex geometries, such as loose garments. Techniques range from end-to-end regression of free-form implicit surfaces [2, 81, 82] to use of geometric priors [33, 42, 103, 104, 117] and implicit shape completion [104]. Notable works such as PIFu [81], ARCH(++) [33, 42], and PaMIR [117] can extract textured models from images, but struggle with depth ambiguities and texture inconsistencies. (3) NeRF-based methods incorporate model-based priors (i.e., SMPL-X) for accurate human reconstruction. Efforts like SHERF [40] and ELICIT [41] improve the reconstruction coherence by addressing 2D observation incompleteness leveraging appearance priors. Most of these 3D clothed human reconstruction and animation works [2, 33, 42, 81, 82, 103, 104, 117] require training on human-specific datasets, which brings another limitation on the availability of such datasets." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.72, + 0.911, + 0.902 + ], + "angle": 0, + "content": "3D Clothing Modeling. Reconstructing clothing from images and videos as a separate layer over the human body poses significant challenges due to the diversity of clothing topologies. Previous efforts relied on either template meshes or implicit surface models, and required extensive, high-quality 3D data from simulations [5, 70, 83, 95] or tailored template meshes [13, 32, 73, 100]. New methods were developed [34, 43] for multi-clothing models and versatile template meshes, respectively, facilitating diverse clothing topology encoding. However, these techniques typically fall short in capturing the clothing texture and appearance. The reliance on predefined clothing style templates further con" + }, + { + "type": "page_number", + "bbox": [ + 0.479, + 0.945, + 0.521, + 0.958 + ], + "angle": 0, + "content": "26332" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.09, + 0.092, + 0.486, + 0.228 + ], + "angle": 0, + "content": "strains their ability to handle real-world clothing variations. Corona et al. [18] addressed these shortcomings by representing clothing layers with deep unsigned distance functions and an auto-decoder for style and cut differentiation, though this often produces overly-smooth reconstructions [18]. On the other hand, SCARF [25] and DELTA [26] significantly enhance the visual fidelity by applying NeRF to clothing layers, but require self-rotating video inputs and considerable processing times." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.247, + 0.236, + 0.261 + ], + "angle": 0, + "content": "2.2. 4D Animation" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.272, + 0.485, + 0.499 + ], + "angle": 0, + "content": "4D Animation. This task aims at capturing dynamic 3D scenes over time. Two primary approaches have emerged: modeling 4D scenes by adding time dimension \\( t \\) or latent codes to spatial coordinates [28, 57, 98]; combining deformation fields with static 3D scenes [20, 58, 68, 69, 75, 92, 114]. Recent efforts in explicit or hybrid representations, like planar decomposition [11, 84, 85], hash representations [94], and other innovative methods [1, 23, 29], have improved reconstruction speed and quality. Gaussian Splatting, especially, stands out for balancing efficiency with quality, with dynamic 3D Gaussians [63] and 4D Gaussian Splatting [97, 110] techniques introducing time-dependent deformations to enhance reconstructions. Notably, DreamGaussian4D [78] stands out by minimizing the optimization time while achieving high-quality 4D reconstructions." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.502, + 0.484, + 0.609 + ], + "angle": 0, + "content": "Human-centric 4D Animation. Recent works leverage Gaussian-based methods [38, 50, 54, 55, 61, 77, 113] for 4D human reconstruction, requiring extensive frame sequences (50-100 frames) and/or multiple viewpoints. Currently there has not been any work on 4D layered human generation and animation from a single image or a video with few images, which will be achieved in this paper." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.631, + 0.226, + 0.649 + ], + "angle": 0, + "content": "3. Methodology" + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.659, + 0.231, + 0.674 + ], + "angle": 0, + "content": "3.1. Preliminaries" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.685, + 0.484, + 0.793 + ], + "angle": 0, + "content": "3D Gaussian Splatting employs explicit 3D Gaussian points as its primary rendering entities. A 3D Gaussian point is defined as a function \\( G(x) = e^{-\\frac{1}{2}(x - \\mu)^T\\Sigma^{-1}(x - \\mu)} \\), where \\( \\mu \\) and \\( \\Sigma \\) are the spatial mean and covariance matrix, respectively. Each Gaussian is also associated with its own rotation \\( r \\), scaling \\( s \\), opacity \\( \\alpha \\), a view-dependent color \\( c \\) represented by spherical harmonic coefficients \\( f \\)." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.796, + 0.484, + 0.902 + ], + "angle": 0, + "content": "SMPL-X parameterization [71] is an extension of the SMPL body model [62] with face and hand, designed to capture a more accurate representation of intricate body movements. SMPL-X is defined as a function \\( M(\\beta, \\theta, \\psi) : \\mathbb{R}^{|\\beta| \\times |\\theta| \\times |\\psi|} \\to \\mathbb{R}^{3N} \\), parametrized by the pose \\( \\theta \\in \\mathbb{R}^{3J} \\) (where \\( J \\) denotes the number of full body joints), body shape \\( \\beta \\in \\mathbb{R}^{|\\beta|} \\) and facial expression \\( \\psi \\in \\mathbb{R}^{|\\psi|} \\)." + }, + { + "type": "title", + "bbox": [ + 0.514, + 0.091, + 0.625, + 0.105 + ], + "angle": 0, + "content": "3.2. Overview" + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.113, + 0.908, + 0.204 + ], + "angle": 0, + "content": "Given a single image, Disco4D generates animatable 3D clothed human avatars in a bottom-up manner, facilitating natural separability. Our generated 3D clothed avatars, denoted as \\( S_{human} \\), are represented as the concatenation of \\( S_{body} \\) and \\( S_{cloth} \\). Inspired by prior works [89, 91], \\( S \\) capitalizes on Gaussian representations:" + }, + { + "type": "equation", + "bbox": [ + 0.633, + 0.217, + 0.907, + 0.234 + ], + "angle": 0, + "content": "\\[\nS = G (\\mu , r, s, \\alpha , c, e), \\tag {1}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.246, + 0.907, + 0.321 + ], + "angle": 0, + "content": "where \\(\\mu, r, s, \\alpha, c\\) and \\(e\\) denote positions, rotation, scaling, opacity, spherical harmonics coefficients and identity encoding, respectively. Different from traditional Gaussian representations, we add identity encoding \\(e\\) to associate each Gaussian with its clothing category." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.322, + 0.909, + 0.488 + ], + "angle": 0, + "content": "Figure 2 depicts our framework. We start by generating colored SMPL-X Gaussians representing the body beneath clothing (Sec. 3.3). We obtain a visual hull for canonicalization and refine Gaussian predictions to align and envelop the SMPL-X mesh (Sec. 3.4). Next, we iteratively optimize canonical clothing Gaussians external to the SMPL-X mesh (Sec. 3.5). Lastly, we showcase the animation and editing of generated clothed avatars (Sec. 3.6). Notably, we leverage diffusion models to refine textures during 3D generation (Sec. 3.5) and extrapolate unseen views during 4D animation (Sec. 3.6)." + }, + { + "type": "title", + "bbox": [ + 0.514, + 0.498, + 0.702, + 0.512 + ], + "angle": 0, + "content": "3.3. SMPL-X Gaussians" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.52, + 0.908, + 0.595 + ], + "angle": 0, + "content": "Given an image, we first estimate coarse SMPL-X parameters with an off-the-shelf model [10], and then refine coarse predictions by fitting on 2D keypoints and clothing segmentation masks [72], obtaining pixel-aligned SMPL-X parameters \\((\\beta, \\theta, \\psi)\\)." + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.597, + 0.909, + 0.839 + ], + "angle": 0, + "content": "Mesh Binding. To convert the SMPL-X [71] mesh \\( M(\\beta, \\theta, \\psi) \\) into Gaussians \\( S_{body} \\) for rendering, flat 3D Gaussians are bound to each mesh triangle, similar to SuGaR [30]. Gaussian means \\( \\mu_{body} \\) are computed using predefined barycentric coordinates, while Gaussian rotations \\( r_{body} \\) derive from surface normals. The initial scaling \\( s_{body} \\) ensures dense mesh coverage, with the last axis set to 0.1 for a uniformly thin surface. For color representation beneath clothing, opacity \\( \\alpha_{body} \\) is set to 1.0, with spherical harmonics \\( c_{body} \\) optimized for each Gaussian. Visible skin color is supervised, while occluded skin color aligns with visible regions. A fixed label \\( e_{body} \\) is assigned for rendering, remaining unchanged during training. When optimizing clothing Gaussians \\( S_{cloth} \\), SMPL-X Gaussians \\( S_{body} \\) parameters stay fixed, preserving the body structure while allowing flexible learning for clothing." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.848, + 0.822, + 0.865 + ], + "angle": 0, + "content": "3.4. Initialization of Clothing Gaussians" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.871, + 0.908, + 0.901 + ], + "angle": 0, + "content": "Cloth styles are diverse, making proper initialization crucial for effective clothing modeling. In synchronization with" + }, + { + "type": "page_number", + "bbox": [ + 0.479, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "26333" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.118, + 0.086, + 0.882, + 0.294 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.089, + 0.296, + 0.908, + 0.395 + ], + "angle": 0, + "content": "Figure 2. Framework Overview of Disco4D. (a) 3D Generation utilizes a single image to obtain disentangled body and clothing Gaussians. Body, face and hand poses are refined to be pixel-aligned. For faster initialization, clothing Gaussians and visual hull are obtained with Gaussian Reconstruction Models. These clothing Gaussians are embedded to SMPL-X mesh and adopt the local coordinate system of the triangle. Subsequently, the iterative optimization process (pruning, identity encoding and densifying) separates the body and garments. The learned identity encodings guide the densification of the clothing Gaussians. (b) 4D Animation is achieved by either direct driving of SMPL-X poses or leveraging video to learn extra clothing deformation (refer to Figure 3 for more details). Various (c) 3D/4D Editing operations can be performed with our disentangled representation." + }, + { + "type": "text", + "bbox": [ + 0.089, + 0.401, + 0.486, + 0.597 + ], + "angle": 0, + "content": "estimating SMPL-X, we first employ the Video Diffusion Model [7] to estimate multi-view images. Subsequently, we leverage Gaussian Reconstruction Models [91] to obtain initial 3D Gaussians and their corresponding visual hull. Yet, the reconstructed 3D outputs often suffer from geometric inaccuracies, such as incorrect poses due to pose ambiguity or missing limbs. To address this, we refine the coarse visual hull to ensure it accurately aligns with and overlays the SMPL-X mesh and encapsulates a good geometry for the clothed figure. With SMPL-X aligned visual hull, we derive the refined Gaussians by adopting properties from their nearest neighbors. The refined visual hull and Gaussians are then canonicalized for the optimization phase." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.598, + 0.486, + 0.84 + ], + "angle": 0, + "content": "Mesh embedding. Each 3D clothing Gaussian is embedded on a triangle of the canonical mesh, defining its position in both canonical and posed spaces. The mean vertex position \\(\\mathbf{O}\\) serves as the origin of the local coordinate system, with the Gaussian positioned by an offset vector \\(\\mathbf{v} = \\sigma \\mathbf{i} + \\beta \\mathbf{j} + \\gamma \\mathbf{k}\\), where \\(\\sigma\\), \\(\\beta\\), and \\(\\gamma\\) are the components of the displacement vector along the tangent \\(\\mathbf{i}\\), bitangent \\(\\mathbf{j}\\), and normal \\(\\mathbf{k}\\). Unlike SplattingAvatar [86], which displaces Gaussians along the normal, our approach allows embedding to the most suitable triangle rather than the nearest one. For example, hair Gaussians are tagged to head faces instead of the nearest face for reposing [39, 86] (Figure 9 in Appendix). In animation, the Gaussian rotates with its embedded triangle face \\((\\delta r)\\), while scaling \\((\\delta s)\\) is adjusted dynamically based on changes in edge lengths. During optimization, Gaussian and embedding parameters \\((\\mathbf{O}, \\mathbf{v}, \\delta r, \\text{and } \\delta s)\\) are jointly updated." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.849, + 0.414, + 0.866 + ], + "angle": 0, + "content": "3.5. Optimization of Separable Gaussians" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.87, + 0.485, + 0.903 + ], + "angle": 0, + "content": "With the SMPL-X Gaussian and initialized clothing Gaussian, we aim to optimize canonical clothing Gaussians \\( S_{cloth} \\)" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.401, + 0.908, + 0.583 + ], + "angle": 0, + "content": "outside the SMPL-X mesh. This involves three steps: 1) we use Signed Distance Function (SDF) loss and pruning to discourage and remove Gaussians that reside within the body; 2) we introduce identity encoding \\( e \\) to attach a clothing label for each clothing Gaussian, by lifting multi-view 2D segmentations of the target object onto the 3D Gaussians; and 3) guided by \\( e_{body} \\) and \\( e_{cloth} \\), we selectively densify only the relevant clothing points while ignoring body points. Once the disentangled clothing is obtained, we use SDS loss to in-paint high-resolution texture from the reference image to individual clothing Gaussians, thereby enriching the details of unseen regions." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.582, + 0.909, + 0.734 + ], + "angle": 0, + "content": "SDF Loss and Pruning. In reality, the clothing is always external to the body. During refinement, we ensure that the clothing Gaussians are positioned externally to the SMPL-X mesh by applying the SDF loss and a pruning strategy. Specifically, the SDF loss \\(\\mathcal{L}_{sdf}\\) penalizes any new densified Gaussians that intrude into the space of the SMPL-X mesh, ensuring that the clothing Gaussians consistently remain outside the body's surface. Pruning is applied at fixed intervals to reinforce this separation, and systematically remove any Gaussians located within the SDF of the SMPL-X mesh." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.733, + 0.909, + 0.855 + ], + "angle": 0, + "content": "Identity encoding. To associate each Gaussian to its clothing category, we introduce Identity Encoding \\((e)\\), a learnable and compact vector of length 15, representing clothing categories from SegFormer [101] segmentation masks1. During training, the encodings are rendered into 2D segmentation masks in a differentiable manner following [111]. For classification, we apply a softmax to the rendered features \\(E_{\\mathrm{id}}\\) and use cross-entropy loss \\(\\mathcal{L}_{2d}\\) for \\((K + 1)\\)-category classifica" + }, + { + "type": "page_footnote", + "bbox": [ + 0.513, + 0.863, + 0.908, + 0.903 + ], + "angle": 0, + "content": "1Categories: 0: \"Background\", 1: \"Hat\", 2: \"Hair\", 3: \"Sunglasses\", 4: \"Upper-clothes\", 5: \"Skirt\", 6: \"Pants\", 7: \"Dress\", 8: \"Belt\", 9: \"Left-shoe\", 10: \"Right-shoe\", 11: \"Face\", 12: \"Skin\", 13: \"Bag\", 14: \"Scarf\"" + }, + { + "type": "page_number", + "bbox": [ + 0.479, + 0.945, + 0.521, + 0.958 + ], + "angle": 0, + "content": "26334" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.09, + 0.092, + 0.485, + 0.273 + ], + "angle": 0, + "content": "tion. An unsupervised 3D regularization loss \\(\\mathcal{L}_{3d}\\) promotes spatial consistency among the top \\(k\\)-nearest 3D Gaussians' Identity Encodings. Consequently, the overall identity loss is \\(\\mathcal{L}_{id} = \\mathcal{L}_{2d} + \\mathcal{L}_{3d}\\). Refer to Appendix 7.2 for more details. Densification of clothing Gaussians. To learn clothing more efficiently, we perform sampling for categorical Gaussians that belong to the same clothing category and embedding. We find the \\(k\\)-nearest Gaussian points for the resampled points and inherit their Gaussian properties (scaling, rotation, opacity, SH properties). By selectively densifying clothing Gaussians, we only add necessary Gaussians while ignoring body Gaussians." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.274, + 0.484, + 0.385 + ], + "angle": 0, + "content": "Anisotropy. To prevent overly-skinny kernels that point outward from the object surface under large deformations, we enforce the anisotropy of Gaussian kernels following [102]. During optimization, we employ \\(\\mathcal{L}_{\\mathrm{ani}} = \\frac{1}{|P|}\\sum_{p\\in P}\\max \\left(\\frac{\\max(s_p)}{\\min(s_p)},\\tau\\right) - \\tau\\), where \\(s_p\\) is the scalings of 3D Gaussians. This loss constrains the ratio between the major and minor axis lengths below threshold \\(\\tau\\)." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.386, + 0.484, + 0.476 + ], + "angle": 0, + "content": "Total loss. To inpaint occluded textures, we use the \\(\\mathcal{L}_{SDS}\\) loss on the Gaussians in the canonical pose after optimizing the front view for 500 steps. Combined with the conventional 3D Gaussian Loss \\(\\mathcal{L}_{ori}\\) on image rendering, the total loss \\(L\\) for end-to-end optimization of clothing Gaussians via network \\(C_N\\) is:" + }, + { + "type": "equation", + "bbox": [ + 0.095, + 0.484, + 0.484, + 0.514 + ], + "angle": 0, + "content": "\\[\n\\mathcal {L} = \\lambda_ {\\text {o r i}} \\mathcal {L} _ {\\text {o r i}} + \\lambda_ {\\text {i d}} \\mathcal {L} _ {\\text {i d}} + \\lambda_ {\\text {a n i}} \\mathcal {L} _ {\\text {a n i}} + \\lambda_ {\\text {s d f}} \\mathcal {L} _ {\\text {s d f}} + \\lambda_ {\\text {S D S}} \\mathcal {L} _ {\\text {S D S}} \\tag {2}\n\\]" + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.518, + 0.395, + 0.535 + ], + "angle": 0, + "content": "3.6. 4D Human Animation and Editing" + }, + { + "type": "image_caption", + "bbox": [ + 0.104, + 0.542, + 0.289, + 0.553 + ], + "angle": 0, + "content": "a) SMPL-X Pose Driven Animation" + }, + { + "type": "image", + "bbox": [ + 0.118, + 0.555, + 0.442, + 0.609 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.104, + 0.612, + 0.27, + 0.623 + ], + "angle": 0, + "content": "b) Video-to-4D Reconstruction" + }, + { + "type": "image", + "bbox": [ + 0.111, + 0.624, + 0.468, + 0.821 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.09, + 0.826, + 0.484, + 0.91 + ], + "angle": 0, + "content": "Figure 3. 4D animation is achieved by (a) driving SMPL-X poses or (b) using video to learn additional clothing deformations. From the first frame, a static 3D disentangled GS model is generated. Pose transformations deform body and clothing Gaussians, and a deformation network is optimized to capture additional clothing deformations over time." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.092, + 0.907, + 0.198 + ], + "angle": 0, + "content": "Disco4D's disentangled representation naturally supports animation and editing. The canonical Gaussians \\( S_{body} \\) and \\( S_{cloth} \\) enable separate deformations for clothing and body, ensuring realistic animation. Besides, individual clothing categories can be easily edited using image or text prompts. The learned clothing can be transferred to different body shapes and poses, for versatile customization." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.199, + 0.909, + 0.336 + ], + "angle": 0, + "content": "Animating Gaussians. As shown in Figure 2, Disco4D enables animation of the canonical human Gaussian via two methods. Firstly, Gaussians can be directly driven using 3D SMPL-X sequences obtained from a motion database or estimated from 2D videos. Secondly, Disco4D enhances the model by learning detailed clothing dynamics from monocular videos. This disentanglement enables the focused modeling of clothing dynamics without altering the underlying human representation." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.337, + 0.909, + 0.61 + ], + "angle": 0, + "content": "To extend static 3D Gaussians into dynamic 4D Gaussians, a deformation network is trained to predict changes in position, rotation, and scale of the reposed clothing Gaussians based on a timestamp, as described in DreamGaussian4D [78]. Unlike [78], which learns deformations for all Gaussians, Disco4D models body Gaussians using the SMPL-X mesh, while clothing Gaussians employ posed transformations and learned deformations. The transformation is defined as \\( S'' = D_N(S', t) \\) where \\( D_N \\) is the deformation network, \\( S' \\) is the spatial descriptions of the reposed 3D clothing Gaussian, \\( t \\) is the timestamp, and \\( S'' \\) is the spatial descriptions of the deformed and reposed 3D clothing Gaussians. Following [78], the deformation model is initialized to predict zero deformation at the start of training to avoid divergence between dynamic and static models. The weights and biases of the final prediction heads are initialized to zero, and skip connections are introduced to enable gradient backpropagation." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.612, + 0.909, + 0.748 + ], + "angle": 0, + "content": "To optimize the deformation field using the reference view video, we minimize the reconstruction loss \\(\\mathcal{L}_{Ref}\\) between the rendered image and video frame at each timestep. To propagate the motion from the reference view to the entire 3D model, we leverage Zero-1-to-3-XL [19] to predict the deformation of the unseen part to calculate \\(\\mathcal{L}_{SDS}\\). Despite per-frame predictions of image diffusion models, the fixed color and opacity of static 3D Gaussians help preserve temporal consistency." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.75, + 0.911, + 0.902 + ], + "angle": 0, + "content": "Editing Clothing Gaussians. We extract the Gaussians corresponding to the specific category and edit them. This allows fine-grained editing and ensures that other Gaussians are not affected. Instead of fine-tuning all 3D Gaussians, we freeze the properties for most of the well-trained Gaussians and only adjust a small part of 3D Gaussians relevant to the target categories. For 3D object removal, we simply delete the 3D Gaussians of the editing target. For 3D object colorization by in-painting or text guidance, we reinitialize the color and tune the color (SH) parameters of the correspond-" + }, + { + "type": "page_number", + "bbox": [ + 0.479, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "26335" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.09, + 0.092, + 0.486, + 0.123 + ], + "angle": 0, + "content": "ing Gaussian group, while fixing the 3D positions and other properties to preserve the learned 3D geometry." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.135, + 0.226, + 0.153 + ], + "angle": 0, + "content": "4. Experiments" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.16, + 0.483, + 0.192 + ], + "angle": 0, + "content": "Our detailed implementation and experiment setup can be found in Appendix 7.3." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.199, + 0.242, + 0.214 + ], + "angle": 0, + "content": "4.1.3D Generation" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.221, + 0.486, + 0.766 + ], + "angle": 0, + "content": "Generation and Disentanglement. Our generation and disentanglement results are presented in Figure 4 and Table 2. We assessed the disentanglement quality using the Synbody [109] and CloSe [3] datasets, rendering 30 and 110 clothed human meshes respectively from four angles and evaluating CLIP-similarity, PSNR, SSIM, and LPIPS for various poses and views within the CloSe dataset. disco4D leverages diffusion models without requiring training on human specific datasets. Therefore, we compare it with DreamGaussian [89] and LGM [91] which reconstruct 3D objects from diffusion models. Additionally, we conducted comparisons with SHERF, a human-centric baseline for evaluating novel poses and views. Figure 4 shows disco4D has higher fidelity and better geometry for body parts such as face and limbs due to the representation using SMPL-X Gaussians. It outperforms DreamGaussian and SHERF on SynBody and CloSe benchmarks. disco4D performs worse than LGM on novel views, likely due to its optimization of Gaussians in canonical space for pose generalization, compromising view-specific detail. Editing. We can edit specific clothing appearance given an image or text prompt, repose the person and transfer person characteristics. The disentanglement allows fine-grained editing and modification of individual assets without affecting other assets, and stacking multiple edits (Figure 4). User study. We conducted a user study to evaluate the generative quality of our image-to-3D Gaussians reconstruction on random in-the-wild images from SHHQ, detailed in Table 3. This study focuses on reference view consistency and overall generation quality, crucial aspects in image reconstruction tasks. We rendered 360-degree rotation videos for 25 images generated by DreamGaussian, LGM, and disco4D. We invited 43 volunteers to rate \\(24\\sim 27\\) mixed samples from these methods on image consistency and overall model quality, yielding 1080 valid scores. As shown in Table 3, disco4D was preferred, demonstrating better alignment with the original image content and superior overall quality." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.774, + 0.238, + 0.789 + ], + "angle": 0, + "content": "4.2. 4D Animation" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.795, + 0.485, + 0.902 + ], + "angle": 0, + "content": "Pose-Driven Animation. disco4D generates canonical Gaussians that can be animated with any pose sequence. Figure 12 in the Appendix demonstrates our animation capabilities and compares them with current SOTA 2D animation methods. Using identical inputs—a single frame and pose sequence—our approach more effectively preserves the body shape and fine details such as facial features and clothing. It" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.092, + 0.908, + 0.182 + ], + "angle": 0, + "content": "surpasses Animate-Anyone [37] and Magic-Animate [106] in accurately modeling fine-grained body parts like hands and faces, and exhibits greater consistency compared to CHAMP [120]. The disentanglement feature of Disco4D further allows for direct manipulation of Clothing Gaussians, as shown in Figure 6." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.185, + 0.909, + 0.365 + ], + "angle": 0, + "content": "4D Reconstruction. For the 4D-Dress Dataset [96], we evaluated 8 sequences, assessing CLIP similarity scores against ground-truth meshes and disentangled assets, along with novel view performance (PSNR, SSIM, LPIPS) from four viewpoints. Table 4 summarizes our quantitative results, benchmarking Disco4D against existing video-to-4D general GS approaches, such as DreamGaussian4D [78], as well as human-centric GS methods, including MonoHuman [113], GART [54], and GaussianAvatar [38]. We evaluate on monocular videos comprising 14 frames, captured from a limited front-view perspective, without full-body visibility across frames." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.368, + 0.909, + 0.519 + ], + "angle": 0, + "content": "Disco4D outperforms MonoHuman [113], GART [54], and GaussianAvatar [38] (Table 4) as these methods reconstruct using known video information, unable to model unseen regions. Consequently, these methods cannot accurately model back views from front-facing videos, leading to artifacts in other perspectives and canonical space (see Figure 5). In contrast, Disco4D first performs reconstruction and subsequently incorporates details, such as clothing deformation, from the input frames, enabling consistent reconstruction even in unseen viewpoints." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.521, + 0.909, + 0.673 + ], + "angle": 0, + "content": "While DreamGaussian4D [78] is capable of modeling back-view information, the details remain coarse. Our results demonstrate that initializing with our model from the first frame (DreamGaussian4D Disco4D-init) significantly outperforms other initialization methods (DreamGaussian4DLGM init, DreamGaussian init) in both fidelity and geometry (Table 4). Nevertheless, without incorporating human priors, DreamGaussian4D [78] still faces challenges, such as missing limbs and difficulty modeling fine details like facial features (see Figure 13 in Appendix)." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.675, + 0.911, + 0.901 + ], + "angle": 0, + "content": "Reposing our canonical avatar enables us to align the body and assets accurately with the inferred postures from the source video, yielding high-quality reconstruction of faces, hands, and garments. Our reposed method surpasses DreamGaussian4D in geometry and fidelity by incorporating human priors. However, reposing alone cannot capture clothing dynamics. To address this, our disentangled approach models clothing deformations on the reposed Gaussians, guided by a diffusion model. As demonstrated in Figure 13 and Table 4, this process enhances the accuracy of clothing resemblance to the ground truth. The combination of asset repositioning and learned deformations improves modeling quality, with repositioning handling pose-driven changes and learned deformations simulating dynamic asset movements as observed in the driving video." + }, + { + "type": "page_number", + "bbox": [ + 0.48, + 0.946, + 0.521, + 0.957 + ], + "angle": 0, + "content": "26336" + } + ], + [ + { + "type": "table_caption", + "bbox": [ + 0.09, + 0.09, + 0.908, + 0.118 + ], + "angle": 0, + "content": "Table 2. CLIP-embedding loss for generated humans and segmented assets, and performance (PSNR, SSIM, LPIPS) comparisons for novel poses and views on the Synbody and CloSe datasets across DreamGaussian, LGM, SHERF, and Disco4D." + }, + { + "type": "table", + "bbox": [ + 0.093, + 0.118, + 0.905, + 0.203 + ], + "angle": 0, + "content": "
MethodSynBodyCloSe
CLIPNovel ViewCLIPNovel ViewNovel Pose
All ↑Pants ↑Shirt ↑Shoes ↑PSNR ↑SSIM ↑LPIPS ↓All ↑Pants ↑Shirt ↑Shoes ↑PSNR ↑SSIM ↑LPIPS ↓PSNR ↑SSIM ↑LPIPS ↓
DreamGaussian0.7510.7150.7100.74913.1180.8830.2290.7340.6930.6740.76720.080.9390.089---
LGM0.8070.7240.7470.76012.8840.8760.2280.8290.7270.7120.77820.500.9390.077---
SHERF0.7660.6490.6360.71415.1890.8520.1890.7770.7850.7290.80118.960.9120.08315.540.8440.165
Disco4D0.8510.7840.7530.80115.6910.8480.1850.8560.8580.8100.84220.100.9180.08117.960.8510.136
" + }, + { + "type": "image", + "bbox": [ + 0.102, + 0.215, + 0.892, + 0.468 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.162, + 0.469, + 0.835, + 0.484 + ], + "angle": 0, + "content": "Figure 4. Qualitative comparison of image generation across DreamGaussian, LGM, SHERF, and Disco4D." + }, + { + "type": "image", + "bbox": [ + 0.113, + 0.492, + 0.882, + 0.862 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.091, + 0.864, + 0.908, + 0.879 + ], + "angle": 0, + "content": "Figure 5. Qualitative comparison of 4D generation between DreamGaussian4D, MonoHuman, GART, GaussianAvatar, and Disco4D." + }, + { + "type": "page_number", + "bbox": [ + 0.479, + 0.946, + 0.52, + 0.957 + ], + "angle": 0, + "content": "26337" + } + ], + [ + { + "type": "table_caption", + "bbox": [ + 0.09, + 0.09, + 0.483, + 0.117 + ], + "angle": 0, + "content": "Table 3. User study rates quality of generated 3D Gaussians from 1-5. The higher the better." + }, + { + "type": "table", + "bbox": [ + 0.152, + 0.118, + 0.421, + 0.173 + ], + "angle": 0, + "content": "
MetricImage Consistency ↑Overall Quality ↑
DreamGaussian2.0171.852
LGM2.3382.017
Disco4D3.1423.037
" + }, + { + "type": "table_caption", + "bbox": [ + 0.09, + 0.182, + 0.486, + 0.224 + ], + "angle": 0, + "content": "Table 4. CLIP-embedding loss for generated humans and segmented assets, and performance (PSNR, SSIM, LPIPS) comparison on 4D-Dress across various video-to-4D methods." + }, + { + "type": "table", + "bbox": [ + 0.092, + 0.228, + 0.485, + 0.326 + ], + "angle": 0, + "content": "
All ↑Assets ↑PSNR ↑SSIM ↑LPIPS ↓
DreamGaussian4D0.7840.76920.540.930.080
MonoHuman0.7620.74320.220.920.086
GART0.8000.77218.810.920.086
GaussianAvatar0.8220.76820.010.930.069
DreamGaussian4D (LGM init)0.8090.79519.160.930.086
DreamGaussian4D (Disco4D init)0.8700.84921.020.930.065
Disco4D (reposed)0.8530.77423.940.950.049
Disco4D (reposed)+learned deformations0.9000.86525.460.960.035
" + }, + { + "type": "image", + "bbox": [ + 0.096, + 0.328, + 0.479, + 0.502 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.09, + 0.506, + 0.484, + 0.535 + ], + "angle": 0, + "content": "Figure 6. First frame Editing and Animation. Betas Editing, Recoloring (Text/Image-guided), Composition (Removal, Swap)." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.549, + 0.484, + 0.641 + ], + "angle": 0, + "content": "4D Editing. For a normal pipeline in character animation, editing the person in the video requires high consistency throughout all frames. For pose-driven animation methods, first frame editing and generation is required. Our method directly edits the Gaussians, which is more straightforward, fine-grained and consistent. This is seen from Figure 6." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.651, + 0.255, + 0.667 + ], + "angle": 0, + "content": "4.3. Ablation Studies" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.674, + 0.483, + 0.854 + ], + "angle": 0, + "content": "Initialization of Clothing Gaussians. This process is crucial for high fidelity reconstruction. As shown in Figure 8 in the Appendix, we evaluate different strategies, including random, surface, and hull-based initialization. Hull-based initialization significantly enhances the model accuracy and realism over other methods. Initialization directly on the SMPL-X surface often leads to inaccurate geometries, particularly with complex or loose garments, creating elongated, thin Gaussians and visual artifacts. In contrast, hull-based initialization captures garment details more effectively and maintains pose consistency, closely aligning with the true geometry of the clothed body." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.856, + 0.484, + 0.901 + ], + "angle": 0, + "content": "Geometry of Clothing Gaussians. Figure 14 in the Appendix highlights the differences in clothing geometry between DreamGaussian [89], LGM [91] and Disco4D. In" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.092, + 0.907, + 0.273 + ], + "angle": 0, + "content": "DreamGaussian, all points are confined within the body geometry, whereas in LGM, about half of the points extend beyond the SMPL-X body. Removing internal points leaves sparse, translucent representations for clothing. This sparsity suggests reliance on internal points for visual representation, failing to accurately depict the object's geometry where appearance should primarily originate from surface points. Often, clothing Gaussian points are incorrectly positioned inside the body's hull rather than on the surface. To better represent clothing geometry, Disco4D positions all clothing Gaussians externally to the SMPL-X body mesh, accurately reflecting the garment's actual physical characteristics." + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.274, + 0.907, + 0.32 + ], + "angle": 0, + "content": "Clothing editing. Figure 14 shows our editing results with the prompt \"Color the top pink\". disco4D allows for precise editing of the targeted clothing without affecting other areas." + }, + { + "type": "title", + "bbox": [ + 0.514, + 0.335, + 0.627, + 0.35 + ], + "angle": 0, + "content": "5. Discussion" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.361, + 0.909, + 0.679 + ], + "angle": 0, + "content": "Despite achieving impressive results, some failure cases still exist, as shown in Figure 7 in the Appendix. disco4D relies on robust and pixel-aligned SMPL-X estimation, which is still an unsolved problem. It occasionally fails for poor visual hull initialization. The extraction of mesh assets from clothing Gaussians using Local Density Query, as per DreamGaussian [89], currently loses fine-grained details. Enhancing the detail level of geometry derived from clothing Gaussians could bolster the utility of reconstructed assets in animation and simulation applications. Furthermore, the initialized visual hulls obtained from multi-view SMPL-X guided images are often of suboptimal quality and suffer from poor side and back views, necessitating refinement. Improving pose guidance models to achieve more accurate visual hulls could alleviate the need for extensive refinement. In addition, future works could look into modeling multi-layered clothing and reconstructing the occluded clothing. disco4D has many positive applications, but it also has the potential to facilitate deepfake avatars and raise IP concerns. Regulations should be built to address these issues alongside its benefits in the entertainment industry." + }, + { + "type": "title", + "bbox": [ + 0.514, + 0.694, + 0.634, + 0.71 + ], + "angle": 0, + "content": "6. Conclusion" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.72, + 0.909, + 0.901 + ], + "angle": 0, + "content": "We propose Disco4D, a novel approach for the generation of 3D animatable clothed human Gaussians from a single image, emphasizing high-fidelity detail and separation of assets. We manage to compositionally generate separate components, such as haircut, accessories, and decoupled outfits. Our core insight is the fixing of SMPL-X Gaussians, fitting segmented Gaussians over SMPL-X Gaussians, and application of diffusion models to enhance 3D reconstruction, including modeling occluded parts not visible in the input image. Its capability to separate assets offers significant advantages, including localized, fine-grained editing of individual assets and enhanced animatability." + }, + { + "type": "page_number", + "bbox": [ + 0.48, + 0.946, + 0.518, + 0.957 + ], + "angle": 0, + "content": "26338" + } + ], + [ + { + "type": "title", + "bbox": [ + 0.092, + 0.091, + 0.259, + 0.108 + ], + "angle": 0, + "content": "Acknowledgements" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.117, + 0.486, + 0.269 + ], + "angle": 0, + "content": "We sincerely thank the anonymous reviewers for their valuable comments on this paper. This research/project is supported by the National Research Foundation, Singapore under its AI Singapore Programme(AISG Award No: AISG3-PhD-2023-08-049T). This study is supported by the Ministry of Education, Singapore, under its MOE AcRF Tier 2 (MOET2EP20221-0012, MOE-T2EP20223-0002), and under the RIE2020 Industry Alignment Fund - Industry Collaboration Projects (IAF-ICP) Funding Initiative, as well as cash and in-kind contribution from the industry partner(s)." + }, + { + "type": "title", + "bbox": [ + 0.093, + 0.285, + 0.188, + 0.301 + ], + "angle": 0, + "content": "References" + }, + { + "type": "ref_text", + "bbox": [ + 0.11, + 0.31, + 0.484, + 0.37 + ], + "angle": 0, + "content": "[1] Jad Abou-Chakra, Feras Dayoub, and Niko Sunderhauf. Particlenerf: Particle based encoding for online neural radiance fields. arXiv preprint arXiv:2211.04041, 2022.3" + }, + { + "type": "ref_text", + "bbox": [ + 0.109, + 0.373, + 0.485, + 0.463 + ], + "angle": 0, + "content": "[2] Thiemo Alldieck, Mihai Zanfir, and Cristian Sminchisescu. Photorealistic Monocular 3D Reconstruction of Humans Wearing Clothing. Proceedings of the IEEE Computer Society Conference on Computer Vision and Pattern Recognition, 2022-June:1496-1505, 2022. 1, 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.11, + 0.466, + 0.487, + 0.54 + ], + "angle": 0, + "content": "[3] Dimitrije Antić, Garvita Tiwari, Batuhan Ozcomlekci, Riccardo Marin, and Gerard Pons-Moll. CloSe: A 3D clothing segmentation dataset and model. In International Conference on 3D Vision (3DV), 2024. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.11, + 0.544, + 0.485, + 0.635 + ], + "angle": 0, + "content": "[4] Omer Bar-Tal, Hila Chefer, Omer Tov, Charles Herrmann, Roni Paiss, Shiran Zada, Ariel Ephrat, Junhwa Hur, Guanghui Liu, Amit Raj, Yanzhen Li, Michael Rubinstein, Tomer Michaeli, Oliver Wang, Deqing Sun, Tali Dekel, and Inbar Mosseri. Lumiere: A spacetime diffusion model for video generation, 2024. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.11, + 0.638, + 0.484, + 0.695 + ], + "angle": 0, + "content": "[5] Hugo Bertiche, Meysam Madadi, and Sergio Escalera. Cloth3d: clothed 3d humans. In European Conference on Computer Vision, pages 344-359. Springer, 2020. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.11, + 0.7, + 0.484, + 0.774 + ], + "angle": 0, + "content": "[6] Bharat Lal Bhatnagar, Garvita Tiwari, Christian Theobalt, and Gerard Pons-Moll. Multi-garment net: Learning to dress 3d people from images. In IEEE International Conference on Computer Vision (ICCV). IEEE, 2019. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.11, + 0.778, + 0.484, + 0.868 + ], + "angle": 0, + "content": "[7] Andreas Blattmann, Tim Dockhorn, Sumith Kulal, Daniel Mendelevitch, Maciej Kilian, Dominik Lorenz, Yam Levi, Zion English, Vikram Voleti, Adam Letts, Varun Jampani, and Robin Rombach. Stable video diffusion: Scaling latent video diffusion models to large datasets, 2023. 2, 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.11, + 0.871, + 0.484, + 0.902 + ], + "angle": 0, + "content": "[8] Andreas Blattmann, Robin Rombach, Huan Ling, Tim Dockhorn, Seung Wook Kim, Sanja Fidler, and" + }, + { + "type": "list", + "bbox": [ + 0.109, + 0.31, + 0.487, + 0.902 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.515, + 0.092, + 0.907, + 0.121 + ], + "angle": 0, + "content": "Karsten Kreis. Align your latents: High-resolution video synthesis with latent diffusion models, 2023." + }, + { + "type": "ref_text", + "bbox": [ + 0.532, + 0.124, + 0.907, + 0.197 + ], + "angle": 0, + "content": "[9] Tim Brooks, Bill Peebles, Connor Holmes, Will DePue, Yufei Guo, Li Jing, David Schnurr, Joe Taylor, Troy Luhman, Eric Luhman, Clarence Ng, Ricky Wang, and Aditya Ramesh. Video generation models as world simulators. 2024. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.524, + 0.2, + 0.907, + 0.288 + ], + "angle": 0, + "content": "[10] Zhongang Cai, Wanqi Yin, Ailing Zeng, Chen Wei, Qingping Sun, Wang Yanjun, Hui En Pang, Haiyi Mei, Mingyuan Zhang, Lei Zhang, et al. Smpler-x: Scaling up expressive human pose and shape estimation. Advances in Neural Information Processing Systems, 36, 2024. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.525, + 0.291, + 0.907, + 0.32 + ], + "angle": 0, + "content": "[11] Ang Cao and Justin Johnson. Hexplane: A fast representation for dynamic scenes. CVPR, 2023. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.525, + 0.322, + 0.907, + 0.442 + ], + "angle": 0, + "content": "[12] Angel X. Chang, Thomas Funkhouser, Leonidas Guibas, Pat Hanrahan, Qixing Huang, Zimo Li, Silvio Savarese, Manolis Savva, Shuran Song, Hao Su, Jianxiong Xiao, Li Yi, and Fisher Yu. ShapeNet: An Information-Rich 3D Model Repository. Technical Report arXiv:1512.03012 [cs.GR], Stanford University — Princeton University — Toyota Technological Institute at Chicago, 2015. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.525, + 0.444, + 0.906, + 0.503 + ], + "angle": 0, + "content": "[13] Xin Chen, Anqi Pang, Yang Wei, Wang Peihao, Lan Xu, and Jingyi Yu. Tightcap: 3d human shape capture with clothing tightness field. ACM Transactions on Graphics (Presented at ACM SIGGRAPH), 2021. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.525, + 0.505, + 0.907, + 0.579 + ], + "angle": 0, + "content": "[14] Zhiqin Chen, Andrea Tagliasacchi, and Hao Zhang. BSP-net: Generating compact meshes via binary space partitioning. Proceedings of IEEE Conference on Computer Vision and Pattern Recognition (CVPR), 2020. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.525, + 0.581, + 0.906, + 0.625 + ], + "angle": 0, + "content": "[15] Zilong Chen, Yikai Wang, Feng Wang, Zhengyi Wang, and Huaping Liu. V3d: Video diffusion models are effective 3d generators, 2024. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.525, + 0.628, + 0.907, + 0.731 + ], + "angle": 0, + "content": "[16] Hongsuk Choi, Gyeongsik Moon, and Kyoung Mu Lee. Pose2Mesh: Graph Convolutional Network for 3D Human Pose and Mesh Recovery from a 2D Human Pose. Lecture Notes in Computer Science (including subseries Lecture Notes in Artificial Intelligence and Lecture Notes in Bioinformatics), 12352 LNCS: 769-787, 2020. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.525, + 0.734, + 0.907, + 0.808 + ], + "angle": 0, + "content": "[17] Vasileios Choutas, Georgios Pavlakos, Timo Bolkart, Dimitrios Tzionas, and Michael J. Black. Monocular expressive body regression through body-driven attention. In European Conference on Computer Vision (ECCV), 2020. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.525, + 0.81, + 0.907, + 0.87 + ], + "angle": 0, + "content": "[18] Enric Corona, Albert Pumarola, Guillem Alenyà, Gerard Pons-Moll, and Francesc Moreno-Noguer. Simplicit: Topology-aware generative model for clothed people. In CVPR, 2021. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.525, + 0.871, + 0.906, + 0.902 + ], + "angle": 0, + "content": "[19] Matt Deitke, Ruoshi Liu, Matthew Wallingford, Huong Ngo, Oscar Michel, Aditya Kusupati, Alan" + }, + { + "type": "list", + "bbox": [ + 0.524, + 0.092, + 0.907, + 0.902 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.48, + 0.945, + 0.52, + 0.957 + ], + "angle": 0, + "content": "26339" + } + ], + [ + { + "type": "ref_text", + "bbox": [ + 0.135, + 0.091, + 0.484, + 0.181 + ], + "angle": 0, + "content": "Fan, Christian Laforte, Vikram Voleti, Samir Yitzhak Gadre, Eli VanderBilt, Aniruddha Kembhavi, Carl Vondrick, Georgia Gkioxari, Kiana Ehsani, Ludwig Schmidt, and Ali Farhadi. Objverse-xl: A universe of \\(10\\mathrm{m} + 3\\mathrm{d}\\) objects. arXiv preprint arXiv:2307.05663, 2023. 2, 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.184, + 0.484, + 0.258 + ], + "angle": 0, + "content": "[20] Yilun Du, Yinan Zhang, Hong-Xing Yu, Joshua B. Tenenbaum, and Jiajun Wu. Neural radiance flow for 4d view synthesis and video processing. In Proceedings of the IEEE/CVF International Conference on Computer Vision, 2021. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.261, + 0.484, + 0.303 + ], + "angle": 0, + "content": "[21] Shivam Duggal and Deepak Pathak. Topologically-aware deformation fields for single-view 3d reconstruction. CVPR, 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.307, + 0.484, + 0.395 + ], + "angle": 0, + "content": "[22] Sai Kumar Dwivedi, Nikos Athanasiou, Muhammed Kocabas, and Michael J. Black. Learning to Regress Bodies from Images using Differentiable Semantic Rendering. In IEEE/CVF International Conference on Computer Vision (ICCV), pages 11230-11239, 2021. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.398, + 0.484, + 0.471 + ], + "angle": 0, + "content": "[23] Jiemin Fang, Taoran Yi, Xinggang Wang, Lingxi Xie, Xiaopeng Zhang, Wenyu Liu, Matthias Nießner, and Qi Tian. Fast dynamic radiance fields with time-aware neural voxels. In SIGGRAPH Asia 2022 Conference Papers. ACM, 2022. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.474, + 0.484, + 0.546 + ], + "angle": 0, + "content": "[24] Yao Feng, Vasileios Choutas, Timo Bolkart, Dimitrios Tzionas, and Michael J. Black. Collaborative regression of expressive bodies using moderation. In International Conference on 3D Vision (3DV), 2021." + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.551, + 0.484, + 0.61 + ], + "angle": 0, + "content": "[25] Yao Feng, Jinlong Yang, Marc Pollefeys, Michael J. Black, and Timo Bolkart. Capturing and animation of body and clothing from monocular video. In SIGGRAPH Asia 2022 Conference Papers, 2022. 1, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.612, + 0.484, + 0.669 + ], + "angle": 0, + "content": "[26] Yao Feng, Weiyang Liu, Timo Bolkart, Jinlong Yang, Marc Pollefeys, and Michael J. Black. Learning disentangled avatars with hybrid 3d representations. arXiv, 2023. 1, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.673, + 0.484, + 0.732 + ], + "angle": 0, + "content": "[27] Valentin Gabeur, Jean-Sebastien Franco, Xavier Martin, Cordelia Schmid, and Gregory Rogez. Moulding humans: Non-parametric 3d human shape estimation from single images, 2019. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.734, + 0.484, + 0.793 + ], + "angle": 0, + "content": "[28] Chen Gao, Ayush Saraf, Johannes Kopf, and Jia-Bin Huang. Dynamic view synthesis from dynamic monocular video. In Proceedings of the IEEE International Conference on Computer Vision, 2021. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.796, + 0.484, + 0.84 + ], + "angle": 0, + "content": "[29] Shanyan Guan, Huayu Deng, Yunbo Wang, and Xiaokang Yang. Neurofluid: Fluid dynamics grounding with particle-driven neural radiance fields, 2022. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.842, + 0.484, + 0.901 + ], + "angle": 0, + "content": "[30] Antoine Guédon and Vincent Lepetit. Sugar: Surface-aligned gaussian splatting for efficient 3d mesh reconstruction and high-quality mesh rendering. arXiv preprint arXiv:2311.12775, 2023. 3" + }, + { + "type": "list", + "bbox": [ + 0.102, + 0.091, + 0.484, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.525, + 0.092, + 0.906, + 0.167 + ], + "angle": 0, + "content": "[31] Yuwei Guo, Ceyuan Yang, Anyi Rao, Zhengyang Liang, Yaohui Wang, Yu Qiao, Maneesh Agrawala, Dahua Lin, and Bo Dai. Animatediff: Animate your personalized text-to-image diffusion models without specific tuning, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.525, + 0.171, + 0.906, + 0.244 + ], + "angle": 0, + "content": "[32] Oshri Halimi, Fabian Prada, Tuur Stuyck, Donglai Xiang, Timur Bagautdinov, He Wen, Ron Kimmel, Takaaki Shiratori, Chenglei Wu, and Yaser Sheikh. Garment avatars: Realistic cloth driving using pattern registration, 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.525, + 0.248, + 0.906, + 0.322 + ], + "angle": 0, + "content": "[33] Tong He, Yuanlu Xu, Shunsuke Saito, Stefano Soatto, and Tony Tung. \\(\\mathsf{ARCH} + +\\): Animation-Ready Clothed Human Reconstruction Revisited. Proceedings of the IEEE International Conference on Computer Vision, pages 11026–11036, 2021. 1, 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.525, + 0.326, + 0.906, + 0.416 + ], + "angle": 0, + "content": "[34] Zhu Heming, Cao Yu, Jin Hang, Chen Weikai, Du Dong, Wang Zhangye, Cui Shuguang, and Han Xiaoguang. Deep fashion3d: A dataset and benchmark for 3d garment reconstruction from single images. In Computer Vision - ECCV 2020, pages 512-530. Springer International Publishing, 2020. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.525, + 0.42, + 0.906, + 0.462 + ], + "angle": 0, + "content": "[35] Jonathan Ho, Ajay Jain, and Pieter Abbeel. Denoising diffusion probabilistic models. arXiv preprint arxiv:2006.11239, 2020. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.525, + 0.467, + 0.906, + 0.54 + ], + "angle": 0, + "content": "[36] Yicong Hong, Kai Zhang, Jiaxiang Gu, Sai Bi, Yang Zhou, Difan Liu, Feng Liu, Kalyan Sunkavalli, Trung Bui, and Hao Tan. Lrm: Large reconstruction model for single image to 3d. arXiv preprint arXiv:2311.04400, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.525, + 0.545, + 0.906, + 0.605 + ], + "angle": 0, + "content": "[37] Li Hu, Xin Gao, Peng Zhang, Ke Sun, Bang Zhang, and Liefeng Bo. Animate anyone: Consistent and controllable image-to-video synthesis for character animation. arXiv preprint arXiv:2311.17117, 2023. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.525, + 0.608, + 0.906, + 0.711 + ], + "angle": 0, + "content": "[38] Liangxiao Hu, Hongwen Zhang, Yuxiang Zhang, Boyao Zhou, Boning Liu, Shengping Zhang, and Liqiang Nie. Gaussianavatar: Towards realistic human avatar modeling from a single video via animatable 3d gaussians. In IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), 2024. 3, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.525, + 0.716, + 0.906, + 0.76 + ], + "angle": 0, + "content": "[39] Shoukang Hu and Ziwei Liu. Gauhuman: Articulated gaussian splatting from monocular human videos. arXiv preprint arXiv:, 2023. 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.525, + 0.764, + 0.906, + 0.822 + ], + "angle": 0, + "content": "[40] Shoukang Hu, Fangzhou Hong, Liang Pan, Haiyi Mei, Lei Yang, and Ziwei Liu. Sherf: Generalizable human nerf from a single image. arXiv preprint arXiv:2303.12791, 2023. 1, 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.525, + 0.826, + 0.906, + 0.901 + ], + "angle": 0, + "content": "[41] Yangyi Huang, Hongwei Yi, Weiyang Liu, Haofan Wang, Boxi Wu, Wenxiao Wang, Binbin Lin, Debing Zhang, and Deng Cai. One-shot implicit animatable avatars with model-based priors. In IEEE Conference on Computer Vision (ICCV), 2023. 2" + }, + { + "type": "list", + "bbox": [ + 0.525, + 0.092, + 0.906, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.48, + 0.945, + 0.52, + 0.957 + ], + "angle": 0, + "content": "26340" + } + ], + [ + { + "type": "ref_text", + "bbox": [ + 0.101, + 0.091, + 0.486, + 0.168 + ], + "angle": 0, + "content": "[42] Zeng Huang, Yuanlu Xu, Christoph Lassner, Hao Li, and Tony Tung. ARCH: Animatable Reconstruction of Clothed Humans. Proceedings of the IEEE Computer Society Conference on Computer Vision and Pattern Recognition, pages 3090-3099, 2020. 1, 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.101, + 0.17, + 0.485, + 0.231 + ], + "angle": 0, + "content": "[43] Boyi Jiang, Juyong Zhang, Yang Hong, Jinhao Luo, Ligang Liu, and Hujun Bao. BCnet: Learning body and cloth shape from a single image. In European Conference on Computer Vision. Springer, 2020. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.232, + 0.484, + 0.307 + ], + "angle": 0, + "content": "[44] Hanbyul Joo, Natalia Neverova, and Andrea Vedaldi. Exemplar Fine-Tuning for 3D Human Model Fitting Towards In-the-Wild 3D Human Pose Estimation. Proceedings - 2021 International Conference on 3D Vision, 3DV 2021, pages 42-52, 2021. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.103, + 0.309, + 0.483, + 0.34 + ], + "angle": 0, + "content": "[45] Heewoo Jun and Alex Nichol. Shap-e: Generating conditional 3d implicit functions, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.342, + 0.485, + 0.417 + ], + "angle": 0, + "content": "[46] Angjoo Kanazawa, Michael J Black, David W Jacobs, and Jitendra Malik. End-to-End Recovery of Human Shape and Pose. In Proceedings of the IEEE Computer Society Conference on Computer Vision and Pattern Recognition, pages 7122–7131, 2018. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.42, + 0.484, + 0.48 + ], + "angle": 0, + "content": "[47] Bernhard Kerbl, Georgios Kopanas, Thomas Leimkuhler, and George Drettakis. 3d gaussian splatting for real-time radiance field rendering. ACM Transactions on Graphics, 42(4), 2023. 2, 15" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.483, + 0.485, + 0.558 + ], + "angle": 0, + "content": "[48] Muhammed Kocabas, Chun-Hao P. Huang, Otmar Hilliges, and Michael J. Black. PARE: Part attention regressor for 3D human body estimation. In Proc. International Conference on Computer Vision (ICCV), pages 11127-11137, 2021. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.561, + 0.485, + 0.649 + ], + "angle": 0, + "content": "[49] Muhammed Kocabas, Chun-Hao P. Huang, Joachim Tesch, Lea Müller, Otmar Hilliges, and Michael J. Black. SPEC: Seeing people in the wild with an estimated camera. In Proc. International Conference on Computer Vision (ICCV), pages 11035-11045, 2021. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.653, + 0.484, + 0.698 + ], + "angle": 0, + "content": "[50] Muhammed Kocabas, Rick Chang, James Gabriel, Oncel Tuzel, and Anurag Ranjan. Hugs: Human gaussian splats, 2023. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.701, + 0.485, + 0.776 + ], + "angle": 0, + "content": "[51] Nikos Kolotouros, Georgios Pavlakos, Michael Black, and Kostas Daniilidis. Learning to reconstruct 3D human pose and shape via model-fitting in the loop. Proceedings of the IEEE International Conference on Computer Vision, 2019-Octob:2252-2261, 2019. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.779, + 0.485, + 0.853 + ], + "angle": 0, + "content": "[52] Georgios Kopanas, Julien Philip, Thomas Leimkuhler, and George Drettakis. Point-based neural rendering with per-view optimization. Computer Graphics Forum (Proceedings of the Eurographics Symposium on Rendering), 40(4), 2021. 15" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.856, + 0.485, + 0.901 + ], + "angle": 0, + "content": "[53] Georgios Kopanas, Thomas Leimkuhler, Gilles Rainer, Clément Jambon, and George Drettakis. Neural point catauastics for novel-view synthesis of re" + }, + { + "type": "list", + "bbox": [ + 0.101, + 0.091, + 0.486, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.553, + 0.092, + 0.907, + 0.121 + ], + "angle": 0, + "content": "flections. ACM Transactions on Graphics, 41(6): Article-201, 2022. 15" + }, + { + "type": "ref_text", + "bbox": [ + 0.525, + 0.124, + 0.906, + 0.169 + ], + "angle": 0, + "content": "[54] Jiahui Lei, Yufu Wang, Georgios Pavlakos, Lingjie Liu, and Kostas Daniilidis. Gart: Gaussian articulated template models, 2023. 3, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.524, + 0.172, + 0.906, + 0.217 + ], + "angle": 0, + "content": "[55] Mengtian Li, Shengxiang Yao, Zhifeng Xie, and Keyu Chen. Gaussianbody:Clothed human reconstruction via 3d gaussian splatting,2024.3" + }, + { + "type": "ref_text", + "bbox": [ + 0.524, + 0.219, + 0.908, + 0.279 + ], + "angle": 0, + "content": "[56] Xuanyi Li, Daquan Zhou, Chenxu Zhang, Shaodong Wei, Qibin Hou, and Ming-Ming Cheng. Sora generates videos with stunning geometrical consistency. arXiv preprint arXiv: 2402.17403, 2024. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.525, + 0.281, + 0.908, + 0.356 + ], + "angle": 0, + "content": "[57] Zhengqi Li, Simon Niklaus, Noah Snavely, and Oliver Wang. Neural scene flow fields for space-time view synthesis of dynamic scenes. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), 2021. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.525, + 0.359, + 0.908, + 0.433 + ], + "angle": 0, + "content": "[58] Zhengqi Li, Qianqian Wang, Forrester Cole, Richard Tucker, and Noah Snavely. Dynibar: Neural dynamic image-based rendering. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), 2023. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.525, + 0.436, + 0.908, + 0.51 + ], + "angle": 0, + "content": "[59] Minghua Liu, Chao Xu, Haian Jin, Linghao Chen, Zexiang Xu, Hao Su, et al. One-2-3-45: Any single image to 3d mesh in 45 seconds without per-shape optimization. arXiv preprint arXiv:2306.16928, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.525, + 0.513, + 0.908, + 0.559 + ], + "angle": 0, + "content": "[60] Ruoshi Liu, Rundi Wu, Basile Van Hoorick, Pavel Tokmakov, Sergey Zakharov, and Carl Vondrick. Zero-1-to-3: Zero-shot one image to 3d object, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.525, + 0.561, + 0.906, + 0.621 + ], + "angle": 0, + "content": "[61] Yang Liu, Xiang Huang, Minghan Qin, Qinwei Lin, and Haoqian Wang. Animatable 3d gaussian: Fast and high-quality reconstruction of multiple human avatars. arXiv preprint arXiv:2311.16482, 2023. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.525, + 0.624, + 0.908, + 0.698 + ], + "angle": 0, + "content": "[62] Matthew Loper, Naureen Mahmood, Javier Romero, Gerard Pons-Moll, and Michael J. Black. SMPL: A skinned multi-person linear model. ACM Trans. Graphics (Proc. SIGGRAPH Asia), 34(6):248:1-248:16, 2015. 2, 3, 15" + }, + { + "type": "ref_text", + "bbox": [ + 0.525, + 0.701, + 0.908, + 0.759 + ], + "angle": 0, + "content": "[63] Jonathon Luiten, Georgios Kopanas, Bastian Leibe, and Deva Ramanan. Dynamic 3d gaussians: Tracking by persistent dynamic view synthesis. In 3DV, 2024. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.525, + 0.763, + 0.908, + 0.839 + ], + "angle": 0, + "content": "[64] Luke Melas-Kyriazi, Iro Laina, Christian Rupprecht, and Andrea Vedaldi. Realfusion: 360deg reconstruction of any object from a single image. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 8446-8455, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.525, + 0.841, + 0.908, + 0.901 + ], + "angle": 0, + "content": "[65] Ben Mildenhall, Pratul P. Srinivasan, Matthew Tancik, Jonathan T. Barron, Ravi Ramamoorthi, and Ren Ng. Nerf: Representing scenes as neural radiance fields for view synthesis. In ECCV, 2020. 2" + }, + { + "type": "list", + "bbox": [ + 0.524, + 0.092, + 0.908, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.48, + 0.945, + 0.518, + 0.957 + ], + "angle": 0, + "content": "26341" + } + ], + [ + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.091, + 0.486, + 0.166 + ], + "angle": 0, + "content": "[66] Gyeongsik Moon, Hongsuk Choi, and Kyoung Mu Lee. Accurate 3d hand pose estimation for whole-body 3d human mesh estimation. In Computer Vision and Pattern Recognition Workshop (CVPRW), 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.169, + 0.484, + 0.228 + ], + "angle": 0, + "content": "[67] Alex Nichol, Heewoo Jun, Prafulla Dhariwal, Pamela Mishkin, and Mark Chen. Point-e: A system for generating 3d point clouds from complex prompts, 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.23, + 0.484, + 0.29 + ], + "angle": 0, + "content": "[68] Keunhong Park, Utkarsh Sinha, Jonathan T. Barron, Sofien Bouaziz, Dan B Goldman, Steven M. Seitz, and Ricardo Martin-Brualla. Nerfies: Deformable neural radiance fields. ICCV, 2021. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.292, + 0.484, + 0.382 + ], + "angle": 0, + "content": "[69] Keunhong Park, Utkarsh Sinha, Peter Hedman, Jonathan T. Barron, Sofien Bouaziz, Dan B Goldman, Ricardo Martin-Brualla, and Steven M. Seitz. Hypernerf: A higher-dimensional representation for topologically varying neural radiance fields. ACM Trans. Graph., 40(6), 2021. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.383, + 0.484, + 0.457 + ], + "angle": 0, + "content": "[70] Chaitanya Patel, Zhouyingcheng Liao, and Gerard Pons-Moll. Tailornet: Predicting clothing in 3d as a function of human pose, shape and garment style. In IEEE Conference on Computer Vision and Pattern Recognition (CVPR). IEEE, 2020. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.459, + 0.486, + 0.564 + ], + "angle": 0, + "content": "[71] Georgios Pavlakos, Vasileios Choutas, Nima Ghorbani, Timo Bolkart, Ahmed A. Osman, Dmitrios Tzionas, and Michael J. Black. Expressive body capture: 3D hands, face, and body from a single image. In Proceedings of the IEEE Computer Society Conference on Computer Vision and Pattern Recognition, pages 10967-10977, 2019. 1, 2, 3, 15" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.565, + 0.486, + 0.67 + ], + "angle": 0, + "content": "[72] Georgios Pavlakos, Vasileios Choutas, Nima Ghorbani, Timo Bolkart, Ahmed AA Osman, Dimitrios Tzionas, and Michael J Black. Expressive body capture: 3d hands, face, and body from a single image. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 10975-10985, 2019. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.672, + 0.486, + 0.748 + ], + "angle": 0, + "content": "[73] Gerard Pons-Moll, Sergi Pujades, Sonny Hu, and Michael Black. Clothcap: Seamless 4d clothing capture and retargeting. ACM Transactions on Graphics, (Proc. SIGGRAPH), 36(4), 2017. Two first authors contributed equally. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.749, + 0.486, + 0.793 + ], + "angle": 0, + "content": "[74] Ben Poole, Ajay Jain, Jonathan T. Barron, and Ben Mildenhall. Dreamfusion: Text-to-3d using 2d diffusion. arXiv, 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.795, + 0.486, + 0.87 + ], + "angle": 0, + "content": "[75] Albert Pumarola, Enric Corona, Gerard Pons-Moll, and Francesc Moreno-Noguer. D-NeRF: Neural Radiance Fields for Dynamic Scenes. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, 2020. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.871, + 0.486, + 0.903 + ], + "angle": 0, + "content": "[76] Guocheng Qian, Jinjie Mai, Abdullah Hamdi, Jian Ren, Aliaksandr Siarohin, Bing Li, Hsin-Ying Lee," + }, + { + "type": "list", + "bbox": [ + 0.102, + 0.091, + 0.486, + 0.903 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.553, + 0.091, + 0.908, + 0.18 + ], + "angle": 0, + "content": "Ivan Skorokhodov, Peter Wonka, Sergey Tulyakov, and Bernard Ghanem. Magic123: One image to high-quality 3d object generation using both 2d and 3d diffusion priors. In The Twelfth International Conference on Learning Representations (ICLR), 2024. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.524, + 0.185, + 0.908, + 0.231 + ], + "angle": 0, + "content": "[77] Zhiyin Qian, Shaofei Wang, Marko Mihajlovic, Andreas Geiger, and Siyu Tang. 3dgs- avatar: Animatable avatars via deformable 3d gaussian splatting. 2024. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.524, + 0.233, + 0.908, + 0.293 + ], + "angle": 0, + "content": "[78] Jiawei Ren, Liang Pan, Jiaxiang Tang, Chi Zhang, Ang Cao, Gang Zeng, and Ziwei Liu. Dreamgaussian4d: Generative 4d gaussian splatting. arXiv preprint arXiv:2312.17142, 2023. 3, 5, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.524, + 0.296, + 0.908, + 0.354 + ], + "angle": 0, + "content": "[79] Robin Rombach, Andreas Blattmann, Dominik Lorenz, Patrick Esser, and Björn Ommer. High-resolution image synthesis with latent diffusion models, 2021. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.524, + 0.359, + 0.908, + 0.434 + ], + "angle": 0, + "content": "[80] Yu Rong, Takaaki Shiratori, and Hanbyul Joo. FrankMocap: A Monocular 3D Whole-Body Pose Estimation System via Regression and Integration. In Proceedings of the IEEE International Conference on Computer Vision, pages 1749–1759, 2021. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.524, + 0.436, + 0.908, + 0.526 + ], + "angle": 0, + "content": "[81] Shunsuke Saito, Zeng Huang, Ryota Natsume, Shigeo Morishima, Angjoo Kanazawa, and Hao Li. Pifu: Pixel-aligned implicit function for high-resolution clothed human digitization. In The IEEE International Conference on Computer Vision (ICCV), 2019. 1, 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.524, + 0.529, + 0.908, + 0.589 + ], + "angle": 0, + "content": "[82] Shunsuke Saito, Tomas Simon, Jason Saragih, and Hanbyul Joo. Pifuhd: Multi-level pixel-aligned implicit function for high-resolution 3d human digitization. In CVPR, 2020. 1, 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.524, + 0.592, + 0.908, + 0.653 + ], + "angle": 0, + "content": "[83] Igor Santesteban, Miguel A. Otaduy, and Dan Casas. Learning-Based Animation of Clothing for Virtual Try-On. Computer Graphics Forum (Proc. Eurographics), 2019. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.524, + 0.655, + 0.908, + 0.715 + ], + "angle": 0, + "content": "[84] Sara Fridovich-Keil and Giacomo Meanti, Frederik Rahbæk Warburg, Benjamin Recht, and Angjoo Kanazawa. K-planes: Explicit radiance fields in space, time, and appearance. In CVPR, 2023. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.524, + 0.718, + 0.908, + 0.808 + ], + "angle": 0, + "content": "[85] Ruizhi Shao, Zerong Zheng, Hanzhang Tu, Boning Liu, Hongwen Zhang, and Yebin Liu. Tensor4d: Efficient neural 4d decomposition for high-fidelity dynamic reconstruction and rendering. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition, 2023. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.524, + 0.81, + 0.908, + 0.901 + ], + "angle": 0, + "content": "[86] Zhijing Shao, Zhaolong Wang, Zhuang Li, Duotun Wang, Xiangru Lin, Yu Zhang, Mingming Fan, and Zeyu Wang. SplattingAvatar: Realistic Real-Time Human Avatars with Mesh-Embedded Gaussian Splatting. In Computer Vision and Pattern Recognition (CVPR), 2024. 4" + }, + { + "type": "list", + "bbox": [ + 0.524, + 0.091, + 0.908, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.48, + 0.945, + 0.52, + 0.957 + ], + "angle": 0, + "content": "26342" + } + ], + [ + { + "type": "ref_text", + "bbox": [ + 0.101, + 0.091, + 0.486, + 0.166 + ], + "angle": 0, + "content": "[87] Ruoxi Shi, Hansheng Chen, Zhuoyang Zhang, Minghua Liu, Chao Xu, Xinyue Wei, Linghao Chen, Chong Zeng, and Hao Su. Zero123++: a single image to consistent multi-view diffusion base model, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.101, + 0.169, + 0.485, + 0.213 + ], + "angle": 0, + "content": "[88] Stanislaw Szymanowicz, Christian Rupprecht, and Andrea Vedaldi. Splatter image: Ultra-fast single-view 3d reconstruction, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.215, + 0.484, + 0.275 + ], + "angle": 0, + "content": "[89] Jiaxiang Tang, Jiawei Ren, Hang Zhou, Ziwei Liu, and Gang Zeng. Dreamgaussian: Generative gaussian splatting for efficient 3d content creation. arXiv preprint arXiv:2309.16653, 2023. 2, 3, 6, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.277, + 0.484, + 0.366 + ], + "angle": 0, + "content": "[90] Junshu Tang, Tengfei Wang, Bo Zhang, Ting Zhang, Ran Yi, Lizhuang Ma, and Dong Chen. Make-it-3d: High-fidelity 3d creation from a single image with diffusion prior. In Proceedings of the IEEE/CVF International Conference on Computer Vision (ICCV), pages 22819–22829, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.368, + 0.484, + 0.442 + ], + "angle": 0, + "content": "[91] Jiaxiang Tang, Zhaoxi Chen, Xiaokang Chen, Tengfei Wang, Gang Zeng, and Ziwei Liu. Lgm: Large multi-view gaussian model for high-resolution 3d content creation. arXiv preprint arXiv:2402.05054, 2024. 2, 3, 4, 6, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.444, + 0.484, + 0.534 + ], + "angle": 0, + "content": "[92] Edgar Tretschk, Ayush Tewari, Vladislav Golyanik, Michael Zollhöfer, Christoph Lassner, and Christian Theobalt. Non-rigid neural radiance fields: Reconstruction and novel view synthesis of a dynamic scene from monocular video. In IEEE International Conference on Computer Vision (ICCV). IEEE, 2021. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.535, + 0.484, + 0.58 + ], + "angle": 0, + "content": "[93] Alex Trevithick and Bo Yang. Grf: Learning a general radiance field for 3d scene representation and rendering. In arXiv:2010.04595, 2020. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.581, + 0.484, + 0.626 + ], + "angle": 0, + "content": "[94] Haithem Turki, Jason Y. Zhang, Francesco Ferroni, and Deva Ramanan. Suits: Scalable urban dynamic scenes, 2023. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.628, + 0.484, + 0.687 + ], + "angle": 0, + "content": "[95] Raquel Vidaurre, Igor Santesteban, Elena Garces, and Dan Casas. Fully Convolutional Graph Neural Networks for Parametric Virtual Try-On. Computer Graphics Forum (Proc. SCA), 2020. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.689, + 0.484, + 0.779 + ], + "angle": 0, + "content": "[96] Wenbo Wang, Hsuan-I Ho, Chen Guo, Boxiang Rong, Artur Grigorev, Jie Song, Juan Jose Zarate, and Otmar Hilliges. 4d-dress: A 4d dataset of real-world human clothing with semantic annotations. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition (CVPR), 2024. 6, 16" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.78, + 0.484, + 0.854 + ], + "angle": 0, + "content": "[97] Guanjun Wu, Taoran Yi, Jiemin Fang, Lingxi Xie, Xiaopeng Zhang, Wei Wei, Wenyu Liu, Qi Tian, and Wang Xinggang. 4d gaussian splatting for real-time dynamic scene rendering. arXiv preprint arXiv:2310.08528, 2023. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.856, + 0.484, + 0.901 + ], + "angle": 0, + "content": "[98] Wenqi Xian, Jia-Bin Huang, Johannes Kopf, and Changil Kim. Space-time neural irradiance fields for free-viewpoint video. In Proceedings of the" + }, + { + "type": "list", + "bbox": [ + 0.101, + 0.091, + 0.486, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.553, + 0.092, + 0.908, + 0.135 + ], + "angle": 0, + "content": "IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 9421-9431, 2021. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.525, + 0.138, + 0.908, + 0.213 + ], + "angle": 0, + "content": "[99] Donglai Xiang, Fabian Prada, Chenglei Wu, and Jessica Hodgins. Monoclothcap: Towards temporally coherent clothing capture from monocular rgb video. In Proceedings of International Conference on 3D Vision (3DV '20), pages 322 - 332, 2020. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.215, + 0.907, + 0.289 + ], + "angle": 0, + "content": "[100] Donglai Xiang, Fabian Prada, Timur Bagautdinov, Weipeng Xu, Yuan Dong, He Wen, Jessica Hodgins, and Chenglei Wu. Modeling clothing as a separate layer for an animatable human avatar. ACM Transactions on Graphics, 40(6):1-15, 2021. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.291, + 0.907, + 0.35 + ], + "angle": 0, + "content": "[101] Enze Xie, Wenhai Wang, Zhiding Yu, Anima Anandkumar, Jose M. Alvarez, and Ping Luo. Segformer: Simple and efficient design for semantic segmentation with transformers. CoRR, abs/2105.15203, 2021. 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.352, + 0.907, + 0.412 + ], + "angle": 0, + "content": "[102] Tianyi Xie, Zeshun Zong, Yuxing Qiu, Xuan Li, Yutao Feng, Yin Yang, and Chenfanfu Jiang. Physgaussian: Physics-integrated 3d gaussians for generative dynamics. arXiv preprint arXiv:2311.12198, 2023. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.414, + 0.907, + 0.503 + ], + "angle": 0, + "content": "[103] Yuliang Xiu, Jinlong Yang, Dimitrios Tzionas, and Michael J. Black. ICON: Implicit Clothed humans Obtained from Normals. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 13296-13306, 2022. 1, 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.505, + 0.908, + 0.594 + ], + "angle": 0, + "content": "[104] Yuliang Xiu, Jinlong Yang, Xu Cao, Dimitrios Tzionas, and Michael J. Black. ECON: Explicit Clothed humans Optimized via Normal integration. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), 2023. 1, 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.597, + 0.908, + 0.672 + ], + "angle": 0, + "content": "[105] Qiangeng Xu, Weiyue Wang, Duygu Ceylan, Radomir Mech, and Ulrich Neumann. Disn: Deep implicit surface network for high-quality single-view 3d reconstruction. In Advances in Neural Information Processing Systems 32, pages 492-502. 2019. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.673, + 0.908, + 0.747 + ], + "angle": 0, + "content": "[106] Zhongcong Xu, Jianfeng Zhang, Jun Hao Liew, Hanshu Yan, Jia-Wei Liu, Chenxu Zhang, Jiashi Feng, and Mike Zheng Shou. Magicanimate: Temporally consistent human image animation using diffusion model. 2023. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.749, + 0.908, + 0.825 + ], + "angle": 0, + "content": "[107] Xueting Yang, Yihao Luo, Yuliang Xiu, Wei Wang, Hao Xu, and Zhaoxin Fan. D-IF: Uncertainty-aware Human Digitization via Implicit Distribution Field. In Proceedings of the IEEE/CVF international conference on computer vision, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.826, + 0.908, + 0.901 + ], + "angle": 0, + "content": "[108] Yifan Yang, Dong Liu, Shuhai Zhang, Zeshuai Deng, Zixiong Huang, and Mingkui Tan. Hilo: Detailed and robust 3d clothed human reconstruction with high- and low-frequency information of parametric models. In Proceedings of the IEEE/CVF Conference" + }, + { + "type": "list", + "bbox": [ + 0.517, + 0.092, + 0.908, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.48, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "26343" + } + ], + [ + { + "type": "ref_text", + "bbox": [ + 0.135, + 0.092, + 0.484, + 0.121 + ], + "angle": 0, + "content": "on Computer Vision and Pattern Recognition, pages 10671-10681, 2024. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.124, + 0.484, + 0.245 + ], + "angle": 0, + "content": "[109] Zhitao Yang, Zhongang Cai, Haiyi Mei, Shuai Liu, Zhaoxi Chen, Weiye Xiao, Yukun Wei, Zhongfei Qing, Chen Wei, Bo Dai, Wayne Wu, Chen Qian, Dahua Lin, Ziwei Liu, and Lei Yang. Synbody: Synthetic dataset with layered human models for 3d human perception and modeling. In Proceedings of the IEEE/CVF International Conference on Computer Vision (ICCV), pages 20282-20292, 2023. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.248, + 0.484, + 0.322 + ], + "angle": 0, + "content": "[110] Ziyi Yang, Xinyu Gao, Wen Zhou, Shaohui Jiao, Yuqing Zhang, and Xiaogang Jin. Deformable 3d gaussians for high-fidelity monocular dynamic scene reconstruction. arXiv preprint arXiv:2309.13101, 2023. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.326, + 0.484, + 0.384 + ], + "angle": 0, + "content": "[111] Mingqiao Ye, Martin Danelljan, Fisher Yu, and Lei Ke. Gaussian grouping: Segment and edit anything in 3d scenes. arXiv preprint arXiv:2312.00732, 2023. 4, 15" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.389, + 0.484, + 0.463 + ], + "angle": 0, + "content": "[112] Xianggang Yu, Mutian Xu, Yidan Zhang, Haolin Liu, Chongjie Ye, Yushuang Wu, Zizheng Yan, Tianyou Liang, Guanying Chen, Shuguang Cui, and Xiaoguang Han. Mvimgnet: A large-scale dataset of multi-view images. In CVPR, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.467, + 0.484, + 0.541 + ], + "angle": 0, + "content": "[113] Zhengming Yu, Wei Cheng, Xian Liu, Wayne Wu, and Kwan-Yee Lin. Monohuman: Animatable human neural field from monocular video. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 16943-16953, 2023. 3, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.545, + 0.484, + 0.634 + ], + "angle": 0, + "content": "[114] Wentao Yuan, Zhaoyang Lv, Tanner Schmidt, and Steven Lovegrove. Star: Self-supervised tracking and reconstruction of rigid objects in motion with neural rendering. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 13144-13152, 2021. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.638, + 0.484, + 0.711 + ], + "angle": 0, + "content": "[115] Ilya Zakharkin, Kirill Mazur, Artur Grigorev, and Victor Lempitsky. Point-based modeling of human clothing. In Proceedings of the IEEE/CVF International Conference on Computer Vision (ICCV), pages 14718-14727, 2021. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.716, + 0.484, + 0.774 + ], + "angle": 0, + "content": "[116] Yan Zeng, Guoqiang Wei, Jiani Zheng, Jiaxin Zou, Yang Wei, Yuchen Zhang, and Hang Li. Make pixels dance: High-dynamic video generation. arXiv:2311.10982, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.778, + 0.484, + 0.837 + ], + "angle": 0, + "content": "[117] Zheng Zerong, Yu Tao, Liu Yebin, and Dai Qionghai. Pamir: Parametric model-conditioned implicit representation for image-based human reconstruction, 2021. 1, 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.841, + 0.484, + 0.901 + ], + "angle": 0, + "content": "[118] Hongwen Zhang, Yating Tian, Yuxiang Zhang, Mengcheng Li, Liang An, Zhenan Sun, and Yebin Liu. Pymaf-x: Towards well-aligned full-body model regression from monocular images. IEEE Transac-" + }, + { + "type": "list", + "bbox": [ + 0.093, + 0.092, + 0.484, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.559, + 0.092, + 0.907, + 0.121 + ], + "angle": 0, + "content": "tions on Pattern Analysis and Machine Intelligence, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.124, + 0.907, + 0.213 + ], + "angle": 0, + "content": "[119] Hao Zhu, Xinxin Zuo, Sen Wang, Xun Cao, and Ruigang Yang. Detailed human shape estimation from a single image by hierarchical mesh deformation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 4491-4500, 2019. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.216, + 0.907, + 0.288 + ], + "angle": 0, + "content": "[120] Shenhao Zhu, Junming Leo Chen, Zuozhuo Dai, Yinghui Xu, Xun Cao, Yao Yao, Hao Zhu, and Siyu Zhu. Champ: Controllable and consistent human image animation with 3d parametric guidance, 2024. 6" + }, + { + "type": "list", + "bbox": [ + 0.517, + 0.092, + 0.907, + 0.288 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.48, + 0.946, + 0.519, + 0.957 + ], + "angle": 0, + "content": "26344" + } + ] +] \ No newline at end of file diff --git a/2025/Disco4D_ Disentangled 4D Human Generation and Animation from a Single Image/5f039bdf-5dd5-496e-b90f-b0a10473c8cd_origin.pdf b/2025/Disco4D_ Disentangled 4D Human Generation and Animation from a Single Image/5f039bdf-5dd5-496e-b90f-b0a10473c8cd_origin.pdf new file mode 100644 index 0000000000000000000000000000000000000000..37fb2ddc638cd17517e65aac9527174a0b256b6e --- /dev/null +++ b/2025/Disco4D_ Disentangled 4D Human Generation and Animation from a Single Image/5f039bdf-5dd5-496e-b90f-b0a10473c8cd_origin.pdf @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:d2256ae7365210481f36f7abf19294641d26ca40249aebbf92568cf1bf35c63b +size 8280745 diff --git a/2025/Disco4D_ Disentangled 4D Human Generation and Animation from a Single Image/full.md b/2025/Disco4D_ Disentangled 4D Human Generation and Animation from a Single Image/full.md new file mode 100644 index 0000000000000000000000000000000000000000..196ad7f881f6dfa08f9e249369ee54e82b195eb2 --- /dev/null +++ b/2025/Disco4D_ Disentangled 4D Human Generation and Animation from a Single Image/full.md @@ -0,0 +1,334 @@ +# Disco4D: Disentangled 4D Human Generation and Animation from a Single Image + +Hui En Pang $^{1}$ , Shuai Liu $^{3}$ , Zhongang Cai $^{1,2,3}$ , Lei Yang $^{2,3}$ , Tianwei Zhang $^{1}$ , Ziwei Liu $^{1}$ + +$^{1}$ S-Lab, Nanyang Technological University $^{2}$ SenseTime Research $^{3}$ Shanghai AI Laboratory + +![](images/95e2c828af253cd8c94fd8208bcc383a17ce69683a910e6599a18759da9fab83.jpg) +Figure 1. Disco4D is a novel Gaussian Splatting framework for 4D disentangled human generation from a single image. Clothing and assets are modeled as Gaussians on top of the SMPL-X body, enabling dynamic animation and flexible editing. + +# Abstract + +We present Disco4D, a novel Gaussian Splatting framework for 4D human generation and animation from a single image. Different from existing methods, Disco4D distinctively disentangles clothings (with Gaussian models) from the human body (with SMPL-X model), significantly enhancing the generation details and flexibility. Specifically, 1) Disco4D learns to efficiently fit the clothing Gaussians over the SMPL-X Gaussians. 2) Next, Disco4D adopts diffusion models to enhance the 3D generation process, e.g., modeling occluded parts not visible in the input image. 3) Finally, Disco4D learns an identity encoding for each clothing Gaussian to facilitate the separation and extraction of clothing assets. Furthermore, Disco4D naturally supports 4D human animation with vivid dynamics. Extensive experiments demonstrate the superiority of Disco4D on 4D human generation and animation tasks. Our code is available at https://github.com/disco-4d/Disco4D. + +# 1. Introduction + +The development of high-fidelity 3D digital humans is increasingly important across a variety of augmented and virtual reality applications. To streamline the creation of these digital avatars from easily accessible in-the-wild images, a multitude of research efforts have been made on reconstructing 3D clothed human models from a single image [2, 33, 40-42, 81, 82, 103, 104, 104, 117]. These works predominantly focus on the simultaneous reconstruction of + +the human body and clothing. Unfortunately, these works have inherent limitations, and integrating them into applications that require virtual try-on or avatar customization poses significant challenges. This is primarily because the models are rendered as single-layer, non-animatable meshes where distinct attributes (e.g., hair, clothing, accessories) are merged into one continuous surface, with underlying layers completely obscured and self-contact areas inseparably connected. Such limitation complicates the re-animation and dynamic customization tasks. Existing works that perform layered reconstruction [25, 26] rely on self-rotating video inputs with extensive frames and viewpoints and involve substantial processing times. + +To address these issues, we propose Disco4D, a novel 4D clothed human reconstruction method that distinctly separates the human body from clothing elements from a single image. It supports human animation as the 4th dimension, which cannot be realized by prior static 3D reconstruction works [2, 33, 42, 81, 82, 103, 104, 117]. To achieve this, it employs the SMPL-X [71] parametric model to represent the human body, capitalizing on its efficacy in capturing body structure and kinematics. Conversely, clothing, along with dynamic and variable elements such as hair and accessories, is represented using Gaussian models, which are able to model the large variability in clothing. By binding Gaussians to a SMPL-X model and fixing it during the training phase, Disco4D ensures the integrity of the body while focusing the learning process on the appearance aspects. To model occluded portions not visible in the input image, diffusion models are used to enhance the 3D generation process. Moreover, Disco4D includes an identity grouping mecha + +nism for the Gaussians, which is instrumental in maintaining the separability and individuality of each clothing asset. + +The independent reconstruction of clothing and body offers several advantages. (1) Enhanced reconstruction fidelity. The SMPL-X body serves as a stable anchor for the clothing to conform to. By isolating the focus to learn clothing Gaussians, we achieve a more refined geometry and intricate detailing in the clothed model. (2) Fine-grained categorization and extraction of clothing items. Disco4D is able to separate clothing Gaussians into their respective categories, which is crucial for the recovery and utilization of individual clothing assets. (3) Extensive editing capabilities. Disco4D supports different editing functions, including the removal of specific items, inpainting (altering color or material), and other modifications. Such rich editing options allow for precise adjustments to individual assets without inadvertently affecting adjacent elements. This level of control is particularly beneficial in applications requiring detailed customization, such as virtual fashion design and digital content creation. (4) Improved animation capabilities. The body Gaussians adhere to the deformations dictated by the SMPL-X model, while clothing Gaussians conform to the underlying body movements but also exhibit behaviors true to their material characteristics. The disentangled deformation allows for nuanced adjustments to clothing behavior in response to complex body movements, thereby elevating the quality of clothed human animation. + +# 2. Related Works + +Table 1 summarizes the relevant 3D/4D generation methods. We describe their details below. + +# 2.1.3D Generation + +Single-image 3D Generation. Single-image reconstruction leverages advanced methods [45, 67] to generate 3D assets in the form of 3D point clouds or NeRF [65] from one image. While earlier efforts using auto-encoders focused on synthetic objects [12, 14, 21, 88, 93, 105], newer approaches treat the task as conditional generation, employing diffusion models [35] for 3D generation from both image and text [19, 35, 60, 64, 74, 76, 79, 90]. One-2-3-45 [59] uses 2D diffusion models [60, 87] to generate multiview images for reconstruction, while LRM [36] adopts transformer-based architecture to scale up the task on large datasets [19, 112]. Gaussian-based methods [47], particularly DreamGaussian [89] and LGM [91], offer efficient, high-resolution 3D model generation from text or images. Recently, video diffusion models have attracted significant attention due to their remarkable ability to generate intricate scenes and complex dynamics with great spatio-temporal consistency [4, 7-9, 31, 56, 116]. They are employed to generate consistent multi-view images, and then reconstruct underlying 3D assets with high quality [15]. + +Table 1. 3D/4D generation methods from a single image. + +
MethodTypeLayeredAnimatable
LGM [91]GeneralXX
PiFU [81]Human-centricXX
DreamFusion [74]GeneralXX
DreamGaussian [89]GeneralXX
PiFU [81]Human-centricXX
D-IF [107]Human-centricXX
HiLo [108]Human-centricXX
ECON [104]Human-centricXX
SHERF [40]Human-centricX✓
Disco4DHuman-centric✓✓
+ +Single-image human-centric 3D Generation. Significant research efforts have been made for 3D human reconstruction, which can be classified into the following categories. (1) Explicit-shape-based methods rely on Human Mesh Recovery (HMR) using parametric models like SMPL [62] and SMPL-X [71] to generate 3D body meshes [16, 17, 22, 24, 44, 46, 48, 49, 51, 66, 80, 118]. To account for 3D garments, several approaches incorporate offsets [99, 119] or templates, utilize deformable garment templates [6, 43], or employ non-parametric forms for clothed figures [27, 104, 115]. Despite their advancements, they face limitations in handling complex outfit variations and loose clothing due to inherent topological constraints. (2) Implicit-function-based methods utilize implicit representations like occupancy or distance fields for modeling clothed humans with complex geometries, such as loose garments. Techniques range from end-to-end regression of free-form implicit surfaces [2, 81, 82] to use of geometric priors [33, 42, 103, 104, 117] and implicit shape completion [104]. Notable works such as PIFu [81], ARCH(++) [33, 42], and PaMIR [117] can extract textured models from images, but struggle with depth ambiguities and texture inconsistencies. (3) NeRF-based methods incorporate model-based priors (i.e., SMPL-X) for accurate human reconstruction. Efforts like SHERF [40] and ELICIT [41] improve the reconstruction coherence by addressing 2D observation incompleteness leveraging appearance priors. Most of these 3D clothed human reconstruction and animation works [2, 33, 42, 81, 82, 103, 104, 117] require training on human-specific datasets, which brings another limitation on the availability of such datasets. + +3D Clothing Modeling. Reconstructing clothing from images and videos as a separate layer over the human body poses significant challenges due to the diversity of clothing topologies. Previous efforts relied on either template meshes or implicit surface models, and required extensive, high-quality 3D data from simulations [5, 70, 83, 95] or tailored template meshes [13, 32, 73, 100]. New methods were developed [34, 43] for multi-clothing models and versatile template meshes, respectively, facilitating diverse clothing topology encoding. However, these techniques typically fall short in capturing the clothing texture and appearance. The reliance on predefined clothing style templates further con + +strains their ability to handle real-world clothing variations. Corona et al. [18] addressed these shortcomings by representing clothing layers with deep unsigned distance functions and an auto-decoder for style and cut differentiation, though this often produces overly-smooth reconstructions [18]. On the other hand, SCARF [25] and DELTA [26] significantly enhance the visual fidelity by applying NeRF to clothing layers, but require self-rotating video inputs and considerable processing times. + +# 2.2. 4D Animation + +4D Animation. This task aims at capturing dynamic 3D scenes over time. Two primary approaches have emerged: modeling 4D scenes by adding time dimension $t$ or latent codes to spatial coordinates [28, 57, 98]; combining deformation fields with static 3D scenes [20, 58, 68, 69, 75, 92, 114]. Recent efforts in explicit or hybrid representations, like planar decomposition [11, 84, 85], hash representations [94], and other innovative methods [1, 23, 29], have improved reconstruction speed and quality. Gaussian Splatting, especially, stands out for balancing efficiency with quality, with dynamic 3D Gaussians [63] and 4D Gaussian Splatting [97, 110] techniques introducing time-dependent deformations to enhance reconstructions. Notably, DreamGaussian4D [78] stands out by minimizing the optimization time while achieving high-quality 4D reconstructions. + +Human-centric 4D Animation. Recent works leverage Gaussian-based methods [38, 50, 54, 55, 61, 77, 113] for 4D human reconstruction, requiring extensive frame sequences (50-100 frames) and/or multiple viewpoints. Currently there has not been any work on 4D layered human generation and animation from a single image or a video with few images, which will be achieved in this paper. + +# 3. Methodology + +# 3.1. Preliminaries + +3D Gaussian Splatting employs explicit 3D Gaussian points as its primary rendering entities. A 3D Gaussian point is defined as a function $G(x) = e^{-\frac{1}{2}(x - \mu)^T\Sigma^{-1}(x - \mu)}$ , where $\mu$ and $\Sigma$ are the spatial mean and covariance matrix, respectively. Each Gaussian is also associated with its own rotation $r$ , scaling $s$ , opacity $\alpha$ , a view-dependent color $c$ represented by spherical harmonic coefficients $f$ . + +SMPL-X parameterization [71] is an extension of the SMPL body model [62] with face and hand, designed to capture a more accurate representation of intricate body movements. SMPL-X is defined as a function $M(\beta, \theta, \psi) : \mathbb{R}^{|\beta| \times |\theta| \times |\psi|} \to \mathbb{R}^{3N}$ , parametrized by the pose $\theta \in \mathbb{R}^{3J}$ (where $J$ denotes the number of full body joints), body shape $\beta \in \mathbb{R}^{|\beta|}$ and facial expression $\psi \in \mathbb{R}^{|\psi|}$ . + +# 3.2. Overview + +Given a single image, Disco4D generates animatable 3D clothed human avatars in a bottom-up manner, facilitating natural separability. Our generated 3D clothed avatars, denoted as $S_{human}$ , are represented as the concatenation of $S_{body}$ and $S_{cloth}$ . Inspired by prior works [89, 91], $S$ capitalizes on Gaussian representations: + +$$ +S = G (\mu , r, s, \alpha , c, e), \tag {1} +$$ + +where $\mu, r, s, \alpha, c$ and $e$ denote positions, rotation, scaling, opacity, spherical harmonics coefficients and identity encoding, respectively. Different from traditional Gaussian representations, we add identity encoding $e$ to associate each Gaussian with its clothing category. + +Figure 2 depicts our framework. We start by generating colored SMPL-X Gaussians representing the body beneath clothing (Sec. 3.3). We obtain a visual hull for canonicalization and refine Gaussian predictions to align and envelop the SMPL-X mesh (Sec. 3.4). Next, we iteratively optimize canonical clothing Gaussians external to the SMPL-X mesh (Sec. 3.5). Lastly, we showcase the animation and editing of generated clothed avatars (Sec. 3.6). Notably, we leverage diffusion models to refine textures during 3D generation (Sec. 3.5) and extrapolate unseen views during 4D animation (Sec. 3.6). + +# 3.3. SMPL-X Gaussians + +Given an image, we first estimate coarse SMPL-X parameters with an off-the-shelf model [10], and then refine coarse predictions by fitting on 2D keypoints and clothing segmentation masks [72], obtaining pixel-aligned SMPL-X parameters $(\beta, \theta, \psi)$ . + +Mesh Binding. To convert the SMPL-X [71] mesh $M(\beta, \theta, \psi)$ into Gaussians $S_{body}$ for rendering, flat 3D Gaussians are bound to each mesh triangle, similar to SuGaR [30]. Gaussian means $\mu_{body}$ are computed using predefined barycentric coordinates, while Gaussian rotations $r_{body}$ derive from surface normals. The initial scaling $s_{body}$ ensures dense mesh coverage, with the last axis set to 0.1 for a uniformly thin surface. For color representation beneath clothing, opacity $\alpha_{body}$ is set to 1.0, with spherical harmonics $c_{body}$ optimized for each Gaussian. Visible skin color is supervised, while occluded skin color aligns with visible regions. A fixed label $e_{body}$ is assigned for rendering, remaining unchanged during training. When optimizing clothing Gaussians $S_{cloth}$ , SMPL-X Gaussians $S_{body}$ parameters stay fixed, preserving the body structure while allowing flexible learning for clothing. + +# 3.4. Initialization of Clothing Gaussians + +Cloth styles are diverse, making proper initialization crucial for effective clothing modeling. In synchronization with + +![](images/a42620725e9459866868a1c7ce133c3e2b8b35f0d94e5e20c14ece6d76cafb14.jpg) +Figure 2. Framework Overview of Disco4D. (a) 3D Generation utilizes a single image to obtain disentangled body and clothing Gaussians. Body, face and hand poses are refined to be pixel-aligned. For faster initialization, clothing Gaussians and visual hull are obtained with Gaussian Reconstruction Models. These clothing Gaussians are embedded to SMPL-X mesh and adopt the local coordinate system of the triangle. Subsequently, the iterative optimization process (pruning, identity encoding and densifying) separates the body and garments. The learned identity encodings guide the densification of the clothing Gaussians. (b) 4D Animation is achieved by either direct driving of SMPL-X poses or leveraging video to learn extra clothing deformation (refer to Figure 3 for more details). Various (c) 3D/4D Editing operations can be performed with our disentangled representation. + +estimating SMPL-X, we first employ the Video Diffusion Model [7] to estimate multi-view images. Subsequently, we leverage Gaussian Reconstruction Models [91] to obtain initial 3D Gaussians and their corresponding visual hull. Yet, the reconstructed 3D outputs often suffer from geometric inaccuracies, such as incorrect poses due to pose ambiguity or missing limbs. To address this, we refine the coarse visual hull to ensure it accurately aligns with and overlays the SMPL-X mesh and encapsulates a good geometry for the clothed figure. With SMPL-X aligned visual hull, we derive the refined Gaussians by adopting properties from their nearest neighbors. The refined visual hull and Gaussians are then canonicalized for the optimization phase. + +Mesh embedding. Each 3D clothing Gaussian is embedded on a triangle of the canonical mesh, defining its position in both canonical and posed spaces. The mean vertex position $\mathbf{O}$ serves as the origin of the local coordinate system, with the Gaussian positioned by an offset vector $\mathbf{v} = \sigma \mathbf{i} + \beta \mathbf{j} + \gamma \mathbf{k}$ , where $\sigma$ , $\beta$ , and $\gamma$ are the components of the displacement vector along the tangent $\mathbf{i}$ , bitangent $\mathbf{j}$ , and normal $\mathbf{k}$ . Unlike SplattingAvatar [86], which displaces Gaussians along the normal, our approach allows embedding to the most suitable triangle rather than the nearest one. For example, hair Gaussians are tagged to head faces instead of the nearest face for reposing [39, 86] (Figure 9 in Appendix). In animation, the Gaussian rotates with its embedded triangle face $(\delta r)$ , while scaling $(\delta s)$ is adjusted dynamically based on changes in edge lengths. During optimization, Gaussian and embedding parameters $(\mathbf{O}, \mathbf{v}, \delta r, \text{and } \delta s)$ are jointly updated. + +# 3.5. Optimization of Separable Gaussians + +With the SMPL-X Gaussian and initialized clothing Gaussian, we aim to optimize canonical clothing Gaussians $S_{cloth}$ + +outside the SMPL-X mesh. This involves three steps: 1) we use Signed Distance Function (SDF) loss and pruning to discourage and remove Gaussians that reside within the body; 2) we introduce identity encoding $e$ to attach a clothing label for each clothing Gaussian, by lifting multi-view 2D segmentations of the target object onto the 3D Gaussians; and 3) guided by $e_{body}$ and $e_{cloth}$ , we selectively densify only the relevant clothing points while ignoring body points. Once the disentangled clothing is obtained, we use SDS loss to in-paint high-resolution texture from the reference image to individual clothing Gaussians, thereby enriching the details of unseen regions. + +SDF Loss and Pruning. In reality, the clothing is always external to the body. During refinement, we ensure that the clothing Gaussians are positioned externally to the SMPL-X mesh by applying the SDF loss and a pruning strategy. Specifically, the SDF loss $\mathcal{L}_{sdf}$ penalizes any new densified Gaussians that intrude into the space of the SMPL-X mesh, ensuring that the clothing Gaussians consistently remain outside the body's surface. Pruning is applied at fixed intervals to reinforce this separation, and systematically remove any Gaussians located within the SDF of the SMPL-X mesh. + +Identity encoding. To associate each Gaussian to its clothing category, we introduce Identity Encoding $(e)$ , a learnable and compact vector of length 15, representing clothing categories from SegFormer [101] segmentation masks1. During training, the encodings are rendered into 2D segmentation masks in a differentiable manner following [111]. For classification, we apply a softmax to the rendered features $E_{\mathrm{id}}$ and use cross-entropy loss $\mathcal{L}_{2d}$ for $(K + 1)$ -category classifica + +tion. An unsupervised 3D regularization loss $\mathcal{L}_{3d}$ promotes spatial consistency among the top $k$ -nearest 3D Gaussians' Identity Encodings. Consequently, the overall identity loss is $\mathcal{L}_{id} = \mathcal{L}_{2d} + \mathcal{L}_{3d}$ . Refer to Appendix 7.2 for more details. Densification of clothing Gaussians. To learn clothing more efficiently, we perform sampling for categorical Gaussians that belong to the same clothing category and embedding. We find the $k$ -nearest Gaussian points for the resampled points and inherit their Gaussian properties (scaling, rotation, opacity, SH properties). By selectively densifying clothing Gaussians, we only add necessary Gaussians while ignoring body Gaussians. + +Anisotropy. To prevent overly-skinny kernels that point outward from the object surface under large deformations, we enforce the anisotropy of Gaussian kernels following [102]. During optimization, we employ $\mathcal{L}_{\mathrm{ani}} = \frac{1}{|P|}\sum_{p\in P}\max \left(\frac{\max(s_p)}{\min(s_p)},\tau\right) - \tau$ , where $s_p$ is the scalings of 3D Gaussians. This loss constrains the ratio between the major and minor axis lengths below threshold $\tau$ . + +Total loss. To inpaint occluded textures, we use the $\mathcal{L}_{SDS}$ loss on the Gaussians in the canonical pose after optimizing the front view for 500 steps. Combined with the conventional 3D Gaussian Loss $\mathcal{L}_{ori}$ on image rendering, the total loss $L$ for end-to-end optimization of clothing Gaussians via network $C_N$ is: + +$$ +\mathcal {L} = \lambda_ {\text {o r i}} \mathcal {L} _ {\text {o r i}} + \lambda_ {\text {i d}} \mathcal {L} _ {\text {i d}} + \lambda_ {\text {a n i}} \mathcal {L} _ {\text {a n i}} + \lambda_ {\text {s d f}} \mathcal {L} _ {\text {s d f}} + \lambda_ {\text {S D S}} \mathcal {L} _ {\text {S D S}} \tag {2} +$$ + +# 3.6. 4D Human Animation and Editing + +![](images/0f4a91071a9f36466d5bc8dae4a308e65ce2cab82b34981bcabd105fe8f64912.jpg) +a) SMPL-X Pose Driven Animation + +![](images/18159098878650bcb17563218212498802af2f353eefda2c9c7cb61d3d527ddb.jpg) +b) Video-to-4D Reconstruction +Figure 3. 4D animation is achieved by (a) driving SMPL-X poses or (b) using video to learn additional clothing deformations. From the first frame, a static 3D disentangled GS model is generated. Pose transformations deform body and clothing Gaussians, and a deformation network is optimized to capture additional clothing deformations over time. + +Disco4D's disentangled representation naturally supports animation and editing. The canonical Gaussians $S_{body}$ and $S_{cloth}$ enable separate deformations for clothing and body, ensuring realistic animation. Besides, individual clothing categories can be easily edited using image or text prompts. The learned clothing can be transferred to different body shapes and poses, for versatile customization. + +Animating Gaussians. As shown in Figure 2, Disco4D enables animation of the canonical human Gaussian via two methods. Firstly, Gaussians can be directly driven using 3D SMPL-X sequences obtained from a motion database or estimated from 2D videos. Secondly, Disco4D enhances the model by learning detailed clothing dynamics from monocular videos. This disentanglement enables the focused modeling of clothing dynamics without altering the underlying human representation. + +To extend static 3D Gaussians into dynamic 4D Gaussians, a deformation network is trained to predict changes in position, rotation, and scale of the reposed clothing Gaussians based on a timestamp, as described in DreamGaussian4D [78]. Unlike [78], which learns deformations for all Gaussians, Disco4D models body Gaussians using the SMPL-X mesh, while clothing Gaussians employ posed transformations and learned deformations. The transformation is defined as $S'' = D_N(S', t)$ where $D_N$ is the deformation network, $S'$ is the spatial descriptions of the reposed 3D clothing Gaussian, $t$ is the timestamp, and $S''$ is the spatial descriptions of the deformed and reposed 3D clothing Gaussians. Following [78], the deformation model is initialized to predict zero deformation at the start of training to avoid divergence between dynamic and static models. The weights and biases of the final prediction heads are initialized to zero, and skip connections are introduced to enable gradient backpropagation. + +To optimize the deformation field using the reference view video, we minimize the reconstruction loss $\mathcal{L}_{Ref}$ between the rendered image and video frame at each timestep. To propagate the motion from the reference view to the entire 3D model, we leverage Zero-1-to-3-XL [19] to predict the deformation of the unseen part to calculate $\mathcal{L}_{SDS}$ . Despite per-frame predictions of image diffusion models, the fixed color and opacity of static 3D Gaussians help preserve temporal consistency. + +Editing Clothing Gaussians. We extract the Gaussians corresponding to the specific category and edit them. This allows fine-grained editing and ensures that other Gaussians are not affected. Instead of fine-tuning all 3D Gaussians, we freeze the properties for most of the well-trained Gaussians and only adjust a small part of 3D Gaussians relevant to the target categories. For 3D object removal, we simply delete the 3D Gaussians of the editing target. For 3D object colorization by in-painting or text guidance, we reinitialize the color and tune the color (SH) parameters of the correspond- + +ing Gaussian group, while fixing the 3D positions and other properties to preserve the learned 3D geometry. + +# 4. Experiments + +Our detailed implementation and experiment setup can be found in Appendix 7.3. + +# 4.1.3D Generation + +Generation and Disentanglement. Our generation and disentanglement results are presented in Figure 4 and Table 2. We assessed the disentanglement quality using the Synbody [109] and CloSe [3] datasets, rendering 30 and 110 clothed human meshes respectively from four angles and evaluating CLIP-similarity, PSNR, SSIM, and LPIPS for various poses and views within the CloSe dataset. disco4D leverages diffusion models without requiring training on human specific datasets. Therefore, we compare it with DreamGaussian [89] and LGM [91] which reconstruct 3D objects from diffusion models. Additionally, we conducted comparisons with SHERF, a human-centric baseline for evaluating novel poses and views. Figure 4 shows disco4D has higher fidelity and better geometry for body parts such as face and limbs due to the representation using SMPL-X Gaussians. It outperforms DreamGaussian and SHERF on SynBody and CloSe benchmarks. disco4D performs worse than LGM on novel views, likely due to its optimization of Gaussians in canonical space for pose generalization, compromising view-specific detail. Editing. We can edit specific clothing appearance given an image or text prompt, repose the person and transfer person characteristics. The disentanglement allows fine-grained editing and modification of individual assets without affecting other assets, and stacking multiple edits (Figure 4). User study. We conducted a user study to evaluate the generative quality of our image-to-3D Gaussians reconstruction on random in-the-wild images from SHHQ, detailed in Table 3. This study focuses on reference view consistency and overall generation quality, crucial aspects in image reconstruction tasks. We rendered 360-degree rotation videos for 25 images generated by DreamGaussian, LGM, and disco4D. We invited 43 volunteers to rate $24\sim 27$ mixed samples from these methods on image consistency and overall model quality, yielding 1080 valid scores. As shown in Table 3, disco4D was preferred, demonstrating better alignment with the original image content and superior overall quality. + +# 4.2. 4D Animation + +Pose-Driven Animation. disco4D generates canonical Gaussians that can be animated with any pose sequence. Figure 12 in the Appendix demonstrates our animation capabilities and compares them with current SOTA 2D animation methods. Using identical inputs—a single frame and pose sequence—our approach more effectively preserves the body shape and fine details such as facial features and clothing. It + +surpasses Animate-Anyone [37] and Magic-Animate [106] in accurately modeling fine-grained body parts like hands and faces, and exhibits greater consistency compared to CHAMP [120]. The disentanglement feature of Disco4D further allows for direct manipulation of Clothing Gaussians, as shown in Figure 6. + +4D Reconstruction. For the 4D-Dress Dataset [96], we evaluated 8 sequences, assessing CLIP similarity scores against ground-truth meshes and disentangled assets, along with novel view performance (PSNR, SSIM, LPIPS) from four viewpoints. Table 4 summarizes our quantitative results, benchmarking Disco4D against existing video-to-4D general GS approaches, such as DreamGaussian4D [78], as well as human-centric GS methods, including MonoHuman [113], GART [54], and GaussianAvatar [38]. We evaluate on monocular videos comprising 14 frames, captured from a limited front-view perspective, without full-body visibility across frames. + +Disco4D outperforms MonoHuman [113], GART [54], and GaussianAvatar [38] (Table 4) as these methods reconstruct using known video information, unable to model unseen regions. Consequently, these methods cannot accurately model back views from front-facing videos, leading to artifacts in other perspectives and canonical space (see Figure 5). In contrast, Disco4D first performs reconstruction and subsequently incorporates details, such as clothing deformation, from the input frames, enabling consistent reconstruction even in unseen viewpoints. + +While DreamGaussian4D [78] is capable of modeling back-view information, the details remain coarse. Our results demonstrate that initializing with our model from the first frame (DreamGaussian4D Disco4D-init) significantly outperforms other initialization methods (DreamGaussian4DLGM init, DreamGaussian init) in both fidelity and geometry (Table 4). Nevertheless, without incorporating human priors, DreamGaussian4D [78] still faces challenges, such as missing limbs and difficulty modeling fine details like facial features (see Figure 13 in Appendix). + +Reposing our canonical avatar enables us to align the body and assets accurately with the inferred postures from the source video, yielding high-quality reconstruction of faces, hands, and garments. Our reposed method surpasses DreamGaussian4D in geometry and fidelity by incorporating human priors. However, reposing alone cannot capture clothing dynamics. To address this, our disentangled approach models clothing deformations on the reposed Gaussians, guided by a diffusion model. As demonstrated in Figure 13 and Table 4, this process enhances the accuracy of clothing resemblance to the ground truth. The combination of asset repositioning and learned deformations improves modeling quality, with repositioning handling pose-driven changes and learned deformations simulating dynamic asset movements as observed in the driving video. + +Table 2. CLIP-embedding loss for generated humans and segmented assets, and performance (PSNR, SSIM, LPIPS) comparisons for novel poses and views on the Synbody and CloSe datasets across DreamGaussian, LGM, SHERF, and Disco4D. + +
MethodSynBodyCloSe
CLIPNovel ViewCLIPNovel ViewNovel Pose
All ↑Pants ↑Shirt ↑Shoes ↑PSNR ↑SSIM ↑LPIPS ↓All ↑Pants ↑Shirt ↑Shoes ↑PSNR ↑SSIM ↑LPIPS ↓PSNR ↑SSIM ↑LPIPS ↓
DreamGaussian0.7510.7150.7100.74913.1180.8830.2290.7340.6930.6740.76720.080.9390.089---
LGM0.8070.7240.7470.76012.8840.8760.2280.8290.7270.7120.77820.500.9390.077---
SHERF0.7660.6490.6360.71415.1890.8520.1890.7770.7850.7290.80118.960.9120.08315.540.8440.165
Disco4D0.8510.7840.7530.80115.6910.8480.1850.8560.8580.8100.84220.100.9180.08117.960.8510.136
+ +![](images/d4c087aa346e36287966f162e97293bec9476662da2b7b9f9bf420bcef256338.jpg) +Figure 4. Qualitative comparison of image generation across DreamGaussian, LGM, SHERF, and Disco4D. + +![](images/57fd506cb5851f16f4b2827cb1b859baacff518a16203678fe9f75a94c0c5543.jpg) +Figure 5. Qualitative comparison of 4D generation between DreamGaussian4D, MonoHuman, GART, GaussianAvatar, and Disco4D. + +Table 3. User study rates quality of generated 3D Gaussians from 1-5. The higher the better. + +
MetricImage Consistency ↑Overall Quality ↑
DreamGaussian2.0171.852
LGM2.3382.017
Disco4D3.1423.037
+ +Table 4. CLIP-embedding loss for generated humans and segmented assets, and performance (PSNR, SSIM, LPIPS) comparison on 4D-Dress across various video-to-4D methods. + +
All ↑Assets ↑PSNR ↑SSIM ↑LPIPS ↓
DreamGaussian4D0.7840.76920.540.930.080
MonoHuman0.7620.74320.220.920.086
GART0.8000.77218.810.920.086
GaussianAvatar0.8220.76820.010.930.069
DreamGaussian4D (LGM init)0.8090.79519.160.930.086
DreamGaussian4D (Disco4D init)0.8700.84921.020.930.065
Disco4D (reposed)0.8530.77423.940.950.049
Disco4D (reposed)+learned deformations0.9000.86525.460.960.035
+ +![](images/465ac7c73588c24ace8492ee9eb936cb4a42e2d7e2432f4d7178a055f7c44cce.jpg) +Figure 6. First frame Editing and Animation. Betas Editing, Recoloring (Text/Image-guided), Composition (Removal, Swap). + +4D Editing. For a normal pipeline in character animation, editing the person in the video requires high consistency throughout all frames. For pose-driven animation methods, first frame editing and generation is required. Our method directly edits the Gaussians, which is more straightforward, fine-grained and consistent. This is seen from Figure 6. + +# 4.3. Ablation Studies + +Initialization of Clothing Gaussians. This process is crucial for high fidelity reconstruction. As shown in Figure 8 in the Appendix, we evaluate different strategies, including random, surface, and hull-based initialization. Hull-based initialization significantly enhances the model accuracy and realism over other methods. Initialization directly on the SMPL-X surface often leads to inaccurate geometries, particularly with complex or loose garments, creating elongated, thin Gaussians and visual artifacts. In contrast, hull-based initialization captures garment details more effectively and maintains pose consistency, closely aligning with the true geometry of the clothed body. + +Geometry of Clothing Gaussians. Figure 14 in the Appendix highlights the differences in clothing geometry between DreamGaussian [89], LGM [91] and Disco4D. In + +DreamGaussian, all points are confined within the body geometry, whereas in LGM, about half of the points extend beyond the SMPL-X body. Removing internal points leaves sparse, translucent representations for clothing. This sparsity suggests reliance on internal points for visual representation, failing to accurately depict the object's geometry where appearance should primarily originate from surface points. Often, clothing Gaussian points are incorrectly positioned inside the body's hull rather than on the surface. To better represent clothing geometry, Disco4D positions all clothing Gaussians externally to the SMPL-X body mesh, accurately reflecting the garment's actual physical characteristics. + +Clothing editing. Figure 14 shows our editing results with the prompt "Color the top pink". disco4D allows for precise editing of the targeted clothing without affecting other areas. + +# 5. Discussion + +Despite achieving impressive results, some failure cases still exist, as shown in Figure 7 in the Appendix. disco4D relies on robust and pixel-aligned SMPL-X estimation, which is still an unsolved problem. It occasionally fails for poor visual hull initialization. The extraction of mesh assets from clothing Gaussians using Local Density Query, as per DreamGaussian [89], currently loses fine-grained details. Enhancing the detail level of geometry derived from clothing Gaussians could bolster the utility of reconstructed assets in animation and simulation applications. Furthermore, the initialized visual hulls obtained from multi-view SMPL-X guided images are often of suboptimal quality and suffer from poor side and back views, necessitating refinement. Improving pose guidance models to achieve more accurate visual hulls could alleviate the need for extensive refinement. In addition, future works could look into modeling multi-layered clothing and reconstructing the occluded clothing. disco4D has many positive applications, but it also has the potential to facilitate deepfake avatars and raise IP concerns. Regulations should be built to address these issues alongside its benefits in the entertainment industry. + +# 6. Conclusion + +We propose Disco4D, a novel approach for the generation of 3D animatable clothed human Gaussians from a single image, emphasizing high-fidelity detail and separation of assets. We manage to compositionally generate separate components, such as haircut, accessories, and decoupled outfits. Our core insight is the fixing of SMPL-X Gaussians, fitting segmented Gaussians over SMPL-X Gaussians, and application of diffusion models to enhance 3D reconstruction, including modeling occluded parts not visible in the input image. Its capability to separate assets offers significant advantages, including localized, fine-grained editing of individual assets and enhanced animatability. + +# Acknowledgements + +We sincerely thank the anonymous reviewers for their valuable comments on this paper. This research/project is supported by the National Research Foundation, Singapore under its AI Singapore Programme(AISG Award No: AISG3-PhD-2023-08-049T). This study is supported by the Ministry of Education, Singapore, under its MOE AcRF Tier 2 (MOET2EP20221-0012, MOE-T2EP20223-0002), and under the RIE2020 Industry Alignment Fund - Industry Collaboration Projects (IAF-ICP) Funding Initiative, as well as cash and in-kind contribution from the industry partner(s). + +# References + +[1] Jad Abou-Chakra, Feras Dayoub, and Niko Sunderhauf. Particlenerf: Particle based encoding for online neural radiance fields. arXiv preprint arXiv:2211.04041, 2022.3 +[2] Thiemo Alldieck, Mihai Zanfir, and Cristian Sminchisescu. Photorealistic Monocular 3D Reconstruction of Humans Wearing Clothing. Proceedings of the IEEE Computer Society Conference on Computer Vision and Pattern Recognition, 2022-June:1496-1505, 2022. 1, 2 +[3] Dimitrije Antić, Garvita Tiwari, Batuhan Ozcomlekci, Riccardo Marin, and Gerard Pons-Moll. CloSe: A 3D clothing segmentation dataset and model. In International Conference on 3D Vision (3DV), 2024. 6 +[4] Omer Bar-Tal, Hila Chefer, Omer Tov, Charles Herrmann, Roni Paiss, Shiran Zada, Ariel Ephrat, Junhwa Hur, Guanghui Liu, Amit Raj, Yanzhen Li, Michael Rubinstein, Tomer Michaeli, Oliver Wang, Deqing Sun, Tali Dekel, and Inbar Mosseri. Lumiere: A spacetime diffusion model for video generation, 2024. 2 +[5] Hugo Bertiche, Meysam Madadi, and Sergio Escalera. Cloth3d: clothed 3d humans. In European Conference on Computer Vision, pages 344-359. Springer, 2020. 2 +[6] Bharat Lal Bhatnagar, Garvita Tiwari, Christian Theobalt, and Gerard Pons-Moll. Multi-garment net: Learning to dress 3d people from images. In IEEE International Conference on Computer Vision (ICCV). IEEE, 2019. 2 +[7] Andreas Blattmann, Tim Dockhorn, Sumith Kulal, Daniel Mendelevitch, Maciej Kilian, Dominik Lorenz, Yam Levi, Zion English, Vikram Voleti, Adam Letts, Varun Jampani, and Robin Rombach. Stable video diffusion: Scaling latent video diffusion models to large datasets, 2023. 2, 4 +[8] Andreas Blattmann, Robin Rombach, Huan Ling, Tim Dockhorn, Seung Wook Kim, Sanja Fidler, and + +Karsten Kreis. Align your latents: High-resolution video synthesis with latent diffusion models, 2023. +[9] Tim Brooks, Bill Peebles, Connor Holmes, Will DePue, Yufei Guo, Li Jing, David Schnurr, Joe Taylor, Troy Luhman, Eric Luhman, Clarence Ng, Ricky Wang, and Aditya Ramesh. Video generation models as world simulators. 2024. 2 +[10] Zhongang Cai, Wanqi Yin, Ailing Zeng, Chen Wei, Qingping Sun, Wang Yanjun, Hui En Pang, Haiyi Mei, Mingyuan Zhang, Lei Zhang, et al. Smpler-x: Scaling up expressive human pose and shape estimation. Advances in Neural Information Processing Systems, 36, 2024. 3 +[11] Ang Cao and Justin Johnson. Hexplane: A fast representation for dynamic scenes. CVPR, 2023. 3 +[12] Angel X. Chang, Thomas Funkhouser, Leonidas Guibas, Pat Hanrahan, Qixing Huang, Zimo Li, Silvio Savarese, Manolis Savva, Shuran Song, Hao Su, Jianxiong Xiao, Li Yi, and Fisher Yu. ShapeNet: An Information-Rich 3D Model Repository. Technical Report arXiv:1512.03012 [cs.GR], Stanford University — Princeton University — Toyota Technological Institute at Chicago, 2015. 2 +[13] Xin Chen, Anqi Pang, Yang Wei, Wang Peihao, Lan Xu, and Jingyi Yu. Tightcap: 3d human shape capture with clothing tightness field. ACM Transactions on Graphics (Presented at ACM SIGGRAPH), 2021. 2 +[14] Zhiqin Chen, Andrea Tagliasacchi, and Hao Zhang. BSP-net: Generating compact meshes via binary space partitioning. Proceedings of IEEE Conference on Computer Vision and Pattern Recognition (CVPR), 2020. 2 +[15] Zilong Chen, Yikai Wang, Feng Wang, Zhengyi Wang, and Huaping Liu. V3d: Video diffusion models are effective 3d generators, 2024. 2 +[16] Hongsuk Choi, Gyeongsik Moon, and Kyoung Mu Lee. Pose2Mesh: Graph Convolutional Network for 3D Human Pose and Mesh Recovery from a 2D Human Pose. Lecture Notes in Computer Science (including subseries Lecture Notes in Artificial Intelligence and Lecture Notes in Bioinformatics), 12352 LNCS: 769-787, 2020. 2 +[17] Vasileios Choutas, Georgios Pavlakos, Timo Bolkart, Dimitrios Tzionas, and Michael J. Black. Monocular expressive body regression through body-driven attention. In European Conference on Computer Vision (ECCV), 2020. 2 +[18] Enric Corona, Albert Pumarola, Guillem Alenyà, Gerard Pons-Moll, and Francesc Moreno-Noguer. Simplicit: Topology-aware generative model for clothed people. In CVPR, 2021. 3 +[19] Matt Deitke, Ruoshi Liu, Matthew Wallingford, Huong Ngo, Oscar Michel, Aditya Kusupati, Alan + +Fan, Christian Laforte, Vikram Voleti, Samir Yitzhak Gadre, Eli VanderBilt, Aniruddha Kembhavi, Carl Vondrick, Georgia Gkioxari, Kiana Ehsani, Ludwig Schmidt, and Ali Farhadi. Objverse-xl: A universe of $10\mathrm{m} + 3\mathrm{d}$ objects. arXiv preprint arXiv:2307.05663, 2023. 2, 5 +[20] Yilun Du, Yinan Zhang, Hong-Xing Yu, Joshua B. Tenenbaum, and Jiajun Wu. Neural radiance flow for 4d view synthesis and video processing. In Proceedings of the IEEE/CVF International Conference on Computer Vision, 2021. 3 +[21] Shivam Duggal and Deepak Pathak. Topologically-aware deformation fields for single-view 3d reconstruction. CVPR, 2022. 2 +[22] Sai Kumar Dwivedi, Nikos Athanasiou, Muhammed Kocabas, and Michael J. Black. Learning to Regress Bodies from Images using Differentiable Semantic Rendering. In IEEE/CVF International Conference on Computer Vision (ICCV), pages 11230-11239, 2021. 2 +[23] Jiemin Fang, Taoran Yi, Xinggang Wang, Lingxi Xie, Xiaopeng Zhang, Wenyu Liu, Matthias Nießner, and Qi Tian. Fast dynamic radiance fields with time-aware neural voxels. In SIGGRAPH Asia 2022 Conference Papers. ACM, 2022. 3 +[24] Yao Feng, Vasileios Choutas, Timo Bolkart, Dimitrios Tzionas, and Michael J. Black. Collaborative regression of expressive bodies using moderation. In International Conference on 3D Vision (3DV), 2021. +[25] Yao Feng, Jinlong Yang, Marc Pollefeys, Michael J. Black, and Timo Bolkart. Capturing and animation of body and clothing from monocular video. In SIGGRAPH Asia 2022 Conference Papers, 2022. 1, 3 +[26] Yao Feng, Weiyang Liu, Timo Bolkart, Jinlong Yang, Marc Pollefeys, and Michael J. Black. Learning disentangled avatars with hybrid 3d representations. arXiv, 2023. 1, 3 +[27] Valentin Gabeur, Jean-Sebastien Franco, Xavier Martin, Cordelia Schmid, and Gregory Rogez. Moulding humans: Non-parametric 3d human shape estimation from single images, 2019. 2 +[28] Chen Gao, Ayush Saraf, Johannes Kopf, and Jia-Bin Huang. Dynamic view synthesis from dynamic monocular video. In Proceedings of the IEEE International Conference on Computer Vision, 2021. 3 +[29] Shanyan Guan, Huayu Deng, Yunbo Wang, and Xiaokang Yang. Neurofluid: Fluid dynamics grounding with particle-driven neural radiance fields, 2022. 3 +[30] Antoine Guédon and Vincent Lepetit. Sugar: Surface-aligned gaussian splatting for efficient 3d mesh reconstruction and high-quality mesh rendering. arXiv preprint arXiv:2311.12775, 2023. 3 + +[31] Yuwei Guo, Ceyuan Yang, Anyi Rao, Zhengyang Liang, Yaohui Wang, Yu Qiao, Maneesh Agrawala, Dahua Lin, and Bo Dai. Animatediff: Animate your personalized text-to-image diffusion models without specific tuning, 2023. 2 +[32] Oshri Halimi, Fabian Prada, Tuur Stuyck, Donglai Xiang, Timur Bagautdinov, He Wen, Ron Kimmel, Takaaki Shiratori, Chenglei Wu, and Yaser Sheikh. Garment avatars: Realistic cloth driving using pattern registration, 2022. 2 +[33] Tong He, Yuanlu Xu, Shunsuke Saito, Stefano Soatto, and Tony Tung. $\mathsf{ARCH} + +$ : Animation-Ready Clothed Human Reconstruction Revisited. Proceedings of the IEEE International Conference on Computer Vision, pages 11026–11036, 2021. 1, 2 +[34] Zhu Heming, Cao Yu, Jin Hang, Chen Weikai, Du Dong, Wang Zhangye, Cui Shuguang, and Han Xiaoguang. Deep fashion3d: A dataset and benchmark for 3d garment reconstruction from single images. In Computer Vision - ECCV 2020, pages 512-530. Springer International Publishing, 2020. 2 +[35] Jonathan Ho, Ajay Jain, and Pieter Abbeel. Denoising diffusion probabilistic models. arXiv preprint arxiv:2006.11239, 2020. 2 +[36] Yicong Hong, Kai Zhang, Jiaxiang Gu, Sai Bi, Yang Zhou, Difan Liu, Feng Liu, Kalyan Sunkavalli, Trung Bui, and Hao Tan. Lrm: Large reconstruction model for single image to 3d. arXiv preprint arXiv:2311.04400, 2023. 2 +[37] Li Hu, Xin Gao, Peng Zhang, Ke Sun, Bang Zhang, and Liefeng Bo. Animate anyone: Consistent and controllable image-to-video synthesis for character animation. arXiv preprint arXiv:2311.17117, 2023. 6 +[38] Liangxiao Hu, Hongwen Zhang, Yuxiang Zhang, Boyao Zhou, Boning Liu, Shengping Zhang, and Liqiang Nie. Gaussianavatar: Towards realistic human avatar modeling from a single video via animatable 3d gaussians. In IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), 2024. 3, 6 +[39] Shoukang Hu and Ziwei Liu. Gauhuman: Articulated gaussian splatting from monocular human videos. arXiv preprint arXiv:, 2023. 4 +[40] Shoukang Hu, Fangzhou Hong, Liang Pan, Haiyi Mei, Lei Yang, and Ziwei Liu. Sherf: Generalizable human nerf from a single image. arXiv preprint arXiv:2303.12791, 2023. 1, 2 +[41] Yangyi Huang, Hongwei Yi, Weiyang Liu, Haofan Wang, Boxi Wu, Wenxiao Wang, Binbin Lin, Debing Zhang, and Deng Cai. One-shot implicit animatable avatars with model-based priors. In IEEE Conference on Computer Vision (ICCV), 2023. 2 + +[42] Zeng Huang, Yuanlu Xu, Christoph Lassner, Hao Li, and Tony Tung. ARCH: Animatable Reconstruction of Clothed Humans. Proceedings of the IEEE Computer Society Conference on Computer Vision and Pattern Recognition, pages 3090-3099, 2020. 1, 2 +[43] Boyi Jiang, Juyong Zhang, Yang Hong, Jinhao Luo, Ligang Liu, and Hujun Bao. BCnet: Learning body and cloth shape from a single image. In European Conference on Computer Vision. Springer, 2020. 2 +[44] Hanbyul Joo, Natalia Neverova, and Andrea Vedaldi. Exemplar Fine-Tuning for 3D Human Model Fitting Towards In-the-Wild 3D Human Pose Estimation. Proceedings - 2021 International Conference on 3D Vision, 3DV 2021, pages 42-52, 2021. 2 +[45] Heewoo Jun and Alex Nichol. Shap-e: Generating conditional 3d implicit functions, 2023. 2 +[46] Angjoo Kanazawa, Michael J Black, David W Jacobs, and Jitendra Malik. End-to-End Recovery of Human Shape and Pose. In Proceedings of the IEEE Computer Society Conference on Computer Vision and Pattern Recognition, pages 7122–7131, 2018. 2 +[47] Bernhard Kerbl, Georgios Kopanas, Thomas Leimkuhler, and George Drettakis. 3d gaussian splatting for real-time radiance field rendering. ACM Transactions on Graphics, 42(4), 2023. 2, 15 +[48] Muhammed Kocabas, Chun-Hao P. Huang, Otmar Hilliges, and Michael J. Black. PARE: Part attention regressor for 3D human body estimation. In Proc. International Conference on Computer Vision (ICCV), pages 11127-11137, 2021. 2 +[49] Muhammed Kocabas, Chun-Hao P. Huang, Joachim Tesch, Lea Müller, Otmar Hilliges, and Michael J. Black. SPEC: Seeing people in the wild with an estimated camera. In Proc. International Conference on Computer Vision (ICCV), pages 11035-11045, 2021. 2 +[50] Muhammed Kocabas, Rick Chang, James Gabriel, Oncel Tuzel, and Anurag Ranjan. Hugs: Human gaussian splats, 2023. 3 +[51] Nikos Kolotouros, Georgios Pavlakos, Michael Black, and Kostas Daniilidis. Learning to reconstruct 3D human pose and shape via model-fitting in the loop. Proceedings of the IEEE International Conference on Computer Vision, 2019-Octob:2252-2261, 2019. 2 +[52] Georgios Kopanas, Julien Philip, Thomas Leimkuhler, and George Drettakis. Point-based neural rendering with per-view optimization. Computer Graphics Forum (Proceedings of the Eurographics Symposium on Rendering), 40(4), 2021. 15 +[53] Georgios Kopanas, Thomas Leimkuhler, Gilles Rainer, Clément Jambon, and George Drettakis. Neural point catauastics for novel-view synthesis of re + +flections. ACM Transactions on Graphics, 41(6): Article-201, 2022. 15 +[54] Jiahui Lei, Yufu Wang, Georgios Pavlakos, Lingjie Liu, and Kostas Daniilidis. Gart: Gaussian articulated template models, 2023. 3, 6 +[55] Mengtian Li, Shengxiang Yao, Zhifeng Xie, and Keyu Chen. Gaussianbody:Clothed human reconstruction via 3d gaussian splatting,2024.3 +[56] Xuanyi Li, Daquan Zhou, Chenxu Zhang, Shaodong Wei, Qibin Hou, and Ming-Ming Cheng. Sora generates videos with stunning geometrical consistency. arXiv preprint arXiv: 2402.17403, 2024. 2 +[57] Zhengqi Li, Simon Niklaus, Noah Snavely, and Oliver Wang. Neural scene flow fields for space-time view synthesis of dynamic scenes. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), 2021. 3 +[58] Zhengqi Li, Qianqian Wang, Forrester Cole, Richard Tucker, and Noah Snavely. Dynibar: Neural dynamic image-based rendering. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), 2023. 3 +[59] Minghua Liu, Chao Xu, Haian Jin, Linghao Chen, Zexiang Xu, Hao Su, et al. One-2-3-45: Any single image to 3d mesh in 45 seconds without per-shape optimization. arXiv preprint arXiv:2306.16928, 2023. 2 +[60] Ruoshi Liu, Rundi Wu, Basile Van Hoorick, Pavel Tokmakov, Sergey Zakharov, and Carl Vondrick. Zero-1-to-3: Zero-shot one image to 3d object, 2023. 2 +[61] Yang Liu, Xiang Huang, Minghan Qin, Qinwei Lin, and Haoqian Wang. Animatable 3d gaussian: Fast and high-quality reconstruction of multiple human avatars. arXiv preprint arXiv:2311.16482, 2023. 3 +[62] Matthew Loper, Naureen Mahmood, Javier Romero, Gerard Pons-Moll, and Michael J. Black. SMPL: A skinned multi-person linear model. ACM Trans. Graphics (Proc. SIGGRAPH Asia), 34(6):248:1-248:16, 2015. 2, 3, 15 +[63] Jonathon Luiten, Georgios Kopanas, Bastian Leibe, and Deva Ramanan. Dynamic 3d gaussians: Tracking by persistent dynamic view synthesis. In 3DV, 2024. 3 +[64] Luke Melas-Kyriazi, Iro Laina, Christian Rupprecht, and Andrea Vedaldi. Realfusion: 360deg reconstruction of any object from a single image. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 8446-8455, 2023. 2 +[65] Ben Mildenhall, Pratul P. Srinivasan, Matthew Tancik, Jonathan T. Barron, Ravi Ramamoorthi, and Ren Ng. Nerf: Representing scenes as neural radiance fields for view synthesis. In ECCV, 2020. 2 + +[66] Gyeongsik Moon, Hongsuk Choi, and Kyoung Mu Lee. Accurate 3d hand pose estimation for whole-body 3d human mesh estimation. In Computer Vision and Pattern Recognition Workshop (CVPRW), 2022. 2 +[67] Alex Nichol, Heewoo Jun, Prafulla Dhariwal, Pamela Mishkin, and Mark Chen. Point-e: A system for generating 3d point clouds from complex prompts, 2022. 2 +[68] Keunhong Park, Utkarsh Sinha, Jonathan T. Barron, Sofien Bouaziz, Dan B Goldman, Steven M. Seitz, and Ricardo Martin-Brualla. Nerfies: Deformable neural radiance fields. ICCV, 2021. 3 +[69] Keunhong Park, Utkarsh Sinha, Peter Hedman, Jonathan T. Barron, Sofien Bouaziz, Dan B Goldman, Ricardo Martin-Brualla, and Steven M. Seitz. Hypernerf: A higher-dimensional representation for topologically varying neural radiance fields. ACM Trans. Graph., 40(6), 2021. 3 +[70] Chaitanya Patel, Zhouyingcheng Liao, and Gerard Pons-Moll. Tailornet: Predicting clothing in 3d as a function of human pose, shape and garment style. In IEEE Conference on Computer Vision and Pattern Recognition (CVPR). IEEE, 2020. 2 +[71] Georgios Pavlakos, Vasileios Choutas, Nima Ghorbani, Timo Bolkart, Ahmed A. Osman, Dmitrios Tzionas, and Michael J. Black. Expressive body capture: 3D hands, face, and body from a single image. In Proceedings of the IEEE Computer Society Conference on Computer Vision and Pattern Recognition, pages 10967-10977, 2019. 1, 2, 3, 15 +[72] Georgios Pavlakos, Vasileios Choutas, Nima Ghorbani, Timo Bolkart, Ahmed AA Osman, Dimitrios Tzionas, and Michael J Black. Expressive body capture: 3d hands, face, and body from a single image. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 10975-10985, 2019. 3 +[73] Gerard Pons-Moll, Sergi Pujades, Sonny Hu, and Michael Black. Clothcap: Seamless 4d clothing capture and retargeting. ACM Transactions on Graphics, (Proc. SIGGRAPH), 36(4), 2017. Two first authors contributed equally. 2 +[74] Ben Poole, Ajay Jain, Jonathan T. Barron, and Ben Mildenhall. Dreamfusion: Text-to-3d using 2d diffusion. arXiv, 2022. 2 +[75] Albert Pumarola, Enric Corona, Gerard Pons-Moll, and Francesc Moreno-Noguer. D-NeRF: Neural Radiance Fields for Dynamic Scenes. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, 2020. 3 +[76] Guocheng Qian, Jinjie Mai, Abdullah Hamdi, Jian Ren, Aliaksandr Siarohin, Bing Li, Hsin-Ying Lee, + +Ivan Skorokhodov, Peter Wonka, Sergey Tulyakov, and Bernard Ghanem. Magic123: One image to high-quality 3d object generation using both 2d and 3d diffusion priors. In The Twelfth International Conference on Learning Representations (ICLR), 2024. 2 +[77] Zhiyin Qian, Shaofei Wang, Marko Mihajlovic, Andreas Geiger, and Siyu Tang. 3dgs- avatar: Animatable avatars via deformable 3d gaussian splatting. 2024. 3 +[78] Jiawei Ren, Liang Pan, Jiaxiang Tang, Chi Zhang, Ang Cao, Gang Zeng, and Ziwei Liu. Dreamgaussian4d: Generative 4d gaussian splatting. arXiv preprint arXiv:2312.17142, 2023. 3, 5, 6 +[79] Robin Rombach, Andreas Blattmann, Dominik Lorenz, Patrick Esser, and Björn Ommer. High-resolution image synthesis with latent diffusion models, 2021. 2 +[80] Yu Rong, Takaaki Shiratori, and Hanbyul Joo. FrankMocap: A Monocular 3D Whole-Body Pose Estimation System via Regression and Integration. In Proceedings of the IEEE International Conference on Computer Vision, pages 1749–1759, 2021. 2 +[81] Shunsuke Saito, Zeng Huang, Ryota Natsume, Shigeo Morishima, Angjoo Kanazawa, and Hao Li. Pifu: Pixel-aligned implicit function for high-resolution clothed human digitization. In The IEEE International Conference on Computer Vision (ICCV), 2019. 1, 2 +[82] Shunsuke Saito, Tomas Simon, Jason Saragih, and Hanbyul Joo. Pifuhd: Multi-level pixel-aligned implicit function for high-resolution 3d human digitization. In CVPR, 2020. 1, 2 +[83] Igor Santesteban, Miguel A. Otaduy, and Dan Casas. Learning-Based Animation of Clothing for Virtual Try-On. Computer Graphics Forum (Proc. Eurographics), 2019. 2 +[84] Sara Fridovich-Keil and Giacomo Meanti, Frederik Rahbæk Warburg, Benjamin Recht, and Angjoo Kanazawa. K-planes: Explicit radiance fields in space, time, and appearance. In CVPR, 2023. 3 +[85] Ruizhi Shao, Zerong Zheng, Hanzhang Tu, Boning Liu, Hongwen Zhang, and Yebin Liu. Tensor4d: Efficient neural 4d decomposition for high-fidelity dynamic reconstruction and rendering. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition, 2023. 3 +[86] Zhijing Shao, Zhaolong Wang, Zhuang Li, Duotun Wang, Xiangru Lin, Yu Zhang, Mingming Fan, and Zeyu Wang. SplattingAvatar: Realistic Real-Time Human Avatars with Mesh-Embedded Gaussian Splatting. In Computer Vision and Pattern Recognition (CVPR), 2024. 4 + +[87] Ruoxi Shi, Hansheng Chen, Zhuoyang Zhang, Minghua Liu, Chao Xu, Xinyue Wei, Linghao Chen, Chong Zeng, and Hao Su. Zero123++: a single image to consistent multi-view diffusion base model, 2023. 2 +[88] Stanislaw Szymanowicz, Christian Rupprecht, and Andrea Vedaldi. Splatter image: Ultra-fast single-view 3d reconstruction, 2023. 2 +[89] Jiaxiang Tang, Jiawei Ren, Hang Zhou, Ziwei Liu, and Gang Zeng. Dreamgaussian: Generative gaussian splatting for efficient 3d content creation. arXiv preprint arXiv:2309.16653, 2023. 2, 3, 6, 8 +[90] Junshu Tang, Tengfei Wang, Bo Zhang, Ting Zhang, Ran Yi, Lizhuang Ma, and Dong Chen. Make-it-3d: High-fidelity 3d creation from a single image with diffusion prior. In Proceedings of the IEEE/CVF International Conference on Computer Vision (ICCV), pages 22819–22829, 2023. 2 +[91] Jiaxiang Tang, Zhaoxi Chen, Xiaokang Chen, Tengfei Wang, Gang Zeng, and Ziwei Liu. Lgm: Large multi-view gaussian model for high-resolution 3d content creation. arXiv preprint arXiv:2402.05054, 2024. 2, 3, 4, 6, 8 +[92] Edgar Tretschk, Ayush Tewari, Vladislav Golyanik, Michael Zollhöfer, Christoph Lassner, and Christian Theobalt. Non-rigid neural radiance fields: Reconstruction and novel view synthesis of a dynamic scene from monocular video. In IEEE International Conference on Computer Vision (ICCV). IEEE, 2021. 3 +[93] Alex Trevithick and Bo Yang. Grf: Learning a general radiance field for 3d scene representation and rendering. In arXiv:2010.04595, 2020. 2 +[94] Haithem Turki, Jason Y. Zhang, Francesco Ferroni, and Deva Ramanan. Suits: Scalable urban dynamic scenes, 2023. 3 +[95] Raquel Vidaurre, Igor Santesteban, Elena Garces, and Dan Casas. Fully Convolutional Graph Neural Networks for Parametric Virtual Try-On. Computer Graphics Forum (Proc. SCA), 2020. 2 +[96] Wenbo Wang, Hsuan-I Ho, Chen Guo, Boxiang Rong, Artur Grigorev, Jie Song, Juan Jose Zarate, and Otmar Hilliges. 4d-dress: A 4d dataset of real-world human clothing with semantic annotations. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition (CVPR), 2024. 6, 16 +[97] Guanjun Wu, Taoran Yi, Jiemin Fang, Lingxi Xie, Xiaopeng Zhang, Wei Wei, Wenyu Liu, Qi Tian, and Wang Xinggang. 4d gaussian splatting for real-time dynamic scene rendering. arXiv preprint arXiv:2310.08528, 2023. 3 +[98] Wenqi Xian, Jia-Bin Huang, Johannes Kopf, and Changil Kim. Space-time neural irradiance fields for free-viewpoint video. In Proceedings of the + +IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 9421-9431, 2021. 3 +[99] Donglai Xiang, Fabian Prada, Chenglei Wu, and Jessica Hodgins. Monoclothcap: Towards temporally coherent clothing capture from monocular rgb video. In Proceedings of International Conference on 3D Vision (3DV '20), pages 322 - 332, 2020. 2 +[100] Donglai Xiang, Fabian Prada, Timur Bagautdinov, Weipeng Xu, Yuan Dong, He Wen, Jessica Hodgins, and Chenglei Wu. Modeling clothing as a separate layer for an animatable human avatar. ACM Transactions on Graphics, 40(6):1-15, 2021. 2 +[101] Enze Xie, Wenhai Wang, Zhiding Yu, Anima Anandkumar, Jose M. Alvarez, and Ping Luo. Segformer: Simple and efficient design for semantic segmentation with transformers. CoRR, abs/2105.15203, 2021. 4 +[102] Tianyi Xie, Zeshun Zong, Yuxing Qiu, Xuan Li, Yutao Feng, Yin Yang, and Chenfanfu Jiang. Physgaussian: Physics-integrated 3d gaussians for generative dynamics. arXiv preprint arXiv:2311.12198, 2023. 5 +[103] Yuliang Xiu, Jinlong Yang, Dimitrios Tzionas, and Michael J. Black. ICON: Implicit Clothed humans Obtained from Normals. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 13296-13306, 2022. 1, 2 +[104] Yuliang Xiu, Jinlong Yang, Xu Cao, Dimitrios Tzionas, and Michael J. Black. ECON: Explicit Clothed humans Optimized via Normal integration. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), 2023. 1, 2 +[105] Qiangeng Xu, Weiyue Wang, Duygu Ceylan, Radomir Mech, and Ulrich Neumann. Disn: Deep implicit surface network for high-quality single-view 3d reconstruction. In Advances in Neural Information Processing Systems 32, pages 492-502. 2019. 2 +[106] Zhongcong Xu, Jianfeng Zhang, Jun Hao Liew, Hanshu Yan, Jia-Wei Liu, Chenxu Zhang, Jiashi Feng, and Mike Zheng Shou. Magicanimate: Temporally consistent human image animation using diffusion model. 2023. 6 +[107] Xueting Yang, Yihao Luo, Yuliang Xiu, Wei Wang, Hao Xu, and Zhaoxin Fan. D-IF: Uncertainty-aware Human Digitization via Implicit Distribution Field. In Proceedings of the IEEE/CVF international conference on computer vision, 2023. 2 +[108] Yifan Yang, Dong Liu, Shuhai Zhang, Zeshuai Deng, Zixiong Huang, and Mingkui Tan. Hilo: Detailed and robust 3d clothed human reconstruction with high- and low-frequency information of parametric models. In Proceedings of the IEEE/CVF Conference + +on Computer Vision and Pattern Recognition, pages 10671-10681, 2024. 2 +[109] Zhitao Yang, Zhongang Cai, Haiyi Mei, Shuai Liu, Zhaoxi Chen, Weiye Xiao, Yukun Wei, Zhongfei Qing, Chen Wei, Bo Dai, Wayne Wu, Chen Qian, Dahua Lin, Ziwei Liu, and Lei Yang. Synbody: Synthetic dataset with layered human models for 3d human perception and modeling. In Proceedings of the IEEE/CVF International Conference on Computer Vision (ICCV), pages 20282-20292, 2023. 6 +[110] Ziyi Yang, Xinyu Gao, Wen Zhou, Shaohui Jiao, Yuqing Zhang, and Xiaogang Jin. Deformable 3d gaussians for high-fidelity monocular dynamic scene reconstruction. arXiv preprint arXiv:2309.13101, 2023. 3 +[111] Mingqiao Ye, Martin Danelljan, Fisher Yu, and Lei Ke. Gaussian grouping: Segment and edit anything in 3d scenes. arXiv preprint arXiv:2312.00732, 2023. 4, 15 +[112] Xianggang Yu, Mutian Xu, Yidan Zhang, Haolin Liu, Chongjie Ye, Yushuang Wu, Zizheng Yan, Tianyou Liang, Guanying Chen, Shuguang Cui, and Xiaoguang Han. Mvimgnet: A large-scale dataset of multi-view images. In CVPR, 2023. 2 +[113] Zhengming Yu, Wei Cheng, Xian Liu, Wayne Wu, and Kwan-Yee Lin. Monohuman: Animatable human neural field from monocular video. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 16943-16953, 2023. 3, 6 +[114] Wentao Yuan, Zhaoyang Lv, Tanner Schmidt, and Steven Lovegrove. Star: Self-supervised tracking and reconstruction of rigid objects in motion with neural rendering. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 13144-13152, 2021. 3 +[115] Ilya Zakharkin, Kirill Mazur, Artur Grigorev, and Victor Lempitsky. Point-based modeling of human clothing. In Proceedings of the IEEE/CVF International Conference on Computer Vision (ICCV), pages 14718-14727, 2021. 2 +[116] Yan Zeng, Guoqiang Wei, Jiani Zheng, Jiaxin Zou, Yang Wei, Yuchen Zhang, and Hang Li. Make pixels dance: High-dynamic video generation. arXiv:2311.10982, 2023. 2 +[117] Zheng Zerong, Yu Tao, Liu Yebin, and Dai Qionghai. Pamir: Parametric model-conditioned implicit representation for image-based human reconstruction, 2021. 1, 2 +[118] Hongwen Zhang, Yating Tian, Yuxiang Zhang, Mengcheng Li, Liang An, Zhenan Sun, and Yebin Liu. Pymaf-x: Towards well-aligned full-body model regression from monocular images. IEEE Transac- + +tions on Pattern Analysis and Machine Intelligence, 2023. 2 +[119] Hao Zhu, Xinxin Zuo, Sen Wang, Xun Cao, and Ruigang Yang. Detailed human shape estimation from a single image by hierarchical mesh deformation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 4491-4500, 2019. 2 +[120] Shenhao Zhu, Junming Leo Chen, Zuozhuo Dai, Yinghui Xu, Xun Cao, Yao Yao, Hao Zhu, and Siyu Zhu. Champ: Controllable and consistent human image animation with 3d parametric guidance, 2024. 6 \ No newline at end of file diff --git a/2025/Disco4D_ Disentangled 4D Human Generation and Animation from a Single Image/images.zip b/2025/Disco4D_ Disentangled 4D Human Generation and Animation from a Single Image/images.zip new file mode 100644 index 0000000000000000000000000000000000000000..5bd466194b8563752e49e8c19cc061be8ff36eb1 --- /dev/null +++ b/2025/Disco4D_ Disentangled 4D Human Generation and Animation from a Single Image/images.zip @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:39fa94d8b09830d7d155ca98fc0d31e3fd95f301fcf269c86e8840274f2c7b2a +size 580338 diff --git a/2025/Disco4D_ Disentangled 4D Human Generation and Animation from a Single Image/layout.json b/2025/Disco4D_ Disentangled 4D Human Generation and Animation from a Single Image/layout.json new file mode 100644 index 0000000000000000000000000000000000000000..e18954f2b11b92b378318b587dd2fdc4b1b13187 --- /dev/null +++ b/2025/Disco4D_ Disentangled 4D Human Generation and Animation from a Single Image/layout.json @@ -0,0 +1,10168 @@ +{ + "pdf_info": [ + { + "para_blocks": [ + { + "bbox": [ + 56, + 103, + 555, + 122 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 103, + 555, + 122 + ], + "spans": [ + { + "bbox": [ + 56, + 103, + 555, + 122 + ], + "type": "text", + "content": "Disco4D: Disentangled 4D Human Generation and Animation from a Single Image" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 145, + 141, + 461, + 171 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 145, + 141, + 461, + 171 + ], + "spans": [ + { + "bbox": [ + 145, + 141, + 461, + 171 + ], + "type": "text", + "content": "Hui En Pang" + }, + { + "bbox": [ + 145, + 141, + 461, + 171 + ], + "type": "inline_equation", + "content": "^{1}" + }, + { + "bbox": [ + 145, + 141, + 461, + 171 + ], + "type": "text", + "content": ", Shuai Liu" + }, + { + "bbox": [ + 145, + 141, + 461, + 171 + ], + "type": "inline_equation", + "content": "^{3}" + }, + { + "bbox": [ + 145, + 141, + 461, + 171 + ], + "type": "text", + "content": ", Zhongang Cai" + }, + { + "bbox": [ + 145, + 141, + 461, + 171 + ], + "type": "inline_equation", + "content": "^{1,2,3}" + }, + { + "bbox": [ + 145, + 141, + 461, + 171 + ], + "type": "text", + "content": ", Lei Yang" + }, + { + "bbox": [ + 145, + 141, + 461, + 171 + ], + "type": "inline_equation", + "content": "^{2,3}" + }, + { + "bbox": [ + 145, + 141, + 461, + 171 + ], + "type": "text", + "content": ", Tianwei Zhang" + }, + { + "bbox": [ + 145, + 141, + 461, + 171 + ], + "type": "inline_equation", + "content": "^{1}" + }, + { + "bbox": [ + 145, + 141, + 461, + 171 + ], + "type": "text", + "content": ", Ziwei Liu" + }, + { + "bbox": [ + 145, + 141, + 461, + 171 + ], + "type": "inline_equation", + "content": "^{1}" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 71, + 171, + 535, + 186 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 71, + 171, + 535, + 186 + ], + "spans": [ + { + "bbox": [ + 71, + 171, + 535, + 186 + ], + "type": "inline_equation", + "content": "^{1}" + }, + { + "bbox": [ + 71, + 171, + 535, + 186 + ], + "type": "text", + "content": "S-Lab, Nanyang Technological University " + }, + { + "bbox": [ + 71, + 171, + 535, + 186 + ], + "type": "inline_equation", + "content": "^{2}" + }, + { + "bbox": [ + 71, + 171, + 535, + 186 + ], + "type": "text", + "content": "SenseTime Research " + }, + { + "bbox": [ + 71, + 171, + 535, + 186 + ], + "type": "inline_equation", + "content": "^{3}" + }, + { + "bbox": [ + 71, + 171, + 535, + 186 + ], + "type": "text", + "content": "Shanghai AI Laboratory" + } + ] + } + ], + "index": 4 + }, + { + "type": "image", + "bbox": [ + 63, + 200, + 547, + 312 + ], + "blocks": [ + { + "bbox": [ + 63, + 200, + 547, + 312 + ], + "lines": [ + { + "bbox": [ + 63, + 200, + 547, + 312 + ], + "spans": [ + { + "bbox": [ + 63, + 200, + 547, + 312 + ], + "type": "image", + "image_path": "95e2c828af253cd8c94fd8208bcc383a17ce69683a910e6599a18759da9fab83.jpg" + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 55, + 313, + 555, + 336 + ], + "lines": [ + { + "bbox": [ + 55, + 313, + 555, + 336 + ], + "spans": [ + { + "bbox": [ + 55, + 313, + 555, + 336 + ], + "type": "text", + "content": "Figure 1. Disco4D is a novel Gaussian Splatting framework for 4D disentangled human generation from a single image. Clothing and assets are modeled as Gaussians on top of the SMPL-X body, enabling dynamic animation and flexible editing." + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "image_caption" + } + ], + "index": 5 + }, + { + "bbox": [ + 152, + 341, + 200, + 353 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 152, + 341, + 200, + 353 + ], + "spans": [ + { + "bbox": [ + 152, + 341, + 200, + 353 + ], + "type": "text", + "content": "Abstract" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 54, + 367, + 297, + 571 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 54, + 367, + 297, + 571 + ], + "spans": [ + { + "bbox": [ + 54, + 367, + 297, + 571 + ], + "type": "text", + "content": "We present Disco4D, a novel Gaussian Splatting framework for 4D human generation and animation from a single image. Different from existing methods, Disco4D distinctively disentangles clothings (with Gaussian models) from the human body (with SMPL-X model), significantly enhancing the generation details and flexibility. Specifically, 1) Disco4D learns to efficiently fit the clothing Gaussians over the SMPL-X Gaussians. 2) Next, Disco4D adopts diffusion models to enhance the 3D generation process, e.g., modeling occluded parts not visible in the input image. 3) Finally, Disco4D learns an identity encoding for each clothing Gaussian to facilitate the separation and extraction of clothing assets. Furthermore, Disco4D naturally supports 4D human animation with vivid dynamics. Extensive experiments demonstrate the superiority of Disco4D on 4D human generation and animation tasks. Our code is available at https://github.com/disco-4d/Disco4D." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 55, + 597, + 135, + 609 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 597, + 135, + 609 + ], + "spans": [ + { + "bbox": [ + 55, + 597, + 135, + 609 + ], + "type": "text", + "content": "1. Introduction" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 54, + 617, + 297, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 54, + 617, + 297, + 714 + ], + "spans": [ + { + "bbox": [ + 54, + 617, + 297, + 714 + ], + "type": "text", + "content": "The development of high-fidelity 3D digital humans is increasingly important across a variety of augmented and virtual reality applications. To streamline the creation of these digital avatars from easily accessible in-the-wild images, a multitude of research efforts have been made on reconstructing 3D clothed human models from a single image [2, 33, 40-42, 81, 82, 103, 104, 104, 117]. These works predominantly focus on the simultaneous reconstruction of" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 313, + 341, + 556, + 498 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 341, + 556, + 498 + ], + "spans": [ + { + "bbox": [ + 313, + 341, + 556, + 498 + ], + "type": "text", + "content": "the human body and clothing. Unfortunately, these works have inherent limitations, and integrating them into applications that require virtual try-on or avatar customization poses significant challenges. This is primarily because the models are rendered as single-layer, non-animatable meshes where distinct attributes (e.g., hair, clothing, accessories) are merged into one continuous surface, with underlying layers completely obscured and self-contact areas inseparably connected. Such limitation complicates the re-animation and dynamic customization tasks. Existing works that perform layered reconstruction [25, 26] rely on self-rotating video inputs with extensive frames and viewpoints and involve substantial processing times." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 313, + 498, + 557, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 498, + 557, + 714 + ], + "spans": [ + { + "bbox": [ + 313, + 498, + 557, + 714 + ], + "type": "text", + "content": "To address these issues, we propose Disco4D, a novel 4D clothed human reconstruction method that distinctly separates the human body from clothing elements from a single image. It supports human animation as the 4th dimension, which cannot be realized by prior static 3D reconstruction works [2, 33, 42, 81, 82, 103, 104, 117]. To achieve this, it employs the SMPL-X [71] parametric model to represent the human body, capitalizing on its efficacy in capturing body structure and kinematics. Conversely, clothing, along with dynamic and variable elements such as hair and accessories, is represented using Gaussian models, which are able to model the large variability in clothing. By binding Gaussians to a SMPL-X model and fixing it during the training phase, Disco4D ensures the integrity of the body while focusing the learning process on the appearance aspects. To model occluded portions not visible in the input image, diffusion models are used to enhance the 3D generation process. Moreover, Disco4D includes an identity grouping mecha" + } + ] + } + ], + "index": 12 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "spans": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "type": "text", + "content": "CVF" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 145, + 0, + 496, + 37 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 145, + 0, + 496, + 37 + ], + "spans": [ + { + "bbox": [ + 145, + 0, + 496, + 37 + ], + "type": "text", + "content": "This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "text", + "content": "26331" + } + ] + } + ], + "index": 13 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 0 + }, + { + "para_blocks": [ + { + "bbox": [ + 55, + 72, + 295, + 95 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 72, + 295, + 95 + ], + "spans": [ + { + "bbox": [ + 55, + 72, + 295, + 95 + ], + "type": "text", + "content": "nism for the Gaussians, which is instrumental in maintaining the separability and individuality of each clothing asset." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 56, + 97, + 297, + 395 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 97, + 297, + 395 + ], + "spans": [ + { + "bbox": [ + 56, + 97, + 297, + 395 + ], + "type": "text", + "content": "The independent reconstruction of clothing and body offers several advantages. (1) Enhanced reconstruction fidelity. The SMPL-X body serves as a stable anchor for the clothing to conform to. By isolating the focus to learn clothing Gaussians, we achieve a more refined geometry and intricate detailing in the clothed model. (2) Fine-grained categorization and extraction of clothing items. Disco4D is able to separate clothing Gaussians into their respective categories, which is crucial for the recovery and utilization of individual clothing assets. (3) Extensive editing capabilities. Disco4D supports different editing functions, including the removal of specific items, inpainting (altering color or material), and other modifications. Such rich editing options allow for precise adjustments to individual assets without inadvertently affecting adjacent elements. This level of control is particularly beneficial in applications requiring detailed customization, such as virtual fashion design and digital content creation. (4) Improved animation capabilities. The body Gaussians adhere to the deformations dictated by the SMPL-X model, while clothing Gaussians conform to the underlying body movements but also exhibit behaviors true to their material characteristics. The disentangled deformation allows for nuanced adjustments to clothing behavior in response to complex body movements, thereby elevating the quality of clothed human animation." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 55, + 405, + 148, + 418 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 405, + 148, + 418 + ], + "spans": [ + { + "bbox": [ + 55, + 405, + 148, + 418 + ], + "type": "text", + "content": "2. Related Works" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 55, + 426, + 297, + 449 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 426, + 297, + 449 + ], + "spans": [ + { + "bbox": [ + 55, + 426, + 297, + 449 + ], + "type": "text", + "content": "Table 1 summarizes the relevant 3D/4D generation methods. We describe their details below." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 55, + 456, + 148, + 468 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 456, + 148, + 468 + ], + "spans": [ + { + "bbox": [ + 55, + 456, + 148, + 468 + ], + "type": "text", + "content": "2.1.3D Generation" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 55, + 474, + 297, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 474, + 297, + 714 + ], + "spans": [ + { + "bbox": [ + 55, + 474, + 297, + 714 + ], + "type": "text", + "content": "Single-image 3D Generation. Single-image reconstruction leverages advanced methods [45, 67] to generate 3D assets in the form of 3D point clouds or NeRF [65] from one image. While earlier efforts using auto-encoders focused on synthetic objects [12, 14, 21, 88, 93, 105], newer approaches treat the task as conditional generation, employing diffusion models [35] for 3D generation from both image and text [19, 35, 60, 64, 74, 76, 79, 90]. One-2-3-45 [59] uses 2D diffusion models [60, 87] to generate multiview images for reconstruction, while LRM [36] adopts transformer-based architecture to scale up the task on large datasets [19, 112]. Gaussian-based methods [47], particularly DreamGaussian [89] and LGM [91], offer efficient, high-resolution 3D model generation from text or images. Recently, video diffusion models have attracted significant attention due to their remarkable ability to generate intricate scenes and complex dynamics with great spatio-temporal consistency [4, 7-9, 31, 56, 116]. They are employed to generate consistent multi-view images, and then reconstruct underlying 3D assets with high quality [15]." + } + ] + } + ], + "index": 5 + }, + { + "type": "table", + "bbox": [ + 346, + 87, + 523, + 183 + ], + "blocks": [ + { + "bbox": [ + 326, + 71, + 542, + 83 + ], + "lines": [ + { + "bbox": [ + 326, + 71, + 542, + 83 + ], + "spans": [ + { + "bbox": [ + 326, + 71, + 542, + 83 + ], + "type": "text", + "content": "Table 1. 3D/4D generation methods from a single image." + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 346, + 87, + 523, + 183 + ], + "lines": [ + { + "bbox": [ + 346, + 87, + 523, + 183 + ], + "spans": [ + { + "bbox": [ + 346, + 87, + 523, + 183 + ], + "type": "table", + "html": "
MethodTypeLayeredAnimatable
LGM [91]GeneralXX
PiFU [81]Human-centricXX
DreamFusion [74]GeneralXX
DreamGaussian [89]GeneralXX
PiFU [81]Human-centricXX
D-IF [107]Human-centricXX
HiLo [108]Human-centricXX
ECON [104]Human-centricXX
SHERF [40]Human-centricX✓
Disco4DHuman-centric✓✓
", + "image_path": "2963fa4e815a0db6e0cb0f849e4b7dc57e671763a5eef8ce332a45006557f049.jpg" + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "table_body" + } + ], + "index": 7 + }, + { + "bbox": [ + 313, + 195, + 556, + 567 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 195, + 556, + 567 + ], + "spans": [ + { + "bbox": [ + 313, + 195, + 556, + 567 + ], + "type": "text", + "content": "Single-image human-centric 3D Generation. Significant research efforts have been made for 3D human reconstruction, which can be classified into the following categories. (1) Explicit-shape-based methods rely on Human Mesh Recovery (HMR) using parametric models like SMPL [62] and SMPL-X [71] to generate 3D body meshes [16, 17, 22, 24, 44, 46, 48, 49, 51, 66, 80, 118]. To account for 3D garments, several approaches incorporate offsets [99, 119] or templates, utilize deformable garment templates [6, 43], or employ non-parametric forms for clothed figures [27, 104, 115]. Despite their advancements, they face limitations in handling complex outfit variations and loose clothing due to inherent topological constraints. (2) Implicit-function-based methods utilize implicit representations like occupancy or distance fields for modeling clothed humans with complex geometries, such as loose garments. Techniques range from end-to-end regression of free-form implicit surfaces [2, 81, 82] to use of geometric priors [33, 42, 103, 104, 117] and implicit shape completion [104]. Notable works such as PIFu [81], ARCH(++) [33, 42], and PaMIR [117] can extract textured models from images, but struggle with depth ambiguities and texture inconsistencies. (3) NeRF-based methods incorporate model-based priors (i.e., SMPL-X) for accurate human reconstruction. Efforts like SHERF [40] and ELICIT [41] improve the reconstruction coherence by addressing 2D observation incompleteness leveraging appearance priors. Most of these 3D clothed human reconstruction and animation works [2, 33, 42, 81, 82, 103, 104, 117] require training on human-specific datasets, which brings another limitation on the availability of such datasets." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 313, + 570, + 557, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 570, + 557, + 714 + ], + "spans": [ + { + "bbox": [ + 313, + 570, + 557, + 714 + ], + "type": "text", + "content": "3D Clothing Modeling. Reconstructing clothing from images and videos as a separate layer over the human body poses significant challenges due to the diversity of clothing topologies. Previous efforts relied on either template meshes or implicit surface models, and required extensive, high-quality 3D data from simulations [5, 70, 83, 95] or tailored template meshes [13, 32, 73, 100]. New methods were developed [34, 43] for multi-clothing models and versatile template meshes, respectively, facilitating diverse clothing topology encoding. However, these techniques typically fall short in capturing the clothing texture and appearance. The reliance on predefined clothing style templates further con" + } + ] + } + ], + "index": 9 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "type": "text", + "content": "26332" + } + ] + } + ], + "index": 10 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 1 + }, + { + "para_blocks": [ + { + "bbox": [ + 55, + 72, + 297, + 180 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 72, + 297, + 180 + ], + "spans": [ + { + "bbox": [ + 55, + 72, + 297, + 180 + ], + "type": "text", + "content": "strains their ability to handle real-world clothing variations. Corona et al. [18] addressed these shortcomings by representing clothing layers with deep unsigned distance functions and an auto-decoder for style and cut differentiation, though this often produces overly-smooth reconstructions [18]. On the other hand, SCARF [25] and DELTA [26] significantly enhance the visual fidelity by applying NeRF to clothing layers, but require self-rotating video inputs and considerable processing times." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 55, + 195, + 144, + 206 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 195, + 144, + 206 + ], + "spans": [ + { + "bbox": [ + 55, + 195, + 144, + 206 + ], + "type": "text", + "content": "2.2. 4D Animation" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 55, + 215, + 296, + 395 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 215, + 296, + 395 + ], + "spans": [ + { + "bbox": [ + 55, + 215, + 296, + 395 + ], + "type": "text", + "content": "4D Animation. This task aims at capturing dynamic 3D scenes over time. Two primary approaches have emerged: modeling 4D scenes by adding time dimension " + }, + { + "bbox": [ + 55, + 215, + 296, + 395 + ], + "type": "inline_equation", + "content": "t" + }, + { + "bbox": [ + 55, + 215, + 296, + 395 + ], + "type": "text", + "content": " or latent codes to spatial coordinates [28, 57, 98]; combining deformation fields with static 3D scenes [20, 58, 68, 69, 75, 92, 114]. Recent efforts in explicit or hybrid representations, like planar decomposition [11, 84, 85], hash representations [94], and other innovative methods [1, 23, 29], have improved reconstruction speed and quality. Gaussian Splatting, especially, stands out for balancing efficiency with quality, with dynamic 3D Gaussians [63] and 4D Gaussian Splatting [97, 110] techniques introducing time-dependent deformations to enhance reconstructions. Notably, DreamGaussian4D [78] stands out by minimizing the optimization time while achieving high-quality 4D reconstructions." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 55, + 397, + 296, + 482 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 397, + 296, + 482 + ], + "spans": [ + { + "bbox": [ + 55, + 397, + 296, + 482 + ], + "type": "text", + "content": "Human-centric 4D Animation. Recent works leverage Gaussian-based methods [38, 50, 54, 55, 61, 77, 113] for 4D human reconstruction, requiring extensive frame sequences (50-100 frames) and/or multiple viewpoints. Currently there has not been any work on 4D layered human generation and animation from a single image or a video with few images, which will be achieved in this paper." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 55, + 499, + 138, + 514 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 499, + 138, + 514 + ], + "spans": [ + { + "bbox": [ + 55, + 499, + 138, + 514 + ], + "type": "text", + "content": "3. Methodology" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 55, + 521, + 141, + 533 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 521, + 141, + 533 + ], + "spans": [ + { + "bbox": [ + 55, + 521, + 141, + 533 + ], + "type": "text", + "content": "3.1. Preliminaries" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 55, + 542, + 296, + 628 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 542, + 296, + 628 + ], + "spans": [ + { + "bbox": [ + 55, + 542, + 296, + 628 + ], + "type": "text", + "content": "3D Gaussian Splatting employs explicit 3D Gaussian points as its primary rendering entities. A 3D Gaussian point is defined as a function " + }, + { + "bbox": [ + 55, + 542, + 296, + 628 + ], + "type": "inline_equation", + "content": "G(x) = e^{-\\frac{1}{2}(x - \\mu)^T\\Sigma^{-1}(x - \\mu)}" + }, + { + "bbox": [ + 55, + 542, + 296, + 628 + ], + "type": "text", + "content": ", where " + }, + { + "bbox": [ + 55, + 542, + 296, + 628 + ], + "type": "inline_equation", + "content": "\\mu" + }, + { + "bbox": [ + 55, + 542, + 296, + 628 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 55, + 542, + 296, + 628 + ], + "type": "inline_equation", + "content": "\\Sigma" + }, + { + "bbox": [ + 55, + 542, + 296, + 628 + ], + "type": "text", + "content": " are the spatial mean and covariance matrix, respectively. Each Gaussian is also associated with its own rotation " + }, + { + "bbox": [ + 55, + 542, + 296, + 628 + ], + "type": "inline_equation", + "content": "r" + }, + { + "bbox": [ + 55, + 542, + 296, + 628 + ], + "type": "text", + "content": ", scaling " + }, + { + "bbox": [ + 55, + 542, + 296, + 628 + ], + "type": "inline_equation", + "content": "s" + }, + { + "bbox": [ + 55, + 542, + 296, + 628 + ], + "type": "text", + "content": ", opacity " + }, + { + "bbox": [ + 55, + 542, + 296, + 628 + ], + "type": "inline_equation", + "content": "\\alpha" + }, + { + "bbox": [ + 55, + 542, + 296, + 628 + ], + "type": "text", + "content": ", a view-dependent color " + }, + { + "bbox": [ + 55, + 542, + 296, + 628 + ], + "type": "inline_equation", + "content": "c" + }, + { + "bbox": [ + 55, + 542, + 296, + 628 + ], + "type": "text", + "content": " represented by spherical harmonic coefficients " + }, + { + "bbox": [ + 55, + 542, + 296, + 628 + ], + "type": "inline_equation", + "content": "f" + }, + { + "bbox": [ + 55, + 542, + 296, + 628 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 55, + 630, + 296, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 630, + 296, + 714 + ], + "spans": [ + { + "bbox": [ + 55, + 630, + 296, + 714 + ], + "type": "text", + "content": "SMPL-X parameterization [71] is an extension of the SMPL body model [62] with face and hand, designed to capture a more accurate representation of intricate body movements. SMPL-X is defined as a function " + }, + { + "bbox": [ + 55, + 630, + 296, + 714 + ], + "type": "inline_equation", + "content": "M(\\beta, \\theta, \\psi) : \\mathbb{R}^{|\\beta| \\times |\\theta| \\times |\\psi|} \\to \\mathbb{R}^{3N}" + }, + { + "bbox": [ + 55, + 630, + 296, + 714 + ], + "type": "text", + "content": ", parametrized by the pose " + }, + { + "bbox": [ + 55, + 630, + 296, + 714 + ], + "type": "inline_equation", + "content": "\\theta \\in \\mathbb{R}^{3J}" + }, + { + "bbox": [ + 55, + 630, + 296, + 714 + ], + "type": "text", + "content": " (where " + }, + { + "bbox": [ + 55, + 630, + 296, + 714 + ], + "type": "inline_equation", + "content": "J" + }, + { + "bbox": [ + 55, + 630, + 296, + 714 + ], + "type": "text", + "content": " denotes the number of full body joints), body shape " + }, + { + "bbox": [ + 55, + 630, + 296, + 714 + ], + "type": "inline_equation", + "content": "\\beta \\in \\mathbb{R}^{|\\beta|}" + }, + { + "bbox": [ + 55, + 630, + 296, + 714 + ], + "type": "text", + "content": " and facial expression " + }, + { + "bbox": [ + 55, + 630, + 296, + 714 + ], + "type": "inline_equation", + "content": "\\psi \\in \\mathbb{R}^{|\\psi|}" + }, + { + "bbox": [ + 55, + 630, + 296, + 714 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 314, + 72, + 382, + 83 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 314, + 72, + 382, + 83 + ], + "spans": [ + { + "bbox": [ + 314, + 72, + 382, + 83 + ], + "type": "text", + "content": "3.2. Overview" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 313, + 89, + 555, + 161 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 89, + 555, + 161 + ], + "spans": [ + { + "bbox": [ + 313, + 89, + 555, + 161 + ], + "type": "text", + "content": "Given a single image, Disco4D generates animatable 3D clothed human avatars in a bottom-up manner, facilitating natural separability. Our generated 3D clothed avatars, denoted as " + }, + { + "bbox": [ + 313, + 89, + 555, + 161 + ], + "type": "inline_equation", + "content": "S_{human}" + }, + { + "bbox": [ + 313, + 89, + 555, + 161 + ], + "type": "text", + "content": ", are represented as the concatenation of " + }, + { + "bbox": [ + 313, + 89, + 555, + 161 + ], + "type": "inline_equation", + "content": "S_{body}" + }, + { + "bbox": [ + 313, + 89, + 555, + 161 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 313, + 89, + 555, + 161 + ], + "type": "inline_equation", + "content": "S_{cloth}" + }, + { + "bbox": [ + 313, + 89, + 555, + 161 + ], + "type": "text", + "content": ". Inspired by prior works [89, 91], " + }, + { + "bbox": [ + 313, + 89, + 555, + 161 + ], + "type": "inline_equation", + "content": "S" + }, + { + "bbox": [ + 313, + 89, + 555, + 161 + ], + "type": "text", + "content": " capitalizes on Gaussian representations:" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 387, + 171, + 555, + 185 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 387, + 171, + 555, + 185 + ], + "spans": [ + { + "bbox": [ + 387, + 171, + 555, + 185 + ], + "type": "interline_equation", + "content": "S = G (\\mu , r, s, \\alpha , c, e), \\tag {1}", + "image_path": "5866d6d4184a6e60442e73bf5daa95c6c19232a54865fcec71cceed46f5bfd65.jpg" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 313, + 194, + 555, + 254 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 194, + 555, + 254 + ], + "spans": [ + { + "bbox": [ + 313, + 194, + 555, + 254 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 313, + 194, + 555, + 254 + ], + "type": "inline_equation", + "content": "\\mu, r, s, \\alpha, c" + }, + { + "bbox": [ + 313, + 194, + 555, + 254 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 313, + 194, + 555, + 254 + ], + "type": "inline_equation", + "content": "e" + }, + { + "bbox": [ + 313, + 194, + 555, + 254 + ], + "type": "text", + "content": " denote positions, rotation, scaling, opacity, spherical harmonics coefficients and identity encoding, respectively. Different from traditional Gaussian representations, we add identity encoding " + }, + { + "bbox": [ + 313, + 194, + 555, + 254 + ], + "type": "inline_equation", + "content": "e" + }, + { + "bbox": [ + 313, + 194, + 555, + 254 + ], + "type": "text", + "content": " to associate each Gaussian with its clothing category." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 313, + 255, + 556, + 386 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 255, + 556, + 386 + ], + "spans": [ + { + "bbox": [ + 313, + 255, + 556, + 386 + ], + "type": "text", + "content": "Figure 2 depicts our framework. We start by generating colored SMPL-X Gaussians representing the body beneath clothing (Sec. 3.3). We obtain a visual hull for canonicalization and refine Gaussian predictions to align and envelop the SMPL-X mesh (Sec. 3.4). Next, we iteratively optimize canonical clothing Gaussians external to the SMPL-X mesh (Sec. 3.5). Lastly, we showcase the animation and editing of generated clothed avatars (Sec. 3.6). Notably, we leverage diffusion models to refine textures during 3D generation (Sec. 3.5) and extrapolate unseen views during 4D animation (Sec. 3.6)." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 314, + 394, + 429, + 405 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 314, + 394, + 429, + 405 + ], + "spans": [ + { + "bbox": [ + 314, + 394, + 429, + 405 + ], + "type": "text", + "content": "3.3. SMPL-X Gaussians" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 313, + 411, + 555, + 471 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 411, + 555, + 471 + ], + "spans": [ + { + "bbox": [ + 313, + 411, + 555, + 471 + ], + "type": "text", + "content": "Given an image, we first estimate coarse SMPL-X parameters with an off-the-shelf model [10], and then refine coarse predictions by fitting on 2D keypoints and clothing segmentation masks [72], obtaining pixel-aligned SMPL-X parameters " + }, + { + "bbox": [ + 313, + 411, + 555, + 471 + ], + "type": "inline_equation", + "content": "(\\beta, \\theta, \\psi)" + }, + { + "bbox": [ + 313, + 411, + 555, + 471 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 313, + 472, + 556, + 664 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 472, + 556, + 664 + ], + "spans": [ + { + "bbox": [ + 313, + 472, + 556, + 664 + ], + "type": "text", + "content": "Mesh Binding. To convert the SMPL-X [71] mesh " + }, + { + "bbox": [ + 313, + 472, + 556, + 664 + ], + "type": "inline_equation", + "content": "M(\\beta, \\theta, \\psi)" + }, + { + "bbox": [ + 313, + 472, + 556, + 664 + ], + "type": "text", + "content": " into Gaussians " + }, + { + "bbox": [ + 313, + 472, + 556, + 664 + ], + "type": "inline_equation", + "content": "S_{body}" + }, + { + "bbox": [ + 313, + 472, + 556, + 664 + ], + "type": "text", + "content": " for rendering, flat 3D Gaussians are bound to each mesh triangle, similar to SuGaR [30]. Gaussian means " + }, + { + "bbox": [ + 313, + 472, + 556, + 664 + ], + "type": "inline_equation", + "content": "\\mu_{body}" + }, + { + "bbox": [ + 313, + 472, + 556, + 664 + ], + "type": "text", + "content": " are computed using predefined barycentric coordinates, while Gaussian rotations " + }, + { + "bbox": [ + 313, + 472, + 556, + 664 + ], + "type": "inline_equation", + "content": "r_{body}" + }, + { + "bbox": [ + 313, + 472, + 556, + 664 + ], + "type": "text", + "content": " derive from surface normals. The initial scaling " + }, + { + "bbox": [ + 313, + 472, + 556, + 664 + ], + "type": "inline_equation", + "content": "s_{body}" + }, + { + "bbox": [ + 313, + 472, + 556, + 664 + ], + "type": "text", + "content": " ensures dense mesh coverage, with the last axis set to 0.1 for a uniformly thin surface. For color representation beneath clothing, opacity " + }, + { + "bbox": [ + 313, + 472, + 556, + 664 + ], + "type": "inline_equation", + "content": "\\alpha_{body}" + }, + { + "bbox": [ + 313, + 472, + 556, + 664 + ], + "type": "text", + "content": " is set to 1.0, with spherical harmonics " + }, + { + "bbox": [ + 313, + 472, + 556, + 664 + ], + "type": "inline_equation", + "content": "c_{body}" + }, + { + "bbox": [ + 313, + 472, + 556, + 664 + ], + "type": "text", + "content": " optimized for each Gaussian. Visible skin color is supervised, while occluded skin color aligns with visible regions. A fixed label " + }, + { + "bbox": [ + 313, + 472, + 556, + 664 + ], + "type": "inline_equation", + "content": "e_{body}" + }, + { + "bbox": [ + 313, + 472, + 556, + 664 + ], + "type": "text", + "content": " is assigned for rendering, remaining unchanged during training. When optimizing clothing Gaussians " + }, + { + "bbox": [ + 313, + 472, + 556, + 664 + ], + "type": "inline_equation", + "content": "S_{cloth}" + }, + { + "bbox": [ + 313, + 472, + 556, + 664 + ], + "type": "text", + "content": ", SMPL-X Gaussians " + }, + { + "bbox": [ + 313, + 472, + 556, + 664 + ], + "type": "inline_equation", + "content": "S_{body}" + }, + { + "bbox": [ + 313, + 472, + 556, + 664 + ], + "type": "text", + "content": " parameters stay fixed, preserving the body structure while allowing flexible learning for clothing." + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 313, + 671, + 503, + 685 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 671, + 503, + 685 + ], + "spans": [ + { + "bbox": [ + 313, + 671, + 503, + 685 + ], + "type": "text", + "content": "3.4. Initialization of Clothing Gaussians" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 313, + 689, + 555, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 689, + 555, + 713 + ], + "spans": [ + { + "bbox": [ + 313, + 689, + 555, + 713 + ], + "type": "text", + "content": "Cloth styles are diverse, making proper initialization crucial for effective clothing modeling. In synchronization with" + } + ] + } + ], + "index": 17 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "text", + "content": "26333" + } + ] + } + ], + "index": 18 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 2 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 72, + 68, + 539, + 232 + ], + "blocks": [ + { + "bbox": [ + 72, + 68, + 539, + 232 + ], + "lines": [ + { + "bbox": [ + 72, + 68, + 539, + 232 + ], + "spans": [ + { + "bbox": [ + 72, + 68, + 539, + 232 + ], + "type": "image", + "image_path": "a42620725e9459866868a1c7ce133c3e2b8b35f0d94e5e20c14ece6d76cafb14.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 54, + 234, + 555, + 312 + ], + "lines": [ + { + "bbox": [ + 54, + 234, + 555, + 312 + ], + "spans": [ + { + "bbox": [ + 54, + 234, + 555, + 312 + ], + "type": "text", + "content": "Figure 2. Framework Overview of Disco4D. (a) 3D Generation utilizes a single image to obtain disentangled body and clothing Gaussians. Body, face and hand poses are refined to be pixel-aligned. For faster initialization, clothing Gaussians and visual hull are obtained with Gaussian Reconstruction Models. These clothing Gaussians are embedded to SMPL-X mesh and adopt the local coordinate system of the triangle. Subsequently, the iterative optimization process (pruning, identity encoding and densifying) separates the body and garments. The learned identity encodings guide the densification of the clothing Gaussians. (b) 4D Animation is achieved by either direct driving of SMPL-X poses or leveraging video to learn extra clothing deformation (refer to Figure 3 for more details). Various (c) 3D/4D Editing operations can be performed with our disentangled representation." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "bbox": [ + 54, + 317, + 297, + 472 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 54, + 317, + 297, + 472 + ], + "spans": [ + { + "bbox": [ + 54, + 317, + 297, + 472 + ], + "type": "text", + "content": "estimating SMPL-X, we first employ the Video Diffusion Model [7] to estimate multi-view images. Subsequently, we leverage Gaussian Reconstruction Models [91] to obtain initial 3D Gaussians and their corresponding visual hull. Yet, the reconstructed 3D outputs often suffer from geometric inaccuracies, such as incorrect poses due to pose ambiguity or missing limbs. To address this, we refine the coarse visual hull to ensure it accurately aligns with and overlays the SMPL-X mesh and encapsulates a good geometry for the clothed figure. With SMPL-X aligned visual hull, we derive the refined Gaussians by adopting properties from their nearest neighbors. The refined visual hull and Gaussians are then canonicalized for the optimization phase." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 55, + 473, + 297, + 665 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 473, + 297, + 665 + ], + "spans": [ + { + "bbox": [ + 55, + 473, + 297, + 665 + ], + "type": "text", + "content": "Mesh embedding. Each 3D clothing Gaussian is embedded on a triangle of the canonical mesh, defining its position in both canonical and posed spaces. The mean vertex position " + }, + { + "bbox": [ + 55, + 473, + 297, + 665 + ], + "type": "inline_equation", + "content": "\\mathbf{O}" + }, + { + "bbox": [ + 55, + 473, + 297, + 665 + ], + "type": "text", + "content": " serves as the origin of the local coordinate system, with the Gaussian positioned by an offset vector " + }, + { + "bbox": [ + 55, + 473, + 297, + 665 + ], + "type": "inline_equation", + "content": "\\mathbf{v} = \\sigma \\mathbf{i} + \\beta \\mathbf{j} + \\gamma \\mathbf{k}" + }, + { + "bbox": [ + 55, + 473, + 297, + 665 + ], + "type": "text", + "content": ", where " + }, + { + "bbox": [ + 55, + 473, + 297, + 665 + ], + "type": "inline_equation", + "content": "\\sigma" + }, + { + "bbox": [ + 55, + 473, + 297, + 665 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 55, + 473, + 297, + 665 + ], + "type": "inline_equation", + "content": "\\beta" + }, + { + "bbox": [ + 55, + 473, + 297, + 665 + ], + "type": "text", + "content": ", and " + }, + { + "bbox": [ + 55, + 473, + 297, + 665 + ], + "type": "inline_equation", + "content": "\\gamma" + }, + { + "bbox": [ + 55, + 473, + 297, + 665 + ], + "type": "text", + "content": " are the components of the displacement vector along the tangent " + }, + { + "bbox": [ + 55, + 473, + 297, + 665 + ], + "type": "inline_equation", + "content": "\\mathbf{i}" + }, + { + "bbox": [ + 55, + 473, + 297, + 665 + ], + "type": "text", + "content": ", bitangent " + }, + { + "bbox": [ + 55, + 473, + 297, + 665 + ], + "type": "inline_equation", + "content": "\\mathbf{j}" + }, + { + "bbox": [ + 55, + 473, + 297, + 665 + ], + "type": "text", + "content": ", and normal " + }, + { + "bbox": [ + 55, + 473, + 297, + 665 + ], + "type": "inline_equation", + "content": "\\mathbf{k}" + }, + { + "bbox": [ + 55, + 473, + 297, + 665 + ], + "type": "text", + "content": ". Unlike SplattingAvatar [86], which displaces Gaussians along the normal, our approach allows embedding to the most suitable triangle rather than the nearest one. For example, hair Gaussians are tagged to head faces instead of the nearest face for reposing [39, 86] (Figure 9 in Appendix). In animation, the Gaussian rotates with its embedded triangle face " + }, + { + "bbox": [ + 55, + 473, + 297, + 665 + ], + "type": "inline_equation", + "content": "(\\delta r)" + }, + { + "bbox": [ + 55, + 473, + 297, + 665 + ], + "type": "text", + "content": ", while scaling " + }, + { + "bbox": [ + 55, + 473, + 297, + 665 + ], + "type": "inline_equation", + "content": "(\\delta s)" + }, + { + "bbox": [ + 55, + 473, + 297, + 665 + ], + "type": "text", + "content": " is adjusted dynamically based on changes in edge lengths. During optimization, Gaussian and embedding parameters " + }, + { + "bbox": [ + 55, + 473, + 297, + 665 + ], + "type": "inline_equation", + "content": "(\\mathbf{O}, \\mathbf{v}, \\delta r, \\text{and } \\delta s)" + }, + { + "bbox": [ + 55, + 473, + 297, + 665 + ], + "type": "text", + "content": " are jointly updated." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 55, + 672, + 253, + 685 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 672, + 253, + 685 + ], + "spans": [ + { + "bbox": [ + 55, + 672, + 253, + 685 + ], + "type": "text", + "content": "3.5. Optimization of Separable Gaussians" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 55, + 689, + 296, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 689, + 296, + 715 + ], + "spans": [ + { + "bbox": [ + 55, + 689, + 296, + 715 + ], + "type": "text", + "content": "With the SMPL-X Gaussian and initialized clothing Gaussian, we aim to optimize canonical clothing Gaussians " + }, + { + "bbox": [ + 55, + 689, + 296, + 715 + ], + "type": "inline_equation", + "content": "S_{cloth}" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 313, + 317, + 555, + 461 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 317, + 555, + 461 + ], + "spans": [ + { + "bbox": [ + 313, + 317, + 555, + 461 + ], + "type": "text", + "content": "outside the SMPL-X mesh. This involves three steps: 1) we use Signed Distance Function (SDF) loss and pruning to discourage and remove Gaussians that reside within the body; 2) we introduce identity encoding " + }, + { + "bbox": [ + 313, + 317, + 555, + 461 + ], + "type": "inline_equation", + "content": "e" + }, + { + "bbox": [ + 313, + 317, + 555, + 461 + ], + "type": "text", + "content": " to attach a clothing label for each clothing Gaussian, by lifting multi-view 2D segmentations of the target object onto the 3D Gaussians; and 3) guided by " + }, + { + "bbox": [ + 313, + 317, + 555, + 461 + ], + "type": "inline_equation", + "content": "e_{body}" + }, + { + "bbox": [ + 313, + 317, + 555, + 461 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 313, + 317, + 555, + 461 + ], + "type": "inline_equation", + "content": "e_{cloth}" + }, + { + "bbox": [ + 313, + 317, + 555, + 461 + ], + "type": "text", + "content": ", we selectively densify only the relevant clothing points while ignoring body points. Once the disentangled clothing is obtained, we use SDS loss to in-paint high-resolution texture from the reference image to individual clothing Gaussians, thereby enriching the details of unseen regions." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 313, + 460, + 556, + 581 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 460, + 556, + 581 + ], + "spans": [ + { + "bbox": [ + 313, + 460, + 556, + 581 + ], + "type": "text", + "content": "SDF Loss and Pruning. In reality, the clothing is always external to the body. During refinement, we ensure that the clothing Gaussians are positioned externally to the SMPL-X mesh by applying the SDF loss and a pruning strategy. Specifically, the SDF loss " + }, + { + "bbox": [ + 313, + 460, + 556, + 581 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{sdf}" + }, + { + "bbox": [ + 313, + 460, + 556, + 581 + ], + "type": "text", + "content": " penalizes any new densified Gaussians that intrude into the space of the SMPL-X mesh, ensuring that the clothing Gaussians consistently remain outside the body's surface. Pruning is applied at fixed intervals to reinforce this separation, and systematically remove any Gaussians located within the SDF of the SMPL-X mesh." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 313, + 580, + 556, + 677 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 580, + 556, + 677 + ], + "spans": [ + { + "bbox": [ + 313, + 580, + 556, + 677 + ], + "type": "text", + "content": "Identity encoding. To associate each Gaussian to its clothing category, we introduce Identity Encoding " + }, + { + "bbox": [ + 313, + 580, + 556, + 677 + ], + "type": "inline_equation", + "content": "(e)" + }, + { + "bbox": [ + 313, + 580, + 556, + 677 + ], + "type": "text", + "content": ", a learnable and compact vector of length 15, representing clothing categories from SegFormer [101] segmentation masks1. During training, the encodings are rendered into 2D segmentation masks in a differentiable manner following [111]. For classification, we apply a softmax to the rendered features " + }, + { + "bbox": [ + 313, + 580, + 556, + 677 + ], + "type": "inline_equation", + "content": "E_{\\mathrm{id}}" + }, + { + "bbox": [ + 313, + 580, + 556, + 677 + ], + "type": "text", + "content": " and use cross-entropy loss " + }, + { + "bbox": [ + 313, + 580, + 556, + 677 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{2d}" + }, + { + "bbox": [ + 313, + 580, + 556, + 677 + ], + "type": "text", + "content": " for " + }, + { + "bbox": [ + 313, + 580, + 556, + 677 + ], + "type": "inline_equation", + "content": "(K + 1)" + }, + { + "bbox": [ + 313, + 580, + 556, + 677 + ], + "type": "text", + "content": "-category classifica" + } + ] + } + ], + "index": 8 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 313, + 683, + 555, + 715 + ], + "type": "page_footnote", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 683, + 555, + 715 + ], + "spans": [ + { + "bbox": [ + 313, + 683, + 555, + 715 + ], + "type": "text", + "content": "1Categories: 0: \"Background\", 1: \"Hat\", 2: \"Hair\", 3: \"Sunglasses\", 4: \"Upper-clothes\", 5: \"Skirt\", 6: \"Pants\", 7: \"Dress\", 8: \"Belt\", 9: \"Left-shoe\", 10: \"Right-shoe\", 11: \"Face\", 12: \"Skin\", 13: \"Bag\", 14: \"Scarf\"" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "type": "text", + "content": "26334" + } + ] + } + ], + "index": 10 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 3 + }, + { + "para_blocks": [ + { + "bbox": [ + 55, + 72, + 296, + 216 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 72, + 296, + 216 + ], + "spans": [ + { + "bbox": [ + 55, + 72, + 296, + 216 + ], + "type": "text", + "content": "tion. An unsupervised 3D regularization loss " + }, + { + "bbox": [ + 55, + 72, + 296, + 216 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{3d}" + }, + { + "bbox": [ + 55, + 72, + 296, + 216 + ], + "type": "text", + "content": " promotes spatial consistency among the top " + }, + { + "bbox": [ + 55, + 72, + 296, + 216 + ], + "type": "inline_equation", + "content": "k" + }, + { + "bbox": [ + 55, + 72, + 296, + 216 + ], + "type": "text", + "content": "-nearest 3D Gaussians' Identity Encodings. Consequently, the overall identity loss is " + }, + { + "bbox": [ + 55, + 72, + 296, + 216 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{id} = \\mathcal{L}_{2d} + \\mathcal{L}_{3d}" + }, + { + "bbox": [ + 55, + 72, + 296, + 216 + ], + "type": "text", + "content": ". Refer to Appendix 7.2 for more details. Densification of clothing Gaussians. To learn clothing more efficiently, we perform sampling for categorical Gaussians that belong to the same clothing category and embedding. We find the " + }, + { + "bbox": [ + 55, + 72, + 296, + 216 + ], + "type": "inline_equation", + "content": "k" + }, + { + "bbox": [ + 55, + 72, + 296, + 216 + ], + "type": "text", + "content": "-nearest Gaussian points for the resampled points and inherit their Gaussian properties (scaling, rotation, opacity, SH properties). By selectively densifying clothing Gaussians, we only add necessary Gaussians while ignoring body Gaussians." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 55, + 217, + 296, + 304 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 217, + 296, + 304 + ], + "spans": [ + { + "bbox": [ + 55, + 217, + 296, + 304 + ], + "type": "text", + "content": "Anisotropy. To prevent overly-skinny kernels that point outward from the object surface under large deformations, we enforce the anisotropy of Gaussian kernels following [102]. During optimization, we employ " + }, + { + "bbox": [ + 55, + 217, + 296, + 304 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{\\mathrm{ani}} = \\frac{1}{|P|}\\sum_{p\\in P}\\max \\left(\\frac{\\max(s_p)}{\\min(s_p)},\\tau\\right) - \\tau" + }, + { + "bbox": [ + 55, + 217, + 296, + 304 + ], + "type": "text", + "content": ", where " + }, + { + "bbox": [ + 55, + 217, + 296, + 304 + ], + "type": "inline_equation", + "content": "s_p" + }, + { + "bbox": [ + 55, + 217, + 296, + 304 + ], + "type": "text", + "content": " is the scalings of 3D Gaussians. This loss constrains the ratio between the major and minor axis lengths below threshold " + }, + { + "bbox": [ + 55, + 217, + 296, + 304 + ], + "type": "inline_equation", + "content": "\\tau" + }, + { + "bbox": [ + 55, + 217, + 296, + 304 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 55, + 305, + 296, + 376 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 305, + 296, + 376 + ], + "spans": [ + { + "bbox": [ + 55, + 305, + 296, + 376 + ], + "type": "text", + "content": "Total loss. To inpaint occluded textures, we use the " + }, + { + "bbox": [ + 55, + 305, + 296, + 376 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{SDS}" + }, + { + "bbox": [ + 55, + 305, + 296, + 376 + ], + "type": "text", + "content": " loss on the Gaussians in the canonical pose after optimizing the front view for 500 steps. Combined with the conventional 3D Gaussian Loss " + }, + { + "bbox": [ + 55, + 305, + 296, + 376 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{ori}" + }, + { + "bbox": [ + 55, + 305, + 296, + 376 + ], + "type": "text", + "content": " on image rendering, the total loss " + }, + { + "bbox": [ + 55, + 305, + 296, + 376 + ], + "type": "inline_equation", + "content": "L" + }, + { + "bbox": [ + 55, + 305, + 296, + 376 + ], + "type": "text", + "content": " for end-to-end optimization of clothing Gaussians via network " + }, + { + "bbox": [ + 55, + 305, + 296, + 376 + ], + "type": "inline_equation", + "content": "C_N" + }, + { + "bbox": [ + 55, + 305, + 296, + 376 + ], + "type": "text", + "content": " is:" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 58, + 383, + 296, + 407 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 58, + 383, + 296, + 407 + ], + "spans": [ + { + "bbox": [ + 58, + 383, + 296, + 407 + ], + "type": "interline_equation", + "content": "\\mathcal {L} = \\lambda_ {\\text {o r i}} \\mathcal {L} _ {\\text {o r i}} + \\lambda_ {\\text {i d}} \\mathcal {L} _ {\\text {i d}} + \\lambda_ {\\text {a n i}} \\mathcal {L} _ {\\text {a n i}} + \\lambda_ {\\text {s d f}} \\mathcal {L} _ {\\text {s d f}} + \\lambda_ {\\text {S D S}} \\mathcal {L} _ {\\text {S D S}} \\tag {2}", + "image_path": "50e4701ff0a0acb3aa31988b2632e617973dce4958776830a00331421f31e4f2.jpg" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 55, + 410, + 241, + 423 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 410, + 241, + 423 + ], + "spans": [ + { + "bbox": [ + 55, + 410, + 241, + 423 + ], + "type": "text", + "content": "3.6. 4D Human Animation and Editing" + } + ] + } + ], + "index": 4 + }, + { + "type": "image", + "bbox": [ + 72, + 439, + 270, + 482 + ], + "blocks": [ + { + "bbox": [ + 63, + 429, + 176, + 437 + ], + "lines": [ + { + "bbox": [ + 63, + 429, + 176, + 437 + ], + "spans": [ + { + "bbox": [ + 63, + 429, + 176, + 437 + ], + "type": "text", + "content": "a) SMPL-X Pose Driven Animation" + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 72, + 439, + 270, + 482 + ], + "lines": [ + { + "bbox": [ + 72, + 439, + 270, + 482 + ], + "spans": [ + { + "bbox": [ + 72, + 439, + 270, + 482 + ], + "type": "image", + "image_path": "0f4a91071a9f36466d5bc8dae4a308e65ce2cab82b34981bcabd105fe8f64912.jpg" + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "image_body" + } + ], + "index": 6 + }, + { + "type": "image", + "bbox": [ + 67, + 494, + 286, + 650 + ], + "blocks": [ + { + "bbox": [ + 63, + 484, + 165, + 493 + ], + "lines": [ + { + "bbox": [ + 63, + 484, + 165, + 493 + ], + "spans": [ + { + "bbox": [ + 63, + 484, + 165, + 493 + ], + "type": "text", + "content": "b) Video-to-4D Reconstruction" + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 67, + 494, + 286, + 650 + ], + "lines": [ + { + "bbox": [ + 67, + 494, + 286, + 650 + ], + "spans": [ + { + "bbox": [ + 67, + 494, + 286, + 650 + ], + "type": "image", + "image_path": "18159098878650bcb17563218212498802af2f353eefda2c9c7cb61d3d527ddb.jpg" + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 55, + 654, + 296, + 720 + ], + "lines": [ + { + "bbox": [ + 55, + 654, + 296, + 720 + ], + "spans": [ + { + "bbox": [ + 55, + 654, + 296, + 720 + ], + "type": "text", + "content": "Figure 3. 4D animation is achieved by (a) driving SMPL-X poses or (b) using video to learn additional clothing deformations. From the first frame, a static 3D disentangled GS model is generated. Pose transformations deform body and clothing Gaussians, and a deformation network is optimized to capture additional clothing deformations over time." + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "image_caption" + } + ], + "index": 8 + }, + { + "bbox": [ + 313, + 72, + 555, + 156 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 72, + 555, + 156 + ], + "spans": [ + { + "bbox": [ + 313, + 72, + 555, + 156 + ], + "type": "text", + "content": "Disco4D's disentangled representation naturally supports animation and editing. The canonical Gaussians " + }, + { + "bbox": [ + 313, + 72, + 555, + 156 + ], + "type": "inline_equation", + "content": "S_{body}" + }, + { + "bbox": [ + 313, + 72, + 555, + 156 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 313, + 72, + 555, + 156 + ], + "type": "inline_equation", + "content": "S_{cloth}" + }, + { + "bbox": [ + 313, + 72, + 555, + 156 + ], + "type": "text", + "content": " enable separate deformations for clothing and body, ensuring realistic animation. Besides, individual clothing categories can be easily edited using image or text prompts. The learned clothing can be transferred to different body shapes and poses, for versatile customization." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 313, + 157, + 556, + 266 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 157, + 556, + 266 + ], + "spans": [ + { + "bbox": [ + 313, + 157, + 556, + 266 + ], + "type": "text", + "content": "Animating Gaussians. As shown in Figure 2, Disco4D enables animation of the canonical human Gaussian via two methods. Firstly, Gaussians can be directly driven using 3D SMPL-X sequences obtained from a motion database or estimated from 2D videos. Secondly, Disco4D enhances the model by learning detailed clothing dynamics from monocular videos. This disentanglement enables the focused modeling of clothing dynamics without altering the underlying human representation." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 313, + 266, + 556, + 483 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 266, + 556, + 483 + ], + "spans": [ + { + "bbox": [ + 313, + 266, + 556, + 483 + ], + "type": "text", + "content": "To extend static 3D Gaussians into dynamic 4D Gaussians, a deformation network is trained to predict changes in position, rotation, and scale of the reposed clothing Gaussians based on a timestamp, as described in DreamGaussian4D [78]. Unlike [78], which learns deformations for all Gaussians, Disco4D models body Gaussians using the SMPL-X mesh, while clothing Gaussians employ posed transformations and learned deformations. The transformation is defined as " + }, + { + "bbox": [ + 313, + 266, + 556, + 483 + ], + "type": "inline_equation", + "content": "S'' = D_N(S', t)" + }, + { + "bbox": [ + 313, + 266, + 556, + 483 + ], + "type": "text", + "content": " where " + }, + { + "bbox": [ + 313, + 266, + 556, + 483 + ], + "type": "inline_equation", + "content": "D_N" + }, + { + "bbox": [ + 313, + 266, + 556, + 483 + ], + "type": "text", + "content": " is the deformation network, " + }, + { + "bbox": [ + 313, + 266, + 556, + 483 + ], + "type": "inline_equation", + "content": "S'" + }, + { + "bbox": [ + 313, + 266, + 556, + 483 + ], + "type": "text", + "content": " is the spatial descriptions of the reposed 3D clothing Gaussian, " + }, + { + "bbox": [ + 313, + 266, + 556, + 483 + ], + "type": "inline_equation", + "content": "t" + }, + { + "bbox": [ + 313, + 266, + 556, + 483 + ], + "type": "text", + "content": " is the timestamp, and " + }, + { + "bbox": [ + 313, + 266, + 556, + 483 + ], + "type": "inline_equation", + "content": "S''" + }, + { + "bbox": [ + 313, + 266, + 556, + 483 + ], + "type": "text", + "content": " is the spatial descriptions of the deformed and reposed 3D clothing Gaussians. Following [78], the deformation model is initialized to predict zero deformation at the start of training to avoid divergence between dynamic and static models. The weights and biases of the final prediction heads are initialized to zero, and skip connections are introduced to enable gradient backpropagation." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 313, + 484, + 556, + 592 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 484, + 556, + 592 + ], + "spans": [ + { + "bbox": [ + 313, + 484, + 556, + 592 + ], + "type": "text", + "content": "To optimize the deformation field using the reference view video, we minimize the reconstruction loss " + }, + { + "bbox": [ + 313, + 484, + 556, + 592 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{Ref}" + }, + { + "bbox": [ + 313, + 484, + 556, + 592 + ], + "type": "text", + "content": " between the rendered image and video frame at each timestep. To propagate the motion from the reference view to the entire 3D model, we leverage Zero-1-to-3-XL [19] to predict the deformation of the unseen part to calculate " + }, + { + "bbox": [ + 313, + 484, + 556, + 592 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{SDS}" + }, + { + "bbox": [ + 313, + 484, + 556, + 592 + ], + "type": "text", + "content": ". Despite per-frame predictions of image diffusion models, the fixed color and opacity of static 3D Gaussians help preserve temporal consistency." + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 313, + 594, + 557, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 594, + 557, + 714 + ], + "spans": [ + { + "bbox": [ + 313, + 594, + 557, + 714 + ], + "type": "text", + "content": "Editing Clothing Gaussians. We extract the Gaussians corresponding to the specific category and edit them. This allows fine-grained editing and ensures that other Gaussians are not affected. Instead of fine-tuning all 3D Gaussians, we freeze the properties for most of the well-trained Gaussians and only adjust a small part of 3D Gaussians relevant to the target categories. For 3D object removal, we simply delete the 3D Gaussians of the editing target. For 3D object colorization by in-painting or text guidance, we reinitialize the color and tune the color (SH) parameters of the correspond-" + } + ] + } + ], + "index": 14 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "text", + "content": "26335" + } + ] + } + ], + "index": 15 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 4 + }, + { + "para_blocks": [ + { + "bbox": [ + 55, + 72, + 297, + 97 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 72, + 297, + 97 + ], + "spans": [ + { + "bbox": [ + 55, + 72, + 297, + 97 + ], + "type": "text", + "content": "ing Gaussian group, while fixing the 3D positions and other properties to preserve the learned 3D geometry." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 55, + 106, + 138, + 121 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 106, + 138, + 121 + ], + "spans": [ + { + "bbox": [ + 55, + 106, + 138, + 121 + ], + "type": "text", + "content": "4. Experiments" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 55, + 126, + 295, + 152 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 126, + 295, + 152 + ], + "spans": [ + { + "bbox": [ + 55, + 126, + 295, + 152 + ], + "type": "text", + "content": "Our detailed implementation and experiment setup can be found in Appendix 7.3." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 55, + 157, + 148, + 169 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 157, + 148, + 169 + ], + "spans": [ + { + "bbox": [ + 55, + 157, + 148, + 169 + ], + "type": "text", + "content": "4.1.3D Generation" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 55, + 175, + 297, + 606 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 175, + 297, + 606 + ], + "spans": [ + { + "bbox": [ + 55, + 175, + 297, + 606 + ], + "type": "text", + "content": "Generation and Disentanglement. Our generation and disentanglement results are presented in Figure 4 and Table 2. We assessed the disentanglement quality using the Synbody [109] and CloSe [3] datasets, rendering 30 and 110 clothed human meshes respectively from four angles and evaluating CLIP-similarity, PSNR, SSIM, and LPIPS for various poses and views within the CloSe dataset. disco4D leverages diffusion models without requiring training on human specific datasets. Therefore, we compare it with DreamGaussian [89] and LGM [91] which reconstruct 3D objects from diffusion models. Additionally, we conducted comparisons with SHERF, a human-centric baseline for evaluating novel poses and views. Figure 4 shows disco4D has higher fidelity and better geometry for body parts such as face and limbs due to the representation using SMPL-X Gaussians. It outperforms DreamGaussian and SHERF on SynBody and CloSe benchmarks. disco4D performs worse than LGM on novel views, likely due to its optimization of Gaussians in canonical space for pose generalization, compromising view-specific detail. Editing. We can edit specific clothing appearance given an image or text prompt, repose the person and transfer person characteristics. The disentanglement allows fine-grained editing and modification of individual assets without affecting other assets, and stacking multiple edits (Figure 4). User study. We conducted a user study to evaluate the generative quality of our image-to-3D Gaussians reconstruction on random in-the-wild images from SHHQ, detailed in Table 3. This study focuses on reference view consistency and overall generation quality, crucial aspects in image reconstruction tasks. We rendered 360-degree rotation videos for 25 images generated by DreamGaussian, LGM, and disco4D. We invited 43 volunteers to rate " + }, + { + "bbox": [ + 55, + 175, + 297, + 606 + ], + "type": "inline_equation", + "content": "24\\sim 27" + }, + { + "bbox": [ + 55, + 175, + 297, + 606 + ], + "type": "text", + "content": " mixed samples from these methods on image consistency and overall model quality, yielding 1080 valid scores. As shown in Table 3, disco4D was preferred, demonstrating better alignment with the original image content and superior overall quality." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 55, + 613, + 145, + 624 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 613, + 145, + 624 + ], + "spans": [ + { + "bbox": [ + 55, + 613, + 145, + 624 + ], + "type": "text", + "content": "4.2. 4D Animation" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 55, + 629, + 296, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 629, + 296, + 714 + ], + "spans": [ + { + "bbox": [ + 55, + 629, + 296, + 714 + ], + "type": "text", + "content": "Pose-Driven Animation. disco4D generates canonical Gaussians that can be animated with any pose sequence. Figure 12 in the Appendix demonstrates our animation capabilities and compares them with current SOTA 2D animation methods. Using identical inputs—a single frame and pose sequence—our approach more effectively preserves the body shape and fine details such as facial features and clothing. It" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 313, + 72, + 555, + 144 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 72, + 555, + 144 + ], + "spans": [ + { + "bbox": [ + 313, + 72, + 555, + 144 + ], + "type": "text", + "content": "surpasses Animate-Anyone [37] and Magic-Animate [106] in accurately modeling fine-grained body parts like hands and faces, and exhibits greater consistency compared to CHAMP [120]. The disentanglement feature of Disco4D further allows for direct manipulation of Clothing Gaussians, as shown in Figure 6." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 313, + 146, + 556, + 289 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 146, + 556, + 289 + ], + "spans": [ + { + "bbox": [ + 313, + 146, + 556, + 289 + ], + "type": "text", + "content": "4D Reconstruction. For the 4D-Dress Dataset [96], we evaluated 8 sequences, assessing CLIP similarity scores against ground-truth meshes and disentangled assets, along with novel view performance (PSNR, SSIM, LPIPS) from four viewpoints. Table 4 summarizes our quantitative results, benchmarking Disco4D against existing video-to-4D general GS approaches, such as DreamGaussian4D [78], as well as human-centric GS methods, including MonoHuman [113], GART [54], and GaussianAvatar [38]. We evaluate on monocular videos comprising 14 frames, captured from a limited front-view perspective, without full-body visibility across frames." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 313, + 291, + 556, + 411 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 291, + 556, + 411 + ], + "spans": [ + { + "bbox": [ + 313, + 291, + 556, + 411 + ], + "type": "text", + "content": "Disco4D outperforms MonoHuman [113], GART [54], and GaussianAvatar [38] (Table 4) as these methods reconstruct using known video information, unable to model unseen regions. Consequently, these methods cannot accurately model back views from front-facing videos, leading to artifacts in other perspectives and canonical space (see Figure 5). In contrast, Disco4D first performs reconstruction and subsequently incorporates details, such as clothing deformation, from the input frames, enabling consistent reconstruction even in unseen viewpoints." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 313, + 412, + 556, + 533 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 412, + 556, + 533 + ], + "spans": [ + { + "bbox": [ + 313, + 412, + 556, + 533 + ], + "type": "text", + "content": "While DreamGaussian4D [78] is capable of modeling back-view information, the details remain coarse. Our results demonstrate that initializing with our model from the first frame (DreamGaussian4D Disco4D-init) significantly outperforms other initialization methods (DreamGaussian4DLGM init, DreamGaussian init) in both fidelity and geometry (Table 4). Nevertheless, without incorporating human priors, DreamGaussian4D [78] still faces challenges, such as missing limbs and difficulty modeling fine details like facial features (see Figure 13 in Appendix)." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 313, + 534, + 557, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 534, + 557, + 713 + ], + "spans": [ + { + "bbox": [ + 313, + 534, + 557, + 713 + ], + "type": "text", + "content": "Reposing our canonical avatar enables us to align the body and assets accurately with the inferred postures from the source video, yielding high-quality reconstruction of faces, hands, and garments. Our reposed method surpasses DreamGaussian4D in geometry and fidelity by incorporating human priors. However, reposing alone cannot capture clothing dynamics. To address this, our disentangled approach models clothing deformations on the reposed Gaussians, guided by a diffusion model. As demonstrated in Figure 13 and Table 4, this process enhances the accuracy of clothing resemblance to the ground truth. The combination of asset repositioning and learned deformations improves modeling quality, with repositioning handling pose-driven changes and learned deformations simulating dynamic asset movements as observed in the driving video." + } + ] + } + ], + "index": 11 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 749, + 318, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 749, + 318, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 749, + 318, + 757 + ], + "type": "text", + "content": "26336" + } + ] + } + ], + "index": 12 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 5 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 56, + 93, + 553, + 160 + ], + "blocks": [ + { + "bbox": [ + 55, + 71, + 555, + 93 + ], + "lines": [ + { + "bbox": [ + 55, + 71, + 555, + 93 + ], + "spans": [ + { + "bbox": [ + 55, + 71, + 555, + 93 + ], + "type": "text", + "content": "Table 2. CLIP-embedding loss for generated humans and segmented assets, and performance (PSNR, SSIM, LPIPS) comparisons for novel poses and views on the Synbody and CloSe datasets across DreamGaussian, LGM, SHERF, and Disco4D." + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 56, + 93, + 553, + 160 + ], + "lines": [ + { + "bbox": [ + 56, + 93, + 553, + 160 + ], + "spans": [ + { + "bbox": [ + 56, + 93, + 553, + 160 + ], + "type": "table", + "html": "
MethodSynBodyCloSe
CLIPNovel ViewCLIPNovel ViewNovel Pose
All ↑Pants ↑Shirt ↑Shoes ↑PSNR ↑SSIM ↑LPIPS ↓All ↑Pants ↑Shirt ↑Shoes ↑PSNR ↑SSIM ↑LPIPS ↓PSNR ↑SSIM ↑LPIPS ↓
DreamGaussian0.7510.7150.7100.74913.1180.8830.2290.7340.6930.6740.76720.080.9390.089---
LGM0.8070.7240.7470.76012.8840.8760.2280.8290.7270.7120.77820.500.9390.077---
SHERF0.7660.6490.6360.71415.1890.8520.1890.7770.7850.7290.80118.960.9120.08315.540.8440.165
Disco4D0.8510.7840.7530.80115.6910.8480.1850.8560.8580.8100.84220.100.9180.08117.960.8510.136
", + "image_path": "f8a529b962d52f736868671611637835ae46b599aa08d4ce47c5180fad51cd45.jpg" + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "table_body" + } + ], + "index": 1 + }, + { + "type": "image", + "bbox": [ + 62, + 170, + 545, + 370 + ], + "blocks": [ + { + "bbox": [ + 62, + 170, + 545, + 370 + ], + "lines": [ + { + "bbox": [ + 62, + 170, + 545, + 370 + ], + "spans": [ + { + "bbox": [ + 62, + 170, + 545, + 370 + ], + "type": "image", + "image_path": "d4c087aa346e36287966f162e97293bec9476662da2b7b9f9bf420bcef256338.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 99, + 371, + 511, + 383 + ], + "lines": [ + { + "bbox": [ + 99, + 371, + 511, + 383 + ], + "spans": [ + { + "bbox": [ + 99, + 371, + 511, + 383 + ], + "type": "text", + "content": "Figure 4. Qualitative comparison of image generation across DreamGaussian, LGM, SHERF, and Disco4D." + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_caption" + } + ], + "index": 2 + }, + { + "type": "image", + "bbox": [ + 69, + 389, + 539, + 682 + ], + "blocks": [ + { + "bbox": [ + 69, + 389, + 539, + 682 + ], + "lines": [ + { + "bbox": [ + 69, + 389, + 539, + 682 + ], + "spans": [ + { + "bbox": [ + 69, + 389, + 539, + 682 + ], + "type": "image", + "image_path": "57fd506cb5851f16f4b2827cb1b859baacff518a16203678fe9f75a94c0c5543.jpg" + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 55, + 684, + 555, + 696 + ], + "lines": [ + { + "bbox": [ + 55, + 684, + 555, + 696 + ], + "spans": [ + { + "bbox": [ + 55, + 684, + 555, + 696 + ], + "type": "text", + "content": "Figure 5. Qualitative comparison of 4D generation between DreamGaussian4D, MonoHuman, GART, GaussianAvatar, and Disco4D." + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "image_caption" + } + ], + "index": 4 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 749, + 318, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 749, + 318, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 749, + 318, + 757 + ], + "type": "text", + "content": "26337" + } + ] + } + ], + "index": 6 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 6 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 93, + 93, + 257, + 137 + ], + "blocks": [ + { + "bbox": [ + 55, + 71, + 295, + 92 + ], + "lines": [ + { + "bbox": [ + 55, + 71, + 295, + 92 + ], + "spans": [ + { + "bbox": [ + 55, + 71, + 295, + 92 + ], + "type": "text", + "content": "Table 3. User study rates quality of generated 3D Gaussians from 1-5. The higher the better." + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 93, + 93, + 257, + 137 + ], + "lines": [ + { + "bbox": [ + 93, + 93, + 257, + 137 + ], + "spans": [ + { + "bbox": [ + 93, + 93, + 257, + 137 + ], + "type": "table", + "html": "
MetricImage Consistency ↑Overall Quality ↑
DreamGaussian2.0171.852
LGM2.3382.017
Disco4D3.1423.037
", + "image_path": "4b2da948bac6ce9f771e239d3de5f968ccd804fd06c3345ed3cba2c13270a4c0.jpg" + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "table_body" + } + ], + "index": 1 + }, + { + "type": "table", + "bbox": [ + 56, + 180, + 296, + 258 + ], + "blocks": [ + { + "bbox": [ + 55, + 144, + 297, + 177 + ], + "lines": [ + { + "bbox": [ + 55, + 144, + 297, + 177 + ], + "spans": [ + { + "bbox": [ + 55, + 144, + 297, + 177 + ], + "type": "text", + "content": "Table 4. CLIP-embedding loss for generated humans and segmented assets, and performance (PSNR, SSIM, LPIPS) comparison on 4D-Dress across various video-to-4D methods." + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 56, + 180, + 296, + 258 + ], + "lines": [ + { + "bbox": [ + 56, + 180, + 296, + 258 + ], + "spans": [ + { + "bbox": [ + 56, + 180, + 296, + 258 + ], + "type": "table", + "html": "
All ↑Assets ↑PSNR ↑SSIM ↑LPIPS ↓
DreamGaussian4D0.7840.76920.540.930.080
MonoHuman0.7620.74320.220.920.086
GART0.8000.77218.810.920.086
GaussianAvatar0.8220.76820.010.930.069
DreamGaussian4D (LGM init)0.8090.79519.160.930.086
DreamGaussian4D (Disco4D init)0.8700.84921.020.930.065
Disco4D (reposed)0.8530.77423.940.950.049
Disco4D (reposed)+learned deformations0.9000.86525.460.960.035
", + "image_path": "dbfe3cda0184f66b403e6885cc745a30396ef77df4d91d224cddfc6d86550869.jpg" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "table_body" + } + ], + "index": 3 + }, + { + "type": "image", + "bbox": [ + 58, + 259, + 293, + 397 + ], + "blocks": [ + { + "bbox": [ + 58, + 259, + 293, + 397 + ], + "lines": [ + { + "bbox": [ + 58, + 259, + 293, + 397 + ], + "spans": [ + { + "bbox": [ + 58, + 259, + 293, + 397 + ], + "type": "image", + "image_path": "465ac7c73588c24ace8492ee9eb936cb4a42e2d7e2432f4d7178a055f7c44cce.jpg" + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 55, + 400, + 296, + 423 + ], + "lines": [ + { + "bbox": [ + 55, + 400, + 296, + 423 + ], + "spans": [ + { + "bbox": [ + 55, + 400, + 296, + 423 + ], + "type": "text", + "content": "Figure 6. First frame Editing and Animation. Betas Editing, Recoloring (Text/Image-guided), Composition (Removal, Swap)." + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "image_caption" + } + ], + "index": 4 + }, + { + "bbox": [ + 55, + 434, + 296, + 507 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 434, + 296, + 507 + ], + "spans": [ + { + "bbox": [ + 55, + 434, + 296, + 507 + ], + "type": "text", + "content": "4D Editing. For a normal pipeline in character animation, editing the person in the video requires high consistency throughout all frames. For pose-driven animation methods, first frame editing and generation is required. Our method directly edits the Gaussians, which is more straightforward, fine-grained and consistent. This is seen from Figure 6." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 55, + 515, + 156, + 528 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 515, + 156, + 528 + ], + "spans": [ + { + "bbox": [ + 55, + 515, + 156, + 528 + ], + "type": "text", + "content": "4.3. Ablation Studies" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 55, + 533, + 295, + 676 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 533, + 295, + 676 + ], + "spans": [ + { + "bbox": [ + 55, + 533, + 295, + 676 + ], + "type": "text", + "content": "Initialization of Clothing Gaussians. This process is crucial for high fidelity reconstruction. As shown in Figure 8 in the Appendix, we evaluate different strategies, including random, surface, and hull-based initialization. Hull-based initialization significantly enhances the model accuracy and realism over other methods. Initialization directly on the SMPL-X surface often leads to inaccurate geometries, particularly with complex or loose garments, creating elongated, thin Gaussians and visual artifacts. In contrast, hull-based initialization captures garment details more effectively and maintains pose consistency, closely aligning with the true geometry of the clothed body." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 55, + 677, + 296, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 677, + 296, + 713 + ], + "spans": [ + { + "bbox": [ + 55, + 677, + 296, + 713 + ], + "type": "text", + "content": "Geometry of Clothing Gaussians. Figure 14 in the Appendix highlights the differences in clothing geometry between DreamGaussian [89], LGM [91] and Disco4D. In" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 313, + 72, + 555, + 216 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 72, + 555, + 216 + ], + "spans": [ + { + "bbox": [ + 313, + 72, + 555, + 216 + ], + "type": "text", + "content": "DreamGaussian, all points are confined within the body geometry, whereas in LGM, about half of the points extend beyond the SMPL-X body. Removing internal points leaves sparse, translucent representations for clothing. This sparsity suggests reliance on internal points for visual representation, failing to accurately depict the object's geometry where appearance should primarily originate from surface points. Often, clothing Gaussian points are incorrectly positioned inside the body's hull rather than on the surface. To better represent clothing geometry, Disco4D positions all clothing Gaussians externally to the SMPL-X body mesh, accurately reflecting the garment's actual physical characteristics." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 313, + 217, + 555, + 253 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 217, + 555, + 253 + ], + "spans": [ + { + "bbox": [ + 313, + 217, + 555, + 253 + ], + "type": "text", + "content": "Clothing editing. Figure 14 shows our editing results with the prompt \"Color the top pink\". disco4D allows for precise editing of the targeted clothing without affecting other areas." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 314, + 265, + 383, + 277 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 314, + 265, + 383, + 277 + ], + "spans": [ + { + "bbox": [ + 314, + 265, + 383, + 277 + ], + "type": "text", + "content": "5. Discussion" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 313, + 285, + 556, + 537 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 285, + 556, + 537 + ], + "spans": [ + { + "bbox": [ + 313, + 285, + 556, + 537 + ], + "type": "text", + "content": "Despite achieving impressive results, some failure cases still exist, as shown in Figure 7 in the Appendix. disco4D relies on robust and pixel-aligned SMPL-X estimation, which is still an unsolved problem. It occasionally fails for poor visual hull initialization. The extraction of mesh assets from clothing Gaussians using Local Density Query, as per DreamGaussian [89], currently loses fine-grained details. Enhancing the detail level of geometry derived from clothing Gaussians could bolster the utility of reconstructed assets in animation and simulation applications. Furthermore, the initialized visual hulls obtained from multi-view SMPL-X guided images are often of suboptimal quality and suffer from poor side and back views, necessitating refinement. Improving pose guidance models to achieve more accurate visual hulls could alleviate the need for extensive refinement. In addition, future works could look into modeling multi-layered clothing and reconstructing the occluded clothing. disco4D has many positive applications, but it also has the potential to facilitate deepfake avatars and raise IP concerns. Regulations should be built to address these issues alongside its benefits in the entertainment industry." + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 314, + 549, + 388, + 562 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 314, + 549, + 388, + 562 + ], + "spans": [ + { + "bbox": [ + 314, + 549, + 388, + 562 + ], + "type": "text", + "content": "6. Conclusion" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 313, + 570, + 556, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 570, + 556, + 713 + ], + "spans": [ + { + "bbox": [ + 313, + 570, + 556, + 713 + ], + "type": "text", + "content": "We propose Disco4D, a novel approach for the generation of 3D animatable clothed human Gaussians from a single image, emphasizing high-fidelity detail and separation of assets. We manage to compositionally generate separate components, such as haircut, accessories, and decoupled outfits. Our core insight is the fixing of SMPL-X Gaussians, fitting segmented Gaussians over SMPL-X Gaussians, and application of diffusion models to enhance 3D reconstruction, including modeling occluded parts not visible in the input image. Its capability to separate assets offers significant advantages, including localized, fine-grained editing of individual assets and enhanced animatability." + } + ] + } + ], + "index": 15 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 749, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 749, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 749, + 317, + 757 + ], + "type": "text", + "content": "26338" + } + ] + } + ], + "index": 16 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 7 + }, + { + "para_blocks": [ + { + "bbox": [ + 56, + 72, + 158, + 85 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 72, + 158, + 85 + ], + "spans": [ + { + "bbox": [ + 56, + 72, + 158, + 85 + ], + "type": "text", + "content": "Acknowledgements" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 55, + 92, + 297, + 213 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 92, + 297, + 213 + ], + "spans": [ + { + "bbox": [ + 55, + 92, + 297, + 213 + ], + "type": "text", + "content": "We sincerely thank the anonymous reviewers for their valuable comments on this paper. This research/project is supported by the National Research Foundation, Singapore under its AI Singapore Programme(AISG Award No: AISG3-PhD-2023-08-049T). This study is supported by the Ministry of Education, Singapore, under its MOE AcRF Tier 2 (MOET2EP20221-0012, MOE-T2EP20223-0002), and under the RIE2020 Industry Alignment Fund - Industry Collaboration Projects (IAF-ICP) Funding Initiative, as well as cash and in-kind contribution from the industry partner(s)." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 56, + 225, + 115, + 238 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 225, + 115, + 238 + ], + "spans": [ + { + "bbox": [ + 56, + 225, + 115, + 238 + ], + "type": "text", + "content": "References" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 66, + 245, + 298, + 714 + ], + "type": "list", + "angle": 0, + "index": 11, + "blocks": [ + { + "bbox": [ + 67, + 245, + 296, + 293 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 67, + 245, + 296, + 293 + ], + "spans": [ + { + "bbox": [ + 67, + 245, + 296, + 293 + ], + "type": "text", + "content": "[1] Jad Abou-Chakra, Feras Dayoub, and Niko Sunderhauf. Particlenerf: Particle based encoding for online neural radiance fields. arXiv preprint arXiv:2211.04041, 2022.3" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 66, + 295, + 296, + 366 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 66, + 295, + 296, + 366 + ], + "spans": [ + { + "bbox": [ + 66, + 295, + 296, + 366 + ], + "type": "text", + "content": "[2] Thiemo Alldieck, Mihai Zanfir, and Cristian Sminchisescu. Photorealistic Monocular 3D Reconstruction of Humans Wearing Clothing. Proceedings of the IEEE Computer Society Conference on Computer Vision and Pattern Recognition, 2022-June:1496-1505, 2022. 1, 2" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 67, + 369, + 298, + 427 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 67, + 369, + 298, + 427 + ], + "spans": [ + { + "bbox": [ + 67, + 369, + 298, + 427 + ], + "type": "text", + "content": "[3] Dimitrije Antić, Garvita Tiwari, Batuhan Ozcomlekci, Riccardo Marin, and Gerard Pons-Moll. CloSe: A 3D clothing segmentation dataset and model. In International Conference on 3D Vision (3DV), 2024. 6" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 67, + 430, + 296, + 502 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 67, + 430, + 296, + 502 + ], + "spans": [ + { + "bbox": [ + 67, + 430, + 296, + 502 + ], + "type": "text", + "content": "[4] Omer Bar-Tal, Hila Chefer, Omer Tov, Charles Herrmann, Roni Paiss, Shiran Zada, Ariel Ephrat, Junhwa Hur, Guanghui Liu, Amit Raj, Yanzhen Li, Michael Rubinstein, Tomer Michaeli, Oliver Wang, Deqing Sun, Tali Dekel, and Inbar Mosseri. Lumiere: A spacetime diffusion model for video generation, 2024. 2" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 67, + 505, + 296, + 550 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 67, + 505, + 296, + 550 + ], + "spans": [ + { + "bbox": [ + 67, + 505, + 296, + 550 + ], + "type": "text", + "content": "[5] Hugo Bertiche, Meysam Madadi, and Sergio Escalera. Cloth3d: clothed 3d humans. In European Conference on Computer Vision, pages 344-359. Springer, 2020. 2" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 67, + 554, + 296, + 613 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 67, + 554, + 296, + 613 + ], + "spans": [ + { + "bbox": [ + 67, + 554, + 296, + 613 + ], + "type": "text", + "content": "[6] Bharat Lal Bhatnagar, Garvita Tiwari, Christian Theobalt, and Gerard Pons-Moll. Multi-garment net: Learning to dress 3d people from images. In IEEE International Conference on Computer Vision (ICCV). IEEE, 2019. 2" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 67, + 616, + 296, + 687 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 67, + 616, + 296, + 687 + ], + "spans": [ + { + "bbox": [ + 67, + 616, + 296, + 687 + ], + "type": "text", + "content": "[7] Andreas Blattmann, Tim Dockhorn, Sumith Kulal, Daniel Mendelevitch, Maciej Kilian, Dominik Lorenz, Yam Levi, Zion English, Vikram Voleti, Adam Letts, Varun Jampani, and Robin Rombach. Stable video diffusion: Scaling latent video diffusion models to large datasets, 2023. 2, 4" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 67, + 689, + 296, + 714 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 67, + 689, + 296, + 714 + ], + "spans": [ + { + "bbox": [ + 67, + 689, + 296, + 714 + ], + "type": "text", + "content": "[8] Andreas Blattmann, Robin Rombach, Huan Ling, Tim Dockhorn, Seung Wook Kim, Sanja Fidler, and" + } + ] + } + ], + "index": 10 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 320, + 72, + 555, + 714 + ], + "type": "list", + "angle": 0, + "index": 24, + "blocks": [ + { + "bbox": [ + 315, + 72, + 555, + 95 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 315, + 72, + 555, + 95 + ], + "spans": [ + { + "bbox": [ + 315, + 72, + 555, + 95 + ], + "type": "text", + "content": "Karsten Kreis. Align your latents: High-resolution video synthesis with latent diffusion models, 2023." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 325, + 98, + 555, + 156 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 325, + 98, + 555, + 156 + ], + "spans": [ + { + "bbox": [ + 325, + 98, + 555, + 156 + ], + "type": "text", + "content": "[9] Tim Brooks, Bill Peebles, Connor Holmes, Will DePue, Yufei Guo, Li Jing, David Schnurr, Joe Taylor, Troy Luhman, Eric Luhman, Clarence Ng, Ricky Wang, and Aditya Ramesh. Video generation models as world simulators. 2024. 2" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 320, + 158, + 555, + 228 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 320, + 158, + 555, + 228 + ], + "spans": [ + { + "bbox": [ + 320, + 158, + 555, + 228 + ], + "type": "text", + "content": "[10] Zhongang Cai, Wanqi Yin, Ailing Zeng, Chen Wei, Qingping Sun, Wang Yanjun, Hui En Pang, Haiyi Mei, Mingyuan Zhang, Lei Zhang, et al. Smpler-x: Scaling up expressive human pose and shape estimation. Advances in Neural Information Processing Systems, 36, 2024. 3" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 321, + 230, + 555, + 253 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 321, + 230, + 555, + 253 + ], + "spans": [ + { + "bbox": [ + 321, + 230, + 555, + 253 + ], + "type": "text", + "content": "[11] Ang Cao and Justin Johnson. Hexplane: A fast representation for dynamic scenes. CVPR, 2023. 3" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 321, + 255, + 555, + 350 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 321, + 255, + 555, + 350 + ], + "spans": [ + { + "bbox": [ + 321, + 255, + 555, + 350 + ], + "type": "text", + "content": "[12] Angel X. Chang, Thomas Funkhouser, Leonidas Guibas, Pat Hanrahan, Qixing Huang, Zimo Li, Silvio Savarese, Manolis Savva, Shuran Song, Hao Su, Jianxiong Xiao, Li Yi, and Fisher Yu. ShapeNet: An Information-Rich 3D Model Repository. Technical Report arXiv:1512.03012 [cs.GR], Stanford University — Princeton University — Toyota Technological Institute at Chicago, 2015. 2" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 321, + 351, + 554, + 398 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 321, + 351, + 554, + 398 + ], + "spans": [ + { + "bbox": [ + 321, + 351, + 554, + 398 + ], + "type": "text", + "content": "[13] Xin Chen, Anqi Pang, Yang Wei, Wang Peihao, Lan Xu, and Jingyi Yu. Tightcap: 3d human shape capture with clothing tightness field. ACM Transactions on Graphics (Presented at ACM SIGGRAPH), 2021. 2" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 321, + 399, + 555, + 458 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 321, + 399, + 555, + 458 + ], + "spans": [ + { + "bbox": [ + 321, + 399, + 555, + 458 + ], + "type": "text", + "content": "[14] Zhiqin Chen, Andrea Tagliasacchi, and Hao Zhang. BSP-net: Generating compact meshes via binary space partitioning. Proceedings of IEEE Conference on Computer Vision and Pattern Recognition (CVPR), 2020. 2" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 321, + 460, + 554, + 495 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 321, + 460, + 554, + 495 + ], + "spans": [ + { + "bbox": [ + 321, + 460, + 554, + 495 + ], + "type": "text", + "content": "[15] Zilong Chen, Yikai Wang, Feng Wang, Zhengyi Wang, and Huaping Liu. V3d: Video diffusion models are effective 3d generators, 2024. 2" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 321, + 497, + 555, + 578 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 321, + 497, + 555, + 578 + ], + "spans": [ + { + "bbox": [ + 321, + 497, + 555, + 578 + ], + "type": "text", + "content": "[16] Hongsuk Choi, Gyeongsik Moon, and Kyoung Mu Lee. Pose2Mesh: Graph Convolutional Network for 3D Human Pose and Mesh Recovery from a 2D Human Pose. Lecture Notes in Computer Science (including subseries Lecture Notes in Artificial Intelligence and Lecture Notes in Bioinformatics), 12352 LNCS: 769-787, 2020. 2" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 321, + 581, + 555, + 639 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 321, + 581, + 555, + 639 + ], + "spans": [ + { + "bbox": [ + 321, + 581, + 555, + 639 + ], + "type": "text", + "content": "[17] Vasileios Choutas, Georgios Pavlakos, Timo Bolkart, Dimitrios Tzionas, and Michael J. Black. Monocular expressive body regression through body-driven attention. In European Conference on Computer Vision (ECCV), 2020. 2" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 321, + 641, + 555, + 689 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 321, + 641, + 555, + 689 + ], + "spans": [ + { + "bbox": [ + 321, + 641, + 555, + 689 + ], + "type": "text", + "content": "[18] Enric Corona, Albert Pumarola, Guillem Alenyà, Gerard Pons-Moll, and Francesc Moreno-Noguer. Simplicit: Topology-aware generative model for clothed people. In CVPR, 2021. 3" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 321, + 689, + 554, + 714 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 321, + 689, + 554, + 714 + ], + "spans": [ + { + "bbox": [ + 321, + 689, + 554, + 714 + ], + "type": "text", + "content": "[19] Matt Deitke, Ruoshi Liu, Matthew Wallingford, Huong Ngo, Oscar Michel, Aditya Kusupati, Alan" + } + ] + } + ], + "index": 23 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 318, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 318, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 318, + 757 + ], + "type": "text", + "content": "26339" + } + ] + } + ], + "index": 25 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 8 + }, + { + "para_blocks": [ + { + "bbox": [ + 62, + 72, + 296, + 713 + ], + "type": "list", + "angle": 0, + "index": 12, + "blocks": [ + { + "bbox": [ + 82, + 72, + 296, + 143 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 82, + 72, + 296, + 143 + ], + "spans": [ + { + "bbox": [ + 82, + 72, + 296, + 143 + ], + "type": "text", + "content": "Fan, Christian Laforte, Vikram Voleti, Samir Yitzhak Gadre, Eli VanderBilt, Aniruddha Kembhavi, Carl Vondrick, Georgia Gkioxari, Kiana Ehsani, Ludwig Schmidt, and Ali Farhadi. Objverse-xl: A universe of " + }, + { + "bbox": [ + 82, + 72, + 296, + 143 + ], + "type": "inline_equation", + "content": "10\\mathrm{m} + 3\\mathrm{d}" + }, + { + "bbox": [ + 82, + 72, + 296, + 143 + ], + "type": "text", + "content": " objects. arXiv preprint arXiv:2307.05663, 2023. 2, 5" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 62, + 145, + 296, + 204 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 145, + 296, + 204 + ], + "spans": [ + { + "bbox": [ + 62, + 145, + 296, + 204 + ], + "type": "text", + "content": "[20] Yilun Du, Yinan Zhang, Hong-Xing Yu, Joshua B. Tenenbaum, and Jiajun Wu. Neural radiance flow for 4d view synthesis and video processing. In Proceedings of the IEEE/CVF International Conference on Computer Vision, 2021. 3" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 62, + 206, + 296, + 239 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 206, + 296, + 239 + ], + "spans": [ + { + "bbox": [ + 62, + 206, + 296, + 239 + ], + "type": "text", + "content": "[21] Shivam Duggal and Deepak Pathak. Topologically-aware deformation fields for single-view 3d reconstruction. CVPR, 2022. 2" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 62, + 243, + 296, + 312 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 243, + 296, + 312 + ], + "spans": [ + { + "bbox": [ + 62, + 243, + 296, + 312 + ], + "type": "text", + "content": "[22] Sai Kumar Dwivedi, Nikos Athanasiou, Muhammed Kocabas, and Michael J. Black. Learning to Regress Bodies from Images using Differentiable Semantic Rendering. In IEEE/CVF International Conference on Computer Vision (ICCV), pages 11230-11239, 2021. 2" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 62, + 315, + 296, + 373 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 315, + 296, + 373 + ], + "spans": [ + { + "bbox": [ + 62, + 315, + 296, + 373 + ], + "type": "text", + "content": "[23] Jiemin Fang, Taoran Yi, Xinggang Wang, Lingxi Xie, Xiaopeng Zhang, Wenyu Liu, Matthias Nießner, and Qi Tian. Fast dynamic radiance fields with time-aware neural voxels. In SIGGRAPH Asia 2022 Conference Papers. ACM, 2022. 3" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 62, + 375, + 296, + 432 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 375, + 296, + 432 + ], + "spans": [ + { + "bbox": [ + 62, + 375, + 296, + 432 + ], + "type": "text", + "content": "[24] Yao Feng, Vasileios Choutas, Timo Bolkart, Dimitrios Tzionas, and Michael J. Black. Collaborative regression of expressive bodies using moderation. In International Conference on 3D Vision (3DV), 2021." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 62, + 436, + 296, + 483 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 436, + 296, + 483 + ], + "spans": [ + { + "bbox": [ + 62, + 436, + 296, + 483 + ], + "type": "text", + "content": "[25] Yao Feng, Jinlong Yang, Marc Pollefeys, Michael J. Black, and Timo Bolkart. Capturing and animation of body and clothing from monocular video. In SIGGRAPH Asia 2022 Conference Papers, 2022. 1, 3" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 62, + 484, + 296, + 529 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 484, + 296, + 529 + ], + "spans": [ + { + "bbox": [ + 62, + 484, + 296, + 529 + ], + "type": "text", + "content": "[26] Yao Feng, Weiyang Liu, Timo Bolkart, Jinlong Yang, Marc Pollefeys, and Michael J. Black. Learning disentangled avatars with hybrid 3d representations. arXiv, 2023. 1, 3" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 62, + 533, + 296, + 579 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 533, + 296, + 579 + ], + "spans": [ + { + "bbox": [ + 62, + 533, + 296, + 579 + ], + "type": "text", + "content": "[27] Valentin Gabeur, Jean-Sebastien Franco, Xavier Martin, Cordelia Schmid, and Gregory Rogez. Moulding humans: Non-parametric 3d human shape estimation from single images, 2019. 2" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 62, + 581, + 296, + 628 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 581, + 296, + 628 + ], + "spans": [ + { + "bbox": [ + 62, + 581, + 296, + 628 + ], + "type": "text", + "content": "[28] Chen Gao, Ayush Saraf, Johannes Kopf, and Jia-Bin Huang. Dynamic view synthesis from dynamic monocular video. In Proceedings of the IEEE International Conference on Computer Vision, 2021. 3" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 62, + 630, + 296, + 665 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 630, + 296, + 665 + ], + "spans": [ + { + "bbox": [ + 62, + 630, + 296, + 665 + ], + "type": "text", + "content": "[29] Shanyan Guan, Huayu Deng, Yunbo Wang, and Xiaokang Yang. Neurofluid: Fluid dynamics grounding with particle-driven neural radiance fields, 2022. 3" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 62, + 666, + 296, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 666, + 296, + 713 + ], + "spans": [ + { + "bbox": [ + 62, + 666, + 296, + 713 + ], + "type": "text", + "content": "[30] Antoine Guédon and Vincent Lepetit. Sugar: Surface-aligned gaussian splatting for efficient 3d mesh reconstruction and high-quality mesh rendering. arXiv preprint arXiv:2311.12775, 2023. 3" + } + ] + } + ], + "index": 11 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 321, + 72, + 554, + 713 + ], + "type": "list", + "angle": 0, + "index": 24, + "blocks": [ + { + "bbox": [ + 321, + 72, + 554, + 132 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 321, + 72, + 554, + 132 + ], + "spans": [ + { + "bbox": [ + 321, + 72, + 554, + 132 + ], + "type": "text", + "content": "[31] Yuwei Guo, Ceyuan Yang, Anyi Rao, Zhengyang Liang, Yaohui Wang, Yu Qiao, Maneesh Agrawala, Dahua Lin, and Bo Dai. Animatediff: Animate your personalized text-to-image diffusion models without specific tuning, 2023. 2" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 321, + 135, + 554, + 193 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 321, + 135, + 554, + 193 + ], + "spans": [ + { + "bbox": [ + 321, + 135, + 554, + 193 + ], + "type": "text", + "content": "[32] Oshri Halimi, Fabian Prada, Tuur Stuyck, Donglai Xiang, Timur Bagautdinov, He Wen, Ron Kimmel, Takaaki Shiratori, Chenglei Wu, and Yaser Sheikh. Garment avatars: Realistic cloth driving using pattern registration, 2022. 2" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 321, + 196, + 554, + 255 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 321, + 196, + 554, + 255 + ], + "spans": [ + { + "bbox": [ + 321, + 196, + 554, + 255 + ], + "type": "text", + "content": "[33] Tong He, Yuanlu Xu, Shunsuke Saito, Stefano Soatto, and Tony Tung. " + }, + { + "bbox": [ + 321, + 196, + 554, + 255 + ], + "type": "inline_equation", + "content": "\\mathsf{ARCH} + +" + }, + { + "bbox": [ + 321, + 196, + 554, + 255 + ], + "type": "text", + "content": ": Animation-Ready Clothed Human Reconstruction Revisited. Proceedings of the IEEE International Conference on Computer Vision, pages 11026–11036, 2021. 1, 2" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 321, + 258, + 554, + 329 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 321, + 258, + 554, + 329 + ], + "spans": [ + { + "bbox": [ + 321, + 258, + 554, + 329 + ], + "type": "text", + "content": "[34] Zhu Heming, Cao Yu, Jin Hang, Chen Weikai, Du Dong, Wang Zhangye, Cui Shuguang, and Han Xiaoguang. Deep fashion3d: A dataset and benchmark for 3d garment reconstruction from single images. In Computer Vision - ECCV 2020, pages 512-530. Springer International Publishing, 2020. 2" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 321, + 332, + 554, + 365 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 321, + 332, + 554, + 365 + ], + "spans": [ + { + "bbox": [ + 321, + 332, + 554, + 365 + ], + "type": "text", + "content": "[35] Jonathan Ho, Ajay Jain, and Pieter Abbeel. Denoising diffusion probabilistic models. arXiv preprint arxiv:2006.11239, 2020. 2" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 321, + 369, + 554, + 427 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 321, + 369, + 554, + 427 + ], + "spans": [ + { + "bbox": [ + 321, + 369, + 554, + 427 + ], + "type": "text", + "content": "[36] Yicong Hong, Kai Zhang, Jiaxiang Gu, Sai Bi, Yang Zhou, Difan Liu, Feng Liu, Kalyan Sunkavalli, Trung Bui, and Hao Tan. Lrm: Large reconstruction model for single image to 3d. arXiv preprint arXiv:2311.04400, 2023. 2" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 321, + 431, + 554, + 479 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 321, + 431, + 554, + 479 + ], + "spans": [ + { + "bbox": [ + 321, + 431, + 554, + 479 + ], + "type": "text", + "content": "[37] Li Hu, Xin Gao, Peng Zhang, Ke Sun, Bang Zhang, and Liefeng Bo. Animate anyone: Consistent and controllable image-to-video synthesis for character animation. arXiv preprint arXiv:2311.17117, 2023. 6" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 321, + 481, + 554, + 563 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 321, + 481, + 554, + 563 + ], + "spans": [ + { + "bbox": [ + 321, + 481, + 554, + 563 + ], + "type": "text", + "content": "[38] Liangxiao Hu, Hongwen Zhang, Yuxiang Zhang, Boyao Zhou, Boning Liu, Shengping Zhang, and Liqiang Nie. Gaussianavatar: Towards realistic human avatar modeling from a single video via animatable 3d gaussians. In IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), 2024. 3, 6" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 321, + 567, + 554, + 601 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 321, + 567, + 554, + 601 + ], + "spans": [ + { + "bbox": [ + 321, + 567, + 554, + 601 + ], + "type": "text", + "content": "[39] Shoukang Hu and Ziwei Liu. Gauhuman: Articulated gaussian splatting from monocular human videos. arXiv preprint arXiv:, 2023. 4" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 321, + 605, + 554, + 651 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 321, + 605, + 554, + 651 + ], + "spans": [ + { + "bbox": [ + 321, + 605, + 554, + 651 + ], + "type": "text", + "content": "[40] Shoukang Hu, Fangzhou Hong, Liang Pan, Haiyi Mei, Lei Yang, and Ziwei Liu. Sherf: Generalizable human nerf from a single image. arXiv preprint arXiv:2303.12791, 2023. 1, 2" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 321, + 654, + 554, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 321, + 654, + 554, + 713 + ], + "spans": [ + { + "bbox": [ + 321, + 654, + 554, + 713 + ], + "type": "text", + "content": "[41] Yangyi Huang, Hongwei Yi, Weiyang Liu, Haofan Wang, Boxi Wu, Wenxiao Wang, Binbin Lin, Debing Zhang, and Deng Cai. One-shot implicit animatable avatars with model-based priors. In IEEE Conference on Computer Vision (ICCV), 2023. 2" + } + ] + } + ], + "index": 23 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 318, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 318, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 318, + 757 + ], + "type": "text", + "content": "26340" + } + ] + } + ], + "index": 25 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 9 + }, + { + "para_blocks": [ + { + "bbox": [ + 61, + 72, + 297, + 713 + ], + "type": "list", + "angle": 0, + "index": 12, + "blocks": [ + { + "bbox": [ + 61, + 72, + 297, + 133 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 61, + 72, + 297, + 133 + ], + "spans": [ + { + "bbox": [ + 61, + 72, + 297, + 133 + ], + "type": "text", + "content": "[42] Zeng Huang, Yuanlu Xu, Christoph Lassner, Hao Li, and Tony Tung. ARCH: Animatable Reconstruction of Clothed Humans. Proceedings of the IEEE Computer Society Conference on Computer Vision and Pattern Recognition, pages 3090-3099, 2020. 1, 2" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 61, + 134, + 296, + 182 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 61, + 134, + 296, + 182 + ], + "spans": [ + { + "bbox": [ + 61, + 134, + 296, + 182 + ], + "type": "text", + "content": "[43] Boyi Jiang, Juyong Zhang, Yang Hong, Jinhao Luo, Ligang Liu, and Hujun Bao. BCnet: Learning body and cloth shape from a single image. In European Conference on Computer Vision. Springer, 2020. 2" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 62, + 183, + 296, + 243 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 183, + 296, + 243 + ], + "spans": [ + { + "bbox": [ + 62, + 183, + 296, + 243 + ], + "type": "text", + "content": "[44] Hanbyul Joo, Natalia Neverova, and Andrea Vedaldi. Exemplar Fine-Tuning for 3D Human Model Fitting Towards In-the-Wild 3D Human Pose Estimation. Proceedings - 2021 International Conference on 3D Vision, 3DV 2021, pages 42-52, 2021. 2" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 63, + 244, + 295, + 269 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 63, + 244, + 295, + 269 + ], + "spans": [ + { + "bbox": [ + 63, + 244, + 295, + 269 + ], + "type": "text", + "content": "[45] Heewoo Jun and Alex Nichol. Shap-e: Generating conditional 3d implicit functions, 2023. 2" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 62, + 270, + 296, + 330 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 270, + 296, + 330 + ], + "spans": [ + { + "bbox": [ + 62, + 270, + 296, + 330 + ], + "type": "text", + "content": "[46] Angjoo Kanazawa, Michael J Black, David W Jacobs, and Jitendra Malik. End-to-End Recovery of Human Shape and Pose. In Proceedings of the IEEE Computer Society Conference on Computer Vision and Pattern Recognition, pages 7122–7131, 2018. 2" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 62, + 332, + 296, + 380 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 332, + 296, + 380 + ], + "spans": [ + { + "bbox": [ + 62, + 332, + 296, + 380 + ], + "type": "text", + "content": "[47] Bernhard Kerbl, Georgios Kopanas, Thomas Leimkuhler, and George Drettakis. 3d gaussian splatting for real-time radiance field rendering. ACM Transactions on Graphics, 42(4), 2023. 2, 15" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 62, + 382, + 296, + 441 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 382, + 296, + 441 + ], + "spans": [ + { + "bbox": [ + 62, + 382, + 296, + 441 + ], + "type": "text", + "content": "[48] Muhammed Kocabas, Chun-Hao P. Huang, Otmar Hilliges, and Michael J. Black. PARE: Part attention regressor for 3D human body estimation. In Proc. International Conference on Computer Vision (ICCV), pages 11127-11137, 2021. 2" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 62, + 444, + 296, + 514 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 444, + 296, + 514 + ], + "spans": [ + { + "bbox": [ + 62, + 444, + 296, + 514 + ], + "type": "text", + "content": "[49] Muhammed Kocabas, Chun-Hao P. Huang, Joachim Tesch, Lea Müller, Otmar Hilliges, and Michael J. Black. SPEC: Seeing people in the wild with an estimated camera. In Proc. International Conference on Computer Vision (ICCV), pages 11035-11045, 2021. 2" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 62, + 517, + 296, + 552 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 517, + 296, + 552 + ], + "spans": [ + { + "bbox": [ + 62, + 517, + 296, + 552 + ], + "type": "text", + "content": "[50] Muhammed Kocabas, Rick Chang, James Gabriel, Oncel Tuzel, and Anurag Ranjan. Hugs: Human gaussian splats, 2023. 3" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 62, + 555, + 296, + 614 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 555, + 296, + 614 + ], + "spans": [ + { + "bbox": [ + 62, + 555, + 296, + 614 + ], + "type": "text", + "content": "[51] Nikos Kolotouros, Georgios Pavlakos, Michael Black, and Kostas Daniilidis. Learning to reconstruct 3D human pose and shape via model-fitting in the loop. Proceedings of the IEEE International Conference on Computer Vision, 2019-Octob:2252-2261, 2019. 2" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 62, + 616, + 296, + 675 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 616, + 296, + 675 + ], + "spans": [ + { + "bbox": [ + 62, + 616, + 296, + 675 + ], + "type": "text", + "content": "[52] Georgios Kopanas, Julien Philip, Thomas Leimkuhler, and George Drettakis. Point-based neural rendering with per-view optimization. Computer Graphics Forum (Proceedings of the Eurographics Symposium on Rendering), 40(4), 2021. 15" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 62, + 677, + 296, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 677, + 296, + 713 + ], + "spans": [ + { + "bbox": [ + 62, + 677, + 296, + 713 + ], + "type": "text", + "content": "[53] Georgios Kopanas, Thomas Leimkuhler, Gilles Rainer, Clément Jambon, and George Drettakis. Neural point catauastics for novel-view synthesis of re" + } + ] + } + ], + "index": 11 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 320, + 72, + 555, + 713 + ], + "type": "list", + "angle": 0, + "index": 26, + "blocks": [ + { + "bbox": [ + 338, + 72, + 555, + 95 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 338, + 72, + 555, + 95 + ], + "spans": [ + { + "bbox": [ + 338, + 72, + 555, + 95 + ], + "type": "text", + "content": "flections. ACM Transactions on Graphics, 41(6): Article-201, 2022. 15" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 321, + 98, + 554, + 133 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 321, + 98, + 554, + 133 + ], + "spans": [ + { + "bbox": [ + 321, + 98, + 554, + 133 + ], + "type": "text", + "content": "[54] Jiahui Lei, Yufu Wang, Georgios Pavlakos, Lingjie Liu, and Kostas Daniilidis. Gart: Gaussian articulated template models, 2023. 3, 6" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 320, + 136, + 554, + 171 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 320, + 136, + 554, + 171 + ], + "spans": [ + { + "bbox": [ + 320, + 136, + 554, + 171 + ], + "type": "text", + "content": "[55] Mengtian Li, Shengxiang Yao, Zhifeng Xie, and Keyu Chen. Gaussianbody:Clothed human reconstruction via 3d gaussian splatting,2024.3" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 320, + 173, + 555, + 220 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 320, + 173, + 555, + 220 + ], + "spans": [ + { + "bbox": [ + 320, + 173, + 555, + 220 + ], + "type": "text", + "content": "[56] Xuanyi Li, Daquan Zhou, Chenxu Zhang, Shaodong Wei, Qibin Hou, and Ming-Ming Cheng. Sora generates videos with stunning geometrical consistency. arXiv preprint arXiv: 2402.17403, 2024. 2" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 321, + 222, + 555, + 281 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 321, + 222, + 555, + 281 + ], + "spans": [ + { + "bbox": [ + 321, + 222, + 555, + 281 + ], + "type": "text", + "content": "[57] Zhengqi Li, Simon Niklaus, Noah Snavely, and Oliver Wang. Neural scene flow fields for space-time view synthesis of dynamic scenes. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), 2021. 3" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 321, + 284, + 555, + 342 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 321, + 284, + 555, + 342 + ], + "spans": [ + { + "bbox": [ + 321, + 284, + 555, + 342 + ], + "type": "text", + "content": "[58] Zhengqi Li, Qianqian Wang, Forrester Cole, Richard Tucker, and Noah Snavely. Dynibar: Neural dynamic image-based rendering. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), 2023. 3" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 321, + 345, + 555, + 403 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 321, + 345, + 555, + 403 + ], + "spans": [ + { + "bbox": [ + 321, + 345, + 555, + 403 + ], + "type": "text", + "content": "[59] Minghua Liu, Chao Xu, Haian Jin, Linghao Chen, Zexiang Xu, Hao Su, et al. One-2-3-45: Any single image to 3d mesh in 45 seconds without per-shape optimization. arXiv preprint arXiv:2306.16928, 2023. 2" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 321, + 406, + 555, + 442 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 321, + 406, + 555, + 442 + ], + "spans": [ + { + "bbox": [ + 321, + 406, + 555, + 442 + ], + "type": "text", + "content": "[60] Ruoshi Liu, Rundi Wu, Basile Van Hoorick, Pavel Tokmakov, Sergey Zakharov, and Carl Vondrick. Zero-1-to-3: Zero-shot one image to 3d object, 2023. 2" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 321, + 444, + 554, + 491 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 321, + 444, + 554, + 491 + ], + "spans": [ + { + "bbox": [ + 321, + 444, + 554, + 491 + ], + "type": "text", + "content": "[61] Yang Liu, Xiang Huang, Minghan Qin, Qinwei Lin, and Haoqian Wang. Animatable 3d gaussian: Fast and high-quality reconstruction of multiple human avatars. arXiv preprint arXiv:2311.16482, 2023. 3" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 321, + 494, + 555, + 552 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 321, + 494, + 555, + 552 + ], + "spans": [ + { + "bbox": [ + 321, + 494, + 555, + 552 + ], + "type": "text", + "content": "[62] Matthew Loper, Naureen Mahmood, Javier Romero, Gerard Pons-Moll, and Michael J. Black. SMPL: A skinned multi-person linear model. ACM Trans. Graphics (Proc. SIGGRAPH Asia), 34(6):248:1-248:16, 2015. 2, 3, 15" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 321, + 555, + 555, + 601 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 321, + 555, + 555, + 601 + ], + "spans": [ + { + "bbox": [ + 321, + 555, + 555, + 601 + ], + "type": "text", + "content": "[63] Jonathon Luiten, Georgios Kopanas, Bastian Leibe, and Deva Ramanan. Dynamic 3d gaussians: Tracking by persistent dynamic view synthesis. In 3DV, 2024. 3" + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 321, + 604, + 555, + 664 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 321, + 604, + 555, + 664 + ], + "spans": [ + { + "bbox": [ + 321, + 604, + 555, + 664 + ], + "type": "text", + "content": "[64] Luke Melas-Kyriazi, Iro Laina, Christian Rupprecht, and Andrea Vedaldi. Realfusion: 360deg reconstruction of any object from a single image. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 8446-8455, 2023. 2" + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 321, + 666, + 555, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 321, + 666, + 555, + 713 + ], + "spans": [ + { + "bbox": [ + 321, + 666, + 555, + 713 + ], + "type": "text", + "content": "[65] Ben Mildenhall, Pratul P. Srinivasan, Matthew Tancik, Jonathan T. Barron, Ravi Ramamoorthi, and Ren Ng. Nerf: Representing scenes as neural radiance fields for view synthesis. In ECCV, 2020. 2" + } + ] + } + ], + "index": 25 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "text", + "content": "26341" + } + ] + } + ], + "index": 27 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 10 + }, + { + "para_blocks": [ + { + "bbox": [ + 62, + 72, + 297, + 715 + ], + "type": "list", + "angle": 0, + "index": 11, + "blocks": [ + { + "bbox": [ + 62, + 72, + 297, + 131 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 72, + 297, + 131 + ], + "spans": [ + { + "bbox": [ + 62, + 72, + 297, + 131 + ], + "type": "text", + "content": "[66] Gyeongsik Moon, Hongsuk Choi, and Kyoung Mu Lee. Accurate 3d hand pose estimation for whole-body 3d human mesh estimation. In Computer Vision and Pattern Recognition Workshop (CVPRW), 2022. 2" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 62, + 133, + 296, + 180 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 133, + 296, + 180 + ], + "spans": [ + { + "bbox": [ + 62, + 133, + 296, + 180 + ], + "type": "text", + "content": "[67] Alex Nichol, Heewoo Jun, Prafulla Dhariwal, Pamela Mishkin, and Mark Chen. Point-e: A system for generating 3d point clouds from complex prompts, 2022. 2" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 62, + 182, + 296, + 229 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 182, + 296, + 229 + ], + "spans": [ + { + "bbox": [ + 62, + 182, + 296, + 229 + ], + "type": "text", + "content": "[68] Keunhong Park, Utkarsh Sinha, Jonathan T. Barron, Sofien Bouaziz, Dan B Goldman, Steven M. Seitz, and Ricardo Martin-Brualla. Nerfies: Deformable neural radiance fields. ICCV, 2021. 3" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 62, + 231, + 296, + 302 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 231, + 296, + 302 + ], + "spans": [ + { + "bbox": [ + 62, + 231, + 296, + 302 + ], + "type": "text", + "content": "[69] Keunhong Park, Utkarsh Sinha, Peter Hedman, Jonathan T. Barron, Sofien Bouaziz, Dan B Goldman, Ricardo Martin-Brualla, and Steven M. Seitz. Hypernerf: A higher-dimensional representation for topologically varying neural radiance fields. ACM Trans. Graph., 40(6), 2021. 3" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 62, + 303, + 296, + 361 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 303, + 296, + 361 + ], + "spans": [ + { + "bbox": [ + 62, + 303, + 296, + 361 + ], + "type": "text", + "content": "[70] Chaitanya Patel, Zhouyingcheng Liao, and Gerard Pons-Moll. Tailornet: Predicting clothing in 3d as a function of human pose, shape and garment style. In IEEE Conference on Computer Vision and Pattern Recognition (CVPR). IEEE, 2020. 2" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 62, + 363, + 297, + 446 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 363, + 297, + 446 + ], + "spans": [ + { + "bbox": [ + 62, + 363, + 297, + 446 + ], + "type": "text", + "content": "[71] Georgios Pavlakos, Vasileios Choutas, Nima Ghorbani, Timo Bolkart, Ahmed A. Osman, Dmitrios Tzionas, and Michael J. Black. Expressive body capture: 3D hands, face, and body from a single image. In Proceedings of the IEEE Computer Society Conference on Computer Vision and Pattern Recognition, pages 10967-10977, 2019. 1, 2, 3, 15" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 62, + 447, + 297, + 530 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 447, + 297, + 530 + ], + "spans": [ + { + "bbox": [ + 62, + 447, + 297, + 530 + ], + "type": "text", + "content": "[72] Georgios Pavlakos, Vasileios Choutas, Nima Ghorbani, Timo Bolkart, Ahmed AA Osman, Dimitrios Tzionas, and Michael J Black. Expressive body capture: 3d hands, face, and body from a single image. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 10975-10985, 2019. 3" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 62, + 532, + 297, + 592 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 532, + 297, + 592 + ], + "spans": [ + { + "bbox": [ + 62, + 532, + 297, + 592 + ], + "type": "text", + "content": "[73] Gerard Pons-Moll, Sergi Pujades, Sonny Hu, and Michael Black. Clothcap: Seamless 4d clothing capture and retargeting. ACM Transactions on Graphics, (Proc. SIGGRAPH), 36(4), 2017. Two first authors contributed equally. 2" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 62, + 593, + 297, + 628 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 593, + 297, + 628 + ], + "spans": [ + { + "bbox": [ + 62, + 593, + 297, + 628 + ], + "type": "text", + "content": "[74] Ben Poole, Ajay Jain, Jonathan T. Barron, and Ben Mildenhall. Dreamfusion: Text-to-3d using 2d diffusion. arXiv, 2022. 2" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 62, + 629, + 297, + 689 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 629, + 297, + 689 + ], + "spans": [ + { + "bbox": [ + 62, + 629, + 297, + 689 + ], + "type": "text", + "content": "[75] Albert Pumarola, Enric Corona, Gerard Pons-Moll, and Francesc Moreno-Noguer. D-NeRF: Neural Radiance Fields for Dynamic Scenes. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, 2020. 3" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 62, + 689, + 297, + 715 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 689, + 297, + 715 + ], + "spans": [ + { + "bbox": [ + 62, + 689, + 297, + 715 + ], + "type": "text", + "content": "[76] Guocheng Qian, Jinjie Mai, Abdullah Hamdi, Jian Ren, Aliaksandr Siarohin, Bing Li, Hsin-Ying Lee," + } + ] + } + ], + "index": 10 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 320, + 72, + 555, + 713 + ], + "type": "list", + "angle": 0, + "index": 23, + "blocks": [ + { + "bbox": [ + 338, + 72, + 555, + 142 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 338, + 72, + 555, + 142 + ], + "spans": [ + { + "bbox": [ + 338, + 72, + 555, + 142 + ], + "type": "text", + "content": "Ivan Skorokhodov, Peter Wonka, Sergey Tulyakov, and Bernard Ghanem. Magic123: One image to high-quality 3d object generation using both 2d and 3d diffusion priors. In The Twelfth International Conference on Learning Representations (ICLR), 2024. 2" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 320, + 146, + 555, + 182 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 320, + 146, + 555, + 182 + ], + "spans": [ + { + "bbox": [ + 320, + 146, + 555, + 182 + ], + "type": "text", + "content": "[77] Zhiyin Qian, Shaofei Wang, Marko Mihajlovic, Andreas Geiger, and Siyu Tang. 3dgs- avatar: Animatable avatars via deformable 3d gaussian splatting. 2024. 3" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 320, + 184, + 555, + 232 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 320, + 184, + 555, + 232 + ], + "spans": [ + { + "bbox": [ + 320, + 184, + 555, + 232 + ], + "type": "text", + "content": "[78] Jiawei Ren, Liang Pan, Jiaxiang Tang, Chi Zhang, Ang Cao, Gang Zeng, and Ziwei Liu. Dreamgaussian4d: Generative 4d gaussian splatting. arXiv preprint arXiv:2312.17142, 2023. 3, 5, 6" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 320, + 234, + 555, + 280 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 320, + 234, + 555, + 280 + ], + "spans": [ + { + "bbox": [ + 320, + 234, + 555, + 280 + ], + "type": "text", + "content": "[79] Robin Rombach, Andreas Blattmann, Dominik Lorenz, Patrick Esser, and Björn Ommer. High-resolution image synthesis with latent diffusion models, 2021. 2" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 320, + 284, + 555, + 343 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 320, + 284, + 555, + 343 + ], + "spans": [ + { + "bbox": [ + 320, + 284, + 555, + 343 + ], + "type": "text", + "content": "[80] Yu Rong, Takaaki Shiratori, and Hanbyul Joo. FrankMocap: A Monocular 3D Whole-Body Pose Estimation System via Regression and Integration. In Proceedings of the IEEE International Conference on Computer Vision, pages 1749–1759, 2021. 2" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 320, + 345, + 555, + 416 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 320, + 345, + 555, + 416 + ], + "spans": [ + { + "bbox": [ + 320, + 345, + 555, + 416 + ], + "type": "text", + "content": "[81] Shunsuke Saito, Zeng Huang, Ryota Natsume, Shigeo Morishima, Angjoo Kanazawa, and Hao Li. Pifu: Pixel-aligned implicit function for high-resolution clothed human digitization. In The IEEE International Conference on Computer Vision (ICCV), 2019. 1, 2" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 320, + 418, + 555, + 466 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 320, + 418, + 555, + 466 + ], + "spans": [ + { + "bbox": [ + 320, + 418, + 555, + 466 + ], + "type": "text", + "content": "[82] Shunsuke Saito, Tomas Simon, Jason Saragih, and Hanbyul Joo. Pifuhd: Multi-level pixel-aligned implicit function for high-resolution 3d human digitization. In CVPR, 2020. 1, 2" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 320, + 468, + 555, + 517 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 320, + 468, + 555, + 517 + ], + "spans": [ + { + "bbox": [ + 320, + 468, + 555, + 517 + ], + "type": "text", + "content": "[83] Igor Santesteban, Miguel A. Otaduy, and Dan Casas. Learning-Based Animation of Clothing for Virtual Try-On. Computer Graphics Forum (Proc. Eurographics), 2019. 2" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 320, + 518, + 555, + 566 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 320, + 518, + 555, + 566 + ], + "spans": [ + { + "bbox": [ + 320, + 518, + 555, + 566 + ], + "type": "text", + "content": "[84] Sara Fridovich-Keil and Giacomo Meanti, Frederik Rahbæk Warburg, Benjamin Recht, and Angjoo Kanazawa. K-planes: Explicit radiance fields in space, time, and appearance. In CVPR, 2023. 3" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 320, + 568, + 555, + 639 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 320, + 568, + 555, + 639 + ], + "spans": [ + { + "bbox": [ + 320, + 568, + 555, + 639 + ], + "type": "text", + "content": "[85] Ruizhi Shao, Zerong Zheng, Hanzhang Tu, Boning Liu, Hongwen Zhang, and Yebin Liu. Tensor4d: Efficient neural 4d decomposition for high-fidelity dynamic reconstruction and rendering. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition, 2023. 3" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 320, + 641, + 555, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 320, + 641, + 555, + 713 + ], + "spans": [ + { + "bbox": [ + 320, + 641, + 555, + 713 + ], + "type": "text", + "content": "[86] Zhijing Shao, Zhaolong Wang, Zhuang Li, Duotun Wang, Xiangru Lin, Yu Zhang, Mingming Fan, and Zeyu Wang. SplattingAvatar: Realistic Real-Time Human Avatars with Mesh-Embedded Gaussian Splatting. In Computer Vision and Pattern Recognition (CVPR), 2024. 4" + } + ] + } + ], + "index": 22 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 318, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 318, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 318, + 757 + ], + "type": "text", + "content": "26342" + } + ] + } + ], + "index": 24 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 11 + }, + { + "para_blocks": [ + { + "bbox": [ + 61, + 72, + 297, + 713 + ], + "type": "list", + "angle": 0, + "index": 12, + "blocks": [ + { + "bbox": [ + 61, + 72, + 297, + 131 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 61, + 72, + 297, + 131 + ], + "spans": [ + { + "bbox": [ + 61, + 72, + 297, + 131 + ], + "type": "text", + "content": "[87] Ruoxi Shi, Hansheng Chen, Zhuoyang Zhang, Minghua Liu, Chao Xu, Xinyue Wei, Linghao Chen, Chong Zeng, and Hao Su. Zero123++: a single image to consistent multi-view diffusion base model, 2023. 2" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 61, + 133, + 296, + 168 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 61, + 133, + 296, + 168 + ], + "spans": [ + { + "bbox": [ + 61, + 133, + 296, + 168 + ], + "type": "text", + "content": "[88] Stanislaw Szymanowicz, Christian Rupprecht, and Andrea Vedaldi. Splatter image: Ultra-fast single-view 3d reconstruction, 2023. 2" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 62, + 170, + 296, + 217 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 170, + 296, + 217 + ], + "spans": [ + { + "bbox": [ + 62, + 170, + 296, + 217 + ], + "type": "text", + "content": "[89] Jiaxiang Tang, Jiawei Ren, Hang Zhou, Ziwei Liu, and Gang Zeng. Dreamgaussian: Generative gaussian splatting for efficient 3d content creation. arXiv preprint arXiv:2309.16653, 2023. 2, 3, 6, 8" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 62, + 219, + 296, + 289 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 219, + 296, + 289 + ], + "spans": [ + { + "bbox": [ + 62, + 219, + 296, + 289 + ], + "type": "text", + "content": "[90] Junshu Tang, Tengfei Wang, Bo Zhang, Ting Zhang, Ran Yi, Lizhuang Ma, and Dong Chen. Make-it-3d: High-fidelity 3d creation from a single image with diffusion prior. In Proceedings of the IEEE/CVF International Conference on Computer Vision (ICCV), pages 22819–22829, 2023. 2" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 62, + 291, + 296, + 350 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 291, + 296, + 350 + ], + "spans": [ + { + "bbox": [ + 62, + 291, + 296, + 350 + ], + "type": "text", + "content": "[91] Jiaxiang Tang, Zhaoxi Chen, Xiaokang Chen, Tengfei Wang, Gang Zeng, and Ziwei Liu. Lgm: Large multi-view gaussian model for high-resolution 3d content creation. arXiv preprint arXiv:2402.05054, 2024. 2, 3, 4, 6, 8" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 62, + 351, + 296, + 422 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 351, + 296, + 422 + ], + "spans": [ + { + "bbox": [ + 62, + 351, + 296, + 422 + ], + "type": "text", + "content": "[92] Edgar Tretschk, Ayush Tewari, Vladislav Golyanik, Michael Zollhöfer, Christoph Lassner, and Christian Theobalt. Non-rigid neural radiance fields: Reconstruction and novel view synthesis of a dynamic scene from monocular video. In IEEE International Conference on Computer Vision (ICCV). IEEE, 2021. 3" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 62, + 423, + 296, + 459 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 423, + 296, + 459 + ], + "spans": [ + { + "bbox": [ + 62, + 423, + 296, + 459 + ], + "type": "text", + "content": "[93] Alex Trevithick and Bo Yang. Grf: Learning a general radiance field for 3d scene representation and rendering. In arXiv:2010.04595, 2020. 2" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 62, + 460, + 296, + 495 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 460, + 296, + 495 + ], + "spans": [ + { + "bbox": [ + 62, + 460, + 296, + 495 + ], + "type": "text", + "content": "[94] Haithem Turki, Jason Y. Zhang, Francesco Ferroni, and Deva Ramanan. Suits: Scalable urban dynamic scenes, 2023. 3" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 62, + 497, + 296, + 544 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 497, + 296, + 544 + ], + "spans": [ + { + "bbox": [ + 62, + 497, + 296, + 544 + ], + "type": "text", + "content": "[95] Raquel Vidaurre, Igor Santesteban, Elena Garces, and Dan Casas. Fully Convolutional Graph Neural Networks for Parametric Virtual Try-On. Computer Graphics Forum (Proc. SCA), 2020. 2" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 62, + 545, + 296, + 616 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 545, + 296, + 616 + ], + "spans": [ + { + "bbox": [ + 62, + 545, + 296, + 616 + ], + "type": "text", + "content": "[96] Wenbo Wang, Hsuan-I Ho, Chen Guo, Boxiang Rong, Artur Grigorev, Jie Song, Juan Jose Zarate, and Otmar Hilliges. 4d-dress: A 4d dataset of real-world human clothing with semantic annotations. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition (CVPR), 2024. 6, 16" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 62, + 617, + 296, + 676 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 617, + 296, + 676 + ], + "spans": [ + { + "bbox": [ + 62, + 617, + 296, + 676 + ], + "type": "text", + "content": "[97] Guanjun Wu, Taoran Yi, Jiemin Fang, Lingxi Xie, Xiaopeng Zhang, Wei Wei, Wenyu Liu, Qi Tian, and Wang Xinggang. 4d gaussian splatting for real-time dynamic scene rendering. arXiv preprint arXiv:2310.08528, 2023. 3" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 62, + 677, + 296, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 677, + 296, + 713 + ], + "spans": [ + { + "bbox": [ + 62, + 677, + 296, + 713 + ], + "type": "text", + "content": "[98] Wenqi Xian, Jia-Bin Huang, Johannes Kopf, and Changil Kim. Space-time neural irradiance fields for free-viewpoint video. In Proceedings of the" + } + ] + } + ], + "index": 11 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 316, + 72, + 555, + 713 + ], + "type": "list", + "angle": 0, + "index": 24, + "blocks": [ + { + "bbox": [ + 338, + 72, + 555, + 106 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 338, + 72, + 555, + 106 + ], + "spans": [ + { + "bbox": [ + 338, + 72, + 555, + 106 + ], + "type": "text", + "content": "IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 9421-9431, 2021. 3" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 321, + 109, + 555, + 168 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 321, + 109, + 555, + 168 + ], + "spans": [ + { + "bbox": [ + 321, + 109, + 555, + 168 + ], + "type": "text", + "content": "[99] Donglai Xiang, Fabian Prada, Chenglei Wu, and Jessica Hodgins. Monoclothcap: Towards temporally coherent clothing capture from monocular rgb video. In Proceedings of International Conference on 3D Vision (3DV '20), pages 322 - 332, 2020. 2" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 316, + 170, + 555, + 228 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 170, + 555, + 228 + ], + "spans": [ + { + "bbox": [ + 316, + 170, + 555, + 228 + ], + "type": "text", + "content": "[100] Donglai Xiang, Fabian Prada, Timur Bagautdinov, Weipeng Xu, Yuan Dong, He Wen, Jessica Hodgins, and Chenglei Wu. Modeling clothing as a separate layer for an animatable human avatar. ACM Transactions on Graphics, 40(6):1-15, 2021. 2" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 316, + 230, + 555, + 277 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 230, + 555, + 277 + ], + "spans": [ + { + "bbox": [ + 316, + 230, + 555, + 277 + ], + "type": "text", + "content": "[101] Enze Xie, Wenhai Wang, Zhiding Yu, Anima Anandkumar, Jose M. Alvarez, and Ping Luo. Segformer: Simple and efficient design for semantic segmentation with transformers. CoRR, abs/2105.15203, 2021. 4" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 316, + 278, + 555, + 326 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 278, + 555, + 326 + ], + "spans": [ + { + "bbox": [ + 316, + 278, + 555, + 326 + ], + "type": "text", + "content": "[102] Tianyi Xie, Zeshun Zong, Yuxing Qiu, Xuan Li, Yutao Feng, Yin Yang, and Chenfanfu Jiang. Physgaussian: Physics-integrated 3d gaussians for generative dynamics. arXiv preprint arXiv:2311.12198, 2023. 5" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 316, + 327, + 555, + 398 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 327, + 555, + 398 + ], + "spans": [ + { + "bbox": [ + 316, + 327, + 555, + 398 + ], + "type": "text", + "content": "[103] Yuliang Xiu, Jinlong Yang, Dimitrios Tzionas, and Michael J. Black. ICON: Implicit Clothed humans Obtained from Normals. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 13296-13306, 2022. 1, 2" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 316, + 399, + 555, + 470 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 399, + 555, + 470 + ], + "spans": [ + { + "bbox": [ + 316, + 399, + 555, + 470 + ], + "type": "text", + "content": "[104] Yuliang Xiu, Jinlong Yang, Xu Cao, Dimitrios Tzionas, and Michael J. Black. ECON: Explicit Clothed humans Optimized via Normal integration. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), 2023. 1, 2" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 316, + 472, + 555, + 532 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 472, + 555, + 532 + ], + "spans": [ + { + "bbox": [ + 316, + 472, + 555, + 532 + ], + "type": "text", + "content": "[105] Qiangeng Xu, Weiyue Wang, Duygu Ceylan, Radomir Mech, and Ulrich Neumann. Disn: Deep implicit surface network for high-quality single-view 3d reconstruction. In Advances in Neural Information Processing Systems 32, pages 492-502. 2019. 2" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 316, + 533, + 555, + 591 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 533, + 555, + 591 + ], + "spans": [ + { + "bbox": [ + 316, + 533, + 555, + 591 + ], + "type": "text", + "content": "[106] Zhongcong Xu, Jianfeng Zhang, Jun Hao Liew, Hanshu Yan, Jia-Wei Liu, Chenxu Zhang, Jiashi Feng, and Mike Zheng Shou. Magicanimate: Temporally consistent human image animation using diffusion model. 2023. 6" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 316, + 593, + 555, + 653 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 593, + 555, + 653 + ], + "spans": [ + { + "bbox": [ + 316, + 593, + 555, + 653 + ], + "type": "text", + "content": "[107] Xueting Yang, Yihao Luo, Yuliang Xiu, Wei Wang, Hao Xu, and Zhaoxin Fan. D-IF: Uncertainty-aware Human Digitization via Implicit Distribution Field. In Proceedings of the IEEE/CVF international conference on computer vision, 2023. 2" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 316, + 654, + 555, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 654, + 555, + 713 + ], + "spans": [ + { + "bbox": [ + 316, + 654, + 555, + 713 + ], + "type": "text", + "content": "[108] Yifan Yang, Dong Liu, Shuhai Zhang, Zeshuai Deng, Zixiong Huang, and Mingkui Tan. Hilo: Detailed and robust 3d clothed human reconstruction with high- and low-frequency information of parametric models. In Proceedings of the IEEE/CVF Conference" + } + ] + } + ], + "index": 23 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "text", + "content": "26343" + } + ] + } + ], + "index": 25 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 12 + }, + { + "para_blocks": [ + { + "bbox": [ + 56, + 72, + 296, + 713 + ], + "type": "list", + "angle": 0, + "index": 11, + "blocks": [ + { + "bbox": [ + 82, + 72, + 296, + 95 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 82, + 72, + 296, + 95 + ], + "spans": [ + { + "bbox": [ + 82, + 72, + 296, + 95 + ], + "type": "text", + "content": "on Computer Vision and Pattern Recognition, pages 10671-10681, 2024. 2" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 56, + 98, + 296, + 194 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 98, + 296, + 194 + ], + "spans": [ + { + "bbox": [ + 56, + 98, + 296, + 194 + ], + "type": "text", + "content": "[109] Zhitao Yang, Zhongang Cai, Haiyi Mei, Shuai Liu, Zhaoxi Chen, Weiye Xiao, Yukun Wei, Zhongfei Qing, Chen Wei, Bo Dai, Wayne Wu, Chen Qian, Dahua Lin, Ziwei Liu, and Lei Yang. Synbody: Synthetic dataset with layered human models for 3d human perception and modeling. In Proceedings of the IEEE/CVF International Conference on Computer Vision (ICCV), pages 20282-20292, 2023. 6" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 56, + 196, + 296, + 255 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 196, + 296, + 255 + ], + "spans": [ + { + "bbox": [ + 56, + 196, + 296, + 255 + ], + "type": "text", + "content": "[110] Ziyi Yang, Xinyu Gao, Wen Zhou, Shaohui Jiao, Yuqing Zhang, and Xiaogang Jin. Deformable 3d gaussians for high-fidelity monocular dynamic scene reconstruction. arXiv preprint arXiv:2309.13101, 2023. 3" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 56, + 258, + 296, + 304 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 258, + 296, + 304 + ], + "spans": [ + { + "bbox": [ + 56, + 258, + 296, + 304 + ], + "type": "text", + "content": "[111] Mingqiao Ye, Martin Danelljan, Fisher Yu, and Lei Ke. Gaussian grouping: Segment and edit anything in 3d scenes. arXiv preprint arXiv:2312.00732, 2023. 4, 15" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 56, + 308, + 296, + 366 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 308, + 296, + 366 + ], + "spans": [ + { + "bbox": [ + 56, + 308, + 296, + 366 + ], + "type": "text", + "content": "[112] Xianggang Yu, Mutian Xu, Yidan Zhang, Haolin Liu, Chongjie Ye, Yushuang Wu, Zizheng Yan, Tianyou Liang, Guanying Chen, Shuguang Cui, and Xiaoguang Han. Mvimgnet: A large-scale dataset of multi-view images. In CVPR, 2023. 2" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 56, + 369, + 296, + 428 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 369, + 296, + 428 + ], + "spans": [ + { + "bbox": [ + 56, + 369, + 296, + 428 + ], + "type": "text", + "content": "[113] Zhengming Yu, Wei Cheng, Xian Liu, Wayne Wu, and Kwan-Yee Lin. Monohuman: Animatable human neural field from monocular video. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 16943-16953, 2023. 3, 6" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 56, + 431, + 296, + 502 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 431, + 296, + 502 + ], + "spans": [ + { + "bbox": [ + 56, + 431, + 296, + 502 + ], + "type": "text", + "content": "[114] Wentao Yuan, Zhaoyang Lv, Tanner Schmidt, and Steven Lovegrove. Star: Self-supervised tracking and reconstruction of rigid objects in motion with neural rendering. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 13144-13152, 2021. 3" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 56, + 505, + 296, + 563 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 505, + 296, + 563 + ], + "spans": [ + { + "bbox": [ + 56, + 505, + 296, + 563 + ], + "type": "text", + "content": "[115] Ilya Zakharkin, Kirill Mazur, Artur Grigorev, and Victor Lempitsky. Point-based modeling of human clothing. In Proceedings of the IEEE/CVF International Conference on Computer Vision (ICCV), pages 14718-14727, 2021. 2" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 56, + 567, + 296, + 613 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 567, + 296, + 613 + ], + "spans": [ + { + "bbox": [ + 56, + 567, + 296, + 613 + ], + "type": "text", + "content": "[116] Yan Zeng, Guoqiang Wei, Jiani Zheng, Jiaxin Zou, Yang Wei, Yuchen Zhang, and Hang Li. Make pixels dance: High-dynamic video generation. arXiv:2311.10982, 2023. 2" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 56, + 616, + 296, + 662 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 616, + 296, + 662 + ], + "spans": [ + { + "bbox": [ + 56, + 616, + 296, + 662 + ], + "type": "text", + "content": "[117] Zheng Zerong, Yu Tao, Liu Yebin, and Dai Qionghai. Pamir: Parametric model-conditioned implicit representation for image-based human reconstruction, 2021. 1, 2" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 56, + 666, + 296, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 666, + 296, + 713 + ], + "spans": [ + { + "bbox": [ + 56, + 666, + 296, + 713 + ], + "type": "text", + "content": "[118] Hongwen Zhang, Yating Tian, Yuxiang Zhang, Mengcheng Li, Liang An, Zhenan Sun, and Yebin Liu. Pymaf-x: Towards well-aligned full-body model regression from monocular images. IEEE Transac-" + } + ] + } + ], + "index": 10 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 316, + 72, + 555, + 228 + ], + "type": "list", + "angle": 0, + "index": 15, + "blocks": [ + { + "bbox": [ + 342, + 72, + 555, + 95 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 342, + 72, + 555, + 95 + ], + "spans": [ + { + "bbox": [ + 342, + 72, + 555, + 95 + ], + "type": "text", + "content": "tions on Pattern Analysis and Machine Intelligence, 2023. 2" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 316, + 98, + 555, + 168 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 98, + 555, + 168 + ], + "spans": [ + { + "bbox": [ + 316, + 98, + 555, + 168 + ], + "type": "text", + "content": "[119] Hao Zhu, Xinxin Zuo, Sen Wang, Xun Cao, and Ruigang Yang. Detailed human shape estimation from a single image by hierarchical mesh deformation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 4491-4500, 2019. 2" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 316, + 171, + 555, + 228 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 171, + 555, + 228 + ], + "spans": [ + { + "bbox": [ + 316, + 171, + 555, + 228 + ], + "type": "text", + "content": "[120] Shenhao Zhu, Junming Leo Chen, Zuozhuo Dai, Yinghui Xu, Xun Cao, Yao Yao, Hao Zhu, and Siyu Zhu. Champ: Controllable and consistent human image animation with 3d parametric guidance, 2024. 6" + } + ] + } + ], + "index": 14 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 749, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 749, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 749, + 317, + 757 + ], + "type": "text", + "content": "26344" + } + ] + } + ], + "index": 16 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 13 + } + ], + "_backend": "vlm", + "_version_name": "2.6.4" +} \ No newline at end of file diff --git a/2025/DiscoVLA_ Discrepancy Reduction in Vision, Language, and Alignment for Parameter-Efficient Video-Text Retrieval/9d7a1765-54c3-41c7-9719-3ce7d4a7b6fa_content_list.json b/2025/DiscoVLA_ Discrepancy Reduction in Vision, Language, and Alignment for Parameter-Efficient Video-Text Retrieval/9d7a1765-54c3-41c7-9719-3ce7d4a7b6fa_content_list.json new file mode 100644 index 0000000000000000000000000000000000000000..27e01d7233bd376e0c31530b61c7f29376712096 --- /dev/null +++ b/2025/DiscoVLA_ Discrepancy Reduction in Vision, Language, and Alignment for Parameter-Efficient Video-Text Retrieval/9d7a1765-54c3-41c7-9719-3ce7d4a7b6fa_content_list.json @@ -0,0 +1,1594 @@ +[ + { + "type": "text", + "text": "DiscoVLA: Discrepancy Reduction in Vision, Language, and Alignment for Parameter-Efficient Video-Text Retrieval", + "text_level": 1, + "bbox": [ + 138, + 128, + 859, + 174 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Leqi Shen $^{1,2,4*}$ Guoqiang Gong $^{5\\dagger}$ Tianxiang Hao $^{1,2}$ Tao He $^{6,7}$ Yifeng Zhang $^{5}$ Pengzhang Liu $^{5}$ Sicheng Zhao $^{2‡}$ Jungong Han $^{3}$ Guiguang Ding $^{1,2‡}$", + "bbox": [ + 147, + 200, + 846, + 239 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "$^{1}$ School of Software $^{2}$ BNrist $^{3}$ Department of Automation, Tsinghua University $^{4}$ Hangzhou Zhuoxi Institute of Brain and Intelligence", + "bbox": [ + 178, + 239, + 818, + 273 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "$^{5}$ JD.com $^{6}$ GRG Banking Equipment Co., Ltd. $^{7}$ South China University of Technology", + "bbox": [ + 140, + 273, + 870, + 292 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Abstract", + "text_level": 1, + "bbox": [ + 246, + 325, + 326, + 342 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "The parameter-efficient adaptation of the image-text pretraining model CLIP for video-text retrieval is a prominent area of research. While CLIP is focused on image-level vision-language matching, video-text retrieval demands comprehensive understanding at the video level. Three key discrepancies emerge in the transfer from image-level to video-level: vision, language, and alignment. However, existing methods mainly focus on vision while neglecting language and alignment. In this paper, we propose Discrepancy Reduction in Vision, Language, and Alignment (DiscoVLA), which simultaneously mitigates all three discrepancies. Specifically, we introduce Image-Video Features Fusion to integrate image-level and video-level features, effectively tackling both vision and language discrepancies. Additionally, we generate pseudo image captions to learn fine-grained image-level alignment. To mitigate alignment discrepancies, we propose Image-to-Video Alignment Distillation, which leverages image-level alignment knowledge to enhance video-level alignment. Extensive experiments demonstrate the superiority of our DiscoVLA. In particular, on MSRVTT with CLIP (ViT-B/16), DiscoVLA outperforms previous methods by $2.2\\%$ R@1 and $7.5\\%$ R@sum. The code is available at https://github.com/LunarShen/DsicoVLA.", + "bbox": [ + 88, + 358, + 483, + 720 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "1. Introduction", + "text_level": 1, + "bbox": [ + 89, + 750, + 220, + 763 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "The rapid expansion of online video content has generated a growing demand for effective video-text retrieval, focusing on matching video content with relevant textual descriptions. Inspired by the success of image-text pretraining approaches such as CLIP [42], recent efforts [9, 14, 24, 37, 48, 49, 51] have increasingly focused on extend", + "bbox": [ + 89, + 773, + 482, + 864 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/bfbe10255f720db020eda19693dec725fa3660b25e37918a8c85685474458566.jpg", + "image_caption": [ + "Figure 1. When parameter-efficient transferring image-level CLIP to video-text retrieval, we identify three key discrepancies: vision, alignment, and language. Unlike previous methods, our DiscoVLA aims to address all these discrepancies. More visualization examples are provided in Figure 4." + ], + "image_footnote": [], + "bbox": [ + 535, + 324, + 888, + 595 + ], + "page_idx": 0 + }, + { + "type": "table", + "img_path": "images/4296e1e353b6b09270b71acaae8eb2dae06abe5ead9805faa297f623ce290e48.jpg", + "table_caption": [], + "table_footnote": [], + "table_body": "
MethodImage-Level PETL Video-Level
VisionLanguageAlignment
Previous Methods✓××
Our DiscoVLA✓✓✓
", + "bbox": [ + 557, + 598, + 864, + 662 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "ing these powerful capabilities to video-text retrieval tasks. However, fully fine-tuning these pretrained models requires updating numerous parameters for every dataset, resulting in large storage overhead [22, 58]. Therefore, we focus on parameter-efficient transfer learning methods, targeting strong performance with minimal trainable parameters.", + "bbox": [ + 511, + 776, + 906, + 867 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Although both image-text pretraining and video-text retrieval involve vision-language (VL) matching, they exhibit", + "bbox": [ + 511, + 869, + 906, + 901 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "CVF", + "bbox": [ + 106, + 2, + 181, + 42 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore.", + "bbox": [ + 236, + 0, + 810, + 46 + ], + "page_idx": 0 + }, + { + "type": "page_footnote", + "text": "* lunarshen@gmail.com. Work done during an internship at JD.com.", + "bbox": [ + 107, + 875, + 468, + 887 + ], + "page_idx": 0 + }, + { + "type": "page_footnote", + "text": "†Project leader. ‡ Corresponding authors.", + "bbox": [ + 109, + 887, + 338, + 898 + ], + "page_idx": 0 + }, + { + "type": "page_number", + "text": "19702", + "bbox": [ + 480, + 944, + 517, + 955 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "substantial discrepancies that hinder straightforward knowledge transfer. To be specific, image-text pretraining focuses on matching images with image captions, which represents image-level VL matching. In contrast, video-text retrieval focuses on matching videos with video captions, which represents video-level VL matching. The challenges of video-level VL matching, which demands understanding motion and time, create significant obstacles for transferring knowledge from image-text to video-text tasks.", + "bbox": [ + 89, + 90, + 480, + 226 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "In this work, we explore the discrepancies between image-level and video-level VL matching across three key aspects: vision, language, and alignment. Videos introduce a temporal dimension, which is absent in images. Similarly, the distinction between image captions and video captions arises from the varying levels of granularity, such as isolated image descriptions versus continuous video narratives. Thus, video-level alignment is inherently more complex than image-level alignment, as it must account for intricate spatio-temporal relationships. In Figure 1, the three images correspond to \"preparing food\", \"pouring\", and \"cutting\", while the entire video demonstrates an instructional cooking process. Image-level VL matching focuses on isolated images, while video-level VL matching emphasizes the overall understanding of continuous segments.", + "bbox": [ + 89, + 229, + 480, + 455 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "However, current methods [5, 22, 25, 58] struggle to fully address the discrepancies mentioned above. Some [5, 22, 25, 58] focus on the vision aspect, modeling temporal relationships to extract video-level features, while others [25, 58] attempt to reduce the modality gap between vision and language through shared parameter mechanisms. Despite these advances, these methods remain limited in both language and alignment discrepancies.", + "bbox": [ + 89, + 458, + 480, + 578 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "To tackle these challenges, we introduce Discrepancy Reduction in Vision, Language, and Alignment, termed DiscoVLA, a novel parameter-efficient method for videotext retrieval. Unlike previous methods that target a single gap, our method aims to simultaneously reduce all significant discrepancies, as illustrated in the table of Figure 1.", + "bbox": [ + 89, + 580, + 480, + 672 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Specifically, we propose an Image-Video Features Fusion module (IVFusion), which addresses both vision and language discrepancies through a unified feature extraction approach. IVFusion effectively combines both image- and video-level features by utilizing a lightweight adapter. In addition to video-level alignment, we propose Pseudo Image-level Alignment (PIImgAlign) which generates pseudo image captions and learns fine-grained image-level alignment. Furthermore, to reduce alignment discrepancy, we propose Image-to-Video Alignment Distillation (AlignDistill), transferring image-level alignment knowledge to improve video-level alignment. All proposed modules prioritize parameter efficiency: IVFusion requires minimal trainable parameters, while PImgAlign and AlignDistill operate without increasing inference parameters.", + "bbox": [ + 89, + 674, + 482, + 900 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "In summary, the main contributions are as follows:", + "bbox": [ + 532, + 90, + 867, + 104 + ], + "page_idx": 1 + }, + { + "type": "list", + "sub_type": "text", + "list_items": [ + "- We reveal, for the first time, the necessity of addressing all three image-to-video discrepancies—vision, language, and alignment—to achieve parameter-efficient VTR.", + "- We introduce IVFusion to fuse image- and video-level features for both vision and language gaps. PImgAlign is introduced for fine-grained image-level alignment. We introduce AlignDistill to minimize alignment gaps.", + "- Our DiscoVLA achieves state-of-the-art results on MSRVTT, LSMDC, ActivityNet, and DiDeMo. Notably, on MSRVTT with CLIP (ViT-B/16), DiscoVLA reaches $50.5\\%$ R@1, surpassing previous methods by $2.2\\%$ ." + ], + "bbox": [ + 513, + 106, + 903, + 272 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "2. Related Work", + "text_level": 1, + "bbox": [ + 513, + 286, + 653, + 301 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Video-Text Retrieval. Advances in vision-language pretraining [7, 12, 32-34, 42, 50, 52, 55] have achieved significant success across various downstream tasks [2, 10, 11, 18, 26, 45, 47, 54, 56, 57]. Leveraging powerful CLIP [42], recent video-text retrieval (VTR) methods [9, 14, 24, 37, 48, 49, 51] have achieved impressive performance enhancements. CLIP4Clip [37] is a pioneering CLIP-based method that aggregates image features via mean-pooling or transformers to obtain video features, inspiring subsequent studies. Additionally, Cap4Video [51] utilizes ZeroCap [46] and GPT-2 [41] to generate video-level captions as supplementary video content during both training and inference. In contrast, our DiscoVLA addresses the image-level and video-level gap by generating image-level captions used only in training to improve video-level alignment. These full fine-tuning methods impose substantial overhead from large trainable parameters.", + "bbox": [ + 511, + 311, + 903, + 566 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Parameter-Efficient Transfer Learning. Parameter-efficient methods aim to adapt pre-trained models to new tasks with minimal fine-tuning. The mainstream methods in natural language processing and image tasks include techniques such as Prompt [23, 27, 28, 61, 62], Adapter [6, 17, 20, 60], and LoRA [21, 60]. Prompt [27] modifies the input space by adding learnable tokens. Adapter [20] introduces a lightweight bottleneck neural network consisting of up- and down-projection layers with a non-linear activation function. LoRA [21] aims to update up- and down-projection layers by learning low-rank decomposition matrices.", + "bbox": [ + 511, + 568, + 903, + 734 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Recently, several studies have explored parameter-efficient VTR methods [5, 22, 25, 58] to incorporate temporal modeling into the frozen image-level CLIP backbone. VoP [22] employs trainable BiLSTMs [16] to produce prompt tokens in the vision encoder. DGL [58] introduces weight-sharing layers to generate prompts for both text and video encoders. MV-Adapter [25] introduces shared bottleneck structures in both video and text branches. However, these methods overlook language and alignment discrepancies. We address these limitations with DiscoVLA, a framework that facilitates robust video-level understanding.", + "bbox": [ + 511, + 734, + 903, + 900 + ], + "page_idx": 1 + }, + { + "type": "page_number", + "text": "19703", + "bbox": [ + 480, + 944, + 517, + 955 + ], + "page_idx": 1 + }, + { + "type": "image", + "img_path": "images/3baf7371eccbd1f136ffa4c85bcbf4d194ecc701dc4f2f62831d41c878f71323.jpg", + "image_caption": [ + "Figure 2. The overall framework of DiscoVLA. Initially, we generate pseudo image captions for each sampled image. In both vision and text encoders, we utilize image-level layers to acquire pretrained image-level knowledge and employ IVFusion layers to enhance spatiotemporal information. The single video caption is encoded through the text encoder, utilizing IVFusion layer as image-level (image-lvl) layer. Finally, AlignDistill is applied to distill image-level alignment to video-level alignment. For fair comparisons with previous methods, we do not generate pseudo image captions during the inference phase." + ], + "image_footnote": [], + "bbox": [ + 96, + 112, + 903, + 335 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3. Methodology", + "text_level": 1, + "bbox": [ + 89, + 435, + 225, + 452 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3.1. Preliminary", + "text_level": 1, + "bbox": [ + 89, + 460, + 220, + 478 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Video-text retrieval is a task that involves retrieving relevant videos based on a given text query, or conversely, retrieving appropriate text descriptions for a video. Consistent with previous approaches [5, 22, 25, 58], we utilize the CLIP model, pre-trained on image-text data, as our backbone.", + "bbox": [ + 89, + 486, + 483, + 561 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "To process a given video, we first sample a sequence of frame images uniformly, which are then input into the CLIP vision encoder to extract image features $v^{\\mathrm{img}}$ . These features are then aggregated through temporal average pooling to get video features $v^{\\mathrm{vid}}$ . For a given video caption, it is passed through the CLIP text encoder to extract video caption features $t^{\\mathrm{vid}}$ . Then, we compute the cross-modal contrastive loss [40], which functions as a video-level alignment term, maximizing the video-level similarity $\\mathrm{Sim}^{\\mathrm{vid}}$ between matched pairs:", + "bbox": [ + 89, + 564, + 483, + 715 + ], + "page_idx": 2 + }, + { + "type": "equation", + "text": "\n$$\n\\operatorname {S i m} _ {i j} ^ {\\mathrm {v i d}} = t _ {i} ^ {\\mathrm {v i d} ^ {T}} v _ {j} ^ {\\mathrm {v i d}}, \\tag {1}\n$$\n", + "text_format": "latex", + "bbox": [ + 217, + 729, + 480, + 752 + ], + "page_idx": 2 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {L} _ {t 2 v} \\left(\\operatorname {S i m} ^ {\\text {v i d}}\\right) = \\frac {1}{B} \\sum_ {i = 1} ^ {B} \\log \\frac {\\exp \\left(\\operatorname {S i m} _ {i i} ^ {\\text {v i d}}\\right) / \\tau)}{\\sum_ {j = 1} ^ {B} \\exp \\left(\\operatorname {S i m} _ {i j} ^ {\\text {v i d}}\\right) / \\tau)}, \\tag {2}\n$$\n", + "text_format": "latex", + "bbox": [ + 106, + 753, + 482, + 796 + ], + "page_idx": 2 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {L} _ {v 2 t} \\left(\\operatorname {S i m} ^ {\\text {v i d}}\\right) = \\frac {1}{B} \\sum_ {i = 1} ^ {B} \\log \\frac {\\exp \\left(\\operatorname {S i m} _ {i i} ^ {\\text {v i d}}\\right) / \\tau)}{\\sum_ {j = 1} ^ {B} \\exp \\left(\\operatorname {S i m} _ {j i} ^ {\\text {v i d}}\\right) / \\tau)}, \\tag {3}\n$$\n", + "text_format": "latex", + "bbox": [ + 106, + 799, + 480, + 840 + ], + "page_idx": 2 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {L} _ {A} (\\operatorname {S i m} ^ {\\mathrm {v i d}}) = - \\frac {1}{2} \\left[ \\mathcal {L} _ {t 2 v} (\\operatorname {S i m} ^ {\\mathrm {v i d}}) + \\mathcal {L} _ {v 2 t} (\\operatorname {S i m} ^ {\\mathrm {v i d}}) \\right], \\tag {4}\n$$\n", + "text_format": "latex", + "bbox": [ + 106, + 842, + 480, + 872 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "where $B$ is the batch size, $\\tau$ is the temperature parameter.", + "bbox": [ + 89, + 885, + 470, + 900 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3.2. DiscoVLA", + "text_level": 1, + "bbox": [ + 513, + 436, + 629, + 450 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "As shown in Figure 2, we propose DiscoVLA to address the challenges of vision, language, and alignment when transferring from image-level to video-level VL matching. First, we introduce Image-Video Features Fusion (IVFusion) in both vision and text encoders to enhance feature extraction. Next, we introduce Pseudo Image-Level Alignment (PIImgAlign) for fine-grained image-level alignment, consisting of Pseudo Image Language Generation and Image-Level Alignment. Finally, Image-to-Video Alignment Distillation (AlignDistill) is proposed to leverage image-level alignment knowledge to enhance video-level alignment.", + "bbox": [ + 511, + 458, + 906, + 625 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Importantly, PImgAlign is only employed during training. For a fair comparison with previous methods and due to its generation inefficiency, PImgAlign is omitted at inference. The final similarity used for retrieval is based on video-level similarity in Eq. (1).", + "bbox": [ + 511, + 625, + 905, + 700 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "We focus on parameter-efficient transfer learning for video-text retrieval, where the large backbone parameters remain fixed, and only a small number of additional parameters are trained. The trainable parameters are embedded in the parameter-efficient attention of each encoder layer, as illustrated in Figure 3.", + "bbox": [ + 511, + 700, + 905, + 792 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3.3. Image-Video Features Fusion", + "text_level": 1, + "bbox": [ + 511, + 801, + 774, + 818 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Image-level and video-level variations exist in both vision and language modalities. Our proposed IVFusion fuses these image-level and video-level features in both the vision and language encoders. Below, we illustrate IVFusion's application within the vision encoder.", + "bbox": [ + 511, + 824, + 905, + 900 + ], + "page_idx": 2 + }, + { + "type": "header", + "text": "CLSTokenofVisionEncoder EOSTokenofTextEncoder -- $\\rightarrow$ UtilizedONLYduringthe training phase.", + "bbox": [ + 96, + 93, + 689, + 107 + ], + "page_idx": 2 + }, + { + "type": "page_number", + "text": "19704", + "bbox": [ + 480, + 944, + 517, + 955 + ], + "page_idx": 2 + }, + { + "type": "image", + "img_path": "images/0704c9a8a0461be70b62fb28713d053f6bc102bbe788c05808941ec59c7cb103.jpg", + "image_caption": [ + "Figure 3. Illustration of the encoder layers for vision and text encoders. (a) Image-Level Attn operates on each of the $F$ images or image captions individually. (b) Video-Level Attn concatenates tokens across all sampled images or image captions. (c) IVFusion Attn employs a lightweight adapter to integrate the efficiency of Image-Level Attn with the effectiveness of Video-Level Attn. Here, we illustrate the application of IVFusion within the vision encoder. The text encoder adopts the same approach for processing pseudo image captions." + ], + "image_footnote": [], + "bbox": [ + 106, + 88, + 496, + 340 + ], + "page_idx": 3 + }, + { + "type": "image", + "img_path": "images/46a2b4606872b486ec8096af7f0399e265e85e9f2ff67c904082ded61a9105d2.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 504, + 89, + 885, + 339 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Image-Level Attn. To achieve parameter-efficient transfer learning, we apply LoRA in the attention module of each encoder layer. As illustrated in Figure 3a, $F \\times (N + 1)$ tokens are processed via Image-level Attn, where $F$ denotes the number of sampled images, and $N + 1$ represents the combination of the global CLS token and patch tokens. Image-Level Attn processes each of the $F$ images individually, using $N + 1$ tokens per image. However, Image-Level Attn does not encode temporal information.", + "bbox": [ + 88, + 422, + 480, + 559 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Video-Level Attn. A straightforward solution is Video-Level Attn, which concatenates tokens across all sampled images, enabling the attention module to learn spatiotemporal information. As shown in Figure 3b, $F(N + 1)$ tokens are simultaneously processed during the attention module. Due to the quadratic complexity of the attention mechanism, Image-Level Attn has a complexity of $O(F(N)^2)$ , while Video-Level Attn has a higher complexity of $O((FN)^2)$ .", + "bbox": [ + 88, + 559, + 482, + 696 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "IVFusion Attn. To combine Image-Level Attn's efficiency with Video-Level Attn's effectiveness, we introduce IV-Fusion Attn in Figure 3c. We extend Image-Level Attn by incorporating a branch that uses all CLS tokens to extract spatio-temporal information, reducing complexity to $O(F(N)^2 + (F)^2 N)$ :", + "bbox": [ + 89, + 696, + 482, + 787 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\left[ c _ {i} ^ {(1)}, \\hat {E} _ {i} \\right] = \\mathrm {A T T N} \\left(Q K V = \\left[ c _ {i} ^ {(0)}, E _ {i} ^ {(0)} \\right]\\right), \\tag {5}\n$$\n", + "text_format": "latex", + "bbox": [ + 148, + 795, + 482, + 816 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\left[ c _ {1: F} ^ {(2)} \\right] = \\operatorname {A T T N} \\left(Q = \\left[ c _ {1: F} ^ {(0)} \\right], K V = \\left[ c _ {1: F} ^ {(0)}, E _ {1: F} ^ {(0)} \\right]\\right), \\tag {6}\n$$\n", + "text_format": "latex", + "bbox": [ + 112, + 818, + 480, + 839 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\hat {c} _ {i} = c _ {i} ^ {(2)} + \\sigma (c _ {i} ^ {(1)} \\mathbf {W} _ {\\mathrm {d o w n}}) \\mathbf {W} _ {\\mathrm {u p}}, \\tag {7}\n$$\n", + "text_format": "latex", + "bbox": [ + 178, + 840, + 480, + 862 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "where $c_{i}\\in \\mathbb{R}^{1\\times D}$ is CLS token of the $i^{\\mathrm{th}}$ image and $c_{1:F}\\in \\mathbb{R}^{F\\times D}$ are CLS tokens of all sampled images.", + "bbox": [ + 89, + 869, + 482, + 901 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Similarly, $E_{i} \\in \\mathbb{R}^{N \\times D}$ are patch tokens of the $i^{\\text{th}}$ image and $E_{1:F} \\in \\mathbb{R}^{FN \\times D}$ are patch tokens of all sampled images. In Eq. (7), we employ a trainable adapter to merge image-level features $c_{i}^{(1)}$ and video-level features $c_{i}^{(2)}$ , where $\\mathbf{W}_{down} \\in \\mathbb{R}^{D \\times r}$ and $\\mathbf{W}_{up} \\in \\mathbb{R}^{r \\times D}$ represent the up- and down-projection layers, and $\\sigma$ denotes the nonlinear GELU activation [19]. Image-level features preserve pretrained spatial details, while video-level features capture temporal information. This fusion mechanism strengthens feature extraction.", + "bbox": [ + 511, + 422, + 903, + 578 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Our parameter-efficient vision encoder comprises both image-level and IVFusion layers. The shallow image-level layers employ Image-Level Attn to capture spatial information, while the top $H_{v}$ IVFusion layers leverage IVFusion Attn to focus on spatio-temporal information. Consequently, our vision encoder outputs enhanced image features $\\left\\{ v_{i}^{\\mathrm{img}} \\mid 1 \\leq i \\leq F \\right\\}$ .", + "bbox": [ + 511, + 580, + 905, + 696 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "3.4. Pseudo Image-Level Alignment", + "text_level": 1, + "bbox": [ + 511, + 710, + 790, + 726 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Unlike previous work focused on video-level alignment, we introduce PImgAlign to learn fine-grained image-level alignment via the decomposition of video content at the image level. To accomplish this, we first generate pseudo image captions in VTR datasets. In contrast to the straightforward sampling of images from a video, sampling image captions is impractical, as video captions generally lack the specific details needed for individual images. As a solution, we utilize the multimodal large language model (MLLM) LLaVA-NeXT [35] to generate pseudo image captions. For each sampled image from a video, we prompt the model", + "bbox": [ + 511, + 734, + 906, + 901 + ], + "page_idx": 3 + }, + { + "type": "page_number", + "text": "19705", + "bbox": [ + 480, + 944, + 517, + 955 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "using the relevant video caption for guidance:", + "bbox": [ + 89, + 90, + 393, + 106 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "The provided image is a frame sampled from the video, which describes {video caption}. Based on the video's content, provide a caption for the provided image.", + "bbox": [ + 120, + 116, + 450, + 176 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Unlike EA-VTR [38], which uses an image captioner, our method employs a strong MLLM. By incorporating video caption guidance, it generates higher-quality image captions that align with the overall video content. Similarly to visual feature extraction, we adopt our proposed IVFusion to extract pseudo image caption features $\\left\\{t_i^{\\mathrm{img}}|1\\leq i\\leq F\\right\\}$ .", + "bbox": [ + 89, + 186, + 482, + 284 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Given image features and pseudo image caption features from a paired video and video caption, we introduce a fine-grained image-level similarity. Since image captions are typically more concise than the images they describe, a single caption may correspond to multiple images, as shown in Figure 4d. Therefore, we calculate the maximum similarity for each image across all image captions and vice versa. Specifically, we computed a similarity matrix between image and image caption features, selecting the maximum value from each row and column to obtain the final image-level similarity:", + "bbox": [ + 89, + 285, + 483, + 450 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\operatorname {S i m} _ {i j - t 2 v} ^ {\\mathrm {i m g}} = \\frac {1}{F} \\sum_ {n = 1} ^ {F} \\max _ {1 \\leq m \\leq F} t _ {i | n} ^ {\\mathrm {i m g} ^ {T}} v _ {j | m} ^ {\\mathrm {i m g}}, \\tag {8}\n$$\n", + "text_format": "latex", + "bbox": [ + 151, + 460, + 482, + 503 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\operatorname {S i m} _ {i j - v 2 t} ^ {\\mathrm {i m g}} = \\frac {1}{F} \\sum_ {m = 1} ^ {F} \\max _ {1 \\leq n \\leq F} t _ {i | n} ^ {\\mathrm {i m g} ^ {T}} v _ {j | m} ^ {\\mathrm {i m g}}, \\tag {9}\n$$\n", + "text_format": "latex", + "bbox": [ + 153, + 506, + 482, + 546 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\operatorname {S i m} _ {i j} ^ {\\mathrm {i m g}} = \\frac {1}{2} \\left[ \\operatorname {S i m} _ {i j - t 2 v} ^ {\\mathrm {i m g}} + \\operatorname {S i m} _ {i j - v 2 t} ^ {\\mathrm {i m g}} \\right], \\tag {10}\n$$\n", + "text_format": "latex", + "bbox": [ + 158, + 549, + 480, + 578 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "where $\\mathrm{Sim}_{ij}^{\\mathrm{img}}$ denotes the image-level similarity between the $i^{\\mathrm{th}}$ video caption and the $j^{\\mathrm{th}}$ video. $t_{i|n}^{\\mathrm{img}}$ denotes the $n^{\\mathrm{th}}$ pseudo image caption features of the $i^{\\mathrm{th}}$ video caption and $v_{j|m}^{\\mathrm{img}}$ denotes the $m^{\\mathrm{th}}$ image features of the $j^{\\mathrm{th}}$ video. PImgAlign achieves video-level alignment from the image level, effectively emphasizing spatial details.", + "bbox": [ + 89, + 589, + 482, + 686 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "However, in the inference stage, the lack of ground truth for video captions prevents pseudo image captions generation. Additionally, MLLM introduces a computational overhead. For fair comparisons with other methods, we exclude PImgAlign during inference.", + "bbox": [ + 89, + 686, + 483, + 763 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "3.5. Image-to-Video Alignment Distillation", + "text_level": 1, + "bbox": [ + 89, + 772, + 419, + 787 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Pretrained CLIP demonstrates a strong capability for image-level alignment. In contrast, video-text retrieval relies on comprehensive spatio-temporal alignment. Therefore, we propose AlignDistill to distill image-level alignment knowledge into video-level alignment, thus mitigating alignment discrepancies. Then, we optimize the Kullback-Leibler divergence [30] between image-level similarity and", + "bbox": [ + 89, + 794, + 483, + 900 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "video-level similarity:", + "bbox": [ + 513, + 90, + 663, + 106 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {S} _ {t 2 v} ^ {\\mathrm {i m g}} = \\left[ \\mathrm {s} _ {i 1} ^ {\\mathrm {i m g}}, \\dots , \\mathrm {s} _ {i N _ {v}} ^ {\\mathrm {i m g}} \\right], \\mathcal {S} _ {v 2 t} ^ {\\mathrm {i m g}} = \\left[ \\mathrm {s} _ {1 i} ^ {\\mathrm {i m g}}, \\dots , \\mathrm {s} _ {N _ {t} i} ^ {\\mathrm {i m g}} \\right], \\tag {11}\n$$\n", + "text_format": "latex", + "bbox": [ + 526, + 117, + 903, + 137 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {S} _ {t 2 v} ^ {\\mathrm {v i d}} = \\left[ \\mathrm {s} _ {i 1} ^ {\\mathrm {v i d}}, \\dots , \\mathrm {s} _ {i N _ {v}} ^ {\\mathrm {v i d}} \\right], \\mathcal {S} _ {v 2 t} ^ {\\mathrm {v i d}} = \\left[ \\mathrm {s} _ {1 i} ^ {\\mathrm {v i d}}, \\dots , \\mathrm {s} _ {N _ {t i}} ^ {\\mathrm {v i d}} \\right], \\tag {12}\n$$\n", + "text_format": "latex", + "bbox": [ + 535, + 138, + 903, + 159 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {L} _ {K L} = \\frac {1}{2} \\left(\\mathbf {K L} \\left(\\mathcal {S} _ {t 2 v} ^ {\\mathrm {i m g}} \\| \\mathcal {S} _ {t 2 v} ^ {\\mathrm {v i d}}\\right) + \\mathbf {K L} \\left(\\mathcal {S} _ {v 2 t} ^ {\\mathrm {i m g}} \\| \\mathcal {S} _ {v 2 t} ^ {\\mathrm {v i d}}\\right)\\right). \\tag {13}\n$$\n", + "text_format": "latex", + "bbox": [ + 524, + 161, + 903, + 189 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "where $\\mathrm{s}_{ij}^{\\mathrm{img}}$ is $\\mathrm{Sim}_{ij}^{\\mathrm{img}}$ in Eq. (10) and $\\mathrm{s}_{ij}^{\\mathrm{vid}}$ is $\\mathrm{Sim}_{ij}^{\\mathrm{vid}}$ in Eq. (1). $S_{t2v}^{\\mathrm{img}}$ and $S_{t2v}^{\\mathrm{vid}}$ denote the probability distributions for the text-to-video task, while $S_{v2t}^{\\mathrm{img}}$ and $S_{v2t}^{\\mathrm{vid}}$ denote those for the video-to-text task. During the inference phase, We employ enhanced video-level similarity for retrieval tasks.", + "bbox": [ + 511, + 200, + 903, + 284 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Finally, the overall training objective, which includes alignment loss (Eq. (4)) and distillation loss (Eq. (13)), is formulated as follows:", + "bbox": [ + 511, + 285, + 903, + 330 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {L} = \\mathcal {L} _ {A} (\\operatorname {S i m} ^ {\\mathrm {v i d}}) + \\alpha \\mathcal {L} _ {A} (\\operatorname {S i m} ^ {\\mathrm {i m g}}) + \\beta \\mathcal {L} _ {K L}, \\tag {14}\n$$\n", + "text_format": "latex", + "bbox": [ + 545, + 342, + 903, + 361 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "where $\\alpha$ and $\\beta$ are used to balance the loss. $\\mathcal{L}_A(\\mathrm{Sim}^{\\mathrm{img}})$ is the image-level alignment loss (see Eq. (4) and (10)).", + "bbox": [ + 511, + 375, + 903, + 407 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "4. Experiments", + "text_level": 1, + "bbox": [ + 511, + 422, + 645, + 439 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "4.1. Experimental Settings", + "text_level": 1, + "bbox": [ + 511, + 448, + 720, + 465 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Following common practice, we perform evaluations on four widely used benchmarks: MSRVTT [53], LSMDC [43], ActivityNet [29], and DiDeMo [1]. We evaluate the performance using common retrieval metrics such as Recall at K (R@K and K = 1, 5, 10), the sum of these recalls (R@sum), and Mean Rank (MnR). In all experiments, the LoRA dimension and the adapter dimension $r$ are set to 8. In Eq. 14, $\\alpha$ and $\\beta$ are set to 0.3 and 1.0, respectively. For the number of IVFusion layers, we set $H_V = 4$ for the vision encoder and $H_L = 2$ for the text encoder. Please see Appendix A for further details.", + "bbox": [ + 511, + 470, + 906, + 637 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "4.2. Comparisons with State-of-the-Art Methods", + "text_level": 1, + "bbox": [ + 511, + 648, + 888, + 665 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "We compare our proposed DiscoVLA with state-of-the-art methods on popular benchmarks such as MSRVTT, LSMDC, ActivityNet, and DiDeMo.", + "bbox": [ + 511, + 672, + 903, + 717 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Full fine-tuning refers to CLIP4Clip-meanP [37], which averages image features along the temporal dimension. All parameter-efficient methods follow this temporal average pooling strategy to obtain the final video features. Although Prompt [27], Adapter [20], and LoRA [21] are widely adopted for parameter-efficient image and NLP tasks, they lack effective temporal modeling for video-level understanding. In contrast, methods like RAP [5], VoP [22], DGL [58], and MV-Adapter [25] focus on temporal modeling in the vision modality.", + "bbox": [ + 511, + 718, + 905, + 869 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "As shown in Table 1, on MSRVTT, our DiscoVLA with CLIP (ViT-B/32) as the backbone achieves $47.0\\%$ R@1 in", + "bbox": [ + 511, + 869, + 905, + 900 + ], + "page_idx": 4 + }, + { + "type": "page_number", + "text": "19706", + "bbox": [ + 480, + 944, + 519, + 955 + ], + "page_idx": 4 + }, + { + "type": "table", + "img_path": "images/811aca382ac40874239e8f66b6a50d4e99c1747b898de85248e9097c9bdde09a.jpg", + "table_caption": [], + "table_footnote": [], + "table_body": "
Method# Params (M)Text-to-VideoVideo-to-Text
R@1↑R@5↑R@10↑R@sum↑MnR↓R@1↑R@5↑R@10↑R@sum↑MnR↓
CLIP (ViT-B/32)
Full fine-tuning123.5443.170.480.8194.316.243.170.581.2194.812.4
Prompt [27]0.0840.466.377.3184.016.742.269.779.2191.112.4
Adapter [20]0.2641.969.978.7190.214.943.669.980.1193.611.5
LoRA [21]0.4943.768.980.4193.016.043.070.282.2195.412.0
RAP [5]1.0644.871.481.5197.714.444.071.982.4198.310.1
VoPF+P [22]0.443.569.379.3192.114.843.671.281.2196.011.0
VoPF+C [22]14.1044.770.579.2194.416.242.170.080.6192.713.4
DGL [58]0.8344.669.980.3194.816.344.570.780.6195.811.5
DiscoVLA0.5647.073.082.8202.814.147.773.683.6204.910.0
CLIP (ViT-B/16)
MV-Adapter [25]3.646.072.082.1200.1-45.674.083.8203.4-
RAP [5]1.0646.573.982.0202.412.145.376.484.8206.59.1
VoPF+P [22]0.447.172.481.8201.312.9-----
VoPF+C [22]14.1047.772.482.2202.312.0-----
DGL [58]0.8348.371.880.6200.713.445.774.082.9202.610.9
TempMe [44]0.5049.074.483.3206.711.947.675.385.4208.39.0
DiscoVLA0.5650.575.683.8209.912.149.276.084.7209.98.6
", + "bbox": [ + 117, + 88, + 875, + 353 + ], + "page_idx": 5 + }, + { + "type": "table", + "img_path": "images/a5a7903817c034caa677995764389cb1cc1e597cb314aacec7b9fc5e3bfc0cca.jpg", + "table_caption": [ + "Table 1. Comparisons with state-of-the-art methods on MSRVTT. # Params denotes the number of trainable parameters. $\\uparrow$ denotes higher values represent better performance and $\\downarrow$ denotes lower values represent better performance. The best value for each metric is highlighted in bold. R@sum is defined as the total of R@1, R@5, and R@10. The gray row represents the fully fine-tuned CLIP4Clip-meanP [37]." + ], + "table_footnote": [], + "table_body": "
Method# Params (M)Text-to-VideoVideo-to-Text
R@1↑R@5↑R@10↑R@sum↑MnR↓R@1↑R@5↑R@10↑R@sum↑MnR↓
Full fine-tuning123.5420.738.947.2106.865.320.639.447.5107.556.7
VoPF+P [22]0.420.740.749.7111.159.121.540.650.7112.850.8
VoPF+C [22]14.1021.140.949.6111.660.122.340.350.7113.351.1
DGL [58]0.8321.439.448.4109.264.3-----
DiscoVLA0.5623.642.052.3117.952.022.841.951.2115.946.6
", + "bbox": [ + 122, + 402, + 870, + 500 + ], + "page_idx": 5 + }, + { + "type": "table", + "img_path": "images/bd378c525d80127c71fa8ac9d0c05c1920522dc51787a2993d585255365a49ab.jpg", + "table_caption": [ + "Table 2. Comparisons with state-of-the-art methods on LSMDC using CLIP (ViT-B/32)." + ], + "table_footnote": [], + "table_body": "
Method# Params (M)Text-to-VideoVideo-to-Text
R@1↑R@5↑R@10↑R@sum↑MnR↓R@1↑R@5↑R@10↑R@sum↑MnR↓
Full fine-tuning123.5440.572.4--7.442.574.185.8202.46.6
RAP [5]1.0640.871.082.2194.08.3-----
VoPF+P [22]0.436.165.578.5180.110.936.365.979.2181.410.1
VoPF+C [22]14.1035.163.777.6176.411.435.665.977.8179.310.4
DGL [58]0.8338.669.281.6189.49.0-----
DiscoVLA0.5641.272.483.6197.27.841.872.884.7199.36.9
", + "bbox": [ + 122, + 518, + 870, + 631 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Table 3. Comparisons with state-of-the-art methods on ActivityNet using CLIP (ViT-B/32).", + "bbox": [ + 225, + 631, + 767, + 645 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "the text-to-video task (t2v) and $47.7\\%$ R@1 in the video-to-text task (v2t), significantly outperforming previous methods. Furthermore, when using CLIP (ViT-B/16), DiscoVLA achieves improvements of $2.2\\%$ R@1 and $7.5\\%$ R@sum. The comparison results on LSMDC, ActivityNet, DiDeMo are shown in Table 2-4. Our method outperforms all existing approaches across these datasets. Existing approaches overlook the critical image-to-video language and alignment discrepancies. In contrast, our proposed DiscoVLA addresses these discrepancies across all three areas: vision, language, and alignment.", + "bbox": [ + 88, + 659, + 482, + 825 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "4.3. Ablation Study", + "text_level": 1, + "bbox": [ + 89, + 833, + 243, + 849 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Ablation study on individual components. We evaluate the proposed components in Table 5. LoRA applies Image-Level Attn (see Figure 3a) in both the vision and text", + "bbox": [ + 89, + 854, + 483, + 902 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "encoders. In B1, IVFusion significantly enhances performance by $8.7\\%$ R@sum in t2v and $6.7\\%$ R@sum in v2t, effectively integrating image-level and video-level features. In B2 and B3, we find that using PImgAlign or AlignDistill independently yields limited improvements. This is largely due to the notable gap between image-level and video-level (see Figure 1). Image-level PImgAlign struggles to impact video-level tasks, while AlignDistill lacks sufficient image-specific learning. When combined, however, they complement each other. PImgAlign captures fine-grained image-level alignments, and AlignDistill effectively transfers this knowledge to the video-level alignment. This improves $1.1\\%$ R@sum in t2v and $2.8\\%$ R@sum in v2t over B1. Finally, our full DiscoVLA improves $9.8\\%$ R@sum in t2v and $9.5\\%$ R@sum in v2t over LoRA.", + "bbox": [ + 511, + 659, + 906, + 885 + ], + "page_idx": 5 + }, + { + "type": "page_number", + "text": "19707", + "bbox": [ + 480, + 944, + 519, + 957 + ], + "page_idx": 5 + }, + { + "type": "table", + "img_path": "images/a95bab51d49738d0843e405888fe37e2f953209995ba6bc1fa3e9b927bce5000.jpg", + "table_caption": [], + "table_footnote": [], + "table_body": "
Method# Params (M)Text-to-VideoVideo-to-Text
R@1↑R@5↑R@10↑R@sum↑MnR↓R@1↑R@5↑R@10↑R@sum↑MnR↓
Full fine-tuning123.5443.470.280.6194.217.542.570.680.2193.311.6
RAP [5]1.0642.670.479.6192.618.0-----
VoPF+P [22]0.445.372.380.4198.013.844.771.281.1197.09.9
VoPF+C [22]14.1046.471.981.5199.813.644.471.881.8198.09.5
DiscoVLA0.5648.474.582.7205.614.047.774.483.8205.99.3
", + "bbox": [ + 122, + 88, + 870, + 186 + ], + "page_idx": 6 + }, + { + "type": "table", + "img_path": "images/4de793244a6ffc5548b625922d297337421431621587b97d1fbed686874bcc6f.jpg", + "table_caption": [ + "Table 4. Comparisons with state-of-the-art methods on DiDeMo using CLIP (ViT-B/32)." + ], + "table_footnote": [], + "table_body": "
MethodsComponents# Params (M)Text-to-VideoVideo-to-Text
IVFusionPIImgAlignAlignDistillR@1↑R@5↑R@10↑R@sum↑R@1↑R@5↑R@10↑R@sum↑
Pretraind CLIPXXX030.853.863.3147.926.650.162.0138.7
LoRAXXX0.4943.768.980.4193.043.070.282.2195.4
B1✓XX0.5446.573.282.0201.746.672.682.9202.1
B2✓✓X0.5647.073.181.3201.445.772.982.4201.0
B3✓X✓0.5446.671.681.0199.245.971.982.2200.0
DiscoVLA✓✓✓0.5647.073.082.8202.847.773.683.6204.9
", + "bbox": [ + 94, + 205, + 903, + 314 + ], + "page_idx": 6 + }, + { + "type": "table", + "img_path": "images/7eb3e177f63cafd83480656c66b7d4ac4ec410146817651e40e015edc4dea50f.jpg", + "table_caption": [ + "Table 5. Ablation study on the contribution of each proposed component on MSRVTT using CLIP (ViT-B/32). Pretrained CLIP denotes the zero-shot performance of CLIP without any additional training. Our proposed methods are built upon LoRA. B1, B2, and B3 denote various combinations of our proposed components." + ], + "table_footnote": [], + "table_body": "
Parameter-Efficient AttentionText-to-Video
R@1↑R@5↑R@10↑R@sum↑
Image-Level Attn43.768.980.4193.0
Video-Level Attn45.770.980.8197.4
IVFusion Attn w/o Adapter46.071.080.4197.4
IVFusion Attn46.573.282.0201.7
", + "bbox": [ + 94, + 366, + 480, + 446 + ], + "page_idx": 6 + }, + { + "type": "table", + "img_path": "images/ef4fc948a9d8c1a36962f76c8d7c80eec8c9313c0e05ca334a81f4e5047037b2.jpg", + "table_caption": [ + "Table 6. Ablation study on IVFusion on MSRVTT using CLIP (ViT-B/32). We compare several implementations against our proposed IVFusion for Parameter-Efficient Attention (see Figure 3). Image-Level Attn represents the LoRA baseline." + ], + "table_footnote": [], + "table_body": "
Similarity in PImgAlignText-to-Video
R@1↑R@5↑R@10↑R@sum↑
None46.671.681.0199.2
Video-Level46.372.381.2199.8
Paired Image-Level46.173.282.4201.7
Fine-grained Image-Level47.073.082.8202.8
", + "bbox": [ + 94, + 508, + 480, + 589 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Analysis of trainable parameters. The trainable parameters of our DiscoVLA consist of the LoRA and Adapter parts. As shown in Table 5, LoRA in both the vision and text encoders contain $\\sim 0.49\\mathrm{M}$ parameters. In B1, an additional $\\sim 0.05\\mathrm{M}$ parameters are introduced by the Adapter in the $H_{V} = 4$ IVFusion layers of the vision encoder. In B2, an additional $\\sim 0.02\\mathrm{M}$ parameters are introduced by the Adapter in the $H_{L} = 2$ IVFusion layers of the text encoder, which are utilized for pseudo-image caption features in PImgAlign. Overall, the total number of trainable parameters of DiscoVLA is $\\sim 0.56\\mathrm{M}$ .", + "bbox": [ + 89, + 674, + 482, + 838 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Ablation study on IVFusion. In Table 6, we compare different strategies for parameter-efficient attention in Figure 3. Video-Level Attn concatenates all tokens across sampled images to capture spatio-temporal information,", + "bbox": [ + 89, + 839, + 483, + 901 + ], + "page_idx": 6 + }, + { + "type": "table", + "img_path": "images/cabb0ee4fb81f45229bc4c29854fd64c6b8412b318c92742c1c3e31c7de73074.jpg", + "table_caption": [ + "Table 7. Ablation study on similarity in PImgAlign on MSRVTT using CLIP (ViT-B/32). None represents DiscoVLA without image-level alignment optimization, shown as B3 in Table 5. Paired Image-Level is based on matched image-caption pairs. Fine-grained Image-Level represents the full DiscoVLA." + ], + "table_footnote": [], + "table_body": "
Image Caption GenerationText-to-Video
R@1↑R@5↑R@10↑R@sum↑
DiscoVLA47.073.082.8202.8
W/O VidCap46.1 (-0.9)73.082.1 (-0.7)201.2 (-1.6)
", + "bbox": [ + 516, + 366, + 903, + 425 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Table 8. Ablation study on image caption generation on MSRVTT using CLIP (ViT-B/32). W/O VidCap denotes generating pseudo image captions without video caption guidance.", + "bbox": [ + 511, + 426, + 903, + 468 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "achieving a notable $4.4\\%$ R@sum improvement. To reduce the computational complexity, IVFusion applies spatiotemporal attention only to global CLS tokens and merges image-level features with video-level features. Without an adapter, averaging image- and video-level features results in a modest performance increase. In contrast, with a learnable adapter, IVFusion merges the features more effectively, achieving $201.7\\%$ R@sum.", + "bbox": [ + 511, + 489, + 906, + 609 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Ablation study on similarity in PImgAlign. Table 7 demonstrates the effectiveness of our proposed fine-grained image-level similarity in PImgAlign. The similarity is computed between sampled images and pseudo image captions from a paired video and video caption. The Video-Level baseline averages the image features and pseudo-image caption features along the temporal dimension to compute the video-level similarity, replacing fine-grained image-level similarity. However, this modification leads to a notable decrease in performance, emphasizing the importance of image-level similarity in PImgAlign. In contrast, we introduce the Paired Image-Level baseline, which computes similarity based on matched image-caption pairs. Compared to Video-Level, Paired Image-Level yields a $1.9\\%$ gain in R@sum. Moreover, we observe in Figure 4d that a single image caption may describe multiple images. Consequently, our fine-grained image-level similarity, as defined in Eq. 10, selects the maximum similarity from all available pairs, further enhancing video-level alignment.", + "bbox": [ + 511, + 613, + 908, + 902 + ], + "page_idx": 6 + }, + { + "type": "page_number", + "text": "19708", + "bbox": [ + 480, + 944, + 517, + 955 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/0b53ce37bcdb38f14d71a1c9fbed0ebd0a0de2a600f4af6fc821dc44e4c1a3bc.jpg", + "image_caption": [ + "Figure 4. Visualization of discrepancies between video-level and image-level data. Each example consists of a paired video and video caption, with four sampled images from the video. Video captions are highlighted with orange solid lines, and pseudo image captions generated by LLaVA-NeXT [35] are indicated with green dashed lines. While video captions convey the general context of the video, image captions focus on the detailed context of each individual frame." + ], + "image_footnote": [], + "bbox": [ + 89, + 102, + 903, + 392 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/7e4c331fad03523b338be4fd02a3c63a13d7db79123d524fa9d44dff7d422ba4.jpg", + "image_caption": [ + "Figure 5. Comparison of image caption generation with (W/) and without (W/O) video caption guidance (VidCap) in PImgAlign." + ], + "image_footnote": [], + "bbox": [ + 93, + 469, + 482, + 585 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Ablation study on image caption generation in PImgAlign. Figure 5 and Table 8 show qualitative and quantitative evaluations of video caption guidance. In Figure 5, video captions serve as the global context that enriches the understanding of individual frames (marked in red). In Table 8, without this guidance, image captions focus only on salient regions while missing globally relevant details, leading to performance drops. In PImgAlign, we employ MLLM to process both frame images and video descriptions, generating image captions that align with the overall video content. These high-quality captions further enhance DiscoVLA's ability to learn robust image-level alignment.", + "bbox": [ + 88, + 641, + 482, + 823 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "More Ablation Analysis. In Appendix B, we conduct additional ablation studies. Specifically, we evaluate the effects of hyperparameters $\\alpha$ , $\\beta$ , $H_V$ and $H_V$ . For fairness, all results in the main paper are reported without post-processing. Complementary evaluations with Q-Norm [4] and DSL [8]", + "bbox": [ + 88, + 824, + 482, + 902 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "as post-processing methods are provided in the appendix.", + "bbox": [ + 511, + 472, + 890, + 488 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "4.4. Qualitative Analysis", + "text_level": 1, + "bbox": [ + 511, + 494, + 705, + 511 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "To compare image-level and video-level data, we conduct a visualization experiment in Figure 4. In Figure 4a and 4b, image-level data alone fails to capture the complete content of the video. Figure 4c and 4d demonstrate a strong correlation between the video-level and image-level content, but differences across images lead to varying image captions. Visually, a video is composed of a sequence of images, incorporating temporal information. Textually, image captions are specific to individual images, while video captions describe the overall video content. These discrepancies hinder the effective transfer of image-level CLIP. Our approach addresses this issue by utilizing image-level alignment to enhance video-level alignment.", + "bbox": [ + 511, + 516, + 906, + 714 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "5. Conclusion", + "text_level": 1, + "bbox": [ + 511, + 724, + 633, + 739 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "In this paper, we identify three major discrepancies in the parameter-efficient transfer of image-level CLIP to videotext retrieval: vision, language, and alignment. To address these challenges, we propose DiscoVLA. Specifically, IV-Fusion fuses image-level and video-level features to tackle the vision and language discrepancies. For alignment discrepancies, we propose PImgAlign and AlignDistill, which learn and distill image-level alignment to enhance video-level alignment. Our DiscoVLA significantly outperforms existing methods, achieving state-of-the-art performance.", + "bbox": [ + 511, + 750, + 906, + 902 + ], + "page_idx": 7 + }, + { + "type": "header", + "text": "Video Caption - Image Caption", + "bbox": [ + 148, + 88, + 352, + 99 + ], + "page_idx": 7 + }, + { + "type": "page_number", + "text": "19709", + "bbox": [ + 480, + 944, + 517, + 955 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Acknowledgment", + "text_level": 1, + "bbox": [ + 91, + 90, + 240, + 107 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "References", + "text_level": 1, + "bbox": [ + 91, + 280, + 187, + 296 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[1] Lisa Anne Hendricks, Oliver Wang, Eli Shechtman, Josef Sivic, Trevor Darrell, and Bryan Russell. Localizing moments in video with natural language. In Proceedings of the International Conference on Computer Vision, 2017. 5, 12", + "[2] Stanislaw Antol, Aishwarya Agrawal, Jiasen Lu, Margaret Mitchell, Dhruv Batra, C Lawrence Zitnick, and Devi Parikh. Vqa: Visual question answering. In Proceedings of the International Conference on Computer Vision, 2015. 2", + "[3] Max Bain, Arsha Nagrani, Gül Varol, and Andrew Zisserman. Frozen in time: A joint video and image encoder for end-to-end retrieval. In Proceedings of the International Conference on Computer Vision, 2021. 12", + "[4] Simion-Vlad Bogolin, Ioana Croitoru, Hailin Jin, Yang Liu, and Samuel Albanie. Cross modal retrieval with querybank normalisation. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition, 2022. 8, 13", + "[5] Meng Cao, Haoran Tang, Jinfa Huang, Peng Jin, Can Zhang, Ruyang Liu, Long Chen, Xiaodan Liang, Li Yuan, and Ge Li. RAP: Efficient text-video retrieval with sparse-and-correlated adapter. In Association for Computational Linguistics, 2024. 2, 3, 5, 6, 7, 12", + "[6] Shoufa Chen, Chongjian Ge, Zhan Tong, Jiangliu Wang, Yibing Song, Jue Wang, and Ping Luo. Adaptformer: Adapting vision transformers for scalable visual recognition. Advances in Neural Information Processing Systems, 2022. 2", + "[7] Sihan Chen, Handong Li, Qunbo Wang, Zijia Zhao, Mingzhen Sun, Xinxin Zhu, and Jing Liu. Vast: A vision-audio-subtitle-text omni-modality foundation model and dataset. Advances in Neural Information Processing Systems, 2023. 2", + "[8] Xing Cheng, Hezheng Lin, Xiangyu Wu, Fan Yang, and Dong Shen. Improving video-text retrieval by multi-stream corpus alignment and dual softmax loss. arXiv preprint arXiv:2109.04290, 2021. 8, 13", + "[9] Chaorui Deng, Qi Chen, Pengda Qin, Da Chen, and Qi Wu. Prompt switch: Efficient clip adaptation for text-video retrieval. In Proceedings of the International Conference on Computer Vision, 2023. 1, 2", + "[10] Xiaohan Ding, Yuchen Guo, Guiguang Ding, and Jungong Han. Acnet: Strengthening the kernel skeletons for powerful cnn via asymmetric convolution blocks. In Proceedings of the International Conference on Computer Vision, 2019. 2" + ], + "bbox": [ + 93, + 305, + 483, + 900 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "This work was supported by Beijing Natural Science Foundation (No. L223023), National Natural Science Foundation of China (Nos. 62441235, 62021002, 62441614), the Key R & D Program of Xinjiang, China (2022B01006), Zhejiang Provincial Natural Science Foundation of China under Grant (No. LDT23F01013F01), CCF-DiDi GAIA Collaborative Research Funds, China Postdoctoral Science Foundation (2024M750565), Guangdong S & T Program (2024B0101040008), and the Key Realm Research and Development Program of Guangzhou (No. 2024B01W0007).", + "[11] Xiaohan Ding, Xiangyu Zhang, Ningning Ma, Jungong Han, Guiguang Ding, and Jian Sun. Repvgg: Making vgg-style convnets great again. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition, 2021. 2", + "[12] Lijie Fan, Dilip Krishnan, Phillip Isola, Dina Katabi, and Yonglong Tian. Improving clip training with language rewrites. Advances in Neural Information Processing Systems, 2024. 2", + "[13] Valentin Gabeur, Chen Sun, Karteek Alahari, and Cordelia Schmid. Multi-modal transformer for video retrieval. In Proceedings of the European Conference on Computer Vision, 2020. 12", + "[14] Satya Krishna Gorti, Noel Vouitsis, Junwei Ma, Keyvan Golestan, Maksims Volkovs, Animesh Garg, and Guangwei Yu. X-pool: Cross-modal language-video attention for text-video retrieval. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition, 2022. 1, 2", + "[15] Priya Goyal, Piotr Dólar, Ross Girshick, Pieter Noordhuis, Lukasz Wesolowski, Aapo Kyrola, Andrew Tulloch, Yangqing Jia, and Kaiming He. Accurate, large minibatch sgd: Training imagenet in 1 hour. arXiv preprint arXiv:1706.02677, 2017. 12", + "[16] Alex Graves and Jürgen Schmidhuber. Framewise phoneme classification with bidirectional LSTM and other neural network architectures. Neural Networks, 2005. 2", + "[17] Junxian He, Chunting Zhou, Xuezhe Ma, Taylor Berg-Kirkpatrick, and Graham Neubig. Towards a unified view of parameter-efficient transfer learning. In Proceedings of the International Conference on Learning Representations, 2021. 2", + "[18] Tao He, Leqi Shen, Yuchen Guo, Guiguang Ding, and Zhenhua Guo. Secret: Self-consistent pseudo label refinement for unsupervised domain adaptive person re-identification. In Proceedings of the AAAI Conference on Artificial Intelligence, 2022. 2", + "[19] Dan Hendrycks and Kevin Gimpel. Gaussian error linear units (gelus). arXiv preprint arXiv:1606.08415, 2016. 4", + "[20] Neil Houlsby, Andrei Giurgiu, Stanislaw Jastrzebski, Bruna Morrone, Quentin De Laroussilhe, Andrea Gesmundo, Mona Attariyan, and Sylvain Gelly. Parameter-efficient transfer learning for nlp. In Proceedings of the International Conference on Machine Learning, 2019. 2, 5, 6", + "[21] Edward J Hu, Yelong Shen, Phillip Wallis, Zeyuan Allen-Zhu, Yuanzhi Li, Shean Wang, Lu Wang, and Weizhu Chen. Lora: Low-rank adaptation of large language models. arXiv preprint arXiv:2106.09685, 2021. 2, 5, 6", + "[22] Siteng Huang, Biao Gong, Yulin Pan, Jianwen Jiang, Yiliang Lv, Yuyuan Li, and Donglin Wang. Vop: Text-video cooperative prompt tuning for cross-modal retrieval. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition, 2023. 1, 2, 3, 5, 6, 7, 12", + "[23] Menglin Jia, Luming Tang, Bor-Chun Chen, Claire Cardie, Serge Belongie, Bharath Hariharan, and Ser-Nam Lim. Visual prompt tuning. In Proceedings of the European Conference on Computer Vision, 2022. 2", + "[24] Peng Jin, Jina Huang, Pengfei Xiong, Shangxuan Tian, Chang Liu, Xiangyang Ji, Li Yuan, and Jie Chen. Videotext as game players: Hierarchical banzhaf interaction for" + ], + "bbox": [ + 93, + 92, + 906, + 900 + ], + "page_idx": 8 + }, + { + "type": "page_number", + "text": "19710", + "bbox": [ + 480, + 944, + 519, + 955 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "cross-modal representation learning. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition, 2023. 1, 2", + "[25] Xiaojie Jin, Bowen Zhang, Weibo Gong, Kai Xu, Xueqing Deng, Peng Wang, Zhao Zhang, Xiaohui Shen, and Jiashi Feng. Mv-adapter: Multimodal video transfer learning for video text retrieval. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition, 2024. 2, 3, 5, 6, 12", + "[26] Andrej Karpathy and Li Fei-Fei. Deep visual-semantic alignments for generating image descriptions. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition, 2015. 2", + "[27] Muhammad Uzair Khattak, Hanoona Rasheed, Muhammad Maaz, Salman Khan, and Fahad Shahbaz Khan. Maple: Multi-modal prompt learning. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition, 2023. 2, 5, 6", + "[28] Muhammad Uzair Khattak, Syed Talal Wasim, Muzammal Naseer, Salman Khan, Ming-Hsuan Yang, and Fahad Shahbaz Khan. Self-regulating prompts: Foundational model adaptation without forgetting. In Proceedings of the International Conference on Computer Vision, 2023. 2", + "[29] Ranjay Krishna, Kenji Hata, Frederic Ren, Li Fei-Fei, and Juan Carlos Niebles. Dense captioning events in videos. In Proceedings of the International Conference on Computer Vision, 2017. 5, 12", + "[30] Solomon Kullback. Information theory and statistics. Courier Corporation, 1997. 5", + "[31] Jie Lei, Linjie Li, Luowei Zhou, Zhe Gan, Tamara L Berg, Mohit Bansal, and Jingjing Liu. Less is more: Clipbert for video-and-language learning via sparse sampling. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition, 2021. 12", + "[32] Junnan Li, Ramprasaath Selvaraju, Akhilesh Gotmare, Shafiq Joty, Caiming Xiong, and Steven Chu Hong Hoi. Align before fuse: Vision and language representation learning with momentum distillation. Advances in Neural Information Processing Systems, 2021. 2", + "[33] Junnan Li, Dongxu Li, Caiming Xiong, and Steven Hoi. Blip: Bootstrapping language-image pre-training for unified vision-language understanding and generation. In Proceedings of the International Conference on Machine Learning, 2022.", + "[34] Kunchang Li, Yali Wang, Yizhuo Li, Yi Wang, Yinan He, Limin Wang, and Yu Qiao. Unmasked teacher: Towards training-efficient video foundation models. In Proceedings of the International Conference on Computer Vision, 2023. 2", + "[35] Haotian Liu, Chunyuan Li, Yuheng Li, Bo Li, Yuanhan Zhang, Sheng Shen, and Yong Jae Lee. Llava next: Improved reasoning,ocr,and world knowledge,2024.4,8", + "[36] Ilya Loshchilov and Frank Hutter. Sgdr: Stochastic gradient descent with warm restarts. In Proceedings of the International Conference on Learning Representations, 2016. 12", + "[37] Huaishao Luo, Lei Ji, Ming Zhong, Yang Chen, Wen Lei, Nan Duan, and Tianrui Li. Clip4clip: An empirical study of" + ], + "bbox": [ + 91, + 90, + 483, + 901 + ], + "page_idx": 9 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "clip for end to end video clip retrieval and captioning. Neurocomputing, 2022. 1, 2, 5, 6", + "[38] Zongyang Ma, Ziqi Zhang, Yuxin Chen, Zhongang Qi, Chunfeng Yuan, Bing Li, Yingmin Luo, Xu Li, Xiaojuan Qi, Ying Shan, et al. Ea-vtr: Event-aware video-text retrieval. In Proceedings of the European Conference on Computer Vision, 2024. 5", + "[39] Antoine Miech, Dimitri Zhukov, Jean-Baptiste Alayrac, Makarand Tapaswi, Ivan Laptev, and Josef Sivic. Howto100m: Learning a text-video embedding by watching hundred million narrated video clips. In Proceedings of the International Conference on Computer Vision, 2019. 12", + "[40] Aaron van den Oord, Yazhe Li, and Oriol Vinyals. Representation learning with contrastive predictive coding. arXiv preprint arXiv:1807.03748, 2018. 3", + "[41] Alec Radford, Jeffrey Wu, Rewon Child, David Luan, Dario Amodei, Ilya Sutskever, et al. Language models are unsupervised multitask learners. OpenAI blog, 2019. 2", + "[42] Alec Radford, Jong Wook Kim, Chris Hallacy, Aditya Ramesh, Gabriel Goh, Sandhini Agarwal, Girish Sastry, Amanda Askell, Pamela Mishkin, Jack Clark, et al. Learning transferable visual models from natural language supervision. In Proceedings of the International Conference on Machine Learning, 2021. 1, 2", + "[43] Anna Rohrbach, Marcus Rohrbach, and Bernt Schiele. The long-short story of movie description. In Pattern Recognition, 2015. 5, 12", + "[44] Leqi Shen, Tianxiang Hao, Sicheng Zhao, Yifeng Zhang, Pengzhang Liu, Yongjun Bao, and Guiguang Ding. Tempme: Video temporal token merging for efficient text-video retrieval. arXiv preprint arXiv:2409.01156, 2024. 6", + "[45] Leqi Shen, Tao He, Sicheng Zhao, Zhelun Shen, Yuchen Guo, Tianshi Xu, and Guiguang Ding. X-reid: Cross-instance transformer for identity-level person re-identification. In IEEE International Conference on Multimedia and Expo, 2024. 2", + "[46] Yoad Tewel, Yoav Shalev, Idan Schwartz, and Lior Wolf. Zerocap: Zero-shot image-to-text generation for visual-semantic arithmetic. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition, 2022. 2", + "[47] Ao Wang, Hui Chen, Lihao Liu, Kai Chen, Zijia Lin, Jungong Han, et al. Yolov10: Real-time end-to-end object detection. Advances in Neural Information Processing Systems, 2024. 2", + "[48] Jiamian Wang, Guohao Sun, Pichao Wang, Dongfang Liu, Sohail Dianat, Majid Rabbani, Raghuveer Rao, and Zhiqiang Tao. Text is mass: Modeling as stochastic embedding for text-video retrieval. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition, 2024. 1, 2", + "[49] Qiang Wang, Yanhao Zhang, Yun Zheng, Pan Pan, and Xi-ansheng Hua. Disentangled representation learning for text-video retrieval. arXiv preprint arXiv:2203.07111, 2022. 1, 2", + "[50] Yi Wang, Kunchang Li, Yizhuo Li, Yinan He, Bingkun Huang, Zhiyu Zhao, Hongjie Zhang, Jilan Xu, Yi Liu, Zun Wang, et al. Intervideo: General video foundation models via generative and discriminative learning. arXiv preprint arXiv:2212.03191, 2022. 2" + ], + "bbox": [ + 516, + 92, + 903, + 898 + ], + "page_idx": 9 + }, + { + "type": "page_number", + "text": "19711", + "bbox": [ + 480, + 944, + 517, + 955 + ], + "page_idx": 9 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[51] Wenhao Wu, Haipeng Luo, Bo Fang, Jingdong Wang, and Wanli Ouyang. Cap4video: What can auxiliary captions do for text-video retrieval? In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition, 2023. 1, 2", + "[52] Hu Xu, Saining Xie, Xiaqing Tan, Po-Yao Huang, Russell Howes, Vasu Sharma, Shang-Wen Li, Gargi Ghosh, Luke Zettlemoyer, and Christoph Feichtenhofer. Demystifying clip data. In Proceedings of the International Conference on Learning Representations, 2024. 2", + "[53] Jun Xu, Tao Mei, Ting Yao, and Yong Rui. Msr-vtt: A large video description dataset for bridging video and language. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition, 2016. 5, 12", + "[54] Kelvin Xu, Jimmy Ba, Ryan Kiros, Kyunghyun Cho, Aaron Courville, Ruslan Salakhudinov, Rich Zemel, and Yoshua Bengio. Show, attend and tell: Neural image caption generation with visual attention. In Proceedings of the International Conference on Machine Learning, 2015. 2", + "[55] Hongwei Xue, Yuchong Sun, Bei Liu, Jianlong Fu, Ruihua Song, Houqiang Li, and Jiebo Luo. Clip-vip: Adapting pre-trained image-text model to video-language alignment. In Proceedings of the International Conference on Learning Representations, 2022. 2", + "[56] Fan Yang, Sicheng Zhao, Yanhao Zhang, Haoxiang Chen, Hui Chen, Wenbo Tang, Haonan Lu, Pengfei Xu, Zhenyu Yang, Jungong Han, et al. Lmi3d: Empowering llm with 3d perception from a single 2d image. arXiv preprint arXiv:2408.07422, 2024. 2", + "[57] Fan Yang, Ru Zhen, Jianing Wang, Yanhao Zhang, Haoxiang Chen, Haonan Lu, Sicheng Zhao, and Guiguang Ding. Heie: Mllm-based hierarchical explainable aigc image implausibility evaluator. arXiv preprint arXiv:2411.17261, 2024. 2", + "[58] Xiangpeng Yang, Linchao Zhu, Xiaohan Wang, and Yi Yang. Dgl: Dynamic global-local prompt tuning for text-video retrieval. Proceedings of the AAAI Conference on Artificial Intelligence, 2024. 1, 2, 3, 5, 6, 12", + "[59] Bowen Zhang, Hexiang Hu, and Fei Sha. Cross-modal and hierarchical modeling of video and text. In Proceedings of the European Conference on Computer Vision, 2018. 12", + "[60] Yuanhan Zhang, Kaiyang Zhou, and Ziwei Liu. Neural prompt search. arXiv preprint arXiv:2206.04673, 2022. 2", + "[61] Kaiyang Zhou, Jingkang Yang, Chen Change Loy, and Ziwei Liu. Conditional prompt learning for vision-language models. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition, 2022. 2", + "[62] Kaiyang Zhou, Jingkang Yang, Chen Change Loy, and Ziwei Liu. Learning to prompt for vision-language models. International Journal of Computer Vision, 2022. 2" + ], + "bbox": [ + 91, + 90, + 480, + 782 + ], + "page_idx": 10 + }, + { + "type": "page_number", + "text": "19712", + "bbox": [ + 480, + 944, + 517, + 955 + ], + "page_idx": 10 + } +] \ No newline at end of file diff --git a/2025/DiscoVLA_ Discrepancy Reduction in Vision, Language, and Alignment for Parameter-Efficient Video-Text Retrieval/9d7a1765-54c3-41c7-9719-3ce7d4a7b6fa_model.json b/2025/DiscoVLA_ Discrepancy Reduction in Vision, Language, and Alignment for Parameter-Efficient Video-Text Retrieval/9d7a1765-54c3-41c7-9719-3ce7d4a7b6fa_model.json new file mode 100644 index 0000000000000000000000000000000000000000..e64cdcfd43bc099153da10e9a290d53eeca0fae9 --- /dev/null +++ b/2025/DiscoVLA_ Discrepancy Reduction in Vision, Language, and Alignment for Parameter-Efficient Video-Text Retrieval/9d7a1765-54c3-41c7-9719-3ce7d4a7b6fa_model.json @@ -0,0 +1,2323 @@ +[ + [ + { + "type": "header", + "bbox": [ + 0.107, + 0.003, + 0.182, + 0.044 + ], + "angle": 0, + "content": "CVF" + }, + { + "type": "header", + "bbox": [ + 0.237, + 0.001, + 0.812, + 0.047 + ], + "angle": 0, + "content": "This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore." + }, + { + "type": "title", + "bbox": [ + 0.139, + 0.13, + 0.861, + 0.175 + ], + "angle": 0, + "content": "DiscoVLA: Discrepancy Reduction in Vision, Language, and Alignment for Parameter-Efficient Video-Text Retrieval" + }, + { + "type": "text", + "bbox": [ + 0.148, + 0.202, + 0.847, + 0.24 + ], + "angle": 0, + "content": "Leqi Shen\\(^{1,2,4*}\\) Guoqiang Gong\\(^{5\\dagger}\\) Tianxiang Hao\\(^{1,2}\\) Tao He\\(^{6,7}\\) Yifeng Zhang\\(^{5}\\) Pengzhang Liu\\(^{5}\\) Sicheng Zhao\\(^{2‡}\\) Jungong Han\\(^{3}\\) Guiguang Ding\\(^{1,2‡}\\)" + }, + { + "type": "text", + "bbox": [ + 0.179, + 0.24, + 0.819, + 0.275 + ], + "angle": 0, + "content": "\\(^{1}\\) School of Software \\(^{2}\\) BNrist \\(^{3}\\) Department of Automation, Tsinghua University \\(^{4}\\) Hangzhou Zhuoxi Institute of Brain and Intelligence" + }, + { + "type": "text", + "bbox": [ + 0.141, + 0.275, + 0.871, + 0.293 + ], + "angle": 0, + "content": "\\(^{5}\\) JD.com \\(^{6}\\) GRG Banking Equipment Co., Ltd. \\(^{7}\\) South China University of Technology" + }, + { + "type": "title", + "bbox": [ + 0.248, + 0.327, + 0.327, + 0.343 + ], + "angle": 0, + "content": "Abstract" + }, + { + "type": "text", + "bbox": [ + 0.089, + 0.359, + 0.485, + 0.721 + ], + "angle": 0, + "content": "The parameter-efficient adaptation of the image-text pretraining model CLIP for video-text retrieval is a prominent area of research. While CLIP is focused on image-level vision-language matching, video-text retrieval demands comprehensive understanding at the video level. Three key discrepancies emerge in the transfer from image-level to video-level: vision, language, and alignment. However, existing methods mainly focus on vision while neglecting language and alignment. In this paper, we propose Discrepancy Reduction in Vision, Language, and Alignment (DiscoVLA), which simultaneously mitigates all three discrepancies. Specifically, we introduce Image-Video Features Fusion to integrate image-level and video-level features, effectively tackling both vision and language discrepancies. Additionally, we generate pseudo image captions to learn fine-grained image-level alignment. To mitigate alignment discrepancies, we propose Image-to-Video Alignment Distillation, which leverages image-level alignment knowledge to enhance video-level alignment. Extensive experiments demonstrate the superiority of our DiscoVLA. In particular, on MSRVTT with CLIP (ViT-B/16), DiscoVLA outperforms previous methods by \\(2.2\\%\\) R@1 and \\(7.5\\%\\) R@sum. The code is available at https://github.com/LunarShen/DsicoVLA." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.75, + 0.222, + 0.765 + ], + "angle": 0, + "content": "1. Introduction" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.775, + 0.483, + 0.866 + ], + "angle": 0, + "content": "The rapid expansion of online video content has generated a growing demand for effective video-text retrieval, focusing on matching video content with relevant textual descriptions. Inspired by the success of image-text pretraining approaches such as CLIP [42], recent efforts [9, 14, 24, 37, 48, 49, 51] have increasingly focused on extend" + }, + { + "type": "image", + "bbox": [ + 0.536, + 0.325, + 0.889, + 0.596 + ], + "angle": 0, + "content": null + }, + { + "type": "table", + "bbox": [ + 0.558, + 0.599, + 0.866, + 0.664 + ], + "angle": 0, + "content": "
MethodImage-Level PETL Video-Level
VisionLanguageAlignment
Previous Methods✓××
Our DiscoVLA✓✓✓
" + }, + { + "type": "image_caption", + "bbox": [ + 0.512, + 0.675, + 0.908, + 0.745 + ], + "angle": 0, + "content": "Figure 1. When parameter-efficient transferring image-level CLIP to video-text retrieval, we identify three key discrepancies: vision, alignment, and language. Unlike previous methods, our DiscoVLA aims to address all these discrepancies. More visualization examples are provided in Figure 4." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.777, + 0.907, + 0.868 + ], + "angle": 0, + "content": "ing these powerful capabilities to video-text retrieval tasks. However, fully fine-tuning these pretrained models requires updating numerous parameters for every dataset, resulting in large storage overhead [22, 58]. Therefore, we focus on parameter-efficient transfer learning methods, targeting strong performance with minimal trainable parameters." + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.871, + 0.907, + 0.902 + ], + "angle": 0, + "content": "Although both image-text pretraining and video-text retrieval involve vision-language (VL) matching, they exhibit" + }, + { + "type": "page_footnote", + "bbox": [ + 0.109, + 0.875, + 0.469, + 0.888 + ], + "angle": 0, + "content": "* lunarshen@gmail.com. Work done during an internship at JD.com." + }, + { + "type": "page_footnote", + "bbox": [ + 0.111, + 0.888, + 0.339, + 0.9 + ], + "angle": 0, + "content": "†Project leader. ‡ Corresponding authors." + }, + { + "type": "list", + "bbox": [ + 0.109, + 0.875, + 0.469, + 0.9 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "19702" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.09, + 0.092, + 0.482, + 0.227 + ], + "angle": 0, + "content": "substantial discrepancies that hinder straightforward knowledge transfer. To be specific, image-text pretraining focuses on matching images with image captions, which represents image-level VL matching. In contrast, video-text retrieval focuses on matching videos with video captions, which represents video-level VL matching. The challenges of video-level VL matching, which demands understanding motion and time, create significant obstacles for transferring knowledge from image-text to video-text tasks." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.23, + 0.482, + 0.456 + ], + "angle": 0, + "content": "In this work, we explore the discrepancies between image-level and video-level VL matching across three key aspects: vision, language, and alignment. Videos introduce a temporal dimension, which is absent in images. Similarly, the distinction between image captions and video captions arises from the varying levels of granularity, such as isolated image descriptions versus continuous video narratives. Thus, video-level alignment is inherently more complex than image-level alignment, as it must account for intricate spatio-temporal relationships. In Figure 1, the three images correspond to \"preparing food\", \"pouring\", and \"cutting\", while the entire video demonstrates an instructional cooking process. Image-level VL matching focuses on isolated images, while video-level VL matching emphasizes the overall understanding of continuous segments." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.459, + 0.482, + 0.579 + ], + "angle": 0, + "content": "However, current methods [5, 22, 25, 58] struggle to fully address the discrepancies mentioned above. Some [5, 22, 25, 58] focus on the vision aspect, modeling temporal relationships to extract video-level features, while others [25, 58] attempt to reduce the modality gap between vision and language through shared parameter mechanisms. Despite these advances, these methods remain limited in both language and alignment discrepancies." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.582, + 0.482, + 0.673 + ], + "angle": 0, + "content": "To tackle these challenges, we introduce Discrepancy Reduction in Vision, Language, and Alignment, termed DiscoVLA, a novel parameter-efficient method for videotext retrieval. Unlike previous methods that target a single gap, our method aims to simultaneously reduce all significant discrepancies, as illustrated in the table of Figure 1." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.675, + 0.483, + 0.901 + ], + "angle": 0, + "content": "Specifically, we propose an Image-Video Features Fusion module (IVFusion), which addresses both vision and language discrepancies through a unified feature extraction approach. IVFusion effectively combines both image- and video-level features by utilizing a lightweight adapter. In addition to video-level alignment, we propose Pseudo Image-level Alignment (PIImgAlign) which generates pseudo image captions and learns fine-grained image-level alignment. Furthermore, to reduce alignment discrepancy, we propose Image-to-Video Alignment Distillation (AlignDistill), transferring image-level alignment knowledge to improve video-level alignment. All proposed modules prioritize parameter efficiency: IVFusion requires minimal trainable parameters, while PImgAlign and AlignDistill operate without increasing inference parameters." + }, + { + "type": "text", + "bbox": [ + 0.533, + 0.092, + 0.869, + 0.106 + ], + "angle": 0, + "content": "In summary, the main contributions are as follows:" + }, + { + "type": "text", + "bbox": [ + 0.514, + 0.107, + 0.905, + 0.151 + ], + "angle": 0, + "content": "- We reveal, for the first time, the necessity of addressing all three image-to-video discrepancies—vision, language, and alignment—to achieve parameter-efficient VTR." + }, + { + "type": "text", + "bbox": [ + 0.514, + 0.153, + 0.905, + 0.213 + ], + "angle": 0, + "content": "- We introduce IVFusion to fuse image- and video-level features for both vision and language gaps. PImgAlign is introduced for fine-grained image-level alignment. We introduce AlignDistill to minimize alignment gaps." + }, + { + "type": "text", + "bbox": [ + 0.514, + 0.213, + 0.905, + 0.273 + ], + "angle": 0, + "content": "- Our DiscoVLA achieves state-of-the-art results on MSRVTT, LSMDC, ActivityNet, and DiDeMo. Notably, on MSRVTT with CLIP (ViT-B/16), DiscoVLA reaches \\(50.5\\%\\) R@1, surpassing previous methods by \\(2.2\\%\\)." + }, + { + "type": "list", + "bbox": [ + 0.514, + 0.107, + 0.905, + 0.273 + ], + "angle": 0, + "content": null + }, + { + "type": "title", + "bbox": [ + 0.514, + 0.287, + 0.655, + 0.303 + ], + "angle": 0, + "content": "2. Related Work" + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.312, + 0.905, + 0.568 + ], + "angle": 0, + "content": "Video-Text Retrieval. Advances in vision-language pretraining [7, 12, 32-34, 42, 50, 52, 55] have achieved significant success across various downstream tasks [2, 10, 11, 18, 26, 45, 47, 54, 56, 57]. Leveraging powerful CLIP [42], recent video-text retrieval (VTR) methods [9, 14, 24, 37, 48, 49, 51] have achieved impressive performance enhancements. CLIP4Clip [37] is a pioneering CLIP-based method that aggregates image features via mean-pooling or transformers to obtain video features, inspiring subsequent studies. Additionally, Cap4Video [51] utilizes ZeroCap [46] and GPT-2 [41] to generate video-level captions as supplementary video content during both training and inference. In contrast, our DiscoVLA addresses the image-level and video-level gap by generating image-level captions used only in training to improve video-level alignment. These full fine-tuning methods impose substantial overhead from large trainable parameters." + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.569, + 0.905, + 0.735 + ], + "angle": 0, + "content": "Parameter-Efficient Transfer Learning. Parameter-efficient methods aim to adapt pre-trained models to new tasks with minimal fine-tuning. The mainstream methods in natural language processing and image tasks include techniques such as Prompt [23, 27, 28, 61, 62], Adapter [6, 17, 20, 60], and LoRA [21, 60]. Prompt [27] modifies the input space by adding learnable tokens. Adapter [20] introduces a lightweight bottleneck neural network consisting of up- and down-projection layers with a non-linear activation function. LoRA [21] aims to update up- and down-projection layers by learning low-rank decomposition matrices." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.735, + 0.905, + 0.901 + ], + "angle": 0, + "content": "Recently, several studies have explored parameter-efficient VTR methods [5, 22, 25, 58] to incorporate temporal modeling into the frozen image-level CLIP backbone. VoP [22] employs trainable BiLSTMs [16] to produce prompt tokens in the vision encoder. DGL [58] introduces weight-sharing layers to generate prompts for both text and video encoders. MV-Adapter [25] introduces shared bottleneck structures in both video and text branches. However, these methods overlook language and alignment discrepancies. We address these limitations with DiscoVLA, a framework that facilitates robust video-level understanding." + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.518, + 0.956 + ], + "angle": 0, + "content": "19703" + } + ], + [ + { + "type": "header", + "bbox": [ + 0.098, + 0.094, + 0.691, + 0.108 + ], + "angle": 0, + "content": "CLSTokenofVisionEncoder EOSTokenofTextEncoder -- \\(\\rightarrow\\) UtilizedONLYduringthe training phase." + }, + { + "type": "image", + "bbox": [ + 0.098, + 0.113, + 0.904, + 0.337 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.09, + 0.341, + 0.907, + 0.412 + ], + "angle": 0, + "content": "Figure 2. The overall framework of DiscoVLA. Initially, we generate pseudo image captions for each sampled image. In both vision and text encoders, we utilize image-level layers to acquire pretrained image-level knowledge and employ IVFusion layers to enhance spatiotemporal information. The single video caption is encoded through the text encoder, utilizing IVFusion layer as image-level (image-lvl) layer. Finally, AlignDistill is applied to distill image-level alignment to video-level alignment. For fair comparisons with previous methods, we do not generate pseudo image captions during the inference phase." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.436, + 0.226, + 0.453 + ], + "angle": 0, + "content": "3. Methodology" + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.462, + 0.221, + 0.479 + ], + "angle": 0, + "content": "3.1. Preliminary" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.487, + 0.484, + 0.563 + ], + "angle": 0, + "content": "Video-text retrieval is a task that involves retrieving relevant videos based on a given text query, or conversely, retrieving appropriate text descriptions for a video. Consistent with previous approaches [5, 22, 25, 58], we utilize the CLIP model, pre-trained on image-text data, as our backbone." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.565, + 0.484, + 0.716 + ], + "angle": 0, + "content": "To process a given video, we first sample a sequence of frame images uniformly, which are then input into the CLIP vision encoder to extract image features \\( v^{\\mathrm{img}} \\). These features are then aggregated through temporal average pooling to get video features \\( v^{\\mathrm{vid}} \\). For a given video caption, it is passed through the CLIP text encoder to extract video caption features \\( t^{\\mathrm{vid}} \\). Then, we compute the cross-modal contrastive loss [40], which functions as a video-level alignment term, maximizing the video-level similarity \\( \\mathrm{Sim}^{\\mathrm{vid}} \\) between matched pairs:" + }, + { + "type": "equation", + "bbox": [ + 0.218, + 0.731, + 0.482, + 0.753 + ], + "angle": 0, + "content": "\\[\n\\operatorname {S i m} _ {i j} ^ {\\mathrm {v i d}} = t _ {i} ^ {\\mathrm {v i d} ^ {T}} v _ {j} ^ {\\mathrm {v i d}}, \\tag {1}\n\\]" + }, + { + "type": "equation", + "bbox": [ + 0.107, + 0.755, + 0.483, + 0.797 + ], + "angle": 0, + "content": "\\[\n\\mathcal {L} _ {t 2 v} \\left(\\operatorname {S i m} ^ {\\text {v i d}}\\right) = \\frac {1}{B} \\sum_ {i = 1} ^ {B} \\log \\frac {\\exp \\left(\\operatorname {S i m} _ {i i} ^ {\\text {v i d}}\\right) / \\tau)}{\\sum_ {j = 1} ^ {B} \\exp \\left(\\operatorname {S i m} _ {i j} ^ {\\text {v i d}}\\right) / \\tau)}, \\tag {2}\n\\]" + }, + { + "type": "equation", + "bbox": [ + 0.107, + 0.8, + 0.482, + 0.841 + ], + "angle": 0, + "content": "\\[\n\\mathcal {L} _ {v 2 t} \\left(\\operatorname {S i m} ^ {\\text {v i d}}\\right) = \\frac {1}{B} \\sum_ {i = 1} ^ {B} \\log \\frac {\\exp \\left(\\operatorname {S i m} _ {i i} ^ {\\text {v i d}}\\right) / \\tau)}{\\sum_ {j = 1} ^ {B} \\exp \\left(\\operatorname {S i m} _ {j i} ^ {\\text {v i d}}\\right) / \\tau)}, \\tag {3}\n\\]" + }, + { + "type": "equation", + "bbox": [ + 0.107, + 0.843, + 0.482, + 0.873 + ], + "angle": 0, + "content": "\\[\n\\mathcal {L} _ {A} (\\operatorname {S i m} ^ {\\mathrm {v i d}}) = - \\frac {1}{2} \\left[ \\mathcal {L} _ {t 2 v} (\\operatorname {S i m} ^ {\\mathrm {v i d}}) + \\mathcal {L} _ {v 2 t} (\\operatorname {S i m} ^ {\\mathrm {v i d}}) \\right], \\tag {4}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.091, + 0.886, + 0.472, + 0.901 + ], + "angle": 0, + "content": "where \\(B\\) is the batch size, \\(\\tau\\) is the temperature parameter." + }, + { + "type": "title", + "bbox": [ + 0.514, + 0.437, + 0.63, + 0.451 + ], + "angle": 0, + "content": "3.2. DiscoVLA" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.459, + 0.907, + 0.625 + ], + "angle": 0, + "content": "As shown in Figure 2, we propose DiscoVLA to address the challenges of vision, language, and alignment when transferring from image-level to video-level VL matching. First, we introduce Image-Video Features Fusion (IVFusion) in both vision and text encoders to enhance feature extraction. Next, we introduce Pseudo Image-Level Alignment (PIImgAlign) for fine-grained image-level alignment, consisting of Pseudo Image Language Generation and Image-Level Alignment. Finally, Image-to-Video Alignment Distillation (AlignDistill) is proposed to leverage image-level alignment knowledge to enhance video-level alignment." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.626, + 0.906, + 0.701 + ], + "angle": 0, + "content": "Importantly, PImgAlign is only employed during training. For a fair comparison with previous methods and due to its generation inefficiency, PImgAlign is omitted at inference. The final similarity used for retrieval is based on video-level similarity in Eq. (1)." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.702, + 0.906, + 0.793 + ], + "angle": 0, + "content": "We focus on parameter-efficient transfer learning for video-text retrieval, where the large backbone parameters remain fixed, and only a small number of additional parameters are trained. The trainable parameters are embedded in the parameter-efficient attention of each encoder layer, as illustrated in Figure 3." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.803, + 0.776, + 0.819 + ], + "angle": 0, + "content": "3.3. Image-Video Features Fusion" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.825, + 0.906, + 0.901 + ], + "angle": 0, + "content": "Image-level and video-level variations exist in both vision and language modalities. Our proposed IVFusion fuses these image-level and video-level features in both the vision and language encoders. Below, we illustrate IVFusion's application within the vision encoder." + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "19704" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.107, + 0.089, + 0.498, + 0.341 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.506, + 0.09, + 0.887, + 0.34 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.089, + 0.342, + 0.908, + 0.4 + ], + "angle": 0, + "content": "Figure 3. Illustration of the encoder layers for vision and text encoders. (a) Image-Level Attn operates on each of the \\(F\\) images or image captions individually. (b) Video-Level Attn concatenates tokens across all sampled images or image captions. (c) IVFusion Attn employs a lightweight adapter to integrate the efficiency of Image-Level Attn with the effectiveness of Video-Level Attn. Here, we illustrate the application of IVFusion within the vision encoder. The text encoder adopts the same approach for processing pseudo image captions." + }, + { + "type": "text", + "bbox": [ + 0.089, + 0.424, + 0.482, + 0.56 + ], + "angle": 0, + "content": "Image-Level Attn. To achieve parameter-efficient transfer learning, we apply LoRA in the attention module of each encoder layer. As illustrated in Figure 3a, \\( F \\times (N + 1) \\) tokens are processed via Image-level Attn, where \\( F \\) denotes the number of sampled images, and \\( N + 1 \\) represents the combination of the global CLS token and patch tokens. Image-Level Attn processes each of the \\( F \\) images individually, using \\( N + 1 \\) tokens per image. However, Image-Level Attn does not encode temporal information." + }, + { + "type": "text", + "bbox": [ + 0.089, + 0.56, + 0.483, + 0.697 + ], + "angle": 0, + "content": "Video-Level Attn. A straightforward solution is Video-Level Attn, which concatenates tokens across all sampled images, enabling the attention module to learn spatiotemporal information. As shown in Figure 3b, \\( F(N + 1) \\) tokens are simultaneously processed during the attention module. Due to the quadratic complexity of the attention mechanism, Image-Level Attn has a complexity of \\( O(F(N)^2) \\), while Video-Level Attn has a higher complexity of \\( O((FN)^2) \\)." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.697, + 0.483, + 0.789 + ], + "angle": 0, + "content": "IVFusion Attn. To combine Image-Level Attn's efficiency with Video-Level Attn's effectiveness, we introduce IV-Fusion Attn in Figure 3c. We extend Image-Level Attn by incorporating a branch that uses all CLS tokens to extract spatio-temporal information, reducing complexity to \\( O(F(N)^2 + (F)^2 N) \\):" + }, + { + "type": "equation", + "bbox": [ + 0.149, + 0.796, + 0.483, + 0.817 + ], + "angle": 0, + "content": "\\[\n\\left[ c _ {i} ^ {(1)}, \\hat {E} _ {i} \\right] = \\mathrm {A T T N} \\left(Q K V = \\left[ c _ {i} ^ {(0)}, E _ {i} ^ {(0)} \\right]\\right), \\tag {5}\n\\]" + }, + { + "type": "equation", + "bbox": [ + 0.113, + 0.819, + 0.482, + 0.84 + ], + "angle": 0, + "content": "\\[\n\\left[ c _ {1: F} ^ {(2)} \\right] = \\operatorname {A T T N} \\left(Q = \\left[ c _ {1: F} ^ {(0)} \\right], K V = \\left[ c _ {1: F} ^ {(0)}, E _ {1: F} ^ {(0)} \\right]\\right), \\tag {6}\n\\]" + }, + { + "type": "equation", + "bbox": [ + 0.179, + 0.842, + 0.482, + 0.863 + ], + "angle": 0, + "content": "\\[\n\\hat {c} _ {i} = c _ {i} ^ {(2)} + \\sigma (c _ {i} ^ {(1)} \\mathbf {W} _ {\\mathrm {d o w n}}) \\mathbf {W} _ {\\mathrm {u p}}, \\tag {7}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.87, + 0.483, + 0.902 + ], + "angle": 0, + "content": "where \\(c_{i}\\in \\mathbb{R}^{1\\times D}\\) is CLS token of the \\(i^{\\mathrm{th}}\\) image and \\(c_{1:F}\\in \\mathbb{R}^{F\\times D}\\) are CLS tokens of all sampled images." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.423, + 0.905, + 0.579 + ], + "angle": 0, + "content": "Similarly, \\(E_{i} \\in \\mathbb{R}^{N \\times D}\\) are patch tokens of the \\(i^{\\text{th}}\\) image and \\(E_{1:F} \\in \\mathbb{R}^{FN \\times D}\\) are patch tokens of all sampled images. In Eq. (7), we employ a trainable adapter to merge image-level features \\(c_{i}^{(1)}\\) and video-level features \\(c_{i}^{(2)}\\), where \\(\\mathbf{W}_{down} \\in \\mathbb{R}^{D \\times r}\\) and \\(\\mathbf{W}_{up} \\in \\mathbb{R}^{r \\times D}\\) represent the up- and down-projection layers, and \\(\\sigma\\) denotes the nonlinear GELU activation [19]. Image-level features preserve pretrained spatial details, while video-level features capture temporal information. This fusion mechanism strengthens feature extraction." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.582, + 0.906, + 0.698 + ], + "angle": 0, + "content": "Our parameter-efficient vision encoder comprises both image-level and IVFusion layers. The shallow image-level layers employ Image-Level Attn to capture spatial information, while the top \\( H_{v} \\) IVFusion layers leverage IVFusion Attn to focus on spatio-temporal information. Consequently, our vision encoder outputs enhanced image features \\( \\left\\{ v_{i}^{\\mathrm{img}} \\mid 1 \\leq i \\leq F \\right\\} \\)." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.711, + 0.792, + 0.727 + ], + "angle": 0, + "content": "3.4. Pseudo Image-Level Alignment" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.735, + 0.907, + 0.902 + ], + "angle": 0, + "content": "Unlike previous work focused on video-level alignment, we introduce PImgAlign to learn fine-grained image-level alignment via the decomposition of video content at the image level. To accomplish this, we first generate pseudo image captions in VTR datasets. In contrast to the straightforward sampling of images from a video, sampling image captions is impractical, as video captions generally lack the specific details needed for individual images. As a solution, we utilize the multimodal large language model (MLLM) LLaVA-NeXT [35] to generate pseudo image captions. For each sampled image from a video, we prompt the model" + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "19705" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.091, + 0.092, + 0.395, + 0.107 + ], + "angle": 0, + "content": "using the relevant video caption for guidance:" + }, + { + "type": "text", + "bbox": [ + 0.122, + 0.117, + 0.452, + 0.178 + ], + "angle": 0, + "content": "The provided image is a frame sampled from the video, which describes {video caption}. Based on the video's content, provide a caption for the provided image." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.187, + 0.483, + 0.285 + ], + "angle": 0, + "content": "Unlike EA-VTR [38], which uses an image captioner, our method employs a strong MLLM. By incorporating video caption guidance, it generates higher-quality image captions that align with the overall video content. Similarly to visual feature extraction, we adopt our proposed IVFusion to extract pseudo image caption features \\(\\left\\{t_i^{\\mathrm{img}}|1\\leq i\\leq F\\right\\}\\)." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.286, + 0.484, + 0.451 + ], + "angle": 0, + "content": "Given image features and pseudo image caption features from a paired video and video caption, we introduce a fine-grained image-level similarity. Since image captions are typically more concise than the images they describe, a single caption may correspond to multiple images, as shown in Figure 4d. Therefore, we calculate the maximum similarity for each image across all image captions and vice versa. Specifically, we computed a similarity matrix between image and image caption features, selecting the maximum value from each row and column to obtain the final image-level similarity:" + }, + { + "type": "equation", + "bbox": [ + 0.153, + 0.462, + 0.483, + 0.504 + ], + "angle": 0, + "content": "\\[\n\\operatorname {S i m} _ {i j - t 2 v} ^ {\\mathrm {i m g}} = \\frac {1}{F} \\sum_ {n = 1} ^ {F} \\max _ {1 \\leq m \\leq F} t _ {i | n} ^ {\\mathrm {i m g} ^ {T}} v _ {j | m} ^ {\\mathrm {i m g}}, \\tag {8}\n\\]" + }, + { + "type": "equation", + "bbox": [ + 0.154, + 0.507, + 0.483, + 0.547 + ], + "angle": 0, + "content": "\\[\n\\operatorname {S i m} _ {i j - v 2 t} ^ {\\mathrm {i m g}} = \\frac {1}{F} \\sum_ {m = 1} ^ {F} \\max _ {1 \\leq n \\leq F} t _ {i | n} ^ {\\mathrm {i m g} ^ {T}} v _ {j | m} ^ {\\mathrm {i m g}}, \\tag {9}\n\\]" + }, + { + "type": "equation", + "bbox": [ + 0.159, + 0.55, + 0.482, + 0.579 + ], + "angle": 0, + "content": "\\[\n\\operatorname {S i m} _ {i j} ^ {\\mathrm {i m g}} = \\frac {1}{2} \\left[ \\operatorname {S i m} _ {i j - t 2 v} ^ {\\mathrm {i m g}} + \\operatorname {S i m} _ {i j - v 2 t} ^ {\\mathrm {i m g}} \\right], \\tag {10}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.59, + 0.483, + 0.688 + ], + "angle": 0, + "content": "where \\(\\mathrm{Sim}_{ij}^{\\mathrm{img}}\\) denotes the image-level similarity between the \\(i^{\\mathrm{th}}\\) video caption and the \\(j^{\\mathrm{th}}\\) video. \\(t_{i|n}^{\\mathrm{img}}\\) denotes the \\(n^{\\mathrm{th}}\\) pseudo image caption features of the \\(i^{\\mathrm{th}}\\) video caption and \\(v_{j|m}^{\\mathrm{img}}\\) denotes the \\(m^{\\mathrm{th}}\\) image features of the \\(j^{\\mathrm{th}}\\) video. PImgAlign achieves video-level alignment from the image level, effectively emphasizing spatial details." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.688, + 0.484, + 0.764 + ], + "angle": 0, + "content": "However, in the inference stage, the lack of ground truth for video captions prevents pseudo image captions generation. Additionally, MLLM introduces a computational overhead. For fair comparisons with other methods, we exclude PImgAlign during inference." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.773, + 0.421, + 0.789 + ], + "angle": 0, + "content": "3.5. Image-to-Video Alignment Distillation" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.795, + 0.484, + 0.901 + ], + "angle": 0, + "content": "Pretrained CLIP demonstrates a strong capability for image-level alignment. In contrast, video-text retrieval relies on comprehensive spatio-temporal alignment. Therefore, we propose AlignDistill to distill image-level alignment knowledge into video-level alignment, thus mitigating alignment discrepancies. Then, we optimize the Kullback-Leibler divergence [30] between image-level similarity and" + }, + { + "type": "text", + "bbox": [ + 0.514, + 0.092, + 0.664, + 0.107 + ], + "angle": 0, + "content": "video-level similarity:" + }, + { + "type": "equation", + "bbox": [ + 0.527, + 0.118, + 0.905, + 0.138 + ], + "angle": 0, + "content": "\\[\n\\mathcal {S} _ {t 2 v} ^ {\\mathrm {i m g}} = \\left[ \\mathrm {s} _ {i 1} ^ {\\mathrm {i m g}}, \\dots , \\mathrm {s} _ {i N _ {v}} ^ {\\mathrm {i m g}} \\right], \\mathcal {S} _ {v 2 t} ^ {\\mathrm {i m g}} = \\left[ \\mathrm {s} _ {1 i} ^ {\\mathrm {i m g}}, \\dots , \\mathrm {s} _ {N _ {t} i} ^ {\\mathrm {i m g}} \\right], \\tag {11}\n\\]" + }, + { + "type": "equation", + "bbox": [ + 0.536, + 0.14, + 0.905, + 0.16 + ], + "angle": 0, + "content": "\\[\n\\mathcal {S} _ {t 2 v} ^ {\\mathrm {v i d}} = \\left[ \\mathrm {s} _ {i 1} ^ {\\mathrm {v i d}}, \\dots , \\mathrm {s} _ {i N _ {v}} ^ {\\mathrm {v i d}} \\right], \\mathcal {S} _ {v 2 t} ^ {\\mathrm {v i d}} = \\left[ \\mathrm {s} _ {1 i} ^ {\\mathrm {v i d}}, \\dots , \\mathrm {s} _ {N _ {t i}} ^ {\\mathrm {v i d}} \\right], \\tag {12}\n\\]" + }, + { + "type": "equation", + "bbox": [ + 0.525, + 0.162, + 0.905, + 0.19 + ], + "angle": 0, + "content": "\\[\n\\mathcal {L} _ {K L} = \\frac {1}{2} \\left(\\mathbf {K L} \\left(\\mathcal {S} _ {t 2 v} ^ {\\mathrm {i m g}} \\| \\mathcal {S} _ {t 2 v} ^ {\\mathrm {v i d}}\\right) + \\mathbf {K L} \\left(\\mathcal {S} _ {v 2 t} ^ {\\mathrm {i m g}} \\| \\mathcal {S} _ {v 2 t} ^ {\\mathrm {v i d}}\\right)\\right). \\tag {13}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.202, + 0.905, + 0.285 + ], + "angle": 0, + "content": "where \\(\\mathrm{s}_{ij}^{\\mathrm{img}}\\) is \\(\\mathrm{Sim}_{ij}^{\\mathrm{img}}\\) in Eq. (10) and \\(\\mathrm{s}_{ij}^{\\mathrm{vid}}\\) is \\(\\mathrm{Sim}_{ij}^{\\mathrm{vid}}\\) in Eq. (1). \\(S_{t2v}^{\\mathrm{img}}\\) and \\(S_{t2v}^{\\mathrm{vid}}\\) denote the probability distributions for the text-to-video task, while \\(S_{v2t}^{\\mathrm{img}}\\) and \\(S_{v2t}^{\\mathrm{vid}}\\) denote those for the video-to-text task. During the inference phase, We employ enhanced video-level similarity for retrieval tasks." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.286, + 0.905, + 0.331 + ], + "angle": 0, + "content": "Finally, the overall training objective, which includes alignment loss (Eq. (4)) and distillation loss (Eq. (13)), is formulated as follows:" + }, + { + "type": "equation", + "bbox": [ + 0.546, + 0.343, + 0.905, + 0.362 + ], + "angle": 0, + "content": "\\[\n\\mathcal {L} = \\mathcal {L} _ {A} (\\operatorname {S i m} ^ {\\mathrm {v i d}}) + \\alpha \\mathcal {L} _ {A} (\\operatorname {S i m} ^ {\\mathrm {i m g}}) + \\beta \\mathcal {L} _ {K L}, \\tag {14}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.376, + 0.905, + 0.408 + ], + "angle": 0, + "content": "where \\(\\alpha\\) and \\(\\beta\\) are used to balance the loss. \\(\\mathcal{L}_A(\\mathrm{Sim}^{\\mathrm{img}})\\) is the image-level alignment loss (see Eq. (4) and (10))." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.424, + 0.646, + 0.44 + ], + "angle": 0, + "content": "4. Experiments" + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.449, + 0.722, + 0.466 + ], + "angle": 0, + "content": "4.1. Experimental Settings" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.472, + 0.907, + 0.638 + ], + "angle": 0, + "content": "Following common practice, we perform evaluations on four widely used benchmarks: MSRVTT [53], LSMDC [43], ActivityNet [29], and DiDeMo [1]. We evaluate the performance using common retrieval metrics such as Recall at K (R@K and K = 1, 5, 10), the sum of these recalls (R@sum), and Mean Rank (MnR). In all experiments, the LoRA dimension and the adapter dimension \\( r \\) are set to 8. In Eq. 14, \\( \\alpha \\) and \\( \\beta \\) are set to 0.3 and 1.0, respectively. For the number of IVFusion layers, we set \\( H_V = 4 \\) for the vision encoder and \\( H_L = 2 \\) for the text encoder. Please see Appendix A for further details." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.65, + 0.889, + 0.666 + ], + "angle": 0, + "content": "4.2. Comparisons with State-of-the-Art Methods" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.673, + 0.905, + 0.718 + ], + "angle": 0, + "content": "We compare our proposed DiscoVLA with state-of-the-art methods on popular benchmarks such as MSRVTT, LSMDC, ActivityNet, and DiDeMo." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.719, + 0.906, + 0.87 + ], + "angle": 0, + "content": "Full fine-tuning refers to CLIP4Clip-meanP [37], which averages image features along the temporal dimension. All parameter-efficient methods follow this temporal average pooling strategy to obtain the final video features. Although Prompt [27], Adapter [20], and LoRA [21] are widely adopted for parameter-efficient image and NLP tasks, they lack effective temporal modeling for video-level understanding. In contrast, methods like RAP [5], VoP [22], DGL [58], and MV-Adapter [25] focus on temporal modeling in the vision modality." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.871, + 0.906, + 0.901 + ], + "angle": 0, + "content": "As shown in Table 1, on MSRVTT, our DiscoVLA with CLIP (ViT-B/32) as the backbone achieves \\(47.0\\%\\) R@1 in" + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.52, + 0.957 + ], + "angle": 0, + "content": "19706" + } + ], + [ + { + "type": "table", + "bbox": [ + 0.119, + 0.089, + 0.877, + 0.354 + ], + "angle": 0, + "content": "
Method# Params (M)Text-to-VideoVideo-to-Text
R@1↑R@5↑R@10↑R@sum↑MnR↓R@1↑R@5↑R@10↑R@sum↑MnR↓
CLIP (ViT-B/32)
Full fine-tuning123.5443.170.480.8194.316.243.170.581.2194.812.4
Prompt [27]0.0840.466.377.3184.016.742.269.779.2191.112.4
Adapter [20]0.2641.969.978.7190.214.943.669.980.1193.611.5
LoRA [21]0.4943.768.980.4193.016.043.070.282.2195.412.0
RAP [5]1.0644.871.481.5197.714.444.071.982.4198.310.1
VoPF+P [22]0.443.569.379.3192.114.843.671.281.2196.011.0
VoPF+C [22]14.1044.770.579.2194.416.242.170.080.6192.713.4
DGL [58]0.8344.669.980.3194.816.344.570.780.6195.811.5
DiscoVLA0.5647.073.082.8202.814.147.773.683.6204.910.0
CLIP (ViT-B/16)
MV-Adapter [25]3.646.072.082.1200.1-45.674.083.8203.4-
RAP [5]1.0646.573.982.0202.412.145.376.484.8206.59.1
VoPF+P [22]0.447.172.481.8201.312.9-----
VoPF+C [22]14.1047.772.482.2202.312.0-----
DGL [58]0.8348.371.880.6200.713.445.774.082.9202.610.9
TempMe [44]0.5049.074.483.3206.711.947.675.385.4208.39.0
DiscoVLA0.5650.575.683.8209.912.149.276.084.7209.98.6
" + }, + { + "type": "table_caption", + "bbox": [ + 0.09, + 0.355, + 0.907, + 0.397 + ], + "angle": 0, + "content": "Table 1. Comparisons with state-of-the-art methods on MSRVTT. # Params denotes the number of trainable parameters. \\(\\uparrow\\) denotes higher values represent better performance and \\(\\downarrow\\) denotes lower values represent better performance. The best value for each metric is highlighted in bold. R@sum is defined as the total of R@1, R@5, and R@10. The gray row represents the fully fine-tuned CLIP4Clip-meanP [37]." + }, + { + "type": "table", + "bbox": [ + 0.124, + 0.403, + 0.872, + 0.501 + ], + "angle": 0, + "content": "
Method# Params (M)Text-to-VideoVideo-to-Text
R@1↑R@5↑R@10↑R@sum↑MnR↓R@1↑R@5↑R@10↑R@sum↑MnR↓
Full fine-tuning123.5420.738.947.2106.865.320.639.447.5107.556.7
VoPF+P [22]0.420.740.749.7111.159.121.540.650.7112.850.8
VoPF+C [22]14.1021.140.949.6111.660.122.340.350.7113.351.1
DGL [58]0.8321.439.448.4109.264.3-----
DiscoVLA0.5623.642.052.3117.952.022.841.951.2115.946.6
" + }, + { + "type": "table_caption", + "bbox": [ + 0.236, + 0.501, + 0.76, + 0.515 + ], + "angle": 0, + "content": "Table 2. Comparisons with state-of-the-art methods on LSMDC using CLIP (ViT-B/32)." + }, + { + "type": "table", + "bbox": [ + 0.124, + 0.52, + 0.872, + 0.632 + ], + "angle": 0, + "content": "
Method# Params (M)Text-to-VideoVideo-to-Text
R@1↑R@5↑R@10↑R@sum↑MnR↓R@1↑R@5↑R@10↑R@sum↑MnR↓
Full fine-tuning123.5440.572.4--7.442.574.185.8202.46.6
RAP [5]1.0640.871.082.2194.08.3-----
VoPF+P [22]0.436.165.578.5180.110.936.365.979.2181.410.1
VoPF+C [22]14.1035.163.777.6176.411.435.665.977.8179.310.4
DGL [58]0.8338.669.281.6189.49.0-----
DiscoVLA0.5641.272.483.6197.27.841.872.884.7199.36.9
" + }, + { + "type": "table_caption", + "bbox": [ + 0.227, + 0.632, + 0.768, + 0.646 + ], + "angle": 0, + "content": "Table 3. Comparisons with state-of-the-art methods on ActivityNet using CLIP (ViT-B/32)." + }, + { + "type": "text", + "bbox": [ + 0.089, + 0.66, + 0.483, + 0.827 + ], + "angle": 0, + "content": "the text-to-video task (t2v) and \\(47.7\\%\\) R@1 in the video-to-text task (v2t), significantly outperforming previous methods. Furthermore, when using CLIP (ViT-B/16), DiscoVLA achieves improvements of \\(2.2\\%\\) R@1 and \\(7.5\\%\\) R@sum. The comparison results on LSMDC, ActivityNet, DiDeMo are shown in Table 2-4. Our method outperforms all existing approaches across these datasets. Existing approaches overlook the critical image-to-video language and alignment discrepancies. In contrast, our proposed DiscoVLA addresses these discrepancies across all three areas: vision, language, and alignment." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.834, + 0.245, + 0.851 + ], + "angle": 0, + "content": "4.3. Ablation Study" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.856, + 0.485, + 0.903 + ], + "angle": 0, + "content": "Ablation study on individual components. We evaluate the proposed components in Table 5. LoRA applies Image-Level Attn (see Figure 3a) in both the vision and text" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.66, + 0.907, + 0.886 + ], + "angle": 0, + "content": "encoders. In B1, IVFusion significantly enhances performance by \\(8.7\\%\\) R@sum in t2v and \\(6.7\\%\\) R@sum in v2t, effectively integrating image-level and video-level features. In B2 and B3, we find that using PImgAlign or AlignDistill independently yields limited improvements. This is largely due to the notable gap between image-level and video-level (see Figure 1). Image-level PImgAlign struggles to impact video-level tasks, while AlignDistill lacks sufficient image-specific learning. When combined, however, they complement each other. PImgAlign captures fine-grained image-level alignments, and AlignDistill effectively transfers this knowledge to the video-level alignment. This improves \\(1.1\\%\\) R@sum in t2v and \\(2.8\\%\\) R@sum in v2t over B1. Finally, our full DiscoVLA improves \\(9.8\\%\\) R@sum in t2v and \\(9.5\\%\\) R@sum in v2t over LoRA." + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.52, + 0.958 + ], + "angle": 0, + "content": "19707" + } + ], + [ + { + "type": "table", + "bbox": [ + 0.124, + 0.089, + 0.872, + 0.187 + ], + "angle": 0, + "content": "
Method# Params (M)Text-to-VideoVideo-to-Text
R@1↑R@5↑R@10↑R@sum↑MnR↓R@1↑R@5↑R@10↑R@sum↑MnR↓
Full fine-tuning123.5443.470.280.6194.217.542.570.680.2193.311.6
RAP [5]1.0642.670.479.6192.618.0-----
VoPF+P [22]0.445.372.380.4198.013.844.771.281.1197.09.9
VoPF+C [22]14.1046.471.981.5199.813.644.471.881.8198.09.5
DiscoVLA0.5648.474.582.7205.614.047.774.483.8205.99.3
" + }, + { + "type": "table_caption", + "bbox": [ + 0.236, + 0.187, + 0.76, + 0.2 + ], + "angle": 0, + "content": "Table 4. Comparisons with state-of-the-art methods on DiDeMo using CLIP (ViT-B/32)." + }, + { + "type": "table", + "bbox": [ + 0.095, + 0.206, + 0.904, + 0.315 + ], + "angle": 0, + "content": "
MethodsComponents# Params (M)Text-to-VideoVideo-to-Text
IVFusionPIImgAlignAlignDistillR@1↑R@5↑R@10↑R@sum↑R@1↑R@5↑R@10↑R@sum↑
Pretraind CLIPXXX030.853.863.3147.926.650.162.0138.7
LoRAXXX0.4943.768.980.4193.043.070.282.2195.4
B1✓XX0.5446.573.282.0201.746.672.682.9202.1
B2✓✓X0.5647.073.181.3201.445.772.982.4201.0
B3✓X✓0.5446.671.681.0199.245.971.982.2200.0
DiscoVLA✓✓✓0.5647.073.082.8202.847.773.683.6204.9
" + }, + { + "type": "table_caption", + "bbox": [ + 0.09, + 0.316, + 0.905, + 0.359 + ], + "angle": 0, + "content": "Table 5. Ablation study on the contribution of each proposed component on MSRVTT using CLIP (ViT-B/32). Pretrained CLIP denotes the zero-shot performance of CLIP without any additional training. Our proposed methods are built upon LoRA. B1, B2, and B3 denote various combinations of our proposed components." + }, + { + "type": "table", + "bbox": [ + 0.095, + 0.367, + 0.481, + 0.447 + ], + "angle": 0, + "content": "
Parameter-Efficient AttentionText-to-Video
R@1↑R@5↑R@10↑R@sum↑
Image-Level Attn43.768.980.4193.0
Video-Level Attn45.770.980.8197.4
IVFusion Attn w/o Adapter46.071.080.4197.4
IVFusion Attn46.573.282.0201.7
" + }, + { + "type": "table_caption", + "bbox": [ + 0.09, + 0.448, + 0.483, + 0.504 + ], + "angle": 0, + "content": "Table 6. Ablation study on IVFusion on MSRVTT using CLIP (ViT-B/32). We compare several implementations against our proposed IVFusion for Parameter-Efficient Attention (see Figure 3). Image-Level Attn represents the LoRA baseline." + }, + { + "type": "table", + "bbox": [ + 0.095, + 0.509, + 0.481, + 0.59 + ], + "angle": 0, + "content": "
Similarity in PImgAlignText-to-Video
R@1↑R@5↑R@10↑R@sum↑
None46.671.681.0199.2
Video-Level46.372.381.2199.8
Paired Image-Level46.173.282.4201.7
Fine-grained Image-Level47.073.082.8202.8
" + }, + { + "type": "table_caption", + "bbox": [ + 0.09, + 0.591, + 0.483, + 0.66 + ], + "angle": 0, + "content": "Table 7. Ablation study on similarity in PImgAlign on MSRVTT using CLIP (ViT-B/32). None represents DiscoVLA without image-level alignment optimization, shown as B3 in Table 5. Paired Image-Level is based on matched image-caption pairs. Fine-grained Image-Level represents the full DiscoVLA." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.675, + 0.483, + 0.839 + ], + "angle": 0, + "content": "Analysis of trainable parameters. The trainable parameters of our DiscoVLA consist of the LoRA and Adapter parts. As shown in Table 5, LoRA in both the vision and text encoders contain \\(\\sim 0.49\\mathrm{M}\\) parameters. In B1, an additional \\(\\sim 0.05\\mathrm{M}\\) parameters are introduced by the Adapter in the \\(H_{V} = 4\\) IVFusion layers of the vision encoder. In B2, an additional \\(\\sim 0.02\\mathrm{M}\\) parameters are introduced by the Adapter in the \\(H_{L} = 2\\) IVFusion layers of the text encoder, which are utilized for pseudo-image caption features in PImgAlign. Overall, the total number of trainable parameters of DiscoVLA is \\(\\sim 0.56\\mathrm{M}\\)." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.84, + 0.484, + 0.902 + ], + "angle": 0, + "content": "Ablation study on IVFusion. In Table 6, we compare different strategies for parameter-efficient attention in Figure 3. Video-Level Attn concatenates all tokens across sampled images to capture spatio-temporal information," + }, + { + "type": "table", + "bbox": [ + 0.517, + 0.367, + 0.905, + 0.426 + ], + "angle": 0, + "content": "
Image Caption GenerationText-to-Video
R@1↑R@5↑R@10↑R@sum↑
DiscoVLA47.073.082.8202.8
W/O VidCap46.1 (-0.9)73.082.1 (-0.7)201.2 (-1.6)
" + }, + { + "type": "table_caption", + "bbox": [ + 0.513, + 0.427, + 0.905, + 0.469 + ], + "angle": 0, + "content": "Table 8. Ablation study on image caption generation on MSRVTT using CLIP (ViT-B/32). W/O VidCap denotes generating pseudo image captions without video caption guidance." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.49, + 0.907, + 0.61 + ], + "angle": 0, + "content": "achieving a notable \\(4.4\\%\\) R@sum improvement. To reduce the computational complexity, IVFusion applies spatiotemporal attention only to global CLS tokens and merges image-level features with video-level features. Without an adapter, averaging image- and video-level features results in a modest performance increase. In contrast, with a learnable adapter, IVFusion merges the features more effectively, achieving \\(201.7\\%\\) R@sum." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.614, + 0.909, + 0.903 + ], + "angle": 0, + "content": "Ablation study on similarity in PImgAlign. Table 7 demonstrates the effectiveness of our proposed fine-grained image-level similarity in PImgAlign. The similarity is computed between sampled images and pseudo image captions from a paired video and video caption. The Video-Level baseline averages the image features and pseudo-image caption features along the temporal dimension to compute the video-level similarity, replacing fine-grained image-level similarity. However, this modification leads to a notable decrease in performance, emphasizing the importance of image-level similarity in PImgAlign. In contrast, we introduce the Paired Image-Level baseline, which computes similarity based on matched image-caption pairs. Compared to Video-Level, Paired Image-Level yields a \\(1.9\\%\\) gain in R@sum. Moreover, we observe in Figure 4d that a single image caption may describe multiple images. Consequently, our fine-grained image-level similarity, as defined in Eq. 10, selects the maximum similarity from all available pairs, further enhancing video-level alignment." + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.519, + 0.956 + ], + "angle": 0, + "content": "19708" + } + ], + [ + { + "type": "header", + "bbox": [ + 0.15, + 0.089, + 0.353, + 0.101 + ], + "angle": 0, + "content": "Video Caption - Image Caption" + }, + { + "type": "image", + "bbox": [ + 0.091, + 0.103, + 0.905, + 0.393 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.09, + 0.395, + 0.907, + 0.451 + ], + "angle": 0, + "content": "Figure 4. Visualization of discrepancies between video-level and image-level data. Each example consists of a paired video and video caption, with four sampled images from the video. Video captions are highlighted with orange solid lines, and pseudo image captions generated by LLaVA-NeXT [35] are indicated with green dashed lines. While video captions convey the general context of the video, image captions focus on the detailed context of each individual frame." + }, + { + "type": "image", + "bbox": [ + 0.094, + 0.47, + 0.483, + 0.586 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.091, + 0.587, + 0.483, + 0.616 + ], + "angle": 0, + "content": "Figure 5. Comparison of image caption generation with (W/) and without (W/O) video caption guidance (VidCap) in PImgAlign." + }, + { + "type": "text", + "bbox": [ + 0.089, + 0.642, + 0.483, + 0.824 + ], + "angle": 0, + "content": "Ablation study on image caption generation in PImgAlign. Figure 5 and Table 8 show qualitative and quantitative evaluations of video caption guidance. In Figure 5, video captions serve as the global context that enriches the understanding of individual frames (marked in red). In Table 8, without this guidance, image captions focus only on salient regions while missing globally relevant details, leading to performance drops. In PImgAlign, we employ MLLM to process both frame images and video descriptions, generating image captions that align with the overall video content. These high-quality captions further enhance DiscoVLA's ability to learn robust image-level alignment." + }, + { + "type": "text", + "bbox": [ + 0.089, + 0.825, + 0.483, + 0.903 + ], + "angle": 0, + "content": "More Ablation Analysis. In Appendix B, we conduct additional ablation studies. Specifically, we evaluate the effects of hyperparameters \\(\\alpha\\), \\(\\beta\\), \\(H_V\\) and \\(H_V\\). For fairness, all results in the main paper are reported without post-processing. Complementary evaluations with Q-Norm [4] and DSL [8]" + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.473, + 0.892, + 0.489 + ], + "angle": 0, + "content": "as post-processing methods are provided in the appendix." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.495, + 0.707, + 0.512 + ], + "angle": 0, + "content": "4.4. Qualitative Analysis" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.517, + 0.907, + 0.715 + ], + "angle": 0, + "content": "To compare image-level and video-level data, we conduct a visualization experiment in Figure 4. In Figure 4a and 4b, image-level data alone fails to capture the complete content of the video. Figure 4c and 4d demonstrate a strong correlation between the video-level and image-level content, but differences across images lead to varying image captions. Visually, a video is composed of a sequence of images, incorporating temporal information. Textually, image captions are specific to individual images, while video captions describe the overall video content. These discrepancies hinder the effective transfer of image-level CLIP. Our approach addresses this issue by utilizing image-level alignment to enhance video-level alignment." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.725, + 0.634, + 0.74 + ], + "angle": 0, + "content": "5. Conclusion" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.75, + 0.907, + 0.903 + ], + "angle": 0, + "content": "In this paper, we identify three major discrepancies in the parameter-efficient transfer of image-level CLIP to videotext retrieval: vision, language, and alignment. To address these challenges, we propose DiscoVLA. Specifically, IV-Fusion fuses image-level and video-level features to tackle the vision and language discrepancies. For alignment discrepancies, we propose PImgAlign and AlignDistill, which learn and distill image-level alignment to enhance video-level alignment. Our DiscoVLA significantly outperforms existing methods, achieving state-of-the-art performance." + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "19709" + } + ], + [ + { + "type": "title", + "bbox": [ + 0.092, + 0.091, + 0.241, + 0.108 + ], + "angle": 0, + "content": "Acknowledgment" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.115, + 0.484, + 0.268 + ], + "angle": 0, + "content": "This work was supported by Beijing Natural Science Foundation (No. L223023), National Natural Science Foundation of China (Nos. 62441235, 62021002, 62441614), the Key R & D Program of Xinjiang, China (2022B01006), Zhejiang Provincial Natural Science Foundation of China under Grant (No. LDT23F01013F01), CCF-DiDi GAIA Collaborative Research Funds, China Postdoctoral Science Foundation (2024M750565), Guangdong S & T Program (2024B0101040008), and the Key Realm Research and Development Program of Guangzhou (No. 2024B01W0007)." + }, + { + "type": "title", + "bbox": [ + 0.093, + 0.281, + 0.188, + 0.297 + ], + "angle": 0, + "content": "References" + }, + { + "type": "ref_text", + "bbox": [ + 0.101, + 0.306, + 0.484, + 0.362 + ], + "angle": 0, + "content": "[1] Lisa Anne Hendricks, Oliver Wang, Eli Shechtman, Josef Sivic, Trevor Darrell, and Bryan Russell. Localizing moments in video with natural language. In Proceedings of the International Conference on Computer Vision, 2017. 5, 12" + }, + { + "type": "ref_text", + "bbox": [ + 0.101, + 0.364, + 0.484, + 0.419 + ], + "angle": 0, + "content": "[2] Stanislaw Antol, Aishwarya Agrawal, Jiasen Lu, Margaret Mitchell, Dhruv Batra, C Lawrence Zitnick, and Devi Parikh. Vqa: Visual question answering. In Proceedings of the International Conference on Computer Vision, 2015. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.421, + 0.484, + 0.476 + ], + "angle": 0, + "content": "[3] Max Bain, Arsha Nagrani, Gül Varol, and Andrew Zisserman. Frozen in time: A joint video and image encoder for end-to-end retrieval. In Proceedings of the International Conference on Computer Vision, 2021. 12" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.478, + 0.482, + 0.532 + ], + "angle": 0, + "content": "[4] Simion-Vlad Bogolin, Ioana Croitoru, Hailin Jin, Yang Liu, and Samuel Albanie. Cross modal retrieval with querybank normalisation. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition, 2022. 8, 13" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.535, + 0.484, + 0.603 + ], + "angle": 0, + "content": "[5] Meng Cao, Haoran Tang, Jinfa Huang, Peng Jin, Can Zhang, Ruyang Liu, Long Chen, Xiaodan Liang, Li Yuan, and Ge Li. RAP: Efficient text-video retrieval with sparse-and-correlated adapter. In Association for Computational Linguistics, 2024. 2, 3, 5, 6, 7, 12" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.605, + 0.484, + 0.659 + ], + "angle": 0, + "content": "[6] Shoufa Chen, Chongjian Ge, Zhan Tong, Jiangliu Wang, Yibing Song, Jue Wang, and Ping Luo. Adaptformer: Adapting vision transformers for scalable visual recognition. Advances in Neural Information Processing Systems, 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.661, + 0.484, + 0.73 + ], + "angle": 0, + "content": "[7] Sihan Chen, Handong Li, Qunbo Wang, Zijia Zhao, Mingzhen Sun, Xinxin Zhu, and Jing Liu. Vast: A vision-audio-subtitle-text omni-modality foundation model and dataset. Advances in Neural Information Processing Systems, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.733, + 0.484, + 0.787 + ], + "angle": 0, + "content": "[8] Xing Cheng, Hezheng Lin, Xiangyu Wu, Fan Yang, and Dong Shen. Improving video-text retrieval by multi-stream corpus alignment and dual softmax loss. arXiv preprint arXiv:2109.04290, 2021. 8, 13" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.789, + 0.484, + 0.844 + ], + "angle": 0, + "content": "[9] Chaorui Deng, Qi Chen, Pengda Qin, Da Chen, and Qi Wu. Prompt switch: Efficient clip adaptation for text-video retrieval. In Proceedings of the International Conference on Computer Vision, 2023. 1, 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.846, + 0.485, + 0.901 + ], + "angle": 0, + "content": "[10] Xiaohan Ding, Yuchen Guo, Guiguang Ding, and Jungong Han. Acnet: Strengthening the kernel skeletons for powerful cnn via asymmetric convolution blocks. In Proceedings of the International Conference on Computer Vision, 2019. 2" + }, + { + "type": "list", + "bbox": [ + 0.094, + 0.306, + 0.485, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.093, + 0.907, + 0.148 + ], + "angle": 0, + "content": "[11] Xiaohan Ding, Xiangyu Zhang, Ningning Ma, Jungong Han, Guiguang Ding, and Jian Sun. Repvgg: Making vgg-style convnets great again. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition, 2021. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.149, + 0.906, + 0.203 + ], + "angle": 0, + "content": "[12] Lijie Fan, Dilip Krishnan, Phillip Isola, Dina Katabi, and Yonglong Tian. Improving clip training with language rewrites. Advances in Neural Information Processing Systems, 2024. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.205, + 0.906, + 0.259 + ], + "angle": 0, + "content": "[13] Valentin Gabeur, Chen Sun, Karteek Alahari, and Cordelia Schmid. Multi-modal transformer for video retrieval. In Proceedings of the European Conference on Computer Vision, 2020. 12" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.261, + 0.906, + 0.329 + ], + "angle": 0, + "content": "[14] Satya Krishna Gorti, Noel Vouitsis, Junwei Ma, Keyvan Golestan, Maksims Volkovs, Animesh Garg, and Guangwei Yu. X-pool: Cross-modal language-video attention for text-video retrieval. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition, 2022. 1, 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.331, + 0.906, + 0.398 + ], + "angle": 0, + "content": "[15] Priya Goyal, Piotr Dólar, Ross Girshick, Pieter Noordhuis, Lukasz Wesolowski, Aapo Kyrola, Andrew Tulloch, Yangqing Jia, and Kaiming He. Accurate, large minibatch sgd: Training imagenet in 1 hour. arXiv preprint arXiv:1706.02677, 2017. 12" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.4, + 0.906, + 0.44 + ], + "angle": 0, + "content": "[16] Alex Graves and Jürgen Schmidhuber. Framewise phoneme classification with bidirectional LSTM and other neural network architectures. Neural Networks, 2005. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.442, + 0.907, + 0.509 + ], + "angle": 0, + "content": "[17] Junxian He, Chunting Zhou, Xuezhe Ma, Taylor Berg-Kirkpatrick, and Graham Neubig. Towards a unified view of parameter-efficient transfer learning. In Proceedings of the International Conference on Learning Representations, 2021. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.511, + 0.906, + 0.58 + ], + "angle": 0, + "content": "[18] Tao He, Leqi Shen, Yuchen Guo, Guiguang Ding, and Zhenhua Guo. Secret: Self-consistent pseudo label refinement for unsupervised domain adaptive person re-identification. In Proceedings of the AAAI Conference on Artificial Intelligence, 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.581, + 0.906, + 0.608 + ], + "angle": 0, + "content": "[19] Dan Hendrycks and Kevin Gimpel. Gaussian error linear units (gelus). arXiv preprint arXiv:1606.08415, 2016. 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.61, + 0.906, + 0.678 + ], + "angle": 0, + "content": "[20] Neil Houlsby, Andrei Giurgiu, Stanislaw Jastrzebski, Bruna Morrone, Quentin De Laroussilhe, Andrea Gesmundo, Mona Attariyan, and Sylvain Gelly. Parameter-efficient transfer learning for nlp. In Proceedings of the International Conference on Machine Learning, 2019. 2, 5, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.679, + 0.906, + 0.734 + ], + "angle": 0, + "content": "[21] Edward J Hu, Yelong Shen, Phillip Wallis, Zeyuan Allen-Zhu, Yuanzhi Li, Shean Wang, Lu Wang, and Weizhu Chen. Lora: Low-rank adaptation of large language models. arXiv preprint arXiv:2106.09685, 2021. 2, 5, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.735, + 0.907, + 0.803 + ], + "angle": 0, + "content": "[22] Siteng Huang, Biao Gong, Yulin Pan, Jianwen Jiang, Yiliang Lv, Yuyuan Li, and Donglin Wang. Vop: Text-video cooperative prompt tuning for cross-modal retrieval. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition, 2023. 1, 2, 3, 5, 6, 7, 12" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.804, + 0.906, + 0.859 + ], + "angle": 0, + "content": "[23] Menglin Jia, Luming Tang, Bor-Chun Chen, Claire Cardie, Serge Belongie, Bharath Hariharan, and Ser-Nam Lim. Visual prompt tuning. In Proceedings of the European Conference on Computer Vision, 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.86, + 0.906, + 0.901 + ], + "angle": 0, + "content": "[24] Peng Jin, Jina Huang, Pengfei Xiong, Shangxuan Tian, Chang Liu, Xiangyang Ji, Li Yuan, and Jie Chen. Videotext as game players: Hierarchical banzhaf interaction for" + }, + { + "type": "list", + "bbox": [ + 0.094, + 0.093, + 0.907, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.52, + 0.957 + ], + "angle": 0, + "content": "19710" + } + ], + [ + { + "type": "ref_text", + "bbox": [ + 0.125, + 0.092, + 0.484, + 0.134 + ], + "angle": 0, + "content": "cross-modal representation learning. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition, 2023. 1, 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.136, + 0.484, + 0.219 + ], + "angle": 0, + "content": "[25] Xiaojie Jin, Bowen Zhang, Weibo Gong, Kai Xu, Xueqing Deng, Peng Wang, Zhao Zhang, Xiaohui Shen, and Jiashi Feng. Mv-adapter: Multimodal video transfer learning for video text retrieval. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition, 2024. 2, 3, 5, 6, 12" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.221, + 0.483, + 0.276 + ], + "angle": 0, + "content": "[26] Andrej Karpathy and Li Fei-Fei. Deep visual-semantic alignments for generating image descriptions. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition, 2015. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.278, + 0.483, + 0.346 + ], + "angle": 0, + "content": "[27] Muhammad Uzair Khattak, Hanoona Rasheed, Muhammad Maaz, Salman Khan, and Fahad Shahbaz Khan. Maple: Multi-modal prompt learning. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition, 2023. 2, 5, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.348, + 0.483, + 0.417 + ], + "angle": 0, + "content": "[28] Muhammad Uzair Khattak, Syed Talal Wasim, Muzammal Naseer, Salman Khan, Ming-Hsuan Yang, and Fahad Shahbaz Khan. Self-regulating prompts: Foundational model adaptation without forgetting. In Proceedings of the International Conference on Computer Vision, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.419, + 0.483, + 0.473 + ], + "angle": 0, + "content": "[29] Ranjay Krishna, Kenji Hata, Frederic Ren, Li Fei-Fei, and Juan Carlos Niebles. Dense captioning events in videos. In Proceedings of the International Conference on Computer Vision, 2017. 5, 12" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.476, + 0.482, + 0.503 + ], + "angle": 0, + "content": "[30] Solomon Kullback. Information theory and statistics. Courier Corporation, 1997. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.505, + 0.483, + 0.574 + ], + "angle": 0, + "content": "[31] Jie Lei, Linjie Li, Luowei Zhou, Zhe Gan, Tamara L Berg, Mohit Bansal, and Jingjing Liu. Less is more: Clipbert for video-and-language learning via sparse sampling. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition, 2021. 12" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.576, + 0.483, + 0.645 + ], + "angle": 0, + "content": "[32] Junnan Li, Ramprasaath Selvaraju, Akhilesh Gotmare, Shafiq Joty, Caiming Xiong, and Steven Chu Hong Hoi. Align before fuse: Vision and language representation learning with momentum distillation. Advances in Neural Information Processing Systems, 2021. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.647, + 0.483, + 0.713 + ], + "angle": 0, + "content": "[33] Junnan Li, Dongxu Li, Caiming Xiong, and Steven Hoi. Blip: Bootstrapping language-image pre-training for unified vision-language understanding and generation. In Proceedings of the International Conference on Machine Learning, 2022." + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.716, + 0.483, + 0.785 + ], + "angle": 0, + "content": "[34] Kunchang Li, Yali Wang, Yizhuo Li, Yi Wang, Yinan He, Limin Wang, and Yu Qiao. Unmasked teacher: Towards training-efficient video foundation models. In Proceedings of the International Conference on Computer Vision, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.788, + 0.483, + 0.829 + ], + "angle": 0, + "content": "[35] Haotian Liu, Chunyuan Li, Yuheng Li, Bo Li, Yuanhan Zhang, Sheng Shen, and Yong Jae Lee. Llava next: Improved reasoning,ocr,and world knowledge,2024.4,8" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.831, + 0.483, + 0.872 + ], + "angle": 0, + "content": "[36] Ilya Loshchilov and Frank Hutter. Sgdr: Stochastic gradient descent with warm restarts. In Proceedings of the International Conference on Learning Representations, 2016. 12" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.874, + 0.483, + 0.902 + ], + "angle": 0, + "content": "[37] Huaishao Luo, Lei Ji, Ming Zhong, Yang Chen, Wen Lei, Nan Duan, and Tianrui Li. Clip4clip: An empirical study of" + }, + { + "type": "list", + "bbox": [ + 0.093, + 0.092, + 0.484, + 0.902 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.545, + 0.093, + 0.905, + 0.12 + ], + "angle": 0, + "content": "clip for end to end video clip retrieval and captioning. Neurocomputing, 2022. 1, 2, 5, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.122, + 0.905, + 0.189 + ], + "angle": 0, + "content": "[38] Zongyang Ma, Ziqi Zhang, Yuxin Chen, Zhongang Qi, Chunfeng Yuan, Bing Li, Yingmin Luo, Xu Li, Xiaojuan Qi, Ying Shan, et al. Ea-vtr: Event-aware video-text retrieval. In Proceedings of the European Conference on Computer Vision, 2024. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.191, + 0.905, + 0.259 + ], + "angle": 0, + "content": "[39] Antoine Miech, Dimitri Zhukov, Jean-Baptiste Alayrac, Makarand Tapaswi, Ivan Laptev, and Josef Sivic. Howto100m: Learning a text-video embedding by watching hundred million narrated video clips. In Proceedings of the International Conference on Computer Vision, 2019. 12" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.261, + 0.905, + 0.302 + ], + "angle": 0, + "content": "[40] Aaron van den Oord, Yazhe Li, and Oriol Vinyals. Representation learning with contrastive predictive coding. arXiv preprint arXiv:1807.03748, 2018. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.304, + 0.905, + 0.344 + ], + "angle": 0, + "content": "[41] Alec Radford, Jeffrey Wu, Rewon Child, David Luan, Dario Amodei, Ilya Sutskever, et al. Language models are unsupervised multitask learners. OpenAI blog, 2019. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.345, + 0.905, + 0.427 + ], + "angle": 0, + "content": "[42] Alec Radford, Jong Wook Kim, Chris Hallacy, Aditya Ramesh, Gabriel Goh, Sandhini Agarwal, Girish Sastry, Amanda Askell, Pamela Mishkin, Jack Clark, et al. Learning transferable visual models from natural language supervision. In Proceedings of the International Conference on Machine Learning, 2021. 1, 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.428, + 0.905, + 0.468 + ], + "angle": 0, + "content": "[43] Anna Rohrbach, Marcus Rohrbach, and Bernt Schiele. The long-short story of movie description. In Pattern Recognition, 2015. 5, 12" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.47, + 0.905, + 0.525 + ], + "angle": 0, + "content": "[44] Leqi Shen, Tianxiang Hao, Sicheng Zhao, Yifeng Zhang, Pengzhang Liu, Yongjun Bao, and Guiguang Ding. Tempme: Video temporal token merging for efficient text-video retrieval. arXiv preprint arXiv:2409.01156, 2024. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.526, + 0.905, + 0.594 + ], + "angle": 0, + "content": "[45] Leqi Shen, Tao He, Sicheng Zhao, Zhelun Shen, Yuchen Guo, Tianshi Xu, and Guiguang Ding. X-reid: Cross-instance transformer for identity-level person re-identification. In IEEE International Conference on Multimedia and Expo, 2024. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.596, + 0.905, + 0.65 + ], + "angle": 0, + "content": "[46] Yoad Tewel, Yoav Shalev, Idan Schwartz, and Lior Wolf. Zerocap: Zero-shot image-to-text generation for visual-semantic arithmetic. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition, 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.651, + 0.905, + 0.704 + ], + "angle": 0, + "content": "[47] Ao Wang, Hui Chen, Lihao Liu, Kai Chen, Zijia Lin, Jungong Han, et al. Yolov10: Real-time end-to-end object detection. Advances in Neural Information Processing Systems, 2024. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.706, + 0.905, + 0.775 + ], + "angle": 0, + "content": "[48] Jiamian Wang, Guohao Sun, Pichao Wang, Dongfang Liu, Sohail Dianat, Majid Rabbani, Raghuveer Rao, and Zhiqiang Tao. Text is mass: Modeling as stochastic embedding for text-video retrieval. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition, 2024. 1, 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.777, + 0.905, + 0.829 + ], + "angle": 0, + "content": "[49] Qiang Wang, Yanhao Zhang, Yun Zheng, Pan Pan, and Xi-ansheng Hua. Disentangled representation learning for text-video retrieval. arXiv preprint arXiv:2203.07111, 2022. 1, 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.832, + 0.905, + 0.9 + ], + "angle": 0, + "content": "[50] Yi Wang, Kunchang Li, Yizhuo Li, Yinan He, Bingkun Huang, Zhiyu Zhao, Hongjie Zhang, Jilan Xu, Yi Liu, Zun Wang, et al. Intervideo: General video foundation models via generative and discriminative learning. arXiv preprint arXiv:2212.03191, 2022. 2" + }, + { + "type": "list", + "bbox": [ + 0.517, + 0.093, + 0.905, + 0.9 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.518, + 0.957 + ], + "angle": 0, + "content": "19711" + } + ], + [ + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.092, + 0.482, + 0.16 + ], + "angle": 0, + "content": "[51] Wenhao Wu, Haipeng Luo, Bo Fang, Jingdong Wang, and Wanli Ouyang. Cap4video: What can auxiliary captions do for text-video retrieval? In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition, 2023. 1, 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.163, + 0.482, + 0.233 + ], + "angle": 0, + "content": "[52] Hu Xu, Saining Xie, Xiaqing Tan, Po-Yao Huang, Russell Howes, Vasu Sharma, Shang-Wen Li, Gargi Ghosh, Luke Zettlemoyer, and Christoph Feichtenhofer. Demystifying clip data. In Proceedings of the International Conference on Learning Representations, 2024. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.234, + 0.482, + 0.289 + ], + "angle": 0, + "content": "[53] Jun Xu, Tao Mei, Ting Yao, and Yong Rui. Msr-vtt: A large video description dataset for bridging video and language. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition, 2016. 5, 12" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.29, + 0.482, + 0.359 + ], + "angle": 0, + "content": "[54] Kelvin Xu, Jimmy Ba, Ryan Kiros, Kyunghyun Cho, Aaron Courville, Ruslan Salakhudinov, Rich Zemel, and Yoshua Bengio. Show, attend and tell: Neural image caption generation with visual attention. In Proceedings of the International Conference on Machine Learning, 2015. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.36, + 0.482, + 0.43 + ], + "angle": 0, + "content": "[55] Hongwei Xue, Yuchong Sun, Bei Liu, Jianlong Fu, Ruihua Song, Houqiang Li, and Jiebo Luo. Clip-vip: Adapting pre-trained image-text model to video-language alignment. In Proceedings of the International Conference on Learning Representations, 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.431, + 0.482, + 0.5 + ], + "angle": 0, + "content": "[56] Fan Yang, Sicheng Zhao, Yanhao Zhang, Haoxiang Chen, Hui Chen, Wenbo Tang, Haonan Lu, Pengfei Xu, Zhenyu Yang, Jungong Han, et al. Lmi3d: Empowering llm with 3d perception from a single 2d image. arXiv preprint arXiv:2408.07422, 2024. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.501, + 0.482, + 0.557 + ], + "angle": 0, + "content": "[57] Fan Yang, Ru Zhen, Jianing Wang, Yanhao Zhang, Haoxiang Chen, Haonan Lu, Sicheng Zhao, and Guiguang Ding. Heie: Mllm-based hierarchical explainable aigc image implausibility evaluator. arXiv preprint arXiv:2411.17261, 2024. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.559, + 0.482, + 0.613 + ], + "angle": 0, + "content": "[58] Xiangpeng Yang, Linchao Zhu, Xiaohan Wang, and Yi Yang. Dgl: Dynamic global-local prompt tuning for text-video retrieval. Proceedings of the AAAI Conference on Artificial Intelligence, 2024. 1, 2, 3, 5, 6, 12" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.615, + 0.482, + 0.655 + ], + "angle": 0, + "content": "[59] Bowen Zhang, Hexiang Hu, and Fei Sha. Cross-modal and hierarchical modeling of video and text. In Proceedings of the European Conference on Computer Vision, 2018. 12" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.657, + 0.482, + 0.685 + ], + "angle": 0, + "content": "[60] Yuanhan Zhang, Kaiyang Zhou, and Ziwei Liu. Neural prompt search. arXiv preprint arXiv:2206.04673, 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.687, + 0.482, + 0.741 + ], + "angle": 0, + "content": "[61] Kaiyang Zhou, Jingkang Yang, Chen Change Loy, and Ziwei Liu. Conditional prompt learning for vision-language models. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition, 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.743, + 0.482, + 0.784 + ], + "angle": 0, + "content": "[62] Kaiyang Zhou, Jingkang Yang, Chen Change Loy, and Ziwei Liu. Learning to prompt for vision-language models. International Journal of Computer Vision, 2022. 2" + }, + { + "type": "list", + "bbox": [ + 0.093, + 0.092, + 0.482, + 0.784 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.519, + 0.956 + ], + "angle": 0, + "content": "19712" + } + ] +] \ No newline at end of file diff --git a/2025/DiscoVLA_ Discrepancy Reduction in Vision, Language, and Alignment for Parameter-Efficient Video-Text Retrieval/9d7a1765-54c3-41c7-9719-3ce7d4a7b6fa_origin.pdf b/2025/DiscoVLA_ Discrepancy Reduction in Vision, Language, and Alignment for Parameter-Efficient Video-Text Retrieval/9d7a1765-54c3-41c7-9719-3ce7d4a7b6fa_origin.pdf new file mode 100644 index 0000000000000000000000000000000000000000..81c001ee9988e3def08e9e90d41bf2ea087b8b3e --- /dev/null +++ b/2025/DiscoVLA_ Discrepancy Reduction in Vision, Language, and Alignment for Parameter-Efficient Video-Text Retrieval/9d7a1765-54c3-41c7-9719-3ce7d4a7b6fa_origin.pdf @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:e7dcb9506a73507eb561a7b4bc44c87d5b66b972b43a7c511253ee0520531144 +size 941611 diff --git a/2025/DiscoVLA_ Discrepancy Reduction in Vision, Language, and Alignment for Parameter-Efficient Video-Text Retrieval/full.md b/2025/DiscoVLA_ Discrepancy Reduction in Vision, Language, and Alignment for Parameter-Efficient Video-Text Retrieval/full.md new file mode 100644 index 0000000000000000000000000000000000000000..7ca5516143a6378dd4e1bb3d3a7d9ef605c7ca6b --- /dev/null +++ b/2025/DiscoVLA_ Discrepancy Reduction in Vision, Language, and Alignment for Parameter-Efficient Video-Text Retrieval/full.md @@ -0,0 +1,330 @@ +# DiscoVLA: Discrepancy Reduction in Vision, Language, and Alignment for Parameter-Efficient Video-Text Retrieval + +Leqi Shen $^{1,2,4*}$ Guoqiang Gong $^{5\dagger}$ Tianxiang Hao $^{1,2}$ Tao He $^{6,7}$ Yifeng Zhang $^{5}$ Pengzhang Liu $^{5}$ Sicheng Zhao $^{2‡}$ Jungong Han $^{3}$ Guiguang Ding $^{1,2‡}$ + +$^{1}$ School of Software $^{2}$ BNrist $^{3}$ Department of Automation, Tsinghua University $^{4}$ Hangzhou Zhuoxi Institute of Brain and Intelligence + +$^{5}$ JD.com $^{6}$ GRG Banking Equipment Co., Ltd. $^{7}$ South China University of Technology + +# Abstract + +The parameter-efficient adaptation of the image-text pretraining model CLIP for video-text retrieval is a prominent area of research. While CLIP is focused on image-level vision-language matching, video-text retrieval demands comprehensive understanding at the video level. Three key discrepancies emerge in the transfer from image-level to video-level: vision, language, and alignment. However, existing methods mainly focus on vision while neglecting language and alignment. In this paper, we propose Discrepancy Reduction in Vision, Language, and Alignment (DiscoVLA), which simultaneously mitigates all three discrepancies. Specifically, we introduce Image-Video Features Fusion to integrate image-level and video-level features, effectively tackling both vision and language discrepancies. Additionally, we generate pseudo image captions to learn fine-grained image-level alignment. To mitigate alignment discrepancies, we propose Image-to-Video Alignment Distillation, which leverages image-level alignment knowledge to enhance video-level alignment. Extensive experiments demonstrate the superiority of our DiscoVLA. In particular, on MSRVTT with CLIP (ViT-B/16), DiscoVLA outperforms previous methods by $2.2\%$ R@1 and $7.5\%$ R@sum. The code is available at https://github.com/LunarShen/DsicoVLA. + +# 1. Introduction + +The rapid expansion of online video content has generated a growing demand for effective video-text retrieval, focusing on matching video content with relevant textual descriptions. Inspired by the success of image-text pretraining approaches such as CLIP [42], recent efforts [9, 14, 24, 37, 48, 49, 51] have increasingly focused on extend + +![](images/bfbe10255f720db020eda19693dec725fa3660b25e37918a8c85685474458566.jpg) +Figure 1. When parameter-efficient transferring image-level CLIP to video-text retrieval, we identify three key discrepancies: vision, alignment, and language. Unlike previous methods, our DiscoVLA aims to address all these discrepancies. More visualization examples are provided in Figure 4. + +
MethodImage-Level PETL Video-Level
VisionLanguageAlignment
Previous Methods✓××
Our DiscoVLA✓✓✓
+ +ing these powerful capabilities to video-text retrieval tasks. However, fully fine-tuning these pretrained models requires updating numerous parameters for every dataset, resulting in large storage overhead [22, 58]. Therefore, we focus on parameter-efficient transfer learning methods, targeting strong performance with minimal trainable parameters. + +Although both image-text pretraining and video-text retrieval involve vision-language (VL) matching, they exhibit + +substantial discrepancies that hinder straightforward knowledge transfer. To be specific, image-text pretraining focuses on matching images with image captions, which represents image-level VL matching. In contrast, video-text retrieval focuses on matching videos with video captions, which represents video-level VL matching. The challenges of video-level VL matching, which demands understanding motion and time, create significant obstacles for transferring knowledge from image-text to video-text tasks. + +In this work, we explore the discrepancies between image-level and video-level VL matching across three key aspects: vision, language, and alignment. Videos introduce a temporal dimension, which is absent in images. Similarly, the distinction between image captions and video captions arises from the varying levels of granularity, such as isolated image descriptions versus continuous video narratives. Thus, video-level alignment is inherently more complex than image-level alignment, as it must account for intricate spatio-temporal relationships. In Figure 1, the three images correspond to "preparing food", "pouring", and "cutting", while the entire video demonstrates an instructional cooking process. Image-level VL matching focuses on isolated images, while video-level VL matching emphasizes the overall understanding of continuous segments. + +However, current methods [5, 22, 25, 58] struggle to fully address the discrepancies mentioned above. Some [5, 22, 25, 58] focus on the vision aspect, modeling temporal relationships to extract video-level features, while others [25, 58] attempt to reduce the modality gap between vision and language through shared parameter mechanisms. Despite these advances, these methods remain limited in both language and alignment discrepancies. + +To tackle these challenges, we introduce Discrepancy Reduction in Vision, Language, and Alignment, termed DiscoVLA, a novel parameter-efficient method for videotext retrieval. Unlike previous methods that target a single gap, our method aims to simultaneously reduce all significant discrepancies, as illustrated in the table of Figure 1. + +Specifically, we propose an Image-Video Features Fusion module (IVFusion), which addresses both vision and language discrepancies through a unified feature extraction approach. IVFusion effectively combines both image- and video-level features by utilizing a lightweight adapter. In addition to video-level alignment, we propose Pseudo Image-level Alignment (PIImgAlign) which generates pseudo image captions and learns fine-grained image-level alignment. Furthermore, to reduce alignment discrepancy, we propose Image-to-Video Alignment Distillation (AlignDistill), transferring image-level alignment knowledge to improve video-level alignment. All proposed modules prioritize parameter efficiency: IVFusion requires minimal trainable parameters, while PImgAlign and AlignDistill operate without increasing inference parameters. + +In summary, the main contributions are as follows: + +- We reveal, for the first time, the necessity of addressing all three image-to-video discrepancies—vision, language, and alignment—to achieve parameter-efficient VTR. +- We introduce IVFusion to fuse image- and video-level features for both vision and language gaps. PImgAlign is introduced for fine-grained image-level alignment. We introduce AlignDistill to minimize alignment gaps. +- Our DiscoVLA achieves state-of-the-art results on MSRVTT, LSMDC, ActivityNet, and DiDeMo. Notably, on MSRVTT with CLIP (ViT-B/16), DiscoVLA reaches $50.5\%$ R@1, surpassing previous methods by $2.2\%$ . + +# 2. Related Work + +Video-Text Retrieval. Advances in vision-language pretraining [7, 12, 32-34, 42, 50, 52, 55] have achieved significant success across various downstream tasks [2, 10, 11, 18, 26, 45, 47, 54, 56, 57]. Leveraging powerful CLIP [42], recent video-text retrieval (VTR) methods [9, 14, 24, 37, 48, 49, 51] have achieved impressive performance enhancements. CLIP4Clip [37] is a pioneering CLIP-based method that aggregates image features via mean-pooling or transformers to obtain video features, inspiring subsequent studies. Additionally, Cap4Video [51] utilizes ZeroCap [46] and GPT-2 [41] to generate video-level captions as supplementary video content during both training and inference. In contrast, our DiscoVLA addresses the image-level and video-level gap by generating image-level captions used only in training to improve video-level alignment. These full fine-tuning methods impose substantial overhead from large trainable parameters. + +Parameter-Efficient Transfer Learning. Parameter-efficient methods aim to adapt pre-trained models to new tasks with minimal fine-tuning. The mainstream methods in natural language processing and image tasks include techniques such as Prompt [23, 27, 28, 61, 62], Adapter [6, 17, 20, 60], and LoRA [21, 60]. Prompt [27] modifies the input space by adding learnable tokens. Adapter [20] introduces a lightweight bottleneck neural network consisting of up- and down-projection layers with a non-linear activation function. LoRA [21] aims to update up- and down-projection layers by learning low-rank decomposition matrices. + +Recently, several studies have explored parameter-efficient VTR methods [5, 22, 25, 58] to incorporate temporal modeling into the frozen image-level CLIP backbone. VoP [22] employs trainable BiLSTMs [16] to produce prompt tokens in the vision encoder. DGL [58] introduces weight-sharing layers to generate prompts for both text and video encoders. MV-Adapter [25] introduces shared bottleneck structures in both video and text branches. However, these methods overlook language and alignment discrepancies. We address these limitations with DiscoVLA, a framework that facilitates robust video-level understanding. + +![](images/3baf7371eccbd1f136ffa4c85bcbf4d194ecc701dc4f2f62831d41c878f71323.jpg) +Figure 2. The overall framework of DiscoVLA. Initially, we generate pseudo image captions for each sampled image. In both vision and text encoders, we utilize image-level layers to acquire pretrained image-level knowledge and employ IVFusion layers to enhance spatiotemporal information. The single video caption is encoded through the text encoder, utilizing IVFusion layer as image-level (image-lvl) layer. Finally, AlignDistill is applied to distill image-level alignment to video-level alignment. For fair comparisons with previous methods, we do not generate pseudo image captions during the inference phase. + +# 3. Methodology + +# 3.1. Preliminary + +Video-text retrieval is a task that involves retrieving relevant videos based on a given text query, or conversely, retrieving appropriate text descriptions for a video. Consistent with previous approaches [5, 22, 25, 58], we utilize the CLIP model, pre-trained on image-text data, as our backbone. + +To process a given video, we first sample a sequence of frame images uniformly, which are then input into the CLIP vision encoder to extract image features $v^{\mathrm{img}}$ . These features are then aggregated through temporal average pooling to get video features $v^{\mathrm{vid}}$ . For a given video caption, it is passed through the CLIP text encoder to extract video caption features $t^{\mathrm{vid}}$ . Then, we compute the cross-modal contrastive loss [40], which functions as a video-level alignment term, maximizing the video-level similarity $\mathrm{Sim}^{\mathrm{vid}}$ between matched pairs: + +$$ +\operatorname {S i m} _ {i j} ^ {\mathrm {v i d}} = t _ {i} ^ {\mathrm {v i d} ^ {T}} v _ {j} ^ {\mathrm {v i d}}, \tag {1} +$$ + +$$ +\mathcal {L} _ {t 2 v} \left(\operatorname {S i m} ^ {\text {v i d}}\right) = \frac {1}{B} \sum_ {i = 1} ^ {B} \log \frac {\exp \left(\operatorname {S i m} _ {i i} ^ {\text {v i d}}\right) / \tau)}{\sum_ {j = 1} ^ {B} \exp \left(\operatorname {S i m} _ {i j} ^ {\text {v i d}}\right) / \tau)}, \tag {2} +$$ + +$$ +\mathcal {L} _ {v 2 t} \left(\operatorname {S i m} ^ {\text {v i d}}\right) = \frac {1}{B} \sum_ {i = 1} ^ {B} \log \frac {\exp \left(\operatorname {S i m} _ {i i} ^ {\text {v i d}}\right) / \tau)}{\sum_ {j = 1} ^ {B} \exp \left(\operatorname {S i m} _ {j i} ^ {\text {v i d}}\right) / \tau)}, \tag {3} +$$ + +$$ +\mathcal {L} _ {A} (\operatorname {S i m} ^ {\mathrm {v i d}}) = - \frac {1}{2} \left[ \mathcal {L} _ {t 2 v} (\operatorname {S i m} ^ {\mathrm {v i d}}) + \mathcal {L} _ {v 2 t} (\operatorname {S i m} ^ {\mathrm {v i d}}) \right], \tag {4} +$$ + +where $B$ is the batch size, $\tau$ is the temperature parameter. + +# 3.2. DiscoVLA + +As shown in Figure 2, we propose DiscoVLA to address the challenges of vision, language, and alignment when transferring from image-level to video-level VL matching. First, we introduce Image-Video Features Fusion (IVFusion) in both vision and text encoders to enhance feature extraction. Next, we introduce Pseudo Image-Level Alignment (PIImgAlign) for fine-grained image-level alignment, consisting of Pseudo Image Language Generation and Image-Level Alignment. Finally, Image-to-Video Alignment Distillation (AlignDistill) is proposed to leverage image-level alignment knowledge to enhance video-level alignment. + +Importantly, PImgAlign is only employed during training. For a fair comparison with previous methods and due to its generation inefficiency, PImgAlign is omitted at inference. The final similarity used for retrieval is based on video-level similarity in Eq. (1). + +We focus on parameter-efficient transfer learning for video-text retrieval, where the large backbone parameters remain fixed, and only a small number of additional parameters are trained. The trainable parameters are embedded in the parameter-efficient attention of each encoder layer, as illustrated in Figure 3. + +# 3.3. Image-Video Features Fusion + +Image-level and video-level variations exist in both vision and language modalities. Our proposed IVFusion fuses these image-level and video-level features in both the vision and language encoders. Below, we illustrate IVFusion's application within the vision encoder. + +![](images/0704c9a8a0461be70b62fb28713d053f6bc102bbe788c05808941ec59c7cb103.jpg) +Figure 3. Illustration of the encoder layers for vision and text encoders. (a) Image-Level Attn operates on each of the $F$ images or image captions individually. (b) Video-Level Attn concatenates tokens across all sampled images or image captions. (c) IVFusion Attn employs a lightweight adapter to integrate the efficiency of Image-Level Attn with the effectiveness of Video-Level Attn. Here, we illustrate the application of IVFusion within the vision encoder. The text encoder adopts the same approach for processing pseudo image captions. + +![](images/46a2b4606872b486ec8096af7f0399e265e85e9f2ff67c904082ded61a9105d2.jpg) + +Image-Level Attn. To achieve parameter-efficient transfer learning, we apply LoRA in the attention module of each encoder layer. As illustrated in Figure 3a, $F \times (N + 1)$ tokens are processed via Image-level Attn, where $F$ denotes the number of sampled images, and $N + 1$ represents the combination of the global CLS token and patch tokens. Image-Level Attn processes each of the $F$ images individually, using $N + 1$ tokens per image. However, Image-Level Attn does not encode temporal information. + +Video-Level Attn. A straightforward solution is Video-Level Attn, which concatenates tokens across all sampled images, enabling the attention module to learn spatiotemporal information. As shown in Figure 3b, $F(N + 1)$ tokens are simultaneously processed during the attention module. Due to the quadratic complexity of the attention mechanism, Image-Level Attn has a complexity of $O(F(N)^2)$ , while Video-Level Attn has a higher complexity of $O((FN)^2)$ . + +IVFusion Attn. To combine Image-Level Attn's efficiency with Video-Level Attn's effectiveness, we introduce IV-Fusion Attn in Figure 3c. We extend Image-Level Attn by incorporating a branch that uses all CLS tokens to extract spatio-temporal information, reducing complexity to $O(F(N)^2 + (F)^2 N)$ : + +$$ +\left[ c _ {i} ^ {(1)}, \hat {E} _ {i} \right] = \mathrm {A T T N} \left(Q K V = \left[ c _ {i} ^ {(0)}, E _ {i} ^ {(0)} \right]\right), \tag {5} +$$ + +$$ +\left[ c _ {1: F} ^ {(2)} \right] = \operatorname {A T T N} \left(Q = \left[ c _ {1: F} ^ {(0)} \right], K V = \left[ c _ {1: F} ^ {(0)}, E _ {1: F} ^ {(0)} \right]\right), \tag {6} +$$ + +$$ +\hat {c} _ {i} = c _ {i} ^ {(2)} + \sigma (c _ {i} ^ {(1)} \mathbf {W} _ {\mathrm {d o w n}}) \mathbf {W} _ {\mathrm {u p}}, \tag {7} +$$ + +where $c_{i}\in \mathbb{R}^{1\times D}$ is CLS token of the $i^{\mathrm{th}}$ image and $c_{1:F}\in \mathbb{R}^{F\times D}$ are CLS tokens of all sampled images. + +Similarly, $E_{i} \in \mathbb{R}^{N \times D}$ are patch tokens of the $i^{\text{th}}$ image and $E_{1:F} \in \mathbb{R}^{FN \times D}$ are patch tokens of all sampled images. In Eq. (7), we employ a trainable adapter to merge image-level features $c_{i}^{(1)}$ and video-level features $c_{i}^{(2)}$ , where $\mathbf{W}_{down} \in \mathbb{R}^{D \times r}$ and $\mathbf{W}_{up} \in \mathbb{R}^{r \times D}$ represent the up- and down-projection layers, and $\sigma$ denotes the nonlinear GELU activation [19]. Image-level features preserve pretrained spatial details, while video-level features capture temporal information. This fusion mechanism strengthens feature extraction. + +Our parameter-efficient vision encoder comprises both image-level and IVFusion layers. The shallow image-level layers employ Image-Level Attn to capture spatial information, while the top $H_{v}$ IVFusion layers leverage IVFusion Attn to focus on spatio-temporal information. Consequently, our vision encoder outputs enhanced image features $\left\{ v_{i}^{\mathrm{img}} \mid 1 \leq i \leq F \right\}$ . + +# 3.4. Pseudo Image-Level Alignment + +Unlike previous work focused on video-level alignment, we introduce PImgAlign to learn fine-grained image-level alignment via the decomposition of video content at the image level. To accomplish this, we first generate pseudo image captions in VTR datasets. In contrast to the straightforward sampling of images from a video, sampling image captions is impractical, as video captions generally lack the specific details needed for individual images. As a solution, we utilize the multimodal large language model (MLLM) LLaVA-NeXT [35] to generate pseudo image captions. For each sampled image from a video, we prompt the model + +using the relevant video caption for guidance: + +The provided image is a frame sampled from the video, which describes {video caption}. Based on the video's content, provide a caption for the provided image. + +Unlike EA-VTR [38], which uses an image captioner, our method employs a strong MLLM. By incorporating video caption guidance, it generates higher-quality image captions that align with the overall video content. Similarly to visual feature extraction, we adopt our proposed IVFusion to extract pseudo image caption features $\left\{t_i^{\mathrm{img}}|1\leq i\leq F\right\}$ . + +Given image features and pseudo image caption features from a paired video and video caption, we introduce a fine-grained image-level similarity. Since image captions are typically more concise than the images they describe, a single caption may correspond to multiple images, as shown in Figure 4d. Therefore, we calculate the maximum similarity for each image across all image captions and vice versa. Specifically, we computed a similarity matrix between image and image caption features, selecting the maximum value from each row and column to obtain the final image-level similarity: + +$$ +\operatorname {S i m} _ {i j - t 2 v} ^ {\mathrm {i m g}} = \frac {1}{F} \sum_ {n = 1} ^ {F} \max _ {1 \leq m \leq F} t _ {i | n} ^ {\mathrm {i m g} ^ {T}} v _ {j | m} ^ {\mathrm {i m g}}, \tag {8} +$$ + +$$ +\operatorname {S i m} _ {i j - v 2 t} ^ {\mathrm {i m g}} = \frac {1}{F} \sum_ {m = 1} ^ {F} \max _ {1 \leq n \leq F} t _ {i | n} ^ {\mathrm {i m g} ^ {T}} v _ {j | m} ^ {\mathrm {i m g}}, \tag {9} +$$ + +$$ +\operatorname {S i m} _ {i j} ^ {\mathrm {i m g}} = \frac {1}{2} \left[ \operatorname {S i m} _ {i j - t 2 v} ^ {\mathrm {i m g}} + \operatorname {S i m} _ {i j - v 2 t} ^ {\mathrm {i m g}} \right], \tag {10} +$$ + +where $\mathrm{Sim}_{ij}^{\mathrm{img}}$ denotes the image-level similarity between the $i^{\mathrm{th}}$ video caption and the $j^{\mathrm{th}}$ video. $t_{i|n}^{\mathrm{img}}$ denotes the $n^{\mathrm{th}}$ pseudo image caption features of the $i^{\mathrm{th}}$ video caption and $v_{j|m}^{\mathrm{img}}$ denotes the $m^{\mathrm{th}}$ image features of the $j^{\mathrm{th}}$ video. PImgAlign achieves video-level alignment from the image level, effectively emphasizing spatial details. + +However, in the inference stage, the lack of ground truth for video captions prevents pseudo image captions generation. Additionally, MLLM introduces a computational overhead. For fair comparisons with other methods, we exclude PImgAlign during inference. + +# 3.5. Image-to-Video Alignment Distillation + +Pretrained CLIP demonstrates a strong capability for image-level alignment. In contrast, video-text retrieval relies on comprehensive spatio-temporal alignment. Therefore, we propose AlignDistill to distill image-level alignment knowledge into video-level alignment, thus mitigating alignment discrepancies. Then, we optimize the Kullback-Leibler divergence [30] between image-level similarity and + +video-level similarity: + +$$ +\mathcal {S} _ {t 2 v} ^ {\mathrm {i m g}} = \left[ \mathrm {s} _ {i 1} ^ {\mathrm {i m g}}, \dots , \mathrm {s} _ {i N _ {v}} ^ {\mathrm {i m g}} \right], \mathcal {S} _ {v 2 t} ^ {\mathrm {i m g}} = \left[ \mathrm {s} _ {1 i} ^ {\mathrm {i m g}}, \dots , \mathrm {s} _ {N _ {t} i} ^ {\mathrm {i m g}} \right], \tag {11} +$$ + +$$ +\mathcal {S} _ {t 2 v} ^ {\mathrm {v i d}} = \left[ \mathrm {s} _ {i 1} ^ {\mathrm {v i d}}, \dots , \mathrm {s} _ {i N _ {v}} ^ {\mathrm {v i d}} \right], \mathcal {S} _ {v 2 t} ^ {\mathrm {v i d}} = \left[ \mathrm {s} _ {1 i} ^ {\mathrm {v i d}}, \dots , \mathrm {s} _ {N _ {t i}} ^ {\mathrm {v i d}} \right], \tag {12} +$$ + +$$ +\mathcal {L} _ {K L} = \frac {1}{2} \left(\mathbf {K L} \left(\mathcal {S} _ {t 2 v} ^ {\mathrm {i m g}} \| \mathcal {S} _ {t 2 v} ^ {\mathrm {v i d}}\right) + \mathbf {K L} \left(\mathcal {S} _ {v 2 t} ^ {\mathrm {i m g}} \| \mathcal {S} _ {v 2 t} ^ {\mathrm {v i d}}\right)\right). \tag {13} +$$ + +where $\mathrm{s}_{ij}^{\mathrm{img}}$ is $\mathrm{Sim}_{ij}^{\mathrm{img}}$ in Eq. (10) and $\mathrm{s}_{ij}^{\mathrm{vid}}$ is $\mathrm{Sim}_{ij}^{\mathrm{vid}}$ in Eq. (1). $S_{t2v}^{\mathrm{img}}$ and $S_{t2v}^{\mathrm{vid}}$ denote the probability distributions for the text-to-video task, while $S_{v2t}^{\mathrm{img}}$ and $S_{v2t}^{\mathrm{vid}}$ denote those for the video-to-text task. During the inference phase, We employ enhanced video-level similarity for retrieval tasks. + +Finally, the overall training objective, which includes alignment loss (Eq. (4)) and distillation loss (Eq. (13)), is formulated as follows: + +$$ +\mathcal {L} = \mathcal {L} _ {A} (\operatorname {S i m} ^ {\mathrm {v i d}}) + \alpha \mathcal {L} _ {A} (\operatorname {S i m} ^ {\mathrm {i m g}}) + \beta \mathcal {L} _ {K L}, \tag {14} +$$ + +where $\alpha$ and $\beta$ are used to balance the loss. $\mathcal{L}_A(\mathrm{Sim}^{\mathrm{img}})$ is the image-level alignment loss (see Eq. (4) and (10)). + +# 4. Experiments + +# 4.1. Experimental Settings + +Following common practice, we perform evaluations on four widely used benchmarks: MSRVTT [53], LSMDC [43], ActivityNet [29], and DiDeMo [1]. We evaluate the performance using common retrieval metrics such as Recall at K (R@K and K = 1, 5, 10), the sum of these recalls (R@sum), and Mean Rank (MnR). In all experiments, the LoRA dimension and the adapter dimension $r$ are set to 8. In Eq. 14, $\alpha$ and $\beta$ are set to 0.3 and 1.0, respectively. For the number of IVFusion layers, we set $H_V = 4$ for the vision encoder and $H_L = 2$ for the text encoder. Please see Appendix A for further details. + +# 4.2. Comparisons with State-of-the-Art Methods + +We compare our proposed DiscoVLA with state-of-the-art methods on popular benchmarks such as MSRVTT, LSMDC, ActivityNet, and DiDeMo. + +Full fine-tuning refers to CLIP4Clip-meanP [37], which averages image features along the temporal dimension. All parameter-efficient methods follow this temporal average pooling strategy to obtain the final video features. Although Prompt [27], Adapter [20], and LoRA [21] are widely adopted for parameter-efficient image and NLP tasks, they lack effective temporal modeling for video-level understanding. In contrast, methods like RAP [5], VoP [22], DGL [58], and MV-Adapter [25] focus on temporal modeling in the vision modality. + +As shown in Table 1, on MSRVTT, our DiscoVLA with CLIP (ViT-B/32) as the backbone achieves $47.0\%$ R@1 in + +
Method# Params (M)Text-to-VideoVideo-to-Text
R@1↑R@5↑R@10↑R@sum↑MnR↓R@1↑R@5↑R@10↑R@sum↑MnR↓
CLIP (ViT-B/32)
Full fine-tuning123.5443.170.480.8194.316.243.170.581.2194.812.4
Prompt [27]0.0840.466.377.3184.016.742.269.779.2191.112.4
Adapter [20]0.2641.969.978.7190.214.943.669.980.1193.611.5
LoRA [21]0.4943.768.980.4193.016.043.070.282.2195.412.0
RAP [5]1.0644.871.481.5197.714.444.071.982.4198.310.1
VoPF+P [22]0.443.569.379.3192.114.843.671.281.2196.011.0
VoPF+C [22]14.1044.770.579.2194.416.242.170.080.6192.713.4
DGL [58]0.8344.669.980.3194.816.344.570.780.6195.811.5
DiscoVLA0.5647.073.082.8202.814.147.773.683.6204.910.0
CLIP (ViT-B/16)
MV-Adapter [25]3.646.072.082.1200.1-45.674.083.8203.4-
RAP [5]1.0646.573.982.0202.412.145.376.484.8206.59.1
VoPF+P [22]0.447.172.481.8201.312.9-----
VoPF+C [22]14.1047.772.482.2202.312.0-----
DGL [58]0.8348.371.880.6200.713.445.774.082.9202.610.9
TempMe [44]0.5049.074.483.3206.711.947.675.385.4208.39.0
DiscoVLA0.5650.575.683.8209.912.149.276.084.7209.98.6
+ +Table 1. Comparisons with state-of-the-art methods on MSRVTT. # Params denotes the number of trainable parameters. $\uparrow$ denotes higher values represent better performance and $\downarrow$ denotes lower values represent better performance. The best value for each metric is highlighted in bold. R@sum is defined as the total of R@1, R@5, and R@10. The gray row represents the fully fine-tuned CLIP4Clip-meanP [37]. + +
Method# Params (M)Text-to-VideoVideo-to-Text
R@1↑R@5↑R@10↑R@sum↑MnR↓R@1↑R@5↑R@10↑R@sum↑MnR↓
Full fine-tuning123.5420.738.947.2106.865.320.639.447.5107.556.7
VoPF+P [22]0.420.740.749.7111.159.121.540.650.7112.850.8
VoPF+C [22]14.1021.140.949.6111.660.122.340.350.7113.351.1
DGL [58]0.8321.439.448.4109.264.3-----
DiscoVLA0.5623.642.052.3117.952.022.841.951.2115.946.6
+ +Table 2. Comparisons with state-of-the-art methods on LSMDC using CLIP (ViT-B/32). + +
Method# Params (M)Text-to-VideoVideo-to-Text
R@1↑R@5↑R@10↑R@sum↑MnR↓R@1↑R@5↑R@10↑R@sum↑MnR↓
Full fine-tuning123.5440.572.4--7.442.574.185.8202.46.6
RAP [5]1.0640.871.082.2194.08.3-----
VoPF+P [22]0.436.165.578.5180.110.936.365.979.2181.410.1
VoPF+C [22]14.1035.163.777.6176.411.435.665.977.8179.310.4
DGL [58]0.8338.669.281.6189.49.0-----
DiscoVLA0.5641.272.483.6197.27.841.872.884.7199.36.9
+ +Table 3. Comparisons with state-of-the-art methods on ActivityNet using CLIP (ViT-B/32). + +the text-to-video task (t2v) and $47.7\%$ R@1 in the video-to-text task (v2t), significantly outperforming previous methods. Furthermore, when using CLIP (ViT-B/16), DiscoVLA achieves improvements of $2.2\%$ R@1 and $7.5\%$ R@sum. The comparison results on LSMDC, ActivityNet, DiDeMo are shown in Table 2-4. Our method outperforms all existing approaches across these datasets. Existing approaches overlook the critical image-to-video language and alignment discrepancies. In contrast, our proposed DiscoVLA addresses these discrepancies across all three areas: vision, language, and alignment. + +# 4.3. Ablation Study + +Ablation study on individual components. We evaluate the proposed components in Table 5. LoRA applies Image-Level Attn (see Figure 3a) in both the vision and text + +encoders. In B1, IVFusion significantly enhances performance by $8.7\%$ R@sum in t2v and $6.7\%$ R@sum in v2t, effectively integrating image-level and video-level features. In B2 and B3, we find that using PImgAlign or AlignDistill independently yields limited improvements. This is largely due to the notable gap between image-level and video-level (see Figure 1). Image-level PImgAlign struggles to impact video-level tasks, while AlignDistill lacks sufficient image-specific learning. When combined, however, they complement each other. PImgAlign captures fine-grained image-level alignments, and AlignDistill effectively transfers this knowledge to the video-level alignment. This improves $1.1\%$ R@sum in t2v and $2.8\%$ R@sum in v2t over B1. Finally, our full DiscoVLA improves $9.8\%$ R@sum in t2v and $9.5\%$ R@sum in v2t over LoRA. + +
Method# Params (M)Text-to-VideoVideo-to-Text
R@1↑R@5↑R@10↑R@sum↑MnR↓R@1↑R@5↑R@10↑R@sum↑MnR↓
Full fine-tuning123.5443.470.280.6194.217.542.570.680.2193.311.6
RAP [5]1.0642.670.479.6192.618.0-----
VoPF+P [22]0.445.372.380.4198.013.844.771.281.1197.09.9
VoPF+C [22]14.1046.471.981.5199.813.644.471.881.8198.09.5
DiscoVLA0.5648.474.582.7205.614.047.774.483.8205.99.3
+ +Table 4. Comparisons with state-of-the-art methods on DiDeMo using CLIP (ViT-B/32). + +
MethodsComponents# Params (M)Text-to-VideoVideo-to-Text
IVFusionPIImgAlignAlignDistillR@1↑R@5↑R@10↑R@sum↑R@1↑R@5↑R@10↑R@sum↑
Pretraind CLIPXXX030.853.863.3147.926.650.162.0138.7
LoRAXXX0.4943.768.980.4193.043.070.282.2195.4
B1✓XX0.5446.573.282.0201.746.672.682.9202.1
B2✓✓X0.5647.073.181.3201.445.772.982.4201.0
B3✓X✓0.5446.671.681.0199.245.971.982.2200.0
DiscoVLA✓✓✓0.5647.073.082.8202.847.773.683.6204.9
+ +Table 5. Ablation study on the contribution of each proposed component on MSRVTT using CLIP (ViT-B/32). Pretrained CLIP denotes the zero-shot performance of CLIP without any additional training. Our proposed methods are built upon LoRA. B1, B2, and B3 denote various combinations of our proposed components. + +
Parameter-Efficient AttentionText-to-Video
R@1↑R@5↑R@10↑R@sum↑
Image-Level Attn43.768.980.4193.0
Video-Level Attn45.770.980.8197.4
IVFusion Attn w/o Adapter46.071.080.4197.4
IVFusion Attn46.573.282.0201.7
+ +Table 6. Ablation study on IVFusion on MSRVTT using CLIP (ViT-B/32). We compare several implementations against our proposed IVFusion for Parameter-Efficient Attention (see Figure 3). Image-Level Attn represents the LoRA baseline. + +
Similarity in PImgAlignText-to-Video
R@1↑R@5↑R@10↑R@sum↑
None46.671.681.0199.2
Video-Level46.372.381.2199.8
Paired Image-Level46.173.282.4201.7
Fine-grained Image-Level47.073.082.8202.8
+ +Analysis of trainable parameters. The trainable parameters of our DiscoVLA consist of the LoRA and Adapter parts. As shown in Table 5, LoRA in both the vision and text encoders contain $\sim 0.49\mathrm{M}$ parameters. In B1, an additional $\sim 0.05\mathrm{M}$ parameters are introduced by the Adapter in the $H_{V} = 4$ IVFusion layers of the vision encoder. In B2, an additional $\sim 0.02\mathrm{M}$ parameters are introduced by the Adapter in the $H_{L} = 2$ IVFusion layers of the text encoder, which are utilized for pseudo-image caption features in PImgAlign. Overall, the total number of trainable parameters of DiscoVLA is $\sim 0.56\mathrm{M}$ . + +Ablation study on IVFusion. In Table 6, we compare different strategies for parameter-efficient attention in Figure 3. Video-Level Attn concatenates all tokens across sampled images to capture spatio-temporal information, + +Table 7. Ablation study on similarity in PImgAlign on MSRVTT using CLIP (ViT-B/32). None represents DiscoVLA without image-level alignment optimization, shown as B3 in Table 5. Paired Image-Level is based on matched image-caption pairs. Fine-grained Image-Level represents the full DiscoVLA. + +
Image Caption GenerationText-to-Video
R@1↑R@5↑R@10↑R@sum↑
DiscoVLA47.073.082.8202.8
W/O VidCap46.1 (-0.9)73.082.1 (-0.7)201.2 (-1.6)
+ +Table 8. Ablation study on image caption generation on MSRVTT using CLIP (ViT-B/32). W/O VidCap denotes generating pseudo image captions without video caption guidance. + +achieving a notable $4.4\%$ R@sum improvement. To reduce the computational complexity, IVFusion applies spatiotemporal attention only to global CLS tokens and merges image-level features with video-level features. Without an adapter, averaging image- and video-level features results in a modest performance increase. In contrast, with a learnable adapter, IVFusion merges the features more effectively, achieving $201.7\%$ R@sum. + +Ablation study on similarity in PImgAlign. Table 7 demonstrates the effectiveness of our proposed fine-grained image-level similarity in PImgAlign. The similarity is computed between sampled images and pseudo image captions from a paired video and video caption. The Video-Level baseline averages the image features and pseudo-image caption features along the temporal dimension to compute the video-level similarity, replacing fine-grained image-level similarity. However, this modification leads to a notable decrease in performance, emphasizing the importance of image-level similarity in PImgAlign. In contrast, we introduce the Paired Image-Level baseline, which computes similarity based on matched image-caption pairs. Compared to Video-Level, Paired Image-Level yields a $1.9\%$ gain in R@sum. Moreover, we observe in Figure 4d that a single image caption may describe multiple images. Consequently, our fine-grained image-level similarity, as defined in Eq. 10, selects the maximum similarity from all available pairs, further enhancing video-level alignment. + +![](images/0b53ce37bcdb38f14d71a1c9fbed0ebd0a0de2a600f4af6fc821dc44e4c1a3bc.jpg) +Figure 4. Visualization of discrepancies between video-level and image-level data. Each example consists of a paired video and video caption, with four sampled images from the video. Video captions are highlighted with orange solid lines, and pseudo image captions generated by LLaVA-NeXT [35] are indicated with green dashed lines. While video captions convey the general context of the video, image captions focus on the detailed context of each individual frame. + +![](images/7e4c331fad03523b338be4fd02a3c63a13d7db79123d524fa9d44dff7d422ba4.jpg) +Figure 5. Comparison of image caption generation with (W/) and without (W/O) video caption guidance (VidCap) in PImgAlign. + +Ablation study on image caption generation in PImgAlign. Figure 5 and Table 8 show qualitative and quantitative evaluations of video caption guidance. In Figure 5, video captions serve as the global context that enriches the understanding of individual frames (marked in red). In Table 8, without this guidance, image captions focus only on salient regions while missing globally relevant details, leading to performance drops. In PImgAlign, we employ MLLM to process both frame images and video descriptions, generating image captions that align with the overall video content. These high-quality captions further enhance DiscoVLA's ability to learn robust image-level alignment. + +More Ablation Analysis. In Appendix B, we conduct additional ablation studies. Specifically, we evaluate the effects of hyperparameters $\alpha$ , $\beta$ , $H_V$ and $H_V$ . For fairness, all results in the main paper are reported without post-processing. Complementary evaluations with Q-Norm [4] and DSL [8] + +as post-processing methods are provided in the appendix. + +# 4.4. Qualitative Analysis + +To compare image-level and video-level data, we conduct a visualization experiment in Figure 4. In Figure 4a and 4b, image-level data alone fails to capture the complete content of the video. Figure 4c and 4d demonstrate a strong correlation between the video-level and image-level content, but differences across images lead to varying image captions. Visually, a video is composed of a sequence of images, incorporating temporal information. Textually, image captions are specific to individual images, while video captions describe the overall video content. These discrepancies hinder the effective transfer of image-level CLIP. Our approach addresses this issue by utilizing image-level alignment to enhance video-level alignment. + +# 5. Conclusion + +In this paper, we identify three major discrepancies in the parameter-efficient transfer of image-level CLIP to videotext retrieval: vision, language, and alignment. To address these challenges, we propose DiscoVLA. Specifically, IV-Fusion fuses image-level and video-level features to tackle the vision and language discrepancies. For alignment discrepancies, we propose PImgAlign and AlignDistill, which learn and distill image-level alignment to enhance video-level alignment. Our DiscoVLA significantly outperforms existing methods, achieving state-of-the-art performance. + +# Acknowledgment + +# References + +[1] Lisa Anne Hendricks, Oliver Wang, Eli Shechtman, Josef Sivic, Trevor Darrell, and Bryan Russell. Localizing moments in video with natural language. In Proceedings of the International Conference on Computer Vision, 2017. 5, 12 +[2] Stanislaw Antol, Aishwarya Agrawal, Jiasen Lu, Margaret Mitchell, Dhruv Batra, C Lawrence Zitnick, and Devi Parikh. Vqa: Visual question answering. In Proceedings of the International Conference on Computer Vision, 2015. 2 +[3] Max Bain, Arsha Nagrani, Gül Varol, and Andrew Zisserman. Frozen in time: A joint video and image encoder for end-to-end retrieval. In Proceedings of the International Conference on Computer Vision, 2021. 12 +[4] Simion-Vlad Bogolin, Ioana Croitoru, Hailin Jin, Yang Liu, and Samuel Albanie. Cross modal retrieval with querybank normalisation. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition, 2022. 8, 13 +[5] Meng Cao, Haoran Tang, Jinfa Huang, Peng Jin, Can Zhang, Ruyang Liu, Long Chen, Xiaodan Liang, Li Yuan, and Ge Li. RAP: Efficient text-video retrieval with sparse-and-correlated adapter. In Association for Computational Linguistics, 2024. 2, 3, 5, 6, 7, 12 +[6] Shoufa Chen, Chongjian Ge, Zhan Tong, Jiangliu Wang, Yibing Song, Jue Wang, and Ping Luo. Adaptformer: Adapting vision transformers for scalable visual recognition. Advances in Neural Information Processing Systems, 2022. 2 +[7] Sihan Chen, Handong Li, Qunbo Wang, Zijia Zhao, Mingzhen Sun, Xinxin Zhu, and Jing Liu. Vast: A vision-audio-subtitle-text omni-modality foundation model and dataset. Advances in Neural Information Processing Systems, 2023. 2 +[8] Xing Cheng, Hezheng Lin, Xiangyu Wu, Fan Yang, and Dong Shen. Improving video-text retrieval by multi-stream corpus alignment and dual softmax loss. arXiv preprint arXiv:2109.04290, 2021. 8, 13 +[9] Chaorui Deng, Qi Chen, Pengda Qin, Da Chen, and Qi Wu. Prompt switch: Efficient clip adaptation for text-video retrieval. In Proceedings of the International Conference on Computer Vision, 2023. 1, 2 +[10] Xiaohan Ding, Yuchen Guo, Guiguang Ding, and Jungong Han. Acnet: Strengthening the kernel skeletons for powerful cnn via asymmetric convolution blocks. In Proceedings of the International Conference on Computer Vision, 2019. 2 + +This work was supported by Beijing Natural Science Foundation (No. L223023), National Natural Science Foundation of China (Nos. 62441235, 62021002, 62441614), the Key R & D Program of Xinjiang, China (2022B01006), Zhejiang Provincial Natural Science Foundation of China under Grant (No. LDT23F01013F01), CCF-DiDi GAIA Collaborative Research Funds, China Postdoctoral Science Foundation (2024M750565), Guangdong S & T Program (2024B0101040008), and the Key Realm Research and Development Program of Guangzhou (No. 2024B01W0007). +[11] Xiaohan Ding, Xiangyu Zhang, Ningning Ma, Jungong Han, Guiguang Ding, and Jian Sun. Repvgg: Making vgg-style convnets great again. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition, 2021. 2 +[12] Lijie Fan, Dilip Krishnan, Phillip Isola, Dina Katabi, and Yonglong Tian. Improving clip training with language rewrites. Advances in Neural Information Processing Systems, 2024. 2 +[13] Valentin Gabeur, Chen Sun, Karteek Alahari, and Cordelia Schmid. Multi-modal transformer for video retrieval. In Proceedings of the European Conference on Computer Vision, 2020. 12 +[14] Satya Krishna Gorti, Noel Vouitsis, Junwei Ma, Keyvan Golestan, Maksims Volkovs, Animesh Garg, and Guangwei Yu. X-pool: Cross-modal language-video attention for text-video retrieval. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition, 2022. 1, 2 +[15] Priya Goyal, Piotr Dólar, Ross Girshick, Pieter Noordhuis, Lukasz Wesolowski, Aapo Kyrola, Andrew Tulloch, Yangqing Jia, and Kaiming He. Accurate, large minibatch sgd: Training imagenet in 1 hour. arXiv preprint arXiv:1706.02677, 2017. 12 +[16] Alex Graves and Jürgen Schmidhuber. Framewise phoneme classification with bidirectional LSTM and other neural network architectures. Neural Networks, 2005. 2 +[17] Junxian He, Chunting Zhou, Xuezhe Ma, Taylor Berg-Kirkpatrick, and Graham Neubig. Towards a unified view of parameter-efficient transfer learning. In Proceedings of the International Conference on Learning Representations, 2021. 2 +[18] Tao He, Leqi Shen, Yuchen Guo, Guiguang Ding, and Zhenhua Guo. Secret: Self-consistent pseudo label refinement for unsupervised domain adaptive person re-identification. In Proceedings of the AAAI Conference on Artificial Intelligence, 2022. 2 +[19] Dan Hendrycks and Kevin Gimpel. Gaussian error linear units (gelus). arXiv preprint arXiv:1606.08415, 2016. 4 +[20] Neil Houlsby, Andrei Giurgiu, Stanislaw Jastrzebski, Bruna Morrone, Quentin De Laroussilhe, Andrea Gesmundo, Mona Attariyan, and Sylvain Gelly. Parameter-efficient transfer learning for nlp. In Proceedings of the International Conference on Machine Learning, 2019. 2, 5, 6 +[21] Edward J Hu, Yelong Shen, Phillip Wallis, Zeyuan Allen-Zhu, Yuanzhi Li, Shean Wang, Lu Wang, and Weizhu Chen. Lora: Low-rank adaptation of large language models. arXiv preprint arXiv:2106.09685, 2021. 2, 5, 6 +[22] Siteng Huang, Biao Gong, Yulin Pan, Jianwen Jiang, Yiliang Lv, Yuyuan Li, and Donglin Wang. Vop: Text-video cooperative prompt tuning for cross-modal retrieval. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition, 2023. 1, 2, 3, 5, 6, 7, 12 +[23] Menglin Jia, Luming Tang, Bor-Chun Chen, Claire Cardie, Serge Belongie, Bharath Hariharan, and Ser-Nam Lim. Visual prompt tuning. In Proceedings of the European Conference on Computer Vision, 2022. 2 +[24] Peng Jin, Jina Huang, Pengfei Xiong, Shangxuan Tian, Chang Liu, Xiangyang Ji, Li Yuan, and Jie Chen. Videotext as game players: Hierarchical banzhaf interaction for + +cross-modal representation learning. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition, 2023. 1, 2 +[25] Xiaojie Jin, Bowen Zhang, Weibo Gong, Kai Xu, Xueqing Deng, Peng Wang, Zhao Zhang, Xiaohui Shen, and Jiashi Feng. Mv-adapter: Multimodal video transfer learning for video text retrieval. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition, 2024. 2, 3, 5, 6, 12 +[26] Andrej Karpathy and Li Fei-Fei. Deep visual-semantic alignments for generating image descriptions. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition, 2015. 2 +[27] Muhammad Uzair Khattak, Hanoona Rasheed, Muhammad Maaz, Salman Khan, and Fahad Shahbaz Khan. Maple: Multi-modal prompt learning. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition, 2023. 2, 5, 6 +[28] Muhammad Uzair Khattak, Syed Talal Wasim, Muzammal Naseer, Salman Khan, Ming-Hsuan Yang, and Fahad Shahbaz Khan. Self-regulating prompts: Foundational model adaptation without forgetting. In Proceedings of the International Conference on Computer Vision, 2023. 2 +[29] Ranjay Krishna, Kenji Hata, Frederic Ren, Li Fei-Fei, and Juan Carlos Niebles. Dense captioning events in videos. In Proceedings of the International Conference on Computer Vision, 2017. 5, 12 +[30] Solomon Kullback. Information theory and statistics. Courier Corporation, 1997. 5 +[31] Jie Lei, Linjie Li, Luowei Zhou, Zhe Gan, Tamara L Berg, Mohit Bansal, and Jingjing Liu. Less is more: Clipbert for video-and-language learning via sparse sampling. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition, 2021. 12 +[32] Junnan Li, Ramprasaath Selvaraju, Akhilesh Gotmare, Shafiq Joty, Caiming Xiong, and Steven Chu Hong Hoi. Align before fuse: Vision and language representation learning with momentum distillation. Advances in Neural Information Processing Systems, 2021. 2 +[33] Junnan Li, Dongxu Li, Caiming Xiong, and Steven Hoi. Blip: Bootstrapping language-image pre-training for unified vision-language understanding and generation. In Proceedings of the International Conference on Machine Learning, 2022. +[34] Kunchang Li, Yali Wang, Yizhuo Li, Yi Wang, Yinan He, Limin Wang, and Yu Qiao. Unmasked teacher: Towards training-efficient video foundation models. In Proceedings of the International Conference on Computer Vision, 2023. 2 +[35] Haotian Liu, Chunyuan Li, Yuheng Li, Bo Li, Yuanhan Zhang, Sheng Shen, and Yong Jae Lee. Llava next: Improved reasoning,ocr,and world knowledge,2024.4,8 +[36] Ilya Loshchilov and Frank Hutter. Sgdr: Stochastic gradient descent with warm restarts. In Proceedings of the International Conference on Learning Representations, 2016. 12 +[37] Huaishao Luo, Lei Ji, Ming Zhong, Yang Chen, Wen Lei, Nan Duan, and Tianrui Li. Clip4clip: An empirical study of + +clip for end to end video clip retrieval and captioning. Neurocomputing, 2022. 1, 2, 5, 6 +[38] Zongyang Ma, Ziqi Zhang, Yuxin Chen, Zhongang Qi, Chunfeng Yuan, Bing Li, Yingmin Luo, Xu Li, Xiaojuan Qi, Ying Shan, et al. Ea-vtr: Event-aware video-text retrieval. In Proceedings of the European Conference on Computer Vision, 2024. 5 +[39] Antoine Miech, Dimitri Zhukov, Jean-Baptiste Alayrac, Makarand Tapaswi, Ivan Laptev, and Josef Sivic. Howto100m: Learning a text-video embedding by watching hundred million narrated video clips. In Proceedings of the International Conference on Computer Vision, 2019. 12 +[40] Aaron van den Oord, Yazhe Li, and Oriol Vinyals. Representation learning with contrastive predictive coding. arXiv preprint arXiv:1807.03748, 2018. 3 +[41] Alec Radford, Jeffrey Wu, Rewon Child, David Luan, Dario Amodei, Ilya Sutskever, et al. Language models are unsupervised multitask learners. OpenAI blog, 2019. 2 +[42] Alec Radford, Jong Wook Kim, Chris Hallacy, Aditya Ramesh, Gabriel Goh, Sandhini Agarwal, Girish Sastry, Amanda Askell, Pamela Mishkin, Jack Clark, et al. Learning transferable visual models from natural language supervision. In Proceedings of the International Conference on Machine Learning, 2021. 1, 2 +[43] Anna Rohrbach, Marcus Rohrbach, and Bernt Schiele. The long-short story of movie description. In Pattern Recognition, 2015. 5, 12 +[44] Leqi Shen, Tianxiang Hao, Sicheng Zhao, Yifeng Zhang, Pengzhang Liu, Yongjun Bao, and Guiguang Ding. Tempme: Video temporal token merging for efficient text-video retrieval. arXiv preprint arXiv:2409.01156, 2024. 6 +[45] Leqi Shen, Tao He, Sicheng Zhao, Zhelun Shen, Yuchen Guo, Tianshi Xu, and Guiguang Ding. X-reid: Cross-instance transformer for identity-level person re-identification. In IEEE International Conference on Multimedia and Expo, 2024. 2 +[46] Yoad Tewel, Yoav Shalev, Idan Schwartz, and Lior Wolf. Zerocap: Zero-shot image-to-text generation for visual-semantic arithmetic. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition, 2022. 2 +[47] Ao Wang, Hui Chen, Lihao Liu, Kai Chen, Zijia Lin, Jungong Han, et al. Yolov10: Real-time end-to-end object detection. Advances in Neural Information Processing Systems, 2024. 2 +[48] Jiamian Wang, Guohao Sun, Pichao Wang, Dongfang Liu, Sohail Dianat, Majid Rabbani, Raghuveer Rao, and Zhiqiang Tao. Text is mass: Modeling as stochastic embedding for text-video retrieval. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition, 2024. 1, 2 +[49] Qiang Wang, Yanhao Zhang, Yun Zheng, Pan Pan, and Xi-ansheng Hua. Disentangled representation learning for text-video retrieval. arXiv preprint arXiv:2203.07111, 2022. 1, 2 +[50] Yi Wang, Kunchang Li, Yizhuo Li, Yinan He, Bingkun Huang, Zhiyu Zhao, Hongjie Zhang, Jilan Xu, Yi Liu, Zun Wang, et al. Intervideo: General video foundation models via generative and discriminative learning. arXiv preprint arXiv:2212.03191, 2022. 2 + +[51] Wenhao Wu, Haipeng Luo, Bo Fang, Jingdong Wang, and Wanli Ouyang. Cap4video: What can auxiliary captions do for text-video retrieval? In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition, 2023. 1, 2 +[52] Hu Xu, Saining Xie, Xiaqing Tan, Po-Yao Huang, Russell Howes, Vasu Sharma, Shang-Wen Li, Gargi Ghosh, Luke Zettlemoyer, and Christoph Feichtenhofer. Demystifying clip data. In Proceedings of the International Conference on Learning Representations, 2024. 2 +[53] Jun Xu, Tao Mei, Ting Yao, and Yong Rui. Msr-vtt: A large video description dataset for bridging video and language. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition, 2016. 5, 12 +[54] Kelvin Xu, Jimmy Ba, Ryan Kiros, Kyunghyun Cho, Aaron Courville, Ruslan Salakhudinov, Rich Zemel, and Yoshua Bengio. Show, attend and tell: Neural image caption generation with visual attention. In Proceedings of the International Conference on Machine Learning, 2015. 2 +[55] Hongwei Xue, Yuchong Sun, Bei Liu, Jianlong Fu, Ruihua Song, Houqiang Li, and Jiebo Luo. Clip-vip: Adapting pre-trained image-text model to video-language alignment. In Proceedings of the International Conference on Learning Representations, 2022. 2 +[56] Fan Yang, Sicheng Zhao, Yanhao Zhang, Haoxiang Chen, Hui Chen, Wenbo Tang, Haonan Lu, Pengfei Xu, Zhenyu Yang, Jungong Han, et al. Lmi3d: Empowering llm with 3d perception from a single 2d image. arXiv preprint arXiv:2408.07422, 2024. 2 +[57] Fan Yang, Ru Zhen, Jianing Wang, Yanhao Zhang, Haoxiang Chen, Haonan Lu, Sicheng Zhao, and Guiguang Ding. Heie: Mllm-based hierarchical explainable aigc image implausibility evaluator. arXiv preprint arXiv:2411.17261, 2024. 2 +[58] Xiangpeng Yang, Linchao Zhu, Xiaohan Wang, and Yi Yang. Dgl: Dynamic global-local prompt tuning for text-video retrieval. Proceedings of the AAAI Conference on Artificial Intelligence, 2024. 1, 2, 3, 5, 6, 12 +[59] Bowen Zhang, Hexiang Hu, and Fei Sha. Cross-modal and hierarchical modeling of video and text. In Proceedings of the European Conference on Computer Vision, 2018. 12 +[60] Yuanhan Zhang, Kaiyang Zhou, and Ziwei Liu. Neural prompt search. arXiv preprint arXiv:2206.04673, 2022. 2 +[61] Kaiyang Zhou, Jingkang Yang, Chen Change Loy, and Ziwei Liu. Conditional prompt learning for vision-language models. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition, 2022. 2 +[62] Kaiyang Zhou, Jingkang Yang, Chen Change Loy, and Ziwei Liu. Learning to prompt for vision-language models. International Journal of Computer Vision, 2022. 2 \ No newline at end of file diff --git a/2025/DiscoVLA_ Discrepancy Reduction in Vision, Language, and Alignment for Parameter-Efficient Video-Text Retrieval/images.zip b/2025/DiscoVLA_ Discrepancy Reduction in Vision, Language, and Alignment for Parameter-Efficient Video-Text Retrieval/images.zip new file mode 100644 index 0000000000000000000000000000000000000000..955d8d5f4659b7783106a94dbadc17f6fba9f235 --- /dev/null +++ b/2025/DiscoVLA_ Discrepancy Reduction in Vision, Language, and Alignment for Parameter-Efficient Video-Text Retrieval/images.zip @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:6ee3f69f423a6aac9d0f12eb600ab5c92e1e668683a9b39039a6ca2788f14b60 +size 892256 diff --git a/2025/DiscoVLA_ Discrepancy Reduction in Vision, Language, and Alignment for Parameter-Efficient Video-Text Retrieval/layout.json b/2025/DiscoVLA_ Discrepancy Reduction in Vision, Language, and Alignment for Parameter-Efficient Video-Text Retrieval/layout.json new file mode 100644 index 0000000000000000000000000000000000000000..0bd18b7e8a8c4700984ffa462573a510fa96ef9a --- /dev/null +++ b/2025/DiscoVLA_ Discrepancy Reduction in Vision, Language, and Alignment for Parameter-Efficient Video-Text Retrieval/layout.json @@ -0,0 +1,9070 @@ +{ + "pdf_info": [ + { + "para_blocks": [ + { + "bbox": [ + 85, + 102, + 526, + 138 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 85, + 102, + 526, + 138 + ], + "spans": [ + { + "bbox": [ + 85, + 102, + 526, + 138 + ], + "type": "text", + "content": "DiscoVLA: Discrepancy Reduction in Vision, Language, and Alignment for Parameter-Efficient Video-Text Retrieval" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 90, + 159, + 518, + 190 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 90, + 159, + 518, + 190 + ], + "spans": [ + { + "bbox": [ + 90, + 159, + 518, + 190 + ], + "type": "text", + "content": "Leqi Shen" + }, + { + "bbox": [ + 90, + 159, + 518, + 190 + ], + "type": "inline_equation", + "content": "^{1,2,4*}" + }, + { + "bbox": [ + 90, + 159, + 518, + 190 + ], + "type": "text", + "content": " Guoqiang Gong" + }, + { + "bbox": [ + 90, + 159, + 518, + 190 + ], + "type": "inline_equation", + "content": "^{5\\dagger}" + }, + { + "bbox": [ + 90, + 159, + 518, + 190 + ], + "type": "text", + "content": " Tianxiang Hao" + }, + { + "bbox": [ + 90, + 159, + 518, + 190 + ], + "type": "inline_equation", + "content": "^{1,2}" + }, + { + "bbox": [ + 90, + 159, + 518, + 190 + ], + "type": "text", + "content": " Tao He" + }, + { + "bbox": [ + 90, + 159, + 518, + 190 + ], + "type": "inline_equation", + "content": "^{6,7}" + }, + { + "bbox": [ + 90, + 159, + 518, + 190 + ], + "type": "text", + "content": " Yifeng Zhang" + }, + { + "bbox": [ + 90, + 159, + 518, + 190 + ], + "type": "inline_equation", + "content": "^{5}" + }, + { + "bbox": [ + 90, + 159, + 518, + 190 + ], + "type": "text", + "content": " Pengzhang Liu" + }, + { + "bbox": [ + 90, + 159, + 518, + 190 + ], + "type": "inline_equation", + "content": "^{5}" + }, + { + "bbox": [ + 90, + 159, + 518, + 190 + ], + "type": "text", + "content": " Sicheng Zhao" + }, + { + "bbox": [ + 90, + 159, + 518, + 190 + ], + "type": "inline_equation", + "content": "^{2‡}" + }, + { + "bbox": [ + 90, + 159, + 518, + 190 + ], + "type": "text", + "content": " Jungong Han" + }, + { + "bbox": [ + 90, + 159, + 518, + 190 + ], + "type": "inline_equation", + "content": "^{3}" + }, + { + "bbox": [ + 90, + 159, + 518, + 190 + ], + "type": "text", + "content": " Guiguang Ding" + }, + { + "bbox": [ + 90, + 159, + 518, + 190 + ], + "type": "inline_equation", + "content": "^{1,2‡}" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 109, + 190, + 501, + 217 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 109, + 190, + 501, + 217 + ], + "spans": [ + { + "bbox": [ + 109, + 190, + 501, + 217 + ], + "type": "inline_equation", + "content": "^{1}" + }, + { + "bbox": [ + 109, + 190, + 501, + 217 + ], + "type": "text", + "content": " School of Software " + }, + { + "bbox": [ + 109, + 190, + 501, + 217 + ], + "type": "inline_equation", + "content": "^{2}" + }, + { + "bbox": [ + 109, + 190, + 501, + 217 + ], + "type": "text", + "content": " BNrist " + }, + { + "bbox": [ + 109, + 190, + 501, + 217 + ], + "type": "inline_equation", + "content": "^{3}" + }, + { + "bbox": [ + 109, + 190, + 501, + 217 + ], + "type": "text", + "content": " Department of Automation, Tsinghua University " + }, + { + "bbox": [ + 109, + 190, + 501, + 217 + ], + "type": "inline_equation", + "content": "^{4}" + }, + { + "bbox": [ + 109, + 190, + 501, + 217 + ], + "type": "text", + "content": " Hangzhou Zhuoxi Institute of Brain and Intelligence" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 86, + 217, + 533, + 232 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 86, + 217, + 533, + 232 + ], + "spans": [ + { + "bbox": [ + 86, + 217, + 533, + 232 + ], + "type": "inline_equation", + "content": "^{5}" + }, + { + "bbox": [ + 86, + 217, + 533, + 232 + ], + "type": "text", + "content": " JD.com " + }, + { + "bbox": [ + 86, + 217, + 533, + 232 + ], + "type": "inline_equation", + "content": "^{6}" + }, + { + "bbox": [ + 86, + 217, + 533, + 232 + ], + "type": "text", + "content": " GRG Banking Equipment Co., Ltd. " + }, + { + "bbox": [ + 86, + 217, + 533, + 232 + ], + "type": "inline_equation", + "content": "^{7}" + }, + { + "bbox": [ + 86, + 217, + 533, + 232 + ], + "type": "text", + "content": " South China University of Technology" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 151, + 258, + 200, + 271 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 151, + 258, + 200, + 271 + ], + "spans": [ + { + "bbox": [ + 151, + 258, + 200, + 271 + ], + "type": "text", + "content": "Abstract" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 54, + 284, + 296, + 571 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 54, + 284, + 296, + 571 + ], + "spans": [ + { + "bbox": [ + 54, + 284, + 296, + 571 + ], + "type": "text", + "content": "The parameter-efficient adaptation of the image-text pretraining model CLIP for video-text retrieval is a prominent area of research. While CLIP is focused on image-level vision-language matching, video-text retrieval demands comprehensive understanding at the video level. Three key discrepancies emerge in the transfer from image-level to video-level: vision, language, and alignment. However, existing methods mainly focus on vision while neglecting language and alignment. In this paper, we propose Discrepancy Reduction in Vision, Language, and Alignment (DiscoVLA), which simultaneously mitigates all three discrepancies. Specifically, we introduce Image-Video Features Fusion to integrate image-level and video-level features, effectively tackling both vision and language discrepancies. Additionally, we generate pseudo image captions to learn fine-grained image-level alignment. To mitigate alignment discrepancies, we propose Image-to-Video Alignment Distillation, which leverages image-level alignment knowledge to enhance video-level alignment. Extensive experiments demonstrate the superiority of our DiscoVLA. In particular, on MSRVTT with CLIP (ViT-B/16), DiscoVLA outperforms previous methods by " + }, + { + "bbox": [ + 54, + 284, + 296, + 571 + ], + "type": "inline_equation", + "content": "2.2\\%" + }, + { + "bbox": [ + 54, + 284, + 296, + 571 + ], + "type": "text", + "content": " R@1 and " + }, + { + "bbox": [ + 54, + 284, + 296, + 571 + ], + "type": "inline_equation", + "content": "7.5\\%" + }, + { + "bbox": [ + 54, + 284, + 296, + 571 + ], + "type": "text", + "content": " R@sum. The code is available at https://github.com/LunarShen/DsicoVLA." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 55, + 594, + 135, + 605 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 594, + 135, + 605 + ], + "spans": [ + { + "bbox": [ + 55, + 594, + 135, + 605 + ], + "type": "text", + "content": "1. Introduction" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 55, + 613, + 295, + 685 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 613, + 295, + 685 + ], + "spans": [ + { + "bbox": [ + 55, + 613, + 295, + 685 + ], + "type": "text", + "content": "The rapid expansion of online video content has generated a growing demand for effective video-text retrieval, focusing on matching video content with relevant textual descriptions. Inspired by the success of image-text pretraining approaches such as CLIP [42], recent efforts [9, 14, 24, 37, 48, 49, 51] have increasingly focused on extend" + } + ] + } + ], + "index": 9 + }, + { + "type": "image", + "bbox": [ + 328, + 257, + 544, + 472 + ], + "blocks": [ + { + "bbox": [ + 328, + 257, + 544, + 472 + ], + "lines": [ + { + "bbox": [ + 328, + 257, + 544, + 472 + ], + "spans": [ + { + "bbox": [ + 328, + 257, + 544, + 472 + ], + "type": "image", + "image_path": "bfbe10255f720db020eda19693dec725fa3660b25e37918a8c85685474458566.jpg" + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 313, + 534, + 555, + 590 + ], + "lines": [ + { + "bbox": [ + 313, + 534, + 555, + 590 + ], + "spans": [ + { + "bbox": [ + 313, + 534, + 555, + 590 + ], + "type": "text", + "content": "Figure 1. When parameter-efficient transferring image-level CLIP to video-text retrieval, we identify three key discrepancies: vision, alignment, and language. Unlike previous methods, our DiscoVLA aims to address all these discrepancies. More visualization examples are provided in Figure 4." + } + ] + } + ], + "index": 12, + "angle": 0, + "type": "image_caption" + } + ], + "index": 10 + }, + { + "type": "table", + "bbox": [ + 341, + 474, + 529, + 525 + ], + "blocks": [ + { + "bbox": [ + 341, + 474, + 529, + 525 + ], + "lines": [ + { + "bbox": [ + 341, + 474, + 529, + 525 + ], + "spans": [ + { + "bbox": [ + 341, + 474, + 529, + 525 + ], + "type": "table", + "html": "
MethodImage-Level PETL Video-Level
VisionLanguageAlignment
Previous Methods✓××
Our DiscoVLA✓✓✓
", + "image_path": "4296e1e353b6b09270b71acaae8eb2dae06abe5ead9805faa297f623ce290e48.jpg" + } + ] + } + ], + "index": 11, + "angle": 0, + "type": "table_body" + } + ], + "index": 11 + }, + { + "bbox": [ + 313, + 615, + 555, + 687 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 615, + 555, + 687 + ], + "spans": [ + { + "bbox": [ + 313, + 615, + 555, + 687 + ], + "type": "text", + "content": "ing these powerful capabilities to video-text retrieval tasks. However, fully fine-tuning these pretrained models requires updating numerous parameters for every dataset, resulting in large storage overhead [22, 58]. Therefore, we focus on parameter-efficient transfer learning methods, targeting strong performance with minimal trainable parameters." + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 313, + 689, + 555, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 689, + 555, + 714 + ], + "spans": [ + { + "bbox": [ + 313, + 689, + 555, + 714 + ], + "type": "text", + "content": "Although both image-text pretraining and video-text retrieval involve vision-language (VL) matching, they exhibit" + } + ] + } + ], + "index": 14 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "spans": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "type": "text", + "content": "CVF" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 145, + 0, + 496, + 37 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 145, + 0, + 496, + 37 + ], + "spans": [ + { + "bbox": [ + 145, + 0, + 496, + 37 + ], + "type": "text", + "content": "This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 66, + 693, + 287, + 703 + ], + "type": "page_footnote", + "angle": 0, + "lines": [ + { + "bbox": [ + 66, + 693, + 287, + 703 + ], + "spans": [ + { + "bbox": [ + 66, + 693, + 287, + 703 + ], + "type": "text", + "content": "* lunarshen@gmail.com. Work done during an internship at JD.com." + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 67, + 703, + 207, + 712 + ], + "type": "page_footnote", + "angle": 0, + "lines": [ + { + "bbox": [ + 67, + 703, + 207, + 712 + ], + "spans": [ + { + "bbox": [ + 67, + 703, + 207, + 712 + ], + "type": "text", + "content": "†Project leader. ‡ Corresponding authors." + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "text", + "content": "19702" + } + ] + } + ], + "index": 18 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 0 + }, + { + "para_blocks": [ + { + "bbox": [ + 55, + 72, + 294, + 179 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 72, + 294, + 179 + ], + "spans": [ + { + "bbox": [ + 55, + 72, + 294, + 179 + ], + "type": "text", + "content": "substantial discrepancies that hinder straightforward knowledge transfer. To be specific, image-text pretraining focuses on matching images with image captions, which represents image-level VL matching. In contrast, video-text retrieval focuses on matching videos with video captions, which represents video-level VL matching. The challenges of video-level VL matching, which demands understanding motion and time, create significant obstacles for transferring knowledge from image-text to video-text tasks." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 55, + 182, + 294, + 361 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 182, + 294, + 361 + ], + "spans": [ + { + "bbox": [ + 55, + 182, + 294, + 361 + ], + "type": "text", + "content": "In this work, we explore the discrepancies between image-level and video-level VL matching across three key aspects: vision, language, and alignment. Videos introduce a temporal dimension, which is absent in images. Similarly, the distinction between image captions and video captions arises from the varying levels of granularity, such as isolated image descriptions versus continuous video narratives. Thus, video-level alignment is inherently more complex than image-level alignment, as it must account for intricate spatio-temporal relationships. In Figure 1, the three images correspond to \"preparing food\", \"pouring\", and \"cutting\", while the entire video demonstrates an instructional cooking process. Image-level VL matching focuses on isolated images, while video-level VL matching emphasizes the overall understanding of continuous segments." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 55, + 363, + 294, + 458 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 363, + 294, + 458 + ], + "spans": [ + { + "bbox": [ + 55, + 363, + 294, + 458 + ], + "type": "text", + "content": "However, current methods [5, 22, 25, 58] struggle to fully address the discrepancies mentioned above. Some [5, 22, 25, 58] focus on the vision aspect, modeling temporal relationships to extract video-level features, while others [25, 58] attempt to reduce the modality gap between vision and language through shared parameter mechanisms. Despite these advances, these methods remain limited in both language and alignment discrepancies." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 55, + 460, + 294, + 533 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 460, + 294, + 533 + ], + "spans": [ + { + "bbox": [ + 55, + 460, + 294, + 533 + ], + "type": "text", + "content": "To tackle these challenges, we introduce Discrepancy Reduction in Vision, Language, and Alignment, termed DiscoVLA, a novel parameter-efficient method for videotext retrieval. Unlike previous methods that target a single gap, our method aims to simultaneously reduce all significant discrepancies, as illustrated in the table of Figure 1." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 55, + 534, + 295, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 534, + 295, + 713 + ], + "spans": [ + { + "bbox": [ + 55, + 534, + 295, + 713 + ], + "type": "text", + "content": "Specifically, we propose an Image-Video Features Fusion module (IVFusion), which addresses both vision and language discrepancies through a unified feature extraction approach. IVFusion effectively combines both image- and video-level features by utilizing a lightweight adapter. In addition to video-level alignment, we propose Pseudo Image-level Alignment (PIImgAlign) which generates pseudo image captions and learns fine-grained image-level alignment. Furthermore, to reduce alignment discrepancy, we propose Image-to-Video Alignment Distillation (AlignDistill), transferring image-level alignment knowledge to improve video-level alignment. All proposed modules prioritize parameter efficiency: IVFusion requires minimal trainable parameters, while PImgAlign and AlignDistill operate without increasing inference parameters." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 326, + 72, + 531, + 83 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 326, + 72, + 531, + 83 + ], + "spans": [ + { + "bbox": [ + 326, + 72, + 531, + 83 + ], + "type": "text", + "content": "In summary, the main contributions are as follows:" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 314, + 84, + 553, + 216 + ], + "type": "list", + "angle": 0, + "index": 9, + "blocks": [ + { + "bbox": [ + 314, + 84, + 553, + 119 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 314, + 84, + 553, + 119 + ], + "spans": [ + { + "bbox": [ + 314, + 84, + 553, + 119 + ], + "type": "text", + "content": "- We reveal, for the first time, the necessity of addressing all three image-to-video discrepancies—vision, language, and alignment—to achieve parameter-efficient VTR." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 314, + 121, + 553, + 168 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 314, + 121, + 553, + 168 + ], + "spans": [ + { + "bbox": [ + 314, + 121, + 553, + 168 + ], + "type": "text", + "content": "- We introduce IVFusion to fuse image- and video-level features for both vision and language gaps. PImgAlign is introduced for fine-grained image-level alignment. We introduce AlignDistill to minimize alignment gaps." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 314, + 168, + 553, + 216 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 314, + 168, + 553, + 216 + ], + "spans": [ + { + "bbox": [ + 314, + 168, + 553, + 216 + ], + "type": "text", + "content": "- Our DiscoVLA achieves state-of-the-art results on MSRVTT, LSMDC, ActivityNet, and DiDeMo. Notably, on MSRVTT with CLIP (ViT-B/16), DiscoVLA reaches " + }, + { + "bbox": [ + 314, + 168, + 553, + 216 + ], + "type": "inline_equation", + "content": "50.5\\%" + }, + { + "bbox": [ + 314, + 168, + 553, + 216 + ], + "type": "text", + "content": " R@1, surpassing previous methods by " + }, + { + "bbox": [ + 314, + 168, + 553, + 216 + ], + "type": "inline_equation", + "content": "2.2\\%" + }, + { + "bbox": [ + 314, + 168, + 553, + 216 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 8 + } + ], + "sub_type": "text" + }, + { + "bbox": [ + 314, + 227, + 400, + 239 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 314, + 227, + 400, + 239 + ], + "spans": [ + { + "bbox": [ + 314, + 227, + 400, + 239 + ], + "type": "text", + "content": "2. Related Work" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 313, + 247, + 553, + 449 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 247, + 553, + 449 + ], + "spans": [ + { + "bbox": [ + 313, + 247, + 553, + 449 + ], + "type": "text", + "content": "Video-Text Retrieval. Advances in vision-language pretraining [7, 12, 32-34, 42, 50, 52, 55] have achieved significant success across various downstream tasks [2, 10, 11, 18, 26, 45, 47, 54, 56, 57]. Leveraging powerful CLIP [42], recent video-text retrieval (VTR) methods [9, 14, 24, 37, 48, 49, 51] have achieved impressive performance enhancements. CLIP4Clip [37] is a pioneering CLIP-based method that aggregates image features via mean-pooling or transformers to obtain video features, inspiring subsequent studies. Additionally, Cap4Video [51] utilizes ZeroCap [46] and GPT-2 [41] to generate video-level captions as supplementary video content during both training and inference. In contrast, our DiscoVLA addresses the image-level and video-level gap by generating image-level captions used only in training to improve video-level alignment. These full fine-tuning methods impose substantial overhead from large trainable parameters." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 313, + 450, + 553, + 582 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 450, + 553, + 582 + ], + "spans": [ + { + "bbox": [ + 313, + 450, + 553, + 582 + ], + "type": "text", + "content": "Parameter-Efficient Transfer Learning. Parameter-efficient methods aim to adapt pre-trained models to new tasks with minimal fine-tuning. The mainstream methods in natural language processing and image tasks include techniques such as Prompt [23, 27, 28, 61, 62], Adapter [6, 17, 20, 60], and LoRA [21, 60]. Prompt [27] modifies the input space by adding learnable tokens. Adapter [20] introduces a lightweight bottleneck neural network consisting of up- and down-projection layers with a non-linear activation function. LoRA [21] aims to update up- and down-projection layers by learning low-rank decomposition matrices." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 313, + 582, + 553, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 582, + 553, + 713 + ], + "spans": [ + { + "bbox": [ + 313, + 582, + 553, + 713 + ], + "type": "text", + "content": "Recently, several studies have explored parameter-efficient VTR methods [5, 22, 25, 58] to incorporate temporal modeling into the frozen image-level CLIP backbone. VoP [22] employs trainable BiLSTMs [16] to produce prompt tokens in the vision encoder. DGL [58] introduces weight-sharing layers to generate prompts for both text and video encoders. MV-Adapter [25] introduces shared bottleneck structures in both video and text branches. However, these methods overlook language and alignment discrepancies. We address these limitations with DiscoVLA, a framework that facilitates robust video-level understanding." + } + ] + } + ], + "index": 13 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "text", + "content": "19703" + } + ] + } + ], + "index": 14 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 1 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 59, + 89, + 553, + 266 + ], + "blocks": [ + { + "bbox": [ + 59, + 89, + 553, + 266 + ], + "lines": [ + { + "bbox": [ + 59, + 89, + 553, + 266 + ], + "spans": [ + { + "bbox": [ + 59, + 89, + 553, + 266 + ], + "type": "image", + "image_path": "3baf7371eccbd1f136ffa4c85bcbf4d194ecc701dc4f2f62831d41c878f71323.jpg" + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 55, + 270, + 555, + 326 + ], + "lines": [ + { + "bbox": [ + 55, + 270, + 555, + 326 + ], + "spans": [ + { + "bbox": [ + 55, + 270, + 555, + 326 + ], + "type": "text", + "content": "Figure 2. The overall framework of DiscoVLA. Initially, we generate pseudo image captions for each sampled image. In both vision and text encoders, we utilize image-level layers to acquire pretrained image-level knowledge and employ IVFusion layers to enhance spatiotemporal information. The single video caption is encoded through the text encoder, utilizing IVFusion layer as image-level (image-lvl) layer. Finally, AlignDistill is applied to distill image-level alignment to video-level alignment. For fair comparisons with previous methods, we do not generate pseudo image captions during the inference phase." + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_caption" + } + ], + "index": 1 + }, + { + "bbox": [ + 55, + 345, + 138, + 358 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 345, + 138, + 358 + ], + "spans": [ + { + "bbox": [ + 55, + 345, + 138, + 358 + ], + "type": "text", + "content": "3. Methodology" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 55, + 365, + 135, + 379 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 365, + 135, + 379 + ], + "spans": [ + { + "bbox": [ + 55, + 365, + 135, + 379 + ], + "type": "text", + "content": "3.1. Preliminary" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 55, + 385, + 296, + 445 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 385, + 296, + 445 + ], + "spans": [ + { + "bbox": [ + 55, + 385, + 296, + 445 + ], + "type": "text", + "content": "Video-text retrieval is a task that involves retrieving relevant videos based on a given text query, or conversely, retrieving appropriate text descriptions for a video. Consistent with previous approaches [5, 22, 25, 58], we utilize the CLIP model, pre-trained on image-text data, as our backbone." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 55, + 447, + 296, + 567 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 447, + 296, + 567 + ], + "spans": [ + { + "bbox": [ + 55, + 447, + 296, + 567 + ], + "type": "text", + "content": "To process a given video, we first sample a sequence of frame images uniformly, which are then input into the CLIP vision encoder to extract image features " + }, + { + "bbox": [ + 55, + 447, + 296, + 567 + ], + "type": "inline_equation", + "content": "v^{\\mathrm{img}}" + }, + { + "bbox": [ + 55, + 447, + 296, + 567 + ], + "type": "text", + "content": ". These features are then aggregated through temporal average pooling to get video features " + }, + { + "bbox": [ + 55, + 447, + 296, + 567 + ], + "type": "inline_equation", + "content": "v^{\\mathrm{vid}}" + }, + { + "bbox": [ + 55, + 447, + 296, + 567 + ], + "type": "text", + "content": ". For a given video caption, it is passed through the CLIP text encoder to extract video caption features " + }, + { + "bbox": [ + 55, + 447, + 296, + 567 + ], + "type": "inline_equation", + "content": "t^{\\mathrm{vid}}" + }, + { + "bbox": [ + 55, + 447, + 296, + 567 + ], + "type": "text", + "content": ". Then, we compute the cross-modal contrastive loss [40], which functions as a video-level alignment term, maximizing the video-level similarity " + }, + { + "bbox": [ + 55, + 447, + 296, + 567 + ], + "type": "inline_equation", + "content": "\\mathrm{Sim}^{\\mathrm{vid}}" + }, + { + "bbox": [ + 55, + 447, + 296, + 567 + ], + "type": "text", + "content": " between matched pairs:" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 133, + 578, + 294, + 596 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 578, + 294, + 596 + ], + "spans": [ + { + "bbox": [ + 133, + 578, + 294, + 596 + ], + "type": "interline_equation", + "content": "\\operatorname {S i m} _ {i j} ^ {\\mathrm {v i d}} = t _ {i} ^ {\\mathrm {v i d} ^ {T}} v _ {j} ^ {\\mathrm {v i d}}, \\tag {1}", + "image_path": "136c483010ac3f2dc5ce7ebfbac8dbb9576247027c3ccbb0833fab75dab9b72f.jpg" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 65, + 597, + 295, + 631 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 65, + 597, + 295, + 631 + ], + "spans": [ + { + "bbox": [ + 65, + 597, + 295, + 631 + ], + "type": "interline_equation", + "content": "\\mathcal {L} _ {t 2 v} \\left(\\operatorname {S i m} ^ {\\text {v i d}}\\right) = \\frac {1}{B} \\sum_ {i = 1} ^ {B} \\log \\frac {\\exp \\left(\\operatorname {S i m} _ {i i} ^ {\\text {v i d}}\\right) / \\tau)}{\\sum_ {j = 1} ^ {B} \\exp \\left(\\operatorname {S i m} _ {i j} ^ {\\text {v i d}}\\right) / \\tau)}, \\tag {2}", + "image_path": "c66f87315dbbfe5ce5b96aaaa3b5b940bd080318c3687b15272bb0eec3c30186.jpg" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 65, + 633, + 294, + 666 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 65, + 633, + 294, + 666 + ], + "spans": [ + { + "bbox": [ + 65, + 633, + 294, + 666 + ], + "type": "interline_equation", + "content": "\\mathcal {L} _ {v 2 t} \\left(\\operatorname {S i m} ^ {\\text {v i d}}\\right) = \\frac {1}{B} \\sum_ {i = 1} ^ {B} \\log \\frac {\\exp \\left(\\operatorname {S i m} _ {i i} ^ {\\text {v i d}}\\right) / \\tau)}{\\sum_ {j = 1} ^ {B} \\exp \\left(\\operatorname {S i m} _ {j i} ^ {\\text {v i d}}\\right) / \\tau)}, \\tag {3}", + "image_path": "d0b814dd629056be75bc2a683c138c74949c11353b3602d9291b90b05ec7037a.jpg" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 65, + 667, + 294, + 691 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 65, + 667, + 294, + 691 + ], + "spans": [ + { + "bbox": [ + 65, + 667, + 294, + 691 + ], + "type": "interline_equation", + "content": "\\mathcal {L} _ {A} (\\operatorname {S i m} ^ {\\mathrm {v i d}}) = - \\frac {1}{2} \\left[ \\mathcal {L} _ {t 2 v} (\\operatorname {S i m} ^ {\\mathrm {v i d}}) + \\mathcal {L} _ {v 2 t} (\\operatorname {S i m} ^ {\\mathrm {v i d}}) \\right], \\tag {4}", + "image_path": "4defd300227dd64aaa6ff5e7d8cf31921ab78fcef85b7c351c4ad9bb8b305e32.jpg" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 55, + 701, + 288, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 701, + 288, + 713 + ], + "spans": [ + { + "bbox": [ + 55, + 701, + 288, + 713 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 55, + 701, + 288, + 713 + ], + "type": "inline_equation", + "content": "B" + }, + { + "bbox": [ + 55, + 701, + 288, + 713 + ], + "type": "text", + "content": " is the batch size, " + }, + { + "bbox": [ + 55, + 701, + 288, + 713 + ], + "type": "inline_equation", + "content": "\\tau" + }, + { + "bbox": [ + 55, + 701, + 288, + 713 + ], + "type": "text", + "content": " is the temperature parameter." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 314, + 346, + 385, + 357 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 314, + 346, + 385, + 357 + ], + "spans": [ + { + "bbox": [ + 314, + 346, + 385, + 357 + ], + "type": "text", + "content": "3.2. DiscoVLA" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 313, + 363, + 555, + 495 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 363, + 555, + 495 + ], + "spans": [ + { + "bbox": [ + 313, + 363, + 555, + 495 + ], + "type": "text", + "content": "As shown in Figure 2, we propose DiscoVLA to address the challenges of vision, language, and alignment when transferring from image-level to video-level VL matching. First, we introduce Image-Video Features Fusion (IVFusion) in both vision and text encoders to enhance feature extraction. Next, we introduce Pseudo Image-Level Alignment (PIImgAlign) for fine-grained image-level alignment, consisting of Pseudo Image Language Generation and Image-Level Alignment. Finally, Image-to-Video Alignment Distillation (AlignDistill) is proposed to leverage image-level alignment knowledge to enhance video-level alignment." + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 313, + 495, + 554, + 555 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 495, + 554, + 555 + ], + "spans": [ + { + "bbox": [ + 313, + 495, + 554, + 555 + ], + "type": "text", + "content": "Importantly, PImgAlign is only employed during training. For a fair comparison with previous methods and due to its generation inefficiency, PImgAlign is omitted at inference. The final similarity used for retrieval is based on video-level similarity in Eq. (1)." + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 313, + 555, + 554, + 628 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 555, + 554, + 628 + ], + "spans": [ + { + "bbox": [ + 313, + 555, + 554, + 628 + ], + "type": "text", + "content": "We focus on parameter-efficient transfer learning for video-text retrieval, where the large backbone parameters remain fixed, and only a small number of additional parameters are trained. The trainable parameters are embedded in the parameter-efficient attention of each encoder layer, as illustrated in Figure 3." + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 313, + 635, + 474, + 648 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 635, + 474, + 648 + ], + "spans": [ + { + "bbox": [ + 313, + 635, + 474, + 648 + ], + "type": "text", + "content": "3.3. Image-Video Features Fusion" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 313, + 653, + 554, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 653, + 554, + 713 + ], + "spans": [ + { + "bbox": [ + 313, + 653, + 554, + 713 + ], + "type": "text", + "content": "Image-level and video-level variations exist in both vision and language modalities. Our proposed IVFusion fuses these image-level and video-level features in both the vision and language encoders. Below, we illustrate IVFusion's application within the vision encoder." + } + ] + } + ], + "index": 17 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 59, + 74, + 422, + 85 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 59, + 74, + 422, + 85 + ], + "spans": [ + { + "bbox": [ + 59, + 74, + 422, + 85 + ], + "type": "text", + "content": "CLSTokenofVisionEncoder EOSTokenofTextEncoder -- " + }, + { + "bbox": [ + 59, + 74, + 422, + 85 + ], + "type": "inline_equation", + "content": "\\rightarrow" + }, + { + "bbox": [ + 59, + 74, + 422, + 85 + ], + "type": "text", + "content": " UtilizedONLYduringthe training phase." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "text", + "content": "19704" + } + ] + } + ], + "index": 18 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 2 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 65, + 70, + 304, + 270 + ], + "blocks": [ + { + "bbox": [ + 65, + 70, + 304, + 270 + ], + "lines": [ + { + "bbox": [ + 65, + 70, + 304, + 270 + ], + "spans": [ + { + "bbox": [ + 65, + 70, + 304, + 270 + ], + "type": "image", + "image_path": "0704c9a8a0461be70b62fb28713d053f6bc102bbe788c05808941ec59c7cb103.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 54, + 270, + 555, + 316 + ], + "lines": [ + { + "bbox": [ + 54, + 270, + 555, + 316 + ], + "spans": [ + { + "bbox": [ + 54, + 270, + 555, + 316 + ], + "type": "text", + "content": "Figure 3. Illustration of the encoder layers for vision and text encoders. (a) Image-Level Attn operates on each of the " + }, + { + "bbox": [ + 54, + 270, + 555, + 316 + ], + "type": "inline_equation", + "content": "F" + }, + { + "bbox": [ + 54, + 270, + 555, + 316 + ], + "type": "text", + "content": " images or image captions individually. (b) Video-Level Attn concatenates tokens across all sampled images or image captions. (c) IVFusion Attn employs a lightweight adapter to integrate the efficiency of Image-Level Attn with the effectiveness of Video-Level Attn. Here, we illustrate the application of IVFusion within the vision encoder. The text encoder adopts the same approach for processing pseudo image captions." + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "type": "image", + "bbox": [ + 309, + 71, + 542, + 269 + ], + "blocks": [ + { + "bbox": [ + 309, + 71, + 542, + 269 + ], + "lines": [ + { + "bbox": [ + 309, + 71, + 542, + 269 + ], + "spans": [ + { + "bbox": [ + 309, + 71, + 542, + 269 + ], + "type": "image", + "image_path": "46a2b4606872b486ec8096af7f0399e265e85e9f2ff67c904082ded61a9105d2.jpg" + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_body" + } + ], + "index": 1 + }, + { + "bbox": [ + 54, + 335, + 294, + 443 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 54, + 335, + 294, + 443 + ], + "spans": [ + { + "bbox": [ + 54, + 335, + 294, + 443 + ], + "type": "text", + "content": "Image-Level Attn. To achieve parameter-efficient transfer learning, we apply LoRA in the attention module of each encoder layer. As illustrated in Figure 3a, " + }, + { + "bbox": [ + 54, + 335, + 294, + 443 + ], + "type": "inline_equation", + "content": "F \\times (N + 1)" + }, + { + "bbox": [ + 54, + 335, + 294, + 443 + ], + "type": "text", + "content": " tokens are processed via Image-level Attn, where " + }, + { + "bbox": [ + 54, + 335, + 294, + 443 + ], + "type": "inline_equation", + "content": "F" + }, + { + "bbox": [ + 54, + 335, + 294, + 443 + ], + "type": "text", + "content": " denotes the number of sampled images, and " + }, + { + "bbox": [ + 54, + 335, + 294, + 443 + ], + "type": "inline_equation", + "content": "N + 1" + }, + { + "bbox": [ + 54, + 335, + 294, + 443 + ], + "type": "text", + "content": " represents the combination of the global CLS token and patch tokens. Image-Level Attn processes each of the " + }, + { + "bbox": [ + 54, + 335, + 294, + 443 + ], + "type": "inline_equation", + "content": "F" + }, + { + "bbox": [ + 54, + 335, + 294, + 443 + ], + "type": "text", + "content": " images individually, using " + }, + { + "bbox": [ + 54, + 335, + 294, + 443 + ], + "type": "inline_equation", + "content": "N + 1" + }, + { + "bbox": [ + 54, + 335, + 294, + 443 + ], + "type": "text", + "content": " tokens per image. However, Image-Level Attn does not encode temporal information." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 54, + 443, + 295, + 552 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 54, + 443, + 295, + 552 + ], + "spans": [ + { + "bbox": [ + 54, + 443, + 295, + 552 + ], + "type": "text", + "content": "Video-Level Attn. A straightforward solution is Video-Level Attn, which concatenates tokens across all sampled images, enabling the attention module to learn spatiotemporal information. As shown in Figure 3b, " + }, + { + "bbox": [ + 54, + 443, + 295, + 552 + ], + "type": "inline_equation", + "content": "F(N + 1)" + }, + { + "bbox": [ + 54, + 443, + 295, + 552 + ], + "type": "text", + "content": " tokens are simultaneously processed during the attention module. Due to the quadratic complexity of the attention mechanism, Image-Level Attn has a complexity of " + }, + { + "bbox": [ + 54, + 443, + 295, + 552 + ], + "type": "inline_equation", + "content": "O(F(N)^2)" + }, + { + "bbox": [ + 54, + 443, + 295, + 552 + ], + "type": "text", + "content": ", while Video-Level Attn has a higher complexity of " + }, + { + "bbox": [ + 54, + 443, + 295, + 552 + ], + "type": "inline_equation", + "content": "O((FN)^2)" + }, + { + "bbox": [ + 54, + 443, + 295, + 552 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 55, + 552, + 295, + 624 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 552, + 295, + 624 + ], + "spans": [ + { + "bbox": [ + 55, + 552, + 295, + 624 + ], + "type": "text", + "content": "IVFusion Attn. To combine Image-Level Attn's efficiency with Video-Level Attn's effectiveness, we introduce IV-Fusion Attn in Figure 3c. We extend Image-Level Attn by incorporating a branch that uses all CLS tokens to extract spatio-temporal information, reducing complexity to " + }, + { + "bbox": [ + 55, + 552, + 295, + 624 + ], + "type": "inline_equation", + "content": "O(F(N)^2 + (F)^2 N)" + }, + { + "bbox": [ + 55, + 552, + 295, + 624 + ], + "type": "text", + "content": ":" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 91, + 630, + 295, + 647 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 91, + 630, + 295, + 647 + ], + "spans": [ + { + "bbox": [ + 91, + 630, + 295, + 647 + ], + "type": "interline_equation", + "content": "\\left[ c _ {i} ^ {(1)}, \\hat {E} _ {i} \\right] = \\mathrm {A T T N} \\left(Q K V = \\left[ c _ {i} ^ {(0)}, E _ {i} ^ {(0)} \\right]\\right), \\tag {5}", + "image_path": "cb287acdb7ca3999e32e24a6f985ba9eaf060837b2151cd4a6b3449fbf1dff0d.jpg" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 69, + 648, + 294, + 665 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 69, + 648, + 294, + 665 + ], + "spans": [ + { + "bbox": [ + 69, + 648, + 294, + 665 + ], + "type": "interline_equation", + "content": "\\left[ c _ {1: F} ^ {(2)} \\right] = \\operatorname {A T T N} \\left(Q = \\left[ c _ {1: F} ^ {(0)} \\right], K V = \\left[ c _ {1: F} ^ {(0)}, E _ {1: F} ^ {(0)} \\right]\\right), \\tag {6}", + "image_path": "75eea03f3cf455dacc9e6ef4b735d66cda5cc8c2cf2afb212bbc4d60097d7721.jpg" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 109, + 666, + 294, + 683 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 109, + 666, + 294, + 683 + ], + "spans": [ + { + "bbox": [ + 109, + 666, + 294, + 683 + ], + "type": "interline_equation", + "content": "\\hat {c} _ {i} = c _ {i} ^ {(2)} + \\sigma (c _ {i} ^ {(1)} \\mathbf {W} _ {\\mathrm {d o w n}}) \\mathbf {W} _ {\\mathrm {u p}}, \\tag {7}", + "image_path": "cfc9b4e8357c6414a2b1e66068848758e70c8000050ab5c2c9b972382e819b34.jpg" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 55, + 689, + 295, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 689, + 295, + 714 + ], + "spans": [ + { + "bbox": [ + 55, + 689, + 295, + 714 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 55, + 689, + 295, + 714 + ], + "type": "inline_equation", + "content": "c_{i}\\in \\mathbb{R}^{1\\times D}" + }, + { + "bbox": [ + 55, + 689, + 295, + 714 + ], + "type": "text", + "content": " is CLS token of the " + }, + { + "bbox": [ + 55, + 689, + 295, + 714 + ], + "type": "inline_equation", + "content": "i^{\\mathrm{th}}" + }, + { + "bbox": [ + 55, + 689, + 295, + 714 + ], + "type": "text", + "content": " image and " + }, + { + "bbox": [ + 55, + 689, + 295, + 714 + ], + "type": "inline_equation", + "content": "c_{1:F}\\in \\mathbb{R}^{F\\times D}" + }, + { + "bbox": [ + 55, + 689, + 295, + 714 + ], + "type": "text", + "content": " are CLS tokens of all sampled images." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 313, + 335, + 553, + 458 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 335, + 553, + 458 + ], + "spans": [ + { + "bbox": [ + 313, + 335, + 553, + 458 + ], + "type": "text", + "content": "Similarly, " + }, + { + "bbox": [ + 313, + 335, + 553, + 458 + ], + "type": "inline_equation", + "content": "E_{i} \\in \\mathbb{R}^{N \\times D}" + }, + { + "bbox": [ + 313, + 335, + 553, + 458 + ], + "type": "text", + "content": " are patch tokens of the " + }, + { + "bbox": [ + 313, + 335, + 553, + 458 + ], + "type": "inline_equation", + "content": "i^{\\text{th}}" + }, + { + "bbox": [ + 313, + 335, + 553, + 458 + ], + "type": "text", + "content": " image and " + }, + { + "bbox": [ + 313, + 335, + 553, + 458 + ], + "type": "inline_equation", + "content": "E_{1:F} \\in \\mathbb{R}^{FN \\times D}" + }, + { + "bbox": [ + 313, + 335, + 553, + 458 + ], + "type": "text", + "content": " are patch tokens of all sampled images. In Eq. (7), we employ a trainable adapter to merge image-level features " + }, + { + "bbox": [ + 313, + 335, + 553, + 458 + ], + "type": "inline_equation", + "content": "c_{i}^{(1)}" + }, + { + "bbox": [ + 313, + 335, + 553, + 458 + ], + "type": "text", + "content": " and video-level features " + }, + { + "bbox": [ + 313, + 335, + 553, + 458 + ], + "type": "inline_equation", + "content": "c_{i}^{(2)}" + }, + { + "bbox": [ + 313, + 335, + 553, + 458 + ], + "type": "text", + "content": ", where " + }, + { + "bbox": [ + 313, + 335, + 553, + 458 + ], + "type": "inline_equation", + "content": "\\mathbf{W}_{down} \\in \\mathbb{R}^{D \\times r}" + }, + { + "bbox": [ + 313, + 335, + 553, + 458 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 313, + 335, + 553, + 458 + ], + "type": "inline_equation", + "content": "\\mathbf{W}_{up} \\in \\mathbb{R}^{r \\times D}" + }, + { + "bbox": [ + 313, + 335, + 553, + 458 + ], + "type": "text", + "content": " represent the up- and down-projection layers, and " + }, + { + "bbox": [ + 313, + 335, + 553, + 458 + ], + "type": "inline_equation", + "content": "\\sigma" + }, + { + "bbox": [ + 313, + 335, + 553, + 458 + ], + "type": "text", + "content": " denotes the nonlinear GELU activation [19]. Image-level features preserve pretrained spatial details, while video-level features capture temporal information. This fusion mechanism strengthens feature extraction." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 313, + 460, + 554, + 552 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 460, + 554, + 552 + ], + "spans": [ + { + "bbox": [ + 313, + 460, + 554, + 552 + ], + "type": "text", + "content": "Our parameter-efficient vision encoder comprises both image-level and IVFusion layers. The shallow image-level layers employ Image-Level Attn to capture spatial information, while the top " + }, + { + "bbox": [ + 313, + 460, + 554, + 552 + ], + "type": "inline_equation", + "content": "H_{v}" + }, + { + "bbox": [ + 313, + 460, + 554, + 552 + ], + "type": "text", + "content": " IVFusion layers leverage IVFusion Attn to focus on spatio-temporal information. Consequently, our vision encoder outputs enhanced image features " + }, + { + "bbox": [ + 313, + 460, + 554, + 552 + ], + "type": "inline_equation", + "content": "\\left\\{ v_{i}^{\\mathrm{img}} \\mid 1 \\leq i \\leq F \\right\\}" + }, + { + "bbox": [ + 313, + 460, + 554, + 552 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 313, + 563, + 484, + 575 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 563, + 484, + 575 + ], + "spans": [ + { + "bbox": [ + 313, + 563, + 484, + 575 + ], + "type": "text", + "content": "3.4. Pseudo Image-Level Alignment" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 313, + 582, + 555, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 582, + 555, + 714 + ], + "spans": [ + { + "bbox": [ + 313, + 582, + 555, + 714 + ], + "type": "text", + "content": "Unlike previous work focused on video-level alignment, we introduce PImgAlign to learn fine-grained image-level alignment via the decomposition of video content at the image level. To accomplish this, we first generate pseudo image captions in VTR datasets. In contrast to the straightforward sampling of images from a video, sampling image captions is impractical, as video captions generally lack the specific details needed for individual images. As a solution, we utilize the multimodal large language model (MLLM) LLaVA-NeXT [35] to generate pseudo image captions. For each sampled image from a video, we prompt the model" + } + ] + } + ], + "index": 13 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "text", + "content": "19705" + } + ] + } + ], + "index": 14 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 3 + }, + { + "para_blocks": [ + { + "bbox": [ + 55, + 72, + 241, + 84 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 72, + 241, + 84 + ], + "spans": [ + { + "bbox": [ + 55, + 72, + 241, + 84 + ], + "type": "text", + "content": "using the relevant video caption for guidance:" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 74, + 92, + 276, + 140 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 74, + 92, + 276, + 140 + ], + "spans": [ + { + "bbox": [ + 74, + 92, + 276, + 140 + ], + "type": "text", + "content": "The provided image is a frame sampled from the video, which describes {video caption}. Based on the video's content, provide a caption for the provided image." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 55, + 148, + 295, + 225 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 148, + 295, + 225 + ], + "spans": [ + { + "bbox": [ + 55, + 148, + 295, + 225 + ], + "type": "text", + "content": "Unlike EA-VTR [38], which uses an image captioner, our method employs a strong MLLM. By incorporating video caption guidance, it generates higher-quality image captions that align with the overall video content. Similarly to visual feature extraction, we adopt our proposed IVFusion to extract pseudo image caption features " + }, + { + "bbox": [ + 55, + 148, + 295, + 225 + ], + "type": "inline_equation", + "content": "\\left\\{t_i^{\\mathrm{img}}|1\\leq i\\leq F\\right\\}" + }, + { + "bbox": [ + 55, + 148, + 295, + 225 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 55, + 226, + 296, + 357 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 226, + 296, + 357 + ], + "spans": [ + { + "bbox": [ + 55, + 226, + 296, + 357 + ], + "type": "text", + "content": "Given image features and pseudo image caption features from a paired video and video caption, we introduce a fine-grained image-level similarity. Since image captions are typically more concise than the images they describe, a single caption may correspond to multiple images, as shown in Figure 4d. Therefore, we calculate the maximum similarity for each image across all image captions and vice versa. Specifically, we computed a similarity matrix between image and image caption features, selecting the maximum value from each row and column to obtain the final image-level similarity:" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 93, + 365, + 295, + 399 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 93, + 365, + 295, + 399 + ], + "spans": [ + { + "bbox": [ + 93, + 365, + 295, + 399 + ], + "type": "interline_equation", + "content": "\\operatorname {S i m} _ {i j - t 2 v} ^ {\\mathrm {i m g}} = \\frac {1}{F} \\sum_ {n = 1} ^ {F} \\max _ {1 \\leq m \\leq F} t _ {i | n} ^ {\\mathrm {i m g} ^ {T}} v _ {j | m} ^ {\\mathrm {i m g}}, \\tag {8}", + "image_path": "1119e4983d8555c735ade650b01fea02aa32697c52b92c176b71a0f405d23b8c.jpg" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 94, + 401, + 295, + 433 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 94, + 401, + 295, + 433 + ], + "spans": [ + { + "bbox": [ + 94, + 401, + 295, + 433 + ], + "type": "interline_equation", + "content": "\\operatorname {S i m} _ {i j - v 2 t} ^ {\\mathrm {i m g}} = \\frac {1}{F} \\sum_ {m = 1} ^ {F} \\max _ {1 \\leq n \\leq F} t _ {i | n} ^ {\\mathrm {i m g} ^ {T}} v _ {j | m} ^ {\\mathrm {i m g}}, \\tag {9}", + "image_path": "9effc40f13ad21b19e2903569cc775e1ea2436bb6dc00d22e6996774bda83353.jpg" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 97, + 435, + 294, + 458 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 97, + 435, + 294, + 458 + ], + "spans": [ + { + "bbox": [ + 97, + 435, + 294, + 458 + ], + "type": "interline_equation", + "content": "\\operatorname {S i m} _ {i j} ^ {\\mathrm {i m g}} = \\frac {1}{2} \\left[ \\operatorname {S i m} _ {i j - t 2 v} ^ {\\mathrm {i m g}} + \\operatorname {S i m} _ {i j - v 2 t} ^ {\\mathrm {i m g}} \\right], \\tag {10}", + "image_path": "611b5a622a131a51a47cfc201f5925c97b05e1e6e73f9968efbe3b50aa4b6d60.jpg" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 55, + 467, + 295, + 544 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 467, + 295, + 544 + ], + "spans": [ + { + "bbox": [ + 55, + 467, + 295, + 544 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 55, + 467, + 295, + 544 + ], + "type": "inline_equation", + "content": "\\mathrm{Sim}_{ij}^{\\mathrm{img}}" + }, + { + "bbox": [ + 55, + 467, + 295, + 544 + ], + "type": "text", + "content": " denotes the image-level similarity between the " + }, + { + "bbox": [ + 55, + 467, + 295, + 544 + ], + "type": "inline_equation", + "content": "i^{\\mathrm{th}}" + }, + { + "bbox": [ + 55, + 467, + 295, + 544 + ], + "type": "text", + "content": " video caption and the " + }, + { + "bbox": [ + 55, + 467, + 295, + 544 + ], + "type": "inline_equation", + "content": "j^{\\mathrm{th}}" + }, + { + "bbox": [ + 55, + 467, + 295, + 544 + ], + "type": "text", + "content": " video. " + }, + { + "bbox": [ + 55, + 467, + 295, + 544 + ], + "type": "inline_equation", + "content": "t_{i|n}^{\\mathrm{img}}" + }, + { + "bbox": [ + 55, + 467, + 295, + 544 + ], + "type": "text", + "content": " denotes the " + }, + { + "bbox": [ + 55, + 467, + 295, + 544 + ], + "type": "inline_equation", + "content": "n^{\\mathrm{th}}" + }, + { + "bbox": [ + 55, + 467, + 295, + 544 + ], + "type": "text", + "content": " pseudo image caption features of the " + }, + { + "bbox": [ + 55, + 467, + 295, + 544 + ], + "type": "inline_equation", + "content": "i^{\\mathrm{th}}" + }, + { + "bbox": [ + 55, + 467, + 295, + 544 + ], + "type": "text", + "content": " video caption and " + }, + { + "bbox": [ + 55, + 467, + 295, + 544 + ], + "type": "inline_equation", + "content": "v_{j|m}^{\\mathrm{img}}" + }, + { + "bbox": [ + 55, + 467, + 295, + 544 + ], + "type": "text", + "content": " denotes the " + }, + { + "bbox": [ + 55, + 467, + 295, + 544 + ], + "type": "inline_equation", + "content": "m^{\\mathrm{th}}" + }, + { + "bbox": [ + 55, + 467, + 295, + 544 + ], + "type": "text", + "content": " image features of the " + }, + { + "bbox": [ + 55, + 467, + 295, + 544 + ], + "type": "inline_equation", + "content": "j^{\\mathrm{th}}" + }, + { + "bbox": [ + 55, + 467, + 295, + 544 + ], + "type": "text", + "content": " video. PImgAlign achieves video-level alignment from the image level, effectively emphasizing spatial details." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 55, + 544, + 296, + 605 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 544, + 296, + 605 + ], + "spans": [ + { + "bbox": [ + 55, + 544, + 296, + 605 + ], + "type": "text", + "content": "However, in the inference stage, the lack of ground truth for video captions prevents pseudo image captions generation. Additionally, MLLM introduces a computational overhead. For fair comparisons with other methods, we exclude PImgAlign during inference." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 55, + 612, + 257, + 624 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 612, + 257, + 624 + ], + "spans": [ + { + "bbox": [ + 55, + 612, + 257, + 624 + ], + "type": "text", + "content": "3.5. Image-to-Video Alignment Distillation" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 55, + 629, + 296, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 629, + 296, + 713 + ], + "spans": [ + { + "bbox": [ + 55, + 629, + 296, + 713 + ], + "type": "text", + "content": "Pretrained CLIP demonstrates a strong capability for image-level alignment. In contrast, video-text retrieval relies on comprehensive spatio-temporal alignment. Therefore, we propose AlignDistill to distill image-level alignment knowledge into video-level alignment, thus mitigating alignment discrepancies. Then, we optimize the Kullback-Leibler divergence [30] between image-level similarity and" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 314, + 72, + 406, + 84 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 314, + 72, + 406, + 84 + ], + "spans": [ + { + "bbox": [ + 314, + 72, + 406, + 84 + ], + "type": "text", + "content": "video-level similarity:" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 322, + 93, + 553, + 109 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 322, + 93, + 553, + 109 + ], + "spans": [ + { + "bbox": [ + 322, + 93, + 553, + 109 + ], + "type": "interline_equation", + "content": "\\mathcal {S} _ {t 2 v} ^ {\\mathrm {i m g}} = \\left[ \\mathrm {s} _ {i 1} ^ {\\mathrm {i m g}}, \\dots , \\mathrm {s} _ {i N _ {v}} ^ {\\mathrm {i m g}} \\right], \\mathcal {S} _ {v 2 t} ^ {\\mathrm {i m g}} = \\left[ \\mathrm {s} _ {1 i} ^ {\\mathrm {i m g}}, \\dots , \\mathrm {s} _ {N _ {t} i} ^ {\\mathrm {i m g}} \\right], \\tag {11}", + "image_path": "eacaba637c12f3bde36f1be23958430c84e1fdf88e74af7b0d12b1e0350716e2.jpg" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 328, + 110, + 553, + 126 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 328, + 110, + 553, + 126 + ], + "spans": [ + { + "bbox": [ + 328, + 110, + 553, + 126 + ], + "type": "interline_equation", + "content": "\\mathcal {S} _ {t 2 v} ^ {\\mathrm {v i d}} = \\left[ \\mathrm {s} _ {i 1} ^ {\\mathrm {v i d}}, \\dots , \\mathrm {s} _ {i N _ {v}} ^ {\\mathrm {v i d}} \\right], \\mathcal {S} _ {v 2 t} ^ {\\mathrm {v i d}} = \\left[ \\mathrm {s} _ {1 i} ^ {\\mathrm {v i d}}, \\dots , \\mathrm {s} _ {N _ {t i}} ^ {\\mathrm {v i d}} \\right], \\tag {12}", + "image_path": "7b034588ddbd5e1f0795c4e7cc0741dd39d8c7f767c87d1aa2f6d97550aa3e2f.jpg" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 321, + 128, + 553, + 150 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 321, + 128, + 553, + 150 + ], + "spans": [ + { + "bbox": [ + 321, + 128, + 553, + 150 + ], + "type": "interline_equation", + "content": "\\mathcal {L} _ {K L} = \\frac {1}{2} \\left(\\mathbf {K L} \\left(\\mathcal {S} _ {t 2 v} ^ {\\mathrm {i m g}} \\| \\mathcal {S} _ {t 2 v} ^ {\\mathrm {v i d}}\\right) + \\mathbf {K L} \\left(\\mathcal {S} _ {v 2 t} ^ {\\mathrm {i m g}} \\| \\mathcal {S} _ {v 2 t} ^ {\\mathrm {v i d}}\\right)\\right). \\tag {13}", + "image_path": "e073ec58b8dfbdd28f1c784c69506030516d62bace17d60e00a8423afc28df18.jpg" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 313, + 159, + 553, + 225 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 159, + 553, + 225 + ], + "spans": [ + { + "bbox": [ + 313, + 159, + 553, + 225 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 313, + 159, + 553, + 225 + ], + "type": "inline_equation", + "content": "\\mathrm{s}_{ij}^{\\mathrm{img}}" + }, + { + "bbox": [ + 313, + 159, + 553, + 225 + ], + "type": "text", + "content": " is " + }, + { + "bbox": [ + 313, + 159, + 553, + 225 + ], + "type": "inline_equation", + "content": "\\mathrm{Sim}_{ij}^{\\mathrm{img}}" + }, + { + "bbox": [ + 313, + 159, + 553, + 225 + ], + "type": "text", + "content": " in Eq. (10) and " + }, + { + "bbox": [ + 313, + 159, + 553, + 225 + ], + "type": "inline_equation", + "content": "\\mathrm{s}_{ij}^{\\mathrm{vid}}" + }, + { + "bbox": [ + 313, + 159, + 553, + 225 + ], + "type": "text", + "content": " is " + }, + { + "bbox": [ + 313, + 159, + 553, + 225 + ], + "type": "inline_equation", + "content": "\\mathrm{Sim}_{ij}^{\\mathrm{vid}}" + }, + { + "bbox": [ + 313, + 159, + 553, + 225 + ], + "type": "text", + "content": " in Eq. (1). " + }, + { + "bbox": [ + 313, + 159, + 553, + 225 + ], + "type": "inline_equation", + "content": "S_{t2v}^{\\mathrm{img}}" + }, + { + "bbox": [ + 313, + 159, + 553, + 225 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 313, + 159, + 553, + 225 + ], + "type": "inline_equation", + "content": "S_{t2v}^{\\mathrm{vid}}" + }, + { + "bbox": [ + 313, + 159, + 553, + 225 + ], + "type": "text", + "content": " denote the probability distributions for the text-to-video task, while " + }, + { + "bbox": [ + 313, + 159, + 553, + 225 + ], + "type": "inline_equation", + "content": "S_{v2t}^{\\mathrm{img}}" + }, + { + "bbox": [ + 313, + 159, + 553, + 225 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 313, + 159, + 553, + 225 + ], + "type": "inline_equation", + "content": "S_{v2t}^{\\mathrm{vid}}" + }, + { + "bbox": [ + 313, + 159, + 553, + 225 + ], + "type": "text", + "content": " denote those for the video-to-text task. During the inference phase, We employ enhanced video-level similarity for retrieval tasks." + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 313, + 226, + 553, + 262 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 226, + 553, + 262 + ], + "spans": [ + { + "bbox": [ + 313, + 226, + 553, + 262 + ], + "type": "text", + "content": "Finally, the overall training objective, which includes alignment loss (Eq. (4)) and distillation loss (Eq. (13)), is formulated as follows:" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 334, + 271, + 553, + 286 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 334, + 271, + 553, + 286 + ], + "spans": [ + { + "bbox": [ + 334, + 271, + 553, + 286 + ], + "type": "interline_equation", + "content": "\\mathcal {L} = \\mathcal {L} _ {A} (\\operatorname {S i m} ^ {\\mathrm {v i d}}) + \\alpha \\mathcal {L} _ {A} (\\operatorname {S i m} ^ {\\mathrm {i m g}}) + \\beta \\mathcal {L} _ {K L}, \\tag {14}", + "image_path": "027375c2bbfe7c5ba3a0b0ec0ffe9d6802072aa78e19d179a5193d812812f65b.jpg" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 313, + 297, + 553, + 323 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 297, + 553, + 323 + ], + "spans": [ + { + "bbox": [ + 313, + 297, + 553, + 323 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 313, + 297, + 553, + 323 + ], + "type": "inline_equation", + "content": "\\alpha" + }, + { + "bbox": [ + 313, + 297, + 553, + 323 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 313, + 297, + 553, + 323 + ], + "type": "inline_equation", + "content": "\\beta" + }, + { + "bbox": [ + 313, + 297, + 553, + 323 + ], + "type": "text", + "content": " are used to balance the loss. " + }, + { + "bbox": [ + 313, + 297, + 553, + 323 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_A(\\mathrm{Sim}^{\\mathrm{img}})" + }, + { + "bbox": [ + 313, + 297, + 553, + 323 + ], + "type": "text", + "content": " is the image-level alignment loss (see Eq. (4) and (10))." + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 313, + 335, + 395, + 348 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 335, + 395, + 348 + ], + "spans": [ + { + "bbox": [ + 313, + 335, + 395, + 348 + ], + "type": "text", + "content": "4. Experiments" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 313, + 355, + 441, + 369 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 355, + 441, + 369 + ], + "spans": [ + { + "bbox": [ + 313, + 355, + 441, + 369 + ], + "type": "text", + "content": "4.1. Experimental Settings" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 313, + 373, + 555, + 505 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 373, + 555, + 505 + ], + "spans": [ + { + "bbox": [ + 313, + 373, + 555, + 505 + ], + "type": "text", + "content": "Following common practice, we perform evaluations on four widely used benchmarks: MSRVTT [53], LSMDC [43], ActivityNet [29], and DiDeMo [1]. We evaluate the performance using common retrieval metrics such as Recall at K (R@K and K = 1, 5, 10), the sum of these recalls (R@sum), and Mean Rank (MnR). In all experiments, the LoRA dimension and the adapter dimension " + }, + { + "bbox": [ + 313, + 373, + 555, + 505 + ], + "type": "inline_equation", + "content": "r" + }, + { + "bbox": [ + 313, + 373, + 555, + 505 + ], + "type": "text", + "content": " are set to 8. In Eq. 14, " + }, + { + "bbox": [ + 313, + 373, + 555, + 505 + ], + "type": "inline_equation", + "content": "\\alpha" + }, + { + "bbox": [ + 313, + 373, + 555, + 505 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 313, + 373, + 555, + 505 + ], + "type": "inline_equation", + "content": "\\beta" + }, + { + "bbox": [ + 313, + 373, + 555, + 505 + ], + "type": "text", + "content": " are set to 0.3 and 1.0, respectively. For the number of IVFusion layers, we set " + }, + { + "bbox": [ + 313, + 373, + 555, + 505 + ], + "type": "inline_equation", + "content": "H_V = 4" + }, + { + "bbox": [ + 313, + 373, + 555, + 505 + ], + "type": "text", + "content": " for the vision encoder and " + }, + { + "bbox": [ + 313, + 373, + 555, + 505 + ], + "type": "inline_equation", + "content": "H_L = 2" + }, + { + "bbox": [ + 313, + 373, + 555, + 505 + ], + "type": "text", + "content": " for the text encoder. Please see Appendix A for further details." + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 313, + 514, + 544, + 527 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 514, + 544, + 527 + ], + "spans": [ + { + "bbox": [ + 313, + 514, + 544, + 527 + ], + "type": "text", + "content": "4.2. Comparisons with State-of-the-Art Methods" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 313, + 533, + 553, + 568 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 533, + 553, + 568 + ], + "spans": [ + { + "bbox": [ + 313, + 533, + 553, + 568 + ], + "type": "text", + "content": "We compare our proposed DiscoVLA with state-of-the-art methods on popular benchmarks such as MSRVTT, LSMDC, ActivityNet, and DiDeMo." + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 313, + 569, + 554, + 689 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 569, + 554, + 689 + ], + "spans": [ + { + "bbox": [ + 313, + 569, + 554, + 689 + ], + "type": "text", + "content": "Full fine-tuning refers to CLIP4Clip-meanP [37], which averages image features along the temporal dimension. All parameter-efficient methods follow this temporal average pooling strategy to obtain the final video features. Although Prompt [27], Adapter [20], and LoRA [21] are widely adopted for parameter-efficient image and NLP tasks, they lack effective temporal modeling for video-level understanding. In contrast, methods like RAP [5], VoP [22], DGL [58], and MV-Adapter [25] focus on temporal modeling in the vision modality." + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 313, + 689, + 554, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 689, + 554, + 713 + ], + "spans": [ + { + "bbox": [ + 313, + 689, + 554, + 713 + ], + "type": "text", + "content": "As shown in Table 1, on MSRVTT, our DiscoVLA with CLIP (ViT-B/32) as the backbone achieves " + }, + { + "bbox": [ + 313, + 689, + 554, + 713 + ], + "type": "inline_equation", + "content": "47.0\\%" + }, + { + "bbox": [ + 313, + 689, + 554, + 713 + ], + "type": "text", + "content": " R@1 in" + } + ] + } + ], + "index": 25 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "type": "text", + "content": "19706" + } + ] + } + ], + "index": 26 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 4 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 72, + 70, + 536, + 280 + ], + "blocks": [ + { + "bbox": [ + 72, + 70, + 536, + 280 + ], + "lines": [ + { + "bbox": [ + 72, + 70, + 536, + 280 + ], + "spans": [ + { + "bbox": [ + 72, + 70, + 536, + 280 + ], + "type": "table", + "html": "
Method# Params (M)Text-to-VideoVideo-to-Text
R@1↑R@5↑R@10↑R@sum↑MnR↓R@1↑R@5↑R@10↑R@sum↑MnR↓
CLIP (ViT-B/32)
Full fine-tuning123.5443.170.480.8194.316.243.170.581.2194.812.4
Prompt [27]0.0840.466.377.3184.016.742.269.779.2191.112.4
Adapter [20]0.2641.969.978.7190.214.943.669.980.1193.611.5
LoRA [21]0.4943.768.980.4193.016.043.070.282.2195.412.0
RAP [5]1.0644.871.481.5197.714.444.071.982.4198.310.1
VoPF+P [22]0.443.569.379.3192.114.843.671.281.2196.011.0
VoPF+C [22]14.1044.770.579.2194.416.242.170.080.6192.713.4
DGL [58]0.8344.669.980.3194.816.344.570.780.6195.811.5
DiscoVLA0.5647.073.082.8202.814.147.773.683.6204.910.0
CLIP (ViT-B/16)
MV-Adapter [25]3.646.072.082.1200.1-45.674.083.8203.4-
RAP [5]1.0646.573.982.0202.412.145.376.484.8206.59.1
VoPF+P [22]0.447.172.481.8201.312.9-----
VoPF+C [22]14.1047.772.482.2202.312.0-----
DGL [58]0.8348.371.880.6200.713.445.774.082.9202.610.9
TempMe [44]0.5049.074.483.3206.711.947.675.385.4208.39.0
DiscoVLA0.5650.575.683.8209.912.149.276.084.7209.98.6
", + "image_path": "811aca382ac40874239e8f66b6a50d4e99c1747b898de85248e9097c9bdde09a.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "table_body" + } + ], + "index": 0 + }, + { + "type": "table", + "bbox": [ + 75, + 319, + 533, + 396 + ], + "blocks": [ + { + "bbox": [ + 55, + 281, + 555, + 314 + ], + "lines": [ + { + "bbox": [ + 55, + 281, + 555, + 314 + ], + "spans": [ + { + "bbox": [ + 55, + 281, + 555, + 314 + ], + "type": "text", + "content": "Table 1. Comparisons with state-of-the-art methods on MSRVTT. # Params denotes the number of trainable parameters. " + }, + { + "bbox": [ + 55, + 281, + 555, + 314 + ], + "type": "inline_equation", + "content": "\\uparrow" + }, + { + "bbox": [ + 55, + 281, + 555, + 314 + ], + "type": "text", + "content": " denotes higher values represent better performance and " + }, + { + "bbox": [ + 55, + 281, + 555, + 314 + ], + "type": "inline_equation", + "content": "\\downarrow" + }, + { + "bbox": [ + 55, + 281, + 555, + 314 + ], + "type": "text", + "content": " denotes lower values represent better performance. The best value for each metric is highlighted in bold. R@sum is defined as the total of R@1, R@5, and R@10. The gray row represents the fully fine-tuned CLIP4Clip-meanP [37]." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 75, + 319, + 533, + 396 + ], + "lines": [ + { + "bbox": [ + 75, + 319, + 533, + 396 + ], + "spans": [ + { + "bbox": [ + 75, + 319, + 533, + 396 + ], + "type": "table", + "html": "
Method# Params (M)Text-to-VideoVideo-to-Text
R@1↑R@5↑R@10↑R@sum↑MnR↓R@1↑R@5↑R@10↑R@sum↑MnR↓
Full fine-tuning123.5420.738.947.2106.865.320.639.447.5107.556.7
VoPF+P [22]0.420.740.749.7111.159.121.540.650.7112.850.8
VoPF+C [22]14.1021.140.949.6111.660.122.340.350.7113.351.1
DGL [58]0.8321.439.448.4109.264.3-----
DiscoVLA0.5623.642.052.3117.952.022.841.951.2115.946.6
", + "image_path": "a5a7903817c034caa677995764389cb1cc1e597cb314aacec7b9fc5e3bfc0cca.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "table_body" + } + ], + "index": 2 + }, + { + "type": "table", + "bbox": [ + 75, + 411, + 533, + 500 + ], + "blocks": [ + { + "bbox": [ + 144, + 396, + 465, + 407 + ], + "lines": [ + { + "bbox": [ + 144, + 396, + 465, + 407 + ], + "spans": [ + { + "bbox": [ + 144, + 396, + 465, + 407 + ], + "type": "text", + "content": "Table 2. Comparisons with state-of-the-art methods on LSMDC using CLIP (ViT-B/32)." + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 75, + 411, + 533, + 500 + ], + "lines": [ + { + "bbox": [ + 75, + 411, + 533, + 500 + ], + "spans": [ + { + "bbox": [ + 75, + 411, + 533, + 500 + ], + "type": "table", + "html": "
Method# Params (M)Text-to-VideoVideo-to-Text
R@1↑R@5↑R@10↑R@sum↑MnR↓R@1↑R@5↑R@10↑R@sum↑MnR↓
Full fine-tuning123.5440.572.4--7.442.574.185.8202.46.6
RAP [5]1.0640.871.082.2194.08.3-----
VoPF+P [22]0.436.165.578.5180.110.936.365.979.2181.410.1
VoPF+C [22]14.1035.163.777.6176.411.435.665.977.8179.310.4
DGL [58]0.8338.669.281.6189.49.0-----
DiscoVLA0.5641.272.483.6197.27.841.872.884.7199.36.9
", + "image_path": "bd378c525d80127c71fa8ac9d0c05c1920522dc51787a2993d585255365a49ab.jpg" + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "table_body" + } + ], + "index": 4 + }, + { + "bbox": [ + 138, + 500, + 470, + 511 + ], + "lines": [ + { + "bbox": [ + 138, + 500, + 470, + 511 + ], + "spans": [ + { + "bbox": [ + 138, + 500, + 470, + 511 + ], + "type": "text", + "content": "Table 3. Comparisons with state-of-the-art methods on ActivityNet using CLIP (ViT-B/32)." + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "text" + }, + { + "bbox": [ + 54, + 522, + 295, + 654 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 54, + 522, + 295, + 654 + ], + "spans": [ + { + "bbox": [ + 54, + 522, + 295, + 654 + ], + "type": "text", + "content": "the text-to-video task (t2v) and " + }, + { + "bbox": [ + 54, + 522, + 295, + 654 + ], + "type": "inline_equation", + "content": "47.7\\%" + }, + { + "bbox": [ + 54, + 522, + 295, + 654 + ], + "type": "text", + "content": " R@1 in the video-to-text task (v2t), significantly outperforming previous methods. Furthermore, when using CLIP (ViT-B/16), DiscoVLA achieves improvements of " + }, + { + "bbox": [ + 54, + 522, + 295, + 654 + ], + "type": "inline_equation", + "content": "2.2\\%" + }, + { + "bbox": [ + 54, + 522, + 295, + 654 + ], + "type": "text", + "content": " R@1 and " + }, + { + "bbox": [ + 54, + 522, + 295, + 654 + ], + "type": "inline_equation", + "content": "7.5\\%" + }, + { + "bbox": [ + 54, + 522, + 295, + 654 + ], + "type": "text", + "content": " R@sum. The comparison results on LSMDC, ActivityNet, DiDeMo are shown in Table 2-4. Our method outperforms all existing approaches across these datasets. Existing approaches overlook the critical image-to-video language and alignment discrepancies. In contrast, our proposed DiscoVLA addresses these discrepancies across all three areas: vision, language, and alignment." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 55, + 660, + 149, + 673 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 660, + 149, + 673 + ], + "spans": [ + { + "bbox": [ + 55, + 660, + 149, + 673 + ], + "type": "text", + "content": "4.3. Ablation Study" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 55, + 677, + 296, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 677, + 296, + 715 + ], + "spans": [ + { + "bbox": [ + 55, + 677, + 296, + 715 + ], + "type": "text", + "content": "Ablation study on individual components. We evaluate the proposed components in Table 5. LoRA applies Image-Level Attn (see Figure 3a) in both the vision and text" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 313, + 522, + 555, + 701 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 522, + 555, + 701 + ], + "spans": [ + { + "bbox": [ + 313, + 522, + 555, + 701 + ], + "type": "text", + "content": "encoders. In B1, IVFusion significantly enhances performance by " + }, + { + "bbox": [ + 313, + 522, + 555, + 701 + ], + "type": "inline_equation", + "content": "8.7\\%" + }, + { + "bbox": [ + 313, + 522, + 555, + 701 + ], + "type": "text", + "content": " R@sum in t2v and " + }, + { + "bbox": [ + 313, + 522, + 555, + 701 + ], + "type": "inline_equation", + "content": "6.7\\%" + }, + { + "bbox": [ + 313, + 522, + 555, + 701 + ], + "type": "text", + "content": " R@sum in v2t, effectively integrating image-level and video-level features. In B2 and B3, we find that using PImgAlign or AlignDistill independently yields limited improvements. This is largely due to the notable gap between image-level and video-level (see Figure 1). Image-level PImgAlign struggles to impact video-level tasks, while AlignDistill lacks sufficient image-specific learning. When combined, however, they complement each other. PImgAlign captures fine-grained image-level alignments, and AlignDistill effectively transfers this knowledge to the video-level alignment. This improves " + }, + { + "bbox": [ + 313, + 522, + 555, + 701 + ], + "type": "inline_equation", + "content": "1.1\\%" + }, + { + "bbox": [ + 313, + 522, + 555, + 701 + ], + "type": "text", + "content": " R@sum in t2v and " + }, + { + "bbox": [ + 313, + 522, + 555, + 701 + ], + "type": "inline_equation", + "content": "2.8\\%" + }, + { + "bbox": [ + 313, + 522, + 555, + 701 + ], + "type": "text", + "content": " R@sum in v2t over B1. Finally, our full DiscoVLA improves " + }, + { + "bbox": [ + 313, + 522, + 555, + 701 + ], + "type": "inline_equation", + "content": "9.8\\%" + }, + { + "bbox": [ + 313, + 522, + 555, + 701 + ], + "type": "text", + "content": " R@sum in t2v and " + }, + { + "bbox": [ + 313, + 522, + 555, + 701 + ], + "type": "inline_equation", + "content": "9.5\\%" + }, + { + "bbox": [ + 313, + 522, + 555, + 701 + ], + "type": "text", + "content": " R@sum in v2t over LoRA." + } + ] + } + ], + "index": 9 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 318, + 758 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 318, + 758 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 318, + 758 + ], + "type": "text", + "content": "19707" + } + ] + } + ], + "index": 10 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 5 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 75, + 70, + 533, + 148 + ], + "blocks": [ + { + "bbox": [ + 75, + 70, + 533, + 148 + ], + "lines": [ + { + "bbox": [ + 75, + 70, + 533, + 148 + ], + "spans": [ + { + "bbox": [ + 75, + 70, + 533, + 148 + ], + "type": "table", + "html": "
Method# Params (M)Text-to-VideoVideo-to-Text
R@1↑R@5↑R@10↑R@sum↑MnR↓R@1↑R@5↑R@10↑R@sum↑MnR↓
Full fine-tuning123.5443.470.280.6194.217.542.570.680.2193.311.6
RAP [5]1.0642.670.479.6192.618.0-----
VoPF+P [22]0.445.372.380.4198.013.844.771.281.1197.09.9
VoPF+C [22]14.1046.471.981.5199.813.644.471.881.8198.09.5
DiscoVLA0.5648.474.582.7205.614.047.774.483.8205.99.3
", + "image_path": "a95bab51d49738d0843e405888fe37e2f953209995ba6bc1fa3e9b927bce5000.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "table_body" + } + ], + "index": 0 + }, + { + "type": "table", + "bbox": [ + 58, + 163, + 553, + 249 + ], + "blocks": [ + { + "bbox": [ + 144, + 148, + 465, + 158 + ], + "lines": [ + { + "bbox": [ + 144, + 148, + 465, + 158 + ], + "spans": [ + { + "bbox": [ + 144, + 148, + 465, + 158 + ], + "type": "text", + "content": "Table 4. Comparisons with state-of-the-art methods on DiDeMo using CLIP (ViT-B/32)." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 58, + 163, + 553, + 249 + ], + "lines": [ + { + "bbox": [ + 58, + 163, + 553, + 249 + ], + "spans": [ + { + "bbox": [ + 58, + 163, + 553, + 249 + ], + "type": "table", + "html": "
MethodsComponents# Params (M)Text-to-VideoVideo-to-Text
IVFusionPIImgAlignAlignDistillR@1↑R@5↑R@10↑R@sum↑R@1↑R@5↑R@10↑R@sum↑
Pretraind CLIPXXX030.853.863.3147.926.650.162.0138.7
LoRAXXX0.4943.768.980.4193.043.070.282.2195.4
B1✓XX0.5446.573.282.0201.746.672.682.9202.1
B2✓✓X0.5647.073.181.3201.445.772.982.4201.0
B3✓X✓0.5446.671.681.0199.245.971.982.2200.0
DiscoVLA✓✓✓0.5647.073.082.8202.847.773.683.6204.9
", + "image_path": "4de793244a6ffc5548b625922d297337421431621587b97d1fbed686874bcc6f.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "table_body" + } + ], + "index": 2 + }, + { + "type": "table", + "bbox": [ + 58, + 290, + 294, + 354 + ], + "blocks": [ + { + "bbox": [ + 55, + 250, + 553, + 284 + ], + "lines": [ + { + "bbox": [ + 55, + 250, + 553, + 284 + ], + "spans": [ + { + "bbox": [ + 55, + 250, + 553, + 284 + ], + "type": "text", + "content": "Table 5. Ablation study on the contribution of each proposed component on MSRVTT using CLIP (ViT-B/32). Pretrained CLIP denotes the zero-shot performance of CLIP without any additional training. Our proposed methods are built upon LoRA. B1, B2, and B3 denote various combinations of our proposed components." + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 58, + 290, + 294, + 354 + ], + "lines": [ + { + "bbox": [ + 58, + 290, + 294, + 354 + ], + "spans": [ + { + "bbox": [ + 58, + 290, + 294, + 354 + ], + "type": "table", + "html": "
Parameter-Efficient AttentionText-to-Video
R@1↑R@5↑R@10↑R@sum↑
Image-Level Attn43.768.980.4193.0
Video-Level Attn45.770.980.8197.4
IVFusion Attn w/o Adapter46.071.080.4197.4
IVFusion Attn46.573.282.0201.7
", + "image_path": "7eb3e177f63cafd83480656c66b7d4ac4ec410146817651e40e015edc4dea50f.jpg" + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "table_body" + } + ], + "index": 4 + }, + { + "type": "table", + "bbox": [ + 58, + 403, + 294, + 467 + ], + "blocks": [ + { + "bbox": [ + 55, + 354, + 295, + 399 + ], + "lines": [ + { + "bbox": [ + 55, + 354, + 295, + 399 + ], + "spans": [ + { + "bbox": [ + 55, + 354, + 295, + 399 + ], + "type": "text", + "content": "Table 6. Ablation study on IVFusion on MSRVTT using CLIP (ViT-B/32). We compare several implementations against our proposed IVFusion for Parameter-Efficient Attention (see Figure 3). Image-Level Attn represents the LoRA baseline." + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 58, + 403, + 294, + 467 + ], + "lines": [ + { + "bbox": [ + 58, + 403, + 294, + 467 + ], + "spans": [ + { + "bbox": [ + 58, + 403, + 294, + 467 + ], + "type": "table", + "html": "
Similarity in PImgAlignText-to-Video
R@1↑R@5↑R@10↑R@sum↑
None46.671.681.0199.2
Video-Level46.372.381.2199.8
Paired Image-Level46.173.282.4201.7
Fine-grained Image-Level47.073.082.8202.8
", + "image_path": "ef4fc948a9d8c1a36962f76c8d7c80eec8c9313c0e05ca334a81f4e5047037b2.jpg" + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "table_body" + } + ], + "index": 6 + }, + { + "bbox": [ + 55, + 534, + 295, + 664 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 534, + 295, + 664 + ], + "spans": [ + { + "bbox": [ + 55, + 534, + 295, + 664 + ], + "type": "text", + "content": "Analysis of trainable parameters. The trainable parameters of our DiscoVLA consist of the LoRA and Adapter parts. As shown in Table 5, LoRA in both the vision and text encoders contain " + }, + { + "bbox": [ + 55, + 534, + 295, + 664 + ], + "type": "inline_equation", + "content": "\\sim 0.49\\mathrm{M}" + }, + { + "bbox": [ + 55, + 534, + 295, + 664 + ], + "type": "text", + "content": " parameters. In B1, an additional " + }, + { + "bbox": [ + 55, + 534, + 295, + 664 + ], + "type": "inline_equation", + "content": "\\sim 0.05\\mathrm{M}" + }, + { + "bbox": [ + 55, + 534, + 295, + 664 + ], + "type": "text", + "content": " parameters are introduced by the Adapter in the " + }, + { + "bbox": [ + 55, + 534, + 295, + 664 + ], + "type": "inline_equation", + "content": "H_{V} = 4" + }, + { + "bbox": [ + 55, + 534, + 295, + 664 + ], + "type": "text", + "content": " IVFusion layers of the vision encoder. In B2, an additional " + }, + { + "bbox": [ + 55, + 534, + 295, + 664 + ], + "type": "inline_equation", + "content": "\\sim 0.02\\mathrm{M}" + }, + { + "bbox": [ + 55, + 534, + 295, + 664 + ], + "type": "text", + "content": " parameters are introduced by the Adapter in the " + }, + { + "bbox": [ + 55, + 534, + 295, + 664 + ], + "type": "inline_equation", + "content": "H_{L} = 2" + }, + { + "bbox": [ + 55, + 534, + 295, + 664 + ], + "type": "text", + "content": " IVFusion layers of the text encoder, which are utilized for pseudo-image caption features in PImgAlign. Overall, the total number of trainable parameters of DiscoVLA is " + }, + { + "bbox": [ + 55, + 534, + 295, + 664 + ], + "type": "inline_equation", + "content": "\\sim 0.56\\mathrm{M}" + }, + { + "bbox": [ + 55, + 534, + 295, + 664 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 55, + 665, + 296, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 665, + 296, + 714 + ], + "spans": [ + { + "bbox": [ + 55, + 665, + 296, + 714 + ], + "type": "text", + "content": "Ablation study on IVFusion. In Table 6, we compare different strategies for parameter-efficient attention in Figure 3. Video-Level Attn concatenates all tokens across sampled images to capture spatio-temporal information," + } + ] + } + ], + "index": 9 + }, + { + "type": "table", + "bbox": [ + 316, + 290, + 553, + 337 + ], + "blocks": [ + { + "bbox": [ + 55, + 468, + 295, + 522 + ], + "lines": [ + { + "bbox": [ + 55, + 468, + 295, + 522 + ], + "spans": [ + { + "bbox": [ + 55, + 468, + 295, + 522 + ], + "type": "text", + "content": "Table 7. Ablation study on similarity in PImgAlign on MSRVTT using CLIP (ViT-B/32). None represents DiscoVLA without image-level alignment optimization, shown as B3 in Table 5. Paired Image-Level is based on matched image-caption pairs. Fine-grained Image-Level represents the full DiscoVLA." + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 316, + 290, + 553, + 337 + ], + "lines": [ + { + "bbox": [ + 316, + 290, + 553, + 337 + ], + "spans": [ + { + "bbox": [ + 316, + 290, + 553, + 337 + ], + "type": "table", + "html": "
Image Caption GenerationText-to-Video
R@1↑R@5↑R@10↑R@sum↑
DiscoVLA47.073.082.8202.8
W/O VidCap46.1 (-0.9)73.082.1 (-0.7)201.2 (-1.6)
", + "image_path": "cabb0ee4fb81f45229bc4c29854fd64c6b8412b318c92742c1c3e31c7de73074.jpg" + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "table_body" + } + ], + "index": 10 + }, + { + "bbox": [ + 313, + 338, + 553, + 371 + ], + "lines": [ + { + "bbox": [ + 313, + 338, + 553, + 371 + ], + "spans": [ + { + "bbox": [ + 313, + 338, + 553, + 371 + ], + "type": "text", + "content": "Table 8. Ablation study on image caption generation on MSRVTT using CLIP (ViT-B/32). W/O VidCap denotes generating pseudo image captions without video caption guidance." + } + ] + } + ], + "index": 11, + "angle": 0, + "type": "text" + }, + { + "bbox": [ + 313, + 388, + 555, + 483 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 388, + 555, + 483 + ], + "spans": [ + { + "bbox": [ + 313, + 388, + 555, + 483 + ], + "type": "text", + "content": "achieving a notable " + }, + { + "bbox": [ + 313, + 388, + 555, + 483 + ], + "type": "inline_equation", + "content": "4.4\\%" + }, + { + "bbox": [ + 313, + 388, + 555, + 483 + ], + "type": "text", + "content": " R@sum improvement. To reduce the computational complexity, IVFusion applies spatiotemporal attention only to global CLS tokens and merges image-level features with video-level features. Without an adapter, averaging image- and video-level features results in a modest performance increase. In contrast, with a learnable adapter, IVFusion merges the features more effectively, achieving " + }, + { + "bbox": [ + 313, + 388, + 555, + 483 + ], + "type": "inline_equation", + "content": "201.7\\%" + }, + { + "bbox": [ + 313, + 388, + 555, + 483 + ], + "type": "text", + "content": " R@sum." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 313, + 486, + 556, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 486, + 556, + 715 + ], + "spans": [ + { + "bbox": [ + 313, + 486, + 556, + 715 + ], + "type": "text", + "content": "Ablation study on similarity in PImgAlign. Table 7 demonstrates the effectiveness of our proposed fine-grained image-level similarity in PImgAlign. The similarity is computed between sampled images and pseudo image captions from a paired video and video caption. The Video-Level baseline averages the image features and pseudo-image caption features along the temporal dimension to compute the video-level similarity, replacing fine-grained image-level similarity. However, this modification leads to a notable decrease in performance, emphasizing the importance of image-level similarity in PImgAlign. In contrast, we introduce the Paired Image-Level baseline, which computes similarity based on matched image-caption pairs. Compared to Video-Level, Paired Image-Level yields a " + }, + { + "bbox": [ + 313, + 486, + 556, + 715 + ], + "type": "inline_equation", + "content": "1.9\\%" + }, + { + "bbox": [ + 313, + 486, + 556, + 715 + ], + "type": "text", + "content": " gain in R@sum. Moreover, we observe in Figure 4d that a single image caption may describe multiple images. Consequently, our fine-grained image-level similarity, as defined in Eq. 10, selects the maximum similarity from all available pairs, further enhancing video-level alignment." + } + ] + } + ], + "index": 13 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "text", + "content": "19708" + } + ] + } + ], + "index": 14 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 6 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 55, + 81, + 553, + 311 + ], + "blocks": [ + { + "bbox": [ + 55, + 81, + 553, + 311 + ], + "lines": [ + { + "bbox": [ + 55, + 81, + 553, + 311 + ], + "spans": [ + { + "bbox": [ + 55, + 81, + 553, + 311 + ], + "type": "image", + "image_path": "0b53ce37bcdb38f14d71a1c9fbed0ebd0a0de2a600f4af6fc821dc44e4c1a3bc.jpg" + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 55, + 312, + 555, + 357 + ], + "lines": [ + { + "bbox": [ + 55, + 312, + 555, + 357 + ], + "spans": [ + { + "bbox": [ + 55, + 312, + 555, + 357 + ], + "type": "text", + "content": "Figure 4. Visualization of discrepancies between video-level and image-level data. Each example consists of a paired video and video caption, with four sampled images from the video. Video captions are highlighted with orange solid lines, and pseudo image captions generated by LLaVA-NeXT [35] are indicated with green dashed lines. While video captions convey the general context of the video, image captions focus on the detailed context of each individual frame." + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_caption" + } + ], + "index": 1 + }, + { + "type": "image", + "bbox": [ + 57, + 372, + 295, + 464 + ], + "blocks": [ + { + "bbox": [ + 57, + 372, + 295, + 464 + ], + "lines": [ + { + "bbox": [ + 57, + 372, + 295, + 464 + ], + "spans": [ + { + "bbox": [ + 57, + 372, + 295, + 464 + ], + "type": "image", + "image_path": "7e4c331fad03523b338be4fd02a3c63a13d7db79123d524fa9d44dff7d422ba4.jpg" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 55, + 464, + 295, + 487 + ], + "lines": [ + { + "bbox": [ + 55, + 464, + 295, + 487 + ], + "spans": [ + { + "bbox": [ + 55, + 464, + 295, + 487 + ], + "type": "text", + "content": "Figure 5. Comparison of image caption generation with (W/) and without (W/O) video caption guidance (VidCap) in PImgAlign." + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_caption" + } + ], + "index": 3 + }, + { + "bbox": [ + 54, + 508, + 295, + 652 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 54, + 508, + 295, + 652 + ], + "spans": [ + { + "bbox": [ + 54, + 508, + 295, + 652 + ], + "type": "text", + "content": "Ablation study on image caption generation in PImgAlign. Figure 5 and Table 8 show qualitative and quantitative evaluations of video caption guidance. In Figure 5, video captions serve as the global context that enriches the understanding of individual frames (marked in red). In Table 8, without this guidance, image captions focus only on salient regions while missing globally relevant details, leading to performance drops. In PImgAlign, we employ MLLM to process both frame images and video descriptions, generating image captions that align with the overall video content. These high-quality captions further enhance DiscoVLA's ability to learn robust image-level alignment." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 54, + 653, + 295, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 54, + 653, + 295, + 715 + ], + "spans": [ + { + "bbox": [ + 54, + 653, + 295, + 715 + ], + "type": "text", + "content": "More Ablation Analysis. In Appendix B, we conduct additional ablation studies. Specifically, we evaluate the effects of hyperparameters " + }, + { + "bbox": [ + 54, + 653, + 295, + 715 + ], + "type": "inline_equation", + "content": "\\alpha" + }, + { + "bbox": [ + 54, + 653, + 295, + 715 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 54, + 653, + 295, + 715 + ], + "type": "inline_equation", + "content": "\\beta" + }, + { + "bbox": [ + 54, + 653, + 295, + 715 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 54, + 653, + 295, + 715 + ], + "type": "inline_equation", + "content": "H_V" + }, + { + "bbox": [ + 54, + 653, + 295, + 715 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 54, + 653, + 295, + 715 + ], + "type": "inline_equation", + "content": "H_V" + }, + { + "bbox": [ + 54, + 653, + 295, + 715 + ], + "type": "text", + "content": ". For fairness, all results in the main paper are reported without post-processing. Complementary evaluations with Q-Norm [4] and DSL [8]" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 313, + 374, + 545, + 387 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 374, + 545, + 387 + ], + "spans": [ + { + "bbox": [ + 313, + 374, + 545, + 387 + ], + "type": "text", + "content": "as post-processing methods are provided in the appendix." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 313, + 392, + 432, + 405 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 392, + 432, + 405 + ], + "spans": [ + { + "bbox": [ + 313, + 392, + 432, + 405 + ], + "type": "text", + "content": "4.4. Qualitative Analysis" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 313, + 409, + 555, + 566 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 409, + 555, + 566 + ], + "spans": [ + { + "bbox": [ + 313, + 409, + 555, + 566 + ], + "type": "text", + "content": "To compare image-level and video-level data, we conduct a visualization experiment in Figure 4. In Figure 4a and 4b, image-level data alone fails to capture the complete content of the video. Figure 4c and 4d demonstrate a strong correlation between the video-level and image-level content, but differences across images lead to varying image captions. Visually, a video is composed of a sequence of images, incorporating temporal information. Textually, image captions are specific to individual images, while video captions describe the overall video content. These discrepancies hinder the effective transfer of image-level CLIP. Our approach addresses this issue by utilizing image-level alignment to enhance video-level alignment." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 313, + 574, + 388, + 586 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 574, + 388, + 586 + ], + "spans": [ + { + "bbox": [ + 313, + 574, + 388, + 586 + ], + "type": "text", + "content": "5. Conclusion" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 313, + 594, + 555, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 594, + 555, + 715 + ], + "spans": [ + { + "bbox": [ + 313, + 594, + 555, + 715 + ], + "type": "text", + "content": "In this paper, we identify three major discrepancies in the parameter-efficient transfer of image-level CLIP to videotext retrieval: vision, language, and alignment. To address these challenges, we propose DiscoVLA. Specifically, IV-Fusion fuses image-level and video-level features to tackle the vision and language discrepancies. For alignment discrepancies, we propose PImgAlign and AlignDistill, which learn and distill image-level alignment to enhance video-level alignment. Our DiscoVLA significantly outperforms existing methods, achieving state-of-the-art performance." + } + ] + } + ], + "index": 11 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 91, + 70, + 216, + 79 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 91, + 70, + 216, + 79 + ], + "spans": [ + { + "bbox": [ + 91, + 70, + 216, + 79 + ], + "type": "text", + "content": "Video Caption - Image Caption" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "text", + "content": "19709" + } + ] + } + ], + "index": 12 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 7 + }, + { + "para_blocks": [ + { + "bbox": [ + 56, + 72, + 147, + 85 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 72, + 147, + 85 + ], + "spans": [ + { + "bbox": [ + 56, + 72, + 147, + 85 + ], + "type": "text", + "content": "Acknowledgment" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 56, + 222, + 115, + 235 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 222, + 115, + 235 + ], + "spans": [ + { + "bbox": [ + 56, + 222, + 115, + 235 + ], + "type": "text", + "content": "References" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 57, + 242, + 296, + 713 + ], + "type": "list", + "angle": 0, + "index": 13, + "blocks": [ + { + "bbox": [ + 61, + 242, + 296, + 286 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 61, + 242, + 296, + 286 + ], + "spans": [ + { + "bbox": [ + 61, + 242, + 296, + 286 + ], + "type": "text", + "content": "[1] Lisa Anne Hendricks, Oliver Wang, Eli Shechtman, Josef Sivic, Trevor Darrell, and Bryan Russell. Localizing moments in video with natural language. In Proceedings of the International Conference on Computer Vision, 2017. 5, 12" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 61, + 288, + 296, + 331 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 61, + 288, + 296, + 331 + ], + "spans": [ + { + "bbox": [ + 61, + 288, + 296, + 331 + ], + "type": "text", + "content": "[2] Stanislaw Antol, Aishwarya Agrawal, Jiasen Lu, Margaret Mitchell, Dhruv Batra, C Lawrence Zitnick, and Devi Parikh. Vqa: Visual question answering. In Proceedings of the International Conference on Computer Vision, 2015. 2" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 62, + 333, + 296, + 376 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 333, + 296, + 376 + ], + "spans": [ + { + "bbox": [ + 62, + 333, + 296, + 376 + ], + "type": "text", + "content": "[3] Max Bain, Arsha Nagrani, Gül Varol, and Andrew Zisserman. Frozen in time: A joint video and image encoder for end-to-end retrieval. In Proceedings of the International Conference on Computer Vision, 2021. 12" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 62, + 378, + 294, + 421 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 378, + 294, + 421 + ], + "spans": [ + { + "bbox": [ + 62, + 378, + 294, + 421 + ], + "type": "text", + "content": "[4] Simion-Vlad Bogolin, Ioana Croitoru, Hailin Jin, Yang Liu, and Samuel Albanie. Cross modal retrieval with querybank normalisation. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition, 2022. 8, 13" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 62, + 423, + 296, + 477 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 423, + 296, + 477 + ], + "spans": [ + { + "bbox": [ + 62, + 423, + 296, + 477 + ], + "type": "text", + "content": "[5] Meng Cao, Haoran Tang, Jinfa Huang, Peng Jin, Can Zhang, Ruyang Liu, Long Chen, Xiaodan Liang, Li Yuan, and Ge Li. RAP: Efficient text-video retrieval with sparse-and-correlated adapter. In Association for Computational Linguistics, 2024. 2, 3, 5, 6, 7, 12" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 62, + 479, + 296, + 521 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 479, + 296, + 521 + ], + "spans": [ + { + "bbox": [ + 62, + 479, + 296, + 521 + ], + "type": "text", + "content": "[6] Shoufa Chen, Chongjian Ge, Zhan Tong, Jiangliu Wang, Yibing Song, Jue Wang, and Ping Luo. Adaptformer: Adapting vision transformers for scalable visual recognition. Advances in Neural Information Processing Systems, 2022. 2" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 62, + 523, + 296, + 578 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 523, + 296, + 578 + ], + "spans": [ + { + "bbox": [ + 62, + 523, + 296, + 578 + ], + "type": "text", + "content": "[7] Sihan Chen, Handong Li, Qunbo Wang, Zijia Zhao, Mingzhen Sun, Xinxin Zhu, and Jing Liu. Vast: A vision-audio-subtitle-text omni-modality foundation model and dataset. Advances in Neural Information Processing Systems, 2023. 2" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 62, + 580, + 296, + 623 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 580, + 296, + 623 + ], + "spans": [ + { + "bbox": [ + 62, + 580, + 296, + 623 + ], + "type": "text", + "content": "[8] Xing Cheng, Hezheng Lin, Xiangyu Wu, Fan Yang, and Dong Shen. Improving video-text retrieval by multi-stream corpus alignment and dual softmax loss. arXiv preprint arXiv:2109.04290, 2021. 8, 13" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 62, + 624, + 296, + 668 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 624, + 296, + 668 + ], + "spans": [ + { + "bbox": [ + 62, + 624, + 296, + 668 + ], + "type": "text", + "content": "[9] Chaorui Deng, Qi Chen, Pengda Qin, Da Chen, and Qi Wu. Prompt switch: Efficient clip adaptation for text-video retrieval. In Proceedings of the International Conference on Computer Vision, 2023. 1, 2" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 57, + 670, + 296, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 670, + 296, + 713 + ], + "spans": [ + { + "bbox": [ + 57, + 670, + 296, + 713 + ], + "type": "text", + "content": "[10] Xiaohan Ding, Yuchen Guo, Guiguang Ding, and Jungong Han. Acnet: Strengthening the kernel skeletons for powerful cnn via asymmetric convolution blocks. In Proceedings of the International Conference on Computer Vision, 2019. 2" + } + ] + } + ], + "index": 12 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 57, + 73, + 555, + 713 + ], + "type": "list", + "angle": 0, + "index": 28, + "blocks": [ + { + "bbox": [ + 55, + 91, + 296, + 212 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 91, + 296, + 212 + ], + "spans": [ + { + "bbox": [ + 55, + 91, + 296, + 212 + ], + "type": "text", + "content": "This work was supported by Beijing Natural Science Foundation (No. L223023), National Natural Science Foundation of China (Nos. 62441235, 62021002, 62441614), the Key R & D Program of Xinjiang, China (2022B01006), Zhejiang Provincial Natural Science Foundation of China under Grant (No. LDT23F01013F01), CCF-DiDi GAIA Collaborative Research Funds, China Postdoctoral Science Foundation (2024M750565), Guangdong S & T Program (2024B0101040008), and the Key Realm Research and Development Program of Guangzhou (No. 2024B01W0007)." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 316, + 73, + 555, + 117 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 73, + 555, + 117 + ], + "spans": [ + { + "bbox": [ + 316, + 73, + 555, + 117 + ], + "type": "text", + "content": "[11] Xiaohan Ding, Xiangyu Zhang, Ningning Ma, Jungong Han, Guiguang Ding, and Jian Sun. Repvgg: Making vgg-style convnets great again. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition, 2021. 2" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 317, + 118, + 554, + 160 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 118, + 554, + 160 + ], + "spans": [ + { + "bbox": [ + 317, + 118, + 554, + 160 + ], + "type": "text", + "content": "[12] Lijie Fan, Dilip Krishnan, Phillip Isola, Dina Katabi, and Yonglong Tian. Improving clip training with language rewrites. Advances in Neural Information Processing Systems, 2024. 2" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 317, + 162, + 554, + 205 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 162, + 554, + 205 + ], + "spans": [ + { + "bbox": [ + 317, + 162, + 554, + 205 + ], + "type": "text", + "content": "[13] Valentin Gabeur, Chen Sun, Karteek Alahari, and Cordelia Schmid. Multi-modal transformer for video retrieval. In Proceedings of the European Conference on Computer Vision, 2020. 12" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 316, + 206, + 554, + 260 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 206, + 554, + 260 + ], + "spans": [ + { + "bbox": [ + 316, + 206, + 554, + 260 + ], + "type": "text", + "content": "[14] Satya Krishna Gorti, Noel Vouitsis, Junwei Ma, Keyvan Golestan, Maksims Volkovs, Animesh Garg, and Guangwei Yu. X-pool: Cross-modal language-video attention for text-video retrieval. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition, 2022. 1, 2" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 316, + 262, + 554, + 315 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 262, + 554, + 315 + ], + "spans": [ + { + "bbox": [ + 316, + 262, + 554, + 315 + ], + "type": "text", + "content": "[15] Priya Goyal, Piotr Dólar, Ross Girshick, Pieter Noordhuis, Lukasz Wesolowski, Aapo Kyrola, Andrew Tulloch, Yangqing Jia, and Kaiming He. Accurate, large minibatch sgd: Training imagenet in 1 hour. arXiv preprint arXiv:1706.02677, 2017. 12" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 316, + 316, + 554, + 348 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 316, + 554, + 348 + ], + "spans": [ + { + "bbox": [ + 316, + 316, + 554, + 348 + ], + "type": "text", + "content": "[16] Alex Graves and Jürgen Schmidhuber. Framewise phoneme classification with bidirectional LSTM and other neural network architectures. Neural Networks, 2005. 2" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 316, + 350, + 555, + 403 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 350, + 555, + 403 + ], + "spans": [ + { + "bbox": [ + 316, + 350, + 555, + 403 + ], + "type": "text", + "content": "[17] Junxian He, Chunting Zhou, Xuezhe Ma, Taylor Berg-Kirkpatrick, and Graham Neubig. Towards a unified view of parameter-efficient transfer learning. In Proceedings of the International Conference on Learning Representations, 2021. 2" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 316, + 404, + 554, + 459 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 404, + 554, + 459 + ], + "spans": [ + { + "bbox": [ + 316, + 404, + 554, + 459 + ], + "type": "text", + "content": "[18] Tao He, Leqi Shen, Yuchen Guo, Guiguang Ding, and Zhenhua Guo. Secret: Self-consistent pseudo label refinement for unsupervised domain adaptive person re-identification. In Proceedings of the AAAI Conference on Artificial Intelligence, 2022. 2" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 316, + 460, + 554, + 481 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 460, + 554, + 481 + ], + "spans": [ + { + "bbox": [ + 316, + 460, + 554, + 481 + ], + "type": "text", + "content": "[19] Dan Hendrycks and Kevin Gimpel. Gaussian error linear units (gelus). arXiv preprint arXiv:1606.08415, 2016. 4" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 317, + 483, + 554, + 536 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 483, + 554, + 536 + ], + "spans": [ + { + "bbox": [ + 317, + 483, + 554, + 536 + ], + "type": "text", + "content": "[20] Neil Houlsby, Andrei Giurgiu, Stanislaw Jastrzebski, Bruna Morrone, Quentin De Laroussilhe, Andrea Gesmundo, Mona Attariyan, and Sylvain Gelly. Parameter-efficient transfer learning for nlp. In Proceedings of the International Conference on Machine Learning, 2019. 2, 5, 6" + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 317, + 537, + 554, + 581 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 537, + 554, + 581 + ], + "spans": [ + { + "bbox": [ + 317, + 537, + 554, + 581 + ], + "type": "text", + "content": "[21] Edward J Hu, Yelong Shen, Phillip Wallis, Zeyuan Allen-Zhu, Yuanzhi Li, Shean Wang, Lu Wang, and Weizhu Chen. Lora: Low-rank adaptation of large language models. arXiv preprint arXiv:2106.09685, 2021. 2, 5, 6" + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 317, + 582, + 555, + 635 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 582, + 555, + 635 + ], + "spans": [ + { + "bbox": [ + 317, + 582, + 555, + 635 + ], + "type": "text", + "content": "[22] Siteng Huang, Biao Gong, Yulin Pan, Jianwen Jiang, Yiliang Lv, Yuyuan Li, and Donglin Wang. Vop: Text-video cooperative prompt tuning for cross-modal retrieval. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition, 2023. 1, 2, 3, 5, 6, 7, 12" + } + ] + } + ], + "index": 25 + }, + { + "bbox": [ + 317, + 636, + 554, + 680 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 636, + 554, + 680 + ], + "spans": [ + { + "bbox": [ + 317, + 636, + 554, + 680 + ], + "type": "text", + "content": "[23] Menglin Jia, Luming Tang, Bor-Chun Chen, Claire Cardie, Serge Belongie, Bharath Hariharan, and Ser-Nam Lim. Visual prompt tuning. In Proceedings of the European Conference on Computer Vision, 2022. 2" + } + ] + } + ], + "index": 26 + }, + { + "bbox": [ + 316, + 681, + 554, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 681, + 554, + 713 + ], + "spans": [ + { + "bbox": [ + 316, + 681, + 554, + 713 + ], + "type": "text", + "content": "[24] Peng Jin, Jina Huang, Pengfei Xiong, Shangxuan Tian, Chang Liu, Xiangyang Ji, Li Yuan, and Jie Chen. Videotext as game players: Hierarchical banzhaf interaction for" + } + ] + } + ], + "index": 27 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "type": "text", + "content": "19710" + } + ] + } + ], + "index": 29 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 8 + }, + { + "para_blocks": [ + { + "bbox": [ + 56, + 72, + 296, + 714 + ], + "type": "list", + "angle": 0, + "index": 14, + "blocks": [ + { + "bbox": [ + 76, + 72, + 296, + 106 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 76, + 72, + 296, + 106 + ], + "spans": [ + { + "bbox": [ + 76, + 72, + 296, + 106 + ], + "type": "text", + "content": "cross-modal representation learning. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition, 2023. 1, 2" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 56, + 107, + 296, + 173 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 107, + 296, + 173 + ], + "spans": [ + { + "bbox": [ + 56, + 107, + 296, + 173 + ], + "type": "text", + "content": "[25] Xiaojie Jin, Bowen Zhang, Weibo Gong, Kai Xu, Xueqing Deng, Peng Wang, Zhao Zhang, Xiaohui Shen, and Jiashi Feng. Mv-adapter: Multimodal video transfer learning for video text retrieval. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition, 2024. 2, 3, 5, 6, 12" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 56, + 175, + 295, + 218 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 175, + 295, + 218 + ], + "spans": [ + { + "bbox": [ + 56, + 175, + 295, + 218 + ], + "type": "text", + "content": "[26] Andrej Karpathy and Li Fei-Fei. Deep visual-semantic alignments for generating image descriptions. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition, 2015. 2" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 56, + 220, + 295, + 274 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 220, + 295, + 274 + ], + "spans": [ + { + "bbox": [ + 56, + 220, + 295, + 274 + ], + "type": "text", + "content": "[27] Muhammad Uzair Khattak, Hanoona Rasheed, Muhammad Maaz, Salman Khan, and Fahad Shahbaz Khan. Maple: Multi-modal prompt learning. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition, 2023. 2, 5, 6" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 56, + 275, + 295, + 330 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 275, + 295, + 330 + ], + "spans": [ + { + "bbox": [ + 56, + 275, + 295, + 330 + ], + "type": "text", + "content": "[28] Muhammad Uzair Khattak, Syed Talal Wasim, Muzammal Naseer, Salman Khan, Ming-Hsuan Yang, and Fahad Shahbaz Khan. Self-regulating prompts: Foundational model adaptation without forgetting. In Proceedings of the International Conference on Computer Vision, 2023. 2" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 56, + 331, + 295, + 374 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 331, + 295, + 374 + ], + "spans": [ + { + "bbox": [ + 56, + 331, + 295, + 374 + ], + "type": "text", + "content": "[29] Ranjay Krishna, Kenji Hata, Frederic Ren, Li Fei-Fei, and Juan Carlos Niebles. Dense captioning events in videos. In Proceedings of the International Conference on Computer Vision, 2017. 5, 12" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 56, + 376, + 294, + 398 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 376, + 294, + 398 + ], + "spans": [ + { + "bbox": [ + 56, + 376, + 294, + 398 + ], + "type": "text", + "content": "[30] Solomon Kullback. Information theory and statistics. Courier Corporation, 1997. 5" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 56, + 399, + 295, + 454 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 399, + 295, + 454 + ], + "spans": [ + { + "bbox": [ + 56, + 399, + 295, + 454 + ], + "type": "text", + "content": "[31] Jie Lei, Linjie Li, Luowei Zhou, Zhe Gan, Tamara L Berg, Mohit Bansal, and Jingjing Liu. Less is more: Clipbert for video-and-language learning via sparse sampling. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition, 2021. 12" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 56, + 456, + 295, + 510 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 456, + 295, + 510 + ], + "spans": [ + { + "bbox": [ + 56, + 456, + 295, + 510 + ], + "type": "text", + "content": "[32] Junnan Li, Ramprasaath Selvaraju, Akhilesh Gotmare, Shafiq Joty, Caiming Xiong, and Steven Chu Hong Hoi. Align before fuse: Vision and language representation learning with momentum distillation. Advances in Neural Information Processing Systems, 2021. 2" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 56, + 512, + 295, + 564 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 512, + 295, + 564 + ], + "spans": [ + { + "bbox": [ + 56, + 512, + 295, + 564 + ], + "type": "text", + "content": "[33] Junnan Li, Dongxu Li, Caiming Xiong, and Steven Hoi. Blip: Bootstrapping language-image pre-training for unified vision-language understanding and generation. In Proceedings of the International Conference on Machine Learning, 2022." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 56, + 567, + 295, + 621 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 567, + 295, + 621 + ], + "spans": [ + { + "bbox": [ + 56, + 567, + 295, + 621 + ], + "type": "text", + "content": "[34] Kunchang Li, Yali Wang, Yizhuo Li, Yi Wang, Yinan He, Limin Wang, and Yu Qiao. Unmasked teacher: Towards training-efficient video foundation models. In Proceedings of the International Conference on Computer Vision, 2023. 2" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 56, + 624, + 295, + 656 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 624, + 295, + 656 + ], + "spans": [ + { + "bbox": [ + 56, + 624, + 295, + 656 + ], + "type": "text", + "content": "[35] Haotian Liu, Chunyuan Li, Yuheng Li, Bo Li, Yuanhan Zhang, Sheng Shen, and Yong Jae Lee. Llava next: Improved reasoning,ocr,and world knowledge,2024.4,8" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 56, + 658, + 295, + 690 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 658, + 295, + 690 + ], + "spans": [ + { + "bbox": [ + 56, + 658, + 295, + 690 + ], + "type": "text", + "content": "[36] Ilya Loshchilov and Frank Hutter. Sgdr: Stochastic gradient descent with warm restarts. In Proceedings of the International Conference on Learning Representations, 2016. 12" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 56, + 692, + 295, + 714 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 692, + 295, + 714 + ], + "spans": [ + { + "bbox": [ + 56, + 692, + 295, + 714 + ], + "type": "text", + "content": "[37] Huaishao Luo, Lei Ji, Ming Zhong, Yang Chen, Wen Lei, Nan Duan, and Tianrui Li. Clip4clip: An empirical study of" + } + ] + } + ], + "index": 13 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 316, + 73, + 553, + 712 + ], + "type": "list", + "angle": 0, + "index": 29, + "blocks": [ + { + "bbox": [ + 333, + 73, + 553, + 95 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 333, + 73, + 553, + 95 + ], + "spans": [ + { + "bbox": [ + 333, + 73, + 553, + 95 + ], + "type": "text", + "content": "clip for end to end video clip retrieval and captioning. Neurocomputing, 2022. 1, 2, 5, 6" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 316, + 96, + 553, + 149 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 96, + 553, + 149 + ], + "spans": [ + { + "bbox": [ + 316, + 96, + 553, + 149 + ], + "type": "text", + "content": "[38] Zongyang Ma, Ziqi Zhang, Yuxin Chen, Zhongang Qi, Chunfeng Yuan, Bing Li, Yingmin Luo, Xu Li, Xiaojuan Qi, Ying Shan, et al. Ea-vtr: Event-aware video-text retrieval. In Proceedings of the European Conference on Computer Vision, 2024. 5" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 316, + 151, + 553, + 205 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 151, + 553, + 205 + ], + "spans": [ + { + "bbox": [ + 316, + 151, + 553, + 205 + ], + "type": "text", + "content": "[39] Antoine Miech, Dimitri Zhukov, Jean-Baptiste Alayrac, Makarand Tapaswi, Ivan Laptev, and Josef Sivic. Howto100m: Learning a text-video embedding by watching hundred million narrated video clips. In Proceedings of the International Conference on Computer Vision, 2019. 12" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 316, + 206, + 553, + 239 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 206, + 553, + 239 + ], + "spans": [ + { + "bbox": [ + 316, + 206, + 553, + 239 + ], + "type": "text", + "content": "[40] Aaron van den Oord, Yazhe Li, and Oriol Vinyals. Representation learning with contrastive predictive coding. arXiv preprint arXiv:1807.03748, 2018. 3" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 316, + 240, + 553, + 272 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 240, + 553, + 272 + ], + "spans": [ + { + "bbox": [ + 316, + 240, + 553, + 272 + ], + "type": "text", + "content": "[41] Alec Radford, Jeffrey Wu, Rewon Child, David Luan, Dario Amodei, Ilya Sutskever, et al. Language models are unsupervised multitask learners. OpenAI blog, 2019. 2" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 316, + 273, + 553, + 338 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 273, + 553, + 338 + ], + "spans": [ + { + "bbox": [ + 316, + 273, + 553, + 338 + ], + "type": "text", + "content": "[42] Alec Radford, Jong Wook Kim, Chris Hallacy, Aditya Ramesh, Gabriel Goh, Sandhini Agarwal, Girish Sastry, Amanda Askell, Pamela Mishkin, Jack Clark, et al. Learning transferable visual models from natural language supervision. In Proceedings of the International Conference on Machine Learning, 2021. 1, 2" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 316, + 338, + 553, + 370 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 338, + 553, + 370 + ], + "spans": [ + { + "bbox": [ + 316, + 338, + 553, + 370 + ], + "type": "text", + "content": "[43] Anna Rohrbach, Marcus Rohrbach, and Bernt Schiele. The long-short story of movie description. In Pattern Recognition, 2015. 5, 12" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 316, + 372, + 553, + 415 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 372, + 553, + 415 + ], + "spans": [ + { + "bbox": [ + 316, + 372, + 553, + 415 + ], + "type": "text", + "content": "[44] Leqi Shen, Tianxiang Hao, Sicheng Zhao, Yifeng Zhang, Pengzhang Liu, Yongjun Bao, and Guiguang Ding. Tempme: Video temporal token merging for efficient text-video retrieval. arXiv preprint arXiv:2409.01156, 2024. 6" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 316, + 416, + 553, + 470 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 416, + 553, + 470 + ], + "spans": [ + { + "bbox": [ + 316, + 416, + 553, + 470 + ], + "type": "text", + "content": "[45] Leqi Shen, Tao He, Sicheng Zhao, Zhelun Shen, Yuchen Guo, Tianshi Xu, and Guiguang Ding. X-reid: Cross-instance transformer for identity-level person re-identification. In IEEE International Conference on Multimedia and Expo, 2024. 2" + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 316, + 472, + 553, + 514 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 472, + 553, + 514 + ], + "spans": [ + { + "bbox": [ + 316, + 472, + 553, + 514 + ], + "type": "text", + "content": "[46] Yoad Tewel, Yoav Shalev, Idan Schwartz, and Lior Wolf. Zerocap: Zero-shot image-to-text generation for visual-semantic arithmetic. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition, 2022. 2" + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 316, + 515, + 553, + 557 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 515, + 553, + 557 + ], + "spans": [ + { + "bbox": [ + 316, + 515, + 553, + 557 + ], + "type": "text", + "content": "[47] Ao Wang, Hui Chen, Lihao Liu, Kai Chen, Zijia Lin, Jungong Han, et al. Yolov10: Real-time end-to-end object detection. Advances in Neural Information Processing Systems, 2024. 2" + } + ] + } + ], + "index": 25 + }, + { + "bbox": [ + 316, + 559, + 553, + 613 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 559, + 553, + 613 + ], + "spans": [ + { + "bbox": [ + 316, + 559, + 553, + 613 + ], + "type": "text", + "content": "[48] Jiamian Wang, Guohao Sun, Pichao Wang, Dongfang Liu, Sohail Dianat, Majid Rabbani, Raghuveer Rao, and Zhiqiang Tao. Text is mass: Modeling as stochastic embedding for text-video retrieval. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition, 2024. 1, 2" + } + ] + } + ], + "index": 26 + }, + { + "bbox": [ + 316, + 615, + 553, + 656 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 615, + 553, + 656 + ], + "spans": [ + { + "bbox": [ + 316, + 615, + 553, + 656 + ], + "type": "text", + "content": "[49] Qiang Wang, Yanhao Zhang, Yun Zheng, Pan Pan, and Xi-ansheng Hua. Disentangled representation learning for text-video retrieval. arXiv preprint arXiv:2203.07111, 2022. 1, 2" + } + ] + } + ], + "index": 27 + }, + { + "bbox": [ + 316, + 658, + 553, + 712 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 658, + 553, + 712 + ], + "spans": [ + { + "bbox": [ + 316, + 658, + 553, + 712 + ], + "type": "text", + "content": "[50] Yi Wang, Kunchang Li, Yizhuo Li, Yinan He, Bingkun Huang, Zhiyu Zhao, Hongjie Zhang, Jilan Xu, Yi Liu, Zun Wang, et al. Intervideo: General video foundation models via generative and discriminative learning. arXiv preprint arXiv:2212.03191, 2022. 2" + } + ] + } + ], + "index": 28 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "text", + "content": "19711" + } + ] + } + ], + "index": 30 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 9 + }, + { + "para_blocks": [ + { + "bbox": [ + 56, + 72, + 294, + 620 + ], + "type": "list", + "angle": 0, + "index": 12, + "blocks": [ + { + "bbox": [ + 56, + 72, + 294, + 126 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 72, + 294, + 126 + ], + "spans": [ + { + "bbox": [ + 56, + 72, + 294, + 126 + ], + "type": "text", + "content": "[51] Wenhao Wu, Haipeng Luo, Bo Fang, Jingdong Wang, and Wanli Ouyang. Cap4video: What can auxiliary captions do for text-video retrieval? In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition, 2023. 1, 2" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 56, + 129, + 294, + 184 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 129, + 294, + 184 + ], + "spans": [ + { + "bbox": [ + 56, + 129, + 294, + 184 + ], + "type": "text", + "content": "[52] Hu Xu, Saining Xie, Xiaqing Tan, Po-Yao Huang, Russell Howes, Vasu Sharma, Shang-Wen Li, Gargi Ghosh, Luke Zettlemoyer, and Christoph Feichtenhofer. Demystifying clip data. In Proceedings of the International Conference on Learning Representations, 2024. 2" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 56, + 185, + 294, + 228 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 185, + 294, + 228 + ], + "spans": [ + { + "bbox": [ + 56, + 185, + 294, + 228 + ], + "type": "text", + "content": "[53] Jun Xu, Tao Mei, Ting Yao, and Yong Rui. Msr-vtt: A large video description dataset for bridging video and language. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition, 2016. 5, 12" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 56, + 229, + 294, + 284 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 229, + 294, + 284 + ], + "spans": [ + { + "bbox": [ + 56, + 229, + 294, + 284 + ], + "type": "text", + "content": "[54] Kelvin Xu, Jimmy Ba, Ryan Kiros, Kyunghyun Cho, Aaron Courville, Ruslan Salakhudinov, Rich Zemel, and Yoshua Bengio. Show, attend and tell: Neural image caption generation with visual attention. In Proceedings of the International Conference on Machine Learning, 2015. 2" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 56, + 285, + 294, + 340 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 285, + 294, + 340 + ], + "spans": [ + { + "bbox": [ + 56, + 285, + 294, + 340 + ], + "type": "text", + "content": "[55] Hongwei Xue, Yuchong Sun, Bei Liu, Jianlong Fu, Ruihua Song, Houqiang Li, and Jiebo Luo. Clip-vip: Adapting pre-trained image-text model to video-language alignment. In Proceedings of the International Conference on Learning Representations, 2022. 2" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 56, + 341, + 294, + 396 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 341, + 294, + 396 + ], + "spans": [ + { + "bbox": [ + 56, + 341, + 294, + 396 + ], + "type": "text", + "content": "[56] Fan Yang, Sicheng Zhao, Yanhao Zhang, Haoxiang Chen, Hui Chen, Wenbo Tang, Haonan Lu, Pengfei Xu, Zhenyu Yang, Jungong Han, et al. Lmi3d: Empowering llm with 3d perception from a single 2d image. arXiv preprint arXiv:2408.07422, 2024. 2" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 56, + 396, + 294, + 441 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 396, + 294, + 441 + ], + "spans": [ + { + "bbox": [ + 56, + 396, + 294, + 441 + ], + "type": "text", + "content": "[57] Fan Yang, Ru Zhen, Jianing Wang, Yanhao Zhang, Haoxiang Chen, Haonan Lu, Sicheng Zhao, and Guiguang Ding. Heie: Mllm-based hierarchical explainable aigc image implausibility evaluator. arXiv preprint arXiv:2411.17261, 2024. 2" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 56, + 442, + 294, + 485 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 442, + 294, + 485 + ], + "spans": [ + { + "bbox": [ + 56, + 442, + 294, + 485 + ], + "type": "text", + "content": "[58] Xiangpeng Yang, Linchao Zhu, Xiaohan Wang, and Yi Yang. Dgl: Dynamic global-local prompt tuning for text-video retrieval. Proceedings of the AAAI Conference on Artificial Intelligence, 2024. 1, 2, 3, 5, 6, 12" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 56, + 487, + 294, + 518 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 487, + 294, + 518 + ], + "spans": [ + { + "bbox": [ + 56, + 487, + 294, + 518 + ], + "type": "text", + "content": "[59] Bowen Zhang, Hexiang Hu, and Fei Sha. Cross-modal and hierarchical modeling of video and text. In Proceedings of the European Conference on Computer Vision, 2018. 12" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 56, + 520, + 294, + 542 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 520, + 294, + 542 + ], + "spans": [ + { + "bbox": [ + 56, + 520, + 294, + 542 + ], + "type": "text", + "content": "[60] Yuanhan Zhang, Kaiyang Zhou, and Ziwei Liu. Neural prompt search. arXiv preprint arXiv:2206.04673, 2022. 2" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 56, + 544, + 294, + 586 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 544, + 294, + 586 + ], + "spans": [ + { + "bbox": [ + 56, + 544, + 294, + 586 + ], + "type": "text", + "content": "[61] Kaiyang Zhou, Jingkang Yang, Chen Change Loy, and Ziwei Liu. Conditional prompt learning for vision-language models. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition, 2022. 2" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 56, + 588, + 294, + 620 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 588, + 294, + 620 + ], + "spans": [ + { + "bbox": [ + 56, + 588, + 294, + 620 + ], + "type": "text", + "content": "[62] Kaiyang Zhou, Jingkang Yang, Chen Change Loy, and Ziwei Liu. Learning to prompt for vision-language models. International Journal of Computer Vision, 2022. 2" + } + ] + } + ], + "index": 11 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "text", + "content": "19712" + } + ] + } + ], + "index": 13 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 10 + } + ], + "_backend": "vlm", + "_version_name": "2.6.4" +} \ No newline at end of file diff --git a/2025/Discovering Fine-Grained Visual-Concept Relations by Disentangled Optimal Transport Concept Bottleneck Models/353387ac-1b53-4c3c-ac08-91f95db61c80_content_list.json b/2025/Discovering Fine-Grained Visual-Concept Relations by Disentangled Optimal Transport Concept Bottleneck Models/353387ac-1b53-4c3c-ac08-91f95db61c80_content_list.json new file mode 100644 index 0000000000000000000000000000000000000000..090f0920825978fadca4a925269b5ca768c940c9 --- /dev/null +++ b/2025/Discovering Fine-Grained Visual-Concept Relations by Disentangled Optimal Transport Concept Bottleneck Models/353387ac-1b53-4c3c-ac08-91f95db61c80_content_list.json @@ -0,0 +1,1677 @@ +[ + { + "type": "text", + "text": "Discovering Fine-Grained Visual-Concept Relations by Disentangled Optimal Transport Concept Bottleneck Models", + "text_level": 1, + "bbox": [ + 109, + 128, + 888, + 176 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Yan Xie $^{1*}$ Zequn Zeng $^{1*}$ Hao Zhang $^{1\\dagger}$", + "bbox": [ + 341, + 202, + 661, + 220 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Yucheng Ding $^{1}$ Yi Wang $^{1}$ Zhengjue Wang $^{2}$ Bo Chen $^{1}$ Hongwei Liu $^{1}$", + "bbox": [ + 197, + 220, + 795, + 239 + ], + "page_idx": 0 + }, + { + "type": "list", + "sub_type": "text", + "list_items": [ + "$^{1}$ National Key Laboratory of Radar Signal Processing, Xidian University, Xi'an, 710071, China", + "$^{2}$ State Key Laboratory of Integrated Service Networks, Xidian University, Xi'an, 710071, China" + ], + "bbox": [ + 120, + 239, + 875, + 273 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "{yanxie0904, zzequn99, zhanghao_xidian}@163.com, bchen@mail.xidian.edu.cn", + "bbox": [ + 176, + 277, + 818, + 292 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Abstract", + "text_level": 1, + "bbox": [ + 246, + 325, + 326, + 342 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Concept Bottleneck Models (CBMs) try to make the decision-making process transparent by exploring an intermediate concept space between the input image and the output prediction. Existing CBMs just learn coarse-grained relations between the whole image and the concepts, less considering local image information, leading to two main drawbacks: i) they often produce spurious visual-concept relations, hence decreasing model reliability; and ii) though CBMs could explain the importance of every concept to the final prediction, it is still challenging to tell which visual region produces the prediction. To solve these problems, this paper proposes a Disentangled Optimal Transport CBM (DOT-CBM) framework to explore fine-grained visual-concept relations between local image patches and concepts. Specifically, we model the concept prediction process as a transportation problem between the patches and concepts, thereby achieving explicit fine-grained feature alignment. We also incorporate orthogonal projection losses within the modality to enhance local feature disentanglement. To further address the shortcut issues caused by statistical biases in the data, we utilize the visual saliency map and concept label statistics as transportation priors. Thus, DOT-CBM can visualize inversion heatmaps, provide more reliable concept predictions, and produce more accurate class predictions. Comprehensive experiments demonstrate that our proposed DOT-CBM achieves SOTA performance on several tasks, including image classification, local part detection and out-of-distribution generalization. Codes are available in supplementary material.", + "bbox": [ + 89, + 357, + 483, + 796 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "1. Introduction", + "text_level": 1, + "bbox": [ + 91, + 823, + 220, + 838 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Explainable artificial intelligence aims to allow human users to comprehend and trust the output created by ma", + "bbox": [ + 89, + 848, + 482, + 878 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/1edd78dc0d730e337b22051b57bb7c9bf8da58a82b39a181dce327c0fe149562.jpg", + "image_caption": [ + "(a) Brief comparison about the coarse-grained design of vanilla CBM and fine-grained design of our proposed DOT-CBM." + ], + "image_footnote": [], + "bbox": [ + 535, + 325, + 887, + 421 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/1a3a251f1f2573376d3ca7c861920e8d68e2f22ae6017dbcf9c8b667e6458e8b.jpg", + "image_caption": [ + "(b) Inversion heatmap visualization.", + "Figure 1. Comparison between vanilla CBM and our proposed DOT-CBM, including architecture design and inversion heatmap for concept predictions. (a) Due to the black-box mapping from images to concepts, Vanilla CBM needs Grad-CAM [37] techniques to locate concept predictions back to image space while DOT-CBM can provide an explicit inversion heatmap visualization. (b) Due to lack of fine-grained alignment, vanilla CBM produces spurious correlations in three levels of granularity (part-background, part-object and part-part level) that mislocalizes the local concept to the background, the whole object, and incorrect local region." + ], + "image_footnote": [], + "bbox": [ + 532, + 459, + 883, + 590 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "chine learning algorithms [6, 14]. Compared with post-hoc explanation methods [37] for black-box models, designing explicit interpretability during modeling and learning processes is beneficial to obtaining more natural explanations [8].", + "bbox": [ + 511, + 777, + 906, + 853 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "The Concept Bottleneck Model (CBM) [17], a representative interpretable image classification model, first predicts a pre-defined set of concepts from images and then", + "bbox": [ + 511, + 854, + 906, + 900 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "CVF", + "bbox": [ + 106, + 2, + 181, + 42 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore.", + "bbox": [ + 236, + 0, + 810, + 46 + ], + "page_idx": 0 + }, + { + "type": "page_footnote", + "text": "*Equal contribution. †Corresponding authors", + "bbox": [ + 107, + 886, + 372, + 900 + ], + "page_idx": 0 + }, + { + "type": "page_number", + "text": "30199", + "bbox": [ + 478, + 944, + 519, + 957 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "maps these concept descriptions to the final output class. Therefore, CBM is good at interpreting the relationship between the input image and output class predictions via intermediate human-understandable concepts. However, vanilla CBM brings these good properties at the expense of classification performance [27]. For this problem, a series of CBM variants are proposed from two lines, construct a more comprehensive and discriminative concept bank [31, 45, 46], and sophisticated architecture design [29, 38, 39] for better concept representation learning.", + "bbox": [ + 89, + 90, + 480, + 242 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Although these variants narrow the performance gap between CBMs and their black-box counterparts, their modeling method between images and concepts still has two disadvantages, as shown in Fig. 1a. Firstly, they predict multiple local concepts (i.e. local attributes like \"head\") from a global image feature, which only learns a coarse-grained relationship between the full image and all concepts. Secondly, the interpretability of prevailing CBMs mainly comes from the linear concept-class relations. The mapping from the original image to all concepts is still modeled as the black-box neural networks, which still need some post-hoc analysis [26, 27] to visualize the learned visual-concept relations. As shown in Fig. 1b, these two disadvantages often bring some visual-concept spurious correlations, where many concept predictions are frequently mis-localized in the images. These spurious correlations inherently undermine the interpretability and generalizing capabilities of CBMs.", + "bbox": [ + 91, + 244, + 483, + 517 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Empirically, we categorized these visual-concept spurious correlations into three levels of granularity: i) part-background spurious correlation. This refers to a misattribution of the concept predictions of local parts to the background. ii) part-object spurious correlation. The concepts of local parts are wrongly associated with the entire object without further distinctions. iii) part-part spurious correlation. In this case, the local concepts are mislocalized to incorrect image regions. We attribute these spurious correlation problems to two key factors. The first one is the lack of fine-grained alignment between the local image regions and corresponding textual concepts, i.e. the fine-grained visual-concept relations. The second one is the overfitting of the data bias such as the statistically high co-occurrence between class and background and between concept and concept, yielding confusion to learn the difference between them. Besides, it is intuitive that the explicit interpretability modeling for the concept predictions from the images is conceivably effective in addressing these visual-concept spurious correlations.", + "bbox": [ + 89, + 520, + 483, + 821 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "To this end, we propose a novel Disentangled Optimal Transport Concept Bottleneck Models, namely DOT-CBM. Compared with current CBMs, our proposed DOT-CBM can provide holistic explicit interpretability on not only the concept-class relations but also the visual region", + "bbox": [ + 89, + 824, + 483, + 902 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "concept relations. Specifically, we introduce a disentangled optimal transport (DOT) framework to ensure the intramodality disentanglement and inter-modality alignment between the image patches and textual concepts. The transport assignment solution of OT framework, is an explicit correlation matrix between image patches and pre-defined concepts, facilitating bidirectional outputs including concept predictions on concept space and concept inversion mask on the image space. Furthermore, to tackle the overfitting to the data bias, we introduce two discrete prior distributions on the patch set and the concept set to penalize the incorrect shortcut alignment between patches and concepts. In summary, our contributions are as follows:", + "bbox": [ + 511, + 90, + 903, + 286 + ], + "page_idx": 1 + }, + { + "type": "list", + "sub_type": "text", + "list_items": [ + "- We empirically investigate the visual-concept spurious correlation phenomenon and pinpoint the two key factors of these limitations, i.e. lack of fine-grained visual-concept alignment and overfitting to the data bias.", + "- We propose a novel disentangled optimal transport framework to model the fine-grained visual-concept alignment and address the overfitting problems. The transport design between image patches and textual concepts provides explicit relations between image and concept predictions, which can produce concept predictions on the concept spaces and concept inversion mask on the image space.", + "- Extensive experiments on image classification, part detection, and out-of-distribution generalization benchmarks demonstrate our proposed DOT-CBM can achieve significant improvements over current state-of-the-art CBMs on classification accuracy, interpretability and generalizing capabilities to new distributions." + ], + "bbox": [ + 511, + 287, + 906, + 544 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "2. Related Work", + "text_level": 1, + "bbox": [ + 513, + 558, + 656, + 574 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Concept Bottleneck Models are a family of models [17] that first predict concepts from input images, then use these concepts to predict a downstream label. This design allows CBMs to (1) provide concept-based explanations via their predicted concepts, and (2) improve their test performance when deployed with experts via concept interventions [2, 11, 40]. Similar to previous interpretable models, CBMs also face the issue of balancing task accuracy and interoperability [10, 27]. Recent works [31, 46] has leveraged extra knowledge graph [38, 46], expert rules [18] or Large Language Models (LLMs) [22, 31, 38, 45, 47] to overcome the limitations of manually defined concept sets and has utilized pre-trained Vision-Language Models (VLMs) [24, 42, 45, 49, 50] to obtain concept predictions. These advancements have elevated the performance of CBMs to levels approaching those of black-box models.", + "bbox": [ + 511, + 583, + 906, + 824 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "The interpretability of prevailing CBMs primarily arises from linear relationships between concepts and classes. However, the mapping from the image to the concepts is still handled by black-box neural networks, requiring post-hoc analysis for visualization of the learned visual-concept", + "bbox": [ + 511, + 825, + 908, + 902 + ], + "page_idx": 1 + }, + { + "type": "page_number", + "text": "30200", + "bbox": [ + 478, + 944, + 519, + 957 + ], + "page_idx": 1 + }, + { + "type": "image", + "img_path": "images/87885aaab748a7eea38412e2b1636bcedb07341b38479ca7175d9dae9818b4f6.jpg", + "image_caption": [ + "Figure 2. Overview of our proposed DOT-CBM. The overall CBM framework proceeds from left to right. In the first stage, the model transforms input images into concept activation vectors. In forward pass, local embeddings from a pre-trained Vision Transformer (ViT) and a text encoder are processed through learnable adapters to generate two feature sets (Sec. 3.2). Two orthogonal projection losses are applied to both feature sets to constrain Adapter training. An Optimal Transport (OT) algorithm optimizes the Assignment Matrix, representing the explicit correlation between concepts and image patches. Concept activation values are derived by combining this matrix with a cost matrix, and the loss is supervised by concept labels for training (Sec. 3.3). To address data bias, we use the Saliency Map from the pre-trained ViT for rough foreground-background separation as the visual prior in OT. A Conditioned Co-occurrence Matrix, based on frequency statistics, serves as the prior for concepts, enhancing the model's ability to distinguish co-occurring concepts (Sec. 3.4). In the second stage, consistent with the vanilla CBM framework, concept activation values are used to predict final class labels via a linear classification network, which is trained using class labels (Sec. 3.5)." + ], + "image_footnote": [], + "bbox": [ + 96, + 89, + 903, + 311 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "relations. This lack of transparency often leads to visual-concept spurious correlations, where concept predictions are frequently mis-localized in the image. We propose a disentangled OT to address these limitations.", + "bbox": [ + 89, + 484, + 482, + 545 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Optimal Transport aims to find the most efficient way to map one distribution onto another by minimizing a transportation cost. In the context of machine learning, OT is widely applied in a variety of domains, including unsupervised domain adaptation [36], label representation [12, 52], cross-modal semantics [3, 4, 20, 21, 23, 25], and beyond.", + "bbox": [ + 89, + 546, + 482, + 636 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "In this work, we further extend OT to a disentangled OT framework, to enforce the fine-grained alignment between image patches and their corresponding concepts while also ensuring disentanglement within each modality.", + "bbox": [ + 89, + 636, + 483, + 696 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3. Method", + "text_level": 1, + "bbox": [ + 89, + 709, + 181, + 724 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "In this section, we begin with a brief review of the CBM problem setup in Sec. 3.1 including two stages of optimization, i.e. patches to concepts $(f_{\\mathcal{X}\\rightarrow \\mathcal{C}})$ and concepts to classes $(f_{\\mathcal{C}\\rightarrow \\mathcal{Y}})$ . In the first stage, to address the limited interpretation of previous CBMs in the patches-to-concepts process, we introduce a disentangled optimal transport alignment framework between patches and concepts. Concretely, in Sec. 3.2, we propose two orthogonal projection losses to disentangle the patch features and concept features. In Sec. 3.3, we introduce an optimal transport (OT) based framework to model the complex relationship", + "bbox": [ + 89, + 734, + 482, + 901 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "between patches and concepts, which can predict concepts from patches and meanwhile inverse the concept to patches to provide explanations of concept prediction. In Sec. 3.4, we integrate two prior distribution modeling to address the data bias problem. In the second stage, in Sec. 3.5, we follow vanilla CBMs and predict classes from concepts. Overall, our proposed DOT-CBM achieves holistic interpretability whether on intermediate concept predictions or on final class predictions. The whole pipeline is shown in Fig. 2.", + "bbox": [ + 511, + 484, + 906, + 621 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3.1. Problem Formulation", + "text_level": 1, + "bbox": [ + 511, + 633, + 718, + 648 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Traditional black-box visual models aim to fit the conditional distribution $P(y|x)$ given an a dataset $\\mathcal{D} = \\{x_i, y_i\\}_{i=1}^N$ , where $x_i \\in \\mathcal{X}, y_i \\in \\mathcal{Y}$ are images and class labels, respectively. Though effective, they still lack transparency to understand the underlying decision-making process to drive the final predictions.", + "bbox": [ + 511, + 657, + 905, + 748 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Concept Bottleneck Models (CBMs) [17] offer an interpretable solution by first predicting human-interpretable high-level concepts before making the final prediction. Formally, the dataset is extended to $\\mathcal{D} = \\{x_i,y_i,\\pmb {a}_i\\}_{i = 1}^N$ , where $\\pmb {a}_i = \\{a_{ij}|a_{ij}\\in \\{0,1\\} \\}_{j = 1}^M$ represents a binary concept activation label. $a_{ij} = 1$ indicates the presence of a specific high-level concept $c_{j}$ in image $x_{i}$ , while $a_{ij} = 0$ means the absence. $c_{j}$ is concept from the pre-defined concept set $\\mathcal{C} = \\{c_j\\}_{j = 1}^M$ . The optimization process of CBMs can be divided into two-stage [26, 35]: I) they first leverage a con", + "bbox": [ + 511, + 750, + 906, + 901 + ], + "page_idx": 2 + }, + { + "type": "page_number", + "text": "30201", + "bbox": [ + 478, + 944, + 517, + 955 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "cept predictor $f_{\\mathcal{X}\\to \\mathcal{C}}:\\mathbb{R}^{H\\times W\\times C}\\to \\mathbb{R}^M$ to map the input data $x$ to the concept activation vector $\\pmb {a}\\in \\mathbb{R}^{M}$ . 2) Subsequently, a target predictor $f_{\\mathcal{C}\\rightarrow \\mathcal{Y}}:\\mathbb{R}^{M}\\to \\mathbb{R}^{N_{y}}$ maps the concept activation vector $\\pmb{a}$ to the final label $y$ . This structure enhances the interpretability of the model, as the intermediate concept predictions provide interpretability into the model's final decision. However, the interpretability from image to concept is un-explored.", + "bbox": [ + 89, + 89, + 483, + 212 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "3.2. Disentanglement on Local representations", + "text_level": 1, + "bbox": [ + 89, + 220, + 449, + 238 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Disentangled visual patch features $\\widetilde{\\mathcal{P}}$ . As previously discussed, current CBMs typically associate multiple local concepts with a global image feature, which leads to coarse-grained correspondences. In this paper, we adopt the pretrained vision transformer (ViT) [1] model, which has been approved to exhibit better spatial location information extraction than CNN [34], to extract local patch features to represent the local image regions.", + "bbox": [ + 89, + 244, + 483, + 366 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Specifically, as shown in the left part of Fig. 2, ViT uniformly divides the full image $x \\in \\mathbb{R}^{H \\times W \\times C}$ into $N$ non-overlapping patches $\\{I_{pi} \\in \\mathbb{R}^{P_h \\times P_w}\\}_{i=1}^N$ . After linearly projecting these patches into a $d$ -dimension embedding space and flattening all patch embeddings into a sequence, we can get an initialized patch embedding matrix $X_P \\in \\mathbb{R}^{N \\times d}$ . To extract the global image feature, ViT concatenates a learnable special token [CLS] with $X_P$ . Besides, in order to preserve the spatial information of images, ViT adds a learnable position embedding $X_{pos} \\in \\mathbb{R}^{(N+1) \\times d}$ and we get the input embeddings $X$ for ViT:", + "bbox": [ + 89, + 367, + 483, + 534 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\nX _ {P} = \\left[ I _ {p 1} \\cdot W _ {p}, \\dots , I _ {p N} \\cdot W _ {p} \\right], \\tag {1}\n$$\n", + "text_format": "latex", + "bbox": [ + 181, + 545, + 482, + 564 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\nX = \\left[ \\left[ \\mathrm {C L S} \\right], X _ {P} \\right] + X _ {\\text {p o s}}, \\tag {2}\n$$\n", + "text_format": "latex", + "bbox": [ + 183, + 565, + 480, + 582 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "where $W_{p}\\in \\mathbb{R}^{(P_{h}\\cdot P_{w}\\cdot C)\\times d}$ is the patch embedding projection weights. Then, the pre-trained ViT $E_I$ to map the input embeddings $X$ into the global feature $g_{I}$ , and patch features $\\mathcal{P} = \\{p_i\\}_{i = 1}^N$ :", + "bbox": [ + 89, + 593, + 483, + 657 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n[ g _ {I}, \\mathcal {P} ] = E _ {I} (X), \\tag {3}\n$$\n", + "text_format": "latex", + "bbox": [ + 225, + 667, + 482, + 686 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "We empirically find that the patch features within the object regions are highly similar which is good for object representation, but bad for distinguishing local parts within the object region. Thus, we fit a learnable visual adapter $A_v$ on the $\\mathcal{P}$ and map the patch features within the foreground region to a disentangled unit hypersphere by introducing a visual orthogonal projection (VOP) regularizer as:", + "bbox": [ + 89, + 696, + 483, + 803 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\nS _ {I} = g _ {I} \\cdot P, \\tag {4}\n$$\n", + "text_format": "latex", + "bbox": [ + 243, + 814, + 482, + 830 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\langle A _ {v} (p _ {i}), A _ {v} (p _ {j}) \\rangle = \\frac {A _ {v} (p _ {i}) \\cdot A _ {v} (p _ {j})}{\\| A _ {v} (p _ {i}) \\| \\| A _ {v} (p _ {j}) \\|}, \\tag {5}\n$$\n", + "text_format": "latex", + "bbox": [ + 148, + 834, + 482, + 867 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {L} _ {\\mathrm {V O P}} = \\underset {\\left(p _ {i}, p _ {j}\\right) \\sim \\mathcal {P}} {\\mathbb {E}} \\left[ S _ {I} (i) \\cdot S _ {I} (j) \\cdot \\left\\langle A _ {v} \\left(p _ {i}\\right), A _ {v} \\left(p _ {j}\\right) \\right\\rangle \\right] \\tag {6}\n$$\n", + "text_format": "latex", + "bbox": [ + 107, + 872, + 482, + 897 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "where $S_{I}$ is the saliency map to distinguish foreground and background, $\\langle \\cdot ,\\cdot \\rangle$ is the cosine similarity. Finally, we get the disentangled patch features $\\widetilde{\\mathcal{P}} = \\{\\widetilde{p_i} = A_v(p_i)\\}_{i = 1}^N$ Disentangled textual concept features $\\widetilde{\\mathcal{T}}$ . For concept feature extraction, following [22, 29, 45, 47], we utilize a pre-trained text encoder $E_{T}$ to compute concept embeddings $\\mathcal{T} = \\{t_j = E_T(c_j)\\}_{j = 1}^M$ . Similarly, we fit a learnable textual adapter $A_{t}$ and introduce a textual orthogonal projection (TOP) regularizer to disentangle the concept features in a unit hyper-sphere, as:", + "bbox": [ + 511, + 90, + 906, + 247 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {L} _ {\\mathrm {T O P}} = \\underset {(t _ {i}, t _ {j}) \\sim \\mathcal {T}} {\\mathbb {E}} [ \\langle A _ {t} (t _ {i}), A _ {t} (t _ {j}) \\rangle ], \\tag {7}\n$$\n", + "text_format": "latex", + "bbox": [ + 591, + 255, + 906, + 281 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Consequently, we get the disentangled concept features $\\widetilde{\\mathcal{T}} = \\{\\widetilde{t_j} = A_t(t_j)\\}_{j = 1}^M.$", + "bbox": [ + 513, + 290, + 906, + 321 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "3.3. Patch-concept alignment via optimal transport", + "text_level": 1, + "bbox": [ + 511, + 328, + 906, + 345 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Now, we have the disentangled patch feature set $\\widetilde{\\mathcal{P}}$ and disentangled concept feature set $\\widetilde{\\mathcal{T}}$ , we need to model the relationship $\\mathbf{Q}$ between these two sets and derive the concept activation vector $\\hat{\\pmb{a}}$ based on the input image. As a complex many-to-many alignment problem, we propose an optimal transport (OT) based binary cross-entropy loss to tackle these problems. Our goal is to correctly align the patch features and concept features with the same semantics.", + "bbox": [ + 511, + 352, + 905, + 473 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Assume that these two sets $\\widetilde{\\mathcal{P}}$ and $\\widetilde{\\mathcal{T}}$ follow two discrete probability distributions $\\Theta \\in \\mathbb{R}^N$ and $\\Gamma \\in \\mathbb{R}^M$ on the $d$ -dimension embedding space, respectively:", + "bbox": [ + 511, + 474, + 905, + 520 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\boldsymbol {\\Theta} = \\sum_ {i = 1} ^ {N} \\theta_ {i} \\delta_ {p _ {i}}, \\quad \\boldsymbol {\\Gamma} = \\sum_ {j = 1} ^ {M} \\gamma_ {j} \\delta_ {c _ {j}}, \\tag {8}\n$$\n", + "text_format": "latex", + "bbox": [ + 596, + 529, + 905, + 571 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "where $\\pmb{\\theta} \\in \\Sigma^{N}$ and $\\pmb{\\gamma} \\in \\Sigma^{M}$ , the simplex of $\\mathbb{R}^N$ and $\\mathbb{R}^{M}$ , denote the probability values of the discrete states with satisfaction that $\\sum_{i=1}^{N} \\theta_{i} = 1$ and $\\sum_{j=1}^{M} \\gamma_{j} = 1$ . $\\delta_{e}$ refers to a point mass located at coordinate $e \\in \\mathbb{R}^{d}$ . The prior modeling of $\\pmb{\\theta}$ and $\\pmb{\\gamma}$ are shown in the Sec. 3.4.", + "bbox": [ + 511, + 582, + 905, + 660 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "To get the transport assignment $\\mathbf{Q}$ between $\\Theta$ and $\\Gamma$ , the objective of OT is formulated as the optimization problem:", + "bbox": [ + 511, + 661, + 905, + 691 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\min _ {\\mathbf {Q} \\in \\mathcal {Q}} \\sum_ {i, j} q _ {i j} d _ {i j} - \\varepsilon H (\\mathbf {Q}) \\tag {9}\n$$\n", + "text_format": "latex", + "bbox": [ + 527, + 700, + 905, + 734 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\text {s . t .} \\mathcal {Q} = \\left\\{\\mathbf {Q} \\in \\mathbb {R} _ {+} ^ {N \\times M} \\mid \\mathbf {Q} \\mathbf {1} _ {M} = \\boldsymbol {\\theta}, \\mathbf {Q} ^ {\\top} \\mathbf {1} _ {N} = \\boldsymbol {\\gamma} \\right\\}.\n$$\n", + "text_format": "latex", + "bbox": [ + 529, + 736, + 870, + 756 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "where $\\mathcal{Q}$ is the set of doubly stochastic matrices $\\mathbf{Q} = [q_{ij}]_{N\\times M}$ , denotes the corresponding transport assignment where $q_{ij}$ represents the probabilities of aligning $p_i$ with $t_j$ . $\\mathbf{1}^M$ is the $M$ -dimensional vector of ones. $d_{ij} = d(p_i,t_j) = (1 - \\langle p_i,t_j\\rangle)\\geq 0$ is the transport cost between the two points $p_i$ and $s_j$ and we use the cosine distance to compute cost matrix $\\mathbf{D}$ . The optimal $\\mathbf{Q}^*$ of (9) has a simple normalized exponential matrix solution by Sinkhorn fixed point iterations [7],", + "bbox": [ + 511, + 763, + 906, + 901 + ], + "page_idx": 3 + }, + { + "type": "footer", + "text": "30202", + "bbox": [ + 478, + 944, + 519, + 955 + ], + "page_idx": 3 + }, + { + "type": "image", + "img_path": "images/ea76a7dba2a4f3ab732113d26928c8cbb74c4ad73f1fd3de7a85d95713606a83.jpg", + "image_caption": [ + "Figure 3. Conditioned co-occurrence matrix of concept labels." + ], + "image_footnote": [], + "bbox": [ + 150, + 88, + 428, + 270 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\mathbf {Q} ^ {*} = \\operatorname {S i n k h o r n} (\\mathbf {D}, \\boldsymbol {\\theta}, \\gamma), \\tag {10}\n$$\n", + "text_format": "latex", + "bbox": [ + 196, + 325, + 480, + 340 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "After obtaining the optimal transport assignment $\\mathbf{Q}^*$ , the OT distance between the two sets is expressed as:", + "bbox": [ + 89, + 349, + 482, + 378 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\nd _ {O T} = \\sum_ {i, j} q _ {i j} d _ {i j} \\tag {11}\n$$\n", + "text_format": "latex", + "bbox": [ + 225, + 391, + 482, + 424 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "For the concept embedding $t_j$ , we only consider the minimum distance $d_{OT_j} = \\min_i(q_{ij}d_{ij})$ between this concept embedding $t_j$ and all patch embeddings to facilitate the concept activation value prediction $(x\\rightarrow \\mathcal{C})$ , the predicted concept activation vector $\\hat{\\pmb{a}}$ are stated as:", + "bbox": [ + 89, + 435, + 483, + 510 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\hat {\\boldsymbol {a}} = \\left\\{\\hat {a} _ {j} \\mid \\hat {a} _ {j} = \\sigma \\left(1 - d _ {O T _ {j}}\\right) \\right\\} _ {j = 1} ^ {M}. \\tag {12}\n$$\n", + "text_format": "latex", + "bbox": [ + 173, + 521, + 480, + 542 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "where $\\sigma$ is the sigmoid function. Moreover, we can inverse specific concepts to the image patches based on $\\mathbf{Q}^*$ , and get the inversion mask $m_j$ of concept $c_j$ , as:", + "bbox": [ + 89, + 551, + 483, + 598 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\nm _ {j} = \\left[ q _ {1 j}, q _ {2 j}, \\dots , q _ {N j} \\right]. \\tag {13}\n$$\n", + "text_format": "latex", + "bbox": [ + 202, + 609, + 480, + 627 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "By utilizing OT distance between patches and concepts as the patch-concept transport, our OT-based binary cross-entropy loss $\\mathcal{L}_{\\mathrm{concept}}$ are as follows:", + "bbox": [ + 89, + 638, + 482, + 683 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {L} _ {\\text {c o n c e p t}} = - \\mathbb {E} [ \\boldsymbol {a} \\log (\\hat {\\boldsymbol {a}}) + (1 - \\boldsymbol {a}) \\log (1 - \\hat {\\boldsymbol {a}}) ] \\tag {14}\n$$\n", + "text_format": "latex", + "bbox": [ + 117, + 695, + 482, + 712 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Now, we can get the total objective of stage 1 $(f_{\\mathcal{X}\\to \\mathcal{C}})$ , as:", + "bbox": [ + 89, + 723, + 473, + 739 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {L} _ {f, x \\rightarrow c} = \\lambda_ {1} \\mathcal {L} _ {\\mathrm {V O P}} + \\lambda_ {2} \\mathcal {L} _ {\\mathrm {T O P}} + \\lambda_ {3} \\mathcal {L} _ {\\text {c o n c e p t}}, \\tag {15}\n$$\n", + "text_format": "latex", + "bbox": [ + 130, + 751, + 482, + 768 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "where $\\lambda_1, \\lambda_2, \\lambda_3$ are the hyperparameters.", + "bbox": [ + 89, + 779, + 369, + 794 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "The obtained assignment $\\mathbf{Q}^{*} = [q_{i,j}]_{N\\times M}$ indicates the relationship between patches and concepts, where $q_{i,j}$ represent correlation between patch $p_i$ and concept $c_{j}$ . Consequently, based on $\\mathbf{Q}^*$ , we can obtain concept activation $\\pmb{a}$ (Eq.(12)) based on the image features $(x\\to \\mathcal{C})$ and meanwhile inverse the concept to the original image patches (Eq.(13)) $(\\mathcal{C}\\rightarrow x)$ to provide explicit interpretation.", + "bbox": [ + 89, + 795, + 483, + 901 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "3.4. Prior distributions Modeling", + "text_level": 1, + "bbox": [ + 513, + 90, + 769, + 107 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "The prior distributions of $\\theta$ and $\\gamma$ of OT are typically uniform distributions, which is sub-optimal for our task. Considering the part-background spurious correlation and part-part spurious correlation problem mentioned in Fig. 1b, we propose two prior distribution modeling in this section to address these limitations, respectively.", + "bbox": [ + 511, + 113, + 903, + 204 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Patch set prior $\\theta$ . We attribute the part-background spurious correlation phenomenon to the high co-occurrence of some classes and backgrounds, such as \"seabird\" and \"water\". To solve this biased association between object concepts and the background, we use saliency map $S_I$ from Eq. (4) to encourage the transports between the patch features within foreground regions and concepts and penalize the transports of background regions. Specifically, the prior distribution of patch set $\\theta$ are formulated as:", + "bbox": [ + 511, + 205, + 905, + 339 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\boldsymbol {\\theta} = \\operatorname {S o f t m a x} \\left(S _ {I}\\right), \\tag {16}\n$$\n", + "text_format": "latex", + "bbox": [ + 643, + 347, + 903, + 363 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Concept set prior $\\gamma$ . As illustrated in Fig. 3, some concepts are highly co-occurrent. These concept label biases lead to the model's confusion between these concepts, yielding the part-part spurious correlation phenomenon. To address this problem, we employ the statistical co-occurrence matrix of the training dataset to facilitate the prior distribution modeling of $\\gamma$ , aiming to penalize those concepts with high co-occurrence with other concepts.", + "bbox": [ + 511, + 369, + 905, + 489 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Specifically, for a dataset with $M$ concepts and $K$ classes, class-concept label form a matrix $L = [l_{i,j}]_{K\\times M}$ where $l_{i,j}$ represents the activation value of the $j$ -th concept in the concept label for the $i$ -th class. First, we count the co-occurrence frequencies of each pair of concepts. Then, we normalize these counts by the total number of times each concept is activated, resulting in the conditional co-occurrence matrix. The conditional co-occurrence vector $r_i$ is calculated as follows:", + "bbox": [ + 511, + 489, + 905, + 625 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\nr _ {i} = \\sum_ {j = 1} ^ {M} \\left(\\frac {\\sum_ {k = 1} ^ {K} L _ {k i} \\cdot L _ {k j}}{\\sum_ {k = 1} ^ {K} L _ {k i}}\\right) \\tag {17}\n$$\n", + "text_format": "latex", + "bbox": [ + 607, + 630, + 903, + 672 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "The concept set prior distribution $\\gamma$ is then derived from the conditional co-occurrence vector as follows:", + "bbox": [ + 511, + 676, + 903, + 705 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\gamma = 1 - \\operatorname {S o f t m a x} \\left(r _ {i}\\right) \\tag {18}\n$$\n", + "text_format": "latex", + "bbox": [ + 635, + 715, + 903, + 731 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "This prior modeling ensures that those concepts with high co-occurrence with other concepts will be penalized.", + "bbox": [ + 511, + 737, + 903, + 768 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "3.5. Concept $\\rightarrow$ class prediction", + "text_level": 1, + "bbox": [ + 513, + 773, + 758, + 790 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Now, we already finished the first stage which obtains the intermediate concept activation predictions $\\pmb{a}$ . In the second stage, our goal is to learn the association between concept and final class prediction similar to vanilla CBM. Specifically, we can map the $\\pmb{a}$ to final output $\\hat{y}$ , as:", + "bbox": [ + 511, + 797, + 905, + 873 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\hat {y} = W _ {F} \\cdot \\boldsymbol {a}, \\tag {19}\n$$\n", + "text_format": "latex", + "bbox": [ + 663, + 881, + 903, + 896 + ], + "page_idx": 4 + }, + { + "type": "page_number", + "text": "30203", + "bbox": [ + 478, + 944, + 517, + 955 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "The classification objective of stage 2 $(f_{\\mathcal{C}\\to \\mathcal{Y}})$ is the cross-entropy loss $\\mathcal{L}_{\\mathrm{CE}}(y,\\hat{y})$", + "bbox": [ + 89, + 90, + 485, + 122 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "4. Experiment", + "text_level": 1, + "bbox": [ + 89, + 133, + 215, + 152 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "To evaluate the effectiveness of our proposed DOT-CBM, we conduct experiments on the image classification task and part detection task. To further demonstrate our method's abilities to reduce spurious correlations, we also conduct experiments on out-of-distribution generalization tasks. For the concept-to-class linear layer, we omit explicit sparse regularization (e.g., L1 penalties) during training. Despite this, the model spontaneously exhibits significant sparsity in the weight matrix due to two orthogonal losses. Comprehensive training details, training efficiency results, quantitative evaluations of sparsity and additional experiments are listed in the Appendix.", + "bbox": [ + 89, + 159, + 483, + 340 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "4.1. Datasets and Metrics", + "text_level": 1, + "bbox": [ + 89, + 348, + 290, + 364 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "For the image classification task. We evaluate our method on the CUB [41] and AwA2 [43] datasets, which are widely adopted in the interpretable community. The CUB dataset contains 11,788 images of 200 bird subcategories, and we use the official train-test split for fine-grained image classification. The AwA2 dataset includes 37,322 images across 50 animal categories, and we follow the same train-test split as in [17]. Additionally, we conduct experiments on the larger and more complex ImageNet-1K[9] and CIFAR-100[19] datasets. On these classification datasets, we leverage the classification accuracy to measure the performance of our method. We also conducted experiments on a more complex dataset (Places365), where the complexity arises from discriminative content being image backgrounds. Quantitative and qualitative results are listed in Appendix.", + "bbox": [ + 89, + 372, + 482, + 613 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "For the part detection task. To evaluate the accuracy of the inversion mask produced by DOT-CBM, we evaluate our method on some part detection benchmarks, including CUB dataset, Part-ImageNet [13] and PASCAL-Part [5] datasets, which are widely used for part segmentation tasks. We also test our model on the RIVAL-10 dataset [28], which is specifically designed for evaluating model interpretability. On the CUB dataset, we obtain the bounding boxes for local parts that are commonly used in part detection tasks[44, 51]. The RIVAL-10 dataset consists of over 26K images retrieved from the ImageNet dataset according to the CIFAR-10 [19] categories, and it includes part attribute segmentation masks for the main subjects in the images. The Part-ImageNet and PASCAL-Part datasets are part segmentation datasets that provide fine-grained attribute segmentation masks. We follow [30] and preprocess these datasets by focusing on the animal data and merging overly fine-grained part attribute masks. The detailed data preprocessing procedures are described in the Appendix. We use the", + "bbox": [ + 89, + 614, + 482, + 901 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "bounding boxes of these masks as the ground truth of part detection. On these datasets, we use the $mAP_{0.5}$ metric to evaluate the precision of predicted concept inversion masks.", + "bbox": [ + 511, + 90, + 906, + 137 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "4.2. Implementation Details", + "text_level": 1, + "bbox": [ + 511, + 146, + 732, + 162 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "We employ the pre-trained DINOV2 ViT-L/14 [32] as the image encoder and pre-trained CLIP text encoder [33] as the text encoder. The feature dimension in both encoders is $d = 768$ . We resize all images to $H \\times W = 252 \\times 252$ as input resolution in both training and testing phases throughout all experiments. For both adapter $A_v$ and $A_t$ , we use a three-layer MLP with ReLU activation functions and a dropout rate of 0.4. The optimization of DOT-CBM is done by AdamW with an initialized learning rate of $1e - 4$ , and decreased every ten epochs by a factor of 0.5. We set weight decay of 10-5. Batch sizes are set as 128 with maximally 200 training epochs. All experiments are conducted on a single RTX 3090 GPU and our model is implemented in PyTorch. We set $\\lambda_1 = 0.15$ , $\\lambda_2 = 0.2$ , $\\lambda_3 = 0.8$ for CUB dataset. Hyperparameters setting on other datasets and the range ablation study are in the Appendix.", + "bbox": [ + 511, + 170, + 906, + 412 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "4.3. Main Results", + "text_level": 1, + "bbox": [ + 511, + 422, + 653, + 436 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Baselines. We compare our proposed DOT-CBM with several state-of-the-art CBM baselines from the literature. Vanilla-CBM [17] is the first CBM model, proposing the design of the intermediate concept layer. Vanilla-CBM introduces three training strategies, i.e. independent, sequential and joint training, in which the independent training can achieve the best interpretation [26]. Our two-stage optimization belongs to the independent training strategy. Additionally, we compare our performance with recent CBM variants including CEM [48], LaBo [45], SparseCBM [38], CoopCBM [39]. Due to the fact that pre-defined concepts and image backbone will significantly affect the final performance. For a fair comparison, we re-implement all baselines with the high-quality concepts produced by the CDL [47] and a pre-trained ViT backbone to extract image features, the same as our method. The performance of all re-implemented baselines exceeds the best performance reported in their papers.", + "bbox": [ + 511, + 446, + 908, + 718 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Image classification results. As shown in Table 1, we compare our proposed DOT-CBM with all baselines on four common image classification benchmarks. The classification results show that the DOT-CBM achieves superior performance on three benchmarks. Specifically, the performance improvements over the state-of-the-art (SOTA) are 0.91 points on ImageNet, 3.29 points on CUB, and 1.08 points on CIFAR100. Our method only lags behind CoopCBM by 0.25 points on the AwA2 dataset.", + "bbox": [ + 511, + 719, + 908, + 854 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Part detection results. A key property of our proposed DOT-CBM model is able to provide explicit explanations for the concept predictions, via the locate the concept to the", + "bbox": [ + 511, + 854, + 908, + 902 + ], + "page_idx": 5 + }, + { + "type": "page_number", + "text": "30204", + "bbox": [ + 478, + 944, + 519, + 957 + ], + "page_idx": 5 + }, + { + "type": "table", + "img_path": "images/2fb6c11fcb69f793cac2a8e72ab4e09ac7d847b3a45326b8cc983657e3f6155e.jpg", + "table_caption": [], + "table_footnote": [], + "table_body": "
MethodClassification Accuracy (↑)Part Detection mAP0.5 (↑)
ImageNetCUBCIFAR100AWA2PartImageNetCUBRIVALPASCAL-Parts
Vanilla-CBM [17]79.1778.3280.0493.1526.9427.0226.7617.32
CEM [48]81.2980.4781.2395.9229.1930.8630.8320.85
LaBo [45]82.9381.3084.1096.9238.2739.2840.3430.17
SparseCBM [38]82.8582.0784.7595.5639.8440.2340.6833.71
CoopCBM [39]82.7382.1084.6697.0841.1745.1643.8235.75
DOT-CBM83.84(+0.91)85.39(+3.29)85.83(+1.08)96.83(-0.25)50.12(+8.95)53.47(+8.31)50.93(+7.11)44.18(+8.43)
", + "bbox": [ + 91, + 88, + 908, + 246 + ], + "page_idx": 6 + }, + { + "type": "table", + "img_path": "images/5448c68cfb65eff58dcded3813da7d0a5836f418d1ce6398a436420e81a34413.jpg", + "table_caption": [ + "Table 1. Performance comparison of Classification Accuracy (%) and Part Detection $\\left( {mA}{P}_{0.5}\\right)$ ." + ], + "table_footnote": [], + "table_body": "
MethodCUBDogs
IDOODIDOOD
Vanilla-CBM [17]86.927.787.329.4
CEM [48]84.034.083.836.5
LaBo [45]85.339.485.941.7
Sparse-CBM [38]84.535.885.838.6
Coop-CBM [39]85.836.286.140.3
DOT-CBM w/o. prior86.536.987.040.8
DOT-CBM w. prior87.249.789.152.4
(+0.3)(+10.3)(+1.8)(+10.7)
", + "bbox": [ + 93, + 276, + 475, + 448 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Table 2. Performance comparison of OOD generalization.", + "bbox": [ + 114, + 450, + 457, + 463 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "original image space, i.e. the concept inversion mask derived from the transport assignment. This concept inversion mask in the image space indicates the most influential image region contributing to the concept predictions. By contrast, all previous CBMs, using a black-box neuron network to model the mapping from global image features to concept predictions, can only inverse the concept prediction to the image space via the post-hoc Grad-CAM. To evaluate the accuracy of concept inversion mask, we conduct experiments on the part detection datasets and use the groundtruth bounding boxes of the parts to assess the localization performance. To measure the similarity with groundtruths, we need to process our concept inversion mask output and heatmap outputs of baselines (derived from Grad-CAM) into rectangle boxes format. Specifically, we set the threshold at 0.6 of the maximum peak value output of the mask. The local regions above this threshold are considered as the corresponding regions of the concept in the original image. We take the bounding box that encloses this region as the predicted bounding box for the concept.", + "bbox": [ + 88, + 477, + 482, + 779 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "From Table. 1, we can observe that DOT-CBM significantly surpasses all baselines (average 8 points improvements), demonstrating the superiority of our design for fine-grained patch-concept alignment. The Coop-CBM achieves performance that is second only to ours due to the introduction of an orthogonality loss similar to our proposed textual orthogonal project $\\mathcal{L}_{\\mathrm{TOP}}$ , which further proves the validity of our proposed disentangled OT framework. Consequently,", + "bbox": [ + 89, + 779, + 482, + 901 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "DOT-CBM can inherently produce accurate and explicit explanations for intermediate concept predictions. We also conduct comparison with other ViT-appropriate attribution method. Results shown in Appendix.", + "bbox": [ + 511, + 277, + 906, + 339 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "4.4. Out-of-distribution generalization", + "text_level": 1, + "bbox": [ + 511, + 347, + 808, + 363 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "As discussed in the previous sections, there exists data bias in many datasets, leading to shortcut learning. For example, classifying a dog as a huskie based on the snowy background due to the high co-occurrence in the training set, which is analogous to the part-background spurious correlation discussed in Fig. 1b. To validate the effectiveness of our method with background shifts, we follow [39] to segment the foreground object in CUB[41] and Dogs[15] datasets using SAM [16] with bounding box annotations and add a colored background to all images. Each class is associated with a randomly generated color background with an $80\\%$ probability in the training set. The in-distribution (ID) test set contains images with a similar color background probability as the training set, while the color similarity of out-of-distribution (OOD) test set is reduced to $30\\%$ . Results are shown in Table 2. As we can see, DOT-CBM can achieve comparable results on ID set and surpass the performance of baselines on the OOD set by a large margin, demonstrating that DOT-CBM can effectively decrease the biased association between object concepts and background. To investigate whether the improvements of tackling part-background spurious correlation are derived from the patch set prior modeling as mentioned in Sec. 3.4, we conduct ablations for the patch set prior and demonstrate the efficacy of patch set prior $\\theta$ .", + "bbox": [ + 511, + 369, + 906, + 750 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "4.5. Ablation Studies", + "text_level": 1, + "bbox": [ + 511, + 756, + 679, + 771 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "In this section, we investigate the effects of each component in our framework. Ablation results are listed in Table 3. The baseline has the same trainable networks as DOT-CBM, including two adapters and a linear classification layer. The baseline replaces the OT assignment with a dot product of patch features and concept features.", + "bbox": [ + 511, + 779, + 906, + 869 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Effect of OT design (OT). For the comparison between the first row and second row in Table. 3, we can find the", + "bbox": [ + 511, + 869, + 906, + 898 + ], + "page_idx": 6 + }, + { + "type": "page_number", + "text": "30205", + "bbox": [ + 478, + 944, + 517, + 955 + ], + "page_idx": 6 + }, + { + "type": "table", + "img_path": "images/d2bbfdaf42f92d062a40b52d41705ee27235cfbf92af7caf2552997f2c329b90.jpg", + "table_caption": [], + "table_footnote": [], + "table_body": "
OTPriorLrDClassification Accuracy (↑)Part Detection mIOU (↑)
ImageNetCUBCIFAR100AWA2PartImageNetCUBRIVALPASCAL-Parts
XXX81.6080.9282.7396.070.350.410.370.28
✓XX82.0381.7582.9196.240.420.490.420.34
✓✓X83.2284.9285.4996.470.500.610.530.46
XX✓82.7382.1084.6697.080.470.580.490.41
✓✓✓83.8485.3985.8396.830.520.660.540.49
", + "bbox": [ + 102, + 88, + 890, + 209 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Table 3. Ablation studies on image classification and part detection. OT, Prior, and LrD denote the optimal transport design, prior distribution modeling, and local representation disentanglement.", + "bbox": [ + 89, + 210, + 903, + 239 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/1d71a38ad4c7dbfe9cb6d33181883381d0462270320f24079438f568e80b6584.jpg", + "image_caption": [ + "Figure 4. Qualitative results for the concept inversion. VanillaCBM and CoopCBM adopt Grad-CAM to provide a concept inversion heatmap, while our proposed DOT-CBM can generate the inversion mask on its own." + ], + "image_footnote": [], + "bbox": [ + 130, + 243, + 862, + 428 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "incorporation of OT distance, mentioned in Sec. 3.3, can slightly improve the classification accuracy and largely enhance the part detection performance. OT provides a better distance metric space for modeling the complex many-to-many correlation between concept features and patch features, compared to the direct dot product similarity used in the baseline model.", + "bbox": [ + 88, + 474, + 482, + 579 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Effect of prior distribution modeling (Prior). The OT design above, setting the two prior distributions $\\theta$ and $\\gamma$ as a uniform distribution. To mitigate the shortcut learning brought by high occurrence between concept and background and between concept and concept, we add the background penalty patch set prior and the co-occurrence penalty concept set prior, mentioned in Sec. 3.4. Comparing the second and third rows of Table. 3, it demonstrates that these two prior distributions can further improve the classification and interpretability.", + "bbox": [ + 88, + 580, + 482, + 733 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Effect of local representation disentanglement (LrD). After integrating the local representation disentanglement into the baseline (the fourth row of Table. 3), the improvements are larger than a single incorporation of OT distance (the second row of Table. 3). The overall DOT-CBM, with a combination of the OT, Prior, and LrD, yields more improvements than individual modules alone.", + "bbox": [ + 88, + 733, + 482, + 839 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Qualitative results. Fig. 4 has shown a qualitative comparison between our method and the baseline methods Vanilla-CBM and CoopCBM. It is observed that our method is able to more accurately locate the region of the", + "bbox": [ + 88, + 839, + 483, + 901 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "concept in the original image compared to the baseline methods. Meanwhile, We also present qualitative results by sequentially ablating each component of our framework. When the local representation disentanglement is ablated, the model's localization accuracy decreases to some extent. Further removal of the priors leads to the emergence of part-background spurious correlations. When the OT design is removed, the model produces incorrect localization results.", + "bbox": [ + 511, + 474, + 906, + 595 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "5. Conclusion.", + "text_level": 1, + "bbox": [ + 511, + 606, + 637, + 622 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "In this paper, we introduced the Disentangled Optimal Transport Concept Bottleneck Model (DOT-CBM) to address key limitations in existing Concept Bottleneck Models. While CBMs offer interpretable models by mapping image features to human-understandable concepts, they suffer from misalignment between the image and concept, leading to spurious correlations that hinder both interpretability and generalization. To overcome these challenges, DOT-CBM introduces a disentangled optimal transport framework that ensures fine-grained alignment between image patches and corresponding textual concepts, thereby providing a more robust and interpretable model. By employing a dual penalty system, we mitigate the overfitting problem caused by data bias. Empirical results demonstrate that DOT-CBM outperforms previous CBM variants in both classification accuracy and interpretability, offering an effective solution to the problem of visual-concept spurious correlation.", + "bbox": [ + 511, + 628, + 906, + 900 + ], + "page_idx": 7 + }, + { + "type": "page_number", + "text": "30206", + "bbox": [ + 478, + 944, + 519, + 957 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "6. Acknowledgement", + "text_level": 1, + "bbox": [ + 89, + 90, + 269, + 107 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "This work was supported in part by the National Natural Science Foundation of China under Grant U21B2006; in part by Shaanxi Youth Innovation Team Project; in part by the Fundamental Research Funds for the Central Universities QTZX24003 and QTZX22160; in part by the 111 Project under Grant B18039; Hao Zhang acknowledges the support of NSFC (62301384); Excellent Young Scientists Fund (Overseas); Foundation of National Key Laboratory of Radar Signal Processing under Grant JKW202308. Zhengjue Wang acknowledges the support of NSFC (62301407).", + "bbox": [ + 89, + 113, + 485, + 268 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "References", + "text_level": 1, + "bbox": [ + 91, + 292, + 187, + 308 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[1] Dosovitskiy Alexey. An image is worth 16x16 words: Transformers for image recognition at scale. arXiv preprint arXiv: 2010.11929, 2020.4", + "[2] Kushal Chauhan, Rishabh Tiwari, Jan Freyberg, Pradeep Shenoy, and Krishnamurthy Dvijotham. Interactive concept bottleneck models. In Proceedings of the AAAI Conference on Artificial Intelligence, pages 5948-5955, 2023. 2", + "[3] Guangyi Chen, Weiran Yao, Xiangchen Song, Xinyue Li, Yongming Rao, and Kun Zhang. Plot: Prompt learning with optimal transport for vision-language models. arXiv preprint arXiv:2210.01253, 2022. 3", + "[4] Liquin Chen, Zhe Gan, Yu Cheng, Linjie Li, Lawrence Carin, and Jingjing Liu. Graph optimal transport for cross-domain alignment. In International Conference on Machine Learning, pages 1542-1553. PMLR, 2020. 3", + "[5] Xianjie Chen, Roozbeh Mottaghi, Xiaobai Liu, Sanja Fidler, Raquel Urtasun, and Alan Yuille. Detect what you can: Detecting and representing objects using holistic models and body parts. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 1971-1978, 2014. 6", + "[6] Molnar Christoph. Interpretable machine learning: A guide for making black box models explainable. Leanpub, 2020. 1", + "[7] Marco Cuturi. Sinkhorn distances: Lightspeed computation of optimal transport. Advances in neural information processing systems, 26, 2013. 4", + "[8] Barry De Ville. Decision trees. Wiley Interdisciplinary Reviews: Computational Statistics, 5(6):448-455, 2013. 1", + "[9] Jia Deng, Wei Dong, Richard Socher, Li-Jia Li, Kai Li, and Li Fei-Fei. Imagenet: A large-scale hierarchical image database. In 2009 IEEE conference on computer vision and pattern recognition, pages 248-255. IEEE, 2009. 6", + "[10] Mateo Espinosa Zarlenga, Pietro Barbiero, Gabriele Ciravegna, Giuseppe Marra, Francesco Giannini, Michelangelo Diligenti, Zohreh Shams, Frederic Precioso, Stefano Melacci, Adrian Weller, et al. Concept embedding models: Beyond the accuracy-explainability trade-off. Advances in Neural Information Processing Systems, 35:21400-21413, 2022. 2", + "[11] Mateo Espinosa Zarlenga, Katie Collins, Krishnamurthy Dvijotham, Adrian Weller, Zohreh Shams, and Mateja Jamnik. Learning to receive help: Intervention-aware concept" + ], + "bbox": [ + 93, + 316, + 482, + 901 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "embedding models. Advances in Neural Information Processing Systems, 36, 2024. 2", + "[12] Charlie Frogner, Chiyuan Zhang, Hossein Mobahi, Mauricio Araya, and Tomaso A Poggio. Learning with a wasserstein loss. Advances in neural information processing systems, 28, 2015. 3", + "[13] Ju He, Shuo Yang, Shaokang Yang, Adam Kortylewski, Xiaoding Yuan, Jie-Neng Chen, Shuai Liu, Cheng Yang, Qihang Yu, and Alan Yuille. Partimagenet: A large, high-quality dataset of parts. In European Conference on Computer Vision, pages 128-145. Springer, 2022. 6", + "[14] Johan Huysmans, Karel Dejaeger, Christophe Mues, Jan Vanthienen, and Bart Baesens. An empirical evaluation of the comprehensibility of decision table, tree and rule based predictive models. Decision Support Systems, 51(1):141-154, 2011. 1", + "[15] Aditya Khosla, Nityananda Jayadevaprakash, Bangpeng Yao, and Li Fei-Fei. Novel dataset for fine-grained image categorization. In First Workshop on Fine-Grained Visual Categorization, IEEE Conference on Computer Vision and Pattern Recognition, Colorado Springs, CO, 2011. 7", + "[16] Alexander Kirillov, Eric Mintun, Nikhila Ravi, Hanzi Mao, Chloe Rolland, Laura Gustafson, Tete Xiao, Spencer Whitehead, Alexander C Berg, Wan-Yen Lo, et al. Segment anything. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 4015-4026, 2023. 7", + "[17] Pang Wei Koh, Thao Nguyen, Yew Siang Tang, Stephen Mussmann, Emma Pierson, Been Kim, and Percy Liang. Concept bottleneck models. In International conference on machine learning, pages 5338-5348. PMLR, 2020. 1, 2, 3, 6, 7", + "[18] Andrei V Konstantinov and Lev V Utkin. Incorporating expert rules into neural networks in the framework of concept-based learning. arXiv preprint arXiv:2402.14726, 2024. 2", + "[19] Alex Krizhevsky, Geoffrey Hinton, et al. Learning multiple layers of features from tiny images. 2009. 6", + "[20] John Lee, Max Dabagia, Eva Dyer, and Christopher Rozell. Hierarchical optimal transport for multimodal distribution alignment. Advances in neural information processing systems, 32, 2019. 3", + "[21] Miaoge Li, Dongsheng Wang, Xinyang Liu, Zequn Zeng, Ruiying Lu, Bo Chen, and Mingyuan Zhou. Patchct: Aligning patch set and label set with conditional transport for multi-label image classification. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 15348-15358, 2023. 3", + "[22] Tang Li, Mengmeng Ma, and Xi Peng. Deal: Disentangle and localize concept-level explanations for vlms. In European Conference on Computer Vision, pages 383-401. Springer, 2025. 2, 4", + "[23] Yijie Lin, Jie Zhang, Zhenyu Huang, Jia Liu, Zujie Wen, and Xi Peng. Multi-granularity correspondence learning from long-term noisy videos. arXiv preprint arXiv:2401.16702, 2024. 3", + "[24] Jiawei Ma, Po-Yao Huang, Saining Xie, Shang-Wen Li, Luke Zettlemoyer, Shih-Fu Chang, Wen-Tau Yih, and Hu Xu. Mode: Clip data experts via clustering. In Proceed-" + ], + "bbox": [ + 516, + 92, + 903, + 901 + ], + "page_idx": 8 + }, + { + "type": "page_number", + "text": "30207", + "bbox": [ + 478, + 944, + 517, + 955 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "ings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 26354-26363, 2024. 2", + "[25] Shweta Mahajan, Teresa Botschen, Iryna Gurevych, and Stefan Roth. Joint wasserstein autoencoders for aligning multimodal embeddings. In Proceedings of the IEEE/CVF International Conference on Computer Vision Workshops, pages 0-0, 2019. 3", + "[26] Andrei Margeloiu, Matthew Ashman, Umang Bhatt, Yanzhi Chen, Mateja Jamnik, and Adrian Weller. Do concept bottleneck models learn as intended? arXiv preprint arXiv:2105.04289, 2021. 2, 3, 6", + "[27] Nesta Midavaine, Gregory Hok Tjoan Go, Diego Canez, Ioana Simion, and Satchit Chatterji. [re] on the reproducibility of post-hoc concept bottleneck models. Transactions on Machine Learning Research. 2", + "[28] Mazda Moayeri, Phillip Pope, Yogesh Balaji, and Soheil Feizi. A comprehensive study of image classification model sensitivity to foregrounds, backgrounds, and visual attributes. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, 2022. 6", + "[29] Mazda Moayeri, Keivan Rezaei, Maziar Sanjabi, and Soheil Feizi. Text2concept: Concept activation vectors directly from text. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR) Workshops, pages 3744-3749, 2023. 2, 4", + "[30] Keval Morabia, Jatin Arora, and Tara Vijaykumar. Attention-based joint detection of object and semantic part. arXiv preprint arXiv:2007.02419, 2020. 6", + "[31] Tuomas Oikarinen, Subhro Das, Lam M Nguyen, and Tsui-Wei Weng. Label-free concept bottleneck models. arXiv preprint arXiv:2304.06129, 2023. 2", + "[32] Maxime Oquab, Timothee Darcet, Theo Moutakanni, Huy Vo, Marc Szafraniec, Vasil Khalidov, Pierre Fernandez, Daniel Haziza, Francisco Massa, Alaaeldin El-Nouby, et al. Dinov2: Learning robust visual features without supervision. arXiv preprint arXiv:2304.07193, 2023. 6", + "[33] Alec Radford, Jong Wook Kim, Chris Hallacy, Aditya Ramesh, Gabriel Goh, Sandhini Agarwal, Girish Sastry, Amanda Askell, Pamela Mishkin, Jack Clark, et al. Learning transferable visual models from natural language supervision. In International conference on machine learning, pages 8748-8763. PMLR, 2021. 6", + "[34] Maithra Raghu, Thomas Unterthiner, Simon Kornblith, Chiyuan Zhang, and Alexey Dosovitskiy. Do vision transformers see like convolutional neural networks? Advances in neural information processing systems, 34:12116-12128, 2021. 4", + "[35] Naveen Raman, Mateo Espinosa Zarlenga, Juyeon Heo, and Mateja Jamnik. Do concept bottleneck models obey locality? In XAI in Action: Past, Present, and Future Applications, 2023. 3", + "[36] Ievgen Redko, Nicolas Courty, Rémi Flamary, and Devis Tuia. Optimal transport for multi-source domain adaptation under target shift. In The 22nd International Conference on artificial intelligence and statistics, pages 849-858. PMLR, 2019. 3", + "[37] Ramprasaath R Selvaraju, Michael Cogswell, Abhishek Das, Ramakrishna Vedantam, Devi Parikh, and Dhruv Batra." + ], + "bbox": [ + 91, + 92, + 482, + 900 + ], + "page_idx": 9 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "Grad-cam: Visual explanations from deep networks via gradient-based localization. In Proceedings of the IEEE international conference on computer vision, pages 618-626, 2017. 1", + "[38] Andrei Semenov, Vladimir Ivanov, Aleksandr Beznosikov, and Alexander Gasnikov. Sparse concept bottleneck models: Gumbel tricks in contrastive learning. arXiv preprint arXiv:2404.03323, 2024. 2, 6, 7", + "[39] Ivaxi Sheth and Samira Ebrahimi Kahou. Auxiliary losses for learning generalizable concept-based models. Advances in Neural Information Processing Systems, 36, 2024. 2, 6, 7", + "[40] Sungbin Shin, Yohan Jo, Sungsoo Ahn, and Namhoon Lee. A closer look at the intervention procedure of concept bottleneck models. In International Conference on Machine Learning, pages 31504-31520. PMLR, 2023. 2", + "[41] Catherine Wah, Steve Branson, Peter Welinder, Pietro Perona, and Serge Belongie. The caltech-ucsd birds-200-2011 dataset. 2011. 6, 7", + "[42] Tiansheng Wen, Yifei Wang, Zequn Zeng, Zhong Peng, Yudi Su, Xinyang Liu, Bo Chen, Hongwei Liu, Stefanie Jegelka, and Chenyu You. Beyond matryoshka: Revisiting sparse coding for adaptive representation. arXiv preprint arXiv:2503.01776, 2025. 2", + "[43] Yongqin Xian, Christoph H Lampert, Bernt Schiele, and Zeynep Akata. Zero-shot learning—a comprehensive evaluation of the good, the bad and the ugly. IEEE transactions on pattern analysis and machine intelligence, 41(9):2251-2265, 2018. 6", + "[44] Wenjia Xu, Yongqin Xian, Jiuniu Wang, Bernt Schiele, and Zeynep Akata. Attribute prototype network for zero-shot learning. Advances in Neural Information Processing Systems, 33:21969-21980, 2020. 6", + "[45] Yue Yang, Artemis Panagopoulou, Shenghao Zhou, Daniel Jin, Chris Callison-Burch, and Mark Yatskar. Language in a bottle: Language model guided concept bottlenecks for interpretable image classification. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 19187-19197, 2023. 2, 4, 6, 7", + "[46] Mert Yuksekgonul, Maggie Wang, and James Zou. Post-hoc concept bottleneck models. arXiv preprint arXiv:2205.15480, 2022. 2", + "[47] Yuan Zang, Tian Yun, Hao Tan, Trung Bui, and Chen Sun. Pre-trained vision-language models learn discoverable visual concepts. arXiv preprint arXiv:2404.12652, 2024. 2, 4, 6", + "[48] Mateo Espinosa Zarlenga, Pietro Barbiero, Gabriele Ciravegna, Giuseppe Marra, Francesco Giannini, Michelangelo Diligenti, Zohreh Shams, Frederic Precioso, Stefano Melacci, Adrian Weller, et al. Concept embedding models: Beyond the accuracy-explainability trade-off. arXiv preprint arXiv:2209.09056, 2022. 6, 7", + "[49] Zequn Zeng, Hao Zhang, Ruiying Lu, Dongsheng Wang, Bo Chen, and Zhengjue Wang. Conzic: Controllable zero-shot image captioning by sampling-based polishing. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 23465-23476, 2023. 2", + "[50] Zequn Zeng, Yan Xie, Hao Zhang, Chiyu Chen, Bo Chen, and Zhengjue Wang. Meacap: Memory-augmented zero" + ], + "bbox": [ + 516, + 92, + 905, + 900 + ], + "page_idx": 9 + }, + { + "type": "page_number", + "text": "30208", + "bbox": [ + 478, + 944, + 517, + 955 + ], + "page_idx": 9 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "shot image captioning. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 14100-14110, 2024. 2", + "[51] Han Zhang, Tao Xu, Mohamed Elhoseiny, Xiaolei Huang, Shaoting Zhang, Ahmed Elgammal, and Dimitris Metaxas. Spda-cnn: Unifying semantic part detection and abstraction for fine-grained recognition. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 1143-1152, 2016. 6", + "[52] Peng Zhao and Zhi-Hua Zhou. Label distribution learning by optimal transport. In Proceedings of the AAAI Conference on Artificial Intelligence, 2018. 3" + ], + "bbox": [ + 91, + 90, + 482, + 260 + ], + "page_idx": 10 + }, + { + "type": "page_number", + "text": "30209", + "bbox": [ + 478, + 945, + 517, + 955 + ], + "page_idx": 10 + } +] \ No newline at end of file diff --git a/2025/Discovering Fine-Grained Visual-Concept Relations by Disentangled Optimal Transport Concept Bottleneck Models/353387ac-1b53-4c3c-ac08-91f95db61c80_model.json b/2025/Discovering Fine-Grained Visual-Concept Relations by Disentangled Optimal Transport Concept Bottleneck Models/353387ac-1b53-4c3c-ac08-91f95db61c80_model.json new file mode 100644 index 0000000000000000000000000000000000000000..b364ef7cb748adedf0f6a8c528f93843667d2a14 --- /dev/null +++ b/2025/Discovering Fine-Grained Visual-Concept Relations by Disentangled Optimal Transport Concept Bottleneck Models/353387ac-1b53-4c3c-ac08-91f95db61c80_model.json @@ -0,0 +1,2301 @@ +[ + [ + { + "type": "header", + "bbox": [ + 0.107, + 0.003, + 0.182, + 0.043 + ], + "angle": 0, + "content": "CVF" + }, + { + "type": "header", + "bbox": [ + 0.237, + 0.001, + 0.812, + 0.047 + ], + "angle": 0, + "content": "This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore." + }, + { + "type": "title", + "bbox": [ + 0.11, + 0.13, + 0.89, + 0.177 + ], + "angle": 0, + "content": "Discovering Fine-Grained Visual-Concept Relations by Disentangled Optimal Transport Concept Bottleneck Models" + }, + { + "type": "text", + "bbox": [ + 0.342, + 0.203, + 0.662, + 0.222 + ], + "angle": 0, + "content": "Yan Xie\\(^{1*}\\) Zequn Zeng\\(^{1*}\\) Hao Zhang\\(^{1\\dagger}\\)" + }, + { + "type": "text", + "bbox": [ + 0.199, + 0.222, + 0.796, + 0.24 + ], + "angle": 0, + "content": "Yucheng Ding\\(^{1}\\) Yi Wang\\(^{1}\\) Zhengjue Wang\\(^{2}\\) Bo Chen\\(^{1}\\) Hongwei Liu\\(^{1}\\)" + }, + { + "type": "text", + "bbox": [ + 0.122, + 0.24, + 0.875, + 0.257 + ], + "angle": 0, + "content": "\\(^{1}\\)National Key Laboratory of Radar Signal Processing, Xidian University, Xi'an, 710071, China" + }, + { + "type": "text", + "bbox": [ + 0.121, + 0.257, + 0.877, + 0.275 + ], + "angle": 0, + "content": "\\(^{2}\\)State Key Laboratory of Integrated Service Networks, Xidian University, Xi'an, 710071, China" + }, + { + "type": "list", + "bbox": [ + 0.121, + 0.24, + 0.877, + 0.275 + ], + "angle": 0, + "content": null + }, + { + "type": "text", + "bbox": [ + 0.178, + 0.278, + 0.819, + 0.293 + ], + "angle": 0, + "content": "{yanxie0904, zzequn99, zhanghao_xidian}@163.com, bchen@mail.xidian.edu.cn" + }, + { + "type": "title", + "bbox": [ + 0.248, + 0.327, + 0.327, + 0.343 + ], + "angle": 0, + "content": "Abstract" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.358, + 0.485, + 0.797 + ], + "angle": 0, + "content": "Concept Bottleneck Models (CBMs) try to make the decision-making process transparent by exploring an intermediate concept space between the input image and the output prediction. Existing CBMs just learn coarse-grained relations between the whole image and the concepts, less considering local image information, leading to two main drawbacks: i) they often produce spurious visual-concept relations, hence decreasing model reliability; and ii) though CBMs could explain the importance of every concept to the final prediction, it is still challenging to tell which visual region produces the prediction. To solve these problems, this paper proposes a Disentangled Optimal Transport CBM (DOT-CBM) framework to explore fine-grained visual-concept relations between local image patches and concepts. Specifically, we model the concept prediction process as a transportation problem between the patches and concepts, thereby achieving explicit fine-grained feature alignment. We also incorporate orthogonal projection losses within the modality to enhance local feature disentanglement. To further address the shortcut issues caused by statistical biases in the data, we utilize the visual saliency map and concept label statistics as transportation priors. Thus, DOT-CBM can visualize inversion heatmaps, provide more reliable concept predictions, and produce more accurate class predictions. Comprehensive experiments demonstrate that our proposed DOT-CBM achieves SOTA performance on several tasks, including image classification, local part detection and out-of-distribution generalization. Codes are available in supplementary material." + }, + { + "type": "title", + "bbox": [ + 0.092, + 0.824, + 0.222, + 0.839 + ], + "angle": 0, + "content": "1. Introduction" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.849, + 0.483, + 0.88 + ], + "angle": 0, + "content": "Explainable artificial intelligence aims to allow human users to comprehend and trust the output created by ma" + }, + { + "type": "image", + "bbox": [ + 0.536, + 0.326, + 0.888, + 0.422 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.534, + 0.423, + 0.886, + 0.448 + ], + "angle": 0, + "content": "(a) Brief comparison about the coarse-grained design of vanilla CBM and fine-grained design of our proposed DOT-CBM." + }, + { + "type": "image", + "bbox": [ + 0.533, + 0.46, + 0.885, + 0.591 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.612, + 0.593, + 0.803, + 0.606 + ], + "angle": 0, + "content": "(b) Inversion heatmap visualization." + }, + { + "type": "image_caption", + "bbox": [ + 0.512, + 0.607, + 0.907, + 0.76 + ], + "angle": 0, + "content": "Figure 1. Comparison between vanilla CBM and our proposed DOT-CBM, including architecture design and inversion heatmap for concept predictions. (a) Due to the black-box mapping from images to concepts, Vanilla CBM needs Grad-CAM [37] techniques to locate concept predictions back to image space while DOT-CBM can provide an explicit inversion heatmap visualization. (b) Due to lack of fine-grained alignment, vanilla CBM produces spurious correlations in three levels of granularity (part-background, part-object and part-part level) that mislocalizes the local concept to the background, the whole object, and incorrect local region." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.779, + 0.907, + 0.854 + ], + "angle": 0, + "content": "chine learning algorithms [6, 14]. Compared with post-hoc explanation methods [37] for black-box models, designing explicit interpretability during modeling and learning processes is beneficial to obtaining more natural explanations [8]." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.856, + 0.907, + 0.901 + ], + "angle": 0, + "content": "The Concept Bottleneck Model (CBM) [17], a representative interpretable image classification model, first predicts a pre-defined set of concepts from images and then" + }, + { + "type": "page_footnote", + "bbox": [ + 0.109, + 0.887, + 0.373, + 0.901 + ], + "angle": 0, + "content": "*Equal contribution. †Corresponding authors" + }, + { + "type": "page_number", + "bbox": [ + 0.479, + 0.945, + 0.52, + 0.958 + ], + "angle": 0, + "content": "30199" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.09, + 0.092, + 0.482, + 0.243 + ], + "angle": 0, + "content": "maps these concept descriptions to the final output class. Therefore, CBM is good at interpreting the relationship between the input image and output class predictions via intermediate human-understandable concepts. However, vanilla CBM brings these good properties at the expense of classification performance [27]. For this problem, a series of CBM variants are proposed from two lines, construct a more comprehensive and discriminative concept bank [31, 45, 46], and sophisticated architecture design [29, 38, 39] for better concept representation learning." + }, + { + "type": "text", + "bbox": [ + 0.093, + 0.246, + 0.485, + 0.518 + ], + "angle": 0, + "content": "Although these variants narrow the performance gap between CBMs and their black-box counterparts, their modeling method between images and concepts still has two disadvantages, as shown in Fig. 1a. Firstly, they predict multiple local concepts (i.e. local attributes like \"head\") from a global image feature, which only learns a coarse-grained relationship between the full image and all concepts. Secondly, the interpretability of prevailing CBMs mainly comes from the linear concept-class relations. The mapping from the original image to all concepts is still modeled as the black-box neural networks, which still need some post-hoc analysis [26, 27] to visualize the learned visual-concept relations. As shown in Fig. 1b, these two disadvantages often bring some visual-concept spurious correlations, where many concept predictions are frequently mis-localized in the images. These spurious correlations inherently undermine the interpretability and generalizing capabilities of CBMs." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.521, + 0.484, + 0.822 + ], + "angle": 0, + "content": "Empirically, we categorized these visual-concept spurious correlations into three levels of granularity: i) part-background spurious correlation. This refers to a misattribution of the concept predictions of local parts to the background. ii) part-object spurious correlation. The concepts of local parts are wrongly associated with the entire object without further distinctions. iii) part-part spurious correlation. In this case, the local concepts are mislocalized to incorrect image regions. We attribute these spurious correlation problems to two key factors. The first one is the lack of fine-grained alignment between the local image regions and corresponding textual concepts, i.e. the fine-grained visual-concept relations. The second one is the overfitting of the data bias such as the statistically high co-occurrence between class and background and between concept and concept, yielding confusion to learn the difference between them. Besides, it is intuitive that the explicit interpretability modeling for the concept predictions from the images is conceivably effective in addressing these visual-concept spurious correlations." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.825, + 0.484, + 0.903 + ], + "angle": 0, + "content": "To this end, we propose a novel Disentangled Optimal Transport Concept Bottleneck Models, namely DOT-CBM. Compared with current CBMs, our proposed DOT-CBM can provide holistic explicit interpretability on not only the concept-class relations but also the visual region" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.092, + 0.905, + 0.287 + ], + "angle": 0, + "content": "concept relations. Specifically, we introduce a disentangled optimal transport (DOT) framework to ensure the intramodality disentanglement and inter-modality alignment between the image patches and textual concepts. The transport assignment solution of OT framework, is an explicit correlation matrix between image patches and pre-defined concepts, facilitating bidirectional outputs including concept predictions on concept space and concept inversion mask on the image space. Furthermore, to tackle the overfitting to the data bias, we introduce two discrete prior distributions on the patch set and the concept set to penalize the incorrect shortcut alignment between patches and concepts. In summary, our contributions are as follows:" + }, + { + "type": "text", + "bbox": [ + 0.514, + 0.289, + 0.905, + 0.348 + ], + "angle": 0, + "content": "- We empirically investigate the visual-concept spurious correlation phenomenon and pinpoint the two key factors of these limitations, i.e. lack of fine-grained visual-concept alignment and overfitting to the data bias." + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.349, + 0.905, + 0.455 + ], + "angle": 0, + "content": "- We propose a novel disentangled optimal transport framework to model the fine-grained visual-concept alignment and address the overfitting problems. The transport design between image patches and textual concepts provides explicit relations between image and concept predictions, which can produce concept predictions on the concept spaces and concept inversion mask on the image space." + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.455, + 0.907, + 0.545 + ], + "angle": 0, + "content": "- Extensive experiments on image classification, part detection, and out-of-distribution generalization benchmarks demonstrate our proposed DOT-CBM can achieve significant improvements over current state-of-the-art CBMs on classification accuracy, interpretability and generalizing capabilities to new distributions." + }, + { + "type": "list", + "bbox": [ + 0.513, + 0.289, + 0.907, + 0.545 + ], + "angle": 0, + "content": null + }, + { + "type": "title", + "bbox": [ + 0.514, + 0.559, + 0.657, + 0.575 + ], + "angle": 0, + "content": "2. Related Work" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.584, + 0.907, + 0.825 + ], + "angle": 0, + "content": "Concept Bottleneck Models are a family of models [17] that first predict concepts from input images, then use these concepts to predict a downstream label. This design allows CBMs to (1) provide concept-based explanations via their predicted concepts, and (2) improve their test performance when deployed with experts via concept interventions [2, 11, 40]. Similar to previous interpretable models, CBMs also face the issue of balancing task accuracy and interoperability [10, 27]. Recent works [31, 46] has leveraged extra knowledge graph [38, 46], expert rules [18] or Large Language Models (LLMs) [22, 31, 38, 45, 47] to overcome the limitations of manually defined concept sets and has utilized pre-trained Vision-Language Models (VLMs) [24, 42, 45, 49, 50] to obtain concept predictions. These advancements have elevated the performance of CBMs to levels approaching those of black-box models." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.826, + 0.909, + 0.903 + ], + "angle": 0, + "content": "The interpretability of prevailing CBMs primarily arises from linear relationships between concepts and classes. However, the mapping from the image to the concepts is still handled by black-box neural networks, requiring post-hoc analysis for visualization of the learned visual-concept" + }, + { + "type": "page_number", + "bbox": [ + 0.48, + 0.945, + 0.521, + 0.958 + ], + "angle": 0, + "content": "30200" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.098, + 0.09, + 0.904, + 0.312 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.09, + 0.32, + 0.907, + 0.461 + ], + "angle": 0, + "content": "Figure 2. Overview of our proposed DOT-CBM. The overall CBM framework proceeds from left to right. In the first stage, the model transforms input images into concept activation vectors. In forward pass, local embeddings from a pre-trained Vision Transformer (ViT) and a text encoder are processed through learnable adapters to generate two feature sets (Sec. 3.2). Two orthogonal projection losses are applied to both feature sets to constrain Adapter training. An Optimal Transport (OT) algorithm optimizes the Assignment Matrix, representing the explicit correlation between concepts and image patches. Concept activation values are derived by combining this matrix with a cost matrix, and the loss is supervised by concept labels for training (Sec. 3.3). To address data bias, we use the Saliency Map from the pre-trained ViT for rough foreground-background separation as the visual prior in OT. A Conditioned Co-occurrence Matrix, based on frequency statistics, serves as the prior for concepts, enhancing the model's ability to distinguish co-occurring concepts (Sec. 3.4). In the second stage, consistent with the vanilla CBM framework, concept activation values are used to predict final class labels via a linear classification network, which is trained using class labels (Sec. 3.5)." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.486, + 0.483, + 0.546 + ], + "angle": 0, + "content": "relations. This lack of transparency often leads to visual-concept spurious correlations, where concept predictions are frequently mis-localized in the image. We propose a disentangled OT to address these limitations." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.547, + 0.483, + 0.637 + ], + "angle": 0, + "content": "Optimal Transport aims to find the most efficient way to map one distribution onto another by minimizing a transportation cost. In the context of machine learning, OT is widely applied in a variety of domains, including unsupervised domain adaptation [36], label representation [12, 52], cross-modal semantics [3, 4, 20, 21, 23, 25], and beyond." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.637, + 0.484, + 0.698 + ], + "angle": 0, + "content": "In this work, we further extend OT to a disentangled OT framework, to enforce the fine-grained alignment between image patches and their corresponding concepts while also ensuring disentanglement within each modality." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.71, + 0.182, + 0.725 + ], + "angle": 0, + "content": "3. Method" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.735, + 0.483, + 0.902 + ], + "angle": 0, + "content": "In this section, we begin with a brief review of the CBM problem setup in Sec. 3.1 including two stages of optimization, i.e. patches to concepts \\((f_{\\mathcal{X}\\rightarrow \\mathcal{C}})\\) and concepts to classes \\((f_{\\mathcal{C}\\rightarrow \\mathcal{Y}})\\). In the first stage, to address the limited interpretation of previous CBMs in the patches-to-concepts process, we introduce a disentangled optimal transport alignment framework between patches and concepts. Concretely, in Sec. 3.2, we propose two orthogonal projection losses to disentangle the patch features and concept features. In Sec. 3.3, we introduce an optimal transport (OT) based framework to model the complex relationship" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.486, + 0.907, + 0.622 + ], + "angle": 0, + "content": "between patches and concepts, which can predict concepts from patches and meanwhile inverse the concept to patches to provide explanations of concept prediction. In Sec. 3.4, we integrate two prior distribution modeling to address the data bias problem. In the second stage, in Sec. 3.5, we follow vanilla CBMs and predict classes from concepts. Overall, our proposed DOT-CBM achieves holistic interpretability whether on intermediate concept predictions or on final class predictions. The whole pipeline is shown in Fig. 2." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.634, + 0.719, + 0.649 + ], + "angle": 0, + "content": "3.1. Problem Formulation" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.659, + 0.906, + 0.749 + ], + "angle": 0, + "content": "Traditional black-box visual models aim to fit the conditional distribution \\( P(y|x) \\) given an a dataset \\( \\mathcal{D} = \\{x_i, y_i\\}_{i=1}^N \\), where \\( x_i \\in \\mathcal{X}, y_i \\in \\mathcal{Y} \\) are images and class labels, respectively. Though effective, they still lack transparency to understand the underlying decision-making process to drive the final predictions." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.75, + 0.907, + 0.902 + ], + "angle": 0, + "content": "Concept Bottleneck Models (CBMs) [17] offer an interpretable solution by first predicting human-interpretable high-level concepts before making the final prediction. Formally, the dataset is extended to \\(\\mathcal{D} = \\{x_i,y_i,\\pmb {a}_i\\}_{i = 1}^N\\) , where \\(\\pmb {a}_i = \\{a_{ij}|a_{ij}\\in \\{0,1\\} \\}_{j = 1}^M\\) represents a binary concept activation label. \\(a_{ij} = 1\\) indicates the presence of a specific high-level concept \\(c_{j}\\) in image \\(x_{i}\\) , while \\(a_{ij} = 0\\) means the absence. \\(c_{j}\\) is concept from the pre-defined concept set \\(\\mathcal{C} = \\{c_j\\}_{j = 1}^M\\) . The optimization process of CBMs can be divided into two-stage [26, 35]: I) they first leverage a con" + }, + { + "type": "page_number", + "bbox": [ + 0.48, + 0.945, + 0.518, + 0.957 + ], + "angle": 0, + "content": "30201" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.09, + 0.09, + 0.485, + 0.213 + ], + "angle": 0, + "content": "cept predictor \\(f_{\\mathcal{X}\\to \\mathcal{C}}:\\mathbb{R}^{H\\times W\\times C}\\to \\mathbb{R}^M\\) to map the input data \\(x\\) to the concept activation vector \\(\\pmb {a}\\in \\mathbb{R}^{M}\\) . 2) Subsequently, a target predictor \\(f_{\\mathcal{C}\\rightarrow \\mathcal{Y}}:\\mathbb{R}^{M}\\to \\mathbb{R}^{N_{y}}\\) maps the concept activation vector \\(\\pmb{a}\\) to the final label \\(y\\) . This structure enhances the interpretability of the model, as the intermediate concept predictions provide interpretability into the model's final decision. However, the interpretability from image to concept is un-explored." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.222, + 0.45, + 0.239 + ], + "angle": 0, + "content": "3.2. Disentanglement on Local representations" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.246, + 0.484, + 0.367 + ], + "angle": 0, + "content": "Disentangled visual patch features \\(\\widetilde{\\mathcal{P}}\\). As previously discussed, current CBMs typically associate multiple local concepts with a global image feature, which leads to coarse-grained correspondences. In this paper, we adopt the pretrained vision transformer (ViT) [1] model, which has been approved to exhibit better spatial location information extraction than CNN [34], to extract local patch features to represent the local image regions." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.368, + 0.484, + 0.535 + ], + "angle": 0, + "content": "Specifically, as shown in the left part of Fig. 2, ViT uniformly divides the full image \\(x \\in \\mathbb{R}^{H \\times W \\times C}\\) into \\(N\\) non-overlapping patches \\(\\{I_{pi} \\in \\mathbb{R}^{P_h \\times P_w}\\}_{i=1}^N\\). After linearly projecting these patches into a \\(d\\)-dimension embedding space and flattening all patch embeddings into a sequence, we can get an initialized patch embedding matrix \\(X_P \\in \\mathbb{R}^{N \\times d}\\). To extract the global image feature, ViT concatenates a learnable special token [CLS] with \\(X_P\\). Besides, in order to preserve the spatial information of images, ViT adds a learnable position embedding \\(X_{pos} \\in \\mathbb{R}^{(N+1) \\times d}\\) and we get the input embeddings \\(X\\) for ViT:" + }, + { + "type": "equation", + "bbox": [ + 0.182, + 0.546, + 0.483, + 0.565 + ], + "angle": 0, + "content": "\\[\nX _ {P} = \\left[ I _ {p 1} \\cdot W _ {p}, \\dots , I _ {p N} \\cdot W _ {p} \\right], \\tag {1}\n\\]" + }, + { + "type": "equation", + "bbox": [ + 0.184, + 0.566, + 0.482, + 0.583 + ], + "angle": 0, + "content": "\\[\nX = \\left[ \\left[ \\mathrm {C L S} \\right], X _ {P} \\right] + X _ {\\text {p o s}}, \\tag {2}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.594, + 0.484, + 0.658 + ], + "angle": 0, + "content": "where \\(W_{p}\\in \\mathbb{R}^{(P_{h}\\cdot P_{w}\\cdot C)\\times d}\\) is the patch embedding projection weights. Then, the pre-trained ViT \\(E_I\\) to map the input embeddings \\(X\\) into the global feature \\(g_{I}\\), and patch features \\(\\mathcal{P} = \\{p_i\\}_{i = 1}^N\\):" + }, + { + "type": "equation", + "bbox": [ + 0.227, + 0.668, + 0.483, + 0.687 + ], + "angle": 0, + "content": "\\[\n[ g _ {I}, \\mathcal {P} ] = E _ {I} (X), \\tag {3}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.697, + 0.484, + 0.804 + ], + "angle": 0, + "content": "We empirically find that the patch features within the object regions are highly similar which is good for object representation, but bad for distinguishing local parts within the object region. Thus, we fit a learnable visual adapter \\( A_v \\) on the \\( \\mathcal{P} \\) and map the patch features within the foreground region to a disentangled unit hypersphere by introducing a visual orthogonal projection (VOP) regularizer as:" + }, + { + "type": "equation", + "bbox": [ + 0.244, + 0.815, + 0.483, + 0.832 + ], + "angle": 0, + "content": "\\[\nS _ {I} = g _ {I} \\cdot P, \\tag {4}\n\\]" + }, + { + "type": "equation", + "bbox": [ + 0.15, + 0.835, + 0.483, + 0.868 + ], + "angle": 0, + "content": "\\[\n\\langle A _ {v} (p _ {i}), A _ {v} (p _ {j}) \\rangle = \\frac {A _ {v} (p _ {i}) \\cdot A _ {v} (p _ {j})}{\\| A _ {v} (p _ {i}) \\| \\| A _ {v} (p _ {j}) \\|}, \\tag {5}\n\\]" + }, + { + "type": "equation", + "bbox": [ + 0.108, + 0.873, + 0.483, + 0.898 + ], + "angle": 0, + "content": "\\[\n\\mathcal {L} _ {\\mathrm {V O P}} = \\underset {\\left(p _ {i}, p _ {j}\\right) \\sim \\mathcal {P}} {\\mathbb {E}} \\left[ S _ {I} (i) \\cdot S _ {I} (j) \\cdot \\left\\langle A _ {v} \\left(p _ {i}\\right), A _ {v} \\left(p _ {j}\\right) \\right\\rangle \\right] \\tag {6}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.092, + 0.907, + 0.248 + ], + "angle": 0, + "content": "where \\( S_{I} \\) is the saliency map to distinguish foreground and background, \\( \\langle \\cdot ,\\cdot \\rangle \\) is the cosine similarity. Finally, we get the disentangled patch features \\( \\widetilde{\\mathcal{P}} = \\{\\widetilde{p_i} = A_v(p_i)\\}_{i = 1}^N \\) Disentangled textual concept features \\( \\widetilde{\\mathcal{T}} \\). For concept feature extraction, following [22, 29, 45, 47], we utilize a pre-trained text encoder \\( E_{T} \\) to compute concept embeddings \\( \\mathcal{T} = \\{t_j = E_T(c_j)\\}_{j = 1}^M \\). Similarly, we fit a learnable textual adapter \\( A_{t} \\) and introduce a textual orthogonal projection (TOP) regularizer to disentangle the concept features in a unit hyper-sphere, as:" + }, + { + "type": "equation", + "bbox": [ + 0.592, + 0.256, + 0.907, + 0.282 + ], + "angle": 0, + "content": "\\[\n\\mathcal {L} _ {\\mathrm {T O P}} = \\underset {(t _ {i}, t _ {j}) \\sim \\mathcal {T}} {\\mathbb {E}} [ \\langle A _ {t} (t _ {i}), A _ {t} (t _ {j}) \\rangle ], \\tag {7}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.514, + 0.291, + 0.907, + 0.323 + ], + "angle": 0, + "content": "Consequently, we get the disentangled concept features \\(\\widetilde{\\mathcal{T}} = \\{\\widetilde{t_j} = A_t(t_j)\\}_{j = 1}^M.\\)" + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.329, + 0.907, + 0.346 + ], + "angle": 0, + "content": "3.3. Patch-concept alignment via optimal transport" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.353, + 0.906, + 0.474 + ], + "angle": 0, + "content": "Now, we have the disentangled patch feature set \\(\\widetilde{\\mathcal{P}}\\) and disentangled concept feature set \\(\\widetilde{\\mathcal{T}}\\), we need to model the relationship \\(\\mathbf{Q}\\) between these two sets and derive the concept activation vector \\(\\hat{\\pmb{a}}\\) based on the input image. As a complex many-to-many alignment problem, we propose an optimal transport (OT) based binary cross-entropy loss to tackle these problems. Our goal is to correctly align the patch features and concept features with the same semantics." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.475, + 0.906, + 0.521 + ], + "angle": 0, + "content": "Assume that these two sets \\(\\widetilde{\\mathcal{P}}\\) and \\(\\widetilde{\\mathcal{T}}\\) follow two discrete probability distributions \\(\\Theta \\in \\mathbb{R}^N\\) and \\(\\Gamma \\in \\mathbb{R}^M\\) on the \\(d\\)-dimension embedding space, respectively:" + }, + { + "type": "equation", + "bbox": [ + 0.598, + 0.53, + 0.906, + 0.573 + ], + "angle": 0, + "content": "\\[\n\\boldsymbol {\\Theta} = \\sum_ {i = 1} ^ {N} \\theta_ {i} \\delta_ {p _ {i}}, \\quad \\boldsymbol {\\Gamma} = \\sum_ {j = 1} ^ {M} \\gamma_ {j} \\delta_ {c _ {j}}, \\tag {8}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.583, + 0.906, + 0.661 + ], + "angle": 0, + "content": "where \\(\\pmb{\\theta} \\in \\Sigma^{N}\\) and \\(\\pmb{\\gamma} \\in \\Sigma^{M}\\), the simplex of \\(\\mathbb{R}^N\\) and \\(\\mathbb{R}^{M}\\), denote the probability values of the discrete states with satisfaction that \\(\\sum_{i=1}^{N} \\theta_{i} = 1\\) and \\(\\sum_{j=1}^{M} \\gamma_{j} = 1\\). \\(\\delta_{e}\\) refers to a point mass located at coordinate \\(e \\in \\mathbb{R}^{d}\\). The prior modeling of \\(\\pmb{\\theta}\\) and \\(\\pmb{\\gamma}\\) are shown in the Sec. 3.4." + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.662, + 0.906, + 0.692 + ], + "angle": 0, + "content": "To get the transport assignment \\(\\mathbf{Q}\\) between \\(\\Theta\\) and \\(\\Gamma\\), the objective of OT is formulated as the optimization problem:" + }, + { + "type": "equation", + "bbox": [ + 0.529, + 0.702, + 0.906, + 0.736 + ], + "angle": 0, + "content": "\\[\n\\min _ {\\mathbf {Q} \\in \\mathcal {Q}} \\sum_ {i, j} q _ {i j} d _ {i j} - \\varepsilon H (\\mathbf {Q}) \\tag {9}\n\\]" + }, + { + "type": "equation", + "bbox": [ + 0.531, + 0.737, + 0.871, + 0.757 + ], + "angle": 0, + "content": "\\[\n\\text {s . t .} \\mathcal {Q} = \\left\\{\\mathbf {Q} \\in \\mathbb {R} _ {+} ^ {N \\times M} \\mid \\mathbf {Q} \\mathbf {1} _ {M} = \\boldsymbol {\\theta}, \\mathbf {Q} ^ {\\top} \\mathbf {1} _ {N} = \\boldsymbol {\\gamma} \\right\\}.\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.765, + 0.907, + 0.902 + ], + "angle": 0, + "content": "where \\(\\mathcal{Q}\\) is the set of doubly stochastic matrices \\(\\mathbf{Q} = [q_{ij}]_{N\\times M}\\), denotes the corresponding transport assignment where \\(q_{ij}\\) represents the probabilities of aligning \\(p_i\\) with \\(t_j\\). \\(\\mathbf{1}^M\\) is the \\(M\\)-dimensional vector of ones. \\(d_{ij} = d(p_i,t_j) = (1 - \\langle p_i,t_j\\rangle)\\geq 0\\) is the transport cost between the two points \\(p_i\\) and \\(s_j\\) and we use the cosine distance to compute cost matrix \\(\\mathbf{D}\\). The optimal \\(\\mathbf{Q}^*\\) of (9) has a simple normalized exponential matrix solution by Sinkhorn fixed point iterations [7]," + }, + { + "type": "footer", + "bbox": [ + 0.479, + 0.945, + 0.521, + 0.957 + ], + "angle": 0, + "content": "30202" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.151, + 0.089, + 0.429, + 0.271 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.101, + 0.281, + 0.471, + 0.295 + ], + "angle": 0, + "content": "Figure 3. Conditioned co-occurrence matrix of concept labels." + }, + { + "type": "equation", + "bbox": [ + 0.197, + 0.327, + 0.482, + 0.342 + ], + "angle": 0, + "content": "\\[\n\\mathbf {Q} ^ {*} = \\operatorname {S i n k h o r n} (\\mathbf {D}, \\boldsymbol {\\theta}, \\gamma), \\tag {10}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.35, + 0.483, + 0.38 + ], + "angle": 0, + "content": "After obtaining the optimal transport assignment \\(\\mathbf{Q}^*\\), the OT distance between the two sets is expressed as:" + }, + { + "type": "equation", + "bbox": [ + 0.226, + 0.392, + 0.483, + 0.425 + ], + "angle": 0, + "content": "\\[\nd _ {O T} = \\sum_ {i, j} q _ {i j} d _ {i j} \\tag {11}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.436, + 0.484, + 0.511 + ], + "angle": 0, + "content": "For the concept embedding \\( t_j \\), we only consider the minimum distance \\( d_{OT_j} = \\min_i(q_{ij}d_{ij}) \\) between this concept embedding \\( t_j \\) and all patch embeddings to facilitate the concept activation value prediction \\( (x\\rightarrow \\mathcal{C}) \\), the predicted concept activation vector \\( \\hat{\\pmb{a}} \\) are stated as:" + }, + { + "type": "equation", + "bbox": [ + 0.174, + 0.522, + 0.482, + 0.543 + ], + "angle": 0, + "content": "\\[\n\\hat {\\boldsymbol {a}} = \\left\\{\\hat {a} _ {j} \\mid \\hat {a} _ {j} = \\sigma \\left(1 - d _ {O T _ {j}}\\right) \\right\\} _ {j = 1} ^ {M}. \\tag {12}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.553, + 0.484, + 0.599 + ], + "angle": 0, + "content": "where \\(\\sigma\\) is the sigmoid function. Moreover, we can inverse specific concepts to the image patches based on \\(\\mathbf{Q}^*\\), and get the inversion mask \\(m_j\\) of concept \\(c_j\\), as:" + }, + { + "type": "equation", + "bbox": [ + 0.204, + 0.61, + 0.482, + 0.628 + ], + "angle": 0, + "content": "\\[\nm _ {j} = \\left[ q _ {1 j}, q _ {2 j}, \\dots , q _ {N j} \\right]. \\tag {13}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.639, + 0.483, + 0.684 + ], + "angle": 0, + "content": "By utilizing OT distance between patches and concepts as the patch-concept transport, our OT-based binary cross-entropy loss \\(\\mathcal{L}_{\\mathrm{concept}}\\) are as follows:" + }, + { + "type": "equation", + "bbox": [ + 0.119, + 0.696, + 0.483, + 0.713 + ], + "angle": 0, + "content": "\\[\n\\mathcal {L} _ {\\text {c o n c e p t}} = - \\mathbb {E} [ \\boldsymbol {a} \\log (\\hat {\\boldsymbol {a}}) + (1 - \\boldsymbol {a}) \\log (1 - \\hat {\\boldsymbol {a}}) ] \\tag {14}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.091, + 0.724, + 0.474, + 0.74 + ], + "angle": 0, + "content": "Now, we can get the total objective of stage 1 \\((f_{\\mathcal{X}\\to \\mathcal{C}})\\), as:" + }, + { + "type": "equation", + "bbox": [ + 0.132, + 0.752, + 0.483, + 0.77 + ], + "angle": 0, + "content": "\\[\n\\mathcal {L} _ {f, x \\rightarrow c} = \\lambda_ {1} \\mathcal {L} _ {\\mathrm {V O P}} + \\lambda_ {2} \\mathcal {L} _ {\\mathrm {T O P}} + \\lambda_ {3} \\mathcal {L} _ {\\text {c o n c e p t}}, \\tag {15}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.091, + 0.78, + 0.37, + 0.795 + ], + "angle": 0, + "content": "where \\(\\lambda_1, \\lambda_2, \\lambda_3\\) are the hyperparameters." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.796, + 0.484, + 0.902 + ], + "angle": 0, + "content": "The obtained assignment \\(\\mathbf{Q}^{*} = [q_{i,j}]_{N\\times M}\\) indicates the relationship between patches and concepts, where \\(q_{i,j}\\) represent correlation between patch \\(p_i\\) and concept \\(c_{j}\\). Consequently, based on \\(\\mathbf{Q}^*\\), we can obtain concept activation \\(\\pmb{a}\\) (Eq.(12)) based on the image features \\((x\\to \\mathcal{C})\\) and meanwhile inverse the concept to the original image patches (Eq.(13)) \\((\\mathcal{C}\\rightarrow x)\\) to provide explicit interpretation." + }, + { + "type": "title", + "bbox": [ + 0.514, + 0.091, + 0.771, + 0.108 + ], + "angle": 0, + "content": "3.4. Prior distributions Modeling" + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.114, + 0.905, + 0.205 + ], + "angle": 0, + "content": "The prior distributions of \\(\\theta\\) and \\(\\gamma\\) of OT are typically uniform distributions, which is sub-optimal for our task. Considering the part-background spurious correlation and part-part spurious correlation problem mentioned in Fig. 1b, we propose two prior distribution modeling in this section to address these limitations, respectively." + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.206, + 0.906, + 0.34 + ], + "angle": 0, + "content": "Patch set prior \\(\\theta\\). We attribute the part-background spurious correlation phenomenon to the high co-occurrence of some classes and backgrounds, such as \"seabird\" and \"water\". To solve this biased association between object concepts and the background, we use saliency map \\(S_I\\) from Eq. (4) to encourage the transports between the patch features within foreground regions and concepts and penalize the transports of background regions. Specifically, the prior distribution of patch set \\(\\theta\\) are formulated as:" + }, + { + "type": "equation", + "bbox": [ + 0.645, + 0.348, + 0.905, + 0.364 + ], + "angle": 0, + "content": "\\[\n\\boldsymbol {\\theta} = \\operatorname {S o f t m a x} \\left(S _ {I}\\right), \\tag {16}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.371, + 0.906, + 0.49 + ], + "angle": 0, + "content": "Concept set prior \\(\\gamma\\). As illustrated in Fig. 3, some concepts are highly co-occurrent. These concept label biases lead to the model's confusion between these concepts, yielding the part-part spurious correlation phenomenon. To address this problem, we employ the statistical co-occurrence matrix of the training dataset to facilitate the prior distribution modeling of \\(\\gamma\\), aiming to penalize those concepts with high co-occurrence with other concepts." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.491, + 0.906, + 0.626 + ], + "angle": 0, + "content": "Specifically, for a dataset with \\(M\\) concepts and \\(K\\) classes, class-concept label form a matrix \\(L = [l_{i,j}]_{K\\times M}\\) where \\(l_{i,j}\\) represents the activation value of the \\(j\\)-th concept in the concept label for the \\(i\\)-th class. First, we count the co-occurrence frequencies of each pair of concepts. Then, we normalize these counts by the total number of times each concept is activated, resulting in the conditional co-occurrence matrix. The conditional co-occurrence vector \\(r_i\\) is calculated as follows:" + }, + { + "type": "equation", + "bbox": [ + 0.608, + 0.631, + 0.905, + 0.673 + ], + "angle": 0, + "content": "\\[\nr _ {i} = \\sum_ {j = 1} ^ {M} \\left(\\frac {\\sum_ {k = 1} ^ {K} L _ {k i} \\cdot L _ {k j}}{\\sum_ {k = 1} ^ {K} L _ {k i}}\\right) \\tag {17}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.678, + 0.905, + 0.707 + ], + "angle": 0, + "content": "The concept set prior distribution \\(\\gamma\\) is then derived from the conditional co-occurrence vector as follows:" + }, + { + "type": "equation", + "bbox": [ + 0.636, + 0.716, + 0.905, + 0.732 + ], + "angle": 0, + "content": "\\[\n\\gamma = 1 - \\operatorname {S o f t m a x} \\left(r _ {i}\\right) \\tag {18}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.738, + 0.905, + 0.769 + ], + "angle": 0, + "content": "This prior modeling ensures that those concepts with high co-occurrence with other concepts will be penalized." + }, + { + "type": "title", + "bbox": [ + 0.514, + 0.775, + 0.759, + 0.791 + ], + "angle": 0, + "content": "3.5. Concept \\(\\rightarrow\\) class prediction" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.799, + 0.906, + 0.874 + ], + "angle": 0, + "content": "Now, we already finished the first stage which obtains the intermediate concept activation predictions \\( \\pmb{a} \\). In the second stage, our goal is to learn the association between concept and final class prediction similar to vanilla CBM. Specifically, we can map the \\( \\pmb{a} \\) to final output \\( \\hat{y} \\), as:" + }, + { + "type": "equation", + "bbox": [ + 0.665, + 0.882, + 0.905, + 0.897 + ], + "angle": 0, + "content": "\\[\n\\hat {y} = W _ {F} \\cdot \\boldsymbol {a}, \\tag {19}\n\\]" + }, + { + "type": "page_number", + "bbox": [ + 0.48, + 0.945, + 0.518, + 0.957 + ], + "angle": 0, + "content": "30203" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.091, + 0.092, + 0.486, + 0.123 + ], + "angle": 0, + "content": "The classification objective of stage 2 \\((f_{\\mathcal{C}\\to \\mathcal{Y}})\\) is the cross-entropy loss \\(\\mathcal{L}_{\\mathrm{CE}}(y,\\hat{y})\\)" + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.134, + 0.216, + 0.153 + ], + "angle": 0, + "content": "4. Experiment" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.16, + 0.485, + 0.342 + ], + "angle": 0, + "content": "To evaluate the effectiveness of our proposed DOT-CBM, we conduct experiments on the image classification task and part detection task. To further demonstrate our method's abilities to reduce spurious correlations, we also conduct experiments on out-of-distribution generalization tasks. For the concept-to-class linear layer, we omit explicit sparse regularization (e.g., L1 penalties) during training. Despite this, the model spontaneously exhibits significant sparsity in the weight matrix due to two orthogonal losses. Comprehensive training details, training efficiency results, quantitative evaluations of sparsity and additional experiments are listed in the Appendix." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.349, + 0.291, + 0.365 + ], + "angle": 0, + "content": "4.1. Datasets and Metrics" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.373, + 0.483, + 0.614 + ], + "angle": 0, + "content": "For the image classification task. We evaluate our method on the CUB [41] and AwA2 [43] datasets, which are widely adopted in the interpretable community. The CUB dataset contains 11,788 images of 200 bird subcategories, and we use the official train-test split for fine-grained image classification. The AwA2 dataset includes 37,322 images across 50 animal categories, and we follow the same train-test split as in [17]. Additionally, we conduct experiments on the larger and more complex ImageNet-1K[9] and CIFAR-100[19] datasets. On these classification datasets, we leverage the classification accuracy to measure the performance of our method. We also conducted experiments on a more complex dataset (Places365), where the complexity arises from discriminative content being image backgrounds. Quantitative and qualitative results are listed in Appendix." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.615, + 0.483, + 0.902 + ], + "angle": 0, + "content": "For the part detection task. To evaluate the accuracy of the inversion mask produced by DOT-CBM, we evaluate our method on some part detection benchmarks, including CUB dataset, Part-ImageNet [13] and PASCAL-Part [5] datasets, which are widely used for part segmentation tasks. We also test our model on the RIVAL-10 dataset [28], which is specifically designed for evaluating model interpretability. On the CUB dataset, we obtain the bounding boxes for local parts that are commonly used in part detection tasks[44, 51]. The RIVAL-10 dataset consists of over 26K images retrieved from the ImageNet dataset according to the CIFAR-10 [19] categories, and it includes part attribute segmentation masks for the main subjects in the images. The Part-ImageNet and PASCAL-Part datasets are part segmentation datasets that provide fine-grained attribute segmentation masks. We follow [30] and preprocess these datasets by focusing on the animal data and merging overly fine-grained part attribute masks. The detailed data preprocessing procedures are described in the Appendix. We use the" + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.092, + 0.907, + 0.138 + ], + "angle": 0, + "content": "bounding boxes of these masks as the ground truth of part detection. On these datasets, we use the \\( mAP_{0.5} \\) metric to evaluate the precision of predicted concept inversion masks." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.147, + 0.733, + 0.164 + ], + "angle": 0, + "content": "4.2. Implementation Details" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.171, + 0.907, + 0.414 + ], + "angle": 0, + "content": "We employ the pre-trained DINOV2 ViT-L/14 [32] as the image encoder and pre-trained CLIP text encoder [33] as the text encoder. The feature dimension in both encoders is \\( d = 768 \\). We resize all images to \\( H \\times W = 252 \\times 252 \\) as input resolution in both training and testing phases throughout all experiments. For both adapter \\( A_v \\) and \\( A_t \\), we use a three-layer MLP with ReLU activation functions and a dropout rate of 0.4. The optimization of DOT-CBM is done by AdamW with an initialized learning rate of \\( 1e - 4 \\), and decreased every ten epochs by a factor of 0.5. We set weight decay of 10-5. Batch sizes are set as 128 with maximally 200 training epochs. All experiments are conducted on a single RTX 3090 GPU and our model is implemented in PyTorch. We set \\( \\lambda_1 = 0.15 \\), \\( \\lambda_2 = 0.2 \\), \\( \\lambda_3 = 0.8 \\) for CUB dataset. Hyperparameters setting on other datasets and the range ablation study are in the Appendix." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.424, + 0.654, + 0.438 + ], + "angle": 0, + "content": "4.3. Main Results" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.447, + 0.909, + 0.719 + ], + "angle": 0, + "content": "Baselines. We compare our proposed DOT-CBM with several state-of-the-art CBM baselines from the literature. Vanilla-CBM [17] is the first CBM model, proposing the design of the intermediate concept layer. Vanilla-CBM introduces three training strategies, i.e. independent, sequential and joint training, in which the independent training can achieve the best interpretation [26]. Our two-stage optimization belongs to the independent training strategy. Additionally, we compare our performance with recent CBM variants including CEM [48], LaBo [45], SparseCBM [38], CoopCBM [39]. Due to the fact that pre-defined concepts and image backbone will significantly affect the final performance. For a fair comparison, we re-implement all baselines with the high-quality concepts produced by the CDL [47] and a pre-trained ViT backbone to extract image features, the same as our method. The performance of all re-implemented baselines exceeds the best performance reported in their papers." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.72, + 0.909, + 0.855 + ], + "angle": 0, + "content": "Image classification results. As shown in Table 1, we compare our proposed DOT-CBM with all baselines on four common image classification benchmarks. The classification results show that the DOT-CBM achieves superior performance on three benchmarks. Specifically, the performance improvements over the state-of-the-art (SOTA) are 0.91 points on ImageNet, 3.29 points on CUB, and 1.08 points on CIFAR100. Our method only lags behind CoopCBM by 0.25 points on the AwA2 dataset." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.856, + 0.909, + 0.903 + ], + "angle": 0, + "content": "Part detection results. A key property of our proposed DOT-CBM model is able to provide explicit explanations for the concept predictions, via the locate the concept to the" + }, + { + "type": "page_number", + "bbox": [ + 0.48, + 0.945, + 0.521, + 0.958 + ], + "angle": 0, + "content": "30204" + } + ], + [ + { + "type": "table", + "bbox": [ + 0.093, + 0.089, + 0.91, + 0.247 + ], + "angle": 0, + "content": "
MethodClassification Accuracy (↑)Part Detection mAP0.5 (↑)
ImageNetCUBCIFAR100AWA2PartImageNetCUBRIVALPASCAL-Parts
Vanilla-CBM [17]79.1778.3280.0493.1526.9427.0226.7617.32
CEM [48]81.2980.4781.2395.9229.1930.8630.8320.85
LaBo [45]82.9381.3084.1096.9238.2739.2840.3430.17
SparseCBM [38]82.8582.0784.7595.5639.8440.2340.6833.71
CoopCBM [39]82.7382.1084.6697.0841.1745.1643.8235.75
DOT-CBM83.84(+0.91)85.39(+3.29)85.83(+1.08)96.83(-0.25)50.12(+8.95)53.47(+8.31)50.93(+7.11)44.18(+8.43)
" + }, + { + "type": "table_caption", + "bbox": [ + 0.214, + 0.25, + 0.782, + 0.264 + ], + "angle": 0, + "content": "Table 1. Performance comparison of Classification Accuracy (%) and Part Detection \\( \\left( {mA}{P}_{0.5}\\right) \\) ." + }, + { + "type": "table", + "bbox": [ + 0.094, + 0.277, + 0.476, + 0.449 + ], + "angle": 0, + "content": "
MethodCUBDogs
IDOODIDOOD
Vanilla-CBM [17]86.927.787.329.4
CEM [48]84.034.083.836.5
LaBo [45]85.339.485.941.7
Sparse-CBM [38]84.535.885.838.6
Coop-CBM [39]85.836.286.140.3
DOT-CBM w/o. prior86.536.987.040.8
DOT-CBM w. prior87.249.789.152.4
(+0.3)(+10.3)(+1.8)(+10.7)
" + }, + { + "type": "table_caption", + "bbox": [ + 0.115, + 0.451, + 0.459, + 0.464 + ], + "angle": 0, + "content": "Table 2. Performance comparison of OOD generalization." + }, + { + "type": "text", + "bbox": [ + 0.089, + 0.478, + 0.483, + 0.78 + ], + "angle": 0, + "content": "original image space, i.e. the concept inversion mask derived from the transport assignment. This concept inversion mask in the image space indicates the most influential image region contributing to the concept predictions. By contrast, all previous CBMs, using a black-box neuron network to model the mapping from global image features to concept predictions, can only inverse the concept prediction to the image space via the post-hoc Grad-CAM. To evaluate the accuracy of concept inversion mask, we conduct experiments on the part detection datasets and use the groundtruth bounding boxes of the parts to assess the localization performance. To measure the similarity with groundtruths, we need to process our concept inversion mask output and heatmap outputs of baselines (derived from Grad-CAM) into rectangle boxes format. Specifically, we set the threshold at 0.6 of the maximum peak value output of the mask. The local regions above this threshold are considered as the corresponding regions of the concept in the original image. We take the bounding box that encloses this region as the predicted bounding box for the concept." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.78, + 0.483, + 0.902 + ], + "angle": 0, + "content": "From Table. 1, we can observe that DOT-CBM significantly surpasses all baselines (average 8 points improvements), demonstrating the superiority of our design for fine-grained patch-concept alignment. The Coop-CBM achieves performance that is second only to ours due to the introduction of an orthogonality loss similar to our proposed textual orthogonal project \\(\\mathcal{L}_{\\mathrm{TOP}}\\), which further proves the validity of our proposed disentangled OT framework. Consequently," + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.279, + 0.907, + 0.34 + ], + "angle": 0, + "content": "DOT-CBM can inherently produce accurate and explicit explanations for intermediate concept predictions. We also conduct comparison with other ViT-appropriate attribution method. Results shown in Appendix." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.348, + 0.81, + 0.364 + ], + "angle": 0, + "content": "4.4. Out-of-distribution generalization" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.371, + 0.908, + 0.75 + ], + "angle": 0, + "content": "As discussed in the previous sections, there exists data bias in many datasets, leading to shortcut learning. For example, classifying a dog as a huskie based on the snowy background due to the high co-occurrence in the training set, which is analogous to the part-background spurious correlation discussed in Fig. 1b. To validate the effectiveness of our method with background shifts, we follow [39] to segment the foreground object in CUB[41] and Dogs[15] datasets using SAM [16] with bounding box annotations and add a colored background to all images. Each class is associated with a randomly generated color background with an \\(80\\%\\) probability in the training set. The in-distribution (ID) test set contains images with a similar color background probability as the training set, while the color similarity of out-of-distribution (OOD) test set is reduced to \\(30\\%\\). Results are shown in Table 2. As we can see, DOT-CBM can achieve comparable results on ID set and surpass the performance of baselines on the OOD set by a large margin, demonstrating that DOT-CBM can effectively decrease the biased association between object concepts and background. To investigate whether the improvements of tackling part-background spurious correlation are derived from the patch set prior modeling as mentioned in Sec. 3.4, we conduct ablations for the patch set prior and demonstrate the efficacy of patch set prior \\(\\theta\\)." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.757, + 0.681, + 0.772 + ], + "angle": 0, + "content": "4.5. Ablation Studies" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.78, + 0.907, + 0.87 + ], + "angle": 0, + "content": "In this section, we investigate the effects of each component in our framework. Ablation results are listed in Table 3. The baseline has the same trainable networks as DOT-CBM, including two adapters and a linear classification layer. The baseline replaces the OT assignment with a dot product of patch features and concept features." + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.871, + 0.907, + 0.9 + ], + "angle": 0, + "content": "Effect of OT design (OT). For the comparison between the first row and second row in Table. 3, we can find the" + }, + { + "type": "page_number", + "bbox": [ + 0.48, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "30205" + } + ], + [ + { + "type": "table", + "bbox": [ + 0.104, + 0.089, + 0.891, + 0.21 + ], + "angle": 0, + "content": "
OTPriorLrDClassification Accuracy (↑)Part Detection mIOU (↑)
ImageNetCUBCIFAR100AWA2PartImageNetCUBRIVALPASCAL-Parts
XXX81.6080.9282.7396.070.350.410.370.28
✓XX82.0381.7582.9196.240.420.490.420.34
✓✓X83.2284.9285.4996.470.500.610.530.46
XX✓82.7382.1084.6697.080.470.580.490.41
✓✓✓83.8485.3985.8396.830.520.660.540.49
" + }, + { + "type": "table_caption", + "bbox": [ + 0.09, + 0.212, + 0.905, + 0.24 + ], + "angle": 0, + "content": "Table 3. Ablation studies on image classification and part detection. OT, Prior, and LrD denote the optimal transport design, prior distribution modeling, and local representation disentanglement." + }, + { + "type": "image", + "bbox": [ + 0.132, + 0.244, + 0.864, + 0.429 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.089, + 0.431, + 0.908, + 0.46 + ], + "angle": 0, + "content": "Figure 4. Qualitative results for the concept inversion. VanillaCBM and CoopCBM adopt Grad-CAM to provide a concept inversion heatmap, while our proposed DOT-CBM can generate the inversion mask on its own." + }, + { + "type": "text", + "bbox": [ + 0.089, + 0.475, + 0.483, + 0.58 + ], + "angle": 0, + "content": "incorporation of OT distance, mentioned in Sec. 3.3, can slightly improve the classification accuracy and largely enhance the part detection performance. OT provides a better distance metric space for modeling the complex many-to-many correlation between concept features and patch features, compared to the direct dot product similarity used in the baseline model." + }, + { + "type": "text", + "bbox": [ + 0.089, + 0.582, + 0.483, + 0.734 + ], + "angle": 0, + "content": "Effect of prior distribution modeling (Prior). The OT design above, setting the two prior distributions \\(\\theta\\) and \\(\\gamma\\) as a uniform distribution. To mitigate the shortcut learning brought by high occurrence between concept and background and between concept and concept, we add the background penalty patch set prior and the co-occurrence penalty concept set prior, mentioned in Sec. 3.4. Comparing the second and third rows of Table. 3, it demonstrates that these two prior distributions can further improve the classification and interpretability." + }, + { + "type": "text", + "bbox": [ + 0.089, + 0.734, + 0.483, + 0.84 + ], + "angle": 0, + "content": "Effect of local representation disentanglement (LrD). After integrating the local representation disentanglement into the baseline (the fourth row of Table. 3), the improvements are larger than a single incorporation of OT distance (the second row of Table. 3). The overall DOT-CBM, with a combination of the OT, Prior, and LrD, yields more improvements than individual modules alone." + }, + { + "type": "text", + "bbox": [ + 0.089, + 0.84, + 0.484, + 0.902 + ], + "angle": 0, + "content": "Qualitative results. Fig. 4 has shown a qualitative comparison between our method and the baseline methods Vanilla-CBM and CoopCBM. It is observed that our method is able to more accurately locate the region of the" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.475, + 0.907, + 0.596 + ], + "angle": 0, + "content": "concept in the original image compared to the baseline methods. Meanwhile, We also present qualitative results by sequentially ablating each component of our framework. When the local representation disentanglement is ablated, the model's localization accuracy decreases to some extent. Further removal of the priors leads to the emergence of part-background spurious correlations. When the OT design is removed, the model produces incorrect localization results." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.607, + 0.638, + 0.623 + ], + "angle": 0, + "content": "5. Conclusion." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.629, + 0.907, + 0.901 + ], + "angle": 0, + "content": "In this paper, we introduced the Disentangled Optimal Transport Concept Bottleneck Model (DOT-CBM) to address key limitations in existing Concept Bottleneck Models. While CBMs offer interpretable models by mapping image features to human-understandable concepts, they suffer from misalignment between the image and concept, leading to spurious correlations that hinder both interpretability and generalization. To overcome these challenges, DOT-CBM introduces a disentangled optimal transport framework that ensures fine-grained alignment between image patches and corresponding textual concepts, thereby providing a more robust and interpretable model. By employing a dual penalty system, we mitigate the overfitting problem caused by data bias. Empirical results demonstrate that DOT-CBM outperforms previous CBM variants in both classification accuracy and interpretability, offering an effective solution to the problem of visual-concept spurious correlation." + }, + { + "type": "page_number", + "bbox": [ + 0.479, + 0.945, + 0.521, + 0.958 + ], + "angle": 0, + "content": "30206" + } + ], + [ + { + "type": "title", + "bbox": [ + 0.091, + 0.091, + 0.271, + 0.108 + ], + "angle": 0, + "content": "6. Acknowledgement" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.114, + 0.486, + 0.269 + ], + "angle": 0, + "content": "This work was supported in part by the National Natural Science Foundation of China under Grant U21B2006; in part by Shaanxi Youth Innovation Team Project; in part by the Fundamental Research Funds for the Central Universities QTZX24003 and QTZX22160; in part by the 111 Project under Grant B18039; Hao Zhang acknowledges the support of NSFC (62301384); Excellent Young Scientists Fund (Overseas); Foundation of National Key Laboratory of Radar Signal Processing under Grant JKW202308. Zhengjue Wang acknowledges the support of NSFC (62301407)." + }, + { + "type": "title", + "bbox": [ + 0.093, + 0.293, + 0.188, + 0.309 + ], + "angle": 0, + "content": "References" + }, + { + "type": "ref_text", + "bbox": [ + 0.101, + 0.318, + 0.482, + 0.358 + ], + "angle": 0, + "content": "[1] Dosovitskiy Alexey. An image is worth 16x16 words: Transformers for image recognition at scale. arXiv preprint arXiv: 2010.11929, 2020.4" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.361, + 0.483, + 0.415 + ], + "angle": 0, + "content": "[2] Kushal Chauhan, Rishabh Tiwari, Jan Freyberg, Pradeep Shenoy, and Krishnamurthy Dvijotham. Interactive concept bottleneck models. In Proceedings of the AAAI Conference on Artificial Intelligence, pages 5948-5955, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.416, + 0.483, + 0.469 + ], + "angle": 0, + "content": "[3] Guangyi Chen, Weiran Yao, Xiangchen Song, Xinyue Li, Yongming Rao, and Kun Zhang. Plot: Prompt learning with optimal transport for vision-language models. arXiv preprint arXiv:2210.01253, 2022. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.472, + 0.482, + 0.526 + ], + "angle": 0, + "content": "[4] Liquin Chen, Zhe Gan, Yu Cheng, Linjie Li, Lawrence Carin, and Jingjing Liu. Graph optimal transport for cross-domain alignment. In International Conference on Machine Learning, pages 1542-1553. PMLR, 2020. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.528, + 0.482, + 0.609 + ], + "angle": 0, + "content": "[5] Xianjie Chen, Roozbeh Mottaghi, Xiaobai Liu, Sanja Fidler, Raquel Urtasun, and Alan Yuille. Detect what you can: Detecting and representing objects using holistic models and body parts. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 1971-1978, 2014. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.611, + 0.482, + 0.638 + ], + "angle": 0, + "content": "[6] Molnar Christoph. Interpretable machine learning: A guide for making black box models explainable. Leanpub, 2020. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.639, + 0.482, + 0.679 + ], + "angle": 0, + "content": "[7] Marco Cuturi. Sinkhorn distances: Lightspeed computation of optimal transport. Advances in neural information processing systems, 26, 2013. 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.68, + 0.482, + 0.706 + ], + "angle": 0, + "content": "[8] Barry De Ville. Decision trees. Wiley Interdisciplinary Reviews: Computational Statistics, 5(6):448-455, 2013. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.708, + 0.483, + 0.762 + ], + "angle": 0, + "content": "[9] Jia Deng, Wei Dong, Richard Socher, Li-Jia Li, Kai Li, and Li Fei-Fei. Imagenet: A large-scale hierarchical image database. In 2009 IEEE conference on computer vision and pattern recognition, pages 248-255. IEEE, 2009. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.763, + 0.483, + 0.857 + ], + "angle": 0, + "content": "[10] Mateo Espinosa Zarlenga, Pietro Barbiero, Gabriele Ciravegna, Giuseppe Marra, Francesco Giannini, Michelangelo Diligenti, Zohreh Shams, Frederic Precioso, Stefano Melacci, Adrian Weller, et al. Concept embedding models: Beyond the accuracy-explainability trade-off. Advances in Neural Information Processing Systems, 35:21400-21413, 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.859, + 0.483, + 0.902 + ], + "angle": 0, + "content": "[11] Mateo Espinosa Zarlenga, Katie Collins, Krishnamurthy Dvijotham, Adrian Weller, Zohreh Shams, and Mateja Jamnik. Learning to receive help: Intervention-aware concept" + }, + { + "type": "list", + "bbox": [ + 0.094, + 0.318, + 0.483, + 0.902 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.548, + 0.093, + 0.905, + 0.12 + ], + "angle": 0, + "content": "embedding models. Advances in Neural Information Processing Systems, 36, 2024. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.123, + 0.905, + 0.176 + ], + "angle": 0, + "content": "[12] Charlie Frogner, Chiyuan Zhang, Hossein Mobahi, Mauricio Araya, and Tomaso A Poggio. Learning with a wasserstein loss. Advances in neural information processing systems, 28, 2015. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.179, + 0.905, + 0.248 + ], + "angle": 0, + "content": "[13] Ju He, Shuo Yang, Shaokang Yang, Adam Kortylewski, Xiaoding Yuan, Jie-Neng Chen, Shuai Liu, Cheng Yang, Qihang Yu, and Alan Yuille. Partimagenet: A large, high-quality dataset of parts. In European Conference on Computer Vision, pages 128-145. Springer, 2022. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.25, + 0.905, + 0.317 + ], + "angle": 0, + "content": "[14] Johan Huysmans, Karel Dejaeger, Christophe Mues, Jan Vanthienen, and Bart Baesens. An empirical evaluation of the comprehensibility of decision table, tree and rule based predictive models. Decision Support Systems, 51(1):141-154, 2011. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.321, + 0.905, + 0.389 + ], + "angle": 0, + "content": "[15] Aditya Khosla, Nityananda Jayadevaprakash, Bangpeng Yao, and Li Fei-Fei. Novel dataset for fine-grained image categorization. In First Workshop on Fine-Grained Visual Categorization, IEEE Conference on Computer Vision and Pattern Recognition, Colorado Springs, CO, 2011. 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.391, + 0.905, + 0.46 + ], + "angle": 0, + "content": "[16] Alexander Kirillov, Eric Mintun, Nikhila Ravi, Hanzi Mao, Chloe Rolland, Laura Gustafson, Tete Xiao, Spencer Whitehead, Alexander C Berg, Wan-Yen Lo, et al. Segment anything. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 4015-4026, 2023. 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.462, + 0.905, + 0.53 + ], + "angle": 0, + "content": "[17] Pang Wei Koh, Thao Nguyen, Yew Siang Tang, Stephen Mussmann, Emma Pierson, Been Kim, and Percy Liang. Concept bottleneck models. In International conference on machine learning, pages 5338-5348. PMLR, 2020. 1, 2, 3, 6, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.532, + 0.905, + 0.573 + ], + "angle": 0, + "content": "[18] Andrei V Konstantinov and Lev V Utkin. Incorporating expert rules into neural networks in the framework of concept-based learning. arXiv preprint arXiv:2402.14726, 2024. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.575, + 0.905, + 0.603 + ], + "angle": 0, + "content": "[19] Alex Krizhevsky, Geoffrey Hinton, et al. Learning multiple layers of features from tiny images. 2009. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.605, + 0.905, + 0.658 + ], + "angle": 0, + "content": "[20] John Lee, Max Dabagia, Eva Dyer, and Christopher Rozell. Hierarchical optimal transport for multimodal distribution alignment. Advances in neural information processing systems, 32, 2019. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.661, + 0.905, + 0.744 + ], + "angle": 0, + "content": "[21] Miaoge Li, Dongsheng Wang, Xinyang Liu, Zequn Zeng, Ruiying Lu, Bo Chen, and Mingyuan Zhou. Patchct: Aligning patch set and label set with conditional transport for multi-label image classification. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 15348-15358, 2023. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.746, + 0.905, + 0.801 + ], + "angle": 0, + "content": "[22] Tang Li, Mengmeng Ma, and Xi Peng. Deal: Disentangle and localize concept-level explanations for vlms. In European Conference on Computer Vision, pages 383-401. Springer, 2025. 2, 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.803, + 0.905, + 0.856 + ], + "angle": 0, + "content": "[23] Yijie Lin, Jie Zhang, Zhenyu Huang, Jia Liu, Zujie Wen, and Xi Peng. Multi-granularity correspondence learning from long-term noisy videos. arXiv preprint arXiv:2401.16702, 2024. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.859, + 0.905, + 0.902 + ], + "angle": 0, + "content": "[24] Jiawei Ma, Po-Yao Huang, Saining Xie, Shang-Wen Li, Luke Zettlemoyer, Shih-Fu Chang, Wen-Tau Yih, and Hu Xu. Mode: Clip data experts via clustering. In Proceed-" + }, + { + "type": "list", + "bbox": [ + 0.517, + 0.093, + 0.905, + 0.902 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.48, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "30207" + } + ], + [ + { + "type": "ref_text", + "bbox": [ + 0.126, + 0.093, + 0.483, + 0.12 + ], + "angle": 0, + "content": "ings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 26354-26363, 2024. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.121, + 0.483, + 0.188 + ], + "angle": 0, + "content": "[25] Shweta Mahajan, Teresa Botschen, Iryna Gurevych, and Stefan Roth. Joint wasserstein autoencoders for aligning multimodal embeddings. In Proceedings of the IEEE/CVF International Conference on Computer Vision Workshops, pages 0-0, 2019. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.19, + 0.482, + 0.245 + ], + "angle": 0, + "content": "[26] Andrei Margeloiu, Matthew Ashman, Umang Bhatt, Yanzhi Chen, Mateja Jamnik, and Adrian Weller. Do concept bottleneck models learn as intended? arXiv preprint arXiv:2105.04289, 2021. 2, 3, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.247, + 0.482, + 0.302 + ], + "angle": 0, + "content": "[27] Nesta Midavaine, Gregory Hok Tjoan Go, Diego Canez, Ioana Simion, and Satchit Chatterji. [re] on the reproducibility of post-hoc concept bottleneck models. Transactions on Machine Learning Research. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.302, + 0.482, + 0.371 + ], + "angle": 0, + "content": "[28] Mazda Moayeri, Phillip Pope, Yogesh Balaji, and Soheil Feizi. A comprehensive study of image classification model sensitivity to foregrounds, backgrounds, and visual attributes. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, 2022. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.372, + 0.482, + 0.441 + ], + "angle": 0, + "content": "[29] Mazda Moayeri, Keivan Rezaei, Maziar Sanjabi, and Soheil Feizi. Text2concept: Concept activation vectors directly from text. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR) Workshops, pages 3744-3749, 2023. 2, 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.442, + 0.482, + 0.483 + ], + "angle": 0, + "content": "[30] Keval Morabia, Jatin Arora, and Tara Vijaykumar. Attention-based joint detection of object and semantic part. arXiv preprint arXiv:2007.02419, 2020. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.484, + 0.482, + 0.525 + ], + "angle": 0, + "content": "[31] Tuomas Oikarinen, Subhro Das, Lam M Nguyen, and Tsui-Wei Weng. Label-free concept bottleneck models. arXiv preprint arXiv:2304.06129, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.526, + 0.482, + 0.594 + ], + "angle": 0, + "content": "[32] Maxime Oquab, Timothee Darcet, Theo Moutakanni, Huy Vo, Marc Szafraniec, Vasil Khalidov, Pierre Fernandez, Daniel Haziza, Francisco Massa, Alaaeldin El-Nouby, et al. Dinov2: Learning robust visual features without supervision. arXiv preprint arXiv:2304.07193, 2023. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.595, + 0.482, + 0.677 + ], + "angle": 0, + "content": "[33] Alec Radford, Jong Wook Kim, Chris Hallacy, Aditya Ramesh, Gabriel Goh, Sandhini Agarwal, Girish Sastry, Amanda Askell, Pamela Mishkin, Jack Clark, et al. Learning transferable visual models from natural language supervision. In International conference on machine learning, pages 8748-8763. PMLR, 2021. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.679, + 0.482, + 0.746 + ], + "angle": 0, + "content": "[34] Maithra Raghu, Thomas Unterthiner, Simon Kornblith, Chiyuan Zhang, and Alexey Dosovitskiy. Do vision transformers see like convolutional neural networks? Advances in neural information processing systems, 34:12116-12128, 2021. 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.748, + 0.482, + 0.802 + ], + "angle": 0, + "content": "[35] Naveen Raman, Mateo Espinosa Zarlenga, Juyeon Heo, and Mateja Jamnik. Do concept bottleneck models obey locality? In XAI in Action: Past, Present, and Future Applications, 2023. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.804, + 0.482, + 0.871 + ], + "angle": 0, + "content": "[36] Ievgen Redko, Nicolas Courty, Rémi Flamary, and Devis Tuia. Optimal transport for multi-source domain adaptation under target shift. In The 22nd International Conference on artificial intelligence and statistics, pages 849-858. PMLR, 2019. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.873, + 0.482, + 0.901 + ], + "angle": 0, + "content": "[37] Ramprasaath R Selvaraju, Michael Cogswell, Abhishek Das, Ramakrishna Vedantam, Devi Parikh, and Dhruv Batra." + }, + { + "type": "list", + "bbox": [ + 0.093, + 0.093, + 0.483, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.545, + 0.093, + 0.905, + 0.147 + ], + "angle": 0, + "content": "Grad-cam: Visual explanations from deep networks via gradient-based localization. In Proceedings of the IEEE international conference on computer vision, pages 618-626, 2017. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.15, + 0.905, + 0.204 + ], + "angle": 0, + "content": "[38] Andrei Semenov, Vladimir Ivanov, Aleksandr Beznosikov, and Alexander Gasnikov. Sparse concept bottleneck models: Gumbel tricks in contrastive learning. arXiv preprint arXiv:2404.03323, 2024. 2, 6, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.207, + 0.905, + 0.248 + ], + "angle": 0, + "content": "[39] Ivaxi Sheth and Samira Ebrahimi Kahou. Auxiliary losses for learning generalizable concept-based models. Advances in Neural Information Processing Systems, 36, 2024. 2, 6, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.25, + 0.904, + 0.305 + ], + "angle": 0, + "content": "[40] Sungbin Shin, Yohan Jo, Sungsoo Ahn, and Namhoon Lee. A closer look at the intervention procedure of concept bottleneck models. In International Conference on Machine Learning, pages 31504-31520. PMLR, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.307, + 0.904, + 0.347 + ], + "angle": 0, + "content": "[41] Catherine Wah, Steve Branson, Peter Welinder, Pietro Perona, and Serge Belongie. The caltech-ucsd birds-200-2011 dataset. 2011. 6, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.349, + 0.905, + 0.417 + ], + "angle": 0, + "content": "[42] Tiansheng Wen, Yifei Wang, Zequn Zeng, Zhong Peng, Yudi Su, Xinyang Liu, Bo Chen, Hongwei Liu, Stefanie Jegelka, and Chenyu You. Beyond matryoshka: Revisiting sparse coding for adaptive representation. arXiv preprint arXiv:2503.01776, 2025. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.42, + 0.905, + 0.488 + ], + "angle": 0, + "content": "[43] Yongqin Xian, Christoph H Lampert, Bernt Schiele, and Zeynep Akata. Zero-shot learning—a comprehensive evaluation of the good, the bad and the ugly. IEEE transactions on pattern analysis and machine intelligence, 41(9):2251-2265, 2018. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.491, + 0.905, + 0.545 + ], + "angle": 0, + "content": "[44] Wenjia Xu, Yongqin Xian, Jiuniu Wang, Bernt Schiele, and Zeynep Akata. Attribute prototype network for zero-shot learning. Advances in Neural Information Processing Systems, 33:21969-21980, 2020. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.548, + 0.906, + 0.63 + ], + "angle": 0, + "content": "[45] Yue Yang, Artemis Panagopoulou, Shenghao Zhou, Daniel Jin, Chris Callison-Burch, and Mark Yatskar. Language in a bottle: Language model guided concept bottlenecks for interpretable image classification. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 19187-19197, 2023. 2, 4, 6, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.633, + 0.905, + 0.673 + ], + "angle": 0, + "content": "[46] Mert Yuksekgonul, Maggie Wang, and James Zou. Post-hoc concept bottleneck models. arXiv preprint arXiv:2205.15480, 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.676, + 0.905, + 0.716 + ], + "angle": 0, + "content": "[47] Yuan Zang, Tian Yun, Hao Tan, Trung Bui, and Chen Sun. Pre-trained vision-language models learn discoverable visual concepts. arXiv preprint arXiv:2404.12652, 2024. 2, 4, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.718, + 0.905, + 0.8 + ], + "angle": 0, + "content": "[48] Mateo Espinosa Zarlenga, Pietro Barbiero, Gabriele Ciravegna, Giuseppe Marra, Francesco Giannini, Michelangelo Diligenti, Zohreh Shams, Frederic Precioso, Stefano Melacci, Adrian Weller, et al. Concept embedding models: Beyond the accuracy-explainability trade-off. arXiv preprint arXiv:2209.09056, 2022. 6, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.803, + 0.905, + 0.871 + ], + "angle": 0, + "content": "[49] Zequn Zeng, Hao Zhang, Ruiying Lu, Dongsheng Wang, Bo Chen, and Zhengjue Wang. Conzic: Controllable zero-shot image captioning by sampling-based polishing. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 23465-23476, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.874, + 0.905, + 0.901 + ], + "angle": 0, + "content": "[50] Zequn Zeng, Yan Xie, Hao Zhang, Chiyu Chen, Bo Chen, and Zhengjue Wang. Meacap: Memory-augmented zero" + }, + { + "type": "list", + "bbox": [ + 0.517, + 0.093, + 0.906, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.48, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "30208" + } + ], + [ + { + "type": "ref_text", + "bbox": [ + 0.123, + 0.092, + 0.482, + 0.134 + ], + "angle": 0, + "content": "shot image captioning. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 14100-14110, 2024. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.136, + 0.483, + 0.219 + ], + "angle": 0, + "content": "[51] Han Zhang, Tao Xu, Mohamed Elhoseiny, Xiaolei Huang, Shaoting Zhang, Ahmed Elgammal, and Dimitris Metaxas. Spda-cnn: Unifying semantic part detection and abstraction for fine-grained recognition. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 1143-1152, 2016. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.22, + 0.482, + 0.261 + ], + "angle": 0, + "content": "[52] Peng Zhao and Zhi-Hua Zhou. Label distribution learning by optimal transport. In Proceedings of the AAAI Conference on Artificial Intelligence, 2018. 3" + }, + { + "type": "list", + "bbox": [ + 0.093, + 0.092, + 0.483, + 0.261 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.48, + 0.946, + 0.519, + 0.957 + ], + "angle": 0, + "content": "30209" + } + ] +] \ No newline at end of file diff --git a/2025/Discovering Fine-Grained Visual-Concept Relations by Disentangled Optimal Transport Concept Bottleneck Models/353387ac-1b53-4c3c-ac08-91f95db61c80_origin.pdf b/2025/Discovering Fine-Grained Visual-Concept Relations by Disentangled Optimal Transport Concept Bottleneck Models/353387ac-1b53-4c3c-ac08-91f95db61c80_origin.pdf new file mode 100644 index 0000000000000000000000000000000000000000..080538c802821fbd2a8b2d9e82593505dc0b4939 --- /dev/null +++ b/2025/Discovering Fine-Grained Visual-Concept Relations by Disentangled Optimal Transport Concept Bottleneck Models/353387ac-1b53-4c3c-ac08-91f95db61c80_origin.pdf @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:a03d351fbdc723b36b8b9594590b5d87eed0cf2780f86ea463212131e4830ba3 +size 2743001 diff --git a/2025/Discovering Fine-Grained Visual-Concept Relations by Disentangled Optimal Transport Concept Bottleneck Models/full.md b/2025/Discovering Fine-Grained Visual-Concept Relations by Disentangled Optimal Transport Concept Bottleneck Models/full.md new file mode 100644 index 0000000000000000000000000000000000000000..0f6e401f38766a26f562730e8261bfe8851b5dd1 --- /dev/null +++ b/2025/Discovering Fine-Grained Visual-Concept Relations by Disentangled Optimal Transport Concept Bottleneck Models/full.md @@ -0,0 +1,351 @@ +# Discovering Fine-Grained Visual-Concept Relations by Disentangled Optimal Transport Concept Bottleneck Models + +Yan Xie $^{1*}$ Zequn Zeng $^{1*}$ Hao Zhang $^{1\dagger}$ + +Yucheng Ding $^{1}$ Yi Wang $^{1}$ Zhengjue Wang $^{2}$ Bo Chen $^{1}$ Hongwei Liu $^{1}$ + +$^{1}$ National Key Laboratory of Radar Signal Processing, Xidian University, Xi'an, 710071, China +$^{2}$ State Key Laboratory of Integrated Service Networks, Xidian University, Xi'an, 710071, China + +{yanxie0904, zzequn99, zhanghao_xidian}@163.com, bchen@mail.xidian.edu.cn + +# Abstract + +Concept Bottleneck Models (CBMs) try to make the decision-making process transparent by exploring an intermediate concept space between the input image and the output prediction. Existing CBMs just learn coarse-grained relations between the whole image and the concepts, less considering local image information, leading to two main drawbacks: i) they often produce spurious visual-concept relations, hence decreasing model reliability; and ii) though CBMs could explain the importance of every concept to the final prediction, it is still challenging to tell which visual region produces the prediction. To solve these problems, this paper proposes a Disentangled Optimal Transport CBM (DOT-CBM) framework to explore fine-grained visual-concept relations between local image patches and concepts. Specifically, we model the concept prediction process as a transportation problem between the patches and concepts, thereby achieving explicit fine-grained feature alignment. We also incorporate orthogonal projection losses within the modality to enhance local feature disentanglement. To further address the shortcut issues caused by statistical biases in the data, we utilize the visual saliency map and concept label statistics as transportation priors. Thus, DOT-CBM can visualize inversion heatmaps, provide more reliable concept predictions, and produce more accurate class predictions. Comprehensive experiments demonstrate that our proposed DOT-CBM achieves SOTA performance on several tasks, including image classification, local part detection and out-of-distribution generalization. Codes are available in supplementary material. + +# 1. Introduction + +Explainable artificial intelligence aims to allow human users to comprehend and trust the output created by ma + +![](images/1edd78dc0d730e337b22051b57bb7c9bf8da58a82b39a181dce327c0fe149562.jpg) +(a) Brief comparison about the coarse-grained design of vanilla CBM and fine-grained design of our proposed DOT-CBM. + +![](images/1a3a251f1f2573376d3ca7c861920e8d68e2f22ae6017dbcf9c8b667e6458e8b.jpg) +(b) Inversion heatmap visualization. +Figure 1. Comparison between vanilla CBM and our proposed DOT-CBM, including architecture design and inversion heatmap for concept predictions. (a) Due to the black-box mapping from images to concepts, Vanilla CBM needs Grad-CAM [37] techniques to locate concept predictions back to image space while DOT-CBM can provide an explicit inversion heatmap visualization. (b) Due to lack of fine-grained alignment, vanilla CBM produces spurious correlations in three levels of granularity (part-background, part-object and part-part level) that mislocalizes the local concept to the background, the whole object, and incorrect local region. + +chine learning algorithms [6, 14]. Compared with post-hoc explanation methods [37] for black-box models, designing explicit interpretability during modeling and learning processes is beneficial to obtaining more natural explanations [8]. + +The Concept Bottleneck Model (CBM) [17], a representative interpretable image classification model, first predicts a pre-defined set of concepts from images and then + +maps these concept descriptions to the final output class. Therefore, CBM is good at interpreting the relationship between the input image and output class predictions via intermediate human-understandable concepts. However, vanilla CBM brings these good properties at the expense of classification performance [27]. For this problem, a series of CBM variants are proposed from two lines, construct a more comprehensive and discriminative concept bank [31, 45, 46], and sophisticated architecture design [29, 38, 39] for better concept representation learning. + +Although these variants narrow the performance gap between CBMs and their black-box counterparts, their modeling method between images and concepts still has two disadvantages, as shown in Fig. 1a. Firstly, they predict multiple local concepts (i.e. local attributes like "head") from a global image feature, which only learns a coarse-grained relationship between the full image and all concepts. Secondly, the interpretability of prevailing CBMs mainly comes from the linear concept-class relations. The mapping from the original image to all concepts is still modeled as the black-box neural networks, which still need some post-hoc analysis [26, 27] to visualize the learned visual-concept relations. As shown in Fig. 1b, these two disadvantages often bring some visual-concept spurious correlations, where many concept predictions are frequently mis-localized in the images. These spurious correlations inherently undermine the interpretability and generalizing capabilities of CBMs. + +Empirically, we categorized these visual-concept spurious correlations into three levels of granularity: i) part-background spurious correlation. This refers to a misattribution of the concept predictions of local parts to the background. ii) part-object spurious correlation. The concepts of local parts are wrongly associated with the entire object without further distinctions. iii) part-part spurious correlation. In this case, the local concepts are mislocalized to incorrect image regions. We attribute these spurious correlation problems to two key factors. The first one is the lack of fine-grained alignment between the local image regions and corresponding textual concepts, i.e. the fine-grained visual-concept relations. The second one is the overfitting of the data bias such as the statistically high co-occurrence between class and background and between concept and concept, yielding confusion to learn the difference between them. Besides, it is intuitive that the explicit interpretability modeling for the concept predictions from the images is conceivably effective in addressing these visual-concept spurious correlations. + +To this end, we propose a novel Disentangled Optimal Transport Concept Bottleneck Models, namely DOT-CBM. Compared with current CBMs, our proposed DOT-CBM can provide holistic explicit interpretability on not only the concept-class relations but also the visual region + +concept relations. Specifically, we introduce a disentangled optimal transport (DOT) framework to ensure the intramodality disentanglement and inter-modality alignment between the image patches and textual concepts. The transport assignment solution of OT framework, is an explicit correlation matrix between image patches and pre-defined concepts, facilitating bidirectional outputs including concept predictions on concept space and concept inversion mask on the image space. Furthermore, to tackle the overfitting to the data bias, we introduce two discrete prior distributions on the patch set and the concept set to penalize the incorrect shortcut alignment between patches and concepts. In summary, our contributions are as follows: + +- We empirically investigate the visual-concept spurious correlation phenomenon and pinpoint the two key factors of these limitations, i.e. lack of fine-grained visual-concept alignment and overfitting to the data bias. +- We propose a novel disentangled optimal transport framework to model the fine-grained visual-concept alignment and address the overfitting problems. The transport design between image patches and textual concepts provides explicit relations between image and concept predictions, which can produce concept predictions on the concept spaces and concept inversion mask on the image space. +- Extensive experiments on image classification, part detection, and out-of-distribution generalization benchmarks demonstrate our proposed DOT-CBM can achieve significant improvements over current state-of-the-art CBMs on classification accuracy, interpretability and generalizing capabilities to new distributions. + +# 2. Related Work + +Concept Bottleneck Models are a family of models [17] that first predict concepts from input images, then use these concepts to predict a downstream label. This design allows CBMs to (1) provide concept-based explanations via their predicted concepts, and (2) improve their test performance when deployed with experts via concept interventions [2, 11, 40]. Similar to previous interpretable models, CBMs also face the issue of balancing task accuracy and interoperability [10, 27]. Recent works [31, 46] has leveraged extra knowledge graph [38, 46], expert rules [18] or Large Language Models (LLMs) [22, 31, 38, 45, 47] to overcome the limitations of manually defined concept sets and has utilized pre-trained Vision-Language Models (VLMs) [24, 42, 45, 49, 50] to obtain concept predictions. These advancements have elevated the performance of CBMs to levels approaching those of black-box models. + +The interpretability of prevailing CBMs primarily arises from linear relationships between concepts and classes. However, the mapping from the image to the concepts is still handled by black-box neural networks, requiring post-hoc analysis for visualization of the learned visual-concept + +![](images/87885aaab748a7eea38412e2b1636bcedb07341b38479ca7175d9dae9818b4f6.jpg) +Figure 2. Overview of our proposed DOT-CBM. The overall CBM framework proceeds from left to right. In the first stage, the model transforms input images into concept activation vectors. In forward pass, local embeddings from a pre-trained Vision Transformer (ViT) and a text encoder are processed through learnable adapters to generate two feature sets (Sec. 3.2). Two orthogonal projection losses are applied to both feature sets to constrain Adapter training. An Optimal Transport (OT) algorithm optimizes the Assignment Matrix, representing the explicit correlation between concepts and image patches. Concept activation values are derived by combining this matrix with a cost matrix, and the loss is supervised by concept labels for training (Sec. 3.3). To address data bias, we use the Saliency Map from the pre-trained ViT for rough foreground-background separation as the visual prior in OT. A Conditioned Co-occurrence Matrix, based on frequency statistics, serves as the prior for concepts, enhancing the model's ability to distinguish co-occurring concepts (Sec. 3.4). In the second stage, consistent with the vanilla CBM framework, concept activation values are used to predict final class labels via a linear classification network, which is trained using class labels (Sec. 3.5). + +relations. This lack of transparency often leads to visual-concept spurious correlations, where concept predictions are frequently mis-localized in the image. We propose a disentangled OT to address these limitations. + +Optimal Transport aims to find the most efficient way to map one distribution onto another by minimizing a transportation cost. In the context of machine learning, OT is widely applied in a variety of domains, including unsupervised domain adaptation [36], label representation [12, 52], cross-modal semantics [3, 4, 20, 21, 23, 25], and beyond. + +In this work, we further extend OT to a disentangled OT framework, to enforce the fine-grained alignment between image patches and their corresponding concepts while also ensuring disentanglement within each modality. + +# 3. Method + +In this section, we begin with a brief review of the CBM problem setup in Sec. 3.1 including two stages of optimization, i.e. patches to concepts $(f_{\mathcal{X}\rightarrow \mathcal{C}})$ and concepts to classes $(f_{\mathcal{C}\rightarrow \mathcal{Y}})$ . In the first stage, to address the limited interpretation of previous CBMs in the patches-to-concepts process, we introduce a disentangled optimal transport alignment framework between patches and concepts. Concretely, in Sec. 3.2, we propose two orthogonal projection losses to disentangle the patch features and concept features. In Sec. 3.3, we introduce an optimal transport (OT) based framework to model the complex relationship + +between patches and concepts, which can predict concepts from patches and meanwhile inverse the concept to patches to provide explanations of concept prediction. In Sec. 3.4, we integrate two prior distribution modeling to address the data bias problem. In the second stage, in Sec. 3.5, we follow vanilla CBMs and predict classes from concepts. Overall, our proposed DOT-CBM achieves holistic interpretability whether on intermediate concept predictions or on final class predictions. The whole pipeline is shown in Fig. 2. + +# 3.1. Problem Formulation + +Traditional black-box visual models aim to fit the conditional distribution $P(y|x)$ given an a dataset $\mathcal{D} = \{x_i, y_i\}_{i=1}^N$ , where $x_i \in \mathcal{X}, y_i \in \mathcal{Y}$ are images and class labels, respectively. Though effective, they still lack transparency to understand the underlying decision-making process to drive the final predictions. + +Concept Bottleneck Models (CBMs) [17] offer an interpretable solution by first predicting human-interpretable high-level concepts before making the final prediction. Formally, the dataset is extended to $\mathcal{D} = \{x_i,y_i,\pmb {a}_i\}_{i = 1}^N$ , where $\pmb {a}_i = \{a_{ij}|a_{ij}\in \{0,1\} \}_{j = 1}^M$ represents a binary concept activation label. $a_{ij} = 1$ indicates the presence of a specific high-level concept $c_{j}$ in image $x_{i}$ , while $a_{ij} = 0$ means the absence. $c_{j}$ is concept from the pre-defined concept set $\mathcal{C} = \{c_j\}_{j = 1}^M$ . The optimization process of CBMs can be divided into two-stage [26, 35]: I) they first leverage a con + +cept predictor $f_{\mathcal{X}\to \mathcal{C}}:\mathbb{R}^{H\times W\times C}\to \mathbb{R}^M$ to map the input data $x$ to the concept activation vector $\pmb {a}\in \mathbb{R}^{M}$ . 2) Subsequently, a target predictor $f_{\mathcal{C}\rightarrow \mathcal{Y}}:\mathbb{R}^{M}\to \mathbb{R}^{N_{y}}$ maps the concept activation vector $\pmb{a}$ to the final label $y$ . This structure enhances the interpretability of the model, as the intermediate concept predictions provide interpretability into the model's final decision. However, the interpretability from image to concept is un-explored. + +# 3.2. Disentanglement on Local representations + +Disentangled visual patch features $\widetilde{\mathcal{P}}$ . As previously discussed, current CBMs typically associate multiple local concepts with a global image feature, which leads to coarse-grained correspondences. In this paper, we adopt the pretrained vision transformer (ViT) [1] model, which has been approved to exhibit better spatial location information extraction than CNN [34], to extract local patch features to represent the local image regions. + +Specifically, as shown in the left part of Fig. 2, ViT uniformly divides the full image $x \in \mathbb{R}^{H \times W \times C}$ into $N$ non-overlapping patches $\{I_{pi} \in \mathbb{R}^{P_h \times P_w}\}_{i=1}^N$ . After linearly projecting these patches into a $d$ -dimension embedding space and flattening all patch embeddings into a sequence, we can get an initialized patch embedding matrix $X_P \in \mathbb{R}^{N \times d}$ . To extract the global image feature, ViT concatenates a learnable special token [CLS] with $X_P$ . Besides, in order to preserve the spatial information of images, ViT adds a learnable position embedding $X_{pos} \in \mathbb{R}^{(N+1) \times d}$ and we get the input embeddings $X$ for ViT: + +$$ +X _ {P} = \left[ I _ {p 1} \cdot W _ {p}, \dots , I _ {p N} \cdot W _ {p} \right], \tag {1} +$$ + +$$ +X = \left[ \left[ \mathrm {C L S} \right], X _ {P} \right] + X _ {\text {p o s}}, \tag {2} +$$ + +where $W_{p}\in \mathbb{R}^{(P_{h}\cdot P_{w}\cdot C)\times d}$ is the patch embedding projection weights. Then, the pre-trained ViT $E_I$ to map the input embeddings $X$ into the global feature $g_{I}$ , and patch features $\mathcal{P} = \{p_i\}_{i = 1}^N$ : + +$$ +[ g _ {I}, \mathcal {P} ] = E _ {I} (X), \tag {3} +$$ + +We empirically find that the patch features within the object regions are highly similar which is good for object representation, but bad for distinguishing local parts within the object region. Thus, we fit a learnable visual adapter $A_v$ on the $\mathcal{P}$ and map the patch features within the foreground region to a disentangled unit hypersphere by introducing a visual orthogonal projection (VOP) regularizer as: + +$$ +S _ {I} = g _ {I} \cdot P, \tag {4} +$$ + +$$ +\langle A _ {v} (p _ {i}), A _ {v} (p _ {j}) \rangle = \frac {A _ {v} (p _ {i}) \cdot A _ {v} (p _ {j})}{\| A _ {v} (p _ {i}) \| \| A _ {v} (p _ {j}) \|}, \tag {5} +$$ + +$$ +\mathcal {L} _ {\mathrm {V O P}} = \underset {\left(p _ {i}, p _ {j}\right) \sim \mathcal {P}} {\mathbb {E}} \left[ S _ {I} (i) \cdot S _ {I} (j) \cdot \left\langle A _ {v} \left(p _ {i}\right), A _ {v} \left(p _ {j}\right) \right\rangle \right] \tag {6} +$$ + +where $S_{I}$ is the saliency map to distinguish foreground and background, $\langle \cdot ,\cdot \rangle$ is the cosine similarity. Finally, we get the disentangled patch features $\widetilde{\mathcal{P}} = \{\widetilde{p_i} = A_v(p_i)\}_{i = 1}^N$ Disentangled textual concept features $\widetilde{\mathcal{T}}$ . For concept feature extraction, following [22, 29, 45, 47], we utilize a pre-trained text encoder $E_{T}$ to compute concept embeddings $\mathcal{T} = \{t_j = E_T(c_j)\}_{j = 1}^M$ . Similarly, we fit a learnable textual adapter $A_{t}$ and introduce a textual orthogonal projection (TOP) regularizer to disentangle the concept features in a unit hyper-sphere, as: + +$$ +\mathcal {L} _ {\mathrm {T O P}} = \underset {(t _ {i}, t _ {j}) \sim \mathcal {T}} {\mathbb {E}} [ \langle A _ {t} (t _ {i}), A _ {t} (t _ {j}) \rangle ], \tag {7} +$$ + +Consequently, we get the disentangled concept features $\widetilde{\mathcal{T}} = \{\widetilde{t_j} = A_t(t_j)\}_{j = 1}^M.$ + +# 3.3. Patch-concept alignment via optimal transport + +Now, we have the disentangled patch feature set $\widetilde{\mathcal{P}}$ and disentangled concept feature set $\widetilde{\mathcal{T}}$ , we need to model the relationship $\mathbf{Q}$ between these two sets and derive the concept activation vector $\hat{\pmb{a}}$ based on the input image. As a complex many-to-many alignment problem, we propose an optimal transport (OT) based binary cross-entropy loss to tackle these problems. Our goal is to correctly align the patch features and concept features with the same semantics. + +Assume that these two sets $\widetilde{\mathcal{P}}$ and $\widetilde{\mathcal{T}}$ follow two discrete probability distributions $\Theta \in \mathbb{R}^N$ and $\Gamma \in \mathbb{R}^M$ on the $d$ -dimension embedding space, respectively: + +$$ +\boldsymbol {\Theta} = \sum_ {i = 1} ^ {N} \theta_ {i} \delta_ {p _ {i}}, \quad \boldsymbol {\Gamma} = \sum_ {j = 1} ^ {M} \gamma_ {j} \delta_ {c _ {j}}, \tag {8} +$$ + +where $\pmb{\theta} \in \Sigma^{N}$ and $\pmb{\gamma} \in \Sigma^{M}$ , the simplex of $\mathbb{R}^N$ and $\mathbb{R}^{M}$ , denote the probability values of the discrete states with satisfaction that $\sum_{i=1}^{N} \theta_{i} = 1$ and $\sum_{j=1}^{M} \gamma_{j} = 1$ . $\delta_{e}$ refers to a point mass located at coordinate $e \in \mathbb{R}^{d}$ . The prior modeling of $\pmb{\theta}$ and $\pmb{\gamma}$ are shown in the Sec. 3.4. + +To get the transport assignment $\mathbf{Q}$ between $\Theta$ and $\Gamma$ , the objective of OT is formulated as the optimization problem: + +$$ +\min _ {\mathbf {Q} \in \mathcal {Q}} \sum_ {i, j} q _ {i j} d _ {i j} - \varepsilon H (\mathbf {Q}) \tag {9} +$$ + +$$ +\text {s . t .} \mathcal {Q} = \left\{\mathbf {Q} \in \mathbb {R} _ {+} ^ {N \times M} \mid \mathbf {Q} \mathbf {1} _ {M} = \boldsymbol {\theta}, \mathbf {Q} ^ {\top} \mathbf {1} _ {N} = \boldsymbol {\gamma} \right\}. +$$ + +where $\mathcal{Q}$ is the set of doubly stochastic matrices $\mathbf{Q} = [q_{ij}]_{N\times M}$ , denotes the corresponding transport assignment where $q_{ij}$ represents the probabilities of aligning $p_i$ with $t_j$ . $\mathbf{1}^M$ is the $M$ -dimensional vector of ones. $d_{ij} = d(p_i,t_j) = (1 - \langle p_i,t_j\rangle)\geq 0$ is the transport cost between the two points $p_i$ and $s_j$ and we use the cosine distance to compute cost matrix $\mathbf{D}$ . The optimal $\mathbf{Q}^*$ of (9) has a simple normalized exponential matrix solution by Sinkhorn fixed point iterations [7], + +![](images/ea76a7dba2a4f3ab732113d26928c8cbb74c4ad73f1fd3de7a85d95713606a83.jpg) +Figure 3. Conditioned co-occurrence matrix of concept labels. + +$$ +\mathbf {Q} ^ {*} = \operatorname {S i n k h o r n} (\mathbf {D}, \boldsymbol {\theta}, \gamma), \tag {10} +$$ + +After obtaining the optimal transport assignment $\mathbf{Q}^*$ , the OT distance between the two sets is expressed as: + +$$ +d _ {O T} = \sum_ {i, j} q _ {i j} d _ {i j} \tag {11} +$$ + +For the concept embedding $t_j$ , we only consider the minimum distance $d_{OT_j} = \min_i(q_{ij}d_{ij})$ between this concept embedding $t_j$ and all patch embeddings to facilitate the concept activation value prediction $(x\rightarrow \mathcal{C})$ , the predicted concept activation vector $\hat{\pmb{a}}$ are stated as: + +$$ +\hat {\boldsymbol {a}} = \left\{\hat {a} _ {j} \mid \hat {a} _ {j} = \sigma \left(1 - d _ {O T _ {j}}\right) \right\} _ {j = 1} ^ {M}. \tag {12} +$$ + +where $\sigma$ is the sigmoid function. Moreover, we can inverse specific concepts to the image patches based on $\mathbf{Q}^*$ , and get the inversion mask $m_j$ of concept $c_j$ , as: + +$$ +m _ {j} = \left[ q _ {1 j}, q _ {2 j}, \dots , q _ {N j} \right]. \tag {13} +$$ + +By utilizing OT distance between patches and concepts as the patch-concept transport, our OT-based binary cross-entropy loss $\mathcal{L}_{\mathrm{concept}}$ are as follows: + +$$ +\mathcal {L} _ {\text {c o n c e p t}} = - \mathbb {E} [ \boldsymbol {a} \log (\hat {\boldsymbol {a}}) + (1 - \boldsymbol {a}) \log (1 - \hat {\boldsymbol {a}}) ] \tag {14} +$$ + +Now, we can get the total objective of stage 1 $(f_{\mathcal{X}\to \mathcal{C}})$ , as: + +$$ +\mathcal {L} _ {f, x \rightarrow c} = \lambda_ {1} \mathcal {L} _ {\mathrm {V O P}} + \lambda_ {2} \mathcal {L} _ {\mathrm {T O P}} + \lambda_ {3} \mathcal {L} _ {\text {c o n c e p t}}, \tag {15} +$$ + +where $\lambda_1, \lambda_2, \lambda_3$ are the hyperparameters. + +The obtained assignment $\mathbf{Q}^{*} = [q_{i,j}]_{N\times M}$ indicates the relationship between patches and concepts, where $q_{i,j}$ represent correlation between patch $p_i$ and concept $c_{j}$ . Consequently, based on $\mathbf{Q}^*$ , we can obtain concept activation $\pmb{a}$ (Eq.(12)) based on the image features $(x\to \mathcal{C})$ and meanwhile inverse the concept to the original image patches (Eq.(13)) $(\mathcal{C}\rightarrow x)$ to provide explicit interpretation. + +# 3.4. Prior distributions Modeling + +The prior distributions of $\theta$ and $\gamma$ of OT are typically uniform distributions, which is sub-optimal for our task. Considering the part-background spurious correlation and part-part spurious correlation problem mentioned in Fig. 1b, we propose two prior distribution modeling in this section to address these limitations, respectively. + +Patch set prior $\theta$ . We attribute the part-background spurious correlation phenomenon to the high co-occurrence of some classes and backgrounds, such as "seabird" and "water". To solve this biased association between object concepts and the background, we use saliency map $S_I$ from Eq. (4) to encourage the transports between the patch features within foreground regions and concepts and penalize the transports of background regions. Specifically, the prior distribution of patch set $\theta$ are formulated as: + +$$ +\boldsymbol {\theta} = \operatorname {S o f t m a x} \left(S _ {I}\right), \tag {16} +$$ + +Concept set prior $\gamma$ . As illustrated in Fig. 3, some concepts are highly co-occurrent. These concept label biases lead to the model's confusion between these concepts, yielding the part-part spurious correlation phenomenon. To address this problem, we employ the statistical co-occurrence matrix of the training dataset to facilitate the prior distribution modeling of $\gamma$ , aiming to penalize those concepts with high co-occurrence with other concepts. + +Specifically, for a dataset with $M$ concepts and $K$ classes, class-concept label form a matrix $L = [l_{i,j}]_{K\times M}$ where $l_{i,j}$ represents the activation value of the $j$ -th concept in the concept label for the $i$ -th class. First, we count the co-occurrence frequencies of each pair of concepts. Then, we normalize these counts by the total number of times each concept is activated, resulting in the conditional co-occurrence matrix. The conditional co-occurrence vector $r_i$ is calculated as follows: + +$$ +r _ {i} = \sum_ {j = 1} ^ {M} \left(\frac {\sum_ {k = 1} ^ {K} L _ {k i} \cdot L _ {k j}}{\sum_ {k = 1} ^ {K} L _ {k i}}\right) \tag {17} +$$ + +The concept set prior distribution $\gamma$ is then derived from the conditional co-occurrence vector as follows: + +$$ +\gamma = 1 - \operatorname {S o f t m a x} \left(r _ {i}\right) \tag {18} +$$ + +This prior modeling ensures that those concepts with high co-occurrence with other concepts will be penalized. + +# 3.5. Concept $\rightarrow$ class prediction + +Now, we already finished the first stage which obtains the intermediate concept activation predictions $\pmb{a}$ . In the second stage, our goal is to learn the association between concept and final class prediction similar to vanilla CBM. Specifically, we can map the $\pmb{a}$ to final output $\hat{y}$ , as: + +$$ +\hat {y} = W _ {F} \cdot \boldsymbol {a}, \tag {19} +$$ + +The classification objective of stage 2 $(f_{\mathcal{C}\to \mathcal{Y}})$ is the cross-entropy loss $\mathcal{L}_{\mathrm{CE}}(y,\hat{y})$ + +# 4. Experiment + +To evaluate the effectiveness of our proposed DOT-CBM, we conduct experiments on the image classification task and part detection task. To further demonstrate our method's abilities to reduce spurious correlations, we also conduct experiments on out-of-distribution generalization tasks. For the concept-to-class linear layer, we omit explicit sparse regularization (e.g., L1 penalties) during training. Despite this, the model spontaneously exhibits significant sparsity in the weight matrix due to two orthogonal losses. Comprehensive training details, training efficiency results, quantitative evaluations of sparsity and additional experiments are listed in the Appendix. + +# 4.1. Datasets and Metrics + +For the image classification task. We evaluate our method on the CUB [41] and AwA2 [43] datasets, which are widely adopted in the interpretable community. The CUB dataset contains 11,788 images of 200 bird subcategories, and we use the official train-test split for fine-grained image classification. The AwA2 dataset includes 37,322 images across 50 animal categories, and we follow the same train-test split as in [17]. Additionally, we conduct experiments on the larger and more complex ImageNet-1K[9] and CIFAR-100[19] datasets. On these classification datasets, we leverage the classification accuracy to measure the performance of our method. We also conducted experiments on a more complex dataset (Places365), where the complexity arises from discriminative content being image backgrounds. Quantitative and qualitative results are listed in Appendix. + +For the part detection task. To evaluate the accuracy of the inversion mask produced by DOT-CBM, we evaluate our method on some part detection benchmarks, including CUB dataset, Part-ImageNet [13] and PASCAL-Part [5] datasets, which are widely used for part segmentation tasks. We also test our model on the RIVAL-10 dataset [28], which is specifically designed for evaluating model interpretability. On the CUB dataset, we obtain the bounding boxes for local parts that are commonly used in part detection tasks[44, 51]. The RIVAL-10 dataset consists of over 26K images retrieved from the ImageNet dataset according to the CIFAR-10 [19] categories, and it includes part attribute segmentation masks for the main subjects in the images. The Part-ImageNet and PASCAL-Part datasets are part segmentation datasets that provide fine-grained attribute segmentation masks. We follow [30] and preprocess these datasets by focusing on the animal data and merging overly fine-grained part attribute masks. The detailed data preprocessing procedures are described in the Appendix. We use the + +bounding boxes of these masks as the ground truth of part detection. On these datasets, we use the $mAP_{0.5}$ metric to evaluate the precision of predicted concept inversion masks. + +# 4.2. Implementation Details + +We employ the pre-trained DINOV2 ViT-L/14 [32] as the image encoder and pre-trained CLIP text encoder [33] as the text encoder. The feature dimension in both encoders is $d = 768$ . We resize all images to $H \times W = 252 \times 252$ as input resolution in both training and testing phases throughout all experiments. For both adapter $A_v$ and $A_t$ , we use a three-layer MLP with ReLU activation functions and a dropout rate of 0.4. The optimization of DOT-CBM is done by AdamW with an initialized learning rate of $1e - 4$ , and decreased every ten epochs by a factor of 0.5. We set weight decay of 10-5. Batch sizes are set as 128 with maximally 200 training epochs. All experiments are conducted on a single RTX 3090 GPU and our model is implemented in PyTorch. We set $\lambda_1 = 0.15$ , $\lambda_2 = 0.2$ , $\lambda_3 = 0.8$ for CUB dataset. Hyperparameters setting on other datasets and the range ablation study are in the Appendix. + +# 4.3. Main Results + +Baselines. We compare our proposed DOT-CBM with several state-of-the-art CBM baselines from the literature. Vanilla-CBM [17] is the first CBM model, proposing the design of the intermediate concept layer. Vanilla-CBM introduces three training strategies, i.e. independent, sequential and joint training, in which the independent training can achieve the best interpretation [26]. Our two-stage optimization belongs to the independent training strategy. Additionally, we compare our performance with recent CBM variants including CEM [48], LaBo [45], SparseCBM [38], CoopCBM [39]. Due to the fact that pre-defined concepts and image backbone will significantly affect the final performance. For a fair comparison, we re-implement all baselines with the high-quality concepts produced by the CDL [47] and a pre-trained ViT backbone to extract image features, the same as our method. The performance of all re-implemented baselines exceeds the best performance reported in their papers. + +Image classification results. As shown in Table 1, we compare our proposed DOT-CBM with all baselines on four common image classification benchmarks. The classification results show that the DOT-CBM achieves superior performance on three benchmarks. Specifically, the performance improvements over the state-of-the-art (SOTA) are 0.91 points on ImageNet, 3.29 points on CUB, and 1.08 points on CIFAR100. Our method only lags behind CoopCBM by 0.25 points on the AwA2 dataset. + +Part detection results. A key property of our proposed DOT-CBM model is able to provide explicit explanations for the concept predictions, via the locate the concept to the + +
MethodClassification Accuracy (↑)Part Detection mAP0.5 (↑)
ImageNetCUBCIFAR100AWA2PartImageNetCUBRIVALPASCAL-Parts
Vanilla-CBM [17]79.1778.3280.0493.1526.9427.0226.7617.32
CEM [48]81.2980.4781.2395.9229.1930.8630.8320.85
LaBo [45]82.9381.3084.1096.9238.2739.2840.3430.17
SparseCBM [38]82.8582.0784.7595.5639.8440.2340.6833.71
CoopCBM [39]82.7382.1084.6697.0841.1745.1643.8235.75
DOT-CBM83.84(+0.91)85.39(+3.29)85.83(+1.08)96.83(-0.25)50.12(+8.95)53.47(+8.31)50.93(+7.11)44.18(+8.43)
+ +Table 1. Performance comparison of Classification Accuracy (%) and Part Detection $\left( {mA}{P}_{0.5}\right)$ . + +
MethodCUBDogs
IDOODIDOOD
Vanilla-CBM [17]86.927.787.329.4
CEM [48]84.034.083.836.5
LaBo [45]85.339.485.941.7
Sparse-CBM [38]84.535.885.838.6
Coop-CBM [39]85.836.286.140.3
DOT-CBM w/o. prior86.536.987.040.8
DOT-CBM w. prior87.249.789.152.4
(+0.3)(+10.3)(+1.8)(+10.7)
+ +Table 2. Performance comparison of OOD generalization. + +original image space, i.e. the concept inversion mask derived from the transport assignment. This concept inversion mask in the image space indicates the most influential image region contributing to the concept predictions. By contrast, all previous CBMs, using a black-box neuron network to model the mapping from global image features to concept predictions, can only inverse the concept prediction to the image space via the post-hoc Grad-CAM. To evaluate the accuracy of concept inversion mask, we conduct experiments on the part detection datasets and use the groundtruth bounding boxes of the parts to assess the localization performance. To measure the similarity with groundtruths, we need to process our concept inversion mask output and heatmap outputs of baselines (derived from Grad-CAM) into rectangle boxes format. Specifically, we set the threshold at 0.6 of the maximum peak value output of the mask. The local regions above this threshold are considered as the corresponding regions of the concept in the original image. We take the bounding box that encloses this region as the predicted bounding box for the concept. + +From Table. 1, we can observe that DOT-CBM significantly surpasses all baselines (average 8 points improvements), demonstrating the superiority of our design for fine-grained patch-concept alignment. The Coop-CBM achieves performance that is second only to ours due to the introduction of an orthogonality loss similar to our proposed textual orthogonal project $\mathcal{L}_{\mathrm{TOP}}$ , which further proves the validity of our proposed disentangled OT framework. Consequently, + +DOT-CBM can inherently produce accurate and explicit explanations for intermediate concept predictions. We also conduct comparison with other ViT-appropriate attribution method. Results shown in Appendix. + +# 4.4. Out-of-distribution generalization + +As discussed in the previous sections, there exists data bias in many datasets, leading to shortcut learning. For example, classifying a dog as a huskie based on the snowy background due to the high co-occurrence in the training set, which is analogous to the part-background spurious correlation discussed in Fig. 1b. To validate the effectiveness of our method with background shifts, we follow [39] to segment the foreground object in CUB[41] and Dogs[15] datasets using SAM [16] with bounding box annotations and add a colored background to all images. Each class is associated with a randomly generated color background with an $80\%$ probability in the training set. The in-distribution (ID) test set contains images with a similar color background probability as the training set, while the color similarity of out-of-distribution (OOD) test set is reduced to $30\%$ . Results are shown in Table 2. As we can see, DOT-CBM can achieve comparable results on ID set and surpass the performance of baselines on the OOD set by a large margin, demonstrating that DOT-CBM can effectively decrease the biased association between object concepts and background. To investigate whether the improvements of tackling part-background spurious correlation are derived from the patch set prior modeling as mentioned in Sec. 3.4, we conduct ablations for the patch set prior and demonstrate the efficacy of patch set prior $\theta$ . + +# 4.5. Ablation Studies + +In this section, we investigate the effects of each component in our framework. Ablation results are listed in Table 3. The baseline has the same trainable networks as DOT-CBM, including two adapters and a linear classification layer. The baseline replaces the OT assignment with a dot product of patch features and concept features. + +Effect of OT design (OT). For the comparison between the first row and second row in Table. 3, we can find the + +
OTPriorLrDClassification Accuracy (↑)Part Detection mIOU (↑)
ImageNetCUBCIFAR100AWA2PartImageNetCUBRIVALPASCAL-Parts
XXX81.6080.9282.7396.070.350.410.370.28
✓XX82.0381.7582.9196.240.420.490.420.34
✓✓X83.2284.9285.4996.470.500.610.530.46
XX✓82.7382.1084.6697.080.470.580.490.41
✓✓✓83.8485.3985.8396.830.520.660.540.49
+ +Table 3. Ablation studies on image classification and part detection. OT, Prior, and LrD denote the optimal transport design, prior distribution modeling, and local representation disentanglement. + +![](images/1d71a38ad4c7dbfe9cb6d33181883381d0462270320f24079438f568e80b6584.jpg) +Figure 4. Qualitative results for the concept inversion. VanillaCBM and CoopCBM adopt Grad-CAM to provide a concept inversion heatmap, while our proposed DOT-CBM can generate the inversion mask on its own. + +incorporation of OT distance, mentioned in Sec. 3.3, can slightly improve the classification accuracy and largely enhance the part detection performance. OT provides a better distance metric space for modeling the complex many-to-many correlation between concept features and patch features, compared to the direct dot product similarity used in the baseline model. + +Effect of prior distribution modeling (Prior). The OT design above, setting the two prior distributions $\theta$ and $\gamma$ as a uniform distribution. To mitigate the shortcut learning brought by high occurrence between concept and background and between concept and concept, we add the background penalty patch set prior and the co-occurrence penalty concept set prior, mentioned in Sec. 3.4. Comparing the second and third rows of Table. 3, it demonstrates that these two prior distributions can further improve the classification and interpretability. + +Effect of local representation disentanglement (LrD). After integrating the local representation disentanglement into the baseline (the fourth row of Table. 3), the improvements are larger than a single incorporation of OT distance (the second row of Table. 3). The overall DOT-CBM, with a combination of the OT, Prior, and LrD, yields more improvements than individual modules alone. + +Qualitative results. Fig. 4 has shown a qualitative comparison between our method and the baseline methods Vanilla-CBM and CoopCBM. It is observed that our method is able to more accurately locate the region of the + +concept in the original image compared to the baseline methods. Meanwhile, We also present qualitative results by sequentially ablating each component of our framework. When the local representation disentanglement is ablated, the model's localization accuracy decreases to some extent. Further removal of the priors leads to the emergence of part-background spurious correlations. When the OT design is removed, the model produces incorrect localization results. + +# 5. Conclusion. + +In this paper, we introduced the Disentangled Optimal Transport Concept Bottleneck Model (DOT-CBM) to address key limitations in existing Concept Bottleneck Models. While CBMs offer interpretable models by mapping image features to human-understandable concepts, they suffer from misalignment between the image and concept, leading to spurious correlations that hinder both interpretability and generalization. To overcome these challenges, DOT-CBM introduces a disentangled optimal transport framework that ensures fine-grained alignment between image patches and corresponding textual concepts, thereby providing a more robust and interpretable model. By employing a dual penalty system, we mitigate the overfitting problem caused by data bias. Empirical results demonstrate that DOT-CBM outperforms previous CBM variants in both classification accuracy and interpretability, offering an effective solution to the problem of visual-concept spurious correlation. + +# 6. Acknowledgement + +This work was supported in part by the National Natural Science Foundation of China under Grant U21B2006; in part by Shaanxi Youth Innovation Team Project; in part by the Fundamental Research Funds for the Central Universities QTZX24003 and QTZX22160; in part by the 111 Project under Grant B18039; Hao Zhang acknowledges the support of NSFC (62301384); Excellent Young Scientists Fund (Overseas); Foundation of National Key Laboratory of Radar Signal Processing under Grant JKW202308. Zhengjue Wang acknowledges the support of NSFC (62301407). + +# References + +[1] Dosovitskiy Alexey. An image is worth 16x16 words: Transformers for image recognition at scale. arXiv preprint arXiv: 2010.11929, 2020.4 +[2] Kushal Chauhan, Rishabh Tiwari, Jan Freyberg, Pradeep Shenoy, and Krishnamurthy Dvijotham. Interactive concept bottleneck models. In Proceedings of the AAAI Conference on Artificial Intelligence, pages 5948-5955, 2023. 2 +[3] Guangyi Chen, Weiran Yao, Xiangchen Song, Xinyue Li, Yongming Rao, and Kun Zhang. Plot: Prompt learning with optimal transport for vision-language models. arXiv preprint arXiv:2210.01253, 2022. 3 +[4] Liquin Chen, Zhe Gan, Yu Cheng, Linjie Li, Lawrence Carin, and Jingjing Liu. Graph optimal transport for cross-domain alignment. In International Conference on Machine Learning, pages 1542-1553. PMLR, 2020. 3 +[5] Xianjie Chen, Roozbeh Mottaghi, Xiaobai Liu, Sanja Fidler, Raquel Urtasun, and Alan Yuille. Detect what you can: Detecting and representing objects using holistic models and body parts. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 1971-1978, 2014. 6 +[6] Molnar Christoph. Interpretable machine learning: A guide for making black box models explainable. Leanpub, 2020. 1 +[7] Marco Cuturi. Sinkhorn distances: Lightspeed computation of optimal transport. Advances in neural information processing systems, 26, 2013. 4 +[8] Barry De Ville. Decision trees. Wiley Interdisciplinary Reviews: Computational Statistics, 5(6):448-455, 2013. 1 +[9] Jia Deng, Wei Dong, Richard Socher, Li-Jia Li, Kai Li, and Li Fei-Fei. Imagenet: A large-scale hierarchical image database. In 2009 IEEE conference on computer vision and pattern recognition, pages 248-255. IEEE, 2009. 6 +[10] Mateo Espinosa Zarlenga, Pietro Barbiero, Gabriele Ciravegna, Giuseppe Marra, Francesco Giannini, Michelangelo Diligenti, Zohreh Shams, Frederic Precioso, Stefano Melacci, Adrian Weller, et al. Concept embedding models: Beyond the accuracy-explainability trade-off. Advances in Neural Information Processing Systems, 35:21400-21413, 2022. 2 +[11] Mateo Espinosa Zarlenga, Katie Collins, Krishnamurthy Dvijotham, Adrian Weller, Zohreh Shams, and Mateja Jamnik. Learning to receive help: Intervention-aware concept + +embedding models. Advances in Neural Information Processing Systems, 36, 2024. 2 +[12] Charlie Frogner, Chiyuan Zhang, Hossein Mobahi, Mauricio Araya, and Tomaso A Poggio. Learning with a wasserstein loss. Advances in neural information processing systems, 28, 2015. 3 +[13] Ju He, Shuo Yang, Shaokang Yang, Adam Kortylewski, Xiaoding Yuan, Jie-Neng Chen, Shuai Liu, Cheng Yang, Qihang Yu, and Alan Yuille. Partimagenet: A large, high-quality dataset of parts. In European Conference on Computer Vision, pages 128-145. Springer, 2022. 6 +[14] Johan Huysmans, Karel Dejaeger, Christophe Mues, Jan Vanthienen, and Bart Baesens. An empirical evaluation of the comprehensibility of decision table, tree and rule based predictive models. Decision Support Systems, 51(1):141-154, 2011. 1 +[15] Aditya Khosla, Nityananda Jayadevaprakash, Bangpeng Yao, and Li Fei-Fei. Novel dataset for fine-grained image categorization. In First Workshop on Fine-Grained Visual Categorization, IEEE Conference on Computer Vision and Pattern Recognition, Colorado Springs, CO, 2011. 7 +[16] Alexander Kirillov, Eric Mintun, Nikhila Ravi, Hanzi Mao, Chloe Rolland, Laura Gustafson, Tete Xiao, Spencer Whitehead, Alexander C Berg, Wan-Yen Lo, et al. Segment anything. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 4015-4026, 2023. 7 +[17] Pang Wei Koh, Thao Nguyen, Yew Siang Tang, Stephen Mussmann, Emma Pierson, Been Kim, and Percy Liang. Concept bottleneck models. In International conference on machine learning, pages 5338-5348. PMLR, 2020. 1, 2, 3, 6, 7 +[18] Andrei V Konstantinov and Lev V Utkin. Incorporating expert rules into neural networks in the framework of concept-based learning. arXiv preprint arXiv:2402.14726, 2024. 2 +[19] Alex Krizhevsky, Geoffrey Hinton, et al. Learning multiple layers of features from tiny images. 2009. 6 +[20] John Lee, Max Dabagia, Eva Dyer, and Christopher Rozell. Hierarchical optimal transport for multimodal distribution alignment. Advances in neural information processing systems, 32, 2019. 3 +[21] Miaoge Li, Dongsheng Wang, Xinyang Liu, Zequn Zeng, Ruiying Lu, Bo Chen, and Mingyuan Zhou. Patchct: Aligning patch set and label set with conditional transport for multi-label image classification. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 15348-15358, 2023. 3 +[22] Tang Li, Mengmeng Ma, and Xi Peng. Deal: Disentangle and localize concept-level explanations for vlms. In European Conference on Computer Vision, pages 383-401. Springer, 2025. 2, 4 +[23] Yijie Lin, Jie Zhang, Zhenyu Huang, Jia Liu, Zujie Wen, and Xi Peng. Multi-granularity correspondence learning from long-term noisy videos. arXiv preprint arXiv:2401.16702, 2024. 3 +[24] Jiawei Ma, Po-Yao Huang, Saining Xie, Shang-Wen Li, Luke Zettlemoyer, Shih-Fu Chang, Wen-Tau Yih, and Hu Xu. Mode: Clip data experts via clustering. In Proceed- + +ings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 26354-26363, 2024. 2 +[25] Shweta Mahajan, Teresa Botschen, Iryna Gurevych, and Stefan Roth. Joint wasserstein autoencoders for aligning multimodal embeddings. In Proceedings of the IEEE/CVF International Conference on Computer Vision Workshops, pages 0-0, 2019. 3 +[26] Andrei Margeloiu, Matthew Ashman, Umang Bhatt, Yanzhi Chen, Mateja Jamnik, and Adrian Weller. Do concept bottleneck models learn as intended? arXiv preprint arXiv:2105.04289, 2021. 2, 3, 6 +[27] Nesta Midavaine, Gregory Hok Tjoan Go, Diego Canez, Ioana Simion, and Satchit Chatterji. [re] on the reproducibility of post-hoc concept bottleneck models. Transactions on Machine Learning Research. 2 +[28] Mazda Moayeri, Phillip Pope, Yogesh Balaji, and Soheil Feizi. A comprehensive study of image classification model sensitivity to foregrounds, backgrounds, and visual attributes. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, 2022. 6 +[29] Mazda Moayeri, Keivan Rezaei, Maziar Sanjabi, and Soheil Feizi. Text2concept: Concept activation vectors directly from text. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR) Workshops, pages 3744-3749, 2023. 2, 4 +[30] Keval Morabia, Jatin Arora, and Tara Vijaykumar. Attention-based joint detection of object and semantic part. arXiv preprint arXiv:2007.02419, 2020. 6 +[31] Tuomas Oikarinen, Subhro Das, Lam M Nguyen, and Tsui-Wei Weng. Label-free concept bottleneck models. arXiv preprint arXiv:2304.06129, 2023. 2 +[32] Maxime Oquab, Timothee Darcet, Theo Moutakanni, Huy Vo, Marc Szafraniec, Vasil Khalidov, Pierre Fernandez, Daniel Haziza, Francisco Massa, Alaaeldin El-Nouby, et al. Dinov2: Learning robust visual features without supervision. arXiv preprint arXiv:2304.07193, 2023. 6 +[33] Alec Radford, Jong Wook Kim, Chris Hallacy, Aditya Ramesh, Gabriel Goh, Sandhini Agarwal, Girish Sastry, Amanda Askell, Pamela Mishkin, Jack Clark, et al. Learning transferable visual models from natural language supervision. In International conference on machine learning, pages 8748-8763. PMLR, 2021. 6 +[34] Maithra Raghu, Thomas Unterthiner, Simon Kornblith, Chiyuan Zhang, and Alexey Dosovitskiy. Do vision transformers see like convolutional neural networks? Advances in neural information processing systems, 34:12116-12128, 2021. 4 +[35] Naveen Raman, Mateo Espinosa Zarlenga, Juyeon Heo, and Mateja Jamnik. Do concept bottleneck models obey locality? In XAI in Action: Past, Present, and Future Applications, 2023. 3 +[36] Ievgen Redko, Nicolas Courty, Rémi Flamary, and Devis Tuia. Optimal transport for multi-source domain adaptation under target shift. In The 22nd International Conference on artificial intelligence and statistics, pages 849-858. PMLR, 2019. 3 +[37] Ramprasaath R Selvaraju, Michael Cogswell, Abhishek Das, Ramakrishna Vedantam, Devi Parikh, and Dhruv Batra. + +Grad-cam: Visual explanations from deep networks via gradient-based localization. In Proceedings of the IEEE international conference on computer vision, pages 618-626, 2017. 1 +[38] Andrei Semenov, Vladimir Ivanov, Aleksandr Beznosikov, and Alexander Gasnikov. Sparse concept bottleneck models: Gumbel tricks in contrastive learning. arXiv preprint arXiv:2404.03323, 2024. 2, 6, 7 +[39] Ivaxi Sheth and Samira Ebrahimi Kahou. Auxiliary losses for learning generalizable concept-based models. Advances in Neural Information Processing Systems, 36, 2024. 2, 6, 7 +[40] Sungbin Shin, Yohan Jo, Sungsoo Ahn, and Namhoon Lee. A closer look at the intervention procedure of concept bottleneck models. In International Conference on Machine Learning, pages 31504-31520. PMLR, 2023. 2 +[41] Catherine Wah, Steve Branson, Peter Welinder, Pietro Perona, and Serge Belongie. The caltech-ucsd birds-200-2011 dataset. 2011. 6, 7 +[42] Tiansheng Wen, Yifei Wang, Zequn Zeng, Zhong Peng, Yudi Su, Xinyang Liu, Bo Chen, Hongwei Liu, Stefanie Jegelka, and Chenyu You. Beyond matryoshka: Revisiting sparse coding for adaptive representation. arXiv preprint arXiv:2503.01776, 2025. 2 +[43] Yongqin Xian, Christoph H Lampert, Bernt Schiele, and Zeynep Akata. Zero-shot learning—a comprehensive evaluation of the good, the bad and the ugly. IEEE transactions on pattern analysis and machine intelligence, 41(9):2251-2265, 2018. 6 +[44] Wenjia Xu, Yongqin Xian, Jiuniu Wang, Bernt Schiele, and Zeynep Akata. Attribute prototype network for zero-shot learning. Advances in Neural Information Processing Systems, 33:21969-21980, 2020. 6 +[45] Yue Yang, Artemis Panagopoulou, Shenghao Zhou, Daniel Jin, Chris Callison-Burch, and Mark Yatskar. Language in a bottle: Language model guided concept bottlenecks for interpretable image classification. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 19187-19197, 2023. 2, 4, 6, 7 +[46] Mert Yuksekgonul, Maggie Wang, and James Zou. Post-hoc concept bottleneck models. arXiv preprint arXiv:2205.15480, 2022. 2 +[47] Yuan Zang, Tian Yun, Hao Tan, Trung Bui, and Chen Sun. Pre-trained vision-language models learn discoverable visual concepts. arXiv preprint arXiv:2404.12652, 2024. 2, 4, 6 +[48] Mateo Espinosa Zarlenga, Pietro Barbiero, Gabriele Ciravegna, Giuseppe Marra, Francesco Giannini, Michelangelo Diligenti, Zohreh Shams, Frederic Precioso, Stefano Melacci, Adrian Weller, et al. Concept embedding models: Beyond the accuracy-explainability trade-off. arXiv preprint arXiv:2209.09056, 2022. 6, 7 +[49] Zequn Zeng, Hao Zhang, Ruiying Lu, Dongsheng Wang, Bo Chen, and Zhengjue Wang. Conzic: Controllable zero-shot image captioning by sampling-based polishing. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 23465-23476, 2023. 2 +[50] Zequn Zeng, Yan Xie, Hao Zhang, Chiyu Chen, Bo Chen, and Zhengjue Wang. Meacap: Memory-augmented zero + +shot image captioning. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 14100-14110, 2024. 2 +[51] Han Zhang, Tao Xu, Mohamed Elhoseiny, Xiaolei Huang, Shaoting Zhang, Ahmed Elgammal, and Dimitris Metaxas. Spda-cnn: Unifying semantic part detection and abstraction for fine-grained recognition. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 1143-1152, 2016. 6 +[52] Peng Zhao and Zhi-Hua Zhou. Label distribution learning by optimal transport. In Proceedings of the AAAI Conference on Artificial Intelligence, 2018. 3 \ No newline at end of file diff --git a/2025/Discovering Fine-Grained Visual-Concept Relations by Disentangled Optimal Transport Concept Bottleneck Models/images.zip b/2025/Discovering Fine-Grained Visual-Concept Relations by Disentangled Optimal Transport Concept Bottleneck Models/images.zip new file mode 100644 index 0000000000000000000000000000000000000000..4f795093bf1425c1c633e48227e7d7ad34301c20 --- /dev/null +++ b/2025/Discovering Fine-Grained Visual-Concept Relations by Disentangled Optimal Transport Concept Bottleneck Models/images.zip @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:044c0894bc247731ee883114f2b7ae6a5d68d81d0f6c7633961df2e8aeacc550 +size 490121 diff --git a/2025/Discovering Fine-Grained Visual-Concept Relations by Disentangled Optimal Transport Concept Bottleneck Models/layout.json b/2025/Discovering Fine-Grained Visual-Concept Relations by Disentangled Optimal Transport Concept Bottleneck Models/layout.json new file mode 100644 index 0000000000000000000000000000000000000000..f0548e3c29f0e87273689a416963c6b24765b5f4 --- /dev/null +++ b/2025/Discovering Fine-Grained Visual-Concept Relations by Disentangled Optimal Transport Concept Bottleneck Models/layout.json @@ -0,0 +1,9784 @@ +{ + "pdf_info": [ + { + "para_blocks": [ + { + "bbox": [ + 67, + 102, + 544, + 140 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 67, + 102, + 544, + 140 + ], + "spans": [ + { + "bbox": [ + 67, + 102, + 544, + 140 + ], + "type": "text", + "content": "Discovering Fine-Grained Visual-Concept Relations by Disentangled Optimal Transport Concept Bottleneck Models" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 209, + 160, + 405, + 175 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 209, + 160, + 405, + 175 + ], + "spans": [ + { + "bbox": [ + 209, + 160, + 405, + 175 + ], + "type": "text", + "content": "Yan Xie" + }, + { + "bbox": [ + 209, + 160, + 405, + 175 + ], + "type": "inline_equation", + "content": "^{1*}" + }, + { + "bbox": [ + 209, + 160, + 405, + 175 + ], + "type": "text", + "content": " Zequn Zeng" + }, + { + "bbox": [ + 209, + 160, + 405, + 175 + ], + "type": "inline_equation", + "content": "^{1*}" + }, + { + "bbox": [ + 209, + 160, + 405, + 175 + ], + "type": "text", + "content": " Hao Zhang" + }, + { + "bbox": [ + 209, + 160, + 405, + 175 + ], + "type": "inline_equation", + "content": "^{1\\dagger}" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 121, + 175, + 487, + 190 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 121, + 175, + 487, + 190 + ], + "spans": [ + { + "bbox": [ + 121, + 175, + 487, + 190 + ], + "type": "text", + "content": "Yucheng Ding" + }, + { + "bbox": [ + 121, + 175, + 487, + 190 + ], + "type": "inline_equation", + "content": "^{1}" + }, + { + "bbox": [ + 121, + 175, + 487, + 190 + ], + "type": "text", + "content": " Yi Wang" + }, + { + "bbox": [ + 121, + 175, + 487, + 190 + ], + "type": "inline_equation", + "content": "^{1}" + }, + { + "bbox": [ + 121, + 175, + 487, + 190 + ], + "type": "text", + "content": " Zhengjue Wang" + }, + { + "bbox": [ + 121, + 175, + 487, + 190 + ], + "type": "inline_equation", + "content": "^{2}" + }, + { + "bbox": [ + 121, + 175, + 487, + 190 + ], + "type": "text", + "content": " Bo Chen" + }, + { + "bbox": [ + 121, + 175, + 487, + 190 + ], + "type": "inline_equation", + "content": "^{1}" + }, + { + "bbox": [ + 121, + 175, + 487, + 190 + ], + "type": "text", + "content": " Hongwei Liu" + }, + { + "bbox": [ + 121, + 175, + 487, + 190 + ], + "type": "inline_equation", + "content": "^{1}" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 74, + 190, + 536, + 217 + ], + "type": "list", + "angle": 0, + "index": 7, + "blocks": [ + { + "bbox": [ + 74, + 190, + 535, + 203 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 74, + 190, + 535, + 203 + ], + "spans": [ + { + "bbox": [ + 74, + 190, + 535, + 203 + ], + "type": "inline_equation", + "content": "^{1}" + }, + { + "bbox": [ + 74, + 190, + 535, + 203 + ], + "type": "text", + "content": "National Key Laboratory of Radar Signal Processing, Xidian University, Xi'an, 710071, China" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 74, + 203, + 536, + 217 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 74, + 203, + 536, + 217 + ], + "spans": [ + { + "bbox": [ + 74, + 203, + 536, + 217 + ], + "type": "inline_equation", + "content": "^{2}" + }, + { + "bbox": [ + 74, + 203, + 536, + 217 + ], + "type": "text", + "content": "State Key Laboratory of Integrated Service Networks, Xidian University, Xi'an, 710071, China" + } + ] + } + ], + "index": 6 + } + ], + "sub_type": "text" + }, + { + "bbox": [ + 108, + 220, + 501, + 232 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 108, + 220, + 501, + 232 + ], + "spans": [ + { + "bbox": [ + 108, + 220, + 501, + 232 + ], + "type": "text", + "content": "{yanxie0904, zzequn99, zhanghao_xidian}@163.com, bchen@mail.xidian.edu.cn" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 151, + 258, + 200, + 271 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 151, + 258, + 200, + 271 + ], + "spans": [ + { + "bbox": [ + 151, + 258, + 200, + 271 + ], + "type": "text", + "content": "Abstract" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 55, + 283, + 296, + 631 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 283, + 296, + 631 + ], + "spans": [ + { + "bbox": [ + 55, + 283, + 296, + 631 + ], + "type": "text", + "content": "Concept Bottleneck Models (CBMs) try to make the decision-making process transparent by exploring an intermediate concept space between the input image and the output prediction. Existing CBMs just learn coarse-grained relations between the whole image and the concepts, less considering local image information, leading to two main drawbacks: i) they often produce spurious visual-concept relations, hence decreasing model reliability; and ii) though CBMs could explain the importance of every concept to the final prediction, it is still challenging to tell which visual region produces the prediction. To solve these problems, this paper proposes a Disentangled Optimal Transport CBM (DOT-CBM) framework to explore fine-grained visual-concept relations between local image patches and concepts. Specifically, we model the concept prediction process as a transportation problem between the patches and concepts, thereby achieving explicit fine-grained feature alignment. We also incorporate orthogonal projection losses within the modality to enhance local feature disentanglement. To further address the shortcut issues caused by statistical biases in the data, we utilize the visual saliency map and concept label statistics as transportation priors. Thus, DOT-CBM can visualize inversion heatmaps, provide more reliable concept predictions, and produce more accurate class predictions. Comprehensive experiments demonstrate that our proposed DOT-CBM achieves SOTA performance on several tasks, including image classification, local part detection and out-of-distribution generalization. Codes are available in supplementary material." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 56, + 652, + 135, + 664 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 652, + 135, + 664 + ], + "spans": [ + { + "bbox": [ + 56, + 652, + 135, + 664 + ], + "type": "text", + "content": "1. Introduction" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 55, + 672, + 295, + 696 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 672, + 295, + 696 + ], + "spans": [ + { + "bbox": [ + 55, + 672, + 295, + 696 + ], + "type": "text", + "content": "Explainable artificial intelligence aims to allow human users to comprehend and trust the output created by ma" + } + ] + } + ], + "index": 12 + }, + { + "type": "image", + "bbox": [ + 328, + 258, + 543, + 334 + ], + "blocks": [ + { + "bbox": [ + 328, + 258, + 543, + 334 + ], + "lines": [ + { + "bbox": [ + 328, + 258, + 543, + 334 + ], + "spans": [ + { + "bbox": [ + 328, + 258, + 543, + 334 + ], + "type": "image", + "image_path": "1edd78dc0d730e337b22051b57bb7c9bf8da58a82b39a181dce327c0fe149562.jpg" + } + ] + } + ], + "index": 13, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 326, + 335, + 542, + 354 + ], + "lines": [ + { + "bbox": [ + 326, + 335, + 542, + 354 + ], + "spans": [ + { + "bbox": [ + 326, + 335, + 542, + 354 + ], + "type": "text", + "content": "(a) Brief comparison about the coarse-grained design of vanilla CBM and fine-grained design of our proposed DOT-CBM." + } + ] + } + ], + "index": 14, + "angle": 0, + "type": "image_caption" + } + ], + "index": 13 + }, + { + "type": "image", + "bbox": [ + 326, + 364, + 541, + 468 + ], + "blocks": [ + { + "bbox": [ + 326, + 364, + 541, + 468 + ], + "lines": [ + { + "bbox": [ + 326, + 364, + 541, + 468 + ], + "spans": [ + { + "bbox": [ + 326, + 364, + 541, + 468 + ], + "type": "image", + "image_path": "1a3a251f1f2573376d3ca7c861920e8d68e2f22ae6017dbcf9c8b667e6458e8b.jpg" + } + ] + } + ], + "index": 15, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 374, + 469, + 491, + 479 + ], + "lines": [ + { + "bbox": [ + 374, + 469, + 491, + 479 + ], + "spans": [ + { + "bbox": [ + 374, + 469, + 491, + 479 + ], + "type": "text", + "content": "(b) Inversion heatmap visualization." + } + ] + } + ], + "index": 16, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 313, + 480, + 555, + 601 + ], + "lines": [ + { + "bbox": [ + 313, + 480, + 555, + 601 + ], + "spans": [ + { + "bbox": [ + 313, + 480, + 555, + 601 + ], + "type": "text", + "content": "Figure 1. Comparison between vanilla CBM and our proposed DOT-CBM, including architecture design and inversion heatmap for concept predictions. (a) Due to the black-box mapping from images to concepts, Vanilla CBM needs Grad-CAM [37] techniques to locate concept predictions back to image space while DOT-CBM can provide an explicit inversion heatmap visualization. (b) Due to lack of fine-grained alignment, vanilla CBM produces spurious correlations in three levels of granularity (part-background, part-object and part-part level) that mislocalizes the local concept to the background, the whole object, and incorrect local region." + } + ] + } + ], + "index": 17, + "angle": 0, + "type": "image_caption" + } + ], + "index": 15 + }, + { + "bbox": [ + 313, + 616, + 555, + 676 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 616, + 555, + 676 + ], + "spans": [ + { + "bbox": [ + 313, + 616, + 555, + 676 + ], + "type": "text", + "content": "chine learning algorithms [6, 14]. Compared with post-hoc explanation methods [37] for black-box models, designing explicit interpretability during modeling and learning processes is beneficial to obtaining more natural explanations [8]." + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 313, + 677, + 555, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 677, + 555, + 713 + ], + "spans": [ + { + "bbox": [ + 313, + 677, + 555, + 713 + ], + "type": "text", + "content": "The Concept Bottleneck Model (CBM) [17], a representative interpretable image classification model, first predicts a pre-defined set of concepts from images and then" + } + ] + } + ], + "index": 19 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "spans": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "type": "text", + "content": "CVF" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 145, + 0, + 496, + 37 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 145, + 0, + 496, + 37 + ], + "spans": [ + { + "bbox": [ + 145, + 0, + 496, + 37 + ], + "type": "text", + "content": "This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 66, + 702, + 228, + 713 + ], + "type": "page_footnote", + "angle": 0, + "lines": [ + { + "bbox": [ + 66, + 702, + 228, + 713 + ], + "spans": [ + { + "bbox": [ + 66, + 702, + 228, + 713 + ], + "type": "text", + "content": "*Equal contribution. †Corresponding authors" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "type": "text", + "content": "30199" + } + ] + } + ], + "index": 21 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 0 + }, + { + "para_blocks": [ + { + "bbox": [ + 55, + 72, + 294, + 192 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 72, + 294, + 192 + ], + "spans": [ + { + "bbox": [ + 55, + 72, + 294, + 192 + ], + "type": "text", + "content": "maps these concept descriptions to the final output class. Therefore, CBM is good at interpreting the relationship between the input image and output class predictions via intermediate human-understandable concepts. However, vanilla CBM brings these good properties at the expense of classification performance [27]. For this problem, a series of CBM variants are proposed from two lines, construct a more comprehensive and discriminative concept bank [31, 45, 46], and sophisticated architecture design [29, 38, 39] for better concept representation learning." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 56, + 194, + 296, + 410 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 194, + 296, + 410 + ], + "spans": [ + { + "bbox": [ + 56, + 194, + 296, + 410 + ], + "type": "text", + "content": "Although these variants narrow the performance gap between CBMs and their black-box counterparts, their modeling method between images and concepts still has two disadvantages, as shown in Fig. 1a. Firstly, they predict multiple local concepts (i.e. local attributes like \"head\") from a global image feature, which only learns a coarse-grained relationship between the full image and all concepts. Secondly, the interpretability of prevailing CBMs mainly comes from the linear concept-class relations. The mapping from the original image to all concepts is still modeled as the black-box neural networks, which still need some post-hoc analysis [26, 27] to visualize the learned visual-concept relations. As shown in Fig. 1b, these two disadvantages often bring some visual-concept spurious correlations, where many concept predictions are frequently mis-localized in the images. These spurious correlations inherently undermine the interpretability and generalizing capabilities of CBMs." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 55, + 412, + 296, + 651 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 412, + 296, + 651 + ], + "spans": [ + { + "bbox": [ + 55, + 412, + 296, + 651 + ], + "type": "text", + "content": "Empirically, we categorized these visual-concept spurious correlations into three levels of granularity: i) part-background spurious correlation. This refers to a misattribution of the concept predictions of local parts to the background. ii) part-object spurious correlation. The concepts of local parts are wrongly associated with the entire object without further distinctions. iii) part-part spurious correlation. In this case, the local concepts are mislocalized to incorrect image regions. We attribute these spurious correlation problems to two key factors. The first one is the lack of fine-grained alignment between the local image regions and corresponding textual concepts, i.e. the fine-grained visual-concept relations. The second one is the overfitting of the data bias such as the statistically high co-occurrence between class and background and between concept and concept, yielding confusion to learn the difference between them. Besides, it is intuitive that the explicit interpretability modeling for the concept predictions from the images is conceivably effective in addressing these visual-concept spurious correlations." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 55, + 653, + 296, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 653, + 296, + 715 + ], + "spans": [ + { + "bbox": [ + 55, + 653, + 296, + 715 + ], + "type": "text", + "content": "To this end, we propose a novel Disentangled Optimal Transport Concept Bottleneck Models, namely DOT-CBM. Compared with current CBMs, our proposed DOT-CBM can provide holistic explicit interpretability on not only the concept-class relations but also the visual region" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 313, + 72, + 553, + 227 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 72, + 553, + 227 + ], + "spans": [ + { + "bbox": [ + 313, + 72, + 553, + 227 + ], + "type": "text", + "content": "concept relations. Specifically, we introduce a disentangled optimal transport (DOT) framework to ensure the intramodality disentanglement and inter-modality alignment between the image patches and textual concepts. The transport assignment solution of OT framework, is an explicit correlation matrix between image patches and pre-defined concepts, facilitating bidirectional outputs including concept predictions on concept space and concept inversion mask on the image space. Furthermore, to tackle the overfitting to the data bias, we introduce two discrete prior distributions on the patch set and the concept set to penalize the incorrect shortcut alignment between patches and concepts. In summary, our contributions are as follows:" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 313, + 228, + 555, + 431 + ], + "type": "list", + "angle": 0, + "index": 8, + "blocks": [ + { + "bbox": [ + 314, + 228, + 553, + 275 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 314, + 228, + 553, + 275 + ], + "spans": [ + { + "bbox": [ + 314, + 228, + 553, + 275 + ], + "type": "text", + "content": "- We empirically investigate the visual-concept spurious correlation phenomenon and pinpoint the two key factors of these limitations, i.e. lack of fine-grained visual-concept alignment and overfitting to the data bias." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 313, + 276, + 553, + 360 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 276, + 553, + 360 + ], + "spans": [ + { + "bbox": [ + 313, + 276, + 553, + 360 + ], + "type": "text", + "content": "- We propose a novel disentangled optimal transport framework to model the fine-grained visual-concept alignment and address the overfitting problems. The transport design between image patches and textual concepts provides explicit relations between image and concept predictions, which can produce concept predictions on the concept spaces and concept inversion mask on the image space." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 313, + 360, + 555, + 431 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 360, + 555, + 431 + ], + "spans": [ + { + "bbox": [ + 313, + 360, + 555, + 431 + ], + "type": "text", + "content": "- Extensive experiments on image classification, part detection, and out-of-distribution generalization benchmarks demonstrate our proposed DOT-CBM can achieve significant improvements over current state-of-the-art CBMs on classification accuracy, interpretability and generalizing capabilities to new distributions." + } + ] + } + ], + "index": 7 + } + ], + "sub_type": "text" + }, + { + "bbox": [ + 314, + 442, + 402, + 455 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 314, + 442, + 402, + 455 + ], + "spans": [ + { + "bbox": [ + 314, + 442, + 402, + 455 + ], + "type": "text", + "content": "2. Related Work" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 313, + 462, + 555, + 653 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 462, + 555, + 653 + ], + "spans": [ + { + "bbox": [ + 313, + 462, + 555, + 653 + ], + "type": "text", + "content": "Concept Bottleneck Models are a family of models [17] that first predict concepts from input images, then use these concepts to predict a downstream label. This design allows CBMs to (1) provide concept-based explanations via their predicted concepts, and (2) improve their test performance when deployed with experts via concept interventions [2, 11, 40]. Similar to previous interpretable models, CBMs also face the issue of balancing task accuracy and interoperability [10, 27]. Recent works [31, 46] has leveraged extra knowledge graph [38, 46], expert rules [18] or Large Language Models (LLMs) [22, 31, 38, 45, 47] to overcome the limitations of manually defined concept sets and has utilized pre-trained Vision-Language Models (VLMs) [24, 42, 45, 49, 50] to obtain concept predictions. These advancements have elevated the performance of CBMs to levels approaching those of black-box models." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 313, + 654, + 556, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 654, + 556, + 715 + ], + "spans": [ + { + "bbox": [ + 313, + 654, + 556, + 715 + ], + "type": "text", + "content": "The interpretability of prevailing CBMs primarily arises from linear relationships between concepts and classes. However, the mapping from the image to the concepts is still handled by black-box neural networks, requiring post-hoc analysis for visualization of the learned visual-concept" + } + ] + } + ], + "index": 11 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "type": "text", + "content": "30200" + } + ] + } + ], + "index": 12 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 1 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 59, + 71, + 553, + 247 + ], + "blocks": [ + { + "bbox": [ + 59, + 71, + 553, + 247 + ], + "lines": [ + { + "bbox": [ + 59, + 71, + 553, + 247 + ], + "spans": [ + { + "bbox": [ + 59, + 71, + 553, + 247 + ], + "type": "image", + "image_path": "87885aaab748a7eea38412e2b1636bcedb07341b38479ca7175d9dae9818b4f6.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 55, + 253, + 555, + 365 + ], + "lines": [ + { + "bbox": [ + 55, + 253, + 555, + 365 + ], + "spans": [ + { + "bbox": [ + 55, + 253, + 555, + 365 + ], + "type": "text", + "content": "Figure 2. Overview of our proposed DOT-CBM. The overall CBM framework proceeds from left to right. In the first stage, the model transforms input images into concept activation vectors. In forward pass, local embeddings from a pre-trained Vision Transformer (ViT) and a text encoder are processed through learnable adapters to generate two feature sets (Sec. 3.2). Two orthogonal projection losses are applied to both feature sets to constrain Adapter training. An Optimal Transport (OT) algorithm optimizes the Assignment Matrix, representing the explicit correlation between concepts and image patches. Concept activation values are derived by combining this matrix with a cost matrix, and the loss is supervised by concept labels for training (Sec. 3.3). To address data bias, we use the Saliency Map from the pre-trained ViT for rough foreground-background separation as the visual prior in OT. A Conditioned Co-occurrence Matrix, based on frequency statistics, serves as the prior for concepts, enhancing the model's ability to distinguish co-occurring concepts (Sec. 3.4). In the second stage, consistent with the vanilla CBM framework, concept activation values are used to predict final class labels via a linear classification network, which is trained using class labels (Sec. 3.5)." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "bbox": [ + 55, + 384, + 295, + 432 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 384, + 295, + 432 + ], + "spans": [ + { + "bbox": [ + 55, + 384, + 295, + 432 + ], + "type": "text", + "content": "relations. This lack of transparency often leads to visual-concept spurious correlations, where concept predictions are frequently mis-localized in the image. We propose a disentangled OT to address these limitations." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 55, + 433, + 295, + 504 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 433, + 295, + 504 + ], + "spans": [ + { + "bbox": [ + 55, + 433, + 295, + 504 + ], + "type": "text", + "content": "Optimal Transport aims to find the most efficient way to map one distribution onto another by minimizing a transportation cost. In the context of machine learning, OT is widely applied in a variety of domains, including unsupervised domain adaptation [36], label representation [12, 52], cross-modal semantics [3, 4, 20, 21, 23, 25], and beyond." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 55, + 504, + 296, + 552 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 504, + 296, + 552 + ], + "spans": [ + { + "bbox": [ + 55, + 504, + 296, + 552 + ], + "type": "text", + "content": "In this work, we further extend OT to a disentangled OT framework, to enforce the fine-grained alignment between image patches and their corresponding concepts while also ensuring disentanglement within each modality." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 55, + 562, + 111, + 574 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 562, + 111, + 574 + ], + "spans": [ + { + "bbox": [ + 55, + 562, + 111, + 574 + ], + "type": "text", + "content": "3. Method" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 55, + 582, + 295, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 582, + 295, + 714 + ], + "spans": [ + { + "bbox": [ + 55, + 582, + 295, + 714 + ], + "type": "text", + "content": "In this section, we begin with a brief review of the CBM problem setup in Sec. 3.1 including two stages of optimization, i.e. patches to concepts " + }, + { + "bbox": [ + 55, + 582, + 295, + 714 + ], + "type": "inline_equation", + "content": "(f_{\\mathcal{X}\\rightarrow \\mathcal{C}})" + }, + { + "bbox": [ + 55, + 582, + 295, + 714 + ], + "type": "text", + "content": " and concepts to classes " + }, + { + "bbox": [ + 55, + 582, + 295, + 714 + ], + "type": "inline_equation", + "content": "(f_{\\mathcal{C}\\rightarrow \\mathcal{Y}})" + }, + { + "bbox": [ + 55, + 582, + 295, + 714 + ], + "type": "text", + "content": ". In the first stage, to address the limited interpretation of previous CBMs in the patches-to-concepts process, we introduce a disentangled optimal transport alignment framework between patches and concepts. Concretely, in Sec. 3.2, we propose two orthogonal projection losses to disentangle the patch features and concept features. In Sec. 3.3, we introduce an optimal transport (OT) based framework to model the complex relationship" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 313, + 384, + 555, + 492 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 384, + 555, + 492 + ], + "spans": [ + { + "bbox": [ + 313, + 384, + 555, + 492 + ], + "type": "text", + "content": "between patches and concepts, which can predict concepts from patches and meanwhile inverse the concept to patches to provide explanations of concept prediction. In Sec. 3.4, we integrate two prior distribution modeling to address the data bias problem. In the second stage, in Sec. 3.5, we follow vanilla CBMs and predict classes from concepts. Overall, our proposed DOT-CBM achieves holistic interpretability whether on intermediate concept predictions or on final class predictions. The whole pipeline is shown in Fig. 2." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 313, + 502, + 440, + 514 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 502, + 440, + 514 + ], + "spans": [ + { + "bbox": [ + 313, + 502, + 440, + 514 + ], + "type": "text", + "content": "3.1. Problem Formulation" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 313, + 521, + 554, + 593 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 521, + 554, + 593 + ], + "spans": [ + { + "bbox": [ + 313, + 521, + 554, + 593 + ], + "type": "text", + "content": "Traditional black-box visual models aim to fit the conditional distribution " + }, + { + "bbox": [ + 313, + 521, + 554, + 593 + ], + "type": "inline_equation", + "content": "P(y|x)" + }, + { + "bbox": [ + 313, + 521, + 554, + 593 + ], + "type": "text", + "content": " given an a dataset " + }, + { + "bbox": [ + 313, + 521, + 554, + 593 + ], + "type": "inline_equation", + "content": "\\mathcal{D} = \\{x_i, y_i\\}_{i=1}^N" + }, + { + "bbox": [ + 313, + 521, + 554, + 593 + ], + "type": "text", + "content": ", where " + }, + { + "bbox": [ + 313, + 521, + 554, + 593 + ], + "type": "inline_equation", + "content": "x_i \\in \\mathcal{X}, y_i \\in \\mathcal{Y}" + }, + { + "bbox": [ + 313, + 521, + 554, + 593 + ], + "type": "text", + "content": " are images and class labels, respectively. Though effective, they still lack transparency to understand the underlying decision-making process to drive the final predictions." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 313, + 594, + 555, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 594, + 555, + 714 + ], + "spans": [ + { + "bbox": [ + 313, + 594, + 555, + 714 + ], + "type": "text", + "content": "Concept Bottleneck Models (CBMs) [17] offer an interpretable solution by first predicting human-interpretable high-level concepts before making the final prediction. Formally, the dataset is extended to " + }, + { + "bbox": [ + 313, + 594, + 555, + 714 + ], + "type": "inline_equation", + "content": "\\mathcal{D} = \\{x_i,y_i,\\pmb {a}_i\\}_{i = 1}^N" + }, + { + "bbox": [ + 313, + 594, + 555, + 714 + ], + "type": "text", + "content": " , where " + }, + { + "bbox": [ + 313, + 594, + 555, + 714 + ], + "type": "inline_equation", + "content": "\\pmb {a}_i = \\{a_{ij}|a_{ij}\\in \\{0,1\\} \\}_{j = 1}^M" + }, + { + "bbox": [ + 313, + 594, + 555, + 714 + ], + "type": "text", + "content": " represents a binary concept activation label. " + }, + { + "bbox": [ + 313, + 594, + 555, + 714 + ], + "type": "inline_equation", + "content": "a_{ij} = 1" + }, + { + "bbox": [ + 313, + 594, + 555, + 714 + ], + "type": "text", + "content": " indicates the presence of a specific high-level concept " + }, + { + "bbox": [ + 313, + 594, + 555, + 714 + ], + "type": "inline_equation", + "content": "c_{j}" + }, + { + "bbox": [ + 313, + 594, + 555, + 714 + ], + "type": "text", + "content": " in image " + }, + { + "bbox": [ + 313, + 594, + 555, + 714 + ], + "type": "inline_equation", + "content": "x_{i}" + }, + { + "bbox": [ + 313, + 594, + 555, + 714 + ], + "type": "text", + "content": " , while " + }, + { + "bbox": [ + 313, + 594, + 555, + 714 + ], + "type": "inline_equation", + "content": "a_{ij} = 0" + }, + { + "bbox": [ + 313, + 594, + 555, + 714 + ], + "type": "text", + "content": " means the absence. " + }, + { + "bbox": [ + 313, + 594, + 555, + 714 + ], + "type": "inline_equation", + "content": "c_{j}" + }, + { + "bbox": [ + 313, + 594, + 555, + 714 + ], + "type": "text", + "content": " is concept from the pre-defined concept set " + }, + { + "bbox": [ + 313, + 594, + 555, + 714 + ], + "type": "inline_equation", + "content": "\\mathcal{C} = \\{c_j\\}_{j = 1}^M" + }, + { + "bbox": [ + 313, + 594, + 555, + 714 + ], + "type": "text", + "content": " . The optimization process of CBMs can be divided into two-stage [26, 35]: I) they first leverage a con" + } + ] + } + ], + "index": 10 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "text", + "content": "30201" + } + ] + } + ], + "index": 11 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 2 + }, + { + "para_blocks": [ + { + "bbox": [ + 55, + 71, + 296, + 168 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 71, + 296, + 168 + ], + "spans": [ + { + "bbox": [ + 55, + 71, + 296, + 168 + ], + "type": "text", + "content": "cept predictor " + }, + { + "bbox": [ + 55, + 71, + 296, + 168 + ], + "type": "inline_equation", + "content": "f_{\\mathcal{X}\\to \\mathcal{C}}:\\mathbb{R}^{H\\times W\\times C}\\to \\mathbb{R}^M" + }, + { + "bbox": [ + 55, + 71, + 296, + 168 + ], + "type": "text", + "content": " to map the input data " + }, + { + "bbox": [ + 55, + 71, + 296, + 168 + ], + "type": "inline_equation", + "content": "x" + }, + { + "bbox": [ + 55, + 71, + 296, + 168 + ], + "type": "text", + "content": " to the concept activation vector " + }, + { + "bbox": [ + 55, + 71, + 296, + 168 + ], + "type": "inline_equation", + "content": "\\pmb {a}\\in \\mathbb{R}^{M}" + }, + { + "bbox": [ + 55, + 71, + 296, + 168 + ], + "type": "text", + "content": " . 2) Subsequently, a target predictor " + }, + { + "bbox": [ + 55, + 71, + 296, + 168 + ], + "type": "inline_equation", + "content": "f_{\\mathcal{C}\\rightarrow \\mathcal{Y}}:\\mathbb{R}^{M}\\to \\mathbb{R}^{N_{y}}" + }, + { + "bbox": [ + 55, + 71, + 296, + 168 + ], + "type": "text", + "content": " maps the concept activation vector " + }, + { + "bbox": [ + 55, + 71, + 296, + 168 + ], + "type": "inline_equation", + "content": "\\pmb{a}" + }, + { + "bbox": [ + 55, + 71, + 296, + 168 + ], + "type": "text", + "content": " to the final label " + }, + { + "bbox": [ + 55, + 71, + 296, + 168 + ], + "type": "inline_equation", + "content": "y" + }, + { + "bbox": [ + 55, + 71, + 296, + 168 + ], + "type": "text", + "content": " . This structure enhances the interpretability of the model, as the intermediate concept predictions provide interpretability into the model's final decision. However, the interpretability from image to concept is un-explored." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 55, + 175, + 275, + 189 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 175, + 275, + 189 + ], + "spans": [ + { + "bbox": [ + 55, + 175, + 275, + 189 + ], + "type": "text", + "content": "3.2. Disentanglement on Local representations" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 55, + 194, + 296, + 290 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 194, + 296, + 290 + ], + "spans": [ + { + "bbox": [ + 55, + 194, + 296, + 290 + ], + "type": "text", + "content": "Disentangled visual patch features " + }, + { + "bbox": [ + 55, + 194, + 296, + 290 + ], + "type": "inline_equation", + "content": "\\widetilde{\\mathcal{P}}" + }, + { + "bbox": [ + 55, + 194, + 296, + 290 + ], + "type": "text", + "content": ". As previously discussed, current CBMs typically associate multiple local concepts with a global image feature, which leads to coarse-grained correspondences. In this paper, we adopt the pretrained vision transformer (ViT) [1] model, which has been approved to exhibit better spatial location information extraction than CNN [34], to extract local patch features to represent the local image regions." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 55, + 291, + 296, + 423 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 291, + 296, + 423 + ], + "spans": [ + { + "bbox": [ + 55, + 291, + 296, + 423 + ], + "type": "text", + "content": "Specifically, as shown in the left part of Fig. 2, ViT uniformly divides the full image " + }, + { + "bbox": [ + 55, + 291, + 296, + 423 + ], + "type": "inline_equation", + "content": "x \\in \\mathbb{R}^{H \\times W \\times C}" + }, + { + "bbox": [ + 55, + 291, + 296, + 423 + ], + "type": "text", + "content": " into " + }, + { + "bbox": [ + 55, + 291, + 296, + 423 + ], + "type": "inline_equation", + "content": "N" + }, + { + "bbox": [ + 55, + 291, + 296, + 423 + ], + "type": "text", + "content": " non-overlapping patches " + }, + { + "bbox": [ + 55, + 291, + 296, + 423 + ], + "type": "inline_equation", + "content": "\\{I_{pi} \\in \\mathbb{R}^{P_h \\times P_w}\\}_{i=1}^N" + }, + { + "bbox": [ + 55, + 291, + 296, + 423 + ], + "type": "text", + "content": ". After linearly projecting these patches into a " + }, + { + "bbox": [ + 55, + 291, + 296, + 423 + ], + "type": "inline_equation", + "content": "d" + }, + { + "bbox": [ + 55, + 291, + 296, + 423 + ], + "type": "text", + "content": "-dimension embedding space and flattening all patch embeddings into a sequence, we can get an initialized patch embedding matrix " + }, + { + "bbox": [ + 55, + 291, + 296, + 423 + ], + "type": "inline_equation", + "content": "X_P \\in \\mathbb{R}^{N \\times d}" + }, + { + "bbox": [ + 55, + 291, + 296, + 423 + ], + "type": "text", + "content": ". To extract the global image feature, ViT concatenates a learnable special token [CLS] with " + }, + { + "bbox": [ + 55, + 291, + 296, + 423 + ], + "type": "inline_equation", + "content": "X_P" + }, + { + "bbox": [ + 55, + 291, + 296, + 423 + ], + "type": "text", + "content": ". Besides, in order to preserve the spatial information of images, ViT adds a learnable position embedding " + }, + { + "bbox": [ + 55, + 291, + 296, + 423 + ], + "type": "inline_equation", + "content": "X_{pos} \\in \\mathbb{R}^{(N+1) \\times d}" + }, + { + "bbox": [ + 55, + 291, + 296, + 423 + ], + "type": "text", + "content": " and we get the input embeddings " + }, + { + "bbox": [ + 55, + 291, + 296, + 423 + ], + "type": "inline_equation", + "content": "X" + }, + { + "bbox": [ + 55, + 291, + 296, + 423 + ], + "type": "text", + "content": " for ViT:" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 111, + 432, + 295, + 447 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 111, + 432, + 295, + 447 + ], + "spans": [ + { + "bbox": [ + 111, + 432, + 295, + 447 + ], + "type": "interline_equation", + "content": "X _ {P} = \\left[ I _ {p 1} \\cdot W _ {p}, \\dots , I _ {p N} \\cdot W _ {p} \\right], \\tag {1}", + "image_path": "986c088d0e01d5176d5d044094cf932df53690b211cc1b3bf45e65a256e1918f.jpg" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 112, + 448, + 294, + 461 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 112, + 448, + 294, + 461 + ], + "spans": [ + { + "bbox": [ + 112, + 448, + 294, + 461 + ], + "type": "interline_equation", + "content": "X = \\left[ \\left[ \\mathrm {C L S} \\right], X _ {P} \\right] + X _ {\\text {p o s}}, \\tag {2}", + "image_path": "0531b7c8bfc5b44dd55879d5fbc8d6054ff3394bcec8e651d15c2309cda338e7.jpg" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 55, + 470, + 296, + 521 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 470, + 296, + 521 + ], + "spans": [ + { + "bbox": [ + 55, + 470, + 296, + 521 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 55, + 470, + 296, + 521 + ], + "type": "inline_equation", + "content": "W_{p}\\in \\mathbb{R}^{(P_{h}\\cdot P_{w}\\cdot C)\\times d}" + }, + { + "bbox": [ + 55, + 470, + 296, + 521 + ], + "type": "text", + "content": " is the patch embedding projection weights. Then, the pre-trained ViT " + }, + { + "bbox": [ + 55, + 470, + 296, + 521 + ], + "type": "inline_equation", + "content": "E_I" + }, + { + "bbox": [ + 55, + 470, + 296, + 521 + ], + "type": "text", + "content": " to map the input embeddings " + }, + { + "bbox": [ + 55, + 470, + 296, + 521 + ], + "type": "inline_equation", + "content": "X" + }, + { + "bbox": [ + 55, + 470, + 296, + 521 + ], + "type": "text", + "content": " into the global feature " + }, + { + "bbox": [ + 55, + 470, + 296, + 521 + ], + "type": "inline_equation", + "content": "g_{I}" + }, + { + "bbox": [ + 55, + 470, + 296, + 521 + ], + "type": "text", + "content": ", and patch features " + }, + { + "bbox": [ + 55, + 470, + 296, + 521 + ], + "type": "inline_equation", + "content": "\\mathcal{P} = \\{p_i\\}_{i = 1}^N" + }, + { + "bbox": [ + 55, + 470, + 296, + 521 + ], + "type": "text", + "content": ":" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 138, + 529, + 295, + 544 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 529, + 295, + 544 + ], + "spans": [ + { + "bbox": [ + 138, + 529, + 295, + 544 + ], + "type": "interline_equation", + "content": "[ g _ {I}, \\mathcal {P} ] = E _ {I} (X), \\tag {3}", + "image_path": "b6fc7c0f1d6598da668eeb43ac3bd16c02f4093ef92be62a876bb4a540960e18.jpg" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 55, + 552, + 296, + 636 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 552, + 296, + 636 + ], + "spans": [ + { + "bbox": [ + 55, + 552, + 296, + 636 + ], + "type": "text", + "content": "We empirically find that the patch features within the object regions are highly similar which is good for object representation, but bad for distinguishing local parts within the object region. Thus, we fit a learnable visual adapter " + }, + { + "bbox": [ + 55, + 552, + 296, + 636 + ], + "type": "inline_equation", + "content": "A_v" + }, + { + "bbox": [ + 55, + 552, + 296, + 636 + ], + "type": "text", + "content": " on the " + }, + { + "bbox": [ + 55, + 552, + 296, + 636 + ], + "type": "inline_equation", + "content": "\\mathcal{P}" + }, + { + "bbox": [ + 55, + 552, + 296, + 636 + ], + "type": "text", + "content": " and map the patch features within the foreground region to a disentangled unit hypersphere by introducing a visual orthogonal projection (VOP) regularizer as:" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 149, + 645, + 295, + 658 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 149, + 645, + 295, + 658 + ], + "spans": [ + { + "bbox": [ + 149, + 645, + 295, + 658 + ], + "type": "interline_equation", + "content": "S _ {I} = g _ {I} \\cdot P, \\tag {4}", + "image_path": "ab54265344978056652a9964dcc086cf5d3f65a8ca0784ea00985dcb93fbfd0f.jpg" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 91, + 661, + 295, + 687 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 91, + 661, + 295, + 687 + ], + "spans": [ + { + "bbox": [ + 91, + 661, + 295, + 687 + ], + "type": "interline_equation", + "content": "\\langle A _ {v} (p _ {i}), A _ {v} (p _ {j}) \\rangle = \\frac {A _ {v} (p _ {i}) \\cdot A _ {v} (p _ {j})}{\\| A _ {v} (p _ {i}) \\| \\| A _ {v} (p _ {j}) \\|}, \\tag {5}", + "image_path": "c7f68eb04e89590ef7d0bc972970eaa98b1fe3bfdf29952f6a16731536976a80.jpg" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 66, + 691, + 295, + 711 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 66, + 691, + 295, + 711 + ], + "spans": [ + { + "bbox": [ + 66, + 691, + 295, + 711 + ], + "type": "interline_equation", + "content": "\\mathcal {L} _ {\\mathrm {V O P}} = \\underset {\\left(p _ {i}, p _ {j}\\right) \\sim \\mathcal {P}} {\\mathbb {E}} \\left[ S _ {I} (i) \\cdot S _ {I} (j) \\cdot \\left\\langle A _ {v} \\left(p _ {i}\\right), A _ {v} \\left(p _ {j}\\right) \\right\\rangle \\right] \\tag {6}", + "image_path": "619e943d7ba6b8838d4b9dc7bfa2df8b6d6d4414a754dcfa71ad4f22ab1a1092.jpg" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 313, + 72, + 555, + 196 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 72, + 555, + 196 + ], + "spans": [ + { + "bbox": [ + 313, + 72, + 555, + 196 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 313, + 72, + 555, + 196 + ], + "type": "inline_equation", + "content": "S_{I}" + }, + { + "bbox": [ + 313, + 72, + 555, + 196 + ], + "type": "text", + "content": " is the saliency map to distinguish foreground and background, " + }, + { + "bbox": [ + 313, + 72, + 555, + 196 + ], + "type": "inline_equation", + "content": "\\langle \\cdot ,\\cdot \\rangle" + }, + { + "bbox": [ + 313, + 72, + 555, + 196 + ], + "type": "text", + "content": " is the cosine similarity. Finally, we get the disentangled patch features " + }, + { + "bbox": [ + 313, + 72, + 555, + 196 + ], + "type": "inline_equation", + "content": "\\widetilde{\\mathcal{P}} = \\{\\widetilde{p_i} = A_v(p_i)\\}_{i = 1}^N" + }, + { + "bbox": [ + 313, + 72, + 555, + 196 + ], + "type": "text", + "content": " Disentangled textual concept features " + }, + { + "bbox": [ + 313, + 72, + 555, + 196 + ], + "type": "inline_equation", + "content": "\\widetilde{\\mathcal{T}}" + }, + { + "bbox": [ + 313, + 72, + 555, + 196 + ], + "type": "text", + "content": ". For concept feature extraction, following [22, 29, 45, 47], we utilize a pre-trained text encoder " + }, + { + "bbox": [ + 313, + 72, + 555, + 196 + ], + "type": "inline_equation", + "content": "E_{T}" + }, + { + "bbox": [ + 313, + 72, + 555, + 196 + ], + "type": "text", + "content": " to compute concept embeddings " + }, + { + "bbox": [ + 313, + 72, + 555, + 196 + ], + "type": "inline_equation", + "content": "\\mathcal{T} = \\{t_j = E_T(c_j)\\}_{j = 1}^M" + }, + { + "bbox": [ + 313, + 72, + 555, + 196 + ], + "type": "text", + "content": ". Similarly, we fit a learnable textual adapter " + }, + { + "bbox": [ + 313, + 72, + 555, + 196 + ], + "type": "inline_equation", + "content": "A_{t}" + }, + { + "bbox": [ + 313, + 72, + 555, + 196 + ], + "type": "text", + "content": " and introduce a textual orthogonal projection (TOP) regularizer to disentangle the concept features in a unit hyper-sphere, as:" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 362, + 202, + 555, + 223 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 362, + 202, + 555, + 223 + ], + "spans": [ + { + "bbox": [ + 362, + 202, + 555, + 223 + ], + "type": "interline_equation", + "content": "\\mathcal {L} _ {\\mathrm {T O P}} = \\underset {(t _ {i}, t _ {j}) \\sim \\mathcal {T}} {\\mathbb {E}} [ \\langle A _ {t} (t _ {i}), A _ {t} (t _ {j}) \\rangle ], \\tag {7}", + "image_path": "19f268efc8accecd686e413319119f77c8068fa5f9a3fa9c79672478605978e3.jpg" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 314, + 230, + 555, + 255 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 314, + 230, + 555, + 255 + ], + "spans": [ + { + "bbox": [ + 314, + 230, + 555, + 255 + ], + "type": "text", + "content": "Consequently, we get the disentangled concept features " + }, + { + "bbox": [ + 314, + 230, + 555, + 255 + ], + "type": "inline_equation", + "content": "\\widetilde{\\mathcal{T}} = \\{\\widetilde{t_j} = A_t(t_j)\\}_{j = 1}^M." + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 313, + 260, + 555, + 274 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 260, + 555, + 274 + ], + "spans": [ + { + "bbox": [ + 313, + 260, + 555, + 274 + ], + "type": "text", + "content": "3.3. Patch-concept alignment via optimal transport" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 313, + 279, + 554, + 375 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 279, + 554, + 375 + ], + "spans": [ + { + "bbox": [ + 313, + 279, + 554, + 375 + ], + "type": "text", + "content": "Now, we have the disentangled patch feature set " + }, + { + "bbox": [ + 313, + 279, + 554, + 375 + ], + "type": "inline_equation", + "content": "\\widetilde{\\mathcal{P}}" + }, + { + "bbox": [ + 313, + 279, + 554, + 375 + ], + "type": "text", + "content": " and disentangled concept feature set " + }, + { + "bbox": [ + 313, + 279, + 554, + 375 + ], + "type": "inline_equation", + "content": "\\widetilde{\\mathcal{T}}" + }, + { + "bbox": [ + 313, + 279, + 554, + 375 + ], + "type": "text", + "content": ", we need to model the relationship " + }, + { + "bbox": [ + 313, + 279, + 554, + 375 + ], + "type": "inline_equation", + "content": "\\mathbf{Q}" + }, + { + "bbox": [ + 313, + 279, + 554, + 375 + ], + "type": "text", + "content": " between these two sets and derive the concept activation vector " + }, + { + "bbox": [ + 313, + 279, + 554, + 375 + ], + "type": "inline_equation", + "content": "\\hat{\\pmb{a}}" + }, + { + "bbox": [ + 313, + 279, + 554, + 375 + ], + "type": "text", + "content": " based on the input image. As a complex many-to-many alignment problem, we propose an optimal transport (OT) based binary cross-entropy loss to tackle these problems. Our goal is to correctly align the patch features and concept features with the same semantics." + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 313, + 376, + 554, + 412 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 376, + 554, + 412 + ], + "spans": [ + { + "bbox": [ + 313, + 376, + 554, + 412 + ], + "type": "text", + "content": "Assume that these two sets " + }, + { + "bbox": [ + 313, + 376, + 554, + 412 + ], + "type": "inline_equation", + "content": "\\widetilde{\\mathcal{P}}" + }, + { + "bbox": [ + 313, + 376, + 554, + 412 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 313, + 376, + 554, + 412 + ], + "type": "inline_equation", + "content": "\\widetilde{\\mathcal{T}}" + }, + { + "bbox": [ + 313, + 376, + 554, + 412 + ], + "type": "text", + "content": " follow two discrete probability distributions " + }, + { + "bbox": [ + 313, + 376, + 554, + 412 + ], + "type": "inline_equation", + "content": "\\Theta \\in \\mathbb{R}^N" + }, + { + "bbox": [ + 313, + 376, + 554, + 412 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 313, + 376, + 554, + 412 + ], + "type": "inline_equation", + "content": "\\Gamma \\in \\mathbb{R}^M" + }, + { + "bbox": [ + 313, + 376, + 554, + 412 + ], + "type": "text", + "content": " on the " + }, + { + "bbox": [ + 313, + 376, + 554, + 412 + ], + "type": "inline_equation", + "content": "d" + }, + { + "bbox": [ + 313, + 376, + 554, + 412 + ], + "type": "text", + "content": "-dimension embedding space, respectively:" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 365, + 419, + 554, + 453 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 365, + 419, + 554, + 453 + ], + "spans": [ + { + "bbox": [ + 365, + 419, + 554, + 453 + ], + "type": "interline_equation", + "content": "\\boldsymbol {\\Theta} = \\sum_ {i = 1} ^ {N} \\theta_ {i} \\delta_ {p _ {i}}, \\quad \\boldsymbol {\\Gamma} = \\sum_ {j = 1} ^ {M} \\gamma_ {j} \\delta_ {c _ {j}}, \\tag {8}", + "image_path": "2b7a2fa93d7e5a4cdbe65d9a1ed90559cf3ca885e4f9e338e3f820fe30a316a1.jpg" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 313, + 461, + 554, + 523 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 461, + 554, + 523 + ], + "spans": [ + { + "bbox": [ + 313, + 461, + 554, + 523 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 313, + 461, + 554, + 523 + ], + "type": "inline_equation", + "content": "\\pmb{\\theta} \\in \\Sigma^{N}" + }, + { + "bbox": [ + 313, + 461, + 554, + 523 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 313, + 461, + 554, + 523 + ], + "type": "inline_equation", + "content": "\\pmb{\\gamma} \\in \\Sigma^{M}" + }, + { + "bbox": [ + 313, + 461, + 554, + 523 + ], + "type": "text", + "content": ", the simplex of " + }, + { + "bbox": [ + 313, + 461, + 554, + 523 + ], + "type": "inline_equation", + "content": "\\mathbb{R}^N" + }, + { + "bbox": [ + 313, + 461, + 554, + 523 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 313, + 461, + 554, + 523 + ], + "type": "inline_equation", + "content": "\\mathbb{R}^{M}" + }, + { + "bbox": [ + 313, + 461, + 554, + 523 + ], + "type": "text", + "content": ", denote the probability values of the discrete states with satisfaction that " + }, + { + "bbox": [ + 313, + 461, + 554, + 523 + ], + "type": "inline_equation", + "content": "\\sum_{i=1}^{N} \\theta_{i} = 1" + }, + { + "bbox": [ + 313, + 461, + 554, + 523 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 313, + 461, + 554, + 523 + ], + "type": "inline_equation", + "content": "\\sum_{j=1}^{M} \\gamma_{j} = 1" + }, + { + "bbox": [ + 313, + 461, + 554, + 523 + ], + "type": "text", + "content": ". " + }, + { + "bbox": [ + 313, + 461, + 554, + 523 + ], + "type": "inline_equation", + "content": "\\delta_{e}" + }, + { + "bbox": [ + 313, + 461, + 554, + 523 + ], + "type": "text", + "content": " refers to a point mass located at coordinate " + }, + { + "bbox": [ + 313, + 461, + 554, + 523 + ], + "type": "inline_equation", + "content": "e \\in \\mathbb{R}^{d}" + }, + { + "bbox": [ + 313, + 461, + 554, + 523 + ], + "type": "text", + "content": ". The prior modeling of " + }, + { + "bbox": [ + 313, + 461, + 554, + 523 + ], + "type": "inline_equation", + "content": "\\pmb{\\theta}" + }, + { + "bbox": [ + 313, + 461, + 554, + 523 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 313, + 461, + 554, + 523 + ], + "type": "inline_equation", + "content": "\\pmb{\\gamma}" + }, + { + "bbox": [ + 313, + 461, + 554, + 523 + ], + "type": "text", + "content": " are shown in the Sec. 3.4." + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 313, + 524, + 554, + 548 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 524, + 554, + 548 + ], + "spans": [ + { + "bbox": [ + 313, + 524, + 554, + 548 + ], + "type": "text", + "content": "To get the transport assignment " + }, + { + "bbox": [ + 313, + 524, + 554, + 548 + ], + "type": "inline_equation", + "content": "\\mathbf{Q}" + }, + { + "bbox": [ + 313, + 524, + 554, + 548 + ], + "type": "text", + "content": " between " + }, + { + "bbox": [ + 313, + 524, + 554, + 548 + ], + "type": "inline_equation", + "content": "\\Theta" + }, + { + "bbox": [ + 313, + 524, + 554, + 548 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 313, + 524, + 554, + 548 + ], + "type": "inline_equation", + "content": "\\Gamma" + }, + { + "bbox": [ + 313, + 524, + 554, + 548 + ], + "type": "text", + "content": ", the objective of OT is formulated as the optimization problem:" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 323, + 555, + 554, + 582 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 323, + 555, + 554, + 582 + ], + "spans": [ + { + "bbox": [ + 323, + 555, + 554, + 582 + ], + "type": "interline_equation", + "content": "\\min _ {\\mathbf {Q} \\in \\mathcal {Q}} \\sum_ {i, j} q _ {i j} d _ {i j} - \\varepsilon H (\\mathbf {Q}) \\tag {9}", + "image_path": "b9e7299aef0c5bb8c32f840d2d4ca81db552905573a881290d865a90535e5507.jpg" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 324, + 583, + 533, + 599 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 324, + 583, + 533, + 599 + ], + "spans": [ + { + "bbox": [ + 324, + 583, + 533, + 599 + ], + "type": "interline_equation", + "content": "\\text {s . t .} \\mathcal {Q} = \\left\\{\\mathbf {Q} \\in \\mathbb {R} _ {+} ^ {N \\times M} \\mid \\mathbf {Q} \\mathbf {1} _ {M} = \\boldsymbol {\\theta}, \\mathbf {Q} ^ {\\top} \\mathbf {1} _ {N} = \\boldsymbol {\\gamma} \\right\\}.", + "image_path": "389b50dbd9ac5e89b4a7cd2e61eaaf40e9ed5619063310509528b27d2adabfb0.jpg" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 313, + 605, + 555, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 605, + 555, + 714 + ], + "spans": [ + { + "bbox": [ + 313, + 605, + 555, + 714 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 313, + 605, + 555, + 714 + ], + "type": "inline_equation", + "content": "\\mathcal{Q}" + }, + { + "bbox": [ + 313, + 605, + 555, + 714 + ], + "type": "text", + "content": " is the set of doubly stochastic matrices " + }, + { + "bbox": [ + 313, + 605, + 555, + 714 + ], + "type": "inline_equation", + "content": "\\mathbf{Q} = [q_{ij}]_{N\\times M}" + }, + { + "bbox": [ + 313, + 605, + 555, + 714 + ], + "type": "text", + "content": ", denotes the corresponding transport assignment where " + }, + { + "bbox": [ + 313, + 605, + 555, + 714 + ], + "type": "inline_equation", + "content": "q_{ij}" + }, + { + "bbox": [ + 313, + 605, + 555, + 714 + ], + "type": "text", + "content": " represents the probabilities of aligning " + }, + { + "bbox": [ + 313, + 605, + 555, + 714 + ], + "type": "inline_equation", + "content": "p_i" + }, + { + "bbox": [ + 313, + 605, + 555, + 714 + ], + "type": "text", + "content": " with " + }, + { + "bbox": [ + 313, + 605, + 555, + 714 + ], + "type": "inline_equation", + "content": "t_j" + }, + { + "bbox": [ + 313, + 605, + 555, + 714 + ], + "type": "text", + "content": ". " + }, + { + "bbox": [ + 313, + 605, + 555, + 714 + ], + "type": "inline_equation", + "content": "\\mathbf{1}^M" + }, + { + "bbox": [ + 313, + 605, + 555, + 714 + ], + "type": "text", + "content": " is the " + }, + { + "bbox": [ + 313, + 605, + 555, + 714 + ], + "type": "inline_equation", + "content": "M" + }, + { + "bbox": [ + 313, + 605, + 555, + 714 + ], + "type": "text", + "content": "-dimensional vector of ones. " + }, + { + "bbox": [ + 313, + 605, + 555, + 714 + ], + "type": "inline_equation", + "content": "d_{ij} = d(p_i,t_j) = (1 - \\langle p_i,t_j\\rangle)\\geq 0" + }, + { + "bbox": [ + 313, + 605, + 555, + 714 + ], + "type": "text", + "content": " is the transport cost between the two points " + }, + { + "bbox": [ + 313, + 605, + 555, + 714 + ], + "type": "inline_equation", + "content": "p_i" + }, + { + "bbox": [ + 313, + 605, + 555, + 714 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 313, + 605, + 555, + 714 + ], + "type": "inline_equation", + "content": "s_j" + }, + { + "bbox": [ + 313, + 605, + 555, + 714 + ], + "type": "text", + "content": " and we use the cosine distance to compute cost matrix " + }, + { + "bbox": [ + 313, + 605, + 555, + 714 + ], + "type": "inline_equation", + "content": "\\mathbf{D}" + }, + { + "bbox": [ + 313, + 605, + 555, + 714 + ], + "type": "text", + "content": ". The optimal " + }, + { + "bbox": [ + 313, + 605, + 555, + 714 + ], + "type": "inline_equation", + "content": "\\mathbf{Q}^*" + }, + { + "bbox": [ + 313, + 605, + 555, + 714 + ], + "type": "text", + "content": " of (9) has a simple normalized exponential matrix solution by Sinkhorn fixed point iterations [7]," + } + ] + } + ], + "index": 23 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 318, + 757 + ], + "type": "footer", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 318, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 318, + 757 + ], + "type": "text", + "content": "30202" + } + ] + } + ], + "index": 24 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 3 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 92, + 70, + 262, + 214 + ], + "blocks": [ + { + "bbox": [ + 92, + 70, + 262, + 214 + ], + "lines": [ + { + "bbox": [ + 92, + 70, + 262, + 214 + ], + "spans": [ + { + "bbox": [ + 92, + 70, + 262, + 214 + ], + "type": "image", + "image_path": "ea76a7dba2a4f3ab732113d26928c8cbb74c4ad73f1fd3de7a85d95713606a83.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 61, + 222, + 288, + 233 + ], + "lines": [ + { + "bbox": [ + 61, + 222, + 288, + 233 + ], + "spans": [ + { + "bbox": [ + 61, + 222, + 288, + 233 + ], + "type": "text", + "content": "Figure 3. Conditioned co-occurrence matrix of concept labels." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "bbox": [ + 120, + 258, + 294, + 270 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 120, + 258, + 294, + 270 + ], + "spans": [ + { + "bbox": [ + 120, + 258, + 294, + 270 + ], + "type": "interline_equation", + "content": "\\mathbf {Q} ^ {*} = \\operatorname {S i n k h o r n} (\\mathbf {D}, \\boldsymbol {\\theta}, \\gamma), \\tag {10}", + "image_path": "316fbbd516fd3d205fbe70e7f2bd8972ed6c650aacb0422a58b8ca02364e2ce3.jpg" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 55, + 277, + 295, + 300 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 277, + 295, + 300 + ], + "spans": [ + { + "bbox": [ + 55, + 277, + 295, + 300 + ], + "type": "text", + "content": "After obtaining the optimal transport assignment " + }, + { + "bbox": [ + 55, + 277, + 295, + 300 + ], + "type": "inline_equation", + "content": "\\mathbf{Q}^*" + }, + { + "bbox": [ + 55, + 277, + 295, + 300 + ], + "type": "text", + "content": ", the OT distance between the two sets is expressed as:" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 138, + 310, + 295, + 336 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 310, + 295, + 336 + ], + "spans": [ + { + "bbox": [ + 138, + 310, + 295, + 336 + ], + "type": "interline_equation", + "content": "d _ {O T} = \\sum_ {i, j} q _ {i j} d _ {i j} \\tag {11}", + "image_path": "a34ab69129cc5079f7cc8cb48909cf7c06a2ad063809c5e6668127457bf7aefc.jpg" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 55, + 345, + 296, + 404 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 345, + 296, + 404 + ], + "spans": [ + { + "bbox": [ + 55, + 345, + 296, + 404 + ], + "type": "text", + "content": "For the concept embedding " + }, + { + "bbox": [ + 55, + 345, + 296, + 404 + ], + "type": "inline_equation", + "content": "t_j" + }, + { + "bbox": [ + 55, + 345, + 296, + 404 + ], + "type": "text", + "content": ", we only consider the minimum distance " + }, + { + "bbox": [ + 55, + 345, + 296, + 404 + ], + "type": "inline_equation", + "content": "d_{OT_j} = \\min_i(q_{ij}d_{ij})" + }, + { + "bbox": [ + 55, + 345, + 296, + 404 + ], + "type": "text", + "content": " between this concept embedding " + }, + { + "bbox": [ + 55, + 345, + 296, + 404 + ], + "type": "inline_equation", + "content": "t_j" + }, + { + "bbox": [ + 55, + 345, + 296, + 404 + ], + "type": "text", + "content": " and all patch embeddings to facilitate the concept activation value prediction " + }, + { + "bbox": [ + 55, + 345, + 296, + 404 + ], + "type": "inline_equation", + "content": "(x\\rightarrow \\mathcal{C})" + }, + { + "bbox": [ + 55, + 345, + 296, + 404 + ], + "type": "text", + "content": ", the predicted concept activation vector " + }, + { + "bbox": [ + 55, + 345, + 296, + 404 + ], + "type": "inline_equation", + "content": "\\hat{\\pmb{a}}" + }, + { + "bbox": [ + 55, + 345, + 296, + 404 + ], + "type": "text", + "content": " are stated as:" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 106, + 413, + 294, + 430 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 106, + 413, + 294, + 430 + ], + "spans": [ + { + "bbox": [ + 106, + 413, + 294, + 430 + ], + "type": "interline_equation", + "content": "\\hat {\\boldsymbol {a}} = \\left\\{\\hat {a} _ {j} \\mid \\hat {a} _ {j} = \\sigma \\left(1 - d _ {O T _ {j}}\\right) \\right\\} _ {j = 1} ^ {M}. \\tag {12}", + "image_path": "2e05589576b916945523eb9f47449ca9a0badac38f79405ce61153669c9cad1e.jpg" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 55, + 437, + 296, + 474 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 437, + 296, + 474 + ], + "spans": [ + { + "bbox": [ + 55, + 437, + 296, + 474 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 55, + 437, + 296, + 474 + ], + "type": "inline_equation", + "content": "\\sigma" + }, + { + "bbox": [ + 55, + 437, + 296, + 474 + ], + "type": "text", + "content": " is the sigmoid function. Moreover, we can inverse specific concepts to the image patches based on " + }, + { + "bbox": [ + 55, + 437, + 296, + 474 + ], + "type": "inline_equation", + "content": "\\mathbf{Q}^*" + }, + { + "bbox": [ + 55, + 437, + 296, + 474 + ], + "type": "text", + "content": ", and get the inversion mask " + }, + { + "bbox": [ + 55, + 437, + 296, + 474 + ], + "type": "inline_equation", + "content": "m_j" + }, + { + "bbox": [ + 55, + 437, + 296, + 474 + ], + "type": "text", + "content": " of concept " + }, + { + "bbox": [ + 55, + 437, + 296, + 474 + ], + "type": "inline_equation", + "content": "c_j" + }, + { + "bbox": [ + 55, + 437, + 296, + 474 + ], + "type": "text", + "content": ", as:" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 124, + 483, + 294, + 497 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 124, + 483, + 294, + 497 + ], + "spans": [ + { + "bbox": [ + 124, + 483, + 294, + 497 + ], + "type": "interline_equation", + "content": "m _ {j} = \\left[ q _ {1 j}, q _ {2 j}, \\dots , q _ {N j} \\right]. \\tag {13}", + "image_path": "6b0b26436fb3e66dc25567c8372e3da25330ad08b45e106b63643a5a54744457.jpg" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 55, + 506, + 295, + 541 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 506, + 295, + 541 + ], + "spans": [ + { + "bbox": [ + 55, + 506, + 295, + 541 + ], + "type": "text", + "content": "By utilizing OT distance between patches and concepts as the patch-concept transport, our OT-based binary cross-entropy loss " + }, + { + "bbox": [ + 55, + 506, + 295, + 541 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{\\mathrm{concept}}" + }, + { + "bbox": [ + 55, + 506, + 295, + 541 + ], + "type": "text", + "content": " are as follows:" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 72, + 551, + 295, + 564 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 72, + 551, + 295, + 564 + ], + "spans": [ + { + "bbox": [ + 72, + 551, + 295, + 564 + ], + "type": "interline_equation", + "content": "\\mathcal {L} _ {\\text {c o n c e p t}} = - \\mathbb {E} [ \\boldsymbol {a} \\log (\\hat {\\boldsymbol {a}}) + (1 - \\boldsymbol {a}) \\log (1 - \\hat {\\boldsymbol {a}}) ] \\tag {14}", + "image_path": "ef3a331ef296c1dc8611326268a5426117a2480107b243854fab41f30a1e0e58.jpg" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 55, + 573, + 290, + 586 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 573, + 290, + 586 + ], + "spans": [ + { + "bbox": [ + 55, + 573, + 290, + 586 + ], + "type": "text", + "content": "Now, we can get the total objective of stage 1 " + }, + { + "bbox": [ + 55, + 573, + 290, + 586 + ], + "type": "inline_equation", + "content": "(f_{\\mathcal{X}\\to \\mathcal{C}})" + }, + { + "bbox": [ + 55, + 573, + 290, + 586 + ], + "type": "text", + "content": ", as:" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 80, + 595, + 295, + 609 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 80, + 595, + 295, + 609 + ], + "spans": [ + { + "bbox": [ + 80, + 595, + 295, + 609 + ], + "type": "interline_equation", + "content": "\\mathcal {L} _ {f, x \\rightarrow c} = \\lambda_ {1} \\mathcal {L} _ {\\mathrm {V O P}} + \\lambda_ {2} \\mathcal {L} _ {\\mathrm {T O P}} + \\lambda_ {3} \\mathcal {L} _ {\\text {c o n c e p t}}, \\tag {15}", + "image_path": "27ae602254f8b3dbceef4909a501f3c6f2fcce544c47f406868227f48e170c7a.jpg" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 55, + 617, + 226, + 629 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 617, + 226, + 629 + ], + "spans": [ + { + "bbox": [ + 55, + 617, + 226, + 629 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 55, + 617, + 226, + 629 + ], + "type": "inline_equation", + "content": "\\lambda_1, \\lambda_2, \\lambda_3" + }, + { + "bbox": [ + 55, + 617, + 226, + 629 + ], + "type": "text", + "content": " are the hyperparameters." + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 55, + 630, + 296, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 630, + 296, + 714 + ], + "spans": [ + { + "bbox": [ + 55, + 630, + 296, + 714 + ], + "type": "text", + "content": "The obtained assignment " + }, + { + "bbox": [ + 55, + 630, + 296, + 714 + ], + "type": "inline_equation", + "content": "\\mathbf{Q}^{*} = [q_{i,j}]_{N\\times M}" + }, + { + "bbox": [ + 55, + 630, + 296, + 714 + ], + "type": "text", + "content": " indicates the relationship between patches and concepts, where " + }, + { + "bbox": [ + 55, + 630, + 296, + 714 + ], + "type": "inline_equation", + "content": "q_{i,j}" + }, + { + "bbox": [ + 55, + 630, + 296, + 714 + ], + "type": "text", + "content": " represent correlation between patch " + }, + { + "bbox": [ + 55, + 630, + 296, + 714 + ], + "type": "inline_equation", + "content": "p_i" + }, + { + "bbox": [ + 55, + 630, + 296, + 714 + ], + "type": "text", + "content": " and concept " + }, + { + "bbox": [ + 55, + 630, + 296, + 714 + ], + "type": "inline_equation", + "content": "c_{j}" + }, + { + "bbox": [ + 55, + 630, + 296, + 714 + ], + "type": "text", + "content": ". Consequently, based on " + }, + { + "bbox": [ + 55, + 630, + 296, + 714 + ], + "type": "inline_equation", + "content": "\\mathbf{Q}^*" + }, + { + "bbox": [ + 55, + 630, + 296, + 714 + ], + "type": "text", + "content": ", we can obtain concept activation " + }, + { + "bbox": [ + 55, + 630, + 296, + 714 + ], + "type": "inline_equation", + "content": "\\pmb{a}" + }, + { + "bbox": [ + 55, + 630, + 296, + 714 + ], + "type": "text", + "content": " (Eq.(12)) based on the image features " + }, + { + "bbox": [ + 55, + 630, + 296, + 714 + ], + "type": "inline_equation", + "content": "(x\\to \\mathcal{C})" + }, + { + "bbox": [ + 55, + 630, + 296, + 714 + ], + "type": "text", + "content": " and meanwhile inverse the concept to the original image patches (Eq.(13)) " + }, + { + "bbox": [ + 55, + 630, + 296, + 714 + ], + "type": "inline_equation", + "content": "(\\mathcal{C}\\rightarrow x)" + }, + { + "bbox": [ + 55, + 630, + 296, + 714 + ], + "type": "text", + "content": " to provide explicit interpretation." + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 314, + 72, + 471, + 85 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 314, + 72, + 471, + 85 + ], + "spans": [ + { + "bbox": [ + 314, + 72, + 471, + 85 + ], + "type": "text", + "content": "3.4. Prior distributions Modeling" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 313, + 90, + 553, + 162 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 90, + 553, + 162 + ], + "spans": [ + { + "bbox": [ + 313, + 90, + 553, + 162 + ], + "type": "text", + "content": "The prior distributions of " + }, + { + "bbox": [ + 313, + 90, + 553, + 162 + ], + "type": "inline_equation", + "content": "\\theta" + }, + { + "bbox": [ + 313, + 90, + 553, + 162 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 313, + 90, + 553, + 162 + ], + "type": "inline_equation", + "content": "\\gamma" + }, + { + "bbox": [ + 313, + 90, + 553, + 162 + ], + "type": "text", + "content": " of OT are typically uniform distributions, which is sub-optimal for our task. Considering the part-background spurious correlation and part-part spurious correlation problem mentioned in Fig. 1b, we propose two prior distribution modeling in this section to address these limitations, respectively." + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 313, + 163, + 554, + 269 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 163, + 554, + 269 + ], + "spans": [ + { + "bbox": [ + 313, + 163, + 554, + 269 + ], + "type": "text", + "content": "Patch set prior " + }, + { + "bbox": [ + 313, + 163, + 554, + 269 + ], + "type": "inline_equation", + "content": "\\theta" + }, + { + "bbox": [ + 313, + 163, + 554, + 269 + ], + "type": "text", + "content": ". We attribute the part-background spurious correlation phenomenon to the high co-occurrence of some classes and backgrounds, such as \"seabird\" and \"water\". To solve this biased association between object concepts and the background, we use saliency map " + }, + { + "bbox": [ + 313, + 163, + 554, + 269 + ], + "type": "inline_equation", + "content": "S_I" + }, + { + "bbox": [ + 313, + 163, + 554, + 269 + ], + "type": "text", + "content": " from Eq. (4) to encourage the transports between the patch features within foreground regions and concepts and penalize the transports of background regions. Specifically, the prior distribution of patch set " + }, + { + "bbox": [ + 313, + 163, + 554, + 269 + ], + "type": "inline_equation", + "content": "\\theta" + }, + { + "bbox": [ + 313, + 163, + 554, + 269 + ], + "type": "text", + "content": " are formulated as:" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 394, + 275, + 553, + 288 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 394, + 275, + 553, + 288 + ], + "spans": [ + { + "bbox": [ + 394, + 275, + 553, + 288 + ], + "type": "interline_equation", + "content": "\\boldsymbol {\\theta} = \\operatorname {S o f t m a x} \\left(S _ {I}\\right), \\tag {16}", + "image_path": "f7102bab3dd4f0a0a00435c2b9e8c2f06cdda99148413de8ccd0be553e7ce7d6.jpg" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 313, + 293, + 554, + 388 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 293, + 554, + 388 + ], + "spans": [ + { + "bbox": [ + 313, + 293, + 554, + 388 + ], + "type": "text", + "content": "Concept set prior " + }, + { + "bbox": [ + 313, + 293, + 554, + 388 + ], + "type": "inline_equation", + "content": "\\gamma" + }, + { + "bbox": [ + 313, + 293, + 554, + 388 + ], + "type": "text", + "content": ". As illustrated in Fig. 3, some concepts are highly co-occurrent. These concept label biases lead to the model's confusion between these concepts, yielding the part-part spurious correlation phenomenon. To address this problem, we employ the statistical co-occurrence matrix of the training dataset to facilitate the prior distribution modeling of " + }, + { + "bbox": [ + 313, + 293, + 554, + 388 + ], + "type": "inline_equation", + "content": "\\gamma" + }, + { + "bbox": [ + 313, + 293, + 554, + 388 + ], + "type": "text", + "content": ", aiming to penalize those concepts with high co-occurrence with other concepts." + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 313, + 388, + 554, + 495 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 388, + 554, + 495 + ], + "spans": [ + { + "bbox": [ + 313, + 388, + 554, + 495 + ], + "type": "text", + "content": "Specifically, for a dataset with " + }, + { + "bbox": [ + 313, + 388, + 554, + 495 + ], + "type": "inline_equation", + "content": "M" + }, + { + "bbox": [ + 313, + 388, + 554, + 495 + ], + "type": "text", + "content": " concepts and " + }, + { + "bbox": [ + 313, + 388, + 554, + 495 + ], + "type": "inline_equation", + "content": "K" + }, + { + "bbox": [ + 313, + 388, + 554, + 495 + ], + "type": "text", + "content": " classes, class-concept label form a matrix " + }, + { + "bbox": [ + 313, + 388, + 554, + 495 + ], + "type": "inline_equation", + "content": "L = [l_{i,j}]_{K\\times M}" + }, + { + "bbox": [ + 313, + 388, + 554, + 495 + ], + "type": "text", + "content": " where " + }, + { + "bbox": [ + 313, + 388, + 554, + 495 + ], + "type": "inline_equation", + "content": "l_{i,j}" + }, + { + "bbox": [ + 313, + 388, + 554, + 495 + ], + "type": "text", + "content": " represents the activation value of the " + }, + { + "bbox": [ + 313, + 388, + 554, + 495 + ], + "type": "inline_equation", + "content": "j" + }, + { + "bbox": [ + 313, + 388, + 554, + 495 + ], + "type": "text", + "content": "-th concept in the concept label for the " + }, + { + "bbox": [ + 313, + 388, + 554, + 495 + ], + "type": "inline_equation", + "content": "i" + }, + { + "bbox": [ + 313, + 388, + 554, + 495 + ], + "type": "text", + "content": "-th class. First, we count the co-occurrence frequencies of each pair of concepts. Then, we normalize these counts by the total number of times each concept is activated, resulting in the conditional co-occurrence matrix. The conditional co-occurrence vector " + }, + { + "bbox": [ + 313, + 388, + 554, + 495 + ], + "type": "inline_equation", + "content": "r_i" + }, + { + "bbox": [ + 313, + 388, + 554, + 495 + ], + "type": "text", + "content": " is calculated as follows:" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 372, + 499, + 553, + 533 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 372, + 499, + 553, + 533 + ], + "spans": [ + { + "bbox": [ + 372, + 499, + 553, + 533 + ], + "type": "interline_equation", + "content": "r _ {i} = \\sum_ {j = 1} ^ {M} \\left(\\frac {\\sum_ {k = 1} ^ {K} L _ {k i} \\cdot L _ {k j}}{\\sum_ {k = 1} ^ {K} L _ {k i}}\\right) \\tag {17}", + "image_path": "5e2d4bdc969cd1423109079042ca9a34279c7710af7ae22da57d63d4f5552609.jpg" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 313, + 536, + 553, + 559 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 536, + 553, + 559 + ], + "spans": [ + { + "bbox": [ + 313, + 536, + 553, + 559 + ], + "type": "text", + "content": "The concept set prior distribution " + }, + { + "bbox": [ + 313, + 536, + 553, + 559 + ], + "type": "inline_equation", + "content": "\\gamma" + }, + { + "bbox": [ + 313, + 536, + 553, + 559 + ], + "type": "text", + "content": " is then derived from the conditional co-occurrence vector as follows:" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 389, + 567, + 553, + 579 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 389, + 567, + 553, + 579 + ], + "spans": [ + { + "bbox": [ + 389, + 567, + 553, + 579 + ], + "type": "interline_equation", + "content": "\\gamma = 1 - \\operatorname {S o f t m a x} \\left(r _ {i}\\right) \\tag {18}", + "image_path": "92e0951073e875795194b7236a55b359afe6d8391ad754e191e5cc1981a0e8ce.jpg" + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 313, + 584, + 553, + 609 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 584, + 553, + 609 + ], + "spans": [ + { + "bbox": [ + 313, + 584, + 553, + 609 + ], + "type": "text", + "content": "This prior modeling ensures that those concepts with high co-occurrence with other concepts will be penalized." + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 314, + 613, + 464, + 626 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 314, + 613, + 464, + 626 + ], + "spans": [ + { + "bbox": [ + 314, + 613, + 464, + 626 + ], + "type": "text", + "content": "3.5. Concept " + }, + { + "bbox": [ + 314, + 613, + 464, + 626 + ], + "type": "inline_equation", + "content": "\\rightarrow" + }, + { + "bbox": [ + 314, + 613, + 464, + 626 + ], + "type": "text", + "content": " class prediction" + } + ] + } + ], + "index": 25 + }, + { + "bbox": [ + 313, + 632, + 554, + 692 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 632, + 554, + 692 + ], + "spans": [ + { + "bbox": [ + 313, + 632, + 554, + 692 + ], + "type": "text", + "content": "Now, we already finished the first stage which obtains the intermediate concept activation predictions " + }, + { + "bbox": [ + 313, + 632, + 554, + 692 + ], + "type": "inline_equation", + "content": "\\pmb{a}" + }, + { + "bbox": [ + 313, + 632, + 554, + 692 + ], + "type": "text", + "content": ". In the second stage, our goal is to learn the association between concept and final class prediction similar to vanilla CBM. Specifically, we can map the " + }, + { + "bbox": [ + 313, + 632, + 554, + 692 + ], + "type": "inline_equation", + "content": "\\pmb{a}" + }, + { + "bbox": [ + 313, + 632, + 554, + 692 + ], + "type": "text", + "content": " to final output " + }, + { + "bbox": [ + 313, + 632, + 554, + 692 + ], + "type": "inline_equation", + "content": "\\hat{y}" + }, + { + "bbox": [ + 313, + 632, + 554, + 692 + ], + "type": "text", + "content": ", as:" + } + ] + } + ], + "index": 26 + }, + { + "bbox": [ + 406, + 698, + 553, + 710 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 406, + 698, + 553, + 710 + ], + "spans": [ + { + "bbox": [ + 406, + 698, + 553, + 710 + ], + "type": "interline_equation", + "content": "\\hat {y} = W _ {F} \\cdot \\boldsymbol {a}, \\tag {19}", + "image_path": "fe97c23677b0223ddcda674ba18a30c9c5c3adfaf4881799efaefb9b06491f43.jpg" + } + ] + } + ], + "index": 27 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "text", + "content": "30203" + } + ] + } + ], + "index": 28 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 4 + }, + { + "para_blocks": [ + { + "bbox": [ + 55, + 72, + 297, + 97 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 72, + 297, + 97 + ], + "spans": [ + { + "bbox": [ + 55, + 72, + 297, + 97 + ], + "type": "text", + "content": "The classification objective of stage 2 " + }, + { + "bbox": [ + 55, + 72, + 297, + 97 + ], + "type": "inline_equation", + "content": "(f_{\\mathcal{C}\\to \\mathcal{Y}})" + }, + { + "bbox": [ + 55, + 72, + 297, + 97 + ], + "type": "text", + "content": " is the cross-entropy loss " + }, + { + "bbox": [ + 55, + 72, + 297, + 97 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{\\mathrm{CE}}(y,\\hat{y})" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 55, + 106, + 132, + 121 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 106, + 132, + 121 + ], + "spans": [ + { + "bbox": [ + 55, + 106, + 132, + 121 + ], + "type": "text", + "content": "4. Experiment" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 55, + 126, + 296, + 270 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 126, + 296, + 270 + ], + "spans": [ + { + "bbox": [ + 55, + 126, + 296, + 270 + ], + "type": "text", + "content": "To evaluate the effectiveness of our proposed DOT-CBM, we conduct experiments on the image classification task and part detection task. To further demonstrate our method's abilities to reduce spurious correlations, we also conduct experiments on out-of-distribution generalization tasks. For the concept-to-class linear layer, we omit explicit sparse regularization (e.g., L1 penalties) during training. Despite this, the model spontaneously exhibits significant sparsity in the weight matrix due to two orthogonal losses. Comprehensive training details, training efficiency results, quantitative evaluations of sparsity and additional experiments are listed in the Appendix." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 55, + 276, + 178, + 289 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 276, + 178, + 289 + ], + "spans": [ + { + "bbox": [ + 55, + 276, + 178, + 289 + ], + "type": "text", + "content": "4.1. Datasets and Metrics" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 55, + 295, + 295, + 486 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 295, + 295, + 486 + ], + "spans": [ + { + "bbox": [ + 55, + 295, + 295, + 486 + ], + "type": "text", + "content": "For the image classification task. We evaluate our method on the CUB [41] and AwA2 [43] datasets, which are widely adopted in the interpretable community. The CUB dataset contains 11,788 images of 200 bird subcategories, and we use the official train-test split for fine-grained image classification. The AwA2 dataset includes 37,322 images across 50 animal categories, and we follow the same train-test split as in [17]. Additionally, we conduct experiments on the larger and more complex ImageNet-1K[9] and CIFAR-100[19] datasets. On these classification datasets, we leverage the classification accuracy to measure the performance of our method. We also conducted experiments on a more complex dataset (Places365), where the complexity arises from discriminative content being image backgrounds. Quantitative and qualitative results are listed in Appendix." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 55, + 487, + 295, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 487, + 295, + 714 + ], + "spans": [ + { + "bbox": [ + 55, + 487, + 295, + 714 + ], + "type": "text", + "content": "For the part detection task. To evaluate the accuracy of the inversion mask produced by DOT-CBM, we evaluate our method on some part detection benchmarks, including CUB dataset, Part-ImageNet [13] and PASCAL-Part [5] datasets, which are widely used for part segmentation tasks. We also test our model on the RIVAL-10 dataset [28], which is specifically designed for evaluating model interpretability. On the CUB dataset, we obtain the bounding boxes for local parts that are commonly used in part detection tasks[44, 51]. The RIVAL-10 dataset consists of over 26K images retrieved from the ImageNet dataset according to the CIFAR-10 [19] categories, and it includes part attribute segmentation masks for the main subjects in the images. The Part-ImageNet and PASCAL-Part datasets are part segmentation datasets that provide fine-grained attribute segmentation masks. We follow [30] and preprocess these datasets by focusing on the animal data and merging overly fine-grained part attribute masks. The detailed data preprocessing procedures are described in the Appendix. We use the" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 313, + 72, + 555, + 109 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 72, + 555, + 109 + ], + "spans": [ + { + "bbox": [ + 313, + 72, + 555, + 109 + ], + "type": "text", + "content": "bounding boxes of these masks as the ground truth of part detection. On these datasets, we use the " + }, + { + "bbox": [ + 313, + 72, + 555, + 109 + ], + "type": "inline_equation", + "content": "mAP_{0.5}" + }, + { + "bbox": [ + 313, + 72, + 555, + 109 + ], + "type": "text", + "content": " metric to evaluate the precision of predicted concept inversion masks." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 313, + 116, + 448, + 129 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 116, + 448, + 129 + ], + "spans": [ + { + "bbox": [ + 313, + 116, + 448, + 129 + ], + "type": "text", + "content": "4.2. Implementation Details" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 313, + 135, + 555, + 327 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 135, + 555, + 327 + ], + "spans": [ + { + "bbox": [ + 313, + 135, + 555, + 327 + ], + "type": "text", + "content": "We employ the pre-trained DINOV2 ViT-L/14 [32] as the image encoder and pre-trained CLIP text encoder [33] as the text encoder. The feature dimension in both encoders is " + }, + { + "bbox": [ + 313, + 135, + 555, + 327 + ], + "type": "inline_equation", + "content": "d = 768" + }, + { + "bbox": [ + 313, + 135, + 555, + 327 + ], + "type": "text", + "content": ". We resize all images to " + }, + { + "bbox": [ + 313, + 135, + 555, + 327 + ], + "type": "inline_equation", + "content": "H \\times W = 252 \\times 252" + }, + { + "bbox": [ + 313, + 135, + 555, + 327 + ], + "type": "text", + "content": " as input resolution in both training and testing phases throughout all experiments. For both adapter " + }, + { + "bbox": [ + 313, + 135, + 555, + 327 + ], + "type": "inline_equation", + "content": "A_v" + }, + { + "bbox": [ + 313, + 135, + 555, + 327 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 313, + 135, + 555, + 327 + ], + "type": "inline_equation", + "content": "A_t" + }, + { + "bbox": [ + 313, + 135, + 555, + 327 + ], + "type": "text", + "content": ", we use a three-layer MLP with ReLU activation functions and a dropout rate of 0.4. The optimization of DOT-CBM is done by AdamW with an initialized learning rate of " + }, + { + "bbox": [ + 313, + 135, + 555, + 327 + ], + "type": "inline_equation", + "content": "1e - 4" + }, + { + "bbox": [ + 313, + 135, + 555, + 327 + ], + "type": "text", + "content": ", and decreased every ten epochs by a factor of 0.5. We set weight decay of 10-5. Batch sizes are set as 128 with maximally 200 training epochs. All experiments are conducted on a single RTX 3090 GPU and our model is implemented in PyTorch. We set " + }, + { + "bbox": [ + 313, + 135, + 555, + 327 + ], + "type": "inline_equation", + "content": "\\lambda_1 = 0.15" + }, + { + "bbox": [ + 313, + 135, + 555, + 327 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 313, + 135, + 555, + 327 + ], + "type": "inline_equation", + "content": "\\lambda_2 = 0.2" + }, + { + "bbox": [ + 313, + 135, + 555, + 327 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 313, + 135, + 555, + 327 + ], + "type": "inline_equation", + "content": "\\lambda_3 = 0.8" + }, + { + "bbox": [ + 313, + 135, + 555, + 327 + ], + "type": "text", + "content": " for CUB dataset. Hyperparameters setting on other datasets and the range ablation study are in the Appendix." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 313, + 335, + 400, + 346 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 335, + 400, + 346 + ], + "spans": [ + { + "bbox": [ + 313, + 335, + 400, + 346 + ], + "type": "text", + "content": "4.3. Main Results" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 313, + 354, + 556, + 569 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 354, + 556, + 569 + ], + "spans": [ + { + "bbox": [ + 313, + 354, + 556, + 569 + ], + "type": "text", + "content": "Baselines. We compare our proposed DOT-CBM with several state-of-the-art CBM baselines from the literature. Vanilla-CBM [17] is the first CBM model, proposing the design of the intermediate concept layer. Vanilla-CBM introduces three training strategies, i.e. independent, sequential and joint training, in which the independent training can achieve the best interpretation [26]. Our two-stage optimization belongs to the independent training strategy. Additionally, we compare our performance with recent CBM variants including CEM [48], LaBo [45], SparseCBM [38], CoopCBM [39]. Due to the fact that pre-defined concepts and image backbone will significantly affect the final performance. For a fair comparison, we re-implement all baselines with the high-quality concepts produced by the CDL [47] and a pre-trained ViT backbone to extract image features, the same as our method. The performance of all re-implemented baselines exceeds the best performance reported in their papers." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 313, + 570, + 556, + 677 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 570, + 556, + 677 + ], + "spans": [ + { + "bbox": [ + 313, + 570, + 556, + 677 + ], + "type": "text", + "content": "Image classification results. As shown in Table 1, we compare our proposed DOT-CBM with all baselines on four common image classification benchmarks. The classification results show that the DOT-CBM achieves superior performance on three benchmarks. Specifically, the performance improvements over the state-of-the-art (SOTA) are 0.91 points on ImageNet, 3.29 points on CUB, and 1.08 points on CIFAR100. Our method only lags behind CoopCBM by 0.25 points on the AwA2 dataset." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 313, + 677, + 556, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 677, + 556, + 715 + ], + "spans": [ + { + "bbox": [ + 313, + 677, + 556, + 715 + ], + "type": "text", + "content": "Part detection results. A key property of our proposed DOT-CBM model is able to provide explicit explanations for the concept predictions, via the locate the concept to the" + } + ] + } + ], + "index": 12 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "type": "text", + "content": "30204" + } + ] + } + ], + "index": 13 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 5 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 56, + 70, + 556, + 195 + ], + "blocks": [ + { + "bbox": [ + 56, + 70, + 556, + 195 + ], + "lines": [ + { + "bbox": [ + 56, + 70, + 556, + 195 + ], + "spans": [ + { + "bbox": [ + 56, + 70, + 556, + 195 + ], + "type": "table", + "html": "
MethodClassification Accuracy (↑)Part Detection mAP0.5 (↑)
ImageNetCUBCIFAR100AWA2PartImageNetCUBRIVALPASCAL-Parts
Vanilla-CBM [17]79.1778.3280.0493.1526.9427.0226.7617.32
CEM [48]81.2980.4781.2395.9229.1930.8630.8320.85
LaBo [45]82.9381.3084.1096.9238.2739.2840.3430.17
SparseCBM [38]82.8582.0784.7595.5639.8440.2340.6833.71
CoopCBM [39]82.7382.1084.6697.0841.1745.1643.8235.75
DOT-CBM83.84(+0.91)85.39(+3.29)85.83(+1.08)96.83(-0.25)50.12(+8.95)53.47(+8.31)50.93(+7.11)44.18(+8.43)
", + "image_path": "2fb6c11fcb69f793cac2a8e72ab4e09ac7d847b3a45326b8cc983657e3f6155e.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "table_body" + } + ], + "index": 0 + }, + { + "type": "table", + "bbox": [ + 57, + 219, + 291, + 355 + ], + "blocks": [ + { + "bbox": [ + 130, + 198, + 478, + 209 + ], + "lines": [ + { + "bbox": [ + 130, + 198, + 478, + 209 + ], + "spans": [ + { + "bbox": [ + 130, + 198, + 478, + 209 + ], + "type": "text", + "content": "Table 1. Performance comparison of Classification Accuracy (%) and Part Detection " + }, + { + "bbox": [ + 130, + 198, + 478, + 209 + ], + "type": "inline_equation", + "content": "\\left( {mA}{P}_{0.5}\\right)" + }, + { + "bbox": [ + 130, + 198, + 478, + 209 + ], + "type": "text", + "content": " ." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 57, + 219, + 291, + 355 + ], + "lines": [ + { + "bbox": [ + 57, + 219, + 291, + 355 + ], + "spans": [ + { + "bbox": [ + 57, + 219, + 291, + 355 + ], + "type": "table", + "html": "
MethodCUBDogs
IDOODIDOOD
Vanilla-CBM [17]86.927.787.329.4
CEM [48]84.034.083.836.5
LaBo [45]85.339.485.941.7
Sparse-CBM [38]84.535.885.838.6
Coop-CBM [39]85.836.286.140.3
DOT-CBM w/o. prior86.536.987.040.8
DOT-CBM w. prior87.249.789.152.4
(+0.3)(+10.3)(+1.8)(+10.7)
", + "image_path": "5448c68cfb65eff58dcded3813da7d0a5836f418d1ce6398a436420e81a34413.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "table_body" + } + ], + "index": 2 + }, + { + "bbox": [ + 70, + 357, + 280, + 367 + ], + "lines": [ + { + "bbox": [ + 70, + 357, + 280, + 367 + ], + "spans": [ + { + "bbox": [ + 70, + 357, + 280, + 367 + ], + "type": "text", + "content": "Table 2. Performance comparison of OOD generalization." + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "text" + }, + { + "bbox": [ + 54, + 378, + 295, + 617 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 54, + 378, + 295, + 617 + ], + "spans": [ + { + "bbox": [ + 54, + 378, + 295, + 617 + ], + "type": "text", + "content": "original image space, i.e. the concept inversion mask derived from the transport assignment. This concept inversion mask in the image space indicates the most influential image region contributing to the concept predictions. By contrast, all previous CBMs, using a black-box neuron network to model the mapping from global image features to concept predictions, can only inverse the concept prediction to the image space via the post-hoc Grad-CAM. To evaluate the accuracy of concept inversion mask, we conduct experiments on the part detection datasets and use the groundtruth bounding boxes of the parts to assess the localization performance. To measure the similarity with groundtruths, we need to process our concept inversion mask output and heatmap outputs of baselines (derived from Grad-CAM) into rectangle boxes format. Specifically, we set the threshold at 0.6 of the maximum peak value output of the mask. The local regions above this threshold are considered as the corresponding regions of the concept in the original image. We take the bounding box that encloses this region as the predicted bounding box for the concept." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 55, + 617, + 295, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 617, + 295, + 714 + ], + "spans": [ + { + "bbox": [ + 55, + 617, + 295, + 714 + ], + "type": "text", + "content": "From Table. 1, we can observe that DOT-CBM significantly surpasses all baselines (average 8 points improvements), demonstrating the superiority of our design for fine-grained patch-concept alignment. The Coop-CBM achieves performance that is second only to ours due to the introduction of an orthogonality loss similar to our proposed textual orthogonal project " + }, + { + "bbox": [ + 55, + 617, + 295, + 714 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{\\mathrm{TOP}}" + }, + { + "bbox": [ + 55, + 617, + 295, + 714 + ], + "type": "text", + "content": ", which further proves the validity of our proposed disentangled OT framework. Consequently," + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 313, + 220, + 555, + 269 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 220, + 555, + 269 + ], + "spans": [ + { + "bbox": [ + 313, + 220, + 555, + 269 + ], + "type": "text", + "content": "DOT-CBM can inherently produce accurate and explicit explanations for intermediate concept predictions. We also conduct comparison with other ViT-appropriate attribution method. Results shown in Appendix." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 313, + 275, + 495, + 288 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 275, + 495, + 288 + ], + "spans": [ + { + "bbox": [ + 313, + 275, + 495, + 288 + ], + "type": "text", + "content": "4.4. Out-of-distribution generalization" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 313, + 293, + 555, + 594 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 293, + 555, + 594 + ], + "spans": [ + { + "bbox": [ + 313, + 293, + 555, + 594 + ], + "type": "text", + "content": "As discussed in the previous sections, there exists data bias in many datasets, leading to shortcut learning. For example, classifying a dog as a huskie based on the snowy background due to the high co-occurrence in the training set, which is analogous to the part-background spurious correlation discussed in Fig. 1b. To validate the effectiveness of our method with background shifts, we follow [39] to segment the foreground object in CUB[41] and Dogs[15] datasets using SAM [16] with bounding box annotations and add a colored background to all images. Each class is associated with a randomly generated color background with an " + }, + { + "bbox": [ + 313, + 293, + 555, + 594 + ], + "type": "inline_equation", + "content": "80\\%" + }, + { + "bbox": [ + 313, + 293, + 555, + 594 + ], + "type": "text", + "content": " probability in the training set. The in-distribution (ID) test set contains images with a similar color background probability as the training set, while the color similarity of out-of-distribution (OOD) test set is reduced to " + }, + { + "bbox": [ + 313, + 293, + 555, + 594 + ], + "type": "inline_equation", + "content": "30\\%" + }, + { + "bbox": [ + 313, + 293, + 555, + 594 + ], + "type": "text", + "content": ". Results are shown in Table 2. As we can see, DOT-CBM can achieve comparable results on ID set and surpass the performance of baselines on the OOD set by a large margin, demonstrating that DOT-CBM can effectively decrease the biased association between object concepts and background. To investigate whether the improvements of tackling part-background spurious correlation are derived from the patch set prior modeling as mentioned in Sec. 3.4, we conduct ablations for the patch set prior and demonstrate the efficacy of patch set prior " + }, + { + "bbox": [ + 313, + 293, + 555, + 594 + ], + "type": "inline_equation", + "content": "\\theta" + }, + { + "bbox": [ + 313, + 293, + 555, + 594 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 313, + 599, + 416, + 611 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 599, + 416, + 611 + ], + "spans": [ + { + "bbox": [ + 313, + 599, + 416, + 611 + ], + "type": "text", + "content": "4.5. Ablation Studies" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 313, + 617, + 555, + 689 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 617, + 555, + 689 + ], + "spans": [ + { + "bbox": [ + 313, + 617, + 555, + 689 + ], + "type": "text", + "content": "In this section, we investigate the effects of each component in our framework. Ablation results are listed in Table 3. The baseline has the same trainable networks as DOT-CBM, including two adapters and a linear classification layer. The baseline replaces the OT assignment with a dot product of patch features and concept features." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 313, + 689, + 555, + 712 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 689, + 555, + 712 + ], + "spans": [ + { + "bbox": [ + 313, + 689, + 555, + 712 + ], + "type": "text", + "content": "Effect of OT design (OT). For the comparison between the first row and second row in Table. 3, we can find the" + } + ] + } + ], + "index": 11 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "text", + "content": "30205" + } + ] + } + ], + "index": 12 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 6 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 63, + 70, + 545, + 166 + ], + "blocks": [ + { + "bbox": [ + 63, + 70, + 545, + 166 + ], + "lines": [ + { + "bbox": [ + 63, + 70, + 545, + 166 + ], + "spans": [ + { + "bbox": [ + 63, + 70, + 545, + 166 + ], + "type": "table", + "html": "
OTPriorLrDClassification Accuracy (↑)Part Detection mIOU (↑)
ImageNetCUBCIFAR100AWA2PartImageNetCUBRIVALPASCAL-Parts
XXX81.6080.9282.7396.070.350.410.370.28
✓XX82.0381.7582.9196.240.420.490.420.34
✓✓X83.2284.9285.4996.470.500.610.530.46
XX✓82.7382.1084.6697.080.470.580.490.41
✓✓✓83.8485.3985.8396.830.520.660.540.49
", + "image_path": "d2bbfdaf42f92d062a40b52d41705ee27235cfbf92af7caf2552997f2c329b90.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "table_body" + } + ], + "index": 0 + }, + { + "bbox": [ + 55, + 167, + 553, + 190 + ], + "lines": [ + { + "bbox": [ + 55, + 167, + 553, + 190 + ], + "spans": [ + { + "bbox": [ + 55, + 167, + 553, + 190 + ], + "type": "text", + "content": "Table 3. Ablation studies on image classification and part detection. OT, Prior, and LrD denote the optimal transport design, prior distribution modeling, and local representation disentanglement." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "text" + }, + { + "type": "image", + "bbox": [ + 80, + 193, + 528, + 339 + ], + "blocks": [ + { + "bbox": [ + 80, + 193, + 528, + 339 + ], + "lines": [ + { + "bbox": [ + 80, + 193, + 528, + 339 + ], + "spans": [ + { + "bbox": [ + 80, + 193, + 528, + 339 + ], + "type": "image", + "image_path": "1d71a38ad4c7dbfe9cb6d33181883381d0462270320f24079438f568e80b6584.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 54, + 341, + 555, + 364 + ], + "lines": [ + { + "bbox": [ + 54, + 341, + 555, + 364 + ], + "spans": [ + { + "bbox": [ + 54, + 341, + 555, + 364 + ], + "type": "text", + "content": "Figure 4. Qualitative results for the concept inversion. VanillaCBM and CoopCBM adopt Grad-CAM to provide a concept inversion heatmap, while our proposed DOT-CBM can generate the inversion mask on its own." + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_caption" + } + ], + "index": 2 + }, + { + "bbox": [ + 54, + 376, + 295, + 459 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 54, + 376, + 295, + 459 + ], + "spans": [ + { + "bbox": [ + 54, + 376, + 295, + 459 + ], + "type": "text", + "content": "incorporation of OT distance, mentioned in Sec. 3.3, can slightly improve the classification accuracy and largely enhance the part detection performance. OT provides a better distance metric space for modeling the complex many-to-many correlation between concept features and patch features, compared to the direct dot product similarity used in the baseline model." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 54, + 460, + 295, + 581 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 54, + 460, + 295, + 581 + ], + "spans": [ + { + "bbox": [ + 54, + 460, + 295, + 581 + ], + "type": "text", + "content": "Effect of prior distribution modeling (Prior). The OT design above, setting the two prior distributions " + }, + { + "bbox": [ + 54, + 460, + 295, + 581 + ], + "type": "inline_equation", + "content": "\\theta" + }, + { + "bbox": [ + 54, + 460, + 295, + 581 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 54, + 460, + 295, + 581 + ], + "type": "inline_equation", + "content": "\\gamma" + }, + { + "bbox": [ + 54, + 460, + 295, + 581 + ], + "type": "text", + "content": " as a uniform distribution. To mitigate the shortcut learning brought by high occurrence between concept and background and between concept and concept, we add the background penalty patch set prior and the co-occurrence penalty concept set prior, mentioned in Sec. 3.4. Comparing the second and third rows of Table. 3, it demonstrates that these two prior distributions can further improve the classification and interpretability." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 54, + 581, + 295, + 665 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 54, + 581, + 295, + 665 + ], + "spans": [ + { + "bbox": [ + 54, + 581, + 295, + 665 + ], + "type": "text", + "content": "Effect of local representation disentanglement (LrD). After integrating the local representation disentanglement into the baseline (the fourth row of Table. 3), the improvements are larger than a single incorporation of OT distance (the second row of Table. 3). The overall DOT-CBM, with a combination of the OT, Prior, and LrD, yields more improvements than individual modules alone." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 54, + 665, + 296, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 54, + 665, + 296, + 714 + ], + "spans": [ + { + "bbox": [ + 54, + 665, + 296, + 714 + ], + "type": "text", + "content": "Qualitative results. Fig. 4 has shown a qualitative comparison between our method and the baseline methods Vanilla-CBM and CoopCBM. It is observed that our method is able to more accurately locate the region of the" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 313, + 376, + 555, + 472 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 376, + 555, + 472 + ], + "spans": [ + { + "bbox": [ + 313, + 376, + 555, + 472 + ], + "type": "text", + "content": "concept in the original image compared to the baseline methods. Meanwhile, We also present qualitative results by sequentially ablating each component of our framework. When the local representation disentanglement is ablated, the model's localization accuracy decreases to some extent. Further removal of the priors leads to the emergence of part-background spurious correlations. When the OT design is removed, the model produces incorrect localization results." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 313, + 480, + 390, + 493 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 480, + 390, + 493 + ], + "spans": [ + { + "bbox": [ + 313, + 480, + 390, + 493 + ], + "type": "text", + "content": "5. Conclusion." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 313, + 498, + 555, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 498, + 555, + 713 + ], + "spans": [ + { + "bbox": [ + 313, + 498, + 555, + 713 + ], + "type": "text", + "content": "In this paper, we introduced the Disentangled Optimal Transport Concept Bottleneck Model (DOT-CBM) to address key limitations in existing Concept Bottleneck Models. While CBMs offer interpretable models by mapping image features to human-understandable concepts, they suffer from misalignment between the image and concept, leading to spurious correlations that hinder both interpretability and generalization. To overcome these challenges, DOT-CBM introduces a disentangled optimal transport framework that ensures fine-grained alignment between image patches and corresponding textual concepts, thereby providing a more robust and interpretable model. By employing a dual penalty system, we mitigate the overfitting problem caused by data bias. Empirical results demonstrate that DOT-CBM outperforms previous CBM variants in both classification accuracy and interpretability, offering an effective solution to the problem of visual-concept spurious correlation." + } + ] + } + ], + "index": 10 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "type": "text", + "content": "30206" + } + ] + } + ], + "index": 11 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 7 + }, + { + "para_blocks": [ + { + "bbox": [ + 55, + 72, + 165, + 85 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 72, + 165, + 85 + ], + "spans": [ + { + "bbox": [ + 55, + 72, + 165, + 85 + ], + "type": "text", + "content": "6. Acknowledgement" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 55, + 90, + 297, + 213 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 90, + 297, + 213 + ], + "spans": [ + { + "bbox": [ + 55, + 90, + 297, + 213 + ], + "type": "text", + "content": "This work was supported in part by the National Natural Science Foundation of China under Grant U21B2006; in part by Shaanxi Youth Innovation Team Project; in part by the Fundamental Research Funds for the Central Universities QTZX24003 and QTZX22160; in part by the 111 Project under Grant B18039; Hao Zhang acknowledges the support of NSFC (62301384); Excellent Young Scientists Fund (Overseas); Foundation of National Key Laboratory of Radar Signal Processing under Grant JKW202308. Zhengjue Wang acknowledges the support of NSFC (62301407)." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 56, + 232, + 115, + 244 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 232, + 115, + 244 + ], + "spans": [ + { + "bbox": [ + 56, + 232, + 115, + 244 + ], + "type": "text", + "content": "References" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 57, + 251, + 295, + 714 + ], + "type": "list", + "angle": 0, + "index": 14, + "blocks": [ + { + "bbox": [ + 61, + 251, + 294, + 283 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 61, + 251, + 294, + 283 + ], + "spans": [ + { + "bbox": [ + 61, + 251, + 294, + 283 + ], + "type": "text", + "content": "[1] Dosovitskiy Alexey. An image is worth 16x16 words: Transformers for image recognition at scale. arXiv preprint arXiv: 2010.11929, 2020.4" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 62, + 285, + 295, + 328 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 285, + 295, + 328 + ], + "spans": [ + { + "bbox": [ + 62, + 285, + 295, + 328 + ], + "type": "text", + "content": "[2] Kushal Chauhan, Rishabh Tiwari, Jan Freyberg, Pradeep Shenoy, and Krishnamurthy Dvijotham. Interactive concept bottleneck models. In Proceedings of the AAAI Conference on Artificial Intelligence, pages 5948-5955, 2023. 2" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 62, + 329, + 295, + 371 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 329, + 295, + 371 + ], + "spans": [ + { + "bbox": [ + 62, + 329, + 295, + 371 + ], + "type": "text", + "content": "[3] Guangyi Chen, Weiran Yao, Xiangchen Song, Xinyue Li, Yongming Rao, and Kun Zhang. Plot: Prompt learning with optimal transport for vision-language models. arXiv preprint arXiv:2210.01253, 2022. 3" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 62, + 373, + 294, + 416 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 373, + 294, + 416 + ], + "spans": [ + { + "bbox": [ + 62, + 373, + 294, + 416 + ], + "type": "text", + "content": "[4] Liquin Chen, Zhe Gan, Yu Cheng, Linjie Li, Lawrence Carin, and Jingjing Liu. Graph optimal transport for cross-domain alignment. In International Conference on Machine Learning, pages 1542-1553. PMLR, 2020. 3" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 62, + 418, + 294, + 482 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 418, + 294, + 482 + ], + "spans": [ + { + "bbox": [ + 62, + 418, + 294, + 482 + ], + "type": "text", + "content": "[5] Xianjie Chen, Roozbeh Mottaghi, Xiaobai Liu, Sanja Fidler, Raquel Urtasun, and Alan Yuille. Detect what you can: Detecting and representing objects using holistic models and body parts. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 1971-1978, 2014. 6" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 62, + 483, + 294, + 505 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 483, + 294, + 505 + ], + "spans": [ + { + "bbox": [ + 62, + 483, + 294, + 505 + ], + "type": "text", + "content": "[6] Molnar Christoph. Interpretable machine learning: A guide for making black box models explainable. Leanpub, 2020. 1" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 62, + 506, + 294, + 537 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 506, + 294, + 537 + ], + "spans": [ + { + "bbox": [ + 62, + 506, + 294, + 537 + ], + "type": "text", + "content": "[7] Marco Cuturi. Sinkhorn distances: Lightspeed computation of optimal transport. Advances in neural information processing systems, 26, 2013. 4" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 62, + 538, + 294, + 559 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 538, + 294, + 559 + ], + "spans": [ + { + "bbox": [ + 62, + 538, + 294, + 559 + ], + "type": "text", + "content": "[8] Barry De Ville. Decision trees. Wiley Interdisciplinary Reviews: Computational Statistics, 5(6):448-455, 2013. 1" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 62, + 560, + 295, + 603 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 560, + 295, + 603 + ], + "spans": [ + { + "bbox": [ + 62, + 560, + 295, + 603 + ], + "type": "text", + "content": "[9] Jia Deng, Wei Dong, Richard Socher, Li-Jia Li, Kai Li, and Li Fei-Fei. Imagenet: A large-scale hierarchical image database. In 2009 IEEE conference on computer vision and pattern recognition, pages 248-255. IEEE, 2009. 6" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 57, + 604, + 295, + 678 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 604, + 295, + 678 + ], + "spans": [ + { + "bbox": [ + 57, + 604, + 295, + 678 + ], + "type": "text", + "content": "[10] Mateo Espinosa Zarlenga, Pietro Barbiero, Gabriele Ciravegna, Giuseppe Marra, Francesco Giannini, Michelangelo Diligenti, Zohreh Shams, Frederic Precioso, Stefano Melacci, Adrian Weller, et al. Concept embedding models: Beyond the accuracy-explainability trade-off. Advances in Neural Information Processing Systems, 35:21400-21413, 2022. 2" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 57, + 680, + 295, + 714 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 680, + 295, + 714 + ], + "spans": [ + { + "bbox": [ + 57, + 680, + 295, + 714 + ], + "type": "text", + "content": "[11] Mateo Espinosa Zarlenga, Katie Collins, Krishnamurthy Dvijotham, Adrian Weller, Zohreh Shams, and Mateja Jamnik. Learning to receive help: Intervention-aware concept" + } + ] + } + ], + "index": 13 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 316, + 73, + 553, + 714 + ], + "type": "list", + "angle": 0, + "index": 29, + "blocks": [ + { + "bbox": [ + 335, + 73, + 553, + 95 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 335, + 73, + 553, + 95 + ], + "spans": [ + { + "bbox": [ + 335, + 73, + 553, + 95 + ], + "type": "text", + "content": "embedding models. Advances in Neural Information Processing Systems, 36, 2024. 2" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 316, + 97, + 553, + 139 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 97, + 553, + 139 + ], + "spans": [ + { + "bbox": [ + 316, + 97, + 553, + 139 + ], + "type": "text", + "content": "[12] Charlie Frogner, Chiyuan Zhang, Hossein Mobahi, Mauricio Araya, and Tomaso A Poggio. Learning with a wasserstein loss. Advances in neural information processing systems, 28, 2015. 3" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 316, + 141, + 553, + 196 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 141, + 553, + 196 + ], + "spans": [ + { + "bbox": [ + 316, + 141, + 553, + 196 + ], + "type": "text", + "content": "[13] Ju He, Shuo Yang, Shaokang Yang, Adam Kortylewski, Xiaoding Yuan, Jie-Neng Chen, Shuai Liu, Cheng Yang, Qihang Yu, and Alan Yuille. Partimagenet: A large, high-quality dataset of parts. In European Conference on Computer Vision, pages 128-145. Springer, 2022. 6" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 316, + 198, + 553, + 251 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 198, + 553, + 251 + ], + "spans": [ + { + "bbox": [ + 316, + 198, + 553, + 251 + ], + "type": "text", + "content": "[14] Johan Huysmans, Karel Dejaeger, Christophe Mues, Jan Vanthienen, and Bart Baesens. An empirical evaluation of the comprehensibility of decision table, tree and rule based predictive models. Decision Support Systems, 51(1):141-154, 2011. 1" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 316, + 254, + 553, + 308 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 254, + 553, + 308 + ], + "spans": [ + { + "bbox": [ + 316, + 254, + 553, + 308 + ], + "type": "text", + "content": "[15] Aditya Khosla, Nityananda Jayadevaprakash, Bangpeng Yao, and Li Fei-Fei. Novel dataset for fine-grained image categorization. In First Workshop on Fine-Grained Visual Categorization, IEEE Conference on Computer Vision and Pattern Recognition, Colorado Springs, CO, 2011. 7" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 316, + 309, + 553, + 364 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 309, + 553, + 364 + ], + "spans": [ + { + "bbox": [ + 316, + 309, + 553, + 364 + ], + "type": "text", + "content": "[16] Alexander Kirillov, Eric Mintun, Nikhila Ravi, Hanzi Mao, Chloe Rolland, Laura Gustafson, Tete Xiao, Spencer Whitehead, Alexander C Berg, Wan-Yen Lo, et al. Segment anything. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 4015-4026, 2023. 7" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 316, + 365, + 553, + 419 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 365, + 553, + 419 + ], + "spans": [ + { + "bbox": [ + 316, + 365, + 553, + 419 + ], + "type": "text", + "content": "[17] Pang Wei Koh, Thao Nguyen, Yew Siang Tang, Stephen Mussmann, Emma Pierson, Been Kim, and Percy Liang. Concept bottleneck models. In International conference on machine learning, pages 5338-5348. PMLR, 2020. 1, 2, 3, 6, 7" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 316, + 421, + 553, + 453 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 421, + 553, + 453 + ], + "spans": [ + { + "bbox": [ + 316, + 421, + 553, + 453 + ], + "type": "text", + "content": "[18] Andrei V Konstantinov and Lev V Utkin. Incorporating expert rules into neural networks in the framework of concept-based learning. arXiv preprint arXiv:2402.14726, 2024. 2" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 316, + 455, + 553, + 477 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 455, + 553, + 477 + ], + "spans": [ + { + "bbox": [ + 316, + 455, + 553, + 477 + ], + "type": "text", + "content": "[19] Alex Krizhevsky, Geoffrey Hinton, et al. Learning multiple layers of features from tiny images. 2009. 6" + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 316, + 479, + 553, + 521 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 479, + 553, + 521 + ], + "spans": [ + { + "bbox": [ + 316, + 479, + 553, + 521 + ], + "type": "text", + "content": "[20] John Lee, Max Dabagia, Eva Dyer, and Christopher Rozell. Hierarchical optimal transport for multimodal distribution alignment. Advances in neural information processing systems, 32, 2019. 3" + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 316, + 523, + 553, + 589 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 523, + 553, + 589 + ], + "spans": [ + { + "bbox": [ + 316, + 523, + 553, + 589 + ], + "type": "text", + "content": "[21] Miaoge Li, Dongsheng Wang, Xinyang Liu, Zequn Zeng, Ruiying Lu, Bo Chen, and Mingyuan Zhou. Patchct: Aligning patch set and label set with conditional transport for multi-label image classification. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 15348-15358, 2023. 3" + } + ] + } + ], + "index": 25 + }, + { + "bbox": [ + 316, + 590, + 553, + 634 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 590, + 553, + 634 + ], + "spans": [ + { + "bbox": [ + 316, + 590, + 553, + 634 + ], + "type": "text", + "content": "[22] Tang Li, Mengmeng Ma, and Xi Peng. Deal: Disentangle and localize concept-level explanations for vlms. In European Conference on Computer Vision, pages 383-401. Springer, 2025. 2, 4" + } + ] + } + ], + "index": 26 + }, + { + "bbox": [ + 316, + 635, + 553, + 677 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 635, + 553, + 677 + ], + "spans": [ + { + "bbox": [ + 316, + 635, + 553, + 677 + ], + "type": "text", + "content": "[23] Yijie Lin, Jie Zhang, Zhenyu Huang, Jia Liu, Zujie Wen, and Xi Peng. Multi-granularity correspondence learning from long-term noisy videos. arXiv preprint arXiv:2401.16702, 2024. 3" + } + ] + } + ], + "index": 27 + }, + { + "bbox": [ + 316, + 680, + 553, + 714 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 680, + 553, + 714 + ], + "spans": [ + { + "bbox": [ + 316, + 680, + 553, + 714 + ], + "type": "text", + "content": "[24] Jiawei Ma, Po-Yao Huang, Saining Xie, Shang-Wen Li, Luke Zettlemoyer, Shih-Fu Chang, Wen-Tau Yih, and Hu Xu. Mode: Clip data experts via clustering. In Proceed-" + } + ] + } + ], + "index": 28 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "text", + "content": "30207" + } + ] + } + ], + "index": 30 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 8 + }, + { + "para_blocks": [ + { + "bbox": [ + 56, + 73, + 295, + 713 + ], + "type": "list", + "angle": 0, + "index": 14, + "blocks": [ + { + "bbox": [ + 77, + 73, + 295, + 95 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 77, + 73, + 295, + 95 + ], + "spans": [ + { + "bbox": [ + 77, + 73, + 295, + 95 + ], + "type": "text", + "content": "ings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 26354-26363, 2024. 2" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 57, + 95, + 295, + 148 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 95, + 295, + 148 + ], + "spans": [ + { + "bbox": [ + 57, + 95, + 295, + 148 + ], + "type": "text", + "content": "[25] Shweta Mahajan, Teresa Botschen, Iryna Gurevych, and Stefan Roth. Joint wasserstein autoencoders for aligning multimodal embeddings. In Proceedings of the IEEE/CVF International Conference on Computer Vision Workshops, pages 0-0, 2019. 3" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 56, + 150, + 294, + 194 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 150, + 294, + 194 + ], + "spans": [ + { + "bbox": [ + 56, + 150, + 294, + 194 + ], + "type": "text", + "content": "[26] Andrei Margeloiu, Matthew Ashman, Umang Bhatt, Yanzhi Chen, Mateja Jamnik, and Adrian Weller. Do concept bottleneck models learn as intended? arXiv preprint arXiv:2105.04289, 2021. 2, 3, 6" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 56, + 195, + 294, + 239 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 195, + 294, + 239 + ], + "spans": [ + { + "bbox": [ + 56, + 195, + 294, + 239 + ], + "type": "text", + "content": "[27] Nesta Midavaine, Gregory Hok Tjoan Go, Diego Canez, Ioana Simion, and Satchit Chatterji. [re] on the reproducibility of post-hoc concept bottleneck models. Transactions on Machine Learning Research. 2" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 56, + 239, + 294, + 293 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 239, + 294, + 293 + ], + "spans": [ + { + "bbox": [ + 56, + 239, + 294, + 293 + ], + "type": "text", + "content": "[28] Mazda Moayeri, Phillip Pope, Yogesh Balaji, and Soheil Feizi. A comprehensive study of image classification model sensitivity to foregrounds, backgrounds, and visual attributes. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, 2022. 6" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 56, + 294, + 294, + 349 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 294, + 294, + 349 + ], + "spans": [ + { + "bbox": [ + 56, + 294, + 294, + 349 + ], + "type": "text", + "content": "[29] Mazda Moayeri, Keivan Rezaei, Maziar Sanjabi, and Soheil Feizi. Text2concept: Concept activation vectors directly from text. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR) Workshops, pages 3744-3749, 2023. 2, 4" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 56, + 350, + 294, + 382 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 350, + 294, + 382 + ], + "spans": [ + { + "bbox": [ + 56, + 350, + 294, + 382 + ], + "type": "text", + "content": "[30] Keval Morabia, Jatin Arora, and Tara Vijaykumar. Attention-based joint detection of object and semantic part. arXiv preprint arXiv:2007.02419, 2020. 6" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 56, + 383, + 294, + 415 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 383, + 294, + 415 + ], + "spans": [ + { + "bbox": [ + 56, + 383, + 294, + 415 + ], + "type": "text", + "content": "[31] Tuomas Oikarinen, Subhro Das, Lam M Nguyen, and Tsui-Wei Weng. Label-free concept bottleneck models. arXiv preprint arXiv:2304.06129, 2023. 2" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 56, + 416, + 294, + 470 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 416, + 294, + 470 + ], + "spans": [ + { + "bbox": [ + 56, + 416, + 294, + 470 + ], + "type": "text", + "content": "[32] Maxime Oquab, Timothee Darcet, Theo Moutakanni, Huy Vo, Marc Szafraniec, Vasil Khalidov, Pierre Fernandez, Daniel Haziza, Francisco Massa, Alaaeldin El-Nouby, et al. Dinov2: Learning robust visual features without supervision. arXiv preprint arXiv:2304.07193, 2023. 6" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 56, + 471, + 294, + 536 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 471, + 294, + 536 + ], + "spans": [ + { + "bbox": [ + 56, + 471, + 294, + 536 + ], + "type": "text", + "content": "[33] Alec Radford, Jong Wook Kim, Chris Hallacy, Aditya Ramesh, Gabriel Goh, Sandhini Agarwal, Girish Sastry, Amanda Askell, Pamela Mishkin, Jack Clark, et al. Learning transferable visual models from natural language supervision. In International conference on machine learning, pages 8748-8763. PMLR, 2021. 6" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 56, + 537, + 294, + 590 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 537, + 294, + 590 + ], + "spans": [ + { + "bbox": [ + 56, + 537, + 294, + 590 + ], + "type": "text", + "content": "[34] Maithra Raghu, Thomas Unterthiner, Simon Kornblith, Chiyuan Zhang, and Alexey Dosovitskiy. Do vision transformers see like convolutional neural networks? Advances in neural information processing systems, 34:12116-12128, 2021. 4" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 56, + 592, + 294, + 635 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 592, + 294, + 635 + ], + "spans": [ + { + "bbox": [ + 56, + 592, + 294, + 635 + ], + "type": "text", + "content": "[35] Naveen Raman, Mateo Espinosa Zarlenga, Juyeon Heo, and Mateja Jamnik. Do concept bottleneck models obey locality? In XAI in Action: Past, Present, and Future Applications, 2023. 3" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 56, + 636, + 294, + 689 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 636, + 294, + 689 + ], + "spans": [ + { + "bbox": [ + 56, + 636, + 294, + 689 + ], + "type": "text", + "content": "[36] Ievgen Redko, Nicolas Courty, Rémi Flamary, and Devis Tuia. Optimal transport for multi-source domain adaptation under target shift. In The 22nd International Conference on artificial intelligence and statistics, pages 849-858. PMLR, 2019. 3" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 56, + 691, + 294, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 691, + 294, + 713 + ], + "spans": [ + { + "bbox": [ + 56, + 691, + 294, + 713 + ], + "type": "text", + "content": "[37] Ramprasaath R Selvaraju, Michael Cogswell, Abhishek Das, Ramakrishna Vedantam, Devi Parikh, and Dhruv Batra." + } + ] + } + ], + "index": 13 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 316, + 73, + 554, + 713 + ], + "type": "list", + "angle": 0, + "index": 29, + "blocks": [ + { + "bbox": [ + 333, + 73, + 553, + 116 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 333, + 73, + 553, + 116 + ], + "spans": [ + { + "bbox": [ + 333, + 73, + 553, + 116 + ], + "type": "text", + "content": "Grad-cam: Visual explanations from deep networks via gradient-based localization. In Proceedings of the IEEE international conference on computer vision, pages 618-626, 2017. 1" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 316, + 118, + 553, + 161 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 118, + 553, + 161 + ], + "spans": [ + { + "bbox": [ + 316, + 118, + 553, + 161 + ], + "type": "text", + "content": "[38] Andrei Semenov, Vladimir Ivanov, Aleksandr Beznosikov, and Alexander Gasnikov. Sparse concept bottleneck models: Gumbel tricks in contrastive learning. arXiv preprint arXiv:2404.03323, 2024. 2, 6, 7" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 316, + 163, + 553, + 196 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 163, + 553, + 196 + ], + "spans": [ + { + "bbox": [ + 316, + 163, + 553, + 196 + ], + "type": "text", + "content": "[39] Ivaxi Sheth and Samira Ebrahimi Kahou. Auxiliary losses for learning generalizable concept-based models. Advances in Neural Information Processing Systems, 36, 2024. 2, 6, 7" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 316, + 198, + 553, + 241 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 198, + 553, + 241 + ], + "spans": [ + { + "bbox": [ + 316, + 198, + 553, + 241 + ], + "type": "text", + "content": "[40] Sungbin Shin, Yohan Jo, Sungsoo Ahn, and Namhoon Lee. A closer look at the intervention procedure of concept bottleneck models. In International Conference on Machine Learning, pages 31504-31520. PMLR, 2023. 2" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 316, + 243, + 553, + 274 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 243, + 553, + 274 + ], + "spans": [ + { + "bbox": [ + 316, + 243, + 553, + 274 + ], + "type": "text", + "content": "[41] Catherine Wah, Steve Branson, Peter Welinder, Pietro Perona, and Serge Belongie. The caltech-ucsd birds-200-2011 dataset. 2011. 6, 7" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 316, + 276, + 553, + 330 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 276, + 553, + 330 + ], + "spans": [ + { + "bbox": [ + 316, + 276, + 553, + 330 + ], + "type": "text", + "content": "[42] Tiansheng Wen, Yifei Wang, Zequn Zeng, Zhong Peng, Yudi Su, Xinyang Liu, Bo Chen, Hongwei Liu, Stefanie Jegelka, and Chenyu You. Beyond matryoshka: Revisiting sparse coding for adaptive representation. arXiv preprint arXiv:2503.01776, 2025. 2" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 316, + 332, + 553, + 386 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 332, + 553, + 386 + ], + "spans": [ + { + "bbox": [ + 316, + 332, + 553, + 386 + ], + "type": "text", + "content": "[43] Yongqin Xian, Christoph H Lampert, Bernt Schiele, and Zeynep Akata. Zero-shot learning—a comprehensive evaluation of the good, the bad and the ugly. IEEE transactions on pattern analysis and machine intelligence, 41(9):2251-2265, 2018. 6" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 316, + 388, + 553, + 431 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 388, + 553, + 431 + ], + "spans": [ + { + "bbox": [ + 316, + 388, + 553, + 431 + ], + "type": "text", + "content": "[44] Wenjia Xu, Yongqin Xian, Jiuniu Wang, Bernt Schiele, and Zeynep Akata. Attribute prototype network for zero-shot learning. Advances in Neural Information Processing Systems, 33:21969-21980, 2020. 6" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 316, + 434, + 554, + 498 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 434, + 554, + 498 + ], + "spans": [ + { + "bbox": [ + 316, + 434, + 554, + 498 + ], + "type": "text", + "content": "[45] Yue Yang, Artemis Panagopoulou, Shenghao Zhou, Daniel Jin, Chris Callison-Burch, and Mark Yatskar. Language in a bottle: Language model guided concept bottlenecks for interpretable image classification. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 19187-19197, 2023. 2, 4, 6, 7" + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 316, + 501, + 553, + 533 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 501, + 553, + 533 + ], + "spans": [ + { + "bbox": [ + 316, + 501, + 553, + 533 + ], + "type": "text", + "content": "[46] Mert Yuksekgonul, Maggie Wang, and James Zou. Post-hoc concept bottleneck models. arXiv preprint arXiv:2205.15480, 2022. 2" + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 316, + 535, + 553, + 567 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 535, + 553, + 567 + ], + "spans": [ + { + "bbox": [ + 316, + 535, + 553, + 567 + ], + "type": "text", + "content": "[47] Yuan Zang, Tian Yun, Hao Tan, Trung Bui, and Chen Sun. Pre-trained vision-language models learn discoverable visual concepts. arXiv preprint arXiv:2404.12652, 2024. 2, 4, 6" + } + ] + } + ], + "index": 25 + }, + { + "bbox": [ + 316, + 568, + 553, + 633 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 568, + 553, + 633 + ], + "spans": [ + { + "bbox": [ + 316, + 568, + 553, + 633 + ], + "type": "text", + "content": "[48] Mateo Espinosa Zarlenga, Pietro Barbiero, Gabriele Ciravegna, Giuseppe Marra, Francesco Giannini, Michelangelo Diligenti, Zohreh Shams, Frederic Precioso, Stefano Melacci, Adrian Weller, et al. Concept embedding models: Beyond the accuracy-explainability trade-off. arXiv preprint arXiv:2209.09056, 2022. 6, 7" + } + ] + } + ], + "index": 26 + }, + { + "bbox": [ + 316, + 635, + 553, + 689 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 635, + 553, + 689 + ], + "spans": [ + { + "bbox": [ + 316, + 635, + 553, + 689 + ], + "type": "text", + "content": "[49] Zequn Zeng, Hao Zhang, Ruiying Lu, Dongsheng Wang, Bo Chen, and Zhengjue Wang. Conzic: Controllable zero-shot image captioning by sampling-based polishing. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 23465-23476, 2023. 2" + } + ] + } + ], + "index": 27 + }, + { + "bbox": [ + 316, + 692, + 553, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 692, + 553, + 713 + ], + "spans": [ + { + "bbox": [ + 316, + 692, + 553, + 713 + ], + "type": "text", + "content": "[50] Zequn Zeng, Yan Xie, Hao Zhang, Chiyu Chen, Bo Chen, and Zhengjue Wang. Meacap: Memory-augmented zero" + } + ] + } + ], + "index": 28 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "text", + "content": "30208" + } + ] + } + ], + "index": 30 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 9 + }, + { + "para_blocks": [ + { + "bbox": [ + 56, + 72, + 295, + 206 + ], + "type": "list", + "angle": 0, + "index": 3, + "blocks": [ + { + "bbox": [ + 75, + 72, + 294, + 106 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 75, + 72, + 294, + 106 + ], + "spans": [ + { + "bbox": [ + 75, + 72, + 294, + 106 + ], + "type": "text", + "content": "shot image captioning. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 14100-14110, 2024. 2" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 56, + 107, + 295, + 173 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 107, + 295, + 173 + ], + "spans": [ + { + "bbox": [ + 56, + 107, + 295, + 173 + ], + "type": "text", + "content": "[51] Han Zhang, Tao Xu, Mohamed Elhoseiny, Xiaolei Huang, Shaoting Zhang, Ahmed Elgammal, and Dimitris Metaxas. Spda-cnn: Unifying semantic part detection and abstraction for fine-grained recognition. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 1143-1152, 2016. 6" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 56, + 174, + 294, + 206 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 174, + 294, + 206 + ], + "spans": [ + { + "bbox": [ + 56, + 174, + 294, + 206 + ], + "type": "text", + "content": "[52] Peng Zhao and Zhi-Hua Zhou. Label distribution learning by optimal transport. In Proceedings of the AAAI Conference on Artificial Intelligence, 2018. 3" + } + ] + } + ], + "index": 2 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 749, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 749, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 749, + 317, + 757 + ], + "type": "text", + "content": "30209" + } + ] + } + ], + "index": 4 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 10 + } + ], + "_backend": "vlm", + "_version_name": "2.6.4" +} \ No newline at end of file diff --git a/2025/Discovering Hidden Visual Concepts Beyond Linguistic Input in Infant Learning/50337d5c-a1b4-4b7f-9096-7ccfa12467ab_content_list.json b/2025/Discovering Hidden Visual Concepts Beyond Linguistic Input in Infant Learning/50337d5c-a1b4-4b7f-9096-7ccfa12467ab_content_list.json new file mode 100644 index 0000000000000000000000000000000000000000..55c208b73909f1955047f5359c6c8f7844d3b7a8 --- /dev/null +++ b/2025/Discovering Hidden Visual Concepts Beyond Linguistic Input in Infant Learning/50337d5c-a1b4-4b7f-9096-7ccfa12467ab_content_list.json @@ -0,0 +1,1554 @@ +[ + { + "type": "text", + "text": "Discovering Hidden Visual Concepts Beyond Linguistic Input in Infant Learning", + "text_level": 1, + "bbox": [ + 91, + 130, + 906, + 154 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Xueyi Ke", + "bbox": [ + 235, + 178, + 321, + 199 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Satoshi Tsutsui", + "bbox": [ + 352, + 180, + 483, + 198 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Yayun Zhang", + "bbox": [ + 514, + 180, + 630, + 198 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Bihan Wen $^{1*}$", + "bbox": [ + 661, + 180, + 764, + 196 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "$^{1}$ Nanyang Technological University", + "bbox": [ + 357, + 205, + 640, + 223 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "2The Max Planck Institute for Psycholinguistics", + "bbox": [ + 307, + 224, + 684, + 241 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/104eed72a52bef9f3192e4c489bda704eb1b560438b594c9093bc850548dd7c8.jpg", + "image_caption": [ + "Figure 1. Inspired by the development of the infant visual system, CVCL [41] is trained on infant egocentric frames and transcribed parental speech (a) and demonstrates object recognition ability within the vocabulary provided by parental speech (c). However, infant visual development is not limited to parental guidance. Thus, we hypothesize that a computational model trained on an infant's daily experiences can similarly acquire visual concepts beyond its training parental speech. To explore this, we perform neuron labeling (b) to identify visual concept neurons, including concepts that were never mentioned in the parental speech vocabulary (e.g., \"rug\"). Based on discovered neurons, we show that the model can recognize objects beyond the training vocabulary, akin to early infant visual development (d). Some images are cited from prior work [41] as the original data is not accessible to us." + ], + "image_footnote": [], + "bbox": [ + 114, + 299, + 491, + 521 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/2c3dd38a2a7a357b90f5463f15ec9af1d9071757f84369491f9cea9a5fc61baa.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 500, + 299, + 885, + 521 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Abstract", + "text_level": 1, + "bbox": [ + 246, + 613, + 326, + 628 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Infants develop complex visual understanding rapidly, even preceding of the acquisition of linguistic skills. As computer vision seeks to replicate the human vision system, understanding infant visual development may offer valuable insights. In this paper, we present an interdisciplinary study exploring this question: can a computational model that imitates the infant learning process develop broader visual concepts that extend beyond the vocabulary it has heard, similar to how infants naturally learn? To investigate this, we analyze a recently published model in Science by Vong et al., which is trained on longitudinal, egocentric images of a single child paired with transcribed parental speech. We perform neuron labeling to identify visual concept neurons hidden in the model's internal representations. We then demonstrate that these neurons can", + "bbox": [ + 88, + 647, + 485, + 875 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "recognize objects beyond the model's original vocabulary. Furthermore, we compare the differences in representation between infant models and those in modern computer vision models, such as CLIP and ImageNet pre-trained model. Ultimately, our work bridges cognitive science and computer vision by analyzing the internal representations of a computational model trained on an infant visual and linguistic inputs. Our code is available at https://github.com/Kexueyi/discover_infant.vis.", + "bbox": [ + 511, + 614, + 908, + 752 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "1. Introduction", + "text_level": 1, + "bbox": [ + 513, + 784, + 645, + 800 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Infants are remarkable learners, sparking interest across various academic disciplines. Computer vision is no exception, with researchers studying infant visual learning from various perspectives [1, 31, 32, 35, 41]. A recent milestone is Child's View for Contrastive Learning (CVCL) by Vong et al.[41], which trains a model from scratch on longitudinal", + "bbox": [ + 511, + 811, + 908, + 905 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "CVF", + "bbox": [ + 106, + 2, + 181, + 42 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore.", + "bbox": [ + 236, + 0, + 810, + 46 + ], + "page_idx": 0 + }, + { + "type": "page_footnote", + "text": "*Corresponding Author.", + "bbox": [ + 107, + 888, + 240, + 901 + ], + "page_idx": 0 + }, + { + "type": "page_number", + "text": "4343", + "bbox": [ + 482, + 944, + 514, + 955 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "egocentric videos of a single infant (6-25 months) paired with transcribed parental speech to learn visual-linguistic associations, similar to CLIP [34]. The resulting model develops object recognition abilities, aligning with developmental psychology findings that infants acquire object names by linking words to visual referents [3, 17]. In this paper, we analyze the internal visual representations of CVCL to better understand its mechanism, similar to how developmental studies observe infants' internal neurons [38].", + "bbox": [ + 89, + 90, + 480, + 241 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "In developmental psychology, research on real infants suggests that their object recognition is deeply influenced by their visual experiences with the world. A headcam study of 8.5 to 10.5-month-old infants [8] showed that the first nouns infants acquire often correspond to objects they see most frequently. While infants may also hear the names of these frequently seen objects, their early visual familiarity plays a critical role in object recognition. Visual understanding may develop before the learning of corresponding names, potentially facilitating the process of word acquisition [33]. For instance, newborn infants have shown an innate ability to recognize visual patterns [13], and their development of visual concepts often precedes the emergence of verbal thought [27].", + "bbox": [ + 89, + 244, + 482, + 455 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Our Hypothesis: Based on these infant studies, we hypothesize that a computational model trained on an infant's daily experiences may similarly acquire visual concepts extending beyond its linguistic training data. While CVCL demonstrates visual-linguistic mappings of objects named in parental speech, its visual recognition capability should not be limited to parental vocabulary supervision. We conjecture that the vision encoder may develop the ability to recognize concepts beyond these linguistic training data, similar to how infants form object recognition before learning object names.", + "bbox": [ + 89, + 458, + 482, + 625 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "To investigate this, we analyze CVCL's internal representations using network dissection [2, 30], or more intuitively, neuron labeling. Furthermore, we implement a neuron-based, training-free classification framework (NeuronClassifier) that leverages visual concepts identified via neuron labeling. Relying solely on these neurons, this approach not only achieves better recognition performance than originally reported [41], but also discovers internal visual concepts extending beyond the model's training vocabulary, supporting our aligned hypothesis.", + "bbox": [ + 89, + 626, + 482, + 777 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "From a developmental psychology perspective, the visual concepts discovered beyond the training vocabulary tend to have a higher age of acquisition (AoA) [26] values (Figure 7). This results supports CVCL's ability to develop visual understanding that precedes explicit labeling, mirroring cognitive studies where infants develop pre-verbal visual concepts [27]. This likely reflects real-world learning dynamics: children first acquire labels for frequent, con", + "bbox": [ + 89, + 780, + 482, + 902 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "crete concepts (captured in CVCL's training vocabulary), while visually grounded representations for unlabeled concepts still form earlier than their eventual linguistic acquisition.", + "bbox": [ + 511, + 90, + 903, + 150 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "To further explore CVCL's internal representations in the context of computer vision, we compare its visual features with widely-used representations such as CLIP and ImageNet. While CVCL is trained on a unique infant dataset with limited exposure to diverse scenes and a much smaller dataset compared to CLIP, it exhibits similar low-level features in its early layers. However, its higher-level features in the final layer differ significantly. These differences also extend to the visual concept neurons in the model's deeper layers.", + "bbox": [ + 511, + 151, + 906, + 303 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Contributions The key contributions of our work are:", + "text_level": 1, + "bbox": [ + 513, + 306, + 872, + 321 + ], + "page_idx": 1 + }, + { + "type": "list", + "sub_type": "text", + "list_items": [ + "- We show that infant model have developed understanding beyond linguistic training inputs, by discovering visual concepts hidden in the model representations, aligning with existing infant studies.", + "- We demonstrate that the discovered visual concept neurons can improve object recognition performance by implementing a training-free framework (NeuronClassifier).", + "- We find that the infant model shares similar low-level representations with ImageNet or CLIP models but diverges in deeper layers due to a lack of diverse higher-level visual concepts." + ], + "bbox": [ + 514, + 325, + 903, + 491 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "2. Related Work", + "text_level": 1, + "bbox": [ + 513, + 506, + 653, + 522 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Learning from children. Modeling how children learn has long been a strategy for advancing artificial intelligence. Instead of directly replicating adult intelligence, Alan Turing suggested, \"why not rather try to produce one which simulates the child's?\" [40] Training models on egocentric videos or multimodal data captured from infant perspectives aligns with this idea. [1, 31, 32, 35, 39, 41], as these videos approximate the input available to human infants during development. Our study established on CVCL [41], which we explain in Section 2.1.", + "bbox": [ + 511, + 532, + 906, + 683 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Interpreting vision model representations. Our goal is to understand the model internal neurons trained on infant data. In this context, techniques for interpreting intermediate representations in deep neural networks are relevant. Beginning with Network Dissection [2], a method that quantifies alignment between hidden neurons and visual concepts, numerous studies have aimed to make black-box models more transparent. These methods enable compositional concept discovery [28], assignment of compositional concepts with statistical quantification [5], open-vocabulary neuron captioning [21], and the use of CLIP's rich embeddings for neuron-concept alignment [22, 30]. In addition to direct neuron dissection, other approaches analyze component functions by decomposing image repre", + "bbox": [ + 511, + 689, + 908, + 901 + ], + "page_idx": 1 + }, + { + "type": "page_number", + "text": "4344", + "bbox": [ + 482, + 944, + 514, + 955 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "sentations to reveal the role of attention heads within multimodal embedding spaces [15] or identifying neurons with similar functions across a diverse model zoo [11].", + "bbox": [ + 89, + 90, + 482, + 137 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "2.1. Preliminary: CVCL", + "text_level": 1, + "bbox": [ + 89, + 150, + 284, + 166 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Training data. CVCL is trained on the SAYCam-S dataset [37], a longitudinal collection of egocentric recordings from a child aged 6 to 25 months, containing around 200 hours of video. To create meaningful image-text pairs for model training, transcripts were pre-processed to retain only child-directed utterances, excluding the child's own vocalizations. Frames were extracted to align with utterance timestamps. The resulting dataset comprises 600,285 frames paired with 37,500 transcribed utterances, forming a multimodal dataset that simulates the sparse and noisy real-world experiences from which children learn.", + "bbox": [ + 89, + 172, + 483, + 339 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Model architecture. Employing a self-supervised contrastive learning approach akin to CLIP [34], CVCL learns to align egocentric visual frames with transcribed parental speech. Co-occurring pairs are treated as positive examples, while non-co-occurring pairs serve as negatives. This method allows the model to develop multimodal representations without external labels, imitating a child's natural learning process.", + "bbox": [ + 89, + 352, + 483, + 473 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Evaluation. For evaluation, CVCL adopts a $n$ -way classification task (Figure 2) in which the model selects the most relevant visual reference from a set comprising one target image and $n - 1$ foil images. This approach is inspired by the intermodal preferential looking paradigm (IPLP) [16, 17], used in infant recognition studies to measure language comprehension through differential visual fixation. By aligning its visual and text encoders, CVCL achieves comparable in-domain test accuracy to models like CLIP. However, CVCL demonstrates relatively weak test performance on the Konkle object dataset [23], which includes naturalistic object categories on a white background, using only classes available in the training data.", + "bbox": [ + 89, + 484, + 483, + 681 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3. Method", + "text_level": 1, + "bbox": [ + 89, + 700, + 181, + 715 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "In this section, we describe how to explore neuron-level concepts and leveraging them in $n$ way classification tasks. We begin by using published neuron labeling techniques to discover visual concepts hidden within CVCL [41]. Then, we introduce our framework to utilize these labeled neuron concepts for $n$ -way classification.", + "bbox": [ + 89, + 727, + 483, + 818 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3.1. Neuron Labeling", + "text_level": 1, + "bbox": [ + 89, + 830, + 259, + 848 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "We follow CLIP-Dissect [30] for internal representation analysis due to its flexibility in concept sets and input image dataset.", + "bbox": [ + 89, + 854, + 483, + 898 + ], + "page_idx": 2 + }, + { + "type": "image", + "img_path": "images/a87f51f0a332bfcb59757dca0a108bda47ef000039658f45f341b2c7b38877d5.jpg", + "image_caption": [ + "Figure 2. CVCL's $n$ -way evaluation [41] poses a classification task of choosing the given target object label (\"apple\") from $n = 4$ images, where only one of them contains the target object. The model feeds the target label into the text encoder and computes the pairwise similarities to each of the $n$ images, selecting the image with the highest similarity to demonstrate object recognition ability. However, this way of recognition limits its ability to the vocabularies in the text encoder. Our framework in Figure 4 overcomes this limitation." + ], + "image_footnote": [], + "bbox": [ + 519, + 89, + 903, + 224 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Preliminary: CILP-dissect. Given a neural network $f(x)$ , where $f$ takes an image $x$ as input and $x \\in \\mathcal{D}_{\\mathrm{probe}}$ with $|\\mathcal{D}_{\\mathrm{probe}}| = N$ , and a concept set $S$ with $|S| = M$ . The algorithm computes the concept-activation matrix $P \\in \\mathbb{R}^{N \\times M}$ .", + "bbox": [ + 511, + 387, + 905, + 448 + ], + "page_idx": 2 + }, + { + "type": "equation", + "text": "\n$$\nP _ {i, j} = I _ {i} \\cdot T _ {j}, \\tag {1}\n$$\n", + "text_format": "latex", + "bbox": [ + 660, + 462, + 903, + 479 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "where $I_{i}$ and $T_{j}$ are the embeddings of the images and concepts, respectively. For each neuron $k \\in K$ , where $K$ denotes the set of all neurons in the network, we summarize activations $A_{k}(x_{i})$ with a scalar function $g$ , producing an activation vector:", + "bbox": [ + 511, + 484, + 905, + 560 + ], + "page_idx": 2 + }, + { + "type": "equation", + "text": "\n$$\nq _ {k} = \\left[ g \\left(A _ {k} \\left(x _ {1}\\right)\\right), \\dots , g \\left(A _ {k} \\left(x _ {N}\\right)\\right) \\right] ^ {\\top} \\in \\mathbb {R} ^ {N}. \\tag {2}\n$$\n", + "text_format": "latex", + "bbox": [ + 560, + 571, + 903, + 590 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "The neuron is labeled with the concept that maximizes similarity:", + "bbox": [ + 511, + 601, + 903, + 631 + ], + "page_idx": 2 + }, + { + "type": "equation", + "text": "\n$$\nl _ {k} = \\arg \\max _ {m} \\operatorname {s i m} \\left(t _ {m}, q _ {k}; P\\right), \\tag {3}\n$$\n", + "text_format": "latex", + "bbox": [ + 604, + 631, + 903, + 652 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "where $\\operatorname{sim}(\\cdot, \\cdot)$ represents the similarity function (e.g. cosine similarity or Soft-WPMI [30, 42]), and $t_m$ denotes the most similar text concept. Collecting the label $l_k$ for each neuron, we define the label vector $\\mathcal{L} = [l_1, l_2, \\ldots, l_K]$ to represent the assigned concepts across the entire model.", + "bbox": [ + 511, + 659, + 905, + 734 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "During this neuron labeling process, we aim to assign meaningful concepts to each neuron. We use CLIP-dissect because:", + "bbox": [ + 511, + 734, + 905, + 777 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "- Concept set $S$ : Instead of allowing an infinite range of possible concepts for neuron labeling, using a fixed concept set narrows this process by constraining it to a limited selection of concepts. Including diverse concepts in this set enables us to identify neurons corresponding to these visual concepts. This setup ensures that each neuron is assigned a specific label, though some labels may be spurious, as illustrated in Figure 3.", + "bbox": [ + 511, + 779, + 906, + 900 + ], + "page_idx": 2 + }, + { + "type": "page_number", + "text": "4345", + "bbox": [ + 482, + 944, + 514, + 955 + ], + "page_idx": 2 + }, + { + "type": "image", + "img_path": "images/e4ac9913c0aff54801f05a019813b1b9dc17b7cc977dbd54941f57501925f334.jpg", + "image_caption": [ + "Figure 3. Spurious labeled neurons in infant model CVCL by CLIP-Dissect, showing top-3 activated Konkle dataset images. This indicates: (1) not all neurons interpretable with human semantic concepts [12]; (2) neuron-labeling can produce spurious labels. To further support our hypothesis, we implement NeuronClassifier." + ], + "image_footnote": [], + "bbox": [ + 96, + 90, + 475, + 213 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "- Probing dataset $\\mathcal{D}_{\\mathrm{probe}}$ : The probing dataset allows neurons to activate specifically in response to the dataset of interest. For instance, when analyzing representations from an infant model, we use images from the Konkle object dataset, which better aligns with infant recognition than the ImageNet [10] validation set. Additionally, for classification tasks, generalization can be achieved by adaptively selecting the probing dataset to focus on relevant concepts within a given dataset.", + "bbox": [ + 89, + 338, + 483, + 474 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Extending beyond vocabulary. The flexibility of the concept set $S$ allowing us to discover visual concepts that the infant model has never encountered in its training data. In this process, CLIP serves as a well-pretrained miner, utilizing its rich image-text embeddings to identify concepts hidden within CVCL. By aligning the activations of CVCL's neurons with CLIP's embeddings, we discover meaningful hidden visual concepts within the infant model, even extending beyond the model's linguistic training data. This approach leverages the diverse vocabulary of the concept set and the rich embeddings of CLIP to reveal visual concepts embedded in CVCL's internal representations, each associated with corresponding neurons.", + "bbox": [ + 89, + 481, + 483, + 679 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "3.2. Neuron-Based Classification", + "text_level": 1, + "bbox": [ + 89, + 689, + 344, + 704 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "How do we ensure that neurons representing concepts beyond the model's original vocabulary truly exist within the network? In this section, we propose NeuronClassifier, a training-free framework that leverages neuron activations to detect and validate such concepts. By discovering neurons with specific visual concept, we aim to confirm the presence of these latent, beyond-vocabulary neurons and use them to perform $n$ -way classification. The framework, illustrated in Figure 4, involves three main steps.", + "bbox": [ + 89, + 710, + 483, + 848 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Step 1: neuron labeling with concept set. Given an image encoder $f(x)$ , we labeled each neuron in the network using a concept set $S$ that contains (but is not limited to) all", + "bbox": [ + 89, + 854, + 483, + 900 + ], + "page_idx": 3 + }, + { + "type": "image", + "img_path": "images/933a9a5319d3c4839ebd5e09ddbd5d3496a5173a40856fc83309dc4b55bb7a9c.jpg", + "image_caption": [ + "Figure 4. Our NeuronClassifier Overview. A training-free $n$ -way framework with three key steps: (1) Label all neurons in the network using a concept set that includes class labels and common words; (2) identify neurons associated with the target concept (e.g., \"rug\"); (3) Evaluate the activations of visual concept neuron across $n$ images (in this example, $n = 4$ ) and select the image with the highest activation as the most relevant to the target label." + ], + "image_footnote": [], + "bbox": [ + 519, + 88, + 903, + 242 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "class labels relevant to the task. The dissection process can be expressed as a function:", + "bbox": [ + 511, + 376, + 906, + 406 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {N} _ {s} = \\text {N e u r o n L a b e l i n g} (f, \\mathcal {S}), \\tag {4}\n$$\n", + "text_format": "latex", + "bbox": [ + 604, + 420, + 903, + 436 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "where $\\mathcal{N}_s$ is the set of neurons labeled with concepts from $S$ . Each neuron $k \\in \\mathcal{N}_s$ is associated with a specific concept, such as \"rug\" or \"calculator\", based on its alignment with concept embeddings obtained during neuron labeling (e.g. similarity in CLIP-Dissect [30]).", + "bbox": [ + 511, + 449, + 905, + 526 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Step 2: identifying visual concept neurons. Given a target label $l_{k} \\in \\mathcal{L}$ , where $\\mathcal{L}$ represents all neuron labels in the model, same as label vector in Section 3.1 (e.g., \"rug\"), we select the subset of neurons labeled with this concept, denoted as $\\mathcal{N}_{l_k} \\subset \\mathcal{N}_{\\mathcal{L}}$ . These neurons are responsible for encoding the target concept.", + "bbox": [ + 511, + 534, + 905, + 625 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "To further refine the labeling and reduce spurious assignments, we select the most similar neuron from $\\mathcal{N}_{l_k}$ to represent the target concept. The similarity measure varies based on the dissection method used. For example, Network Dissection [2] employs Intersection over Union (IoU) for similarity, while CLIP-Dissect [30] supports multiple similarity metrics:", + "bbox": [ + 511, + 625, + 905, + 729 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\nk ^ {*} = \\arg \\max _ {k \\in \\mathcal {N} _ {\\mathbf {L}}} \\operatorname {s i m} \\left(t _ {l _ {k}}, q _ {k}; P\\right), \\tag {5}\n$$\n", + "text_format": "latex", + "bbox": [ + 601, + 744, + 903, + 767 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "where $t_{l_k}$ is the embedding of the target label $l_k$ , $q_k$ is the neuron activation value, same in Section 3.1. This step ensures that the neuron most aligned with the concept is selected, minimizing the possibility of spurious labeling.", + "bbox": [ + 511, + 780, + 905, + 840 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "For each selected neuron $k \\in \\mathcal{N}_{l_k}$ , its activation value on an input image $x_i$ is computed as", + "bbox": [ + 511, + 840, + 905, + 871 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\nq _ {k} \\left(x _ {i}\\right) = g \\left(A _ {k} \\left(x _ {i}\\right)\\right), \\tag {6}\n$$\n", + "text_format": "latex", + "bbox": [ + 635, + 885, + 903, + 901 + ], + "page_idx": 3 + }, + { + "type": "page_number", + "text": "4346", + "bbox": [ + 482, + 944, + 514, + 955 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "where $A_{k}(x_{i})$ is the raw activation map, and $g(\\cdot)$ is a summary function (e.g., spatial mean) that reduces it to a scalar representing the neuron's response strength.", + "bbox": [ + 89, + 90, + 482, + 136 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Step 3: selecting the most relevant image. Given $n$ candidate images $\\{x_1, x_2, \\ldots, x_n\\}$ in an $n$ -way trial, we compute the activation values $q_k(x_i)$ for neuron $k^*$ with highest similarity across all images. The image with the highest activation is selected as the most relevant to the target concept:", + "bbox": [ + 89, + 140, + 482, + 215 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\nx ^ {*} = \\arg \\max _ {x _ {i}} q _ {k ^ {*}} \\left(x _ {i}\\right). \\tag {7}\n$$\n", + "text_format": "latex", + "bbox": [ + 207, + 229, + 480, + 251 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "In the example shown in Figure 4, the target concept is \"rug\", and we select the image with the highest activation from the four candidates as the closest match to the concept.", + "bbox": [ + 89, + 256, + 482, + 301 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "4. Neuron-wise Representation Analysis", + "text_level": 1, + "bbox": [ + 89, + 313, + 428, + 330 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "In this section, we conduct a neuron-wise analysis of infant models and provide implementation details. First, we perform neuron labeling on the infant CVCL model. Then, we use our NeuronClassifier framework to leverage visual concept neurons identified through neuron labeling, resulting in better recognition performance than the original approach [41] while also discovering internal visual concepts beyond the model's training vocabulary. These results support our hypothesis.", + "bbox": [ + 89, + 338, + 483, + 474 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "4.1. Setup", + "text_level": 1, + "bbox": [ + 89, + 482, + 171, + 498 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Datasets. We use the Konkle object dataset [23], as introduced in Section 2.1. This dataset consists of 3,406 images, each featuring a single object on a clean white background, including 406 test items across 200 classes. Each trial comprises $n$ images: one target image and the remaining as foils, with foil images randomly sampled from classes other than the target class. For each class, we generate 5 trials, each containing $n$ images. Following the previous work on CVCL, we use $n = 4$ in our main experiments, with one target and three foils per trial.", + "bbox": [ + 89, + 503, + 483, + 655 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Neuron labeling. We utilize CLIP-Dissect [30] for neuron labeling, which assigns visual concepts to each neuron in the network. As we aim to perform classification on the Konkle object dataset, we use the same dataset as a part of $\\mathcal{D}_{\\mathrm{probe}}$ . Additionally, to avoid limiting the search space only around class names and ensure comprehensive neuron labeling, we employ a combined concept, consisting of three components:", + "bbox": [ + 89, + 659, + 482, + 779 + ], + "page_idx": 4 + }, + { + "type": "list", + "sub_type": "text", + "list_items": [ + "- SAYCam-S vocabulary: We clean the original vocabulary by removing noisy child speech and retaining meaningful words.", + "- Common English words: We chose the top 30,000 most common English words based on a 1-gram frequency analysis by Peter Norvig [29, 30].", + "- Class in Konkle object dataset: All class labels from the Konkle object dataset are included." + ], + "bbox": [ + 89, + 780, + 482, + 900 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "We combine these three sources, ensuring no duplicates, resulting in a final concept set containing 30,427 words.", + "bbox": [ + 511, + 90, + 903, + 121 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Models. Our primary focus is the CVCL-ResNeXt50 [41], trained on SAY-Cam-S [37] dataset. For comprehensive analysis, we apply our framework to the following models:", + "bbox": [ + 511, + 126, + 903, + 186 + ], + "page_idx": 4 + }, + { + "type": "list", + "sub_type": "text", + "list_items": [ + "- Infant models: CVCL is trained on unique infant data, using both egocentric frames and transcribed parent speech. We also take DINO-S-ResNeXt50 [31] as reference model compare recognition ability derived from same visual experience. It trained with DINO[6] self-supervised approach on the infant same dataset.", + "- Broadly-trained models: To establish an upper bound, we include CLIP [34] and ResNeXt50 [43], which are broadly trained on large scale Internet images. Although CLIP uses a ResNet50-based vision encoder rather than ResNeXt, we select CLIP-ResNet50 due to the architectural similarity between ResNeXt [43] and ResNet [20].", + "- Randomized model: As a lower bound, we introduce a randomized version of the CVCL-ResNeXt50 model. In this setup, the convolution layer weights in the vision encoder are initialized using Kaiming Initialization [19]." + ], + "bbox": [ + 513, + 189, + 903, + 431 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "4.2. Results", + "text_level": 1, + "bbox": [ + 513, + 441, + 607, + 455 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "In this section, we present our results and findings from applying neuron labeling and our NeuronClassifier framework to the infant CVCL model, compared with other reference models. We define two class types for our analysis:", + "bbox": [ + 511, + 463, + 903, + 523 + ], + "page_idx": 4 + }, + { + "type": "list", + "sub_type": "text", + "list_items": [ + "- In-vocabulary classes: Object classes present in the model's training linguistic input, which also appear in test object class and are detectable in internal representations.", + "- Out-of-vocabulary classes: Object classes not included in the training linguistic input but detected in test object class and internal representations." + ], + "bbox": [ + 513, + 527, + 903, + 617 + ], + "page_idx": 4 + }, + { + "type": "image", + "img_path": "images/425151cb957ec1ad32b744485555faa6b42d59f1096e4fb78e3be27b764e78e0.jpg", + "image_caption": [ + "Figure 5. In-vocabulary and out-of-vocabulary relationships visualized using a Venn diagram." + ], + "image_footnote": [], + "bbox": [ + 583, + 633, + 839, + 708 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Our results demonstrate that the proposed framework effectively discovers meaningful neurons that represent concepts beyond the model's training vocabulary. These findings align with the cognitive perspective of vocabulary acquisition in infant development. We analyze the models' performance across different $n$ -way classification settings, showing that our method yields strong out-of-vocabulary classification performance while simultaneously improving in-vocabulary classification accuracy.", + "bbox": [ + 511, + 763, + 906, + 901 + ], + "page_idx": 4 + }, + { + "type": "page_number", + "text": "4347", + "bbox": [ + 480, + 944, + 514, + 955 + ], + "page_idx": 4 + }, + { + "type": "image", + "img_path": "images/ffe366a0f8ed821e4fce48429054d773f759c535135c1bfd371e6e2db9f5647d.jpg", + "image_caption": [ + "Figure 6. Class coverage in visual concept neurons. Percentage of Konkle object dataset [23] classes identified through neuron labeling. Broadly pre-trained models (CLIP, ResNeXt) achieve over $50\\%$ coverage, while developmentally inspired infant models (CVCL [41], DINO [6]) show lower coverage. CVCL-Randomized provides a lower-bound comparison." + ], + "image_footnote": [], + "bbox": [ + 98, + 88, + 472, + 244 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Class coverage in visual concept neurons. How well do visual concept neurons identified through neuron labeling correspond to specific class names (e.g., \"rug\") rather than general descriptive attributes (e.g., \"red\"? Figure 6 shows the percentage of classes in the Konkle object dataset that are discovered in visual concept neurons from each model during the neuron labeling process. The results indicate that well-pretrained models, such as CLIP-ResNet50 and ResNeXt50, demonstrate broader class name coverage. While CVCL performs slightly weaker, it still maintains coverage of slightly less than $50\\%$ . In contrast, CVCL-Randomized achieves only around $28\\%$ coverage. This class coverage metric reflects the models' capacity to form class-corresponding meaningful representations during the neuron labeling process.", + "bbox": [ + 88, + 366, + 482, + 592 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Age of Acquisition (AoA) ratings. Age of acquisition (AoA) is used to indicate when, and in what sequence, words are learned, and it is often assessed through ratings or observations reported by adults. This indirect method generally correlates well with other metrics indicating when children acquire vocabulary. Previous developmental work has shown that infants' early visual familiarity with common objects helps with object recognition, which subsequently helps support the process of learning the names of those objects. We next examined how early words in our models are learned and whether there is an AoA difference between in-vocab and out-of-vocab words.", + "bbox": [ + 89, + 598, + 482, + 777 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "We used AoA ratings from a dataset compiled by Kuperman, Stadthagen-Gonzalez, and Brysbaert [25], which includes norms for over 30,000 English words gathered via Amazon Mechanical Turk. Each participant estimated the age in years at which they believed they first understood each word, even if they did not actively use them. This dataset is comparable to previously reported AoA norms [36] gathered in laboratory settings.", + "bbox": [ + 89, + 779, + 482, + 901 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/5bf13d735d09342d50dc52a7e8341121f4445cd0f761898d1ff137794446a05c.jpg", + "image_caption": [ + "Figure 7. Age of acquisition (AoA) ratings for in- and out-of-vocabulary visual concepts. Comparison of word acquisition timing [25] between in-vocabulary and out-of-vocabulary concepts. Out-of-vocabulary concepts tend to have a higher estimated acquisition ages in the infant-inspired CVCL model, indicating development of visual understanding beyond explicit linguist inputs." + ], + "image_footnote": [], + "bbox": [ + 516, + 89, + 885, + 263 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Using this set of AoA norms, we compared mean AoA between in-vocab and out-of-vocab words discovered in CVCL's internal representations. As shown in Figure 7, we found a significant difference between in-vocab and out-of-vocab AoA rating ( $t(82) = 4.64$ , $p < 0.0001$ ), in-vocab words (mean AoA = 4.99) are learned earlier than out-of-vocab words (mean AoA = 6.82). This pattern suggests that: (1) both sets of words are learned quite early, around later preschool and school years, with or without supervised labeling; (2) the difference in AoA between in-vocabulary and out-of-vocabulary words indicates that the infant model has developed a basic visual understanding of concepts with higher AoA. This foundational knowledge may lay the groundwork for word learning once corresponding parental speech is introduced.", + "bbox": [ + 511, + 388, + 906, + 616 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Neuron-based classification performance. We evaluated the hidden potential of infant models' vision encoders by applying our NeuronClassifier framework, summarized in Table 1. Despite being trained on infant egocentric data with limited amount and diversity, CVCL demonstrates the ability to recognize similarly as nature of infant learning, revealing strong out-of-vocabulary classification performance. This result suggests that this infant model developed broader visual concepts that extend beyond linguistic input, similar to how infants naturally learn. We also applied our method to in-vocabulary classification, where it outperformed the vanilla method previously used in CVCL [41], as introduced in Figure 2. \"All\" representing the combined performance on both \"in-vocab\" and \"out-of-vocab\". These results support our hypothesis.", + "bbox": [ + 511, + 627, + 908, + 853 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "We include additional model comparisons: (1) The DINO-S-ResNeXt50 infant model [31], trained on the same dataset without text supervision, achieves comparable per", + "bbox": [ + 511, + 854, + 906, + 902 + ], + "page_idx": 5 + }, + { + "type": "page_number", + "text": "4348", + "bbox": [ + 482, + 944, + 514, + 955 + ], + "page_idx": 5 + }, + { + "type": "table", + "img_path": "images/f284228f2cb39339ff83cec0e68f61d9d0ead3a72b09c012f455c3afb6b24433.jpg", + "table_caption": [ + "Table 1. Neuron-based classification results in 4-way evaluation among models in Section 4.1. \"Vanilla\" refers to classification based on image-text similarity (Figure 2). \"X\" denotes cases where direct classification on the Konkle dataset [23] is not possible due to missing text encoder or need fine-tuning. By leveraging neurons discovered in the representation, NeuronClassifier enables broader recognition, particularly in CVCL (bolded for emphasis), achieving improved recognition in both in-vocabulary and out-of-vocabulary, supporting our hypothesis. \"All\" represents the combined performance on both in- and out-of-vocabulary." + ], + "table_footnote": [], + "table_body": "
MethodModelIn-vocabOut-of-vocabAll
VanillaCLIP-ResNet5098.81±0.1696.93±0.0697.42±0.05
ResNeXt50×××
CVCL-ResNeXt5036.18±0.91××
DINO-S-ResNeXt50×××
Neuron ClassifierCLIP-ResNet5091.59±0.5288.66±0.3589.79±0.38
ResNeXt5088.17±0.4593.28±0.3691.88±0.15
CVCL-ResNeXt5079.50±0.7876.81±0.3577.79±0.40
DINO-S-ResNeXt5077.53±0.2477.96±0.2777.65±0.21
", + "bbox": [ + 91, + 238, + 483, + 366 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "formance in visual representations. This suggests that models trained on identical data distributions with different self-supervised methods may yield similar representational outcomes. (2) Broadly-trained models such as ResNeXt50 and CLIP establish performance upper bounds. However, CLIP underperforms in the vanilla setting, relying exclusively on visual neurons without text encoder guidance. While this work does not aim to advance zero-shot learning methods, it reveals visual concepts in infant models that emerge independently of linguistic inputs.", + "bbox": [ + 88, + 392, + 482, + 542 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "In classification on the Konkle object dataset, both DINO-S and ImageNet ResNeXt50 required fine-tuning for this task. Our framework enables neuron-based classification without downstream fine-tuning, and providing a training-free qualitative inspection of internal representations.", + "bbox": [ + 89, + 544, + 483, + 633 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Analysis across $n$ -way settings. We evaluate the models under various $n$ -way classification setups. Figure 8 illustrates the performance trends for in- and out-of-vocabulary class classification accuracy applying our NeuronClassifier.", + "bbox": [ + 89, + 643, + 482, + 703 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Our method not only enables out-of-vocabulary classification but also significantly improves in-vocab performance compared to previous results [41], further supporting the presence of beyond-vocabulary potential in infant models. However, due to limited class coverage, CVCL with our method can classify a maximum of 31 classes (see Appendix A.3). These findings support our hypothesis that the infant model has acquired visual concepts beyond its initial vocabulary. These results show that leveraging the model's internal representations for classification that go beyond its vocabulary is not only feasible but also robust across different $n$ -way settings. These findings support our hypothesis that the infant model has acquired visual concepts beyond", + "bbox": [ + 89, + 704, + 482, + 901 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/42c71236126c2ca2c87f41d16d2a98c32bb84d27134c7a9a5a5694e72ecaa499.jpg", + "image_caption": [ + "Figure 8. In- and out-of-vocabulary class performance across $n$ -way settings using NeuronClassifier. (a) For out-of-vocabulary classification (left), our method enables classification without additional training. The infant model CVCL ( $\\clubsuit$ ) maintains robust performance as $n$ increases. (b) For in-vocabulary classification (right), \"ours\" (NeuronClassifier) enhances CVCL recognition ability compared to the \"vanilla\" setting. However, CLIP ( $\\clubsuit$ ) declines, as it relies solely on neurons without text encoder input. Random ones serve as lower bounds." + ], + "image_footnote": [], + "bbox": [ + 516, + 90, + 712, + 311 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/34575b3d787034d64dcf712c3f7a84750558d28b1dd93679d3db2b0f4d1e2944.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 720, + 90, + 880, + 311 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "its initial vocabulary.", + "bbox": [ + 511, + 474, + 653, + 489 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "However, as $n$ increases exponentially, performance gradually declines. This decline is reasonable, as the task becomes increasingly difficult by nature as $n$ increases. It may also be attributed to dimensionality reduction in activation maps, leading to coarser classification. CLIPRN50 and ResNeXt50 perform well with NeuronClassifier, though not as effectively as their direct or fine-tuned versions, since our method is designed to reveal latent concepts rather than to perform fully optimized classification. In in-vocab settings, the \"vanilla\" approach represents direct classification as shown in Figure 2.", + "bbox": [ + 511, + 491, + 906, + 657 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "5. Layer-wise Representation Analysis", + "text_level": 1, + "bbox": [ + 511, + 672, + 836, + 690 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "How does the representation learned from infant data differ from other representations widely used in the computer vision community? To explore this, we perform a layer-wise representation analysis using Centered Kernel Alignment (CKA) [24]. We compute the similarity between the representations of the infant model(CVCL), ImageNet-pretrained, and CLIP. Additionally, we apply neuron labeling techniques from a layer-wise perspective to identify unique visual concepts discovered at each layer between ImageNet and infant models.", + "bbox": [ + 511, + 696, + 906, + 849 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Layer similarity analysis. CKA [24] applies HSIC [18] (see Appendix B) over a set of images to provide layer-wise similarity scores between the representations of different", + "bbox": [ + 511, + 854, + 908, + 902 + ], + "page_idx": 6 + }, + { + "type": "page_number", + "text": "4349", + "bbox": [ + 482, + 944, + 514, + 955 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/9ad3b396e855efda1aed02449395d95d28c514c860fa7bfcb0ceb0e97b2bc720.jpg", + "image_caption": [ + "Figure 9. CKA layer-wise similarity between CVCL and common models. Using the ImageNet validation set as input, CVCL (y-axis) exhibits similarity to CLIP-RN50 (x-axis, left) and ResNeXt50 (x-axis, right) in the shallow layers (lower-level features) but diverges significantly in the final layer (higher-level features). Notably, Layer 4 of CVCL shows very low similarity to all layers of both common models." + ], + "image_footnote": [], + "bbox": [ + 91, + 90, + 480, + 234 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "neural networks. A higher CKA score indicates more similar representations between two models at the given layer. We use the ImageNet validation set [10] as input for the networks, and compute the CKA similarity between the infant model (CVCL), the ImageNet-pretrained ResNeXt50, and CLIP-ResNet50. The results are presented as matrices in Figure 9. The lower layers of CVCL exhibit greater similarity to larger-scale models than its deeper layers. In larger-scale models, shallow layers are known to capture low-complexity features, while deeper layers progressively specialize in capturing higher-level concepts [7, 14]. Therefore, CVCL – the model trained on infant data – successfully develops lower-level representations comparable to those in common pre-trained models. However, the divergence in deeper layers suggests a lack of the diverse higher-level representations typically observed in models trained on common datasets.", + "bbox": [ + 88, + 372, + 482, + 628 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Neuron-based analysis. To investigate the characteristics of each layer, we apply network dissection to identify neurons that are aligned with specific visual concepts. Essentially, this is an extension of the neuron labeling process, where the Broden [2] dataset provides category labels for each visual concept neuron. We count the number of unique visual concepts discovered in each category and perform layer-wise comparisons to gain a broader view of the differences between models trained from ImageNet and infant data. In Figure 10, we visualize the number of unique visual concept neurons across layers for each model. The results show that early layers in both models predominantly have neurons of low-level features like color, with minimal differences between models. However, as we move to deeper layers, higher-level concepts such as objects and scenes become more prominent, and the disparities between models become clearer. CVCL exhibits fewer unique visual con", + "bbox": [ + 89, + 643, + 482, + 900 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/1d9df6f63c2f7251555648f664f46debbcd23323a72a3a93e1573922be641f1c.jpg", + "image_caption": [ + "Figure 10. Number of visual concepts in ImageNet and infant models. Concepts are categorized using Broden dataset [2]. Neurons in deeper model layers capture increasingly complex concepts. Early layers primarily detect lower-level features like color and texture, while higher-level concepts such as objects and scenes emerge in deeper layers. CVCL exhibits fewer visual concepts than the ImageNet model, especially for higher-level visual concepts (e.g., objects and scenes)." + ], + "image_footnote": [], + "bbox": [ + 513, + 90, + 893, + 244 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "cepts in these higher-level categories compared to ImageNet model. This finding aligns with the layer similarity analysis.", + "bbox": [ + 511, + 393, + 906, + 426 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "6. Conclusion", + "text_level": 1, + "bbox": [ + 511, + 438, + 633, + 454 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "In this paper, we explored whether an infant model (CVCL), trained on infant egocentric video frames and linguistic inputs, can acquire broader visual concepts extending beyond its initial training vocabulary. By introducing NeuronClassifier, a training-free framework to discover and leverage visual concepts hidden in representations, we unlocked the CVCL visual encoder's ability to recognize out-of-vocabulary concepts, establishing its potential as a strong classifier. Our findings also reveal that while CVCL, trained on a unique infant dataset with limited exposure to diverse scenes, it representations capture low-level features similar to those in common pre-trained models, they diverge significantly in higher-level representations, contributing to the observed performance differences.", + "bbox": [ + 509, + 463, + 906, + 672 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Overall, our approach bridges cognitive science and computer vision, providing insights into how infant models develop visual concepts that precede linguistic inputs, aligning with the natural way infants explore the world through sight.", + "bbox": [ + 511, + 674, + 906, + 751 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Limitations and future work. Our study did not analyze the infant training data, as we were unable to access it due to strict access controls. Instead, we analyze models trained on infant egocentric data, finding developmental alignment with cognitive studies. However, we did not extend this analysis to adult egocentric data. While the study focuses on infant data for developmental process, the framework can be applied to adult models, which remains a direction for future research.", + "bbox": [ + 511, + 753, + 908, + 888 + ], + "page_idx": 7 + }, + { + "type": "page_number", + "text": "4350", + "bbox": [ + 482, + 944, + 514, + 955 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Acknowledgements", + "text_level": 1, + "bbox": [ + 91, + 90, + 258, + 107 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "This work was partially supported by the National Research Foundation Singapore Competitive Research Program (award number CRP29-2022-0003). We thank Wai Keen Vong for invaluable discussion and CVCL's pretrained weights. We thank Jingyi Lin for figure discussions. We thank the anonymous reviewers for their constructive feedback.", + "bbox": [ + 89, + 114, + 483, + 219 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "References", + "text_level": 1, + "bbox": [ + 91, + 234, + 187, + 251 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[1] Sven Bambach, David Crandall, Linda Smith, and Chen Yu. Toddler-inspired visual object learning. In Neural Information Processing Systems, 2018. 1, 2", + "[2] David Bau, Bolei Zhou, Aditya Khosla, Aude Oliva, and Antonio Torralba. Network dissection: Quantifying interpretability of deep visual representations. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 6541-6549, 2017. 2, 4, 8", + "[3] Erika Bergelson and Daniel Swingley. At 6-9 months, human infants know the meanings of many common nouns. Proceedings of the National Academy of Sciences, 109(9): 3253-3258, 2012. 2", + "[4] Marc Brysbaert and Boris New. Moving beyond kucera and francis: A critical evaluation of current word frequency norms and the introduction of a new and improved word frequency measure for american english. Behavior Research Methods, 41(4):977-990, 2009. 1", + "[5] Kirill Bykov, Laura Kopf, Shinichi Nakajima, Marius Kloft, and Marina Höhne. Labeling neural representations with inverse recognition. Advances in Neural Information Processing Systems, 36, 2024. 2", + "[6] Mathilde Caron, Hugo Touvron, Ishan Misra, Hervé Jégou, Julien Mairal, Piotr Bojanowski, and Armand Joulin. Emerging properties in self-supervised vision transformers. In Proceedings of the IEEE/CVF international conference on computer vision, pages 9650-9660, 2021. 5, 6", + "[7] Yixiong Chen, Alan Yuille, and Zongwei Zhou. Which layer is learning faster? a systematic exploration of layerwise convergence rate for deep neural networks. In The Eleventh International Conference on Learning Representations, 2023. 8", + "[8] Elizabeth M Clerkin, Elizabeth Hart, James M Rehg, Chen Yu, and Linda B Smith. Real-world visual statistics and infants' first-learned object names. Philosophical Transactions of the Royal Society B: Biological Sciences, 372(1711): 20160055, 2017. 2", + "[9] Michael J Cortese and Maya M Khanna. Age of acquisition ratings for 3,000 monosyllabic words. Behavior Research Methods, 40(3):791-794, 2008. 1", + "[10] Jia Deng, Wei Dong, Richard Socher, Li-Jia Li, Kai Li, and Li Fei-Fei. Imagenet: A large-scale hierarchical image database. In 2009 IEEE conference on computer vision and pattern recognition, pages 248-255. IEEE, 2009. 4, 8", + "[11] Amir Dravid, Yossi Gandelsman, Alexei A Efros, and Assaf Shocher. Rosetta neurons: Mining the common units in a" + ], + "bbox": [ + 93, + 260, + 483, + 900 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "model zoo. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 1934-1943, 2023. 3", + "[12] Nelson Elhage, Tristan Hume, Catherine Olsson, Nicholas Schiefer, Tom Henighan, Shauna Kravec, Zac Hatfield-Dodds, Robert Lasenby, Dawn Drain, Carol Chen, et al. Toy models of superposition. arXiv preprint arXiv:2209.10652, 2022.4", + "[13] Robert L Fantz. Pattern vision in newborn infants. Science, 140(3564):296-297, 1963. 2", + "[14] Thomas Fel, Louis Bethune, Andrew Kyle Lampinen, Thomas Serre, and Katherine Hermann. Understanding visual feature reliance through the lens of complexity. arXiv preprint arXiv:2407.06076, 2024. 8", + "[15] Yossi Gandelsman, Alexei A Efros, and Jacob Steinhardt. Interpreting clip's image representation via text-based decomposition. arXiv preprint arXiv:2310.05916, 2023. 3", + "[16] Roberta Michnick Golinkoff, Kathryn Hirsh-Pasek, Kathleen M Cauley, and Laura Gordon. The eyes have it: Lexical and syntactic comprehension in a new paradigm. Journal of child language, 14(1):23-45, 1987. 3", + "[17] Roberta Michnick Golinkoff, Weiyi Ma, Lulu Song, and Kathy Hirsh-Pasek. Twenty-five years using the intermodal preferential looking paradigm to study language acquisition: What have we learned? Perspectives on Psychological Science, 8(3):316-339, 2013. 2, 3", + "[18] Arthur Gretton, Ralf Herbrich, Alexander Smola, Olivier Bousquet, Bernhard Schölkopf, and Aapo Hyvarinen. Kernel methods for measuring independence. Journal of Machine Learning Research, 6(12), 2005. 7, 3", + "[19] Kaiming He, Xiangyu Zhang, Shaoqing Ren, and Jian Sun. Delving deep into rectifiers: Surpassing human-level performance on imagenet classification. In Proceedings of the IEEE international conference on computer vision, pages 1026-1034, 2015. 5", + "[20] Kaiming He, Xiangyu Zhang, Shaoqing Ren, and Jian Sun. Deep residual learning for image recognition. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 770-778, 2016. 5", + "[21] Evan Hernandez, Sarah Schwettmann, David Bau, Teona Bagashvili, Antonio Torralba, and Jacob Andreas. Natural language descriptions of deep visual features. In International Conference on Learning Representations, 2021. 2", + "[22] Neha Kalibhat, Shweta Bhardwaj, C Bayan Bruss, Hamed Firooz, Maziar Sanjabi, and Soheil Feizi. Identifying interpretable subspaces in image representations. In International Conference on Machine Learning, pages 15623-15638. PMLR, 2023. 2", + "[23] Talia Konkle, Timothy F Brady, George A Alvarez, and Aude Oliva. Conceptual distinctiveness supports detailed visual long-term memory for real-world objects. Journal of experimental Psychology: general, 139(3):558, 2010. 3, 5, 6, 7, 1, 2", + "[24] Simon Kornblith, Mohammad Norouzi, Honglak Lee, and Geoffrey Hinton. Similarity of neural network representations revisited. In International conference on machine learning, pages 3519-3529. PMLR, 2019. 7" + ], + "bbox": [ + 516, + 92, + 906, + 900 + ], + "page_idx": 8 + }, + { + "type": "page_number", + "text": "4351", + "bbox": [ + 482, + 944, + 513, + 955 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[25] Victor Kuperman, Hans Stadthagen-Gonzalez, and Marc Brysbaert. Age-of-acquisition ratings for 30,000 english words. Behavior research methods, 44:978-990, 2012. 6", + "[26] Victor Kuperman, Hans Stadthagen-Gonzalez, and Marc Brysbaert. Age-of-acquisition ratings for 30,000 english words. Behavior Research Methods, 44(4):978-990, 2012. 2, 1", + "[27] Jean M Mandler. How to build a baby: II. conceptual primitives. Psychological review, 99(4):587, 1992. 2", + "[28] Jesse Mu and Jacob Andreas. Compositional explanations of neurons. Advances in Neural Information Processing Systems, 33:17153-17163, 2020. 2", + "[29] Peter Norvig. Natural language corpus data: Beautiful data, 2009. Accessed: 2024-10-27. 5", + "[30] Tuomas Oikarinen and Tsui-Wei Weng. Clip-dissect: Automatic description of neuron representations in deep vision networks. arXiv preprint arXiv:2204.10965, 2022. 2, 3, 4, 5", + "[31] A Emin Orhan and Brenden M Lake. Learning high-level visual representations from a child's perspective without strong inductive biases. Nature Machine Intelligence, 6(3):271-283, 2024. 1, 2, 5, 6", + "[32] Emin Orhan, Vaibhav Gupta, and Brenden M Lake. Self-supervised learning through the eyes of a child. Advances in Neural Information Processing Systems, 33:9960-9971, 2020. 1, 2", + "[33] Deborah A Phillips and Jack P Shonkoff. From neurons to neighborhoods: The science of early childhood development. 2000. 2", + "[34] Alec Radford, Jong Wook Kim, Chris Hallacy, Aditya Ramesh, Gabriel Goh, Sandhini Agarwal, Girish Sastry, Amanda Askell, Pamela Mishkin, Jack Clark, et al. Learning transferable visual models from natural language supervision. In International conference on machine learning, pages 8748-8763. PMLR, 2021. 2, 3, 5", + "[35] Saber Sheybani, Himanshu Hansaria, Justin Wood, Linda Smith, and Zoran Tiganj. Curriculum learning with infant egocentric videos. Advances in Neural Information Processing Systems, 36, 2024. 1, 2", + "[36] Hans Stadthagen-Gonzalez and Colin J Davis. The bristol norms for age of acquisition, imageability, and familiarity. Behavior Research Methods, 38(4):598-605, 2006. 6, 1", + "[37] Jessica Sullivan, Michelle Mei, Andrew Perfors, Erica Wojcik, and Michael C Frank. Saycam: A large, longitudinal audiovisual dataset recorded from the infant's perspective. Open mind, 5:20-29, 2021. 3, 5", + "[38] Gentaro Taga, Kayo Asakawa, Atsushi Maki, Yukuo Konishi, and Hideaki Koizumi. Brain imaging in awake infants by near-infrared optical topography. Proceedings of the National Academy of Sciences, 100(19):10722-10727, 2003. 2", + "[39] Satoshi Tsutsui, Arjun Chandrasekaran, Md Alimoor Reza, David Crandall, and Chen Yu. A computational model of early word learning from the infant's point of view. CogSci, 2020. 2", + "[40] Alan M Turing. Computing machinery and intelligence. Mind, LIX(236):433-460, 1950. 2", + "[41] Wai Keen Vong, Wentao Wang, A Emin Orhan, and Brenden M Lake. Grounded language acquisition through the" + ], + "bbox": [ + 91, + 92, + 480, + 900 + ], + "page_idx": 9 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "eyes and ears of a single child. Science, 383(6682):504-511, 2024. 1, 2, 3, 5, 6, 7", + "[42] Zeyu Wang, Berthy Feng, Karthik Narasimhan, and Olga Russakovsky. Towards unique and informative captioning of images. In Computer Vision-ECCV 2020: 16th European Conference, Glasgow, UK, August 23-28, 2020, Proceedings, Part VII 16, pages 629-644. Springer, 2020. 3", + "[43] Saining Xie, Ross Girshick, Piotr Dólár, Zhuowen Tu, and Kaiming He. Aggregated residual transformations for deep neural networks. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 1492-1500, 2017. 5" + ], + "bbox": [ + 516, + 92, + 903, + 258 + ], + "page_idx": 9 + }, + { + "type": "page_number", + "text": "4352", + "bbox": [ + 482, + 944, + 514, + 955 + ], + "page_idx": 9 + } +] \ No newline at end of file diff --git a/2025/Discovering Hidden Visual Concepts Beyond Linguistic Input in Infant Learning/50337d5c-a1b4-4b7f-9096-7ccfa12467ab_model.json b/2025/Discovering Hidden Visual Concepts Beyond Linguistic Input in Infant Learning/50337d5c-a1b4-4b7f-9096-7ccfa12467ab_model.json new file mode 100644 index 0000000000000000000000000000000000000000..2af90f3cfcd9f84a424e04b0f7188fa1c89579d0 --- /dev/null +++ b/2025/Discovering Hidden Visual Concepts Beyond Linguistic Input in Infant Learning/50337d5c-a1b4-4b7f-9096-7ccfa12467ab_model.json @@ -0,0 +1,2167 @@ +[ + [ + { + "type": "header", + "bbox": [ + 0.107, + 0.003, + 0.182, + 0.044 + ], + "angle": 0, + "content": "CVF" + }, + { + "type": "header", + "bbox": [ + 0.237, + 0.001, + 0.812, + 0.047 + ], + "angle": 0, + "content": "This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore." + }, + { + "type": "title", + "bbox": [ + 0.093, + 0.131, + 0.907, + 0.155 + ], + "angle": 0, + "content": "Discovering Hidden Visual Concepts Beyond Linguistic Input in Infant Learning" + }, + { + "type": "text", + "bbox": [ + 0.236, + 0.179, + 0.323, + 0.2 + ], + "angle": 0, + "content": "Xueyi Ke" + }, + { + "type": "text", + "bbox": [ + 0.354, + 0.181, + 0.484, + 0.199 + ], + "angle": 0, + "content": "Satoshi Tsutsui" + }, + { + "type": "text", + "bbox": [ + 0.516, + 0.181, + 0.632, + 0.199 + ], + "angle": 0, + "content": "Yayun Zhang" + }, + { + "type": "text", + "bbox": [ + 0.663, + 0.181, + 0.766, + 0.198 + ], + "angle": 0, + "content": "Bihan Wen\\(^{1*}\\)" + }, + { + "type": "text", + "bbox": [ + 0.358, + 0.206, + 0.641, + 0.224 + ], + "angle": 0, + "content": "\\(^{1}\\)Nanyang Technological University" + }, + { + "type": "text", + "bbox": [ + 0.308, + 0.225, + 0.686, + 0.242 + ], + "angle": 0, + "content": "2The Max Planck Institute for Psycholinguistics" + }, + { + "type": "image", + "bbox": [ + 0.116, + 0.3, + 0.492, + 0.522 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.501, + 0.3, + 0.887, + 0.522 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.09, + 0.522, + 0.907, + 0.596 + ], + "angle": 0, + "content": "Figure 1. Inspired by the development of the infant visual system, CVCL [41] is trained on infant egocentric frames and transcribed parental speech (a) and demonstrates object recognition ability within the vocabulary provided by parental speech (c). However, infant visual development is not limited to parental guidance. Thus, we hypothesize that a computational model trained on an infant's daily experiences can similarly acquire visual concepts beyond its training parental speech. To explore this, we perform neuron labeling (b) to identify visual concept neurons, including concepts that were never mentioned in the parental speech vocabulary (e.g., \"rug\"). Based on discovered neurons, we show that the model can recognize objects beyond the training vocabulary, akin to early infant visual development (d). Some images are cited from prior work [41] as the original data is not accessible to us." + }, + { + "type": "title", + "bbox": [ + 0.248, + 0.614, + 0.327, + 0.63 + ], + "angle": 0, + "content": "Abstract" + }, + { + "type": "text", + "bbox": [ + 0.089, + 0.648, + 0.486, + 0.875 + ], + "angle": 0, + "content": "Infants develop complex visual understanding rapidly, even preceding of the acquisition of linguistic skills. As computer vision seeks to replicate the human vision system, understanding infant visual development may offer valuable insights. In this paper, we present an interdisciplinary study exploring this question: can a computational model that imitates the infant learning process develop broader visual concepts that extend beyond the vocabulary it has heard, similar to how infants naturally learn? To investigate this, we analyze a recently published model in Science by Vong et al., which is trained on longitudinal, egocentric images of a single child paired with transcribed parental speech. We perform neuron labeling to identify visual concept neurons hidden in the model's internal representations. We then demonstrate that these neurons can" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.616, + 0.909, + 0.753 + ], + "angle": 0, + "content": "recognize objects beyond the model's original vocabulary. Furthermore, we compare the differences in representation between infant models and those in modern computer vision models, such as CLIP and ImageNet pre-trained model. Ultimately, our work bridges cognitive science and computer vision by analyzing the internal representations of a computational model trained on an infant visual and linguistic inputs. Our code is available at https://github.com/Kexueyi/discover_infant.vis." + }, + { + "type": "title", + "bbox": [ + 0.514, + 0.785, + 0.646, + 0.801 + ], + "angle": 0, + "content": "1. Introduction" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.812, + 0.909, + 0.906 + ], + "angle": 0, + "content": "Infants are remarkable learners, sparking interest across various academic disciplines. Computer vision is no exception, with researchers studying infant visual learning from various perspectives [1, 31, 32, 35, 41]. A recent milestone is Child's View for Contrastive Learning (CVCL) by Vong et al.[41], which trains a model from scratch on longitudinal" + }, + { + "type": "page_footnote", + "bbox": [ + 0.109, + 0.889, + 0.241, + 0.902 + ], + "angle": 0, + "content": "*Corresponding Author." + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.516, + 0.957 + ], + "angle": 0, + "content": "4343" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.09, + 0.092, + 0.482, + 0.242 + ], + "angle": 0, + "content": "egocentric videos of a single infant (6-25 months) paired with transcribed parental speech to learn visual-linguistic associations, similar to CLIP [34]. The resulting model develops object recognition abilities, aligning with developmental psychology findings that infants acquire object names by linking words to visual referents [3, 17]. In this paper, we analyze the internal visual representations of CVCL to better understand its mechanism, similar to how developmental studies observe infants' internal neurons [38]." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.245, + 0.483, + 0.456 + ], + "angle": 0, + "content": "In developmental psychology, research on real infants suggests that their object recognition is deeply influenced by their visual experiences with the world. A headcam study of 8.5 to 10.5-month-old infants [8] showed that the first nouns infants acquire often correspond to objects they see most frequently. While infants may also hear the names of these frequently seen objects, their early visual familiarity plays a critical role in object recognition. Visual understanding may develop before the learning of corresponding names, potentially facilitating the process of word acquisition [33]. For instance, newborn infants have shown an innate ability to recognize visual patterns [13], and their development of visual concepts often precedes the emergence of verbal thought [27]." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.459, + 0.483, + 0.625 + ], + "angle": 0, + "content": "Our Hypothesis: Based on these infant studies, we hypothesize that a computational model trained on an infant's daily experiences may similarly acquire visual concepts extending beyond its linguistic training data. While CVCL demonstrates visual-linguistic mappings of objects named in parental speech, its visual recognition capability should not be limited to parental vocabulary supervision. We conjecture that the vision encoder may develop the ability to recognize concepts beyond these linguistic training data, similar to how infants form object recognition before learning object names." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.627, + 0.483, + 0.779 + ], + "angle": 0, + "content": "To investigate this, we analyze CVCL's internal representations using network dissection [2, 30], or more intuitively, neuron labeling. Furthermore, we implement a neuron-based, training-free classification framework (NeuronClassifier) that leverages visual concepts identified via neuron labeling. Relying solely on these neurons, this approach not only achieves better recognition performance than originally reported [41], but also discovers internal visual concepts extending beyond the model's training vocabulary, supporting our aligned hypothesis." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.781, + 0.483, + 0.903 + ], + "angle": 0, + "content": "From a developmental psychology perspective, the visual concepts discovered beyond the training vocabulary tend to have a higher age of acquisition (AoA) [26] values (Figure 7). This results supports CVCL's ability to develop visual understanding that precedes explicit labeling, mirroring cognitive studies where infants develop pre-verbal visual concepts [27]. This likely reflects real-world learning dynamics: children first acquire labels for frequent, con" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.092, + 0.905, + 0.151 + ], + "angle": 0, + "content": "crete concepts (captured in CVCL's training vocabulary), while visually grounded representations for unlabeled concepts still form earlier than their eventual linguistic acquisition." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.152, + 0.907, + 0.304 + ], + "angle": 0, + "content": "To further explore CVCL's internal representations in the context of computer vision, we compare its visual features with widely-used representations such as CLIP and ImageNet. While CVCL is trained on a unique infant dataset with limited exposure to diverse scenes and a much smaller dataset compared to CLIP, it exhibits similar low-level features in its early layers. However, its higher-level features in the final layer differ significantly. These differences also extend to the visual concept neurons in the model's deeper layers." + }, + { + "type": "title", + "bbox": [ + 0.514, + 0.308, + 0.874, + 0.323 + ], + "angle": 0, + "content": "Contributions The key contributions of our work are:" + }, + { + "type": "text", + "bbox": [ + 0.515, + 0.327, + 0.905, + 0.386 + ], + "angle": 0, + "content": "- We show that infant model have developed understanding beyond linguistic training inputs, by discovering visual concepts hidden in the model representations, aligning with existing infant studies." + }, + { + "type": "text", + "bbox": [ + 0.515, + 0.387, + 0.905, + 0.432 + ], + "angle": 0, + "content": "- We demonstrate that the discovered visual concept neurons can improve object recognition performance by implementing a training-free framework (NeuronClassifier)." + }, + { + "type": "text", + "bbox": [ + 0.515, + 0.433, + 0.905, + 0.492 + ], + "angle": 0, + "content": "- We find that the infant model shares similar low-level representations with ImageNet or CLIP models but diverges in deeper layers due to a lack of diverse higher-level visual concepts." + }, + { + "type": "list", + "bbox": [ + 0.515, + 0.327, + 0.905, + 0.492 + ], + "angle": 0, + "content": null + }, + { + "type": "title", + "bbox": [ + 0.514, + 0.507, + 0.655, + 0.523 + ], + "angle": 0, + "content": "2. Related Work" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.533, + 0.907, + 0.684 + ], + "angle": 0, + "content": "Learning from children. Modeling how children learn has long been a strategy for advancing artificial intelligence. Instead of directly replicating adult intelligence, Alan Turing suggested, \"why not rather try to produce one which simulates the child's?\" [40] Training models on egocentric videos or multimodal data captured from infant perspectives aligns with this idea. [1, 31, 32, 35, 39, 41], as these videos approximate the input available to human infants during development. Our study established on CVCL [41], which we explain in Section 2.1." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.69, + 0.909, + 0.902 + ], + "angle": 0, + "content": "Interpreting vision model representations. Our goal is to understand the model internal neurons trained on infant data. In this context, techniques for interpreting intermediate representations in deep neural networks are relevant. Beginning with Network Dissection [2], a method that quantifies alignment between hidden neurons and visual concepts, numerous studies have aimed to make black-box models more transparent. These methods enable compositional concept discovery [28], assignment of compositional concepts with statistical quantification [5], open-vocabulary neuron captioning [21], and the use of CLIP's rich embeddings for neuron-concept alignment [22, 30]. In addition to direct neuron dissection, other approaches analyze component functions by decomposing image repre" + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.516, + 0.956 + ], + "angle": 0, + "content": "4344" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.09, + 0.092, + 0.483, + 0.138 + ], + "angle": 0, + "content": "sentations to reveal the role of attention heads within multimodal embedding spaces [15] or identifying neurons with similar functions across a diverse model zoo [11]." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.151, + 0.285, + 0.167 + ], + "angle": 0, + "content": "2.1. Preliminary: CVCL" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.174, + 0.484, + 0.34 + ], + "angle": 0, + "content": "Training data. CVCL is trained on the SAYCam-S dataset [37], a longitudinal collection of egocentric recordings from a child aged 6 to 25 months, containing around 200 hours of video. To create meaningful image-text pairs for model training, transcripts were pre-processed to retain only child-directed utterances, excluding the child's own vocalizations. Frames were extracted to align with utterance timestamps. The resulting dataset comprises 600,285 frames paired with 37,500 transcribed utterances, forming a multimodal dataset that simulates the sparse and noisy real-world experiences from which children learn." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.353, + 0.484, + 0.474 + ], + "angle": 0, + "content": "Model architecture. Employing a self-supervised contrastive learning approach akin to CLIP [34], CVCL learns to align egocentric visual frames with transcribed parental speech. Co-occurring pairs are treated as positive examples, while non-co-occurring pairs serve as negatives. This method allows the model to develop multimodal representations without external labels, imitating a child's natural learning process." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.486, + 0.484, + 0.683 + ], + "angle": 0, + "content": "Evaluation. For evaluation, CVCL adopts a \\( n \\)-way classification task (Figure 2) in which the model selects the most relevant visual reference from a set comprising one target image and \\( n - 1 \\) foil images. This approach is inspired by the intermodal preferential looking paradigm (IPLP) [16, 17], used in infant recognition studies to measure language comprehension through differential visual fixation. By aligning its visual and text encoders, CVCL achieves comparable in-domain test accuracy to models like CLIP. However, CVCL demonstrates relatively weak test performance on the Konkle object dataset [23], which includes naturalistic object categories on a white background, using only classes available in the training data." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.701, + 0.182, + 0.716 + ], + "angle": 0, + "content": "3. Method" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.728, + 0.484, + 0.819 + ], + "angle": 0, + "content": "In this section, we describe how to explore neuron-level concepts and leveraging them in \\( n \\) way classification tasks. We begin by using published neuron labeling techniques to discover visual concepts hidden within CVCL [41]. Then, we introduce our framework to utilize these labeled neuron concepts for \\( n \\)-way classification." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.832, + 0.26, + 0.849 + ], + "angle": 0, + "content": "3.1. Neuron Labeling" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.856, + 0.484, + 0.9 + ], + "angle": 0, + "content": "We follow CLIP-Dissect [30] for internal representation analysis due to its flexibility in concept sets and input image dataset." + }, + { + "type": "image", + "bbox": [ + 0.52, + 0.09, + 0.905, + 0.226 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.512, + 0.236, + 0.907, + 0.361 + ], + "angle": 0, + "content": "Figure 2. CVCL's \\( n \\)-way evaluation [41] poses a classification task of choosing the given target object label (\"apple\") from \\( n = 4 \\) images, where only one of them contains the target object. The model feeds the target label into the text encoder and computes the pairwise similarities to each of the \\( n \\) images, selecting the image with the highest similarity to demonstrate object recognition ability. However, this way of recognition limits its ability to the vocabularies in the text encoder. Our framework in Figure 4 overcomes this limitation." + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.388, + 0.906, + 0.449 + ], + "angle": 0, + "content": "Preliminary: CILP-dissect. Given a neural network \\( f(x) \\), where \\( f \\) takes an image \\( x \\) as input and \\( x \\in \\mathcal{D}_{\\mathrm{probe}} \\) with \\( |\\mathcal{D}_{\\mathrm{probe}}| = N \\), and a concept set \\( S \\) with \\( |S| = M \\). The algorithm computes the concept-activation matrix \\( P \\in \\mathbb{R}^{N \\times M} \\)." + }, + { + "type": "equation", + "bbox": [ + 0.661, + 0.463, + 0.905, + 0.48 + ], + "angle": 0, + "content": "\\[\nP _ {i, j} = I _ {i} \\cdot T _ {j}, \\tag {1}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.486, + 0.906, + 0.561 + ], + "angle": 0, + "content": "where \\( I_{i} \\) and \\( T_{j} \\) are the embeddings of the images and concepts, respectively. For each neuron \\( k \\in K \\), where \\( K \\) denotes the set of all neurons in the network, we summarize activations \\( A_{k}(x_{i}) \\) with a scalar function \\( g \\), producing an activation vector:" + }, + { + "type": "equation", + "bbox": [ + 0.562, + 0.572, + 0.905, + 0.591 + ], + "angle": 0, + "content": "\\[\nq _ {k} = \\left[ g \\left(A _ {k} \\left(x _ {1}\\right)\\right), \\dots , g \\left(A _ {k} \\left(x _ {N}\\right)\\right) \\right] ^ {\\top} \\in \\mathbb {R} ^ {N}. \\tag {2}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.602, + 0.904, + 0.632 + ], + "angle": 0, + "content": "The neuron is labeled with the concept that maximizes similarity:" + }, + { + "type": "equation", + "bbox": [ + 0.606, + 0.632, + 0.905, + 0.653 + ], + "angle": 0, + "content": "\\[\nl _ {k} = \\arg \\max _ {m} \\operatorname {s i m} \\left(t _ {m}, q _ {k}; P\\right), \\tag {3}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.66, + 0.906, + 0.735 + ], + "angle": 0, + "content": "where \\(\\operatorname{sim}(\\cdot, \\cdot)\\) represents the similarity function (e.g. cosine similarity or Soft-WPMI [30, 42]), and \\(t_m\\) denotes the most similar text concept. Collecting the label \\(l_k\\) for each neuron, we define the label vector \\(\\mathcal{L} = [l_1, l_2, \\ldots, l_K]\\) to represent the assigned concepts across the entire model." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.735, + 0.906, + 0.779 + ], + "angle": 0, + "content": "During this neuron labeling process, we aim to assign meaningful concepts to each neuron. We use CLIP-dissect because:" + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.78, + 0.907, + 0.901 + ], + "angle": 0, + "content": "- Concept set \\( S \\): Instead of allowing an infinite range of possible concepts for neuron labeling, using a fixed concept set narrows this process by constraining it to a limited selection of concepts. Including diverse concepts in this set enables us to identify neurons corresponding to these visual concepts. This setup ensures that each neuron is assigned a specific label, though some labels may be spurious, as illustrated in Figure 3." + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.516, + 0.957 + ], + "angle": 0, + "content": "4345" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.097, + 0.092, + 0.477, + 0.214 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.09, + 0.227, + 0.483, + 0.311 + ], + "angle": 0, + "content": "Figure 3. Spurious labeled neurons in infant model CVCL by CLIP-Dissect, showing top-3 activated Konkle dataset images. This indicates: (1) not all neurons interpretable with human semantic concepts [12]; (2) neuron-labeling can produce spurious labels. To further support our hypothesis, we implement NeuronClassifier." + }, + { + "type": "text", + "bbox": [ + 0.091, + 0.339, + 0.484, + 0.476 + ], + "angle": 0, + "content": "- Probing dataset \\(\\mathcal{D}_{\\mathrm{probe}}\\): The probing dataset allows neurons to activate specifically in response to the dataset of interest. For instance, when analyzing representations from an infant model, we use images from the Konkle object dataset, which better aligns with infant recognition than the ImageNet [10] validation set. Additionally, for classification tasks, generalization can be achieved by adaptively selecting the probing dataset to focus on relevant concepts within a given dataset." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.482, + 0.484, + 0.68 + ], + "angle": 0, + "content": "Extending beyond vocabulary. The flexibility of the concept set \\( S \\) allowing us to discover visual concepts that the infant model has never encountered in its training data. In this process, CLIP serves as a well-pretrained miner, utilizing its rich image-text embeddings to identify concepts hidden within CVCL. By aligning the activations of CVCL's neurons with CLIP's embeddings, we discover meaningful hidden visual concepts within the infant model, even extending beyond the model's linguistic training data. This approach leverages the diverse vocabulary of the concept set and the rich embeddings of CLIP to reveal visual concepts embedded in CVCL's internal representations, each associated with corresponding neurons." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.69, + 0.345, + 0.705 + ], + "angle": 0, + "content": "3.2. Neuron-Based Classification" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.712, + 0.484, + 0.849 + ], + "angle": 0, + "content": "How do we ensure that neurons representing concepts beyond the model's original vocabulary truly exist within the network? In this section, we propose NeuronClassifier, a training-free framework that leverages neuron activations to detect and validate such concepts. By discovering neurons with specific visual concept, we aim to confirm the presence of these latent, beyond-vocabulary neurons and use them to perform \\( n \\)-way classification. The framework, illustrated in Figure 4, involves three main steps." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.856, + 0.484, + 0.901 + ], + "angle": 0, + "content": "Step 1: neuron labeling with concept set. Given an image encoder \\( f(x) \\), we labeled each neuron in the network using a concept set \\( S \\) that contains (but is not limited to) all" + }, + { + "type": "image", + "bbox": [ + 0.52, + 0.089, + 0.905, + 0.243 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.512, + 0.251, + 0.907, + 0.349 + ], + "angle": 0, + "content": "Figure 4. Our NeuronClassifier Overview. A training-free \\( n \\)-way framework with three key steps: (1) Label all neurons in the network using a concept set that includes class labels and common words; (2) identify neurons associated with the target concept (e.g., \"rug\"); (3) Evaluate the activations of visual concept neuron across \\( n \\) images (in this example, \\( n = 4 \\)) and select the image with the highest activation as the most relevant to the target label." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.377, + 0.907, + 0.407 + ], + "angle": 0, + "content": "class labels relevant to the task. The dissection process can be expressed as a function:" + }, + { + "type": "equation", + "bbox": [ + 0.606, + 0.421, + 0.905, + 0.437 + ], + "angle": 0, + "content": "\\[\n\\mathcal {N} _ {s} = \\text {N e u r o n L a b e l i n g} (f, \\mathcal {S}), \\tag {4}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.45, + 0.906, + 0.527 + ], + "angle": 0, + "content": "where \\(\\mathcal{N}_s\\) is the set of neurons labeled with concepts from \\(S\\). Each neuron \\(k \\in \\mathcal{N}_s\\) is associated with a specific concept, such as \"rug\" or \"calculator\", based on its alignment with concept embeddings obtained during neuron labeling (e.g. similarity in CLIP-Dissect [30])." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.535, + 0.906, + 0.626 + ], + "angle": 0, + "content": "Step 2: identifying visual concept neurons. Given a target label \\( l_{k} \\in \\mathcal{L} \\), where \\( \\mathcal{L} \\) represents all neuron labels in the model, same as label vector in Section 3.1 (e.g., \"rug\"), we select the subset of neurons labeled with this concept, denoted as \\( \\mathcal{N}_{l_k} \\subset \\mathcal{N}_{\\mathcal{L}} \\). These neurons are responsible for encoding the target concept." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.626, + 0.906, + 0.73 + ], + "angle": 0, + "content": "To further refine the labeling and reduce spurious assignments, we select the most similar neuron from \\(\\mathcal{N}_{l_k}\\) to represent the target concept. The similarity measure varies based on the dissection method used. For example, Network Dissection [2] employs Intersection over Union (IoU) for similarity, while CLIP-Dissect [30] supports multiple similarity metrics:" + }, + { + "type": "equation", + "bbox": [ + 0.602, + 0.745, + 0.905, + 0.768 + ], + "angle": 0, + "content": "\\[\nk ^ {*} = \\arg \\max _ {k \\in \\mathcal {N} _ {\\mathbf {L}}} \\operatorname {s i m} \\left(t _ {l _ {k}}, q _ {k}; P\\right), \\tag {5}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.781, + 0.906, + 0.841 + ], + "angle": 0, + "content": "where \\(t_{l_k}\\) is the embedding of the target label \\(l_k\\), \\(q_k\\) is the neuron activation value, same in Section 3.1. This step ensures that the neuron most aligned with the concept is selected, minimizing the possibility of spurious labeling." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.842, + 0.906, + 0.872 + ], + "angle": 0, + "content": "For each selected neuron \\(k \\in \\mathcal{N}_{l_k}\\), its activation value on an input image \\(x_i\\) is computed as" + }, + { + "type": "equation", + "bbox": [ + 0.636, + 0.886, + 0.905, + 0.902 + ], + "angle": 0, + "content": "\\[\nq _ {k} \\left(x _ {i}\\right) = g \\left(A _ {k} \\left(x _ {i}\\right)\\right), \\tag {6}\n\\]" + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.516, + 0.956 + ], + "angle": 0, + "content": "4346" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.09, + 0.092, + 0.483, + 0.137 + ], + "angle": 0, + "content": "where \\( A_{k}(x_{i}) \\) is the raw activation map, and \\( g(\\cdot) \\) is a summary function (e.g., spatial mean) that reduces it to a scalar representing the neuron's response strength." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.141, + 0.483, + 0.217 + ], + "angle": 0, + "content": "Step 3: selecting the most relevant image. Given \\( n \\) candidate images \\( \\{x_1, x_2, \\ldots, x_n\\} \\) in an \\( n \\)-way trial, we compute the activation values \\( q_k(x_i) \\) for neuron \\( k^* \\) with highest similarity across all images. The image with the highest activation is selected as the most relevant to the target concept:" + }, + { + "type": "equation", + "bbox": [ + 0.208, + 0.231, + 0.482, + 0.252 + ], + "angle": 0, + "content": "\\[\nx ^ {*} = \\arg \\max _ {x _ {i}} q _ {k ^ {*}} \\left(x _ {i}\\right). \\tag {7}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.091, + 0.257, + 0.483, + 0.302 + ], + "angle": 0, + "content": "In the example shown in Figure 4, the target concept is \"rug\", and we select the image with the highest activation from the four candidates as the closest match to the concept." + }, + { + "type": "title", + "bbox": [ + 0.09, + 0.314, + 0.429, + 0.332 + ], + "angle": 0, + "content": "4. Neuron-wise Representation Analysis" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.339, + 0.484, + 0.475 + ], + "angle": 0, + "content": "In this section, we conduct a neuron-wise analysis of infant models and provide implementation details. First, we perform neuron labeling on the infant CVCL model. Then, we use our NeuronClassifier framework to leverage visual concept neurons identified through neuron labeling, resulting in better recognition performance than the original approach [41] while also discovering internal visual concepts beyond the model's training vocabulary. These results support our hypothesis." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.483, + 0.173, + 0.499 + ], + "angle": 0, + "content": "4.1. Setup" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.505, + 0.484, + 0.656 + ], + "angle": 0, + "content": "Datasets. We use the Konkle object dataset [23], as introduced in Section 2.1. This dataset consists of 3,406 images, each featuring a single object on a clean white background, including 406 test items across 200 classes. Each trial comprises \\( n \\) images: one target image and the remaining as foils, with foil images randomly sampled from classes other than the target class. For each class, we generate 5 trials, each containing \\( n \\) images. Following the previous work on CVCL, we use \\( n = 4 \\) in our main experiments, with one target and three foils per trial." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.66, + 0.483, + 0.78 + ], + "angle": 0, + "content": "Neuron labeling. We utilize CLIP-Dissect [30] for neuron labeling, which assigns visual concepts to each neuron in the network. As we aim to perform classification on the Konkle object dataset, we use the same dataset as a part of \\(\\mathcal{D}_{\\mathrm{probe}}\\). Additionally, to avoid limiting the search space only around class names and ensure comprehensive neuron labeling, we employ a combined concept, consisting of three components:" + }, + { + "type": "text", + "bbox": [ + 0.091, + 0.781, + 0.483, + 0.825 + ], + "angle": 0, + "content": "- SAYCam-S vocabulary: We clean the original vocabulary by removing noisy child speech and retaining meaningful words." + }, + { + "type": "text", + "bbox": [ + 0.091, + 0.826, + 0.483, + 0.871 + ], + "angle": 0, + "content": "- Common English words: We chose the top 30,000 most common English words based on a 1-gram frequency analysis by Peter Norvig [29, 30]." + }, + { + "type": "text", + "bbox": [ + 0.091, + 0.871, + 0.483, + 0.901 + ], + "angle": 0, + "content": "- Class in Konkle object dataset: All class labels from the Konkle object dataset are included." + }, + { + "type": "list", + "bbox": [ + 0.091, + 0.781, + 0.483, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.092, + 0.905, + 0.122 + ], + "angle": 0, + "content": "We combine these three sources, ensuring no duplicates, resulting in a final concept set containing 30,427 words." + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.127, + 0.905, + 0.187 + ], + "angle": 0, + "content": "Models. Our primary focus is the CVCL-ResNeXt50 [41], trained on SAY-Cam-S [37] dataset. For comprehensive analysis, we apply our framework to the following models:" + }, + { + "type": "text", + "bbox": [ + 0.514, + 0.19, + 0.905, + 0.28 + ], + "angle": 0, + "content": "- Infant models: CVCL is trained on unique infant data, using both egocentric frames and transcribed parent speech. We also take DINO-S-ResNeXt50 [31] as reference model compare recognition ability derived from same visual experience. It trained with DINO[6] self-supervised approach on the infant same dataset." + }, + { + "type": "text", + "bbox": [ + 0.514, + 0.281, + 0.905, + 0.371 + ], + "angle": 0, + "content": "- Broadly-trained models: To establish an upper bound, we include CLIP [34] and ResNeXt50 [43], which are broadly trained on large scale Internet images. Although CLIP uses a ResNet50-based vision encoder rather than ResNeXt, we select CLIP-ResNet50 due to the architectural similarity between ResNeXt [43] and ResNet [20]." + }, + { + "type": "text", + "bbox": [ + 0.514, + 0.372, + 0.905, + 0.432 + ], + "angle": 0, + "content": "- Randomized model: As a lower bound, we introduce a randomized version of the CVCL-ResNeXt50 model. In this setup, the convolution layer weights in the vision encoder are initialized using Kaiming Initialization [19]." + }, + { + "type": "list", + "bbox": [ + 0.514, + 0.19, + 0.905, + 0.432 + ], + "angle": 0, + "content": null + }, + { + "type": "title", + "bbox": [ + 0.514, + 0.442, + 0.608, + 0.457 + ], + "angle": 0, + "content": "4.2. Results" + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.464, + 0.905, + 0.525 + ], + "angle": 0, + "content": "In this section, we present our results and findings from applying neuron labeling and our NeuronClassifier framework to the infant CVCL model, compared with other reference models. We define two class types for our analysis:" + }, + { + "type": "text", + "bbox": [ + 0.514, + 0.528, + 0.905, + 0.572 + ], + "angle": 0, + "content": "- In-vocabulary classes: Object classes present in the model's training linguistic input, which also appear in test object class and are detectable in internal representations." + }, + { + "type": "text", + "bbox": [ + 0.514, + 0.573, + 0.905, + 0.618 + ], + "angle": 0, + "content": "- Out-of-vocabulary classes: Object classes not included in the training linguistic input but detected in test object class and internal representations." + }, + { + "type": "list", + "bbox": [ + 0.514, + 0.528, + 0.905, + 0.618 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.584, + 0.634, + 0.84, + 0.709 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.513, + 0.72, + 0.905, + 0.748 + ], + "angle": 0, + "content": "Figure 5. In-vocabulary and out-of-vocabulary relationships visualized using a Venn diagram." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.765, + 0.907, + 0.902 + ], + "angle": 0, + "content": "Our results demonstrate that the proposed framework effectively discovers meaningful neurons that represent concepts beyond the model's training vocabulary. These findings align with the cognitive perspective of vocabulary acquisition in infant development. We analyze the models' performance across different \\( n \\)-way classification settings, showing that our method yields strong out-of-vocabulary classification performance while simultaneously improving in-vocabulary classification accuracy." + }, + { + "type": "page_number", + "bbox": [ + 0.482, + 0.945, + 0.516, + 0.956 + ], + "angle": 0, + "content": "4347" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.099, + 0.089, + 0.473, + 0.246 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.09, + 0.256, + 0.483, + 0.34 + ], + "angle": 0, + "content": "Figure 6. Class coverage in visual concept neurons. Percentage of Konkle object dataset [23] classes identified through neuron labeling. Broadly pre-trained models (CLIP, ResNeXt) achieve over \\(50\\%\\) coverage, while developmentally inspired infant models (CVCL [41], DINO [6]) show lower coverage. CVCL-Randomized provides a lower-bound comparison." + }, + { + "type": "text", + "bbox": [ + 0.089, + 0.367, + 0.483, + 0.593 + ], + "angle": 0, + "content": "Class coverage in visual concept neurons. How well do visual concept neurons identified through neuron labeling correspond to specific class names (e.g., \"rug\") rather than general descriptive attributes (e.g., \"red\"? Figure 6 shows the percentage of classes in the Konkle object dataset that are discovered in visual concept neurons from each model during the neuron labeling process. The results indicate that well-pretrained models, such as CLIP-ResNet50 and ResNeXt50, demonstrate broader class name coverage. While CVCL performs slightly weaker, it still maintains coverage of slightly less than \\(50\\%\\). In contrast, CVCL-Randomized achieves only around \\(28\\%\\) coverage. This class coverage metric reflects the models' capacity to form class-corresponding meaningful representations during the neuron labeling process." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.599, + 0.483, + 0.779 + ], + "angle": 0, + "content": "Age of Acquisition (AoA) ratings. Age of acquisition (AoA) is used to indicate when, and in what sequence, words are learned, and it is often assessed through ratings or observations reported by adults. This indirect method generally correlates well with other metrics indicating when children acquire vocabulary. Previous developmental work has shown that infants' early visual familiarity with common objects helps with object recognition, which subsequently helps support the process of learning the names of those objects. We next examined how early words in our models are learned and whether there is an AoA difference between in-vocab and out-of-vocab words." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.78, + 0.483, + 0.902 + ], + "angle": 0, + "content": "We used AoA ratings from a dataset compiled by Kuperman, Stadthagen-Gonzalez, and Brysbaert [25], which includes norms for over 30,000 English words gathered via Amazon Mechanical Turk. Each participant estimated the age in years at which they believed they first understood each word, even if they did not actively use them. This dataset is comparable to previously reported AoA norms [36] gathered in laboratory settings." + }, + { + "type": "image", + "bbox": [ + 0.517, + 0.09, + 0.887, + 0.265 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.513, + 0.275, + 0.906, + 0.361 + ], + "angle": 0, + "content": "Figure 7. Age of acquisition (AoA) ratings for in- and out-of-vocabulary visual concepts. Comparison of word acquisition timing [25] between in-vocabulary and out-of-vocabulary concepts. Out-of-vocabulary concepts tend to have a higher estimated acquisition ages in the infant-inspired CVCL model, indicating development of visual understanding beyond explicit linguist inputs." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.389, + 0.908, + 0.617 + ], + "angle": 0, + "content": "Using this set of AoA norms, we compared mean AoA between in-vocab and out-of-vocab words discovered in CVCL's internal representations. As shown in Figure 7, we found a significant difference between in-vocab and out-of-vocab AoA rating (\\( t(82) = 4.64 \\), \\( p < 0.0001 \\)), in-vocab words (mean AoA = 4.99) are learned earlier than out-of-vocab words (mean AoA = 6.82). This pattern suggests that: (1) both sets of words are learned quite early, around later preschool and school years, with or without supervised labeling; (2) the difference in AoA between in-vocabulary and out-of-vocabulary words indicates that the infant model has developed a basic visual understanding of concepts with higher AoA. This foundational knowledge may lay the groundwork for word learning once corresponding parental speech is introduced." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.628, + 0.909, + 0.854 + ], + "angle": 0, + "content": "Neuron-based classification performance. We evaluated the hidden potential of infant models' vision encoders by applying our NeuronClassifier framework, summarized in Table 1. Despite being trained on infant egocentric data with limited amount and diversity, CVCL demonstrates the ability to recognize similarly as nature of infant learning, revealing strong out-of-vocabulary classification performance. This result suggests that this infant model developed broader visual concepts that extend beyond linguistic input, similar to how infants naturally learn. We also applied our method to in-vocabulary classification, where it outperformed the vanilla method previously used in CVCL [41], as introduced in Figure 2. \"All\" representing the combined performance on both \"in-vocab\" and \"out-of-vocab\". These results support our hypothesis." + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.856, + 0.908, + 0.903 + ], + "angle": 0, + "content": "We include additional model comparisons: (1) The DINO-S-ResNeXt50 infant model [31], trained on the same dataset without text supervision, achieves comparable per" + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.516, + 0.956 + ], + "angle": 0, + "content": "4348" + } + ], + [ + { + "type": "table_caption", + "bbox": [ + 0.09, + 0.089, + 0.485, + 0.23 + ], + "angle": 0, + "content": "Table 1. Neuron-based classification results in 4-way evaluation among models in Section 4.1. \"Vanilla\" refers to classification based on image-text similarity (Figure 2). \"X\" denotes cases where direct classification on the Konkle dataset [23] is not possible due to missing text encoder or need fine-tuning. By leveraging neurons discovered in the representation, NeuronClassifier enables broader recognition, particularly in CVCL (bolded for emphasis), achieving improved recognition in both in-vocabulary and out-of-vocabulary, supporting our hypothesis. \"All\" represents the combined performance on both in- and out-of-vocabulary." + }, + { + "type": "table", + "bbox": [ + 0.093, + 0.239, + 0.485, + 0.367 + ], + "angle": 0, + "content": "
MethodModelIn-vocabOut-of-vocabAll
VanillaCLIP-ResNet5098.81±0.1696.93±0.0697.42±0.05
ResNeXt50×××
CVCL-ResNeXt5036.18±0.91××
DINO-S-ResNeXt50×××
Neuron ClassifierCLIP-ResNet5091.59±0.5288.66±0.3589.79±0.38
ResNeXt5088.17±0.4593.28±0.3691.88±0.15
CVCL-ResNeXt5079.50±0.7876.81±0.3577.79±0.40
DINO-S-ResNeXt5077.53±0.2477.96±0.2777.65±0.21
" + }, + { + "type": "text", + "bbox": [ + 0.089, + 0.393, + 0.483, + 0.544 + ], + "angle": 0, + "content": "formance in visual representations. This suggests that models trained on identical data distributions with different self-supervised methods may yield similar representational outcomes. (2) Broadly-trained models such as ResNeXt50 and CLIP establish performance upper bounds. However, CLIP underperforms in the vanilla setting, relying exclusively on visual neurons without text encoder guidance. While this work does not aim to advance zero-shot learning methods, it reveals visual concepts in infant models that emerge independently of linguistic inputs." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.545, + 0.484, + 0.635 + ], + "angle": 0, + "content": "In classification on the Konkle object dataset, both DINO-S and ImageNet ResNeXt50 required fine-tuning for this task. Our framework enables neuron-based classification without downstream fine-tuning, and providing a training-free qualitative inspection of internal representations." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.644, + 0.483, + 0.704 + ], + "angle": 0, + "content": "Analysis across \\( n \\)-way settings. We evaluate the models under various \\( n \\)-way classification setups. Figure 8 illustrates the performance trends for in- and out-of-vocabulary class classification accuracy applying our NeuronClassifier." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.705, + 0.483, + 0.902 + ], + "angle": 0, + "content": "Our method not only enables out-of-vocabulary classification but also significantly improves in-vocab performance compared to previous results [41], further supporting the presence of beyond-vocabulary potential in infant models. However, due to limited class coverage, CVCL with our method can classify a maximum of 31 classes (see Appendix A.3). These findings support our hypothesis that the infant model has acquired visual concepts beyond its initial vocabulary. These results show that leveraging the model's internal representations for classification that go beyond its vocabulary is not only feasible but also robust across different \\( n \\)-way settings. These findings support our hypothesis that the infant model has acquired visual concepts beyond" + }, + { + "type": "image", + "bbox": [ + 0.517, + 0.091, + 0.714, + 0.313 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.722, + 0.091, + 0.882, + 0.312 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.512, + 0.323, + 0.908, + 0.448 + ], + "angle": 0, + "content": "Figure 8. In- and out-of-vocabulary class performance across \\( n \\)-way settings using NeuronClassifier. (a) For out-of-vocabulary classification (left), our method enables classification without additional training. The infant model CVCL (\\(\\clubsuit\\)) maintains robust performance as \\( n \\) increases. (b) For in-vocabulary classification (right), \"ours\" (NeuronClassifier) enhances CVCL recognition ability compared to the \"vanilla\" setting. However, CLIP (\\(\\clubsuit\\)) declines, as it relies solely on neurons without text encoder input. Random ones serve as lower bounds." + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.476, + 0.655, + 0.491 + ], + "angle": 0, + "content": "its initial vocabulary." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.492, + 0.907, + 0.658 + ], + "angle": 0, + "content": "However, as \\( n \\) increases exponentially, performance gradually declines. This decline is reasonable, as the task becomes increasingly difficult by nature as \\( n \\) increases. It may also be attributed to dimensionality reduction in activation maps, leading to coarser classification. CLIPRN50 and ResNeXt50 perform well with NeuronClassifier, though not as effectively as their direct or fine-tuned versions, since our method is designed to reveal latent concepts rather than to perform fully optimized classification. In in-vocab settings, the \"vanilla\" approach represents direct classification as shown in Figure 2." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.673, + 0.838, + 0.691 + ], + "angle": 0, + "content": "5. Layer-wise Representation Analysis" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.698, + 0.907, + 0.85 + ], + "angle": 0, + "content": "How does the representation learned from infant data differ from other representations widely used in the computer vision community? To explore this, we perform a layer-wise representation analysis using Centered Kernel Alignment (CKA) [24]. We compute the similarity between the representations of the infant model(CVCL), ImageNet-pretrained, and CLIP. Additionally, we apply neuron labeling techniques from a layer-wise perspective to identify unique visual concepts discovered at each layer between ImageNet and infant models." + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.856, + 0.909, + 0.903 + ], + "angle": 0, + "content": "Layer similarity analysis. CKA [24] applies HSIC [18] (see Appendix B) over a set of images to provide layer-wise similarity scores between the representations of different" + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.516, + 0.956 + ], + "angle": 0, + "content": "4349" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.093, + 0.092, + 0.482, + 0.236 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.09, + 0.245, + 0.485, + 0.345 + ], + "angle": 0, + "content": "Figure 9. CKA layer-wise similarity between CVCL and common models. Using the ImageNet validation set as input, CVCL (y-axis) exhibits similarity to CLIP-RN50 (x-axis, left) and ResNeXt50 (x-axis, right) in the shallow layers (lower-level features) but diverges significantly in the final layer (higher-level features). Notably, Layer 4 of CVCL shows very low similarity to all layers of both common models." + }, + { + "type": "text", + "bbox": [ + 0.089, + 0.373, + 0.483, + 0.63 + ], + "angle": 0, + "content": "neural networks. A higher CKA score indicates more similar representations between two models at the given layer. We use the ImageNet validation set [10] as input for the networks, and compute the CKA similarity between the infant model (CVCL), the ImageNet-pretrained ResNeXt50, and CLIP-ResNet50. The results are presented as matrices in Figure 9. The lower layers of CVCL exhibit greater similarity to larger-scale models than its deeper layers. In larger-scale models, shallow layers are known to capture low-complexity features, while deeper layers progressively specialize in capturing higher-level concepts [7, 14]. Therefore, CVCL – the model trained on infant data – successfully develops lower-level representations comparable to those in common pre-trained models. However, the divergence in deeper layers suggests a lack of the diverse higher-level representations typically observed in models trained on common datasets." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.644, + 0.483, + 0.901 + ], + "angle": 0, + "content": "Neuron-based analysis. To investigate the characteristics of each layer, we apply network dissection to identify neurons that are aligned with specific visual concepts. Essentially, this is an extension of the neuron labeling process, where the Broden [2] dataset provides category labels for each visual concept neuron. We count the number of unique visual concepts discovered in each category and perform layer-wise comparisons to gain a broader view of the differences between models trained from ImageNet and infant data. In Figure 10, we visualize the number of unique visual concept neurons across layers for each model. The results show that early layers in both models predominantly have neurons of low-level features like color, with minimal differences between models. However, as we move to deeper layers, higher-level concepts such as objects and scenes become more prominent, and the disparities between models become clearer. CVCL exhibits fewer unique visual con" + }, + { + "type": "image", + "bbox": [ + 0.514, + 0.091, + 0.895, + 0.246 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.512, + 0.257, + 0.908, + 0.368 + ], + "angle": 0, + "content": "Figure 10. Number of visual concepts in ImageNet and infant models. Concepts are categorized using Broden dataset [2]. Neurons in deeper model layers capture increasingly complex concepts. Early layers primarily detect lower-level features like color and texture, while higher-level concepts such as objects and scenes emerge in deeper layers. CVCL exhibits fewer visual concepts than the ImageNet model, especially for higher-level visual concepts (e.g., objects and scenes)." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.395, + 0.907, + 0.427 + ], + "angle": 0, + "content": "cepts in these higher-level categories compared to ImageNet model. This finding aligns with the layer similarity analysis." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.439, + 0.634, + 0.455 + ], + "angle": 0, + "content": "6. Conclusion" + }, + { + "type": "text", + "bbox": [ + 0.511, + 0.464, + 0.907, + 0.674 + ], + "angle": 0, + "content": "In this paper, we explored whether an infant model (CVCL), trained on infant egocentric video frames and linguistic inputs, can acquire broader visual concepts extending beyond its initial training vocabulary. By introducing NeuronClassifier, a training-free framework to discover and leverage visual concepts hidden in representations, we unlocked the CVCL visual encoder's ability to recognize out-of-vocabulary concepts, establishing its potential as a strong classifier. Our findings also reveal that while CVCL, trained on a unique infant dataset with limited exposure to diverse scenes, it representations capture low-level features similar to those in common pre-trained models, they diverge significantly in higher-level representations, contributing to the observed performance differences." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.675, + 0.907, + 0.752 + ], + "angle": 0, + "content": "Overall, our approach bridges cognitive science and computer vision, providing insights into how infant models develop visual concepts that precede linguistic inputs, aligning with the natural way infants explore the world through sight." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.755, + 0.909, + 0.89 + ], + "angle": 0, + "content": "Limitations and future work. Our study did not analyze the infant training data, as we were unable to access it due to strict access controls. Instead, we analyze models trained on infant egocentric data, finding developmental alignment with cognitive studies. However, we did not extend this analysis to adult egocentric data. While the study focuses on infant data for developmental process, the framework can be applied to adult models, which remains a direction for future research." + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.516, + 0.956 + ], + "angle": 0, + "content": "4350" + } + ], + [ + { + "type": "title", + "bbox": [ + 0.092, + 0.091, + 0.259, + 0.108 + ], + "angle": 0, + "content": "Acknowledgements" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.115, + 0.484, + 0.22 + ], + "angle": 0, + "content": "This work was partially supported by the National Research Foundation Singapore Competitive Research Program (award number CRP29-2022-0003). We thank Wai Keen Vong for invaluable discussion and CVCL's pretrained weights. We thank Jingyi Lin for figure discussions. We thank the anonymous reviewers for their constructive feedback." + }, + { + "type": "title", + "bbox": [ + 0.093, + 0.236, + 0.188, + 0.252 + ], + "angle": 0, + "content": "References" + }, + { + "type": "ref_text", + "bbox": [ + 0.101, + 0.261, + 0.484, + 0.304 + ], + "angle": 0, + "content": "[1] Sven Bambach, David Crandall, Linda Smith, and Chen Yu. Toddler-inspired visual object learning. In Neural Information Processing Systems, 2018. 1, 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.101, + 0.304, + 0.484, + 0.375 + ], + "angle": 0, + "content": "[2] David Bau, Bolei Zhou, Aditya Khosla, Aude Oliva, and Antonio Torralba. Network dissection: Quantifying interpretability of deep visual representations. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 6541-6549, 2017. 2, 4, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.376, + 0.483, + 0.431 + ], + "angle": 0, + "content": "[3] Erika Bergelson and Daniel Swingley. At 6-9 months, human infants know the meanings of many common nouns. Proceedings of the National Academy of Sciences, 109(9): 3253-3258, 2012. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.433, + 0.483, + 0.502 + ], + "angle": 0, + "content": "[4] Marc Brysbaert and Boris New. Moving beyond kucera and francis: A critical evaluation of current word frequency norms and the introduction of a new and improved word frequency measure for american english. Behavior Research Methods, 41(4):977-990, 2009. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.504, + 0.483, + 0.559 + ], + "angle": 0, + "content": "[5] Kirill Bykov, Laura Kopf, Shinichi Nakajima, Marius Kloft, and Marina Höhne. Labeling neural representations with inverse recognition. Advances in Neural Information Processing Systems, 36, 2024. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.561, + 0.483, + 0.63 + ], + "angle": 0, + "content": "[6] Mathilde Caron, Hugo Touvron, Ishan Misra, Hervé Jégou, Julien Mairal, Piotr Bojanowski, and Armand Joulin. Emerging properties in self-supervised vision transformers. In Proceedings of the IEEE/CVF international conference on computer vision, pages 9650-9660, 2021. 5, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.632, + 0.483, + 0.699 + ], + "angle": 0, + "content": "[7] Yixiong Chen, Alan Yuille, and Zongwei Zhou. Which layer is learning faster? a systematic exploration of layerwise convergence rate for deep neural networks. In The Eleventh International Conference on Learning Representations, 2023. 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.702, + 0.483, + 0.77 + ], + "angle": 0, + "content": "[8] Elizabeth M Clerkin, Elizabeth Hart, James M Rehg, Chen Yu, and Linda B Smith. Real-world visual statistics and infants' first-learned object names. Philosophical Transactions of the Royal Society B: Biological Sciences, 372(1711): 20160055, 2017. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.773, + 0.483, + 0.814 + ], + "angle": 0, + "content": "[9] Michael J Cortese and Maya M Khanna. Age of acquisition ratings for 3,000 monosyllabic words. Behavior Research Methods, 40(3):791-794, 2008. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.817, + 0.483, + 0.871 + ], + "angle": 0, + "content": "[10] Jia Deng, Wei Dong, Richard Socher, Li-Jia Li, Kai Li, and Li Fei-Fei. Imagenet: A large-scale hierarchical image database. In 2009 IEEE conference on computer vision and pattern recognition, pages 248-255. IEEE, 2009. 4, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.873, + 0.483, + 0.901 + ], + "angle": 0, + "content": "[11] Amir Dravid, Yossi Gandelsman, Alexei A Efros, and Assaf Shocher. Rosetta neurons: Mining the common units in a" + }, + { + "type": "list", + "bbox": [ + 0.094, + 0.261, + 0.484, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.545, + 0.093, + 0.906, + 0.121 + ], + "angle": 0, + "content": "model zoo. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 1934-1943, 2023. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.124, + 0.906, + 0.191 + ], + "angle": 0, + "content": "[12] Nelson Elhage, Tristan Hume, Catherine Olsson, Nicholas Schiefer, Tom Henighan, Shauna Kravec, Zac Hatfield-Dodds, Robert Lasenby, Dawn Drain, Carol Chen, et al. Toy models of superposition. arXiv preprint arXiv:2209.10652, 2022.4" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.194, + 0.905, + 0.223 + ], + "angle": 0, + "content": "[13] Robert L Fantz. Pattern vision in newborn infants. Science, 140(3564):296-297, 1963. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.225, + 0.906, + 0.28 + ], + "angle": 0, + "content": "[14] Thomas Fel, Louis Bethune, Andrew Kyle Lampinen, Thomas Serre, and Katherine Hermann. Understanding visual feature reliance through the lens of complexity. arXiv preprint arXiv:2407.06076, 2024. 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.283, + 0.905, + 0.325 + ], + "angle": 0, + "content": "[15] Yossi Gandelsman, Alexei A Efros, and Jacob Steinhardt. Interpreting clip's image representation via text-based decomposition. arXiv preprint arXiv:2310.05916, 2023. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.327, + 0.908, + 0.383 + ], + "angle": 0, + "content": "[16] Roberta Michnick Golinkoff, Kathryn Hirsh-Pasek, Kathleen M Cauley, and Laura Gordon. The eyes have it: Lexical and syntactic comprehension in a new paradigm. Journal of child language, 14(1):23-45, 1987. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.385, + 0.906, + 0.453 + ], + "angle": 0, + "content": "[17] Roberta Michnick Golinkoff, Weiyi Ma, Lulu Song, and Kathy Hirsh-Pasek. Twenty-five years using the intermodal preferential looking paradigm to study language acquisition: What have we learned? Perspectives on Psychological Science, 8(3):316-339, 2013. 2, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.457, + 0.906, + 0.512 + ], + "angle": 0, + "content": "[18] Arthur Gretton, Ralf Herbrich, Alexander Smola, Olivier Bousquet, Bernhard Schölkopf, and Aapo Hyvarinen. Kernel methods for measuring independence. Journal of Machine Learning Research, 6(12), 2005. 7, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.515, + 0.906, + 0.583 + ], + "angle": 0, + "content": "[19] Kaiming He, Xiangyu Zhang, Shaoqing Ren, and Jian Sun. Delving deep into rectifiers: Surpassing human-level performance on imagenet classification. In Proceedings of the IEEE international conference on computer vision, pages 1026-1034, 2015. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.586, + 0.905, + 0.642 + ], + "angle": 0, + "content": "[20] Kaiming He, Xiangyu Zhang, Shaoqing Ren, and Jian Sun. Deep residual learning for image recognition. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 770-778, 2016. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.644, + 0.905, + 0.699 + ], + "angle": 0, + "content": "[21] Evan Hernandez, Sarah Schwettmann, David Bau, Teona Bagashvili, Antonio Torralba, and Jacob Andreas. Natural language descriptions of deep visual features. In International Conference on Learning Representations, 2021. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.702, + 0.906, + 0.77 + ], + "angle": 0, + "content": "[22] Neha Kalibhat, Shweta Bhardwaj, C Bayan Bruss, Hamed Firooz, Maziar Sanjabi, and Soheil Feizi. Identifying interpretable subspaces in image representations. In International Conference on Machine Learning, pages 15623-15638. PMLR, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.773, + 0.906, + 0.843 + ], + "angle": 0, + "content": "[23] Talia Konkle, Timothy F Brady, George A Alvarez, and Aude Oliva. Conceptual distinctiveness supports detailed visual long-term memory for real-world objects. Journal of experimental Psychology: general, 139(3):558, 2010. 3, 5, 6, 7, 1, 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.846, + 0.906, + 0.901 + ], + "angle": 0, + "content": "[24] Simon Kornblith, Mohammad Norouzi, Honglak Lee, and Geoffrey Hinton. Similarity of neural network representations revisited. In International conference on machine learning, pages 3519-3529. PMLR, 2019. 7" + }, + { + "type": "list", + "bbox": [ + 0.517, + 0.093, + 0.908, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.514, + 0.957 + ], + "angle": 0, + "content": "4351" + } + ], + [ + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.093, + 0.482, + 0.134 + ], + "angle": 0, + "content": "[25] Victor Kuperman, Hans Stadthagen-Gonzalez, and Marc Brysbaert. Age-of-acquisition ratings for 30,000 english words. Behavior research methods, 44:978-990, 2012. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.136, + 0.482, + 0.19 + ], + "angle": 0, + "content": "[26] Victor Kuperman, Hans Stadthagen-Gonzalez, and Marc Brysbaert. Age-of-acquisition ratings for 30,000 english words. Behavior Research Methods, 44(4):978-990, 2012. 2, 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.192, + 0.482, + 0.22 + ], + "angle": 0, + "content": "[27] Jean M Mandler. How to build a baby: II. conceptual primitives. Psychological review, 99(4):587, 1992. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.222, + 0.482, + 0.262 + ], + "angle": 0, + "content": "[28] Jesse Mu and Jacob Andreas. Compositional explanations of neurons. Advances in Neural Information Processing Systems, 33:17153-17163, 2020. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.264, + 0.482, + 0.29 + ], + "angle": 0, + "content": "[29] Peter Norvig. Natural language corpus data: Beautiful data, 2009. Accessed: 2024-10-27. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.293, + 0.482, + 0.334 + ], + "angle": 0, + "content": "[30] Tuomas Oikarinen and Tsui-Wei Weng. Clip-dissect: Automatic description of neuron representations in deep vision networks. arXiv preprint arXiv:2204.10965, 2022. 2, 3, 4, 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.336, + 0.482, + 0.389 + ], + "angle": 0, + "content": "[31] A Emin Orhan and Brenden M Lake. Learning high-level visual representations from a child's perspective without strong inductive biases. Nature Machine Intelligence, 6(3):271-283, 2024. 1, 2, 5, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.392, + 0.482, + 0.446 + ], + "angle": 0, + "content": "[32] Emin Orhan, Vaibhav Gupta, and Brenden M Lake. Self-supervised learning through the eyes of a child. Advances in Neural Information Processing Systems, 33:9960-9971, 2020. 1, 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.449, + 0.482, + 0.489 + ], + "angle": 0, + "content": "[33] Deborah A Phillips and Jack P Shonkoff. From neurons to neighborhoods: The science of early childhood development. 2000. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.491, + 0.482, + 0.573 + ], + "angle": 0, + "content": "[34] Alec Radford, Jong Wook Kim, Chris Hallacy, Aditya Ramesh, Gabriel Goh, Sandhini Agarwal, Girish Sastry, Amanda Askell, Pamela Mishkin, Jack Clark, et al. Learning transferable visual models from natural language supervision. In International conference on machine learning, pages 8748-8763. PMLR, 2021. 2, 3, 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.576, + 0.482, + 0.63 + ], + "angle": 0, + "content": "[35] Saber Sheybani, Himanshu Hansaria, Justin Wood, Linda Smith, and Zoran Tiganj. Curriculum learning with infant egocentric videos. Advances in Neural Information Processing Systems, 36, 2024. 1, 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.633, + 0.482, + 0.673 + ], + "angle": 0, + "content": "[36] Hans Stadthagen-Gonzalez and Colin J Davis. The bristol norms for age of acquisition, imageability, and familiarity. Behavior Research Methods, 38(4):598-605, 2006. 6, 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.676, + 0.482, + 0.73 + ], + "angle": 0, + "content": "[37] Jessica Sullivan, Michelle Mei, Andrew Perfors, Erica Wojcik, and Michael C Frank. Saycam: A large, longitudinal audiovisual dataset recorded from the infant's perspective. Open mind, 5:20-29, 2021. 3, 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.732, + 0.482, + 0.786 + ], + "angle": 0, + "content": "[38] Gentaro Taga, Kayo Asakawa, Atsushi Maki, Yukuo Konishi, and Hideaki Koizumi. Brain imaging in awake infants by near-infrared optical topography. Proceedings of the National Academy of Sciences, 100(19):10722-10727, 2003. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.789, + 0.482, + 0.842 + ], + "angle": 0, + "content": "[39] Satoshi Tsutsui, Arjun Chandrasekaran, Md Alimoor Reza, David Crandall, and Chen Yu. A computational model of early word learning from the infant's point of view. CogSci, 2020. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.845, + 0.482, + 0.871 + ], + "angle": 0, + "content": "[40] Alan M Turing. Computing machinery and intelligence. Mind, LIX(236):433-460, 1950. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.874, + 0.482, + 0.901 + ], + "angle": 0, + "content": "[41] Wai Keen Vong, Wentao Wang, A Emin Orhan, and Brenden M Lake. Grounded language acquisition through the" + }, + { + "type": "list", + "bbox": [ + 0.093, + 0.093, + 0.482, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.548, + 0.093, + 0.905, + 0.12 + ], + "angle": 0, + "content": "eyes and ears of a single child. Science, 383(6682):504-511, 2024. 1, 2, 3, 5, 6, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.122, + 0.905, + 0.19 + ], + "angle": 0, + "content": "[42] Zeyu Wang, Berthy Feng, Karthik Narasimhan, and Olga Russakovsky. Towards unique and informative captioning of images. In Computer Vision-ECCV 2020: 16th European Conference, Glasgow, UK, August 23-28, 2020, Proceedings, Part VII 16, pages 629-644. Springer, 2020. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.193, + 0.905, + 0.26 + ], + "angle": 0, + "content": "[43] Saining Xie, Ross Girshick, Piotr Dólár, Zhuowen Tu, and Kaiming He. Aggregated residual transformations for deep neural networks. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 1492-1500, 2017. 5" + }, + { + "type": "list", + "bbox": [ + 0.517, + 0.093, + 0.905, + 0.26 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.516, + 0.956 + ], + "angle": 0, + "content": "4352" + } + ] +] \ No newline at end of file diff --git a/2025/Discovering Hidden Visual Concepts Beyond Linguistic Input in Infant Learning/50337d5c-a1b4-4b7f-9096-7ccfa12467ab_origin.pdf b/2025/Discovering Hidden Visual Concepts Beyond Linguistic Input in Infant Learning/50337d5c-a1b4-4b7f-9096-7ccfa12467ab_origin.pdf new file mode 100644 index 0000000000000000000000000000000000000000..701c57d6b391bb562623fa5db7c515623f6aa4ca --- /dev/null +++ b/2025/Discovering Hidden Visual Concepts Beyond Linguistic Input in Infant Learning/50337d5c-a1b4-4b7f-9096-7ccfa12467ab_origin.pdf @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:0ae787a67156362f62ee09320ac20b8f59d3b7550b23d5c565f7ab16794e9b28 +size 5181448 diff --git a/2025/Discovering Hidden Visual Concepts Beyond Linguistic Input in Infant Learning/full.md b/2025/Discovering Hidden Visual Concepts Beyond Linguistic Input in Infant Learning/full.md new file mode 100644 index 0000000000000000000000000000000000000000..3579dae8b1c26f41db251bf0e8dc420c5bdd437e --- /dev/null +++ b/2025/Discovering Hidden Visual Concepts Beyond Linguistic Input in Infant Learning/full.md @@ -0,0 +1,303 @@ +# Discovering Hidden Visual Concepts Beyond Linguistic Input in Infant Learning + +Xueyi Ke + +Satoshi Tsutsui + +Yayun Zhang + +Bihan Wen $^{1*}$ + +$^{1}$ Nanyang Technological University + +2The Max Planck Institute for Psycholinguistics + +![](images/104eed72a52bef9f3192e4c489bda704eb1b560438b594c9093bc850548dd7c8.jpg) +Figure 1. Inspired by the development of the infant visual system, CVCL [41] is trained on infant egocentric frames and transcribed parental speech (a) and demonstrates object recognition ability within the vocabulary provided by parental speech (c). However, infant visual development is not limited to parental guidance. Thus, we hypothesize that a computational model trained on an infant's daily experiences can similarly acquire visual concepts beyond its training parental speech. To explore this, we perform neuron labeling (b) to identify visual concept neurons, including concepts that were never mentioned in the parental speech vocabulary (e.g., "rug"). Based on discovered neurons, we show that the model can recognize objects beyond the training vocabulary, akin to early infant visual development (d). Some images are cited from prior work [41] as the original data is not accessible to us. + +![](images/2c3dd38a2a7a357b90f5463f15ec9af1d9071757f84369491f9cea9a5fc61baa.jpg) + +# Abstract + +Infants develop complex visual understanding rapidly, even preceding of the acquisition of linguistic skills. As computer vision seeks to replicate the human vision system, understanding infant visual development may offer valuable insights. In this paper, we present an interdisciplinary study exploring this question: can a computational model that imitates the infant learning process develop broader visual concepts that extend beyond the vocabulary it has heard, similar to how infants naturally learn? To investigate this, we analyze a recently published model in Science by Vong et al., which is trained on longitudinal, egocentric images of a single child paired with transcribed parental speech. We perform neuron labeling to identify visual concept neurons hidden in the model's internal representations. We then demonstrate that these neurons can + +recognize objects beyond the model's original vocabulary. Furthermore, we compare the differences in representation between infant models and those in modern computer vision models, such as CLIP and ImageNet pre-trained model. Ultimately, our work bridges cognitive science and computer vision by analyzing the internal representations of a computational model trained on an infant visual and linguistic inputs. Our code is available at https://github.com/Kexueyi/discover_infant.vis. + +# 1. Introduction + +Infants are remarkable learners, sparking interest across various academic disciplines. Computer vision is no exception, with researchers studying infant visual learning from various perspectives [1, 31, 32, 35, 41]. A recent milestone is Child's View for Contrastive Learning (CVCL) by Vong et al.[41], which trains a model from scratch on longitudinal + +egocentric videos of a single infant (6-25 months) paired with transcribed parental speech to learn visual-linguistic associations, similar to CLIP [34]. The resulting model develops object recognition abilities, aligning with developmental psychology findings that infants acquire object names by linking words to visual referents [3, 17]. In this paper, we analyze the internal visual representations of CVCL to better understand its mechanism, similar to how developmental studies observe infants' internal neurons [38]. + +In developmental psychology, research on real infants suggests that their object recognition is deeply influenced by their visual experiences with the world. A headcam study of 8.5 to 10.5-month-old infants [8] showed that the first nouns infants acquire often correspond to objects they see most frequently. While infants may also hear the names of these frequently seen objects, their early visual familiarity plays a critical role in object recognition. Visual understanding may develop before the learning of corresponding names, potentially facilitating the process of word acquisition [33]. For instance, newborn infants have shown an innate ability to recognize visual patterns [13], and their development of visual concepts often precedes the emergence of verbal thought [27]. + +Our Hypothesis: Based on these infant studies, we hypothesize that a computational model trained on an infant's daily experiences may similarly acquire visual concepts extending beyond its linguistic training data. While CVCL demonstrates visual-linguistic mappings of objects named in parental speech, its visual recognition capability should not be limited to parental vocabulary supervision. We conjecture that the vision encoder may develop the ability to recognize concepts beyond these linguistic training data, similar to how infants form object recognition before learning object names. + +To investigate this, we analyze CVCL's internal representations using network dissection [2, 30], or more intuitively, neuron labeling. Furthermore, we implement a neuron-based, training-free classification framework (NeuronClassifier) that leverages visual concepts identified via neuron labeling. Relying solely on these neurons, this approach not only achieves better recognition performance than originally reported [41], but also discovers internal visual concepts extending beyond the model's training vocabulary, supporting our aligned hypothesis. + +From a developmental psychology perspective, the visual concepts discovered beyond the training vocabulary tend to have a higher age of acquisition (AoA) [26] values (Figure 7). This results supports CVCL's ability to develop visual understanding that precedes explicit labeling, mirroring cognitive studies where infants develop pre-verbal visual concepts [27]. This likely reflects real-world learning dynamics: children first acquire labels for frequent, con + +crete concepts (captured in CVCL's training vocabulary), while visually grounded representations for unlabeled concepts still form earlier than their eventual linguistic acquisition. + +To further explore CVCL's internal representations in the context of computer vision, we compare its visual features with widely-used representations such as CLIP and ImageNet. While CVCL is trained on a unique infant dataset with limited exposure to diverse scenes and a much smaller dataset compared to CLIP, it exhibits similar low-level features in its early layers. However, its higher-level features in the final layer differ significantly. These differences also extend to the visual concept neurons in the model's deeper layers. + +# Contributions The key contributions of our work are: + +- We show that infant model have developed understanding beyond linguistic training inputs, by discovering visual concepts hidden in the model representations, aligning with existing infant studies. +- We demonstrate that the discovered visual concept neurons can improve object recognition performance by implementing a training-free framework (NeuronClassifier). +- We find that the infant model shares similar low-level representations with ImageNet or CLIP models but diverges in deeper layers due to a lack of diverse higher-level visual concepts. + +# 2. Related Work + +Learning from children. Modeling how children learn has long been a strategy for advancing artificial intelligence. Instead of directly replicating adult intelligence, Alan Turing suggested, "why not rather try to produce one which simulates the child's?" [40] Training models on egocentric videos or multimodal data captured from infant perspectives aligns with this idea. [1, 31, 32, 35, 39, 41], as these videos approximate the input available to human infants during development. Our study established on CVCL [41], which we explain in Section 2.1. + +Interpreting vision model representations. Our goal is to understand the model internal neurons trained on infant data. In this context, techniques for interpreting intermediate representations in deep neural networks are relevant. Beginning with Network Dissection [2], a method that quantifies alignment between hidden neurons and visual concepts, numerous studies have aimed to make black-box models more transparent. These methods enable compositional concept discovery [28], assignment of compositional concepts with statistical quantification [5], open-vocabulary neuron captioning [21], and the use of CLIP's rich embeddings for neuron-concept alignment [22, 30]. In addition to direct neuron dissection, other approaches analyze component functions by decomposing image repre + +sentations to reveal the role of attention heads within multimodal embedding spaces [15] or identifying neurons with similar functions across a diverse model zoo [11]. + +# 2.1. Preliminary: CVCL + +Training data. CVCL is trained on the SAYCam-S dataset [37], a longitudinal collection of egocentric recordings from a child aged 6 to 25 months, containing around 200 hours of video. To create meaningful image-text pairs for model training, transcripts were pre-processed to retain only child-directed utterances, excluding the child's own vocalizations. Frames were extracted to align with utterance timestamps. The resulting dataset comprises 600,285 frames paired with 37,500 transcribed utterances, forming a multimodal dataset that simulates the sparse and noisy real-world experiences from which children learn. + +Model architecture. Employing a self-supervised contrastive learning approach akin to CLIP [34], CVCL learns to align egocentric visual frames with transcribed parental speech. Co-occurring pairs are treated as positive examples, while non-co-occurring pairs serve as negatives. This method allows the model to develop multimodal representations without external labels, imitating a child's natural learning process. + +Evaluation. For evaluation, CVCL adopts a $n$ -way classification task (Figure 2) in which the model selects the most relevant visual reference from a set comprising one target image and $n - 1$ foil images. This approach is inspired by the intermodal preferential looking paradigm (IPLP) [16, 17], used in infant recognition studies to measure language comprehension through differential visual fixation. By aligning its visual and text encoders, CVCL achieves comparable in-domain test accuracy to models like CLIP. However, CVCL demonstrates relatively weak test performance on the Konkle object dataset [23], which includes naturalistic object categories on a white background, using only classes available in the training data. + +# 3. Method + +In this section, we describe how to explore neuron-level concepts and leveraging them in $n$ way classification tasks. We begin by using published neuron labeling techniques to discover visual concepts hidden within CVCL [41]. Then, we introduce our framework to utilize these labeled neuron concepts for $n$ -way classification. + +# 3.1. Neuron Labeling + +We follow CLIP-Dissect [30] for internal representation analysis due to its flexibility in concept sets and input image dataset. + +![](images/a87f51f0a332bfcb59757dca0a108bda47ef000039658f45f341b2c7b38877d5.jpg) +Figure 2. CVCL's $n$ -way evaluation [41] poses a classification task of choosing the given target object label ("apple") from $n = 4$ images, where only one of them contains the target object. The model feeds the target label into the text encoder and computes the pairwise similarities to each of the $n$ images, selecting the image with the highest similarity to demonstrate object recognition ability. However, this way of recognition limits its ability to the vocabularies in the text encoder. Our framework in Figure 4 overcomes this limitation. + +Preliminary: CILP-dissect. Given a neural network $f(x)$ , where $f$ takes an image $x$ as input and $x \in \mathcal{D}_{\mathrm{probe}}$ with $|\mathcal{D}_{\mathrm{probe}}| = N$ , and a concept set $S$ with $|S| = M$ . The algorithm computes the concept-activation matrix $P \in \mathbb{R}^{N \times M}$ . + +$$ +P _ {i, j} = I _ {i} \cdot T _ {j}, \tag {1} +$$ + +where $I_{i}$ and $T_{j}$ are the embeddings of the images and concepts, respectively. For each neuron $k \in K$ , where $K$ denotes the set of all neurons in the network, we summarize activations $A_{k}(x_{i})$ with a scalar function $g$ , producing an activation vector: + +$$ +q _ {k} = \left[ g \left(A _ {k} \left(x _ {1}\right)\right), \dots , g \left(A _ {k} \left(x _ {N}\right)\right) \right] ^ {\top} \in \mathbb {R} ^ {N}. \tag {2} +$$ + +The neuron is labeled with the concept that maximizes similarity: + +$$ +l _ {k} = \arg \max _ {m} \operatorname {s i m} \left(t _ {m}, q _ {k}; P\right), \tag {3} +$$ + +where $\operatorname{sim}(\cdot, \cdot)$ represents the similarity function (e.g. cosine similarity or Soft-WPMI [30, 42]), and $t_m$ denotes the most similar text concept. Collecting the label $l_k$ for each neuron, we define the label vector $\mathcal{L} = [l_1, l_2, \ldots, l_K]$ to represent the assigned concepts across the entire model. + +During this neuron labeling process, we aim to assign meaningful concepts to each neuron. We use CLIP-dissect because: + +- Concept set $S$ : Instead of allowing an infinite range of possible concepts for neuron labeling, using a fixed concept set narrows this process by constraining it to a limited selection of concepts. Including diverse concepts in this set enables us to identify neurons corresponding to these visual concepts. This setup ensures that each neuron is assigned a specific label, though some labels may be spurious, as illustrated in Figure 3. + +![](images/e4ac9913c0aff54801f05a019813b1b9dc17b7cc977dbd54941f57501925f334.jpg) +Figure 3. Spurious labeled neurons in infant model CVCL by CLIP-Dissect, showing top-3 activated Konkle dataset images. This indicates: (1) not all neurons interpretable with human semantic concepts [12]; (2) neuron-labeling can produce spurious labels. To further support our hypothesis, we implement NeuronClassifier. + +- Probing dataset $\mathcal{D}_{\mathrm{probe}}$ : The probing dataset allows neurons to activate specifically in response to the dataset of interest. For instance, when analyzing representations from an infant model, we use images from the Konkle object dataset, which better aligns with infant recognition than the ImageNet [10] validation set. Additionally, for classification tasks, generalization can be achieved by adaptively selecting the probing dataset to focus on relevant concepts within a given dataset. + +Extending beyond vocabulary. The flexibility of the concept set $S$ allowing us to discover visual concepts that the infant model has never encountered in its training data. In this process, CLIP serves as a well-pretrained miner, utilizing its rich image-text embeddings to identify concepts hidden within CVCL. By aligning the activations of CVCL's neurons with CLIP's embeddings, we discover meaningful hidden visual concepts within the infant model, even extending beyond the model's linguistic training data. This approach leverages the diverse vocabulary of the concept set and the rich embeddings of CLIP to reveal visual concepts embedded in CVCL's internal representations, each associated with corresponding neurons. + +# 3.2. Neuron-Based Classification + +How do we ensure that neurons representing concepts beyond the model's original vocabulary truly exist within the network? In this section, we propose NeuronClassifier, a training-free framework that leverages neuron activations to detect and validate such concepts. By discovering neurons with specific visual concept, we aim to confirm the presence of these latent, beyond-vocabulary neurons and use them to perform $n$ -way classification. The framework, illustrated in Figure 4, involves three main steps. + +Step 1: neuron labeling with concept set. Given an image encoder $f(x)$ , we labeled each neuron in the network using a concept set $S$ that contains (but is not limited to) all + +![](images/933a9a5319d3c4839ebd5e09ddbd5d3496a5173a40856fc83309dc4b55bb7a9c.jpg) +Figure 4. Our NeuronClassifier Overview. A training-free $n$ -way framework with three key steps: (1) Label all neurons in the network using a concept set that includes class labels and common words; (2) identify neurons associated with the target concept (e.g., "rug"); (3) Evaluate the activations of visual concept neuron across $n$ images (in this example, $n = 4$ ) and select the image with the highest activation as the most relevant to the target label. + +class labels relevant to the task. The dissection process can be expressed as a function: + +$$ +\mathcal {N} _ {s} = \text {N e u r o n L a b e l i n g} (f, \mathcal {S}), \tag {4} +$$ + +where $\mathcal{N}_s$ is the set of neurons labeled with concepts from $S$ . Each neuron $k \in \mathcal{N}_s$ is associated with a specific concept, such as "rug" or "calculator", based on its alignment with concept embeddings obtained during neuron labeling (e.g. similarity in CLIP-Dissect [30]). + +Step 2: identifying visual concept neurons. Given a target label $l_{k} \in \mathcal{L}$ , where $\mathcal{L}$ represents all neuron labels in the model, same as label vector in Section 3.1 (e.g., "rug"), we select the subset of neurons labeled with this concept, denoted as $\mathcal{N}_{l_k} \subset \mathcal{N}_{\mathcal{L}}$ . These neurons are responsible for encoding the target concept. + +To further refine the labeling and reduce spurious assignments, we select the most similar neuron from $\mathcal{N}_{l_k}$ to represent the target concept. The similarity measure varies based on the dissection method used. For example, Network Dissection [2] employs Intersection over Union (IoU) for similarity, while CLIP-Dissect [30] supports multiple similarity metrics: + +$$ +k ^ {*} = \arg \max _ {k \in \mathcal {N} _ {\mathbf {L}}} \operatorname {s i m} \left(t _ {l _ {k}}, q _ {k}; P\right), \tag {5} +$$ + +where $t_{l_k}$ is the embedding of the target label $l_k$ , $q_k$ is the neuron activation value, same in Section 3.1. This step ensures that the neuron most aligned with the concept is selected, minimizing the possibility of spurious labeling. + +For each selected neuron $k \in \mathcal{N}_{l_k}$ , its activation value on an input image $x_i$ is computed as + +$$ +q _ {k} \left(x _ {i}\right) = g \left(A _ {k} \left(x _ {i}\right)\right), \tag {6} +$$ + +where $A_{k}(x_{i})$ is the raw activation map, and $g(\cdot)$ is a summary function (e.g., spatial mean) that reduces it to a scalar representing the neuron's response strength. + +Step 3: selecting the most relevant image. Given $n$ candidate images $\{x_1, x_2, \ldots, x_n\}$ in an $n$ -way trial, we compute the activation values $q_k(x_i)$ for neuron $k^*$ with highest similarity across all images. The image with the highest activation is selected as the most relevant to the target concept: + +$$ +x ^ {*} = \arg \max _ {x _ {i}} q _ {k ^ {*}} \left(x _ {i}\right). \tag {7} +$$ + +In the example shown in Figure 4, the target concept is "rug", and we select the image with the highest activation from the four candidates as the closest match to the concept. + +# 4. Neuron-wise Representation Analysis + +In this section, we conduct a neuron-wise analysis of infant models and provide implementation details. First, we perform neuron labeling on the infant CVCL model. Then, we use our NeuronClassifier framework to leverage visual concept neurons identified through neuron labeling, resulting in better recognition performance than the original approach [41] while also discovering internal visual concepts beyond the model's training vocabulary. These results support our hypothesis. + +# 4.1. Setup + +Datasets. We use the Konkle object dataset [23], as introduced in Section 2.1. This dataset consists of 3,406 images, each featuring a single object on a clean white background, including 406 test items across 200 classes. Each trial comprises $n$ images: one target image and the remaining as foils, with foil images randomly sampled from classes other than the target class. For each class, we generate 5 trials, each containing $n$ images. Following the previous work on CVCL, we use $n = 4$ in our main experiments, with one target and three foils per trial. + +Neuron labeling. We utilize CLIP-Dissect [30] for neuron labeling, which assigns visual concepts to each neuron in the network. As we aim to perform classification on the Konkle object dataset, we use the same dataset as a part of $\mathcal{D}_{\mathrm{probe}}$ . Additionally, to avoid limiting the search space only around class names and ensure comprehensive neuron labeling, we employ a combined concept, consisting of three components: + +- SAYCam-S vocabulary: We clean the original vocabulary by removing noisy child speech and retaining meaningful words. +- Common English words: We chose the top 30,000 most common English words based on a 1-gram frequency analysis by Peter Norvig [29, 30]. +- Class in Konkle object dataset: All class labels from the Konkle object dataset are included. + +We combine these three sources, ensuring no duplicates, resulting in a final concept set containing 30,427 words. + +Models. Our primary focus is the CVCL-ResNeXt50 [41], trained on SAY-Cam-S [37] dataset. For comprehensive analysis, we apply our framework to the following models: + +- Infant models: CVCL is trained on unique infant data, using both egocentric frames and transcribed parent speech. We also take DINO-S-ResNeXt50 [31] as reference model compare recognition ability derived from same visual experience. It trained with DINO[6] self-supervised approach on the infant same dataset. +- Broadly-trained models: To establish an upper bound, we include CLIP [34] and ResNeXt50 [43], which are broadly trained on large scale Internet images. Although CLIP uses a ResNet50-based vision encoder rather than ResNeXt, we select CLIP-ResNet50 due to the architectural similarity between ResNeXt [43] and ResNet [20]. +- Randomized model: As a lower bound, we introduce a randomized version of the CVCL-ResNeXt50 model. In this setup, the convolution layer weights in the vision encoder are initialized using Kaiming Initialization [19]. + +# 4.2. Results + +In this section, we present our results and findings from applying neuron labeling and our NeuronClassifier framework to the infant CVCL model, compared with other reference models. We define two class types for our analysis: + +- In-vocabulary classes: Object classes present in the model's training linguistic input, which also appear in test object class and are detectable in internal representations. +- Out-of-vocabulary classes: Object classes not included in the training linguistic input but detected in test object class and internal representations. + +![](images/425151cb957ec1ad32b744485555faa6b42d59f1096e4fb78e3be27b764e78e0.jpg) +Figure 5. In-vocabulary and out-of-vocabulary relationships visualized using a Venn diagram. + +Our results demonstrate that the proposed framework effectively discovers meaningful neurons that represent concepts beyond the model's training vocabulary. These findings align with the cognitive perspective of vocabulary acquisition in infant development. We analyze the models' performance across different $n$ -way classification settings, showing that our method yields strong out-of-vocabulary classification performance while simultaneously improving in-vocabulary classification accuracy. + +![](images/ffe366a0f8ed821e4fce48429054d773f759c535135c1bfd371e6e2db9f5647d.jpg) +Figure 6. Class coverage in visual concept neurons. Percentage of Konkle object dataset [23] classes identified through neuron labeling. Broadly pre-trained models (CLIP, ResNeXt) achieve over $50\%$ coverage, while developmentally inspired infant models (CVCL [41], DINO [6]) show lower coverage. CVCL-Randomized provides a lower-bound comparison. + +Class coverage in visual concept neurons. How well do visual concept neurons identified through neuron labeling correspond to specific class names (e.g., "rug") rather than general descriptive attributes (e.g., "red"? Figure 6 shows the percentage of classes in the Konkle object dataset that are discovered in visual concept neurons from each model during the neuron labeling process. The results indicate that well-pretrained models, such as CLIP-ResNet50 and ResNeXt50, demonstrate broader class name coverage. While CVCL performs slightly weaker, it still maintains coverage of slightly less than $50\%$ . In contrast, CVCL-Randomized achieves only around $28\%$ coverage. This class coverage metric reflects the models' capacity to form class-corresponding meaningful representations during the neuron labeling process. + +Age of Acquisition (AoA) ratings. Age of acquisition (AoA) is used to indicate when, and in what sequence, words are learned, and it is often assessed through ratings or observations reported by adults. This indirect method generally correlates well with other metrics indicating when children acquire vocabulary. Previous developmental work has shown that infants' early visual familiarity with common objects helps with object recognition, which subsequently helps support the process of learning the names of those objects. We next examined how early words in our models are learned and whether there is an AoA difference between in-vocab and out-of-vocab words. + +We used AoA ratings from a dataset compiled by Kuperman, Stadthagen-Gonzalez, and Brysbaert [25], which includes norms for over 30,000 English words gathered via Amazon Mechanical Turk. Each participant estimated the age in years at which they believed they first understood each word, even if they did not actively use them. This dataset is comparable to previously reported AoA norms [36] gathered in laboratory settings. + +![](images/5bf13d735d09342d50dc52a7e8341121f4445cd0f761898d1ff137794446a05c.jpg) +Figure 7. Age of acquisition (AoA) ratings for in- and out-of-vocabulary visual concepts. Comparison of word acquisition timing [25] between in-vocabulary and out-of-vocabulary concepts. Out-of-vocabulary concepts tend to have a higher estimated acquisition ages in the infant-inspired CVCL model, indicating development of visual understanding beyond explicit linguist inputs. + +Using this set of AoA norms, we compared mean AoA between in-vocab and out-of-vocab words discovered in CVCL's internal representations. As shown in Figure 7, we found a significant difference between in-vocab and out-of-vocab AoA rating ( $t(82) = 4.64$ , $p < 0.0001$ ), in-vocab words (mean AoA = 4.99) are learned earlier than out-of-vocab words (mean AoA = 6.82). This pattern suggests that: (1) both sets of words are learned quite early, around later preschool and school years, with or without supervised labeling; (2) the difference in AoA between in-vocabulary and out-of-vocabulary words indicates that the infant model has developed a basic visual understanding of concepts with higher AoA. This foundational knowledge may lay the groundwork for word learning once corresponding parental speech is introduced. + +Neuron-based classification performance. We evaluated the hidden potential of infant models' vision encoders by applying our NeuronClassifier framework, summarized in Table 1. Despite being trained on infant egocentric data with limited amount and diversity, CVCL demonstrates the ability to recognize similarly as nature of infant learning, revealing strong out-of-vocabulary classification performance. This result suggests that this infant model developed broader visual concepts that extend beyond linguistic input, similar to how infants naturally learn. We also applied our method to in-vocabulary classification, where it outperformed the vanilla method previously used in CVCL [41], as introduced in Figure 2. "All" representing the combined performance on both "in-vocab" and "out-of-vocab". These results support our hypothesis. + +We include additional model comparisons: (1) The DINO-S-ResNeXt50 infant model [31], trained on the same dataset without text supervision, achieves comparable per + +Table 1. Neuron-based classification results in 4-way evaluation among models in Section 4.1. "Vanilla" refers to classification based on image-text similarity (Figure 2). "X" denotes cases where direct classification on the Konkle dataset [23] is not possible due to missing text encoder or need fine-tuning. By leveraging neurons discovered in the representation, NeuronClassifier enables broader recognition, particularly in CVCL (bolded for emphasis), achieving improved recognition in both in-vocabulary and out-of-vocabulary, supporting our hypothesis. "All" represents the combined performance on both in- and out-of-vocabulary. + +
MethodModelIn-vocabOut-of-vocabAll
VanillaCLIP-ResNet5098.81±0.1696.93±0.0697.42±0.05
ResNeXt50×××
CVCL-ResNeXt5036.18±0.91××
DINO-S-ResNeXt50×××
Neuron ClassifierCLIP-ResNet5091.59±0.5288.66±0.3589.79±0.38
ResNeXt5088.17±0.4593.28±0.3691.88±0.15
CVCL-ResNeXt5079.50±0.7876.81±0.3577.79±0.40
DINO-S-ResNeXt5077.53±0.2477.96±0.2777.65±0.21
+ +formance in visual representations. This suggests that models trained on identical data distributions with different self-supervised methods may yield similar representational outcomes. (2) Broadly-trained models such as ResNeXt50 and CLIP establish performance upper bounds. However, CLIP underperforms in the vanilla setting, relying exclusively on visual neurons without text encoder guidance. While this work does not aim to advance zero-shot learning methods, it reveals visual concepts in infant models that emerge independently of linguistic inputs. + +In classification on the Konkle object dataset, both DINO-S and ImageNet ResNeXt50 required fine-tuning for this task. Our framework enables neuron-based classification without downstream fine-tuning, and providing a training-free qualitative inspection of internal representations. + +Analysis across $n$ -way settings. We evaluate the models under various $n$ -way classification setups. Figure 8 illustrates the performance trends for in- and out-of-vocabulary class classification accuracy applying our NeuronClassifier. + +Our method not only enables out-of-vocabulary classification but also significantly improves in-vocab performance compared to previous results [41], further supporting the presence of beyond-vocabulary potential in infant models. However, due to limited class coverage, CVCL with our method can classify a maximum of 31 classes (see Appendix A.3). These findings support our hypothesis that the infant model has acquired visual concepts beyond its initial vocabulary. These results show that leveraging the model's internal representations for classification that go beyond its vocabulary is not only feasible but also robust across different $n$ -way settings. These findings support our hypothesis that the infant model has acquired visual concepts beyond + +![](images/42c71236126c2ca2c87f41d16d2a98c32bb84d27134c7a9a5a5694e72ecaa499.jpg) +Figure 8. In- and out-of-vocabulary class performance across $n$ -way settings using NeuronClassifier. (a) For out-of-vocabulary classification (left), our method enables classification without additional training. The infant model CVCL ( $\clubsuit$ ) maintains robust performance as $n$ increases. (b) For in-vocabulary classification (right), "ours" (NeuronClassifier) enhances CVCL recognition ability compared to the "vanilla" setting. However, CLIP ( $\clubsuit$ ) declines, as it relies solely on neurons without text encoder input. Random ones serve as lower bounds. + +![](images/34575b3d787034d64dcf712c3f7a84750558d28b1dd93679d3db2b0f4d1e2944.jpg) + +its initial vocabulary. + +However, as $n$ increases exponentially, performance gradually declines. This decline is reasonable, as the task becomes increasingly difficult by nature as $n$ increases. It may also be attributed to dimensionality reduction in activation maps, leading to coarser classification. CLIPRN50 and ResNeXt50 perform well with NeuronClassifier, though not as effectively as their direct or fine-tuned versions, since our method is designed to reveal latent concepts rather than to perform fully optimized classification. In in-vocab settings, the "vanilla" approach represents direct classification as shown in Figure 2. + +# 5. Layer-wise Representation Analysis + +How does the representation learned from infant data differ from other representations widely used in the computer vision community? To explore this, we perform a layer-wise representation analysis using Centered Kernel Alignment (CKA) [24]. We compute the similarity between the representations of the infant model(CVCL), ImageNet-pretrained, and CLIP. Additionally, we apply neuron labeling techniques from a layer-wise perspective to identify unique visual concepts discovered at each layer between ImageNet and infant models. + +Layer similarity analysis. CKA [24] applies HSIC [18] (see Appendix B) over a set of images to provide layer-wise similarity scores between the representations of different + +![](images/9ad3b396e855efda1aed02449395d95d28c514c860fa7bfcb0ceb0e97b2bc720.jpg) +Figure 9. CKA layer-wise similarity between CVCL and common models. Using the ImageNet validation set as input, CVCL (y-axis) exhibits similarity to CLIP-RN50 (x-axis, left) and ResNeXt50 (x-axis, right) in the shallow layers (lower-level features) but diverges significantly in the final layer (higher-level features). Notably, Layer 4 of CVCL shows very low similarity to all layers of both common models. + +neural networks. A higher CKA score indicates more similar representations between two models at the given layer. We use the ImageNet validation set [10] as input for the networks, and compute the CKA similarity between the infant model (CVCL), the ImageNet-pretrained ResNeXt50, and CLIP-ResNet50. The results are presented as matrices in Figure 9. The lower layers of CVCL exhibit greater similarity to larger-scale models than its deeper layers. In larger-scale models, shallow layers are known to capture low-complexity features, while deeper layers progressively specialize in capturing higher-level concepts [7, 14]. Therefore, CVCL – the model trained on infant data – successfully develops lower-level representations comparable to those in common pre-trained models. However, the divergence in deeper layers suggests a lack of the diverse higher-level representations typically observed in models trained on common datasets. + +Neuron-based analysis. To investigate the characteristics of each layer, we apply network dissection to identify neurons that are aligned with specific visual concepts. Essentially, this is an extension of the neuron labeling process, where the Broden [2] dataset provides category labels for each visual concept neuron. We count the number of unique visual concepts discovered in each category and perform layer-wise comparisons to gain a broader view of the differences between models trained from ImageNet and infant data. In Figure 10, we visualize the number of unique visual concept neurons across layers for each model. The results show that early layers in both models predominantly have neurons of low-level features like color, with minimal differences between models. However, as we move to deeper layers, higher-level concepts such as objects and scenes become more prominent, and the disparities between models become clearer. CVCL exhibits fewer unique visual con + +![](images/1d9df6f63c2f7251555648f664f46debbcd23323a72a3a93e1573922be641f1c.jpg) +Figure 10. Number of visual concepts in ImageNet and infant models. Concepts are categorized using Broden dataset [2]. Neurons in deeper model layers capture increasingly complex concepts. Early layers primarily detect lower-level features like color and texture, while higher-level concepts such as objects and scenes emerge in deeper layers. CVCL exhibits fewer visual concepts than the ImageNet model, especially for higher-level visual concepts (e.g., objects and scenes). + +cepts in these higher-level categories compared to ImageNet model. This finding aligns with the layer similarity analysis. + +# 6. Conclusion + +In this paper, we explored whether an infant model (CVCL), trained on infant egocentric video frames and linguistic inputs, can acquire broader visual concepts extending beyond its initial training vocabulary. By introducing NeuronClassifier, a training-free framework to discover and leverage visual concepts hidden in representations, we unlocked the CVCL visual encoder's ability to recognize out-of-vocabulary concepts, establishing its potential as a strong classifier. Our findings also reveal that while CVCL, trained on a unique infant dataset with limited exposure to diverse scenes, it representations capture low-level features similar to those in common pre-trained models, they diverge significantly in higher-level representations, contributing to the observed performance differences. + +Overall, our approach bridges cognitive science and computer vision, providing insights into how infant models develop visual concepts that precede linguistic inputs, aligning with the natural way infants explore the world through sight. + +Limitations and future work. Our study did not analyze the infant training data, as we were unable to access it due to strict access controls. Instead, we analyze models trained on infant egocentric data, finding developmental alignment with cognitive studies. However, we did not extend this analysis to adult egocentric data. While the study focuses on infant data for developmental process, the framework can be applied to adult models, which remains a direction for future research. + +# Acknowledgements + +This work was partially supported by the National Research Foundation Singapore Competitive Research Program (award number CRP29-2022-0003). We thank Wai Keen Vong for invaluable discussion and CVCL's pretrained weights. We thank Jingyi Lin for figure discussions. We thank the anonymous reviewers for their constructive feedback. + +# References + +[1] Sven Bambach, David Crandall, Linda Smith, and Chen Yu. Toddler-inspired visual object learning. In Neural Information Processing Systems, 2018. 1, 2 +[2] David Bau, Bolei Zhou, Aditya Khosla, Aude Oliva, and Antonio Torralba. Network dissection: Quantifying interpretability of deep visual representations. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 6541-6549, 2017. 2, 4, 8 +[3] Erika Bergelson and Daniel Swingley. At 6-9 months, human infants know the meanings of many common nouns. Proceedings of the National Academy of Sciences, 109(9): 3253-3258, 2012. 2 +[4] Marc Brysbaert and Boris New. Moving beyond kucera and francis: A critical evaluation of current word frequency norms and the introduction of a new and improved word frequency measure for american english. Behavior Research Methods, 41(4):977-990, 2009. 1 +[5] Kirill Bykov, Laura Kopf, Shinichi Nakajima, Marius Kloft, and Marina Höhne. Labeling neural representations with inverse recognition. Advances in Neural Information Processing Systems, 36, 2024. 2 +[6] Mathilde Caron, Hugo Touvron, Ishan Misra, Hervé Jégou, Julien Mairal, Piotr Bojanowski, and Armand Joulin. Emerging properties in self-supervised vision transformers. In Proceedings of the IEEE/CVF international conference on computer vision, pages 9650-9660, 2021. 5, 6 +[7] Yixiong Chen, Alan Yuille, and Zongwei Zhou. Which layer is learning faster? a systematic exploration of layerwise convergence rate for deep neural networks. In The Eleventh International Conference on Learning Representations, 2023. 8 +[8] Elizabeth M Clerkin, Elizabeth Hart, James M Rehg, Chen Yu, and Linda B Smith. Real-world visual statistics and infants' first-learned object names. Philosophical Transactions of the Royal Society B: Biological Sciences, 372(1711): 20160055, 2017. 2 +[9] Michael J Cortese and Maya M Khanna. Age of acquisition ratings for 3,000 monosyllabic words. Behavior Research Methods, 40(3):791-794, 2008. 1 +[10] Jia Deng, Wei Dong, Richard Socher, Li-Jia Li, Kai Li, and Li Fei-Fei. Imagenet: A large-scale hierarchical image database. In 2009 IEEE conference on computer vision and pattern recognition, pages 248-255. IEEE, 2009. 4, 8 +[11] Amir Dravid, Yossi Gandelsman, Alexei A Efros, and Assaf Shocher. Rosetta neurons: Mining the common units in a + +model zoo. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 1934-1943, 2023. 3 +[12] Nelson Elhage, Tristan Hume, Catherine Olsson, Nicholas Schiefer, Tom Henighan, Shauna Kravec, Zac Hatfield-Dodds, Robert Lasenby, Dawn Drain, Carol Chen, et al. Toy models of superposition. arXiv preprint arXiv:2209.10652, 2022.4 +[13] Robert L Fantz. Pattern vision in newborn infants. Science, 140(3564):296-297, 1963. 2 +[14] Thomas Fel, Louis Bethune, Andrew Kyle Lampinen, Thomas Serre, and Katherine Hermann. Understanding visual feature reliance through the lens of complexity. arXiv preprint arXiv:2407.06076, 2024. 8 +[15] Yossi Gandelsman, Alexei A Efros, and Jacob Steinhardt. Interpreting clip's image representation via text-based decomposition. arXiv preprint arXiv:2310.05916, 2023. 3 +[16] Roberta Michnick Golinkoff, Kathryn Hirsh-Pasek, Kathleen M Cauley, and Laura Gordon. The eyes have it: Lexical and syntactic comprehension in a new paradigm. Journal of child language, 14(1):23-45, 1987. 3 +[17] Roberta Michnick Golinkoff, Weiyi Ma, Lulu Song, and Kathy Hirsh-Pasek. Twenty-five years using the intermodal preferential looking paradigm to study language acquisition: What have we learned? Perspectives on Psychological Science, 8(3):316-339, 2013. 2, 3 +[18] Arthur Gretton, Ralf Herbrich, Alexander Smola, Olivier Bousquet, Bernhard Schölkopf, and Aapo Hyvarinen. Kernel methods for measuring independence. Journal of Machine Learning Research, 6(12), 2005. 7, 3 +[19] Kaiming He, Xiangyu Zhang, Shaoqing Ren, and Jian Sun. Delving deep into rectifiers: Surpassing human-level performance on imagenet classification. In Proceedings of the IEEE international conference on computer vision, pages 1026-1034, 2015. 5 +[20] Kaiming He, Xiangyu Zhang, Shaoqing Ren, and Jian Sun. Deep residual learning for image recognition. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 770-778, 2016. 5 +[21] Evan Hernandez, Sarah Schwettmann, David Bau, Teona Bagashvili, Antonio Torralba, and Jacob Andreas. Natural language descriptions of deep visual features. In International Conference on Learning Representations, 2021. 2 +[22] Neha Kalibhat, Shweta Bhardwaj, C Bayan Bruss, Hamed Firooz, Maziar Sanjabi, and Soheil Feizi. Identifying interpretable subspaces in image representations. In International Conference on Machine Learning, pages 15623-15638. PMLR, 2023. 2 +[23] Talia Konkle, Timothy F Brady, George A Alvarez, and Aude Oliva. Conceptual distinctiveness supports detailed visual long-term memory for real-world objects. Journal of experimental Psychology: general, 139(3):558, 2010. 3, 5, 6, 7, 1, 2 +[24] Simon Kornblith, Mohammad Norouzi, Honglak Lee, and Geoffrey Hinton. Similarity of neural network representations revisited. In International conference on machine learning, pages 3519-3529. PMLR, 2019. 7 + +[25] Victor Kuperman, Hans Stadthagen-Gonzalez, and Marc Brysbaert. Age-of-acquisition ratings for 30,000 english words. Behavior research methods, 44:978-990, 2012. 6 +[26] Victor Kuperman, Hans Stadthagen-Gonzalez, and Marc Brysbaert. Age-of-acquisition ratings for 30,000 english words. Behavior Research Methods, 44(4):978-990, 2012. 2, 1 +[27] Jean M Mandler. How to build a baby: II. conceptual primitives. Psychological review, 99(4):587, 1992. 2 +[28] Jesse Mu and Jacob Andreas. Compositional explanations of neurons. Advances in Neural Information Processing Systems, 33:17153-17163, 2020. 2 +[29] Peter Norvig. Natural language corpus data: Beautiful data, 2009. Accessed: 2024-10-27. 5 +[30] Tuomas Oikarinen and Tsui-Wei Weng. Clip-dissect: Automatic description of neuron representations in deep vision networks. arXiv preprint arXiv:2204.10965, 2022. 2, 3, 4, 5 +[31] A Emin Orhan and Brenden M Lake. Learning high-level visual representations from a child's perspective without strong inductive biases. Nature Machine Intelligence, 6(3):271-283, 2024. 1, 2, 5, 6 +[32] Emin Orhan, Vaibhav Gupta, and Brenden M Lake. Self-supervised learning through the eyes of a child. Advances in Neural Information Processing Systems, 33:9960-9971, 2020. 1, 2 +[33] Deborah A Phillips and Jack P Shonkoff. From neurons to neighborhoods: The science of early childhood development. 2000. 2 +[34] Alec Radford, Jong Wook Kim, Chris Hallacy, Aditya Ramesh, Gabriel Goh, Sandhini Agarwal, Girish Sastry, Amanda Askell, Pamela Mishkin, Jack Clark, et al. Learning transferable visual models from natural language supervision. In International conference on machine learning, pages 8748-8763. PMLR, 2021. 2, 3, 5 +[35] Saber Sheybani, Himanshu Hansaria, Justin Wood, Linda Smith, and Zoran Tiganj. Curriculum learning with infant egocentric videos. Advances in Neural Information Processing Systems, 36, 2024. 1, 2 +[36] Hans Stadthagen-Gonzalez and Colin J Davis. The bristol norms for age of acquisition, imageability, and familiarity. Behavior Research Methods, 38(4):598-605, 2006. 6, 1 +[37] Jessica Sullivan, Michelle Mei, Andrew Perfors, Erica Wojcik, and Michael C Frank. Saycam: A large, longitudinal audiovisual dataset recorded from the infant's perspective. Open mind, 5:20-29, 2021. 3, 5 +[38] Gentaro Taga, Kayo Asakawa, Atsushi Maki, Yukuo Konishi, and Hideaki Koizumi. Brain imaging in awake infants by near-infrared optical topography. Proceedings of the National Academy of Sciences, 100(19):10722-10727, 2003. 2 +[39] Satoshi Tsutsui, Arjun Chandrasekaran, Md Alimoor Reza, David Crandall, and Chen Yu. A computational model of early word learning from the infant's point of view. CogSci, 2020. 2 +[40] Alan M Turing. Computing machinery and intelligence. Mind, LIX(236):433-460, 1950. 2 +[41] Wai Keen Vong, Wentao Wang, A Emin Orhan, and Brenden M Lake. Grounded language acquisition through the + +eyes and ears of a single child. Science, 383(6682):504-511, 2024. 1, 2, 3, 5, 6, 7 +[42] Zeyu Wang, Berthy Feng, Karthik Narasimhan, and Olga Russakovsky. Towards unique and informative captioning of images. In Computer Vision-ECCV 2020: 16th European Conference, Glasgow, UK, August 23-28, 2020, Proceedings, Part VII 16, pages 629-644. Springer, 2020. 3 +[43] Saining Xie, Ross Girshick, Piotr Dólár, Zhuowen Tu, and Kaiming He. Aggregated residual transformations for deep neural networks. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 1492-1500, 2017. 5 \ No newline at end of file diff --git a/2025/Discovering Hidden Visual Concepts Beyond Linguistic Input in Infant Learning/images.zip b/2025/Discovering Hidden Visual Concepts Beyond Linguistic Input in Infant Learning/images.zip new file mode 100644 index 0000000000000000000000000000000000000000..39b222cf250e789d4e04d265f809fc5ece4d3487 --- /dev/null +++ b/2025/Discovering Hidden Visual Concepts Beyond Linguistic Input in Infant Learning/images.zip @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:ad253e1983be3b509fd46da520c266137e89aa4f33e945a7eb4be8bf29be6d0a +size 431879 diff --git a/2025/Discovering Hidden Visual Concepts Beyond Linguistic Input in Infant Learning/layout.json b/2025/Discovering Hidden Visual Concepts Beyond Linguistic Input in Infant Learning/layout.json new file mode 100644 index 0000000000000000000000000000000000000000..5ee678e11f4e6825d16182bba27b61afd4580f0a --- /dev/null +++ b/2025/Discovering Hidden Visual Concepts Beyond Linguistic Input in Infant Learning/layout.json @@ -0,0 +1,8023 @@ +{ + "pdf_info": [ + { + "para_blocks": [ + { + "bbox": [ + 56, + 103, + 555, + 122 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 103, + 555, + 122 + ], + "spans": [ + { + "bbox": [ + 56, + 103, + 555, + 122 + ], + "type": "text", + "content": "Discovering Hidden Visual Concepts Beyond Linguistic Input in Infant Learning" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 144, + 141, + 197, + 158 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 144, + 141, + 197, + 158 + ], + "spans": [ + { + "bbox": [ + 144, + 141, + 197, + 158 + ], + "type": "text", + "content": "Xueyi Ke" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 216, + 143, + 296, + 157 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 216, + 143, + 296, + 157 + ], + "spans": [ + { + "bbox": [ + 216, + 143, + 296, + 157 + ], + "type": "text", + "content": "Satoshi Tsutsui" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 315, + 143, + 386, + 157 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 315, + 143, + 386, + 157 + ], + "spans": [ + { + "bbox": [ + 315, + 143, + 386, + 157 + ], + "type": "text", + "content": "Yayun Zhang" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 405, + 143, + 468, + 156 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 405, + 143, + 468, + 156 + ], + "spans": [ + { + "bbox": [ + 405, + 143, + 468, + 156 + ], + "type": "text", + "content": "Bihan Wen" + }, + { + "bbox": [ + 405, + 143, + 468, + 156 + ], + "type": "inline_equation", + "content": "^{1*}" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 219, + 163, + 392, + 177 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 219, + 163, + 392, + 177 + ], + "spans": [ + { + "bbox": [ + 219, + 163, + 392, + 177 + ], + "type": "inline_equation", + "content": "^{1}" + }, + { + "bbox": [ + 219, + 163, + 392, + 177 + ], + "type": "text", + "content": "Nanyang Technological University" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 188, + 178, + 419, + 191 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 188, + 178, + 419, + 191 + ], + "spans": [ + { + "bbox": [ + 188, + 178, + 419, + 191 + ], + "type": "text", + "content": "2The Max Planck Institute for Psycholinguistics" + } + ] + } + ], + "index": 8 + }, + { + "type": "image", + "bbox": [ + 70, + 237, + 301, + 413 + ], + "blocks": [ + { + "bbox": [ + 70, + 237, + 301, + 413 + ], + "lines": [ + { + "bbox": [ + 70, + 237, + 301, + 413 + ], + "spans": [ + { + "bbox": [ + 70, + 237, + 301, + 413 + ], + "type": "image", + "image_path": "104eed72a52bef9f3192e4c489bda704eb1b560438b594c9093bc850548dd7c8.jpg" + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 55, + 413, + 555, + 472 + ], + "lines": [ + { + "bbox": [ + 55, + 413, + 555, + 472 + ], + "spans": [ + { + "bbox": [ + 55, + 413, + 555, + 472 + ], + "type": "text", + "content": "Figure 1. Inspired by the development of the infant visual system, CVCL [41] is trained on infant egocentric frames and transcribed parental speech (a) and demonstrates object recognition ability within the vocabulary provided by parental speech (c). However, infant visual development is not limited to parental guidance. Thus, we hypothesize that a computational model trained on an infant's daily experiences can similarly acquire visual concepts beyond its training parental speech. To explore this, we perform neuron labeling (b) to identify visual concept neurons, including concepts that were never mentioned in the parental speech vocabulary (e.g., \"rug\"). Based on discovered neurons, we show that the model can recognize objects beyond the training vocabulary, akin to early infant visual development (d). Some images are cited from prior work [41] as the original data is not accessible to us." + } + ] + } + ], + "index": 11, + "angle": 0, + "type": "image_caption" + } + ], + "index": 9 + }, + { + "type": "image", + "bbox": [ + 306, + 237, + 542, + 413 + ], + "blocks": [ + { + "bbox": [ + 306, + 237, + 542, + 413 + ], + "lines": [ + { + "bbox": [ + 306, + 237, + 542, + 413 + ], + "spans": [ + { + "bbox": [ + 306, + 237, + 542, + 413 + ], + "type": "image", + "image_path": "2c3dd38a2a7a357b90f5463f15ec9af1d9071757f84369491f9cea9a5fc61baa.jpg" + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "image_body" + } + ], + "index": 10 + }, + { + "bbox": [ + 151, + 486, + 200, + 498 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 151, + 486, + 200, + 498 + ], + "spans": [ + { + "bbox": [ + 151, + 486, + 200, + 498 + ], + "type": "text", + "content": "Abstract" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 54, + 513, + 297, + 693 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 54, + 513, + 297, + 693 + ], + "spans": [ + { + "bbox": [ + 54, + 513, + 297, + 693 + ], + "type": "text", + "content": "Infants develop complex visual understanding rapidly, even preceding of the acquisition of linguistic skills. As computer vision seeks to replicate the human vision system, understanding infant visual development may offer valuable insights. In this paper, we present an interdisciplinary study exploring this question: can a computational model that imitates the infant learning process develop broader visual concepts that extend beyond the vocabulary it has heard, similar to how infants naturally learn? To investigate this, we analyze a recently published model in Science by Vong et al., which is trained on longitudinal, egocentric images of a single child paired with transcribed parental speech. We perform neuron labeling to identify visual concept neurons hidden in the model's internal representations. We then demonstrate that these neurons can" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 313, + 487, + 556, + 596 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 487, + 556, + 596 + ], + "spans": [ + { + "bbox": [ + 313, + 487, + 556, + 596 + ], + "type": "text", + "content": "recognize objects beyond the model's original vocabulary. Furthermore, we compare the differences in representation between infant models and those in modern computer vision models, such as CLIP and ImageNet pre-trained model. Ultimately, our work bridges cognitive science and computer vision by analyzing the internal representations of a computational model trained on an infant visual and linguistic inputs. Our code is available at https://github.com/Kexueyi/discover_infant.vis." + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 314, + 621, + 395, + 634 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 314, + 621, + 395, + 634 + ], + "spans": [ + { + "bbox": [ + 314, + 621, + 395, + 634 + ], + "type": "text", + "content": "1. Introduction" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 313, + 643, + 556, + 717 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 643, + 556, + 717 + ], + "spans": [ + { + "bbox": [ + 313, + 643, + 556, + 717 + ], + "type": "text", + "content": "Infants are remarkable learners, sparking interest across various academic disciplines. Computer vision is no exception, with researchers studying infant visual learning from various perspectives [1, 31, 32, 35, 41]. A recent milestone is Child's View for Contrastive Learning (CVCL) by Vong et al.[41], which trains a model from scratch on longitudinal" + } + ] + } + ], + "index": 16 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "spans": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "type": "text", + "content": "CVF" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 145, + 0, + 496, + 37 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 145, + 0, + 496, + 37 + ], + "spans": [ + { + "bbox": [ + 145, + 0, + 496, + 37 + ], + "type": "text", + "content": "This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 66, + 704, + 147, + 714 + ], + "type": "page_footnote", + "angle": 0, + "lines": [ + { + "bbox": [ + 66, + 704, + 147, + 714 + ], + "spans": [ + { + "bbox": [ + 66, + 704, + 147, + 714 + ], + "type": "text", + "content": "*Corresponding Author." + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "text", + "content": "4343" + } + ] + } + ], + "index": 18 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 0 + }, + { + "para_blocks": [ + { + "bbox": [ + 55, + 72, + 294, + 191 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 72, + 294, + 191 + ], + "spans": [ + { + "bbox": [ + 55, + 72, + 294, + 191 + ], + "type": "text", + "content": "egocentric videos of a single infant (6-25 months) paired with transcribed parental speech to learn visual-linguistic associations, similar to CLIP [34]. The resulting model develops object recognition abilities, aligning with developmental psychology findings that infants acquire object names by linking words to visual referents [3, 17]. In this paper, we analyze the internal visual representations of CVCL to better understand its mechanism, similar to how developmental studies observe infants' internal neurons [38]." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 55, + 194, + 295, + 361 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 194, + 295, + 361 + ], + "spans": [ + { + "bbox": [ + 55, + 194, + 295, + 361 + ], + "type": "text", + "content": "In developmental psychology, research on real infants suggests that their object recognition is deeply influenced by their visual experiences with the world. A headcam study of 8.5 to 10.5-month-old infants [8] showed that the first nouns infants acquire often correspond to objects they see most frequently. While infants may also hear the names of these frequently seen objects, their early visual familiarity plays a critical role in object recognition. Visual understanding may develop before the learning of corresponding names, potentially facilitating the process of word acquisition [33]. For instance, newborn infants have shown an innate ability to recognize visual patterns [13], and their development of visual concepts often precedes the emergence of verbal thought [27]." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 55, + 363, + 295, + 495 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 363, + 295, + 495 + ], + "spans": [ + { + "bbox": [ + 55, + 363, + 295, + 495 + ], + "type": "text", + "content": "Our Hypothesis: Based on these infant studies, we hypothesize that a computational model trained on an infant's daily experiences may similarly acquire visual concepts extending beyond its linguistic training data. While CVCL demonstrates visual-linguistic mappings of objects named in parental speech, its visual recognition capability should not be limited to parental vocabulary supervision. We conjecture that the vision encoder may develop the ability to recognize concepts beyond these linguistic training data, similar to how infants form object recognition before learning object names." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 55, + 496, + 295, + 616 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 496, + 295, + 616 + ], + "spans": [ + { + "bbox": [ + 55, + 496, + 295, + 616 + ], + "type": "text", + "content": "To investigate this, we analyze CVCL's internal representations using network dissection [2, 30], or more intuitively, neuron labeling. Furthermore, we implement a neuron-based, training-free classification framework (NeuronClassifier) that leverages visual concepts identified via neuron labeling. Relying solely on these neurons, this approach not only achieves better recognition performance than originally reported [41], but also discovers internal visual concepts extending beyond the model's training vocabulary, supporting our aligned hypothesis." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 55, + 618, + 295, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 618, + 295, + 715 + ], + "spans": [ + { + "bbox": [ + 55, + 618, + 295, + 715 + ], + "type": "text", + "content": "From a developmental psychology perspective, the visual concepts discovered beyond the training vocabulary tend to have a higher age of acquisition (AoA) [26] values (Figure 7). This results supports CVCL's ability to develop visual understanding that precedes explicit labeling, mirroring cognitive studies where infants develop pre-verbal visual concepts [27]. This likely reflects real-world learning dynamics: children first acquire labels for frequent, con" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 313, + 72, + 553, + 119 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 72, + 553, + 119 + ], + "spans": [ + { + "bbox": [ + 313, + 72, + 553, + 119 + ], + "type": "text", + "content": "crete concepts (captured in CVCL's training vocabulary), while visually grounded representations for unlabeled concepts still form earlier than their eventual linguistic acquisition." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 313, + 120, + 555, + 240 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 120, + 555, + 240 + ], + "spans": [ + { + "bbox": [ + 313, + 120, + 555, + 240 + ], + "type": "text", + "content": "To further explore CVCL's internal representations in the context of computer vision, we compare its visual features with widely-used representations such as CLIP and ImageNet. While CVCL is trained on a unique infant dataset with limited exposure to diverse scenes and a much smaller dataset compared to CLIP, it exhibits similar low-level features in its early layers. However, its higher-level features in the final layer differ significantly. These differences also extend to the visual concept neurons in the model's deeper layers." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 314, + 243, + 534, + 255 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 314, + 243, + 534, + 255 + ], + "spans": [ + { + "bbox": [ + 314, + 243, + 534, + 255 + ], + "type": "text", + "content": "Contributions The key contributions of our work are:" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 315, + 258, + 553, + 389 + ], + "type": "list", + "angle": 0, + "index": 11, + "blocks": [ + { + "bbox": [ + 315, + 258, + 553, + 305 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 315, + 258, + 553, + 305 + ], + "spans": [ + { + "bbox": [ + 315, + 258, + 553, + 305 + ], + "type": "text", + "content": "- We show that infant model have developed understanding beyond linguistic training inputs, by discovering visual concepts hidden in the model representations, aligning with existing infant studies." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 315, + 306, + 553, + 342 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 315, + 306, + 553, + 342 + ], + "spans": [ + { + "bbox": [ + 315, + 306, + 553, + 342 + ], + "type": "text", + "content": "- We demonstrate that the discovered visual concept neurons can improve object recognition performance by implementing a training-free framework (NeuronClassifier)." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 315, + 342, + 553, + 389 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 315, + 342, + 553, + 389 + ], + "spans": [ + { + "bbox": [ + 315, + 342, + 553, + 389 + ], + "type": "text", + "content": "- We find that the infant model shares similar low-level representations with ImageNet or CLIP models but diverges in deeper layers due to a lack of diverse higher-level visual concepts." + } + ] + } + ], + "index": 10 + } + ], + "sub_type": "text" + }, + { + "bbox": [ + 314, + 401, + 400, + 414 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 314, + 401, + 400, + 414 + ], + "spans": [ + { + "bbox": [ + 314, + 401, + 400, + 414 + ], + "type": "text", + "content": "2. Related Work" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 313, + 422, + 555, + 541 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 422, + 555, + 541 + ], + "spans": [ + { + "bbox": [ + 313, + 422, + 555, + 541 + ], + "type": "text", + "content": "Learning from children. Modeling how children learn has long been a strategy for advancing artificial intelligence. Instead of directly replicating adult intelligence, Alan Turing suggested, \"why not rather try to produce one which simulates the child's?\" [40] Training models on egocentric videos or multimodal data captured from infant perspectives aligns with this idea. [1, 31, 32, 35, 39, 41], as these videos approximate the input available to human infants during development. Our study established on CVCL [41], which we explain in Section 2.1." + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 313, + 546, + 556, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 546, + 556, + 714 + ], + "spans": [ + { + "bbox": [ + 313, + 546, + 556, + 714 + ], + "type": "text", + "content": "Interpreting vision model representations. Our goal is to understand the model internal neurons trained on infant data. In this context, techniques for interpreting intermediate representations in deep neural networks are relevant. Beginning with Network Dissection [2], a method that quantifies alignment between hidden neurons and visual concepts, numerous studies have aimed to make black-box models more transparent. These methods enable compositional concept discovery [28], assignment of compositional concepts with statistical quantification [5], open-vocabulary neuron captioning [21], and the use of CLIP's rich embeddings for neuron-concept alignment [22, 30]. In addition to direct neuron dissection, other approaches analyze component functions by decomposing image repre" + } + ] + } + ], + "index": 14 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "text", + "content": "4344" + } + ] + } + ], + "index": 15 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 1 + }, + { + "para_blocks": [ + { + "bbox": [ + 55, + 72, + 295, + 109 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 72, + 295, + 109 + ], + "spans": [ + { + "bbox": [ + 55, + 72, + 295, + 109 + ], + "type": "text", + "content": "sentations to reveal the role of attention heads within multimodal embedding spaces [15] or identifying neurons with similar functions across a diverse model zoo [11]." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 55, + 119, + 174, + 132 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 119, + 174, + 132 + ], + "spans": [ + { + "bbox": [ + 55, + 119, + 174, + 132 + ], + "type": "text", + "content": "2.1. Preliminary: CVCL" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 55, + 137, + 296, + 269 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 137, + 296, + 269 + ], + "spans": [ + { + "bbox": [ + 55, + 137, + 296, + 269 + ], + "type": "text", + "content": "Training data. CVCL is trained on the SAYCam-S dataset [37], a longitudinal collection of egocentric recordings from a child aged 6 to 25 months, containing around 200 hours of video. To create meaningful image-text pairs for model training, transcripts were pre-processed to retain only child-directed utterances, excluding the child's own vocalizations. Frames were extracted to align with utterance timestamps. The resulting dataset comprises 600,285 frames paired with 37,500 transcribed utterances, forming a multimodal dataset that simulates the sparse and noisy real-world experiences from which children learn." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 55, + 279, + 296, + 375 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 279, + 296, + 375 + ], + "spans": [ + { + "bbox": [ + 55, + 279, + 296, + 375 + ], + "type": "text", + "content": "Model architecture. Employing a self-supervised contrastive learning approach akin to CLIP [34], CVCL learns to align egocentric visual frames with transcribed parental speech. Co-occurring pairs are treated as positive examples, while non-co-occurring pairs serve as negatives. This method allows the model to develop multimodal representations without external labels, imitating a child's natural learning process." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 55, + 384, + 296, + 540 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 384, + 296, + 540 + ], + "spans": [ + { + "bbox": [ + 55, + 384, + 296, + 540 + ], + "type": "text", + "content": "Evaluation. For evaluation, CVCL adopts a " + }, + { + "bbox": [ + 55, + 384, + 296, + 540 + ], + "type": "inline_equation", + "content": "n" + }, + { + "bbox": [ + 55, + 384, + 296, + 540 + ], + "type": "text", + "content": "-way classification task (Figure 2) in which the model selects the most relevant visual reference from a set comprising one target image and " + }, + { + "bbox": [ + 55, + 384, + 296, + 540 + ], + "type": "inline_equation", + "content": "n - 1" + }, + { + "bbox": [ + 55, + 384, + 296, + 540 + ], + "type": "text", + "content": " foil images. This approach is inspired by the intermodal preferential looking paradigm (IPLP) [16, 17], used in infant recognition studies to measure language comprehension through differential visual fixation. By aligning its visual and text encoders, CVCL achieves comparable in-domain test accuracy to models like CLIP. However, CVCL demonstrates relatively weak test performance on the Konkle object dataset [23], which includes naturalistic object categories on a white background, using only classes available in the training data." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 55, + 555, + 111, + 567 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 555, + 111, + 567 + ], + "spans": [ + { + "bbox": [ + 55, + 555, + 111, + 567 + ], + "type": "text", + "content": "3. Method" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 55, + 576, + 296, + 648 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 576, + 296, + 648 + ], + "spans": [ + { + "bbox": [ + 55, + 576, + 296, + 648 + ], + "type": "text", + "content": "In this section, we describe how to explore neuron-level concepts and leveraging them in " + }, + { + "bbox": [ + 55, + 576, + 296, + 648 + ], + "type": "inline_equation", + "content": "n" + }, + { + "bbox": [ + 55, + 576, + 296, + 648 + ], + "type": "text", + "content": " way classification tasks. We begin by using published neuron labeling techniques to discover visual concepts hidden within CVCL [41]. Then, we introduce our framework to utilize these labeled neuron concepts for " + }, + { + "bbox": [ + 55, + 576, + 296, + 648 + ], + "type": "inline_equation", + "content": "n" + }, + { + "bbox": [ + 55, + 576, + 296, + 648 + ], + "type": "text", + "content": "-way classification." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 55, + 658, + 159, + 672 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 658, + 159, + 672 + ], + "spans": [ + { + "bbox": [ + 55, + 658, + 159, + 672 + ], + "type": "text", + "content": "3.1. Neuron Labeling" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 55, + 677, + 296, + 712 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 677, + 296, + 712 + ], + "spans": [ + { + "bbox": [ + 55, + 677, + 296, + 712 + ], + "type": "text", + "content": "We follow CLIP-Dissect [30] for internal representation analysis due to its flexibility in concept sets and input image dataset." + } + ] + } + ], + "index": 8 + }, + { + "type": "image", + "bbox": [ + 318, + 71, + 553, + 178 + ], + "blocks": [ + { + "bbox": [ + 318, + 71, + 553, + 178 + ], + "lines": [ + { + "bbox": [ + 318, + 71, + 553, + 178 + ], + "spans": [ + { + "bbox": [ + 318, + 71, + 553, + 178 + ], + "type": "image", + "image_path": "a87f51f0a332bfcb59757dca0a108bda47ef000039658f45f341b2c7b38877d5.jpg" + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 313, + 186, + 555, + 285 + ], + "lines": [ + { + "bbox": [ + 313, + 186, + 555, + 285 + ], + "spans": [ + { + "bbox": [ + 313, + 186, + 555, + 285 + ], + "type": "text", + "content": "Figure 2. CVCL's " + }, + { + "bbox": [ + 313, + 186, + 555, + 285 + ], + "type": "inline_equation", + "content": "n" + }, + { + "bbox": [ + 313, + 186, + 555, + 285 + ], + "type": "text", + "content": "-way evaluation [41] poses a classification task of choosing the given target object label (\"apple\") from " + }, + { + "bbox": [ + 313, + 186, + 555, + 285 + ], + "type": "inline_equation", + "content": "n = 4" + }, + { + "bbox": [ + 313, + 186, + 555, + 285 + ], + "type": "text", + "content": " images, where only one of them contains the target object. The model feeds the target label into the text encoder and computes the pairwise similarities to each of the " + }, + { + "bbox": [ + 313, + 186, + 555, + 285 + ], + "type": "inline_equation", + "content": "n" + }, + { + "bbox": [ + 313, + 186, + 555, + 285 + ], + "type": "text", + "content": " images, selecting the image with the highest similarity to demonstrate object recognition ability. However, this way of recognition limits its ability to the vocabularies in the text encoder. Our framework in Figure 4 overcomes this limitation." + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "image_caption" + } + ], + "index": 9 + }, + { + "bbox": [ + 313, + 307, + 554, + 355 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 307, + 554, + 355 + ], + "spans": [ + { + "bbox": [ + 313, + 307, + 554, + 355 + ], + "type": "text", + "content": "Preliminary: CILP-dissect. Given a neural network " + }, + { + "bbox": [ + 313, + 307, + 554, + 355 + ], + "type": "inline_equation", + "content": "f(x)" + }, + { + "bbox": [ + 313, + 307, + 554, + 355 + ], + "type": "text", + "content": ", where " + }, + { + "bbox": [ + 313, + 307, + 554, + 355 + ], + "type": "inline_equation", + "content": "f" + }, + { + "bbox": [ + 313, + 307, + 554, + 355 + ], + "type": "text", + "content": " takes an image " + }, + { + "bbox": [ + 313, + 307, + 554, + 355 + ], + "type": "inline_equation", + "content": "x" + }, + { + "bbox": [ + 313, + 307, + 554, + 355 + ], + "type": "text", + "content": " as input and " + }, + { + "bbox": [ + 313, + 307, + 554, + 355 + ], + "type": "inline_equation", + "content": "x \\in \\mathcal{D}_{\\mathrm{probe}}" + }, + { + "bbox": [ + 313, + 307, + 554, + 355 + ], + "type": "text", + "content": " with " + }, + { + "bbox": [ + 313, + 307, + 554, + 355 + ], + "type": "inline_equation", + "content": "|\\mathcal{D}_{\\mathrm{probe}}| = N" + }, + { + "bbox": [ + 313, + 307, + 554, + 355 + ], + "type": "text", + "content": ", and a concept set " + }, + { + "bbox": [ + 313, + 307, + 554, + 355 + ], + "type": "inline_equation", + "content": "S" + }, + { + "bbox": [ + 313, + 307, + 554, + 355 + ], + "type": "text", + "content": " with " + }, + { + "bbox": [ + 313, + 307, + 554, + 355 + ], + "type": "inline_equation", + "content": "|S| = M" + }, + { + "bbox": [ + 313, + 307, + 554, + 355 + ], + "type": "text", + "content": ". The algorithm computes the concept-activation matrix " + }, + { + "bbox": [ + 313, + 307, + 554, + 355 + ], + "type": "inline_equation", + "content": "P \\in \\mathbb{R}^{N \\times M}" + }, + { + "bbox": [ + 313, + 307, + 554, + 355 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 404, + 366, + 553, + 380 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 404, + 366, + 553, + 380 + ], + "spans": [ + { + "bbox": [ + 404, + 366, + 553, + 380 + ], + "type": "interline_equation", + "content": "P _ {i, j} = I _ {i} \\cdot T _ {j}, \\tag {1}", + "image_path": "da7e5c2eb631362bfd553d8c056fb28adaeb66230109ac25f0c3bd8ac4ba828a.jpg" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 313, + 384, + 554, + 444 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 384, + 554, + 444 + ], + "spans": [ + { + "bbox": [ + 313, + 384, + 554, + 444 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 313, + 384, + 554, + 444 + ], + "type": "inline_equation", + "content": "I_{i}" + }, + { + "bbox": [ + 313, + 384, + 554, + 444 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 313, + 384, + 554, + 444 + ], + "type": "inline_equation", + "content": "T_{j}" + }, + { + "bbox": [ + 313, + 384, + 554, + 444 + ], + "type": "text", + "content": " are the embeddings of the images and concepts, respectively. For each neuron " + }, + { + "bbox": [ + 313, + 384, + 554, + 444 + ], + "type": "inline_equation", + "content": "k \\in K" + }, + { + "bbox": [ + 313, + 384, + 554, + 444 + ], + "type": "text", + "content": ", where " + }, + { + "bbox": [ + 313, + 384, + 554, + 444 + ], + "type": "inline_equation", + "content": "K" + }, + { + "bbox": [ + 313, + 384, + 554, + 444 + ], + "type": "text", + "content": " denotes the set of all neurons in the network, we summarize activations " + }, + { + "bbox": [ + 313, + 384, + 554, + 444 + ], + "type": "inline_equation", + "content": "A_{k}(x_{i})" + }, + { + "bbox": [ + 313, + 384, + 554, + 444 + ], + "type": "text", + "content": " with a scalar function " + }, + { + "bbox": [ + 313, + 384, + 554, + 444 + ], + "type": "inline_equation", + "content": "g" + }, + { + "bbox": [ + 313, + 384, + 554, + 444 + ], + "type": "text", + "content": ", producing an activation vector:" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 343, + 453, + 553, + 468 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 343, + 453, + 553, + 468 + ], + "spans": [ + { + "bbox": [ + 343, + 453, + 553, + 468 + ], + "type": "interline_equation", + "content": "q _ {k} = \\left[ g \\left(A _ {k} \\left(x _ {1}\\right)\\right), \\dots , g \\left(A _ {k} \\left(x _ {N}\\right)\\right) \\right] ^ {\\top} \\in \\mathbb {R} ^ {N}. \\tag {2}", + "image_path": "ebc1eb544fea3acc20b8cfad305b4740447039877ed768be4402faabae9e1391.jpg" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 313, + 476, + 553, + 500 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 476, + 553, + 500 + ], + "spans": [ + { + "bbox": [ + 313, + 476, + 553, + 500 + ], + "type": "text", + "content": "The neuron is labeled with the concept that maximizes similarity:" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 370, + 500, + 553, + 517 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 370, + 500, + 553, + 517 + ], + "spans": [ + { + "bbox": [ + 370, + 500, + 553, + 517 + ], + "type": "interline_equation", + "content": "l _ {k} = \\arg \\max _ {m} \\operatorname {s i m} \\left(t _ {m}, q _ {k}; P\\right), \\tag {3}", + "image_path": "6be1e23367590c9a3735269098170ca08e1f618a46f54cc155cb75f0c822556c.jpg" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 313, + 522, + 554, + 582 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 522, + 554, + 582 + ], + "spans": [ + { + "bbox": [ + 313, + 522, + 554, + 582 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 313, + 522, + 554, + 582 + ], + "type": "inline_equation", + "content": "\\operatorname{sim}(\\cdot, \\cdot)" + }, + { + "bbox": [ + 313, + 522, + 554, + 582 + ], + "type": "text", + "content": " represents the similarity function (e.g. cosine similarity or Soft-WPMI [30, 42]), and " + }, + { + "bbox": [ + 313, + 522, + 554, + 582 + ], + "type": "inline_equation", + "content": "t_m" + }, + { + "bbox": [ + 313, + 522, + 554, + 582 + ], + "type": "text", + "content": " denotes the most similar text concept. Collecting the label " + }, + { + "bbox": [ + 313, + 522, + 554, + 582 + ], + "type": "inline_equation", + "content": "l_k" + }, + { + "bbox": [ + 313, + 522, + 554, + 582 + ], + "type": "text", + "content": " for each neuron, we define the label vector " + }, + { + "bbox": [ + 313, + 522, + 554, + 582 + ], + "type": "inline_equation", + "content": "\\mathcal{L} = [l_1, l_2, \\ldots, l_K]" + }, + { + "bbox": [ + 313, + 522, + 554, + 582 + ], + "type": "text", + "content": " to represent the assigned concepts across the entire model." + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 313, + 582, + 554, + 616 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 582, + 554, + 616 + ], + "spans": [ + { + "bbox": [ + 313, + 582, + 554, + 616 + ], + "type": "text", + "content": "During this neuron labeling process, we aim to assign meaningful concepts to each neuron. We use CLIP-dissect because:" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 313, + 617, + 555, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 617, + 555, + 713 + ], + "spans": [ + { + "bbox": [ + 313, + 617, + 555, + 713 + ], + "type": "text", + "content": "- Concept set " + }, + { + "bbox": [ + 313, + 617, + 555, + 713 + ], + "type": "inline_equation", + "content": "S" + }, + { + "bbox": [ + 313, + 617, + 555, + 713 + ], + "type": "text", + "content": ": Instead of allowing an infinite range of possible concepts for neuron labeling, using a fixed concept set narrows this process by constraining it to a limited selection of concepts. Including diverse concepts in this set enables us to identify neurons corresponding to these visual concepts. This setup ensures that each neuron is assigned a specific label, though some labels may be spurious, as illustrated in Figure 3." + } + ] + } + ], + "index": 19 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "text", + "content": "4345" + } + ] + } + ], + "index": 20 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 2 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 59, + 72, + 291, + 169 + ], + "blocks": [ + { + "bbox": [ + 59, + 72, + 291, + 169 + ], + "lines": [ + { + "bbox": [ + 59, + 72, + 291, + 169 + ], + "spans": [ + { + "bbox": [ + 59, + 72, + 291, + 169 + ], + "type": "image", + "image_path": "e4ac9913c0aff54801f05a019813b1b9dc17b7cc977dbd54941f57501925f334.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 55, + 179, + 295, + 246 + ], + "lines": [ + { + "bbox": [ + 55, + 179, + 295, + 246 + ], + "spans": [ + { + "bbox": [ + 55, + 179, + 295, + 246 + ], + "type": "text", + "content": "Figure 3. Spurious labeled neurons in infant model CVCL by CLIP-Dissect, showing top-3 activated Konkle dataset images. This indicates: (1) not all neurons interpretable with human semantic concepts [12]; (2) neuron-labeling can produce spurious labels. To further support our hypothesis, we implement NeuronClassifier." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "bbox": [ + 55, + 268, + 296, + 376 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 268, + 296, + 376 + ], + "spans": [ + { + "bbox": [ + 55, + 268, + 296, + 376 + ], + "type": "text", + "content": "- Probing dataset " + }, + { + "bbox": [ + 55, + 268, + 296, + 376 + ], + "type": "inline_equation", + "content": "\\mathcal{D}_{\\mathrm{probe}}" + }, + { + "bbox": [ + 55, + 268, + 296, + 376 + ], + "type": "text", + "content": ": The probing dataset allows neurons to activate specifically in response to the dataset of interest. For instance, when analyzing representations from an infant model, we use images from the Konkle object dataset, which better aligns with infant recognition than the ImageNet [10] validation set. Additionally, for classification tasks, generalization can be achieved by adaptively selecting the probing dataset to focus on relevant concepts within a given dataset." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 55, + 381, + 296, + 538 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 381, + 296, + 538 + ], + "spans": [ + { + "bbox": [ + 55, + 381, + 296, + 538 + ], + "type": "text", + "content": "Extending beyond vocabulary. The flexibility of the concept set " + }, + { + "bbox": [ + 55, + 381, + 296, + 538 + ], + "type": "inline_equation", + "content": "S" + }, + { + "bbox": [ + 55, + 381, + 296, + 538 + ], + "type": "text", + "content": " allowing us to discover visual concepts that the infant model has never encountered in its training data. In this process, CLIP serves as a well-pretrained miner, utilizing its rich image-text embeddings to identify concepts hidden within CVCL. By aligning the activations of CVCL's neurons with CLIP's embeddings, we discover meaningful hidden visual concepts within the infant model, even extending beyond the model's linguistic training data. This approach leverages the diverse vocabulary of the concept set and the rich embeddings of CLIP to reveal visual concepts embedded in CVCL's internal representations, each associated with corresponding neurons." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 55, + 546, + 211, + 558 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 546, + 211, + 558 + ], + "spans": [ + { + "bbox": [ + 55, + 546, + 211, + 558 + ], + "type": "text", + "content": "3.2. Neuron-Based Classification" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 55, + 563, + 296, + 672 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 563, + 296, + 672 + ], + "spans": [ + { + "bbox": [ + 55, + 563, + 296, + 672 + ], + "type": "text", + "content": "How do we ensure that neurons representing concepts beyond the model's original vocabulary truly exist within the network? In this section, we propose NeuronClassifier, a training-free framework that leverages neuron activations to detect and validate such concepts. By discovering neurons with specific visual concept, we aim to confirm the presence of these latent, beyond-vocabulary neurons and use them to perform " + }, + { + "bbox": [ + 55, + 563, + 296, + 672 + ], + "type": "inline_equation", + "content": "n" + }, + { + "bbox": [ + 55, + 563, + 296, + 672 + ], + "type": "text", + "content": "-way classification. The framework, illustrated in Figure 4, involves three main steps." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 55, + 677, + 296, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 677, + 296, + 713 + ], + "spans": [ + { + "bbox": [ + 55, + 677, + 296, + 713 + ], + "type": "text", + "content": "Step 1: neuron labeling with concept set. Given an image encoder " + }, + { + "bbox": [ + 55, + 677, + 296, + 713 + ], + "type": "inline_equation", + "content": "f(x)" + }, + { + "bbox": [ + 55, + 677, + 296, + 713 + ], + "type": "text", + "content": ", we labeled each neuron in the network using a concept set " + }, + { + "bbox": [ + 55, + 677, + 296, + 713 + ], + "type": "inline_equation", + "content": "S" + }, + { + "bbox": [ + 55, + 677, + 296, + 713 + ], + "type": "text", + "content": " that contains (but is not limited to) all" + } + ] + } + ], + "index": 6 + }, + { + "type": "image", + "bbox": [ + 318, + 70, + 553, + 192 + ], + "blocks": [ + { + "bbox": [ + 318, + 70, + 553, + 192 + ], + "lines": [ + { + "bbox": [ + 318, + 70, + 553, + 192 + ], + "spans": [ + { + "bbox": [ + 318, + 70, + 553, + 192 + ], + "type": "image", + "image_path": "933a9a5319d3c4839ebd5e09ddbd5d3496a5173a40856fc83309dc4b55bb7a9c.jpg" + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 313, + 198, + 555, + 276 + ], + "lines": [ + { + "bbox": [ + 313, + 198, + 555, + 276 + ], + "spans": [ + { + "bbox": [ + 313, + 198, + 555, + 276 + ], + "type": "text", + "content": "Figure 4. Our NeuronClassifier Overview. A training-free " + }, + { + "bbox": [ + 313, + 198, + 555, + 276 + ], + "type": "inline_equation", + "content": "n" + }, + { + "bbox": [ + 313, + 198, + 555, + 276 + ], + "type": "text", + "content": "-way framework with three key steps: (1) Label all neurons in the network using a concept set that includes class labels and common words; (2) identify neurons associated with the target concept (e.g., \"rug\"); (3) Evaluate the activations of visual concept neuron across " + }, + { + "bbox": [ + 313, + 198, + 555, + 276 + ], + "type": "inline_equation", + "content": "n" + }, + { + "bbox": [ + 313, + 198, + 555, + 276 + ], + "type": "text", + "content": " images (in this example, " + }, + { + "bbox": [ + 313, + 198, + 555, + 276 + ], + "type": "inline_equation", + "content": "n = 4" + }, + { + "bbox": [ + 313, + 198, + 555, + 276 + ], + "type": "text", + "content": ") and select the image with the highest activation as the most relevant to the target label." + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "image_caption" + } + ], + "index": 7 + }, + { + "bbox": [ + 313, + 298, + 555, + 322 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 298, + 555, + 322 + ], + "spans": [ + { + "bbox": [ + 313, + 298, + 555, + 322 + ], + "type": "text", + "content": "class labels relevant to the task. The dissection process can be expressed as a function:" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 370, + 333, + 553, + 346 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 370, + 333, + 553, + 346 + ], + "spans": [ + { + "bbox": [ + 370, + 333, + 553, + 346 + ], + "type": "interline_equation", + "content": "\\mathcal {N} _ {s} = \\text {N e u r o n L a b e l i n g} (f, \\mathcal {S}), \\tag {4}", + "image_path": "e0193b67955f1b3935fdbdac73a9b35a025f354be9d96a244492bdf556dde810.jpg" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 313, + 356, + 554, + 417 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 356, + 554, + 417 + ], + "spans": [ + { + "bbox": [ + 313, + 356, + 554, + 417 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 313, + 356, + 554, + 417 + ], + "type": "inline_equation", + "content": "\\mathcal{N}_s" + }, + { + "bbox": [ + 313, + 356, + 554, + 417 + ], + "type": "text", + "content": " is the set of neurons labeled with concepts from " + }, + { + "bbox": [ + 313, + 356, + 554, + 417 + ], + "type": "inline_equation", + "content": "S" + }, + { + "bbox": [ + 313, + 356, + 554, + 417 + ], + "type": "text", + "content": ". Each neuron " + }, + { + "bbox": [ + 313, + 356, + 554, + 417 + ], + "type": "inline_equation", + "content": "k \\in \\mathcal{N}_s" + }, + { + "bbox": [ + 313, + 356, + 554, + 417 + ], + "type": "text", + "content": " is associated with a specific concept, such as \"rug\" or \"calculator\", based on its alignment with concept embeddings obtained during neuron labeling (e.g. similarity in CLIP-Dissect [30])." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 313, + 423, + 554, + 495 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 423, + 554, + 495 + ], + "spans": [ + { + "bbox": [ + 313, + 423, + 554, + 495 + ], + "type": "text", + "content": "Step 2: identifying visual concept neurons. Given a target label " + }, + { + "bbox": [ + 313, + 423, + 554, + 495 + ], + "type": "inline_equation", + "content": "l_{k} \\in \\mathcal{L}" + }, + { + "bbox": [ + 313, + 423, + 554, + 495 + ], + "type": "text", + "content": ", where " + }, + { + "bbox": [ + 313, + 423, + 554, + 495 + ], + "type": "inline_equation", + "content": "\\mathcal{L}" + }, + { + "bbox": [ + 313, + 423, + 554, + 495 + ], + "type": "text", + "content": " represents all neuron labels in the model, same as label vector in Section 3.1 (e.g., \"rug\"), we select the subset of neurons labeled with this concept, denoted as " + }, + { + "bbox": [ + 313, + 423, + 554, + 495 + ], + "type": "inline_equation", + "content": "\\mathcal{N}_{l_k} \\subset \\mathcal{N}_{\\mathcal{L}}" + }, + { + "bbox": [ + 313, + 423, + 554, + 495 + ], + "type": "text", + "content": ". These neurons are responsible for encoding the target concept." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 313, + 495, + 554, + 578 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 495, + 554, + 578 + ], + "spans": [ + { + "bbox": [ + 313, + 495, + 554, + 578 + ], + "type": "text", + "content": "To further refine the labeling and reduce spurious assignments, we select the most similar neuron from " + }, + { + "bbox": [ + 313, + 495, + 554, + 578 + ], + "type": "inline_equation", + "content": "\\mathcal{N}_{l_k}" + }, + { + "bbox": [ + 313, + 495, + 554, + 578 + ], + "type": "text", + "content": " to represent the target concept. The similarity measure varies based on the dissection method used. For example, Network Dissection [2] employs Intersection over Union (IoU) for similarity, while CLIP-Dissect [30] supports multiple similarity metrics:" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 368, + 590, + 553, + 608 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 368, + 590, + 553, + 608 + ], + "spans": [ + { + "bbox": [ + 368, + 590, + 553, + 608 + ], + "type": "interline_equation", + "content": "k ^ {*} = \\arg \\max _ {k \\in \\mathcal {N} _ {\\mathbf {L}}} \\operatorname {s i m} \\left(t _ {l _ {k}}, q _ {k}; P\\right), \\tag {5}", + "image_path": "df9181ac08d4210c6491d3063011bb3c55efbd3c4f43b2e3606902aec648659d.jpg" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 313, + 618, + 554, + 666 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 618, + 554, + 666 + ], + "spans": [ + { + "bbox": [ + 313, + 618, + 554, + 666 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 313, + 618, + 554, + 666 + ], + "type": "inline_equation", + "content": "t_{l_k}" + }, + { + "bbox": [ + 313, + 618, + 554, + 666 + ], + "type": "text", + "content": " is the embedding of the target label " + }, + { + "bbox": [ + 313, + 618, + 554, + 666 + ], + "type": "inline_equation", + "content": "l_k" + }, + { + "bbox": [ + 313, + 618, + 554, + 666 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 313, + 618, + 554, + 666 + ], + "type": "inline_equation", + "content": "q_k" + }, + { + "bbox": [ + 313, + 618, + 554, + 666 + ], + "type": "text", + "content": " is the neuron activation value, same in Section 3.1. This step ensures that the neuron most aligned with the concept is selected, minimizing the possibility of spurious labeling." + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 313, + 666, + 554, + 690 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 666, + 554, + 690 + ], + "spans": [ + { + "bbox": [ + 313, + 666, + 554, + 690 + ], + "type": "text", + "content": "For each selected neuron " + }, + { + "bbox": [ + 313, + 666, + 554, + 690 + ], + "type": "inline_equation", + "content": "k \\in \\mathcal{N}_{l_k}" + }, + { + "bbox": [ + 313, + 666, + 554, + 690 + ], + "type": "text", + "content": ", its activation value on an input image " + }, + { + "bbox": [ + 313, + 666, + 554, + 690 + ], + "type": "inline_equation", + "content": "x_i" + }, + { + "bbox": [ + 313, + 666, + 554, + 690 + ], + "type": "text", + "content": " is computed as" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 389, + 701, + 553, + 714 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 389, + 701, + 553, + 714 + ], + "spans": [ + { + "bbox": [ + 389, + 701, + 553, + 714 + ], + "type": "interline_equation", + "content": "q _ {k} \\left(x _ {i}\\right) = g \\left(A _ {k} \\left(x _ {i}\\right)\\right), \\tag {6}", + "image_path": "27e69f372462821834e11295e536c4f87a32f4eaf38acf201d6849875da784b1.jpg" + } + ] + } + ], + "index": 17 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "text", + "content": "4346" + } + ] + } + ], + "index": 18 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 3 + }, + { + "para_blocks": [ + { + "bbox": [ + 55, + 72, + 295, + 108 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 72, + 295, + 108 + ], + "spans": [ + { + "bbox": [ + 55, + 72, + 295, + 108 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 55, + 72, + 295, + 108 + ], + "type": "inline_equation", + "content": "A_{k}(x_{i})" + }, + { + "bbox": [ + 55, + 72, + 295, + 108 + ], + "type": "text", + "content": " is the raw activation map, and " + }, + { + "bbox": [ + 55, + 72, + 295, + 108 + ], + "type": "inline_equation", + "content": "g(\\cdot)" + }, + { + "bbox": [ + 55, + 72, + 295, + 108 + ], + "type": "text", + "content": " is a summary function (e.g., spatial mean) that reduces it to a scalar representing the neuron's response strength." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 55, + 111, + 295, + 171 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 111, + 295, + 171 + ], + "spans": [ + { + "bbox": [ + 55, + 111, + 295, + 171 + ], + "type": "text", + "content": "Step 3: selecting the most relevant image. Given " + }, + { + "bbox": [ + 55, + 111, + 295, + 171 + ], + "type": "inline_equation", + "content": "n" + }, + { + "bbox": [ + 55, + 111, + 295, + 171 + ], + "type": "text", + "content": " candidate images " + }, + { + "bbox": [ + 55, + 111, + 295, + 171 + ], + "type": "inline_equation", + "content": "\\{x_1, x_2, \\ldots, x_n\\}" + }, + { + "bbox": [ + 55, + 111, + 295, + 171 + ], + "type": "text", + "content": " in an " + }, + { + "bbox": [ + 55, + 111, + 295, + 171 + ], + "type": "inline_equation", + "content": "n" + }, + { + "bbox": [ + 55, + 111, + 295, + 171 + ], + "type": "text", + "content": "-way trial, we compute the activation values " + }, + { + "bbox": [ + 55, + 111, + 295, + 171 + ], + "type": "inline_equation", + "content": "q_k(x_i)" + }, + { + "bbox": [ + 55, + 111, + 295, + 171 + ], + "type": "text", + "content": " for neuron " + }, + { + "bbox": [ + 55, + 111, + 295, + 171 + ], + "type": "inline_equation", + "content": "k^*" + }, + { + "bbox": [ + 55, + 111, + 295, + 171 + ], + "type": "text", + "content": " with highest similarity across all images. The image with the highest activation is selected as the most relevant to the target concept:" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 127, + 182, + 294, + 199 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 127, + 182, + 294, + 199 + ], + "spans": [ + { + "bbox": [ + 127, + 182, + 294, + 199 + ], + "type": "interline_equation", + "content": "x ^ {*} = \\arg \\max _ {x _ {i}} q _ {k ^ {*}} \\left(x _ {i}\\right). \\tag {7}", + "image_path": "9ea8ff5f47d11b319b371708b1fbcb1fa8911cae13a9bcf08381a6961f59d327.jpg" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 55, + 203, + 295, + 239 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 203, + 295, + 239 + ], + "spans": [ + { + "bbox": [ + 55, + 203, + 295, + 239 + ], + "type": "text", + "content": "In the example shown in Figure 4, the target concept is \"rug\", and we select the image with the highest activation from the four candidates as the closest match to the concept." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 55, + 248, + 262, + 262 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 248, + 262, + 262 + ], + "spans": [ + { + "bbox": [ + 55, + 248, + 262, + 262 + ], + "type": "text", + "content": "4. Neuron-wise Representation Analysis" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 55, + 268, + 296, + 376 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 268, + 296, + 376 + ], + "spans": [ + { + "bbox": [ + 55, + 268, + 296, + 376 + ], + "type": "text", + "content": "In this section, we conduct a neuron-wise analysis of infant models and provide implementation details. First, we perform neuron labeling on the infant CVCL model. Then, we use our NeuronClassifier framework to leverage visual concept neurons identified through neuron labeling, resulting in better recognition performance than the original approach [41] while also discovering internal visual concepts beyond the model's training vocabulary. These results support our hypothesis." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 55, + 382, + 105, + 395 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 382, + 105, + 395 + ], + "spans": [ + { + "bbox": [ + 55, + 382, + 105, + 395 + ], + "type": "text", + "content": "4.1. Setup" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 55, + 399, + 296, + 519 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 399, + 296, + 519 + ], + "spans": [ + { + "bbox": [ + 55, + 399, + 296, + 519 + ], + "type": "text", + "content": "Datasets. We use the Konkle object dataset [23], as introduced in Section 2.1. This dataset consists of 3,406 images, each featuring a single object on a clean white background, including 406 test items across 200 classes. Each trial comprises " + }, + { + "bbox": [ + 55, + 399, + 296, + 519 + ], + "type": "inline_equation", + "content": "n" + }, + { + "bbox": [ + 55, + 399, + 296, + 519 + ], + "type": "text", + "content": " images: one target image and the remaining as foils, with foil images randomly sampled from classes other than the target class. For each class, we generate 5 trials, each containing " + }, + { + "bbox": [ + 55, + 399, + 296, + 519 + ], + "type": "inline_equation", + "content": "n" + }, + { + "bbox": [ + 55, + 399, + 296, + 519 + ], + "type": "text", + "content": " images. Following the previous work on CVCL, we use " + }, + { + "bbox": [ + 55, + 399, + 296, + 519 + ], + "type": "inline_equation", + "content": "n = 4" + }, + { + "bbox": [ + 55, + 399, + 296, + 519 + ], + "type": "text", + "content": " in our main experiments, with one target and three foils per trial." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 55, + 522, + 295, + 617 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 522, + 295, + 617 + ], + "spans": [ + { + "bbox": [ + 55, + 522, + 295, + 617 + ], + "type": "text", + "content": "Neuron labeling. We utilize CLIP-Dissect [30] for neuron labeling, which assigns visual concepts to each neuron in the network. As we aim to perform classification on the Konkle object dataset, we use the same dataset as a part of " + }, + { + "bbox": [ + 55, + 522, + 295, + 617 + ], + "type": "inline_equation", + "content": "\\mathcal{D}_{\\mathrm{probe}}" + }, + { + "bbox": [ + 55, + 522, + 295, + 617 + ], + "type": "text", + "content": ". Additionally, to avoid limiting the search space only around class names and ensure comprehensive neuron labeling, we employ a combined concept, consisting of three components:" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 55, + 618, + 295, + 713 + ], + "type": "list", + "angle": 0, + "index": 12, + "blocks": [ + { + "bbox": [ + 55, + 618, + 295, + 653 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 618, + 295, + 653 + ], + "spans": [ + { + "bbox": [ + 55, + 618, + 295, + 653 + ], + "type": "text", + "content": "- SAYCam-S vocabulary: We clean the original vocabulary by removing noisy child speech and retaining meaningful words." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 55, + 654, + 295, + 689 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 654, + 295, + 689 + ], + "spans": [ + { + "bbox": [ + 55, + 654, + 295, + 689 + ], + "type": "text", + "content": "- Common English words: We chose the top 30,000 most common English words based on a 1-gram frequency analysis by Peter Norvig [29, 30]." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 55, + 689, + 295, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 689, + 295, + 713 + ], + "spans": [ + { + "bbox": [ + 55, + 689, + 295, + 713 + ], + "type": "text", + "content": "- Class in Konkle object dataset: All class labels from the Konkle object dataset are included." + } + ] + } + ], + "index": 11 + } + ], + "sub_type": "text" + }, + { + "bbox": [ + 313, + 72, + 553, + 96 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 72, + 553, + 96 + ], + "spans": [ + { + "bbox": [ + 313, + 72, + 553, + 96 + ], + "type": "text", + "content": "We combine these three sources, ensuring no duplicates, resulting in a final concept set containing 30,427 words." + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 313, + 100, + 553, + 148 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 100, + 553, + 148 + ], + "spans": [ + { + "bbox": [ + 313, + 100, + 553, + 148 + ], + "type": "text", + "content": "Models. Our primary focus is the CVCL-ResNeXt50 [41], trained on SAY-Cam-S [37] dataset. For comprehensive analysis, we apply our framework to the following models:" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 314, + 150, + 553, + 342 + ], + "type": "list", + "angle": 0, + "index": 18, + "blocks": [ + { + "bbox": [ + 314, + 150, + 553, + 221 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 314, + 150, + 553, + 221 + ], + "spans": [ + { + "bbox": [ + 314, + 150, + 553, + 221 + ], + "type": "text", + "content": "- Infant models: CVCL is trained on unique infant data, using both egocentric frames and transcribed parent speech. We also take DINO-S-ResNeXt50 [31] as reference model compare recognition ability derived from same visual experience. It trained with DINO[6] self-supervised approach on the infant same dataset." + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 314, + 222, + 553, + 293 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 314, + 222, + 553, + 293 + ], + "spans": [ + { + "bbox": [ + 314, + 222, + 553, + 293 + ], + "type": "text", + "content": "- Broadly-trained models: To establish an upper bound, we include CLIP [34] and ResNeXt50 [43], which are broadly trained on large scale Internet images. Although CLIP uses a ResNet50-based vision encoder rather than ResNeXt, we select CLIP-ResNet50 due to the architectural similarity between ResNeXt [43] and ResNet [20]." + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 314, + 294, + 553, + 342 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 314, + 294, + 553, + 342 + ], + "spans": [ + { + "bbox": [ + 314, + 294, + 553, + 342 + ], + "type": "text", + "content": "- Randomized model: As a lower bound, we introduce a randomized version of the CVCL-ResNeXt50 model. In this setup, the convolution layer weights in the vision encoder are initialized using Kaiming Initialization [19]." + } + ] + } + ], + "index": 17 + } + ], + "sub_type": "text" + }, + { + "bbox": [ + 314, + 350, + 372, + 361 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 314, + 350, + 372, + 361 + ], + "spans": [ + { + "bbox": [ + 314, + 350, + 372, + 361 + ], + "type": "text", + "content": "4.2. Results" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 313, + 367, + 553, + 415 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 367, + 553, + 415 + ], + "spans": [ + { + "bbox": [ + 313, + 367, + 553, + 415 + ], + "type": "text", + "content": "In this section, we present our results and findings from applying neuron labeling and our NeuronClassifier framework to the infant CVCL model, compared with other reference models. We define two class types for our analysis:" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 314, + 418, + 553, + 489 + ], + "type": "list", + "angle": 0, + "index": 23, + "blocks": [ + { + "bbox": [ + 314, + 418, + 553, + 453 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 314, + 418, + 553, + 453 + ], + "spans": [ + { + "bbox": [ + 314, + 418, + 553, + 453 + ], + "type": "text", + "content": "- In-vocabulary classes: Object classes present in the model's training linguistic input, which also appear in test object class and are detectable in internal representations." + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 314, + 453, + 553, + 489 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 314, + 453, + 553, + 489 + ], + "spans": [ + { + "bbox": [ + 314, + 453, + 553, + 489 + ], + "type": "text", + "content": "- Out-of-vocabulary classes: Object classes not included in the training linguistic input but detected in test object class and internal representations." + } + ] + } + ], + "index": 22 + } + ], + "sub_type": "text" + }, + { + "type": "image", + "bbox": [ + 357, + 502, + 514, + 561 + ], + "blocks": [ + { + "bbox": [ + 357, + 502, + 514, + 561 + ], + "lines": [ + { + "bbox": [ + 357, + 502, + 514, + 561 + ], + "spans": [ + { + "bbox": [ + 357, + 502, + 514, + 561 + ], + "type": "image", + "image_path": "425151cb957ec1ad32b744485555faa6b42d59f1096e4fb78e3be27b764e78e0.jpg" + } + ] + } + ], + "index": 24, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 313, + 570, + 553, + 592 + ], + "lines": [ + { + "bbox": [ + 313, + 570, + 553, + 592 + ], + "spans": [ + { + "bbox": [ + 313, + 570, + 553, + 592 + ], + "type": "text", + "content": "Figure 5. In-vocabulary and out-of-vocabulary relationships visualized using a Venn diagram." + } + ] + } + ], + "index": 25, + "angle": 0, + "type": "image_caption" + } + ], + "index": 24 + }, + { + "bbox": [ + 313, + 605, + 555, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 605, + 555, + 714 + ], + "spans": [ + { + "bbox": [ + 313, + 605, + 555, + 714 + ], + "type": "text", + "content": "Our results demonstrate that the proposed framework effectively discovers meaningful neurons that represent concepts beyond the model's training vocabulary. These findings align with the cognitive perspective of vocabulary acquisition in infant development. We analyze the models' performance across different " + }, + { + "bbox": [ + 313, + 605, + 555, + 714 + ], + "type": "inline_equation", + "content": "n" + }, + { + "bbox": [ + 313, + 605, + 555, + 714 + ], + "type": "text", + "content": "-way classification settings, showing that our method yields strong out-of-vocabulary classification performance while simultaneously improving in-vocabulary classification accuracy." + } + ] + } + ], + "index": 26 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 315, + 757 + ], + "type": "text", + "content": "4347" + } + ] + } + ], + "index": 27 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 4 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 60, + 70, + 289, + 194 + ], + "blocks": [ + { + "bbox": [ + 60, + 70, + 289, + 194 + ], + "lines": [ + { + "bbox": [ + 60, + 70, + 289, + 194 + ], + "spans": [ + { + "bbox": [ + 60, + 70, + 289, + 194 + ], + "type": "image", + "image_path": "ffe366a0f8ed821e4fce48429054d773f759c535135c1bfd371e6e2db9f5647d.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 55, + 202, + 295, + 269 + ], + "lines": [ + { + "bbox": [ + 55, + 202, + 295, + 269 + ], + "spans": [ + { + "bbox": [ + 55, + 202, + 295, + 269 + ], + "type": "text", + "content": "Figure 6. Class coverage in visual concept neurons. Percentage of Konkle object dataset [23] classes identified through neuron labeling. Broadly pre-trained models (CLIP, ResNeXt) achieve over " + }, + { + "bbox": [ + 55, + 202, + 295, + 269 + ], + "type": "inline_equation", + "content": "50\\%" + }, + { + "bbox": [ + 55, + 202, + 295, + 269 + ], + "type": "text", + "content": " coverage, while developmentally inspired infant models (CVCL [41], DINO [6]) show lower coverage. CVCL-Randomized provides a lower-bound comparison." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "bbox": [ + 54, + 290, + 295, + 469 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 54, + 290, + 295, + 469 + ], + "spans": [ + { + "bbox": [ + 54, + 290, + 295, + 469 + ], + "type": "text", + "content": "Class coverage in visual concept neurons. How well do visual concept neurons identified through neuron labeling correspond to specific class names (e.g., \"rug\") rather than general descriptive attributes (e.g., \"red\"? Figure 6 shows the percentage of classes in the Konkle object dataset that are discovered in visual concept neurons from each model during the neuron labeling process. The results indicate that well-pretrained models, such as CLIP-ResNet50 and ResNeXt50, demonstrate broader class name coverage. While CVCL performs slightly weaker, it still maintains coverage of slightly less than " + }, + { + "bbox": [ + 54, + 290, + 295, + 469 + ], + "type": "inline_equation", + "content": "50\\%" + }, + { + "bbox": [ + 54, + 290, + 295, + 469 + ], + "type": "text", + "content": ". In contrast, CVCL-Randomized achieves only around " + }, + { + "bbox": [ + 54, + 290, + 295, + 469 + ], + "type": "inline_equation", + "content": "28\\%" + }, + { + "bbox": [ + 54, + 290, + 295, + 469 + ], + "type": "text", + "content": " coverage. This class coverage metric reflects the models' capacity to form class-corresponding meaningful representations during the neuron labeling process." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 55, + 474, + 295, + 616 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 474, + 295, + 616 + ], + "spans": [ + { + "bbox": [ + 55, + 474, + 295, + 616 + ], + "type": "text", + "content": "Age of Acquisition (AoA) ratings. Age of acquisition (AoA) is used to indicate when, and in what sequence, words are learned, and it is often assessed through ratings or observations reported by adults. This indirect method generally correlates well with other metrics indicating when children acquire vocabulary. Previous developmental work has shown that infants' early visual familiarity with common objects helps with object recognition, which subsequently helps support the process of learning the names of those objects. We next examined how early words in our models are learned and whether there is an AoA difference between in-vocab and out-of-vocab words." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 55, + 617, + 295, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 617, + 295, + 714 + ], + "spans": [ + { + "bbox": [ + 55, + 617, + 295, + 714 + ], + "type": "text", + "content": "We used AoA ratings from a dataset compiled by Kuperman, Stadthagen-Gonzalez, and Brysbaert [25], which includes norms for over 30,000 English words gathered via Amazon Mechanical Turk. Each participant estimated the age in years at which they believed they first understood each word, even if they did not actively use them. This dataset is comparable to previously reported AoA norms [36] gathered in laboratory settings." + } + ] + } + ], + "index": 4 + }, + { + "type": "image", + "bbox": [ + 316, + 71, + 542, + 209 + ], + "blocks": [ + { + "bbox": [ + 316, + 71, + 542, + 209 + ], + "lines": [ + { + "bbox": [ + 316, + 71, + 542, + 209 + ], + "spans": [ + { + "bbox": [ + 316, + 71, + 542, + 209 + ], + "type": "image", + "image_path": "5bf13d735d09342d50dc52a7e8341121f4445cd0f761898d1ff137794446a05c.jpg" + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 313, + 217, + 554, + 285 + ], + "lines": [ + { + "bbox": [ + 313, + 217, + 554, + 285 + ], + "spans": [ + { + "bbox": [ + 313, + 217, + 554, + 285 + ], + "type": "text", + "content": "Figure 7. Age of acquisition (AoA) ratings for in- and out-of-vocabulary visual concepts. Comparison of word acquisition timing [25] between in-vocabulary and out-of-vocabulary concepts. Out-of-vocabulary concepts tend to have a higher estimated acquisition ages in the infant-inspired CVCL model, indicating development of visual understanding beyond explicit linguist inputs." + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "image_caption" + } + ], + "index": 5 + }, + { + "bbox": [ + 313, + 308, + 555, + 488 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 308, + 555, + 488 + ], + "spans": [ + { + "bbox": [ + 313, + 308, + 555, + 488 + ], + "type": "text", + "content": "Using this set of AoA norms, we compared mean AoA between in-vocab and out-of-vocab words discovered in CVCL's internal representations. As shown in Figure 7, we found a significant difference between in-vocab and out-of-vocab AoA rating (" + }, + { + "bbox": [ + 313, + 308, + 555, + 488 + ], + "type": "inline_equation", + "content": "t(82) = 4.64" + }, + { + "bbox": [ + 313, + 308, + 555, + 488 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 313, + 308, + 555, + 488 + ], + "type": "inline_equation", + "content": "p < 0.0001" + }, + { + "bbox": [ + 313, + 308, + 555, + 488 + ], + "type": "text", + "content": "), in-vocab words (mean AoA = 4.99) are learned earlier than out-of-vocab words (mean AoA = 6.82). This pattern suggests that: (1) both sets of words are learned quite early, around later preschool and school years, with or without supervised labeling; (2) the difference in AoA between in-vocabulary and out-of-vocabulary words indicates that the infant model has developed a basic visual understanding of concepts with higher AoA. This foundational knowledge may lay the groundwork for word learning once corresponding parental speech is introduced." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 313, + 497, + 556, + 676 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 497, + 556, + 676 + ], + "spans": [ + { + "bbox": [ + 313, + 497, + 556, + 676 + ], + "type": "text", + "content": "Neuron-based classification performance. We evaluated the hidden potential of infant models' vision encoders by applying our NeuronClassifier framework, summarized in Table 1. Despite being trained on infant egocentric data with limited amount and diversity, CVCL demonstrates the ability to recognize similarly as nature of infant learning, revealing strong out-of-vocabulary classification performance. This result suggests that this infant model developed broader visual concepts that extend beyond linguistic input, similar to how infants naturally learn. We also applied our method to in-vocabulary classification, where it outperformed the vanilla method previously used in CVCL [41], as introduced in Figure 2. \"All\" representing the combined performance on both \"in-vocab\" and \"out-of-vocab\". These results support our hypothesis." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 313, + 677, + 555, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 677, + 555, + 715 + ], + "spans": [ + { + "bbox": [ + 313, + 677, + 555, + 715 + ], + "type": "text", + "content": "We include additional model comparisons: (1) The DINO-S-ResNeXt50 infant model [31], trained on the same dataset without text supervision, achieves comparable per" + } + ] + } + ], + "index": 9 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "text", + "content": "4348" + } + ] + } + ], + "index": 10 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 5 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 56, + 189, + 296, + 290 + ], + "blocks": [ + { + "bbox": [ + 55, + 70, + 296, + 182 + ], + "lines": [ + { + "bbox": [ + 55, + 70, + 296, + 182 + ], + "spans": [ + { + "bbox": [ + 55, + 70, + 296, + 182 + ], + "type": "text", + "content": "Table 1. Neuron-based classification results in 4-way evaluation among models in Section 4.1. \"Vanilla\" refers to classification based on image-text similarity (Figure 2). \"X\" denotes cases where direct classification on the Konkle dataset [23] is not possible due to missing text encoder or need fine-tuning. By leveraging neurons discovered in the representation, NeuronClassifier enables broader recognition, particularly in CVCL (bolded for emphasis), achieving improved recognition in both in-vocabulary and out-of-vocabulary, supporting our hypothesis. \"All\" represents the combined performance on both in- and out-of-vocabulary." + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 56, + 189, + 296, + 290 + ], + "lines": [ + { + "bbox": [ + 56, + 189, + 296, + 290 + ], + "spans": [ + { + "bbox": [ + 56, + 189, + 296, + 290 + ], + "type": "table", + "html": "
MethodModelIn-vocabOut-of-vocabAll
VanillaCLIP-ResNet5098.81±0.1696.93±0.0697.42±0.05
ResNeXt50×××
CVCL-ResNeXt5036.18±0.91××
DINO-S-ResNeXt50×××
Neuron ClassifierCLIP-ResNet5091.59±0.5288.66±0.3589.79±0.38
ResNeXt5088.17±0.4593.28±0.3691.88±0.15
CVCL-ResNeXt5079.50±0.7876.81±0.3577.79±0.40
DINO-S-ResNeXt5077.53±0.2477.96±0.2777.65±0.21
", + "image_path": "f284228f2cb39339ff83cec0e68f61d9d0ead3a72b09c012f455c3afb6b24433.jpg" + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "table_body" + } + ], + "index": 1 + }, + { + "bbox": [ + 54, + 311, + 295, + 430 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 54, + 311, + 295, + 430 + ], + "spans": [ + { + "bbox": [ + 54, + 311, + 295, + 430 + ], + "type": "text", + "content": "formance in visual representations. This suggests that models trained on identical data distributions with different self-supervised methods may yield similar representational outcomes. (2) Broadly-trained models such as ResNeXt50 and CLIP establish performance upper bounds. However, CLIP underperforms in the vanilla setting, relying exclusively on visual neurons without text encoder guidance. While this work does not aim to advance zero-shot learning methods, it reveals visual concepts in infant models that emerge independently of linguistic inputs." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 55, + 431, + 296, + 502 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 431, + 296, + 502 + ], + "spans": [ + { + "bbox": [ + 55, + 431, + 296, + 502 + ], + "type": "text", + "content": "In classification on the Konkle object dataset, both DINO-S and ImageNet ResNeXt50 required fine-tuning for this task. Our framework enables neuron-based classification without downstream fine-tuning, and providing a training-free qualitative inspection of internal representations." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 55, + 510, + 295, + 557 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 510, + 295, + 557 + ], + "spans": [ + { + "bbox": [ + 55, + 510, + 295, + 557 + ], + "type": "text", + "content": "Analysis across " + }, + { + "bbox": [ + 55, + 510, + 295, + 557 + ], + "type": "inline_equation", + "content": "n" + }, + { + "bbox": [ + 55, + 510, + 295, + 557 + ], + "type": "text", + "content": "-way settings. We evaluate the models under various " + }, + { + "bbox": [ + 55, + 510, + 295, + 557 + ], + "type": "inline_equation", + "content": "n" + }, + { + "bbox": [ + 55, + 510, + 295, + 557 + ], + "type": "text", + "content": "-way classification setups. Figure 8 illustrates the performance trends for in- and out-of-vocabulary class classification accuracy applying our NeuronClassifier." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 55, + 558, + 295, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 558, + 295, + 714 + ], + "spans": [ + { + "bbox": [ + 55, + 558, + 295, + 714 + ], + "type": "text", + "content": "Our method not only enables out-of-vocabulary classification but also significantly improves in-vocab performance compared to previous results [41], further supporting the presence of beyond-vocabulary potential in infant models. However, due to limited class coverage, CVCL with our method can classify a maximum of 31 classes (see Appendix A.3). These findings support our hypothesis that the infant model has acquired visual concepts beyond its initial vocabulary. These results show that leveraging the model's internal representations for classification that go beyond its vocabulary is not only feasible but also robust across different " + }, + { + "bbox": [ + 55, + 558, + 295, + 714 + ], + "type": "inline_equation", + "content": "n" + }, + { + "bbox": [ + 55, + 558, + 295, + 714 + ], + "type": "text", + "content": "-way settings. These findings support our hypothesis that the infant model has acquired visual concepts beyond" + } + ] + } + ], + "index": 5 + }, + { + "type": "image", + "bbox": [ + 316, + 72, + 436, + 247 + ], + "blocks": [ + { + "bbox": [ + 316, + 72, + 436, + 247 + ], + "lines": [ + { + "bbox": [ + 316, + 72, + 436, + 247 + ], + "spans": [ + { + "bbox": [ + 316, + 72, + 436, + 247 + ], + "type": "image", + "image_path": "42c71236126c2ca2c87f41d16d2a98c32bb84d27134c7a9a5a5694e72ecaa499.jpg" + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 313, + 255, + 555, + 354 + ], + "lines": [ + { + "bbox": [ + 313, + 255, + 555, + 354 + ], + "spans": [ + { + "bbox": [ + 313, + 255, + 555, + 354 + ], + "type": "text", + "content": "Figure 8. In- and out-of-vocabulary class performance across " + }, + { + "bbox": [ + 313, + 255, + 555, + 354 + ], + "type": "inline_equation", + "content": "n" + }, + { + "bbox": [ + 313, + 255, + 555, + 354 + ], + "type": "text", + "content": "-way settings using NeuronClassifier. (a) For out-of-vocabulary classification (left), our method enables classification without additional training. The infant model CVCL (" + }, + { + "bbox": [ + 313, + 255, + 555, + 354 + ], + "type": "inline_equation", + "content": "\\clubsuit" + }, + { + "bbox": [ + 313, + 255, + 555, + 354 + ], + "type": "text", + "content": ") maintains robust performance as " + }, + { + "bbox": [ + 313, + 255, + 555, + 354 + ], + "type": "inline_equation", + "content": "n" + }, + { + "bbox": [ + 313, + 255, + 555, + 354 + ], + "type": "text", + "content": " increases. (b) For in-vocabulary classification (right), \"ours\" (NeuronClassifier) enhances CVCL recognition ability compared to the \"vanilla\" setting. However, CLIP (" + }, + { + "bbox": [ + 313, + 255, + 555, + 354 + ], + "type": "inline_equation", + "content": "\\clubsuit" + }, + { + "bbox": [ + 313, + 255, + 555, + 354 + ], + "type": "text", + "content": ") declines, as it relies solely on neurons without text encoder input. Random ones serve as lower bounds." + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "image_caption" + } + ], + "index": 6 + }, + { + "type": "image", + "bbox": [ + 441, + 72, + 539, + 247 + ], + "blocks": [ + { + "bbox": [ + 441, + 72, + 539, + 247 + ], + "lines": [ + { + "bbox": [ + 441, + 72, + 539, + 247 + ], + "spans": [ + { + "bbox": [ + 441, + 72, + 539, + 247 + ], + "type": "image", + "image_path": "34575b3d787034d64dcf712c3f7a84750558d28b1dd93679d3db2b0f4d1e2944.jpg" + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "image_body" + } + ], + "index": 7 + }, + { + "bbox": [ + 313, + 376, + 400, + 388 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 376, + 400, + 388 + ], + "spans": [ + { + "bbox": [ + 313, + 376, + 400, + 388 + ], + "type": "text", + "content": "its initial vocabulary." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 313, + 389, + 555, + 521 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 389, + 555, + 521 + ], + "spans": [ + { + "bbox": [ + 313, + 389, + 555, + 521 + ], + "type": "text", + "content": "However, as " + }, + { + "bbox": [ + 313, + 389, + 555, + 521 + ], + "type": "inline_equation", + "content": "n" + }, + { + "bbox": [ + 313, + 389, + 555, + 521 + ], + "type": "text", + "content": " increases exponentially, performance gradually declines. This decline is reasonable, as the task becomes increasingly difficult by nature as " + }, + { + "bbox": [ + 313, + 389, + 555, + 521 + ], + "type": "inline_equation", + "content": "n" + }, + { + "bbox": [ + 313, + 389, + 555, + 521 + ], + "type": "text", + "content": " increases. It may also be attributed to dimensionality reduction in activation maps, leading to coarser classification. CLIPRN50 and ResNeXt50 perform well with NeuronClassifier, though not as effectively as their direct or fine-tuned versions, since our method is designed to reveal latent concepts rather than to perform fully optimized classification. In in-vocab settings, the \"vanilla\" approach represents direct classification as shown in Figure 2." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 313, + 533, + 512, + 547 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 533, + 512, + 547 + ], + "spans": [ + { + "bbox": [ + 313, + 533, + 512, + 547 + ], + "type": "text", + "content": "5. Layer-wise Representation Analysis" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 313, + 552, + 555, + 673 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 552, + 555, + 673 + ], + "spans": [ + { + "bbox": [ + 313, + 552, + 555, + 673 + ], + "type": "text", + "content": "How does the representation learned from infant data differ from other representations widely used in the computer vision community? To explore this, we perform a layer-wise representation analysis using Centered Kernel Alignment (CKA) [24]. We compute the similarity between the representations of the infant model(CVCL), ImageNet-pretrained, and CLIP. Additionally, we apply neuron labeling techniques from a layer-wise perspective to identify unique visual concepts discovered at each layer between ImageNet and infant models." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 313, + 677, + 556, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 677, + 556, + 715 + ], + "spans": [ + { + "bbox": [ + 313, + 677, + 556, + 715 + ], + "type": "text", + "content": "Layer similarity analysis. CKA [24] applies HSIC [18] (see Appendix B) over a set of images to provide layer-wise similarity scores between the representations of different" + } + ] + } + ], + "index": 13 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "text", + "content": "4349" + } + ] + } + ], + "index": 14 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 6 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 56, + 72, + 294, + 186 + ], + "blocks": [ + { + "bbox": [ + 56, + 72, + 294, + 186 + ], + "lines": [ + { + "bbox": [ + 56, + 72, + 294, + 186 + ], + "spans": [ + { + "bbox": [ + 56, + 72, + 294, + 186 + ], + "type": "image", + "image_path": "9ad3b396e855efda1aed02449395d95d28c514c860fa7bfcb0ceb0e97b2bc720.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 55, + 194, + 296, + 273 + ], + "lines": [ + { + "bbox": [ + 55, + 194, + 296, + 273 + ], + "spans": [ + { + "bbox": [ + 55, + 194, + 296, + 273 + ], + "type": "text", + "content": "Figure 9. CKA layer-wise similarity between CVCL and common models. Using the ImageNet validation set as input, CVCL (y-axis) exhibits similarity to CLIP-RN50 (x-axis, left) and ResNeXt50 (x-axis, right) in the shallow layers (lower-level features) but diverges significantly in the final layer (higher-level features). Notably, Layer 4 of CVCL shows very low similarity to all layers of both common models." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "bbox": [ + 54, + 295, + 295, + 498 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 54, + 295, + 295, + 498 + ], + "spans": [ + { + "bbox": [ + 54, + 295, + 295, + 498 + ], + "type": "text", + "content": "neural networks. A higher CKA score indicates more similar representations between two models at the given layer. We use the ImageNet validation set [10] as input for the networks, and compute the CKA similarity between the infant model (CVCL), the ImageNet-pretrained ResNeXt50, and CLIP-ResNet50. The results are presented as matrices in Figure 9. The lower layers of CVCL exhibit greater similarity to larger-scale models than its deeper layers. In larger-scale models, shallow layers are known to capture low-complexity features, while deeper layers progressively specialize in capturing higher-level concepts [7, 14]. Therefore, CVCL – the model trained on infant data – successfully develops lower-level representations comparable to those in common pre-trained models. However, the divergence in deeper layers suggests a lack of the diverse higher-level representations typically observed in models trained on common datasets." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 55, + 510, + 295, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 510, + 295, + 713 + ], + "spans": [ + { + "bbox": [ + 55, + 510, + 295, + 713 + ], + "type": "text", + "content": "Neuron-based analysis. To investigate the characteristics of each layer, we apply network dissection to identify neurons that are aligned with specific visual concepts. Essentially, this is an extension of the neuron labeling process, where the Broden [2] dataset provides category labels for each visual concept neuron. We count the number of unique visual concepts discovered in each category and perform layer-wise comparisons to gain a broader view of the differences between models trained from ImageNet and infant data. In Figure 10, we visualize the number of unique visual concept neurons across layers for each model. The results show that early layers in both models predominantly have neurons of low-level features like color, with minimal differences between models. However, as we move to deeper layers, higher-level concepts such as objects and scenes become more prominent, and the disparities between models become clearer. CVCL exhibits fewer unique visual con" + } + ] + } + ], + "index": 3 + }, + { + "type": "image", + "bbox": [ + 314, + 72, + 547, + 194 + ], + "blocks": [ + { + "bbox": [ + 314, + 72, + 547, + 194 + ], + "lines": [ + { + "bbox": [ + 314, + 72, + 547, + 194 + ], + "spans": [ + { + "bbox": [ + 314, + 72, + 547, + 194 + ], + "type": "image", + "image_path": "1d9df6f63c2f7251555648f664f46debbcd23323a72a3a93e1573922be641f1c.jpg" + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 313, + 203, + 555, + 291 + ], + "lines": [ + { + "bbox": [ + 313, + 203, + 555, + 291 + ], + "spans": [ + { + "bbox": [ + 313, + 203, + 555, + 291 + ], + "type": "text", + "content": "Figure 10. Number of visual concepts in ImageNet and infant models. Concepts are categorized using Broden dataset [2]. Neurons in deeper model layers capture increasingly complex concepts. Early layers primarily detect lower-level features like color and texture, while higher-level concepts such as objects and scenes emerge in deeper layers. CVCL exhibits fewer visual concepts than the ImageNet model, especially for higher-level visual concepts (e.g., objects and scenes)." + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "image_caption" + } + ], + "index": 4 + }, + { + "bbox": [ + 313, + 312, + 555, + 338 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 312, + 555, + 338 + ], + "spans": [ + { + "bbox": [ + 313, + 312, + 555, + 338 + ], + "type": "text", + "content": "cepts in these higher-level categories compared to ImageNet model. This finding aligns with the layer similarity analysis." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 313, + 347, + 388, + 360 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 347, + 388, + 360 + ], + "spans": [ + { + "bbox": [ + 313, + 347, + 388, + 360 + ], + "type": "text", + "content": "6. Conclusion" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 312, + 367, + 555, + 533 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 312, + 367, + 555, + 533 + ], + "spans": [ + { + "bbox": [ + 312, + 367, + 555, + 533 + ], + "type": "text", + "content": "In this paper, we explored whether an infant model (CVCL), trained on infant egocentric video frames and linguistic inputs, can acquire broader visual concepts extending beyond its initial training vocabulary. By introducing NeuronClassifier, a training-free framework to discover and leverage visual concepts hidden in representations, we unlocked the CVCL visual encoder's ability to recognize out-of-vocabulary concepts, establishing its potential as a strong classifier. Our findings also reveal that while CVCL, trained on a unique infant dataset with limited exposure to diverse scenes, it representations capture low-level features similar to those in common pre-trained models, they diverge significantly in higher-level representations, contributing to the observed performance differences." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 313, + 534, + 555, + 595 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 534, + 555, + 595 + ], + "spans": [ + { + "bbox": [ + 313, + 534, + 555, + 595 + ], + "type": "text", + "content": "Overall, our approach bridges cognitive science and computer vision, providing insights into how infant models develop visual concepts that precede linguistic inputs, aligning with the natural way infants explore the world through sight." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 313, + 597, + 556, + 704 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 597, + 556, + 704 + ], + "spans": [ + { + "bbox": [ + 313, + 597, + 556, + 704 + ], + "type": "text", + "content": "Limitations and future work. Our study did not analyze the infant training data, as we were unable to access it due to strict access controls. Instead, we analyze models trained on infant egocentric data, finding developmental alignment with cognitive studies. However, we did not extend this analysis to adult egocentric data. While the study focuses on infant data for developmental process, the framework can be applied to adult models, which remains a direction for future research." + } + ] + } + ], + "index": 10 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "text", + "content": "4350" + } + ] + } + ], + "index": 11 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 7 + }, + { + "para_blocks": [ + { + "bbox": [ + 56, + 72, + 158, + 85 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 72, + 158, + 85 + ], + "spans": [ + { + "bbox": [ + 56, + 72, + 158, + 85 + ], + "type": "text", + "content": "Acknowledgements" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 55, + 91, + 296, + 174 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 91, + 296, + 174 + ], + "spans": [ + { + "bbox": [ + 55, + 91, + 296, + 174 + ], + "type": "text", + "content": "This work was partially supported by the National Research Foundation Singapore Competitive Research Program (award number CRP29-2022-0003). We thank Wai Keen Vong for invaluable discussion and CVCL's pretrained weights. We thank Jingyi Lin for figure discussions. We thank the anonymous reviewers for their constructive feedback." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 56, + 186, + 115, + 199 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 186, + 115, + 199 + ], + "spans": [ + { + "bbox": [ + 56, + 186, + 115, + 199 + ], + "type": "text", + "content": "References" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 57, + 206, + 296, + 713 + ], + "type": "list", + "angle": 0, + "index": 14, + "blocks": [ + { + "bbox": [ + 61, + 206, + 296, + 240 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 61, + 206, + 296, + 240 + ], + "spans": [ + { + "bbox": [ + 61, + 206, + 296, + 240 + ], + "type": "text", + "content": "[1] Sven Bambach, David Crandall, Linda Smith, and Chen Yu. Toddler-inspired visual object learning. In Neural Information Processing Systems, 2018. 1, 2" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 61, + 240, + 296, + 297 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 61, + 240, + 296, + 297 + ], + "spans": [ + { + "bbox": [ + 61, + 240, + 296, + 297 + ], + "type": "text", + "content": "[2] David Bau, Bolei Zhou, Aditya Khosla, Aude Oliva, and Antonio Torralba. Network dissection: Quantifying interpretability of deep visual representations. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 6541-6549, 2017. 2, 4, 8" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 62, + 297, + 295, + 341 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 297, + 295, + 341 + ], + "spans": [ + { + "bbox": [ + 62, + 297, + 295, + 341 + ], + "type": "text", + "content": "[3] Erika Bergelson and Daniel Swingley. At 6-9 months, human infants know the meanings of many common nouns. Proceedings of the National Academy of Sciences, 109(9): 3253-3258, 2012. 2" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 62, + 342, + 295, + 397 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 342, + 295, + 397 + ], + "spans": [ + { + "bbox": [ + 62, + 342, + 295, + 397 + ], + "type": "text", + "content": "[4] Marc Brysbaert and Boris New. Moving beyond kucera and francis: A critical evaluation of current word frequency norms and the introduction of a new and improved word frequency measure for american english. Behavior Research Methods, 41(4):977-990, 2009. 1" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 62, + 399, + 295, + 442 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 399, + 295, + 442 + ], + "spans": [ + { + "bbox": [ + 62, + 399, + 295, + 442 + ], + "type": "text", + "content": "[5] Kirill Bykov, Laura Kopf, Shinichi Nakajima, Marius Kloft, and Marina Höhne. Labeling neural representations with inverse recognition. Advances in Neural Information Processing Systems, 36, 2024. 2" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 62, + 444, + 295, + 498 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 444, + 295, + 498 + ], + "spans": [ + { + "bbox": [ + 62, + 444, + 295, + 498 + ], + "type": "text", + "content": "[6] Mathilde Caron, Hugo Touvron, Ishan Misra, Hervé Jégou, Julien Mairal, Piotr Bojanowski, and Armand Joulin. Emerging properties in self-supervised vision transformers. In Proceedings of the IEEE/CVF international conference on computer vision, pages 9650-9660, 2021. 5, 6" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 62, + 500, + 295, + 553 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 500, + 295, + 553 + ], + "spans": [ + { + "bbox": [ + 62, + 500, + 295, + 553 + ], + "type": "text", + "content": "[7] Yixiong Chen, Alan Yuille, and Zongwei Zhou. Which layer is learning faster? a systematic exploration of layerwise convergence rate for deep neural networks. In The Eleventh International Conference on Learning Representations, 2023. 8" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 62, + 555, + 295, + 609 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 555, + 295, + 609 + ], + "spans": [ + { + "bbox": [ + 62, + 555, + 295, + 609 + ], + "type": "text", + "content": "[8] Elizabeth M Clerkin, Elizabeth Hart, James M Rehg, Chen Yu, and Linda B Smith. Real-world visual statistics and infants' first-learned object names. Philosophical Transactions of the Royal Society B: Biological Sciences, 372(1711): 20160055, 2017. 2" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 62, + 612, + 295, + 644 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 612, + 295, + 644 + ], + "spans": [ + { + "bbox": [ + 62, + 612, + 295, + 644 + ], + "type": "text", + "content": "[9] Michael J Cortese and Maya M Khanna. Age of acquisition ratings for 3,000 monosyllabic words. Behavior Research Methods, 40(3):791-794, 2008. 1" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 57, + 647, + 295, + 689 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 647, + 295, + 689 + ], + "spans": [ + { + "bbox": [ + 57, + 647, + 295, + 689 + ], + "type": "text", + "content": "[10] Jia Deng, Wei Dong, Richard Socher, Li-Jia Li, Kai Li, and Li Fei-Fei. Imagenet: A large-scale hierarchical image database. In 2009 IEEE conference on computer vision and pattern recognition, pages 248-255. IEEE, 2009. 4, 8" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 57, + 691, + 295, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 691, + 295, + 713 + ], + "spans": [ + { + "bbox": [ + 57, + 691, + 295, + 713 + ], + "type": "text", + "content": "[11] Amir Dravid, Yossi Gandelsman, Alexei A Efros, and Assaf Shocher. Rosetta neurons: Mining the common units in a" + } + ] + } + ], + "index": 13 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 316, + 73, + 555, + 713 + ], + "type": "list", + "angle": 0, + "index": 29, + "blocks": [ + { + "bbox": [ + 333, + 73, + 554, + 95 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 333, + 73, + 554, + 95 + ], + "spans": [ + { + "bbox": [ + 333, + 73, + 554, + 95 + ], + "type": "text", + "content": "model zoo. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 1934-1943, 2023. 3" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 316, + 98, + 554, + 151 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 98, + 554, + 151 + ], + "spans": [ + { + "bbox": [ + 316, + 98, + 554, + 151 + ], + "type": "text", + "content": "[12] Nelson Elhage, Tristan Hume, Catherine Olsson, Nicholas Schiefer, Tom Henighan, Shauna Kravec, Zac Hatfield-Dodds, Robert Lasenby, Dawn Drain, Carol Chen, et al. Toy models of superposition. arXiv preprint arXiv:2209.10652, 2022.4" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 316, + 153, + 553, + 176 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 153, + 553, + 176 + ], + "spans": [ + { + "bbox": [ + 316, + 153, + 553, + 176 + ], + "type": "text", + "content": "[13] Robert L Fantz. Pattern vision in newborn infants. Science, 140(3564):296-297, 1963. 2" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 317, + 178, + 554, + 221 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 178, + 554, + 221 + ], + "spans": [ + { + "bbox": [ + 317, + 178, + 554, + 221 + ], + "type": "text", + "content": "[14] Thomas Fel, Louis Bethune, Andrew Kyle Lampinen, Thomas Serre, and Katherine Hermann. Understanding visual feature reliance through the lens of complexity. arXiv preprint arXiv:2407.06076, 2024. 8" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 316, + 224, + 553, + 257 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 224, + 553, + 257 + ], + "spans": [ + { + "bbox": [ + 316, + 224, + 553, + 257 + ], + "type": "text", + "content": "[15] Yossi Gandelsman, Alexei A Efros, and Jacob Steinhardt. Interpreting clip's image representation via text-based decomposition. arXiv preprint arXiv:2310.05916, 2023. 3" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 316, + 258, + 555, + 303 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 258, + 555, + 303 + ], + "spans": [ + { + "bbox": [ + 316, + 258, + 555, + 303 + ], + "type": "text", + "content": "[16] Roberta Michnick Golinkoff, Kathryn Hirsh-Pasek, Kathleen M Cauley, and Laura Gordon. The eyes have it: Lexical and syntactic comprehension in a new paradigm. Journal of child language, 14(1):23-45, 1987. 3" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 316, + 304, + 554, + 358 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 304, + 554, + 358 + ], + "spans": [ + { + "bbox": [ + 316, + 304, + 554, + 358 + ], + "type": "text", + "content": "[17] Roberta Michnick Golinkoff, Weiyi Ma, Lulu Song, and Kathy Hirsh-Pasek. Twenty-five years using the intermodal preferential looking paradigm to study language acquisition: What have we learned? Perspectives on Psychological Science, 8(3):316-339, 2013. 2, 3" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 316, + 361, + 554, + 405 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 361, + 554, + 405 + ], + "spans": [ + { + "bbox": [ + 316, + 361, + 554, + 405 + ], + "type": "text", + "content": "[18] Arthur Gretton, Ralf Herbrich, Alexander Smola, Olivier Bousquet, Bernhard Schölkopf, and Aapo Hyvarinen. Kernel methods for measuring independence. Journal of Machine Learning Research, 6(12), 2005. 7, 3" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 317, + 407, + 554, + 461 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 407, + 554, + 461 + ], + "spans": [ + { + "bbox": [ + 317, + 407, + 554, + 461 + ], + "type": "text", + "content": "[19] Kaiming He, Xiangyu Zhang, Shaoqing Ren, and Jian Sun. Delving deep into rectifiers: Surpassing human-level performance on imagenet classification. In Proceedings of the IEEE international conference on computer vision, pages 1026-1034, 2015. 5" + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 316, + 464, + 553, + 508 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 464, + 553, + 508 + ], + "spans": [ + { + "bbox": [ + 316, + 464, + 553, + 508 + ], + "type": "text", + "content": "[20] Kaiming He, Xiangyu Zhang, Shaoqing Ren, and Jian Sun. Deep residual learning for image recognition. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 770-778, 2016. 5" + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 316, + 510, + 553, + 553 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 510, + 553, + 553 + ], + "spans": [ + { + "bbox": [ + 316, + 510, + 553, + 553 + ], + "type": "text", + "content": "[21] Evan Hernandez, Sarah Schwettmann, David Bau, Teona Bagashvili, Antonio Torralba, and Jacob Andreas. Natural language descriptions of deep visual features. In International Conference on Learning Representations, 2021. 2" + } + ] + } + ], + "index": 25 + }, + { + "bbox": [ + 317, + 555, + 554, + 609 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 555, + 554, + 609 + ], + "spans": [ + { + "bbox": [ + 317, + 555, + 554, + 609 + ], + "type": "text", + "content": "[22] Neha Kalibhat, Shweta Bhardwaj, C Bayan Bruss, Hamed Firooz, Maziar Sanjabi, and Soheil Feizi. Identifying interpretable subspaces in image representations. In International Conference on Machine Learning, pages 15623-15638. PMLR, 2023. 2" + } + ] + } + ], + "index": 26 + }, + { + "bbox": [ + 316, + 612, + 554, + 667 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 612, + 554, + 667 + ], + "spans": [ + { + "bbox": [ + 316, + 612, + 554, + 667 + ], + "type": "text", + "content": "[23] Talia Konkle, Timothy F Brady, George A Alvarez, and Aude Oliva. Conceptual distinctiveness supports detailed visual long-term memory for real-world objects. Journal of experimental Psychology: general, 139(3):558, 2010. 3, 5, 6, 7, 1, 2" + } + ] + } + ], + "index": 27 + }, + { + "bbox": [ + 316, + 670, + 554, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 670, + 554, + 713 + ], + "spans": [ + { + "bbox": [ + 316, + 670, + 554, + 713 + ], + "type": "text", + "content": "[24] Simon Kornblith, Mohammad Norouzi, Honglak Lee, and Geoffrey Hinton. Similarity of neural network representations revisited. In International conference on machine learning, pages 3519-3529. PMLR, 2019. 7" + } + ] + } + ], + "index": 28 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 748, + 314, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 314, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 314, + 757 + ], + "type": "text", + "content": "4351" + } + ] + } + ], + "index": 30 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 8 + }, + { + "para_blocks": [ + { + "bbox": [ + 56, + 73, + 294, + 713 + ], + "type": "list", + "angle": 0, + "index": 17, + "blocks": [ + { + "bbox": [ + 56, + 73, + 294, + 106 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 73, + 294, + 106 + ], + "spans": [ + { + "bbox": [ + 56, + 73, + 294, + 106 + ], + "type": "text", + "content": "[25] Victor Kuperman, Hans Stadthagen-Gonzalez, and Marc Brysbaert. Age-of-acquisition ratings for 30,000 english words. Behavior research methods, 44:978-990, 2012. 6" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 56, + 107, + 294, + 150 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 107, + 294, + 150 + ], + "spans": [ + { + "bbox": [ + 56, + 107, + 294, + 150 + ], + "type": "text", + "content": "[26] Victor Kuperman, Hans Stadthagen-Gonzalez, and Marc Brysbaert. Age-of-acquisition ratings for 30,000 english words. Behavior Research Methods, 44(4):978-990, 2012. 2, 1" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 56, + 152, + 294, + 174 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 152, + 294, + 174 + ], + "spans": [ + { + "bbox": [ + 56, + 152, + 294, + 174 + ], + "type": "text", + "content": "[27] Jean M Mandler. How to build a baby: II. conceptual primitives. Psychological review, 99(4):587, 1992. 2" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 56, + 175, + 294, + 207 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 175, + 294, + 207 + ], + "spans": [ + { + "bbox": [ + 56, + 175, + 294, + 207 + ], + "type": "text", + "content": "[28] Jesse Mu and Jacob Andreas. Compositional explanations of neurons. Advances in Neural Information Processing Systems, 33:17153-17163, 2020. 2" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 56, + 209, + 294, + 229 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 209, + 294, + 229 + ], + "spans": [ + { + "bbox": [ + 56, + 209, + 294, + 229 + ], + "type": "text", + "content": "[29] Peter Norvig. Natural language corpus data: Beautiful data, 2009. Accessed: 2024-10-27. 5" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 56, + 232, + 294, + 264 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 232, + 294, + 264 + ], + "spans": [ + { + "bbox": [ + 56, + 232, + 294, + 264 + ], + "type": "text", + "content": "[30] Tuomas Oikarinen and Tsui-Wei Weng. Clip-dissect: Automatic description of neuron representations in deep vision networks. arXiv preprint arXiv:2204.10965, 2022. 2, 3, 4, 5" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 56, + 266, + 294, + 308 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 266, + 294, + 308 + ], + "spans": [ + { + "bbox": [ + 56, + 266, + 294, + 308 + ], + "type": "text", + "content": "[31] A Emin Orhan and Brenden M Lake. Learning high-level visual representations from a child's perspective without strong inductive biases. Nature Machine Intelligence, 6(3):271-283, 2024. 1, 2, 5, 6" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 56, + 310, + 294, + 353 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 310, + 294, + 353 + ], + "spans": [ + { + "bbox": [ + 56, + 310, + 294, + 353 + ], + "type": "text", + "content": "[32] Emin Orhan, Vaibhav Gupta, and Brenden M Lake. Self-supervised learning through the eyes of a child. Advances in Neural Information Processing Systems, 33:9960-9971, 2020. 1, 2" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 56, + 355, + 294, + 387 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 355, + 294, + 387 + ], + "spans": [ + { + "bbox": [ + 56, + 355, + 294, + 387 + ], + "type": "text", + "content": "[33] Deborah A Phillips and Jack P Shonkoff. From neurons to neighborhoods: The science of early childhood development. 2000. 2" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 56, + 388, + 294, + 453 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 388, + 294, + 453 + ], + "spans": [ + { + "bbox": [ + 56, + 388, + 294, + 453 + ], + "type": "text", + "content": "[34] Alec Radford, Jong Wook Kim, Chris Hallacy, Aditya Ramesh, Gabriel Goh, Sandhini Agarwal, Girish Sastry, Amanda Askell, Pamela Mishkin, Jack Clark, et al. Learning transferable visual models from natural language supervision. In International conference on machine learning, pages 8748-8763. PMLR, 2021. 2, 3, 5" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 56, + 456, + 294, + 498 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 456, + 294, + 498 + ], + "spans": [ + { + "bbox": [ + 56, + 456, + 294, + 498 + ], + "type": "text", + "content": "[35] Saber Sheybani, Himanshu Hansaria, Justin Wood, Linda Smith, and Zoran Tiganj. Curriculum learning with infant egocentric videos. Advances in Neural Information Processing Systems, 36, 2024. 1, 2" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 56, + 501, + 294, + 533 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 501, + 294, + 533 + ], + "spans": [ + { + "bbox": [ + 56, + 501, + 294, + 533 + ], + "type": "text", + "content": "[36] Hans Stadthagen-Gonzalez and Colin J Davis. The bristol norms for age of acquisition, imageability, and familiarity. Behavior Research Methods, 38(4):598-605, 2006. 6, 1" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 56, + 535, + 294, + 578 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 535, + 294, + 578 + ], + "spans": [ + { + "bbox": [ + 56, + 535, + 294, + 578 + ], + "type": "text", + "content": "[37] Jessica Sullivan, Michelle Mei, Andrew Perfors, Erica Wojcik, and Michael C Frank. Saycam: A large, longitudinal audiovisual dataset recorded from the infant's perspective. Open mind, 5:20-29, 2021. 3, 5" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 56, + 579, + 294, + 622 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 579, + 294, + 622 + ], + "spans": [ + { + "bbox": [ + 56, + 579, + 294, + 622 + ], + "type": "text", + "content": "[38] Gentaro Taga, Kayo Asakawa, Atsushi Maki, Yukuo Konishi, and Hideaki Koizumi. Brain imaging in awake infants by near-infrared optical topography. Proceedings of the National Academy of Sciences, 100(19):10722-10727, 2003. 2" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 56, + 624, + 294, + 666 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 624, + 294, + 666 + ], + "spans": [ + { + "bbox": [ + 56, + 624, + 294, + 666 + ], + "type": "text", + "content": "[39] Satoshi Tsutsui, Arjun Chandrasekaran, Md Alimoor Reza, David Crandall, and Chen Yu. A computational model of early word learning from the infant's point of view. CogSci, 2020. 2" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 56, + 669, + 294, + 689 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 669, + 294, + 689 + ], + "spans": [ + { + "bbox": [ + 56, + 669, + 294, + 689 + ], + "type": "text", + "content": "[40] Alan M Turing. Computing machinery and intelligence. Mind, LIX(236):433-460, 1950. 2" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 56, + 692, + 294, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 692, + 294, + 713 + ], + "spans": [ + { + "bbox": [ + 56, + 692, + 294, + 713 + ], + "type": "text", + "content": "[41] Wai Keen Vong, Wentao Wang, A Emin Orhan, and Brenden M Lake. Grounded language acquisition through the" + } + ] + } + ], + "index": 16 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 316, + 73, + 553, + 205 + ], + "type": "list", + "angle": 0, + "index": 21, + "blocks": [ + { + "bbox": [ + 335, + 73, + 553, + 95 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 335, + 73, + 553, + 95 + ], + "spans": [ + { + "bbox": [ + 335, + 73, + 553, + 95 + ], + "type": "text", + "content": "eyes and ears of a single child. Science, 383(6682):504-511, 2024. 1, 2, 3, 5, 6, 7" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 316, + 96, + 553, + 150 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 96, + 553, + 150 + ], + "spans": [ + { + "bbox": [ + 316, + 96, + 553, + 150 + ], + "type": "text", + "content": "[42] Zeyu Wang, Berthy Feng, Karthik Narasimhan, and Olga Russakovsky. Towards unique and informative captioning of images. In Computer Vision-ECCV 2020: 16th European Conference, Glasgow, UK, August 23-28, 2020, Proceedings, Part VII 16, pages 629-644. Springer, 2020. 3" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 316, + 152, + 553, + 205 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 152, + 553, + 205 + ], + "spans": [ + { + "bbox": [ + 316, + 152, + 553, + 205 + ], + "type": "text", + "content": "[43] Saining Xie, Ross Girshick, Piotr Dólár, Zhuowen Tu, and Kaiming He. Aggregated residual transformations for deep neural networks. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 1492-1500, 2017. 5" + } + ] + } + ], + "index": 20 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "text", + "content": "4352" + } + ] + } + ], + "index": 22 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 9 + } + ], + "_backend": "vlm", + "_version_name": "2.6.4" +} \ No newline at end of file diff --git a/2025/Discrete to Continuous_ Generating Smooth Transition Poses from Sign Language Observations/4e14bfdb-c7fc-42ce-bc5c-af5b00971815_content_list.json b/2025/Discrete to Continuous_ Generating Smooth Transition Poses from Sign Language Observations/4e14bfdb-c7fc-42ce-bc5c-af5b00971815_content_list.json new file mode 100644 index 0000000000000000000000000000000000000000..d7a74f7d2cd007e6ac7121d2249195541883b821 --- /dev/null +++ b/2025/Discrete to Continuous_ Generating Smooth Transition Poses from Sign Language Observations/4e14bfdb-c7fc-42ce-bc5c-af5b00971815_content_list.json @@ -0,0 +1,1557 @@ +[ + { + "type": "text", + "text": "Discrete to Continuous: Generating Smooth Transition Poses from Sign Language Observations", + "text_level": 1, + "bbox": [ + 191, + 130, + 805, + 176 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Shengeng Tang Jiayi He Lechao Cheng* Jingjing Wu Dan Guo Richang Hong* \nSchool of Computer Science and Information Engineering, Hefei University of Technology \n{tangsg, chenglc, wujingjing, guodan, hongrc}@h Fut.edu.cn, hejy4396@mail.hfut.edu.cn", + "bbox": [ + 127, + 203, + 867, + 257 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Abstract", + "text_level": 1, + "bbox": [ + 246, + 291, + 326, + 306 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Generating continuous sign language videos from discrete segments is challenging due to the need for smooth transitions that preserve natural flow and meaning. Traditional approaches that simply concatenate isolated signs often result in abrupt transitions, disrupting video coherence. To address this, we propose a novel framework, Sign-D2C, that employs a conditional diffusion model to synthesize contextually smooth transition frames, enabling the seamless construction of continuous sign language sequences. Our approach transforms the unsupervised problem of transition frame generation into a supervised training task by simulating the absence of transition frames through random masking of segments in long-duration sign videos. The model learns to predict these masked frames by denoising Gaussian noise, conditioned on the surrounding sign observations, allowing it to handle complex, unstructured transitions. During inference, we apply a linearly interpolating padding strategy that initializes missing frames through interpolation between boundary frames, providing a stable foundation for iterative refinement by the diffusion model. Extensive experiments on the PHOENIX14T, USTCCSL100, and USTC-SLR500 datasets demonstrate the effectiveness of our method in producing continuous, natural sign language videos.", + "bbox": [ + 89, + 323, + 483, + 686 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "1. Introduction", + "text_level": 1, + "bbox": [ + 91, + 715, + 220, + 729 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Sign language [3, 9, 13, 15, 16, 22, 42, 54], as a unique form of visual communication, presents an intricate challenge for data collection, especially when scaling datasets for continuous, long-duration sign sentences. Unlike textual language, which can be assembled seamlessly from words to sentences, sign language requires a fluid connection between individual signs to convey meaning accurately. This connection involves not only the specific signs but also the transitions between them, which convey tem-", + "bbox": [ + 89, + 741, + 482, + 876 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/a34c3d1d814e76cbdfae282031a899ea45cf3fc0faaafb7adb1ff387663d6df7.jpg", + "image_caption": [ + "Figure 1. Task and key steps. Our work aims to generate continuous sign videos by creating transition poses between discrete segments. In training, random masking simulates missing transitions, and the model learns to recover these frames (steps 1-3). During inference, padding initializes missing transitions, which the model refines to generate smooth, coherent sequences (steps 4-6)." + ], + "image_footnote": [], + "bbox": [ + 517, + 290, + 903, + 534 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "poral and contextual coherence vital to understanding the intended message. As a result, while word-level sign language data [5, 7, 24, 53] has been collected extensively, there remains a critical gap in developing robust, large-scale datasets for sentence-level sign language [2, 8, 18], particularly for continuous sequences.", + "bbox": [ + 511, + 657, + 906, + 748 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "A possible workaround involves pre-collecting word-level sign data akin to a dictionary [1, 6, 20] and then attempting to form sentences by joining these discrete signs. However, assembling sign sentences is fundamentally different from assembling textual sentences due to the need for visually smooth transitions between signs, which maintain natural flow and meaning. Direct concatenation of individual signs into sentences often results in abrupt and unnatural transitions, disrupting the continuity of the video and potentially obscuring meaning. Consequently, generating smooth", + "bbox": [ + 511, + 750, + 908, + 902 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "CVF", + "bbox": [ + 106, + 2, + 181, + 42 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore.", + "bbox": [ + 236, + 0, + 810, + 46 + ], + "page_idx": 0 + }, + { + "type": "page_footnote", + "text": "*Corresponding authors.", + "bbox": [ + 109, + 887, + 241, + 898 + ], + "page_idx": 0 + }, + { + "type": "page_number", + "text": "3481", + "bbox": [ + 482, + 944, + 513, + 955 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "transition frames between sign segments is essential to synthesizing continuous sign language videos [41, 51, 52], yet presents significant challenges due to the lack of supervised reference data for the missing transition frames.", + "bbox": [ + 89, + 90, + 480, + 151 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "To address this issue, we propose a smooth transition pose generation framework, named Sign-D2C, to synthesize Continuous sign language videos from Discrete sign segments. This framework is based on a diffusion model, which learns to generate missing frames by observing surrounding sign frames. Specifically, our method leverages a conditional diffusion model trained to predict the missing transition frames by sampling from Gaussian noise and being conditioned by the observed video frames. As illustrated in Fig. 1, our method consists of a two-phase process designed to overcome the limitations of direct sign concatenation. In the training phase, we simulate the transition frame generation task by randomly masking frames within long sign language videos. This random masking effectively removes video segments, mimicking real-world scenarios where transition frames are absent. The model is then trained to reconstruct these missing frames based on the observable surrounding frames. By doing so, the model learns to generate plausible transitions in a supervised manner, enhancing its robustness and adaptability to complex, unstructured transitions. In the inference phase, we introduce a linearly interpolating padding strategy to initialize the missing transition frames. This strategy begins by selecting the last frame of the preceding observed segment and the first frame of the following observed segment as the boundary frames for the missing transition segment. The intermediate frames within the missing segment are then initialized by linearly interpolating between these boundary frames, creating an initial linearly interpolating padding that provides a stable foundation for generating smooth transitions. The conditional diffusion model then iteratively refines these frames to produce naturalistic transition frames that connect the observed segments seamlessly. The contributions of this work are summarized as follows:", + "bbox": [ + 91, + 151, + 486, + 667 + ], + "page_idx": 1 + }, + { + "type": "list", + "sub_type": "text", + "list_items": [ + "- We propose a novel conditional diffusion framework, Sign-D2C, specifically designed to synthesize smooth and contextually accurate transition frames for continuous sign videos, overcoming the limitations of traditional word-to-sentence assembly techniques in sign language.", + "- To simulate the generation of transition frames, we introduce a random masking strategy that removes sections of long sign language videos during training, enabling the model to predict missing frames in a supervised manner. This approach enhances the model's ability to handle diverse, complex transitions.", + "- We design a linearly interpolating padding strategy for transition initialization in the inference phase. This strategy utilizes the boundary frames of observed segments to initialize the missing poses with linear interpolation," + ], + "bbox": [ + 89, + 674, + 480, + 900 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "which provides a stable basis for generating visually seamless transitions.", + "bbox": [ + 526, + 90, + 903, + 119 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "- Extensive experiments on three standard benchmarks (i.e., Phoenix14T, UCST-CSL100, and USTC-SLR500) demonstrate the robustness and effectiveness of our approach in producing smooth, contextually accurate transition poses for continuous sign language videos.", + "bbox": [ + 513, + 121, + 903, + 196 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "2. Related Work", + "text_level": 1, + "bbox": [ + 513, + 212, + 653, + 228 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "2.1. Sign Language Production & Synthesis", + "text_level": 1, + "bbox": [ + 511, + 237, + 849, + 255 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Sign language plays a vital role in daily life and has become a prominent topic in artificial intelligence. Early research primarily focused on Sign Language Recognition (SLR) [10, 15, 16, 22, 48, 54] and Sign Language Translation (SLT) [3, 9, 12, 25]. However, with the growing need for effective communication between deaf and hearing individuals, Sign Language Production (SLP) has gained significant attention in recent years.", + "bbox": [ + 511, + 260, + 903, + 381 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Early approaches to SLP utilize synthetic animation techniques [28, 29, 37] to convert sentences into sign animations. These techniques, however, incur high collection costs and are limited to specific terms. With advancements in deep learning models, SLP has seen more extensive exploration. Stoll et al. [40] proposes a multi-step approach for generating sign language videos from text, involving the transformation of text into gloss, gloss into poses, and finally, poses into video. Building on this, Saunders et al. [35] introduces a model that generates sign language poses through autoregression. Diffusion-based solutions like G2P-DDM [47] and GCDM [43] further advance SLP by generating sign poses under semantic guidance.", + "bbox": [ + 511, + 382, + 905, + 579 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Despite significant progress in these methods, the limitations of existing data have become a major bottleneck in the development of sign language technology. Previous Sign Language Synthesis (SLS) approaches [51] focus on extracting gestures from continuous sign language videos. However, these methods still face challenges in synthesizing continuous, long-duration sign language sentences. Our method addresses this gap by starting with word-level sign language pose data, generating smooth transition frames between individual signs, and combining these segments into sentence-level sequences, thereby overcoming the limitations of existing methods.", + "bbox": [ + 511, + 579, + 905, + 760 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "2.2. Diffusion-based Generative Models", + "text_level": 1, + "bbox": [ + 511, + 771, + 820, + 787 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Diffusion models have gained considerable attention in generative modeling since they were first introduced by Sohl-Dickstein et al. [39]. Its forward-backward diffusion process has proven highly effective in generating realistic and diverse outputs, enabling diffusion models to achieve competitive performance in various generative tasks, including image synthesis [34], video generation [23, 49], and 3D shape", + "bbox": [ + 511, + 794, + 903, + 902 + ], + "page_idx": 1 + }, + { + "type": "page_number", + "text": "3482", + "bbox": [ + 482, + 944, + 514, + 955 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "generation [31]. One of the strengths of diffusion models is their iterative refinement capability, allowing for flexible and high-quality outputs that align well with tasks requiring fine control over generative processes. This quality has led to their adoption in areas such as image restoration, where they have been used for denoising and superresolution [30, 45, 50], as well as pose estimation, where they can handle uncertainty in human poses [38]. Recent advancements have extended their use to more complex generative tasks, such as controlled text-to-image synthesis [46] and conditional video generation [32], further demonstrating their adaptability and robustness.", + "bbox": [ + 89, + 90, + 480, + 272 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "In Sign Language Production (SLP), diffusion models remain relatively unexplored but have shown promise in recent studies. For instance, G2P-DDM [47] uses a discrete diffusion model to translate the continuous pose generation task into a discrete space, simplifying the complexity of generating sign language poses while maintaining semantic alignment with gloss-based inputs. This approach allows for finer control over pose accuracy, making it particularly suitable for generating static sign poses from linguistic cues. Similarly, GCDM [43] applies a multi-hypothesis diffusion framework to generate diverse sign poses from gloss, accommodating multiple interpretations and improving the naturalness of synthesized gestures. These models represent pioneering steps in applying diffusion-based approaches to SLP, leveraging the unique properties of diffusion models to generate realistic sign poses under semantic guidance.", + "bbox": [ + 91, + 291, + 483, + 534 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Despite these advancements, current diffusion-based methods for SLP are primarily focused on generating sign poses according to textual signals rather than addressing the temporal continuity required for longer sign language sequences. In sign language synthesis, especially when producing smooth transitions in continuous sign language videos, maintaining smooth transitions between gestures is essential for ensuring that the sequences are fluid and natural. This need for temporally smooth transitions has led to research on methods that can generate these connecting frames between segments, although it remains an area with limited exploration in SLP. Beyond SLP, diffusion models have been successfully applied to tasks requiring temporal coherence and frame interpolation in video synthesis. For example, diffusion-based models have been employed in video generation to ensure smooth transitions between frames [26], and similar techniques have been explored in action synthesis and motion completion [21], where maintaining continuity across frames is crucial. Such applications highlight the potential of diffusion models to handle tasks involving complex temporal dependencies, suggesting that similar techniques could be valuable in SLP for generating continuous sign language videos.", + "bbox": [ + 91, + 553, + 483, + 901 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3. Methodology", + "text_level": 1, + "bbox": [ + 513, + 89, + 648, + 107 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "In this work, we propose a framework for generating smooth transition frames between discrete sign segments to create continuous sign pose videos. Our approach transforms an unsupervised transition generation task into a supervised framework using a random masking strategy based on conditional diffusion. This section will outline the preliminaries (Sec. 3.1), followed by our core idea of realizing the task (Sec. 3.2), the framework structure (Sec. 3.3), and the training process (Sec. 3.4).", + "bbox": [ + 511, + 114, + 906, + 252 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3.1. Preliminary", + "text_level": 1, + "bbox": [ + 511, + 262, + 643, + 277 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "The diffusion model [14] is highly effective in generating high-quality images by combining diffusion processes, latent space modeling, and conditional inputs. In this work, we employ a conditional diffusion model to synthesize smooth transition frames between discrete sign segments, forming continuous sign language videos. Here, the model is trained on complete long-duration sign language pose sequences, progressively adding noise to simulate missing transitions. During testing, the model starts from Gaussian noise samples and generates the required transitions between observed sign frames.", + "bbox": [ + 511, + 285, + 906, + 450 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "In the forward process, noise is incrementally added to the full sign language pose sequence $x_0$ to obtain a noisy version $x_T$ , mimicking the degradation of visual continuity. This process is described by:", + "bbox": [ + 511, + 452, + 905, + 513 + ], + "page_idx": 2 + }, + { + "type": "equation", + "text": "\n$$\nx _ {t} = \\alpha_ {t} \\cdot x _ {t - 1} + \\sqrt {1 - \\alpha_ {t}} \\cdot \\epsilon_ {t}, \\tag {1}\n$$\n", + "text_format": "latex", + "bbox": [ + 586, + 523, + 903, + 542 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "where $\\alpha_{t}$ controls the level of noise at each step $t$ , and $\\epsilon_{t}$ represents Gaussian noise. As $t$ increases, the original sequence $x_0$ transitions to pure noise $x_{T}$ .", + "bbox": [ + 511, + 555, + 905, + 599 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "The reverse process reconstructs the continuous sequence by progressively removing noise, where a neural network is trained to predict each denoised frame based on previous noisy frames. The denoising step at each iteration is defined as:", + "bbox": [ + 511, + 601, + 905, + 675 + ], + "page_idx": 2 + }, + { + "type": "equation", + "text": "\n$$\np \\left(x _ {t - 1} \\mid x _ {t}, x _ {\\text {o b s}}\\right) = \\mathcal {N} \\left(x _ {t - 1}; \\mu_ {\\theta} \\left(x _ {t}, t, x _ {\\text {o b s}}\\right), \\Sigma_ {t}\\right), \\tag {2}\n$$\n", + "text_format": "latex", + "bbox": [ + 531, + 689, + 903, + 705 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "where $x_{\\mathrm{obs}}$ denotes the observed sign language frames that serve as conditional inputs, and $\\mu_{\\theta}(x_t,t,x_{\\mathrm{obs}})$ is the predicted mean for the denoising process. The observed frames condition the diffusion model, guiding it to produce smooth, contextually consistent transition frames.", + "bbox": [ + 511, + 719, + 905, + 792 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "In our model, the training input is a complete long-duration sign language pose video, gradually noised to simulate missing transitions. During testing, Gaussian noise serves as the input, and the model generates a continuous video with smooth transition frames conditioned on observed frames $x_{\\mathrm{obs}}$ , effectively bridging discrete sign segments into a fluid sign language sequence.", + "bbox": [ + 511, + 795, + 905, + 901 + ], + "page_idx": 2 + }, + { + "type": "page_number", + "text": "3483", + "bbox": [ + 482, + 944, + 514, + 955 + ], + "page_idx": 2 + }, + { + "type": "image", + "img_path": "images/10f8ad0591c27acc5272163fac9914faf71f533ade0ba56830fcad116f74e3bf.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 96, + 90, + 898, + 210 + ], + "page_idx": 3 + }, + { + "type": "image", + "img_path": "images/eb02065371f8d11e017dca2ef7790a6ca4f0f5e84baa494082c6b1e828f7207d.jpg", + "image_caption": [ + "Figure 2. Overview. The proposed framework for generating continuous sign language videos with smooth transitions between discrete segments. In the training phase (top), a long-duration sign video undergoes random masking to create gaps, simulating missing transitions. An encoder processes the video, and a conditional diffusion model denoises and predicts the masked segments based on observable frames. The recovered video is produced through a decoder, with both encoder and decoder pre-trained for effective representation. In the inference phase (bottom), discrete sign videos are connected using a linearly interpolating padding strategy to initialize missing transitions. Gaussian noise is applied, and the model refines these transitions, generating a continuous sign video with coherent flow." + ], + "image_footnote": [], + "bbox": [ + 96, + 215, + 898, + 329 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "3.2. Core Idea: Unsupervised to Supervised", + "text_level": 1, + "bbox": [ + 89, + 453, + 428, + 469 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Generating continuous sign language videos by adding transitions between discrete segments is challenging, as these segments typically start and end with specific gestures that make direct concatenation impractical. Additionally, this transition generation task is inherently unsupervised, as real transition frames are generally unavailable for training. To address this, we convert the unsupervised problem of generating transitions into a supervised learning framework by leveraging existing continuous sign language videos.", + "bbox": [ + 89, + 486, + 482, + 622 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Given a complete continuous sign language sequence $X = \\{x_{1}, x_{2}, \\dots, x_{n}\\}$ , we simulate missing transitions by randomly masking portions of $X$ , creating gaps that mimic real-world transition absences. Let the masked frames be denoted as $M \\subset X$ , and the observable pose frames as $X_{\\mathrm{obs}} = (X \\setminus X_{\\mathrm{mask}})$ . This setup frames the task as predicting $X_{\\mathrm{mask}}$ based on $X_{\\mathrm{obs}}$ , effectively generating supervised training data for transition prediction.", + "bbox": [ + 89, + 633, + 482, + 753 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "By training the model to predict the masked frames $X_{\\mathrm{mask}}$ conditioned on the observable context $X_{\\mathrm{obs}}$ , it learns to generate realistic transitions that smoothly bridge the gap. This innovative masking strategy allows our model to leverage the context of surrounding frames provided by $X_{\\mathrm{obs}}$ to produce temporally coherent connections, transforming the inherently unsupervised transition generation task into a supervised one and enabling effective continuous sign language synthesis.", + "bbox": [ + 89, + 763, + 482, + 901 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "3.3. Sign-D2C Framework", + "text_level": 1, + "bbox": [ + 513, + 453, + 718, + 469 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Our approach leverages a conditional diffusion model to generate transition frames, enabling smooth continuity between discrete sign segments. The model architecture consists of three main components: an encoder, a decoder, and a denoiser. The first two are composed of two self-attention layers and the denoiser consists of two blocks, each containing a self-attention layer, a cross-attention layer, and a feed-forward layer. This setup allows the model to effectively capture temporal dependencies and complex contextual information needed for sign language synthesis.", + "bbox": [ + 511, + 474, + 906, + 626 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "The Encoder first processes the input sequence $X = \\{x_{1}, x_{2}, \\ldots, x_{n}\\}$ , where $x_{i}$ represents individual sign language pose frames. By encoding the long-duration video input, the Encoder transforms $X$ into a latent representation $Z$ , which serves as the foundation for both the diffusion and denoising processes:", + "bbox": [ + 511, + 626, + 906, + 717 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\nZ = \\operatorname {E n c o d e r} (X). \\tag {3}\n$$\n", + "text_format": "latex", + "bbox": [ + 630, + 723, + 903, + 739 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "During the forward diffusion, Gaussian noise is gradually added to $Z$ at each time step $t$ , resulting in a noisy representation $Z_{t}$ . This process can be transferred from Eq. 1, formulated as:", + "bbox": [ + 511, + 746, + 903, + 805 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\nZ _ {t} = \\alpha_ {t} \\cdot Z _ {t - 1} + \\sqrt {1 - \\alpha_ {t} ^ {2}} \\cdot \\epsilon_ {t}, \\tag {4}\n$$\n", + "text_format": "latex", + "bbox": [ + 584, + 810, + 903, + 835 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "where $\\alpha_{t}$ controls the noise schedule, and $\\epsilon_{t}$ is sampled from a Gaussian distribution. The diffusion process progressively transforms the latent representation towards pure noise.", + "bbox": [ + 511, + 840, + 906, + 898 + ], + "page_idx": 3 + }, + { + "type": "page_number", + "text": "3484", + "bbox": [ + 482, + 944, + 514, + 955 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "The denoiser then reconstructs the original representation by iteratively removing the noise. Here, the denoiser is conditioned on sign observations $X_{\\mathrm{obs}}$ , which serve as contextual priors. These observations are incorporated into the denoiser via cross-attention, allowing the model to focus on the relevant parts of $X_{\\mathrm{obs}}$ at each denoising step. For each time step $t$ in the reverse process, the denoised output $\\hat{Z}_{t - 1}$ is given by:", + "bbox": [ + 89, + 90, + 483, + 212 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\hat {Z} _ {t - 1} = \\operatorname {D e n o i s e r} \\left(Z _ {t}, X _ {\\mathrm {o b s}}\\right), \\tag {5}\n$$\n", + "text_format": "latex", + "bbox": [ + 174, + 220, + 482, + 241 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "where $X_{\\mathrm{obs}}$ provides the model with the necessary context to predict the masked segments accurately.", + "bbox": [ + 89, + 250, + 482, + 280 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Finally, the decoder, also structured with three self-attention layers, takes the denoised latent representation and transforms it back into the original sign language video space, outputting the recovered or synthesized video $\\hat{X}$ :", + "bbox": [ + 89, + 280, + 483, + 340 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\hat {X} = \\operatorname {D e c o d e r} \\left(\\hat {Z} _ {0}\\right). \\tag {6}\n$$\n", + "text_format": "latex", + "bbox": [ + 204, + 349, + 482, + 368 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "By using self-attention within both the encoder and decoder, and cross-attention in the denoiser to incorporate contextual observations, the model effectively learns to generate temporally coherent transitions, resulting in a smooth, continuous sign language video output.", + "bbox": [ + 89, + 378, + 483, + 455 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "The proposed Sign-D2C incorporates two key strategies to enhance transition generation: a random masking strategy during training and linearly interpolating padding initialization for inference. The following provides details.", + "bbox": [ + 89, + 454, + 483, + 515 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Random Masking Strategy. To generate smooth transitional poses and capture relationships between different segments, we employ a random masking strategy during training, as illustrated in the training phase of Fig. 2.", + "bbox": [ + 89, + 515, + 483, + 575 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Given a long-duration video $p \\in \\mathbb{R}^{F \\times (N \\times 3)}$ with $F$ frames, each containing $N$ joints in 3D space, the video is first encoded by a pre-trained encoder into a latent representation $x \\in \\mathbb{R}^{F \\times W}$ , where $W$ denotes the dimension of the latent feature $x$ . The mask $m$ is constructed by randomly selecting frame indices, extending across the dimensions of the latent feature $x$ . This mask $m$ is then applied to selectively retain or mask out parts of $x$ , resulting in a masked latent representation:", + "bbox": [ + 89, + 575, + 483, + 712 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\nx ^ {m} = x \\odot m, x \\in \\mathbb {R} ^ {F \\times W}, \\tag {7}\n$$\n", + "text_format": "latex", + "bbox": [ + 179, + 720, + 482, + 739 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "where $x^{m}$ serves as a conditional input to the model, distinguishing between occluded and visible movements. This masking strategy enables the model to predict missing frames based on observable context, facilitating smooth transition generation.", + "bbox": [ + 89, + 750, + 482, + 824 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Linear Interpolation Padding. In the inference phase, generating smooth transitions between discrete motion segments is challenging due to the gaps between observed frames. We leverage the natural correlation between consecutive frames in sign language sequences by introducing", + "bbox": [ + 89, + 825, + 483, + 901 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "a linear interpolation padding initialization strategy, as illustrated in Fig. 2.", + "bbox": [ + 511, + 90, + 903, + 121 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "This strategy starts by using the last pose of the preceding observed segment, $p_{s}$ , and the first pose of the following observed segment, $p_{e}$ , as boundary frames for the missing transition segment. The goal is to generate intermediate frames $p_{c}$ that provide a smooth transition from $p_{s}$ to $p_{e}$ . Using linear interpolation, we initialize these missing frames as follows:", + "bbox": [ + 511, + 121, + 905, + 226 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\left\\{ \\begin{array}{l} f (i) = \\sum_ {k = 0} ^ {i} \\frac {1}{(k + 1)}, \\quad i = 0, \\dots , (e - s), \\\\ p _ {c} (i) = p _ {s} + \\frac {p _ {e} - p _ {s}}{e - s + 1} \\cdot f (i). \\end{array} \\right. \\tag {8}\n$$\n", + "text_format": "latex", + "bbox": [ + 534, + 229, + 903, + 271 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "This interpolation scheme gradually fills in the key points at each time step, creating intermediate poses that ensure a coherent and natural flow between segments. This strategy facilitates smoother transitions and provides a stable foundation for later refinement, resulting in seamless sign language video synthesis.", + "bbox": [ + 511, + 276, + 906, + 367 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "3.4. Training", + "text_level": 1, + "bbox": [ + 513, + 375, + 617, + 390 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Pre-training of Encoder & decoder. To enable coherent video generation, we pre-train the encoder and decoder to ensure effective latent representation processing. The encoder extracts latent features $x$ from input videos, while the decoder reconstructs realistic sign language poses from these latent representations. For training the diffusion model, we freeze the encoder and decoder parameters, allowing the model to focus on learning smooth transitions based on these pre-trained features.", + "bbox": [ + 511, + 396, + 905, + 532 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Pre-training is conducted end-to-end with a reconstruction loss, optimizing the encoder-decoder pair to accurately represent and recover video details:", + "bbox": [ + 511, + 532, + 905, + 578 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {L} _ {\\text {p r e}} = \\frac {1}{F} \\sum_ {f = 1} ^ {F} | p _ {f} - \\hat {p} _ {f} |, \\tag {9}\n$$\n", + "text_format": "latex", + "bbox": [ + 609, + 584, + 903, + 627 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "where $p$ represents the input video with $F$ frames, and $\\hat{p}$ is the reconstructed output. This process ensures the encoder captures essential features, while the decoder faithfully reconstructs realistic sign language poses, setting a solid foundation for high-quality transition generation.", + "bbox": [ + 511, + 633, + 905, + 708 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Loss Optimization. According to Eq. 5, we can obtain the denoiser prediction $\\hat{X} = \\{\\hat{x}_1,\\hat{x}_2,\\dots ,\\hat{x}_F\\}$ . During training, the loss function constrains the difference between the original latent feature $x$ and the denoiser's prediction $\\hat{x}$ :", + "bbox": [ + 511, + 709, + 905, + 770 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {L} _ {\\text {d i f f}} = \\frac {1}{F} \\sum_ {f = 1} ^ {F} | x _ {f} - \\hat {x} _ {f} |, \\tag {10}\n$$\n", + "text_format": "latex", + "bbox": [ + 607, + 776, + 903, + 819 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "where $F$ denotes the number of frames in the video sequence. This loss encourages the model to accurately recover the original latent representation, improving the quality of the generated transitions and ensuring alignment with the encoded video features.", + "bbox": [ + 511, + 824, + 905, + 898 + ], + "page_idx": 4 + }, + { + "type": "page_number", + "text": "3485", + "bbox": [ + 482, + 944, + 514, + 955 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "4. Experiments", + "text_level": 1, + "bbox": [ + 89, + 89, + 223, + 107 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "4.1. Experimental Settings", + "text_level": 1, + "bbox": [ + 89, + 114, + 297, + 132 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Datasets. We conduct the experiments on three datasets, i.e., PHOENIX14T, USTC-CSL100, and USTC-SLR500. The first two are sentence-level continuous sign language datasets, which are used to evaluate supervised transition generation in controlled settings. While the latter one is a word-level isolated sign language dataset, allowing us to assess the model's performance in generating transitions within real-world, discrete sign scenarios.", + "bbox": [ + 89, + 138, + 482, + 261 + ], + "page_idx": 5 + }, + { + "type": "list", + "sub_type": "text", + "list_items": [ + "- PHOENIX14T [2] comprises 8,257 sentence-level samples, which covers 2,887 German words. For preprocessing, we follow the settings in previous sign language production studies [19, 36], using OpenPose [4] to extract 2D coordinates from the videos, which are then refined to 3D coordinates through a skeleton correction model [51].", + "- USTC-CSL100 [18] covers 100 unique sentences performed by 50 signers. We use the split-I [11] approach to divide the dataset into 4,000 training examples and 1,000 test examples. The above two datasets enable us to evaluate the model's effectiveness on continuous sign language sequences in a structured environment.", + "- USTC-SLR500 [17] is an isolated sign language dataset containing 500 Chinese sign words, each performed five times by 50 signers. Following [17], we split the dataset by using videos from 36 signers for training and the remaining 14 signers for testing." + ], + "bbox": [ + 89, + 262, + 482, + 520 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Implementation Details. Since generating continuous sign language videos from discrete fragments is inherently an unsupervised task, there are no standardized evaluation metrics. To effectively assess our model's performance, we analyzed transition actions between two gestures on the sentence-level datasets (i.e., PHOENIX14T and USTC-CSL100), observing that transitions typically span 10 to 20 frames. Based on these findings, we designed experiments using two masking strategies: masking 10 frames every 20 frames and masking 20 frames every 10 frames. Additionally, to further validate the effectiveness of our approach, we used the isolated sign language dataset (i.e., USTC-SLR500) to assemble discrete vocabulary videos into continuous sequences, demonstrating the model's capability to produce smooth transitions in a real-world setting.", + "bbox": [ + 88, + 521, + 482, + 748 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Evaluation Metrics. We evaluate our performance on two sentence-level datasets in terms of both semantics and coherence. Following prior work in sign language generation [19, 35, 36, 44], we use the NSLT method [2] to convert generated sign language videos back into text. We then compare this text with reference data to calculate semantic accuracy metrics, including BLEU [33], ROUGE [27], and WER. Additionally, we use FID, DTW, and MPJPE to assess the consistency and smoothness of the generated transitions, measuring the coherence of the results.", + "bbox": [ + 89, + 750, + 482, + 901 + ], + "page_idx": 5 + }, + { + "type": "table", + "img_path": "images/5472886423c56c48a382123e04956c58b2a519e9451a6d2bd1bbde7f8cd15c76.jpg", + "table_caption": [ + "Table 1. Ablation results of different masking ratios on PHOENIX14T (observation 20 frames - prediction 10 frames)." + ], + "table_footnote": [], + "table_body": "
Masking Ratio rDEVTEST
B1↑B4↑DTW↓B1↑B4↑DTW↓
r = 0.122.178.040.6122.478.670.60
r = 0.322.468.060.6222.368.730.61
r = 0.522.308.060.7722.028.350.75
", + "bbox": [ + 517, + 117, + 903, + 193 + ], + "page_idx": 5 + }, + { + "type": "table", + "img_path": "images/8c731ebddf3fe11838b94b9a4645b5deb2da69b7aaeff6e918b4486abc66c580.jpg", + "table_caption": [ + "Table 2. Ablation results of different padding strategies on PHOENIX14T (observation 20 frames - prediction 10 frames)." + ], + "table_footnote": [], + "table_body": "
Padding StrategyDEVTEST
DTW↓FID↓MPJPE↓DTW↓FID↓MPJPE↓
w/o Padding11.462.7469.6011.603.0687.57
Front Padding0.722.6847.900.692.9960.05
Linear Padding0.622.6847.300.612.9859.36
Back Padding0.672.6847.890.662.9959.91
", + "bbox": [ + 517, + 231, + 903, + 311 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "4.2. Ablation Study", + "text_level": 1, + "bbox": [ + 511, + 330, + 666, + 345 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "In this subsection, we explore the effects of different random masking ratios in the training phase and various padding strategies during inference. The experiments in this part are all conducted under the condition of predicting 10 transition frames for every 20 observed sign frames to evaluate their impact on both semantic accuracy and transition coherence.", + "bbox": [ + 511, + 353, + 906, + 458 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Effect of Masking Ratio. Table 1 presents the results of different random masking ratios $r$ on semantic preservation and temporal coherence. With a lower masking ratio $(r = 0.1)$ , the model achieves the best coherence, reflected by the lowest DTW scores of 0.61 on DEV and 0.60 on TEST. This indicates that smaller masking ratios allow the model to maintain smoother transitions, as more contextual frames are available to guide the generation of missing frames. However, as the masking ratio increases to $r = 0.3$ , we observe a slight improvement in semantic accuracy, with BLEU-4 scores rising to 8.06 on DEV and 8.73 on TEST. This suggests that a higher masking ratio may encourage the model to focus more on the sign semantic content, as it learns to reconstruct more extensive missing segments.", + "bbox": [ + 511, + 460, + 908, + 672 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Interestingly, further increasing the masking ratio to $r = 0.5$ leads to a drop in both semantic accuracy and coherence, with DTW increasing and BLEU decreasing slightly. This highlights a trade-off: while a moderate masking ratio improves semantic preservation, excessive masking may reduce the model's ability to maintain smooth transitions and semantic accuracy. Based on these observations, we select $r = 0.3$ for subsequent experiments, as it provides a good balance between semantic retention and transition smoothness.", + "bbox": [ + 511, + 672, + 908, + 808 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Influence of Padding Strategy. Table 2 summarizes the results of different padding strategies applied during inference. We experiment with four approaches: w/o Padding (no initialization for missing frames), Front Padding (using the last frame of the previous observation segment), Back Padding (using the first frame of the following observation", + "bbox": [ + 511, + 809, + 908, + 901 + ], + "page_idx": 5 + }, + { + "type": "page_number", + "text": "3486", + "bbox": [ + 482, + 944, + 514, + 955 + ], + "page_idx": 5 + }, + { + "type": "table", + "img_path": "images/e2ca61446809ef5dfc7c7b0c89a596c49bdb264a185d9e7940d97d04c01e6d78.jpg", + "table_caption": [ + "Table 3. Performance comparison on PHOENIX14T (under setting: observation 20 frames - prediction 10 frames)." + ], + "table_footnote": [], + "table_body": "
MethodsDEVTEST
B1↑B4↑ROUGE↑WER↓DTW↓FID↓MPIJE↓B1↑B4↑ROUGE↑WER↓DTW↓FID↓MPIJE↓
Ground Truth29.7712.1329.6074.170.000.000.0029.7611.9328.9871.940.000.000.00
G2P-DDM [47]8.250.997.1798.9312.712.8555.029.521.298.1298.5412.573.1967.81
VQ-GCDM [43]9.310.877.5998.6912.112.8653.879.850.967.9598.5911.983.2166.51
Ours22.468.0622.0780.730.622.6847.3022.368.7322.3679.060.612.9859.36
", + "bbox": [ + 94, + 104, + 903, + 202 + ], + "page_idx": 6 + }, + { + "type": "table", + "img_path": "images/954bc67869cdb04a76a2cb068cb8ad5d0ced6ca672a04c7e6a221232b36c5387.jpg", + "table_caption": [ + "Table 4. Performance comparison on PHOENIX14T (under setting: observation 10 frames - prediction 20 frames)." + ], + "table_footnote": [], + "table_body": "
MethodsDEVTEST
B1↑B4↑ROUGE↑WER↓DTW↓FID↓MPJPE↓B1↑B4↑ROUGE↑WER↓DTW↓FID↓MPJPE↓
Ground Truth29.7712.1329.6074.170.000.000.0029.7611.9328.9871.940.000.000.00
G2P-DDM [47]8.980.637.4498.6115.952.8955.219.130.747.7598.6415.583.2268.18
VQ-GCDM [43]8.640.897.0898.4812.403.0548.669.021.367.2298.6112.223.3760.27
Ours19.325.9218.3387.300.702.7548.0419.596.2618.5986.80.683.0559.81
", + "bbox": [ + 94, + 222, + 903, + 321 + ], + "page_idx": 6 + }, + { + "type": "table", + "img_path": "images/ffd62c6eb05b8f9275a1998af7519eeef62ae08e5107721f3f74a043252e4a17.jpg", + "table_caption": [ + "Table 5. Performance comparison on USTC-CSL100." + ], + "table_footnote": [], + "table_body": "
MethodsObservation 20 Frames - Prediction 10 FramesObservation 10 Frames - Prediction 20 Frames
B1↑B4↑ROUGE↑WER↓DTW↓FID↓MPJPE↓B1↑B4↑ROUGE↑WER↓DTW↓FID↓MPJPE↓
Ground Truth69.1059.1268.5347.380.000.000.0069.1059.1268.5347.380.000.000.00
G2P-DDM [47]13.742.1113.8592.047.000.53258.4313.431.7413.6292.157.470.53258.43
VQ-GCDM [43]30.7110.6829.1777.875.100.4391.1018.772.0816.6492.6710.890.54104.27
Ours69.6458.9869.2646.661.920.3177.0467.9556.6167.3849.241.790.3178.62
", + "bbox": [ + 94, + 339, + 903, + 438 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "segment), and Linear Padding (performing linear interpolation between the two boundary poses, as Eq. 8). The \"w/o Padding\" approach yields the highest DTW scores (11.46 - DEV and 10.36 - TEST), indicating poor coherence and unnatural transitions. Both Front Padding and Back Padding improve the smoothness of transitions, as shown by lower DTW scores (0.67 - DEV and 0.66 - TEST), but their FID and MPJPE remain better compared to Linear Padding.", + "bbox": [ + 88, + 453, + 480, + 574 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "The Linear Padding strategy achieves the best overall performance, with the lowest DTW scores (0.62 on DEV and 0.66 on TEST) and MPJPE values, reflecting both smoothness and accuracy in pose reconstruction. By initializing missing frames through linear interpolation, Linear Padding provides a stable and effective foundation for generating realistic and seamless transitions, closely resembling real-world signing behavior.", + "bbox": [ + 88, + 575, + 480, + 696 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "4.3. Comparison with State-of-the-Arts", + "text_level": 1, + "bbox": [ + 89, + 709, + 393, + 724 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "We compare our method with two state-of-the-art models: G2P-DDM [47] and VQ-GCDM [43]. Both G2P-DDM and VQ-GCDM are primarily designed for SLP. Due to the absence of established benchmarks for our specific task, we adapted these models to serve as comparative baselines.", + "bbox": [ + 89, + 732, + 482, + 808 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Comparison on PHOENIX14T. As shown in Table 4, under the \"Observation 20 Frames - Prediction 10 Frames\" setting, our method achieves superior performance across both semantic and coherence metrics. Specifically, we attain a BLEU-1 score of $22.46\\%$ and $22.36\\%$ on the DEV and TEST sets, respectively, outperforming G2P-DDM and", + "bbox": [ + 88, + 809, + 482, + 901 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "VQ-GCDM. This improvement in BLEU scores indicates that our method more accurately preserves semantic content compared to existing methods. Furthermore, our method shows significant gains in temporal coherence, with DTW scores of 0.62 on DEV and 0.61 on TEST, respectively. These results confirm that our model generates smoother and more realistic transitions than the baseline methods.", + "bbox": [ + 511, + 453, + 903, + 559 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Under the more challenging \"Observation 10 Frames - Prediction 20 Frames\" setting, our method continues to demonstrate robust performance despite the increased prediction length. We achieve BLEU-1 scores of $19.32\\%$ on DEV and $19.59\\%$ on TEST, considerably higher than the baseline methods. Our method also maintains lower DTW and FID scores, demonstrating its ability to generate coherent and semantically accurate sequences even with limited observed data. The MPJPE scores further support the effectiveness of our approach, with scores of 48.04 on DEV and 59.81 on TEST, indicating a closer alignment to ground truth poses than the other methods.", + "bbox": [ + 511, + 564, + 906, + 743 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Comparison on USTC-CSL100. Table 5 shows the results on the USTC-CSL100 dataset, which evaluates our method's ability to generate sign language sequences. In the \"Observation 20 Frames - Prediction 10 Frames\" setting, our model achieves near-ground-truth semantic accuracy, with BLEU-1 and BLEU-4 scores of $69.64\\%$ and $58.98\\%$ on DEV, closely approaching the ground truth scores of $69.10\\%$ and $59.12\\%$ . Additionally, our DTW score of 1.922 and FID score of 3.21 indicate high temporal consistency and low distributional divergence from the real data. Com", + "bbox": [ + 511, + 750, + 908, + 900 + ], + "page_idx": 6 + }, + { + "type": "page_number", + "text": "3487", + "bbox": [ + 482, + 944, + 514, + 955 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/cad7faac2655967240f13d92e536dbe0310e82462e52c30edee06bdaf17c4681.jpg", + "image_caption": [ + "Figure 3. Visualization examples of generating 10-frame transition pose under 20-frame observations on PHOENIX14T. We compare our method with G2P-DDM and VQ-GCDM, and the Ground Truth is attached. Green: Transitions. Orange and Blue: Observations." + ], + "image_footnote": [], + "bbox": [ + 94, + 89, + 903, + 251 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/b904cb3598f50b51614234e478fd5f038d405c24b8a61c541dbc538df5cf9062.jpg", + "image_caption": [ + "Figure 4. Visualization examples on the USTC-SLR500, demonstrate the generation of transition poses according to two discrete sign segments. Green: Transitions. Orange and Blue: Observations." + ], + "image_footnote": [], + "bbox": [ + 94, + 297, + 901, + 455 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "pared to G2P-DDM and VQ-GCDM, which yield higher DTW and MPJPE scores, our model achieves smoother and more accurate pose transitions.", + "bbox": [ + 89, + 513, + 482, + 559 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "In the \"Observation 10 Frames - Prediction 20 Frames\" scenario, our method continues to demonstrate resilience in generating long sequences. Our BLEU-1 and BLEU-4 scores on DEV are $67.95\\%$ and $56.61\\%$ , respectively, showing superior semantic preservation compared to the baseline methods. Additionally, the DTW score of 1.79 and MPJPE score of 78.62 confirm our model's ability to produce coherent and realistic transitions, outperforming the baselines in alignment with true pose sequences.", + "bbox": [ + 89, + 561, + 482, + 696 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "4.4. Qualitative Results", + "text_level": 1, + "bbox": [ + 89, + 710, + 272, + 726 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Visualization Results. Figure 3 presents a comparative demonstration of our method against other approaches (i.e., G2P-DDM [47] and VQ-GCDM [43]) in generating transition sign poses. This example clearly shows that our method significantly outperforms G2P-DDM and VQ-GCDM in the quality of interpolated movements, especially excelling in overall motion coherence. Furthermore, our method meticulously fine-tunes the observations (the Blue and Orange sections), effectively avoiding coherence issues that may arise from directly concatenating discrete sign segments.", + "bbox": [ + 89, + 733, + 482, + 883 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Verification in Realistic Scenarios. Figure 4 demonstrates", + "bbox": [ + 89, + 885, + 482, + 900 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "visualization examples for synthesizing continuous sign videos from discrete segments in a realistic, word-level scenario using the USTC-SLR500 dataset. Here, the observed frames from two isolated sign segments are marked in blue and orange, while the generated transitions are shown in green. The results show that our method not only generates smooth transitions between discrete sign language segments but also fine-tunes the observed poses based on the overall semantics, resulting in a coherent sign language video.", + "bbox": [ + 511, + 515, + 906, + 651 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "5. Conclusions", + "text_level": 1, + "bbox": [ + 513, + 675, + 638, + 690 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "This paper proposes a method to generate continuous sign language videos by synthesizing transition frames between discrete segments, transforming an unsupervised task into a supervised one through random masking. Using a conditional diffusion model, the approach learns to predict missing frames based on observable context, allowing for smooth and temporally coherent outputs. By incorporating a linearly interpolating padding strategy during inference, the model initializes transitions effectively, further refining them into seamless sequences. Extensive experiments demonstrate the effectiveness of leveraging context for natural transitions, advancing sign language synthesis by enabling fluid, continuous sequences.", + "bbox": [ + 511, + 704, + 906, + 900 + ], + "page_idx": 7 + }, + { + "type": "page_number", + "text": "3488", + "bbox": [ + 482, + 944, + 514, + 955 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "6. Acknowledgments", + "text_level": 1, + "bbox": [ + 91, + 90, + 269, + 107 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "This work was supported by the National Natural Science Foundation of China (Grants No. U23B2031, 62472139, 62302142, 62272144), the Anhui Provincial Natural Science Foundation, China (Grant No. 2408085QF191), the Fundamental Research Funds for the Central Universities (Grants No. JZ2024HGTA0178, JZ2023HGQA0097), and the Major Project of Anhui Province (Grant No. 202423k09020001).", + "bbox": [ + 89, + 114, + 483, + 234 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "References", + "text_level": 1, + "bbox": [ + 91, + 247, + 187, + 263 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[1] Ahmed H Aliwy and A Alethary Ahmed. Development of arabic sign language dictionary using 3d avatar technologies. Indonesian Journal of Electrical Engineering and Computer Science, 21(1):609-616, 2021. 1", + "[2] Necati Cihan Camgoz, Simon Hadfield, Oscar Koller, Hermann Ney, and Richard Bowden. Neural sign language translation. In Computer Vision and Pattern Recognition, pages 7784-7793, 2018. 1, 6", + "[3] Necati Cihan Camgoz, Oscar Koller, Simon Hadfield, and Richard Bowden. Multi-channel transformers for multi-articulatory sign language translation. In European Conference on Computer Vision, pages 301-319, 2020. 1, 2", + "[4] Zhe Cao, Tomas Simon, Shih-En Wei, and Yaser Sheikh. Realtime multi-person 2d pose estimation using part affinity fields. In Computer Vision and Pattern Recognition, pages 7291-7299, 2017. 6", + "[5] Xiujuan Chai, Hanjie Wang, and Xilin Chen. The devisesign large vocabulary of Chinese sign language database and baseline evaluations. In Technical report, Key Lab of Intelligent Information Processing of CAS. Institute of Computing Technology, 2014. 1", + "[6] Mathieu De Coster and Joni Dambre. Querying a sign language dictionary with videos using dense vector search. In 2023 IEEE International Conference on Acoustics, Speech, and Signal Processing Workshops, pages 1-5. IEEE, 2023. 1", + "[7] Philippe Dreuw, Carol Neidle, Vassilis Athitsos, Stan Sclaroff, and Hermann Ney. Benchmark databases for videobased automatic sign language recognition. In International Conference on Language Resources and Evaluation, 2008. 1", + "[8] Amanda Cardoso Duarte. Cross-modal neural sign language translation. In ACM International Conference on Multimedia, pages 1650-1654, 2019. 1", + "[9] Jia Gong, Lin Geng Foo, Yixuan He, Hossein Rahmani, and Jun Liu. Llms are good sign language translators. In Computer Vision and Pattern Recognition, pages 18362-18372, 2024. 1, 2", + "[10] Dan Guo, Wengang Zhou, Houqiang Li, and Meng Wang. Online early-late fusion based on adaptive hmm for sign language recognition. ACM Transactions on Multimedia Computing, Communications, and Applications, pages 1-18, 2017. 2", + "[11] Dan Guo, Wengang Zhou, Houqiang Li, and Meng Wang. Hierarchical LSTM for sign language translation. In AAAI Conference on Artificial Intelligence, pages 6845-6852, 2018. 6" + ], + "bbox": [ + 93, + 272, + 483, + 898 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[12] Dan Guo, Shengeng Tang, and Meng Wang. Connectionist temporal modeling of video and language: A joint model for translation and sign labeling. In International Joint Conference on Artificial Intelligence, pages 751-757, 2019. 2", + "[13] Dan Guo, Shengeng Tang, Richang Hong, and Meng Wang. Sign language recognition. Multimedia for Accessible Human Computer Interfaces, pages 23-59, 2021. 1", + "[14] Jonathan Ho, Ajay Jain, and Pieter Abbeel. Denoising diffusion probabilistic models. Conference and Workshop on Neural Information Processing Systems, 33:6840-6851, 2020. 3", + "[15] Hezhen Hu, Junfu Pu, Wengang Zhou, and Houqiang Li. Collaborative multilingual continuous sign language recognition: A unified framework. IEEE Transactions on Multimedia, pages 7559-7570, 2022. 1, 2", + "[16] Hezhen Hu, Weichao Zhao, Wengang Zhou, and Houqiang Li. Signbert+: Hand-model-aware self-supervised pretraining for sign language understanding. IEEE Transactions on Pattern Analysis and Machine Intelligence, pages 11221-11239, 2023. 1, 2", + "[17] Jie Huang, Wengang Zhou, Houqiang Li, and Weiping Li. Attention-based 3d-cnns for large-vocabulary sign language recognition. IEEE Transactions on Circuits and Systems for Video Technology, pages 2822-2832, 2018. 6", + "[18] Jie Huang, Wengang Zhou, Qilin Zhang, Houqiang Li, and Weiping Li. Video-based sign language recognition without temporal segmentation. In AAAI Conference on Artificial Intelligence, pages 2257-2264, 2018. 1, 6", + "[19] Wencan Huang, Wenwen Pan, Zhou Zhao, and Qi Tian. Towards fast and high-quality sign language production. In ACM International Conference on Multimedia, pages 3172-3181, 2021. 6", + "[20] Jestin Joy, Kannan Balakrishnan, and Sreeraj Madhavankutty. Developing a bilingual mobile dictionary for indian sign language and gathering users experience with signdict. Assistive Technology, 2020. 1", + "[21] Korrawe Karunratanakul, Konpat Preechakul, Supasorn Suwajanakorn, and Siyu Tang. Guided motion diffusion for controllable human motion synthesis. In International Conference on Computer Vision, pages 2151-2162, 2023. 3", + "[22] Oscar Koller. Quantitative survey of the state of the art in sign language recognition. ArXiv Preprint ArXiv:2008.09918, pages ArXiv-2008, 2020. 1, 2", + "[23] Taegyeong Lee, Soyeong Kwon, and Taehwan Kim. Grid diffusion models for text-to-video generation. In Computer Vision and Pattern Recognition, pages 8734-8743, 2024. 2", + "[24] Dongxu Li, Cristian Rodriguez, Xin Yu, and Hongdong Li. Word-level deep sign language recognition from video: A new large-scale dataset and methods comparison. In IEEE Winter Conference on Applications of Computer Vision, pages 1459-1469, 2020. 1", + "[25] Dongxu Li, Chenchen Xu, Xin Yu, Kaihao Zhang, Benjamin Swift, Hanna Suominen, and Hongdong Li. Tspnet: Hierarchical feature learning via temporal semantic pyramid for sign language translation. Neural Information Processing Systems, pages 12034-12045, 2020. 2", + "[26] Jingyun Liang, Yuchen Fan, Kai Zhang, Radu Timofte, Luc Van Gool, and Rakesh Ranjan. Movideo: Motion-aware" + ], + "bbox": [ + 516, + 92, + 905, + 900 + ], + "page_idx": 8 + }, + { + "type": "page_number", + "text": "3489", + "bbox": [ + 482, + 944, + 514, + 955 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "video generation with diffusion model. In European Conference on Computer Vision, pages 56-74. Springer, 2025. 3", + "[27] Chin-Yew Lin. Rouge: A package for automatic evaluation of summaries. In Text Summarization Branches Out, pages 74-81, 2004. 6", + "[28] Seshadri Mazumder, Rudrabha Mukhopadhyay, Vinay P Namboodiri, and CV Jawahar. Translating sign language videos to talking faces. In Indian Conference on Computer Vision, Graphics and Image Processing, pages 1-10, 2021. 2", + "[29] John McDonald, Rosalee Wolfe, Jerry Schnegg, Julie Hochgesang, Diana Gorman Jamrozik, Marie Stumbo, Larwan Berke, Melissa Bialek, and Farah Thomas. An automated technique for real-time production of lifelike animations of american sign language. Universal Access in the Information Society, pages 551-566, 2016. 2", + "[30] Yuchun Miao, Lefei Zhang, Liangpei Zhang, and Dacheng Tao. Dds2m: Self-supervised denoising diffusion spati-spectral model for hyperspectral image restoration. In International Conference on Computer Vision, pages 12086-12096, 2023. 3", + "[31] Shentong Mo, Enze Xie, Ruihang Chu, Lanqing Hong, Matthias Niessner, and Zhenguo Li. Dit-3d: Exploring plain diffusion transformers for 3d shape generation. Neural Information Processing Systems, 36:67960-67971, 2023. 3", + "[32] Haomiao Ni, Changhao Shi, Kai Li, Sharon X Huang, and Martin Renqiang Min. Conditional image-to-video generation with latent flow diffusion models. In Computer Vision and Pattern Recognition, pages 18444-18455, 2023. 3", + "[33] Kishore Papineni, Salim Roukos, Todd Ward, and Wei-Jing Zhu. Bleu: a method for automatic evaluation of machine translation. In Association for Computational Linguistics, pages 311-318, 2002. 6", + "[34] Robin Rombach, Andreas Blattmann, Dominik Lorenz, Patrick Esser, and Björn Ommer. High-resolution image synthesis with latent diffusion models. In Computer Vision and Pattern Recognition, pages 10684-10695, 2022. 2", + "[35] Ben Saunders, Necati Cihan Camgoz, and Richard Bowden. Progressive transformers for end-to-end sign language production. In European Conference on Computer Vision, pages 687-705, 2020. 2, 6", + "[36] Ben Saunders, Necati Cihan Camgoz, and Richard Bowden. Mixed signals: Sign language production via a mixture of motion primitives. In International Conference on Computer Vision, pages 1919-1929, 2021. 6", + "[37] Jérémie Segouat. A study of sign language coarticulation. ACM Sigaccess Accessibility and Computing, pages 31-38, 2009. 2", + "[38] Wenkang Shan, Zhenhua Liu, Xinfeng Zhang, Zhao Wang, Kai Han, Shanshe Wang, Siwei Ma, and Wen Gao. Diffusion-based 3d human pose estimation with multi-hypothesis aggregation. In International Conference on Computer Vision, pages 14761-14771, 2023. 3", + "[39] Jascha Sohl-Dickstein, Eric Weiss, Niru Maheswaranathan, and Surya Ganguli. Deep unsupervised learning using nonequilibrium thermodynamics. In International Conference on Machine Learning, pages 2256-2265, 2015. 2" + ], + "bbox": [ + 91, + 90, + 482, + 900 + ], + "page_idx": 9 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[40] Stephanie Stoll, Necati Cihan Camgoz, Simon Hadfield, and Richard Bowden. Text2sign: Towards sign language production using neural machine translation and generative adversarial networks. International Journal of Computer Vision, pages 891–908, 2020. 2", + "[41] Sugandhi, Parteek Kumar, and Sanmeet Kaur. Sign language generation system based on indian sign language grammar. ACM Transactions on Asian and Low-Resource Language Information Processing, 19(4):1-26, 2020. 2", + "[42] Shengeng Tang, Dan Guo, Richang Hong, and Meng Wang. Graph-based multimodal sequential embedding for sign language translation. IEEE Transactions on Multimedia, 24: 4433-4445, 2022. 1", + "[43] Shengeng Tang, Feng Xue, Jingjing Wu, Shuo Wang, and Richang Hong. Gloss-driven conditional diffusion models for sign language production. ACM Transactions on Multimedia Computing, Communications, and Applications, 2024. 2, 3, 7, 8", + "[44] Carla Viegas, Mert Inan, Lorna Quandt, and Malihe Alikhani. Including facial expressions in contextual embeddings for sign language generation. In Joint Conference on Lexical and Computational Semantics, pages 1-10, 2023. 6", + "[45] Bin Xia, Yulun Zhang, Shiyin Wang, Yitong Wang, Xinglong Wu, Yapeng Tian, Wenming Yang, and Luc Van Gool. Diffir: Efficient diffusion model for image restoration. In International Conference on Computer Vision, pages 13095-13105, 2023. 3", + "[46] Jinheng Xie, Yuexiang Li, Yawen Huang, Haozhe Liu, Wentian Zhang, Yefeng Zheng, and Mike Zheng Shou. Boxdiff: Text-to-image synthesis with training-free box-constrained diffusion. In International Conference on Computer Vision, pages 7452-7461, 2023. 3", + "[47] Pan Xie, Qipeng Zhang, Peng Taiying, Hao Tang, Yao Du, and Zexian Li. G2p-ddm: Generating sign pose sequence from gloss sequence with discrete diffusion model. In AAAI Conference on Artificial Intelligence, pages 6234-6242, 2024. 2, 3, 7, 8", + "[48] Wanli Xue, Jingze Liu, Siyi Yan, Yuxi Zhou, Tiantian Yuan, and Qing Guo. Alleviating data insufficiency for Chinese sign language recognition. Visual Intelligence, page 26, 2023. 2", + "[49] Shuai Yang, Yifan Zhou, Ziwei Liu, and Chen Change Loy. Fresco: Spatial-temporal correspondence for zero-shot video translation. In Computer Vision and Pattern Recognition, pages 8703-8712, 2024. 2", + "[50] Xunpeng Yi, Han Xu, Hao Zhang, Linfeng Tang, and Jiayi Ma. Diff-retinex: Rethinking low-light image enhancement with a generative diffusion model. In International Conference on Computer Vision, pages 12302-12311, 2023. 3", + "[51] Jan Zelinka and Jakub Kanis. Neural sign language synthesis: Words are our glosses. In Winter Conference on Applications of Computer Vision, pages 3395-3403, 2020. 2, 6", + "[52] Ni Zeng, Yiqiang Chen, Yang Gu, Dongdong Liu, and Yunbing Xing. Highly fluent sign language synthesis based on variable motion frame interpolation. In 2020 IEEE International Conference on Systems, Man, and Cybernetics, pages 1772-1777. IEEE, 2020. 2" + ], + "bbox": [ + 516, + 90, + 905, + 898 + ], + "page_idx": 9 + }, + { + "type": "page_number", + "text": "3490", + "bbox": [ + 482, + 944, + 516, + 955 + ], + "page_idx": 9 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[53] Jihai Zhang, Wengang Zhou, Chao Xie, Junfu Pu, and Houqiang Li. Chinese sign language recognition with adaptive hmm. In IEEE International Conference on Multimedia and Expo, pages 1-6. IEEE, 2016. 1", + "[54] Weichao Zhao, Hezhen Hu, Wengang Zhou, Jiaxin Shi, and Houqiang Li. Best: Bert pre-training for sign language recognition with coupling tokenization. In AAAI Conference on Artificial Intelligence, pages 3597-3605, 2023. 1, 2" + ], + "bbox": [ + 91, + 92, + 480, + 204 + ], + "page_idx": 10 + }, + { + "type": "page_number", + "text": "3491", + "bbox": [ + 482, + 945, + 511, + 955 + ], + "page_idx": 10 + } +] \ No newline at end of file diff --git a/2025/Discrete to Continuous_ Generating Smooth Transition Poses from Sign Language Observations/4e14bfdb-c7fc-42ce-bc5c-af5b00971815_model.json b/2025/Discrete to Continuous_ Generating Smooth Transition Poses from Sign Language Observations/4e14bfdb-c7fc-42ce-bc5c-af5b00971815_model.json new file mode 100644 index 0000000000000000000000000000000000000000..326a42ea38f33c2750d11f210cdb4427d7553e77 --- /dev/null +++ b/2025/Discrete to Continuous_ Generating Smooth Transition Poses from Sign Language Observations/4e14bfdb-c7fc-42ce-bc5c-af5b00971815_model.json @@ -0,0 +1,2202 @@ +[ + [ + { + "type": "header", + "bbox": [ + 0.107, + 0.003, + 0.182, + 0.044 + ], + "angle": 0, + "content": "CVF" + }, + { + "type": "header", + "bbox": [ + 0.237, + 0.001, + 0.812, + 0.047 + ], + "angle": 0, + "content": "This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore." + }, + { + "type": "title", + "bbox": [ + 0.192, + 0.131, + 0.806, + 0.177 + ], + "angle": 0, + "content": "Discrete to Continuous: Generating Smooth Transition Poses from Sign Language Observations" + }, + { + "type": "text", + "bbox": [ + 0.128, + 0.204, + 0.868, + 0.258 + ], + "angle": 0, + "content": "Shengeng Tang Jiayi He Lechao Cheng* Jingjing Wu Dan Guo Richang Hong* \nSchool of Computer Science and Information Engineering, Hefei University of Technology \n{tangsg, chenglc, wujingjing, guodan, hongrc}@h Fut.edu.cn, hejy4396@mail.hfut.edu.cn" + }, + { + "type": "title", + "bbox": [ + 0.248, + 0.292, + 0.327, + 0.308 + ], + "angle": 0, + "content": "Abstract" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.324, + 0.485, + 0.687 + ], + "angle": 0, + "content": "Generating continuous sign language videos from discrete segments is challenging due to the need for smooth transitions that preserve natural flow and meaning. Traditional approaches that simply concatenate isolated signs often result in abrupt transitions, disrupting video coherence. To address this, we propose a novel framework, Sign-D2C, that employs a conditional diffusion model to synthesize contextually smooth transition frames, enabling the seamless construction of continuous sign language sequences. Our approach transforms the unsupervised problem of transition frame generation into a supervised training task by simulating the absence of transition frames through random masking of segments in long-duration sign videos. The model learns to predict these masked frames by denoising Gaussian noise, conditioned on the surrounding sign observations, allowing it to handle complex, unstructured transitions. During inference, we apply a linearly interpolating padding strategy that initializes missing frames through interpolation between boundary frames, providing a stable foundation for iterative refinement by the diffusion model. Extensive experiments on the PHOENIX14T, USTCCSL100, and USTC-SLR500 datasets demonstrate the effectiveness of our method in producing continuous, natural sign language videos." + }, + { + "type": "title", + "bbox": [ + 0.092, + 0.716, + 0.222, + 0.731 + ], + "angle": 0, + "content": "1. Introduction" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.742, + 0.483, + 0.877 + ], + "angle": 0, + "content": "Sign language [3, 9, 13, 15, 16, 22, 42, 54], as a unique form of visual communication, presents an intricate challenge for data collection, especially when scaling datasets for continuous, long-duration sign sentences. Unlike textual language, which can be assembled seamlessly from words to sentences, sign language requires a fluid connection between individual signs to convey meaning accurately. This connection involves not only the specific signs but also the transitions between them, which convey tem-" + }, + { + "type": "image", + "bbox": [ + 0.518, + 0.291, + 0.905, + 0.535 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.512, + 0.546, + 0.908, + 0.63 + ], + "angle": 0, + "content": "Figure 1. Task and key steps. Our work aims to generate continuous sign videos by creating transition poses between discrete segments. In training, random masking simulates missing transitions, and the model learns to recover these frames (steps 1-3). During inference, padding initializes missing transitions, which the model refines to generate smooth, coherent sequences (steps 4-6)." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.658, + 0.907, + 0.749 + ], + "angle": 0, + "content": "poral and contextual coherence vital to understanding the intended message. As a result, while word-level sign language data [5, 7, 24, 53] has been collected extensively, there remains a critical gap in developing robust, large-scale datasets for sentence-level sign language [2, 8, 18], particularly for continuous sequences." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.75, + 0.909, + 0.903 + ], + "angle": 0, + "content": "A possible workaround involves pre-collecting word-level sign data akin to a dictionary [1, 6, 20] and then attempting to form sentences by joining these discrete signs. However, assembling sign sentences is fundamentally different from assembling textual sentences due to the need for visually smooth transitions between signs, which maintain natural flow and meaning. Direct concatenation of individual signs into sentences often results in abrupt and unnatural transitions, disrupting the continuity of the video and potentially obscuring meaning. Consequently, generating smooth" + }, + { + "type": "page_footnote", + "bbox": [ + 0.11, + 0.888, + 0.242, + 0.9 + ], + "angle": 0, + "content": "*Corresponding authors." + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.514, + 0.957 + ], + "angle": 0, + "content": "3481" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.09, + 0.092, + 0.482, + 0.152 + ], + "angle": 0, + "content": "transition frames between sign segments is essential to synthesizing continuous sign language videos [41, 51, 52], yet presents significant challenges due to the lack of supervised reference data for the missing transition frames." + }, + { + "type": "text", + "bbox": [ + 0.093, + 0.152, + 0.487, + 0.668 + ], + "angle": 0, + "content": "To address this issue, we propose a smooth transition pose generation framework, named Sign-D2C, to synthesize Continuous sign language videos from Discrete sign segments. This framework is based on a diffusion model, which learns to generate missing frames by observing surrounding sign frames. Specifically, our method leverages a conditional diffusion model trained to predict the missing transition frames by sampling from Gaussian noise and being conditioned by the observed video frames. As illustrated in Fig. 1, our method consists of a two-phase process designed to overcome the limitations of direct sign concatenation. In the training phase, we simulate the transition frame generation task by randomly masking frames within long sign language videos. This random masking effectively removes video segments, mimicking real-world scenarios where transition frames are absent. The model is then trained to reconstruct these missing frames based on the observable surrounding frames. By doing so, the model learns to generate plausible transitions in a supervised manner, enhancing its robustness and adaptability to complex, unstructured transitions. In the inference phase, we introduce a linearly interpolating padding strategy to initialize the missing transition frames. This strategy begins by selecting the last frame of the preceding observed segment and the first frame of the following observed segment as the boundary frames for the missing transition segment. The intermediate frames within the missing segment are then initialized by linearly interpolating between these boundary frames, creating an initial linearly interpolating padding that provides a stable foundation for generating smooth transitions. The conditional diffusion model then iteratively refines these frames to produce naturalistic transition frames that connect the observed segments seamlessly. The contributions of this work are summarized as follows:" + }, + { + "type": "text", + "bbox": [ + 0.091, + 0.675, + 0.482, + 0.75 + ], + "angle": 0, + "content": "- We propose a novel conditional diffusion framework, Sign-D2C, specifically designed to synthesize smooth and contextually accurate transition frames for continuous sign videos, overcoming the limitations of traditional word-to-sentence assembly techniques in sign language." + }, + { + "type": "text", + "bbox": [ + 0.091, + 0.751, + 0.482, + 0.84 + ], + "angle": 0, + "content": "- To simulate the generation of transition frames, we introduce a random masking strategy that removes sections of long sign language videos during training, enabling the model to predict missing frames in a supervised manner. This approach enhances the model's ability to handle diverse, complex transitions." + }, + { + "type": "text", + "bbox": [ + 0.092, + 0.84, + 0.482, + 0.901 + ], + "angle": 0, + "content": "- We design a linearly interpolating padding strategy for transition initialization in the inference phase. This strategy utilizes the boundary frames of observed segments to initialize the missing poses with linear interpolation," + }, + { + "type": "list", + "bbox": [ + 0.091, + 0.675, + 0.482, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "text", + "bbox": [ + 0.527, + 0.092, + 0.905, + 0.121 + ], + "angle": 0, + "content": "which provides a stable basis for generating visually seamless transitions." + }, + { + "type": "text", + "bbox": [ + 0.514, + 0.122, + 0.905, + 0.198 + ], + "angle": 0, + "content": "- Extensive experiments on three standard benchmarks (i.e., Phoenix14T, UCST-CSL100, and USTC-SLR500) demonstrate the robustness and effectiveness of our approach in producing smooth, contextually accurate transition poses for continuous sign language videos." + }, + { + "type": "title", + "bbox": [ + 0.514, + 0.213, + 0.655, + 0.229 + ], + "angle": 0, + "content": "2. Related Work" + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.238, + 0.851, + 0.256 + ], + "angle": 0, + "content": "2.1. Sign Language Production & Synthesis" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.261, + 0.905, + 0.382 + ], + "angle": 0, + "content": "Sign language plays a vital role in daily life and has become a prominent topic in artificial intelligence. Early research primarily focused on Sign Language Recognition (SLR) [10, 15, 16, 22, 48, 54] and Sign Language Translation (SLT) [3, 9, 12, 25]. However, with the growing need for effective communication between deaf and hearing individuals, Sign Language Production (SLP) has gained significant attention in recent years." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.383, + 0.906, + 0.58 + ], + "angle": 0, + "content": "Early approaches to SLP utilize synthetic animation techniques [28, 29, 37] to convert sentences into sign animations. These techniques, however, incur high collection costs and are limited to specific terms. With advancements in deep learning models, SLP has seen more extensive exploration. Stoll et al. [40] proposes a multi-step approach for generating sign language videos from text, involving the transformation of text into gloss, gloss into poses, and finally, poses into video. Building on this, Saunders et al. [35] introduces a model that generates sign language poses through autoregression. Diffusion-based solutions like G2P-DDM [47] and GCDM [43] further advance SLP by generating sign poses under semantic guidance." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.58, + 0.906, + 0.761 + ], + "angle": 0, + "content": "Despite significant progress in these methods, the limitations of existing data have become a major bottleneck in the development of sign language technology. Previous Sign Language Synthesis (SLS) approaches [51] focus on extracting gestures from continuous sign language videos. However, these methods still face challenges in synthesizing continuous, long-duration sign language sentences. Our method addresses this gap by starting with word-level sign language pose data, generating smooth transition frames between individual signs, and combining these segments into sentence-level sequences, thereby overcoming the limitations of existing methods." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.772, + 0.821, + 0.788 + ], + "angle": 0, + "content": "2.2. Diffusion-based Generative Models" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.795, + 0.905, + 0.903 + ], + "angle": 0, + "content": "Diffusion models have gained considerable attention in generative modeling since they were first introduced by Sohl-Dickstein et al. [39]. Its forward-backward diffusion process has proven highly effective in generating realistic and diverse outputs, enabling diffusion models to achieve competitive performance in various generative tasks, including image synthesis [34], video generation [23, 49], and 3D shape" + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.516, + 0.956 + ], + "angle": 0, + "content": "3482" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.09, + 0.092, + 0.482, + 0.273 + ], + "angle": 0, + "content": "generation [31]. One of the strengths of diffusion models is their iterative refinement capability, allowing for flexible and high-quality outputs that align well with tasks requiring fine control over generative processes. This quality has led to their adoption in areas such as image restoration, where they have been used for denoising and superresolution [30, 45, 50], as well as pose estimation, where they can handle uncertainty in human poses [38]. Recent advancements have extended their use to more complex generative tasks, such as controlled text-to-image synthesis [46] and conditional video generation [32], further demonstrating their adaptability and robustness." + }, + { + "type": "text", + "bbox": [ + 0.093, + 0.292, + 0.485, + 0.535 + ], + "angle": 0, + "content": "In Sign Language Production (SLP), diffusion models remain relatively unexplored but have shown promise in recent studies. For instance, G2P-DDM [47] uses a discrete diffusion model to translate the continuous pose generation task into a discrete space, simplifying the complexity of generating sign language poses while maintaining semantic alignment with gloss-based inputs. This approach allows for finer control over pose accuracy, making it particularly suitable for generating static sign poses from linguistic cues. Similarly, GCDM [43] applies a multi-hypothesis diffusion framework to generate diverse sign poses from gloss, accommodating multiple interpretations and improving the naturalness of synthesized gestures. These models represent pioneering steps in applying diffusion-based approaches to SLP, leveraging the unique properties of diffusion models to generate realistic sign poses under semantic guidance." + }, + { + "type": "text", + "bbox": [ + 0.093, + 0.554, + 0.484, + 0.902 + ], + "angle": 0, + "content": "Despite these advancements, current diffusion-based methods for SLP are primarily focused on generating sign poses according to textual signals rather than addressing the temporal continuity required for longer sign language sequences. In sign language synthesis, especially when producing smooth transitions in continuous sign language videos, maintaining smooth transitions between gestures is essential for ensuring that the sequences are fluid and natural. This need for temporally smooth transitions has led to research on methods that can generate these connecting frames between segments, although it remains an area with limited exploration in SLP. Beyond SLP, diffusion models have been successfully applied to tasks requiring temporal coherence and frame interpolation in video synthesis. For example, diffusion-based models have been employed in video generation to ensure smooth transitions between frames [26], and similar techniques have been explored in action synthesis and motion completion [21], where maintaining continuity across frames is crucial. Such applications highlight the potential of diffusion models to handle tasks involving complex temporal dependencies, suggesting that similar techniques could be valuable in SLP for generating continuous sign language videos." + }, + { + "type": "title", + "bbox": [ + 0.514, + 0.09, + 0.65, + 0.108 + ], + "angle": 0, + "content": "3. Methodology" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.116, + 0.907, + 0.253 + ], + "angle": 0, + "content": "In this work, we propose a framework for generating smooth transition frames between discrete sign segments to create continuous sign pose videos. Our approach transforms an unsupervised transition generation task into a supervised framework using a random masking strategy based on conditional diffusion. This section will outline the preliminaries (Sec. 3.1), followed by our core idea of realizing the task (Sec. 3.2), the framework structure (Sec. 3.3), and the training process (Sec. 3.4)." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.263, + 0.644, + 0.279 + ], + "angle": 0, + "content": "3.1. Preliminary" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.286, + 0.907, + 0.451 + ], + "angle": 0, + "content": "The diffusion model [14] is highly effective in generating high-quality images by combining diffusion processes, latent space modeling, and conditional inputs. In this work, we employ a conditional diffusion model to synthesize smooth transition frames between discrete sign segments, forming continuous sign language videos. Here, the model is trained on complete long-duration sign language pose sequences, progressively adding noise to simulate missing transitions. During testing, the model starts from Gaussian noise samples and generates the required transitions between observed sign frames." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.453, + 0.906, + 0.514 + ], + "angle": 0, + "content": "In the forward process, noise is incrementally added to the full sign language pose sequence \\( x_0 \\) to obtain a noisy version \\( x_T \\), mimicking the degradation of visual continuity. This process is described by:" + }, + { + "type": "equation", + "bbox": [ + 0.588, + 0.525, + 0.905, + 0.543 + ], + "angle": 0, + "content": "\\[\nx _ {t} = \\alpha_ {t} \\cdot x _ {t - 1} + \\sqrt {1 - \\alpha_ {t}} \\cdot \\epsilon_ {t}, \\tag {1}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.556, + 0.906, + 0.601 + ], + "angle": 0, + "content": "where \\(\\alpha_{t}\\) controls the level of noise at each step \\(t\\), and \\(\\epsilon_{t}\\) represents Gaussian noise. As \\(t\\) increases, the original sequence \\(x_0\\) transitions to pure noise \\(x_{T}\\)." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.602, + 0.906, + 0.676 + ], + "angle": 0, + "content": "The reverse process reconstructs the continuous sequence by progressively removing noise, where a neural network is trained to predict each denoised frame based on previous noisy frames. The denoising step at each iteration is defined as:" + }, + { + "type": "equation", + "bbox": [ + 0.532, + 0.69, + 0.905, + 0.707 + ], + "angle": 0, + "content": "\\[\np \\left(x _ {t - 1} \\mid x _ {t}, x _ {\\text {o b s}}\\right) = \\mathcal {N} \\left(x _ {t - 1}; \\mu_ {\\theta} \\left(x _ {t}, t, x _ {\\text {o b s}}\\right), \\Sigma_ {t}\\right), \\tag {2}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.72, + 0.906, + 0.794 + ], + "angle": 0, + "content": "where \\( x_{\\mathrm{obs}} \\) denotes the observed sign language frames that serve as conditional inputs, and \\( \\mu_{\\theta}(x_t,t,x_{\\mathrm{obs}}) \\) is the predicted mean for the denoising process. The observed frames condition the diffusion model, guiding it to produce smooth, contextually consistent transition frames." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.796, + 0.906, + 0.902 + ], + "angle": 0, + "content": "In our model, the training input is a complete long-duration sign language pose video, gradually noised to simulate missing transitions. During testing, Gaussian noise serves as the input, and the model generates a continuous video with smooth transition frames conditioned on observed frames \\( x_{\\mathrm{obs}} \\), effectively bridging discrete sign segments into a fluid sign language sequence." + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.516, + 0.957 + ], + "angle": 0, + "content": "3483" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.098, + 0.092, + 0.9, + 0.212 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.097, + 0.217, + 0.9, + 0.33 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.09, + 0.344, + 0.907, + 0.429 + ], + "angle": 0, + "content": "Figure 2. Overview. The proposed framework for generating continuous sign language videos with smooth transitions between discrete segments. In the training phase (top), a long-duration sign video undergoes random masking to create gaps, simulating missing transitions. An encoder processes the video, and a conditional diffusion model denoises and predicts the masked segments based on observable frames. The recovered video is produced through a decoder, with both encoder and decoder pre-trained for effective representation. In the inference phase (bottom), discrete sign videos are connected using a linearly interpolating padding strategy to initialize missing transitions. Gaussian noise is applied, and the model refines these transitions, generating a continuous sign video with coherent flow." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.454, + 0.429, + 0.47 + ], + "angle": 0, + "content": "3.2. Core Idea: Unsupervised to Supervised" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.487, + 0.483, + 0.623 + ], + "angle": 0, + "content": "Generating continuous sign language videos by adding transitions between discrete segments is challenging, as these segments typically start and end with specific gestures that make direct concatenation impractical. Additionally, this transition generation task is inherently unsupervised, as real transition frames are generally unavailable for training. To address this, we convert the unsupervised problem of generating transitions into a supervised learning framework by leveraging existing continuous sign language videos." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.634, + 0.483, + 0.755 + ], + "angle": 0, + "content": "Given a complete continuous sign language sequence \\( X = \\{x_{1}, x_{2}, \\dots, x_{n}\\} \\), we simulate missing transitions by randomly masking portions of \\( X \\), creating gaps that mimic real-world transition absences. Let the masked frames be denoted as \\( M \\subset X \\), and the observable pose frames as \\( X_{\\mathrm{obs}} = (X \\setminus X_{\\mathrm{mask}}) \\). This setup frames the task as predicting \\( X_{\\mathrm{mask}} \\) based on \\( X_{\\mathrm{obs}} \\), effectively generating supervised training data for transition prediction." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.765, + 0.483, + 0.902 + ], + "angle": 0, + "content": "By training the model to predict the masked frames \\( X_{\\mathrm{mask}} \\) conditioned on the observable context \\( X_{\\mathrm{obs}} \\), it learns to generate realistic transitions that smoothly bridge the gap. This innovative masking strategy allows our model to leverage the context of surrounding frames provided by \\( X_{\\mathrm{obs}} \\) to produce temporally coherent connections, transforming the inherently unsupervised transition generation task into a supervised one and enabling effective continuous sign language synthesis." + }, + { + "type": "title", + "bbox": [ + 0.514, + 0.454, + 0.72, + 0.47 + ], + "angle": 0, + "content": "3.3. Sign-D2C Framework" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.476, + 0.907, + 0.627 + ], + "angle": 0, + "content": "Our approach leverages a conditional diffusion model to generate transition frames, enabling smooth continuity between discrete sign segments. The model architecture consists of three main components: an encoder, a decoder, and a denoiser. The first two are composed of two self-attention layers and the denoiser consists of two blocks, each containing a self-attention layer, a cross-attention layer, and a feed-forward layer. This setup allows the model to effectively capture temporal dependencies and complex contextual information needed for sign language synthesis." + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.627, + 0.907, + 0.718 + ], + "angle": 0, + "content": "The Encoder first processes the input sequence \\( X = \\{x_{1}, x_{2}, \\ldots, x_{n}\\} \\), where \\( x_{i} \\) represents individual sign language pose frames. By encoding the long-duration video input, the Encoder transforms \\( X \\) into a latent representation \\( Z \\), which serves as the foundation for both the diffusion and denoising processes:" + }, + { + "type": "equation", + "bbox": [ + 0.631, + 0.724, + 0.905, + 0.74 + ], + "angle": 0, + "content": "\\[\nZ = \\operatorname {E n c o d e r} (X). \\tag {3}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.747, + 0.905, + 0.806 + ], + "angle": 0, + "content": "During the forward diffusion, Gaussian noise is gradually added to \\( Z \\) at each time step \\( t \\), resulting in a noisy representation \\( Z_{t} \\). This process can be transferred from Eq. 1, formulated as:" + }, + { + "type": "equation", + "bbox": [ + 0.585, + 0.811, + 0.905, + 0.836 + ], + "angle": 0, + "content": "\\[\nZ _ {t} = \\alpha_ {t} \\cdot Z _ {t - 1} + \\sqrt {1 - \\alpha_ {t} ^ {2}} \\cdot \\epsilon_ {t}, \\tag {4}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.841, + 0.907, + 0.899 + ], + "angle": 0, + "content": "where \\(\\alpha_{t}\\) controls the noise schedule, and \\(\\epsilon_{t}\\) is sampled from a Gaussian distribution. The diffusion process progressively transforms the latent representation towards pure noise." + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.516, + 0.956 + ], + "angle": 0, + "content": "3484" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.09, + 0.092, + 0.484, + 0.213 + ], + "angle": 0, + "content": "The denoiser then reconstructs the original representation by iteratively removing the noise. Here, the denoiser is conditioned on sign observations \\( X_{\\mathrm{obs}} \\), which serve as contextual priors. These observations are incorporated into the denoiser via cross-attention, allowing the model to focus on the relevant parts of \\( X_{\\mathrm{obs}} \\) at each denoising step. For each time step \\( t \\) in the reverse process, the denoised output \\( \\hat{Z}_{t - 1} \\) is given by:" + }, + { + "type": "equation", + "bbox": [ + 0.175, + 0.222, + 0.483, + 0.242 + ], + "angle": 0, + "content": "\\[\n\\hat {Z} _ {t - 1} = \\operatorname {D e n o i s e r} \\left(Z _ {t}, X _ {\\mathrm {o b s}}\\right), \\tag {5}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.251, + 0.483, + 0.281 + ], + "angle": 0, + "content": "where \\( X_{\\mathrm{obs}} \\) provides the model with the necessary context to predict the masked segments accurately." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.281, + 0.484, + 0.342 + ], + "angle": 0, + "content": "Finally, the decoder, also structured with three self-attention layers, takes the denoised latent representation and transforms it back into the original sign language video space, outputting the recovered or synthesized video \\(\\hat{X}\\):" + }, + { + "type": "equation", + "bbox": [ + 0.205, + 0.351, + 0.483, + 0.369 + ], + "angle": 0, + "content": "\\[\n\\hat {X} = \\operatorname {D e c o d e r} \\left(\\hat {Z} _ {0}\\right). \\tag {6}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.38, + 0.484, + 0.456 + ], + "angle": 0, + "content": "By using self-attention within both the encoder and decoder, and cross-attention in the denoiser to incorporate contextual observations, the model effectively learns to generate temporally coherent transitions, resulting in a smooth, continuous sign language video output." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.455, + 0.484, + 0.516 + ], + "angle": 0, + "content": "The proposed Sign-D2C incorporates two key strategies to enhance transition generation: a random masking strategy during training and linearly interpolating padding initialization for inference. The following provides details." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.516, + 0.484, + 0.576 + ], + "angle": 0, + "content": "Random Masking Strategy. To generate smooth transitional poses and capture relationships between different segments, we employ a random masking strategy during training, as illustrated in the training phase of Fig. 2." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.576, + 0.484, + 0.713 + ], + "angle": 0, + "content": "Given a long-duration video \\( p \\in \\mathbb{R}^{F \\times (N \\times 3)} \\) with \\( F \\) frames, each containing \\( N \\) joints in 3D space, the video is first encoded by a pre-trained encoder into a latent representation \\( x \\in \\mathbb{R}^{F \\times W} \\), where \\( W \\) denotes the dimension of the latent feature \\( x \\). The mask \\( m \\) is constructed by randomly selecting frame indices, extending across the dimensions of the latent feature \\( x \\). This mask \\( m \\) is then applied to selectively retain or mask out parts of \\( x \\), resulting in a masked latent representation:" + }, + { + "type": "equation", + "bbox": [ + 0.18, + 0.721, + 0.483, + 0.74 + ], + "angle": 0, + "content": "\\[\nx ^ {m} = x \\odot m, x \\in \\mathbb {R} ^ {F \\times W}, \\tag {7}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.75, + 0.483, + 0.825 + ], + "angle": 0, + "content": "where \\( x^{m} \\) serves as a conditional input to the model, distinguishing between occluded and visible movements. This masking strategy enables the model to predict missing frames based on observable context, facilitating smooth transition generation." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.826, + 0.484, + 0.902 + ], + "angle": 0, + "content": "Linear Interpolation Padding. In the inference phase, generating smooth transitions between discrete motion segments is challenging due to the gaps between observed frames. We leverage the natural correlation between consecutive frames in sign language sequences by introducing" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.092, + 0.905, + 0.122 + ], + "angle": 0, + "content": "a linear interpolation padding initialization strategy, as illustrated in Fig. 2." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.122, + 0.906, + 0.227 + ], + "angle": 0, + "content": "This strategy starts by using the last pose of the preceding observed segment, \\( p_{s} \\), and the first pose of the following observed segment, \\( p_{e} \\), as boundary frames for the missing transition segment. The goal is to generate intermediate frames \\( p_{c} \\) that provide a smooth transition from \\( p_{s} \\) to \\( p_{e} \\). Using linear interpolation, we initialize these missing frames as follows:" + }, + { + "type": "equation", + "bbox": [ + 0.535, + 0.231, + 0.905, + 0.272 + ], + "angle": 0, + "content": "\\[\n\\left\\{ \\begin{array}{l} f (i) = \\sum_ {k = 0} ^ {i} \\frac {1}{(k + 1)}, \\quad i = 0, \\dots , (e - s), \\\\ p _ {c} (i) = p _ {s} + \\frac {p _ {e} - p _ {s}}{e - s + 1} \\cdot f (i). \\end{array} \\right. \\tag {8}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.277, + 0.907, + 0.368 + ], + "angle": 0, + "content": "This interpolation scheme gradually fills in the key points at each time step, creating intermediate poses that ensure a coherent and natural flow between segments. This strategy facilitates smoother transitions and provides a stable foundation for later refinement, resulting in seamless sign language video synthesis." + }, + { + "type": "title", + "bbox": [ + 0.514, + 0.375, + 0.619, + 0.391 + ], + "angle": 0, + "content": "3.4. Training" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.397, + 0.906, + 0.533 + ], + "angle": 0, + "content": "Pre-training of Encoder & decoder. To enable coherent video generation, we pre-train the encoder and decoder to ensure effective latent representation processing. The encoder extracts latent features \\( x \\) from input videos, while the decoder reconstructs realistic sign language poses from these latent representations. For training the diffusion model, we freeze the encoder and decoder parameters, allowing the model to focus on learning smooth transitions based on these pre-trained features." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.534, + 0.906, + 0.579 + ], + "angle": 0, + "content": "Pre-training is conducted end-to-end with a reconstruction loss, optimizing the encoder-decoder pair to accurately represent and recover video details:" + }, + { + "type": "equation", + "bbox": [ + 0.61, + 0.585, + 0.905, + 0.628 + ], + "angle": 0, + "content": "\\[\n\\mathcal {L} _ {\\text {p r e}} = \\frac {1}{F} \\sum_ {f = 1} ^ {F} | p _ {f} - \\hat {p} _ {f} |, \\tag {9}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.634, + 0.906, + 0.709 + ], + "angle": 0, + "content": "where \\( p \\) represents the input video with \\( F \\) frames, and \\( \\hat{p} \\) is the reconstructed output. This process ensures the encoder captures essential features, while the decoder faithfully reconstructs realistic sign language poses, setting a solid foundation for high-quality transition generation." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.71, + 0.906, + 0.771 + ], + "angle": 0, + "content": "Loss Optimization. According to Eq. 5, we can obtain the denoiser prediction \\(\\hat{X} = \\{\\hat{x}_1,\\hat{x}_2,\\dots ,\\hat{x}_F\\}\\). During training, the loss function constrains the difference between the original latent feature \\(x\\) and the denoiser's prediction \\(\\hat{x}\\):" + }, + { + "type": "equation", + "bbox": [ + 0.608, + 0.777, + 0.905, + 0.82 + ], + "angle": 0, + "content": "\\[\n\\mathcal {L} _ {\\text {d i f f}} = \\frac {1}{F} \\sum_ {f = 1} ^ {F} | x _ {f} - \\hat {x} _ {f} |, \\tag {10}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.825, + 0.906, + 0.9 + ], + "angle": 0, + "content": "where \\(F\\) denotes the number of frames in the video sequence. This loss encourages the model to accurately recover the original latent representation, improving the quality of the generated transitions and ensuring alignment with the encoded video features." + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.516, + 0.957 + ], + "angle": 0, + "content": "3485" + } + ], + [ + { + "type": "title", + "bbox": [ + 0.091, + 0.09, + 0.224, + 0.108 + ], + "angle": 0, + "content": "4. Experiments" + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.116, + 0.299, + 0.133 + ], + "angle": 0, + "content": "4.1. Experimental Settings" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.14, + 0.483, + 0.262 + ], + "angle": 0, + "content": "Datasets. We conduct the experiments on three datasets, i.e., PHOENIX14T, USTC-CSL100, and USTC-SLR500. The first two are sentence-level continuous sign language datasets, which are used to evaluate supervised transition generation in controlled settings. While the latter one is a word-level isolated sign language dataset, allowing us to assess the model's performance in generating transitions within real-world, discrete sign scenarios." + }, + { + "type": "text", + "bbox": [ + 0.091, + 0.263, + 0.483, + 0.354 + ], + "angle": 0, + "content": "- PHOENIX14T [2] comprises 8,257 sentence-level samples, which covers 2,887 German words. For preprocessing, we follow the settings in previous sign language production studies [19, 36], using OpenPose [4] to extract 2D coordinates from the videos, which are then refined to 3D coordinates through a skeleton correction model [51]." + }, + { + "type": "text", + "bbox": [ + 0.092, + 0.355, + 0.483, + 0.444 + ], + "angle": 0, + "content": "- USTC-CSL100 [18] covers 100 unique sentences performed by 50 signers. We use the split-I [11] approach to divide the dataset into 4,000 training examples and 1,000 test examples. The above two datasets enable us to evaluate the model's effectiveness on continuous sign language sequences in a structured environment." + }, + { + "type": "text", + "bbox": [ + 0.092, + 0.445, + 0.483, + 0.521 + ], + "angle": 0, + "content": "- USTC-SLR500 [17] is an isolated sign language dataset containing 500 Chinese sign words, each performed five times by 50 signers. Following [17], we split the dataset by using videos from 36 signers for training and the remaining 14 signers for testing." + }, + { + "type": "list", + "bbox": [ + 0.091, + 0.263, + 0.483, + 0.521 + ], + "angle": 0, + "content": null + }, + { + "type": "text", + "bbox": [ + 0.089, + 0.522, + 0.483, + 0.749 + ], + "angle": 0, + "content": "Implementation Details. Since generating continuous sign language videos from discrete fragments is inherently an unsupervised task, there are no standardized evaluation metrics. To effectively assess our model's performance, we analyzed transition actions between two gestures on the sentence-level datasets (i.e., PHOENIX14T and USTC-CSL100), observing that transitions typically span 10 to 20 frames. Based on these findings, we designed experiments using two masking strategies: masking 10 frames every 20 frames and masking 20 frames every 10 frames. Additionally, to further validate the effectiveness of our approach, we used the isolated sign language dataset (i.e., USTC-SLR500) to assemble discrete vocabulary videos into continuous sequences, demonstrating the model's capability to produce smooth transitions in a real-world setting." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.75, + 0.483, + 0.902 + ], + "angle": 0, + "content": "Evaluation Metrics. We evaluate our performance on two sentence-level datasets in terms of both semantics and coherence. Following prior work in sign language generation [19, 35, 36, 44], we use the NSLT method [2] to convert generated sign language videos back into text. We then compare this text with reference data to calculate semantic accuracy metrics, including BLEU [33], ROUGE [27], and WER. Additionally, we use FID, DTW, and MPJPE to assess the consistency and smoothness of the generated transitions, measuring the coherence of the results." + }, + { + "type": "table_caption", + "bbox": [ + 0.514, + 0.09, + 0.905, + 0.117 + ], + "angle": 0, + "content": "Table 1. Ablation results of different masking ratios on PHOENIX14T (observation 20 frames - prediction 10 frames)." + }, + { + "type": "table", + "bbox": [ + 0.518, + 0.118, + 0.905, + 0.194 + ], + "angle": 0, + "content": "
Masking Ratio rDEVTEST
B1↑B4↑DTW↓B1↑B4↑DTW↓
r = 0.122.178.040.6122.478.670.60
r = 0.322.468.060.6222.368.730.61
r = 0.522.308.060.7722.028.350.75
" + }, + { + "type": "table_caption", + "bbox": [ + 0.514, + 0.203, + 0.905, + 0.231 + ], + "angle": 0, + "content": "Table 2. Ablation results of different padding strategies on PHOENIX14T (observation 20 frames - prediction 10 frames)." + }, + { + "type": "table", + "bbox": [ + 0.518, + 0.232, + 0.905, + 0.312 + ], + "angle": 0, + "content": "
Padding StrategyDEVTEST
DTW↓FID↓MPJPE↓DTW↓FID↓MPJPE↓
w/o Padding11.462.7469.6011.603.0687.57
Front Padding0.722.6847.900.692.9960.05
Linear Padding0.622.6847.300.612.9859.36
Back Padding0.672.6847.890.662.9959.91
" + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.331, + 0.667, + 0.347 + ], + "angle": 0, + "content": "4.2. Ablation Study" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.354, + 0.907, + 0.459 + ], + "angle": 0, + "content": "In this subsection, we explore the effects of different random masking ratios in the training phase and various padding strategies during inference. The experiments in this part are all conducted under the condition of predicting 10 transition frames for every 20 observed sign frames to evaluate their impact on both semantic accuracy and transition coherence." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.461, + 0.909, + 0.674 + ], + "angle": 0, + "content": "Effect of Masking Ratio. Table 1 presents the results of different random masking ratios \\( r \\) on semantic preservation and temporal coherence. With a lower masking ratio \\( (r = 0.1) \\), the model achieves the best coherence, reflected by the lowest DTW scores of 0.61 on DEV and 0.60 on TEST. This indicates that smaller masking ratios allow the model to maintain smoother transitions, as more contextual frames are available to guide the generation of missing frames. However, as the masking ratio increases to \\( r = 0.3 \\), we observe a slight improvement in semantic accuracy, with BLEU-4 scores rising to 8.06 on DEV and 8.73 on TEST. This suggests that a higher masking ratio may encourage the model to focus more on the sign semantic content, as it learns to reconstruct more extensive missing segments." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.674, + 0.909, + 0.809 + ], + "angle": 0, + "content": "Interestingly, further increasing the masking ratio to \\( r = 0.5 \\) leads to a drop in both semantic accuracy and coherence, with DTW increasing and BLEU decreasing slightly. This highlights a trade-off: while a moderate masking ratio improves semantic preservation, excessive masking may reduce the model's ability to maintain smooth transitions and semantic accuracy. Based on these observations, we select \\( r = 0.3 \\) for subsequent experiments, as it provides a good balance between semantic retention and transition smoothness." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.81, + 0.909, + 0.902 + ], + "angle": 0, + "content": "Influence of Padding Strategy. Table 2 summarizes the results of different padding strategies applied during inference. We experiment with four approaches: w/o Padding (no initialization for missing frames), Front Padding (using the last frame of the previous observation segment), Back Padding (using the first frame of the following observation" + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.516, + 0.956 + ], + "angle": 0, + "content": "3486" + } + ], + [ + { + "type": "table_caption", + "bbox": [ + 0.159, + 0.09, + 0.837, + 0.104 + ], + "angle": 0, + "content": "Table 3. Performance comparison on PHOENIX14T (under setting: observation 20 frames - prediction 10 frames)." + }, + { + "type": "table", + "bbox": [ + 0.095, + 0.105, + 0.905, + 0.203 + ], + "angle": 0, + "content": "
MethodsDEVTEST
B1↑B4↑ROUGE↑WER↓DTW↓FID↓MPIJE↓B1↑B4↑ROUGE↑WER↓DTW↓FID↓MPIJE↓
Ground Truth29.7712.1329.6074.170.000.000.0029.7611.9328.9871.940.000.000.00
G2P-DDM [47]8.250.997.1798.9312.712.8555.029.521.298.1298.5412.573.1967.81
VQ-GCDM [43]9.310.877.5998.6912.112.8653.879.850.967.9598.5911.983.2166.51
Ours22.468.0622.0780.730.622.6847.3022.368.7322.3679.060.612.9859.36
" + }, + { + "type": "table_caption", + "bbox": [ + 0.16, + 0.208, + 0.837, + 0.221 + ], + "angle": 0, + "content": "Table 4. Performance comparison on PHOENIX14T (under setting: observation 10 frames - prediction 20 frames)." + }, + { + "type": "table", + "bbox": [ + 0.095, + 0.223, + 0.905, + 0.322 + ], + "angle": 0, + "content": "
MethodsDEVTEST
B1↑B4↑ROUGE↑WER↓DTW↓FID↓MPJPE↓B1↑B4↑ROUGE↑WER↓DTW↓FID↓MPJPE↓
Ground Truth29.7712.1329.6074.170.000.000.0029.7611.9328.9871.940.000.000.00
G2P-DDM [47]8.980.637.4498.6115.952.8955.219.130.747.7598.6415.583.2268.18
VQ-GCDM [43]8.640.897.0898.4812.403.0548.669.021.367.2298.6112.223.3760.27
Ours19.325.9218.3387.300.702.7548.0419.596.2618.5986.80.683.0559.81
" + }, + { + "type": "table_caption", + "bbox": [ + 0.339, + 0.326, + 0.658, + 0.338 + ], + "angle": 0, + "content": "Table 5. Performance comparison on USTC-CSL100." + }, + { + "type": "table", + "bbox": [ + 0.095, + 0.34, + 0.905, + 0.439 + ], + "angle": 0, + "content": "
MethodsObservation 20 Frames - Prediction 10 FramesObservation 10 Frames - Prediction 20 Frames
B1↑B4↑ROUGE↑WER↓DTW↓FID↓MPJPE↓B1↑B4↑ROUGE↑WER↓DTW↓FID↓MPJPE↓
Ground Truth69.1059.1268.5347.380.000.000.0069.1059.1268.5347.380.000.000.00
G2P-DDM [47]13.742.1113.8592.047.000.53258.4313.431.7413.6292.157.470.53258.43
VQ-GCDM [43]30.7110.6829.1777.875.100.4391.1018.772.0816.6492.6710.890.54104.27
Ours69.6458.9869.2646.661.920.3177.0467.9556.6167.3849.241.790.3178.62
" + }, + { + "type": "text", + "bbox": [ + 0.089, + 0.454, + 0.482, + 0.575 + ], + "angle": 0, + "content": "segment), and Linear Padding (performing linear interpolation between the two boundary poses, as Eq. 8). The \"w/o Padding\" approach yields the highest DTW scores (11.46 - DEV and 10.36 - TEST), indicating poor coherence and unnatural transitions. Both Front Padding and Back Padding improve the smoothness of transitions, as shown by lower DTW scores (0.67 - DEV and 0.66 - TEST), but their FID and MPJPE remain better compared to Linear Padding." + }, + { + "type": "text", + "bbox": [ + 0.089, + 0.576, + 0.482, + 0.698 + ], + "angle": 0, + "content": "The Linear Padding strategy achieves the best overall performance, with the lowest DTW scores (0.62 on DEV and 0.66 on TEST) and MPJPE values, reflecting both smoothness and accuracy in pose reconstruction. By initializing missing frames through linear interpolation, Linear Padding provides a stable and effective foundation for generating realistic and seamless transitions, closely resembling real-world signing behavior." + }, + { + "type": "title", + "bbox": [ + 0.09, + 0.71, + 0.395, + 0.726 + ], + "angle": 0, + "content": "4.3. Comparison with State-of-the-Arts" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.733, + 0.483, + 0.809 + ], + "angle": 0, + "content": "We compare our method with two state-of-the-art models: G2P-DDM [47] and VQ-GCDM [43]. Both G2P-DDM and VQ-GCDM are primarily designed for SLP. Due to the absence of established benchmarks for our specific task, we adapted these models to serve as comparative baselines." + }, + { + "type": "text", + "bbox": [ + 0.089, + 0.81, + 0.483, + 0.902 + ], + "angle": 0, + "content": "Comparison on PHOENIX14T. As shown in Table 4, under the \"Observation 20 Frames - Prediction 10 Frames\" setting, our method achieves superior performance across both semantic and coherence metrics. Specifically, we attain a BLEU-1 score of \\(22.46\\%\\) and \\(22.36\\%\\) on the DEV and TEST sets, respectively, outperforming G2P-DDM and" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.454, + 0.905, + 0.56 + ], + "angle": 0, + "content": "VQ-GCDM. This improvement in BLEU scores indicates that our method more accurately preserves semantic content compared to existing methods. Furthermore, our method shows significant gains in temporal coherence, with DTW scores of 0.62 on DEV and 0.61 on TEST, respectively. These results confirm that our model generates smoother and more realistic transitions than the baseline methods." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.565, + 0.907, + 0.744 + ], + "angle": 0, + "content": "Under the more challenging \"Observation 10 Frames - Prediction 20 Frames\" setting, our method continues to demonstrate robust performance despite the increased prediction length. We achieve BLEU-1 scores of \\(19.32\\%\\) on DEV and \\(19.59\\%\\) on TEST, considerably higher than the baseline methods. Our method also maintains lower DTW and FID scores, demonstrating its ability to generate coherent and semantically accurate sequences even with limited observed data. The MPJPE scores further support the effectiveness of our approach, with scores of 48.04 on DEV and 59.81 on TEST, indicating a closer alignment to ground truth poses than the other methods." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.75, + 0.909, + 0.901 + ], + "angle": 0, + "content": "Comparison on USTC-CSL100. Table 5 shows the results on the USTC-CSL100 dataset, which evaluates our method's ability to generate sign language sequences. In the \"Observation 20 Frames - Prediction 10 Frames\" setting, our model achieves near-ground-truth semantic accuracy, with BLEU-1 and BLEU-4 scores of \\(69.64\\%\\) and \\(58.98\\%\\) on DEV, closely approaching the ground truth scores of \\(69.10\\%\\) and \\(59.12\\%\\). Additionally, our DTW score of 1.922 and FID score of 3.21 indicate high temporal consistency and low distributional divergence from the real data. Com" + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.516, + 0.956 + ], + "angle": 0, + "content": "3487" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.095, + 0.09, + 0.904, + 0.252 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.09, + 0.254, + 0.908, + 0.283 + ], + "angle": 0, + "content": "Figure 3. Visualization examples of generating 10-frame transition pose under 20-frame observations on PHOENIX14T. We compare our method with G2P-DDM and VQ-GCDM, and the Ground Truth is attached. Green: Transitions. Orange and Blue: Observations." + }, + { + "type": "image", + "bbox": [ + 0.095, + 0.298, + 0.902, + 0.457 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.09, + 0.46, + 0.908, + 0.489 + ], + "angle": 0, + "content": "Figure 4. Visualization examples on the USTC-SLR500, demonstrate the generation of transition poses according to two discrete sign segments. Green: Transitions. Orange and Blue: Observations." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.515, + 0.483, + 0.56 + ], + "angle": 0, + "content": "pared to G2P-DDM and VQ-GCDM, which yield higher DTW and MPJPE scores, our model achieves smoother and more accurate pose transitions." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.562, + 0.483, + 0.698 + ], + "angle": 0, + "content": "In the \"Observation 10 Frames - Prediction 20 Frames\" scenario, our method continues to demonstrate resilience in generating long sequences. Our BLEU-1 and BLEU-4 scores on DEV are \\(67.95\\%\\) and \\(56.61\\%\\), respectively, showing superior semantic preservation compared to the baseline methods. Additionally, the DTW score of 1.79 and MPJPE score of 78.62 confirm our model's ability to produce coherent and realistic transitions, outperforming the baselines in alignment with true pose sequences." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.711, + 0.274, + 0.727 + ], + "angle": 0, + "content": "4.4. Qualitative Results" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.734, + 0.483, + 0.884 + ], + "angle": 0, + "content": "Visualization Results. Figure 3 presents a comparative demonstration of our method against other approaches (i.e., G2P-DDM [47] and VQ-GCDM [43]) in generating transition sign poses. This example clearly shows that our method significantly outperforms G2P-DDM and VQ-GCDM in the quality of interpolated movements, especially excelling in overall motion coherence. Furthermore, our method meticulously fine-tunes the observations (the Blue and Orange sections), effectively avoiding coherence issues that may arise from directly concatenating discrete sign segments." + }, + { + "type": "text", + "bbox": [ + 0.091, + 0.886, + 0.483, + 0.901 + ], + "angle": 0, + "content": "Verification in Realistic Scenarios. Figure 4 demonstrates" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.516, + 0.907, + 0.652 + ], + "angle": 0, + "content": "visualization examples for synthesizing continuous sign videos from discrete segments in a realistic, word-level scenario using the USTC-SLR500 dataset. Here, the observed frames from two isolated sign segments are marked in blue and orange, while the generated transitions are shown in green. The results show that our method not only generates smooth transitions between discrete sign language segments but also fine-tunes the observed poses based on the overall semantics, resulting in a coherent sign language video." + }, + { + "type": "title", + "bbox": [ + 0.514, + 0.676, + 0.64, + 0.691 + ], + "angle": 0, + "content": "5. Conclusions" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.705, + 0.907, + 0.901 + ], + "angle": 0, + "content": "This paper proposes a method to generate continuous sign language videos by synthesizing transition frames between discrete segments, transforming an unsupervised task into a supervised one through random masking. Using a conditional diffusion model, the approach learns to predict missing frames based on observable context, allowing for smooth and temporally coherent outputs. By incorporating a linearly interpolating padding strategy during inference, the model initializes transitions effectively, further refining them into seamless sequences. Extensive experiments demonstrate the effectiveness of leveraging context for natural transitions, advancing sign language synthesis by enabling fluid, continuous sequences." + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.516, + 0.956 + ], + "angle": 0, + "content": "3488" + } + ], + [ + { + "type": "title", + "bbox": [ + 0.092, + 0.091, + 0.27, + 0.108 + ], + "angle": 0, + "content": "6. Acknowledgments" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.115, + 0.484, + 0.236 + ], + "angle": 0, + "content": "This work was supported by the National Natural Science Foundation of China (Grants No. U23B2031, 62472139, 62302142, 62272144), the Anhui Provincial Natural Science Foundation, China (Grant No. 2408085QF191), the Fundamental Research Funds for the Central Universities (Grants No. JZ2024HGTA0178, JZ2023HGQA0097), and the Major Project of Anhui Province (Grant No. 202423k09020001)." + }, + { + "type": "title", + "bbox": [ + 0.093, + 0.248, + 0.188, + 0.264 + ], + "angle": 0, + "content": "References" + }, + { + "type": "ref_text", + "bbox": [ + 0.101, + 0.273, + 0.484, + 0.328 + ], + "angle": 0, + "content": "[1] Ahmed H Aliwy and A Alethary Ahmed. Development of arabic sign language dictionary using 3d avatar technologies. Indonesian Journal of Electrical Engineering and Computer Science, 21(1):609-616, 2021. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.33, + 0.484, + 0.384 + ], + "angle": 0, + "content": "[2] Necati Cihan Camgoz, Simon Hadfield, Oscar Koller, Hermann Ney, and Richard Bowden. Neural sign language translation. In Computer Vision and Pattern Recognition, pages 7784-7793, 2018. 1, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.385, + 0.483, + 0.44 + ], + "angle": 0, + "content": "[3] Necati Cihan Camgoz, Oscar Koller, Simon Hadfield, and Richard Bowden. Multi-channel transformers for multi-articulatory sign language translation. In European Conference on Computer Vision, pages 301-319, 2020. 1, 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.442, + 0.482, + 0.495 + ], + "angle": 0, + "content": "[4] Zhe Cao, Tomas Simon, Shih-En Wei, and Yaser Sheikh. Realtime multi-person 2d pose estimation using part affinity fields. In Computer Vision and Pattern Recognition, pages 7291-7299, 2017. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.497, + 0.483, + 0.566 + ], + "angle": 0, + "content": "[5] Xiujuan Chai, Hanjie Wang, and Xilin Chen. The devisesign large vocabulary of Chinese sign language database and baseline evaluations. In Technical report, Key Lab of Intelligent Information Processing of CAS. Institute of Computing Technology, 2014. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.567, + 0.482, + 0.622 + ], + "angle": 0, + "content": "[6] Mathieu De Coster and Joni Dambre. Querying a sign language dictionary with videos using dense vector search. In 2023 IEEE International Conference on Acoustics, Speech, and Signal Processing Workshops, pages 1-5. IEEE, 2023. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.623, + 0.483, + 0.678 + ], + "angle": 0, + "content": "[7] Philippe Dreuw, Carol Neidle, Vassilis Athitsos, Stan Sclaroff, and Hermann Ney. Benchmark databases for videobased automatic sign language recognition. In International Conference on Language Resources and Evaluation, 2008. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.679, + 0.482, + 0.719 + ], + "angle": 0, + "content": "[8] Amanda Cardoso Duarte. Cross-modal neural sign language translation. In ACM International Conference on Multimedia, pages 1650-1654, 2019. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.721, + 0.483, + 0.774 + ], + "angle": 0, + "content": "[9] Jia Gong, Lin Geng Foo, Yixuan He, Hossein Rahmani, and Jun Liu. Llms are good sign language translators. In Computer Vision and Pattern Recognition, pages 18362-18372, 2024. 1, 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.776, + 0.483, + 0.844 + ], + "angle": 0, + "content": "[10] Dan Guo, Wengang Zhou, Houqiang Li, and Meng Wang. Online early-late fusion based on adaptive hmm for sign language recognition. ACM Transactions on Multimedia Computing, Communications, and Applications, pages 1-18, 2017. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.846, + 0.483, + 0.899 + ], + "angle": 0, + "content": "[11] Dan Guo, Wengang Zhou, Houqiang Li, and Meng Wang. Hierarchical LSTM for sign language translation. In AAAI Conference on Artificial Intelligence, pages 6845-6852, 2018. 6" + }, + { + "type": "list", + "bbox": [ + 0.094, + 0.273, + 0.484, + 0.899 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.093, + 0.906, + 0.148 + ], + "angle": 0, + "content": "[12] Dan Guo, Shengeng Tang, and Meng Wang. Connectionist temporal modeling of video and language: A joint model for translation and sign labeling. In International Joint Conference on Artificial Intelligence, pages 751-757, 2019. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.149, + 0.905, + 0.19 + ], + "angle": 0, + "content": "[13] Dan Guo, Shengeng Tang, Richang Hong, and Meng Wang. Sign language recognition. Multimedia for Accessible Human Computer Interfaces, pages 23-59, 2021. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.191, + 0.905, + 0.244 + ], + "angle": 0, + "content": "[14] Jonathan Ho, Ajay Jain, and Pieter Abbeel. Denoising diffusion probabilistic models. Conference and Workshop on Neural Information Processing Systems, 33:6840-6851, 2020. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.246, + 0.905, + 0.301 + ], + "angle": 0, + "content": "[15] Hezhen Hu, Junfu Pu, Wengang Zhou, and Houqiang Li. Collaborative multilingual continuous sign language recognition: A unified framework. IEEE Transactions on Multimedia, pages 7559-7570, 2022. 1, 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.302, + 0.905, + 0.37 + ], + "angle": 0, + "content": "[16] Hezhen Hu, Weichao Zhao, Wengang Zhou, and Houqiang Li. Signbert+: Hand-model-aware self-supervised pretraining for sign language understanding. IEEE Transactions on Pattern Analysis and Machine Intelligence, pages 11221-11239, 2023. 1, 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.372, + 0.905, + 0.427 + ], + "angle": 0, + "content": "[17] Jie Huang, Wengang Zhou, Houqiang Li, and Weiping Li. Attention-based 3d-cnns for large-vocabulary sign language recognition. IEEE Transactions on Circuits and Systems for Video Technology, pages 2822-2832, 2018. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.428, + 0.905, + 0.482 + ], + "angle": 0, + "content": "[18] Jie Huang, Wengang Zhou, Qilin Zhang, Houqiang Li, and Weiping Li. Video-based sign language recognition without temporal segmentation. In AAAI Conference on Artificial Intelligence, pages 2257-2264, 2018. 1, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.484, + 0.905, + 0.537 + ], + "angle": 0, + "content": "[19] Wencan Huang, Wenwen Pan, Zhou Zhao, and Qi Tian. Towards fast and high-quality sign language production. In ACM International Conference on Multimedia, pages 3172-3181, 2021. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.539, + 0.905, + 0.594 + ], + "angle": 0, + "content": "[20] Jestin Joy, Kannan Balakrishnan, and Sreeraj Madhavankutty. Developing a bilingual mobile dictionary for indian sign language and gathering users experience with signdict. Assistive Technology, 2020. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.595, + 0.905, + 0.65 + ], + "angle": 0, + "content": "[21] Korrawe Karunratanakul, Konpat Preechakul, Supasorn Suwajanakorn, and Siyu Tang. Guided motion diffusion for controllable human motion synthesis. In International Conference on Computer Vision, pages 2151-2162, 2023. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.651, + 0.905, + 0.692 + ], + "angle": 0, + "content": "[22] Oscar Koller. Quantitative survey of the state of the art in sign language recognition. ArXiv Preprint ArXiv:2008.09918, pages ArXiv-2008, 2020. 1, 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.692, + 0.905, + 0.734 + ], + "angle": 0, + "content": "[23] Taegyeong Lee, Soyeong Kwon, and Taehwan Kim. Grid diffusion models for text-to-video generation. In Computer Vision and Pattern Recognition, pages 8734-8743, 2024. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.735, + 0.905, + 0.803 + ], + "angle": 0, + "content": "[24] Dongxu Li, Cristian Rodriguez, Xin Yu, and Hongdong Li. Word-level deep sign language recognition from video: A new large-scale dataset and methods comparison. In IEEE Winter Conference on Applications of Computer Vision, pages 1459-1469, 2020. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.804, + 0.905, + 0.872 + ], + "angle": 0, + "content": "[25] Dongxu Li, Chenchen Xu, Xin Yu, Kaihao Zhang, Benjamin Swift, Hanna Suominen, and Hongdong Li. Tspnet: Hierarchical feature learning via temporal semantic pyramid for sign language translation. Neural Information Processing Systems, pages 12034-12045, 2020. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.873, + 0.905, + 0.901 + ], + "angle": 0, + "content": "[26] Jingyun Liang, Yuchen Fan, Kai Zhang, Radu Timofte, Luc Van Gool, and Rakesh Ranjan. Movideo: Motion-aware" + }, + { + "type": "list", + "bbox": [ + 0.517, + 0.093, + 0.906, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.516, + 0.956 + ], + "angle": 0, + "content": "3489" + } + ], + [ + { + "type": "ref_text", + "bbox": [ + 0.123, + 0.092, + 0.482, + 0.133 + ], + "angle": 0, + "content": "video generation with diffusion model. In European Conference on Computer Vision, pages 56-74. Springer, 2025. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.134, + 0.483, + 0.175 + ], + "angle": 0, + "content": "[27] Chin-Yew Lin. Rouge: A package for automatic evaluation of summaries. In Text Summarization Branches Out, pages 74-81, 2004. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.177, + 0.483, + 0.244 + ], + "angle": 0, + "content": "[28] Seshadri Mazumder, Rudrabha Mukhopadhyay, Vinay P Namboodiri, and CV Jawahar. Translating sign language videos to talking faces. In Indian Conference on Computer Vision, Graphics and Image Processing, pages 1-10, 2021. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.246, + 0.483, + 0.329 + ], + "angle": 0, + "content": "[29] John McDonald, Rosalee Wolfe, Jerry Schnegg, Julie Hochgesang, Diana Gorman Jamrozik, Marie Stumbo, Larwan Berke, Melissa Bialek, and Farah Thomas. An automated technique for real-time production of lifelike animations of american sign language. Universal Access in the Information Society, pages 551-566, 2016. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.33, + 0.483, + 0.398 + ], + "angle": 0, + "content": "[30] Yuchun Miao, Lefei Zhang, Liangpei Zhang, and Dacheng Tao. Dds2m: Self-supervised denoising diffusion spati-spectral model for hyperspectral image restoration. In International Conference on Computer Vision, pages 12086-12096, 2023. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.4, + 0.483, + 0.455 + ], + "angle": 0, + "content": "[31] Shentong Mo, Enze Xie, Ruihang Chu, Lanqing Hong, Matthias Niessner, and Zhenguo Li. Dit-3d: Exploring plain diffusion transformers for 3d shape generation. Neural Information Processing Systems, 36:67960-67971, 2023. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.457, + 0.483, + 0.51 + ], + "angle": 0, + "content": "[32] Haomiao Ni, Changhao Shi, Kai Li, Sharon X Huang, and Martin Renqiang Min. Conditional image-to-video generation with latent flow diffusion models. In Computer Vision and Pattern Recognition, pages 18444-18455, 2023. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.511, + 0.483, + 0.566 + ], + "angle": 0, + "content": "[33] Kishore Papineni, Salim Roukos, Todd Ward, and Wei-Jing Zhu. Bleu: a method for automatic evaluation of machine translation. In Association for Computational Linguistics, pages 311-318, 2002. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.567, + 0.483, + 0.622 + ], + "angle": 0, + "content": "[34] Robin Rombach, Andreas Blattmann, Dominik Lorenz, Patrick Esser, and Björn Ommer. High-resolution image synthesis with latent diffusion models. In Computer Vision and Pattern Recognition, pages 10684-10695, 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.623, + 0.483, + 0.677 + ], + "angle": 0, + "content": "[35] Ben Saunders, Necati Cihan Camgoz, and Richard Bowden. Progressive transformers for end-to-end sign language production. In European Conference on Computer Vision, pages 687-705, 2020. 2, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.679, + 0.483, + 0.734 + ], + "angle": 0, + "content": "[36] Ben Saunders, Necati Cihan Camgoz, and Richard Bowden. Mixed signals: Sign language production via a mixture of motion primitives. In International Conference on Computer Vision, pages 1919-1929, 2021. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.735, + 0.483, + 0.774 + ], + "angle": 0, + "content": "[37] Jérémie Segouat. A study of sign language coarticulation. ACM Sigaccess Accessibility and Computing, pages 31-38, 2009. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.776, + 0.483, + 0.845 + ], + "angle": 0, + "content": "[38] Wenkang Shan, Zhenhua Liu, Xinfeng Zhang, Zhao Wang, Kai Han, Shanshe Wang, Siwei Ma, and Wen Gao. Diffusion-based 3d human pose estimation with multi-hypothesis aggregation. In International Conference on Computer Vision, pages 14761-14771, 2023. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.846, + 0.483, + 0.901 + ], + "angle": 0, + "content": "[39] Jascha Sohl-Dickstein, Eric Weiss, Niru Maheswaranathan, and Surya Ganguli. Deep unsupervised learning using nonequilibrium thermodynamics. In International Conference on Machine Learning, pages 2256-2265, 2015. 2" + }, + { + "type": "list", + "bbox": [ + 0.093, + 0.092, + 0.483, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.092, + 0.906, + 0.162 + ], + "angle": 0, + "content": "[40] Stephanie Stoll, Necati Cihan Camgoz, Simon Hadfield, and Richard Bowden. Text2sign: Towards sign language production using neural machine translation and generative adversarial networks. International Journal of Computer Vision, pages 891–908, 2020. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.165, + 0.906, + 0.22 + ], + "angle": 0, + "content": "[41] Sugandhi, Parteek Kumar, and Sanmeet Kaur. Sign language generation system based on indian sign language grammar. ACM Transactions on Asian and Low-Resource Language Information Processing, 19(4):1-26, 2020. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.223, + 0.906, + 0.277 + ], + "angle": 0, + "content": "[42] Shengeng Tang, Dan Guo, Richang Hong, and Meng Wang. Graph-based multimodal sequential embedding for sign language translation. IEEE Transactions on Multimedia, 24: 4433-4445, 2022. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.281, + 0.906, + 0.349 + ], + "angle": 0, + "content": "[43] Shengeng Tang, Feng Xue, Jingjing Wu, Shuo Wang, and Richang Hong. Gloss-driven conditional diffusion models for sign language production. ACM Transactions on Multimedia Computing, Communications, and Applications, 2024. 2, 3, 7, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.353, + 0.906, + 0.408 + ], + "angle": 0, + "content": "[44] Carla Viegas, Mert Inan, Lorna Quandt, and Malihe Alikhani. Including facial expressions in contextual embeddings for sign language generation. In Joint Conference on Lexical and Computational Semantics, pages 1-10, 2023. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.411, + 0.906, + 0.479 + ], + "angle": 0, + "content": "[45] Bin Xia, Yulun Zhang, Shiyin Wang, Yitong Wang, Xinglong Wu, Yapeng Tian, Wenming Yang, and Luc Van Gool. Diffir: Efficient diffusion model for image restoration. In International Conference on Computer Vision, pages 13095-13105, 2023. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.483, + 0.906, + 0.552 + ], + "angle": 0, + "content": "[46] Jinheng Xie, Yuexiang Li, Yawen Huang, Haozhe Liu, Wentian Zhang, Yefeng Zheng, and Mike Zheng Shou. Boxdiff: Text-to-image synthesis with training-free box-constrained diffusion. In International Conference on Computer Vision, pages 7452-7461, 2023. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.555, + 0.906, + 0.623 + ], + "angle": 0, + "content": "[47] Pan Xie, Qipeng Zhang, Peng Taiying, Hao Tang, Yao Du, and Zexian Li. G2p-ddm: Generating sign pose sequence from gloss sequence with discrete diffusion model. In AAAI Conference on Artificial Intelligence, pages 6234-6242, 2024. 2, 3, 7, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.627, + 0.906, + 0.669 + ], + "angle": 0, + "content": "[48] Wanli Xue, Jingze Liu, Siyi Yan, Yuxi Zhou, Tiantian Yuan, and Qing Guo. Alleviating data insufficiency for Chinese sign language recognition. Visual Intelligence, page 26, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.672, + 0.906, + 0.727 + ], + "angle": 0, + "content": "[49] Shuai Yang, Yifan Zhou, Ziwei Liu, and Chen Change Loy. Fresco: Spatial-temporal correspondence for zero-shot video translation. In Computer Vision and Pattern Recognition, pages 8703-8712, 2024. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.729, + 0.906, + 0.784 + ], + "angle": 0, + "content": "[50] Xunpeng Yi, Han Xu, Hao Zhang, Linfeng Tang, and Jiayi Ma. Diff-retinex: Rethinking low-light image enhancement with a generative diffusion model. In International Conference on Computer Vision, pages 12302-12311, 2023. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.787, + 0.906, + 0.829 + ], + "angle": 0, + "content": "[51] Jan Zelinka and Jakub Kanis. Neural sign language synthesis: Words are our glosses. In Winter Conference on Applications of Computer Vision, pages 3395-3403, 2020. 2, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.832, + 0.906, + 0.9 + ], + "angle": 0, + "content": "[52] Ni Zeng, Yiqiang Chen, Yang Gu, Dongdong Liu, and Yunbing Xing. Highly fluent sign language synthesis based on variable motion frame interpolation. In 2020 IEEE International Conference on Systems, Man, and Cybernetics, pages 1772-1777. IEEE, 2020. 2" + }, + { + "type": "list", + "bbox": [ + 0.517, + 0.092, + 0.906, + 0.9 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.517, + 0.957 + ], + "angle": 0, + "content": "3490" + } + ], + [ + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.093, + 0.482, + 0.147 + ], + "angle": 0, + "content": "[53] Jihai Zhang, Wengang Zhou, Chao Xie, Junfu Pu, and Houqiang Li. Chinese sign language recognition with adaptive hmm. In IEEE International Conference on Multimedia and Expo, pages 1-6. IEEE, 2016. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.149, + 0.482, + 0.205 + ], + "angle": 0, + "content": "[54] Weichao Zhao, Hezhen Hu, Wengang Zhou, Jiaxin Shi, and Houqiang Li. Best: Bert pre-training for sign language recognition with coupling tokenization. In AAAI Conference on Artificial Intelligence, pages 3597-3605, 2023. 1, 2" + }, + { + "type": "list", + "bbox": [ + 0.093, + 0.093, + 0.482, + 0.205 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.946, + 0.513, + 0.956 + ], + "angle": 0, + "content": "3491" + } + ] +] \ No newline at end of file diff --git a/2025/Discrete to Continuous_ Generating Smooth Transition Poses from Sign Language Observations/4e14bfdb-c7fc-42ce-bc5c-af5b00971815_origin.pdf b/2025/Discrete to Continuous_ Generating Smooth Transition Poses from Sign Language Observations/4e14bfdb-c7fc-42ce-bc5c-af5b00971815_origin.pdf new file mode 100644 index 0000000000000000000000000000000000000000..a1bb7a5768540d6de890faa749d318bb5290dd65 --- /dev/null +++ b/2025/Discrete to Continuous_ Generating Smooth Transition Poses from Sign Language Observations/4e14bfdb-c7fc-42ce-bc5c-af5b00971815_origin.pdf @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:05b7a0bce7aed1435787361041a6016609741a86696361a5deb21fa6196c2d1a +size 915374 diff --git a/2025/Discrete to Continuous_ Generating Smooth Transition Poses from Sign Language Observations/full.md b/2025/Discrete to Continuous_ Generating Smooth Transition Poses from Sign Language Observations/full.md new file mode 100644 index 0000000000000000000000000000000000000000..b78c57f8b7855e7585b563809228f1091bd29d0a --- /dev/null +++ b/2025/Discrete to Continuous_ Generating Smooth Transition Poses from Sign Language Observations/full.md @@ -0,0 +1,317 @@ +# Discrete to Continuous: Generating Smooth Transition Poses from Sign Language Observations + +Shengeng Tang Jiayi He Lechao Cheng* Jingjing Wu Dan Guo Richang Hong* +School of Computer Science and Information Engineering, Hefei University of Technology +{tangsg, chenglc, wujingjing, guodan, hongrc}@h Fut.edu.cn, hejy4396@mail.hfut.edu.cn + +# Abstract + +Generating continuous sign language videos from discrete segments is challenging due to the need for smooth transitions that preserve natural flow and meaning. Traditional approaches that simply concatenate isolated signs often result in abrupt transitions, disrupting video coherence. To address this, we propose a novel framework, Sign-D2C, that employs a conditional diffusion model to synthesize contextually smooth transition frames, enabling the seamless construction of continuous sign language sequences. Our approach transforms the unsupervised problem of transition frame generation into a supervised training task by simulating the absence of transition frames through random masking of segments in long-duration sign videos. The model learns to predict these masked frames by denoising Gaussian noise, conditioned on the surrounding sign observations, allowing it to handle complex, unstructured transitions. During inference, we apply a linearly interpolating padding strategy that initializes missing frames through interpolation between boundary frames, providing a stable foundation for iterative refinement by the diffusion model. Extensive experiments on the PHOENIX14T, USTCCSL100, and USTC-SLR500 datasets demonstrate the effectiveness of our method in producing continuous, natural sign language videos. + +# 1. Introduction + +Sign language [3, 9, 13, 15, 16, 22, 42, 54], as a unique form of visual communication, presents an intricate challenge for data collection, especially when scaling datasets for continuous, long-duration sign sentences. Unlike textual language, which can be assembled seamlessly from words to sentences, sign language requires a fluid connection between individual signs to convey meaning accurately. This connection involves not only the specific signs but also the transitions between them, which convey tem- + +![](images/a34c3d1d814e76cbdfae282031a899ea45cf3fc0faaafb7adb1ff387663d6df7.jpg) +Figure 1. Task and key steps. Our work aims to generate continuous sign videos by creating transition poses between discrete segments. In training, random masking simulates missing transitions, and the model learns to recover these frames (steps 1-3). During inference, padding initializes missing transitions, which the model refines to generate smooth, coherent sequences (steps 4-6). + +poral and contextual coherence vital to understanding the intended message. As a result, while word-level sign language data [5, 7, 24, 53] has been collected extensively, there remains a critical gap in developing robust, large-scale datasets for sentence-level sign language [2, 8, 18], particularly for continuous sequences. + +A possible workaround involves pre-collecting word-level sign data akin to a dictionary [1, 6, 20] and then attempting to form sentences by joining these discrete signs. However, assembling sign sentences is fundamentally different from assembling textual sentences due to the need for visually smooth transitions between signs, which maintain natural flow and meaning. Direct concatenation of individual signs into sentences often results in abrupt and unnatural transitions, disrupting the continuity of the video and potentially obscuring meaning. Consequently, generating smooth + +transition frames between sign segments is essential to synthesizing continuous sign language videos [41, 51, 52], yet presents significant challenges due to the lack of supervised reference data for the missing transition frames. + +To address this issue, we propose a smooth transition pose generation framework, named Sign-D2C, to synthesize Continuous sign language videos from Discrete sign segments. This framework is based on a diffusion model, which learns to generate missing frames by observing surrounding sign frames. Specifically, our method leverages a conditional diffusion model trained to predict the missing transition frames by sampling from Gaussian noise and being conditioned by the observed video frames. As illustrated in Fig. 1, our method consists of a two-phase process designed to overcome the limitations of direct sign concatenation. In the training phase, we simulate the transition frame generation task by randomly masking frames within long sign language videos. This random masking effectively removes video segments, mimicking real-world scenarios where transition frames are absent. The model is then trained to reconstruct these missing frames based on the observable surrounding frames. By doing so, the model learns to generate plausible transitions in a supervised manner, enhancing its robustness and adaptability to complex, unstructured transitions. In the inference phase, we introduce a linearly interpolating padding strategy to initialize the missing transition frames. This strategy begins by selecting the last frame of the preceding observed segment and the first frame of the following observed segment as the boundary frames for the missing transition segment. The intermediate frames within the missing segment are then initialized by linearly interpolating between these boundary frames, creating an initial linearly interpolating padding that provides a stable foundation for generating smooth transitions. The conditional diffusion model then iteratively refines these frames to produce naturalistic transition frames that connect the observed segments seamlessly. The contributions of this work are summarized as follows: + +- We propose a novel conditional diffusion framework, Sign-D2C, specifically designed to synthesize smooth and contextually accurate transition frames for continuous sign videos, overcoming the limitations of traditional word-to-sentence assembly techniques in sign language. +- To simulate the generation of transition frames, we introduce a random masking strategy that removes sections of long sign language videos during training, enabling the model to predict missing frames in a supervised manner. This approach enhances the model's ability to handle diverse, complex transitions. +- We design a linearly interpolating padding strategy for transition initialization in the inference phase. This strategy utilizes the boundary frames of observed segments to initialize the missing poses with linear interpolation, + +which provides a stable basis for generating visually seamless transitions. + +- Extensive experiments on three standard benchmarks (i.e., Phoenix14T, UCST-CSL100, and USTC-SLR500) demonstrate the robustness and effectiveness of our approach in producing smooth, contextually accurate transition poses for continuous sign language videos. + +# 2. Related Work + +# 2.1. Sign Language Production & Synthesis + +Sign language plays a vital role in daily life and has become a prominent topic in artificial intelligence. Early research primarily focused on Sign Language Recognition (SLR) [10, 15, 16, 22, 48, 54] and Sign Language Translation (SLT) [3, 9, 12, 25]. However, with the growing need for effective communication between deaf and hearing individuals, Sign Language Production (SLP) has gained significant attention in recent years. + +Early approaches to SLP utilize synthetic animation techniques [28, 29, 37] to convert sentences into sign animations. These techniques, however, incur high collection costs and are limited to specific terms. With advancements in deep learning models, SLP has seen more extensive exploration. Stoll et al. [40] proposes a multi-step approach for generating sign language videos from text, involving the transformation of text into gloss, gloss into poses, and finally, poses into video. Building on this, Saunders et al. [35] introduces a model that generates sign language poses through autoregression. Diffusion-based solutions like G2P-DDM [47] and GCDM [43] further advance SLP by generating sign poses under semantic guidance. + +Despite significant progress in these methods, the limitations of existing data have become a major bottleneck in the development of sign language technology. Previous Sign Language Synthesis (SLS) approaches [51] focus on extracting gestures from continuous sign language videos. However, these methods still face challenges in synthesizing continuous, long-duration sign language sentences. Our method addresses this gap by starting with word-level sign language pose data, generating smooth transition frames between individual signs, and combining these segments into sentence-level sequences, thereby overcoming the limitations of existing methods. + +# 2.2. Diffusion-based Generative Models + +Diffusion models have gained considerable attention in generative modeling since they were first introduced by Sohl-Dickstein et al. [39]. Its forward-backward diffusion process has proven highly effective in generating realistic and diverse outputs, enabling diffusion models to achieve competitive performance in various generative tasks, including image synthesis [34], video generation [23, 49], and 3D shape + +generation [31]. One of the strengths of diffusion models is their iterative refinement capability, allowing for flexible and high-quality outputs that align well with tasks requiring fine control over generative processes. This quality has led to their adoption in areas such as image restoration, where they have been used for denoising and superresolution [30, 45, 50], as well as pose estimation, where they can handle uncertainty in human poses [38]. Recent advancements have extended their use to more complex generative tasks, such as controlled text-to-image synthesis [46] and conditional video generation [32], further demonstrating their adaptability and robustness. + +In Sign Language Production (SLP), diffusion models remain relatively unexplored but have shown promise in recent studies. For instance, G2P-DDM [47] uses a discrete diffusion model to translate the continuous pose generation task into a discrete space, simplifying the complexity of generating sign language poses while maintaining semantic alignment with gloss-based inputs. This approach allows for finer control over pose accuracy, making it particularly suitable for generating static sign poses from linguistic cues. Similarly, GCDM [43] applies a multi-hypothesis diffusion framework to generate diverse sign poses from gloss, accommodating multiple interpretations and improving the naturalness of synthesized gestures. These models represent pioneering steps in applying diffusion-based approaches to SLP, leveraging the unique properties of diffusion models to generate realistic sign poses under semantic guidance. + +Despite these advancements, current diffusion-based methods for SLP are primarily focused on generating sign poses according to textual signals rather than addressing the temporal continuity required for longer sign language sequences. In sign language synthesis, especially when producing smooth transitions in continuous sign language videos, maintaining smooth transitions between gestures is essential for ensuring that the sequences are fluid and natural. This need for temporally smooth transitions has led to research on methods that can generate these connecting frames between segments, although it remains an area with limited exploration in SLP. Beyond SLP, diffusion models have been successfully applied to tasks requiring temporal coherence and frame interpolation in video synthesis. For example, diffusion-based models have been employed in video generation to ensure smooth transitions between frames [26], and similar techniques have been explored in action synthesis and motion completion [21], where maintaining continuity across frames is crucial. Such applications highlight the potential of diffusion models to handle tasks involving complex temporal dependencies, suggesting that similar techniques could be valuable in SLP for generating continuous sign language videos. + +# 3. Methodology + +In this work, we propose a framework for generating smooth transition frames between discrete sign segments to create continuous sign pose videos. Our approach transforms an unsupervised transition generation task into a supervised framework using a random masking strategy based on conditional diffusion. This section will outline the preliminaries (Sec. 3.1), followed by our core idea of realizing the task (Sec. 3.2), the framework structure (Sec. 3.3), and the training process (Sec. 3.4). + +# 3.1. Preliminary + +The diffusion model [14] is highly effective in generating high-quality images by combining diffusion processes, latent space modeling, and conditional inputs. In this work, we employ a conditional diffusion model to synthesize smooth transition frames between discrete sign segments, forming continuous sign language videos. Here, the model is trained on complete long-duration sign language pose sequences, progressively adding noise to simulate missing transitions. During testing, the model starts from Gaussian noise samples and generates the required transitions between observed sign frames. + +In the forward process, noise is incrementally added to the full sign language pose sequence $x_0$ to obtain a noisy version $x_T$ , mimicking the degradation of visual continuity. This process is described by: + +$$ +x _ {t} = \alpha_ {t} \cdot x _ {t - 1} + \sqrt {1 - \alpha_ {t}} \cdot \epsilon_ {t}, \tag {1} +$$ + +where $\alpha_{t}$ controls the level of noise at each step $t$ , and $\epsilon_{t}$ represents Gaussian noise. As $t$ increases, the original sequence $x_0$ transitions to pure noise $x_{T}$ . + +The reverse process reconstructs the continuous sequence by progressively removing noise, where a neural network is trained to predict each denoised frame based on previous noisy frames. The denoising step at each iteration is defined as: + +$$ +p \left(x _ {t - 1} \mid x _ {t}, x _ {\text {o b s}}\right) = \mathcal {N} \left(x _ {t - 1}; \mu_ {\theta} \left(x _ {t}, t, x _ {\text {o b s}}\right), \Sigma_ {t}\right), \tag {2} +$$ + +where $x_{\mathrm{obs}}$ denotes the observed sign language frames that serve as conditional inputs, and $\mu_{\theta}(x_t,t,x_{\mathrm{obs}})$ is the predicted mean for the denoising process. The observed frames condition the diffusion model, guiding it to produce smooth, contextually consistent transition frames. + +In our model, the training input is a complete long-duration sign language pose video, gradually noised to simulate missing transitions. During testing, Gaussian noise serves as the input, and the model generates a continuous video with smooth transition frames conditioned on observed frames $x_{\mathrm{obs}}$ , effectively bridging discrete sign segments into a fluid sign language sequence. + +![](images/10f8ad0591c27acc5272163fac9914faf71f533ade0ba56830fcad116f74e3bf.jpg) + +![](images/eb02065371f8d11e017dca2ef7790a6ca4f0f5e84baa494082c6b1e828f7207d.jpg) +Figure 2. Overview. The proposed framework for generating continuous sign language videos with smooth transitions between discrete segments. In the training phase (top), a long-duration sign video undergoes random masking to create gaps, simulating missing transitions. An encoder processes the video, and a conditional diffusion model denoises and predicts the masked segments based on observable frames. The recovered video is produced through a decoder, with both encoder and decoder pre-trained for effective representation. In the inference phase (bottom), discrete sign videos are connected using a linearly interpolating padding strategy to initialize missing transitions. Gaussian noise is applied, and the model refines these transitions, generating a continuous sign video with coherent flow. + +# 3.2. Core Idea: Unsupervised to Supervised + +Generating continuous sign language videos by adding transitions between discrete segments is challenging, as these segments typically start and end with specific gestures that make direct concatenation impractical. Additionally, this transition generation task is inherently unsupervised, as real transition frames are generally unavailable for training. To address this, we convert the unsupervised problem of generating transitions into a supervised learning framework by leveraging existing continuous sign language videos. + +Given a complete continuous sign language sequence $X = \{x_{1}, x_{2}, \dots, x_{n}\}$ , we simulate missing transitions by randomly masking portions of $X$ , creating gaps that mimic real-world transition absences. Let the masked frames be denoted as $M \subset X$ , and the observable pose frames as $X_{\mathrm{obs}} = (X \setminus X_{\mathrm{mask}})$ . This setup frames the task as predicting $X_{\mathrm{mask}}$ based on $X_{\mathrm{obs}}$ , effectively generating supervised training data for transition prediction. + +By training the model to predict the masked frames $X_{\mathrm{mask}}$ conditioned on the observable context $X_{\mathrm{obs}}$ , it learns to generate realistic transitions that smoothly bridge the gap. This innovative masking strategy allows our model to leverage the context of surrounding frames provided by $X_{\mathrm{obs}}$ to produce temporally coherent connections, transforming the inherently unsupervised transition generation task into a supervised one and enabling effective continuous sign language synthesis. + +# 3.3. Sign-D2C Framework + +Our approach leverages a conditional diffusion model to generate transition frames, enabling smooth continuity between discrete sign segments. The model architecture consists of three main components: an encoder, a decoder, and a denoiser. The first two are composed of two self-attention layers and the denoiser consists of two blocks, each containing a self-attention layer, a cross-attention layer, and a feed-forward layer. This setup allows the model to effectively capture temporal dependencies and complex contextual information needed for sign language synthesis. + +The Encoder first processes the input sequence $X = \{x_{1}, x_{2}, \ldots, x_{n}\}$ , where $x_{i}$ represents individual sign language pose frames. By encoding the long-duration video input, the Encoder transforms $X$ into a latent representation $Z$ , which serves as the foundation for both the diffusion and denoising processes: + +$$ +Z = \operatorname {E n c o d e r} (X). \tag {3} +$$ + +During the forward diffusion, Gaussian noise is gradually added to $Z$ at each time step $t$ , resulting in a noisy representation $Z_{t}$ . This process can be transferred from Eq. 1, formulated as: + +$$ +Z _ {t} = \alpha_ {t} \cdot Z _ {t - 1} + \sqrt {1 - \alpha_ {t} ^ {2}} \cdot \epsilon_ {t}, \tag {4} +$$ + +where $\alpha_{t}$ controls the noise schedule, and $\epsilon_{t}$ is sampled from a Gaussian distribution. The diffusion process progressively transforms the latent representation towards pure noise. + +The denoiser then reconstructs the original representation by iteratively removing the noise. Here, the denoiser is conditioned on sign observations $X_{\mathrm{obs}}$ , which serve as contextual priors. These observations are incorporated into the denoiser via cross-attention, allowing the model to focus on the relevant parts of $X_{\mathrm{obs}}$ at each denoising step. For each time step $t$ in the reverse process, the denoised output $\hat{Z}_{t - 1}$ is given by: + +$$ +\hat {Z} _ {t - 1} = \operatorname {D e n o i s e r} \left(Z _ {t}, X _ {\mathrm {o b s}}\right), \tag {5} +$$ + +where $X_{\mathrm{obs}}$ provides the model with the necessary context to predict the masked segments accurately. + +Finally, the decoder, also structured with three self-attention layers, takes the denoised latent representation and transforms it back into the original sign language video space, outputting the recovered or synthesized video $\hat{X}$ : + +$$ +\hat {X} = \operatorname {D e c o d e r} \left(\hat {Z} _ {0}\right). \tag {6} +$$ + +By using self-attention within both the encoder and decoder, and cross-attention in the denoiser to incorporate contextual observations, the model effectively learns to generate temporally coherent transitions, resulting in a smooth, continuous sign language video output. + +The proposed Sign-D2C incorporates two key strategies to enhance transition generation: a random masking strategy during training and linearly interpolating padding initialization for inference. The following provides details. + +Random Masking Strategy. To generate smooth transitional poses and capture relationships between different segments, we employ a random masking strategy during training, as illustrated in the training phase of Fig. 2. + +Given a long-duration video $p \in \mathbb{R}^{F \times (N \times 3)}$ with $F$ frames, each containing $N$ joints in 3D space, the video is first encoded by a pre-trained encoder into a latent representation $x \in \mathbb{R}^{F \times W}$ , where $W$ denotes the dimension of the latent feature $x$ . The mask $m$ is constructed by randomly selecting frame indices, extending across the dimensions of the latent feature $x$ . This mask $m$ is then applied to selectively retain or mask out parts of $x$ , resulting in a masked latent representation: + +$$ +x ^ {m} = x \odot m, x \in \mathbb {R} ^ {F \times W}, \tag {7} +$$ + +where $x^{m}$ serves as a conditional input to the model, distinguishing between occluded and visible movements. This masking strategy enables the model to predict missing frames based on observable context, facilitating smooth transition generation. + +Linear Interpolation Padding. In the inference phase, generating smooth transitions between discrete motion segments is challenging due to the gaps between observed frames. We leverage the natural correlation between consecutive frames in sign language sequences by introducing + +a linear interpolation padding initialization strategy, as illustrated in Fig. 2. + +This strategy starts by using the last pose of the preceding observed segment, $p_{s}$ , and the first pose of the following observed segment, $p_{e}$ , as boundary frames for the missing transition segment. The goal is to generate intermediate frames $p_{c}$ that provide a smooth transition from $p_{s}$ to $p_{e}$ . Using linear interpolation, we initialize these missing frames as follows: + +$$ +\left\{ \begin{array}{l} f (i) = \sum_ {k = 0} ^ {i} \frac {1}{(k + 1)}, \quad i = 0, \dots , (e - s), \\ p _ {c} (i) = p _ {s} + \frac {p _ {e} - p _ {s}}{e - s + 1} \cdot f (i). \end{array} \right. \tag {8} +$$ + +This interpolation scheme gradually fills in the key points at each time step, creating intermediate poses that ensure a coherent and natural flow between segments. This strategy facilitates smoother transitions and provides a stable foundation for later refinement, resulting in seamless sign language video synthesis. + +# 3.4. Training + +Pre-training of Encoder & decoder. To enable coherent video generation, we pre-train the encoder and decoder to ensure effective latent representation processing. The encoder extracts latent features $x$ from input videos, while the decoder reconstructs realistic sign language poses from these latent representations. For training the diffusion model, we freeze the encoder and decoder parameters, allowing the model to focus on learning smooth transitions based on these pre-trained features. + +Pre-training is conducted end-to-end with a reconstruction loss, optimizing the encoder-decoder pair to accurately represent and recover video details: + +$$ +\mathcal {L} _ {\text {p r e}} = \frac {1}{F} \sum_ {f = 1} ^ {F} | p _ {f} - \hat {p} _ {f} |, \tag {9} +$$ + +where $p$ represents the input video with $F$ frames, and $\hat{p}$ is the reconstructed output. This process ensures the encoder captures essential features, while the decoder faithfully reconstructs realistic sign language poses, setting a solid foundation for high-quality transition generation. + +Loss Optimization. According to Eq. 5, we can obtain the denoiser prediction $\hat{X} = \{\hat{x}_1,\hat{x}_2,\dots ,\hat{x}_F\}$ . During training, the loss function constrains the difference between the original latent feature $x$ and the denoiser's prediction $\hat{x}$ : + +$$ +\mathcal {L} _ {\text {d i f f}} = \frac {1}{F} \sum_ {f = 1} ^ {F} | x _ {f} - \hat {x} _ {f} |, \tag {10} +$$ + +where $F$ denotes the number of frames in the video sequence. This loss encourages the model to accurately recover the original latent representation, improving the quality of the generated transitions and ensuring alignment with the encoded video features. + +# 4. Experiments + +# 4.1. Experimental Settings + +Datasets. We conduct the experiments on three datasets, i.e., PHOENIX14T, USTC-CSL100, and USTC-SLR500. The first two are sentence-level continuous sign language datasets, which are used to evaluate supervised transition generation in controlled settings. While the latter one is a word-level isolated sign language dataset, allowing us to assess the model's performance in generating transitions within real-world, discrete sign scenarios. + +- PHOENIX14T [2] comprises 8,257 sentence-level samples, which covers 2,887 German words. For preprocessing, we follow the settings in previous sign language production studies [19, 36], using OpenPose [4] to extract 2D coordinates from the videos, which are then refined to 3D coordinates through a skeleton correction model [51]. +- USTC-CSL100 [18] covers 100 unique sentences performed by 50 signers. We use the split-I [11] approach to divide the dataset into 4,000 training examples and 1,000 test examples. The above two datasets enable us to evaluate the model's effectiveness on continuous sign language sequences in a structured environment. +- USTC-SLR500 [17] is an isolated sign language dataset containing 500 Chinese sign words, each performed five times by 50 signers. Following [17], we split the dataset by using videos from 36 signers for training and the remaining 14 signers for testing. + +Implementation Details. Since generating continuous sign language videos from discrete fragments is inherently an unsupervised task, there are no standardized evaluation metrics. To effectively assess our model's performance, we analyzed transition actions between two gestures on the sentence-level datasets (i.e., PHOENIX14T and USTC-CSL100), observing that transitions typically span 10 to 20 frames. Based on these findings, we designed experiments using two masking strategies: masking 10 frames every 20 frames and masking 20 frames every 10 frames. Additionally, to further validate the effectiveness of our approach, we used the isolated sign language dataset (i.e., USTC-SLR500) to assemble discrete vocabulary videos into continuous sequences, demonstrating the model's capability to produce smooth transitions in a real-world setting. + +Evaluation Metrics. We evaluate our performance on two sentence-level datasets in terms of both semantics and coherence. Following prior work in sign language generation [19, 35, 36, 44], we use the NSLT method [2] to convert generated sign language videos back into text. We then compare this text with reference data to calculate semantic accuracy metrics, including BLEU [33], ROUGE [27], and WER. Additionally, we use FID, DTW, and MPJPE to assess the consistency and smoothness of the generated transitions, measuring the coherence of the results. + +Table 1. Ablation results of different masking ratios on PHOENIX14T (observation 20 frames - prediction 10 frames). + +
Masking Ratio rDEVTEST
B1↑B4↑DTW↓B1↑B4↑DTW↓
r = 0.122.178.040.6122.478.670.60
r = 0.322.468.060.6222.368.730.61
r = 0.522.308.060.7722.028.350.75
+ +Table 2. Ablation results of different padding strategies on PHOENIX14T (observation 20 frames - prediction 10 frames). + +
Padding StrategyDEVTEST
DTW↓FID↓MPJPE↓DTW↓FID↓MPJPE↓
w/o Padding11.462.7469.6011.603.0687.57
Front Padding0.722.6847.900.692.9960.05
Linear Padding0.622.6847.300.612.9859.36
Back Padding0.672.6847.890.662.9959.91
+ +# 4.2. Ablation Study + +In this subsection, we explore the effects of different random masking ratios in the training phase and various padding strategies during inference. The experiments in this part are all conducted under the condition of predicting 10 transition frames for every 20 observed sign frames to evaluate their impact on both semantic accuracy and transition coherence. + +Effect of Masking Ratio. Table 1 presents the results of different random masking ratios $r$ on semantic preservation and temporal coherence. With a lower masking ratio $(r = 0.1)$ , the model achieves the best coherence, reflected by the lowest DTW scores of 0.61 on DEV and 0.60 on TEST. This indicates that smaller masking ratios allow the model to maintain smoother transitions, as more contextual frames are available to guide the generation of missing frames. However, as the masking ratio increases to $r = 0.3$ , we observe a slight improvement in semantic accuracy, with BLEU-4 scores rising to 8.06 on DEV and 8.73 on TEST. This suggests that a higher masking ratio may encourage the model to focus more on the sign semantic content, as it learns to reconstruct more extensive missing segments. + +Interestingly, further increasing the masking ratio to $r = 0.5$ leads to a drop in both semantic accuracy and coherence, with DTW increasing and BLEU decreasing slightly. This highlights a trade-off: while a moderate masking ratio improves semantic preservation, excessive masking may reduce the model's ability to maintain smooth transitions and semantic accuracy. Based on these observations, we select $r = 0.3$ for subsequent experiments, as it provides a good balance between semantic retention and transition smoothness. + +Influence of Padding Strategy. Table 2 summarizes the results of different padding strategies applied during inference. We experiment with four approaches: w/o Padding (no initialization for missing frames), Front Padding (using the last frame of the previous observation segment), Back Padding (using the first frame of the following observation + +Table 3. Performance comparison on PHOENIX14T (under setting: observation 20 frames - prediction 10 frames). + +
MethodsDEVTEST
B1↑B4↑ROUGE↑WER↓DTW↓FID↓MPIJE↓B1↑B4↑ROUGE↑WER↓DTW↓FID↓MPIJE↓
Ground Truth29.7712.1329.6074.170.000.000.0029.7611.9328.9871.940.000.000.00
G2P-DDM [47]8.250.997.1798.9312.712.8555.029.521.298.1298.5412.573.1967.81
VQ-GCDM [43]9.310.877.5998.6912.112.8653.879.850.967.9598.5911.983.2166.51
Ours22.468.0622.0780.730.622.6847.3022.368.7322.3679.060.612.9859.36
+ +Table 4. Performance comparison on PHOENIX14T (under setting: observation 10 frames - prediction 20 frames). + +
MethodsDEVTEST
B1↑B4↑ROUGE↑WER↓DTW↓FID↓MPJPE↓B1↑B4↑ROUGE↑WER↓DTW↓FID↓MPJPE↓
Ground Truth29.7712.1329.6074.170.000.000.0029.7611.9328.9871.940.000.000.00
G2P-DDM [47]8.980.637.4498.6115.952.8955.219.130.747.7598.6415.583.2268.18
VQ-GCDM [43]8.640.897.0898.4812.403.0548.669.021.367.2298.6112.223.3760.27
Ours19.325.9218.3387.300.702.7548.0419.596.2618.5986.80.683.0559.81
+ +Table 5. Performance comparison on USTC-CSL100. + +
MethodsObservation 20 Frames - Prediction 10 FramesObservation 10 Frames - Prediction 20 Frames
B1↑B4↑ROUGE↑WER↓DTW↓FID↓MPJPE↓B1↑B4↑ROUGE↑WER↓DTW↓FID↓MPJPE↓
Ground Truth69.1059.1268.5347.380.000.000.0069.1059.1268.5347.380.000.000.00
G2P-DDM [47]13.742.1113.8592.047.000.53258.4313.431.7413.6292.157.470.53258.43
VQ-GCDM [43]30.7110.6829.1777.875.100.4391.1018.772.0816.6492.6710.890.54104.27
Ours69.6458.9869.2646.661.920.3177.0467.9556.6167.3849.241.790.3178.62
+ +segment), and Linear Padding (performing linear interpolation between the two boundary poses, as Eq. 8). The "w/o Padding" approach yields the highest DTW scores (11.46 - DEV and 10.36 - TEST), indicating poor coherence and unnatural transitions. Both Front Padding and Back Padding improve the smoothness of transitions, as shown by lower DTW scores (0.67 - DEV and 0.66 - TEST), but their FID and MPJPE remain better compared to Linear Padding. + +The Linear Padding strategy achieves the best overall performance, with the lowest DTW scores (0.62 on DEV and 0.66 on TEST) and MPJPE values, reflecting both smoothness and accuracy in pose reconstruction. By initializing missing frames through linear interpolation, Linear Padding provides a stable and effective foundation for generating realistic and seamless transitions, closely resembling real-world signing behavior. + +# 4.3. Comparison with State-of-the-Arts + +We compare our method with two state-of-the-art models: G2P-DDM [47] and VQ-GCDM [43]. Both G2P-DDM and VQ-GCDM are primarily designed for SLP. Due to the absence of established benchmarks for our specific task, we adapted these models to serve as comparative baselines. + +Comparison on PHOENIX14T. As shown in Table 4, under the "Observation 20 Frames - Prediction 10 Frames" setting, our method achieves superior performance across both semantic and coherence metrics. Specifically, we attain a BLEU-1 score of $22.46\%$ and $22.36\%$ on the DEV and TEST sets, respectively, outperforming G2P-DDM and + +VQ-GCDM. This improvement in BLEU scores indicates that our method more accurately preserves semantic content compared to existing methods. Furthermore, our method shows significant gains in temporal coherence, with DTW scores of 0.62 on DEV and 0.61 on TEST, respectively. These results confirm that our model generates smoother and more realistic transitions than the baseline methods. + +Under the more challenging "Observation 10 Frames - Prediction 20 Frames" setting, our method continues to demonstrate robust performance despite the increased prediction length. We achieve BLEU-1 scores of $19.32\%$ on DEV and $19.59\%$ on TEST, considerably higher than the baseline methods. Our method also maintains lower DTW and FID scores, demonstrating its ability to generate coherent and semantically accurate sequences even with limited observed data. The MPJPE scores further support the effectiveness of our approach, with scores of 48.04 on DEV and 59.81 on TEST, indicating a closer alignment to ground truth poses than the other methods. + +Comparison on USTC-CSL100. Table 5 shows the results on the USTC-CSL100 dataset, which evaluates our method's ability to generate sign language sequences. In the "Observation 20 Frames - Prediction 10 Frames" setting, our model achieves near-ground-truth semantic accuracy, with BLEU-1 and BLEU-4 scores of $69.64\%$ and $58.98\%$ on DEV, closely approaching the ground truth scores of $69.10\%$ and $59.12\%$ . Additionally, our DTW score of 1.922 and FID score of 3.21 indicate high temporal consistency and low distributional divergence from the real data. Com + +![](images/cad7faac2655967240f13d92e536dbe0310e82462e52c30edee06bdaf17c4681.jpg) +Figure 3. Visualization examples of generating 10-frame transition pose under 20-frame observations on PHOENIX14T. We compare our method with G2P-DDM and VQ-GCDM, and the Ground Truth is attached. Green: Transitions. Orange and Blue: Observations. + +![](images/b904cb3598f50b51614234e478fd5f038d405c24b8a61c541dbc538df5cf9062.jpg) +Figure 4. Visualization examples on the USTC-SLR500, demonstrate the generation of transition poses according to two discrete sign segments. Green: Transitions. Orange and Blue: Observations. + +pared to G2P-DDM and VQ-GCDM, which yield higher DTW and MPJPE scores, our model achieves smoother and more accurate pose transitions. + +In the "Observation 10 Frames - Prediction 20 Frames" scenario, our method continues to demonstrate resilience in generating long sequences. Our BLEU-1 and BLEU-4 scores on DEV are $67.95\%$ and $56.61\%$ , respectively, showing superior semantic preservation compared to the baseline methods. Additionally, the DTW score of 1.79 and MPJPE score of 78.62 confirm our model's ability to produce coherent and realistic transitions, outperforming the baselines in alignment with true pose sequences. + +# 4.4. Qualitative Results + +Visualization Results. Figure 3 presents a comparative demonstration of our method against other approaches (i.e., G2P-DDM [47] and VQ-GCDM [43]) in generating transition sign poses. This example clearly shows that our method significantly outperforms G2P-DDM and VQ-GCDM in the quality of interpolated movements, especially excelling in overall motion coherence. Furthermore, our method meticulously fine-tunes the observations (the Blue and Orange sections), effectively avoiding coherence issues that may arise from directly concatenating discrete sign segments. + +Verification in Realistic Scenarios. Figure 4 demonstrates + +visualization examples for synthesizing continuous sign videos from discrete segments in a realistic, word-level scenario using the USTC-SLR500 dataset. Here, the observed frames from two isolated sign segments are marked in blue and orange, while the generated transitions are shown in green. The results show that our method not only generates smooth transitions between discrete sign language segments but also fine-tunes the observed poses based on the overall semantics, resulting in a coherent sign language video. + +# 5. Conclusions + +This paper proposes a method to generate continuous sign language videos by synthesizing transition frames between discrete segments, transforming an unsupervised task into a supervised one through random masking. Using a conditional diffusion model, the approach learns to predict missing frames based on observable context, allowing for smooth and temporally coherent outputs. By incorporating a linearly interpolating padding strategy during inference, the model initializes transitions effectively, further refining them into seamless sequences. Extensive experiments demonstrate the effectiveness of leveraging context for natural transitions, advancing sign language synthesis by enabling fluid, continuous sequences. + +# 6. Acknowledgments + +This work was supported by the National Natural Science Foundation of China (Grants No. U23B2031, 62472139, 62302142, 62272144), the Anhui Provincial Natural Science Foundation, China (Grant No. 2408085QF191), the Fundamental Research Funds for the Central Universities (Grants No. JZ2024HGTA0178, JZ2023HGQA0097), and the Major Project of Anhui Province (Grant No. 202423k09020001). + +# References + +[1] Ahmed H Aliwy and A Alethary Ahmed. Development of arabic sign language dictionary using 3d avatar technologies. Indonesian Journal of Electrical Engineering and Computer Science, 21(1):609-616, 2021. 1 +[2] Necati Cihan Camgoz, Simon Hadfield, Oscar Koller, Hermann Ney, and Richard Bowden. Neural sign language translation. In Computer Vision and Pattern Recognition, pages 7784-7793, 2018. 1, 6 +[3] Necati Cihan Camgoz, Oscar Koller, Simon Hadfield, and Richard Bowden. Multi-channel transformers for multi-articulatory sign language translation. In European Conference on Computer Vision, pages 301-319, 2020. 1, 2 +[4] Zhe Cao, Tomas Simon, Shih-En Wei, and Yaser Sheikh. Realtime multi-person 2d pose estimation using part affinity fields. In Computer Vision and Pattern Recognition, pages 7291-7299, 2017. 6 +[5] Xiujuan Chai, Hanjie Wang, and Xilin Chen. The devisesign large vocabulary of Chinese sign language database and baseline evaluations. In Technical report, Key Lab of Intelligent Information Processing of CAS. Institute of Computing Technology, 2014. 1 +[6] Mathieu De Coster and Joni Dambre. Querying a sign language dictionary with videos using dense vector search. In 2023 IEEE International Conference on Acoustics, Speech, and Signal Processing Workshops, pages 1-5. IEEE, 2023. 1 +[7] Philippe Dreuw, Carol Neidle, Vassilis Athitsos, Stan Sclaroff, and Hermann Ney. Benchmark databases for videobased automatic sign language recognition. In International Conference on Language Resources and Evaluation, 2008. 1 +[8] Amanda Cardoso Duarte. Cross-modal neural sign language translation. In ACM International Conference on Multimedia, pages 1650-1654, 2019. 1 +[9] Jia Gong, Lin Geng Foo, Yixuan He, Hossein Rahmani, and Jun Liu. Llms are good sign language translators. In Computer Vision and Pattern Recognition, pages 18362-18372, 2024. 1, 2 +[10] Dan Guo, Wengang Zhou, Houqiang Li, and Meng Wang. Online early-late fusion based on adaptive hmm for sign language recognition. ACM Transactions on Multimedia Computing, Communications, and Applications, pages 1-18, 2017. 2 +[11] Dan Guo, Wengang Zhou, Houqiang Li, and Meng Wang. Hierarchical LSTM for sign language translation. In AAAI Conference on Artificial Intelligence, pages 6845-6852, 2018. 6 + +[12] Dan Guo, Shengeng Tang, and Meng Wang. Connectionist temporal modeling of video and language: A joint model for translation and sign labeling. In International Joint Conference on Artificial Intelligence, pages 751-757, 2019. 2 +[13] Dan Guo, Shengeng Tang, Richang Hong, and Meng Wang. Sign language recognition. Multimedia for Accessible Human Computer Interfaces, pages 23-59, 2021. 1 +[14] Jonathan Ho, Ajay Jain, and Pieter Abbeel. Denoising diffusion probabilistic models. Conference and Workshop on Neural Information Processing Systems, 33:6840-6851, 2020. 3 +[15] Hezhen Hu, Junfu Pu, Wengang Zhou, and Houqiang Li. Collaborative multilingual continuous sign language recognition: A unified framework. IEEE Transactions on Multimedia, pages 7559-7570, 2022. 1, 2 +[16] Hezhen Hu, Weichao Zhao, Wengang Zhou, and Houqiang Li. Signbert+: Hand-model-aware self-supervised pretraining for sign language understanding. IEEE Transactions on Pattern Analysis and Machine Intelligence, pages 11221-11239, 2023. 1, 2 +[17] Jie Huang, Wengang Zhou, Houqiang Li, and Weiping Li. Attention-based 3d-cnns for large-vocabulary sign language recognition. IEEE Transactions on Circuits and Systems for Video Technology, pages 2822-2832, 2018. 6 +[18] Jie Huang, Wengang Zhou, Qilin Zhang, Houqiang Li, and Weiping Li. Video-based sign language recognition without temporal segmentation. In AAAI Conference on Artificial Intelligence, pages 2257-2264, 2018. 1, 6 +[19] Wencan Huang, Wenwen Pan, Zhou Zhao, and Qi Tian. Towards fast and high-quality sign language production. In ACM International Conference on Multimedia, pages 3172-3181, 2021. 6 +[20] Jestin Joy, Kannan Balakrishnan, and Sreeraj Madhavankutty. Developing a bilingual mobile dictionary for indian sign language and gathering users experience with signdict. Assistive Technology, 2020. 1 +[21] Korrawe Karunratanakul, Konpat Preechakul, Supasorn Suwajanakorn, and Siyu Tang. Guided motion diffusion for controllable human motion synthesis. In International Conference on Computer Vision, pages 2151-2162, 2023. 3 +[22] Oscar Koller. Quantitative survey of the state of the art in sign language recognition. ArXiv Preprint ArXiv:2008.09918, pages ArXiv-2008, 2020. 1, 2 +[23] Taegyeong Lee, Soyeong Kwon, and Taehwan Kim. Grid diffusion models for text-to-video generation. In Computer Vision and Pattern Recognition, pages 8734-8743, 2024. 2 +[24] Dongxu Li, Cristian Rodriguez, Xin Yu, and Hongdong Li. Word-level deep sign language recognition from video: A new large-scale dataset and methods comparison. In IEEE Winter Conference on Applications of Computer Vision, pages 1459-1469, 2020. 1 +[25] Dongxu Li, Chenchen Xu, Xin Yu, Kaihao Zhang, Benjamin Swift, Hanna Suominen, and Hongdong Li. Tspnet: Hierarchical feature learning via temporal semantic pyramid for sign language translation. Neural Information Processing Systems, pages 12034-12045, 2020. 2 +[26] Jingyun Liang, Yuchen Fan, Kai Zhang, Radu Timofte, Luc Van Gool, and Rakesh Ranjan. Movideo: Motion-aware + +video generation with diffusion model. In European Conference on Computer Vision, pages 56-74. Springer, 2025. 3 +[27] Chin-Yew Lin. Rouge: A package for automatic evaluation of summaries. In Text Summarization Branches Out, pages 74-81, 2004. 6 +[28] Seshadri Mazumder, Rudrabha Mukhopadhyay, Vinay P Namboodiri, and CV Jawahar. Translating sign language videos to talking faces. In Indian Conference on Computer Vision, Graphics and Image Processing, pages 1-10, 2021. 2 +[29] John McDonald, Rosalee Wolfe, Jerry Schnegg, Julie Hochgesang, Diana Gorman Jamrozik, Marie Stumbo, Larwan Berke, Melissa Bialek, and Farah Thomas. An automated technique for real-time production of lifelike animations of american sign language. Universal Access in the Information Society, pages 551-566, 2016. 2 +[30] Yuchun Miao, Lefei Zhang, Liangpei Zhang, and Dacheng Tao. Dds2m: Self-supervised denoising diffusion spati-spectral model for hyperspectral image restoration. In International Conference on Computer Vision, pages 12086-12096, 2023. 3 +[31] Shentong Mo, Enze Xie, Ruihang Chu, Lanqing Hong, Matthias Niessner, and Zhenguo Li. Dit-3d: Exploring plain diffusion transformers for 3d shape generation. Neural Information Processing Systems, 36:67960-67971, 2023. 3 +[32] Haomiao Ni, Changhao Shi, Kai Li, Sharon X Huang, and Martin Renqiang Min. Conditional image-to-video generation with latent flow diffusion models. In Computer Vision and Pattern Recognition, pages 18444-18455, 2023. 3 +[33] Kishore Papineni, Salim Roukos, Todd Ward, and Wei-Jing Zhu. Bleu: a method for automatic evaluation of machine translation. In Association for Computational Linguistics, pages 311-318, 2002. 6 +[34] Robin Rombach, Andreas Blattmann, Dominik Lorenz, Patrick Esser, and Björn Ommer. High-resolution image synthesis with latent diffusion models. In Computer Vision and Pattern Recognition, pages 10684-10695, 2022. 2 +[35] Ben Saunders, Necati Cihan Camgoz, and Richard Bowden. Progressive transformers for end-to-end sign language production. In European Conference on Computer Vision, pages 687-705, 2020. 2, 6 +[36] Ben Saunders, Necati Cihan Camgoz, and Richard Bowden. Mixed signals: Sign language production via a mixture of motion primitives. In International Conference on Computer Vision, pages 1919-1929, 2021. 6 +[37] Jérémie Segouat. A study of sign language coarticulation. ACM Sigaccess Accessibility and Computing, pages 31-38, 2009. 2 +[38] Wenkang Shan, Zhenhua Liu, Xinfeng Zhang, Zhao Wang, Kai Han, Shanshe Wang, Siwei Ma, and Wen Gao. Diffusion-based 3d human pose estimation with multi-hypothesis aggregation. In International Conference on Computer Vision, pages 14761-14771, 2023. 3 +[39] Jascha Sohl-Dickstein, Eric Weiss, Niru Maheswaranathan, and Surya Ganguli. Deep unsupervised learning using nonequilibrium thermodynamics. In International Conference on Machine Learning, pages 2256-2265, 2015. 2 + +[40] Stephanie Stoll, Necati Cihan Camgoz, Simon Hadfield, and Richard Bowden. Text2sign: Towards sign language production using neural machine translation and generative adversarial networks. International Journal of Computer Vision, pages 891–908, 2020. 2 +[41] Sugandhi, Parteek Kumar, and Sanmeet Kaur. Sign language generation system based on indian sign language grammar. ACM Transactions on Asian and Low-Resource Language Information Processing, 19(4):1-26, 2020. 2 +[42] Shengeng Tang, Dan Guo, Richang Hong, and Meng Wang. Graph-based multimodal sequential embedding for sign language translation. IEEE Transactions on Multimedia, 24: 4433-4445, 2022. 1 +[43] Shengeng Tang, Feng Xue, Jingjing Wu, Shuo Wang, and Richang Hong. Gloss-driven conditional diffusion models for sign language production. ACM Transactions on Multimedia Computing, Communications, and Applications, 2024. 2, 3, 7, 8 +[44] Carla Viegas, Mert Inan, Lorna Quandt, and Malihe Alikhani. Including facial expressions in contextual embeddings for sign language generation. In Joint Conference on Lexical and Computational Semantics, pages 1-10, 2023. 6 +[45] Bin Xia, Yulun Zhang, Shiyin Wang, Yitong Wang, Xinglong Wu, Yapeng Tian, Wenming Yang, and Luc Van Gool. Diffir: Efficient diffusion model for image restoration. In International Conference on Computer Vision, pages 13095-13105, 2023. 3 +[46] Jinheng Xie, Yuexiang Li, Yawen Huang, Haozhe Liu, Wentian Zhang, Yefeng Zheng, and Mike Zheng Shou. Boxdiff: Text-to-image synthesis with training-free box-constrained diffusion. In International Conference on Computer Vision, pages 7452-7461, 2023. 3 +[47] Pan Xie, Qipeng Zhang, Peng Taiying, Hao Tang, Yao Du, and Zexian Li. G2p-ddm: Generating sign pose sequence from gloss sequence with discrete diffusion model. In AAAI Conference on Artificial Intelligence, pages 6234-6242, 2024. 2, 3, 7, 8 +[48] Wanli Xue, Jingze Liu, Siyi Yan, Yuxi Zhou, Tiantian Yuan, and Qing Guo. Alleviating data insufficiency for Chinese sign language recognition. Visual Intelligence, page 26, 2023. 2 +[49] Shuai Yang, Yifan Zhou, Ziwei Liu, and Chen Change Loy. Fresco: Spatial-temporal correspondence for zero-shot video translation. In Computer Vision and Pattern Recognition, pages 8703-8712, 2024. 2 +[50] Xunpeng Yi, Han Xu, Hao Zhang, Linfeng Tang, and Jiayi Ma. Diff-retinex: Rethinking low-light image enhancement with a generative diffusion model. In International Conference on Computer Vision, pages 12302-12311, 2023. 3 +[51] Jan Zelinka and Jakub Kanis. Neural sign language synthesis: Words are our glosses. In Winter Conference on Applications of Computer Vision, pages 3395-3403, 2020. 2, 6 +[52] Ni Zeng, Yiqiang Chen, Yang Gu, Dongdong Liu, and Yunbing Xing. Highly fluent sign language synthesis based on variable motion frame interpolation. In 2020 IEEE International Conference on Systems, Man, and Cybernetics, pages 1772-1777. IEEE, 2020. 2 + +[53] Jihai Zhang, Wengang Zhou, Chao Xie, Junfu Pu, and Houqiang Li. Chinese sign language recognition with adaptive hmm. In IEEE International Conference on Multimedia and Expo, pages 1-6. IEEE, 2016. 1 +[54] Weichao Zhao, Hezhen Hu, Wengang Zhou, Jiaxin Shi, and Houqiang Li. Best: Bert pre-training for sign language recognition with coupling tokenization. In AAAI Conference on Artificial Intelligence, pages 3597-3605, 2023. 1, 2 \ No newline at end of file diff --git a/2025/Discrete to Continuous_ Generating Smooth Transition Poses from Sign Language Observations/images.zip b/2025/Discrete to Continuous_ Generating Smooth Transition Poses from Sign Language Observations/images.zip new file mode 100644 index 0000000000000000000000000000000000000000..6cdb6de5a87fc43818b0a2fa1bd147590d1f7ff7 --- /dev/null +++ b/2025/Discrete to Continuous_ Generating Smooth Transition Poses from Sign Language Observations/images.zip @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:6d8e072838449a6124cca7456038b21b57dcb9aafc4aafd23dda1cacda4cb4f7 +size 569142 diff --git a/2025/Discrete to Continuous_ Generating Smooth Transition Poses from Sign Language Observations/layout.json b/2025/Discrete to Continuous_ Generating Smooth Transition Poses from Sign Language Observations/layout.json new file mode 100644 index 0000000000000000000000000000000000000000..f5ba5bccc79cc181fcab1d8fd940a1183f9a3046 --- /dev/null +++ b/2025/Discrete to Continuous_ Generating Smooth Transition Poses from Sign Language Observations/layout.json @@ -0,0 +1,8143 @@ +{ + "pdf_info": [ + { + "para_blocks": [ + { + "bbox": [ + 117, + 103, + 493, + 140 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 117, + 103, + 493, + 140 + ], + "spans": [ + { + "bbox": [ + 117, + 103, + 493, + 140 + ], + "type": "text", + "content": "Discrete to Continuous: Generating Smooth Transition Poses from Sign Language Observations" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 78, + 161, + 531, + 204 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 78, + 161, + 531, + 204 + ], + "spans": [ + { + "bbox": [ + 78, + 161, + 531, + 204 + ], + "type": "text", + "content": "Shengeng Tang Jiayi He Lechao Cheng* Jingjing Wu Dan Guo Richang Hong* \nSchool of Computer Science and Information Engineering, Hefei University of Technology \n{tangsg, chenglc, wujingjing, guodan, hongrc}@h Fut.edu.cn, hejy4396@mail.hfut.edu.cn" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 151, + 231, + 200, + 243 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 151, + 231, + 200, + 243 + ], + "spans": [ + { + "bbox": [ + 151, + 231, + 200, + 243 + ], + "type": "text", + "content": "Abstract" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 55, + 256, + 296, + 544 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 256, + 296, + 544 + ], + "spans": [ + { + "bbox": [ + 55, + 256, + 296, + 544 + ], + "type": "text", + "content": "Generating continuous sign language videos from discrete segments is challenging due to the need for smooth transitions that preserve natural flow and meaning. Traditional approaches that simply concatenate isolated signs often result in abrupt transitions, disrupting video coherence. To address this, we propose a novel framework, Sign-D2C, that employs a conditional diffusion model to synthesize contextually smooth transition frames, enabling the seamless construction of continuous sign language sequences. Our approach transforms the unsupervised problem of transition frame generation into a supervised training task by simulating the absence of transition frames through random masking of segments in long-duration sign videos. The model learns to predict these masked frames by denoising Gaussian noise, conditioned on the surrounding sign observations, allowing it to handle complex, unstructured transitions. During inference, we apply a linearly interpolating padding strategy that initializes missing frames through interpolation between boundary frames, providing a stable foundation for iterative refinement by the diffusion model. Extensive experiments on the PHOENIX14T, USTCCSL100, and USTC-SLR500 datasets demonstrate the effectiveness of our method in producing continuous, natural sign language videos." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 56, + 567, + 135, + 578 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 567, + 135, + 578 + ], + "spans": [ + { + "bbox": [ + 56, + 567, + 135, + 578 + ], + "type": "text", + "content": "1. Introduction" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 55, + 587, + 295, + 694 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 587, + 295, + 694 + ], + "spans": [ + { + "bbox": [ + 55, + 587, + 295, + 694 + ], + "type": "text", + "content": "Sign language [3, 9, 13, 15, 16, 22, 42, 54], as a unique form of visual communication, presents an intricate challenge for data collection, especially when scaling datasets for continuous, long-duration sign sentences. Unlike textual language, which can be assembled seamlessly from words to sentences, sign language requires a fluid connection between individual signs to convey meaning accurately. This connection involves not only the specific signs but also the transitions between them, which convey tem-" + } + ] + } + ], + "index": 7 + }, + { + "type": "image", + "bbox": [ + 317, + 230, + 553, + 423 + ], + "blocks": [ + { + "bbox": [ + 317, + 230, + 553, + 423 + ], + "lines": [ + { + "bbox": [ + 317, + 230, + 553, + 423 + ], + "spans": [ + { + "bbox": [ + 317, + 230, + 553, + 423 + ], + "type": "image", + "image_path": "a34c3d1d814e76cbdfae282031a899ea45cf3fc0faaafb7adb1ff387663d6df7.jpg" + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 313, + 432, + 555, + 498 + ], + "lines": [ + { + "bbox": [ + 313, + 432, + 555, + 498 + ], + "spans": [ + { + "bbox": [ + 313, + 432, + 555, + 498 + ], + "type": "text", + "content": "Figure 1. Task and key steps. Our work aims to generate continuous sign videos by creating transition poses between discrete segments. In training, random masking simulates missing transitions, and the model learns to recover these frames (steps 1-3). During inference, padding initializes missing transitions, which the model refines to generate smooth, coherent sequences (steps 4-6)." + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "image_caption" + } + ], + "index": 8 + }, + { + "bbox": [ + 313, + 521, + 555, + 593 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 521, + 555, + 593 + ], + "spans": [ + { + "bbox": [ + 313, + 521, + 555, + 593 + ], + "type": "text", + "content": "poral and contextual coherence vital to understanding the intended message. As a result, while word-level sign language data [5, 7, 24, 53] has been collected extensively, there remains a critical gap in developing robust, large-scale datasets for sentence-level sign language [2, 8, 18], particularly for continuous sequences." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 313, + 594, + 556, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 594, + 556, + 715 + ], + "spans": [ + { + "bbox": [ + 313, + 594, + 556, + 715 + ], + "type": "text", + "content": "A possible workaround involves pre-collecting word-level sign data akin to a dictionary [1, 6, 20] and then attempting to form sentences by joining these discrete signs. However, assembling sign sentences is fundamentally different from assembling textual sentences due to the need for visually smooth transitions between signs, which maintain natural flow and meaning. Direct concatenation of individual signs into sentences often results in abrupt and unnatural transitions, disrupting the continuity of the video and potentially obscuring meaning. Consequently, generating smooth" + } + ] + } + ], + "index": 11 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "spans": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "type": "text", + "content": "CVF" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 145, + 0, + 496, + 37 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 145, + 0, + 496, + 37 + ], + "spans": [ + { + "bbox": [ + 145, + 0, + 496, + 37 + ], + "type": "text", + "content": "This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 67, + 703, + 148, + 712 + ], + "type": "page_footnote", + "angle": 0, + "lines": [ + { + "bbox": [ + 67, + 703, + 148, + 712 + ], + "spans": [ + { + "bbox": [ + 67, + 703, + 148, + 712 + ], + "type": "text", + "content": "*Corresponding authors." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 295, + 748, + 314, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 314, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 314, + 757 + ], + "type": "text", + "content": "3481" + } + ] + } + ], + "index": 13 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 0 + }, + { + "para_blocks": [ + { + "bbox": [ + 55, + 72, + 294, + 120 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 72, + 294, + 120 + ], + "spans": [ + { + "bbox": [ + 55, + 72, + 294, + 120 + ], + "type": "text", + "content": "transition frames between sign segments is essential to synthesizing continuous sign language videos [41, 51, 52], yet presents significant challenges due to the lack of supervised reference data for the missing transition frames." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 56, + 120, + 298, + 529 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 120, + 298, + 529 + ], + "spans": [ + { + "bbox": [ + 56, + 120, + 298, + 529 + ], + "type": "text", + "content": "To address this issue, we propose a smooth transition pose generation framework, named Sign-D2C, to synthesize Continuous sign language videos from Discrete sign segments. This framework is based on a diffusion model, which learns to generate missing frames by observing surrounding sign frames. Specifically, our method leverages a conditional diffusion model trained to predict the missing transition frames by sampling from Gaussian noise and being conditioned by the observed video frames. As illustrated in Fig. 1, our method consists of a two-phase process designed to overcome the limitations of direct sign concatenation. In the training phase, we simulate the transition frame generation task by randomly masking frames within long sign language videos. This random masking effectively removes video segments, mimicking real-world scenarios where transition frames are absent. The model is then trained to reconstruct these missing frames based on the observable surrounding frames. By doing so, the model learns to generate plausible transitions in a supervised manner, enhancing its robustness and adaptability to complex, unstructured transitions. In the inference phase, we introduce a linearly interpolating padding strategy to initialize the missing transition frames. This strategy begins by selecting the last frame of the preceding observed segment and the first frame of the following observed segment as the boundary frames for the missing transition segment. The intermediate frames within the missing segment are then initialized by linearly interpolating between these boundary frames, creating an initial linearly interpolating padding that provides a stable foundation for generating smooth transitions. The conditional diffusion model then iteratively refines these frames to produce naturalistic transition frames that connect the observed segments seamlessly. The contributions of this work are summarized as follows:" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 55, + 534, + 294, + 713 + ], + "type": "list", + "angle": 0, + "index": 5, + "blocks": [ + { + "bbox": [ + 55, + 534, + 294, + 594 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 534, + 294, + 594 + ], + "spans": [ + { + "bbox": [ + 55, + 534, + 294, + 594 + ], + "type": "text", + "content": "- We propose a novel conditional diffusion framework, Sign-D2C, specifically designed to synthesize smooth and contextually accurate transition frames for continuous sign videos, overcoming the limitations of traditional word-to-sentence assembly techniques in sign language." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 55, + 594, + 294, + 665 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 594, + 294, + 665 + ], + "spans": [ + { + "bbox": [ + 55, + 594, + 294, + 665 + ], + "type": "text", + "content": "- To simulate the generation of transition frames, we introduce a random masking strategy that removes sections of long sign language videos during training, enabling the model to predict missing frames in a supervised manner. This approach enhances the model's ability to handle diverse, complex transitions." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 56, + 665, + 294, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 665, + 294, + 713 + ], + "spans": [ + { + "bbox": [ + 56, + 665, + 294, + 713 + ], + "type": "text", + "content": "- We design a linearly interpolating padding strategy for transition initialization in the inference phase. This strategy utilizes the boundary frames of observed segments to initialize the missing poses with linear interpolation," + } + ] + } + ], + "index": 4 + } + ], + "sub_type": "text" + }, + { + "bbox": [ + 322, + 72, + 553, + 95 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 322, + 72, + 553, + 95 + ], + "spans": [ + { + "bbox": [ + 322, + 72, + 553, + 95 + ], + "type": "text", + "content": "which provides a stable basis for generating visually seamless transitions." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 314, + 96, + 553, + 156 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 314, + 96, + 553, + 156 + ], + "spans": [ + { + "bbox": [ + 314, + 96, + 553, + 156 + ], + "type": "text", + "content": "- Extensive experiments on three standard benchmarks (i.e., Phoenix14T, UCST-CSL100, and USTC-SLR500) demonstrate the robustness and effectiveness of our approach in producing smooth, contextually accurate transition poses for continuous sign language videos." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 314, + 168, + 400, + 181 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 314, + 168, + 400, + 181 + ], + "spans": [ + { + "bbox": [ + 314, + 168, + 400, + 181 + ], + "type": "text", + "content": "2. Related Work" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 313, + 188, + 520, + 202 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 188, + 520, + 202 + ], + "spans": [ + { + "bbox": [ + 313, + 188, + 520, + 202 + ], + "type": "text", + "content": "2.1. Sign Language Production & Synthesis" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 313, + 206, + 553, + 302 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 206, + 553, + 302 + ], + "spans": [ + { + "bbox": [ + 313, + 206, + 553, + 302 + ], + "type": "text", + "content": "Sign language plays a vital role in daily life and has become a prominent topic in artificial intelligence. Early research primarily focused on Sign Language Recognition (SLR) [10, 15, 16, 22, 48, 54] and Sign Language Translation (SLT) [3, 9, 12, 25]. However, with the growing need for effective communication between deaf and hearing individuals, Sign Language Production (SLP) has gained significant attention in recent years." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 313, + 303, + 554, + 459 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 303, + 554, + 459 + ], + "spans": [ + { + "bbox": [ + 313, + 303, + 554, + 459 + ], + "type": "text", + "content": "Early approaches to SLP utilize synthetic animation techniques [28, 29, 37] to convert sentences into sign animations. These techniques, however, incur high collection costs and are limited to specific terms. With advancements in deep learning models, SLP has seen more extensive exploration. Stoll et al. [40] proposes a multi-step approach for generating sign language videos from text, involving the transformation of text into gloss, gloss into poses, and finally, poses into video. Building on this, Saunders et al. [35] introduces a model that generates sign language poses through autoregression. Diffusion-based solutions like G2P-DDM [47] and GCDM [43] further advance SLP by generating sign poses under semantic guidance." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 313, + 459, + 554, + 602 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 459, + 554, + 602 + ], + "spans": [ + { + "bbox": [ + 313, + 459, + 554, + 602 + ], + "type": "text", + "content": "Despite significant progress in these methods, the limitations of existing data have become a major bottleneck in the development of sign language technology. Previous Sign Language Synthesis (SLS) approaches [51] focus on extracting gestures from continuous sign language videos. However, these methods still face challenges in synthesizing continuous, long-duration sign language sentences. Our method addresses this gap by starting with word-level sign language pose data, generating smooth transition frames between individual signs, and combining these segments into sentence-level sequences, thereby overcoming the limitations of existing methods." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 313, + 611, + 502, + 624 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 611, + 502, + 624 + ], + "spans": [ + { + "bbox": [ + 313, + 611, + 502, + 624 + ], + "type": "text", + "content": "2.2. Diffusion-based Generative Models" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 313, + 629, + 553, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 629, + 553, + 715 + ], + "spans": [ + { + "bbox": [ + 313, + 629, + 553, + 715 + ], + "type": "text", + "content": "Diffusion models have gained considerable attention in generative modeling since they were first introduced by Sohl-Dickstein et al. [39]. Its forward-backward diffusion process has proven highly effective in generating realistic and diverse outputs, enabling diffusion models to achieve competitive performance in various generative tasks, including image synthesis [34], video generation [23, 49], and 3D shape" + } + ] + } + ], + "index": 14 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "text", + "content": "3482" + } + ] + } + ], + "index": 15 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 1 + }, + { + "para_blocks": [ + { + "bbox": [ + 55, + 72, + 294, + 216 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 72, + 294, + 216 + ], + "spans": [ + { + "bbox": [ + 55, + 72, + 294, + 216 + ], + "type": "text", + "content": "generation [31]. One of the strengths of diffusion models is their iterative refinement capability, allowing for flexible and high-quality outputs that align well with tasks requiring fine control over generative processes. This quality has led to their adoption in areas such as image restoration, where they have been used for denoising and superresolution [30, 45, 50], as well as pose estimation, where they can handle uncertainty in human poses [38]. Recent advancements have extended their use to more complex generative tasks, such as controlled text-to-image synthesis [46] and conditional video generation [32], further demonstrating their adaptability and robustness." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 56, + 231, + 296, + 423 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 231, + 296, + 423 + ], + "spans": [ + { + "bbox": [ + 56, + 231, + 296, + 423 + ], + "type": "text", + "content": "In Sign Language Production (SLP), diffusion models remain relatively unexplored but have shown promise in recent studies. For instance, G2P-DDM [47] uses a discrete diffusion model to translate the continuous pose generation task into a discrete space, simplifying the complexity of generating sign language poses while maintaining semantic alignment with gloss-based inputs. This approach allows for finer control over pose accuracy, making it particularly suitable for generating static sign poses from linguistic cues. Similarly, GCDM [43] applies a multi-hypothesis diffusion framework to generate diverse sign poses from gloss, accommodating multiple interpretations and improving the naturalness of synthesized gestures. These models represent pioneering steps in applying diffusion-based approaches to SLP, leveraging the unique properties of diffusion models to generate realistic sign poses under semantic guidance." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 56, + 438, + 296, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 438, + 296, + 714 + ], + "spans": [ + { + "bbox": [ + 56, + 438, + 296, + 714 + ], + "type": "text", + "content": "Despite these advancements, current diffusion-based methods for SLP are primarily focused on generating sign poses according to textual signals rather than addressing the temporal continuity required for longer sign language sequences. In sign language synthesis, especially when producing smooth transitions in continuous sign language videos, maintaining smooth transitions between gestures is essential for ensuring that the sequences are fluid and natural. This need for temporally smooth transitions has led to research on methods that can generate these connecting frames between segments, although it remains an area with limited exploration in SLP. Beyond SLP, diffusion models have been successfully applied to tasks requiring temporal coherence and frame interpolation in video synthesis. For example, diffusion-based models have been employed in video generation to ensure smooth transitions between frames [26], and similar techniques have been explored in action synthesis and motion completion [21], where maintaining continuity across frames is crucial. Such applications highlight the potential of diffusion models to handle tasks involving complex temporal dependencies, suggesting that similar techniques could be valuable in SLP for generating continuous sign language videos." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 314, + 71, + 397, + 85 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 314, + 71, + 397, + 85 + ], + "spans": [ + { + "bbox": [ + 314, + 71, + 397, + 85 + ], + "type": "text", + "content": "3. Methodology" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 313, + 91, + 555, + 200 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 91, + 555, + 200 + ], + "spans": [ + { + "bbox": [ + 313, + 91, + 555, + 200 + ], + "type": "text", + "content": "In this work, we propose a framework for generating smooth transition frames between discrete sign segments to create continuous sign pose videos. Our approach transforms an unsupervised transition generation task into a supervised framework using a random masking strategy based on conditional diffusion. This section will outline the preliminaries (Sec. 3.1), followed by our core idea of realizing the task (Sec. 3.2), the framework structure (Sec. 3.3), and the training process (Sec. 3.4)." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 313, + 208, + 394, + 220 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 208, + 394, + 220 + ], + "spans": [ + { + "bbox": [ + 313, + 208, + 394, + 220 + ], + "type": "text", + "content": "3.1. Preliminary" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 313, + 226, + 555, + 357 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 226, + 555, + 357 + ], + "spans": [ + { + "bbox": [ + 313, + 226, + 555, + 357 + ], + "type": "text", + "content": "The diffusion model [14] is highly effective in generating high-quality images by combining diffusion processes, latent space modeling, and conditional inputs. In this work, we employ a conditional diffusion model to synthesize smooth transition frames between discrete sign segments, forming continuous sign language videos. Here, the model is trained on complete long-duration sign language pose sequences, progressively adding noise to simulate missing transitions. During testing, the model starts from Gaussian noise samples and generates the required transitions between observed sign frames." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 313, + 358, + 554, + 407 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 358, + 554, + 407 + ], + "spans": [ + { + "bbox": [ + 313, + 358, + 554, + 407 + ], + "type": "text", + "content": "In the forward process, noise is incrementally added to the full sign language pose sequence " + }, + { + "bbox": [ + 313, + 358, + 554, + 407 + ], + "type": "inline_equation", + "content": "x_0" + }, + { + "bbox": [ + 313, + 358, + 554, + 407 + ], + "type": "text", + "content": " to obtain a noisy version " + }, + { + "bbox": [ + 313, + 358, + 554, + 407 + ], + "type": "inline_equation", + "content": "x_T" + }, + { + "bbox": [ + 313, + 358, + 554, + 407 + ], + "type": "text", + "content": ", mimicking the degradation of visual continuity. This process is described by:" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 359, + 415, + 553, + 430 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 359, + 415, + 553, + 430 + ], + "spans": [ + { + "bbox": [ + 359, + 415, + 553, + 430 + ], + "type": "interline_equation", + "content": "x _ {t} = \\alpha_ {t} \\cdot x _ {t - 1} + \\sqrt {1 - \\alpha_ {t}} \\cdot \\epsilon_ {t}, \\tag {1}", + "image_path": "4cf30feb55960d70408fdbe69f81fb86bce67ea136cbd91dddea76875c5a4caf.jpg" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 313, + 440, + 554, + 475 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 440, + 554, + 475 + ], + "spans": [ + { + "bbox": [ + 313, + 440, + 554, + 475 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 313, + 440, + 554, + 475 + ], + "type": "inline_equation", + "content": "\\alpha_{t}" + }, + { + "bbox": [ + 313, + 440, + 554, + 475 + ], + "type": "text", + "content": " controls the level of noise at each step " + }, + { + "bbox": [ + 313, + 440, + 554, + 475 + ], + "type": "inline_equation", + "content": "t" + }, + { + "bbox": [ + 313, + 440, + 554, + 475 + ], + "type": "text", + "content": ", and " + }, + { + "bbox": [ + 313, + 440, + 554, + 475 + ], + "type": "inline_equation", + "content": "\\epsilon_{t}" + }, + { + "bbox": [ + 313, + 440, + 554, + 475 + ], + "type": "text", + "content": " represents Gaussian noise. As " + }, + { + "bbox": [ + 313, + 440, + 554, + 475 + ], + "type": "inline_equation", + "content": "t" + }, + { + "bbox": [ + 313, + 440, + 554, + 475 + ], + "type": "text", + "content": " increases, the original sequence " + }, + { + "bbox": [ + 313, + 440, + 554, + 475 + ], + "type": "inline_equation", + "content": "x_0" + }, + { + "bbox": [ + 313, + 440, + 554, + 475 + ], + "type": "text", + "content": " transitions to pure noise " + }, + { + "bbox": [ + 313, + 440, + 554, + 475 + ], + "type": "inline_equation", + "content": "x_{T}" + }, + { + "bbox": [ + 313, + 440, + 554, + 475 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 313, + 476, + 554, + 535 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 476, + 554, + 535 + ], + "spans": [ + { + "bbox": [ + 313, + 476, + 554, + 535 + ], + "type": "text", + "content": "The reverse process reconstructs the continuous sequence by progressively removing noise, where a neural network is trained to predict each denoised frame based on previous noisy frames. The denoising step at each iteration is defined as:" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 325, + 546, + 553, + 559 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 325, + 546, + 553, + 559 + ], + "spans": [ + { + "bbox": [ + 325, + 546, + 553, + 559 + ], + "type": "interline_equation", + "content": "p \\left(x _ {t - 1} \\mid x _ {t}, x _ {\\text {o b s}}\\right) = \\mathcal {N} \\left(x _ {t - 1}; \\mu_ {\\theta} \\left(x _ {t}, t, x _ {\\text {o b s}}\\right), \\Sigma_ {t}\\right), \\tag {2}", + "image_path": "af9916ef7e6cb822ecbf37353cc355380958bf8e8324e9d199640d20ce5a32f3.jpg" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 313, + 570, + 554, + 628 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 570, + 554, + 628 + ], + "spans": [ + { + "bbox": [ + 313, + 570, + 554, + 628 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 313, + 570, + 554, + 628 + ], + "type": "inline_equation", + "content": "x_{\\mathrm{obs}}" + }, + { + "bbox": [ + 313, + 570, + 554, + 628 + ], + "type": "text", + "content": " denotes the observed sign language frames that serve as conditional inputs, and " + }, + { + "bbox": [ + 313, + 570, + 554, + 628 + ], + "type": "inline_equation", + "content": "\\mu_{\\theta}(x_t,t,x_{\\mathrm{obs}})" + }, + { + "bbox": [ + 313, + 570, + 554, + 628 + ], + "type": "text", + "content": " is the predicted mean for the denoising process. The observed frames condition the diffusion model, guiding it to produce smooth, contextually consistent transition frames." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 313, + 630, + 554, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 630, + 554, + 714 + ], + "spans": [ + { + "bbox": [ + 313, + 630, + 554, + 714 + ], + "type": "text", + "content": "In our model, the training input is a complete long-duration sign language pose video, gradually noised to simulate missing transitions. During testing, Gaussian noise serves as the input, and the model generates a continuous video with smooth transition frames conditioned on observed frames " + }, + { + "bbox": [ + 313, + 630, + 554, + 714 + ], + "type": "inline_equation", + "content": "x_{\\mathrm{obs}}" + }, + { + "bbox": [ + 313, + 630, + 554, + 714 + ], + "type": "text", + "content": ", effectively bridging discrete sign segments into a fluid sign language sequence." + } + ] + } + ], + "index": 13 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "text", + "content": "3483" + } + ] + } + ], + "index": 14 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 2 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 59, + 72, + 550, + 167 + ], + "blocks": [ + { + "bbox": [ + 59, + 72, + 550, + 167 + ], + "lines": [ + { + "bbox": [ + 59, + 72, + 550, + 167 + ], + "spans": [ + { + "bbox": [ + 59, + 72, + 550, + 167 + ], + "type": "image", + "image_path": "10f8ad0591c27acc5272163fac9914faf71f533ade0ba56830fcad116f74e3bf.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + } + ], + "index": 0 + }, + { + "type": "image", + "bbox": [ + 59, + 171, + 550, + 261 + ], + "blocks": [ + { + "bbox": [ + 59, + 171, + 550, + 261 + ], + "lines": [ + { + "bbox": [ + 59, + 171, + 550, + 261 + ], + "spans": [ + { + "bbox": [ + 59, + 171, + 550, + 261 + ], + "type": "image", + "image_path": "eb02065371f8d11e017dca2ef7790a6ca4f0f5e84baa494082c6b1e828f7207d.jpg" + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 55, + 272, + 555, + 339 + ], + "lines": [ + { + "bbox": [ + 55, + 272, + 555, + 339 + ], + "spans": [ + { + "bbox": [ + 55, + 272, + 555, + 339 + ], + "type": "text", + "content": "Figure 2. Overview. The proposed framework for generating continuous sign language videos with smooth transitions between discrete segments. In the training phase (top), a long-duration sign video undergoes random masking to create gaps, simulating missing transitions. An encoder processes the video, and a conditional diffusion model denoises and predicts the masked segments based on observable frames. The recovered video is produced through a decoder, with both encoder and decoder pre-trained for effective representation. In the inference phase (bottom), discrete sign videos are connected using a linearly interpolating padding strategy to initialize missing transitions. Gaussian noise is applied, and the model refines these transitions, generating a continuous sign video with coherent flow." + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_caption" + } + ], + "index": 1 + }, + { + "bbox": [ + 55, + 359, + 262, + 372 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 359, + 262, + 372 + ], + "spans": [ + { + "bbox": [ + 55, + 359, + 262, + 372 + ], + "type": "text", + "content": "3.2. Core Idea: Unsupervised to Supervised" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 55, + 385, + 295, + 493 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 385, + 295, + 493 + ], + "spans": [ + { + "bbox": [ + 55, + 385, + 295, + 493 + ], + "type": "text", + "content": "Generating continuous sign language videos by adding transitions between discrete segments is challenging, as these segments typically start and end with specific gestures that make direct concatenation impractical. Additionally, this transition generation task is inherently unsupervised, as real transition frames are generally unavailable for training. To address this, we convert the unsupervised problem of generating transitions into a supervised learning framework by leveraging existing continuous sign language videos." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 55, + 502, + 295, + 597 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 502, + 295, + 597 + ], + "spans": [ + { + "bbox": [ + 55, + 502, + 295, + 597 + ], + "type": "text", + "content": "Given a complete continuous sign language sequence " + }, + { + "bbox": [ + 55, + 502, + 295, + 597 + ], + "type": "inline_equation", + "content": "X = \\{x_{1}, x_{2}, \\dots, x_{n}\\}" + }, + { + "bbox": [ + 55, + 502, + 295, + 597 + ], + "type": "text", + "content": ", we simulate missing transitions by randomly masking portions of " + }, + { + "bbox": [ + 55, + 502, + 295, + 597 + ], + "type": "inline_equation", + "content": "X" + }, + { + "bbox": [ + 55, + 502, + 295, + 597 + ], + "type": "text", + "content": ", creating gaps that mimic real-world transition absences. Let the masked frames be denoted as " + }, + { + "bbox": [ + 55, + 502, + 295, + 597 + ], + "type": "inline_equation", + "content": "M \\subset X" + }, + { + "bbox": [ + 55, + 502, + 295, + 597 + ], + "type": "text", + "content": ", and the observable pose frames as " + }, + { + "bbox": [ + 55, + 502, + 295, + 597 + ], + "type": "inline_equation", + "content": "X_{\\mathrm{obs}} = (X \\setminus X_{\\mathrm{mask}})" + }, + { + "bbox": [ + 55, + 502, + 295, + 597 + ], + "type": "text", + "content": ". This setup frames the task as predicting " + }, + { + "bbox": [ + 55, + 502, + 295, + 597 + ], + "type": "inline_equation", + "content": "X_{\\mathrm{mask}}" + }, + { + "bbox": [ + 55, + 502, + 295, + 597 + ], + "type": "text", + "content": " based on " + }, + { + "bbox": [ + 55, + 502, + 295, + 597 + ], + "type": "inline_equation", + "content": "X_{\\mathrm{obs}}" + }, + { + "bbox": [ + 55, + 502, + 295, + 597 + ], + "type": "text", + "content": ", effectively generating supervised training data for transition prediction." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 55, + 605, + 295, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 605, + 295, + 714 + ], + "spans": [ + { + "bbox": [ + 55, + 605, + 295, + 714 + ], + "type": "text", + "content": "By training the model to predict the masked frames " + }, + { + "bbox": [ + 55, + 605, + 295, + 714 + ], + "type": "inline_equation", + "content": "X_{\\mathrm{mask}}" + }, + { + "bbox": [ + 55, + 605, + 295, + 714 + ], + "type": "text", + "content": " conditioned on the observable context " + }, + { + "bbox": [ + 55, + 605, + 295, + 714 + ], + "type": "inline_equation", + "content": "X_{\\mathrm{obs}}" + }, + { + "bbox": [ + 55, + 605, + 295, + 714 + ], + "type": "text", + "content": ", it learns to generate realistic transitions that smoothly bridge the gap. This innovative masking strategy allows our model to leverage the context of surrounding frames provided by " + }, + { + "bbox": [ + 55, + 605, + 295, + 714 + ], + "type": "inline_equation", + "content": "X_{\\mathrm{obs}}" + }, + { + "bbox": [ + 55, + 605, + 295, + 714 + ], + "type": "text", + "content": " to produce temporally coherent connections, transforming the inherently unsupervised transition generation task into a supervised one and enabling effective continuous sign language synthesis." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 314, + 359, + 440, + 372 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 314, + 359, + 440, + 372 + ], + "spans": [ + { + "bbox": [ + 314, + 359, + 440, + 372 + ], + "type": "text", + "content": "3.3. Sign-D2C Framework" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 313, + 376, + 555, + 496 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 376, + 555, + 496 + ], + "spans": [ + { + "bbox": [ + 313, + 376, + 555, + 496 + ], + "type": "text", + "content": "Our approach leverages a conditional diffusion model to generate transition frames, enabling smooth continuity between discrete sign segments. The model architecture consists of three main components: an encoder, a decoder, and a denoiser. The first two are composed of two self-attention layers and the denoiser consists of two blocks, each containing a self-attention layer, a cross-attention layer, and a feed-forward layer. This setup allows the model to effectively capture temporal dependencies and complex contextual information needed for sign language synthesis." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 313, + 496, + 555, + 568 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 496, + 555, + 568 + ], + "spans": [ + { + "bbox": [ + 313, + 496, + 555, + 568 + ], + "type": "text", + "content": "The Encoder first processes the input sequence " + }, + { + "bbox": [ + 313, + 496, + 555, + 568 + ], + "type": "inline_equation", + "content": "X = \\{x_{1}, x_{2}, \\ldots, x_{n}\\}" + }, + { + "bbox": [ + 313, + 496, + 555, + 568 + ], + "type": "text", + "content": ", where " + }, + { + "bbox": [ + 313, + 496, + 555, + 568 + ], + "type": "inline_equation", + "content": "x_{i}" + }, + { + "bbox": [ + 313, + 496, + 555, + 568 + ], + "type": "text", + "content": " represents individual sign language pose frames. By encoding the long-duration video input, the Encoder transforms " + }, + { + "bbox": [ + 313, + 496, + 555, + 568 + ], + "type": "inline_equation", + "content": "X" + }, + { + "bbox": [ + 313, + 496, + 555, + 568 + ], + "type": "text", + "content": " into a latent representation " + }, + { + "bbox": [ + 313, + 496, + 555, + 568 + ], + "type": "inline_equation", + "content": "Z" + }, + { + "bbox": [ + 313, + 496, + 555, + 568 + ], + "type": "text", + "content": ", which serves as the foundation for both the diffusion and denoising processes:" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 386, + 573, + 553, + 586 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 386, + 573, + 553, + 586 + ], + "spans": [ + { + "bbox": [ + 386, + 573, + 553, + 586 + ], + "type": "interline_equation", + "content": "Z = \\operatorname {E n c o d e r} (X). \\tag {3}", + "image_path": "2cfecf43ee1b826ebcd2f09d093d9a5e916ef3db9fb5e14c19088186da6679d0.jpg" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 313, + 591, + 553, + 638 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 591, + 553, + 638 + ], + "spans": [ + { + "bbox": [ + 313, + 591, + 553, + 638 + ], + "type": "text", + "content": "During the forward diffusion, Gaussian noise is gradually added to " + }, + { + "bbox": [ + 313, + 591, + 553, + 638 + ], + "type": "inline_equation", + "content": "Z" + }, + { + "bbox": [ + 313, + 591, + 553, + 638 + ], + "type": "text", + "content": " at each time step " + }, + { + "bbox": [ + 313, + 591, + 553, + 638 + ], + "type": "inline_equation", + "content": "t" + }, + { + "bbox": [ + 313, + 591, + 553, + 638 + ], + "type": "text", + "content": ", resulting in a noisy representation " + }, + { + "bbox": [ + 313, + 591, + 553, + 638 + ], + "type": "inline_equation", + "content": "Z_{t}" + }, + { + "bbox": [ + 313, + 591, + 553, + 638 + ], + "type": "text", + "content": ". This process can be transferred from Eq. 1, formulated as:" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 358, + 642, + 553, + 662 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 358, + 642, + 553, + 662 + ], + "spans": [ + { + "bbox": [ + 358, + 642, + 553, + 662 + ], + "type": "interline_equation", + "content": "Z _ {t} = \\alpha_ {t} \\cdot Z _ {t - 1} + \\sqrt {1 - \\alpha_ {t} ^ {2}} \\cdot \\epsilon_ {t}, \\tag {4}", + "image_path": "a5bdc7d313cb20a86029d231c14dec3c4c3c2a3348d42186c14b5d3823c64698.jpg" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 313, + 666, + 555, + 712 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 666, + 555, + 712 + ], + "spans": [ + { + "bbox": [ + 313, + 666, + 555, + 712 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 313, + 666, + 555, + 712 + ], + "type": "inline_equation", + "content": "\\alpha_{t}" + }, + { + "bbox": [ + 313, + 666, + 555, + 712 + ], + "type": "text", + "content": " controls the noise schedule, and " + }, + { + "bbox": [ + 313, + 666, + 555, + 712 + ], + "type": "inline_equation", + "content": "\\epsilon_{t}" + }, + { + "bbox": [ + 313, + 666, + 555, + 712 + ], + "type": "text", + "content": " is sampled from a Gaussian distribution. The diffusion process progressively transforms the latent representation towards pure noise." + } + ] + } + ], + "index": 13 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "text", + "content": "3484" + } + ] + } + ], + "index": 14 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 3 + }, + { + "para_blocks": [ + { + "bbox": [ + 55, + 72, + 296, + 168 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 72, + 296, + 168 + ], + "spans": [ + { + "bbox": [ + 55, + 72, + 296, + 168 + ], + "type": "text", + "content": "The denoiser then reconstructs the original representation by iteratively removing the noise. Here, the denoiser is conditioned on sign observations " + }, + { + "bbox": [ + 55, + 72, + 296, + 168 + ], + "type": "inline_equation", + "content": "X_{\\mathrm{obs}}" + }, + { + "bbox": [ + 55, + 72, + 296, + 168 + ], + "type": "text", + "content": ", which serve as contextual priors. These observations are incorporated into the denoiser via cross-attention, allowing the model to focus on the relevant parts of " + }, + { + "bbox": [ + 55, + 72, + 296, + 168 + ], + "type": "inline_equation", + "content": "X_{\\mathrm{obs}}" + }, + { + "bbox": [ + 55, + 72, + 296, + 168 + ], + "type": "text", + "content": " at each denoising step. For each time step " + }, + { + "bbox": [ + 55, + 72, + 296, + 168 + ], + "type": "inline_equation", + "content": "t" + }, + { + "bbox": [ + 55, + 72, + 296, + 168 + ], + "type": "text", + "content": " in the reverse process, the denoised output " + }, + { + "bbox": [ + 55, + 72, + 296, + 168 + ], + "type": "inline_equation", + "content": "\\hat{Z}_{t - 1}" + }, + { + "bbox": [ + 55, + 72, + 296, + 168 + ], + "type": "text", + "content": " is given by:" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 107, + 175, + 295, + 191 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 107, + 175, + 295, + 191 + ], + "spans": [ + { + "bbox": [ + 107, + 175, + 295, + 191 + ], + "type": "interline_equation", + "content": "\\hat {Z} _ {t - 1} = \\operatorname {D e n o i s e r} \\left(Z _ {t}, X _ {\\mathrm {o b s}}\\right), \\tag {5}", + "image_path": "e81397e5cc9056919627218e1979f44fd05776e4a1e2528922d5e25660244ab3.jpg" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 55, + 198, + 295, + 222 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 198, + 295, + 222 + ], + "spans": [ + { + "bbox": [ + 55, + 198, + 295, + 222 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 55, + 198, + 295, + 222 + ], + "type": "inline_equation", + "content": "X_{\\mathrm{obs}}" + }, + { + "bbox": [ + 55, + 198, + 295, + 222 + ], + "type": "text", + "content": " provides the model with the necessary context to predict the masked segments accurately." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 55, + 222, + 296, + 270 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 222, + 296, + 270 + ], + "spans": [ + { + "bbox": [ + 55, + 222, + 296, + 270 + ], + "type": "text", + "content": "Finally, the decoder, also structured with three self-attention layers, takes the denoised latent representation and transforms it back into the original sign language video space, outputting the recovered or synthesized video " + }, + { + "bbox": [ + 55, + 222, + 296, + 270 + ], + "type": "inline_equation", + "content": "\\hat{X}" + }, + { + "bbox": [ + 55, + 222, + 296, + 270 + ], + "type": "text", + "content": ":" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 125, + 277, + 295, + 292 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 125, + 277, + 295, + 292 + ], + "spans": [ + { + "bbox": [ + 125, + 277, + 295, + 292 + ], + "type": "interline_equation", + "content": "\\hat {X} = \\operatorname {D e c o d e r} \\left(\\hat {Z} _ {0}\\right). \\tag {6}", + "image_path": "bd86444ee49cce0144873a02364ba3bf0255f0fb6912d63fb45fb9e9aec5dfae.jpg" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 55, + 300, + 296, + 361 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 300, + 296, + 361 + ], + "spans": [ + { + "bbox": [ + 55, + 300, + 296, + 361 + ], + "type": "text", + "content": "By using self-attention within both the encoder and decoder, and cross-attention in the denoiser to incorporate contextual observations, the model effectively learns to generate temporally coherent transitions, resulting in a smooth, continuous sign language video output." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 55, + 360, + 296, + 408 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 360, + 296, + 408 + ], + "spans": [ + { + "bbox": [ + 55, + 360, + 296, + 408 + ], + "type": "text", + "content": "The proposed Sign-D2C incorporates two key strategies to enhance transition generation: a random masking strategy during training and linearly interpolating padding initialization for inference. The following provides details." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 55, + 408, + 296, + 456 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 408, + 296, + 456 + ], + "spans": [ + { + "bbox": [ + 55, + 408, + 296, + 456 + ], + "type": "text", + "content": "Random Masking Strategy. To generate smooth transitional poses and capture relationships between different segments, we employ a random masking strategy during training, as illustrated in the training phase of Fig. 2." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 55, + 456, + 296, + 564 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 456, + 296, + 564 + ], + "spans": [ + { + "bbox": [ + 55, + 456, + 296, + 564 + ], + "type": "text", + "content": "Given a long-duration video " + }, + { + "bbox": [ + 55, + 456, + 296, + 564 + ], + "type": "inline_equation", + "content": "p \\in \\mathbb{R}^{F \\times (N \\times 3)}" + }, + { + "bbox": [ + 55, + 456, + 296, + 564 + ], + "type": "text", + "content": " with " + }, + { + "bbox": [ + 55, + 456, + 296, + 564 + ], + "type": "inline_equation", + "content": "F" + }, + { + "bbox": [ + 55, + 456, + 296, + 564 + ], + "type": "text", + "content": " frames, each containing " + }, + { + "bbox": [ + 55, + 456, + 296, + 564 + ], + "type": "inline_equation", + "content": "N" + }, + { + "bbox": [ + 55, + 456, + 296, + 564 + ], + "type": "text", + "content": " joints in 3D space, the video is first encoded by a pre-trained encoder into a latent representation " + }, + { + "bbox": [ + 55, + 456, + 296, + 564 + ], + "type": "inline_equation", + "content": "x \\in \\mathbb{R}^{F \\times W}" + }, + { + "bbox": [ + 55, + 456, + 296, + 564 + ], + "type": "text", + "content": ", where " + }, + { + "bbox": [ + 55, + 456, + 296, + 564 + ], + "type": "inline_equation", + "content": "W" + }, + { + "bbox": [ + 55, + 456, + 296, + 564 + ], + "type": "text", + "content": " denotes the dimension of the latent feature " + }, + { + "bbox": [ + 55, + 456, + 296, + 564 + ], + "type": "inline_equation", + "content": "x" + }, + { + "bbox": [ + 55, + 456, + 296, + 564 + ], + "type": "text", + "content": ". The mask " + }, + { + "bbox": [ + 55, + 456, + 296, + 564 + ], + "type": "inline_equation", + "content": "m" + }, + { + "bbox": [ + 55, + 456, + 296, + 564 + ], + "type": "text", + "content": " is constructed by randomly selecting frame indices, extending across the dimensions of the latent feature " + }, + { + "bbox": [ + 55, + 456, + 296, + 564 + ], + "type": "inline_equation", + "content": "x" + }, + { + "bbox": [ + 55, + 456, + 296, + 564 + ], + "type": "text", + "content": ". This mask " + }, + { + "bbox": [ + 55, + 456, + 296, + 564 + ], + "type": "inline_equation", + "content": "m" + }, + { + "bbox": [ + 55, + 456, + 296, + 564 + ], + "type": "text", + "content": " is then applied to selectively retain or mask out parts of " + }, + { + "bbox": [ + 55, + 456, + 296, + 564 + ], + "type": "inline_equation", + "content": "x" + }, + { + "bbox": [ + 55, + 456, + 296, + 564 + ], + "type": "text", + "content": ", resulting in a masked latent representation:" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 110, + 571, + 295, + 586 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 110, + 571, + 295, + 586 + ], + "spans": [ + { + "bbox": [ + 110, + 571, + 295, + 586 + ], + "type": "interline_equation", + "content": "x ^ {m} = x \\odot m, x \\in \\mathbb {R} ^ {F \\times W}, \\tag {7}", + "image_path": "15ad7c6c532afd2bc8520b020b7a8a82fbcccfecfb98ce285c4de26026a83ee7.jpg" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 55, + 594, + 295, + 653 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 594, + 295, + 653 + ], + "spans": [ + { + "bbox": [ + 55, + 594, + 295, + 653 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 55, + 594, + 295, + 653 + ], + "type": "inline_equation", + "content": "x^{m}" + }, + { + "bbox": [ + 55, + 594, + 295, + 653 + ], + "type": "text", + "content": " serves as a conditional input to the model, distinguishing between occluded and visible movements. This masking strategy enables the model to predict missing frames based on observable context, facilitating smooth transition generation." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 55, + 654, + 296, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 654, + 296, + 714 + ], + "spans": [ + { + "bbox": [ + 55, + 654, + 296, + 714 + ], + "type": "text", + "content": "Linear Interpolation Padding. In the inference phase, generating smooth transitions between discrete motion segments is challenging due to the gaps between observed frames. We leverage the natural correlation between consecutive frames in sign language sequences by introducing" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 313, + 72, + 553, + 96 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 72, + 553, + 96 + ], + "spans": [ + { + "bbox": [ + 313, + 72, + 553, + 96 + ], + "type": "text", + "content": "a linear interpolation padding initialization strategy, as illustrated in Fig. 2." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 313, + 96, + 554, + 179 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 96, + 554, + 179 + ], + "spans": [ + { + "bbox": [ + 313, + 96, + 554, + 179 + ], + "type": "text", + "content": "This strategy starts by using the last pose of the preceding observed segment, " + }, + { + "bbox": [ + 313, + 96, + 554, + 179 + ], + "type": "inline_equation", + "content": "p_{s}" + }, + { + "bbox": [ + 313, + 96, + 554, + 179 + ], + "type": "text", + "content": ", and the first pose of the following observed segment, " + }, + { + "bbox": [ + 313, + 96, + 554, + 179 + ], + "type": "inline_equation", + "content": "p_{e}" + }, + { + "bbox": [ + 313, + 96, + 554, + 179 + ], + "type": "text", + "content": ", as boundary frames for the missing transition segment. The goal is to generate intermediate frames " + }, + { + "bbox": [ + 313, + 96, + 554, + 179 + ], + "type": "inline_equation", + "content": "p_{c}" + }, + { + "bbox": [ + 313, + 96, + 554, + 179 + ], + "type": "text", + "content": " that provide a smooth transition from " + }, + { + "bbox": [ + 313, + 96, + 554, + 179 + ], + "type": "inline_equation", + "content": "p_{s}" + }, + { + "bbox": [ + 313, + 96, + 554, + 179 + ], + "type": "text", + "content": " to " + }, + { + "bbox": [ + 313, + 96, + 554, + 179 + ], + "type": "inline_equation", + "content": "p_{e}" + }, + { + "bbox": [ + 313, + 96, + 554, + 179 + ], + "type": "text", + "content": ". Using linear interpolation, we initialize these missing frames as follows:" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 327, + 182, + 553, + 215 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 327, + 182, + 553, + 215 + ], + "spans": [ + { + "bbox": [ + 327, + 182, + 553, + 215 + ], + "type": "interline_equation", + "content": "\\left\\{ \\begin{array}{l} f (i) = \\sum_ {k = 0} ^ {i} \\frac {1}{(k + 1)}, \\quad i = 0, \\dots , (e - s), \\\\ p _ {c} (i) = p _ {s} + \\frac {p _ {e} - p _ {s}}{e - s + 1} \\cdot f (i). \\end{array} \\right. \\tag {8}", + "image_path": "7962f6c5e1fffd71fe57510b47516a99d93ac61c1910716a2f5f383335163b31.jpg" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 313, + 219, + 555, + 291 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 219, + 555, + 291 + ], + "spans": [ + { + "bbox": [ + 313, + 219, + 555, + 291 + ], + "type": "text", + "content": "This interpolation scheme gradually fills in the key points at each time step, creating intermediate poses that ensure a coherent and natural flow between segments. This strategy facilitates smoother transitions and provides a stable foundation for later refinement, resulting in seamless sign language video synthesis." + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 314, + 297, + 378, + 309 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 314, + 297, + 378, + 309 + ], + "spans": [ + { + "bbox": [ + 314, + 297, + 378, + 309 + ], + "type": "text", + "content": "3.4. Training" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 313, + 314, + 554, + 422 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 314, + 554, + 422 + ], + "spans": [ + { + "bbox": [ + 313, + 314, + 554, + 422 + ], + "type": "text", + "content": "Pre-training of Encoder & decoder. To enable coherent video generation, we pre-train the encoder and decoder to ensure effective latent representation processing. The encoder extracts latent features " + }, + { + "bbox": [ + 313, + 314, + 554, + 422 + ], + "type": "inline_equation", + "content": "x" + }, + { + "bbox": [ + 313, + 314, + 554, + 422 + ], + "type": "text", + "content": " from input videos, while the decoder reconstructs realistic sign language poses from these latent representations. For training the diffusion model, we freeze the encoder and decoder parameters, allowing the model to focus on learning smooth transitions based on these pre-trained features." + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 313, + 422, + 554, + 458 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 422, + 554, + 458 + ], + "spans": [ + { + "bbox": [ + 313, + 422, + 554, + 458 + ], + "type": "text", + "content": "Pre-training is conducted end-to-end with a reconstruction loss, optimizing the encoder-decoder pair to accurately represent and recover video details:" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 373, + 463, + 553, + 497 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 373, + 463, + 553, + 497 + ], + "spans": [ + { + "bbox": [ + 373, + 463, + 553, + 497 + ], + "type": "interline_equation", + "content": "\\mathcal {L} _ {\\text {p r e}} = \\frac {1}{F} \\sum_ {f = 1} ^ {F} | p _ {f} - \\hat {p} _ {f} |, \\tag {9}", + "image_path": "951a3bc1194199940551a1c4be8e865858ac30cac16b2b074500fe2001829fc0.jpg" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 313, + 502, + 554, + 561 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 502, + 554, + 561 + ], + "spans": [ + { + "bbox": [ + 313, + 502, + 554, + 561 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 313, + 502, + 554, + 561 + ], + "type": "inline_equation", + "content": "p" + }, + { + "bbox": [ + 313, + 502, + 554, + 561 + ], + "type": "text", + "content": " represents the input video with " + }, + { + "bbox": [ + 313, + 502, + 554, + 561 + ], + "type": "inline_equation", + "content": "F" + }, + { + "bbox": [ + 313, + 502, + 554, + 561 + ], + "type": "text", + "content": " frames, and " + }, + { + "bbox": [ + 313, + 502, + 554, + 561 + ], + "type": "inline_equation", + "content": "\\hat{p}" + }, + { + "bbox": [ + 313, + 502, + 554, + 561 + ], + "type": "text", + "content": " is the reconstructed output. This process ensures the encoder captures essential features, while the decoder faithfully reconstructs realistic sign language poses, setting a solid foundation for high-quality transition generation." + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 313, + 562, + 554, + 610 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 562, + 554, + 610 + ], + "spans": [ + { + "bbox": [ + 313, + 562, + 554, + 610 + ], + "type": "text", + "content": "Loss Optimization. According to Eq. 5, we can obtain the denoiser prediction " + }, + { + "bbox": [ + 313, + 562, + 554, + 610 + ], + "type": "inline_equation", + "content": "\\hat{X} = \\{\\hat{x}_1,\\hat{x}_2,\\dots ,\\hat{x}_F\\}" + }, + { + "bbox": [ + 313, + 562, + 554, + 610 + ], + "type": "text", + "content": ". During training, the loss function constrains the difference between the original latent feature " + }, + { + "bbox": [ + 313, + 562, + 554, + 610 + ], + "type": "inline_equation", + "content": "x" + }, + { + "bbox": [ + 313, + 562, + 554, + 610 + ], + "type": "text", + "content": " and the denoiser's prediction " + }, + { + "bbox": [ + 313, + 562, + 554, + 610 + ], + "type": "inline_equation", + "content": "\\hat{x}" + }, + { + "bbox": [ + 313, + 562, + 554, + 610 + ], + "type": "text", + "content": ":" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 372, + 615, + 553, + 649 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 372, + 615, + 553, + 649 + ], + "spans": [ + { + "bbox": [ + 372, + 615, + 553, + 649 + ], + "type": "interline_equation", + "content": "\\mathcal {L} _ {\\text {d i f f}} = \\frac {1}{F} \\sum_ {f = 1} ^ {F} | x _ {f} - \\hat {x} _ {f} |, \\tag {10}", + "image_path": "7fb90a024b355e77f672eca9165770ed1a1d9ae2a30dcd284c92c2ee9f70dec4.jpg" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 313, + 653, + 554, + 712 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 653, + 554, + 712 + ], + "spans": [ + { + "bbox": [ + 313, + 653, + 554, + 712 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 313, + 653, + 554, + 712 + ], + "type": "inline_equation", + "content": "F" + }, + { + "bbox": [ + 313, + 653, + 554, + 712 + ], + "type": "text", + "content": " denotes the number of frames in the video sequence. This loss encourages the model to accurately recover the original latent representation, improving the quality of the generated transitions and ensuring alignment with the encoded video features." + } + ] + } + ], + "index": 23 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "text", + "content": "3485" + } + ] + } + ], + "index": 24 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 4 + }, + { + "para_blocks": [ + { + "bbox": [ + 55, + 71, + 137, + 85 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 71, + 137, + 85 + ], + "spans": [ + { + "bbox": [ + 55, + 71, + 137, + 85 + ], + "type": "text", + "content": "4. Experiments" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 55, + 91, + 182, + 105 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 91, + 182, + 105 + ], + "spans": [ + { + "bbox": [ + 55, + 91, + 182, + 105 + ], + "type": "text", + "content": "4.1. Experimental Settings" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 55, + 110, + 295, + 207 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 110, + 295, + 207 + ], + "spans": [ + { + "bbox": [ + 55, + 110, + 295, + 207 + ], + "type": "text", + "content": "Datasets. We conduct the experiments on three datasets, i.e., PHOENIX14T, USTC-CSL100, and USTC-SLR500. The first two are sentence-level continuous sign language datasets, which are used to evaluate supervised transition generation in controlled settings. While the latter one is a word-level isolated sign language dataset, allowing us to assess the model's performance in generating transitions within real-world, discrete sign scenarios." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 55, + 208, + 295, + 412 + ], + "type": "list", + "angle": 0, + "index": 6, + "blocks": [ + { + "bbox": [ + 55, + 208, + 295, + 280 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 208, + 295, + 280 + ], + "spans": [ + { + "bbox": [ + 55, + 208, + 295, + 280 + ], + "type": "text", + "content": "- PHOENIX14T [2] comprises 8,257 sentence-level samples, which covers 2,887 German words. For preprocessing, we follow the settings in previous sign language production studies [19, 36], using OpenPose [4] to extract 2D coordinates from the videos, which are then refined to 3D coordinates through a skeleton correction model [51]." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 56, + 281, + 295, + 351 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 281, + 295, + 351 + ], + "spans": [ + { + "bbox": [ + 56, + 281, + 295, + 351 + ], + "type": "text", + "content": "- USTC-CSL100 [18] covers 100 unique sentences performed by 50 signers. We use the split-I [11] approach to divide the dataset into 4,000 training examples and 1,000 test examples. The above two datasets enable us to evaluate the model's effectiveness on continuous sign language sequences in a structured environment." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 56, + 352, + 295, + 412 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 352, + 295, + 412 + ], + "spans": [ + { + "bbox": [ + 56, + 352, + 295, + 412 + ], + "type": "text", + "content": "- USTC-SLR500 [17] is an isolated sign language dataset containing 500 Chinese sign words, each performed five times by 50 signers. Following [17], we split the dataset by using videos from 36 signers for training and the remaining 14 signers for testing." + } + ] + } + ], + "index": 5 + } + ], + "sub_type": "text" + }, + { + "bbox": [ + 54, + 413, + 295, + 593 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 54, + 413, + 295, + 593 + ], + "spans": [ + { + "bbox": [ + 54, + 413, + 295, + 593 + ], + "type": "text", + "content": "Implementation Details. Since generating continuous sign language videos from discrete fragments is inherently an unsupervised task, there are no standardized evaluation metrics. To effectively assess our model's performance, we analyzed transition actions between two gestures on the sentence-level datasets (i.e., PHOENIX14T and USTC-CSL100), observing that transitions typically span 10 to 20 frames. Based on these findings, we designed experiments using two masking strategies: masking 10 frames every 20 frames and masking 20 frames every 10 frames. Additionally, to further validate the effectiveness of our approach, we used the isolated sign language dataset (i.e., USTC-SLR500) to assemble discrete vocabulary videos into continuous sequences, demonstrating the model's capability to produce smooth transitions in a real-world setting." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 55, + 594, + 295, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 594, + 295, + 714 + ], + "spans": [ + { + "bbox": [ + 55, + 594, + 295, + 714 + ], + "type": "text", + "content": "Evaluation Metrics. We evaluate our performance on two sentence-level datasets in terms of both semantics and coherence. Following prior work in sign language generation [19, 35, 36, 44], we use the NSLT method [2] to convert generated sign language videos back into text. We then compare this text with reference data to calculate semantic accuracy metrics, including BLEU [33], ROUGE [27], and WER. Additionally, we use FID, DTW, and MPJPE to assess the consistency and smoothness of the generated transitions, measuring the coherence of the results." + } + ] + } + ], + "index": 8 + }, + { + "type": "table", + "bbox": [ + 317, + 93, + 553, + 153 + ], + "blocks": [ + { + "bbox": [ + 314, + 71, + 553, + 92 + ], + "lines": [ + { + "bbox": [ + 314, + 71, + 553, + 92 + ], + "spans": [ + { + "bbox": [ + 314, + 71, + 553, + 92 + ], + "type": "text", + "content": "Table 1. Ablation results of different masking ratios on PHOENIX14T (observation 20 frames - prediction 10 frames)." + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 317, + 93, + 553, + 153 + ], + "lines": [ + { + "bbox": [ + 317, + 93, + 553, + 153 + ], + "spans": [ + { + "bbox": [ + 317, + 93, + 553, + 153 + ], + "type": "table", + "html": "
Masking Ratio rDEVTEST
B1↑B4↑DTW↓B1↑B4↑DTW↓
r = 0.122.178.040.6122.478.670.60
r = 0.322.468.060.6222.368.730.61
r = 0.522.308.060.7722.028.350.75
", + "image_path": "5472886423c56c48a382123e04956c58b2a519e9451a6d2bd1bbde7f8cd15c76.jpg" + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "table_body" + } + ], + "index": 10 + }, + { + "type": "table", + "bbox": [ + 317, + 183, + 553, + 247 + ], + "blocks": [ + { + "bbox": [ + 314, + 160, + 553, + 182 + ], + "lines": [ + { + "bbox": [ + 314, + 160, + 553, + 182 + ], + "spans": [ + { + "bbox": [ + 314, + 160, + 553, + 182 + ], + "type": "text", + "content": "Table 2. Ablation results of different padding strategies on PHOENIX14T (observation 20 frames - prediction 10 frames)." + } + ] + } + ], + "index": 11, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 317, + 183, + 553, + 247 + ], + "lines": [ + { + "bbox": [ + 317, + 183, + 553, + 247 + ], + "spans": [ + { + "bbox": [ + 317, + 183, + 553, + 247 + ], + "type": "table", + "html": "
Padding StrategyDEVTEST
DTW↓FID↓MPJPE↓DTW↓FID↓MPJPE↓
w/o Padding11.462.7469.6011.603.0687.57
Front Padding0.722.6847.900.692.9960.05
Linear Padding0.622.6847.300.612.9859.36
Back Padding0.672.6847.890.662.9959.91
", + "image_path": "8c731ebddf3fe11838b94b9a4645b5deb2da69b7aaeff6e918b4486abc66c580.jpg" + } + ] + } + ], + "index": 12, + "angle": 0, + "type": "table_body" + } + ], + "index": 12 + }, + { + "bbox": [ + 313, + 262, + 408, + 274 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 262, + 408, + 274 + ], + "spans": [ + { + "bbox": [ + 313, + 262, + 408, + 274 + ], + "type": "text", + "content": "4.2. Ablation Study" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 313, + 280, + 555, + 363 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 280, + 555, + 363 + ], + "spans": [ + { + "bbox": [ + 313, + 280, + 555, + 363 + ], + "type": "text", + "content": "In this subsection, we explore the effects of different random masking ratios in the training phase and various padding strategies during inference. The experiments in this part are all conducted under the condition of predicting 10 transition frames for every 20 observed sign frames to evaluate their impact on both semantic accuracy and transition coherence." + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 313, + 365, + 556, + 533 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 365, + 556, + 533 + ], + "spans": [ + { + "bbox": [ + 313, + 365, + 556, + 533 + ], + "type": "text", + "content": "Effect of Masking Ratio. Table 1 presents the results of different random masking ratios " + }, + { + "bbox": [ + 313, + 365, + 556, + 533 + ], + "type": "inline_equation", + "content": "r" + }, + { + "bbox": [ + 313, + 365, + 556, + 533 + ], + "type": "text", + "content": " on semantic preservation and temporal coherence. With a lower masking ratio " + }, + { + "bbox": [ + 313, + 365, + 556, + 533 + ], + "type": "inline_equation", + "content": "(r = 0.1)" + }, + { + "bbox": [ + 313, + 365, + 556, + 533 + ], + "type": "text", + "content": ", the model achieves the best coherence, reflected by the lowest DTW scores of 0.61 on DEV and 0.60 on TEST. This indicates that smaller masking ratios allow the model to maintain smoother transitions, as more contextual frames are available to guide the generation of missing frames. However, as the masking ratio increases to " + }, + { + "bbox": [ + 313, + 365, + 556, + 533 + ], + "type": "inline_equation", + "content": "r = 0.3" + }, + { + "bbox": [ + 313, + 365, + 556, + 533 + ], + "type": "text", + "content": ", we observe a slight improvement in semantic accuracy, with BLEU-4 scores rising to 8.06 on DEV and 8.73 on TEST. This suggests that a higher masking ratio may encourage the model to focus more on the sign semantic content, as it learns to reconstruct more extensive missing segments." + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 313, + 533, + 556, + 640 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 533, + 556, + 640 + ], + "spans": [ + { + "bbox": [ + 313, + 533, + 556, + 640 + ], + "type": "text", + "content": "Interestingly, further increasing the masking ratio to " + }, + { + "bbox": [ + 313, + 533, + 556, + 640 + ], + "type": "inline_equation", + "content": "r = 0.5" + }, + { + "bbox": [ + 313, + 533, + 556, + 640 + ], + "type": "text", + "content": " leads to a drop in both semantic accuracy and coherence, with DTW increasing and BLEU decreasing slightly. This highlights a trade-off: while a moderate masking ratio improves semantic preservation, excessive masking may reduce the model's ability to maintain smooth transitions and semantic accuracy. Based on these observations, we select " + }, + { + "bbox": [ + 313, + 533, + 556, + 640 + ], + "type": "inline_equation", + "content": "r = 0.3" + }, + { + "bbox": [ + 313, + 533, + 556, + 640 + ], + "type": "text", + "content": " for subsequent experiments, as it provides a good balance between semantic retention and transition smoothness." + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 313, + 641, + 556, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 641, + 556, + 714 + ], + "spans": [ + { + "bbox": [ + 313, + 641, + 556, + 714 + ], + "type": "text", + "content": "Influence of Padding Strategy. Table 2 summarizes the results of different padding strategies applied during inference. We experiment with four approaches: w/o Padding (no initialization for missing frames), Front Padding (using the last frame of the previous observation segment), Back Padding (using the first frame of the following observation" + } + ] + } + ], + "index": 17 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "text", + "content": "3486" + } + ] + } + ], + "index": 18 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 5 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 58, + 83, + 553, + 160 + ], + "blocks": [ + { + "bbox": [ + 97, + 71, + 512, + 82 + ], + "lines": [ + { + "bbox": [ + 97, + 71, + 512, + 82 + ], + "spans": [ + { + "bbox": [ + 97, + 71, + 512, + 82 + ], + "type": "text", + "content": "Table 3. Performance comparison on PHOENIX14T (under setting: observation 20 frames - prediction 10 frames)." + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 58, + 83, + 553, + 160 + ], + "lines": [ + { + "bbox": [ + 58, + 83, + 553, + 160 + ], + "spans": [ + { + "bbox": [ + 58, + 83, + 553, + 160 + ], + "type": "table", + "html": "
MethodsDEVTEST
B1↑B4↑ROUGE↑WER↓DTW↓FID↓MPIJE↓B1↑B4↑ROUGE↑WER↓DTW↓FID↓MPIJE↓
Ground Truth29.7712.1329.6074.170.000.000.0029.7611.9328.9871.940.000.000.00
G2P-DDM [47]8.250.997.1798.9312.712.8555.029.521.298.1298.5412.573.1967.81
VQ-GCDM [43]9.310.877.5998.6912.112.8653.879.850.967.9598.5911.983.2166.51
Ours22.468.0622.0780.730.622.6847.3022.368.7322.3679.060.612.9859.36
", + "image_path": "e2ca61446809ef5dfc7c7b0c89a596c49bdb264a185d9e7940d97d04c01e6d78.jpg" + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "table_body" + } + ], + "index": 1 + }, + { + "type": "table", + "bbox": [ + 58, + 176, + 553, + 255 + ], + "blocks": [ + { + "bbox": [ + 97, + 164, + 512, + 175 + ], + "lines": [ + { + "bbox": [ + 97, + 164, + 512, + 175 + ], + "spans": [ + { + "bbox": [ + 97, + 164, + 512, + 175 + ], + "type": "text", + "content": "Table 4. Performance comparison on PHOENIX14T (under setting: observation 10 frames - prediction 20 frames)." + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 58, + 176, + 553, + 255 + ], + "lines": [ + { + "bbox": [ + 58, + 176, + 553, + 255 + ], + "spans": [ + { + "bbox": [ + 58, + 176, + 553, + 255 + ], + "type": "table", + "html": "
MethodsDEVTEST
B1↑B4↑ROUGE↑WER↓DTW↓FID↓MPJPE↓B1↑B4↑ROUGE↑WER↓DTW↓FID↓MPJPE↓
Ground Truth29.7712.1329.6074.170.000.000.0029.7611.9328.9871.940.000.000.00
G2P-DDM [47]8.980.637.4498.6115.952.8955.219.130.747.7598.6415.583.2268.18
VQ-GCDM [43]8.640.897.0898.4812.403.0548.669.021.367.2298.6112.223.3760.27
Ours19.325.9218.3387.300.702.7548.0419.596.2618.5986.80.683.0559.81
", + "image_path": "954bc67869cdb04a76a2cb068cb8ad5d0ced6ca672a04c7e6a221232b36c5387.jpg" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "table_body" + } + ], + "index": 3 + }, + { + "type": "table", + "bbox": [ + 58, + 269, + 553, + 347 + ], + "blocks": [ + { + "bbox": [ + 207, + 258, + 402, + 267 + ], + "lines": [ + { + "bbox": [ + 207, + 258, + 402, + 267 + ], + "spans": [ + { + "bbox": [ + 207, + 258, + 402, + 267 + ], + "type": "text", + "content": "Table 5. Performance comparison on USTC-CSL100." + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 58, + 269, + 553, + 347 + ], + "lines": [ + { + "bbox": [ + 58, + 269, + 553, + 347 + ], + "spans": [ + { + "bbox": [ + 58, + 269, + 553, + 347 + ], + "type": "table", + "html": "
MethodsObservation 20 Frames - Prediction 10 FramesObservation 10 Frames - Prediction 20 Frames
B1↑B4↑ROUGE↑WER↓DTW↓FID↓MPJPE↓B1↑B4↑ROUGE↑WER↓DTW↓FID↓MPJPE↓
Ground Truth69.1059.1268.5347.380.000.000.0069.1059.1268.5347.380.000.000.00
G2P-DDM [47]13.742.1113.8592.047.000.53258.4313.431.7413.6292.157.470.53258.43
VQ-GCDM [43]30.7110.6829.1777.875.100.4391.1018.772.0816.6492.6710.890.54104.27
Ours69.6458.9869.2646.661.920.3177.0467.9556.6167.3849.241.790.3178.62
", + "image_path": "ffd62c6eb05b8f9275a1998af7519eeef62ae08e5107721f3f74a043252e4a17.jpg" + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "table_body" + } + ], + "index": 5 + }, + { + "bbox": [ + 54, + 359, + 294, + 455 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 54, + 359, + 294, + 455 + ], + "spans": [ + { + "bbox": [ + 54, + 359, + 294, + 455 + ], + "type": "text", + "content": "segment), and Linear Padding (performing linear interpolation between the two boundary poses, as Eq. 8). The \"w/o Padding\" approach yields the highest DTW scores (11.46 - DEV and 10.36 - TEST), indicating poor coherence and unnatural transitions. Both Front Padding and Back Padding improve the smoothness of transitions, as shown by lower DTW scores (0.67 - DEV and 0.66 - TEST), but their FID and MPJPE remain better compared to Linear Padding." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 54, + 456, + 294, + 552 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 54, + 456, + 294, + 552 + ], + "spans": [ + { + "bbox": [ + 54, + 456, + 294, + 552 + ], + "type": "text", + "content": "The Linear Padding strategy achieves the best overall performance, with the lowest DTW scores (0.62 on DEV and 0.66 on TEST) and MPJPE values, reflecting both smoothness and accuracy in pose reconstruction. By initializing missing frames through linear interpolation, Linear Padding provides a stable and effective foundation for generating realistic and seamless transitions, closely resembling real-world signing behavior." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 55, + 562, + 241, + 574 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 562, + 241, + 574 + ], + "spans": [ + { + "bbox": [ + 55, + 562, + 241, + 574 + ], + "type": "text", + "content": "4.3. Comparison with State-of-the-Arts" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 55, + 580, + 295, + 640 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 580, + 295, + 640 + ], + "spans": [ + { + "bbox": [ + 55, + 580, + 295, + 640 + ], + "type": "text", + "content": "We compare our method with two state-of-the-art models: G2P-DDM [47] and VQ-GCDM [43]. Both G2P-DDM and VQ-GCDM are primarily designed for SLP. Due to the absence of established benchmarks for our specific task, we adapted these models to serve as comparative baselines." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 54, + 641, + 295, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 54, + 641, + 295, + 714 + ], + "spans": [ + { + "bbox": [ + 54, + 641, + 295, + 714 + ], + "type": "text", + "content": "Comparison on PHOENIX14T. As shown in Table 4, under the \"Observation 20 Frames - Prediction 10 Frames\" setting, our method achieves superior performance across both semantic and coherence metrics. Specifically, we attain a BLEU-1 score of " + }, + { + "bbox": [ + 54, + 641, + 295, + 714 + ], + "type": "inline_equation", + "content": "22.46\\%" + }, + { + "bbox": [ + 54, + 641, + 295, + 714 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 54, + 641, + 295, + 714 + ], + "type": "inline_equation", + "content": "22.36\\%" + }, + { + "bbox": [ + 54, + 641, + 295, + 714 + ], + "type": "text", + "content": " on the DEV and TEST sets, respectively, outperforming G2P-DDM and" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 313, + 359, + 553, + 443 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 359, + 553, + 443 + ], + "spans": [ + { + "bbox": [ + 313, + 359, + 553, + 443 + ], + "type": "text", + "content": "VQ-GCDM. This improvement in BLEU scores indicates that our method more accurately preserves semantic content compared to existing methods. Furthermore, our method shows significant gains in temporal coherence, with DTW scores of 0.62 on DEV and 0.61 on TEST, respectively. These results confirm that our model generates smoother and more realistic transitions than the baseline methods." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 313, + 447, + 555, + 589 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 447, + 555, + 589 + ], + "spans": [ + { + "bbox": [ + 313, + 447, + 555, + 589 + ], + "type": "text", + "content": "Under the more challenging \"Observation 10 Frames - Prediction 20 Frames\" setting, our method continues to demonstrate robust performance despite the increased prediction length. We achieve BLEU-1 scores of " + }, + { + "bbox": [ + 313, + 447, + 555, + 589 + ], + "type": "inline_equation", + "content": "19.32\\%" + }, + { + "bbox": [ + 313, + 447, + 555, + 589 + ], + "type": "text", + "content": " on DEV and " + }, + { + "bbox": [ + 313, + 447, + 555, + 589 + ], + "type": "inline_equation", + "content": "19.59\\%" + }, + { + "bbox": [ + 313, + 447, + 555, + 589 + ], + "type": "text", + "content": " on TEST, considerably higher than the baseline methods. Our method also maintains lower DTW and FID scores, demonstrating its ability to generate coherent and semantically accurate sequences even with limited observed data. The MPJPE scores further support the effectiveness of our approach, with scores of 48.04 on DEV and 59.81 on TEST, indicating a closer alignment to ground truth poses than the other methods." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 313, + 594, + 556, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 594, + 556, + 713 + ], + "spans": [ + { + "bbox": [ + 313, + 594, + 556, + 713 + ], + "type": "text", + "content": "Comparison on USTC-CSL100. Table 5 shows the results on the USTC-CSL100 dataset, which evaluates our method's ability to generate sign language sequences. In the \"Observation 20 Frames - Prediction 10 Frames\" setting, our model achieves near-ground-truth semantic accuracy, with BLEU-1 and BLEU-4 scores of " + }, + { + "bbox": [ + 313, + 594, + 556, + 713 + ], + "type": "inline_equation", + "content": "69.64\\%" + }, + { + "bbox": [ + 313, + 594, + 556, + 713 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 313, + 594, + 556, + 713 + ], + "type": "inline_equation", + "content": "58.98\\%" + }, + { + "bbox": [ + 313, + 594, + 556, + 713 + ], + "type": "text", + "content": " on DEV, closely approaching the ground truth scores of " + }, + { + "bbox": [ + 313, + 594, + 556, + 713 + ], + "type": "inline_equation", + "content": "69.10\\%" + }, + { + "bbox": [ + 313, + 594, + 556, + 713 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 313, + 594, + 556, + 713 + ], + "type": "inline_equation", + "content": "59.12\\%" + }, + { + "bbox": [ + 313, + 594, + 556, + 713 + ], + "type": "text", + "content": ". Additionally, our DTW score of 1.922 and FID score of 3.21 indicate high temporal consistency and low distributional divergence from the real data. Com" + } + ] + } + ], + "index": 13 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "text", + "content": "3487" + } + ] + } + ], + "index": 14 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 6 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 58, + 71, + 553, + 199 + ], + "blocks": [ + { + "bbox": [ + 58, + 71, + 553, + 199 + ], + "lines": [ + { + "bbox": [ + 58, + 71, + 553, + 199 + ], + "spans": [ + { + "bbox": [ + 58, + 71, + 553, + 199 + ], + "type": "image", + "image_path": "cad7faac2655967240f13d92e536dbe0310e82462e52c30edee06bdaf17c4681.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 55, + 201, + 555, + 224 + ], + "lines": [ + { + "bbox": [ + 55, + 201, + 555, + 224 + ], + "spans": [ + { + "bbox": [ + 55, + 201, + 555, + 224 + ], + "type": "text", + "content": "Figure 3. Visualization examples of generating 10-frame transition pose under 20-frame observations on PHOENIX14T. We compare our method with G2P-DDM and VQ-GCDM, and the Ground Truth is attached. Green: Transitions. Orange and Blue: Observations." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "type": "image", + "bbox": [ + 58, + 236, + 552, + 361 + ], + "blocks": [ + { + "bbox": [ + 58, + 236, + 552, + 361 + ], + "lines": [ + { + "bbox": [ + 58, + 236, + 552, + 361 + ], + "spans": [ + { + "bbox": [ + 58, + 236, + 552, + 361 + ], + "type": "image", + "image_path": "b904cb3598f50b51614234e478fd5f038d405c24b8a61c541dbc538df5cf9062.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 55, + 364, + 555, + 387 + ], + "lines": [ + { + "bbox": [ + 55, + 364, + 555, + 387 + ], + "spans": [ + { + "bbox": [ + 55, + 364, + 555, + 387 + ], + "type": "text", + "content": "Figure 4. Visualization examples on the USTC-SLR500, demonstrate the generation of transition poses according to two discrete sign segments. Green: Transitions. Orange and Blue: Observations." + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_caption" + } + ], + "index": 2 + }, + { + "bbox": [ + 55, + 407, + 295, + 443 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 407, + 295, + 443 + ], + "spans": [ + { + "bbox": [ + 55, + 407, + 295, + 443 + ], + "type": "text", + "content": "pared to G2P-DDM and VQ-GCDM, which yield higher DTW and MPJPE scores, our model achieves smoother and more accurate pose transitions." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 55, + 445, + 295, + 552 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 445, + 295, + 552 + ], + "spans": [ + { + "bbox": [ + 55, + 445, + 295, + 552 + ], + "type": "text", + "content": "In the \"Observation 10 Frames - Prediction 20 Frames\" scenario, our method continues to demonstrate resilience in generating long sequences. Our BLEU-1 and BLEU-4 scores on DEV are " + }, + { + "bbox": [ + 55, + 445, + 295, + 552 + ], + "type": "inline_equation", + "content": "67.95\\%" + }, + { + "bbox": [ + 55, + 445, + 295, + 552 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 55, + 445, + 295, + 552 + ], + "type": "inline_equation", + "content": "56.61\\%" + }, + { + "bbox": [ + 55, + 445, + 295, + 552 + ], + "type": "text", + "content": ", respectively, showing superior semantic preservation compared to the baseline methods. Additionally, the DTW score of 1.79 and MPJPE score of 78.62 confirm our model's ability to produce coherent and realistic transitions, outperforming the baselines in alignment with true pose sequences." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 55, + 563, + 167, + 575 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 563, + 167, + 575 + ], + "spans": [ + { + "bbox": [ + 55, + 563, + 167, + 575 + ], + "type": "text", + "content": "4.4. Qualitative Results" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 55, + 581, + 295, + 700 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 581, + 295, + 700 + ], + "spans": [ + { + "bbox": [ + 55, + 581, + 295, + 700 + ], + "type": "text", + "content": "Visualization Results. Figure 3 presents a comparative demonstration of our method against other approaches (i.e., G2P-DDM [47] and VQ-GCDM [43]) in generating transition sign poses. This example clearly shows that our method significantly outperforms G2P-DDM and VQ-GCDM in the quality of interpolated movements, especially excelling in overall motion coherence. Furthermore, our method meticulously fine-tunes the observations (the Blue and Orange sections), effectively avoiding coherence issues that may arise from directly concatenating discrete sign segments." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 55, + 701, + 295, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 701, + 295, + 713 + ], + "spans": [ + { + "bbox": [ + 55, + 701, + 295, + 713 + ], + "type": "text", + "content": "Verification in Realistic Scenarios. Figure 4 demonstrates" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 313, + 408, + 555, + 516 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 408, + 555, + 516 + ], + "spans": [ + { + "bbox": [ + 313, + 408, + 555, + 516 + ], + "type": "text", + "content": "visualization examples for synthesizing continuous sign videos from discrete segments in a realistic, word-level scenario using the USTC-SLR500 dataset. Here, the observed frames from two isolated sign segments are marked in blue and orange, while the generated transitions are shown in green. The results show that our method not only generates smooth transitions between discrete sign language segments but also fine-tunes the observed poses based on the overall semantics, resulting in a coherent sign language video." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 314, + 535, + 391, + 547 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 314, + 535, + 391, + 547 + ], + "spans": [ + { + "bbox": [ + 314, + 535, + 391, + 547 + ], + "type": "text", + "content": "5. Conclusions" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 313, + 558, + 555, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 558, + 555, + 713 + ], + "spans": [ + { + "bbox": [ + 313, + 558, + 555, + 713 + ], + "type": "text", + "content": "This paper proposes a method to generate continuous sign language videos by synthesizing transition frames between discrete segments, transforming an unsupervised task into a supervised one through random masking. Using a conditional diffusion model, the approach learns to predict missing frames based on observable context, allowing for smooth and temporally coherent outputs. By incorporating a linearly interpolating padding strategy during inference, the model initializes transitions effectively, further refining them into seamless sequences. Extensive experiments demonstrate the effectiveness of leveraging context for natural transitions, advancing sign language synthesis by enabling fluid, continuous sequences." + } + ] + } + ], + "index": 11 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "text", + "content": "3488" + } + ] + } + ], + "index": 12 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 7 + }, + { + "para_blocks": [ + { + "bbox": [ + 56, + 72, + 165, + 85 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 72, + 165, + 85 + ], + "spans": [ + { + "bbox": [ + 56, + 72, + 165, + 85 + ], + "type": "text", + "content": "6. Acknowledgments" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 55, + 91, + 296, + 186 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 91, + 296, + 186 + ], + "spans": [ + { + "bbox": [ + 55, + 91, + 296, + 186 + ], + "type": "text", + "content": "This work was supported by the National Natural Science Foundation of China (Grants No. U23B2031, 62472139, 62302142, 62272144), the Anhui Provincial Natural Science Foundation, China (Grant No. 2408085QF191), the Fundamental Research Funds for the Central Universities (Grants No. JZ2024HGTA0178, JZ2023HGQA0097), and the Major Project of Anhui Province (Grant No. 202423k09020001)." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 56, + 196, + 115, + 209 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 196, + 115, + 209 + ], + "spans": [ + { + "bbox": [ + 56, + 196, + 115, + 209 + ], + "type": "text", + "content": "References" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 57, + 216, + 296, + 712 + ], + "type": "list", + "angle": 0, + "index": 14, + "blocks": [ + { + "bbox": [ + 61, + 216, + 296, + 259 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 61, + 216, + 296, + 259 + ], + "spans": [ + { + "bbox": [ + 61, + 216, + 296, + 259 + ], + "type": "text", + "content": "[1] Ahmed H Aliwy and A Alethary Ahmed. Development of arabic sign language dictionary using 3d avatar technologies. Indonesian Journal of Electrical Engineering and Computer Science, 21(1):609-616, 2021. 1" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 62, + 261, + 296, + 304 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 261, + 296, + 304 + ], + "spans": [ + { + "bbox": [ + 62, + 261, + 296, + 304 + ], + "type": "text", + "content": "[2] Necati Cihan Camgoz, Simon Hadfield, Oscar Koller, Hermann Ney, and Richard Bowden. Neural sign language translation. In Computer Vision and Pattern Recognition, pages 7784-7793, 2018. 1, 6" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 62, + 304, + 295, + 348 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 304, + 295, + 348 + ], + "spans": [ + { + "bbox": [ + 62, + 304, + 295, + 348 + ], + "type": "text", + "content": "[3] Necati Cihan Camgoz, Oscar Koller, Simon Hadfield, and Richard Bowden. Multi-channel transformers for multi-articulatory sign language translation. In European Conference on Computer Vision, pages 301-319, 2020. 1, 2" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 62, + 350, + 294, + 392 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 350, + 294, + 392 + ], + "spans": [ + { + "bbox": [ + 62, + 350, + 294, + 392 + ], + "type": "text", + "content": "[4] Zhe Cao, Tomas Simon, Shih-En Wei, and Yaser Sheikh. Realtime multi-person 2d pose estimation using part affinity fields. In Computer Vision and Pattern Recognition, pages 7291-7299, 2017. 6" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 62, + 393, + 295, + 448 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 393, + 295, + 448 + ], + "spans": [ + { + "bbox": [ + 62, + 393, + 295, + 448 + ], + "type": "text", + "content": "[5] Xiujuan Chai, Hanjie Wang, and Xilin Chen. The devisesign large vocabulary of Chinese sign language database and baseline evaluations. In Technical report, Key Lab of Intelligent Information Processing of CAS. Institute of Computing Technology, 2014. 1" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 62, + 449, + 294, + 492 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 449, + 294, + 492 + ], + "spans": [ + { + "bbox": [ + 62, + 449, + 294, + 492 + ], + "type": "text", + "content": "[6] Mathieu De Coster and Joni Dambre. Querying a sign language dictionary with videos using dense vector search. In 2023 IEEE International Conference on Acoustics, Speech, and Signal Processing Workshops, pages 1-5. IEEE, 2023. 1" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 62, + 493, + 295, + 536 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 493, + 295, + 536 + ], + "spans": [ + { + "bbox": [ + 62, + 493, + 295, + 536 + ], + "type": "text", + "content": "[7] Philippe Dreuw, Carol Neidle, Vassilis Athitsos, Stan Sclaroff, and Hermann Ney. Benchmark databases for videobased automatic sign language recognition. In International Conference on Language Resources and Evaluation, 2008. 1" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 62, + 537, + 294, + 569 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 537, + 294, + 569 + ], + "spans": [ + { + "bbox": [ + 62, + 537, + 294, + 569 + ], + "type": "text", + "content": "[8] Amanda Cardoso Duarte. Cross-modal neural sign language translation. In ACM International Conference on Multimedia, pages 1650-1654, 2019. 1" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 62, + 571, + 295, + 613 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 571, + 295, + 613 + ], + "spans": [ + { + "bbox": [ + 62, + 571, + 295, + 613 + ], + "type": "text", + "content": "[9] Jia Gong, Lin Geng Foo, Yixuan He, Hossein Rahmani, and Jun Liu. Llms are good sign language translators. In Computer Vision and Pattern Recognition, pages 18362-18372, 2024. 1, 2" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 57, + 614, + 295, + 668 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 614, + 295, + 668 + ], + "spans": [ + { + "bbox": [ + 57, + 614, + 295, + 668 + ], + "type": "text", + "content": "[10] Dan Guo, Wengang Zhou, Houqiang Li, and Meng Wang. Online early-late fusion based on adaptive hmm for sign language recognition. ACM Transactions on Multimedia Computing, Communications, and Applications, pages 1-18, 2017. 2" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 57, + 670, + 295, + 712 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 670, + 295, + 712 + ], + "spans": [ + { + "bbox": [ + 57, + 670, + 295, + 712 + ], + "type": "text", + "content": "[11] Dan Guo, Wengang Zhou, Houqiang Li, and Meng Wang. Hierarchical LSTM for sign language translation. In AAAI Conference on Artificial Intelligence, pages 6845-6852, 2018. 6" + } + ] + } + ], + "index": 13 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 316, + 73, + 554, + 713 + ], + "type": "list", + "angle": 0, + "index": 30, + "blocks": [ + { + "bbox": [ + 316, + 73, + 554, + 117 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 73, + 554, + 117 + ], + "spans": [ + { + "bbox": [ + 316, + 73, + 554, + 117 + ], + "type": "text", + "content": "[12] Dan Guo, Shengeng Tang, and Meng Wang. Connectionist temporal modeling of video and language: A joint model for translation and sign labeling. In International Joint Conference on Artificial Intelligence, pages 751-757, 2019. 2" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 317, + 118, + 553, + 150 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 118, + 553, + 150 + ], + "spans": [ + { + "bbox": [ + 317, + 118, + 553, + 150 + ], + "type": "text", + "content": "[13] Dan Guo, Shengeng Tang, Richang Hong, and Meng Wang. Sign language recognition. Multimedia for Accessible Human Computer Interfaces, pages 23-59, 2021. 1" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 317, + 151, + 553, + 193 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 151, + 553, + 193 + ], + "spans": [ + { + "bbox": [ + 317, + 151, + 553, + 193 + ], + "type": "text", + "content": "[14] Jonathan Ho, Ajay Jain, and Pieter Abbeel. Denoising diffusion probabilistic models. Conference and Workshop on Neural Information Processing Systems, 33:6840-6851, 2020. 3" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 316, + 194, + 553, + 238 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 194, + 553, + 238 + ], + "spans": [ + { + "bbox": [ + 316, + 194, + 553, + 238 + ], + "type": "text", + "content": "[15] Hezhen Hu, Junfu Pu, Wengang Zhou, and Houqiang Li. Collaborative multilingual continuous sign language recognition: A unified framework. IEEE Transactions on Multimedia, pages 7559-7570, 2022. 1, 2" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 316, + 239, + 553, + 293 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 239, + 553, + 293 + ], + "spans": [ + { + "bbox": [ + 316, + 239, + 553, + 293 + ], + "type": "text", + "content": "[16] Hezhen Hu, Weichao Zhao, Wengang Zhou, and Houqiang Li. Signbert+: Hand-model-aware self-supervised pretraining for sign language understanding. IEEE Transactions on Pattern Analysis and Machine Intelligence, pages 11221-11239, 2023. 1, 2" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 316, + 294, + 553, + 338 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 294, + 553, + 338 + ], + "spans": [ + { + "bbox": [ + 316, + 294, + 553, + 338 + ], + "type": "text", + "content": "[17] Jie Huang, Wengang Zhou, Houqiang Li, and Weiping Li. Attention-based 3d-cnns for large-vocabulary sign language recognition. IEEE Transactions on Circuits and Systems for Video Technology, pages 2822-2832, 2018. 6" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 316, + 338, + 553, + 381 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 338, + 553, + 381 + ], + "spans": [ + { + "bbox": [ + 316, + 338, + 553, + 381 + ], + "type": "text", + "content": "[18] Jie Huang, Wengang Zhou, Qilin Zhang, Houqiang Li, and Weiping Li. Video-based sign language recognition without temporal segmentation. In AAAI Conference on Artificial Intelligence, pages 2257-2264, 2018. 1, 6" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 316, + 383, + 553, + 425 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 383, + 553, + 425 + ], + "spans": [ + { + "bbox": [ + 316, + 383, + 553, + 425 + ], + "type": "text", + "content": "[19] Wencan Huang, Wenwen Pan, Zhou Zhao, and Qi Tian. Towards fast and high-quality sign language production. In ACM International Conference on Multimedia, pages 3172-3181, 2021. 6" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 316, + 426, + 553, + 470 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 426, + 553, + 470 + ], + "spans": [ + { + "bbox": [ + 316, + 426, + 553, + 470 + ], + "type": "text", + "content": "[20] Jestin Joy, Kannan Balakrishnan, and Sreeraj Madhavankutty. Developing a bilingual mobile dictionary for indian sign language and gathering users experience with signdict. Assistive Technology, 2020. 1" + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 316, + 471, + 553, + 514 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 471, + 553, + 514 + ], + "spans": [ + { + "bbox": [ + 316, + 471, + 553, + 514 + ], + "type": "text", + "content": "[21] Korrawe Karunratanakul, Konpat Preechakul, Supasorn Suwajanakorn, and Siyu Tang. Guided motion diffusion for controllable human motion synthesis. In International Conference on Computer Vision, pages 2151-2162, 2023. 3" + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 316, + 515, + 553, + 548 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 515, + 553, + 548 + ], + "spans": [ + { + "bbox": [ + 316, + 515, + 553, + 548 + ], + "type": "text", + "content": "[22] Oscar Koller. Quantitative survey of the state of the art in sign language recognition. ArXiv Preprint ArXiv:2008.09918, pages ArXiv-2008, 2020. 1, 2" + } + ] + } + ], + "index": 25 + }, + { + "bbox": [ + 317, + 548, + 553, + 581 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 548, + 553, + 581 + ], + "spans": [ + { + "bbox": [ + 317, + 548, + 553, + 581 + ], + "type": "text", + "content": "[23] Taegyeong Lee, Soyeong Kwon, and Taehwan Kim. Grid diffusion models for text-to-video generation. In Computer Vision and Pattern Recognition, pages 8734-8743, 2024. 2" + } + ] + } + ], + "index": 26 + }, + { + "bbox": [ + 317, + 582, + 553, + 635 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 582, + 553, + 635 + ], + "spans": [ + { + "bbox": [ + 317, + 582, + 553, + 635 + ], + "type": "text", + "content": "[24] Dongxu Li, Cristian Rodriguez, Xin Yu, and Hongdong Li. Word-level deep sign language recognition from video: A new large-scale dataset and methods comparison. In IEEE Winter Conference on Applications of Computer Vision, pages 1459-1469, 2020. 1" + } + ] + } + ], + "index": 27 + }, + { + "bbox": [ + 317, + 636, + 553, + 690 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 636, + 553, + 690 + ], + "spans": [ + { + "bbox": [ + 317, + 636, + 553, + 690 + ], + "type": "text", + "content": "[25] Dongxu Li, Chenchen Xu, Xin Yu, Kaihao Zhang, Benjamin Swift, Hanna Suominen, and Hongdong Li. Tspnet: Hierarchical feature learning via temporal semantic pyramid for sign language translation. Neural Information Processing Systems, pages 12034-12045, 2020. 2" + } + ] + } + ], + "index": 28 + }, + { + "bbox": [ + 316, + 691, + 553, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 691, + 553, + 713 + ], + "spans": [ + { + "bbox": [ + 316, + 691, + 553, + 713 + ], + "type": "text", + "content": "[26] Jingyun Liang, Yuchen Fan, Kai Zhang, Radu Timofte, Luc Van Gool, and Rakesh Ranjan. Movideo: Motion-aware" + } + ] + } + ], + "index": 29 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "text", + "content": "3489" + } + ] + } + ], + "index": 31 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 8 + }, + { + "para_blocks": [ + { + "bbox": [ + 56, + 72, + 295, + 713 + ], + "type": "list", + "angle": 0, + "index": 14, + "blocks": [ + { + "bbox": [ + 75, + 72, + 294, + 105 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 75, + 72, + 294, + 105 + ], + "spans": [ + { + "bbox": [ + 75, + 72, + 294, + 105 + ], + "type": "text", + "content": "video generation with diffusion model. In European Conference on Computer Vision, pages 56-74. Springer, 2025. 3" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 56, + 106, + 295, + 138 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 106, + 295, + 138 + ], + "spans": [ + { + "bbox": [ + 56, + 106, + 295, + 138 + ], + "type": "text", + "content": "[27] Chin-Yew Lin. Rouge: A package for automatic evaluation of summaries. In Text Summarization Branches Out, pages 74-81, 2004. 6" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 56, + 140, + 295, + 193 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 140, + 295, + 193 + ], + "spans": [ + { + "bbox": [ + 56, + 140, + 295, + 193 + ], + "type": "text", + "content": "[28] Seshadri Mazumder, Rudrabha Mukhopadhyay, Vinay P Namboodiri, and CV Jawahar. Translating sign language videos to talking faces. In Indian Conference on Computer Vision, Graphics and Image Processing, pages 1-10, 2021. 2" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 56, + 194, + 295, + 260 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 194, + 295, + 260 + ], + "spans": [ + { + "bbox": [ + 56, + 194, + 295, + 260 + ], + "type": "text", + "content": "[29] John McDonald, Rosalee Wolfe, Jerry Schnegg, Julie Hochgesang, Diana Gorman Jamrozik, Marie Stumbo, Larwan Berke, Melissa Bialek, and Farah Thomas. An automated technique for real-time production of lifelike animations of american sign language. Universal Access in the Information Society, pages 551-566, 2016. 2" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 56, + 261, + 295, + 315 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 261, + 295, + 315 + ], + "spans": [ + { + "bbox": [ + 56, + 261, + 295, + 315 + ], + "type": "text", + "content": "[30] Yuchun Miao, Lefei Zhang, Liangpei Zhang, and Dacheng Tao. Dds2m: Self-supervised denoising diffusion spati-spectral model for hyperspectral image restoration. In International Conference on Computer Vision, pages 12086-12096, 2023. 3" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 56, + 316, + 295, + 360 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 316, + 295, + 360 + ], + "spans": [ + { + "bbox": [ + 56, + 316, + 295, + 360 + ], + "type": "text", + "content": "[31] Shentong Mo, Enze Xie, Ruihang Chu, Lanqing Hong, Matthias Niessner, and Zhenguo Li. Dit-3d: Exploring plain diffusion transformers for 3d shape generation. Neural Information Processing Systems, 36:67960-67971, 2023. 3" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 56, + 361, + 295, + 403 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 361, + 295, + 403 + ], + "spans": [ + { + "bbox": [ + 56, + 361, + 295, + 403 + ], + "type": "text", + "content": "[32] Haomiao Ni, Changhao Shi, Kai Li, Sharon X Huang, and Martin Renqiang Min. Conditional image-to-video generation with latent flow diffusion models. In Computer Vision and Pattern Recognition, pages 18444-18455, 2023. 3" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 56, + 404, + 295, + 448 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 404, + 295, + 448 + ], + "spans": [ + { + "bbox": [ + 56, + 404, + 295, + 448 + ], + "type": "text", + "content": "[33] Kishore Papineni, Salim Roukos, Todd Ward, and Wei-Jing Zhu. Bleu: a method for automatic evaluation of machine translation. In Association for Computational Linguistics, pages 311-318, 2002. 6" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 56, + 449, + 295, + 492 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 449, + 295, + 492 + ], + "spans": [ + { + "bbox": [ + 56, + 449, + 295, + 492 + ], + "type": "text", + "content": "[34] Robin Rombach, Andreas Blattmann, Dominik Lorenz, Patrick Esser, and Björn Ommer. High-resolution image synthesis with latent diffusion models. In Computer Vision and Pattern Recognition, pages 10684-10695, 2022. 2" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 56, + 493, + 295, + 536 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 493, + 295, + 536 + ], + "spans": [ + { + "bbox": [ + 56, + 493, + 295, + 536 + ], + "type": "text", + "content": "[35] Ben Saunders, Necati Cihan Camgoz, and Richard Bowden. Progressive transformers for end-to-end sign language production. In European Conference on Computer Vision, pages 687-705, 2020. 2, 6" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 56, + 537, + 295, + 581 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 537, + 295, + 581 + ], + "spans": [ + { + "bbox": [ + 56, + 537, + 295, + 581 + ], + "type": "text", + "content": "[36] Ben Saunders, Necati Cihan Camgoz, and Richard Bowden. Mixed signals: Sign language production via a mixture of motion primitives. In International Conference on Computer Vision, pages 1919-1929, 2021. 6" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 56, + 582, + 295, + 613 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 582, + 295, + 613 + ], + "spans": [ + { + "bbox": [ + 56, + 582, + 295, + 613 + ], + "type": "text", + "content": "[37] Jérémie Segouat. A study of sign language coarticulation. ACM Sigaccess Accessibility and Computing, pages 31-38, 2009. 2" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 56, + 614, + 295, + 669 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 614, + 295, + 669 + ], + "spans": [ + { + "bbox": [ + 56, + 614, + 295, + 669 + ], + "type": "text", + "content": "[38] Wenkang Shan, Zhenhua Liu, Xinfeng Zhang, Zhao Wang, Kai Han, Shanshe Wang, Siwei Ma, and Wen Gao. Diffusion-based 3d human pose estimation with multi-hypothesis aggregation. In International Conference on Computer Vision, pages 14761-14771, 2023. 3" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 56, + 670, + 295, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 670, + 295, + 713 + ], + "spans": [ + { + "bbox": [ + 56, + 670, + 295, + 713 + ], + "type": "text", + "content": "[39] Jascha Sohl-Dickstein, Eric Weiss, Niru Maheswaranathan, and Surya Ganguli. Deep unsupervised learning using nonequilibrium thermodynamics. In International Conference on Machine Learning, pages 2256-2265, 2015. 2" + } + ] + } + ], + "index": 13 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 316, + 72, + 554, + 712 + ], + "type": "list", + "angle": 0, + "index": 28, + "blocks": [ + { + "bbox": [ + 316, + 72, + 554, + 128 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 72, + 554, + 128 + ], + "spans": [ + { + "bbox": [ + 316, + 72, + 554, + 128 + ], + "type": "text", + "content": "[40] Stephanie Stoll, Necati Cihan Camgoz, Simon Hadfield, and Richard Bowden. Text2sign: Towards sign language production using neural machine translation and generative adversarial networks. International Journal of Computer Vision, pages 891–908, 2020. 2" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 316, + 130, + 554, + 174 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 130, + 554, + 174 + ], + "spans": [ + { + "bbox": [ + 316, + 130, + 554, + 174 + ], + "type": "text", + "content": "[41] Sugandhi, Parteek Kumar, and Sanmeet Kaur. Sign language generation system based on indian sign language grammar. ACM Transactions on Asian and Low-Resource Language Information Processing, 19(4):1-26, 2020. 2" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 316, + 176, + 554, + 219 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 176, + 554, + 219 + ], + "spans": [ + { + "bbox": [ + 316, + 176, + 554, + 219 + ], + "type": "text", + "content": "[42] Shengeng Tang, Dan Guo, Richang Hong, and Meng Wang. Graph-based multimodal sequential embedding for sign language translation. IEEE Transactions on Multimedia, 24: 4433-4445, 2022. 1" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 316, + 222, + 554, + 276 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 222, + 554, + 276 + ], + "spans": [ + { + "bbox": [ + 316, + 222, + 554, + 276 + ], + "type": "text", + "content": "[43] Shengeng Tang, Feng Xue, Jingjing Wu, Shuo Wang, and Richang Hong. Gloss-driven conditional diffusion models for sign language production. ACM Transactions on Multimedia Computing, Communications, and Applications, 2024. 2, 3, 7, 8" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 316, + 279, + 554, + 323 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 279, + 554, + 323 + ], + "spans": [ + { + "bbox": [ + 316, + 279, + 554, + 323 + ], + "type": "text", + "content": "[44] Carla Viegas, Mert Inan, Lorna Quandt, and Malihe Alikhani. Including facial expressions in contextual embeddings for sign language generation. In Joint Conference on Lexical and Computational Semantics, pages 1-10, 2023. 6" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 316, + 325, + 554, + 379 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 325, + 554, + 379 + ], + "spans": [ + { + "bbox": [ + 316, + 325, + 554, + 379 + ], + "type": "text", + "content": "[45] Bin Xia, Yulun Zhang, Shiyin Wang, Yitong Wang, Xinglong Wu, Yapeng Tian, Wenming Yang, and Luc Van Gool. Diffir: Efficient diffusion model for image restoration. In International Conference on Computer Vision, pages 13095-13105, 2023. 3" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 316, + 382, + 554, + 437 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 382, + 554, + 437 + ], + "spans": [ + { + "bbox": [ + 316, + 382, + 554, + 437 + ], + "type": "text", + "content": "[46] Jinheng Xie, Yuexiang Li, Yawen Huang, Haozhe Liu, Wentian Zhang, Yefeng Zheng, and Mike Zheng Shou. Boxdiff: Text-to-image synthesis with training-free box-constrained diffusion. In International Conference on Computer Vision, pages 7452-7461, 2023. 3" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 316, + 439, + 554, + 493 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 439, + 554, + 493 + ], + "spans": [ + { + "bbox": [ + 316, + 439, + 554, + 493 + ], + "type": "text", + "content": "[47] Pan Xie, Qipeng Zhang, Peng Taiying, Hao Tang, Yao Du, and Zexian Li. G2p-ddm: Generating sign pose sequence from gloss sequence with discrete diffusion model. In AAAI Conference on Artificial Intelligence, pages 6234-6242, 2024. 2, 3, 7, 8" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 316, + 496, + 554, + 529 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 496, + 554, + 529 + ], + "spans": [ + { + "bbox": [ + 316, + 496, + 554, + 529 + ], + "type": "text", + "content": "[48] Wanli Xue, Jingze Liu, Siyi Yan, Yuxi Zhou, Tiantian Yuan, and Qing Guo. Alleviating data insufficiency for Chinese sign language recognition. Visual Intelligence, page 26, 2023. 2" + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 316, + 532, + 554, + 575 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 532, + 554, + 575 + ], + "spans": [ + { + "bbox": [ + 316, + 532, + 554, + 575 + ], + "type": "text", + "content": "[49] Shuai Yang, Yifan Zhou, Ziwei Liu, and Chen Change Loy. Fresco: Spatial-temporal correspondence for zero-shot video translation. In Computer Vision and Pattern Recognition, pages 8703-8712, 2024. 2" + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 316, + 577, + 554, + 620 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 577, + 554, + 620 + ], + "spans": [ + { + "bbox": [ + 316, + 577, + 554, + 620 + ], + "type": "text", + "content": "[50] Xunpeng Yi, Han Xu, Hao Zhang, Linfeng Tang, and Jiayi Ma. Diff-retinex: Rethinking low-light image enhancement with a generative diffusion model. In International Conference on Computer Vision, pages 12302-12311, 2023. 3" + } + ] + } + ], + "index": 25 + }, + { + "bbox": [ + 316, + 623, + 554, + 656 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 623, + 554, + 656 + ], + "spans": [ + { + "bbox": [ + 316, + 623, + 554, + 656 + ], + "type": "text", + "content": "[51] Jan Zelinka and Jakub Kanis. Neural sign language synthesis: Words are our glosses. In Winter Conference on Applications of Computer Vision, pages 3395-3403, 2020. 2, 6" + } + ] + } + ], + "index": 26 + }, + { + "bbox": [ + 316, + 658, + 554, + 712 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 658, + 554, + 712 + ], + "spans": [ + { + "bbox": [ + 316, + 658, + 554, + 712 + ], + "type": "text", + "content": "[52] Ni Zeng, Yiqiang Chen, Yang Gu, Dongdong Liu, and Yunbing Xing. Highly fluent sign language synthesis based on variable motion frame interpolation. In 2020 IEEE International Conference on Systems, Man, and Cybernetics, pages 1772-1777. IEEE, 2020. 2" + } + ] + } + ], + "index": 27 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 748, + 316, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 316, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 316, + 757 + ], + "type": "text", + "content": "3490" + } + ] + } + ], + "index": 29 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 9 + }, + { + "para_blocks": [ + { + "bbox": [ + 56, + 73, + 294, + 162 + ], + "type": "list", + "angle": 0, + "index": 2, + "blocks": [ + { + "bbox": [ + 56, + 73, + 294, + 116 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 73, + 294, + 116 + ], + "spans": [ + { + "bbox": [ + 56, + 73, + 294, + 116 + ], + "type": "text", + "content": "[53] Jihai Zhang, Wengang Zhou, Chao Xie, Junfu Pu, and Houqiang Li. Chinese sign language recognition with adaptive hmm. In IEEE International Conference on Multimedia and Expo, pages 1-6. IEEE, 2016. 1" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 56, + 118, + 294, + 162 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 118, + 294, + 162 + ], + "spans": [ + { + "bbox": [ + 56, + 118, + 294, + 162 + ], + "type": "text", + "content": "[54] Weichao Zhao, Hezhen Hu, Wengang Zhou, Jiaxin Shi, and Houqiang Li. Best: Bert pre-training for sign language recognition with coupling tokenization. In AAAI Conference on Artificial Intelligence, pages 3597-3605, 2023. 1, 2" + } + ] + } + ], + "index": 1 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 749, + 313, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 749, + 313, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 749, + 313, + 757 + ], + "type": "text", + "content": "3491" + } + ] + } + ], + "index": 3 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 10 + } + ], + "_backend": "vlm", + "_version_name": "2.6.4" +} \ No newline at end of file diff --git a/2025/Disentangled Pose and Appearance Guidance for Multi-Pose Generation/eaadccfc-d365-4d2d-83e8-71935f857eca_content_list.json b/2025/Disentangled Pose and Appearance Guidance for Multi-Pose Generation/eaadccfc-d365-4d2d-83e8-71935f857eca_content_list.json new file mode 100644 index 0000000000000000000000000000000000000000..3fdc73cdc2929531d048a7e960b9d1d4c1f164b1 --- /dev/null +++ b/2025/Disentangled Pose and Appearance Guidance for Multi-Pose Generation/eaadccfc-d365-4d2d-83e8-71935f857eca_content_list.json @@ -0,0 +1,1333 @@ +[ + { + "type": "text", + "text": "Disentangled Pose and Appearance Guidance for Multi-Pose Generation", + "text_level": 1, + "bbox": [ + 133, + 130, + 864, + 152 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Tengfei Xiao $^{1,4}$ , Yue Wu $^{\\dagger,*,1,4}$ , Yuelong Li $^{\\dagger,*,2,5}$ , Can Qin $^{3}$ , Maoguo Gong $^{1,4}$ , Qiguang Miao $^{1,4}$ , Wenping Ma $^{1}$ $^{1}$ Xidian University $^{2}$ Tiangong University $^{3}$ Northeastern University $^{4}$ MoE Key Lab of Collaborative Intelligence Systems, Xidian University \n $^{5}$ Tianjin Key Laboratory of Autonomous Intelligence Technology and Systems, Tiangong University tfxiao@stu.xidian.edu.cn, {ywu, qgmiao}@xidian.edu.cn, liyuelong@pku.edu.cn, qin.ca@northeastern.edu, gong@ieee.org, wpma@mail.xidian.edu.cn", + "bbox": [ + 101, + 178, + 895, + 305 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/0206018f2f10f04df1d5cff292ffedde2c0bf9b7ab97c5366686992084d8ca15.jpg", + "image_caption": [ + "Figure 1. Given a reference image and different target poses, our method can generate multiple images of the person in various poses." + ], + "image_footnote": [], + "bbox": [ + 99, + 311, + 900, + 631 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Abstract", + "text_level": 1, + "bbox": [ + 248, + 661, + 326, + 676 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Human pose generation is a complex task due to the non-rigid and highly variable nature of human body structures and appearances. However, existing methods often overlook the fundamental differences between spatial transformations of poses and texture generation for appearance, which makes them prone to overfitting. To address this issue, we propose a multi-pose generation framework driven by disentangled pose and appearance guidance. Our approach includes a Global-aware Pose Generation module that iteratively generates pose embeddings, enabling effective control over non-rigid body deformations. Addition-", + "bbox": [ + 89, + 694, + 483, + 861 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "ally, we introduce the Global-aware Transformer Decoder, which leverages similarity queries and attention mechanisms to achieve spatial transformations and enhance pose consistency through a Global-aware block. In the appearance generation phase, we condition a diffusion model on pose embeddings produced in the initial stage and introduce an Appearance Adapter that extracts high-level contextual semantic information from multi-scale features, enabling further refinement of pose appearance textures and providing appearance guidance. Extensive experiments on the UBC Fashion and TikTok datasets demonstrate that our framework achieves state-of-the-art results in both quality and fidelity, establishing it as a powerful approach for complex pose generation tasks.", + "bbox": [ + 511, + 662, + 906, + 875 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "CVF", + "bbox": [ + 106, + 2, + 181, + 42 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore.", + "bbox": [ + 236, + 0, + 810, + 46 + ], + "page_idx": 0 + }, + { + "type": "page_footnote", + "text": "$\\dagger$ Equal contributions in no particular order.", + "bbox": [ + 112, + 875, + 346, + 887 + ], + "page_idx": 0 + }, + { + "type": "page_footnote", + "text": "* Corresponding author.", + "bbox": [ + 112, + 887, + 245, + 898 + ], + "page_idx": 0 + }, + { + "type": "page_number", + "text": "5646", + "bbox": [ + 482, + 944, + 516, + 955 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "1. Introduction", + "text_level": 1, + "bbox": [ + 91, + 89, + 222, + 106 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "The inherent complexity of human body poses, with their non-rigid, dynamic structures and diverse appearances, makes pose generation a challenging task. This process requires transforming a person's reference pose into any target pose while preserving the individual's identity and appearance with high fidelity. This capability has substantial applications in areas such as virtual reality (VR) [1], augmented reality (AR) [2], and person re-identification [3]. Pose generation, however, involves two distinct challenges: effectively controlling the non-rigid deformations of human body structures and maintaining appearance consistency across different poses and viewpoints. Most existing methods directly generate human poses without addressing the fundamental differences between pose control and appearance generation.", + "bbox": [ + 89, + 114, + 482, + 340 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Advancements in deep learning have introduced promising solutions to pose generation, primarily through two classes of techniques: GAN-based methods and diffusion-based methods. GAN-based methods [4-8] typically align the reference pose with the target pose directly, learning a simple mapping from the appearance features of the reference pose to the target pose, which results in poor generalization. Diffusion models [9], as parameterized Markov chains trained via variational inference, offer straightforward optimization objectives and training efficiency, making them a viable alternative to GANs. Recent works [10-12] have applied diffusion models to pose generation, achieving promising results. However, we find that these methods still face limitations in terms of generation quality, identity preservation, or efficiency.", + "bbox": [ + 89, + 340, + 482, + 568 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "In this paper, we present a novel framework for multi-posede generation based on disentangled pose and appearance guidance. In our approach, pose control primarily addresses spatial geometry, while appearance generation leverages semantic and contextual information. By assigning each task to a distinct module, the model efficiently addresses each objective independently. We introduce a Global-aware Pose Generation module that iteratively produces pose embeddings. Each embedding is a compact representation within the feature space, containing the spatially transformed pose along with appearance priors from the reference image. To generate pose embeddings, we propose the Global-aware Transformer Decoder (GTD), which employs similarity queries and attention mechanisms for spatial transformation of poses. An additional Global-aware block within the GTD models poses generated in previous iterations, integrating enriched global information to enhance pose consistency. During the appearance generation phase, the diffusion model is conditioned on multiple pose embeddings produced in the initial stage to refine appearance details. To further enhance appearance fidelity, we implement an Appearance Adapter module that encodes", + "bbox": [ + 89, + 569, + 482, + 901 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "multi-scale features to extract high-level semantic information, providing fine-grained appearance guidance to the diffusion model.", + "bbox": [ + 511, + 90, + 903, + 135 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "In summary, our main contributions include:", + "bbox": [ + 532, + 136, + 828, + 151 + ], + "page_idx": 1 + }, + { + "type": "list", + "sub_type": "text", + "list_items": [ + "- Introducing a novel multi-pose generation framework based on disentangled pose and appearance guidance.", + "- Developing a global-aware pose generation module that models all generated poses, facilitating feature interactions among poses and integrating global information for improved pose generation.", + "- Designing an appearance adapter module to extract high-level contextual semantic information from multi-scale features, providing fine-grained appearance guidance to the diffusion model.", + "- Conducting extensive experiments on the UBC Fashion [13] and TikTok [14] datasets, achieving state-of-the-art results in both quality and fidelity." + ], + "bbox": [ + 511, + 152, + 903, + 349 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "2. Related Works", + "text_level": 1, + "bbox": [ + 513, + 362, + 663, + 378 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Generative Adversarial Networks. Over the past decades, Generative Adversarial Networks (GANs) [15] have been extensively applied to human pose generation. Leveraging an adversarial learning framework, GANs can effectively model complex pose spatial transformations while producing realistic and detailed textures. Despite significant progress in recent years, challenges remain in predicting invisible regions, generating complex appearance textures, and maintaining identity consistency.", + "bbox": [ + 511, + 387, + 903, + 523 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "To achieve accurate human pose synthesis with consistent clothing appearance and style, PISE [8] introduces a two-stage generative model that decouples appearance control from pose generation, enabling the creation of realistic human images. BiGraphGAN [4] addresses the challenges of pose deformation by incorporating a bipartite graph reasoning (BGR) block, which captures long-range relations between source and target poses in a bipartite graph structure. Additionally, BiGraphGAN includes an interaction- and-aggregation (IA) block to improve the feature representation of both body shape and appearance.", + "bbox": [ + 511, + 523, + 903, + 689 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Recently, attention mechanisms have been integrated into pose generation tasks. DPTN [5] employs a transformer-based siamese network, which supports pose generation by reconstructing reference images, thereby maintaining consistency in the target pose's appearance. NTED [7] presents a neural texture extraction and distribution module based on dual attention, extracting semantic textures from source features and distributing them according to the spatial layout of the target pose. IEPG [6] proposes a progressive pose generation framework, designed with an incremental evolution approach to address pose synthesis challenges. This multi-step generation strategy effectively handles occlusions, significantly enhancing the realism and coherence of the generated appearances.", + "bbox": [ + 511, + 689, + 903, + 900 + ], + "page_idx": 1 + }, + { + "type": "page_number", + "text": "5647", + "bbox": [ + 482, + 944, + 514, + 955 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Diffusion models. Diffusion models (DM) [9] have emerged as a promising alternative to GANs, decomposing the generation process into a series of forward-backward diffusion steps to learn plausible transfer trajectories. This iterative refinement, starting with a low-quality noisy signal and progressively enhancing it, enables the generation of high-quality, contextually consistent samples, making diffusion models particularly effective for generating high-fidelity results. PIDM [12] introduces a texture diffusion module based on cross-attention, which accurately models the correspondences between pose and appearance information at the pixel level across source and target images. Additionally, PIDM employs a disentangled classifier-free guidance strategy [16] to ensure the conditional input and generated output closely match both pose and appearance. While PIDM achieves state-of-the-art synthesis results, it operates directly in pixel space, which makes its optimization and inference computationally expensive. MagicPose [11] adopts a two-stage training strategy to separately handle character appearance control and pose control, disentangling the two processes. Additionally, it incorporates facial keypoints, finger keypoints, and sparse body keypoints to achieve more precise pose control. Although these methods have made promising progress, they condition the pose generation on the global features of the reference pose and delegate both the spatial transformation of the pose and appearance texture generation to the diffusion model, causing the diffusion model to handle two distinct tasks simultaneously. In this paper, we develop a novel paradigm for multi-pose generation that disentangles pose control from appearance guidance. This approach allows the model to focus on a single objective within each module, thereby improving the consistency and controllability of the generated human pose images.", + "bbox": [ + 89, + 90, + 483, + 607 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3. Method", + "text_level": 1, + "bbox": [ + 89, + 623, + 181, + 638 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3.1. Architecture and Overview", + "text_level": 1, + "bbox": [ + 89, + 648, + 334, + 664 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Given a reference image $\\mathcal{X}_{ref}$ and multiple target poses $\\mathcal{P}^{1:N} = [p_1,p_2\\dots ,p_N]$ , where $N$ is the number of target poses. The objective of our method is to generate a series of realistic images $\\mathcal{Y}^{1:N} = [y_1,y_2,\\dots,y_N]$ while ensuring that both the poses and the appearance are consistent with the given poses $\\mathcal{P}^{1:N}$ and the reference image $\\mathcal{X}_{ref}$ , respectively. The overall pipeline of our approach is illustrated in Fig. 2. To address the distinct nature of spatial transformations in pose generation and the fine-grained synthesis of appearance textures, we propose disentangling these two processes. This separation enables independent optimization of pose control and appearance refinement, leading to more accurate and natural image generation. To implement this, we use a multi-stage image encoder, such as Swin Transformer [17], denoted as $\\mathcal{M}_{ref}$ , to embed the ref", + "bbox": [ + 89, + 674, + 483, + 901 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "erence image $\\mathcal{X}_{ref}$ into multi-scale appearance embeddings $\\mathcal{F}_{ref} = [f_1, f_2, f_3, f_4]$ . These embeddings allow the network to perform multi-scale appearance control. For pose control, we developed a lightweight pose control network $\\mathcal{M}_p$ , based on ResNet [18], which extracts the pose features $\\mathcal{F}_p^{1:N}$ from the multiple target poses.", + "bbox": [ + 511, + 90, + 903, + 181 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Next, more importantly, we designed the Global-aware Pose Generation module, which is trained to generate target pose embedding $\\mathcal{F}_{tp}^{1:N}$ . Unlike previous methods, we leverage the Global-aware Pose Generation module to perform spatial transformations of the pose while also encouraging the module to learn appearance priors from the reference image. This approach naturally embeds some appearance knowledge into the generated pose, simplifying subsequent appearance refinement or occluded region completion by the diffusion model. Additionally, to enhance the realism of textures, we employ a multi-scale Appearance Adapter (AD) module to provide multi-scale finer-grained appearance features to the U-net, further improving the quality of the generated images.", + "bbox": [ + 511, + 181, + 906, + 393 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3.2. Global-aware Pose Generation", + "text_level": 1, + "bbox": [ + 511, + 401, + 784, + 416 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Existing multi-person pose generation frameworks predominantly optimize individual target poses in isolation, which inherently limits their ability to model the spatiotemporal correlations among concurrent human motions. This limitation restricts these models' capacity to capture broader contextual information, potentially leading to inconsistencies or unnatural transitions between generated poses. To address this, we propose a unified module for multi-pose generation (Fig. 2) that leverages the Transformer architecture for its effective modeling capabilities, enabling coherent pose generation. Central to this module is the Global-aware Transformer Decoder (GTD) Block, which iteratively generates each target pose by integrating information from both the reference image and diverse target poses. Using the reference image features $f_{4}$ from $\\mathcal{X}_{ref}$ and the $N$ -th target pose embeddings from $\\mathcal{P}^{N}$ , the GTD Block outputs the target pose $\\mathcal{F}_{tp}^{N}$ for the $N$ -th pose. To ensure coherence across poses, we enhance cross-pose consistency by incorporating previously generated poses (i.e., $\\mathcal{F}_{tp}^{1:N-1}$ in Eq. 1) into each subsequent generation step. The iterative process for generating multiple pose $\\mathcal{F}_{tp}^{1:N}$ and the specific generation for the $N$ -th GTD iteration are defined as follows:", + "bbox": [ + 511, + 424, + 906, + 755 + ], + "page_idx": 2 + }, + { + "type": "equation", + "text": "\n$$\n\\begin{array}{l} \\mathcal {F} _ {t p} ^ {1: N} = \\varphi (f _ {4}, \\mathcal {F} _ {p} ^ {1: N}), \\\\ \\mathcal {F} _ {t p} ^ {1: N} = \\mathcal {F} _ {t p} ^ {1: N} + \\frac {1}{N} \\sum_ {i = 1} ^ {N} \\mathcal {F} _ {t p} ^ {i}, \\\\ \\mathcal {F} _ {t p} ^ {N} = G T D \\left(f _ {4}, \\mathcal {F} _ {p} ^ {1: N}, \\mathcal {F} _ {t p} ^ {1: N - 1}\\right), \\tag {1} \\\\ \\end{array}\n$$\n", + "text_format": "latex", + "bbox": [ + 589, + 762, + 903, + 849 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "where $\\varphi$ represents the global-aware pose generation function, and $\\mathcal{F}_{tp}^{1:N - 1}$ is the generated result from the previous $N - 1$ iterations of GTD. To further reinforce contextual", + "bbox": [ + 511, + 854, + 906, + 900 + ], + "page_idx": 2 + }, + { + "type": "page_number", + "text": "5648", + "bbox": [ + 482, + 944, + 514, + 955 + ], + "page_idx": 2 + }, + { + "type": "image", + "img_path": "images/ac61b8f68c98a2ff1c8b76e777bf3a308e8081337e6061079e0e416f5e096b74.jpg", + "image_caption": [ + "Figure 2. Overview of the architecture of the proposed pipeline. The Image Encoder encodes a reference image to obtain multi-scale features, while the Pose Encoder encodes multiple target poses. Pose control is achieved by the Global-aware Pose Generation module, which iteratively generates pose embeddings from the final layer features of the image encoder and the target pose features. A diffusion model then refines the pose appearance, conditioned on these pose embeddings. The Appearance Adapter further enhances appearance detail by refining the multi-scale features from the image encoder and injecting them into the diffusion model." + ], + "image_footnote": [], + "bbox": [ + 94, + 77, + 906, + 435 + ], + "page_idx": 3 + }, + { + "type": "image", + "img_path": "images/9f4b059a545121584618136c86f900160cea41e876dbbe718ae242e68a1c077c.jpg", + "image_caption": [ + "Figure 3. Comparison of the proposed Global-aware Transformer Decoder architecture (c) with two alternative decoder architectures (a) and (b). For clarity, we have omitted the residual connections in the cross-attention and self-attention layers." + ], + "image_footnote": [], + "bbox": [ + 94, + 530, + 480, + 659 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "consistency, we include a pooling layer to capture global pose features, strengthening the model's understanding of overall pose structure across generated poses.", + "bbox": [ + 89, + 744, + 483, + 790 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Global-aware Transformer Decoder. For efficient target pose generation, we propose a Transformer-based decoder architecture capable of iterative pose generation. In this design, we developed three distinct Transformer decoder architectures, as shown in detail in Fig. 3. The specifics of these architectures will be discussed in the ablation experiments and the Appendix. Unlike traditional convolu", + "bbox": [ + 89, + 795, + 483, + 900 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "tional neural networks (CNNs) commonly applied in visual tasks, attention-based Transformers lack certain inductive biases inherent to conventional visual models [19]. To improve the decoder's ability to incorporate pose-specific inductive biases, we introduce a global-aware block that models the entire pose holistically. This global-aware block is a flexible, plug-and-play addition compatible with any Transformer architecture. Distinct from the decoder architectures (a) and (b), which employ cross-attention layers to capture semantic relationships. Specifically, we enhance the target pose feature $\\mathcal{F}_p^N$ by computing a similarity matrix between $\\mathcal{F}_p^N$ and $\\mathcal{F}_{tp}^{1:N - 1}$ , which integrates relevant information from prior poses, ensuring consistency within the generated poses. Next, a self-attention layer and feed-forward network (FFN) further refine $\\mathcal{F}_p^N$ to enhance its expressiveness. The global-aware block is formalized as follows:", + "bbox": [ + 511, + 532, + 906, + 776 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {F} _ {o u t} = \\phi (\\mathcal {F} _ {p} ^ {N} \\cdot \\mathcal {F} _ {t p} ^ {N - 1 ^ {T}}) \\cdot \\mathcal {F} _ {t p} ^ {N - 1} + \\mathcal {F} _ {p} ^ {N},\n$$\n", + "text_format": "latex", + "bbox": [ + 570, + 786, + 846, + 809 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {F} _ {o u t} = \\mathrm {M H S A} (\\tilde {\\mathcal {F}} _ {o u t}) + \\mathcal {F} _ {o u t},\n$$\n", + "text_format": "latex", + "bbox": [ + 571, + 811, + 777, + 829 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {F} _ {\\text {o u t}} = \\operatorname {F F N} \\left(\\hat {\\mathcal {F}} _ {\\text {o u t}}\\right) + \\mathcal {F} _ {\\text {o u t}}, \\tag {2}\n$$\n", + "text_format": "latex", + "bbox": [ + 571, + 833, + 906, + 849 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "where $\\phi = \\frac{e^{s_{ij}}}{\\sum_k e^{s_{ik}}}$ aggregates pose information. Here, MHSA denotes the multi-head self-attention layer. Finally,", + "bbox": [ + 511, + 862, + 905, + 900 + ], + "page_idx": 3 + }, + { + "type": "page_number", + "text": "5649", + "bbox": [ + 482, + 944, + 514, + 955 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "the reference image features $\\mathcal{F}_{ref}$ interact with the globally enhanced target pose features $\\mathcal{F}_{out}$ through a series of attention-based modules, producing the pose used as a conditional embedding input for the U-Net.", + "bbox": [ + 89, + 90, + 483, + 151 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "3.3. Fine-grained Appearance Generation", + "text_level": 1, + "bbox": [ + 89, + 160, + 415, + 176 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Latent Diffusion Model. Latent diffusion models (LDMs) [20] have recently gained prominence as robust alternatives to generative adversarial networks (GANs) for various generative tasks. Building upon the foundational Markov chain framework of DDPM [9], LDMs adhere to a dual-phase paradigm comprising progressive noise injection and learned reconstruction. As formalized in [20], the forward diffusion operator systematically perturbs input data through a predefined variance schedule:", + "bbox": [ + 89, + 183, + 483, + 319 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\nq \\left(z _ {t} \\mid z _ {t - 1}\\right) = \\mathcal {N} \\left(z _ {t}; \\sqrt {1 - \\beta_ {t}} z _ {t - 1}, \\beta_ {t} I\\right), \\tag {3}\n$$\n", + "text_format": "latex", + "bbox": [ + 143, + 330, + 482, + 356 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "which transforms the latent distribution of the data $z_0$ into a noise distribution $z_{t}$ (e.g., Gaussian distribution). The Gaussian variance, $\\beta_{t}$ , can be learned or derived from a predefined schedule. In the reverse diffusion process, LDMs learn to predict and remove noise, effectively reversing the transformation to reconstruct the original data distribution:", + "bbox": [ + 89, + 363, + 483, + 455 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\np _ {\\theta} \\left(z _ {t - 1} \\mid z _ {t}\\right) = \\mathcal {N} \\left(z _ {t - 1}; \\mu_ {\\theta} \\left(z _ {t}, t, c\\right), \\sigma^ {2} I\\right), \\tag {4}\n$$\n", + "text_format": "latex", + "bbox": [ + 138, + 465, + 482, + 483 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "where $\\mu_{\\theta}$ denotes the learned mean parameterized by $\\theta$ as a function of the current noise state $z_{t}$ , time step $t$ , and condition $c$ . A more detailed formulation of diffusion models is provided in the Appendix.", + "bbox": [ + 89, + 493, + 483, + 553 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "In this paper, we utilize the powerful LDM model to generate human appearances. To achieve appearance guidance, we disentangle pose control and appearance guidance. Unlike prior approaches [11, 21], which rely on CLIP's [22] image encoder or a U-Net-like appearance encoder, we condition the diffusion model on multi-pose embeddings generated by the Global-aware Pose Generation module. Additionally, to ensure coherence across the poses and to generate the entire sequence simultaneously, we extend the 2D diffusion model to a 3D setting, inspired by [21]. For improved efficiency, we fine-tune only selected weights in the U-Net, as shown in Fig. 2. Here, the \"lock\" symbol denotes frozen weights, while the \"flame\" symbol indicates fine-tuned parameters.", + "bbox": [ + 89, + 553, + 483, + 763 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Appearance Adapter. To enable precise control over the character's appearance details, we introduce the Appearance Adapter (AD), shown in Fig. 2. The AD operates hierarchically between multi-scale upsampling blocks in the diffusion framework, extracting pose-agnostic appearance features from reference images through cross-scale feature learning. This architecture enables explicit disentanglement of high-level semantic understanding between appearance attributes and pose informations.", + "bbox": [ + 89, + 763, + 483, + 901 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Specifically, our implementation with $\\mathrm{k} = 3$ Appearance Embedding (AE) layers is augmented by two convolutional layers for dimensional alignment. Adopting the zero-initialization strategy from [23], these convolutional layers progressively smooth feature representations while suppressing high-frequency noise, thereby allowing stable learning of fine-grained appearance patterns. The AE layers enhance feature refinement through global attention mechanisms, followed by a Feed-Forward Network (FFN) that non-linearly reinforces the discriminative power of the processed features.", + "bbox": [ + 511, + 90, + 906, + 257 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "3.4. Optimization", + "text_level": 1, + "bbox": [ + 513, + 267, + 653, + 282 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "We train our proposed model in an end-to-end manner. Our overall optimization objective is:", + "bbox": [ + 511, + 290, + 906, + 321 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {L} _ {\\text {t o t a l}} = \\mathcal {L} _ {m g} + \\mathcal {L} _ {\\text {r e c}}, \\tag {5}\n$$\n", + "text_format": "latex", + "bbox": [ + 633, + 333, + 903, + 349 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "where $\\mathcal{L}_{mg}$ is used to optimize the multi-pose generation process. The optimization objective is:", + "bbox": [ + 511, + 361, + 906, + 392 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\begin{array}{l} \\mathcal {L} _ {m g} = \\mathbb {E} _ {z _ {0} ^ {1: N}, \\mathcal {X} _ {r e f}, \\mathcal {P} ^ {1: N}} [ \\| \\epsilon - \\\\ \\epsilon_ {\\theta} \\left(z _ {t} ^ {1: N}, t, \\varphi \\left(\\mathcal {F} _ {r e f}, \\mathcal {F} _ {p} ^ {1: N}\\right)\\right) | | _ {2} ^ {2} ], \\tag {6} \\\\ \\end{array}\n$$\n", + "text_format": "latex", + "bbox": [ + 519, + 402, + 903, + 454 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "where $\\mathcal{X}_{ref}$ denote the reference image, with $\\mathcal{F}_{ref}$ representing its features. The target poses are given by $\\mathcal{P}^{1:N}$ , and $\\mathcal{F}_p^{1:N} = \\mathcal{M}_p(\\mathcal{P}^{1:N})$ represents the features of the target poses. Additionally, $z_0^{1:N} = \\mathcal{E}(\\mathcal{V}^{1:N})$ denotes the initial latent features, while $z_t^{1:N}$ is the noisy latent representation derived from $z_0^{1:N}$ at timestep $t$ , where $\\mathcal{E}$ is VAE encoder. In addition, inspired by [5], we introduced an additional reconstruction task for the reference image during training to ensure the consistency between the appearance of the generated target poses and the reference image. The optimization objective of the reconstruction task is:", + "bbox": [ + 511, + 455, + 906, + 623 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {L} _ {r e c} = \\mathbb {E} _ {z _ {0}, \\chi_ {r e f}, \\mathcal {P} ^ {0}} [ | | \\epsilon - \\epsilon_ {\\theta} (z _ {t}, t, \\mathcal {F} _ {r e f})) | | _ {2} ^ {2} ]. \\qquad (7)\n$$\n", + "text_format": "latex", + "bbox": [ + 558, + 633, + 906, + 652 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Sampling. At inference time, we sample $z_{t}$ from the Gaussian noise and iteratively sample from the previous step until $z_{0}$ , then we decode $z_{0}$ to image space get the final human pose image. To improve the sample quality during the generation process, we use a classifier-free guidance strategy. To implement classifier-free guidance, we randomly drop the reference image used as the appearance condition with a probability of $5\\%$ during training.", + "bbox": [ + 511, + 662, + 906, + 785 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\hat {\\epsilon} _ {\\theta} \\left(z _ {t}, t, \\varphi\\right) = \\lambda \\epsilon_ {\\theta} \\left(z _ {t}, t, \\varphi\\right) + (1 - \\lambda) \\epsilon_ {\\theta} \\left(z _ {t}, t\\right), \\tag {8}\n$$\n", + "text_format": "latex", + "bbox": [ + 540, + 797, + 903, + 813 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "where $\\lambda$ is a hyperparameter used to control the strength of guidance during the inference process, while $\\epsilon_{\\theta}(z_t,t,\\varphi)$ and $\\epsilon_{\\theta}(z_t,t)$ represent conditional prediction and unconditional prediction, respectively. To handle the case when the reference image is missing, inspired by [10], we use a", + "bbox": [ + 511, + 825, + 906, + 901 + ], + "page_idx": 4 + }, + { + "type": "page_number", + "text": "5650", + "bbox": [ + 482, + 944, + 514, + 955 + ], + "page_idx": 4 + }, + { + "type": "table", + "img_path": "images/6535856aab0e335496b4c906b030c727deabe9d69e815d26e06e0c29adb786b5.jpg", + "table_caption": [], + "table_footnote": [], + "table_body": "
ApproachesUBC FashionTikTok
SSIM ↑PSNR ↑FID ↓LPIPS ↓\\( L_1 \\) ↓SSIM ↑PSNR ↑FID ↓LPIPS ↓\\( L_1 \\) ↓
PISE [8] (CVPR'21)0.86717.11618.6570.07520.0460.4299.947120.6480.30100.186
DPTN [5] (CVPR'22)0.83315.93719.9570.08760.0500.1885.696143.1100.45170.449
NTED [7] (CVPR'22)0.82415.68319.7840.09480.0530.3398.551140.3030.37270.228
BiGraphGAN [4] (IJCV'23)0.84416.40121.0120.08060.0470.4459.916147.5210.26760.177
PIDM [12] (CVPR'23)0.81218.87014.9510.16350.0920.3077.558130.3920.42770.262
CFLD [10] (CVPR'24)0.88922.14212.8100.07180.0260.72917.520105.6310.21360.078
MagicPose [11] (ICML'24)0.83120.74214.6490.08660.0300.67717.260108.0670.25070.083
Ours0.94125.65511.4870.04970.0140.75318.466103.7080.23110.072
", + "bbox": [ + 96, + 77, + 901, + 224 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/350dbc4f2801db176651fffe8185f801abfb2c9134521767c98eff0684cc875f.jpg", + "image_caption": [ + "Figure 4. The pose synthesis accuracy of the proposed method on the UBC Fashion [13] dataset as the number of generated images increases, with 8 as the training setting." + ], + "image_footnote": [], + "bbox": [ + 93, + 268, + 475, + 357 + ], + "page_idx": 5 + }, + { + "type": "table", + "img_path": "images/a95065c579c03055c2ca7f4dda417482686f35952a6f481f4603fca0e7e5aaec.jpg", + "table_caption": [ + "Table 1. Quantitative comparison of pose sequence synthesis quality on the UBC Fashion [13], and TikTok [14] dataset." + ], + "table_footnote": [], + "table_body": "
MethodsPIDM [12] (CVPR'23)CFLD [10] (CVPR'24)MagicPose [11] (ICML'24)Ours(8-24)
#Param131M248M2081M367M
Inference Time28.64s35.8s67.8s5.5-24.1s
", + "bbox": [ + 94, + 424, + 478, + 481 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Table 2. Comparison of parameter count and inference time between the proposed method and other SOTA diffusion models.", + "bbox": [ + 89, + 487, + 482, + 516 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "learnable vector as the conditional embedding. We concatenate the learnable vector with the reference image features to form a compact conditional embedding for the generation process, encouraging the learnable vector to capture more information beyond the reference image during training. This allows the model to receive additional appearance information when predicting invisible regions.", + "bbox": [ + 89, + 521, + 483, + 628 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "4. Experiments", + "text_level": 1, + "bbox": [ + 89, + 641, + 223, + 657 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "4.1. Setup", + "text_level": 1, + "bbox": [ + 89, + 665, + 171, + 681 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Dataset. We validate the effectiveness of our proposed method on both the UBC Fashion [13] and TikTok [14] datasets. The UBC Fashion dataset is a collection of data featuring various poses and outfits, containing 500 training and 100 test videos. Compared to UBC Fashion, the TikTok dataset has more complex backgrounds and greater diversity in the proportions of people, making the task more challenging. The TikTok dataset comprises 350 dance videos featuring individual performers, each lasting between 10 and 15 seconds. To ensure a fair comparison with state-of-the-art methods, we use the same test set as [11] for evaluation. We set the image resolution to $256 \\times 256$ and randomly sample $N$ frames (in this paper, $N$ is set to 8) from each training video for training. Additionally, we use Open-", + "bbox": [ + 89, + 688, + 483, + 900 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Pose [24] and PGN [25] to extract human poses and human parsing maps. Due to category redundancy in the human parsing maps, we reduce the number of categories from 20 to 8 as [8].", + "bbox": [ + 511, + 270, + 906, + 330 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Metrics. To evaluate the quality of the generated pose sequences, we report several metrics: Structural Similarity Index Measure (SSIM) [26], Peak Signal to Noise Ratio (PSNR), Learned Perceptual Image Patch Similarity (LPIPS) [27], and Fréchet Inception Distance (FID) [28]. SSIM and PSNR assess pixel-level similarity by measuring statistical differences between the generated images and the ground truth (GT). However, these metrics are limited to low-level image properties and may not fully capture perceptual quality. To address this limitation, we also include LPIPS and FID, which evaluate deep feature-based similarities. FID evaluates the distribution differences between the generated images and GT using Inception-v3 [29] features, while LPIPS measures the perceptual similarity between images using a VGG [30] network, making it more aligned with human perception of image similarity.", + "bbox": [ + 511, + 343, + 906, + 587 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Implementation details. The proposed method is implemented based on PyTorch [31] and Diffusers [32], trained in an end-to-end manner with a batch size of 16 across 4 NVIDIA A100 GPUs. The model was trained for 100 epochs utilizing the Adam [33] optimizer, starting with an initial learning rate of $5 \\times 10^{-5}$ . A learning rate warmup strategy was implemented for the first 1000 training steps, followed by a decay to $1 \\times 10^{-5}$ at the 50th epoch. To accelerate model convergence, we initialize the image encoder using Swin-B [17] pretrained on ImageNet [34] and the U-net using version 1.5 of Stable Diffusion [20]. Additionally, to address the issue of poor facial realism in generated pose, we followed strategy [20] by fine-tuning the decoder of the VAE using both MSE reconstruction loss and perceptual loss [35]. During the fine-tuning process, we placed greater emphasis on the MSE reconstruction loss. Further details on the fine-tuning process can be found in the Appendix. For classifier-free guidance, we set the parameter $\\lambda$ to 2.0 during sampling, and the reference condition $\\mathcal{X}_{ref}$ was dropped with a probability of $5\\%$ throughout training.", + "bbox": [ + 511, + 598, + 908, + 902 + ], + "page_idx": 5 + }, + { + "type": "page_number", + "text": "5651", + "bbox": [ + 482, + 945, + 513, + 955 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/940da2dad2d15f7a32bc66e786cb4110fe0a69046f25020bc1e07820d96020fc.jpg", + "image_caption": [ + "Figure 5. Qualitative comparison of our method with SOTA methods on UBC Fashion [13] and TikTok[14] dataset. Our approach can generate more realistic details consistent with the reference image, such as garment folds, fine-grained textures, and facial identity features." + ], + "image_footnote": [], + "bbox": [ + 99, + 78, + 263, + 481 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/16720a2c89453f62e1f068e850405c517837a50e90b36ae484ed2234a30ec063.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 267, + 78, + 341, + 479 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/f255f3d716debeb8e1f22c3b7641bf930997bdde57013ac8a2d5c77c8cb8b3ca.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 343, + 78, + 496, + 479 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/193a2c86f2ef0f2099d29c0557c1fb5091c6a78d8ca6cdea9ea36ed502c25063.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 500, + 78, + 653, + 479 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/c6361b96483922b507213ca94bbbd03bcafc642cfa6517337a5e80f5af8a2238.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 656, + 78, + 733, + 479 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/ff74e153d5ac47b00c93ee7b9674ff31c1dba7812ca51c67387f32c2c8d2d400.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 736, + 78, + 810, + 479 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/5016a54b521fd3e4d512c5dcd7604d75f527278c4b9f963211e847fedc2915f2.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 813, + 79, + 900, + 479 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "4.2. Quantitative Comparison", + "text_level": 1, + "bbox": [ + 89, + 542, + 323, + 559 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "We conducted a quantitative comparison of our method against state-of-the-art GAN-based and diffusion-based models on the UBC Fashion [13] and TikTok [14] datasets, as shown in Tab. 1. On the majority of metrics, our method outperforms the advanced methods on both datasets. Specifically, compared with four other diffusion models [10-12, 36], our approach achieves more stable generation performance and higher efficiency, with specific results shown in the Tab. 2, owing to the disentangling of pose control from appearance generation. Additionally, this design enables us to generate more human poses during inference (up to 3 times the training quantity) without additional training. The Fig. 4 illustrates the performance of our method as the generation count gradually increases. As shown in this figure, although an increase in generation quantity may lead to a slight performance decline, the results still serve as acceptable human image outputs, outperforming other SOTA methods.", + "bbox": [ + 89, + 566, + 483, + 838 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "4.3. Qualitative Comparison", + "text_level": 1, + "bbox": [ + 89, + 847, + 312, + 864 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "The qualitative comparison with state-of-the-art methods is illustrated in Fig. 5 and 6. When using these other meth", + "bbox": [ + 89, + 869, + 483, + 901 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "ods, both GAN-based and diffusion-based approaches show overfitting issues, resulting in the loss of person identity and clothing texture details. Additionally, when generating more complex clothing or unseen regions, issues like clothing blurriness, unnatural texture transitions, and substantial differences from the source image become apparent. Benefiting from our proposed strategy of disentangling pose control and appearance generation, our method mitigates overfitting to some extent. Furthermore, with the high-consistency pose embedding provided by Global-aware Pose Generation and the fine-grained, multi-scale features from Appearance Adapter, our approach yields more realistic results in maintaining person identity and capturing intricate clothing patterns. In particular, Fig. 6 provides an additional qualitative comparison with existing SOTA diffusion models for generating multiple poses. This scenario places higher demands on cross-view consistency in person identity and appearance. As observed in the Fig. 6, our method achieves higher consistency across different viewpoints of generated person images, thanks to Global-aware Transformer Decoder's global pose awareness in the pose generation process. More qualitative comparisons can be found in the Appendix.", + "bbox": [ + 511, + 544, + 906, + 892 + ], + "page_idx": 6 + }, + { + "type": "page_number", + "text": "5652", + "bbox": [ + 482, + 944, + 514, + 955 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/88735d997250453678ac179b5f765a723919163e1ab74e3f11e0c560697d60d4.jpg", + "image_caption": [ + "Figure 6. Qualitative comparison with other SOTA methods on various poses, with drawbacks highlighted by red circles." + ], + "image_footnote": [], + "bbox": [ + 93, + 70, + 480, + 397 + ], + "page_idx": 7 + }, + { + "type": "table", + "img_path": "images/4d3bdde5d99a998ffea95294ad1d45edd573c8feb1c94b2b2cf3eb3792ce234f.jpg", + "table_caption": [], + "table_footnote": [], + "table_body": "
ModelSSIM ↑PSNR ↑FID ↓LPIPS ↓
w/o Global-aware Pose Generation (GPG)0.936225.02411.8650.0532
w/o Multi-stage Image Encoder (MIE)0.934424.89511.2840.0554
w/o Appearance Adapter (AD)0.938425.20511.9340.0517
w/o Fine-tune VAE (FVAE)0.938625.46411.6270.0514
w/o Lrec0.940325.53911.5470.0516
Ours + (a)0.938925.44311.4870.0503
Ours + (b)0.938125.44711.5720.0506
Ours0.941125.65511.4870.0497
", + "bbox": [ + 93, + 448, + 480, + 556 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "4.4. Ablation Study", + "text_level": 1, + "bbox": [ + 89, + 604, + 243, + 622 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Tab. 3 and Fig. 7 show quantitative and qualitative results, respectively. It can be observed that both GPG and AD are essential for pose synthesis performance, with their absence potentially leading to a performance drop of up to $3.8\\%$ (FID). This highlights the importance of separating pose control and appearance generation strategies. We also replaced MIE with a CLIP image encoder, resulting in a $0.7\\%$ (SSIM) and $2.9\\%$ (PSNR) decline, though with an improvement in FID performance. As shown in the first row of Fig. 7, the CLIP image encoder fails to retain intricate clothing details, leading to a loss of appearance-related information. Additionally, fine-tuning the VAE for the specific dataset helps preserve realistic details of faces and clothing to a certain degree. Finally, we also experimented with two alternative Transformer decoder architectures in Fig. 3, both of which led to varying degrees of performance decline. Detailed ablation experiments can be found in the Appendix.", + "bbox": [ + 88, + 628, + 485, + 887 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/2eeb1baa38a4468e016a694a995f14930a2fabb6846ce9ffc262718546e84fac.jpg", + "image_caption": [ + "Figure 7. Qualitative ablation study on the UBC Fashion [13] dataset." + ], + "image_footnote": [], + "bbox": [ + 521, + 70, + 897, + 218 + ], + "page_idx": 7 + }, + { + "type": "table", + "img_path": "images/c2f6c121f82d6ec19683e6849f5db647983e8031855facbc1386971f451896f3.jpg", + "table_caption": [ + "Table 3. Ablation study of different components, where (a) and (b) represent different decoder architectures in Fig. 3." + ], + "table_footnote": [], + "table_body": "
MethodsPIDM [12] (CVPR'23)CFLD [10] (CVPR'24)MagicPose [11] (ICML'24)Ours
Average13.6%16.7%8.3%61.4%
", + "bbox": [ + 517, + 258, + 903, + 311 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Table 4. User study on the UBC Fashion [13] dataset. We followed [11] to collect vote data from 30 volunteers. The participants believed that our proposed method can generate visually realistic and identity-consistent human poses.", + "bbox": [ + 511, + 316, + 906, + 375 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "4.5. User Study", + "text_level": 1, + "bbox": [ + 511, + 382, + 635, + 398 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "To more accurately evaluate the effectiveness of our proposed method, we designed a user study. In this experiment, we collected images of 9 subjects from the test set for comparative evaluation. One subject was designated as a warm-up, while the remaining eight were used as formal test. For each subject, eight distinct pose images were randomly selected to comparison. We presented images generated by various methods [10-12] to users, instructing them to evaluate the images based on visual quality, pose accuracy, and detail consistency, and to select the method they preferred. We present the averaged vote result in Tab. 4.", + "bbox": [ + 511, + 406, + 906, + 571 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "5. Conclusions", + "text_level": 1, + "bbox": [ + 511, + 587, + 640, + 603 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "In this paper, we propose a unified framework for diverse multi-pose generation driven by disentangled pose and appearance guidance. This framework separately addresses complex human pose control and appearance generation to produce realistic and varied human poses. To achieve precise pose control and realistic appearance generation, we introduce Global-aware Pose Generation and Appearance Adapter module. These modules enhance interactions among poses during the generation process, leveraging all available mutual information to provide global guidance, while also learning high-level semantic information from multi-scale features to preserve texture details as much as possible. Extensive experiments demonstrate that our proposed approach achieves significant advantages over other state-of-the-art methods.", + "bbox": [ + 511, + 613, + 906, + 839 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Acknowledgments. This work was supported by the National Natural Science Foundation of China (62036006, 62276200) and the Tianjin Natural Science Foundation (24JCYBJC00310).", + "bbox": [ + 511, + 839, + 908, + 900 + ], + "page_idx": 7 + }, + { + "type": "page_number", + "text": "5653", + "bbox": [ + 482, + 944, + 514, + 955 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "References", + "text_level": 1, + "bbox": [ + 91, + 90, + 187, + 104 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[1] Grigore C Burdea and Philippe Coiffet. *Virtual reality technology*. John Wiley & Sons, 2003. 2", + "[2] Julie Carmigniani, Borko Furht, Marco Anisetti, Paolo Ceravolo, Ernesto Damiani, and Misa Ivkovic. Augmented reality technologies, systems and applications. Multimedia tools and applications, 51:341-377, 2011. 2", + "[3] Liang Zheng, Liyue Shen, Lu Tian, Shengjin Wang, Jingdong Wang, and Qi Tian. Scalable person re-identification: A benchmark. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 1116-1124, 2015. 2", + "[4] Hao Tang, Ling Shao, Philip HS Torr, and Nicu Sebe. Bipartite graph reasoning gans for person pose and facial image synthesis. International Journal of Computer Vision, 131(3):644-658, 2023. 2, 6", + "[5] Pengze Zhang, Lingxiao Yang, Jian-Huang Lai, and Xiaohua Xie. Exploring dual-task correlation for pose guided person image generation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 7713-7722, 2022. 2, 5, 6", + "[6] Yuelong Li, Tengfei Xiao, Lei Geng, and Jianming Wang. Direct may not be the best: An incremental evolution view of pose generation. In Proceedings of the AAAI Conference on Artificial Intelligence, volume 38, pages 3270-3278, 2024. 2", + "[7] Yurui Ren, Xiaqing Fan, Ge Li, Shan Liu, and Thomas H Li. Neural texture extraction and distribution for controllable person image synthesis. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 13535-13544, 2022. 2, 6", + "[8] Jinsong Zhang, Kun Li, Yu-Kun Lai, and Jingyu Yang. Pise: Person image synthesis and editing with decoupled gan. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 7982-7990, 2021. 2, 6", + "[9] Jonathan Ho, Ajay Jain, and Pieter Abbeel. Denoising diffusion probabilistic models. Advances in Neural Information Processing Systems, 33:6840-6851, 2020. 2, 3, 5", + "[10] Yanzuo Lu, Manlin Zhang, Andy J Ma, Xiaohua Xie, and Jianhuang Lai. Coarse-to-fine latent diffusion for pose-guided person image synthesis. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 6420-6429, 2024. 2, 5, 6, 7, 8", + "[11] Di Chang, Yichun Shi, Quankai Gao, Hongyi Xu, Jessica Fu, Guoxian Song, Qing Yan, Yizhe Zhu, Xiao Yang, and Mohammad Soleymani. Magicpose: Realistic human poses and facial expressions retargeting with identity-aware diffusion. In Proceedings of the International Conference on Machine Learning, 2023. 3, 5, 6, 8", + "[12] Ankan Kumar Bhunia, Salman Khan, Hisham Cholakkal, Rao Muhammad Anwer, Jorma Laaksonen, Mubarak Shah, and Fahad Shahbaz Khan. Person image synthesis via denoising diffusion model. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 5968-5976, 2023. 2, 3, 6, 7, 8", + "[13] Polina Zablotskaia, Aliaksandr Siarohin, Bo Zhao, and Leonid Sigal. Dwnet: Dense warp-based network for" + ], + "bbox": [ + 93, + 114, + 482, + 900 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "pose-guided human video generation. arXiv preprint arXiv:1910.09139, 2019. 2, 6, 7, 8", + "[14] Yasamin Jafarian and Hyun Soo Park. Learning high fidelity depths of dressed humans by watching social media dance videos. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 12753-12762, 2021. 2, 6, 7", + "[15] Ian Goodfellow, Jean Pouget-Abadie, Mehdi Mirza, Bing Xu, David Warde-Farley, Sherjil Ozair, Aaron Courville, and Yoshua Bengio. Generative adversarial networks. Communications of the ACM, 63(11):139-144, 2020. 2", + "[16] Jonathan Ho and Tim Salimans. Classifier-free diffusion guidance. In NeurIPS 2021 Workshop on Deep Generative Models and Downstream Applications, 2021. 3", + "[17] Ze Liu, Yutong Lin, Yue Cao, Han Hu, Yixuan Wei, Zheng Zhang, Stephen Lin, and Baining Guo. Swin transformer: Hierarchical vision transformer using shifted windows. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 10012-10022, 2021. 3, 6", + "[18] Kaiming He, Xiangyu Zhang, Shaoqing Ren, and Jian Sun. Deep residual learning for image recognition. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 770-778, 2016. 3", + "[19] Xumin Yu, Yongming Rao, Ziyi Wang, Zuyan Liu, Jiwen Lu, and Jie Zhou. Pointr: Diverse point cloud completion with geometry-aware transformers. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 12498-12507, 2021. 4", + "[20] Robin Rombach, Andreas Blattmann, Dominik Lorenz, Patrick Esser, and Björn Ommer. High-resolution image synthesis with latent diffusion models. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 10684-10695, 2022. 5, 6", + "[21] Zhongcong Xu, Jianfeng Zhang, Jun Hao Liew, Hanshu Yan, Jia-Wei Liu, Chenxu Zhang, Jiashi Feng, and Mike Zheng Shou. Magicanimate: Temporally consistent human image animation using diffusion model. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 1481–1490, 2024. 5", + "[22] Alec Radford, Jong Wook Kim, Chris Hallacy, Aditya Ramesh, Gabriel Goh, Sandhini Agarwal, Girish Sastry, Amanda Askell, Pamela Mishkin, Jack Clark, et al. Learning transferable visual models from natural language supervision. In Proceedings of the International Conference on Machine Learning, pages 8748-8763. PMLR, 2021. 5", + "[23] Lvmin Zhang, Anyi Rao, and Maneesh Agrawala. Adding conditional control to text-to-image diffusion models. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 3836-3847, 2023. 5", + "[24] Riza Alp Güler, Natalia Neverova, and Iasonas Kokkinos. Densesepose: Dense human pose estimation in the wild. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 7297-7306, 2018. 6", + "[25] Ke Gong, Xiaodan Liang, Yicheng Li, Yimin Chen, Ming Yang, and Liang Lin. Instance-level human parsing via part grouping network. In Proceedings of the European Conference on Computer Vision, pages 770-785, 2018. 6" + ], + "bbox": [ + 517, + 93, + 906, + 900 + ], + "page_idx": 8 + }, + { + "type": "page_number", + "text": "5654", + "bbox": [ + 482, + 944, + 514, + 955 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[26] Zhou Wang, Alan C Bovik, Hamid R Sheikh, and Eero P Simoncelli. Image quality assessment: from error visibility to structural similarity. IEEE transactions on image processing, 13(4):600-612, 2004. 6", + "[27] Richard Zhang, Phillip Isola, Alexei A Efros, Eli Shechtman, and Oliver Wang. The unreasonable effectiveness of deep features as a perceptual metric. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 586-595, 2018. 6", + "[28] Martin Heusel, Hubert Ramsauer, Thomas Unterthiner, Bernhard Nessler, and Sepp Hochreiter. Gans trained by a two time-scale update rule converge to a local nash equilibrium. Advances in Neural Information Processing Systems, 30, 2017. 6", + "[29] Christian Szegedy, Vincent Vanhoucke, Sergey Ioffe, Jon Shlens, and Zbigniew Wojna. Rethinking the inception architecture for computer vision. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 2818-2826, 2016. 6", + "[30] Karen Simonyan and Andrew Zisserman. Very deep convolutional networks for large-scale image recognition. arXiv preprint arXiv:1409.1556, 2014. 6", + "[31] Adam Paszke, Sam Gross, Francisco Massa, Adam Lerer, James Bradbury, Gregory Chanan, Trevor Killeen, Zeming Lin, Natalia Gimelshein, Luca Antiga, et al. Pytorch: An imperative style, high-performance deep learning library. Advances in Neural Information Processing Systems, 32, 2019. 6", + "[32] Patrick von Platen, Suraj Patil, Anton Lozhkov, Pedro Cuenca, Nathan Lambert, Kashif Rasul, Mishig Davaadorj, Dhruv Nair, Sayak Paul, William Berman, Yiyi Xu, Steven Liu, and Thomas Wolf. Diffusers: State-of-the-art diffusion models. https://github.com/huggingface/diffusers, 2022.6", + "[33] Diederik P Kingma. Adam: A method for stochastic optimization. arXiv preprint arXiv:1412.6980, 2014. 6", + "[34] Jia Deng, Wei Dong, Richard Socher, Li-Jia Li, Kai Li, and Li Fei-Fei. Imagenet: A large-scale hierarchical image database. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition, pages 248–255. IEEE, 2009. 6", + "[35] Justin Johnson, Alexandre Alahi, and Li Fei-Fei. Perceptual losses for real-time style transfer and super-resolution. In Computer Vision-ECCV 2016: 14th European Conference, Amsterdam, The Netherlands, October 11-14, 2016, Proceedings, Part II 14, pages 694-711. Springer, 2016. 6", + "[36] Tan Wang, Linjie Li, Kevin Lin, Yuanhao Zhai, Chung-Ching Lin, Zhengyuan Yang, Hanwang Zhang, Zicheng Liu, and Lijuan Wang. Disco: Disentangled control for realistic human dance generation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 9326-9336, 2024. 7" + ], + "bbox": [ + 91, + 92, + 482, + 824 + ], + "page_idx": 9 + }, + { + "type": "page_number", + "text": "5655", + "bbox": [ + 482, + 945, + 514, + 955 + ], + "page_idx": 9 + } +] \ No newline at end of file diff --git a/2025/Disentangled Pose and Appearance Guidance for Multi-Pose Generation/eaadccfc-d365-4d2d-83e8-71935f857eca_model.json b/2025/Disentangled Pose and Appearance Guidance for Multi-Pose Generation/eaadccfc-d365-4d2d-83e8-71935f857eca_model.json new file mode 100644 index 0000000000000000000000000000000000000000..1cf8651389f1b2e0b295a4dd3aabf40f87655a7d --- /dev/null +++ b/2025/Disentangled Pose and Appearance Guidance for Multi-Pose Generation/eaadccfc-d365-4d2d-83e8-71935f857eca_model.json @@ -0,0 +1,1782 @@ +[ + [ + { + "type": "header", + "bbox": [ + 0.107, + 0.003, + 0.182, + 0.043 + ], + "angle": 0, + "content": "CVF" + }, + { + "type": "header", + "bbox": [ + 0.237, + 0.0, + 0.812, + 0.047 + ], + "angle": 0, + "content": "This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore." + }, + { + "type": "title", + "bbox": [ + 0.135, + 0.131, + 0.865, + 0.154 + ], + "angle": 0, + "content": "Disentangled Pose and Appearance Guidance for Multi-Pose Generation" + }, + { + "type": "text", + "bbox": [ + 0.102, + 0.179, + 0.896, + 0.306 + ], + "angle": 0, + "content": "Tengfei Xiao\\(^{1,4}\\), Yue Wu\\(^{\\dagger,*,1,4}\\), Yuelong Li\\(^{\\dagger,*,2,5}\\), Can Qin\\(^{3}\\), Maoguo Gong\\(^{1,4}\\), Qiguang Miao\\(^{1,4}\\), Wenping Ma\\(^{1}\\) \n\\(^{1}\\)Xidian University \\(^{2}\\)Tiangong University \\(^{3}\\)Northeastern University \\(^{4}\\)MoE Key Lab of Collaborative Intelligence Systems, Xidian University \n\\(^{5}\\)Tianjin Key Laboratory of Autonomous Intelligence Technology and Systems, Tiangong University tfxiao@stu.xidian.edu.cn, {ywu, qgmiao}@xidian.edu.cn, liyuelong@pku.edu.cn, qin.ca@northeastern.edu, gong@ieee.org, wpma@mail.xidian.edu.cn" + }, + { + "type": "image", + "bbox": [ + 0.1, + 0.312, + 0.901, + 0.632 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.103, + 0.634, + 0.893, + 0.65 + ], + "angle": 0, + "content": "Figure 1. Given a reference image and different target poses, our method can generate multiple images of the person in various poses." + }, + { + "type": "title", + "bbox": [ + 0.249, + 0.662, + 0.327, + 0.677 + ], + "angle": 0, + "content": "Abstract" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.695, + 0.484, + 0.862 + ], + "angle": 0, + "content": "Human pose generation is a complex task due to the non-rigid and highly variable nature of human body structures and appearances. However, existing methods often overlook the fundamental differences between spatial transformations of poses and texture generation for appearance, which makes them prone to overfitting. To address this issue, we propose a multi-pose generation framework driven by disentangled pose and appearance guidance. Our approach includes a Global-aware Pose Generation module that iteratively generates pose embeddings, enabling effective control over non-rigid body deformations. Addition-" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.663, + 0.907, + 0.875 + ], + "angle": 0, + "content": "ally, we introduce the Global-aware Transformer Decoder, which leverages similarity queries and attention mechanisms to achieve spatial transformations and enhance pose consistency through a Global-aware block. In the appearance generation phase, we condition a diffusion model on pose embeddings produced in the initial stage and introduce an Appearance Adapter that extracts high-level contextual semantic information from multi-scale features, enabling further refinement of pose appearance textures and providing appearance guidance. Extensive experiments on the UBC Fashion and TikTok datasets demonstrate that our framework achieves state-of-the-art results in both quality and fidelity, establishing it as a powerful approach for complex pose generation tasks." + }, + { + "type": "page_footnote", + "bbox": [ + 0.114, + 0.875, + 0.348, + 0.888 + ], + "angle": 0, + "content": "\\(\\dagger\\) Equal contributions in no particular order." + }, + { + "type": "page_footnote", + "bbox": [ + 0.114, + 0.888, + 0.246, + 0.9 + ], + "angle": 0, + "content": "* Corresponding author." + }, + { + "type": "list", + "bbox": [ + 0.114, + 0.875, + 0.348, + 0.9 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.517, + 0.957 + ], + "angle": 0, + "content": "5646" + } + ], + [ + { + "type": "title", + "bbox": [ + 0.093, + 0.09, + 0.223, + 0.107 + ], + "angle": 0, + "content": "1. Introduction" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.115, + 0.483, + 0.341 + ], + "angle": 0, + "content": "The inherent complexity of human body poses, with their non-rigid, dynamic structures and diverse appearances, makes pose generation a challenging task. This process requires transforming a person's reference pose into any target pose while preserving the individual's identity and appearance with high fidelity. This capability has substantial applications in areas such as virtual reality (VR) [1], augmented reality (AR) [2], and person re-identification [3]. Pose generation, however, involves two distinct challenges: effectively controlling the non-rigid deformations of human body structures and maintaining appearance consistency across different poses and viewpoints. Most existing methods directly generate human poses without addressing the fundamental differences between pose control and appearance generation." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.342, + 0.483, + 0.569 + ], + "angle": 0, + "content": "Advancements in deep learning have introduced promising solutions to pose generation, primarily through two classes of techniques: GAN-based methods and diffusion-based methods. GAN-based methods [4-8] typically align the reference pose with the target pose directly, learning a simple mapping from the appearance features of the reference pose to the target pose, which results in poor generalization. Diffusion models [9], as parameterized Markov chains trained via variational inference, offer straightforward optimization objectives and training efficiency, making them a viable alternative to GANs. Recent works [10-12] have applied diffusion models to pose generation, achieving promising results. However, we find that these methods still face limitations in terms of generation quality, identity preservation, or efficiency." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.57, + 0.483, + 0.902 + ], + "angle": 0, + "content": "In this paper, we present a novel framework for multi-posede generation based on disentangled pose and appearance guidance. In our approach, pose control primarily addresses spatial geometry, while appearance generation leverages semantic and contextual information. By assigning each task to a distinct module, the model efficiently addresses each objective independently. We introduce a Global-aware Pose Generation module that iteratively produces pose embeddings. Each embedding is a compact representation within the feature space, containing the spatially transformed pose along with appearance priors from the reference image. To generate pose embeddings, we propose the Global-aware Transformer Decoder (GTD), which employs similarity queries and attention mechanisms for spatial transformation of poses. An additional Global-aware block within the GTD models poses generated in previous iterations, integrating enriched global information to enhance pose consistency. During the appearance generation phase, the diffusion model is conditioned on multiple pose embeddings produced in the initial stage to refine appearance details. To further enhance appearance fidelity, we implement an Appearance Adapter module that encodes" + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.092, + 0.905, + 0.136 + ], + "angle": 0, + "content": "multi-scale features to extract high-level semantic information, providing fine-grained appearance guidance to the diffusion model." + }, + { + "type": "text", + "bbox": [ + 0.534, + 0.137, + 0.829, + 0.152 + ], + "angle": 0, + "content": "In summary, our main contributions include:" + }, + { + "type": "text", + "bbox": [ + 0.514, + 0.154, + 0.905, + 0.184 + ], + "angle": 0, + "content": "- Introducing a novel multi-pose generation framework based on disentangled pose and appearance guidance." + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.185, + 0.905, + 0.244 + ], + "angle": 0, + "content": "- Developing a global-aware pose generation module that models all generated poses, facilitating feature interactions among poses and integrating global information for improved pose generation." + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.245, + 0.905, + 0.303 + ], + "angle": 0, + "content": "- Designing an appearance adapter module to extract high-level contextual semantic information from multi-scale features, providing fine-grained appearance guidance to the diffusion model." + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.305, + 0.905, + 0.351 + ], + "angle": 0, + "content": "- Conducting extensive experiments on the UBC Fashion [13] and TikTok [14] datasets, achieving state-of-the-art results in both quality and fidelity." + }, + { + "type": "list", + "bbox": [ + 0.513, + 0.154, + 0.905, + 0.351 + ], + "angle": 0, + "content": null + }, + { + "type": "title", + "bbox": [ + 0.514, + 0.363, + 0.665, + 0.379 + ], + "angle": 0, + "content": "2. Related Works" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.388, + 0.905, + 0.524 + ], + "angle": 0, + "content": "Generative Adversarial Networks. Over the past decades, Generative Adversarial Networks (GANs) [15] have been extensively applied to human pose generation. Leveraging an adversarial learning framework, GANs can effectively model complex pose spatial transformations while producing realistic and detailed textures. Despite significant progress in recent years, challenges remain in predicting invisible regions, generating complex appearance textures, and maintaining identity consistency." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.524, + 0.905, + 0.69 + ], + "angle": 0, + "content": "To achieve accurate human pose synthesis with consistent clothing appearance and style, PISE [8] introduces a two-stage generative model that decouples appearance control from pose generation, enabling the creation of realistic human images. BiGraphGAN [4] addresses the challenges of pose deformation by incorporating a bipartite graph reasoning (BGR) block, which captures long-range relations between source and target poses in a bipartite graph structure. Additionally, BiGraphGAN includes an interaction- and-aggregation (IA) block to improve the feature representation of both body shape and appearance." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.69, + 0.905, + 0.901 + ], + "angle": 0, + "content": "Recently, attention mechanisms have been integrated into pose generation tasks. DPTN [5] employs a transformer-based siamese network, which supports pose generation by reconstructing reference images, thereby maintaining consistency in the target pose's appearance. NTED [7] presents a neural texture extraction and distribution module based on dual attention, extracting semantic textures from source features and distributing them according to the spatial layout of the target pose. IEPG [6] proposes a progressive pose generation framework, designed with an incremental evolution approach to address pose synthesis challenges. This multi-step generation strategy effectively handles occlusions, significantly enhancing the realism and coherence of the generated appearances." + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.516, + 0.957 + ], + "angle": 0, + "content": "5647" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.09, + 0.091, + 0.485, + 0.608 + ], + "angle": 0, + "content": "Diffusion models. Diffusion models (DM) [9] have emerged as a promising alternative to GANs, decomposing the generation process into a series of forward-backward diffusion steps to learn plausible transfer trajectories. This iterative refinement, starting with a low-quality noisy signal and progressively enhancing it, enables the generation of high-quality, contextually consistent samples, making diffusion models particularly effective for generating high-fidelity results. PIDM [12] introduces a texture diffusion module based on cross-attention, which accurately models the correspondences between pose and appearance information at the pixel level across source and target images. Additionally, PIDM employs a disentangled classifier-free guidance strategy [16] to ensure the conditional input and generated output closely match both pose and appearance. While PIDM achieves state-of-the-art synthesis results, it operates directly in pixel space, which makes its optimization and inference computationally expensive. MagicPose [11] adopts a two-stage training strategy to separately handle character appearance control and pose control, disentangling the two processes. Additionally, it incorporates facial keypoints, finger keypoints, and sparse body keypoints to achieve more precise pose control. Although these methods have made promising progress, they condition the pose generation on the global features of the reference pose and delegate both the spatial transformation of the pose and appearance texture generation to the diffusion model, causing the diffusion model to handle two distinct tasks simultaneously. In this paper, we develop a novel paradigm for multi-pose generation that disentangles pose control from appearance guidance. This approach allows the model to focus on a single objective within each module, thereby improving the consistency and controllability of the generated human pose images." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.624, + 0.182, + 0.639 + ], + "angle": 0, + "content": "3. Method" + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.65, + 0.336, + 0.665 + ], + "angle": 0, + "content": "3.1. Architecture and Overview" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.675, + 0.485, + 0.902 + ], + "angle": 0, + "content": "Given a reference image \\(\\mathcal{X}_{ref}\\) and multiple target poses \\(\\mathcal{P}^{1:N} = [p_1,p_2\\dots ,p_N]\\), where \\(N\\) is the number of target poses. The objective of our method is to generate a series of realistic images \\(\\mathcal{Y}^{1:N} = [y_1,y_2,\\dots,y_N]\\) while ensuring that both the poses and the appearance are consistent with the given poses \\(\\mathcal{P}^{1:N}\\) and the reference image \\(\\mathcal{X}_{ref}\\), respectively. The overall pipeline of our approach is illustrated in Fig. 2. To address the distinct nature of spatial transformations in pose generation and the fine-grained synthesis of appearance textures, we propose disentangling these two processes. This separation enables independent optimization of pose control and appearance refinement, leading to more accurate and natural image generation. To implement this, we use a multi-stage image encoder, such as Swin Transformer [17], denoted as \\(\\mathcal{M}_{ref}\\), to embed the ref" + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.092, + 0.905, + 0.183 + ], + "angle": 0, + "content": "erence image \\(\\mathcal{X}_{ref}\\) into multi-scale appearance embeddings \\(\\mathcal{F}_{ref} = [f_1, f_2, f_3, f_4]\\). These embeddings allow the network to perform multi-scale appearance control. For pose control, we developed a lightweight pose control network \\(\\mathcal{M}_p\\), based on ResNet [18], which extracts the pose features \\(\\mathcal{F}_p^{1:N}\\) from the multiple target poses." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.183, + 0.907, + 0.394 + ], + "angle": 0, + "content": "Next, more importantly, we designed the Global-aware Pose Generation module, which is trained to generate target pose embedding \\(\\mathcal{F}_{tp}^{1:N}\\). Unlike previous methods, we leverage the Global-aware Pose Generation module to perform spatial transformations of the pose while also encouraging the module to learn appearance priors from the reference image. This approach naturally embeds some appearance knowledge into the generated pose, simplifying subsequent appearance refinement or occluded region completion by the diffusion model. Additionally, to enhance the realism of textures, we employ a multi-scale Appearance Adapter (AD) module to provide multi-scale finer-grained appearance features to the U-net, further improving the quality of the generated images." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.402, + 0.785, + 0.417 + ], + "angle": 0, + "content": "3.2. Global-aware Pose Generation" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.425, + 0.907, + 0.756 + ], + "angle": 0, + "content": "Existing multi-person pose generation frameworks predominantly optimize individual target poses in isolation, which inherently limits their ability to model the spatiotemporal correlations among concurrent human motions. This limitation restricts these models' capacity to capture broader contextual information, potentially leading to inconsistencies or unnatural transitions between generated poses. To address this, we propose a unified module for multi-pose generation (Fig. 2) that leverages the Transformer architecture for its effective modeling capabilities, enabling coherent pose generation. Central to this module is the Global-aware Transformer Decoder (GTD) Block, which iteratively generates each target pose by integrating information from both the reference image and diverse target poses. Using the reference image features \\( f_{4} \\) from \\( \\mathcal{X}_{ref} \\) and the \\( N \\)-th target pose embeddings from \\( \\mathcal{P}^{N} \\), the GTD Block outputs the target pose \\( \\mathcal{F}_{tp}^{N} \\) for the \\( N \\)-th pose. To ensure coherence across poses, we enhance cross-pose consistency by incorporating previously generated poses (i.e., \\( \\mathcal{F}_{tp}^{1:N-1} \\) in Eq. 1) into each subsequent generation step. The iterative process for generating multiple pose \\( \\mathcal{F}_{tp}^{1:N} \\) and the specific generation for the \\( N \\)-th GTD iteration are defined as follows:" + }, + { + "type": "equation", + "bbox": [ + 0.59, + 0.763, + 0.905, + 0.85 + ], + "angle": 0, + "content": "\\[\n\\begin{array}{l} \\mathcal {F} _ {t p} ^ {1: N} = \\varphi (f _ {4}, \\mathcal {F} _ {p} ^ {1: N}), \\\\ \\mathcal {F} _ {t p} ^ {1: N} = \\mathcal {F} _ {t p} ^ {1: N} + \\frac {1}{N} \\sum_ {i = 1} ^ {N} \\mathcal {F} _ {t p} ^ {i}, \\\\ \\mathcal {F} _ {t p} ^ {N} = G T D \\left(f _ {4}, \\mathcal {F} _ {p} ^ {1: N}, \\mathcal {F} _ {t p} ^ {1: N - 1}\\right), \\tag {1} \\\\ \\end{array}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.856, + 0.907, + 0.901 + ], + "angle": 0, + "content": "where \\(\\varphi\\) represents the global-aware pose generation function, and \\(\\mathcal{F}_{tp}^{1:N - 1}\\) is the generated result from the previous \\(N - 1\\) iterations of GTD. To further reinforce contextual" + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.516, + 0.956 + ], + "angle": 0, + "content": "5648" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.096, + 0.078, + 0.907, + 0.436 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.089, + 0.447, + 0.907, + 0.517 + ], + "angle": 0, + "content": "Figure 2. Overview of the architecture of the proposed pipeline. The Image Encoder encodes a reference image to obtain multi-scale features, while the Pose Encoder encodes multiple target poses. Pose control is achieved by the Global-aware Pose Generation module, which iteratively generates pose embeddings from the final layer features of the image encoder and the target pose features. A diffusion model then refines the pose appearance, conditioned on these pose embeddings. The Appearance Adapter further enhances appearance detail by refining the multi-scale features from the image encoder and injecting them into the diffusion model." + }, + { + "type": "image", + "bbox": [ + 0.095, + 0.531, + 0.481, + 0.66 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.09, + 0.67, + 0.483, + 0.727 + ], + "angle": 0, + "content": "Figure 3. Comparison of the proposed Global-aware Transformer Decoder architecture (c) with two alternative decoder architectures (a) and (b). For clarity, we have omitted the residual connections in the cross-attention and self-attention layers." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.745, + 0.484, + 0.791 + ], + "angle": 0, + "content": "consistency, we include a pooling layer to capture global pose features, strengthening the model's understanding of overall pose structure across generated poses." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.796, + 0.484, + 0.901 + ], + "angle": 0, + "content": "Global-aware Transformer Decoder. For efficient target pose generation, we propose a Transformer-based decoder architecture capable of iterative pose generation. In this design, we developed three distinct Transformer decoder architectures, as shown in detail in Fig. 3. The specifics of these architectures will be discussed in the ablation experiments and the Appendix. Unlike traditional convolu" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.533, + 0.907, + 0.777 + ], + "angle": 0, + "content": "tional neural networks (CNNs) commonly applied in visual tasks, attention-based Transformers lack certain inductive biases inherent to conventional visual models [19]. To improve the decoder's ability to incorporate pose-specific inductive biases, we introduce a global-aware block that models the entire pose holistically. This global-aware block is a flexible, plug-and-play addition compatible with any Transformer architecture. Distinct from the decoder architectures (a) and (b), which employ cross-attention layers to capture semantic relationships. Specifically, we enhance the target pose feature \\(\\mathcal{F}_p^N\\) by computing a similarity matrix between \\(\\mathcal{F}_p^N\\) and \\(\\mathcal{F}_{tp}^{1:N - 1}\\), which integrates relevant information from prior poses, ensuring consistency within the generated poses. Next, a self-attention layer and feed-forward network (FFN) further refine \\(\\mathcal{F}_p^N\\) to enhance its expressiveness. The global-aware block is formalized as follows:" + }, + { + "type": "equation", + "bbox": [ + 0.571, + 0.787, + 0.848, + 0.81 + ], + "angle": 0, + "content": "\\[\n\\mathcal {F} _ {o u t} = \\phi (\\mathcal {F} _ {p} ^ {N} \\cdot \\mathcal {F} _ {t p} ^ {N - 1 ^ {T}}) \\cdot \\mathcal {F} _ {t p} ^ {N - 1} + \\mathcal {F} _ {p} ^ {N},\n\\]" + }, + { + "type": "equation", + "bbox": [ + 0.573, + 0.812, + 0.778, + 0.83 + ], + "angle": 0, + "content": "\\[\n\\mathcal {F} _ {o u t} = \\mathrm {M H S A} (\\tilde {\\mathcal {F}} _ {o u t}) + \\mathcal {F} _ {o u t},\n\\]" + }, + { + "type": "equation", + "bbox": [ + 0.573, + 0.834, + 0.907, + 0.851 + ], + "angle": 0, + "content": "\\[\n\\mathcal {F} _ {\\text {o u t}} = \\operatorname {F F N} \\left(\\hat {\\mathcal {F}} _ {\\text {o u t}}\\right) + \\mathcal {F} _ {\\text {o u t}}, \\tag {2}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.863, + 0.906, + 0.901 + ], + "angle": 0, + "content": "where \\(\\phi = \\frac{e^{s_{ij}}}{\\sum_k e^{s_{ik}}}\\) aggregates pose information. Here, MHSA denotes the multi-head self-attention layer. Finally," + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.516, + 0.956 + ], + "angle": 0, + "content": "5649" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.09, + 0.092, + 0.485, + 0.152 + ], + "angle": 0, + "content": "the reference image features \\(\\mathcal{F}_{ref}\\) interact with the globally enhanced target pose features \\(\\mathcal{F}_{out}\\) through a series of attention-based modules, producing the pose used as a conditional embedding input for the U-Net." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.161, + 0.416, + 0.178 + ], + "angle": 0, + "content": "3.3. Fine-grained Appearance Generation" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.184, + 0.485, + 0.32 + ], + "angle": 0, + "content": "Latent Diffusion Model. Latent diffusion models (LDMs) [20] have recently gained prominence as robust alternatives to generative adversarial networks (GANs) for various generative tasks. Building upon the foundational Markov chain framework of DDPM [9], LDMs adhere to a dual-phase paradigm comprising progressive noise injection and learned reconstruction. As formalized in [20], the forward diffusion operator systematically perturbs input data through a predefined variance schedule:" + }, + { + "type": "equation", + "bbox": [ + 0.144, + 0.331, + 0.483, + 0.357 + ], + "angle": 0, + "content": "\\[\nq \\left(z _ {t} \\mid z _ {t - 1}\\right) = \\mathcal {N} \\left(z _ {t}; \\sqrt {1 - \\beta_ {t}} z _ {t - 1}, \\beta_ {t} I\\right), \\tag {3}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.364, + 0.484, + 0.456 + ], + "angle": 0, + "content": "which transforms the latent distribution of the data \\( z_0 \\) into a noise distribution \\( z_{t} \\) (e.g., Gaussian distribution). The Gaussian variance, \\( \\beta_{t} \\), can be learned or derived from a predefined schedule. In the reverse diffusion process, LDMs learn to predict and remove noise, effectively reversing the transformation to reconstruct the original data distribution:" + }, + { + "type": "equation", + "bbox": [ + 0.14, + 0.466, + 0.483, + 0.484 + ], + "angle": 0, + "content": "\\[\np _ {\\theta} \\left(z _ {t - 1} \\mid z _ {t}\\right) = \\mathcal {N} \\left(z _ {t - 1}; \\mu_ {\\theta} \\left(z _ {t}, t, c\\right), \\sigma^ {2} I\\right), \\tag {4}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.494, + 0.484, + 0.554 + ], + "angle": 0, + "content": "where \\(\\mu_{\\theta}\\) denotes the learned mean parameterized by \\(\\theta\\) as a function of the current noise state \\(z_{t}\\), time step \\(t\\), and condition \\(c\\). A more detailed formulation of diffusion models is provided in the Appendix." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.554, + 0.484, + 0.765 + ], + "angle": 0, + "content": "In this paper, we utilize the powerful LDM model to generate human appearances. To achieve appearance guidance, we disentangle pose control and appearance guidance. Unlike prior approaches [11, 21], which rely on CLIP's [22] image encoder or a U-Net-like appearance encoder, we condition the diffusion model on multi-pose embeddings generated by the Global-aware Pose Generation module. Additionally, to ensure coherence across the poses and to generate the entire sequence simultaneously, we extend the 2D diffusion model to a 3D setting, inspired by [21]. For improved efficiency, we fine-tune only selected weights in the U-Net, as shown in Fig. 2. Here, the \"lock\" symbol denotes frozen weights, while the \"flame\" symbol indicates fine-tuned parameters." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.765, + 0.484, + 0.902 + ], + "angle": 0, + "content": "Appearance Adapter. To enable precise control over the character's appearance details, we introduce the Appearance Adapter (AD), shown in Fig. 2. The AD operates hierarchically between multi-scale upsampling blocks in the diffusion framework, extracting pose-agnostic appearance features from reference images through cross-scale feature learning. This architecture enables explicit disentanglement of high-level semantic understanding between appearance attributes and pose informations." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.092, + 0.907, + 0.258 + ], + "angle": 0, + "content": "Specifically, our implementation with \\( \\mathrm{k} = 3 \\) Appearance Embedding (AE) layers is augmented by two convolutional layers for dimensional alignment. Adopting the zero-initialization strategy from [23], these convolutional layers progressively smooth feature representations while suppressing high-frequency noise, thereby allowing stable learning of fine-grained appearance patterns. The AE layers enhance feature refinement through global attention mechanisms, followed by a Feed-Forward Network (FFN) that non-linearly reinforces the discriminative power of the processed features." + }, + { + "type": "title", + "bbox": [ + 0.514, + 0.268, + 0.654, + 0.284 + ], + "angle": 0, + "content": "3.4. Optimization" + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.291, + 0.907, + 0.322 + ], + "angle": 0, + "content": "We train our proposed model in an end-to-end manner. Our overall optimization objective is:" + }, + { + "type": "equation", + "bbox": [ + 0.634, + 0.334, + 0.905, + 0.35 + ], + "angle": 0, + "content": "\\[\n\\mathcal {L} _ {\\text {t o t a l}} = \\mathcal {L} _ {m g} + \\mathcal {L} _ {\\text {r e c}}, \\tag {5}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.362, + 0.907, + 0.393 + ], + "angle": 0, + "content": "where \\(\\mathcal{L}_{mg}\\) is used to optimize the multi-pose generation process. The optimization objective is:" + }, + { + "type": "equation", + "bbox": [ + 0.521, + 0.403, + 0.905, + 0.455 + ], + "angle": 0, + "content": "\\[\n\\begin{array}{l} \\mathcal {L} _ {m g} = \\mathbb {E} _ {z _ {0} ^ {1: N}, \\mathcal {X} _ {r e f}, \\mathcal {P} ^ {1: N}} [ \\| \\epsilon - \\\\ \\epsilon_ {\\theta} \\left(z _ {t} ^ {1: N}, t, \\varphi \\left(\\mathcal {F} _ {r e f}, \\mathcal {F} _ {p} ^ {1: N}\\right)\\right) | | _ {2} ^ {2} ], \\tag {6} \\\\ \\end{array}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.456, + 0.907, + 0.624 + ], + "angle": 0, + "content": "where \\(\\mathcal{X}_{ref}\\) denote the reference image, with \\(\\mathcal{F}_{ref}\\) representing its features. The target poses are given by \\(\\mathcal{P}^{1:N}\\), and \\(\\mathcal{F}_p^{1:N} = \\mathcal{M}_p(\\mathcal{P}^{1:N})\\) represents the features of the target poses. Additionally, \\(z_0^{1:N} = \\mathcal{E}(\\mathcal{V}^{1:N})\\) denotes the initial latent features, while \\(z_t^{1:N}\\) is the noisy latent representation derived from \\(z_0^{1:N}\\) at timestep \\(t\\), where \\(\\mathcal{E}\\) is VAE encoder. In addition, inspired by [5], we introduced an additional reconstruction task for the reference image during training to ensure the consistency between the appearance of the generated target poses and the reference image. The optimization objective of the reconstruction task is:" + }, + { + "type": "equation", + "bbox": [ + 0.56, + 0.635, + 0.907, + 0.654 + ], + "angle": 0, + "content": "\\[\n\\mathcal {L} _ {r e c} = \\mathbb {E} _ {z _ {0}, \\chi_ {r e f}, \\mathcal {P} ^ {0}} [ | | \\epsilon - \\epsilon_ {\\theta} (z _ {t}, t, \\mathcal {F} _ {r e f})) | | _ {2} ^ {2} ]. \\qquad (7)\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.664, + 0.907, + 0.786 + ], + "angle": 0, + "content": "Sampling. At inference time, we sample \\( z_{t} \\) from the Gaussian noise and iteratively sample from the previous step until \\( z_{0} \\), then we decode \\( z_{0} \\) to image space get the final human pose image. To improve the sample quality during the generation process, we use a classifier-free guidance strategy. To implement classifier-free guidance, we randomly drop the reference image used as the appearance condition with a probability of \\( 5\\% \\) during training." + }, + { + "type": "equation", + "bbox": [ + 0.541, + 0.798, + 0.905, + 0.814 + ], + "angle": 0, + "content": "\\[\n\\hat {\\epsilon} _ {\\theta} \\left(z _ {t}, t, \\varphi\\right) = \\lambda \\epsilon_ {\\theta} \\left(z _ {t}, t, \\varphi\\right) + (1 - \\lambda) \\epsilon_ {\\theta} \\left(z _ {t}, t\\right), \\tag {8}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.826, + 0.907, + 0.902 + ], + "angle": 0, + "content": "where \\(\\lambda\\) is a hyperparameter used to control the strength of guidance during the inference process, while \\(\\epsilon_{\\theta}(z_t,t,\\varphi)\\) and \\(\\epsilon_{\\theta}(z_t,t)\\) represent conditional prediction and unconditional prediction, respectively. To handle the case when the reference image is missing, inspired by [10], we use a" + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.516, + 0.957 + ], + "angle": 0, + "content": "5650" + } + ], + [ + { + "type": "table", + "bbox": [ + 0.097, + 0.078, + 0.902, + 0.226 + ], + "angle": 0, + "content": "
ApproachesUBC FashionTikTok
SSIM ↑PSNR ↑FID ↓LPIPS ↓\\( L_1 \\) ↓SSIM ↑PSNR ↑FID ↓LPIPS ↓\\( L_1 \\) ↓
PISE [8] (CVPR'21)0.86717.11618.6570.07520.0460.4299.947120.6480.30100.186
DPTN [5] (CVPR'22)0.83315.93719.9570.08760.0500.1885.696143.1100.45170.449
NTED [7] (CVPR'22)0.82415.68319.7840.09480.0530.3398.551140.3030.37270.228
BiGraphGAN [4] (IJCV'23)0.84416.40121.0120.08060.0470.4459.916147.5210.26760.177
PIDM [12] (CVPR'23)0.81218.87014.9510.16350.0920.3077.558130.3920.42770.262
CFLD [10] (CVPR'24)0.88922.14212.8100.07180.0260.72917.520105.6310.21360.078
MagicPose [11] (ICML'24)0.83120.74214.6490.08660.0300.67717.260108.0670.25070.083
Ours0.94125.65511.4870.04970.0140.75318.466103.7080.23110.072
" + }, + { + "type": "table_caption", + "bbox": [ + 0.143, + 0.237, + 0.851, + 0.251 + ], + "angle": 0, + "content": "Table 1. Quantitative comparison of pose sequence synthesis quality on the UBC Fashion [13], and TikTok [14] dataset." + }, + { + "type": "image", + "bbox": [ + 0.094, + 0.27, + 0.476, + 0.358 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.09, + 0.369, + 0.483, + 0.412 + ], + "angle": 0, + "content": "Figure 4. The pose synthesis accuracy of the proposed method on the UBC Fashion [13] dataset as the number of generated images increases, with 8 as the training setting." + }, + { + "type": "table", + "bbox": [ + 0.095, + 0.425, + 0.48, + 0.482 + ], + "angle": 0, + "content": "
MethodsPIDM [12] (CVPR'23)CFLD [10] (CVPR'24)MagicPose [11] (ICML'24)Ours(8-24)
#Param131M248M2081M367M
Inference Time28.64s35.8s67.8s5.5-24.1s
" + }, + { + "type": "table_caption", + "bbox": [ + 0.09, + 0.488, + 0.483, + 0.517 + ], + "angle": 0, + "content": "Table 2. Comparison of parameter count and inference time between the proposed method and other SOTA diffusion models." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.522, + 0.484, + 0.629 + ], + "angle": 0, + "content": "learnable vector as the conditional embedding. We concatenate the learnable vector with the reference image features to form a compact conditional embedding for the generation process, encouraging the learnable vector to capture more information beyond the reference image during training. This allows the model to receive additional appearance information when predicting invisible regions." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.642, + 0.224, + 0.659 + ], + "angle": 0, + "content": "4. Experiments" + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.666, + 0.173, + 0.683 + ], + "angle": 0, + "content": "4.1. Setup" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.689, + 0.484, + 0.901 + ], + "angle": 0, + "content": "Dataset. We validate the effectiveness of our proposed method on both the UBC Fashion [13] and TikTok [14] datasets. The UBC Fashion dataset is a collection of data featuring various poses and outfits, containing 500 training and 100 test videos. Compared to UBC Fashion, the TikTok dataset has more complex backgrounds and greater diversity in the proportions of people, making the task more challenging. The TikTok dataset comprises 350 dance videos featuring individual performers, each lasting between 10 and 15 seconds. To ensure a fair comparison with state-of-the-art methods, we use the same test set as [11] for evaluation. We set the image resolution to \\(256 \\times 256\\) and randomly sample \\(N\\) frames (in this paper, \\(N\\) is set to 8) from each training video for training. Additionally, we use Open-" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.271, + 0.907, + 0.332 + ], + "angle": 0, + "content": "Pose [24] and PGN [25] to extract human poses and human parsing maps. Due to category redundancy in the human parsing maps, we reduce the number of categories from 20 to 8 as [8]." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.344, + 0.908, + 0.588 + ], + "angle": 0, + "content": "Metrics. To evaluate the quality of the generated pose sequences, we report several metrics: Structural Similarity Index Measure (SSIM) [26], Peak Signal to Noise Ratio (PSNR), Learned Perceptual Image Patch Similarity (LPIPS) [27], and Fréchet Inception Distance (FID) [28]. SSIM and PSNR assess pixel-level similarity by measuring statistical differences between the generated images and the ground truth (GT). However, these metrics are limited to low-level image properties and may not fully capture perceptual quality. To address this limitation, we also include LPIPS and FID, which evaluate deep feature-based similarities. FID evaluates the distribution differences between the generated images and GT using Inception-v3 [29] features, while LPIPS measures the perceptual similarity between images using a VGG [30] network, making it more aligned with human perception of image similarity." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.599, + 0.909, + 0.903 + ], + "angle": 0, + "content": "Implementation details. The proposed method is implemented based on PyTorch [31] and Diffusers [32], trained in an end-to-end manner with a batch size of 16 across 4 NVIDIA A100 GPUs. The model was trained for 100 epochs utilizing the Adam [33] optimizer, starting with an initial learning rate of \\(5 \\times 10^{-5}\\). A learning rate warmup strategy was implemented for the first 1000 training steps, followed by a decay to \\(1 \\times 10^{-5}\\) at the 50th epoch. To accelerate model convergence, we initialize the image encoder using Swin-B [17] pretrained on ImageNet [34] and the U-net using version 1.5 of Stable Diffusion [20]. Additionally, to address the issue of poor facial realism in generated pose, we followed strategy [20] by fine-tuning the decoder of the VAE using both MSE reconstruction loss and perceptual loss [35]. During the fine-tuning process, we placed greater emphasis on the MSE reconstruction loss. Further details on the fine-tuning process can be found in the Appendix. For classifier-free guidance, we set the parameter \\(\\lambda\\) to 2.0 during sampling, and the reference condition \\(\\mathcal{X}_{ref}\\) was dropped with a probability of \\(5\\%\\) throughout training." + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.946, + 0.514, + 0.957 + ], + "angle": 0, + "content": "5651" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.101, + 0.079, + 0.264, + 0.482 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.269, + 0.079, + 0.342, + 0.481 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.344, + 0.079, + 0.498, + 0.481 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.501, + 0.079, + 0.655, + 0.481 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.658, + 0.079, + 0.735, + 0.481 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.738, + 0.079, + 0.811, + 0.481 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.814, + 0.08, + 0.901, + 0.481 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.09, + 0.497, + 0.908, + 0.527 + ], + "angle": 0, + "content": "Figure 5. Qualitative comparison of our method with SOTA methods on UBC Fashion [13] and TikTok[14] dataset. Our approach can generate more realistic details consistent with the reference image, such as garment folds, fine-grained textures, and facial identity features." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.544, + 0.325, + 0.56 + ], + "angle": 0, + "content": "4.2. Quantitative Comparison" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.567, + 0.484, + 0.839 + ], + "angle": 0, + "content": "We conducted a quantitative comparison of our method against state-of-the-art GAN-based and diffusion-based models on the UBC Fashion [13] and TikTok [14] datasets, as shown in Tab. 1. On the majority of metrics, our method outperforms the advanced methods on both datasets. Specifically, compared with four other diffusion models [10-12, 36], our approach achieves more stable generation performance and higher efficiency, with specific results shown in the Tab. 2, owing to the disentangling of pose control from appearance generation. Additionally, this design enables us to generate more human poses during inference (up to 3 times the training quantity) without additional training. The Fig. 4 illustrates the performance of our method as the generation count gradually increases. As shown in this figure, although an increase in generation quantity may lead to a slight performance decline, the results still serve as acceptable human image outputs, outperforming other SOTA methods." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.848, + 0.313, + 0.865 + ], + "angle": 0, + "content": "4.3. Qualitative Comparison" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.871, + 0.484, + 0.902 + ], + "angle": 0, + "content": "The qualitative comparison with state-of-the-art methods is illustrated in Fig. 5 and 6. When using these other meth" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.545, + 0.907, + 0.893 + ], + "angle": 0, + "content": "ods, both GAN-based and diffusion-based approaches show overfitting issues, resulting in the loss of person identity and clothing texture details. Additionally, when generating more complex clothing or unseen regions, issues like clothing blurriness, unnatural texture transitions, and substantial differences from the source image become apparent. Benefiting from our proposed strategy of disentangling pose control and appearance generation, our method mitigates overfitting to some extent. Furthermore, with the high-consistency pose embedding provided by Global-aware Pose Generation and the fine-grained, multi-scale features from Appearance Adapter, our approach yields more realistic results in maintaining person identity and capturing intricate clothing patterns. In particular, Fig. 6 provides an additional qualitative comparison with existing SOTA diffusion models for generating multiple poses. This scenario places higher demands on cross-view consistency in person identity and appearance. As observed in the Fig. 6, our method achieves higher consistency across different viewpoints of generated person images, thanks to Global-aware Transformer Decoder's global pose awareness in the pose generation process. More qualitative comparisons can be found in the Appendix." + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.516, + 0.957 + ], + "angle": 0, + "content": "5652" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.094, + 0.071, + 0.482, + 0.398 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.09, + 0.406, + 0.483, + 0.435 + ], + "angle": 0, + "content": "Figure 6. Qualitative comparison with other SOTA methods on various poses, with drawbacks highlighted by red circles." + }, + { + "type": "table", + "bbox": [ + 0.094, + 0.449, + 0.481, + 0.558 + ], + "angle": 0, + "content": "
ModelSSIM ↑PSNR ↑FID ↓LPIPS ↓
w/o Global-aware Pose Generation (GPG)0.936225.02411.8650.0532
w/o Multi-stage Image Encoder (MIE)0.934424.89511.2840.0554
w/o Appearance Adapter (AD)0.938425.20511.9340.0517
w/o Fine-tune VAE (FVAE)0.938625.46411.6270.0514
w/o Lrec0.940325.53911.5470.0516
Ours + (a)0.938925.44311.4870.0503
Ours + (b)0.938125.44711.5720.0506
Ours0.941125.65511.4870.0497
" + }, + { + "type": "table_caption", + "bbox": [ + 0.09, + 0.565, + 0.483, + 0.594 + ], + "angle": 0, + "content": "Table 3. Ablation study of different components, where (a) and (b) represent different decoder architectures in Fig. 3." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.606, + 0.245, + 0.623 + ], + "angle": 0, + "content": "4.4. Ablation Study" + }, + { + "type": "text", + "bbox": [ + 0.089, + 0.629, + 0.486, + 0.888 + ], + "angle": 0, + "content": "Tab. 3 and Fig. 7 show quantitative and qualitative results, respectively. It can be observed that both GPG and AD are essential for pose synthesis performance, with their absence potentially leading to a performance drop of up to \\(3.8\\%\\) (FID). This highlights the importance of separating pose control and appearance generation strategies. We also replaced MIE with a CLIP image encoder, resulting in a \\(0.7\\%\\) (SSIM) and \\(2.9\\%\\) (PSNR) decline, though with an improvement in FID performance. As shown in the first row of Fig. 7, the CLIP image encoder fails to retain intricate clothing details, leading to a loss of appearance-related information. Additionally, fine-tuning the VAE for the specific dataset helps preserve realistic details of faces and clothing to a certain degree. Finally, we also experimented with two alternative Transformer decoder architectures in Fig. 3, both of which led to varying degrees of performance decline. Detailed ablation experiments can be found in the Appendix." + }, + { + "type": "image", + "bbox": [ + 0.522, + 0.071, + 0.898, + 0.219 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.513, + 0.226, + 0.907, + 0.253 + ], + "angle": 0, + "content": "Figure 7. Qualitative ablation study on the UBC Fashion [13] dataset." + }, + { + "type": "table", + "bbox": [ + 0.518, + 0.259, + 0.905, + 0.312 + ], + "angle": 0, + "content": "
MethodsPIDM [12] (CVPR'23)CFLD [10] (CVPR'24)MagicPose [11] (ICML'24)Ours
Average13.6%16.7%8.3%61.4%
" + }, + { + "type": "table_caption", + "bbox": [ + 0.513, + 0.318, + 0.907, + 0.375 + ], + "angle": 0, + "content": "Table 4. User study on the UBC Fashion [13] dataset. We followed [11] to collect vote data from 30 volunteers. The participants believed that our proposed method can generate visually realistic and identity-consistent human poses." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.383, + 0.637, + 0.399 + ], + "angle": 0, + "content": "4.5. User Study" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.407, + 0.907, + 0.573 + ], + "angle": 0, + "content": "To more accurately evaluate the effectiveness of our proposed method, we designed a user study. In this experiment, we collected images of 9 subjects from the test set for comparative evaluation. One subject was designated as a warm-up, while the remaining eight were used as formal test. For each subject, eight distinct pose images were randomly selected to comparison. We presented images generated by various methods [10-12] to users, instructing them to evaluate the images based on visual quality, pose accuracy, and detail consistency, and to select the method they preferred. We present the averaged vote result in Tab. 4." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.588, + 0.642, + 0.604 + ], + "angle": 0, + "content": "5. Conclusions" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.614, + 0.907, + 0.84 + ], + "angle": 0, + "content": "In this paper, we propose a unified framework for diverse multi-pose generation driven by disentangled pose and appearance guidance. This framework separately addresses complex human pose control and appearance generation to produce realistic and varied human poses. To achieve precise pose control and realistic appearance generation, we introduce Global-aware Pose Generation and Appearance Adapter module. These modules enhance interactions among poses during the generation process, leveraging all available mutual information to provide global guidance, while also learning high-level semantic information from multi-scale features to preserve texture details as much as possible. Extensive experiments demonstrate that our proposed approach achieves significant advantages over other state-of-the-art methods." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.84, + 0.909, + 0.901 + ], + "angle": 0, + "content": "Acknowledgments. This work was supported by the National Natural Science Foundation of China (62036006, 62276200) and the Tianjin Natural Science Foundation (24JCYBJC00310)." + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.516, + 0.957 + ], + "angle": 0, + "content": "5653" + } + ], + [ + { + "type": "title", + "bbox": [ + 0.093, + 0.091, + 0.188, + 0.106 + ], + "angle": 0, + "content": "References" + }, + { + "type": "ref_text", + "bbox": [ + 0.101, + 0.116, + 0.482, + 0.143 + ], + "angle": 0, + "content": "[1] Grigore C Burdea and Philippe Coiffet. *Virtual reality technology*. John Wiley & Sons, 2003. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.101, + 0.145, + 0.483, + 0.199 + ], + "angle": 0, + "content": "[2] Julie Carmigniani, Borko Furht, Marco Anisetti, Paolo Ceravolo, Ernesto Damiani, and Misa Ivkovic. Augmented reality technologies, systems and applications. Multimedia tools and applications, 51:341-377, 2011. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.201, + 0.483, + 0.268 + ], + "angle": 0, + "content": "[3] Liang Zheng, Liyue Shen, Lu Tian, Shengjin Wang, Jingdong Wang, and Qi Tian. Scalable person re-identification: A benchmark. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 1116-1124, 2015. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.271, + 0.482, + 0.325 + ], + "angle": 0, + "content": "[4] Hao Tang, Ling Shao, Philip HS Torr, and Nicu Sebe. Bipartite graph reasoning gans for person pose and facial image synthesis. International Journal of Computer Vision, 131(3):644-658, 2023. 2, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.328, + 0.483, + 0.395 + ], + "angle": 0, + "content": "[5] Pengze Zhang, Lingxiao Yang, Jian-Huang Lai, and Xiaohua Xie. Exploring dual-task correlation for pose guided person image generation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 7713-7722, 2022. 2, 5, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.398, + 0.483, + 0.452 + ], + "angle": 0, + "content": "[6] Yuelong Li, Tengfei Xiao, Lei Geng, and Jianming Wang. Direct may not be the best: An incremental evolution view of pose generation. In Proceedings of the AAAI Conference on Artificial Intelligence, volume 38, pages 3270-3278, 2024. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.454, + 0.483, + 0.522 + ], + "angle": 0, + "content": "[7] Yurui Ren, Xiaqing Fan, Ge Li, Shan Liu, and Thomas H Li. Neural texture extraction and distribution for controllable person image synthesis. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 13535-13544, 2022. 2, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.524, + 0.483, + 0.591 + ], + "angle": 0, + "content": "[8] Jinsong Zhang, Kun Li, Yu-Kun Lai, and Jingyu Yang. Pise: Person image synthesis and editing with decoupled gan. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 7982-7990, 2021. 2, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.594, + 0.483, + 0.635 + ], + "angle": 0, + "content": "[9] Jonathan Ho, Ajay Jain, and Pieter Abbeel. Denoising diffusion probabilistic models. Advances in Neural Information Processing Systems, 33:6840-6851, 2020. 2, 3, 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.636, + 0.483, + 0.704 + ], + "angle": 0, + "content": "[10] Yanzuo Lu, Manlin Zhang, Andy J Ma, Xiaohua Xie, and Jianhuang Lai. Coarse-to-fine latent diffusion for pose-guided person image synthesis. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 6420-6429, 2024. 2, 5, 6, 7, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.706, + 0.483, + 0.788 + ], + "angle": 0, + "content": "[11] Di Chang, Yichun Shi, Quankai Gao, Hongyi Xu, Jessica Fu, Guoxian Song, Qing Yan, Yizhe Zhu, Xiao Yang, and Mohammad Soleymani. Magicpose: Realistic human poses and facial expressions retargeting with identity-aware diffusion. In Proceedings of the International Conference on Machine Learning, 2023. 3, 5, 6, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.79, + 0.483, + 0.872 + ], + "angle": 0, + "content": "[12] Ankan Kumar Bhunia, Salman Khan, Hisham Cholakkal, Rao Muhammad Anwer, Jorma Laaksonen, Mubarak Shah, and Fahad Shahbaz Khan. Person image synthesis via denoising diffusion model. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 5968-5976, 2023. 2, 3, 6, 7, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.874, + 0.483, + 0.901 + ], + "angle": 0, + "content": "[13] Polina Zablotskaia, Aliaksandr Siarohin, Bo Zhao, and Leonid Sigal. Dwnet: Dense warp-based network for" + }, + { + "type": "list", + "bbox": [ + 0.094, + 0.116, + 0.483, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.548, + 0.094, + 0.905, + 0.12 + ], + "angle": 0, + "content": "pose-guided human video generation. arXiv preprint arXiv:1910.09139, 2019. 2, 6, 7, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.123, + 0.905, + 0.19 + ], + "angle": 0, + "content": "[14] Yasamin Jafarian and Hyun Soo Park. Learning high fidelity depths of dressed humans by watching social media dance videos. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 12753-12762, 2021. 2, 6, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.193, + 0.905, + 0.247 + ], + "angle": 0, + "content": "[15] Ian Goodfellow, Jean Pouget-Abadie, Mehdi Mirza, Bing Xu, David Warde-Farley, Sherjil Ozair, Aaron Courville, and Yoshua Bengio. Generative adversarial networks. Communications of the ACM, 63(11):139-144, 2020. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.25, + 0.905, + 0.291 + ], + "angle": 0, + "content": "[16] Jonathan Ho and Tim Salimans. Classifier-free diffusion guidance. In NeurIPS 2021 Workshop on Deep Generative Models and Downstream Applications, 2021. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.294, + 0.905, + 0.362 + ], + "angle": 0, + "content": "[17] Ze Liu, Yutong Lin, Yue Cao, Han Hu, Yixuan Wei, Zheng Zhang, Stephen Lin, and Baining Guo. Swin transformer: Hierarchical vision transformer using shifted windows. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 10012-10022, 2021. 3, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.365, + 0.905, + 0.419 + ], + "angle": 0, + "content": "[18] Kaiming He, Xiangyu Zhang, Shaoqing Ren, and Jian Sun. Deep residual learning for image recognition. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 770-778, 2016. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.422, + 0.905, + 0.49 + ], + "angle": 0, + "content": "[19] Xumin Yu, Yongming Rao, Ziyi Wang, Zuyan Liu, Jiwen Lu, and Jie Zhou. Pointr: Diverse point cloud completion with geometry-aware transformers. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 12498-12507, 2021. 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.492, + 0.907, + 0.561 + ], + "angle": 0, + "content": "[20] Robin Rombach, Andreas Blattmann, Dominik Lorenz, Patrick Esser, and Björn Ommer. High-resolution image synthesis with latent diffusion models. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 10684-10695, 2022. 5, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.563, + 0.907, + 0.645 + ], + "angle": 0, + "content": "[21] Zhongcong Xu, Jianfeng Zhang, Jun Hao Liew, Hanshu Yan, Jia-Wei Liu, Chenxu Zhang, Jiashi Feng, and Mike Zheng Shou. Magicanimate: Temporally consistent human image animation using diffusion model. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 1481–1490, 2024. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.648, + 0.905, + 0.73 + ], + "angle": 0, + "content": "[22] Alec Radford, Jong Wook Kim, Chris Hallacy, Aditya Ramesh, Gabriel Goh, Sandhini Agarwal, Girish Sastry, Amanda Askell, Pamela Mishkin, Jack Clark, et al. Learning transferable visual models from natural language supervision. In Proceedings of the International Conference on Machine Learning, pages 8748-8763. PMLR, 2021. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.733, + 0.905, + 0.787 + ], + "angle": 0, + "content": "[23] Lvmin Zhang, Anyi Rao, and Maneesh Agrawala. Adding conditional control to text-to-image diffusion models. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 3836-3847, 2023. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.789, + 0.905, + 0.845 + ], + "angle": 0, + "content": "[24] Riza Alp Güler, Natalia Neverova, and Iasonas Kokkinos. Densesepose: Dense human pose estimation in the wild. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 7297-7306, 2018. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.847, + 0.905, + 0.901 + ], + "angle": 0, + "content": "[25] Ke Gong, Xiaodan Liang, Yicheng Li, Yimin Chen, Ming Yang, and Liang Lin. Instance-level human parsing via part grouping network. In Proceedings of the European Conference on Computer Vision, pages 770-785, 2018. 6" + }, + { + "type": "list", + "bbox": [ + 0.518, + 0.094, + 0.907, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.516, + 0.957 + ], + "angle": 0, + "content": "5654" + } + ], + [ + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.093, + 0.482, + 0.147 + ], + "angle": 0, + "content": "[26] Zhou Wang, Alan C Bovik, Hamid R Sheikh, and Eero P Simoncelli. Image quality assessment: from error visibility to structural similarity. IEEE transactions on image processing, 13(4):600-612, 2004. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.149, + 0.483, + 0.219 + ], + "angle": 0, + "content": "[27] Richard Zhang, Phillip Isola, Alexei A Efros, Eli Shechtman, and Oliver Wang. The unreasonable effectiveness of deep features as a perceptual metric. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 586-595, 2018. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.22, + 0.482, + 0.288 + ], + "angle": 0, + "content": "[28] Martin Heusel, Hubert Ramsauer, Thomas Unterthiner, Bernhard Nessler, and Sepp Hochreiter. Gans trained by a two time-scale update rule converge to a local nash equilibrium. Advances in Neural Information Processing Systems, 30, 2017. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.29, + 0.482, + 0.359 + ], + "angle": 0, + "content": "[29] Christian Szegedy, Vincent Vanhoucke, Sergey Ioffe, Jon Shlens, and Zbigniew Wojna. Rethinking the inception architecture for computer vision. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 2818-2826, 2016. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.361, + 0.482, + 0.402 + ], + "angle": 0, + "content": "[30] Karen Simonyan and Andrew Zisserman. Very deep convolutional networks for large-scale image recognition. arXiv preprint arXiv:1409.1556, 2014. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.403, + 0.482, + 0.485 + ], + "angle": 0, + "content": "[31] Adam Paszke, Sam Gross, Francisco Massa, Adam Lerer, James Bradbury, Gregory Chanan, Trevor Killeen, Zeming Lin, Natalia Gimelshein, Luca Antiga, et al. Pytorch: An imperative style, high-performance deep learning library. Advances in Neural Information Processing Systems, 32, 2019. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.488, + 0.482, + 0.57 + ], + "angle": 0, + "content": "[32] Patrick von Platen, Suraj Patil, Anton Lozhkov, Pedro Cuenca, Nathan Lambert, Kashif Rasul, Mishig Davaadorj, Dhruv Nair, Sayak Paul, William Berman, Yiyi Xu, Steven Liu, and Thomas Wolf. Diffusers: State-of-the-art diffusion models. https://github.com/huggingface/diffusers, 2022.6" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.572, + 0.482, + 0.599 + ], + "angle": 0, + "content": "[33] Diederik P Kingma. Adam: A method for stochastic optimization. arXiv preprint arXiv:1412.6980, 2014. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.601, + 0.482, + 0.668 + ], + "angle": 0, + "content": "[34] Jia Deng, Wei Dong, Richard Socher, Li-Jia Li, Kai Li, and Li Fei-Fei. Imagenet: A large-scale hierarchical image database. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition, pages 248–255. IEEE, 2009. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.671, + 0.482, + 0.74 + ], + "angle": 0, + "content": "[35] Justin Johnson, Alexandre Alahi, and Li Fei-Fei. Perceptual losses for real-time style transfer and super-resolution. In Computer Vision-ECCV 2016: 14th European Conference, Amsterdam, The Netherlands, October 11-14, 2016, Proceedings, Part II 14, pages 694-711. Springer, 2016. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.742, + 0.482, + 0.825 + ], + "angle": 0, + "content": "[36] Tan Wang, Linjie Li, Kevin Lin, Yuanhao Zhai, Chung-Ching Lin, Zhengyuan Yang, Hanwang Zhang, Zicheng Liu, and Lijuan Wang. Disco: Disentangled control for realistic human dance generation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 9326-9336, 2024. 7" + }, + { + "type": "list", + "bbox": [ + 0.093, + 0.093, + 0.483, + 0.825 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.946, + 0.515, + 0.956 + ], + "angle": 0, + "content": "5655" + } + ] +] \ No newline at end of file diff --git a/2025/Disentangled Pose and Appearance Guidance for Multi-Pose Generation/eaadccfc-d365-4d2d-83e8-71935f857eca_origin.pdf b/2025/Disentangled Pose and Appearance Guidance for Multi-Pose Generation/eaadccfc-d365-4d2d-83e8-71935f857eca_origin.pdf new file mode 100644 index 0000000000000000000000000000000000000000..cd627c5d027884961b7598f096aff872f46aecfb --- /dev/null +++ b/2025/Disentangled Pose and Appearance Guidance for Multi-Pose Generation/eaadccfc-d365-4d2d-83e8-71935f857eca_origin.pdf @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:57574d18407a4e4f2ce1cc0369e7574592188db90b15d2f133a9d42c1f8bbb70 +size 2247529 diff --git a/2025/Disentangled Pose and Appearance Guidance for Multi-Pose Generation/full.md b/2025/Disentangled Pose and Appearance Guidance for Multi-Pose Generation/full.md new file mode 100644 index 0000000000000000000000000000000000000000..3c4d1ebb880eb91c304b09ef01702b642d23e911 --- /dev/null +++ b/2025/Disentangled Pose and Appearance Guidance for Multi-Pose Generation/full.md @@ -0,0 +1,258 @@ +# Disentangled Pose and Appearance Guidance for Multi-Pose Generation + +Tengfei Xiao $^{1,4}$ , Yue Wu $^{\dagger,*,1,4}$ , Yuelong Li $^{\dagger,*,2,5}$ , Can Qin $^{3}$ , Maoguo Gong $^{1,4}$ , Qiguang Miao $^{1,4}$ , Wenping Ma $^{1}$ $^{1}$ Xidian University $^{2}$ Tiangong University $^{3}$ Northeastern University $^{4}$ MoE Key Lab of Collaborative Intelligence Systems, Xidian University + $^{5}$ Tianjin Key Laboratory of Autonomous Intelligence Technology and Systems, Tiangong University tfxiao@stu.xidian.edu.cn, {ywu, qgmiao}@xidian.edu.cn, liyuelong@pku.edu.cn, qin.ca@northeastern.edu, gong@ieee.org, wpma@mail.xidian.edu.cn + +![](images/0206018f2f10f04df1d5cff292ffedde2c0bf9b7ab97c5366686992084d8ca15.jpg) +Figure 1. Given a reference image and different target poses, our method can generate multiple images of the person in various poses. + +# Abstract + +Human pose generation is a complex task due to the non-rigid and highly variable nature of human body structures and appearances. However, existing methods often overlook the fundamental differences between spatial transformations of poses and texture generation for appearance, which makes them prone to overfitting. To address this issue, we propose a multi-pose generation framework driven by disentangled pose and appearance guidance. Our approach includes a Global-aware Pose Generation module that iteratively generates pose embeddings, enabling effective control over non-rigid body deformations. Addition- + +ally, we introduce the Global-aware Transformer Decoder, which leverages similarity queries and attention mechanisms to achieve spatial transformations and enhance pose consistency through a Global-aware block. In the appearance generation phase, we condition a diffusion model on pose embeddings produced in the initial stage and introduce an Appearance Adapter that extracts high-level contextual semantic information from multi-scale features, enabling further refinement of pose appearance textures and providing appearance guidance. Extensive experiments on the UBC Fashion and TikTok datasets demonstrate that our framework achieves state-of-the-art results in both quality and fidelity, establishing it as a powerful approach for complex pose generation tasks. + +# 1. Introduction + +The inherent complexity of human body poses, with their non-rigid, dynamic structures and diverse appearances, makes pose generation a challenging task. This process requires transforming a person's reference pose into any target pose while preserving the individual's identity and appearance with high fidelity. This capability has substantial applications in areas such as virtual reality (VR) [1], augmented reality (AR) [2], and person re-identification [3]. Pose generation, however, involves two distinct challenges: effectively controlling the non-rigid deformations of human body structures and maintaining appearance consistency across different poses and viewpoints. Most existing methods directly generate human poses without addressing the fundamental differences between pose control and appearance generation. + +Advancements in deep learning have introduced promising solutions to pose generation, primarily through two classes of techniques: GAN-based methods and diffusion-based methods. GAN-based methods [4-8] typically align the reference pose with the target pose directly, learning a simple mapping from the appearance features of the reference pose to the target pose, which results in poor generalization. Diffusion models [9], as parameterized Markov chains trained via variational inference, offer straightforward optimization objectives and training efficiency, making them a viable alternative to GANs. Recent works [10-12] have applied diffusion models to pose generation, achieving promising results. However, we find that these methods still face limitations in terms of generation quality, identity preservation, or efficiency. + +In this paper, we present a novel framework for multi-posede generation based on disentangled pose and appearance guidance. In our approach, pose control primarily addresses spatial geometry, while appearance generation leverages semantic and contextual information. By assigning each task to a distinct module, the model efficiently addresses each objective independently. We introduce a Global-aware Pose Generation module that iteratively produces pose embeddings. Each embedding is a compact representation within the feature space, containing the spatially transformed pose along with appearance priors from the reference image. To generate pose embeddings, we propose the Global-aware Transformer Decoder (GTD), which employs similarity queries and attention mechanisms for spatial transformation of poses. An additional Global-aware block within the GTD models poses generated in previous iterations, integrating enriched global information to enhance pose consistency. During the appearance generation phase, the diffusion model is conditioned on multiple pose embeddings produced in the initial stage to refine appearance details. To further enhance appearance fidelity, we implement an Appearance Adapter module that encodes + +multi-scale features to extract high-level semantic information, providing fine-grained appearance guidance to the diffusion model. + +In summary, our main contributions include: + +- Introducing a novel multi-pose generation framework based on disentangled pose and appearance guidance. +- Developing a global-aware pose generation module that models all generated poses, facilitating feature interactions among poses and integrating global information for improved pose generation. +- Designing an appearance adapter module to extract high-level contextual semantic information from multi-scale features, providing fine-grained appearance guidance to the diffusion model. +- Conducting extensive experiments on the UBC Fashion [13] and TikTok [14] datasets, achieving state-of-the-art results in both quality and fidelity. + +# 2. Related Works + +Generative Adversarial Networks. Over the past decades, Generative Adversarial Networks (GANs) [15] have been extensively applied to human pose generation. Leveraging an adversarial learning framework, GANs can effectively model complex pose spatial transformations while producing realistic and detailed textures. Despite significant progress in recent years, challenges remain in predicting invisible regions, generating complex appearance textures, and maintaining identity consistency. + +To achieve accurate human pose synthesis with consistent clothing appearance and style, PISE [8] introduces a two-stage generative model that decouples appearance control from pose generation, enabling the creation of realistic human images. BiGraphGAN [4] addresses the challenges of pose deformation by incorporating a bipartite graph reasoning (BGR) block, which captures long-range relations between source and target poses in a bipartite graph structure. Additionally, BiGraphGAN includes an interaction- and-aggregation (IA) block to improve the feature representation of both body shape and appearance. + +Recently, attention mechanisms have been integrated into pose generation tasks. DPTN [5] employs a transformer-based siamese network, which supports pose generation by reconstructing reference images, thereby maintaining consistency in the target pose's appearance. NTED [7] presents a neural texture extraction and distribution module based on dual attention, extracting semantic textures from source features and distributing them according to the spatial layout of the target pose. IEPG [6] proposes a progressive pose generation framework, designed with an incremental evolution approach to address pose synthesis challenges. This multi-step generation strategy effectively handles occlusions, significantly enhancing the realism and coherence of the generated appearances. + +Diffusion models. Diffusion models (DM) [9] have emerged as a promising alternative to GANs, decomposing the generation process into a series of forward-backward diffusion steps to learn plausible transfer trajectories. This iterative refinement, starting with a low-quality noisy signal and progressively enhancing it, enables the generation of high-quality, contextually consistent samples, making diffusion models particularly effective for generating high-fidelity results. PIDM [12] introduces a texture diffusion module based on cross-attention, which accurately models the correspondences between pose and appearance information at the pixel level across source and target images. Additionally, PIDM employs a disentangled classifier-free guidance strategy [16] to ensure the conditional input and generated output closely match both pose and appearance. While PIDM achieves state-of-the-art synthesis results, it operates directly in pixel space, which makes its optimization and inference computationally expensive. MagicPose [11] adopts a two-stage training strategy to separately handle character appearance control and pose control, disentangling the two processes. Additionally, it incorporates facial keypoints, finger keypoints, and sparse body keypoints to achieve more precise pose control. Although these methods have made promising progress, they condition the pose generation on the global features of the reference pose and delegate both the spatial transformation of the pose and appearance texture generation to the diffusion model, causing the diffusion model to handle two distinct tasks simultaneously. In this paper, we develop a novel paradigm for multi-pose generation that disentangles pose control from appearance guidance. This approach allows the model to focus on a single objective within each module, thereby improving the consistency and controllability of the generated human pose images. + +# 3. Method + +# 3.1. Architecture and Overview + +Given a reference image $\mathcal{X}_{ref}$ and multiple target poses $\mathcal{P}^{1:N} = [p_1,p_2\dots ,p_N]$ , where $N$ is the number of target poses. The objective of our method is to generate a series of realistic images $\mathcal{Y}^{1:N} = [y_1,y_2,\dots,y_N]$ while ensuring that both the poses and the appearance are consistent with the given poses $\mathcal{P}^{1:N}$ and the reference image $\mathcal{X}_{ref}$ , respectively. The overall pipeline of our approach is illustrated in Fig. 2. To address the distinct nature of spatial transformations in pose generation and the fine-grained synthesis of appearance textures, we propose disentangling these two processes. This separation enables independent optimization of pose control and appearance refinement, leading to more accurate and natural image generation. To implement this, we use a multi-stage image encoder, such as Swin Transformer [17], denoted as $\mathcal{M}_{ref}$ , to embed the ref + +erence image $\mathcal{X}_{ref}$ into multi-scale appearance embeddings $\mathcal{F}_{ref} = [f_1, f_2, f_3, f_4]$ . These embeddings allow the network to perform multi-scale appearance control. For pose control, we developed a lightweight pose control network $\mathcal{M}_p$ , based on ResNet [18], which extracts the pose features $\mathcal{F}_p^{1:N}$ from the multiple target poses. + +Next, more importantly, we designed the Global-aware Pose Generation module, which is trained to generate target pose embedding $\mathcal{F}_{tp}^{1:N}$ . Unlike previous methods, we leverage the Global-aware Pose Generation module to perform spatial transformations of the pose while also encouraging the module to learn appearance priors from the reference image. This approach naturally embeds some appearance knowledge into the generated pose, simplifying subsequent appearance refinement or occluded region completion by the diffusion model. Additionally, to enhance the realism of textures, we employ a multi-scale Appearance Adapter (AD) module to provide multi-scale finer-grained appearance features to the U-net, further improving the quality of the generated images. + +# 3.2. Global-aware Pose Generation + +Existing multi-person pose generation frameworks predominantly optimize individual target poses in isolation, which inherently limits their ability to model the spatiotemporal correlations among concurrent human motions. This limitation restricts these models' capacity to capture broader contextual information, potentially leading to inconsistencies or unnatural transitions between generated poses. To address this, we propose a unified module for multi-pose generation (Fig. 2) that leverages the Transformer architecture for its effective modeling capabilities, enabling coherent pose generation. Central to this module is the Global-aware Transformer Decoder (GTD) Block, which iteratively generates each target pose by integrating information from both the reference image and diverse target poses. Using the reference image features $f_{4}$ from $\mathcal{X}_{ref}$ and the $N$ -th target pose embeddings from $\mathcal{P}^{N}$ , the GTD Block outputs the target pose $\mathcal{F}_{tp}^{N}$ for the $N$ -th pose. To ensure coherence across poses, we enhance cross-pose consistency by incorporating previously generated poses (i.e., $\mathcal{F}_{tp}^{1:N-1}$ in Eq. 1) into each subsequent generation step. The iterative process for generating multiple pose $\mathcal{F}_{tp}^{1:N}$ and the specific generation for the $N$ -th GTD iteration are defined as follows: + +$$ +\begin{array}{l} \mathcal {F} _ {t p} ^ {1: N} = \varphi (f _ {4}, \mathcal {F} _ {p} ^ {1: N}), \\ \mathcal {F} _ {t p} ^ {1: N} = \mathcal {F} _ {t p} ^ {1: N} + \frac {1}{N} \sum_ {i = 1} ^ {N} \mathcal {F} _ {t p} ^ {i}, \\ \mathcal {F} _ {t p} ^ {N} = G T D \left(f _ {4}, \mathcal {F} _ {p} ^ {1: N}, \mathcal {F} _ {t p} ^ {1: N - 1}\right), \tag {1} \\ \end{array} +$$ + +where $\varphi$ represents the global-aware pose generation function, and $\mathcal{F}_{tp}^{1:N - 1}$ is the generated result from the previous $N - 1$ iterations of GTD. To further reinforce contextual + +![](images/ac61b8f68c98a2ff1c8b76e777bf3a308e8081337e6061079e0e416f5e096b74.jpg) +Figure 2. Overview of the architecture of the proposed pipeline. The Image Encoder encodes a reference image to obtain multi-scale features, while the Pose Encoder encodes multiple target poses. Pose control is achieved by the Global-aware Pose Generation module, which iteratively generates pose embeddings from the final layer features of the image encoder and the target pose features. A diffusion model then refines the pose appearance, conditioned on these pose embeddings. The Appearance Adapter further enhances appearance detail by refining the multi-scale features from the image encoder and injecting them into the diffusion model. + +![](images/9f4b059a545121584618136c86f900160cea41e876dbbe718ae242e68a1c077c.jpg) +Figure 3. Comparison of the proposed Global-aware Transformer Decoder architecture (c) with two alternative decoder architectures (a) and (b). For clarity, we have omitted the residual connections in the cross-attention and self-attention layers. + +consistency, we include a pooling layer to capture global pose features, strengthening the model's understanding of overall pose structure across generated poses. + +Global-aware Transformer Decoder. For efficient target pose generation, we propose a Transformer-based decoder architecture capable of iterative pose generation. In this design, we developed three distinct Transformer decoder architectures, as shown in detail in Fig. 3. The specifics of these architectures will be discussed in the ablation experiments and the Appendix. Unlike traditional convolu + +tional neural networks (CNNs) commonly applied in visual tasks, attention-based Transformers lack certain inductive biases inherent to conventional visual models [19]. To improve the decoder's ability to incorporate pose-specific inductive biases, we introduce a global-aware block that models the entire pose holistically. This global-aware block is a flexible, plug-and-play addition compatible with any Transformer architecture. Distinct from the decoder architectures (a) and (b), which employ cross-attention layers to capture semantic relationships. Specifically, we enhance the target pose feature $\mathcal{F}_p^N$ by computing a similarity matrix between $\mathcal{F}_p^N$ and $\mathcal{F}_{tp}^{1:N - 1}$ , which integrates relevant information from prior poses, ensuring consistency within the generated poses. Next, a self-attention layer and feed-forward network (FFN) further refine $\mathcal{F}_p^N$ to enhance its expressiveness. The global-aware block is formalized as follows: + +$$ +\mathcal {F} _ {o u t} = \phi (\mathcal {F} _ {p} ^ {N} \cdot \mathcal {F} _ {t p} ^ {N - 1 ^ {T}}) \cdot \mathcal {F} _ {t p} ^ {N - 1} + \mathcal {F} _ {p} ^ {N}, +$$ + +$$ +\mathcal {F} _ {o u t} = \mathrm {M H S A} (\tilde {\mathcal {F}} _ {o u t}) + \mathcal {F} _ {o u t}, +$$ + +$$ +\mathcal {F} _ {\text {o u t}} = \operatorname {F F N} \left(\hat {\mathcal {F}} _ {\text {o u t}}\right) + \mathcal {F} _ {\text {o u t}}, \tag {2} +$$ + +where $\phi = \frac{e^{s_{ij}}}{\sum_k e^{s_{ik}}}$ aggregates pose information. Here, MHSA denotes the multi-head self-attention layer. Finally, + +the reference image features $\mathcal{F}_{ref}$ interact with the globally enhanced target pose features $\mathcal{F}_{out}$ through a series of attention-based modules, producing the pose used as a conditional embedding input for the U-Net. + +# 3.3. Fine-grained Appearance Generation + +Latent Diffusion Model. Latent diffusion models (LDMs) [20] have recently gained prominence as robust alternatives to generative adversarial networks (GANs) for various generative tasks. Building upon the foundational Markov chain framework of DDPM [9], LDMs adhere to a dual-phase paradigm comprising progressive noise injection and learned reconstruction. As formalized in [20], the forward diffusion operator systematically perturbs input data through a predefined variance schedule: + +$$ +q \left(z _ {t} \mid z _ {t - 1}\right) = \mathcal {N} \left(z _ {t}; \sqrt {1 - \beta_ {t}} z _ {t - 1}, \beta_ {t} I\right), \tag {3} +$$ + +which transforms the latent distribution of the data $z_0$ into a noise distribution $z_{t}$ (e.g., Gaussian distribution). The Gaussian variance, $\beta_{t}$ , can be learned or derived from a predefined schedule. In the reverse diffusion process, LDMs learn to predict and remove noise, effectively reversing the transformation to reconstruct the original data distribution: + +$$ +p _ {\theta} \left(z _ {t - 1} \mid z _ {t}\right) = \mathcal {N} \left(z _ {t - 1}; \mu_ {\theta} \left(z _ {t}, t, c\right), \sigma^ {2} I\right), \tag {4} +$$ + +where $\mu_{\theta}$ denotes the learned mean parameterized by $\theta$ as a function of the current noise state $z_{t}$ , time step $t$ , and condition $c$ . A more detailed formulation of diffusion models is provided in the Appendix. + +In this paper, we utilize the powerful LDM model to generate human appearances. To achieve appearance guidance, we disentangle pose control and appearance guidance. Unlike prior approaches [11, 21], which rely on CLIP's [22] image encoder or a U-Net-like appearance encoder, we condition the diffusion model on multi-pose embeddings generated by the Global-aware Pose Generation module. Additionally, to ensure coherence across the poses and to generate the entire sequence simultaneously, we extend the 2D diffusion model to a 3D setting, inspired by [21]. For improved efficiency, we fine-tune only selected weights in the U-Net, as shown in Fig. 2. Here, the "lock" symbol denotes frozen weights, while the "flame" symbol indicates fine-tuned parameters. + +Appearance Adapter. To enable precise control over the character's appearance details, we introduce the Appearance Adapter (AD), shown in Fig. 2. The AD operates hierarchically between multi-scale upsampling blocks in the diffusion framework, extracting pose-agnostic appearance features from reference images through cross-scale feature learning. This architecture enables explicit disentanglement of high-level semantic understanding between appearance attributes and pose informations. + +Specifically, our implementation with $\mathrm{k} = 3$ Appearance Embedding (AE) layers is augmented by two convolutional layers for dimensional alignment. Adopting the zero-initialization strategy from [23], these convolutional layers progressively smooth feature representations while suppressing high-frequency noise, thereby allowing stable learning of fine-grained appearance patterns. The AE layers enhance feature refinement through global attention mechanisms, followed by a Feed-Forward Network (FFN) that non-linearly reinforces the discriminative power of the processed features. + +# 3.4. Optimization + +We train our proposed model in an end-to-end manner. Our overall optimization objective is: + +$$ +\mathcal {L} _ {\text {t o t a l}} = \mathcal {L} _ {m g} + \mathcal {L} _ {\text {r e c}}, \tag {5} +$$ + +where $\mathcal{L}_{mg}$ is used to optimize the multi-pose generation process. The optimization objective is: + +$$ +\begin{array}{l} \mathcal {L} _ {m g} = \mathbb {E} _ {z _ {0} ^ {1: N}, \mathcal {X} _ {r e f}, \mathcal {P} ^ {1: N}} [ \| \epsilon - \\ \epsilon_ {\theta} \left(z _ {t} ^ {1: N}, t, \varphi \left(\mathcal {F} _ {r e f}, \mathcal {F} _ {p} ^ {1: N}\right)\right) | | _ {2} ^ {2} ], \tag {6} \\ \end{array} +$$ + +where $\mathcal{X}_{ref}$ denote the reference image, with $\mathcal{F}_{ref}$ representing its features. The target poses are given by $\mathcal{P}^{1:N}$ , and $\mathcal{F}_p^{1:N} = \mathcal{M}_p(\mathcal{P}^{1:N})$ represents the features of the target poses. Additionally, $z_0^{1:N} = \mathcal{E}(\mathcal{V}^{1:N})$ denotes the initial latent features, while $z_t^{1:N}$ is the noisy latent representation derived from $z_0^{1:N}$ at timestep $t$ , where $\mathcal{E}$ is VAE encoder. In addition, inspired by [5], we introduced an additional reconstruction task for the reference image during training to ensure the consistency between the appearance of the generated target poses and the reference image. The optimization objective of the reconstruction task is: + +$$ +\mathcal {L} _ {r e c} = \mathbb {E} _ {z _ {0}, \chi_ {r e f}, \mathcal {P} ^ {0}} [ | | \epsilon - \epsilon_ {\theta} (z _ {t}, t, \mathcal {F} _ {r e f})) | | _ {2} ^ {2} ]. \qquad (7) +$$ + +Sampling. At inference time, we sample $z_{t}$ from the Gaussian noise and iteratively sample from the previous step until $z_{0}$ , then we decode $z_{0}$ to image space get the final human pose image. To improve the sample quality during the generation process, we use a classifier-free guidance strategy. To implement classifier-free guidance, we randomly drop the reference image used as the appearance condition with a probability of $5\%$ during training. + +$$ +\hat {\epsilon} _ {\theta} \left(z _ {t}, t, \varphi\right) = \lambda \epsilon_ {\theta} \left(z _ {t}, t, \varphi\right) + (1 - \lambda) \epsilon_ {\theta} \left(z _ {t}, t\right), \tag {8} +$$ + +where $\lambda$ is a hyperparameter used to control the strength of guidance during the inference process, while $\epsilon_{\theta}(z_t,t,\varphi)$ and $\epsilon_{\theta}(z_t,t)$ represent conditional prediction and unconditional prediction, respectively. To handle the case when the reference image is missing, inspired by [10], we use a + +
ApproachesUBC FashionTikTok
SSIM ↑PSNR ↑FID ↓LPIPS ↓\( L_1 \) ↓SSIM ↑PSNR ↑FID ↓LPIPS ↓\( L_1 \) ↓
PISE [8] (CVPR'21)0.86717.11618.6570.07520.0460.4299.947120.6480.30100.186
DPTN [5] (CVPR'22)0.83315.93719.9570.08760.0500.1885.696143.1100.45170.449
NTED [7] (CVPR'22)0.82415.68319.7840.09480.0530.3398.551140.3030.37270.228
BiGraphGAN [4] (IJCV'23)0.84416.40121.0120.08060.0470.4459.916147.5210.26760.177
PIDM [12] (CVPR'23)0.81218.87014.9510.16350.0920.3077.558130.3920.42770.262
CFLD [10] (CVPR'24)0.88922.14212.8100.07180.0260.72917.520105.6310.21360.078
MagicPose [11] (ICML'24)0.83120.74214.6490.08660.0300.67717.260108.0670.25070.083
Ours0.94125.65511.4870.04970.0140.75318.466103.7080.23110.072
+ +![](images/350dbc4f2801db176651fffe8185f801abfb2c9134521767c98eff0684cc875f.jpg) +Figure 4. The pose synthesis accuracy of the proposed method on the UBC Fashion [13] dataset as the number of generated images increases, with 8 as the training setting. + +Table 1. Quantitative comparison of pose sequence synthesis quality on the UBC Fashion [13], and TikTok [14] dataset. + +
MethodsPIDM [12] (CVPR'23)CFLD [10] (CVPR'24)MagicPose [11] (ICML'24)Ours(8-24)
#Param131M248M2081M367M
Inference Time28.64s35.8s67.8s5.5-24.1s
+ +Table 2. Comparison of parameter count and inference time between the proposed method and other SOTA diffusion models. + +learnable vector as the conditional embedding. We concatenate the learnable vector with the reference image features to form a compact conditional embedding for the generation process, encouraging the learnable vector to capture more information beyond the reference image during training. This allows the model to receive additional appearance information when predicting invisible regions. + +# 4. Experiments + +# 4.1. Setup + +Dataset. We validate the effectiveness of our proposed method on both the UBC Fashion [13] and TikTok [14] datasets. The UBC Fashion dataset is a collection of data featuring various poses and outfits, containing 500 training and 100 test videos. Compared to UBC Fashion, the TikTok dataset has more complex backgrounds and greater diversity in the proportions of people, making the task more challenging. The TikTok dataset comprises 350 dance videos featuring individual performers, each lasting between 10 and 15 seconds. To ensure a fair comparison with state-of-the-art methods, we use the same test set as [11] for evaluation. We set the image resolution to $256 \times 256$ and randomly sample $N$ frames (in this paper, $N$ is set to 8) from each training video for training. Additionally, we use Open- + +Pose [24] and PGN [25] to extract human poses and human parsing maps. Due to category redundancy in the human parsing maps, we reduce the number of categories from 20 to 8 as [8]. + +Metrics. To evaluate the quality of the generated pose sequences, we report several metrics: Structural Similarity Index Measure (SSIM) [26], Peak Signal to Noise Ratio (PSNR), Learned Perceptual Image Patch Similarity (LPIPS) [27], and Fréchet Inception Distance (FID) [28]. SSIM and PSNR assess pixel-level similarity by measuring statistical differences between the generated images and the ground truth (GT). However, these metrics are limited to low-level image properties and may not fully capture perceptual quality. To address this limitation, we also include LPIPS and FID, which evaluate deep feature-based similarities. FID evaluates the distribution differences between the generated images and GT using Inception-v3 [29] features, while LPIPS measures the perceptual similarity between images using a VGG [30] network, making it more aligned with human perception of image similarity. + +Implementation details. The proposed method is implemented based on PyTorch [31] and Diffusers [32], trained in an end-to-end manner with a batch size of 16 across 4 NVIDIA A100 GPUs. The model was trained for 100 epochs utilizing the Adam [33] optimizer, starting with an initial learning rate of $5 \times 10^{-5}$ . A learning rate warmup strategy was implemented for the first 1000 training steps, followed by a decay to $1 \times 10^{-5}$ at the 50th epoch. To accelerate model convergence, we initialize the image encoder using Swin-B [17] pretrained on ImageNet [34] and the U-net using version 1.5 of Stable Diffusion [20]. Additionally, to address the issue of poor facial realism in generated pose, we followed strategy [20] by fine-tuning the decoder of the VAE using both MSE reconstruction loss and perceptual loss [35]. During the fine-tuning process, we placed greater emphasis on the MSE reconstruction loss. Further details on the fine-tuning process can be found in the Appendix. For classifier-free guidance, we set the parameter $\lambda$ to 2.0 during sampling, and the reference condition $\mathcal{X}_{ref}$ was dropped with a probability of $5\%$ throughout training. + +![](images/940da2dad2d15f7a32bc66e786cb4110fe0a69046f25020bc1e07820d96020fc.jpg) +Figure 5. Qualitative comparison of our method with SOTA methods on UBC Fashion [13] and TikTok[14] dataset. Our approach can generate more realistic details consistent with the reference image, such as garment folds, fine-grained textures, and facial identity features. + +![](images/16720a2c89453f62e1f068e850405c517837a50e90b36ae484ed2234a30ec063.jpg) + +![](images/f255f3d716debeb8e1f22c3b7641bf930997bdde57013ac8a2d5c77c8cb8b3ca.jpg) + +![](images/193a2c86f2ef0f2099d29c0557c1fb5091c6a78d8ca6cdea9ea36ed502c25063.jpg) + +![](images/c6361b96483922b507213ca94bbbd03bcafc642cfa6517337a5e80f5af8a2238.jpg) + +![](images/ff74e153d5ac47b00c93ee7b9674ff31c1dba7812ca51c67387f32c2c8d2d400.jpg) + +![](images/5016a54b521fd3e4d512c5dcd7604d75f527278c4b9f963211e847fedc2915f2.jpg) + +# 4.2. Quantitative Comparison + +We conducted a quantitative comparison of our method against state-of-the-art GAN-based and diffusion-based models on the UBC Fashion [13] and TikTok [14] datasets, as shown in Tab. 1. On the majority of metrics, our method outperforms the advanced methods on both datasets. Specifically, compared with four other diffusion models [10-12, 36], our approach achieves more stable generation performance and higher efficiency, with specific results shown in the Tab. 2, owing to the disentangling of pose control from appearance generation. Additionally, this design enables us to generate more human poses during inference (up to 3 times the training quantity) without additional training. The Fig. 4 illustrates the performance of our method as the generation count gradually increases. As shown in this figure, although an increase in generation quantity may lead to a slight performance decline, the results still serve as acceptable human image outputs, outperforming other SOTA methods. + +# 4.3. Qualitative Comparison + +The qualitative comparison with state-of-the-art methods is illustrated in Fig. 5 and 6. When using these other meth + +ods, both GAN-based and diffusion-based approaches show overfitting issues, resulting in the loss of person identity and clothing texture details. Additionally, when generating more complex clothing or unseen regions, issues like clothing blurriness, unnatural texture transitions, and substantial differences from the source image become apparent. Benefiting from our proposed strategy of disentangling pose control and appearance generation, our method mitigates overfitting to some extent. Furthermore, with the high-consistency pose embedding provided by Global-aware Pose Generation and the fine-grained, multi-scale features from Appearance Adapter, our approach yields more realistic results in maintaining person identity and capturing intricate clothing patterns. In particular, Fig. 6 provides an additional qualitative comparison with existing SOTA diffusion models for generating multiple poses. This scenario places higher demands on cross-view consistency in person identity and appearance. As observed in the Fig. 6, our method achieves higher consistency across different viewpoints of generated person images, thanks to Global-aware Transformer Decoder's global pose awareness in the pose generation process. More qualitative comparisons can be found in the Appendix. + +![](images/88735d997250453678ac179b5f765a723919163e1ab74e3f11e0c560697d60d4.jpg) +Figure 6. Qualitative comparison with other SOTA methods on various poses, with drawbacks highlighted by red circles. + +
ModelSSIM ↑PSNR ↑FID ↓LPIPS ↓
w/o Global-aware Pose Generation (GPG)0.936225.02411.8650.0532
w/o Multi-stage Image Encoder (MIE)0.934424.89511.2840.0554
w/o Appearance Adapter (AD)0.938425.20511.9340.0517
w/o Fine-tune VAE (FVAE)0.938625.46411.6270.0514
w/o Lrec0.940325.53911.5470.0516
Ours + (a)0.938925.44311.4870.0503
Ours + (b)0.938125.44711.5720.0506
Ours0.941125.65511.4870.0497
+ +# 4.4. Ablation Study + +Tab. 3 and Fig. 7 show quantitative and qualitative results, respectively. It can be observed that both GPG and AD are essential for pose synthesis performance, with their absence potentially leading to a performance drop of up to $3.8\%$ (FID). This highlights the importance of separating pose control and appearance generation strategies. We also replaced MIE with a CLIP image encoder, resulting in a $0.7\%$ (SSIM) and $2.9\%$ (PSNR) decline, though with an improvement in FID performance. As shown in the first row of Fig. 7, the CLIP image encoder fails to retain intricate clothing details, leading to a loss of appearance-related information. Additionally, fine-tuning the VAE for the specific dataset helps preserve realistic details of faces and clothing to a certain degree. Finally, we also experimented with two alternative Transformer decoder architectures in Fig. 3, both of which led to varying degrees of performance decline. Detailed ablation experiments can be found in the Appendix. + +![](images/2eeb1baa38a4468e016a694a995f14930a2fabb6846ce9ffc262718546e84fac.jpg) +Figure 7. Qualitative ablation study on the UBC Fashion [13] dataset. + +Table 3. Ablation study of different components, where (a) and (b) represent different decoder architectures in Fig. 3. + +
MethodsPIDM [12] (CVPR'23)CFLD [10] (CVPR'24)MagicPose [11] (ICML'24)Ours
Average13.6%16.7%8.3%61.4%
+ +Table 4. User study on the UBC Fashion [13] dataset. We followed [11] to collect vote data from 30 volunteers. The participants believed that our proposed method can generate visually realistic and identity-consistent human poses. + +# 4.5. User Study + +To more accurately evaluate the effectiveness of our proposed method, we designed a user study. In this experiment, we collected images of 9 subjects from the test set for comparative evaluation. One subject was designated as a warm-up, while the remaining eight were used as formal test. For each subject, eight distinct pose images were randomly selected to comparison. We presented images generated by various methods [10-12] to users, instructing them to evaluate the images based on visual quality, pose accuracy, and detail consistency, and to select the method they preferred. We present the averaged vote result in Tab. 4. + +# 5. Conclusions + +In this paper, we propose a unified framework for diverse multi-pose generation driven by disentangled pose and appearance guidance. This framework separately addresses complex human pose control and appearance generation to produce realistic and varied human poses. To achieve precise pose control and realistic appearance generation, we introduce Global-aware Pose Generation and Appearance Adapter module. These modules enhance interactions among poses during the generation process, leveraging all available mutual information to provide global guidance, while also learning high-level semantic information from multi-scale features to preserve texture details as much as possible. Extensive experiments demonstrate that our proposed approach achieves significant advantages over other state-of-the-art methods. + +Acknowledgments. This work was supported by the National Natural Science Foundation of China (62036006, 62276200) and the Tianjin Natural Science Foundation (24JCYBJC00310). + +# References + +[1] Grigore C Burdea and Philippe Coiffet. *Virtual reality technology*. John Wiley & Sons, 2003. 2 +[2] Julie Carmigniani, Borko Furht, Marco Anisetti, Paolo Ceravolo, Ernesto Damiani, and Misa Ivkovic. Augmented reality technologies, systems and applications. Multimedia tools and applications, 51:341-377, 2011. 2 +[3] Liang Zheng, Liyue Shen, Lu Tian, Shengjin Wang, Jingdong Wang, and Qi Tian. Scalable person re-identification: A benchmark. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 1116-1124, 2015. 2 +[4] Hao Tang, Ling Shao, Philip HS Torr, and Nicu Sebe. Bipartite graph reasoning gans for person pose and facial image synthesis. International Journal of Computer Vision, 131(3):644-658, 2023. 2, 6 +[5] Pengze Zhang, Lingxiao Yang, Jian-Huang Lai, and Xiaohua Xie. Exploring dual-task correlation for pose guided person image generation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 7713-7722, 2022. 2, 5, 6 +[6] Yuelong Li, Tengfei Xiao, Lei Geng, and Jianming Wang. Direct may not be the best: An incremental evolution view of pose generation. In Proceedings of the AAAI Conference on Artificial Intelligence, volume 38, pages 3270-3278, 2024. 2 +[7] Yurui Ren, Xiaqing Fan, Ge Li, Shan Liu, and Thomas H Li. Neural texture extraction and distribution for controllable person image synthesis. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 13535-13544, 2022. 2, 6 +[8] Jinsong Zhang, Kun Li, Yu-Kun Lai, and Jingyu Yang. Pise: Person image synthesis and editing with decoupled gan. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 7982-7990, 2021. 2, 6 +[9] Jonathan Ho, Ajay Jain, and Pieter Abbeel. Denoising diffusion probabilistic models. Advances in Neural Information Processing Systems, 33:6840-6851, 2020. 2, 3, 5 +[10] Yanzuo Lu, Manlin Zhang, Andy J Ma, Xiaohua Xie, and Jianhuang Lai. Coarse-to-fine latent diffusion for pose-guided person image synthesis. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 6420-6429, 2024. 2, 5, 6, 7, 8 +[11] Di Chang, Yichun Shi, Quankai Gao, Hongyi Xu, Jessica Fu, Guoxian Song, Qing Yan, Yizhe Zhu, Xiao Yang, and Mohammad Soleymani. Magicpose: Realistic human poses and facial expressions retargeting with identity-aware diffusion. In Proceedings of the International Conference on Machine Learning, 2023. 3, 5, 6, 8 +[12] Ankan Kumar Bhunia, Salman Khan, Hisham Cholakkal, Rao Muhammad Anwer, Jorma Laaksonen, Mubarak Shah, and Fahad Shahbaz Khan. Person image synthesis via denoising diffusion model. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 5968-5976, 2023. 2, 3, 6, 7, 8 +[13] Polina Zablotskaia, Aliaksandr Siarohin, Bo Zhao, and Leonid Sigal. Dwnet: Dense warp-based network for + +pose-guided human video generation. arXiv preprint arXiv:1910.09139, 2019. 2, 6, 7, 8 +[14] Yasamin Jafarian and Hyun Soo Park. Learning high fidelity depths of dressed humans by watching social media dance videos. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 12753-12762, 2021. 2, 6, 7 +[15] Ian Goodfellow, Jean Pouget-Abadie, Mehdi Mirza, Bing Xu, David Warde-Farley, Sherjil Ozair, Aaron Courville, and Yoshua Bengio. Generative adversarial networks. Communications of the ACM, 63(11):139-144, 2020. 2 +[16] Jonathan Ho and Tim Salimans. Classifier-free diffusion guidance. In NeurIPS 2021 Workshop on Deep Generative Models and Downstream Applications, 2021. 3 +[17] Ze Liu, Yutong Lin, Yue Cao, Han Hu, Yixuan Wei, Zheng Zhang, Stephen Lin, and Baining Guo. Swin transformer: Hierarchical vision transformer using shifted windows. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 10012-10022, 2021. 3, 6 +[18] Kaiming He, Xiangyu Zhang, Shaoqing Ren, and Jian Sun. Deep residual learning for image recognition. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 770-778, 2016. 3 +[19] Xumin Yu, Yongming Rao, Ziyi Wang, Zuyan Liu, Jiwen Lu, and Jie Zhou. Pointr: Diverse point cloud completion with geometry-aware transformers. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 12498-12507, 2021. 4 +[20] Robin Rombach, Andreas Blattmann, Dominik Lorenz, Patrick Esser, and Björn Ommer. High-resolution image synthesis with latent diffusion models. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 10684-10695, 2022. 5, 6 +[21] Zhongcong Xu, Jianfeng Zhang, Jun Hao Liew, Hanshu Yan, Jia-Wei Liu, Chenxu Zhang, Jiashi Feng, and Mike Zheng Shou. Magicanimate: Temporally consistent human image animation using diffusion model. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 1481–1490, 2024. 5 +[22] Alec Radford, Jong Wook Kim, Chris Hallacy, Aditya Ramesh, Gabriel Goh, Sandhini Agarwal, Girish Sastry, Amanda Askell, Pamela Mishkin, Jack Clark, et al. Learning transferable visual models from natural language supervision. In Proceedings of the International Conference on Machine Learning, pages 8748-8763. PMLR, 2021. 5 +[23] Lvmin Zhang, Anyi Rao, and Maneesh Agrawala. Adding conditional control to text-to-image diffusion models. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 3836-3847, 2023. 5 +[24] Riza Alp Güler, Natalia Neverova, and Iasonas Kokkinos. Densesepose: Dense human pose estimation in the wild. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 7297-7306, 2018. 6 +[25] Ke Gong, Xiaodan Liang, Yicheng Li, Yimin Chen, Ming Yang, and Liang Lin. Instance-level human parsing via part grouping network. In Proceedings of the European Conference on Computer Vision, pages 770-785, 2018. 6 + +[26] Zhou Wang, Alan C Bovik, Hamid R Sheikh, and Eero P Simoncelli. Image quality assessment: from error visibility to structural similarity. IEEE transactions on image processing, 13(4):600-612, 2004. 6 +[27] Richard Zhang, Phillip Isola, Alexei A Efros, Eli Shechtman, and Oliver Wang. The unreasonable effectiveness of deep features as a perceptual metric. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 586-595, 2018. 6 +[28] Martin Heusel, Hubert Ramsauer, Thomas Unterthiner, Bernhard Nessler, and Sepp Hochreiter. Gans trained by a two time-scale update rule converge to a local nash equilibrium. Advances in Neural Information Processing Systems, 30, 2017. 6 +[29] Christian Szegedy, Vincent Vanhoucke, Sergey Ioffe, Jon Shlens, and Zbigniew Wojna. Rethinking the inception architecture for computer vision. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 2818-2826, 2016. 6 +[30] Karen Simonyan and Andrew Zisserman. Very deep convolutional networks for large-scale image recognition. arXiv preprint arXiv:1409.1556, 2014. 6 +[31] Adam Paszke, Sam Gross, Francisco Massa, Adam Lerer, James Bradbury, Gregory Chanan, Trevor Killeen, Zeming Lin, Natalia Gimelshein, Luca Antiga, et al. Pytorch: An imperative style, high-performance deep learning library. Advances in Neural Information Processing Systems, 32, 2019. 6 +[32] Patrick von Platen, Suraj Patil, Anton Lozhkov, Pedro Cuenca, Nathan Lambert, Kashif Rasul, Mishig Davaadorj, Dhruv Nair, Sayak Paul, William Berman, Yiyi Xu, Steven Liu, and Thomas Wolf. Diffusers: State-of-the-art diffusion models. https://github.com/huggingface/diffusers, 2022.6 +[33] Diederik P Kingma. Adam: A method for stochastic optimization. arXiv preprint arXiv:1412.6980, 2014. 6 +[34] Jia Deng, Wei Dong, Richard Socher, Li-Jia Li, Kai Li, and Li Fei-Fei. Imagenet: A large-scale hierarchical image database. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition, pages 248–255. IEEE, 2009. 6 +[35] Justin Johnson, Alexandre Alahi, and Li Fei-Fei. Perceptual losses for real-time style transfer and super-resolution. In Computer Vision-ECCV 2016: 14th European Conference, Amsterdam, The Netherlands, October 11-14, 2016, Proceedings, Part II 14, pages 694-711. Springer, 2016. 6 +[36] Tan Wang, Linjie Li, Kevin Lin, Yuanhao Zhai, Chung-Ching Lin, Zhengyuan Yang, Hanwang Zhang, Zicheng Liu, and Lijuan Wang. Disco: Disentangled control for realistic human dance generation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 9326-9336, 2024. 7 \ No newline at end of file diff --git a/2025/Disentangled Pose and Appearance Guidance for Multi-Pose Generation/images.zip b/2025/Disentangled Pose and Appearance Guidance for Multi-Pose Generation/images.zip new file mode 100644 index 0000000000000000000000000000000000000000..52e6eb830f82127340e1430179d0dd3f2301de2d --- /dev/null +++ b/2025/Disentangled Pose and Appearance Guidance for Multi-Pose Generation/images.zip @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:b53c73072ea0fa1718cdcc0a45d2f4ffdf78f0d53321e17936783470defab0af +size 702333 diff --git a/2025/Disentangled Pose and Appearance Guidance for Multi-Pose Generation/layout.json b/2025/Disentangled Pose and Appearance Guidance for Multi-Pose Generation/layout.json new file mode 100644 index 0000000000000000000000000000000000000000..4710f3398d96c1db7f5348528f5c4f7feb518c76 --- /dev/null +++ b/2025/Disentangled Pose and Appearance Guidance for Multi-Pose Generation/layout.json @@ -0,0 +1,7105 @@ +{ + "pdf_info": [ + { + "para_blocks": [ + { + "bbox": [ + 82, + 103, + 529, + 121 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 82, + 103, + 529, + 121 + ], + "spans": [ + { + "bbox": [ + 82, + 103, + 529, + 121 + ], + "type": "text", + "content": "Disentangled Pose and Appearance Guidance for Multi-Pose Generation" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 62, + 141, + 548, + 242 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 141, + 548, + 242 + ], + "spans": [ + { + "bbox": [ + 62, + 141, + 548, + 242 + ], + "type": "text", + "content": "Tengfei Xiao" + }, + { + "bbox": [ + 62, + 141, + 548, + 242 + ], + "type": "inline_equation", + "content": "^{1,4}" + }, + { + "bbox": [ + 62, + 141, + 548, + 242 + ], + "type": "text", + "content": ", Yue Wu" + }, + { + "bbox": [ + 62, + 141, + 548, + 242 + ], + "type": "inline_equation", + "content": "^{\\dagger,*,1,4}" + }, + { + "bbox": [ + 62, + 141, + 548, + 242 + ], + "type": "text", + "content": ", Yuelong Li" + }, + { + "bbox": [ + 62, + 141, + 548, + 242 + ], + "type": "inline_equation", + "content": "^{\\dagger,*,2,5}" + }, + { + "bbox": [ + 62, + 141, + 548, + 242 + ], + "type": "text", + "content": ", Can Qin" + }, + { + "bbox": [ + 62, + 141, + 548, + 242 + ], + "type": "inline_equation", + "content": "^{3}" + }, + { + "bbox": [ + 62, + 141, + 548, + 242 + ], + "type": "text", + "content": ", Maoguo Gong" + }, + { + "bbox": [ + 62, + 141, + 548, + 242 + ], + "type": "inline_equation", + "content": "^{1,4}" + }, + { + "bbox": [ + 62, + 141, + 548, + 242 + ], + "type": "text", + "content": ", Qiguang Miao" + }, + { + "bbox": [ + 62, + 141, + 548, + 242 + ], + "type": "inline_equation", + "content": "^{1,4}" + }, + { + "bbox": [ + 62, + 141, + 548, + 242 + ], + "type": "text", + "content": ", Wenping Ma" + }, + { + "bbox": [ + 62, + 141, + 548, + 242 + ], + "type": "inline_equation", + "content": "^{1}" + }, + { + "bbox": [ + 62, + 141, + 548, + 242 + ], + "type": "inline_equation", + "content": "^{1}" + }, + { + "bbox": [ + 62, + 141, + 548, + 242 + ], + "type": "text", + "content": "Xidian University " + }, + { + "bbox": [ + 62, + 141, + 548, + 242 + ], + "type": "inline_equation", + "content": "^{2}" + }, + { + "bbox": [ + 62, + 141, + 548, + 242 + ], + "type": "text", + "content": "Tiangong University " + }, + { + "bbox": [ + 62, + 141, + 548, + 242 + ], + "type": "inline_equation", + "content": "^{3}" + }, + { + "bbox": [ + 62, + 141, + 548, + 242 + ], + "type": "text", + "content": "Northeastern University " + }, + { + "bbox": [ + 62, + 141, + 548, + 242 + ], + "type": "inline_equation", + "content": "^{4}" + }, + { + "bbox": [ + 62, + 141, + 548, + 242 + ], + "type": "text", + "content": "MoE Key Lab of Collaborative Intelligence Systems, Xidian University \n" + }, + { + "bbox": [ + 62, + 141, + 548, + 242 + ], + "type": "inline_equation", + "content": "^{5}" + }, + { + "bbox": [ + 62, + 141, + 548, + 242 + ], + "type": "text", + "content": "Tianjin Key Laboratory of Autonomous Intelligence Technology and Systems, Tiangong University tfxiao@stu.xidian.edu.cn, {ywu, qgmiao}@xidian.edu.cn, liyuelong@pku.edu.cn, qin.ca@northeastern.edu, gong@ieee.org, wpma@mail.xidian.edu.cn" + } + ] + } + ], + "index": 3 + }, + { + "type": "image", + "bbox": [ + 61, + 247, + 551, + 500 + ], + "blocks": [ + { + "bbox": [ + 61, + 247, + 551, + 500 + ], + "lines": [ + { + "bbox": [ + 61, + 247, + 551, + 500 + ], + "spans": [ + { + "bbox": [ + 61, + 247, + 551, + 500 + ], + "type": "image", + "image_path": "0206018f2f10f04df1d5cff292ffedde2c0bf9b7ab97c5366686992084d8ca15.jpg" + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 63, + 502, + 546, + 514 + ], + "lines": [ + { + "bbox": [ + 63, + 502, + 546, + 514 + ], + "spans": [ + { + "bbox": [ + 63, + 502, + 546, + 514 + ], + "type": "text", + "content": "Figure 1. Given a reference image and different target poses, our method can generate multiple images of the person in various poses." + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "image_caption" + } + ], + "index": 4 + }, + { + "bbox": [ + 152, + 524, + 200, + 536 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 152, + 524, + 200, + 536 + ], + "spans": [ + { + "bbox": [ + 152, + 524, + 200, + 536 + ], + "type": "text", + "content": "Abstract" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 55, + 550, + 296, + 682 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 550, + 296, + 682 + ], + "spans": [ + { + "bbox": [ + 55, + 550, + 296, + 682 + ], + "type": "text", + "content": "Human pose generation is a complex task due to the non-rigid and highly variable nature of human body structures and appearances. However, existing methods often overlook the fundamental differences between spatial transformations of poses and texture generation for appearance, which makes them prone to overfitting. To address this issue, we propose a multi-pose generation framework driven by disentangled pose and appearance guidance. Our approach includes a Global-aware Pose Generation module that iteratively generates pose embeddings, enabling effective control over non-rigid body deformations. Addition-" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 313, + 525, + 555, + 693 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 525, + 555, + 693 + ], + "spans": [ + { + "bbox": [ + 313, + 525, + 555, + 693 + ], + "type": "text", + "content": "ally, we introduce the Global-aware Transformer Decoder, which leverages similarity queries and attention mechanisms to achieve spatial transformations and enhance pose consistency through a Global-aware block. In the appearance generation phase, we condition a diffusion model on pose embeddings produced in the initial stage and introduce an Appearance Adapter that extracts high-level contextual semantic information from multi-scale features, enabling further refinement of pose appearance textures and providing appearance guidance. Extensive experiments on the UBC Fashion and TikTok datasets demonstrate that our framework achieves state-of-the-art results in both quality and fidelity, establishing it as a powerful approach for complex pose generation tasks." + } + ] + } + ], + "index": 8 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "spans": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "type": "text", + "content": "CVF" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 145, + 0, + 496, + 37 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 145, + 0, + 496, + 37 + ], + "spans": [ + { + "bbox": [ + 145, + 0, + 496, + 37 + ], + "type": "text", + "content": "This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 69, + 693, + 212, + 703 + ], + "type": "page_footnote", + "angle": 0, + "lines": [ + { + "bbox": [ + 69, + 693, + 212, + 703 + ], + "spans": [ + { + "bbox": [ + 69, + 693, + 212, + 703 + ], + "type": "inline_equation", + "content": "\\dagger" + }, + { + "bbox": [ + 69, + 693, + 212, + 703 + ], + "type": "text", + "content": " Equal contributions in no particular order." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 69, + 703, + 150, + 712 + ], + "type": "page_footnote", + "angle": 0, + "lines": [ + { + "bbox": [ + 69, + 703, + 150, + 712 + ], + "spans": [ + { + "bbox": [ + 69, + 703, + 150, + 712 + ], + "type": "text", + "content": "* Corresponding author." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 295, + 748, + 316, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 316, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 316, + 757 + ], + "type": "text", + "content": "5646" + } + ] + } + ], + "index": 12 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 0 + }, + { + "para_blocks": [ + { + "bbox": [ + 56, + 71, + 136, + 84 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 71, + 136, + 84 + ], + "spans": [ + { + "bbox": [ + 56, + 71, + 136, + 84 + ], + "type": "text", + "content": "1. Introduction" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 55, + 91, + 295, + 270 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 91, + 295, + 270 + ], + "spans": [ + { + "bbox": [ + 55, + 91, + 295, + 270 + ], + "type": "text", + "content": "The inherent complexity of human body poses, with their non-rigid, dynamic structures and diverse appearances, makes pose generation a challenging task. This process requires transforming a person's reference pose into any target pose while preserving the individual's identity and appearance with high fidelity. This capability has substantial applications in areas such as virtual reality (VR) [1], augmented reality (AR) [2], and person re-identification [3]. Pose generation, however, involves two distinct challenges: effectively controlling the non-rigid deformations of human body structures and maintaining appearance consistency across different poses and viewpoints. Most existing methods directly generate human poses without addressing the fundamental differences between pose control and appearance generation." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 55, + 270, + 295, + 450 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 270, + 295, + 450 + ], + "spans": [ + { + "bbox": [ + 55, + 270, + 295, + 450 + ], + "type": "text", + "content": "Advancements in deep learning have introduced promising solutions to pose generation, primarily through two classes of techniques: GAN-based methods and diffusion-based methods. GAN-based methods [4-8] typically align the reference pose with the target pose directly, learning a simple mapping from the appearance features of the reference pose to the target pose, which results in poor generalization. Diffusion models [9], as parameterized Markov chains trained via variational inference, offer straightforward optimization objectives and training efficiency, making them a viable alternative to GANs. Recent works [10-12] have applied diffusion models to pose generation, achieving promising results. However, we find that these methods still face limitations in terms of generation quality, identity preservation, or efficiency." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 55, + 451, + 295, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 451, + 295, + 714 + ], + "spans": [ + { + "bbox": [ + 55, + 451, + 295, + 714 + ], + "type": "text", + "content": "In this paper, we present a novel framework for multi-posede generation based on disentangled pose and appearance guidance. In our approach, pose control primarily addresses spatial geometry, while appearance generation leverages semantic and contextual information. By assigning each task to a distinct module, the model efficiently addresses each objective independently. We introduce a Global-aware Pose Generation module that iteratively produces pose embeddings. Each embedding is a compact representation within the feature space, containing the spatially transformed pose along with appearance priors from the reference image. To generate pose embeddings, we propose the Global-aware Transformer Decoder (GTD), which employs similarity queries and attention mechanisms for spatial transformation of poses. An additional Global-aware block within the GTD models poses generated in previous iterations, integrating enriched global information to enhance pose consistency. During the appearance generation phase, the diffusion model is conditioned on multiple pose embeddings produced in the initial stage to refine appearance details. To further enhance appearance fidelity, we implement an Appearance Adapter module that encodes" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 313, + 72, + 553, + 107 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 72, + 553, + 107 + ], + "spans": [ + { + "bbox": [ + 313, + 72, + 553, + 107 + ], + "type": "text", + "content": "multi-scale features to extract high-level semantic information, providing fine-grained appearance guidance to the diffusion model." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 326, + 108, + 507, + 120 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 326, + 108, + 507, + 120 + ], + "spans": [ + { + "bbox": [ + 326, + 108, + 507, + 120 + ], + "type": "text", + "content": "In summary, our main contributions include:" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 313, + 121, + 553, + 277 + ], + "type": "list", + "angle": 0, + "index": 10, + "blocks": [ + { + "bbox": [ + 314, + 121, + 553, + 145 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 314, + 121, + 553, + 145 + ], + "spans": [ + { + "bbox": [ + 314, + 121, + 553, + 145 + ], + "type": "text", + "content": "- Introducing a novel multi-pose generation framework based on disentangled pose and appearance guidance." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 313, + 146, + 553, + 193 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 146, + 553, + 193 + ], + "spans": [ + { + "bbox": [ + 313, + 146, + 553, + 193 + ], + "type": "text", + "content": "- Developing a global-aware pose generation module that models all generated poses, facilitating feature interactions among poses and integrating global information for improved pose generation." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 313, + 194, + 553, + 239 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 194, + 553, + 239 + ], + "spans": [ + { + "bbox": [ + 313, + 194, + 553, + 239 + ], + "type": "text", + "content": "- Designing an appearance adapter module to extract high-level contextual semantic information from multi-scale features, providing fine-grained appearance guidance to the diffusion model." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 313, + 241, + 553, + 277 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 241, + 553, + 277 + ], + "spans": [ + { + "bbox": [ + 313, + 241, + 553, + 277 + ], + "type": "text", + "content": "- Conducting extensive experiments on the UBC Fashion [13] and TikTok [14] datasets, achieving state-of-the-art results in both quality and fidelity." + } + ] + } + ], + "index": 9 + } + ], + "sub_type": "text" + }, + { + "bbox": [ + 314, + 287, + 406, + 300 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 314, + 287, + 406, + 300 + ], + "spans": [ + { + "bbox": [ + 314, + 287, + 406, + 300 + ], + "type": "text", + "content": "2. Related Works" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 313, + 307, + 553, + 415 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 307, + 553, + 415 + ], + "spans": [ + { + "bbox": [ + 313, + 307, + 553, + 415 + ], + "type": "text", + "content": "Generative Adversarial Networks. Over the past decades, Generative Adversarial Networks (GANs) [15] have been extensively applied to human pose generation. Leveraging an adversarial learning framework, GANs can effectively model complex pose spatial transformations while producing realistic and detailed textures. Despite significant progress in recent years, challenges remain in predicting invisible regions, generating complex appearance textures, and maintaining identity consistency." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 313, + 415, + 553, + 546 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 415, + 553, + 546 + ], + "spans": [ + { + "bbox": [ + 313, + 415, + 553, + 546 + ], + "type": "text", + "content": "To achieve accurate human pose synthesis with consistent clothing appearance and style, PISE [8] introduces a two-stage generative model that decouples appearance control from pose generation, enabling the creation of realistic human images. BiGraphGAN [4] addresses the challenges of pose deformation by incorporating a bipartite graph reasoning (BGR) block, which captures long-range relations between source and target poses in a bipartite graph structure. Additionally, BiGraphGAN includes an interaction- and-aggregation (IA) block to improve the feature representation of both body shape and appearance." + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 313, + 546, + 553, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 546, + 553, + 713 + ], + "spans": [ + { + "bbox": [ + 313, + 546, + 553, + 713 + ], + "type": "text", + "content": "Recently, attention mechanisms have been integrated into pose generation tasks. DPTN [5] employs a transformer-based siamese network, which supports pose generation by reconstructing reference images, thereby maintaining consistency in the target pose's appearance. NTED [7] presents a neural texture extraction and distribution module based on dual attention, extracting semantic textures from source features and distributing them according to the spatial layout of the target pose. IEPG [6] proposes a progressive pose generation framework, designed with an incremental evolution approach to address pose synthesis challenges. This multi-step generation strategy effectively handles occlusions, significantly enhancing the realism and coherence of the generated appearances." + } + ] + } + ], + "index": 14 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "text", + "content": "5647" + } + ] + } + ], + "index": 15 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 1 + }, + { + "para_blocks": [ + { + "bbox": [ + 55, + 72, + 296, + 481 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 72, + 296, + 481 + ], + "spans": [ + { + "bbox": [ + 55, + 72, + 296, + 481 + ], + "type": "text", + "content": "Diffusion models. Diffusion models (DM) [9] have emerged as a promising alternative to GANs, decomposing the generation process into a series of forward-backward diffusion steps to learn plausible transfer trajectories. This iterative refinement, starting with a low-quality noisy signal and progressively enhancing it, enables the generation of high-quality, contextually consistent samples, making diffusion models particularly effective for generating high-fidelity results. PIDM [12] introduces a texture diffusion module based on cross-attention, which accurately models the correspondences between pose and appearance information at the pixel level across source and target images. Additionally, PIDM employs a disentangled classifier-free guidance strategy [16] to ensure the conditional input and generated output closely match both pose and appearance. While PIDM achieves state-of-the-art synthesis results, it operates directly in pixel space, which makes its optimization and inference computationally expensive. MagicPose [11] adopts a two-stage training strategy to separately handle character appearance control and pose control, disentangling the two processes. Additionally, it incorporates facial keypoints, finger keypoints, and sparse body keypoints to achieve more precise pose control. Although these methods have made promising progress, they condition the pose generation on the global features of the reference pose and delegate both the spatial transformation of the pose and appearance texture generation to the diffusion model, causing the diffusion model to handle two distinct tasks simultaneously. In this paper, we develop a novel paradigm for multi-pose generation that disentangles pose control from appearance guidance. This approach allows the model to focus on a single objective within each module, thereby improving the consistency and controllability of the generated human pose images." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 55, + 494, + 111, + 506 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 494, + 111, + 506 + ], + "spans": [ + { + "bbox": [ + 55, + 494, + 111, + 506 + ], + "type": "text", + "content": "3. Method" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 55, + 514, + 205, + 526 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 514, + 205, + 526 + ], + "spans": [ + { + "bbox": [ + 55, + 514, + 205, + 526 + ], + "type": "text", + "content": "3.1. Architecture and Overview" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 55, + 534, + 296, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 534, + 296, + 714 + ], + "spans": [ + { + "bbox": [ + 55, + 534, + 296, + 714 + ], + "type": "text", + "content": "Given a reference image " + }, + { + "bbox": [ + 55, + 534, + 296, + 714 + ], + "type": "inline_equation", + "content": "\\mathcal{X}_{ref}" + }, + { + "bbox": [ + 55, + 534, + 296, + 714 + ], + "type": "text", + "content": " and multiple target poses " + }, + { + "bbox": [ + 55, + 534, + 296, + 714 + ], + "type": "inline_equation", + "content": "\\mathcal{P}^{1:N} = [p_1,p_2\\dots ,p_N]" + }, + { + "bbox": [ + 55, + 534, + 296, + 714 + ], + "type": "text", + "content": ", where " + }, + { + "bbox": [ + 55, + 534, + 296, + 714 + ], + "type": "inline_equation", + "content": "N" + }, + { + "bbox": [ + 55, + 534, + 296, + 714 + ], + "type": "text", + "content": " is the number of target poses. The objective of our method is to generate a series of realistic images " + }, + { + "bbox": [ + 55, + 534, + 296, + 714 + ], + "type": "inline_equation", + "content": "\\mathcal{Y}^{1:N} = [y_1,y_2,\\dots,y_N]" + }, + { + "bbox": [ + 55, + 534, + 296, + 714 + ], + "type": "text", + "content": " while ensuring that both the poses and the appearance are consistent with the given poses " + }, + { + "bbox": [ + 55, + 534, + 296, + 714 + ], + "type": "inline_equation", + "content": "\\mathcal{P}^{1:N}" + }, + { + "bbox": [ + 55, + 534, + 296, + 714 + ], + "type": "text", + "content": " and the reference image " + }, + { + "bbox": [ + 55, + 534, + 296, + 714 + ], + "type": "inline_equation", + "content": "\\mathcal{X}_{ref}" + }, + { + "bbox": [ + 55, + 534, + 296, + 714 + ], + "type": "text", + "content": ", respectively. The overall pipeline of our approach is illustrated in Fig. 2. To address the distinct nature of spatial transformations in pose generation and the fine-grained synthesis of appearance textures, we propose disentangling these two processes. This separation enables independent optimization of pose control and appearance refinement, leading to more accurate and natural image generation. To implement this, we use a multi-stage image encoder, such as Swin Transformer [17], denoted as " + }, + { + "bbox": [ + 55, + 534, + 296, + 714 + ], + "type": "inline_equation", + "content": "\\mathcal{M}_{ref}" + }, + { + "bbox": [ + 55, + 534, + 296, + 714 + ], + "type": "text", + "content": ", to embed the ref" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 313, + 72, + 553, + 144 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 72, + 553, + 144 + ], + "spans": [ + { + "bbox": [ + 313, + 72, + 553, + 144 + ], + "type": "text", + "content": "erence image " + }, + { + "bbox": [ + 313, + 72, + 553, + 144 + ], + "type": "inline_equation", + "content": "\\mathcal{X}_{ref}" + }, + { + "bbox": [ + 313, + 72, + 553, + 144 + ], + "type": "text", + "content": " into multi-scale appearance embeddings " + }, + { + "bbox": [ + 313, + 72, + 553, + 144 + ], + "type": "inline_equation", + "content": "\\mathcal{F}_{ref} = [f_1, f_2, f_3, f_4]" + }, + { + "bbox": [ + 313, + 72, + 553, + 144 + ], + "type": "text", + "content": ". These embeddings allow the network to perform multi-scale appearance control. For pose control, we developed a lightweight pose control network " + }, + { + "bbox": [ + 313, + 72, + 553, + 144 + ], + "type": "inline_equation", + "content": "\\mathcal{M}_p" + }, + { + "bbox": [ + 313, + 72, + 553, + 144 + ], + "type": "text", + "content": ", based on ResNet [18], which extracts the pose features " + }, + { + "bbox": [ + 313, + 72, + 553, + 144 + ], + "type": "inline_equation", + "content": "\\mathcal{F}_p^{1:N}" + }, + { + "bbox": [ + 313, + 72, + 553, + 144 + ], + "type": "text", + "content": " from the multiple target poses." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 313, + 144, + 555, + 312 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 144, + 555, + 312 + ], + "spans": [ + { + "bbox": [ + 313, + 144, + 555, + 312 + ], + "type": "text", + "content": "Next, more importantly, we designed the Global-aware Pose Generation module, which is trained to generate target pose embedding " + }, + { + "bbox": [ + 313, + 144, + 555, + 312 + ], + "type": "inline_equation", + "content": "\\mathcal{F}_{tp}^{1:N}" + }, + { + "bbox": [ + 313, + 144, + 555, + 312 + ], + "type": "text", + "content": ". Unlike previous methods, we leverage the Global-aware Pose Generation module to perform spatial transformations of the pose while also encouraging the module to learn appearance priors from the reference image. This approach naturally embeds some appearance knowledge into the generated pose, simplifying subsequent appearance refinement or occluded region completion by the diffusion model. Additionally, to enhance the realism of textures, we employ a multi-scale Appearance Adapter (AD) module to provide multi-scale finer-grained appearance features to the U-net, further improving the quality of the generated images." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 313, + 318, + 480, + 330 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 318, + 480, + 330 + ], + "spans": [ + { + "bbox": [ + 313, + 318, + 480, + 330 + ], + "type": "text", + "content": "3.2. Global-aware Pose Generation" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 313, + 336, + 555, + 598 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 336, + 555, + 598 + ], + "spans": [ + { + "bbox": [ + 313, + 336, + 555, + 598 + ], + "type": "text", + "content": "Existing multi-person pose generation frameworks predominantly optimize individual target poses in isolation, which inherently limits their ability to model the spatiotemporal correlations among concurrent human motions. This limitation restricts these models' capacity to capture broader contextual information, potentially leading to inconsistencies or unnatural transitions between generated poses. To address this, we propose a unified module for multi-pose generation (Fig. 2) that leverages the Transformer architecture for its effective modeling capabilities, enabling coherent pose generation. Central to this module is the Global-aware Transformer Decoder (GTD) Block, which iteratively generates each target pose by integrating information from both the reference image and diverse target poses. Using the reference image features " + }, + { + "bbox": [ + 313, + 336, + 555, + 598 + ], + "type": "inline_equation", + "content": "f_{4}" + }, + { + "bbox": [ + 313, + 336, + 555, + 598 + ], + "type": "text", + "content": " from " + }, + { + "bbox": [ + 313, + 336, + 555, + 598 + ], + "type": "inline_equation", + "content": "\\mathcal{X}_{ref}" + }, + { + "bbox": [ + 313, + 336, + 555, + 598 + ], + "type": "text", + "content": " and the " + }, + { + "bbox": [ + 313, + 336, + 555, + 598 + ], + "type": "inline_equation", + "content": "N" + }, + { + "bbox": [ + 313, + 336, + 555, + 598 + ], + "type": "text", + "content": "-th target pose embeddings from " + }, + { + "bbox": [ + 313, + 336, + 555, + 598 + ], + "type": "inline_equation", + "content": "\\mathcal{P}^{N}" + }, + { + "bbox": [ + 313, + 336, + 555, + 598 + ], + "type": "text", + "content": ", the GTD Block outputs the target pose " + }, + { + "bbox": [ + 313, + 336, + 555, + 598 + ], + "type": "inline_equation", + "content": "\\mathcal{F}_{tp}^{N}" + }, + { + "bbox": [ + 313, + 336, + 555, + 598 + ], + "type": "text", + "content": " for the " + }, + { + "bbox": [ + 313, + 336, + 555, + 598 + ], + "type": "inline_equation", + "content": "N" + }, + { + "bbox": [ + 313, + 336, + 555, + 598 + ], + "type": "text", + "content": "-th pose. To ensure coherence across poses, we enhance cross-pose consistency by incorporating previously generated poses (i.e., " + }, + { + "bbox": [ + 313, + 336, + 555, + 598 + ], + "type": "inline_equation", + "content": "\\mathcal{F}_{tp}^{1:N-1}" + }, + { + "bbox": [ + 313, + 336, + 555, + 598 + ], + "type": "text", + "content": " in Eq. 1) into each subsequent generation step. The iterative process for generating multiple pose " + }, + { + "bbox": [ + 313, + 336, + 555, + 598 + ], + "type": "inline_equation", + "content": "\\mathcal{F}_{tp}^{1:N}" + }, + { + "bbox": [ + 313, + 336, + 555, + 598 + ], + "type": "text", + "content": " and the specific generation for the " + }, + { + "bbox": [ + 313, + 336, + 555, + 598 + ], + "type": "inline_equation", + "content": "N" + }, + { + "bbox": [ + 313, + 336, + 555, + 598 + ], + "type": "text", + "content": "-th GTD iteration are defined as follows:" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 361, + 604, + 553, + 673 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 361, + 604, + 553, + 673 + ], + "spans": [ + { + "bbox": [ + 361, + 604, + 553, + 673 + ], + "type": "interline_equation", + "content": "\\begin{array}{l} \\mathcal {F} _ {t p} ^ {1: N} = \\varphi (f _ {4}, \\mathcal {F} _ {p} ^ {1: N}), \\\\ \\mathcal {F} _ {t p} ^ {1: N} = \\mathcal {F} _ {t p} ^ {1: N} + \\frac {1}{N} \\sum_ {i = 1} ^ {N} \\mathcal {F} _ {t p} ^ {i}, \\\\ \\mathcal {F} _ {t p} ^ {N} = G T D \\left(f _ {4}, \\mathcal {F} _ {p} ^ {1: N}, \\mathcal {F} _ {t p} ^ {1: N - 1}\\right), \\tag {1} \\\\ \\end{array}", + "image_path": "a2c1156c2615d35f303de3baa252486ac406c336a121c48d0e04a5e9350ba61c.jpg" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 313, + 677, + 555, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 677, + 555, + 713 + ], + "spans": [ + { + "bbox": [ + 313, + 677, + 555, + 713 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 313, + 677, + 555, + 713 + ], + "type": "inline_equation", + "content": "\\varphi" + }, + { + "bbox": [ + 313, + 677, + 555, + 713 + ], + "type": "text", + "content": " represents the global-aware pose generation function, and " + }, + { + "bbox": [ + 313, + 677, + 555, + 713 + ], + "type": "inline_equation", + "content": "\\mathcal{F}_{tp}^{1:N - 1}" + }, + { + "bbox": [ + 313, + 677, + 555, + 713 + ], + "type": "text", + "content": " is the generated result from the previous " + }, + { + "bbox": [ + 313, + 677, + 555, + 713 + ], + "type": "inline_equation", + "content": "N - 1" + }, + { + "bbox": [ + 313, + 677, + 555, + 713 + ], + "type": "text", + "content": " iterations of GTD. To further reinforce contextual" + } + ] + } + ], + "index": 9 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "text", + "content": "5648" + } + ] + } + ], + "index": 10 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 2 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 58, + 61, + 555, + 345 + ], + "blocks": [ + { + "bbox": [ + 58, + 61, + 555, + 345 + ], + "lines": [ + { + "bbox": [ + 58, + 61, + 555, + 345 + ], + "spans": [ + { + "bbox": [ + 58, + 61, + 555, + 345 + ], + "type": "image", + "image_path": "ac61b8f68c98a2ff1c8b76e777bf3a308e8081337e6061079e0e416f5e096b74.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 54, + 354, + 555, + 409 + ], + "lines": [ + { + "bbox": [ + 54, + 354, + 555, + 409 + ], + "spans": [ + { + "bbox": [ + 54, + 354, + 555, + 409 + ], + "type": "text", + "content": "Figure 2. Overview of the architecture of the proposed pipeline. The Image Encoder encodes a reference image to obtain multi-scale features, while the Pose Encoder encodes multiple target poses. Pose control is achieved by the Global-aware Pose Generation module, which iteratively generates pose embeddings from the final layer features of the image encoder and the target pose features. A diffusion model then refines the pose appearance, conditioned on these pose embeddings. The Appearance Adapter further enhances appearance detail by refining the multi-scale features from the image encoder and injecting them into the diffusion model." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "type": "image", + "bbox": [ + 58, + 420, + 294, + 522 + ], + "blocks": [ + { + "bbox": [ + 58, + 420, + 294, + 522 + ], + "lines": [ + { + "bbox": [ + 58, + 420, + 294, + 522 + ], + "spans": [ + { + "bbox": [ + 58, + 420, + 294, + 522 + ], + "type": "image", + "image_path": "9f4b059a545121584618136c86f900160cea41e876dbbe718ae242e68a1c077c.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 55, + 530, + 295, + 575 + ], + "lines": [ + { + "bbox": [ + 55, + 530, + 295, + 575 + ], + "spans": [ + { + "bbox": [ + 55, + 530, + 295, + 575 + ], + "type": "text", + "content": "Figure 3. Comparison of the proposed Global-aware Transformer Decoder architecture (c) with two alternative decoder architectures (a) and (b). For clarity, we have omitted the residual connections in the cross-attention and self-attention layers." + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_caption" + } + ], + "index": 2 + }, + { + "bbox": [ + 55, + 590, + 296, + 626 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 590, + 296, + 626 + ], + "spans": [ + { + "bbox": [ + 55, + 590, + 296, + 626 + ], + "type": "text", + "content": "consistency, we include a pooling layer to capture global pose features, strengthening the model's understanding of overall pose structure across generated poses." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 55, + 630, + 296, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 630, + 296, + 713 + ], + "spans": [ + { + "bbox": [ + 55, + 630, + 296, + 713 + ], + "type": "text", + "content": "Global-aware Transformer Decoder. For efficient target pose generation, we propose a Transformer-based decoder architecture capable of iterative pose generation. In this design, we developed three distinct Transformer decoder architectures, as shown in detail in Fig. 3. The specifics of these architectures will be discussed in the ablation experiments and the Appendix. Unlike traditional convolu" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 313, + 422, + 555, + 615 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 422, + 555, + 615 + ], + "spans": [ + { + "bbox": [ + 313, + 422, + 555, + 615 + ], + "type": "text", + "content": "tional neural networks (CNNs) commonly applied in visual tasks, attention-based Transformers lack certain inductive biases inherent to conventional visual models [19]. To improve the decoder's ability to incorporate pose-specific inductive biases, we introduce a global-aware block that models the entire pose holistically. This global-aware block is a flexible, plug-and-play addition compatible with any Transformer architecture. Distinct from the decoder architectures (a) and (b), which employ cross-attention layers to capture semantic relationships. Specifically, we enhance the target pose feature " + }, + { + "bbox": [ + 313, + 422, + 555, + 615 + ], + "type": "inline_equation", + "content": "\\mathcal{F}_p^N" + }, + { + "bbox": [ + 313, + 422, + 555, + 615 + ], + "type": "text", + "content": " by computing a similarity matrix between " + }, + { + "bbox": [ + 313, + 422, + 555, + 615 + ], + "type": "inline_equation", + "content": "\\mathcal{F}_p^N" + }, + { + "bbox": [ + 313, + 422, + 555, + 615 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 313, + 422, + 555, + 615 + ], + "type": "inline_equation", + "content": "\\mathcal{F}_{tp}^{1:N - 1}" + }, + { + "bbox": [ + 313, + 422, + 555, + 615 + ], + "type": "text", + "content": ", which integrates relevant information from prior poses, ensuring consistency within the generated poses. Next, a self-attention layer and feed-forward network (FFN) further refine " + }, + { + "bbox": [ + 313, + 422, + 555, + 615 + ], + "type": "inline_equation", + "content": "\\mathcal{F}_p^N" + }, + { + "bbox": [ + 313, + 422, + 555, + 615 + ], + "type": "text", + "content": " to enhance its expressiveness. The global-aware block is formalized as follows:" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 349, + 623, + 518, + 641 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 349, + 623, + 518, + 641 + ], + "spans": [ + { + "bbox": [ + 349, + 623, + 518, + 641 + ], + "type": "interline_equation", + "content": "\\mathcal {F} _ {o u t} = \\phi (\\mathcal {F} _ {p} ^ {N} \\cdot \\mathcal {F} _ {t p} ^ {N - 1 ^ {T}}) \\cdot \\mathcal {F} _ {t p} ^ {N - 1} + \\mathcal {F} _ {p} ^ {N},", + "image_path": "67f11b152f44aa65cd1f3bdd7d175c6ae3b016fcc654c92edeb9bff5c6a767a5.jpg" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 350, + 643, + 476, + 657 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 350, + 643, + 476, + 657 + ], + "spans": [ + { + "bbox": [ + 350, + 643, + 476, + 657 + ], + "type": "interline_equation", + "content": "\\mathcal {F} _ {o u t} = \\mathrm {M H S A} (\\tilde {\\mathcal {F}} _ {o u t}) + \\mathcal {F} _ {o u t},", + "image_path": "cb08f56407dbd9dbe0333307552902aca06db4d478412eaa7e1e7e9a0e534449.jpg" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 350, + 660, + 555, + 673 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 350, + 660, + 555, + 673 + ], + "spans": [ + { + "bbox": [ + 350, + 660, + 555, + 673 + ], + "type": "interline_equation", + "content": "\\mathcal {F} _ {\\text {o u t}} = \\operatorname {F F N} \\left(\\hat {\\mathcal {F}} _ {\\text {o u t}}\\right) + \\mathcal {F} _ {\\text {o u t}}, \\tag {2}", + "image_path": "ace190ab8ce6baad3289fb182902a8491687426031136cdce13cfd16f6d954a4.jpg" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 313, + 683, + 554, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 683, + 554, + 713 + ], + "spans": [ + { + "bbox": [ + 313, + 683, + 554, + 713 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 313, + 683, + 554, + 713 + ], + "type": "inline_equation", + "content": "\\phi = \\frac{e^{s_{ij}}}{\\sum_k e^{s_{ik}}}" + }, + { + "bbox": [ + 313, + 683, + 554, + 713 + ], + "type": "text", + "content": " aggregates pose information. Here, MHSA denotes the multi-head self-attention layer. Finally," + } + ] + } + ], + "index": 10 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "text", + "content": "5649" + } + ] + } + ], + "index": 11 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 3 + }, + { + "para_blocks": [ + { + "bbox": [ + 55, + 72, + 296, + 120 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 72, + 296, + 120 + ], + "spans": [ + { + "bbox": [ + 55, + 72, + 296, + 120 + ], + "type": "text", + "content": "the reference image features " + }, + { + "bbox": [ + 55, + 72, + 296, + 120 + ], + "type": "inline_equation", + "content": "\\mathcal{F}_{ref}" + }, + { + "bbox": [ + 55, + 72, + 296, + 120 + ], + "type": "text", + "content": " interact with the globally enhanced target pose features " + }, + { + "bbox": [ + 55, + 72, + 296, + 120 + ], + "type": "inline_equation", + "content": "\\mathcal{F}_{out}" + }, + { + "bbox": [ + 55, + 72, + 296, + 120 + ], + "type": "text", + "content": " through a series of attention-based modules, producing the pose used as a conditional embedding input for the U-Net." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 55, + 127, + 254, + 140 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 127, + 254, + 140 + ], + "spans": [ + { + "bbox": [ + 55, + 127, + 254, + 140 + ], + "type": "text", + "content": "3.3. Fine-grained Appearance Generation" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 55, + 145, + 296, + 253 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 145, + 296, + 253 + ], + "spans": [ + { + "bbox": [ + 55, + 145, + 296, + 253 + ], + "type": "text", + "content": "Latent Diffusion Model. Latent diffusion models (LDMs) [20] have recently gained prominence as robust alternatives to generative adversarial networks (GANs) for various generative tasks. Building upon the foundational Markov chain framework of DDPM [9], LDMs adhere to a dual-phase paradigm comprising progressive noise injection and learned reconstruction. As formalized in [20], the forward diffusion operator systematically perturbs input data through a predefined variance schedule:" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 88, + 262, + 295, + 282 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 88, + 262, + 295, + 282 + ], + "spans": [ + { + "bbox": [ + 88, + 262, + 295, + 282 + ], + "type": "interline_equation", + "content": "q \\left(z _ {t} \\mid z _ {t - 1}\\right) = \\mathcal {N} \\left(z _ {t}; \\sqrt {1 - \\beta_ {t}} z _ {t - 1}, \\beta_ {t} I\\right), \\tag {3}", + "image_path": "1975b1bf7227dd298611bf5916c2e2ff74ae4156a4a93f406d80e584bf6e1540.jpg" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 55, + 288, + 296, + 361 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 288, + 296, + 361 + ], + "spans": [ + { + "bbox": [ + 55, + 288, + 296, + 361 + ], + "type": "text", + "content": "which transforms the latent distribution of the data " + }, + { + "bbox": [ + 55, + 288, + 296, + 361 + ], + "type": "inline_equation", + "content": "z_0" + }, + { + "bbox": [ + 55, + 288, + 296, + 361 + ], + "type": "text", + "content": " into a noise distribution " + }, + { + "bbox": [ + 55, + 288, + 296, + 361 + ], + "type": "inline_equation", + "content": "z_{t}" + }, + { + "bbox": [ + 55, + 288, + 296, + 361 + ], + "type": "text", + "content": " (e.g., Gaussian distribution). The Gaussian variance, " + }, + { + "bbox": [ + 55, + 288, + 296, + 361 + ], + "type": "inline_equation", + "content": "\\beta_{t}" + }, + { + "bbox": [ + 55, + 288, + 296, + 361 + ], + "type": "text", + "content": ", can be learned or derived from a predefined schedule. In the reverse diffusion process, LDMs learn to predict and remove noise, effectively reversing the transformation to reconstruct the original data distribution:" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 85, + 369, + 295, + 383 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 85, + 369, + 295, + 383 + ], + "spans": [ + { + "bbox": [ + 85, + 369, + 295, + 383 + ], + "type": "interline_equation", + "content": "p _ {\\theta} \\left(z _ {t - 1} \\mid z _ {t}\\right) = \\mathcal {N} \\left(z _ {t - 1}; \\mu_ {\\theta} \\left(z _ {t}, t, c\\right), \\sigma^ {2} I\\right), \\tag {4}", + "image_path": "abee16091e80ade0c30d2be49f8072fa2373d432ad7818caea500e050ac5fb46.jpg" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 55, + 391, + 296, + 438 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 391, + 296, + 438 + ], + "spans": [ + { + "bbox": [ + 55, + 391, + 296, + 438 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 55, + 391, + 296, + 438 + ], + "type": "inline_equation", + "content": "\\mu_{\\theta}" + }, + { + "bbox": [ + 55, + 391, + 296, + 438 + ], + "type": "text", + "content": " denotes the learned mean parameterized by " + }, + { + "bbox": [ + 55, + 391, + 296, + 438 + ], + "type": "inline_equation", + "content": "\\theta" + }, + { + "bbox": [ + 55, + 391, + 296, + 438 + ], + "type": "text", + "content": " as a function of the current noise state " + }, + { + "bbox": [ + 55, + 391, + 296, + 438 + ], + "type": "inline_equation", + "content": "z_{t}" + }, + { + "bbox": [ + 55, + 391, + 296, + 438 + ], + "type": "text", + "content": ", time step " + }, + { + "bbox": [ + 55, + 391, + 296, + 438 + ], + "type": "inline_equation", + "content": "t" + }, + { + "bbox": [ + 55, + 391, + 296, + 438 + ], + "type": "text", + "content": ", and condition " + }, + { + "bbox": [ + 55, + 391, + 296, + 438 + ], + "type": "inline_equation", + "content": "c" + }, + { + "bbox": [ + 55, + 391, + 296, + 438 + ], + "type": "text", + "content": ". A more detailed formulation of diffusion models is provided in the Appendix." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 55, + 438, + 296, + 605 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 438, + 296, + 605 + ], + "spans": [ + { + "bbox": [ + 55, + 438, + 296, + 605 + ], + "type": "text", + "content": "In this paper, we utilize the powerful LDM model to generate human appearances. To achieve appearance guidance, we disentangle pose control and appearance guidance. Unlike prior approaches [11, 21], which rely on CLIP's [22] image encoder or a U-Net-like appearance encoder, we condition the diffusion model on multi-pose embeddings generated by the Global-aware Pose Generation module. Additionally, to ensure coherence across the poses and to generate the entire sequence simultaneously, we extend the 2D diffusion model to a 3D setting, inspired by [21]. For improved efficiency, we fine-tune only selected weights in the U-Net, as shown in Fig. 2. Here, the \"lock\" symbol denotes frozen weights, while the \"flame\" symbol indicates fine-tuned parameters." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 55, + 605, + 296, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 605, + 296, + 714 + ], + "spans": [ + { + "bbox": [ + 55, + 605, + 296, + 714 + ], + "type": "text", + "content": "Appearance Adapter. To enable precise control over the character's appearance details, we introduce the Appearance Adapter (AD), shown in Fig. 2. The AD operates hierarchically between multi-scale upsampling blocks in the diffusion framework, extracting pose-agnostic appearance features from reference images through cross-scale feature learning. This architecture enables explicit disentanglement of high-level semantic understanding between appearance attributes and pose informations." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 313, + 72, + 555, + 204 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 72, + 555, + 204 + ], + "spans": [ + { + "bbox": [ + 313, + 72, + 555, + 204 + ], + "type": "text", + "content": "Specifically, our implementation with " + }, + { + "bbox": [ + 313, + 72, + 555, + 204 + ], + "type": "inline_equation", + "content": "\\mathrm{k} = 3" + }, + { + "bbox": [ + 313, + 72, + 555, + 204 + ], + "type": "text", + "content": " Appearance Embedding (AE) layers is augmented by two convolutional layers for dimensional alignment. Adopting the zero-initialization strategy from [23], these convolutional layers progressively smooth feature representations while suppressing high-frequency noise, thereby allowing stable learning of fine-grained appearance patterns. The AE layers enhance feature refinement through global attention mechanisms, followed by a Feed-Forward Network (FFN) that non-linearly reinforces the discriminative power of the processed features." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 314, + 212, + 400, + 224 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 314, + 212, + 400, + 224 + ], + "spans": [ + { + "bbox": [ + 314, + 212, + 400, + 224 + ], + "type": "text", + "content": "3.4. Optimization" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 313, + 230, + 555, + 255 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 230, + 555, + 255 + ], + "spans": [ + { + "bbox": [ + 313, + 230, + 555, + 255 + ], + "type": "text", + "content": "We train our proposed model in an end-to-end manner. Our overall optimization objective is:" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 388, + 264, + 553, + 277 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 388, + 264, + 553, + 277 + ], + "spans": [ + { + "bbox": [ + 388, + 264, + 553, + 277 + ], + "type": "interline_equation", + "content": "\\mathcal {L} _ {\\text {t o t a l}} = \\mathcal {L} _ {m g} + \\mathcal {L} _ {\\text {r e c}}, \\tag {5}", + "image_path": "4812b25a16c6334894e7d42c5bd129fa3eedff2e14d2917a544ce6cf205708d0.jpg" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 313, + 286, + 555, + 311 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 286, + 555, + 311 + ], + "spans": [ + { + "bbox": [ + 313, + 286, + 555, + 311 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 313, + 286, + 555, + 311 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{mg}" + }, + { + "bbox": [ + 313, + 286, + 555, + 311 + ], + "type": "text", + "content": " is used to optimize the multi-pose generation process. The optimization objective is:" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 318, + 319, + 553, + 360 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 318, + 319, + 553, + 360 + ], + "spans": [ + { + "bbox": [ + 318, + 319, + 553, + 360 + ], + "type": "interline_equation", + "content": "\\begin{array}{l} \\mathcal {L} _ {m g} = \\mathbb {E} _ {z _ {0} ^ {1: N}, \\mathcal {X} _ {r e f}, \\mathcal {P} ^ {1: N}} [ \\| \\epsilon - \\\\ \\epsilon_ {\\theta} \\left(z _ {t} ^ {1: N}, t, \\varphi \\left(\\mathcal {F} _ {r e f}, \\mathcal {F} _ {p} ^ {1: N}\\right)\\right) | | _ {2} ^ {2} ], \\tag {6} \\\\ \\end{array}", + "image_path": "00da66ad7552f0f353593e150b3a65aec37d414777787d1bf54dc02fbeeb70d0.jpg" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 313, + 361, + 555, + 494 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 361, + 555, + 494 + ], + "spans": [ + { + "bbox": [ + 313, + 361, + 555, + 494 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 313, + 361, + 555, + 494 + ], + "type": "inline_equation", + "content": "\\mathcal{X}_{ref}" + }, + { + "bbox": [ + 313, + 361, + 555, + 494 + ], + "type": "text", + "content": " denote the reference image, with " + }, + { + "bbox": [ + 313, + 361, + 555, + 494 + ], + "type": "inline_equation", + "content": "\\mathcal{F}_{ref}" + }, + { + "bbox": [ + 313, + 361, + 555, + 494 + ], + "type": "text", + "content": " representing its features. The target poses are given by " + }, + { + "bbox": [ + 313, + 361, + 555, + 494 + ], + "type": "inline_equation", + "content": "\\mathcal{P}^{1:N}" + }, + { + "bbox": [ + 313, + 361, + 555, + 494 + ], + "type": "text", + "content": ", and " + }, + { + "bbox": [ + 313, + 361, + 555, + 494 + ], + "type": "inline_equation", + "content": "\\mathcal{F}_p^{1:N} = \\mathcal{M}_p(\\mathcal{P}^{1:N})" + }, + { + "bbox": [ + 313, + 361, + 555, + 494 + ], + "type": "text", + "content": " represents the features of the target poses. Additionally, " + }, + { + "bbox": [ + 313, + 361, + 555, + 494 + ], + "type": "inline_equation", + "content": "z_0^{1:N} = \\mathcal{E}(\\mathcal{V}^{1:N})" + }, + { + "bbox": [ + 313, + 361, + 555, + 494 + ], + "type": "text", + "content": " denotes the initial latent features, while " + }, + { + "bbox": [ + 313, + 361, + 555, + 494 + ], + "type": "inline_equation", + "content": "z_t^{1:N}" + }, + { + "bbox": [ + 313, + 361, + 555, + 494 + ], + "type": "text", + "content": " is the noisy latent representation derived from " + }, + { + "bbox": [ + 313, + 361, + 555, + 494 + ], + "type": "inline_equation", + "content": "z_0^{1:N}" + }, + { + "bbox": [ + 313, + 361, + 555, + 494 + ], + "type": "text", + "content": " at timestep " + }, + { + "bbox": [ + 313, + 361, + 555, + 494 + ], + "type": "inline_equation", + "content": "t" + }, + { + "bbox": [ + 313, + 361, + 555, + 494 + ], + "type": "text", + "content": ", where " + }, + { + "bbox": [ + 313, + 361, + 555, + 494 + ], + "type": "inline_equation", + "content": "\\mathcal{E}" + }, + { + "bbox": [ + 313, + 361, + 555, + 494 + ], + "type": "text", + "content": " is VAE encoder. In addition, inspired by [5], we introduced an additional reconstruction task for the reference image during training to ensure the consistency between the appearance of the generated target poses and the reference image. The optimization objective of the reconstruction task is:" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 342, + 502, + 555, + 517 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 342, + 502, + 555, + 517 + ], + "spans": [ + { + "bbox": [ + 342, + 502, + 555, + 517 + ], + "type": "interline_equation", + "content": "\\mathcal {L} _ {r e c} = \\mathbb {E} _ {z _ {0}, \\chi_ {r e f}, \\mathcal {P} ^ {0}} [ | | \\epsilon - \\epsilon_ {\\theta} (z _ {t}, t, \\mathcal {F} _ {r e f})) | | _ {2} ^ {2} ]. \\qquad (7)", + "image_path": "b30fae345534e7d00a4b16d5dd28ebe50644222c3a1ce92dc5d410e384e86dd3.jpg" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 313, + 525, + 555, + 622 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 525, + 555, + 622 + ], + "spans": [ + { + "bbox": [ + 313, + 525, + 555, + 622 + ], + "type": "text", + "content": "Sampling. At inference time, we sample " + }, + { + "bbox": [ + 313, + 525, + 555, + 622 + ], + "type": "inline_equation", + "content": "z_{t}" + }, + { + "bbox": [ + 313, + 525, + 555, + 622 + ], + "type": "text", + "content": " from the Gaussian noise and iteratively sample from the previous step until " + }, + { + "bbox": [ + 313, + 525, + 555, + 622 + ], + "type": "inline_equation", + "content": "z_{0}" + }, + { + "bbox": [ + 313, + 525, + 555, + 622 + ], + "type": "text", + "content": ", then we decode " + }, + { + "bbox": [ + 313, + 525, + 555, + 622 + ], + "type": "inline_equation", + "content": "z_{0}" + }, + { + "bbox": [ + 313, + 525, + 555, + 622 + ], + "type": "text", + "content": " to image space get the final human pose image. To improve the sample quality during the generation process, we use a classifier-free guidance strategy. To implement classifier-free guidance, we randomly drop the reference image used as the appearance condition with a probability of " + }, + { + "bbox": [ + 313, + 525, + 555, + 622 + ], + "type": "inline_equation", + "content": "5\\%" + }, + { + "bbox": [ + 313, + 525, + 555, + 622 + ], + "type": "text", + "content": " during training." + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 331, + 632, + 553, + 644 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 331, + 632, + 553, + 644 + ], + "spans": [ + { + "bbox": [ + 331, + 632, + 553, + 644 + ], + "type": "interline_equation", + "content": "\\hat {\\epsilon} _ {\\theta} \\left(z _ {t}, t, \\varphi\\right) = \\lambda \\epsilon_ {\\theta} \\left(z _ {t}, t, \\varphi\\right) + (1 - \\lambda) \\epsilon_ {\\theta} \\left(z _ {t}, t\\right), \\tag {8}", + "image_path": "6549f7f71b5d30c53a35addf53ac76c194a4255752968eef4409f9ec97e87189.jpg" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 313, + 654, + 555, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 654, + 555, + 714 + ], + "spans": [ + { + "bbox": [ + 313, + 654, + 555, + 714 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 313, + 654, + 555, + 714 + ], + "type": "inline_equation", + "content": "\\lambda" + }, + { + "bbox": [ + 313, + 654, + 555, + 714 + ], + "type": "text", + "content": " is a hyperparameter used to control the strength of guidance during the inference process, while " + }, + { + "bbox": [ + 313, + 654, + 555, + 714 + ], + "type": "inline_equation", + "content": "\\epsilon_{\\theta}(z_t,t,\\varphi)" + }, + { + "bbox": [ + 313, + 654, + 555, + 714 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 313, + 654, + 555, + 714 + ], + "type": "inline_equation", + "content": "\\epsilon_{\\theta}(z_t,t)" + }, + { + "bbox": [ + 313, + 654, + 555, + 714 + ], + "type": "text", + "content": " represent conditional prediction and unconditional prediction, respectively. To handle the case when the reference image is missing, inspired by [10], we use a" + } + ] + } + ], + "index": 19 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "text", + "content": "5650" + } + ] + } + ], + "index": 20 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 4 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 59, + 61, + 552, + 178 + ], + "blocks": [ + { + "bbox": [ + 59, + 61, + 552, + 178 + ], + "lines": [ + { + "bbox": [ + 59, + 61, + 552, + 178 + ], + "spans": [ + { + "bbox": [ + 59, + 61, + 552, + 178 + ], + "type": "table", + "html": "
ApproachesUBC FashionTikTok
SSIM ↑PSNR ↑FID ↓LPIPS ↓\\( L_1 \\) ↓SSIM ↑PSNR ↑FID ↓LPIPS ↓\\( L_1 \\) ↓
PISE [8] (CVPR'21)0.86717.11618.6570.07520.0460.4299.947120.6480.30100.186
DPTN [5] (CVPR'22)0.83315.93719.9570.08760.0500.1885.696143.1100.45170.449
NTED [7] (CVPR'22)0.82415.68319.7840.09480.0530.3398.551140.3030.37270.228
BiGraphGAN [4] (IJCV'23)0.84416.40121.0120.08060.0470.4459.916147.5210.26760.177
PIDM [12] (CVPR'23)0.81218.87014.9510.16350.0920.3077.558130.3920.42770.262
CFLD [10] (CVPR'24)0.88922.14212.8100.07180.0260.72917.520105.6310.21360.078
MagicPose [11] (ICML'24)0.83120.74214.6490.08660.0300.67717.260108.0670.25070.083
Ours0.94125.65511.4870.04970.0140.75318.466103.7080.23110.072
", + "image_path": "6535856aab0e335496b4c906b030c727deabe9d69e815d26e06e0c29adb786b5.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "table_body" + } + ], + "index": 0 + }, + { + "type": "image", + "bbox": [ + 57, + 213, + 291, + 283 + ], + "blocks": [ + { + "bbox": [ + 57, + 213, + 291, + 283 + ], + "lines": [ + { + "bbox": [ + 57, + 213, + 291, + 283 + ], + "spans": [ + { + "bbox": [ + 57, + 213, + 291, + 283 + ], + "type": "image", + "image_path": "350dbc4f2801db176651fffe8185f801abfb2c9134521767c98eff0684cc875f.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 55, + 292, + 295, + 326 + ], + "lines": [ + { + "bbox": [ + 55, + 292, + 295, + 326 + ], + "spans": [ + { + "bbox": [ + 55, + 292, + 295, + 326 + ], + "type": "text", + "content": "Figure 4. The pose synthesis accuracy of the proposed method on the UBC Fashion [13] dataset as the number of generated images increases, with 8 as the training setting." + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_caption" + } + ], + "index": 2 + }, + { + "type": "table", + "bbox": [ + 58, + 336, + 293, + 381 + ], + "blocks": [ + { + "bbox": [ + 87, + 187, + 520, + 198 + ], + "lines": [ + { + "bbox": [ + 87, + 187, + 520, + 198 + ], + "spans": [ + { + "bbox": [ + 87, + 187, + 520, + 198 + ], + "type": "text", + "content": "Table 1. Quantitative comparison of pose sequence synthesis quality on the UBC Fashion [13], and TikTok [14] dataset." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 58, + 336, + 293, + 381 + ], + "lines": [ + { + "bbox": [ + 58, + 336, + 293, + 381 + ], + "spans": [ + { + "bbox": [ + 58, + 336, + 293, + 381 + ], + "type": "table", + "html": "
MethodsPIDM [12] (CVPR'23)CFLD [10] (CVPR'24)MagicPose [11] (ICML'24)Ours(8-24)
#Param131M248M2081M367M
Inference Time28.64s35.8s67.8s5.5-24.1s
", + "image_path": "a95065c579c03055c2ca7f4dda417482686f35952a6f481f4603fca0e7e5aaec.jpg" + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "table_body" + } + ], + "index": 4 + }, + { + "bbox": [ + 55, + 386, + 295, + 409 + ], + "lines": [ + { + "bbox": [ + 55, + 386, + 295, + 409 + ], + "spans": [ + { + "bbox": [ + 55, + 386, + 295, + 409 + ], + "type": "text", + "content": "Table 2. Comparison of parameter count and inference time between the proposed method and other SOTA diffusion models." + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "text" + }, + { + "bbox": [ + 55, + 413, + 296, + 498 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 413, + 296, + 498 + ], + "spans": [ + { + "bbox": [ + 55, + 413, + 296, + 498 + ], + "type": "text", + "content": "learnable vector as the conditional embedding. We concatenate the learnable vector with the reference image features to form a compact conditional embedding for the generation process, encouraging the learnable vector to capture more information beyond the reference image during training. This allows the model to receive additional appearance information when predicting invisible regions." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 55, + 508, + 137, + 521 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 508, + 137, + 521 + ], + "spans": [ + { + "bbox": [ + 55, + 508, + 137, + 521 + ], + "type": "text", + "content": "4. Experiments" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 55, + 527, + 105, + 540 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 527, + 105, + 540 + ], + "spans": [ + { + "bbox": [ + 55, + 527, + 105, + 540 + ], + "type": "text", + "content": "4.1. Setup" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 55, + 545, + 296, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 545, + 296, + 713 + ], + "spans": [ + { + "bbox": [ + 55, + 545, + 296, + 713 + ], + "type": "text", + "content": "Dataset. We validate the effectiveness of our proposed method on both the UBC Fashion [13] and TikTok [14] datasets. The UBC Fashion dataset is a collection of data featuring various poses and outfits, containing 500 training and 100 test videos. Compared to UBC Fashion, the TikTok dataset has more complex backgrounds and greater diversity in the proportions of people, making the task more challenging. The TikTok dataset comprises 350 dance videos featuring individual performers, each lasting between 10 and 15 seconds. To ensure a fair comparison with state-of-the-art methods, we use the same test set as [11] for evaluation. We set the image resolution to " + }, + { + "bbox": [ + 55, + 545, + 296, + 713 + ], + "type": "inline_equation", + "content": "256 \\times 256" + }, + { + "bbox": [ + 55, + 545, + 296, + 713 + ], + "type": "text", + "content": " and randomly sample " + }, + { + "bbox": [ + 55, + 545, + 296, + 713 + ], + "type": "inline_equation", + "content": "N" + }, + { + "bbox": [ + 55, + 545, + 296, + 713 + ], + "type": "text", + "content": " frames (in this paper, " + }, + { + "bbox": [ + 55, + 545, + 296, + 713 + ], + "type": "inline_equation", + "content": "N" + }, + { + "bbox": [ + 55, + 545, + 296, + 713 + ], + "type": "text", + "content": " is set to 8) from each training video for training. Additionally, we use Open-" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 313, + 214, + 555, + 262 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 214, + 555, + 262 + ], + "spans": [ + { + "bbox": [ + 313, + 214, + 555, + 262 + ], + "type": "text", + "content": "Pose [24] and PGN [25] to extract human poses and human parsing maps. Due to category redundancy in the human parsing maps, we reduce the number of categories from 20 to 8 as [8]." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 313, + 272, + 555, + 465 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 272, + 555, + 465 + ], + "spans": [ + { + "bbox": [ + 313, + 272, + 555, + 465 + ], + "type": "text", + "content": "Metrics. To evaluate the quality of the generated pose sequences, we report several metrics: Structural Similarity Index Measure (SSIM) [26], Peak Signal to Noise Ratio (PSNR), Learned Perceptual Image Patch Similarity (LPIPS) [27], and Fréchet Inception Distance (FID) [28]. SSIM and PSNR assess pixel-level similarity by measuring statistical differences between the generated images and the ground truth (GT). However, these metrics are limited to low-level image properties and may not fully capture perceptual quality. To address this limitation, we also include LPIPS and FID, which evaluate deep feature-based similarities. FID evaluates the distribution differences between the generated images and GT using Inception-v3 [29] features, while LPIPS measures the perceptual similarity between images using a VGG [30] network, making it more aligned with human perception of image similarity." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 313, + 474, + 556, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 474, + 556, + 715 + ], + "spans": [ + { + "bbox": [ + 313, + 474, + 556, + 715 + ], + "type": "text", + "content": "Implementation details. The proposed method is implemented based on PyTorch [31] and Diffusers [32], trained in an end-to-end manner with a batch size of 16 across 4 NVIDIA A100 GPUs. The model was trained for 100 epochs utilizing the Adam [33] optimizer, starting with an initial learning rate of " + }, + { + "bbox": [ + 313, + 474, + 556, + 715 + ], + "type": "inline_equation", + "content": "5 \\times 10^{-5}" + }, + { + "bbox": [ + 313, + 474, + 556, + 715 + ], + "type": "text", + "content": ". A learning rate warmup strategy was implemented for the first 1000 training steps, followed by a decay to " + }, + { + "bbox": [ + 313, + 474, + 556, + 715 + ], + "type": "inline_equation", + "content": "1 \\times 10^{-5}" + }, + { + "bbox": [ + 313, + 474, + 556, + 715 + ], + "type": "text", + "content": " at the 50th epoch. To accelerate model convergence, we initialize the image encoder using Swin-B [17] pretrained on ImageNet [34] and the U-net using version 1.5 of Stable Diffusion [20]. Additionally, to address the issue of poor facial realism in generated pose, we followed strategy [20] by fine-tuning the decoder of the VAE using both MSE reconstruction loss and perceptual loss [35]. During the fine-tuning process, we placed greater emphasis on the MSE reconstruction loss. Further details on the fine-tuning process can be found in the Appendix. For classifier-free guidance, we set the parameter " + }, + { + "bbox": [ + 313, + 474, + 556, + 715 + ], + "type": "inline_equation", + "content": "\\lambda" + }, + { + "bbox": [ + 313, + 474, + 556, + 715 + ], + "type": "text", + "content": " to 2.0 during sampling, and the reference condition " + }, + { + "bbox": [ + 313, + 474, + 556, + 715 + ], + "type": "inline_equation", + "content": "\\mathcal{X}_{ref}" + }, + { + "bbox": [ + 313, + 474, + 556, + 715 + ], + "type": "text", + "content": " was dropped with a probability of " + }, + { + "bbox": [ + 313, + 474, + 556, + 715 + ], + "type": "inline_equation", + "content": "5\\%" + }, + { + "bbox": [ + 313, + 474, + 556, + 715 + ], + "type": "text", + "content": " throughout training." + } + ] + } + ], + "index": 12 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 749, + 314, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 749, + 314, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 749, + 314, + 757 + ], + "type": "text", + "content": "5651" + } + ] + } + ], + "index": 13 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 5 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 61, + 62, + 161, + 381 + ], + "blocks": [ + { + "bbox": [ + 61, + 62, + 161, + 381 + ], + "lines": [ + { + "bbox": [ + 61, + 62, + 161, + 381 + ], + "spans": [ + { + "bbox": [ + 61, + 62, + 161, + 381 + ], + "type": "image", + "image_path": "940da2dad2d15f7a32bc66e786cb4110fe0a69046f25020bc1e07820d96020fc.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 55, + 393, + 555, + 417 + ], + "lines": [ + { + "bbox": [ + 55, + 393, + 555, + 417 + ], + "spans": [ + { + "bbox": [ + 55, + 393, + 555, + 417 + ], + "type": "text", + "content": "Figure 5. Qualitative comparison of our method with SOTA methods on UBC Fashion [13] and TikTok[14] dataset. Our approach can generate more realistic details consistent with the reference image, such as garment folds, fine-grained textures, and facial identity features." + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "type": "image", + "bbox": [ + 164, + 62, + 209, + 380 + ], + "blocks": [ + { + "bbox": [ + 164, + 62, + 209, + 380 + ], + "lines": [ + { + "bbox": [ + 164, + 62, + 209, + 380 + ], + "spans": [ + { + "bbox": [ + 164, + 62, + 209, + 380 + ], + "type": "image", + "image_path": "16720a2c89453f62e1f068e850405c517837a50e90b36ae484ed2234a30ec063.jpg" + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_body" + } + ], + "index": 1 + }, + { + "type": "image", + "bbox": [ + 210, + 62, + 304, + 380 + ], + "blocks": [ + { + "bbox": [ + 210, + 62, + 304, + 380 + ], + "lines": [ + { + "bbox": [ + 210, + 62, + 304, + 380 + ], + "spans": [ + { + "bbox": [ + 210, + 62, + 304, + 380 + ], + "type": "image", + "image_path": "f255f3d716debeb8e1f22c3b7641bf930997bdde57013ac8a2d5c77c8cb8b3ca.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + } + ], + "index": 2 + }, + { + "type": "image", + "bbox": [ + 306, + 62, + 400, + 380 + ], + "blocks": [ + { + "bbox": [ + 306, + 62, + 400, + 380 + ], + "lines": [ + { + "bbox": [ + 306, + 62, + 400, + 380 + ], + "spans": [ + { + "bbox": [ + 306, + 62, + 400, + 380 + ], + "type": "image", + "image_path": "193a2c86f2ef0f2099d29c0557c1fb5091c6a78d8ca6cdea9ea36ed502c25063.jpg" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_body" + } + ], + "index": 3 + }, + { + "type": "image", + "bbox": [ + 402, + 62, + 449, + 380 + ], + "blocks": [ + { + "bbox": [ + 402, + 62, + 449, + 380 + ], + "lines": [ + { + "bbox": [ + 402, + 62, + 449, + 380 + ], + "spans": [ + { + "bbox": [ + 402, + 62, + 449, + 380 + ], + "type": "image", + "image_path": "c6361b96483922b507213ca94bbbd03bcafc642cfa6517337a5e80f5af8a2238.jpg" + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_body" + } + ], + "index": 4 + }, + { + "type": "image", + "bbox": [ + 451, + 62, + 496, + 380 + ], + "blocks": [ + { + "bbox": [ + 451, + 62, + 496, + 380 + ], + "lines": [ + { + "bbox": [ + 451, + 62, + 496, + 380 + ], + "spans": [ + { + "bbox": [ + 451, + 62, + 496, + 380 + ], + "type": "image", + "image_path": "ff74e153d5ac47b00c93ee7b9674ff31c1dba7812ca51c67387f32c2c8d2d400.jpg" + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "image_body" + } + ], + "index": 5 + }, + { + "type": "image", + "bbox": [ + 498, + 63, + 551, + 380 + ], + "blocks": [ + { + "bbox": [ + 498, + 63, + 551, + 380 + ], + "lines": [ + { + "bbox": [ + 498, + 63, + 551, + 380 + ], + "spans": [ + { + "bbox": [ + 498, + 63, + 551, + 380 + ], + "type": "image", + "image_path": "5016a54b521fd3e4d512c5dcd7604d75f527278c4b9f963211e847fedc2915f2.jpg" + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "image_body" + } + ], + "index": 6 + }, + { + "bbox": [ + 55, + 430, + 198, + 443 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 430, + 198, + 443 + ], + "spans": [ + { + "bbox": [ + 55, + 430, + 198, + 443 + ], + "type": "text", + "content": "4.2. Quantitative Comparison" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 55, + 449, + 296, + 664 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 449, + 296, + 664 + ], + "spans": [ + { + "bbox": [ + 55, + 449, + 296, + 664 + ], + "type": "text", + "content": "We conducted a quantitative comparison of our method against state-of-the-art GAN-based and diffusion-based models on the UBC Fashion [13] and TikTok [14] datasets, as shown in Tab. 1. On the majority of metrics, our method outperforms the advanced methods on both datasets. Specifically, compared with four other diffusion models [10-12, 36], our approach achieves more stable generation performance and higher efficiency, with specific results shown in the Tab. 2, owing to the disentangling of pose control from appearance generation. Additionally, this design enables us to generate more human poses during inference (up to 3 times the training quantity) without additional training. The Fig. 4 illustrates the performance of our method as the generation count gradually increases. As shown in this figure, although an increase in generation quantity may lead to a slight performance decline, the results still serve as acceptable human image outputs, outperforming other SOTA methods." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 55, + 671, + 191, + 685 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 671, + 191, + 685 + ], + "spans": [ + { + "bbox": [ + 55, + 671, + 191, + 685 + ], + "type": "text", + "content": "4.3. Qualitative Comparison" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 55, + 689, + 296, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 689, + 296, + 714 + ], + "spans": [ + { + "bbox": [ + 55, + 689, + 296, + 714 + ], + "type": "text", + "content": "The qualitative comparison with state-of-the-art methods is illustrated in Fig. 5 and 6. When using these other meth" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 313, + 431, + 555, + 707 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 431, + 555, + 707 + ], + "spans": [ + { + "bbox": [ + 313, + 431, + 555, + 707 + ], + "type": "text", + "content": "ods, both GAN-based and diffusion-based approaches show overfitting issues, resulting in the loss of person identity and clothing texture details. Additionally, when generating more complex clothing or unseen regions, issues like clothing blurriness, unnatural texture transitions, and substantial differences from the source image become apparent. Benefiting from our proposed strategy of disentangling pose control and appearance generation, our method mitigates overfitting to some extent. Furthermore, with the high-consistency pose embedding provided by Global-aware Pose Generation and the fine-grained, multi-scale features from Appearance Adapter, our approach yields more realistic results in maintaining person identity and capturing intricate clothing patterns. In particular, Fig. 6 provides an additional qualitative comparison with existing SOTA diffusion models for generating multiple poses. This scenario places higher demands on cross-view consistency in person identity and appearance. As observed in the Fig. 6, our method achieves higher consistency across different viewpoints of generated person images, thanks to Global-aware Transformer Decoder's global pose awareness in the pose generation process. More qualitative comparisons can be found in the Appendix." + } + ] + } + ], + "index": 12 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "text", + "content": "5652" + } + ] + } + ], + "index": 13 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 6 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 57, + 56, + 294, + 315 + ], + "blocks": [ + { + "bbox": [ + 57, + 56, + 294, + 315 + ], + "lines": [ + { + "bbox": [ + 57, + 56, + 294, + 315 + ], + "spans": [ + { + "bbox": [ + 57, + 56, + 294, + 315 + ], + "type": "image", + "image_path": "88735d997250453678ac179b5f765a723919163e1ab74e3f11e0c560697d60d4.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 55, + 321, + 295, + 344 + ], + "lines": [ + { + "bbox": [ + 55, + 321, + 295, + 344 + ], + "spans": [ + { + "bbox": [ + 55, + 321, + 295, + 344 + ], + "type": "text", + "content": "Figure 6. Qualitative comparison with other SOTA methods on various poses, with drawbacks highlighted by red circles." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "type": "table", + "bbox": [ + 57, + 355, + 294, + 441 + ], + "blocks": [ + { + "bbox": [ + 57, + 355, + 294, + 441 + ], + "lines": [ + { + "bbox": [ + 57, + 355, + 294, + 441 + ], + "spans": [ + { + "bbox": [ + 57, + 355, + 294, + 441 + ], + "type": "table", + "html": "
ModelSSIM ↑PSNR ↑FID ↓LPIPS ↓
w/o Global-aware Pose Generation (GPG)0.936225.02411.8650.0532
w/o Multi-stage Image Encoder (MIE)0.934424.89511.2840.0554
w/o Appearance Adapter (AD)0.938425.20511.9340.0517
w/o Fine-tune VAE (FVAE)0.938625.46411.6270.0514
w/o Lrec0.940325.53911.5470.0516
Ours + (a)0.938925.44311.4870.0503
Ours + (b)0.938125.44711.5720.0506
Ours0.941125.65511.4870.0497
", + "image_path": "4d3bdde5d99a998ffea95294ad1d45edd573c8feb1c94b2b2cf3eb3792ce234f.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "table_body" + } + ], + "index": 2 + }, + { + "bbox": [ + 55, + 479, + 149, + 493 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 479, + 149, + 493 + ], + "spans": [ + { + "bbox": [ + 55, + 479, + 149, + 493 + ], + "type": "text", + "content": "4.4. Ablation Study" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 54, + 498, + 297, + 703 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 54, + 498, + 297, + 703 + ], + "spans": [ + { + "bbox": [ + 54, + 498, + 297, + 703 + ], + "type": "text", + "content": "Tab. 3 and Fig. 7 show quantitative and qualitative results, respectively. It can be observed that both GPG and AD are essential for pose synthesis performance, with their absence potentially leading to a performance drop of up to " + }, + { + "bbox": [ + 54, + 498, + 297, + 703 + ], + "type": "inline_equation", + "content": "3.8\\%" + }, + { + "bbox": [ + 54, + 498, + 297, + 703 + ], + "type": "text", + "content": " (FID). This highlights the importance of separating pose control and appearance generation strategies. We also replaced MIE with a CLIP image encoder, resulting in a " + }, + { + "bbox": [ + 54, + 498, + 297, + 703 + ], + "type": "inline_equation", + "content": "0.7\\%" + }, + { + "bbox": [ + 54, + 498, + 297, + 703 + ], + "type": "text", + "content": " (SSIM) and " + }, + { + "bbox": [ + 54, + 498, + 297, + 703 + ], + "type": "inline_equation", + "content": "2.9\\%" + }, + { + "bbox": [ + 54, + 498, + 297, + 703 + ], + "type": "text", + "content": " (PSNR) decline, though with an improvement in FID performance. As shown in the first row of Fig. 7, the CLIP image encoder fails to retain intricate clothing details, leading to a loss of appearance-related information. Additionally, fine-tuning the VAE for the specific dataset helps preserve realistic details of faces and clothing to a certain degree. Finally, we also experimented with two alternative Transformer decoder architectures in Fig. 3, both of which led to varying degrees of performance decline. Detailed ablation experiments can be found in the Appendix." + } + ] + } + ], + "index": 5 + }, + { + "type": "image", + "bbox": [ + 319, + 56, + 549, + 173 + ], + "blocks": [ + { + "bbox": [ + 319, + 56, + 549, + 173 + ], + "lines": [ + { + "bbox": [ + 319, + 56, + 549, + 173 + ], + "spans": [ + { + "bbox": [ + 319, + 56, + 549, + 173 + ], + "type": "image", + "image_path": "2eeb1baa38a4468e016a694a995f14930a2fabb6846ce9ffc262718546e84fac.jpg" + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 313, + 178, + 555, + 200 + ], + "lines": [ + { + "bbox": [ + 313, + 178, + 555, + 200 + ], + "spans": [ + { + "bbox": [ + 313, + 178, + 555, + 200 + ], + "type": "text", + "content": "Figure 7. Qualitative ablation study on the UBC Fashion [13] dataset." + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "image_caption" + } + ], + "index": 6 + }, + { + "type": "table", + "bbox": [ + 317, + 205, + 553, + 247 + ], + "blocks": [ + { + "bbox": [ + 55, + 447, + 295, + 470 + ], + "lines": [ + { + "bbox": [ + 55, + 447, + 295, + 470 + ], + "spans": [ + { + "bbox": [ + 55, + 447, + 295, + 470 + ], + "type": "text", + "content": "Table 3. Ablation study of different components, where (a) and (b) represent different decoder architectures in Fig. 3." + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 317, + 205, + 553, + 247 + ], + "lines": [ + { + "bbox": [ + 317, + 205, + 553, + 247 + ], + "spans": [ + { + "bbox": [ + 317, + 205, + 553, + 247 + ], + "type": "table", + "html": "
MethodsPIDM [12] (CVPR'23)CFLD [10] (CVPR'24)MagicPose [11] (ICML'24)Ours
Average13.6%16.7%8.3%61.4%
", + "image_path": "c2f6c121f82d6ec19683e6849f5db647983e8031855facbc1386971f451896f3.jpg" + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "table_body" + } + ], + "index": 8 + }, + { + "bbox": [ + 313, + 251, + 555, + 297 + ], + "lines": [ + { + "bbox": [ + 313, + 251, + 555, + 297 + ], + "spans": [ + { + "bbox": [ + 313, + 251, + 555, + 297 + ], + "type": "text", + "content": "Table 4. User study on the UBC Fashion [13] dataset. We followed [11] to collect vote data from 30 volunteers. The participants believed that our proposed method can generate visually realistic and identity-consistent human poses." + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "text" + }, + { + "bbox": [ + 313, + 303, + 389, + 316 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 303, + 389, + 316 + ], + "spans": [ + { + "bbox": [ + 313, + 303, + 389, + 316 + ], + "type": "text", + "content": "4.5. User Study" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 313, + 322, + 555, + 453 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 322, + 555, + 453 + ], + "spans": [ + { + "bbox": [ + 313, + 322, + 555, + 453 + ], + "type": "text", + "content": "To more accurately evaluate the effectiveness of our proposed method, we designed a user study. In this experiment, we collected images of 9 subjects from the test set for comparative evaluation. One subject was designated as a warm-up, while the remaining eight were used as formal test. For each subject, eight distinct pose images were randomly selected to comparison. We presented images generated by various methods [10-12] to users, instructing them to evaluate the images based on visual quality, pose accuracy, and detail consistency, and to select the method they preferred. We present the averaged vote result in Tab. 4." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 313, + 465, + 392, + 478 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 465, + 392, + 478 + ], + "spans": [ + { + "bbox": [ + 313, + 465, + 392, + 478 + ], + "type": "text", + "content": "5. Conclusions" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 313, + 486, + 555, + 665 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 486, + 555, + 665 + ], + "spans": [ + { + "bbox": [ + 313, + 486, + 555, + 665 + ], + "type": "text", + "content": "In this paper, we propose a unified framework for diverse multi-pose generation driven by disentangled pose and appearance guidance. This framework separately addresses complex human pose control and appearance generation to produce realistic and varied human poses. To achieve precise pose control and realistic appearance generation, we introduce Global-aware Pose Generation and Appearance Adapter module. These modules enhance interactions among poses during the generation process, leveraging all available mutual information to provide global guidance, while also learning high-level semantic information from multi-scale features to preserve texture details as much as possible. Extensive experiments demonstrate that our proposed approach achieves significant advantages over other state-of-the-art methods." + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 313, + 665, + 556, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 665, + 556, + 713 + ], + "spans": [ + { + "bbox": [ + 313, + 665, + 556, + 713 + ], + "type": "text", + "content": "Acknowledgments. This work was supported by the National Natural Science Foundation of China (62036006, 62276200) and the Tianjin Natural Science Foundation (24JCYBJC00310)." + } + ] + } + ], + "index": 14 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "text", + "content": "5653" + } + ] + } + ], + "index": 15 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 7 + }, + { + "para_blocks": [ + { + "bbox": [ + 56, + 72, + 115, + 83 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 72, + 115, + 83 + ], + "spans": [ + { + "bbox": [ + 56, + 72, + 115, + 83 + ], + "type": "text", + "content": "References" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 57, + 91, + 295, + 713 + ], + "type": "list", + "angle": 0, + "index": 14, + "blocks": [ + { + "bbox": [ + 61, + 91, + 294, + 113 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 61, + 91, + 294, + 113 + ], + "spans": [ + { + "bbox": [ + 61, + 91, + 294, + 113 + ], + "type": "text", + "content": "[1] Grigore C Burdea and Philippe Coiffet. *Virtual reality technology*. John Wiley & Sons, 2003. 2" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 61, + 114, + 295, + 157 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 61, + 114, + 295, + 157 + ], + "spans": [ + { + "bbox": [ + 61, + 114, + 295, + 157 + ], + "type": "text", + "content": "[2] Julie Carmigniani, Borko Furht, Marco Anisetti, Paolo Ceravolo, Ernesto Damiani, and Misa Ivkovic. Augmented reality technologies, systems and applications. Multimedia tools and applications, 51:341-377, 2011. 2" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 62, + 159, + 295, + 212 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 159, + 295, + 212 + ], + "spans": [ + { + "bbox": [ + 62, + 159, + 295, + 212 + ], + "type": "text", + "content": "[3] Liang Zheng, Liyue Shen, Lu Tian, Shengjin Wang, Jingdong Wang, and Qi Tian. Scalable person re-identification: A benchmark. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 1116-1124, 2015. 2" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 62, + 214, + 294, + 257 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 214, + 294, + 257 + ], + "spans": [ + { + "bbox": [ + 62, + 214, + 294, + 257 + ], + "type": "text", + "content": "[4] Hao Tang, Ling Shao, Philip HS Torr, and Nicu Sebe. Bipartite graph reasoning gans for person pose and facial image synthesis. International Journal of Computer Vision, 131(3):644-658, 2023. 2, 6" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 62, + 259, + 295, + 312 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 259, + 295, + 312 + ], + "spans": [ + { + "bbox": [ + 62, + 259, + 295, + 312 + ], + "type": "text", + "content": "[5] Pengze Zhang, Lingxiao Yang, Jian-Huang Lai, and Xiaohua Xie. Exploring dual-task correlation for pose guided person image generation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 7713-7722, 2022. 2, 5, 6" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 62, + 315, + 295, + 357 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 315, + 295, + 357 + ], + "spans": [ + { + "bbox": [ + 62, + 315, + 295, + 357 + ], + "type": "text", + "content": "[6] Yuelong Li, Tengfei Xiao, Lei Geng, and Jianming Wang. Direct may not be the best: An incremental evolution view of pose generation. In Proceedings of the AAAI Conference on Artificial Intelligence, volume 38, pages 3270-3278, 2024. 2" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 62, + 359, + 295, + 413 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 359, + 295, + 413 + ], + "spans": [ + { + "bbox": [ + 62, + 359, + 295, + 413 + ], + "type": "text", + "content": "[7] Yurui Ren, Xiaqing Fan, Ge Li, Shan Liu, and Thomas H Li. Neural texture extraction and distribution for controllable person image synthesis. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 13535-13544, 2022. 2, 6" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 62, + 415, + 295, + 468 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 415, + 295, + 468 + ], + "spans": [ + { + "bbox": [ + 62, + 415, + 295, + 468 + ], + "type": "text", + "content": "[8] Jinsong Zhang, Kun Li, Yu-Kun Lai, and Jingyu Yang. Pise: Person image synthesis and editing with decoupled gan. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 7982-7990, 2021. 2, 6" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 62, + 470, + 295, + 502 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 470, + 295, + 502 + ], + "spans": [ + { + "bbox": [ + 62, + 470, + 295, + 502 + ], + "type": "text", + "content": "[9] Jonathan Ho, Ajay Jain, and Pieter Abbeel. Denoising diffusion probabilistic models. Advances in Neural Information Processing Systems, 33:6840-6851, 2020. 2, 3, 5" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 57, + 503, + 295, + 557 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 503, + 295, + 557 + ], + "spans": [ + { + "bbox": [ + 57, + 503, + 295, + 557 + ], + "type": "text", + "content": "[10] Yanzuo Lu, Manlin Zhang, Andy J Ma, Xiaohua Xie, and Jianhuang Lai. Coarse-to-fine latent diffusion for pose-guided person image synthesis. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 6420-6429, 2024. 2, 5, 6, 7, 8" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 57, + 559, + 295, + 624 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 559, + 295, + 624 + ], + "spans": [ + { + "bbox": [ + 57, + 559, + 295, + 624 + ], + "type": "text", + "content": "[11] Di Chang, Yichun Shi, Quankai Gao, Hongyi Xu, Jessica Fu, Guoxian Song, Qing Yan, Yizhe Zhu, Xiao Yang, and Mohammad Soleymani. Magicpose: Realistic human poses and facial expressions retargeting with identity-aware diffusion. In Proceedings of the International Conference on Machine Learning, 2023. 3, 5, 6, 8" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 57, + 625, + 295, + 690 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 625, + 295, + 690 + ], + "spans": [ + { + "bbox": [ + 57, + 625, + 295, + 690 + ], + "type": "text", + "content": "[12] Ankan Kumar Bhunia, Salman Khan, Hisham Cholakkal, Rao Muhammad Anwer, Jorma Laaksonen, Mubarak Shah, and Fahad Shahbaz Khan. Person image synthesis via denoising diffusion model. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 5968-5976, 2023. 2, 3, 6, 7, 8" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 57, + 692, + 295, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 692, + 295, + 713 + ], + "spans": [ + { + "bbox": [ + 57, + 692, + 295, + 713 + ], + "type": "text", + "content": "[13] Polina Zablotskaia, Aliaksandr Siarohin, Bo Zhao, and Leonid Sigal. Dwnet: Dense warp-based network for" + } + ] + } + ], + "index": 13 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 317, + 74, + 555, + 713 + ], + "type": "list", + "angle": 0, + "index": 28, + "blocks": [ + { + "bbox": [ + 335, + 74, + 553, + 95 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 335, + 74, + 553, + 95 + ], + "spans": [ + { + "bbox": [ + 335, + 74, + 553, + 95 + ], + "type": "text", + "content": "pose-guided human video generation. arXiv preprint arXiv:1910.09139, 2019. 2, 6, 7, 8" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 317, + 97, + 553, + 150 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 97, + 553, + 150 + ], + "spans": [ + { + "bbox": [ + 317, + 97, + 553, + 150 + ], + "type": "text", + "content": "[14] Yasamin Jafarian and Hyun Soo Park. Learning high fidelity depths of dressed humans by watching social media dance videos. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 12753-12762, 2021. 2, 6, 7" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 317, + 152, + 553, + 195 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 152, + 553, + 195 + ], + "spans": [ + { + "bbox": [ + 317, + 152, + 553, + 195 + ], + "type": "text", + "content": "[15] Ian Goodfellow, Jean Pouget-Abadie, Mehdi Mirza, Bing Xu, David Warde-Farley, Sherjil Ozair, Aaron Courville, and Yoshua Bengio. Generative adversarial networks. Communications of the ACM, 63(11):139-144, 2020. 2" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 317, + 198, + 553, + 230 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 198, + 553, + 230 + ], + "spans": [ + { + "bbox": [ + 317, + 198, + 553, + 230 + ], + "type": "text", + "content": "[16] Jonathan Ho and Tim Salimans. Classifier-free diffusion guidance. In NeurIPS 2021 Workshop on Deep Generative Models and Downstream Applications, 2021. 3" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 317, + 232, + 553, + 286 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 232, + 553, + 286 + ], + "spans": [ + { + "bbox": [ + 317, + 232, + 553, + 286 + ], + "type": "text", + "content": "[17] Ze Liu, Yutong Lin, Yue Cao, Han Hu, Yixuan Wei, Zheng Zhang, Stephen Lin, and Baining Guo. Swin transformer: Hierarchical vision transformer using shifted windows. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 10012-10022, 2021. 3, 6" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 317, + 289, + 553, + 331 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 289, + 553, + 331 + ], + "spans": [ + { + "bbox": [ + 317, + 289, + 553, + 331 + ], + "type": "text", + "content": "[18] Kaiming He, Xiangyu Zhang, Shaoqing Ren, and Jian Sun. Deep residual learning for image recognition. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 770-778, 2016. 3" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 317, + 334, + 553, + 388 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 334, + 553, + 388 + ], + "spans": [ + { + "bbox": [ + 317, + 334, + 553, + 388 + ], + "type": "text", + "content": "[19] Xumin Yu, Yongming Rao, Ziyi Wang, Zuyan Liu, Jiwen Lu, and Jie Zhou. Pointr: Diverse point cloud completion with geometry-aware transformers. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 12498-12507, 2021. 4" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 317, + 389, + 555, + 444 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 389, + 555, + 444 + ], + "spans": [ + { + "bbox": [ + 317, + 389, + 555, + 444 + ], + "type": "text", + "content": "[20] Robin Rombach, Andreas Blattmann, Dominik Lorenz, Patrick Esser, and Björn Ommer. High-resolution image synthesis with latent diffusion models. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 10684-10695, 2022. 5, 6" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 317, + 445, + 555, + 510 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 445, + 555, + 510 + ], + "spans": [ + { + "bbox": [ + 317, + 445, + 555, + 510 + ], + "type": "text", + "content": "[21] Zhongcong Xu, Jianfeng Zhang, Jun Hao Liew, Hanshu Yan, Jia-Wei Liu, Chenxu Zhang, Jiashi Feng, and Mike Zheng Shou. Magicanimate: Temporally consistent human image animation using diffusion model. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 1481–1490, 2024. 5" + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 317, + 513, + 553, + 578 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 513, + 553, + 578 + ], + "spans": [ + { + "bbox": [ + 317, + 513, + 553, + 578 + ], + "type": "text", + "content": "[22] Alec Radford, Jong Wook Kim, Chris Hallacy, Aditya Ramesh, Gabriel Goh, Sandhini Agarwal, Girish Sastry, Amanda Askell, Pamela Mishkin, Jack Clark, et al. Learning transferable visual models from natural language supervision. In Proceedings of the International Conference on Machine Learning, pages 8748-8763. PMLR, 2021. 5" + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 317, + 580, + 553, + 623 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 580, + 553, + 623 + ], + "spans": [ + { + "bbox": [ + 317, + 580, + 553, + 623 + ], + "type": "text", + "content": "[23] Lvmin Zhang, Anyi Rao, and Maneesh Agrawala. Adding conditional control to text-to-image diffusion models. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 3836-3847, 2023. 5" + } + ] + } + ], + "index": 25 + }, + { + "bbox": [ + 317, + 624, + 553, + 669 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 624, + 553, + 669 + ], + "spans": [ + { + "bbox": [ + 317, + 624, + 553, + 669 + ], + "type": "text", + "content": "[24] Riza Alp Güler, Natalia Neverova, and Iasonas Kokkinos. Densesepose: Dense human pose estimation in the wild. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 7297-7306, 2018. 6" + } + ] + } + ], + "index": 26 + }, + { + "bbox": [ + 317, + 670, + 553, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 670, + 553, + 713 + ], + "spans": [ + { + "bbox": [ + 317, + 670, + 553, + 713 + ], + "type": "text", + "content": "[25] Ke Gong, Xiaodan Liang, Yicheng Li, Yimin Chen, Ming Yang, and Liang Lin. Instance-level human parsing via part grouping network. In Proceedings of the European Conference on Computer Vision, pages 770-785, 2018. 6" + } + ] + } + ], + "index": 27 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "text", + "content": "5654" + } + ] + } + ], + "index": 29 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 8 + }, + { + "para_blocks": [ + { + "bbox": [ + 56, + 73, + 295, + 653 + ], + "type": "list", + "angle": 0, + "index": 11, + "blocks": [ + { + "bbox": [ + 56, + 73, + 294, + 116 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 73, + 294, + 116 + ], + "spans": [ + { + "bbox": [ + 56, + 73, + 294, + 116 + ], + "type": "text", + "content": "[26] Zhou Wang, Alan C Bovik, Hamid R Sheikh, and Eero P Simoncelli. Image quality assessment: from error visibility to structural similarity. IEEE transactions on image processing, 13(4):600-612, 2004. 6" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 56, + 118, + 295, + 173 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 118, + 295, + 173 + ], + "spans": [ + { + "bbox": [ + 56, + 118, + 295, + 173 + ], + "type": "text", + "content": "[27] Richard Zhang, Phillip Isola, Alexei A Efros, Eli Shechtman, and Oliver Wang. The unreasonable effectiveness of deep features as a perceptual metric. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 586-595, 2018. 6" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 56, + 174, + 294, + 228 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 174, + 294, + 228 + ], + "spans": [ + { + "bbox": [ + 56, + 174, + 294, + 228 + ], + "type": "text", + "content": "[28] Martin Heusel, Hubert Ramsauer, Thomas Unterthiner, Bernhard Nessler, and Sepp Hochreiter. Gans trained by a two time-scale update rule converge to a local nash equilibrium. Advances in Neural Information Processing Systems, 30, 2017. 6" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 56, + 229, + 294, + 284 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 229, + 294, + 284 + ], + "spans": [ + { + "bbox": [ + 56, + 229, + 294, + 284 + ], + "type": "text", + "content": "[29] Christian Szegedy, Vincent Vanhoucke, Sergey Ioffe, Jon Shlens, and Zbigniew Wojna. Rethinking the inception architecture for computer vision. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 2818-2826, 2016. 6" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 56, + 285, + 294, + 318 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 285, + 294, + 318 + ], + "spans": [ + { + "bbox": [ + 56, + 285, + 294, + 318 + ], + "type": "text", + "content": "[30] Karen Simonyan and Andrew Zisserman. Very deep convolutional networks for large-scale image recognition. arXiv preprint arXiv:1409.1556, 2014. 6" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 56, + 319, + 294, + 384 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 319, + 294, + 384 + ], + "spans": [ + { + "bbox": [ + 56, + 319, + 294, + 384 + ], + "type": "text", + "content": "[31] Adam Paszke, Sam Gross, Francisco Massa, Adam Lerer, James Bradbury, Gregory Chanan, Trevor Killeen, Zeming Lin, Natalia Gimelshein, Luca Antiga, et al. Pytorch: An imperative style, high-performance deep learning library. Advances in Neural Information Processing Systems, 32, 2019. 6" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 56, + 386, + 294, + 451 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 386, + 294, + 451 + ], + "spans": [ + { + "bbox": [ + 56, + 386, + 294, + 451 + ], + "type": "text", + "content": "[32] Patrick von Platen, Suraj Patil, Anton Lozhkov, Pedro Cuenca, Nathan Lambert, Kashif Rasul, Mishig Davaadorj, Dhruv Nair, Sayak Paul, William Berman, Yiyi Xu, Steven Liu, and Thomas Wolf. Diffusers: State-of-the-art diffusion models. https://github.com/huggingface/diffusers, 2022.6" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 56, + 453, + 294, + 474 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 453, + 294, + 474 + ], + "spans": [ + { + "bbox": [ + 56, + 453, + 294, + 474 + ], + "type": "text", + "content": "[33] Diederik P Kingma. Adam: A method for stochastic optimization. arXiv preprint arXiv:1412.6980, 2014. 6" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 56, + 475, + 294, + 529 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 475, + 294, + 529 + ], + "spans": [ + { + "bbox": [ + 56, + 475, + 294, + 529 + ], + "type": "text", + "content": "[34] Jia Deng, Wei Dong, Richard Socher, Li-Jia Li, Kai Li, and Li Fei-Fei. Imagenet: A large-scale hierarchical image database. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition, pages 248–255. IEEE, 2009. 6" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 56, + 531, + 294, + 586 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 531, + 294, + 586 + ], + "spans": [ + { + "bbox": [ + 56, + 531, + 294, + 586 + ], + "type": "text", + "content": "[35] Justin Johnson, Alexandre Alahi, and Li Fei-Fei. Perceptual losses for real-time style transfer and super-resolution. In Computer Vision-ECCV 2016: 14th European Conference, Amsterdam, The Netherlands, October 11-14, 2016, Proceedings, Part II 14, pages 694-711. Springer, 2016. 6" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 56, + 587, + 294, + 653 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 587, + 294, + 653 + ], + "spans": [ + { + "bbox": [ + 56, + 587, + 294, + 653 + ], + "type": "text", + "content": "[36] Tan Wang, Linjie Li, Kevin Lin, Yuanhao Zhai, Chung-Ching Lin, Zhengyuan Yang, Hanwang Zhang, Zicheng Liu, and Lijuan Wang. Disco: Disentangled control for realistic human dance generation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 9326-9336, 2024. 7" + } + ] + } + ], + "index": 10 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 749, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 749, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 749, + 315, + 757 + ], + "type": "text", + "content": "5655" + } + ] + } + ], + "index": 12 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 9 + } + ], + "_backend": "vlm", + "_version_name": "2.6.4" +} \ No newline at end of file diff --git a/2025/Disentangling Safe and Unsafe Image Corruptions via Anisotropy and Locality/3a52d538-e6a0-40a8-904e-0b81bce10c6a_content_list.json b/2025/Disentangling Safe and Unsafe Image Corruptions via Anisotropy and Locality/3a52d538-e6a0-40a8-904e-0b81bce10c6a_content_list.json new file mode 100644 index 0000000000000000000000000000000000000000..ce28a85c8c8a550cae22a83cd6d730bcd96d8ae7 --- /dev/null +++ b/2025/Disentangling Safe and Unsafe Image Corruptions via Anisotropy and Locality/3a52d538-e6a0-40a8-904e-0b81bce10c6a_content_list.json @@ -0,0 +1,1878 @@ +[ + { + "type": "text", + "text": "Disentangling Safe and Unsafe Image Corruptions via Anisotropy and Locality", + "text_level": 1, + "bbox": [ + 101, + 128, + 897, + 154 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Ramchandran Muthukumar*¹ Ambar Pal²† Jeremias Sulam¹ René Vidal³ \n¹Johns Hopkins University ²Amazon Web Services ³University of Pennsylvania", + "bbox": [ + 163, + 179, + 828, + 217 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Abstract", + "text_level": 1, + "bbox": [ + 246, + 251, + 326, + 266 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "State-of-the-art machine learning systems are vulnerable to small perturbations to their input, where \"small\" is defined according to a threat model that assigns a positive threat to each perturbation. Most prior works define a task-agnostic, isotropic, and global threat, like the $\\ell_p$ norm, where the magnitude of the perturbation fully determines the degree of the threat and neither the direction of the attack nor its position in space matter. However, common corruptions in computer vision, such as blur, compression, or occlusions, are not well captured by such threat models. This paper proposes a novel threat model called Projected Displacement (PD) to study robustness beyond existing isotropic and global threat models. The proposed threat model measures the threat of a perturbation via its alignment with unsafe directions, defined as directions in the input space along which a perturbation of sufficient magnitude changes the ground truth class label. Unsafe directions are identified locally for each input based on observed training data. In this way, the PD-threat model exhibits anisotropy and locality. Experiments on Imagenet-1k data indicate that, for any input, the set of perturbations with small PD threat includes safe perturbations of large $\\ell_p$ norm that preserve the true label, such as noise, blur and compression, while simultaneously excluding unsafe perturbations that alter the true label. Unlike perceptual threat models based on embeddings of large-vision models, the PD-threat model can be readily computed for arbitrary classification tasks without pre-training or finetuning. Further additional task information such as sensitivity to image regions or concept hierarchies can be easily integrated into the assessment of threat and thus the PD threat model presents practitioners with a flexible, task-driven threat specification that alleviates the limitations of $\\ell_p$ -threat models.", + "bbox": [ + 88, + 282, + 485, + 782 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "1. Introduction", + "text_level": 1, + "bbox": [ + 91, + 809, + 220, + 825 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Modern machine learning (ML) systems are nearing widespread deployment in civilian life, and hence a com", + "bbox": [ + 89, + 835, + 483, + 866 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "prehensive understanding of their security vulnerabilities is necessary [46]. One such vulnerability concerns the ability of a malicious adversary to tamper with predictions by adding small imperceptible corruptions referred to as adversarial perturbations [22, 42]. There is by now overwhelming evidence that carefully crafted adversarial perturbations can foil the prediction of state-of-the-art machine learning classifiers, i.e. they are not adversarily robust [6, 8]. Such vulnerabilities have been observed in a wide variety of applications like computer vision [1, 49], speech recognition [37, 55], autonomous driving [5, 45], and more. The goal of designing safe and reliable ML systems remains incomplete [4, 11] despite significant investment [23, 24, 33]. Often, strategies that intend to foil performance, i.e. adversarial attacks [6, 7, 22, 31, 44], have proven more successful than strategies aimed at mitigating vulnerabilities, i.e. adversarial defenses [3, 9, 34, 48, 52].", + "bbox": [ + 511, + 252, + 906, + 508 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Meaningfully evaluating the adversarial robustness of a machine learning system requires a formal specification of a threat function, and an associated threat model that limits the scope of malicious adversaries [20]. Informally, a threat function $d: \\mathbb{R}^d \\times \\mathbb{R}^d \\to \\mathbb{R}^{\\geq 0}$ measures the threat represented by a corruption $\\delta \\in \\mathbb{R}^d$ towards altering the true label at input $\\mathbf{x} \\in \\mathbb{R}^d$ as $d(\\mathbf{x}, \\delta)$ . We let $S(\\mathbf{x}, d, \\varepsilon) \\coloneqq \\{\\delta \\in \\mathbb{R}^d \\mid d(\\mathbf{x}, \\delta) \\leq \\varepsilon\\}$ denote the $\\varepsilon$ -sublevel set of perturbations where the threat $d$ is measured w.r.t. input $\\mathbf{x}$ . A threat model $(d, \\varepsilon)$ is a pair of a threat function $d$ and a permissible threshold $\\varepsilon$ that together define the set of permissible perturbations $S(\\mathbf{x}, d, \\varepsilon)$ . Under the threat model $(d, \\varepsilon)$ , the robust accuracy of a classifier $h \\in \\mathcal{H}$ is the probability that the label prediction at $\\mathbf{x}$ is locally invariant to corruptions within the permissible set $S(\\mathbf{x}, d, \\varepsilon)$ , see Definition 2 for an explicit definition.", + "bbox": [ + 511, + 510, + 908, + 750 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "One of the most commonly used threat models is the $\\ell_p$ -threat model $(d_p, \\epsilon)$ , corresponding to the choice $d_p(\\mathbf{x}, \\delta) \\coloneqq \\| \\delta \\|_p$ . The threat function $d_p$ is task-agnostic, easy to evaluate, and induces a compact sub-level set $S(\\mathbf{x}, d_p, \\varepsilon)$ that allows efficient projection. Hence, the $\\ell_p$ -threat model $(d_p, \\varepsilon)$ presents a natural starting point for in", + "bbox": [ + 511, + 752, + 908, + 844 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "CVF", + "bbox": [ + 106, + 2, + 181, + 42 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore.", + "bbox": [ + 236, + 0, + 810, + 46 + ], + "page_idx": 0 + }, + { + "type": "page_footnote", + "text": "1The flexibility to define threat w.r.t. input $\\mathbf{x}$ is critical to our work.", + "bbox": [ + 532, + 854, + 887, + 868 + ], + "page_idx": 0 + }, + { + "type": "page_footnote", + "text": "2For any vector $\\pmb {\\delta}\\in \\mathbb{R}^d$ , the $\\ell_p$ norm $\\| \\cdot \\| _p$ for $p\\geq 1$ is defined as $\\| \\pmb {\\delta}\\| _p\\coloneqq \\left(\\sum_i|\\delta_i|^p\\right)^{\\frac{1}{p}}.$", + "bbox": [ + 514, + 868, + 903, + 902 + ], + "page_idx": 0 + }, + { + "type": "page_footnote", + "text": "* Corresponding author: rmuthukl@jhu.edu", + "bbox": [ + 107, + 875, + 400, + 887 + ], + "page_idx": 0 + }, + { + "type": "page_footnote", + "text": "This work is not related to AP's position at Amazon", + "bbox": [ + 109, + 887, + 390, + 900 + ], + "page_idx": 0 + }, + { + "type": "page_number", + "text": "9954", + "bbox": [ + 482, + 946, + 514, + 957 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/c3652cb8b0020bc5a6e618359cc909da7e2e65b34bf021d70c7774d12e0b2112.jpg", + "image_caption": [ + "Figure 1. Corruptions with equal $\\ell_{\\infty}$ -threat, $\\|\\delta_1\\|_{\\infty} = \\|\\delta_2\\|_{\\infty} = \\|\\delta_3\\|_{\\infty} = \\|\\delta_4\\|_{\\infty}$ , but varying PD-threat." + ], + "image_footnote": [], + "bbox": [ + 150, + 87, + 428, + 265 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "vestigating robustness. RobustBench [8] maintains an up-to-date leaderboard of the robust accuracies of benchmark models under $(d_p, \\varepsilon)$ . Unfortunately, the progress towards achieving perfect adversarial robustness (i.e. $100\\%$ robust test accuracy) in RobustBench has plateaued in recent years. Unlike supervised learning on clean data, scaling data, model size and computing resources might be insufficient to bridge the gap [4, 11]. Below we expand further on the fundamental limitations of isotropic and global threat models.", + "bbox": [ + 89, + 334, + 483, + 470 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "1.1. Motivation: Specification of Threat Model", + "text_level": 1, + "bbox": [ + 89, + 483, + 452, + 500 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "For image-based datasets, it is widely recognized that $\\ell_p$ norms are neither necessary nor sufficient for capturing perceptual similarity [39, 40], which poses significant challenges for accurately evaluating robustness [26, 43]. In Figure 1, $\\mathbf{x}$ (at the center) is an image of class ROSE HIP from the Imagenet-1k dataset, and $\\mathbf{x} + \\delta_1$ , $\\mathbf{x} + \\delta_2$ are two corrupted images equidistant (w.r.t. $\\ell_{\\infty}$ ) from $\\mathbf{x}$ , i.e. $\\| \\delta_1 \\|_{\\infty} = \\| \\delta_2 \\|_{\\infty} = \\varepsilon$ . Intuitively, both $\\mathbf{x} + \\delta_1$ and $\\mathbf{x} + \\delta_2$ share the same label as $\\mathbf{x}$ , since its appearance still depicts the fruit ROSE HIP. We refer to such perturbations that preserve the class label as safe, while those that change it are termed unsafe. Consider now moving further along these directions to $\\mathbf{x} + t\\delta_1$ and $\\mathbf{x} + t\\delta_2$ , for $t > 1$ . These perturbed points are again equidistant from $\\mathbf{x}$ , but while $t\\delta_1$ does not alter the true label (i.e. it is safe), $t\\delta_2$ does (and it is unsafe). As a result, any threat model $(d_{\\infty}, t' \\varepsilon)$ with $t' \\geq t$ necessarily incurs misspecification: any classifier $h$ stable to all perturbations in $S(\\mathbf{x}, d_{\\infty}, t' \\varepsilon)$ produces an incorrect prediction at $\\mathbf{x} + t\\delta_2$ , i.e. $h(\\mathbf{x} + t\\delta_2) = \\text{ROSE HIP}$ . On the other hand, shrinking the threat model is also futile, as then the safe perturbation $t\\delta_1$ is necessarily excluded.", + "bbox": [ + 89, + 506, + 483, + 823 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "This illustrates that standard $\\ell_p$ -threat models are unable to distinguish between the safe perturbations that preserve the true label and the unsafe perturbations that alter the true label. Hence, perfect robust accuracy under $\\ell_p$ threat models might be neither achievable nor desirable. Attempting to", + "bbox": [ + 89, + 824, + 483, + 901 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "resolve these limitations, Laidlaw et al. [32] formulate perceptual threat models via neural perceptual metric based on neural representations. Unfortunately (but unsurprisingly), all neural perceptual distance metrics are themselves vulnerable to adversarial attacks [17-19, 29].", + "bbox": [ + 511, + 90, + 906, + 167 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "1.2. Summary of Contributions", + "text_level": 1, + "bbox": [ + 513, + 179, + 756, + 194 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "We propose a novel threat function called Projected Displacement (PD), denoted by $d_{\\mathrm{PD}}$ (see Definition 7 for an explicit description), that accounts for the local directional statistics of the data via unsafe directions3 $\\mathcal{U}(\\mathbf{x})$ (Definition 6) at each input $\\mathbf{x}$ . For an input $\\mathbf{x}$ with label $y$ , any input $\\tilde{\\mathbf{x}}$ with label $\\tilde{y} \\neq y$ represents the unsafe direction $\\mathbf{u} := \\frac{\\tilde{\\mathbf{x}} - \\mathbf{x}}{\\|\\tilde{\\mathbf{x}} - \\mathbf{x}\\|_2}$ , which will be estimated from observed training data. Then, the degree of threat of each perturbation $\\delta$ is the maximal alignment with the set of unsafe directions: A large threat $d_{\\mathrm{PD}}(\\mathbf{x},\\delta)$ indicates that $\\delta$ is well-aligned with an unsafe direction $\\mathbf{u} \\in \\mathcal{U}(\\mathbf{x})$ . We highlight some key properties.", + "bbox": [ + 511, + 200, + 906, + 369 + ], + "page_idx": 1 + }, + { + "type": "list", + "sub_type": "text", + "list_items": [ + "1. Disentangling safe vs unsafe perturbations. PD-threat is able to distinguish between safe and unsafe perturbations of equal $\\ell_p$ norms. From Figure 1, different corruptions (Gaussian noise $t\\delta_1$ , motion blur $t\\delta_3$ and saturation $t\\delta_4$ ) applied to the original image $\\mathbf{x}$ all have equal $\\ell_{\\infty}$ norm of 240/255. Despite this, the true label of these perturbed samples does not change, unlike that of corruption $t\\delta_2$ , which results in an image $\\mathbf{x} + t\\delta_2$ with a different class. For each perturbation $t\\delta_i$ , the computed values of PD threat in Figure 1 naturally reflect how close they are to changing the class label. PD threat model is competent with state-of-the-art neural perceptual threat model DreamSim (denoted $d_{\\mathrm{DS}}$ ) [15] for distinguishing safe and unsafe perturbations (see Section 4).", + "2. Ease of use. As we will show, the sub-level set $S(\\mathbf{x}, d_{\\mathrm{PD}}, \\varepsilon)$ at each input $\\mathbf{x}$ is convex, allowing for efficient projections onto it. This enables a straightforward plug-and-play mechanism for adapting existing adversarial attacks to the PD-threat model. We incorporate such an extension of AutoAttack [7] to evaluate the robustness under the PD-threat model ( $d_{\\mathrm{PD}}, \\varepsilon$ ) for state-of-the-art robust classifiers registered at RobustBench [8].", + "3. Task-dependency. PD-threat models can readily integrate additional task-relevant information via label annotation, such as inter-class hierarchy (e.g. the Wordnet hierarchy) or pixel annotation (e.g. segmentation masks)—see Table 1—allowing for variations in the form of robustness while retaining the above benefits (anisotropy, locality, convex sub-level sets, and efficient projection)." + ], + "bbox": [ + 514, + 369, + 908, + 808 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "1.3. Related Work", + "text_level": 1, + "bbox": [ + 514, + 820, + 656, + 834 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "In this section, we highlight prior work that explores robustness beyond $\\ell_p$ -threat models.", + "bbox": [ + 511, + 843, + 906, + 873 + ], + "page_idx": 1 + }, + { + "type": "page_footnote", + "text": "3Directions refer to vectors $\\mathbf{u}\\in \\mathbb{R}^d$ with unit $\\ell_2$ norm.", + "bbox": [ + 529, + 886, + 823, + 898 + ], + "page_idx": 1 + }, + { + "type": "page_number", + "text": "9955", + "bbox": [ + 482, + 946, + 514, + 957 + ], + "page_idx": 1 + }, + { + "type": "table", + "img_path": "images/d7da485e19ce5281492dd24d0c0b0a4723f27000460e96412a8054fd79194af3.jpg", + "table_caption": [], + "table_footnote": [], + "table_body": "
PROPERTYFEATURESdpdDS[15]dPD
Degree of ThreatAnisotropy, Locality×✓✓
Label AnnotationInter-class Distance×~✓
Pixel AnnotationSegmentation✓×✓
", + "bbox": [ + 91, + 88, + 480, + 157 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Table 1. Comparison of Threat Models", + "bbox": [ + 168, + 167, + 406, + 181 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Common Corruptions. Neural networks are vulnerable to image distortions (e.g. translation, rotation), common corruptions (e.g. blur, noise), changes in lighting, view, depth, etc. [12, 16, 25, 27, 28, 36, 50]. Note that these are known safe perturbations as they preserve the true label, even if they have large $\\ell_p$ norms. To measure this vulnerability, [25] formalized the notion of corruption robustness and presented a standardized benchmark dataset, Imagenet-C containing 19 styles of common image corruptions (categorized into noise, blur, weather, and digital). Training on Imagenet-C, or on other augmentations, can result in over-fitting, particularly to specific types of corruptions [36]. Additionally, [36] introduced a newdataset, Imagenet-C, with 10 new corruption styles identified from a larger set of real-world corruptions. These will be useful later in our experiments.", + "bbox": [ + 88, + 199, + 485, + 425 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Perceptual Distance Metrics. Since $\\ell_p$ provides inaccurate approximations to perceptual distances, several works have attempted to propose alternatives, such as SSIM [47], FSIM [53], and HDR-VDP-2 [35]. With the advent of learning data-driven representations, [15, 54] hail the effectiveness of internal representations of neural classifiers in capturing perceptual similarity. DreamSim [15], the current state-of-the-art among the neural perceptual distance metrics, is finetuned on human perceptual similarity judgements.", + "bbox": [ + 89, + 426, + 483, + 563 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "2. Key Definitions", + "text_level": 1, + "bbox": [ + 89, + 575, + 243, + 592 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Notation. We denote scalar quantities in Roman letters and vectors in boldface Roman letters. The empty set is given by $\\emptyset$ . For any set $\\mathcal{V}$ , we denote by $2^{\\mathcal{V}}$ the power set of all its subsets. For any set $\\mathcal{V} \\subset \\mathbb{R}^d$ and a vector $\\mathbf{v} \\in \\mathbb{R}^d$ , $\\mathcal{V} - \\{\\mathbf{v}\\}$ is the Minkowski difference, i.e. the set $\\{\\mathbf{w} - \\mathbf{v} \\mid \\forall \\mathbf{w} \\in \\mathcal{V}\\}$ (subtracting $\\mathbf{v}$ from all elements of $\\mathcal{V}$ ). We denote the unit ball and the unit sphere in $\\mathbb{R}^d$ w.r.t. the $\\ell_p$ norm by $\\mathbb{B}_p^d := \\{\\mathbf{v} \\in \\mathbb{R}^d \\mid \\| \\mathbf{v} \\|_p \\leq 1\\}$ and $\\mathbb{S}_p^d := \\{\\mathbf{v} \\in \\mathbb{R}^d \\mid \\| \\mathbf{v} \\|_p = 1\\}$ , respectively. A direction is an element of the unit sphere $\\mathbb{S}_2^d$ .", + "bbox": [ + 89, + 601, + 485, + 743 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "2.1. Supervised Learning", + "text_level": 1, + "bbox": [ + 89, + 751, + 287, + 767 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "We consider a supervised learning task with a bounded input domain $\\mathcal{X} \\subset \\mathbb{R}^d$ and labels $\\mathcal{Y} = \\{1, \\dots, C\\}$ such that each input $\\mathbf{x}$ is assigned a label by a deterministic true $^4$ labeling function $h^\\star : \\mathcal{X} \\to \\mathcal{Y}$ . The true labeling function $h^\\star$ partitions the input domain $\\mathcal{X} := \\cup_y \\mathcal{X}_y$ , where $\\mathcal{X}_y$ is the set of inputs assigned label $y$ , i.e. $\\mathcal{X}_y = \\{\\mathbf{x} \\in \\mathcal{X} \\mid h^\\star(\\mathbf{x}) = y\\}$ .", + "bbox": [ + 89, + 773, + 483, + 864 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "We assume that the partition sets $\\mathcal{X}_1, \\ldots, \\mathcal{X}_C$ are nonempty and open in order to avoid degeneracies.", + "bbox": [ + 511, + 90, + 906, + 119 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "We let $\\mathcal{D}_{\\mathcal{X}}$ be a data distribution over the input domain $\\mathcal{X}$ and $\\mathcal{D}_{\\mathcal{Z}}$ the extension of $\\mathcal{D}_{\\mathcal{X}}$ to a distribution over the joint domain $\\mathcal{Z} \\coloneqq \\mathcal{X} \\times \\mathcal{Y}$ using the true labeling function $h^{\\star}$ . The true labeling function $h^{\\star}$ and the marginal input distribution $\\mathcal{D}_{\\mathcal{X}}$ are unknown, but we observe finite labeled training data $\\mathrm{S} \\coloneqq \\left\\{(\\mathbf{x}^{(1)}, y^{(1)}), \\ldots, (\\mathbf{x}^{(m)}, y^{(m)})\\right\\} \\stackrel{\\mathrm{i.i.d.}}{\\sim} (\\mathcal{D}_{\\mathcal{Z}})^{m}$ . For any classifier $h: \\mathcal{X} \\to \\mathcal{Y}$ , accuracy is the probability of correctly labeling a random input, i.e. $\\mathrm{Acc}(h) \\coloneqq \\mathrm{Prob}_{\\mathbf{x} \\sim \\mathcal{D}_{\\mathcal{X}}} [h(\\mathbf{x}) = h^{\\star}(\\mathbf{x})]$ . The task of supervised learning seeks a classifier $\\bar{h}$ in a hypothesis class $\\mathcal{H} \\subset \\{h: \\mathcal{X} \\to \\mathcal{Y}\\}$ with high accuracy.", + "bbox": [ + 511, + 119, + 908, + 277 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "2.2. Adversarial Perturbations", + "text_level": 1, + "bbox": [ + 511, + 286, + 751, + 300 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "In this article, we formally define adversarial perturbations through the lens of the partition sets $\\mathcal{X}_y$ .", + "bbox": [ + 511, + 306, + 905, + 339 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Definition 1 (Adversarial Perturbation). A perturbation $\\delta \\in \\mathbb{R}^d$ is adversarial for predictor $h$ at input $\\mathbf{x}$ if it is,", + "bbox": [ + 511, + 347, + 905, + 378 + ], + "page_idx": 2 + }, + { + "type": "list", + "sub_type": "text", + "list_items": [ + "1. Domain constrained: $\\mathbf{x} + \\delta \\in \\mathcal{X}$ , i.e. $\\delta \\in \\mathcal{X} - \\{\\mathbf{x}\\}$ .", + "2. Label invariant: $\\mathbf{x} + \\boldsymbol{\\delta}$ has the same true label as $\\mathbf{x}$ , i.e. $h^{\\star}(\\mathbf{x} + \\boldsymbol{\\delta}) = h^{\\star}(\\mathbf{x})$ .", + "3. Adversarial: $\\mathbf{x} + \\boldsymbol{\\delta}$ is misclassified by $h$ , i.e. $h(\\mathbf{x} + \\boldsymbol{\\delta}) \\neq h^{\\star}(\\mathbf{x} + \\boldsymbol{\\delta})$ ." + ], + "bbox": [ + 513, + 378, + 906, + 454 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "The labels assigned by $h^{\\star}$ are considered the ground truth and thus $h^{\\star}$ has no adversarial perturbation at any input. The presence of adversarial perturbations indicates an imperfect labeling function $h$ . Based on Definition 1, perturbations $\\delta \\in \\mathbb{R}^{d}$ such that $\\mathbf{x} + \\delta \\notin \\mathcal{X}$ are not deemed adversarial as the classifier $h$ is not required to predict labels outside the input domain. On the other hand, for $\\mathbf{x}$ with true label $y$ , a perturbation $\\delta$ such that $h^{\\star}(\\mathbf{x} + \\delta) \\neq y$ cannot be adversarial at $\\mathbf{x}$ as the corrupted input $\\mathbf{x} + \\delta$ has a different true label. Finally, if $h$ predicts the label at $\\mathbf{x}$ incorrectly, then $\\mathbf{0}$ is already an adversarial perturbation.", + "bbox": [ + 511, + 464, + 908, + 630 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "2.2.1. Adversarial Robustness via Threat Specification", + "text_level": 1, + "bbox": [ + 511, + 638, + 893, + 652 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "A threat function $d: \\mathcal{X} \\times \\mathbb{R}^d \\to \\mathbb{R}^{\\geq 0}$ presents practitioners a way to specify the kind of robustness suitable for a particular learning task. The $\\varepsilon$ -sublevel set of a threat function $d$ at an input $\\mathbf{x}$ , $S(\\mathbf{x}, d, \\varepsilon) := \\{\\delta \\in \\mathbb{R}^d \\mid d(\\mathbf{x}, \\delta) \\leq \\varepsilon\\}$ , defines the set of perturbations with a threat at most $\\varepsilon$ .", + "bbox": [ + 511, + 656, + 905, + 733 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Definition 2 (Threat Model, Robust Accuracy). A threat model $(d,\\varepsilon)$ is a pair of threat function $d$ and permissible threshold $\\varepsilon$ that together define the set of permissible perturbations at each input $\\mathbf{x}$ as the sublevel set $S(\\mathbf{x},d,\\varepsilon)$ . Under the threat model $(d,\\varepsilon)$ , the robust accuracy RobustAcc $(h,(d,\\varepsilon))$ of a classifier $h\\in \\mathcal{H}$ is the probability of stable prediction upon corruptions within the permissible set, i.e. $\\mathrm{Prob}_{\\mathbf{x}}\\big[\\forall \\delta \\in S(\\mathbf{x},d,\\varepsilon), h(\\mathbf{x} + \\delta) = h^{\\star}(\\mathbf{x})\\big]$ .", + "bbox": [ + 511, + 743, + 908, + 866 + ], + "page_idx": 2 + }, + { + "type": "page_footnote", + "text": "${}^{5}$ Sampling $\\mathbf{z} \\sim {\\mathcal{D}}_{\\mathcal{Z}}$ is equivalent to sampling(x,h*(x))where $\\mathbf{x} \\sim$ ${\\mathcal{D}}_{\\mathcal{X}}$ .", + "bbox": [ + 511, + 875, + 908, + 900 + ], + "page_idx": 2 + }, + { + "type": "page_footnote", + "text": "4Realizability in the context of adversarial robustness has also been studied in Awasthi et al. [2], Tramér et al. [43].", + "bbox": [ + 89, + 875, + 482, + 900 + ], + "page_idx": 2 + }, + { + "type": "page_number", + "text": "9956", + "bbox": [ + 482, + 946, + 514, + 957 + ], + "page_idx": 2 + }, + { + "type": "image", + "img_path": "images/729f79c773047969e68a073fc968d08f975b0727244bac0404b0de377195067e.jpg", + "image_caption": [ + "Figure 2. An illustration of unsafe directions, and sub-level sets of the PD-threat." + ], + "image_footnote": [], + "bbox": [ + 150, + 85, + 428, + 272 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Thus we define robustness as the stability of a classifier's prediction under perturbations within the permissible sets. The task of robust supervised learning seeks a classifier in $\\mathcal{H}$ with high robust accuracy.", + "bbox": [ + 89, + 338, + 483, + 398 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Definition 3 ( $\\varepsilon$ -robust). A classifier $h \\in \\mathcal{H}$ is $\\varepsilon$ -robust at $\\mathbf{x}$ w.r.t. a threat function $d$ if, $\\forall \\delta \\in S(\\mathbf{x}, d, \\varepsilon) \\cap (\\mathcal{X} - \\{\\mathbf{x}\\})$ , we have $h(\\mathbf{x} + \\delta) = h^{\\star}(\\mathbf{x})$ .", + "bbox": [ + 89, + 409, + 483, + 455 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "An $\\varepsilon$ -robust classifier predicts the same label $h^{\\star}(\\mathbf{x})$ at $\\mathbf{x} + \\delta$ for any corruption $\\delta$ within the $\\varepsilon$ -sublevel set $S(\\mathbf{x}, d, \\varepsilon)$ . We emphasize that robust accuracy and $\\varepsilon$ -robustness are measures of stability rather than correctness. In this way, each threat model $(d, \\varepsilon)$ implicitly encodes a trade-off between stability and correctness. [13]. Requiring more stability than necessary can lead to incorrect predictions as in Figure 1.", + "bbox": [ + 89, + 465, + 483, + 571 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "3. Threat specification beyond $\\ell_p$ norms", + "text_level": 1, + "bbox": [ + 89, + 585, + 426, + 604 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "In this section, we design a principled threat model from first principles that is local and anisotropic. We first theoretically develop our exact PD\\*-threat model assuming knowledge of the decision boundaries of the true labeling function $h^\\star$ (Section 3.1), and later relax this assumption by developing an approximation leveraging the observed training data and obtaining a practical PD-threat (Section 3.2).", + "bbox": [ + 89, + 612, + 483, + 717 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "3.1. Measuring threat with class partition", + "text_level": 1, + "bbox": [ + 89, + 728, + 413, + 744 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Definition 4 (Unsafe Directions). At each input $\\mathbf{x}$ , a direction $\\mathbf{u} \\in \\mathbb{S}_2^d$ is called unsafe if there exists a step size $t \\geq 0$ such that $\\mathbf{x} + t\\mathbf{u} \\in \\mathcal{X}$ and $h^\\star(\\mathbf{x} + t\\mathbf{u}) \\neq h^\\star(\\mathbf{x})$ . We denote the set of all unsafe directions at $\\mathbf{x}$ as $\\mathcal{U}^\\star(\\mathbf{x})$ .", + "bbox": [ + 89, + 750, + 483, + 811 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "The minimum step size needed to alter the true label can vary across unsafe directions. Further, $\\mathcal{U}^{\\star}(\\mathbf{x})$ varies", + "bbox": [ + 89, + 821, + 483, + 853 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "with input. Directions $\\mathbf{v} \\in \\mathcal{B}_2^d \\cap (\\mathcal{U}^\\star(\\mathbf{x}))^c$ are called safe, since moving along them (whilst domain constrained) cannot alter the true label. We note that adversarial perturbations (Definition 1) are a characteristic of a learned predictor $h \\in \\mathcal{H}$ while unsafe directions are a characteristic of the true labeling function $h^\\star$ . We propose to measure threat of a perturbations $\\delta$ using $\\mathcal{U}^\\star(\\mathbf{x})$ .", + "bbox": [ + 511, + 90, + 906, + 196 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Definition 5 (PD\\* -threat). Let $\\mathbf{x}\\in \\mathcal{X}$ and let $\\delta \\in \\mathbb{R}^d$ be a perturbation. The exact projected displacement threat function $d_{PD}^{\\star}$ is defined as the maximum scaled displacement of the Euclidean projection of the perturbation $\\delta$ over all unsafe directions,", + "bbox": [ + 511, + 205, + 906, + 282 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\nd _ {\\mathrm {P D}} ^ {\\star} (\\mathbf {x}, \\boldsymbol {\\delta}) := \\sup _ {\\mathbf {u} \\in \\mathcal {U} ^ {\\star} (\\mathbf {x})} \\frac {1}{g ^ {\\star} (\\mathbf {x} , \\mathbf {u})} \\max \\left(\\langle \\boldsymbol {\\delta}, \\mathbf {u} \\rangle , 0\\right) \\quad (1)\n$$\n", + "text_format": "latex", + "bbox": [ + 534, + 292, + 906, + 327 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "where $g^{\\star}(\\mathbf{u},\\mathbf{x})$ is called the normalization function,", + "bbox": [ + 511, + 338, + 857, + 354 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\ng ^ {\\star} (\\mathbf {x}, \\mathbf {u}) := \\sup _ {\\mathbb {R} ^ {\\geq 0}} M \\text {s . t .} \\mathbf {x} + t \\mathbf {u} \\in \\mathcal {X} _ {h ^ {\\star} (\\mathbf {x})} \\forall t \\in [ 0, M ].\n$$\n", + "text_format": "latex", + "bbox": [ + 522, + 364, + 893, + 391 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "At each $\\mathbf{x}$ , the normalization function $g^{*}(\\mathbf{x},\\mathbf{u})$ captures the distance along the direction $\\mathbf{u}$ to the boundary of the set $\\mathcal{X}_{h^{\\star}(\\mathbf{x})}$ and enables a comparison of the threat along different unsafe directions (see Figure 2). In (1), the supremum ensures that we do not underestimate the threat along a direction that potentially alters the true label. We highlight two key characteristics.", + "bbox": [ + 511, + 402, + 906, + 508 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Anisotropy. The threat function $d_{\\mathrm{PD}}^{\\star}$ is anisotropic (the threat of a perturbation $\\delta$ at input $\\mathbf{x}$ depends on both the direction and magnitude of the perturbation) and does not correspond to a norm in general, unlike $\\ell_p$ threat models. For a fixed $\\mathbf{x}$ , the threat along different directions $\\mathbf{u}$ varies based on the alignment with unsafe directions $\\mathcal{U}^{\\star}(\\mathbf{x})$ . In particular, $d_{\\mathrm{PD}}^{\\star}$ can differentiate between the perturbations $\\delta$ and $-\\delta$ .", + "bbox": [ + 511, + 508, + 906, + 614 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Locality. The threat function $d_{\\mathrm{PD}}^{\\star}$ exhibits locality; i.e. a fixed perturbation $\\delta$ has varying threat $d_{\\mathrm{PD}}^{\\star}(\\mathbf{x}, \\delta)$ depending on the input $\\mathbf{x}$ . Thus the permissible set varies with input7, enabling an input-aware assessment of threat. We note that weighted $\\ell_p$ norms [14] can be anisotropic but not local.", + "bbox": [ + 511, + 614, + 906, + 691 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "3.1.1. Stability vs Correctness Trade-off", + "text_level": 1, + "bbox": [ + 511, + 698, + 795, + 713 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Since the normalization function $g^{\\star}$ accounts for the class partition $\\{\\mathcal{X}_c\\}_{c=1}^C$ , the threat function $d_{\\mathrm{PD}}^{\\star}$ has a principled trade-off between stability and correctness. Specifically, the true labeling function $h^{\\star}$ exhibits a high level of robustness under the PD $^{\\star}$ -threat8.", + "bbox": [ + 511, + 717, + 905, + 792 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Theorem 1. The true labeling function $h^{\\star}$ is 1-robust at any input $\\mathbf{x} \\in \\mathcal{X}$ w.r.t. the threat $d_{\\mathrm{PD}}^{\\star}$ . Additionally, if a classifier $h$ is not 1-robust at all inputs then there exists an input $\\mathbf{x}$ misclassified by $h$ .", + "bbox": [ + 511, + 804, + 906, + 864 + ], + "page_idx": 3 + }, + { + "type": "page_footnote", + "text": "Unlike $\\ell_p$ threat models where $S(\\mathbf{x},d_p,\\varepsilon) = \\varepsilon \\mathcal{B}_p^d\\forall \\mathbf{x}\\in \\mathcal{X}$", + "bbox": [ + 529, + 873, + 867, + 887 + ], + "page_idx": 3 + }, + { + "type": "page_footnote", + "text": "The proofs of all our results can be found in Section 8.", + "bbox": [ + 532, + 887, + 825, + 898 + ], + "page_idx": 3 + }, + { + "type": "page_footnote", + "text": "Note as per Definition 2, robust accuracy measures stability rather than correctness since we aren't evaluating probability of the event $\\{\\forall \\delta \\in S(\\mathbf{x},d,\\varepsilon), h(\\mathbf{x} + \\delta) = h^{\\star}(\\mathbf{x} + \\delta)\\}$ .", + "bbox": [ + 89, + 862, + 483, + 901 + ], + "page_idx": 3 + }, + { + "type": "page_number", + "text": "9957", + "bbox": [ + 482, + 946, + 514, + 957 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "In other words, for any pairs of inputs $\\mathbf{x},\\tilde{\\mathbf{x}}\\in \\mathcal{X}$ with different labels, the threat $d_{\\mathrm{PD}}^{\\star}(\\mathbf{x},\\tilde{\\mathbf{x}} -\\mathbf{x})$ is always larger than 1. Thus seeking $h\\in \\mathcal{H}$ that is 1-robust at any input under $d_{\\mathrm{PD}}^{\\star}$ is not at odds with learning $h^{\\star}$ .", + "bbox": [ + 89, + 90, + 483, + 154 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "3.2. Measuring threats with observed data", + "text_level": 1, + "bbox": [ + 89, + 160, + 418, + 176 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Computing $d_{\\mathrm{PD}}^{\\star}(\\mathbf{x},\\cdot)$ from Definition 5 requires explicitly characterizing the set of unsafe directions $\\mathcal{U}^{\\star}(\\mathbf{x})$ and the normalization function $g^{\\star}(\\mathbf{u},\\mathbf{x})$ , both of which need knowledge of the partition sets $\\{\\mathcal{X}_c\\}_{c = 1}^C$ induced by $h^\\star$ . In this subsection, we propose a heuristic empirical approximation.", + "bbox": [ + 89, + 181, + 483, + 257 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Approximating $\\mathcal{U}^{\\star}(\\mathbf{x})$ . We have access to training data S and incomplete information on the partition sets $\\{\\mathcal{X}_c\\}_{c=1}^C$ via the partition $S = \\{S_c\\}_{c=1}^C$ , where $S_c := \\{\\mathbf{x} \\in \\mathcal{X} \\mid (\\mathbf{x}, c) \\in S\\} \\subset \\mathcal{X}_c$ . At every training input $\\mathbf{x} \\in S_y$ for each input $\\tilde{\\mathbf{x}} \\in S_c$ (for $c \\neq y$ ), the direction $\\mathbf{u} = \\frac{\\tilde{\\mathbf{x}} - \\mathbf{x}}{\\|\\tilde{\\mathbf{x}} - \\mathbf{x}\\|_2}$ is an unsafe direction. For computational considerations, we choose a collection $\\bigcup_{c} S_{c,k}$ of representative subsets $S_{c,k} \\subset S_c$ of size $k$ . As a heuristic choice, we select subsets $S_{c,k}$ by solving a discrete-k-center optimization problem (detailed in Sec. 9.1). We are now ready to define an observed subset of the unsafe directions based on $S_{c,k}$ .", + "bbox": [ + 89, + 258, + 483, + 426 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Definition 6 (Observed Unsafe Directions). For each $\\mathbf{x} \\in S_y$ , we identify a set of $(C - 1)k$ -observed unsafe directions based on the representative subsets $S_{c,k}$ ,", + "bbox": [ + 89, + 434, + 483, + 481 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {U} _ {k} (\\mathbf {x}) := \\left\\{\\frac {\\tilde {\\mathbf {x}} - \\mathbf {x}}{\\| \\tilde {\\mathbf {x}} - \\mathbf {x} \\| _ {2}} \\Big | \\tilde {\\mathbf {x}} \\in \\cup_ {c \\neq y} \\mathbb {S} _ {c, k} \\right\\} \\subset \\mathcal {U} ^ {\\star} (\\mathbf {x}).\n$$\n", + "text_format": "latex", + "bbox": [ + 120, + 489, + 449, + 525 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Approximating $g^{\\star}(\\mathbf{x},\\mathbf{u})$ . At any $\\mathbf{x}\\in \\mathcal{X}_y$ , for each $\\mathbf{u}\\in \\mathcal{U}_k(\\mathbf{x})$ , there exists $\\tilde{\\mathbf{x}}\\in \\cup_{c\\neq y}\\mathrm{S}_{c,k}$ such that $\\mathbf{u} = \\frac{\\tilde{\\mathbf{x}} - \\mathbf{x}}{\\|\\tilde{\\mathbf{x}} - \\mathbf{x}\\|_2}$ . normalization function can be bounded as $g^{\\star}(\\mathbf{x},\\mathbf{u})\\leq \\| \\tilde{\\mathbf{x}} -\\mathbf{x}\\| _2$ , since $h^\\star (\\tilde{\\mathbf{x}})\\neq h^\\star (\\mathbf{x})$ . In practice, we use the heuristic $g_{\\beta}(\\mathbf{u},\\mathbf{x})\\coloneqq \\beta \\| \\tilde{\\mathbf{x}} -\\mathbf{x}\\| _2$ , where $\\beta \\in (0,1)$ is a scaling hyper-parameter.", + "bbox": [ + 89, + 539, + 483, + 633 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Definition 7 $((k,\\beta)$ -PD threat). Let $\\mathbf{x}\\in \\mathcal{X}$ and let $\\delta \\in \\mathbb{R}^d$ be a perturbation. The projected displacement threat is defined as", + "bbox": [ + 89, + 642, + 483, + 686 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\nd _ {\\mathrm {P D}, k, \\beta} (\\mathbf {x}, \\boldsymbol {\\delta}) := \\max _ {\\mathbf {u} \\in \\mathcal {U} _ {k} (\\mathbf {x})} \\frac {1}{g _ {\\beta} (\\mathbf {x} , \\mathbf {u})} \\max \\left(\\langle \\boldsymbol {\\delta}, \\mathbf {u} \\rangle , 0\\right).\n$$\n", + "text_format": "latex", + "bbox": [ + 104, + 693, + 467, + 726 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "The quality of the approximation of the exact threat $d_{\\mathrm{PD}}^{\\star}$ by $d_{\\mathrm{PD,k},\\beta}$ depends on the choice of representative unsafe directions $\\mathcal{U}_k(\\cdot)$ via the $k$ -subset $\\mathbb{S}_{c,k}$ and on the heuristic choice of the approximate normalization $g_{\\beta}$ , via the scaling hyper-parameter $\\beta$ . In this article, we fix $(k,\\beta) = (50,\\frac{1}{2})$ as the default hyper-parameters (see Section 9.3 for an expanded discussion on choosing hyper-parameters $(k,\\beta)$ for any learning task). For brevity, we refer to the observed unsafe directions (Definition 6), the heuristic normalizations and the PD threat (Definition 7) as $\\mathcal{U}(\\mathbf{x}), g(\\mathbf{x},\\delta)$ and $d_{\\mathrm{PD}}(\\mathbf{x},\\delta)$ , respectively.", + "bbox": [ + 89, + 734, + 485, + 902 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "We now state a few important properties of our proposed empirical approximation $d_{\\mathrm{PD}}$ . Clearly $d_{\\mathrm{PD}}$ inherits the anisotropy and locality properties of the exact threat $d_{\\mathrm{PD}}^{\\star}$ .", + "bbox": [ + 511, + 90, + 903, + 137 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Growth and Sensitivity. The growth of PD threat along any direction is linear, i.e., $d_{\\mathrm{PD}}(\\mathbf{x},t\\boldsymbol {\\delta}) = td_{\\mathrm{PD}}(\\mathbf{x},\\boldsymbol {\\delta})$ , and the rate of growth scales inversely with the (approximate) normalization. In particular, for a fixed $\\mathbf{x}$ $d_{\\mathrm{PD}}(\\mathbf{x},\\cdot)$ is $(\\max_{u\\in \\mathcal{U}(\\mathbf{x})}1 / g(\\mathbf{x},\\mathbf{u}))$ -Lipschitz w.r.t $\\ell_2$ norm.", + "bbox": [ + 511, + 137, + 906, + 214 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Projection onto sublevel sets. For any threat model $(d,\\varepsilon)$ , the ability to efficiently project a perturbation $\\delta$ onto the sub-level sets $S(d,\\mathbf{x},\\varepsilon)$ enables one to leverage a wide literature on gradient-based attacks to conduct a rigorous evaluations of adversarial robustness. The sublevel sets of PD threat model, i.e., $S(\\mathbf{x},d,\\varepsilon)$ , are convex for all $\\mathbf{x}$ for all $\\epsilon \\geq 0$ . In particular, $\\mathcal{S}$ is the intersection of at most $k\\cdot C$ half-spaces characterized by unsafe directions $\\mathcal{U}(\\mathbf{x})$ as $S(\\mathbf{x},d_{\\mathrm{PD}},\\varepsilon) = \\bigcap_{\\mathbf{u}\\in \\mathcal{U}(\\mathbf{x})}\\{\\pmb {\\delta}\\in \\mathbb{R}^d\\mid \\langle \\pmb {\\delta},\\mathbf{u}\\rangle \\leq \\varepsilon \\cdot g(\\mathbf{x},\\mathbf{u})\\}$ .", + "bbox": [ + 511, + 214, + 908, + 358 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Further, the sub-level sets are monotonic w.r.t. the threshold, i.e., for $\\varepsilon_1\\geq \\varepsilon_2\\geq 0,S(\\mathbf{x},d_{\\mathrm{PD}},\\varepsilon_2)\\subseteq S(\\mathbf{x},d_{\\mathrm{PD}},\\varepsilon_1)$", + "bbox": [ + 511, + 357, + 906, + 387 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Linearity of growth and convexity of the sub-level sets together provide a straightforward approximate projection algorithm, since for any perturbation $\\delta$ , the scaled perturbation $\\frac{\\varepsilon}{d_{\\mathrm{PD}}(\\mathbf{x},\\delta)}\\delta$ lies in the permissible set $S(\\mathbf{x},d_{\\mathrm{PD}},\\varepsilon)$ . Further, due to the convexity of sublevel sets, an exact projection (in $\\ell_2$ distance) can be computed using the iterative greedy scheme outlined in the Algorithm 2 (deferred to the appendix). Thus the PD threat provides the practitioner with the flexibility of a fast approximate projection or an exact projection at the expense of increased computation. Together, these projection algorithms enable immediate adaptation of existing adversarial attacks to our proposed threat model.", + "bbox": [ + 511, + 388, + 908, + 568 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "We note that efficient exact projection algorithms are also known for $\\ell_p$ threat models. However, for neural perceptual threat models such as DreamSim [15], the sub-level sets are non-convex and a projection can only be approximately computed [32] and is typically not computationally efficient.", + "bbox": [ + 511, + 569, + 908, + 646 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "4. Experiments on Real World Data", + "text_level": 1, + "bbox": [ + 511, + 661, + 816, + 679 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "In this section we illustrate the qualitative and quantitative characteristics of the PD-threat in comparison to existing threat models. We evaluate 4 threat functions: $d_{\\infty}, d_{2}, d_{\\mathrm{DS}}$ , and our proposed PD threat function $d_{\\mathrm{PD}}$ , where $d_{\\mathrm{DS}}$ is the neural perceptual threat function $d_{\\mathrm{DS}}(\\mathbf{x}, \\tilde{\\mathbf{x}} - \\mathbf{x}) \\coloneqq \\mathrm{DreamSim}(\\mathbf{x}, \\tilde{\\mathbf{x}})$ from DreamSim [15]. Note that we focus on DreamSim due to its superior performance over other neural perceptual threat models [15], but the following comparison applies to any modern neural perceptual threat model.", + "bbox": [ + 511, + 688, + 908, + 824 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Experimental Setup. We seek to measure the ability of threat functions to distinguish between safe and unsafe corruptions. The threat models $d_{\\mathrm{PD}}$ and $d_{\\mathrm{DS}}$ depend on the Imagenet-1k training dataset S. All threat models are evaluated on the Imagenet-1k validation dataset $\\mathbf{S}_{\\mathrm{val}}$ . For any pair", + "bbox": [ + 511, + 825, + 908, + 902 + ], + "page_idx": 4 + }, + { + "type": "page_number", + "text": "9958", + "bbox": [ + 480, + 946, + 514, + 957 + ], + "page_idx": 4 + }, + { + "type": "table", + "img_path": "images/20affae325cbe6b4e64be8104246c37515168946fc039dfcf72fd6b064f3cccb.jpg", + "table_caption": [], + "table_footnote": [], + "table_body": "
CATEGORYCORRUPTION STYLES
NoiseGaussian, Shot, Impulse, Speckle, Blue, Brown, Perlin, Single-Frequency, Plasma
BlurGaussian, Defocus, Glass, Motion, Concentric Sine Waves, Caustic Refraction
CompressionPixelate, JPEG
DigitalBrightness, Contrast, Saturate, Elastic Transform
WeatherFrost, Fog, Snow
OcclusionSpatter, Checkerboard Cutout, Sparkles, Inverse Sparkles
", + "bbox": [ + 91, + 88, + 483, + 271 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Table 2. Categories of corruptions studied, from Imagenet-C [25] and Imagenet- $\\tilde{\\mathbf{C}}$ [36].", + "bbox": [ + 89, + 281, + 483, + 310 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "of labelled inputs $(\\mathbf{x},y),(\\tilde{\\mathbf{x}},c)\\in S_{\\mathrm{val}}$ with distinct labels $(c\\neq y)$ , the perturbation $\\delta \\coloneqq \\tilde{\\mathbf{x}} -\\mathbf{x}$ is an unsafe perturbation at $\\mathbf{x}$ . This way, we can evaluate the statistics of threat under unsafe perturbations for each threat function. To characterize threat under safe perturbations, we consider a set $\\Omega$ of 150 distinct common corruptions consisting of 30 styles $\\times 5$ severity levels, i.e., severity $\\in \\{1,2,3,4,5\\}$ - the supplementary material contains visual examples (see Figures 8 and 9), sourced from Imagenet- $C$ [25] and Imagenet- $\\bar{\\mathbf{C}}$ [36] accompanied with their threat assessment.", + "bbox": [ + 89, + 335, + 483, + 487 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "For a specific corruption style such as Gaussian noise and a severity level 5, we let $\\Omega_{\\mathrm{GaussianNoise},5}$ denote the corresponding corruption function. To aid visualization, we group the corruption styles into 6 distinct categories (listed in Table 2) so that $\\Omega := \\cup_{i \\in [5]} \\Omega_i$ where $\\Omega_i$ is the set of all corruption styles at severity level $i$ ,", + "bbox": [ + 89, + 488, + 483, + 580 + ], + "page_idx": 5 + }, + { + "type": "equation", + "text": "\n$$\n\\Omega_ {i} := \\left(\\Omega_ {\\text {n o i s e}, i} \\cup \\Omega_ {\\text {b l u r}, i} \\cup \\Omega_ {\\text {c o m p r e s s i o n}, i} \\cup \\Omega_ {\\text {d i g i t a l}, i} \\cup \\Omega_ {\\text {w e a t h e r}, i} \\cup \\Omega_ {\\text {o c c l u s i o n}, i}\\right).\n$$\n", + "text_format": "latex", + "bbox": [ + 89, + 589, + 491, + 616 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Each individual corruption $\\omega \\in \\Omega$ is applied to a subset $\\bar{\\mathbf{S}}\\subset \\mathbf{S}_{\\mathrm{val}}$ of 5,000 images chosen uniformly at random. For each threat function $d\\in \\{d_{\\infty},d_2,d_{\\mathrm{PD}},d_{\\mathrm{DS}}\\}$ and corruption $\\omega \\in \\Omega$ , for each image $\\mathbf{x}\\in \\bar{\\mathbf{S}}$ , we denote by $\\mathrm{avg}(d,\\omega)$ , the average threat statistic, $\\mathrm{avg}(d,\\omega)\\coloneqq \\frac{1}{|\\bar{s}|}\\sum_{(\\mathbf{x},y)\\in S}d(\\mathbf{x},\\omega (\\mathbf{x}) - \\mathbf{x})$ , where $\\omega (\\mathbf{x})$ is the corrupted input and $\\omega (\\mathbf{x}) - \\mathbf{x}$ is the safe perturbation at $\\mathbf{x}$ .", + "bbox": [ + 89, + 625, + 486, + 733 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "4.1. Disentangling Safe and Unsafe Corruption", + "text_level": 1, + "bbox": [ + 89, + 742, + 454, + 758 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "We now measure the ability of threat models to distinguish safe and unsafe corruptions. Each corruption $\\omega \\in \\Omega$ defines a point in Figure 3a having as the horizontal coordinate $\\mathrm{avg}(d_{\\infty},\\omega)$ and $\\mathrm{avg}(d_{\\mathrm{PD}},\\omega)$ as the vertical coordinate, producing a comparison of our PD threat to the $\\ell_{\\infty}$ -threat model. Similarly, Figure 3b compares PD threat to the DreamSIM-threat model.", + "bbox": [ + 89, + 763, + 483, + 868 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "A specification of threat model $(d, \\varepsilon)$ corresponds to the choice of a horizontal line for $d_{\\mathrm{PD}}$ or a vertical line for $(d_{\\infty}$", + "bbox": [ + 89, + 869, + 483, + 902 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "or $d_{\\mathrm{DS}}$ ) in Figure 3. We note that a size-able set of common corruptions incur large $d_{\\infty}$ threat, in particular as large as the unsafe corruptions (as indicated in Figure 1), thus any vertical line in Figure 3a is bound to exclude such common corruptions. In contrast, both $d_{\\mathrm{PD}}$ and $d_{\\mathrm{DS}}$ admit choices of $\\varepsilon$ that exclude unsafe perturbations while including most common corruptions. Hence, PD threat and DS threat are able to disentangle safe and unsafe corruptions.", + "bbox": [ + 511, + 90, + 903, + 210 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "For each of the comparisons in Figure 3, we identify 4 quadrants I-IV, distinguishing between low-threat and high-threat corruptions as measured by the corresponding threat models. We define the quadrants by choosing thresholds of 1, 0.5, and 0.25 for $d_{\\mathrm{PD}}$ , $d_{\\infty}$ , and $d_{\\mathrm{DS}}$ , respectively9, and comment on the qualitative behavior observed:", + "bbox": [ + 511, + 213, + 906, + 304 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "- Quadrant IV contains corruptions that are characterized as low-threat by both threat models. As expected, almost all corruptions having a low severity (1) lie in this region. However, we already see deficiencies of the $d_{\\infty}$ threat, as some blur, noise, and occlusion corruptions of severity 1 in Figure 3a are not characterized as low-threat by at least one of the threat models.", + "bbox": [ + 513, + 305, + 906, + 411 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "- Quadrant II contains corruptions that are characterized as high-threat by both threat models. These include digital and weather corruptions of high severity (5) as they tend to remove most of the signal in the input.", + "bbox": [ + 513, + 411, + 906, + 470 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "- Quadrant I contains corruptions having low PD-threat but high $d_{\\infty}$ or $d_{\\mathrm{DS}}$ and showcases the resilience of our PD-threat model to natural corruptions like blur and Gaussian noise, which retain a lot of the information in the image relevant to the classification task. Indeed, a visual inspection of these corruptions (see Section 10 in the Appendix) reveals that even at a high severity, a human is able to discern task-relevant information from the corrupted images, and hence these corruptions should be assigned a low threat.", + "bbox": [ + 513, + 472, + 908, + 619 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "- Quadrant III contains corruptions with high PD threat but low $d_{\\infty}$ or $d_{\\mathrm{DS}}$ threat. We note that corruptions rate as high threat by $d_{\\mathrm{PD}}$ are largely also rated highly by $d_{\\mathrm{DS}}$ . Since the PD and DS threats are aligned, the corruptions in Quadrant III in Figure 3a indicate that low $d_{\\infty}$ is not sufficient to characterize hard corruptions.", + "bbox": [ + 513, + 622, + 908, + 714 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "As a final point, we note that while the size of markers in Figure 3 indicate the severity levels, different corruptions of similar severity levels are qualitatively different. For e.g, weather corruptions of severity level 5 are uniformly harder than noise corruptions such as shot noise, Gaussian noise. Figure 11 illustrates heatmap of average threat for each corruption category across increasing severity. We observe that both PD and DS threat exhibit largely monotonic", + "bbox": [ + 511, + 715, + 908, + 835 + ], + "page_idx": 5 + }, + { + "type": "page_footnote", + "text": "9These thresholds were chosen to approximately be half of the average threat of unsafe perturbations that alter the true label. Admittedly, this binary threshold is coarse, and the Appendix contains fine-grained qualitative examples of corruptions of varying threat.", + "bbox": [ + 511, + 849, + 906, + 902 + ], + "page_idx": 5 + }, + { + "type": "page_number", + "text": "9959", + "bbox": [ + 482, + 946, + 514, + 957 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "increase across severity unlike $d_{\\infty}$ . At this point, we will pause to summarize the benefits of PD-threat.", + "bbox": [ + 89, + 90, + 485, + 121 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/920cca1a7bf6f7190f4a454ce4244bf8e2d3d453e4edd1c6663d204ac5f4d784.jpg", + "image_caption": [ + "(a)" + ], + "image_footnote": [], + "bbox": [ + 127, + 135, + 444, + 386 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/c1b23194f51f9e38ce127382832edc10c7be4637cd15b6dcd3e547ebe6e6624c.jpg", + "image_caption": [ + "(b)" + ], + "image_footnote": [], + "bbox": [ + 127, + 400, + 444, + 630 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "PD vs DreamSim. The PD-threat compares well to the DS-threat but does not require pre-trained large vision models or fine-tuning on a curated dataset of human perceptual similarity judgements. For real-world classification tasks where perceptual similarity between pairs of input is harder to quantify even with expert human annotators, instantiation of high-quality neural perceptual distance metrics is harder. Further, unlike perceptual distance metrics [15, 17, 18, 32, 54], the growth of the PD-threat is linear and interpretable along any direction, the sub-level sets induced by the PD-threat model are convex, and admit efficient projection algorithms and hence PD threat presents a promising alternative.", + "bbox": [ + 89, + 704, + 485, + 898 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "5. Measuring threat with Task Annotation", + "text_level": 1, + "bbox": [ + 511, + 90, + 870, + 107 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "We described two practical design choices for PD-threat in Section 3.2: (1) a choice of the set of unsafe directions $\\mathcal{U}(\\mathbf{x})$ at each input, and (2) a choice of normalization $g(\\mathbf{x},\\mathbf{u})$ for each unsafe direction. Next, we incorporate additional task-relevant information to refine PD-threat further.", + "bbox": [ + 511, + 114, + 906, + 189 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "5.1. Segmentation-aware Threat Specification", + "text_level": 1, + "bbox": [ + 511, + 199, + 864, + 215 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Segmentation masks isolate regions of the input that contain semantic information relevant to the class label, and can be generated efficiently following recent advances in foundational models for image segmentation [30, 38]. We demonstrate a straightforward method to incorporate semantic masks in the assessment of threat.", + "bbox": [ + 511, + 222, + 906, + 311 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Definition 8 (PD-S threat). Let $\\mathbf{x} \\in \\mathcal{X}$ and let $\\delta \\in \\mathbb{R}^d$ . Let $\\mathbf{a} \\in \\{0,1\\}^d$ be a boolean mask. The segmented PD threat,", + "bbox": [ + 511, + 320, + 906, + 352 + ], + "page_idx": 6 + }, + { + "type": "equation", + "text": "\n$$\nd _ {\\mathrm {P D - S}} (\\mathbf {x}, \\mathbf {a}, \\boldsymbol {\\delta}) := \\max _ {\\mathbf {u} \\in \\mathcal {U} (\\mathbf {x})} \\frac {1}{g _ {\\boldsymbol {\\beta}} (\\mathbf {x} , \\mathbf {u})} \\max \\left(\\langle \\boldsymbol {\\delta} [ \\mathbf {a} ], \\mathbf {u} [ \\mathbf {a} ] \\rangle , 0\\right).\n$$\n", + "text_format": "latex", + "bbox": [ + 511, + 373, + 913, + 406 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Here the $\\mathbf{u}[\\mathbf{a}]$ denotes a sub-vector whose indices are selected by the boolean mask $\\mathbf{a}$ . The threat model $(d_{\\mathrm{PD-S}}, \\varepsilon)$ is aligned with Xiao et al. [51]'s proposal to learn models robust to adversarially chosen backgrounds. For each image, we generate automatic masks by prompting SAM10 with the center pixel coordinates. Figure 4 illustrates the difference between PD and PD-S threat on 2 corruptions that are background-only and foreground-only respectively. PD-S threat is oblivious to background corruption but more sensitive to foreground corruption.", + "bbox": [ + 511, + 415, + 908, + 566 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/5462da7c42ae9bf65700ad81708555cfdd3bd5f026366f2bb6f22c619c2eb01e.jpg", + "image_caption": [ + "Figure 3. Comparison of Threat models w.r.t. various common corruptions. The size of a marker represents the severity level.", + "Original Image", + "Figure 4. Comparison of threat on background vs foreground." + ], + "image_footnote": [], + "bbox": [ + 531, + 579, + 617, + 647 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/24e99c7989721abcb7274ffba9b6228b921d7bcd540f9bead52f9a64c2bb8d3d.jpg", + "image_caption": [ + "$\\mathrm{PD} = 1.4$", + "PD-S=0.0" + ], + "image_footnote": [], + "bbox": [ + 666, + 579, + 751, + 646 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/c664177ece929bf1a2a5d80009090d35a775d5cd81f42704bbb36807eab55069.jpg", + "image_caption": [ + "$\\mathrm{PD} = 0.68$", + "PD-S=1.78" + ], + "image_footnote": [], + "bbox": [ + 800, + 579, + 887, + 646 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "We note that $\\ell_p$ threat models are also capable of integrating segmentation masks as they are pixel-based, but it is unclear if perceptual threat can account for pixel annotation.", + "bbox": [ + 511, + 710, + 906, + 756 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "5.2. Concept-aware Threat Specification", + "text_level": 1, + "bbox": [ + 511, + 765, + 825, + 782 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "The PD-threat assumes no hierarchy of class labels; i.e. all labels are distinct to the same degree. However, classification tasks are often accompanied with an additional hierarchy that identifies groups of class labels based on similar semantic content. We propose to account for such a hierarchy by refining our choice of normalization. Suppose $\\mathbf{x}$ has label", + "bbox": [ + 511, + 787, + 906, + 878 + ], + "page_idx": 6 + }, + { + "type": "page_footnote", + "text": "10In particular we use the default ViT-H SAM checkpoint.", + "bbox": [ + 527, + 886, + 834, + 900 + ], + "page_idx": 6 + }, + { + "type": "page_number", + "text": "9960", + "bbox": [ + 482, + 946, + 514, + 957 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "$y$ and $\\mathbf{u} = \\frac{\\tilde{\\mathbf{x}} - \\mathbf{x}}{\\|\\tilde{\\mathbf{x}} - \\mathbf{x}\\|_2}$ where $\\tilde{\\mathbf{x}}\\in S_{c,k}$ has label c. Let $W:[C]\\times [C]\\in [0,1]$ denote a relative distance between class labels based on Wordnet-hierarchy (refer to Section 9.5 for explicit details). We propose the weighted normalization $g_{\\beta ,W}(\\mathbf{x},\\mathbf{u})\\coloneqq \\boldsymbol {\\beta}\\cdot W(y,c)\\| \\tilde{\\mathbf{x}} -\\mathbf{x}\\| _2$ so that perturbation between nearby classes are weighted higher in threat.", + "bbox": [ + 89, + 90, + 483, + 185 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Definition 9 (PD-W threat). Let $\\mathbf{x} \\in \\mathcal{X}$ with label $y$ and let $\\pmb{\\delta} \\in \\mathbb{R}^{d}$ . The weighted PD threat is defined as,", + "bbox": [ + 89, + 191, + 483, + 222 + ], + "page_idx": 7 + }, + { + "type": "equation", + "text": "\n$$\nd_{\\mathrm{PD - W}}(\\mathbf{x},\\boldsymbol {\\delta}):= \\max_{\\mathbf{u}\\in \\mathcal{U}(\\mathbf{x})}\\frac{1}{g_{\\beta,W}(\\mathbf{x},\\mathbf{u})}\\max \\left(\\langle \\boldsymbol {\\delta},\\mathbf{u}\\rangle ,0\\right).\n$$\n", + "text_format": "latex", + "bbox": [ + 99, + 227, + 473, + 260 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "The threat model $(d_{\\mathrm{PD - W}},\\varepsilon)$ is weaker than $(d_{\\mathrm{PD}},\\varepsilon)$ , since $S(\\mathbf{x},d_{\\mathrm{PD - W}},\\varepsilon)\\subseteq S(\\mathbf{x},d_{\\mathrm{PD}},\\varepsilon)$ . Thus PD-W is a relaxation of PD threat with a softer requirement of stability between nearby classes. Figure 5 depicts the average relative $^{12}$ threat vs relative distance $W(\\cdot ,\\cdot)$ of class labels on the Imagenet-1k validation dataset. The decreasing PD-W threat for increasing class distance is aligned with the following intuition - failure to distinguish between semantically distant classes such as ENGLISH FOXHOUND and FIRE TRUCK should incur lower robust accuracy than failure to distinguish between similar classes such as ENGLISH FOXHOUND and IRISH WOLFHOUND.", + "bbox": [ + 89, + 266, + 486, + 446 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/37f4b90c34d11b1ef600464461f5cc220b8ed0f11c89285aa67e83458ed08837.jpg", + "image_caption": [ + "Figure 5. Relative Threat vs relative inter-class distances" + ], + "image_footnote": [], + "bbox": [ + 148, + 467, + 400, + 621 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "In summary our novel threat specification framework enables evaluation of robustness to adversarial perturbations and common corruptions while enabling sensitivity to semantic regions of an image and class hierarchy by incorporating additional task annotation. As a consequence, our threat model unifies adversarial robustness with variants of corruption robustness suggested by several independent benchmarks [25, 36, 51].", + "bbox": [ + 89, + 667, + 485, + 787 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "6. Evaluation of benchmark $\\ell_p$ robust models", + "text_level": 1, + "bbox": [ + 89, + 800, + 472, + 818 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Finally, we evaluate state-of-the-art robust models against our proposed thread models. Note that, as per Theorem 1,", + "bbox": [ + 89, + 825, + 483, + 856 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "one should hope for classifiers that are 1-robust to corruptions under the PD-threat. The corresponding permissible set of corruptions cover regions IV and I in Figure 3. We evaluate robustness to corruptions in the set $S(\\mathbf{x}, d_{\\mathrm{PD}}, \\varepsilon) \\cap S(\\mathbf{x}, d_{\\infty}, \\varepsilon)$ . We denote the corresponding threat model $(d_{\\infty} \\bigcap d_{\\mathrm{PD}}, \\varepsilon)$ . A robust evaluation for such a threat model can be executed in practice by generating adversarial perturbations with standard AutoAttack [7] for $(d_{\\infty}, \\varepsilon)$ followed by projection onto $S(\\mathbf{x}, d_{\\mathrm{PD}}, \\varepsilon)$ . Table 3 shows the robustness of benchmark classifiers registered at RobustBench order by $d_{\\infty}$ robustness (column 1) when evaluated against $(d_{\\infty} \\bigcap d_{\\mathrm{PD}}, 16/255)$ threat model for the Imagenet-1k classification task.", + "bbox": [ + 511, + 90, + 906, + 287 + ], + "page_idx": 7 + }, + { + "type": "table", + "img_path": "images/b53163e26126063e2fe08fef8735ec813786c135dde85d6f95aa49a37b4d333b.jpg", + "table_caption": [], + "table_footnote": [], + "table_body": "
Benchmark Modeld∞PDPD-SPD-W
ConvNeXt-L [33]0.250.320.310.31
Swin-B [33]0.120.210.200.21
ConvNeXt-B [33]0.10.190.180.18
ConvNeXt-B-ConvStem [41]0.090.190.190.19
ViT-S-ConvStem [41]0.050.130.120.12
XCiT-L12 [10]0.040.170.180.18
", + "bbox": [ + 521, + 299, + 898, + 422 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Table 3. Robust Evaluation for $\\varepsilon = {16}/{255}$ .", + "bbox": [ + 575, + 431, + 841, + 446 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "We note that robust accuracies uniformly improve under our threat specification in comparison to $(d_{\\infty}, \\varepsilon)$ , indicating a meaningful restriction of unsafe corruptions. The ordering of robust accuracies highlights diversity, e.g. [10]'s XCiT-L12 fares better under PD threat than suggested by the $d_{\\infty}$ leaderboard. The indistinguishability across PD, PD-S and PD-w is surprising. Similar levels of robustness w.r.t PD-W and PD indicate that models are unable to distinguish distant classes, while similar levels under PD and PD-S indicate adversarial attacks are often aligned with foreground masks. We note that adversarial attacks adapted to the threat specification often lead to a more accurate pessimistic evaluation of robustness.", + "bbox": [ + 511, + 463, + 908, + 659 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "7. Conclusion", + "text_level": 1, + "bbox": [ + 513, + 672, + 633, + 689 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "This work proposes a novel task-dependent threat specification Projected Displacement threat that is adapted to the shape of decision boundaries based on observed training data. Unlike standard $\\ell_p$ -threat models, the proposed threat model exhibits anisotropy and locality and is thus able to distinguish between safe and unsafe perturbations. The proposed threat specification framework is flexible and can effectively account for additional task information, such as image regions or label hierarchies. We invite the community to suggest further adaptive attacks and effective training strategies to encourage robustness in PD threat.", + "bbox": [ + 511, + 699, + 908, + 864 + ], + "page_idx": 7 + }, + { + "type": "page_footnote", + "text": "11 Relative since we require a normalized value in [0,1]", + "bbox": [ + 104, + 862, + 393, + 875 + ], + "page_idx": 7 + }, + { + "type": "page_footnote", + "text": "$^{12}$ Due to different scaling of PD-W and PD, we visualize average threat in each relative to maximum value on data.", + "bbox": [ + 91, + 875, + 482, + 898 + ], + "page_idx": 7 + }, + { + "type": "page_number", + "text": "9961", + "bbox": [ + 482, + 946, + 513, + 957 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Acknowledgements", + "text_level": 1, + "bbox": [ + 91, + 90, + 258, + 107 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "This research was supported by NSF CISE 2212457 and NSF CAREER Award CCF 2239787", + "bbox": [ + 89, + 113, + 485, + 142 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "References", + "text_level": 1, + "bbox": [ + 91, + 154, + 187, + 169 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[1] Naveed Akhtar, Ajmal Mian, Navid Kardan, and Mubarak Shah. Advances in adversarial attacks and defenses in computer vision: A survey. IEEE Access, 9:155161-155196, 2021. 1", + "[2] Pranjal Awasthi, Anqi Mao, Mehryar Mohri, and Yutao Zhong. Theoretically grounded loss functions and algorithms for adversarial robustness. In Proceedings of The 26th International Conference on Artificial Intelligence and Statistics, pages 10077-10094. PMLR, 2023. 3", + "[3] Yatong Bai, Brendon G. Anderson, Aerin Kim, and Somayeh Sojoudi. Improving the accuracy-robustness trade-off of classifiers via adaptive smoothing, 2024. 1", + "[4] Brian R. Bartoldson, James Diffenderfer, Konstantinos Parasyris, and Bhavya Kailkhura. Adversarial robustness limits via scaling-law and human-alignment studies. In Proceedings of the 41st International Conference on Machine Learning, pages 3046-3072. PMLR, 2024. 1, 2", + "[5] Yulong Cao, Chaowei Xiao, Benjamin Cyr, Yimeng Zhou, Wonseok Park, Sara Rampazzi, Qi Alfred Chen, Kevin Fu, and Z. Morley Mao. Adversarial sensor attack on lidar-based perception in autonomous driving. Proceedings of the 2019 ACM SIGSAC Conference on Computer and Communications Security, 2019. 1", + "[6] Nicholas Carlini and David A. Wagner. Adversarial examples are not easily detected: Bypassing ten detection methods. Proceedings of the 10th ACM Workshop on Artificial Intelligence and Security, 2017. 1", + "[7] Francesco Croce and Matthias Hein. Reliable evaluation of adversarial robustness with an ensemble of diverse parameter-free attacks. In Proceedings of the 37th International Conference on Machine Learning. JMLR.org, 2020. 1, 2, 8", + "[8] Francesco Croce, Maksym Andriushchenko, Vikash Sehwag, Edoardo Debenedetti, Nicolas Flammarion, Mung Chiang, Prateek Mittal, and Matthias Hein. Robustbench: a standardized adversarial robustness benchmark. In Thirty-fifth Conference on Neural Information Processing Systems Datasets and Benchmarks Track, 2021. 1, 2", + "[9] Jiequan Cui, Zhuotao Tian, Zhisheng Zhong, Xiaojuan Qi, Bei Yu, and Hanwang Zhang. Decoupled kullback-leibler divergence loss, 2023. 1", + "[10] Edoardo Debenedetti, Vikash Sehwag, and Prateek Mittal. A light recipe to train robust vision transformers, 2023. 8", + "[11] Edoardo Debenedetti, Zishen Wan, Maksym Andriushchenko, Vikash Sehwag, Kshitij Bhardwaj, and Bhavya Kailkhura. Scaling compute is not all you need for adversarial robustness. In ICLR 2024 Workshop on Reliable and Responsible Foundation Models, 2024. 1, 2", + "[12] Samuel Dodge and Lina Karam. A study and comparison of human and deep learning recognition performance under visual distortions. In 2017 26th international conference on computer communication and networks (ICCCN), pages 1-7. IEEE, 2017. 3" + ], + "bbox": [ + 93, + 179, + 485, + 898 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[13] Elvis Dohmatob. Generalized no free lunch theorem for adversarial robustness. In ICML, 2019. 4", + "[14] Ecenaz Erdemir, Jeffrey Bickford, Luca Melis, and Sergul Aydore. Adversarial robustness with non-uniform perturbations, 2021. 4", + "[15] Stephanie Fu, Netanel Tamir, Shobhita Sundaram, Lucy Chai, Richard Zhang, Tali Dekel, and Phillip Isola. Dreamsim: Learning new dimensions of human visual similarity using synthetic data. Advances in Neural Information Processing Systems, 36, 2024. 2, 3, 5, 7", + "[16] Robert Geirhos, Carlos R. M. Temme, Jonas Rauber, Heiko H. Schütt, Matthias Bethge, and Felix A. Wichmann. Generalisation in humans and deep neural networks. In Advances in Neural Information Processing Systems. Curran Associates, Inc., 2018. 3", + "[17] Sara Ghazanfari, Alexandre Araujo, Prashanth Krishnamurthy, Farshad Khorrami, and Siddharth Garg. Lipsim: A provably robust perceptual similarity metric. ArXiv, abs/2310.18274, 2023. 2, 7", + "[18] Sara Ghazanfari, Siddharth Garg, Prashanth Krishnamurthy, Farshad Khorrami, and Alexandre Araujo. R-LPIPS: An adversarially robust perceptual similarity metric. In The Second Workshop on New Frontiers in Adversarial Machine Learning, 2023. 7", + "[19] Abhijay Ghildyal and Feng Liu. Attacking perceptual similarity metrics. Trans. Mach. Learn. Res., 2023, 2023. 2", + "[20] Justin Gilmer, Ryan P. Adams, Ian Goodfellow, David Andersen, and George E. Dahl. Motivating the rules of the game for adversarial example research, 2018. 1", + "[21] Teofilo F. Gonzalez. Clustering to minimize the maximum intercluster distance. Theoretical Computer Science, 38:293-306, 1985. 11", + "[22] Ian J. Goodfellow, Jonathon Shlens, and Christian Szegedy. Explaining and harnessing adversarial examples. CoRR, abs/1412.6572, 2015. 1", + "[23] Sven Gowal, Chongli Qin, Jonathan Uesato, Timothy Mann, and Pushmeet Kohli. Uncovering the limits of adversarial training against norm-bounded adversarial examples, 2021. 1", + "[24] Sven Gowal, Sylvestre-Alvise Rebuffi, Olivia Wiles, Florian Stimberg, Dan Andrei Calian, and Timothy Mann. Improving robustness using generated data, 2021. 1", + "[25] Dan Hendrycks and Thomas Dietterich. Benchmarking neural network robustness to common corruptions and perturbations. Proceedings of the International Conference on Learning Representations, 2019. 3, 6, 8, 16", + "[26] Joern-Henrik Jacobsen, Jens Behrmann, Richard Zemel, and Matthias Bethge. Excessive invariance causes adversarial vulnerability. In International Conference on Learning Representations, 2019. 2", + "[27] Can Kanbak, Seyed-Mohsen Moosavi-Dezfooli, and Pascal Frossard. Geometric Robustness of Deep Networks: Analysis and Improvement. In 2018 IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 4441-4449, Los Alamitos, CA, USA, 2018. IEEE Computer Society. 3", + "[28] Oğuzhan Fatih Kar, Teresa Yeo, Andrei Atanov, and Amir Zamir. 3d common corruptions and data augmentation. In" + ], + "bbox": [ + 516, + 92, + 906, + 900 + ], + "page_idx": 8 + }, + { + "type": "page_number", + "text": "9962", + "bbox": [ + 482, + 946, + 514, + 957 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 18963-18974, 2022. 3", + "[29] Markus Kettunen, Erik Härkönen, and Jaakko Lehtinen. ELPips: Robust perceptual image similarity via random transformation ensembles, 2019. 2", + "[30] Alexander Kirillov, Eric Mintun, Nikhila Ravi, Hanzi Mao, Chloe Rolland, Laura Gustafson, Tete Xiao, Spencer Whitehead, Alexander C. Berg, Wan-Yen Lo, Piotr Dólar, and Ross Girshick. Segment anything. arXiv:2304.02643, 2023. 7", + "[31] Alexey Kurakin, Ian J. Goodfellow, and Samy Bengio. Adversarial examples in the physical world. ArXiv, abs/1607.02533, 2017. 1", + "[32] Cassidy Laidlaw, Sahil Singla, and Soheil Feizi. Perceptual adversarial robustness: Defense against unseen threat models, 2021. 2, 5, 7", + "[33] Chang Liu, Yinpeng Dong, Wenzhao Xiang, Xiao Yang, Hang Su, Jun Zhu, Yuefeng Chen, Yuan He, Hui Xue, and Shibao Zheng. A comprehensive study on robustness of image classification models: Benchmarking and rethinking, 2023. 1, 8, 13", + "[34] Aleksander Madry, Aleksandar Makelov, Ludwig Schmidt, Dimitris Tsipras, and Adrian Vladu. Towards deep learning models resistant to adversarial attacks. ArXiv, abs/1706.06083, 2018. 1", + "[35] Rafat K. Mantiuk, Kil Joong Kim, Allan G. Rempel, and Wolfgang Heidrich. Hdr-vdp-2: a calibrated visual metric for visibility and quality predictions in all luminance conditions. ACM SIGGRAPH 2011 papers, 2011. 3", + "[36] Eric Mintun, Alexander Kirillov, and Saining Xie. On interaction between augmentations and corruptions in natural corruption robustness. In Advances in Neural Information Processing Systems, pages 3571-3583. Curran Associates, Inc., 2021. 3, 6, 8, 16", + "[37] Yao Qin, Nicholas Carlini, Garrison Cottrell, Ian Goodfellow, and Colin Raffel. Imperceptible, robust, and targeted adversarial examples for automatic speech recognition. In International conference on machine learning, pages 5231-5240. PMLR, 2019. 1", + "[38] Nikhila Ravi, Valentin Gabeur, Yuan-Ting Hu, Ronghang Hu, Chaitanya Ryali, Tengyu Ma, Haitham Khedr, Roman Rädle, Chloe Rolland, Laura Gustafson, Eric Mintun, Junting Pan, Kalyan Vasudev Alwala, Nicolas Carion, Chao-Yuan Wu, Ross Girshick, Piotr Dólár, and Christoph Feichtenhofer. Sam 2: Segment anything in images and videos. arXiv preprint arXiv:2408.00714, 2024. 7", + "[39] Ayon Sen, Xiaojin Zhu, Erin Marshall, and Robert Nowak. Popular imperceptibility measures in visual adversarial attacks are far from human perception. In Decision and Game Theory for Security, pages 188-199, Cham, 2020. Springer International Publishing. 2", + "[40] Mahmood Sharif, Lujo Bauer, and Michael K. Reiter. On the suitability of lp-norms for creating and preventing adversarial examples. 2018 IEEE/CVF Conference on Computer Vision and Pattern Recognition Workshops (CVPRW), pages 1686-16868, 2018. 2", + "[41] Naman D Singh, Francesco Croce, and Matthias Hein. Revisiting adversarial training forImagenet: Architectures, training and generalization across threat models, 2023. 8" + ], + "bbox": [ + 91, + 92, + 485, + 900 + ], + "page_idx": 9 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[42] Christian Szegedy, Wojciech Zaremba, Ilya Sutskever, Joan Bruna, D. Erhan, Ian J. Goodfellow, and Rob Fergus. Intriguing properties of neural networks. CoRR, abs/1312.6199, 2014. 1", + "[43] Florian Tramér, Jens Behrmann, Nicholas Carlini, Nicolas Papernot, and Jörn-Henrik Jacobsen. Fundamental tradeoffs between invariance and sensitivity to adversarial perturbations. In Proceedings of the 37th International Conference on Machine Learning, ICML 2020, 13-18 July 2020, Virtual Event, pages 9561-9571. PMLR, 2020. 2, 3", + "[44] Florian Tramér, Nicholas Carlini, Wieland Brendel, and Aleksander Madry. On adaptive attacks to adversarial example defenses. ArXiv, abs/2002.08347, 2020. 1", + "[45] James Tu, Huichen Li, Xinchen Yan, Mengye Ren, Yun Chen, Ming Liang, Eilyan Bitar, Ersin Yumer, and Raquel Urtasun. Exploring adversarial robustness of multi-sensor perception systems in self driving. In Proceedings of the 5th Conference on Robot Learning, pages 1013-1024. PMLR, 2022. 1", + "[46] Apostol Vassilev, Alina Oprea, Alie Fordyce, and Hyrum Andersen. Adversarial machine learning: A taxonomy and terminology of attacks and mitigations, 2024. 1", + "[47] Zhou Wang, A.C. Bovik, H.R. Sheikh, and E.P. Simoncelli. Image quality assessment: from error visibility to structural similarity. IEEE Transactions on Image Processing, 13(4): 600-612, 2004. 3", + "[48] Zekai Wang, Tianyu Pang, Chao Du, Min Lin, Weiwei Liu, and Shuicheng Yan. Better diffusion models further improve adversarial training, 2023. 1", + "[49] Zuxuan Wu, Ser-Nam Lim, Larry Davis, and Tom Goldstein. Making an invisibility cloak: Real world adversarial attacks on object detectors, 2020. 1", + "[50] Chaowei Xiao, Jun-Yan Zhu, Bo Li, Warren He, Mingyan Liu, and Dawn Song. Spatially transformed adversarial examples. In International Conference on Learning Representations, 2018. 3", + "[51] Kai Yuanqing Xiao, Logan Engstrom, Andrew Ilyas, and Aleksander Madry. Noise or signal: The role of image backgrounds in object recognition. In International Conference on Learning Representations, 2021. 7, 8", + "[52] Hongyang Zhang, Yaodong Yu, Jiantao Jiao, Eric Xing, Laurent El Ghaoui, and Michael Jordan. Theoretically principled trade-off between robustness and accuracy. In Proceedings of the 36th International Conference on Machine Learning, pages 7472-7482. PMLR, 2019. 1", + "[53] Lin Zhang, Lei Zhang, Xuanqin Mou, and David Zhang. Fsim: A feature similarity index for image quality assessment. IEEE Transactions on Image Processing, 20(8):2378-2386, 2011. 3", + "[54] Richard Zhang, Phillip Isola, Alexei A. Efros, Eli Shechtman, and Oliver Wang. The unreasonable effectiveness of deep features as a perceptual metric. In 2018 IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 586-595, Los Alamitos, CA, USA, 2018. IEEE Computer Society. 3, 7", + "[55] Piotr Želasko, Sonal Joshi, Yiwen Shao, Jesus Villalba, Jan Trmal, Najim Dehak, and Sanjeev Khudanpur. Adversarial attacks and defenses for speech recognition systems, 2021." + ], + "bbox": [ + 516, + 92, + 906, + 897 + ], + "page_idx": 9 + }, + { + "type": "page_number", + "text": "9963", + "bbox": [ + 482, + 946, + 514, + 957 + ], + "page_idx": 9 + } +] \ No newline at end of file diff --git a/2025/Disentangling Safe and Unsafe Image Corruptions via Anisotropy and Locality/3a52d538-e6a0-40a8-904e-0b81bce10c6a_model.json b/2025/Disentangling Safe and Unsafe Image Corruptions via Anisotropy and Locality/3a52d538-e6a0-40a8-904e-0b81bce10c6a_model.json new file mode 100644 index 0000000000000000000000000000000000000000..59a998a7d43a39938bb0e5e0ba565e36fe45944b --- /dev/null +++ b/2025/Disentangling Safe and Unsafe Image Corruptions via Anisotropy and Locality/3a52d538-e6a0-40a8-904e-0b81bce10c6a_model.json @@ -0,0 +1,2607 @@ +[ + [ + { + "type": "header", + "bbox": [ + 0.107, + 0.003, + 0.182, + 0.043 + ], + "angle": 0, + "content": "CVF" + }, + { + "type": "header", + "bbox": [ + 0.237, + 0.001, + 0.812, + 0.047 + ], + "angle": 0, + "content": "This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore." + }, + { + "type": "title", + "bbox": [ + 0.102, + 0.13, + 0.898, + 0.155 + ], + "angle": 0, + "content": "Disentangling Safe and Unsafe Image Corruptions via Anisotropy and Locality" + }, + { + "type": "text", + "bbox": [ + 0.165, + 0.18, + 0.829, + 0.218 + ], + "angle": 0, + "content": "Ramchandran Muthukumar*¹ Ambar Pal²† Jeremias Sulam¹ René Vidal³ \n¹Johns Hopkins University ²Amazon Web Services ³University of Pennsylvania" + }, + { + "type": "title", + "bbox": [ + 0.248, + 0.252, + 0.327, + 0.267 + ], + "angle": 0, + "content": "Abstract" + }, + { + "type": "text", + "bbox": [ + 0.089, + 0.283, + 0.486, + 0.784 + ], + "angle": 0, + "content": "State-of-the-art machine learning systems are vulnerable to small perturbations to their input, where \"small\" is defined according to a threat model that assigns a positive threat to each perturbation. Most prior works define a task-agnostic, isotropic, and global threat, like the \\(\\ell_p\\) norm, where the magnitude of the perturbation fully determines the degree of the threat and neither the direction of the attack nor its position in space matter. However, common corruptions in computer vision, such as blur, compression, or occlusions, are not well captured by such threat models. This paper proposes a novel threat model called Projected Displacement (PD) to study robustness beyond existing isotropic and global threat models. The proposed threat model measures the threat of a perturbation via its alignment with unsafe directions, defined as directions in the input space along which a perturbation of sufficient magnitude changes the ground truth class label. Unsafe directions are identified locally for each input based on observed training data. In this way, the PD-threat model exhibits anisotropy and locality. Experiments on Imagenet-1k data indicate that, for any input, the set of perturbations with small PD threat includes safe perturbations of large \\(\\ell_p\\) norm that preserve the true label, such as noise, blur and compression, while simultaneously excluding unsafe perturbations that alter the true label. Unlike perceptual threat models based on embeddings of large-vision models, the PD-threat model can be readily computed for arbitrary classification tasks without pre-training or finetuning. Further additional task information such as sensitivity to image regions or concept hierarchies can be easily integrated into the assessment of threat and thus the PD threat model presents practitioners with a flexible, task-driven threat specification that alleviates the limitations of \\(\\ell_p\\)-threat models." + }, + { + "type": "title", + "bbox": [ + 0.092, + 0.81, + 0.222, + 0.826 + ], + "angle": 0, + "content": "1. Introduction" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.836, + 0.485, + 0.867 + ], + "angle": 0, + "content": "Modern machine learning (ML) systems are nearing widespread deployment in civilian life, and hence a com" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.253, + 0.908, + 0.509 + ], + "angle": 0, + "content": "prehensive understanding of their security vulnerabilities is necessary [46]. One such vulnerability concerns the ability of a malicious adversary to tamper with predictions by adding small imperceptible corruptions referred to as adversarial perturbations [22, 42]. There is by now overwhelming evidence that carefully crafted adversarial perturbations can foil the prediction of state-of-the-art machine learning classifiers, i.e. they are not adversarily robust [6, 8]. Such vulnerabilities have been observed in a wide variety of applications like computer vision [1, 49], speech recognition [37, 55], autonomous driving [5, 45], and more. The goal of designing safe and reliable ML systems remains incomplete [4, 11] despite significant investment [23, 24, 33]. Often, strategies that intend to foil performance, i.e. adversarial attacks [6, 7, 22, 31, 44], have proven more successful than strategies aimed at mitigating vulnerabilities, i.e. adversarial defenses [3, 9, 34, 48, 52]." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.511, + 0.909, + 0.751 + ], + "angle": 0, + "content": "Meaningfully evaluating the adversarial robustness of a machine learning system requires a formal specification of a threat function, and an associated threat model that limits the scope of malicious adversaries [20]. Informally, a threat function \\( d: \\mathbb{R}^d \\times \\mathbb{R}^d \\to \\mathbb{R}^{\\geq 0} \\) measures the threat represented by a corruption \\( \\delta \\in \\mathbb{R}^d \\) towards altering the true label at input \\( \\mathbf{x} \\in \\mathbb{R}^d \\) as \\( d(\\mathbf{x}, \\delta) \\). We let \\( S(\\mathbf{x}, d, \\varepsilon) \\coloneqq \\{\\delta \\in \\mathbb{R}^d \\mid d(\\mathbf{x}, \\delta) \\leq \\varepsilon\\} \\) denote the \\( \\varepsilon \\)-sublevel set of perturbations where the threat \\( d \\) is measured w.r.t. input \\( \\mathbf{x} \\). A threat model \\( (d, \\varepsilon) \\) is a pair of a threat function \\( d \\) and a permissible threshold \\( \\varepsilon \\) that together define the set of permissible perturbations \\( S(\\mathbf{x}, d, \\varepsilon) \\). Under the threat model \\( (d, \\varepsilon) \\), the robust accuracy of a classifier \\( h \\in \\mathcal{H} \\) is the probability that the label prediction at \\( \\mathbf{x} \\) is locally invariant to corruptions within the permissible set \\( S(\\mathbf{x}, d, \\varepsilon) \\), see Definition 2 for an explicit definition." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.753, + 0.909, + 0.845 + ], + "angle": 0, + "content": "One of the most commonly used threat models is the \\(\\ell_p\\)-threat model \\((d_p, \\epsilon)\\), corresponding to the choice \\(d_p(\\mathbf{x}, \\delta) \\coloneqq \\| \\delta \\|_p\\). The threat function \\(d_p\\) is task-agnostic, easy to evaluate, and induces a compact sub-level set \\(S(\\mathbf{x}, d_p, \\varepsilon)\\) that allows efficient projection. Hence, the \\(\\ell_p\\)-threat model \\((d_p, \\varepsilon)\\) presents a natural starting point for in" + }, + { + "type": "page_footnote", + "bbox": [ + 0.533, + 0.856, + 0.888, + 0.869 + ], + "angle": 0, + "content": "1The flexibility to define threat w.r.t. input \\(\\mathbf{x}\\) is critical to our work." + }, + { + "type": "page_footnote", + "bbox": [ + 0.516, + 0.869, + 0.905, + 0.904 + ], + "angle": 0, + "content": "2For any vector \\(\\pmb {\\delta}\\in \\mathbb{R}^d\\) , the \\(\\ell_p\\) norm \\(\\| \\cdot \\| _p\\) for \\(p\\geq 1\\) is defined as \\(\\| \\pmb {\\delta}\\| _p\\coloneqq \\left(\\sum_i|\\delta_i|^p\\right)^{\\frac{1}{p}}.\\)" + }, + { + "type": "list", + "bbox": [ + 0.516, + 0.856, + 0.905, + 0.904 + ], + "angle": 0, + "content": null + }, + { + "type": "page_footnote", + "bbox": [ + 0.109, + 0.875, + 0.401, + 0.888 + ], + "angle": 0, + "content": "* Corresponding author: rmuthukl@jhu.edu" + }, + { + "type": "page_footnote", + "bbox": [ + 0.111, + 0.888, + 0.392, + 0.901 + ], + "angle": 0, + "content": "This work is not related to AP's position at Amazon" + }, + { + "type": "list", + "bbox": [ + 0.109, + 0.875, + 0.401, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.947, + 0.516, + 0.958 + ], + "angle": 0, + "content": "9954" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.151, + 0.088, + 0.429, + 0.266 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.091, + 0.277, + 0.483, + 0.308 + ], + "angle": 0, + "content": "Figure 1. Corruptions with equal \\(\\ell_{\\infty}\\)-threat, \\(\\|\\delta_1\\|_{\\infty} = \\|\\delta_2\\|_{\\infty} = \\|\\delta_3\\|_{\\infty} = \\|\\delta_4\\|_{\\infty}\\), but varying PD-threat." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.335, + 0.484, + 0.471 + ], + "angle": 0, + "content": "vestigating robustness. RobustBench [8] maintains an up-to-date leaderboard of the robust accuracies of benchmark models under \\((d_p, \\varepsilon)\\). Unfortunately, the progress towards achieving perfect adversarial robustness (i.e. \\(100\\%\\) robust test accuracy) in RobustBench has plateaued in recent years. Unlike supervised learning on clean data, scaling data, model size and computing resources might be insufficient to bridge the gap [4, 11]. Below we expand further on the fundamental limitations of isotropic and global threat models." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.484, + 0.453, + 0.5 + ], + "angle": 0, + "content": "1.1. Motivation: Specification of Threat Model" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.507, + 0.484, + 0.824 + ], + "angle": 0, + "content": "For image-based datasets, it is widely recognized that \\(\\ell_p\\) norms are neither necessary nor sufficient for capturing perceptual similarity [39, 40], which poses significant challenges for accurately evaluating robustness [26, 43]. In Figure 1, \\(\\mathbf{x}\\) (at the center) is an image of class ROSE HIP from the Imagenet-1k dataset, and \\(\\mathbf{x} + \\delta_1\\), \\(\\mathbf{x} + \\delta_2\\) are two corrupted images equidistant (w.r.t. \\(\\ell_{\\infty}\\)) from \\(\\mathbf{x}\\), i.e. \\(\\| \\delta_1 \\|_{\\infty} = \\| \\delta_2 \\|_{\\infty} = \\varepsilon\\). Intuitively, both \\(\\mathbf{x} + \\delta_1\\) and \\(\\mathbf{x} + \\delta_2\\) share the same label as \\(\\mathbf{x}\\), since its appearance still depicts the fruit ROSE HIP. We refer to such perturbations that preserve the class label as safe, while those that change it are termed unsafe. Consider now moving further along these directions to \\(\\mathbf{x} + t\\delta_1\\) and \\(\\mathbf{x} + t\\delta_2\\), for \\(t > 1\\). These perturbed points are again equidistant from \\(\\mathbf{x}\\), but while \\(t\\delta_1\\) does not alter the true label (i.e. it is safe), \\(t\\delta_2\\) does (and it is unsafe). As a result, any threat model \\((d_{\\infty}, t' \\varepsilon)\\) with \\(t' \\geq t\\) necessarily incurs misspecification: any classifier \\(h\\) stable to all perturbations in \\(S(\\mathbf{x}, d_{\\infty}, t' \\varepsilon)\\) produces an incorrect prediction at \\(\\mathbf{x} + t\\delta_2\\), i.e. \\(h(\\mathbf{x} + t\\delta_2) = \\text{ROSE HIP}\\). On the other hand, shrinking the threat model is also futile, as then the safe perturbation \\(t\\delta_1\\) is necessarily excluded." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.825, + 0.484, + 0.902 + ], + "angle": 0, + "content": "This illustrates that standard \\(\\ell_p\\)-threat models are unable to distinguish between the safe perturbations that preserve the true label and the unsafe perturbations that alter the true label. Hence, perfect robust accuracy under \\(\\ell_p\\) threat models might be neither achievable nor desirable. Attempting to" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.092, + 0.908, + 0.168 + ], + "angle": 0, + "content": "resolve these limitations, Laidlaw et al. [32] formulate perceptual threat models via neural perceptual metric based on neural representations. Unfortunately (but unsurprisingly), all neural perceptual distance metrics are themselves vulnerable to adversarial attacks [17-19, 29]." + }, + { + "type": "title", + "bbox": [ + 0.514, + 0.18, + 0.757, + 0.195 + ], + "angle": 0, + "content": "1.2. Summary of Contributions" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.202, + 0.907, + 0.37 + ], + "angle": 0, + "content": "We propose a novel threat function called Projected Displacement (PD), denoted by \\( d_{\\mathrm{PD}} \\) (see Definition 7 for an explicit description), that accounts for the local directional statistics of the data via unsafe directions3 \\( \\mathcal{U}(\\mathbf{x}) \\) (Definition 6) at each input \\( \\mathbf{x} \\). For an input \\( \\mathbf{x} \\) with label \\( y \\), any input \\( \\tilde{\\mathbf{x}} \\) with label \\( \\tilde{y} \\neq y \\) represents the unsafe direction \\( \\mathbf{u} := \\frac{\\tilde{\\mathbf{x}} - \\mathbf{x}}{\\|\\tilde{\\mathbf{x}} - \\mathbf{x}\\|_2} \\), which will be estimated from observed training data. Then, the degree of threat of each perturbation \\( \\delta \\) is the maximal alignment with the set of unsafe directions: A large threat \\( d_{\\mathrm{PD}}(\\mathbf{x},\\delta) \\) indicates that \\( \\delta \\) is well-aligned with an unsafe direction \\( \\mathbf{u} \\in \\mathcal{U}(\\mathbf{x}) \\). We highlight some key properties." + }, + { + "type": "text", + "bbox": [ + 0.515, + 0.371, + 0.909, + 0.581 + ], + "angle": 0, + "content": "1. Disentangling safe vs unsafe perturbations. PD-threat is able to distinguish between safe and unsafe perturbations of equal \\(\\ell_p\\) norms. From Figure 1, different corruptions (Gaussian noise \\(t\\delta_1\\), motion blur \\(t\\delta_3\\) and saturation \\(t\\delta_4\\)) applied to the original image \\(\\mathbf{x}\\) all have equal \\(\\ell_{\\infty}\\) norm of 240/255. Despite this, the true label of these perturbed samples does not change, unlike that of corruption \\(t\\delta_2\\), which results in an image \\(\\mathbf{x} + t\\delta_2\\) with a different class. For each perturbation \\(t\\delta_i\\), the computed values of PD threat in Figure 1 naturally reflect how close they are to changing the class label. PD threat model is competent with state-of-the-art neural perceptual threat model DreamSim (denoted \\(d_{\\mathrm{DS}}\\)) [15] for distinguishing safe and unsafe perturbations (see Section 4)." + }, + { + "type": "text", + "bbox": [ + 0.515, + 0.582, + 0.909, + 0.702 + ], + "angle": 0, + "content": "2. Ease of use. As we will show, the sub-level set \\( S(\\mathbf{x}, d_{\\mathrm{PD}}, \\varepsilon) \\) at each input \\( \\mathbf{x} \\) is convex, allowing for efficient projections onto it. This enables a straightforward plug-and-play mechanism for adapting existing adversarial attacks to the PD-threat model. We incorporate such an extension of AutoAttack [7] to evaluate the robustness under the PD-threat model (\\( d_{\\mathrm{PD}}, \\varepsilon \\)) for state-of-the-art robust classifiers registered at RobustBench [8]." + }, + { + "type": "text", + "bbox": [ + 0.515, + 0.703, + 0.909, + 0.809 + ], + "angle": 0, + "content": "3. Task-dependency. PD-threat models can readily integrate additional task-relevant information via label annotation, such as inter-class hierarchy (e.g. the Wordnet hierarchy) or pixel annotation (e.g. segmentation masks)—see Table 1—allowing for variations in the form of robustness while retaining the above benefits (anisotropy, locality, convex sub-level sets, and efficient projection)." + }, + { + "type": "list", + "bbox": [ + 0.515, + 0.371, + 0.909, + 0.809 + ], + "angle": 0, + "content": null + }, + { + "type": "title", + "bbox": [ + 0.515, + 0.821, + 0.658, + 0.835 + ], + "angle": 0, + "content": "1.3. Related Work" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.844, + 0.908, + 0.874 + ], + "angle": 0, + "content": "In this section, we highlight prior work that explores robustness beyond \\(\\ell_p\\)-threat models." + }, + { + "type": "page_footnote", + "bbox": [ + 0.531, + 0.887, + 0.824, + 0.9 + ], + "angle": 0, + "content": "3Directions refer to vectors \\(\\mathbf{u}\\in \\mathbb{R}^d\\) with unit \\(\\ell_2\\) norm." + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.947, + 0.516, + 0.958 + ], + "angle": 0, + "content": "9955" + } + ], + [ + { + "type": "table", + "bbox": [ + 0.092, + 0.089, + 0.482, + 0.158 + ], + "angle": 0, + "content": "
PROPERTYFEATURESdpdDS[15]dPD
Degree of ThreatAnisotropy, Locality×✓✓
Label AnnotationInter-class Distance×~✓
Pixel AnnotationSegmentation✓×✓
" + }, + { + "type": "table_caption", + "bbox": [ + 0.169, + 0.168, + 0.407, + 0.182 + ], + "angle": 0, + "content": "Table 1. Comparison of Threat Models" + }, + { + "type": "text", + "bbox": [ + 0.089, + 0.2, + 0.486, + 0.426 + ], + "angle": 0, + "content": "Common Corruptions. Neural networks are vulnerable to image distortions (e.g. translation, rotation), common corruptions (e.g. blur, noise), changes in lighting, view, depth, etc. [12, 16, 25, 27, 28, 36, 50]. Note that these are known safe perturbations as they preserve the true label, even if they have large \\(\\ell_p\\) norms. To measure this vulnerability, [25] formalized the notion of corruption robustness and presented a standardized benchmark dataset, Imagenet-C containing 19 styles of common image corruptions (categorized into noise, blur, weather, and digital). Training on Imagenet-C, or on other augmentations, can result in over-fitting, particularly to specific types of corruptions [36]. Additionally, [36] introduced a newdataset, Imagenet-C, with 10 new corruption styles identified from a larger set of real-world corruptions. These will be useful later in our experiments." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.427, + 0.485, + 0.564 + ], + "angle": 0, + "content": "Perceptual Distance Metrics. Since \\(\\ell_p\\) provides inaccurate approximations to perceptual distances, several works have attempted to propose alternatives, such as SSIM [47], FSIM [53], and HDR-VDP-2 [35]. With the advent of learning data-driven representations, [15, 54] hail the effectiveness of internal representations of neural classifiers in capturing perceptual similarity. DreamSim [15], the current state-of-the-art among the neural perceptual distance metrics, is finetuned on human perceptual similarity judgements." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.577, + 0.245, + 0.593 + ], + "angle": 0, + "content": "2. Key Definitions" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.602, + 0.486, + 0.744 + ], + "angle": 0, + "content": "Notation. We denote scalar quantities in Roman letters and vectors in boldface Roman letters. The empty set is given by \\(\\emptyset\\). For any set \\(\\mathcal{V}\\), we denote by \\(2^{\\mathcal{V}}\\) the power set of all its subsets. For any set \\(\\mathcal{V} \\subset \\mathbb{R}^d\\) and a vector \\(\\mathbf{v} \\in \\mathbb{R}^d\\), \\(\\mathcal{V} - \\{\\mathbf{v}\\}\\) is the Minkowski difference, i.e. the set \\(\\{\\mathbf{w} - \\mathbf{v} \\mid \\forall \\mathbf{w} \\in \\mathcal{V}\\}\\) (subtracting \\(\\mathbf{v}\\) from all elements of \\(\\mathcal{V}\\)). We denote the unit ball and the unit sphere in \\(\\mathbb{R}^d\\) w.r.t. the \\(\\ell_p\\) norm by \\(\\mathbb{B}_p^d := \\{\\mathbf{v} \\in \\mathbb{R}^d \\mid \\| \\mathbf{v} \\|_p \\leq 1\\}\\) and \\(\\mathbb{S}_p^d := \\{\\mathbf{v} \\in \\mathbb{R}^d \\mid \\| \\mathbf{v} \\|_p = 1\\}\\), respectively. A direction is an element of the unit sphere \\(\\mathbb{S}_2^d\\)." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.752, + 0.289, + 0.768 + ], + "angle": 0, + "content": "2.1. Supervised Learning" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.774, + 0.485, + 0.866 + ], + "angle": 0, + "content": "We consider a supervised learning task with a bounded input domain \\(\\mathcal{X} \\subset \\mathbb{R}^d\\) and labels \\(\\mathcal{Y} = \\{1, \\dots, C\\}\\) such that each input \\(\\mathbf{x}\\) is assigned a label by a deterministic true \\(^4\\) labeling function \\(h^\\star : \\mathcal{X} \\to \\mathcal{Y}\\). The true labeling function \\(h^\\star\\) partitions the input domain \\(\\mathcal{X} := \\cup_y \\mathcal{X}_y\\), where \\(\\mathcal{X}_y\\) is the set of inputs assigned label \\(y\\), i.e. \\(\\mathcal{X}_y = \\{\\mathbf{x} \\in \\mathcal{X} \\mid h^\\star(\\mathbf{x}) = y\\}\\)." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.092, + 0.907, + 0.121 + ], + "angle": 0, + "content": "We assume that the partition sets \\(\\mathcal{X}_1, \\ldots, \\mathcal{X}_C\\) are nonempty and open in order to avoid degeneracies." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.121, + 0.909, + 0.278 + ], + "angle": 0, + "content": "We let \\(\\mathcal{D}_{\\mathcal{X}}\\) be a data distribution over the input domain \\(\\mathcal{X}\\) and \\(\\mathcal{D}_{\\mathcal{Z}}\\) the extension of \\(\\mathcal{D}_{\\mathcal{X}}\\) to a distribution over the joint domain \\(\\mathcal{Z} \\coloneqq \\mathcal{X} \\times \\mathcal{Y}\\) using the true labeling function \\(h^{\\star}\\). The true labeling function \\(h^{\\star}\\) and the marginal input distribution \\(\\mathcal{D}_{\\mathcal{X}}\\) are unknown, but we observe finite labeled training data \\(\\mathrm{S} \\coloneqq \\left\\{(\\mathbf{x}^{(1)}, y^{(1)}), \\ldots, (\\mathbf{x}^{(m)}, y^{(m)})\\right\\} \\stackrel{\\mathrm{i.i.d.}}{\\sim} (\\mathcal{D}_{\\mathcal{Z}})^{m}\\). For any classifier \\(h: \\mathcal{X} \\to \\mathcal{Y}\\), accuracy is the probability of correctly labeling a random input, i.e. \\(\\mathrm{Acc}(h) \\coloneqq \\mathrm{Prob}_{\\mathbf{x} \\sim \\mathcal{D}_{\\mathcal{X}}} [h(\\mathbf{x}) = h^{\\star}(\\mathbf{x})]\\). The task of supervised learning seeks a classifier \\(\\bar{h}\\) in a hypothesis class \\(\\mathcal{H} \\subset \\{h: \\mathcal{X} \\to \\mathcal{Y}\\}\\) with high accuracy." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.287, + 0.753, + 0.301 + ], + "angle": 0, + "content": "2.2. Adversarial Perturbations" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.308, + 0.906, + 0.34 + ], + "angle": 0, + "content": "In this article, we formally define adversarial perturbations through the lens of the partition sets \\(\\mathcal{X}_y\\)." + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.348, + 0.906, + 0.379 + ], + "angle": 0, + "content": "Definition 1 (Adversarial Perturbation). A perturbation \\(\\delta \\in \\mathbb{R}^d\\) is adversarial for predictor \\(h\\) at input \\(\\mathbf{x}\\) if it is," + }, + { + "type": "text", + "bbox": [ + 0.515, + 0.379, + 0.888, + 0.394 + ], + "angle": 0, + "content": "1. Domain constrained: \\(\\mathbf{x} + \\delta \\in \\mathcal{X}\\), i.e. \\(\\delta \\in \\mathcal{X} - \\{\\mathbf{x}\\}\\)." + }, + { + "type": "text", + "bbox": [ + 0.514, + 0.395, + 0.907, + 0.424 + ], + "angle": 0, + "content": "2. Label invariant: \\(\\mathbf{x} + \\boldsymbol{\\delta}\\) has the same true label as \\(\\mathbf{x}\\), i.e. \\(h^{\\star}(\\mathbf{x} + \\boldsymbol{\\delta}) = h^{\\star}(\\mathbf{x})\\)." + }, + { + "type": "text", + "bbox": [ + 0.514, + 0.424, + 0.906, + 0.455 + ], + "angle": 0, + "content": "3. Adversarial: \\(\\mathbf{x} + \\boldsymbol{\\delta}\\) is misclassified by \\(h\\), i.e. \\(h(\\mathbf{x} + \\boldsymbol{\\delta}) \\neq h^{\\star}(\\mathbf{x} + \\boldsymbol{\\delta})\\)." + }, + { + "type": "list", + "bbox": [ + 0.514, + 0.379, + 0.907, + 0.455 + ], + "angle": 0, + "content": null + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.465, + 0.909, + 0.631 + ], + "angle": 0, + "content": "The labels assigned by \\( h^{\\star} \\) are considered the ground truth and thus \\( h^{\\star} \\) has no adversarial perturbation at any input. The presence of adversarial perturbations indicates an imperfect labeling function \\( h \\). Based on Definition 1, perturbations \\( \\delta \\in \\mathbb{R}^{d} \\) such that \\( \\mathbf{x} + \\delta \\notin \\mathcal{X} \\) are not deemed adversarial as the classifier \\( h \\) is not required to predict labels outside the input domain. On the other hand, for \\( \\mathbf{x} \\) with true label \\( y \\), a perturbation \\( \\delta \\) such that \\( h^{\\star}(\\mathbf{x} + \\delta) \\neq y \\) cannot be adversarial at \\( \\mathbf{x} \\) as the corrupted input \\( \\mathbf{x} + \\delta \\) has a different true label. Finally, if \\( h \\) predicts the label at \\( \\mathbf{x} \\) incorrectly, then \\( \\mathbf{0} \\) is already an adversarial perturbation." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.639, + 0.895, + 0.654 + ], + "angle": 0, + "content": "2.2.1. Adversarial Robustness via Threat Specification" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.657, + 0.906, + 0.734 + ], + "angle": 0, + "content": "A threat function \\( d: \\mathcal{X} \\times \\mathbb{R}^d \\to \\mathbb{R}^{\\geq 0} \\) presents practitioners a way to specify the kind of robustness suitable for a particular learning task. The \\( \\varepsilon \\)-sublevel set of a threat function \\( d \\) at an input \\( \\mathbf{x} \\), \\( S(\\mathbf{x}, d, \\varepsilon) := \\{\\delta \\in \\mathbb{R}^d \\mid d(\\mathbf{x}, \\delta) \\leq \\varepsilon\\} \\), defines the set of perturbations with a threat at most \\( \\varepsilon \\)." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.744, + 0.909, + 0.867 + ], + "angle": 0, + "content": "Definition 2 (Threat Model, Robust Accuracy). A threat model \\((d,\\varepsilon)\\) is a pair of threat function \\(d\\) and permissible threshold \\(\\varepsilon\\) that together define the set of permissible perturbations at each input \\(\\mathbf{x}\\) as the sublevel set \\(S(\\mathbf{x},d,\\varepsilon)\\). Under the threat model \\((d,\\varepsilon)\\), the robust accuracy RobustAcc \\((h,(d,\\varepsilon))\\) of a classifier \\(h\\in \\mathcal{H}\\) is the probability of stable prediction upon corruptions within the permissible set, i.e. \\(\\mathrm{Prob}_{\\mathbf{x}}\\big[\\forall \\delta \\in S(\\mathbf{x},d,\\varepsilon), h(\\mathbf{x} + \\delta) = h^{\\star}(\\mathbf{x})\\big]\\)." + }, + { + "type": "page_footnote", + "bbox": [ + 0.513, + 0.875, + 0.909, + 0.901 + ], + "angle": 0, + "content": "\\( {}^{5} \\) Sampling \\( \\mathbf{z} \\sim {\\mathcal{D}}_{\\mathcal{Z}} \\) is equivalent to sampling(x,h*(x))where \\( \\mathbf{x} \\sim \\) \\( {\\mathcal{D}}_{\\mathcal{X}} \\) ." + }, + { + "type": "page_footnote", + "bbox": [ + 0.091, + 0.875, + 0.483, + 0.901 + ], + "angle": 0, + "content": "4Realizability in the context of adversarial robustness has also been studied in Awasthi et al. [2], Tramér et al. [43]." + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.947, + 0.516, + 0.958 + ], + "angle": 0, + "content": "9956" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.151, + 0.087, + 0.429, + 0.273 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.09, + 0.283, + 0.486, + 0.311 + ], + "angle": 0, + "content": "Figure 2. An illustration of unsafe directions, and sub-level sets of the PD-threat." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.339, + 0.484, + 0.4 + ], + "angle": 0, + "content": "Thus we define robustness as the stability of a classifier's prediction under perturbations within the permissible sets. The task of robust supervised learning seeks a classifier in \\(\\mathcal{H}\\) with high robust accuracy." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.41, + 0.484, + 0.456 + ], + "angle": 0, + "content": "Definition 3 (\\(\\varepsilon\\)-robust). A classifier \\(h \\in \\mathcal{H}\\) is \\(\\varepsilon\\)-robust at \\(\\mathbf{x}\\) w.r.t. a threat function \\(d\\) if, \\(\\forall \\delta \\in S(\\mathbf{x}, d, \\varepsilon) \\cap (\\mathcal{X} - \\{\\mathbf{x}\\})\\), we have \\(h(\\mathbf{x} + \\delta) = h^{\\star}(\\mathbf{x})\\)." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.467, + 0.484, + 0.573 + ], + "angle": 0, + "content": "An \\(\\varepsilon\\)-robust classifier predicts the same label \\(h^{\\star}(\\mathbf{x})\\) at \\(\\mathbf{x} + \\delta\\) for any corruption \\(\\delta\\) within the \\(\\varepsilon\\)-sublevel set \\(S(\\mathbf{x}, d, \\varepsilon)\\). We emphasize that robust accuracy and \\(\\varepsilon\\)-robustness are measures of stability rather than correctness. In this way, each threat model \\((d, \\varepsilon)\\) implicitly encodes a trade-off between stability and correctness. [13]. Requiring more stability than necessary can lead to incorrect predictions as in Figure 1." + }, + { + "type": "title", + "bbox": [ + 0.09, + 0.587, + 0.427, + 0.606 + ], + "angle": 0, + "content": "3. Threat specification beyond \\(\\ell_p\\) norms" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.613, + 0.484, + 0.718 + ], + "angle": 0, + "content": "In this section, we design a principled threat model from first principles that is local and anisotropic. We first theoretically develop our exact PD\\*-threat model assuming knowledge of the decision boundaries of the true labeling function \\(h^\\star\\) (Section 3.1), and later relax this assumption by developing an approximation leveraging the observed training data and obtaining a practical PD-threat (Section 3.2)." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.729, + 0.414, + 0.745 + ], + "angle": 0, + "content": "3.1. Measuring threat with class partition" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.751, + 0.485, + 0.813 + ], + "angle": 0, + "content": "Definition 4 (Unsafe Directions). At each input \\(\\mathbf{x}\\), a direction \\(\\mathbf{u} \\in \\mathbb{S}_2^d\\) is called unsafe if there exists a step size \\(t \\geq 0\\) such that \\(\\mathbf{x} + t\\mathbf{u} \\in \\mathcal{X}\\) and \\(h^\\star(\\mathbf{x} + t\\mathbf{u}) \\neq h^\\star(\\mathbf{x})\\). We denote the set of all unsafe directions at \\(\\mathbf{x}\\) as \\(\\mathcal{U}^\\star(\\mathbf{x})\\)." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.823, + 0.484, + 0.854 + ], + "angle": 0, + "content": "The minimum step size needed to alter the true label can vary across unsafe directions. Further, \\(\\mathcal{U}^{\\star}(\\mathbf{x})\\) varies" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.091, + 0.907, + 0.198 + ], + "angle": 0, + "content": "with input. Directions \\(\\mathbf{v} \\in \\mathcal{B}_2^d \\cap (\\mathcal{U}^\\star(\\mathbf{x}))^c\\) are called safe, since moving along them (whilst domain constrained) cannot alter the true label. We note that adversarial perturbations (Definition 1) are a characteristic of a learned predictor \\(h \\in \\mathcal{H}\\) while unsafe directions are a characteristic of the true labeling function \\(h^\\star\\). We propose to measure threat of a perturbations \\(\\delta\\) using \\(\\mathcal{U}^\\star(\\mathbf{x})\\)." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.207, + 0.908, + 0.283 + ], + "angle": 0, + "content": "Definition 5 (PD\\* -threat). Let \\(\\mathbf{x}\\in \\mathcal{X}\\) and let \\(\\delta \\in \\mathbb{R}^d\\) be a perturbation. The exact projected displacement threat function \\(d_{PD}^{\\star}\\) is defined as the maximum scaled displacement of the Euclidean projection of the perturbation \\(\\delta\\) over all unsafe directions," + }, + { + "type": "equation", + "bbox": [ + 0.535, + 0.293, + 0.907, + 0.328 + ], + "angle": 0, + "content": "\\[\nd _ {\\mathrm {P D}} ^ {\\star} (\\mathbf {x}, \\boldsymbol {\\delta}) := \\sup _ {\\mathbf {u} \\in \\mathcal {U} ^ {\\star} (\\mathbf {x})} \\frac {1}{g ^ {\\star} (\\mathbf {x} , \\mathbf {u})} \\max \\left(\\langle \\boldsymbol {\\delta}, \\mathbf {u} \\rangle , 0\\right) \\quad (1)\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.339, + 0.859, + 0.355 + ], + "angle": 0, + "content": "where \\(g^{\\star}(\\mathbf{u},\\mathbf{x})\\) is called the normalization function," + }, + { + "type": "equation", + "bbox": [ + 0.523, + 0.366, + 0.895, + 0.392 + ], + "angle": 0, + "content": "\\[\ng ^ {\\star} (\\mathbf {x}, \\mathbf {u}) := \\sup _ {\\mathbb {R} ^ {\\geq 0}} M \\text {s . t .} \\mathbf {x} + t \\mathbf {u} \\in \\mathcal {X} _ {h ^ {\\star} (\\mathbf {x})} \\forall t \\in [ 0, M ].\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.404, + 0.907, + 0.509 + ], + "angle": 0, + "content": "At each \\(\\mathbf{x}\\), the normalization function \\(g^{*}(\\mathbf{x},\\mathbf{u})\\) captures the distance along the direction \\(\\mathbf{u}\\) to the boundary of the set \\(\\mathcal{X}_{h^{\\star}(\\mathbf{x})}\\) and enables a comparison of the threat along different unsafe directions (see Figure 2). In (1), the supremum ensures that we do not underestimate the threat along a direction that potentially alters the true label. We highlight two key characteristics." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.51, + 0.907, + 0.615 + ], + "angle": 0, + "content": "Anisotropy. The threat function \\(d_{\\mathrm{PD}}^{\\star}\\) is anisotropic (the threat of a perturbation \\(\\delta\\) at input \\(\\mathbf{x}\\) depends on both the direction and magnitude of the perturbation) and does not correspond to a norm in general, unlike \\(\\ell_p\\) threat models. For a fixed \\(\\mathbf{x}\\), the threat along different directions \\(\\mathbf{u}\\) varies based on the alignment with unsafe directions \\(\\mathcal{U}^{\\star}(\\mathbf{x})\\). In particular, \\(d_{\\mathrm{PD}}^{\\star}\\) can differentiate between the perturbations \\(\\delta\\) and \\(-\\delta\\)." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.616, + 0.907, + 0.692 + ], + "angle": 0, + "content": "Locality. The threat function \\( d_{\\mathrm{PD}}^{\\star} \\) exhibits locality; i.e. a fixed perturbation \\( \\delta \\) has varying threat \\( d_{\\mathrm{PD}}^{\\star}(\\mathbf{x}, \\delta) \\) depending on the input \\( \\mathbf{x} \\). Thus the permissible set varies with input7, enabling an input-aware assessment of threat. We note that weighted \\( \\ell_p \\) norms [14] can be anisotropic but not local." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.699, + 0.797, + 0.714 + ], + "angle": 0, + "content": "3.1.1. Stability vs Correctness Trade-off" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.718, + 0.906, + 0.793 + ], + "angle": 0, + "content": "Since the normalization function \\( g^{\\star} \\) accounts for the class partition \\( \\{\\mathcal{X}_c\\}_{c=1}^C \\), the threat function \\( d_{\\mathrm{PD}}^{\\star} \\) has a principled trade-off between stability and correctness. Specifically, the true labeling function \\( h^{\\star} \\) exhibits a high level of robustness under the PD\\(^{\\star}\\)-threat8." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.805, + 0.907, + 0.865 + ], + "angle": 0, + "content": "Theorem 1. The true labeling function \\( h^{\\star} \\) is 1-robust at any input \\( \\mathbf{x} \\in \\mathcal{X} \\) w.r.t. the threat \\( d_{\\mathrm{PD}}^{\\star} \\). Additionally, if a classifier \\( h \\) is not 1-robust at all inputs then there exists an input \\( \\mathbf{x} \\) misclassified by \\( h \\)." + }, + { + "type": "page_footnote", + "bbox": [ + 0.53, + 0.874, + 0.868, + 0.888 + ], + "angle": 0, + "content": "Unlike \\(\\ell_p\\) threat models where \\(S(\\mathbf{x},d_p,\\varepsilon) = \\varepsilon \\mathcal{B}_p^d\\forall \\mathbf{x}\\in \\mathcal{X}\\)" + }, + { + "type": "page_footnote", + "bbox": [ + 0.533, + 0.888, + 0.826, + 0.9 + ], + "angle": 0, + "content": "The proofs of all our results can be found in Section 8." + }, + { + "type": "list", + "bbox": [ + 0.53, + 0.874, + 0.868, + 0.9 + ], + "angle": 0, + "content": null + }, + { + "type": "page_footnote", + "bbox": [ + 0.09, + 0.863, + 0.484, + 0.902 + ], + "angle": 0, + "content": "Note as per Definition 2, robust accuracy measures stability rather than correctness since we aren't evaluating probability of the event \\(\\{\\forall \\delta \\in S(\\mathbf{x},d,\\varepsilon), h(\\mathbf{x} + \\delta) = h^{\\star}(\\mathbf{x} + \\delta)\\}\\)." + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.947, + 0.516, + 0.958 + ], + "angle": 0, + "content": "9957" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.09, + 0.092, + 0.485, + 0.155 + ], + "angle": 0, + "content": "In other words, for any pairs of inputs \\(\\mathbf{x},\\tilde{\\mathbf{x}}\\in \\mathcal{X}\\) with different labels, the threat \\(d_{\\mathrm{PD}}^{\\star}(\\mathbf{x},\\tilde{\\mathbf{x}} -\\mathbf{x})\\) is always larger than 1. Thus seeking \\(h\\in \\mathcal{H}\\) that is 1-robust at any input under \\(d_{\\mathrm{PD}}^{\\star}\\) is not at odds with learning \\(h^{\\star}\\)." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.161, + 0.419, + 0.177 + ], + "angle": 0, + "content": "3.2. Measuring threats with observed data" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.182, + 0.484, + 0.258 + ], + "angle": 0, + "content": "Computing \\(d_{\\mathrm{PD}}^{\\star}(\\mathbf{x},\\cdot)\\) from Definition 5 requires explicitly characterizing the set of unsafe directions \\(\\mathcal{U}^{\\star}(\\mathbf{x})\\) and the normalization function \\(g^{\\star}(\\mathbf{u},\\mathbf{x})\\), both of which need knowledge of the partition sets \\(\\{\\mathcal{X}_c\\}_{c = 1}^C\\) induced by \\(h^\\star\\). In this subsection, we propose a heuristic empirical approximation." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.259, + 0.484, + 0.428 + ], + "angle": 0, + "content": "Approximating \\(\\mathcal{U}^{\\star}(\\mathbf{x})\\). We have access to training data S and incomplete information on the partition sets \\(\\{\\mathcal{X}_c\\}_{c=1}^C\\) via the partition \\(S = \\{S_c\\}_{c=1}^C\\), where \\(S_c := \\{\\mathbf{x} \\in \\mathcal{X} \\mid (\\mathbf{x}, c) \\in S\\} \\subset \\mathcal{X}_c\\). At every training input \\(\\mathbf{x} \\in S_y\\) for each input \\(\\tilde{\\mathbf{x}} \\in S_c\\) (for \\(c \\neq y\\)), the direction \\(\\mathbf{u} = \\frac{\\tilde{\\mathbf{x}} - \\mathbf{x}}{\\|\\tilde{\\mathbf{x}} - \\mathbf{x}\\|_2}\\) is an unsafe direction. For computational considerations, we choose a collection \\(\\bigcup_{c} S_{c,k}\\) of representative subsets \\(S_{c,k} \\subset S_c\\) of size \\(k\\). As a heuristic choice, we select subsets \\(S_{c,k}\\) by solving a discrete-k-center optimization problem (detailed in Sec. 9.1). We are now ready to define an observed subset of the unsafe directions based on \\(S_{c,k}\\)." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.435, + 0.484, + 0.482 + ], + "angle": 0, + "content": "Definition 6 (Observed Unsafe Directions). For each \\(\\mathbf{x} \\in S_y\\), we identify a set of \\((C - 1)k\\)-observed unsafe directions based on the representative subsets \\(S_{c,k}\\)," + }, + { + "type": "equation", + "bbox": [ + 0.122, + 0.49, + 0.45, + 0.526 + ], + "angle": 0, + "content": "\\[\n\\mathcal {U} _ {k} (\\mathbf {x}) := \\left\\{\\frac {\\tilde {\\mathbf {x}} - \\mathbf {x}}{\\| \\tilde {\\mathbf {x}} - \\mathbf {x} \\| _ {2}} \\Big | \\tilde {\\mathbf {x}} \\in \\cup_ {c \\neq y} \\mathbb {S} _ {c, k} \\right\\} \\subset \\mathcal {U} ^ {\\star} (\\mathbf {x}).\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.54, + 0.485, + 0.635 + ], + "angle": 0, + "content": "Approximating \\( g^{\\star}(\\mathbf{x},\\mathbf{u}) \\). At any \\( \\mathbf{x}\\in \\mathcal{X}_y \\), for each \\( \\mathbf{u}\\in \\mathcal{U}_k(\\mathbf{x}) \\), there exists \\( \\tilde{\\mathbf{x}}\\in \\cup_{c\\neq y}\\mathrm{S}_{c,k} \\) such that \\( \\mathbf{u} = \\frac{\\tilde{\\mathbf{x}} - \\mathbf{x}}{\\|\\tilde{\\mathbf{x}} - \\mathbf{x}\\|_2} \\). normalization function can be bounded as \\( g^{\\star}(\\mathbf{x},\\mathbf{u})\\leq \\| \\tilde{\\mathbf{x}} -\\mathbf{x}\\| _2 \\), since \\( h^\\star (\\tilde{\\mathbf{x}})\\neq h^\\star (\\mathbf{x}) \\). In practice, we use the heuristic \\( g_{\\beta}(\\mathbf{u},\\mathbf{x})\\coloneqq \\beta \\| \\tilde{\\mathbf{x}} -\\mathbf{x}\\| _2 \\), where \\( \\beta \\in (0,1) \\) is a scaling hyper-parameter." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.643, + 0.484, + 0.687 + ], + "angle": 0, + "content": "Definition 7 \\(((k,\\beta)\\) -PD threat). Let \\(\\mathbf{x}\\in \\mathcal{X}\\) and let \\(\\delta \\in \\mathbb{R}^d\\) be a perturbation. The projected displacement threat is defined as" + }, + { + "type": "equation", + "bbox": [ + 0.106, + 0.694, + 0.468, + 0.727 + ], + "angle": 0, + "content": "\\[\nd _ {\\mathrm {P D}, k, \\beta} (\\mathbf {x}, \\boldsymbol {\\delta}) := \\max _ {\\mathbf {u} \\in \\mathcal {U} _ {k} (\\mathbf {x})} \\frac {1}{g _ {\\beta} (\\mathbf {x} , \\mathbf {u})} \\max \\left(\\langle \\boldsymbol {\\delta}, \\mathbf {u} \\rangle , 0\\right).\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.735, + 0.486, + 0.903 + ], + "angle": 0, + "content": "The quality of the approximation of the exact threat \\(d_{\\mathrm{PD}}^{\\star}\\) by \\(d_{\\mathrm{PD,k},\\beta}\\) depends on the choice of representative unsafe directions \\(\\mathcal{U}_k(\\cdot)\\) via the \\(k\\)-subset \\(\\mathbb{S}_{c,k}\\) and on the heuristic choice of the approximate normalization \\(g_{\\beta}\\), via the scaling hyper-parameter \\(\\beta\\). In this article, we fix \\((k,\\beta) = (50,\\frac{1}{2})\\) as the default hyper-parameters (see Section 9.3 for an expanded discussion on choosing hyper-parameters \\((k,\\beta)\\) for any learning task). For brevity, we refer to the observed unsafe directions (Definition 6), the heuristic normalizations and the PD threat (Definition 7) as \\(\\mathcal{U}(\\mathbf{x}), g(\\mathbf{x},\\delta)\\) and \\(d_{\\mathrm{PD}}(\\mathbf{x},\\delta)\\), respectively." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.092, + 0.905, + 0.138 + ], + "angle": 0, + "content": "We now state a few important properties of our proposed empirical approximation \\( d_{\\mathrm{PD}} \\). Clearly \\( d_{\\mathrm{PD}} \\) inherits the anisotropy and locality properties of the exact threat \\( d_{\\mathrm{PD}}^{\\star} \\)." + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.138, + 0.907, + 0.215 + ], + "angle": 0, + "content": "Growth and Sensitivity. The growth of PD threat along any direction is linear, i.e., \\(d_{\\mathrm{PD}}(\\mathbf{x},t\\boldsymbol {\\delta}) = td_{\\mathrm{PD}}(\\mathbf{x},\\boldsymbol {\\delta})\\) , and the rate of growth scales inversely with the (approximate) normalization. In particular, for a fixed \\(\\mathbf{x}\\) \\(d_{\\mathrm{PD}}(\\mathbf{x},\\cdot)\\) is \\((\\max_{u\\in \\mathcal{U}(\\mathbf{x})}1 / g(\\mathbf{x},\\mathbf{u}))\\) -Lipschitz w.r.t \\(\\ell_2\\) norm." + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.215, + 0.909, + 0.359 + ], + "angle": 0, + "content": "Projection onto sublevel sets. For any threat model \\((d,\\varepsilon)\\), the ability to efficiently project a perturbation \\(\\delta\\) onto the sub-level sets \\(S(d,\\mathbf{x},\\varepsilon)\\) enables one to leverage a wide literature on gradient-based attacks to conduct a rigorous evaluations of adversarial robustness. The sublevel sets of PD threat model, i.e., \\(S(\\mathbf{x},d,\\varepsilon)\\), are convex for all \\(\\mathbf{x}\\) for all \\(\\epsilon \\geq 0\\). In particular, \\(\\mathcal{S}\\) is the intersection of at most \\(k\\cdot C\\) half-spaces characterized by unsafe directions \\(\\mathcal{U}(\\mathbf{x})\\) as \\(S(\\mathbf{x},d_{\\mathrm{PD}},\\varepsilon) = \\bigcap_{\\mathbf{u}\\in \\mathcal{U}(\\mathbf{x})}\\{\\pmb {\\delta}\\in \\mathbb{R}^d\\mid \\langle \\pmb {\\delta},\\mathbf{u}\\rangle \\leq \\varepsilon \\cdot g(\\mathbf{x},\\mathbf{u})\\}\\)." + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.358, + 0.907, + 0.388 + ], + "angle": 0, + "content": "Further, the sub-level sets are monotonic w.r.t. the threshold, i.e., for \\(\\varepsilon_1\\geq \\varepsilon_2\\geq 0,S(\\mathbf{x},d_{\\mathrm{PD}},\\varepsilon_2)\\subseteq S(\\mathbf{x},d_{\\mathrm{PD}},\\varepsilon_1)\\)" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.389, + 0.909, + 0.569 + ], + "angle": 0, + "content": "Linearity of growth and convexity of the sub-level sets together provide a straightforward approximate projection algorithm, since for any perturbation \\(\\delta\\), the scaled perturbation \\(\\frac{\\varepsilon}{d_{\\mathrm{PD}}(\\mathbf{x},\\delta)}\\delta\\) lies in the permissible set \\(S(\\mathbf{x},d_{\\mathrm{PD}},\\varepsilon)\\). Further, due to the convexity of sublevel sets, an exact projection (in \\(\\ell_2\\) distance) can be computed using the iterative greedy scheme outlined in the Algorithm 2 (deferred to the appendix). Thus the PD threat provides the practitioner with the flexibility of a fast approximate projection or an exact projection at the expense of increased computation. Together, these projection algorithms enable immediate adaptation of existing adversarial attacks to our proposed threat model." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.57, + 0.909, + 0.647 + ], + "angle": 0, + "content": "We note that efficient exact projection algorithms are also known for \\(\\ell_p\\) threat models. However, for neural perceptual threat models such as DreamSim [15], the sub-level sets are non-convex and a projection can only be approximately computed [32] and is typically not computationally efficient." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.662, + 0.818, + 0.68 + ], + "angle": 0, + "content": "4. Experiments on Real World Data" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.689, + 0.909, + 0.825 + ], + "angle": 0, + "content": "In this section we illustrate the qualitative and quantitative characteristics of the PD-threat in comparison to existing threat models. We evaluate 4 threat functions: \\( d_{\\infty}, d_{2}, d_{\\mathrm{DS}} \\), and our proposed PD threat function \\( d_{\\mathrm{PD}} \\), where \\( d_{\\mathrm{DS}} \\) is the neural perceptual threat function \\( d_{\\mathrm{DS}}(\\mathbf{x}, \\tilde{\\mathbf{x}} - \\mathbf{x}) \\coloneqq \\mathrm{DreamSim}(\\mathbf{x}, \\tilde{\\mathbf{x}}) \\) from DreamSim [15]. Note that we focus on DreamSim due to its superior performance over other neural perceptual threat models [15], but the following comparison applies to any modern neural perceptual threat model." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.826, + 0.909, + 0.903 + ], + "angle": 0, + "content": "Experimental Setup. We seek to measure the ability of threat functions to distinguish between safe and unsafe corruptions. The threat models \\(d_{\\mathrm{PD}}\\) and \\(d_{\\mathrm{DS}}\\) depend on the Imagenet-1k training dataset S. All threat models are evaluated on the Imagenet-1k validation dataset \\(\\mathbf{S}_{\\mathrm{val}}\\). For any pair" + }, + { + "type": "page_number", + "bbox": [ + 0.482, + 0.947, + 0.516, + 0.958 + ], + "angle": 0, + "content": "9958" + } + ], + [ + { + "type": "table", + "bbox": [ + 0.093, + 0.089, + 0.485, + 0.272 + ], + "angle": 0, + "content": "
CATEGORYCORRUPTION STYLES
NoiseGaussian, Shot, Impulse, Speckle, Blue, Brown, Perlin, Single-Frequency, Plasma
BlurGaussian, Defocus, Glass, Motion, Concentric Sine Waves, Caustic Refraction
CompressionPixelate, JPEG
DigitalBrightness, Contrast, Saturate, Elastic Transform
WeatherFrost, Fog, Snow
OcclusionSpatter, Checkerboard Cutout, Sparkles, Inverse Sparkles
" + }, + { + "type": "table_caption", + "bbox": [ + 0.09, + 0.282, + 0.484, + 0.311 + ], + "angle": 0, + "content": "Table 2. Categories of corruptions studied, from Imagenet-C [25] and Imagenet- \\(\\tilde{\\mathbf{C}}\\) [36]." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.337, + 0.484, + 0.488 + ], + "angle": 0, + "content": "of labelled inputs \\((\\mathbf{x},y),(\\tilde{\\mathbf{x}},c)\\in S_{\\mathrm{val}}\\) with distinct labels \\((c\\neq y)\\), the perturbation \\(\\delta \\coloneqq \\tilde{\\mathbf{x}} -\\mathbf{x}\\) is an unsafe perturbation at \\(\\mathbf{x}\\). This way, we can evaluate the statistics of threat under unsafe perturbations for each threat function. To characterize threat under safe perturbations, we consider a set \\(\\Omega\\) of 150 distinct common corruptions consisting of 30 styles \\(\\times 5\\) severity levels, i.e., severity \\(\\in \\{1,2,3,4,5\\}\\) - the supplementary material contains visual examples (see Figures 8 and 9), sourced from Imagenet- \\(C\\) [25] and Imagenet- \\(\\bar{\\mathbf{C}}\\) [36] accompanied with their threat assessment." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.489, + 0.484, + 0.581 + ], + "angle": 0, + "content": "For a specific corruption style such as Gaussian noise and a severity level 5, we let \\(\\Omega_{\\mathrm{GaussianNoise},5}\\) denote the corresponding corruption function. To aid visualization, we group the corruption styles into 6 distinct categories (listed in Table 2) so that \\(\\Omega := \\cup_{i \\in [5]} \\Omega_i\\) where \\(\\Omega_i\\) is the set of all corruption styles at severity level \\(i\\)," + }, + { + "type": "equation", + "bbox": [ + 0.09, + 0.59, + 0.493, + 0.617 + ], + "angle": 0, + "content": "\\[\n\\Omega_ {i} := \\left(\\Omega_ {\\text {n o i s e}, i} \\cup \\Omega_ {\\text {b l u r}, i} \\cup \\Omega_ {\\text {c o m p r e s s i o n}, i} \\cup \\Omega_ {\\text {d i g i t a l}, i} \\cup \\Omega_ {\\text {w e a t h e r}, i} \\cup \\Omega_ {\\text {o c c l u s i o n}, i}\\right).\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.625, + 0.487, + 0.734 + ], + "angle": 0, + "content": "Each individual corruption \\(\\omega \\in \\Omega\\) is applied to a subset \\(\\bar{\\mathbf{S}}\\subset \\mathbf{S}_{\\mathrm{val}}\\) of 5,000 images chosen uniformly at random. For each threat function \\(d\\in \\{d_{\\infty},d_2,d_{\\mathrm{PD}},d_{\\mathrm{DS}}\\}\\) and corruption \\(\\omega \\in \\Omega\\), for each image \\(\\mathbf{x}\\in \\bar{\\mathbf{S}}\\), we denote by \\(\\mathrm{avg}(d,\\omega)\\), the average threat statistic, \\(\\mathrm{avg}(d,\\omega)\\coloneqq \\frac{1}{|\\bar{s}|}\\sum_{(\\mathbf{x},y)\\in S}d(\\mathbf{x},\\omega (\\mathbf{x}) - \\mathbf{x})\\), where \\(\\omega (\\mathbf{x})\\) is the corrupted input and \\(\\omega (\\mathbf{x}) - \\mathbf{x}\\) is the safe perturbation at \\(\\mathbf{x}\\)." + }, + { + "type": "title", + "bbox": [ + 0.09, + 0.743, + 0.455, + 0.76 + ], + "angle": 0, + "content": "4.1. Disentangling Safe and Unsafe Corruption" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.765, + 0.484, + 0.869 + ], + "angle": 0, + "content": "We now measure the ability of threat models to distinguish safe and unsafe corruptions. Each corruption \\(\\omega \\in \\Omega\\) defines a point in Figure 3a having as the horizontal coordinate \\(\\mathrm{avg}(d_{\\infty},\\omega)\\) and \\(\\mathrm{avg}(d_{\\mathrm{PD}},\\omega)\\) as the vertical coordinate, producing a comparison of our PD threat to the \\(\\ell_{\\infty}\\)-threat model. Similarly, Figure 3b compares PD threat to the DreamSIM-threat model." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.871, + 0.484, + 0.903 + ], + "angle": 0, + "content": "A specification of threat model \\((d, \\varepsilon)\\) corresponds to the choice of a horizontal line for \\(d_{\\mathrm{PD}}\\) or a vertical line for \\((d_{\\infty}\\)" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.092, + 0.905, + 0.212 + ], + "angle": 0, + "content": "or \\( d_{\\mathrm{DS}} \\)) in Figure 3. We note that a size-able set of common corruptions incur large \\( d_{\\infty} \\) threat, in particular as large as the unsafe corruptions (as indicated in Figure 1), thus any vertical line in Figure 3a is bound to exclude such common corruptions. In contrast, both \\( d_{\\mathrm{PD}} \\) and \\( d_{\\mathrm{DS}} \\) admit choices of \\( \\varepsilon \\) that exclude unsafe perturbations while including most common corruptions. Hence, PD threat and DS threat are able to disentangle safe and unsafe corruptions." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.214, + 0.908, + 0.305 + ], + "angle": 0, + "content": "For each of the comparisons in Figure 3, we identify 4 quadrants I-IV, distinguishing between low-threat and high-threat corruptions as measured by the corresponding threat models. We define the quadrants by choosing thresholds of 1, 0.5, and 0.25 for \\( d_{\\mathrm{PD}} \\), \\( d_{\\infty} \\), and \\( d_{\\mathrm{DS}} \\), respectively9, and comment on the qualitative behavior observed:" + }, + { + "type": "text", + "bbox": [ + 0.514, + 0.306, + 0.907, + 0.412 + ], + "angle": 0, + "content": "- Quadrant IV contains corruptions that are characterized as low-threat by both threat models. As expected, almost all corruptions having a low severity (1) lie in this region. However, we already see deficiencies of the \\(d_{\\infty}\\) threat, as some blur, noise, and occlusion corruptions of severity 1 in Figure 3a are not characterized as low-threat by at least one of the threat models." + }, + { + "type": "text", + "bbox": [ + 0.514, + 0.412, + 0.907, + 0.472 + ], + "angle": 0, + "content": "- Quadrant II contains corruptions that are characterized as high-threat by both threat models. These include digital and weather corruptions of high severity (5) as they tend to remove most of the signal in the input." + }, + { + "type": "text", + "bbox": [ + 0.514, + 0.473, + 0.909, + 0.621 + ], + "angle": 0, + "content": "- Quadrant I contains corruptions having low PD-threat but high \\(d_{\\infty}\\) or \\(d_{\\mathrm{DS}}\\) and showcases the resilience of our PD-threat model to natural corruptions like blur and Gaussian noise, which retain a lot of the information in the image relevant to the classification task. Indeed, a visual inspection of these corruptions (see Section 10 in the Appendix) reveals that even at a high severity, a human is able to discern task-relevant information from the corrupted images, and hence these corruptions should be assigned a low threat." + }, + { + "type": "text", + "bbox": [ + 0.514, + 0.623, + 0.909, + 0.715 + ], + "angle": 0, + "content": "- Quadrant III contains corruptions with high PD threat but low \\( d_{\\infty} \\) or \\( d_{\\mathrm{DS}} \\) threat. We note that corruptions rate as high threat by \\( d_{\\mathrm{PD}} \\) are largely also rated highly by \\( d_{\\mathrm{DS}} \\). Since the PD and DS threats are aligned, the corruptions in Quadrant III in Figure 3a indicate that low \\( d_{\\infty} \\) is not sufficient to characterize hard corruptions." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.716, + 0.909, + 0.837 + ], + "angle": 0, + "content": "As a final point, we note that while the size of markers in Figure 3 indicate the severity levels, different corruptions of similar severity levels are qualitatively different. For e.g, weather corruptions of severity level 5 are uniformly harder than noise corruptions such as shot noise, Gaussian noise. Figure 11 illustrates heatmap of average threat for each corruption category across increasing severity. We observe that both PD and DS threat exhibit largely monotonic" + }, + { + "type": "page_footnote", + "bbox": [ + 0.512, + 0.851, + 0.908, + 0.903 + ], + "angle": 0, + "content": "9These thresholds were chosen to approximately be half of the average threat of unsafe perturbations that alter the true label. Admittedly, this binary threshold is coarse, and the Appendix contains fine-grained qualitative examples of corruptions of varying threat." + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.947, + 0.516, + 0.958 + ], + "angle": 0, + "content": "9959" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.09, + 0.092, + 0.486, + 0.122 + ], + "angle": 0, + "content": "increase across severity unlike \\( d_{\\infty} \\). At this point, we will pause to summarize the benefits of PD-threat." + }, + { + "type": "image", + "bbox": [ + 0.129, + 0.136, + 0.446, + 0.387 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.279, + 0.388, + 0.297, + 0.399 + ], + "angle": 0, + "content": "(a)" + }, + { + "type": "image", + "bbox": [ + 0.128, + 0.401, + 0.445, + 0.631 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.279, + 0.632, + 0.297, + 0.643 + ], + "angle": 0, + "content": "(b)" + }, + { + "type": "image_caption", + "bbox": [ + 0.09, + 0.657, + 0.483, + 0.685 + ], + "angle": 0, + "content": "Figure 3. Comparison of Threat models w.r.t. various common corruptions. The size of a marker represents the severity level." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.705, + 0.486, + 0.9 + ], + "angle": 0, + "content": "PD vs DreamSim. The PD-threat compares well to the DS-threat but does not require pre-trained large vision models or fine-tuning on a curated dataset of human perceptual similarity judgements. For real-world classification tasks where perceptual similarity between pairs of input is harder to quantify even with expert human annotators, instantiation of high-quality neural perceptual distance metrics is harder. Further, unlike perceptual distance metrics [15, 17, 18, 32, 54], the growth of the PD-threat is linear and interpretable along any direction, the sub-level sets induced by the PD-threat model are convex, and admit efficient projection algorithms and hence PD threat presents a promising alternative." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.091, + 0.871, + 0.108 + ], + "angle": 0, + "content": "5. Measuring threat with Task Annotation" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.115, + 0.907, + 0.19 + ], + "angle": 0, + "content": "We described two practical design choices for PD-threat in Section 3.2: (1) a choice of the set of unsafe directions \\(\\mathcal{U}(\\mathbf{x})\\) at each input, and (2) a choice of normalization \\(g(\\mathbf{x},\\mathbf{u})\\) for each unsafe direction. Next, we incorporate additional task-relevant information to refine PD-threat further." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.2, + 0.866, + 0.216 + ], + "angle": 0, + "content": "5.1. Segmentation-aware Threat Specification" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.223, + 0.908, + 0.312 + ], + "angle": 0, + "content": "Segmentation masks isolate regions of the input that contain semantic information relevant to the class label, and can be generated efficiently following recent advances in foundational models for image segmentation [30, 38]. We demonstrate a straightforward method to incorporate semantic masks in the assessment of threat." + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.321, + 0.908, + 0.353 + ], + "angle": 0, + "content": "Definition 8 (PD-S threat). Let \\(\\mathbf{x} \\in \\mathcal{X}\\) and let \\(\\delta \\in \\mathbb{R}^d\\). Let \\(\\mathbf{a} \\in \\{0,1\\}^d\\) be a boolean mask. The segmented PD threat," + }, + { + "type": "equation", + "bbox": [ + 0.513, + 0.374, + 0.915, + 0.407 + ], + "angle": 0, + "content": "\\[\nd _ {\\mathrm {P D - S}} (\\mathbf {x}, \\mathbf {a}, \\boldsymbol {\\delta}) := \\max _ {\\mathbf {u} \\in \\mathcal {U} (\\mathbf {x})} \\frac {1}{g _ {\\boldsymbol {\\beta}} (\\mathbf {x} , \\mathbf {u})} \\max \\left(\\langle \\boldsymbol {\\delta} [ \\mathbf {a} ], \\mathbf {u} [ \\mathbf {a} ] \\rangle , 0\\right).\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.416, + 0.909, + 0.568 + ], + "angle": 0, + "content": "Here the \\(\\mathbf{u}[\\mathbf{a}]\\) denotes a sub-vector whose indices are selected by the boolean mask \\(\\mathbf{a}\\). The threat model \\((d_{\\mathrm{PD-S}}, \\varepsilon)\\) is aligned with Xiao et al. [51]'s proposal to learn models robust to adversarially chosen backgrounds. For each image, we generate automatic masks by prompting SAM10 with the center pixel coordinates. Figure 4 illustrates the difference between PD and PD-S threat on 2 corruptions that are background-only and foreground-only respectively. PD-S threat is oblivious to background corruption but more sensitive to foreground corruption." + }, + { + "type": "image", + "bbox": [ + 0.532, + 0.58, + 0.619, + 0.648 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.549, + 0.654, + 0.602, + 0.662 + ], + "angle": 0, + "content": "Original Image" + }, + { + "type": "image", + "bbox": [ + 0.668, + 0.58, + 0.753, + 0.647 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.693, + 0.654, + 0.728, + 0.662 + ], + "angle": 0, + "content": "\\(\\mathrm{PD} = 1.4\\)" + }, + { + "type": "image_caption", + "bbox": [ + 0.691, + 0.662, + 0.73, + 0.669 + ], + "angle": 0, + "content": "PD-S=0.0" + }, + { + "type": "image", + "bbox": [ + 0.802, + 0.58, + 0.888, + 0.647 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.826, + 0.654, + 0.864, + 0.662 + ], + "angle": 0, + "content": "\\(\\mathrm{PD} = 0.68\\)" + }, + { + "type": "image_caption", + "bbox": [ + 0.825, + 0.662, + 0.867, + 0.669 + ], + "angle": 0, + "content": "PD-S=1.78" + }, + { + "type": "image_caption", + "bbox": [ + 0.524, + 0.682, + 0.895, + 0.696 + ], + "angle": 0, + "content": "Figure 4. Comparison of threat on background vs foreground." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.712, + 0.908, + 0.757 + ], + "angle": 0, + "content": "We note that \\(\\ell_p\\) threat models are also capable of integrating segmentation masks as they are pixel-based, but it is unclear if perceptual threat can account for pixel annotation." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.766, + 0.826, + 0.783 + ], + "angle": 0, + "content": "5.2. Concept-aware Threat Specification" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.788, + 0.907, + 0.879 + ], + "angle": 0, + "content": "The PD-threat assumes no hierarchy of class labels; i.e. all labels are distinct to the same degree. However, classification tasks are often accompanied with an additional hierarchy that identifies groups of class labels based on similar semantic content. We propose to account for such a hierarchy by refining our choice of normalization. Suppose \\( \\mathbf{x} \\) has label" + }, + { + "type": "page_footnote", + "bbox": [ + 0.528, + 0.887, + 0.836, + 0.901 + ], + "angle": 0, + "content": "10In particular we use the default ViT-H SAM checkpoint." + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.947, + 0.516, + 0.958 + ], + "angle": 0, + "content": "9960" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.09, + 0.091, + 0.485, + 0.186 + ], + "angle": 0, + "content": "\\(y\\) and \\(\\mathbf{u} = \\frac{\\tilde{\\mathbf{x}} - \\mathbf{x}}{\\|\\tilde{\\mathbf{x}} - \\mathbf{x}\\|_2}\\) where \\(\\tilde{\\mathbf{x}}\\in S_{c,k}\\) has label c. Let \\(W:[C]\\times [C]\\in [0,1]\\) denote a relative distance between class labels based on Wordnet-hierarchy (refer to Section 9.5 for explicit details). We propose the weighted normalization \\(g_{\\beta ,W}(\\mathbf{x},\\mathbf{u})\\coloneqq \\boldsymbol {\\beta}\\cdot W(y,c)\\| \\tilde{\\mathbf{x}} -\\mathbf{x}\\| _2\\) so that perturbation between nearby classes are weighted higher in threat." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.192, + 0.484, + 0.223 + ], + "angle": 0, + "content": "Definition 9 (PD-W threat). Let \\(\\mathbf{x} \\in \\mathcal{X}\\) with label \\(y\\) and let \\(\\pmb{\\delta} \\in \\mathbb{R}^{d}\\). The weighted PD threat is defined as," + }, + { + "type": "equation", + "bbox": [ + 0.1, + 0.228, + 0.474, + 0.261 + ], + "angle": 0, + "content": "\\[\nd_{\\mathrm{PD - W}}(\\mathbf{x},\\boldsymbol {\\delta}):= \\max_{\\mathbf{u}\\in \\mathcal{U}(\\mathbf{x})}\\frac{1}{g_{\\beta,W}(\\mathbf{x},\\mathbf{u})}\\max \\left(\\langle \\boldsymbol {\\delta},\\mathbf{u}\\rangle ,0\\right).\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.267, + 0.487, + 0.448 + ], + "angle": 0, + "content": "The threat model \\((d_{\\mathrm{PD - W}},\\varepsilon)\\) is weaker than \\((d_{\\mathrm{PD}},\\varepsilon)\\), since \\(S(\\mathbf{x},d_{\\mathrm{PD - W}},\\varepsilon)\\subseteq S(\\mathbf{x},d_{\\mathrm{PD}},\\varepsilon)\\). Thus PD-W is a relaxation of PD threat with a softer requirement of stability between nearby classes. Figure 5 depicts the average relative\\(^{12}\\) threat vs relative distance \\(W(\\cdot ,\\cdot)\\) of class labels on the Imagenet-1k validation dataset. The decreasing PD-W threat for increasing class distance is aligned with the following intuition - failure to distinguish between semantically distant classes such as ENGLISH FOXHOUND and FIRE TRUCK should incur lower robust accuracy than failure to distinguish between similar classes such as ENGLISH FOXHOUND and IRISH WOLFHOUND." + }, + { + "type": "image", + "bbox": [ + 0.149, + 0.468, + 0.401, + 0.622 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.116, + 0.637, + 0.458, + 0.651 + ], + "angle": 0, + "content": "Figure 5. Relative Threat vs relative inter-class distances" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.668, + 0.486, + 0.789 + ], + "angle": 0, + "content": "In summary our novel threat specification framework enables evaluation of robustness to adversarial perturbations and common corruptions while enabling sensitivity to semantic regions of an image and class hierarchy by incorporating additional task annotation. As a consequence, our threat model unifies adversarial robustness with variants of corruption robustness suggested by several independent benchmarks [25, 36, 51]." + }, + { + "type": "title", + "bbox": [ + 0.09, + 0.801, + 0.473, + 0.819 + ], + "angle": 0, + "content": "6. Evaluation of benchmark \\(\\ell_p\\) robust models" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.826, + 0.485, + 0.857 + ], + "angle": 0, + "content": "Finally, we evaluate state-of-the-art robust models against our proposed thread models. Note that, as per Theorem 1," + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.092, + 0.908, + 0.289 + ], + "angle": 0, + "content": "one should hope for classifiers that are 1-robust to corruptions under the PD-threat. The corresponding permissible set of corruptions cover regions IV and I in Figure 3. We evaluate robustness to corruptions in the set \\( S(\\mathbf{x}, d_{\\mathrm{PD}}, \\varepsilon) \\cap S(\\mathbf{x}, d_{\\infty}, \\varepsilon) \\). We denote the corresponding threat model \\( (d_{\\infty} \\bigcap d_{\\mathrm{PD}}, \\varepsilon) \\). A robust evaluation for such a threat model can be executed in practice by generating adversarial perturbations with standard AutoAttack [7] for \\( (d_{\\infty}, \\varepsilon) \\) followed by projection onto \\( S(\\mathbf{x}, d_{\\mathrm{PD}}, \\varepsilon) \\). Table 3 shows the robustness of benchmark classifiers registered at RobustBench order by \\( d_{\\infty} \\) robustness (column 1) when evaluated against \\( (d_{\\infty} \\bigcap d_{\\mathrm{PD}}, 16/255) \\) threat model for the Imagenet-1k classification task." + }, + { + "type": "table", + "bbox": [ + 0.522, + 0.3, + 0.9, + 0.424 + ], + "angle": 0, + "content": "
Benchmark Modeld∞PDPD-SPD-W
ConvNeXt-L [33]0.250.320.310.31
Swin-B [33]0.120.210.200.21
ConvNeXt-B [33]0.10.190.180.18
ConvNeXt-B-ConvStem [41]0.090.190.190.19
ViT-S-ConvStem [41]0.050.130.120.12
XCiT-L12 [10]0.040.170.180.18
" + }, + { + "type": "table_caption", + "bbox": [ + 0.576, + 0.433, + 0.843, + 0.448 + ], + "angle": 0, + "content": "Table 3. Robust Evaluation for \\( \\varepsilon = {16}/{255} \\) ." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.464, + 0.909, + 0.66 + ], + "angle": 0, + "content": "We note that robust accuracies uniformly improve under our threat specification in comparison to \\((d_{\\infty}, \\varepsilon)\\), indicating a meaningful restriction of unsafe corruptions. The ordering of robust accuracies highlights diversity, e.g. [10]'s XCiT-L12 fares better under PD threat than suggested by the \\(d_{\\infty}\\) leaderboard. The indistinguishability across PD, PD-S and PD-w is surprising. Similar levels of robustness w.r.t PD-W and PD indicate that models are unable to distinguish distant classes, while similar levels under PD and PD-S indicate adversarial attacks are often aligned with foreground masks. We note that adversarial attacks adapted to the threat specification often lead to a more accurate pessimistic evaluation of robustness." + }, + { + "type": "title", + "bbox": [ + 0.514, + 0.674, + 0.634, + 0.69 + ], + "angle": 0, + "content": "7. Conclusion" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.7, + 0.909, + 0.866 + ], + "angle": 0, + "content": "This work proposes a novel task-dependent threat specification Projected Displacement threat that is adapted to the shape of decision boundaries based on observed training data. Unlike standard \\(\\ell_p\\)-threat models, the proposed threat model exhibits anisotropy and locality and is thus able to distinguish between safe and unsafe perturbations. The proposed threat specification framework is flexible and can effectively account for additional task information, such as image regions or label hierarchies. We invite the community to suggest further adaptive attacks and effective training strategies to encourage robustness in PD threat." + }, + { + "type": "page_footnote", + "bbox": [ + 0.106, + 0.863, + 0.395, + 0.876 + ], + "angle": 0, + "content": "11 Relative since we require a normalized value in [0,1]" + }, + { + "type": "page_footnote", + "bbox": [ + 0.093, + 0.876, + 0.483, + 0.9 + ], + "angle": 0, + "content": "\\(^{12}\\)Due to different scaling of PD-W and PD, we visualize average threat in each relative to maximum value on data." + }, + { + "type": "list", + "bbox": [ + 0.093, + 0.863, + 0.483, + 0.9 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.947, + 0.514, + 0.958 + ], + "angle": 0, + "content": "9961" + } + ], + [ + { + "type": "title", + "bbox": [ + 0.092, + 0.091, + 0.259, + 0.108 + ], + "angle": 0, + "content": "Acknowledgements" + }, + { + "type": "text", + "bbox": [ + 0.091, + 0.114, + 0.486, + 0.143 + ], + "angle": 0, + "content": "This research was supported by NSF CISE 2212457 and NSF CAREER Award CCF 2239787" + }, + { + "type": "title", + "bbox": [ + 0.093, + 0.155, + 0.188, + 0.17 + ], + "angle": 0, + "content": "References" + }, + { + "type": "ref_text", + "bbox": [ + 0.101, + 0.18, + 0.486, + 0.235 + ], + "angle": 0, + "content": "[1] Naveed Akhtar, Ajmal Mian, Navid Kardan, and Mubarak Shah. Advances in adversarial attacks and defenses in computer vision: A survey. IEEE Access, 9:155161-155196, 2021. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.1, + 0.236, + 0.485, + 0.304 + ], + "angle": 0, + "content": "[2] Pranjal Awasthi, Anqi Mao, Mehryar Mohri, and Yutao Zhong. Theoretically grounded loss functions and algorithms for adversarial robustness. In Proceedings of The 26th International Conference on Artificial Intelligence and Statistics, pages 10077-10094. PMLR, 2023. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.305, + 0.484, + 0.346 + ], + "angle": 0, + "content": "[3] Yatong Bai, Brendon G. Anderson, Aerin Kim, and Somayeh Sojoudi. Improving the accuracy-robustness trade-off of classifiers via adaptive smoothing, 2024. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.347, + 0.484, + 0.415 + ], + "angle": 0, + "content": "[4] Brian R. Bartoldson, James Diffenderfer, Konstantinos Parasyris, and Bhavya Kailkhura. Adversarial robustness limits via scaling-law and human-alignment studies. In Proceedings of the 41st International Conference on Machine Learning, pages 3046-3072. PMLR, 2024. 1, 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.416, + 0.484, + 0.497 + ], + "angle": 0, + "content": "[5] Yulong Cao, Chaowei Xiao, Benjamin Cyr, Yimeng Zhou, Wonseok Park, Sara Rampazzi, Qi Alfred Chen, Kevin Fu, and Z. Morley Mao. Adversarial sensor attack on lidar-based perception in autonomous driving. Proceedings of the 2019 ACM SIGSAC Conference on Computer and Communications Security, 2019. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.499, + 0.484, + 0.553 + ], + "angle": 0, + "content": "[6] Nicholas Carlini and David A. Wagner. Adversarial examples are not easily detected: Bypassing ten detection methods. Proceedings of the 10th ACM Workshop on Artificial Intelligence and Security, 2017. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.555, + 0.484, + 0.609 + ], + "angle": 0, + "content": "[7] Francesco Croce and Matthias Hein. Reliable evaluation of adversarial robustness with an ensemble of diverse parameter-free attacks. In Proceedings of the 37th International Conference on Machine Learning. JMLR.org, 2020. 1, 2, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.61, + 0.484, + 0.691 + ], + "angle": 0, + "content": "[8] Francesco Croce, Maksym Andriushchenko, Vikash Sehwag, Edoardo Debenedetti, Nicolas Flammarion, Mung Chiang, Prateek Mittal, and Matthias Hein. Robustbench: a standardized adversarial robustness benchmark. In Thirty-fifth Conference on Neural Information Processing Systems Datasets and Benchmarks Track, 2021. 1, 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.693, + 0.484, + 0.734 + ], + "angle": 0, + "content": "[9] Jiequan Cui, Zhuotao Tian, Zhisheng Zhong, Xiaojuan Qi, Bei Yu, and Hanwang Zhang. Decoupled kullback-leibler divergence loss, 2023. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.735, + 0.483, + 0.762 + ], + "angle": 0, + "content": "[10] Edoardo Debenedetti, Vikash Sehwag, and Prateek Mittal. A light recipe to train robust vision transformers, 2023. 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.763, + 0.484, + 0.831 + ], + "angle": 0, + "content": "[11] Edoardo Debenedetti, Zishen Wan, Maksym Andriushchenko, Vikash Sehwag, Kshitij Bhardwaj, and Bhavya Kailkhura. Scaling compute is not all you need for adversarial robustness. In ICLR 2024 Workshop on Reliable and Responsible Foundation Models, 2024. 1, 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.832, + 0.484, + 0.9 + ], + "angle": 0, + "content": "[12] Samuel Dodge and Lina Karam. A study and comparison of human and deep learning recognition performance under visual distortions. In 2017 26th international conference on computer communication and networks (ICCCN), pages 1-7. IEEE, 2017. 3" + }, + { + "type": "list", + "bbox": [ + 0.094, + 0.18, + 0.486, + 0.9 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.093, + 0.907, + 0.12 + ], + "angle": 0, + "content": "[13] Elvis Dohmatob. Generalized no free lunch theorem for adversarial robustness. In ICML, 2019. 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.122, + 0.908, + 0.162 + ], + "angle": 0, + "content": "[14] Ecenaz Erdemir, Jeffrey Bickford, Luca Melis, and Sergul Aydore. Adversarial robustness with non-uniform perturbations, 2021. 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.165, + 0.907, + 0.233 + ], + "angle": 0, + "content": "[15] Stephanie Fu, Netanel Tamir, Shobhita Sundaram, Lucy Chai, Richard Zhang, Tali Dekel, and Phillip Isola. Dreamsim: Learning new dimensions of human visual similarity using synthetic data. Advances in Neural Information Processing Systems, 36, 2024. 2, 3, 5, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.235, + 0.908, + 0.303 + ], + "angle": 0, + "content": "[16] Robert Geirhos, Carlos R. M. Temme, Jonas Rauber, Heiko H. Schütt, Matthias Bethge, and Felix A. Wichmann. Generalisation in humans and deep neural networks. In Advances in Neural Information Processing Systems. Curran Associates, Inc., 2018. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.305, + 0.907, + 0.36 + ], + "angle": 0, + "content": "[17] Sara Ghazanfari, Alexandre Araujo, Prashanth Krishnamurthy, Farshad Khorrami, and Siddharth Garg. Lipsim: A provably robust perceptual similarity metric. ArXiv, abs/2310.18274, 2023. 2, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.362, + 0.907, + 0.431 + ], + "angle": 0, + "content": "[18] Sara Ghazanfari, Siddharth Garg, Prashanth Krishnamurthy, Farshad Khorrami, and Alexandre Araujo. R-LPIPS: An adversarially robust perceptual similarity metric. In The Second Workshop on New Frontiers in Adversarial Machine Learning, 2023. 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.433, + 0.907, + 0.46 + ], + "angle": 0, + "content": "[19] Abhijay Ghildyal and Feng Liu. Attacking perceptual similarity metrics. Trans. Mach. Learn. Res., 2023, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.461, + 0.906, + 0.503 + ], + "angle": 0, + "content": "[20] Justin Gilmer, Ryan P. Adams, Ian Goodfellow, David Andersen, and George E. Dahl. Motivating the rules of the game for adversarial example research, 2018. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.504, + 0.907, + 0.545 + ], + "angle": 0, + "content": "[21] Teofilo F. Gonzalez. Clustering to minimize the maximum intercluster distance. Theoretical Computer Science, 38:293-306, 1985. 11" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.547, + 0.907, + 0.588 + ], + "angle": 0, + "content": "[22] Ian J. Goodfellow, Jonathon Shlens, and Christian Szegedy. Explaining and harnessing adversarial examples. CoRR, abs/1412.6572, 2015. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.59, + 0.907, + 0.632 + ], + "angle": 0, + "content": "[23] Sven Gowal, Chongli Qin, Jonathan Uesato, Timothy Mann, and Pushmeet Kohli. Uncovering the limits of adversarial training against norm-bounded adversarial examples, 2021. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.633, + 0.906, + 0.674 + ], + "angle": 0, + "content": "[24] Sven Gowal, Sylvestre-Alvise Rebuffi, Olivia Wiles, Florian Stimberg, Dan Andrei Calian, and Timothy Mann. Improving robustness using generated data, 2021. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.676, + 0.907, + 0.731 + ], + "angle": 0, + "content": "[25] Dan Hendrycks and Thomas Dietterich. Benchmarking neural network robustness to common corruptions and perturbations. Proceedings of the International Conference on Learning Representations, 2019. 3, 6, 8, 16" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.732, + 0.907, + 0.787 + ], + "angle": 0, + "content": "[26] Joern-Henrik Jacobsen, Jens Behrmann, Richard Zemel, and Matthias Bethge. Excessive invariance causes adversarial vulnerability. In International Conference on Learning Representations, 2019. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.789, + 0.907, + 0.872 + ], + "angle": 0, + "content": "[27] Can Kanbak, Seyed-Mohsen Moosavi-Dezfooli, and Pascal Frossard. Geometric Robustness of Deep Networks: Analysis and Improvement. In 2018 IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 4441-4449, Los Alamitos, CA, USA, 2018. IEEE Computer Society. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.873, + 0.906, + 0.901 + ], + "angle": 0, + "content": "[28] Oğuzhan Fatih Kar, Teresa Yeo, Andrei Atanov, and Amir Zamir. 3d common corruptions and data augmentation. In" + }, + { + "type": "list", + "bbox": [ + 0.517, + 0.093, + 0.908, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.947, + 0.516, + 0.958 + ], + "angle": 0, + "content": "9962" + } + ], + [ + { + "type": "ref_text", + "bbox": [ + 0.126, + 0.093, + 0.484, + 0.12 + ], + "angle": 0, + "content": "Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 18963-18974, 2022. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.121, + 0.486, + 0.162 + ], + "angle": 0, + "content": "[29] Markus Kettunen, Erik Härkönen, and Jaakko Lehtinen. ELPips: Robust perceptual image similarity via random transformation ensembles, 2019. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.163, + 0.484, + 0.219 + ], + "angle": 0, + "content": "[30] Alexander Kirillov, Eric Mintun, Nikhila Ravi, Hanzi Mao, Chloe Rolland, Laura Gustafson, Tete Xiao, Spencer Whitehead, Alexander C. Berg, Wan-Yen Lo, Piotr Dólar, and Ross Girshick. Segment anything. arXiv:2304.02643, 2023. 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.219, + 0.484, + 0.259 + ], + "angle": 0, + "content": "[31] Alexey Kurakin, Ian J. Goodfellow, and Samy Bengio. Adversarial examples in the physical world. ArXiv, abs/1607.02533, 2017. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.26, + 0.484, + 0.302 + ], + "angle": 0, + "content": "[32] Cassidy Laidlaw, Sahil Singla, and Soheil Feizi. Perceptual adversarial robustness: Defense against unseen threat models, 2021. 2, 5, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.302, + 0.484, + 0.37 + ], + "angle": 0, + "content": "[33] Chang Liu, Yinpeng Dong, Wenzhao Xiang, Xiao Yang, Hang Su, Jun Zhu, Yuefeng Chen, Yuan He, Hui Xue, and Shibao Zheng. A comprehensive study on robustness of image classification models: Benchmarking and rethinking, 2023. 1, 8, 13" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.372, + 0.484, + 0.427 + ], + "angle": 0, + "content": "[34] Aleksander Madry, Aleksandar Makelov, Ludwig Schmidt, Dimitris Tsipras, and Adrian Vladu. Towards deep learning models resistant to adversarial attacks. ArXiv, abs/1706.06083, 2018. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.428, + 0.484, + 0.483 + ], + "angle": 0, + "content": "[35] Rafat K. Mantiuk, Kil Joong Kim, Allan G. Rempel, and Wolfgang Heidrich. Hdr-vdp-2: a calibrated visual metric for visibility and quality predictions in all luminance conditions. ACM SIGGRAPH 2011 papers, 2011. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.484, + 0.484, + 0.552 + ], + "angle": 0, + "content": "[36] Eric Mintun, Alexander Kirillov, and Saining Xie. On interaction between augmentations and corruptions in natural corruption robustness. In Advances in Neural Information Processing Systems, pages 3571-3583. Curran Associates, Inc., 2021. 3, 6, 8, 16" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.553, + 0.484, + 0.621 + ], + "angle": 0, + "content": "[37] Yao Qin, Nicholas Carlini, Garrison Cottrell, Ian Goodfellow, and Colin Raffel. Imperceptible, robust, and targeted adversarial examples for automatic speech recognition. In International conference on machine learning, pages 5231-5240. PMLR, 2019. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.622, + 0.484, + 0.719 + ], + "angle": 0, + "content": "[38] Nikhila Ravi, Valentin Gabeur, Yuan-Ting Hu, Ronghang Hu, Chaitanya Ryali, Tengyu Ma, Haitham Khedr, Roman Rädle, Chloe Rolland, Laura Gustafson, Eric Mintun, Junting Pan, Kalyan Vasudev Alwala, Nicolas Carion, Chao-Yuan Wu, Ross Girshick, Piotr Dólár, and Christoph Feichtenhofer. Sam 2: Segment anything in images and videos. arXiv preprint arXiv:2408.00714, 2024. 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.72, + 0.484, + 0.789 + ], + "angle": 0, + "content": "[39] Ayon Sen, Xiaojin Zhu, Erin Marshall, and Robert Nowak. Popular imperceptibility measures in visual adversarial attacks are far from human perception. In Decision and Game Theory for Security, pages 188-199, Cham, 2020. Springer International Publishing. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.79, + 0.484, + 0.858 + ], + "angle": 0, + "content": "[40] Mahmood Sharif, Lujo Bauer, and Michael K. Reiter. On the suitability of lp-norms for creating and preventing adversarial examples. 2018 IEEE/CVF Conference on Computer Vision and Pattern Recognition Workshops (CVPRW), pages 1686-16868, 2018. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.859, + 0.484, + 0.901 + ], + "angle": 0, + "content": "[41] Naman D Singh, Francesco Croce, and Matthias Hein. Revisiting adversarial training forImagenet: Architectures, training and generalization across threat models, 2023. 8" + }, + { + "type": "list", + "bbox": [ + 0.093, + 0.093, + 0.486, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.093, + 0.908, + 0.147 + ], + "angle": 0, + "content": "[42] Christian Szegedy, Wojciech Zaremba, Ilya Sutskever, Joan Bruna, D. Erhan, Ian J. Goodfellow, and Rob Fergus. Intriguing properties of neural networks. CoRR, abs/1312.6199, 2014. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.149, + 0.908, + 0.232 + ], + "angle": 0, + "content": "[43] Florian Tramér, Jens Behrmann, Nicholas Carlini, Nicolas Papernot, and Jörn-Henrik Jacobsen. Fundamental tradeoffs between invariance and sensitivity to adversarial perturbations. In Proceedings of the 37th International Conference on Machine Learning, ICML 2020, 13-18 July 2020, Virtual Event, pages 9561-9571. PMLR, 2020. 2, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.233, + 0.907, + 0.275 + ], + "angle": 0, + "content": "[44] Florian Tramér, Nicholas Carlini, Wieland Brendel, and Aleksander Madry. On adaptive attacks to adversarial example defenses. ArXiv, abs/2002.08347, 2020. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.277, + 0.908, + 0.346 + ], + "angle": 0, + "content": "[45] James Tu, Huichen Li, Xinchen Yan, Mengye Ren, Yun Chen, Ming Liang, Eilyan Bitar, Ersin Yumer, and Raquel Urtasun. Exploring adversarial robustness of multi-sensor perception systems in self driving. In Proceedings of the 5th Conference on Robot Learning, pages 1013-1024. PMLR, 2022. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.347, + 0.906, + 0.389 + ], + "angle": 0, + "content": "[46] Apostol Vassilev, Alina Oprea, Alie Fordyce, and Hyrum Andersen. Adversarial machine learning: A taxonomy and terminology of attacks and mitigations, 2024. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.39, + 0.908, + 0.444 + ], + "angle": 0, + "content": "[47] Zhou Wang, A.C. Bovik, H.R. Sheikh, and E.P. Simoncelli. Image quality assessment: from error visibility to structural similarity. IEEE Transactions on Image Processing, 13(4): 600-612, 2004. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.446, + 0.906, + 0.488 + ], + "angle": 0, + "content": "[48] Zekai Wang, Tianyu Pang, Chao Du, Min Lin, Weiwei Liu, and Shuicheng Yan. Better diffusion models further improve adversarial training, 2023. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.489, + 0.906, + 0.531 + ], + "angle": 0, + "content": "[49] Zuxuan Wu, Ser-Nam Lim, Larry Davis, and Tom Goldstein. Making an invisibility cloak: Real world adversarial attacks on object detectors, 2020. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.531, + 0.908, + 0.586 + ], + "angle": 0, + "content": "[50] Chaowei Xiao, Jun-Yan Zhu, Bo Li, Warren He, Mingyan Liu, and Dawn Song. Spatially transformed adversarial examples. In International Conference on Learning Representations, 2018. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.588, + 0.908, + 0.643 + ], + "angle": 0, + "content": "[51] Kai Yuanqing Xiao, Logan Engstrom, Andrew Ilyas, and Aleksander Madry. Noise or signal: The role of image backgrounds in object recognition. In International Conference on Learning Representations, 2021. 7, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.645, + 0.908, + 0.714 + ], + "angle": 0, + "content": "[52] Hongyang Zhang, Yaodong Yu, Jiantao Jiao, Eric Xing, Laurent El Ghaoui, and Michael Jordan. Theoretically principled trade-off between robustness and accuracy. In Proceedings of the 36th International Conference on Machine Learning, pages 7472-7482. PMLR, 2019. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.715, + 0.908, + 0.769 + ], + "angle": 0, + "content": "[53] Lin Zhang, Lei Zhang, Xuanqin Mou, and David Zhang. Fsim: A feature similarity index for image quality assessment. IEEE Transactions on Image Processing, 20(8):2378-2386, 2011. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.771, + 0.908, + 0.855 + ], + "angle": 0, + "content": "[54] Richard Zhang, Phillip Isola, Alexei A. Efros, Eli Shechtman, and Oliver Wang. The unreasonable effectiveness of deep features as a perceptual metric. In 2018 IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 586-595, Los Alamitos, CA, USA, 2018. IEEE Computer Society. 3, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.856, + 0.906, + 0.898 + ], + "angle": 0, + "content": "[55] Piotr Želasko, Sonal Joshi, Yiwen Shao, Jesus Villalba, Jan Trmal, Najim Dehak, and Sanjeev Khudanpur. Adversarial attacks and defenses for speech recognition systems, 2021." + }, + { + "type": "list", + "bbox": [ + 0.517, + 0.093, + 0.908, + 0.898 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.947, + 0.516, + 0.958 + ], + "angle": 0, + "content": "9963" + } + ] +] \ No newline at end of file diff --git a/2025/Disentangling Safe and Unsafe Image Corruptions via Anisotropy and Locality/3a52d538-e6a0-40a8-904e-0b81bce10c6a_origin.pdf b/2025/Disentangling Safe and Unsafe Image Corruptions via Anisotropy and Locality/3a52d538-e6a0-40a8-904e-0b81bce10c6a_origin.pdf new file mode 100644 index 0000000000000000000000000000000000000000..a95fbecbd111624fbd42983142ab8dc153f31af8 --- /dev/null +++ b/2025/Disentangling Safe and Unsafe Image Corruptions via Anisotropy and Locality/3a52d538-e6a0-40a8-904e-0b81bce10c6a_origin.pdf @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:3e95ea44725fa754858658cd524b9b12af7ecf8b44fbb658ee8a803b560be054 +size 4882233 diff --git a/2025/Disentangling Safe and Unsafe Image Corruptions via Anisotropy and Locality/full.md b/2025/Disentangling Safe and Unsafe Image Corruptions via Anisotropy and Locality/full.md new file mode 100644 index 0000000000000000000000000000000000000000..e6391c94f4ff8fe2e7cb151961b16e7d6eaf62ac --- /dev/null +++ b/2025/Disentangling Safe and Unsafe Image Corruptions via Anisotropy and Locality/full.md @@ -0,0 +1,344 @@ +# Disentangling Safe and Unsafe Image Corruptions via Anisotropy and Locality + +Ramchandran Muthukumar*¹ Ambar Pal²† Jeremias Sulam¹ René Vidal³ +¹Johns Hopkins University ²Amazon Web Services ³University of Pennsylvania + +# Abstract + +State-of-the-art machine learning systems are vulnerable to small perturbations to their input, where "small" is defined according to a threat model that assigns a positive threat to each perturbation. Most prior works define a task-agnostic, isotropic, and global threat, like the $\ell_p$ norm, where the magnitude of the perturbation fully determines the degree of the threat and neither the direction of the attack nor its position in space matter. However, common corruptions in computer vision, such as blur, compression, or occlusions, are not well captured by such threat models. This paper proposes a novel threat model called Projected Displacement (PD) to study robustness beyond existing isotropic and global threat models. The proposed threat model measures the threat of a perturbation via its alignment with unsafe directions, defined as directions in the input space along which a perturbation of sufficient magnitude changes the ground truth class label. Unsafe directions are identified locally for each input based on observed training data. In this way, the PD-threat model exhibits anisotropy and locality. Experiments on Imagenet-1k data indicate that, for any input, the set of perturbations with small PD threat includes safe perturbations of large $\ell_p$ norm that preserve the true label, such as noise, blur and compression, while simultaneously excluding unsafe perturbations that alter the true label. Unlike perceptual threat models based on embeddings of large-vision models, the PD-threat model can be readily computed for arbitrary classification tasks without pre-training or finetuning. Further additional task information such as sensitivity to image regions or concept hierarchies can be easily integrated into the assessment of threat and thus the PD threat model presents practitioners with a flexible, task-driven threat specification that alleviates the limitations of $\ell_p$ -threat models. + +# 1. Introduction + +Modern machine learning (ML) systems are nearing widespread deployment in civilian life, and hence a com + +prehensive understanding of their security vulnerabilities is necessary [46]. One such vulnerability concerns the ability of a malicious adversary to tamper with predictions by adding small imperceptible corruptions referred to as adversarial perturbations [22, 42]. There is by now overwhelming evidence that carefully crafted adversarial perturbations can foil the prediction of state-of-the-art machine learning classifiers, i.e. they are not adversarily robust [6, 8]. Such vulnerabilities have been observed in a wide variety of applications like computer vision [1, 49], speech recognition [37, 55], autonomous driving [5, 45], and more. The goal of designing safe and reliable ML systems remains incomplete [4, 11] despite significant investment [23, 24, 33]. Often, strategies that intend to foil performance, i.e. adversarial attacks [6, 7, 22, 31, 44], have proven more successful than strategies aimed at mitigating vulnerabilities, i.e. adversarial defenses [3, 9, 34, 48, 52]. + +Meaningfully evaluating the adversarial robustness of a machine learning system requires a formal specification of a threat function, and an associated threat model that limits the scope of malicious adversaries [20]. Informally, a threat function $d: \mathbb{R}^d \times \mathbb{R}^d \to \mathbb{R}^{\geq 0}$ measures the threat represented by a corruption $\delta \in \mathbb{R}^d$ towards altering the true label at input $\mathbf{x} \in \mathbb{R}^d$ as $d(\mathbf{x}, \delta)$ . We let $S(\mathbf{x}, d, \varepsilon) \coloneqq \{\delta \in \mathbb{R}^d \mid d(\mathbf{x}, \delta) \leq \varepsilon\}$ denote the $\varepsilon$ -sublevel set of perturbations where the threat $d$ is measured w.r.t. input $\mathbf{x}$ . A threat model $(d, \varepsilon)$ is a pair of a threat function $d$ and a permissible threshold $\varepsilon$ that together define the set of permissible perturbations $S(\mathbf{x}, d, \varepsilon)$ . Under the threat model $(d, \varepsilon)$ , the robust accuracy of a classifier $h \in \mathcal{H}$ is the probability that the label prediction at $\mathbf{x}$ is locally invariant to corruptions within the permissible set $S(\mathbf{x}, d, \varepsilon)$ , see Definition 2 for an explicit definition. + +One of the most commonly used threat models is the $\ell_p$ -threat model $(d_p, \epsilon)$ , corresponding to the choice $d_p(\mathbf{x}, \delta) \coloneqq \| \delta \|_p$ . The threat function $d_p$ is task-agnostic, easy to evaluate, and induces a compact sub-level set $S(\mathbf{x}, d_p, \varepsilon)$ that allows efficient projection. Hence, the $\ell_p$ -threat model $(d_p, \varepsilon)$ presents a natural starting point for in + +![](images/c3652cb8b0020bc5a6e618359cc909da7e2e65b34bf021d70c7774d12e0b2112.jpg) +Figure 1. Corruptions with equal $\ell_{\infty}$ -threat, $\|\delta_1\|_{\infty} = \|\delta_2\|_{\infty} = \|\delta_3\|_{\infty} = \|\delta_4\|_{\infty}$ , but varying PD-threat. + +vestigating robustness. RobustBench [8] maintains an up-to-date leaderboard of the robust accuracies of benchmark models under $(d_p, \varepsilon)$ . Unfortunately, the progress towards achieving perfect adversarial robustness (i.e. $100\%$ robust test accuracy) in RobustBench has plateaued in recent years. Unlike supervised learning on clean data, scaling data, model size and computing resources might be insufficient to bridge the gap [4, 11]. Below we expand further on the fundamental limitations of isotropic and global threat models. + +# 1.1. Motivation: Specification of Threat Model + +For image-based datasets, it is widely recognized that $\ell_p$ norms are neither necessary nor sufficient for capturing perceptual similarity [39, 40], which poses significant challenges for accurately evaluating robustness [26, 43]. In Figure 1, $\mathbf{x}$ (at the center) is an image of class ROSE HIP from the Imagenet-1k dataset, and $\mathbf{x} + \delta_1$ , $\mathbf{x} + \delta_2$ are two corrupted images equidistant (w.r.t. $\ell_{\infty}$ ) from $\mathbf{x}$ , i.e. $\| \delta_1 \|_{\infty} = \| \delta_2 \|_{\infty} = \varepsilon$ . Intuitively, both $\mathbf{x} + \delta_1$ and $\mathbf{x} + \delta_2$ share the same label as $\mathbf{x}$ , since its appearance still depicts the fruit ROSE HIP. We refer to such perturbations that preserve the class label as safe, while those that change it are termed unsafe. Consider now moving further along these directions to $\mathbf{x} + t\delta_1$ and $\mathbf{x} + t\delta_2$ , for $t > 1$ . These perturbed points are again equidistant from $\mathbf{x}$ , but while $t\delta_1$ does not alter the true label (i.e. it is safe), $t\delta_2$ does (and it is unsafe). As a result, any threat model $(d_{\infty}, t' \varepsilon)$ with $t' \geq t$ necessarily incurs misspecification: any classifier $h$ stable to all perturbations in $S(\mathbf{x}, d_{\infty}, t' \varepsilon)$ produces an incorrect prediction at $\mathbf{x} + t\delta_2$ , i.e. $h(\mathbf{x} + t\delta_2) = \text{ROSE HIP}$ . On the other hand, shrinking the threat model is also futile, as then the safe perturbation $t\delta_1$ is necessarily excluded. + +This illustrates that standard $\ell_p$ -threat models are unable to distinguish between the safe perturbations that preserve the true label and the unsafe perturbations that alter the true label. Hence, perfect robust accuracy under $\ell_p$ threat models might be neither achievable nor desirable. Attempting to + +resolve these limitations, Laidlaw et al. [32] formulate perceptual threat models via neural perceptual metric based on neural representations. Unfortunately (but unsurprisingly), all neural perceptual distance metrics are themselves vulnerable to adversarial attacks [17-19, 29]. + +# 1.2. Summary of Contributions + +We propose a novel threat function called Projected Displacement (PD), denoted by $d_{\mathrm{PD}}$ (see Definition 7 for an explicit description), that accounts for the local directional statistics of the data via unsafe directions3 $\mathcal{U}(\mathbf{x})$ (Definition 6) at each input $\mathbf{x}$ . For an input $\mathbf{x}$ with label $y$ , any input $\tilde{\mathbf{x}}$ with label $\tilde{y} \neq y$ represents the unsafe direction $\mathbf{u} := \frac{\tilde{\mathbf{x}} - \mathbf{x}}{\|\tilde{\mathbf{x}} - \mathbf{x}\|_2}$ , which will be estimated from observed training data. Then, the degree of threat of each perturbation $\delta$ is the maximal alignment with the set of unsafe directions: A large threat $d_{\mathrm{PD}}(\mathbf{x},\delta)$ indicates that $\delta$ is well-aligned with an unsafe direction $\mathbf{u} \in \mathcal{U}(\mathbf{x})$ . We highlight some key properties. + +1. Disentangling safe vs unsafe perturbations. PD-threat is able to distinguish between safe and unsafe perturbations of equal $\ell_p$ norms. From Figure 1, different corruptions (Gaussian noise $t\delta_1$ , motion blur $t\delta_3$ and saturation $t\delta_4$ ) applied to the original image $\mathbf{x}$ all have equal $\ell_{\infty}$ norm of 240/255. Despite this, the true label of these perturbed samples does not change, unlike that of corruption $t\delta_2$ , which results in an image $\mathbf{x} + t\delta_2$ with a different class. For each perturbation $t\delta_i$ , the computed values of PD threat in Figure 1 naturally reflect how close they are to changing the class label. PD threat model is competent with state-of-the-art neural perceptual threat model DreamSim (denoted $d_{\mathrm{DS}}$ ) [15] for distinguishing safe and unsafe perturbations (see Section 4). +2. Ease of use. As we will show, the sub-level set $S(\mathbf{x}, d_{\mathrm{PD}}, \varepsilon)$ at each input $\mathbf{x}$ is convex, allowing for efficient projections onto it. This enables a straightforward plug-and-play mechanism for adapting existing adversarial attacks to the PD-threat model. We incorporate such an extension of AutoAttack [7] to evaluate the robustness under the PD-threat model ( $d_{\mathrm{PD}}, \varepsilon$ ) for state-of-the-art robust classifiers registered at RobustBench [8]. +3. Task-dependency. PD-threat models can readily integrate additional task-relevant information via label annotation, such as inter-class hierarchy (e.g. the Wordnet hierarchy) or pixel annotation (e.g. segmentation masks)—see Table 1—allowing for variations in the form of robustness while retaining the above benefits (anisotropy, locality, convex sub-level sets, and efficient projection). + +# 1.3. Related Work + +In this section, we highlight prior work that explores robustness beyond $\ell_p$ -threat models. + +
PROPERTYFEATURESdpdDS[15]dPD
Degree of ThreatAnisotropy, Locality×✓✓
Label AnnotationInter-class Distance×~✓
Pixel AnnotationSegmentation✓×✓
+ +Table 1. Comparison of Threat Models + +Common Corruptions. Neural networks are vulnerable to image distortions (e.g. translation, rotation), common corruptions (e.g. blur, noise), changes in lighting, view, depth, etc. [12, 16, 25, 27, 28, 36, 50]. Note that these are known safe perturbations as they preserve the true label, even if they have large $\ell_p$ norms. To measure this vulnerability, [25] formalized the notion of corruption robustness and presented a standardized benchmark dataset, Imagenet-C containing 19 styles of common image corruptions (categorized into noise, blur, weather, and digital). Training on Imagenet-C, or on other augmentations, can result in over-fitting, particularly to specific types of corruptions [36]. Additionally, [36] introduced a newdataset, Imagenet-C, with 10 new corruption styles identified from a larger set of real-world corruptions. These will be useful later in our experiments. + +Perceptual Distance Metrics. Since $\ell_p$ provides inaccurate approximations to perceptual distances, several works have attempted to propose alternatives, such as SSIM [47], FSIM [53], and HDR-VDP-2 [35]. With the advent of learning data-driven representations, [15, 54] hail the effectiveness of internal representations of neural classifiers in capturing perceptual similarity. DreamSim [15], the current state-of-the-art among the neural perceptual distance metrics, is finetuned on human perceptual similarity judgements. + +# 2. Key Definitions + +Notation. We denote scalar quantities in Roman letters and vectors in boldface Roman letters. The empty set is given by $\emptyset$ . For any set $\mathcal{V}$ , we denote by $2^{\mathcal{V}}$ the power set of all its subsets. For any set $\mathcal{V} \subset \mathbb{R}^d$ and a vector $\mathbf{v} \in \mathbb{R}^d$ , $\mathcal{V} - \{\mathbf{v}\}$ is the Minkowski difference, i.e. the set $\{\mathbf{w} - \mathbf{v} \mid \forall \mathbf{w} \in \mathcal{V}\}$ (subtracting $\mathbf{v}$ from all elements of $\mathcal{V}$ ). We denote the unit ball and the unit sphere in $\mathbb{R}^d$ w.r.t. the $\ell_p$ norm by $\mathbb{B}_p^d := \{\mathbf{v} \in \mathbb{R}^d \mid \| \mathbf{v} \|_p \leq 1\}$ and $\mathbb{S}_p^d := \{\mathbf{v} \in \mathbb{R}^d \mid \| \mathbf{v} \|_p = 1\}$ , respectively. A direction is an element of the unit sphere $\mathbb{S}_2^d$ . + +# 2.1. Supervised Learning + +We consider a supervised learning task with a bounded input domain $\mathcal{X} \subset \mathbb{R}^d$ and labels $\mathcal{Y} = \{1, \dots, C\}$ such that each input $\mathbf{x}$ is assigned a label by a deterministic true $^4$ labeling function $h^\star : \mathcal{X} \to \mathcal{Y}$ . The true labeling function $h^\star$ partitions the input domain $\mathcal{X} := \cup_y \mathcal{X}_y$ , where $\mathcal{X}_y$ is the set of inputs assigned label $y$ , i.e. $\mathcal{X}_y = \{\mathbf{x} \in \mathcal{X} \mid h^\star(\mathbf{x}) = y\}$ . + +We assume that the partition sets $\mathcal{X}_1, \ldots, \mathcal{X}_C$ are nonempty and open in order to avoid degeneracies. + +We let $\mathcal{D}_{\mathcal{X}}$ be a data distribution over the input domain $\mathcal{X}$ and $\mathcal{D}_{\mathcal{Z}}$ the extension of $\mathcal{D}_{\mathcal{X}}$ to a distribution over the joint domain $\mathcal{Z} \coloneqq \mathcal{X} \times \mathcal{Y}$ using the true labeling function $h^{\star}$ . The true labeling function $h^{\star}$ and the marginal input distribution $\mathcal{D}_{\mathcal{X}}$ are unknown, but we observe finite labeled training data $\mathrm{S} \coloneqq \left\{(\mathbf{x}^{(1)}, y^{(1)}), \ldots, (\mathbf{x}^{(m)}, y^{(m)})\right\} \stackrel{\mathrm{i.i.d.}}{\sim} (\mathcal{D}_{\mathcal{Z}})^{m}$ . For any classifier $h: \mathcal{X} \to \mathcal{Y}$ , accuracy is the probability of correctly labeling a random input, i.e. $\mathrm{Acc}(h) \coloneqq \mathrm{Prob}_{\mathbf{x} \sim \mathcal{D}_{\mathcal{X}}} [h(\mathbf{x}) = h^{\star}(\mathbf{x})]$ . The task of supervised learning seeks a classifier $\bar{h}$ in a hypothesis class $\mathcal{H} \subset \{h: \mathcal{X} \to \mathcal{Y}\}$ with high accuracy. + +# 2.2. Adversarial Perturbations + +In this article, we formally define adversarial perturbations through the lens of the partition sets $\mathcal{X}_y$ . + +Definition 1 (Adversarial Perturbation). A perturbation $\delta \in \mathbb{R}^d$ is adversarial for predictor $h$ at input $\mathbf{x}$ if it is, + +1. Domain constrained: $\mathbf{x} + \delta \in \mathcal{X}$ , i.e. $\delta \in \mathcal{X} - \{\mathbf{x}\}$ . +2. Label invariant: $\mathbf{x} + \boldsymbol{\delta}$ has the same true label as $\mathbf{x}$ , i.e. $h^{\star}(\mathbf{x} + \boldsymbol{\delta}) = h^{\star}(\mathbf{x})$ . +3. Adversarial: $\mathbf{x} + \boldsymbol{\delta}$ is misclassified by $h$ , i.e. $h(\mathbf{x} + \boldsymbol{\delta}) \neq h^{\star}(\mathbf{x} + \boldsymbol{\delta})$ . + +The labels assigned by $h^{\star}$ are considered the ground truth and thus $h^{\star}$ has no adversarial perturbation at any input. The presence of adversarial perturbations indicates an imperfect labeling function $h$ . Based on Definition 1, perturbations $\delta \in \mathbb{R}^{d}$ such that $\mathbf{x} + \delta \notin \mathcal{X}$ are not deemed adversarial as the classifier $h$ is not required to predict labels outside the input domain. On the other hand, for $\mathbf{x}$ with true label $y$ , a perturbation $\delta$ such that $h^{\star}(\mathbf{x} + \delta) \neq y$ cannot be adversarial at $\mathbf{x}$ as the corrupted input $\mathbf{x} + \delta$ has a different true label. Finally, if $h$ predicts the label at $\mathbf{x}$ incorrectly, then $\mathbf{0}$ is already an adversarial perturbation. + +# 2.2.1. Adversarial Robustness via Threat Specification + +A threat function $d: \mathcal{X} \times \mathbb{R}^d \to \mathbb{R}^{\geq 0}$ presents practitioners a way to specify the kind of robustness suitable for a particular learning task. The $\varepsilon$ -sublevel set of a threat function $d$ at an input $\mathbf{x}$ , $S(\mathbf{x}, d, \varepsilon) := \{\delta \in \mathbb{R}^d \mid d(\mathbf{x}, \delta) \leq \varepsilon\}$ , defines the set of perturbations with a threat at most $\varepsilon$ . + +Definition 2 (Threat Model, Robust Accuracy). A threat model $(d,\varepsilon)$ is a pair of threat function $d$ and permissible threshold $\varepsilon$ that together define the set of permissible perturbations at each input $\mathbf{x}$ as the sublevel set $S(\mathbf{x},d,\varepsilon)$ . Under the threat model $(d,\varepsilon)$ , the robust accuracy RobustAcc $(h,(d,\varepsilon))$ of a classifier $h\in \mathcal{H}$ is the probability of stable prediction upon corruptions within the permissible set, i.e. $\mathrm{Prob}_{\mathbf{x}}\big[\forall \delta \in S(\mathbf{x},d,\varepsilon), h(\mathbf{x} + \delta) = h^{\star}(\mathbf{x})\big]$ . + +![](images/729f79c773047969e68a073fc968d08f975b0727244bac0404b0de377195067e.jpg) +Figure 2. An illustration of unsafe directions, and sub-level sets of the PD-threat. + +Thus we define robustness as the stability of a classifier's prediction under perturbations within the permissible sets. The task of robust supervised learning seeks a classifier in $\mathcal{H}$ with high robust accuracy. + +Definition 3 ( $\varepsilon$ -robust). A classifier $h \in \mathcal{H}$ is $\varepsilon$ -robust at $\mathbf{x}$ w.r.t. a threat function $d$ if, $\forall \delta \in S(\mathbf{x}, d, \varepsilon) \cap (\mathcal{X} - \{\mathbf{x}\})$ , we have $h(\mathbf{x} + \delta) = h^{\star}(\mathbf{x})$ . + +An $\varepsilon$ -robust classifier predicts the same label $h^{\star}(\mathbf{x})$ at $\mathbf{x} + \delta$ for any corruption $\delta$ within the $\varepsilon$ -sublevel set $S(\mathbf{x}, d, \varepsilon)$ . We emphasize that robust accuracy and $\varepsilon$ -robustness are measures of stability rather than correctness. In this way, each threat model $(d, \varepsilon)$ implicitly encodes a trade-off between stability and correctness. [13]. Requiring more stability than necessary can lead to incorrect predictions as in Figure 1. + +# 3. Threat specification beyond $\ell_p$ norms + +In this section, we design a principled threat model from first principles that is local and anisotropic. We first theoretically develop our exact PD\*-threat model assuming knowledge of the decision boundaries of the true labeling function $h^\star$ (Section 3.1), and later relax this assumption by developing an approximation leveraging the observed training data and obtaining a practical PD-threat (Section 3.2). + +# 3.1. Measuring threat with class partition + +Definition 4 (Unsafe Directions). At each input $\mathbf{x}$ , a direction $\mathbf{u} \in \mathbb{S}_2^d$ is called unsafe if there exists a step size $t \geq 0$ such that $\mathbf{x} + t\mathbf{u} \in \mathcal{X}$ and $h^\star(\mathbf{x} + t\mathbf{u}) \neq h^\star(\mathbf{x})$ . We denote the set of all unsafe directions at $\mathbf{x}$ as $\mathcal{U}^\star(\mathbf{x})$ . + +The minimum step size needed to alter the true label can vary across unsafe directions. Further, $\mathcal{U}^{\star}(\mathbf{x})$ varies + +with input. Directions $\mathbf{v} \in \mathcal{B}_2^d \cap (\mathcal{U}^\star(\mathbf{x}))^c$ are called safe, since moving along them (whilst domain constrained) cannot alter the true label. We note that adversarial perturbations (Definition 1) are a characteristic of a learned predictor $h \in \mathcal{H}$ while unsafe directions are a characteristic of the true labeling function $h^\star$ . We propose to measure threat of a perturbations $\delta$ using $\mathcal{U}^\star(\mathbf{x})$ . + +Definition 5 (PD\* -threat). Let $\mathbf{x}\in \mathcal{X}$ and let $\delta \in \mathbb{R}^d$ be a perturbation. The exact projected displacement threat function $d_{PD}^{\star}$ is defined as the maximum scaled displacement of the Euclidean projection of the perturbation $\delta$ over all unsafe directions, + +$$ +d _ {\mathrm {P D}} ^ {\star} (\mathbf {x}, \boldsymbol {\delta}) := \sup _ {\mathbf {u} \in \mathcal {U} ^ {\star} (\mathbf {x})} \frac {1}{g ^ {\star} (\mathbf {x} , \mathbf {u})} \max \left(\langle \boldsymbol {\delta}, \mathbf {u} \rangle , 0\right) \quad (1) +$$ + +where $g^{\star}(\mathbf{u},\mathbf{x})$ is called the normalization function, + +$$ +g ^ {\star} (\mathbf {x}, \mathbf {u}) := \sup _ {\mathbb {R} ^ {\geq 0}} M \text {s . t .} \mathbf {x} + t \mathbf {u} \in \mathcal {X} _ {h ^ {\star} (\mathbf {x})} \forall t \in [ 0, M ]. +$$ + +At each $\mathbf{x}$ , the normalization function $g^{*}(\mathbf{x},\mathbf{u})$ captures the distance along the direction $\mathbf{u}$ to the boundary of the set $\mathcal{X}_{h^{\star}(\mathbf{x})}$ and enables a comparison of the threat along different unsafe directions (see Figure 2). In (1), the supremum ensures that we do not underestimate the threat along a direction that potentially alters the true label. We highlight two key characteristics. + +Anisotropy. The threat function $d_{\mathrm{PD}}^{\star}$ is anisotropic (the threat of a perturbation $\delta$ at input $\mathbf{x}$ depends on both the direction and magnitude of the perturbation) and does not correspond to a norm in general, unlike $\ell_p$ threat models. For a fixed $\mathbf{x}$ , the threat along different directions $\mathbf{u}$ varies based on the alignment with unsafe directions $\mathcal{U}^{\star}(\mathbf{x})$ . In particular, $d_{\mathrm{PD}}^{\star}$ can differentiate between the perturbations $\delta$ and $-\delta$ . + +Locality. The threat function $d_{\mathrm{PD}}^{\star}$ exhibits locality; i.e. a fixed perturbation $\delta$ has varying threat $d_{\mathrm{PD}}^{\star}(\mathbf{x}, \delta)$ depending on the input $\mathbf{x}$ . Thus the permissible set varies with input7, enabling an input-aware assessment of threat. We note that weighted $\ell_p$ norms [14] can be anisotropic but not local. + +# 3.1.1. Stability vs Correctness Trade-off + +Since the normalization function $g^{\star}$ accounts for the class partition $\{\mathcal{X}_c\}_{c=1}^C$ , the threat function $d_{\mathrm{PD}}^{\star}$ has a principled trade-off between stability and correctness. Specifically, the true labeling function $h^{\star}$ exhibits a high level of robustness under the PD $^{\star}$ -threat8. + +Theorem 1. The true labeling function $h^{\star}$ is 1-robust at any input $\mathbf{x} \in \mathcal{X}$ w.r.t. the threat $d_{\mathrm{PD}}^{\star}$ . Additionally, if a classifier $h$ is not 1-robust at all inputs then there exists an input $\mathbf{x}$ misclassified by $h$ . + +In other words, for any pairs of inputs $\mathbf{x},\tilde{\mathbf{x}}\in \mathcal{X}$ with different labels, the threat $d_{\mathrm{PD}}^{\star}(\mathbf{x},\tilde{\mathbf{x}} -\mathbf{x})$ is always larger than 1. Thus seeking $h\in \mathcal{H}$ that is 1-robust at any input under $d_{\mathrm{PD}}^{\star}$ is not at odds with learning $h^{\star}$ . + +# 3.2. Measuring threats with observed data + +Computing $d_{\mathrm{PD}}^{\star}(\mathbf{x},\cdot)$ from Definition 5 requires explicitly characterizing the set of unsafe directions $\mathcal{U}^{\star}(\mathbf{x})$ and the normalization function $g^{\star}(\mathbf{u},\mathbf{x})$ , both of which need knowledge of the partition sets $\{\mathcal{X}_c\}_{c = 1}^C$ induced by $h^\star$ . In this subsection, we propose a heuristic empirical approximation. + +Approximating $\mathcal{U}^{\star}(\mathbf{x})$ . We have access to training data S and incomplete information on the partition sets $\{\mathcal{X}_c\}_{c=1}^C$ via the partition $S = \{S_c\}_{c=1}^C$ , where $S_c := \{\mathbf{x} \in \mathcal{X} \mid (\mathbf{x}, c) \in S\} \subset \mathcal{X}_c$ . At every training input $\mathbf{x} \in S_y$ for each input $\tilde{\mathbf{x}} \in S_c$ (for $c \neq y$ ), the direction $\mathbf{u} = \frac{\tilde{\mathbf{x}} - \mathbf{x}}{\|\tilde{\mathbf{x}} - \mathbf{x}\|_2}$ is an unsafe direction. For computational considerations, we choose a collection $\bigcup_{c} S_{c,k}$ of representative subsets $S_{c,k} \subset S_c$ of size $k$ . As a heuristic choice, we select subsets $S_{c,k}$ by solving a discrete-k-center optimization problem (detailed in Sec. 9.1). We are now ready to define an observed subset of the unsafe directions based on $S_{c,k}$ . + +Definition 6 (Observed Unsafe Directions). For each $\mathbf{x} \in S_y$ , we identify a set of $(C - 1)k$ -observed unsafe directions based on the representative subsets $S_{c,k}$ , + +$$ +\mathcal {U} _ {k} (\mathbf {x}) := \left\{\frac {\tilde {\mathbf {x}} - \mathbf {x}}{\| \tilde {\mathbf {x}} - \mathbf {x} \| _ {2}} \Big | \tilde {\mathbf {x}} \in \cup_ {c \neq y} \mathbb {S} _ {c, k} \right\} \subset \mathcal {U} ^ {\star} (\mathbf {x}). +$$ + +Approximating $g^{\star}(\mathbf{x},\mathbf{u})$ . At any $\mathbf{x}\in \mathcal{X}_y$ , for each $\mathbf{u}\in \mathcal{U}_k(\mathbf{x})$ , there exists $\tilde{\mathbf{x}}\in \cup_{c\neq y}\mathrm{S}_{c,k}$ such that $\mathbf{u} = \frac{\tilde{\mathbf{x}} - \mathbf{x}}{\|\tilde{\mathbf{x}} - \mathbf{x}\|_2}$ . normalization function can be bounded as $g^{\star}(\mathbf{x},\mathbf{u})\leq \| \tilde{\mathbf{x}} -\mathbf{x}\| _2$ , since $h^\star (\tilde{\mathbf{x}})\neq h^\star (\mathbf{x})$ . In practice, we use the heuristic $g_{\beta}(\mathbf{u},\mathbf{x})\coloneqq \beta \| \tilde{\mathbf{x}} -\mathbf{x}\| _2$ , where $\beta \in (0,1)$ is a scaling hyper-parameter. + +Definition 7 $((k,\beta)$ -PD threat). Let $\mathbf{x}\in \mathcal{X}$ and let $\delta \in \mathbb{R}^d$ be a perturbation. The projected displacement threat is defined as + +$$ +d _ {\mathrm {P D}, k, \beta} (\mathbf {x}, \boldsymbol {\delta}) := \max _ {\mathbf {u} \in \mathcal {U} _ {k} (\mathbf {x})} \frac {1}{g _ {\beta} (\mathbf {x} , \mathbf {u})} \max \left(\langle \boldsymbol {\delta}, \mathbf {u} \rangle , 0\right). +$$ + +The quality of the approximation of the exact threat $d_{\mathrm{PD}}^{\star}$ by $d_{\mathrm{PD,k},\beta}$ depends on the choice of representative unsafe directions $\mathcal{U}_k(\cdot)$ via the $k$ -subset $\mathbb{S}_{c,k}$ and on the heuristic choice of the approximate normalization $g_{\beta}$ , via the scaling hyper-parameter $\beta$ . In this article, we fix $(k,\beta) = (50,\frac{1}{2})$ as the default hyper-parameters (see Section 9.3 for an expanded discussion on choosing hyper-parameters $(k,\beta)$ for any learning task). For brevity, we refer to the observed unsafe directions (Definition 6), the heuristic normalizations and the PD threat (Definition 7) as $\mathcal{U}(\mathbf{x}), g(\mathbf{x},\delta)$ and $d_{\mathrm{PD}}(\mathbf{x},\delta)$ , respectively. + +We now state a few important properties of our proposed empirical approximation $d_{\mathrm{PD}}$ . Clearly $d_{\mathrm{PD}}$ inherits the anisotropy and locality properties of the exact threat $d_{\mathrm{PD}}^{\star}$ . + +Growth and Sensitivity. The growth of PD threat along any direction is linear, i.e., $d_{\mathrm{PD}}(\mathbf{x},t\boldsymbol {\delta}) = td_{\mathrm{PD}}(\mathbf{x},\boldsymbol {\delta})$ , and the rate of growth scales inversely with the (approximate) normalization. In particular, for a fixed $\mathbf{x}$ $d_{\mathrm{PD}}(\mathbf{x},\cdot)$ is $(\max_{u\in \mathcal{U}(\mathbf{x})}1 / g(\mathbf{x},\mathbf{u}))$ -Lipschitz w.r.t $\ell_2$ norm. + +Projection onto sublevel sets. For any threat model $(d,\varepsilon)$ , the ability to efficiently project a perturbation $\delta$ onto the sub-level sets $S(d,\mathbf{x},\varepsilon)$ enables one to leverage a wide literature on gradient-based attacks to conduct a rigorous evaluations of adversarial robustness. The sublevel sets of PD threat model, i.e., $S(\mathbf{x},d,\varepsilon)$ , are convex for all $\mathbf{x}$ for all $\epsilon \geq 0$ . In particular, $\mathcal{S}$ is the intersection of at most $k\cdot C$ half-spaces characterized by unsafe directions $\mathcal{U}(\mathbf{x})$ as $S(\mathbf{x},d_{\mathrm{PD}},\varepsilon) = \bigcap_{\mathbf{u}\in \mathcal{U}(\mathbf{x})}\{\pmb {\delta}\in \mathbb{R}^d\mid \langle \pmb {\delta},\mathbf{u}\rangle \leq \varepsilon \cdot g(\mathbf{x},\mathbf{u})\}$ . + +Further, the sub-level sets are monotonic w.r.t. the threshold, i.e., for $\varepsilon_1\geq \varepsilon_2\geq 0,S(\mathbf{x},d_{\mathrm{PD}},\varepsilon_2)\subseteq S(\mathbf{x},d_{\mathrm{PD}},\varepsilon_1)$ + +Linearity of growth and convexity of the sub-level sets together provide a straightforward approximate projection algorithm, since for any perturbation $\delta$ , the scaled perturbation $\frac{\varepsilon}{d_{\mathrm{PD}}(\mathbf{x},\delta)}\delta$ lies in the permissible set $S(\mathbf{x},d_{\mathrm{PD}},\varepsilon)$ . Further, due to the convexity of sublevel sets, an exact projection (in $\ell_2$ distance) can be computed using the iterative greedy scheme outlined in the Algorithm 2 (deferred to the appendix). Thus the PD threat provides the practitioner with the flexibility of a fast approximate projection or an exact projection at the expense of increased computation. Together, these projection algorithms enable immediate adaptation of existing adversarial attacks to our proposed threat model. + +We note that efficient exact projection algorithms are also known for $\ell_p$ threat models. However, for neural perceptual threat models such as DreamSim [15], the sub-level sets are non-convex and a projection can only be approximately computed [32] and is typically not computationally efficient. + +# 4. Experiments on Real World Data + +In this section we illustrate the qualitative and quantitative characteristics of the PD-threat in comparison to existing threat models. We evaluate 4 threat functions: $d_{\infty}, d_{2}, d_{\mathrm{DS}}$ , and our proposed PD threat function $d_{\mathrm{PD}}$ , where $d_{\mathrm{DS}}$ is the neural perceptual threat function $d_{\mathrm{DS}}(\mathbf{x}, \tilde{\mathbf{x}} - \mathbf{x}) \coloneqq \mathrm{DreamSim}(\mathbf{x}, \tilde{\mathbf{x}})$ from DreamSim [15]. Note that we focus on DreamSim due to its superior performance over other neural perceptual threat models [15], but the following comparison applies to any modern neural perceptual threat model. + +Experimental Setup. We seek to measure the ability of threat functions to distinguish between safe and unsafe corruptions. The threat models $d_{\mathrm{PD}}$ and $d_{\mathrm{DS}}$ depend on the Imagenet-1k training dataset S. All threat models are evaluated on the Imagenet-1k validation dataset $\mathbf{S}_{\mathrm{val}}$ . For any pair + +
CATEGORYCORRUPTION STYLES
NoiseGaussian, Shot, Impulse, Speckle, Blue, Brown, Perlin, Single-Frequency, Plasma
BlurGaussian, Defocus, Glass, Motion, Concentric Sine Waves, Caustic Refraction
CompressionPixelate, JPEG
DigitalBrightness, Contrast, Saturate, Elastic Transform
WeatherFrost, Fog, Snow
OcclusionSpatter, Checkerboard Cutout, Sparkles, Inverse Sparkles
+ +Table 2. Categories of corruptions studied, from Imagenet-C [25] and Imagenet- $\tilde{\mathbf{C}}$ [36]. + +of labelled inputs $(\mathbf{x},y),(\tilde{\mathbf{x}},c)\in S_{\mathrm{val}}$ with distinct labels $(c\neq y)$ , the perturbation $\delta \coloneqq \tilde{\mathbf{x}} -\mathbf{x}$ is an unsafe perturbation at $\mathbf{x}$ . This way, we can evaluate the statistics of threat under unsafe perturbations for each threat function. To characterize threat under safe perturbations, we consider a set $\Omega$ of 150 distinct common corruptions consisting of 30 styles $\times 5$ severity levels, i.e., severity $\in \{1,2,3,4,5\}$ - the supplementary material contains visual examples (see Figures 8 and 9), sourced from Imagenet- $C$ [25] and Imagenet- $\bar{\mathbf{C}}$ [36] accompanied with their threat assessment. + +For a specific corruption style such as Gaussian noise and a severity level 5, we let $\Omega_{\mathrm{GaussianNoise},5}$ denote the corresponding corruption function. To aid visualization, we group the corruption styles into 6 distinct categories (listed in Table 2) so that $\Omega := \cup_{i \in [5]} \Omega_i$ where $\Omega_i$ is the set of all corruption styles at severity level $i$ , + +$$ +\Omega_ {i} := \left(\Omega_ {\text {n o i s e}, i} \cup \Omega_ {\text {b l u r}, i} \cup \Omega_ {\text {c o m p r e s s i o n}, i} \cup \Omega_ {\text {d i g i t a l}, i} \cup \Omega_ {\text {w e a t h e r}, i} \cup \Omega_ {\text {o c c l u s i o n}, i}\right). +$$ + +Each individual corruption $\omega \in \Omega$ is applied to a subset $\bar{\mathbf{S}}\subset \mathbf{S}_{\mathrm{val}}$ of 5,000 images chosen uniformly at random. For each threat function $d\in \{d_{\infty},d_2,d_{\mathrm{PD}},d_{\mathrm{DS}}\}$ and corruption $\omega \in \Omega$ , for each image $\mathbf{x}\in \bar{\mathbf{S}}$ , we denote by $\mathrm{avg}(d,\omega)$ , the average threat statistic, $\mathrm{avg}(d,\omega)\coloneqq \frac{1}{|\bar{s}|}\sum_{(\mathbf{x},y)\in S}d(\mathbf{x},\omega (\mathbf{x}) - \mathbf{x})$ , where $\omega (\mathbf{x})$ is the corrupted input and $\omega (\mathbf{x}) - \mathbf{x}$ is the safe perturbation at $\mathbf{x}$ . + +# 4.1. Disentangling Safe and Unsafe Corruption + +We now measure the ability of threat models to distinguish safe and unsafe corruptions. Each corruption $\omega \in \Omega$ defines a point in Figure 3a having as the horizontal coordinate $\mathrm{avg}(d_{\infty},\omega)$ and $\mathrm{avg}(d_{\mathrm{PD}},\omega)$ as the vertical coordinate, producing a comparison of our PD threat to the $\ell_{\infty}$ -threat model. Similarly, Figure 3b compares PD threat to the DreamSIM-threat model. + +A specification of threat model $(d, \varepsilon)$ corresponds to the choice of a horizontal line for $d_{\mathrm{PD}}$ or a vertical line for $(d_{\infty}$ + +or $d_{\mathrm{DS}}$ ) in Figure 3. We note that a size-able set of common corruptions incur large $d_{\infty}$ threat, in particular as large as the unsafe corruptions (as indicated in Figure 1), thus any vertical line in Figure 3a is bound to exclude such common corruptions. In contrast, both $d_{\mathrm{PD}}$ and $d_{\mathrm{DS}}$ admit choices of $\varepsilon$ that exclude unsafe perturbations while including most common corruptions. Hence, PD threat and DS threat are able to disentangle safe and unsafe corruptions. + +For each of the comparisons in Figure 3, we identify 4 quadrants I-IV, distinguishing between low-threat and high-threat corruptions as measured by the corresponding threat models. We define the quadrants by choosing thresholds of 1, 0.5, and 0.25 for $d_{\mathrm{PD}}$ , $d_{\infty}$ , and $d_{\mathrm{DS}}$ , respectively9, and comment on the qualitative behavior observed: + +- Quadrant IV contains corruptions that are characterized as low-threat by both threat models. As expected, almost all corruptions having a low severity (1) lie in this region. However, we already see deficiencies of the $d_{\infty}$ threat, as some blur, noise, and occlusion corruptions of severity 1 in Figure 3a are not characterized as low-threat by at least one of the threat models. + +- Quadrant II contains corruptions that are characterized as high-threat by both threat models. These include digital and weather corruptions of high severity (5) as they tend to remove most of the signal in the input. + +- Quadrant I contains corruptions having low PD-threat but high $d_{\infty}$ or $d_{\mathrm{DS}}$ and showcases the resilience of our PD-threat model to natural corruptions like blur and Gaussian noise, which retain a lot of the information in the image relevant to the classification task. Indeed, a visual inspection of these corruptions (see Section 10 in the Appendix) reveals that even at a high severity, a human is able to discern task-relevant information from the corrupted images, and hence these corruptions should be assigned a low threat. + +- Quadrant III contains corruptions with high PD threat but low $d_{\infty}$ or $d_{\mathrm{DS}}$ threat. We note that corruptions rate as high threat by $d_{\mathrm{PD}}$ are largely also rated highly by $d_{\mathrm{DS}}$ . Since the PD and DS threats are aligned, the corruptions in Quadrant III in Figure 3a indicate that low $d_{\infty}$ is not sufficient to characterize hard corruptions. + +As a final point, we note that while the size of markers in Figure 3 indicate the severity levels, different corruptions of similar severity levels are qualitatively different. For e.g, weather corruptions of severity level 5 are uniformly harder than noise corruptions such as shot noise, Gaussian noise. Figure 11 illustrates heatmap of average threat for each corruption category across increasing severity. We observe that both PD and DS threat exhibit largely monotonic + +increase across severity unlike $d_{\infty}$ . At this point, we will pause to summarize the benefits of PD-threat. + +![](images/920cca1a7bf6f7190f4a454ce4244bf8e2d3d453e4edd1c6663d204ac5f4d784.jpg) +(a) + +![](images/c1b23194f51f9e38ce127382832edc10c7be4637cd15b6dcd3e547ebe6e6624c.jpg) +(b) + +PD vs DreamSim. The PD-threat compares well to the DS-threat but does not require pre-trained large vision models or fine-tuning on a curated dataset of human perceptual similarity judgements. For real-world classification tasks where perceptual similarity between pairs of input is harder to quantify even with expert human annotators, instantiation of high-quality neural perceptual distance metrics is harder. Further, unlike perceptual distance metrics [15, 17, 18, 32, 54], the growth of the PD-threat is linear and interpretable along any direction, the sub-level sets induced by the PD-threat model are convex, and admit efficient projection algorithms and hence PD threat presents a promising alternative. + +# 5. Measuring threat with Task Annotation + +We described two practical design choices for PD-threat in Section 3.2: (1) a choice of the set of unsafe directions $\mathcal{U}(\mathbf{x})$ at each input, and (2) a choice of normalization $g(\mathbf{x},\mathbf{u})$ for each unsafe direction. Next, we incorporate additional task-relevant information to refine PD-threat further. + +# 5.1. Segmentation-aware Threat Specification + +Segmentation masks isolate regions of the input that contain semantic information relevant to the class label, and can be generated efficiently following recent advances in foundational models for image segmentation [30, 38]. We demonstrate a straightforward method to incorporate semantic masks in the assessment of threat. + +Definition 8 (PD-S threat). Let $\mathbf{x} \in \mathcal{X}$ and let $\delta \in \mathbb{R}^d$ . Let $\mathbf{a} \in \{0,1\}^d$ be a boolean mask. The segmented PD threat, + +$$ +d _ {\mathrm {P D - S}} (\mathbf {x}, \mathbf {a}, \boldsymbol {\delta}) := \max _ {\mathbf {u} \in \mathcal {U} (\mathbf {x})} \frac {1}{g _ {\boldsymbol {\beta}} (\mathbf {x} , \mathbf {u})} \max \left(\langle \boldsymbol {\delta} [ \mathbf {a} ], \mathbf {u} [ \mathbf {a} ] \rangle , 0\right). +$$ + +Here the $\mathbf{u}[\mathbf{a}]$ denotes a sub-vector whose indices are selected by the boolean mask $\mathbf{a}$ . The threat model $(d_{\mathrm{PD-S}}, \varepsilon)$ is aligned with Xiao et al. [51]'s proposal to learn models robust to adversarially chosen backgrounds. For each image, we generate automatic masks by prompting SAM10 with the center pixel coordinates. Figure 4 illustrates the difference between PD and PD-S threat on 2 corruptions that are background-only and foreground-only respectively. PD-S threat is oblivious to background corruption but more sensitive to foreground corruption. + +![](images/5462da7c42ae9bf65700ad81708555cfdd3bd5f026366f2bb6f22c619c2eb01e.jpg) +Figure 3. Comparison of Threat models w.r.t. various common corruptions. The size of a marker represents the severity level. +Original Image +Figure 4. Comparison of threat on background vs foreground. + +![](images/24e99c7989721abcb7274ffba9b6228b921d7bcd540f9bead52f9a64c2bb8d3d.jpg) +$\mathrm{PD} = 1.4$ +PD-S=0.0 + +![](images/c664177ece929bf1a2a5d80009090d35a775d5cd81f42704bbb36807eab55069.jpg) +$\mathrm{PD} = 0.68$ +PD-S=1.78 + +We note that $\ell_p$ threat models are also capable of integrating segmentation masks as they are pixel-based, but it is unclear if perceptual threat can account for pixel annotation. + +# 5.2. Concept-aware Threat Specification + +The PD-threat assumes no hierarchy of class labels; i.e. all labels are distinct to the same degree. However, classification tasks are often accompanied with an additional hierarchy that identifies groups of class labels based on similar semantic content. We propose to account for such a hierarchy by refining our choice of normalization. Suppose $\mathbf{x}$ has label + +$y$ and $\mathbf{u} = \frac{\tilde{\mathbf{x}} - \mathbf{x}}{\|\tilde{\mathbf{x}} - \mathbf{x}\|_2}$ where $\tilde{\mathbf{x}}\in S_{c,k}$ has label c. Let $W:[C]\times [C]\in [0,1]$ denote a relative distance between class labels based on Wordnet-hierarchy (refer to Section 9.5 for explicit details). We propose the weighted normalization $g_{\beta ,W}(\mathbf{x},\mathbf{u})\coloneqq \boldsymbol {\beta}\cdot W(y,c)\| \tilde{\mathbf{x}} -\mathbf{x}\| _2$ so that perturbation between nearby classes are weighted higher in threat. + +Definition 9 (PD-W threat). Let $\mathbf{x} \in \mathcal{X}$ with label $y$ and let $\pmb{\delta} \in \mathbb{R}^{d}$ . The weighted PD threat is defined as, + +$$ +d_{\mathrm{PD - W}}(\mathbf{x},\boldsymbol {\delta}):= \max_{\mathbf{u}\in \mathcal{U}(\mathbf{x})}\frac{1}{g_{\beta,W}(\mathbf{x},\mathbf{u})}\max \left(\langle \boldsymbol {\delta},\mathbf{u}\rangle ,0\right). +$$ + +The threat model $(d_{\mathrm{PD - W}},\varepsilon)$ is weaker than $(d_{\mathrm{PD}},\varepsilon)$ , since $S(\mathbf{x},d_{\mathrm{PD - W}},\varepsilon)\subseteq S(\mathbf{x},d_{\mathrm{PD}},\varepsilon)$ . Thus PD-W is a relaxation of PD threat with a softer requirement of stability between nearby classes. Figure 5 depicts the average relative $^{12}$ threat vs relative distance $W(\cdot ,\cdot)$ of class labels on the Imagenet-1k validation dataset. The decreasing PD-W threat for increasing class distance is aligned with the following intuition - failure to distinguish between semantically distant classes such as ENGLISH FOXHOUND and FIRE TRUCK should incur lower robust accuracy than failure to distinguish between similar classes such as ENGLISH FOXHOUND and IRISH WOLFHOUND. + +![](images/37f4b90c34d11b1ef600464461f5cc220b8ed0f11c89285aa67e83458ed08837.jpg) +Figure 5. Relative Threat vs relative inter-class distances + +In summary our novel threat specification framework enables evaluation of robustness to adversarial perturbations and common corruptions while enabling sensitivity to semantic regions of an image and class hierarchy by incorporating additional task annotation. As a consequence, our threat model unifies adversarial robustness with variants of corruption robustness suggested by several independent benchmarks [25, 36, 51]. + +# 6. Evaluation of benchmark $\ell_p$ robust models + +Finally, we evaluate state-of-the-art robust models against our proposed thread models. Note that, as per Theorem 1, + +one should hope for classifiers that are 1-robust to corruptions under the PD-threat. The corresponding permissible set of corruptions cover regions IV and I in Figure 3. We evaluate robustness to corruptions in the set $S(\mathbf{x}, d_{\mathrm{PD}}, \varepsilon) \cap S(\mathbf{x}, d_{\infty}, \varepsilon)$ . We denote the corresponding threat model $(d_{\infty} \bigcap d_{\mathrm{PD}}, \varepsilon)$ . A robust evaluation for such a threat model can be executed in practice by generating adversarial perturbations with standard AutoAttack [7] for $(d_{\infty}, \varepsilon)$ followed by projection onto $S(\mathbf{x}, d_{\mathrm{PD}}, \varepsilon)$ . Table 3 shows the robustness of benchmark classifiers registered at RobustBench order by $d_{\infty}$ robustness (column 1) when evaluated against $(d_{\infty} \bigcap d_{\mathrm{PD}}, 16/255)$ threat model for the Imagenet-1k classification task. + +
Benchmark Modeld∞PDPD-SPD-W
ConvNeXt-L [33]0.250.320.310.31
Swin-B [33]0.120.210.200.21
ConvNeXt-B [33]0.10.190.180.18
ConvNeXt-B-ConvStem [41]0.090.190.190.19
ViT-S-ConvStem [41]0.050.130.120.12
XCiT-L12 [10]0.040.170.180.18
+ +Table 3. Robust Evaluation for $\varepsilon = {16}/{255}$ . + +We note that robust accuracies uniformly improve under our threat specification in comparison to $(d_{\infty}, \varepsilon)$ , indicating a meaningful restriction of unsafe corruptions. The ordering of robust accuracies highlights diversity, e.g. [10]'s XCiT-L12 fares better under PD threat than suggested by the $d_{\infty}$ leaderboard. The indistinguishability across PD, PD-S and PD-w is surprising. Similar levels of robustness w.r.t PD-W and PD indicate that models are unable to distinguish distant classes, while similar levels under PD and PD-S indicate adversarial attacks are often aligned with foreground masks. We note that adversarial attacks adapted to the threat specification often lead to a more accurate pessimistic evaluation of robustness. + +# 7. Conclusion + +This work proposes a novel task-dependent threat specification Projected Displacement threat that is adapted to the shape of decision boundaries based on observed training data. Unlike standard $\ell_p$ -threat models, the proposed threat model exhibits anisotropy and locality and is thus able to distinguish between safe and unsafe perturbations. The proposed threat specification framework is flexible and can effectively account for additional task information, such as image regions or label hierarchies. We invite the community to suggest further adaptive attacks and effective training strategies to encourage robustness in PD threat. + +# Acknowledgements + +This research was supported by NSF CISE 2212457 and NSF CAREER Award CCF 2239787 + +# References + +[1] Naveed Akhtar, Ajmal Mian, Navid Kardan, and Mubarak Shah. Advances in adversarial attacks and defenses in computer vision: A survey. IEEE Access, 9:155161-155196, 2021. 1 +[2] Pranjal Awasthi, Anqi Mao, Mehryar Mohri, and Yutao Zhong. Theoretically grounded loss functions and algorithms for adversarial robustness. In Proceedings of The 26th International Conference on Artificial Intelligence and Statistics, pages 10077-10094. PMLR, 2023. 3 +[3] Yatong Bai, Brendon G. Anderson, Aerin Kim, and Somayeh Sojoudi. Improving the accuracy-robustness trade-off of classifiers via adaptive smoothing, 2024. 1 +[4] Brian R. Bartoldson, James Diffenderfer, Konstantinos Parasyris, and Bhavya Kailkhura. Adversarial robustness limits via scaling-law and human-alignment studies. In Proceedings of the 41st International Conference on Machine Learning, pages 3046-3072. PMLR, 2024. 1, 2 +[5] Yulong Cao, Chaowei Xiao, Benjamin Cyr, Yimeng Zhou, Wonseok Park, Sara Rampazzi, Qi Alfred Chen, Kevin Fu, and Z. Morley Mao. Adversarial sensor attack on lidar-based perception in autonomous driving. Proceedings of the 2019 ACM SIGSAC Conference on Computer and Communications Security, 2019. 1 +[6] Nicholas Carlini and David A. Wagner. Adversarial examples are not easily detected: Bypassing ten detection methods. Proceedings of the 10th ACM Workshop on Artificial Intelligence and Security, 2017. 1 +[7] Francesco Croce and Matthias Hein. Reliable evaluation of adversarial robustness with an ensemble of diverse parameter-free attacks. In Proceedings of the 37th International Conference on Machine Learning. JMLR.org, 2020. 1, 2, 8 +[8] Francesco Croce, Maksym Andriushchenko, Vikash Sehwag, Edoardo Debenedetti, Nicolas Flammarion, Mung Chiang, Prateek Mittal, and Matthias Hein. Robustbench: a standardized adversarial robustness benchmark. In Thirty-fifth Conference on Neural Information Processing Systems Datasets and Benchmarks Track, 2021. 1, 2 +[9] Jiequan Cui, Zhuotao Tian, Zhisheng Zhong, Xiaojuan Qi, Bei Yu, and Hanwang Zhang. Decoupled kullback-leibler divergence loss, 2023. 1 +[10] Edoardo Debenedetti, Vikash Sehwag, and Prateek Mittal. A light recipe to train robust vision transformers, 2023. 8 +[11] Edoardo Debenedetti, Zishen Wan, Maksym Andriushchenko, Vikash Sehwag, Kshitij Bhardwaj, and Bhavya Kailkhura. Scaling compute is not all you need for adversarial robustness. In ICLR 2024 Workshop on Reliable and Responsible Foundation Models, 2024. 1, 2 +[12] Samuel Dodge and Lina Karam. A study and comparison of human and deep learning recognition performance under visual distortions. In 2017 26th international conference on computer communication and networks (ICCCN), pages 1-7. IEEE, 2017. 3 + +[13] Elvis Dohmatob. Generalized no free lunch theorem for adversarial robustness. In ICML, 2019. 4 +[14] Ecenaz Erdemir, Jeffrey Bickford, Luca Melis, and Sergul Aydore. Adversarial robustness with non-uniform perturbations, 2021. 4 +[15] Stephanie Fu, Netanel Tamir, Shobhita Sundaram, Lucy Chai, Richard Zhang, Tali Dekel, and Phillip Isola. Dreamsim: Learning new dimensions of human visual similarity using synthetic data. Advances in Neural Information Processing Systems, 36, 2024. 2, 3, 5, 7 +[16] Robert Geirhos, Carlos R. M. Temme, Jonas Rauber, Heiko H. Schütt, Matthias Bethge, and Felix A. Wichmann. Generalisation in humans and deep neural networks. In Advances in Neural Information Processing Systems. Curran Associates, Inc., 2018. 3 +[17] Sara Ghazanfari, Alexandre Araujo, Prashanth Krishnamurthy, Farshad Khorrami, and Siddharth Garg. Lipsim: A provably robust perceptual similarity metric. ArXiv, abs/2310.18274, 2023. 2, 7 +[18] Sara Ghazanfari, Siddharth Garg, Prashanth Krishnamurthy, Farshad Khorrami, and Alexandre Araujo. R-LPIPS: An adversarially robust perceptual similarity metric. In The Second Workshop on New Frontiers in Adversarial Machine Learning, 2023. 7 +[19] Abhijay Ghildyal and Feng Liu. Attacking perceptual similarity metrics. Trans. Mach. Learn. Res., 2023, 2023. 2 +[20] Justin Gilmer, Ryan P. Adams, Ian Goodfellow, David Andersen, and George E. Dahl. Motivating the rules of the game for adversarial example research, 2018. 1 +[21] Teofilo F. Gonzalez. Clustering to minimize the maximum intercluster distance. Theoretical Computer Science, 38:293-306, 1985. 11 +[22] Ian J. Goodfellow, Jonathon Shlens, and Christian Szegedy. Explaining and harnessing adversarial examples. CoRR, abs/1412.6572, 2015. 1 +[23] Sven Gowal, Chongli Qin, Jonathan Uesato, Timothy Mann, and Pushmeet Kohli. Uncovering the limits of adversarial training against norm-bounded adversarial examples, 2021. 1 +[24] Sven Gowal, Sylvestre-Alvise Rebuffi, Olivia Wiles, Florian Stimberg, Dan Andrei Calian, and Timothy Mann. Improving robustness using generated data, 2021. 1 +[25] Dan Hendrycks and Thomas Dietterich. Benchmarking neural network robustness to common corruptions and perturbations. Proceedings of the International Conference on Learning Representations, 2019. 3, 6, 8, 16 +[26] Joern-Henrik Jacobsen, Jens Behrmann, Richard Zemel, and Matthias Bethge. Excessive invariance causes adversarial vulnerability. In International Conference on Learning Representations, 2019. 2 +[27] Can Kanbak, Seyed-Mohsen Moosavi-Dezfooli, and Pascal Frossard. Geometric Robustness of Deep Networks: Analysis and Improvement. In 2018 IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 4441-4449, Los Alamitos, CA, USA, 2018. IEEE Computer Society. 3 +[28] Oğuzhan Fatih Kar, Teresa Yeo, Andrei Atanov, and Amir Zamir. 3d common corruptions and data augmentation. In + +Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 18963-18974, 2022. 3 +[29] Markus Kettunen, Erik Härkönen, and Jaakko Lehtinen. ELPips: Robust perceptual image similarity via random transformation ensembles, 2019. 2 +[30] Alexander Kirillov, Eric Mintun, Nikhila Ravi, Hanzi Mao, Chloe Rolland, Laura Gustafson, Tete Xiao, Spencer Whitehead, Alexander C. Berg, Wan-Yen Lo, Piotr Dólar, and Ross Girshick. Segment anything. arXiv:2304.02643, 2023. 7 +[31] Alexey Kurakin, Ian J. Goodfellow, and Samy Bengio. Adversarial examples in the physical world. ArXiv, abs/1607.02533, 2017. 1 +[32] Cassidy Laidlaw, Sahil Singla, and Soheil Feizi. Perceptual adversarial robustness: Defense against unseen threat models, 2021. 2, 5, 7 +[33] Chang Liu, Yinpeng Dong, Wenzhao Xiang, Xiao Yang, Hang Su, Jun Zhu, Yuefeng Chen, Yuan He, Hui Xue, and Shibao Zheng. A comprehensive study on robustness of image classification models: Benchmarking and rethinking, 2023. 1, 8, 13 +[34] Aleksander Madry, Aleksandar Makelov, Ludwig Schmidt, Dimitris Tsipras, and Adrian Vladu. Towards deep learning models resistant to adversarial attacks. ArXiv, abs/1706.06083, 2018. 1 +[35] Rafat K. Mantiuk, Kil Joong Kim, Allan G. Rempel, and Wolfgang Heidrich. Hdr-vdp-2: a calibrated visual metric for visibility and quality predictions in all luminance conditions. ACM SIGGRAPH 2011 papers, 2011. 3 +[36] Eric Mintun, Alexander Kirillov, and Saining Xie. On interaction between augmentations and corruptions in natural corruption robustness. In Advances in Neural Information Processing Systems, pages 3571-3583. Curran Associates, Inc., 2021. 3, 6, 8, 16 +[37] Yao Qin, Nicholas Carlini, Garrison Cottrell, Ian Goodfellow, and Colin Raffel. Imperceptible, robust, and targeted adversarial examples for automatic speech recognition. In International conference on machine learning, pages 5231-5240. PMLR, 2019. 1 +[38] Nikhila Ravi, Valentin Gabeur, Yuan-Ting Hu, Ronghang Hu, Chaitanya Ryali, Tengyu Ma, Haitham Khedr, Roman Rädle, Chloe Rolland, Laura Gustafson, Eric Mintun, Junting Pan, Kalyan Vasudev Alwala, Nicolas Carion, Chao-Yuan Wu, Ross Girshick, Piotr Dólár, and Christoph Feichtenhofer. Sam 2: Segment anything in images and videos. arXiv preprint arXiv:2408.00714, 2024. 7 +[39] Ayon Sen, Xiaojin Zhu, Erin Marshall, and Robert Nowak. Popular imperceptibility measures in visual adversarial attacks are far from human perception. In Decision and Game Theory for Security, pages 188-199, Cham, 2020. Springer International Publishing. 2 +[40] Mahmood Sharif, Lujo Bauer, and Michael K. Reiter. On the suitability of lp-norms for creating and preventing adversarial examples. 2018 IEEE/CVF Conference on Computer Vision and Pattern Recognition Workshops (CVPRW), pages 1686-16868, 2018. 2 +[41] Naman D Singh, Francesco Croce, and Matthias Hein. Revisiting adversarial training forImagenet: Architectures, training and generalization across threat models, 2023. 8 + +[42] Christian Szegedy, Wojciech Zaremba, Ilya Sutskever, Joan Bruna, D. Erhan, Ian J. Goodfellow, and Rob Fergus. Intriguing properties of neural networks. CoRR, abs/1312.6199, 2014. 1 +[43] Florian Tramér, Jens Behrmann, Nicholas Carlini, Nicolas Papernot, and Jörn-Henrik Jacobsen. Fundamental tradeoffs between invariance and sensitivity to adversarial perturbations. In Proceedings of the 37th International Conference on Machine Learning, ICML 2020, 13-18 July 2020, Virtual Event, pages 9561-9571. PMLR, 2020. 2, 3 +[44] Florian Tramér, Nicholas Carlini, Wieland Brendel, and Aleksander Madry. On adaptive attacks to adversarial example defenses. ArXiv, abs/2002.08347, 2020. 1 +[45] James Tu, Huichen Li, Xinchen Yan, Mengye Ren, Yun Chen, Ming Liang, Eilyan Bitar, Ersin Yumer, and Raquel Urtasun. Exploring adversarial robustness of multi-sensor perception systems in self driving. In Proceedings of the 5th Conference on Robot Learning, pages 1013-1024. PMLR, 2022. 1 +[46] Apostol Vassilev, Alina Oprea, Alie Fordyce, and Hyrum Andersen. Adversarial machine learning: A taxonomy and terminology of attacks and mitigations, 2024. 1 +[47] Zhou Wang, A.C. Bovik, H.R. Sheikh, and E.P. Simoncelli. Image quality assessment: from error visibility to structural similarity. IEEE Transactions on Image Processing, 13(4): 600-612, 2004. 3 +[48] Zekai Wang, Tianyu Pang, Chao Du, Min Lin, Weiwei Liu, and Shuicheng Yan. Better diffusion models further improve adversarial training, 2023. 1 +[49] Zuxuan Wu, Ser-Nam Lim, Larry Davis, and Tom Goldstein. Making an invisibility cloak: Real world adversarial attacks on object detectors, 2020. 1 +[50] Chaowei Xiao, Jun-Yan Zhu, Bo Li, Warren He, Mingyan Liu, and Dawn Song. Spatially transformed adversarial examples. In International Conference on Learning Representations, 2018. 3 +[51] Kai Yuanqing Xiao, Logan Engstrom, Andrew Ilyas, and Aleksander Madry. Noise or signal: The role of image backgrounds in object recognition. In International Conference on Learning Representations, 2021. 7, 8 +[52] Hongyang Zhang, Yaodong Yu, Jiantao Jiao, Eric Xing, Laurent El Ghaoui, and Michael Jordan. Theoretically principled trade-off between robustness and accuracy. In Proceedings of the 36th International Conference on Machine Learning, pages 7472-7482. PMLR, 2019. 1 +[53] Lin Zhang, Lei Zhang, Xuanqin Mou, and David Zhang. Fsim: A feature similarity index for image quality assessment. IEEE Transactions on Image Processing, 20(8):2378-2386, 2011. 3 +[54] Richard Zhang, Phillip Isola, Alexei A. Efros, Eli Shechtman, and Oliver Wang. The unreasonable effectiveness of deep features as a perceptual metric. In 2018 IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 586-595, Los Alamitos, CA, USA, 2018. IEEE Computer Society. 3, 7 +[55] Piotr Želasko, Sonal Joshi, Yiwen Shao, Jesus Villalba, Jan Trmal, Najim Dehak, and Sanjeev Khudanpur. Adversarial attacks and defenses for speech recognition systems, 2021. \ No newline at end of file diff --git a/2025/Disentangling Safe and Unsafe Image Corruptions via Anisotropy and Locality/images.zip b/2025/Disentangling Safe and Unsafe Image Corruptions via Anisotropy and Locality/images.zip new file mode 100644 index 0000000000000000000000000000000000000000..eaadf0e25a4f318ef22b1bb821cbd62c192d77bd --- /dev/null +++ b/2025/Disentangling Safe and Unsafe Image Corruptions via Anisotropy and Locality/images.zip @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:5eb49a654f22570bcd45e3a014d1ec671185d217cfeb4dc793759b295803d850 +size 271288 diff --git a/2025/Disentangling Safe and Unsafe Image Corruptions via Anisotropy and Locality/layout.json b/2025/Disentangling Safe and Unsafe Image Corruptions via Anisotropy and Locality/layout.json new file mode 100644 index 0000000000000000000000000000000000000000..5611bccc5552e28ae913dc2d537ae5f188b1ddf2 --- /dev/null +++ b/2025/Disentangling Safe and Unsafe Image Corruptions via Anisotropy and Locality/layout.json @@ -0,0 +1,15817 @@ +{ + "pdf_info": [ + { + "para_blocks": [ + { + "bbox": [ + 62, + 102, + 549, + 122 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 102, + 549, + 122 + ], + "spans": [ + { + "bbox": [ + 62, + 102, + 549, + 122 + ], + "type": "text", + "content": "Disentangling Safe and Unsafe Image Corruptions via Anisotropy and Locality" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 100, + 142, + 507, + 172 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 100, + 142, + 507, + 172 + ], + "spans": [ + { + "bbox": [ + 100, + 142, + 507, + 172 + ], + "type": "text", + "content": "Ramchandran Muthukumar*¹ Ambar Pal²† Jeremias Sulam¹ René Vidal³ \n¹Johns Hopkins University ²Amazon Web Services ³University of Pennsylvania" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 151, + 199, + 200, + 211 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 151, + 199, + 200, + 211 + ], + "spans": [ + { + "bbox": [ + 151, + 199, + 200, + 211 + ], + "type": "text", + "content": "Abstract" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 54, + 224, + 297, + 620 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 54, + 224, + 297, + 620 + ], + "spans": [ + { + "bbox": [ + 54, + 224, + 297, + 620 + ], + "type": "text", + "content": "State-of-the-art machine learning systems are vulnerable to small perturbations to their input, where \"small\" is defined according to a threat model that assigns a positive threat to each perturbation. Most prior works define a task-agnostic, isotropic, and global threat, like the " + }, + { + "bbox": [ + 54, + 224, + 297, + 620 + ], + "type": "inline_equation", + "content": "\\ell_p" + }, + { + "bbox": [ + 54, + 224, + 297, + 620 + ], + "type": "text", + "content": " norm, where the magnitude of the perturbation fully determines the degree of the threat and neither the direction of the attack nor its position in space matter. However, common corruptions in computer vision, such as blur, compression, or occlusions, are not well captured by such threat models. This paper proposes a novel threat model called Projected Displacement (PD) to study robustness beyond existing isotropic and global threat models. The proposed threat model measures the threat of a perturbation via its alignment with unsafe directions, defined as directions in the input space along which a perturbation of sufficient magnitude changes the ground truth class label. Unsafe directions are identified locally for each input based on observed training data. In this way, the PD-threat model exhibits anisotropy and locality. Experiments on Imagenet-1k data indicate that, for any input, the set of perturbations with small PD threat includes safe perturbations of large " + }, + { + "bbox": [ + 54, + 224, + 297, + 620 + ], + "type": "inline_equation", + "content": "\\ell_p" + }, + { + "bbox": [ + 54, + 224, + 297, + 620 + ], + "type": "text", + "content": " norm that preserve the true label, such as noise, blur and compression, while simultaneously excluding unsafe perturbations that alter the true label. Unlike perceptual threat models based on embeddings of large-vision models, the PD-threat model can be readily computed for arbitrary classification tasks without pre-training or finetuning. Further additional task information such as sensitivity to image regions or concept hierarchies can be easily integrated into the assessment of threat and thus the PD threat model presents practitioners with a flexible, task-driven threat specification that alleviates the limitations of " + }, + { + "bbox": [ + 54, + 224, + 297, + 620 + ], + "type": "inline_equation", + "content": "\\ell_p" + }, + { + "bbox": [ + 54, + 224, + 297, + 620 + ], + "type": "text", + "content": "-threat models." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 56, + 641, + 135, + 654 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 641, + 135, + 654 + ], + "spans": [ + { + "bbox": [ + 56, + 641, + 135, + 654 + ], + "type": "text", + "content": "1. Introduction" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 55, + 662, + 296, + 686 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 662, + 296, + 686 + ], + "spans": [ + { + "bbox": [ + 55, + 662, + 296, + 686 + ], + "type": "text", + "content": "Modern machine learning (ML) systems are nearing widespread deployment in civilian life, and hence a com" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 313, + 200, + 555, + 403 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 200, + 555, + 403 + ], + "spans": [ + { + "bbox": [ + 313, + 200, + 555, + 403 + ], + "type": "text", + "content": "prehensive understanding of their security vulnerabilities is necessary [46]. One such vulnerability concerns the ability of a malicious adversary to tamper with predictions by adding small imperceptible corruptions referred to as adversarial perturbations [22, 42]. There is by now overwhelming evidence that carefully crafted adversarial perturbations can foil the prediction of state-of-the-art machine learning classifiers, i.e. they are not adversarily robust [6, 8]. Such vulnerabilities have been observed in a wide variety of applications like computer vision [1, 49], speech recognition [37, 55], autonomous driving [5, 45], and more. The goal of designing safe and reliable ML systems remains incomplete [4, 11] despite significant investment [23, 24, 33]. Often, strategies that intend to foil performance, i.e. adversarial attacks [6, 7, 22, 31, 44], have proven more successful than strategies aimed at mitigating vulnerabilities, i.e. adversarial defenses [3, 9, 34, 48, 52]." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 313, + 404, + 556, + 594 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 404, + 556, + 594 + ], + "spans": [ + { + "bbox": [ + 313, + 404, + 556, + 594 + ], + "type": "text", + "content": "Meaningfully evaluating the adversarial robustness of a machine learning system requires a formal specification of a threat function, and an associated threat model that limits the scope of malicious adversaries [20]. Informally, a threat function " + }, + { + "bbox": [ + 313, + 404, + 556, + 594 + ], + "type": "inline_equation", + "content": "d: \\mathbb{R}^d \\times \\mathbb{R}^d \\to \\mathbb{R}^{\\geq 0}" + }, + { + "bbox": [ + 313, + 404, + 556, + 594 + ], + "type": "text", + "content": " measures the threat represented by a corruption " + }, + { + "bbox": [ + 313, + 404, + 556, + 594 + ], + "type": "inline_equation", + "content": "\\delta \\in \\mathbb{R}^d" + }, + { + "bbox": [ + 313, + 404, + 556, + 594 + ], + "type": "text", + "content": " towards altering the true label at input " + }, + { + "bbox": [ + 313, + 404, + 556, + 594 + ], + "type": "inline_equation", + "content": "\\mathbf{x} \\in \\mathbb{R}^d" + }, + { + "bbox": [ + 313, + 404, + 556, + 594 + ], + "type": "text", + "content": " as " + }, + { + "bbox": [ + 313, + 404, + 556, + 594 + ], + "type": "inline_equation", + "content": "d(\\mathbf{x}, \\delta)" + }, + { + "bbox": [ + 313, + 404, + 556, + 594 + ], + "type": "text", + "content": ". We let " + }, + { + "bbox": [ + 313, + 404, + 556, + 594 + ], + "type": "inline_equation", + "content": "S(\\mathbf{x}, d, \\varepsilon) \\coloneqq \\{\\delta \\in \\mathbb{R}^d \\mid d(\\mathbf{x}, \\delta) \\leq \\varepsilon\\}" + }, + { + "bbox": [ + 313, + 404, + 556, + 594 + ], + "type": "text", + "content": " denote the " + }, + { + "bbox": [ + 313, + 404, + 556, + 594 + ], + "type": "inline_equation", + "content": "\\varepsilon" + }, + { + "bbox": [ + 313, + 404, + 556, + 594 + ], + "type": "text", + "content": "-sublevel set of perturbations where the threat " + }, + { + "bbox": [ + 313, + 404, + 556, + 594 + ], + "type": "inline_equation", + "content": "d" + }, + { + "bbox": [ + 313, + 404, + 556, + 594 + ], + "type": "text", + "content": " is measured w.r.t. input " + }, + { + "bbox": [ + 313, + 404, + 556, + 594 + ], + "type": "inline_equation", + "content": "\\mathbf{x}" + }, + { + "bbox": [ + 313, + 404, + 556, + 594 + ], + "type": "text", + "content": ". A threat model " + }, + { + "bbox": [ + 313, + 404, + 556, + 594 + ], + "type": "inline_equation", + "content": "(d, \\varepsilon)" + }, + { + "bbox": [ + 313, + 404, + 556, + 594 + ], + "type": "text", + "content": " is a pair of a threat function " + }, + { + "bbox": [ + 313, + 404, + 556, + 594 + ], + "type": "inline_equation", + "content": "d" + }, + { + "bbox": [ + 313, + 404, + 556, + 594 + ], + "type": "text", + "content": " and a permissible threshold " + }, + { + "bbox": [ + 313, + 404, + 556, + 594 + ], + "type": "inline_equation", + "content": "\\varepsilon" + }, + { + "bbox": [ + 313, + 404, + 556, + 594 + ], + "type": "text", + "content": " that together define the set of permissible perturbations " + }, + { + "bbox": [ + 313, + 404, + 556, + 594 + ], + "type": "inline_equation", + "content": "S(\\mathbf{x}, d, \\varepsilon)" + }, + { + "bbox": [ + 313, + 404, + 556, + 594 + ], + "type": "text", + "content": ". Under the threat model " + }, + { + "bbox": [ + 313, + 404, + 556, + 594 + ], + "type": "inline_equation", + "content": "(d, \\varepsilon)" + }, + { + "bbox": [ + 313, + 404, + 556, + 594 + ], + "type": "text", + "content": ", the robust accuracy of a classifier " + }, + { + "bbox": [ + 313, + 404, + 556, + 594 + ], + "type": "inline_equation", + "content": "h \\in \\mathcal{H}" + }, + { + "bbox": [ + 313, + 404, + 556, + 594 + ], + "type": "text", + "content": " is the probability that the label prediction at " + }, + { + "bbox": [ + 313, + 404, + 556, + 594 + ], + "type": "inline_equation", + "content": "\\mathbf{x}" + }, + { + "bbox": [ + 313, + 404, + 556, + 594 + ], + "type": "text", + "content": " is locally invariant to corruptions within the permissible set " + }, + { + "bbox": [ + 313, + 404, + 556, + 594 + ], + "type": "inline_equation", + "content": "S(\\mathbf{x}, d, \\varepsilon)" + }, + { + "bbox": [ + 313, + 404, + 556, + 594 + ], + "type": "text", + "content": ", see Definition 2 for an explicit definition." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 313, + 596, + 556, + 669 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 596, + 556, + 669 + ], + "spans": [ + { + "bbox": [ + 313, + 596, + 556, + 669 + ], + "type": "text", + "content": "One of the most commonly used threat models is the " + }, + { + "bbox": [ + 313, + 596, + 556, + 669 + ], + "type": "inline_equation", + "content": "\\ell_p" + }, + { + "bbox": [ + 313, + 596, + 556, + 669 + ], + "type": "text", + "content": "-threat model " + }, + { + "bbox": [ + 313, + 596, + 556, + 669 + ], + "type": "inline_equation", + "content": "(d_p, \\epsilon)" + }, + { + "bbox": [ + 313, + 596, + 556, + 669 + ], + "type": "text", + "content": ", corresponding to the choice " + }, + { + "bbox": [ + 313, + 596, + 556, + 669 + ], + "type": "inline_equation", + "content": "d_p(\\mathbf{x}, \\delta) \\coloneqq \\| \\delta \\|_p" + }, + { + "bbox": [ + 313, + 596, + 556, + 669 + ], + "type": "text", + "content": ". The threat function " + }, + { + "bbox": [ + 313, + 596, + 556, + 669 + ], + "type": "inline_equation", + "content": "d_p" + }, + { + "bbox": [ + 313, + 596, + 556, + 669 + ], + "type": "text", + "content": " is task-agnostic, easy to evaluate, and induces a compact sub-level set " + }, + { + "bbox": [ + 313, + 596, + 556, + 669 + ], + "type": "inline_equation", + "content": "S(\\mathbf{x}, d_p, \\varepsilon)" + }, + { + "bbox": [ + 313, + 596, + 556, + 669 + ], + "type": "text", + "content": " that allows efficient projection. Hence, the " + }, + { + "bbox": [ + 313, + 596, + 556, + 669 + ], + "type": "inline_equation", + "content": "\\ell_p" + }, + { + "bbox": [ + 313, + 596, + 556, + 669 + ], + "type": "text", + "content": "-threat model " + }, + { + "bbox": [ + 313, + 596, + 556, + 669 + ], + "type": "inline_equation", + "content": "(d_p, \\varepsilon)" + }, + { + "bbox": [ + 313, + 596, + 556, + 669 + ], + "type": "text", + "content": " presents a natural starting point for in" + } + ] + } + ], + "index": 10 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "spans": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "type": "text", + "content": "CVF" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 145, + 0, + 496, + 37 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 145, + 0, + 496, + 37 + ], + "spans": [ + { + "bbox": [ + 145, + 0, + 496, + 37 + ], + "type": "text", + "content": "This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 326, + 677, + 543, + 688 + ], + "type": "page_footnote", + "angle": 0, + "lines": [ + { + "bbox": [ + 326, + 677, + 543, + 688 + ], + "spans": [ + { + "bbox": [ + 326, + 677, + 543, + 688 + ], + "type": "text", + "content": "1The flexibility to define threat w.r.t. input " + }, + { + "bbox": [ + 326, + 677, + 543, + 688 + ], + "type": "inline_equation", + "content": "\\mathbf{x}" + }, + { + "bbox": [ + 326, + 677, + 543, + 688 + ], + "type": "text", + "content": " is critical to our work." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 315, + 688, + 553, + 715 + ], + "type": "page_footnote", + "angle": 0, + "lines": [ + { + "bbox": [ + 315, + 688, + 553, + 715 + ], + "spans": [ + { + "bbox": [ + 315, + 688, + 553, + 715 + ], + "type": "text", + "content": "2For any vector " + }, + { + "bbox": [ + 315, + 688, + 553, + 715 + ], + "type": "inline_equation", + "content": "\\pmb {\\delta}\\in \\mathbb{R}^d" + }, + { + "bbox": [ + 315, + 688, + 553, + 715 + ], + "type": "text", + "content": " , the " + }, + { + "bbox": [ + 315, + 688, + 553, + 715 + ], + "type": "inline_equation", + "content": "\\ell_p" + }, + { + "bbox": [ + 315, + 688, + 553, + 715 + ], + "type": "text", + "content": " norm " + }, + { + "bbox": [ + 315, + 688, + 553, + 715 + ], + "type": "inline_equation", + "content": "\\| \\cdot \\| _p" + }, + { + "bbox": [ + 315, + 688, + 553, + 715 + ], + "type": "text", + "content": " for " + }, + { + "bbox": [ + 315, + 688, + 553, + 715 + ], + "type": "inline_equation", + "content": "p\\geq 1" + }, + { + "bbox": [ + 315, + 688, + 553, + 715 + ], + "type": "text", + "content": " is defined as " + }, + { + "bbox": [ + 315, + 688, + 553, + 715 + ], + "type": "inline_equation", + "content": "\\| \\pmb {\\delta}\\| _p\\coloneqq \\left(\\sum_i|\\delta_i|^p\\right)^{\\frac{1}{p}}." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 66, + 693, + 245, + 703 + ], + "type": "page_footnote", + "angle": 0, + "lines": [ + { + "bbox": [ + 66, + 693, + 245, + 703 + ], + "spans": [ + { + "bbox": [ + 66, + 693, + 245, + 703 + ], + "type": "text", + "content": "* Corresponding author: rmuthukl@jhu.edu" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 67, + 703, + 239, + 713 + ], + "type": "page_footnote", + "angle": 0, + "lines": [ + { + "bbox": [ + 67, + 703, + 239, + 713 + ], + "spans": [ + { + "bbox": [ + 67, + 703, + 239, + 713 + ], + "type": "text", + "content": "This work is not related to AP's position at Amazon" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 295, + 750, + 315, + 758 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 750, + 315, + 758 + ], + "spans": [ + { + "bbox": [ + 295, + 750, + 315, + 758 + ], + "type": "text", + "content": "9954" + } + ] + } + ], + "index": 17 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 0 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 92, + 69, + 262, + 210 + ], + "blocks": [ + { + "bbox": [ + 92, + 69, + 262, + 210 + ], + "lines": [ + { + "bbox": [ + 92, + 69, + 262, + 210 + ], + "spans": [ + { + "bbox": [ + 92, + 69, + 262, + 210 + ], + "type": "image", + "image_path": "c3652cb8b0020bc5a6e618359cc909da7e2e65b34bf021d70c7774d12e0b2112.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 55, + 219, + 295, + 243 + ], + "lines": [ + { + "bbox": [ + 55, + 219, + 295, + 243 + ], + "spans": [ + { + "bbox": [ + 55, + 219, + 295, + 243 + ], + "type": "text", + "content": "Figure 1. Corruptions with equal " + }, + { + "bbox": [ + 55, + 219, + 295, + 243 + ], + "type": "inline_equation", + "content": "\\ell_{\\infty}" + }, + { + "bbox": [ + 55, + 219, + 295, + 243 + ], + "type": "text", + "content": "-threat, " + }, + { + "bbox": [ + 55, + 219, + 295, + 243 + ], + "type": "inline_equation", + "content": "\\|\\delta_1\\|_{\\infty} = \\|\\delta_2\\|_{\\infty} = \\|\\delta_3\\|_{\\infty} = \\|\\delta_4\\|_{\\infty}" + }, + { + "bbox": [ + 55, + 219, + 295, + 243 + ], + "type": "text", + "content": ", but varying PD-threat." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "bbox": [ + 55, + 265, + 296, + 373 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 265, + 296, + 373 + ], + "spans": [ + { + "bbox": [ + 55, + 265, + 296, + 373 + ], + "type": "text", + "content": "vestigating robustness. RobustBench [8] maintains an up-to-date leaderboard of the robust accuracies of benchmark models under " + }, + { + "bbox": [ + 55, + 265, + 296, + 373 + ], + "type": "inline_equation", + "content": "(d_p, \\varepsilon)" + }, + { + "bbox": [ + 55, + 265, + 296, + 373 + ], + "type": "text", + "content": ". Unfortunately, the progress towards achieving perfect adversarial robustness (i.e. " + }, + { + "bbox": [ + 55, + 265, + 296, + 373 + ], + "type": "inline_equation", + "content": "100\\%" + }, + { + "bbox": [ + 55, + 265, + 296, + 373 + ], + "type": "text", + "content": " robust test accuracy) in RobustBench has plateaued in recent years. Unlike supervised learning on clean data, scaling data, model size and computing resources might be insufficient to bridge the gap [4, 11]. Below we expand further on the fundamental limitations of isotropic and global threat models." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 55, + 383, + 277, + 396 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 383, + 277, + 396 + ], + "spans": [ + { + "bbox": [ + 55, + 383, + 277, + 396 + ], + "type": "text", + "content": "1.1. Motivation: Specification of Threat Model" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 55, + 401, + 296, + 652 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 401, + 296, + 652 + ], + "spans": [ + { + "bbox": [ + 55, + 401, + 296, + 652 + ], + "type": "text", + "content": "For image-based datasets, it is widely recognized that " + }, + { + "bbox": [ + 55, + 401, + 296, + 652 + ], + "type": "inline_equation", + "content": "\\ell_p" + }, + { + "bbox": [ + 55, + 401, + 296, + 652 + ], + "type": "text", + "content": " norms are neither necessary nor sufficient for capturing perceptual similarity [39, 40], which poses significant challenges for accurately evaluating robustness [26, 43]. In Figure 1, " + }, + { + "bbox": [ + 55, + 401, + 296, + 652 + ], + "type": "inline_equation", + "content": "\\mathbf{x}" + }, + { + "bbox": [ + 55, + 401, + 296, + 652 + ], + "type": "text", + "content": " (at the center) is an image of class ROSE HIP from the Imagenet-1k dataset, and " + }, + { + "bbox": [ + 55, + 401, + 296, + 652 + ], + "type": "inline_equation", + "content": "\\mathbf{x} + \\delta_1" + }, + { + "bbox": [ + 55, + 401, + 296, + 652 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 55, + 401, + 296, + 652 + ], + "type": "inline_equation", + "content": "\\mathbf{x} + \\delta_2" + }, + { + "bbox": [ + 55, + 401, + 296, + 652 + ], + "type": "text", + "content": " are two corrupted images equidistant (w.r.t. " + }, + { + "bbox": [ + 55, + 401, + 296, + 652 + ], + "type": "inline_equation", + "content": "\\ell_{\\infty}" + }, + { + "bbox": [ + 55, + 401, + 296, + 652 + ], + "type": "text", + "content": ") from " + }, + { + "bbox": [ + 55, + 401, + 296, + 652 + ], + "type": "inline_equation", + "content": "\\mathbf{x}" + }, + { + "bbox": [ + 55, + 401, + 296, + 652 + ], + "type": "text", + "content": ", i.e. " + }, + { + "bbox": [ + 55, + 401, + 296, + 652 + ], + "type": "inline_equation", + "content": "\\| \\delta_1 \\|_{\\infty} = \\| \\delta_2 \\|_{\\infty} = \\varepsilon" + }, + { + "bbox": [ + 55, + 401, + 296, + 652 + ], + "type": "text", + "content": ". Intuitively, both " + }, + { + "bbox": [ + 55, + 401, + 296, + 652 + ], + "type": "inline_equation", + "content": "\\mathbf{x} + \\delta_1" + }, + { + "bbox": [ + 55, + 401, + 296, + 652 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 55, + 401, + 296, + 652 + ], + "type": "inline_equation", + "content": "\\mathbf{x} + \\delta_2" + }, + { + "bbox": [ + 55, + 401, + 296, + 652 + ], + "type": "text", + "content": " share the same label as " + }, + { + "bbox": [ + 55, + 401, + 296, + 652 + ], + "type": "inline_equation", + "content": "\\mathbf{x}" + }, + { + "bbox": [ + 55, + 401, + 296, + 652 + ], + "type": "text", + "content": ", since its appearance still depicts the fruit ROSE HIP. We refer to such perturbations that preserve the class label as safe, while those that change it are termed unsafe. Consider now moving further along these directions to " + }, + { + "bbox": [ + 55, + 401, + 296, + 652 + ], + "type": "inline_equation", + "content": "\\mathbf{x} + t\\delta_1" + }, + { + "bbox": [ + 55, + 401, + 296, + 652 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 55, + 401, + 296, + 652 + ], + "type": "inline_equation", + "content": "\\mathbf{x} + t\\delta_2" + }, + { + "bbox": [ + 55, + 401, + 296, + 652 + ], + "type": "text", + "content": ", for " + }, + { + "bbox": [ + 55, + 401, + 296, + 652 + ], + "type": "inline_equation", + "content": "t > 1" + }, + { + "bbox": [ + 55, + 401, + 296, + 652 + ], + "type": "text", + "content": ". These perturbed points are again equidistant from " + }, + { + "bbox": [ + 55, + 401, + 296, + 652 + ], + "type": "inline_equation", + "content": "\\mathbf{x}" + }, + { + "bbox": [ + 55, + 401, + 296, + 652 + ], + "type": "text", + "content": ", but while " + }, + { + "bbox": [ + 55, + 401, + 296, + 652 + ], + "type": "inline_equation", + "content": "t\\delta_1" + }, + { + "bbox": [ + 55, + 401, + 296, + 652 + ], + "type": "text", + "content": " does not alter the true label (i.e. it is safe), " + }, + { + "bbox": [ + 55, + 401, + 296, + 652 + ], + "type": "inline_equation", + "content": "t\\delta_2" + }, + { + "bbox": [ + 55, + 401, + 296, + 652 + ], + "type": "text", + "content": " does (and it is unsafe). As a result, any threat model " + }, + { + "bbox": [ + 55, + 401, + 296, + 652 + ], + "type": "inline_equation", + "content": "(d_{\\infty}, t' \\varepsilon)" + }, + { + "bbox": [ + 55, + 401, + 296, + 652 + ], + "type": "text", + "content": " with " + }, + { + "bbox": [ + 55, + 401, + 296, + 652 + ], + "type": "inline_equation", + "content": "t' \\geq t" + }, + { + "bbox": [ + 55, + 401, + 296, + 652 + ], + "type": "text", + "content": " necessarily incurs misspecification: any classifier " + }, + { + "bbox": [ + 55, + 401, + 296, + 652 + ], + "type": "inline_equation", + "content": "h" + }, + { + "bbox": [ + 55, + 401, + 296, + 652 + ], + "type": "text", + "content": " stable to all perturbations in " + }, + { + "bbox": [ + 55, + 401, + 296, + 652 + ], + "type": "inline_equation", + "content": "S(\\mathbf{x}, d_{\\infty}, t' \\varepsilon)" + }, + { + "bbox": [ + 55, + 401, + 296, + 652 + ], + "type": "text", + "content": " produces an incorrect prediction at " + }, + { + "bbox": [ + 55, + 401, + 296, + 652 + ], + "type": "inline_equation", + "content": "\\mathbf{x} + t\\delta_2" + }, + { + "bbox": [ + 55, + 401, + 296, + 652 + ], + "type": "text", + "content": ", i.e. " + }, + { + "bbox": [ + 55, + 401, + 296, + 652 + ], + "type": "inline_equation", + "content": "h(\\mathbf{x} + t\\delta_2) = \\text{ROSE HIP}" + }, + { + "bbox": [ + 55, + 401, + 296, + 652 + ], + "type": "text", + "content": ". On the other hand, shrinking the threat model is also futile, as then the safe perturbation " + }, + { + "bbox": [ + 55, + 401, + 296, + 652 + ], + "type": "inline_equation", + "content": "t\\delta_1" + }, + { + "bbox": [ + 55, + 401, + 296, + 652 + ], + "type": "text", + "content": " is necessarily excluded." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 55, + 653, + 296, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 653, + 296, + 714 + ], + "spans": [ + { + "bbox": [ + 55, + 653, + 296, + 714 + ], + "type": "text", + "content": "This illustrates that standard " + }, + { + "bbox": [ + 55, + 653, + 296, + 714 + ], + "type": "inline_equation", + "content": "\\ell_p" + }, + { + "bbox": [ + 55, + 653, + 296, + 714 + ], + "type": "text", + "content": "-threat models are unable to distinguish between the safe perturbations that preserve the true label and the unsafe perturbations that alter the true label. Hence, perfect robust accuracy under " + }, + { + "bbox": [ + 55, + 653, + 296, + 714 + ], + "type": "inline_equation", + "content": "\\ell_p" + }, + { + "bbox": [ + 55, + 653, + 296, + 714 + ], + "type": "text", + "content": " threat models might be neither achievable nor desirable. Attempting to" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 313, + 72, + 555, + 133 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 72, + 555, + 133 + ], + "spans": [ + { + "bbox": [ + 313, + 72, + 555, + 133 + ], + "type": "text", + "content": "resolve these limitations, Laidlaw et al. [32] formulate perceptual threat models via neural perceptual metric based on neural representations. Unfortunately (but unsurprisingly), all neural perceptual distance metrics are themselves vulnerable to adversarial attacks [17-19, 29]." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 314, + 142, + 463, + 154 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 314, + 142, + 463, + 154 + ], + "spans": [ + { + "bbox": [ + 314, + 142, + 463, + 154 + ], + "type": "text", + "content": "1.2. Summary of Contributions" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 313, + 159, + 555, + 293 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 159, + 555, + 293 + ], + "spans": [ + { + "bbox": [ + 313, + 159, + 555, + 293 + ], + "type": "text", + "content": "We propose a novel threat function called Projected Displacement (PD), denoted by " + }, + { + "bbox": [ + 313, + 159, + 555, + 293 + ], + "type": "inline_equation", + "content": "d_{\\mathrm{PD}}" + }, + { + "bbox": [ + 313, + 159, + 555, + 293 + ], + "type": "text", + "content": " (see Definition 7 for an explicit description), that accounts for the local directional statistics of the data via unsafe directions3 " + }, + { + "bbox": [ + 313, + 159, + 555, + 293 + ], + "type": "inline_equation", + "content": "\\mathcal{U}(\\mathbf{x})" + }, + { + "bbox": [ + 313, + 159, + 555, + 293 + ], + "type": "text", + "content": " (Definition 6) at each input " + }, + { + "bbox": [ + 313, + 159, + 555, + 293 + ], + "type": "inline_equation", + "content": "\\mathbf{x}" + }, + { + "bbox": [ + 313, + 159, + 555, + 293 + ], + "type": "text", + "content": ". For an input " + }, + { + "bbox": [ + 313, + 159, + 555, + 293 + ], + "type": "inline_equation", + "content": "\\mathbf{x}" + }, + { + "bbox": [ + 313, + 159, + 555, + 293 + ], + "type": "text", + "content": " with label " + }, + { + "bbox": [ + 313, + 159, + 555, + 293 + ], + "type": "inline_equation", + "content": "y" + }, + { + "bbox": [ + 313, + 159, + 555, + 293 + ], + "type": "text", + "content": ", any input " + }, + { + "bbox": [ + 313, + 159, + 555, + 293 + ], + "type": "inline_equation", + "content": "\\tilde{\\mathbf{x}}" + }, + { + "bbox": [ + 313, + 159, + 555, + 293 + ], + "type": "text", + "content": " with label " + }, + { + "bbox": [ + 313, + 159, + 555, + 293 + ], + "type": "inline_equation", + "content": "\\tilde{y} \\neq y" + }, + { + "bbox": [ + 313, + 159, + 555, + 293 + ], + "type": "text", + "content": " represents the unsafe direction " + }, + { + "bbox": [ + 313, + 159, + 555, + 293 + ], + "type": "inline_equation", + "content": "\\mathbf{u} := \\frac{\\tilde{\\mathbf{x}} - \\mathbf{x}}{\\|\\tilde{\\mathbf{x}} - \\mathbf{x}\\|_2}" + }, + { + "bbox": [ + 313, + 159, + 555, + 293 + ], + "type": "text", + "content": ", which will be estimated from observed training data. Then, the degree of threat of each perturbation " + }, + { + "bbox": [ + 313, + 159, + 555, + 293 + ], + "type": "inline_equation", + "content": "\\delta" + }, + { + "bbox": [ + 313, + 159, + 555, + 293 + ], + "type": "text", + "content": " is the maximal alignment with the set of unsafe directions: A large threat " + }, + { + "bbox": [ + 313, + 159, + 555, + 293 + ], + "type": "inline_equation", + "content": "d_{\\mathrm{PD}}(\\mathbf{x},\\delta)" + }, + { + "bbox": [ + 313, + 159, + 555, + 293 + ], + "type": "text", + "content": " indicates that " + }, + { + "bbox": [ + 313, + 159, + 555, + 293 + ], + "type": "inline_equation", + "content": "\\delta" + }, + { + "bbox": [ + 313, + 159, + 555, + 293 + ], + "type": "text", + "content": " is well-aligned with an unsafe direction " + }, + { + "bbox": [ + 313, + 159, + 555, + 293 + ], + "type": "inline_equation", + "content": "\\mathbf{u} \\in \\mathcal{U}(\\mathbf{x})" + }, + { + "bbox": [ + 313, + 159, + 555, + 293 + ], + "type": "text", + "content": ". We highlight some key properties." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 315, + 293, + 556, + 640 + ], + "type": "list", + "angle": 0, + "index": 12, + "blocks": [ + { + "bbox": [ + 315, + 293, + 556, + 460 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 315, + 293, + 556, + 460 + ], + "spans": [ + { + "bbox": [ + 315, + 293, + 556, + 460 + ], + "type": "text", + "content": "1. Disentangling safe vs unsafe perturbations. PD-threat is able to distinguish between safe and unsafe perturbations of equal " + }, + { + "bbox": [ + 315, + 293, + 556, + 460 + ], + "type": "inline_equation", + "content": "\\ell_p" + }, + { + "bbox": [ + 315, + 293, + 556, + 460 + ], + "type": "text", + "content": " norms. From Figure 1, different corruptions (Gaussian noise " + }, + { + "bbox": [ + 315, + 293, + 556, + 460 + ], + "type": "inline_equation", + "content": "t\\delta_1" + }, + { + "bbox": [ + 315, + 293, + 556, + 460 + ], + "type": "text", + "content": ", motion blur " + }, + { + "bbox": [ + 315, + 293, + 556, + 460 + ], + "type": "inline_equation", + "content": "t\\delta_3" + }, + { + "bbox": [ + 315, + 293, + 556, + 460 + ], + "type": "text", + "content": " and saturation " + }, + { + "bbox": [ + 315, + 293, + 556, + 460 + ], + "type": "inline_equation", + "content": "t\\delta_4" + }, + { + "bbox": [ + 315, + 293, + 556, + 460 + ], + "type": "text", + "content": ") applied to the original image " + }, + { + "bbox": [ + 315, + 293, + 556, + 460 + ], + "type": "inline_equation", + "content": "\\mathbf{x}" + }, + { + "bbox": [ + 315, + 293, + 556, + 460 + ], + "type": "text", + "content": " all have equal " + }, + { + "bbox": [ + 315, + 293, + 556, + 460 + ], + "type": "inline_equation", + "content": "\\ell_{\\infty}" + }, + { + "bbox": [ + 315, + 293, + 556, + 460 + ], + "type": "text", + "content": " norm of 240/255. Despite this, the true label of these perturbed samples does not change, unlike that of corruption " + }, + { + "bbox": [ + 315, + 293, + 556, + 460 + ], + "type": "inline_equation", + "content": "t\\delta_2" + }, + { + "bbox": [ + 315, + 293, + 556, + 460 + ], + "type": "text", + "content": ", which results in an image " + }, + { + "bbox": [ + 315, + 293, + 556, + 460 + ], + "type": "inline_equation", + "content": "\\mathbf{x} + t\\delta_2" + }, + { + "bbox": [ + 315, + 293, + 556, + 460 + ], + "type": "text", + "content": " with a different class. For each perturbation " + }, + { + "bbox": [ + 315, + 293, + 556, + 460 + ], + "type": "inline_equation", + "content": "t\\delta_i" + }, + { + "bbox": [ + 315, + 293, + 556, + 460 + ], + "type": "text", + "content": ", the computed values of PD threat in Figure 1 naturally reflect how close they are to changing the class label. PD threat model is competent with state-of-the-art neural perceptual threat model DreamSim (denoted " + }, + { + "bbox": [ + 315, + 293, + 556, + 460 + ], + "type": "inline_equation", + "content": "d_{\\mathrm{DS}}" + }, + { + "bbox": [ + 315, + 293, + 556, + 460 + ], + "type": "text", + "content": ") [15] for distinguishing safe and unsafe perturbations (see Section 4)." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 315, + 460, + 556, + 555 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 315, + 460, + 556, + 555 + ], + "spans": [ + { + "bbox": [ + 315, + 460, + 556, + 555 + ], + "type": "text", + "content": "2. Ease of use. As we will show, the sub-level set " + }, + { + "bbox": [ + 315, + 460, + 556, + 555 + ], + "type": "inline_equation", + "content": "S(\\mathbf{x}, d_{\\mathrm{PD}}, \\varepsilon)" + }, + { + "bbox": [ + 315, + 460, + 556, + 555 + ], + "type": "text", + "content": " at each input " + }, + { + "bbox": [ + 315, + 460, + 556, + 555 + ], + "type": "inline_equation", + "content": "\\mathbf{x}" + }, + { + "bbox": [ + 315, + 460, + 556, + 555 + ], + "type": "text", + "content": " is convex, allowing for efficient projections onto it. This enables a straightforward plug-and-play mechanism for adapting existing adversarial attacks to the PD-threat model. We incorporate such an extension of AutoAttack [7] to evaluate the robustness under the PD-threat model (" + }, + { + "bbox": [ + 315, + 460, + 556, + 555 + ], + "type": "inline_equation", + "content": "d_{\\mathrm{PD}}, \\varepsilon" + }, + { + "bbox": [ + 315, + 460, + 556, + 555 + ], + "type": "text", + "content": ") for state-of-the-art robust classifiers registered at RobustBench [8]." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 315, + 556, + 556, + 640 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 315, + 556, + 556, + 640 + ], + "spans": [ + { + "bbox": [ + 315, + 556, + 556, + 640 + ], + "type": "text", + "content": "3. Task-dependency. PD-threat models can readily integrate additional task-relevant information via label annotation, such as inter-class hierarchy (e.g. the Wordnet hierarchy) or pixel annotation (e.g. segmentation masks)—see Table 1—allowing for variations in the form of robustness while retaining the above benefits (anisotropy, locality, convex sub-level sets, and efficient projection)." + } + ] + } + ], + "index": 11 + } + ], + "sub_type": "text" + }, + { + "bbox": [ + 315, + 650, + 402, + 661 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 315, + 650, + 402, + 661 + ], + "spans": [ + { + "bbox": [ + 315, + 650, + 402, + 661 + ], + "type": "text", + "content": "1.3. Related Work" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 313, + 668, + 555, + 692 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 668, + 555, + 692 + ], + "spans": [ + { + "bbox": [ + 313, + 668, + 555, + 692 + ], + "type": "text", + "content": "In this section, we highlight prior work that explores robustness beyond " + }, + { + "bbox": [ + 313, + 668, + 555, + 692 + ], + "type": "inline_equation", + "content": "\\ell_p" + }, + { + "bbox": [ + 313, + 668, + 555, + 692 + ], + "type": "text", + "content": "-threat models." + } + ] + } + ], + "index": 14 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 324, + 702, + 504, + 712 + ], + "type": "page_footnote", + "angle": 0, + "lines": [ + { + "bbox": [ + 324, + 702, + 504, + 712 + ], + "spans": [ + { + "bbox": [ + 324, + 702, + 504, + 712 + ], + "type": "text", + "content": "3Directions refer to vectors " + }, + { + "bbox": [ + 324, + 702, + 504, + 712 + ], + "type": "inline_equation", + "content": "\\mathbf{u}\\in \\mathbb{R}^d" + }, + { + "bbox": [ + 324, + 702, + 504, + 712 + ], + "type": "text", + "content": " with unit " + }, + { + "bbox": [ + 324, + 702, + 504, + 712 + ], + "type": "inline_equation", + "content": "\\ell_2" + }, + { + "bbox": [ + 324, + 702, + 504, + 712 + ], + "type": "text", + "content": " norm." + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 295, + 750, + 315, + 758 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 750, + 315, + 758 + ], + "spans": [ + { + "bbox": [ + 295, + 750, + 315, + 758 + ], + "type": "text", + "content": "9955" + } + ] + } + ], + "index": 16 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 1 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 56, + 70, + 294, + 125 + ], + "blocks": [ + { + "bbox": [ + 56, + 70, + 294, + 125 + ], + "lines": [ + { + "bbox": [ + 56, + 70, + 294, + 125 + ], + "spans": [ + { + "bbox": [ + 56, + 70, + 294, + 125 + ], + "type": "table", + "html": "
PROPERTYFEATURESdpdDS[15]dPD
Degree of ThreatAnisotropy, Locality×✓✓
Label AnnotationInter-class Distance×~✓
Pixel AnnotationSegmentation✓×✓
", + "image_path": "d7da485e19ce5281492dd24d0c0b0a4723f27000460e96412a8054fd79194af3.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "table_body" + } + ], + "index": 0 + }, + { + "bbox": [ + 103, + 133, + 249, + 144 + ], + "lines": [ + { + "bbox": [ + 103, + 133, + 249, + 144 + ], + "spans": [ + { + "bbox": [ + 103, + 133, + 249, + 144 + ], + "type": "text", + "content": "Table 1. Comparison of Threat Models" + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "text" + }, + { + "bbox": [ + 54, + 158, + 297, + 337 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 54, + 158, + 297, + 337 + ], + "spans": [ + { + "bbox": [ + 54, + 158, + 297, + 337 + ], + "type": "text", + "content": "Common Corruptions. Neural networks are vulnerable to image distortions (e.g. translation, rotation), common corruptions (e.g. blur, noise), changes in lighting, view, depth, etc. [12, 16, 25, 27, 28, 36, 50]. Note that these are known safe perturbations as they preserve the true label, even if they have large " + }, + { + "bbox": [ + 54, + 158, + 297, + 337 + ], + "type": "inline_equation", + "content": "\\ell_p" + }, + { + "bbox": [ + 54, + 158, + 297, + 337 + ], + "type": "text", + "content": " norms. To measure this vulnerability, [25] formalized the notion of corruption robustness and presented a standardized benchmark dataset, Imagenet-C containing 19 styles of common image corruptions (categorized into noise, blur, weather, and digital). Training on Imagenet-C, or on other augmentations, can result in over-fitting, particularly to specific types of corruptions [36]. Additionally, [36] introduced a newdataset, Imagenet-C, with 10 new corruption styles identified from a larger set of real-world corruptions. These will be useful later in our experiments." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 55, + 338, + 296, + 446 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 338, + 296, + 446 + ], + "spans": [ + { + "bbox": [ + 55, + 338, + 296, + 446 + ], + "type": "text", + "content": "Perceptual Distance Metrics. Since " + }, + { + "bbox": [ + 55, + 338, + 296, + 446 + ], + "type": "inline_equation", + "content": "\\ell_p" + }, + { + "bbox": [ + 55, + 338, + 296, + 446 + ], + "type": "text", + "content": " provides inaccurate approximations to perceptual distances, several works have attempted to propose alternatives, such as SSIM [47], FSIM [53], and HDR-VDP-2 [35]. With the advent of learning data-driven representations, [15, 54] hail the effectiveness of internal representations of neural classifiers in capturing perceptual similarity. DreamSim [15], the current state-of-the-art among the neural perceptual distance metrics, is finetuned on human perceptual similarity judgements." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 55, + 456, + 149, + 469 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 456, + 149, + 469 + ], + "spans": [ + { + "bbox": [ + 55, + 456, + 149, + 469 + ], + "type": "text", + "content": "2. Key Definitions" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 55, + 476, + 297, + 589 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 476, + 297, + 589 + ], + "spans": [ + { + "bbox": [ + 55, + 476, + 297, + 589 + ], + "type": "text", + "content": "Notation. We denote scalar quantities in Roman letters and vectors in boldface Roman letters. The empty set is given by " + }, + { + "bbox": [ + 55, + 476, + 297, + 589 + ], + "type": "inline_equation", + "content": "\\emptyset" + }, + { + "bbox": [ + 55, + 476, + 297, + 589 + ], + "type": "text", + "content": ". For any set " + }, + { + "bbox": [ + 55, + 476, + 297, + 589 + ], + "type": "inline_equation", + "content": "\\mathcal{V}" + }, + { + "bbox": [ + 55, + 476, + 297, + 589 + ], + "type": "text", + "content": ", we denote by " + }, + { + "bbox": [ + 55, + 476, + 297, + 589 + ], + "type": "inline_equation", + "content": "2^{\\mathcal{V}}" + }, + { + "bbox": [ + 55, + 476, + 297, + 589 + ], + "type": "text", + "content": " the power set of all its subsets. For any set " + }, + { + "bbox": [ + 55, + 476, + 297, + 589 + ], + "type": "inline_equation", + "content": "\\mathcal{V} \\subset \\mathbb{R}^d" + }, + { + "bbox": [ + 55, + 476, + 297, + 589 + ], + "type": "text", + "content": " and a vector " + }, + { + "bbox": [ + 55, + 476, + 297, + 589 + ], + "type": "inline_equation", + "content": "\\mathbf{v} \\in \\mathbb{R}^d" + }, + { + "bbox": [ + 55, + 476, + 297, + 589 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 55, + 476, + 297, + 589 + ], + "type": "inline_equation", + "content": "\\mathcal{V} - \\{\\mathbf{v}\\}" + }, + { + "bbox": [ + 55, + 476, + 297, + 589 + ], + "type": "text", + "content": " is the Minkowski difference, i.e. the set " + }, + { + "bbox": [ + 55, + 476, + 297, + 589 + ], + "type": "inline_equation", + "content": "\\{\\mathbf{w} - \\mathbf{v} \\mid \\forall \\mathbf{w} \\in \\mathcal{V}\\}" + }, + { + "bbox": [ + 55, + 476, + 297, + 589 + ], + "type": "text", + "content": " (subtracting " + }, + { + "bbox": [ + 55, + 476, + 297, + 589 + ], + "type": "inline_equation", + "content": "\\mathbf{v}" + }, + { + "bbox": [ + 55, + 476, + 297, + 589 + ], + "type": "text", + "content": " from all elements of " + }, + { + "bbox": [ + 55, + 476, + 297, + 589 + ], + "type": "inline_equation", + "content": "\\mathcal{V}" + }, + { + "bbox": [ + 55, + 476, + 297, + 589 + ], + "type": "text", + "content": "). We denote the unit ball and the unit sphere in " + }, + { + "bbox": [ + 55, + 476, + 297, + 589 + ], + "type": "inline_equation", + "content": "\\mathbb{R}^d" + }, + { + "bbox": [ + 55, + 476, + 297, + 589 + ], + "type": "text", + "content": " w.r.t. the " + }, + { + "bbox": [ + 55, + 476, + 297, + 589 + ], + "type": "inline_equation", + "content": "\\ell_p" + }, + { + "bbox": [ + 55, + 476, + 297, + 589 + ], + "type": "text", + "content": " norm by " + }, + { + "bbox": [ + 55, + 476, + 297, + 589 + ], + "type": "inline_equation", + "content": "\\mathbb{B}_p^d := \\{\\mathbf{v} \\in \\mathbb{R}^d \\mid \\| \\mathbf{v} \\|_p \\leq 1\\}" + }, + { + "bbox": [ + 55, + 476, + 297, + 589 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 55, + 476, + 297, + 589 + ], + "type": "inline_equation", + "content": "\\mathbb{S}_p^d := \\{\\mathbf{v} \\in \\mathbb{R}^d \\mid \\| \\mathbf{v} \\|_p = 1\\}" + }, + { + "bbox": [ + 55, + 476, + 297, + 589 + ], + "type": "text", + "content": ", respectively. A direction is an element of the unit sphere " + }, + { + "bbox": [ + 55, + 476, + 297, + 589 + ], + "type": "inline_equation", + "content": "\\mathbb{S}_2^d" + }, + { + "bbox": [ + 55, + 476, + 297, + 589 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 55, + 595, + 176, + 608 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 595, + 176, + 608 + ], + "spans": [ + { + "bbox": [ + 55, + 595, + 176, + 608 + ], + "type": "text", + "content": "2.1. Supervised Learning" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 55, + 613, + 296, + 685 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 613, + 296, + 685 + ], + "spans": [ + { + "bbox": [ + 55, + 613, + 296, + 685 + ], + "type": "text", + "content": "We consider a supervised learning task with a bounded input domain " + }, + { + "bbox": [ + 55, + 613, + 296, + 685 + ], + "type": "inline_equation", + "content": "\\mathcal{X} \\subset \\mathbb{R}^d" + }, + { + "bbox": [ + 55, + 613, + 296, + 685 + ], + "type": "text", + "content": " and labels " + }, + { + "bbox": [ + 55, + 613, + 296, + 685 + ], + "type": "inline_equation", + "content": "\\mathcal{Y} = \\{1, \\dots, C\\}" + }, + { + "bbox": [ + 55, + 613, + 296, + 685 + ], + "type": "text", + "content": " such that each input " + }, + { + "bbox": [ + 55, + 613, + 296, + 685 + ], + "type": "inline_equation", + "content": "\\mathbf{x}" + }, + { + "bbox": [ + 55, + 613, + 296, + 685 + ], + "type": "text", + "content": " is assigned a label by a deterministic true " + }, + { + "bbox": [ + 55, + 613, + 296, + 685 + ], + "type": "inline_equation", + "content": "^4" + }, + { + "bbox": [ + 55, + 613, + 296, + 685 + ], + "type": "text", + "content": " labeling function " + }, + { + "bbox": [ + 55, + 613, + 296, + 685 + ], + "type": "inline_equation", + "content": "h^\\star : \\mathcal{X} \\to \\mathcal{Y}" + }, + { + "bbox": [ + 55, + 613, + 296, + 685 + ], + "type": "text", + "content": ". The true labeling function " + }, + { + "bbox": [ + 55, + 613, + 296, + 685 + ], + "type": "inline_equation", + "content": "h^\\star" + }, + { + "bbox": [ + 55, + 613, + 296, + 685 + ], + "type": "text", + "content": " partitions the input domain " + }, + { + "bbox": [ + 55, + 613, + 296, + 685 + ], + "type": "inline_equation", + "content": "\\mathcal{X} := \\cup_y \\mathcal{X}_y" + }, + { + "bbox": [ + 55, + 613, + 296, + 685 + ], + "type": "text", + "content": ", where " + }, + { + "bbox": [ + 55, + 613, + 296, + 685 + ], + "type": "inline_equation", + "content": "\\mathcal{X}_y" + }, + { + "bbox": [ + 55, + 613, + 296, + 685 + ], + "type": "text", + "content": " is the set of inputs assigned label " + }, + { + "bbox": [ + 55, + 613, + 296, + 685 + ], + "type": "inline_equation", + "content": "y" + }, + { + "bbox": [ + 55, + 613, + 296, + 685 + ], + "type": "text", + "content": ", i.e. " + }, + { + "bbox": [ + 55, + 613, + 296, + 685 + ], + "type": "inline_equation", + "content": "\\mathcal{X}_y = \\{\\mathbf{x} \\in \\mathcal{X} \\mid h^\\star(\\mathbf{x}) = y\\}" + }, + { + "bbox": [ + 55, + 613, + 296, + 685 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 313, + 72, + 555, + 95 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 72, + 555, + 95 + ], + "spans": [ + { + "bbox": [ + 313, + 72, + 555, + 95 + ], + "type": "text", + "content": "We assume that the partition sets " + }, + { + "bbox": [ + 313, + 72, + 555, + 95 + ], + "type": "inline_equation", + "content": "\\mathcal{X}_1, \\ldots, \\mathcal{X}_C" + }, + { + "bbox": [ + 313, + 72, + 555, + 95 + ], + "type": "text", + "content": " are nonempty and open in order to avoid degeneracies." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 313, + 95, + 556, + 220 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 95, + 556, + 220 + ], + "spans": [ + { + "bbox": [ + 313, + 95, + 556, + 220 + ], + "type": "text", + "content": "We let " + }, + { + "bbox": [ + 313, + 95, + 556, + 220 + ], + "type": "inline_equation", + "content": "\\mathcal{D}_{\\mathcal{X}}" + }, + { + "bbox": [ + 313, + 95, + 556, + 220 + ], + "type": "text", + "content": " be a data distribution over the input domain " + }, + { + "bbox": [ + 313, + 95, + 556, + 220 + ], + "type": "inline_equation", + "content": "\\mathcal{X}" + }, + { + "bbox": [ + 313, + 95, + 556, + 220 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 313, + 95, + 556, + 220 + ], + "type": "inline_equation", + "content": "\\mathcal{D}_{\\mathcal{Z}}" + }, + { + "bbox": [ + 313, + 95, + 556, + 220 + ], + "type": "text", + "content": " the extension of " + }, + { + "bbox": [ + 313, + 95, + 556, + 220 + ], + "type": "inline_equation", + "content": "\\mathcal{D}_{\\mathcal{X}}" + }, + { + "bbox": [ + 313, + 95, + 556, + 220 + ], + "type": "text", + "content": " to a distribution over the joint domain " + }, + { + "bbox": [ + 313, + 95, + 556, + 220 + ], + "type": "inline_equation", + "content": "\\mathcal{Z} \\coloneqq \\mathcal{X} \\times \\mathcal{Y}" + }, + { + "bbox": [ + 313, + 95, + 556, + 220 + ], + "type": "text", + "content": " using the true labeling function " + }, + { + "bbox": [ + 313, + 95, + 556, + 220 + ], + "type": "inline_equation", + "content": "h^{\\star}" + }, + { + "bbox": [ + 313, + 95, + 556, + 220 + ], + "type": "text", + "content": ". The true labeling function " + }, + { + "bbox": [ + 313, + 95, + 556, + 220 + ], + "type": "inline_equation", + "content": "h^{\\star}" + }, + { + "bbox": [ + 313, + 95, + 556, + 220 + ], + "type": "text", + "content": " and the marginal input distribution " + }, + { + "bbox": [ + 313, + 95, + 556, + 220 + ], + "type": "inline_equation", + "content": "\\mathcal{D}_{\\mathcal{X}}" + }, + { + "bbox": [ + 313, + 95, + 556, + 220 + ], + "type": "text", + "content": " are unknown, but we observe finite labeled training data " + }, + { + "bbox": [ + 313, + 95, + 556, + 220 + ], + "type": "inline_equation", + "content": "\\mathrm{S} \\coloneqq \\left\\{(\\mathbf{x}^{(1)}, y^{(1)}), \\ldots, (\\mathbf{x}^{(m)}, y^{(m)})\\right\\} \\stackrel{\\mathrm{i.i.d.}}{\\sim} (\\mathcal{D}_{\\mathcal{Z}})^{m}" + }, + { + "bbox": [ + 313, + 95, + 556, + 220 + ], + "type": "text", + "content": ". For any classifier " + }, + { + "bbox": [ + 313, + 95, + 556, + 220 + ], + "type": "inline_equation", + "content": "h: \\mathcal{X} \\to \\mathcal{Y}" + }, + { + "bbox": [ + 313, + 95, + 556, + 220 + ], + "type": "text", + "content": ", accuracy is the probability of correctly labeling a random input, i.e. " + }, + { + "bbox": [ + 313, + 95, + 556, + 220 + ], + "type": "inline_equation", + "content": "\\mathrm{Acc}(h) \\coloneqq \\mathrm{Prob}_{\\mathbf{x} \\sim \\mathcal{D}_{\\mathcal{X}}} [h(\\mathbf{x}) = h^{\\star}(\\mathbf{x})]" + }, + { + "bbox": [ + 313, + 95, + 556, + 220 + ], + "type": "text", + "content": ". The task of supervised learning seeks a classifier " + }, + { + "bbox": [ + 313, + 95, + 556, + 220 + ], + "type": "inline_equation", + "content": "\\bar{h}" + }, + { + "bbox": [ + 313, + 95, + 556, + 220 + ], + "type": "text", + "content": " in a hypothesis class " + }, + { + "bbox": [ + 313, + 95, + 556, + 220 + ], + "type": "inline_equation", + "content": "\\mathcal{H} \\subset \\{h: \\mathcal{X} \\to \\mathcal{Y}\\}" + }, + { + "bbox": [ + 313, + 95, + 556, + 220 + ], + "type": "text", + "content": " with high accuracy." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 313, + 227, + 460, + 238 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 227, + 460, + 238 + ], + "spans": [ + { + "bbox": [ + 313, + 227, + 460, + 238 + ], + "type": "text", + "content": "2.2. Adversarial Perturbations" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 313, + 243, + 554, + 269 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 243, + 554, + 269 + ], + "spans": [ + { + "bbox": [ + 313, + 243, + 554, + 269 + ], + "type": "text", + "content": "In this article, we formally define adversarial perturbations through the lens of the partition sets " + }, + { + "bbox": [ + 313, + 243, + 554, + 269 + ], + "type": "inline_equation", + "content": "\\mathcal{X}_y" + }, + { + "bbox": [ + 313, + 243, + 554, + 269 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 313, + 275, + 554, + 300 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 275, + 554, + 300 + ], + "spans": [ + { + "bbox": [ + 313, + 275, + 554, + 300 + ], + "type": "text", + "content": "Definition 1 (Adversarial Perturbation). A perturbation " + }, + { + "bbox": [ + 313, + 275, + 554, + 300 + ], + "type": "inline_equation", + "content": "\\delta \\in \\mathbb{R}^d" + }, + { + "bbox": [ + 313, + 275, + 554, + 300 + ], + "type": "text", + "content": " is adversarial for predictor " + }, + { + "bbox": [ + 313, + 275, + 554, + 300 + ], + "type": "inline_equation", + "content": "h" + }, + { + "bbox": [ + 313, + 275, + 554, + 300 + ], + "type": "text", + "content": " at input " + }, + { + "bbox": [ + 313, + 275, + 554, + 300 + ], + "type": "inline_equation", + "content": "\\mathbf{x}" + }, + { + "bbox": [ + 313, + 275, + 554, + 300 + ], + "type": "text", + "content": " if it is," + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 314, + 300, + 555, + 360 + ], + "type": "list", + "angle": 0, + "index": 16, + "blocks": [ + { + "bbox": [ + 315, + 300, + 543, + 312 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 315, + 300, + 543, + 312 + ], + "spans": [ + { + "bbox": [ + 315, + 300, + 543, + 312 + ], + "type": "text", + "content": "1. Domain constrained: " + }, + { + "bbox": [ + 315, + 300, + 543, + 312 + ], + "type": "inline_equation", + "content": "\\mathbf{x} + \\delta \\in \\mathcal{X}" + }, + { + "bbox": [ + 315, + 300, + 543, + 312 + ], + "type": "text", + "content": ", i.e. " + }, + { + "bbox": [ + 315, + 300, + 543, + 312 + ], + "type": "inline_equation", + "content": "\\delta \\in \\mathcal{X} - \\{\\mathbf{x}\\}" + }, + { + "bbox": [ + 315, + 300, + 543, + 312 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 314, + 312, + 555, + 335 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 314, + 312, + 555, + 335 + ], + "spans": [ + { + "bbox": [ + 314, + 312, + 555, + 335 + ], + "type": "text", + "content": "2. Label invariant: " + }, + { + "bbox": [ + 314, + 312, + 555, + 335 + ], + "type": "inline_equation", + "content": "\\mathbf{x} + \\boldsymbol{\\delta}" + }, + { + "bbox": [ + 314, + 312, + 555, + 335 + ], + "type": "text", + "content": " has the same true label as " + }, + { + "bbox": [ + 314, + 312, + 555, + 335 + ], + "type": "inline_equation", + "content": "\\mathbf{x}" + }, + { + "bbox": [ + 314, + 312, + 555, + 335 + ], + "type": "text", + "content": ", i.e. " + }, + { + "bbox": [ + 314, + 312, + 555, + 335 + ], + "type": "inline_equation", + "content": "h^{\\star}(\\mathbf{x} + \\boldsymbol{\\delta}) = h^{\\star}(\\mathbf{x})" + }, + { + "bbox": [ + 314, + 312, + 555, + 335 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 314, + 335, + 554, + 360 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 314, + 335, + 554, + 360 + ], + "spans": [ + { + "bbox": [ + 314, + 335, + 554, + 360 + ], + "type": "text", + "content": "3. Adversarial: " + }, + { + "bbox": [ + 314, + 335, + 554, + 360 + ], + "type": "inline_equation", + "content": "\\mathbf{x} + \\boldsymbol{\\delta}" + }, + { + "bbox": [ + 314, + 335, + 554, + 360 + ], + "type": "text", + "content": " is misclassified by " + }, + { + "bbox": [ + 314, + 335, + 554, + 360 + ], + "type": "inline_equation", + "content": "h" + }, + { + "bbox": [ + 314, + 335, + 554, + 360 + ], + "type": "text", + "content": ", i.e. " + }, + { + "bbox": [ + 314, + 335, + 554, + 360 + ], + "type": "inline_equation", + "content": "h(\\mathbf{x} + \\boldsymbol{\\delta}) \\neq h^{\\star}(\\mathbf{x} + \\boldsymbol{\\delta})" + }, + { + "bbox": [ + 314, + 335, + 554, + 360 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 15 + } + ], + "sub_type": "text" + }, + { + "bbox": [ + 313, + 368, + 556, + 499 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 368, + 556, + 499 + ], + "spans": [ + { + "bbox": [ + 313, + 368, + 556, + 499 + ], + "type": "text", + "content": "The labels assigned by " + }, + { + "bbox": [ + 313, + 368, + 556, + 499 + ], + "type": "inline_equation", + "content": "h^{\\star}" + }, + { + "bbox": [ + 313, + 368, + 556, + 499 + ], + "type": "text", + "content": " are considered the ground truth and thus " + }, + { + "bbox": [ + 313, + 368, + 556, + 499 + ], + "type": "inline_equation", + "content": "h^{\\star}" + }, + { + "bbox": [ + 313, + 368, + 556, + 499 + ], + "type": "text", + "content": " has no adversarial perturbation at any input. The presence of adversarial perturbations indicates an imperfect labeling function " + }, + { + "bbox": [ + 313, + 368, + 556, + 499 + ], + "type": "inline_equation", + "content": "h" + }, + { + "bbox": [ + 313, + 368, + 556, + 499 + ], + "type": "text", + "content": ". Based on Definition 1, perturbations " + }, + { + "bbox": [ + 313, + 368, + 556, + 499 + ], + "type": "inline_equation", + "content": "\\delta \\in \\mathbb{R}^{d}" + }, + { + "bbox": [ + 313, + 368, + 556, + 499 + ], + "type": "text", + "content": " such that " + }, + { + "bbox": [ + 313, + 368, + 556, + 499 + ], + "type": "inline_equation", + "content": "\\mathbf{x} + \\delta \\notin \\mathcal{X}" + }, + { + "bbox": [ + 313, + 368, + 556, + 499 + ], + "type": "text", + "content": " are not deemed adversarial as the classifier " + }, + { + "bbox": [ + 313, + 368, + 556, + 499 + ], + "type": "inline_equation", + "content": "h" + }, + { + "bbox": [ + 313, + 368, + 556, + 499 + ], + "type": "text", + "content": " is not required to predict labels outside the input domain. On the other hand, for " + }, + { + "bbox": [ + 313, + 368, + 556, + 499 + ], + "type": "inline_equation", + "content": "\\mathbf{x}" + }, + { + "bbox": [ + 313, + 368, + 556, + 499 + ], + "type": "text", + "content": " with true label " + }, + { + "bbox": [ + 313, + 368, + 556, + 499 + ], + "type": "inline_equation", + "content": "y" + }, + { + "bbox": [ + 313, + 368, + 556, + 499 + ], + "type": "text", + "content": ", a perturbation " + }, + { + "bbox": [ + 313, + 368, + 556, + 499 + ], + "type": "inline_equation", + "content": "\\delta" + }, + { + "bbox": [ + 313, + 368, + 556, + 499 + ], + "type": "text", + "content": " such that " + }, + { + "bbox": [ + 313, + 368, + 556, + 499 + ], + "type": "inline_equation", + "content": "h^{\\star}(\\mathbf{x} + \\delta) \\neq y" + }, + { + "bbox": [ + 313, + 368, + 556, + 499 + ], + "type": "text", + "content": " cannot be adversarial at " + }, + { + "bbox": [ + 313, + 368, + 556, + 499 + ], + "type": "inline_equation", + "content": "\\mathbf{x}" + }, + { + "bbox": [ + 313, + 368, + 556, + 499 + ], + "type": "text", + "content": " as the corrupted input " + }, + { + "bbox": [ + 313, + 368, + 556, + 499 + ], + "type": "inline_equation", + "content": "\\mathbf{x} + \\delta" + }, + { + "bbox": [ + 313, + 368, + 556, + 499 + ], + "type": "text", + "content": " has a different true label. Finally, if " + }, + { + "bbox": [ + 313, + 368, + 556, + 499 + ], + "type": "inline_equation", + "content": "h" + }, + { + "bbox": [ + 313, + 368, + 556, + 499 + ], + "type": "text", + "content": " predicts the label at " + }, + { + "bbox": [ + 313, + 368, + 556, + 499 + ], + "type": "inline_equation", + "content": "\\mathbf{x}" + }, + { + "bbox": [ + 313, + 368, + 556, + 499 + ], + "type": "text", + "content": " incorrectly, then " + }, + { + "bbox": [ + 313, + 368, + 556, + 499 + ], + "type": "inline_equation", + "content": "\\mathbf{0}" + }, + { + "bbox": [ + 313, + 368, + 556, + 499 + ], + "type": "text", + "content": " is already an adversarial perturbation." + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 313, + 506, + 547, + 517 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 506, + 547, + 517 + ], + "spans": [ + { + "bbox": [ + 313, + 506, + 547, + 517 + ], + "type": "text", + "content": "2.2.1. Adversarial Robustness via Threat Specification" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 313, + 520, + 554, + 581 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 520, + 554, + 581 + ], + "spans": [ + { + "bbox": [ + 313, + 520, + 554, + 581 + ], + "type": "text", + "content": "A threat function " + }, + { + "bbox": [ + 313, + 520, + 554, + 581 + ], + "type": "inline_equation", + "content": "d: \\mathcal{X} \\times \\mathbb{R}^d \\to \\mathbb{R}^{\\geq 0}" + }, + { + "bbox": [ + 313, + 520, + 554, + 581 + ], + "type": "text", + "content": " presents practitioners a way to specify the kind of robustness suitable for a particular learning task. The " + }, + { + "bbox": [ + 313, + 520, + 554, + 581 + ], + "type": "inline_equation", + "content": "\\varepsilon" + }, + { + "bbox": [ + 313, + 520, + 554, + 581 + ], + "type": "text", + "content": "-sublevel set of a threat function " + }, + { + "bbox": [ + 313, + 520, + 554, + 581 + ], + "type": "inline_equation", + "content": "d" + }, + { + "bbox": [ + 313, + 520, + 554, + 581 + ], + "type": "text", + "content": " at an input " + }, + { + "bbox": [ + 313, + 520, + 554, + 581 + ], + "type": "inline_equation", + "content": "\\mathbf{x}" + }, + { + "bbox": [ + 313, + 520, + 554, + 581 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 313, + 520, + 554, + 581 + ], + "type": "inline_equation", + "content": "S(\\mathbf{x}, d, \\varepsilon) := \\{\\delta \\in \\mathbb{R}^d \\mid d(\\mathbf{x}, \\delta) \\leq \\varepsilon\\}" + }, + { + "bbox": [ + 313, + 520, + 554, + 581 + ], + "type": "text", + "content": ", defines the set of perturbations with a threat at most " + }, + { + "bbox": [ + 313, + 520, + 554, + 581 + ], + "type": "inline_equation", + "content": "\\varepsilon" + }, + { + "bbox": [ + 313, + 520, + 554, + 581 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 313, + 589, + 556, + 686 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 589, + 556, + 686 + ], + "spans": [ + { + "bbox": [ + 313, + 589, + 556, + 686 + ], + "type": "text", + "content": "Definition 2 (Threat Model, Robust Accuracy). A threat model " + }, + { + "bbox": [ + 313, + 589, + 556, + 686 + ], + "type": "inline_equation", + "content": "(d,\\varepsilon)" + }, + { + "bbox": [ + 313, + 589, + 556, + 686 + ], + "type": "text", + "content": " is a pair of threat function " + }, + { + "bbox": [ + 313, + 589, + 556, + 686 + ], + "type": "inline_equation", + "content": "d" + }, + { + "bbox": [ + 313, + 589, + 556, + 686 + ], + "type": "text", + "content": " and permissible threshold " + }, + { + "bbox": [ + 313, + 589, + 556, + 686 + ], + "type": "inline_equation", + "content": "\\varepsilon" + }, + { + "bbox": [ + 313, + 589, + 556, + 686 + ], + "type": "text", + "content": " that together define the set of permissible perturbations at each input " + }, + { + "bbox": [ + 313, + 589, + 556, + 686 + ], + "type": "inline_equation", + "content": "\\mathbf{x}" + }, + { + "bbox": [ + 313, + 589, + 556, + 686 + ], + "type": "text", + "content": " as the sublevel set " + }, + { + "bbox": [ + 313, + 589, + 556, + 686 + ], + "type": "inline_equation", + "content": "S(\\mathbf{x},d,\\varepsilon)" + }, + { + "bbox": [ + 313, + 589, + 556, + 686 + ], + "type": "text", + "content": ". Under the threat model " + }, + { + "bbox": [ + 313, + 589, + 556, + 686 + ], + "type": "inline_equation", + "content": "(d,\\varepsilon)" + }, + { + "bbox": [ + 313, + 589, + 556, + 686 + ], + "type": "text", + "content": ", the robust accuracy RobustAcc " + }, + { + "bbox": [ + 313, + 589, + 556, + 686 + ], + "type": "inline_equation", + "content": "(h,(d,\\varepsilon))" + }, + { + "bbox": [ + 313, + 589, + 556, + 686 + ], + "type": "text", + "content": " of a classifier " + }, + { + "bbox": [ + 313, + 589, + 556, + 686 + ], + "type": "inline_equation", + "content": "h\\in \\mathcal{H}" + }, + { + "bbox": [ + 313, + 589, + 556, + 686 + ], + "type": "text", + "content": " is the probability of stable prediction upon corruptions within the permissible set, i.e. " + }, + { + "bbox": [ + 313, + 589, + 556, + 686 + ], + "type": "inline_equation", + "content": "\\mathrm{Prob}_{\\mathbf{x}}\\big[\\forall \\delta \\in S(\\mathbf{x},d,\\varepsilon), h(\\mathbf{x} + \\delta) = h^{\\star}(\\mathbf{x})\\big]" + }, + { + "bbox": [ + 313, + 589, + 556, + 686 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 20 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 313, + 693, + 556, + 713 + ], + "type": "page_footnote", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 693, + 556, + 713 + ], + "spans": [ + { + "bbox": [ + 313, + 693, + 556, + 713 + ], + "type": "inline_equation", + "content": "{}^{5}" + }, + { + "bbox": [ + 313, + 693, + 556, + 713 + ], + "type": "text", + "content": " Sampling " + }, + { + "bbox": [ + 313, + 693, + 556, + 713 + ], + "type": "inline_equation", + "content": "\\mathbf{z} \\sim {\\mathcal{D}}_{\\mathcal{Z}}" + }, + { + "bbox": [ + 313, + 693, + 556, + 713 + ], + "type": "text", + "content": " is equivalent to sampling(x,h*(x))where " + }, + { + "bbox": [ + 313, + 693, + 556, + 713 + ], + "type": "inline_equation", + "content": "\\mathbf{x} \\sim" + }, + { + "bbox": [ + 313, + 693, + 556, + 713 + ], + "type": "inline_equation", + "content": "{\\mathcal{D}}_{\\mathcal{X}}" + }, + { + "bbox": [ + 313, + 693, + 556, + 713 + ], + "type": "text", + "content": " ." + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 55, + 693, + 295, + 713 + ], + "type": "page_footnote", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 693, + 295, + 713 + ], + "spans": [ + { + "bbox": [ + 55, + 693, + 295, + 713 + ], + "type": "text", + "content": "4Realizability in the context of adversarial robustness has also been studied in Awasthi et al. [2], Tramér et al. [43]." + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 295, + 750, + 315, + 758 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 750, + 315, + 758 + ], + "spans": [ + { + "bbox": [ + 295, + 750, + 315, + 758 + ], + "type": "text", + "content": "9956" + } + ] + } + ], + "index": 23 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 2 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 92, + 68, + 262, + 216 + ], + "blocks": [ + { + "bbox": [ + 92, + 68, + 262, + 216 + ], + "lines": [ + { + "bbox": [ + 92, + 68, + 262, + 216 + ], + "spans": [ + { + "bbox": [ + 92, + 68, + 262, + 216 + ], + "type": "image", + "image_path": "729f79c773047969e68a073fc968d08f975b0727244bac0404b0de377195067e.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 55, + 224, + 297, + 246 + ], + "lines": [ + { + "bbox": [ + 55, + 224, + 297, + 246 + ], + "spans": [ + { + "bbox": [ + 55, + 224, + 297, + 246 + ], + "type": "text", + "content": "Figure 2. An illustration of unsafe directions, and sub-level sets of the PD-threat." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "bbox": [ + 55, + 268, + 296, + 316 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 268, + 296, + 316 + ], + "spans": [ + { + "bbox": [ + 55, + 268, + 296, + 316 + ], + "type": "text", + "content": "Thus we define robustness as the stability of a classifier's prediction under perturbations within the permissible sets. The task of robust supervised learning seeks a classifier in " + }, + { + "bbox": [ + 55, + 268, + 296, + 316 + ], + "type": "inline_equation", + "content": "\\mathcal{H}" + }, + { + "bbox": [ + 55, + 268, + 296, + 316 + ], + "type": "text", + "content": " with high robust accuracy." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 55, + 324, + 296, + 361 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 324, + 296, + 361 + ], + "spans": [ + { + "bbox": [ + 55, + 324, + 296, + 361 + ], + "type": "text", + "content": "Definition 3 (" + }, + { + "bbox": [ + 55, + 324, + 296, + 361 + ], + "type": "inline_equation", + "content": "\\varepsilon" + }, + { + "bbox": [ + 55, + 324, + 296, + 361 + ], + "type": "text", + "content": "-robust). A classifier " + }, + { + "bbox": [ + 55, + 324, + 296, + 361 + ], + "type": "inline_equation", + "content": "h \\in \\mathcal{H}" + }, + { + "bbox": [ + 55, + 324, + 296, + 361 + ], + "type": "text", + "content": " is " + }, + { + "bbox": [ + 55, + 324, + 296, + 361 + ], + "type": "inline_equation", + "content": "\\varepsilon" + }, + { + "bbox": [ + 55, + 324, + 296, + 361 + ], + "type": "text", + "content": "-robust at " + }, + { + "bbox": [ + 55, + 324, + 296, + 361 + ], + "type": "inline_equation", + "content": "\\mathbf{x}" + }, + { + "bbox": [ + 55, + 324, + 296, + 361 + ], + "type": "text", + "content": " w.r.t. a threat function " + }, + { + "bbox": [ + 55, + 324, + 296, + 361 + ], + "type": "inline_equation", + "content": "d" + }, + { + "bbox": [ + 55, + 324, + 296, + 361 + ], + "type": "text", + "content": " if, " + }, + { + "bbox": [ + 55, + 324, + 296, + 361 + ], + "type": "inline_equation", + "content": "\\forall \\delta \\in S(\\mathbf{x}, d, \\varepsilon) \\cap (\\mathcal{X} - \\{\\mathbf{x}\\})" + }, + { + "bbox": [ + 55, + 324, + 296, + 361 + ], + "type": "text", + "content": ", we have " + }, + { + "bbox": [ + 55, + 324, + 296, + 361 + ], + "type": "inline_equation", + "content": "h(\\mathbf{x} + \\delta) = h^{\\star}(\\mathbf{x})" + }, + { + "bbox": [ + 55, + 324, + 296, + 361 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 55, + 369, + 296, + 453 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 369, + 296, + 453 + ], + "spans": [ + { + "bbox": [ + 55, + 369, + 296, + 453 + ], + "type": "text", + "content": "An " + }, + { + "bbox": [ + 55, + 369, + 296, + 453 + ], + "type": "inline_equation", + "content": "\\varepsilon" + }, + { + "bbox": [ + 55, + 369, + 296, + 453 + ], + "type": "text", + "content": "-robust classifier predicts the same label " + }, + { + "bbox": [ + 55, + 369, + 296, + 453 + ], + "type": "inline_equation", + "content": "h^{\\star}(\\mathbf{x})" + }, + { + "bbox": [ + 55, + 369, + 296, + 453 + ], + "type": "text", + "content": " at " + }, + { + "bbox": [ + 55, + 369, + 296, + 453 + ], + "type": "inline_equation", + "content": "\\mathbf{x} + \\delta" + }, + { + "bbox": [ + 55, + 369, + 296, + 453 + ], + "type": "text", + "content": " for any corruption " + }, + { + "bbox": [ + 55, + 369, + 296, + 453 + ], + "type": "inline_equation", + "content": "\\delta" + }, + { + "bbox": [ + 55, + 369, + 296, + 453 + ], + "type": "text", + "content": " within the " + }, + { + "bbox": [ + 55, + 369, + 296, + 453 + ], + "type": "inline_equation", + "content": "\\varepsilon" + }, + { + "bbox": [ + 55, + 369, + 296, + 453 + ], + "type": "text", + "content": "-sublevel set " + }, + { + "bbox": [ + 55, + 369, + 296, + 453 + ], + "type": "inline_equation", + "content": "S(\\mathbf{x}, d, \\varepsilon)" + }, + { + "bbox": [ + 55, + 369, + 296, + 453 + ], + "type": "text", + "content": ". We emphasize that robust accuracy and " + }, + { + "bbox": [ + 55, + 369, + 296, + 453 + ], + "type": "inline_equation", + "content": "\\varepsilon" + }, + { + "bbox": [ + 55, + 369, + 296, + 453 + ], + "type": "text", + "content": "-robustness are measures of stability rather than correctness. In this way, each threat model " + }, + { + "bbox": [ + 55, + 369, + 296, + 453 + ], + "type": "inline_equation", + "content": "(d, \\varepsilon)" + }, + { + "bbox": [ + 55, + 369, + 296, + 453 + ], + "type": "text", + "content": " implicitly encodes a trade-off between stability and correctness. [13]. Requiring more stability than necessary can lead to incorrect predictions as in Figure 1." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 55, + 464, + 261, + 479 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 464, + 261, + 479 + ], + "spans": [ + { + "bbox": [ + 55, + 464, + 261, + 479 + ], + "type": "text", + "content": "3. Threat specification beyond " + }, + { + "bbox": [ + 55, + 464, + 261, + 479 + ], + "type": "inline_equation", + "content": "\\ell_p" + }, + { + "bbox": [ + 55, + 464, + 261, + 479 + ], + "type": "text", + "content": " norms" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 55, + 485, + 296, + 568 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 485, + 296, + 568 + ], + "spans": [ + { + "bbox": [ + 55, + 485, + 296, + 568 + ], + "type": "text", + "content": "In this section, we design a principled threat model from first principles that is local and anisotropic. We first theoretically develop our exact PD\\*-threat model assuming knowledge of the decision boundaries of the true labeling function " + }, + { + "bbox": [ + 55, + 485, + 296, + 568 + ], + "type": "inline_equation", + "content": "h^\\star" + }, + { + "bbox": [ + 55, + 485, + 296, + 568 + ], + "type": "text", + "content": " (Section 3.1), and later relax this assumption by developing an approximation leveraging the observed training data and obtaining a practical PD-threat (Section 3.2)." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 55, + 577, + 253, + 590 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 577, + 253, + 590 + ], + "spans": [ + { + "bbox": [ + 55, + 577, + 253, + 590 + ], + "type": "text", + "content": "3.1. Measuring threat with class partition" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 55, + 594, + 296, + 643 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 594, + 296, + 643 + ], + "spans": [ + { + "bbox": [ + 55, + 594, + 296, + 643 + ], + "type": "text", + "content": "Definition 4 (Unsafe Directions). At each input " + }, + { + "bbox": [ + 55, + 594, + 296, + 643 + ], + "type": "inline_equation", + "content": "\\mathbf{x}" + }, + { + "bbox": [ + 55, + 594, + 296, + 643 + ], + "type": "text", + "content": ", a direction " + }, + { + "bbox": [ + 55, + 594, + 296, + 643 + ], + "type": "inline_equation", + "content": "\\mathbf{u} \\in \\mathbb{S}_2^d" + }, + { + "bbox": [ + 55, + 594, + 296, + 643 + ], + "type": "text", + "content": " is called unsafe if there exists a step size " + }, + { + "bbox": [ + 55, + 594, + 296, + 643 + ], + "type": "inline_equation", + "content": "t \\geq 0" + }, + { + "bbox": [ + 55, + 594, + 296, + 643 + ], + "type": "text", + "content": " such that " + }, + { + "bbox": [ + 55, + 594, + 296, + 643 + ], + "type": "inline_equation", + "content": "\\mathbf{x} + t\\mathbf{u} \\in \\mathcal{X}" + }, + { + "bbox": [ + 55, + 594, + 296, + 643 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 55, + 594, + 296, + 643 + ], + "type": "inline_equation", + "content": "h^\\star(\\mathbf{x} + t\\mathbf{u}) \\neq h^\\star(\\mathbf{x})" + }, + { + "bbox": [ + 55, + 594, + 296, + 643 + ], + "type": "text", + "content": ". We denote the set of all unsafe directions at " + }, + { + "bbox": [ + 55, + 594, + 296, + 643 + ], + "type": "inline_equation", + "content": "\\mathbf{x}" + }, + { + "bbox": [ + 55, + 594, + 296, + 643 + ], + "type": "text", + "content": " as " + }, + { + "bbox": [ + 55, + 594, + 296, + 643 + ], + "type": "inline_equation", + "content": "\\mathcal{U}^\\star(\\mathbf{x})" + }, + { + "bbox": [ + 55, + 594, + 296, + 643 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 55, + 651, + 296, + 676 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 651, + 296, + 676 + ], + "spans": [ + { + "bbox": [ + 55, + 651, + 296, + 676 + ], + "type": "text", + "content": "The minimum step size needed to alter the true label can vary across unsafe directions. Further, " + }, + { + "bbox": [ + 55, + 651, + 296, + 676 + ], + "type": "inline_equation", + "content": "\\mathcal{U}^{\\star}(\\mathbf{x})" + }, + { + "bbox": [ + 55, + 651, + 296, + 676 + ], + "type": "text", + "content": " varies" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 313, + 72, + 555, + 156 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 72, + 555, + 156 + ], + "spans": [ + { + "bbox": [ + 313, + 72, + 555, + 156 + ], + "type": "text", + "content": "with input. Directions " + }, + { + "bbox": [ + 313, + 72, + 555, + 156 + ], + "type": "inline_equation", + "content": "\\mathbf{v} \\in \\mathcal{B}_2^d \\cap (\\mathcal{U}^\\star(\\mathbf{x}))^c" + }, + { + "bbox": [ + 313, + 72, + 555, + 156 + ], + "type": "text", + "content": " are called safe, since moving along them (whilst domain constrained) cannot alter the true label. We note that adversarial perturbations (Definition 1) are a characteristic of a learned predictor " + }, + { + "bbox": [ + 313, + 72, + 555, + 156 + ], + "type": "inline_equation", + "content": "h \\in \\mathcal{H}" + }, + { + "bbox": [ + 313, + 72, + 555, + 156 + ], + "type": "text", + "content": " while unsafe directions are a characteristic of the true labeling function " + }, + { + "bbox": [ + 313, + 72, + 555, + 156 + ], + "type": "inline_equation", + "content": "h^\\star" + }, + { + "bbox": [ + 313, + 72, + 555, + 156 + ], + "type": "text", + "content": ". We propose to measure threat of a perturbations " + }, + { + "bbox": [ + 313, + 72, + 555, + 156 + ], + "type": "inline_equation", + "content": "\\delta" + }, + { + "bbox": [ + 313, + 72, + 555, + 156 + ], + "type": "text", + "content": " using " + }, + { + "bbox": [ + 313, + 72, + 555, + 156 + ], + "type": "inline_equation", + "content": "\\mathcal{U}^\\star(\\mathbf{x})" + }, + { + "bbox": [ + 313, + 72, + 555, + 156 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 313, + 163, + 555, + 224 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 163, + 555, + 224 + ], + "spans": [ + { + "bbox": [ + 313, + 163, + 555, + 224 + ], + "type": "text", + "content": "Definition 5 (PD\\* -threat). Let " + }, + { + "bbox": [ + 313, + 163, + 555, + 224 + ], + "type": "inline_equation", + "content": "\\mathbf{x}\\in \\mathcal{X}" + }, + { + "bbox": [ + 313, + 163, + 555, + 224 + ], + "type": "text", + "content": " and let " + }, + { + "bbox": [ + 313, + 163, + 555, + 224 + ], + "type": "inline_equation", + "content": "\\delta \\in \\mathbb{R}^d" + }, + { + "bbox": [ + 313, + 163, + 555, + 224 + ], + "type": "text", + "content": " be a perturbation. The exact projected displacement threat function " + }, + { + "bbox": [ + 313, + 163, + 555, + 224 + ], + "type": "inline_equation", + "content": "d_{PD}^{\\star}" + }, + { + "bbox": [ + 313, + 163, + 555, + 224 + ], + "type": "text", + "content": " is defined as the maximum scaled displacement of the Euclidean projection of the perturbation " + }, + { + "bbox": [ + 313, + 163, + 555, + 224 + ], + "type": "inline_equation", + "content": "\\delta" + }, + { + "bbox": [ + 313, + 163, + 555, + 224 + ], + "type": "text", + "content": " over all unsafe directions," + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 327, + 232, + 555, + 259 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 327, + 232, + 555, + 259 + ], + "spans": [ + { + "bbox": [ + 327, + 232, + 555, + 259 + ], + "type": "interline_equation", + "content": "d _ {\\mathrm {P D}} ^ {\\star} (\\mathbf {x}, \\boldsymbol {\\delta}) := \\sup _ {\\mathbf {u} \\in \\mathcal {U} ^ {\\star} (\\mathbf {x})} \\frac {1}{g ^ {\\star} (\\mathbf {x} , \\mathbf {u})} \\max \\left(\\langle \\boldsymbol {\\delta}, \\mathbf {u} \\rangle , 0\\right) \\quad (1)", + "image_path": "c7ce28ddf2103941e5c1f1865b20277e92315935381ae69ab1ec620b3e391854.jpg" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 313, + 268, + 525, + 281 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 268, + 525, + 281 + ], + "spans": [ + { + "bbox": [ + 313, + 268, + 525, + 281 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 313, + 268, + 525, + 281 + ], + "type": "inline_equation", + "content": "g^{\\star}(\\mathbf{u},\\mathbf{x})" + }, + { + "bbox": [ + 313, + 268, + 525, + 281 + ], + "type": "text", + "content": " is called the normalization function," + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 320, + 289, + 547, + 310 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 320, + 289, + 547, + 310 + ], + "spans": [ + { + "bbox": [ + 320, + 289, + 547, + 310 + ], + "type": "interline_equation", + "content": "g ^ {\\star} (\\mathbf {x}, \\mathbf {u}) := \\sup _ {\\mathbb {R} ^ {\\geq 0}} M \\text {s . t .} \\mathbf {x} + t \\mathbf {u} \\in \\mathcal {X} _ {h ^ {\\star} (\\mathbf {x})} \\forall t \\in [ 0, M ].", + "image_path": "c5c61f88dbbaca682c8afc6fa4a9c012d4ec92b03b07b326e81d3617de6cd14c.jpg" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 313, + 319, + 555, + 403 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 319, + 555, + 403 + ], + "spans": [ + { + "bbox": [ + 313, + 319, + 555, + 403 + ], + "type": "text", + "content": "At each " + }, + { + "bbox": [ + 313, + 319, + 555, + 403 + ], + "type": "inline_equation", + "content": "\\mathbf{x}" + }, + { + "bbox": [ + 313, + 319, + 555, + 403 + ], + "type": "text", + "content": ", the normalization function " + }, + { + "bbox": [ + 313, + 319, + 555, + 403 + ], + "type": "inline_equation", + "content": "g^{*}(\\mathbf{x},\\mathbf{u})" + }, + { + "bbox": [ + 313, + 319, + 555, + 403 + ], + "type": "text", + "content": " captures the distance along the direction " + }, + { + "bbox": [ + 313, + 319, + 555, + 403 + ], + "type": "inline_equation", + "content": "\\mathbf{u}" + }, + { + "bbox": [ + 313, + 319, + 555, + 403 + ], + "type": "text", + "content": " to the boundary of the set " + }, + { + "bbox": [ + 313, + 319, + 555, + 403 + ], + "type": "inline_equation", + "content": "\\mathcal{X}_{h^{\\star}(\\mathbf{x})}" + }, + { + "bbox": [ + 313, + 319, + 555, + 403 + ], + "type": "text", + "content": " and enables a comparison of the threat along different unsafe directions (see Figure 2). In (1), the supremum ensures that we do not underestimate the threat along a direction that potentially alters the true label. We highlight two key characteristics." + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 313, + 403, + 555, + 487 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 403, + 555, + 487 + ], + "spans": [ + { + "bbox": [ + 313, + 403, + 555, + 487 + ], + "type": "text", + "content": "Anisotropy. The threat function " + }, + { + "bbox": [ + 313, + 403, + 555, + 487 + ], + "type": "inline_equation", + "content": "d_{\\mathrm{PD}}^{\\star}" + }, + { + "bbox": [ + 313, + 403, + 555, + 487 + ], + "type": "text", + "content": " is anisotropic (the threat of a perturbation " + }, + { + "bbox": [ + 313, + 403, + 555, + 487 + ], + "type": "inline_equation", + "content": "\\delta" + }, + { + "bbox": [ + 313, + 403, + 555, + 487 + ], + "type": "text", + "content": " at input " + }, + { + "bbox": [ + 313, + 403, + 555, + 487 + ], + "type": "inline_equation", + "content": "\\mathbf{x}" + }, + { + "bbox": [ + 313, + 403, + 555, + 487 + ], + "type": "text", + "content": " depends on both the direction and magnitude of the perturbation) and does not correspond to a norm in general, unlike " + }, + { + "bbox": [ + 313, + 403, + 555, + 487 + ], + "type": "inline_equation", + "content": "\\ell_p" + }, + { + "bbox": [ + 313, + 403, + 555, + 487 + ], + "type": "text", + "content": " threat models. For a fixed " + }, + { + "bbox": [ + 313, + 403, + 555, + 487 + ], + "type": "inline_equation", + "content": "\\mathbf{x}" + }, + { + "bbox": [ + 313, + 403, + 555, + 487 + ], + "type": "text", + "content": ", the threat along different directions " + }, + { + "bbox": [ + 313, + 403, + 555, + 487 + ], + "type": "inline_equation", + "content": "\\mathbf{u}" + }, + { + "bbox": [ + 313, + 403, + 555, + 487 + ], + "type": "text", + "content": " varies based on the alignment with unsafe directions " + }, + { + "bbox": [ + 313, + 403, + 555, + 487 + ], + "type": "inline_equation", + "content": "\\mathcal{U}^{\\star}(\\mathbf{x})" + }, + { + "bbox": [ + 313, + 403, + 555, + 487 + ], + "type": "text", + "content": ". In particular, " + }, + { + "bbox": [ + 313, + 403, + 555, + 487 + ], + "type": "inline_equation", + "content": "d_{\\mathrm{PD}}^{\\star}" + }, + { + "bbox": [ + 313, + 403, + 555, + 487 + ], + "type": "text", + "content": " can differentiate between the perturbations " + }, + { + "bbox": [ + 313, + 403, + 555, + 487 + ], + "type": "inline_equation", + "content": "\\delta" + }, + { + "bbox": [ + 313, + 403, + 555, + 487 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 313, + 403, + 555, + 487 + ], + "type": "inline_equation", + "content": "-\\delta" + }, + { + "bbox": [ + 313, + 403, + 555, + 487 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 313, + 487, + 555, + 548 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 487, + 555, + 548 + ], + "spans": [ + { + "bbox": [ + 313, + 487, + 555, + 548 + ], + "type": "text", + "content": "Locality. The threat function " + }, + { + "bbox": [ + 313, + 487, + 555, + 548 + ], + "type": "inline_equation", + "content": "d_{\\mathrm{PD}}^{\\star}" + }, + { + "bbox": [ + 313, + 487, + 555, + 548 + ], + "type": "text", + "content": " exhibits locality; i.e. a fixed perturbation " + }, + { + "bbox": [ + 313, + 487, + 555, + 548 + ], + "type": "inline_equation", + "content": "\\delta" + }, + { + "bbox": [ + 313, + 487, + 555, + 548 + ], + "type": "text", + "content": " has varying threat " + }, + { + "bbox": [ + 313, + 487, + 555, + 548 + ], + "type": "inline_equation", + "content": "d_{\\mathrm{PD}}^{\\star}(\\mathbf{x}, \\delta)" + }, + { + "bbox": [ + 313, + 487, + 555, + 548 + ], + "type": "text", + "content": " depending on the input " + }, + { + "bbox": [ + 313, + 487, + 555, + 548 + ], + "type": "inline_equation", + "content": "\\mathbf{x}" + }, + { + "bbox": [ + 313, + 487, + 555, + 548 + ], + "type": "text", + "content": ". Thus the permissible set varies with input7, enabling an input-aware assessment of threat. We note that weighted " + }, + { + "bbox": [ + 313, + 487, + 555, + 548 + ], + "type": "inline_equation", + "content": "\\ell_p" + }, + { + "bbox": [ + 313, + 487, + 555, + 548 + ], + "type": "text", + "content": " norms [14] can be anisotropic but not local." + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 313, + 553, + 487, + 565 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 553, + 487, + 565 + ], + "spans": [ + { + "bbox": [ + 313, + 553, + 487, + 565 + ], + "type": "text", + "content": "3.1.1. Stability vs Correctness Trade-off" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 313, + 568, + 554, + 628 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 568, + 554, + 628 + ], + "spans": [ + { + "bbox": [ + 313, + 568, + 554, + 628 + ], + "type": "text", + "content": "Since the normalization function " + }, + { + "bbox": [ + 313, + 568, + 554, + 628 + ], + "type": "inline_equation", + "content": "g^{\\star}" + }, + { + "bbox": [ + 313, + 568, + 554, + 628 + ], + "type": "text", + "content": " accounts for the class partition " + }, + { + "bbox": [ + 313, + 568, + 554, + 628 + ], + "type": "inline_equation", + "content": "\\{\\mathcal{X}_c\\}_{c=1}^C" + }, + { + "bbox": [ + 313, + 568, + 554, + 628 + ], + "type": "text", + "content": ", the threat function " + }, + { + "bbox": [ + 313, + 568, + 554, + 628 + ], + "type": "inline_equation", + "content": "d_{\\mathrm{PD}}^{\\star}" + }, + { + "bbox": [ + 313, + 568, + 554, + 628 + ], + "type": "text", + "content": " has a principled trade-off between stability and correctness. Specifically, the true labeling function " + }, + { + "bbox": [ + 313, + 568, + 554, + 628 + ], + "type": "inline_equation", + "content": "h^{\\star}" + }, + { + "bbox": [ + 313, + 568, + 554, + 628 + ], + "type": "text", + "content": " exhibits a high level of robustness under the PD" + }, + { + "bbox": [ + 313, + 568, + 554, + 628 + ], + "type": "inline_equation", + "content": "^{\\star}" + }, + { + "bbox": [ + 313, + 568, + 554, + 628 + ], + "type": "text", + "content": "-threat8." + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 313, + 637, + 555, + 685 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 637, + 555, + 685 + ], + "spans": [ + { + "bbox": [ + 313, + 637, + 555, + 685 + ], + "type": "text", + "content": "Theorem 1. The true labeling function " + }, + { + "bbox": [ + 313, + 637, + 555, + 685 + ], + "type": "inline_equation", + "content": "h^{\\star}" + }, + { + "bbox": [ + 313, + 637, + 555, + 685 + ], + "type": "text", + "content": " is 1-robust at any input " + }, + { + "bbox": [ + 313, + 637, + 555, + 685 + ], + "type": "inline_equation", + "content": "\\mathbf{x} \\in \\mathcal{X}" + }, + { + "bbox": [ + 313, + 637, + 555, + 685 + ], + "type": "text", + "content": " w.r.t. the threat " + }, + { + "bbox": [ + 313, + 637, + 555, + 685 + ], + "type": "inline_equation", + "content": "d_{\\mathrm{PD}}^{\\star}" + }, + { + "bbox": [ + 313, + 637, + 555, + 685 + ], + "type": "text", + "content": ". Additionally, if a classifier " + }, + { + "bbox": [ + 313, + 637, + 555, + 685 + ], + "type": "inline_equation", + "content": "h" + }, + { + "bbox": [ + 313, + 637, + 555, + 685 + ], + "type": "text", + "content": " is not 1-robust at all inputs then there exists an input " + }, + { + "bbox": [ + 313, + 637, + 555, + 685 + ], + "type": "inline_equation", + "content": "\\mathbf{x}" + }, + { + "bbox": [ + 313, + 637, + 555, + 685 + ], + "type": "text", + "content": " misclassified by " + }, + { + "bbox": [ + 313, + 637, + 555, + 685 + ], + "type": "inline_equation", + "content": "h" + }, + { + "bbox": [ + 313, + 637, + 555, + 685 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 20 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 324, + 692, + 531, + 703 + ], + "type": "page_footnote", + "angle": 0, + "lines": [ + { + "bbox": [ + 324, + 692, + 531, + 703 + ], + "spans": [ + { + "bbox": [ + 324, + 692, + 531, + 703 + ], + "type": "text", + "content": "Unlike " + }, + { + "bbox": [ + 324, + 692, + 531, + 703 + ], + "type": "inline_equation", + "content": "\\ell_p" + }, + { + "bbox": [ + 324, + 692, + 531, + 703 + ], + "type": "text", + "content": " threat models where " + }, + { + "bbox": [ + 324, + 692, + 531, + 703 + ], + "type": "inline_equation", + "content": "S(\\mathbf{x},d_p,\\varepsilon) = \\varepsilon \\mathcal{B}_p^d\\forall \\mathbf{x}\\in \\mathcal{X}" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 326, + 703, + 505, + 712 + ], + "type": "page_footnote", + "angle": 0, + "lines": [ + { + "bbox": [ + 326, + 703, + 505, + 712 + ], + "spans": [ + { + "bbox": [ + 326, + 703, + 505, + 712 + ], + "type": "text", + "content": "The proofs of all our results can be found in Section 8." + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 55, + 683, + 296, + 714 + ], + "type": "page_footnote", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 683, + 296, + 714 + ], + "spans": [ + { + "bbox": [ + 55, + 683, + 296, + 714 + ], + "type": "text", + "content": "Note as per Definition 2, robust accuracy measures stability rather than correctness since we aren't evaluating probability of the event " + }, + { + "bbox": [ + 55, + 683, + 296, + 714 + ], + "type": "inline_equation", + "content": "\\{\\forall \\delta \\in S(\\mathbf{x},d,\\varepsilon), h(\\mathbf{x} + \\delta) = h^{\\star}(\\mathbf{x} + \\delta)\\}" + }, + { + "bbox": [ + 55, + 683, + 296, + 714 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 295, + 750, + 315, + 758 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 750, + 315, + 758 + ], + "spans": [ + { + "bbox": [ + 295, + 750, + 315, + 758 + ], + "type": "text", + "content": "9957" + } + ] + } + ], + "index": 25 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 3 + }, + { + "para_blocks": [ + { + "bbox": [ + 55, + 72, + 296, + 122 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 72, + 296, + 122 + ], + "spans": [ + { + "bbox": [ + 55, + 72, + 296, + 122 + ], + "type": "text", + "content": "In other words, for any pairs of inputs " + }, + { + "bbox": [ + 55, + 72, + 296, + 122 + ], + "type": "inline_equation", + "content": "\\mathbf{x},\\tilde{\\mathbf{x}}\\in \\mathcal{X}" + }, + { + "bbox": [ + 55, + 72, + 296, + 122 + ], + "type": "text", + "content": " with different labels, the threat " + }, + { + "bbox": [ + 55, + 72, + 296, + 122 + ], + "type": "inline_equation", + "content": "d_{\\mathrm{PD}}^{\\star}(\\mathbf{x},\\tilde{\\mathbf{x}} -\\mathbf{x})" + }, + { + "bbox": [ + 55, + 72, + 296, + 122 + ], + "type": "text", + "content": " is always larger than 1. Thus seeking " + }, + { + "bbox": [ + 55, + 72, + 296, + 122 + ], + "type": "inline_equation", + "content": "h\\in \\mathcal{H}" + }, + { + "bbox": [ + 55, + 72, + 296, + 122 + ], + "type": "text", + "content": " that is 1-robust at any input under " + }, + { + "bbox": [ + 55, + 72, + 296, + 122 + ], + "type": "inline_equation", + "content": "d_{\\mathrm{PD}}^{\\star}" + }, + { + "bbox": [ + 55, + 72, + 296, + 122 + ], + "type": "text", + "content": " is not at odds with learning " + }, + { + "bbox": [ + 55, + 72, + 296, + 122 + ], + "type": "inline_equation", + "content": "h^{\\star}" + }, + { + "bbox": [ + 55, + 72, + 296, + 122 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 55, + 127, + 256, + 140 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 127, + 256, + 140 + ], + "spans": [ + { + "bbox": [ + 55, + 127, + 256, + 140 + ], + "type": "text", + "content": "3.2. Measuring threats with observed data" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 55, + 144, + 296, + 204 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 144, + 296, + 204 + ], + "spans": [ + { + "bbox": [ + 55, + 144, + 296, + 204 + ], + "type": "text", + "content": "Computing " + }, + { + "bbox": [ + 55, + 144, + 296, + 204 + ], + "type": "inline_equation", + "content": "d_{\\mathrm{PD}}^{\\star}(\\mathbf{x},\\cdot)" + }, + { + "bbox": [ + 55, + 144, + 296, + 204 + ], + "type": "text", + "content": " from Definition 5 requires explicitly characterizing the set of unsafe directions " + }, + { + "bbox": [ + 55, + 144, + 296, + 204 + ], + "type": "inline_equation", + "content": "\\mathcal{U}^{\\star}(\\mathbf{x})" + }, + { + "bbox": [ + 55, + 144, + 296, + 204 + ], + "type": "text", + "content": " and the normalization function " + }, + { + "bbox": [ + 55, + 144, + 296, + 204 + ], + "type": "inline_equation", + "content": "g^{\\star}(\\mathbf{u},\\mathbf{x})" + }, + { + "bbox": [ + 55, + 144, + 296, + 204 + ], + "type": "text", + "content": ", both of which need knowledge of the partition sets " + }, + { + "bbox": [ + 55, + 144, + 296, + 204 + ], + "type": "inline_equation", + "content": "\\{\\mathcal{X}_c\\}_{c = 1}^C" + }, + { + "bbox": [ + 55, + 144, + 296, + 204 + ], + "type": "text", + "content": " induced by " + }, + { + "bbox": [ + 55, + 144, + 296, + 204 + ], + "type": "inline_equation", + "content": "h^\\star" + }, + { + "bbox": [ + 55, + 144, + 296, + 204 + ], + "type": "text", + "content": ". In this subsection, we propose a heuristic empirical approximation." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 55, + 205, + 296, + 338 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 205, + 296, + 338 + ], + "spans": [ + { + "bbox": [ + 55, + 205, + 296, + 338 + ], + "type": "text", + "content": "Approximating " + }, + { + "bbox": [ + 55, + 205, + 296, + 338 + ], + "type": "inline_equation", + "content": "\\mathcal{U}^{\\star}(\\mathbf{x})" + }, + { + "bbox": [ + 55, + 205, + 296, + 338 + ], + "type": "text", + "content": ". We have access to training data S and incomplete information on the partition sets " + }, + { + "bbox": [ + 55, + 205, + 296, + 338 + ], + "type": "inline_equation", + "content": "\\{\\mathcal{X}_c\\}_{c=1}^C" + }, + { + "bbox": [ + 55, + 205, + 296, + 338 + ], + "type": "text", + "content": " via the partition " + }, + { + "bbox": [ + 55, + 205, + 296, + 338 + ], + "type": "inline_equation", + "content": "S = \\{S_c\\}_{c=1}^C" + }, + { + "bbox": [ + 55, + 205, + 296, + 338 + ], + "type": "text", + "content": ", where " + }, + { + "bbox": [ + 55, + 205, + 296, + 338 + ], + "type": "inline_equation", + "content": "S_c := \\{\\mathbf{x} \\in \\mathcal{X} \\mid (\\mathbf{x}, c) \\in S\\} \\subset \\mathcal{X}_c" + }, + { + "bbox": [ + 55, + 205, + 296, + 338 + ], + "type": "text", + "content": ". At every training input " + }, + { + "bbox": [ + 55, + 205, + 296, + 338 + ], + "type": "inline_equation", + "content": "\\mathbf{x} \\in S_y" + }, + { + "bbox": [ + 55, + 205, + 296, + 338 + ], + "type": "text", + "content": " for each input " + }, + { + "bbox": [ + 55, + 205, + 296, + 338 + ], + "type": "inline_equation", + "content": "\\tilde{\\mathbf{x}} \\in S_c" + }, + { + "bbox": [ + 55, + 205, + 296, + 338 + ], + "type": "text", + "content": " (for " + }, + { + "bbox": [ + 55, + 205, + 296, + 338 + ], + "type": "inline_equation", + "content": "c \\neq y" + }, + { + "bbox": [ + 55, + 205, + 296, + 338 + ], + "type": "text", + "content": "), the direction " + }, + { + "bbox": [ + 55, + 205, + 296, + 338 + ], + "type": "inline_equation", + "content": "\\mathbf{u} = \\frac{\\tilde{\\mathbf{x}} - \\mathbf{x}}{\\|\\tilde{\\mathbf{x}} - \\mathbf{x}\\|_2}" + }, + { + "bbox": [ + 55, + 205, + 296, + 338 + ], + "type": "text", + "content": " is an unsafe direction. For computational considerations, we choose a collection " + }, + { + "bbox": [ + 55, + 205, + 296, + 338 + ], + "type": "inline_equation", + "content": "\\bigcup_{c} S_{c,k}" + }, + { + "bbox": [ + 55, + 205, + 296, + 338 + ], + "type": "text", + "content": " of representative subsets " + }, + { + "bbox": [ + 55, + 205, + 296, + 338 + ], + "type": "inline_equation", + "content": "S_{c,k} \\subset S_c" + }, + { + "bbox": [ + 55, + 205, + 296, + 338 + ], + "type": "text", + "content": " of size " + }, + { + "bbox": [ + 55, + 205, + 296, + 338 + ], + "type": "inline_equation", + "content": "k" + }, + { + "bbox": [ + 55, + 205, + 296, + 338 + ], + "type": "text", + "content": ". As a heuristic choice, we select subsets " + }, + { + "bbox": [ + 55, + 205, + 296, + 338 + ], + "type": "inline_equation", + "content": "S_{c,k}" + }, + { + "bbox": [ + 55, + 205, + 296, + 338 + ], + "type": "text", + "content": " by solving a discrete-k-center optimization problem (detailed in Sec. 9.1). We are now ready to define an observed subset of the unsafe directions based on " + }, + { + "bbox": [ + 55, + 205, + 296, + 338 + ], + "type": "inline_equation", + "content": "S_{c,k}" + }, + { + "bbox": [ + 55, + 205, + 296, + 338 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 55, + 344, + 296, + 381 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 344, + 296, + 381 + ], + "spans": [ + { + "bbox": [ + 55, + 344, + 296, + 381 + ], + "type": "text", + "content": "Definition 6 (Observed Unsafe Directions). For each " + }, + { + "bbox": [ + 55, + 344, + 296, + 381 + ], + "type": "inline_equation", + "content": "\\mathbf{x} \\in S_y" + }, + { + "bbox": [ + 55, + 344, + 296, + 381 + ], + "type": "text", + "content": ", we identify a set of " + }, + { + "bbox": [ + 55, + 344, + 296, + 381 + ], + "type": "inline_equation", + "content": "(C - 1)k" + }, + { + "bbox": [ + 55, + 344, + 296, + 381 + ], + "type": "text", + "content": "-observed unsafe directions based on the representative subsets " + }, + { + "bbox": [ + 55, + 344, + 296, + 381 + ], + "type": "inline_equation", + "content": "S_{c,k}" + }, + { + "bbox": [ + 55, + 344, + 296, + 381 + ], + "type": "text", + "content": "," + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 74, + 388, + 275, + 416 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 74, + 388, + 275, + 416 + ], + "spans": [ + { + "bbox": [ + 74, + 388, + 275, + 416 + ], + "type": "interline_equation", + "content": "\\mathcal {U} _ {k} (\\mathbf {x}) := \\left\\{\\frac {\\tilde {\\mathbf {x}} - \\mathbf {x}}{\\| \\tilde {\\mathbf {x}} - \\mathbf {x} \\| _ {2}} \\Big | \\tilde {\\mathbf {x}} \\in \\cup_ {c \\neq y} \\mathbb {S} _ {c, k} \\right\\} \\subset \\mathcal {U} ^ {\\star} (\\mathbf {x}).", + "image_path": "727e44c656d070f879aeb67c378c22b152039d0d983911af21364ce5eb18635c.jpg" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 55, + 427, + 296, + 502 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 427, + 296, + 502 + ], + "spans": [ + { + "bbox": [ + 55, + 427, + 296, + 502 + ], + "type": "text", + "content": "Approximating " + }, + { + "bbox": [ + 55, + 427, + 296, + 502 + ], + "type": "inline_equation", + "content": "g^{\\star}(\\mathbf{x},\\mathbf{u})" + }, + { + "bbox": [ + 55, + 427, + 296, + 502 + ], + "type": "text", + "content": ". At any " + }, + { + "bbox": [ + 55, + 427, + 296, + 502 + ], + "type": "inline_equation", + "content": "\\mathbf{x}\\in \\mathcal{X}_y" + }, + { + "bbox": [ + 55, + 427, + 296, + 502 + ], + "type": "text", + "content": ", for each " + }, + { + "bbox": [ + 55, + 427, + 296, + 502 + ], + "type": "inline_equation", + "content": "\\mathbf{u}\\in \\mathcal{U}_k(\\mathbf{x})" + }, + { + "bbox": [ + 55, + 427, + 296, + 502 + ], + "type": "text", + "content": ", there exists " + }, + { + "bbox": [ + 55, + 427, + 296, + 502 + ], + "type": "inline_equation", + "content": "\\tilde{\\mathbf{x}}\\in \\cup_{c\\neq y}\\mathrm{S}_{c,k}" + }, + { + "bbox": [ + 55, + 427, + 296, + 502 + ], + "type": "text", + "content": " such that " + }, + { + "bbox": [ + 55, + 427, + 296, + 502 + ], + "type": "inline_equation", + "content": "\\mathbf{u} = \\frac{\\tilde{\\mathbf{x}} - \\mathbf{x}}{\\|\\tilde{\\mathbf{x}} - \\mathbf{x}\\|_2}" + }, + { + "bbox": [ + 55, + 427, + 296, + 502 + ], + "type": "text", + "content": ". normalization function can be bounded as " + }, + { + "bbox": [ + 55, + 427, + 296, + 502 + ], + "type": "inline_equation", + "content": "g^{\\star}(\\mathbf{x},\\mathbf{u})\\leq \\| \\tilde{\\mathbf{x}} -\\mathbf{x}\\| _2" + }, + { + "bbox": [ + 55, + 427, + 296, + 502 + ], + "type": "text", + "content": ", since " + }, + { + "bbox": [ + 55, + 427, + 296, + 502 + ], + "type": "inline_equation", + "content": "h^\\star (\\tilde{\\mathbf{x}})\\neq h^\\star (\\mathbf{x})" + }, + { + "bbox": [ + 55, + 427, + 296, + 502 + ], + "type": "text", + "content": ". In practice, we use the heuristic " + }, + { + "bbox": [ + 55, + 427, + 296, + 502 + ], + "type": "inline_equation", + "content": "g_{\\beta}(\\mathbf{u},\\mathbf{x})\\coloneqq \\beta \\| \\tilde{\\mathbf{x}} -\\mathbf{x}\\| _2" + }, + { + "bbox": [ + 55, + 427, + 296, + 502 + ], + "type": "text", + "content": ", where " + }, + { + "bbox": [ + 55, + 427, + 296, + 502 + ], + "type": "inline_equation", + "content": "\\beta \\in (0,1)" + }, + { + "bbox": [ + 55, + 427, + 296, + 502 + ], + "type": "text", + "content": " is a scaling hyper-parameter." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 55, + 509, + 296, + 544 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 509, + 296, + 544 + ], + "spans": [ + { + "bbox": [ + 55, + 509, + 296, + 544 + ], + "type": "text", + "content": "Definition 7 " + }, + { + "bbox": [ + 55, + 509, + 296, + 544 + ], + "type": "inline_equation", + "content": "((k,\\beta)" + }, + { + "bbox": [ + 55, + 509, + 296, + 544 + ], + "type": "text", + "content": " -PD threat). Let " + }, + { + "bbox": [ + 55, + 509, + 296, + 544 + ], + "type": "inline_equation", + "content": "\\mathbf{x}\\in \\mathcal{X}" + }, + { + "bbox": [ + 55, + 509, + 296, + 544 + ], + "type": "text", + "content": " and let " + }, + { + "bbox": [ + 55, + 509, + 296, + 544 + ], + "type": "inline_equation", + "content": "\\delta \\in \\mathbb{R}^d" + }, + { + "bbox": [ + 55, + 509, + 296, + 544 + ], + "type": "text", + "content": " be a perturbation. The projected displacement threat is defined as" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 64, + 549, + 286, + 575 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 64, + 549, + 286, + 575 + ], + "spans": [ + { + "bbox": [ + 64, + 549, + 286, + 575 + ], + "type": "interline_equation", + "content": "d _ {\\mathrm {P D}, k, \\beta} (\\mathbf {x}, \\boldsymbol {\\delta}) := \\max _ {\\mathbf {u} \\in \\mathcal {U} _ {k} (\\mathbf {x})} \\frac {1}{g _ {\\beta} (\\mathbf {x} , \\mathbf {u})} \\max \\left(\\langle \\boldsymbol {\\delta}, \\mathbf {u} \\rangle , 0\\right).", + "image_path": "a422ab6f6b6c8f61a2fb2c8fe3c7129da93bc4f355adb1028cc81b7ae2f415ca.jpg" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 55, + 582, + 297, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 582, + 297, + 715 + ], + "spans": [ + { + "bbox": [ + 55, + 582, + 297, + 715 + ], + "type": "text", + "content": "The quality of the approximation of the exact threat " + }, + { + "bbox": [ + 55, + 582, + 297, + 715 + ], + "type": "inline_equation", + "content": "d_{\\mathrm{PD}}^{\\star}" + }, + { + "bbox": [ + 55, + 582, + 297, + 715 + ], + "type": "text", + "content": " by " + }, + { + "bbox": [ + 55, + 582, + 297, + 715 + ], + "type": "inline_equation", + "content": "d_{\\mathrm{PD,k},\\beta}" + }, + { + "bbox": [ + 55, + 582, + 297, + 715 + ], + "type": "text", + "content": " depends on the choice of representative unsafe directions " + }, + { + "bbox": [ + 55, + 582, + 297, + 715 + ], + "type": "inline_equation", + "content": "\\mathcal{U}_k(\\cdot)" + }, + { + "bbox": [ + 55, + 582, + 297, + 715 + ], + "type": "text", + "content": " via the " + }, + { + "bbox": [ + 55, + 582, + 297, + 715 + ], + "type": "inline_equation", + "content": "k" + }, + { + "bbox": [ + 55, + 582, + 297, + 715 + ], + "type": "text", + "content": "-subset " + }, + { + "bbox": [ + 55, + 582, + 297, + 715 + ], + "type": "inline_equation", + "content": "\\mathbb{S}_{c,k}" + }, + { + "bbox": [ + 55, + 582, + 297, + 715 + ], + "type": "text", + "content": " and on the heuristic choice of the approximate normalization " + }, + { + "bbox": [ + 55, + 582, + 297, + 715 + ], + "type": "inline_equation", + "content": "g_{\\beta}" + }, + { + "bbox": [ + 55, + 582, + 297, + 715 + ], + "type": "text", + "content": ", via the scaling hyper-parameter " + }, + { + "bbox": [ + 55, + 582, + 297, + 715 + ], + "type": "inline_equation", + "content": "\\beta" + }, + { + "bbox": [ + 55, + 582, + 297, + 715 + ], + "type": "text", + "content": ". In this article, we fix " + }, + { + "bbox": [ + 55, + 582, + 297, + 715 + ], + "type": "inline_equation", + "content": "(k,\\beta) = (50,\\frac{1}{2})" + }, + { + "bbox": [ + 55, + 582, + 297, + 715 + ], + "type": "text", + "content": " as the default hyper-parameters (see Section 9.3 for an expanded discussion on choosing hyper-parameters " + }, + { + "bbox": [ + 55, + 582, + 297, + 715 + ], + "type": "inline_equation", + "content": "(k,\\beta)" + }, + { + "bbox": [ + 55, + 582, + 297, + 715 + ], + "type": "text", + "content": " for any learning task). For brevity, we refer to the observed unsafe directions (Definition 6), the heuristic normalizations and the PD threat (Definition 7) as " + }, + { + "bbox": [ + 55, + 582, + 297, + 715 + ], + "type": "inline_equation", + "content": "\\mathcal{U}(\\mathbf{x}), g(\\mathbf{x},\\delta)" + }, + { + "bbox": [ + 55, + 582, + 297, + 715 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 55, + 582, + 297, + 715 + ], + "type": "inline_equation", + "content": "d_{\\mathrm{PD}}(\\mathbf{x},\\delta)" + }, + { + "bbox": [ + 55, + 582, + 297, + 715 + ], + "type": "text", + "content": ", respectively." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 313, + 72, + 553, + 109 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 72, + 553, + 109 + ], + "spans": [ + { + "bbox": [ + 313, + 72, + 553, + 109 + ], + "type": "text", + "content": "We now state a few important properties of our proposed empirical approximation " + }, + { + "bbox": [ + 313, + 72, + 553, + 109 + ], + "type": "inline_equation", + "content": "d_{\\mathrm{PD}}" + }, + { + "bbox": [ + 313, + 72, + 553, + 109 + ], + "type": "text", + "content": ". Clearly " + }, + { + "bbox": [ + 313, + 72, + 553, + 109 + ], + "type": "inline_equation", + "content": "d_{\\mathrm{PD}}" + }, + { + "bbox": [ + 313, + 72, + 553, + 109 + ], + "type": "text", + "content": " inherits the anisotropy and locality properties of the exact threat " + }, + { + "bbox": [ + 313, + 72, + 553, + 109 + ], + "type": "inline_equation", + "content": "d_{\\mathrm{PD}}^{\\star}" + }, + { + "bbox": [ + 313, + 72, + 553, + 109 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 313, + 109, + 555, + 170 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 109, + 555, + 170 + ], + "spans": [ + { + "bbox": [ + 313, + 109, + 555, + 170 + ], + "type": "text", + "content": "Growth and Sensitivity. The growth of PD threat along any direction is linear, i.e., " + }, + { + "bbox": [ + 313, + 109, + 555, + 170 + ], + "type": "inline_equation", + "content": "d_{\\mathrm{PD}}(\\mathbf{x},t\\boldsymbol {\\delta}) = td_{\\mathrm{PD}}(\\mathbf{x},\\boldsymbol {\\delta})" + }, + { + "bbox": [ + 313, + 109, + 555, + 170 + ], + "type": "text", + "content": " , and the rate of growth scales inversely with the (approximate) normalization. In particular, for a fixed " + }, + { + "bbox": [ + 313, + 109, + 555, + 170 + ], + "type": "inline_equation", + "content": "\\mathbf{x}" + }, + { + "bbox": [ + 313, + 109, + 555, + 170 + ], + "type": "inline_equation", + "content": "d_{\\mathrm{PD}}(\\mathbf{x},\\cdot)" + }, + { + "bbox": [ + 313, + 109, + 555, + 170 + ], + "type": "text", + "content": " is " + }, + { + "bbox": [ + 313, + 109, + 555, + 170 + ], + "type": "inline_equation", + "content": "(\\max_{u\\in \\mathcal{U}(\\mathbf{x})}1 / g(\\mathbf{x},\\mathbf{u}))" + }, + { + "bbox": [ + 313, + 109, + 555, + 170 + ], + "type": "text", + "content": " -Lipschitz w.r.t " + }, + { + "bbox": [ + 313, + 109, + 555, + 170 + ], + "type": "inline_equation", + "content": "\\ell_2" + }, + { + "bbox": [ + 313, + 109, + 555, + 170 + ], + "type": "text", + "content": " norm." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 313, + 170, + 556, + 284 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 170, + 556, + 284 + ], + "spans": [ + { + "bbox": [ + 313, + 170, + 556, + 284 + ], + "type": "text", + "content": "Projection onto sublevel sets. For any threat model " + }, + { + "bbox": [ + 313, + 170, + 556, + 284 + ], + "type": "inline_equation", + "content": "(d,\\varepsilon)" + }, + { + "bbox": [ + 313, + 170, + 556, + 284 + ], + "type": "text", + "content": ", the ability to efficiently project a perturbation " + }, + { + "bbox": [ + 313, + 170, + 556, + 284 + ], + "type": "inline_equation", + "content": "\\delta" + }, + { + "bbox": [ + 313, + 170, + 556, + 284 + ], + "type": "text", + "content": " onto the sub-level sets " + }, + { + "bbox": [ + 313, + 170, + 556, + 284 + ], + "type": "inline_equation", + "content": "S(d,\\mathbf{x},\\varepsilon)" + }, + { + "bbox": [ + 313, + 170, + 556, + 284 + ], + "type": "text", + "content": " enables one to leverage a wide literature on gradient-based attacks to conduct a rigorous evaluations of adversarial robustness. The sublevel sets of PD threat model, i.e., " + }, + { + "bbox": [ + 313, + 170, + 556, + 284 + ], + "type": "inline_equation", + "content": "S(\\mathbf{x},d,\\varepsilon)" + }, + { + "bbox": [ + 313, + 170, + 556, + 284 + ], + "type": "text", + "content": ", are convex for all " + }, + { + "bbox": [ + 313, + 170, + 556, + 284 + ], + "type": "inline_equation", + "content": "\\mathbf{x}" + }, + { + "bbox": [ + 313, + 170, + 556, + 284 + ], + "type": "text", + "content": " for all " + }, + { + "bbox": [ + 313, + 170, + 556, + 284 + ], + "type": "inline_equation", + "content": "\\epsilon \\geq 0" + }, + { + "bbox": [ + 313, + 170, + 556, + 284 + ], + "type": "text", + "content": ". In particular, " + }, + { + "bbox": [ + 313, + 170, + 556, + 284 + ], + "type": "inline_equation", + "content": "\\mathcal{S}" + }, + { + "bbox": [ + 313, + 170, + 556, + 284 + ], + "type": "text", + "content": " is the intersection of at most " + }, + { + "bbox": [ + 313, + 170, + 556, + 284 + ], + "type": "inline_equation", + "content": "k\\cdot C" + }, + { + "bbox": [ + 313, + 170, + 556, + 284 + ], + "type": "text", + "content": " half-spaces characterized by unsafe directions " + }, + { + "bbox": [ + 313, + 170, + 556, + 284 + ], + "type": "inline_equation", + "content": "\\mathcal{U}(\\mathbf{x})" + }, + { + "bbox": [ + 313, + 170, + 556, + 284 + ], + "type": "text", + "content": " as " + }, + { + "bbox": [ + 313, + 170, + 556, + 284 + ], + "type": "inline_equation", + "content": "S(\\mathbf{x},d_{\\mathrm{PD}},\\varepsilon) = \\bigcap_{\\mathbf{u}\\in \\mathcal{U}(\\mathbf{x})}\\{\\pmb {\\delta}\\in \\mathbb{R}^d\\mid \\langle \\pmb {\\delta},\\mathbf{u}\\rangle \\leq \\varepsilon \\cdot g(\\mathbf{x},\\mathbf{u})\\}" + }, + { + "bbox": [ + 313, + 170, + 556, + 284 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 313, + 283, + 555, + 307 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 283, + 555, + 307 + ], + "spans": [ + { + "bbox": [ + 313, + 283, + 555, + 307 + ], + "type": "text", + "content": "Further, the sub-level sets are monotonic w.r.t. the threshold, i.e., for " + }, + { + "bbox": [ + 313, + 283, + 555, + 307 + ], + "type": "inline_equation", + "content": "\\varepsilon_1\\geq \\varepsilon_2\\geq 0,S(\\mathbf{x},d_{\\mathrm{PD}},\\varepsilon_2)\\subseteq S(\\mathbf{x},d_{\\mathrm{PD}},\\varepsilon_1)" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 313, + 308, + 556, + 450 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 308, + 556, + 450 + ], + "spans": [ + { + "bbox": [ + 313, + 308, + 556, + 450 + ], + "type": "text", + "content": "Linearity of growth and convexity of the sub-level sets together provide a straightforward approximate projection algorithm, since for any perturbation " + }, + { + "bbox": [ + 313, + 308, + 556, + 450 + ], + "type": "inline_equation", + "content": "\\delta" + }, + { + "bbox": [ + 313, + 308, + 556, + 450 + ], + "type": "text", + "content": ", the scaled perturbation " + }, + { + "bbox": [ + 313, + 308, + 556, + 450 + ], + "type": "inline_equation", + "content": "\\frac{\\varepsilon}{d_{\\mathrm{PD}}(\\mathbf{x},\\delta)}\\delta" + }, + { + "bbox": [ + 313, + 308, + 556, + 450 + ], + "type": "text", + "content": " lies in the permissible set " + }, + { + "bbox": [ + 313, + 308, + 556, + 450 + ], + "type": "inline_equation", + "content": "S(\\mathbf{x},d_{\\mathrm{PD}},\\varepsilon)" + }, + { + "bbox": [ + 313, + 308, + 556, + 450 + ], + "type": "text", + "content": ". Further, due to the convexity of sublevel sets, an exact projection (in " + }, + { + "bbox": [ + 313, + 308, + 556, + 450 + ], + "type": "inline_equation", + "content": "\\ell_2" + }, + { + "bbox": [ + 313, + 308, + 556, + 450 + ], + "type": "text", + "content": " distance) can be computed using the iterative greedy scheme outlined in the Algorithm 2 (deferred to the appendix). Thus the PD threat provides the practitioner with the flexibility of a fast approximate projection or an exact projection at the expense of increased computation. Together, these projection algorithms enable immediate adaptation of existing adversarial attacks to our proposed threat model." + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 313, + 451, + 556, + 512 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 451, + 556, + 512 + ], + "spans": [ + { + "bbox": [ + 313, + 451, + 556, + 512 + ], + "type": "text", + "content": "We note that efficient exact projection algorithms are also known for " + }, + { + "bbox": [ + 313, + 451, + 556, + 512 + ], + "type": "inline_equation", + "content": "\\ell_p" + }, + { + "bbox": [ + 313, + 451, + 556, + 512 + ], + "type": "text", + "content": " threat models. However, for neural perceptual threat models such as DreamSim [15], the sub-level sets are non-convex and a projection can only be approximately computed [32] and is typically not computationally efficient." + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 313, + 524, + 500, + 538 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 524, + 500, + 538 + ], + "spans": [ + { + "bbox": [ + 313, + 524, + 500, + 538 + ], + "type": "text", + "content": "4. Experiments on Real World Data" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 313, + 545, + 556, + 653 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 545, + 556, + 653 + ], + "spans": [ + { + "bbox": [ + 313, + 545, + 556, + 653 + ], + "type": "text", + "content": "In this section we illustrate the qualitative and quantitative characteristics of the PD-threat in comparison to existing threat models. We evaluate 4 threat functions: " + }, + { + "bbox": [ + 313, + 545, + 556, + 653 + ], + "type": "inline_equation", + "content": "d_{\\infty}, d_{2}, d_{\\mathrm{DS}}" + }, + { + "bbox": [ + 313, + 545, + 556, + 653 + ], + "type": "text", + "content": ", and our proposed PD threat function " + }, + { + "bbox": [ + 313, + 545, + 556, + 653 + ], + "type": "inline_equation", + "content": "d_{\\mathrm{PD}}" + }, + { + "bbox": [ + 313, + 545, + 556, + 653 + ], + "type": "text", + "content": ", where " + }, + { + "bbox": [ + 313, + 545, + 556, + 653 + ], + "type": "inline_equation", + "content": "d_{\\mathrm{DS}}" + }, + { + "bbox": [ + 313, + 545, + 556, + 653 + ], + "type": "text", + "content": " is the neural perceptual threat function " + }, + { + "bbox": [ + 313, + 545, + 556, + 653 + ], + "type": "inline_equation", + "content": "d_{\\mathrm{DS}}(\\mathbf{x}, \\tilde{\\mathbf{x}} - \\mathbf{x}) \\coloneqq \\mathrm{DreamSim}(\\mathbf{x}, \\tilde{\\mathbf{x}})" + }, + { + "bbox": [ + 313, + 545, + 556, + 653 + ], + "type": "text", + "content": " from DreamSim [15]. Note that we focus on DreamSim due to its superior performance over other neural perceptual threat models [15], but the following comparison applies to any modern neural perceptual threat model." + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 313, + 654, + 556, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 654, + 556, + 715 + ], + "spans": [ + { + "bbox": [ + 313, + 654, + 556, + 715 + ], + "type": "text", + "content": "Experimental Setup. We seek to measure the ability of threat functions to distinguish between safe and unsafe corruptions. The threat models " + }, + { + "bbox": [ + 313, + 654, + 556, + 715 + ], + "type": "inline_equation", + "content": "d_{\\mathrm{PD}}" + }, + { + "bbox": [ + 313, + 654, + 556, + 715 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 313, + 654, + 556, + 715 + ], + "type": "inline_equation", + "content": "d_{\\mathrm{DS}}" + }, + { + "bbox": [ + 313, + 654, + 556, + 715 + ], + "type": "text", + "content": " depend on the Imagenet-1k training dataset S. All threat models are evaluated on the Imagenet-1k validation dataset " + }, + { + "bbox": [ + 313, + 654, + 556, + 715 + ], + "type": "inline_equation", + "content": "\\mathbf{S}_{\\mathrm{val}}" + }, + { + "bbox": [ + 313, + 654, + 556, + 715 + ], + "type": "text", + "content": ". For any pair" + } + ] + } + ], + "index": 18 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 750, + 315, + 758 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 750, + 315, + 758 + ], + "spans": [ + { + "bbox": [ + 294, + 750, + 315, + 758 + ], + "type": "text", + "content": "9958" + } + ] + } + ], + "index": 19 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 4 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 56, + 70, + 296, + 215 + ], + "blocks": [ + { + "bbox": [ + 56, + 70, + 296, + 215 + ], + "lines": [ + { + "bbox": [ + 56, + 70, + 296, + 215 + ], + "spans": [ + { + "bbox": [ + 56, + 70, + 296, + 215 + ], + "type": "table", + "html": "
CATEGORYCORRUPTION STYLES
NoiseGaussian, Shot, Impulse, Speckle, Blue, Brown, Perlin, Single-Frequency, Plasma
BlurGaussian, Defocus, Glass, Motion, Concentric Sine Waves, Caustic Refraction
CompressionPixelate, JPEG
DigitalBrightness, Contrast, Saturate, Elastic Transform
WeatherFrost, Fog, Snow
OcclusionSpatter, Checkerboard Cutout, Sparkles, Inverse Sparkles
", + "image_path": "20affae325cbe6b4e64be8104246c37515168946fc039dfcf72fd6b064f3cccb.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "table_body" + } + ], + "index": 0 + }, + { + "bbox": [ + 55, + 223, + 296, + 246 + ], + "lines": [ + { + "bbox": [ + 55, + 223, + 296, + 246 + ], + "spans": [ + { + "bbox": [ + 55, + 223, + 296, + 246 + ], + "type": "text", + "content": "Table 2. Categories of corruptions studied, from Imagenet-C [25] and Imagenet- " + }, + { + "bbox": [ + 55, + 223, + 296, + 246 + ], + "type": "inline_equation", + "content": "\\tilde{\\mathbf{C}}" + }, + { + "bbox": [ + 55, + 223, + 296, + 246 + ], + "type": "text", + "content": " [36]." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "text" + }, + { + "bbox": [ + 55, + 266, + 296, + 386 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 266, + 296, + 386 + ], + "spans": [ + { + "bbox": [ + 55, + 266, + 296, + 386 + ], + "type": "text", + "content": "of labelled inputs " + }, + { + "bbox": [ + 55, + 266, + 296, + 386 + ], + "type": "inline_equation", + "content": "(\\mathbf{x},y),(\\tilde{\\mathbf{x}},c)\\in S_{\\mathrm{val}}" + }, + { + "bbox": [ + 55, + 266, + 296, + 386 + ], + "type": "text", + "content": " with distinct labels " + }, + { + "bbox": [ + 55, + 266, + 296, + 386 + ], + "type": "inline_equation", + "content": "(c\\neq y)" + }, + { + "bbox": [ + 55, + 266, + 296, + 386 + ], + "type": "text", + "content": ", the perturbation " + }, + { + "bbox": [ + 55, + 266, + 296, + 386 + ], + "type": "inline_equation", + "content": "\\delta \\coloneqq \\tilde{\\mathbf{x}} -\\mathbf{x}" + }, + { + "bbox": [ + 55, + 266, + 296, + 386 + ], + "type": "text", + "content": " is an unsafe perturbation at " + }, + { + "bbox": [ + 55, + 266, + 296, + 386 + ], + "type": "inline_equation", + "content": "\\mathbf{x}" + }, + { + "bbox": [ + 55, + 266, + 296, + 386 + ], + "type": "text", + "content": ". This way, we can evaluate the statistics of threat under unsafe perturbations for each threat function. To characterize threat under safe perturbations, we consider a set " + }, + { + "bbox": [ + 55, + 266, + 296, + 386 + ], + "type": "inline_equation", + "content": "\\Omega" + }, + { + "bbox": [ + 55, + 266, + 296, + 386 + ], + "type": "text", + "content": " of 150 distinct common corruptions consisting of 30 styles " + }, + { + "bbox": [ + 55, + 266, + 296, + 386 + ], + "type": "inline_equation", + "content": "\\times 5" + }, + { + "bbox": [ + 55, + 266, + 296, + 386 + ], + "type": "text", + "content": " severity levels, i.e., severity " + }, + { + "bbox": [ + 55, + 266, + 296, + 386 + ], + "type": "inline_equation", + "content": "\\in \\{1,2,3,4,5\\}" + }, + { + "bbox": [ + 55, + 266, + 296, + 386 + ], + "type": "text", + "content": " - the supplementary material contains visual examples (see Figures 8 and 9), sourced from Imagenet- " + }, + { + "bbox": [ + 55, + 266, + 296, + 386 + ], + "type": "inline_equation", + "content": "C" + }, + { + "bbox": [ + 55, + 266, + 296, + 386 + ], + "type": "text", + "content": " [25] and Imagenet- " + }, + { + "bbox": [ + 55, + 266, + 296, + 386 + ], + "type": "inline_equation", + "content": "\\bar{\\mathbf{C}}" + }, + { + "bbox": [ + 55, + 266, + 296, + 386 + ], + "type": "text", + "content": " [36] accompanied with their threat assessment." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 55, + 387, + 296, + 460 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 387, + 296, + 460 + ], + "spans": [ + { + "bbox": [ + 55, + 387, + 296, + 460 + ], + "type": "text", + "content": "For a specific corruption style such as Gaussian noise and a severity level 5, we let " + }, + { + "bbox": [ + 55, + 387, + 296, + 460 + ], + "type": "inline_equation", + "content": "\\Omega_{\\mathrm{GaussianNoise},5}" + }, + { + "bbox": [ + 55, + 387, + 296, + 460 + ], + "type": "text", + "content": " denote the corresponding corruption function. To aid visualization, we group the corruption styles into 6 distinct categories (listed in Table 2) so that " + }, + { + "bbox": [ + 55, + 387, + 296, + 460 + ], + "type": "inline_equation", + "content": "\\Omega := \\cup_{i \\in [5]} \\Omega_i" + }, + { + "bbox": [ + 55, + 387, + 296, + 460 + ], + "type": "text", + "content": " where " + }, + { + "bbox": [ + 55, + 387, + 296, + 460 + ], + "type": "inline_equation", + "content": "\\Omega_i" + }, + { + "bbox": [ + 55, + 387, + 296, + 460 + ], + "type": "text", + "content": " is the set of all corruption styles at severity level " + }, + { + "bbox": [ + 55, + 387, + 296, + 460 + ], + "type": "inline_equation", + "content": "i" + }, + { + "bbox": [ + 55, + 387, + 296, + 460 + ], + "type": "text", + "content": "," + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 55, + 467, + 301, + 488 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 467, + 301, + 488 + ], + "spans": [ + { + "bbox": [ + 55, + 467, + 301, + 488 + ], + "type": "interline_equation", + "content": "\\Omega_ {i} := \\left(\\Omega_ {\\text {n o i s e}, i} \\cup \\Omega_ {\\text {b l u r}, i} \\cup \\Omega_ {\\text {c o m p r e s s i o n}, i} \\cup \\Omega_ {\\text {d i g i t a l}, i} \\cup \\Omega_ {\\text {w e a t h e r}, i} \\cup \\Omega_ {\\text {o c c l u s i o n}, i}\\right).", + "image_path": "be4dc448a707044daae5bc1f5a1ecf45db4cf10b768b80385609c309a2822c32.jpg" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 55, + 495, + 298, + 581 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 495, + 298, + 581 + ], + "spans": [ + { + "bbox": [ + 55, + 495, + 298, + 581 + ], + "type": "text", + "content": "Each individual corruption " + }, + { + "bbox": [ + 55, + 495, + 298, + 581 + ], + "type": "inline_equation", + "content": "\\omega \\in \\Omega" + }, + { + "bbox": [ + 55, + 495, + 298, + 581 + ], + "type": "text", + "content": " is applied to a subset " + }, + { + "bbox": [ + 55, + 495, + 298, + 581 + ], + "type": "inline_equation", + "content": "\\bar{\\mathbf{S}}\\subset \\mathbf{S}_{\\mathrm{val}}" + }, + { + "bbox": [ + 55, + 495, + 298, + 581 + ], + "type": "text", + "content": " of 5,000 images chosen uniformly at random. For each threat function " + }, + { + "bbox": [ + 55, + 495, + 298, + 581 + ], + "type": "inline_equation", + "content": "d\\in \\{d_{\\infty},d_2,d_{\\mathrm{PD}},d_{\\mathrm{DS}}\\}" + }, + { + "bbox": [ + 55, + 495, + 298, + 581 + ], + "type": "text", + "content": " and corruption " + }, + { + "bbox": [ + 55, + 495, + 298, + 581 + ], + "type": "inline_equation", + "content": "\\omega \\in \\Omega" + }, + { + "bbox": [ + 55, + 495, + 298, + 581 + ], + "type": "text", + "content": ", for each image " + }, + { + "bbox": [ + 55, + 495, + 298, + 581 + ], + "type": "inline_equation", + "content": "\\mathbf{x}\\in \\bar{\\mathbf{S}}" + }, + { + "bbox": [ + 55, + 495, + 298, + 581 + ], + "type": "text", + "content": ", we denote by " + }, + { + "bbox": [ + 55, + 495, + 298, + 581 + ], + "type": "inline_equation", + "content": "\\mathrm{avg}(d,\\omega)" + }, + { + "bbox": [ + 55, + 495, + 298, + 581 + ], + "type": "text", + "content": ", the average threat statistic, " + }, + { + "bbox": [ + 55, + 495, + 298, + 581 + ], + "type": "inline_equation", + "content": "\\mathrm{avg}(d,\\omega)\\coloneqq \\frac{1}{|\\bar{s}|}\\sum_{(\\mathbf{x},y)\\in S}d(\\mathbf{x},\\omega (\\mathbf{x}) - \\mathbf{x})" + }, + { + "bbox": [ + 55, + 495, + 298, + 581 + ], + "type": "text", + "content": ", where " + }, + { + "bbox": [ + 55, + 495, + 298, + 581 + ], + "type": "inline_equation", + "content": "\\omega (\\mathbf{x})" + }, + { + "bbox": [ + 55, + 495, + 298, + 581 + ], + "type": "text", + "content": " is the corrupted input and " + }, + { + "bbox": [ + 55, + 495, + 298, + 581 + ], + "type": "inline_equation", + "content": "\\omega (\\mathbf{x}) - \\mathbf{x}" + }, + { + "bbox": [ + 55, + 495, + 298, + 581 + ], + "type": "text", + "content": " is the safe perturbation at " + }, + { + "bbox": [ + 55, + 495, + 298, + 581 + ], + "type": "inline_equation", + "content": "\\mathbf{x}" + }, + { + "bbox": [ + 55, + 495, + 298, + 581 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 55, + 588, + 278, + 601 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 588, + 278, + 601 + ], + "spans": [ + { + "bbox": [ + 55, + 588, + 278, + 601 + ], + "type": "text", + "content": "4.1. Disentangling Safe and Unsafe Corruption" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 55, + 605, + 296, + 688 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 605, + 296, + 688 + ], + "spans": [ + { + "bbox": [ + 55, + 605, + 296, + 688 + ], + "type": "text", + "content": "We now measure the ability of threat models to distinguish safe and unsafe corruptions. Each corruption " + }, + { + "bbox": [ + 55, + 605, + 296, + 688 + ], + "type": "inline_equation", + "content": "\\omega \\in \\Omega" + }, + { + "bbox": [ + 55, + 605, + 296, + 688 + ], + "type": "text", + "content": " defines a point in Figure 3a having as the horizontal coordinate " + }, + { + "bbox": [ + 55, + 605, + 296, + 688 + ], + "type": "inline_equation", + "content": "\\mathrm{avg}(d_{\\infty},\\omega)" + }, + { + "bbox": [ + 55, + 605, + 296, + 688 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 55, + 605, + 296, + 688 + ], + "type": "inline_equation", + "content": "\\mathrm{avg}(d_{\\mathrm{PD}},\\omega)" + }, + { + "bbox": [ + 55, + 605, + 296, + 688 + ], + "type": "text", + "content": " as the vertical coordinate, producing a comparison of our PD threat to the " + }, + { + "bbox": [ + 55, + 605, + 296, + 688 + ], + "type": "inline_equation", + "content": "\\ell_{\\infty}" + }, + { + "bbox": [ + 55, + 605, + 296, + 688 + ], + "type": "text", + "content": "-threat model. Similarly, Figure 3b compares PD threat to the DreamSIM-threat model." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 55, + 689, + 296, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 689, + 296, + 715 + ], + "spans": [ + { + "bbox": [ + 55, + 689, + 296, + 715 + ], + "type": "text", + "content": "A specification of threat model " + }, + { + "bbox": [ + 55, + 689, + 296, + 715 + ], + "type": "inline_equation", + "content": "(d, \\varepsilon)" + }, + { + "bbox": [ + 55, + 689, + 296, + 715 + ], + "type": "text", + "content": " corresponds to the choice of a horizontal line for " + }, + { + "bbox": [ + 55, + 689, + 296, + 715 + ], + "type": "inline_equation", + "content": "d_{\\mathrm{PD}}" + }, + { + "bbox": [ + 55, + 689, + 296, + 715 + ], + "type": "text", + "content": " or a vertical line for " + }, + { + "bbox": [ + 55, + 689, + 296, + 715 + ], + "type": "inline_equation", + "content": "(d_{\\infty}" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 313, + 72, + 553, + 167 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 72, + 553, + 167 + ], + "spans": [ + { + "bbox": [ + 313, + 72, + 553, + 167 + ], + "type": "text", + "content": "or " + }, + { + "bbox": [ + 313, + 72, + 553, + 167 + ], + "type": "inline_equation", + "content": "d_{\\mathrm{DS}}" + }, + { + "bbox": [ + 313, + 72, + 553, + 167 + ], + "type": "text", + "content": ") in Figure 3. We note that a size-able set of common corruptions incur large " + }, + { + "bbox": [ + 313, + 72, + 553, + 167 + ], + "type": "inline_equation", + "content": "d_{\\infty}" + }, + { + "bbox": [ + 313, + 72, + 553, + 167 + ], + "type": "text", + "content": " threat, in particular as large as the unsafe corruptions (as indicated in Figure 1), thus any vertical line in Figure 3a is bound to exclude such common corruptions. In contrast, both " + }, + { + "bbox": [ + 313, + 72, + 553, + 167 + ], + "type": "inline_equation", + "content": "d_{\\mathrm{PD}}" + }, + { + "bbox": [ + 313, + 72, + 553, + 167 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 313, + 72, + 553, + 167 + ], + "type": "inline_equation", + "content": "d_{\\mathrm{DS}}" + }, + { + "bbox": [ + 313, + 72, + 553, + 167 + ], + "type": "text", + "content": " admit choices of " + }, + { + "bbox": [ + 313, + 72, + 553, + 167 + ], + "type": "inline_equation", + "content": "\\varepsilon" + }, + { + "bbox": [ + 313, + 72, + 553, + 167 + ], + "type": "text", + "content": " that exclude unsafe perturbations while including most common corruptions. Hence, PD threat and DS threat are able to disentangle safe and unsafe corruptions." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 313, + 169, + 555, + 241 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 169, + 555, + 241 + ], + "spans": [ + { + "bbox": [ + 313, + 169, + 555, + 241 + ], + "type": "text", + "content": "For each of the comparisons in Figure 3, we identify 4 quadrants I-IV, distinguishing between low-threat and high-threat corruptions as measured by the corresponding threat models. We define the quadrants by choosing thresholds of 1, 0.5, and 0.25 for " + }, + { + "bbox": [ + 313, + 169, + 555, + 241 + ], + "type": "inline_equation", + "content": "d_{\\mathrm{PD}}" + }, + { + "bbox": [ + 313, + 169, + 555, + 241 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 313, + 169, + 555, + 241 + ], + "type": "inline_equation", + "content": "d_{\\infty}" + }, + { + "bbox": [ + 313, + 169, + 555, + 241 + ], + "type": "text", + "content": ", and " + }, + { + "bbox": [ + 313, + 169, + 555, + 241 + ], + "type": "inline_equation", + "content": "d_{\\mathrm{DS}}" + }, + { + "bbox": [ + 313, + 169, + 555, + 241 + ], + "type": "text", + "content": ", respectively9, and comment on the qualitative behavior observed:" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 314, + 242, + 555, + 326 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 314, + 242, + 555, + 326 + ], + "spans": [ + { + "bbox": [ + 314, + 242, + 555, + 326 + ], + "type": "text", + "content": "- Quadrant IV contains corruptions that are characterized as low-threat by both threat models. As expected, almost all corruptions having a low severity (1) lie in this region. However, we already see deficiencies of the " + }, + { + "bbox": [ + 314, + 242, + 555, + 326 + ], + "type": "inline_equation", + "content": "d_{\\infty}" + }, + { + "bbox": [ + 314, + 242, + 555, + 326 + ], + "type": "text", + "content": " threat, as some blur, noise, and occlusion corruptions of severity 1 in Figure 3a are not characterized as low-threat by at least one of the threat models." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 314, + 326, + 555, + 373 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 314, + 326, + 555, + 373 + ], + "spans": [ + { + "bbox": [ + 314, + 326, + 555, + 373 + ], + "type": "text", + "content": "- Quadrant II contains corruptions that are characterized as high-threat by both threat models. These include digital and weather corruptions of high severity (5) as they tend to remove most of the signal in the input." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 314, + 374, + 556, + 491 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 314, + 374, + 556, + 491 + ], + "spans": [ + { + "bbox": [ + 314, + 374, + 556, + 491 + ], + "type": "text", + "content": "- Quadrant I contains corruptions having low PD-threat but high " + }, + { + "bbox": [ + 314, + 374, + 556, + 491 + ], + "type": "inline_equation", + "content": "d_{\\infty}" + }, + { + "bbox": [ + 314, + 374, + 556, + 491 + ], + "type": "text", + "content": " or " + }, + { + "bbox": [ + 314, + 374, + 556, + 491 + ], + "type": "inline_equation", + "content": "d_{\\mathrm{DS}}" + }, + { + "bbox": [ + 314, + 374, + 556, + 491 + ], + "type": "text", + "content": " and showcases the resilience of our PD-threat model to natural corruptions like blur and Gaussian noise, which retain a lot of the information in the image relevant to the classification task. Indeed, a visual inspection of these corruptions (see Section 10 in the Appendix) reveals that even at a high severity, a human is able to discern task-relevant information from the corrupted images, and hence these corruptions should be assigned a low threat." + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 314, + 493, + 556, + 566 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 314, + 493, + 556, + 566 + ], + "spans": [ + { + "bbox": [ + 314, + 493, + 556, + 566 + ], + "type": "text", + "content": "- Quadrant III contains corruptions with high PD threat but low " + }, + { + "bbox": [ + 314, + 493, + 556, + 566 + ], + "type": "inline_equation", + "content": "d_{\\infty}" + }, + { + "bbox": [ + 314, + 493, + 556, + 566 + ], + "type": "text", + "content": " or " + }, + { + "bbox": [ + 314, + 493, + 556, + 566 + ], + "type": "inline_equation", + "content": "d_{\\mathrm{DS}}" + }, + { + "bbox": [ + 314, + 493, + 556, + 566 + ], + "type": "text", + "content": " threat. We note that corruptions rate as high threat by " + }, + { + "bbox": [ + 314, + 493, + 556, + 566 + ], + "type": "inline_equation", + "content": "d_{\\mathrm{PD}}" + }, + { + "bbox": [ + 314, + 493, + 556, + 566 + ], + "type": "text", + "content": " are largely also rated highly by " + }, + { + "bbox": [ + 314, + 493, + 556, + 566 + ], + "type": "inline_equation", + "content": "d_{\\mathrm{DS}}" + }, + { + "bbox": [ + 314, + 493, + 556, + 566 + ], + "type": "text", + "content": ". Since the PD and DS threats are aligned, the corruptions in Quadrant III in Figure 3a indicate that low " + }, + { + "bbox": [ + 314, + 493, + 556, + 566 + ], + "type": "inline_equation", + "content": "d_{\\infty}" + }, + { + "bbox": [ + 314, + 493, + 556, + 566 + ], + "type": "text", + "content": " is not sufficient to characterize hard corruptions." + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 313, + 567, + 556, + 662 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 567, + 556, + 662 + ], + "spans": [ + { + "bbox": [ + 313, + 567, + 556, + 662 + ], + "type": "text", + "content": "As a final point, we note that while the size of markers in Figure 3 indicate the severity levels, different corruptions of similar severity levels are qualitatively different. For e.g, weather corruptions of severity level 5 are uniformly harder than noise corruptions such as shot noise, Gaussian noise. Figure 11 illustrates heatmap of average threat for each corruption category across increasing severity. We observe that both PD and DS threat exhibit largely monotonic" + } + ] + } + ], + "index": 15 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 313, + 673, + 555, + 715 + ], + "type": "page_footnote", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 673, + 555, + 715 + ], + "spans": [ + { + "bbox": [ + 313, + 673, + 555, + 715 + ], + "type": "text", + "content": "9These thresholds were chosen to approximately be half of the average threat of unsafe perturbations that alter the true label. Admittedly, this binary threshold is coarse, and the Appendix contains fine-grained qualitative examples of corruptions of varying threat." + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 295, + 750, + 315, + 758 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 750, + 315, + 758 + ], + "spans": [ + { + "bbox": [ + 295, + 750, + 315, + 758 + ], + "type": "text", + "content": "9959" + } + ] + } + ], + "index": 17 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 5 + }, + { + "para_blocks": [ + { + "bbox": [ + 55, + 72, + 297, + 96 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 72, + 297, + 96 + ], + "spans": [ + { + "bbox": [ + 55, + 72, + 297, + 96 + ], + "type": "text", + "content": "increase across severity unlike " + }, + { + "bbox": [ + 55, + 72, + 297, + 96 + ], + "type": "inline_equation", + "content": "d_{\\infty}" + }, + { + "bbox": [ + 55, + 72, + 297, + 96 + ], + "type": "text", + "content": ". At this point, we will pause to summarize the benefits of PD-threat." + } + ] + } + ], + "index": 0 + }, + { + "type": "image", + "bbox": [ + 78, + 107, + 272, + 306 + ], + "blocks": [ + { + "bbox": [ + 78, + 107, + 272, + 306 + ], + "lines": [ + { + "bbox": [ + 78, + 107, + 272, + 306 + ], + "spans": [ + { + "bbox": [ + 78, + 107, + 272, + 306 + ], + "type": "image", + "image_path": "920cca1a7bf6f7190f4a454ce4244bf8e2d3d453e4edd1c6663d204ac5f4d784.jpg" + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 170, + 307, + 181, + 316 + ], + "lines": [ + { + "bbox": [ + 170, + 307, + 181, + 316 + ], + "spans": [ + { + "bbox": [ + 170, + 307, + 181, + 316 + ], + "type": "text", + "content": "(a)" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_caption" + } + ], + "index": 1 + }, + { + "type": "image", + "bbox": [ + 78, + 317, + 272, + 499 + ], + "blocks": [ + { + "bbox": [ + 78, + 317, + 272, + 499 + ], + "lines": [ + { + "bbox": [ + 78, + 317, + 272, + 499 + ], + "spans": [ + { + "bbox": [ + 78, + 317, + 272, + 499 + ], + "type": "image", + "image_path": "c1b23194f51f9e38ce127382832edc10c7be4637cd15b6dcd3e547ebe6e6624c.jpg" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 170, + 500, + 181, + 509 + ], + "lines": [ + { + "bbox": [ + 170, + 500, + 181, + 509 + ], + "spans": [ + { + "bbox": [ + 170, + 500, + 181, + 509 + ], + "type": "text", + "content": "(b)" + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_caption" + } + ], + "index": 3 + }, + { + "bbox": [ + 55, + 558, + 297, + 712 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 558, + 297, + 712 + ], + "spans": [ + { + "bbox": [ + 55, + 558, + 297, + 712 + ], + "type": "text", + "content": "PD vs DreamSim. The PD-threat compares well to the DS-threat but does not require pre-trained large vision models or fine-tuning on a curated dataset of human perceptual similarity judgements. For real-world classification tasks where perceptual similarity between pairs of input is harder to quantify even with expert human annotators, instantiation of high-quality neural perceptual distance metrics is harder. Further, unlike perceptual distance metrics [15, 17, 18, 32, 54], the growth of the PD-threat is linear and interpretable along any direction, the sub-level sets induced by the PD-threat model are convex, and admit efficient projection algorithms and hence PD threat presents a promising alternative." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 313, + 72, + 533, + 85 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 72, + 533, + 85 + ], + "spans": [ + { + "bbox": [ + 313, + 72, + 533, + 85 + ], + "type": "text", + "content": "5. Measuring threat with Task Annotation" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 313, + 91, + 555, + 150 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 91, + 555, + 150 + ], + "spans": [ + { + "bbox": [ + 313, + 91, + 555, + 150 + ], + "type": "text", + "content": "We described two practical design choices for PD-threat in Section 3.2: (1) a choice of the set of unsafe directions " + }, + { + "bbox": [ + 313, + 91, + 555, + 150 + ], + "type": "inline_equation", + "content": "\\mathcal{U}(\\mathbf{x})" + }, + { + "bbox": [ + 313, + 91, + 555, + 150 + ], + "type": "text", + "content": " at each input, and (2) a choice of normalization " + }, + { + "bbox": [ + 313, + 91, + 555, + 150 + ], + "type": "inline_equation", + "content": "g(\\mathbf{x},\\mathbf{u})" + }, + { + "bbox": [ + 313, + 91, + 555, + 150 + ], + "type": "text", + "content": " for each unsafe direction. Next, we incorporate additional task-relevant information to refine PD-threat further." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 313, + 158, + 529, + 171 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 158, + 529, + 171 + ], + "spans": [ + { + "bbox": [ + 313, + 158, + 529, + 171 + ], + "type": "text", + "content": "5.1. Segmentation-aware Threat Specification" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 313, + 176, + 555, + 247 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 176, + 555, + 247 + ], + "spans": [ + { + "bbox": [ + 313, + 176, + 555, + 247 + ], + "type": "text", + "content": "Segmentation masks isolate regions of the input that contain semantic information relevant to the class label, and can be generated efficiently following recent advances in foundational models for image segmentation [30, 38]. We demonstrate a straightforward method to incorporate semantic masks in the assessment of threat." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 313, + 254, + 555, + 279 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 254, + 555, + 279 + ], + "spans": [ + { + "bbox": [ + 313, + 254, + 555, + 279 + ], + "type": "text", + "content": "Definition 8 (PD-S threat). Let " + }, + { + "bbox": [ + 313, + 254, + 555, + 279 + ], + "type": "inline_equation", + "content": "\\mathbf{x} \\in \\mathcal{X}" + }, + { + "bbox": [ + 313, + 254, + 555, + 279 + ], + "type": "text", + "content": " and let " + }, + { + "bbox": [ + 313, + 254, + 555, + 279 + ], + "type": "inline_equation", + "content": "\\delta \\in \\mathbb{R}^d" + }, + { + "bbox": [ + 313, + 254, + 555, + 279 + ], + "type": "text", + "content": ". Let " + }, + { + "bbox": [ + 313, + 254, + 555, + 279 + ], + "type": "inline_equation", + "content": "\\mathbf{a} \\in \\{0,1\\}^d" + }, + { + "bbox": [ + 313, + 254, + 555, + 279 + ], + "type": "text", + "content": " be a boolean mask. The segmented PD threat," + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 313, + 296, + 559, + 322 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 296, + 559, + 322 + ], + "spans": [ + { + "bbox": [ + 313, + 296, + 559, + 322 + ], + "type": "interline_equation", + "content": "d _ {\\mathrm {P D - S}} (\\mathbf {x}, \\mathbf {a}, \\boldsymbol {\\delta}) := \\max _ {\\mathbf {u} \\in \\mathcal {U} (\\mathbf {x})} \\frac {1}{g _ {\\boldsymbol {\\beta}} (\\mathbf {x} , \\mathbf {u})} \\max \\left(\\langle \\boldsymbol {\\delta} [ \\mathbf {a} ], \\mathbf {u} [ \\mathbf {a} ] \\rangle , 0\\right).", + "image_path": "613363ff1d93f2f3ba25c2d5f2b4e964730cebb9345716aa1c7984744e4926b5.jpg" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 313, + 329, + 556, + 449 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 329, + 556, + 449 + ], + "spans": [ + { + "bbox": [ + 313, + 329, + 556, + 449 + ], + "type": "text", + "content": "Here the " + }, + { + "bbox": [ + 313, + 329, + 556, + 449 + ], + "type": "inline_equation", + "content": "\\mathbf{u}[\\mathbf{a}]" + }, + { + "bbox": [ + 313, + 329, + 556, + 449 + ], + "type": "text", + "content": " denotes a sub-vector whose indices are selected by the boolean mask " + }, + { + "bbox": [ + 313, + 329, + 556, + 449 + ], + "type": "inline_equation", + "content": "\\mathbf{a}" + }, + { + "bbox": [ + 313, + 329, + 556, + 449 + ], + "type": "text", + "content": ". The threat model " + }, + { + "bbox": [ + 313, + 329, + 556, + 449 + ], + "type": "inline_equation", + "content": "(d_{\\mathrm{PD-S}}, \\varepsilon)" + }, + { + "bbox": [ + 313, + 329, + 556, + 449 + ], + "type": "text", + "content": " is aligned with Xiao et al. [51]'s proposal to learn models robust to adversarially chosen backgrounds. For each image, we generate automatic masks by prompting SAM10 with the center pixel coordinates. Figure 4 illustrates the difference between PD and PD-S threat on 2 corruptions that are background-only and foreground-only respectively. PD-S threat is oblivious to background corruption but more sensitive to foreground corruption." + } + ] + } + ], + "index": 13 + }, + { + "type": "image", + "bbox": [ + 325, + 459, + 378, + 513 + ], + "blocks": [ + { + "bbox": [ + 55, + 520, + 295, + 542 + ], + "lines": [ + { + "bbox": [ + 55, + 520, + 295, + 542 + ], + "spans": [ + { + "bbox": [ + 55, + 520, + 295, + 542 + ], + "type": "text", + "content": "Figure 3. Comparison of Threat models w.r.t. various common corruptions. The size of a marker represents the severity level." + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 325, + 459, + 378, + 513 + ], + "lines": [ + { + "bbox": [ + 325, + 459, + 378, + 513 + ], + "spans": [ + { + "bbox": [ + 325, + 459, + 378, + 513 + ], + "type": "image", + "image_path": "5462da7c42ae9bf65700ad81708555cfdd3bd5f026366f2bb6f22c619c2eb01e.jpg" + } + ] + } + ], + "index": 14, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 335, + 517, + 368, + 524 + ], + "lines": [ + { + "bbox": [ + 335, + 517, + 368, + 524 + ], + "spans": [ + { + "bbox": [ + 335, + 517, + 368, + 524 + ], + "type": "text", + "content": "Original Image" + } + ] + } + ], + "index": 15, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 320, + 540, + 547, + 551 + ], + "lines": [ + { + "bbox": [ + 320, + 540, + 547, + 551 + ], + "spans": [ + { + "bbox": [ + 320, + 540, + 547, + 551 + ], + "type": "text", + "content": "Figure 4. Comparison of threat on background vs foreground." + } + ] + } + ], + "index": 22, + "angle": 0, + "type": "image_caption" + } + ], + "index": 14 + }, + { + "type": "image", + "bbox": [ + 408, + 459, + 460, + 512 + ], + "blocks": [ + { + "bbox": [ + 408, + 459, + 460, + 512 + ], + "lines": [ + { + "bbox": [ + 408, + 459, + 460, + 512 + ], + "spans": [ + { + "bbox": [ + 408, + 459, + 460, + 512 + ], + "type": "image", + "image_path": "24e99c7989721abcb7274ffba9b6228b921d7bcd540f9bead52f9a64c2bb8d3d.jpg" + } + ] + } + ], + "index": 16, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 424, + 517, + 445, + 524 + ], + "lines": [ + { + "bbox": [ + 424, + 517, + 445, + 524 + ], + "spans": [ + { + "bbox": [ + 424, + 517, + 445, + 524 + ], + "type": "inline_equation", + "content": "\\mathrm{PD} = 1.4" + } + ] + } + ], + "index": 17, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 422, + 524, + 446, + 529 + ], + "lines": [ + { + "bbox": [ + 422, + 524, + 446, + 529 + ], + "spans": [ + { + "bbox": [ + 422, + 524, + 446, + 529 + ], + "type": "text", + "content": "PD-S=0.0" + } + ] + } + ], + "index": 18, + "angle": 0, + "type": "image_caption" + } + ], + "index": 16 + }, + { + "type": "image", + "bbox": [ + 490, + 459, + 543, + 512 + ], + "blocks": [ + { + "bbox": [ + 490, + 459, + 543, + 512 + ], + "lines": [ + { + "bbox": [ + 490, + 459, + 543, + 512 + ], + "spans": [ + { + "bbox": [ + 490, + 459, + 543, + 512 + ], + "type": "image", + "image_path": "c664177ece929bf1a2a5d80009090d35a775d5cd81f42704bbb36807eab55069.jpg" + } + ] + } + ], + "index": 19, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 505, + 517, + 528, + 524 + ], + "lines": [ + { + "bbox": [ + 505, + 517, + 528, + 524 + ], + "spans": [ + { + "bbox": [ + 505, + 517, + 528, + 524 + ], + "type": "inline_equation", + "content": "\\mathrm{PD} = 0.68" + } + ] + } + ], + "index": 20, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 504, + 524, + 530, + 529 + ], + "lines": [ + { + "bbox": [ + 504, + 524, + 530, + 529 + ], + "spans": [ + { + "bbox": [ + 504, + 524, + 530, + 529 + ], + "type": "text", + "content": "PD-S=1.78" + } + ] + } + ], + "index": 21, + "angle": 0, + "type": "image_caption" + } + ], + "index": 19 + }, + { + "bbox": [ + 313, + 563, + 555, + 599 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 563, + 555, + 599 + ], + "spans": [ + { + "bbox": [ + 313, + 563, + 555, + 599 + ], + "type": "text", + "content": "We note that " + }, + { + "bbox": [ + 313, + 563, + 555, + 599 + ], + "type": "inline_equation", + "content": "\\ell_p" + }, + { + "bbox": [ + 313, + 563, + 555, + 599 + ], + "type": "text", + "content": " threat models are also capable of integrating segmentation masks as they are pixel-based, but it is unclear if perceptual threat can account for pixel annotation." + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 313, + 606, + 505, + 620 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 606, + 505, + 620 + ], + "spans": [ + { + "bbox": [ + 313, + 606, + 505, + 620 + ], + "type": "text", + "content": "5.2. Concept-aware Threat Specification" + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 313, + 624, + 555, + 696 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 624, + 555, + 696 + ], + "spans": [ + { + "bbox": [ + 313, + 624, + 555, + 696 + ], + "type": "text", + "content": "The PD-threat assumes no hierarchy of class labels; i.e. all labels are distinct to the same degree. However, classification tasks are often accompanied with an additional hierarchy that identifies groups of class labels based on similar semantic content. We propose to account for such a hierarchy by refining our choice of normalization. Suppose " + }, + { + "bbox": [ + 313, + 624, + 555, + 696 + ], + "type": "inline_equation", + "content": "\\mathbf{x}" + }, + { + "bbox": [ + 313, + 624, + 555, + 696 + ], + "type": "text", + "content": " has label" + } + ] + } + ], + "index": 25 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 323, + 702, + 511, + 713 + ], + "type": "page_footnote", + "angle": 0, + "lines": [ + { + "bbox": [ + 323, + 702, + 511, + 713 + ], + "spans": [ + { + "bbox": [ + 323, + 702, + 511, + 713 + ], + "type": "text", + "content": "10In particular we use the default ViT-H SAM checkpoint." + } + ] + } + ], + "index": 26 + }, + { + "bbox": [ + 295, + 750, + 315, + 758 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 750, + 315, + 758 + ], + "spans": [ + { + "bbox": [ + 295, + 750, + 315, + 758 + ], + "type": "text", + "content": "9960" + } + ] + } + ], + "index": 27 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 6 + }, + { + "para_blocks": [ + { + "bbox": [ + 55, + 72, + 296, + 147 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 72, + 296, + 147 + ], + "spans": [ + { + "bbox": [ + 55, + 72, + 296, + 147 + ], + "type": "inline_equation", + "content": "y" + }, + { + "bbox": [ + 55, + 72, + 296, + 147 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 55, + 72, + 296, + 147 + ], + "type": "inline_equation", + "content": "\\mathbf{u} = \\frac{\\tilde{\\mathbf{x}} - \\mathbf{x}}{\\|\\tilde{\\mathbf{x}} - \\mathbf{x}\\|_2}" + }, + { + "bbox": [ + 55, + 72, + 296, + 147 + ], + "type": "text", + "content": " where " + }, + { + "bbox": [ + 55, + 72, + 296, + 147 + ], + "type": "inline_equation", + "content": "\\tilde{\\mathbf{x}}\\in S_{c,k}" + }, + { + "bbox": [ + 55, + 72, + 296, + 147 + ], + "type": "text", + "content": " has label c. Let " + }, + { + "bbox": [ + 55, + 72, + 296, + 147 + ], + "type": "inline_equation", + "content": "W:[C]\\times [C]\\in [0,1]" + }, + { + "bbox": [ + 55, + 72, + 296, + 147 + ], + "type": "text", + "content": " denote a relative distance between class labels based on Wordnet-hierarchy (refer to Section 9.5 for explicit details). We propose the weighted normalization " + }, + { + "bbox": [ + 55, + 72, + 296, + 147 + ], + "type": "inline_equation", + "content": "g_{\\beta ,W}(\\mathbf{x},\\mathbf{u})\\coloneqq \\boldsymbol {\\beta}\\cdot W(y,c)\\| \\tilde{\\mathbf{x}} -\\mathbf{x}\\| _2" + }, + { + "bbox": [ + 55, + 72, + 296, + 147 + ], + "type": "text", + "content": " so that perturbation between nearby classes are weighted higher in threat." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 55, + 152, + 296, + 176 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 152, + 296, + 176 + ], + "spans": [ + { + "bbox": [ + 55, + 152, + 296, + 176 + ], + "type": "text", + "content": "Definition 9 (PD-W threat). Let " + }, + { + "bbox": [ + 55, + 152, + 296, + 176 + ], + "type": "inline_equation", + "content": "\\mathbf{x} \\in \\mathcal{X}" + }, + { + "bbox": [ + 55, + 152, + 296, + 176 + ], + "type": "text", + "content": " with label " + }, + { + "bbox": [ + 55, + 152, + 296, + 176 + ], + "type": "inline_equation", + "content": "y" + }, + { + "bbox": [ + 55, + 152, + 296, + 176 + ], + "type": "text", + "content": " and let " + }, + { + "bbox": [ + 55, + 152, + 296, + 176 + ], + "type": "inline_equation", + "content": "\\pmb{\\delta} \\in \\mathbb{R}^{d}" + }, + { + "bbox": [ + 55, + 152, + 296, + 176 + ], + "type": "text", + "content": ". The weighted PD threat is defined as," + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 61, + 180, + 290, + 206 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 61, + 180, + 290, + 206 + ], + "spans": [ + { + "bbox": [ + 61, + 180, + 290, + 206 + ], + "type": "interline_equation", + "content": "d_{\\mathrm{PD - W}}(\\mathbf{x},\\boldsymbol {\\delta}):= \\max_{\\mathbf{u}\\in \\mathcal{U}(\\mathbf{x})}\\frac{1}{g_{\\beta,W}(\\mathbf{x},\\mathbf{u})}\\max \\left(\\langle \\boldsymbol {\\delta},\\mathbf{u}\\rangle ,0\\right).", + "image_path": "f642b4e15b20c895289237490a6776a7a216c00305c34334b66118a27f652eb9.jpg" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 55, + 211, + 298, + 354 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 211, + 298, + 354 + ], + "spans": [ + { + "bbox": [ + 55, + 211, + 298, + 354 + ], + "type": "text", + "content": "The threat model " + }, + { + "bbox": [ + 55, + 211, + 298, + 354 + ], + "type": "inline_equation", + "content": "(d_{\\mathrm{PD - W}},\\varepsilon)" + }, + { + "bbox": [ + 55, + 211, + 298, + 354 + ], + "type": "text", + "content": " is weaker than " + }, + { + "bbox": [ + 55, + 211, + 298, + 354 + ], + "type": "inline_equation", + "content": "(d_{\\mathrm{PD}},\\varepsilon)" + }, + { + "bbox": [ + 55, + 211, + 298, + 354 + ], + "type": "text", + "content": ", since " + }, + { + "bbox": [ + 55, + 211, + 298, + 354 + ], + "type": "inline_equation", + "content": "S(\\mathbf{x},d_{\\mathrm{PD - W}},\\varepsilon)\\subseteq S(\\mathbf{x},d_{\\mathrm{PD}},\\varepsilon)" + }, + { + "bbox": [ + 55, + 211, + 298, + 354 + ], + "type": "text", + "content": ". Thus PD-W is a relaxation of PD threat with a softer requirement of stability between nearby classes. Figure 5 depicts the average relative" + }, + { + "bbox": [ + 55, + 211, + 298, + 354 + ], + "type": "inline_equation", + "content": "^{12}" + }, + { + "bbox": [ + 55, + 211, + 298, + 354 + ], + "type": "text", + "content": " threat vs relative distance " + }, + { + "bbox": [ + 55, + 211, + 298, + 354 + ], + "type": "inline_equation", + "content": "W(\\cdot ,\\cdot)" + }, + { + "bbox": [ + 55, + 211, + 298, + 354 + ], + "type": "text", + "content": " of class labels on the Imagenet-1k validation dataset. The decreasing PD-W threat for increasing class distance is aligned with the following intuition - failure to distinguish between semantically distant classes such as ENGLISH FOXHOUND and FIRE TRUCK should incur lower robust accuracy than failure to distinguish between similar classes such as ENGLISH FOXHOUND and IRISH WOLFHOUND." + } + ] + } + ], + "index": 3 + }, + { + "type": "image", + "bbox": [ + 91, + 370, + 245, + 492 + ], + "blocks": [ + { + "bbox": [ + 91, + 370, + 245, + 492 + ], + "lines": [ + { + "bbox": [ + 91, + 370, + 245, + 492 + ], + "spans": [ + { + "bbox": [ + 91, + 370, + 245, + 492 + ], + "type": "image", + "image_path": "37f4b90c34d11b1ef600464461f5cc220b8ed0f11c89285aa67e83458ed08837.jpg" + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 70, + 504, + 280, + 515 + ], + "lines": [ + { + "bbox": [ + 70, + 504, + 280, + 515 + ], + "spans": [ + { + "bbox": [ + 70, + 504, + 280, + 515 + ], + "type": "text", + "content": "Figure 5. Relative Threat vs relative inter-class distances" + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "image_caption" + } + ], + "index": 4 + }, + { + "bbox": [ + 55, + 529, + 297, + 624 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 529, + 297, + 624 + ], + "spans": [ + { + "bbox": [ + 55, + 529, + 297, + 624 + ], + "type": "text", + "content": "In summary our novel threat specification framework enables evaluation of robustness to adversarial perturbations and common corruptions while enabling sensitivity to semantic regions of an image and class hierarchy by incorporating additional task annotation. As a consequence, our threat model unifies adversarial robustness with variants of corruption robustness suggested by several independent benchmarks [25, 36, 51]." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 55, + 634, + 289, + 648 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 634, + 289, + 648 + ], + "spans": [ + { + "bbox": [ + 55, + 634, + 289, + 648 + ], + "type": "text", + "content": "6. Evaluation of benchmark " + }, + { + "bbox": [ + 55, + 634, + 289, + 648 + ], + "type": "inline_equation", + "content": "\\ell_p" + }, + { + "bbox": [ + 55, + 634, + 289, + 648 + ], + "type": "text", + "content": " robust models" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 55, + 654, + 296, + 678 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 654, + 296, + 678 + ], + "spans": [ + { + "bbox": [ + 55, + 654, + 296, + 678 + ], + "type": "text", + "content": "Finally, we evaluate state-of-the-art robust models against our proposed thread models. Note that, as per Theorem 1," + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 313, + 72, + 555, + 228 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 72, + 555, + 228 + ], + "spans": [ + { + "bbox": [ + 313, + 72, + 555, + 228 + ], + "type": "text", + "content": "one should hope for classifiers that are 1-robust to corruptions under the PD-threat. The corresponding permissible set of corruptions cover regions IV and I in Figure 3. We evaluate robustness to corruptions in the set " + }, + { + "bbox": [ + 313, + 72, + 555, + 228 + ], + "type": "inline_equation", + "content": "S(\\mathbf{x}, d_{\\mathrm{PD}}, \\varepsilon) \\cap S(\\mathbf{x}, d_{\\infty}, \\varepsilon)" + }, + { + "bbox": [ + 313, + 72, + 555, + 228 + ], + "type": "text", + "content": ". We denote the corresponding threat model " + }, + { + "bbox": [ + 313, + 72, + 555, + 228 + ], + "type": "inline_equation", + "content": "(d_{\\infty} \\bigcap d_{\\mathrm{PD}}, \\varepsilon)" + }, + { + "bbox": [ + 313, + 72, + 555, + 228 + ], + "type": "text", + "content": ". A robust evaluation for such a threat model can be executed in practice by generating adversarial perturbations with standard AutoAttack [7] for " + }, + { + "bbox": [ + 313, + 72, + 555, + 228 + ], + "type": "inline_equation", + "content": "(d_{\\infty}, \\varepsilon)" + }, + { + "bbox": [ + 313, + 72, + 555, + 228 + ], + "type": "text", + "content": " followed by projection onto " + }, + { + "bbox": [ + 313, + 72, + 555, + 228 + ], + "type": "inline_equation", + "content": "S(\\mathbf{x}, d_{\\mathrm{PD}}, \\varepsilon)" + }, + { + "bbox": [ + 313, + 72, + 555, + 228 + ], + "type": "text", + "content": ". Table 3 shows the robustness of benchmark classifiers registered at RobustBench order by " + }, + { + "bbox": [ + 313, + 72, + 555, + 228 + ], + "type": "inline_equation", + "content": "d_{\\infty}" + }, + { + "bbox": [ + 313, + 72, + 555, + 228 + ], + "type": "text", + "content": " robustness (column 1) when evaluated against " + }, + { + "bbox": [ + 313, + 72, + 555, + 228 + ], + "type": "inline_equation", + "content": "(d_{\\infty} \\bigcap d_{\\mathrm{PD}}, 16/255)" + }, + { + "bbox": [ + 313, + 72, + 555, + 228 + ], + "type": "text", + "content": " threat model for the Imagenet-1k classification task." + } + ] + } + ], + "index": 9 + }, + { + "type": "table", + "bbox": [ + 319, + 237, + 550, + 335 + ], + "blocks": [ + { + "bbox": [ + 319, + 237, + 550, + 335 + ], + "lines": [ + { + "bbox": [ + 319, + 237, + 550, + 335 + ], + "spans": [ + { + "bbox": [ + 319, + 237, + 550, + 335 + ], + "type": "table", + "html": "
Benchmark Modeld∞PDPD-SPD-W
ConvNeXt-L [33]0.250.320.310.31
Swin-B [33]0.120.210.200.21
ConvNeXt-B [33]0.10.190.180.18
ConvNeXt-B-ConvStem [41]0.090.190.190.19
ViT-S-ConvStem [41]0.050.130.120.12
XCiT-L12 [10]0.040.170.180.18
", + "image_path": "b53163e26126063e2fe08fef8735ec813786c135dde85d6f95aa49a37b4d333b.jpg" + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "table_body" + } + ], + "index": 10 + }, + { + "bbox": [ + 352, + 342, + 515, + 354 + ], + "lines": [ + { + "bbox": [ + 352, + 342, + 515, + 354 + ], + "spans": [ + { + "bbox": [ + 352, + 342, + 515, + 354 + ], + "type": "text", + "content": "Table 3. Robust Evaluation for " + }, + { + "bbox": [ + 352, + 342, + 515, + 354 + ], + "type": "inline_equation", + "content": "\\varepsilon = {16}/{255}" + }, + { + "bbox": [ + 352, + 342, + 515, + 354 + ], + "type": "text", + "content": " ." + } + ] + } + ], + "index": 11, + "angle": 0, + "type": "text" + }, + { + "bbox": [ + 313, + 367, + 556, + 522 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 367, + 556, + 522 + ], + "spans": [ + { + "bbox": [ + 313, + 367, + 556, + 522 + ], + "type": "text", + "content": "We note that robust accuracies uniformly improve under our threat specification in comparison to " + }, + { + "bbox": [ + 313, + 367, + 556, + 522 + ], + "type": "inline_equation", + "content": "(d_{\\infty}, \\varepsilon)" + }, + { + "bbox": [ + 313, + 367, + 556, + 522 + ], + "type": "text", + "content": ", indicating a meaningful restriction of unsafe corruptions. The ordering of robust accuracies highlights diversity, e.g. [10]'s XCiT-L12 fares better under PD threat than suggested by the " + }, + { + "bbox": [ + 313, + 367, + 556, + 522 + ], + "type": "inline_equation", + "content": "d_{\\infty}" + }, + { + "bbox": [ + 313, + 367, + 556, + 522 + ], + "type": "text", + "content": " leaderboard. The indistinguishability across PD, PD-S and PD-w is surprising. Similar levels of robustness w.r.t PD-W and PD indicate that models are unable to distinguish distant classes, while similar levels under PD and PD-S indicate adversarial attacks are often aligned with foreground masks. We note that adversarial attacks adapted to the threat specification often lead to a more accurate pessimistic evaluation of robustness." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 314, + 533, + 388, + 546 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 314, + 533, + 388, + 546 + ], + "spans": [ + { + "bbox": [ + 314, + 533, + 388, + 546 + ], + "type": "text", + "content": "7. Conclusion" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 313, + 554, + 556, + 685 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 554, + 556, + 685 + ], + "spans": [ + { + "bbox": [ + 313, + 554, + 556, + 685 + ], + "type": "text", + "content": "This work proposes a novel task-dependent threat specification Projected Displacement threat that is adapted to the shape of decision boundaries based on observed training data. Unlike standard " + }, + { + "bbox": [ + 313, + 554, + 556, + 685 + ], + "type": "inline_equation", + "content": "\\ell_p" + }, + { + "bbox": [ + 313, + 554, + 556, + 685 + ], + "type": "text", + "content": "-threat models, the proposed threat model exhibits anisotropy and locality and is thus able to distinguish between safe and unsafe perturbations. The proposed threat specification framework is flexible and can effectively account for additional task information, such as image regions or label hierarchies. We invite the community to suggest further adaptive attacks and effective training strategies to encourage robustness in PD threat." + } + ] + } + ], + "index": 14 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 64, + 683, + 241, + 693 + ], + "type": "page_footnote", + "angle": 0, + "lines": [ + { + "bbox": [ + 64, + 683, + 241, + 693 + ], + "spans": [ + { + "bbox": [ + 64, + 683, + 241, + 693 + ], + "type": "text", + "content": "11 Relative since we require a normalized value in [0,1]" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 56, + 693, + 295, + 712 + ], + "type": "page_footnote", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 693, + 295, + 712 + ], + "spans": [ + { + "bbox": [ + 56, + 693, + 295, + 712 + ], + "type": "inline_equation", + "content": "^{12}" + }, + { + "bbox": [ + 56, + 693, + 295, + 712 + ], + "type": "text", + "content": "Due to different scaling of PD-W and PD, we visualize average threat in each relative to maximum value on data." + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 295, + 750, + 314, + 758 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 750, + 314, + 758 + ], + "spans": [ + { + "bbox": [ + 295, + 750, + 314, + 758 + ], + "type": "text", + "content": "9961" + } + ] + } + ], + "index": 18 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 7 + }, + { + "para_blocks": [ + { + "bbox": [ + 56, + 72, + 158, + 85 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 72, + 158, + 85 + ], + "spans": [ + { + "bbox": [ + 56, + 72, + 158, + 85 + ], + "type": "text", + "content": "Acknowledgements" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 55, + 90, + 297, + 113 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 90, + 297, + 113 + ], + "spans": [ + { + "bbox": [ + 55, + 90, + 297, + 113 + ], + "type": "text", + "content": "This research was supported by NSF CISE 2212457 and NSF CAREER Award CCF 2239787" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 56, + 122, + 115, + 134 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 122, + 115, + 134 + ], + "spans": [ + { + "bbox": [ + 56, + 122, + 115, + 134 + ], + "type": "text", + "content": "References" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 57, + 142, + 297, + 712 + ], + "type": "list", + "angle": 0, + "index": 15, + "blocks": [ + { + "bbox": [ + 61, + 142, + 297, + 186 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 61, + 142, + 297, + 186 + ], + "spans": [ + { + "bbox": [ + 61, + 142, + 297, + 186 + ], + "type": "text", + "content": "[1] Naveed Akhtar, Ajmal Mian, Navid Kardan, and Mubarak Shah. Advances in adversarial attacks and defenses in computer vision: A survey. IEEE Access, 9:155161-155196, 2021. 1" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 61, + 186, + 296, + 240 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 61, + 186, + 296, + 240 + ], + "spans": [ + { + "bbox": [ + 61, + 186, + 296, + 240 + ], + "type": "text", + "content": "[2] Pranjal Awasthi, Anqi Mao, Mehryar Mohri, and Yutao Zhong. Theoretically grounded loss functions and algorithms for adversarial robustness. In Proceedings of The 26th International Conference on Artificial Intelligence and Statistics, pages 10077-10094. PMLR, 2023. 3" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 62, + 241, + 296, + 274 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 241, + 296, + 274 + ], + "spans": [ + { + "bbox": [ + 62, + 241, + 296, + 274 + ], + "type": "text", + "content": "[3] Yatong Bai, Brendon G. Anderson, Aerin Kim, and Somayeh Sojoudi. Improving the accuracy-robustness trade-off of classifiers via adaptive smoothing, 2024. 1" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 62, + 274, + 296, + 328 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 274, + 296, + 328 + ], + "spans": [ + { + "bbox": [ + 62, + 274, + 296, + 328 + ], + "type": "text", + "content": "[4] Brian R. Bartoldson, James Diffenderfer, Konstantinos Parasyris, and Bhavya Kailkhura. Adversarial robustness limits via scaling-law and human-alignment studies. In Proceedings of the 41st International Conference on Machine Learning, pages 3046-3072. PMLR, 2024. 1, 2" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 62, + 329, + 296, + 393 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 329, + 296, + 393 + ], + "spans": [ + { + "bbox": [ + 62, + 329, + 296, + 393 + ], + "type": "text", + "content": "[5] Yulong Cao, Chaowei Xiao, Benjamin Cyr, Yimeng Zhou, Wonseok Park, Sara Rampazzi, Qi Alfred Chen, Kevin Fu, and Z. Morley Mao. Adversarial sensor attack on lidar-based perception in autonomous driving. Proceedings of the 2019 ACM SIGSAC Conference on Computer and Communications Security, 2019. 1" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 62, + 395, + 296, + 437 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 395, + 296, + 437 + ], + "spans": [ + { + "bbox": [ + 62, + 395, + 296, + 437 + ], + "type": "text", + "content": "[6] Nicholas Carlini and David A. Wagner. Adversarial examples are not easily detected: Bypassing ten detection methods. Proceedings of the 10th ACM Workshop on Artificial Intelligence and Security, 2017. 1" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 62, + 439, + 296, + 482 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 439, + 296, + 482 + ], + "spans": [ + { + "bbox": [ + 62, + 439, + 296, + 482 + ], + "type": "text", + "content": "[7] Francesco Croce and Matthias Hein. Reliable evaluation of adversarial robustness with an ensemble of diverse parameter-free attacks. In Proceedings of the 37th International Conference on Machine Learning. JMLR.org, 2020. 1, 2, 8" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 62, + 483, + 296, + 547 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 483, + 296, + 547 + ], + "spans": [ + { + "bbox": [ + 62, + 483, + 296, + 547 + ], + "type": "text", + "content": "[8] Francesco Croce, Maksym Andriushchenko, Vikash Sehwag, Edoardo Debenedetti, Nicolas Flammarion, Mung Chiang, Prateek Mittal, and Matthias Hein. Robustbench: a standardized adversarial robustness benchmark. In Thirty-fifth Conference on Neural Information Processing Systems Datasets and Benchmarks Track, 2021. 1, 2" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 62, + 548, + 296, + 581 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 548, + 296, + 581 + ], + "spans": [ + { + "bbox": [ + 62, + 548, + 296, + 581 + ], + "type": "text", + "content": "[9] Jiequan Cui, Zhuotao Tian, Zhisheng Zhong, Xiaojuan Qi, Bei Yu, and Hanwang Zhang. Decoupled kullback-leibler divergence loss, 2023. 1" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 57, + 582, + 295, + 603 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 582, + 295, + 603 + ], + "spans": [ + { + "bbox": [ + 57, + 582, + 295, + 603 + ], + "type": "text", + "content": "[10] Edoardo Debenedetti, Vikash Sehwag, and Prateek Mittal. A light recipe to train robust vision transformers, 2023. 8" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 57, + 604, + 296, + 658 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 604, + 296, + 658 + ], + "spans": [ + { + "bbox": [ + 57, + 604, + 296, + 658 + ], + "type": "text", + "content": "[11] Edoardo Debenedetti, Zishen Wan, Maksym Andriushchenko, Vikash Sehwag, Kshitij Bhardwaj, and Bhavya Kailkhura. Scaling compute is not all you need for adversarial robustness. In ICLR 2024 Workshop on Reliable and Responsible Foundation Models, 2024. 1, 2" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 57, + 658, + 296, + 712 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 658, + 296, + 712 + ], + "spans": [ + { + "bbox": [ + 57, + 658, + 296, + 712 + ], + "type": "text", + "content": "[12] Samuel Dodge and Lina Karam. A study and comparison of human and deep learning recognition performance under visual distortions. In 2017 26th international conference on computer communication and networks (ICCCN), pages 1-7. IEEE, 2017. 3" + } + ] + } + ], + "index": 14 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 316, + 73, + 555, + 713 + ], + "type": "list", + "angle": 0, + "index": 32, + "blocks": [ + { + "bbox": [ + 316, + 73, + 555, + 95 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 73, + 555, + 95 + ], + "spans": [ + { + "bbox": [ + 316, + 73, + 555, + 95 + ], + "type": "text", + "content": "[13] Elvis Dohmatob. Generalized no free lunch theorem for adversarial robustness. In ICML, 2019. 4" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 317, + 96, + 555, + 128 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 96, + 555, + 128 + ], + "spans": [ + { + "bbox": [ + 317, + 96, + 555, + 128 + ], + "type": "text", + "content": "[14] Ecenaz Erdemir, Jeffrey Bickford, Luca Melis, and Sergul Aydore. Adversarial robustness with non-uniform perturbations, 2021. 4" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 316, + 130, + 555, + 184 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 130, + 555, + 184 + ], + "spans": [ + { + "bbox": [ + 316, + 130, + 555, + 184 + ], + "type": "text", + "content": "[15] Stephanie Fu, Netanel Tamir, Shobhita Sundaram, Lucy Chai, Richard Zhang, Tali Dekel, and Phillip Isola. Dreamsim: Learning new dimensions of human visual similarity using synthetic data. Advances in Neural Information Processing Systems, 36, 2024. 2, 3, 5, 7" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 317, + 186, + 555, + 239 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 186, + 555, + 239 + ], + "spans": [ + { + "bbox": [ + 317, + 186, + 555, + 239 + ], + "type": "text", + "content": "[16] Robert Geirhos, Carlos R. M. Temme, Jonas Rauber, Heiko H. Schütt, Matthias Bethge, and Felix A. Wichmann. Generalisation in humans and deep neural networks. In Advances in Neural Information Processing Systems. Curran Associates, Inc., 2018. 3" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 316, + 241, + 555, + 285 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 241, + 555, + 285 + ], + "spans": [ + { + "bbox": [ + 316, + 241, + 555, + 285 + ], + "type": "text", + "content": "[17] Sara Ghazanfari, Alexandre Araujo, Prashanth Krishnamurthy, Farshad Khorrami, and Siddharth Garg. Lipsim: A provably robust perceptual similarity metric. ArXiv, abs/2310.18274, 2023. 2, 7" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 316, + 286, + 555, + 341 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 286, + 555, + 341 + ], + "spans": [ + { + "bbox": [ + 316, + 286, + 555, + 341 + ], + "type": "text", + "content": "[18] Sara Ghazanfari, Siddharth Garg, Prashanth Krishnamurthy, Farshad Khorrami, and Alexandre Araujo. R-LPIPS: An adversarially robust perceptual similarity metric. In The Second Workshop on New Frontiers in Adversarial Machine Learning, 2023. 7" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 316, + 342, + 555, + 364 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 342, + 555, + 364 + ], + "spans": [ + { + "bbox": [ + 316, + 342, + 555, + 364 + ], + "type": "text", + "content": "[19] Abhijay Ghildyal and Feng Liu. Attacking perceptual similarity metrics. Trans. Mach. Learn. Res., 2023, 2023. 2" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 316, + 365, + 554, + 398 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 365, + 554, + 398 + ], + "spans": [ + { + "bbox": [ + 316, + 365, + 554, + 398 + ], + "type": "text", + "content": "[20] Justin Gilmer, Ryan P. Adams, Ian Goodfellow, David Andersen, and George E. Dahl. Motivating the rules of the game for adversarial example research, 2018. 1" + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 316, + 399, + 555, + 431 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 399, + 555, + 431 + ], + "spans": [ + { + "bbox": [ + 316, + 399, + 555, + 431 + ], + "type": "text", + "content": "[21] Teofilo F. Gonzalez. Clustering to minimize the maximum intercluster distance. Theoretical Computer Science, 38:293-306, 1985. 11" + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 316, + 433, + 555, + 465 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 433, + 555, + 465 + ], + "spans": [ + { + "bbox": [ + 316, + 433, + 555, + 465 + ], + "type": "text", + "content": "[22] Ian J. Goodfellow, Jonathon Shlens, and Christian Szegedy. Explaining and harnessing adversarial examples. CoRR, abs/1412.6572, 2015. 1" + } + ] + } + ], + "index": 25 + }, + { + "bbox": [ + 317, + 467, + 555, + 500 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 467, + 555, + 500 + ], + "spans": [ + { + "bbox": [ + 317, + 467, + 555, + 500 + ], + "type": "text", + "content": "[23] Sven Gowal, Chongli Qin, Jonathan Uesato, Timothy Mann, and Pushmeet Kohli. Uncovering the limits of adversarial training against norm-bounded adversarial examples, 2021. 1" + } + ] + } + ], + "index": 26 + }, + { + "bbox": [ + 316, + 501, + 554, + 533 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 501, + 554, + 533 + ], + "spans": [ + { + "bbox": [ + 316, + 501, + 554, + 533 + ], + "type": "text", + "content": "[24] Sven Gowal, Sylvestre-Alvise Rebuffi, Olivia Wiles, Florian Stimberg, Dan Andrei Calian, and Timothy Mann. Improving robustness using generated data, 2021. 1" + } + ] + } + ], + "index": 27 + }, + { + "bbox": [ + 317, + 535, + 555, + 578 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 535, + 555, + 578 + ], + "spans": [ + { + "bbox": [ + 317, + 535, + 555, + 578 + ], + "type": "text", + "content": "[25] Dan Hendrycks and Thomas Dietterich. Benchmarking neural network robustness to common corruptions and perturbations. Proceedings of the International Conference on Learning Representations, 2019. 3, 6, 8, 16" + } + ] + } + ], + "index": 28 + }, + { + "bbox": [ + 316, + 579, + 555, + 623 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 579, + 555, + 623 + ], + "spans": [ + { + "bbox": [ + 316, + 579, + 555, + 623 + ], + "type": "text", + "content": "[26] Joern-Henrik Jacobsen, Jens Behrmann, Richard Zemel, and Matthias Bethge. Excessive invariance causes adversarial vulnerability. In International Conference on Learning Representations, 2019. 2" + } + ] + } + ], + "index": 29 + }, + { + "bbox": [ + 316, + 624, + 555, + 690 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 624, + 555, + 690 + ], + "spans": [ + { + "bbox": [ + 316, + 624, + 555, + 690 + ], + "type": "text", + "content": "[27] Can Kanbak, Seyed-Mohsen Moosavi-Dezfooli, and Pascal Frossard. Geometric Robustness of Deep Networks: Analysis and Improvement. In 2018 IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 4441-4449, Los Alamitos, CA, USA, 2018. IEEE Computer Society. 3" + } + ] + } + ], + "index": 30 + }, + { + "bbox": [ + 316, + 691, + 554, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 691, + 554, + 713 + ], + "spans": [ + { + "bbox": [ + 316, + 691, + 554, + 713 + ], + "type": "text", + "content": "[28] Oğuzhan Fatih Kar, Teresa Yeo, Andrei Atanov, and Amir Zamir. 3d common corruptions and data augmentation. In" + } + ] + } + ], + "index": 31 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 750, + 315, + 758 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 750, + 315, + 758 + ], + "spans": [ + { + "bbox": [ + 295, + 750, + 315, + 758 + ], + "type": "text", + "content": "9962" + } + ] + } + ], + "index": 33 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 8 + }, + { + "para_blocks": [ + { + "bbox": [ + 56, + 73, + 297, + 713 + ], + "type": "list", + "angle": 0, + "index": 14, + "blocks": [ + { + "bbox": [ + 77, + 73, + 296, + 95 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 77, + 73, + 296, + 95 + ], + "spans": [ + { + "bbox": [ + 77, + 73, + 296, + 95 + ], + "type": "text", + "content": "Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 18963-18974, 2022. 3" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 57, + 95, + 297, + 128 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 95, + 297, + 128 + ], + "spans": [ + { + "bbox": [ + 57, + 95, + 297, + 128 + ], + "type": "text", + "content": "[29] Markus Kettunen, Erik Härkönen, and Jaakko Lehtinen. ELPips: Robust perceptual image similarity via random transformation ensembles, 2019. 2" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 56, + 129, + 296, + 173 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 129, + 296, + 173 + ], + "spans": [ + { + "bbox": [ + 56, + 129, + 296, + 173 + ], + "type": "text", + "content": "[30] Alexander Kirillov, Eric Mintun, Nikhila Ravi, Hanzi Mao, Chloe Rolland, Laura Gustafson, Tete Xiao, Spencer Whitehead, Alexander C. Berg, Wan-Yen Lo, Piotr Dólar, and Ross Girshick. Segment anything. arXiv:2304.02643, 2023. 7" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 56, + 173, + 296, + 205 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 173, + 296, + 205 + ], + "spans": [ + { + "bbox": [ + 56, + 173, + 296, + 205 + ], + "type": "text", + "content": "[31] Alexey Kurakin, Ian J. Goodfellow, and Samy Bengio. Adversarial examples in the physical world. ArXiv, abs/1607.02533, 2017. 1" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 56, + 205, + 296, + 239 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 205, + 296, + 239 + ], + "spans": [ + { + "bbox": [ + 56, + 205, + 296, + 239 + ], + "type": "text", + "content": "[32] Cassidy Laidlaw, Sahil Singla, and Soheil Feizi. Perceptual adversarial robustness: Defense against unseen threat models, 2021. 2, 5, 7" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 56, + 239, + 296, + 293 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 239, + 296, + 293 + ], + "spans": [ + { + "bbox": [ + 56, + 239, + 296, + 293 + ], + "type": "text", + "content": "[33] Chang Liu, Yinpeng Dong, Wenzhao Xiang, Xiao Yang, Hang Su, Jun Zhu, Yuefeng Chen, Yuan He, Hui Xue, and Shibao Zheng. A comprehensive study on robustness of image classification models: Benchmarking and rethinking, 2023. 1, 8, 13" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 56, + 294, + 296, + 338 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 294, + 296, + 338 + ], + "spans": [ + { + "bbox": [ + 56, + 294, + 296, + 338 + ], + "type": "text", + "content": "[34] Aleksander Madry, Aleksandar Makelov, Ludwig Schmidt, Dimitris Tsipras, and Adrian Vladu. Towards deep learning models resistant to adversarial attacks. ArXiv, abs/1706.06083, 2018. 1" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 56, + 338, + 296, + 382 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 338, + 296, + 382 + ], + "spans": [ + { + "bbox": [ + 56, + 338, + 296, + 382 + ], + "type": "text", + "content": "[35] Rafat K. Mantiuk, Kil Joong Kim, Allan G. Rempel, and Wolfgang Heidrich. Hdr-vdp-2: a calibrated visual metric for visibility and quality predictions in all luminance conditions. ACM SIGGRAPH 2011 papers, 2011. 3" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 56, + 383, + 296, + 437 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 383, + 296, + 437 + ], + "spans": [ + { + "bbox": [ + 56, + 383, + 296, + 437 + ], + "type": "text", + "content": "[36] Eric Mintun, Alexander Kirillov, and Saining Xie. On interaction between augmentations and corruptions in natural corruption robustness. In Advances in Neural Information Processing Systems, pages 3571-3583. Curran Associates, Inc., 2021. 3, 6, 8, 16" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 56, + 437, + 296, + 491 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 437, + 296, + 491 + ], + "spans": [ + { + "bbox": [ + 56, + 437, + 296, + 491 + ], + "type": "text", + "content": "[37] Yao Qin, Nicholas Carlini, Garrison Cottrell, Ian Goodfellow, and Colin Raffel. Imperceptible, robust, and targeted adversarial examples for automatic speech recognition. In International conference on machine learning, pages 5231-5240. PMLR, 2019. 1" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 56, + 492, + 296, + 569 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 492, + 296, + 569 + ], + "spans": [ + { + "bbox": [ + 56, + 492, + 296, + 569 + ], + "type": "text", + "content": "[38] Nikhila Ravi, Valentin Gabeur, Yuan-Ting Hu, Ronghang Hu, Chaitanya Ryali, Tengyu Ma, Haitham Khedr, Roman Rädle, Chloe Rolland, Laura Gustafson, Eric Mintun, Junting Pan, Kalyan Vasudev Alwala, Nicolas Carion, Chao-Yuan Wu, Ross Girshick, Piotr Dólár, and Christoph Feichtenhofer. Sam 2: Segment anything in images and videos. arXiv preprint arXiv:2408.00714, 2024. 7" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 56, + 570, + 296, + 624 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 570, + 296, + 624 + ], + "spans": [ + { + "bbox": [ + 56, + 570, + 296, + 624 + ], + "type": "text", + "content": "[39] Ayon Sen, Xiaojin Zhu, Erin Marshall, and Robert Nowak. Popular imperceptibility measures in visual adversarial attacks are far from human perception. In Decision and Game Theory for Security, pages 188-199, Cham, 2020. Springer International Publishing. 2" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 56, + 625, + 296, + 679 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 625, + 296, + 679 + ], + "spans": [ + { + "bbox": [ + 56, + 625, + 296, + 679 + ], + "type": "text", + "content": "[40] Mahmood Sharif, Lujo Bauer, and Michael K. Reiter. On the suitability of lp-norms for creating and preventing adversarial examples. 2018 IEEE/CVF Conference on Computer Vision and Pattern Recognition Workshops (CVPRW), pages 1686-16868, 2018. 2" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 56, + 680, + 296, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 680, + 296, + 713 + ], + "spans": [ + { + "bbox": [ + 56, + 680, + 296, + 713 + ], + "type": "text", + "content": "[41] Naman D Singh, Francesco Croce, and Matthias Hein. Revisiting adversarial training forImagenet: Architectures, training and generalization across threat models, 2023. 8" + } + ] + } + ], + "index": 13 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 316, + 73, + 555, + 711 + ], + "type": "list", + "angle": 0, + "index": 29, + "blocks": [ + { + "bbox": [ + 316, + 73, + 555, + 116 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 73, + 555, + 116 + ], + "spans": [ + { + "bbox": [ + 316, + 73, + 555, + 116 + ], + "type": "text", + "content": "[42] Christian Szegedy, Wojciech Zaremba, Ilya Sutskever, Joan Bruna, D. Erhan, Ian J. Goodfellow, and Rob Fergus. Intriguing properties of neural networks. CoRR, abs/1312.6199, 2014. 1" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 316, + 118, + 555, + 183 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 118, + 555, + 183 + ], + "spans": [ + { + "bbox": [ + 316, + 118, + 555, + 183 + ], + "type": "text", + "content": "[43] Florian Tramér, Jens Behrmann, Nicholas Carlini, Nicolas Papernot, and Jörn-Henrik Jacobsen. Fundamental tradeoffs between invariance and sensitivity to adversarial perturbations. In Proceedings of the 37th International Conference on Machine Learning, ICML 2020, 13-18 July 2020, Virtual Event, pages 9561-9571. PMLR, 2020. 2, 3" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 316, + 184, + 555, + 217 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 184, + 555, + 217 + ], + "spans": [ + { + "bbox": [ + 316, + 184, + 555, + 217 + ], + "type": "text", + "content": "[44] Florian Tramér, Nicholas Carlini, Wieland Brendel, and Aleksander Madry. On adaptive attacks to adversarial example defenses. ArXiv, abs/2002.08347, 2020. 1" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 316, + 219, + 555, + 274 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 219, + 555, + 274 + ], + "spans": [ + { + "bbox": [ + 316, + 219, + 555, + 274 + ], + "type": "text", + "content": "[45] James Tu, Huichen Li, Xinchen Yan, Mengye Ren, Yun Chen, Ming Liang, Eilyan Bitar, Ersin Yumer, and Raquel Urtasun. Exploring adversarial robustness of multi-sensor perception systems in self driving. In Proceedings of the 5th Conference on Robot Learning, pages 1013-1024. PMLR, 2022. 1" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 316, + 274, + 554, + 308 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 274, + 554, + 308 + ], + "spans": [ + { + "bbox": [ + 316, + 274, + 554, + 308 + ], + "type": "text", + "content": "[46] Apostol Vassilev, Alina Oprea, Alie Fordyce, and Hyrum Andersen. Adversarial machine learning: A taxonomy and terminology of attacks and mitigations, 2024. 1" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 316, + 308, + 555, + 351 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 308, + 555, + 351 + ], + "spans": [ + { + "bbox": [ + 316, + 308, + 555, + 351 + ], + "type": "text", + "content": "[47] Zhou Wang, A.C. Bovik, H.R. Sheikh, and E.P. Simoncelli. Image quality assessment: from error visibility to structural similarity. IEEE Transactions on Image Processing, 13(4): 600-612, 2004. 3" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 316, + 353, + 554, + 386 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 353, + 554, + 386 + ], + "spans": [ + { + "bbox": [ + 316, + 353, + 554, + 386 + ], + "type": "text", + "content": "[48] Zekai Wang, Tianyu Pang, Chao Du, Min Lin, Weiwei Liu, and Shuicheng Yan. Better diffusion models further improve adversarial training, 2023. 1" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 316, + 387, + 554, + 420 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 387, + 554, + 420 + ], + "spans": [ + { + "bbox": [ + 316, + 387, + 554, + 420 + ], + "type": "text", + "content": "[49] Zuxuan Wu, Ser-Nam Lim, Larry Davis, and Tom Goldstein. Making an invisibility cloak: Real world adversarial attacks on object detectors, 2020. 1" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 316, + 420, + 555, + 464 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 420, + 555, + 464 + ], + "spans": [ + { + "bbox": [ + 316, + 420, + 555, + 464 + ], + "type": "text", + "content": "[50] Chaowei Xiao, Jun-Yan Zhu, Bo Li, Warren He, Mingyan Liu, and Dawn Song. Spatially transformed adversarial examples. In International Conference on Learning Representations, 2018. 3" + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 316, + 465, + 555, + 509 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 465, + 555, + 509 + ], + "spans": [ + { + "bbox": [ + 316, + 465, + 555, + 509 + ], + "type": "text", + "content": "[51] Kai Yuanqing Xiao, Logan Engstrom, Andrew Ilyas, and Aleksander Madry. Noise or signal: The role of image backgrounds in object recognition. In International Conference on Learning Representations, 2021. 7, 8" + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 316, + 510, + 555, + 565 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 510, + 555, + 565 + ], + "spans": [ + { + "bbox": [ + 316, + 510, + 555, + 565 + ], + "type": "text", + "content": "[52] Hongyang Zhang, Yaodong Yu, Jiantao Jiao, Eric Xing, Laurent El Ghaoui, and Michael Jordan. Theoretically principled trade-off between robustness and accuracy. In Proceedings of the 36th International Conference on Machine Learning, pages 7472-7482. PMLR, 2019. 1" + } + ] + } + ], + "index": 25 + }, + { + "bbox": [ + 316, + 566, + 555, + 609 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 566, + 555, + 609 + ], + "spans": [ + { + "bbox": [ + 316, + 566, + 555, + 609 + ], + "type": "text", + "content": "[53] Lin Zhang, Lei Zhang, Xuanqin Mou, and David Zhang. Fsim: A feature similarity index for image quality assessment. IEEE Transactions on Image Processing, 20(8):2378-2386, 2011. 3" + } + ] + } + ], + "index": 26 + }, + { + "bbox": [ + 316, + 610, + 555, + 677 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 610, + 555, + 677 + ], + "spans": [ + { + "bbox": [ + 316, + 610, + 555, + 677 + ], + "type": "text", + "content": "[54] Richard Zhang, Phillip Isola, Alexei A. Efros, Eli Shechtman, and Oliver Wang. The unreasonable effectiveness of deep features as a perceptual metric. In 2018 IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 586-595, Los Alamitos, CA, USA, 2018. IEEE Computer Society. 3, 7" + } + ] + } + ], + "index": 27 + }, + { + "bbox": [ + 316, + 677, + 554, + 711 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 677, + 554, + 711 + ], + "spans": [ + { + "bbox": [ + 316, + 677, + 554, + 711 + ], + "type": "text", + "content": "[55] Piotr Želasko, Sonal Joshi, Yiwen Shao, Jesus Villalba, Jan Trmal, Najim Dehak, and Sanjeev Khudanpur. Adversarial attacks and defenses for speech recognition systems, 2021." + } + ] + } + ], + "index": 28 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 750, + 315, + 758 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 750, + 315, + 758 + ], + "spans": [ + { + "bbox": [ + 295, + 750, + 315, + 758 + ], + "type": "text", + "content": "9963" + } + ] + } + ], + "index": 30 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 9 + } + ], + "_backend": "vlm", + "_version_name": "2.6.4" +} \ No newline at end of file diff --git a/2025/DiskVPS_ Vanishing Point Detector via Hough Transform in a Disk Region/80e52809-79f5-4a63-9312-c7df737fcd10_content_list.json b/2025/DiskVPS_ Vanishing Point Detector via Hough Transform in a Disk Region/80e52809-79f5-4a63-9312-c7df737fcd10_content_list.json new file mode 100644 index 0000000000000000000000000000000000000000..fe7b93a2367643977456beca1a4fb6616011616e --- /dev/null +++ b/2025/DiskVPS_ Vanishing Point Detector via Hough Transform in a Disk Region/80e52809-79f5-4a63-9312-c7df737fcd10_content_list.json @@ -0,0 +1,1828 @@ +[ + { + "type": "text", + "text": "DiskVPS: Vanishing Point Detector via Hough Transform in a Disk Region", + "text_level": 1, + "bbox": [ + 109, + 128, + 861, + 152 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Jianping Wu \nSuzhou Vocational University \nSuzhou, China 215104 \nwujp@jssvc.edu.cn", + "bbox": [ + 364, + 181, + 604, + 252 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Abstract", + "text_level": 1, + "bbox": [ + 233, + 286, + 313, + 301 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "DiskVPS is a novel and robust vanishing point (VP) detection scheme based on Hough Transform (HT) over an image-plane-mapped disk region. In the DiskVPS, the image plane is first mapped into a disk region, which is then partitioned into concentrical ring-shaped subspaces. Each subspace is further partitioned into approximately equal probability Hough cells. By using individual edges rather than edge pairs as voters, DiskVPS can achieve high accuracy with extreme efficiency. Involving no calibration parameters, DiskVPS is particularly suitable for VP detection in uncalibrated images and thus can be applied in image calibration. A comparative experimental study demonstrates that the basic DiskVPS model without parameter optimization achieved significantly better performance over the SOTA in detection accuracy and processing speed with real-world images. The study also shows that DiskVPS is robust against parameter changes.", + "bbox": [ + 75, + 316, + 472, + 575 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "1. Introduction", + "text_level": 1, + "bbox": [ + 76, + 602, + 209, + 618 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "VP detection has a variety of applications, including camera calibration ([1, 3, 5, 15]), 3D reconstruction ([25, 40]), virtual reality [30], robot navigation [12], scene tracking [6] and vehicle automatic driving [13]. Depending on the direction of the parallel lines and the setup of the capturing camera, a VP can appear anywhere, including infinity, in the image plane. A triplet of orthogonal VPs is formed by 3 sets of mutually orthogonal parallel lines in the man-made world commonly called Manhattan world [5]. According to [3], the information of any 2 orthogonal VPs in an image is sufficient to obtain the focal length of the capturing camera. Moreover, a triplet of orthogonal VPs fully determines the camera's intrinsic and extrinsic parameters.", + "bbox": [ + 75, + 627, + 468, + 824 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "1.1. Related Works", + "text_level": 1, + "bbox": [ + 76, + 832, + 228, + 847 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "VP detection has long been a key research topic attracting great attention in computer vision. However, it remains a challenging issue far from soundly resolved. The key is-", + "bbox": [ + 75, + 854, + 468, + 902 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "sue in VP detection is the unboundedness of the image plane in which VPs reside. In VP detection, it is essential to map the image plane into some bounded space before practical techniques such as Hough Transform (HT, [9]) can be deployed.", + "bbox": [ + 496, + 287, + 893, + 364 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "The first such attempt came from [29] that divided the image plane into 3 bounded subspaces to perform the HT. Since its division method is rather arbitrary without offsetting the probability biases in cell partitioning, it failed to yield satisfactory results. For calibrated images, Barnard [2] proposed a VP detection algorithm that mapped the image plane into a unit sphere called Gaussian sphere (GS) for HT, which is widely used in VP detection ([18, 24]) and is also universally applied in deep learning VP detection algorithms([4, 10, 11, 16, 23, 27, 32, 39]). However, GS-based approaches have limitations. Since building the GS requires the known parameters of the capturing camera such as the focal length and coordinates of the optical axis in the image plane, GS approaches are unsuitable for uncalibrated images, which account for a substantial fraction of images of interest. GS's voting scheme using edge pairs can significantly amplify discretization and distortion errors in edge detection and thus cause a high likelihood of spurious VP detection. Accordingly, GS models are prone to spurious detection ([16, 24, 39]). Moreover, the process of partitioning the GS into Hough cells itself is complicated and delicate. The non-planar shape of the GS is not amenable for partitioning of the GS into planar Hough cells while the Hough cell sizes on the GS are wildly non-uniform and sensitive to the errors of calibration parameters. With the limitations, the GS baseline model did not draw wide attention until the recent rapid development of deep learning technology.", + "bbox": [ + 496, + 367, + 893, + 790 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "As an important alternative to HT in VP detection, random sample consensus (RANSAC [8]) is another extensively explored scheme in VP detection ([17, 19, 21, 22, 26, 28, 33, 35, 38]). RANSAC can work directly in the image plane and is suitable for VP detection in uncalibrated images; moreover, RANSAC outperforms GS-based VP detection models in practice with scenes where many edges", + "bbox": [ + 496, + 795, + 893, + 902 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "CVF", + "bbox": [ + 106, + 2, + 181, + 42 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore.", + "bbox": [ + 236, + 0, + 807, + 46 + ], + "page_idx": 0 + }, + { + "type": "page_number", + "text": "27049", + "bbox": [ + 478, + 944, + 519, + 957 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "belong to VP inliers. The main drawback of RANSAC is its susceptibility to false negative detection when the size of samples accounts for only a tiny portion of the full set of edge pairs used for VP detection (3 or more edges belonging to a set of 2 or 3 orthogonal VPs are sometimes proposed in [19]). Additionally, RANSAC is computationally expensive, particularly for images of cluttered scenes. Recently, [34] proposed an algorithm that integrated RANSAC and HT to achieve real-time VP detection. There are other methods like using dual spaces ([14]).", + "bbox": [ + 75, + 90, + 472, + 241 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Despite its limitations, GS is the only proven VP detection model available to transform an unbounded space (image plane) into a bounded space (GS), enabling its deployment in deep learning neural networks. Therefore, the GS remains the chosen technique in deep-learning-based VP detection algorithms([4, 10, 11, 16, 23, 27, 32, 37, 39]). While the baseline model of GS-based VP detection is not optimal in both accuracy and processing speed, GS-based deep learning approaches making use of advanced high-level-feature extraction schemes and high-performance hardware have achieved the performance comparable to the SOTA, a level far superior to the GS's baseline model. However, deep learning approaches have limitations. The costs of data annotation and training needed by deep learning are expensive. Additionally, limited datasets usually do not generalize well in deep learning, which can fail in unexpected ways.", + "bbox": [ + 75, + 243, + 473, + 501 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "1.2. Contribution", + "text_level": 1, + "bbox": [ + 76, + 513, + 215, + 527 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "DiskVPS is suitable for conducting highly reliable and real-time VP detection. We use Cartesian coordinates $(u, v)$ to denote the normalized coordinates with the origin set at the image center and both image width and height normalized to 1 (see next section for how $(u, v)$ is mapped from the pixel-based image plane coordinates). DiskVPS maps an image plane (unbounded space) (Fig. 1(a)) into a disk-shaped space (Fig. 1(b)). Specifically, it maps an unbounded polar coordinate system $(r, \\theta)$ $(r = \\sqrt{u^2 + v^2}$ and $\\theta = \\tan^{-1} \\frac{u}{v}$ into a bounded polar coordinate system $(\\tan^{-1}(4r), \\theta)$ with the latter's radial boundary located at $\\frac{\\pi}{2}$ . With the approach, a straight line that passes the origin in the image plane is mapped into a diametrical line segment in the disk. In contrast, a straight line not passing the origin in the image plane is mapped into an arc-shaped segment with 2 ends overlapped with its origin-passing parallel counterpart at the disk boundary. Consequently, parallel lines in the image plane intersect on the disk boundary, whereas non-parallel lines in the image plane intersect each other only once inside the disk. In the scheme, HT is reduced to a probabilistic method to partition the disk into a set of approximately equal-probability Hough cells. Fig. 2 shows a set of edges detected in a typical image (see Fig. 2(a)) and its corresponding Hough accumulation map generated by the", + "bbox": [ + 75, + 537, + 472, + 902 + ], + "page_idx": 1 + }, + { + "type": "image", + "img_path": "images/dfed0fd5dc4e2f4dad963cf38f9f921999e201678cb0408e3617ca24ea33a895.jpg", + "image_caption": [ + "(a)" + ], + "image_footnote": [], + "bbox": [ + 524, + 95, + 686, + 223 + ], + "page_idx": 1 + }, + { + "type": "image", + "img_path": "images/9bea78ddb7cea411f96e05c6197cfbc4716ef2f130ad127ece26ab9fb3a73d7f.jpg", + "image_caption": [ + "(b)" + ], + "image_footnote": [], + "bbox": [ + 696, + 95, + 859, + 224 + ], + "page_idx": 1 + }, + { + "type": "image", + "img_path": "images/05fc8c78f9eadcfd2cd1bb6057c271d2b84b5f299744ac83a3fd2e798c3407c0.jpg", + "image_caption": [ + "Figure 1. Mapping from an image plane (Cartesian coordinates $(u, v)$ (a)) to a disk-shaped Hough space (polar coordinates $(\\tan^{-1}(4r), \\theta)$ (b)) with $r = \\sqrt{u^2 + v^2}$ and $\\theta = \\tan^{-1} \\frac{v}{u}$ .", + "(a)" + ], + "image_footnote": [], + "bbox": [ + 537, + 316, + 681, + 398 + ], + "page_idx": 1 + }, + { + "type": "image", + "img_path": "images/7f28b618a869c4a5ed9ad512be42ecabacff4fd00b7155b7b6c6d82921be608d.jpg", + "image_caption": [ + "(b)" + ], + "image_footnote": [], + "bbox": [ + 705, + 323, + 859, + 402 + ], + "page_idx": 1 + }, + { + "type": "image", + "img_path": "images/330a8000a1b133d2c18f962ebb4fa0c4c928ca2160f1a1765f0887765e85887a.jpg", + "image_caption": [ + "Figure 2. Detected edges of a typical image (a) and its corresponding Hough accumulation map (b).", + "Figure 3. Intersection $P$ of the extension line of a random edge and a given line defined by the distance $d$ and the angle $\\phi$ ." + ], + "image_footnote": [], + "bbox": [ + 617, + 488, + 774, + 597 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "edges using DiskVPS (see Fig. 2(b)). Five sharp and salient peaks represent 3 mutually orthogonal VPs, 2 far-field VPs generated by nearly parallel edges and one near-field VP generated by non-parallel edges. The proposed approach successfully manifests all valid VPs in the image without yielding any spurious VPs in Fig. 2b.", + "bbox": [ + 496, + 675, + 893, + 767 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "2. Algorithm Description", + "text_level": 1, + "bbox": [ + 500, + 781, + 714, + 799 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "2.1. Theoretical Foundation", + "text_level": 1, + "bbox": [ + 500, + 806, + 718, + 821 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Assume $(x,y)$ is the pixel-based coordinates of a point in the image plane and the image of interest is encoded by 2 coordinates in pixels: $(x_{\\min},y_{\\min})$ and $(x_{\\max},y_{\\max})$ , which refer to the pixel-based coordinates of the image's bottom left corner and top right corner respectively. To ease analy-", + "bbox": [ + 496, + 829, + 893, + 902 + ], + "page_idx": 1 + }, + { + "type": "page_number", + "text": "27050", + "bbox": [ + 478, + 944, + 519, + 957 + ], + "page_idx": 1 + }, + { + "type": "image", + "img_path": "images/3d0d52e89e5c6ced96176238e73654e014726575ed37294b3f10cf1e49cf97ea.jpg", + "image_caption": [ + "Figure 4. Illustration of the 2 special directions along which the PDF is considered, the radial and tangential directions." + ], + "image_footnote": [], + "bbox": [ + 174, + 90, + 370, + 231 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "sis, $(x,y)$ is transformed to a normalized coordinate system $(u,v)$ . Their relationship is written as:", + "bbox": [ + 76, + 300, + 468, + 330 + ], + "page_idx": 2 + }, + { + "type": "equation", + "text": "\n$$\n\\left\\{ \\begin{array}{l} u = \\frac {x - 0 . 5 \\left(x _ {\\min } + x _ {\\max }\\right)}{x _ {\\max } - x _ {\\min }} \\\\ v = \\frac {y - 0 . 5 \\left(y _ {\\min } + y _ {\\max }\\right)}{y _ {\\max } - y _ {\\min }} \\end{array} \\right. \\tag {1}\n$$\n", + "text_format": "latex", + "bbox": [ + 194, + 344, + 468, + 383 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "In the $(u,v)$ system, the origin $O$ is fixed to the image center, and the bottom left corner and top right corner of the image are moved to $(-0.5, -0.5)$ and $(0.5, 0.5)$ , respectively (see Fig. 3). The image region turns into an Origin-centered unit square in the $(u,v)$ system.", + "bbox": [ + 76, + 398, + 468, + 474 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "[34] proposed a statistical model to estimate the probabilistic distribution of $P$ , the intersection of a given line and the extension line of a random edge, along the very given line. According to [34], the midpoint of the random edge is distributed uniformly in the unit square, and the direction of the random edge is also uniformly distributed over $[0, \\pi]$ . Such a random edge can be used to simulate the statistical distribution of the edges extracted from a sufficiently large number of images. We cite the conclusion of [34] in this paper. As shown in Fig. 3, if the given line is defined by $d$ (the distance from $O$ on the given line) and $\\phi$ (the directional angle pointing from $O$ to its perpendicular foot to the given line), then [34] gives:", + "bbox": [ + 75, + 476, + 472, + 657 + ], + "page_idx": 2 + }, + { + "type": "equation", + "text": "\n$$\nf _ {d, \\phi} (t) = \\frac {1}{\\pi} \\int_ {- 0. 5 - 0. 5} ^ {0. 5} \\int_ {0} ^ {0. 5} \\frac {\\left| u \\cos \\phi + v \\sin \\phi - d \\right| d u d v}{(u - d \\cos \\phi - t \\sin \\phi) ^ {2} + (v - d \\sin \\phi + t \\cos \\phi) ^ {2}} \\tag {2}\n$$\n", + "text_format": "latex", + "bbox": [ + 76, + 671, + 495, + 728 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "where $t$ is the coordinate of $P$ on a new axis created along the given line with the origin of the axis set at the perpendicular foot of $O$ on the given line, as shown in Fig. 3.", + "bbox": [ + 76, + 729, + 468, + 773 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Here, we are interested in Eq.2's solutions for 2 special cases. As shown in Fig.4, one is for the case in which the given line is along the radial direction in the $(u,v)$ coordinates, and the other is for the case in which the given line is in the azimuthal or tangential direction. $f_{\\mathrm{rad}}(r,\\theta)$ and $f_{\\mathrm{tag}}(r,\\theta)$ are used to denote the PDFs of $P$ along the given line along the radial direction and the given line along the tangential direction respectively, where $\\theta$ is the directional angle of $OP$ and $r$ is the length of $OP$ , both in the $(u,v)$", + "bbox": [ + 76, + 775, + 468, + 901 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "coordinate system. Accordingly, we have:", + "bbox": [ + 498, + 90, + 779, + 107 + ], + "page_idx": 2 + }, + { + "type": "equation", + "text": "\n$$\n\\begin{array}{l} f _ {\\mathrm {r a d}} (r, \\theta) = f _ {0, \\theta + \\frac {\\pi}{2}} (r) \\\\ = \\frac {1}{\\pi} \\int_ {- 0. 5 - 0. 5} ^ {0. 5} \\int_ {0} ^ {0. 5} \\frac {\\left| - u \\sin \\theta + v \\cos \\theta \\right| d u d v}{(u - r \\cos \\theta) ^ {2} + (v - r \\sin \\theta) ^ {2}} \\tag {3} \\\\ \\end{array}\n$$\n", + "text_format": "latex", + "bbox": [ + 532, + 114, + 890, + 179 + ], + "page_idx": 2 + }, + { + "type": "equation", + "text": "\n$$\n\\begin{array}{l} f _ {\\mathrm {t a g}} (r, \\theta) = f _ {r, \\theta + \\frac {\\pi}{2}} (0) \\\\ = \\frac {1}{\\pi} \\int_ {- 0. 5} ^ {0. 5} \\int_ {- 0. 5} ^ {0. 5} \\frac {\\left| u \\sin \\theta - v \\cos \\theta + r \\right| d u d v}{(u + r \\sin \\theta) ^ {2} + (v - r \\cos \\theta) ^ {2}} \\tag {4} \\\\ \\end{array}\n$$\n", + "text_format": "latex", + "bbox": [ + 517, + 191, + 890, + 255 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Since $f_{\\mathrm{rad}}(r,\\theta)$ and $f_{\\mathrm{tag}}(r,\\theta)$ 's dependence on the variable $\\theta$ is insignificant (see Fig.5), they are conveniently approximated by 2 less complex and more manageable forms $g_{\\mathrm{rad}}(r) = f_{\\mathrm{rad}}(r,0)$ and $g_{\\mathrm{tag}}(r) = f_{\\mathrm{tag}}(r,0)$ to facilitate subsequent construction of Hough space. They are:", + "bbox": [ + 498, + 260, + 890, + 330 + ], + "page_idx": 2 + }, + { + "type": "equation", + "text": "\n$$\ng _ {\\text {r a d}} (r) = \\frac {1}{\\pi} \\int_ {- 0. 5 - 0. 5} ^ {0. 5} \\int_ {0} ^ {0. 5} \\frac {| v | d u d v}{(u - r) ^ {2} + v ^ {2}} \\tag {5}\n$$\n", + "text_format": "latex", + "bbox": [ + 586, + 340, + 890, + 385 + ], + "page_idx": 2 + }, + { + "type": "equation", + "text": "\n$$\ng _ {\\text {a g}} (r) = \\frac {1}{\\pi} \\int_ {- 0. 5 - 0. 5} ^ {0. 5} \\int_ {- 0. 5} ^ {0. 5} \\frac {| v - r | d u d v}{u ^ {2} + (v - r) ^ {2}} \\tag {6}\n$$\n", + "text_format": "latex", + "bbox": [ + 584, + 397, + 890, + 441 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "The analytical solutions of $g_{\\mathrm{rad}}(r)$ and $g_{\\mathrm{tag}}(r)$ are:", + "bbox": [ + 519, + 446, + 849, + 462 + ], + "page_idx": 2 + }, + { + "type": "equation", + "text": "\n$$\ng _ {\\mathrm {r a d}} (r) = \\left\\{ \\begin{array}{l l} \\frac {\\tan^ {- 1} (2 r + 1) - \\tan^ {- 1} (2 r - 1)}{\\pi} & \\\\ + \\frac {1}{2 \\pi} \\ln \\frac {[ (2 r - 1) ^ {2} + 1 ] [ (2 r + 1) ^ {2} + 1 ]}{(2 r - 1) ^ {2} (2 r + 1) ^ {2}} & r \\neq 0. 5 \\\\ + \\frac {r}{\\pi} \\ln \\frac {(2 r - 1) ^ {2} [ (2 r + 1) ^ {2} + 1 ]}{(2 r + 1) ^ {2} [ (2 r - 1) ^ {2} + 1 ]} & \\\\ \\frac {\\tan^ {- 1} 2 + \\ln 5 - \\ln 4}{\\pi} & r = 0. 5 \\end{array} \\right. \\tag {7}\n$$\n", + "text_format": "latex", + "bbox": [ + 500, + 470, + 890, + 607 + ], + "page_idx": 2 + }, + { + "type": "equation", + "text": "\n$$\ng _ {\\text {t a g}} (r) = \\left\\{ \\begin{array}{l l} 1 - \\frac {\\tan^ {- 1} (2 r + 1) - \\tan^ {- 1} (2 r - 1)}{\\pi} & r \\leq 0. 5 \\\\ - \\frac {2 r \\left[ \\tan^ {- 1} (2 r + 1) + \\tan^ {- 1} (2 r - 1) \\right]}{\\pi} & \\\\ + \\frac {\\ln [ 1 + (2 r + 1) ^ {2} ] + \\ln [ 1 + (2 r - 1) ^ {2} ]}{2 \\pi} & \\\\ 1 - \\frac {(2 r + 1) \\tan^ {- 1} (2 r + 1)}{\\pi} & \\\\ - \\frac {(2 r - 1) \\tan^ {- 1} (2 r - 1)}{\\pi} & r > 0. 5 \\\\ + \\frac {\\ln [ 1 + (2 r + 1) ^ {2} ] - \\ln [ 1 + (2 r - 1) ^ {2} ]}{2 \\pi} & \\end{array} \\right. \\tag {8}\n$$\n", + "text_format": "latex", + "bbox": [ + 500, + 627, + 890, + 820 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "As shown in Fig. 5, both $\\frac{g_{\\mathrm{rad}}(r)}{f_{\\mathrm{rad}}(r,\\theta)}$ and $\\frac{g_{\\mathrm{tag}}(r)}{f_{\\mathrm{tag}}(r,\\theta)}$ lie between 0.88 and 1.13 for $r \\leq 1.7$ and they move closer to 1 with the increase of $r$ for $r \\geq 1.0$ . Overall, using $g_{\\mathrm{rad}}(r)$ to approximate $f_{\\mathrm{rad}}(r,\\theta)$ and using $g_{\\mathrm{tag}}(r)$ to approximate $f_{\\mathrm{tag}}(r,\\theta)$ do not lead to significant errors.", + "bbox": [ + 498, + 821, + 890, + 900 + ], + "page_idx": 2 + }, + { + "type": "page_number", + "text": "27051", + "bbox": [ + 478, + 944, + 517, + 955 + ], + "page_idx": 2 + }, + { + "type": "image", + "img_path": "images/b0c19e335660e103c499d6f7dc6cdb1fff7dc3dbe868829abc0d2f41abb8b4ed.jpg", + "image_caption": [ + "(a)" + ], + "image_footnote": [], + "bbox": [ + 80, + 94, + 261, + 202 + ], + "page_idx": 3 + }, + { + "type": "image", + "img_path": "images/19fd581a7f960b722199ec8be233b5a09acf49e9900fdf9d641e6cb28e294cea.jpg", + "image_caption": [ + "(b)" + ], + "image_footnote": [], + "bbox": [ + 272, + 93, + 452, + 202 + ], + "page_idx": 3 + }, + { + "type": "image", + "img_path": "images/9df6ead6121bd57f5c9992989bbf3706cea7f665e818bc692db34bc50fc351e2.jpg", + "image_caption": [ + "Figure 5. Proximity of $g_{\\mathrm{rad}}(r)$ to $f_{\\mathrm{rad}}(r,\\theta)$ (a) and proximity of $g_{\\mathrm{tag}}(r)$ to $f_{\\mathrm{tag}}(r,\\theta)$ (b).", + "Figure 6. $\\frac{1}{\\pi} \\frac{4}{(4r)^2 + 1}$ 's proximity to $f_{\\mathrm{rad}}(r, \\theta)$ for chosen $r$ values." + ], + "image_footnote": [], + "bbox": [ + 176, + 280, + 356, + 388 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "The previous analysis provides a theoretical basis for building a Hough space from the $(r,\\theta)$ . However, since the radial dimension $(r)$ remains unbounded, it is not feasible to create a Hough space directly from its current form (Eq.7 and 8).", + "bbox": [ + 75, + 449, + 468, + 523 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "So it makes sense to transform $r$ to a bounded space and derive the PDF in the new space. Note that when $r \\gg 1$ , Eq.5 leads to $g_{\\mathrm{rad}}(r) \\approx \\frac{1}{\\pi r^2} \\int_{-0.5}^{0.5} \\int_{-0.5}^{0.5} |v| du dv = \\frac{1}{4\\pi r^2}$ . Since $\\frac{1}{4\\pi r^2} \\approx \\frac{1}{\\pi} \\frac{4}{(4r)^2 + 1}$ for $r \\gg 1$ , Eq. 5 can be further approximated as:", + "bbox": [ + 76, + 525, + 468, + 617 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\ng _ {\\text {r a d}} (r) \\approx \\frac {1}{\\pi} \\frac {4}{(4 r) ^ {2} + 1} \\text {f o r} r \\gg 1 \\tag {9}\n$$\n", + "text_format": "latex", + "bbox": [ + 169, + 617, + 468, + 646 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Here $g_{\\mathrm{rad}}(r)\\approx \\frac{1}{\\pi}\\frac{4}{(4r)^2 + 1}$ is chosen over other alternatives because the integration of $\\frac{1}{\\pi}\\frac{4}{(4r)^2 + 1}$ over a full radial line is equal to 1, which matches the probability of a radial line intersecting the extension line of the random edge. Note $\\frac{1}{\\pi}\\int \\limits_{r = -\\infty}^{r = +\\infty}\\frac{4dr}{(4r)^2 + 1} = \\frac{1}{\\pi}\\int \\limits_{r = -\\infty}^{r = +\\infty}d\\tan^{-1}(4r) = \\frac{1}{\\pi}\\times 2\\times \\frac{\\pi}{2} = 1.$ As shown in Fig.6, when $r\\geq 1.5$ Eq.9 is already a good approximation of $f_{\\mathrm{rad}}(r,\\theta)$ as $\\frac{\\frac{1}{\\pi}\\frac{4}{(4r)^2 + 1}}{f_{\\mathrm{rad}}(r,\\theta)}$ lies consistently between 0.93 and 1.08 for $r\\geq 1.5$", + "bbox": [ + 76, + 651, + 468, + 800 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Eq.9 creates a route to transform $r$ to a bounded space. To achieve the goal, we introduce a new function $\\psi(r)$ :", + "bbox": [ + 76, + 801, + 467, + 830 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\psi (r) = \\tan^ {- 1} (4 r) \\tag {10}\n$$\n", + "text_format": "latex", + "bbox": [ + 212, + 842, + 468, + 859 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Unlike $r$ , $\\psi$ is defined in a bounded space since $\\psi \\in [0, \\frac{\\pi}{2}]$ . We denote the PDF along $\\psi$ as $h_{\\mathrm{rad}}(\\psi)$ , which is", + "bbox": [ + 76, + 871, + 468, + 902 + ], + "page_idx": 3 + }, + { + "type": "image", + "img_path": "images/cf5df42cbcdb12186be01f3c61c73856cc3c5fb4d22b9f1dfb9d5f6d16bcaef5.jpg", + "image_caption": [ + "(a)" + ], + "image_footnote": [], + "bbox": [ + 531, + 93, + 686, + 226 + ], + "page_idx": 3 + }, + { + "type": "image", + "img_path": "images/17141d096b5248c572f005a65b8098d255e036018a31aae4b1fb0e4dd5a66a13.jpg", + "image_caption": [ + "(b)", + "Figure 7. Depiction of DiskVPS's Hough space. (a) The structure of the Hough space; (b) The parameters of a Hough cell in the $i$ th subspace." + ], + "image_footnote": [], + "bbox": [ + 694, + 89, + 872, + 227 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "related to $g_{\\mathrm{rad}}(r)$ through the following formula:", + "bbox": [ + 500, + 324, + 818, + 340 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\nh _ {\\mathrm {r a d}} (\\psi) = g _ {\\mathrm {r a d}} (r) \\left| \\frac {d r}{d \\psi} \\right| \\tag {11}\n$$\n", + "text_format": "latex", + "bbox": [ + 625, + 349, + 890, + 376 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Combining Eq.9, Eq.10 and Eq.11, and noting $\\frac{d(\\tan^{-1}x)}{dx} = \\frac{1}{x^2 + 1}$ , we have:", + "bbox": [ + 500, + 387, + 890, + 425 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\nh _ {\\mathrm {r a d}} (\\psi) \\approx \\frac {1}{\\pi} \\frac {4}{(4 r) ^ {2} + 1} \\div \\frac {4}{(4 r) ^ {2} + 1} = \\frac {1}{\\pi} \\tag {12}\n$$\n", + "text_format": "latex", + "bbox": [ + 563, + 433, + 890, + 462 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Eq.12 indicates that the PDF along $\\psi$ is approximately constant $(= \\frac{1}{\\pi})$ in the space of $(\\psi, \\theta)$ (for $r \\geq 1.5$ ).", + "bbox": [ + 498, + 470, + 890, + 502 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "2.2. Construction of Hough Space", + "text_level": 1, + "bbox": [ + 500, + 510, + 761, + 527 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Structurally, the Hough space is partitioned into a set of $M$ concentric ring-shaped subspaces, each of which is further partitioned into a set of at most $N$ fan-shaped cells, as shown in Fig.7(a). The principle is to ensure that each subspace has a roughly equal probability of intersecting the extension line of the random edge along the radial direction and each cell has a roughly equal probability of intersecting the line in the tangential direction.", + "bbox": [ + 496, + 534, + 890, + 652 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Our first goal is to obtain each subspace's inner radius and radial span using the principle above. We denote the inner radius and the radial span of the $i$ th subspace as $r_i$ and $\\delta r_i$ respectively, where $i = 1,2,\\ldots,M$ is the order number of the subspace counting from the innermost outwards. Noting the outer radius of the $i$ th subspace is the inner radius of the $i + 1$ th subspace, there exist the following relations throughout the subspaces:", + "bbox": [ + 496, + 652, + 890, + 765 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\begin{array}{l} r _ {1} = 0 \\\\ r _ {2} = r _ {1} + \\delta r _ {1} \\\\ \\dots \\tag {13} \\\\ r _ {i + 1} = r _ {i} + \\delta r _ {i} \\\\ \\dots \\\\ \\end{array}\n$$\n", + "text_format": "latex", + "bbox": [ + 647, + 777, + 890, + 858 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Since a full radial line intersects the extension line of the random edge exactly once (the special case where the 2", + "bbox": [ + 500, + 869, + 890, + 900 + ], + "page_idx": 3 + }, + { + "type": "page_number", + "text": "27052", + "bbox": [ + 478, + 944, + 517, + 955 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "lines are parallel and do not intersect has infinitesimal probability and is therefore ignored), the probability of the radial span of each subspace intersecting the extension line of the random edge is $\\frac{1}{M}$ .", + "bbox": [ + 75, + 90, + 468, + 151 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "In addition to setting the inner radius and the radial span of the $i$ th subspace as $r_i$ and $\\delta r_i$ for near-field subspaces, we also set the inner radius and the radial span of the subspace as $\\psi_i = \\tan^{-1}(4r_i)$ and $\\delta \\psi_i$ for far-field subspaces. Assuming that the PDF along the radial span of the $i$ th space is constant ( $= g_{\\mathrm{rad}(r_i)}$ for near-field subspace and $= h_{\\mathrm{rad}}(\\psi_i)$ for far-field subspace), then the probability $\\frac{1}{M}$ must match with $g_{\\mathrm{rad}}(r_i)\\delta r_i$ if the $i$ th subspace is in the near field or $h_{\\mathrm{rad}}(\\psi_i)\\delta \\psi_i$ if the $i$ th subspace is in the far field. Since $h_{\\mathrm{rad}}(\\psi) \\approx \\frac{1}{\\pi}$ (see Eq.12), we have:", + "bbox": [ + 75, + 151, + 473, + 297 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\delta r _ {i} = \\frac {1}{M g _ {\\mathrm {r a d}} \\left(r _ {i}\\right)} \\quad \\text {f o r} r _ {i} \\leq 1. 5 \\tag {14}\n$$\n", + "text_format": "latex", + "bbox": [ + 169, + 306, + 468, + 339 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\delta \\psi_ {i} = \\frac {\\pi}{M} \\quad \\text {f o r} r _ {i} > 1. 5\n$$\n", + "text_format": "latex", + "bbox": [ + 173, + 337, + 372, + 361 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Combining $\\delta \\psi_{i} = \\psi_{i + 1} - \\psi_{i} = \\tan^{-1}(4r_{i + 1}) - \\tan^{-1}(4r_{i})$ and $\\delta r_i = r_{i + 1} - r_i$ , the 2nd term of Eq.14 turns into $\\delta r_{i} = \\frac{\\tan(\\frac{\\pi}{M} + \\tan^{-1}4r_{i})}{4} -r_{i}$ . And Eq.14 can be rewritten as:", + "bbox": [ + 76, + 375, + 468, + 436 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\delta r _ {i} = \\left\\{ \\begin{array}{l l} \\frac {1}{M g _ {\\text {r a d}} \\left(r _ {i}\\right)} & \\text {f o r} r _ {i} \\leq 1. 5 \\\\ \\frac {\\tan \\left(\\frac {\\pi}{M} + \\tan^ {- 1} 4 r _ {i}\\right)}{4} - r _ {i} & \\text {f o r} r _ {i} > 1. 5 \\end{array} \\right. \\tag {15}\n$$\n", + "text_format": "latex", + "bbox": [ + 127, + 446, + 468, + 503 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "The Hough space partitioning starts from $i = 1$ . Eq.13 and Eq.15 work together to obtain the inner radius $r_i$ and the radial span $\\delta r_i$ of the $i$ th subspace iteratively ( $i = i + 1$ after each iteration) until the $i$ th subspace comes to the Hough space boundary that is indicated by $\\psi_{i+1} \\geq \\frac{\\pi}{2}$ or $\\tan^{-1}(4r_i) + \\frac{\\pi}{M} \\geq \\frac{\\pi}{2}$ . The final count of the subspaces is denoted as $M^* = i$ at the end of the last iteration. Since the PDF approximations are used in the Hough space partitioning, $M^*$ is generally not equal, but very close, to $M$ .", + "bbox": [ + 75, + 516, + 468, + 651 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Afterward, each subspace is partitioned into multiple equal-angle-span cells. For the $i$ th subspace, the angle span of a cell is denoted as $\\delta \\theta_{i}$ ( $i = 1,2,\\ldots M^{*}$ ), as shown in Fig.7(b). We observe that the inner circle of the outermost (the $M^{*}$ th) subspace intersects the extension line of the random edge exactly twice. In the meantime, the very circle is partitioned into $N$ equal-angle-span arcs, each of which forms the inner arc of a cell. Consequently, the probability of the inner arc of a cell intersecting the random edge's extension line is $\\frac{2}{N}$ . To make sure that each cell in any subspace gets the same probability $\\frac{2}{N}$ of intersecting the random edge's extension line, this probability must match the probability obtained through the product of the cell's inner arc length ( $= \\delta \\theta_{i}r_{i}$ ) and the PDF along the arc ( $= g_{\\mathrm{tag}}(r_i)$ ) throughout the entire Hough space. Accordingly, we have:", + "bbox": [ + 75, + 652, + 470, + 864 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\delta \\theta_ {i} = \\frac {2}{N r _ {i} g _ {\\text {t a g}} \\left(r _ {i}\\right)} \\tag {16}\n$$\n", + "text_format": "latex", + "bbox": [ + 212, + 875, + 468, + 905 + ], + "page_idx": 4 + }, + { + "type": "image", + "img_path": "images/7bb5ef4a48a765b106c5788f8f1a7a43d46679b255cc6bc5031e7b72b91be9a3.jpg", + "image_caption": [ + "Figure 8. A typical scenario of a voting edge casting its votes in the $i$ th subspace with 1 or 0 inside each cell referring to the number of votes each cell receives." + ], + "image_footnote": [], + "bbox": [ + 527, + 94, + 864, + 277 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "The number of cells the $i$ th subspace is partitioned into is denoted as $n_i (i = 1, 2, \\ldots, M^*)$ . It is written as:", + "bbox": [ + 498, + 361, + 890, + 390 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\nn _ {i} = \\operatorname {r o u n d} \\left(\\frac {2 \\pi}{\\delta \\theta_ {i}}\\right) \\tag {17}\n$$\n", + "text_format": "latex", + "bbox": [ + 638, + 398, + 890, + 426 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "where $2\\pi$ is the angle span of the $i$ th subspace, and $\\delta \\theta_{i}$ is the angle span of a cell in the $i$ th subspace, and round refers to the namesake maths function.", + "bbox": [ + 498, + 433, + 890, + 477 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Finally, $\\delta \\theta_{i}$ needs to be slightly modified due to $n_i$ 's discrete nature. Its adjusted form, $\\delta \\theta_{i}^{*}$ , is written as:", + "bbox": [ + 498, + 478, + 890, + 507 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\delta \\theta_ {i} ^ {*} = \\frac {2 \\pi}{n _ {i}} \\tag {18}\n$$\n", + "text_format": "latex", + "bbox": [ + 661, + 515, + 890, + 542 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "This way, the $i$ th subspace is partitioned into $n_i$ equal-angle-span cells, each covering an angle span of $\\delta \\theta_i^*$ . Via iterating $i$ from 1 through $M^*$ , all subspaces are thus partitioned into approximately equally probabilistic cells in terms of intersecting the extension line of the random edge.", + "bbox": [ + 498, + 550, + 890, + 627 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "2.3. Implementation of Hough Transform", + "text_level": 1, + "bbox": [ + 500, + 635, + 821, + 652 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "A Hough accumulation map, each of whose elements is mapped from a corresponding Hough cell, is constructed and the magnitude of each element is initialized to zero.", + "bbox": [ + 498, + 659, + 890, + 703 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "The VP detection consists of a voting process and an iterative VP detection process.", + "bbox": [ + 498, + 704, + 890, + 733 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "In the voting process, each edge casts its votes based on whether its extension line intersects any of the subspaces, and if yes then which specific cell(s) of the subspace the line intersects, with each intersecting cell receiving exactly 1 vote. Assume the extension line of the voting edge is defined by $d$ , the distance from $O$ to the line, and $\\varphi$ , the azimuthal angle of $O$ 's perpendicular foot on the line. For an edge's voting in the $i$ th subspace, there exist 3 cases, namely $d < r_i$ , $r_i \\leq d < r_i + \\delta r_i$ , and $d \\geq r_i + \\delta r_i$ respectively, where $r_i$ and $\\delta r_i$ are the inner radius and radial span of the $i$ th subspace.", + "bbox": [ + 496, + 734, + 890, + 901 + ], + "page_idx": 4 + }, + { + "type": "page_number", + "text": "27053", + "bbox": [ + 478, + 944, + 517, + 955 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "A scenario of the 1st case ( $d < r_i$ ) is shown in Fig.8. The extension line of the voting edge intersects the $i$ th subspace boundary at 4 points, namely $P_1, P_2, P_3$ , and $P_4$ . Assume $\\Psi_{\\overrightarrow{OP_1}}$ , $\\Psi_{\\overrightarrow{OP_2}}$ , $\\Psi_{\\overrightarrow{OP_3}}$ , and $\\Psi_{\\overrightarrow{OP_4}}$ are the azimuthal angles of vectors $\\overrightarrow{OP_1}$ , $\\overrightarrow{OP_2}$ , $\\overrightarrow{OP_3}$ and $\\overrightarrow{OP_4}$ , respectively. Then we have:", + "bbox": [ + 76, + 90, + 472, + 179 + ], + "page_idx": 5 + }, + { + "type": "equation", + "text": "\n$$\n\\left\\{ \\begin{array}{l} \\Psi_ {\\overrightarrow {O P _ {1}}} = \\varphi + \\cos^ {- 1} \\frac {d}{r _ {i} + \\delta r _ {i}} \\\\ \\Psi_ {\\overrightarrow {O P _ {2}}} = \\varphi + \\cos^ {- 1} \\frac {d}{r _ {i}} \\\\ \\Psi_ {\\overrightarrow {O P _ {3}}} = \\varphi - \\cos^ {- 1} \\frac {d}{r _ {i}} \\\\ \\Psi_ {\\overrightarrow {O P _ {4}}} = \\varphi - \\cos^ {- 1} \\frac {d}{r _ {i} + \\delta r _ {i}} \\end{array} \\right. \\tag {19}\n$$\n", + "text_format": "latex", + "bbox": [ + 176, + 181, + 468, + 252 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "As shown in Fig.8, $P_{1}P_{2}$ and $P_{3}P_{4}$ are the line's 2 sections intersecting the $i$ th subspace. Assume the $n_{i}$ cells of the $i$ th subspace are labeled 1 through $n_{i}$ and a set of consecutive cells that intersect $P_{1}P_{2}$ have the labels numbered $j$ through $k$ , then $j$ and $k$ are written as:", + "bbox": [ + 76, + 256, + 472, + 328 + ], + "page_idx": 5 + }, + { + "type": "equation", + "text": "\n$$\n\\left\\{\\begin{array}{l}j = \\left\\lceil \\frac {\\Psi_ {\\overrightarrow {O P _ {2}}}}{\\delta \\theta_ {i} ^ {*}} \\right\\rceil\\\\k = \\left\\lceil \\frac {\\Psi_ {\\overrightarrow {O P _ {1}}}}{\\delta \\theta_ {i} ^ {*}} \\right\\rceil\\end{array}\\right. \\tag {20}\n$$\n", + "text_format": "latex", + "bbox": [ + 217, + 337, + 468, + 385 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "where $\\delta \\theta_{i}^{*}$ is defined in Eq.18, and $\\lceil \\rceil$ refers to maths function ceil.", + "bbox": [ + 76, + 395, + 468, + 424 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Similarly, we can obtain the set of consecutive cells intersecting $P_{3}P_{4}$ . In voting, each cell in the two sets receives 1 vote from the voting edge while the rest of the $n_{i}$ cells receive no votes.", + "bbox": [ + 76, + 426, + 468, + 484 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "For the 2nd case, voting is performed similarly to the first, but only the angle span between $\\Psi_{\\overrightarrow{OP_1}}$ and $\\Psi_{\\overrightarrow{OP_4}}$ needs to be handled as neither $P_{2}$ nor $P_{3}$ exists.", + "bbox": [ + 76, + 487, + 468, + 532 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "For the 3rd case, the subspace does not intersect the line; thus, no cell receives any vote.", + "bbox": [ + 76, + 532, + 468, + 563 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "The voting process generally yields a Hough accumulation map with a small number of salient peaks scattered over the map similar to Fig.2(b).", + "bbox": [ + 76, + 564, + 468, + 609 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Then, VP detection is as simple as finding salient peaks in the Hough accumulation map. In practice, at most $N_{\\mathrm{VP}}$ dominant VPs are detected via at most $N_{\\mathrm{VP}}$ iterations with each iteration extracting exactly 1 VP. In each iteration, the following steps are taken:", + "bbox": [ + 76, + 609, + 468, + 685 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Step 1: Search the global peak in the current Hough accumulation space.", + "bbox": [ + 76, + 686, + 468, + 715 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Step 2: Obtain the location of the peak and set it as the initial position of the current VP.", + "bbox": [ + 76, + 717, + 468, + 746 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Step 3: Search all of the inlier edges of the current VP. An edge is regarded as an inlier of the VP only if the angle between its extension line and the line that passes both the midpoint of the edge and the VP is no more than $\\Theta$ .", + "bbox": [ + 76, + 748, + 468, + 808 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Step 4: Update the Hough accumulation map by removing all of the votes cast by the inliers of the current VP. This ensures that inliers of the dominant VPs will not affect the subsequent detection of less dominant VPs.", + "bbox": [ + 76, + 809, + 468, + 868 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Step 5: If the peak value in the updated accumulation map is at least 3 times as much as the average and the total", + "bbox": [ + 76, + 869, + 468, + 900 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/54e99d8984c49e7841feb5ac62f320c5f60e03c8ef77bb0a0b7a3b566ee561b7.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 524, + 93, + 620, + 167 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/0d4cfc94c3f34d16d19216fbbd291954c62918572ea9cdad60cbba5d06706227.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 647, + 93, + 743, + 167 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/69e503b3f0d39b524e2cc06fc27f50b365571dd4ecc2d7494726e0e3745f08fe.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 769, + 90, + 870, + 169 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/f5e6e581506cd4cd1c9f95e693bf9cdbceec2df7c1ddb26fe8ab5051e7428679.jpg", + "image_caption": [ + "(a)", + "(f)", + "Figure 9. Graphical demonstration of the steps in DiskVPS' VP detection: (a) is the original image; (b) is the detected edge set; (c) is the initial Hough accumulation colormap with its global peak (also the position of the 1st VP) marked in red; (d) is the updated Hough accumulation map after removing the votes cast by the inliers of the 1st VP with 1 detected VP and the updated peak marked in red; (e) is the updated Hough accumulation map after the 2nd iteration with 2 detected VPs and the updated peak marked in red; (f) is the updated Hough accumulation map after the 3rd iteration with 3 detected VPs marked in red." + ], + "image_footnote": [], + "bbox": [ + 522, + 194, + 622, + 273 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/95cda287321aac94349ba9684fdd23d9d505f5fabb1d6cb7701238de7b921853.jpg", + "image_caption": [ + "(b)", + "(e)" + ], + "image_footnote": [], + "bbox": [ + 645, + 194, + 746, + 273 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/22938e7e0ac996026b6d9b6a8fb5c6f96270ca47eaf70c5ed3930b8648ec7f24.jpg", + "image_caption": [ + "(c)", + "(d)" + ], + "image_footnote": [], + "bbox": [ + 769, + 195, + 870, + 273 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "number of detected VPs is less than $N_{\\mathrm{VP}}$ , then go to Step 1. Otherwise, stop iteration and output all detected VPs.", + "bbox": [ + 498, + 474, + 888, + 505 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "The VP detection process is demonstrated in Fig.9. The zenith, which is the convergent point of nearly vertical edges, is located at the local maximum around the south pole of the disk-shaped colormap and the inliers of the zenith also lead to the significantly enhanced distribution around the north pole; the $1^{st}$ horizontal VP at the east pole is formed by nearly parallel edges which also contribute to the enhanced accumulation around the west pole; non-parallel edges form the $2^{nd}$ horizontal VP and generate a corresponding peak in the interior of the map. Some local maxima are noises caused by combined voting of inliers of different VPs and outliers of the VPs. However, after 3 iterations almost all local maxima unrelated to any valid VPs in the original accumulation map disappear (see Fig. 9(f)). It demonstrates that the vote removal process of the inliers of the detected VPs in Step 4 effectively minimizes false positive detection.", + "bbox": [ + 496, + 506, + 890, + 763 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "3. Experiment", + "text_level": 1, + "bbox": [ + 500, + 781, + 624, + 799 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "The proposed approach is examined and evaluated against the state-of-the-art (SOTA) in VP detection. It was implemented in $\\mathrm{C + + }$ and tested on a PC powered by an Intel i7 CPU (3.5 GHZ).", + "bbox": [ + 498, + 808, + 890, + 868 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "DiskVPS was evaluated via 3 benchmark datasets, the York Urban Dataset (YUD [7]), SVVP([27]), and", + "bbox": [ + 498, + 869, + 890, + 900 + ], + "page_idx": 5 + }, + { + "type": "page_number", + "text": "27054", + "bbox": [ + 478, + 944, + 519, + 955 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/be74e2064883f1b56355d02d5b697024eff9531bc97f4d51040ada2f1512dbf1.jpg", + "image_caption": [ + "Figure 10. Example images from YUD [7] (the 1st row), SVVP[27] (the 3rd row), and SU3 [40] (the 5th row) datasets with each even row immediately below displaying their generated disk-shaped Hough colormaps and the VP detection results. The ground truth VPs are marked with white \"+\" and the detected VPs are marked with red \"×\"." + ], + "image_footnote": [], + "bbox": [ + 91, + 87, + 176, + 441 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/0356ac0da5725a5538177d36637d26ff7f5fd4115e2bb9bf291d10d95c7af390.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 184, + 88, + 267, + 441 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/d8db0a2de0ab0692331d601d0eb65bb959c8985fdbe37521994cdc660101da9e.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 277, + 88, + 361, + 441 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/02588881492730d419bc21fa1d800f353bf60745ac3a642168258aa836c4e5be.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 372, + 88, + 454, + 441 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "SU3([40]). YUD contains 102 calibrated real-world images of $640 \\times 480$ pixels. SVVP contains 500 calibrated real-world images of $512 \\times 512$ pixels, and SU3 contains a photorealistic dataset containing 23,000 calibrated synthetic outdoor images of $512 \\times 512$ pixels. The commonness of the 3 benchmark datasets is that they all satisfy the Manhattanworld assumption [5] as there exists a triplet of orthogonal VPs in each image. So our experiment is focused on how accurately DiskVPS performs VP detections on the triplets of orthogonal VPs in each dataset compared to the SOTA. Please note that while the images used in the experiment were calibrated, DiskVPS treated those images as uncalibrated and performed VP detection without using any calibration info until checking accuracy against the ground truth.", + "bbox": [ + 75, + 565, + 467, + 790 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "DiskVPS uses edges as input. In preprocessing, there are many algorithms to perform edge detection (see LSD([31]), [20], and [36], etc.). A similar algorithm to LSD was selected to perform edge detection. Table1 lists the 4 parameters and their respective magnitude that our implementation used for evaluation. $M$ is the number of concentric subspaces that the disk-shaped Hough space is roughly parti", + "bbox": [ + 75, + 795, + 467, + 902 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "tioned into, $N$ is the maximum number of roughly equal-probability cells that each of the subspaces is partitioned into, $N_{\\mathrm{VP}}$ is the maximum number of dominant VPs that are initially extracted from each image, and $\\Theta$ is the criterion angle to separate the inliers of a VP from its outliers. The parameters are fixed throughout the experiment despite the differences among the 3 datasets.", + "bbox": [ + 496, + 90, + 890, + 196 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Fig.10 lists some example images of the 3 datasets and their VP detection results of our approach. It demonstrates that the disk-shaped Hough accumulation map not only manifests almost all valid VPs but also generates no noticeable spurious VPs. Most detected VPs are either very close to or precisely overlapped with their ground truth counterparts. The only exception happens in the 6th row and the 3rd column, in which the zenith's ground truth lies around the north pole (corresponding to $(0, +\\infty)$ in $(u, v)$ coordinates) while its detected location is around the south pole $(0, -\\infty)$ in $(u, v)$ . The discrepancy is reconciled as the VP at $(0, +\\infty)$ and the VP at $(0, -\\infty)$ in the $(u, v)$ coordinates are formed by the same group of vertical edges.", + "bbox": [ + 496, + 198, + 890, + 393 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "In the experiment, DiskVPS is compared with six SOTA algorithms ([11, 16, 26, 27, 34, 39]) using a protocol adopted by [11, 16, 27]. In the protocol, a triple of orthogonal VPs are detected from each image, then 3 angular differences in the camera space between the 3 detected VPs and their respective ground truth are obtained, with the largest of the 3 differences adopted as the angular difference of the image. According to the protocol, the percentage of a dataset's images that have a smaller angular difference than a given threshold is used as the angular accuracy (AA) criterion. In the experiment, AA@3°, AA@5°, and AA@10° represent AA thresholds of 3°, 5°, and 10° respectively. For example, an algorithm achieving 80% on AA@5° with a dataset means that the algorithm achieves an angular difference of 5° or less on 80% of the images in the dataset.", + "bbox": [ + 496, + 393, + 890, + 619 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "In DiskVPS implementation, a triplet of orthogonal VPs is picked from the initial $N_{\\mathrm{VP}}$ VPs based on their mutual orthogonality (see [3] for details of VP orthogonality). However, there is a small fraction of images in which only 2 of the 3 orthogonal VPs are captured in the initial $N_{\\mathrm{VP}}$ VPs. This happens when more dominant non-orthogonal VPs suppress a less salient orthogonal VP and thus miss the initial detection. In the latter case, DiskVPS detects the 3rd orthogonal VP by searching for the local maximum in", + "bbox": [ + 496, + 621, + 890, + 758 + ], + "page_idx": 6 + }, + { + "type": "table", + "img_path": "images/2d064063661c9fb332937969954b02825faa1eee66e16a786737a89ad27b7dfc.jpg", + "table_caption": [], + "table_footnote": [], + "table_body": "
NameMagnitude
M200
N200
NVP4
Θ3°
", + "bbox": [ + 620, + 781, + 772, + 873 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Table 1. Parameters used in the implementation of our algorithm.", + "bbox": [ + 500, + 883, + 885, + 898 + ], + "page_idx": 6 + }, + { + "type": "page_number", + "text": "27055", + "bbox": [ + 478, + 944, + 517, + 955 + ], + "page_idx": 6 + }, + { + "type": "table", + "img_path": "images/111051d5aa184f1e371a909c7b6b16e0d9993abf2e60b963bec4294936ac2613.jpg", + "table_caption": [], + "table_footnote": [], + "table_body": "
Datasets MetricsFPSYUD [7]SVVP[27]SU3 [40]
AA@3°AA@5°AA@10°AA@3°AA@5°AA@10°AA@3°AA@5°
J-Linkage[26]1.060.871.881.527.343.260.882.087.2
Wu et al. [34]3958.373.482.243.860.175.782.387.6
CONSAC [11]3.061.773.384.443.856.569.486.290.4
NeurVPS [39]0.560.871.881.532.845.760.293.996.3
Lin et al. [16]2360.774.386.351.868.782.284.490.6
Tong et al. [27]2565.577.187.451.667.782.691.394.6
Ours (DiskVPS)15874.590.299.058.381.698.484.391.7
", + "bbox": [ + 106, + 88, + 867, + 241 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "the Hough space around the 3rd VP's predicted location derived from the 2 known orthogonal VPs. In the experiment, it was observed that when $N_{\\mathrm{VP}} = 4$ , DiskVPS successfully captured a triple of orthogonal VPs in the initial $N_{\\mathrm{VP}}$ VPs in 90 (88% of the total) images in YUD, while it captured 2 of the triplet of orthogonal VPs in 11 (11% of the total) images of YUD in the initial $N_{\\mathrm{VP}}$ . After executing the 3rd VP detection on the 11 images, 101 of the 102 images in YUD have their respective triples of orthogonal VPs successfully detected.", + "bbox": [ + 75, + 316, + 472, + 467 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "The comparison results are shown in Table 2. Our approach achieves $74.5\\%$ at AA@3° on YUD, $9\\%$ higher than the second best at $65.5\\%$ achieved by [27], a significant improvement over the SOTA. Similar patterns are observed at AA@5° and AA@10° on YUD and SVVP. Notably, our approach achieves $99.0\\%$ on YUD and $98.4\\%$ on SVVP at AA@10° whereas none of any other approaches reaches $90\\%$ or higher on this respect. Overall, for real-world images, our approach significantly outperforms the SOTA as far as angular accuracy is concerned.", + "bbox": [ + 75, + 470, + 472, + 623 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "On the synthetic SU3, DiskVPS is also competitive. Note that the number of images of SU3 is roughly 40 times as large as YUD and SVD combined and the two approaches ([27, 39]) that outperformed DiskVPS are both deep-learning (DL) methods. We believe the large size of SU3 provides substantially more training and helps significantly optimize the DL algorithms. Considering that DiskVPS is a non-optimized basic VP detection model involving no calibration information which the DL models require, the accuracy of DiskVPS is still impressive.", + "bbox": [ + 75, + 625, + 472, + 776 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Moreover, we increased $M$ from 200 to 300 and $N$ from 100 to 180 (both at stepsize of 5) and observed the angular accuracy (AA) distribution at AA@ $3^{\\circ}$ , AA@ $5^{\\circ}$ , and AA@ $10^{\\circ}$ . The results are shown in Table 3. Since the standard deviations (STDEV) are consistently below $3\\%$ of their respective means, it is evident that DiskVPS is reasonably robust against changing parameters.", + "bbox": [ + 75, + 777, + 472, + 883 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Lastly, we also compare our approach's efficiency with", + "bbox": [ + 96, + 885, + 470, + 901 + ], + "page_idx": 7 + }, + { + "type": "table", + "img_path": "images/233a49d316afeebfef3a06cbf8515f81b794a09dd211451a36dedf7442fb6703.jpg", + "table_caption": [ + "Table 2. Processing speed and angular accuracy (AA) comparison on YUD( [7]), SVVP([27]) and SU3( [40]). Our algorithm achieves by far the best average processing speed on the 3 datasets. Moreover, our approach significantly outperforms all other approaches on real-world YUD and SVVP, and is competitive on synthetic SU3." + ], + "table_footnote": [], + "table_body": "
CriterionMinimumMaximumMeanSTDEV
AA@3°70.678.475.32.2
AA@5°89.294.191.51.3
AA@10°99.099.099.00
", + "bbox": [ + 503, + 314, + 890, + 393 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Table 3. Angular accuracy statistics over $200\\leq M\\leq 300$ and $100\\le N\\le 180$", + "bbox": [ + 500, + 402, + 890, + 430 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "other algorithms. Our approach achieves 158 frames per second (fps) or 6.3 milliseconds per frame excluding image decoding and edge detection. It is around 4 times as fast as the second best at 39 fps achieved by [34]. DiskVPS's high efficiency lies in its voting mechanism. In the voting process, an edge line has 0, 1, or 2 sections crossing a subspace (see, Fig.8) and has around 1 section crossing any subspace on average. The 2 azimuthal angles of the line section are used to derive which series of the consecutively numbered cells of the subspace should receive the votes (see 20). The time complexity of the voting process is $M \\times n$ where $M$ is the total number of subspaces and $n$ is the number of detected edges of a typical image. Assuming $M = 200$ and a typical image contains around 500 edges, the computational cost of the voting process is roughly $2 \\times 200 \\times 500 = 200,000$ math function calls, which can be processed in around 2 milliseconds by the 3.5GHz CPU we used. Unlike a typical DL approach that generally employs millions of hidden parameters and deals with time-consuming multilayer processing, DiskVPS has none and thus can reach much higher efficiency.", + "bbox": [ + 496, + 459, + 890, + 777 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "4. Conclusion", + "text_level": 1, + "bbox": [ + 500, + 797, + 619, + 811 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "By using a more precise voting mechanism than the Gaussian sphere model currently adopted by most of the SOTA, DiskVPS has achieved significantly better performance than the SOTA in detection accuracy and processing speed for real-world images.", + "bbox": [ + 496, + 825, + 890, + 902 + ], + "page_idx": 7 + }, + { + "type": "page_number", + "text": "27056", + "bbox": [ + 478, + 944, + 519, + 957 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "References", + "text_level": 1, + "bbox": [ + 78, + 89, + 173, + 104 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[1] Michel Antunes and Joao P Barreto. A global approach for the detection of vanishing points and mutually orthogonal vanishing directions. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition, pages 1336-1343, 2013. 1", + "[2] Stephen T Barnard. Interpreting perspective images. Artificial intelligence, 21(4):435-462, 1983. 1", + "[3] Bruno Caprile and Vincent Torre. Using vanishing points for camera calibration. International journal of computer vision, 4(2):127-139, 1990. 1, 7", + "[4] Taiyan Chen, Xianghua Ying, Jinfa Yang, Ruibin Wang, Ruohao Guo, Bowei Xing, and Ji Shi. Vpdetr: end-to-end vanishing point detection transformers. In Proceedings of the AAAI Conference on Artificial Intelligence, volume 38, pages 1192-1200, 2024. 1, 2", + "[5] James M Coughlan and Alan L Yuille. Manhattan world: Compass direction from a single image by bayesian inference. In Proceedings of the seventh IEEE international conference on computer vision, volume 2, pages 941-947. IEEE, 1999. 1, 7", + "[6] James M Coughlan and Alan L Yuille. The Manhattan world assumption: Regularities in scene statistics which enable bayesian inference. In NIPS, volume 2, page 3, 2000. 1", + "[7] Patrick Denis, James H Elder, and Francisco J Estrada. Efficient edge-based methods for estimating Manhattan frames in urban imagery. In European conference on computer vision, pages 197-210. Springer, 2008. 6, 7, 8", + "[8] Martin A Fischler and Robert C Bolles. Random sample consensus: a paradigm for model fitting with applications to image analysis and automated cartography. Communications of the ACM, 24(6):381-395, 1981. 1", + "[9] Paul VC Hough. Machine analysis of bubble chamber pictures. In Proc. of the International Conference on High Energy Accelerators and Instrumentation, Sept. 1959, pages 554-556, 1959. 1", + "[10] Florian Kluger, Hanno Ackermann, Michael Ying Yang, and Bodo Rosenhahn. Deep learning for vanishing point detection using an inverse gnomonic projection. In German Conference on Pattern Recognition, pages 17-28. Springer, 2017. 1, 2", + "[11] Florian Kluger, Eric Brachmann, Hanno Ackermann, Carsten Rother, Michael Ying Yang, and Bodo Rosenhahn. Consac: Robust multi-model fitting by conditional sample consensus. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 4634-4643, 2020. 1, 2, 7, 8", + "[12] Till Kroeger, Dengxin Dai, and Luc Van Gool. Joint vanishing point extraction and tracking. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition, pages 2449-2457, 2015. 1", + "[13] Seokju Lee, Junsik Kim, Jae Shin Yoon, Seunghak Shin, Oleksandr Bailo, Namil Kim, Tae-Hee Lee, Hyun Seok Hong, Seung-Hoon Han, and In So Kweon. Vpgnet: Vanishing point guided network for lane and road marking detection and recognition. In Proceedings of the IEEE inter" + ], + "bbox": [ + 78, + 114, + 470, + 900 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "national conference on computer vision, pages 1947-1955, 2017. 1", + "[14] José Lezama, Rafael Grompone von Gioi, Gregory Randall, and Jean-Michel Morel. Finding vanishing points via point alignments in image primal and dual domains. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition, pages 509-515, 2014. 2", + "[15] David Liebowitz and Andrew Zisserman. Combining scene and auto-calibration constraints. In Proceedings of the Seventh IEEE International Conference on Computer Vision, volume 1, pages 293-300. IEEE, 1999. 1", + "[16] Yancong Lin, Ruben Wiersma, Silvia L. Pintea, Klaus Hildebrandt, Elmar Eisemann, and Jan C. van Gemert. Deep vanishing point detection: Geometric priors make dataset variations vanish. In 2022 IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 6093-6103, 2022. 1, 2, 7, 8", + "[17] Xiaohu Lu, Jian Yaoy, Haoang Li, Yahui Liu, and Xiaofeng Zhang. 2-line exhaustive searching for real-time vanishing point estimation in manhattan world. In 2017 IEEE Winter Conference on Applications of Computer Vision (WACV), pages 345-353. IEEE, 2017. 1", + "[18] Evelyne Lutton, Henri Maitre, and Jaime Lopez-Krahe. Contribution to the determination of vanishing points using hough transform. IEEE transactions on pattern analysis and machine intelligence, 16(4):430-438, 1994. 1", + "[19] Faraz M Mirzaei and Stergios I Roumeliotis. Optimal estimation of vanishing points in a Manhattan world. In 2011 International Conference on Computer Vision, pages 2454-2461. IEEE, 2011. 1, 2", + "[20] Van Nhan Nguyen, Robert Jenssen, and Davide Roverso. Lsnet: Fast single-shot line-segment detector. Machine Vision and Applications, 32(1):12, 2021. 7", + "[21] Rémi Pautrat, Shaohui Liu, Petr Hruby, Marc Pollefeys, and Daniel Barath. Vanishing point estimation in uncalibrated images with prior gravity direction. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 14118-14127, 2023. 1", + "[22] Carsten Rother. A new approach to vanishing point detection in architectural environments. Image and Vision Computing, 20(9-10):647-655, 2002. 1", + "[23] Alexander Sheshkus, Anastasia Ingacheva, Vladimir Arlazarov, and Dmitry Nikolaev. Houghnet: neural network architecture for vanishing points detection. In 2019 International Conference on Document Analysis and Recognition (ICDAR), pages 844-849. IEEE, 2019. 1, 2", + "[24] Jefferey A Shufelt. Performance evaluation and analysis of vanishing point detection techniques. IEEE transactions on pattern analysis and machine intelligence, 21(3):282-288, 1999. 1", + "[25] Sudipta Sinha, Drew Steedly, and Rick Szeliski. Piecewise planar stereo for image-based rendering. 2009. 1", + "[26] Jean-Philippe Tardif. Non-iterative approach for fast and accurate vanishing point detection. In 2009 IEEE 12th International Conference on Computer Vision, pages 1250–1257. IEEE, 2009. 1, 7, 8", + "[27] Xin Tong, Xianghua Ying, Yongjie Shi, Ruibin Wang, and Jinfa Yang. Transformer based line segment classifier with" + ], + "bbox": [ + 501, + 92, + 893, + 900 + ], + "page_idx": 8 + }, + { + "type": "page_number", + "text": "27057", + "bbox": [ + 478, + 944, + 517, + 955 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "image context for real-time vanishing point detection in man-. hhattan world. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 6093- 6102, 2022. 1, 2, 6, 7, 8", + "[28] Elena Tretyak, Olga Barinova, Pushmeet Kohli, and Victor Lempitsky. Geometric image parsing in man-made environments. International Journal of Computer Vision, 97(3):305-321, 2012. 1", + "[29] Tinne Tuytelaars, Luc Van Gool, Marc Proesmans, and Theo Moons. The cascaded hough transform as an aid in aerial image interpretation. In Sixth International Conference on Computer Vision (IEEE Cat. No. 98CH36271), pages 67-72. IEEE, 1998. 1", + "[30] Jonathan Ventura and Tobias Höllerer. Structure and motion in urban environments using upright panoramas. *Virtual Reality*, 17(2):147-156, 2013. 1", + "[31] Rafael Grompone Von Gioi, Jeremie Jakubowicz, Jean-Michel Morel, and Gregory Randall. Lsd: A fast line segment detector with a false detection control. IEEE transactions on pattern analysis and machine intelligence, 32(4):722-732, 2008. 7", + "[32] Yukun Wang, Shuo Gu, Yinbo Liu, and Hui Kong. Semi-supervised vanishing point detection with contrastive learning. Pattern Recognition, 153:110518, 2024. 1, 2", + "[33] Horst Wildenauer and Allan Hanbury. Robust camera self-calibration from monocular images of manhattan worlds. In 2012 IEEE Conference on Computer Vision and Pattern Recognition, pages 2831-2838. IEEE, 2012. 1", + "[34] Jianping Wu, Liang Zhang, Ye Liu, and Ke Chen. Real-time vanishing point detector integrating under-parameterized ransac and hough transform. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 3732-3741, 2021. 2, 3, 7, 8", + "[35] Yiliang Xu, Sangmin Oh, and Anthony Hoogs. A minimum error vanishing point detection approach for uncalibrated monocular images of man-made environments. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition, pages 1376-1383, 2013. 1", + "[36] Yifan Xu, Weijian Xu, David Cheung, and Zhuowen Tu. Line segment detection using transformers without edges. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 4257-4266, 2021. 7", + "[37] Menghua Zhai, Scott Workman, and Nathan Jacobs. Detecting vanishing points using global image context in a non-manhattan world. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition, pages 5657-5665, 2016. 2", + "[38] Lilian Zhang, Huimin Lu, Xiaoping Hu, and Reinhard Koch. Vanishing point estimation and line classification in a Manhattan world with a unifying camera model. International Journal of Computer Vision, 117:111-130, 2016. 1", + "[39] Yichao Zhou, Haozhi Qi, Jingwei Huang, and Yi Ma. Neurvps: Neural vanishing point scanning via conic convolution. arXiv preprint arXiv:1910.06316, 2019. 1, 2, 7, 8", + "[40] Yichao Zhou, Haozhi Qi, Yuexiang Zhai, Qi Sun, Zhili Chen, Li-Yi Wei, and Yi Ma. Learning to reconstruct 3d Manhattan wireframes from a single image. In Proceedings of the" + ], + "bbox": [ + 78, + 92, + 468, + 900 + ], + "page_idx": 9 + }, + { + "type": "text", + "text": "IEEE/CVF International Conference on Computer Vision, pages 7698-7707, 2019. 1, 7, 8", + "bbox": [ + 532, + 92, + 890, + 119 + ], + "page_idx": 9 + }, + { + "type": "page_number", + "text": "27058", + "bbox": [ + 478, + 945, + 517, + 955 + ], + "page_idx": 9 + } +] \ No newline at end of file diff --git a/2025/DiskVPS_ Vanishing Point Detector via Hough Transform in a Disk Region/80e52809-79f5-4a63-9312-c7df737fcd10_model.json b/2025/DiskVPS_ Vanishing Point Detector via Hough Transform in a Disk Region/80e52809-79f5-4a63-9312-c7df737fcd10_model.json new file mode 100644 index 0000000000000000000000000000000000000000..2fdc5b63e69df18f7796ec8f0a67d933078f105b --- /dev/null +++ b/2025/DiskVPS_ Vanishing Point Detector via Hough Transform in a Disk Region/80e52809-79f5-4a63-9312-c7df737fcd10_model.json @@ -0,0 +1,2409 @@ +[ + [ + { + "type": "header", + "bbox": [ + 0.107, + 0.003, + 0.182, + 0.043 + ], + "angle": 0, + "content": "CVF" + }, + { + "type": "header", + "bbox": [ + 0.237, + 0.001, + 0.808, + 0.047 + ], + "angle": 0, + "content": "This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore." + }, + { + "type": "title", + "bbox": [ + 0.11, + 0.13, + 0.862, + 0.154 + ], + "angle": 0, + "content": "DiskVPS: Vanishing Point Detector via Hough Transform in a Disk Region" + }, + { + "type": "text", + "bbox": [ + 0.366, + 0.182, + 0.605, + 0.253 + ], + "angle": 0, + "content": "Jianping Wu \nSuzhou Vocational University \nSuzhou, China 215104 \nwujp@jssvc.edu.cn" + }, + { + "type": "title", + "bbox": [ + 0.235, + 0.287, + 0.314, + 0.303 + ], + "angle": 0, + "content": "Abstract" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.318, + 0.473, + 0.576 + ], + "angle": 0, + "content": "DiskVPS is a novel and robust vanishing point (VP) detection scheme based on Hough Transform (HT) over an image-plane-mapped disk region. In the DiskVPS, the image plane is first mapped into a disk region, which is then partitioned into concentrical ring-shaped subspaces. Each subspace is further partitioned into approximately equal probability Hough cells. By using individual edges rather than edge pairs as voters, DiskVPS can achieve high accuracy with extreme efficiency. Involving no calibration parameters, DiskVPS is particularly suitable for VP detection in uncalibrated images and thus can be applied in image calibration. A comparative experimental study demonstrates that the basic DiskVPS model without parameter optimization achieved significantly better performance over the SOTA in detection accuracy and processing speed with real-world images. The study also shows that DiskVPS is robust against parameter changes." + }, + { + "type": "title", + "bbox": [ + 0.078, + 0.603, + 0.21, + 0.619 + ], + "angle": 0, + "content": "1. Introduction" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.628, + 0.47, + 0.825 + ], + "angle": 0, + "content": "VP detection has a variety of applications, including camera calibration ([1, 3, 5, 15]), 3D reconstruction ([25, 40]), virtual reality [30], robot navigation [12], scene tracking [6] and vehicle automatic driving [13]. Depending on the direction of the parallel lines and the setup of the capturing camera, a VP can appear anywhere, including infinity, in the image plane. A triplet of orthogonal VPs is formed by 3 sets of mutually orthogonal parallel lines in the man-made world commonly called Manhattan world [5]. According to [3], the information of any 2 orthogonal VPs in an image is sufficient to obtain the focal length of the capturing camera. Moreover, a triplet of orthogonal VPs fully determines the camera's intrinsic and extrinsic parameters." + }, + { + "type": "title", + "bbox": [ + 0.078, + 0.833, + 0.229, + 0.848 + ], + "angle": 0, + "content": "1.1. Related Works" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.856, + 0.47, + 0.903 + ], + "angle": 0, + "content": "VP detection has long been a key research topic attracting great attention in computer vision. However, it remains a challenging issue far from soundly resolved. The key is-" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.288, + 0.895, + 0.365 + ], + "angle": 0, + "content": "sue in VP detection is the unboundedness of the image plane in which VPs reside. In VP detection, it is essential to map the image plane into some bounded space before practical techniques such as Hough Transform (HT, [9]) can be deployed." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.368, + 0.895, + 0.791 + ], + "angle": 0, + "content": "The first such attempt came from [29] that divided the image plane into 3 bounded subspaces to perform the HT. Since its division method is rather arbitrary without offsetting the probability biases in cell partitioning, it failed to yield satisfactory results. For calibrated images, Barnard [2] proposed a VP detection algorithm that mapped the image plane into a unit sphere called Gaussian sphere (GS) for HT, which is widely used in VP detection ([18, 24]) and is also universally applied in deep learning VP detection algorithms([4, 10, 11, 16, 23, 27, 32, 39]). However, GS-based approaches have limitations. Since building the GS requires the known parameters of the capturing camera such as the focal length and coordinates of the optical axis in the image plane, GS approaches are unsuitable for uncalibrated images, which account for a substantial fraction of images of interest. GS's voting scheme using edge pairs can significantly amplify discretization and distortion errors in edge detection and thus cause a high likelihood of spurious VP detection. Accordingly, GS models are prone to spurious detection ([16, 24, 39]). Moreover, the process of partitioning the GS into Hough cells itself is complicated and delicate. The non-planar shape of the GS is not amenable for partitioning of the GS into planar Hough cells while the Hough cell sizes on the GS are wildly non-uniform and sensitive to the errors of calibration parameters. With the limitations, the GS baseline model did not draw wide attention until the recent rapid development of deep learning technology." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.796, + 0.895, + 0.903 + ], + "angle": 0, + "content": "As an important alternative to HT in VP detection, random sample consensus (RANSAC [8]) is another extensively explored scheme in VP detection ([17, 19, 21, 22, 26, 28, 33, 35, 38]). RANSAC can work directly in the image plane and is suitable for VP detection in uncalibrated images; moreover, RANSAC outperforms GS-based VP detection models in practice with scenes where many edges" + }, + { + "type": "page_number", + "bbox": [ + 0.479, + 0.945, + 0.521, + 0.958 + ], + "angle": 0, + "content": "27049" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.076, + 0.091, + 0.473, + 0.242 + ], + "angle": 0, + "content": "belong to VP inliers. The main drawback of RANSAC is its susceptibility to false negative detection when the size of samples accounts for only a tiny portion of the full set of edge pairs used for VP detection (3 or more edges belonging to a set of 2 or 3 orthogonal VPs are sometimes proposed in [19]). Additionally, RANSAC is computationally expensive, particularly for images of cluttered scenes. Recently, [34] proposed an algorithm that integrated RANSAC and HT to achieve real-time VP detection. There are other methods like using dual spaces ([14])." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.244, + 0.474, + 0.502 + ], + "angle": 0, + "content": "Despite its limitations, GS is the only proven VP detection model available to transform an unbounded space (image plane) into a bounded space (GS), enabling its deployment in deep learning neural networks. Therefore, the GS remains the chosen technique in deep-learning-based VP detection algorithms([4, 10, 11, 16, 23, 27, 32, 37, 39]). While the baseline model of GS-based VP detection is not optimal in both accuracy and processing speed, GS-based deep learning approaches making use of advanced high-level-feature extraction schemes and high-performance hardware have achieved the performance comparable to the SOTA, a level far superior to the GS's baseline model. However, deep learning approaches have limitations. The costs of data annotation and training needed by deep learning are expensive. Additionally, limited datasets usually do not generalize well in deep learning, which can fail in unexpected ways." + }, + { + "type": "title", + "bbox": [ + 0.078, + 0.514, + 0.216, + 0.529 + ], + "angle": 0, + "content": "1.2. Contribution" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.538, + 0.473, + 0.903 + ], + "angle": 0, + "content": "DiskVPS is suitable for conducting highly reliable and real-time VP detection. We use Cartesian coordinates \\((u, v)\\) to denote the normalized coordinates with the origin set at the image center and both image width and height normalized to 1 (see next section for how \\((u, v)\\) is mapped from the pixel-based image plane coordinates). DiskVPS maps an image plane (unbounded space) (Fig. 1(a)) into a disk-shaped space (Fig. 1(b)). Specifically, it maps an unbounded polar coordinate system \\((r, \\theta)\\) \\((r = \\sqrt{u^2 + v^2}\\) and \\(\\theta = \\tan^{-1} \\frac{u}{v}\\) into a bounded polar coordinate system \\((\\tan^{-1}(4r), \\theta)\\) with the latter's radial boundary located at \\(\\frac{\\pi}{2}\\). With the approach, a straight line that passes the origin in the image plane is mapped into a diametrical line segment in the disk. In contrast, a straight line not passing the origin in the image plane is mapped into an arc-shaped segment with 2 ends overlapped with its origin-passing parallel counterpart at the disk boundary. Consequently, parallel lines in the image plane intersect on the disk boundary, whereas non-parallel lines in the image plane intersect each other only once inside the disk. In the scheme, HT is reduced to a probabilistic method to partition the disk into a set of approximately equal-probability Hough cells. Fig. 2 shows a set of edges detected in a typical image (see Fig. 2(a)) and its corresponding Hough accumulation map generated by the" + }, + { + "type": "image", + "bbox": [ + 0.525, + 0.096, + 0.687, + 0.224 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.587, + 0.229, + 0.61, + 0.244 + ], + "angle": 0, + "content": "(a)" + }, + { + "type": "image", + "bbox": [ + 0.697, + 0.096, + 0.861, + 0.226 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.785, + 0.229, + 0.807, + 0.244 + ], + "angle": 0, + "content": "(b)" + }, + { + "type": "image_caption", + "bbox": [ + 0.5, + 0.256, + 0.895, + 0.3 + ], + "angle": 0, + "content": "Figure 1. Mapping from an image plane (Cartesian coordinates \\((u, v)\\) (a)) to a disk-shaped Hough space (polar coordinates \\((\\tan^{-1}(4r), \\theta)\\) (b)) with \\(r = \\sqrt{u^2 + v^2}\\) and \\(\\theta = \\tan^{-1} \\frac{v}{u}\\)." + }, + { + "type": "image", + "bbox": [ + 0.539, + 0.317, + 0.682, + 0.4 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.587, + 0.414, + 0.61, + 0.428 + ], + "angle": 0, + "content": "(a)" + }, + { + "type": "image", + "bbox": [ + 0.706, + 0.324, + 0.861, + 0.403 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.785, + 0.413, + 0.807, + 0.427 + ], + "angle": 0, + "content": "(b)" + }, + { + "type": "image_caption", + "bbox": [ + 0.5, + 0.439, + 0.892, + 0.468 + ], + "angle": 0, + "content": "Figure 2. Detected edges of a typical image (a) and its corresponding Hough accumulation map (b)." + }, + { + "type": "image", + "bbox": [ + 0.619, + 0.489, + 0.776, + 0.598 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.5, + 0.622, + 0.892, + 0.651 + ], + "angle": 0, + "content": "Figure 3. Intersection \\( P \\) of the extension line of a random edge and a given line defined by the distance \\( d \\) and the angle \\( \\phi \\)." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.676, + 0.894, + 0.768 + ], + "angle": 0, + "content": "edges using DiskVPS (see Fig. 2(b)). Five sharp and salient peaks represent 3 mutually orthogonal VPs, 2 far-field VPs generated by nearly parallel edges and one near-field VP generated by non-parallel edges. The proposed approach successfully manifests all valid VPs in the image without yielding any spurious VPs in Fig. 2b." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.782, + 0.715, + 0.8 + ], + "angle": 0, + "content": "2. Algorithm Description" + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.808, + 0.719, + 0.823 + ], + "angle": 0, + "content": "2.1. Theoretical Foundation" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.83, + 0.894, + 0.903 + ], + "angle": 0, + "content": "Assume \\((x,y)\\) is the pixel-based coordinates of a point in the image plane and the image of interest is encoded by 2 coordinates in pixels: \\((x_{\\min},y_{\\min})\\) and \\((x_{\\max},y_{\\max})\\), which refer to the pixel-based coordinates of the image's bottom left corner and top right corner respectively. To ease analy-" + }, + { + "type": "page_number", + "bbox": [ + 0.479, + 0.945, + 0.521, + 0.958 + ], + "angle": 0, + "content": "27050" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.175, + 0.091, + 0.372, + 0.232 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.077, + 0.244, + 0.47, + 0.273 + ], + "angle": 0, + "content": "Figure 4. Illustration of the 2 special directions along which the PDF is considered, the radial and tangential directions." + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.301, + 0.469, + 0.332 + ], + "angle": 0, + "content": "sis, \\((x,y)\\) is transformed to a normalized coordinate system \\((u,v)\\). Their relationship is written as:" + }, + { + "type": "equation", + "bbox": [ + 0.195, + 0.345, + 0.47, + 0.385 + ], + "angle": 0, + "content": "\\[\n\\left\\{ \\begin{array}{l} u = \\frac {x - 0 . 5 \\left(x _ {\\min } + x _ {\\max }\\right)}{x _ {\\max } - x _ {\\min }} \\\\ v = \\frac {y - 0 . 5 \\left(y _ {\\min } + y _ {\\max }\\right)}{y _ {\\max } - y _ {\\min }} \\end{array} \\right. \\tag {1}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.4, + 0.47, + 0.476 + ], + "angle": 0, + "content": "In the \\((u,v)\\) system, the origin \\(O\\) is fixed to the image center, and the bottom left corner and top right corner of the image are moved to \\((-0.5, -0.5)\\) and \\((0.5, 0.5)\\), respectively (see Fig. 3). The image region turns into an Origin-centered unit square in the \\((u,v)\\) system." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.477, + 0.473, + 0.658 + ], + "angle": 0, + "content": "[34] proposed a statistical model to estimate the probabilistic distribution of \\( P \\), the intersection of a given line and the extension line of a random edge, along the very given line. According to [34], the midpoint of the random edge is distributed uniformly in the unit square, and the direction of the random edge is also uniformly distributed over \\( [0, \\pi] \\). Such a random edge can be used to simulate the statistical distribution of the edges extracted from a sufficiently large number of images. We cite the conclusion of [34] in this paper. As shown in Fig. 3, if the given line is defined by \\( d \\) (the distance from \\( O \\) on the given line) and \\( \\phi \\) (the directional angle pointing from \\( O \\) to its perpendicular foot to the given line), then [34] gives:" + }, + { + "type": "equation", + "bbox": [ + 0.077, + 0.672, + 0.496, + 0.729 + ], + "angle": 0, + "content": "\\[\nf _ {d, \\phi} (t) = \\frac {1}{\\pi} \\int_ {- 0. 5 - 0. 5} ^ {0. 5} \\int_ {0} ^ {0. 5} \\frac {\\left| u \\cos \\phi + v \\sin \\phi - d \\right| d u d v}{(u - d \\cos \\phi - t \\sin \\phi) ^ {2} + (v - d \\sin \\phi + t \\cos \\phi) ^ {2}} \\tag {2}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.73, + 0.469, + 0.775 + ], + "angle": 0, + "content": "where \\( t \\) is the coordinate of \\( P \\) on a new axis created along the given line with the origin of the axis set at the perpendicular foot of \\( O \\) on the given line, as shown in Fig. 3." + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.776, + 0.47, + 0.902 + ], + "angle": 0, + "content": "Here, we are interested in Eq.2's solutions for 2 special cases. As shown in Fig.4, one is for the case in which the given line is along the radial direction in the \\((u,v)\\) coordinates, and the other is for the case in which the given line is in the azimuthal or tangential direction. \\(f_{\\mathrm{rad}}(r,\\theta)\\) and \\(f_{\\mathrm{tag}}(r,\\theta)\\) are used to denote the PDFs of \\(P\\) along the given line along the radial direction and the given line along the tangential direction respectively, where \\(\\theta\\) is the directional angle of \\(OP\\) and \\(r\\) is the length of \\(OP\\), both in the \\((u,v)\\)" + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.092, + 0.78, + 0.108 + ], + "angle": 0, + "content": "coordinate system. Accordingly, we have:" + }, + { + "type": "equation", + "bbox": [ + 0.534, + 0.116, + 0.892, + 0.18 + ], + "angle": 0, + "content": "\\[\n\\begin{array}{l} f _ {\\mathrm {r a d}} (r, \\theta) = f _ {0, \\theta + \\frac {\\pi}{2}} (r) \\\\ = \\frac {1}{\\pi} \\int_ {- 0. 5 - 0. 5} ^ {0. 5} \\int_ {0} ^ {0. 5} \\frac {\\left| - u \\sin \\theta + v \\cos \\theta \\right| d u d v}{(u - r \\cos \\theta) ^ {2} + (v - r \\sin \\theta) ^ {2}} \\tag {3} \\\\ \\end{array}\n\\]" + }, + { + "type": "equation", + "bbox": [ + 0.518, + 0.192, + 0.892, + 0.256 + ], + "angle": 0, + "content": "\\[\n\\begin{array}{l} f _ {\\mathrm {t a g}} (r, \\theta) = f _ {r, \\theta + \\frac {\\pi}{2}} (0) \\\\ = \\frac {1}{\\pi} \\int_ {- 0. 5} ^ {0. 5} \\int_ {- 0. 5} ^ {0. 5} \\frac {\\left| u \\sin \\theta - v \\cos \\theta + r \\right| d u d v}{(u + r \\sin \\theta) ^ {2} + (v - r \\cos \\theta) ^ {2}} \\tag {4} \\\\ \\end{array}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.261, + 0.892, + 0.332 + ], + "angle": 0, + "content": "Since \\( f_{\\mathrm{rad}}(r,\\theta) \\) and \\( f_{\\mathrm{tag}}(r,\\theta) \\)'s dependence on the variable \\( \\theta \\) is insignificant (see Fig.5), they are conveniently approximated by 2 less complex and more manageable forms \\( g_{\\mathrm{rad}}(r) = f_{\\mathrm{rad}}(r,0) \\) and \\( g_{\\mathrm{tag}}(r) = f_{\\mathrm{tag}}(r,0) \\) to facilitate subsequent construction of Hough space. They are:" + }, + { + "type": "equation", + "bbox": [ + 0.588, + 0.341, + 0.892, + 0.386 + ], + "angle": 0, + "content": "\\[\ng _ {\\text {r a d}} (r) = \\frac {1}{\\pi} \\int_ {- 0. 5 - 0. 5} ^ {0. 5} \\int_ {0} ^ {0. 5} \\frac {| v | d u d v}{(u - r) ^ {2} + v ^ {2}} \\tag {5}\n\\]" + }, + { + "type": "equation", + "bbox": [ + 0.585, + 0.398, + 0.892, + 0.442 + ], + "angle": 0, + "content": "\\[\ng _ {\\text {a g}} (r) = \\frac {1}{\\pi} \\int_ {- 0. 5 - 0. 5} ^ {0. 5} \\int_ {- 0. 5} ^ {0. 5} \\frac {| v - r | d u d v}{u ^ {2} + (v - r) ^ {2}} \\tag {6}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.52, + 0.447, + 0.85, + 0.463 + ], + "angle": 0, + "content": "The analytical solutions of \\( g_{\\mathrm{rad}}(r) \\) and \\( g_{\\mathrm{tag}}(r) \\) are:" + }, + { + "type": "equation", + "bbox": [ + 0.501, + 0.471, + 0.892, + 0.608 + ], + "angle": 0, + "content": "\\[\ng _ {\\mathrm {r a d}} (r) = \\left\\{ \\begin{array}{l l} \\frac {\\tan^ {- 1} (2 r + 1) - \\tan^ {- 1} (2 r - 1)}{\\pi} & \\\\ + \\frac {1}{2 \\pi} \\ln \\frac {[ (2 r - 1) ^ {2} + 1 ] [ (2 r + 1) ^ {2} + 1 ]}{(2 r - 1) ^ {2} (2 r + 1) ^ {2}} & r \\neq 0. 5 \\\\ + \\frac {r}{\\pi} \\ln \\frac {(2 r - 1) ^ {2} [ (2 r + 1) ^ {2} + 1 ]}{(2 r + 1) ^ {2} [ (2 r - 1) ^ {2} + 1 ]} & \\\\ \\frac {\\tan^ {- 1} 2 + \\ln 5 - \\ln 4}{\\pi} & r = 0. 5 \\end{array} \\right. \\tag {7}\n\\]" + }, + { + "type": "equation", + "bbox": [ + 0.501, + 0.628, + 0.892, + 0.821 + ], + "angle": 0, + "content": "\\[\ng _ {\\text {t a g}} (r) = \\left\\{ \\begin{array}{l l} 1 - \\frac {\\tan^ {- 1} (2 r + 1) - \\tan^ {- 1} (2 r - 1)}{\\pi} & r \\leq 0. 5 \\\\ - \\frac {2 r \\left[ \\tan^ {- 1} (2 r + 1) + \\tan^ {- 1} (2 r - 1) \\right]}{\\pi} & \\\\ + \\frac {\\ln [ 1 + (2 r + 1) ^ {2} ] + \\ln [ 1 + (2 r - 1) ^ {2} ]}{2 \\pi} & \\\\ 1 - \\frac {(2 r + 1) \\tan^ {- 1} (2 r + 1)}{\\pi} & \\\\ - \\frac {(2 r - 1) \\tan^ {- 1} (2 r - 1)}{\\pi} & r > 0. 5 \\\\ + \\frac {\\ln [ 1 + (2 r + 1) ^ {2} ] - \\ln [ 1 + (2 r - 1) ^ {2} ]}{2 \\pi} & \\end{array} \\right. \\tag {8}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.822, + 0.892, + 0.901 + ], + "angle": 0, + "content": "As shown in Fig. 5, both \\(\\frac{g_{\\mathrm{rad}}(r)}{f_{\\mathrm{rad}}(r,\\theta)}\\) and \\(\\frac{g_{\\mathrm{tag}}(r)}{f_{\\mathrm{tag}}(r,\\theta)}\\) lie between 0.88 and 1.13 for \\(r \\leq 1.7\\) and they move closer to 1 with the increase of \\(r\\) for \\(r \\geq 1.0\\). Overall, using \\(g_{\\mathrm{rad}}(r)\\) to approximate \\(f_{\\mathrm{rad}}(r,\\theta)\\) and using \\(g_{\\mathrm{tag}}(r)\\) to approximate \\(f_{\\mathrm{tag}}(r,\\theta)\\) do not lead to significant errors." + }, + { + "type": "page_number", + "bbox": [ + 0.479, + 0.945, + 0.518, + 0.957 + ], + "angle": 0, + "content": "27051" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.081, + 0.095, + 0.262, + 0.203 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.165, + 0.208, + 0.185, + 0.22 + ], + "angle": 0, + "content": "(a)" + }, + { + "type": "image", + "bbox": [ + 0.274, + 0.094, + 0.454, + 0.203 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.362, + 0.207, + 0.383, + 0.22 + ], + "angle": 0, + "content": "(b)" + }, + { + "type": "image_caption", + "bbox": [ + 0.077, + 0.233, + 0.471, + 0.262 + ], + "angle": 0, + "content": "Figure 5. Proximity of \\( g_{\\mathrm{rad}}(r) \\) to \\( f_{\\mathrm{rad}}(r,\\theta) \\) (a) and proximity of \\( g_{\\mathrm{tag}}(r) \\) to \\( f_{\\mathrm{tag}}(r,\\theta) \\) (b)." + }, + { + "type": "image", + "bbox": [ + 0.177, + 0.281, + 0.357, + 0.39 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.077, + 0.404, + 0.461, + 0.423 + ], + "angle": 0, + "content": "Figure 6. \\(\\frac{1}{\\pi} \\frac{4}{(4r)^2 + 1}\\) 's proximity to \\(f_{\\mathrm{rad}}(r, \\theta)\\) for chosen \\(r\\) values." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.45, + 0.469, + 0.525 + ], + "angle": 0, + "content": "The previous analysis provides a theoretical basis for building a Hough space from the \\((r,\\theta)\\). However, since the radial dimension \\((r)\\) remains unbounded, it is not feasible to create a Hough space directly from its current form (Eq.7 and 8)." + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.526, + 0.469, + 0.618 + ], + "angle": 0, + "content": "So it makes sense to transform \\( r \\) to a bounded space and derive the PDF in the new space. Note that when \\( r \\gg 1 \\), Eq.5 leads to \\( g_{\\mathrm{rad}}(r) \\approx \\frac{1}{\\pi r^2} \\int_{-0.5}^{0.5} \\int_{-0.5}^{0.5} |v| du dv = \\frac{1}{4\\pi r^2} \\). Since \\( \\frac{1}{4\\pi r^2} \\approx \\frac{1}{\\pi} \\frac{4}{(4r)^2 + 1} \\) for \\( r \\gg 1 \\), Eq. 5 can be further approximated as:" + }, + { + "type": "equation", + "bbox": [ + 0.171, + 0.618, + 0.469, + 0.647 + ], + "angle": 0, + "content": "\\[\ng _ {\\text {r a d}} (r) \\approx \\frac {1}{\\pi} \\frac {4}{(4 r) ^ {2} + 1} \\text {f o r} r \\gg 1 \\tag {9}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.652, + 0.469, + 0.801 + ], + "angle": 0, + "content": "Here \\(g_{\\mathrm{rad}}(r)\\approx \\frac{1}{\\pi}\\frac{4}{(4r)^2 + 1}\\) is chosen over other alternatives because the integration of \\(\\frac{1}{\\pi}\\frac{4}{(4r)^2 + 1}\\) over a full radial line is equal to 1, which matches the probability of a radial line intersecting the extension line of the random edge. Note \\(\\frac{1}{\\pi}\\int \\limits_{r = -\\infty}^{r = +\\infty}\\frac{4dr}{(4r)^2 + 1} = \\frac{1}{\\pi}\\int \\limits_{r = -\\infty}^{r = +\\infty}d\\tan^{-1}(4r) = \\frac{1}{\\pi}\\times 2\\times \\frac{\\pi}{2} = 1.\\) As shown in Fig.6, when \\(r\\geq 1.5\\) Eq.9 is already a good approximation of \\(f_{\\mathrm{rad}}(r,\\theta)\\) as \\(\\frac{\\frac{1}{\\pi}\\frac{4}{(4r)^2 + 1}}{f_{\\mathrm{rad}}(r,\\theta)}\\) lies consistently between 0.93 and 1.08 for \\(r\\geq 1.5\\)" + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.802, + 0.468, + 0.831 + ], + "angle": 0, + "content": "Eq.9 creates a route to transform \\( r \\) to a bounded space. To achieve the goal, we introduce a new function \\( \\psi(r) \\):" + }, + { + "type": "equation", + "bbox": [ + 0.214, + 0.843, + 0.469, + 0.86 + ], + "angle": 0, + "content": "\\[\n\\psi (r) = \\tan^ {- 1} (4 r) \\tag {10}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.078, + 0.872, + 0.469, + 0.903 + ], + "angle": 0, + "content": "Unlike \\(r\\), \\(\\psi\\) is defined in a bounded space since \\(\\psi \\in [0, \\frac{\\pi}{2}]\\). We denote the PDF along \\(\\psi\\) as \\(h_{\\mathrm{rad}}(\\psi)\\), which is" + }, + { + "type": "image", + "bbox": [ + 0.532, + 0.094, + 0.687, + 0.227 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.596, + 0.232, + 0.616, + 0.245 + ], + "angle": 0, + "content": "(a)" + }, + { + "type": "image", + "bbox": [ + 0.695, + 0.09, + 0.874, + 0.228 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.777, + 0.231, + 0.798, + 0.245 + ], + "angle": 0, + "content": "(b)" + }, + { + "type": "image_caption", + "bbox": [ + 0.499, + 0.257, + 0.892, + 0.3 + ], + "angle": 0, + "content": "Figure 7. Depiction of DiskVPS's Hough space. (a) The structure of the Hough space; (b) The parameters of a Hough cell in the \\(i\\)th subspace." + }, + { + "type": "text", + "bbox": [ + 0.5, + 0.325, + 0.82, + 0.341 + ], + "angle": 0, + "content": "related to \\(g_{\\mathrm{rad}}(r)\\) through the following formula:" + }, + { + "type": "equation", + "bbox": [ + 0.626, + 0.35, + 0.891, + 0.377 + ], + "angle": 0, + "content": "\\[\nh _ {\\mathrm {r a d}} (\\psi) = g _ {\\mathrm {r a d}} (r) \\left| \\frac {d r}{d \\psi} \\right| \\tag {11}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.501, + 0.388, + 0.891, + 0.426 + ], + "angle": 0, + "content": "Combining Eq.9, Eq.10 and Eq.11, and noting \\(\\frac{d(\\tan^{-1}x)}{dx} = \\frac{1}{x^2 + 1}\\), we have:" + }, + { + "type": "equation", + "bbox": [ + 0.565, + 0.434, + 0.891, + 0.463 + ], + "angle": 0, + "content": "\\[\nh _ {\\mathrm {r a d}} (\\psi) \\approx \\frac {1}{\\pi} \\frac {4}{(4 r) ^ {2} + 1} \\div \\frac {4}{(4 r) ^ {2} + 1} = \\frac {1}{\\pi} \\tag {12}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.471, + 0.891, + 0.503 + ], + "angle": 0, + "content": "Eq.12 indicates that the PDF along \\(\\psi\\) is approximately constant \\((= \\frac{1}{\\pi})\\) in the space of \\((\\psi, \\theta)\\) (for \\(r \\geq 1.5\\))." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.511, + 0.763, + 0.528 + ], + "angle": 0, + "content": "2.2. Construction of Hough Space" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.535, + 0.892, + 0.654 + ], + "angle": 0, + "content": "Structurally, the Hough space is partitioned into a set of \\(M\\) concentric ring-shaped subspaces, each of which is further partitioned into a set of at most \\(N\\) fan-shaped cells, as shown in Fig.7(a). The principle is to ensure that each subspace has a roughly equal probability of intersecting the extension line of the random edge along the radial direction and each cell has a roughly equal probability of intersecting the line in the tangential direction." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.654, + 0.892, + 0.766 + ], + "angle": 0, + "content": "Our first goal is to obtain each subspace's inner radius and radial span using the principle above. We denote the inner radius and the radial span of the \\(i\\)th subspace as \\(r_i\\) and \\(\\delta r_i\\) respectively, where \\(i = 1,2,\\ldots,M\\) is the order number of the subspace counting from the innermost outwards. Noting the outer radius of the \\(i\\)th subspace is the inner radius of the \\(i + 1\\)th subspace, there exist the following relations throughout the subspaces:" + }, + { + "type": "equation", + "bbox": [ + 0.648, + 0.778, + 0.891, + 0.859 + ], + "angle": 0, + "content": "\\[\n\\begin{array}{l} r _ {1} = 0 \\\\ r _ {2} = r _ {1} + \\delta r _ {1} \\\\ \\dots \\tag {13} \\\\ r _ {i + 1} = r _ {i} + \\delta r _ {i} \\\\ \\dots \\\\ \\end{array}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.5, + 0.871, + 0.892, + 0.901 + ], + "angle": 0, + "content": "Since a full radial line intersects the extension line of the random edge exactly once (the special case where the 2" + }, + { + "type": "page_number", + "bbox": [ + 0.479, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "27052" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.076, + 0.092, + 0.47, + 0.152 + ], + "angle": 0, + "content": "lines are parallel and do not intersect has infinitesimal probability and is therefore ignored), the probability of the radial span of each subspace intersecting the extension line of the random edge is \\(\\frac{1}{M}\\)." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.152, + 0.474, + 0.298 + ], + "angle": 0, + "content": "In addition to setting the inner radius and the radial span of the \\(i\\)th subspace as \\(r_i\\) and \\(\\delta r_i\\) for near-field subspaces, we also set the inner radius and the radial span of the subspace as \\(\\psi_i = \\tan^{-1}(4r_i)\\) and \\(\\delta \\psi_i\\) for far-field subspaces. Assuming that the PDF along the radial span of the \\(i\\)th space is constant (\\(= g_{\\mathrm{rad}(r_i)}\\) for near-field subspace and \\(= h_{\\mathrm{rad}}(\\psi_i)\\) for far-field subspace), then the probability \\(\\frac{1}{M}\\) must match with \\(g_{\\mathrm{rad}}(r_i)\\delta r_i\\) if the \\(i\\)th subspace is in the near field or \\(h_{\\mathrm{rad}}(\\psi_i)\\delta \\psi_i\\) if the \\(i\\)th subspace is in the far field. Since \\(h_{\\mathrm{rad}}(\\psi) \\approx \\frac{1}{\\pi}\\) (see Eq.12), we have:" + }, + { + "type": "equation", + "bbox": [ + 0.171, + 0.307, + 0.47, + 0.34 + ], + "angle": 0, + "content": "\\[\n\\delta r _ {i} = \\frac {1}{M g _ {\\mathrm {r a d}} \\left(r _ {i}\\right)} \\quad \\text {f o r} r _ {i} \\leq 1. 5 \\tag {14}\n\\]" + }, + { + "type": "equation", + "bbox": [ + 0.174, + 0.338, + 0.374, + 0.362 + ], + "angle": 0, + "content": "\\[\n\\delta \\psi_ {i} = \\frac {\\pi}{M} \\quad \\text {f o r} r _ {i} > 1. 5\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.375, + 0.47, + 0.437 + ], + "angle": 0, + "content": "Combining \\(\\delta \\psi_{i} = \\psi_{i + 1} - \\psi_{i} = \\tan^{-1}(4r_{i + 1}) - \\tan^{-1}(4r_{i})\\) and \\(\\delta r_i = r_{i + 1} - r_i\\), the 2nd term of Eq.14 turns into \\(\\delta r_{i} = \\frac{\\tan(\\frac{\\pi}{M} + \\tan^{-1}4r_{i})}{4} -r_{i}\\). And Eq.14 can be rewritten as:" + }, + { + "type": "equation", + "bbox": [ + 0.129, + 0.447, + 0.469, + 0.505 + ], + "angle": 0, + "content": "\\[\n\\delta r _ {i} = \\left\\{ \\begin{array}{l l} \\frac {1}{M g _ {\\text {r a d}} \\left(r _ {i}\\right)} & \\text {f o r} r _ {i} \\leq 1. 5 \\\\ \\frac {\\tan \\left(\\frac {\\pi}{M} + \\tan^ {- 1} 4 r _ {i}\\right)}{4} - r _ {i} & \\text {f o r} r _ {i} > 1. 5 \\end{array} \\right. \\tag {15}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.517, + 0.47, + 0.652 + ], + "angle": 0, + "content": "The Hough space partitioning starts from \\( i = 1 \\). Eq.13 and Eq.15 work together to obtain the inner radius \\( r_i \\) and the radial span \\( \\delta r_i \\) of the \\( i \\)th subspace iteratively (\\( i = i + 1 \\) after each iteration) until the \\( i \\)th subspace comes to the Hough space boundary that is indicated by \\( \\psi_{i+1} \\geq \\frac{\\pi}{2} \\) or \\( \\tan^{-1}(4r_i) + \\frac{\\pi}{M} \\geq \\frac{\\pi}{2} \\). The final count of the subspaces is denoted as \\( M^* = i \\) at the end of the last iteration. Since the PDF approximations are used in the Hough space partitioning, \\( M^* \\) is generally not equal, but very close, to \\( M \\)." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.653, + 0.471, + 0.865 + ], + "angle": 0, + "content": "Afterward, each subspace is partitioned into multiple equal-angle-span cells. For the \\(i\\)th subspace, the angle span of a cell is denoted as \\(\\delta \\theta_{i}\\) (\\(i = 1,2,\\ldots M^{*}\\)), as shown in Fig.7(b). We observe that the inner circle of the outermost (the \\(M^{*}\\)th) subspace intersects the extension line of the random edge exactly twice. In the meantime, the very circle is partitioned into \\(N\\) equal-angle-span arcs, each of which forms the inner arc of a cell. Consequently, the probability of the inner arc of a cell intersecting the random edge's extension line is \\(\\frac{2}{N}\\). To make sure that each cell in any subspace gets the same probability \\(\\frac{2}{N}\\) of intersecting the random edge's extension line, this probability must match the probability obtained through the product of the cell's inner arc length (\\(= \\delta \\theta_{i}r_{i}\\)) and the PDF along the arc (\\(= g_{\\mathrm{tag}}(r_i)\\)) throughout the entire Hough space. Accordingly, we have:" + }, + { + "type": "equation", + "bbox": [ + 0.214, + 0.875, + 0.47, + 0.906 + ], + "angle": 0, + "content": "\\[\n\\delta \\theta_ {i} = \\frac {2}{N r _ {i} g _ {\\text {t a g}} \\left(r _ {i}\\right)} \\tag {16}\n\\]" + }, + { + "type": "image", + "bbox": [ + 0.529, + 0.095, + 0.865, + 0.279 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.499, + 0.298, + 0.893, + 0.339 + ], + "angle": 0, + "content": "Figure 8. A typical scenario of a voting edge casting its votes in the \\(i\\)th subspace with 1 or 0 inside each cell referring to the number of votes each cell receives." + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.362, + 0.892, + 0.391 + ], + "angle": 0, + "content": "The number of cells the \\(i\\)th subspace is partitioned into is denoted as \\(n_i (i = 1, 2, \\ldots, M^*)\\). It is written as:" + }, + { + "type": "equation", + "bbox": [ + 0.639, + 0.399, + 0.892, + 0.427 + ], + "angle": 0, + "content": "\\[\nn _ {i} = \\operatorname {r o u n d} \\left(\\frac {2 \\pi}{\\delta \\theta_ {i}}\\right) \\tag {17}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.434, + 0.892, + 0.478 + ], + "angle": 0, + "content": "where \\(2\\pi\\) is the angle span of the \\(i\\)th subspace, and \\(\\delta \\theta_{i}\\) is the angle span of a cell in the \\(i\\)th subspace, and round refers to the namesake maths function." + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.479, + 0.892, + 0.508 + ], + "angle": 0, + "content": "Finally, \\(\\delta \\theta_{i}\\) needs to be slightly modified due to \\(n_i\\)'s discrete nature. Its adjusted form, \\(\\delta \\theta_{i}^{*}\\), is written as:" + }, + { + "type": "equation", + "bbox": [ + 0.662, + 0.516, + 0.892, + 0.544 + ], + "angle": 0, + "content": "\\[\n\\delta \\theta_ {i} ^ {*} = \\frac {2 \\pi}{n _ {i}} \\tag {18}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.551, + 0.892, + 0.628 + ], + "angle": 0, + "content": "This way, the \\(i\\)th subspace is partitioned into \\(n_i\\) equal-angle-span cells, each covering an angle span of \\(\\delta \\theta_i^*\\). Via iterating \\(i\\) from 1 through \\(M^*\\), all subspaces are thus partitioned into approximately equally probabilistic cells in terms of intersecting the extension line of the random edge." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.636, + 0.822, + 0.653 + ], + "angle": 0, + "content": "2.3. Implementation of Hough Transform" + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.66, + 0.892, + 0.704 + ], + "angle": 0, + "content": "A Hough accumulation map, each of whose elements is mapped from a corresponding Hough cell, is constructed and the magnitude of each element is initialized to zero." + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.705, + 0.892, + 0.734 + ], + "angle": 0, + "content": "The VP detection consists of a voting process and an iterative VP detection process." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.735, + 0.892, + 0.902 + ], + "angle": 0, + "content": "In the voting process, each edge casts its votes based on whether its extension line intersects any of the subspaces, and if yes then which specific cell(s) of the subspace the line intersects, with each intersecting cell receiving exactly 1 vote. Assume the extension line of the voting edge is defined by \\( d \\), the distance from \\( O \\) to the line, and \\( \\varphi \\), the azimuthal angle of \\( O \\)'s perpendicular foot on the line. For an edge's voting in the \\( i \\)th subspace, there exist 3 cases, namely \\( d < r_i \\), \\( r_i \\leq d < r_i + \\delta r_i \\), and \\( d \\geq r_i + \\delta r_i \\) respectively, where \\( r_i \\) and \\( \\delta r_i \\) are the inner radius and radial span of the \\( i \\)th subspace." + }, + { + "type": "page_number", + "bbox": [ + 0.479, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "27053" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.077, + 0.092, + 0.473, + 0.18 + ], + "angle": 0, + "content": "A scenario of the 1st case (\\(d < r_i\\)) is shown in Fig.8. The extension line of the voting edge intersects the \\(i\\)th subspace boundary at 4 points, namely \\(P_1, P_2, P_3\\), and \\(P_4\\). Assume \\(\\Psi_{\\overrightarrow{OP_1}}\\), \\(\\Psi_{\\overrightarrow{OP_2}}\\), \\(\\Psi_{\\overrightarrow{OP_3}}\\), and \\(\\Psi_{\\overrightarrow{OP_4}}\\) are the azimuthal angles of vectors \\(\\overrightarrow{OP_1}\\), \\(\\overrightarrow{OP_2}\\), \\(\\overrightarrow{OP_3}\\) and \\(\\overrightarrow{OP_4}\\), respectively. Then we have:" + }, + { + "type": "equation", + "bbox": [ + 0.177, + 0.182, + 0.47, + 0.253 + ], + "angle": 0, + "content": "\\[\n\\left\\{ \\begin{array}{l} \\Psi_ {\\overrightarrow {O P _ {1}}} = \\varphi + \\cos^ {- 1} \\frac {d}{r _ {i} + \\delta r _ {i}} \\\\ \\Psi_ {\\overrightarrow {O P _ {2}}} = \\varphi + \\cos^ {- 1} \\frac {d}{r _ {i}} \\\\ \\Psi_ {\\overrightarrow {O P _ {3}}} = \\varphi - \\cos^ {- 1} \\frac {d}{r _ {i}} \\\\ \\Psi_ {\\overrightarrow {O P _ {4}}} = \\varphi - \\cos^ {- 1} \\frac {d}{r _ {i} + \\delta r _ {i}} \\end{array} \\right. \\tag {19}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.257, + 0.473, + 0.329 + ], + "angle": 0, + "content": "As shown in Fig.8, \\( P_{1}P_{2} \\) and \\( P_{3}P_{4} \\) are the line's 2 sections intersecting the \\( i \\)th subspace. Assume the \\( n_{i} \\) cells of the \\( i \\)th subspace are labeled 1 through \\( n_{i} \\) and a set of consecutive cells that intersect \\( P_{1}P_{2} \\) have the labels numbered \\( j \\) through \\( k \\), then \\( j \\) and \\( k \\) are written as:" + }, + { + "type": "equation", + "bbox": [ + 0.218, + 0.338, + 0.47, + 0.386 + ], + "angle": 0, + "content": "\\[\n\\left\\{\\begin{array}{l}j = \\left\\lceil \\frac {\\Psi_ {\\overrightarrow {O P _ {2}}}}{\\delta \\theta_ {i} ^ {*}} \\right\\rceil\\\\k = \\left\\lceil \\frac {\\Psi_ {\\overrightarrow {O P _ {1}}}}{\\delta \\theta_ {i} ^ {*}} \\right\\rceil\\end{array}\\right. \\tag {20}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.078, + 0.396, + 0.47, + 0.425 + ], + "angle": 0, + "content": "where \\(\\delta \\theta_{i}^{*}\\) is defined in Eq.18, and \\(\\lceil \\rceil\\) refers to maths function ceil." + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.427, + 0.47, + 0.485 + ], + "angle": 0, + "content": "Similarly, we can obtain the set of consecutive cells intersecting \\( P_{3}P_{4} \\). In voting, each cell in the two sets receives 1 vote from the voting edge while the rest of the \\( n_{i} \\) cells receive no votes." + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.488, + 0.469, + 0.533 + ], + "angle": 0, + "content": "For the 2nd case, voting is performed similarly to the first, but only the angle span between \\(\\Psi_{\\overrightarrow{OP_1}}\\) and \\(\\Psi_{\\overrightarrow{OP_4}}\\) needs to be handled as neither \\(P_{2}\\) nor \\(P_{3}\\) exists." + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.534, + 0.469, + 0.564 + ], + "angle": 0, + "content": "For the 3rd case, the subspace does not intersect the line; thus, no cell receives any vote." + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.565, + 0.469, + 0.61 + ], + "angle": 0, + "content": "The voting process generally yields a Hough accumulation map with a small number of salient peaks scattered over the map similar to Fig.2(b)." + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.611, + 0.469, + 0.686 + ], + "angle": 0, + "content": "Then, VP detection is as simple as finding salient peaks in the Hough accumulation map. In practice, at most \\( N_{\\mathrm{VP}} \\) dominant VPs are detected via at most \\( N_{\\mathrm{VP}} \\) iterations with each iteration extracting exactly 1 VP. In each iteration, the following steps are taken:" + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.687, + 0.469, + 0.717 + ], + "angle": 0, + "content": "Step 1: Search the global peak in the current Hough accumulation space." + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.718, + 0.469, + 0.747 + ], + "angle": 0, + "content": "Step 2: Obtain the location of the peak and set it as the initial position of the current VP." + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.749, + 0.469, + 0.809 + ], + "angle": 0, + "content": "Step 3: Search all of the inlier edges of the current VP. An edge is regarded as an inlier of the VP only if the angle between its extension line and the line that passes both the midpoint of the edge and the VP is no more than \\(\\Theta\\)." + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.81, + 0.469, + 0.869 + ], + "angle": 0, + "content": "Step 4: Update the Hough accumulation map by removing all of the votes cast by the inliers of the current VP. This ensures that inliers of the dominant VPs will not affect the subsequent detection of less dominant VPs." + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.871, + 0.469, + 0.901 + ], + "angle": 0, + "content": "Step 5: If the peak value in the updated accumulation map is at least 3 times as much as the average and the total" + }, + { + "type": "image", + "bbox": [ + 0.526, + 0.094, + 0.622, + 0.169 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.554, + 0.181, + 0.575, + 0.194 + ], + "angle": 0, + "content": "(a)" + }, + { + "type": "image", + "bbox": [ + 0.648, + 0.094, + 0.745, + 0.169 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.687, + 0.181, + 0.708, + 0.194 + ], + "angle": 0, + "content": "(b)" + }, + { + "type": "image", + "bbox": [ + 0.77, + 0.092, + 0.871, + 0.17 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.82, + 0.181, + 0.84, + 0.194 + ], + "angle": 0, + "content": "(c)" + }, + { + "type": "image", + "bbox": [ + 0.523, + 0.195, + 0.624, + 0.275 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.555, + 0.284, + 0.574, + 0.298 + ], + "angle": 0, + "content": "(f)" + }, + { + "type": "image", + "bbox": [ + 0.647, + 0.195, + 0.747, + 0.275 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.686, + 0.284, + 0.706, + 0.298 + ], + "angle": 0, + "content": "(e)" + }, + { + "type": "image", + "bbox": [ + 0.77, + 0.196, + 0.871, + 0.275 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.818, + 0.284, + 0.839, + 0.298 + ], + "angle": 0, + "content": "(d)" + }, + { + "type": "image_caption", + "bbox": [ + 0.499, + 0.31, + 0.892, + 0.449 + ], + "angle": 0, + "content": "Figure 9. Graphical demonstration of the steps in DiskVPS' VP detection: (a) is the original image; (b) is the detected edge set; (c) is the initial Hough accumulation colormap with its global peak (also the position of the 1st VP) marked in red; (d) is the updated Hough accumulation map after removing the votes cast by the inliers of the 1st VP with 1 detected VP and the updated peak marked in red; (e) is the updated Hough accumulation map after the 2nd iteration with 2 detected VPs and the updated peak marked in red; (f) is the updated Hough accumulation map after the 3rd iteration with 3 detected VPs marked in red." + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.475, + 0.89, + 0.506 + ], + "angle": 0, + "content": "number of detected VPs is less than \\( N_{\\mathrm{VP}} \\), then go to Step 1. Otherwise, stop iteration and output all detected VPs." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.507, + 0.892, + 0.764 + ], + "angle": 0, + "content": "The VP detection process is demonstrated in Fig.9. The zenith, which is the convergent point of nearly vertical edges, is located at the local maximum around the south pole of the disk-shaped colormap and the inliers of the zenith also lead to the significantly enhanced distribution around the north pole; the \\(1^{st}\\) horizontal VP at the east pole is formed by nearly parallel edges which also contribute to the enhanced accumulation around the west pole; non-parallel edges form the \\(2^{nd}\\) horizontal VP and generate a corresponding peak in the interior of the map. Some local maxima are noises caused by combined voting of inliers of different VPs and outliers of the VPs. However, after 3 iterations almost all local maxima unrelated to any valid VPs in the original accumulation map disappear (see Fig. 9(f)). It demonstrates that the vote removal process of the inliers of the detected VPs in Step 4 effectively minimizes false positive detection." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.782, + 0.625, + 0.8 + ], + "angle": 0, + "content": "3. Experiment" + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.809, + 0.892, + 0.869 + ], + "angle": 0, + "content": "The proposed approach is examined and evaluated against the state-of-the-art (SOTA) in VP detection. It was implemented in \\(\\mathrm{C + + }\\) and tested on a PC powered by an Intel i7 CPU (3.5 GHZ)." + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.871, + 0.892, + 0.901 + ], + "angle": 0, + "content": "DiskVPS was evaluated via 3 benchmark datasets, the York Urban Dataset (YUD [7]), SVVP([27]), and" + }, + { + "type": "page_number", + "bbox": [ + 0.479, + 0.945, + 0.52, + 0.957 + ], + "angle": 0, + "content": "27054" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.092, + 0.088, + 0.177, + 0.442 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.185, + 0.089, + 0.268, + 0.442 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.279, + 0.089, + 0.362, + 0.442 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.373, + 0.089, + 0.455, + 0.442 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.076, + 0.455, + 0.47, + 0.54 + ], + "angle": 0, + "content": "Figure 10. Example images from YUD [7] (the 1st row), SVVP[27] (the 3rd row), and SU3 [40] (the 5th row) datasets with each even row immediately below displaying their generated disk-shaped Hough colormaps and the VP detection results. The ground truth VPs are marked with white \"+\" and the detected VPs are marked with red \"×\"." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.566, + 0.468, + 0.791 + ], + "angle": 0, + "content": "SU3([40]). YUD contains 102 calibrated real-world images of \\(640 \\times 480\\) pixels. SVVP contains 500 calibrated real-world images of \\(512 \\times 512\\) pixels, and SU3 contains a photorealistic dataset containing 23,000 calibrated synthetic outdoor images of \\(512 \\times 512\\) pixels. The commonness of the 3 benchmark datasets is that they all satisfy the Manhattanworld assumption [5] as there exists a triplet of orthogonal VPs in each image. So our experiment is focused on how accurately DiskVPS performs VP detections on the triplets of orthogonal VPs in each dataset compared to the SOTA. Please note that while the images used in the experiment were calibrated, DiskVPS treated those images as uncalibrated and performed VP detection without using any calibration info until checking accuracy against the ground truth." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.796, + 0.468, + 0.903 + ], + "angle": 0, + "content": "DiskVPS uses edges as input. In preprocessing, there are many algorithms to perform edge detection (see LSD([31]), [20], and [36], etc.). A similar algorithm to LSD was selected to perform edge detection. Table1 lists the 4 parameters and their respective magnitude that our implementation used for evaluation. \\(M\\) is the number of concentric subspaces that the disk-shaped Hough space is roughly parti" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.092, + 0.892, + 0.198 + ], + "angle": 0, + "content": "tioned into, \\(N\\) is the maximum number of roughly equal-probability cells that each of the subspaces is partitioned into, \\(N_{\\mathrm{VP}}\\) is the maximum number of dominant VPs that are initially extracted from each image, and \\(\\Theta\\) is the criterion angle to separate the inliers of a VP from its outliers. The parameters are fixed throughout the experiment despite the differences among the 3 datasets." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.199, + 0.892, + 0.394 + ], + "angle": 0, + "content": "Fig.10 lists some example images of the 3 datasets and their VP detection results of our approach. It demonstrates that the disk-shaped Hough accumulation map not only manifests almost all valid VPs but also generates no noticeable spurious VPs. Most detected VPs are either very close to or precisely overlapped with their ground truth counterparts. The only exception happens in the 6th row and the 3rd column, in which the zenith's ground truth lies around the north pole (corresponding to \\((0, +\\infty)\\) in \\((u, v)\\) coordinates) while its detected location is around the south pole \\((0, -\\infty)\\) in \\((u, v)\\). The discrepancy is reconciled as the VP at \\((0, +\\infty)\\) and the VP at \\((0, -\\infty)\\) in the \\((u, v)\\) coordinates are formed by the same group of vertical edges." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.395, + 0.892, + 0.621 + ], + "angle": 0, + "content": "In the experiment, DiskVPS is compared with six SOTA algorithms ([11, 16, 26, 27, 34, 39]) using a protocol adopted by [11, 16, 27]. In the protocol, a triple of orthogonal VPs are detected from each image, then 3 angular differences in the camera space between the 3 detected VPs and their respective ground truth are obtained, with the largest of the 3 differences adopted as the angular difference of the image. According to the protocol, the percentage of a dataset's images that have a smaller angular difference than a given threshold is used as the angular accuracy (AA) criterion. In the experiment, AA@3°, AA@5°, and AA@10° represent AA thresholds of 3°, 5°, and 10° respectively. For example, an algorithm achieving 80% on AA@5° with a dataset means that the algorithm achieves an angular difference of 5° or less on 80% of the images in the dataset." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.622, + 0.892, + 0.759 + ], + "angle": 0, + "content": "In DiskVPS implementation, a triplet of orthogonal VPs is picked from the initial \\( N_{\\mathrm{VP}} \\) VPs based on their mutual orthogonality (see [3] for details of VP orthogonality). However, there is a small fraction of images in which only 2 of the 3 orthogonal VPs are captured in the initial \\( N_{\\mathrm{VP}} \\) VPs. This happens when more dominant non-orthogonal VPs suppress a less salient orthogonal VP and thus miss the initial detection. In the latter case, DiskVPS detects the 3rd orthogonal VP by searching for the local maximum in" + }, + { + "type": "table", + "bbox": [ + 0.621, + 0.782, + 0.774, + 0.874 + ], + "angle": 0, + "content": "
NameMagnitude
M200
N200
NVP4
Θ3°
" + }, + { + "type": "table_caption", + "bbox": [ + 0.5, + 0.884, + 0.886, + 0.899 + ], + "angle": 0, + "content": "Table 1. Parameters used in the implementation of our algorithm." + }, + { + "type": "page_number", + "bbox": [ + 0.479, + 0.945, + 0.518, + 0.957 + ], + "angle": 0, + "content": "27055" + } + ], + [ + { + "type": "table", + "bbox": [ + 0.107, + 0.089, + 0.868, + 0.242 + ], + "angle": 0, + "content": "
Datasets MetricsFPSYUD [7]SVVP[27]SU3 [40]
AA@3°AA@5°AA@10°AA@3°AA@5°AA@10°AA@3°AA@5°
J-Linkage[26]1.060.871.881.527.343.260.882.087.2
Wu et al. [34]3958.373.482.243.860.175.782.387.6
CONSAC [11]3.061.773.384.443.856.569.486.290.4
NeurVPS [39]0.560.871.881.532.845.760.293.996.3
Lin et al. [16]2360.774.386.351.868.782.284.490.6
Tong et al. [27]2565.577.187.451.667.782.691.394.6
Ours (DiskVPS)15874.590.299.058.381.698.484.391.7
" + }, + { + "type": "table_caption", + "bbox": [ + 0.076, + 0.251, + 0.895, + 0.296 + ], + "angle": 0, + "content": "Table 2. Processing speed and angular accuracy (AA) comparison on YUD( [7]), SVVP([27]) and SU3( [40]). Our algorithm achieves by far the best average processing speed on the 3 datasets. Moreover, our approach significantly outperforms all other approaches on real-world YUD and SVVP, and is competitive on synthetic SU3." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.318, + 0.473, + 0.468 + ], + "angle": 0, + "content": "the Hough space around the 3rd VP's predicted location derived from the 2 known orthogonal VPs. In the experiment, it was observed that when \\( N_{\\mathrm{VP}} = 4 \\), DiskVPS successfully captured a triple of orthogonal VPs in the initial \\( N_{\\mathrm{VP}} \\) VPs in 90 (88% of the total) images in YUD, while it captured 2 of the triplet of orthogonal VPs in 11 (11% of the total) images of YUD in the initial \\( N_{\\mathrm{VP}} \\). After executing the 3rd VP detection on the 11 images, 101 of the 102 images in YUD have their respective triples of orthogonal VPs successfully detected." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.471, + 0.473, + 0.624 + ], + "angle": 0, + "content": "The comparison results are shown in Table 2. Our approach achieves \\(74.5\\%\\) at AA@3° on YUD, \\(9\\%\\) higher than the second best at \\(65.5\\%\\) achieved by [27], a significant improvement over the SOTA. Similar patterns are observed at AA@5° and AA@10° on YUD and SVVP. Notably, our approach achieves \\(99.0\\%\\) on YUD and \\(98.4\\%\\) on SVVP at AA@10° whereas none of any other approaches reaches \\(90\\%\\) or higher on this respect. Overall, for real-world images, our approach significantly outperforms the SOTA as far as angular accuracy is concerned." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.625, + 0.473, + 0.777 + ], + "angle": 0, + "content": "On the synthetic SU3, DiskVPS is also competitive. Note that the number of images of SU3 is roughly 40 times as large as YUD and SVD combined and the two approaches ([27, 39]) that outperformed DiskVPS are both deep-learning (DL) methods. We believe the large size of SU3 provides substantially more training and helps significantly optimize the DL algorithms. Considering that DiskVPS is a non-optimized basic VP detection model involving no calibration information which the DL models require, the accuracy of DiskVPS is still impressive." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.778, + 0.473, + 0.884 + ], + "angle": 0, + "content": "Moreover, we increased \\(M\\) from 200 to 300 and \\(N\\) from 100 to 180 (both at stepsize of 5) and observed the angular accuracy (AA) distribution at AA@\\(3^{\\circ}\\), AA@\\(5^{\\circ}\\), and AA@\\(10^{\\circ}\\). The results are shown in Table 3. Since the standard deviations (STDEV) are consistently below \\(3\\%\\) of their respective means, it is evident that DiskVPS is reasonably robust against changing parameters." + }, + { + "type": "text", + "bbox": [ + 0.097, + 0.886, + 0.471, + 0.902 + ], + "angle": 0, + "content": "Lastly, we also compare our approach's efficiency with" + }, + { + "type": "table", + "bbox": [ + 0.504, + 0.315, + 0.892, + 0.394 + ], + "angle": 0, + "content": "
CriterionMinimumMaximumMeanSTDEV
AA@3°70.678.475.32.2
AA@5°89.294.191.51.3
AA@10°99.099.099.00
" + }, + { + "type": "table_caption", + "bbox": [ + 0.5, + 0.403, + 0.892, + 0.431 + ], + "angle": 0, + "content": "Table 3. Angular accuracy statistics over \\(200\\leq M\\leq 300\\) and \\(100\\le N\\le 180\\)" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.46, + 0.892, + 0.779 + ], + "angle": 0, + "content": "other algorithms. Our approach achieves 158 frames per second (fps) or 6.3 milliseconds per frame excluding image decoding and edge detection. It is around 4 times as fast as the second best at 39 fps achieved by [34]. DiskVPS's high efficiency lies in its voting mechanism. In the voting process, an edge line has 0, 1, or 2 sections crossing a subspace (see, Fig.8) and has around 1 section crossing any subspace on average. The 2 azimuthal angles of the line section are used to derive which series of the consecutively numbered cells of the subspace should receive the votes (see 20). The time complexity of the voting process is \\( M \\times n \\) where \\( M \\) is the total number of subspaces and \\( n \\) is the number of detected edges of a typical image. Assuming \\( M = 200 \\) and a typical image contains around 500 edges, the computational cost of the voting process is roughly \\( 2 \\times 200 \\times 500 = 200,000 \\) math function calls, which can be processed in around 2 milliseconds by the 3.5GHz CPU we used. Unlike a typical DL approach that generally employs millions of hidden parameters and deals with time-consuming multilayer processing, DiskVPS has none and thus can reach much higher efficiency." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.798, + 0.62, + 0.813 + ], + "angle": 0, + "content": "4. Conclusion" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.826, + 0.892, + 0.903 + ], + "angle": 0, + "content": "By using a more precise voting mechanism than the Gaussian sphere model currently adopted by most of the SOTA, DiskVPS has achieved significantly better performance than the SOTA in detection accuracy and processing speed for real-world images." + }, + { + "type": "page_number", + "bbox": [ + 0.479, + 0.945, + 0.521, + 0.958 + ], + "angle": 0, + "content": "27056" + } + ], + [ + { + "type": "title", + "bbox": [ + 0.08, + 0.09, + 0.174, + 0.105 + ], + "angle": 0, + "content": "References" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.116, + 0.47, + 0.185 + ], + "angle": 0, + "content": "[1] Michel Antunes and Joao P Barreto. A global approach for the detection of vanishing points and mutually orthogonal vanishing directions. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition, pages 1336-1343, 2013. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.187, + 0.47, + 0.215 + ], + "angle": 0, + "content": "[2] Stephen T Barnard. Interpreting perspective images. Artificial intelligence, 21(4):435-462, 1983. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.217, + 0.47, + 0.258 + ], + "angle": 0, + "content": "[3] Bruno Caprile and Vincent Torre. Using vanishing points for camera calibration. International journal of computer vision, 4(2):127-139, 1990. 1, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.261, + 0.471, + 0.329 + ], + "angle": 0, + "content": "[4] Taiyan Chen, Xianghua Ying, Jinfa Yang, Ruibin Wang, Ruohao Guo, Bowei Xing, and Ji Shi. Vpdetr: end-to-end vanishing point detection transformers. In Proceedings of the AAAI Conference on Artificial Intelligence, volume 38, pages 1192-1200, 2024. 1, 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.331, + 0.47, + 0.4 + ], + "angle": 0, + "content": "[5] James M Coughlan and Alan L Yuille. Manhattan world: Compass direction from a single image by bayesian inference. In Proceedings of the seventh IEEE international conference on computer vision, volume 2, pages 941-947. IEEE, 1999. 1, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.402, + 0.47, + 0.445 + ], + "angle": 0, + "content": "[6] James M Coughlan and Alan L Yuille. The Manhattan world assumption: Regularities in scene statistics which enable bayesian inference. In NIPS, volume 2, page 3, 2000. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.446, + 0.47, + 0.502 + ], + "angle": 0, + "content": "[7] Patrick Denis, James H Elder, and Francisco J Estrada. Efficient edge-based methods for estimating Manhattan frames in urban imagery. In European conference on computer vision, pages 197-210. Springer, 2008. 6, 7, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.504, + 0.47, + 0.559 + ], + "angle": 0, + "content": "[8] Martin A Fischler and Robert C Bolles. Random sample consensus: a paradigm for model fitting with applications to image analysis and automated cartography. Communications of the ACM, 24(6):381-395, 1981. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.561, + 0.47, + 0.616 + ], + "angle": 0, + "content": "[9] Paul VC Hough. Machine analysis of bubble chamber pictures. In Proc. of the International Conference on High Energy Accelerators and Instrumentation, Sept. 1959, pages 554-556, 1959. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.618, + 0.47, + 0.687 + ], + "angle": 0, + "content": "[10] Florian Kluger, Hanno Ackermann, Michael Ying Yang, and Bodo Rosenhahn. Deep learning for vanishing point detection using an inverse gnomonic projection. In German Conference on Pattern Recognition, pages 17-28. Springer, 2017. 1, 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.689, + 0.47, + 0.771 + ], + "angle": 0, + "content": "[11] Florian Kluger, Eric Brachmann, Hanno Ackermann, Carsten Rother, Michael Ying Yang, and Bodo Rosenhahn. Consac: Robust multi-model fitting by conditional sample consensus. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 4634-4643, 2020. 1, 2, 7, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.774, + 0.47, + 0.829 + ], + "angle": 0, + "content": "[12] Till Kroeger, Dengxin Dai, and Luc Van Gool. Joint vanishing point extraction and tracking. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition, pages 2449-2457, 2015. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.832, + 0.47, + 0.901 + ], + "angle": 0, + "content": "[13] Seokju Lee, Junsik Kim, Jae Shin Yoon, Seunghak Shin, Oleksandr Bailo, Namil Kim, Tae-Hee Lee, Hyun Seok Hong, Seung-Hoon Han, and In So Kweon. Vpgnet: Vanishing point guided network for lane and road marking detection and recognition. In Proceedings of the IEEE inter" + }, + { + "type": "list", + "bbox": [ + 0.08, + 0.116, + 0.471, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.533, + 0.093, + 0.892, + 0.119 + ], + "angle": 0, + "content": "national conference on computer vision, pages 1947-1955, 2017. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.121, + 0.894, + 0.189 + ], + "angle": 0, + "content": "[14] José Lezama, Rafael Grompone von Gioi, Gregory Randall, and Jean-Michel Morel. Finding vanishing points via point alignments in image primal and dual domains. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition, pages 509-515, 2014. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.191, + 0.892, + 0.245 + ], + "angle": 0, + "content": "[15] David Liebowitz and Andrew Zisserman. Combining scene and auto-calibration constraints. In Proceedings of the Seventh IEEE International Conference on Computer Vision, volume 1, pages 293-300. IEEE, 1999. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.247, + 0.894, + 0.328 + ], + "angle": 0, + "content": "[16] Yancong Lin, Ruben Wiersma, Silvia L. Pintea, Klaus Hildebrandt, Elmar Eisemann, and Jan C. van Gemert. Deep vanishing point detection: Geometric priors make dataset variations vanish. In 2022 IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 6093-6103, 2022. 1, 2, 7, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.33, + 0.892, + 0.399 + ], + "angle": 0, + "content": "[17] Xiaohu Lu, Jian Yaoy, Haoang Li, Yahui Liu, and Xiaofeng Zhang. 2-line exhaustive searching for real-time vanishing point estimation in manhattan world. In 2017 IEEE Winter Conference on Applications of Computer Vision (WACV), pages 345-353. IEEE, 2017. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.4, + 0.892, + 0.455 + ], + "angle": 0, + "content": "[18] Evelyne Lutton, Henri Maitre, and Jaime Lopez-Krahe. Contribution to the determination of vanishing points using hough transform. IEEE transactions on pattern analysis and machine intelligence, 16(4):430-438, 1994. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.455, + 0.892, + 0.509 + ], + "angle": 0, + "content": "[19] Faraz M Mirzaei and Stergios I Roumeliotis. Optimal estimation of vanishing points in a Manhattan world. In 2011 International Conference on Computer Vision, pages 2454-2461. IEEE, 2011. 1, 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.511, + 0.892, + 0.552 + ], + "angle": 0, + "content": "[20] Van Nhan Nguyen, Robert Jenssen, and Davide Roverso. Lsnet: Fast single-shot line-segment detector. Machine Vision and Applications, 32(1):12, 2021. 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.553, + 0.892, + 0.621 + ], + "angle": 0, + "content": "[21] Rémi Pautrat, Shaohui Liu, Petr Hruby, Marc Pollefeys, and Daniel Barath. Vanishing point estimation in uncalibrated images with prior gravity direction. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 14118-14127, 2023. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.622, + 0.892, + 0.663 + ], + "angle": 0, + "content": "[22] Carsten Rother. A new approach to vanishing point detection in architectural environments. Image and Vision Computing, 20(9-10):647-655, 2002. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.664, + 0.892, + 0.734 + ], + "angle": 0, + "content": "[23] Alexander Sheshkus, Anastasia Ingacheva, Vladimir Arlazarov, and Dmitry Nikolaev. Houghnet: neural network architecture for vanishing points detection. In 2019 International Conference on Document Analysis and Recognition (ICDAR), pages 844-849. IEEE, 2019. 1, 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.735, + 0.892, + 0.788 + ], + "angle": 0, + "content": "[24] Jefferey A Shufelt. Performance evaluation and analysis of vanishing point detection techniques. IEEE transactions on pattern analysis and machine intelligence, 21(3):282-288, 1999. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.789, + 0.892, + 0.817 + ], + "angle": 0, + "content": "[25] Sudipta Sinha, Drew Steedly, and Rick Szeliski. Piecewise planar stereo for image-based rendering. 2009. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.818, + 0.892, + 0.871 + ], + "angle": 0, + "content": "[26] Jean-Philippe Tardif. Non-iterative approach for fast and accurate vanishing point detection. In 2009 IEEE 12th International Conference on Computer Vision, pages 1250–1257. IEEE, 2009. 1, 7, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.873, + 0.892, + 0.901 + ], + "angle": 0, + "content": "[27] Xin Tong, Xianghua Ying, Yongjie Shi, Ruibin Wang, and Jinfa Yang. Transformer based line segment classifier with" + }, + { + "type": "list", + "bbox": [ + 0.503, + 0.093, + 0.894, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.48, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "27057" + } + ], + [ + { + "type": "ref_text", + "bbox": [ + 0.108, + 0.093, + 0.47, + 0.147 + ], + "angle": 0, + "content": "image context for real-time vanishing point detection in man-. hhattan world. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 6093- 6102, 2022. 1, 2, 6, 7, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.15, + 0.47, + 0.205 + ], + "angle": 0, + "content": "[28] Elena Tretyak, Olga Barinova, Pushmeet Kohli, and Victor Lempitsky. Geometric image parsing in man-made environments. International Journal of Computer Vision, 97(3):305-321, 2012. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.207, + 0.47, + 0.275 + ], + "angle": 0, + "content": "[29] Tinne Tuytelaars, Luc Van Gool, Marc Proesmans, and Theo Moons. The cascaded hough transform as an aid in aerial image interpretation. In Sixth International Conference on Computer Vision (IEEE Cat. No. 98CH36271), pages 67-72. IEEE, 1998. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.277, + 0.47, + 0.319 + ], + "angle": 0, + "content": "[30] Jonathan Ventura and Tobias Höllerer. Structure and motion in urban environments using upright panoramas. *Virtual Reality*, 17(2):147-156, 2013. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.321, + 0.47, + 0.389 + ], + "angle": 0, + "content": "[31] Rafael Grompone Von Gioi, Jeremie Jakubowicz, Jean-Michel Morel, and Gregory Randall. Lsd: A fast line segment detector with a false detection control. IEEE transactions on pattern analysis and machine intelligence, 32(4):722-732, 2008. 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.391, + 0.47, + 0.433 + ], + "angle": 0, + "content": "[32] Yukun Wang, Shuo Gu, Yinbo Liu, and Hui Kong. Semi-supervised vanishing point detection with contrastive learning. Pattern Recognition, 153:110518, 2024. 1, 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.435, + 0.47, + 0.489 + ], + "angle": 0, + "content": "[33] Horst Wildenauer and Allan Hanbury. Robust camera self-calibration from monocular images of manhattan worlds. In 2012 IEEE Conference on Computer Vision and Pattern Recognition, pages 2831-2838. IEEE, 2012. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.491, + 0.47, + 0.56 + ], + "angle": 0, + "content": "[34] Jianping Wu, Liang Zhang, Ye Liu, and Ke Chen. Real-time vanishing point detector integrating under-parameterized ransac and hough transform. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 3732-3741, 2021. 2, 3, 7, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.562, + 0.47, + 0.631 + ], + "angle": 0, + "content": "[35] Yiliang Xu, Sangmin Oh, and Anthony Hoogs. A minimum error vanishing point detection approach for uncalibrated monocular images of man-made environments. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition, pages 1376-1383, 2013. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.633, + 0.47, + 0.688 + ], + "angle": 0, + "content": "[36] Yifan Xu, Weijian Xu, David Cheung, and Zhuowen Tu. Line segment detection using transformers without edges. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 4257-4266, 2021. 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.689, + 0.47, + 0.757 + ], + "angle": 0, + "content": "[37] Menghua Zhai, Scott Workman, and Nathan Jacobs. Detecting vanishing points using global image context in a non-manhattan world. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition, pages 5657-5665, 2016. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.76, + 0.47, + 0.815 + ], + "angle": 0, + "content": "[38] Lilian Zhang, Huimin Lu, Xiaoping Hu, and Reinhard Koch. Vanishing point estimation and line classification in a Manhattan world with a unifying camera model. International Journal of Computer Vision, 117:111-130, 2016. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.817, + 0.47, + 0.858 + ], + "angle": 0, + "content": "[39] Yichao Zhou, Haozhi Qi, Jingwei Huang, and Yi Ma. Neurvps: Neural vanishing point scanning via conic convolution. arXiv preprint arXiv:1910.06316, 2019. 1, 2, 7, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.86, + 0.47, + 0.901 + ], + "angle": 0, + "content": "[40] Yichao Zhou, Haozhi Qi, Yuexiang Zhai, Qi Sun, Zhili Chen, Li-Yi Wei, and Yi Ma. Learning to reconstruct 3d Manhattan wireframes from a single image. In Proceedings of the" + }, + { + "type": "list", + "bbox": [ + 0.08, + 0.093, + 0.47, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "text", + "bbox": [ + 0.534, + 0.093, + 0.892, + 0.121 + ], + "angle": 0, + "content": "IEEE/CVF International Conference on Computer Vision, pages 7698-7707, 2019. 1, 7, 8" + }, + { + "type": "page_number", + "bbox": [ + 0.48, + 0.946, + 0.519, + 0.957 + ], + "angle": 0, + "content": "27058" + } + ] +] \ No newline at end of file diff --git a/2025/DiskVPS_ Vanishing Point Detector via Hough Transform in a Disk Region/80e52809-79f5-4a63-9312-c7df737fcd10_origin.pdf b/2025/DiskVPS_ Vanishing Point Detector via Hough Transform in a Disk Region/80e52809-79f5-4a63-9312-c7df737fcd10_origin.pdf new file mode 100644 index 0000000000000000000000000000000000000000..0e86761ec2b54f4cb7a0cb4ad9735fa5a982a10b --- /dev/null +++ b/2025/DiskVPS_ Vanishing Point Detector via Hough Transform in a Disk Region/80e52809-79f5-4a63-9312-c7df737fcd10_origin.pdf @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:56ab1afa07fa5e695d91816b72d088a963b2023d7a8f61fffe483803c4071d75 +size 10048869 diff --git a/2025/DiskVPS_ Vanishing Point Detector via Hough Transform in a Disk Region/full.md b/2025/DiskVPS_ Vanishing Point Detector via Hough Transform in a Disk Region/full.md new file mode 100644 index 0000000000000000000000000000000000000000..018e5199f983cde5931b41456b87d0f032f49586 --- /dev/null +++ b/2025/DiskVPS_ Vanishing Point Detector via Hough Transform in a Disk Region/full.md @@ -0,0 +1,385 @@ +# DiskVPS: Vanishing Point Detector via Hough Transform in a Disk Region + +Jianping Wu +Suzhou Vocational University +Suzhou, China 215104 +wujp@jssvc.edu.cn + +# Abstract + +DiskVPS is a novel and robust vanishing point (VP) detection scheme based on Hough Transform (HT) over an image-plane-mapped disk region. In the DiskVPS, the image plane is first mapped into a disk region, which is then partitioned into concentrical ring-shaped subspaces. Each subspace is further partitioned into approximately equal probability Hough cells. By using individual edges rather than edge pairs as voters, DiskVPS can achieve high accuracy with extreme efficiency. Involving no calibration parameters, DiskVPS is particularly suitable for VP detection in uncalibrated images and thus can be applied in image calibration. A comparative experimental study demonstrates that the basic DiskVPS model without parameter optimization achieved significantly better performance over the SOTA in detection accuracy and processing speed with real-world images. The study also shows that DiskVPS is robust against parameter changes. + +# 1. Introduction + +VP detection has a variety of applications, including camera calibration ([1, 3, 5, 15]), 3D reconstruction ([25, 40]), virtual reality [30], robot navigation [12], scene tracking [6] and vehicle automatic driving [13]. Depending on the direction of the parallel lines and the setup of the capturing camera, a VP can appear anywhere, including infinity, in the image plane. A triplet of orthogonal VPs is formed by 3 sets of mutually orthogonal parallel lines in the man-made world commonly called Manhattan world [5]. According to [3], the information of any 2 orthogonal VPs in an image is sufficient to obtain the focal length of the capturing camera. Moreover, a triplet of orthogonal VPs fully determines the camera's intrinsic and extrinsic parameters. + +# 1.1. Related Works + +VP detection has long been a key research topic attracting great attention in computer vision. However, it remains a challenging issue far from soundly resolved. The key is- + +sue in VP detection is the unboundedness of the image plane in which VPs reside. In VP detection, it is essential to map the image plane into some bounded space before practical techniques such as Hough Transform (HT, [9]) can be deployed. + +The first such attempt came from [29] that divided the image plane into 3 bounded subspaces to perform the HT. Since its division method is rather arbitrary without offsetting the probability biases in cell partitioning, it failed to yield satisfactory results. For calibrated images, Barnard [2] proposed a VP detection algorithm that mapped the image plane into a unit sphere called Gaussian sphere (GS) for HT, which is widely used in VP detection ([18, 24]) and is also universally applied in deep learning VP detection algorithms([4, 10, 11, 16, 23, 27, 32, 39]). However, GS-based approaches have limitations. Since building the GS requires the known parameters of the capturing camera such as the focal length and coordinates of the optical axis in the image plane, GS approaches are unsuitable for uncalibrated images, which account for a substantial fraction of images of interest. GS's voting scheme using edge pairs can significantly amplify discretization and distortion errors in edge detection and thus cause a high likelihood of spurious VP detection. Accordingly, GS models are prone to spurious detection ([16, 24, 39]). Moreover, the process of partitioning the GS into Hough cells itself is complicated and delicate. The non-planar shape of the GS is not amenable for partitioning of the GS into planar Hough cells while the Hough cell sizes on the GS are wildly non-uniform and sensitive to the errors of calibration parameters. With the limitations, the GS baseline model did not draw wide attention until the recent rapid development of deep learning technology. + +As an important alternative to HT in VP detection, random sample consensus (RANSAC [8]) is another extensively explored scheme in VP detection ([17, 19, 21, 22, 26, 28, 33, 35, 38]). RANSAC can work directly in the image plane and is suitable for VP detection in uncalibrated images; moreover, RANSAC outperforms GS-based VP detection models in practice with scenes where many edges + +belong to VP inliers. The main drawback of RANSAC is its susceptibility to false negative detection when the size of samples accounts for only a tiny portion of the full set of edge pairs used for VP detection (3 or more edges belonging to a set of 2 or 3 orthogonal VPs are sometimes proposed in [19]). Additionally, RANSAC is computationally expensive, particularly for images of cluttered scenes. Recently, [34] proposed an algorithm that integrated RANSAC and HT to achieve real-time VP detection. There are other methods like using dual spaces ([14]). + +Despite its limitations, GS is the only proven VP detection model available to transform an unbounded space (image plane) into a bounded space (GS), enabling its deployment in deep learning neural networks. Therefore, the GS remains the chosen technique in deep-learning-based VP detection algorithms([4, 10, 11, 16, 23, 27, 32, 37, 39]). While the baseline model of GS-based VP detection is not optimal in both accuracy and processing speed, GS-based deep learning approaches making use of advanced high-level-feature extraction schemes and high-performance hardware have achieved the performance comparable to the SOTA, a level far superior to the GS's baseline model. However, deep learning approaches have limitations. The costs of data annotation and training needed by deep learning are expensive. Additionally, limited datasets usually do not generalize well in deep learning, which can fail in unexpected ways. + +# 1.2. Contribution + +DiskVPS is suitable for conducting highly reliable and real-time VP detection. We use Cartesian coordinates $(u, v)$ to denote the normalized coordinates with the origin set at the image center and both image width and height normalized to 1 (see next section for how $(u, v)$ is mapped from the pixel-based image plane coordinates). DiskVPS maps an image plane (unbounded space) (Fig. 1(a)) into a disk-shaped space (Fig. 1(b)). Specifically, it maps an unbounded polar coordinate system $(r, \theta)$ $(r = \sqrt{u^2 + v^2}$ and $\theta = \tan^{-1} \frac{u}{v}$ into a bounded polar coordinate system $(\tan^{-1}(4r), \theta)$ with the latter's radial boundary located at $\frac{\pi}{2}$ . With the approach, a straight line that passes the origin in the image plane is mapped into a diametrical line segment in the disk. In contrast, a straight line not passing the origin in the image plane is mapped into an arc-shaped segment with 2 ends overlapped with its origin-passing parallel counterpart at the disk boundary. Consequently, parallel lines in the image plane intersect on the disk boundary, whereas non-parallel lines in the image plane intersect each other only once inside the disk. In the scheme, HT is reduced to a probabilistic method to partition the disk into a set of approximately equal-probability Hough cells. Fig. 2 shows a set of edges detected in a typical image (see Fig. 2(a)) and its corresponding Hough accumulation map generated by the + +![](images/dfed0fd5dc4e2f4dad963cf38f9f921999e201678cb0408e3617ca24ea33a895.jpg) +(a) + +![](images/9bea78ddb7cea411f96e05c6197cfbc4716ef2f130ad127ece26ab9fb3a73d7f.jpg) +(b) + +![](images/05fc8c78f9eadcfd2cd1bb6057c271d2b84b5f299744ac83a3fd2e798c3407c0.jpg) +Figure 1. Mapping from an image plane (Cartesian coordinates $(u, v)$ (a)) to a disk-shaped Hough space (polar coordinates $(\tan^{-1}(4r), \theta)$ (b)) with $r = \sqrt{u^2 + v^2}$ and $\theta = \tan^{-1} \frac{v}{u}$ . +(a) + +![](images/7f28b618a869c4a5ed9ad512be42ecabacff4fd00b7155b7b6c6d82921be608d.jpg) +(b) + +![](images/330a8000a1b133d2c18f962ebb4fa0c4c928ca2160f1a1765f0887765e85887a.jpg) +Figure 2. Detected edges of a typical image (a) and its corresponding Hough accumulation map (b). +Figure 3. Intersection $P$ of the extension line of a random edge and a given line defined by the distance $d$ and the angle $\phi$ . + +edges using DiskVPS (see Fig. 2(b)). Five sharp and salient peaks represent 3 mutually orthogonal VPs, 2 far-field VPs generated by nearly parallel edges and one near-field VP generated by non-parallel edges. The proposed approach successfully manifests all valid VPs in the image without yielding any spurious VPs in Fig. 2b. + +# 2. Algorithm Description + +# 2.1. Theoretical Foundation + +Assume $(x,y)$ is the pixel-based coordinates of a point in the image plane and the image of interest is encoded by 2 coordinates in pixels: $(x_{\min},y_{\min})$ and $(x_{\max},y_{\max})$ , which refer to the pixel-based coordinates of the image's bottom left corner and top right corner respectively. To ease analy- + +![](images/3d0d52e89e5c6ced96176238e73654e014726575ed37294b3f10cf1e49cf97ea.jpg) +Figure 4. Illustration of the 2 special directions along which the PDF is considered, the radial and tangential directions. + +sis, $(x,y)$ is transformed to a normalized coordinate system $(u,v)$ . Their relationship is written as: + +$$ +\left\{ \begin{array}{l} u = \frac {x - 0 . 5 \left(x _ {\min } + x _ {\max }\right)}{x _ {\max } - x _ {\min }} \\ v = \frac {y - 0 . 5 \left(y _ {\min } + y _ {\max }\right)}{y _ {\max } - y _ {\min }} \end{array} \right. \tag {1} +$$ + +In the $(u,v)$ system, the origin $O$ is fixed to the image center, and the bottom left corner and top right corner of the image are moved to $(-0.5, -0.5)$ and $(0.5, 0.5)$ , respectively (see Fig. 3). The image region turns into an Origin-centered unit square in the $(u,v)$ system. + +[34] proposed a statistical model to estimate the probabilistic distribution of $P$ , the intersection of a given line and the extension line of a random edge, along the very given line. According to [34], the midpoint of the random edge is distributed uniformly in the unit square, and the direction of the random edge is also uniformly distributed over $[0, \pi]$ . Such a random edge can be used to simulate the statistical distribution of the edges extracted from a sufficiently large number of images. We cite the conclusion of [34] in this paper. As shown in Fig. 3, if the given line is defined by $d$ (the distance from $O$ on the given line) and $\phi$ (the directional angle pointing from $O$ to its perpendicular foot to the given line), then [34] gives: + +$$ +f _ {d, \phi} (t) = \frac {1}{\pi} \int_ {- 0. 5 - 0. 5} ^ {0. 5} \int_ {0} ^ {0. 5} \frac {\left| u \cos \phi + v \sin \phi - d \right| d u d v}{(u - d \cos \phi - t \sin \phi) ^ {2} + (v - d \sin \phi + t \cos \phi) ^ {2}} \tag {2} +$$ + +where $t$ is the coordinate of $P$ on a new axis created along the given line with the origin of the axis set at the perpendicular foot of $O$ on the given line, as shown in Fig. 3. + +Here, we are interested in Eq.2's solutions for 2 special cases. As shown in Fig.4, one is for the case in which the given line is along the radial direction in the $(u,v)$ coordinates, and the other is for the case in which the given line is in the azimuthal or tangential direction. $f_{\mathrm{rad}}(r,\theta)$ and $f_{\mathrm{tag}}(r,\theta)$ are used to denote the PDFs of $P$ along the given line along the radial direction and the given line along the tangential direction respectively, where $\theta$ is the directional angle of $OP$ and $r$ is the length of $OP$ , both in the $(u,v)$ + +coordinate system. Accordingly, we have: + +$$ +\begin{array}{l} f _ {\mathrm {r a d}} (r, \theta) = f _ {0, \theta + \frac {\pi}{2}} (r) \\ = \frac {1}{\pi} \int_ {- 0. 5 - 0. 5} ^ {0. 5} \int_ {0} ^ {0. 5} \frac {\left| - u \sin \theta + v \cos \theta \right| d u d v}{(u - r \cos \theta) ^ {2} + (v - r \sin \theta) ^ {2}} \tag {3} \\ \end{array} +$$ + +$$ +\begin{array}{l} f _ {\mathrm {t a g}} (r, \theta) = f _ {r, \theta + \frac {\pi}{2}} (0) \\ = \frac {1}{\pi} \int_ {- 0. 5} ^ {0. 5} \int_ {- 0. 5} ^ {0. 5} \frac {\left| u \sin \theta - v \cos \theta + r \right| d u d v}{(u + r \sin \theta) ^ {2} + (v - r \cos \theta) ^ {2}} \tag {4} \\ \end{array} +$$ + +Since $f_{\mathrm{rad}}(r,\theta)$ and $f_{\mathrm{tag}}(r,\theta)$ 's dependence on the variable $\theta$ is insignificant (see Fig.5), they are conveniently approximated by 2 less complex and more manageable forms $g_{\mathrm{rad}}(r) = f_{\mathrm{rad}}(r,0)$ and $g_{\mathrm{tag}}(r) = f_{\mathrm{tag}}(r,0)$ to facilitate subsequent construction of Hough space. They are: + +$$ +g _ {\text {r a d}} (r) = \frac {1}{\pi} \int_ {- 0. 5 - 0. 5} ^ {0. 5} \int_ {0} ^ {0. 5} \frac {| v | d u d v}{(u - r) ^ {2} + v ^ {2}} \tag {5} +$$ + +$$ +g _ {\text {a g}} (r) = \frac {1}{\pi} \int_ {- 0. 5 - 0. 5} ^ {0. 5} \int_ {- 0. 5} ^ {0. 5} \frac {| v - r | d u d v}{u ^ {2} + (v - r) ^ {2}} \tag {6} +$$ + +The analytical solutions of $g_{\mathrm{rad}}(r)$ and $g_{\mathrm{tag}}(r)$ are: + +$$ +g _ {\mathrm {r a d}} (r) = \left\{ \begin{array}{l l} \frac {\tan^ {- 1} (2 r + 1) - \tan^ {- 1} (2 r - 1)}{\pi} & \\ + \frac {1}{2 \pi} \ln \frac {[ (2 r - 1) ^ {2} + 1 ] [ (2 r + 1) ^ {2} + 1 ]}{(2 r - 1) ^ {2} (2 r + 1) ^ {2}} & r \neq 0. 5 \\ + \frac {r}{\pi} \ln \frac {(2 r - 1) ^ {2} [ (2 r + 1) ^ {2} + 1 ]}{(2 r + 1) ^ {2} [ (2 r - 1) ^ {2} + 1 ]} & \\ \frac {\tan^ {- 1} 2 + \ln 5 - \ln 4}{\pi} & r = 0. 5 \end{array} \right. \tag {7} +$$ + +$$ +g _ {\text {t a g}} (r) = \left\{ \begin{array}{l l} 1 - \frac {\tan^ {- 1} (2 r + 1) - \tan^ {- 1} (2 r - 1)}{\pi} & r \leq 0. 5 \\ - \frac {2 r \left[ \tan^ {- 1} (2 r + 1) + \tan^ {- 1} (2 r - 1) \right]}{\pi} & \\ + \frac {\ln [ 1 + (2 r + 1) ^ {2} ] + \ln [ 1 + (2 r - 1) ^ {2} ]}{2 \pi} & \\ 1 - \frac {(2 r + 1) \tan^ {- 1} (2 r + 1)}{\pi} & \\ - \frac {(2 r - 1) \tan^ {- 1} (2 r - 1)}{\pi} & r > 0. 5 \\ + \frac {\ln [ 1 + (2 r + 1) ^ {2} ] - \ln [ 1 + (2 r - 1) ^ {2} ]}{2 \pi} & \end{array} \right. \tag {8} +$$ + +As shown in Fig. 5, both $\frac{g_{\mathrm{rad}}(r)}{f_{\mathrm{rad}}(r,\theta)}$ and $\frac{g_{\mathrm{tag}}(r)}{f_{\mathrm{tag}}(r,\theta)}$ lie between 0.88 and 1.13 for $r \leq 1.7$ and they move closer to 1 with the increase of $r$ for $r \geq 1.0$ . Overall, using $g_{\mathrm{rad}}(r)$ to approximate $f_{\mathrm{rad}}(r,\theta)$ and using $g_{\mathrm{tag}}(r)$ to approximate $f_{\mathrm{tag}}(r,\theta)$ do not lead to significant errors. + +![](images/b0c19e335660e103c499d6f7dc6cdb1fff7dc3dbe868829abc0d2f41abb8b4ed.jpg) +(a) + +![](images/19fd581a7f960b722199ec8be233b5a09acf49e9900fdf9d641e6cb28e294cea.jpg) +(b) + +![](images/9df6ead6121bd57f5c9992989bbf3706cea7f665e818bc692db34bc50fc351e2.jpg) +Figure 5. Proximity of $g_{\mathrm{rad}}(r)$ to $f_{\mathrm{rad}}(r,\theta)$ (a) and proximity of $g_{\mathrm{tag}}(r)$ to $f_{\mathrm{tag}}(r,\theta)$ (b). +Figure 6. $\frac{1}{\pi} \frac{4}{(4r)^2 + 1}$ 's proximity to $f_{\mathrm{rad}}(r, \theta)$ for chosen $r$ values. + +The previous analysis provides a theoretical basis for building a Hough space from the $(r,\theta)$ . However, since the radial dimension $(r)$ remains unbounded, it is not feasible to create a Hough space directly from its current form (Eq.7 and 8). + +So it makes sense to transform $r$ to a bounded space and derive the PDF in the new space. Note that when $r \gg 1$ , Eq.5 leads to $g_{\mathrm{rad}}(r) \approx \frac{1}{\pi r^2} \int_{-0.5}^{0.5} \int_{-0.5}^{0.5} |v| du dv = \frac{1}{4\pi r^2}$ . Since $\frac{1}{4\pi r^2} \approx \frac{1}{\pi} \frac{4}{(4r)^2 + 1}$ for $r \gg 1$ , Eq. 5 can be further approximated as: + +$$ +g _ {\text {r a d}} (r) \approx \frac {1}{\pi} \frac {4}{(4 r) ^ {2} + 1} \text {f o r} r \gg 1 \tag {9} +$$ + +Here $g_{\mathrm{rad}}(r)\approx \frac{1}{\pi}\frac{4}{(4r)^2 + 1}$ is chosen over other alternatives because the integration of $\frac{1}{\pi}\frac{4}{(4r)^2 + 1}$ over a full radial line is equal to 1, which matches the probability of a radial line intersecting the extension line of the random edge. Note $\frac{1}{\pi}\int \limits_{r = -\infty}^{r = +\infty}\frac{4dr}{(4r)^2 + 1} = \frac{1}{\pi}\int \limits_{r = -\infty}^{r = +\infty}d\tan^{-1}(4r) = \frac{1}{\pi}\times 2\times \frac{\pi}{2} = 1.$ As shown in Fig.6, when $r\geq 1.5$ Eq.9 is already a good approximation of $f_{\mathrm{rad}}(r,\theta)$ as $\frac{\frac{1}{\pi}\frac{4}{(4r)^2 + 1}}{f_{\mathrm{rad}}(r,\theta)}$ lies consistently between 0.93 and 1.08 for $r\geq 1.5$ + +Eq.9 creates a route to transform $r$ to a bounded space. To achieve the goal, we introduce a new function $\psi(r)$ : + +$$ +\psi (r) = \tan^ {- 1} (4 r) \tag {10} +$$ + +Unlike $r$ , $\psi$ is defined in a bounded space since $\psi \in [0, \frac{\pi}{2}]$ . We denote the PDF along $\psi$ as $h_{\mathrm{rad}}(\psi)$ , which is + +![](images/cf5df42cbcdb12186be01f3c61c73856cc3c5fb4d22b9f1dfb9d5f6d16bcaef5.jpg) +(a) + +![](images/17141d096b5248c572f005a65b8098d255e036018a31aae4b1fb0e4dd5a66a13.jpg) +(b) +Figure 7. Depiction of DiskVPS's Hough space. (a) The structure of the Hough space; (b) The parameters of a Hough cell in the $i$ th subspace. + +related to $g_{\mathrm{rad}}(r)$ through the following formula: + +$$ +h _ {\mathrm {r a d}} (\psi) = g _ {\mathrm {r a d}} (r) \left| \frac {d r}{d \psi} \right| \tag {11} +$$ + +Combining Eq.9, Eq.10 and Eq.11, and noting $\frac{d(\tan^{-1}x)}{dx} = \frac{1}{x^2 + 1}$ , we have: + +$$ +h _ {\mathrm {r a d}} (\psi) \approx \frac {1}{\pi} \frac {4}{(4 r) ^ {2} + 1} \div \frac {4}{(4 r) ^ {2} + 1} = \frac {1}{\pi} \tag {12} +$$ + +Eq.12 indicates that the PDF along $\psi$ is approximately constant $(= \frac{1}{\pi})$ in the space of $(\psi, \theta)$ (for $r \geq 1.5$ ). + +# 2.2. Construction of Hough Space + +Structurally, the Hough space is partitioned into a set of $M$ concentric ring-shaped subspaces, each of which is further partitioned into a set of at most $N$ fan-shaped cells, as shown in Fig.7(a). The principle is to ensure that each subspace has a roughly equal probability of intersecting the extension line of the random edge along the radial direction and each cell has a roughly equal probability of intersecting the line in the tangential direction. + +Our first goal is to obtain each subspace's inner radius and radial span using the principle above. We denote the inner radius and the radial span of the $i$ th subspace as $r_i$ and $\delta r_i$ respectively, where $i = 1,2,\ldots,M$ is the order number of the subspace counting from the innermost outwards. Noting the outer radius of the $i$ th subspace is the inner radius of the $i + 1$ th subspace, there exist the following relations throughout the subspaces: + +$$ +\begin{array}{l} r _ {1} = 0 \\ r _ {2} = r _ {1} + \delta r _ {1} \\ \dots \tag {13} \\ r _ {i + 1} = r _ {i} + \delta r _ {i} \\ \dots \\ \end{array} +$$ + +Since a full radial line intersects the extension line of the random edge exactly once (the special case where the 2 + +lines are parallel and do not intersect has infinitesimal probability and is therefore ignored), the probability of the radial span of each subspace intersecting the extension line of the random edge is $\frac{1}{M}$ . + +In addition to setting the inner radius and the radial span of the $i$ th subspace as $r_i$ and $\delta r_i$ for near-field subspaces, we also set the inner radius and the radial span of the subspace as $\psi_i = \tan^{-1}(4r_i)$ and $\delta \psi_i$ for far-field subspaces. Assuming that the PDF along the radial span of the $i$ th space is constant ( $= g_{\mathrm{rad}(r_i)}$ for near-field subspace and $= h_{\mathrm{rad}}(\psi_i)$ for far-field subspace), then the probability $\frac{1}{M}$ must match with $g_{\mathrm{rad}}(r_i)\delta r_i$ if the $i$ th subspace is in the near field or $h_{\mathrm{rad}}(\psi_i)\delta \psi_i$ if the $i$ th subspace is in the far field. Since $h_{\mathrm{rad}}(\psi) \approx \frac{1}{\pi}$ (see Eq.12), we have: + +$$ +\delta r _ {i} = \frac {1}{M g _ {\mathrm {r a d}} \left(r _ {i}\right)} \quad \text {f o r} r _ {i} \leq 1. 5 \tag {14} +$$ + +$$ +\delta \psi_ {i} = \frac {\pi}{M} \quad \text {f o r} r _ {i} > 1. 5 +$$ + +Combining $\delta \psi_{i} = \psi_{i + 1} - \psi_{i} = \tan^{-1}(4r_{i + 1}) - \tan^{-1}(4r_{i})$ and $\delta r_i = r_{i + 1} - r_i$ , the 2nd term of Eq.14 turns into $\delta r_{i} = \frac{\tan(\frac{\pi}{M} + \tan^{-1}4r_{i})}{4} -r_{i}$ . And Eq.14 can be rewritten as: + +$$ +\delta r _ {i} = \left\{ \begin{array}{l l} \frac {1}{M g _ {\text {r a d}} \left(r _ {i}\right)} & \text {f o r} r _ {i} \leq 1. 5 \\ \frac {\tan \left(\frac {\pi}{M} + \tan^ {- 1} 4 r _ {i}\right)}{4} - r _ {i} & \text {f o r} r _ {i} > 1. 5 \end{array} \right. \tag {15} +$$ + +The Hough space partitioning starts from $i = 1$ . Eq.13 and Eq.15 work together to obtain the inner radius $r_i$ and the radial span $\delta r_i$ of the $i$ th subspace iteratively ( $i = i + 1$ after each iteration) until the $i$ th subspace comes to the Hough space boundary that is indicated by $\psi_{i+1} \geq \frac{\pi}{2}$ or $\tan^{-1}(4r_i) + \frac{\pi}{M} \geq \frac{\pi}{2}$ . The final count of the subspaces is denoted as $M^* = i$ at the end of the last iteration. Since the PDF approximations are used in the Hough space partitioning, $M^*$ is generally not equal, but very close, to $M$ . + +Afterward, each subspace is partitioned into multiple equal-angle-span cells. For the $i$ th subspace, the angle span of a cell is denoted as $\delta \theta_{i}$ ( $i = 1,2,\ldots M^{*}$ ), as shown in Fig.7(b). We observe that the inner circle of the outermost (the $M^{*}$ th) subspace intersects the extension line of the random edge exactly twice. In the meantime, the very circle is partitioned into $N$ equal-angle-span arcs, each of which forms the inner arc of a cell. Consequently, the probability of the inner arc of a cell intersecting the random edge's extension line is $\frac{2}{N}$ . To make sure that each cell in any subspace gets the same probability $\frac{2}{N}$ of intersecting the random edge's extension line, this probability must match the probability obtained through the product of the cell's inner arc length ( $= \delta \theta_{i}r_{i}$ ) and the PDF along the arc ( $= g_{\mathrm{tag}}(r_i)$ ) throughout the entire Hough space. Accordingly, we have: + +$$ +\delta \theta_ {i} = \frac {2}{N r _ {i} g _ {\text {t a g}} \left(r _ {i}\right)} \tag {16} +$$ + +![](images/7bb5ef4a48a765b106c5788f8f1a7a43d46679b255cc6bc5031e7b72b91be9a3.jpg) +Figure 8. A typical scenario of a voting edge casting its votes in the $i$ th subspace with 1 or 0 inside each cell referring to the number of votes each cell receives. + +The number of cells the $i$ th subspace is partitioned into is denoted as $n_i (i = 1, 2, \ldots, M^*)$ . It is written as: + +$$ +n _ {i} = \operatorname {r o u n d} \left(\frac {2 \pi}{\delta \theta_ {i}}\right) \tag {17} +$$ + +where $2\pi$ is the angle span of the $i$ th subspace, and $\delta \theta_{i}$ is the angle span of a cell in the $i$ th subspace, and round refers to the namesake maths function. + +Finally, $\delta \theta_{i}$ needs to be slightly modified due to $n_i$ 's discrete nature. Its adjusted form, $\delta \theta_{i}^{*}$ , is written as: + +$$ +\delta \theta_ {i} ^ {*} = \frac {2 \pi}{n _ {i}} \tag {18} +$$ + +This way, the $i$ th subspace is partitioned into $n_i$ equal-angle-span cells, each covering an angle span of $\delta \theta_i^*$ . Via iterating $i$ from 1 through $M^*$ , all subspaces are thus partitioned into approximately equally probabilistic cells in terms of intersecting the extension line of the random edge. + +# 2.3. Implementation of Hough Transform + +A Hough accumulation map, each of whose elements is mapped from a corresponding Hough cell, is constructed and the magnitude of each element is initialized to zero. + +The VP detection consists of a voting process and an iterative VP detection process. + +In the voting process, each edge casts its votes based on whether its extension line intersects any of the subspaces, and if yes then which specific cell(s) of the subspace the line intersects, with each intersecting cell receiving exactly 1 vote. Assume the extension line of the voting edge is defined by $d$ , the distance from $O$ to the line, and $\varphi$ , the azimuthal angle of $O$ 's perpendicular foot on the line. For an edge's voting in the $i$ th subspace, there exist 3 cases, namely $d < r_i$ , $r_i \leq d < r_i + \delta r_i$ , and $d \geq r_i + \delta r_i$ respectively, where $r_i$ and $\delta r_i$ are the inner radius and radial span of the $i$ th subspace. + +A scenario of the 1st case ( $d < r_i$ ) is shown in Fig.8. The extension line of the voting edge intersects the $i$ th subspace boundary at 4 points, namely $P_1, P_2, P_3$ , and $P_4$ . Assume $\Psi_{\overrightarrow{OP_1}}$ , $\Psi_{\overrightarrow{OP_2}}$ , $\Psi_{\overrightarrow{OP_3}}$ , and $\Psi_{\overrightarrow{OP_4}}$ are the azimuthal angles of vectors $\overrightarrow{OP_1}$ , $\overrightarrow{OP_2}$ , $\overrightarrow{OP_3}$ and $\overrightarrow{OP_4}$ , respectively. Then we have: + +$$ +\left\{ \begin{array}{l} \Psi_ {\overrightarrow {O P _ {1}}} = \varphi + \cos^ {- 1} \frac {d}{r _ {i} + \delta r _ {i}} \\ \Psi_ {\overrightarrow {O P _ {2}}} = \varphi + \cos^ {- 1} \frac {d}{r _ {i}} \\ \Psi_ {\overrightarrow {O P _ {3}}} = \varphi - \cos^ {- 1} \frac {d}{r _ {i}} \\ \Psi_ {\overrightarrow {O P _ {4}}} = \varphi - \cos^ {- 1} \frac {d}{r _ {i} + \delta r _ {i}} \end{array} \right. \tag {19} +$$ + +As shown in Fig.8, $P_{1}P_{2}$ and $P_{3}P_{4}$ are the line's 2 sections intersecting the $i$ th subspace. Assume the $n_{i}$ cells of the $i$ th subspace are labeled 1 through $n_{i}$ and a set of consecutive cells that intersect $P_{1}P_{2}$ have the labels numbered $j$ through $k$ , then $j$ and $k$ are written as: + +$$ +\left\{\begin{array}{l}j = \left\lceil \frac {\Psi_ {\overrightarrow {O P _ {2}}}}{\delta \theta_ {i} ^ {*}} \right\rceil\\k = \left\lceil \frac {\Psi_ {\overrightarrow {O P _ {1}}}}{\delta \theta_ {i} ^ {*}} \right\rceil\end{array}\right. \tag {20} +$$ + +where $\delta \theta_{i}^{*}$ is defined in Eq.18, and $\lceil \rceil$ refers to maths function ceil. + +Similarly, we can obtain the set of consecutive cells intersecting $P_{3}P_{4}$ . In voting, each cell in the two sets receives 1 vote from the voting edge while the rest of the $n_{i}$ cells receive no votes. + +For the 2nd case, voting is performed similarly to the first, but only the angle span between $\Psi_{\overrightarrow{OP_1}}$ and $\Psi_{\overrightarrow{OP_4}}$ needs to be handled as neither $P_{2}$ nor $P_{3}$ exists. + +For the 3rd case, the subspace does not intersect the line; thus, no cell receives any vote. + +The voting process generally yields a Hough accumulation map with a small number of salient peaks scattered over the map similar to Fig.2(b). + +Then, VP detection is as simple as finding salient peaks in the Hough accumulation map. In practice, at most $N_{\mathrm{VP}}$ dominant VPs are detected via at most $N_{\mathrm{VP}}$ iterations with each iteration extracting exactly 1 VP. In each iteration, the following steps are taken: + +Step 1: Search the global peak in the current Hough accumulation space. + +Step 2: Obtain the location of the peak and set it as the initial position of the current VP. + +Step 3: Search all of the inlier edges of the current VP. An edge is regarded as an inlier of the VP only if the angle between its extension line and the line that passes both the midpoint of the edge and the VP is no more than $\Theta$ . + +Step 4: Update the Hough accumulation map by removing all of the votes cast by the inliers of the current VP. This ensures that inliers of the dominant VPs will not affect the subsequent detection of less dominant VPs. + +Step 5: If the peak value in the updated accumulation map is at least 3 times as much as the average and the total + +![](images/54e99d8984c49e7841feb5ac62f320c5f60e03c8ef77bb0a0b7a3b566ee561b7.jpg) + +![](images/0d4cfc94c3f34d16d19216fbbd291954c62918572ea9cdad60cbba5d06706227.jpg) + +![](images/69e503b3f0d39b524e2cc06fc27f50b365571dd4ecc2d7494726e0e3745f08fe.jpg) + +![](images/f5e6e581506cd4cd1c9f95e693bf9cdbceec2df7c1ddb26fe8ab5051e7428679.jpg) +(a) +(f) +Figure 9. Graphical demonstration of the steps in DiskVPS' VP detection: (a) is the original image; (b) is the detected edge set; (c) is the initial Hough accumulation colormap with its global peak (also the position of the 1st VP) marked in red; (d) is the updated Hough accumulation map after removing the votes cast by the inliers of the 1st VP with 1 detected VP and the updated peak marked in red; (e) is the updated Hough accumulation map after the 2nd iteration with 2 detected VPs and the updated peak marked in red; (f) is the updated Hough accumulation map after the 3rd iteration with 3 detected VPs marked in red. + +![](images/95cda287321aac94349ba9684fdd23d9d505f5fabb1d6cb7701238de7b921853.jpg) +(b) +(e) + +![](images/22938e7e0ac996026b6d9b6a8fb5c6f96270ca47eaf70c5ed3930b8648ec7f24.jpg) +(c) +(d) + +number of detected VPs is less than $N_{\mathrm{VP}}$ , then go to Step 1. Otherwise, stop iteration and output all detected VPs. + +The VP detection process is demonstrated in Fig.9. The zenith, which is the convergent point of nearly vertical edges, is located at the local maximum around the south pole of the disk-shaped colormap and the inliers of the zenith also lead to the significantly enhanced distribution around the north pole; the $1^{st}$ horizontal VP at the east pole is formed by nearly parallel edges which also contribute to the enhanced accumulation around the west pole; non-parallel edges form the $2^{nd}$ horizontal VP and generate a corresponding peak in the interior of the map. Some local maxima are noises caused by combined voting of inliers of different VPs and outliers of the VPs. However, after 3 iterations almost all local maxima unrelated to any valid VPs in the original accumulation map disappear (see Fig. 9(f)). It demonstrates that the vote removal process of the inliers of the detected VPs in Step 4 effectively minimizes false positive detection. + +# 3. Experiment + +The proposed approach is examined and evaluated against the state-of-the-art (SOTA) in VP detection. It was implemented in $\mathrm{C + + }$ and tested on a PC powered by an Intel i7 CPU (3.5 GHZ). + +DiskVPS was evaluated via 3 benchmark datasets, the York Urban Dataset (YUD [7]), SVVP([27]), and + +![](images/be74e2064883f1b56355d02d5b697024eff9531bc97f4d51040ada2f1512dbf1.jpg) +Figure 10. Example images from YUD [7] (the 1st row), SVVP[27] (the 3rd row), and SU3 [40] (the 5th row) datasets with each even row immediately below displaying their generated disk-shaped Hough colormaps and the VP detection results. The ground truth VPs are marked with white "+" and the detected VPs are marked with red "×". + +![](images/0356ac0da5725a5538177d36637d26ff7f5fd4115e2bb9bf291d10d95c7af390.jpg) + +![](images/d8db0a2de0ab0692331d601d0eb65bb959c8985fdbe37521994cdc660101da9e.jpg) + +![](images/02588881492730d419bc21fa1d800f353bf60745ac3a642168258aa836c4e5be.jpg) + +SU3([40]). YUD contains 102 calibrated real-world images of $640 \times 480$ pixels. SVVP contains 500 calibrated real-world images of $512 \times 512$ pixels, and SU3 contains a photorealistic dataset containing 23,000 calibrated synthetic outdoor images of $512 \times 512$ pixels. The commonness of the 3 benchmark datasets is that they all satisfy the Manhattanworld assumption [5] as there exists a triplet of orthogonal VPs in each image. So our experiment is focused on how accurately DiskVPS performs VP detections on the triplets of orthogonal VPs in each dataset compared to the SOTA. Please note that while the images used in the experiment were calibrated, DiskVPS treated those images as uncalibrated and performed VP detection without using any calibration info until checking accuracy against the ground truth. + +DiskVPS uses edges as input. In preprocessing, there are many algorithms to perform edge detection (see LSD([31]), [20], and [36], etc.). A similar algorithm to LSD was selected to perform edge detection. Table1 lists the 4 parameters and their respective magnitude that our implementation used for evaluation. $M$ is the number of concentric subspaces that the disk-shaped Hough space is roughly parti + +tioned into, $N$ is the maximum number of roughly equal-probability cells that each of the subspaces is partitioned into, $N_{\mathrm{VP}}$ is the maximum number of dominant VPs that are initially extracted from each image, and $\Theta$ is the criterion angle to separate the inliers of a VP from its outliers. The parameters are fixed throughout the experiment despite the differences among the 3 datasets. + +Fig.10 lists some example images of the 3 datasets and their VP detection results of our approach. It demonstrates that the disk-shaped Hough accumulation map not only manifests almost all valid VPs but also generates no noticeable spurious VPs. Most detected VPs are either very close to or precisely overlapped with their ground truth counterparts. The only exception happens in the 6th row and the 3rd column, in which the zenith's ground truth lies around the north pole (corresponding to $(0, +\infty)$ in $(u, v)$ coordinates) while its detected location is around the south pole $(0, -\infty)$ in $(u, v)$ . The discrepancy is reconciled as the VP at $(0, +\infty)$ and the VP at $(0, -\infty)$ in the $(u, v)$ coordinates are formed by the same group of vertical edges. + +In the experiment, DiskVPS is compared with six SOTA algorithms ([11, 16, 26, 27, 34, 39]) using a protocol adopted by [11, 16, 27]. In the protocol, a triple of orthogonal VPs are detected from each image, then 3 angular differences in the camera space between the 3 detected VPs and their respective ground truth are obtained, with the largest of the 3 differences adopted as the angular difference of the image. According to the protocol, the percentage of a dataset's images that have a smaller angular difference than a given threshold is used as the angular accuracy (AA) criterion. In the experiment, AA@3°, AA@5°, and AA@10° represent AA thresholds of 3°, 5°, and 10° respectively. For example, an algorithm achieving 80% on AA@5° with a dataset means that the algorithm achieves an angular difference of 5° or less on 80% of the images in the dataset. + +In DiskVPS implementation, a triplet of orthogonal VPs is picked from the initial $N_{\mathrm{VP}}$ VPs based on their mutual orthogonality (see [3] for details of VP orthogonality). However, there is a small fraction of images in which only 2 of the 3 orthogonal VPs are captured in the initial $N_{\mathrm{VP}}$ VPs. This happens when more dominant non-orthogonal VPs suppress a less salient orthogonal VP and thus miss the initial detection. In the latter case, DiskVPS detects the 3rd orthogonal VP by searching for the local maximum in + +
NameMagnitude
M200
N200
NVP4
Θ3°
+ +Table 1. Parameters used in the implementation of our algorithm. + +
Datasets MetricsFPSYUD [7]SVVP[27]SU3 [40]
AA@3°AA@5°AA@10°AA@3°AA@5°AA@10°AA@3°AA@5°
J-Linkage[26]1.060.871.881.527.343.260.882.087.2
Wu et al. [34]3958.373.482.243.860.175.782.387.6
CONSAC [11]3.061.773.384.443.856.569.486.290.4
NeurVPS [39]0.560.871.881.532.845.760.293.996.3
Lin et al. [16]2360.774.386.351.868.782.284.490.6
Tong et al. [27]2565.577.187.451.667.782.691.394.6
Ours (DiskVPS)15874.590.299.058.381.698.484.391.7
+ +the Hough space around the 3rd VP's predicted location derived from the 2 known orthogonal VPs. In the experiment, it was observed that when $N_{\mathrm{VP}} = 4$ , DiskVPS successfully captured a triple of orthogonal VPs in the initial $N_{\mathrm{VP}}$ VPs in 90 (88% of the total) images in YUD, while it captured 2 of the triplet of orthogonal VPs in 11 (11% of the total) images of YUD in the initial $N_{\mathrm{VP}}$ . After executing the 3rd VP detection on the 11 images, 101 of the 102 images in YUD have their respective triples of orthogonal VPs successfully detected. + +The comparison results are shown in Table 2. Our approach achieves $74.5\%$ at AA@3° on YUD, $9\%$ higher than the second best at $65.5\%$ achieved by [27], a significant improvement over the SOTA. Similar patterns are observed at AA@5° and AA@10° on YUD and SVVP. Notably, our approach achieves $99.0\%$ on YUD and $98.4\%$ on SVVP at AA@10° whereas none of any other approaches reaches $90\%$ or higher on this respect. Overall, for real-world images, our approach significantly outperforms the SOTA as far as angular accuracy is concerned. + +On the synthetic SU3, DiskVPS is also competitive. Note that the number of images of SU3 is roughly 40 times as large as YUD and SVD combined and the two approaches ([27, 39]) that outperformed DiskVPS are both deep-learning (DL) methods. We believe the large size of SU3 provides substantially more training and helps significantly optimize the DL algorithms. Considering that DiskVPS is a non-optimized basic VP detection model involving no calibration information which the DL models require, the accuracy of DiskVPS is still impressive. + +Moreover, we increased $M$ from 200 to 300 and $N$ from 100 to 180 (both at stepsize of 5) and observed the angular accuracy (AA) distribution at AA@ $3^{\circ}$ , AA@ $5^{\circ}$ , and AA@ $10^{\circ}$ . The results are shown in Table 3. Since the standard deviations (STDEV) are consistently below $3\%$ of their respective means, it is evident that DiskVPS is reasonably robust against changing parameters. + +Lastly, we also compare our approach's efficiency with + +Table 2. Processing speed and angular accuracy (AA) comparison on YUD( [7]), SVVP([27]) and SU3( [40]). Our algorithm achieves by far the best average processing speed on the 3 datasets. Moreover, our approach significantly outperforms all other approaches on real-world YUD and SVVP, and is competitive on synthetic SU3. + +
CriterionMinimumMaximumMeanSTDEV
AA@3°70.678.475.32.2
AA@5°89.294.191.51.3
AA@10°99.099.099.00
+ +Table 3. Angular accuracy statistics over $200\leq M\leq 300$ and $100\le N\le 180$ + +other algorithms. Our approach achieves 158 frames per second (fps) or 6.3 milliseconds per frame excluding image decoding and edge detection. It is around 4 times as fast as the second best at 39 fps achieved by [34]. DiskVPS's high efficiency lies in its voting mechanism. In the voting process, an edge line has 0, 1, or 2 sections crossing a subspace (see, Fig.8) and has around 1 section crossing any subspace on average. The 2 azimuthal angles of the line section are used to derive which series of the consecutively numbered cells of the subspace should receive the votes (see 20). The time complexity of the voting process is $M \times n$ where $M$ is the total number of subspaces and $n$ is the number of detected edges of a typical image. Assuming $M = 200$ and a typical image contains around 500 edges, the computational cost of the voting process is roughly $2 \times 200 \times 500 = 200,000$ math function calls, which can be processed in around 2 milliseconds by the 3.5GHz CPU we used. Unlike a typical DL approach that generally employs millions of hidden parameters and deals with time-consuming multilayer processing, DiskVPS has none and thus can reach much higher efficiency. + +# 4. Conclusion + +By using a more precise voting mechanism than the Gaussian sphere model currently adopted by most of the SOTA, DiskVPS has achieved significantly better performance than the SOTA in detection accuracy and processing speed for real-world images. + +# References + +[1] Michel Antunes and Joao P Barreto. A global approach for the detection of vanishing points and mutually orthogonal vanishing directions. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition, pages 1336-1343, 2013. 1 +[2] Stephen T Barnard. Interpreting perspective images. Artificial intelligence, 21(4):435-462, 1983. 1 +[3] Bruno Caprile and Vincent Torre. Using vanishing points for camera calibration. International journal of computer vision, 4(2):127-139, 1990. 1, 7 +[4] Taiyan Chen, Xianghua Ying, Jinfa Yang, Ruibin Wang, Ruohao Guo, Bowei Xing, and Ji Shi. Vpdetr: end-to-end vanishing point detection transformers. In Proceedings of the AAAI Conference on Artificial Intelligence, volume 38, pages 1192-1200, 2024. 1, 2 +[5] James M Coughlan and Alan L Yuille. Manhattan world: Compass direction from a single image by bayesian inference. In Proceedings of the seventh IEEE international conference on computer vision, volume 2, pages 941-947. IEEE, 1999. 1, 7 +[6] James M Coughlan and Alan L Yuille. The Manhattan world assumption: Regularities in scene statistics which enable bayesian inference. In NIPS, volume 2, page 3, 2000. 1 +[7] Patrick Denis, James H Elder, and Francisco J Estrada. Efficient edge-based methods for estimating Manhattan frames in urban imagery. In European conference on computer vision, pages 197-210. Springer, 2008. 6, 7, 8 +[8] Martin A Fischler and Robert C Bolles. Random sample consensus: a paradigm for model fitting with applications to image analysis and automated cartography. Communications of the ACM, 24(6):381-395, 1981. 1 +[9] Paul VC Hough. Machine analysis of bubble chamber pictures. In Proc. of the International Conference on High Energy Accelerators and Instrumentation, Sept. 1959, pages 554-556, 1959. 1 +[10] Florian Kluger, Hanno Ackermann, Michael Ying Yang, and Bodo Rosenhahn. Deep learning for vanishing point detection using an inverse gnomonic projection. In German Conference on Pattern Recognition, pages 17-28. Springer, 2017. 1, 2 +[11] Florian Kluger, Eric Brachmann, Hanno Ackermann, Carsten Rother, Michael Ying Yang, and Bodo Rosenhahn. Consac: Robust multi-model fitting by conditional sample consensus. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 4634-4643, 2020. 1, 2, 7, 8 +[12] Till Kroeger, Dengxin Dai, and Luc Van Gool. Joint vanishing point extraction and tracking. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition, pages 2449-2457, 2015. 1 +[13] Seokju Lee, Junsik Kim, Jae Shin Yoon, Seunghak Shin, Oleksandr Bailo, Namil Kim, Tae-Hee Lee, Hyun Seok Hong, Seung-Hoon Han, and In So Kweon. Vpgnet: Vanishing point guided network for lane and road marking detection and recognition. In Proceedings of the IEEE inter + +national conference on computer vision, pages 1947-1955, 2017. 1 +[14] José Lezama, Rafael Grompone von Gioi, Gregory Randall, and Jean-Michel Morel. Finding vanishing points via point alignments in image primal and dual domains. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition, pages 509-515, 2014. 2 +[15] David Liebowitz and Andrew Zisserman. Combining scene and auto-calibration constraints. In Proceedings of the Seventh IEEE International Conference on Computer Vision, volume 1, pages 293-300. IEEE, 1999. 1 +[16] Yancong Lin, Ruben Wiersma, Silvia L. Pintea, Klaus Hildebrandt, Elmar Eisemann, and Jan C. van Gemert. Deep vanishing point detection: Geometric priors make dataset variations vanish. In 2022 IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 6093-6103, 2022. 1, 2, 7, 8 +[17] Xiaohu Lu, Jian Yaoy, Haoang Li, Yahui Liu, and Xiaofeng Zhang. 2-line exhaustive searching for real-time vanishing point estimation in manhattan world. In 2017 IEEE Winter Conference on Applications of Computer Vision (WACV), pages 345-353. IEEE, 2017. 1 +[18] Evelyne Lutton, Henri Maitre, and Jaime Lopez-Krahe. Contribution to the determination of vanishing points using hough transform. IEEE transactions on pattern analysis and machine intelligence, 16(4):430-438, 1994. 1 +[19] Faraz M Mirzaei and Stergios I Roumeliotis. Optimal estimation of vanishing points in a Manhattan world. In 2011 International Conference on Computer Vision, pages 2454-2461. IEEE, 2011. 1, 2 +[20] Van Nhan Nguyen, Robert Jenssen, and Davide Roverso. Lsnet: Fast single-shot line-segment detector. Machine Vision and Applications, 32(1):12, 2021. 7 +[21] Rémi Pautrat, Shaohui Liu, Petr Hruby, Marc Pollefeys, and Daniel Barath. Vanishing point estimation in uncalibrated images with prior gravity direction. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 14118-14127, 2023. 1 +[22] Carsten Rother. A new approach to vanishing point detection in architectural environments. Image and Vision Computing, 20(9-10):647-655, 2002. 1 +[23] Alexander Sheshkus, Anastasia Ingacheva, Vladimir Arlazarov, and Dmitry Nikolaev. Houghnet: neural network architecture for vanishing points detection. In 2019 International Conference on Document Analysis and Recognition (ICDAR), pages 844-849. IEEE, 2019. 1, 2 +[24] Jefferey A Shufelt. Performance evaluation and analysis of vanishing point detection techniques. IEEE transactions on pattern analysis and machine intelligence, 21(3):282-288, 1999. 1 +[25] Sudipta Sinha, Drew Steedly, and Rick Szeliski. Piecewise planar stereo for image-based rendering. 2009. 1 +[26] Jean-Philippe Tardif. Non-iterative approach for fast and accurate vanishing point detection. In 2009 IEEE 12th International Conference on Computer Vision, pages 1250–1257. IEEE, 2009. 1, 7, 8 +[27] Xin Tong, Xianghua Ying, Yongjie Shi, Ruibin Wang, and Jinfa Yang. Transformer based line segment classifier with + +image context for real-time vanishing point detection in man-. hhattan world. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 6093- 6102, 2022. 1, 2, 6, 7, 8 +[28] Elena Tretyak, Olga Barinova, Pushmeet Kohli, and Victor Lempitsky. Geometric image parsing in man-made environments. International Journal of Computer Vision, 97(3):305-321, 2012. 1 +[29] Tinne Tuytelaars, Luc Van Gool, Marc Proesmans, and Theo Moons. The cascaded hough transform as an aid in aerial image interpretation. In Sixth International Conference on Computer Vision (IEEE Cat. No. 98CH36271), pages 67-72. IEEE, 1998. 1 +[30] Jonathan Ventura and Tobias Höllerer. Structure and motion in urban environments using upright panoramas. *Virtual Reality*, 17(2):147-156, 2013. 1 +[31] Rafael Grompone Von Gioi, Jeremie Jakubowicz, Jean-Michel Morel, and Gregory Randall. Lsd: A fast line segment detector with a false detection control. IEEE transactions on pattern analysis and machine intelligence, 32(4):722-732, 2008. 7 +[32] Yukun Wang, Shuo Gu, Yinbo Liu, and Hui Kong. Semi-supervised vanishing point detection with contrastive learning. Pattern Recognition, 153:110518, 2024. 1, 2 +[33] Horst Wildenauer and Allan Hanbury. Robust camera self-calibration from monocular images of manhattan worlds. In 2012 IEEE Conference on Computer Vision and Pattern Recognition, pages 2831-2838. IEEE, 2012. 1 +[34] Jianping Wu, Liang Zhang, Ye Liu, and Ke Chen. Real-time vanishing point detector integrating under-parameterized ransac and hough transform. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 3732-3741, 2021. 2, 3, 7, 8 +[35] Yiliang Xu, Sangmin Oh, and Anthony Hoogs. A minimum error vanishing point detection approach for uncalibrated monocular images of man-made environments. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition, pages 1376-1383, 2013. 1 +[36] Yifan Xu, Weijian Xu, David Cheung, and Zhuowen Tu. Line segment detection using transformers without edges. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 4257-4266, 2021. 7 +[37] Menghua Zhai, Scott Workman, and Nathan Jacobs. Detecting vanishing points using global image context in a non-manhattan world. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition, pages 5657-5665, 2016. 2 +[38] Lilian Zhang, Huimin Lu, Xiaoping Hu, and Reinhard Koch. Vanishing point estimation and line classification in a Manhattan world with a unifying camera model. International Journal of Computer Vision, 117:111-130, 2016. 1 +[39] Yichao Zhou, Haozhi Qi, Jingwei Huang, and Yi Ma. Neurvps: Neural vanishing point scanning via conic convolution. arXiv preprint arXiv:1910.06316, 2019. 1, 2, 7, 8 +[40] Yichao Zhou, Haozhi Qi, Yuexiang Zhai, Qi Sun, Zhili Chen, Li-Yi Wei, and Yi Ma. Learning to reconstruct 3d Manhattan wireframes from a single image. In Proceedings of the + +IEEE/CVF International Conference on Computer Vision, pages 7698-7707, 2019. 1, 7, 8 \ No newline at end of file diff --git a/2025/DiskVPS_ Vanishing Point Detector via Hough Transform in a Disk Region/images.zip b/2025/DiskVPS_ Vanishing Point Detector via Hough Transform in a Disk Region/images.zip new file mode 100644 index 0000000000000000000000000000000000000000..f026e750b6d30e223e67d69459c2a2a3e27c58a7 --- /dev/null +++ b/2025/DiskVPS_ Vanishing Point Detector via Hough Transform in a Disk Region/images.zip @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:60fb5b8a4df8ab05752b7f61bec4265ac91db401c40eae8efbcc6065e9aab75b +size 513152 diff --git a/2025/DiskVPS_ Vanishing Point Detector via Hough Transform in a Disk Region/layout.json b/2025/DiskVPS_ Vanishing Point Detector via Hough Transform in a Disk Region/layout.json new file mode 100644 index 0000000000000000000000000000000000000000..492b13617720706694b0764761f2839c5211879d --- /dev/null +++ b/2025/DiskVPS_ Vanishing Point Detector via Hough Transform in a Disk Region/layout.json @@ -0,0 +1,12324 @@ +{ + "pdf_info": [ + { + "para_blocks": [ + { + "bbox": [ + 67, + 102, + 527, + 121 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 67, + 102, + 527, + 121 + ], + "spans": [ + { + "bbox": [ + 67, + 102, + 527, + 121 + ], + "type": "text", + "content": "DiskVPS: Vanishing Point Detector via Hough Transform in a Disk Region" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 223, + 144, + 370, + 200 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 223, + 144, + 370, + 200 + ], + "spans": [ + { + "bbox": [ + 223, + 144, + 370, + 200 + ], + "type": "text", + "content": "Jianping Wu \nSuzhou Vocational University \nSuzhou, China 215104 \nwujp@jssvc.edu.cn" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 143, + 227, + 192, + 239 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 143, + 227, + 192, + 239 + ], + "spans": [ + { + "bbox": [ + 143, + 227, + 192, + 239 + ], + "type": "text", + "content": "Abstract" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 46, + 251, + 289, + 456 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 251, + 289, + 456 + ], + "spans": [ + { + "bbox": [ + 46, + 251, + 289, + 456 + ], + "type": "text", + "content": "DiskVPS is a novel and robust vanishing point (VP) detection scheme based on Hough Transform (HT) over an image-plane-mapped disk region. In the DiskVPS, the image plane is first mapped into a disk region, which is then partitioned into concentrical ring-shaped subspaces. Each subspace is further partitioned into approximately equal probability Hough cells. By using individual edges rather than edge pairs as voters, DiskVPS can achieve high accuracy with extreme efficiency. Involving no calibration parameters, DiskVPS is particularly suitable for VP detection in uncalibrated images and thus can be applied in image calibration. A comparative experimental study demonstrates that the basic DiskVPS model without parameter optimization achieved significantly better performance over the SOTA in detection accuracy and processing speed with real-world images. The study also shows that DiskVPS is robust against parameter changes." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 47, + 477, + 128, + 490 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 477, + 128, + 490 + ], + "spans": [ + { + "bbox": [ + 47, + 477, + 128, + 490 + ], + "type": "text", + "content": "1. Introduction" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 46, + 497, + 287, + 653 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 497, + 287, + 653 + ], + "spans": [ + { + "bbox": [ + 46, + 497, + 287, + 653 + ], + "type": "text", + "content": "VP detection has a variety of applications, including camera calibration ([1, 3, 5, 15]), 3D reconstruction ([25, 40]), virtual reality [30], robot navigation [12], scene tracking [6] and vehicle automatic driving [13]. Depending on the direction of the parallel lines and the setup of the capturing camera, a VP can appear anywhere, including infinity, in the image plane. A triplet of orthogonal VPs is formed by 3 sets of mutually orthogonal parallel lines in the man-made world commonly called Manhattan world [5]. According to [3], the information of any 2 orthogonal VPs in an image is sufficient to obtain the focal length of the capturing camera. Moreover, a triplet of orthogonal VPs fully determines the camera's intrinsic and extrinsic parameters." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 47, + 659, + 140, + 671 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 659, + 140, + 671 + ], + "spans": [ + { + "bbox": [ + 47, + 659, + 140, + 671 + ], + "type": "text", + "content": "1.1. Related Works" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 46, + 677, + 287, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 677, + 287, + 715 + ], + "spans": [ + { + "bbox": [ + 46, + 677, + 287, + 715 + ], + "type": "text", + "content": "VP detection has long been a key research topic attracting great attention in computer vision. However, it remains a challenging issue far from soundly resolved. The key is-" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 304, + 228, + 547, + 289 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 228, + 547, + 289 + ], + "spans": [ + { + "bbox": [ + 304, + 228, + 547, + 289 + ], + "type": "text", + "content": "sue in VP detection is the unboundedness of the image plane in which VPs reside. In VP detection, it is essential to map the image plane into some bounded space before practical techniques such as Hough Transform (HT, [9]) can be deployed." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 304, + 291, + 547, + 626 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 291, + 547, + 626 + ], + "spans": [ + { + "bbox": [ + 304, + 291, + 547, + 626 + ], + "type": "text", + "content": "The first such attempt came from [29] that divided the image plane into 3 bounded subspaces to perform the HT. Since its division method is rather arbitrary without offsetting the probability biases in cell partitioning, it failed to yield satisfactory results. For calibrated images, Barnard [2] proposed a VP detection algorithm that mapped the image plane into a unit sphere called Gaussian sphere (GS) for HT, which is widely used in VP detection ([18, 24]) and is also universally applied in deep learning VP detection algorithms([4, 10, 11, 16, 23, 27, 32, 39]). However, GS-based approaches have limitations. Since building the GS requires the known parameters of the capturing camera such as the focal length and coordinates of the optical axis in the image plane, GS approaches are unsuitable for uncalibrated images, which account for a substantial fraction of images of interest. GS's voting scheme using edge pairs can significantly amplify discretization and distortion errors in edge detection and thus cause a high likelihood of spurious VP detection. Accordingly, GS models are prone to spurious detection ([16, 24, 39]). Moreover, the process of partitioning the GS into Hough cells itself is complicated and delicate. The non-planar shape of the GS is not amenable for partitioning of the GS into planar Hough cells while the Hough cell sizes on the GS are wildly non-uniform and sensitive to the errors of calibration parameters. With the limitations, the GS baseline model did not draw wide attention until the recent rapid development of deep learning technology." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 304, + 630, + 547, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 630, + 547, + 715 + ], + "spans": [ + { + "bbox": [ + 304, + 630, + 547, + 715 + ], + "type": "text", + "content": "As an important alternative to HT in VP detection, random sample consensus (RANSAC [8]) is another extensively explored scheme in VP detection ([17, 19, 21, 22, 26, 28, 33, 35, 38]). RANSAC can work directly in the image plane and is suitable for VP detection in uncalibrated images; moreover, RANSAC outperforms GS-based VP detection models in practice with scenes where many edges" + } + ] + } + ], + "index": 12 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "spans": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "type": "text", + "content": "CVF" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 145, + 0, + 494, + 37 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 145, + 0, + 494, + 37 + ], + "spans": [ + { + "bbox": [ + 145, + 0, + 494, + 37 + ], + "type": "text", + "content": "This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "type": "text", + "content": "27049" + } + ] + } + ], + "index": 13 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 0 + }, + { + "para_blocks": [ + { + "bbox": [ + 46, + 72, + 289, + 191 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 72, + 289, + 191 + ], + "spans": [ + { + "bbox": [ + 46, + 72, + 289, + 191 + ], + "type": "text", + "content": "belong to VP inliers. The main drawback of RANSAC is its susceptibility to false negative detection when the size of samples accounts for only a tiny portion of the full set of edge pairs used for VP detection (3 or more edges belonging to a set of 2 or 3 orthogonal VPs are sometimes proposed in [19]). Additionally, RANSAC is computationally expensive, particularly for images of cluttered scenes. Recently, [34] proposed an algorithm that integrated RANSAC and HT to achieve real-time VP detection. There are other methods like using dual spaces ([14])." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 46, + 193, + 290, + 397 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 193, + 290, + 397 + ], + "spans": [ + { + "bbox": [ + 46, + 193, + 290, + 397 + ], + "type": "text", + "content": "Despite its limitations, GS is the only proven VP detection model available to transform an unbounded space (image plane) into a bounded space (GS), enabling its deployment in deep learning neural networks. Therefore, the GS remains the chosen technique in deep-learning-based VP detection algorithms([4, 10, 11, 16, 23, 27, 32, 37, 39]). While the baseline model of GS-based VP detection is not optimal in both accuracy and processing speed, GS-based deep learning approaches making use of advanced high-level-feature extraction schemes and high-performance hardware have achieved the performance comparable to the SOTA, a level far superior to the GS's baseline model. However, deep learning approaches have limitations. The costs of data annotation and training needed by deep learning are expensive. Additionally, limited datasets usually do not generalize well in deep learning, which can fail in unexpected ways." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 47, + 407, + 132, + 418 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 407, + 132, + 418 + ], + "spans": [ + { + "bbox": [ + 47, + 407, + 132, + 418 + ], + "type": "text", + "content": "1.2. Contribution" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 46, + 426, + 289, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 426, + 289, + 715 + ], + "spans": [ + { + "bbox": [ + 46, + 426, + 289, + 715 + ], + "type": "text", + "content": "DiskVPS is suitable for conducting highly reliable and real-time VP detection. We use Cartesian coordinates " + }, + { + "bbox": [ + 46, + 426, + 289, + 715 + ], + "type": "inline_equation", + "content": "(u, v)" + }, + { + "bbox": [ + 46, + 426, + 289, + 715 + ], + "type": "text", + "content": " to denote the normalized coordinates with the origin set at the image center and both image width and height normalized to 1 (see next section for how " + }, + { + "bbox": [ + 46, + 426, + 289, + 715 + ], + "type": "inline_equation", + "content": "(u, v)" + }, + { + "bbox": [ + 46, + 426, + 289, + 715 + ], + "type": "text", + "content": " is mapped from the pixel-based image plane coordinates). DiskVPS maps an image plane (unbounded space) (Fig. 1(a)) into a disk-shaped space (Fig. 1(b)). Specifically, it maps an unbounded polar coordinate system " + }, + { + "bbox": [ + 46, + 426, + 289, + 715 + ], + "type": "inline_equation", + "content": "(r, \\theta)" + }, + { + "bbox": [ + 46, + 426, + 289, + 715 + ], + "type": "inline_equation", + "content": "(r = \\sqrt{u^2 + v^2}" + }, + { + "bbox": [ + 46, + 426, + 289, + 715 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 46, + 426, + 289, + 715 + ], + "type": "inline_equation", + "content": "\\theta = \\tan^{-1} \\frac{u}{v}" + }, + { + "bbox": [ + 46, + 426, + 289, + 715 + ], + "type": "text", + "content": " into a bounded polar coordinate system " + }, + { + "bbox": [ + 46, + 426, + 289, + 715 + ], + "type": "inline_equation", + "content": "(\\tan^{-1}(4r), \\theta)" + }, + { + "bbox": [ + 46, + 426, + 289, + 715 + ], + "type": "text", + "content": " with the latter's radial boundary located at " + }, + { + "bbox": [ + 46, + 426, + 289, + 715 + ], + "type": "inline_equation", + "content": "\\frac{\\pi}{2}" + }, + { + "bbox": [ + 46, + 426, + 289, + 715 + ], + "type": "text", + "content": ". With the approach, a straight line that passes the origin in the image plane is mapped into a diametrical line segment in the disk. In contrast, a straight line not passing the origin in the image plane is mapped into an arc-shaped segment with 2 ends overlapped with its origin-passing parallel counterpart at the disk boundary. Consequently, parallel lines in the image plane intersect on the disk boundary, whereas non-parallel lines in the image plane intersect each other only once inside the disk. In the scheme, HT is reduced to a probabilistic method to partition the disk into a set of approximately equal-probability Hough cells. Fig. 2 shows a set of edges detected in a typical image (see Fig. 2(a)) and its corresponding Hough accumulation map generated by the" + } + ] + } + ], + "index": 3 + }, + { + "type": "image", + "bbox": [ + 321, + 76, + 420, + 177 + ], + "blocks": [ + { + "bbox": [ + 321, + 76, + 420, + 177 + ], + "lines": [ + { + "bbox": [ + 321, + 76, + 420, + 177 + ], + "spans": [ + { + "bbox": [ + 321, + 76, + 420, + 177 + ], + "type": "image", + "image_path": "dfed0fd5dc4e2f4dad963cf38f9f921999e201678cb0408e3617ca24ea33a895.jpg" + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 359, + 181, + 373, + 193 + ], + "lines": [ + { + "bbox": [ + 359, + 181, + 373, + 193 + ], + "spans": [ + { + "bbox": [ + 359, + 181, + 373, + 193 + ], + "type": "text", + "content": "(a)" + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "image_caption" + } + ], + "index": 4 + }, + { + "type": "image", + "bbox": [ + 426, + 76, + 526, + 178 + ], + "blocks": [ + { + "bbox": [ + 426, + 76, + 526, + 178 + ], + "lines": [ + { + "bbox": [ + 426, + 76, + 526, + 178 + ], + "spans": [ + { + "bbox": [ + 426, + 76, + 526, + 178 + ], + "type": "image", + "image_path": "9bea78ddb7cea411f96e05c6197cfbc4716ef2f130ad127ece26ab9fb3a73d7f.jpg" + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 480, + 181, + 493, + 193 + ], + "lines": [ + { + "bbox": [ + 480, + 181, + 493, + 193 + ], + "spans": [ + { + "bbox": [ + 480, + 181, + 493, + 193 + ], + "type": "text", + "content": "(b)" + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "image_caption" + } + ], + "index": 6 + }, + { + "type": "image", + "bbox": [ + 329, + 251, + 417, + 316 + ], + "blocks": [ + { + "bbox": [ + 306, + 202, + 547, + 237 + ], + "lines": [ + { + "bbox": [ + 306, + 202, + 547, + 237 + ], + "spans": [ + { + "bbox": [ + 306, + 202, + 547, + 237 + ], + "type": "text", + "content": "Figure 1. Mapping from an image plane (Cartesian coordinates " + }, + { + "bbox": [ + 306, + 202, + 547, + 237 + ], + "type": "inline_equation", + "content": "(u, v)" + }, + { + "bbox": [ + 306, + 202, + 547, + 237 + ], + "type": "text", + "content": " (a)) to a disk-shaped Hough space (polar coordinates " + }, + { + "bbox": [ + 306, + 202, + 547, + 237 + ], + "type": "inline_equation", + "content": "(\\tan^{-1}(4r), \\theta)" + }, + { + "bbox": [ + 306, + 202, + 547, + 237 + ], + "type": "text", + "content": " (b)) with " + }, + { + "bbox": [ + 306, + 202, + 547, + 237 + ], + "type": "inline_equation", + "content": "r = \\sqrt{u^2 + v^2}" + }, + { + "bbox": [ + 306, + 202, + 547, + 237 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 306, + 202, + 547, + 237 + ], + "type": "inline_equation", + "content": "\\theta = \\tan^{-1} \\frac{v}{u}" + }, + { + "bbox": [ + 306, + 202, + 547, + 237 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 329, + 251, + 417, + 316 + ], + "lines": [ + { + "bbox": [ + 329, + 251, + 417, + 316 + ], + "spans": [ + { + "bbox": [ + 329, + 251, + 417, + 316 + ], + "type": "image", + "image_path": "05fc8c78f9eadcfd2cd1bb6057c271d2b84b5f299744ac83a3fd2e798c3407c0.jpg" + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 359, + 327, + 373, + 338 + ], + "lines": [ + { + "bbox": [ + 359, + 327, + 373, + 338 + ], + "spans": [ + { + "bbox": [ + 359, + 327, + 373, + 338 + ], + "type": "text", + "content": "(a)" + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "image_caption" + } + ], + "index": 9 + }, + { + "type": "image", + "bbox": [ + 432, + 256, + 526, + 319 + ], + "blocks": [ + { + "bbox": [ + 432, + 256, + 526, + 319 + ], + "lines": [ + { + "bbox": [ + 432, + 256, + 526, + 319 + ], + "spans": [ + { + "bbox": [ + 432, + 256, + 526, + 319 + ], + "type": "image", + "image_path": "7f28b618a869c4a5ed9ad512be42ecabacff4fd00b7155b7b6c6d82921be608d.jpg" + } + ] + } + ], + "index": 11, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 480, + 327, + 493, + 338 + ], + "lines": [ + { + "bbox": [ + 480, + 327, + 493, + 338 + ], + "spans": [ + { + "bbox": [ + 480, + 327, + 493, + 338 + ], + "type": "text", + "content": "(b)" + } + ] + } + ], + "index": 12, + "angle": 0, + "type": "image_caption" + } + ], + "index": 11 + }, + { + "type": "image", + "bbox": [ + 378, + 387, + 474, + 473 + ], + "blocks": [ + { + "bbox": [ + 306, + 347, + 545, + 370 + ], + "lines": [ + { + "bbox": [ + 306, + 347, + 545, + 370 + ], + "spans": [ + { + "bbox": [ + 306, + 347, + 545, + 370 + ], + "type": "text", + "content": "Figure 2. Detected edges of a typical image (a) and its corresponding Hough accumulation map (b)." + } + ] + } + ], + "index": 13, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 378, + 387, + 474, + 473 + ], + "lines": [ + { + "bbox": [ + 378, + 387, + 474, + 473 + ], + "spans": [ + { + "bbox": [ + 378, + 387, + 474, + 473 + ], + "type": "image", + "image_path": "330a8000a1b133d2c18f962ebb4fa0c4c928ca2160f1a1765f0887765e85887a.jpg" + } + ] + } + ], + "index": 14, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 306, + 492, + 545, + 515 + ], + "lines": [ + { + "bbox": [ + 306, + 492, + 545, + 515 + ], + "spans": [ + { + "bbox": [ + 306, + 492, + 545, + 515 + ], + "type": "text", + "content": "Figure 3. Intersection " + }, + { + "bbox": [ + 306, + 492, + 545, + 515 + ], + "type": "inline_equation", + "content": "P" + }, + { + "bbox": [ + 306, + 492, + 545, + 515 + ], + "type": "text", + "content": " of the extension line of a random edge and a given line defined by the distance " + }, + { + "bbox": [ + 306, + 492, + 545, + 515 + ], + "type": "inline_equation", + "content": "d" + }, + { + "bbox": [ + 306, + 492, + 545, + 515 + ], + "type": "text", + "content": " and the angle " + }, + { + "bbox": [ + 306, + 492, + 545, + 515 + ], + "type": "inline_equation", + "content": "\\phi" + }, + { + "bbox": [ + 306, + 492, + 545, + 515 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 15, + "angle": 0, + "type": "image_caption" + } + ], + "index": 14 + }, + { + "bbox": [ + 304, + 535, + 547, + 608 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 535, + 547, + 608 + ], + "spans": [ + { + "bbox": [ + 304, + 535, + 547, + 608 + ], + "type": "text", + "content": "edges using DiskVPS (see Fig. 2(b)). Five sharp and salient peaks represent 3 mutually orthogonal VPs, 2 far-field VPs generated by nearly parallel edges and one near-field VP generated by non-parallel edges. The proposed approach successfully manifests all valid VPs in the image without yielding any spurious VPs in Fig. 2b." + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 306, + 619, + 437, + 633 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 619, + 437, + 633 + ], + "spans": [ + { + "bbox": [ + 306, + 619, + 437, + 633 + ], + "type": "text", + "content": "2. Algorithm Description" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 306, + 639, + 440, + 651 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 639, + 440, + 651 + ], + "spans": [ + { + "bbox": [ + 306, + 639, + 440, + 651 + ], + "type": "text", + "content": "2.1. Theoretical Foundation" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 304, + 657, + 547, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 657, + 547, + 715 + ], + "spans": [ + { + "bbox": [ + 304, + 657, + 547, + 715 + ], + "type": "text", + "content": "Assume " + }, + { + "bbox": [ + 304, + 657, + 547, + 715 + ], + "type": "inline_equation", + "content": "(x,y)" + }, + { + "bbox": [ + 304, + 657, + 547, + 715 + ], + "type": "text", + "content": " is the pixel-based coordinates of a point in the image plane and the image of interest is encoded by 2 coordinates in pixels: " + }, + { + "bbox": [ + 304, + 657, + 547, + 715 + ], + "type": "inline_equation", + "content": "(x_{\\min},y_{\\min})" + }, + { + "bbox": [ + 304, + 657, + 547, + 715 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 304, + 657, + 547, + 715 + ], + "type": "inline_equation", + "content": "(x_{\\max},y_{\\max})" + }, + { + "bbox": [ + 304, + 657, + 547, + 715 + ], + "type": "text", + "content": ", which refer to the pixel-based coordinates of the image's bottom left corner and top right corner respectively. To ease analy-" + } + ] + } + ], + "index": 19 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "type": "text", + "content": "27050" + } + ] + } + ], + "index": 20 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 1 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 107, + 72, + 227, + 183 + ], + "blocks": [ + { + "bbox": [ + 107, + 72, + 227, + 183 + ], + "lines": [ + { + "bbox": [ + 107, + 72, + 227, + 183 + ], + "spans": [ + { + "bbox": [ + 107, + 72, + 227, + 183 + ], + "type": "image", + "image_path": "3d0d52e89e5c6ced96176238e73654e014726575ed37294b3f10cf1e49cf97ea.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 47, + 193, + 287, + 216 + ], + "lines": [ + { + "bbox": [ + 47, + 193, + 287, + 216 + ], + "spans": [ + { + "bbox": [ + 47, + 193, + 287, + 216 + ], + "type": "text", + "content": "Figure 4. Illustration of the 2 special directions along which the PDF is considered, the radial and tangential directions." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "bbox": [ + 47, + 238, + 287, + 262 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 238, + 287, + 262 + ], + "spans": [ + { + "bbox": [ + 47, + 238, + 287, + 262 + ], + "type": "text", + "content": "sis, " + }, + { + "bbox": [ + 47, + 238, + 287, + 262 + ], + "type": "inline_equation", + "content": "(x,y)" + }, + { + "bbox": [ + 47, + 238, + 287, + 262 + ], + "type": "text", + "content": " is transformed to a normalized coordinate system " + }, + { + "bbox": [ + 47, + 238, + 287, + 262 + ], + "type": "inline_equation", + "content": "(u,v)" + }, + { + "bbox": [ + 47, + 238, + 287, + 262 + ], + "type": "text", + "content": ". Their relationship is written as:" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 119, + 273, + 287, + 304 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 119, + 273, + 287, + 304 + ], + "spans": [ + { + "bbox": [ + 119, + 273, + 287, + 304 + ], + "type": "interline_equation", + "content": "\\left\\{ \\begin{array}{l} u = \\frac {x - 0 . 5 \\left(x _ {\\min } + x _ {\\max }\\right)}{x _ {\\max } - x _ {\\min }} \\\\ v = \\frac {y - 0 . 5 \\left(y _ {\\min } + y _ {\\max }\\right)}{y _ {\\max } - y _ {\\min }} \\end{array} \\right. \\tag {1}", + "image_path": "9f433c4883c9228cb4130a6d8bdf0a7753042715b87756fd9945d178656ee088.jpg" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 47, + 316, + 287, + 376 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 316, + 287, + 376 + ], + "spans": [ + { + "bbox": [ + 47, + 316, + 287, + 376 + ], + "type": "text", + "content": "In the " + }, + { + "bbox": [ + 47, + 316, + 287, + 376 + ], + "type": "inline_equation", + "content": "(u,v)" + }, + { + "bbox": [ + 47, + 316, + 287, + 376 + ], + "type": "text", + "content": " system, the origin " + }, + { + "bbox": [ + 47, + 316, + 287, + 376 + ], + "type": "inline_equation", + "content": "O" + }, + { + "bbox": [ + 47, + 316, + 287, + 376 + ], + "type": "text", + "content": " is fixed to the image center, and the bottom left corner and top right corner of the image are moved to " + }, + { + "bbox": [ + 47, + 316, + 287, + 376 + ], + "type": "inline_equation", + "content": "(-0.5, -0.5)" + }, + { + "bbox": [ + 47, + 316, + 287, + 376 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 47, + 316, + 287, + 376 + ], + "type": "inline_equation", + "content": "(0.5, 0.5)" + }, + { + "bbox": [ + 47, + 316, + 287, + 376 + ], + "type": "text", + "content": ", respectively (see Fig. 3). The image region turns into an Origin-centered unit square in the " + }, + { + "bbox": [ + 47, + 316, + 287, + 376 + ], + "type": "inline_equation", + "content": "(u,v)" + }, + { + "bbox": [ + 47, + 316, + 287, + 376 + ], + "type": "text", + "content": " system." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 46, + 377, + 289, + 521 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 377, + 289, + 521 + ], + "spans": [ + { + "bbox": [ + 46, + 377, + 289, + 521 + ], + "type": "text", + "content": "[34] proposed a statistical model to estimate the probabilistic distribution of " + }, + { + "bbox": [ + 46, + 377, + 289, + 521 + ], + "type": "inline_equation", + "content": "P" + }, + { + "bbox": [ + 46, + 377, + 289, + 521 + ], + "type": "text", + "content": ", the intersection of a given line and the extension line of a random edge, along the very given line. According to [34], the midpoint of the random edge is distributed uniformly in the unit square, and the direction of the random edge is also uniformly distributed over " + }, + { + "bbox": [ + 46, + 377, + 289, + 521 + ], + "type": "inline_equation", + "content": "[0, \\pi]" + }, + { + "bbox": [ + 46, + 377, + 289, + 521 + ], + "type": "text", + "content": ". Such a random edge can be used to simulate the statistical distribution of the edges extracted from a sufficiently large number of images. We cite the conclusion of [34] in this paper. As shown in Fig. 3, if the given line is defined by " + }, + { + "bbox": [ + 46, + 377, + 289, + 521 + ], + "type": "inline_equation", + "content": "d" + }, + { + "bbox": [ + 46, + 377, + 289, + 521 + ], + "type": "text", + "content": " (the distance from " + }, + { + "bbox": [ + 46, + 377, + 289, + 521 + ], + "type": "inline_equation", + "content": "O" + }, + { + "bbox": [ + 46, + 377, + 289, + 521 + ], + "type": "text", + "content": " on the given line) and " + }, + { + "bbox": [ + 46, + 377, + 289, + 521 + ], + "type": "inline_equation", + "content": "\\phi" + }, + { + "bbox": [ + 46, + 377, + 289, + 521 + ], + "type": "text", + "content": " (the directional angle pointing from " + }, + { + "bbox": [ + 46, + 377, + 289, + 521 + ], + "type": "inline_equation", + "content": "O" + }, + { + "bbox": [ + 46, + 377, + 289, + 521 + ], + "type": "text", + "content": " to its perpendicular foot to the given line), then [34] gives:" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 47, + 532, + 303, + 577 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 532, + 303, + 577 + ], + "spans": [ + { + "bbox": [ + 47, + 532, + 303, + 577 + ], + "type": "interline_equation", + "content": "f _ {d, \\phi} (t) = \\frac {1}{\\pi} \\int_ {- 0. 5 - 0. 5} ^ {0. 5} \\int_ {0} ^ {0. 5} \\frac {\\left| u \\cos \\phi + v \\sin \\phi - d \\right| d u d v}{(u - d \\cos \\phi - t \\sin \\phi) ^ {2} + (v - d \\sin \\phi + t \\cos \\phi) ^ {2}} \\tag {2}", + "image_path": "31884ac94550003ac922d14c0b0a44b5357570c6f7c3fb0c9d3278ccb898edf8.jpg" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 47, + 578, + 287, + 613 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 578, + 287, + 613 + ], + "spans": [ + { + "bbox": [ + 47, + 578, + 287, + 613 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 47, + 578, + 287, + 613 + ], + "type": "inline_equation", + "content": "t" + }, + { + "bbox": [ + 47, + 578, + 287, + 613 + ], + "type": "text", + "content": " is the coordinate of " + }, + { + "bbox": [ + 47, + 578, + 287, + 613 + ], + "type": "inline_equation", + "content": "P" + }, + { + "bbox": [ + 47, + 578, + 287, + 613 + ], + "type": "text", + "content": " on a new axis created along the given line with the origin of the axis set at the perpendicular foot of " + }, + { + "bbox": [ + 47, + 578, + 287, + 613 + ], + "type": "inline_equation", + "content": "O" + }, + { + "bbox": [ + 47, + 578, + 287, + 613 + ], + "type": "text", + "content": " on the given line, as shown in Fig. 3." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 47, + 614, + 287, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 614, + 287, + 714 + ], + "spans": [ + { + "bbox": [ + 47, + 614, + 287, + 714 + ], + "type": "text", + "content": "Here, we are interested in Eq.2's solutions for 2 special cases. As shown in Fig.4, one is for the case in which the given line is along the radial direction in the " + }, + { + "bbox": [ + 47, + 614, + 287, + 714 + ], + "type": "inline_equation", + "content": "(u,v)" + }, + { + "bbox": [ + 47, + 614, + 287, + 714 + ], + "type": "text", + "content": " coordinates, and the other is for the case in which the given line is in the azimuthal or tangential direction. " + }, + { + "bbox": [ + 47, + 614, + 287, + 714 + ], + "type": "inline_equation", + "content": "f_{\\mathrm{rad}}(r,\\theta)" + }, + { + "bbox": [ + 47, + 614, + 287, + 714 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 47, + 614, + 287, + 714 + ], + "type": "inline_equation", + "content": "f_{\\mathrm{tag}}(r,\\theta)" + }, + { + "bbox": [ + 47, + 614, + 287, + 714 + ], + "type": "text", + "content": " are used to denote the PDFs of " + }, + { + "bbox": [ + 47, + 614, + 287, + 714 + ], + "type": "inline_equation", + "content": "P" + }, + { + "bbox": [ + 47, + 614, + 287, + 714 + ], + "type": "text", + "content": " along the given line along the radial direction and the given line along the tangential direction respectively, where " + }, + { + "bbox": [ + 47, + 614, + 287, + 714 + ], + "type": "inline_equation", + "content": "\\theta" + }, + { + "bbox": [ + 47, + 614, + 287, + 714 + ], + "type": "text", + "content": " is the directional angle of " + }, + { + "bbox": [ + 47, + 614, + 287, + 714 + ], + "type": "inline_equation", + "content": "OP" + }, + { + "bbox": [ + 47, + 614, + 287, + 714 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 47, + 614, + 287, + 714 + ], + "type": "inline_equation", + "content": "r" + }, + { + "bbox": [ + 47, + 614, + 287, + 714 + ], + "type": "text", + "content": " is the length of " + }, + { + "bbox": [ + 47, + 614, + 287, + 714 + ], + "type": "inline_equation", + "content": "OP" + }, + { + "bbox": [ + 47, + 614, + 287, + 714 + ], + "type": "text", + "content": ", both in the " + }, + { + "bbox": [ + 47, + 614, + 287, + 714 + ], + "type": "inline_equation", + "content": "(u,v)" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 305, + 72, + 477, + 85 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 72, + 477, + 85 + ], + "spans": [ + { + "bbox": [ + 305, + 72, + 477, + 85 + ], + "type": "text", + "content": "coordinate system. Accordingly, we have:" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 326, + 91, + 545, + 142 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 326, + 91, + 545, + 142 + ], + "spans": [ + { + "bbox": [ + 326, + 91, + 545, + 142 + ], + "type": "interline_equation", + "content": "\\begin{array}{l} f _ {\\mathrm {r a d}} (r, \\theta) = f _ {0, \\theta + \\frac {\\pi}{2}} (r) \\\\ = \\frac {1}{\\pi} \\int_ {- 0. 5 - 0. 5} ^ {0. 5} \\int_ {0} ^ {0. 5} \\frac {\\left| - u \\sin \\theta + v \\cos \\theta \\right| d u d v}{(u - r \\cos \\theta) ^ {2} + (v - r \\sin \\theta) ^ {2}} \\tag {3} \\\\ \\end{array}", + "image_path": "656066f71fe96b75845d08bebbfef7cc98c94c33998e509309f4ef205df5056a.jpg" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 317, + 152, + 545, + 202 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 152, + 545, + 202 + ], + "spans": [ + { + "bbox": [ + 317, + 152, + 545, + 202 + ], + "type": "interline_equation", + "content": "\\begin{array}{l} f _ {\\mathrm {t a g}} (r, \\theta) = f _ {r, \\theta + \\frac {\\pi}{2}} (0) \\\\ = \\frac {1}{\\pi} \\int_ {- 0. 5} ^ {0. 5} \\int_ {- 0. 5} ^ {0. 5} \\frac {\\left| u \\sin \\theta - v \\cos \\theta + r \\right| d u d v}{(u + r \\sin \\theta) ^ {2} + (v - r \\cos \\theta) ^ {2}} \\tag {4} \\\\ \\end{array}", + "image_path": "274c13ad412fca15ac46d667b4d95f4a8f3461304a15ffd1a0ca0afc92459ffb.jpg" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 305, + 206, + 545, + 262 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 206, + 545, + 262 + ], + "spans": [ + { + "bbox": [ + 305, + 206, + 545, + 262 + ], + "type": "text", + "content": "Since " + }, + { + "bbox": [ + 305, + 206, + 545, + 262 + ], + "type": "inline_equation", + "content": "f_{\\mathrm{rad}}(r,\\theta)" + }, + { + "bbox": [ + 305, + 206, + 545, + 262 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 305, + 206, + 545, + 262 + ], + "type": "inline_equation", + "content": "f_{\\mathrm{tag}}(r,\\theta)" + }, + { + "bbox": [ + 305, + 206, + 545, + 262 + ], + "type": "text", + "content": "'s dependence on the variable " + }, + { + "bbox": [ + 305, + 206, + 545, + 262 + ], + "type": "inline_equation", + "content": "\\theta" + }, + { + "bbox": [ + 305, + 206, + 545, + 262 + ], + "type": "text", + "content": " is insignificant (see Fig.5), they are conveniently approximated by 2 less complex and more manageable forms " + }, + { + "bbox": [ + 305, + 206, + 545, + 262 + ], + "type": "inline_equation", + "content": "g_{\\mathrm{rad}}(r) = f_{\\mathrm{rad}}(r,0)" + }, + { + "bbox": [ + 305, + 206, + 545, + 262 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 305, + 206, + 545, + 262 + ], + "type": "inline_equation", + "content": "g_{\\mathrm{tag}}(r) = f_{\\mathrm{tag}}(r,0)" + }, + { + "bbox": [ + 305, + 206, + 545, + 262 + ], + "type": "text", + "content": " to facilitate subsequent construction of Hough space. They are:" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 359, + 270, + 545, + 305 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 359, + 270, + 545, + 305 + ], + "spans": [ + { + "bbox": [ + 359, + 270, + 545, + 305 + ], + "type": "interline_equation", + "content": "g _ {\\text {r a d}} (r) = \\frac {1}{\\pi} \\int_ {- 0. 5 - 0. 5} ^ {0. 5} \\int_ {0} ^ {0. 5} \\frac {| v | d u d v}{(u - r) ^ {2} + v ^ {2}} \\tag {5}", + "image_path": "97bff9d0132bc4a7048b10c97b8ba405372c733e90ea34ac7811dc835373f7b2.jpg" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 358, + 315, + 545, + 350 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 358, + 315, + 545, + 350 + ], + "spans": [ + { + "bbox": [ + 358, + 315, + 545, + 350 + ], + "type": "interline_equation", + "content": "g _ {\\text {a g}} (r) = \\frac {1}{\\pi} \\int_ {- 0. 5 - 0. 5} ^ {0. 5} \\int_ {- 0. 5} ^ {0. 5} \\frac {| v - r | d u d v}{u ^ {2} + (v - r) ^ {2}} \\tag {6}", + "image_path": "e9af42652d9071a44df9382af4edad2381e99f36b29c652b6d690a9f1c4fc993.jpg" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 318, + 354, + 520, + 366 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 318, + 354, + 520, + 366 + ], + "spans": [ + { + "bbox": [ + 318, + 354, + 520, + 366 + ], + "type": "text", + "content": "The analytical solutions of " + }, + { + "bbox": [ + 318, + 354, + 520, + 366 + ], + "type": "inline_equation", + "content": "g_{\\mathrm{rad}}(r)" + }, + { + "bbox": [ + 318, + 354, + 520, + 366 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 318, + 354, + 520, + 366 + ], + "type": "inline_equation", + "content": "g_{\\mathrm{tag}}(r)" + }, + { + "bbox": [ + 318, + 354, + 520, + 366 + ], + "type": "text", + "content": " are:" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 306, + 373, + 545, + 481 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 373, + 545, + 481 + ], + "spans": [ + { + "bbox": [ + 306, + 373, + 545, + 481 + ], + "type": "interline_equation", + "content": "g _ {\\mathrm {r a d}} (r) = \\left\\{ \\begin{array}{l l} \\frac {\\tan^ {- 1} (2 r + 1) - \\tan^ {- 1} (2 r - 1)}{\\pi} & \\\\ + \\frac {1}{2 \\pi} \\ln \\frac {[ (2 r - 1) ^ {2} + 1 ] [ (2 r + 1) ^ {2} + 1 ]}{(2 r - 1) ^ {2} (2 r + 1) ^ {2}} & r \\neq 0. 5 \\\\ + \\frac {r}{\\pi} \\ln \\frac {(2 r - 1) ^ {2} [ (2 r + 1) ^ {2} + 1 ]}{(2 r + 1) ^ {2} [ (2 r - 1) ^ {2} + 1 ]} & \\\\ \\frac {\\tan^ {- 1} 2 + \\ln 5 - \\ln 4}{\\pi} & r = 0. 5 \\end{array} \\right. \\tag {7}", + "image_path": "be8b16827cbc88b6e8f57587455286861464013528863aa66e1dab7453702b0d.jpg" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 306, + 497, + 545, + 650 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 497, + 545, + 650 + ], + "spans": [ + { + "bbox": [ + 306, + 497, + 545, + 650 + ], + "type": "interline_equation", + "content": "g _ {\\text {t a g}} (r) = \\left\\{ \\begin{array}{l l} 1 - \\frac {\\tan^ {- 1} (2 r + 1) - \\tan^ {- 1} (2 r - 1)}{\\pi} & r \\leq 0. 5 \\\\ - \\frac {2 r \\left[ \\tan^ {- 1} (2 r + 1) + \\tan^ {- 1} (2 r - 1) \\right]}{\\pi} & \\\\ + \\frac {\\ln [ 1 + (2 r + 1) ^ {2} ] + \\ln [ 1 + (2 r - 1) ^ {2} ]}{2 \\pi} & \\\\ 1 - \\frac {(2 r + 1) \\tan^ {- 1} (2 r + 1)}{\\pi} & \\\\ - \\frac {(2 r - 1) \\tan^ {- 1} (2 r - 1)}{\\pi} & r > 0. 5 \\\\ + \\frac {\\ln [ 1 + (2 r + 1) ^ {2} ] - \\ln [ 1 + (2 r - 1) ^ {2} ]}{2 \\pi} & \\end{array} \\right. \\tag {8}", + "image_path": "80a974141b899e7772a90e701d3614f435a440809633adb6c108c1466faa9ab5.jpg" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 305, + 651, + 545, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 651, + 545, + 713 + ], + "spans": [ + { + "bbox": [ + 305, + 651, + 545, + 713 + ], + "type": "text", + "content": "As shown in Fig. 5, both " + }, + { + "bbox": [ + 305, + 651, + 545, + 713 + ], + "type": "inline_equation", + "content": "\\frac{g_{\\mathrm{rad}}(r)}{f_{\\mathrm{rad}}(r,\\theta)}" + }, + { + "bbox": [ + 305, + 651, + 545, + 713 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 305, + 651, + 545, + 713 + ], + "type": "inline_equation", + "content": "\\frac{g_{\\mathrm{tag}}(r)}{f_{\\mathrm{tag}}(r,\\theta)}" + }, + { + "bbox": [ + 305, + 651, + 545, + 713 + ], + "type": "text", + "content": " lie between 0.88 and 1.13 for " + }, + { + "bbox": [ + 305, + 651, + 545, + 713 + ], + "type": "inline_equation", + "content": "r \\leq 1.7" + }, + { + "bbox": [ + 305, + 651, + 545, + 713 + ], + "type": "text", + "content": " and they move closer to 1 with the increase of " + }, + { + "bbox": [ + 305, + 651, + 545, + 713 + ], + "type": "inline_equation", + "content": "r" + }, + { + "bbox": [ + 305, + 651, + 545, + 713 + ], + "type": "text", + "content": " for " + }, + { + "bbox": [ + 305, + 651, + 545, + 713 + ], + "type": "inline_equation", + "content": "r \\geq 1.0" + }, + { + "bbox": [ + 305, + 651, + 545, + 713 + ], + "type": "text", + "content": ". Overall, using " + }, + { + "bbox": [ + 305, + 651, + 545, + 713 + ], + "type": "inline_equation", + "content": "g_{\\mathrm{rad}}(r)" + }, + { + "bbox": [ + 305, + 651, + 545, + 713 + ], + "type": "text", + "content": " to approximate " + }, + { + "bbox": [ + 305, + 651, + 545, + 713 + ], + "type": "inline_equation", + "content": "f_{\\mathrm{rad}}(r,\\theta)" + }, + { + "bbox": [ + 305, + 651, + 545, + 713 + ], + "type": "text", + "content": " and using " + }, + { + "bbox": [ + 305, + 651, + 545, + 713 + ], + "type": "inline_equation", + "content": "g_{\\mathrm{tag}}(r)" + }, + { + "bbox": [ + 305, + 651, + 545, + 713 + ], + "type": "text", + "content": " to approximate " + }, + { + "bbox": [ + 305, + 651, + 545, + 713 + ], + "type": "inline_equation", + "content": "f_{\\mathrm{tag}}(r,\\theta)" + }, + { + "bbox": [ + 305, + 651, + 545, + 713 + ], + "type": "text", + "content": " do not lead to significant errors." + } + ] + } + ], + "index": 18 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "text", + "content": "27051" + } + ] + } + ], + "index": 19 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 2 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 49, + 75, + 160, + 160 + ], + "blocks": [ + { + "bbox": [ + 49, + 75, + 160, + 160 + ], + "lines": [ + { + "bbox": [ + 49, + 75, + 160, + 160 + ], + "spans": [ + { + "bbox": [ + 49, + 75, + 160, + 160 + ], + "type": "image", + "image_path": "b0c19e335660e103c499d6f7dc6cdb1fff7dc3dbe868829abc0d2f41abb8b4ed.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 100, + 164, + 113, + 174 + ], + "lines": [ + { + "bbox": [ + 100, + 164, + 113, + 174 + ], + "spans": [ + { + "bbox": [ + 100, + 164, + 113, + 174 + ], + "type": "text", + "content": "(a)" + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "type": "image", + "bbox": [ + 167, + 74, + 277, + 160 + ], + "blocks": [ + { + "bbox": [ + 167, + 74, + 277, + 160 + ], + "lines": [ + { + "bbox": [ + 167, + 74, + 277, + 160 + ], + "spans": [ + { + "bbox": [ + 167, + 74, + 277, + 160 + ], + "type": "image", + "image_path": "19fd581a7f960b722199ec8be233b5a09acf49e9900fdf9d641e6cb28e294cea.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 221, + 163, + 234, + 174 + ], + "lines": [ + { + "bbox": [ + 221, + 163, + 234, + 174 + ], + "spans": [ + { + "bbox": [ + 221, + 163, + 234, + 174 + ], + "type": "text", + "content": "(b)" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_caption" + } + ], + "index": 2 + }, + { + "type": "image", + "bbox": [ + 108, + 222, + 218, + 308 + ], + "blocks": [ + { + "bbox": [ + 47, + 184, + 288, + 207 + ], + "lines": [ + { + "bbox": [ + 47, + 184, + 288, + 207 + ], + "spans": [ + { + "bbox": [ + 47, + 184, + 288, + 207 + ], + "type": "text", + "content": "Figure 5. Proximity of " + }, + { + "bbox": [ + 47, + 184, + 288, + 207 + ], + "type": "inline_equation", + "content": "g_{\\mathrm{rad}}(r)" + }, + { + "bbox": [ + 47, + 184, + 288, + 207 + ], + "type": "text", + "content": " to " + }, + { + "bbox": [ + 47, + 184, + 288, + 207 + ], + "type": "inline_equation", + "content": "f_{\\mathrm{rad}}(r,\\theta)" + }, + { + "bbox": [ + 47, + 184, + 288, + 207 + ], + "type": "text", + "content": " (a) and proximity of " + }, + { + "bbox": [ + 47, + 184, + 288, + 207 + ], + "type": "inline_equation", + "content": "g_{\\mathrm{tag}}(r)" + }, + { + "bbox": [ + 47, + 184, + 288, + 207 + ], + "type": "text", + "content": " to " + }, + { + "bbox": [ + 47, + 184, + 288, + 207 + ], + "type": "inline_equation", + "content": "f_{\\mathrm{tag}}(r,\\theta)" + }, + { + "bbox": [ + 47, + 184, + 288, + 207 + ], + "type": "text", + "content": " (b)." + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 108, + 222, + 218, + 308 + ], + "lines": [ + { + "bbox": [ + 108, + 222, + 218, + 308 + ], + "spans": [ + { + "bbox": [ + 108, + 222, + 218, + 308 + ], + "type": "image", + "image_path": "9df6ead6121bd57f5c9992989bbf3706cea7f665e818bc692db34bc50fc351e2.jpg" + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 47, + 319, + 282, + 335 + ], + "lines": [ + { + "bbox": [ + 47, + 319, + 282, + 335 + ], + "spans": [ + { + "bbox": [ + 47, + 319, + 282, + 335 + ], + "type": "text", + "content": "Figure 6. " + }, + { + "bbox": [ + 47, + 319, + 282, + 335 + ], + "type": "inline_equation", + "content": "\\frac{1}{\\pi} \\frac{4}{(4r)^2 + 1}" + }, + { + "bbox": [ + 47, + 319, + 282, + 335 + ], + "type": "text", + "content": " 's proximity to " + }, + { + "bbox": [ + 47, + 319, + 282, + 335 + ], + "type": "inline_equation", + "content": "f_{\\mathrm{rad}}(r, \\theta)" + }, + { + "bbox": [ + 47, + 319, + 282, + 335 + ], + "type": "text", + "content": " for chosen " + }, + { + "bbox": [ + 47, + 319, + 282, + 335 + ], + "type": "inline_equation", + "content": "r" + }, + { + "bbox": [ + 47, + 319, + 282, + 335 + ], + "type": "text", + "content": " values." + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "image_caption" + } + ], + "index": 5 + }, + { + "bbox": [ + 46, + 356, + 287, + 415 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 356, + 287, + 415 + ], + "spans": [ + { + "bbox": [ + 46, + 356, + 287, + 415 + ], + "type": "text", + "content": "The previous analysis provides a theoretical basis for building a Hough space from the " + }, + { + "bbox": [ + 46, + 356, + 287, + 415 + ], + "type": "inline_equation", + "content": "(r,\\theta)" + }, + { + "bbox": [ + 46, + 356, + 287, + 415 + ], + "type": "text", + "content": ". However, since the radial dimension " + }, + { + "bbox": [ + 46, + 356, + 287, + 415 + ], + "type": "inline_equation", + "content": "(r)" + }, + { + "bbox": [ + 46, + 356, + 287, + 415 + ], + "type": "text", + "content": " remains unbounded, it is not feasible to create a Hough space directly from its current form (Eq.7 and 8)." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 47, + 416, + 287, + 489 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 416, + 287, + 489 + ], + "spans": [ + { + "bbox": [ + 47, + 416, + 287, + 489 + ], + "type": "text", + "content": "So it makes sense to transform " + }, + { + "bbox": [ + 47, + 416, + 287, + 489 + ], + "type": "inline_equation", + "content": "r" + }, + { + "bbox": [ + 47, + 416, + 287, + 489 + ], + "type": "text", + "content": " to a bounded space and derive the PDF in the new space. Note that when " + }, + { + "bbox": [ + 47, + 416, + 287, + 489 + ], + "type": "inline_equation", + "content": "r \\gg 1" + }, + { + "bbox": [ + 47, + 416, + 287, + 489 + ], + "type": "text", + "content": ", Eq.5 leads to " + }, + { + "bbox": [ + 47, + 416, + 287, + 489 + ], + "type": "inline_equation", + "content": "g_{\\mathrm{rad}}(r) \\approx \\frac{1}{\\pi r^2} \\int_{-0.5}^{0.5} \\int_{-0.5}^{0.5} |v| du dv = \\frac{1}{4\\pi r^2}" + }, + { + "bbox": [ + 47, + 416, + 287, + 489 + ], + "type": "text", + "content": ". Since " + }, + { + "bbox": [ + 47, + 416, + 287, + 489 + ], + "type": "inline_equation", + "content": "\\frac{1}{4\\pi r^2} \\approx \\frac{1}{\\pi} \\frac{4}{(4r)^2 + 1}" + }, + { + "bbox": [ + 47, + 416, + 287, + 489 + ], + "type": "text", + "content": " for " + }, + { + "bbox": [ + 47, + 416, + 287, + 489 + ], + "type": "inline_equation", + "content": "r \\gg 1" + }, + { + "bbox": [ + 47, + 416, + 287, + 489 + ], + "type": "text", + "content": ", Eq. 5 can be further approximated as:" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 104, + 489, + 287, + 512 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 104, + 489, + 287, + 512 + ], + "spans": [ + { + "bbox": [ + 104, + 489, + 287, + 512 + ], + "type": "interline_equation", + "content": "g _ {\\text {r a d}} (r) \\approx \\frac {1}{\\pi} \\frac {4}{(4 r) ^ {2} + 1} \\text {f o r} r \\gg 1 \\tag {9}", + "image_path": "9396d5a6c4b831c58622024fddbb5505c6ce96ad11a37f9cc55538bcac8348e0.jpg" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 47, + 516, + 287, + 634 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 516, + 287, + 634 + ], + "spans": [ + { + "bbox": [ + 47, + 516, + 287, + 634 + ], + "type": "text", + "content": "Here " + }, + { + "bbox": [ + 47, + 516, + 287, + 634 + ], + "type": "inline_equation", + "content": "g_{\\mathrm{rad}}(r)\\approx \\frac{1}{\\pi}\\frac{4}{(4r)^2 + 1}" + }, + { + "bbox": [ + 47, + 516, + 287, + 634 + ], + "type": "text", + "content": " is chosen over other alternatives because the integration of " + }, + { + "bbox": [ + 47, + 516, + 287, + 634 + ], + "type": "inline_equation", + "content": "\\frac{1}{\\pi}\\frac{4}{(4r)^2 + 1}" + }, + { + "bbox": [ + 47, + 516, + 287, + 634 + ], + "type": "text", + "content": " over a full radial line is equal to 1, which matches the probability of a radial line intersecting the extension line of the random edge. Note " + }, + { + "bbox": [ + 47, + 516, + 287, + 634 + ], + "type": "inline_equation", + "content": "\\frac{1}{\\pi}\\int \\limits_{r = -\\infty}^{r = +\\infty}\\frac{4dr}{(4r)^2 + 1} = \\frac{1}{\\pi}\\int \\limits_{r = -\\infty}^{r = +\\infty}d\\tan^{-1}(4r) = \\frac{1}{\\pi}\\times 2\\times \\frac{\\pi}{2} = 1." + }, + { + "bbox": [ + 47, + 516, + 287, + 634 + ], + "type": "text", + "content": " As shown in Fig.6, when " + }, + { + "bbox": [ + 47, + 516, + 287, + 634 + ], + "type": "inline_equation", + "content": "r\\geq 1.5" + }, + { + "bbox": [ + 47, + 516, + 287, + 634 + ], + "type": "text", + "content": " Eq.9 is already a good approximation of " + }, + { + "bbox": [ + 47, + 516, + 287, + 634 + ], + "type": "inline_equation", + "content": "f_{\\mathrm{rad}}(r,\\theta)" + }, + { + "bbox": [ + 47, + 516, + 287, + 634 + ], + "type": "text", + "content": " as " + }, + { + "bbox": [ + 47, + 516, + 287, + 634 + ], + "type": "inline_equation", + "content": "\\frac{\\frac{1}{\\pi}\\frac{4}{(4r)^2 + 1}}{f_{\\mathrm{rad}}(r,\\theta)}" + }, + { + "bbox": [ + 47, + 516, + 287, + 634 + ], + "type": "text", + "content": " lies consistently between 0.93 and 1.08 for " + }, + { + "bbox": [ + 47, + 516, + 287, + 634 + ], + "type": "inline_equation", + "content": "r\\geq 1.5" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 47, + 635, + 286, + 658 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 635, + 286, + 658 + ], + "spans": [ + { + "bbox": [ + 47, + 635, + 286, + 658 + ], + "type": "text", + "content": "Eq.9 creates a route to transform " + }, + { + "bbox": [ + 47, + 635, + 286, + 658 + ], + "type": "inline_equation", + "content": "r" + }, + { + "bbox": [ + 47, + 635, + 286, + 658 + ], + "type": "text", + "content": " to a bounded space. To achieve the goal, we introduce a new function " + }, + { + "bbox": [ + 47, + 635, + 286, + 658 + ], + "type": "inline_equation", + "content": "\\psi(r)" + }, + { + "bbox": [ + 47, + 635, + 286, + 658 + ], + "type": "text", + "content": ":" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 130, + 667, + 287, + 681 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 667, + 287, + 681 + ], + "spans": [ + { + "bbox": [ + 130, + 667, + 287, + 681 + ], + "type": "interline_equation", + "content": "\\psi (r) = \\tan^ {- 1} (4 r) \\tag {10}", + "image_path": "77bf5294e3d3397acbe1c9981d74a1d8874e027921c82f8aec76da4e3fbec372.jpg" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 47, + 690, + 287, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 690, + 287, + 715 + ], + "spans": [ + { + "bbox": [ + 47, + 690, + 287, + 715 + ], + "type": "text", + "content": "Unlike " + }, + { + "bbox": [ + 47, + 690, + 287, + 715 + ], + "type": "inline_equation", + "content": "r" + }, + { + "bbox": [ + 47, + 690, + 287, + 715 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 47, + 690, + 287, + 715 + ], + "type": "inline_equation", + "content": "\\psi" + }, + { + "bbox": [ + 47, + 690, + 287, + 715 + ], + "type": "text", + "content": " is defined in a bounded space since " + }, + { + "bbox": [ + 47, + 690, + 287, + 715 + ], + "type": "inline_equation", + "content": "\\psi \\in [0, \\frac{\\pi}{2}]" + }, + { + "bbox": [ + 47, + 690, + 287, + 715 + ], + "type": "text", + "content": ". We denote the PDF along " + }, + { + "bbox": [ + 47, + 690, + 287, + 715 + ], + "type": "inline_equation", + "content": "\\psi" + }, + { + "bbox": [ + 47, + 690, + 287, + 715 + ], + "type": "text", + "content": " as " + }, + { + "bbox": [ + 47, + 690, + 287, + 715 + ], + "type": "inline_equation", + "content": "h_{\\mathrm{rad}}(\\psi)" + }, + { + "bbox": [ + 47, + 690, + 287, + 715 + ], + "type": "text", + "content": ", which is" + } + ] + } + ], + "index": 13 + }, + { + "type": "image", + "bbox": [ + 325, + 74, + 420, + 179 + ], + "blocks": [ + { + "bbox": [ + 325, + 74, + 420, + 179 + ], + "lines": [ + { + "bbox": [ + 325, + 74, + 420, + 179 + ], + "spans": [ + { + "bbox": [ + 325, + 74, + 420, + 179 + ], + "type": "image", + "image_path": "cf5df42cbcdb12186be01f3c61c73856cc3c5fb4d22b9f1dfb9d5f6d16bcaef5.jpg" + } + ] + } + ], + "index": 14, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 364, + 183, + 376, + 194 + ], + "lines": [ + { + "bbox": [ + 364, + 183, + 376, + 194 + ], + "spans": [ + { + "bbox": [ + 364, + 183, + 376, + 194 + ], + "type": "text", + "content": "(a)" + } + ] + } + ], + "index": 15, + "angle": 0, + "type": "image_caption" + } + ], + "index": 14 + }, + { + "type": "image", + "bbox": [ + 425, + 71, + 534, + 180 + ], + "blocks": [ + { + "bbox": [ + 425, + 71, + 534, + 180 + ], + "lines": [ + { + "bbox": [ + 425, + 71, + 534, + 180 + ], + "spans": [ + { + "bbox": [ + 425, + 71, + 534, + 180 + ], + "type": "image", + "image_path": "17141d096b5248c572f005a65b8098d255e036018a31aae4b1fb0e4dd5a66a13.jpg" + } + ] + } + ], + "index": 16, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 475, + 182, + 488, + 194 + ], + "lines": [ + { + "bbox": [ + 475, + 182, + 488, + 194 + ], + "spans": [ + { + "bbox": [ + 475, + 182, + 488, + 194 + ], + "type": "text", + "content": "(b)" + } + ] + } + ], + "index": 17, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 305, + 203, + 545, + 237 + ], + "lines": [ + { + "bbox": [ + 305, + 203, + 545, + 237 + ], + "spans": [ + { + "bbox": [ + 305, + 203, + 545, + 237 + ], + "type": "text", + "content": "Figure 7. Depiction of DiskVPS's Hough space. (a) The structure of the Hough space; (b) The parameters of a Hough cell in the " + }, + { + "bbox": [ + 305, + 203, + 545, + 237 + ], + "type": "inline_equation", + "content": "i" + }, + { + "bbox": [ + 305, + 203, + 545, + 237 + ], + "type": "text", + "content": "th subspace." + } + ] + } + ], + "index": 18, + "angle": 0, + "type": "image_caption" + } + ], + "index": 16 + }, + { + "bbox": [ + 306, + 257, + 501, + 270 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 257, + 501, + 270 + ], + "spans": [ + { + "bbox": [ + 306, + 257, + 501, + 270 + ], + "type": "text", + "content": "related to " + }, + { + "bbox": [ + 306, + 257, + 501, + 270 + ], + "type": "inline_equation", + "content": "g_{\\mathrm{rad}}(r)" + }, + { + "bbox": [ + 306, + 257, + 501, + 270 + ], + "type": "text", + "content": " through the following formula:" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 383, + 277, + 545, + 298 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 383, + 277, + 545, + 298 + ], + "spans": [ + { + "bbox": [ + 383, + 277, + 545, + 298 + ], + "type": "interline_equation", + "content": "h _ {\\mathrm {r a d}} (\\psi) = g _ {\\mathrm {r a d}} (r) \\left| \\frac {d r}{d \\psi} \\right| \\tag {11}", + "image_path": "9b42459182d3b4fa97e8da47e13efe5b5803feda7f5ee637ef154189f3d8826c.jpg" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 306, + 307, + 545, + 337 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 307, + 545, + 337 + ], + "spans": [ + { + "bbox": [ + 306, + 307, + 545, + 337 + ], + "type": "text", + "content": "Combining Eq.9, Eq.10 and Eq.11, and noting " + }, + { + "bbox": [ + 306, + 307, + 545, + 337 + ], + "type": "inline_equation", + "content": "\\frac{d(\\tan^{-1}x)}{dx} = \\frac{1}{x^2 + 1}" + }, + { + "bbox": [ + 306, + 307, + 545, + 337 + ], + "type": "text", + "content": ", we have:" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 345, + 343, + 545, + 366 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 345, + 343, + 545, + 366 + ], + "spans": [ + { + "bbox": [ + 345, + 343, + 545, + 366 + ], + "type": "interline_equation", + "content": "h _ {\\mathrm {r a d}} (\\psi) \\approx \\frac {1}{\\pi} \\frac {4}{(4 r) ^ {2} + 1} \\div \\frac {4}{(4 r) ^ {2} + 1} = \\frac {1}{\\pi} \\tag {12}", + "image_path": "7d066eaa72266f05ea89d6d13e904d9cdd7c088c1923eed0ac06155c2e6e3f4a.jpg" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 305, + 373, + 545, + 398 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 373, + 545, + 398 + ], + "spans": [ + { + "bbox": [ + 305, + 373, + 545, + 398 + ], + "type": "text", + "content": "Eq.12 indicates that the PDF along " + }, + { + "bbox": [ + 305, + 373, + 545, + 398 + ], + "type": "inline_equation", + "content": "\\psi" + }, + { + "bbox": [ + 305, + 373, + 545, + 398 + ], + "type": "text", + "content": " is approximately constant " + }, + { + "bbox": [ + 305, + 373, + 545, + 398 + ], + "type": "inline_equation", + "content": "(= \\frac{1}{\\pi})" + }, + { + "bbox": [ + 305, + 373, + 545, + 398 + ], + "type": "text", + "content": " in the space of " + }, + { + "bbox": [ + 305, + 373, + 545, + 398 + ], + "type": "inline_equation", + "content": "(\\psi, \\theta)" + }, + { + "bbox": [ + 305, + 373, + 545, + 398 + ], + "type": "text", + "content": " (for " + }, + { + "bbox": [ + 305, + 373, + 545, + 398 + ], + "type": "inline_equation", + "content": "r \\geq 1.5" + }, + { + "bbox": [ + 305, + 373, + 545, + 398 + ], + "type": "text", + "content": ")." + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 306, + 404, + 466, + 418 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 404, + 466, + 418 + ], + "spans": [ + { + "bbox": [ + 306, + 404, + 466, + 418 + ], + "type": "text", + "content": "2.2. Construction of Hough Space" + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 304, + 423, + 545, + 517 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 423, + 545, + 517 + ], + "spans": [ + { + "bbox": [ + 304, + 423, + 545, + 517 + ], + "type": "text", + "content": "Structurally, the Hough space is partitioned into a set of " + }, + { + "bbox": [ + 304, + 423, + 545, + 517 + ], + "type": "inline_equation", + "content": "M" + }, + { + "bbox": [ + 304, + 423, + 545, + 517 + ], + "type": "text", + "content": " concentric ring-shaped subspaces, each of which is further partitioned into a set of at most " + }, + { + "bbox": [ + 304, + 423, + 545, + 517 + ], + "type": "inline_equation", + "content": "N" + }, + { + "bbox": [ + 304, + 423, + 545, + 517 + ], + "type": "text", + "content": " fan-shaped cells, as shown in Fig.7(a). The principle is to ensure that each subspace has a roughly equal probability of intersecting the extension line of the random edge along the radial direction and each cell has a roughly equal probability of intersecting the line in the tangential direction." + } + ] + } + ], + "index": 25 + }, + { + "bbox": [ + 304, + 517, + 545, + 606 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 517, + 545, + 606 + ], + "spans": [ + { + "bbox": [ + 304, + 517, + 545, + 606 + ], + "type": "text", + "content": "Our first goal is to obtain each subspace's inner radius and radial span using the principle above. We denote the inner radius and the radial span of the " + }, + { + "bbox": [ + 304, + 517, + 545, + 606 + ], + "type": "inline_equation", + "content": "i" + }, + { + "bbox": [ + 304, + 517, + 545, + 606 + ], + "type": "text", + "content": "th subspace as " + }, + { + "bbox": [ + 304, + 517, + 545, + 606 + ], + "type": "inline_equation", + "content": "r_i" + }, + { + "bbox": [ + 304, + 517, + 545, + 606 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 304, + 517, + 545, + 606 + ], + "type": "inline_equation", + "content": "\\delta r_i" + }, + { + "bbox": [ + 304, + 517, + 545, + 606 + ], + "type": "text", + "content": " respectively, where " + }, + { + "bbox": [ + 304, + 517, + 545, + 606 + ], + "type": "inline_equation", + "content": "i = 1,2,\\ldots,M" + }, + { + "bbox": [ + 304, + 517, + 545, + 606 + ], + "type": "text", + "content": " is the order number of the subspace counting from the innermost outwards. Noting the outer radius of the " + }, + { + "bbox": [ + 304, + 517, + 545, + 606 + ], + "type": "inline_equation", + "content": "i" + }, + { + "bbox": [ + 304, + 517, + 545, + 606 + ], + "type": "text", + "content": "th subspace is the inner radius of the " + }, + { + "bbox": [ + 304, + 517, + 545, + 606 + ], + "type": "inline_equation", + "content": "i + 1" + }, + { + "bbox": [ + 304, + 517, + 545, + 606 + ], + "type": "text", + "content": "th subspace, there exist the following relations throughout the subspaces:" + } + ] + } + ], + "index": 26 + }, + { + "bbox": [ + 396, + 616, + 545, + 680 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 396, + 616, + 545, + 680 + ], + "spans": [ + { + "bbox": [ + 396, + 616, + 545, + 680 + ], + "type": "interline_equation", + "content": "\\begin{array}{l} r _ {1} = 0 \\\\ r _ {2} = r _ {1} + \\delta r _ {1} \\\\ \\dots \\tag {13} \\\\ r _ {i + 1} = r _ {i} + \\delta r _ {i} \\\\ \\dots \\\\ \\end{array}", + "image_path": "34e0ec5bab2ec7d88430991c52f47d959afcf3910d053d434bbd71a092685858.jpg" + } + ] + } + ], + "index": 27 + }, + { + "bbox": [ + 306, + 689, + 545, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 689, + 545, + 713 + ], + "spans": [ + { + "bbox": [ + 306, + 689, + 545, + 713 + ], + "type": "text", + "content": "Since a full radial line intersects the extension line of the random edge exactly once (the special case where the 2" + } + ] + } + ], + "index": 28 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "text", + "content": "27052" + } + ] + } + ], + "index": 29 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 3 + }, + { + "para_blocks": [ + { + "bbox": [ + 46, + 72, + 287, + 120 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 72, + 287, + 120 + ], + "spans": [ + { + "bbox": [ + 46, + 72, + 287, + 120 + ], + "type": "text", + "content": "lines are parallel and do not intersect has infinitesimal probability and is therefore ignored), the probability of the radial span of each subspace intersecting the extension line of the random edge is " + }, + { + "bbox": [ + 46, + 72, + 287, + 120 + ], + "type": "inline_equation", + "content": "\\frac{1}{M}" + }, + { + "bbox": [ + 46, + 72, + 287, + 120 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 46, + 120, + 290, + 236 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 120, + 290, + 236 + ], + "spans": [ + { + "bbox": [ + 46, + 120, + 290, + 236 + ], + "type": "text", + "content": "In addition to setting the inner radius and the radial span of the " + }, + { + "bbox": [ + 46, + 120, + 290, + 236 + ], + "type": "inline_equation", + "content": "i" + }, + { + "bbox": [ + 46, + 120, + 290, + 236 + ], + "type": "text", + "content": "th subspace as " + }, + { + "bbox": [ + 46, + 120, + 290, + 236 + ], + "type": "inline_equation", + "content": "r_i" + }, + { + "bbox": [ + 46, + 120, + 290, + 236 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 46, + 120, + 290, + 236 + ], + "type": "inline_equation", + "content": "\\delta r_i" + }, + { + "bbox": [ + 46, + 120, + 290, + 236 + ], + "type": "text", + "content": " for near-field subspaces, we also set the inner radius and the radial span of the subspace as " + }, + { + "bbox": [ + 46, + 120, + 290, + 236 + ], + "type": "inline_equation", + "content": "\\psi_i = \\tan^{-1}(4r_i)" + }, + { + "bbox": [ + 46, + 120, + 290, + 236 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 46, + 120, + 290, + 236 + ], + "type": "inline_equation", + "content": "\\delta \\psi_i" + }, + { + "bbox": [ + 46, + 120, + 290, + 236 + ], + "type": "text", + "content": " for far-field subspaces. Assuming that the PDF along the radial span of the " + }, + { + "bbox": [ + 46, + 120, + 290, + 236 + ], + "type": "inline_equation", + "content": "i" + }, + { + "bbox": [ + 46, + 120, + 290, + 236 + ], + "type": "text", + "content": "th space is constant (" + }, + { + "bbox": [ + 46, + 120, + 290, + 236 + ], + "type": "inline_equation", + "content": "= g_{\\mathrm{rad}(r_i)}" + }, + { + "bbox": [ + 46, + 120, + 290, + 236 + ], + "type": "text", + "content": " for near-field subspace and " + }, + { + "bbox": [ + 46, + 120, + 290, + 236 + ], + "type": "inline_equation", + "content": "= h_{\\mathrm{rad}}(\\psi_i)" + }, + { + "bbox": [ + 46, + 120, + 290, + 236 + ], + "type": "text", + "content": " for far-field subspace), then the probability " + }, + { + "bbox": [ + 46, + 120, + 290, + 236 + ], + "type": "inline_equation", + "content": "\\frac{1}{M}" + }, + { + "bbox": [ + 46, + 120, + 290, + 236 + ], + "type": "text", + "content": " must match with " + }, + { + "bbox": [ + 46, + 120, + 290, + 236 + ], + "type": "inline_equation", + "content": "g_{\\mathrm{rad}}(r_i)\\delta r_i" + }, + { + "bbox": [ + 46, + 120, + 290, + 236 + ], + "type": "text", + "content": " if the " + }, + { + "bbox": [ + 46, + 120, + 290, + 236 + ], + "type": "inline_equation", + "content": "i" + }, + { + "bbox": [ + 46, + 120, + 290, + 236 + ], + "type": "text", + "content": "th subspace is in the near field or " + }, + { + "bbox": [ + 46, + 120, + 290, + 236 + ], + "type": "inline_equation", + "content": "h_{\\mathrm{rad}}(\\psi_i)\\delta \\psi_i" + }, + { + "bbox": [ + 46, + 120, + 290, + 236 + ], + "type": "text", + "content": " if the " + }, + { + "bbox": [ + 46, + 120, + 290, + 236 + ], + "type": "inline_equation", + "content": "i" + }, + { + "bbox": [ + 46, + 120, + 290, + 236 + ], + "type": "text", + "content": "th subspace is in the far field. Since " + }, + { + "bbox": [ + 46, + 120, + 290, + 236 + ], + "type": "inline_equation", + "content": "h_{\\mathrm{rad}}(\\psi) \\approx \\frac{1}{\\pi}" + }, + { + "bbox": [ + 46, + 120, + 290, + 236 + ], + "type": "text", + "content": " (see Eq.12), we have:" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 104, + 243, + 287, + 269 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 104, + 243, + 287, + 269 + ], + "spans": [ + { + "bbox": [ + 104, + 243, + 287, + 269 + ], + "type": "interline_equation", + "content": "\\delta r _ {i} = \\frac {1}{M g _ {\\mathrm {r a d}} \\left(r _ {i}\\right)} \\quad \\text {f o r} r _ {i} \\leq 1. 5 \\tag {14}", + "image_path": "93f19279e2014bb1bb7e6c6874b36d1f9a04beddd4989f4f5e5d3d9372d7910a.jpg" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 106, + 267, + 228, + 286 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 106, + 267, + 228, + 286 + ], + "spans": [ + { + "bbox": [ + 106, + 267, + 228, + 286 + ], + "type": "interline_equation", + "content": "\\delta \\psi_ {i} = \\frac {\\pi}{M} \\quad \\text {f o r} r _ {i} > 1. 5", + "image_path": "4cdf8a7f5c645f1c504989e61e83cea4a1eeca952e70368fc0adfb770e7a0313.jpg" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 47, + 297, + 287, + 346 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 297, + 287, + 346 + ], + "spans": [ + { + "bbox": [ + 47, + 297, + 287, + 346 + ], + "type": "text", + "content": "Combining " + }, + { + "bbox": [ + 47, + 297, + 287, + 346 + ], + "type": "inline_equation", + "content": "\\delta \\psi_{i} = \\psi_{i + 1} - \\psi_{i} = \\tan^{-1}(4r_{i + 1}) - \\tan^{-1}(4r_{i})" + }, + { + "bbox": [ + 47, + 297, + 287, + 346 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 47, + 297, + 287, + 346 + ], + "type": "inline_equation", + "content": "\\delta r_i = r_{i + 1} - r_i" + }, + { + "bbox": [ + 47, + 297, + 287, + 346 + ], + "type": "text", + "content": ", the 2nd term of Eq.14 turns into " + }, + { + "bbox": [ + 47, + 297, + 287, + 346 + ], + "type": "inline_equation", + "content": "\\delta r_{i} = \\frac{\\tan(\\frac{\\pi}{M} + \\tan^{-1}4r_{i})}{4} -r_{i}" + }, + { + "bbox": [ + 47, + 297, + 287, + 346 + ], + "type": "text", + "content": ". And Eq.14 can be rewritten as:" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 78, + 354, + 287, + 399 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 78, + 354, + 287, + 399 + ], + "spans": [ + { + "bbox": [ + 78, + 354, + 287, + 399 + ], + "type": "interline_equation", + "content": "\\delta r _ {i} = \\left\\{ \\begin{array}{l l} \\frac {1}{M g _ {\\text {r a d}} \\left(r _ {i}\\right)} & \\text {f o r} r _ {i} \\leq 1. 5 \\\\ \\frac {\\tan \\left(\\frac {\\pi}{M} + \\tan^ {- 1} 4 r _ {i}\\right)}{4} - r _ {i} & \\text {f o r} r _ {i} > 1. 5 \\end{array} \\right. \\tag {15}", + "image_path": "cd4f4bfdc8035cb74f6cd4bb9ba180fa1ec5d7a33e043f240906d107b21cfa52.jpg" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 46, + 409, + 287, + 516 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 409, + 287, + 516 + ], + "spans": [ + { + "bbox": [ + 46, + 409, + 287, + 516 + ], + "type": "text", + "content": "The Hough space partitioning starts from " + }, + { + "bbox": [ + 46, + 409, + 287, + 516 + ], + "type": "inline_equation", + "content": "i = 1" + }, + { + "bbox": [ + 46, + 409, + 287, + 516 + ], + "type": "text", + "content": ". Eq.13 and Eq.15 work together to obtain the inner radius " + }, + { + "bbox": [ + 46, + 409, + 287, + 516 + ], + "type": "inline_equation", + "content": "r_i" + }, + { + "bbox": [ + 46, + 409, + 287, + 516 + ], + "type": "text", + "content": " and the radial span " + }, + { + "bbox": [ + 46, + 409, + 287, + 516 + ], + "type": "inline_equation", + "content": "\\delta r_i" + }, + { + "bbox": [ + 46, + 409, + 287, + 516 + ], + "type": "text", + "content": " of the " + }, + { + "bbox": [ + 46, + 409, + 287, + 516 + ], + "type": "inline_equation", + "content": "i" + }, + { + "bbox": [ + 46, + 409, + 287, + 516 + ], + "type": "text", + "content": "th subspace iteratively (" + }, + { + "bbox": [ + 46, + 409, + 287, + 516 + ], + "type": "inline_equation", + "content": "i = i + 1" + }, + { + "bbox": [ + 46, + 409, + 287, + 516 + ], + "type": "text", + "content": " after each iteration) until the " + }, + { + "bbox": [ + 46, + 409, + 287, + 516 + ], + "type": "inline_equation", + "content": "i" + }, + { + "bbox": [ + 46, + 409, + 287, + 516 + ], + "type": "text", + "content": "th subspace comes to the Hough space boundary that is indicated by " + }, + { + "bbox": [ + 46, + 409, + 287, + 516 + ], + "type": "inline_equation", + "content": "\\psi_{i+1} \\geq \\frac{\\pi}{2}" + }, + { + "bbox": [ + 46, + 409, + 287, + 516 + ], + "type": "text", + "content": " or " + }, + { + "bbox": [ + 46, + 409, + 287, + 516 + ], + "type": "inline_equation", + "content": "\\tan^{-1}(4r_i) + \\frac{\\pi}{M} \\geq \\frac{\\pi}{2}" + }, + { + "bbox": [ + 46, + 409, + 287, + 516 + ], + "type": "text", + "content": ". The final count of the subspaces is denoted as " + }, + { + "bbox": [ + 46, + 409, + 287, + 516 + ], + "type": "inline_equation", + "content": "M^* = i" + }, + { + "bbox": [ + 46, + 409, + 287, + 516 + ], + "type": "text", + "content": " at the end of the last iteration. Since the PDF approximations are used in the Hough space partitioning, " + }, + { + "bbox": [ + 46, + 409, + 287, + 516 + ], + "type": "inline_equation", + "content": "M^*" + }, + { + "bbox": [ + 46, + 409, + 287, + 516 + ], + "type": "text", + "content": " is generally not equal, but very close, to " + }, + { + "bbox": [ + 46, + 409, + 287, + 516 + ], + "type": "inline_equation", + "content": "M" + }, + { + "bbox": [ + 46, + 409, + 287, + 516 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 46, + 517, + 288, + 685 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 517, + 288, + 685 + ], + "spans": [ + { + "bbox": [ + 46, + 517, + 288, + 685 + ], + "type": "text", + "content": "Afterward, each subspace is partitioned into multiple equal-angle-span cells. For the " + }, + { + "bbox": [ + 46, + 517, + 288, + 685 + ], + "type": "inline_equation", + "content": "i" + }, + { + "bbox": [ + 46, + 517, + 288, + 685 + ], + "type": "text", + "content": "th subspace, the angle span of a cell is denoted as " + }, + { + "bbox": [ + 46, + 517, + 288, + 685 + ], + "type": "inline_equation", + "content": "\\delta \\theta_{i}" + }, + { + "bbox": [ + 46, + 517, + 288, + 685 + ], + "type": "text", + "content": " (" + }, + { + "bbox": [ + 46, + 517, + 288, + 685 + ], + "type": "inline_equation", + "content": "i = 1,2,\\ldots M^{*}" + }, + { + "bbox": [ + 46, + 517, + 288, + 685 + ], + "type": "text", + "content": "), as shown in Fig.7(b). We observe that the inner circle of the outermost (the " + }, + { + "bbox": [ + 46, + 517, + 288, + 685 + ], + "type": "inline_equation", + "content": "M^{*}" + }, + { + "bbox": [ + 46, + 517, + 288, + 685 + ], + "type": "text", + "content": "th) subspace intersects the extension line of the random edge exactly twice. In the meantime, the very circle is partitioned into " + }, + { + "bbox": [ + 46, + 517, + 288, + 685 + ], + "type": "inline_equation", + "content": "N" + }, + { + "bbox": [ + 46, + 517, + 288, + 685 + ], + "type": "text", + "content": " equal-angle-span arcs, each of which forms the inner arc of a cell. Consequently, the probability of the inner arc of a cell intersecting the random edge's extension line is " + }, + { + "bbox": [ + 46, + 517, + 288, + 685 + ], + "type": "inline_equation", + "content": "\\frac{2}{N}" + }, + { + "bbox": [ + 46, + 517, + 288, + 685 + ], + "type": "text", + "content": ". To make sure that each cell in any subspace gets the same probability " + }, + { + "bbox": [ + 46, + 517, + 288, + 685 + ], + "type": "inline_equation", + "content": "\\frac{2}{N}" + }, + { + "bbox": [ + 46, + 517, + 288, + 685 + ], + "type": "text", + "content": " of intersecting the random edge's extension line, this probability must match the probability obtained through the product of the cell's inner arc length (" + }, + { + "bbox": [ + 46, + 517, + 288, + 685 + ], + "type": "inline_equation", + "content": "= \\delta \\theta_{i}r_{i}" + }, + { + "bbox": [ + 46, + 517, + 288, + 685 + ], + "type": "text", + "content": ") and the PDF along the arc (" + }, + { + "bbox": [ + 46, + 517, + 288, + 685 + ], + "type": "inline_equation", + "content": "= g_{\\mathrm{tag}}(r_i)" + }, + { + "bbox": [ + 46, + 517, + 288, + 685 + ], + "type": "text", + "content": ") throughout the entire Hough space. Accordingly, we have:" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 130, + 693, + 287, + 717 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 693, + 287, + 717 + ], + "spans": [ + { + "bbox": [ + 130, + 693, + 287, + 717 + ], + "type": "interline_equation", + "content": "\\delta \\theta_ {i} = \\frac {2}{N r _ {i} g _ {\\text {t a g}} \\left(r _ {i}\\right)} \\tag {16}", + "image_path": "396a5106a3b28f63fae85b82aa66329b9c614ae590f95d245e72e92d98dcb48b.jpg" + } + ] + } + ], + "index": 8 + }, + { + "type": "image", + "bbox": [ + 323, + 75, + 529, + 220 + ], + "blocks": [ + { + "bbox": [ + 323, + 75, + 529, + 220 + ], + "lines": [ + { + "bbox": [ + 323, + 75, + 529, + 220 + ], + "spans": [ + { + "bbox": [ + 323, + 75, + 529, + 220 + ], + "type": "image", + "image_path": "7bb5ef4a48a765b106c5788f8f1a7a43d46679b255cc6bc5031e7b72b91be9a3.jpg" + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 305, + 236, + 546, + 268 + ], + "lines": [ + { + "bbox": [ + 305, + 236, + 546, + 268 + ], + "spans": [ + { + "bbox": [ + 305, + 236, + 546, + 268 + ], + "type": "text", + "content": "Figure 8. A typical scenario of a voting edge casting its votes in the " + }, + { + "bbox": [ + 305, + 236, + 546, + 268 + ], + "type": "inline_equation", + "content": "i" + }, + { + "bbox": [ + 305, + 236, + 546, + 268 + ], + "type": "text", + "content": "th subspace with 1 or 0 inside each cell referring to the number of votes each cell receives." + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "image_caption" + } + ], + "index": 9 + }, + { + "bbox": [ + 305, + 286, + 545, + 309 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 286, + 545, + 309 + ], + "spans": [ + { + "bbox": [ + 305, + 286, + 545, + 309 + ], + "type": "text", + "content": "The number of cells the " + }, + { + "bbox": [ + 305, + 286, + 545, + 309 + ], + "type": "inline_equation", + "content": "i" + }, + { + "bbox": [ + 305, + 286, + 545, + 309 + ], + "type": "text", + "content": "th subspace is partitioned into is denoted as " + }, + { + "bbox": [ + 305, + 286, + 545, + 309 + ], + "type": "inline_equation", + "content": "n_i (i = 1, 2, \\ldots, M^*)" + }, + { + "bbox": [ + 305, + 286, + 545, + 309 + ], + "type": "text", + "content": ". It is written as:" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 391, + 316, + 545, + 338 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 391, + 316, + 545, + 338 + ], + "spans": [ + { + "bbox": [ + 391, + 316, + 545, + 338 + ], + "type": "interline_equation", + "content": "n _ {i} = \\operatorname {r o u n d} \\left(\\frac {2 \\pi}{\\delta \\theta_ {i}}\\right) \\tag {17}", + "image_path": "5fbc0ba02690f23be0bca763165575a4ae106cc518b76d6db5d7954589eba190.jpg" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 305, + 343, + 545, + 378 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 343, + 545, + 378 + ], + "spans": [ + { + "bbox": [ + 305, + 343, + 545, + 378 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 305, + 343, + 545, + 378 + ], + "type": "inline_equation", + "content": "2\\pi" + }, + { + "bbox": [ + 305, + 343, + 545, + 378 + ], + "type": "text", + "content": " is the angle span of the " + }, + { + "bbox": [ + 305, + 343, + 545, + 378 + ], + "type": "inline_equation", + "content": "i" + }, + { + "bbox": [ + 305, + 343, + 545, + 378 + ], + "type": "text", + "content": "th subspace, and " + }, + { + "bbox": [ + 305, + 343, + 545, + 378 + ], + "type": "inline_equation", + "content": "\\delta \\theta_{i}" + }, + { + "bbox": [ + 305, + 343, + 545, + 378 + ], + "type": "text", + "content": " is the angle span of a cell in the " + }, + { + "bbox": [ + 305, + 343, + 545, + 378 + ], + "type": "inline_equation", + "content": "i" + }, + { + "bbox": [ + 305, + 343, + 545, + 378 + ], + "type": "text", + "content": "th subspace, and round refers to the namesake maths function." + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 305, + 379, + 545, + 402 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 379, + 545, + 402 + ], + "spans": [ + { + "bbox": [ + 305, + 379, + 545, + 402 + ], + "type": "text", + "content": "Finally, " + }, + { + "bbox": [ + 305, + 379, + 545, + 402 + ], + "type": "inline_equation", + "content": "\\delta \\theta_{i}" + }, + { + "bbox": [ + 305, + 379, + 545, + 402 + ], + "type": "text", + "content": " needs to be slightly modified due to " + }, + { + "bbox": [ + 305, + 379, + 545, + 402 + ], + "type": "inline_equation", + "content": "n_i" + }, + { + "bbox": [ + 305, + 379, + 545, + 402 + ], + "type": "text", + "content": "'s discrete nature. Its adjusted form, " + }, + { + "bbox": [ + 305, + 379, + 545, + 402 + ], + "type": "inline_equation", + "content": "\\delta \\theta_{i}^{*}" + }, + { + "bbox": [ + 305, + 379, + 545, + 402 + ], + "type": "text", + "content": ", is written as:" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 405, + 408, + 545, + 430 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 405, + 408, + 545, + 430 + ], + "spans": [ + { + "bbox": [ + 405, + 408, + 545, + 430 + ], + "type": "interline_equation", + "content": "\\delta \\theta_ {i} ^ {*} = \\frac {2 \\pi}{n _ {i}} \\tag {18}", + "image_path": "2a321bf82d9ed2de131264ced1146139ea70be403067175dcb8d194f6339dc5e.jpg" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 305, + 436, + 545, + 497 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 436, + 545, + 497 + ], + "spans": [ + { + "bbox": [ + 305, + 436, + 545, + 497 + ], + "type": "text", + "content": "This way, the " + }, + { + "bbox": [ + 305, + 436, + 545, + 497 + ], + "type": "inline_equation", + "content": "i" + }, + { + "bbox": [ + 305, + 436, + 545, + 497 + ], + "type": "text", + "content": "th subspace is partitioned into " + }, + { + "bbox": [ + 305, + 436, + 545, + 497 + ], + "type": "inline_equation", + "content": "n_i" + }, + { + "bbox": [ + 305, + 436, + 545, + 497 + ], + "type": "text", + "content": " equal-angle-span cells, each covering an angle span of " + }, + { + "bbox": [ + 305, + 436, + 545, + 497 + ], + "type": "inline_equation", + "content": "\\delta \\theta_i^*" + }, + { + "bbox": [ + 305, + 436, + 545, + 497 + ], + "type": "text", + "content": ". Via iterating " + }, + { + "bbox": [ + 305, + 436, + 545, + 497 + ], + "type": "inline_equation", + "content": "i" + }, + { + "bbox": [ + 305, + 436, + 545, + 497 + ], + "type": "text", + "content": " from 1 through " + }, + { + "bbox": [ + 305, + 436, + 545, + 497 + ], + "type": "inline_equation", + "content": "M^*" + }, + { + "bbox": [ + 305, + 436, + 545, + 497 + ], + "type": "text", + "content": ", all subspaces are thus partitioned into approximately equally probabilistic cells in terms of intersecting the extension line of the random edge." + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 306, + 503, + 503, + 517 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 503, + 503, + 517 + ], + "spans": [ + { + "bbox": [ + 306, + 503, + 503, + 517 + ], + "type": "text", + "content": "2.3. Implementation of Hough Transform" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 305, + 522, + 545, + 557 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 522, + 545, + 557 + ], + "spans": [ + { + "bbox": [ + 305, + 522, + 545, + 557 + ], + "type": "text", + "content": "A Hough accumulation map, each of whose elements is mapped from a corresponding Hough cell, is constructed and the magnitude of each element is initialized to zero." + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 305, + 558, + 545, + 581 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 558, + 545, + 581 + ], + "spans": [ + { + "bbox": [ + 305, + 558, + 545, + 581 + ], + "type": "text", + "content": "The VP detection consists of a voting process and an iterative VP detection process." + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 304, + 582, + 545, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 582, + 545, + 714 + ], + "spans": [ + { + "bbox": [ + 304, + 582, + 545, + 714 + ], + "type": "text", + "content": "In the voting process, each edge casts its votes based on whether its extension line intersects any of the subspaces, and if yes then which specific cell(s) of the subspace the line intersects, with each intersecting cell receiving exactly 1 vote. Assume the extension line of the voting edge is defined by " + }, + { + "bbox": [ + 304, + 582, + 545, + 714 + ], + "type": "inline_equation", + "content": "d" + }, + { + "bbox": [ + 304, + 582, + 545, + 714 + ], + "type": "text", + "content": ", the distance from " + }, + { + "bbox": [ + 304, + 582, + 545, + 714 + ], + "type": "inline_equation", + "content": "O" + }, + { + "bbox": [ + 304, + 582, + 545, + 714 + ], + "type": "text", + "content": " to the line, and " + }, + { + "bbox": [ + 304, + 582, + 545, + 714 + ], + "type": "inline_equation", + "content": "\\varphi" + }, + { + "bbox": [ + 304, + 582, + 545, + 714 + ], + "type": "text", + "content": ", the azimuthal angle of " + }, + { + "bbox": [ + 304, + 582, + 545, + 714 + ], + "type": "inline_equation", + "content": "O" + }, + { + "bbox": [ + 304, + 582, + 545, + 714 + ], + "type": "text", + "content": "'s perpendicular foot on the line. For an edge's voting in the " + }, + { + "bbox": [ + 304, + 582, + 545, + 714 + ], + "type": "inline_equation", + "content": "i" + }, + { + "bbox": [ + 304, + 582, + 545, + 714 + ], + "type": "text", + "content": "th subspace, there exist 3 cases, namely " + }, + { + "bbox": [ + 304, + 582, + 545, + 714 + ], + "type": "inline_equation", + "content": "d < r_i" + }, + { + "bbox": [ + 304, + 582, + 545, + 714 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 304, + 582, + 545, + 714 + ], + "type": "inline_equation", + "content": "r_i \\leq d < r_i + \\delta r_i" + }, + { + "bbox": [ + 304, + 582, + 545, + 714 + ], + "type": "text", + "content": ", and " + }, + { + "bbox": [ + 304, + 582, + 545, + 714 + ], + "type": "inline_equation", + "content": "d \\geq r_i + \\delta r_i" + }, + { + "bbox": [ + 304, + 582, + 545, + 714 + ], + "type": "text", + "content": " respectively, where " + }, + { + "bbox": [ + 304, + 582, + 545, + 714 + ], + "type": "inline_equation", + "content": "r_i" + }, + { + "bbox": [ + 304, + 582, + 545, + 714 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 304, + 582, + 545, + 714 + ], + "type": "inline_equation", + "content": "\\delta r_i" + }, + { + "bbox": [ + 304, + 582, + 545, + 714 + ], + "type": "text", + "content": " are the inner radius and radial span of the " + }, + { + "bbox": [ + 304, + 582, + 545, + 714 + ], + "type": "inline_equation", + "content": "i" + }, + { + "bbox": [ + 304, + 582, + 545, + 714 + ], + "type": "text", + "content": "th subspace." + } + ] + } + ], + "index": 20 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "text", + "content": "27053" + } + ] + } + ], + "index": 21 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 4 + }, + { + "para_blocks": [ + { + "bbox": [ + 47, + 72, + 289, + 142 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 72, + 289, + 142 + ], + "spans": [ + { + "bbox": [ + 47, + 72, + 289, + 142 + ], + "type": "text", + "content": "A scenario of the 1st case (" + }, + { + "bbox": [ + 47, + 72, + 289, + 142 + ], + "type": "inline_equation", + "content": "d < r_i" + }, + { + "bbox": [ + 47, + 72, + 289, + 142 + ], + "type": "text", + "content": ") is shown in Fig.8. The extension line of the voting edge intersects the " + }, + { + "bbox": [ + 47, + 72, + 289, + 142 + ], + "type": "inline_equation", + "content": "i" + }, + { + "bbox": [ + 47, + 72, + 289, + 142 + ], + "type": "text", + "content": "th subspace boundary at 4 points, namely " + }, + { + "bbox": [ + 47, + 72, + 289, + 142 + ], + "type": "inline_equation", + "content": "P_1, P_2, P_3" + }, + { + "bbox": [ + 47, + 72, + 289, + 142 + ], + "type": "text", + "content": ", and " + }, + { + "bbox": [ + 47, + 72, + 289, + 142 + ], + "type": "inline_equation", + "content": "P_4" + }, + { + "bbox": [ + 47, + 72, + 289, + 142 + ], + "type": "text", + "content": ". Assume " + }, + { + "bbox": [ + 47, + 72, + 289, + 142 + ], + "type": "inline_equation", + "content": "\\Psi_{\\overrightarrow{OP_1}}" + }, + { + "bbox": [ + 47, + 72, + 289, + 142 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 47, + 72, + 289, + 142 + ], + "type": "inline_equation", + "content": "\\Psi_{\\overrightarrow{OP_2}}" + }, + { + "bbox": [ + 47, + 72, + 289, + 142 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 47, + 72, + 289, + 142 + ], + "type": "inline_equation", + "content": "\\Psi_{\\overrightarrow{OP_3}}" + }, + { + "bbox": [ + 47, + 72, + 289, + 142 + ], + "type": "text", + "content": ", and " + }, + { + "bbox": [ + 47, + 72, + 289, + 142 + ], + "type": "inline_equation", + "content": "\\Psi_{\\overrightarrow{OP_4}}" + }, + { + "bbox": [ + 47, + 72, + 289, + 142 + ], + "type": "text", + "content": " are the azimuthal angles of vectors " + }, + { + "bbox": [ + 47, + 72, + 289, + 142 + ], + "type": "inline_equation", + "content": "\\overrightarrow{OP_1}" + }, + { + "bbox": [ + 47, + 72, + 289, + 142 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 47, + 72, + 289, + 142 + ], + "type": "inline_equation", + "content": "\\overrightarrow{OP_2}" + }, + { + "bbox": [ + 47, + 72, + 289, + 142 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 47, + 72, + 289, + 142 + ], + "type": "inline_equation", + "content": "\\overrightarrow{OP_3}" + }, + { + "bbox": [ + 47, + 72, + 289, + 142 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 47, + 72, + 289, + 142 + ], + "type": "inline_equation", + "content": "\\overrightarrow{OP_4}" + }, + { + "bbox": [ + 47, + 72, + 289, + 142 + ], + "type": "text", + "content": ", respectively. Then we have:" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 108, + 144, + 287, + 200 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 108, + 144, + 287, + 200 + ], + "spans": [ + { + "bbox": [ + 108, + 144, + 287, + 200 + ], + "type": "interline_equation", + "content": "\\left\\{ \\begin{array}{l} \\Psi_ {\\overrightarrow {O P _ {1}}} = \\varphi + \\cos^ {- 1} \\frac {d}{r _ {i} + \\delta r _ {i}} \\\\ \\Psi_ {\\overrightarrow {O P _ {2}}} = \\varphi + \\cos^ {- 1} \\frac {d}{r _ {i}} \\\\ \\Psi_ {\\overrightarrow {O P _ {3}}} = \\varphi - \\cos^ {- 1} \\frac {d}{r _ {i}} \\\\ \\Psi_ {\\overrightarrow {O P _ {4}}} = \\varphi - \\cos^ {- 1} \\frac {d}{r _ {i} + \\delta r _ {i}} \\end{array} \\right. \\tag {19}", + "image_path": "32209fe2d5380f892a4d4feca6259ef87b04b3d00b37d5ceafbe56d587605235.jpg" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 47, + 203, + 289, + 260 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 203, + 289, + 260 + ], + "spans": [ + { + "bbox": [ + 47, + 203, + 289, + 260 + ], + "type": "text", + "content": "As shown in Fig.8, " + }, + { + "bbox": [ + 47, + 203, + 289, + 260 + ], + "type": "inline_equation", + "content": "P_{1}P_{2}" + }, + { + "bbox": [ + 47, + 203, + 289, + 260 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 47, + 203, + 289, + 260 + ], + "type": "inline_equation", + "content": "P_{3}P_{4}" + }, + { + "bbox": [ + 47, + 203, + 289, + 260 + ], + "type": "text", + "content": " are the line's 2 sections intersecting the " + }, + { + "bbox": [ + 47, + 203, + 289, + 260 + ], + "type": "inline_equation", + "content": "i" + }, + { + "bbox": [ + 47, + 203, + 289, + 260 + ], + "type": "text", + "content": "th subspace. Assume the " + }, + { + "bbox": [ + 47, + 203, + 289, + 260 + ], + "type": "inline_equation", + "content": "n_{i}" + }, + { + "bbox": [ + 47, + 203, + 289, + 260 + ], + "type": "text", + "content": " cells of the " + }, + { + "bbox": [ + 47, + 203, + 289, + 260 + ], + "type": "inline_equation", + "content": "i" + }, + { + "bbox": [ + 47, + 203, + 289, + 260 + ], + "type": "text", + "content": "th subspace are labeled 1 through " + }, + { + "bbox": [ + 47, + 203, + 289, + 260 + ], + "type": "inline_equation", + "content": "n_{i}" + }, + { + "bbox": [ + 47, + 203, + 289, + 260 + ], + "type": "text", + "content": " and a set of consecutive cells that intersect " + }, + { + "bbox": [ + 47, + 203, + 289, + 260 + ], + "type": "inline_equation", + "content": "P_{1}P_{2}" + }, + { + "bbox": [ + 47, + 203, + 289, + 260 + ], + "type": "text", + "content": " have the labels numbered " + }, + { + "bbox": [ + 47, + 203, + 289, + 260 + ], + "type": "inline_equation", + "content": "j" + }, + { + "bbox": [ + 47, + 203, + 289, + 260 + ], + "type": "text", + "content": " through " + }, + { + "bbox": [ + 47, + 203, + 289, + 260 + ], + "type": "inline_equation", + "content": "k" + }, + { + "bbox": [ + 47, + 203, + 289, + 260 + ], + "type": "text", + "content": ", then " + }, + { + "bbox": [ + 47, + 203, + 289, + 260 + ], + "type": "inline_equation", + "content": "j" + }, + { + "bbox": [ + 47, + 203, + 289, + 260 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 47, + 203, + 289, + 260 + ], + "type": "inline_equation", + "content": "k" + }, + { + "bbox": [ + 47, + 203, + 289, + 260 + ], + "type": "text", + "content": " are written as:" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 133, + 267, + 287, + 305 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 267, + 287, + 305 + ], + "spans": [ + { + "bbox": [ + 133, + 267, + 287, + 305 + ], + "type": "interline_equation", + "content": "\\left\\{\\begin{array}{l}j = \\left\\lceil \\frac {\\Psi_ {\\overrightarrow {O P _ {2}}}}{\\delta \\theta_ {i} ^ {*}} \\right\\rceil\\\\k = \\left\\lceil \\frac {\\Psi_ {\\overrightarrow {O P _ {1}}}}{\\delta \\theta_ {i} ^ {*}} \\right\\rceil\\end{array}\\right. \\tag {20}", + "image_path": "dff3b917735316dfade6f235b613b4b2848452b57cd260f3745a4260f88ccf6e.jpg" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 47, + 313, + 287, + 336 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 313, + 287, + 336 + ], + "spans": [ + { + "bbox": [ + 47, + 313, + 287, + 336 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 47, + 313, + 287, + 336 + ], + "type": "inline_equation", + "content": "\\delta \\theta_{i}^{*}" + }, + { + "bbox": [ + 47, + 313, + 287, + 336 + ], + "type": "text", + "content": " is defined in Eq.18, and " + }, + { + "bbox": [ + 47, + 313, + 287, + 336 + ], + "type": "inline_equation", + "content": "\\lceil \\rceil" + }, + { + "bbox": [ + 47, + 313, + 287, + 336 + ], + "type": "text", + "content": " refers to maths function ceil." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 47, + 338, + 287, + 384 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 338, + 287, + 384 + ], + "spans": [ + { + "bbox": [ + 47, + 338, + 287, + 384 + ], + "type": "text", + "content": "Similarly, we can obtain the set of consecutive cells intersecting " + }, + { + "bbox": [ + 47, + 338, + 287, + 384 + ], + "type": "inline_equation", + "content": "P_{3}P_{4}" + }, + { + "bbox": [ + 47, + 338, + 287, + 384 + ], + "type": "text", + "content": ". In voting, each cell in the two sets receives 1 vote from the voting edge while the rest of the " + }, + { + "bbox": [ + 47, + 338, + 287, + 384 + ], + "type": "inline_equation", + "content": "n_{i}" + }, + { + "bbox": [ + 47, + 338, + 287, + 384 + ], + "type": "text", + "content": " cells receive no votes." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 47, + 386, + 287, + 422 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 386, + 287, + 422 + ], + "spans": [ + { + "bbox": [ + 47, + 386, + 287, + 422 + ], + "type": "text", + "content": "For the 2nd case, voting is performed similarly to the first, but only the angle span between " + }, + { + "bbox": [ + 47, + 386, + 287, + 422 + ], + "type": "inline_equation", + "content": "\\Psi_{\\overrightarrow{OP_1}}" + }, + { + "bbox": [ + 47, + 386, + 287, + 422 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 47, + 386, + 287, + 422 + ], + "type": "inline_equation", + "content": "\\Psi_{\\overrightarrow{OP_4}}" + }, + { + "bbox": [ + 47, + 386, + 287, + 422 + ], + "type": "text", + "content": " needs to be handled as neither " + }, + { + "bbox": [ + 47, + 386, + 287, + 422 + ], + "type": "inline_equation", + "content": "P_{2}" + }, + { + "bbox": [ + 47, + 386, + 287, + 422 + ], + "type": "text", + "content": " nor " + }, + { + "bbox": [ + 47, + 386, + 287, + 422 + ], + "type": "inline_equation", + "content": "P_{3}" + }, + { + "bbox": [ + 47, + 386, + 287, + 422 + ], + "type": "text", + "content": " exists." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 47, + 422, + 287, + 446 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 422, + 287, + 446 + ], + "spans": [ + { + "bbox": [ + 47, + 422, + 287, + 446 + ], + "type": "text", + "content": "For the 3rd case, the subspace does not intersect the line; thus, no cell receives any vote." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 47, + 447, + 287, + 483 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 447, + 287, + 483 + ], + "spans": [ + { + "bbox": [ + 47, + 447, + 287, + 483 + ], + "type": "text", + "content": "The voting process generally yields a Hough accumulation map with a small number of salient peaks scattered over the map similar to Fig.2(b)." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 47, + 483, + 287, + 543 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 483, + 287, + 543 + ], + "spans": [ + { + "bbox": [ + 47, + 483, + 287, + 543 + ], + "type": "text", + "content": "Then, VP detection is as simple as finding salient peaks in the Hough accumulation map. In practice, at most " + }, + { + "bbox": [ + 47, + 483, + 287, + 543 + ], + "type": "inline_equation", + "content": "N_{\\mathrm{VP}}" + }, + { + "bbox": [ + 47, + 483, + 287, + 543 + ], + "type": "text", + "content": " dominant VPs are detected via at most " + }, + { + "bbox": [ + 47, + 483, + 287, + 543 + ], + "type": "inline_equation", + "content": "N_{\\mathrm{VP}}" + }, + { + "bbox": [ + 47, + 483, + 287, + 543 + ], + "type": "text", + "content": " iterations with each iteration extracting exactly 1 VP. In each iteration, the following steps are taken:" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 47, + 544, + 287, + 567 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 544, + 287, + 567 + ], + "spans": [ + { + "bbox": [ + 47, + 544, + 287, + 567 + ], + "type": "text", + "content": "Step 1: Search the global peak in the current Hough accumulation space." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 47, + 568, + 287, + 591 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 568, + 287, + 591 + ], + "spans": [ + { + "bbox": [ + 47, + 568, + 287, + 591 + ], + "type": "text", + "content": "Step 2: Obtain the location of the peak and set it as the initial position of the current VP." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 47, + 593, + 287, + 640 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 593, + 287, + 640 + ], + "spans": [ + { + "bbox": [ + 47, + 593, + 287, + 640 + ], + "type": "text", + "content": "Step 3: Search all of the inlier edges of the current VP. An edge is regarded as an inlier of the VP only if the angle between its extension line and the line that passes both the midpoint of the edge and the VP is no more than " + }, + { + "bbox": [ + 47, + 593, + 287, + 640 + ], + "type": "inline_equation", + "content": "\\Theta" + }, + { + "bbox": [ + 47, + 593, + 287, + 640 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 47, + 641, + 287, + 688 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 641, + 287, + 688 + ], + "spans": [ + { + "bbox": [ + 47, + 641, + 287, + 688 + ], + "type": "text", + "content": "Step 4: Update the Hough accumulation map by removing all of the votes cast by the inliers of the current VP. This ensures that inliers of the dominant VPs will not affect the subsequent detection of less dominant VPs." + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 47, + 689, + 287, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 689, + 287, + 713 + ], + "spans": [ + { + "bbox": [ + 47, + 689, + 287, + 713 + ], + "type": "text", + "content": "Step 5: If the peak value in the updated accumulation map is at least 3 times as much as the average and the total" + } + ] + } + ], + "index": 14 + }, + { + "type": "image", + "bbox": [ + 321, + 74, + 380, + 133 + ], + "blocks": [ + { + "bbox": [ + 321, + 74, + 380, + 133 + ], + "lines": [ + { + "bbox": [ + 321, + 74, + 380, + 133 + ], + "spans": [ + { + "bbox": [ + 321, + 74, + 380, + 133 + ], + "type": "image", + "image_path": "54e99d8984c49e7841feb5ac62f320c5f60e03c8ef77bb0a0b7a3b566ee561b7.jpg" + } + ] + } + ], + "index": 15, + "angle": 0, + "type": "image_body" + } + ], + "index": 15 + }, + { + "type": "image", + "bbox": [ + 396, + 74, + 455, + 133 + ], + "blocks": [ + { + "bbox": [ + 396, + 74, + 455, + 133 + ], + "lines": [ + { + "bbox": [ + 396, + 74, + 455, + 133 + ], + "spans": [ + { + "bbox": [ + 396, + 74, + 455, + 133 + ], + "type": "image", + "image_path": "0d4cfc94c3f34d16d19216fbbd291954c62918572ea9cdad60cbba5d06706227.jpg" + } + ] + } + ], + "index": 17, + "angle": 0, + "type": "image_body" + } + ], + "index": 17 + }, + { + "type": "image", + "bbox": [ + 471, + 72, + 533, + 134 + ], + "blocks": [ + { + "bbox": [ + 471, + 72, + 533, + 134 + ], + "lines": [ + { + "bbox": [ + 471, + 72, + 533, + 134 + ], + "spans": [ + { + "bbox": [ + 471, + 72, + 533, + 134 + ], + "type": "image", + "image_path": "69e503b3f0d39b524e2cc06fc27f50b365571dd4ecc2d7494726e0e3745f08fe.jpg" + } + ] + } + ], + "index": 19, + "angle": 0, + "type": "image_body" + } + ], + "index": 19 + }, + { + "type": "image", + "bbox": [ + 320, + 154, + 381, + 217 + ], + "blocks": [ + { + "bbox": [ + 339, + 143, + 351, + 153 + ], + "lines": [ + { + "bbox": [ + 339, + 143, + 351, + 153 + ], + "spans": [ + { + "bbox": [ + 339, + 143, + 351, + 153 + ], + "type": "text", + "content": "(a)" + } + ] + } + ], + "index": 16, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 320, + 154, + 381, + 217 + ], + "lines": [ + { + "bbox": [ + 320, + 154, + 381, + 217 + ], + "spans": [ + { + "bbox": [ + 320, + 154, + 381, + 217 + ], + "type": "image", + "image_path": "f5e6e581506cd4cd1c9f95e693bf9cdbceec2df7c1ddb26fe8ab5051e7428679.jpg" + } + ] + } + ], + "index": 21, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 339, + 224, + 351, + 236 + ], + "lines": [ + { + "bbox": [ + 339, + 224, + 351, + 236 + ], + "spans": [ + { + "bbox": [ + 339, + 224, + 351, + 236 + ], + "type": "text", + "content": "(f)" + } + ] + } + ], + "index": 22, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 305, + 245, + 545, + 355 + ], + "lines": [ + { + "bbox": [ + 305, + 245, + 545, + 355 + ], + "spans": [ + { + "bbox": [ + 305, + 245, + 545, + 355 + ], + "type": "text", + "content": "Figure 9. Graphical demonstration of the steps in DiskVPS' VP detection: (a) is the original image; (b) is the detected edge set; (c) is the initial Hough accumulation colormap with its global peak (also the position of the 1st VP) marked in red; (d) is the updated Hough accumulation map after removing the votes cast by the inliers of the 1st VP with 1 detected VP and the updated peak marked in red; (e) is the updated Hough accumulation map after the 2nd iteration with 2 detected VPs and the updated peak marked in red; (f) is the updated Hough accumulation map after the 3rd iteration with 3 detected VPs marked in red." + } + ] + } + ], + "index": 27, + "angle": 0, + "type": "image_caption" + } + ], + "index": 21 + }, + { + "type": "image", + "bbox": [ + 395, + 154, + 457, + 217 + ], + "blocks": [ + { + "bbox": [ + 420, + 143, + 433, + 153 + ], + "lines": [ + { + "bbox": [ + 420, + 143, + 433, + 153 + ], + "spans": [ + { + "bbox": [ + 420, + 143, + 433, + 153 + ], + "type": "text", + "content": "(b)" + } + ] + } + ], + "index": 18, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 395, + 154, + 457, + 217 + ], + "lines": [ + { + "bbox": [ + 395, + 154, + 457, + 217 + ], + "spans": [ + { + "bbox": [ + 395, + 154, + 457, + 217 + ], + "type": "image", + "image_path": "95cda287321aac94349ba9684fdd23d9d505f5fabb1d6cb7701238de7b921853.jpg" + } + ] + } + ], + "index": 23, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 419, + 224, + 432, + 236 + ], + "lines": [ + { + "bbox": [ + 419, + 224, + 432, + 236 + ], + "spans": [ + { + "bbox": [ + 419, + 224, + 432, + 236 + ], + "type": "text", + "content": "(e)" + } + ] + } + ], + "index": 24, + "angle": 0, + "type": "image_caption" + } + ], + "index": 23 + }, + { + "type": "image", + "bbox": [ + 471, + 155, + 533, + 217 + ], + "blocks": [ + { + "bbox": [ + 501, + 143, + 514, + 153 + ], + "lines": [ + { + "bbox": [ + 501, + 143, + 514, + 153 + ], + "spans": [ + { + "bbox": [ + 501, + 143, + 514, + 153 + ], + "type": "text", + "content": "(c)" + } + ] + } + ], + "index": 20, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 471, + 155, + 533, + 217 + ], + "lines": [ + { + "bbox": [ + 471, + 155, + 533, + 217 + ], + "spans": [ + { + "bbox": [ + 471, + 155, + 533, + 217 + ], + "type": "image", + "image_path": "22938e7e0ac996026b6d9b6a8fb5c6f96270ca47eaf70c5ed3930b8648ec7f24.jpg" + } + ] + } + ], + "index": 25, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 500, + 224, + 513, + 236 + ], + "lines": [ + { + "bbox": [ + 500, + 224, + 513, + 236 + ], + "spans": [ + { + "bbox": [ + 500, + 224, + 513, + 236 + ], + "type": "text", + "content": "(d)" + } + ] + } + ], + "index": 26, + "angle": 0, + "type": "image_caption" + } + ], + "index": 25 + }, + { + "bbox": [ + 305, + 376, + 544, + 400 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 376, + 544, + 400 + ], + "spans": [ + { + "bbox": [ + 305, + 376, + 544, + 400 + ], + "type": "text", + "content": "number of detected VPs is less than " + }, + { + "bbox": [ + 305, + 376, + 544, + 400 + ], + "type": "inline_equation", + "content": "N_{\\mathrm{VP}}" + }, + { + "bbox": [ + 305, + 376, + 544, + 400 + ], + "type": "text", + "content": ", then go to Step 1. Otherwise, stop iteration and output all detected VPs." + } + ] + } + ], + "index": 28 + }, + { + "bbox": [ + 304, + 401, + 545, + 605 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 401, + 545, + 605 + ], + "spans": [ + { + "bbox": [ + 304, + 401, + 545, + 605 + ], + "type": "text", + "content": "The VP detection process is demonstrated in Fig.9. The zenith, which is the convergent point of nearly vertical edges, is located at the local maximum around the south pole of the disk-shaped colormap and the inliers of the zenith also lead to the significantly enhanced distribution around the north pole; the " + }, + { + "bbox": [ + 304, + 401, + 545, + 605 + ], + "type": "inline_equation", + "content": "1^{st}" + }, + { + "bbox": [ + 304, + 401, + 545, + 605 + ], + "type": "text", + "content": " horizontal VP at the east pole is formed by nearly parallel edges which also contribute to the enhanced accumulation around the west pole; non-parallel edges form the " + }, + { + "bbox": [ + 304, + 401, + 545, + 605 + ], + "type": "inline_equation", + "content": "2^{nd}" + }, + { + "bbox": [ + 304, + 401, + 545, + 605 + ], + "type": "text", + "content": " horizontal VP and generate a corresponding peak in the interior of the map. Some local maxima are noises caused by combined voting of inliers of different VPs and outliers of the VPs. However, after 3 iterations almost all local maxima unrelated to any valid VPs in the original accumulation map disappear (see Fig. 9(f)). It demonstrates that the vote removal process of the inliers of the detected VPs in Step 4 effectively minimizes false positive detection." + } + ] + } + ], + "index": 29 + }, + { + "bbox": [ + 306, + 619, + 382, + 633 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 619, + 382, + 633 + ], + "spans": [ + { + "bbox": [ + 306, + 619, + 382, + 633 + ], + "type": "text", + "content": "3. Experiment" + } + ] + } + ], + "index": 30 + }, + { + "bbox": [ + 305, + 640, + 545, + 688 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 640, + 545, + 688 + ], + "spans": [ + { + "bbox": [ + 305, + 640, + 545, + 688 + ], + "type": "text", + "content": "The proposed approach is examined and evaluated against the state-of-the-art (SOTA) in VP detection. It was implemented in " + }, + { + "bbox": [ + 305, + 640, + 545, + 688 + ], + "type": "inline_equation", + "content": "\\mathrm{C + + }" + }, + { + "bbox": [ + 305, + 640, + 545, + 688 + ], + "type": "text", + "content": " and tested on a PC powered by an Intel i7 CPU (3.5 GHZ)." + } + ] + } + ], + "index": 31 + }, + { + "bbox": [ + 305, + 689, + 545, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 689, + 545, + 713 + ], + "spans": [ + { + "bbox": [ + 305, + 689, + 545, + 713 + ], + "type": "text", + "content": "DiskVPS was evaluated via 3 benchmark datasets, the York Urban Dataset (YUD [7]), SVVP([27]), and" + } + ] + } + ], + "index": 32 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 318, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 318, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 318, + 757 + ], + "type": "text", + "content": "27054" + } + ] + } + ], + "index": 33 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 5 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 56, + 69, + 108, + 350 + ], + "blocks": [ + { + "bbox": [ + 56, + 69, + 108, + 350 + ], + "lines": [ + { + "bbox": [ + 56, + 69, + 108, + 350 + ], + "spans": [ + { + "bbox": [ + 56, + 69, + 108, + 350 + ], + "type": "image", + "image_path": "be74e2064883f1b56355d02d5b697024eff9531bc97f4d51040ada2f1512dbf1.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 46, + 360, + 287, + 427 + ], + "lines": [ + { + "bbox": [ + 46, + 360, + 287, + 427 + ], + "spans": [ + { + "bbox": [ + 46, + 360, + 287, + 427 + ], + "type": "text", + "content": "Figure 10. Example images from YUD [7] (the 1st row), SVVP[27] (the 3rd row), and SU3 [40] (the 5th row) datasets with each even row immediately below displaying their generated disk-shaped Hough colormaps and the VP detection results. The ground truth VPs are marked with white \"+\" and the detected VPs are marked with red \"×\"." + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "type": "image", + "bbox": [ + 113, + 70, + 164, + 350 + ], + "blocks": [ + { + "bbox": [ + 113, + 70, + 164, + 350 + ], + "lines": [ + { + "bbox": [ + 113, + 70, + 164, + 350 + ], + "spans": [ + { + "bbox": [ + 113, + 70, + 164, + 350 + ], + "type": "image", + "image_path": "0356ac0da5725a5538177d36637d26ff7f5fd4115e2bb9bf291d10d95c7af390.jpg" + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_body" + } + ], + "index": 1 + }, + { + "type": "image", + "bbox": [ + 170, + 70, + 221, + 350 + ], + "blocks": [ + { + "bbox": [ + 170, + 70, + 221, + 350 + ], + "lines": [ + { + "bbox": [ + 170, + 70, + 221, + 350 + ], + "spans": [ + { + "bbox": [ + 170, + 70, + 221, + 350 + ], + "type": "image", + "image_path": "d8db0a2de0ab0692331d601d0eb65bb959c8985fdbe37521994cdc660101da9e.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + } + ], + "index": 2 + }, + { + "type": "image", + "bbox": [ + 228, + 70, + 278, + 350 + ], + "blocks": [ + { + "bbox": [ + 228, + 70, + 278, + 350 + ], + "lines": [ + { + "bbox": [ + 228, + 70, + 278, + 350 + ], + "spans": [ + { + "bbox": [ + 228, + 70, + 278, + 350 + ], + "type": "image", + "image_path": "02588881492730d419bc21fa1d800f353bf60745ac3a642168258aa836c4e5be.jpg" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_body" + } + ], + "index": 3 + }, + { + "bbox": [ + 46, + 448, + 286, + 626 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 448, + 286, + 626 + ], + "spans": [ + { + "bbox": [ + 46, + 448, + 286, + 626 + ], + "type": "text", + "content": "SU3([40]). YUD contains 102 calibrated real-world images of " + }, + { + "bbox": [ + 46, + 448, + 286, + 626 + ], + "type": "inline_equation", + "content": "640 \\times 480" + }, + { + "bbox": [ + 46, + 448, + 286, + 626 + ], + "type": "text", + "content": " pixels. SVVP contains 500 calibrated real-world images of " + }, + { + "bbox": [ + 46, + 448, + 286, + 626 + ], + "type": "inline_equation", + "content": "512 \\times 512" + }, + { + "bbox": [ + 46, + 448, + 286, + 626 + ], + "type": "text", + "content": " pixels, and SU3 contains a photorealistic dataset containing 23,000 calibrated synthetic outdoor images of " + }, + { + "bbox": [ + 46, + 448, + 286, + 626 + ], + "type": "inline_equation", + "content": "512 \\times 512" + }, + { + "bbox": [ + 46, + 448, + 286, + 626 + ], + "type": "text", + "content": " pixels. The commonness of the 3 benchmark datasets is that they all satisfy the Manhattanworld assumption [5] as there exists a triplet of orthogonal VPs in each image. So our experiment is focused on how accurately DiskVPS performs VP detections on the triplets of orthogonal VPs in each dataset compared to the SOTA. Please note that while the images used in the experiment were calibrated, DiskVPS treated those images as uncalibrated and performed VP detection without using any calibration info until checking accuracy against the ground truth." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 46, + 630, + 286, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 630, + 286, + 715 + ], + "spans": [ + { + "bbox": [ + 46, + 630, + 286, + 715 + ], + "type": "text", + "content": "DiskVPS uses edges as input. In preprocessing, there are many algorithms to perform edge detection (see LSD([31]), [20], and [36], etc.). A similar algorithm to LSD was selected to perform edge detection. Table1 lists the 4 parameters and their respective magnitude that our implementation used for evaluation. " + }, + { + "bbox": [ + 46, + 630, + 286, + 715 + ], + "type": "inline_equation", + "content": "M" + }, + { + "bbox": [ + 46, + 630, + 286, + 715 + ], + "type": "text", + "content": " is the number of concentric subspaces that the disk-shaped Hough space is roughly parti" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 304, + 72, + 545, + 156 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 72, + 545, + 156 + ], + "spans": [ + { + "bbox": [ + 304, + 72, + 545, + 156 + ], + "type": "text", + "content": "tioned into, " + }, + { + "bbox": [ + 304, + 72, + 545, + 156 + ], + "type": "inline_equation", + "content": "N" + }, + { + "bbox": [ + 304, + 72, + 545, + 156 + ], + "type": "text", + "content": " is the maximum number of roughly equal-probability cells that each of the subspaces is partitioned into, " + }, + { + "bbox": [ + 304, + 72, + 545, + 156 + ], + "type": "inline_equation", + "content": "N_{\\mathrm{VP}}" + }, + { + "bbox": [ + 304, + 72, + 545, + 156 + ], + "type": "text", + "content": " is the maximum number of dominant VPs that are initially extracted from each image, and " + }, + { + "bbox": [ + 304, + 72, + 545, + 156 + ], + "type": "inline_equation", + "content": "\\Theta" + }, + { + "bbox": [ + 304, + 72, + 545, + 156 + ], + "type": "text", + "content": " is the criterion angle to separate the inliers of a VP from its outliers. The parameters are fixed throughout the experiment despite the differences among the 3 datasets." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 304, + 157, + 545, + 312 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 157, + 545, + 312 + ], + "spans": [ + { + "bbox": [ + 304, + 157, + 545, + 312 + ], + "type": "text", + "content": "Fig.10 lists some example images of the 3 datasets and their VP detection results of our approach. It demonstrates that the disk-shaped Hough accumulation map not only manifests almost all valid VPs but also generates no noticeable spurious VPs. Most detected VPs are either very close to or precisely overlapped with their ground truth counterparts. The only exception happens in the 6th row and the 3rd column, in which the zenith's ground truth lies around the north pole (corresponding to " + }, + { + "bbox": [ + 304, + 157, + 545, + 312 + ], + "type": "inline_equation", + "content": "(0, +\\infty)" + }, + { + "bbox": [ + 304, + 157, + 545, + 312 + ], + "type": "text", + "content": " in " + }, + { + "bbox": [ + 304, + 157, + 545, + 312 + ], + "type": "inline_equation", + "content": "(u, v)" + }, + { + "bbox": [ + 304, + 157, + 545, + 312 + ], + "type": "text", + "content": " coordinates) while its detected location is around the south pole " + }, + { + "bbox": [ + 304, + 157, + 545, + 312 + ], + "type": "inline_equation", + "content": "(0, -\\infty)" + }, + { + "bbox": [ + 304, + 157, + 545, + 312 + ], + "type": "text", + "content": " in " + }, + { + "bbox": [ + 304, + 157, + 545, + 312 + ], + "type": "inline_equation", + "content": "(u, v)" + }, + { + "bbox": [ + 304, + 157, + 545, + 312 + ], + "type": "text", + "content": ". The discrepancy is reconciled as the VP at " + }, + { + "bbox": [ + 304, + 157, + 545, + 312 + ], + "type": "inline_equation", + "content": "(0, +\\infty)" + }, + { + "bbox": [ + 304, + 157, + 545, + 312 + ], + "type": "text", + "content": " and the VP at " + }, + { + "bbox": [ + 304, + 157, + 545, + 312 + ], + "type": "inline_equation", + "content": "(0, -\\infty)" + }, + { + "bbox": [ + 304, + 157, + 545, + 312 + ], + "type": "text", + "content": " in the " + }, + { + "bbox": [ + 304, + 157, + 545, + 312 + ], + "type": "inline_equation", + "content": "(u, v)" + }, + { + "bbox": [ + 304, + 157, + 545, + 312 + ], + "type": "text", + "content": " coordinates are formed by the same group of vertical edges." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 304, + 312, + 545, + 491 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 312, + 545, + 491 + ], + "spans": [ + { + "bbox": [ + 304, + 312, + 545, + 491 + ], + "type": "text", + "content": "In the experiment, DiskVPS is compared with six SOTA algorithms ([11, 16, 26, 27, 34, 39]) using a protocol adopted by [11, 16, 27]. In the protocol, a triple of orthogonal VPs are detected from each image, then 3 angular differences in the camera space between the 3 detected VPs and their respective ground truth are obtained, with the largest of the 3 differences adopted as the angular difference of the image. According to the protocol, the percentage of a dataset's images that have a smaller angular difference than a given threshold is used as the angular accuracy (AA) criterion. In the experiment, AA@3°, AA@5°, and AA@10° represent AA thresholds of 3°, 5°, and 10° respectively. For example, an algorithm achieving 80% on AA@5° with a dataset means that the algorithm achieves an angular difference of 5° or less on 80% of the images in the dataset." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 304, + 492, + 545, + 601 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 492, + 545, + 601 + ], + "spans": [ + { + "bbox": [ + 304, + 492, + 545, + 601 + ], + "type": "text", + "content": "In DiskVPS implementation, a triplet of orthogonal VPs is picked from the initial " + }, + { + "bbox": [ + 304, + 492, + 545, + 601 + ], + "type": "inline_equation", + "content": "N_{\\mathrm{VP}}" + }, + { + "bbox": [ + 304, + 492, + 545, + 601 + ], + "type": "text", + "content": " VPs based on their mutual orthogonality (see [3] for details of VP orthogonality). However, there is a small fraction of images in which only 2 of the 3 orthogonal VPs are captured in the initial " + }, + { + "bbox": [ + 304, + 492, + 545, + 601 + ], + "type": "inline_equation", + "content": "N_{\\mathrm{VP}}" + }, + { + "bbox": [ + 304, + 492, + 545, + 601 + ], + "type": "text", + "content": " VPs. This happens when more dominant non-orthogonal VPs suppress a less salient orthogonal VP and thus miss the initial detection. In the latter case, DiskVPS detects the 3rd orthogonal VP by searching for the local maximum in" + } + ] + } + ], + "index": 10 + }, + { + "type": "table", + "bbox": [ + 380, + 619, + 473, + 692 + ], + "blocks": [ + { + "bbox": [ + 380, + 619, + 473, + 692 + ], + "lines": [ + { + "bbox": [ + 380, + 619, + 473, + 692 + ], + "spans": [ + { + "bbox": [ + 380, + 619, + 473, + 692 + ], + "type": "table", + "html": "
NameMagnitude
M200
N200
NVP4
Θ3°
", + "image_path": "2d064063661c9fb332937969954b02825faa1eee66e16a786737a89ad27b7dfc.jpg" + } + ] + } + ], + "index": 11, + "angle": 0, + "type": "table_body" + } + ], + "index": 11 + }, + { + "bbox": [ + 306, + 700, + 542, + 712 + ], + "lines": [ + { + "bbox": [ + 306, + 700, + 542, + 712 + ], + "spans": [ + { + "bbox": [ + 306, + 700, + 542, + 712 + ], + "type": "text", + "content": "Table 1. Parameters used in the implementation of our algorithm." + } + ] + } + ], + "index": 12, + "angle": 0, + "type": "text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "text", + "content": "27055" + } + ] + } + ], + "index": 13 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 6 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 65, + 70, + 531, + 191 + ], + "blocks": [ + { + "bbox": [ + 65, + 70, + 531, + 191 + ], + "lines": [ + { + "bbox": [ + 65, + 70, + 531, + 191 + ], + "spans": [ + { + "bbox": [ + 65, + 70, + 531, + 191 + ], + "type": "table", + "html": "
Datasets MetricsFPSYUD [7]SVVP[27]SU3 [40]
AA@3°AA@5°AA@10°AA@3°AA@5°AA@10°AA@3°AA@5°
J-Linkage[26]1.060.871.881.527.343.260.882.087.2
Wu et al. [34]3958.373.482.243.860.175.782.387.6
CONSAC [11]3.061.773.384.443.856.569.486.290.4
NeurVPS [39]0.560.871.881.532.845.760.293.996.3
Lin et al. [16]2360.774.386.351.868.782.284.490.6
Tong et al. [27]2565.577.187.451.667.782.691.394.6
Ours (DiskVPS)15874.590.299.058.381.698.484.391.7
", + "image_path": "111051d5aa184f1e371a909c7b6b16e0d9993abf2e60b963bec4294936ac2613.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "table_body" + } + ], + "index": 0 + }, + { + "bbox": [ + 46, + 251, + 289, + 370 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 251, + 289, + 370 + ], + "spans": [ + { + "bbox": [ + 46, + 251, + 289, + 370 + ], + "type": "text", + "content": "the Hough space around the 3rd VP's predicted location derived from the 2 known orthogonal VPs. In the experiment, it was observed that when " + }, + { + "bbox": [ + 46, + 251, + 289, + 370 + ], + "type": "inline_equation", + "content": "N_{\\mathrm{VP}} = 4" + }, + { + "bbox": [ + 46, + 251, + 289, + 370 + ], + "type": "text", + "content": ", DiskVPS successfully captured a triple of orthogonal VPs in the initial " + }, + { + "bbox": [ + 46, + 251, + 289, + 370 + ], + "type": "inline_equation", + "content": "N_{\\mathrm{VP}}" + }, + { + "bbox": [ + 46, + 251, + 289, + 370 + ], + "type": "text", + "content": " VPs in 90 (88% of the total) images in YUD, while it captured 2 of the triplet of orthogonal VPs in 11 (11% of the total) images of YUD in the initial " + }, + { + "bbox": [ + 46, + 251, + 289, + 370 + ], + "type": "inline_equation", + "content": "N_{\\mathrm{VP}}" + }, + { + "bbox": [ + 46, + 251, + 289, + 370 + ], + "type": "text", + "content": ". After executing the 3rd VP detection on the 11 images, 101 of the 102 images in YUD have their respective triples of orthogonal VPs successfully detected." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 46, + 373, + 289, + 494 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 373, + 289, + 494 + ], + "spans": [ + { + "bbox": [ + 46, + 373, + 289, + 494 + ], + "type": "text", + "content": "The comparison results are shown in Table 2. Our approach achieves " + }, + { + "bbox": [ + 46, + 373, + 289, + 494 + ], + "type": "inline_equation", + "content": "74.5\\%" + }, + { + "bbox": [ + 46, + 373, + 289, + 494 + ], + "type": "text", + "content": " at AA@3° on YUD, " + }, + { + "bbox": [ + 46, + 373, + 289, + 494 + ], + "type": "inline_equation", + "content": "9\\%" + }, + { + "bbox": [ + 46, + 373, + 289, + 494 + ], + "type": "text", + "content": " higher than the second best at " + }, + { + "bbox": [ + 46, + 373, + 289, + 494 + ], + "type": "inline_equation", + "content": "65.5\\%" + }, + { + "bbox": [ + 46, + 373, + 289, + 494 + ], + "type": "text", + "content": " achieved by [27], a significant improvement over the SOTA. Similar patterns are observed at AA@5° and AA@10° on YUD and SVVP. Notably, our approach achieves " + }, + { + "bbox": [ + 46, + 373, + 289, + 494 + ], + "type": "inline_equation", + "content": "99.0\\%" + }, + { + "bbox": [ + 46, + 373, + 289, + 494 + ], + "type": "text", + "content": " on YUD and " + }, + { + "bbox": [ + 46, + 373, + 289, + 494 + ], + "type": "inline_equation", + "content": "98.4\\%" + }, + { + "bbox": [ + 46, + 373, + 289, + 494 + ], + "type": "text", + "content": " on SVVP at AA@10° whereas none of any other approaches reaches " + }, + { + "bbox": [ + 46, + 373, + 289, + 494 + ], + "type": "inline_equation", + "content": "90\\%" + }, + { + "bbox": [ + 46, + 373, + 289, + 494 + ], + "type": "text", + "content": " or higher on this respect. Overall, for real-world images, our approach significantly outperforms the SOTA as far as angular accuracy is concerned." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 46, + 495, + 289, + 615 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 495, + 289, + 615 + ], + "spans": [ + { + "bbox": [ + 46, + 495, + 289, + 615 + ], + "type": "text", + "content": "On the synthetic SU3, DiskVPS is also competitive. Note that the number of images of SU3 is roughly 40 times as large as YUD and SVD combined and the two approaches ([27, 39]) that outperformed DiskVPS are both deep-learning (DL) methods. We believe the large size of SU3 provides substantially more training and helps significantly optimize the DL algorithms. Considering that DiskVPS is a non-optimized basic VP detection model involving no calibration information which the DL models require, the accuracy of DiskVPS is still impressive." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 46, + 616, + 289, + 700 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 616, + 289, + 700 + ], + "spans": [ + { + "bbox": [ + 46, + 616, + 289, + 700 + ], + "type": "text", + "content": "Moreover, we increased " + }, + { + "bbox": [ + 46, + 616, + 289, + 700 + ], + "type": "inline_equation", + "content": "M" + }, + { + "bbox": [ + 46, + 616, + 289, + 700 + ], + "type": "text", + "content": " from 200 to 300 and " + }, + { + "bbox": [ + 46, + 616, + 289, + 700 + ], + "type": "inline_equation", + "content": "N" + }, + { + "bbox": [ + 46, + 616, + 289, + 700 + ], + "type": "text", + "content": " from 100 to 180 (both at stepsize of 5) and observed the angular accuracy (AA) distribution at AA@" + }, + { + "bbox": [ + 46, + 616, + 289, + 700 + ], + "type": "inline_equation", + "content": "3^{\\circ}" + }, + { + "bbox": [ + 46, + 616, + 289, + 700 + ], + "type": "text", + "content": ", AA@" + }, + { + "bbox": [ + 46, + 616, + 289, + 700 + ], + "type": "inline_equation", + "content": "5^{\\circ}" + }, + { + "bbox": [ + 46, + 616, + 289, + 700 + ], + "type": "text", + "content": ", and AA@" + }, + { + "bbox": [ + 46, + 616, + 289, + 700 + ], + "type": "inline_equation", + "content": "10^{\\circ}" + }, + { + "bbox": [ + 46, + 616, + 289, + 700 + ], + "type": "text", + "content": ". The results are shown in Table 3. Since the standard deviations (STDEV) are consistently below " + }, + { + "bbox": [ + 46, + 616, + 289, + 700 + ], + "type": "inline_equation", + "content": "3\\%" + }, + { + "bbox": [ + 46, + 616, + 289, + 700 + ], + "type": "text", + "content": " of their respective means, it is evident that DiskVPS is reasonably robust against changing parameters." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 59, + 701, + 288, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 59, + 701, + 288, + 714 + ], + "spans": [ + { + "bbox": [ + 59, + 701, + 288, + 714 + ], + "type": "text", + "content": "Lastly, we also compare our approach's efficiency with" + } + ] + } + ], + "index": 6 + }, + { + "type": "table", + "bbox": [ + 308, + 249, + 545, + 312 + ], + "blocks": [ + { + "bbox": [ + 46, + 198, + 547, + 234 + ], + "lines": [ + { + "bbox": [ + 46, + 198, + 547, + 234 + ], + "spans": [ + { + "bbox": [ + 46, + 198, + 547, + 234 + ], + "type": "text", + "content": "Table 2. Processing speed and angular accuracy (AA) comparison on YUD( [7]), SVVP([27]) and SU3( [40]). Our algorithm achieves by far the best average processing speed on the 3 datasets. Moreover, our approach significantly outperforms all other approaches on real-world YUD and SVVP, and is competitive on synthetic SU3." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 308, + 249, + 545, + 312 + ], + "lines": [ + { + "bbox": [ + 308, + 249, + 545, + 312 + ], + "spans": [ + { + "bbox": [ + 308, + 249, + 545, + 312 + ], + "type": "table", + "html": "
CriterionMinimumMaximumMeanSTDEV
AA@3°70.678.475.32.2
AA@5°89.294.191.51.3
AA@10°99.099.099.00
", + "image_path": "233a49d316afeebfef3a06cbf8515f81b794a09dd211451a36dedf7442fb6703.jpg" + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "table_body" + } + ], + "index": 7 + }, + { + "bbox": [ + 306, + 319, + 545, + 341 + ], + "lines": [ + { + "bbox": [ + 306, + 319, + 545, + 341 + ], + "spans": [ + { + "bbox": [ + 306, + 319, + 545, + 341 + ], + "type": "text", + "content": "Table 3. Angular accuracy statistics over " + }, + { + "bbox": [ + 306, + 319, + 545, + 341 + ], + "type": "inline_equation", + "content": "200\\leq M\\leq 300" + }, + { + "bbox": [ + 306, + 319, + 545, + 341 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 306, + 319, + 545, + 341 + ], + "type": "inline_equation", + "content": "100\\le N\\le 180" + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "text" + }, + { + "bbox": [ + 304, + 364, + 545, + 616 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 364, + 545, + 616 + ], + "spans": [ + { + "bbox": [ + 304, + 364, + 545, + 616 + ], + "type": "text", + "content": "other algorithms. Our approach achieves 158 frames per second (fps) or 6.3 milliseconds per frame excluding image decoding and edge detection. It is around 4 times as fast as the second best at 39 fps achieved by [34]. DiskVPS's high efficiency lies in its voting mechanism. In the voting process, an edge line has 0, 1, or 2 sections crossing a subspace (see, Fig.8) and has around 1 section crossing any subspace on average. The 2 azimuthal angles of the line section are used to derive which series of the consecutively numbered cells of the subspace should receive the votes (see 20). The time complexity of the voting process is " + }, + { + "bbox": [ + 304, + 364, + 545, + 616 + ], + "type": "inline_equation", + "content": "M \\times n" + }, + { + "bbox": [ + 304, + 364, + 545, + 616 + ], + "type": "text", + "content": " where " + }, + { + "bbox": [ + 304, + 364, + 545, + 616 + ], + "type": "inline_equation", + "content": "M" + }, + { + "bbox": [ + 304, + 364, + 545, + 616 + ], + "type": "text", + "content": " is the total number of subspaces and " + }, + { + "bbox": [ + 304, + 364, + 545, + 616 + ], + "type": "inline_equation", + "content": "n" + }, + { + "bbox": [ + 304, + 364, + 545, + 616 + ], + "type": "text", + "content": " is the number of detected edges of a typical image. Assuming " + }, + { + "bbox": [ + 304, + 364, + 545, + 616 + ], + "type": "inline_equation", + "content": "M = 200" + }, + { + "bbox": [ + 304, + 364, + 545, + 616 + ], + "type": "text", + "content": " and a typical image contains around 500 edges, the computational cost of the voting process is roughly " + }, + { + "bbox": [ + 304, + 364, + 545, + 616 + ], + "type": "inline_equation", + "content": "2 \\times 200 \\times 500 = 200,000" + }, + { + "bbox": [ + 304, + 364, + 545, + 616 + ], + "type": "text", + "content": " math function calls, which can be processed in around 2 milliseconds by the 3.5GHz CPU we used. Unlike a typical DL approach that generally employs millions of hidden parameters and deals with time-consuming multilayer processing, DiskVPS has none and thus can reach much higher efficiency." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 306, + 632, + 379, + 643 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 632, + 379, + 643 + ], + "spans": [ + { + "bbox": [ + 306, + 632, + 379, + 643 + ], + "type": "text", + "content": "4. Conclusion" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 304, + 654, + 545, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 654, + 545, + 715 + ], + "spans": [ + { + "bbox": [ + 304, + 654, + 545, + 715 + ], + "type": "text", + "content": "By using a more precise voting mechanism than the Gaussian sphere model currently adopted by most of the SOTA, DiskVPS has achieved significantly better performance than the SOTA in detection accuracy and processing speed for real-world images." + } + ] + } + ], + "index": 11 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "type": "text", + "content": "27056" + } + ] + } + ], + "index": 12 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 7 + }, + { + "para_blocks": [ + { + "bbox": [ + 48, + 71, + 106, + 83 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 71, + 106, + 83 + ], + "spans": [ + { + "bbox": [ + 48, + 71, + 106, + 83 + ], + "type": "text", + "content": "References" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 48, + 91, + 288, + 713 + ], + "type": "list", + "angle": 0, + "index": 14, + "blocks": [ + { + "bbox": [ + 53, + 91, + 287, + 146 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 91, + 287, + 146 + ], + "spans": [ + { + "bbox": [ + 53, + 91, + 287, + 146 + ], + "type": "text", + "content": "[1] Michel Antunes and Joao P Barreto. A global approach for the detection of vanishing points and mutually orthogonal vanishing directions. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition, pages 1336-1343, 2013. 1" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 53, + 148, + 287, + 170 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 148, + 287, + 170 + ], + "spans": [ + { + "bbox": [ + 53, + 148, + 287, + 170 + ], + "type": "text", + "content": "[2] Stephen T Barnard. Interpreting perspective images. Artificial intelligence, 21(4):435-462, 1983. 1" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 53, + 171, + 287, + 204 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 171, + 287, + 204 + ], + "spans": [ + { + "bbox": [ + 53, + 171, + 287, + 204 + ], + "type": "text", + "content": "[3] Bruno Caprile and Vincent Torre. Using vanishing points for camera calibration. International journal of computer vision, 4(2):127-139, 1990. 1, 7" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 53, + 206, + 288, + 260 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 206, + 288, + 260 + ], + "spans": [ + { + "bbox": [ + 53, + 206, + 288, + 260 + ], + "type": "text", + "content": "[4] Taiyan Chen, Xianghua Ying, Jinfa Yang, Ruibin Wang, Ruohao Guo, Bowei Xing, and Ji Shi. Vpdetr: end-to-end vanishing point detection transformers. In Proceedings of the AAAI Conference on Artificial Intelligence, volume 38, pages 1192-1200, 2024. 1, 2" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 53, + 262, + 287, + 316 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 262, + 287, + 316 + ], + "spans": [ + { + "bbox": [ + 53, + 262, + 287, + 316 + ], + "type": "text", + "content": "[5] James M Coughlan and Alan L Yuille. Manhattan world: Compass direction from a single image by bayesian inference. In Proceedings of the seventh IEEE international conference on computer vision, volume 2, pages 941-947. IEEE, 1999. 1, 7" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 53, + 318, + 287, + 352 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 318, + 287, + 352 + ], + "spans": [ + { + "bbox": [ + 53, + 318, + 287, + 352 + ], + "type": "text", + "content": "[6] James M Coughlan and Alan L Yuille. The Manhattan world assumption: Regularities in scene statistics which enable bayesian inference. In NIPS, volume 2, page 3, 2000. 1" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 53, + 353, + 287, + 397 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 353, + 287, + 397 + ], + "spans": [ + { + "bbox": [ + 53, + 353, + 287, + 397 + ], + "type": "text", + "content": "[7] Patrick Denis, James H Elder, and Francisco J Estrada. Efficient edge-based methods for estimating Manhattan frames in urban imagery. In European conference on computer vision, pages 197-210. Springer, 2008. 6, 7, 8" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 53, + 399, + 287, + 442 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 399, + 287, + 442 + ], + "spans": [ + { + "bbox": [ + 53, + 399, + 287, + 442 + ], + "type": "text", + "content": "[8] Martin A Fischler and Robert C Bolles. Random sample consensus: a paradigm for model fitting with applications to image analysis and automated cartography. Communications of the ACM, 24(6):381-395, 1981. 1" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 53, + 444, + 287, + 487 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 444, + 287, + 487 + ], + "spans": [ + { + "bbox": [ + 53, + 444, + 287, + 487 + ], + "type": "text", + "content": "[9] Paul VC Hough. Machine analysis of bubble chamber pictures. In Proc. of the International Conference on High Energy Accelerators and Instrumentation, Sept. 1959, pages 554-556, 1959. 1" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 48, + 489, + 287, + 544 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 489, + 287, + 544 + ], + "spans": [ + { + "bbox": [ + 48, + 489, + 287, + 544 + ], + "type": "text", + "content": "[10] Florian Kluger, Hanno Ackermann, Michael Ying Yang, and Bodo Rosenhahn. Deep learning for vanishing point detection using an inverse gnomonic projection. In German Conference on Pattern Recognition, pages 17-28. Springer, 2017. 1, 2" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 48, + 545, + 287, + 610 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 545, + 287, + 610 + ], + "spans": [ + { + "bbox": [ + 48, + 545, + 287, + 610 + ], + "type": "text", + "content": "[11] Florian Kluger, Eric Brachmann, Hanno Ackermann, Carsten Rother, Michael Ying Yang, and Bodo Rosenhahn. Consac: Robust multi-model fitting by conditional sample consensus. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 4634-4643, 2020. 1, 2, 7, 8" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 48, + 613, + 287, + 656 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 613, + 287, + 656 + ], + "spans": [ + { + "bbox": [ + 48, + 613, + 287, + 656 + ], + "type": "text", + "content": "[12] Till Kroeger, Dengxin Dai, and Luc Van Gool. Joint vanishing point extraction and tracking. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition, pages 2449-2457, 2015. 1" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 48, + 658, + 287, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 658, + 287, + 713 + ], + "spans": [ + { + "bbox": [ + 48, + 658, + 287, + 713 + ], + "type": "text", + "content": "[13] Seokju Lee, Junsik Kim, Jae Shin Yoon, Seunghak Shin, Oleksandr Bailo, Namil Kim, Tae-Hee Lee, Hyun Seok Hong, Seung-Hoon Han, and In So Kweon. Vpgnet: Vanishing point guided network for lane and road marking detection and recognition. In Proceedings of the IEEE inter" + } + ] + } + ], + "index": 13 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 307, + 73, + 547, + 713 + ], + "type": "list", + "angle": 0, + "index": 30, + "blocks": [ + { + "bbox": [ + 326, + 73, + 545, + 94 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 326, + 73, + 545, + 94 + ], + "spans": [ + { + "bbox": [ + 326, + 73, + 545, + 94 + ], + "type": "text", + "content": "national conference on computer vision, pages 1947-1955, 2017. 1" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 307, + 95, + 547, + 149 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 95, + 547, + 149 + ], + "spans": [ + { + "bbox": [ + 307, + 95, + 547, + 149 + ], + "type": "text", + "content": "[14] José Lezama, Rafael Grompone von Gioi, Gregory Randall, and Jean-Michel Morel. Finding vanishing points via point alignments in image primal and dual domains. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition, pages 509-515, 2014. 2" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 307, + 151, + 545, + 194 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 151, + 545, + 194 + ], + "spans": [ + { + "bbox": [ + 307, + 151, + 545, + 194 + ], + "type": "text", + "content": "[15] David Liebowitz and Andrew Zisserman. Combining scene and auto-calibration constraints. In Proceedings of the Seventh IEEE International Conference on Computer Vision, volume 1, pages 293-300. IEEE, 1999. 1" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 307, + 195, + 547, + 259 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 195, + 547, + 259 + ], + "spans": [ + { + "bbox": [ + 307, + 195, + 547, + 259 + ], + "type": "text", + "content": "[16] Yancong Lin, Ruben Wiersma, Silvia L. Pintea, Klaus Hildebrandt, Elmar Eisemann, and Jan C. van Gemert. Deep vanishing point detection: Geometric priors make dataset variations vanish. In 2022 IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 6093-6103, 2022. 1, 2, 7, 8" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 307, + 261, + 545, + 316 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 261, + 545, + 316 + ], + "spans": [ + { + "bbox": [ + 307, + 261, + 545, + 316 + ], + "type": "text", + "content": "[17] Xiaohu Lu, Jian Yaoy, Haoang Li, Yahui Liu, and Xiaofeng Zhang. 2-line exhaustive searching for real-time vanishing point estimation in manhattan world. In 2017 IEEE Winter Conference on Applications of Computer Vision (WACV), pages 345-353. IEEE, 2017. 1" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 307, + 316, + 545, + 360 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 316, + 545, + 360 + ], + "spans": [ + { + "bbox": [ + 307, + 316, + 545, + 360 + ], + "type": "text", + "content": "[18] Evelyne Lutton, Henri Maitre, and Jaime Lopez-Krahe. Contribution to the determination of vanishing points using hough transform. IEEE transactions on pattern analysis and machine intelligence, 16(4):430-438, 1994. 1" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 307, + 360, + 545, + 403 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 360, + 545, + 403 + ], + "spans": [ + { + "bbox": [ + 307, + 360, + 545, + 403 + ], + "type": "text", + "content": "[19] Faraz M Mirzaei and Stergios I Roumeliotis. Optimal estimation of vanishing points in a Manhattan world. In 2011 International Conference on Computer Vision, pages 2454-2461. IEEE, 2011. 1, 2" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 307, + 404, + 545, + 437 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 404, + 545, + 437 + ], + "spans": [ + { + "bbox": [ + 307, + 404, + 545, + 437 + ], + "type": "text", + "content": "[20] Van Nhan Nguyen, Robert Jenssen, and Davide Roverso. Lsnet: Fast single-shot line-segment detector. Machine Vision and Applications, 32(1):12, 2021. 7" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 307, + 437, + 545, + 491 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 437, + 545, + 491 + ], + "spans": [ + { + "bbox": [ + 307, + 437, + 545, + 491 + ], + "type": "text", + "content": "[21] Rémi Pautrat, Shaohui Liu, Petr Hruby, Marc Pollefeys, and Daniel Barath. Vanishing point estimation in uncalibrated images with prior gravity direction. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 14118-14127, 2023. 1" + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 307, + 492, + 545, + 525 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 492, + 545, + 525 + ], + "spans": [ + { + "bbox": [ + 307, + 492, + 545, + 525 + ], + "type": "text", + "content": "[22] Carsten Rother. A new approach to vanishing point detection in architectural environments. Image and Vision Computing, 20(9-10):647-655, 2002. 1" + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 307, + 525, + 545, + 581 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 525, + 545, + 581 + ], + "spans": [ + { + "bbox": [ + 307, + 525, + 545, + 581 + ], + "type": "text", + "content": "[23] Alexander Sheshkus, Anastasia Ingacheva, Vladimir Arlazarov, and Dmitry Nikolaev. Houghnet: neural network architecture for vanishing points detection. In 2019 International Conference on Document Analysis and Recognition (ICDAR), pages 844-849. IEEE, 2019. 1, 2" + } + ] + } + ], + "index": 25 + }, + { + "bbox": [ + 307, + 582, + 545, + 624 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 582, + 545, + 624 + ], + "spans": [ + { + "bbox": [ + 307, + 582, + 545, + 624 + ], + "type": "text", + "content": "[24] Jefferey A Shufelt. Performance evaluation and analysis of vanishing point detection techniques. IEEE transactions on pattern analysis and machine intelligence, 21(3):282-288, 1999. 1" + } + ] + } + ], + "index": 26 + }, + { + "bbox": [ + 307, + 624, + 545, + 647 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 624, + 545, + 647 + ], + "spans": [ + { + "bbox": [ + 307, + 624, + 545, + 647 + ], + "type": "text", + "content": "[25] Sudipta Sinha, Drew Steedly, and Rick Szeliski. Piecewise planar stereo for image-based rendering. 2009. 1" + } + ] + } + ], + "index": 27 + }, + { + "bbox": [ + 307, + 647, + 545, + 689 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 647, + 545, + 689 + ], + "spans": [ + { + "bbox": [ + 307, + 647, + 545, + 689 + ], + "type": "text", + "content": "[26] Jean-Philippe Tardif. Non-iterative approach for fast and accurate vanishing point detection. In 2009 IEEE 12th International Conference on Computer Vision, pages 1250–1257. IEEE, 2009. 1, 7, 8" + } + ] + } + ], + "index": 28 + }, + { + "bbox": [ + 307, + 691, + 545, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 691, + 545, + 713 + ], + "spans": [ + { + "bbox": [ + 307, + 691, + 545, + 713 + ], + "type": "text", + "content": "[27] Xin Tong, Xianghua Ying, Yongjie Shi, Ruibin Wang, and Jinfa Yang. Transformer based line segment classifier with" + } + ] + } + ], + "index": 29 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "text", + "content": "27057" + } + ] + } + ], + "index": 31 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 8 + }, + { + "para_blocks": [ + { + "bbox": [ + 48, + 73, + 287, + 713 + ], + "type": "list", + "angle": 0, + "index": 14, + "blocks": [ + { + "bbox": [ + 66, + 73, + 287, + 116 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 66, + 73, + 287, + 116 + ], + "spans": [ + { + "bbox": [ + 66, + 73, + 287, + 116 + ], + "type": "text", + "content": "image context for real-time vanishing point detection in man-. hhattan world. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 6093- 6102, 2022. 1, 2, 6, 7, 8" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 48, + 118, + 287, + 162 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 118, + 287, + 162 + ], + "spans": [ + { + "bbox": [ + 48, + 118, + 287, + 162 + ], + "type": "text", + "content": "[28] Elena Tretyak, Olga Barinova, Pushmeet Kohli, and Victor Lempitsky. Geometric image parsing in man-made environments. International Journal of Computer Vision, 97(3):305-321, 2012. 1" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 48, + 163, + 287, + 217 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 163, + 287, + 217 + ], + "spans": [ + { + "bbox": [ + 48, + 163, + 287, + 217 + ], + "type": "text", + "content": "[29] Tinne Tuytelaars, Luc Van Gool, Marc Proesmans, and Theo Moons. The cascaded hough transform as an aid in aerial image interpretation. In Sixth International Conference on Computer Vision (IEEE Cat. No. 98CH36271), pages 67-72. IEEE, 1998. 1" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 48, + 219, + 287, + 252 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 219, + 287, + 252 + ], + "spans": [ + { + "bbox": [ + 48, + 219, + 287, + 252 + ], + "type": "text", + "content": "[30] Jonathan Ventura and Tobias Höllerer. Structure and motion in urban environments using upright panoramas. *Virtual Reality*, 17(2):147-156, 2013. 1" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 48, + 254, + 287, + 308 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 254, + 287, + 308 + ], + "spans": [ + { + "bbox": [ + 48, + 254, + 287, + 308 + ], + "type": "text", + "content": "[31] Rafael Grompone Von Gioi, Jeremie Jakubowicz, Jean-Michel Morel, and Gregory Randall. Lsd: A fast line segment detector with a false detection control. IEEE transactions on pattern analysis and machine intelligence, 32(4):722-732, 2008. 7" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 48, + 309, + 287, + 342 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 309, + 287, + 342 + ], + "spans": [ + { + "bbox": [ + 48, + 309, + 287, + 342 + ], + "type": "text", + "content": "[32] Yukun Wang, Shuo Gu, Yinbo Liu, and Hui Kong. Semi-supervised vanishing point detection with contrastive learning. Pattern Recognition, 153:110518, 2024. 1, 2" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 48, + 344, + 287, + 387 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 344, + 287, + 387 + ], + "spans": [ + { + "bbox": [ + 48, + 344, + 287, + 387 + ], + "type": "text", + "content": "[33] Horst Wildenauer and Allan Hanbury. Robust camera self-calibration from monocular images of manhattan worlds. In 2012 IEEE Conference on Computer Vision and Pattern Recognition, pages 2831-2838. IEEE, 2012. 1" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 48, + 388, + 287, + 443 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 388, + 287, + 443 + ], + "spans": [ + { + "bbox": [ + 48, + 388, + 287, + 443 + ], + "type": "text", + "content": "[34] Jianping Wu, Liang Zhang, Ye Liu, and Ke Chen. Real-time vanishing point detector integrating under-parameterized ransac and hough transform. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 3732-3741, 2021. 2, 3, 7, 8" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 48, + 445, + 287, + 499 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 445, + 287, + 499 + ], + "spans": [ + { + "bbox": [ + 48, + 445, + 287, + 499 + ], + "type": "text", + "content": "[35] Yiliang Xu, Sangmin Oh, and Anthony Hoogs. A minimum error vanishing point detection approach for uncalibrated monocular images of man-made environments. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition, pages 1376-1383, 2013. 1" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 48, + 501, + 287, + 544 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 501, + 287, + 544 + ], + "spans": [ + { + "bbox": [ + 48, + 501, + 287, + 544 + ], + "type": "text", + "content": "[36] Yifan Xu, Weijian Xu, David Cheung, and Zhuowen Tu. Line segment detection using transformers without edges. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 4257-4266, 2021. 7" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 48, + 545, + 287, + 599 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 545, + 287, + 599 + ], + "spans": [ + { + "bbox": [ + 48, + 545, + 287, + 599 + ], + "type": "text", + "content": "[37] Menghua Zhai, Scott Workman, and Nathan Jacobs. Detecting vanishing points using global image context in a non-manhattan world. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition, pages 5657-5665, 2016. 2" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 48, + 601, + 287, + 645 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 601, + 287, + 645 + ], + "spans": [ + { + "bbox": [ + 48, + 601, + 287, + 645 + ], + "type": "text", + "content": "[38] Lilian Zhang, Huimin Lu, Xiaoping Hu, and Reinhard Koch. Vanishing point estimation and line classification in a Manhattan world with a unifying camera model. International Journal of Computer Vision, 117:111-130, 2016. 1" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 48, + 647, + 287, + 679 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 647, + 287, + 679 + ], + "spans": [ + { + "bbox": [ + 48, + 647, + 287, + 679 + ], + "type": "text", + "content": "[39] Yichao Zhou, Haozhi Qi, Jingwei Huang, and Yi Ma. Neurvps: Neural vanishing point scanning via conic convolution. arXiv preprint arXiv:1910.06316, 2019. 1, 2, 7, 8" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 48, + 681, + 287, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 681, + 287, + 713 + ], + "spans": [ + { + "bbox": [ + 48, + 681, + 287, + 713 + ], + "type": "text", + "content": "[40] Yichao Zhou, Haozhi Qi, Yuexiang Zhai, Qi Sun, Zhili Chen, Li-Yi Wei, and Yi Ma. Learning to reconstruct 3d Manhattan wireframes from a single image. In Proceedings of the" + } + ] + } + ], + "index": 13 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 326, + 73, + 545, + 95 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 326, + 73, + 545, + 95 + ], + "spans": [ + { + "bbox": [ + 326, + 73, + 545, + 95 + ], + "type": "text", + "content": "IEEE/CVF International Conference on Computer Vision, pages 7698-7707, 2019. 1, 7, 8" + } + ] + } + ], + "index": 15 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 749, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 749, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 749, + 317, + 757 + ], + "type": "text", + "content": "27058" + } + ] + } + ], + "index": 16 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 9 + } + ], + "_backend": "vlm", + "_version_name": "2.6.4" +} \ No newline at end of file diff --git a/2025/Dispider_ Enabling Video LLMs with Active Real-Time Interaction via Disentangled Perception, Decision, and Reaction/509357f7-50b0-4a3e-be02-5d53b2c39c5b_content_list.json b/2025/Dispider_ Enabling Video LLMs with Active Real-Time Interaction via Disentangled Perception, Decision, and Reaction/509357f7-50b0-4a3e-be02-5d53b2c39c5b_content_list.json new file mode 100644 index 0000000000000000000000000000000000000000..a250fce1b7748b7e7320938540366158d284368b --- /dev/null +++ b/2025/Dispider_ Enabling Video LLMs with Active Real-Time Interaction via Disentangled Perception, Decision, and Reaction/509357f7-50b0-4a3e-be02-5d53b2c39c5b_content_list.json @@ -0,0 +1,1706 @@ +[ + { + "type": "text", + "text": "Dispider: Enabling Video LLMs with Active Real-Time Interaction via Disentangled Perception, Decision, and Reaction", + "text_level": 1, + "bbox": [ + 142, + 128, + 854, + 176 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Rui Qian $^{1*}$ Shuangrui Ding $^{1*}$ Xiaoyi Dong $^{1,2\\dagger}$ Pan Zhang $^{2}$ Yuhang Zang $^{2}$ Yuhang Cao $^{2}$ Dahua Lin $^{1,2,4}$ Jiaqi Wang $^{2,3\\dagger}$", + "bbox": [ + 240, + 202, + 759, + 239 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "1 The Chinese University of Hong Kong 2 Shanghai AI Laboratory \n3 Shanghai Innovation Institute 4 CPII under InnoHK", + "bbox": [ + 236, + 239, + 759, + 273 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "{qr021, ds023}@ie.cuhk.edu.hk", + "bbox": [ + 367, + 277, + 624, + 291 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Abstract", + "text_level": 1, + "bbox": [ + 246, + 325, + 326, + 343 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Active real-time interaction with video LLMs introduces a new paradigm for human-computer interaction, where the model not only understands user intent but also responds while continuously processing streaming video on the fly. Unlike offline video LLMs, which analyze the entire video before answering questions, active real-time interaction requires three capabilities: 1) Perception: real-time video monitoring and interaction capturing. 2) Decision: raising proactive interaction in proper situations, 3) Reaction: continuous interaction with users. However, inherent conflicts exist among the desired capabilities. The Decision and Reaction require a contrary Perception scale and grain, and the autoregressive decoding blocks the real-time Perception and Decision during the Reaction. To unify the conflicted capabilities within a harmonious system, we present Dispider, a solution built on a Disentangled Perception, Decision, and Reaction framework. Dispider features a lightweight Proactive Streaming Video Processing module that tracks the video stream and identifies optimal moments for interaction. Once the interaction is triggered, an asynchronous Precise Interaction module provides detailed responses, while the processing module continues to monitor the video in the meantime. Our disentangled and asynchronous design ensures timely, contextually accurate, and computationally efficient responses, making Dispider ideal for active real-time interaction for long-duration video streams. Experiments prove that Dispider outperforms existing methods not only in its superior understanding of video content in conventional video QA settings, but also in proactive response capability and temporal awareness under the streaming setting. The code and model are released at https://github.com/Mark12Ding/Dispider.", + "bbox": [ + 89, + 361, + 485, + 845 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "1. Introduction", + "text_level": 1, + "bbox": [ + 514, + 325, + 643, + 342 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "The recent surge in multimodal large language models (MLLMs) has brought considerable attention to video understanding tasks [14, 15, 17, 47]. Given a video, these models aim to accurately comprehend its content and provide responses aligned with human expectations. However, the majority of current video LLMs are designed around an offline setting, where models are required to view the entire video before generating a single response. While effective for certain applications, this offline approach is impractical for real-time $^{1}$ , interactive scenarios where users expect continuous feedback as video streams.", + "bbox": [ + 511, + 352, + 906, + 517 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Consider a scenario where a user performs pull-ups with the assistance of a visual model. As illustrated in Figure 1, current offline video LLMs need to process the entire video before providing the total count of pull-ups. However, users expect real-time feedback, with each pull-up counted as it occurs: \"One... Two... Three...\" This real-time interaction requires the model to recognize each completed action instantly and synchronize its responses with the user's movements. The challenge lies in determining the precise moment to output the count along with the ongoing video stream. To this end, traditional offline video LLMs lack the flexibility to handle such demands, as they are not designed for continuous, dynamic video streams.", + "bbox": [ + 511, + 517, + 908, + 714 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "VideoLLM-online [5] has been a pioneering effort in addressing this issue. By leveraging the LLaMA model [32], they enable proactive response updates during streaming video inputs. However, it has a critical limitation: since it uses a single LLM for video processing and response generation, it cannot perform perception and answer reaction simultaneously. The autoregressive nature of next-token prediction in LLMs forces VideoLLM-online to alternate between perception (processing frames) and response generation, rather than handling both in parallel. This restriction", + "bbox": [ + 511, + 714, + 908, + 867 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "CVF", + "bbox": [ + 106, + 2, + 181, + 42 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore.", + "bbox": [ + 236, + 0, + 810, + 46 + ], + "page_idx": 0 + }, + { + "type": "page_footnote", + "text": "1We use \"real-time\" loosely to emphasize online interaction, not strict frame-rate requirements.", + "bbox": [ + 511, + 875, + 908, + 901 + ], + "page_idx": 0 + }, + { + "type": "page_footnote", + "text": "*Equal Contribution.", + "bbox": [ + 107, + 875, + 222, + 887 + ], + "page_idx": 0 + }, + { + "type": "page_footnote", + "text": "† Corresponding Author.", + "bbox": [ + 109, + 887, + 238, + 898 + ], + "page_idx": 0 + }, + { + "type": "page_number", + "text": "24045", + "bbox": [ + 478, + 944, + 517, + 955 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/255603027e52059f38adf1a94e4fabdbdb4af1fd96859265d3a78ff0e17ecf36.jpg", + "image_caption": [ + "Figure 1. Comparison between offline and active real-time Video LLMs in processing a video sequence of a man performing pull-ups. The user asks the model to count the number of pull-ups. The offline Video LLMs process the entire video before providing a final count of three pull-ups. In contrast, the real-time Video LLMs provide immediate feedback after each pull-up, first counting one, then two, and finally reaching three, while remaining silent between responses." + ], + "image_footnote": [], + "bbox": [ + 94, + 137, + 841, + 242 + ], + "page_idx": 1 + }, + { + "type": "image", + "img_path": "images/7a6a8509236a8b0757bb9626b112faf76911731080344add1aeb9eed2bbfe84c.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 106, + 260, + 127, + 276 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Real-time Video LLM", + "bbox": [ + 130, + 263, + 259, + 276 + ], + "page_idx": 1 + }, + { + "type": "image", + "img_path": "images/c18cb3313382d5bb3999b464a74a28f65af0717b9ec441139da3a37aa9bba761.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 295, + 262, + 321, + 281 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": ": The man did 1 pull-up.", + "bbox": [ + 323, + 266, + 439, + 277 + ], + "page_idx": 1 + }, + { + "type": "image", + "img_path": "images/8cc534a32a9d74bf06415d6300be7c567de6a3244ffe56c129325efbdf41b17f.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 482, + 263, + 504, + 281 + ], + "page_idx": 1 + }, + { + "type": "image", + "img_path": "images/826a2d6569ce3904afa977fb792d52416a35c55c4a58f80e5c0f0d248a8fc6a6.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 542, + 266, + 568, + 279 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": ": The man did 2 pull-ups.", + "bbox": [ + 568, + 267, + 689, + 279 + ], + "page_idx": 1 + }, + { + "type": "image", + "img_path": "images/61f5e1be321d1ef4ad7ce90747054fdf5249166d1d91f08af0a33d88b13e4f40.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 702, + 261, + 730, + 281 + ], + "page_idx": 1 + }, + { + "type": "image", + "img_path": "images/56b5b4d7e1e9970610d47a564691468b520acff2e16c08f0452729651286867e.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 740, + 266, + 766, + 279 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "The man did 3 pull-ups.", + "bbox": [ + 769, + 266, + 885, + 277 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "delays the video inputs during responding, reduces responsiveness and hinders its ability to engage in real-time interactions effectively, especially during long-duration video streams that require continuous, uninterrupted processing.", + "bbox": [ + 89, + 390, + 482, + 450 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "To address these limitations, we present Dispider, a novel framework designed for active real-time interaction with streaming videos. Dispider resolves the conflicts between perception, decision, and reaction through a disentangled, asynchronous design, where each component can operate in parallel. Specifically, we implement these functions in the Proactive Streaming Video Processing module and the Precise Interaction module. In the video processing module, with the streaming video input, we dynamically segment the video stream into non-uniform clips based on scene boundaries. These clips, combined with relevant historical context, are then processed by a real-time response decision module, which determines whether the model should generate a response or wait for additional video content. When an interaction is triggered, the Precise Interaction module generates context-aware responses without interrupting the ongoing video processing.", + "bbox": [ + 89, + 455, + 483, + 714 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "This disentangled design ensures that the Dispider can provide timely, accurate, and computationally efficient responses, even for long-duration video streams. We evaluate Dispider's ability in both real-time video stream interactions, and conventional offline Video LLMs across long-video benchmarks (EgoSchema, VideoMME, MLVU) and time-sensitive tasks (ETBench). Dispider demonstrated strong performance, particularly excelling in temporal reasoning on EgoSchema and handling diverse video lengths effectively. In streaming tasks, it significantly surpasses VideoLLM-online [5] in temporal grounding, proactive response generation, and multi-step reasoning.", + "bbox": [ + 89, + 719, + 483, + 901 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "2. Related Work", + "text_level": 1, + "bbox": [ + 513, + 388, + 653, + 402 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Large Language Models. LLMs have revolutionized natural language processing, achieving remarkable performance across a wide range of tasks. Early models such as BERT [12] and T5 [52] used masked language modeling for pre-training. The shift to decoder-only models like GPT [50] introduced scalable architectures that enhanced language generation. Recent models, including PaLM [10], LLaMA [56], and GPT-4 [45], continue to push the boundaries with massive parameters and extensive training data. Techniques such as supervised fine-tuning and reinforcement learning from human feedback [2-4, 16, 22, 23, 27, 32, 46, 62] have further improved these models' ability to generate coherent, contextually relevant responses. Inspired by the reasoning capabilities of LLMs, we extend these models to the domain of streaming video understanding, where real-time interaction presents unique challenges.", + "bbox": [ + 511, + 415, + 906, + 656 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Video Large Language Models. The progress of multimodal LLMs has been significant, particularly with image-based models like InstructBLIP [11], LLaVA [37-39], and Flamingo [1], which integrate vision and language models. Extending this to video introduces challenges in managing both frame sequences and the context length of LLMs. Recent works on video LLMs [6-9, 18, 21, 33, 36, 42, 43, 55, 57-59, 61, 64, 66, 68] present novel strategies for processing long-form videos while maintaining effective reasoning. Notable approaches include TimeChat [53], which emphasizes temporal relationships across video frames, and MovieChat [55], which uses sparse memory to handle long videos. VideoChat [30] adopts a chat-centric approach, integrating video foundation models with LLMs to excel at spatiotemporal reasoning, event localization, and causal inference. Building on these video LLMs, our work intro", + "bbox": [ + 511, + 657, + 908, + 900 + ], + "page_idx": 1 + }, + { + "type": "header", + "text": "#", + "bbox": [ + 116, + 101, + 138, + 119 + ], + "page_idx": 1 + }, + { + "type": "header", + "text": ": How many pull-ups has the man done?", + "bbox": [ + 142, + 104, + 338, + 118 + ], + "page_idx": 1 + }, + { + "type": "header", + "text": "Offline Video LLM", + "bbox": [ + 570, + 102, + 712, + 118 + ], + "page_idx": 1 + }, + { + "type": "header", + "text": "#", + "bbox": [ + 717, + 101, + 743, + 119 + ], + "page_idx": 1 + }, + { + "type": "header", + "text": ": The man has done 3 pull-ups.", + "bbox": [ + 743, + 104, + 893, + 117 + ], + "page_idx": 1 + }, + { + "type": "page_number", + "text": "24046", + "bbox": [ + 478, + 945, + 517, + 955 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "duces a pipeline designed to process streaming video inputs and generate real-time outputs.", + "bbox": [ + 89, + 90, + 480, + 119 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Streaming Video Understanding. In practice, users typically expect models to operate online and interactively in real time. This is known as streaming video understanding, which involves processing continuous video streams while ensuring long-term temporal consistency and enabling interactive responses. Despite its practical significance, only a few works have explored this area [34, 67]. VideoLLM-online [5] introduces the LIVE framework for streaming dialogue, but it lacks an efficient module for handling long-term video inputs over extended periods and does not prioritize real-time interactivity. VideoStream [48] proposes memory-propagated encoding for long videos, yet its focus is on offline processing rather than real-time streaming. Similarly, Flash-VStream [65] addresses inference latency and memory efficiency for long video streams but overlooks real-time user interaction. In contrast, our approach centers on developing a real-time visual assistant for streaming video, emphasizing both long-context handling and interactive, real-time responses, thereby bridging the gap left by previous methods.", + "bbox": [ + 91, + 122, + 483, + 422 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3. Method", + "text_level": 1, + "bbox": [ + 89, + 436, + 181, + 452 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3.1. Problem Formulation", + "text_level": 1, + "bbox": [ + 89, + 462, + 294, + 477 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "In contrast to previous offline video LLMs, which generate responses only after processing the entire video, our approach requires the model to operate in real-time, providing continuous and interactive responses as the video is processed. Specifically, the model actively determines when sufficient information has been observed to provide a complete response, allowing it to produce answers promptly without waiting for the entire video to finish. Following the formulation of VideoLLM-online [5], we formulate this new setting as below.", + "bbox": [ + 89, + 483, + 482, + 633 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Given a continuous video stream $V = \\{v_{1}, v_{2}, \\dots, v_{T}\\}$ , where $v_{i}$ represents the $i$ -th video clip, and a context sequence $C_{t}$ up to a specific time $t$ (which includes prior vision-language interactions such as historical user queries, assistant responses, and video frames), our goal is to generate timely and accurate dialogue responses $R_{t}$ without processing the entire video sequence.", + "bbox": [ + 89, + 635, + 482, + 739 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "At each time $t$ , the model observes $C_t$ and the video frames up to that time $V_{[0,t]} = \\{v_1, v_2, \\ldots, v_t\\}$ . The model must decide whether it has sufficient information to generate a response. We define a decision function $\\pi$ and a response function $f$ :", + "bbox": [ + 89, + 741, + 482, + 816 + ], + "page_idx": 2 + }, + { + "type": "equation", + "text": "\n$$\na _ {t} = \\pi \\left(C _ {t}, V _ {[ 0, t ]}\\right) \\in \\{\\text {w a i t , r e s p o n d} \\}.\n$$\n", + "text_format": "latex", + "bbox": [ + 161, + 830, + 411, + 848 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "If $a_{t} =$ respond, the model generates a response:", + "bbox": [ + 109, + 854, + 434, + 869 + ], + "page_idx": 2 + }, + { + "type": "equation", + "text": "\n$$\nR _ {t} = f \\left(C _ {t}, V _ {[ 0, t ]}\\right)\n$$\n", + "text_format": "latex", + "bbox": [ + 223, + 883, + 349, + 902 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Otherwise, the model keeps silent for more information.", + "bbox": [ + 511, + 90, + 883, + 106 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "In this work, we implement these functions using a disentangled framework composed of three distinct modules: Perception, Decision, and Reaction. These modules are designed to handle the unique challenges of real-time video understanding and dialogue generation. The Perception module focuses on continuous video monitoring, while the Decision module assesses when to act, and the Reaction module generates responses without waiting for the entire video sequence to finish. This approach allows the system to remain responsive to new information while ensuring that the generated answers are based on the most relevant and up-to-date context.", + "bbox": [ + 511, + 107, + 903, + 287 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3.2. Proactive Streaming Video Processing", + "text_level": 1, + "bbox": [ + 511, + 297, + 841, + 314 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "To enable active real-time response decisions, we propose a proactive streaming video processing approach that dynamically segments the video stream and evaluates whether to generate a response as the video is being processed. This approach balances computational efficiency with the need for timely and accurate responses, ensuring that the model processes the most relevant parts of the video while avoiding unnecessary redundancy. The system is composed of two key modules: the Scene-based Perception Module and the Real-time Response Decision Module.", + "bbox": [ + 511, + 320, + 903, + 470 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Scene-based Perception Module. To ensure efficient processing of long video streams, we begin by adaptively segmenting the video into non-uniform clips based on scene boundaries. This segmentation strategy preserves the structural information of the video, allowing the model to focus on the most informative parts while removing redundancy and maintaining context.", + "bbox": [ + 511, + 474, + 903, + 580 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "We begin by sampling frames at a regular interval (one frame per second) and extract L2-normalized feature embeddings using the pre-trained SigLip model [63]. By computing the cosine similarity between these embeddings, we can identify significant visual changes, which indicate potential scene boundaries. These boundaries help divide the video into meaningful segments. To prevent excessively short clips, we introduce an exclusion window around the boundaries, ensuring that the resulting clips are of sufficient length to contain relevant information.", + "bbox": [ + 511, + 580, + 903, + 731 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Each clip $v_{i}$ is then processed by the video encoder to produce the clip-wise feature representation $F_{i}$ , along with a special clip indicator $\\hat{F}_{i}$ . These clip features are used in the subsequent decision-making process to determine if enough information has been gathered to respond.", + "bbox": [ + 511, + 732, + 903, + 806 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Real-time Response Decision Module. The Real-time Response Decision Module evaluates whether the model should generate a response or continue waiting for more video content. This decision is made based on the video content observed so far and the historical context, including the user's prior questions and assistant responses. We illus", + "bbox": [ + 511, + 810, + 903, + 900 + ], + "page_idx": 2 + }, + { + "type": "page_number", + "text": "24047", + "bbox": [ + 478, + 944, + 519, + 955 + ], + "page_idx": 2 + }, + { + "type": "image", + "img_path": "images/e900973b60bc3bc785f02a22db4251ad3402fc3ca6864d6478513fb87f0cd223.jpg", + "image_caption": [ + "Figure 2. Illustration of the Response Decision Module in a proactive streaming video processing system. The module dynamically determines when to respond during video streaming by segmenting the video into non-uniform clips and utilizing historical memory to capture context. The module then combines memory features, clip features, question text, and special tokens, $\\langle \\mathrm{TODO}\\rangle$ and $\\langle \\mathrm{ANS}\\rangle$ , to decide on response timing." + ], + "image_footnote": [], + "bbox": [ + 94, + 85, + 898, + 440 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "trate this whole process in Figure 2. To effectively combine these multi-modal inputs, we use an interleaved sequence format, which integrates video features, question information, and decision tokens.", + "bbox": [ + 88, + 523, + 480, + 583 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "The process begins with the sequence of video clip representations up to the point of the user's question, followed by the question $Q$ and a special token, $\\langle \\mathrm{TODO} \\rangle$ , which indicates that a response decision is yet to be made. Formally, the input sequence at this stage is:", + "bbox": [ + 89, + 584, + 482, + 659 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\nF [ 0: q _ {\\text {p o s}} ] \\oplus Q \\oplus \\langle \\text {T O D O} \\rangle ,\n$$\n", + "text_format": "latex", + "bbox": [ + 192, + 669, + 380, + 686 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "where $q_{\\mathrm{pos}}$ is the timestamp of the question, and $\\oplus$ denotes concatenation.", + "bbox": [ + 89, + 696, + 482, + 724 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "After this initial stage, the model aggregates the features up to the question timestamp into a global memory $M$ through pooling. Then, we extend the sequence to include this global memory that captures historical context, the question texts, and the clip features from the question timestamp to the current point, ending with $\\langle \\mathrm{TODO}\\rangle$ :", + "bbox": [ + 89, + 727, + 482, + 818 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\nM [ 0: q _ {\\mathrm {p o s}} ] \\oplus Q \\oplus F [ q _ {\\mathrm {p o s}}: ] \\oplus \\langle \\mathrm {T O D O} \\rangle .\n$$\n", + "text_format": "latex", + "bbox": [ + 151, + 828, + 419, + 845 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "When the model decides to respond, the $\\langle \\mathrm{ANS} \\rangle$ token is used to mark the response action. Specifically, with the video continuing to stream, we further integrate global", + "bbox": [ + 89, + 854, + 483, + 901 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "memory from the question timestamp to the position of the first answer, insert the special answer token $\\langle \\mathrm{ANS}\\rangle$ , and include later clips for potential subsequent responses. This sequence concludes with the current clip memory segment and a $\\langle \\mathrm{TODO}\\rangle$ token. Thus, the ultimate input format with multiple answers can be written as:", + "bbox": [ + 511, + 523, + 906, + 613 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\nM [ 0: q _ {\\mathrm {p o s}} ] \\oplus Q \\oplus\n$$\n", + "text_format": "latex", + "bbox": [ + 620, + 623, + 746, + 641 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\nM \\left[ q _ {\\text {p o s}}: a _ {\\text {p o s}} ^ {1} \\right] \\oplus \\langle \\mathrm {A N S} \\rangle \\oplus\n$$\n", + "text_format": "latex", + "bbox": [ + 624, + 642, + 795, + 662 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\nM \\left[ a _ {\\text {p o s}} ^ {1}: a _ {\\text {p o s}} ^ {2} \\right] \\oplus \\langle \\mathrm {A N S} \\rangle \\oplus\n$$\n", + "text_format": "latex", + "bbox": [ + 624, + 662, + 795, + 683 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\cdot \\cdot \\cdot\n$$\n", + "text_format": "latex", + "bbox": [ + 627, + 689, + 648, + 698 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\nF \\left[ a _ {\\text {p o s}} ^ {k}: \\right] \\oplus \\langle \\text {T O D O} \\rangle ,\n$$\n", + "text_format": "latex", + "bbox": [ + 622, + 704, + 764, + 724 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "where $a_{\\mathrm{pos}}^k$ stands for the timestamp for the $k$ -th answer.", + "bbox": [ + 511, + 734, + 883, + 750 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "We feed the interleaved sequence into a compact LLM and adopt a binary classification head on top of the final-layer embedding of the $\\langle \\mathrm{TODO}\\rangle$ token. This head is trained to predict whether the model should respond at each timestep. We use standard binary cross-entropy loss to supervise the model's decision-making process. When the model determines that a response is required, the relevant historical clips are retrieved based on their cosine similarity with the $\\langle \\mathrm{TODO}\\rangle$ token's embedding.", + "bbox": [ + 511, + 750, + 905, + 885 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "To train the model's capacity on temporal retrieval, we", + "bbox": [ + 532, + 885, + 903, + 901 + ], + "page_idx": 3 + }, + { + "type": "page_number", + "text": "24048", + "bbox": [ + 478, + 944, + 517, + 955 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "calculate the cosine similarity between the embedding of the $\\langle \\mathrm{TODO}\\rangle$ token and each historical clip indicator $\\hat{F}_i$ . This similarity yields a probability distribution $\\hat{P}_{\\mathrm{pred}}(i)$ representing the relevance of $i$ -th historical clip to the current response.", + "bbox": [ + 89, + 90, + 482, + 165 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "For the ground truth, we use related clips to create a true temporal distribution. These clips are encoded as a binary mask over the clip intervals relevant to the response, defined as:", + "bbox": [ + 89, + 167, + 482, + 224 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\hat {P} _ {\\text {t r u e}} (i) = \\frac {1}{| R _ {Q} |} \\quad \\text {i f} \\quad i \\in R _ {Q} \\quad \\text {e l s e} \\quad 0,\n$$\n", + "text_format": "latex", + "bbox": [ + 148, + 224, + 423, + 256 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "where $R_{Q}$ is the ground-truth set of relevant clips for particular question $Q$ .", + "bbox": [ + 89, + 261, + 482, + 291 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "We then calculate the KL divergence loss between the true distribution, $\\hat{P}_{\\mathrm{true}}$ , and the predicted distribution, $\\hat{P}_{\\mathrm{pred}}$ , as:", + "bbox": [ + 89, + 292, + 482, + 335 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {L} _ {\\mathrm {K L}} = - \\frac {1}{| R _ {Q} |} \\sum_ {i = 1} ^ {|} R _ {Q} | \\hat {P} _ {\\mathrm {t r u e}} (i) \\log \\hat {P} _ {\\mathrm {p r e d}} (i).\n$$\n", + "text_format": "latex", + "bbox": [ + 140, + 337, + 431, + 377 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "This KL divergence loss encourages the model to align its predictions with the true temporal relevance of the clips, thereby improving its ability to correctly retrieve and ground the relevant video segments for generating accurate, timely responses.", + "bbox": [ + 89, + 383, + 482, + 460 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "3.3. Precise Interaction with Streaming Video", + "text_level": 1, + "bbox": [ + 89, + 469, + 442, + 484 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Once the interaction is triggered, an asynchronous precise interaction module generates a response with fine-grained detail, tailored to the specific context of the video stream. That is, for the $k$ -th answer outputs, the precise interaction module receives the question, the former $k - 1$ history answers as well as the grounded clip features at the current timestamp. This design allows the system to generate appropriate responses by incorporating rich context from both the video content and the ongoing conversation in a streaming setting.", + "bbox": [ + 89, + 492, + 482, + 642 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "A key advantage of this asynchronous approach is that it decouples the response generation from the video stream processing. This separation allows the streaming video to continue being processed in parallel, without waiting for response generation to complete. As a result, the system remains highly efficient and responsive, ensuring that the video content is continuously monitored and processed while still generating contextually accurate responses at the appropriate moments.", + "bbox": [ + 89, + 643, + 482, + 779 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "While the Response Decision Module may determine that a response is needed, there is a chance that no actual output is required. To ensure accuracy in real-time interactions, we enhance the model's responsiveness by training it on a streaming QA dataset that simulates such real-time scenarios. These data include both positive samples (where a response is required) and negative samples (where no response is necessary and the model should keep silent). By", + "bbox": [ + 89, + 780, + 482, + 901 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "dynamically sampling these positive and negative instances, the model learns to either produce the responses that are incremental to the historical answers, summarizing the newly emerging cues in the video stream or output the special $\\langle \\mathrm{SILENT} \\rangle$ token denoting keeping silent.", + "bbox": [ + 511, + 90, + 903, + 167 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "This adaptive reasoning approach ensures that the model remains responsive to new information while avoiding redundancy. By focusing on unanswered content and utilizing past interactions, the model delivers timely and relevant responses, enhancing the overall efficiency and user experience in streaming dialogue generation.", + "bbox": [ + 511, + 167, + 906, + 258 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "4. Experiments", + "text_level": 1, + "bbox": [ + 511, + 273, + 645, + 290 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "4.1. Datasets", + "text_level": 1, + "bbox": [ + 511, + 299, + 614, + 313 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "The training process comprises two parts: response decision with temporal grounding and LLM reasoning. For training the streaming video processor and response decision module, we combine the GroundVQA [13] and ET-Instruct [41] datasets, enriched with temporal annotations to provide supervision for streaming responses and temporal grounding labels. Additionally, to enhance the model's ability to handle questions that require simple reasoning, we curate a set of 50K implicit QA pairs with time labels, guiding the model to locate relevant cues within the video stream.", + "bbox": [ + 511, + 321, + 903, + 470 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "For LLM reasoning, we construct a dataset of 122K streaming video QA pairs based on timestamped QA pairs from ET-Instruct [41] and integrate conventional video QA data from VideoChatGPT [43] and LLaVA-Next-Video [29].", + "bbox": [ + 511, + 473, + 903, + 547 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "4.2. Benchmarks", + "text_level": 1, + "bbox": [ + 511, + 560, + 648, + 574 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "For evaluation, we adopt several established long-video QA benchmarks, including EgoSchema [44], VideoMME [20], and MLVU [71]. EgoSchema consists of over 5K videos, each approximately 3 minutes long, while VideoMME and MLVU include videos of varying lengths, from a few minutes to several hours. We report accuracy on multiple-choice questions for comparison across these benchmarks.", + "bbox": [ + 511, + 582, + 903, + 688 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "In addition, we create a streaming video QA benchmark using a subset of ETBench to measure the model's proactive response capability in real-time video interactions. Specifically, we select six subtasks from ETBench suitable for streaming response evaluation: step localization (SLC), dense video captioning (DVC), temporal action localization (TAL), temporal video grounding (TVG), episodic memory (EPM), and video highlight detection (VHD), encompassing a total of 4,460 videos. For this benchmark, the instruction is provided at the beginning of each video, and the model is required to produce the correct responses at appropriate timestamps as the video plays. We follow the evaluation metrics specified in the original benchmarks to ensure a fair comparison.", + "bbox": [ + 511, + 689, + 906, + 900 + ], + "page_idx": 4 + }, + { + "type": "page_number", + "text": "24049", + "bbox": [ + 478, + 944, + 517, + 955 + ], + "page_idx": 4 + }, + { + "type": "table", + "img_path": "images/bd713f174a7bfb103e19fe4a6f6fdb8b823e56507214b6964d502a9b9baa9e28.jpg", + "table_caption": [], + "table_footnote": [], + "table_body": "
MethodLLM SizeFramesEgoSchemaMLVUVideoMME
Video-LLaVA [35]7B838.447.339.9
Chat-UniVi [28]7B64--40.6
LLaMA-VID [33]7B1 FPS38.533.2-
TimeChat [53]7B9633.030.930.2
MovieChat [55]7B204853.525.838.2
Video-LLaMA2 [9]7B1651.748.547.9
LLaVA-Next-Video [70]7B3243.9-46.6
ShareGPT4Video [6]8B16-46.439.9
VideoChat2 [31]7B1654.447.939.5
LongVA [69]7B128-56.352.6
Kangaroo [40]8B64-61.056.0
Video-CCAM [19]14B96-63.153.2
VideoXL [54]7B128-64.955.5
Ours7B1 FPS55.661.757.2
", + "bbox": [ + 199, + 88, + 795, + 329 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Table 1. Comparison on long video benchmarks. We report the accuracy on the EgoSchema full set, MLVU multiple-choice questions, and the VideoMME overall set without subtitles. For a fair comparison, we also present the model size of the LLM and the number of sampled frames.", + "bbox": [ + 89, + 340, + 906, + 383 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "4.3. Implementation Details", + "text_level": 1, + "bbox": [ + 89, + 397, + 307, + 412 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "The architecture consists of a vision encoder, a compact LLM as the proactive streaming video processor for response decisions, and a larger LLM for the precision interaction module. In practice, the input video frames are resized to $224 \\times 224$ , and a CLIP-L/14 [51] is employed to extract frame-wise features. Following the token compression techniques in [48], we concatenate adjacent tokens and use the compact LLM, instantiated as Qwen2-1.5B [49], to produce time-aware compressed clip-wise features along with clip indicators. Next, we reuse this compact LLM to process the sequence consisting of global memory, question texts, and clip features for response decisions. The final LLM, instantiated as Qwen2-7B [49], receives the grounded clips and global memory to generate responses at the necessary timestamps.", + "bbox": [ + 88, + 421, + 482, + 648 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Training is conducted in two stages. In the first stage, we train the response decision module and the video encoder. In the second stage, we freeze these components and train the final LLM. During training, we insert instructions at arbitrary timestamps within the video to enhance the model's ability to handle varied circumstances. For inference, on conventional video benchmarks, we place the question at the end of the video to ensure fair comparisons with existing works. For streaming evaluation, the question is posed at the beginning to enable proactive responses.", + "bbox": [ + 89, + 648, + 482, + 801 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "4.4. Conventional Video Understanding", + "text_level": 1, + "bbox": [ + 89, + 816, + 397, + 832 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "In this section, we compare our model with existing video LLMs on conventional video QA benchmarks, where the model is required to provide one answer after watching the complete video.", + "bbox": [ + 89, + 839, + 482, + 901 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "We first present a comparative analysis across three challenging long-video benchmarks. As shown in Table 1, we report the accuracy on the EgoSchema full set, MLVU multiple-choice questions, and the VideoMME overall set without subtitles. Generally, our disentangled architecture handles this conventional scenario well and achieves competitive performance. Notably, on EgoSchema, which requires long temporal reasoning, our method achieves a leading performance of 55.6, demonstrating strong temporal perception.", + "bbox": [ + 511, + 398, + 906, + 551 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "In terms of MLVU and VideoMME, which consist of videos ranging from minutes to hours long, our model's promising results highlight its ability to efficiently process contextual information across varying temporal lengths. This capability is crucial for streaming video interactions, as it allows the model to accurately summarize context when questions are inserted at arbitrary temporal positions.", + "bbox": [ + 511, + 559, + 908, + 666 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Additionally, we adopt a recent time-aware benchmark, ETBench, to evaluate temporal awareness. We select six subtasks that require the model to predict explicit event timestamps and report the results in Table 2. Following [41], we report the F1 score for temporal grounding evaluation and the sentence similarity score for caption evaluation. In the conventional setting, our method is able to capture the timestamps corresponding to specific questions. Without a specialized design for time tokens, our model achieves the highest F1 score in temporal video grounding and episodic memory subtasks, even surpassing models with specialized temporal tokens. The promising performance on dense video captioning and step localization further demonstrates the model's ability in both accurate temporal grounding and precise visual perception.", + "bbox": [ + 511, + 674, + 908, + 902 + ], + "page_idx": 5 + }, + { + "type": "page_number", + "text": "24050", + "bbox": [ + 478, + 944, + 517, + 955 + ], + "page_idx": 5 + }, + { + "type": "table", + "img_path": "images/6f854a17f657ab432b2da9ed51c16f3f14c94b74e230f0fb91d2b191fa987152.jpg", + "table_caption": [], + "table_footnote": [], + "table_body": "
MethodFramesTVGF1EPMF1TALF1VHDF1DVCF1DVCsimSLCF1SLCsim
Conventional video QA inference.
w/ specialized time tokens
VTG-LLM [24]9615.93.714.448.240.218.620.814.4
LITA [26]10022.24.618.023.939.717.221.012.2
ETChat [41]1 FPS38.610.230.862.538.419.724.414.6
w/o specialized time tokens
VideoChatGPT [43]1007.01.315.128.88.811.35.710.2
Video-LLaVA [35]87.01.915.028.928.015.00.98.3
LLaMA-VID [33]1 FPS5.51.28.030.027.112.65.211.1
Video-LLaMA2 [9]80.10.00.01,50.614.50.015.2
PLLAVA [60]166.91.15.728.913.310.69.711.8
VTimeLLM [25]1007.61.918.228.912.413.18.76.4
TimeChat [53]9626.23.910.140.516.612.55.69.2
Ours1 FPS43.617.229.951.531.617.814.111.7
Streaming video QA inference.
VideoLLM-Online [5]2 FPS13.23.89.122.424.013.49.910.1
Ours1 FPS36.115.527.354.233.818.918.812.4
", + "bbox": [ + 122, + 87, + 877, + 375 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Table 2. Comparison on ETBench. We present the results for two different settings. In the conventional video QA setting, the model is required to answer the question after watching the entire video. In the streaming setting, the question is placed at the beginning of the video, and the model is expected to provide real-time responses. We report performance on six subtasks that are suitable for both settings.", + "bbox": [ + 89, + 383, + 906, + 428 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "4.5. Streaming Video Understanding", + "text_level": 1, + "bbox": [ + 89, + 439, + 375, + 455 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "In addition to the conventional video QA evaluation, we also assess the performance of our model in streaming video interactions. Specifically, we place the question at the beginning of the video and expect the model to produce a response once it detects relevant clues while the video stream plays in real-time. The model should remain silent when no new information is available. In the bottom lines of Table 2, we compare our model with VideoLLM-online [5], the only existing video LLM capable of streaming video interactions. Our model consistently outperforms VideoLLM-online across all metrics, with significant advantages in temporal grounding. This demonstrates that the disentangled design makes our model have far more powerful temporal awareness, enabling more agile proactive responses aligned with specific instructions.", + "bbox": [ + 88, + 462, + 482, + 686 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Notably, on dense video captioning and step localization tasks, our model achieves both more precise temporal grounding and more comprehensive descriptions in streaming mode than the conventional setting. This indicates that our disentangled architecture can effectively monitor the video stream in real-time and proactively generate informative responses according to the instructions. And the streaming interaction has the potential to achieve more powerful video understanding.", + "bbox": [ + 88, + 688, + 482, + 824 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Additionally, we present a quantitative comparison between our model and VideoLLM-online in Fig. 3. From this comparison, when faced with questions requiring multi-step reasoning, our model can gradually identify the necessary clues from the video stream and generate informative", + "bbox": [ + 88, + 825, + 483, + 901 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "answers step by step, while VideoLLM-online [5] fails to do so. For example, from the 'thirsty' in the question, our model can associate it with the drinks appearing in the video stream, and infer what actions are needed based on the context. In contrast, VideoLLM-online only gives simple descriptions of the scene or ongoing actions.", + "bbox": [ + 511, + 440, + 906, + 530 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "A key advantage of our method is the disentangled perception, decision, and reaction architecture, which enables the model to simultaneously process the streaming video input and generate responses in a non-blocking manner. However, VideoLLM-online has to experience interruptions in the video stream during answer generation.", + "bbox": [ + 511, + 531, + 906, + 622 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "4.6. Ablation Study", + "text_level": 1, + "bbox": [ + 511, + 633, + 666, + 648 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Clip Segmentation Strategy. We first present an ablation study on clip segmentation. We compare the standard uniform clip segmentation (16 frames per clip sampled at 1 FPS) with our scene-based non-uniform segmentation. We report conventional QA accuracy on MLVU and VideoMME, as well as streaming metrics on temporal video grounding and dense video captioning on ETBench in Table 3. The scene-based segmentation yields superior results in both conventional and streaming settings. This aligns with our motivation that scene-based segmentation preserves more structural video information, which facilitates better model learning and more timely responses.", + "bbox": [ + 511, + 655, + 906, + 837 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Special Token Design. We also ablate three special tokens in our architecture, i.e., $\\langle \\mathrm{ANS} \\rangle$ and $\\langle \\mathrm{TODO} \\rangle$ in the streaming video processor, $\\langle \\mathrm{SILENT} \\rangle$ token in the final LLM. These three special tokens mainly impact the stream-", + "bbox": [ + 511, + 839, + 908, + 902 + ], + "page_idx": 6 + }, + { + "type": "page_number", + "text": "24051", + "bbox": [ + 478, + 944, + 517, + 955 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/4f71b6ac993e850e18f9a7bf89d33ecf0a440e0fc85f0e25fcbc7c5644522f71.jpg", + "image_caption": [ + "Figure 3. Quantitative comparison between Dispider and VideoLLM-online in streaming video understanding. The question is inserted at the beginning, and we show the model's response in either answer texts or the state of keeping silent." + ], + "image_footnote": [], + "bbox": [ + 116, + 87, + 883, + 311 + ], + "page_idx": 7 + }, + { + "type": "table", + "img_path": "images/3f2aa94061b82804b44d4e0ed489a5bc6c7afdb7869a03883bdebb4a6248ae62.jpg", + "table_caption": [], + "table_footnote": [], + "table_body": "
ClipMLVUV-MMETVG\\( _{F1} \\)DVC\\( _{F1} \\)DVC\\( _{Sim} \\)
Uniform59.855.434.533.118.1
Scene-based61.757.236.133.818.9
", + "bbox": [ + 91, + 375, + 485, + 433 + ], + "page_idx": 7 + }, + { + "type": "table", + "img_path": "images/d243d7ab0d36239c4fd3015321423d252696386104e0e69ea6709879eeb8735c.jpg", + "table_caption": [ + "Table 3. Ablation study on the clip segmentation. We compare uniform 16-frame clip segmentation and our scene-based segmentation with SigLip." + ], + "table_footnote": [], + "table_body": "
<ANS><TODO><SILENT>TVG\\( _{F1} \\)DVC\\( _{F1} \\)DVC\\( _{Sim} \\)
XXX20.119.712.3
XX✓26.324.913.1
✓X✓35.231.017.2
X✓✓28.725.614.5
✓✓X35.530.216.8
✓✓✓36.133.818.9
", + "bbox": [ + 94, + 493, + 478, + 606 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Table 4. Ablation study on the special token designs. We respectively explore the effectiveness of ⟨ANS⟩ and ⟨TODO⟩ in streaming video processor as well as ⟨SILENT⟩ in the final LLM.", + "bbox": [ + 89, + 617, + 482, + 659 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "ing video interactions, so we report the streaming evaluation metrics on temporal video grounding and dense video captioning in Table 4. There are three observations. First, the absence of $\\langle \\mathrm{ANS}\\rangle$ token can lead the model to be unaware of the timestamps at which a response has been given. As a result, if relevant clues have appeared in the video contexts, the model tends to produce a response, resulting in a high recall but a low accuracy. Second, the $\\langle \\mathrm{TODO}\\rangle$ serves as an indicator that reminds the streaming processor to decide whether to respond. The performance slightly degrades without this special token. Third, the $\\langle \\mathrm{SILENT}\\rangle$ token in the final LLM serves as a secondary filter for response decision. If the preceding streaming processor incorrectly identifies a timestamp as requiring a response, the $\\langle \\mathrm{SILENT}\\rangle$ token enables the LLM to rethink whether an answer is needed", + "bbox": [ + 89, + 674, + 483, + 900 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "referring to the historical QA interaction contexts.", + "bbox": [ + 511, + 377, + 844, + 393 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "5. Conclusion", + "text_level": 1, + "bbox": [ + 511, + 406, + 633, + 422 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "In this work, we introduced Dispider, a novel framework designed to enable active real-time interaction with video LLMs. By disentangling the key capabilities of perception, decision, and reaction, and adopting an asynchronous processing approach, Dispider overcomes the inherent conflicts that hinder real-time interaction in traditional video LLMs. Our approach ensures that the model can continuously process the video stream while providing timely, contextually accurate, and precise responses to user interactions. The Proactive Streaming Video Processing module optimizes the video analysis by focusing on relevant content, while the Precise Interaction module generates detailed responses asynchronously, allowing for uninterrupted video processing. We demonstrated the effectiveness of Dispider through extensive experiments on both conventional and streaming video QA benchmarks, where it outperformed existing methods in proactive response capabilities, temporal awareness, and computational efficiency. Dispider's disentangled architecture and ability to handle long-duration video streams make it an ideal solution for real-time interactive applications.", + "bbox": [ + 511, + 431, + 906, + 750 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Acknowledgement", + "text_level": 1, + "bbox": [ + 511, + 762, + 671, + 779 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "This project is funded in part by Shanghai Artificial Intelligence Laboratory, Shanghai Innovation Institute, the National Key R&D Program of China (2022ZD0160201), the Centre for Perceptual and Interactive Intelligence (CPII) Ltd under the Innovation and Technology Commission (ITC)'s InnoHK. Dahua Lin is a PI of CPII under the InnoHK.", + "bbox": [ + 511, + 787, + 906, + 892 + ], + "page_idx": 7 + }, + { + "type": "page_number", + "text": "24052", + "bbox": [ + 478, + 944, + 517, + 955 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "References", + "text_level": 1, + "bbox": [ + 91, + 89, + 187, + 104 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[1] Jean-Baptiste Alayrac, Jeff Donahue, Pauline Luc, Antoine Miech, Iain Barr, Yana Hasson, Karel Lenc, Arthur Mensch, Katherine Millican, Malcolm Reynolds, et al. Flamingo: a visual language model for few-shot learning. Advances in neural information processing systems, 35:23716-23736, 2022. 2", + "[2] Jinze Bai, Shuai Bai, Shusheng Yang, Shijie Wang, Sinan Tan, Peng Wang, Junyang Lin, Chang Zhou, and Jingren Zhou. Qwen-vl: A frontier large vision-language model with versatile abilities. arXiv preprint arXiv:2308.12966, 2023. 2", + "[3] Yuntao Bai, Andy Jones, Kamal Ndousse, Amanda Askell, Anna Chen, Nova DasSarma, Dawn Drain, Stanislav Fort, Deep Ganguli, Tom Henighan, et al. Training a helpful and harmless assistant with reinforcement learning from human feedback. arXiv preprint arXiv:2204.05862, 2022.", + "[4] Zheng Cai, Maosong Cao, Haojiong Chen, Kai Chen, Keyu Chen, Xin Chen, Xun Chen, Zehui Chen, Zhi Chen, Pei Chu, et al. Internl m2 technical report. arXiv preprint arXiv:2403.17297, 2024. 2", + "[5] Joya Chen, Zhaoyang Lv, Shiwei Wu, Kevin Qinghong Lin, Chenan Song, Difei Gao, Jia-Wei Liu, Ziteng Gao, Dongxing Mao, and Mike Zheng Shou. Videollm-online: Online video large language model for streaming video. In CVPR, 2024. 1, 2, 3, 7", + "[6] Lin Chen, Xilin Wei, Jinsong Li, Xiaoyi Dong, Pan Zhang, Yuhang Zang, Zehui Chen, Haodong Duan, Bin Lin, Zhenyu Tang, et al. Sharegpt4video: Improving video understanding and generation with better captions. arXiv preprint arXiv:2406.04325, 2024. 2, 6", + "[7] Ling-Hao Chen, Shunlin Lu, Ailing Zeng, Hao Zhang, Benyou Wang, Ruimao Zhang, and Lei Zhang. Motionllm: Understanding human behaviors from human motions and videos. arXiv preprint arXiv:2405.20340, 2024.", + "[8] Zhe Chen, Jiannan Wu, Wenhai Wang, Weijie Su, Guo Chen, Sen Xing, Muyan Zhong, Qinglong Zhang, Xizhou Zhu, Lewei Lu, Bin Li, Ping Luo, Tong Lu, Yu Qiao, and Jifeng Dai. Internvl: Scaling up vision foundation models and aligning for generic visual-linguistic tasks. arXiv preprint arXiv:2312.14238, 2023.", + "[9] Zesen Cheng, Sicong Leng, Hang Zhang, Yifei Xin, Xin Li, Guanzheng Chen, Yongxin Zhu, Wenqi Zhang, Ziyang Luo, Deli Zhao, et al. Videollama 2: Advancing spatial-temporal modeling and audio understanding in video-llms. arXiv preprint arXiv:2406.07476, 2024. 2, 6, 7", + "[10] Aakanksha Chowdhery et al. Palm: Scaling language modeling with pathways. arXiv preprint arXiv:2303.12345, 2023. 2", + "[11] Wenliang Dai, Junnan Li, Dongxu Li, Anthony Meng Huat Tiong, Junqi Zhao, Weisheng Wang, Boyang Li, Pascale Fung, and Steven Hoi. Instructclip: Towards general-purpose vision-language models with instruction tuning, 2023. 2", + "[12] Jacob Devlin, Ming-Wei Chang, Kenton Lee, and Kristina Toutanova. Bert: Pre-training of deep bidirectional transformers for language understanding. arXiv preprint arXiv:1810.04805, 2018. 2" + ], + "bbox": [ + 93, + 114, + 483, + 898 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[13] Shangzhe Di and Weidi Xie. Grounded question-answering in long egocentric videos. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 12934–12943, 2024. 5", + "[14] Shuangrui Ding, Maomao Li, Tianyu Yang, Rui Qian, Haohang Xu, Qingyi Chen, Jue Wang, and Hongkai Xiong. Motion-aware contrastive video representation learning via foreground-background merging. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 9716-9726, 2022. 1", + "[15] Shuangrui Ding, Peisen Zhao, Xiaopeng Zhang, Rui Qian, Hongkai Xiong, and Qi Tian. Prune spatio-temporal tokens by semantic-aware temporal accumulation. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 16945-16956, 2023. 1", + "[16] Shuangrui Ding, Zihan Liu, Xiaoyi Dong, Pan Zhang, Rui Qian, Conghui He, Dahua Lin, and Jiaqi Wang. Songcomposer: A large language model for lyric and melody composition in song generation. arXiv preprint arXiv:2402.17645, 2024. 2", + "[17] Shuangrui Ding, Rui Qian, Haohang Xu, Dahua Lin, and Hongkai Xiong. Betrayed by attention: A simple yet effective approach for self-supervised video object segmentation. In European Conference on Computer Vision, pages 215-233. Springer, 2024. 1", + "[18] Yue Fan, Xiaojian Ma, Rujie Wu, Yuntao Du, Jiaqi Li, Zhi Gao, and Qing Li. Videoagent: A memory-augmented multimodal agent for video understanding. arXiv preprint arXiv:2403.11481, 2024. 2", + "[19] Jiajun Fei, Dian Li, Zhidong Deng, Zekun Wang, Gang Liu, and Hui Wang. Video-ccam: Enhancing video-language understanding with causal cross-attention masks for short and long videos. arXiv preprint arXiv:2408.14023, 2024.6", + "[20] Chaoyou Fu, Yuhan Dai, Yondong Luo, Lei Li, Shuhuai Ren, Renrui Zhang, Zihan Wang, Chenyu Zhou, Yunhang Shen, Mengdan Zhang, et al. Video-mme: The first-ever comprehensive evaluation benchmark of multi-modal llms in video analysis. arXiv preprint arXiv:2405.21075, 2024. 5", + "[21] Chaoyou Fu, Haojia Lin, Zuwei Long, Yunhang Shen, Meng Zhao, Yifan Zhang, Xiong Wang, Di Yin, Long Ma, Xiawu Zheng, Ran He, Rongrong Ji, Yunsheng Wu, Caifeng Shan, and Xing Sun. Vita: Towards open-source interactive omni multimodal lIm. arXiv preprint arXiv:2408.05211, 2024. 2", + "[22] Gemini Team Google. Gemini: a family of highly capable multimodal models. arXiv preprint arXiv:2312.11805, 2023. 2", + "[23] Gemini Team Google. Gemini 1.5: Unlocking multimodal understanding across millions of tokens of context. arXiv preprint arXiv:2403.05530, 2024. 2", + "[24] Yongxin Guo, Jingyu Liu, Mingda Li, Xiaoying Tang, Xi Chen, and Bo Zhao. Vtg-llm: Integrating timestamp knowledge into video llms for enhanced video temporal grounding. arXiv preprint arXiv:2405.13382, 2024. 7", + "[25] Bin Huang, Xin Wang, Hong Chen, Zihan Song, and Wenwu Zhu. Vtimellm: Empower llm to grasp video moments. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 14271-14280, 2024. 7" + ], + "bbox": [ + 516, + 92, + 903, + 898 + ], + "page_idx": 8 + }, + { + "type": "page_number", + "text": "24053", + "bbox": [ + 478, + 944, + 517, + 955 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[26] De-An Huang, Shijia Liao, Subhashree Radhakrishnan, Hongxu Yin, Pavlo Molchanov, Zhiding Yu, and Jan Kautz. Lita: Language instructed temporal-localization assistant. In European Conference on Computer Vision, pages 202-218. Springer, 2025. 7", + "[27] Albert Q Jiang, Alexandre Sablayrolles, Arthur Mensch, Chris Bamford, Devendra Singh Chaplot, Diego de las Casas, Florian Bressand, Gianna Lengyel, Guillaume Lample, Lucile Saulnier, et al. Mistral 7b. arXiv preprint arXiv:2310.06825, 2023. 2", + "[28] Peng Jin, Ryuichi Takanobu, Caiwan Zhang, Xiaochun Cao, and Li Yuan. Chat-univi: Unified visual representation empowers large language models with image and video understanding. arXiv preprint arXiv:2311.08046, 2023. 6", + "[29] Bo Li, Yuanhan Zhang, Dong Guo, Renrui Zhang, Feng Li, Hao Zhang, Kaichen Zhang, Yanwei Li, Ziwei Liu, and Chunyuan Li. Llava-onevision: Easy visual task transfer. arXiv preprint arXiv:2408.03326, 2024. 5", + "[30] KunChang Li, Yinan He, Yi Wang, Yizhuo Li, Wenhai Wang, Ping Luo, Yali Wang, Limin Wang, and Yu Qiao. Videochat: Chat-centric video understanding. arXiv preprint arXiv:2305.06355, 2023. 2", + "[31] Kunchang Li, Yali Wang, Yinan He, Yizhuo Li, Yi Wang, Yi Liu, Zun Wang, Jilan Xu, Guo Chen, Ping Luo, et al. Mvbench: A comprehensive multi-modal video understanding benchmark. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 22195-22206, 2024. 6", + "[32] Yanwei Li, Chengyao Wang, and Jiaya Jia. Llama-vid: An image is worth 2 tokens in large language models. arXiv preprint arXiv:2311.17043, 2023. 1, 2", + "[33] Yanwei Li, Chengyao Wang, and Jiaya Jia. Llama-vid: An image is worth 2 tokens in large language models. 2024. 2, 6, 7", + "[34] Yifei Li, Junbo Niu, Ziyang Miao, Chunjiang Ge, Yuanhang Zhou, Qihao He, Xiaoyi Dong, Haodong Duan, Shuangrui Ding, Rui Qian, et al. Ovo-bench: How far is your videolms from real-world online video understanding? arXiv preprint arXiv:2501.05510, 2025. 3", + "[35] Bin Lin, Bin Zhu, Yang Ye, Munan Ning, Peng Jin, and Li Yuan. Video-llava: Learning united visual representation by alignment before projection. arXiv preprint arXiv:2311.10122, 2023. 6, 7", + "[36] Kevin Lin, Faisal Ahmed, Linjie Li, Chung-Ching Lin, Ehsan Azarnasab, Zhengyuan Yang, Jianfeng Wang, Lin Liang, Zicheng Liu, Yumao Lu, et al. Mm-vid: Advancing video understanding with gpt-4v (ision). arXiv preprint arXiv:2310.19773, 2023. 2", + "[37] Haotian Liu, Chunyuan Li, Yuheng Li, and Yong Jae Lee. Improved baselines with visual instruction tuning. arXiv preprint arXiv:2310.03744, 2023. 2", + "[38] Haotian Liu, Chunyuan Li, Qingyang Wu, and Yong Jae Lee. Visual instruction tuning, 2023.", + "[39] Haotian Liu, Chunyuan Li, Yuheng Li, Bo Li, Yuanhan Zhang, Sheng Shen, and Yong Jae Lee. Llava next: Improved reasoning,OCR, and world knowledge, 2024. 2", + "[40] Jiajun Liu, Yibing Wang, Hanghang Ma, Xiaoping Wu, Xiaqi Ma, Xiaoming Wei, Jianbin Jiao, Enhua Wu, and Jie Hu." + ], + "bbox": [ + 91, + 90, + 480, + 900 + ], + "page_idx": 9 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "Kangaroo: A powerful video-language model supporting long-context video input. arXiv preprint arXiv:2408.15542, 2024.6", + "[41] Ye Liu, Zongyang Ma, Zhongang Qi, Yang Wu, Chang Wen Chen, and Ying Shan. E.t. bench: Towards open-ended event-level video-language understanding. In Neural Information Processing Systems (NeurIPS), 2024. 5, 6, 7", + "[42] Xudong Lu, Yinghao Chen, Cheng Chen, Hui Tan, Boheng Chen, Yina Xie, Rui Hu, Guanxin Tan, Renshou Wu, Yan Hu, et al. Bluem-v-3b: Algorithm and system co-design for multimodal large language models on mobile devices. arXiv preprint arXiv:2411.10640, 2024. 2", + "[43] Muhammad Maaz, Hanoona Rasheed, Salman Khan, and Fahad Shahbaz Khan. Video-chatgpt: Towards detailed video understanding via large vision and language models. arXiv preprint arXiv:2306.05424, 2023. 2, 5, 7", + "[44] Karttikeya Mangalam, Raiymbek Akshulakov, and Jitendra Malik. Egoschema: A diagnostic benchmark for very long-form video language understanding. Advances in Neural Information Processing Systems, 36, 2024. 5", + "[45] OpenAI. Gpt-4 technical report. 2023. 2", + "[46] Long Ouyang, Jeffrey Wu, Xu Jiang, Diogo Almeida, Carroll Wainwright, Pamela Mishkin, Chong Zhang, Sandhini Agarwal, Katarina Slama, Alex Ray, et al. Training language models to follow instructions with human feedback. Advances in neural information processing systems, 35:27730-27744, 2022. 2", + "[47] Rui Qian, Shuangrui Ding, and Dahua Lin. Rethinking image-to-video adaptation: An object-centric perspective. In European Conference on Computer Vision, pages 329-348. Springer, 2024. 1", + "[48] Rui Qian, Xiaoyi Dong, Pan Zhang, Yuhang Zang, Shuangrui Ding, Dahua Lin, and Jiaqi Wang. Streaming long video understanding with large language models. In The Thirty-eighth Annual Conference on Neural Information Processing Systems, 2024. 3, 6", + "[49] Team Qwen. Qwen2 technical report, 2024. 6", + "[50] Alec Radford, Karthik Narasimhan, Tim Salimans, and Ilya Sutskever. Improving language understanding by generative pre-training. 2018. 2", + "[51] Alec Radford, Jong Wook Kim, Chris Hallacy, Aditya Ramesh, Gabriel Goh, Sandhini Agarwal, Girish Sastry, Amanda Askell, Pamela Mishkin, Jack Clark, et al. Learning transferable visual models from natural language supervision. In International conference on machine learning, pages 8748-8763, 2021. 6", + "[52] Colin Raffel, Noam Shazeer, Adam Roberts, Katherine Lee, Sharan Narang, Michael Matena, Yanqi Zhou, Wei Li, and Peter J Liu. Exploring the limits of transfer learning with a unified text-to-text transformer. JMLR, 2020. 2", + "[53] Shuhuai Ren, Linli Yao, Shicheng Li, Xu Sun, and Lu Hou. Timechat: A time-sensitive multimodal large language model for long video understanding. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 14313-14323, 2024. 2, 6, 7", + "[54] Yan Shu, Peitian Zhang, Zheng Liu, Minghao Qin, Junjie Zhou, Tiejun Huang, and Bo Zhao. Video-xl: Extra-long" + ], + "bbox": [ + 516, + 92, + 903, + 900 + ], + "page_idx": 9 + }, + { + "type": "page_number", + "text": "24054", + "bbox": [ + 478, + 944, + 519, + 955 + ], + "page_idx": 9 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "vision language model for hour-scale video understanding. arXiv preprint arXiv:2409.14485, 2024. 6", + "[55] Enxin Song, Wenhao Chai, Guanhong Wang, Yucheng Zhang, Haoyang Zhou, Feiyang Wu, Haozhe Chi, Xun Guo, Tian Ye, Yanting Zhang, et al. Moviechat: From dense token to sparse memory for long video understanding. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 18221-18232, 2024. 2, 6", + "[56] Hugo Touvron, Thibaut Lavril, Gautier Izacard, Xavier Martinet, Marie-Anne Lachaux, Timothée Lacroix, Baptiste Rozière, Naman Goyal, Eric Hambro, Faisal Azhar, et al. Llama: Open and efficient foundation language models. arXiv preprint arXiv:2302.13971, 2023. 2", + "[57] Ziyang Wang, Shoubin Yu, Elias Stengel-Eskin, Jaehong Yoon, Feng Cheng, Gedas Bertasius, and Mohit Bansal. Videotree: Adaptive tree-based video representation for llm reasoning on long videos. arXiv preprint arXiv:2405.19209, 2024. 2", + "[58] Qi Wu, Yubo Zhao, Yifan Wang, Yu-Wing Tai, and Chikeung Tang. Motionllm: Multimodal motion-language learning with large language models. arXiv preprint arXiv:2405.17013, 2024.", + "[59] Haiyang Xu, Qinghao Ye, Mingshi Yan, Yaya Shi, Jiabo Ye, Yuanhong Xu, Chenliang Li, Bin Bi, Qiuchen Qian, Wei Wang, Guohai Xu, Ji Zhang, Songfang Huang, Feiran Huang, and Jingren Zhou. mplug-2: A modularized multimodal foundation model across text, image and video. arXiv preprint arXiv:2302.00402, 2023. 2", + "[60] Lin Xu, Yilin Zhao, Daquan Zhou, Zhijie Lin, See Kiong Ng, and Jiashi Feng. Pllava: Parameter-free llava extension from images to videos for video dense captioning. arXiv preprint arXiv:2404.16994, 2024. 7", + "[61] Qinghao Ye, Haiyang Xu, Jiabo Ye, Ming Yan, Anwen Hu, Haowei Liu, Qi Qian, Ji Zhang, and Fei Huang. mplug-owl2: Revolutionizing multi-modal large language model with modality collaboration. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 13040–13051, 2024. 2", + "[62] Alex Young, Bei Chen, Chao Li, Chengen Huang, Ge Zhang, Guanwei Zhang, Heng Li, Jiangcheng Zhu, Jianqun Chen, Jing Chang, et al. Yi: Open foundation models by 01. ai. arXiv preprint arXiv:2403.04652, 2024. 2", + "[63] Xiaohua Zhai, Basil Mustafa, Alexander Kolesnikov, and Lucas Beyer. Sigmoid loss for language image pre-training. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 11975-11986, 2023. 3", + "[64] Ce Zhang, Taixi Lu, Md Mohaiminul Islam, Ziyang Wang, Shoubin Yu, Mohit Bansal, and Gedas Bertasius. A simple llm framework for long-range video question-answering. arXiv preprint arXiv:2312.17235, 2023. 2", + "[65] Haoji Zhang, Yiqin Wang, Yansong Tang, Yong Liu, Jiashi Feng, Jifeng Dai, and Xiaojie Jin. Flash-vstream: Memory-based real-time understanding for long video streams. 2024. 3", + "[66] Pan Zhang, Xiaoyi Dong, Bin Wang, Yuhang Cao, Chao Xu, Linke Ouyang, Zhiyuan Zhao, Haodong Duan, Songyang Zhang, Shuangrui Ding, et al. Internlmxcomposer: A vision-language large model for advanced" + ], + "bbox": [ + 91, + 92, + 482, + 900 + ], + "page_idx": 10 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "text-image comprehension and composition. arXiv preprint arXiv:2309.15112, 2023. 2", + "[67] Pan Zhang, Xiaoyi Dong, Yuhang Cao, Yuhang Zang, Rui Qian, Xilin Wei, Lin Chen, Yifei Li, Junbo Niu, Shuangrui Ding, et al. Internlm-xcomposer2. 5-omnilive: A comprehensive multimodal system for long-term streaming video and audio interactions. arXiv preprint arXiv:2412.09596, 2024.3", + "[68] Pan Zhang, Xiaoyi Dong, Yuhang Zang, Yuhang Cao, Rui Qian, Lin Chen, Qipeng Guo, Haodong Duan, Bin Wang, Linke Ouyang, et al. Internlm-xcomposer-2.5: A versatile large vision language model supporting long-contextual input and output. arXiv preprint arXiv:2407.03320, 2024. 2", + "[69] Peiyuan Zhang, Kaichen Zhang, Bo Li, Guangtao Zeng, Jingkang Yang, Yuanhan Zhang, Ziyue Wang, Haoran Tan, Chunyuan Li, and Ziwei Liu. Long context transfer from language to vision. arXiv preprint arXiv:2406.16852, 2024. 6", + "[70] Yuanhan Zhang, Bo Li, haotian Liu, Yong jae Lee, Liangke Gui, Di Fu, Jiashi Feng, Ziwei Liu, and Chunyuan Li. Llavanext: A strong zero-shot video understanding model, 2024. 6", + "[71] Junjie Zhou, Yan Shu, Bo Zhao, Boya Wu, Shitao Xiao, Xi Yang, Yongping Xiong, Bo Zhang, Tiejun Huang, and Zheng Liu. Mlvu: A comprehensive benchmark for multi-task long video understanding. arXiv preprint arXiv:2406.04264, 2024.5" + ], + "bbox": [ + 516, + 92, + 903, + 470 + ], + "page_idx": 10 + }, + { + "type": "page_number", + "text": "24055", + "bbox": [ + 478, + 944, + 517, + 955 + ], + "page_idx": 10 + } +] \ No newline at end of file diff --git a/2025/Dispider_ Enabling Video LLMs with Active Real-Time Interaction via Disentangled Perception, Decision, and Reaction/509357f7-50b0-4a3e-be02-5d53b2c39c5b_model.json b/2025/Dispider_ Enabling Video LLMs with Active Real-Time Interaction via Disentangled Perception, Decision, and Reaction/509357f7-50b0-4a3e-be02-5d53b2c39c5b_model.json new file mode 100644 index 0000000000000000000000000000000000000000..5f2c8237940a7041e025cdbbd259484a4dbb3c3e --- /dev/null +++ b/2025/Dispider_ Enabling Video LLMs with Active Real-Time Interaction via Disentangled Perception, Decision, and Reaction/509357f7-50b0-4a3e-be02-5d53b2c39c5b_model.json @@ -0,0 +1,2433 @@ +[ + [ + { + "type": "header", + "bbox": [ + 0.107, + 0.003, + 0.182, + 0.043 + ], + "angle": 0, + "content": "CVF" + }, + { + "type": "header", + "bbox": [ + 0.237, + 0.001, + 0.812, + 0.047 + ], + "angle": 0, + "content": "This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore." + }, + { + "type": "title", + "bbox": [ + 0.143, + 0.13, + 0.856, + 0.177 + ], + "angle": 0, + "content": "Dispider: Enabling Video LLMs with Active Real-Time Interaction via Disentangled Perception, Decision, and Reaction" + }, + { + "type": "text", + "bbox": [ + 0.241, + 0.203, + 0.761, + 0.24 + ], + "angle": 0, + "content": "Rui Qian\\(^{1*}\\) Shuangrui Ding\\(^{1*}\\) Xiaoyi Dong\\(^{1,2\\dagger}\\) Pan Zhang\\(^{2}\\) Yuhang Zang\\(^{2}\\) Yuhang Cao\\(^{2}\\) Dahua Lin\\(^{1,2,4}\\) Jiaqi Wang\\(^{2,3\\dagger}\\)" + }, + { + "type": "text", + "bbox": [ + 0.237, + 0.24, + 0.761, + 0.275 + ], + "angle": 0, + "content": "1 The Chinese University of Hong Kong 2 Shanghai AI Laboratory \n3 Shanghai Innovation Institute 4 CPII under InnoHK" + }, + { + "type": "text", + "bbox": [ + 0.368, + 0.278, + 0.625, + 0.292 + ], + "angle": 0, + "content": "{qr021, ds023}@ie.cuhk.edu.hk" + }, + { + "type": "title", + "bbox": [ + 0.248, + 0.327, + 0.327, + 0.344 + ], + "angle": 0, + "content": "Abstract" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.362, + 0.486, + 0.846 + ], + "angle": 0, + "content": "Active real-time interaction with video LLMs introduces a new paradigm for human-computer interaction, where the model not only understands user intent but also responds while continuously processing streaming video on the fly. Unlike offline video LLMs, which analyze the entire video before answering questions, active real-time interaction requires three capabilities: 1) Perception: real-time video monitoring and interaction capturing. 2) Decision: raising proactive interaction in proper situations, 3) Reaction: continuous interaction with users. However, inherent conflicts exist among the desired capabilities. The Decision and Reaction require a contrary Perception scale and grain, and the autoregressive decoding blocks the real-time Perception and Decision during the Reaction. To unify the conflicted capabilities within a harmonious system, we present Dispider, a solution built on a Disentangled Perception, Decision, and Reaction framework. Dispider features a lightweight Proactive Streaming Video Processing module that tracks the video stream and identifies optimal moments for interaction. Once the interaction is triggered, an asynchronous Precise Interaction module provides detailed responses, while the processing module continues to monitor the video in the meantime. Our disentangled and asynchronous design ensures timely, contextually accurate, and computationally efficient responses, making Dispider ideal for active real-time interaction for long-duration video streams. Experiments prove that Dispider outperforms existing methods not only in its superior understanding of video content in conventional video QA settings, but also in proactive response capability and temporal awareness under the streaming setting. The code and model are released at https://github.com/Mark12Ding/Dispider." + }, + { + "type": "title", + "bbox": [ + 0.515, + 0.327, + 0.645, + 0.343 + ], + "angle": 0, + "content": "1. Introduction" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.353, + 0.907, + 0.518 + ], + "angle": 0, + "content": "The recent surge in multimodal large language models (MLLMs) has brought considerable attention to video understanding tasks [14, 15, 17, 47]. Given a video, these models aim to accurately comprehend its content and provide responses aligned with human expectations. However, the majority of current video LLMs are designed around an offline setting, where models are required to view the entire video before generating a single response. While effective for certain applications, this offline approach is impractical for real-time\\(^{1}\\), interactive scenarios where users expect continuous feedback as video streams." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.518, + 0.909, + 0.715 + ], + "angle": 0, + "content": "Consider a scenario where a user performs pull-ups with the assistance of a visual model. As illustrated in Figure 1, current offline video LLMs need to process the entire video before providing the total count of pull-ups. However, users expect real-time feedback, with each pull-up counted as it occurs: \"One... Two... Three...\" This real-time interaction requires the model to recognize each completed action instantly and synchronize its responses with the user's movements. The challenge lies in determining the precise moment to output the count along with the ongoing video stream. To this end, traditional offline video LLMs lack the flexibility to handle such demands, as they are not designed for continuous, dynamic video streams." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.715, + 0.909, + 0.868 + ], + "angle": 0, + "content": "VideoLLM-online [5] has been a pioneering effort in addressing this issue. By leveraging the LLaMA model [32], they enable proactive response updates during streaming video inputs. However, it has a critical limitation: since it uses a single LLM for video processing and response generation, it cannot perform perception and answer reaction simultaneously. The autoregressive nature of next-token prediction in LLMs forces VideoLLM-online to alternate between perception (processing frames) and response generation, rather than handling both in parallel. This restriction" + }, + { + "type": "page_footnote", + "bbox": [ + 0.513, + 0.875, + 0.909, + 0.902 + ], + "angle": 0, + "content": "1We use \"real-time\" loosely to emphasize online interaction, not strict frame-rate requirements." + }, + { + "type": "page_footnote", + "bbox": [ + 0.109, + 0.875, + 0.223, + 0.888 + ], + "angle": 0, + "content": "*Equal Contribution." + }, + { + "type": "page_footnote", + "bbox": [ + 0.111, + 0.888, + 0.24, + 0.9 + ], + "angle": 0, + "content": "† Corresponding Author." + }, + { + "type": "page_number", + "bbox": [ + 0.479, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "24045" + } + ], + [ + { + "type": "header", + "bbox": [ + 0.117, + 0.102, + 0.139, + 0.12 + ], + "angle": 0, + "content": "#" + }, + { + "type": "header", + "bbox": [ + 0.143, + 0.106, + 0.339, + 0.119 + ], + "angle": 0, + "content": ": How many pull-ups has the man done?" + }, + { + "type": "header", + "bbox": [ + 0.571, + 0.103, + 0.713, + 0.119 + ], + "angle": 0, + "content": "Offline Video LLM" + }, + { + "type": "header", + "bbox": [ + 0.718, + 0.102, + 0.744, + 0.121 + ], + "angle": 0, + "content": "#" + }, + { + "type": "header", + "bbox": [ + 0.745, + 0.106, + 0.894, + 0.118 + ], + "angle": 0, + "content": ": The man has done 3 pull-ups." + }, + { + "type": "image", + "bbox": [ + 0.095, + 0.138, + 0.842, + 0.243 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.107, + 0.261, + 0.128, + 0.277 + ], + "angle": 0, + "content": null + }, + { + "type": "text", + "bbox": [ + 0.131, + 0.265, + 0.261, + 0.277 + ], + "angle": 0, + "content": "Real-time Video LLM" + }, + { + "type": "image", + "bbox": [ + 0.296, + 0.263, + 0.322, + 0.282 + ], + "angle": 0, + "content": null + }, + { + "type": "text", + "bbox": [ + 0.325, + 0.267, + 0.44, + 0.279 + ], + "angle": 0, + "content": ": The man did 1 pull-up." + }, + { + "type": "image", + "bbox": [ + 0.483, + 0.264, + 0.506, + 0.282 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.544, + 0.267, + 0.569, + 0.28 + ], + "angle": 0, + "content": null + }, + { + "type": "text", + "bbox": [ + 0.57, + 0.268, + 0.69, + 0.28 + ], + "angle": 0, + "content": ": The man did 2 pull-ups." + }, + { + "type": "image", + "bbox": [ + 0.704, + 0.262, + 0.731, + 0.282 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.741, + 0.267, + 0.767, + 0.28 + ], + "angle": 0, + "content": null + }, + { + "type": "text", + "bbox": [ + 0.77, + 0.267, + 0.886, + 0.279 + ], + "angle": 0, + "content": "The man did 3 pull-ups." + }, + { + "type": "image_caption", + "bbox": [ + 0.09, + 0.308, + 0.907, + 0.364 + ], + "angle": 0, + "content": "Figure 1. Comparison between offline and active real-time Video LLMs in processing a video sequence of a man performing pull-ups. The user asks the model to count the number of pull-ups. The offline Video LLMs process the entire video before providing a final count of three pull-ups. In contrast, the real-time Video LLMs provide immediate feedback after each pull-up, first counting one, then two, and finally reaching three, while remaining silent between responses." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.391, + 0.483, + 0.452 + ], + "angle": 0, + "content": "delays the video inputs during responding, reduces responsiveness and hinders its ability to engage in real-time interactions effectively, especially during long-duration video streams that require continuous, uninterrupted processing." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.457, + 0.484, + 0.715 + ], + "angle": 0, + "content": "To address these limitations, we present Dispider, a novel framework designed for active real-time interaction with streaming videos. Dispider resolves the conflicts between perception, decision, and reaction through a disentangled, asynchronous design, where each component can operate in parallel. Specifically, we implement these functions in the Proactive Streaming Video Processing module and the Precise Interaction module. In the video processing module, with the streaming video input, we dynamically segment the video stream into non-uniform clips based on scene boundaries. These clips, combined with relevant historical context, are then processed by a real-time response decision module, which determines whether the model should generate a response or wait for additional video content. When an interaction is triggered, the Precise Interaction module generates context-aware responses without interrupting the ongoing video processing." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.72, + 0.484, + 0.902 + ], + "angle": 0, + "content": "This disentangled design ensures that the Dispider can provide timely, accurate, and computationally efficient responses, even for long-duration video streams. We evaluate Dispider's ability in both real-time video stream interactions, and conventional offline Video LLMs across long-video benchmarks (EgoSchema, VideoMME, MLVU) and time-sensitive tasks (ETBench). Dispider demonstrated strong performance, particularly excelling in temporal reasoning on EgoSchema and handling diverse video lengths effectively. In streaming tasks, it significantly surpasses VideoLLM-online [5] in temporal grounding, proactive response generation, and multi-step reasoning." + }, + { + "type": "title", + "bbox": [ + 0.514, + 0.389, + 0.655, + 0.404 + ], + "angle": 0, + "content": "2. Related Work" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.416, + 0.907, + 0.657 + ], + "angle": 0, + "content": "Large Language Models. LLMs have revolutionized natural language processing, achieving remarkable performance across a wide range of tasks. Early models such as BERT [12] and T5 [52] used masked language modeling for pre-training. The shift to decoder-only models like GPT [50] introduced scalable architectures that enhanced language generation. Recent models, including PaLM [10], LLaMA [56], and GPT-4 [45], continue to push the boundaries with massive parameters and extensive training data. Techniques such as supervised fine-tuning and reinforcement learning from human feedback [2-4, 16, 22, 23, 27, 32, 46, 62] have further improved these models' ability to generate coherent, contextually relevant responses. Inspired by the reasoning capabilities of LLMs, we extend these models to the domain of streaming video understanding, where real-time interaction presents unique challenges." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.659, + 0.909, + 0.901 + ], + "angle": 0, + "content": "Video Large Language Models. The progress of multimodal LLMs has been significant, particularly with image-based models like InstructBLIP [11], LLaVA [37-39], and Flamingo [1], which integrate vision and language models. Extending this to video introduces challenges in managing both frame sequences and the context length of LLMs. Recent works on video LLMs [6-9, 18, 21, 33, 36, 42, 43, 55, 57-59, 61, 64, 66, 68] present novel strategies for processing long-form videos while maintaining effective reasoning. Notable approaches include TimeChat [53], which emphasizes temporal relationships across video frames, and MovieChat [55], which uses sparse memory to handle long videos. VideoChat [30] adopts a chat-centric approach, integrating video foundation models with LLMs to excel at spatiotemporal reasoning, event localization, and causal inference. Building on these video LLMs, our work intro" + }, + { + "type": "page_number", + "bbox": [ + 0.48, + 0.946, + 0.519, + 0.957 + ], + "angle": 0, + "content": "24046" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.09, + 0.092, + 0.482, + 0.121 + ], + "angle": 0, + "content": "duces a pipeline designed to process streaming video inputs and generate real-time outputs." + }, + { + "type": "text", + "bbox": [ + 0.093, + 0.123, + 0.484, + 0.424 + ], + "angle": 0, + "content": "Streaming Video Understanding. In practice, users typically expect models to operate online and interactively in real time. This is known as streaming video understanding, which involves processing continuous video streams while ensuring long-term temporal consistency and enabling interactive responses. Despite its practical significance, only a few works have explored this area [34, 67]. VideoLLM-online [5] introduces the LIVE framework for streaming dialogue, but it lacks an efficient module for handling long-term video inputs over extended periods and does not prioritize real-time interactivity. VideoStream [48] proposes memory-propagated encoding for long videos, yet its focus is on offline processing rather than real-time streaming. Similarly, Flash-VStream [65] addresses inference latency and memory efficiency for long video streams but overlooks real-time user interaction. In contrast, our approach centers on developing a real-time visual assistant for streaming video, emphasizing both long-context handling and interactive, real-time responses, thereby bridging the gap left by previous methods." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.438, + 0.182, + 0.453 + ], + "angle": 0, + "content": "3. Method" + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.463, + 0.295, + 0.478 + ], + "angle": 0, + "content": "3.1. Problem Formulation" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.484, + 0.483, + 0.635 + ], + "angle": 0, + "content": "In contrast to previous offline video LLMs, which generate responses only after processing the entire video, our approach requires the model to operate in real-time, providing continuous and interactive responses as the video is processed. Specifically, the model actively determines when sufficient information has been observed to provide a complete response, allowing it to produce answers promptly without waiting for the entire video to finish. Following the formulation of VideoLLM-online [5], we formulate this new setting as below." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.636, + 0.483, + 0.741 + ], + "angle": 0, + "content": "Given a continuous video stream \\( V = \\{v_{1}, v_{2}, \\dots, v_{T}\\} \\), where \\( v_{i} \\) represents the \\( i \\)-th video clip, and a context sequence \\( C_{t} \\) up to a specific time \\( t \\) (which includes prior vision-language interactions such as historical user queries, assistant responses, and video frames), our goal is to generate timely and accurate dialogue responses \\( R_{t} \\) without processing the entire video sequence." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.742, + 0.483, + 0.817 + ], + "angle": 0, + "content": "At each time \\( t \\), the model observes \\( C_t \\) and the video frames up to that time \\( V_{[0,t]} = \\{v_1, v_2, \\ldots, v_t\\} \\). The model must decide whether it has sufficient information to generate a response. We define a decision function \\( \\pi \\) and a response function \\( f \\):" + }, + { + "type": "equation", + "bbox": [ + 0.162, + 0.832, + 0.412, + 0.849 + ], + "angle": 0, + "content": "\\[\na _ {t} = \\pi \\left(C _ {t}, V _ {[ 0, t ]}\\right) \\in \\{\\text {w a i t , r e s p o n d} \\}.\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.11, + 0.856, + 0.435, + 0.871 + ], + "angle": 0, + "content": "If \\( a_{t} = \\) respond, the model generates a response:" + }, + { + "type": "equation", + "bbox": [ + 0.225, + 0.885, + 0.35, + 0.903 + ], + "angle": 0, + "content": "\\[\nR _ {t} = f \\left(C _ {t}, V _ {[ 0, t ]}\\right)\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.092, + 0.884, + 0.107 + ], + "angle": 0, + "content": "Otherwise, the model keeps silent for more information." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.108, + 0.905, + 0.289 + ], + "angle": 0, + "content": "In this work, we implement these functions using a disentangled framework composed of three distinct modules: Perception, Decision, and Reaction. These modules are designed to handle the unique challenges of real-time video understanding and dialogue generation. The Perception module focuses on continuous video monitoring, while the Decision module assesses when to act, and the Reaction module generates responses without waiting for the entire video sequence to finish. This approach allows the system to remain responsive to new information while ensuring that the generated answers are based on the most relevant and up-to-date context." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.299, + 0.842, + 0.315 + ], + "angle": 0, + "content": "3.2. Proactive Streaming Video Processing" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.321, + 0.905, + 0.472 + ], + "angle": 0, + "content": "To enable active real-time response decisions, we propose a proactive streaming video processing approach that dynamically segments the video stream and evaluates whether to generate a response as the video is being processed. This approach balances computational efficiency with the need for timely and accurate responses, ensuring that the model processes the most relevant parts of the video while avoiding unnecessary redundancy. The system is composed of two key modules: the Scene-based Perception Module and the Real-time Response Decision Module." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.475, + 0.905, + 0.581 + ], + "angle": 0, + "content": "Scene-based Perception Module. To ensure efficient processing of long video streams, we begin by adaptively segmenting the video into non-uniform clips based on scene boundaries. This segmentation strategy preserves the structural information of the video, allowing the model to focus on the most informative parts while removing redundancy and maintaining context." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.581, + 0.905, + 0.732 + ], + "angle": 0, + "content": "We begin by sampling frames at a regular interval (one frame per second) and extract L2-normalized feature embeddings using the pre-trained SigLip model [63]. By computing the cosine similarity between these embeddings, we can identify significant visual changes, which indicate potential scene boundaries. These boundaries help divide the video into meaningful segments. To prevent excessively short clips, we introduce an exclusion window around the boundaries, ensuring that the resulting clips are of sufficient length to contain relevant information." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.733, + 0.905, + 0.807 + ], + "angle": 0, + "content": "Each clip \\( v_{i} \\) is then processed by the video encoder to produce the clip-wise feature representation \\( F_{i} \\), along with a special clip indicator \\( \\hat{F}_{i} \\). These clip features are used in the subsequent decision-making process to determine if enough information has been gathered to respond." + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.811, + 0.905, + 0.901 + ], + "angle": 0, + "content": "Real-time Response Decision Module. The Real-time Response Decision Module evaluates whether the model should generate a response or continue waiting for more video content. This decision is made based on the video content observed so far and the historical context, including the user's prior questions and assistant responses. We illus" + }, + { + "type": "page_number", + "bbox": [ + 0.479, + 0.945, + 0.521, + 0.957 + ], + "angle": 0, + "content": "24047" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.095, + 0.087, + 0.9, + 0.441 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.089, + 0.453, + 0.907, + 0.511 + ], + "angle": 0, + "content": "Figure 2. Illustration of the Response Decision Module in a proactive streaming video processing system. The module dynamically determines when to respond during video streaming by segmenting the video into non-uniform clips and utilizing historical memory to capture context. The module then combines memory features, clip features, question text, and special tokens, \\(\\langle \\mathrm{TODO}\\rangle\\) and \\(\\langle \\mathrm{ANS}\\rangle\\), to decide on response timing." + }, + { + "type": "text", + "bbox": [ + 0.089, + 0.524, + 0.482, + 0.584 + ], + "angle": 0, + "content": "trate this whole process in Figure 2. To effectively combine these multi-modal inputs, we use an interleaved sequence format, which integrates video features, question information, and decision tokens." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.585, + 0.483, + 0.66 + ], + "angle": 0, + "content": "The process begins with the sequence of video clip representations up to the point of the user's question, followed by the question \\( Q \\) and a special token, \\( \\langle \\mathrm{TODO} \\rangle \\), which indicates that a response decision is yet to be made. Formally, the input sequence at this stage is:" + }, + { + "type": "equation", + "bbox": [ + 0.193, + 0.67, + 0.381, + 0.687 + ], + "angle": 0, + "content": "\\[\nF [ 0: q _ {\\text {p o s}} ] \\oplus Q \\oplus \\langle \\text {T O D O} \\rangle ,\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.698, + 0.483, + 0.725 + ], + "angle": 0, + "content": "where \\(q_{\\mathrm{pos}}\\) is the timestamp of the question, and \\(\\oplus\\) denotes concatenation." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.728, + 0.483, + 0.819 + ], + "angle": 0, + "content": "After this initial stage, the model aggregates the features up to the question timestamp into a global memory \\(M\\) through pooling. Then, we extend the sequence to include this global memory that captures historical context, the question texts, and the clip features from the question timestamp to the current point, ending with \\(\\langle \\mathrm{TODO}\\rangle\\):" + }, + { + "type": "equation", + "bbox": [ + 0.153, + 0.829, + 0.421, + 0.847 + ], + "angle": 0, + "content": "\\[\nM [ 0: q _ {\\mathrm {p o s}} ] \\oplus Q \\oplus F [ q _ {\\mathrm {p o s}}: ] \\oplus \\langle \\mathrm {T O D O} \\rangle .\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.856, + 0.484, + 0.902 + ], + "angle": 0, + "content": "When the model decides to respond, the \\(\\langle \\mathrm{ANS} \\rangle\\) token is used to mark the response action. Specifically, with the video continuing to stream, we further integrate global" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.524, + 0.907, + 0.614 + ], + "angle": 0, + "content": "memory from the question timestamp to the position of the first answer, insert the special answer token \\(\\langle \\mathrm{ANS}\\rangle\\), and include later clips for potential subsequent responses. This sequence concludes with the current clip memory segment and a \\(\\langle \\mathrm{TODO}\\rangle\\) token. Thus, the ultimate input format with multiple answers can be written as:" + }, + { + "type": "equation", + "bbox": [ + 0.622, + 0.624, + 0.747, + 0.642 + ], + "angle": 0, + "content": "\\[\nM [ 0: q _ {\\mathrm {p o s}} ] \\oplus Q \\oplus\n\\]" + }, + { + "type": "equation", + "bbox": [ + 0.625, + 0.643, + 0.796, + 0.663 + ], + "angle": 0, + "content": "\\[\nM \\left[ q _ {\\text {p o s}}: a _ {\\text {p o s}} ^ {1} \\right] \\oplus \\langle \\mathrm {A N S} \\rangle \\oplus\n\\]" + }, + { + "type": "equation", + "bbox": [ + 0.625, + 0.664, + 0.796, + 0.684 + ], + "angle": 0, + "content": "\\[\nM \\left[ a _ {\\text {p o s}} ^ {1}: a _ {\\text {p o s}} ^ {2} \\right] \\oplus \\langle \\mathrm {A N S} \\rangle \\oplus\n\\]" + }, + { + "type": "equation", + "bbox": [ + 0.628, + 0.69, + 0.649, + 0.699 + ], + "angle": 0, + "content": "\\[\n\\cdot \\cdot \\cdot\n\\]" + }, + { + "type": "equation", + "bbox": [ + 0.624, + 0.705, + 0.766, + 0.725 + ], + "angle": 0, + "content": "\\[\nF \\left[ a _ {\\text {p o s}} ^ {k}: \\right] \\oplus \\langle \\text {T O D O} \\rangle ,\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.735, + 0.884, + 0.751 + ], + "angle": 0, + "content": "where \\(a_{\\mathrm{pos}}^k\\) stands for the timestamp for the \\(k\\)-th answer." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.751, + 0.906, + 0.886 + ], + "angle": 0, + "content": "We feed the interleaved sequence into a compact LLM and adopt a binary classification head on top of the final-layer embedding of the \\(\\langle \\mathrm{TODO}\\rangle\\) token. This head is trained to predict whether the model should respond at each timestep. We use standard binary cross-entropy loss to supervise the model's decision-making process. When the model determines that a response is required, the relevant historical clips are retrieved based on their cosine similarity with the \\(\\langle \\mathrm{TODO}\\rangle\\) token's embedding." + }, + { + "type": "text", + "bbox": [ + 0.533, + 0.886, + 0.905, + 0.902 + ], + "angle": 0, + "content": "To train the model's capacity on temporal retrieval, we" + }, + { + "type": "page_number", + "bbox": [ + 0.479, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "24048" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.09, + 0.092, + 0.483, + 0.166 + ], + "angle": 0, + "content": "calculate the cosine similarity between the embedding of the \\(\\langle \\mathrm{TODO}\\rangle\\) token and each historical clip indicator \\(\\hat{F}_i\\). This similarity yields a probability distribution \\(\\hat{P}_{\\mathrm{pred}}(i)\\) representing the relevance of \\(i\\)-th historical clip to the current response." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.168, + 0.483, + 0.226 + ], + "angle": 0, + "content": "For the ground truth, we use related clips to create a true temporal distribution. These clips are encoded as a binary mask over the clip intervals relevant to the response, defined as:" + }, + { + "type": "equation", + "bbox": [ + 0.15, + 0.226, + 0.424, + 0.257 + ], + "angle": 0, + "content": "\\[\n\\hat {P} _ {\\text {t r u e}} (i) = \\frac {1}{| R _ {Q} |} \\quad \\text {i f} \\quad i \\in R _ {Q} \\quad \\text {e l s e} \\quad 0,\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.262, + 0.483, + 0.292 + ], + "angle": 0, + "content": "where \\( R_{Q} \\) is the ground-truth set of relevant clips for particular question \\( Q \\)." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.293, + 0.483, + 0.336 + ], + "angle": 0, + "content": "We then calculate the KL divergence loss between the true distribution, \\(\\hat{P}_{\\mathrm{true}}\\), and the predicted distribution, \\(\\hat{P}_{\\mathrm{pred}}\\), as:" + }, + { + "type": "equation", + "bbox": [ + 0.142, + 0.338, + 0.432, + 0.378 + ], + "angle": 0, + "content": "\\[\n\\mathcal {L} _ {\\mathrm {K L}} = - \\frac {1}{| R _ {Q} |} \\sum_ {i = 1} ^ {|} R _ {Q} | \\hat {P} _ {\\mathrm {t r u e}} (i) \\log \\hat {P} _ {\\mathrm {p r e d}} (i).\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.385, + 0.483, + 0.461 + ], + "angle": 0, + "content": "This KL divergence loss encourages the model to align its predictions with the true temporal relevance of the clips, thereby improving its ability to correctly retrieve and ground the relevant video segments for generating accurate, timely responses." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.47, + 0.444, + 0.486 + ], + "angle": 0, + "content": "3.3. Precise Interaction with Streaming Video" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.493, + 0.483, + 0.643 + ], + "angle": 0, + "content": "Once the interaction is triggered, an asynchronous precise interaction module generates a response with fine-grained detail, tailored to the specific context of the video stream. That is, for the \\(k\\)-th answer outputs, the precise interaction module receives the question, the former \\(k - 1\\) history answers as well as the grounded clip features at the current timestamp. This design allows the system to generate appropriate responses by incorporating rich context from both the video content and the ongoing conversation in a streaming setting." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.645, + 0.483, + 0.78 + ], + "angle": 0, + "content": "A key advantage of this asynchronous approach is that it decouples the response generation from the video stream processing. This separation allows the streaming video to continue being processed in parallel, without waiting for response generation to complete. As a result, the system remains highly efficient and responsive, ensuring that the video content is continuously monitored and processed while still generating contextually accurate responses at the appropriate moments." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.781, + 0.483, + 0.902 + ], + "angle": 0, + "content": "While the Response Decision Module may determine that a response is needed, there is a chance that no actual output is required. To ensure accuracy in real-time interactions, we enhance the model's responsiveness by training it on a streaming QA dataset that simulates such real-time scenarios. These data include both positive samples (where a response is required) and negative samples (where no response is necessary and the model should keep silent). By" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.092, + 0.905, + 0.168 + ], + "angle": 0, + "content": "dynamically sampling these positive and negative instances, the model learns to either produce the responses that are incremental to the historical answers, summarizing the newly emerging cues in the video stream or output the special \\(\\langle \\mathrm{SILENT} \\rangle\\) token denoting keeping silent." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.168, + 0.907, + 0.259 + ], + "angle": 0, + "content": "This adaptive reasoning approach ensures that the model remains responsive to new information while avoiding redundancy. By focusing on unanswered content and utilizing past interactions, the model delivers timely and relevant responses, enhancing the overall efficiency and user experience in streaming dialogue generation." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.274, + 0.646, + 0.291 + ], + "angle": 0, + "content": "4. Experiments" + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.3, + 0.615, + 0.314 + ], + "angle": 0, + "content": "4.1. Datasets" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.323, + 0.905, + 0.472 + ], + "angle": 0, + "content": "The training process comprises two parts: response decision with temporal grounding and LLM reasoning. For training the streaming video processor and response decision module, we combine the GroundVQA [13] and ET-Instruct [41] datasets, enriched with temporal annotations to provide supervision for streaming responses and temporal grounding labels. Additionally, to enhance the model's ability to handle questions that require simple reasoning, we curate a set of 50K implicit QA pairs with time labels, guiding the model to locate relevant cues within the video stream." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.474, + 0.905, + 0.549 + ], + "angle": 0, + "content": "For LLM reasoning, we construct a dataset of 122K streaming video QA pairs based on timestamped QA pairs from ET-Instruct [41] and integrate conventional video QA data from VideoChatGPT [43] and LLaVA-Next-Video [29]." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.561, + 0.649, + 0.575 + ], + "angle": 0, + "content": "4.2. Benchmarks" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.583, + 0.905, + 0.689 + ], + "angle": 0, + "content": "For evaluation, we adopt several established long-video QA benchmarks, including EgoSchema [44], VideoMME [20], and MLVU [71]. EgoSchema consists of over 5K videos, each approximately 3 minutes long, while VideoMME and MLVU include videos of varying lengths, from a few minutes to several hours. We report accuracy on multiple-choice questions for comparison across these benchmarks." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.69, + 0.907, + 0.901 + ], + "angle": 0, + "content": "In addition, we create a streaming video QA benchmark using a subset of ETBench to measure the model's proactive response capability in real-time video interactions. Specifically, we select six subtasks from ETBench suitable for streaming response evaluation: step localization (SLC), dense video captioning (DVC), temporal action localization (TAL), temporal video grounding (TVG), episodic memory (EPM), and video highlight detection (VHD), encompassing a total of 4,460 videos. For this benchmark, the instruction is provided at the beginning of each video, and the model is required to produce the correct responses at appropriate timestamps as the video plays. We follow the evaluation metrics specified in the original benchmarks to ensure a fair comparison." + }, + { + "type": "page_number", + "bbox": [ + 0.48, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "24049" + } + ], + [ + { + "type": "table", + "bbox": [ + 0.2, + 0.089, + 0.797, + 0.33 + ], + "angle": 0, + "content": "
MethodLLM SizeFramesEgoSchemaMLVUVideoMME
Video-LLaVA [35]7B838.447.339.9
Chat-UniVi [28]7B64--40.6
LLaMA-VID [33]7B1 FPS38.533.2-
TimeChat [53]7B9633.030.930.2
MovieChat [55]7B204853.525.838.2
Video-LLaMA2 [9]7B1651.748.547.9
LLaVA-Next-Video [70]7B3243.9-46.6
ShareGPT4Video [6]8B16-46.439.9
VideoChat2 [31]7B1654.447.939.5
LongVA [69]7B128-56.352.6
Kangaroo [40]8B64-61.056.0
Video-CCAM [19]14B96-63.153.2
VideoXL [54]7B128-64.955.5
Ours7B1 FPS55.661.757.2
" + }, + { + "type": "table_caption", + "bbox": [ + 0.09, + 0.342, + 0.907, + 0.385 + ], + "angle": 0, + "content": "Table 1. Comparison on long video benchmarks. We report the accuracy on the EgoSchema full set, MLVU multiple-choice questions, and the VideoMME overall set without subtitles. For a fair comparison, we also present the model size of the LLM and the number of sampled frames." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.398, + 0.308, + 0.414 + ], + "angle": 0, + "content": "4.3. Implementation Details" + }, + { + "type": "text", + "bbox": [ + 0.089, + 0.422, + 0.483, + 0.649 + ], + "angle": 0, + "content": "The architecture consists of a vision encoder, a compact LLM as the proactive streaming video processor for response decisions, and a larger LLM for the precision interaction module. In practice, the input video frames are resized to \\(224 \\times 224\\), and a CLIP-L/14 [51] is employed to extract frame-wise features. Following the token compression techniques in [48], we concatenate adjacent tokens and use the compact LLM, instantiated as Qwen2-1.5B [49], to produce time-aware compressed clip-wise features along with clip indicators. Next, we reuse this compact LLM to process the sequence consisting of global memory, question texts, and clip features for response decisions. The final LLM, instantiated as Qwen2-7B [49], receives the grounded clips and global memory to generate responses at the necessary timestamps." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.65, + 0.483, + 0.803 + ], + "angle": 0, + "content": "Training is conducted in two stages. In the first stage, we train the response decision module and the video encoder. In the second stage, we freeze these components and train the final LLM. During training, we insert instructions at arbitrary timestamps within the video to enhance the model's ability to handle varied circumstances. For inference, on conventional video benchmarks, we place the question at the end of the video to ensure fair comparisons with existing works. For streaming evaluation, the question is posed at the beginning to enable proactive responses." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.817, + 0.398, + 0.833 + ], + "angle": 0, + "content": "4.4. Conventional Video Understanding" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.84, + 0.483, + 0.902 + ], + "angle": 0, + "content": "In this section, we compare our model with existing video LLMs on conventional video QA benchmarks, where the model is required to provide one answer after watching the complete video." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.399, + 0.907, + 0.552 + ], + "angle": 0, + "content": "We first present a comparative analysis across three challenging long-video benchmarks. As shown in Table 1, we report the accuracy on the EgoSchema full set, MLVU multiple-choice questions, and the VideoMME overall set without subtitles. Generally, our disentangled architecture handles this conventional scenario well and achieves competitive performance. Notably, on EgoSchema, which requires long temporal reasoning, our method achieves a leading performance of 55.6, demonstrating strong temporal perception." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.56, + 0.909, + 0.667 + ], + "angle": 0, + "content": "In terms of MLVU and VideoMME, which consist of videos ranging from minutes to hours long, our model's promising results highlight its ability to efficiently process contextual information across varying temporal lengths. This capability is crucial for streaming video interactions, as it allows the model to accurately summarize context when questions are inserted at arbitrary temporal positions." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.675, + 0.909, + 0.903 + ], + "angle": 0, + "content": "Additionally, we adopt a recent time-aware benchmark, ETBench, to evaluate temporal awareness. We select six subtasks that require the model to predict explicit event timestamps and report the results in Table 2. Following [41], we report the F1 score for temporal grounding evaluation and the sentence similarity score for caption evaluation. In the conventional setting, our method is able to capture the timestamps corresponding to specific questions. Without a specialized design for time tokens, our model achieves the highest F1 score in temporal video grounding and episodic memory subtasks, even surpassing models with specialized temporal tokens. The promising performance on dense video captioning and step localization further demonstrates the model's ability in both accurate temporal grounding and precise visual perception." + }, + { + "type": "page_number", + "bbox": [ + 0.48, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "24050" + } + ], + [ + { + "type": "table", + "bbox": [ + 0.123, + 0.088, + 0.878, + 0.375 + ], + "angle": 0, + "content": "
MethodFramesTVGF1EPMF1TALF1VHDF1DVCF1DVCsimSLCF1SLCsim
Conventional video QA inference.
w/ specialized time tokens
VTG-LLM [24]9615.93.714.448.240.218.620.814.4
LITA [26]10022.24.618.023.939.717.221.012.2
ETChat [41]1 FPS38.610.230.862.538.419.724.414.6
w/o specialized time tokens
VideoChatGPT [43]1007.01.315.128.88.811.35.710.2
Video-LLaVA [35]87.01.915.028.928.015.00.98.3
LLaMA-VID [33]1 FPS5.51.28.030.027.112.65.211.1
Video-LLaMA2 [9]80.10.00.01,50.614.50.015.2
PLLAVA [60]166.91.15.728.913.310.69.711.8
VTimeLLM [25]1007.61.918.228.912.413.18.76.4
TimeChat [53]9626.23.910.140.516.612.55.69.2
Ours1 FPS43.617.229.951.531.617.814.111.7
Streaming video QA inference.
VideoLLM-Online [5]2 FPS13.23.89.122.424.013.49.910.1
Ours1 FPS36.115.527.354.233.818.918.812.4
" + }, + { + "type": "table_caption", + "bbox": [ + 0.09, + 0.385, + 0.907, + 0.429 + ], + "angle": 0, + "content": "Table 2. Comparison on ETBench. We present the results for two different settings. In the conventional video QA setting, the model is required to answer the question after watching the entire video. In the streaming setting, the question is placed at the beginning of the video, and the model is expected to provide real-time responses. We report performance on six subtasks that are suitable for both settings." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.44, + 0.377, + 0.456 + ], + "angle": 0, + "content": "4.5. Streaming Video Understanding" + }, + { + "type": "text", + "bbox": [ + 0.089, + 0.463, + 0.483, + 0.688 + ], + "angle": 0, + "content": "In addition to the conventional video QA evaluation, we also assess the performance of our model in streaming video interactions. Specifically, we place the question at the beginning of the video and expect the model to produce a response once it detects relevant clues while the video stream plays in real-time. The model should remain silent when no new information is available. In the bottom lines of Table 2, we compare our model with VideoLLM-online [5], the only existing video LLM capable of streaming video interactions. Our model consistently outperforms VideoLLM-online across all metrics, with significant advantages in temporal grounding. This demonstrates that the disentangled design makes our model have far more powerful temporal awareness, enabling more agile proactive responses aligned with specific instructions." + }, + { + "type": "text", + "bbox": [ + 0.089, + 0.689, + 0.483, + 0.825 + ], + "angle": 0, + "content": "Notably, on dense video captioning and step localization tasks, our model achieves both more precise temporal grounding and more comprehensive descriptions in streaming mode than the conventional setting. This indicates that our disentangled architecture can effectively monitor the video stream in real-time and proactively generate informative responses according to the instructions. And the streaming interaction has the potential to achieve more powerful video understanding." + }, + { + "type": "text", + "bbox": [ + 0.089, + 0.826, + 0.484, + 0.902 + ], + "angle": 0, + "content": "Additionally, we present a quantitative comparison between our model and VideoLLM-online in Fig. 3. From this comparison, when faced with questions requiring multi-step reasoning, our model can gradually identify the necessary clues from the video stream and generate informative" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.441, + 0.907, + 0.531 + ], + "angle": 0, + "content": "answers step by step, while VideoLLM-online [5] fails to do so. For example, from the 'thirsty' in the question, our model can associate it with the drinks appearing in the video stream, and infer what actions are needed based on the context. In contrast, VideoLLM-online only gives simple descriptions of the scene or ongoing actions." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.532, + 0.907, + 0.623 + ], + "angle": 0, + "content": "A key advantage of our method is the disentangled perception, decision, and reaction architecture, which enables the model to simultaneously process the streaming video input and generate responses in a non-blocking manner. However, VideoLLM-online has to experience interruptions in the video stream during answer generation." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.634, + 0.667, + 0.65 + ], + "angle": 0, + "content": "4.6. Ablation Study" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.656, + 0.907, + 0.838 + ], + "angle": 0, + "content": "Clip Segmentation Strategy. We first present an ablation study on clip segmentation. We compare the standard uniform clip segmentation (16 frames per clip sampled at 1 FPS) with our scene-based non-uniform segmentation. We report conventional QA accuracy on MLVU and VideoMME, as well as streaming metrics on temporal video grounding and dense video captioning on ETBench in Table 3. The scene-based segmentation yields superior results in both conventional and streaming settings. This aligns with our motivation that scene-based segmentation preserves more structural video information, which facilitates better model learning and more timely responses." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.84, + 0.909, + 0.903 + ], + "angle": 0, + "content": "Special Token Design. We also ablate three special tokens in our architecture, i.e., \\(\\langle \\mathrm{ANS} \\rangle\\) and \\(\\langle \\mathrm{TODO} \\rangle\\) in the streaming video processor, \\(\\langle \\mathrm{SILENT} \\rangle\\) token in the final LLM. These three special tokens mainly impact the stream-" + }, + { + "type": "page_number", + "bbox": [ + 0.479, + 0.945, + 0.518, + 0.957 + ], + "angle": 0, + "content": "24051" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.117, + 0.088, + 0.885, + 0.313 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.09, + 0.329, + 0.908, + 0.358 + ], + "angle": 0, + "content": "Figure 3. Quantitative comparison between Dispider and VideoLLM-online in streaming video understanding. The question is inserted at the beginning, and we show the model's response in either answer texts or the state of keeping silent." + }, + { + "type": "table", + "bbox": [ + 0.093, + 0.375, + 0.486, + 0.434 + ], + "angle": 0, + "content": "
ClipMLVUV-MMETVG\\( _{F1} \\)DVC\\( _{F1} \\)DVC\\( _{Sim} \\)
Uniform59.855.434.533.118.1
Scene-based61.757.236.133.818.9
" + }, + { + "type": "table_caption", + "bbox": [ + 0.09, + 0.444, + 0.483, + 0.486 + ], + "angle": 0, + "content": "Table 3. Ablation study on the clip segmentation. We compare uniform 16-frame clip segmentation and our scene-based segmentation with SigLip." + }, + { + "type": "table", + "bbox": [ + 0.095, + 0.494, + 0.48, + 0.607 + ], + "angle": 0, + "content": "
<ANS><TODO><SILENT>TVG\\( _{F1} \\)DVC\\( _{F1} \\)DVC\\( _{Sim} \\)
XXX20.119.712.3
XX✓26.324.913.1
✓X✓35.231.017.2
X✓✓28.725.614.5
✓✓X35.530.216.8
✓✓✓36.133.818.9
" + }, + { + "type": "table_caption", + "bbox": [ + 0.09, + 0.618, + 0.483, + 0.66 + ], + "angle": 0, + "content": "Table 4. Ablation study on the special token designs. We respectively explore the effectiveness of ⟨ANS⟩ and ⟨TODO⟩ in streaming video processor as well as ⟨SILENT⟩ in the final LLM." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.675, + 0.484, + 0.901 + ], + "angle": 0, + "content": "ing video interactions, so we report the streaming evaluation metrics on temporal video grounding and dense video captioning in Table 4. There are three observations. First, the absence of \\(\\langle \\mathrm{ANS}\\rangle\\) token can lead the model to be unaware of the timestamps at which a response has been given. As a result, if relevant clues have appeared in the video contexts, the model tends to produce a response, resulting in a high recall but a low accuracy. Second, the \\(\\langle \\mathrm{TODO}\\rangle\\) serves as an indicator that reminds the streaming processor to decide whether to respond. The performance slightly degrades without this special token. Third, the \\(\\langle \\mathrm{SILENT}\\rangle\\) token in the final LLM serves as a secondary filter for response decision. If the preceding streaming processor incorrectly identifies a timestamp as requiring a response, the \\(\\langle \\mathrm{SILENT}\\rangle\\) token enables the LLM to rethink whether an answer is needed" + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.378, + 0.845, + 0.394 + ], + "angle": 0, + "content": "referring to the historical QA interaction contexts." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.407, + 0.634, + 0.423 + ], + "angle": 0, + "content": "5. Conclusion" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.432, + 0.907, + 0.75 + ], + "angle": 0, + "content": "In this work, we introduced Dispider, a novel framework designed to enable active real-time interaction with video LLMs. By disentangling the key capabilities of perception, decision, and reaction, and adopting an asynchronous processing approach, Dispider overcomes the inherent conflicts that hinder real-time interaction in traditional video LLMs. Our approach ensures that the model can continuously process the video stream while providing timely, contextually accurate, and precise responses to user interactions. The Proactive Streaming Video Processing module optimizes the video analysis by focusing on relevant content, while the Precise Interaction module generates detailed responses asynchronously, allowing for uninterrupted video processing. We demonstrated the effectiveness of Dispider through extensive experiments on both conventional and streaming video QA benchmarks, where it outperformed existing methods in proactive response capabilities, temporal awareness, and computational efficiency. Dispider's disentangled architecture and ability to handle long-duration video streams make it an ideal solution for real-time interactive applications." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.763, + 0.673, + 0.78 + ], + "angle": 0, + "content": "Acknowledgement" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.788, + 0.907, + 0.893 + ], + "angle": 0, + "content": "This project is funded in part by Shanghai Artificial Intelligence Laboratory, Shanghai Innovation Institute, the National Key R&D Program of China (2022ZD0160201), the Centre for Perceptual and Interactive Intelligence (CPII) Ltd under the Innovation and Technology Commission (ITC)'s InnoHK. Dahua Lin is a PI of CPII under the InnoHK." + }, + { + "type": "page_number", + "bbox": [ + 0.48, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "24052" + } + ], + [ + { + "type": "title", + "bbox": [ + 0.093, + 0.09, + 0.188, + 0.106 + ], + "angle": 0, + "content": "References" + }, + { + "type": "ref_text", + "bbox": [ + 0.101, + 0.115, + 0.484, + 0.197 + ], + "angle": 0, + "content": "[1] Jean-Baptiste Alayrac, Jeff Donahue, Pauline Luc, Antoine Miech, Iain Barr, Yana Hasson, Karel Lenc, Arthur Mensch, Katherine Millican, Malcolm Reynolds, et al. Flamingo: a visual language model for few-shot learning. Advances in neural information processing systems, 35:23716-23736, 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.101, + 0.2, + 0.484, + 0.255 + ], + "angle": 0, + "content": "[2] Jinze Bai, Shuai Bai, Shusheng Yang, Shijie Wang, Sinan Tan, Peng Wang, Junyang Lin, Chang Zhou, and Jingren Zhou. Qwen-vl: A frontier large vision-language model with versatile abilities. arXiv preprint arXiv:2308.12966, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.257, + 0.483, + 0.325 + ], + "angle": 0, + "content": "[3] Yuntao Bai, Andy Jones, Kamal Ndousse, Amanda Askell, Anna Chen, Nova DasSarma, Dawn Drain, Stanislav Fort, Deep Ganguli, Tom Henighan, et al. Training a helpful and harmless assistant with reinforcement learning from human feedback. arXiv preprint arXiv:2204.05862, 2022." + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.327, + 0.483, + 0.38 + ], + "angle": 0, + "content": "[4] Zheng Cai, Maosong Cao, Haojiong Chen, Kai Chen, Keyu Chen, Xin Chen, Xun Chen, Zehui Chen, Zhi Chen, Pei Chu, et al. Internl m2 technical report. arXiv preprint arXiv:2403.17297, 2024. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.382, + 0.482, + 0.45 + ], + "angle": 0, + "content": "[5] Joya Chen, Zhaoyang Lv, Shiwei Wu, Kevin Qinghong Lin, Chenan Song, Difei Gao, Jia-Wei Liu, Ziteng Gao, Dongxing Mao, and Mike Zheng Shou. Videollm-online: Online video large language model for streaming video. In CVPR, 2024. 1, 2, 3, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.453, + 0.483, + 0.521 + ], + "angle": 0, + "content": "[6] Lin Chen, Xilin Wei, Jinsong Li, Xiaoyi Dong, Pan Zhang, Yuhang Zang, Zehui Chen, Haodong Duan, Bin Lin, Zhenyu Tang, et al. Sharegpt4video: Improving video understanding and generation with better captions. arXiv preprint arXiv:2406.04325, 2024. 2, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.523, + 0.482, + 0.577 + ], + "angle": 0, + "content": "[7] Ling-Hao Chen, Shunlin Lu, Ailing Zeng, Hao Zhang, Benyou Wang, Ruimao Zhang, and Lei Zhang. Motionllm: Understanding human behaviors from human motions and videos. arXiv preprint arXiv:2405.20340, 2024." + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.58, + 0.482, + 0.661 + ], + "angle": 0, + "content": "[8] Zhe Chen, Jiannan Wu, Wenhai Wang, Weijie Su, Guo Chen, Sen Xing, Muyan Zhong, Qinglong Zhang, Xizhou Zhu, Lewei Lu, Bin Li, Ping Luo, Tong Lu, Yu Qiao, and Jifeng Dai. Internvl: Scaling up vision foundation models and aligning for generic visual-linguistic tasks. arXiv preprint arXiv:2312.14238, 2023." + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.664, + 0.482, + 0.732 + ], + "angle": 0, + "content": "[9] Zesen Cheng, Sicong Leng, Hang Zhang, Yifei Xin, Xin Li, Guanzheng Chen, Yongxin Zhu, Wenqi Zhang, Ziyang Luo, Deli Zhao, et al. Videollama 2: Advancing spatial-temporal modeling and audio understanding in video-llms. arXiv preprint arXiv:2406.07476, 2024. 2, 6, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.734, + 0.482, + 0.773 + ], + "angle": 0, + "content": "[10] Aakanksha Chowdhery et al. Palm: Scaling language modeling with pathways. arXiv preprint arXiv:2303.12345, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.776, + 0.482, + 0.843 + ], + "angle": 0, + "content": "[11] Wenliang Dai, Junnan Li, Dongxu Li, Anthony Meng Huat Tiong, Junqi Zhao, Weisheng Wang, Boyang Li, Pascale Fung, and Steven Hoi. Instructclip: Towards general-purpose vision-language models with instruction tuning, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.846, + 0.482, + 0.9 + ], + "angle": 0, + "content": "[12] Jacob Devlin, Ming-Wei Chang, Kenton Lee, and Kristina Toutanova. Bert: Pre-training of deep bidirectional transformers for language understanding. arXiv preprint arXiv:1810.04805, 2018. 2" + }, + { + "type": "list", + "bbox": [ + 0.094, + 0.115, + 0.484, + 0.9 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.093, + 0.905, + 0.147 + ], + "angle": 0, + "content": "[13] Shangzhe Di and Weidi Xie. Grounded question-answering in long egocentric videos. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 12934–12943, 2024. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.151, + 0.905, + 0.233 + ], + "angle": 0, + "content": "[14] Shuangrui Ding, Maomao Li, Tianyu Yang, Rui Qian, Haohang Xu, Qingyi Chen, Jue Wang, and Hongkai Xiong. Motion-aware contrastive video representation learning via foreground-background merging. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 9716-9726, 2022. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.235, + 0.905, + 0.303 + ], + "angle": 0, + "content": "[15] Shuangrui Ding, Peisen Zhao, Xiaopeng Zhang, Rui Qian, Hongkai Xiong, and Qi Tian. Prune spatio-temporal tokens by semantic-aware temporal accumulation. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 16945-16956, 2023. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.306, + 0.905, + 0.373 + ], + "angle": 0, + "content": "[16] Shuangrui Ding, Zihan Liu, Xiaoyi Dong, Pan Zhang, Rui Qian, Conghui He, Dahua Lin, and Jiaqi Wang. Songcomposer: A large language model for lyric and melody composition in song generation. arXiv preprint arXiv:2402.17645, 2024. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.376, + 0.905, + 0.445 + ], + "angle": 0, + "content": "[17] Shuangrui Ding, Rui Qian, Haohang Xu, Dahua Lin, and Hongkai Xiong. Betrayed by attention: A simple yet effective approach for self-supervised video object segmentation. In European Conference on Computer Vision, pages 215-233. Springer, 2024. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.447, + 0.905, + 0.501 + ], + "angle": 0, + "content": "[18] Yue Fan, Xiaojian Ma, Rujie Wu, Yuntao Du, Jiaqi Li, Zhi Gao, and Qing Li. Videoagent: A memory-augmented multimodal agent for video understanding. arXiv preprint arXiv:2403.11481, 2024. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.504, + 0.905, + 0.559 + ], + "angle": 0, + "content": "[19] Jiajun Fei, Dian Li, Zhidong Deng, Zekun Wang, Gang Liu, and Hui Wang. Video-ccam: Enhancing video-language understanding with causal cross-attention masks for short and long videos. arXiv preprint arXiv:2408.14023, 2024.6" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.562, + 0.905, + 0.63 + ], + "angle": 0, + "content": "[20] Chaoyou Fu, Yuhan Dai, Yondong Luo, Lei Li, Shuhuai Ren, Renrui Zhang, Zihan Wang, Chenyu Zhou, Yunhang Shen, Mengdan Zhang, et al. Video-mme: The first-ever comprehensive evaluation benchmark of multi-modal llms in video analysis. arXiv preprint arXiv:2405.21075, 2024. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.633, + 0.905, + 0.701 + ], + "angle": 0, + "content": "[21] Chaoyou Fu, Haojia Lin, Zuwei Long, Yunhang Shen, Meng Zhao, Yifan Zhang, Xiong Wang, Di Yin, Long Ma, Xiawu Zheng, Ran He, Rongrong Ji, Yunsheng Wu, Caifeng Shan, and Xing Sun. Vita: Towards open-source interactive omni multimodal lIm. arXiv preprint arXiv:2408.05211, 2024. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.703, + 0.905, + 0.742 + ], + "angle": 0, + "content": "[22] Gemini Team Google. Gemini: a family of highly capable multimodal models. arXiv preprint arXiv:2312.11805, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.746, + 0.905, + 0.787 + ], + "angle": 0, + "content": "[23] Gemini Team Google. Gemini 1.5: Unlocking multimodal understanding across millions of tokens of context. arXiv preprint arXiv:2403.05530, 2024. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.789, + 0.905, + 0.844 + ], + "angle": 0, + "content": "[24] Yongxin Guo, Jingyu Liu, Mingda Li, Xiaoying Tang, Xi Chen, and Bo Zhao. Vtg-llm: Integrating timestamp knowledge into video llms for enhanced video temporal grounding. arXiv preprint arXiv:2405.13382, 2024. 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.846, + 0.905, + 0.9 + ], + "angle": 0, + "content": "[25] Bin Huang, Xin Wang, Hong Chen, Zihan Song, and Wenwu Zhu. Vtimellm: Empower llm to grasp video moments. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 14271-14280, 2024. 7" + }, + { + "type": "list", + "bbox": [ + 0.517, + 0.093, + 0.905, + 0.9 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.48, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "24053" + } + ], + [ + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.092, + 0.482, + 0.161 + ], + "angle": 0, + "content": "[26] De-An Huang, Shijia Liao, Subhashree Radhakrishnan, Hongxu Yin, Pavlo Molchanov, Zhiding Yu, and Jan Kautz. Lita: Language instructed temporal-localization assistant. In European Conference on Computer Vision, pages 202-218. Springer, 2025. 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.162, + 0.482, + 0.232 + ], + "angle": 0, + "content": "[27] Albert Q Jiang, Alexandre Sablayrolles, Arthur Mensch, Chris Bamford, Devendra Singh Chaplot, Diego de las Casas, Florian Bressand, Gianna Lengyel, Guillaume Lample, Lucile Saulnier, et al. Mistral 7b. arXiv preprint arXiv:2310.06825, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.233, + 0.482, + 0.287 + ], + "angle": 0, + "content": "[28] Peng Jin, Ryuichi Takanobu, Caiwan Zhang, Xiaochun Cao, and Li Yuan. Chat-univi: Unified visual representation empowers large language models with image and video understanding. arXiv preprint arXiv:2311.08046, 2023. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.288, + 0.482, + 0.343 + ], + "angle": 0, + "content": "[29] Bo Li, Yuanhan Zhang, Dong Guo, Renrui Zhang, Feng Li, Hao Zhang, Kaichen Zhang, Yanwei Li, Ziwei Liu, and Chunyuan Li. Llava-onevision: Easy visual task transfer. arXiv preprint arXiv:2408.03326, 2024. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.344, + 0.482, + 0.398 + ], + "angle": 0, + "content": "[30] KunChang Li, Yinan He, Yi Wang, Yizhuo Li, Wenhai Wang, Ping Luo, Yali Wang, Limin Wang, and Yu Qiao. Videochat: Chat-centric video understanding. arXiv preprint arXiv:2305.06355, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.4, + 0.482, + 0.481 + ], + "angle": 0, + "content": "[31] Kunchang Li, Yali Wang, Yinan He, Yizhuo Li, Yi Wang, Yi Liu, Zun Wang, Jilan Xu, Guo Chen, Ping Luo, et al. Mvbench: A comprehensive multi-modal video understanding benchmark. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 22195-22206, 2024. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.483, + 0.482, + 0.524 + ], + "angle": 0, + "content": "[32] Yanwei Li, Chengyao Wang, and Jiaya Jia. Llama-vid: An image is worth 2 tokens in large language models. arXiv preprint arXiv:2311.17043, 2023. 1, 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.525, + 0.482, + 0.564 + ], + "angle": 0, + "content": "[33] Yanwei Li, Chengyao Wang, and Jiaya Jia. Llama-vid: An image is worth 2 tokens in large language models. 2024. 2, 6, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.567, + 0.482, + 0.635 + ], + "angle": 0, + "content": "[34] Yifei Li, Junbo Niu, Ziyang Miao, Chunjiang Ge, Yuanhang Zhou, Qihao He, Xiaoyi Dong, Haodong Duan, Shuangrui Ding, Rui Qian, et al. Ovo-bench: How far is your videolms from real-world online video understanding? arXiv preprint arXiv:2501.05510, 2025. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.637, + 0.482, + 0.69 + ], + "angle": 0, + "content": "[35] Bin Lin, Bin Zhu, Yang Ye, Munan Ning, Peng Jin, and Li Yuan. Video-llava: Learning united visual representation by alignment before projection. arXiv preprint arXiv:2311.10122, 2023. 6, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.692, + 0.482, + 0.761 + ], + "angle": 0, + "content": "[36] Kevin Lin, Faisal Ahmed, Linjie Li, Chung-Ching Lin, Ehsan Azarnasab, Zhengyuan Yang, Jianfeng Wang, Lin Liang, Zicheng Liu, Yumao Lu, et al. Mm-vid: Advancing video understanding with gpt-4v (ision). arXiv preprint arXiv:2310.19773, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.762, + 0.482, + 0.803 + ], + "angle": 0, + "content": "[37] Haotian Liu, Chunyuan Li, Yuheng Li, and Yong Jae Lee. Improved baselines with visual instruction tuning. arXiv preprint arXiv:2310.03744, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.804, + 0.482, + 0.831 + ], + "angle": 0, + "content": "[38] Haotian Liu, Chunyuan Li, Qingyang Wu, and Yong Jae Lee. Visual instruction tuning, 2023." + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.832, + 0.482, + 0.873 + ], + "angle": 0, + "content": "[39] Haotian Liu, Chunyuan Li, Yuheng Li, Bo Li, Yuanhan Zhang, Sheng Shen, and Yong Jae Lee. Llava next: Improved reasoning,OCR, and world knowledge, 2024. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.874, + 0.482, + 0.901 + ], + "angle": 0, + "content": "[40] Jiajun Liu, Yibing Wang, Hanghang Ma, Xiaoping Wu, Xiaqi Ma, Xiaoming Wei, Jianbin Jiao, Enhua Wu, and Jie Hu." + }, + { + "type": "list", + "bbox": [ + 0.093, + 0.092, + 0.482, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.545, + 0.093, + 0.905, + 0.133 + ], + "angle": 0, + "content": "Kangaroo: A powerful video-language model supporting long-context video input. arXiv preprint arXiv:2408.15542, 2024.6" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.136, + 0.905, + 0.19 + ], + "angle": 0, + "content": "[41] Ye Liu, Zongyang Ma, Zhongang Qi, Yang Wu, Chang Wen Chen, and Ying Shan. E.t. bench: Towards open-ended event-level video-language understanding. In Neural Information Processing Systems (NeurIPS), 2024. 5, 6, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.193, + 0.905, + 0.262 + ], + "angle": 0, + "content": "[42] Xudong Lu, Yinghao Chen, Cheng Chen, Hui Tan, Boheng Chen, Yina Xie, Rui Hu, Guanxin Tan, Renshou Wu, Yan Hu, et al. Bluem-v-3b: Algorithm and system co-design for multimodal large language models on mobile devices. arXiv preprint arXiv:2411.10640, 2024. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.263, + 0.905, + 0.318 + ], + "angle": 0, + "content": "[43] Muhammad Maaz, Hanoona Rasheed, Salman Khan, and Fahad Shahbaz Khan. Video-chatgpt: Towards detailed video understanding via large vision and language models. arXiv preprint arXiv:2306.05424, 2023. 2, 5, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.32, + 0.905, + 0.375 + ], + "angle": 0, + "content": "[44] Karttikeya Mangalam, Raiymbek Akshulakov, and Jitendra Malik. Egoschema: A diagnostic benchmark for very long-form video language understanding. Advances in Neural Information Processing Systems, 36, 2024. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.377, + 0.792, + 0.39 + ], + "angle": 0, + "content": "[45] OpenAI. Gpt-4 technical report. 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.392, + 0.905, + 0.473 + ], + "angle": 0, + "content": "[46] Long Ouyang, Jeffrey Wu, Xu Jiang, Diogo Almeida, Carroll Wainwright, Pamela Mishkin, Chong Zhang, Sandhini Agarwal, Katarina Slama, Alex Ray, et al. Training language models to follow instructions with human feedback. Advances in neural information processing systems, 35:27730-27744, 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.476, + 0.905, + 0.532 + ], + "angle": 0, + "content": "[47] Rui Qian, Shuangrui Ding, and Dahua Lin. Rethinking image-to-video adaptation: An object-centric perspective. In European Conference on Computer Vision, pages 329-348. Springer, 2024. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.533, + 0.905, + 0.602 + ], + "angle": 0, + "content": "[48] Rui Qian, Xiaoyi Dong, Pan Zhang, Yuhang Zang, Shuangrui Ding, Dahua Lin, and Jiaqi Wang. Streaming long video understanding with large language models. In The Thirty-eighth Annual Conference on Neural Information Processing Systems, 2024. 3, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.604, + 0.821, + 0.617 + ], + "angle": 0, + "content": "[49] Team Qwen. Qwen2 technical report, 2024. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.619, + 0.905, + 0.66 + ], + "angle": 0, + "content": "[50] Alec Radford, Karthik Narasimhan, Tim Salimans, and Ilya Sutskever. Improving language understanding by generative pre-training. 2018. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.661, + 0.905, + 0.743 + ], + "angle": 0, + "content": "[51] Alec Radford, Jong Wook Kim, Chris Hallacy, Aditya Ramesh, Gabriel Goh, Sandhini Agarwal, Girish Sastry, Amanda Askell, Pamela Mishkin, Jack Clark, et al. Learning transferable visual models from natural language supervision. In International conference on machine learning, pages 8748-8763, 2021. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.746, + 0.905, + 0.801 + ], + "angle": 0, + "content": "[52] Colin Raffel, Noam Shazeer, Adam Roberts, Katherine Lee, Sharan Narang, Michael Matena, Yanqi Zhou, Wei Li, and Peter J Liu. Exploring the limits of transfer learning with a unified text-to-text transformer. JMLR, 2020. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.802, + 0.905, + 0.871 + ], + "angle": 0, + "content": "[53] Shuhuai Ren, Linli Yao, Shicheng Li, Xu Sun, and Lu Hou. Timechat: A time-sensitive multimodal large language model for long video understanding. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 14313-14323, 2024. 2, 6, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.873, + 0.905, + 0.901 + ], + "angle": 0, + "content": "[54] Yan Shu, Peitian Zhang, Zheng Liu, Minghao Qin, Junjie Zhou, Tiejun Huang, and Bo Zhao. Video-xl: Extra-long" + }, + { + "type": "list", + "bbox": [ + 0.517, + 0.093, + 0.905, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.48, + 0.945, + 0.52, + 0.957 + ], + "angle": 0, + "content": "24054" + } + ], + [ + { + "type": "ref_text", + "bbox": [ + 0.125, + 0.093, + 0.482, + 0.12 + ], + "angle": 0, + "content": "vision language model for hour-scale video understanding. arXiv preprint arXiv:2409.14485, 2024. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.122, + 0.483, + 0.205 + ], + "angle": 0, + "content": "[55] Enxin Song, Wenhao Chai, Guanhong Wang, Yucheng Zhang, Haoyang Zhou, Feiyang Wu, Haozhe Chi, Xun Guo, Tian Ye, Yanting Zhang, et al. Moviechat: From dense token to sparse memory for long video understanding. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 18221-18232, 2024. 2, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.205, + 0.482, + 0.273 + ], + "angle": 0, + "content": "[56] Hugo Touvron, Thibaut Lavril, Gautier Izacard, Xavier Martinet, Marie-Anne Lachaux, Timothée Lacroix, Baptiste Rozière, Naman Goyal, Eric Hambro, Faisal Azhar, et al. Llama: Open and efficient foundation language models. arXiv preprint arXiv:2302.13971, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.274, + 0.482, + 0.342 + ], + "angle": 0, + "content": "[57] Ziyang Wang, Shoubin Yu, Elias Stengel-Eskin, Jaehong Yoon, Feng Cheng, Gedas Bertasius, and Mohit Bansal. Videotree: Adaptive tree-based video representation for llm reasoning on long videos. arXiv preprint arXiv:2405.19209, 2024. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.344, + 0.482, + 0.398 + ], + "angle": 0, + "content": "[58] Qi Wu, Yubo Zhao, Yifan Wang, Yu-Wing Tai, and Chikeung Tang. Motionllm: Multimodal motion-language learning with large language models. arXiv preprint arXiv:2405.17013, 2024." + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.4, + 0.482, + 0.482 + ], + "angle": 0, + "content": "[59] Haiyang Xu, Qinghao Ye, Mingshi Yan, Yaya Shi, Jiabo Ye, Yuanhong Xu, Chenliang Li, Bin Bi, Qiuchen Qian, Wei Wang, Guohai Xu, Ji Zhang, Songfang Huang, Feiran Huang, and Jingren Zhou. mplug-2: A modularized multimodal foundation model across text, image and video. arXiv preprint arXiv:2302.00402, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.484, + 0.482, + 0.537 + ], + "angle": 0, + "content": "[60] Lin Xu, Yilin Zhao, Daquan Zhou, Zhijie Lin, See Kiong Ng, and Jiashi Feng. Pllava: Parameter-free llava extension from images to videos for video dense captioning. arXiv preprint arXiv:2404.16994, 2024. 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.539, + 0.482, + 0.621 + ], + "angle": 0, + "content": "[61] Qinghao Ye, Haiyang Xu, Jiabo Ye, Ming Yan, Anwen Hu, Haowei Liu, Qi Qian, Ji Zhang, and Fei Huang. mplug-owl2: Revolutionizing multi-modal large language model with modality collaboration. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 13040–13051, 2024. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.623, + 0.482, + 0.678 + ], + "angle": 0, + "content": "[62] Alex Young, Bei Chen, Chao Li, Chengen Huang, Ge Zhang, Guanwei Zhang, Heng Li, Jiangcheng Zhu, Jianqun Chen, Jing Chang, et al. Yi: Open foundation models by 01. ai. arXiv preprint arXiv:2403.04652, 2024. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.679, + 0.482, + 0.734 + ], + "angle": 0, + "content": "[63] Xiaohua Zhai, Basil Mustafa, Alexander Kolesnikov, and Lucas Beyer. Sigmoid loss for language image pre-training. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 11975-11986, 2023. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.735, + 0.482, + 0.789 + ], + "angle": 0, + "content": "[64] Ce Zhang, Taixi Lu, Md Mohaiminul Islam, Ziyang Wang, Shoubin Yu, Mohit Bansal, and Gedas Bertasius. A simple llm framework for long-range video question-answering. arXiv preprint arXiv:2312.17235, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.79, + 0.482, + 0.843 + ], + "angle": 0, + "content": "[65] Haoji Zhang, Yiqin Wang, Yansong Tang, Yong Liu, Jiashi Feng, Jifeng Dai, and Xiaojie Jin. Flash-vstream: Memory-based real-time understanding for long video streams. 2024. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.845, + 0.482, + 0.901 + ], + "angle": 0, + "content": "[66] Pan Zhang, Xiaoyi Dong, Bin Wang, Yuhang Cao, Chao Xu, Linke Ouyang, Zhiyuan Zhao, Haodong Duan, Songyang Zhang, Shuangrui Ding, et al. Internlmxcomposer: A vision-language large model for advanced" + }, + { + "type": "list", + "bbox": [ + 0.093, + 0.093, + 0.483, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.545, + 0.093, + 0.905, + 0.12 + ], + "angle": 0, + "content": "text-image comprehension and composition. arXiv preprint arXiv:2309.15112, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.122, + 0.905, + 0.204 + ], + "angle": 0, + "content": "[67] Pan Zhang, Xiaoyi Dong, Yuhang Cao, Yuhang Zang, Rui Qian, Xilin Wei, Lin Chen, Yifei Li, Junbo Niu, Shuangrui Ding, et al. Internlm-xcomposer2. 5-omnilive: A comprehensive multimodal system for long-term streaming video and audio interactions. arXiv preprint arXiv:2412.09596, 2024.3" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.207, + 0.905, + 0.275 + ], + "angle": 0, + "content": "[68] Pan Zhang, Xiaoyi Dong, Yuhang Zang, Yuhang Cao, Rui Qian, Lin Chen, Qipeng Guo, Haodong Duan, Bin Wang, Linke Ouyang, et al. Internlm-xcomposer-2.5: A versatile large vision language model supporting long-contextual input and output. arXiv preprint arXiv:2407.03320, 2024. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.277, + 0.905, + 0.343 + ], + "angle": 0, + "content": "[69] Peiyuan Zhang, Kaichen Zhang, Bo Li, Guangtao Zeng, Jingkang Yang, Yuanhan Zhang, Ziyue Wang, Haoran Tan, Chunyuan Li, and Ziwei Liu. Long context transfer from language to vision. arXiv preprint arXiv:2406.16852, 2024. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.346, + 0.905, + 0.4 + ], + "angle": 0, + "content": "[70] Yuanhan Zhang, Bo Li, haotian Liu, Yong jae Lee, Liangke Gui, Di Fu, Jiashi Feng, Ziwei Liu, and Chunyuan Li. Llavanext: A strong zero-shot video understanding model, 2024. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.403, + 0.905, + 0.471 + ], + "angle": 0, + "content": "[71] Junjie Zhou, Yan Shu, Bo Zhao, Boya Wu, Shitao Xiao, Xi Yang, Yongping Xiong, Bo Zhang, Tiejun Huang, and Zheng Liu. Mlvu: A comprehensive benchmark for multi-task long video understanding. arXiv preprint arXiv:2406.04264, 2024.5" + }, + { + "type": "list", + "bbox": [ + 0.517, + 0.093, + 0.905, + 0.471 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.48, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "24055" + } + ] +] \ No newline at end of file diff --git a/2025/Dispider_ Enabling Video LLMs with Active Real-Time Interaction via Disentangled Perception, Decision, and Reaction/509357f7-50b0-4a3e-be02-5d53b2c39c5b_origin.pdf b/2025/Dispider_ Enabling Video LLMs with Active Real-Time Interaction via Disentangled Perception, Decision, and Reaction/509357f7-50b0-4a3e-be02-5d53b2c39c5b_origin.pdf new file mode 100644 index 0000000000000000000000000000000000000000..6025329d9b04a15a0b24d1db5452a204a25daafb --- /dev/null +++ b/2025/Dispider_ Enabling Video LLMs with Active Real-Time Interaction via Disentangled Perception, Decision, and Reaction/509357f7-50b0-4a3e-be02-5d53b2c39c5b_origin.pdf @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:3367d615cac697a784a6c6fab2dab9a2ba9b5b0cd34bcb7b3e0c3f542db69936 +size 5089276 diff --git a/2025/Dispider_ Enabling Video LLMs with Active Real-Time Interaction via Disentangled Perception, Decision, and Reaction/full.md b/2025/Dispider_ Enabling Video LLMs with Active Real-Time Interaction via Disentangled Perception, Decision, and Reaction/full.md new file mode 100644 index 0000000000000000000000000000000000000000..219a0c2dac413b95ee871b73e482b77e5f9caff7 --- /dev/null +++ b/2025/Dispider_ Enabling Video LLMs with Active Real-Time Interaction via Disentangled Perception, Decision, and Reaction/full.md @@ -0,0 +1,335 @@ +# Dispider: Enabling Video LLMs with Active Real-Time Interaction via Disentangled Perception, Decision, and Reaction + +Rui Qian $^{1*}$ Shuangrui Ding $^{1*}$ Xiaoyi Dong $^{1,2\dagger}$ Pan Zhang $^{2}$ Yuhang Zang $^{2}$ Yuhang Cao $^{2}$ Dahua Lin $^{1,2,4}$ Jiaqi Wang $^{2,3\dagger}$ + +1 The Chinese University of Hong Kong 2 Shanghai AI Laboratory +3 Shanghai Innovation Institute 4 CPII under InnoHK + +{qr021, ds023}@ie.cuhk.edu.hk + +# Abstract + +Active real-time interaction with video LLMs introduces a new paradigm for human-computer interaction, where the model not only understands user intent but also responds while continuously processing streaming video on the fly. Unlike offline video LLMs, which analyze the entire video before answering questions, active real-time interaction requires three capabilities: 1) Perception: real-time video monitoring and interaction capturing. 2) Decision: raising proactive interaction in proper situations, 3) Reaction: continuous interaction with users. However, inherent conflicts exist among the desired capabilities. The Decision and Reaction require a contrary Perception scale and grain, and the autoregressive decoding blocks the real-time Perception and Decision during the Reaction. To unify the conflicted capabilities within a harmonious system, we present Dispider, a solution built on a Disentangled Perception, Decision, and Reaction framework. Dispider features a lightweight Proactive Streaming Video Processing module that tracks the video stream and identifies optimal moments for interaction. Once the interaction is triggered, an asynchronous Precise Interaction module provides detailed responses, while the processing module continues to monitor the video in the meantime. Our disentangled and asynchronous design ensures timely, contextually accurate, and computationally efficient responses, making Dispider ideal for active real-time interaction for long-duration video streams. Experiments prove that Dispider outperforms existing methods not only in its superior understanding of video content in conventional video QA settings, but also in proactive response capability and temporal awareness under the streaming setting. The code and model are released at https://github.com/Mark12Ding/Dispider. + +# 1. Introduction + +The recent surge in multimodal large language models (MLLMs) has brought considerable attention to video understanding tasks [14, 15, 17, 47]. Given a video, these models aim to accurately comprehend its content and provide responses aligned with human expectations. However, the majority of current video LLMs are designed around an offline setting, where models are required to view the entire video before generating a single response. While effective for certain applications, this offline approach is impractical for real-time $^{1}$ , interactive scenarios where users expect continuous feedback as video streams. + +Consider a scenario where a user performs pull-ups with the assistance of a visual model. As illustrated in Figure 1, current offline video LLMs need to process the entire video before providing the total count of pull-ups. However, users expect real-time feedback, with each pull-up counted as it occurs: "One... Two... Three..." This real-time interaction requires the model to recognize each completed action instantly and synchronize its responses with the user's movements. The challenge lies in determining the precise moment to output the count along with the ongoing video stream. To this end, traditional offline video LLMs lack the flexibility to handle such demands, as they are not designed for continuous, dynamic video streams. + +VideoLLM-online [5] has been a pioneering effort in addressing this issue. By leveraging the LLaMA model [32], they enable proactive response updates during streaming video inputs. However, it has a critical limitation: since it uses a single LLM for video processing and response generation, it cannot perform perception and answer reaction simultaneously. The autoregressive nature of next-token prediction in LLMs forces VideoLLM-online to alternate between perception (processing frames) and response generation, rather than handling both in parallel. This restriction + +![](images/255603027e52059f38adf1a94e4fabdbdb4af1fd96859265d3a78ff0e17ecf36.jpg) +Figure 1. Comparison between offline and active real-time Video LLMs in processing a video sequence of a man performing pull-ups. The user asks the model to count the number of pull-ups. The offline Video LLMs process the entire video before providing a final count of three pull-ups. In contrast, the real-time Video LLMs provide immediate feedback after each pull-up, first counting one, then two, and finally reaching three, while remaining silent between responses. + +![](images/7a6a8509236a8b0757bb9626b112faf76911731080344add1aeb9eed2bbfe84c.jpg) + +Real-time Video LLM + +![](images/c18cb3313382d5bb3999b464a74a28f65af0717b9ec441139da3a37aa9bba761.jpg) + +: The man did 1 pull-up. + +![](images/8cc534a32a9d74bf06415d6300be7c567de6a3244ffe56c129325efbdf41b17f.jpg) + +![](images/826a2d6569ce3904afa977fb792d52416a35c55c4a58f80e5c0f0d248a8fc6a6.jpg) + +: The man did 2 pull-ups. + +![](images/61f5e1be321d1ef4ad7ce90747054fdf5249166d1d91f08af0a33d88b13e4f40.jpg) + +![](images/56b5b4d7e1e9970610d47a564691468b520acff2e16c08f0452729651286867e.jpg) + +The man did 3 pull-ups. + +delays the video inputs during responding, reduces responsiveness and hinders its ability to engage in real-time interactions effectively, especially during long-duration video streams that require continuous, uninterrupted processing. + +To address these limitations, we present Dispider, a novel framework designed for active real-time interaction with streaming videos. Dispider resolves the conflicts between perception, decision, and reaction through a disentangled, asynchronous design, where each component can operate in parallel. Specifically, we implement these functions in the Proactive Streaming Video Processing module and the Precise Interaction module. In the video processing module, with the streaming video input, we dynamically segment the video stream into non-uniform clips based on scene boundaries. These clips, combined with relevant historical context, are then processed by a real-time response decision module, which determines whether the model should generate a response or wait for additional video content. When an interaction is triggered, the Precise Interaction module generates context-aware responses without interrupting the ongoing video processing. + +This disentangled design ensures that the Dispider can provide timely, accurate, and computationally efficient responses, even for long-duration video streams. We evaluate Dispider's ability in both real-time video stream interactions, and conventional offline Video LLMs across long-video benchmarks (EgoSchema, VideoMME, MLVU) and time-sensitive tasks (ETBench). Dispider demonstrated strong performance, particularly excelling in temporal reasoning on EgoSchema and handling diverse video lengths effectively. In streaming tasks, it significantly surpasses VideoLLM-online [5] in temporal grounding, proactive response generation, and multi-step reasoning. + +# 2. Related Work + +Large Language Models. LLMs have revolutionized natural language processing, achieving remarkable performance across a wide range of tasks. Early models such as BERT [12] and T5 [52] used masked language modeling for pre-training. The shift to decoder-only models like GPT [50] introduced scalable architectures that enhanced language generation. Recent models, including PaLM [10], LLaMA [56], and GPT-4 [45], continue to push the boundaries with massive parameters and extensive training data. Techniques such as supervised fine-tuning and reinforcement learning from human feedback [2-4, 16, 22, 23, 27, 32, 46, 62] have further improved these models' ability to generate coherent, contextually relevant responses. Inspired by the reasoning capabilities of LLMs, we extend these models to the domain of streaming video understanding, where real-time interaction presents unique challenges. + +Video Large Language Models. The progress of multimodal LLMs has been significant, particularly with image-based models like InstructBLIP [11], LLaVA [37-39], and Flamingo [1], which integrate vision and language models. Extending this to video introduces challenges in managing both frame sequences and the context length of LLMs. Recent works on video LLMs [6-9, 18, 21, 33, 36, 42, 43, 55, 57-59, 61, 64, 66, 68] present novel strategies for processing long-form videos while maintaining effective reasoning. Notable approaches include TimeChat [53], which emphasizes temporal relationships across video frames, and MovieChat [55], which uses sparse memory to handle long videos. VideoChat [30] adopts a chat-centric approach, integrating video foundation models with LLMs to excel at spatiotemporal reasoning, event localization, and causal inference. Building on these video LLMs, our work intro + +duces a pipeline designed to process streaming video inputs and generate real-time outputs. + +Streaming Video Understanding. In practice, users typically expect models to operate online and interactively in real time. This is known as streaming video understanding, which involves processing continuous video streams while ensuring long-term temporal consistency and enabling interactive responses. Despite its practical significance, only a few works have explored this area [34, 67]. VideoLLM-online [5] introduces the LIVE framework for streaming dialogue, but it lacks an efficient module for handling long-term video inputs over extended periods and does not prioritize real-time interactivity. VideoStream [48] proposes memory-propagated encoding for long videos, yet its focus is on offline processing rather than real-time streaming. Similarly, Flash-VStream [65] addresses inference latency and memory efficiency for long video streams but overlooks real-time user interaction. In contrast, our approach centers on developing a real-time visual assistant for streaming video, emphasizing both long-context handling and interactive, real-time responses, thereby bridging the gap left by previous methods. + +# 3. Method + +# 3.1. Problem Formulation + +In contrast to previous offline video LLMs, which generate responses only after processing the entire video, our approach requires the model to operate in real-time, providing continuous and interactive responses as the video is processed. Specifically, the model actively determines when sufficient information has been observed to provide a complete response, allowing it to produce answers promptly without waiting for the entire video to finish. Following the formulation of VideoLLM-online [5], we formulate this new setting as below. + +Given a continuous video stream $V = \{v_{1}, v_{2}, \dots, v_{T}\}$ , where $v_{i}$ represents the $i$ -th video clip, and a context sequence $C_{t}$ up to a specific time $t$ (which includes prior vision-language interactions such as historical user queries, assistant responses, and video frames), our goal is to generate timely and accurate dialogue responses $R_{t}$ without processing the entire video sequence. + +At each time $t$ , the model observes $C_t$ and the video frames up to that time $V_{[0,t]} = \{v_1, v_2, \ldots, v_t\}$ . The model must decide whether it has sufficient information to generate a response. We define a decision function $\pi$ and a response function $f$ : + +$$ +a _ {t} = \pi \left(C _ {t}, V _ {[ 0, t ]}\right) \in \{\text {w a i t , r e s p o n d} \}. +$$ + +If $a_{t} =$ respond, the model generates a response: + +$$ +R _ {t} = f \left(C _ {t}, V _ {[ 0, t ]}\right) +$$ + +Otherwise, the model keeps silent for more information. + +In this work, we implement these functions using a disentangled framework composed of three distinct modules: Perception, Decision, and Reaction. These modules are designed to handle the unique challenges of real-time video understanding and dialogue generation. The Perception module focuses on continuous video monitoring, while the Decision module assesses when to act, and the Reaction module generates responses without waiting for the entire video sequence to finish. This approach allows the system to remain responsive to new information while ensuring that the generated answers are based on the most relevant and up-to-date context. + +# 3.2. Proactive Streaming Video Processing + +To enable active real-time response decisions, we propose a proactive streaming video processing approach that dynamically segments the video stream and evaluates whether to generate a response as the video is being processed. This approach balances computational efficiency with the need for timely and accurate responses, ensuring that the model processes the most relevant parts of the video while avoiding unnecessary redundancy. The system is composed of two key modules: the Scene-based Perception Module and the Real-time Response Decision Module. + +Scene-based Perception Module. To ensure efficient processing of long video streams, we begin by adaptively segmenting the video into non-uniform clips based on scene boundaries. This segmentation strategy preserves the structural information of the video, allowing the model to focus on the most informative parts while removing redundancy and maintaining context. + +We begin by sampling frames at a regular interval (one frame per second) and extract L2-normalized feature embeddings using the pre-trained SigLip model [63]. By computing the cosine similarity between these embeddings, we can identify significant visual changes, which indicate potential scene boundaries. These boundaries help divide the video into meaningful segments. To prevent excessively short clips, we introduce an exclusion window around the boundaries, ensuring that the resulting clips are of sufficient length to contain relevant information. + +Each clip $v_{i}$ is then processed by the video encoder to produce the clip-wise feature representation $F_{i}$ , along with a special clip indicator $\hat{F}_{i}$ . These clip features are used in the subsequent decision-making process to determine if enough information has been gathered to respond. + +Real-time Response Decision Module. The Real-time Response Decision Module evaluates whether the model should generate a response or continue waiting for more video content. This decision is made based on the video content observed so far and the historical context, including the user's prior questions and assistant responses. We illus + +![](images/e900973b60bc3bc785f02a22db4251ad3402fc3ca6864d6478513fb87f0cd223.jpg) +Figure 2. Illustration of the Response Decision Module in a proactive streaming video processing system. The module dynamically determines when to respond during video streaming by segmenting the video into non-uniform clips and utilizing historical memory to capture context. The module then combines memory features, clip features, question text, and special tokens, $\langle \mathrm{TODO}\rangle$ and $\langle \mathrm{ANS}\rangle$ , to decide on response timing. + +trate this whole process in Figure 2. To effectively combine these multi-modal inputs, we use an interleaved sequence format, which integrates video features, question information, and decision tokens. + +The process begins with the sequence of video clip representations up to the point of the user's question, followed by the question $Q$ and a special token, $\langle \mathrm{TODO} \rangle$ , which indicates that a response decision is yet to be made. Formally, the input sequence at this stage is: + +$$ +F [ 0: q _ {\text {p o s}} ] \oplus Q \oplus \langle \text {T O D O} \rangle , +$$ + +where $q_{\mathrm{pos}}$ is the timestamp of the question, and $\oplus$ denotes concatenation. + +After this initial stage, the model aggregates the features up to the question timestamp into a global memory $M$ through pooling. Then, we extend the sequence to include this global memory that captures historical context, the question texts, and the clip features from the question timestamp to the current point, ending with $\langle \mathrm{TODO}\rangle$ : + +$$ +M [ 0: q _ {\mathrm {p o s}} ] \oplus Q \oplus F [ q _ {\mathrm {p o s}}: ] \oplus \langle \mathrm {T O D O} \rangle . +$$ + +When the model decides to respond, the $\langle \mathrm{ANS} \rangle$ token is used to mark the response action. Specifically, with the video continuing to stream, we further integrate global + +memory from the question timestamp to the position of the first answer, insert the special answer token $\langle \mathrm{ANS}\rangle$ , and include later clips for potential subsequent responses. This sequence concludes with the current clip memory segment and a $\langle \mathrm{TODO}\rangle$ token. Thus, the ultimate input format with multiple answers can be written as: + +$$ +M [ 0: q _ {\mathrm {p o s}} ] \oplus Q \oplus +$$ + +$$ +M \left[ q _ {\text {p o s}}: a _ {\text {p o s}} ^ {1} \right] \oplus \langle \mathrm {A N S} \rangle \oplus +$$ + +$$ +M \left[ a _ {\text {p o s}} ^ {1}: a _ {\text {p o s}} ^ {2} \right] \oplus \langle \mathrm {A N S} \rangle \oplus +$$ + +$$ +\cdot \cdot \cdot +$$ + +$$ +F \left[ a _ {\text {p o s}} ^ {k}: \right] \oplus \langle \text {T O D O} \rangle , +$$ + +where $a_{\mathrm{pos}}^k$ stands for the timestamp for the $k$ -th answer. + +We feed the interleaved sequence into a compact LLM and adopt a binary classification head on top of the final-layer embedding of the $\langle \mathrm{TODO}\rangle$ token. This head is trained to predict whether the model should respond at each timestep. We use standard binary cross-entropy loss to supervise the model's decision-making process. When the model determines that a response is required, the relevant historical clips are retrieved based on their cosine similarity with the $\langle \mathrm{TODO}\rangle$ token's embedding. + +To train the model's capacity on temporal retrieval, we + +calculate the cosine similarity between the embedding of the $\langle \mathrm{TODO}\rangle$ token and each historical clip indicator $\hat{F}_i$ . This similarity yields a probability distribution $\hat{P}_{\mathrm{pred}}(i)$ representing the relevance of $i$ -th historical clip to the current response. + +For the ground truth, we use related clips to create a true temporal distribution. These clips are encoded as a binary mask over the clip intervals relevant to the response, defined as: + +$$ +\hat {P} _ {\text {t r u e}} (i) = \frac {1}{| R _ {Q} |} \quad \text {i f} \quad i \in R _ {Q} \quad \text {e l s e} \quad 0, +$$ + +where $R_{Q}$ is the ground-truth set of relevant clips for particular question $Q$ . + +We then calculate the KL divergence loss between the true distribution, $\hat{P}_{\mathrm{true}}$ , and the predicted distribution, $\hat{P}_{\mathrm{pred}}$ , as: + +$$ +\mathcal {L} _ {\mathrm {K L}} = - \frac {1}{| R _ {Q} |} \sum_ {i = 1} ^ {|} R _ {Q} | \hat {P} _ {\mathrm {t r u e}} (i) \log \hat {P} _ {\mathrm {p r e d}} (i). +$$ + +This KL divergence loss encourages the model to align its predictions with the true temporal relevance of the clips, thereby improving its ability to correctly retrieve and ground the relevant video segments for generating accurate, timely responses. + +# 3.3. Precise Interaction with Streaming Video + +Once the interaction is triggered, an asynchronous precise interaction module generates a response with fine-grained detail, tailored to the specific context of the video stream. That is, for the $k$ -th answer outputs, the precise interaction module receives the question, the former $k - 1$ history answers as well as the grounded clip features at the current timestamp. This design allows the system to generate appropriate responses by incorporating rich context from both the video content and the ongoing conversation in a streaming setting. + +A key advantage of this asynchronous approach is that it decouples the response generation from the video stream processing. This separation allows the streaming video to continue being processed in parallel, without waiting for response generation to complete. As a result, the system remains highly efficient and responsive, ensuring that the video content is continuously monitored and processed while still generating contextually accurate responses at the appropriate moments. + +While the Response Decision Module may determine that a response is needed, there is a chance that no actual output is required. To ensure accuracy in real-time interactions, we enhance the model's responsiveness by training it on a streaming QA dataset that simulates such real-time scenarios. These data include both positive samples (where a response is required) and negative samples (where no response is necessary and the model should keep silent). By + +dynamically sampling these positive and negative instances, the model learns to either produce the responses that are incremental to the historical answers, summarizing the newly emerging cues in the video stream or output the special $\langle \mathrm{SILENT} \rangle$ token denoting keeping silent. + +This adaptive reasoning approach ensures that the model remains responsive to new information while avoiding redundancy. By focusing on unanswered content and utilizing past interactions, the model delivers timely and relevant responses, enhancing the overall efficiency and user experience in streaming dialogue generation. + +# 4. Experiments + +# 4.1. Datasets + +The training process comprises two parts: response decision with temporal grounding and LLM reasoning. For training the streaming video processor and response decision module, we combine the GroundVQA [13] and ET-Instruct [41] datasets, enriched with temporal annotations to provide supervision for streaming responses and temporal grounding labels. Additionally, to enhance the model's ability to handle questions that require simple reasoning, we curate a set of 50K implicit QA pairs with time labels, guiding the model to locate relevant cues within the video stream. + +For LLM reasoning, we construct a dataset of 122K streaming video QA pairs based on timestamped QA pairs from ET-Instruct [41] and integrate conventional video QA data from VideoChatGPT [43] and LLaVA-Next-Video [29]. + +# 4.2. Benchmarks + +For evaluation, we adopt several established long-video QA benchmarks, including EgoSchema [44], VideoMME [20], and MLVU [71]. EgoSchema consists of over 5K videos, each approximately 3 minutes long, while VideoMME and MLVU include videos of varying lengths, from a few minutes to several hours. We report accuracy on multiple-choice questions for comparison across these benchmarks. + +In addition, we create a streaming video QA benchmark using a subset of ETBench to measure the model's proactive response capability in real-time video interactions. Specifically, we select six subtasks from ETBench suitable for streaming response evaluation: step localization (SLC), dense video captioning (DVC), temporal action localization (TAL), temporal video grounding (TVG), episodic memory (EPM), and video highlight detection (VHD), encompassing a total of 4,460 videos. For this benchmark, the instruction is provided at the beginning of each video, and the model is required to produce the correct responses at appropriate timestamps as the video plays. We follow the evaluation metrics specified in the original benchmarks to ensure a fair comparison. + +
MethodLLM SizeFramesEgoSchemaMLVUVideoMME
Video-LLaVA [35]7B838.447.339.9
Chat-UniVi [28]7B64--40.6
LLaMA-VID [33]7B1 FPS38.533.2-
TimeChat [53]7B9633.030.930.2
MovieChat [55]7B204853.525.838.2
Video-LLaMA2 [9]7B1651.748.547.9
LLaVA-Next-Video [70]7B3243.9-46.6
ShareGPT4Video [6]8B16-46.439.9
VideoChat2 [31]7B1654.447.939.5
LongVA [69]7B128-56.352.6
Kangaroo [40]8B64-61.056.0
Video-CCAM [19]14B96-63.153.2
VideoXL [54]7B128-64.955.5
Ours7B1 FPS55.661.757.2
+ +Table 1. Comparison on long video benchmarks. We report the accuracy on the EgoSchema full set, MLVU multiple-choice questions, and the VideoMME overall set without subtitles. For a fair comparison, we also present the model size of the LLM and the number of sampled frames. + +# 4.3. Implementation Details + +The architecture consists of a vision encoder, a compact LLM as the proactive streaming video processor for response decisions, and a larger LLM for the precision interaction module. In practice, the input video frames are resized to $224 \times 224$ , and a CLIP-L/14 [51] is employed to extract frame-wise features. Following the token compression techniques in [48], we concatenate adjacent tokens and use the compact LLM, instantiated as Qwen2-1.5B [49], to produce time-aware compressed clip-wise features along with clip indicators. Next, we reuse this compact LLM to process the sequence consisting of global memory, question texts, and clip features for response decisions. The final LLM, instantiated as Qwen2-7B [49], receives the grounded clips and global memory to generate responses at the necessary timestamps. + +Training is conducted in two stages. In the first stage, we train the response decision module and the video encoder. In the second stage, we freeze these components and train the final LLM. During training, we insert instructions at arbitrary timestamps within the video to enhance the model's ability to handle varied circumstances. For inference, on conventional video benchmarks, we place the question at the end of the video to ensure fair comparisons with existing works. For streaming evaluation, the question is posed at the beginning to enable proactive responses. + +# 4.4. Conventional Video Understanding + +In this section, we compare our model with existing video LLMs on conventional video QA benchmarks, where the model is required to provide one answer after watching the complete video. + +We first present a comparative analysis across three challenging long-video benchmarks. As shown in Table 1, we report the accuracy on the EgoSchema full set, MLVU multiple-choice questions, and the VideoMME overall set without subtitles. Generally, our disentangled architecture handles this conventional scenario well and achieves competitive performance. Notably, on EgoSchema, which requires long temporal reasoning, our method achieves a leading performance of 55.6, demonstrating strong temporal perception. + +In terms of MLVU and VideoMME, which consist of videos ranging from minutes to hours long, our model's promising results highlight its ability to efficiently process contextual information across varying temporal lengths. This capability is crucial for streaming video interactions, as it allows the model to accurately summarize context when questions are inserted at arbitrary temporal positions. + +Additionally, we adopt a recent time-aware benchmark, ETBench, to evaluate temporal awareness. We select six subtasks that require the model to predict explicit event timestamps and report the results in Table 2. Following [41], we report the F1 score for temporal grounding evaluation and the sentence similarity score for caption evaluation. In the conventional setting, our method is able to capture the timestamps corresponding to specific questions. Without a specialized design for time tokens, our model achieves the highest F1 score in temporal video grounding and episodic memory subtasks, even surpassing models with specialized temporal tokens. The promising performance on dense video captioning and step localization further demonstrates the model's ability in both accurate temporal grounding and precise visual perception. + +
MethodFramesTVGF1EPMF1TALF1VHDF1DVCF1DVCsimSLCF1SLCsim
Conventional video QA inference.
w/ specialized time tokens
VTG-LLM [24]9615.93.714.448.240.218.620.814.4
LITA [26]10022.24.618.023.939.717.221.012.2
ETChat [41]1 FPS38.610.230.862.538.419.724.414.6
w/o specialized time tokens
VideoChatGPT [43]1007.01.315.128.88.811.35.710.2
Video-LLaVA [35]87.01.915.028.928.015.00.98.3
LLaMA-VID [33]1 FPS5.51.28.030.027.112.65.211.1
Video-LLaMA2 [9]80.10.00.01,50.614.50.015.2
PLLAVA [60]166.91.15.728.913.310.69.711.8
VTimeLLM [25]1007.61.918.228.912.413.18.76.4
TimeChat [53]9626.23.910.140.516.612.55.69.2
Ours1 FPS43.617.229.951.531.617.814.111.7
Streaming video QA inference.
VideoLLM-Online [5]2 FPS13.23.89.122.424.013.49.910.1
Ours1 FPS36.115.527.354.233.818.918.812.4
+ +Table 2. Comparison on ETBench. We present the results for two different settings. In the conventional video QA setting, the model is required to answer the question after watching the entire video. In the streaming setting, the question is placed at the beginning of the video, and the model is expected to provide real-time responses. We report performance on six subtasks that are suitable for both settings. + +# 4.5. Streaming Video Understanding + +In addition to the conventional video QA evaluation, we also assess the performance of our model in streaming video interactions. Specifically, we place the question at the beginning of the video and expect the model to produce a response once it detects relevant clues while the video stream plays in real-time. The model should remain silent when no new information is available. In the bottom lines of Table 2, we compare our model with VideoLLM-online [5], the only existing video LLM capable of streaming video interactions. Our model consistently outperforms VideoLLM-online across all metrics, with significant advantages in temporal grounding. This demonstrates that the disentangled design makes our model have far more powerful temporal awareness, enabling more agile proactive responses aligned with specific instructions. + +Notably, on dense video captioning and step localization tasks, our model achieves both more precise temporal grounding and more comprehensive descriptions in streaming mode than the conventional setting. This indicates that our disentangled architecture can effectively monitor the video stream in real-time and proactively generate informative responses according to the instructions. And the streaming interaction has the potential to achieve more powerful video understanding. + +Additionally, we present a quantitative comparison between our model and VideoLLM-online in Fig. 3. From this comparison, when faced with questions requiring multi-step reasoning, our model can gradually identify the necessary clues from the video stream and generate informative + +answers step by step, while VideoLLM-online [5] fails to do so. For example, from the 'thirsty' in the question, our model can associate it with the drinks appearing in the video stream, and infer what actions are needed based on the context. In contrast, VideoLLM-online only gives simple descriptions of the scene or ongoing actions. + +A key advantage of our method is the disentangled perception, decision, and reaction architecture, which enables the model to simultaneously process the streaming video input and generate responses in a non-blocking manner. However, VideoLLM-online has to experience interruptions in the video stream during answer generation. + +# 4.6. Ablation Study + +Clip Segmentation Strategy. We first present an ablation study on clip segmentation. We compare the standard uniform clip segmentation (16 frames per clip sampled at 1 FPS) with our scene-based non-uniform segmentation. We report conventional QA accuracy on MLVU and VideoMME, as well as streaming metrics on temporal video grounding and dense video captioning on ETBench in Table 3. The scene-based segmentation yields superior results in both conventional and streaming settings. This aligns with our motivation that scene-based segmentation preserves more structural video information, which facilitates better model learning and more timely responses. + +Special Token Design. We also ablate three special tokens in our architecture, i.e., $\langle \mathrm{ANS} \rangle$ and $\langle \mathrm{TODO} \rangle$ in the streaming video processor, $\langle \mathrm{SILENT} \rangle$ token in the final LLM. These three special tokens mainly impact the stream- + +![](images/4f71b6ac993e850e18f9a7bf89d33ecf0a440e0fc85f0e25fcbc7c5644522f71.jpg) +Figure 3. Quantitative comparison between Dispider and VideoLLM-online in streaming video understanding. The question is inserted at the beginning, and we show the model's response in either answer texts or the state of keeping silent. + +
ClipMLVUV-MMETVG\( _{F1} \)DVC\( _{F1} \)DVC\( _{Sim} \)
Uniform59.855.434.533.118.1
Scene-based61.757.236.133.818.9
+ +Table 3. Ablation study on the clip segmentation. We compare uniform 16-frame clip segmentation and our scene-based segmentation with SigLip. + +
<ANS><TODO><SILENT>TVG\( _{F1} \)DVC\( _{F1} \)DVC\( _{Sim} \)
XXX20.119.712.3
XX✓26.324.913.1
✓X✓35.231.017.2
X✓✓28.725.614.5
✓✓X35.530.216.8
✓✓✓36.133.818.9
+ +Table 4. Ablation study on the special token designs. We respectively explore the effectiveness of ⟨ANS⟩ and ⟨TODO⟩ in streaming video processor as well as ⟨SILENT⟩ in the final LLM. + +ing video interactions, so we report the streaming evaluation metrics on temporal video grounding and dense video captioning in Table 4. There are three observations. First, the absence of $\langle \mathrm{ANS}\rangle$ token can lead the model to be unaware of the timestamps at which a response has been given. As a result, if relevant clues have appeared in the video contexts, the model tends to produce a response, resulting in a high recall but a low accuracy. Second, the $\langle \mathrm{TODO}\rangle$ serves as an indicator that reminds the streaming processor to decide whether to respond. The performance slightly degrades without this special token. Third, the $\langle \mathrm{SILENT}\rangle$ token in the final LLM serves as a secondary filter for response decision. If the preceding streaming processor incorrectly identifies a timestamp as requiring a response, the $\langle \mathrm{SILENT}\rangle$ token enables the LLM to rethink whether an answer is needed + +referring to the historical QA interaction contexts. + +# 5. Conclusion + +In this work, we introduced Dispider, a novel framework designed to enable active real-time interaction with video LLMs. By disentangling the key capabilities of perception, decision, and reaction, and adopting an asynchronous processing approach, Dispider overcomes the inherent conflicts that hinder real-time interaction in traditional video LLMs. Our approach ensures that the model can continuously process the video stream while providing timely, contextually accurate, and precise responses to user interactions. The Proactive Streaming Video Processing module optimizes the video analysis by focusing on relevant content, while the Precise Interaction module generates detailed responses asynchronously, allowing for uninterrupted video processing. We demonstrated the effectiveness of Dispider through extensive experiments on both conventional and streaming video QA benchmarks, where it outperformed existing methods in proactive response capabilities, temporal awareness, and computational efficiency. Dispider's disentangled architecture and ability to handle long-duration video streams make it an ideal solution for real-time interactive applications. + +# Acknowledgement + +This project is funded in part by Shanghai Artificial Intelligence Laboratory, Shanghai Innovation Institute, the National Key R&D Program of China (2022ZD0160201), the Centre for Perceptual and Interactive Intelligence (CPII) Ltd under the Innovation and Technology Commission (ITC)'s InnoHK. Dahua Lin is a PI of CPII under the InnoHK. + +# References + +[1] Jean-Baptiste Alayrac, Jeff Donahue, Pauline Luc, Antoine Miech, Iain Barr, Yana Hasson, Karel Lenc, Arthur Mensch, Katherine Millican, Malcolm Reynolds, et al. Flamingo: a visual language model for few-shot learning. Advances in neural information processing systems, 35:23716-23736, 2022. 2 +[2] Jinze Bai, Shuai Bai, Shusheng Yang, Shijie Wang, Sinan Tan, Peng Wang, Junyang Lin, Chang Zhou, and Jingren Zhou. Qwen-vl: A frontier large vision-language model with versatile abilities. arXiv preprint arXiv:2308.12966, 2023. 2 +[3] Yuntao Bai, Andy Jones, Kamal Ndousse, Amanda Askell, Anna Chen, Nova DasSarma, Dawn Drain, Stanislav Fort, Deep Ganguli, Tom Henighan, et al. Training a helpful and harmless assistant with reinforcement learning from human feedback. arXiv preprint arXiv:2204.05862, 2022. +[4] Zheng Cai, Maosong Cao, Haojiong Chen, Kai Chen, Keyu Chen, Xin Chen, Xun Chen, Zehui Chen, Zhi Chen, Pei Chu, et al. Internl m2 technical report. arXiv preprint arXiv:2403.17297, 2024. 2 +[5] Joya Chen, Zhaoyang Lv, Shiwei Wu, Kevin Qinghong Lin, Chenan Song, Difei Gao, Jia-Wei Liu, Ziteng Gao, Dongxing Mao, and Mike Zheng Shou. Videollm-online: Online video large language model for streaming video. In CVPR, 2024. 1, 2, 3, 7 +[6] Lin Chen, Xilin Wei, Jinsong Li, Xiaoyi Dong, Pan Zhang, Yuhang Zang, Zehui Chen, Haodong Duan, Bin Lin, Zhenyu Tang, et al. Sharegpt4video: Improving video understanding and generation with better captions. arXiv preprint arXiv:2406.04325, 2024. 2, 6 +[7] Ling-Hao Chen, Shunlin Lu, Ailing Zeng, Hao Zhang, Benyou Wang, Ruimao Zhang, and Lei Zhang. Motionllm: Understanding human behaviors from human motions and videos. arXiv preprint arXiv:2405.20340, 2024. +[8] Zhe Chen, Jiannan Wu, Wenhai Wang, Weijie Su, Guo Chen, Sen Xing, Muyan Zhong, Qinglong Zhang, Xizhou Zhu, Lewei Lu, Bin Li, Ping Luo, Tong Lu, Yu Qiao, and Jifeng Dai. Internvl: Scaling up vision foundation models and aligning for generic visual-linguistic tasks. arXiv preprint arXiv:2312.14238, 2023. +[9] Zesen Cheng, Sicong Leng, Hang Zhang, Yifei Xin, Xin Li, Guanzheng Chen, Yongxin Zhu, Wenqi Zhang, Ziyang Luo, Deli Zhao, et al. Videollama 2: Advancing spatial-temporal modeling and audio understanding in video-llms. arXiv preprint arXiv:2406.07476, 2024. 2, 6, 7 +[10] Aakanksha Chowdhery et al. Palm: Scaling language modeling with pathways. arXiv preprint arXiv:2303.12345, 2023. 2 +[11] Wenliang Dai, Junnan Li, Dongxu Li, Anthony Meng Huat Tiong, Junqi Zhao, Weisheng Wang, Boyang Li, Pascale Fung, and Steven Hoi. Instructclip: Towards general-purpose vision-language models with instruction tuning, 2023. 2 +[12] Jacob Devlin, Ming-Wei Chang, Kenton Lee, and Kristina Toutanova. Bert: Pre-training of deep bidirectional transformers for language understanding. arXiv preprint arXiv:1810.04805, 2018. 2 + +[13] Shangzhe Di and Weidi Xie. Grounded question-answering in long egocentric videos. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 12934–12943, 2024. 5 +[14] Shuangrui Ding, Maomao Li, Tianyu Yang, Rui Qian, Haohang Xu, Qingyi Chen, Jue Wang, and Hongkai Xiong. Motion-aware contrastive video representation learning via foreground-background merging. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 9716-9726, 2022. 1 +[15] Shuangrui Ding, Peisen Zhao, Xiaopeng Zhang, Rui Qian, Hongkai Xiong, and Qi Tian. Prune spatio-temporal tokens by semantic-aware temporal accumulation. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 16945-16956, 2023. 1 +[16] Shuangrui Ding, Zihan Liu, Xiaoyi Dong, Pan Zhang, Rui Qian, Conghui He, Dahua Lin, and Jiaqi Wang. Songcomposer: A large language model for lyric and melody composition in song generation. arXiv preprint arXiv:2402.17645, 2024. 2 +[17] Shuangrui Ding, Rui Qian, Haohang Xu, Dahua Lin, and Hongkai Xiong. Betrayed by attention: A simple yet effective approach for self-supervised video object segmentation. In European Conference on Computer Vision, pages 215-233. Springer, 2024. 1 +[18] Yue Fan, Xiaojian Ma, Rujie Wu, Yuntao Du, Jiaqi Li, Zhi Gao, and Qing Li. Videoagent: A memory-augmented multimodal agent for video understanding. arXiv preprint arXiv:2403.11481, 2024. 2 +[19] Jiajun Fei, Dian Li, Zhidong Deng, Zekun Wang, Gang Liu, and Hui Wang. Video-ccam: Enhancing video-language understanding with causal cross-attention masks for short and long videos. arXiv preprint arXiv:2408.14023, 2024.6 +[20] Chaoyou Fu, Yuhan Dai, Yondong Luo, Lei Li, Shuhuai Ren, Renrui Zhang, Zihan Wang, Chenyu Zhou, Yunhang Shen, Mengdan Zhang, et al. Video-mme: The first-ever comprehensive evaluation benchmark of multi-modal llms in video analysis. arXiv preprint arXiv:2405.21075, 2024. 5 +[21] Chaoyou Fu, Haojia Lin, Zuwei Long, Yunhang Shen, Meng Zhao, Yifan Zhang, Xiong Wang, Di Yin, Long Ma, Xiawu Zheng, Ran He, Rongrong Ji, Yunsheng Wu, Caifeng Shan, and Xing Sun. Vita: Towards open-source interactive omni multimodal lIm. arXiv preprint arXiv:2408.05211, 2024. 2 +[22] Gemini Team Google. Gemini: a family of highly capable multimodal models. arXiv preprint arXiv:2312.11805, 2023. 2 +[23] Gemini Team Google. Gemini 1.5: Unlocking multimodal understanding across millions of tokens of context. arXiv preprint arXiv:2403.05530, 2024. 2 +[24] Yongxin Guo, Jingyu Liu, Mingda Li, Xiaoying Tang, Xi Chen, and Bo Zhao. Vtg-llm: Integrating timestamp knowledge into video llms for enhanced video temporal grounding. arXiv preprint arXiv:2405.13382, 2024. 7 +[25] Bin Huang, Xin Wang, Hong Chen, Zihan Song, and Wenwu Zhu. Vtimellm: Empower llm to grasp video moments. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 14271-14280, 2024. 7 + +[26] De-An Huang, Shijia Liao, Subhashree Radhakrishnan, Hongxu Yin, Pavlo Molchanov, Zhiding Yu, and Jan Kautz. Lita: Language instructed temporal-localization assistant. In European Conference on Computer Vision, pages 202-218. Springer, 2025. 7 +[27] Albert Q Jiang, Alexandre Sablayrolles, Arthur Mensch, Chris Bamford, Devendra Singh Chaplot, Diego de las Casas, Florian Bressand, Gianna Lengyel, Guillaume Lample, Lucile Saulnier, et al. Mistral 7b. arXiv preprint arXiv:2310.06825, 2023. 2 +[28] Peng Jin, Ryuichi Takanobu, Caiwan Zhang, Xiaochun Cao, and Li Yuan. Chat-univi: Unified visual representation empowers large language models with image and video understanding. arXiv preprint arXiv:2311.08046, 2023. 6 +[29] Bo Li, Yuanhan Zhang, Dong Guo, Renrui Zhang, Feng Li, Hao Zhang, Kaichen Zhang, Yanwei Li, Ziwei Liu, and Chunyuan Li. Llava-onevision: Easy visual task transfer. arXiv preprint arXiv:2408.03326, 2024. 5 +[30] KunChang Li, Yinan He, Yi Wang, Yizhuo Li, Wenhai Wang, Ping Luo, Yali Wang, Limin Wang, and Yu Qiao. Videochat: Chat-centric video understanding. arXiv preprint arXiv:2305.06355, 2023. 2 +[31] Kunchang Li, Yali Wang, Yinan He, Yizhuo Li, Yi Wang, Yi Liu, Zun Wang, Jilan Xu, Guo Chen, Ping Luo, et al. Mvbench: A comprehensive multi-modal video understanding benchmark. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 22195-22206, 2024. 6 +[32] Yanwei Li, Chengyao Wang, and Jiaya Jia. Llama-vid: An image is worth 2 tokens in large language models. arXiv preprint arXiv:2311.17043, 2023. 1, 2 +[33] Yanwei Li, Chengyao Wang, and Jiaya Jia. Llama-vid: An image is worth 2 tokens in large language models. 2024. 2, 6, 7 +[34] Yifei Li, Junbo Niu, Ziyang Miao, Chunjiang Ge, Yuanhang Zhou, Qihao He, Xiaoyi Dong, Haodong Duan, Shuangrui Ding, Rui Qian, et al. Ovo-bench: How far is your videolms from real-world online video understanding? arXiv preprint arXiv:2501.05510, 2025. 3 +[35] Bin Lin, Bin Zhu, Yang Ye, Munan Ning, Peng Jin, and Li Yuan. Video-llava: Learning united visual representation by alignment before projection. arXiv preprint arXiv:2311.10122, 2023. 6, 7 +[36] Kevin Lin, Faisal Ahmed, Linjie Li, Chung-Ching Lin, Ehsan Azarnasab, Zhengyuan Yang, Jianfeng Wang, Lin Liang, Zicheng Liu, Yumao Lu, et al. Mm-vid: Advancing video understanding with gpt-4v (ision). arXiv preprint arXiv:2310.19773, 2023. 2 +[37] Haotian Liu, Chunyuan Li, Yuheng Li, and Yong Jae Lee. Improved baselines with visual instruction tuning. arXiv preprint arXiv:2310.03744, 2023. 2 +[38] Haotian Liu, Chunyuan Li, Qingyang Wu, and Yong Jae Lee. Visual instruction tuning, 2023. +[39] Haotian Liu, Chunyuan Li, Yuheng Li, Bo Li, Yuanhan Zhang, Sheng Shen, and Yong Jae Lee. Llava next: Improved reasoning,OCR, and world knowledge, 2024. 2 +[40] Jiajun Liu, Yibing Wang, Hanghang Ma, Xiaoping Wu, Xiaqi Ma, Xiaoming Wei, Jianbin Jiao, Enhua Wu, and Jie Hu. + +Kangaroo: A powerful video-language model supporting long-context video input. arXiv preprint arXiv:2408.15542, 2024.6 +[41] Ye Liu, Zongyang Ma, Zhongang Qi, Yang Wu, Chang Wen Chen, and Ying Shan. E.t. bench: Towards open-ended event-level video-language understanding. In Neural Information Processing Systems (NeurIPS), 2024. 5, 6, 7 +[42] Xudong Lu, Yinghao Chen, Cheng Chen, Hui Tan, Boheng Chen, Yina Xie, Rui Hu, Guanxin Tan, Renshou Wu, Yan Hu, et al. Bluem-v-3b: Algorithm and system co-design for multimodal large language models on mobile devices. arXiv preprint arXiv:2411.10640, 2024. 2 +[43] Muhammad Maaz, Hanoona Rasheed, Salman Khan, and Fahad Shahbaz Khan. Video-chatgpt: Towards detailed video understanding via large vision and language models. arXiv preprint arXiv:2306.05424, 2023. 2, 5, 7 +[44] Karttikeya Mangalam, Raiymbek Akshulakov, and Jitendra Malik. Egoschema: A diagnostic benchmark for very long-form video language understanding. Advances in Neural Information Processing Systems, 36, 2024. 5 +[45] OpenAI. Gpt-4 technical report. 2023. 2 +[46] Long Ouyang, Jeffrey Wu, Xu Jiang, Diogo Almeida, Carroll Wainwright, Pamela Mishkin, Chong Zhang, Sandhini Agarwal, Katarina Slama, Alex Ray, et al. Training language models to follow instructions with human feedback. Advances in neural information processing systems, 35:27730-27744, 2022. 2 +[47] Rui Qian, Shuangrui Ding, and Dahua Lin. Rethinking image-to-video adaptation: An object-centric perspective. In European Conference on Computer Vision, pages 329-348. Springer, 2024. 1 +[48] Rui Qian, Xiaoyi Dong, Pan Zhang, Yuhang Zang, Shuangrui Ding, Dahua Lin, and Jiaqi Wang. Streaming long video understanding with large language models. In The Thirty-eighth Annual Conference on Neural Information Processing Systems, 2024. 3, 6 +[49] Team Qwen. Qwen2 technical report, 2024. 6 +[50] Alec Radford, Karthik Narasimhan, Tim Salimans, and Ilya Sutskever. Improving language understanding by generative pre-training. 2018. 2 +[51] Alec Radford, Jong Wook Kim, Chris Hallacy, Aditya Ramesh, Gabriel Goh, Sandhini Agarwal, Girish Sastry, Amanda Askell, Pamela Mishkin, Jack Clark, et al. Learning transferable visual models from natural language supervision. In International conference on machine learning, pages 8748-8763, 2021. 6 +[52] Colin Raffel, Noam Shazeer, Adam Roberts, Katherine Lee, Sharan Narang, Michael Matena, Yanqi Zhou, Wei Li, and Peter J Liu. Exploring the limits of transfer learning with a unified text-to-text transformer. JMLR, 2020. 2 +[53] Shuhuai Ren, Linli Yao, Shicheng Li, Xu Sun, and Lu Hou. Timechat: A time-sensitive multimodal large language model for long video understanding. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 14313-14323, 2024. 2, 6, 7 +[54] Yan Shu, Peitian Zhang, Zheng Liu, Minghao Qin, Junjie Zhou, Tiejun Huang, and Bo Zhao. Video-xl: Extra-long + +vision language model for hour-scale video understanding. arXiv preprint arXiv:2409.14485, 2024. 6 +[55] Enxin Song, Wenhao Chai, Guanhong Wang, Yucheng Zhang, Haoyang Zhou, Feiyang Wu, Haozhe Chi, Xun Guo, Tian Ye, Yanting Zhang, et al. Moviechat: From dense token to sparse memory for long video understanding. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 18221-18232, 2024. 2, 6 +[56] Hugo Touvron, Thibaut Lavril, Gautier Izacard, Xavier Martinet, Marie-Anne Lachaux, Timothée Lacroix, Baptiste Rozière, Naman Goyal, Eric Hambro, Faisal Azhar, et al. Llama: Open and efficient foundation language models. arXiv preprint arXiv:2302.13971, 2023. 2 +[57] Ziyang Wang, Shoubin Yu, Elias Stengel-Eskin, Jaehong Yoon, Feng Cheng, Gedas Bertasius, and Mohit Bansal. Videotree: Adaptive tree-based video representation for llm reasoning on long videos. arXiv preprint arXiv:2405.19209, 2024. 2 +[58] Qi Wu, Yubo Zhao, Yifan Wang, Yu-Wing Tai, and Chikeung Tang. Motionllm: Multimodal motion-language learning with large language models. arXiv preprint arXiv:2405.17013, 2024. +[59] Haiyang Xu, Qinghao Ye, Mingshi Yan, Yaya Shi, Jiabo Ye, Yuanhong Xu, Chenliang Li, Bin Bi, Qiuchen Qian, Wei Wang, Guohai Xu, Ji Zhang, Songfang Huang, Feiran Huang, and Jingren Zhou. mplug-2: A modularized multimodal foundation model across text, image and video. arXiv preprint arXiv:2302.00402, 2023. 2 +[60] Lin Xu, Yilin Zhao, Daquan Zhou, Zhijie Lin, See Kiong Ng, and Jiashi Feng. Pllava: Parameter-free llava extension from images to videos for video dense captioning. arXiv preprint arXiv:2404.16994, 2024. 7 +[61] Qinghao Ye, Haiyang Xu, Jiabo Ye, Ming Yan, Anwen Hu, Haowei Liu, Qi Qian, Ji Zhang, and Fei Huang. mplug-owl2: Revolutionizing multi-modal large language model with modality collaboration. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 13040–13051, 2024. 2 +[62] Alex Young, Bei Chen, Chao Li, Chengen Huang, Ge Zhang, Guanwei Zhang, Heng Li, Jiangcheng Zhu, Jianqun Chen, Jing Chang, et al. Yi: Open foundation models by 01. ai. arXiv preprint arXiv:2403.04652, 2024. 2 +[63] Xiaohua Zhai, Basil Mustafa, Alexander Kolesnikov, and Lucas Beyer. Sigmoid loss for language image pre-training. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 11975-11986, 2023. 3 +[64] Ce Zhang, Taixi Lu, Md Mohaiminul Islam, Ziyang Wang, Shoubin Yu, Mohit Bansal, and Gedas Bertasius. A simple llm framework for long-range video question-answering. arXiv preprint arXiv:2312.17235, 2023. 2 +[65] Haoji Zhang, Yiqin Wang, Yansong Tang, Yong Liu, Jiashi Feng, Jifeng Dai, and Xiaojie Jin. Flash-vstream: Memory-based real-time understanding for long video streams. 2024. 3 +[66] Pan Zhang, Xiaoyi Dong, Bin Wang, Yuhang Cao, Chao Xu, Linke Ouyang, Zhiyuan Zhao, Haodong Duan, Songyang Zhang, Shuangrui Ding, et al. Internlmxcomposer: A vision-language large model for advanced + +text-image comprehension and composition. arXiv preprint arXiv:2309.15112, 2023. 2 +[67] Pan Zhang, Xiaoyi Dong, Yuhang Cao, Yuhang Zang, Rui Qian, Xilin Wei, Lin Chen, Yifei Li, Junbo Niu, Shuangrui Ding, et al. Internlm-xcomposer2. 5-omnilive: A comprehensive multimodal system for long-term streaming video and audio interactions. arXiv preprint arXiv:2412.09596, 2024.3 +[68] Pan Zhang, Xiaoyi Dong, Yuhang Zang, Yuhang Cao, Rui Qian, Lin Chen, Qipeng Guo, Haodong Duan, Bin Wang, Linke Ouyang, et al. Internlm-xcomposer-2.5: A versatile large vision language model supporting long-contextual input and output. arXiv preprint arXiv:2407.03320, 2024. 2 +[69] Peiyuan Zhang, Kaichen Zhang, Bo Li, Guangtao Zeng, Jingkang Yang, Yuanhan Zhang, Ziyue Wang, Haoran Tan, Chunyuan Li, and Ziwei Liu. Long context transfer from language to vision. arXiv preprint arXiv:2406.16852, 2024. 6 +[70] Yuanhan Zhang, Bo Li, haotian Liu, Yong jae Lee, Liangke Gui, Di Fu, Jiashi Feng, Ziwei Liu, and Chunyuan Li. Llavanext: A strong zero-shot video understanding model, 2024. 6 +[71] Junjie Zhou, Yan Shu, Bo Zhao, Boya Wu, Shitao Xiao, Xi Yang, Yongping Xiong, Bo Zhang, Tiejun Huang, and Zheng Liu. Mlvu: A comprehensive benchmark for multi-task long video understanding. arXiv preprint arXiv:2406.04264, 2024.5 \ No newline at end of file diff --git a/2025/Dispider_ Enabling Video LLMs with Active Real-Time Interaction via Disentangled Perception, Decision, and Reaction/images.zip b/2025/Dispider_ Enabling Video LLMs with Active Real-Time Interaction via Disentangled Perception, Decision, and Reaction/images.zip new file mode 100644 index 0000000000000000000000000000000000000000..38415c40fb969be4da3b31cb799b7e5862b2ac79 --- /dev/null +++ b/2025/Dispider_ Enabling Video LLMs with Active Real-Time Interaction via Disentangled Perception, Decision, and Reaction/images.zip @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:a52586ed8b732493c4ce4b3c9db71bb12db81056ed52506c3844fa436cb2e9b2 +size 504490 diff --git a/2025/Dispider_ Enabling Video LLMs with Active Real-Time Interaction via Disentangled Perception, Decision, and Reaction/layout.json b/2025/Dispider_ Enabling Video LLMs with Active Real-Time Interaction via Disentangled Perception, Decision, and Reaction/layout.json new file mode 100644 index 0000000000000000000000000000000000000000..7776e846f062859a4d64c017d3a2260a7880c18b --- /dev/null +++ b/2025/Dispider_ Enabling Video LLMs with Active Real-Time Interaction via Disentangled Perception, Decision, and Reaction/layout.json @@ -0,0 +1,8561 @@ +{ + "pdf_info": [ + { + "para_blocks": [ + { + "bbox": [ + 87, + 102, + 523, + 140 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 87, + 102, + 523, + 140 + ], + "spans": [ + { + "bbox": [ + 87, + 102, + 523, + 140 + ], + "type": "text", + "content": "Dispider: Enabling Video LLMs with Active Real-Time Interaction via Disentangled Perception, Decision, and Reaction" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 147, + 160, + 465, + 190 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 147, + 160, + 465, + 190 + ], + "spans": [ + { + "bbox": [ + 147, + 160, + 465, + 190 + ], + "type": "text", + "content": "Rui Qian" + }, + { + "bbox": [ + 147, + 160, + 465, + 190 + ], + "type": "inline_equation", + "content": "^{1*}" + }, + { + "bbox": [ + 147, + 160, + 465, + 190 + ], + "type": "text", + "content": " Shuangrui Ding" + }, + { + "bbox": [ + 147, + 160, + 465, + 190 + ], + "type": "inline_equation", + "content": "^{1*}" + }, + { + "bbox": [ + 147, + 160, + 465, + 190 + ], + "type": "text", + "content": " Xiaoyi Dong" + }, + { + "bbox": [ + 147, + 160, + 465, + 190 + ], + "type": "inline_equation", + "content": "^{1,2\\dagger}" + }, + { + "bbox": [ + 147, + 160, + 465, + 190 + ], + "type": "text", + "content": " Pan Zhang" + }, + { + "bbox": [ + 147, + 160, + 465, + 190 + ], + "type": "inline_equation", + "content": "^{2}" + }, + { + "bbox": [ + 147, + 160, + 465, + 190 + ], + "type": "text", + "content": " Yuhang Zang" + }, + { + "bbox": [ + 147, + 160, + 465, + 190 + ], + "type": "inline_equation", + "content": "^{2}" + }, + { + "bbox": [ + 147, + 160, + 465, + 190 + ], + "type": "text", + "content": " Yuhang Cao" + }, + { + "bbox": [ + 147, + 160, + 465, + 190 + ], + "type": "inline_equation", + "content": "^{2}" + }, + { + "bbox": [ + 147, + 160, + 465, + 190 + ], + "type": "text", + "content": " Dahua Lin" + }, + { + "bbox": [ + 147, + 160, + 465, + 190 + ], + "type": "inline_equation", + "content": "^{1,2,4}" + }, + { + "bbox": [ + 147, + 160, + 465, + 190 + ], + "type": "text", + "content": " Jiaqi Wang" + }, + { + "bbox": [ + 147, + 160, + 465, + 190 + ], + "type": "inline_equation", + "content": "^{2,3\\dagger}" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 145, + 190, + 465, + 217 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 145, + 190, + 465, + 217 + ], + "spans": [ + { + "bbox": [ + 145, + 190, + 465, + 217 + ], + "type": "text", + "content": "1 The Chinese University of Hong Kong 2 Shanghai AI Laboratory \n3 Shanghai Innovation Institute 4 CPII under InnoHK" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 225, + 220, + 382, + 231 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 225, + 220, + 382, + 231 + ], + "spans": [ + { + "bbox": [ + 225, + 220, + 382, + 231 + ], + "type": "text", + "content": "{qr021, ds023}@ie.cuhk.edu.hk" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 151, + 258, + 200, + 272 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 151, + 258, + 200, + 272 + ], + "spans": [ + { + "bbox": [ + 151, + 258, + 200, + 272 + ], + "type": "text", + "content": "Abstract" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 55, + 286, + 297, + 670 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 286, + 297, + 670 + ], + "spans": [ + { + "bbox": [ + 55, + 286, + 297, + 670 + ], + "type": "text", + "content": "Active real-time interaction with video LLMs introduces a new paradigm for human-computer interaction, where the model not only understands user intent but also responds while continuously processing streaming video on the fly. Unlike offline video LLMs, which analyze the entire video before answering questions, active real-time interaction requires three capabilities: 1) Perception: real-time video monitoring and interaction capturing. 2) Decision: raising proactive interaction in proper situations, 3) Reaction: continuous interaction with users. However, inherent conflicts exist among the desired capabilities. The Decision and Reaction require a contrary Perception scale and grain, and the autoregressive decoding blocks the real-time Perception and Decision during the Reaction. To unify the conflicted capabilities within a harmonious system, we present Dispider, a solution built on a Disentangled Perception, Decision, and Reaction framework. Dispider features a lightweight Proactive Streaming Video Processing module that tracks the video stream and identifies optimal moments for interaction. Once the interaction is triggered, an asynchronous Precise Interaction module provides detailed responses, while the processing module continues to monitor the video in the meantime. Our disentangled and asynchronous design ensures timely, contextually accurate, and computationally efficient responses, making Dispider ideal for active real-time interaction for long-duration video streams. Experiments prove that Dispider outperforms existing methods not only in its superior understanding of video content in conventional video QA settings, but also in proactive response capability and temporal awareness under the streaming setting. The code and model are released at https://github.com/Mark12Ding/Dispider." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 315, + 258, + 394, + 271 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 315, + 258, + 394, + 271 + ], + "spans": [ + { + "bbox": [ + 315, + 258, + 394, + 271 + ], + "type": "text", + "content": "1. Introduction" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 313, + 279, + 555, + 410 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 279, + 555, + 410 + ], + "spans": [ + { + "bbox": [ + 313, + 279, + 555, + 410 + ], + "type": "text", + "content": "The recent surge in multimodal large language models (MLLMs) has brought considerable attention to video understanding tasks [14, 15, 17, 47]. Given a video, these models aim to accurately comprehend its content and provide responses aligned with human expectations. However, the majority of current video LLMs are designed around an offline setting, where models are required to view the entire video before generating a single response. While effective for certain applications, this offline approach is impractical for real-time" + }, + { + "bbox": [ + 313, + 279, + 555, + 410 + ], + "type": "inline_equation", + "content": "^{1}" + }, + { + "bbox": [ + 313, + 279, + 555, + 410 + ], + "type": "text", + "content": ", interactive scenarios where users expect continuous feedback as video streams." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 313, + 410, + 556, + 566 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 410, + 556, + 566 + ], + "spans": [ + { + "bbox": [ + 313, + 410, + 556, + 566 + ], + "type": "text", + "content": "Consider a scenario where a user performs pull-ups with the assistance of a visual model. As illustrated in Figure 1, current offline video LLMs need to process the entire video before providing the total count of pull-ups. However, users expect real-time feedback, with each pull-up counted as it occurs: \"One... Two... Three...\" This real-time interaction requires the model to recognize each completed action instantly and synchronize its responses with the user's movements. The challenge lies in determining the precise moment to output the count along with the ongoing video stream. To this end, traditional offline video LLMs lack the flexibility to handle such demands, as they are not designed for continuous, dynamic video streams." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 313, + 566, + 556, + 687 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 566, + 556, + 687 + ], + "spans": [ + { + "bbox": [ + 313, + 566, + 556, + 687 + ], + "type": "text", + "content": "VideoLLM-online [5] has been a pioneering effort in addressing this issue. By leveraging the LLaMA model [32], they enable proactive response updates during streaming video inputs. However, it has a critical limitation: since it uses a single LLM for video processing and response generation, it cannot perform perception and answer reaction simultaneously. The autoregressive nature of next-token prediction in LLMs forces VideoLLM-online to alternate between perception (processing frames) and response generation, rather than handling both in parallel. This restriction" + } + ] + } + ], + "index": 11 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "spans": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "type": "text", + "content": "CVF" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 145, + 0, + 496, + 37 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 145, + 0, + 496, + 37 + ], + "spans": [ + { + "bbox": [ + 145, + 0, + 496, + 37 + ], + "type": "text", + "content": "This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 313, + 693, + 556, + 714 + ], + "type": "page_footnote", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 693, + 556, + 714 + ], + "spans": [ + { + "bbox": [ + 313, + 693, + 556, + 714 + ], + "type": "text", + "content": "1We use \"real-time\" loosely to emphasize online interaction, not strict frame-rate requirements." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 66, + 693, + 136, + 703 + ], + "type": "page_footnote", + "angle": 0, + "lines": [ + { + "bbox": [ + 66, + 693, + 136, + 703 + ], + "spans": [ + { + "bbox": [ + 66, + 693, + 136, + 703 + ], + "type": "text", + "content": "*Equal Contribution." + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 67, + 703, + 146, + 712 + ], + "type": "page_footnote", + "angle": 0, + "lines": [ + { + "bbox": [ + 67, + 703, + 146, + 712 + ], + "spans": [ + { + "bbox": [ + 67, + 703, + 146, + 712 + ], + "type": "text", + "content": "† Corresponding Author." + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "text", + "content": "24045" + } + ] + } + ], + "index": 15 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 0 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 58, + 109, + 515, + 192 + ], + "blocks": [ + { + "bbox": [ + 58, + 109, + 515, + 192 + ], + "lines": [ + { + "bbox": [ + 58, + 109, + 515, + 192 + ], + "spans": [ + { + "bbox": [ + 58, + 109, + 515, + 192 + ], + "type": "image", + "image_path": "255603027e52059f38adf1a94e4fabdbdb4af1fd96859265d3a78ff0e17ecf36.jpg" + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 55, + 243, + 555, + 288 + ], + "lines": [ + { + "bbox": [ + 55, + 243, + 555, + 288 + ], + "spans": [ + { + "bbox": [ + 55, + 243, + 555, + 288 + ], + "type": "text", + "content": "Figure 1. Comparison between offline and active real-time Video LLMs in processing a video sequence of a man performing pull-ups. The user asks the model to count the number of pull-ups. The offline Video LLMs process the entire video before providing a final count of three pull-ups. In contrast, the real-time Video LLMs provide immediate feedback after each pull-up, first counting one, then two, and finally reaching three, while remaining silent between responses." + } + ] + } + ], + "index": 16, + "angle": 0, + "type": "image_caption" + } + ], + "index": 5 + }, + { + "type": "image", + "bbox": [ + 65, + 206, + 78, + 219 + ], + "blocks": [ + { + "bbox": [ + 65, + 206, + 78, + 219 + ], + "lines": [ + { + "bbox": [ + 65, + 206, + 78, + 219 + ], + "spans": [ + { + "bbox": [ + 65, + 206, + 78, + 219 + ], + "type": "image", + "image_path": "7a6a8509236a8b0757bb9626b112faf76911731080344add1aeb9eed2bbfe84c.jpg" + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "image_body" + } + ], + "index": 6 + }, + { + "bbox": [ + 80, + 209, + 159, + 219 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 80, + 209, + 159, + 219 + ], + "spans": [ + { + "bbox": [ + 80, + 209, + 159, + 219 + ], + "type": "text", + "content": "Real-time Video LLM" + } + ] + } + ], + "index": 7 + }, + { + "type": "image", + "bbox": [ + 181, + 208, + 197, + 223 + ], + "blocks": [ + { + "bbox": [ + 181, + 208, + 197, + 223 + ], + "lines": [ + { + "bbox": [ + 181, + 208, + 197, + 223 + ], + "spans": [ + { + "bbox": [ + 181, + 208, + 197, + 223 + ], + "type": "image", + "image_path": "c18cb3313382d5bb3999b464a74a28f65af0717b9ec441139da3a37aa9bba761.jpg" + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "image_body" + } + ], + "index": 8 + }, + { + "bbox": [ + 198, + 211, + 269, + 220 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 198, + 211, + 269, + 220 + ], + "spans": [ + { + "bbox": [ + 198, + 211, + 269, + 220 + ], + "type": "text", + "content": ": The man did 1 pull-up." + } + ] + } + ], + "index": 9 + }, + { + "type": "image", + "bbox": [ + 295, + 209, + 309, + 223 + ], + "blocks": [ + { + "bbox": [ + 295, + 209, + 309, + 223 + ], + "lines": [ + { + "bbox": [ + 295, + 209, + 309, + 223 + ], + "spans": [ + { + "bbox": [ + 295, + 209, + 309, + 223 + ], + "type": "image", + "image_path": "8cc534a32a9d74bf06415d6300be7c567de6a3244ffe56c129325efbdf41b17f.jpg" + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "image_body" + } + ], + "index": 10 + }, + { + "type": "image", + "bbox": [ + 332, + 211, + 348, + 221 + ], + "blocks": [ + { + "bbox": [ + 332, + 211, + 348, + 221 + ], + "lines": [ + { + "bbox": [ + 332, + 211, + 348, + 221 + ], + "spans": [ + { + "bbox": [ + 332, + 211, + 348, + 221 + ], + "type": "image", + "image_path": "826a2d6569ce3904afa977fb792d52416a35c55c4a58f80e5c0f0d248a8fc6a6.jpg" + } + ] + } + ], + "index": 11, + "angle": 0, + "type": "image_body" + } + ], + "index": 11 + }, + { + "bbox": [ + 348, + 212, + 422, + 221 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 348, + 212, + 422, + 221 + ], + "spans": [ + { + "bbox": [ + 348, + 212, + 422, + 221 + ], + "type": "text", + "content": ": The man did 2 pull-ups." + } + ] + } + ], + "index": 12 + }, + { + "type": "image", + "bbox": [ + 430, + 207, + 447, + 223 + ], + "blocks": [ + { + "bbox": [ + 430, + 207, + 447, + 223 + ], + "lines": [ + { + "bbox": [ + 430, + 207, + 447, + 223 + ], + "spans": [ + { + "bbox": [ + 430, + 207, + 447, + 223 + ], + "type": "image", + "image_path": "61f5e1be321d1ef4ad7ce90747054fdf5249166d1d91f08af0a33d88b13e4f40.jpg" + } + ] + } + ], + "index": 13, + "angle": 0, + "type": "image_body" + } + ], + "index": 13 + }, + { + "type": "image", + "bbox": [ + 453, + 211, + 469, + 221 + ], + "blocks": [ + { + "bbox": [ + 453, + 211, + 469, + 221 + ], + "lines": [ + { + "bbox": [ + 453, + 211, + 469, + 221 + ], + "spans": [ + { + "bbox": [ + 453, + 211, + 469, + 221 + ], + "type": "image", + "image_path": "56b5b4d7e1e9970610d47a564691468b520acff2e16c08f0452729651286867e.jpg" + } + ] + } + ], + "index": 14, + "angle": 0, + "type": "image_body" + } + ], + "index": 14 + }, + { + "bbox": [ + 471, + 211, + 542, + 220 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 471, + 211, + 542, + 220 + ], + "spans": [ + { + "bbox": [ + 471, + 211, + 542, + 220 + ], + "type": "text", + "content": "The man did 3 pull-ups." + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 55, + 309, + 295, + 357 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 309, + 295, + 357 + ], + "spans": [ + { + "bbox": [ + 55, + 309, + 295, + 357 + ], + "type": "text", + "content": "delays the video inputs during responding, reduces responsiveness and hinders its ability to engage in real-time interactions effectively, especially during long-duration video streams that require continuous, uninterrupted processing." + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 55, + 361, + 296, + 566 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 361, + 296, + 566 + ], + "spans": [ + { + "bbox": [ + 55, + 361, + 296, + 566 + ], + "type": "text", + "content": "To address these limitations, we present Dispider, a novel framework designed for active real-time interaction with streaming videos. Dispider resolves the conflicts between perception, decision, and reaction through a disentangled, asynchronous design, where each component can operate in parallel. Specifically, we implement these functions in the Proactive Streaming Video Processing module and the Precise Interaction module. In the video processing module, with the streaming video input, we dynamically segment the video stream into non-uniform clips based on scene boundaries. These clips, combined with relevant historical context, are then processed by a real-time response decision module, which determines whether the model should generate a response or wait for additional video content. When an interaction is triggered, the Precise Interaction module generates context-aware responses without interrupting the ongoing video processing." + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 55, + 570, + 296, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 570, + 296, + 714 + ], + "spans": [ + { + "bbox": [ + 55, + 570, + 296, + 714 + ], + "type": "text", + "content": "This disentangled design ensures that the Dispider can provide timely, accurate, and computationally efficient responses, even for long-duration video streams. We evaluate Dispider's ability in both real-time video stream interactions, and conventional offline Video LLMs across long-video benchmarks (EgoSchema, VideoMME, MLVU) and time-sensitive tasks (ETBench). Dispider demonstrated strong performance, particularly excelling in temporal reasoning on EgoSchema and handling diverse video lengths effectively. In streaming tasks, it significantly surpasses VideoLLM-online [5] in temporal grounding, proactive response generation, and multi-step reasoning." + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 314, + 308, + 400, + 319 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 314, + 308, + 400, + 319 + ], + "spans": [ + { + "bbox": [ + 314, + 308, + 400, + 319 + ], + "type": "text", + "content": "2. Related Work" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 313, + 329, + 555, + 520 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 329, + 555, + 520 + ], + "spans": [ + { + "bbox": [ + 313, + 329, + 555, + 520 + ], + "type": "text", + "content": "Large Language Models. LLMs have revolutionized natural language processing, achieving remarkable performance across a wide range of tasks. Early models such as BERT [12] and T5 [52] used masked language modeling for pre-training. The shift to decoder-only models like GPT [50] introduced scalable architectures that enhanced language generation. Recent models, including PaLM [10], LLaMA [56], and GPT-4 [45], continue to push the boundaries with massive parameters and extensive training data. Techniques such as supervised fine-tuning and reinforcement learning from human feedback [2-4, 16, 22, 23, 27, 32, 46, 62] have further improved these models' ability to generate coherent, contextually relevant responses. Inspired by the reasoning capabilities of LLMs, we extend these models to the domain of streaming video understanding, where real-time interaction presents unique challenges." + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 313, + 521, + 556, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 521, + 556, + 713 + ], + "spans": [ + { + "bbox": [ + 313, + 521, + 556, + 713 + ], + "type": "text", + "content": "Video Large Language Models. The progress of multimodal LLMs has been significant, particularly with image-based models like InstructBLIP [11], LLaVA [37-39], and Flamingo [1], which integrate vision and language models. Extending this to video introduces challenges in managing both frame sequences and the context length of LLMs. Recent works on video LLMs [6-9, 18, 21, 33, 36, 42, 43, 55, 57-59, 61, 64, 66, 68] present novel strategies for processing long-form videos while maintaining effective reasoning. Notable approaches include TimeChat [53], which emphasizes temporal relationships across video frames, and MovieChat [55], which uses sparse memory to handle long videos. VideoChat [30] adopts a chat-centric approach, integrating video foundation models with LLMs to excel at spatiotemporal reasoning, event localization, and causal inference. Building on these video LLMs, our work intro" + } + ] + } + ], + "index": 22 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 71, + 80, + 85, + 95 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 71, + 80, + 85, + 95 + ], + "spans": [ + { + "bbox": [ + 71, + 80, + 85, + 95 + ], + "type": "text", + "content": "#" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 87, + 83, + 207, + 94 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 87, + 83, + 207, + 94 + ], + "spans": [ + { + "bbox": [ + 87, + 83, + 207, + 94 + ], + "type": "text", + "content": ": How many pull-ups has the man done?" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 349, + 81, + 436, + 94 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 349, + 81, + 436, + 94 + ], + "spans": [ + { + "bbox": [ + 349, + 81, + 436, + 94 + ], + "type": "text", + "content": "Offline Video LLM" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 439, + 80, + 455, + 95 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 439, + 80, + 455, + 95 + ], + "spans": [ + { + "bbox": [ + 439, + 80, + 455, + 95 + ], + "type": "text", + "content": "#" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 455, + 83, + 547, + 93 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 455, + 83, + 547, + 93 + ], + "spans": [ + { + "bbox": [ + 455, + 83, + 547, + 93 + ], + "type": "text", + "content": ": The man has done 3 pull-ups." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 293, + 749, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 749, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 749, + 317, + 757 + ], + "type": "text", + "content": "24046" + } + ] + } + ], + "index": 23 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 1 + }, + { + "para_blocks": [ + { + "bbox": [ + 55, + 72, + 294, + 95 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 72, + 294, + 95 + ], + "spans": [ + { + "bbox": [ + 55, + 72, + 294, + 95 + ], + "type": "text", + "content": "duces a pipeline designed to process streaming video inputs and generate real-time outputs." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 56, + 97, + 296, + 335 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 97, + 296, + 335 + ], + "spans": [ + { + "bbox": [ + 56, + 97, + 296, + 335 + ], + "type": "text", + "content": "Streaming Video Understanding. In practice, users typically expect models to operate online and interactively in real time. This is known as streaming video understanding, which involves processing continuous video streams while ensuring long-term temporal consistency and enabling interactive responses. Despite its practical significance, only a few works have explored this area [34, 67]. VideoLLM-online [5] introduces the LIVE framework for streaming dialogue, but it lacks an efficient module for handling long-term video inputs over extended periods and does not prioritize real-time interactivity. VideoStream [48] proposes memory-propagated encoding for long videos, yet its focus is on offline processing rather than real-time streaming. Similarly, Flash-VStream [65] addresses inference latency and memory efficiency for long video streams but overlooks real-time user interaction. In contrast, our approach centers on developing a real-time visual assistant for streaming video, emphasizing both long-context handling and interactive, real-time responses, thereby bridging the gap left by previous methods." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 55, + 346, + 111, + 358 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 346, + 111, + 358 + ], + "spans": [ + { + "bbox": [ + 55, + 346, + 111, + 358 + ], + "type": "text", + "content": "3. Method" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 55, + 366, + 180, + 378 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 366, + 180, + 378 + ], + "spans": [ + { + "bbox": [ + 55, + 366, + 180, + 378 + ], + "type": "text", + "content": "3.1. Problem Formulation" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 55, + 383, + 295, + 502 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 383, + 295, + 502 + ], + "spans": [ + { + "bbox": [ + 55, + 383, + 295, + 502 + ], + "type": "text", + "content": "In contrast to previous offline video LLMs, which generate responses only after processing the entire video, our approach requires the model to operate in real-time, providing continuous and interactive responses as the video is processed. Specifically, the model actively determines when sufficient information has been observed to provide a complete response, allowing it to produce answers promptly without waiting for the entire video to finish. Following the formulation of VideoLLM-online [5], we formulate this new setting as below." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 55, + 503, + 295, + 586 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 503, + 295, + 586 + ], + "spans": [ + { + "bbox": [ + 55, + 503, + 295, + 586 + ], + "type": "text", + "content": "Given a continuous video stream " + }, + { + "bbox": [ + 55, + 503, + 295, + 586 + ], + "type": "inline_equation", + "content": "V = \\{v_{1}, v_{2}, \\dots, v_{T}\\}" + }, + { + "bbox": [ + 55, + 503, + 295, + 586 + ], + "type": "text", + "content": ", where " + }, + { + "bbox": [ + 55, + 503, + 295, + 586 + ], + "type": "inline_equation", + "content": "v_{i}" + }, + { + "bbox": [ + 55, + 503, + 295, + 586 + ], + "type": "text", + "content": " represents the " + }, + { + "bbox": [ + 55, + 503, + 295, + 586 + ], + "type": "inline_equation", + "content": "i" + }, + { + "bbox": [ + 55, + 503, + 295, + 586 + ], + "type": "text", + "content": "-th video clip, and a context sequence " + }, + { + "bbox": [ + 55, + 503, + 295, + 586 + ], + "type": "inline_equation", + "content": "C_{t}" + }, + { + "bbox": [ + 55, + 503, + 295, + 586 + ], + "type": "text", + "content": " up to a specific time " + }, + { + "bbox": [ + 55, + 503, + 295, + 586 + ], + "type": "inline_equation", + "content": "t" + }, + { + "bbox": [ + 55, + 503, + 295, + 586 + ], + "type": "text", + "content": " (which includes prior vision-language interactions such as historical user queries, assistant responses, and video frames), our goal is to generate timely and accurate dialogue responses " + }, + { + "bbox": [ + 55, + 503, + 295, + 586 + ], + "type": "inline_equation", + "content": "R_{t}" + }, + { + "bbox": [ + 55, + 503, + 295, + 586 + ], + "type": "text", + "content": " without processing the entire video sequence." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 55, + 587, + 295, + 647 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 587, + 295, + 647 + ], + "spans": [ + { + "bbox": [ + 55, + 587, + 295, + 647 + ], + "type": "text", + "content": "At each time " + }, + { + "bbox": [ + 55, + 587, + 295, + 647 + ], + "type": "inline_equation", + "content": "t" + }, + { + "bbox": [ + 55, + 587, + 295, + 647 + ], + "type": "text", + "content": ", the model observes " + }, + { + "bbox": [ + 55, + 587, + 295, + 647 + ], + "type": "inline_equation", + "content": "C_t" + }, + { + "bbox": [ + 55, + 587, + 295, + 647 + ], + "type": "text", + "content": " and the video frames up to that time " + }, + { + "bbox": [ + 55, + 587, + 295, + 647 + ], + "type": "inline_equation", + "content": "V_{[0,t]} = \\{v_1, v_2, \\ldots, v_t\\}" + }, + { + "bbox": [ + 55, + 587, + 295, + 647 + ], + "type": "text", + "content": ". The model must decide whether it has sufficient information to generate a response. We define a decision function " + }, + { + "bbox": [ + 55, + 587, + 295, + 647 + ], + "type": "inline_equation", + "content": "\\pi" + }, + { + "bbox": [ + 55, + 587, + 295, + 647 + ], + "type": "text", + "content": " and a response function " + }, + { + "bbox": [ + 55, + 587, + 295, + 647 + ], + "type": "inline_equation", + "content": "f" + }, + { + "bbox": [ + 55, + 587, + 295, + 647 + ], + "type": "text", + "content": ":" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 99, + 658, + 252, + 672 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 99, + 658, + 252, + 672 + ], + "spans": [ + { + "bbox": [ + 99, + 658, + 252, + 672 + ], + "type": "interline_equation", + "content": "a _ {t} = \\pi \\left(C _ {t}, V _ {[ 0, t ]}\\right) \\in \\{\\text {w a i t , r e s p o n d} \\}.", + "image_path": "1c33600f075a5a52004ecf0d6da16b9783264d28c6fedf9e7d2dd5c8237a16d3.jpg" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 67, + 677, + 266, + 689 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 67, + 677, + 266, + 689 + ], + "spans": [ + { + "bbox": [ + 67, + 677, + 266, + 689 + ], + "type": "text", + "content": "If " + }, + { + "bbox": [ + 67, + 677, + 266, + 689 + ], + "type": "inline_equation", + "content": "a_{t} =" + }, + { + "bbox": [ + 67, + 677, + 266, + 689 + ], + "type": "text", + "content": " respond, the model generates a response:" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 137, + 700, + 214, + 715 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 137, + 700, + 214, + 715 + ], + "spans": [ + { + "bbox": [ + 137, + 700, + 214, + 715 + ], + "type": "interline_equation", + "content": "R _ {t} = f \\left(C _ {t}, V _ {[ 0, t ]}\\right)", + "image_path": "b645649427597f2e31c6632986d925e4be0cfba50b9180787487165eb3431b6b.jpg" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 313, + 72, + 541, + 84 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 72, + 541, + 84 + ], + "spans": [ + { + "bbox": [ + 313, + 72, + 541, + 84 + ], + "type": "text", + "content": "Otherwise, the model keeps silent for more information." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 313, + 85, + 553, + 228 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 85, + 553, + 228 + ], + "spans": [ + { + "bbox": [ + 313, + 85, + 553, + 228 + ], + "type": "text", + "content": "In this work, we implement these functions using a disentangled framework composed of three distinct modules: Perception, Decision, and Reaction. These modules are designed to handle the unique challenges of real-time video understanding and dialogue generation. The Perception module focuses on continuous video monitoring, while the Decision module assesses when to act, and the Reaction module generates responses without waiting for the entire video sequence to finish. This approach allows the system to remain responsive to new information while ensuring that the generated answers are based on the most relevant and up-to-date context." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 313, + 236, + 515, + 249 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 236, + 515, + 249 + ], + "spans": [ + { + "bbox": [ + 313, + 236, + 515, + 249 + ], + "type": "text", + "content": "3.2. Proactive Streaming Video Processing" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 313, + 254, + 553, + 373 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 254, + 553, + 373 + ], + "spans": [ + { + "bbox": [ + 313, + 254, + 553, + 373 + ], + "type": "text", + "content": "To enable active real-time response decisions, we propose a proactive streaming video processing approach that dynamically segments the video stream and evaluates whether to generate a response as the video is being processed. This approach balances computational efficiency with the need for timely and accurate responses, ensuring that the model processes the most relevant parts of the video while avoiding unnecessary redundancy. The system is composed of two key modules: the Scene-based Perception Module and the Real-time Response Decision Module." + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 313, + 376, + 553, + 460 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 376, + 553, + 460 + ], + "spans": [ + { + "bbox": [ + 313, + 376, + 553, + 460 + ], + "type": "text", + "content": "Scene-based Perception Module. To ensure efficient processing of long video streams, we begin by adaptively segmenting the video into non-uniform clips based on scene boundaries. This segmentation strategy preserves the structural information of the video, allowing the model to focus on the most informative parts while removing redundancy and maintaining context." + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 313, + 460, + 553, + 579 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 460, + 553, + 579 + ], + "spans": [ + { + "bbox": [ + 313, + 460, + 553, + 579 + ], + "type": "text", + "content": "We begin by sampling frames at a regular interval (one frame per second) and extract L2-normalized feature embeddings using the pre-trained SigLip model [63]. By computing the cosine similarity between these embeddings, we can identify significant visual changes, which indicate potential scene boundaries. These boundaries help divide the video into meaningful segments. To prevent excessively short clips, we introduce an exclusion window around the boundaries, ensuring that the resulting clips are of sufficient length to contain relevant information." + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 313, + 580, + 553, + 639 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 580, + 553, + 639 + ], + "spans": [ + { + "bbox": [ + 313, + 580, + 553, + 639 + ], + "type": "text", + "content": "Each clip " + }, + { + "bbox": [ + 313, + 580, + 553, + 639 + ], + "type": "inline_equation", + "content": "v_{i}" + }, + { + "bbox": [ + 313, + 580, + 553, + 639 + ], + "type": "text", + "content": " is then processed by the video encoder to produce the clip-wise feature representation " + }, + { + "bbox": [ + 313, + 580, + 553, + 639 + ], + "type": "inline_equation", + "content": "F_{i}" + }, + { + "bbox": [ + 313, + 580, + 553, + 639 + ], + "type": "text", + "content": ", along with a special clip indicator " + }, + { + "bbox": [ + 313, + 580, + 553, + 639 + ], + "type": "inline_equation", + "content": "\\hat{F}_{i}" + }, + { + "bbox": [ + 313, + 580, + 553, + 639 + ], + "type": "text", + "content": ". These clip features are used in the subsequent decision-making process to determine if enough information has been gathered to respond." + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 313, + 642, + 553, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 642, + 553, + 713 + ], + "spans": [ + { + "bbox": [ + 313, + 642, + 553, + 713 + ], + "type": "text", + "content": "Real-time Response Decision Module. The Real-time Response Decision Module evaluates whether the model should generate a response or continue waiting for more video content. This decision is made based on the video content observed so far and the historical context, including the user's prior questions and assistant responses. We illus" + } + ] + } + ], + "index": 17 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 318, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 318, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 318, + 757 + ], + "type": "text", + "content": "24047" + } + ] + } + ], + "index": 18 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 2 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 58, + 68, + 550, + 349 + ], + "blocks": [ + { + "bbox": [ + 58, + 68, + 550, + 349 + ], + "lines": [ + { + "bbox": [ + 58, + 68, + 550, + 349 + ], + "spans": [ + { + "bbox": [ + 58, + 68, + 550, + 349 + ], + "type": "image", + "image_path": "e900973b60bc3bc785f02a22db4251ad3402fc3ca6864d6478513fb87f0cd223.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 54, + 358, + 555, + 404 + ], + "lines": [ + { + "bbox": [ + 54, + 358, + 555, + 404 + ], + "spans": [ + { + "bbox": [ + 54, + 358, + 555, + 404 + ], + "type": "text", + "content": "Figure 2. Illustration of the Response Decision Module in a proactive streaming video processing system. The module dynamically determines when to respond during video streaming by segmenting the video into non-uniform clips and utilizing historical memory to capture context. The module then combines memory features, clip features, question text, and special tokens, " + }, + { + "bbox": [ + 54, + 358, + 555, + 404 + ], + "type": "inline_equation", + "content": "\\langle \\mathrm{TODO}\\rangle" + }, + { + "bbox": [ + 54, + 358, + 555, + 404 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 54, + 358, + 555, + 404 + ], + "type": "inline_equation", + "content": "\\langle \\mathrm{ANS}\\rangle" + }, + { + "bbox": [ + 54, + 358, + 555, + 404 + ], + "type": "text", + "content": ", to decide on response timing." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "bbox": [ + 54, + 415, + 294, + 462 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 54, + 415, + 294, + 462 + ], + "spans": [ + { + "bbox": [ + 54, + 415, + 294, + 462 + ], + "type": "text", + "content": "trate this whole process in Figure 2. To effectively combine these multi-modal inputs, we use an interleaved sequence format, which integrates video features, question information, and decision tokens." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 55, + 463, + 295, + 522 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 463, + 295, + 522 + ], + "spans": [ + { + "bbox": [ + 55, + 463, + 295, + 522 + ], + "type": "text", + "content": "The process begins with the sequence of video clip representations up to the point of the user's question, followed by the question " + }, + { + "bbox": [ + 55, + 463, + 295, + 522 + ], + "type": "inline_equation", + "content": "Q" + }, + { + "bbox": [ + 55, + 463, + 295, + 522 + ], + "type": "text", + "content": " and a special token, " + }, + { + "bbox": [ + 55, + 463, + 295, + 522 + ], + "type": "inline_equation", + "content": "\\langle \\mathrm{TODO} \\rangle" + }, + { + "bbox": [ + 55, + 463, + 295, + 522 + ], + "type": "text", + "content": ", which indicates that a response decision is yet to be made. Formally, the input sequence at this stage is:" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 118, + 530, + 233, + 544 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 118, + 530, + 233, + 544 + ], + "spans": [ + { + "bbox": [ + 118, + 530, + 233, + 544 + ], + "type": "interline_equation", + "content": "F [ 0: q _ {\\text {p o s}} ] \\oplus Q \\oplus \\langle \\text {T O D O} \\rangle ,", + "image_path": "e84b3db416c3becda38ab953c3145aaec075d4c4ca0728be6a0ad7566452e3d5.jpg" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 55, + 552, + 295, + 574 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 552, + 295, + 574 + ], + "spans": [ + { + "bbox": [ + 55, + 552, + 295, + 574 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 55, + 552, + 295, + 574 + ], + "type": "inline_equation", + "content": "q_{\\mathrm{pos}}" + }, + { + "bbox": [ + 55, + 552, + 295, + 574 + ], + "type": "text", + "content": " is the timestamp of the question, and " + }, + { + "bbox": [ + 55, + 552, + 295, + 574 + ], + "type": "inline_equation", + "content": "\\oplus" + }, + { + "bbox": [ + 55, + 552, + 295, + 574 + ], + "type": "text", + "content": " denotes concatenation." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 55, + 576, + 295, + 648 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 576, + 295, + 648 + ], + "spans": [ + { + "bbox": [ + 55, + 576, + 295, + 648 + ], + "type": "text", + "content": "After this initial stage, the model aggregates the features up to the question timestamp into a global memory " + }, + { + "bbox": [ + 55, + 576, + 295, + 648 + ], + "type": "inline_equation", + "content": "M" + }, + { + "bbox": [ + 55, + 576, + 295, + 648 + ], + "type": "text", + "content": " through pooling. Then, we extend the sequence to include this global memory that captures historical context, the question texts, and the clip features from the question timestamp to the current point, ending with " + }, + { + "bbox": [ + 55, + 576, + 295, + 648 + ], + "type": "inline_equation", + "content": "\\langle \\mathrm{TODO}\\rangle" + }, + { + "bbox": [ + 55, + 576, + 295, + 648 + ], + "type": "text", + "content": ":" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 93, + 656, + 257, + 670 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 93, + 656, + 257, + 670 + ], + "spans": [ + { + "bbox": [ + 93, + 656, + 257, + 670 + ], + "type": "interline_equation", + "content": "M [ 0: q _ {\\mathrm {p o s}} ] \\oplus Q \\oplus F [ q _ {\\mathrm {p o s}}: ] \\oplus \\langle \\mathrm {T O D O} \\rangle .", + "image_path": "2fa1e6089b20cb76afc8b094b42a1369322137755f341fec2e84fe2772dfe030.jpg" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 55, + 677, + 296, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 677, + 296, + 714 + ], + "spans": [ + { + "bbox": [ + 55, + 677, + 296, + 714 + ], + "type": "text", + "content": "When the model decides to respond, the " + }, + { + "bbox": [ + 55, + 677, + 296, + 714 + ], + "type": "inline_equation", + "content": "\\langle \\mathrm{ANS} \\rangle" + }, + { + "bbox": [ + 55, + 677, + 296, + 714 + ], + "type": "text", + "content": " token is used to mark the response action. Specifically, with the video continuing to stream, we further integrate global" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 313, + 415, + 555, + 486 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 415, + 555, + 486 + ], + "spans": [ + { + "bbox": [ + 313, + 415, + 555, + 486 + ], + "type": "text", + "content": "memory from the question timestamp to the position of the first answer, insert the special answer token " + }, + { + "bbox": [ + 313, + 415, + 555, + 486 + ], + "type": "inline_equation", + "content": "\\langle \\mathrm{ANS}\\rangle" + }, + { + "bbox": [ + 313, + 415, + 555, + 486 + ], + "type": "text", + "content": ", and include later clips for potential subsequent responses. This sequence concludes with the current clip memory segment and a " + }, + { + "bbox": [ + 313, + 415, + 555, + 486 + ], + "type": "inline_equation", + "content": "\\langle \\mathrm{TODO}\\rangle" + }, + { + "bbox": [ + 313, + 415, + 555, + 486 + ], + "type": "text", + "content": " token. Thus, the ultimate input format with multiple answers can be written as:" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 380, + 494, + 457, + 508 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 380, + 494, + 457, + 508 + ], + "spans": [ + { + "bbox": [ + 380, + 494, + 457, + 508 + ], + "type": "interline_equation", + "content": "M [ 0: q _ {\\mathrm {p o s}} ] \\oplus Q \\oplus", + "image_path": "589f9b388499c2917787312a462a7544aa803c3513cdc207c6a566f6638ab384.jpg" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 382, + 509, + 487, + 525 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 382, + 509, + 487, + 525 + ], + "spans": [ + { + "bbox": [ + 382, + 509, + 487, + 525 + ], + "type": "interline_equation", + "content": "M \\left[ q _ {\\text {p o s}}: a _ {\\text {p o s}} ^ {1} \\right] \\oplus \\langle \\mathrm {A N S} \\rangle \\oplus", + "image_path": "8b38785dc2facc2b226e704ed08fc132c3db685f75625fd6ee3eae5a43807e6a.jpg" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 382, + 525, + 487, + 541 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 382, + 525, + 487, + 541 + ], + "spans": [ + { + "bbox": [ + 382, + 525, + 487, + 541 + ], + "type": "interline_equation", + "content": "M \\left[ a _ {\\text {p o s}} ^ {1}: a _ {\\text {p o s}} ^ {2} \\right] \\oplus \\langle \\mathrm {A N S} \\rangle \\oplus", + "image_path": "3536f31f1689235b70ca737f2a409359541420fb258dc778d37f50ce692edb2f.jpg" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 384, + 546, + 397, + 553 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 384, + 546, + 397, + 553 + ], + "spans": [ + { + "bbox": [ + 384, + 546, + 397, + 553 + ], + "type": "interline_equation", + "content": "\\cdot \\cdot \\cdot", + "image_path": "cd0e049480f5e9061dcadda5631cfcf34f8d92ff35f227c9f618c72e943435e1.jpg" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 381, + 558, + 468, + 574 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 381, + 558, + 468, + 574 + ], + "spans": [ + { + "bbox": [ + 381, + 558, + 468, + 574 + ], + "type": "interline_equation", + "content": "F \\left[ a _ {\\text {p o s}} ^ {k}: \\right] \\oplus \\langle \\text {T O D O} \\rangle ,", + "image_path": "2ded6004a3ba6dbb533af16878efdb991040a69d6861209951c9b0c8b74e6430.jpg" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 313, + 582, + 541, + 594 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 582, + 541, + 594 + ], + "spans": [ + { + "bbox": [ + 313, + 582, + 541, + 594 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 313, + 582, + 541, + 594 + ], + "type": "inline_equation", + "content": "a_{\\mathrm{pos}}^k" + }, + { + "bbox": [ + 313, + 582, + 541, + 594 + ], + "type": "text", + "content": " stands for the timestamp for the " + }, + { + "bbox": [ + 313, + 582, + 541, + 594 + ], + "type": "inline_equation", + "content": "k" + }, + { + "bbox": [ + 313, + 582, + 541, + 594 + ], + "type": "text", + "content": "-th answer." + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 313, + 594, + 554, + 701 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 594, + 554, + 701 + ], + "spans": [ + { + "bbox": [ + 313, + 594, + 554, + 701 + ], + "type": "text", + "content": "We feed the interleaved sequence into a compact LLM and adopt a binary classification head on top of the final-layer embedding of the " + }, + { + "bbox": [ + 313, + 594, + 554, + 701 + ], + "type": "inline_equation", + "content": "\\langle \\mathrm{TODO}\\rangle" + }, + { + "bbox": [ + 313, + 594, + 554, + 701 + ], + "type": "text", + "content": " token. This head is trained to predict whether the model should respond at each timestep. We use standard binary cross-entropy loss to supervise the model's decision-making process. When the model determines that a response is required, the relevant historical clips are retrieved based on their cosine similarity with the " + }, + { + "bbox": [ + 313, + 594, + 554, + 701 + ], + "type": "inline_equation", + "content": "\\langle \\mathrm{TODO}\\rangle" + }, + { + "bbox": [ + 313, + 594, + 554, + 701 + ], + "type": "text", + "content": " token's embedding." + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 326, + 701, + 553, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 326, + 701, + 553, + 714 + ], + "spans": [ + { + "bbox": [ + 326, + 701, + 553, + 714 + ], + "type": "text", + "content": "To train the model's capacity on temporal retrieval, we" + } + ] + } + ], + "index": 17 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "text", + "content": "24048" + } + ] + } + ], + "index": 18 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 3 + }, + { + "para_blocks": [ + { + "bbox": [ + 55, + 72, + 295, + 131 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 72, + 295, + 131 + ], + "spans": [ + { + "bbox": [ + 55, + 72, + 295, + 131 + ], + "type": "text", + "content": "calculate the cosine similarity between the embedding of the " + }, + { + "bbox": [ + 55, + 72, + 295, + 131 + ], + "type": "inline_equation", + "content": "\\langle \\mathrm{TODO}\\rangle" + }, + { + "bbox": [ + 55, + 72, + 295, + 131 + ], + "type": "text", + "content": " token and each historical clip indicator " + }, + { + "bbox": [ + 55, + 72, + 295, + 131 + ], + "type": "inline_equation", + "content": "\\hat{F}_i" + }, + { + "bbox": [ + 55, + 72, + 295, + 131 + ], + "type": "text", + "content": ". This similarity yields a probability distribution " + }, + { + "bbox": [ + 55, + 72, + 295, + 131 + ], + "type": "inline_equation", + "content": "\\hat{P}_{\\mathrm{pred}}(i)" + }, + { + "bbox": [ + 55, + 72, + 295, + 131 + ], + "type": "text", + "content": " representing the relevance of " + }, + { + "bbox": [ + 55, + 72, + 295, + 131 + ], + "type": "inline_equation", + "content": "i" + }, + { + "bbox": [ + 55, + 72, + 295, + 131 + ], + "type": "text", + "content": "-th historical clip to the current response." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 55, + 133, + 295, + 178 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 133, + 295, + 178 + ], + "spans": [ + { + "bbox": [ + 55, + 133, + 295, + 178 + ], + "type": "text", + "content": "For the ground truth, we use related clips to create a true temporal distribution. These clips are encoded as a binary mask over the clip intervals relevant to the response, defined as:" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 91, + 178, + 259, + 203 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 91, + 178, + 259, + 203 + ], + "spans": [ + { + "bbox": [ + 91, + 178, + 259, + 203 + ], + "type": "interline_equation", + "content": "\\hat {P} _ {\\text {t r u e}} (i) = \\frac {1}{| R _ {Q} |} \\quad \\text {i f} \\quad i \\in R _ {Q} \\quad \\text {e l s e} \\quad 0,", + "image_path": "ad1401d5e998b3c4f4810da1903d852958945b9bcf298892fcca2ba30f8f1a05.jpg" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 55, + 207, + 295, + 231 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 207, + 295, + 231 + ], + "spans": [ + { + "bbox": [ + 55, + 207, + 295, + 231 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 55, + 207, + 295, + 231 + ], + "type": "inline_equation", + "content": "R_{Q}" + }, + { + "bbox": [ + 55, + 207, + 295, + 231 + ], + "type": "text", + "content": " is the ground-truth set of relevant clips for particular question " + }, + { + "bbox": [ + 55, + 207, + 295, + 231 + ], + "type": "inline_equation", + "content": "Q" + }, + { + "bbox": [ + 55, + 207, + 295, + 231 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 55, + 232, + 295, + 266 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 232, + 295, + 266 + ], + "spans": [ + { + "bbox": [ + 55, + 232, + 295, + 266 + ], + "type": "text", + "content": "We then calculate the KL divergence loss between the true distribution, " + }, + { + "bbox": [ + 55, + 232, + 295, + 266 + ], + "type": "inline_equation", + "content": "\\hat{P}_{\\mathrm{true}}" + }, + { + "bbox": [ + 55, + 232, + 295, + 266 + ], + "type": "text", + "content": ", and the predicted distribution, " + }, + { + "bbox": [ + 55, + 232, + 295, + 266 + ], + "type": "inline_equation", + "content": "\\hat{P}_{\\mathrm{pred}}" + }, + { + "bbox": [ + 55, + 232, + 295, + 266 + ], + "type": "text", + "content": ", as:" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 86, + 267, + 264, + 299 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 86, + 267, + 264, + 299 + ], + "spans": [ + { + "bbox": [ + 86, + 267, + 264, + 299 + ], + "type": "interline_equation", + "content": "\\mathcal {L} _ {\\mathrm {K L}} = - \\frac {1}{| R _ {Q} |} \\sum_ {i = 1} ^ {|} R _ {Q} | \\hat {P} _ {\\mathrm {t r u e}} (i) \\log \\hat {P} _ {\\mathrm {p r e d}} (i).", + "image_path": "8c026eb6026babe5022ca31645eb739bff6e3a46dafb15e2fa5167fbd60b9f59.jpg" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 55, + 304, + 295, + 365 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 304, + 295, + 365 + ], + "spans": [ + { + "bbox": [ + 55, + 304, + 295, + 365 + ], + "type": "text", + "content": "This KL divergence loss encourages the model to align its predictions with the true temporal relevance of the clips, thereby improving its ability to correctly retrieve and ground the relevant video segments for generating accurate, timely responses." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 55, + 372, + 271, + 384 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 372, + 271, + 384 + ], + "spans": [ + { + "bbox": [ + 55, + 372, + 271, + 384 + ], + "type": "text", + "content": "3.3. Precise Interaction with Streaming Video" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 55, + 390, + 295, + 509 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 390, + 295, + 509 + ], + "spans": [ + { + "bbox": [ + 55, + 390, + 295, + 509 + ], + "type": "text", + "content": "Once the interaction is triggered, an asynchronous precise interaction module generates a response with fine-grained detail, tailored to the specific context of the video stream. That is, for the " + }, + { + "bbox": [ + 55, + 390, + 295, + 509 + ], + "type": "inline_equation", + "content": "k" + }, + { + "bbox": [ + 55, + 390, + 295, + 509 + ], + "type": "text", + "content": "-th answer outputs, the precise interaction module receives the question, the former " + }, + { + "bbox": [ + 55, + 390, + 295, + 509 + ], + "type": "inline_equation", + "content": "k - 1" + }, + { + "bbox": [ + 55, + 390, + 295, + 509 + ], + "type": "text", + "content": " history answers as well as the grounded clip features at the current timestamp. This design allows the system to generate appropriate responses by incorporating rich context from both the video content and the ongoing conversation in a streaming setting." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 55, + 510, + 295, + 617 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 510, + 295, + 617 + ], + "spans": [ + { + "bbox": [ + 55, + 510, + 295, + 617 + ], + "type": "text", + "content": "A key advantage of this asynchronous approach is that it decouples the response generation from the video stream processing. This separation allows the streaming video to continue being processed in parallel, without waiting for response generation to complete. As a result, the system remains highly efficient and responsive, ensuring that the video content is continuously monitored and processed while still generating contextually accurate responses at the appropriate moments." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 55, + 618, + 295, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 618, + 295, + 714 + ], + "spans": [ + { + "bbox": [ + 55, + 618, + 295, + 714 + ], + "type": "text", + "content": "While the Response Decision Module may determine that a response is needed, there is a chance that no actual output is required. To ensure accuracy in real-time interactions, we enhance the model's responsiveness by training it on a streaming QA dataset that simulates such real-time scenarios. These data include both positive samples (where a response is required) and negative samples (where no response is necessary and the model should keep silent). By" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 313, + 72, + 553, + 133 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 72, + 553, + 133 + ], + "spans": [ + { + "bbox": [ + 313, + 72, + 553, + 133 + ], + "type": "text", + "content": "dynamically sampling these positive and negative instances, the model learns to either produce the responses that are incremental to the historical answers, summarizing the newly emerging cues in the video stream or output the special " + }, + { + "bbox": [ + 313, + 72, + 553, + 133 + ], + "type": "inline_equation", + "content": "\\langle \\mathrm{SILENT} \\rangle" + }, + { + "bbox": [ + 313, + 72, + 553, + 133 + ], + "type": "text", + "content": " token denoting keeping silent." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 313, + 133, + 555, + 205 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 133, + 555, + 205 + ], + "spans": [ + { + "bbox": [ + 313, + 133, + 555, + 205 + ], + "type": "text", + "content": "This adaptive reasoning approach ensures that the model remains responsive to new information while avoiding redundancy. By focusing on unanswered content and utilizing past interactions, the model delivers timely and relevant responses, enhancing the overall efficiency and user experience in streaming dialogue generation." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 313, + 217, + 395, + 230 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 217, + 395, + 230 + ], + "spans": [ + { + "bbox": [ + 313, + 217, + 395, + 230 + ], + "type": "text", + "content": "4. Experiments" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 313, + 237, + 376, + 248 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 237, + 376, + 248 + ], + "spans": [ + { + "bbox": [ + 313, + 237, + 376, + 248 + ], + "type": "text", + "content": "4.1. Datasets" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 313, + 255, + 553, + 373 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 255, + 553, + 373 + ], + "spans": [ + { + "bbox": [ + 313, + 255, + 553, + 373 + ], + "type": "text", + "content": "The training process comprises two parts: response decision with temporal grounding and LLM reasoning. For training the streaming video processor and response decision module, we combine the GroundVQA [13] and ET-Instruct [41] datasets, enriched with temporal annotations to provide supervision for streaming responses and temporal grounding labels. Additionally, to enhance the model's ability to handle questions that require simple reasoning, we curate a set of 50K implicit QA pairs with time labels, guiding the model to locate relevant cues within the video stream." + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 313, + 375, + 553, + 434 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 375, + 553, + 434 + ], + "spans": [ + { + "bbox": [ + 313, + 375, + 553, + 434 + ], + "type": "text", + "content": "For LLM reasoning, we construct a dataset of 122K streaming video QA pairs based on timestamped QA pairs from ET-Instruct [41] and integrate conventional video QA data from VideoChatGPT [43] and LLaVA-Next-Video [29]." + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 313, + 444, + 397, + 455 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 444, + 397, + 455 + ], + "spans": [ + { + "bbox": [ + 313, + 444, + 397, + 455 + ], + "type": "text", + "content": "4.2. Benchmarks" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 313, + 461, + 553, + 545 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 461, + 553, + 545 + ], + "spans": [ + { + "bbox": [ + 313, + 461, + 553, + 545 + ], + "type": "text", + "content": "For evaluation, we adopt several established long-video QA benchmarks, including EgoSchema [44], VideoMME [20], and MLVU [71]. EgoSchema consists of over 5K videos, each approximately 3 minutes long, while VideoMME and MLVU include videos of varying lengths, from a few minutes to several hours. We report accuracy on multiple-choice questions for comparison across these benchmarks." + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 313, + 546, + 555, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 546, + 555, + 713 + ], + "spans": [ + { + "bbox": [ + 313, + 546, + 555, + 713 + ], + "type": "text", + "content": "In addition, we create a streaming video QA benchmark using a subset of ETBench to measure the model's proactive response capability in real-time video interactions. Specifically, we select six subtasks from ETBench suitable for streaming response evaluation: step localization (SLC), dense video captioning (DVC), temporal action localization (TAL), temporal video grounding (TVG), episodic memory (EPM), and video highlight detection (VHD), encompassing a total of 4,460 videos. For this benchmark, the instruction is provided at the beginning of each video, and the model is required to produce the correct responses at appropriate timestamps as the video plays. We follow the evaluation metrics specified in the original benchmarks to ensure a fair comparison." + } + ] + } + ], + "index": 19 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "text", + "content": "24049" + } + ] + } + ], + "index": 20 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 4 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 122, + 70, + 487, + 261 + ], + "blocks": [ + { + "bbox": [ + 122, + 70, + 487, + 261 + ], + "lines": [ + { + "bbox": [ + 122, + 70, + 487, + 261 + ], + "spans": [ + { + "bbox": [ + 122, + 70, + 487, + 261 + ], + "type": "table", + "html": "
MethodLLM SizeFramesEgoSchemaMLVUVideoMME
Video-LLaVA [35]7B838.447.339.9
Chat-UniVi [28]7B64--40.6
LLaMA-VID [33]7B1 FPS38.533.2-
TimeChat [53]7B9633.030.930.2
MovieChat [55]7B204853.525.838.2
Video-LLaMA2 [9]7B1651.748.547.9
LLaVA-Next-Video [70]7B3243.9-46.6
ShareGPT4Video [6]8B16-46.439.9
VideoChat2 [31]7B1654.447.939.5
LongVA [69]7B128-56.352.6
Kangaroo [40]8B64-61.056.0
Video-CCAM [19]14B96-63.153.2
VideoXL [54]7B128-64.955.5
Ours7B1 FPS55.661.757.2
", + "image_path": "bd713f174a7bfb103e19fe4a6f6fdb8b823e56507214b6964d502a9b9baa9e28.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "table_body" + } + ], + "index": 0 + }, + { + "bbox": [ + 55, + 270, + 555, + 304 + ], + "lines": [ + { + "bbox": [ + 55, + 270, + 555, + 304 + ], + "spans": [ + { + "bbox": [ + 55, + 270, + 555, + 304 + ], + "type": "text", + "content": "Table 1. Comparison on long video benchmarks. We report the accuracy on the EgoSchema full set, MLVU multiple-choice questions, and the VideoMME overall set without subtitles. For a fair comparison, we also present the model size of the LLM and the number of sampled frames." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "text" + }, + { + "bbox": [ + 55, + 315, + 188, + 327 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 315, + 188, + 327 + ], + "spans": [ + { + "bbox": [ + 55, + 315, + 188, + 327 + ], + "type": "text", + "content": "4.3. Implementation Details" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 54, + 334, + 295, + 514 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 54, + 334, + 295, + 514 + ], + "spans": [ + { + "bbox": [ + 54, + 334, + 295, + 514 + ], + "type": "text", + "content": "The architecture consists of a vision encoder, a compact LLM as the proactive streaming video processor for response decisions, and a larger LLM for the precision interaction module. In practice, the input video frames are resized to " + }, + { + "bbox": [ + 54, + 334, + 295, + 514 + ], + "type": "inline_equation", + "content": "224 \\times 224" + }, + { + "bbox": [ + 54, + 334, + 295, + 514 + ], + "type": "text", + "content": ", and a CLIP-L/14 [51] is employed to extract frame-wise features. Following the token compression techniques in [48], we concatenate adjacent tokens and use the compact LLM, instantiated as Qwen2-1.5B [49], to produce time-aware compressed clip-wise features along with clip indicators. Next, we reuse this compact LLM to process the sequence consisting of global memory, question texts, and clip features for response decisions. The final LLM, instantiated as Qwen2-7B [49], receives the grounded clips and global memory to generate responses at the necessary timestamps." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 55, + 514, + 295, + 635 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 514, + 295, + 635 + ], + "spans": [ + { + "bbox": [ + 55, + 514, + 295, + 635 + ], + "type": "text", + "content": "Training is conducted in two stages. In the first stage, we train the response decision module and the video encoder. In the second stage, we freeze these components and train the final LLM. During training, we insert instructions at arbitrary timestamps within the video to enhance the model's ability to handle varied circumstances. For inference, on conventional video benchmarks, we place the question at the end of the video to ensure fair comparisons with existing works. For streaming evaluation, the question is posed at the beginning to enable proactive responses." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 55, + 647, + 243, + 659 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 647, + 243, + 659 + ], + "spans": [ + { + "bbox": [ + 55, + 647, + 243, + 659 + ], + "type": "text", + "content": "4.4. Conventional Video Understanding" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 55, + 665, + 295, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 665, + 295, + 714 + ], + "spans": [ + { + "bbox": [ + 55, + 665, + 295, + 714 + ], + "type": "text", + "content": "In this section, we compare our model with existing video LLMs on conventional video QA benchmarks, where the model is required to provide one answer after watching the complete video." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 313, + 316, + 555, + 437 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 316, + 555, + 437 + ], + "spans": [ + { + "bbox": [ + 313, + 316, + 555, + 437 + ], + "type": "text", + "content": "We first present a comparative analysis across three challenging long-video benchmarks. As shown in Table 1, we report the accuracy on the EgoSchema full set, MLVU multiple-choice questions, and the VideoMME overall set without subtitles. Generally, our disentangled architecture handles this conventional scenario well and achieves competitive performance. Notably, on EgoSchema, which requires long temporal reasoning, our method achieves a leading performance of 55.6, demonstrating strong temporal perception." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 313, + 443, + 556, + 528 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 443, + 556, + 528 + ], + "spans": [ + { + "bbox": [ + 313, + 443, + 556, + 528 + ], + "type": "text", + "content": "In terms of MLVU and VideoMME, which consist of videos ranging from minutes to hours long, our model's promising results highlight its ability to efficiently process contextual information across varying temporal lengths. This capability is crucial for streaming video interactions, as it allows the model to accurately summarize context when questions are inserted at arbitrary temporal positions." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 313, + 534, + 556, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 534, + 556, + 715 + ], + "spans": [ + { + "bbox": [ + 313, + 534, + 556, + 715 + ], + "type": "text", + "content": "Additionally, we adopt a recent time-aware benchmark, ETBench, to evaluate temporal awareness. We select six subtasks that require the model to predict explicit event timestamps and report the results in Table 2. Following [41], we report the F1 score for temporal grounding evaluation and the sentence similarity score for caption evaluation. In the conventional setting, our method is able to capture the timestamps corresponding to specific questions. Without a specialized design for time tokens, our model achieves the highest F1 score in temporal video grounding and episodic memory subtasks, even surpassing models with specialized temporal tokens. The promising performance on dense video captioning and step localization further demonstrates the model's ability in both accurate temporal grounding and precise visual perception." + } + ] + } + ], + "index": 9 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "text", + "content": "24050" + } + ] + } + ], + "index": 10 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 5 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 75, + 69, + 537, + 297 + ], + "blocks": [ + { + "bbox": [ + 75, + 69, + 537, + 297 + ], + "lines": [ + { + "bbox": [ + 75, + 69, + 537, + 297 + ], + "spans": [ + { + "bbox": [ + 75, + 69, + 537, + 297 + ], + "type": "table", + "html": "
MethodFramesTVGF1EPMF1TALF1VHDF1DVCF1DVCsimSLCF1SLCsim
Conventional video QA inference.
w/ specialized time tokens
VTG-LLM [24]9615.93.714.448.240.218.620.814.4
LITA [26]10022.24.618.023.939.717.221.012.2
ETChat [41]1 FPS38.610.230.862.538.419.724.414.6
w/o specialized time tokens
VideoChatGPT [43]1007.01.315.128.88.811.35.710.2
Video-LLaVA [35]87.01.915.028.928.015.00.98.3
LLaMA-VID [33]1 FPS5.51.28.030.027.112.65.211.1
Video-LLaMA2 [9]80.10.00.01,50.614.50.015.2
PLLAVA [60]166.91.15.728.913.310.69.711.8
VTimeLLM [25]1007.61.918.228.912.413.18.76.4
TimeChat [53]9626.23.910.140.516.612.55.69.2
Ours1 FPS43.617.229.951.531.617.814.111.7
Streaming video QA inference.
VideoLLM-Online [5]2 FPS13.23.89.122.424.013.49.910.1
Ours1 FPS36.115.527.354.233.818.918.812.4
", + "image_path": "6f854a17f657ab432b2da9ed51c16f3f14c94b74e230f0fb91d2b191fa987152.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "table_body" + } + ], + "index": 0 + }, + { + "bbox": [ + 55, + 304, + 555, + 339 + ], + "lines": [ + { + "bbox": [ + 55, + 304, + 555, + 339 + ], + "spans": [ + { + "bbox": [ + 55, + 304, + 555, + 339 + ], + "type": "text", + "content": "Table 2. Comparison on ETBench. We present the results for two different settings. In the conventional video QA setting, the model is required to answer the question after watching the entire video. In the streaming setting, the question is placed at the beginning of the video, and the model is expected to provide real-time responses. We report performance on six subtasks that are suitable for both settings." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "text" + }, + { + "bbox": [ + 55, + 348, + 230, + 361 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 348, + 230, + 361 + ], + "spans": [ + { + "bbox": [ + 55, + 348, + 230, + 361 + ], + "type": "text", + "content": "4.5. Streaming Video Understanding" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 54, + 366, + 295, + 544 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 54, + 366, + 295, + 544 + ], + "spans": [ + { + "bbox": [ + 54, + 366, + 295, + 544 + ], + "type": "text", + "content": "In addition to the conventional video QA evaluation, we also assess the performance of our model in streaming video interactions. Specifically, we place the question at the beginning of the video and expect the model to produce a response once it detects relevant clues while the video stream plays in real-time. The model should remain silent when no new information is available. In the bottom lines of Table 2, we compare our model with VideoLLM-online [5], the only existing video LLM capable of streaming video interactions. Our model consistently outperforms VideoLLM-online across all metrics, with significant advantages in temporal grounding. This demonstrates that the disentangled design makes our model have far more powerful temporal awareness, enabling more agile proactive responses aligned with specific instructions." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 54, + 545, + 295, + 653 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 54, + 545, + 295, + 653 + ], + "spans": [ + { + "bbox": [ + 54, + 545, + 295, + 653 + ], + "type": "text", + "content": "Notably, on dense video captioning and step localization tasks, our model achieves both more precise temporal grounding and more comprehensive descriptions in streaming mode than the conventional setting. This indicates that our disentangled architecture can effectively monitor the video stream in real-time and proactively generate informative responses according to the instructions. And the streaming interaction has the potential to achieve more powerful video understanding." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 54, + 654, + 296, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 54, + 654, + 296, + 714 + ], + "spans": [ + { + "bbox": [ + 54, + 654, + 296, + 714 + ], + "type": "text", + "content": "Additionally, we present a quantitative comparison between our model and VideoLLM-online in Fig. 3. From this comparison, when faced with questions requiring multi-step reasoning, our model can gradually identify the necessary clues from the video stream and generate informative" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 313, + 349, + 555, + 420 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 349, + 555, + 420 + ], + "spans": [ + { + "bbox": [ + 313, + 349, + 555, + 420 + ], + "type": "text", + "content": "answers step by step, while VideoLLM-online [5] fails to do so. For example, from the 'thirsty' in the question, our model can associate it with the drinks appearing in the video stream, and infer what actions are needed based on the context. In contrast, VideoLLM-online only gives simple descriptions of the scene or ongoing actions." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 313, + 421, + 555, + 493 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 421, + 555, + 493 + ], + "spans": [ + { + "bbox": [ + 313, + 421, + 555, + 493 + ], + "type": "text", + "content": "A key advantage of our method is the disentangled perception, decision, and reaction architecture, which enables the model to simultaneously process the streaming video input and generate responses in a non-blocking manner. However, VideoLLM-online has to experience interruptions in the video stream during answer generation." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 313, + 502, + 408, + 514 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 502, + 408, + 514 + ], + "spans": [ + { + "bbox": [ + 313, + 502, + 408, + 514 + ], + "type": "text", + "content": "4.6. Ablation Study" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 313, + 519, + 555, + 663 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 519, + 555, + 663 + ], + "spans": [ + { + "bbox": [ + 313, + 519, + 555, + 663 + ], + "type": "text", + "content": "Clip Segmentation Strategy. We first present an ablation study on clip segmentation. We compare the standard uniform clip segmentation (16 frames per clip sampled at 1 FPS) with our scene-based non-uniform segmentation. We report conventional QA accuracy on MLVU and VideoMME, as well as streaming metrics on temporal video grounding and dense video captioning on ETBench in Table 3. The scene-based segmentation yields superior results in both conventional and streaming settings. This aligns with our motivation that scene-based segmentation preserves more structural video information, which facilitates better model learning and more timely responses." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 313, + 665, + 556, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 665, + 556, + 715 + ], + "spans": [ + { + "bbox": [ + 313, + 665, + 556, + 715 + ], + "type": "text", + "content": "Special Token Design. We also ablate three special tokens in our architecture, i.e., " + }, + { + "bbox": [ + 313, + 665, + 556, + 715 + ], + "type": "inline_equation", + "content": "\\langle \\mathrm{ANS} \\rangle" + }, + { + "bbox": [ + 313, + 665, + 556, + 715 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 313, + 665, + 556, + 715 + ], + "type": "inline_equation", + "content": "\\langle \\mathrm{TODO} \\rangle" + }, + { + "bbox": [ + 313, + 665, + 556, + 715 + ], + "type": "text", + "content": " in the streaming video processor, " + }, + { + "bbox": [ + 313, + 665, + 556, + 715 + ], + "type": "inline_equation", + "content": "\\langle \\mathrm{SILENT} \\rangle" + }, + { + "bbox": [ + 313, + 665, + 556, + 715 + ], + "type": "text", + "content": " token in the final LLM. These three special tokens mainly impact the stream-" + } + ] + } + ], + "index": 10 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "text", + "content": "24051" + } + ] + } + ], + "index": 11 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 6 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 71, + 69, + 541, + 247 + ], + "blocks": [ + { + "bbox": [ + 71, + 69, + 541, + 247 + ], + "lines": [ + { + "bbox": [ + 71, + 69, + 541, + 247 + ], + "spans": [ + { + "bbox": [ + 71, + 69, + 541, + 247 + ], + "type": "image", + "image_path": "4f71b6ac993e850e18f9a7bf89d33ecf0a440e0fc85f0e25fcbc7c5644522f71.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 55, + 260, + 555, + 283 + ], + "lines": [ + { + "bbox": [ + 55, + 260, + 555, + 283 + ], + "spans": [ + { + "bbox": [ + 55, + 260, + 555, + 283 + ], + "type": "text", + "content": "Figure 3. Quantitative comparison between Dispider and VideoLLM-online in streaming video understanding. The question is inserted at the beginning, and we show the model's response in either answer texts or the state of keeping silent." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "type": "table", + "bbox": [ + 56, + 297, + 297, + 343 + ], + "blocks": [ + { + "bbox": [ + 56, + 297, + 297, + 343 + ], + "lines": [ + { + "bbox": [ + 56, + 297, + 297, + 343 + ], + "spans": [ + { + "bbox": [ + 56, + 297, + 297, + 343 + ], + "type": "table", + "html": "
ClipMLVUV-MMETVG\\( _{F1} \\)DVC\\( _{F1} \\)DVC\\( _{Sim} \\)
Uniform59.855.434.533.118.1
Scene-based61.757.236.133.818.9
", + "image_path": "3f2aa94061b82804b44d4e0ed489a5bc6c7afdb7869a03883bdebb4a6248ae62.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "table_body" + } + ], + "index": 2 + }, + { + "type": "table", + "bbox": [ + 58, + 391, + 293, + 480 + ], + "blocks": [ + { + "bbox": [ + 55, + 351, + 295, + 384 + ], + "lines": [ + { + "bbox": [ + 55, + 351, + 295, + 384 + ], + "spans": [ + { + "bbox": [ + 55, + 351, + 295, + 384 + ], + "type": "text", + "content": "Table 3. Ablation study on the clip segmentation. We compare uniform 16-frame clip segmentation and our scene-based segmentation with SigLip." + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 58, + 391, + 293, + 480 + ], + "lines": [ + { + "bbox": [ + 58, + 391, + 293, + 480 + ], + "spans": [ + { + "bbox": [ + 58, + 391, + 293, + 480 + ], + "type": "table", + "html": "
<ANS><TODO><SILENT>TVG\\( _{F1} \\)DVC\\( _{F1} \\)DVC\\( _{Sim} \\)
XXX20.119.712.3
XX✓26.324.913.1
✓X✓35.231.017.2
X✓✓28.725.614.5
✓✓X35.530.216.8
✓✓✓36.133.818.9
", + "image_path": "d243d7ab0d36239c4fd3015321423d252696386104e0e69ea6709879eeb8735c.jpg" + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "table_body" + } + ], + "index": 4 + }, + { + "bbox": [ + 55, + 489, + 295, + 522 + ], + "lines": [ + { + "bbox": [ + 55, + 489, + 295, + 522 + ], + "spans": [ + { + "bbox": [ + 55, + 489, + 295, + 522 + ], + "type": "text", + "content": "Table 4. Ablation study on the special token designs. We respectively explore the effectiveness of ⟨ANS⟩ and ⟨TODO⟩ in streaming video processor as well as ⟨SILENT⟩ in the final LLM." + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "text" + }, + { + "bbox": [ + 55, + 534, + 296, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 534, + 296, + 713 + ], + "spans": [ + { + "bbox": [ + 55, + 534, + 296, + 713 + ], + "type": "text", + "content": "ing video interactions, so we report the streaming evaluation metrics on temporal video grounding and dense video captioning in Table 4. There are three observations. First, the absence of " + }, + { + "bbox": [ + 55, + 534, + 296, + 713 + ], + "type": "inline_equation", + "content": "\\langle \\mathrm{ANS}\\rangle" + }, + { + "bbox": [ + 55, + 534, + 296, + 713 + ], + "type": "text", + "content": " token can lead the model to be unaware of the timestamps at which a response has been given. As a result, if relevant clues have appeared in the video contexts, the model tends to produce a response, resulting in a high recall but a low accuracy. Second, the " + }, + { + "bbox": [ + 55, + 534, + 296, + 713 + ], + "type": "inline_equation", + "content": "\\langle \\mathrm{TODO}\\rangle" + }, + { + "bbox": [ + 55, + 534, + 296, + 713 + ], + "type": "text", + "content": " serves as an indicator that reminds the streaming processor to decide whether to respond. The performance slightly degrades without this special token. Third, the " + }, + { + "bbox": [ + 55, + 534, + 296, + 713 + ], + "type": "inline_equation", + "content": "\\langle \\mathrm{SILENT}\\rangle" + }, + { + "bbox": [ + 55, + 534, + 296, + 713 + ], + "type": "text", + "content": " token in the final LLM serves as a secondary filter for response decision. If the preceding streaming processor incorrectly identifies a timestamp as requiring a response, the " + }, + { + "bbox": [ + 55, + 534, + 296, + 713 + ], + "type": "inline_equation", + "content": "\\langle \\mathrm{SILENT}\\rangle" + }, + { + "bbox": [ + 55, + 534, + 296, + 713 + ], + "type": "text", + "content": " token enables the LLM to rethink whether an answer is needed" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 313, + 299, + 517, + 312 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 299, + 517, + 312 + ], + "spans": [ + { + "bbox": [ + 313, + 299, + 517, + 312 + ], + "type": "text", + "content": "referring to the historical QA interaction contexts." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 313, + 322, + 388, + 335 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 322, + 388, + 335 + ], + "spans": [ + { + "bbox": [ + 313, + 322, + 388, + 335 + ], + "type": "text", + "content": "5. Conclusion" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 313, + 342, + 555, + 594 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 342, + 555, + 594 + ], + "spans": [ + { + "bbox": [ + 313, + 342, + 555, + 594 + ], + "type": "text", + "content": "In this work, we introduced Dispider, a novel framework designed to enable active real-time interaction with video LLMs. By disentangling the key capabilities of perception, decision, and reaction, and adopting an asynchronous processing approach, Dispider overcomes the inherent conflicts that hinder real-time interaction in traditional video LLMs. Our approach ensures that the model can continuously process the video stream while providing timely, contextually accurate, and precise responses to user interactions. The Proactive Streaming Video Processing module optimizes the video analysis by focusing on relevant content, while the Precise Interaction module generates detailed responses asynchronously, allowing for uninterrupted video processing. We demonstrated the effectiveness of Dispider through extensive experiments on both conventional and streaming video QA benchmarks, where it outperformed existing methods in proactive response capabilities, temporal awareness, and computational efficiency. Dispider's disentangled architecture and ability to handle long-duration video streams make it an ideal solution for real-time interactive applications." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 313, + 604, + 411, + 617 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 604, + 411, + 617 + ], + "spans": [ + { + "bbox": [ + 313, + 604, + 411, + 617 + ], + "type": "text", + "content": "Acknowledgement" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 313, + 624, + 555, + 707 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 624, + 555, + 707 + ], + "spans": [ + { + "bbox": [ + 313, + 624, + 555, + 707 + ], + "type": "text", + "content": "This project is funded in part by Shanghai Artificial Intelligence Laboratory, Shanghai Innovation Institute, the National Key R&D Program of China (2022ZD0160201), the Centre for Perceptual and Interactive Intelligence (CPII) Ltd under the Innovation and Technology Commission (ITC)'s InnoHK. Dahua Lin is a PI of CPII under the InnoHK." + } + ] + } + ], + "index": 11 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "text", + "content": "24052" + } + ] + } + ], + "index": 12 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 7 + }, + { + "para_blocks": [ + { + "bbox": [ + 56, + 71, + 115, + 83 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 71, + 115, + 83 + ], + "spans": [ + { + "bbox": [ + 56, + 71, + 115, + 83 + ], + "type": "text", + "content": "References" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 57, + 91, + 296, + 712 + ], + "type": "list", + "angle": 0, + "index": 13, + "blocks": [ + { + "bbox": [ + 61, + 91, + 296, + 156 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 61, + 91, + 296, + 156 + ], + "spans": [ + { + "bbox": [ + 61, + 91, + 296, + 156 + ], + "type": "text", + "content": "[1] Jean-Baptiste Alayrac, Jeff Donahue, Pauline Luc, Antoine Miech, Iain Barr, Yana Hasson, Karel Lenc, Arthur Mensch, Katherine Millican, Malcolm Reynolds, et al. Flamingo: a visual language model for few-shot learning. Advances in neural information processing systems, 35:23716-23736, 2022. 2" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 61, + 158, + 296, + 201 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 61, + 158, + 296, + 201 + ], + "spans": [ + { + "bbox": [ + 61, + 158, + 296, + 201 + ], + "type": "text", + "content": "[2] Jinze Bai, Shuai Bai, Shusheng Yang, Shijie Wang, Sinan Tan, Peng Wang, Junyang Lin, Chang Zhou, and Jingren Zhou. Qwen-vl: A frontier large vision-language model with versatile abilities. arXiv preprint arXiv:2308.12966, 2023. 2" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 62, + 203, + 295, + 257 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 203, + 295, + 257 + ], + "spans": [ + { + "bbox": [ + 62, + 203, + 295, + 257 + ], + "type": "text", + "content": "[3] Yuntao Bai, Andy Jones, Kamal Ndousse, Amanda Askell, Anna Chen, Nova DasSarma, Dawn Drain, Stanislav Fort, Deep Ganguli, Tom Henighan, et al. Training a helpful and harmless assistant with reinforcement learning from human feedback. arXiv preprint arXiv:2204.05862, 2022." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 62, + 258, + 295, + 300 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 258, + 295, + 300 + ], + "spans": [ + { + "bbox": [ + 62, + 258, + 295, + 300 + ], + "type": "text", + "content": "[4] Zheng Cai, Maosong Cao, Haojiong Chen, Kai Chen, Keyu Chen, Xin Chen, Xun Chen, Zehui Chen, Zhi Chen, Pei Chu, et al. Internl m2 technical report. arXiv preprint arXiv:2403.17297, 2024. 2" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 62, + 302, + 294, + 356 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 302, + 294, + 356 + ], + "spans": [ + { + "bbox": [ + 62, + 302, + 294, + 356 + ], + "type": "text", + "content": "[5] Joya Chen, Zhaoyang Lv, Shiwei Wu, Kevin Qinghong Lin, Chenan Song, Difei Gao, Jia-Wei Liu, Ziteng Gao, Dongxing Mao, and Mike Zheng Shou. Videollm-online: Online video large language model for streaming video. In CVPR, 2024. 1, 2, 3, 7" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 62, + 358, + 295, + 412 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 358, + 295, + 412 + ], + "spans": [ + { + "bbox": [ + 62, + 358, + 295, + 412 + ], + "type": "text", + "content": "[6] Lin Chen, Xilin Wei, Jinsong Li, Xiaoyi Dong, Pan Zhang, Yuhang Zang, Zehui Chen, Haodong Duan, Bin Lin, Zhenyu Tang, et al. Sharegpt4video: Improving video understanding and generation with better captions. arXiv preprint arXiv:2406.04325, 2024. 2, 6" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 62, + 414, + 294, + 456 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 414, + 294, + 456 + ], + "spans": [ + { + "bbox": [ + 62, + 414, + 294, + 456 + ], + "type": "text", + "content": "[7] Ling-Hao Chen, Shunlin Lu, Ailing Zeng, Hao Zhang, Benyou Wang, Ruimao Zhang, and Lei Zhang. Motionllm: Understanding human behaviors from human motions and videos. arXiv preprint arXiv:2405.20340, 2024." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 62, + 459, + 294, + 523 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 459, + 294, + 523 + ], + "spans": [ + { + "bbox": [ + 62, + 459, + 294, + 523 + ], + "type": "text", + "content": "[8] Zhe Chen, Jiannan Wu, Wenhai Wang, Weijie Su, Guo Chen, Sen Xing, Muyan Zhong, Qinglong Zhang, Xizhou Zhu, Lewei Lu, Bin Li, Ping Luo, Tong Lu, Yu Qiao, and Jifeng Dai. Internvl: Scaling up vision foundation models and aligning for generic visual-linguistic tasks. arXiv preprint arXiv:2312.14238, 2023." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 62, + 525, + 294, + 579 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 525, + 294, + 579 + ], + "spans": [ + { + "bbox": [ + 62, + 525, + 294, + 579 + ], + "type": "text", + "content": "[9] Zesen Cheng, Sicong Leng, Hang Zhang, Yifei Xin, Xin Li, Guanzheng Chen, Yongxin Zhu, Wenqi Zhang, Ziyang Luo, Deli Zhao, et al. Videollama 2: Advancing spatial-temporal modeling and audio understanding in video-llms. arXiv preprint arXiv:2406.07476, 2024. 2, 6, 7" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 57, + 581, + 294, + 612 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 581, + 294, + 612 + ], + "spans": [ + { + "bbox": [ + 57, + 581, + 294, + 612 + ], + "type": "text", + "content": "[10] Aakanksha Chowdhery et al. Palm: Scaling language modeling with pathways. arXiv preprint arXiv:2303.12345, 2023. 2" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 57, + 614, + 294, + 667 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 614, + 294, + 667 + ], + "spans": [ + { + "bbox": [ + 57, + 614, + 294, + 667 + ], + "type": "text", + "content": "[11] Wenliang Dai, Junnan Li, Dongxu Li, Anthony Meng Huat Tiong, Junqi Zhao, Weisheng Wang, Boyang Li, Pascale Fung, and Steven Hoi. Instructclip: Towards general-purpose vision-language models with instruction tuning, 2023. 2" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 57, + 670, + 294, + 712 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 670, + 294, + 712 + ], + "spans": [ + { + "bbox": [ + 57, + 670, + 294, + 712 + ], + "type": "text", + "content": "[12] Jacob Devlin, Ming-Wei Chang, Kenton Lee, and Kristina Toutanova. Bert: Pre-training of deep bidirectional transformers for language understanding. arXiv preprint arXiv:1810.04805, 2018. 2" + } + ] + } + ], + "index": 12 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 316, + 73, + 553, + 712 + ], + "type": "list", + "angle": 0, + "index": 27, + "blocks": [ + { + "bbox": [ + 316, + 73, + 553, + 116 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 73, + 553, + 116 + ], + "spans": [ + { + "bbox": [ + 316, + 73, + 553, + 116 + ], + "type": "text", + "content": "[13] Shangzhe Di and Weidi Xie. Grounded question-answering in long egocentric videos. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 12934–12943, 2024. 5" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 317, + 119, + 553, + 184 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 119, + 553, + 184 + ], + "spans": [ + { + "bbox": [ + 317, + 119, + 553, + 184 + ], + "type": "text", + "content": "[14] Shuangrui Ding, Maomao Li, Tianyu Yang, Rui Qian, Haohang Xu, Qingyi Chen, Jue Wang, and Hongkai Xiong. Motion-aware contrastive video representation learning via foreground-background merging. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 9716-9726, 2022. 1" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 317, + 186, + 553, + 239 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 186, + 553, + 239 + ], + "spans": [ + { + "bbox": [ + 317, + 186, + 553, + 239 + ], + "type": "text", + "content": "[15] Shuangrui Ding, Peisen Zhao, Xiaopeng Zhang, Rui Qian, Hongkai Xiong, and Qi Tian. Prune spatio-temporal tokens by semantic-aware temporal accumulation. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 16945-16956, 2023. 1" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 317, + 242, + 553, + 295 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 242, + 553, + 295 + ], + "spans": [ + { + "bbox": [ + 317, + 242, + 553, + 295 + ], + "type": "text", + "content": "[16] Shuangrui Ding, Zihan Liu, Xiaoyi Dong, Pan Zhang, Rui Qian, Conghui He, Dahua Lin, and Jiaqi Wang. Songcomposer: A large language model for lyric and melody composition in song generation. arXiv preprint arXiv:2402.17645, 2024. 2" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 317, + 297, + 553, + 352 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 297, + 553, + 352 + ], + "spans": [ + { + "bbox": [ + 317, + 297, + 553, + 352 + ], + "type": "text", + "content": "[17] Shuangrui Ding, Rui Qian, Haohang Xu, Dahua Lin, and Hongkai Xiong. Betrayed by attention: A simple yet effective approach for self-supervised video object segmentation. In European Conference on Computer Vision, pages 215-233. Springer, 2024. 1" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 317, + 354, + 553, + 396 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 354, + 553, + 396 + ], + "spans": [ + { + "bbox": [ + 317, + 354, + 553, + 396 + ], + "type": "text", + "content": "[18] Yue Fan, Xiaojian Ma, Rujie Wu, Yuntao Du, Jiaqi Li, Zhi Gao, and Qing Li. Videoagent: A memory-augmented multimodal agent for video understanding. arXiv preprint arXiv:2403.11481, 2024. 2" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 317, + 399, + 553, + 442 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 399, + 553, + 442 + ], + "spans": [ + { + "bbox": [ + 317, + 399, + 553, + 442 + ], + "type": "text", + "content": "[19] Jiajun Fei, Dian Li, Zhidong Deng, Zekun Wang, Gang Liu, and Hui Wang. Video-ccam: Enhancing video-language understanding with causal cross-attention masks for short and long videos. arXiv preprint arXiv:2408.14023, 2024.6" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 317, + 445, + 553, + 498 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 445, + 553, + 498 + ], + "spans": [ + { + "bbox": [ + 317, + 445, + 553, + 498 + ], + "type": "text", + "content": "[20] Chaoyou Fu, Yuhan Dai, Yondong Luo, Lei Li, Shuhuai Ren, Renrui Zhang, Zihan Wang, Chenyu Zhou, Yunhang Shen, Mengdan Zhang, et al. Video-mme: The first-ever comprehensive evaluation benchmark of multi-modal llms in video analysis. arXiv preprint arXiv:2405.21075, 2024. 5" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 317, + 501, + 553, + 555 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 501, + 553, + 555 + ], + "spans": [ + { + "bbox": [ + 317, + 501, + 553, + 555 + ], + "type": "text", + "content": "[21] Chaoyou Fu, Haojia Lin, Zuwei Long, Yunhang Shen, Meng Zhao, Yifan Zhang, Xiong Wang, Di Yin, Long Ma, Xiawu Zheng, Ran He, Rongrong Ji, Yunsheng Wu, Caifeng Shan, and Xing Sun. Vita: Towards open-source interactive omni multimodal lIm. arXiv preprint arXiv:2408.05211, 2024. 2" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 317, + 556, + 553, + 587 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 556, + 553, + 587 + ], + "spans": [ + { + "bbox": [ + 317, + 556, + 553, + 587 + ], + "type": "text", + "content": "[22] Gemini Team Google. Gemini: a family of highly capable multimodal models. arXiv preprint arXiv:2312.11805, 2023. 2" + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 317, + 590, + 553, + 623 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 590, + 553, + 623 + ], + "spans": [ + { + "bbox": [ + 317, + 590, + 553, + 623 + ], + "type": "text", + "content": "[23] Gemini Team Google. Gemini 1.5: Unlocking multimodal understanding across millions of tokens of context. arXiv preprint arXiv:2403.05530, 2024. 2" + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 317, + 624, + 553, + 668 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 624, + 553, + 668 + ], + "spans": [ + { + "bbox": [ + 317, + 624, + 553, + 668 + ], + "type": "text", + "content": "[24] Yongxin Guo, Jingyu Liu, Mingda Li, Xiaoying Tang, Xi Chen, and Bo Zhao. Vtg-llm: Integrating timestamp knowledge into video llms for enhanced video temporal grounding. arXiv preprint arXiv:2405.13382, 2024. 7" + } + ] + } + ], + "index": 25 + }, + { + "bbox": [ + 317, + 670, + 553, + 712 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 670, + 553, + 712 + ], + "spans": [ + { + "bbox": [ + 317, + 670, + 553, + 712 + ], + "type": "text", + "content": "[25] Bin Huang, Xin Wang, Hong Chen, Zihan Song, and Wenwu Zhu. Vtimellm: Empower llm to grasp video moments. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 14271-14280, 2024. 7" + } + ] + } + ], + "index": 26 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "text", + "content": "24053" + } + ] + } + ], + "index": 28 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 8 + }, + { + "para_blocks": [ + { + "bbox": [ + 56, + 72, + 294, + 713 + ], + "type": "list", + "angle": 0, + "index": 15, + "blocks": [ + { + "bbox": [ + 56, + 72, + 294, + 127 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 72, + 294, + 127 + ], + "spans": [ + { + "bbox": [ + 56, + 72, + 294, + 127 + ], + "type": "text", + "content": "[26] De-An Huang, Shijia Liao, Subhashree Radhakrishnan, Hongxu Yin, Pavlo Molchanov, Zhiding Yu, and Jan Kautz. Lita: Language instructed temporal-localization assistant. In European Conference on Computer Vision, pages 202-218. Springer, 2025. 7" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 56, + 128, + 294, + 183 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 128, + 294, + 183 + ], + "spans": [ + { + "bbox": [ + 56, + 128, + 294, + 183 + ], + "type": "text", + "content": "[27] Albert Q Jiang, Alexandre Sablayrolles, Arthur Mensch, Chris Bamford, Devendra Singh Chaplot, Diego de las Casas, Florian Bressand, Gianna Lengyel, Guillaume Lample, Lucile Saulnier, et al. Mistral 7b. arXiv preprint arXiv:2310.06825, 2023. 2" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 56, + 184, + 294, + 227 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 184, + 294, + 227 + ], + "spans": [ + { + "bbox": [ + 56, + 184, + 294, + 227 + ], + "type": "text", + "content": "[28] Peng Jin, Ryuichi Takanobu, Caiwan Zhang, Xiaochun Cao, and Li Yuan. Chat-univi: Unified visual representation empowers large language models with image and video understanding. arXiv preprint arXiv:2311.08046, 2023. 6" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 56, + 228, + 294, + 271 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 228, + 294, + 271 + ], + "spans": [ + { + "bbox": [ + 56, + 228, + 294, + 271 + ], + "type": "text", + "content": "[29] Bo Li, Yuanhan Zhang, Dong Guo, Renrui Zhang, Feng Li, Hao Zhang, Kaichen Zhang, Yanwei Li, Ziwei Liu, and Chunyuan Li. Llava-onevision: Easy visual task transfer. arXiv preprint arXiv:2408.03326, 2024. 5" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 56, + 272, + 294, + 315 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 272, + 294, + 315 + ], + "spans": [ + { + "bbox": [ + 56, + 272, + 294, + 315 + ], + "type": "text", + "content": "[30] KunChang Li, Yinan He, Yi Wang, Yizhuo Li, Wenhai Wang, Ping Luo, Yali Wang, Limin Wang, and Yu Qiao. Videochat: Chat-centric video understanding. arXiv preprint arXiv:2305.06355, 2023. 2" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 56, + 316, + 294, + 380 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 316, + 294, + 380 + ], + "spans": [ + { + "bbox": [ + 56, + 316, + 294, + 380 + ], + "type": "text", + "content": "[31] Kunchang Li, Yali Wang, Yinan He, Yizhuo Li, Yi Wang, Yi Liu, Zun Wang, Jilan Xu, Guo Chen, Ping Luo, et al. Mvbench: A comprehensive multi-modal video understanding benchmark. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 22195-22206, 2024. 6" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 56, + 382, + 294, + 415 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 382, + 294, + 415 + ], + "spans": [ + { + "bbox": [ + 56, + 382, + 294, + 415 + ], + "type": "text", + "content": "[32] Yanwei Li, Chengyao Wang, and Jiaya Jia. Llama-vid: An image is worth 2 tokens in large language models. arXiv preprint arXiv:2311.17043, 2023. 1, 2" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 56, + 415, + 294, + 446 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 415, + 294, + 446 + ], + "spans": [ + { + "bbox": [ + 56, + 415, + 294, + 446 + ], + "type": "text", + "content": "[33] Yanwei Li, Chengyao Wang, and Jiaya Jia. Llama-vid: An image is worth 2 tokens in large language models. 2024. 2, 6, 7" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 56, + 449, + 294, + 502 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 449, + 294, + 502 + ], + "spans": [ + { + "bbox": [ + 56, + 449, + 294, + 502 + ], + "type": "text", + "content": "[34] Yifei Li, Junbo Niu, Ziyang Miao, Chunjiang Ge, Yuanhang Zhou, Qihao He, Xiaoyi Dong, Haodong Duan, Shuangrui Ding, Rui Qian, et al. Ovo-bench: How far is your videolms from real-world online video understanding? arXiv preprint arXiv:2501.05510, 2025. 3" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 56, + 504, + 294, + 546 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 504, + 294, + 546 + ], + "spans": [ + { + "bbox": [ + 56, + 504, + 294, + 546 + ], + "type": "text", + "content": "[35] Bin Lin, Bin Zhu, Yang Ye, Munan Ning, Peng Jin, and Li Yuan. Video-llava: Learning united visual representation by alignment before projection. arXiv preprint arXiv:2311.10122, 2023. 6, 7" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 56, + 548, + 294, + 602 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 548, + 294, + 602 + ], + "spans": [ + { + "bbox": [ + 56, + 548, + 294, + 602 + ], + "type": "text", + "content": "[36] Kevin Lin, Faisal Ahmed, Linjie Li, Chung-Ching Lin, Ehsan Azarnasab, Zhengyuan Yang, Jianfeng Wang, Lin Liang, Zicheng Liu, Yumao Lu, et al. Mm-vid: Advancing video understanding with gpt-4v (ision). arXiv preprint arXiv:2310.19773, 2023. 2" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 56, + 603, + 294, + 635 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 603, + 294, + 635 + ], + "spans": [ + { + "bbox": [ + 56, + 603, + 294, + 635 + ], + "type": "text", + "content": "[37] Haotian Liu, Chunyuan Li, Yuheng Li, and Yong Jae Lee. Improved baselines with visual instruction tuning. arXiv preprint arXiv:2310.03744, 2023. 2" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 56, + 636, + 294, + 658 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 636, + 294, + 658 + ], + "spans": [ + { + "bbox": [ + 56, + 636, + 294, + 658 + ], + "type": "text", + "content": "[38] Haotian Liu, Chunyuan Li, Qingyang Wu, and Yong Jae Lee. Visual instruction tuning, 2023." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 56, + 658, + 294, + 691 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 658, + 294, + 691 + ], + "spans": [ + { + "bbox": [ + 56, + 658, + 294, + 691 + ], + "type": "text", + "content": "[39] Haotian Liu, Chunyuan Li, Yuheng Li, Bo Li, Yuanhan Zhang, Sheng Shen, and Yong Jae Lee. Llava next: Improved reasoning,OCR, and world knowledge, 2024. 2" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 56, + 692, + 294, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 692, + 294, + 713 + ], + "spans": [ + { + "bbox": [ + 56, + 692, + 294, + 713 + ], + "type": "text", + "content": "[40] Jiajun Liu, Yibing Wang, Hanghang Ma, Xiaoping Wu, Xiaqi Ma, Xiaoming Wei, Jianbin Jiao, Enhua Wu, and Jie Hu." + } + ] + } + ], + "index": 14 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 316, + 73, + 553, + 713 + ], + "type": "list", + "angle": 0, + "index": 31, + "blocks": [ + { + "bbox": [ + 333, + 73, + 553, + 105 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 333, + 73, + 553, + 105 + ], + "spans": [ + { + "bbox": [ + 333, + 73, + 553, + 105 + ], + "type": "text", + "content": "Kangaroo: A powerful video-language model supporting long-context video input. arXiv preprint arXiv:2408.15542, 2024.6" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 316, + 107, + 553, + 150 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 107, + 553, + 150 + ], + "spans": [ + { + "bbox": [ + 316, + 107, + 553, + 150 + ], + "type": "text", + "content": "[41] Ye Liu, Zongyang Ma, Zhongang Qi, Yang Wu, Chang Wen Chen, and Ying Shan. E.t. bench: Towards open-ended event-level video-language understanding. In Neural Information Processing Systems (NeurIPS), 2024. 5, 6, 7" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 316, + 152, + 553, + 207 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 152, + 553, + 207 + ], + "spans": [ + { + "bbox": [ + 316, + 152, + 553, + 207 + ], + "type": "text", + "content": "[42] Xudong Lu, Yinghao Chen, Cheng Chen, Hui Tan, Boheng Chen, Yina Xie, Rui Hu, Guanxin Tan, Renshou Wu, Yan Hu, et al. Bluem-v-3b: Algorithm and system co-design for multimodal large language models on mobile devices. arXiv preprint arXiv:2411.10640, 2024. 2" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 316, + 208, + 553, + 251 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 208, + 553, + 251 + ], + "spans": [ + { + "bbox": [ + 316, + 208, + 553, + 251 + ], + "type": "text", + "content": "[43] Muhammad Maaz, Hanoona Rasheed, Salman Khan, and Fahad Shahbaz Khan. Video-chatgpt: Towards detailed video understanding via large vision and language models. arXiv preprint arXiv:2306.05424, 2023. 2, 5, 7" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 316, + 253, + 553, + 297 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 253, + 553, + 297 + ], + "spans": [ + { + "bbox": [ + 316, + 253, + 553, + 297 + ], + "type": "text", + "content": "[44] Karttikeya Mangalam, Raiymbek Akshulakov, and Jitendra Malik. Egoschema: A diagnostic benchmark for very long-form video language understanding. Advances in Neural Information Processing Systems, 36, 2024. 5" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 316, + 298, + 484, + 308 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 298, + 484, + 308 + ], + "spans": [ + { + "bbox": [ + 316, + 298, + 484, + 308 + ], + "type": "text", + "content": "[45] OpenAI. Gpt-4 technical report. 2023. 2" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 316, + 310, + 553, + 374 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 310, + 553, + 374 + ], + "spans": [ + { + "bbox": [ + 316, + 310, + 553, + 374 + ], + "type": "text", + "content": "[46] Long Ouyang, Jeffrey Wu, Xu Jiang, Diogo Almeida, Carroll Wainwright, Pamela Mishkin, Chong Zhang, Sandhini Agarwal, Katarina Slama, Alex Ray, et al. Training language models to follow instructions with human feedback. Advances in neural information processing systems, 35:27730-27744, 2022. 2" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 316, + 376, + 553, + 421 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 376, + 553, + 421 + ], + "spans": [ + { + "bbox": [ + 316, + 376, + 553, + 421 + ], + "type": "text", + "content": "[47] Rui Qian, Shuangrui Ding, and Dahua Lin. Rethinking image-to-video adaptation: An object-centric perspective. In European Conference on Computer Vision, pages 329-348. Springer, 2024. 1" + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 316, + 422, + 553, + 476 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 422, + 553, + 476 + ], + "spans": [ + { + "bbox": [ + 316, + 422, + 553, + 476 + ], + "type": "text", + "content": "[48] Rui Qian, Xiaoyi Dong, Pan Zhang, Yuhang Zang, Shuangrui Ding, Dahua Lin, and Jiaqi Wang. Streaming long video understanding with large language models. In The Thirty-eighth Annual Conference on Neural Information Processing Systems, 2024. 3, 6" + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 316, + 478, + 502, + 488 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 478, + 502, + 488 + ], + "spans": [ + { + "bbox": [ + 316, + 478, + 502, + 488 + ], + "type": "text", + "content": "[49] Team Qwen. Qwen2 technical report, 2024. 6" + } + ] + } + ], + "index": 25 + }, + { + "bbox": [ + 316, + 490, + 553, + 522 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 490, + 553, + 522 + ], + "spans": [ + { + "bbox": [ + 316, + 490, + 553, + 522 + ], + "type": "text", + "content": "[50] Alec Radford, Karthik Narasimhan, Tim Salimans, and Ilya Sutskever. Improving language understanding by generative pre-training. 2018. 2" + } + ] + } + ], + "index": 26 + }, + { + "bbox": [ + 316, + 523, + 553, + 588 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 523, + 553, + 588 + ], + "spans": [ + { + "bbox": [ + 316, + 523, + 553, + 588 + ], + "type": "text", + "content": "[51] Alec Radford, Jong Wook Kim, Chris Hallacy, Aditya Ramesh, Gabriel Goh, Sandhini Agarwal, Girish Sastry, Amanda Askell, Pamela Mishkin, Jack Clark, et al. Learning transferable visual models from natural language supervision. In International conference on machine learning, pages 8748-8763, 2021. 6" + } + ] + } + ], + "index": 27 + }, + { + "bbox": [ + 316, + 590, + 553, + 634 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 590, + 553, + 634 + ], + "spans": [ + { + "bbox": [ + 316, + 590, + 553, + 634 + ], + "type": "text", + "content": "[52] Colin Raffel, Noam Shazeer, Adam Roberts, Katherine Lee, Sharan Narang, Michael Matena, Yanqi Zhou, Wei Li, and Peter J Liu. Exploring the limits of transfer learning with a unified text-to-text transformer. JMLR, 2020. 2" + } + ] + } + ], + "index": 28 + }, + { + "bbox": [ + 316, + 635, + 553, + 689 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 635, + 553, + 689 + ], + "spans": [ + { + "bbox": [ + 316, + 635, + 553, + 689 + ], + "type": "text", + "content": "[53] Shuhuai Ren, Linli Yao, Shicheng Li, Xu Sun, and Lu Hou. Timechat: A time-sensitive multimodal large language model for long video understanding. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 14313-14323, 2024. 2, 6, 7" + } + ] + } + ], + "index": 29 + }, + { + "bbox": [ + 316, + 691, + 553, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 691, + 553, + 713 + ], + "spans": [ + { + "bbox": [ + 316, + 691, + 553, + 713 + ], + "type": "text", + "content": "[54] Yan Shu, Peitian Zhang, Zheng Liu, Minghao Qin, Junjie Zhou, Tiejun Huang, and Bo Zhao. Video-xl: Extra-long" + } + ] + } + ], + "index": 30 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 318, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 318, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 318, + 757 + ], + "type": "text", + "content": "24054" + } + ] + } + ], + "index": 32 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 9 + }, + { + "para_blocks": [ + { + "bbox": [ + 56, + 73, + 295, + 713 + ], + "type": "list", + "angle": 0, + "index": 13, + "blocks": [ + { + "bbox": [ + 76, + 73, + 294, + 95 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 76, + 73, + 294, + 95 + ], + "spans": [ + { + "bbox": [ + 76, + 73, + 294, + 95 + ], + "type": "text", + "content": "vision language model for hour-scale video understanding. arXiv preprint arXiv:2409.14485, 2024. 6" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 56, + 96, + 295, + 162 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 96, + 295, + 162 + ], + "spans": [ + { + "bbox": [ + 56, + 96, + 295, + 162 + ], + "type": "text", + "content": "[55] Enxin Song, Wenhao Chai, Guanhong Wang, Yucheng Zhang, Haoyang Zhou, Feiyang Wu, Haozhe Chi, Xun Guo, Tian Ye, Yanting Zhang, et al. Moviechat: From dense token to sparse memory for long video understanding. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 18221-18232, 2024. 2, 6" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 56, + 162, + 294, + 216 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 162, + 294, + 216 + ], + "spans": [ + { + "bbox": [ + 56, + 162, + 294, + 216 + ], + "type": "text", + "content": "[56] Hugo Touvron, Thibaut Lavril, Gautier Izacard, Xavier Martinet, Marie-Anne Lachaux, Timothée Lacroix, Baptiste Rozière, Naman Goyal, Eric Hambro, Faisal Azhar, et al. Llama: Open and efficient foundation language models. arXiv preprint arXiv:2302.13971, 2023. 2" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 56, + 217, + 294, + 270 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 217, + 294, + 270 + ], + "spans": [ + { + "bbox": [ + 56, + 217, + 294, + 270 + ], + "type": "text", + "content": "[57] Ziyang Wang, Shoubin Yu, Elias Stengel-Eskin, Jaehong Yoon, Feng Cheng, Gedas Bertasius, and Mohit Bansal. Videotree: Adaptive tree-based video representation for llm reasoning on long videos. arXiv preprint arXiv:2405.19209, 2024. 2" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 56, + 272, + 294, + 315 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 272, + 294, + 315 + ], + "spans": [ + { + "bbox": [ + 56, + 272, + 294, + 315 + ], + "type": "text", + "content": "[58] Qi Wu, Yubo Zhao, Yifan Wang, Yu-Wing Tai, and Chikeung Tang. Motionllm: Multimodal motion-language learning with large language models. arXiv preprint arXiv:2405.17013, 2024." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 56, + 316, + 294, + 381 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 316, + 294, + 381 + ], + "spans": [ + { + "bbox": [ + 56, + 316, + 294, + 381 + ], + "type": "text", + "content": "[59] Haiyang Xu, Qinghao Ye, Mingshi Yan, Yaya Shi, Jiabo Ye, Yuanhong Xu, Chenliang Li, Bin Bi, Qiuchen Qian, Wei Wang, Guohai Xu, Ji Zhang, Songfang Huang, Feiran Huang, and Jingren Zhou. mplug-2: A modularized multimodal foundation model across text, image and video. arXiv preprint arXiv:2302.00402, 2023. 2" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 56, + 383, + 294, + 425 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 383, + 294, + 425 + ], + "spans": [ + { + "bbox": [ + 56, + 383, + 294, + 425 + ], + "type": "text", + "content": "[60] Lin Xu, Yilin Zhao, Daquan Zhou, Zhijie Lin, See Kiong Ng, and Jiashi Feng. Pllava: Parameter-free llava extension from images to videos for video dense captioning. arXiv preprint arXiv:2404.16994, 2024. 7" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 56, + 426, + 294, + 491 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 426, + 294, + 491 + ], + "spans": [ + { + "bbox": [ + 56, + 426, + 294, + 491 + ], + "type": "text", + "content": "[61] Qinghao Ye, Haiyang Xu, Jiabo Ye, Ming Yan, Anwen Hu, Haowei Liu, Qi Qian, Ji Zhang, and Fei Huang. mplug-owl2: Revolutionizing multi-modal large language model with modality collaboration. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 13040–13051, 2024. 2" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 56, + 493, + 294, + 536 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 493, + 294, + 536 + ], + "spans": [ + { + "bbox": [ + 56, + 493, + 294, + 536 + ], + "type": "text", + "content": "[62] Alex Young, Bei Chen, Chao Li, Chengen Huang, Ge Zhang, Guanwei Zhang, Heng Li, Jiangcheng Zhu, Jianqun Chen, Jing Chang, et al. Yi: Open foundation models by 01. ai. arXiv preprint arXiv:2403.04652, 2024. 2" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 56, + 537, + 294, + 581 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 537, + 294, + 581 + ], + "spans": [ + { + "bbox": [ + 56, + 537, + 294, + 581 + ], + "type": "text", + "content": "[63] Xiaohua Zhai, Basil Mustafa, Alexander Kolesnikov, and Lucas Beyer. Sigmoid loss for language image pre-training. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 11975-11986, 2023. 3" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 56, + 582, + 294, + 624 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 582, + 294, + 624 + ], + "spans": [ + { + "bbox": [ + 56, + 582, + 294, + 624 + ], + "type": "text", + "content": "[64] Ce Zhang, Taixi Lu, Md Mohaiminul Islam, Ziyang Wang, Shoubin Yu, Mohit Bansal, and Gedas Bertasius. A simple llm framework for long-range video question-answering. arXiv preprint arXiv:2312.17235, 2023. 2" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 56, + 625, + 294, + 667 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 625, + 294, + 667 + ], + "spans": [ + { + "bbox": [ + 56, + 625, + 294, + 667 + ], + "type": "text", + "content": "[65] Haoji Zhang, Yiqin Wang, Yansong Tang, Yong Liu, Jiashi Feng, Jifeng Dai, and Xiaojie Jin. Flash-vstream: Memory-based real-time understanding for long video streams. 2024. 3" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 56, + 669, + 294, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 669, + 294, + 713 + ], + "spans": [ + { + "bbox": [ + 56, + 669, + 294, + 713 + ], + "type": "text", + "content": "[66] Pan Zhang, Xiaoyi Dong, Bin Wang, Yuhang Cao, Chao Xu, Linke Ouyang, Zhiyuan Zhao, Haodong Duan, Songyang Zhang, Shuangrui Ding, et al. Internlmxcomposer: A vision-language large model for advanced" + } + ] + } + ], + "index": 12 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 316, + 73, + 553, + 373 + ], + "type": "list", + "angle": 0, + "index": 20, + "blocks": [ + { + "bbox": [ + 333, + 73, + 553, + 95 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 333, + 73, + 553, + 95 + ], + "spans": [ + { + "bbox": [ + 333, + 73, + 553, + 95 + ], + "type": "text", + "content": "text-image comprehension and composition. arXiv preprint arXiv:2309.15112, 2023. 2" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 316, + 96, + 553, + 161 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 96, + 553, + 161 + ], + "spans": [ + { + "bbox": [ + 316, + 96, + 553, + 161 + ], + "type": "text", + "content": "[67] Pan Zhang, Xiaoyi Dong, Yuhang Cao, Yuhang Zang, Rui Qian, Xilin Wei, Lin Chen, Yifei Li, Junbo Niu, Shuangrui Ding, et al. Internlm-xcomposer2. 5-omnilive: A comprehensive multimodal system for long-term streaming video and audio interactions. arXiv preprint arXiv:2412.09596, 2024.3" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 316, + 163, + 553, + 217 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 163, + 553, + 217 + ], + "spans": [ + { + "bbox": [ + 316, + 163, + 553, + 217 + ], + "type": "text", + "content": "[68] Pan Zhang, Xiaoyi Dong, Yuhang Zang, Yuhang Cao, Rui Qian, Lin Chen, Qipeng Guo, Haodong Duan, Bin Wang, Linke Ouyang, et al. Internlm-xcomposer-2.5: A versatile large vision language model supporting long-contextual input and output. arXiv preprint arXiv:2407.03320, 2024. 2" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 316, + 219, + 553, + 271 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 219, + 553, + 271 + ], + "spans": [ + { + "bbox": [ + 316, + 219, + 553, + 271 + ], + "type": "text", + "content": "[69] Peiyuan Zhang, Kaichen Zhang, Bo Li, Guangtao Zeng, Jingkang Yang, Yuanhan Zhang, Ziyue Wang, Haoran Tan, Chunyuan Li, and Ziwei Liu. Long context transfer from language to vision. arXiv preprint arXiv:2406.16852, 2024. 6" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 316, + 274, + 553, + 316 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 274, + 553, + 316 + ], + "spans": [ + { + "bbox": [ + 316, + 274, + 553, + 316 + ], + "type": "text", + "content": "[70] Yuanhan Zhang, Bo Li, haotian Liu, Yong jae Lee, Liangke Gui, Di Fu, Jiashi Feng, Ziwei Liu, and Chunyuan Li. Llavanext: A strong zero-shot video understanding model, 2024. 6" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 316, + 319, + 553, + 373 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 319, + 553, + 373 + ], + "spans": [ + { + "bbox": [ + 316, + 319, + 553, + 373 + ], + "type": "text", + "content": "[71] Junjie Zhou, Yan Shu, Bo Zhao, Boya Wu, Shitao Xiao, Xi Yang, Yongping Xiong, Bo Zhang, Tiejun Huang, and Zheng Liu. Mlvu: A comprehensive benchmark for multi-task long video understanding. arXiv preprint arXiv:2406.04264, 2024.5" + } + ] + } + ], + "index": 19 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "text", + "content": "24055" + } + ] + } + ], + "index": 21 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 10 + } + ], + "_backend": "vlm", + "_version_name": "2.6.4" +} \ No newline at end of file diff --git a/2025/Dissecting and Mitigating Diffusion Bias via Mechanistic Interpretability/3f305fc0-29c2-45c0-bbbf-2f6d947ae1f1_content_list.json b/2025/Dissecting and Mitigating Diffusion Bias via Mechanistic Interpretability/3f305fc0-29c2-45c0-bbbf-2f6d947ae1f1_content_list.json new file mode 100644 index 0000000000000000000000000000000000000000..fba2f26ac988ffeb2f7971a65e1b69266ab4bafb --- /dev/null +++ b/2025/Dissecting and Mitigating Diffusion Bias via Mechanistic Interpretability/3f305fc0-29c2-45c0-bbbf-2f6d947ae1f1_content_list.json @@ -0,0 +1,1600 @@ +[ + { + "type": "text", + "text": "Dissecting and Mitigating Diffusion Bias via Mechanistic Interpretability", + "text_level": 1, + "bbox": [ + 132, + 130, + 867, + 152 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Yingdong Shi $^{1*}$ Changming Li $^{1*}$ Yifan Wang $^{2}$ Yongxiang Zhao $^{1}$ Anqi Pang $^{3}$ Sibei Yang $^{1}$ Jingyi Yu $^{1}$ Kan Ren $^{1\\dagger}$", + "bbox": [ + 228, + 179, + 767, + 217 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "$^{1}$ ShanghaiTech University $^{2}$ Stony Brook University $^{3}$ Tencent PCG", + "bbox": [ + 215, + 222, + 781, + 241 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "{shiyd2023,lichm2024,renkan}@shanghaitech.edu.cn", + "bbox": [ + 284, + 244, + 709, + 258 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Abstract", + "text_level": 1, + "bbox": [ + 246, + 292, + 326, + 309 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Diffusion models have demonstrated impressive capabilities in synthesizing diverse content. However, despite their high-quality outputs, these models often perpetuate social biases, including those related to gender and race. These biases can potentially contribute to harmful real-world consequences, reinforcing stereotypes and exacerbating inequalities in various social contexts. While existing research on diffusion bias mitigation has predominantly focused on guiding content generation, it often neglects the intrinsic mechanisms within diffusion models that causally drive biased outputs. In this paper, we investigate the internal processes of diffusion models, identifying specific decision-making mechanisms, termed bias features, embedded within the model architecture. By directly manipulating these features, our method precisely isolates and adjusts the elements responsible for bias generation, permitting granular control over the bias levels in the generated content. Through experiments on both unconditional and conditional diffusion models across various social bias attributes, we demonstrate our method's efficacy in managing generation distribution while preserving image quality. We also dissect the discovered model mechanism, revealing different intrinsic features controlling fine-grained aspects of generation, boosting further research on mechanistic interpretability of diffusion models. The project website is at https://foundation-model-research.github.io/difflens.", + "bbox": [ + 89, + 325, + 485, + 717 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "1. Introduction", + "text_level": 1, + "bbox": [ + 89, + 746, + 220, + 762 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Diffusion models have emerged as powerful generative tools capable of generating high-quality content by iteratively denoising random noise [23, 55]. These models are now widely credited in applications such as content synthesis [49], data manipulation [29], and data augmentation [37], impacting numerous aspects of human society. However, diffusion models have been criticized for generating", + "bbox": [ + 89, + 771, + 483, + 878 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "biased content [34, 43], often resulting in skewed representations of sensitive social attributes such as gender, race, and age. While biases may originate from the training data [36, 47], they can be further amplified by the learned model itself [47]. Given the extensive deployment of diffusion models, these biases may lead to detrimental effects in real-world contexts, such as reinforcing stereotypes in media representations or perpetuating inequalities in automated decision-making systems.", + "bbox": [ + 511, + 295, + 906, + 430 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Existing approaches to debias diffusion models generally fall into two main strategies. The first attempts to train unbiased models from scratch on biased datasets [30] or fine-tune the target model [54], both of which are resource-intensive and can significantly impact the model's original performance. The second aims to guide or edit the generation process [17, 42, 43], but these methods undermine the distinct internal workings of diffusion models, risking overcorrecting the model behavior or distorting non-target attributes, which affecting generation quality. In contrast, few studies have explored the intrinsic mechanisms within diffusion models that causally drive biased outcomes.", + "bbox": [ + 511, + 431, + 908, + 611 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "In this work, we hypothesize that neurons collectively play a pivotal role in driving the generation of biased concepts. These neurons exhibit specific features which we identify as intrinsic mechanisms within the model that contribute to social bias. In particular, we set out to investigate these mechanisms to better understand and mitigate bias within diffusion models.", + "bbox": [ + 511, + 613, + 908, + 717 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Our approach also aligns with the latest efforts on understanding the inner workings of neural networks, which has drawn huge interest within the community [7, 44], not only due to the interpretability benefits that reveal model decision patterns [31, 59], but also for its potential to enhance control over model behaviors [4, 38]. However, the intrinsic neuron activities within diffusion models remain largely unexplored. A significant challenge stems from the generative nature of diffusion models, which complicates, if not completely prohibits, fine-grained analysis at the neuron level. The neuron polysemanticity [4], which links individual neurons to multiple unrelated concepts additionally, fur", + "bbox": [ + 511, + 719, + 910, + 900 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "CVF", + "bbox": [ + 106, + 2, + 181, + 42 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore.", + "bbox": [ + 236, + 0, + 810, + 46 + ], + "page_idx": 0 + }, + { + "type": "page_footnote", + "text": "*Equal contribution. †Corresponding author.", + "bbox": [ + 107, + 886, + 352, + 900 + ], + "page_idx": 0 + }, + { + "type": "page_number", + "text": "8192", + "bbox": [ + 482, + 944, + 514, + 955 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "ther presents obstacles on mechanistic study of large-scale diffusion models.", + "bbox": [ + 89, + 90, + 480, + 119 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "To investigate the underlying mechanisms of bias generation in diffusion models, we present DIFFLENS, a comprehensive framework that analyzes the architectural components to identify the most influential elements within the model. Specifically, DIFFLENS first disentangles the hidden neurons of the diffusion model into a sparse semantic feature space via a sparse autoencoder [35], which captures the patterns of neurons responsible for bias features such as gender and age. Next, DIFFLENS employs a gradient-based approach to pinpoint key components in the interpretable semantic space, identifying a minimal set of bias-specific features. We can then further control or even mitigate the bias by adjusting the effects of these bias-related features.", + "bbox": [ + 88, + 119, + 482, + 316 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "DIFFLENS offers two key advantages. First, via elaborately analyzing and identifying the inner mechanisms of diffusion process, it enables fine-grained control of bias with minimal impact on other semantic attributes. Second, the identified bias features are intrinsic to the diffusion model, revealing its underlying contribution to biased generation. Both have been demonstrated in our experiments.", + "bbox": [ + 88, + 316, + 482, + 422 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "In summary, our work makes several key contributions to the field of diffusion models. (i) To the best of our knowledge, this is the first work to address bias mitigation in diffusion models through the lens of mechanistic interpretability, uncovering the underlying patterns of bias generation. Our method provides new insights into how specific neuron activities contribute to biased outcomes, offering a deeper understanding beyond standard latent space analysis. (ii) We introduce a novel framework that includes semantic space construction, identification of influential features, and fine-grained control of bias generation based on the discovered features. This approach is not only effective for bias mitigation, but may also be applied to other aspects of diffusion model analysis. (iii) We conduct extensive experiments on two pretrained diffusion models, demonstrating the effectiveness of our proposed mechanism discovery in comparison to several strong baselines for diffusion bias mitigation.", + "bbox": [ + 91, + 424, + 482, + 680 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "2. Related Work", + "text_level": 1, + "bbox": [ + 89, + 694, + 232, + 709 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Bias in Diffusion Models. Diffusion-based architectures often reflect and amplify social biases [2, 34, 51], leading to potentially harmful outputs. Specifically, unconditional diffusion models exacerbate the biases present in the training data related to gender, race, and age [47]. Text-to-image models, such as Stable Diffusion [49], tend to generate images of specific gender or stereotypes when prompted with occupational terms or terms with character traits and other descriptors [5, 9, 34, 53].", + "bbox": [ + 89, + 719, + 482, + 854 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Mitigating Diffusion Bias. Developing techniques to mitigate social biases in generative models has become increasingly crucial [33, 43]. Many current approaches fo", + "bbox": [ + 89, + 854, + 482, + 900 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "cus on modifying prompts, such as adding gender-specific cues or ethical information, or using prompt learning techniques [1, 11, 16, 28]. While these methods are effective for text-to-image systems, they are limited to scenarios involving textual prompts. Another strategy involves fine-tuning diffusion models using curated datasets [54], which can effectively reduce bias but require extensive training and expensive high-quality data. Guided generation and model editing based methods [17, 20, 42, 43] provide an alternative by introducing external guidance in the model's processing [20, 43] or manipulating attention weight [17, 42]. Although these techniques avoid retraining, they undermine the intrinsic mechanisms within the model specifically related to bias generation, often resulting in imprecise edits that inadvertently affect other parts of the image and degrade quality. In contrast, our approach allows for precise identification of bias-generating features, preserving overall image quality and maintaining other attributes, thereby providing a more reliable solution for bias mitigation.", + "bbox": [ + 511, + 90, + 903, + 378 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Interpretability of Diffusion Models. Most of the work interpreting diffusion models focus on explaining how the input affects the generation via attribution methods and attention maps [8, 19, 21, 57], but they do not reveal the underlying mechanisms by which these attributes are processed and synthesized into a coherent output. Recent research has revealed that diffusion models possess a semantically rich latent space [32], inspiring various studies on customized image editing through latent space manipulation [18, 33, 45]. Although these methods enable more flexible editing, they still do not provide a deep understanding of the model's inherent features, leading to unexpected results. Our approach directly interacts with the internal components of diffusion models, enabling us to identify and precisely modify key features like bias-specific ones. This allows for more accurate control and targeted editing within the model itself, also providing efficient tools for interpreting diffusion models.", + "bbox": [ + 511, + 378, + 906, + 635 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "3. Preliminary", + "text_level": 1, + "bbox": [ + 511, + 650, + 638, + 667 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "In this section, we provide an overview of diffusion model architecture. Diffusion models operate through two key processes, the forward and backward diffusion steps [23]. Taking image generation as an example. In the forward process, noise is incrementally added to the input, transforming a real image $\\mathbf{x}_0$ into pure noise over a series of timesteps $t\\in [0,T)$ where $T$ is the total number of timesteps. Mathematically, the forward process can be represented as", + "bbox": [ + 511, + 676, + 903, + 797 + ], + "page_idx": 1 + }, + { + "type": "equation", + "text": "\n$$\nq \\left(\\mathbf {x} _ {t} \\mid \\mathbf {x} _ {0}\\right) = \\mathcal {N} \\left(\\mathbf {x} _ {t}; \\sqrt {\\bar {\\alpha} _ {t}} \\mathbf {x} _ {0}, (1 - \\bar {\\alpha} _ {t}) I\\right). \\tag {1}\n$$\n", + "text_format": "latex", + "bbox": [ + 576, + 809, + 903, + 828 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Here, $\\bar{\\alpha}_t\\coloneqq \\prod_{i = 1}^t\\alpha_i$ controls the amount of noise injected at each timestep $t$ , with $\\mathbf{x}_T$ becoming a near-random noise vector after the final timestep. In the backward process, the goal is to reverse this noising process and generate a clean", + "bbox": [ + 511, + 838, + 903, + 902 + ], + "page_idx": 1 + }, + { + "type": "page_number", + "text": "8193", + "bbox": [ + 480, + 944, + 514, + 955 + ], + "page_idx": 1 + }, + { + "type": "image", + "img_path": "images/f7139f10dd518f79fc8fa4a8c35d3448900402f917d4d54ba02be2610450e446.jpg", + "image_caption": [ + "Figure 1. Framework of DIFFLENS exploring the inner working of diffusion models for bias mitigation. Neurons that are activated in the semantic space are defined as fired features and inactivated ones as unfired features. Target features are either suppressed or amplified to control bias level. We divide the framework into three parts (a), (b) and (c) which correspond to Secs. 4.1 to 4.3, respectively." + ], + "image_footnote": [], + "bbox": [ + 114, + 87, + 883, + 320 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "image from the noisy sample. At each step, the diffusion network, e.g., U-Net [50], models $p_{\\theta}(\\mathbf{x}_{t - 1}|\\mathbf{x}_t)$ with parameter $\\theta$ predicting the denoised version given the current sample $\\mathbf{x}_t$ . This reverse process is modeled as", + "bbox": [ + 88, + 386, + 480, + 448 + ], + "page_idx": 2 + }, + { + "type": "equation", + "text": "\n$$\np _ {\\theta} (\\mathbf {x} _ {t - 1} | \\mathbf {x} _ {t}) = \\mathcal {N} (\\mathbf {x} _ {t - 1}; \\mu_ {\\theta} (\\mathbf {x} _ {t}, t), \\Sigma_ {\\theta} (\\mathbf {x} _ {t}, t)), \\tag {2}\n$$\n", + "text_format": "latex", + "bbox": [ + 130, + 462, + 482, + 479 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "where $\\mu_{\\theta}$ and $\\Sigma_{\\theta}$ represent the mean and variance of the predicted denoised sample.", + "bbox": [ + 89, + 492, + 482, + 523 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Our work specifically focuses on the inner working of model itself, such as U-Net [50] architecture in the diffusion model, to investigate the mechanisms underlying bias generation. We aim to examine the model's decision-making process through these representations to better understand the biases that emerge during image generation, providing deeper insights into the model's internal functions and the precise factors driving its behavior [4].", + "bbox": [ + 89, + 523, + 483, + 645 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "4. Methodology", + "text_level": 1, + "bbox": [ + 89, + 662, + 225, + 680 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "In this section, we propose DIFFLENS, consisting of three stages aimed at uncovering internal mechanisms of bias generation in diffusion models and leveraging this understanding for bias mitigation. Specifically, our goal is to discover the neuron activities that controls the generation of biased contents. To achieve this, DIFFLENS first disentangles the hidden states of neuron activation from the model layer into a sparse semantic space, capturing meaningful features that represent various attributes. Next, it identifies the semantic features specifically related to biased content generation using a gradient-based attribution method. Finally, by adjusting the scale of these identified features, DIFFLENS allows for control over the level of bias in generated content, providing an effective approach for bias mitigation.", + "bbox": [ + 89, + 689, + 482, + 900 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "The overall framework has been illustrated in Fig. 1 and Algorithms 1 and 2 in Appendix A.", + "bbox": [ + 511, + 386, + 903, + 417 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "4.1. Disentangling Polysemantic Neurons", + "text_level": 1, + "bbox": [ + 511, + 424, + 831, + 441 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "We study the hidden states $\\mathbf{h} = [h_1, h_2, \\dots, h_n] \\in \\mathbb{R}^n$ from the diffusion model, representing the activations of the $n$ neurons at each network layer. We assume that these neuron activations inherently capture the semantic information of the generation process. Similar observations have been made in language models [12, 38], though little attention has been given to diffusion models in this regard. Kwon et al. [32] demonstrated that diffusion models exhibit a semantic space within the U-Net [50] architecture, specifically at the bottleneck layer. However, previous works have not precisely analyzed the different attributes within this space. In contrast, our method explicitly identifies semantic features and pinpoints the specific dimensions of neuron activations responsible for bias generation. This enables more precise and targeted control over bias, as demonstrated in our experiments in Figs. 3 and 4.", + "bbox": [ + 509, + 446, + 906, + 688 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Note that, neurons are often polysemantic, meaning that each neuron often corresponds to multiple, unrelated concepts, as discussed in [14, 39, 41]. This places obstacles for understanding and interpreting the mechanism of the model, as the network's learned features are intertwined in a complex manner [4]. Nevertheless, it also presents an opportunity that these features can potentially be disentangled by identifying directions in the activation space, where each activation vector can be expressed as a sparse linear combination of underlying features [24]. Thus, we leverage k-Sparse Autoencoders (k-SAE) [35] to analyze and interpret the semantic features encoded within the hidden states $\\mathbf{h}$ . It helps the method focus on salient features, disentangling polysemantic activations and isolating semantic attributes", + "bbox": [ + 511, + 689, + 908, + 900 + ], + "page_idx": 2 + }, + { + "type": "page_number", + "text": "8194", + "bbox": [ + 482, + 944, + 514, + 955 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "within the latent space, making it effective for understanding and manipulating key features in diffusion models. Part (a) of Fig. 1 has illustrated this processing stage.", + "bbox": [ + 89, + 90, + 482, + 136 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Specifically, let $\\mathbf{H}^n = \\{\\mathbf{h}\\}$ be the original latent feature space with dimension number $n$ . We transform this feature space to a sparse high-dimensional space, denoted as $\\mathbf{S}^m = \\{\\mathbf{s}\\}$ with dimension $m$ that is much larger than $n$ . In the space $\\mathbf{S}^m$ , the features $\\mathbf{s}$ are monosemantic, i.e., each feature represents a distinct, non-overlapping semantic concept (discussions are in Appendix E.2). The mapping $\\Phi : \\mathbf{H}^n \\to \\mathbf{S}^m$ can be expressed by", + "bbox": [ + 89, + 136, + 483, + 257 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\mathbf {s} = \\left[ s _ {1}, s _ {2}, \\dots , s _ {m} \\right] = \\Phi (\\mathbf {h}) = \\operatorname {T o p K} \\left(\\mathbf {W} _ {\\text {e n c}} \\left(\\mathbf {h} - \\mathbf {b} _ {\\text {p r e}}\\right)\\right). \\tag {3}\n$$\n", + "text_format": "latex", + "bbox": [ + 96, + 263, + 483, + 294 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "We map the vector $\\mathbf{h}$ to a sparse vector $\\mathbf{s}$ with $\\| \\mathbf{s} \\|_0 = k$ and $k < m$ , indicating that $\\mathbf{s}$ is sparse with only $k$ features activated. This is achieved by the encoding process of a k-SAE where $\\mathbf{W}_{\\mathrm{enc}} \\in \\mathbb{R}^{m \\times n}$ and $\\mathbf{b}_{\\mathrm{pre}} \\in \\mathbb{R}^n$ is the bias term. $\\operatorname{TopK}(\\cdot)$ is an operation that selects the top $K = k$ largest features, referred to as \"fired\" features, while setting the remaining features to zero, which are \"unfired\" features as shown in Fig. 1. For the decoding process, we have", + "bbox": [ + 89, + 294, + 483, + 416 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\hat {\\mathbf {h}} = \\Phi^ {- 1} (\\mathbf {s}) = \\mathbf {W} _ {\\mathrm {d e c}} \\mathbf {s} + \\mathbf {b} _ {\\mathrm {p r e}}, \\tag {4}\n$$\n", + "text_format": "latex", + "bbox": [ + 181, + 420, + 482, + 439 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "where $\\hat{\\mathbf{h}}\\in \\mathbb{R}^n$ is the reconstructed latent vector and $\\mathbf{W}_{\\mathrm{dec}}\\in \\mathbb{R}^{n\\times m}$ . $\\Phi^{-1}$ represents the inverse mapping of $\\Phi$ since $\\mathbf{W}_{\\mathrm{dec}}$ and $\\mathbf{W}_{\\mathrm{enc}}$ are initialized as transposes of each other. We train the k-SAE using the loss function defined in Eq. (5), with the objective of minimizing the reconstruction error within the original hidden space, expressed as", + "bbox": [ + 89, + 446, + 483, + 537 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {L} (\\mathbf {h}) = \\left\\| \\mathbf {h} - \\hat {\\mathbf {h}} \\right\\| _ {2} ^ {2}, \\tag {5}\n$$\n", + "text_format": "latex", + "bbox": [ + 218, + 542, + 482, + 560 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "where $\\mathbf{h}$ and $\\hat{\\mathbf{h}}$ refers to the original and the reconstructed hidden vectors, respectively. We present results and implementation details in Appendix B.2, demonstrating that we reconstruct the original latent space with few influences on the target model. The encoding and decoding processes are presented in Algorithm 2 at Appendix A.", + "bbox": [ + 89, + 566, + 483, + 657 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Disentangling the original latent space of diffusion model shares several advantages. On one hand, transforming it into a sparse semantic space can well handle the polysemistic issue of the internal neurons within the neural network [24]. On the other hand, it enables more precise locating on the target feature for generation content attributes, such as identifying the bias features in the subsequent sections in our paper. In addition, editing the whole state in the original hidden space like that in [18, 32, 33, 43] could inadvertently alter irrelevant attributes, caused by the polysemantic properties of the neuron in the model as shown by Figs. 2 and 3 in experiments.", + "bbox": [ + 89, + 659, + 483, + 840 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "4.2. Identifying Bias Features", + "text_level": 1, + "bbox": [ + 89, + 847, + 320, + 863 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "The next step is to identify the most influential features within the sparse semantic space that are associated with", + "bbox": [ + 89, + 869, + 483, + 900 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "particular function, i.e., bias generation mechanism in this paper. This need arises because accurately identifying the features that correspond to specific attributes can significantly enhance our ability to control and manipulate generated content responsibly.", + "bbox": [ + 511, + 90, + 903, + 167 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Inspired by [13, 56], we propose a gradient-based bias attribution method, to locate the features that correlates to the bias concepts. We visualize this process in part (b) of Fig. 1. Intuitively, the neuron activations that influence the outcome of gender diverse in different aspects, which controls different characteristics such as hair style in Fig. 6.", + "bbox": [ + 511, + 167, + 905, + 257 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Our goal is to measure the influence of the semantic feature $s_i \\in \\mathbf{s}$ within $\\mathbf{S}^m$ driving the biased contents on the target outcome. We define $F_{\\mathbf{x}}: \\mathbf{S}^m \\to \\mathbb{R}$ as a measure of interest, assessing the outcome of the model w.r.t. the target $\\mathbf{s}$ given the input $\\mathbf{x}$ . Specifically, we quantify the bias level as the probability of the social attribute class $y$ as $F_{\\mathbf{x}}(\\mathbf{s}) = \\operatorname*{Pr}(y|\\mathbf{s})$ regarding the disentangled feature vectors derived from Eq. (3). Here $y$ represents different classes such as \"male\" or \"female\" in gender bias. To locate the influential features, the internal mechanisms of diffusion model, we propose the attribution score of bias generation to the disentangled feature $s_i$ as", + "bbox": [ + 511, + 257, + 906, + 439 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\nS (s _ {i}; \\mathbf {x}) = (s _ {i} - s _ {i} ^ {\\prime}) \\cdot \\int_ {\\alpha = 0} ^ {1} \\frac {\\partial F _ {\\mathbf {x}} (\\mathbf {s} ^ {\\prime} + \\alpha (\\mathbf {s} - \\mathbf {s} ^ {\\prime}))}{\\partial s _ {i}} d \\alpha , \\tag {6}\n$$\n", + "text_format": "latex", + "bbox": [ + 519, + 448, + 905, + 484 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "where $\\mathbf{s}' = [s_1',\\dots,s_m']$ is a relative baseline of $\\mathbf{s}$ and $\\partial F_{\\mathbf{x}}(\\mathbf{s}' + \\alpha (\\mathbf{s} - \\mathbf{s}')) / \\partial s_i$ is the gradient of the bias measure given the input image $\\mathbf{x}$ , to the target feature space $\\mathbf{s}$ . It is achieved by a light-weight classifier $F_{\\mathbf{x}}(\\cdot ;\\psi)$ with parameter $\\psi$ . The classifier construction follows [43] and we provide more training details in Appendix B.3. Choices of baselines $s'$ are in Appendix A.1.", + "bbox": [ + 511, + 492, + 905, + 598 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "In Eq. (6), the attribution score $S(s_{i};\\mathbf{x})$ aggregates the gradients along the path of the individual semantic feature $s_i$ ranging from zero effect $(\\alpha = 0)$ to full effect $(\\alpha = 1)$ , given the input $\\mathbf{x}$ , which is the generated content from the diffusion model. If a feature $s_i$ significantly impacts the representation of a particular attribute, its corresponding gradient will be prominent, resulting in higher integrated values, i.e., higher attribution score $S$ .", + "bbox": [ + 511, + 598, + 905, + 718 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Further, to assess the overall influence of the semantic feature $s_i$ to the bias contents, we aggregate the attribution score as $S(s_i; X) = \\sum_{j=1}^{N} S(s_i; \\mathbf{x}_j)$ where $\\mathbf{x}_j \\in X$ with $X$ as the support dataset of $N$ generated samples. DIFFLENS then selects the top $\\tau$ features with the largest attribution scores as bias features, whose index set is $\\mathbf{A} = \\{i_1, \\ldots, i_\\tau\\}$ , i.e., these are the most influential semantic features within the diffusion model driving the bias contents during generation. The complete process of bias attribution is outlined in Algorithm 1 at Appendix A. Notably, this step is required only once for one specific diffusion model.", + "bbox": [ + 511, + 719, + 905, + 883 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Additionally, we observe that the target can extend be", + "bbox": [ + 531, + 885, + 903, + 900 + ], + "page_idx": 3 + }, + { + "type": "page_number", + "text": "8195", + "bbox": [ + 480, + 944, + 514, + 955 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "yond s to include the input $\\mathbf{x}$ or any component within the target model itself. This broader approach enables us to examine how these elements impact the behavior of the underlying model concerning the measure $F_{\\mathbf{x}}$ . To the best of our knowledge, we are the first to incorporate feature disentanglement and attribution methods to interpret the bias-related components within diffusion models, providing deeper insights into the internal workings of diffusion process.", + "bbox": [ + 89, + 90, + 480, + 212 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "4.3. Intervening in Bias Features for Debiasing", + "text_level": 1, + "bbox": [ + 89, + 219, + 452, + 236 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "To control and mitigate social bias in diffusion models, we propose an effective method via intervening in the above discovered bias features indexed with $i \\in \\mathbf{A}$ , where $\\mathbf{A}$ consists of indexes of the identified bias features. It is possible to suppress or amplify these features to exert control over the model's behavior [12]. Unlike other methods [32, 33, 43], we only intervene in the identified features indexed with $\\mathbf{A}$ rather than the entire semantic space. Our intervention approach are illustrated in part (c) of Fig. 1.", + "bbox": [ + 89, + 243, + 482, + 378 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Specifically, we apply an intervention operation to the target features in the sparse semantic space. Then these features are reconstructed into the original latent space for subsequent image generation and more details are in Appendix A. For each feature indexed within $\\mathbf{A}$ , we can either apply a scalar multiplication or adding an additive bias as", + "bbox": [ + 89, + 378, + 482, + 470 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\ns _ {i} = \\operatorname {I n t e r v e n e} \\left(s _ {i}\\right) = \\left\\{ \\begin{array}{l l} \\beta s _ {i} & (\\text {S c a l i n g}), \\text {o r} \\\\ s _ {i} + \\beta & (\\text {A d d i n g}). \\forall i \\in \\mathbf {A}, \\beta \\in \\mathbb {R}. \\end{array} \\right. \\tag {7}\n$$\n", + "text_format": "latex", + "bbox": [ + 89, + 477, + 482, + 529 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "By applying a distinct parameter $\\beta$ , we can suppress or amplify the target bias feature to adjust the bias level or even mitigate it for improved balance.", + "bbox": [ + 89, + 530, + 482, + 575 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Our DIFFLENS demonstrates that it can control over the generation and representation of biased features in the output images as shown in Sec. 5.4. Furthermore, in Sec. 5.6, we gain more valuable insights into how different bias features are embedded within the inner working of diffusion model, allowing us not only to debias effectively but also to better understand the model's decision-making processes.", + "bbox": [ + 89, + 575, + 482, + 681 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "5. Experiments", + "text_level": 1, + "bbox": [ + 89, + 693, + 220, + 709 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "In this section, we evaluate our DIFFLENS qualitatively and quantitatively, following the below research questions.", + "bbox": [ + 89, + 718, + 482, + 747 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "RQ1: How effectively does DIFFLENS mitigate social bias while maintaining image quality? RQ2: Can DIFFLENS accurately identify intrinsic mechanism of bias generation?", + "bbox": [ + 89, + 748, + 482, + 792 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "RQ3: How well does it control the bias level through intervening in the found bias features? RQ4: Is it able to interpret bias feature in a fine-grained way?", + "bbox": [ + 89, + 794, + 482, + 839 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "5.1. Experimental Settings", + "text_level": 1, + "bbox": [ + 89, + 847, + 297, + 863 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Diffusion Model Architecture. We evaluate our proposed method on two pretrained diffusion models, (i) P2 [10],", + "bbox": [ + 89, + 869, + 482, + 900 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "an unconditional diffusion model trained on CelebA-HQ dataset [26]; (ii) Stable Diffusion [49], a text-conditional diffusion model trained on LAION dataset [52].", + "bbox": [ + 511, + 90, + 905, + 136 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Baselines. We compare our method against several baselines that are categorized into interpretability and guidance-based methods. For interpretability methods, we construct Activation [6] trying to locate influential neurons according to the neuron activation values. For guidance-based methods, we include Latent Editing [32] and H-Distribution [43] (abbreviated as \"L.E.\" and \"H-Dist.\"), trying to learn a latent vector to guide the unbiased generation within the bottleneck layer of U-Net. We additionally compare two methods, specifically for text-conditional diffusion models, which are Latent Direction [33] and Finetuning [54]. Due to Activation's limited debiasing effectiveness across both diffusion models and its restricted control over bias levels, we report its results only in Tab. 1 and Fig. 2. More details about baselines are in Appendix C.3.", + "bbox": [ + 511, + 137, + 906, + 364 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Implementation Details. To ensure fair comparison, and without loss of generality, we constrain our method, DIFFLENS, and other benchmark methods to operate on the bottleneck layer of the U-Net [50] in the diffusion model, adhering to the baseline settings in [32, 43]. Notably, DIFFLENS can easily extend the targeted mechanism to broader regions within the model. Following Sec. 4, we train kSAEs on the target latent space in the diffusion model. For the training of the light-weight classifier, we follow the settings in [43]. For the support set of samples used for identifying bias feature, we leverage the diffusion models to generate 1,000 sample for each attribute. We provide more details of implementation and dataset we use in Appendix B. The reproducible codes are in supplementary material.", + "bbox": [ + 511, + 364, + 906, + 575 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Evaluation Metrics. To measure the bias level of generated content, we use the Fairness Discrepancy (FD) introduced in [43] which calculates the Euclidean distance between a reference distribution (typically uniform) and the bias distribution of generation, representing class distributions within an attribute (e.g., male and female for gender attribute). A lower FD score indicates more balanced outputs. We also measure image quality using Fréchet Inception Distance (FID) [22] on the generated images. In addition, we follow [54] to measure (i) CLIP-T, which calculates the CLIP [48] based semantic similarity between the generated image and the input text prompt, and (ii) CLIP-I, which assesses the similarity between originally generated images and images after debiasing. More details are in Appendix C.", + "bbox": [ + 511, + 577, + 905, + 789 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "5.2. Debiasing Diffusion Models", + "text_level": 1, + "bbox": [ + 511, + 815, + 761, + 830 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "We define \"bias mitigation\" task as adjusting the diffusion model to produce balanced outputs across specific social attributes (e.g., gender). We focus on the task across gender, age, and race attributes. Gender includes male and female", + "bbox": [ + 511, + 839, + 905, + 900 + ], + "page_idx": 4 + }, + { + "type": "page_number", + "text": "8196", + "bbox": [ + 480, + 944, + 514, + 955 + ], + "page_idx": 4 + }, + { + "type": "image", + "img_path": "images/ef162e8a748da7c069f87047ccb63ada1e4d13cb0c633ef6806d0ac3c8ad980b.jpg", + "image_caption": [ + "Original M:F(%) 33.9:66.1", + "Figure 2. Comparison of randomly sampled original and debiased images from various baseline methods. The minority group (male) is highlighted with red bounding boxes for easier viewing. \"M:F\" refers to the male-to-female ratio. Our DIFFLENS effectively mitigates bias while preserving high generation quality, whereas other methods either struggle to maintain balance or produce images with artifacts." + ], + "image_footnote": [], + "bbox": [ + 94, + 88, + 250, + 210 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/8841639032a35a07dc5242762b1ace1bef02c29205d095bc9d7f72e9c11119b4.jpg", + "image_caption": [ + "Activation M:F(%) 36.6:63.4" + ], + "image_footnote": [], + "bbox": [ + 256, + 88, + 413, + 210 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/f7812aa5cffae3918b5bebd952688e68798759e0dcacbcff820fb0006f5e8a34.jpg", + "image_caption": [ + "Latent Editing \nM:F(%) 50.2:49.8" + ], + "image_footnote": [], + "bbox": [ + 419, + 88, + 575, + 210 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/2aee04b24988774fc2260a7f3108ff5f4d5ea4f261257f12495e76b7965a1e9c.jpg", + "image_caption": [ + "H-Distribution M:F(%) 46.6:53.4" + ], + "image_footnote": [], + "bbox": [ + 583, + 88, + 738, + 210 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/c549bbed29f7d5628145f7046c788246fd39d0cb60077e44eec72a9f3199aed6.jpg", + "image_caption": [ + "DiffLens (Ours) M:F(%) 49.9:50.1" + ], + "image_footnote": [], + "bbox": [ + 746, + 88, + 901, + 210 + ], + "page_idx": 5 + }, + { + "type": "table", + "img_path": "images/21a90c0ea43f1b6f1ae6b78ab4b63e47b99a74bbffdfde984f9d79280b72201f.jpg", + "table_caption": [], + "table_footnote": [], + "table_body": "
CategoryMethodGender (2)Age (3)Race (4)
FD ↓FID ↓CLIP-I ↑FD ↓FID ↓CLIP-I ↑FD ↓FID ↓CLIP-I ↑
Original0.22633.38-0.59233.38-0.71833.38-
Guidance-basedLatent Editing [32]0.00329.850.94740.60633.710.90920.31734.610.9081
H-Distribution [43]0.04831.310.94400.51134.210.85940.49436.190.8762
InterpretabilityActivation [6]0.19034.270.80600.54448.910.77930.70046.130.7846
DIFFLENS (Ours)0.00231.930.94790.40131.710.94140.44733.470.9111
", + "bbox": [ + 133, + 297, + 862, + 398 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Table 1. Evaluation of bias mitigation in unconditional diffusion model P2 [10]. The numbers beside each attribute indicate the number of classes for that attribute. We highlight the best results in bold and the second best with underlined format (excluding the \"original\").", + "bbox": [ + 89, + 407, + 906, + 436 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "classes; age includes young, adult, and old; and race includes white, black, Asian, and Indian classes.", + "bbox": [ + 89, + 450, + 482, + 479 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Intuitively, bias mitigation should yield balanced generation without compromising the overall semantics of the generated content. To comprehensively evaluate the effectiveness of bias mitigation (RQ1) across different methods, we assess (i) the bias level using the FD score of generated images, (ii) the generation quality via the FID score, and (iii) the preservation of overall image semantics after bias mitigation using CLIP-I/T scores. Beyond evaluating on the unconditional generative model, we also extend our analysis to a text-conditional one to showcase the broader applicability and generalizability of our method.", + "bbox": [ + 88, + 484, + 482, + 650 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Qualitative Results on Bias Mitigation. We present batches of generated examples before and after debiasing gender attribute to visually evaluate bias level and generation quality, as shown in Fig. 2. Our DIFFLENS effectively neutralizes gender bias while ensuring the generated images remain realistic and intact. In comparison, Latent Editing may generate unrealistic images, for example, the second one in the first row in Fig. 2, H-Distribution struggles to generate a well-balanced output, and Activation produces imbalanced, distorted images. We also provide visual results for text-to-image diffusion model Stable Diffusion [49] in Appendix D.2.", + "bbox": [ + 89, + 655, + 482, + 835 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Quantitative Analysis on Bias Mitigation. We assess baseline methods on debiasing across various social attributes (e.g., gender) by generating balanced outputs. As shown in Tab. 1, our DIFFLENS demonstrates competitively", + "bbox": [ + 89, + 839, + 482, + 901 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "strong results in (i) bias mitigation (measured by FD), (ii) in preserving semantic feature coherence (measured by CLIP-I) and (iii) maintaining generation quality (measured by FID). Activation shows poor ability to mitigate social bias. Although Latent Editing performs well in mitigating gender bias and achieves favorable FID, it can produce images with artifacts, as shown in Figs. 2 and 3. H-Distribution struggles to generate well-balanced outputs, as reflected by its poor FD scores.", + "bbox": [ + 511, + 450, + 906, + 585 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "In Tab. 2, we show results on a text-to-image model, Stable Diffusion [49]. Our DIFFLENS effectively mitigates bias, especially for gender and age, while effectively maintaining generation quality and image semantics among methods that are effective in bias mitigation. Latent Direction achieves good debiasing for age and race but often overcorrects, reducing CLIP-I scores and compromising the balance between target generation ratio and image quality. Finetuning, demanding extensive effort, excels in gender and race but struggles with age bias mitigation. Other baselines either handle multi-class attributes poorly or lack generalizability to text-to-image diffusion models.", + "bbox": [ + 511, + 588, + 908, + 771 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "5.3. Accurate Identification of Bias Mechanism", + "text_level": 1, + "bbox": [ + 511, + 784, + 877, + 799 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Accurate bias mechanism discovery could provide feasible control on the generated contents of diffusion, regarding the level of bias attributes. For instance, unlike the original diffusion model P2 [10], which generates a male-to-female gender ratio of approximately 3:7, our method easily attains the reverse ratio of 7:3. This capability goes beyond merely", + "bbox": [ + 511, + 809, + 906, + 902 + ], + "page_idx": 5 + }, + { + "type": "page_number", + "text": "8197", + "bbox": [ + 482, + 944, + 514, + 955 + ], + "page_idx": 5 + }, + { + "type": "table", + "img_path": "images/3242417ed5cdcaa408d7209c10620022d6ff6f13604af755afca26739dcc6a52.jpg", + "table_caption": [], + "table_footnote": [], + "table_body": "
MethodGender (2)Age (3)Race (4)
FD ↓FID ↓CLIP-I ↑CLIP-T ↑FD ↓FID ↓CLIP-I ↑CLIP-T ↑FD ↓FID ↓CLIP-I ↑CLIP-T ↑
Original0.564120.06-0.61550.752120.06-0.61550.558120.06-0.6155
Latent Editing [32]0.408166.110.82530.60050.682200.900.85270.61220.524153.050.88040.6086
H-Distribution [43]0.222151.680.84750.60870.506147.710.83450.60980.544126.900.82550.6100
Latent Direction [33]0.305129.370.80580.60910.052113.810.81510.60670.175128.300.82110.6132
Fintuning [54]0.050161.470.87790.60950.746161.470.87790.60950.198161.470.87790.6095
DIFFLENS (Ours)0.046112.830.85010.60900.04999.170.87780.60570.401119.860.90960.6149
", + "bbox": [ + 93, + 87, + 903, + 193 + ], + "page_idx": 6 + }, + { + "type": "table", + "img_path": "images/564746958e8aace1a0fe66c9c721ebd8e0ffabafa964aa24f1fd3c1c5055dc2d.jpg", + "table_caption": [ + "Table 2. Evaluation of bias mitigation in text-to-image diffusion model Stable Diffusion [49], based on average performance across four occupations (see Appendix B). We highlight the best results in bold and the second-best with underlined text (excluding the \"original\")." + ], + "table_footnote": [], + "table_body": "
OriginalL.E.H-Dist.DiffLensOriginalL.E.H-Dist.DiffLensOriginalL.E.H-Dist.DiffLens
", + "bbox": [ + 93, + 237, + 903, + 308 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "balancing generated images in Sec. 5.2; it allows us to examine how effectively each method edits the target attribute, e.g., gender, while preserving non-target semantic features, e.g., eyeglass and smile (RQ2). For illustration, we visualize the generated contents of compared methods upon the reversed ratio of gender attribute in Fig. 3. More examples and details are provided in Appendix D.4.1. As is shown in Fig. 3, our method can preserve non-target attributes such as smile, background coherence, and even eyeglasses when debiasing, showcasing accurate identification of the bias generation mechanism by our method. While the baseline methods often introduce distortions or compromise detail in these areas, without looking into inner working of bias generation. Additionally, from Tabs. 1 and 2, our DIFFLENS consistently achieves high CLIP-I scores, demonstrating its ability to accurately identify and modify target bias specific features with minimal impact on unrelated content.", + "bbox": [ + 88, + 359, + 483, + 617 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "5.4. Fine-grained Bias Level Control and Editing", + "text_level": 1, + "bbox": [ + 89, + 633, + 468, + 650 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Benefiting from bias mechanism identification, we can control bias level with finer granularity and across a broader range (RQ3). Figure 5 illustrates how different methods progressively modulate image generation in diffusion along two opposite gender directions. More examples are in Appendix D.4.2. Our DIFFLENS produces natural transitions, consistently preserving semantic features such as facial expression and maintaining high generation quality. Examples about control in multi-attributes are in Appendix D.4.3. The results of varying gender ratios generated by each method are shown in Fig. 4. Latent Editing and H-Distribution perform well when the gender ratio is around balanced, but their fidelity and visual coherence deteriorate at more skewed ratios, indicating a limitation in handling highly imbalanced outputs. By adjusting the value of intervention parameter incorporated in Eq. (7), our DIFFLENS can flexi", + "bbox": [ + 88, + 659, + 482, + 900 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "bly control bias feature and generate different gender ratios, ranging from balanced to imbalanced ones (almost male or female), illustrating powerful controllability over the representations of bias. Such highly imbalanced ratios also have wide applications in real world such as data augmentation for under-represented subgroups [43].", + "bbox": [ + 511, + 359, + 906, + 450 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/07fe781ce4742c76f11b0d5ebf3a3260fdff06c5d88d23725b097dd14937480f.jpg", + "image_caption": [ + "Figure 3. Comparison in accurate identification of bias features. Our DIFFLENS preserves overall image semantics such as smile and eyeglasses while other methods frequently introduce distortions or lose important details.", + "Figure 4. Achievable range of bias level control for each method. The Log Gender Ratio reflects the log of male to female ratio in the generated images, with 0 indicating balance. Our DIFFLENS offers broader bias control, delivering higher generation quality (left plot) and improved semantic coherence (right plot)." + ], + "image_footnote": [], + "bbox": [ + 514, + 462, + 903, + 564 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "5.5. Ablation Study", + "text_level": 1, + "bbox": [ + 511, + 650, + 668, + 667 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "We conduct an ablation study to reveal how each component in our DIFFLENS takes effect. Recall that, it dissects the diffusion models by (i) disentangling the latent space to a sparse semantic space via k-SAE and (ii) identifying the bias features using bias attribution. In Tab. 3, directly selecting neurons with the highest activation values as bias features (Activation method) performs poorly, supporting the claim in Sec. 4.1. Comparing \"Act@L\" with \"Act@S\" and \"Attr@L\" with DIFFLENS, using k-SAE improves visual coherence, as seen in the higher CLIP-I scores (0.8597 to 0.9479 and 0.8060 to 0.9368), while avoiding direct edits on polysemantic neurons mitigates FID degradation (41.68 to 31.93). Furthermore, comparing \"Act@L\" with \"Attr@L\" and \"Act@\" with DIFFLENS, bias attribution enables more precise localization of bias features, significantly lowering", + "bbox": [ + 509, + 674, + 906, + 902 + ], + "page_idx": 6 + }, + { + "type": "page_number", + "text": "8198", + "bbox": [ + 482, + 944, + 514, + 955 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/1549d9f196ea442434fa0c59b828050caaf9c0e983f0c1b3783a33831c269a97.jpg", + "image_caption": [ + "Figure 5. Comparison of individual image transformations along the gender axis, with columns displaying images sampled from progressively shifting ratios as in Fig. 4. Our DIFFLENS achieves smooth and consistent transitions that preserve semantic features like facial expressions, while other methods show distortions and loss of detail at higher imbalance ratios." + ], + "image_footnote": [], + "bbox": [ + 91, + 87, + 467, + 229 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/10016405e196c19f863a164b7f019b0a16fa3eb6995a03ac73672703899351ef.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 475, + 88, + 905, + 229 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "FD (0.190 to 0.072 and 0.303 to 0.002). DIFFLENS incorporating both shows superiority in bias mitigation with concise localization and intervention of bias features.", + "bbox": [ + 89, + 297, + 482, + 342 + ], + "page_idx": 7 + }, + { + "type": "table", + "img_path": "images/15ac8f29c2f21cf9d13b5a43971024c087bcdba9ac830793e08779e2c15e6b40.jpg", + "table_caption": [], + "table_footnote": [ + "- ①, ②: Bias attribution allows precise bias mitigation.", + "- ③, ④: K-SAE improves visual feature coherence and maintains generation quality, via disentangling original latent space." + ], + "table_body": "
ComponentsOrig.Act@LAct@SAttr@LDIFFLENS
ActivationX✓✓XX
K-SAEXX✓X✓
Bias Attr.XXX✓✓
FD ↓0.2260.1900.3030.0720.002
FID ↓33.3834.2732.7041.6831.93
CLIP-I ↑-0.80600.93860.85970.9479
", + "bbox": [ + 93, + 353, + 478, + 448 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Table 3. Ablation study of DIFFLENS in P2 [10] for gender attribute. We abbreviate Original as \"Orig.\", neuron activations as \"Act.\", and bias attributions as \"Attr\". \"@L\" denotes operations on the original latent space, and \"@S\" on sparse semantic space.", + "bbox": [ + 89, + 492, + 483, + 547 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "5.6. Wider Investigation: Bias Mechanism", + "text_level": 1, + "bbox": [ + 89, + 559, + 418, + 574 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "We further look into the bias mechanism, i.e., the identified bias features $\\{s_i|i\\in \\mathbf{A}\\}$ , to reveal its granular aspects on social attribute generation and how these features distinguishably affect the biased contents, shedding some lights on the inner workings of diffusion models (RQ4). We observe from Fig. 6 that (i) multiple features may jointly influence a single social bias attribute, such as \"short hair\" and \"mustache\" contributing to gender perception; (ii) some features overlap across biases, for instance, \"side pose\" being associated with both age and gender biases as it often appears in images of older males; and (iii) poses like \"side pose\" and \"head down\" reflect in photography, often aligning with gender norms that men are depicted in dynamic or indirect poses, while women appear younger and more polished, reflecting social expectations [15, 58, 60]. These stereotypes mirror underlying social biases, impacting not only how individuals are portrayed but also how their images are interpreted. Our DIFFLENS serves as a lens that uncovers social biases and stereotypes embedded in generated images, providing insight into how attributes such as gender, race, and age are systematically portrayed.", + "bbox": [ + 88, + 583, + 482, + 901 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "6. Conclusion", + "text_level": 1, + "bbox": [ + 513, + 296, + 632, + 311 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "In this work, we introduced a novel approach DIFFLENS, for mitigating social biases in diffusion models by dissecting, analyzing and intervening in the internal mechanisms of the diffusion model. Leveraging k-SAE and gradient-based attribution method, we disentangle, identify, and control bias-related features with precision, allowing targeted bias mitigation while preserving non-target attributes. A wide range of evaluations across social attributes such as gender, age, and race validate effectiveness of DIFFLENS in bias control and content preservation. Our work offers an interpretable solution to bias mitigation in diffusion models, though challenges remain, including the precise identification of certain racial features and a quality decline in intensified edits. Future directions include expanding this approach to other bias types, incorporating additional fairness perspectives, and exploring application across diverse model architectures like Diffusion Transformers [46] for broader impact in ethical generative AI deployment.", + "bbox": [ + 511, + 321, + 906, + 595 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/7654e5edfa8f6ef7606eec14cae9d7684146b2dd906c714e3c874a3b7ab0caf1.jpg", + "image_caption": [ + "Figure 6. Manually analyzing the discovered bias features in diffusion models. Each column represents one specific feature $s_i$ , $i \\in \\mathbf{A}$ and each row originates from the same raw image." + ], + "image_footnote": [], + "bbox": [ + 542, + 607, + 874, + 859 + ], + "page_idx": 7 + }, + { + "type": "page_number", + "text": "8199", + "bbox": [ + 482, + 944, + 514, + 955 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "7. Acknowledgment", + "text_level": 1, + "bbox": [ + 91, + 90, + 261, + 107 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "The research received partial support from National Natural Science Foundation of China (Grant No. 62406193). The authors also gratefully acknowledge further assistance provided by Shanghai Frontiers Science Center of Human-centered Artificial Intelligence, MoE Key Lab of Intelligent Perception and Human-Machine Collaboration, and HPC Platform of ShanghaiTech University.", + "bbox": [ + 89, + 113, + 483, + 212 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "References", + "text_level": 1, + "bbox": [ + 91, + 239, + 187, + 255 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[1] Hritik Bansal, Da Yin, Masoud Monajatipoor, and Kai-Wei Chang. How well can text-to-image generative models understand ethical natural language interventions? In Proceedings of the 2022 Conference on Empirical Methods in Natural Language Processing, pages 1358–1370, 2022. 2", + "[2] Abhipsa Basu, R Venkatesh Babu, and Danish Pruthi. Inspecting the geographical representativeness of images from text-to-image models. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 5136-5147, 2023. 2", + "[3] Sebastian Benthall and Bruce D. Haynes. Racial categories in machine learning. In Proceedings of the Conference on Fairness, Accountability, and Transparency, page 289-298, New York, NY, USA, 2019. Association for Computing Machinery. 11", + "[4] Leonard Bereska and Stratis Gavves. Mechanistic interpretability for AI safety - a review. Transactions on Machine Learning Research, 2024. Survey Certification, Expert Certification. 1, 3, 11", + "[5] Federico Bianchi, Pratyusha Kalluri, Esin Durmus, Faisal Ladhak, Myra Cheng, Debora Nozza, Tatsunori Hashimoto, Dan Jurafsky, James Zou, and Aylin Caliskan. Easily accessible text-to-image generation amplifies demographic stereotypes at large scale. In Proceedings of the 2023 ACM Conference on Fairness, Accountability, and Transparency, pages 1493-1504, 2023. 2", + "[6] Steven Bills, Nick Cammarata, Dan Mossing, Henk Tillman, Leo Gao, Gabriel Goh, Ilya Sutskever, Jan Leike, Jeff Wu, and William Saunders. Language models can explain neurons in language models. https://openaipublic.blob.core.windows.net/neuron-explainer/paper/index.html, 2023.5, 6", + "[7] Hila Chefer, Oran Lang, Mor Geva, Volodymyr Polosukhin, Assaf Shocher, Inbar Mosseri, Lior Wolf, et al. The hidden language of diffusion models. In *The Twelfth International Conference on Learning Representations*. 1", + "[8] Hila Chefer, Yuval Alaluf, Yael Vinker, Lior Wolf, and Daniel Cohen-Or. Attend-and-excite: Attention-based semantic guidance for text-to-image diffusion models. ACM Transactions on Graphics (TOG), 42(4):1-10, 2023. 2", + "[9] Jaemin Cho, Abhay Zala, and Mohit Bansal. Dall-eval: Probing the reasoning skills and social biases of text-to-image generation models. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 3043-3054, 2023. 2" + ], + "bbox": [ + 99, + 263, + 483, + 898 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[10] Jooyoung Choi, Jungbeom Lee, Chaehun Shin, Sungwon Kim, Hyunwoo Kim, and Sungroh Yoon. Perception prioritized training of diffusion models. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 11472-11481, 2022. 5, 6, 8, 1, 2, 3, 4, 9", + "[11] Ching-Yao Chuang, Varun Jampani, Yuanzhen Li, Antonio Torralba, and Stefanie Jegelka. Debiasing vision-language models via biased prompts. arXiv preprint arXiv:2302.00070, 2023. 2", + "[12] Damai Dai, Li Dong, Yaru Hao, Zhifang Sui, Baobao Chang, and Furu Wei. Knowledge neurons in pretrained transformers. In Proceedings of the 60th Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers), pages 8493-8502, Dublin, Ireland, 2022. Association for Computational Linguistics. 3, 5", + "[13] Kedar Dhamhere, Mukund Sundararajan, and Qiqi Yan. How important is a neuron. In International Conference on Learning Representations, 2019. 4", + "[14] Nelson Elhage, Tristan Hume, Catherine Olsson, Nicholas Schiefer, Tom Henighan, Shauna Kravec, Zac Hatfield-Dodds, Robert Lasenby, Dawn Drain, Carol Chen, et al. Toy models of superposition. arXiv preprint arXiv:2209.10652, 2022.3", + "[15] Emilio Ferrara. Fairness and bias in artificial intelligence: A brief survey of sources, impacts, and mitigation strategies. Sci, 6(1), 2024. 8", + "[16] Felix Friedrich, Manuel Brack, Lukas Struppek, Dominik Hintersdorf, Patrick Schramowski, Sasha Luccioni, and Kristian Kersting. Fair diffusion: Instructing text-to-image generation models on fairness. arXiv preprint arXiv:2302.10893, 2023. 2", + "[17] Rohit Gandikota, Hadas Orgad, Yonatan Belinkov, Joanna Materzyńska, and David Bau. Unified concept editing in diffusion models. In Proceedings of the IEEE/CVF Winter Conference on Applications of Computer Vision, pages 5111-5120, 2024. 1, 2", + "[18] René Haas, Inbar Huberman-Spiegelglas, Rotem Mulayoff, Stella Graßhof, Sami S Brandt, and Tomer Michaeli. Discovering interpretable directions in the semantic latent space of diffusion models. In 2024 IEEE 18th International Conference on Automatic Face and Gesture Recognition (FG), pages 1-9. IEEE, 2024. 2, 4", + "[19] Ligong Han, Yinxiao Li, Han Zhang, Peyman Milanfar, Dimitris Metaxas, and Feng Yang. Svdiff: Compact parameter space for diffusion fine-tuning. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 7323-7334, 2023. 2", + "[20] Ruifei He, Chuhui Xue, Haoru Tan, Wenqing Zhang, Yingchen Yu, Song Bai, and Xiaojuan Qi. Debiasing text-to-image diffusion models. In Proceedings of the 1st ACM Multimedia Workshop on Multi-Modal Misinformation Governance in the Era of Foundation Models, page 29–36, New York, NY, USA, 2024. Association for Computing Machinery. 2", + "[21] Amir Hertz, Ron Mokady, Jay Tenenbaum, Kfir Aberman, Yael Pritch, and Daniel Cohen-or. Prompt-to-prompt image editing with cross-attention control. In The Eleventh International Conference on Learning Representations, 2023. 2" + ], + "bbox": [ + 516, + 92, + 906, + 900 + ], + "page_idx": 8 + }, + { + "type": "page_number", + "text": "8200", + "bbox": [ + 482, + 944, + 514, + 955 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[22] Martin Heusel, Hubert Ramsauer, Thomas Unterthiner, Bernhard Nessler, and Sepp Hochreiter. Gans trained by a two time-scale update rule converge to a local nash equilibrium. Advances in neural information processing systems, 30, 2017. 5, 11", + "[23] Jonathan Ho, Ajay Jain, and Pieter Abbeel. Denoising diffusion probabilistic models. In Advances in Neural Information Processing Systems, pages 6840-6851. Curran Associates, Inc., 2020. 1, 2", + "[24] Robert Huben, Hoagy Cunningham, Logan Riggs Smith, Aidan Ewart, and Lee Sharkey. Sparse autoencoders find highly interpretable features in language models. In The Twelfth International Conference on Learning Representations, 2024. 3, 4, 1", + "[25] Kimmo Karkkainen and Jungseock Joo. Fairface: Face attribute dataset for balanced race, gender, and age for bias measurement and mitigation. In Proceedings of the IEEE/CVF winter conference on applications of computer vision, pages 1548-1558, 2021. 2, 4", + "[26] Tero Karras, Timo Aila, Samuli Laine, and Jaakko Lehtinen. Progressive growing of GANs for improved quality, stability, and variation. In International Conference on Learning Representations, 2018. 5", + "[27] Tero Karras, Samuli Laine, and Timo Aila. A style-based generator architecture for generative adversarial networks. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 4401-4410, 2019. 5", + "[28] Eunji Kim, Siwon Kim, Chaehun Shin, and Sungroh Yoon. De-stereotyping text-to-image models through prompt tuning. In ICML Workshop on Challenges in Deployable Generative AI, 2023. 2", + "[29] Gwanghyun Kim, Taesung Kwon, and Jong Chul Ye. Diffus ionclip: Text-guided diffusion models for robust image man i plication. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 2426-2435, 2022. 1", + "[30] Yeongmin Kim, Byeonghu Na, Minsang Park, JoonHo Jang, Dongjun Kim, Wanmo Kang, and Il chul Moon. Training unbiased diffusion models from biased dataset. In The Twelfth International Conference on Learning Representations, 2024. 1", + "[31] Xianghao Kong, Ollie Liu, Han Li, Dani Yogatama, and Greg Ver Steeg. Interpretable diffusion via information decomposition. In The Twelfth International Conference on Learning Representations. 1", + "[32] Mingi Kwon, Jaeseok Jeong, and Youngjung Uh. Diffusion models already have a semantic latent space. In The Eleventh International Conference on Learning Representations, 2023. 2, 3, 4, 5, 6, 7, 9", + "[33] Hang Li, Chengzhi Shen, Philip Torr, Volker Tresp, and Jindong Gu. Self-discovering interpretable diffusion latent directions for responsible text-to-image generation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 12006-12016, 2024. 2, 4, 5, 7", + "[34] Sasha Luccioni, Christopher Akiki, Margaret Mitchell, and Yacine Jernite. Stable bias: Evaluating societal representa" + ], + "bbox": [ + 91, + 92, + 480, + 900 + ], + "page_idx": 9 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "tions in diffusion models. Advances in Neural Information Processing Systems, 36, 2024. 1, 2, 11", + "[35] Alireza Makhzani and Brendan Frey. K-sparse autoencoders. In International Conference on Learning Representations, 2014. 2, 3, 1, 6, 11", + "[36] Vongani H Maluleke, Neerja Thakkar, Tim Brooks, Ethan Weber, Trevor Darrell, Alexei A Efros, Angjoo Kanazawa, and Devin Guillery. Studying bias in gans through the lens of race. In European Conference on Computer Vision, pages 344-360. Springer, 2022. 1", + "[37] Pietro Melzi, Christian Rathgeb, Ruben Tolosana, Ruben Vera-Rodriguez, Dominik Lawatsch, Florian Domin, and Maxim Schaubert. Gandifface: Controllable generation of synthetic datasets for face recognition with realistic variations. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 3086-3095, 2023. 1", + "[38] Kevin Meng, David Bau, Alex Andonian, and Yonatan Belinkov. Locating and editing factual associations in gpt. Advances in Neural Information Processing Systems, 35: 17359-17372, 2022. 1, 3", + "[39] Jesse Mu and Jacob Andreas. Compositional explanations of neurons. In Advances in Neural Information Processing Systems, pages 17153-17163. Curran Associates, Inc., 2020. 3", + "[40] Ranjita Naik and Besmira Nushi. Social biases through the text-to-image generation lens. In Proceedings of the 2023 AAAI/ACM Conference on AI, Ethics, and Society, pages 786-808, 2023. 2, 4", + "[41] Chris Olah, Nick Cammarata, Ludwig Schubert, Gabriel Goh, Michael Petrov, and Shan Carter. Zoom in: An introduction to circuits. Distill, 5(3):e00024-001, 2020. 3", + "[42] Hadas Orgad, Bahjat Kawar, and Yonatan Belinkov. Editing implicit assumptions in text-to-image diffusion models. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 7053-7061, 2023. 1, 2", + "[43] Rishubh Parihar, Abhijnya Bhat, Abhipsa Basu, Saswat Mallick, Jogendra Nath Kundu, and R Venkatesh Babu. Balancing act: Distribution-guided debiasing in diffusion models. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 6668-6678, 2024. 1, 2, 4, 5, 6, 7, 9, 11", + "[44] Ji-Hoon Park, Yeong-Joon Ju, and Seong-Whan Lee. Explaining generative diffusion models via visual analysis for interpretable decision-making process. Expert Systems with Applications, 248:123231, 2024. 1", + "[45] Yong-Hyun Park, Mingi Kwon, Jaewoong Choi, Junghyo Jo, and Youngjung Uh. Understanding the latent space of diffusion models through the lens of riemannian geometry. In Thirty-seventh Conference on Neural Information Processing Systems, 2023. 2", + "[46] William Peebles and Saining Xie. Scalable diffusion models with transformers. In Proceedings of the IEEE/CVF International Conference on Computer Vision (ICCV), pages 4195-4205, 2023. 8", + "[47] Malsha V Perera and Vishal M Patel. Analyzing bias in diffusion-based face generation models. In 2023 IEEE International Joint Conference on Biometrics (IJCB), pages 1-10. IEEE, 2023. 1, 2" + ], + "bbox": [ + 516, + 92, + 903, + 898 + ], + "page_idx": 9 + }, + { + "type": "page_number", + "text": "8201", + "bbox": [ + 482, + 944, + 513, + 955 + ], + "page_idx": 9 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[48] Alec Radford, Jong Wook Kim, Chris Hallacy, Aditya Ramesh, Gabriel Goh, Sandhini Agarwal, Girish Sastry, Amanda Askell, Pamela Mishkin, Jack Clark, et al. Learning transferable visual models from natural language supervision. In International conference on machine learning, pages 8748-8763. PMLR, 2021. 5", + "[49] Robin Rombach, Andreas Blattmann, Dominik Lorenz, Patrick Esser, and Björn Ommer. High-resolution image synthesis with latent diffusion models. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 10684-10695, 2022. 1, 2, 5, 6, 7, 3, 4", + "[50] Olaf Ronneberger, Philipp Fischer, and Thomas Brox. U-net: Convolutional networks for biomedical image segmentation. In Medical image computing and computer-assisted intervention-MICCAI 2015: 18th international conference, Munich, Germany, October 5-9, 2015, proceedings, part III 18, pages 234-241. Springer, 2015. 3, 5, 1, 2, 4, 6", + "[51] Harrison Rosenberg, Shimaa Ahmed, Guruprasad V Ramesh, Ramya Korlakai Vinayak, and Kassem Fawaz. Unbiased face synthesis with diffusion models: Are we there yet? arXiv preprint arXiv:2309.07277, 2023. 2", + "[52] Christoph Schuhmann, Romain Beaumont, Richard Vencu, Cade Gordon, Ross Wightman, Mehdi Cherti, Theo Coombes, Aarush Katta, Clayton Mullis, Mitchell Wortsman, et al. Laion-5b: An open large-scale dataset for training next generation image-text models. Advances in Neural Information Processing Systems, 35:25278-25294, 2022. 5", + "[53] Preethi Seshadri, Sameer Singh, and Yanai Elazar. The bias amplification paradox in text-to-image generation. In Proceedings of the 2024 Conference of the North American Chapter of the Association for Computational Linguistics: Human Language Technologies (Volume 1: Long Papers), pages 6367-6384, 2024. 2", + "[54] Xudong Shen, Chao Du, Tianyu Pang, Min Lin, Yongkang Wong, and Mohan Kankanhalli. Finetuning text-to-image diffusion models for fairness. In The Twelfth International Conference on Learning Representations, 2024. 1, 2, 5, 7, 6, 11", + "[55] Jiaming Song, Chenlin Meng, and Stefano Ermon. Denoising diffusion implicit models. In International Conference on Learning Representations, 2021. 1, 2, 3, 4", + "[56] Mukund Sundararajan, Ankur Taly, and Qiqi Yan. Axiomatic attribution for deep networks. In Proceedings of the 34th International Conference on Machine Learning-Volume 70, pages 3319–3328, 2017. 4", + "[57] Raphael Tang, Linqing Liu, Akshit Pandey, Zhiying Jiang, Gefei Yang, Karun Kumar, Pontus Stenetorp, Jimmy Lin, and Ferhan Ture. What the DAAM: Interpreting stable diffusion using cross attention. In Proceedings of the 61st Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers), pages 5644-5659, Toronto, Canada, 2023. Association for Computational Linguistics. 2", + "[58] Mike Zajko. Conservative ai and social inequality: conceptualizing alternatives to bias through social theory. *Ai & Society*, 36(3):1047–1056, 2021. 8", + "[59] Haiyan Zhao, Hanjie Chen, Fan Yang, Ninghao Liu, Huiqi Deng, Hengyi Cai, Shuaiqiang Wang, Dawei Yin, and Mengnan Du. Explainability for large language models: A survey." + ], + "bbox": [ + 89, + 90, + 485, + 901 + ], + "page_idx": 10 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "ACM Transactions on Intelligent Systems and Technology, 15(2):1-38, 2024. 1", + "[60] Mi Zhou, Vibhanshu Abhishek, Timothy Derdenger, Jaymo Kim, and Kannan Srinivasan. Bias in generative ai. arXiv preprint arXiv:2403.02726, 2024. 8" + ], + "bbox": [ + 516, + 92, + 906, + 162 + ], + "page_idx": 10 + }, + { + "type": "page_number", + "text": "8202", + "bbox": [ + 482, + 944, + 516, + 955 + ], + "page_idx": 10 + } +] \ No newline at end of file diff --git a/2025/Dissecting and Mitigating Diffusion Bias via Mechanistic Interpretability/3f305fc0-29c2-45c0-bbbf-2f6d947ae1f1_model.json b/2025/Dissecting and Mitigating Diffusion Bias via Mechanistic Interpretability/3f305fc0-29c2-45c0-bbbf-2f6d947ae1f1_model.json new file mode 100644 index 0000000000000000000000000000000000000000..5421e7c43af06f8d90ca5c1ffda26c98fda65c0d --- /dev/null +++ b/2025/Dissecting and Mitigating Diffusion Bias via Mechanistic Interpretability/3f305fc0-29c2-45c0-bbbf-2f6d947ae1f1_model.json @@ -0,0 +1,2312 @@ +[ + [ + { + "type": "header", + "bbox": [ + 0.107, + 0.003, + 0.182, + 0.043 + ], + "angle": 0, + "content": "CVF" + }, + { + "type": "header", + "bbox": [ + 0.237, + 0.001, + 0.812, + 0.047 + ], + "angle": 0, + "content": "This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore." + }, + { + "type": "title", + "bbox": [ + 0.133, + 0.131, + 0.868, + 0.154 + ], + "angle": 0, + "content": "Dissecting and Mitigating Diffusion Bias via Mechanistic Interpretability" + }, + { + "type": "text", + "bbox": [ + 0.23, + 0.18, + 0.768, + 0.218 + ], + "angle": 0, + "content": "Yingdong Shi\\(^{1*}\\) Changming Li\\(^{1*}\\) Yifan Wang\\(^{2}\\) Yongxiang Zhao\\(^{1}\\) Anqi Pang\\(^{3}\\) Sibei Yang\\(^{1}\\) Jingyi Yu\\(^{1}\\) Kan Ren\\(^{1\\dagger}\\)" + }, + { + "type": "text", + "bbox": [ + 0.216, + 0.223, + 0.782, + 0.242 + ], + "angle": 0, + "content": "\\(^{1}\\)ShanghaiTech University \\(^{2}\\)Stony Brook University \\(^{3}\\)Tencent PCG" + }, + { + "type": "text", + "bbox": [ + 0.285, + 0.245, + 0.71, + 0.26 + ], + "angle": 0, + "content": "{shiyd2023,lichm2024,renkan}@shanghaitech.edu.cn" + }, + { + "type": "title", + "bbox": [ + 0.248, + 0.294, + 0.327, + 0.31 + ], + "angle": 0, + "content": "Abstract" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.326, + 0.486, + 0.718 + ], + "angle": 0, + "content": "Diffusion models have demonstrated impressive capabilities in synthesizing diverse content. However, despite their high-quality outputs, these models often perpetuate social biases, including those related to gender and race. These biases can potentially contribute to harmful real-world consequences, reinforcing stereotypes and exacerbating inequalities in various social contexts. While existing research on diffusion bias mitigation has predominantly focused on guiding content generation, it often neglects the intrinsic mechanisms within diffusion models that causally drive biased outputs. In this paper, we investigate the internal processes of diffusion models, identifying specific decision-making mechanisms, termed bias features, embedded within the model architecture. By directly manipulating these features, our method precisely isolates and adjusts the elements responsible for bias generation, permitting granular control over the bias levels in the generated content. Through experiments on both unconditional and conditional diffusion models across various social bias attributes, we demonstrate our method's efficacy in managing generation distribution while preserving image quality. We also dissect the discovered model mechanism, revealing different intrinsic features controlling fine-grained aspects of generation, boosting further research on mechanistic interpretability of diffusion models. The project website is at https://foundation-model-research.github.io/difflens." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.747, + 0.222, + 0.763 + ], + "angle": 0, + "content": "1. Introduction" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.772, + 0.485, + 0.879 + ], + "angle": 0, + "content": "Diffusion models have emerged as powerful generative tools capable of generating high-quality content by iteratively denoising random noise [23, 55]. These models are now widely credited in applications such as content synthesis [49], data manipulation [29], and data augmentation [37], impacting numerous aspects of human society. However, diffusion models have been criticized for generating" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.296, + 0.908, + 0.431 + ], + "angle": 0, + "content": "biased content [34, 43], often resulting in skewed representations of sensitive social attributes such as gender, race, and age. While biases may originate from the training data [36, 47], they can be further amplified by the learned model itself [47]. Given the extensive deployment of diffusion models, these biases may lead to detrimental effects in real-world contexts, such as reinforcing stereotypes in media representations or perpetuating inequalities in automated decision-making systems." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.433, + 0.909, + 0.612 + ], + "angle": 0, + "content": "Existing approaches to debias diffusion models generally fall into two main strategies. The first attempts to train unbiased models from scratch on biased datasets [30] or fine-tune the target model [54], both of which are resource-intensive and can significantly impact the model's original performance. The second aims to guide or edit the generation process [17, 42, 43], but these methods undermine the distinct internal workings of diffusion models, risking overcorrecting the model behavior or distorting non-target attributes, which affecting generation quality. In contrast, few studies have explored the intrinsic mechanisms within diffusion models that causally drive biased outcomes." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.614, + 0.909, + 0.718 + ], + "angle": 0, + "content": "In this work, we hypothesize that neurons collectively play a pivotal role in driving the generation of biased concepts. These neurons exhibit specific features which we identify as intrinsic mechanisms within the model that contribute to social bias. In particular, we set out to investigate these mechanisms to better understand and mitigate bias within diffusion models." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.72, + 0.911, + 0.901 + ], + "angle": 0, + "content": "Our approach also aligns with the latest efforts on understanding the inner workings of neural networks, which has drawn huge interest within the community [7, 44], not only due to the interpretability benefits that reveal model decision patterns [31, 59], but also for its potential to enhance control over model behaviors [4, 38]. However, the intrinsic neuron activities within diffusion models remain largely unexplored. A significant challenge stems from the generative nature of diffusion models, which complicates, if not completely prohibits, fine-grained analysis at the neuron level. The neuron polysemanticity [4], which links individual neurons to multiple unrelated concepts additionally, fur" + }, + { + "type": "page_footnote", + "bbox": [ + 0.109, + 0.887, + 0.353, + 0.901 + ], + "angle": 0, + "content": "*Equal contribution. †Corresponding author." + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.516, + 0.957 + ], + "angle": 0, + "content": "8192" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.09, + 0.092, + 0.482, + 0.121 + ], + "angle": 0, + "content": "ther presents obstacles on mechanistic study of large-scale diffusion models." + }, + { + "type": "text", + "bbox": [ + 0.089, + 0.121, + 0.483, + 0.318 + ], + "angle": 0, + "content": "To investigate the underlying mechanisms of bias generation in diffusion models, we present DIFFLENS, a comprehensive framework that analyzes the architectural components to identify the most influential elements within the model. Specifically, DIFFLENS first disentangles the hidden neurons of the diffusion model into a sparse semantic feature space via a sparse autoencoder [35], which captures the patterns of neurons responsible for bias features such as gender and age. Next, DIFFLENS employs a gradient-based approach to pinpoint key components in the interpretable semantic space, identifying a minimal set of bias-specific features. We can then further control or even mitigate the bias by adjusting the effects of these bias-related features." + }, + { + "type": "text", + "bbox": [ + 0.089, + 0.318, + 0.483, + 0.424 + ], + "angle": 0, + "content": "DIFFLENS offers two key advantages. First, via elaborately analyzing and identifying the inner mechanisms of diffusion process, it enables fine-grained control of bias with minimal impact on other semantic attributes. Second, the identified bias features are intrinsic to the diffusion model, revealing its underlying contribution to biased generation. Both have been demonstrated in our experiments." + }, + { + "type": "text", + "bbox": [ + 0.093, + 0.425, + 0.483, + 0.681 + ], + "angle": 0, + "content": "In summary, our work makes several key contributions to the field of diffusion models. (i) To the best of our knowledge, this is the first work to address bias mitigation in diffusion models through the lens of mechanistic interpretability, uncovering the underlying patterns of bias generation. Our method provides new insights into how specific neuron activities contribute to biased outcomes, offering a deeper understanding beyond standard latent space analysis. (ii) We introduce a novel framework that includes semantic space construction, identification of influential features, and fine-grained control of bias generation based on the discovered features. This approach is not only effective for bias mitigation, but may also be applied to other aspects of diffusion model analysis. (iii) We conduct extensive experiments on two pretrained diffusion models, demonstrating the effectiveness of our proposed mechanism discovery in comparison to several strong baselines for diffusion bias mitigation." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.695, + 0.233, + 0.71 + ], + "angle": 0, + "content": "2. Related Work" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.72, + 0.483, + 0.856 + ], + "angle": 0, + "content": "Bias in Diffusion Models. Diffusion-based architectures often reflect and amplify social biases [2, 34, 51], leading to potentially harmful outputs. Specifically, unconditional diffusion models exacerbate the biases present in the training data related to gender, race, and age [47]. Text-to-image models, such as Stable Diffusion [49], tend to generate images of specific gender or stereotypes when prompted with occupational terms or terms with character traits and other descriptors [5, 9, 34, 53]." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.856, + 0.483, + 0.901 + ], + "angle": 0, + "content": "Mitigating Diffusion Bias. Developing techniques to mitigate social biases in generative models has become increasingly crucial [33, 43]. Many current approaches fo" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.092, + 0.905, + 0.379 + ], + "angle": 0, + "content": "cus on modifying prompts, such as adding gender-specific cues or ethical information, or using prompt learning techniques [1, 11, 16, 28]. While these methods are effective for text-to-image systems, they are limited to scenarios involving textual prompts. Another strategy involves fine-tuning diffusion models using curated datasets [54], which can effectively reduce bias but require extensive training and expensive high-quality data. Guided generation and model editing based methods [17, 20, 42, 43] provide an alternative by introducing external guidance in the model's processing [20, 43] or manipulating attention weight [17, 42]. Although these techniques avoid retraining, they undermine the intrinsic mechanisms within the model specifically related to bias generation, often resulting in imprecise edits that inadvertently affect other parts of the image and degrade quality. In contrast, our approach allows for precise identification of bias-generating features, preserving overall image quality and maintaining other attributes, thereby providing a more reliable solution for bias mitigation." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.379, + 0.907, + 0.636 + ], + "angle": 0, + "content": "Interpretability of Diffusion Models. Most of the work interpreting diffusion models focus on explaining how the input affects the generation via attribution methods and attention maps [8, 19, 21, 57], but they do not reveal the underlying mechanisms by which these attributes are processed and synthesized into a coherent output. Recent research has revealed that diffusion models possess a semantically rich latent space [32], inspiring various studies on customized image editing through latent space manipulation [18, 33, 45]. Although these methods enable more flexible editing, they still do not provide a deep understanding of the model's inherent features, leading to unexpected results. Our approach directly interacts with the internal components of diffusion models, enabling us to identify and precisely modify key features like bias-specific ones. This allows for more accurate control and targeted editing within the model itself, also providing efficient tools for interpreting diffusion models." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.651, + 0.64, + 0.669 + ], + "angle": 0, + "content": "3. Preliminary" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.677, + 0.905, + 0.799 + ], + "angle": 0, + "content": "In this section, we provide an overview of diffusion model architecture. Diffusion models operate through two key processes, the forward and backward diffusion steps [23]. Taking image generation as an example. In the forward process, noise is incrementally added to the input, transforming a real image \\(\\mathbf{x}_0\\) into pure noise over a series of timesteps \\(t\\in [0,T)\\) where \\(T\\) is the total number of timesteps. Mathematically, the forward process can be represented as" + }, + { + "type": "equation", + "bbox": [ + 0.577, + 0.81, + 0.905, + 0.829 + ], + "angle": 0, + "content": "\\[\nq \\left(\\mathbf {x} _ {t} \\mid \\mathbf {x} _ {0}\\right) = \\mathcal {N} \\left(\\mathbf {x} _ {t}; \\sqrt {\\bar {\\alpha} _ {t}} \\mathbf {x} _ {0}, (1 - \\bar {\\alpha} _ {t}) I\\right). \\tag {1}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.839, + 0.905, + 0.903 + ], + "angle": 0, + "content": "Here, \\(\\bar{\\alpha}_t\\coloneqq \\prod_{i = 1}^t\\alpha_i\\) controls the amount of noise injected at each timestep \\(t\\), with \\(\\mathbf{x}_T\\) becoming a near-random noise vector after the final timestep. In the backward process, the goal is to reverse this noising process and generate a clean" + }, + { + "type": "page_number", + "bbox": [ + 0.482, + 0.945, + 0.516, + 0.957 + ], + "angle": 0, + "content": "8193" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.115, + 0.088, + 0.885, + 0.321 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.089, + 0.331, + 0.908, + 0.377 + ], + "angle": 0, + "content": "Figure 1. Framework of DIFFLENS exploring the inner working of diffusion models for bias mitigation. Neurons that are activated in the semantic space are defined as fired features and inactivated ones as unfired features. Target features are either suppressed or amplified to control bias level. We divide the framework into three parts (a), (b) and (c) which correspond to Secs. 4.1 to 4.3, respectively." + }, + { + "type": "text", + "bbox": [ + 0.089, + 0.387, + 0.482, + 0.449 + ], + "angle": 0, + "content": "image from the noisy sample. At each step, the diffusion network, e.g., U-Net [50], models \\( p_{\\theta}(\\mathbf{x}_{t - 1}|\\mathbf{x}_t) \\) with parameter \\( \\theta \\) predicting the denoised version given the current sample \\( \\mathbf{x}_t \\). This reverse process is modeled as" + }, + { + "type": "equation", + "bbox": [ + 0.131, + 0.463, + 0.483, + 0.48 + ], + "angle": 0, + "content": "\\[\np _ {\\theta} (\\mathbf {x} _ {t - 1} | \\mathbf {x} _ {t}) = \\mathcal {N} (\\mathbf {x} _ {t - 1}; \\mu_ {\\theta} (\\mathbf {x} _ {t}, t), \\Sigma_ {\\theta} (\\mathbf {x} _ {t}, t)), \\tag {2}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.493, + 0.483, + 0.524 + ], + "angle": 0, + "content": "where \\(\\mu_{\\theta}\\) and \\(\\Sigma_{\\theta}\\) represent the mean and variance of the predicted denoised sample." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.525, + 0.484, + 0.646 + ], + "angle": 0, + "content": "Our work specifically focuses on the inner working of model itself, such as U-Net [50] architecture in the diffusion model, to investigate the mechanisms underlying bias generation. We aim to examine the model's decision-making process through these representations to better understand the biases that emerge during image generation, providing deeper insights into the model's internal functions and the precise factors driving its behavior [4]." + }, + { + "type": "title", + "bbox": [ + 0.09, + 0.663, + 0.226, + 0.681 + ], + "angle": 0, + "content": "4. Methodology" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.69, + 0.483, + 0.901 + ], + "angle": 0, + "content": "In this section, we propose DIFFLENS, consisting of three stages aimed at uncovering internal mechanisms of bias generation in diffusion models and leveraging this understanding for bias mitigation. Specifically, our goal is to discover the neuron activities that controls the generation of biased contents. To achieve this, DIFFLENS first disentangles the hidden states of neuron activation from the model layer into a sparse semantic space, capturing meaningful features that represent various attributes. Next, it identifies the semantic features specifically related to biased content generation using a gradient-based attribution method. Finally, by adjusting the scale of these identified features, DIFFLENS allows for control over the level of bias in generated content, providing an effective approach for bias mitigation." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.387, + 0.905, + 0.419 + ], + "angle": 0, + "content": "The overall framework has been illustrated in Fig. 1 and Algorithms 1 and 2 in Appendix A." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.425, + 0.833, + 0.442 + ], + "angle": 0, + "content": "4.1. Disentangling Polysemantic Neurons" + }, + { + "type": "text", + "bbox": [ + 0.511, + 0.448, + 0.907, + 0.689 + ], + "angle": 0, + "content": "We study the hidden states \\(\\mathbf{h} = [h_1, h_2, \\dots, h_n] \\in \\mathbb{R}^n\\) from the diffusion model, representing the activations of the \\(n\\) neurons at each network layer. We assume that these neuron activations inherently capture the semantic information of the generation process. Similar observations have been made in language models [12, 38], though little attention has been given to diffusion models in this regard. Kwon et al. [32] demonstrated that diffusion models exhibit a semantic space within the U-Net [50] architecture, specifically at the bottleneck layer. However, previous works have not precisely analyzed the different attributes within this space. In contrast, our method explicitly identifies semantic features and pinpoints the specific dimensions of neuron activations responsible for bias generation. This enables more precise and targeted control over bias, as demonstrated in our experiments in Figs. 3 and 4." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.69, + 0.909, + 0.901 + ], + "angle": 0, + "content": "Note that, neurons are often polysemantic, meaning that each neuron often corresponds to multiple, unrelated concepts, as discussed in [14, 39, 41]. This places obstacles for understanding and interpreting the mechanism of the model, as the network's learned features are intertwined in a complex manner [4]. Nevertheless, it also presents an opportunity that these features can potentially be disentangled by identifying directions in the activation space, where each activation vector can be expressed as a sparse linear combination of underlying features [24]. Thus, we leverage k-Sparse Autoencoders (k-SAE) [35] to analyze and interpret the semantic features encoded within the hidden states \\(\\mathbf{h}\\). It helps the method focus on salient features, disentangling polysemantic activations and isolating semantic attributes" + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.516, + 0.957 + ], + "angle": 0, + "content": "8194" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.09, + 0.092, + 0.483, + 0.137 + ], + "angle": 0, + "content": "within the latent space, making it effective for understanding and manipulating key features in diffusion models. Part (a) of Fig. 1 has illustrated this processing stage." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.137, + 0.484, + 0.258 + ], + "angle": 0, + "content": "Specifically, let \\(\\mathbf{H}^n = \\{\\mathbf{h}\\}\\) be the original latent feature space with dimension number \\(n\\). We transform this feature space to a sparse high-dimensional space, denoted as \\(\\mathbf{S}^m = \\{\\mathbf{s}\\}\\) with dimension \\(m\\) that is much larger than \\(n\\). In the space \\(\\mathbf{S}^m\\), the features \\(\\mathbf{s}\\) are monosemantic, i.e., each feature represents a distinct, non-overlapping semantic concept (discussions are in Appendix E.2). The mapping \\(\\Phi : \\mathbf{H}^n \\to \\mathbf{S}^m\\) can be expressed by" + }, + { + "type": "equation", + "bbox": [ + 0.097, + 0.264, + 0.484, + 0.295 + ], + "angle": 0, + "content": "\\[\n\\mathbf {s} = \\left[ s _ {1}, s _ {2}, \\dots , s _ {m} \\right] = \\Phi (\\mathbf {h}) = \\operatorname {T o p K} \\left(\\mathbf {W} _ {\\text {e n c}} \\left(\\mathbf {h} - \\mathbf {b} _ {\\text {p r e}}\\right)\\right). \\tag {3}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.295, + 0.484, + 0.417 + ], + "angle": 0, + "content": "We map the vector \\(\\mathbf{h}\\) to a sparse vector \\(\\mathbf{s}\\) with \\(\\| \\mathbf{s} \\|_0 = k\\) and \\(k < m\\), indicating that \\(\\mathbf{s}\\) is sparse with only \\(k\\) features activated. This is achieved by the encoding process of a k-SAE where \\(\\mathbf{W}_{\\mathrm{enc}} \\in \\mathbb{R}^{m \\times n}\\) and \\(\\mathbf{b}_{\\mathrm{pre}} \\in \\mathbb{R}^n\\) is the bias term. \\(\\operatorname{TopK}(\\cdot)\\) is an operation that selects the top \\(K = k\\) largest features, referred to as \"fired\" features, while setting the remaining features to zero, which are \"unfired\" features as shown in Fig. 1. For the decoding process, we have" + }, + { + "type": "equation", + "bbox": [ + 0.183, + 0.421, + 0.483, + 0.44 + ], + "angle": 0, + "content": "\\[\n\\hat {\\mathbf {h}} = \\Phi^ {- 1} (\\mathbf {s}) = \\mathbf {W} _ {\\mathrm {d e c}} \\mathbf {s} + \\mathbf {b} _ {\\mathrm {p r e}}, \\tag {4}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.447, + 0.484, + 0.538 + ], + "angle": 0, + "content": "where \\(\\hat{\\mathbf{h}}\\in \\mathbb{R}^n\\) is the reconstructed latent vector and \\(\\mathbf{W}_{\\mathrm{dec}}\\in \\mathbb{R}^{n\\times m}\\). \\(\\Phi^{-1}\\) represents the inverse mapping of \\(\\Phi\\) since \\(\\mathbf{W}_{\\mathrm{dec}}\\) and \\(\\mathbf{W}_{\\mathrm{enc}}\\) are initialized as transposes of each other. We train the k-SAE using the loss function defined in Eq. (5), with the objective of minimizing the reconstruction error within the original hidden space, expressed as" + }, + { + "type": "equation", + "bbox": [ + 0.22, + 0.543, + 0.483, + 0.561 + ], + "angle": 0, + "content": "\\[\n\\mathcal {L} (\\mathbf {h}) = \\left\\| \\mathbf {h} - \\hat {\\mathbf {h}} \\right\\| _ {2} ^ {2}, \\tag {5}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.568, + 0.484, + 0.659 + ], + "angle": 0, + "content": "where \\(\\mathbf{h}\\) and \\(\\hat{\\mathbf{h}}\\) refers to the original and the reconstructed hidden vectors, respectively. We present results and implementation details in Appendix B.2, demonstrating that we reconstruct the original latent space with few influences on the target model. The encoding and decoding processes are presented in Algorithm 2 at Appendix A." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.66, + 0.484, + 0.841 + ], + "angle": 0, + "content": "Disentangling the original latent space of diffusion model shares several advantages. On one hand, transforming it into a sparse semantic space can well handle the polysemistic issue of the internal neurons within the neural network [24]. On the other hand, it enables more precise locating on the target feature for generation content attributes, such as identifying the bias features in the subsequent sections in our paper. In addition, editing the whole state in the original hidden space like that in [18, 32, 33, 43] could inadvertently alter irrelevant attributes, caused by the polysemantic properties of the neuron in the model as shown by Figs. 2 and 3 in experiments." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.848, + 0.321, + 0.864 + ], + "angle": 0, + "content": "4.2. Identifying Bias Features" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.871, + 0.484, + 0.901 + ], + "angle": 0, + "content": "The next step is to identify the most influential features within the sparse semantic space that are associated with" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.092, + 0.905, + 0.168 + ], + "angle": 0, + "content": "particular function, i.e., bias generation mechanism in this paper. This need arises because accurately identifying the features that correspond to specific attributes can significantly enhance our ability to control and manipulate generated content responsibly." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.168, + 0.906, + 0.258 + ], + "angle": 0, + "content": "Inspired by [13, 56], we propose a gradient-based bias attribution method, to locate the features that correlates to the bias concepts. We visualize this process in part (b) of Fig. 1. Intuitively, the neuron activations that influence the outcome of gender diverse in different aspects, which controls different characteristics such as hair style in Fig. 6." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.258, + 0.907, + 0.44 + ], + "angle": 0, + "content": "Our goal is to measure the influence of the semantic feature \\( s_i \\in \\mathbf{s} \\) within \\( \\mathbf{S}^m \\) driving the biased contents on the target outcome. We define \\( F_{\\mathbf{x}}: \\mathbf{S}^m \\to \\mathbb{R} \\) as a measure of interest, assessing the outcome of the model w.r.t. the target \\( \\mathbf{s} \\) given the input \\( \\mathbf{x} \\). Specifically, we quantify the bias level as the probability of the social attribute class \\( y \\) as \\( F_{\\mathbf{x}}(\\mathbf{s}) = \\operatorname*{Pr}(y|\\mathbf{s}) \\) regarding the disentangled feature vectors derived from Eq. (3). Here \\( y \\) represents different classes such as \"male\" or \"female\" in gender bias. To locate the influential features, the internal mechanisms of diffusion model, we propose the attribution score of bias generation to the disentangled feature \\( s_i \\) as" + }, + { + "type": "equation", + "bbox": [ + 0.521, + 0.449, + 0.906, + 0.485 + ], + "angle": 0, + "content": "\\[\nS (s _ {i}; \\mathbf {x}) = (s _ {i} - s _ {i} ^ {\\prime}) \\cdot \\int_ {\\alpha = 0} ^ {1} \\frac {\\partial F _ {\\mathbf {x}} (\\mathbf {s} ^ {\\prime} + \\alpha (\\mathbf {s} - \\mathbf {s} ^ {\\prime}))}{\\partial s _ {i}} d \\alpha , \\tag {6}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.493, + 0.906, + 0.599 + ], + "angle": 0, + "content": "where \\(\\mathbf{s}' = [s_1',\\dots,s_m']\\) is a relative baseline of \\(\\mathbf{s}\\) and \\(\\partial F_{\\mathbf{x}}(\\mathbf{s}' + \\alpha (\\mathbf{s} - \\mathbf{s}')) / \\partial s_i\\) is the gradient of the bias measure given the input image \\(\\mathbf{x}\\), to the target feature space \\(\\mathbf{s}\\). It is achieved by a light-weight classifier \\(F_{\\mathbf{x}}(\\cdot ;\\psi)\\) with parameter \\(\\psi\\). The classifier construction follows [43] and we provide more training details in Appendix B.3. Choices of baselines \\(s'\\) are in Appendix A.1." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.599, + 0.906, + 0.719 + ], + "angle": 0, + "content": "In Eq. (6), the attribution score \\( S(s_{i};\\mathbf{x}) \\) aggregates the gradients along the path of the individual semantic feature \\( s_i \\) ranging from zero effect \\( (\\alpha = 0) \\) to full effect \\( (\\alpha = 1) \\), given the input \\( \\mathbf{x} \\), which is the generated content from the diffusion model. If a feature \\( s_i \\) significantly impacts the representation of a particular attribute, its corresponding gradient will be prominent, resulting in higher integrated values, i.e., higher attribution score \\( S \\)." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.72, + 0.906, + 0.885 + ], + "angle": 0, + "content": "Further, to assess the overall influence of the semantic feature \\( s_i \\) to the bias contents, we aggregate the attribution score as \\( S(s_i; X) = \\sum_{j=1}^{N} S(s_i; \\mathbf{x}_j) \\) where \\( \\mathbf{x}_j \\in X \\) with \\( X \\) as the support dataset of \\( N \\) generated samples. DIFFLENS then selects the top \\( \\tau \\) features with the largest attribution scores as bias features, whose index set is \\( \\mathbf{A} = \\{i_1, \\ldots, i_\\tau\\} \\), i.e., these are the most influential semantic features within the diffusion model driving the bias contents during generation. The complete process of bias attribution is outlined in Algorithm 1 at Appendix A. Notably, this step is required only once for one specific diffusion model." + }, + { + "type": "text", + "bbox": [ + 0.532, + 0.886, + 0.905, + 0.901 + ], + "angle": 0, + "content": "Additionally, we observe that the target can extend be" + }, + { + "type": "page_number", + "bbox": [ + 0.482, + 0.945, + 0.516, + 0.957 + ], + "angle": 0, + "content": "8195" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.09, + 0.092, + 0.482, + 0.213 + ], + "angle": 0, + "content": "yond s to include the input \\(\\mathbf{x}\\) or any component within the target model itself. This broader approach enables us to examine how these elements impact the behavior of the underlying model concerning the measure \\(F_{\\mathbf{x}}\\). To the best of our knowledge, we are the first to incorporate feature disentanglement and attribution methods to interpret the bias-related components within diffusion models, providing deeper insights into the internal workings of diffusion process." + }, + { + "type": "title", + "bbox": [ + 0.09, + 0.22, + 0.453, + 0.237 + ], + "angle": 0, + "content": "4.3. Intervening in Bias Features for Debiasing" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.244, + 0.483, + 0.379 + ], + "angle": 0, + "content": "To control and mitigate social bias in diffusion models, we propose an effective method via intervening in the above discovered bias features indexed with \\( i \\in \\mathbf{A} \\), where \\( \\mathbf{A} \\) consists of indexes of the identified bias features. It is possible to suppress or amplify these features to exert control over the model's behavior [12]. Unlike other methods [32, 33, 43], we only intervene in the identified features indexed with \\( \\mathbf{A} \\) rather than the entire semantic space. Our intervention approach are illustrated in part (c) of Fig. 1." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.38, + 0.483, + 0.471 + ], + "angle": 0, + "content": "Specifically, we apply an intervention operation to the target features in the sparse semantic space. Then these features are reconstructed into the original latent space for subsequent image generation and more details are in Appendix A. For each feature indexed within \\(\\mathbf{A}\\), we can either apply a scalar multiplication or adding an additive bias as" + }, + { + "type": "equation", + "bbox": [ + 0.09, + 0.478, + 0.483, + 0.53 + ], + "angle": 0, + "content": "\\[\ns _ {i} = \\operatorname {I n t e r v e n e} \\left(s _ {i}\\right) = \\left\\{ \\begin{array}{l l} \\beta s _ {i} & (\\text {S c a l i n g}), \\text {o r} \\\\ s _ {i} + \\beta & (\\text {A d d i n g}). \\forall i \\in \\mathbf {A}, \\beta \\in \\mathbb {R}. \\end{array} \\right. \\tag {7}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.531, + 0.483, + 0.576 + ], + "angle": 0, + "content": "By applying a distinct parameter \\(\\beta\\), we can suppress or amplify the target bias feature to adjust the bias level or even mitigate it for improved balance." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.577, + 0.483, + 0.682 + ], + "angle": 0, + "content": "Our DIFFLENS demonstrates that it can control over the generation and representation of biased features in the output images as shown in Sec. 5.4. Furthermore, in Sec. 5.6, we gain more valuable insights into how different bias features are embedded within the inner working of diffusion model, allowing us not only to debias effectively but also to better understand the model's decision-making processes." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.694, + 0.222, + 0.71 + ], + "angle": 0, + "content": "5. Experiments" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.719, + 0.483, + 0.748 + ], + "angle": 0, + "content": "In this section, we evaluate our DIFFLENS qualitatively and quantitatively, following the below research questions." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.749, + 0.483, + 0.794 + ], + "angle": 0, + "content": "RQ1: How effectively does DIFFLENS mitigate social bias while maintaining image quality? RQ2: Can DIFFLENS accurately identify intrinsic mechanism of bias generation?" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.795, + 0.483, + 0.84 + ], + "angle": 0, + "content": "RQ3: How well does it control the bias level through intervening in the found bias features? RQ4: Is it able to interpret bias feature in a fine-grained way?" + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.848, + 0.298, + 0.864 + ], + "angle": 0, + "content": "5.1. Experimental Settings" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.871, + 0.483, + 0.901 + ], + "angle": 0, + "content": "Diffusion Model Architecture. We evaluate our proposed method on two pretrained diffusion models, (i) P2 [10]," + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.092, + 0.906, + 0.137 + ], + "angle": 0, + "content": "an unconditional diffusion model trained on CelebA-HQ dataset [26]; (ii) Stable Diffusion [49], a text-conditional diffusion model trained on LAION dataset [52]." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.138, + 0.907, + 0.365 + ], + "angle": 0, + "content": "Baselines. We compare our method against several baselines that are categorized into interpretability and guidance-based methods. For interpretability methods, we construct Activation [6] trying to locate influential neurons according to the neuron activation values. For guidance-based methods, we include Latent Editing [32] and H-Distribution [43] (abbreviated as \"L.E.\" and \"H-Dist.\"), trying to learn a latent vector to guide the unbiased generation within the bottleneck layer of U-Net. We additionally compare two methods, specifically for text-conditional diffusion models, which are Latent Direction [33] and Finetuning [54]. Due to Activation's limited debiasing effectiveness across both diffusion models and its restricted control over bias levels, we report its results only in Tab. 1 and Fig. 2. More details about baselines are in Appendix C.3." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.366, + 0.907, + 0.577 + ], + "angle": 0, + "content": "Implementation Details. To ensure fair comparison, and without loss of generality, we constrain our method, DIFFLENS, and other benchmark methods to operate on the bottleneck layer of the U-Net [50] in the diffusion model, adhering to the baseline settings in [32, 43]. Notably, DIFFLENS can easily extend the targeted mechanism to broader regions within the model. Following Sec. 4, we train kSAEs on the target latent space in the diffusion model. For the training of the light-weight classifier, we follow the settings in [43]. For the support set of samples used for identifying bias feature, we leverage the diffusion models to generate 1,000 sample for each attribute. We provide more details of implementation and dataset we use in Appendix B. The reproducible codes are in supplementary material." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.578, + 0.906, + 0.79 + ], + "angle": 0, + "content": "Evaluation Metrics. To measure the bias level of generated content, we use the Fairness Discrepancy (FD) introduced in [43] which calculates the Euclidean distance between a reference distribution (typically uniform) and the bias distribution of generation, representing class distributions within an attribute (e.g., male and female for gender attribute). A lower FD score indicates more balanced outputs. We also measure image quality using Fréchet Inception Distance (FID) [22] on the generated images. In addition, we follow [54] to measure (i) CLIP-T, which calculates the CLIP [48] based semantic similarity between the generated image and the input text prompt, and (ii) CLIP-I, which assesses the similarity between originally generated images and images after debiasing. More details are in Appendix C." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.816, + 0.762, + 0.832 + ], + "angle": 0, + "content": "5.2. Debiasing Diffusion Models" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.84, + 0.906, + 0.901 + ], + "angle": 0, + "content": "We define \"bias mitigation\" task as adjusting the diffusion model to produce balanced outputs across specific social attributes (e.g., gender). We focus on the task across gender, age, and race attributes. Gender includes male and female" + }, + { + "type": "page_number", + "bbox": [ + 0.482, + 0.945, + 0.516, + 0.957 + ], + "angle": 0, + "content": "8196" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.095, + 0.089, + 0.251, + 0.211 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.121, + 0.212, + 0.222, + 0.231 + ], + "angle": 0, + "content": "Original M:F(%) 33.9:66.1" + }, + { + "type": "image", + "bbox": [ + 0.258, + 0.089, + 0.414, + 0.211 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.285, + 0.212, + 0.386, + 0.231 + ], + "angle": 0, + "content": "Activation M:F(%) 36.6:63.4" + }, + { + "type": "image", + "bbox": [ + 0.42, + 0.089, + 0.576, + 0.211 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.448, + 0.212, + 0.549, + 0.231 + ], + "angle": 0, + "content": "Latent Editing \nM:F(%) 50.2:49.8" + }, + { + "type": "image", + "bbox": [ + 0.584, + 0.089, + 0.74, + 0.211 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.611, + 0.212, + 0.712, + 0.231 + ], + "angle": 0, + "content": "H-Distribution M:F(%) 46.6:53.4" + }, + { + "type": "image", + "bbox": [ + 0.747, + 0.089, + 0.903, + 0.211 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.775, + 0.212, + 0.875, + 0.231 + ], + "angle": 0, + "content": "DiffLens (Ours) M:F(%) 49.9:50.1" + }, + { + "type": "image_caption", + "bbox": [ + 0.09, + 0.243, + 0.907, + 0.285 + ], + "angle": 0, + "content": "Figure 2. Comparison of randomly sampled original and debiased images from various baseline methods. The minority group (male) is highlighted with red bounding boxes for easier viewing. \"M:F\" refers to the male-to-female ratio. Our DIFFLENS effectively mitigates bias while preserving high generation quality, whereas other methods either struggle to maintain balance or produce images with artifacts." + }, + { + "type": "table", + "bbox": [ + 0.135, + 0.298, + 0.863, + 0.399 + ], + "angle": 0, + "content": "
CategoryMethodGender (2)Age (3)Race (4)
FD ↓FID ↓CLIP-I ↑FD ↓FID ↓CLIP-I ↑FD ↓FID ↓CLIP-I ↑
Original0.22633.38-0.59233.38-0.71833.38-
Guidance-basedLatent Editing [32]0.00329.850.94740.60633.710.90920.31734.610.9081
H-Distribution [43]0.04831.310.94400.51134.210.85940.49436.190.8762
InterpretabilityActivation [6]0.19034.270.80600.54448.910.77930.70046.130.7846
DIFFLENS (Ours)0.00231.930.94790.40131.710.94140.44733.470.9111
" + }, + { + "type": "table_caption", + "bbox": [ + 0.09, + 0.409, + 0.907, + 0.438 + ], + "angle": 0, + "content": "Table 1. Evaluation of bias mitigation in unconditional diffusion model P2 [10]. The numbers beside each attribute indicate the number of classes for that attribute. We highlight the best results in bold and the second best with underlined format (excluding the \"original\")." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.451, + 0.483, + 0.481 + ], + "angle": 0, + "content": "classes; age includes young, adult, and old; and race includes white, black, Asian, and Indian classes." + }, + { + "type": "text", + "bbox": [ + 0.089, + 0.485, + 0.483, + 0.651 + ], + "angle": 0, + "content": "Intuitively, bias mitigation should yield balanced generation without compromising the overall semantics of the generated content. To comprehensively evaluate the effectiveness of bias mitigation (RQ1) across different methods, we assess (i) the bias level using the FD score of generated images, (ii) the generation quality via the FID score, and (iii) the preservation of overall image semantics after bias mitigation using CLIP-I/T scores. Beyond evaluating on the unconditional generative model, we also extend our analysis to a text-conditional one to showcase the broader applicability and generalizability of our method." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.656, + 0.483, + 0.837 + ], + "angle": 0, + "content": "Qualitative Results on Bias Mitigation. We present batches of generated examples before and after debiasing gender attribute to visually evaluate bias level and generation quality, as shown in Fig. 2. Our DIFFLENS effectively neutralizes gender bias while ensuring the generated images remain realistic and intact. In comparison, Latent Editing may generate unrealistic images, for example, the second one in the first row in Fig. 2, H-Distribution struggles to generate a well-balanced output, and Activation produces imbalanced, distorted images. We also provide visual results for text-to-image diffusion model Stable Diffusion [49] in Appendix D.2." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.84, + 0.483, + 0.902 + ], + "angle": 0, + "content": "Quantitative Analysis on Bias Mitigation. We assess baseline methods on debiasing across various social attributes (e.g., gender) by generating balanced outputs. As shown in Tab. 1, our DIFFLENS demonstrates competitively" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.451, + 0.907, + 0.587 + ], + "angle": 0, + "content": "strong results in (i) bias mitigation (measured by FD), (ii) in preserving semantic feature coherence (measured by CLIP-I) and (iii) maintaining generation quality (measured by FID). Activation shows poor ability to mitigate social bias. Although Latent Editing performs well in mitigating gender bias and achieves favorable FID, it can produce images with artifacts, as shown in Figs. 2 and 3. H-Distribution struggles to generate well-balanced outputs, as reflected by its poor FD scores." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.589, + 0.909, + 0.772 + ], + "angle": 0, + "content": "In Tab. 2, we show results on a text-to-image model, Stable Diffusion [49]. Our DIFFLENS effectively mitigates bias, especially for gender and age, while effectively maintaining generation quality and image semantics among methods that are effective in bias mitigation. Latent Direction achieves good debiasing for age and race but often overcorrects, reducing CLIP-I scores and compromising the balance between target generation ratio and image quality. Finetuning, demanding extensive effort, excels in gender and race but struggles with age bias mitigation. Other baselines either handle multi-class attributes poorly or lack generalizability to text-to-image diffusion models." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.785, + 0.878, + 0.8 + ], + "angle": 0, + "content": "5.3. Accurate Identification of Bias Mechanism" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.81, + 0.907, + 0.903 + ], + "angle": 0, + "content": "Accurate bias mechanism discovery could provide feasible control on the generated contents of diffusion, regarding the level of bias attributes. For instance, unlike the original diffusion model P2 [10], which generates a male-to-female gender ratio of approximately 3:7, our method easily attains the reverse ratio of 7:3. This capability goes beyond merely" + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.516, + 0.957 + ], + "angle": 0, + "content": "8197" + } + ], + [ + { + "type": "table", + "bbox": [ + 0.094, + 0.088, + 0.905, + 0.194 + ], + "angle": 0, + "content": "
MethodGender (2)Age (3)Race (4)
FD ↓FID ↓CLIP-I ↑CLIP-T ↑FD ↓FID ↓CLIP-I ↑CLIP-T ↑FD ↓FID ↓CLIP-I ↑CLIP-T ↑
Original0.564120.06-0.61550.752120.06-0.61550.558120.06-0.6155
Latent Editing [32]0.408166.110.82530.60050.682200.900.85270.61220.524153.050.88040.6086
H-Distribution [43]0.222151.680.84750.60870.506147.710.83450.60980.544126.900.82550.6100
Latent Direction [33]0.305129.370.80580.60910.052113.810.81510.60670.175128.300.82110.6132
Fintuning [54]0.050161.470.87790.60950.746161.470.87790.60950.198161.470.87790.6095
DIFFLENS (Ours)0.046112.830.85010.60900.04999.170.87780.60570.401119.860.90960.6149
" + }, + { + "type": "table_caption", + "bbox": [ + 0.09, + 0.204, + 0.907, + 0.233 + ], + "angle": 0, + "content": "Table 2. Evaluation of bias mitigation in text-to-image diffusion model Stable Diffusion [49], based on average performance across four occupations (see Appendix B). We highlight the best results in bold and the second-best with underlined text (excluding the \"original\")." + }, + { + "type": "table", + "bbox": [ + 0.094, + 0.238, + 0.905, + 0.309 + ], + "angle": 0, + "content": "
OriginalL.E.H-Dist.DiffLensOriginalL.E.H-Dist.DiffLensOriginalL.E.H-Dist.DiffLens
" + }, + { + "type": "image_caption", + "bbox": [ + 0.089, + 0.318, + 0.907, + 0.348 + ], + "angle": 0, + "content": "Figure 3. Comparison in accurate identification of bias features. Our DIFFLENS preserves overall image semantics such as smile and eyeglasses while other methods frequently introduce distortions or lose important details." + }, + { + "type": "text", + "bbox": [ + 0.089, + 0.36, + 0.485, + 0.618 + ], + "angle": 0, + "content": "balancing generated images in Sec. 5.2; it allows us to examine how effectively each method edits the target attribute, e.g., gender, while preserving non-target semantic features, e.g., eyeglass and smile (RQ2). For illustration, we visualize the generated contents of compared methods upon the reversed ratio of gender attribute in Fig. 3. More examples and details are provided in Appendix D.4.1. As is shown in Fig. 3, our method can preserve non-target attributes such as smile, background coherence, and even eyeglasses when debiasing, showcasing accurate identification of the bias generation mechanism by our method. While the baseline methods often introduce distortions or compromise detail in these areas, without looking into inner working of bias generation. Additionally, from Tabs. 1 and 2, our DIFFLENS consistently achieves high CLIP-I scores, demonstrating its ability to accurately identify and modify target bias specific features with minimal impact on unrelated content." + }, + { + "type": "title", + "bbox": [ + 0.09, + 0.634, + 0.47, + 0.651 + ], + "angle": 0, + "content": "5.4. Fine-grained Bias Level Control and Editing" + }, + { + "type": "text", + "bbox": [ + 0.089, + 0.66, + 0.483, + 0.901 + ], + "angle": 0, + "content": "Benefiting from bias mechanism identification, we can control bias level with finer granularity and across a broader range (RQ3). Figure 5 illustrates how different methods progressively modulate image generation in diffusion along two opposite gender directions. More examples are in Appendix D.4.2. Our DIFFLENS produces natural transitions, consistently preserving semantic features such as facial expression and maintaining high generation quality. Examples about control in multi-attributes are in Appendix D.4.3. The results of varying gender ratios generated by each method are shown in Fig. 4. Latent Editing and H-Distribution perform well when the gender ratio is around balanced, but their fidelity and visual coherence deteriorate at more skewed ratios, indicating a limitation in handling highly imbalanced outputs. By adjusting the value of intervention parameter incorporated in Eq. (7), our DIFFLENS can flexi" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.36, + 0.907, + 0.452 + ], + "angle": 0, + "content": "bly control bias feature and generate different gender ratios, ranging from balanced to imbalanced ones (almost male or female), illustrating powerful controllability over the representations of bias. Such highly imbalanced ratios also have wide applications in real world such as data augmentation for under-represented subgroups [43]." + }, + { + "type": "image", + "bbox": [ + 0.515, + 0.463, + 0.905, + 0.565 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.512, + 0.574, + 0.907, + 0.645 + ], + "angle": 0, + "content": "Figure 4. Achievable range of bias level control for each method. The Log Gender Ratio reflects the log of male to female ratio in the generated images, with 0 indicating balance. Our DIFFLENS offers broader bias control, delivering higher generation quality (left plot) and improved semantic coherence (right plot)." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.651, + 0.669, + 0.668 + ], + "angle": 0, + "content": "5.5. Ablation Study" + }, + { + "type": "text", + "bbox": [ + 0.511, + 0.675, + 0.907, + 0.903 + ], + "angle": 0, + "content": "We conduct an ablation study to reveal how each component in our DIFFLENS takes effect. Recall that, it dissects the diffusion models by (i) disentangling the latent space to a sparse semantic space via k-SAE and (ii) identifying the bias features using bias attribution. In Tab. 3, directly selecting neurons with the highest activation values as bias features (Activation method) performs poorly, supporting the claim in Sec. 4.1. Comparing \"Act@L\" with \"Act@S\" and \"Attr@L\" with DIFFLENS, using k-SAE improves visual coherence, as seen in the higher CLIP-I scores (0.8597 to 0.9479 and 0.8060 to 0.9368), while avoiding direct edits on polysemantic neurons mitigates FID degradation (41.68 to 31.93). Furthermore, comparing \"Act@L\" with \"Attr@L\" and \"Act@\" with DIFFLENS, bias attribution enables more precise localization of bias features, significantly lowering" + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.516, + 0.957 + ], + "angle": 0, + "content": "8198" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.093, + 0.088, + 0.468, + 0.23 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.476, + 0.089, + 0.906, + 0.23 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.09, + 0.243, + 0.908, + 0.286 + ], + "angle": 0, + "content": "Figure 5. Comparison of individual image transformations along the gender axis, with columns displaying images sampled from progressively shifting ratios as in Fig. 4. Our DIFFLENS achieves smooth and consistent transitions that preserve semantic features like facial expressions, while other methods show distortions and loss of detail at higher imbalance ratios." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.298, + 0.483, + 0.343 + ], + "angle": 0, + "content": "FD (0.190 to 0.072 and 0.303 to 0.002). DIFFLENS incorporating both shows superiority in bias mitigation with concise localization and intervention of bias features." + }, + { + "type": "table", + "bbox": [ + 0.094, + 0.354, + 0.48, + 0.449 + ], + "angle": 0, + "content": "
ComponentsOrig.Act@LAct@SAttr@LDIFFLENS
ActivationX✓✓XX
K-SAEXX✓X✓
Bias Attr.XXX✓✓
FD ↓0.2260.1900.3030.0720.002
FID ↓33.3834.2732.7041.6831.93
CLIP-I ↑-0.80600.93860.85970.9479
" + }, + { + "type": "table_footnote", + "bbox": [ + 0.094, + 0.45, + 0.429, + 0.462 + ], + "angle": 0, + "content": "- ①, ②: Bias attribution allows precise bias mitigation." + }, + { + "type": "table_footnote", + "bbox": [ + 0.094, + 0.462, + 0.482, + 0.49 + ], + "angle": 0, + "content": "- ③, ④: K-SAE improves visual feature coherence and maintains generation quality, via disentangling original latent space." + }, + { + "type": "list", + "bbox": [ + 0.094, + 0.45, + 0.482, + 0.49 + ], + "angle": 0, + "content": null + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.493, + 0.484, + 0.549 + ], + "angle": 0, + "content": "Table 3. Ablation study of DIFFLENS in P2 [10] for gender attribute. We abbreviate Original as \"Orig.\", neuron activations as \"Act.\", and bias attributions as \"Attr\". \"@L\" denotes operations on the original latent space, and \"@S\" on sparse semantic space." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.56, + 0.419, + 0.575 + ], + "angle": 0, + "content": "5.6. Wider Investigation: Bias Mechanism" + }, + { + "type": "text", + "bbox": [ + 0.089, + 0.584, + 0.483, + 0.902 + ], + "angle": 0, + "content": "We further look into the bias mechanism, i.e., the identified bias features \\(\\{s_i|i\\in \\mathbf{A}\\}\\), to reveal its granular aspects on social attribute generation and how these features distinguishably affect the biased contents, shedding some lights on the inner workings of diffusion models (RQ4). We observe from Fig. 6 that (i) multiple features may jointly influence a single social bias attribute, such as \"short hair\" and \"mustache\" contributing to gender perception; (ii) some features overlap across biases, for instance, \"side pose\" being associated with both age and gender biases as it often appears in images of older males; and (iii) poses like \"side pose\" and \"head down\" reflect in photography, often aligning with gender norms that men are depicted in dynamic or indirect poses, while women appear younger and more polished, reflecting social expectations [15, 58, 60]. These stereotypes mirror underlying social biases, impacting not only how individuals are portrayed but also how their images are interpreted. Our DIFFLENS serves as a lens that uncovers social biases and stereotypes embedded in generated images, providing insight into how attributes such as gender, race, and age are systematically portrayed." + }, + { + "type": "title", + "bbox": [ + 0.514, + 0.297, + 0.633, + 0.312 + ], + "angle": 0, + "content": "6. Conclusion" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.322, + 0.907, + 0.596 + ], + "angle": 0, + "content": "In this work, we introduced a novel approach DIFFLENS, for mitigating social biases in diffusion models by dissecting, analyzing and intervening in the internal mechanisms of the diffusion model. Leveraging k-SAE and gradient-based attribution method, we disentangle, identify, and control bias-related features with precision, allowing targeted bias mitigation while preserving non-target attributes. A wide range of evaluations across social attributes such as gender, age, and race validate effectiveness of DIFFLENS in bias control and content preservation. Our work offers an interpretable solution to bias mitigation in diffusion models, though challenges remain, including the precise identification of certain racial features and a quality decline in intensified edits. Future directions include expanding this approach to other bias types, incorporating additional fairness perspectives, and exploring application across diverse model architectures like Diffusion Transformers [46] for broader impact in ethical generative AI deployment." + }, + { + "type": "image", + "bbox": [ + 0.543, + 0.608, + 0.875, + 0.86 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.513, + 0.86, + 0.908, + 0.901 + ], + "angle": 0, + "content": "Figure 6. Manually analyzing the discovered bias features in diffusion models. Each column represents one specific feature \\( s_i \\), \\( i \\in \\mathbf{A} \\) and each row originates from the same raw image." + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.516, + 0.957 + ], + "angle": 0, + "content": "8199" + } + ], + [ + { + "type": "title", + "bbox": [ + 0.092, + 0.091, + 0.262, + 0.108 + ], + "angle": 0, + "content": "7. Acknowledgment" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.114, + 0.484, + 0.213 + ], + "angle": 0, + "content": "The research received partial support from National Natural Science Foundation of China (Grant No. 62406193). The authors also gratefully acknowledge further assistance provided by Shanghai Frontiers Science Center of Human-centered Artificial Intelligence, MoE Key Lab of Intelligent Perception and Human-Machine Collaboration, and HPC Platform of ShanghaiTech University." + }, + { + "type": "title", + "bbox": [ + 0.093, + 0.241, + 0.188, + 0.256 + ], + "angle": 0, + "content": "References" + }, + { + "type": "ref_text", + "bbox": [ + 0.101, + 0.265, + 0.483, + 0.336 + ], + "angle": 0, + "content": "[1] Hritik Bansal, Da Yin, Masoud Monajatipoor, and Kai-Wei Chang. How well can text-to-image generative models understand ethical natural language interventions? In Proceedings of the 2022 Conference on Empirical Methods in Natural Language Processing, pages 1358–1370, 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.338, + 0.484, + 0.405 + ], + "angle": 0, + "content": "[2] Abhipsa Basu, R Venkatesh Babu, and Danish Pruthi. Inspecting the geographical representativeness of images from text-to-image models. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 5136-5147, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.408, + 0.483, + 0.478 + ], + "angle": 0, + "content": "[3] Sebastian Benthall and Bruce D. Haynes. Racial categories in machine learning. In Proceedings of the Conference on Fairness, Accountability, and Transparency, page 289-298, New York, NY, USA, 2019. Association for Computing Machinery. 11" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.479, + 0.483, + 0.533 + ], + "angle": 0, + "content": "[4] Leonard Bereska and Stratis Gavves. Mechanistic interpretability for AI safety - a review. Transactions on Machine Learning Research, 2024. Survey Certification, Expert Certification. 1, 3, 11" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.535, + 0.484, + 0.631 + ], + "angle": 0, + "content": "[5] Federico Bianchi, Pratyusha Kalluri, Esin Durmus, Faisal Ladhak, Myra Cheng, Debora Nozza, Tatsunori Hashimoto, Dan Jurafsky, James Zou, and Aylin Caliskan. Easily accessible text-to-image generation amplifies demographic stereotypes at large scale. In Proceedings of the 2023 ACM Conference on Fairness, Accountability, and Transparency, pages 1493-1504, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.633, + 0.482, + 0.716 + ], + "angle": 0, + "content": "[6] Steven Bills, Nick Cammarata, Dan Mossing, Henk Tillman, Leo Gao, Gabriel Goh, Ilya Sutskever, Jan Leike, Jeff Wu, and William Saunders. Language models can explain neurons in language models. https://openaipublic.blob.core.windows.net/neuron-explainer/paper/index.html, 2023.5, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.718, + 0.483, + 0.773 + ], + "angle": 0, + "content": "[7] Hila Chefer, Oran Lang, Mor Geva, Volodymyr Polosukhin, Assaf Shocher, Inbar Mosseri, Lior Wolf, et al. The hidden language of diffusion models. In *The Twelfth International Conference on Learning Representations*. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.775, + 0.483, + 0.83 + ], + "angle": 0, + "content": "[8] Hila Chefer, Yuval Alaluf, Yael Vinker, Lior Wolf, and Daniel Cohen-Or. Attend-and-excite: Attention-based semantic guidance for text-to-image diffusion models. ACM Transactions on Graphics (TOG), 42(4):1-10, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.832, + 0.483, + 0.9 + ], + "angle": 0, + "content": "[9] Jaemin Cho, Abhay Zala, and Mohit Bansal. Dall-eval: Probing the reasoning skills and social biases of text-to-image generation models. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 3043-3054, 2023. 2" + }, + { + "type": "list", + "bbox": [ + 0.101, + 0.265, + 0.484, + 0.9 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.093, + 0.907, + 0.162 + ], + "angle": 0, + "content": "[10] Jooyoung Choi, Jungbeom Lee, Chaehun Shin, Sungwon Kim, Hyunwoo Kim, and Sungroh Yoon. Perception prioritized training of diffusion models. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 11472-11481, 2022. 5, 6, 8, 1, 2, 3, 4, 9" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.164, + 0.906, + 0.217 + ], + "angle": 0, + "content": "[11] Ching-Yao Chuang, Varun Jampani, Yuanzhen Li, Antonio Torralba, and Stefanie Jegelka. Debiasing vision-language models via biased prompts. arXiv preprint arXiv:2302.00070, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.219, + 0.906, + 0.302 + ], + "angle": 0, + "content": "[12] Damai Dai, Li Dong, Yaru Hao, Zhifang Sui, Baobao Chang, and Furu Wei. Knowledge neurons in pretrained transformers. In Proceedings of the 60th Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers), pages 8493-8502, Dublin, Ireland, 2022. Association for Computational Linguistics. 3, 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.303, + 0.906, + 0.344 + ], + "angle": 0, + "content": "[13] Kedar Dhamhere, Mukund Sundararajan, and Qiqi Yan. How important is a neuron. In International Conference on Learning Representations, 2019. 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.345, + 0.906, + 0.412 + ], + "angle": 0, + "content": "[14] Nelson Elhage, Tristan Hume, Catherine Olsson, Nicholas Schiefer, Tom Henighan, Shauna Kravec, Zac Hatfield-Dodds, Robert Lasenby, Dawn Drain, Carol Chen, et al. Toy models of superposition. arXiv preprint arXiv:2209.10652, 2022.3" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.414, + 0.905, + 0.455 + ], + "angle": 0, + "content": "[15] Emilio Ferrara. Fairness and bias in artificial intelligence: A brief survey of sources, impacts, and mitigation strategies. Sci, 6(1), 2024. 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.457, + 0.906, + 0.524 + ], + "angle": 0, + "content": "[16] Felix Friedrich, Manuel Brack, Lukas Struppek, Dominik Hintersdorf, Patrick Schramowski, Sasha Luccioni, and Kristian Kersting. Fair diffusion: Instructing text-to-image generation models on fairness. arXiv preprint arXiv:2302.10893, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.526, + 0.906, + 0.593 + ], + "angle": 0, + "content": "[17] Rohit Gandikota, Hadas Orgad, Yonatan Belinkov, Joanna Materzyńska, and David Bau. Unified concept editing in diffusion models. In Proceedings of the IEEE/CVF Winter Conference on Applications of Computer Vision, pages 5111-5120, 2024. 1, 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.596, + 0.906, + 0.678 + ], + "angle": 0, + "content": "[18] René Haas, Inbar Huberman-Spiegelglas, Rotem Mulayoff, Stella Graßhof, Sami S Brandt, and Tomer Michaeli. Discovering interpretable directions in the semantic latent space of diffusion models. In 2024 IEEE 18th International Conference on Automatic Face and Gesture Recognition (FG), pages 1-9. IEEE, 2024. 2, 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.679, + 0.906, + 0.748 + ], + "angle": 0, + "content": "[19] Ligong Han, Yinxiao Li, Han Zhang, Peyman Milanfar, Dimitris Metaxas, and Feng Yang. Svdiff: Compact parameter space for diffusion fine-tuning. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 7323-7334, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.749, + 0.906, + 0.845 + ], + "angle": 0, + "content": "[20] Ruifei He, Chuhui Xue, Haoru Tan, Wenqing Zhang, Yingchen Yu, Song Bai, and Xiaojuan Qi. Debiasing text-to-image diffusion models. In Proceedings of the 1st ACM Multimedia Workshop on Multi-Modal Misinformation Governance in the Era of Foundation Models, page 29–36, New York, NY, USA, 2024. Association for Computing Machinery. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.847, + 0.905, + 0.901 + ], + "angle": 0, + "content": "[21] Amir Hertz, Ron Mokady, Jay Tenenbaum, Kfir Aberman, Yael Pritch, and Daniel Cohen-or. Prompt-to-prompt image editing with cross-attention control. In The Eleventh International Conference on Learning Representations, 2023. 2" + }, + { + "type": "list", + "bbox": [ + 0.517, + 0.093, + 0.907, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.516, + 0.956 + ], + "angle": 0, + "content": "8200" + } + ], + [ + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.093, + 0.482, + 0.161 + ], + "angle": 0, + "content": "[22] Martin Heusel, Hubert Ramsauer, Thomas Unterthiner, Bernhard Nessler, and Sepp Hochreiter. Gans trained by a two time-scale update rule converge to a local nash equilibrium. Advances in neural information processing systems, 30, 2017. 5, 11" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.164, + 0.482, + 0.219 + ], + "angle": 0, + "content": "[23] Jonathan Ho, Ajay Jain, and Pieter Abbeel. Denoising diffusion probabilistic models. In Advances in Neural Information Processing Systems, pages 6840-6851. Curran Associates, Inc., 2020. 1, 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.221, + 0.482, + 0.29 + ], + "angle": 0, + "content": "[24] Robert Huben, Hoagy Cunningham, Logan Riggs Smith, Aidan Ewart, and Lee Sharkey. Sparse autoencoders find highly interpretable features in language models. In The Twelfth International Conference on Learning Representations, 2024. 3, 4, 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.292, + 0.482, + 0.36 + ], + "angle": 0, + "content": "[25] Kimmo Karkkainen and Jungseock Joo. Fairface: Face attribute dataset for balanced race, gender, and age for bias measurement and mitigation. In Proceedings of the IEEE/CVF winter conference on applications of computer vision, pages 1548-1558, 2021. 2, 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.362, + 0.482, + 0.417 + ], + "angle": 0, + "content": "[26] Tero Karras, Timo Aila, Samuli Laine, and Jaakko Lehtinen. Progressive growing of GANs for improved quality, stability, and variation. In International Conference on Learning Representations, 2018. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.42, + 0.482, + 0.475 + ], + "angle": 0, + "content": "[27] Tero Karras, Samuli Laine, and Timo Aila. A style-based generator architecture for generative adversarial networks. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 4401-4410, 2019. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.477, + 0.482, + 0.531 + ], + "angle": 0, + "content": "[28] Eunji Kim, Siwon Kim, Chaehun Shin, and Sungroh Yoon. De-stereotyping text-to-image models through prompt tuning. In ICML Workshop on Challenges in Deployable Generative AI, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.533, + 0.482, + 0.601 + ], + "angle": 0, + "content": "[29] Gwanghyun Kim, Taesung Kwon, and Jong Chul Ye. Diffus ionclip: Text-guided diffusion models for robust image man i plication. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 2426-2435, 2022. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.604, + 0.482, + 0.672 + ], + "angle": 0, + "content": "[30] Yeongmin Kim, Byeonghu Na, Minsang Park, JoonHo Jang, Dongjun Kim, Wanmo Kang, and Il chul Moon. Training unbiased diffusion models from biased dataset. In The Twelfth International Conference on Learning Representations, 2024. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.675, + 0.482, + 0.731 + ], + "angle": 0, + "content": "[31] Xianghao Kong, Ollie Liu, Han Li, Dani Yogatama, and Greg Ver Steeg. Interpretable diffusion via information decomposition. In The Twelfth International Conference on Learning Representations. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.732, + 0.482, + 0.787 + ], + "angle": 0, + "content": "[32] Mingi Kwon, Jaeseok Jeong, and Youngjung Uh. Diffusion models already have a semantic latent space. In The Eleventh International Conference on Learning Representations, 2023. 2, 3, 4, 5, 6, 7, 9" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.789, + 0.482, + 0.87 + ], + "angle": 0, + "content": "[33] Hang Li, Chengzhi Shen, Philip Torr, Volker Tresp, and Jindong Gu. Self-discovering interpretable diffusion latent directions for responsible text-to-image generation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 12006-12016, 2024. 2, 4, 5, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.873, + 0.482, + 0.901 + ], + "angle": 0, + "content": "[34] Sasha Luccioni, Christopher Akiki, Margaret Mitchell, and Yacine Jernite. Stable bias: Evaluating societal representa" + }, + { + "type": "list", + "bbox": [ + 0.093, + 0.093, + 0.482, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.545, + 0.093, + 0.905, + 0.12 + ], + "angle": 0, + "content": "tions in diffusion models. Advances in Neural Information Processing Systems, 36, 2024. 1, 2, 11" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.122, + 0.905, + 0.162 + ], + "angle": 0, + "content": "[35] Alireza Makhzani and Brendan Frey. K-sparse autoencoders. In International Conference on Learning Representations, 2014. 2, 3, 1, 6, 11" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.164, + 0.905, + 0.232 + ], + "angle": 0, + "content": "[36] Vongani H Maluleke, Neerja Thakkar, Tim Brooks, Ethan Weber, Trevor Darrell, Alexei A Efros, Angjoo Kanazawa, and Devin Guillery. Studying bias in gans through the lens of race. In European Conference on Computer Vision, pages 344-360. Springer, 2022. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.234, + 0.905, + 0.315 + ], + "angle": 0, + "content": "[37] Pietro Melzi, Christian Rathgeb, Ruben Tolosana, Ruben Vera-Rodriguez, Dominik Lawatsch, Florian Domin, and Maxim Schaubert. Gandifface: Controllable generation of synthetic datasets for face recognition with realistic variations. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 3086-3095, 2023. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.317, + 0.905, + 0.37 + ], + "angle": 0, + "content": "[38] Kevin Meng, David Bau, Alex Andonian, and Yonatan Belinkov. Locating and editing factual associations in gpt. Advances in Neural Information Processing Systems, 35: 17359-17372, 2022. 1, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.372, + 0.905, + 0.425 + ], + "angle": 0, + "content": "[39] Jesse Mu and Jacob Andreas. Compositional explanations of neurons. In Advances in Neural Information Processing Systems, pages 17153-17163. Curran Associates, Inc., 2020. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.428, + 0.905, + 0.482 + ], + "angle": 0, + "content": "[40] Ranjita Naik and Besmira Nushi. Social biases through the text-to-image generation lens. In Proceedings of the 2023 AAAI/ACM Conference on AI, Ethics, and Society, pages 786-808, 2023. 2, 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.484, + 0.905, + 0.524 + ], + "angle": 0, + "content": "[41] Chris Olah, Nick Cammarata, Ludwig Schubert, Gabriel Goh, Michael Petrov, and Shan Carter. Zoom in: An introduction to circuits. Distill, 5(3):e00024-001, 2020. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.526, + 0.905, + 0.58 + ], + "angle": 0, + "content": "[42] Hadas Orgad, Bahjat Kawar, and Yonatan Belinkov. Editing implicit assumptions in text-to-image diffusion models. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 7053-7061, 2023. 1, 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.582, + 0.905, + 0.663 + ], + "angle": 0, + "content": "[43] Rishubh Parihar, Abhijnya Bhat, Abhipsa Basu, Saswat Mallick, Jogendra Nath Kundu, and R Venkatesh Babu. Balancing act: Distribution-guided debiasing in diffusion models. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 6668-6678, 2024. 1, 2, 4, 5, 6, 7, 9, 11" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.665, + 0.905, + 0.719 + ], + "angle": 0, + "content": "[44] Ji-Hoon Park, Yeong-Joon Ju, and Seong-Whan Lee. Explaining generative diffusion models via visual analysis for interpretable decision-making process. Expert Systems with Applications, 248:123231, 2024. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.721, + 0.905, + 0.789 + ], + "angle": 0, + "content": "[45] Yong-Hyun Park, Mingi Kwon, Jaewoong Choi, Junghyo Jo, and Youngjung Uh. Understanding the latent space of diffusion models through the lens of riemannian geometry. In Thirty-seventh Conference on Neural Information Processing Systems, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.79, + 0.905, + 0.844 + ], + "angle": 0, + "content": "[46] William Peebles and Saining Xie. Scalable diffusion models with transformers. In Proceedings of the IEEE/CVF International Conference on Computer Vision (ICCV), pages 4195-4205, 2023. 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.846, + 0.905, + 0.9 + ], + "angle": 0, + "content": "[47] Malsha V Perera and Vishal M Patel. Analyzing bias in diffusion-based face generation models. In 2023 IEEE International Joint Conference on Biometrics (IJCB), pages 1-10. IEEE, 2023. 1, 2" + }, + { + "type": "list", + "bbox": [ + 0.517, + 0.093, + 0.905, + 0.9 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.514, + 0.957 + ], + "angle": 0, + "content": "8201" + } + ], + [ + { + "type": "ref_text", + "bbox": [ + 0.092, + 0.092, + 0.486, + 0.176 + ], + "angle": 0, + "content": "[48] Alec Radford, Jong Wook Kim, Chris Hallacy, Aditya Ramesh, Gabriel Goh, Sandhini Agarwal, Girish Sastry, Amanda Askell, Pamela Mishkin, Jack Clark, et al. Learning transferable visual models from natural language supervision. In International conference on machine learning, pages 8748-8763. PMLR, 2021. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.091, + 0.177, + 0.486, + 0.247 + ], + "angle": 0, + "content": "[49] Robin Rombach, Andreas Blattmann, Dominik Lorenz, Patrick Esser, and Björn Ommer. High-resolution image synthesis with latent diffusion models. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 10684-10695, 2022. 1, 2, 5, 6, 7, 3, 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.092, + 0.247, + 0.485, + 0.329 + ], + "angle": 0, + "content": "[50] Olaf Ronneberger, Philipp Fischer, and Thomas Brox. U-net: Convolutional networks for biomedical image segmentation. In Medical image computing and computer-assisted intervention-MICCAI 2015: 18th international conference, Munich, Germany, October 5-9, 2015, proceedings, part III 18, pages 234-241. Springer, 2015. 3, 5, 1, 2, 4, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.092, + 0.33, + 0.484, + 0.385 + ], + "angle": 0, + "content": "[51] Harrison Rosenberg, Shimaa Ahmed, Guruprasad V Ramesh, Ramya Korlakai Vinayak, and Kassem Fawaz. Unbiased face synthesis with diffusion models: Are we there yet? arXiv preprint arXiv:2309.07277, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.092, + 0.386, + 0.485, + 0.469 + ], + "angle": 0, + "content": "[52] Christoph Schuhmann, Romain Beaumont, Richard Vencu, Cade Gordon, Ross Wightman, Mehdi Cherti, Theo Coombes, Aarush Katta, Clayton Mullis, Mitchell Wortsman, et al. Laion-5b: An open large-scale dataset for training next generation image-text models. Advances in Neural Information Processing Systems, 35:25278-25294, 2022. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.092, + 0.47, + 0.485, + 0.552 + ], + "angle": 0, + "content": "[53] Preethi Seshadri, Sameer Singh, and Yanai Elazar. The bias amplification paradox in text-to-image generation. In Proceedings of the 2024 Conference of the North American Chapter of the Association for Computational Linguistics: Human Language Technologies (Volume 1: Long Papers), pages 6367-6384, 2024. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.092, + 0.553, + 0.485, + 0.621 + ], + "angle": 0, + "content": "[54] Xudong Shen, Chao Du, Tianyu Pang, Min Lin, Yongkang Wong, and Mohan Kankanhalli. Finetuning text-to-image diffusion models for fairness. In The Twelfth International Conference on Learning Representations, 2024. 1, 2, 5, 7, 6, 11" + }, + { + "type": "ref_text", + "bbox": [ + 0.092, + 0.622, + 0.485, + 0.664 + ], + "angle": 0, + "content": "[55] Jiaming Song, Chenlin Meng, and Stefano Ermon. Denoising diffusion implicit models. In International Conference on Learning Representations, 2021. 1, 2, 3, 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.092, + 0.665, + 0.485, + 0.719 + ], + "angle": 0, + "content": "[56] Mukund Sundararajan, Ankur Taly, and Qiqi Yan. Axiomatic attribution for deep networks. In Proceedings of the 34th International Conference on Machine Learning-Volume 70, pages 3319–3328, 2017. 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.092, + 0.72, + 0.485, + 0.817 + ], + "angle": 0, + "content": "[57] Raphael Tang, Linqing Liu, Akshit Pandey, Zhiying Jiang, Gefei Yang, Karun Kumar, Pontus Stenetorp, Jimmy Lin, and Ferhan Ture. What the DAAM: Interpreting stable diffusion using cross attention. In Proceedings of the 61st Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers), pages 5644-5659, Toronto, Canada, 2023. Association for Computational Linguistics. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.092, + 0.818, + 0.485, + 0.859 + ], + "angle": 0, + "content": "[58] Mike Zajko. Conservative ai and social inequality: conceptualizing alternatives to bias through social theory. *Ai & Society*, 36(3):1047–1056, 2021. 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.092, + 0.86, + 0.485, + 0.902 + ], + "angle": 0, + "content": "[59] Haiyan Zhao, Hanjie Chen, Fan Yang, Ninghao Liu, Huiqi Deng, Hengyi Cai, Shuaiqiang Wang, Dawei Yin, and Mengnan Du. Explainability for large language models: A survey." + }, + { + "type": "list", + "bbox": [ + 0.091, + 0.092, + 0.486, + 0.902 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.547, + 0.093, + 0.905, + 0.12 + ], + "angle": 0, + "content": "ACM Transactions on Intelligent Systems and Technology, 15(2):1-38, 2024. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.122, + 0.907, + 0.164 + ], + "angle": 0, + "content": "[60] Mi Zhou, Vibhanshu Abhishek, Timothy Derdenger, Jaymo Kim, and Kannan Srinivasan. Bias in generative ai. arXiv preprint arXiv:2403.02726, 2024. 8" + }, + { + "type": "list", + "bbox": [ + 0.517, + 0.093, + 0.907, + 0.164 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.517, + 0.957 + ], + "angle": 0, + "content": "8202" + } + ] +] \ No newline at end of file diff --git a/2025/Dissecting and Mitigating Diffusion Bias via Mechanistic Interpretability/3f305fc0-29c2-45c0-bbbf-2f6d947ae1f1_origin.pdf b/2025/Dissecting and Mitigating Diffusion Bias via Mechanistic Interpretability/3f305fc0-29c2-45c0-bbbf-2f6d947ae1f1_origin.pdf new file mode 100644 index 0000000000000000000000000000000000000000..e92f2b5b04fddc4f653afb2a54a6cae576ca6892 --- /dev/null +++ b/2025/Dissecting and Mitigating Diffusion Bias via Mechanistic Interpretability/3f305fc0-29c2-45c0-bbbf-2f6d947ae1f1_origin.pdf @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:acf47e557894309443e0bb44730e37539960ad7f3906c2a26f8da588f7898a76 +size 7731132 diff --git a/2025/Dissecting and Mitigating Diffusion Bias via Mechanistic Interpretability/full.md b/2025/Dissecting and Mitigating Diffusion Bias via Mechanistic Interpretability/full.md new file mode 100644 index 0000000000000000000000000000000000000000..6b30fa10b7fc77c703f5ffe7b7f73abd93837e0c --- /dev/null +++ b/2025/Dissecting and Mitigating Diffusion Bias via Mechanistic Interpretability/full.md @@ -0,0 +1,320 @@ +# Dissecting and Mitigating Diffusion Bias via Mechanistic Interpretability + +Yingdong Shi $^{1*}$ Changming Li $^{1*}$ Yifan Wang $^{2}$ Yongxiang Zhao $^{1}$ Anqi Pang $^{3}$ Sibei Yang $^{1}$ Jingyi Yu $^{1}$ Kan Ren $^{1\dagger}$ + +$^{1}$ ShanghaiTech University $^{2}$ Stony Brook University $^{3}$ Tencent PCG + +{shiyd2023,lichm2024,renkan}@shanghaitech.edu.cn + +# Abstract + +Diffusion models have demonstrated impressive capabilities in synthesizing diverse content. However, despite their high-quality outputs, these models often perpetuate social biases, including those related to gender and race. These biases can potentially contribute to harmful real-world consequences, reinforcing stereotypes and exacerbating inequalities in various social contexts. While existing research on diffusion bias mitigation has predominantly focused on guiding content generation, it often neglects the intrinsic mechanisms within diffusion models that causally drive biased outputs. In this paper, we investigate the internal processes of diffusion models, identifying specific decision-making mechanisms, termed bias features, embedded within the model architecture. By directly manipulating these features, our method precisely isolates and adjusts the elements responsible for bias generation, permitting granular control over the bias levels in the generated content. Through experiments on both unconditional and conditional diffusion models across various social bias attributes, we demonstrate our method's efficacy in managing generation distribution while preserving image quality. We also dissect the discovered model mechanism, revealing different intrinsic features controlling fine-grained aspects of generation, boosting further research on mechanistic interpretability of diffusion models. The project website is at https://foundation-model-research.github.io/difflens. + +# 1. Introduction + +Diffusion models have emerged as powerful generative tools capable of generating high-quality content by iteratively denoising random noise [23, 55]. These models are now widely credited in applications such as content synthesis [49], data manipulation [29], and data augmentation [37], impacting numerous aspects of human society. However, diffusion models have been criticized for generating + +biased content [34, 43], often resulting in skewed representations of sensitive social attributes such as gender, race, and age. While biases may originate from the training data [36, 47], they can be further amplified by the learned model itself [47]. Given the extensive deployment of diffusion models, these biases may lead to detrimental effects in real-world contexts, such as reinforcing stereotypes in media representations or perpetuating inequalities in automated decision-making systems. + +Existing approaches to debias diffusion models generally fall into two main strategies. The first attempts to train unbiased models from scratch on biased datasets [30] or fine-tune the target model [54], both of which are resource-intensive and can significantly impact the model's original performance. The second aims to guide or edit the generation process [17, 42, 43], but these methods undermine the distinct internal workings of diffusion models, risking overcorrecting the model behavior or distorting non-target attributes, which affecting generation quality. In contrast, few studies have explored the intrinsic mechanisms within diffusion models that causally drive biased outcomes. + +In this work, we hypothesize that neurons collectively play a pivotal role in driving the generation of biased concepts. These neurons exhibit specific features which we identify as intrinsic mechanisms within the model that contribute to social bias. In particular, we set out to investigate these mechanisms to better understand and mitigate bias within diffusion models. + +Our approach also aligns with the latest efforts on understanding the inner workings of neural networks, which has drawn huge interest within the community [7, 44], not only due to the interpretability benefits that reveal model decision patterns [31, 59], but also for its potential to enhance control over model behaviors [4, 38]. However, the intrinsic neuron activities within diffusion models remain largely unexplored. A significant challenge stems from the generative nature of diffusion models, which complicates, if not completely prohibits, fine-grained analysis at the neuron level. The neuron polysemanticity [4], which links individual neurons to multiple unrelated concepts additionally, fur + +ther presents obstacles on mechanistic study of large-scale diffusion models. + +To investigate the underlying mechanisms of bias generation in diffusion models, we present DIFFLENS, a comprehensive framework that analyzes the architectural components to identify the most influential elements within the model. Specifically, DIFFLENS first disentangles the hidden neurons of the diffusion model into a sparse semantic feature space via a sparse autoencoder [35], which captures the patterns of neurons responsible for bias features such as gender and age. Next, DIFFLENS employs a gradient-based approach to pinpoint key components in the interpretable semantic space, identifying a minimal set of bias-specific features. We can then further control or even mitigate the bias by adjusting the effects of these bias-related features. + +DIFFLENS offers two key advantages. First, via elaborately analyzing and identifying the inner mechanisms of diffusion process, it enables fine-grained control of bias with minimal impact on other semantic attributes. Second, the identified bias features are intrinsic to the diffusion model, revealing its underlying contribution to biased generation. Both have been demonstrated in our experiments. + +In summary, our work makes several key contributions to the field of diffusion models. (i) To the best of our knowledge, this is the first work to address bias mitigation in diffusion models through the lens of mechanistic interpretability, uncovering the underlying patterns of bias generation. Our method provides new insights into how specific neuron activities contribute to biased outcomes, offering a deeper understanding beyond standard latent space analysis. (ii) We introduce a novel framework that includes semantic space construction, identification of influential features, and fine-grained control of bias generation based on the discovered features. This approach is not only effective for bias mitigation, but may also be applied to other aspects of diffusion model analysis. (iii) We conduct extensive experiments on two pretrained diffusion models, demonstrating the effectiveness of our proposed mechanism discovery in comparison to several strong baselines for diffusion bias mitigation. + +# 2. Related Work + +Bias in Diffusion Models. Diffusion-based architectures often reflect and amplify social biases [2, 34, 51], leading to potentially harmful outputs. Specifically, unconditional diffusion models exacerbate the biases present in the training data related to gender, race, and age [47]. Text-to-image models, such as Stable Diffusion [49], tend to generate images of specific gender or stereotypes when prompted with occupational terms or terms with character traits and other descriptors [5, 9, 34, 53]. + +Mitigating Diffusion Bias. Developing techniques to mitigate social biases in generative models has become increasingly crucial [33, 43]. Many current approaches fo + +cus on modifying prompts, such as adding gender-specific cues or ethical information, or using prompt learning techniques [1, 11, 16, 28]. While these methods are effective for text-to-image systems, they are limited to scenarios involving textual prompts. Another strategy involves fine-tuning diffusion models using curated datasets [54], which can effectively reduce bias but require extensive training and expensive high-quality data. Guided generation and model editing based methods [17, 20, 42, 43] provide an alternative by introducing external guidance in the model's processing [20, 43] or manipulating attention weight [17, 42]. Although these techniques avoid retraining, they undermine the intrinsic mechanisms within the model specifically related to bias generation, often resulting in imprecise edits that inadvertently affect other parts of the image and degrade quality. In contrast, our approach allows for precise identification of bias-generating features, preserving overall image quality and maintaining other attributes, thereby providing a more reliable solution for bias mitigation. + +Interpretability of Diffusion Models. Most of the work interpreting diffusion models focus on explaining how the input affects the generation via attribution methods and attention maps [8, 19, 21, 57], but they do not reveal the underlying mechanisms by which these attributes are processed and synthesized into a coherent output. Recent research has revealed that diffusion models possess a semantically rich latent space [32], inspiring various studies on customized image editing through latent space manipulation [18, 33, 45]. Although these methods enable more flexible editing, they still do not provide a deep understanding of the model's inherent features, leading to unexpected results. Our approach directly interacts with the internal components of diffusion models, enabling us to identify and precisely modify key features like bias-specific ones. This allows for more accurate control and targeted editing within the model itself, also providing efficient tools for interpreting diffusion models. + +# 3. Preliminary + +In this section, we provide an overview of diffusion model architecture. Diffusion models operate through two key processes, the forward and backward diffusion steps [23]. Taking image generation as an example. In the forward process, noise is incrementally added to the input, transforming a real image $\mathbf{x}_0$ into pure noise over a series of timesteps $t\in [0,T)$ where $T$ is the total number of timesteps. Mathematically, the forward process can be represented as + +$$ +q \left(\mathbf {x} _ {t} \mid \mathbf {x} _ {0}\right) = \mathcal {N} \left(\mathbf {x} _ {t}; \sqrt {\bar {\alpha} _ {t}} \mathbf {x} _ {0}, (1 - \bar {\alpha} _ {t}) I\right). \tag {1} +$$ + +Here, $\bar{\alpha}_t\coloneqq \prod_{i = 1}^t\alpha_i$ controls the amount of noise injected at each timestep $t$ , with $\mathbf{x}_T$ becoming a near-random noise vector after the final timestep. In the backward process, the goal is to reverse this noising process and generate a clean + +![](images/f7139f10dd518f79fc8fa4a8c35d3448900402f917d4d54ba02be2610450e446.jpg) +Figure 1. Framework of DIFFLENS exploring the inner working of diffusion models for bias mitigation. Neurons that are activated in the semantic space are defined as fired features and inactivated ones as unfired features. Target features are either suppressed or amplified to control bias level. We divide the framework into three parts (a), (b) and (c) which correspond to Secs. 4.1 to 4.3, respectively. + +image from the noisy sample. At each step, the diffusion network, e.g., U-Net [50], models $p_{\theta}(\mathbf{x}_{t - 1}|\mathbf{x}_t)$ with parameter $\theta$ predicting the denoised version given the current sample $\mathbf{x}_t$ . This reverse process is modeled as + +$$ +p _ {\theta} (\mathbf {x} _ {t - 1} | \mathbf {x} _ {t}) = \mathcal {N} (\mathbf {x} _ {t - 1}; \mu_ {\theta} (\mathbf {x} _ {t}, t), \Sigma_ {\theta} (\mathbf {x} _ {t}, t)), \tag {2} +$$ + +where $\mu_{\theta}$ and $\Sigma_{\theta}$ represent the mean and variance of the predicted denoised sample. + +Our work specifically focuses on the inner working of model itself, such as U-Net [50] architecture in the diffusion model, to investigate the mechanisms underlying bias generation. We aim to examine the model's decision-making process through these representations to better understand the biases that emerge during image generation, providing deeper insights into the model's internal functions and the precise factors driving its behavior [4]. + +# 4. Methodology + +In this section, we propose DIFFLENS, consisting of three stages aimed at uncovering internal mechanisms of bias generation in diffusion models and leveraging this understanding for bias mitigation. Specifically, our goal is to discover the neuron activities that controls the generation of biased contents. To achieve this, DIFFLENS first disentangles the hidden states of neuron activation from the model layer into a sparse semantic space, capturing meaningful features that represent various attributes. Next, it identifies the semantic features specifically related to biased content generation using a gradient-based attribution method. Finally, by adjusting the scale of these identified features, DIFFLENS allows for control over the level of bias in generated content, providing an effective approach for bias mitigation. + +The overall framework has been illustrated in Fig. 1 and Algorithms 1 and 2 in Appendix A. + +# 4.1. Disentangling Polysemantic Neurons + +We study the hidden states $\mathbf{h} = [h_1, h_2, \dots, h_n] \in \mathbb{R}^n$ from the diffusion model, representing the activations of the $n$ neurons at each network layer. We assume that these neuron activations inherently capture the semantic information of the generation process. Similar observations have been made in language models [12, 38], though little attention has been given to diffusion models in this regard. Kwon et al. [32] demonstrated that diffusion models exhibit a semantic space within the U-Net [50] architecture, specifically at the bottleneck layer. However, previous works have not precisely analyzed the different attributes within this space. In contrast, our method explicitly identifies semantic features and pinpoints the specific dimensions of neuron activations responsible for bias generation. This enables more precise and targeted control over bias, as demonstrated in our experiments in Figs. 3 and 4. + +Note that, neurons are often polysemantic, meaning that each neuron often corresponds to multiple, unrelated concepts, as discussed in [14, 39, 41]. This places obstacles for understanding and interpreting the mechanism of the model, as the network's learned features are intertwined in a complex manner [4]. Nevertheless, it also presents an opportunity that these features can potentially be disentangled by identifying directions in the activation space, where each activation vector can be expressed as a sparse linear combination of underlying features [24]. Thus, we leverage k-Sparse Autoencoders (k-SAE) [35] to analyze and interpret the semantic features encoded within the hidden states $\mathbf{h}$ . It helps the method focus on salient features, disentangling polysemantic activations and isolating semantic attributes + +within the latent space, making it effective for understanding and manipulating key features in diffusion models. Part (a) of Fig. 1 has illustrated this processing stage. + +Specifically, let $\mathbf{H}^n = \{\mathbf{h}\}$ be the original latent feature space with dimension number $n$ . We transform this feature space to a sparse high-dimensional space, denoted as $\mathbf{S}^m = \{\mathbf{s}\}$ with dimension $m$ that is much larger than $n$ . In the space $\mathbf{S}^m$ , the features $\mathbf{s}$ are monosemantic, i.e., each feature represents a distinct, non-overlapping semantic concept (discussions are in Appendix E.2). The mapping $\Phi : \mathbf{H}^n \to \mathbf{S}^m$ can be expressed by + +$$ +\mathbf {s} = \left[ s _ {1}, s _ {2}, \dots , s _ {m} \right] = \Phi (\mathbf {h}) = \operatorname {T o p K} \left(\mathbf {W} _ {\text {e n c}} \left(\mathbf {h} - \mathbf {b} _ {\text {p r e}}\right)\right). \tag {3} +$$ + +We map the vector $\mathbf{h}$ to a sparse vector $\mathbf{s}$ with $\| \mathbf{s} \|_0 = k$ and $k < m$ , indicating that $\mathbf{s}$ is sparse with only $k$ features activated. This is achieved by the encoding process of a k-SAE where $\mathbf{W}_{\mathrm{enc}} \in \mathbb{R}^{m \times n}$ and $\mathbf{b}_{\mathrm{pre}} \in \mathbb{R}^n$ is the bias term. $\operatorname{TopK}(\cdot)$ is an operation that selects the top $K = k$ largest features, referred to as "fired" features, while setting the remaining features to zero, which are "unfired" features as shown in Fig. 1. For the decoding process, we have + +$$ +\hat {\mathbf {h}} = \Phi^ {- 1} (\mathbf {s}) = \mathbf {W} _ {\mathrm {d e c}} \mathbf {s} + \mathbf {b} _ {\mathrm {p r e}}, \tag {4} +$$ + +where $\hat{\mathbf{h}}\in \mathbb{R}^n$ is the reconstructed latent vector and $\mathbf{W}_{\mathrm{dec}}\in \mathbb{R}^{n\times m}$ . $\Phi^{-1}$ represents the inverse mapping of $\Phi$ since $\mathbf{W}_{\mathrm{dec}}$ and $\mathbf{W}_{\mathrm{enc}}$ are initialized as transposes of each other. We train the k-SAE using the loss function defined in Eq. (5), with the objective of minimizing the reconstruction error within the original hidden space, expressed as + +$$ +\mathcal {L} (\mathbf {h}) = \left\| \mathbf {h} - \hat {\mathbf {h}} \right\| _ {2} ^ {2}, \tag {5} +$$ + +where $\mathbf{h}$ and $\hat{\mathbf{h}}$ refers to the original and the reconstructed hidden vectors, respectively. We present results and implementation details in Appendix B.2, demonstrating that we reconstruct the original latent space with few influences on the target model. The encoding and decoding processes are presented in Algorithm 2 at Appendix A. + +Disentangling the original latent space of diffusion model shares several advantages. On one hand, transforming it into a sparse semantic space can well handle the polysemistic issue of the internal neurons within the neural network [24]. On the other hand, it enables more precise locating on the target feature for generation content attributes, such as identifying the bias features in the subsequent sections in our paper. In addition, editing the whole state in the original hidden space like that in [18, 32, 33, 43] could inadvertently alter irrelevant attributes, caused by the polysemantic properties of the neuron in the model as shown by Figs. 2 and 3 in experiments. + +# 4.2. Identifying Bias Features + +The next step is to identify the most influential features within the sparse semantic space that are associated with + +particular function, i.e., bias generation mechanism in this paper. This need arises because accurately identifying the features that correspond to specific attributes can significantly enhance our ability to control and manipulate generated content responsibly. + +Inspired by [13, 56], we propose a gradient-based bias attribution method, to locate the features that correlates to the bias concepts. We visualize this process in part (b) of Fig. 1. Intuitively, the neuron activations that influence the outcome of gender diverse in different aspects, which controls different characteristics such as hair style in Fig. 6. + +Our goal is to measure the influence of the semantic feature $s_i \in \mathbf{s}$ within $\mathbf{S}^m$ driving the biased contents on the target outcome. We define $F_{\mathbf{x}}: \mathbf{S}^m \to \mathbb{R}$ as a measure of interest, assessing the outcome of the model w.r.t. the target $\mathbf{s}$ given the input $\mathbf{x}$ . Specifically, we quantify the bias level as the probability of the social attribute class $y$ as $F_{\mathbf{x}}(\mathbf{s}) = \operatorname*{Pr}(y|\mathbf{s})$ regarding the disentangled feature vectors derived from Eq. (3). Here $y$ represents different classes such as "male" or "female" in gender bias. To locate the influential features, the internal mechanisms of diffusion model, we propose the attribution score of bias generation to the disentangled feature $s_i$ as + +$$ +S (s _ {i}; \mathbf {x}) = (s _ {i} - s _ {i} ^ {\prime}) \cdot \int_ {\alpha = 0} ^ {1} \frac {\partial F _ {\mathbf {x}} (\mathbf {s} ^ {\prime} + \alpha (\mathbf {s} - \mathbf {s} ^ {\prime}))}{\partial s _ {i}} d \alpha , \tag {6} +$$ + +where $\mathbf{s}' = [s_1',\dots,s_m']$ is a relative baseline of $\mathbf{s}$ and $\partial F_{\mathbf{x}}(\mathbf{s}' + \alpha (\mathbf{s} - \mathbf{s}')) / \partial s_i$ is the gradient of the bias measure given the input image $\mathbf{x}$ , to the target feature space $\mathbf{s}$ . It is achieved by a light-weight classifier $F_{\mathbf{x}}(\cdot ;\psi)$ with parameter $\psi$ . The classifier construction follows [43] and we provide more training details in Appendix B.3. Choices of baselines $s'$ are in Appendix A.1. + +In Eq. (6), the attribution score $S(s_{i};\mathbf{x})$ aggregates the gradients along the path of the individual semantic feature $s_i$ ranging from zero effect $(\alpha = 0)$ to full effect $(\alpha = 1)$ , given the input $\mathbf{x}$ , which is the generated content from the diffusion model. If a feature $s_i$ significantly impacts the representation of a particular attribute, its corresponding gradient will be prominent, resulting in higher integrated values, i.e., higher attribution score $S$ . + +Further, to assess the overall influence of the semantic feature $s_i$ to the bias contents, we aggregate the attribution score as $S(s_i; X) = \sum_{j=1}^{N} S(s_i; \mathbf{x}_j)$ where $\mathbf{x}_j \in X$ with $X$ as the support dataset of $N$ generated samples. DIFFLENS then selects the top $\tau$ features with the largest attribution scores as bias features, whose index set is $\mathbf{A} = \{i_1, \ldots, i_\tau\}$ , i.e., these are the most influential semantic features within the diffusion model driving the bias contents during generation. The complete process of bias attribution is outlined in Algorithm 1 at Appendix A. Notably, this step is required only once for one specific diffusion model. + +Additionally, we observe that the target can extend be + +yond s to include the input $\mathbf{x}$ or any component within the target model itself. This broader approach enables us to examine how these elements impact the behavior of the underlying model concerning the measure $F_{\mathbf{x}}$ . To the best of our knowledge, we are the first to incorporate feature disentanglement and attribution methods to interpret the bias-related components within diffusion models, providing deeper insights into the internal workings of diffusion process. + +# 4.3. Intervening in Bias Features for Debiasing + +To control and mitigate social bias in diffusion models, we propose an effective method via intervening in the above discovered bias features indexed with $i \in \mathbf{A}$ , where $\mathbf{A}$ consists of indexes of the identified bias features. It is possible to suppress or amplify these features to exert control over the model's behavior [12]. Unlike other methods [32, 33, 43], we only intervene in the identified features indexed with $\mathbf{A}$ rather than the entire semantic space. Our intervention approach are illustrated in part (c) of Fig. 1. + +Specifically, we apply an intervention operation to the target features in the sparse semantic space. Then these features are reconstructed into the original latent space for subsequent image generation and more details are in Appendix A. For each feature indexed within $\mathbf{A}$ , we can either apply a scalar multiplication or adding an additive bias as + +$$ +s _ {i} = \operatorname {I n t e r v e n e} \left(s _ {i}\right) = \left\{ \begin{array}{l l} \beta s _ {i} & (\text {S c a l i n g}), \text {o r} \\ s _ {i} + \beta & (\text {A d d i n g}). \forall i \in \mathbf {A}, \beta \in \mathbb {R}. \end{array} \right. \tag {7} +$$ + +By applying a distinct parameter $\beta$ , we can suppress or amplify the target bias feature to adjust the bias level or even mitigate it for improved balance. + +Our DIFFLENS demonstrates that it can control over the generation and representation of biased features in the output images as shown in Sec. 5.4. Furthermore, in Sec. 5.6, we gain more valuable insights into how different bias features are embedded within the inner working of diffusion model, allowing us not only to debias effectively but also to better understand the model's decision-making processes. + +# 5. Experiments + +In this section, we evaluate our DIFFLENS qualitatively and quantitatively, following the below research questions. + +RQ1: How effectively does DIFFLENS mitigate social bias while maintaining image quality? RQ2: Can DIFFLENS accurately identify intrinsic mechanism of bias generation? + +RQ3: How well does it control the bias level through intervening in the found bias features? RQ4: Is it able to interpret bias feature in a fine-grained way? + +# 5.1. Experimental Settings + +Diffusion Model Architecture. We evaluate our proposed method on two pretrained diffusion models, (i) P2 [10], + +an unconditional diffusion model trained on CelebA-HQ dataset [26]; (ii) Stable Diffusion [49], a text-conditional diffusion model trained on LAION dataset [52]. + +Baselines. We compare our method against several baselines that are categorized into interpretability and guidance-based methods. For interpretability methods, we construct Activation [6] trying to locate influential neurons according to the neuron activation values. For guidance-based methods, we include Latent Editing [32] and H-Distribution [43] (abbreviated as "L.E." and "H-Dist."), trying to learn a latent vector to guide the unbiased generation within the bottleneck layer of U-Net. We additionally compare two methods, specifically for text-conditional diffusion models, which are Latent Direction [33] and Finetuning [54]. Due to Activation's limited debiasing effectiveness across both diffusion models and its restricted control over bias levels, we report its results only in Tab. 1 and Fig. 2. More details about baselines are in Appendix C.3. + +Implementation Details. To ensure fair comparison, and without loss of generality, we constrain our method, DIFFLENS, and other benchmark methods to operate on the bottleneck layer of the U-Net [50] in the diffusion model, adhering to the baseline settings in [32, 43]. Notably, DIFFLENS can easily extend the targeted mechanism to broader regions within the model. Following Sec. 4, we train kSAEs on the target latent space in the diffusion model. For the training of the light-weight classifier, we follow the settings in [43]. For the support set of samples used for identifying bias feature, we leverage the diffusion models to generate 1,000 sample for each attribute. We provide more details of implementation and dataset we use in Appendix B. The reproducible codes are in supplementary material. + +Evaluation Metrics. To measure the bias level of generated content, we use the Fairness Discrepancy (FD) introduced in [43] which calculates the Euclidean distance between a reference distribution (typically uniform) and the bias distribution of generation, representing class distributions within an attribute (e.g., male and female for gender attribute). A lower FD score indicates more balanced outputs. We also measure image quality using Fréchet Inception Distance (FID) [22] on the generated images. In addition, we follow [54] to measure (i) CLIP-T, which calculates the CLIP [48] based semantic similarity between the generated image and the input text prompt, and (ii) CLIP-I, which assesses the similarity between originally generated images and images after debiasing. More details are in Appendix C. + +# 5.2. Debiasing Diffusion Models + +We define "bias mitigation" task as adjusting the diffusion model to produce balanced outputs across specific social attributes (e.g., gender). We focus on the task across gender, age, and race attributes. Gender includes male and female + +![](images/ef162e8a748da7c069f87047ccb63ada1e4d13cb0c633ef6806d0ac3c8ad980b.jpg) +Original M:F(%) 33.9:66.1 +Figure 2. Comparison of randomly sampled original and debiased images from various baseline methods. The minority group (male) is highlighted with red bounding boxes for easier viewing. "M:F" refers to the male-to-female ratio. Our DIFFLENS effectively mitigates bias while preserving high generation quality, whereas other methods either struggle to maintain balance or produce images with artifacts. + +![](images/8841639032a35a07dc5242762b1ace1bef02c29205d095bc9d7f72e9c11119b4.jpg) +Activation M:F(%) 36.6:63.4 + +![](images/f7812aa5cffae3918b5bebd952688e68798759e0dcacbcff820fb0006f5e8a34.jpg) +Latent Editing +M:F(%) 50.2:49.8 + +![](images/2aee04b24988774fc2260a7f3108ff5f4d5ea4f261257f12495e76b7965a1e9c.jpg) +H-Distribution M:F(%) 46.6:53.4 + +![](images/c549bbed29f7d5628145f7046c788246fd39d0cb60077e44eec72a9f3199aed6.jpg) +DiffLens (Ours) M:F(%) 49.9:50.1 + +
CategoryMethodGender (2)Age (3)Race (4)
FD ↓FID ↓CLIP-I ↑FD ↓FID ↓CLIP-I ↑FD ↓FID ↓CLIP-I ↑
Original0.22633.38-0.59233.38-0.71833.38-
Guidance-basedLatent Editing [32]0.00329.850.94740.60633.710.90920.31734.610.9081
H-Distribution [43]0.04831.310.94400.51134.210.85940.49436.190.8762
InterpretabilityActivation [6]0.19034.270.80600.54448.910.77930.70046.130.7846
DIFFLENS (Ours)0.00231.930.94790.40131.710.94140.44733.470.9111
+ +Table 1. Evaluation of bias mitigation in unconditional diffusion model P2 [10]. The numbers beside each attribute indicate the number of classes for that attribute. We highlight the best results in bold and the second best with underlined format (excluding the "original"). + +classes; age includes young, adult, and old; and race includes white, black, Asian, and Indian classes. + +Intuitively, bias mitigation should yield balanced generation without compromising the overall semantics of the generated content. To comprehensively evaluate the effectiveness of bias mitigation (RQ1) across different methods, we assess (i) the bias level using the FD score of generated images, (ii) the generation quality via the FID score, and (iii) the preservation of overall image semantics after bias mitigation using CLIP-I/T scores. Beyond evaluating on the unconditional generative model, we also extend our analysis to a text-conditional one to showcase the broader applicability and generalizability of our method. + +Qualitative Results on Bias Mitigation. We present batches of generated examples before and after debiasing gender attribute to visually evaluate bias level and generation quality, as shown in Fig. 2. Our DIFFLENS effectively neutralizes gender bias while ensuring the generated images remain realistic and intact. In comparison, Latent Editing may generate unrealistic images, for example, the second one in the first row in Fig. 2, H-Distribution struggles to generate a well-balanced output, and Activation produces imbalanced, distorted images. We also provide visual results for text-to-image diffusion model Stable Diffusion [49] in Appendix D.2. + +Quantitative Analysis on Bias Mitigation. We assess baseline methods on debiasing across various social attributes (e.g., gender) by generating balanced outputs. As shown in Tab. 1, our DIFFLENS demonstrates competitively + +strong results in (i) bias mitigation (measured by FD), (ii) in preserving semantic feature coherence (measured by CLIP-I) and (iii) maintaining generation quality (measured by FID). Activation shows poor ability to mitigate social bias. Although Latent Editing performs well in mitigating gender bias and achieves favorable FID, it can produce images with artifacts, as shown in Figs. 2 and 3. H-Distribution struggles to generate well-balanced outputs, as reflected by its poor FD scores. + +In Tab. 2, we show results on a text-to-image model, Stable Diffusion [49]. Our DIFFLENS effectively mitigates bias, especially for gender and age, while effectively maintaining generation quality and image semantics among methods that are effective in bias mitigation. Latent Direction achieves good debiasing for age and race but often overcorrects, reducing CLIP-I scores and compromising the balance between target generation ratio and image quality. Finetuning, demanding extensive effort, excels in gender and race but struggles with age bias mitigation. Other baselines either handle multi-class attributes poorly or lack generalizability to text-to-image diffusion models. + +# 5.3. Accurate Identification of Bias Mechanism + +Accurate bias mechanism discovery could provide feasible control on the generated contents of diffusion, regarding the level of bias attributes. For instance, unlike the original diffusion model P2 [10], which generates a male-to-female gender ratio of approximately 3:7, our method easily attains the reverse ratio of 7:3. This capability goes beyond merely + +
MethodGender (2)Age (3)Race (4)
FD ↓FID ↓CLIP-I ↑CLIP-T ↑FD ↓FID ↓CLIP-I ↑CLIP-T ↑FD ↓FID ↓CLIP-I ↑CLIP-T ↑
Original0.564120.06-0.61550.752120.06-0.61550.558120.06-0.6155
Latent Editing [32]0.408166.110.82530.60050.682200.900.85270.61220.524153.050.88040.6086
H-Distribution [43]0.222151.680.84750.60870.506147.710.83450.60980.544126.900.82550.6100
Latent Direction [33]0.305129.370.80580.60910.052113.810.81510.60670.175128.300.82110.6132
Fintuning [54]0.050161.470.87790.60950.746161.470.87790.60950.198161.470.87790.6095
DIFFLENS (Ours)0.046112.830.85010.60900.04999.170.87780.60570.401119.860.90960.6149
+ +Table 2. Evaluation of bias mitigation in text-to-image diffusion model Stable Diffusion [49], based on average performance across four occupations (see Appendix B). We highlight the best results in bold and the second-best with underlined text (excluding the "original"). + +
OriginalL.E.H-Dist.DiffLensOriginalL.E.H-Dist.DiffLensOriginalL.E.H-Dist.DiffLens
+ +balancing generated images in Sec. 5.2; it allows us to examine how effectively each method edits the target attribute, e.g., gender, while preserving non-target semantic features, e.g., eyeglass and smile (RQ2). For illustration, we visualize the generated contents of compared methods upon the reversed ratio of gender attribute in Fig. 3. More examples and details are provided in Appendix D.4.1. As is shown in Fig. 3, our method can preserve non-target attributes such as smile, background coherence, and even eyeglasses when debiasing, showcasing accurate identification of the bias generation mechanism by our method. While the baseline methods often introduce distortions or compromise detail in these areas, without looking into inner working of bias generation. Additionally, from Tabs. 1 and 2, our DIFFLENS consistently achieves high CLIP-I scores, demonstrating its ability to accurately identify and modify target bias specific features with minimal impact on unrelated content. + +# 5.4. Fine-grained Bias Level Control and Editing + +Benefiting from bias mechanism identification, we can control bias level with finer granularity and across a broader range (RQ3). Figure 5 illustrates how different methods progressively modulate image generation in diffusion along two opposite gender directions. More examples are in Appendix D.4.2. Our DIFFLENS produces natural transitions, consistently preserving semantic features such as facial expression and maintaining high generation quality. Examples about control in multi-attributes are in Appendix D.4.3. The results of varying gender ratios generated by each method are shown in Fig. 4. Latent Editing and H-Distribution perform well when the gender ratio is around balanced, but their fidelity and visual coherence deteriorate at more skewed ratios, indicating a limitation in handling highly imbalanced outputs. By adjusting the value of intervention parameter incorporated in Eq. (7), our DIFFLENS can flexi + +bly control bias feature and generate different gender ratios, ranging from balanced to imbalanced ones (almost male or female), illustrating powerful controllability over the representations of bias. Such highly imbalanced ratios also have wide applications in real world such as data augmentation for under-represented subgroups [43]. + +![](images/07fe781ce4742c76f11b0d5ebf3a3260fdff06c5d88d23725b097dd14937480f.jpg) +Figure 3. Comparison in accurate identification of bias features. Our DIFFLENS preserves overall image semantics such as smile and eyeglasses while other methods frequently introduce distortions or lose important details. +Figure 4. Achievable range of bias level control for each method. The Log Gender Ratio reflects the log of male to female ratio in the generated images, with 0 indicating balance. Our DIFFLENS offers broader bias control, delivering higher generation quality (left plot) and improved semantic coherence (right plot). + +# 5.5. Ablation Study + +We conduct an ablation study to reveal how each component in our DIFFLENS takes effect. Recall that, it dissects the diffusion models by (i) disentangling the latent space to a sparse semantic space via k-SAE and (ii) identifying the bias features using bias attribution. In Tab. 3, directly selecting neurons with the highest activation values as bias features (Activation method) performs poorly, supporting the claim in Sec. 4.1. Comparing "Act@L" with "Act@S" and "Attr@L" with DIFFLENS, using k-SAE improves visual coherence, as seen in the higher CLIP-I scores (0.8597 to 0.9479 and 0.8060 to 0.9368), while avoiding direct edits on polysemantic neurons mitigates FID degradation (41.68 to 31.93). Furthermore, comparing "Act@L" with "Attr@L" and "Act@" with DIFFLENS, bias attribution enables more precise localization of bias features, significantly lowering + +![](images/1549d9f196ea442434fa0c59b828050caaf9c0e983f0c1b3783a33831c269a97.jpg) +Figure 5. Comparison of individual image transformations along the gender axis, with columns displaying images sampled from progressively shifting ratios as in Fig. 4. Our DIFFLENS achieves smooth and consistent transitions that preserve semantic features like facial expressions, while other methods show distortions and loss of detail at higher imbalance ratios. + +![](images/10016405e196c19f863a164b7f019b0a16fa3eb6995a03ac73672703899351ef.jpg) + +FD (0.190 to 0.072 and 0.303 to 0.002). DIFFLENS incorporating both shows superiority in bias mitigation with concise localization and intervention of bias features. + +
ComponentsOrig.Act@LAct@SAttr@LDIFFLENS
ActivationX✓✓XX
K-SAEXX✓X✓
Bias Attr.XXX✓✓
FD ↓0.2260.1900.3030.0720.002
FID ↓33.3834.2732.7041.6831.93
CLIP-I ↑-0.80600.93860.85970.9479
+ +- ①, ②: Bias attribution allows precise bias mitigation. +- ③, ④: K-SAE improves visual feature coherence and maintains generation quality, via disentangling original latent space. + +Table 3. Ablation study of DIFFLENS in P2 [10] for gender attribute. We abbreviate Original as "Orig.", neuron activations as "Act.", and bias attributions as "Attr". "@L" denotes operations on the original latent space, and "@S" on sparse semantic space. + +# 5.6. Wider Investigation: Bias Mechanism + +We further look into the bias mechanism, i.e., the identified bias features $\{s_i|i\in \mathbf{A}\}$ , to reveal its granular aspects on social attribute generation and how these features distinguishably affect the biased contents, shedding some lights on the inner workings of diffusion models (RQ4). We observe from Fig. 6 that (i) multiple features may jointly influence a single social bias attribute, such as "short hair" and "mustache" contributing to gender perception; (ii) some features overlap across biases, for instance, "side pose" being associated with both age and gender biases as it often appears in images of older males; and (iii) poses like "side pose" and "head down" reflect in photography, often aligning with gender norms that men are depicted in dynamic or indirect poses, while women appear younger and more polished, reflecting social expectations [15, 58, 60]. These stereotypes mirror underlying social biases, impacting not only how individuals are portrayed but also how their images are interpreted. Our DIFFLENS serves as a lens that uncovers social biases and stereotypes embedded in generated images, providing insight into how attributes such as gender, race, and age are systematically portrayed. + +# 6. Conclusion + +In this work, we introduced a novel approach DIFFLENS, for mitigating social biases in diffusion models by dissecting, analyzing and intervening in the internal mechanisms of the diffusion model. Leveraging k-SAE and gradient-based attribution method, we disentangle, identify, and control bias-related features with precision, allowing targeted bias mitigation while preserving non-target attributes. A wide range of evaluations across social attributes such as gender, age, and race validate effectiveness of DIFFLENS in bias control and content preservation. Our work offers an interpretable solution to bias mitigation in diffusion models, though challenges remain, including the precise identification of certain racial features and a quality decline in intensified edits. Future directions include expanding this approach to other bias types, incorporating additional fairness perspectives, and exploring application across diverse model architectures like Diffusion Transformers [46] for broader impact in ethical generative AI deployment. + +![](images/7654e5edfa8f6ef7606eec14cae9d7684146b2dd906c714e3c874a3b7ab0caf1.jpg) +Figure 6. Manually analyzing the discovered bias features in diffusion models. Each column represents one specific feature $s_i$ , $i \in \mathbf{A}$ and each row originates from the same raw image. + +# 7. Acknowledgment + +The research received partial support from National Natural Science Foundation of China (Grant No. 62406193). The authors also gratefully acknowledge further assistance provided by Shanghai Frontiers Science Center of Human-centered Artificial Intelligence, MoE Key Lab of Intelligent Perception and Human-Machine Collaboration, and HPC Platform of ShanghaiTech University. + +# References + +[1] Hritik Bansal, Da Yin, Masoud Monajatipoor, and Kai-Wei Chang. How well can text-to-image generative models understand ethical natural language interventions? In Proceedings of the 2022 Conference on Empirical Methods in Natural Language Processing, pages 1358–1370, 2022. 2 +[2] Abhipsa Basu, R Venkatesh Babu, and Danish Pruthi. Inspecting the geographical representativeness of images from text-to-image models. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 5136-5147, 2023. 2 +[3] Sebastian Benthall and Bruce D. Haynes. Racial categories in machine learning. In Proceedings of the Conference on Fairness, Accountability, and Transparency, page 289-298, New York, NY, USA, 2019. Association for Computing Machinery. 11 +[4] Leonard Bereska and Stratis Gavves. Mechanistic interpretability for AI safety - a review. Transactions on Machine Learning Research, 2024. Survey Certification, Expert Certification. 1, 3, 11 +[5] Federico Bianchi, Pratyusha Kalluri, Esin Durmus, Faisal Ladhak, Myra Cheng, Debora Nozza, Tatsunori Hashimoto, Dan Jurafsky, James Zou, and Aylin Caliskan. Easily accessible text-to-image generation amplifies demographic stereotypes at large scale. In Proceedings of the 2023 ACM Conference on Fairness, Accountability, and Transparency, pages 1493-1504, 2023. 2 +[6] Steven Bills, Nick Cammarata, Dan Mossing, Henk Tillman, Leo Gao, Gabriel Goh, Ilya Sutskever, Jan Leike, Jeff Wu, and William Saunders. Language models can explain neurons in language models. https://openaipublic.blob.core.windows.net/neuron-explainer/paper/index.html, 2023.5, 6 +[7] Hila Chefer, Oran Lang, Mor Geva, Volodymyr Polosukhin, Assaf Shocher, Inbar Mosseri, Lior Wolf, et al. The hidden language of diffusion models. In *The Twelfth International Conference on Learning Representations*. 1 +[8] Hila Chefer, Yuval Alaluf, Yael Vinker, Lior Wolf, and Daniel Cohen-Or. Attend-and-excite: Attention-based semantic guidance for text-to-image diffusion models. ACM Transactions on Graphics (TOG), 42(4):1-10, 2023. 2 +[9] Jaemin Cho, Abhay Zala, and Mohit Bansal. Dall-eval: Probing the reasoning skills and social biases of text-to-image generation models. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 3043-3054, 2023. 2 + +[10] Jooyoung Choi, Jungbeom Lee, Chaehun Shin, Sungwon Kim, Hyunwoo Kim, and Sungroh Yoon. Perception prioritized training of diffusion models. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 11472-11481, 2022. 5, 6, 8, 1, 2, 3, 4, 9 +[11] Ching-Yao Chuang, Varun Jampani, Yuanzhen Li, Antonio Torralba, and Stefanie Jegelka. Debiasing vision-language models via biased prompts. arXiv preprint arXiv:2302.00070, 2023. 2 +[12] Damai Dai, Li Dong, Yaru Hao, Zhifang Sui, Baobao Chang, and Furu Wei. Knowledge neurons in pretrained transformers. In Proceedings of the 60th Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers), pages 8493-8502, Dublin, Ireland, 2022. Association for Computational Linguistics. 3, 5 +[13] Kedar Dhamhere, Mukund Sundararajan, and Qiqi Yan. How important is a neuron. In International Conference on Learning Representations, 2019. 4 +[14] Nelson Elhage, Tristan Hume, Catherine Olsson, Nicholas Schiefer, Tom Henighan, Shauna Kravec, Zac Hatfield-Dodds, Robert Lasenby, Dawn Drain, Carol Chen, et al. Toy models of superposition. arXiv preprint arXiv:2209.10652, 2022.3 +[15] Emilio Ferrara. Fairness and bias in artificial intelligence: A brief survey of sources, impacts, and mitigation strategies. Sci, 6(1), 2024. 8 +[16] Felix Friedrich, Manuel Brack, Lukas Struppek, Dominik Hintersdorf, Patrick Schramowski, Sasha Luccioni, and Kristian Kersting. Fair diffusion: Instructing text-to-image generation models on fairness. arXiv preprint arXiv:2302.10893, 2023. 2 +[17] Rohit Gandikota, Hadas Orgad, Yonatan Belinkov, Joanna Materzyńska, and David Bau. Unified concept editing in diffusion models. In Proceedings of the IEEE/CVF Winter Conference on Applications of Computer Vision, pages 5111-5120, 2024. 1, 2 +[18] René Haas, Inbar Huberman-Spiegelglas, Rotem Mulayoff, Stella Graßhof, Sami S Brandt, and Tomer Michaeli. Discovering interpretable directions in the semantic latent space of diffusion models. In 2024 IEEE 18th International Conference on Automatic Face and Gesture Recognition (FG), pages 1-9. IEEE, 2024. 2, 4 +[19] Ligong Han, Yinxiao Li, Han Zhang, Peyman Milanfar, Dimitris Metaxas, and Feng Yang. Svdiff: Compact parameter space for diffusion fine-tuning. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 7323-7334, 2023. 2 +[20] Ruifei He, Chuhui Xue, Haoru Tan, Wenqing Zhang, Yingchen Yu, Song Bai, and Xiaojuan Qi. Debiasing text-to-image diffusion models. In Proceedings of the 1st ACM Multimedia Workshop on Multi-Modal Misinformation Governance in the Era of Foundation Models, page 29–36, New York, NY, USA, 2024. Association for Computing Machinery. 2 +[21] Amir Hertz, Ron Mokady, Jay Tenenbaum, Kfir Aberman, Yael Pritch, and Daniel Cohen-or. Prompt-to-prompt image editing with cross-attention control. In The Eleventh International Conference on Learning Representations, 2023. 2 + +[22] Martin Heusel, Hubert Ramsauer, Thomas Unterthiner, Bernhard Nessler, and Sepp Hochreiter. Gans trained by a two time-scale update rule converge to a local nash equilibrium. Advances in neural information processing systems, 30, 2017. 5, 11 +[23] Jonathan Ho, Ajay Jain, and Pieter Abbeel. Denoising diffusion probabilistic models. In Advances in Neural Information Processing Systems, pages 6840-6851. Curran Associates, Inc., 2020. 1, 2 +[24] Robert Huben, Hoagy Cunningham, Logan Riggs Smith, Aidan Ewart, and Lee Sharkey. Sparse autoencoders find highly interpretable features in language models. In The Twelfth International Conference on Learning Representations, 2024. 3, 4, 1 +[25] Kimmo Karkkainen and Jungseock Joo. Fairface: Face attribute dataset for balanced race, gender, and age for bias measurement and mitigation. In Proceedings of the IEEE/CVF winter conference on applications of computer vision, pages 1548-1558, 2021. 2, 4 +[26] Tero Karras, Timo Aila, Samuli Laine, and Jaakko Lehtinen. Progressive growing of GANs for improved quality, stability, and variation. In International Conference on Learning Representations, 2018. 5 +[27] Tero Karras, Samuli Laine, and Timo Aila. A style-based generator architecture for generative adversarial networks. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 4401-4410, 2019. 5 +[28] Eunji Kim, Siwon Kim, Chaehun Shin, and Sungroh Yoon. De-stereotyping text-to-image models through prompt tuning. In ICML Workshop on Challenges in Deployable Generative AI, 2023. 2 +[29] Gwanghyun Kim, Taesung Kwon, and Jong Chul Ye. Diffus ionclip: Text-guided diffusion models for robust image man i plication. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 2426-2435, 2022. 1 +[30] Yeongmin Kim, Byeonghu Na, Minsang Park, JoonHo Jang, Dongjun Kim, Wanmo Kang, and Il chul Moon. Training unbiased diffusion models from biased dataset. In The Twelfth International Conference on Learning Representations, 2024. 1 +[31] Xianghao Kong, Ollie Liu, Han Li, Dani Yogatama, and Greg Ver Steeg. Interpretable diffusion via information decomposition. In The Twelfth International Conference on Learning Representations. 1 +[32] Mingi Kwon, Jaeseok Jeong, and Youngjung Uh. Diffusion models already have a semantic latent space. In The Eleventh International Conference on Learning Representations, 2023. 2, 3, 4, 5, 6, 7, 9 +[33] Hang Li, Chengzhi Shen, Philip Torr, Volker Tresp, and Jindong Gu. Self-discovering interpretable diffusion latent directions for responsible text-to-image generation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 12006-12016, 2024. 2, 4, 5, 7 +[34] Sasha Luccioni, Christopher Akiki, Margaret Mitchell, and Yacine Jernite. Stable bias: Evaluating societal representa + +tions in diffusion models. Advances in Neural Information Processing Systems, 36, 2024. 1, 2, 11 +[35] Alireza Makhzani and Brendan Frey. K-sparse autoencoders. In International Conference on Learning Representations, 2014. 2, 3, 1, 6, 11 +[36] Vongani H Maluleke, Neerja Thakkar, Tim Brooks, Ethan Weber, Trevor Darrell, Alexei A Efros, Angjoo Kanazawa, and Devin Guillery. Studying bias in gans through the lens of race. In European Conference on Computer Vision, pages 344-360. Springer, 2022. 1 +[37] Pietro Melzi, Christian Rathgeb, Ruben Tolosana, Ruben Vera-Rodriguez, Dominik Lawatsch, Florian Domin, and Maxim Schaubert. Gandifface: Controllable generation of synthetic datasets for face recognition with realistic variations. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 3086-3095, 2023. 1 +[38] Kevin Meng, David Bau, Alex Andonian, and Yonatan Belinkov. Locating and editing factual associations in gpt. Advances in Neural Information Processing Systems, 35: 17359-17372, 2022. 1, 3 +[39] Jesse Mu and Jacob Andreas. Compositional explanations of neurons. In Advances in Neural Information Processing Systems, pages 17153-17163. Curran Associates, Inc., 2020. 3 +[40] Ranjita Naik and Besmira Nushi. Social biases through the text-to-image generation lens. In Proceedings of the 2023 AAAI/ACM Conference on AI, Ethics, and Society, pages 786-808, 2023. 2, 4 +[41] Chris Olah, Nick Cammarata, Ludwig Schubert, Gabriel Goh, Michael Petrov, and Shan Carter. Zoom in: An introduction to circuits. Distill, 5(3):e00024-001, 2020. 3 +[42] Hadas Orgad, Bahjat Kawar, and Yonatan Belinkov. Editing implicit assumptions in text-to-image diffusion models. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 7053-7061, 2023. 1, 2 +[43] Rishubh Parihar, Abhijnya Bhat, Abhipsa Basu, Saswat Mallick, Jogendra Nath Kundu, and R Venkatesh Babu. Balancing act: Distribution-guided debiasing in diffusion models. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 6668-6678, 2024. 1, 2, 4, 5, 6, 7, 9, 11 +[44] Ji-Hoon Park, Yeong-Joon Ju, and Seong-Whan Lee. Explaining generative diffusion models via visual analysis for interpretable decision-making process. Expert Systems with Applications, 248:123231, 2024. 1 +[45] Yong-Hyun Park, Mingi Kwon, Jaewoong Choi, Junghyo Jo, and Youngjung Uh. Understanding the latent space of diffusion models through the lens of riemannian geometry. In Thirty-seventh Conference on Neural Information Processing Systems, 2023. 2 +[46] William Peebles and Saining Xie. Scalable diffusion models with transformers. In Proceedings of the IEEE/CVF International Conference on Computer Vision (ICCV), pages 4195-4205, 2023. 8 +[47] Malsha V Perera and Vishal M Patel. Analyzing bias in diffusion-based face generation models. In 2023 IEEE International Joint Conference on Biometrics (IJCB), pages 1-10. IEEE, 2023. 1, 2 + +[48] Alec Radford, Jong Wook Kim, Chris Hallacy, Aditya Ramesh, Gabriel Goh, Sandhini Agarwal, Girish Sastry, Amanda Askell, Pamela Mishkin, Jack Clark, et al. Learning transferable visual models from natural language supervision. In International conference on machine learning, pages 8748-8763. PMLR, 2021. 5 +[49] Robin Rombach, Andreas Blattmann, Dominik Lorenz, Patrick Esser, and Björn Ommer. High-resolution image synthesis with latent diffusion models. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 10684-10695, 2022. 1, 2, 5, 6, 7, 3, 4 +[50] Olaf Ronneberger, Philipp Fischer, and Thomas Brox. U-net: Convolutional networks for biomedical image segmentation. In Medical image computing and computer-assisted intervention-MICCAI 2015: 18th international conference, Munich, Germany, October 5-9, 2015, proceedings, part III 18, pages 234-241. Springer, 2015. 3, 5, 1, 2, 4, 6 +[51] Harrison Rosenberg, Shimaa Ahmed, Guruprasad V Ramesh, Ramya Korlakai Vinayak, and Kassem Fawaz. Unbiased face synthesis with diffusion models: Are we there yet? arXiv preprint arXiv:2309.07277, 2023. 2 +[52] Christoph Schuhmann, Romain Beaumont, Richard Vencu, Cade Gordon, Ross Wightman, Mehdi Cherti, Theo Coombes, Aarush Katta, Clayton Mullis, Mitchell Wortsman, et al. Laion-5b: An open large-scale dataset for training next generation image-text models. Advances in Neural Information Processing Systems, 35:25278-25294, 2022. 5 +[53] Preethi Seshadri, Sameer Singh, and Yanai Elazar. The bias amplification paradox in text-to-image generation. In Proceedings of the 2024 Conference of the North American Chapter of the Association for Computational Linguistics: Human Language Technologies (Volume 1: Long Papers), pages 6367-6384, 2024. 2 +[54] Xudong Shen, Chao Du, Tianyu Pang, Min Lin, Yongkang Wong, and Mohan Kankanhalli. Finetuning text-to-image diffusion models for fairness. In The Twelfth International Conference on Learning Representations, 2024. 1, 2, 5, 7, 6, 11 +[55] Jiaming Song, Chenlin Meng, and Stefano Ermon. Denoising diffusion implicit models. In International Conference on Learning Representations, 2021. 1, 2, 3, 4 +[56] Mukund Sundararajan, Ankur Taly, and Qiqi Yan. Axiomatic attribution for deep networks. In Proceedings of the 34th International Conference on Machine Learning-Volume 70, pages 3319–3328, 2017. 4 +[57] Raphael Tang, Linqing Liu, Akshit Pandey, Zhiying Jiang, Gefei Yang, Karun Kumar, Pontus Stenetorp, Jimmy Lin, and Ferhan Ture. What the DAAM: Interpreting stable diffusion using cross attention. In Proceedings of the 61st Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers), pages 5644-5659, Toronto, Canada, 2023. Association for Computational Linguistics. 2 +[58] Mike Zajko. Conservative ai and social inequality: conceptualizing alternatives to bias through social theory. *Ai & Society*, 36(3):1047–1056, 2021. 8 +[59] Haiyan Zhao, Hanjie Chen, Fan Yang, Ninghao Liu, Huiqi Deng, Hengyi Cai, Shuaiqiang Wang, Dawei Yin, and Mengnan Du. Explainability for large language models: A survey. + +ACM Transactions on Intelligent Systems and Technology, 15(2):1-38, 2024. 1 +[60] Mi Zhou, Vibhanshu Abhishek, Timothy Derdenger, Jaymo Kim, and Kannan Srinivasan. Bias in generative ai. arXiv preprint arXiv:2403.02726, 2024. 8 \ No newline at end of file diff --git a/2025/Dissecting and Mitigating Diffusion Bias via Mechanistic Interpretability/images.zip b/2025/Dissecting and Mitigating Diffusion Bias via Mechanistic Interpretability/images.zip new file mode 100644 index 0000000000000000000000000000000000000000..43d098ee63d5510c4c1100fe9a2d7631b4e20354 --- /dev/null +++ b/2025/Dissecting and Mitigating Diffusion Bias via Mechanistic Interpretability/images.zip @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:17950ce3b6628af46385abd8d21ede5bb3fc512f852d4760982cb7414663a26d +size 589064 diff --git a/2025/Dissecting and Mitigating Diffusion Bias via Mechanistic Interpretability/layout.json b/2025/Dissecting and Mitigating Diffusion Bias via Mechanistic Interpretability/layout.json new file mode 100644 index 0000000000000000000000000000000000000000..b1be90c174a264fbda6e1fc6f47dc97180e79128 --- /dev/null +++ b/2025/Dissecting and Mitigating Diffusion Bias via Mechanistic Interpretability/layout.json @@ -0,0 +1,8843 @@ +{ + "pdf_info": [ + { + "para_blocks": [ + { + "bbox": [ + 81, + 103, + 531, + 121 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 81, + 103, + 531, + 121 + ], + "spans": [ + { + "bbox": [ + 81, + 103, + 531, + 121 + ], + "type": "text", + "content": "Dissecting and Mitigating Diffusion Bias via Mechanistic Interpretability" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 140, + 142, + 470, + 172 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 140, + 142, + 470, + 172 + ], + "spans": [ + { + "bbox": [ + 140, + 142, + 470, + 172 + ], + "type": "text", + "content": "Yingdong Shi" + }, + { + "bbox": [ + 140, + 142, + 470, + 172 + ], + "type": "inline_equation", + "content": "^{1*}" + }, + { + "bbox": [ + 140, + 142, + 470, + 172 + ], + "type": "text", + "content": " Changming Li" + }, + { + "bbox": [ + 140, + 142, + 470, + 172 + ], + "type": "inline_equation", + "content": "^{1*}" + }, + { + "bbox": [ + 140, + 142, + 470, + 172 + ], + "type": "text", + "content": " Yifan Wang" + }, + { + "bbox": [ + 140, + 142, + 470, + 172 + ], + "type": "inline_equation", + "content": "^{2}" + }, + { + "bbox": [ + 140, + 142, + 470, + 172 + ], + "type": "text", + "content": " Yongxiang Zhao" + }, + { + "bbox": [ + 140, + 142, + 470, + 172 + ], + "type": "inline_equation", + "content": "^{1}" + }, + { + "bbox": [ + 140, + 142, + 470, + 172 + ], + "type": "text", + "content": " Anqi Pang" + }, + { + "bbox": [ + 140, + 142, + 470, + 172 + ], + "type": "inline_equation", + "content": "^{3}" + }, + { + "bbox": [ + 140, + 142, + 470, + 172 + ], + "type": "text", + "content": " Sibei Yang" + }, + { + "bbox": [ + 140, + 142, + 470, + 172 + ], + "type": "inline_equation", + "content": "^{1}" + }, + { + "bbox": [ + 140, + 142, + 470, + 172 + ], + "type": "text", + "content": " Jingyi Yu" + }, + { + "bbox": [ + 140, + 142, + 470, + 172 + ], + "type": "inline_equation", + "content": "^{1}" + }, + { + "bbox": [ + 140, + 142, + 470, + 172 + ], + "type": "text", + "content": " Kan Ren" + }, + { + "bbox": [ + 140, + 142, + 470, + 172 + ], + "type": "inline_equation", + "content": "^{1\\dagger}" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 132, + 176, + 478, + 191 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 176, + 478, + 191 + ], + "spans": [ + { + "bbox": [ + 132, + 176, + 478, + 191 + ], + "type": "inline_equation", + "content": "^{1}" + }, + { + "bbox": [ + 132, + 176, + 478, + 191 + ], + "type": "text", + "content": "ShanghaiTech University " + }, + { + "bbox": [ + 132, + 176, + 478, + 191 + ], + "type": "inline_equation", + "content": "^{2}" + }, + { + "bbox": [ + 132, + 176, + 478, + 191 + ], + "type": "text", + "content": "Stony Brook University " + }, + { + "bbox": [ + 132, + 176, + 478, + 191 + ], + "type": "inline_equation", + "content": "^{3}" + }, + { + "bbox": [ + 132, + 176, + 478, + 191 + ], + "type": "text", + "content": "Tencent PCG" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 174, + 194, + 434, + 205 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 174, + 194, + 434, + 205 + ], + "spans": [ + { + "bbox": [ + 174, + 194, + 434, + 205 + ], + "type": "text", + "content": "{shiyd2023,lichm2024,renkan}@shanghaitech.edu.cn" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 151, + 232, + 200, + 245 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 151, + 232, + 200, + 245 + ], + "spans": [ + { + "bbox": [ + 151, + 232, + 200, + 245 + ], + "type": "text", + "content": "Abstract" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 55, + 258, + 297, + 568 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 258, + 297, + 568 + ], + "spans": [ + { + "bbox": [ + 55, + 258, + 297, + 568 + ], + "type": "text", + "content": "Diffusion models have demonstrated impressive capabilities in synthesizing diverse content. However, despite their high-quality outputs, these models often perpetuate social biases, including those related to gender and race. These biases can potentially contribute to harmful real-world consequences, reinforcing stereotypes and exacerbating inequalities in various social contexts. While existing research on diffusion bias mitigation has predominantly focused on guiding content generation, it often neglects the intrinsic mechanisms within diffusion models that causally drive biased outputs. In this paper, we investigate the internal processes of diffusion models, identifying specific decision-making mechanisms, termed bias features, embedded within the model architecture. By directly manipulating these features, our method precisely isolates and adjusts the elements responsible for bias generation, permitting granular control over the bias levels in the generated content. Through experiments on both unconditional and conditional diffusion models across various social bias attributes, we demonstrate our method's efficacy in managing generation distribution while preserving image quality. We also dissect the discovered model mechanism, revealing different intrinsic features controlling fine-grained aspects of generation, boosting further research on mechanistic interpretability of diffusion models. The project website is at https://foundation-model-research.github.io/difflens." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 55, + 591, + 135, + 604 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 591, + 135, + 604 + ], + "spans": [ + { + "bbox": [ + 55, + 591, + 135, + 604 + ], + "type": "text", + "content": "1. Introduction" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 55, + 611, + 296, + 696 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 611, + 296, + 696 + ], + "spans": [ + { + "bbox": [ + 55, + 611, + 296, + 696 + ], + "type": "text", + "content": "Diffusion models have emerged as powerful generative tools capable of generating high-quality content by iteratively denoising random noise [23, 55]. These models are now widely credited in applications such as content synthesis [49], data manipulation [29], and data augmentation [37], impacting numerous aspects of human society. However, diffusion models have been criticized for generating" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 313, + 234, + 555, + 341 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 234, + 555, + 341 + ], + "spans": [ + { + "bbox": [ + 313, + 234, + 555, + 341 + ], + "type": "text", + "content": "biased content [34, 43], often resulting in skewed representations of sensitive social attributes such as gender, race, and age. While biases may originate from the training data [36, 47], they can be further amplified by the learned model itself [47]. Given the extensive deployment of diffusion models, these biases may lead to detrimental effects in real-world contexts, such as reinforcing stereotypes in media representations or perpetuating inequalities in automated decision-making systems." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 313, + 342, + 556, + 484 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 342, + 556, + 484 + ], + "spans": [ + { + "bbox": [ + 313, + 342, + 556, + 484 + ], + "type": "text", + "content": "Existing approaches to debias diffusion models generally fall into two main strategies. The first attempts to train unbiased models from scratch on biased datasets [30] or fine-tune the target model [54], both of which are resource-intensive and can significantly impact the model's original performance. The second aims to guide or edit the generation process [17, 42, 43], but these methods undermine the distinct internal workings of diffusion models, risking overcorrecting the model behavior or distorting non-target attributes, which affecting generation quality. In contrast, few studies have explored the intrinsic mechanisms within diffusion models that causally drive biased outcomes." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 313, + 486, + 556, + 568 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 486, + 556, + 568 + ], + "spans": [ + { + "bbox": [ + 313, + 486, + 556, + 568 + ], + "type": "text", + "content": "In this work, we hypothesize that neurons collectively play a pivotal role in driving the generation of biased concepts. These neurons exhibit specific features which we identify as intrinsic mechanisms within the model that contribute to social bias. In particular, we set out to investigate these mechanisms to better understand and mitigate bias within diffusion models." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 313, + 570, + 557, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 570, + 557, + 713 + ], + "spans": [ + { + "bbox": [ + 313, + 570, + 557, + 713 + ], + "type": "text", + "content": "Our approach also aligns with the latest efforts on understanding the inner workings of neural networks, which has drawn huge interest within the community [7, 44], not only due to the interpretability benefits that reveal model decision patterns [31, 59], but also for its potential to enhance control over model behaviors [4, 38]. However, the intrinsic neuron activities within diffusion models remain largely unexplored. A significant challenge stems from the generative nature of diffusion models, which complicates, if not completely prohibits, fine-grained analysis at the neuron level. The neuron polysemanticity [4], which links individual neurons to multiple unrelated concepts additionally, fur" + } + ] + } + ], + "index": 13 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "spans": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "type": "text", + "content": "CVF" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 145, + 0, + 496, + 37 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 145, + 0, + 496, + 37 + ], + "spans": [ + { + "bbox": [ + 145, + 0, + 496, + 37 + ], + "type": "text", + "content": "This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 66, + 702, + 216, + 713 + ], + "type": "page_footnote", + "angle": 0, + "lines": [ + { + "bbox": [ + 66, + 702, + 216, + 713 + ], + "spans": [ + { + "bbox": [ + 66, + 702, + 216, + 713 + ], + "type": "text", + "content": "*Equal contribution. †Corresponding author." + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "text", + "content": "8192" + } + ] + } + ], + "index": 15 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 0 + }, + { + "para_blocks": [ + { + "bbox": [ + 55, + 72, + 294, + 95 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 72, + 294, + 95 + ], + "spans": [ + { + "bbox": [ + 55, + 72, + 294, + 95 + ], + "type": "text", + "content": "ther presents obstacles on mechanistic study of large-scale diffusion models." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 54, + 95, + 295, + 251 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 54, + 95, + 295, + 251 + ], + "spans": [ + { + "bbox": [ + 54, + 95, + 295, + 251 + ], + "type": "text", + "content": "To investigate the underlying mechanisms of bias generation in diffusion models, we present DIFFLENS, a comprehensive framework that analyzes the architectural components to identify the most influential elements within the model. Specifically, DIFFLENS first disentangles the hidden neurons of the diffusion model into a sparse semantic feature space via a sparse autoencoder [35], which captures the patterns of neurons responsible for bias features such as gender and age. Next, DIFFLENS employs a gradient-based approach to pinpoint key components in the interpretable semantic space, identifying a minimal set of bias-specific features. We can then further control or even mitigate the bias by adjusting the effects of these bias-related features." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 54, + 251, + 295, + 335 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 54, + 251, + 295, + 335 + ], + "spans": [ + { + "bbox": [ + 54, + 251, + 295, + 335 + ], + "type": "text", + "content": "DIFFLENS offers two key advantages. First, via elaborately analyzing and identifying the inner mechanisms of diffusion process, it enables fine-grained control of bias with minimal impact on other semantic attributes. Second, the identified bias features are intrinsic to the diffusion model, revealing its underlying contribution to biased generation. Both have been demonstrated in our experiments." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 56, + 336, + 295, + 539 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 336, + 295, + 539 + ], + "spans": [ + { + "bbox": [ + 56, + 336, + 295, + 539 + ], + "type": "text", + "content": "In summary, our work makes several key contributions to the field of diffusion models. (i) To the best of our knowledge, this is the first work to address bias mitigation in diffusion models through the lens of mechanistic interpretability, uncovering the underlying patterns of bias generation. Our method provides new insights into how specific neuron activities contribute to biased outcomes, offering a deeper understanding beyond standard latent space analysis. (ii) We introduce a novel framework that includes semantic space construction, identification of influential features, and fine-grained control of bias generation based on the discovered features. This approach is not only effective for bias mitigation, but may also be applied to other aspects of diffusion model analysis. (iii) We conduct extensive experiments on two pretrained diffusion models, demonstrating the effectiveness of our proposed mechanism discovery in comparison to several strong baselines for diffusion bias mitigation." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 55, + 550, + 142, + 562 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 550, + 142, + 562 + ], + "spans": [ + { + "bbox": [ + 55, + 550, + 142, + 562 + ], + "type": "text", + "content": "2. Related Work" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 55, + 570, + 295, + 677 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 570, + 295, + 677 + ], + "spans": [ + { + "bbox": [ + 55, + 570, + 295, + 677 + ], + "type": "text", + "content": "Bias in Diffusion Models. Diffusion-based architectures often reflect and amplify social biases [2, 34, 51], leading to potentially harmful outputs. Specifically, unconditional diffusion models exacerbate the biases present in the training data related to gender, race, and age [47]. Text-to-image models, such as Stable Diffusion [49], tend to generate images of specific gender or stereotypes when prompted with occupational terms or terms with character traits and other descriptors [5, 9, 34, 53]." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 55, + 677, + 295, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 677, + 295, + 713 + ], + "spans": [ + { + "bbox": [ + 55, + 677, + 295, + 713 + ], + "type": "text", + "content": "Mitigating Diffusion Bias. Developing techniques to mitigate social biases in generative models has become increasingly crucial [33, 43]. Many current approaches fo" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 313, + 72, + 553, + 300 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 72, + 553, + 300 + ], + "spans": [ + { + "bbox": [ + 313, + 72, + 553, + 300 + ], + "type": "text", + "content": "cus on modifying prompts, such as adding gender-specific cues or ethical information, or using prompt learning techniques [1, 11, 16, 28]. While these methods are effective for text-to-image systems, they are limited to scenarios involving textual prompts. Another strategy involves fine-tuning diffusion models using curated datasets [54], which can effectively reduce bias but require extensive training and expensive high-quality data. Guided generation and model editing based methods [17, 20, 42, 43] provide an alternative by introducing external guidance in the model's processing [20, 43] or manipulating attention weight [17, 42]. Although these techniques avoid retraining, they undermine the intrinsic mechanisms within the model specifically related to bias generation, often resulting in imprecise edits that inadvertently affect other parts of the image and degrade quality. In contrast, our approach allows for precise identification of bias-generating features, preserving overall image quality and maintaining other attributes, thereby providing a more reliable solution for bias mitigation." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 313, + 300, + 555, + 503 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 300, + 555, + 503 + ], + "spans": [ + { + "bbox": [ + 313, + 300, + 555, + 503 + ], + "type": "text", + "content": "Interpretability of Diffusion Models. Most of the work interpreting diffusion models focus on explaining how the input affects the generation via attribution methods and attention maps [8, 19, 21, 57], but they do not reveal the underlying mechanisms by which these attributes are processed and synthesized into a coherent output. Recent research has revealed that diffusion models possess a semantically rich latent space [32], inspiring various studies on customized image editing through latent space manipulation [18, 33, 45]. Although these methods enable more flexible editing, they still do not provide a deep understanding of the model's inherent features, leading to unexpected results. Our approach directly interacts with the internal components of diffusion models, enabling us to identify and precisely modify key features like bias-specific ones. This allows for more accurate control and targeted editing within the model itself, also providing efficient tools for interpreting diffusion models." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 313, + 515, + 391, + 529 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 515, + 391, + 529 + ], + "spans": [ + { + "bbox": [ + 313, + 515, + 391, + 529 + ], + "type": "text", + "content": "3. Preliminary" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 313, + 536, + 553, + 632 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 536, + 553, + 632 + ], + "spans": [ + { + "bbox": [ + 313, + 536, + 553, + 632 + ], + "type": "text", + "content": "In this section, we provide an overview of diffusion model architecture. Diffusion models operate through two key processes, the forward and backward diffusion steps [23]. Taking image generation as an example. In the forward process, noise is incrementally added to the input, transforming a real image " + }, + { + "bbox": [ + 313, + 536, + 553, + 632 + ], + "type": "inline_equation", + "content": "\\mathbf{x}_0" + }, + { + "bbox": [ + 313, + 536, + 553, + 632 + ], + "type": "text", + "content": " into pure noise over a series of timesteps " + }, + { + "bbox": [ + 313, + 536, + 553, + 632 + ], + "type": "inline_equation", + "content": "t\\in [0,T)" + }, + { + "bbox": [ + 313, + 536, + 553, + 632 + ], + "type": "text", + "content": " where " + }, + { + "bbox": [ + 313, + 536, + 553, + 632 + ], + "type": "inline_equation", + "content": "T" + }, + { + "bbox": [ + 313, + 536, + 553, + 632 + ], + "type": "text", + "content": " is the total number of timesteps. Mathematically, the forward process can be represented as" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 353, + 641, + 553, + 656 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 353, + 641, + 553, + 656 + ], + "spans": [ + { + "bbox": [ + 353, + 641, + 553, + 656 + ], + "type": "interline_equation", + "content": "q \\left(\\mathbf {x} _ {t} \\mid \\mathbf {x} _ {0}\\right) = \\mathcal {N} \\left(\\mathbf {x} _ {t}; \\sqrt {\\bar {\\alpha} _ {t}} \\mathbf {x} _ {0}, (1 - \\bar {\\alpha} _ {t}) I\\right). \\tag {1}", + "image_path": "63b092667dd666e8b57cd0c7d4d31b9ddda6e709cc2950aaca57f5470367309f.jpg" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 313, + 664, + 553, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 664, + 553, + 715 + ], + "spans": [ + { + "bbox": [ + 313, + 664, + 553, + 715 + ], + "type": "text", + "content": "Here, " + }, + { + "bbox": [ + 313, + 664, + 553, + 715 + ], + "type": "inline_equation", + "content": "\\bar{\\alpha}_t\\coloneqq \\prod_{i = 1}^t\\alpha_i" + }, + { + "bbox": [ + 313, + 664, + 553, + 715 + ], + "type": "text", + "content": " controls the amount of noise injected at each timestep " + }, + { + "bbox": [ + 313, + 664, + 553, + 715 + ], + "type": "inline_equation", + "content": "t" + }, + { + "bbox": [ + 313, + 664, + 553, + 715 + ], + "type": "text", + "content": ", with " + }, + { + "bbox": [ + 313, + 664, + 553, + 715 + ], + "type": "inline_equation", + "content": "\\mathbf{x}_T" + }, + { + "bbox": [ + 313, + 664, + 553, + 715 + ], + "type": "text", + "content": " becoming a near-random noise vector after the final timestep. In the backward process, the goal is to reverse this noising process and generate a clean" + } + ] + } + ], + "index": 12 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 315, + 757 + ], + "type": "text", + "content": "8193" + } + ] + } + ], + "index": 13 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 1 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 70, + 69, + 541, + 254 + ], + "blocks": [ + { + "bbox": [ + 70, + 69, + 541, + 254 + ], + "lines": [ + { + "bbox": [ + 70, + 69, + 541, + 254 + ], + "spans": [ + { + "bbox": [ + 70, + 69, + 541, + 254 + ], + "type": "image", + "image_path": "f7139f10dd518f79fc8fa4a8c35d3448900402f917d4d54ba02be2610450e446.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 54, + 262, + 555, + 298 + ], + "lines": [ + { + "bbox": [ + 54, + 262, + 555, + 298 + ], + "spans": [ + { + "bbox": [ + 54, + 262, + 555, + 298 + ], + "type": "text", + "content": "Figure 1. Framework of DIFFLENS exploring the inner working of diffusion models for bias mitigation. Neurons that are activated in the semantic space are defined as fired features and inactivated ones as unfired features. Target features are either suppressed or amplified to control bias level. We divide the framework into three parts (a), (b) and (c) which correspond to Secs. 4.1 to 4.3, respectively." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "bbox": [ + 54, + 306, + 294, + 355 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 54, + 306, + 294, + 355 + ], + "spans": [ + { + "bbox": [ + 54, + 306, + 294, + 355 + ], + "type": "text", + "content": "image from the noisy sample. At each step, the diffusion network, e.g., U-Net [50], models " + }, + { + "bbox": [ + 54, + 306, + 294, + 355 + ], + "type": "inline_equation", + "content": "p_{\\theta}(\\mathbf{x}_{t - 1}|\\mathbf{x}_t)" + }, + { + "bbox": [ + 54, + 306, + 294, + 355 + ], + "type": "text", + "content": " with parameter " + }, + { + "bbox": [ + 54, + 306, + 294, + 355 + ], + "type": "inline_equation", + "content": "\\theta" + }, + { + "bbox": [ + 54, + 306, + 294, + 355 + ], + "type": "text", + "content": " predicting the denoised version given the current sample " + }, + { + "bbox": [ + 54, + 306, + 294, + 355 + ], + "type": "inline_equation", + "content": "\\mathbf{x}_t" + }, + { + "bbox": [ + 54, + 306, + 294, + 355 + ], + "type": "text", + "content": ". This reverse process is modeled as" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 80, + 366, + 295, + 380 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 80, + 366, + 295, + 380 + ], + "spans": [ + { + "bbox": [ + 80, + 366, + 295, + 380 + ], + "type": "interline_equation", + "content": "p _ {\\theta} (\\mathbf {x} _ {t - 1} | \\mathbf {x} _ {t}) = \\mathcal {N} (\\mathbf {x} _ {t - 1}; \\mu_ {\\theta} (\\mathbf {x} _ {t}, t), \\Sigma_ {\\theta} (\\mathbf {x} _ {t}, t)), \\tag {2}", + "image_path": "8de349abef6a7f5471d721e1b569dfbbbf0071b3e97feffc62dd6a61343d1197.jpg" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 55, + 390, + 295, + 415 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 390, + 295, + 415 + ], + "spans": [ + { + "bbox": [ + 55, + 390, + 295, + 415 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 55, + 390, + 295, + 415 + ], + "type": "inline_equation", + "content": "\\mu_{\\theta}" + }, + { + "bbox": [ + 55, + 390, + 295, + 415 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 55, + 390, + 295, + 415 + ], + "type": "inline_equation", + "content": "\\Sigma_{\\theta}" + }, + { + "bbox": [ + 55, + 390, + 295, + 415 + ], + "type": "text", + "content": " represent the mean and variance of the predicted denoised sample." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 55, + 415, + 296, + 511 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 415, + 296, + 511 + ], + "spans": [ + { + "bbox": [ + 55, + 415, + 296, + 511 + ], + "type": "text", + "content": "Our work specifically focuses on the inner working of model itself, such as U-Net [50] architecture in the diffusion model, to investigate the mechanisms underlying bias generation. We aim to examine the model's decision-making process through these representations to better understand the biases that emerge during image generation, providing deeper insights into the model's internal functions and the precise factors driving its behavior [4]." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 55, + 525, + 138, + 539 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 525, + 138, + 539 + ], + "spans": [ + { + "bbox": [ + 55, + 525, + 138, + 539 + ], + "type": "text", + "content": "4. Methodology" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 55, + 546, + 295, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 546, + 295, + 713 + ], + "spans": [ + { + "bbox": [ + 55, + 546, + 295, + 713 + ], + "type": "text", + "content": "In this section, we propose DIFFLENS, consisting of three stages aimed at uncovering internal mechanisms of bias generation in diffusion models and leveraging this understanding for bias mitigation. Specifically, our goal is to discover the neuron activities that controls the generation of biased contents. To achieve this, DIFFLENS first disentangles the hidden states of neuron activation from the model layer into a sparse semantic space, capturing meaningful features that represent various attributes. Next, it identifies the semantic features specifically related to biased content generation using a gradient-based attribution method. Finally, by adjusting the scale of these identified features, DIFFLENS allows for control over the level of bias in generated content, providing an effective approach for bias mitigation." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 313, + 306, + 553, + 331 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 306, + 553, + 331 + ], + "spans": [ + { + "bbox": [ + 313, + 306, + 553, + 331 + ], + "type": "text", + "content": "The overall framework has been illustrated in Fig. 1 and Algorithms 1 and 2 in Appendix A." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 313, + 336, + 509, + 350 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 336, + 509, + 350 + ], + "spans": [ + { + "bbox": [ + 313, + 336, + 509, + 350 + ], + "type": "text", + "content": "4.1. Disentangling Polysemantic Neurons" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 312, + 354, + 555, + 545 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 312, + 354, + 555, + 545 + ], + "spans": [ + { + "bbox": [ + 312, + 354, + 555, + 545 + ], + "type": "text", + "content": "We study the hidden states " + }, + { + "bbox": [ + 312, + 354, + 555, + 545 + ], + "type": "inline_equation", + "content": "\\mathbf{h} = [h_1, h_2, \\dots, h_n] \\in \\mathbb{R}^n" + }, + { + "bbox": [ + 312, + 354, + 555, + 545 + ], + "type": "text", + "content": " from the diffusion model, representing the activations of the " + }, + { + "bbox": [ + 312, + 354, + 555, + 545 + ], + "type": "inline_equation", + "content": "n" + }, + { + "bbox": [ + 312, + 354, + 555, + 545 + ], + "type": "text", + "content": " neurons at each network layer. We assume that these neuron activations inherently capture the semantic information of the generation process. Similar observations have been made in language models [12, 38], though little attention has been given to diffusion models in this regard. Kwon et al. [32] demonstrated that diffusion models exhibit a semantic space within the U-Net [50] architecture, specifically at the bottleneck layer. However, previous works have not precisely analyzed the different attributes within this space. In contrast, our method explicitly identifies semantic features and pinpoints the specific dimensions of neuron activations responsible for bias generation. This enables more precise and targeted control over bias, as demonstrated in our experiments in Figs. 3 and 4." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 313, + 546, + 556, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 546, + 556, + 713 + ], + "spans": [ + { + "bbox": [ + 313, + 546, + 556, + 713 + ], + "type": "text", + "content": "Note that, neurons are often polysemantic, meaning that each neuron often corresponds to multiple, unrelated concepts, as discussed in [14, 39, 41]. This places obstacles for understanding and interpreting the mechanism of the model, as the network's learned features are intertwined in a complex manner [4]. Nevertheless, it also presents an opportunity that these features can potentially be disentangled by identifying directions in the activation space, where each activation vector can be expressed as a sparse linear combination of underlying features [24]. Thus, we leverage k-Sparse Autoencoders (k-SAE) [35] to analyze and interpret the semantic features encoded within the hidden states " + }, + { + "bbox": [ + 313, + 546, + 556, + 713 + ], + "type": "inline_equation", + "content": "\\mathbf{h}" + }, + { + "bbox": [ + 313, + 546, + 556, + 713 + ], + "type": "text", + "content": ". It helps the method focus on salient features, disentangling polysemantic activations and isolating semantic attributes" + } + ] + } + ], + "index": 11 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "text", + "content": "8194" + } + ] + } + ], + "index": 12 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 2 + }, + { + "para_blocks": [ + { + "bbox": [ + 55, + 72, + 295, + 108 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 72, + 295, + 108 + ], + "spans": [ + { + "bbox": [ + 55, + 72, + 295, + 108 + ], + "type": "text", + "content": "within the latent space, making it effective for understanding and manipulating key features in diffusion models. Part (a) of Fig. 1 has illustrated this processing stage." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 55, + 108, + 296, + 204 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 108, + 296, + 204 + ], + "spans": [ + { + "bbox": [ + 55, + 108, + 296, + 204 + ], + "type": "text", + "content": "Specifically, let " + }, + { + "bbox": [ + 55, + 108, + 296, + 204 + ], + "type": "inline_equation", + "content": "\\mathbf{H}^n = \\{\\mathbf{h}\\}" + }, + { + "bbox": [ + 55, + 108, + 296, + 204 + ], + "type": "text", + "content": " be the original latent feature space with dimension number " + }, + { + "bbox": [ + 55, + 108, + 296, + 204 + ], + "type": "inline_equation", + "content": "n" + }, + { + "bbox": [ + 55, + 108, + 296, + 204 + ], + "type": "text", + "content": ". We transform this feature space to a sparse high-dimensional space, denoted as " + }, + { + "bbox": [ + 55, + 108, + 296, + 204 + ], + "type": "inline_equation", + "content": "\\mathbf{S}^m = \\{\\mathbf{s}\\}" + }, + { + "bbox": [ + 55, + 108, + 296, + 204 + ], + "type": "text", + "content": " with dimension " + }, + { + "bbox": [ + 55, + 108, + 296, + 204 + ], + "type": "inline_equation", + "content": "m" + }, + { + "bbox": [ + 55, + 108, + 296, + 204 + ], + "type": "text", + "content": " that is much larger than " + }, + { + "bbox": [ + 55, + 108, + 296, + 204 + ], + "type": "inline_equation", + "content": "n" + }, + { + "bbox": [ + 55, + 108, + 296, + 204 + ], + "type": "text", + "content": ". In the space " + }, + { + "bbox": [ + 55, + 108, + 296, + 204 + ], + "type": "inline_equation", + "content": "\\mathbf{S}^m" + }, + { + "bbox": [ + 55, + 108, + 296, + 204 + ], + "type": "text", + "content": ", the features " + }, + { + "bbox": [ + 55, + 108, + 296, + 204 + ], + "type": "inline_equation", + "content": "\\mathbf{s}" + }, + { + "bbox": [ + 55, + 108, + 296, + 204 + ], + "type": "text", + "content": " are monosemantic, i.e., each feature represents a distinct, non-overlapping semantic concept (discussions are in Appendix E.2). The mapping " + }, + { + "bbox": [ + 55, + 108, + 296, + 204 + ], + "type": "inline_equation", + "content": "\\Phi : \\mathbf{H}^n \\to \\mathbf{S}^m" + }, + { + "bbox": [ + 55, + 108, + 296, + 204 + ], + "type": "text", + "content": " can be expressed by" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 59, + 209, + 296, + 233 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 59, + 209, + 296, + 233 + ], + "spans": [ + { + "bbox": [ + 59, + 209, + 296, + 233 + ], + "type": "interline_equation", + "content": "\\mathbf {s} = \\left[ s _ {1}, s _ {2}, \\dots , s _ {m} \\right] = \\Phi (\\mathbf {h}) = \\operatorname {T o p K} \\left(\\mathbf {W} _ {\\text {e n c}} \\left(\\mathbf {h} - \\mathbf {b} _ {\\text {p r e}}\\right)\\right). \\tag {3}", + "image_path": "4c470e1ea509d86e92fa3a9e8ee25cda81c2b0af51b81b362c867436d935f118.jpg" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 55, + 233, + 296, + 330 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 233, + 296, + 330 + ], + "spans": [ + { + "bbox": [ + 55, + 233, + 296, + 330 + ], + "type": "text", + "content": "We map the vector " + }, + { + "bbox": [ + 55, + 233, + 296, + 330 + ], + "type": "inline_equation", + "content": "\\mathbf{h}" + }, + { + "bbox": [ + 55, + 233, + 296, + 330 + ], + "type": "text", + "content": " to a sparse vector " + }, + { + "bbox": [ + 55, + 233, + 296, + 330 + ], + "type": "inline_equation", + "content": "\\mathbf{s}" + }, + { + "bbox": [ + 55, + 233, + 296, + 330 + ], + "type": "text", + "content": " with " + }, + { + "bbox": [ + 55, + 233, + 296, + 330 + ], + "type": "inline_equation", + "content": "\\| \\mathbf{s} \\|_0 = k" + }, + { + "bbox": [ + 55, + 233, + 296, + 330 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 55, + 233, + 296, + 330 + ], + "type": "inline_equation", + "content": "k < m" + }, + { + "bbox": [ + 55, + 233, + 296, + 330 + ], + "type": "text", + "content": ", indicating that " + }, + { + "bbox": [ + 55, + 233, + 296, + 330 + ], + "type": "inline_equation", + "content": "\\mathbf{s}" + }, + { + "bbox": [ + 55, + 233, + 296, + 330 + ], + "type": "text", + "content": " is sparse with only " + }, + { + "bbox": [ + 55, + 233, + 296, + 330 + ], + "type": "inline_equation", + "content": "k" + }, + { + "bbox": [ + 55, + 233, + 296, + 330 + ], + "type": "text", + "content": " features activated. This is achieved by the encoding process of a k-SAE where " + }, + { + "bbox": [ + 55, + 233, + 296, + 330 + ], + "type": "inline_equation", + "content": "\\mathbf{W}_{\\mathrm{enc}} \\in \\mathbb{R}^{m \\times n}" + }, + { + "bbox": [ + 55, + 233, + 296, + 330 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 55, + 233, + 296, + 330 + ], + "type": "inline_equation", + "content": "\\mathbf{b}_{\\mathrm{pre}} \\in \\mathbb{R}^n" + }, + { + "bbox": [ + 55, + 233, + 296, + 330 + ], + "type": "text", + "content": " is the bias term. " + }, + { + "bbox": [ + 55, + 233, + 296, + 330 + ], + "type": "inline_equation", + "content": "\\operatorname{TopK}(\\cdot)" + }, + { + "bbox": [ + 55, + 233, + 296, + 330 + ], + "type": "text", + "content": " is an operation that selects the top " + }, + { + "bbox": [ + 55, + 233, + 296, + 330 + ], + "type": "inline_equation", + "content": "K = k" + }, + { + "bbox": [ + 55, + 233, + 296, + 330 + ], + "type": "text", + "content": " largest features, referred to as \"fired\" features, while setting the remaining features to zero, which are \"unfired\" features as shown in Fig. 1. For the decoding process, we have" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 111, + 333, + 295, + 348 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 111, + 333, + 295, + 348 + ], + "spans": [ + { + "bbox": [ + 111, + 333, + 295, + 348 + ], + "type": "interline_equation", + "content": "\\hat {\\mathbf {h}} = \\Phi^ {- 1} (\\mathbf {s}) = \\mathbf {W} _ {\\mathrm {d e c}} \\mathbf {s} + \\mathbf {b} _ {\\mathrm {p r e}}, \\tag {4}", + "image_path": "d1d980fec997825cfd5c0534b209ddfb607316b148a6effeff633aee700fe4d8.jpg" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 55, + 354, + 296, + 426 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 354, + 296, + 426 + ], + "spans": [ + { + "bbox": [ + 55, + 354, + 296, + 426 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 55, + 354, + 296, + 426 + ], + "type": "inline_equation", + "content": "\\hat{\\mathbf{h}}\\in \\mathbb{R}^n" + }, + { + "bbox": [ + 55, + 354, + 296, + 426 + ], + "type": "text", + "content": " is the reconstructed latent vector and " + }, + { + "bbox": [ + 55, + 354, + 296, + 426 + ], + "type": "inline_equation", + "content": "\\mathbf{W}_{\\mathrm{dec}}\\in \\mathbb{R}^{n\\times m}" + }, + { + "bbox": [ + 55, + 354, + 296, + 426 + ], + "type": "text", + "content": ". " + }, + { + "bbox": [ + 55, + 354, + 296, + 426 + ], + "type": "inline_equation", + "content": "\\Phi^{-1}" + }, + { + "bbox": [ + 55, + 354, + 296, + 426 + ], + "type": "text", + "content": " represents the inverse mapping of " + }, + { + "bbox": [ + 55, + 354, + 296, + 426 + ], + "type": "inline_equation", + "content": "\\Phi" + }, + { + "bbox": [ + 55, + 354, + 296, + 426 + ], + "type": "text", + "content": " since " + }, + { + "bbox": [ + 55, + 354, + 296, + 426 + ], + "type": "inline_equation", + "content": "\\mathbf{W}_{\\mathrm{dec}}" + }, + { + "bbox": [ + 55, + 354, + 296, + 426 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 55, + 354, + 296, + 426 + ], + "type": "inline_equation", + "content": "\\mathbf{W}_{\\mathrm{enc}}" + }, + { + "bbox": [ + 55, + 354, + 296, + 426 + ], + "type": "text", + "content": " are initialized as transposes of each other. We train the k-SAE using the loss function defined in Eq. (5), with the objective of minimizing the reconstruction error within the original hidden space, expressed as" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 134, + 430, + 295, + 444 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 134, + 430, + 295, + 444 + ], + "spans": [ + { + "bbox": [ + 134, + 430, + 295, + 444 + ], + "type": "interline_equation", + "content": "\\mathcal {L} (\\mathbf {h}) = \\left\\| \\mathbf {h} - \\hat {\\mathbf {h}} \\right\\| _ {2} ^ {2}, \\tag {5}", + "image_path": "4ae1d25c8ec66d7e80daa00537fdb64be2a2e28f205aeb34ae5d2283321dccf3.jpg" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 55, + 449, + 296, + 521 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 449, + 296, + 521 + ], + "spans": [ + { + "bbox": [ + 55, + 449, + 296, + 521 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 55, + 449, + 296, + 521 + ], + "type": "inline_equation", + "content": "\\mathbf{h}" + }, + { + "bbox": [ + 55, + 449, + 296, + 521 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 55, + 449, + 296, + 521 + ], + "type": "inline_equation", + "content": "\\hat{\\mathbf{h}}" + }, + { + "bbox": [ + 55, + 449, + 296, + 521 + ], + "type": "text", + "content": " refers to the original and the reconstructed hidden vectors, respectively. We present results and implementation details in Appendix B.2, demonstrating that we reconstruct the original latent space with few influences on the target model. The encoding and decoding processes are presented in Algorithm 2 at Appendix A." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 55, + 522, + 296, + 666 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 522, + 296, + 666 + ], + "spans": [ + { + "bbox": [ + 55, + 522, + 296, + 666 + ], + "type": "text", + "content": "Disentangling the original latent space of diffusion model shares several advantages. On one hand, transforming it into a sparse semantic space can well handle the polysemistic issue of the internal neurons within the neural network [24]. On the other hand, it enables more precise locating on the target feature for generation content attributes, such as identifying the bias features in the subsequent sections in our paper. In addition, editing the whole state in the original hidden space like that in [18, 32, 33, 43] could inadvertently alter irrelevant attributes, caused by the polysemantic properties of the neuron in the model as shown by Figs. 2 and 3 in experiments." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 55, + 671, + 196, + 684 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 671, + 196, + 684 + ], + "spans": [ + { + "bbox": [ + 55, + 671, + 196, + 684 + ], + "type": "text", + "content": "4.2. Identifying Bias Features" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 55, + 689, + 296, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 689, + 296, + 713 + ], + "spans": [ + { + "bbox": [ + 55, + 689, + 296, + 713 + ], + "type": "text", + "content": "The next step is to identify the most influential features within the sparse semantic space that are associated with" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 313, + 72, + 553, + 133 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 72, + 553, + 133 + ], + "spans": [ + { + "bbox": [ + 313, + 72, + 553, + 133 + ], + "type": "text", + "content": "particular function, i.e., bias generation mechanism in this paper. This need arises because accurately identifying the features that correspond to specific attributes can significantly enhance our ability to control and manipulate generated content responsibly." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 313, + 133, + 554, + 204 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 133, + 554, + 204 + ], + "spans": [ + { + "bbox": [ + 313, + 133, + 554, + 204 + ], + "type": "text", + "content": "Inspired by [13, 56], we propose a gradient-based bias attribution method, to locate the features that correlates to the bias concepts. We visualize this process in part (b) of Fig. 1. Intuitively, the neuron activations that influence the outcome of gender diverse in different aspects, which controls different characteristics such as hair style in Fig. 6." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 313, + 204, + 555, + 348 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 204, + 555, + 348 + ], + "spans": [ + { + "bbox": [ + 313, + 204, + 555, + 348 + ], + "type": "text", + "content": "Our goal is to measure the influence of the semantic feature " + }, + { + "bbox": [ + 313, + 204, + 555, + 348 + ], + "type": "inline_equation", + "content": "s_i \\in \\mathbf{s}" + }, + { + "bbox": [ + 313, + 204, + 555, + 348 + ], + "type": "text", + "content": " within " + }, + { + "bbox": [ + 313, + 204, + 555, + 348 + ], + "type": "inline_equation", + "content": "\\mathbf{S}^m" + }, + { + "bbox": [ + 313, + 204, + 555, + 348 + ], + "type": "text", + "content": " driving the biased contents on the target outcome. We define " + }, + { + "bbox": [ + 313, + 204, + 555, + 348 + ], + "type": "inline_equation", + "content": "F_{\\mathbf{x}}: \\mathbf{S}^m \\to \\mathbb{R}" + }, + { + "bbox": [ + 313, + 204, + 555, + 348 + ], + "type": "text", + "content": " as a measure of interest, assessing the outcome of the model w.r.t. the target " + }, + { + "bbox": [ + 313, + 204, + 555, + 348 + ], + "type": "inline_equation", + "content": "\\mathbf{s}" + }, + { + "bbox": [ + 313, + 204, + 555, + 348 + ], + "type": "text", + "content": " given the input " + }, + { + "bbox": [ + 313, + 204, + 555, + 348 + ], + "type": "inline_equation", + "content": "\\mathbf{x}" + }, + { + "bbox": [ + 313, + 204, + 555, + 348 + ], + "type": "text", + "content": ". Specifically, we quantify the bias level as the probability of the social attribute class " + }, + { + "bbox": [ + 313, + 204, + 555, + 348 + ], + "type": "inline_equation", + "content": "y" + }, + { + "bbox": [ + 313, + 204, + 555, + 348 + ], + "type": "text", + "content": " as " + }, + { + "bbox": [ + 313, + 204, + 555, + 348 + ], + "type": "inline_equation", + "content": "F_{\\mathbf{x}}(\\mathbf{s}) = \\operatorname*{Pr}(y|\\mathbf{s})" + }, + { + "bbox": [ + 313, + 204, + 555, + 348 + ], + "type": "text", + "content": " regarding the disentangled feature vectors derived from Eq. (3). Here " + }, + { + "bbox": [ + 313, + 204, + 555, + 348 + ], + "type": "inline_equation", + "content": "y" + }, + { + "bbox": [ + 313, + 204, + 555, + 348 + ], + "type": "text", + "content": " represents different classes such as \"male\" or \"female\" in gender bias. To locate the influential features, the internal mechanisms of diffusion model, we propose the attribution score of bias generation to the disentangled feature " + }, + { + "bbox": [ + 313, + 204, + 555, + 348 + ], + "type": "inline_equation", + "content": "s_i" + }, + { + "bbox": [ + 313, + 204, + 555, + 348 + ], + "type": "text", + "content": " as" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 318, + 355, + 554, + 384 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 318, + 355, + 554, + 384 + ], + "spans": [ + { + "bbox": [ + 318, + 355, + 554, + 384 + ], + "type": "interline_equation", + "content": "S (s _ {i}; \\mathbf {x}) = (s _ {i} - s _ {i} ^ {\\prime}) \\cdot \\int_ {\\alpha = 0} ^ {1} \\frac {\\partial F _ {\\mathbf {x}} (\\mathbf {s} ^ {\\prime} + \\alpha (\\mathbf {s} - \\mathbf {s} ^ {\\prime}))}{\\partial s _ {i}} d \\alpha , \\tag {6}", + "image_path": "6fbdb64ccce64370ceeb8c7b2a623019ce58d1d573a695e0dd1a82b5a9e9fba7.jpg" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 313, + 390, + 554, + 474 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 390, + 554, + 474 + ], + "spans": [ + { + "bbox": [ + 313, + 390, + 554, + 474 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 313, + 390, + 554, + 474 + ], + "type": "inline_equation", + "content": "\\mathbf{s}' = [s_1',\\dots,s_m']" + }, + { + "bbox": [ + 313, + 390, + 554, + 474 + ], + "type": "text", + "content": " is a relative baseline of " + }, + { + "bbox": [ + 313, + 390, + 554, + 474 + ], + "type": "inline_equation", + "content": "\\mathbf{s}" + }, + { + "bbox": [ + 313, + 390, + 554, + 474 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 313, + 390, + 554, + 474 + ], + "type": "inline_equation", + "content": "\\partial F_{\\mathbf{x}}(\\mathbf{s}' + \\alpha (\\mathbf{s} - \\mathbf{s}')) / \\partial s_i" + }, + { + "bbox": [ + 313, + 390, + 554, + 474 + ], + "type": "text", + "content": " is the gradient of the bias measure given the input image " + }, + { + "bbox": [ + 313, + 390, + 554, + 474 + ], + "type": "inline_equation", + "content": "\\mathbf{x}" + }, + { + "bbox": [ + 313, + 390, + 554, + 474 + ], + "type": "text", + "content": ", to the target feature space " + }, + { + "bbox": [ + 313, + 390, + 554, + 474 + ], + "type": "inline_equation", + "content": "\\mathbf{s}" + }, + { + "bbox": [ + 313, + 390, + 554, + 474 + ], + "type": "text", + "content": ". It is achieved by a light-weight classifier " + }, + { + "bbox": [ + 313, + 390, + 554, + 474 + ], + "type": "inline_equation", + "content": "F_{\\mathbf{x}}(\\cdot ;\\psi)" + }, + { + "bbox": [ + 313, + 390, + 554, + 474 + ], + "type": "text", + "content": " with parameter " + }, + { + "bbox": [ + 313, + 390, + 554, + 474 + ], + "type": "inline_equation", + "content": "\\psi" + }, + { + "bbox": [ + 313, + 390, + 554, + 474 + ], + "type": "text", + "content": ". The classifier construction follows [43] and we provide more training details in Appendix B.3. Choices of baselines " + }, + { + "bbox": [ + 313, + 390, + 554, + 474 + ], + "type": "inline_equation", + "content": "s'" + }, + { + "bbox": [ + 313, + 390, + 554, + 474 + ], + "type": "text", + "content": " are in Appendix A.1." + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 313, + 474, + 554, + 569 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 474, + 554, + 569 + ], + "spans": [ + { + "bbox": [ + 313, + 474, + 554, + 569 + ], + "type": "text", + "content": "In Eq. (6), the attribution score " + }, + { + "bbox": [ + 313, + 474, + 554, + 569 + ], + "type": "inline_equation", + "content": "S(s_{i};\\mathbf{x})" + }, + { + "bbox": [ + 313, + 474, + 554, + 569 + ], + "type": "text", + "content": " aggregates the gradients along the path of the individual semantic feature " + }, + { + "bbox": [ + 313, + 474, + 554, + 569 + ], + "type": "inline_equation", + "content": "s_i" + }, + { + "bbox": [ + 313, + 474, + 554, + 569 + ], + "type": "text", + "content": " ranging from zero effect " + }, + { + "bbox": [ + 313, + 474, + 554, + 569 + ], + "type": "inline_equation", + "content": "(\\alpha = 0)" + }, + { + "bbox": [ + 313, + 474, + 554, + 569 + ], + "type": "text", + "content": " to full effect " + }, + { + "bbox": [ + 313, + 474, + 554, + 569 + ], + "type": "inline_equation", + "content": "(\\alpha = 1)" + }, + { + "bbox": [ + 313, + 474, + 554, + 569 + ], + "type": "text", + "content": ", given the input " + }, + { + "bbox": [ + 313, + 474, + 554, + 569 + ], + "type": "inline_equation", + "content": "\\mathbf{x}" + }, + { + "bbox": [ + 313, + 474, + 554, + 569 + ], + "type": "text", + "content": ", which is the generated content from the diffusion model. If a feature " + }, + { + "bbox": [ + 313, + 474, + 554, + 569 + ], + "type": "inline_equation", + "content": "s_i" + }, + { + "bbox": [ + 313, + 474, + 554, + 569 + ], + "type": "text", + "content": " significantly impacts the representation of a particular attribute, its corresponding gradient will be prominent, resulting in higher integrated values, i.e., higher attribution score " + }, + { + "bbox": [ + 313, + 474, + 554, + 569 + ], + "type": "inline_equation", + "content": "S" + }, + { + "bbox": [ + 313, + 474, + 554, + 569 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 313, + 570, + 554, + 700 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 570, + 554, + 700 + ], + "spans": [ + { + "bbox": [ + 313, + 570, + 554, + 700 + ], + "type": "text", + "content": "Further, to assess the overall influence of the semantic feature " + }, + { + "bbox": [ + 313, + 570, + 554, + 700 + ], + "type": "inline_equation", + "content": "s_i" + }, + { + "bbox": [ + 313, + 570, + 554, + 700 + ], + "type": "text", + "content": " to the bias contents, we aggregate the attribution score as " + }, + { + "bbox": [ + 313, + 570, + 554, + 700 + ], + "type": "inline_equation", + "content": "S(s_i; X) = \\sum_{j=1}^{N} S(s_i; \\mathbf{x}_j)" + }, + { + "bbox": [ + 313, + 570, + 554, + 700 + ], + "type": "text", + "content": " where " + }, + { + "bbox": [ + 313, + 570, + 554, + 700 + ], + "type": "inline_equation", + "content": "\\mathbf{x}_j \\in X" + }, + { + "bbox": [ + 313, + 570, + 554, + 700 + ], + "type": "text", + "content": " with " + }, + { + "bbox": [ + 313, + 570, + 554, + 700 + ], + "type": "inline_equation", + "content": "X" + }, + { + "bbox": [ + 313, + 570, + 554, + 700 + ], + "type": "text", + "content": " as the support dataset of " + }, + { + "bbox": [ + 313, + 570, + 554, + 700 + ], + "type": "inline_equation", + "content": "N" + }, + { + "bbox": [ + 313, + 570, + 554, + 700 + ], + "type": "text", + "content": " generated samples. DIFFLENS then selects the top " + }, + { + "bbox": [ + 313, + 570, + 554, + 700 + ], + "type": "inline_equation", + "content": "\\tau" + }, + { + "bbox": [ + 313, + 570, + 554, + 700 + ], + "type": "text", + "content": " features with the largest attribution scores as bias features, whose index set is " + }, + { + "bbox": [ + 313, + 570, + 554, + 700 + ], + "type": "inline_equation", + "content": "\\mathbf{A} = \\{i_1, \\ldots, i_\\tau\\}" + }, + { + "bbox": [ + 313, + 570, + 554, + 700 + ], + "type": "text", + "content": ", i.e., these are the most influential semantic features within the diffusion model driving the bias contents during generation. The complete process of bias attribution is outlined in Algorithm 1 at Appendix A. Notably, this step is required only once for one specific diffusion model." + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 325, + 701, + 553, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 325, + 701, + 553, + 713 + ], + "spans": [ + { + "bbox": [ + 325, + 701, + 553, + 713 + ], + "type": "text", + "content": "Additionally, we observe that the target can extend be" + } + ] + } + ], + "index": 18 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 315, + 757 + ], + "type": "text", + "content": "8195" + } + ] + } + ], + "index": 19 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 3 + }, + { + "para_blocks": [ + { + "bbox": [ + 55, + 72, + 294, + 168 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 72, + 294, + 168 + ], + "spans": [ + { + "bbox": [ + 55, + 72, + 294, + 168 + ], + "type": "text", + "content": "yond s to include the input " + }, + { + "bbox": [ + 55, + 72, + 294, + 168 + ], + "type": "inline_equation", + "content": "\\mathbf{x}" + }, + { + "bbox": [ + 55, + 72, + 294, + 168 + ], + "type": "text", + "content": " or any component within the target model itself. This broader approach enables us to examine how these elements impact the behavior of the underlying model concerning the measure " + }, + { + "bbox": [ + 55, + 72, + 294, + 168 + ], + "type": "inline_equation", + "content": "F_{\\mathbf{x}}" + }, + { + "bbox": [ + 55, + 72, + 294, + 168 + ], + "type": "text", + "content": ". To the best of our knowledge, we are the first to incorporate feature disentanglement and attribution methods to interpret the bias-related components within diffusion models, providing deeper insights into the internal workings of diffusion process." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 55, + 174, + 277, + 187 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 174, + 277, + 187 + ], + "spans": [ + { + "bbox": [ + 55, + 174, + 277, + 187 + ], + "type": "text", + "content": "4.3. Intervening in Bias Features for Debiasing" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 55, + 193, + 295, + 300 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 193, + 295, + 300 + ], + "spans": [ + { + "bbox": [ + 55, + 193, + 295, + 300 + ], + "type": "text", + "content": "To control and mitigate social bias in diffusion models, we propose an effective method via intervening in the above discovered bias features indexed with " + }, + { + "bbox": [ + 55, + 193, + 295, + 300 + ], + "type": "inline_equation", + "content": "i \\in \\mathbf{A}" + }, + { + "bbox": [ + 55, + 193, + 295, + 300 + ], + "type": "text", + "content": ", where " + }, + { + "bbox": [ + 55, + 193, + 295, + 300 + ], + "type": "inline_equation", + "content": "\\mathbf{A}" + }, + { + "bbox": [ + 55, + 193, + 295, + 300 + ], + "type": "text", + "content": " consists of indexes of the identified bias features. It is possible to suppress or amplify these features to exert control over the model's behavior [12]. Unlike other methods [32, 33, 43], we only intervene in the identified features indexed with " + }, + { + "bbox": [ + 55, + 193, + 295, + 300 + ], + "type": "inline_equation", + "content": "\\mathbf{A}" + }, + { + "bbox": [ + 55, + 193, + 295, + 300 + ], + "type": "text", + "content": " rather than the entire semantic space. Our intervention approach are illustrated in part (c) of Fig. 1." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 55, + 300, + 295, + 373 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 300, + 295, + 373 + ], + "spans": [ + { + "bbox": [ + 55, + 300, + 295, + 373 + ], + "type": "text", + "content": "Specifically, we apply an intervention operation to the target features in the sparse semantic space. Then these features are reconstructed into the original latent space for subsequent image generation and more details are in Appendix A. For each feature indexed within " + }, + { + "bbox": [ + 55, + 300, + 295, + 373 + ], + "type": "inline_equation", + "content": "\\mathbf{A}" + }, + { + "bbox": [ + 55, + 300, + 295, + 373 + ], + "type": "text", + "content": ", we can either apply a scalar multiplication or adding an additive bias as" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 55, + 378, + 295, + 419 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 378, + 295, + 419 + ], + "spans": [ + { + "bbox": [ + 55, + 378, + 295, + 419 + ], + "type": "interline_equation", + "content": "s _ {i} = \\operatorname {I n t e r v e n e} \\left(s _ {i}\\right) = \\left\\{ \\begin{array}{l l} \\beta s _ {i} & (\\text {S c a l i n g}), \\text {o r} \\\\ s _ {i} + \\beta & (\\text {A d d i n g}). \\forall i \\in \\mathbf {A}, \\beta \\in \\mathbb {R}. \\end{array} \\right. \\tag {7}", + "image_path": "7750c617946272c69ebc80c7785bdbad6ef064f9f52cb2de7fd889f0d7dc6e20.jpg" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 55, + 420, + 295, + 456 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 420, + 295, + 456 + ], + "spans": [ + { + "bbox": [ + 55, + 420, + 295, + 456 + ], + "type": "text", + "content": "By applying a distinct parameter " + }, + { + "bbox": [ + 55, + 420, + 295, + 456 + ], + "type": "inline_equation", + "content": "\\beta" + }, + { + "bbox": [ + 55, + 420, + 295, + 456 + ], + "type": "text", + "content": ", we can suppress or amplify the target bias feature to adjust the bias level or even mitigate it for improved balance." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 55, + 456, + 295, + 540 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 456, + 295, + 540 + ], + "spans": [ + { + "bbox": [ + 55, + 456, + 295, + 540 + ], + "type": "text", + "content": "Our DIFFLENS demonstrates that it can control over the generation and representation of biased features in the output images as shown in Sec. 5.4. Furthermore, in Sec. 5.6, we gain more valuable insights into how different bias features are embedded within the inner working of diffusion model, allowing us not only to debias effectively but also to better understand the model's decision-making processes." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 55, + 549, + 135, + 562 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 549, + 135, + 562 + ], + "spans": [ + { + "bbox": [ + 55, + 549, + 135, + 562 + ], + "type": "text", + "content": "5. Experiments" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 55, + 569, + 295, + 592 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 569, + 295, + 592 + ], + "spans": [ + { + "bbox": [ + 55, + 569, + 295, + 592 + ], + "type": "text", + "content": "In this section, we evaluate our DIFFLENS qualitatively and quantitatively, following the below research questions." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 55, + 593, + 295, + 628 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 593, + 295, + 628 + ], + "spans": [ + { + "bbox": [ + 55, + 593, + 295, + 628 + ], + "type": "text", + "content": "RQ1: How effectively does DIFFLENS mitigate social bias while maintaining image quality? RQ2: Can DIFFLENS accurately identify intrinsic mechanism of bias generation?" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 55, + 629, + 295, + 665 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 629, + 295, + 665 + ], + "spans": [ + { + "bbox": [ + 55, + 629, + 295, + 665 + ], + "type": "text", + "content": "RQ3: How well does it control the bias level through intervening in the found bias features? RQ4: Is it able to interpret bias feature in a fine-grained way?" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 55, + 671, + 182, + 684 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 671, + 182, + 684 + ], + "spans": [ + { + "bbox": [ + 55, + 671, + 182, + 684 + ], + "type": "text", + "content": "5.1. Experimental Settings" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 55, + 689, + 295, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 689, + 295, + 713 + ], + "spans": [ + { + "bbox": [ + 55, + 689, + 295, + 713 + ], + "type": "text", + "content": "Diffusion Model Architecture. We evaluate our proposed method on two pretrained diffusion models, (i) P2 [10]," + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 313, + 72, + 554, + 108 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 72, + 554, + 108 + ], + "spans": [ + { + "bbox": [ + 313, + 72, + 554, + 108 + ], + "type": "text", + "content": "an unconditional diffusion model trained on CelebA-HQ dataset [26]; (ii) Stable Diffusion [49], a text-conditional diffusion model trained on LAION dataset [52]." + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 313, + 109, + 555, + 289 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 109, + 555, + 289 + ], + "spans": [ + { + "bbox": [ + 313, + 109, + 555, + 289 + ], + "type": "text", + "content": "Baselines. We compare our method against several baselines that are categorized into interpretability and guidance-based methods. For interpretability methods, we construct Activation [6] trying to locate influential neurons according to the neuron activation values. For guidance-based methods, we include Latent Editing [32] and H-Distribution [43] (abbreviated as \"L.E.\" and \"H-Dist.\"), trying to learn a latent vector to guide the unbiased generation within the bottleneck layer of U-Net. We additionally compare two methods, specifically for text-conditional diffusion models, which are Latent Direction [33] and Finetuning [54]. Due to Activation's limited debiasing effectiveness across both diffusion models and its restricted control over bias levels, we report its results only in Tab. 1 and Fig. 2. More details about baselines are in Appendix C.3." + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 313, + 289, + 555, + 456 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 289, + 555, + 456 + ], + "spans": [ + { + "bbox": [ + 313, + 289, + 555, + 456 + ], + "type": "text", + "content": "Implementation Details. To ensure fair comparison, and without loss of generality, we constrain our method, DIFFLENS, and other benchmark methods to operate on the bottleneck layer of the U-Net [50] in the diffusion model, adhering to the baseline settings in [32, 43]. Notably, DIFFLENS can easily extend the targeted mechanism to broader regions within the model. Following Sec. 4, we train kSAEs on the target latent space in the diffusion model. For the training of the light-weight classifier, we follow the settings in [43]. For the support set of samples used for identifying bias feature, we leverage the diffusion models to generate 1,000 sample for each attribute. We provide more details of implementation and dataset we use in Appendix B. The reproducible codes are in supplementary material." + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 313, + 457, + 554, + 625 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 457, + 554, + 625 + ], + "spans": [ + { + "bbox": [ + 313, + 457, + 554, + 625 + ], + "type": "text", + "content": "Evaluation Metrics. To measure the bias level of generated content, we use the Fairness Discrepancy (FD) introduced in [43] which calculates the Euclidean distance between a reference distribution (typically uniform) and the bias distribution of generation, representing class distributions within an attribute (e.g., male and female for gender attribute). A lower FD score indicates more balanced outputs. We also measure image quality using Fréchet Inception Distance (FID) [22] on the generated images. In addition, we follow [54] to measure (i) CLIP-T, which calculates the CLIP [48] based semantic similarity between the generated image and the input text prompt, and (ii) CLIP-I, which assesses the similarity between originally generated images and images after debiasing. More details are in Appendix C." + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 313, + 646, + 466, + 658 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 646, + 466, + 658 + ], + "spans": [ + { + "bbox": [ + 313, + 646, + 466, + 658 + ], + "type": "text", + "content": "5.2. Debiasing Diffusion Models" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 313, + 665, + 554, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 665, + 554, + 713 + ], + "spans": [ + { + "bbox": [ + 313, + 665, + 554, + 713 + ], + "type": "text", + "content": "We define \"bias mitigation\" task as adjusting the diffusion model to produce balanced outputs across specific social attributes (e.g., gender). We focus on the task across gender, age, and race attributes. Gender includes male and female" + } + ] + } + ], + "index": 18 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 315, + 757 + ], + "type": "text", + "content": "8196" + } + ] + } + ], + "index": 19 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 4 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 58, + 70, + 153, + 167 + ], + "blocks": [ + { + "bbox": [ + 58, + 70, + 153, + 167 + ], + "lines": [ + { + "bbox": [ + 58, + 70, + 153, + 167 + ], + "spans": [ + { + "bbox": [ + 58, + 70, + 153, + 167 + ], + "type": "image", + "image_path": "ef162e8a748da7c069f87047ccb63ada1e4d13cb0c633ef6806d0ac3c8ad980b.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 74, + 167, + 135, + 182 + ], + "lines": [ + { + "bbox": [ + 74, + 167, + 135, + 182 + ], + "spans": [ + { + "bbox": [ + 74, + 167, + 135, + 182 + ], + "type": "text", + "content": "Original M:F(%) 33.9:66.1" + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 55, + 192, + 555, + 225 + ], + "lines": [ + { + "bbox": [ + 55, + 192, + 555, + 225 + ], + "spans": [ + { + "bbox": [ + 55, + 192, + 555, + 225 + ], + "type": "text", + "content": "Figure 2. Comparison of randomly sampled original and debiased images from various baseline methods. The minority group (male) is highlighted with red bounding boxes for easier viewing. \"M:F\" refers to the male-to-female ratio. Our DIFFLENS effectively mitigates bias while preserving high generation quality, whereas other methods either struggle to maintain balance or produce images with artifacts." + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "type": "image", + "bbox": [ + 157, + 70, + 253, + 167 + ], + "blocks": [ + { + "bbox": [ + 157, + 70, + 253, + 167 + ], + "lines": [ + { + "bbox": [ + 157, + 70, + 253, + 167 + ], + "spans": [ + { + "bbox": [ + 157, + 70, + 253, + 167 + ], + "type": "image", + "image_path": "8841639032a35a07dc5242762b1ace1bef02c29205d095bc9d7f72e9c11119b4.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 174, + 167, + 236, + 182 + ], + "lines": [ + { + "bbox": [ + 174, + 167, + 236, + 182 + ], + "spans": [ + { + "bbox": [ + 174, + 167, + 236, + 182 + ], + "type": "text", + "content": "Activation M:F(%) 36.6:63.4" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_caption" + } + ], + "index": 2 + }, + { + "type": "image", + "bbox": [ + 257, + 70, + 352, + 167 + ], + "blocks": [ + { + "bbox": [ + 257, + 70, + 352, + 167 + ], + "lines": [ + { + "bbox": [ + 257, + 70, + 352, + 167 + ], + "spans": [ + { + "bbox": [ + 257, + 70, + 352, + 167 + ], + "type": "image", + "image_path": "f7812aa5cffae3918b5bebd952688e68798759e0dcacbcff820fb0006f5e8a34.jpg" + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 274, + 167, + 335, + 182 + ], + "lines": [ + { + "bbox": [ + 274, + 167, + 335, + 182 + ], + "spans": [ + { + "bbox": [ + 274, + 167, + 335, + 182 + ], + "type": "text", + "content": "Latent Editing \nM:F(%) 50.2:49.8" + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "image_caption" + } + ], + "index": 4 + }, + { + "type": "image", + "bbox": [ + 357, + 70, + 452, + 167 + ], + "blocks": [ + { + "bbox": [ + 357, + 70, + 452, + 167 + ], + "lines": [ + { + "bbox": [ + 357, + 70, + 452, + 167 + ], + "spans": [ + { + "bbox": [ + 357, + 70, + 452, + 167 + ], + "type": "image", + "image_path": "2aee04b24988774fc2260a7f3108ff5f4d5ea4f261257f12495e76b7965a1e9c.jpg" + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 373, + 167, + 435, + 182 + ], + "lines": [ + { + "bbox": [ + 373, + 167, + 435, + 182 + ], + "spans": [ + { + "bbox": [ + 373, + 167, + 435, + 182 + ], + "type": "text", + "content": "H-Distribution M:F(%) 46.6:53.4" + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "image_caption" + } + ], + "index": 6 + }, + { + "type": "image", + "bbox": [ + 457, + 70, + 552, + 167 + ], + "blocks": [ + { + "bbox": [ + 457, + 70, + 552, + 167 + ], + "lines": [ + { + "bbox": [ + 457, + 70, + 552, + 167 + ], + "spans": [ + { + "bbox": [ + 457, + 70, + 552, + 167 + ], + "type": "image", + "image_path": "c549bbed29f7d5628145f7046c788246fd39d0cb60077e44eec72a9f3199aed6.jpg" + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 474, + 167, + 535, + 182 + ], + "lines": [ + { + "bbox": [ + 474, + 167, + 535, + 182 + ], + "spans": [ + { + "bbox": [ + 474, + 167, + 535, + 182 + ], + "type": "text", + "content": "DiffLens (Ours) M:F(%) 49.9:50.1" + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "image_caption" + } + ], + "index": 8 + }, + { + "type": "table", + "bbox": [ + 82, + 236, + 528, + 316 + ], + "blocks": [ + { + "bbox": [ + 82, + 236, + 528, + 316 + ], + "lines": [ + { + "bbox": [ + 82, + 236, + 528, + 316 + ], + "spans": [ + { + "bbox": [ + 82, + 236, + 528, + 316 + ], + "type": "table", + "html": "
CategoryMethodGender (2)Age (3)Race (4)
FD ↓FID ↓CLIP-I ↑FD ↓FID ↓CLIP-I ↑FD ↓FID ↓CLIP-I ↑
Original0.22633.38-0.59233.38-0.71833.38-
Guidance-basedLatent Editing [32]0.00329.850.94740.60633.710.90920.31734.610.9081
H-Distribution [43]0.04831.310.94400.51134.210.85940.49436.190.8762
InterpretabilityActivation [6]0.19034.270.80600.54448.910.77930.70046.130.7846
DIFFLENS (Ours)0.00231.930.94790.40131.710.94140.44733.470.9111
", + "image_path": "21a90c0ea43f1b6f1ae6b78ab4b63e47b99a74bbffdfde984f9d79280b72201f.jpg" + } + ] + } + ], + "index": 11, + "angle": 0, + "type": "table_body" + } + ], + "index": 11 + }, + { + "bbox": [ + 55, + 323, + 555, + 346 + ], + "lines": [ + { + "bbox": [ + 55, + 323, + 555, + 346 + ], + "spans": [ + { + "bbox": [ + 55, + 323, + 555, + 346 + ], + "type": "text", + "content": "Table 1. Evaluation of bias mitigation in unconditional diffusion model P2 [10]. The numbers beside each attribute indicate the number of classes for that attribute. We highlight the best results in bold and the second best with underlined format (excluding the \"original\")." + } + ] + } + ], + "index": 12, + "angle": 0, + "type": "text" + }, + { + "bbox": [ + 55, + 357, + 295, + 380 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 357, + 295, + 380 + ], + "spans": [ + { + "bbox": [ + 55, + 357, + 295, + 380 + ], + "type": "text", + "content": "classes; age includes young, adult, and old; and race includes white, black, Asian, and Indian classes." + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 54, + 384, + 295, + 515 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 54, + 384, + 295, + 515 + ], + "spans": [ + { + "bbox": [ + 54, + 384, + 295, + 515 + ], + "type": "text", + "content": "Intuitively, bias mitigation should yield balanced generation without compromising the overall semantics of the generated content. To comprehensively evaluate the effectiveness of bias mitigation (RQ1) across different methods, we assess (i) the bias level using the FD score of generated images, (ii) the generation quality via the FID score, and (iii) the preservation of overall image semantics after bias mitigation using CLIP-I/T scores. Beyond evaluating on the unconditional generative model, we also extend our analysis to a text-conditional one to showcase the broader applicability and generalizability of our method." + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 55, + 519, + 295, + 662 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 519, + 295, + 662 + ], + "spans": [ + { + "bbox": [ + 55, + 519, + 295, + 662 + ], + "type": "text", + "content": "Qualitative Results on Bias Mitigation. We present batches of generated examples before and after debiasing gender attribute to visually evaluate bias level and generation quality, as shown in Fig. 2. Our DIFFLENS effectively neutralizes gender bias while ensuring the generated images remain realistic and intact. In comparison, Latent Editing may generate unrealistic images, for example, the second one in the first row in Fig. 2, H-Distribution struggles to generate a well-balanced output, and Activation produces imbalanced, distorted images. We also provide visual results for text-to-image diffusion model Stable Diffusion [49] in Appendix D.2." + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 55, + 665, + 295, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 665, + 295, + 714 + ], + "spans": [ + { + "bbox": [ + 55, + 665, + 295, + 714 + ], + "type": "text", + "content": "Quantitative Analysis on Bias Mitigation. We assess baseline methods on debiasing across various social attributes (e.g., gender) by generating balanced outputs. As shown in Tab. 1, our DIFFLENS demonstrates competitively" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 313, + 357, + 555, + 464 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 357, + 555, + 464 + ], + "spans": [ + { + "bbox": [ + 313, + 357, + 555, + 464 + ], + "type": "text", + "content": "strong results in (i) bias mitigation (measured by FD), (ii) in preserving semantic feature coherence (measured by CLIP-I) and (iii) maintaining generation quality (measured by FID). Activation shows poor ability to mitigate social bias. Although Latent Editing performs well in mitigating gender bias and achieves favorable FID, it can produce images with artifacts, as shown in Figs. 2 and 3. H-Distribution struggles to generate well-balanced outputs, as reflected by its poor FD scores." + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 313, + 466, + 556, + 611 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 466, + 556, + 611 + ], + "spans": [ + { + "bbox": [ + 313, + 466, + 556, + 611 + ], + "type": "text", + "content": "In Tab. 2, we show results on a text-to-image model, Stable Diffusion [49]. Our DIFFLENS effectively mitigates bias, especially for gender and age, while effectively maintaining generation quality and image semantics among methods that are effective in bias mitigation. Latent Direction achieves good debiasing for age and race but often overcorrects, reducing CLIP-I scores and compromising the balance between target generation ratio and image quality. Finetuning, demanding extensive effort, excels in gender and race but struggles with age bias mitigation. Other baselines either handle multi-class attributes poorly or lack generalizability to text-to-image diffusion models." + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 313, + 621, + 537, + 633 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 621, + 537, + 633 + ], + "spans": [ + { + "bbox": [ + 313, + 621, + 537, + 633 + ], + "type": "text", + "content": "5.3. Accurate Identification of Bias Mechanism" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 313, + 641, + 555, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 641, + 555, + 715 + ], + "spans": [ + { + "bbox": [ + 313, + 641, + 555, + 715 + ], + "type": "text", + "content": "Accurate bias mechanism discovery could provide feasible control on the generated contents of diffusion, regarding the level of bias attributes. For instance, unlike the original diffusion model P2 [10], which generates a male-to-female gender ratio of approximately 3:7, our method easily attains the reverse ratio of 7:3. This capability goes beyond merely" + } + ] + } + ], + "index": 20 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "text", + "content": "8197" + } + ] + } + ], + "index": 21 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 5 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 57, + 69, + 553, + 153 + ], + "blocks": [ + { + "bbox": [ + 57, + 69, + 553, + 153 + ], + "lines": [ + { + "bbox": [ + 57, + 69, + 553, + 153 + ], + "spans": [ + { + "bbox": [ + 57, + 69, + 553, + 153 + ], + "type": "table", + "html": "
MethodGender (2)Age (3)Race (4)
FD ↓FID ↓CLIP-I ↑CLIP-T ↑FD ↓FID ↓CLIP-I ↑CLIP-T ↑FD ↓FID ↓CLIP-I ↑CLIP-T ↑
Original0.564120.06-0.61550.752120.06-0.61550.558120.06-0.6155
Latent Editing [32]0.408166.110.82530.60050.682200.900.85270.61220.524153.050.88040.6086
H-Distribution [43]0.222151.680.84750.60870.506147.710.83450.60980.544126.900.82550.6100
Latent Direction [33]0.305129.370.80580.60910.052113.810.81510.60670.175128.300.82110.6132
Fintuning [54]0.050161.470.87790.60950.746161.470.87790.60950.198161.470.87790.6095
DIFFLENS (Ours)0.046112.830.85010.60900.04999.170.87780.60570.401119.860.90960.6149
", + "image_path": "3242417ed5cdcaa408d7209c10620022d6ff6f13604af755afca26739dcc6a52.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "table_body" + } + ], + "index": 0 + }, + { + "type": "table", + "bbox": [ + 57, + 188, + 553, + 244 + ], + "blocks": [ + { + "bbox": [ + 55, + 161, + 555, + 184 + ], + "lines": [ + { + "bbox": [ + 55, + 161, + 555, + 184 + ], + "spans": [ + { + "bbox": [ + 55, + 161, + 555, + 184 + ], + "type": "text", + "content": "Table 2. Evaluation of bias mitigation in text-to-image diffusion model Stable Diffusion [49], based on average performance across four occupations (see Appendix B). We highlight the best results in bold and the second-best with underlined text (excluding the \"original\")." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 57, + 188, + 553, + 244 + ], + "lines": [ + { + "bbox": [ + 57, + 188, + 553, + 244 + ], + "spans": [ + { + "bbox": [ + 57, + 188, + 553, + 244 + ], + "type": "table", + "html": "
OriginalL.E.H-Dist.DiffLensOriginalL.E.H-Dist.DiffLensOriginalL.E.H-Dist.DiffLens
", + "image_path": "564746958e8aace1a0fe66c9c721ebd8e0ffabafa964aa24f1fd3c1c5055dc2d.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "table_body" + } + ], + "index": 2 + }, + { + "bbox": [ + 54, + 285, + 296, + 489 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 54, + 285, + 296, + 489 + ], + "spans": [ + { + "bbox": [ + 54, + 285, + 296, + 489 + ], + "type": "text", + "content": "balancing generated images in Sec. 5.2; it allows us to examine how effectively each method edits the target attribute, e.g., gender, while preserving non-target semantic features, e.g., eyeglass and smile (RQ2). For illustration, we visualize the generated contents of compared methods upon the reversed ratio of gender attribute in Fig. 3. More examples and details are provided in Appendix D.4.1. As is shown in Fig. 3, our method can preserve non-target attributes such as smile, background coherence, and even eyeglasses when debiasing, showcasing accurate identification of the bias generation mechanism by our method. While the baseline methods often introduce distortions or compromise detail in these areas, without looking into inner working of bias generation. Additionally, from Tabs. 1 and 2, our DIFFLENS consistently achieves high CLIP-I scores, demonstrating its ability to accurately identify and modify target bias specific features with minimal impact on unrelated content." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 55, + 502, + 287, + 515 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 502, + 287, + 515 + ], + "spans": [ + { + "bbox": [ + 55, + 502, + 287, + 515 + ], + "type": "text", + "content": "5.4. Fine-grained Bias Level Control and Editing" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 54, + 522, + 295, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 54, + 522, + 295, + 713 + ], + "spans": [ + { + "bbox": [ + 54, + 522, + 295, + 713 + ], + "type": "text", + "content": "Benefiting from bias mechanism identification, we can control bias level with finer granularity and across a broader range (RQ3). Figure 5 illustrates how different methods progressively modulate image generation in diffusion along two opposite gender directions. More examples are in Appendix D.4.2. Our DIFFLENS produces natural transitions, consistently preserving semantic features such as facial expression and maintaining high generation quality. Examples about control in multi-attributes are in Appendix D.4.3. The results of varying gender ratios generated by each method are shown in Fig. 4. Latent Editing and H-Distribution perform well when the gender ratio is around balanced, but their fidelity and visual coherence deteriorate at more skewed ratios, indicating a limitation in handling highly imbalanced outputs. By adjusting the value of intervention parameter incorporated in Eq. (7), our DIFFLENS can flexi" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 313, + 285, + 555, + 357 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 285, + 555, + 357 + ], + "spans": [ + { + "bbox": [ + 313, + 285, + 555, + 357 + ], + "type": "text", + "content": "bly control bias feature and generate different gender ratios, ranging from balanced to imbalanced ones (almost male or female), illustrating powerful controllability over the representations of bias. Such highly imbalanced ratios also have wide applications in real world such as data augmentation for under-represented subgroups [43]." + } + ] + } + ], + "index": 7 + }, + { + "type": "image", + "bbox": [ + 315, + 366, + 553, + 447 + ], + "blocks": [ + { + "bbox": [ + 54, + 251, + 555, + 275 + ], + "lines": [ + { + "bbox": [ + 54, + 251, + 555, + 275 + ], + "spans": [ + { + "bbox": [ + 54, + 251, + 555, + 275 + ], + "type": "text", + "content": "Figure 3. Comparison in accurate identification of bias features. Our DIFFLENS preserves overall image semantics such as smile and eyeglasses while other methods frequently introduce distortions or lose important details." + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 315, + 366, + 553, + 447 + ], + "lines": [ + { + "bbox": [ + 315, + 366, + 553, + 447 + ], + "spans": [ + { + "bbox": [ + 315, + 366, + 553, + 447 + ], + "type": "image", + "image_path": "07fe781ce4742c76f11b0d5ebf3a3260fdff06c5d88d23725b097dd14937480f.jpg" + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 313, + 454, + 555, + 510 + ], + "lines": [ + { + "bbox": [ + 313, + 454, + 555, + 510 + ], + "spans": [ + { + "bbox": [ + 313, + 454, + 555, + 510 + ], + "type": "text", + "content": "Figure 4. Achievable range of bias level control for each method. The Log Gender Ratio reflects the log of male to female ratio in the generated images, with 0 indicating balance. Our DIFFLENS offers broader bias control, delivering higher generation quality (left plot) and improved semantic coherence (right plot)." + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "image_caption" + } + ], + "index": 8 + }, + { + "bbox": [ + 313, + 515, + 409, + 529 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 515, + 409, + 529 + ], + "spans": [ + { + "bbox": [ + 313, + 515, + 409, + 529 + ], + "type": "text", + "content": "5.5. Ablation Study" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 312, + 534, + 555, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 312, + 534, + 555, + 715 + ], + "spans": [ + { + "bbox": [ + 312, + 534, + 555, + 715 + ], + "type": "text", + "content": "We conduct an ablation study to reveal how each component in our DIFFLENS takes effect. Recall that, it dissects the diffusion models by (i) disentangling the latent space to a sparse semantic space via k-SAE and (ii) identifying the bias features using bias attribution. In Tab. 3, directly selecting neurons with the highest activation values as bias features (Activation method) performs poorly, supporting the claim in Sec. 4.1. Comparing \"Act@L\" with \"Act@S\" and \"Attr@L\" with DIFFLENS, using k-SAE improves visual coherence, as seen in the higher CLIP-I scores (0.8597 to 0.9479 and 0.8060 to 0.9368), while avoiding direct edits on polysemantic neurons mitigates FID degradation (41.68 to 31.93). Furthermore, comparing \"Act@L\" with \"Attr@L\" and \"Act@\" with DIFFLENS, bias attribution enables more precise localization of bias features, significantly lowering" + } + ] + } + ], + "index": 11 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "text", + "content": "8198" + } + ] + } + ], + "index": 12 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 6 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 56, + 69, + 286, + 182 + ], + "blocks": [ + { + "bbox": [ + 56, + 69, + 286, + 182 + ], + "lines": [ + { + "bbox": [ + 56, + 69, + 286, + 182 + ], + "spans": [ + { + "bbox": [ + 56, + 69, + 286, + 182 + ], + "type": "image", + "image_path": "1549d9f196ea442434fa0c59b828050caaf9c0e983f0c1b3783a33831c269a97.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 55, + 192, + 555, + 226 + ], + "lines": [ + { + "bbox": [ + 55, + 192, + 555, + 226 + ], + "spans": [ + { + "bbox": [ + 55, + 192, + 555, + 226 + ], + "type": "text", + "content": "Figure 5. Comparison of individual image transformations along the gender axis, with columns displaying images sampled from progressively shifting ratios as in Fig. 4. Our DIFFLENS achieves smooth and consistent transitions that preserve semantic features like facial expressions, while other methods show distortions and loss of detail at higher imbalance ratios." + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "type": "image", + "bbox": [ + 291, + 70, + 554, + 182 + ], + "blocks": [ + { + "bbox": [ + 291, + 70, + 554, + 182 + ], + "lines": [ + { + "bbox": [ + 291, + 70, + 554, + 182 + ], + "spans": [ + { + "bbox": [ + 291, + 70, + 554, + 182 + ], + "type": "image", + "image_path": "10016405e196c19f863a164b7f019b0a16fa3eb6995a03ac73672703899351ef.jpg" + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_body" + } + ], + "index": 1 + }, + { + "bbox": [ + 55, + 236, + 295, + 271 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 236, + 295, + 271 + ], + "spans": [ + { + "bbox": [ + 55, + 236, + 295, + 271 + ], + "type": "text", + "content": "FD (0.190 to 0.072 and 0.303 to 0.002). DIFFLENS incorporating both shows superiority in bias mitigation with concise localization and intervention of bias features." + } + ] + } + ], + "index": 3 + }, + { + "type": "table", + "bbox": [ + 57, + 280, + 293, + 355 + ], + "blocks": [ + { + "bbox": [ + 57, + 280, + 293, + 355 + ], + "lines": [ + { + "bbox": [ + 57, + 280, + 293, + 355 + ], + "spans": [ + { + "bbox": [ + 57, + 280, + 293, + 355 + ], + "type": "table", + "html": "
ComponentsOrig.Act@LAct@SAttr@LDIFFLENS
ActivationX✓✓XX
K-SAEXX✓X✓
Bias Attr.XXX✓✓
FD ↓0.2260.1900.3030.0720.002
FID ↓33.3834.2732.7041.6831.93
CLIP-I ↑-0.80600.93860.85970.9479
", + "image_path": "15ac8f29c2f21cf9d13b5a43971024c087bcdba9ac830793e08779e2c15e6b40.jpg" + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "table_body" + }, + { + "bbox": [ + 57, + 356, + 262, + 365 + ], + "lines": [ + { + "bbox": [ + 57, + 356, + 262, + 365 + ], + "spans": [ + { + "bbox": [ + 57, + 356, + 262, + 365 + ], + "type": "text", + "content": "- ①, ②: Bias attribution allows precise bias mitigation." + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "table_footnote" + }, + { + "bbox": [ + 57, + 365, + 294, + 388 + ], + "lines": [ + { + "bbox": [ + 57, + 365, + 294, + 388 + ], + "spans": [ + { + "bbox": [ + 57, + 365, + 294, + 388 + ], + "type": "text", + "content": "- ③, ④: K-SAE improves visual feature coherence and maintains generation quality, via disentangling original latent space." + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "table_footnote" + } + ], + "index": 4 + }, + { + "bbox": [ + 55, + 390, + 296, + 434 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 390, + 296, + 434 + ], + "spans": [ + { + "bbox": [ + 55, + 390, + 296, + 434 + ], + "type": "text", + "content": "Table 3. Ablation study of DIFFLENS in P2 [10] for gender attribute. We abbreviate Original as \"Orig.\", neuron activations as \"Act.\", and bias attributions as \"Attr\". \"@L\" denotes operations on the original latent space, and \"@S\" on sparse semantic space." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 55, + 443, + 256, + 455 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 443, + 256, + 455 + ], + "spans": [ + { + "bbox": [ + 55, + 443, + 256, + 455 + ], + "type": "text", + "content": "5.6. Wider Investigation: Bias Mechanism" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 54, + 462, + 295, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 54, + 462, + 295, + 714 + ], + "spans": [ + { + "bbox": [ + 54, + 462, + 295, + 714 + ], + "type": "text", + "content": "We further look into the bias mechanism, i.e., the identified bias features " + }, + { + "bbox": [ + 54, + 462, + 295, + 714 + ], + "type": "inline_equation", + "content": "\\{s_i|i\\in \\mathbf{A}\\}" + }, + { + "bbox": [ + 54, + 462, + 295, + 714 + ], + "type": "text", + "content": ", to reveal its granular aspects on social attribute generation and how these features distinguishably affect the biased contents, shedding some lights on the inner workings of diffusion models (RQ4). We observe from Fig. 6 that (i) multiple features may jointly influence a single social bias attribute, such as \"short hair\" and \"mustache\" contributing to gender perception; (ii) some features overlap across biases, for instance, \"side pose\" being associated with both age and gender biases as it often appears in images of older males; and (iii) poses like \"side pose\" and \"head down\" reflect in photography, often aligning with gender norms that men are depicted in dynamic or indirect poses, while women appear younger and more polished, reflecting social expectations [15, 58, 60]. These stereotypes mirror underlying social biases, impacting not only how individuals are portrayed but also how their images are interpreted. Our DIFFLENS serves as a lens that uncovers social biases and stereotypes embedded in generated images, providing insight into how attributes such as gender, race, and age are systematically portrayed." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 314, + 235, + 387, + 247 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 314, + 235, + 387, + 247 + ], + "spans": [ + { + "bbox": [ + 314, + 235, + 387, + 247 + ], + "type": "text", + "content": "6. Conclusion" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 313, + 255, + 555, + 472 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 255, + 555, + 472 + ], + "spans": [ + { + "bbox": [ + 313, + 255, + 555, + 472 + ], + "type": "text", + "content": "In this work, we introduced a novel approach DIFFLENS, for mitigating social biases in diffusion models by dissecting, analyzing and intervening in the internal mechanisms of the diffusion model. Leveraging k-SAE and gradient-based attribution method, we disentangle, identify, and control bias-related features with precision, allowing targeted bias mitigation while preserving non-target attributes. A wide range of evaluations across social attributes such as gender, age, and race validate effectiveness of DIFFLENS in bias control and content preservation. Our work offers an interpretable solution to bias mitigation in diffusion models, though challenges remain, including the precise identification of certain racial features and a quality decline in intensified edits. Future directions include expanding this approach to other bias types, incorporating additional fairness perspectives, and exploring application across diverse model architectures like Diffusion Transformers [46] for broader impact in ethical generative AI deployment." + } + ] + } + ], + "index": 12 + }, + { + "type": "image", + "bbox": [ + 332, + 481, + 535, + 681 + ], + "blocks": [ + { + "bbox": [ + 332, + 481, + 535, + 681 + ], + "lines": [ + { + "bbox": [ + 332, + 481, + 535, + 681 + ], + "spans": [ + { + "bbox": [ + 332, + 481, + 535, + 681 + ], + "type": "image", + "image_path": "7654e5edfa8f6ef7606eec14cae9d7684146b2dd906c714e3c874a3b7ab0caf1.jpg" + } + ] + } + ], + "index": 13, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 313, + 681, + 555, + 713 + ], + "lines": [ + { + "bbox": [ + 313, + 681, + 555, + 713 + ], + "spans": [ + { + "bbox": [ + 313, + 681, + 555, + 713 + ], + "type": "text", + "content": "Figure 6. Manually analyzing the discovered bias features in diffusion models. Each column represents one specific feature " + }, + { + "bbox": [ + 313, + 681, + 555, + 713 + ], + "type": "inline_equation", + "content": "s_i" + }, + { + "bbox": [ + 313, + 681, + 555, + 713 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 313, + 681, + 555, + 713 + ], + "type": "inline_equation", + "content": "i \\in \\mathbf{A}" + }, + { + "bbox": [ + 313, + 681, + 555, + 713 + ], + "type": "text", + "content": " and each row originates from the same raw image." + } + ] + } + ], + "index": 14, + "angle": 0, + "type": "image_caption" + } + ], + "index": 13 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "text", + "content": "8199" + } + ] + } + ], + "index": 15 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 7 + }, + { + "para_blocks": [ + { + "bbox": [ + 56, + 72, + 160, + 85 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 72, + 160, + 85 + ], + "spans": [ + { + "bbox": [ + 56, + 72, + 160, + 85 + ], + "type": "text", + "content": "7. Acknowledgment" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 55, + 90, + 296, + 168 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 90, + 296, + 168 + ], + "spans": [ + { + "bbox": [ + 55, + 90, + 296, + 168 + ], + "type": "text", + "content": "The research received partial support from National Natural Science Foundation of China (Grant No. 62406193). The authors also gratefully acknowledge further assistance provided by Shanghai Frontiers Science Center of Human-centered Artificial Intelligence, MoE Key Lab of Intelligent Perception and Human-Machine Collaboration, and HPC Platform of ShanghaiTech University." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 56, + 190, + 115, + 202 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 190, + 115, + 202 + ], + "spans": [ + { + "bbox": [ + 56, + 190, + 115, + 202 + ], + "type": "text", + "content": "References" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 61, + 209, + 296, + 712 + ], + "type": "list", + "angle": 0, + "index": 12, + "blocks": [ + { + "bbox": [ + 61, + 209, + 295, + 266 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 61, + 209, + 295, + 266 + ], + "spans": [ + { + "bbox": [ + 61, + 209, + 295, + 266 + ], + "type": "text", + "content": "[1] Hritik Bansal, Da Yin, Masoud Monajatipoor, and Kai-Wei Chang. How well can text-to-image generative models understand ethical natural language interventions? In Proceedings of the 2022 Conference on Empirical Methods in Natural Language Processing, pages 1358–1370, 2022. 2" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 62, + 267, + 296, + 320 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 267, + 296, + 320 + ], + "spans": [ + { + "bbox": [ + 62, + 267, + 296, + 320 + ], + "type": "text", + "content": "[2] Abhipsa Basu, R Venkatesh Babu, and Danish Pruthi. Inspecting the geographical representativeness of images from text-to-image models. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 5136-5147, 2023. 2" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 62, + 323, + 295, + 378 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 323, + 295, + 378 + ], + "spans": [ + { + "bbox": [ + 62, + 323, + 295, + 378 + ], + "type": "text", + "content": "[3] Sebastian Benthall and Bruce D. Haynes. Racial categories in machine learning. In Proceedings of the Conference on Fairness, Accountability, and Transparency, page 289-298, New York, NY, USA, 2019. Association for Computing Machinery. 11" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 62, + 379, + 295, + 422 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 379, + 295, + 422 + ], + "spans": [ + { + "bbox": [ + 62, + 379, + 295, + 422 + ], + "type": "text", + "content": "[4] Leonard Bereska and Stratis Gavves. Mechanistic interpretability for AI safety - a review. Transactions on Machine Learning Research, 2024. Survey Certification, Expert Certification. 1, 3, 11" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 62, + 423, + 296, + 499 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 423, + 296, + 499 + ], + "spans": [ + { + "bbox": [ + 62, + 423, + 296, + 499 + ], + "type": "text", + "content": "[5] Federico Bianchi, Pratyusha Kalluri, Esin Durmus, Faisal Ladhak, Myra Cheng, Debora Nozza, Tatsunori Hashimoto, Dan Jurafsky, James Zou, and Aylin Caliskan. Easily accessible text-to-image generation amplifies demographic stereotypes at large scale. In Proceedings of the 2023 ACM Conference on Fairness, Accountability, and Transparency, pages 1493-1504, 2023. 2" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 62, + 501, + 294, + 567 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 501, + 294, + 567 + ], + "spans": [ + { + "bbox": [ + 62, + 501, + 294, + 567 + ], + "type": "text", + "content": "[6] Steven Bills, Nick Cammarata, Dan Mossing, Henk Tillman, Leo Gao, Gabriel Goh, Ilya Sutskever, Jan Leike, Jeff Wu, and William Saunders. Language models can explain neurons in language models. https://openaipublic.blob.core.windows.net/neuron-explainer/paper/index.html, 2023.5, 6" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 62, + 568, + 295, + 612 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 568, + 295, + 612 + ], + "spans": [ + { + "bbox": [ + 62, + 568, + 295, + 612 + ], + "type": "text", + "content": "[7] Hila Chefer, Oran Lang, Mor Geva, Volodymyr Polosukhin, Assaf Shocher, Inbar Mosseri, Lior Wolf, et al. The hidden language of diffusion models. In *The Twelfth International Conference on Learning Representations*. 1" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 62, + 613, + 295, + 657 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 613, + 295, + 657 + ], + "spans": [ + { + "bbox": [ + 62, + 613, + 295, + 657 + ], + "type": "text", + "content": "[8] Hila Chefer, Yuval Alaluf, Yael Vinker, Lior Wolf, and Daniel Cohen-Or. Attend-and-excite: Attention-based semantic guidance for text-to-image diffusion models. ACM Transactions on Graphics (TOG), 42(4):1-10, 2023. 2" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 62, + 658, + 295, + 712 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 658, + 295, + 712 + ], + "spans": [ + { + "bbox": [ + 62, + 658, + 295, + 712 + ], + "type": "text", + "content": "[9] Jaemin Cho, Abhay Zala, and Mohit Bansal. Dall-eval: Probing the reasoning skills and social biases of text-to-image generation models. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 3043-3054, 2023. 2" + } + ] + } + ], + "index": 11 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 316, + 73, + 555, + 713 + ], + "type": "list", + "angle": 0, + "index": 25, + "blocks": [ + { + "bbox": [ + 316, + 73, + 555, + 128 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 73, + 555, + 128 + ], + "spans": [ + { + "bbox": [ + 316, + 73, + 555, + 128 + ], + "type": "text", + "content": "[10] Jooyoung Choi, Jungbeom Lee, Chaehun Shin, Sungwon Kim, Hyunwoo Kim, and Sungroh Yoon. Perception prioritized training of diffusion models. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 11472-11481, 2022. 5, 6, 8, 1, 2, 3, 4, 9" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 317, + 129, + 554, + 171 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 129, + 554, + 171 + ], + "spans": [ + { + "bbox": [ + 317, + 129, + 554, + 171 + ], + "type": "text", + "content": "[11] Ching-Yao Chuang, Varun Jampani, Yuanzhen Li, Antonio Torralba, and Stefanie Jegelka. Debiasing vision-language models via biased prompts. arXiv preprint arXiv:2302.00070, 2023. 2" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 317, + 173, + 554, + 239 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 173, + 554, + 239 + ], + "spans": [ + { + "bbox": [ + 317, + 173, + 554, + 239 + ], + "type": "text", + "content": "[12] Damai Dai, Li Dong, Yaru Hao, Zhifang Sui, Baobao Chang, and Furu Wei. Knowledge neurons in pretrained transformers. In Proceedings of the 60th Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers), pages 8493-8502, Dublin, Ireland, 2022. Association for Computational Linguistics. 3, 5" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 316, + 239, + 554, + 272 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 239, + 554, + 272 + ], + "spans": [ + { + "bbox": [ + 316, + 239, + 554, + 272 + ], + "type": "text", + "content": "[13] Kedar Dhamhere, Mukund Sundararajan, and Qiqi Yan. How important is a neuron. In International Conference on Learning Representations, 2019. 4" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 317, + 273, + 554, + 326 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 273, + 554, + 326 + ], + "spans": [ + { + "bbox": [ + 317, + 273, + 554, + 326 + ], + "type": "text", + "content": "[14] Nelson Elhage, Tristan Hume, Catherine Olsson, Nicholas Schiefer, Tom Henighan, Shauna Kravec, Zac Hatfield-Dodds, Robert Lasenby, Dawn Drain, Carol Chen, et al. Toy models of superposition. arXiv preprint arXiv:2209.10652, 2022.3" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 316, + 327, + 553, + 360 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 327, + 553, + 360 + ], + "spans": [ + { + "bbox": [ + 316, + 327, + 553, + 360 + ], + "type": "text", + "content": "[15] Emilio Ferrara. Fairness and bias in artificial intelligence: A brief survey of sources, impacts, and mitigation strategies. Sci, 6(1), 2024. 8" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 317, + 361, + 554, + 415 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 361, + 554, + 415 + ], + "spans": [ + { + "bbox": [ + 317, + 361, + 554, + 415 + ], + "type": "text", + "content": "[16] Felix Friedrich, Manuel Brack, Lukas Struppek, Dominik Hintersdorf, Patrick Schramowski, Sasha Luccioni, and Kristian Kersting. Fair diffusion: Instructing text-to-image generation models on fairness. arXiv preprint arXiv:2302.10893, 2023. 2" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 317, + 416, + 554, + 469 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 416, + 554, + 469 + ], + "spans": [ + { + "bbox": [ + 317, + 416, + 554, + 469 + ], + "type": "text", + "content": "[17] Rohit Gandikota, Hadas Orgad, Yonatan Belinkov, Joanna Materzyńska, and David Bau. Unified concept editing in diffusion models. In Proceedings of the IEEE/CVF Winter Conference on Applications of Computer Vision, pages 5111-5120, 2024. 1, 2" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 317, + 472, + 554, + 536 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 472, + 554, + 536 + ], + "spans": [ + { + "bbox": [ + 317, + 472, + 554, + 536 + ], + "type": "text", + "content": "[18] René Haas, Inbar Huberman-Spiegelglas, Rotem Mulayoff, Stella Graßhof, Sami S Brandt, and Tomer Michaeli. Discovering interpretable directions in the semantic latent space of diffusion models. In 2024 IEEE 18th International Conference on Automatic Face and Gesture Recognition (FG), pages 1-9. IEEE, 2024. 2, 4" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 317, + 537, + 554, + 592 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 537, + 554, + 592 + ], + "spans": [ + { + "bbox": [ + 317, + 537, + 554, + 592 + ], + "type": "text", + "content": "[19] Ligong Han, Yinxiao Li, Han Zhang, Peyman Milanfar, Dimitris Metaxas, and Feng Yang. Svdiff: Compact parameter space for diffusion fine-tuning. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 7323-7334, 2023. 2" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 317, + 593, + 554, + 669 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 593, + 554, + 669 + ], + "spans": [ + { + "bbox": [ + 317, + 593, + 554, + 669 + ], + "type": "text", + "content": "[20] Ruifei He, Chuhui Xue, Haoru Tan, Wenqing Zhang, Yingchen Yu, Song Bai, and Xiaojuan Qi. Debiasing text-to-image diffusion models. In Proceedings of the 1st ACM Multimedia Workshop on Multi-Modal Misinformation Governance in the Era of Foundation Models, page 29–36, New York, NY, USA, 2024. Association for Computing Machinery. 2" + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 316, + 670, + 553, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 670, + 553, + 713 + ], + "spans": [ + { + "bbox": [ + 316, + 670, + 553, + 713 + ], + "type": "text", + "content": "[21] Amir Hertz, Ron Mokady, Jay Tenenbaum, Kfir Aberman, Yael Pritch, and Daniel Cohen-or. Prompt-to-prompt image editing with cross-attention control. In The Eleventh International Conference on Learning Representations, 2023. 2" + } + ] + } + ], + "index": 24 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "text", + "content": "8200" + } + ] + } + ], + "index": 26 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 8 + }, + { + "para_blocks": [ + { + "bbox": [ + 56, + 73, + 294, + 713 + ], + "type": "list", + "angle": 0, + "index": 13, + "blocks": [ + { + "bbox": [ + 56, + 73, + 294, + 127 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 73, + 294, + 127 + ], + "spans": [ + { + "bbox": [ + 56, + 73, + 294, + 127 + ], + "type": "text", + "content": "[22] Martin Heusel, Hubert Ramsauer, Thomas Unterthiner, Bernhard Nessler, and Sepp Hochreiter. Gans trained by a two time-scale update rule converge to a local nash equilibrium. Advances in neural information processing systems, 30, 2017. 5, 11" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 56, + 129, + 294, + 173 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 129, + 294, + 173 + ], + "spans": [ + { + "bbox": [ + 56, + 129, + 294, + 173 + ], + "type": "text", + "content": "[23] Jonathan Ho, Ajay Jain, and Pieter Abbeel. Denoising diffusion probabilistic models. In Advances in Neural Information Processing Systems, pages 6840-6851. Curran Associates, Inc., 2020. 1, 2" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 56, + 175, + 294, + 229 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 175, + 294, + 229 + ], + "spans": [ + { + "bbox": [ + 56, + 175, + 294, + 229 + ], + "type": "text", + "content": "[24] Robert Huben, Hoagy Cunningham, Logan Riggs Smith, Aidan Ewart, and Lee Sharkey. Sparse autoencoders find highly interpretable features in language models. In The Twelfth International Conference on Learning Representations, 2024. 3, 4, 1" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 56, + 231, + 294, + 285 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 231, + 294, + 285 + ], + "spans": [ + { + "bbox": [ + 56, + 231, + 294, + 285 + ], + "type": "text", + "content": "[25] Kimmo Karkkainen and Jungseock Joo. Fairface: Face attribute dataset for balanced race, gender, and age for bias measurement and mitigation. In Proceedings of the IEEE/CVF winter conference on applications of computer vision, pages 1548-1558, 2021. 2, 4" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 56, + 286, + 294, + 330 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 286, + 294, + 330 + ], + "spans": [ + { + "bbox": [ + 56, + 286, + 294, + 330 + ], + "type": "text", + "content": "[26] Tero Karras, Timo Aila, Samuli Laine, and Jaakko Lehtinen. Progressive growing of GANs for improved quality, stability, and variation. In International Conference on Learning Representations, 2018. 5" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 56, + 332, + 294, + 376 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 332, + 294, + 376 + ], + "spans": [ + { + "bbox": [ + 56, + 332, + 294, + 376 + ], + "type": "text", + "content": "[27] Tero Karras, Samuli Laine, and Timo Aila. A style-based generator architecture for generative adversarial networks. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 4401-4410, 2019. 5" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 56, + 377, + 294, + 420 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 377, + 294, + 420 + ], + "spans": [ + { + "bbox": [ + 56, + 377, + 294, + 420 + ], + "type": "text", + "content": "[28] Eunji Kim, Siwon Kim, Chaehun Shin, and Sungroh Yoon. De-stereotyping text-to-image models through prompt tuning. In ICML Workshop on Challenges in Deployable Generative AI, 2023. 2" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 56, + 422, + 294, + 475 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 422, + 294, + 475 + ], + "spans": [ + { + "bbox": [ + 56, + 422, + 294, + 475 + ], + "type": "text", + "content": "[29] Gwanghyun Kim, Taesung Kwon, and Jong Chul Ye. Diffus ionclip: Text-guided diffusion models for robust image man i plication. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 2426-2435, 2022. 1" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 56, + 478, + 294, + 532 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 478, + 294, + 532 + ], + "spans": [ + { + "bbox": [ + 56, + 478, + 294, + 532 + ], + "type": "text", + "content": "[30] Yeongmin Kim, Byeonghu Na, Minsang Park, JoonHo Jang, Dongjun Kim, Wanmo Kang, and Il chul Moon. Training unbiased diffusion models from biased dataset. In The Twelfth International Conference on Learning Representations, 2024. 1" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 56, + 534, + 294, + 578 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 534, + 294, + 578 + ], + "spans": [ + { + "bbox": [ + 56, + 534, + 294, + 578 + ], + "type": "text", + "content": "[31] Xianghao Kong, Ollie Liu, Han Li, Dani Yogatama, and Greg Ver Steeg. Interpretable diffusion via information decomposition. In The Twelfth International Conference on Learning Representations. 1" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 56, + 579, + 294, + 623 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 579, + 294, + 623 + ], + "spans": [ + { + "bbox": [ + 56, + 579, + 294, + 623 + ], + "type": "text", + "content": "[32] Mingi Kwon, Jaeseok Jeong, and Youngjung Uh. Diffusion models already have a semantic latent space. In The Eleventh International Conference on Learning Representations, 2023. 2, 3, 4, 5, 6, 7, 9" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 56, + 624, + 294, + 689 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 624, + 294, + 689 + ], + "spans": [ + { + "bbox": [ + 56, + 624, + 294, + 689 + ], + "type": "text", + "content": "[33] Hang Li, Chengzhi Shen, Philip Torr, Volker Tresp, and Jindong Gu. Self-discovering interpretable diffusion latent directions for responsible text-to-image generation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 12006-12016, 2024. 2, 4, 5, 7" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 56, + 691, + 294, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 691, + 294, + 713 + ], + "spans": [ + { + "bbox": [ + 56, + 691, + 294, + 713 + ], + "type": "text", + "content": "[34] Sasha Luccioni, Christopher Akiki, Margaret Mitchell, and Yacine Jernite. Stable bias: Evaluating societal representa" + } + ] + } + ], + "index": 12 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 316, + 73, + 553, + 712 + ], + "type": "list", + "angle": 0, + "index": 28, + "blocks": [ + { + "bbox": [ + 333, + 73, + 553, + 95 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 333, + 73, + 553, + 95 + ], + "spans": [ + { + "bbox": [ + 333, + 73, + 553, + 95 + ], + "type": "text", + "content": "tions in diffusion models. Advances in Neural Information Processing Systems, 36, 2024. 1, 2, 11" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 316, + 96, + 553, + 128 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 96, + 553, + 128 + ], + "spans": [ + { + "bbox": [ + 316, + 96, + 553, + 128 + ], + "type": "text", + "content": "[35] Alireza Makhzani and Brendan Frey. K-sparse autoencoders. In International Conference on Learning Representations, 2014. 2, 3, 1, 6, 11" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 316, + 129, + 553, + 183 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 129, + 553, + 183 + ], + "spans": [ + { + "bbox": [ + 316, + 129, + 553, + 183 + ], + "type": "text", + "content": "[36] Vongani H Maluleke, Neerja Thakkar, Tim Brooks, Ethan Weber, Trevor Darrell, Alexei A Efros, Angjoo Kanazawa, and Devin Guillery. Studying bias in gans through the lens of race. In European Conference on Computer Vision, pages 344-360. Springer, 2022. 1" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 316, + 185, + 553, + 249 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 185, + 553, + 249 + ], + "spans": [ + { + "bbox": [ + 316, + 185, + 553, + 249 + ], + "type": "text", + "content": "[37] Pietro Melzi, Christian Rathgeb, Ruben Tolosana, Ruben Vera-Rodriguez, Dominik Lawatsch, Florian Domin, and Maxim Schaubert. Gandifface: Controllable generation of synthetic datasets for face recognition with realistic variations. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 3086-3095, 2023. 1" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 316, + 251, + 553, + 293 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 251, + 553, + 293 + ], + "spans": [ + { + "bbox": [ + 316, + 251, + 553, + 293 + ], + "type": "text", + "content": "[38] Kevin Meng, David Bau, Alex Andonian, and Yonatan Belinkov. Locating and editing factual associations in gpt. Advances in Neural Information Processing Systems, 35: 17359-17372, 2022. 1, 3" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 316, + 294, + 553, + 336 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 294, + 553, + 336 + ], + "spans": [ + { + "bbox": [ + 316, + 294, + 553, + 336 + ], + "type": "text", + "content": "[39] Jesse Mu and Jacob Andreas. Compositional explanations of neurons. In Advances in Neural Information Processing Systems, pages 17153-17163. Curran Associates, Inc., 2020. 3" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 316, + 338, + 553, + 381 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 338, + 553, + 381 + ], + "spans": [ + { + "bbox": [ + 316, + 338, + 553, + 381 + ], + "type": "text", + "content": "[40] Ranjita Naik and Besmira Nushi. Social biases through the text-to-image generation lens. In Proceedings of the 2023 AAAI/ACM Conference on AI, Ethics, and Society, pages 786-808, 2023. 2, 4" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 316, + 383, + 553, + 415 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 383, + 553, + 415 + ], + "spans": [ + { + "bbox": [ + 316, + 383, + 553, + 415 + ], + "type": "text", + "content": "[41] Chris Olah, Nick Cammarata, Ludwig Schubert, Gabriel Goh, Michael Petrov, and Shan Carter. Zoom in: An introduction to circuits. Distill, 5(3):e00024-001, 2020. 3" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 316, + 416, + 553, + 459 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 416, + 553, + 459 + ], + "spans": [ + { + "bbox": [ + 316, + 416, + 553, + 459 + ], + "type": "text", + "content": "[42] Hadas Orgad, Bahjat Kawar, and Yonatan Belinkov. Editing implicit assumptions in text-to-image diffusion models. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 7053-7061, 2023. 1, 2" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 316, + 460, + 553, + 525 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 460, + 553, + 525 + ], + "spans": [ + { + "bbox": [ + 316, + 460, + 553, + 525 + ], + "type": "text", + "content": "[43] Rishubh Parihar, Abhijnya Bhat, Abhipsa Basu, Saswat Mallick, Jogendra Nath Kundu, and R Venkatesh Babu. Balancing act: Distribution-guided debiasing in diffusion models. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 6668-6678, 2024. 1, 2, 4, 5, 6, 7, 9, 11" + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 316, + 526, + 553, + 569 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 526, + 553, + 569 + ], + "spans": [ + { + "bbox": [ + 316, + 526, + 553, + 569 + ], + "type": "text", + "content": "[44] Ji-Hoon Park, Yeong-Joon Ju, and Seong-Whan Lee. Explaining generative diffusion models via visual analysis for interpretable decision-making process. Expert Systems with Applications, 248:123231, 2024. 1" + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 316, + 571, + 553, + 624 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 571, + 553, + 624 + ], + "spans": [ + { + "bbox": [ + 316, + 571, + 553, + 624 + ], + "type": "text", + "content": "[45] Yong-Hyun Park, Mingi Kwon, Jaewoong Choi, Junghyo Jo, and Youngjung Uh. Understanding the latent space of diffusion models through the lens of riemannian geometry. In Thirty-seventh Conference on Neural Information Processing Systems, 2023. 2" + } + ] + } + ], + "index": 25 + }, + { + "bbox": [ + 316, + 625, + 553, + 668 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 625, + 553, + 668 + ], + "spans": [ + { + "bbox": [ + 316, + 625, + 553, + 668 + ], + "type": "text", + "content": "[46] William Peebles and Saining Xie. Scalable diffusion models with transformers. In Proceedings of the IEEE/CVF International Conference on Computer Vision (ICCV), pages 4195-4205, 2023. 8" + } + ] + } + ], + "index": 26 + }, + { + "bbox": [ + 316, + 670, + 553, + 712 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 670, + 553, + 712 + ], + "spans": [ + { + "bbox": [ + 316, + 670, + 553, + 712 + ], + "type": "text", + "content": "[47] Malsha V Perera and Vishal M Patel. Analyzing bias in diffusion-based face generation models. In 2023 IEEE International Joint Conference on Biometrics (IJCB), pages 1-10. IEEE, 2023. 1, 2" + } + ] + } + ], + "index": 27 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 748, + 314, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 314, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 314, + 757 + ], + "type": "text", + "content": "8201" + } + ] + } + ], + "index": 29 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 9 + }, + { + "para_blocks": [ + { + "bbox": [ + 55, + 72, + 297, + 714 + ], + "type": "list", + "angle": 0, + "index": 12, + "blocks": [ + { + "bbox": [ + 56, + 72, + 297, + 139 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 72, + 297, + 139 + ], + "spans": [ + { + "bbox": [ + 56, + 72, + 297, + 139 + ], + "type": "text", + "content": "[48] Alec Radford, Jong Wook Kim, Chris Hallacy, Aditya Ramesh, Gabriel Goh, Sandhini Agarwal, Girish Sastry, Amanda Askell, Pamela Mishkin, Jack Clark, et al. Learning transferable visual models from natural language supervision. In International conference on machine learning, pages 8748-8763. PMLR, 2021. 5" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 55, + 140, + 297, + 195 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 140, + 297, + 195 + ], + "spans": [ + { + "bbox": [ + 55, + 140, + 297, + 195 + ], + "type": "text", + "content": "[49] Robin Rombach, Andreas Blattmann, Dominik Lorenz, Patrick Esser, and Björn Ommer. High-resolution image synthesis with latent diffusion models. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 10684-10695, 2022. 1, 2, 5, 6, 7, 3, 4" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 56, + 195, + 296, + 260 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 195, + 296, + 260 + ], + "spans": [ + { + "bbox": [ + 56, + 195, + 296, + 260 + ], + "type": "text", + "content": "[50] Olaf Ronneberger, Philipp Fischer, and Thomas Brox. U-net: Convolutional networks for biomedical image segmentation. In Medical image computing and computer-assisted intervention-MICCAI 2015: 18th international conference, Munich, Germany, October 5-9, 2015, proceedings, part III 18, pages 234-241. Springer, 2015. 3, 5, 1, 2, 4, 6" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 56, + 261, + 296, + 304 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 261, + 296, + 304 + ], + "spans": [ + { + "bbox": [ + 56, + 261, + 296, + 304 + ], + "type": "text", + "content": "[51] Harrison Rosenberg, Shimaa Ahmed, Guruprasad V Ramesh, Ramya Korlakai Vinayak, and Kassem Fawaz. Unbiased face synthesis with diffusion models: Are we there yet? arXiv preprint arXiv:2309.07277, 2023. 2" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 56, + 305, + 296, + 371 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 305, + 296, + 371 + ], + "spans": [ + { + "bbox": [ + 56, + 305, + 296, + 371 + ], + "type": "text", + "content": "[52] Christoph Schuhmann, Romain Beaumont, Richard Vencu, Cade Gordon, Ross Wightman, Mehdi Cherti, Theo Coombes, Aarush Katta, Clayton Mullis, Mitchell Wortsman, et al. Laion-5b: An open large-scale dataset for training next generation image-text models. Advances in Neural Information Processing Systems, 35:25278-25294, 2022. 5" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 56, + 372, + 296, + 437 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 372, + 296, + 437 + ], + "spans": [ + { + "bbox": [ + 56, + 372, + 296, + 437 + ], + "type": "text", + "content": "[53] Preethi Seshadri, Sameer Singh, and Yanai Elazar. The bias amplification paradox in text-to-image generation. In Proceedings of the 2024 Conference of the North American Chapter of the Association for Computational Linguistics: Human Language Technologies (Volume 1: Long Papers), pages 6367-6384, 2024. 2" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 56, + 437, + 296, + 491 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 437, + 296, + 491 + ], + "spans": [ + { + "bbox": [ + 56, + 437, + 296, + 491 + ], + "type": "text", + "content": "[54] Xudong Shen, Chao Du, Tianyu Pang, Min Lin, Yongkang Wong, and Mohan Kankanhalli. Finetuning text-to-image diffusion models for fairness. In The Twelfth International Conference on Learning Representations, 2024. 1, 2, 5, 7, 6, 11" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 56, + 492, + 296, + 525 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 492, + 296, + 525 + ], + "spans": [ + { + "bbox": [ + 56, + 492, + 296, + 525 + ], + "type": "text", + "content": "[55] Jiaming Song, Chenlin Meng, and Stefano Ermon. Denoising diffusion implicit models. In International Conference on Learning Representations, 2021. 1, 2, 3, 4" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 56, + 526, + 296, + 569 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 526, + 296, + 569 + ], + "spans": [ + { + "bbox": [ + 56, + 526, + 296, + 569 + ], + "type": "text", + "content": "[56] Mukund Sundararajan, Ankur Taly, and Qiqi Yan. Axiomatic attribution for deep networks. In Proceedings of the 34th International Conference on Machine Learning-Volume 70, pages 3319–3328, 2017. 4" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 56, + 570, + 296, + 647 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 570, + 296, + 647 + ], + "spans": [ + { + "bbox": [ + 56, + 570, + 296, + 647 + ], + "type": "text", + "content": "[57] Raphael Tang, Linqing Liu, Akshit Pandey, Zhiying Jiang, Gefei Yang, Karun Kumar, Pontus Stenetorp, Jimmy Lin, and Ferhan Ture. What the DAAM: Interpreting stable diffusion using cross attention. In Proceedings of the 61st Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers), pages 5644-5659, Toronto, Canada, 2023. Association for Computational Linguistics. 2" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 56, + 647, + 296, + 680 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 647, + 296, + 680 + ], + "spans": [ + { + "bbox": [ + 56, + 647, + 296, + 680 + ], + "type": "text", + "content": "[58] Mike Zajko. Conservative ai and social inequality: conceptualizing alternatives to bias through social theory. *Ai & Society*, 36(3):1047–1056, 2021. 8" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 56, + 681, + 296, + 714 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 681, + 296, + 714 + ], + "spans": [ + { + "bbox": [ + 56, + 681, + 296, + 714 + ], + "type": "text", + "content": "[59] Haiyan Zhao, Hanjie Chen, Fan Yang, Ninghao Liu, Huiqi Deng, Hengyi Cai, Shuaiqiang Wang, Dawei Yin, and Mengnan Du. Explainability for large language models: A survey." + } + ] + } + ], + "index": 11 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 316, + 73, + 555, + 129 + ], + "type": "list", + "angle": 0, + "index": 15, + "blocks": [ + { + "bbox": [ + 334, + 73, + 553, + 95 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 334, + 73, + 553, + 95 + ], + "spans": [ + { + "bbox": [ + 334, + 73, + 553, + 95 + ], + "type": "text", + "content": "ACM Transactions on Intelligent Systems and Technology, 15(2):1-38, 2024. 1" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 316, + 96, + 555, + 129 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 96, + 555, + 129 + ], + "spans": [ + { + "bbox": [ + 316, + 96, + 555, + 129 + ], + "type": "text", + "content": "[60] Mi Zhou, Vibhanshu Abhishek, Timothy Derdenger, Jaymo Kim, and Kannan Srinivasan. Bias in generative ai. arXiv preprint arXiv:2403.02726, 2024. 8" + } + ] + } + ], + "index": 14 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 748, + 316, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 316, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 316, + 757 + ], + "type": "text", + "content": "8202" + } + ] + } + ], + "index": 16 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 10 + } + ], + "_backend": "vlm", + "_version_name": "2.6.4" +} \ No newline at end of file diff --git a/2025/Distilled Prompt Learning for Incomplete Multimodal Survival Prediction/5aae416b-ca2c-40f8-8583-f2902bfc3349_content_list.json b/2025/Distilled Prompt Learning for Incomplete Multimodal Survival Prediction/5aae416b-ca2c-40f8-8583-f2902bfc3349_content_list.json new file mode 100644 index 0000000000000000000000000000000000000000..b9064e94075e894676938a7d2cb8a528cae987fe --- /dev/null +++ b/2025/Distilled Prompt Learning for Incomplete Multimodal Survival Prediction/5aae416b-ca2c-40f8-8583-f2902bfc3349_content_list.json @@ -0,0 +1,1445 @@ +[ + { + "type": "text", + "text": "Distilled Prompt Learning for Incomplete Multimodal Survival Prediction", + "text_level": 1, + "bbox": [ + 125, + 130, + 870, + 152 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Yingxue Xu $^{1}$ Fengtao Zhou $^{1}$ Chenyu Zhao $^{1}$ Yihui Wang $^{1}$ Can Yang $^{2}$ Hao Chen $^{1\\boxtimes}$ $^{1}$ Department of Computer Science and Engineering, $^{2}$ Department of Mathematics The Hong Kong University of Science and Technology", + "bbox": [ + 127, + 179, + 869, + 233 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "{yxueb,fzhouaf,cyzhao,ywangrm}@connect.ust.hk，{macyang,jhc}@ust.hk", + "bbox": [ + 200, + 236, + 790, + 252 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Abstract", + "text_level": 1, + "bbox": [ + 246, + 286, + 326, + 301 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "The integration of multimodal data including pathology images and gene profiles is widely applied in precise survival prediction. Despite recent advances in multimodal survival models, collecting complete modalities for multimodal fusion still poses a significant challenge, hindering their application in clinical settings. Current approaches tackling incomplete modalities often fall short, as they typically compensate for only a limited part of the knowledge of missing modalities. To address this issue, we propose a Distilled Prompt Learning framework (DisPro) to utilize the strong robustness of Large Language Models (LLMs) to missing modalities, which employs two-stage prompting for compensation of comprehensive information for missing modalities. In the first stage, Unimodal Prompt (UniPro) distills the knowledge distribution of each modality, preparing for supplementing modality-specific knowledge of the missing modality in the subsequent stage. In the second stage, Multimodal Prompt (MultiPro) leverages available modalities as prompts for LLMs to infer the missing modality, which provides modality-common information. Simultaneously, the unimodal knowledge acquired in the first stage is injected into multimodal inference to compensate for the modality-specific knowledge of the missing modality. Extensive experiments covering various missing scenarios demonstrated the superiority of the proposed method. The code is available at https://github.com/Innse/DisPro.", + "bbox": [ + 88, + 316, + 483, + 724 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "1. Introduction", + "text_level": 1, + "bbox": [ + 89, + 753, + 220, + 768 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Multimodal Survival Prediction typically incorporates pathology slides and genomic data complementary to each other for precise prognosis analysis. Pathology data can provide qualitative morphological insights, while quantitative molecular information can be obtained from genomic data [5, 35, 36]. Despite progress [5, 10, 36] in multimodal survival analysis, these models typically require complete modalities data for reliable performance. However, the ex", + "bbox": [ + 89, + 779, + 483, + 900 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/d84997a7d97f611945b7c7e55fdbc05e48e28a07581278184108b361e3588ef3.jpg", + "image_caption": [ + "Figure 1. Insights for existing incomplete multimodal learning and comparison to the proposed method. (a) Generation-based Imputation and Imputation-free approaches, (b) Retrieved-based Imputation and (c) Ours." + ], + "image_footnote": [], + "bbox": [ + 535, + 286, + 879, + 426 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "pectation of having full access to complete modalities for integration is not always achievable, especially in clinical scenarios, due to various factors such as the cost of data gathering and privacy issues. For example, acquiring genomic data still comes at a high cost due to the technology and infrastructure required for gene sequencing, especially in underdeveloped areas [43]. Inevitably, it is challenging to adapt previously established multimodal survival models [30, 42] to real-world clinical settings. As a result, building a multimodal survival model that is robust to incomplete modalities data is an urgent issue for clinical practice.", + "bbox": [ + 509, + 503, + 906, + 670 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Currently, mainstream approaches addressing incomplete modalities issues include imputation-based and imputation-free methods [8, 34]. The former approaches aim to impute the missing modality based on observed modalities, primarily achieved in two ways including generation [3, 23, 29] and retrieval [4, 40]. In terms of generation-based imputation, generative models are typically employed for synthesizing features or raw data of missing modalities from available modalities. As shown in Fig. 1 (a), these methods usually can only impute modality-common information, as the modality-specific information unique to the missing modality cannot be generated out of thin air. Moreover, the goal of retrieval-based imputation methods is to fill in missing modalities by seeking the most similar sample of the missing modality from the training set. This way tends", + "bbox": [ + 511, + 674, + 908, + 902 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "CVF", + "bbox": [ + 106, + 2, + 181, + 42 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore.", + "bbox": [ + 236, + 0, + 810, + 46 + ], + "page_idx": 0 + }, + { + "type": "page_number", + "text": "5102", + "bbox": [ + 482, + 944, + 514, + 955 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "to compensate only a limited part of modality-common and modality-specific information, as shown in Fig. 1 (b). The reason is that an individual sample exhibits considerable randomness and is challenging to fully capture the unique knowledge inherent in the missing modality.", + "bbox": [ + 89, + 90, + 480, + 167 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Unlike imputation-based approaches, imputation-free approaches [8, 34] primarily focus on how to minimize performance degradation resulting from missing modalities. These approaches make great efforts to learn robust multimodal representations that are less sensitive to missing modalities by capturing modality-invariant information. Recently, with the rise of Large Language Models (LLMs) [1, 20, 33], the researcher [17] attempts to leverage the strong robustness of LLMs benefit from extensive training data to alleviate performance decreases caused by incomplete modalities. They design some sophisticated prompts to inform LLMs of different distributions of inputs. However, they only capture modality-common information as well (Fig. 1a), ignoring distinct modality-specific knowledge of the missing modality. As such, existing methods still struggle to effectively incorporate the learning of both common and specific knowledge of missing modalities.", + "bbox": [ + 89, + 169, + 482, + 425 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Based on these insights, to bridge the aforementioned gap, we propose a Distilled Prompt Learning (DisPro) framework to unlock LLMs' robustness to incomplete modalities, which aims to simultaneously compensate for both specific and common information of missing modalities. It comprises two-stage prompt learning including Unimodal Prompting (UniPro) and Multimodal Prompting (MultiPro) for capturing modality-specific and modality-common information of missing modalities, respectively.", + "bbox": [ + 89, + 428, + 482, + 564 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "In the first stage, to prepare for supplementing modality-specific information of missing modalities, we first learn the knowledge distribution of the missing modality by distilling unimodal knowledge into a set of learnable unimodal prompts for each modality from all remaining samples available of this modality. Specifically, inspired by CoOp [44], a set of learnable prompts for different classes is employed to learn knowledge about how to describe different categories of an individual modality. For example, in survival prediction, learned prompts can be intuitively interpreted as how different risk groups will reflect in pathological morphology or gene expression profiles. However, due to the huge size (e.g., $100,000 \\times 100,000$ ) of pathological WSIs, we cannot simply apply CoOp designed for small natural images to WSIs. Therefore, in UniPro, we extend CoOp into the multiple instance learning (MIL) paradigm that is commonly used for large-scale WSIs [11, 13, 37].", + "bbox": [ + 89, + 566, + 482, + 823 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "In the second stage, to capture modality-common information for missing modalities, we leverage available modalities as multimodal prompts of LLMs to infer representations of missing modalities using LLMs' strong inference ability. Simultaneously, to inject the learned unimodal", + "bbox": [ + 89, + 825, + 482, + 900 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "knowledge in stage 1, we propose UniPro Distillation to enforce the inferred representations of missing modalities to get as close as possible to the learned unimodal prompts of the corresponding class. Additionally, to assist in learning modality-common knowledge from WSIs and genes with high information redundancy [30, 42], we propose UniPro Scoring to re-use UniPro as a grader to select discriminative and relevant tokens. As a result, both common and specific information are compensated for missing modalities under the collaboration of uni- and multimodal prompts, thus mitigating performance declines led by incomplete modalities.", + "bbox": [ + 511, + 90, + 903, + 257 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "It is worth noting that DisPro is proposed as a framework for addressing missing modality issues by leveraging LLMs' power based on prompts, capable of generalizing to a wider range of domains. The contributions of this work are summarized as follows:", + "bbox": [ + 511, + 258, + 903, + 333 + ], + "page_idx": 1 + }, + { + "type": "list", + "sub_type": "text", + "list_items": [ + "- We propose a Distilled Prompt Learning (DisPro) framework for incomplete multimodal survival prediction, where LLMs' robustness to incomplete modalities is explored under two-stage prompting.", + "- We propose two-stage prompt learning including Unimodal Prompting and Multimodal Prompting to compensate for both modality-specific and modality-common information of missing modalities, respectively.", + "- Extensive experiments on five benchmark survival prediction datasets are conducted, which demonstrates significantly superior to state-of-the-art approaches." + ], + "bbox": [ + 513, + 335, + 903, + 501 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "2. Related Work", + "text_level": 1, + "bbox": [ + 513, + 518, + 653, + 534 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "2.1. Incomplete Multimodal Learning", + "text_level": 1, + "bbox": [ + 511, + 545, + 805, + 561 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "As stated above, incomplete multimodal learning [43] can be categorized into two types: imputation-based [4, 12, 38] and imputation-free methods [21, 25]. For example, with the success of the diffusion model [7], the research [23] attempted to apply it to MRI synthesis for missing modalities. Instead of generating raw modality data, SMIL [22] leveraged Bayesian meta-learning to impute features of missing modalities. For retrieval-based approaches, M3Care [40] filled in the missing modalities by retrieving the most similar samples from the training set. On the other hand, imputation-free approaches provided more flexible solutions, which aim to enhance the robustness of fusion and alleviate performance decrease. [8] proposed a segmentation-based regularizer to reduce the sensitivity to missing modalities. MUSE [34] utilized patients' relationships based on graph networks to perform contrastive learning for improving the robustness to missing modalities. GTP-4o [19] proposed modality-prompted heterogeneous graph to embed each modality into a unified space and complete missing modality through a graph prompting mechanism. Recently, MAP [17] proposed missing-aware prompts to unlock the robustness of LLMs to missing", + "bbox": [ + 511, + 568, + 903, + 902 + ], + "page_idx": 1 + }, + { + "type": "page_number", + "text": "5103", + "bbox": [ + 482, + 944, + 514, + 955 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "modalities by indicating LLMs different missing scenarios. However, these methods typically supplemented modality-common knowledge for missing modalities.", + "bbox": [ + 89, + 90, + 482, + 137 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "2.2. Multimodal Survival Prediction", + "text_level": 1, + "bbox": [ + 89, + 147, + 372, + 162 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "With the success of digital pathology [18, 24, 41] and sequencing techniques [26, 31], pathology and genomic data are commonly integrated to analyze survival outcomes, which have proven the superiority on precision in previous works [5, 36]. MCAT [5] proposed to leverage the interactions across modalities to guide the learning of pathology data. Based on this work, MOTCat [36] further leveraged optimal transport to capture global structure consistency across modalities. In parallel, CMTA [10] investigated the intrinsic cross-modal interactions and potential complementary information by cross-modal translation and alignment. From the perspective of information bottleneck, PIBD [42] proposed a prototypical information bottleneck for the MIL setting to select discriminative instances and disentangle multimodal knowledge for redundancy reduction. Similarly, MMP [30] utilized Gaussian Mixture Model to summarize morphological prototypes and biological pathway prototypes for these two modalities. However, all these methods assume all modalities available, inevitably leading to the difficulty of employment in clinical settings.", + "bbox": [ + 89, + 169, + 483, + 473 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3. Method", + "text_level": 1, + "bbox": [ + 89, + 487, + 181, + 502 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "This work aims to explore LLMs' robustness to incomplete modalities by two-stage prompting, Unimodal Prompting (UniPro) and Multimodal Prompting (MultiPro), which capture modality-specific and modality-common information, respectively. The overview framework of the proposed method is shown in Fig. 2. First, we will introduce the problem formulation in Sec. 3.1. Then, we will detail the process of UniPro and MultiPro in Sec. 3.2 and 3.3, respectively.", + "bbox": [ + 89, + 512, + 483, + 633 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3.1. Problem Formulation", + "text_level": 1, + "bbox": [ + 89, + 645, + 294, + 659 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Given a multimodal dataset consisting of $M = 2$ modalities, we use a 2-tuple $\\mathbf{X}_n = (\\mathbf{X}_n^p,\\mathbf{X}_n^g)$ to represent the $i$ -th patient data, where $\\mathbf{X}_n^p$ and $\\mathbf{X}_n^g$ are pathology and genomic data, respectively. In particular, we use the notation of tilde hat to indicate the modality is missing, e.g., $\\widetilde{\\mathbf{X}}_n^p$ .", + "bbox": [ + 89, + 667, + 482, + 744 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Following the conventional setting for pathological WSIs and genomic pathways in previous works [14, 42], we formulate a WSI $\\mathbf{X}_n^p$ and a set of genomic pathways $\\mathbf{X}_n^g$ as the \"bag\" of instances in MIL paradigm. We denote $\\mathbf{X}_n^p = \\{\\mathbf{x}_{n,i}^p\\}_{i=1}^{M_p}$ where each instance $\\mathbf{x}_{n,i}^p$ refers to a patch image cropped from the WSI $\\mathbf{X}_n^p$ and $M_p$ is the patch number of this WSI. Similarly, the bag of genomic pathways can be formulated as $\\mathbf{X}_n^g = \\{\\mathbf{x}_{n,i}^g\\}_{i=1}^{M_g}$ , in which an instance $\\mathbf{x}_{n,i}^g$ represents a pathway consisting of various gene expression values.", + "bbox": [ + 89, + 746, + 483, + 901 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "In survival prediction, the goal is to estimate the risk probability of an outcome event taking place before a specific time. However, this event may not always be observed, resulting in a right-censored occurrence. Therefore, $c_{n} \\in \\{0,1\\}$ is used to denote censorship status indicating whether the outcome event is right-censored (i.e., alive) ( $c = 1$ ) or not ( $c = 0$ ) (i.e., dead), and $t_{n} \\in \\mathbb{R}^{+}$ represents the overall survival time (in months). These two components make up its set of labels $Y_{n} = (c_{n},t_{n})$ . Following the previous setting [5, 36], the continuous survival time would be discretized into $I_{t}$ time intervals corresponding to $I_{t}$ risk bands. As a result, the label set $(c_{n},t_{n})$ can be re-written as $y_{n}$ , resulting in $2I_{t}$ labels.", + "bbox": [ + 511, + 90, + 903, + 287 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Given a set of predicted hazard probabilities $\\mathbf{h}_n$ across all time intervals, the hazard probability $\\mathbf{h}_n^{(j)}$ represents the probability of death event occurring at the $j$ -th time interval. Then, the cumulative survival probability for each time interval can be estimated as follows:", + "bbox": [ + 511, + 287, + 903, + 362 + ], + "page_idx": 2 + }, + { + "type": "equation", + "text": "\n$$\nS _ {n} ^ {(j)} = \\prod_ {z = 1} ^ {j} \\left(1 - \\mathbf {h} _ {n} ^ {(z)}\\right) \\tag {1}\n$$\n", + "text_format": "latex", + "bbox": [ + 635, + 373, + 903, + 414 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Following previous works [36, 42], given the ground truth time interval $\\tau$ , we use the NLL loss [39] for survival prediction, where $N$ is the sample number of the training set:", + "bbox": [ + 511, + 417, + 903, + 464 + ], + "page_idx": 2 + }, + { + "type": "equation", + "text": "\n$$\n\\begin{array}{l} \\mathcal {L} _ {s u r v} = - \\sum_ {n = 1} ^ {N} \\left[ c _ {n} \\log S _ {n} ^ {(\\tau)} + \\left(1 - c _ {n}\\right) \\log S _ {n} ^ {(\\tau - 1)} \\right. \\tag {2} \\\\ \\left. + \\left(1 - c _ {n}\\right) \\log {\\bf h} _ {n} ^ {(\\tau)} \\right] \\\\ \\end{array}\n$$\n", + "text_format": "latex", + "bbox": [ + 532, + 474, + 903, + 539 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3.2. Stage 1: Unimodal Prompting (UniPro)", + "text_level": 1, + "bbox": [ + 511, + 550, + 852, + 566 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Given a WSI $\\mathbf{X}_n^p = \\{\\mathbf{x}_{n,i}^p\\}_{i=1}^{M_p}$ or a bag of genomic pathways $\\mathbf{X}_n^g = \\{\\mathbf{x}_{n,i}^g\\}_{i=1}^{M_g}$ , we formulate their learning process as a MIL task. Inspired by CoOp [44] employing a CLIP-style architecture, we adapt it to the MIL setting to learn how to characterize each modality using a set of learnable prompts for modality-specific knowledge.", + "bbox": [ + 511, + 571, + 903, + 669 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "In terms of pathological WSIs data, following the common setting of the WSI processing [28], as shown in the UniPro-P of Fig 2, we first employ a pretrained and frozen patch encoder $g_{E}(\\cdot)$ to embed every patch images $\\pmb{x}_{n,i}^{p}$ from a WSI into an instance-level embedding $p_n^i = g_E(\\pmb{x}_{n,i}^p) \\in \\mathbb{R}^d$ , where $d$ is the patch feature dimension. To align pathology features with the text space, we apply an adapter $f_{p}(\\cdot)$ for each patch feature to get the aligned patch feature $\\mathbf{p}_n^i = f_p(g_E(\\pmb{x}_{n,i}^p)) \\in \\mathbb{R}^D$ , implemented by a linear layer followed by ReLU [2].", + "bbox": [ + 511, + 670, + 905, + 824 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Text-wise, the input comprises three components: pathology context tokens, prefix tokens andclassname tokens. Forclassname tokens $[C]_{1\\dots s}$ , we map every time interval of overall survival into a unique description, such as \"median risk\", and denote the censorship status by \"alive\"", + "bbox": [ + 511, + 825, + 903, + 901 + ], + "page_idx": 2 + }, + { + "type": "page_number", + "text": "5104", + "bbox": [ + 482, + 944, + 514, + 955 + ], + "page_idx": 2 + }, + { + "type": "image", + "img_path": "images/c9d0d3110b8395a40f7190f7be8aaf2b8a09e9428d281820174d23735f364bc1.jpg", + "image_caption": [ + "Stage 1: Unimodal Prompting (UniPro)" + ], + "image_footnote": [], + "bbox": [ + 114, + 111, + 480, + 281 + ], + "page_idx": 3 + }, + { + "type": "image", + "img_path": "images/42bfdeca25ae2ec5fdb16a3b3239c882f7290e04e8e0ef29aad92b99bd4fe3fb.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 496, + 111, + 870, + 277 + ], + "page_idx": 3 + }, + { + "type": "image", + "img_path": "images/789b7be1dbdc1a4a23d8857ae7e979c760ba2645578cf76a799657f73a7189ef.jpg", + "image_caption": [ + "Stage 2: Multimodal Prompting (MultiPro)" + ], + "image_footnote": [ + "d.1. $\\mathcal{L}_{\\mathrm{surv}}^{\\mathrm{cls}}$" + ], + "bbox": [ + 116, + 282, + 434, + 515 + ], + "page_idx": 3 + }, + { + "type": "image", + "img_path": "images/04683aa816c6afdce9c8c5cd6cc98c9abc1b9f6f554670ef5d86a633349677a8.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 444, + 282, + 875, + 431 + ], + "page_idx": 3 + }, + { + "type": "image", + "img_path": "images/f6fa08541a803969da1d94c9a70c4f5638350c2c1c413bf4593afc11db48c9e2.jpg", + "image_caption": [ + "Figure 2. Overview of DisPro. Stage 1: Unimodal Prompting aims to distill the knowledge distribution for each modality and prepare to supplement modality-specific knowledge for the missing modality. Stage 2: Multimodal Prompting utilizes the available modality as prompts to infer the representations of the missing one, compensating for modality-common knowledge. Simultaneously, the learned UniPro of Stage 1 supervises the learning of imputed representations to compensate for modality-specific knowledge. UniPro Scoring re-uses learned prompts to assist the LLM in capturing modality-common knowledge by selecting discriminative and relevant tokens." + ], + "image_footnote": [], + "bbox": [ + 444, + 435, + 679, + 545 + ], + "page_idx": 3 + }, + { + "type": "image", + "img_path": "images/5b5c49cdbcaf79c879a2444571ccf23c50a90f1fdc1c2d8725afb85c71c6fad1.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 679, + 452, + 877, + 540 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "or \"dead\". The prefix tokens $\\left[\\mathrm{PREFIX}\\right]_{1\\dots m}$ briefly describe the modality, e.g., \"This is a pathology slide image from the patient with overall survival of $[C]_{1\\dots s}$ \". Prefix and classname tokens are obtained using a frozen Tokenizer pretrained on a large-scale biomedical corpus to embed these words. To learn how to describe the modality for a specific risk band (class), we attach a set of learnable context tokens $[P]_{1\\dots k}$ to prefix and classname tokens:", + "bbox": [ + 88, + 642, + 482, + 763 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\boldsymbol {T} _ {p} = [ P ] _ {1} \\dots [ P ] _ {k} [ \\text {P R E F I X} ] _ {1} \\dots [ \\text {P R E F I X} ] _ {m} [ C ] _ {1} \\dots [ C ] _ {s} \\tag {3}\n$$\n", + "text_format": "latex", + "bbox": [ + 94, + 768, + 482, + 786 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "where $k$ , $m$ and $s$ are the token numbers of pathology context tokens, prefix words and classname words, respectively. Each token is a vector with the same dimension $D$ as word embeddings of the LLM (i.e., 768 for BERT). In this way, we can design input tokens for each risk band:", + "bbox": [ + 89, + 787, + 482, + 863 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\boldsymbol {T} _ {p} ^ {(j)} = [ P ] _ {1} ^ {(j)} \\dots [ P ] _ {k} ^ {(j)} [ \\text {P R E F I X} ] _ {1} \\dots [ \\text {P R E F I X} ] _ {m} [ C ] _ {1} ^ {(j)} \\dots [ C ] _ {s} ^ {(j)} \\tag {4}\n$$\n", + "text_format": "latex", + "bbox": [ + 89, + 869, + 491, + 901 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "in which $j \\in [1, 2I_t]$ and prefix tokens are shared across all risk bands. By forwarding $T_p^{(j)}$ of each class to the text encoder $\\mathcal{T}(\\cdot)$ (i.e., LLM), we can obtain a global textual class representation $t_p^{(j)} \\in \\mathbb{R}^D$ for every class.", + "bbox": [ + 511, + 642, + 906, + 708 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "In the MIL setting, to obtain the predicted probability for each class, we can use topK patches to produce the slide-level prediction based on similarity scores. To this end, we first compute the similarity score between every instance and every class:", + "bbox": [ + 511, + 714, + 906, + 790 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\boldsymbol {s} _ {n, p} ^ {(i, j)} = \\boldsymbol {p} _ {n} ^ {i} \\cdot \\boldsymbol {t} _ {p} ^ {(j)} \\tag {5}\n$$\n", + "text_format": "latex", + "bbox": [ + 651, + 809, + 906, + 830 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "As a result, we can get a similarity matrix $S_{n,p} \\in \\mathbb{R}^{M_p \\times 2I_t}$ for a WSI. Then, we apply a TopK max-pooling operator $h(\\cdot)$ to obtain the slide-level predictions, depend", + "bbox": [ + 511, + 854, + 906, + 902 + ], + "page_idx": 3 + }, + { + "type": "page_number", + "text": "5105", + "bbox": [ + 482, + 944, + 514, + 955 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "ing on top-K similarities for each class:", + "bbox": [ + 89, + 90, + 354, + 106 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\nh \\left(\\mathcal {S} _ {n, p}\\right) = \\frac {1}{K} \\left[ \\sum_ {i = 1} ^ {K} \\hat {\\boldsymbol {s}} _ {n, p} ^ {(i, 1)}, \\sum_ {i = 1} ^ {K} \\hat {\\boldsymbol {s}} _ {n, p} ^ {(i, 2)}, \\dots , \\sum_ {i = 1} ^ {K} \\hat {\\boldsymbol {s}} _ {n, p} ^ {(i, 2 \\boldsymbol {I} _ {t})} \\right] \\tag {6}\n$$\n", + "text_format": "latex", + "bbox": [ + 106, + 112, + 480, + 166 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "where $\\hat{s}_{n,p}$ is the sorted $s_{n,p}$ in a descending order across all instances. Next, we can get a set of hazard probabilities $\\mathbf{h}_n^p$ across all time intervals through a sigmoid function.", + "bbox": [ + 89, + 167, + 482, + 213 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\mathbf {h} _ {n} ^ {p} = \\operatorname {S i g m o i d} \\left(h \\left(S _ {n, p}\\right)\\right) \\tag {7}\n$$\n", + "text_format": "latex", + "bbox": [ + 202, + 220, + 482, + 239 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "By computing the survival loss according to Eq. 2 and backforwarding gradients, we can optimize the modality adapter and learnable prompts consisting of context tokens.", + "bbox": [ + 89, + 244, + 482, + 290 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Similarly, we acquire the set of hazard probabilities $\\mathbf{h}_n^g$ for genomic pathways and optimize its learnable prompts $T_{g}$ and gene encoder $f_{g}(\\cdot)$ . Here we use the subscript $p$ and $g$ to represent pathology and genomic data. Every pathway is also encoded into a $D$ -dimensional embedding.", + "bbox": [ + 89, + 290, + 483, + 367 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "3.3. Stage 2: Multimodal Prompting (MultiPro)", + "text_level": 1, + "bbox": [ + 89, + 373, + 460, + 390 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "In this stage, we aim to leverage LLMs' strong inference ability and robustness to missing modalities to predict the representation of missing modalities, as shown in the MultiPro of Fig. 2.", + "bbox": [ + 89, + 395, + 483, + 455 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "To simplify explanations, we take the example input of pathology available and genomics being missing. By prompting the LLM with available pathology data, the Bert-style LLM can capture modality-common knowledge for genomics due to its nature of self-attention. To compensate for modality-specific knowledge, we re-use the unimodal prompts $T_{g}$ learned in stage 1 with all remaining samples available of genomics data to supervise the learning of imputed representations, where unimodal prompts characterize the knowledge distribution of an individual modality.", + "bbox": [ + 89, + 455, + 483, + 606 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "To this end, we first need to construct an interleaved input for the learning of modality-common information. To fully leverage the capabilities of LLMs, every modality are supposed to be embedded into the aligned text space. Consequently, we re-use the aligned modality adapter or encoder learned in the first stage to embed each modality, and we continue to finetune them in the second stage. Note that the LLM used in stage 2 is the same as the one in stage 1. As a result, we can get a bag of aligned feature tokens $\\{\\mathbf{p}_n^i\\}_{i = 1}^{M_p}$ or $\\{\\mathbf{g}_n^i\\}_{i = 1}^{M_g}$ for pathology and genomic data, respectively. Here we assume genomic data $\\{\\tilde{\\mathbf{g}}_n^i\\}_{i = 1}^{M_g}$ is missing.", + "bbox": [ + 89, + 607, + 483, + 780 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Due to the limitation of input length for general LLMs (e.g., 512 for Bert-style LLMs) and high information redundancy in WSIs or genomic pathways [30, 42], more discriminative and relevant feature tokens are expected to be selected for better capturing modality-common knowledge. Therefore, we propose a UniPro Scoring module by re-using UniPro as a grader to estimate a relevance score for each feature token.", + "bbox": [ + 89, + 780, + 483, + 900 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "UniPro Scoring. The score of every feature token is determined by three parts: relevance scores estimated by UniPro-P and UniPro-G, and self-scoring, as shown in Fig. 2.", + "bbox": [ + 511, + 90, + 903, + 136 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Given a set of pathology feature tokens $\\{\\mathbf{p}_n^i\\}_{i = 1}^{M_p}$ , to obtain discriminative tokens, we compute the similarity score $s_{n,p}^{(i,\\tau)}$ between the ground-truth textual class representation of $t_p^{(\\tau)}$ and each feature token $\\mathbf{p}_n^i$ by re-using Eq. 5. Through the sigmoid function, we can get the score $\\mathbf{s}_{n,p}^{(i,\\tau)}$ estimated by UniPro-P. To capture cross-modal modality-common information for better compensation of missing modality, we similarly obtain the score $\\mathbf{s}_{n,g}^{(i,\\tau)}$ estimated by UniPro-G. Since textual class representations of UniPro are fixed once finishing the training of stage 1, the selected tokens will also become fixed accordingly. To make it diverse while dynamically adapting to the current input, we employ a learnable self-scoring mechanism implemented by attention layers [13] to assign a score $\\mathbf{a}_{n,p}^{(i)}$ for each feature token. In the end, the final score $\\mathbf{s}_{n,\\#}^{(i)}$ for each feature token can be computed by their summarization:", + "bbox": [ + 511, + 137, + 906, + 393 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\mathbf {s} _ {n, \\#} ^ {(i)} = \\mathbf {s} _ {n, p} ^ {(i, \\tau)} + \\mathbf {s} _ {n, g} ^ {(i, \\tau)} + \\mathbf {a} _ {n, p} ^ {(i)} \\tag {8}\n$$\n", + "text_format": "latex", + "bbox": [ + 612, + 402, + 903, + 424 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "For the inference phase when the ground-truth class cannot be accessed, we compute similarity scores with all textual class representations and apply Top-K max-pooling to determine which class $j$ these tokens belong to:", + "bbox": [ + 511, + 425, + 905, + 486 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\tau = \\arg \\max _ {j \\in [ 1, 2 I _ {t} ]} h \\left(\\mathcal {S} _ {n, p} ^ {(j)} \\oplus \\mathcal {S} _ {n, g} ^ {(j)}\\right) \\tag {9}\n$$\n", + "text_format": "latex", + "bbox": [ + 599, + 494, + 903, + 520 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "where $\\oplus$ refers to the element-wise addition and $h(\\cdot)$ can be computed by Eq. 6.", + "bbox": [ + 511, + 523, + 903, + 553 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "By selecting feature tokens with the top- $K_{p}$ largest scores, we can get the input tokens $\\{\\mathbf{p}_n^i\\}_{i=1}^{K_p}$ for available pathology modality. Then, we employ a set of learnable tokens to act as placeholders $[H_p]_{1\\dots K_p}$ and $[H_g]_{1\\dots K_g}$ for pathology and genomic data, where $K_{p}$ and $K_{g}$ are their token number of the input for the LLM in stage 2, respectively. They are initialized with modality indicators. For the available pathology, we add the selected tokens to its placeholders, leading to the input of pathology $\\{\\hat{\\mathbf{p}}_n^i\\}_{i=1}^{K_p}$ . At last, we concatenate them with placeholders of the missing genomics as well as a [CLS] token, resulting in the final input being forwarded into a LLM $\\mathcal{T}(\\cdot)$ :", + "bbox": [ + 511, + 553, + 905, + 739 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n[ C L S ] [ \\hat {\\mathbf {p}} _ {n} ] _ {1 \\dots K _ {p}} [ \\tilde {\\mathbf {g}} _ {n} ] _ {1 \\dots K _ {g}} = \\mathcal {T} ([ \\mathrm {C L S} ] | | \\{\\hat {\\mathbf {p}} _ {n} ^ {i} \\} _ {i = 1} ^ {K _ {p}} | | [ H _ {g} ] _ {1 \\dots K _ {g}}) \\tag {10}\n$$\n", + "text_format": "latex", + "bbox": [ + 521, + 747, + 903, + 777 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "UniPro Distillation. To inject modality-specific knowledge for the compensation of missing modality, we utilize UniPro to supervise this process, where UniPro has learned the knowledge distribution regarding the missing modality based on other observed samples in stage 1.", + "bbox": [ + 511, + 777, + 905, + 852 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Given the output (Eq. 10) of the LLM in stage 2, we take the part of the missing modality $[\\tilde{\\mathbf{g}}_n]_{1\\dots K_g}$ and compute similarity scores $\\tilde{\\mathbf{h}}_g$ between $[\\tilde{\\mathbf{g}}_n]_{1\\dots K_g}$ and the textual class", + "bbox": [ + 511, + 853, + 903, + 902 + ], + "page_idx": 4 + }, + { + "type": "page_number", + "text": "5106", + "bbox": [ + 482, + 944, + 514, + 955 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "representation by Eq. 5 - 7. By substituting $[\\tilde{\\mathbf{g}}_n]_{1\\ldots K_g}$ into the survival loss function of Eq. 2, we can supervise the learning of compensation for the missing modality:", + "bbox": [ + 89, + 90, + 485, + 137 + ], + "page_idx": 5 + }, + { + "type": "equation", + "text": "\n$$\n\\begin{array}{l} \\mathcal {L} _ {u d} ^ {g} = - \\sum_ {n = 1} ^ {\\tilde {N} _ {g}} \\left[ c _ {n} \\log S _ {n} ^ {(\\tau)} + \\left(1 - c _ {n}\\right) \\log S _ {n} ^ {(\\tau - 1)} \\right. \\tag {11} \\\\ \\left. + \\left(1 - c _ {n}\\right) \\log \\tilde {\\mathbf {h}} _ {n} ^ {g, (\\tau)} \\right] \\\\ \\end{array}\n$$\n", + "text_format": "latex", + "bbox": [ + 112, + 148, + 482, + 215 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "where $\\tilde{N}_g$ is the number of missing genomic modalities across all data pairs. Similarly, we can get the loss function $\\mathcal{L}_{ud}^p$ when the pathology modality is missing. Lastly, [CLS] token is used for calculating the loss function for survival prediction $\\mathcal{L}_{surv}^{cls}$ in stage 2, and the final loss can be combined linearly with coefficient factors $\\alpha_{1}$ and $\\alpha_{2}$ :", + "bbox": [ + 89, + 217, + 483, + 309 + ], + "page_idx": 5 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {L} = \\mathcal {L} _ {s u r v} ^ {c l s} + \\alpha_ {1} \\mathcal {L} _ {u d} ^ {p} + \\alpha_ {2} \\mathcal {L} _ {u d} ^ {g} \\tag {12}\n$$\n", + "text_format": "latex", + "bbox": [ + 183, + 319, + 482, + 339 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "4. Experiments", + "text_level": 1, + "bbox": [ + 89, + 349, + 223, + 366 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "4.1. Datasets and Settings", + "text_level": 1, + "bbox": [ + 89, + 375, + 290, + 390 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Datasets. To demonstrate the efficacy of the proposed method, we conducted a series of experiments across five public cancer datasets sourced from The Cancer Genome Atlas (TCGA) $^{1}$ . The case number of each dataset is demonstrated next to the dataset name in Tab. 1. Regarding genomic data, we follow the grouping of biological pathways provided by [14] and remove genes absent in the cBioPortal $^{2}$ database, resulting in 330 pathways. More details of datasets can be found in supplementary materials.", + "bbox": [ + 89, + 396, + 483, + 531 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Evaluation Settings on Incomplete Modalities. Following the common setting in previous works [22, 32], to systematically assess performances on various missing settings, we first manually construct a series of missing scenarios under different missing ratios by randomly dropping them for training sets on samples with complete modality pairs. Specifically, we set a total missing rate (60%) for two modalities and then construct 5 combinations for pathology and genomics, i.e., $\\{(0\\%, 60\\%), (20\\%, 40\\%), (30\\%, 30\\%), (40\\%, 20\\%), (60\\%, 0\\%)\\}$ . For inference, all methods are evaluated on three scenarios: pathology only, genomics only, and complete modalities. All experiments are conducted on 5-fold cross-validation following previous works [5, 36] in survival prediction, and the concordance index (C-Index) and its standard deviation (std) are reported to evaluate performances. The LLM is implemented by BioBERT-v1.2 [16]. The supplementary materials contain more experiment results and details on the implementation.", + "bbox": [ + 89, + 532, + 483, + 804 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "4.2. Comparisons with SOTAs", + "text_level": 1, + "bbox": [ + 89, + 813, + 326, + 829 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "We first present unimodal approaches, and SOTA multimodal methods trained with complete modalities (all", + "bbox": [ + 89, + 835, + 482, + 864 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "pathology and genomics data) to serve as the upper bound of current complete modality approaches. Then the average performances on all missing combinations are demonstrated under the total training missing rate of $60\\%$ , evaluated on three inference scenarios. Details on various combinations are shown in Fig. 3 and we discuss them in Sec. 4.3. Lastly, we investigate the upper bound of each model tailored to incomplete modalities by training them with all data (without missing data).", + "bbox": [ + 511, + 90, + 903, + 227 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "From results in Tab. 1, we observed that (1) When applying $60\\%$ training missing rate, DisPro consistently outperformed SOTAs across five datasets by a large margin (on average, $+1.76\\%$ for pathology only, $+1.57\\%$ for genomics only and $2.33\\%$ for complete data). This indicates the superiority of DisPro over the incomplete SOTA methods. (2) When testing on the complete scenario, on 3 out of 5 datasets and overall, DisPro even surpassed the SOTA multimodal survival model (SurvPath) trained on all data. Notably, DisPro achieved this while only utilizing incomplete training data with a $60\\%$ missing rate. It suggests that harnessing the powerful capabilities of LLMs is a promising solution for addressing the classical question of survival analysis. (3) When using complete training data, DisPro consistently performed best on average for various test scenarios as well ( $2.28\\%$ on pathology only, $1.33\\%$ on genomics only and $2.57\\%$ on complete test data). Under the same condition, DisPro further increased by $1.31\\%$ compared with the SOTA complete multimodal method, further demonstrating the strong power of LLMs. (4) In contrast to unimodal methods trained on full unimodal datasets, DisPro, using only incomplete multimodal data with a $60\\%$ missing rate, achieved superior or similar performance when inferring from individual modalities. This showcases DisPro's efficient utilization of multimodal data.", + "bbox": [ + 511, + 228, + 906, + 604 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "4.3. Ablation Study", + "text_level": 1, + "bbox": [ + 511, + 619, + 666, + 636 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Component Validation. In Tab. 2, we ablate each component in our method to investigate their effectiveness. (1) To demonstrate how the LLM performs on missing modalities without any prompts, we establish a baseline by employing a linear classifier head on the same LLM and randomly selecting feature tokens for its input. Results show that the performance significantly decreased by $2.62\\%$ on pathology only, $1.55\\%$ on genomics only and $2.05\\%$ on complete test data. This indicates the importance of appropriate prompts for LLMs. (2) When supplementing modality-specific information for missing modalities, the performance $(+UD)$ consistently improves on all datasets and increases by around $+1\\%$ on various inference settings, which validates the effectiveness of compensating for modality-specific knowledge of the missing modality. (3) When reusing UniPro to select relevant tokens $(+US)$ , the performance increased by about $+1\\%$ overall, which bene-", + "bbox": [ + 511, + 643, + 906, + 900 + ], + "page_idx": 5 + }, + { + "type": "page_footnote", + "text": "1https://portal.qdc.cancer.gov/", + "bbox": [ + 107, + 873, + 352, + 886 + ], + "page_idx": 5 + }, + { + "type": "page_footnote", + "text": "2https://www.cbiportal.org/", + "bbox": [ + 107, + 887, + 331, + 900 + ], + "page_idx": 5 + }, + { + "type": "page_number", + "text": "5107", + "bbox": [ + 482, + 944, + 514, + 955 + ], + "page_idx": 5 + }, + { + "type": "table", + "img_path": "images/725c610a214a123865532340a3d3fd3d37c66d2f12ece7b2abb97c90716f1207.jpg", + "table_caption": [], + "table_footnote": [], + "table_body": "
TrainingMiRateTestP GMethodsBLCA(372)BRCA(1007)COADREAD(533)LUAD(443)UCEC(478)Avg
N/A†TransMIL [28]0.5903±0.0250.6671±0.0540.6093±0.0180.6195±0.0350.6978±0.0460.6368
†CoOp_BioBert [44]0.6014±0.0470.6784±0.0320.6681±0.0260.6543±0.0520.6791±0.0850.6563
†SNN [15]0.6831±0.0400.6567±0.0380.6312±0.0580.6379±0.0710.7267±0.0730.6671
†CoOp_BioBert [44]0.6712±0.0360.6922±0.0350.6743±0.0480.6514±0.0650.7171±0.0600.6812
‡MOTCat [36]0.6274±0.0460.6721±0.0380.6501±0.0350.6753±0.0430.7207±0.0740.6691
‡SurvPath [14]0.6572±0.0270.7071±0.0390.7079±0.0480.6798±0.0490.7391±0.0650.6982
60%•○COM [27]0.6024±0.0130.6740±0.0110.6779±0.0140.6336±0.0110.6985±0.0120.6573
•○M3Care [40]0.6208±0.0140.6690±0.0180.6570±0.0230.6219±0.0070.7031±0.0120.6543
•○HGCN [9]0.5999±0.0090.6831±0.0120.6792±0.0090.6465±0.0050.7011±0.0070.6620
•○MUSE [34]0.6207±0.0130.6514±0.0140.6547±0.0090.6354±0.0120.7048±0.0230.6534
•○MAP [17]0.5918±0.0120.6278±0.0060.5965±0.0170.6489±0.0200.6929±0.0210.6316
•○DisPro (Ours)0.6319±0.0140.6895±0.0110.6880±0.0100.6612±0.0140.7272±0.0180.6796
○•COM [27]0.6415±0.0160.6653±0.0180.6593±0.0300.6456±0.0080.7105±0.0070.6645
○•M3Care [40]0.6420±0.0150.6642±0.0290.6491±0.0080.6462±0.0120.7032±0.0130.6609
○•HGCN [9]0.6225±0.0080.6729±0.0090.6752±0.0150.6418±0.0150.7015±0.0100.6628
○•MUSE [34]0.6299±0.0120.6427±0.0250.6391±0.0230.6462±0.0100.6893±0.0270.6494
○•MAP [17]0.5937±0.0160.6269±0.0060.5971±0.0170.6442±0.0090.6934±0.0180.6311
○•DisPro (Ours)0.6547±0.0120.6841±0.0180.6804±0.0240.6548±0.0120.7271±0.0170.6802
••COM [27]0.6192±0.0130.6846±0.0090.6874±0.0050.6548±0.0120.7136±0.0160.6719
••M3Care [40]0.6513±0.0050.6951±0.0240.6698±0.0160.6491±0.0130.7288±0.0110.6788
••HGCN [9]0.6332±0.0060.6907±0.0120.6861±0.0170.6545±0.0110.7077±0.0080.6745
••MUSE [34]0.6207±0.0120.6555±0.0130.6539±0.0120.6415±0.0080.7164±0.0130.6576
••MAP [17]0.5906±0.0130.6277±0.0040.5990±0.0160.6625±0.0080.6901±0.0180.6340
••DisPro (Ours)0.6638±0.0060.7219±0.0150.7029±0.0160.6741±0.0070.7476±0.0200.7021
0%••COM [27]0.6154±0.0410.6871±0.0270.6925±0.0440.6427±0.0460.6806±0.0610.6637
••M3Care [40]0.6268±0.0240.6775±0.0160.6850±0.0520.6299±0.0380.6964±0.0610.6631
••HGCN [9]0.6055±0.0480.6716±0.0220.6734±0.0480.6421±0.0520.6915±0.0530.6568
••MUSE [34]0.6208±0.0260.6767±0.0400.6710±0.0410.6253±0.0410.7099±0.0630.6607
••MAP [17]0.6205±0.0330.6300±0.0530.5930±0.0650.6662±0.0550.6925±0.0640.6404
••DisPro (Ours)0.6315±0.0280.6945±0.0440.7216±0.0280.6747±0.0520.7104±0.0560.6865
○•COM [27]0.6681±0.0300.6966±0.0470.6770±0.0240.6454±0.0500.7095±0.0590.6793
○•M3Care [40]0.6706±0.0390.6611±0.0400.6563±0.0210.6512±0.0610.7047±0.0600.6688
○•HGCN [9]0.6233±0.0330.6807±0.0230.6777±0.0370.6421±0.0450.7091±0.0430.6666
○•MUSE [34]0.6256±0.0290.6401±0.0240.6208±0.0300.6547±0.0750.6824±0.0390.6447
○•MAP [17]0.6093±0.0260.6305±0.0510.5842±0.0650.6653±0.0450.6841±0.0640.6347
○•DisPro (Ours)0.6745±0.0390.6996±0.0280.7000±0.0210.6751±0.0680.7140±0.0600.6926
••COM [27]0.6376±0.0430.7087±0.0240.6990±0.0450.6659±0.0340.7112±0.0510.6845
••M3Care [40]0.6605±0.0230.6808±0.0410.6768±0.0260.6740±0.0460.7359±0.0330.6856
••HGCN [9]0.6365±0.0210.7058±0.0330.7136±0.0250.6692±0.0500.7017±0.0440.6854
••MUSE [34]0.6282±0.0390.7017±0.0200.6693±0.0420.6248±0.0400.7237±0.0620.6695
••MAP [17]0.6034±0.0260.6328±0.0490.5895±0.0630.6904±0.0470.6827±0.0650.6398
••DisPro (Ours)0.6719±0.0510.7358±0.0510.7553±0.0210.6767±0.0410.7169±0.0710.7113
", + "bbox": [ + 109, + 87, + 888, + 719 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Table 1. C-Index of Comparison to SOTA models. \"MiRate\" represents Missing Rate. \"P\" and \"G\" indicate Pathology and Genomics, respectively. $\\circ$ and $\\bullet$ mean the missing modality and the available modality,respectively. † and ‡ indicate using unimodal and complete multimodal data, respectively, while the rest are for missing modalities. The best results are in bold. The second-best results are underlined.", + "bbox": [ + 89, + 727, + 906, + 770 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "fits from the assistance of US in learning modality-common knowledge from WSIs and genes with high redundancy.", + "bbox": [ + 89, + 786, + 482, + 816 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Various Missing Scenarios. To illustrate how incomplete multimodal approaches handle varying missing settings, we display different combinations of training missing rates for each inference scenario, as shown in Fig. 3. Results showcase that DisPro achieved considerably consistent perfor", + "bbox": [ + 89, + 825, + 482, + 902 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "mance superiority in various missing settings over other incomplete multimodal approaches by a clear margin.", + "bbox": [ + 511, + 786, + 906, + 816 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "5. Conclusion", + "text_level": 1, + "bbox": [ + 513, + 829, + 633, + 845 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "This work proposed a Distilled Prompt Learning (DisPro) framework to unlock the potential of LLMs for incomplete multimodal learning by two-stage prompting, in which", + "bbox": [ + 511, + 854, + 906, + 902 + ], + "page_idx": 6 + }, + { + "type": "page_number", + "text": "5108", + "bbox": [ + 482, + 945, + 514, + 955 + ], + "page_idx": 6 + }, + { + "type": "table", + "img_path": "images/021044f3074909bce32dd4ceb8e19993e3b90d88780b3ecffd6f5210ee921581.jpg", + "table_caption": [], + "table_footnote": [], + "table_body": "
VariantsTestModulesBLCA (372)BRCA (1007)COADREAD (533)LUAD (443)UCEC (478)Avg
PGUSUD
baseline•○0.6068±0.0210.6668±0.0060.6583±0.0120.6270±0.0180.7081±0.0190.6534
+ UD•○✓0.6213±0.0100.6829±0.0080.6740±0.0220.6327±0.0180.7152±0.0110.6652
+ US•○✓0.6285±0.0210.6884±0.0160.6653±0.0260.6546±0.0130.7203±0.0070.6714
DisPro (full)•○✓✓0.6319±0.0140.6895±0.0110.6880±0.0100.6612±0.0140.7272±0.0180.6796
baseline○•0.6435±0.0160.6605±0.0330.6601±0.0180.6467±0.0160.7127±0.0070.6647
+ UD○•✓0.6473±0.0170.6836±0.0140.6707±0.0220.6420±0.0200.7226±0.0130.6732
+ US○•✓0.6452±0.0200.6666±0.0170.6763±0.0230.6471±0.0230.7151±0.0090.6701
DisPro (full)○•✓✓0.6547±0.0120.6841±0.0180.6804±0.0240.6548±0.0120.7271±0.0170.6802
baseline••0.6522±0.0100.7075±0.0120.6785±0.0200.6541±0.0090.7157±0.0030.6816
+ UD••✓0.6585±0.0070.7194±0.0130.6963±0.0090.6635±0.0140.7250±0.0120.6926
+ US••✓0.6593±0.0080.7125±0.0100.7005±0.0180.6717±0.0070.7274±0.0100.6943
DisPro (full)••✓✓0.6638±0.0060.7219±0.0150.7029±0.0160.6741±0.0070.7476±0.0200.7021
", + "bbox": [ + 109, + 88, + 893, + 310 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Table 2. C-Index of Ablation Study on the Setting of $60\\%$ Training Missing Rate.\"US\" and \"UD\" refer to the UniPro Scoring and UniPro Distillation modules, respectively. \"P\" and \"G\" indicate Pathology and Genomics, respectively. $\\circ$ means the modality is missing, while $\\bullet$ refers to the available modality.", + "bbox": [ + 89, + 321, + 906, + 362 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/4918a135959d8a2aaa250d2f0febc2b07cefb70695a1228d78a63e9cb9a6efb9.jpg", + "image_caption": [ + "Figure 3. Performance on various combinations under $60\\%$ training missing rate for different test scenarios." + ], + "image_footnote": [], + "bbox": [ + 117, + 363, + 879, + 714 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Unimodal Prompting (UniPro) and Multimodal Prompting (MultiPro) were proposed for supplementing modality-specific and modality-common knowledge for missing modalities. In the first stage, UniPro extended CoOp into the MIL setting and distilled knowledge distribution to a set of learnable prompts for each modality, which prepped for compensation of modality-specific knowledge. In the second stage, considering available modalities as prompts, we leveraged an LLM to impute common knowledge for miss", + "bbox": [ + 88, + 760, + 480, + 896 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "ing modalities. To inject the specific knowledge of missing modalities, the distilled prompts in UniPro are used to supervise the learning of imputing representations for missing modalities, leading to more comprehensive knowledge compensation in incomplete multimodal learning. Extensive experiments on five cancer benchmark datasets demonstrated the effectiveness of the proposed method over state-of-the-art approaches.", + "bbox": [ + 511, + 760, + 906, + 882 + ], + "page_idx": 7 + }, + { + "type": "page_number", + "text": "5109", + "bbox": [ + 482, + 944, + 514, + 955 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Acknowledgement", + "text_level": 1, + "bbox": [ + 91, + 90, + 250, + 107 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "This work was supported by the National Natural Science Foundation of China (No. 62202403), Hong Kong Innovation and Technology Commission (Project No. MHP/002/22 and ITCPD/17-9), and Research Grants Council of the Hong Kong Special Administrative Region, China (Project No. R6003-22 and C4024-22GF).", + "bbox": [ + 89, + 113, + 483, + 200 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "References", + "text_level": 1, + "bbox": [ + 91, + 224, + 187, + 239 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[1] Som S Biswas. Role of chat gpt in public health. Annals of biomedical engineering, 51(5):868-869, 2023. 2", + "[2] Jason Brownlee. A gentle introduction to the rectified linear unit (relu). Machine learning mastery, 6, 2019. 3, 1", + "[3] Agisilaos Chartsias, Thomas Joyce, Mario Valerio Giuffrida, and Sotirios A Tsaftaris. Multimodal mr synthesis via modality-invariant latent representation. IEEE transactions on medical imaging, 37(3):803-814, 2017. 1", + "[4] Jiayi Chen and Aidong Zhang. Hgmf: heterogeneous graph-based fusion for multimodal data with incompleteness. In Proceedings of the 26th ACM SIGKDD international conference on knowledge discovery & data mining, pages 1295-1305, 2020. 1, 2", + "[5] Richard J Chen, Ming Y Lu, Wei-Hung Weng, Tiffany Y Chen, Drew FK Williamson, Trevor Manz, Maha Shady, and Faisal Mahmood. Multimodal co-attention transformer for survival prediction in gigapixel whole slide images. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 4015–4025, 2021. 1, 3, 6", + "[6] Richard J Chen, Tong Ding, Ming Y Lu, Drew FK Williamson, Guillaume Jaume, Andrew H Song, Bowen Chen, Andrew Zhang, Daniel Shao, Muhammad Shaban, et al. Towards a general-purpose foundation model for computational pathology. Nature Medicine, 30(3):850-862, 2024. 1", + "[7] Florinel-Alin Croitoru, Vlad Hondru, Radu Tudor Ionescu, and Mubarak Shah. Diffusion models in vision: A survey. IEEE Transactions on Pattern Analysis and Machine Intelligence, 45(9):10850-10869, 2023. 2", + "[8] Yuhang Ding, Xin Yu, and Yi Yang. Rfnet: Region-aware fusion network for incomplete multi-modal brain tumor segmentation. In Proceedings of the IEEE/CVF international conference on computer vision, pages 3975-3984, 2021. 1, 2", + "[9] Hou et al. Hybrid graph convolutional network with online masked autoencoder for robust multimodal cancer survival prediction. IEEE Transactions on Medical Imaging, 2023. 7", + "[10] Zhou et al. Cross-modal translation and alignment for survival analysis. In ICCV, pages 21485-21494, 2023. 1, 3", + "[11] Michael Gadermayr and Maximilian Tschuchnig. Multiple instance learning for digital pathology: A review of the state-of-the-art, limitations & future potential. Computerized Medical Imaging and Graphics, page 102337, 2024. 2", + "[12] Mohammad Hamghalam, Alejandro F Frangi, Baiying Lei, and Amber L Simpson. Modality completion via gaussian process prior variational autoencoders for multi-modal" + ], + "bbox": [ + 93, + 250, + 485, + 900 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "glioma segmentation. In Medical Image Computing and Computer Assisted Intervention-MICCAI 2021: 24th International Conference, Strasbourg, France, September 27-October 1, 2021, Proceedings, Part VII 24, pages 442-452. Springer, 2021. 2", + "[13] Maximilian Ilse, Jakub Tomczak, and Max Welling. Attention-based deep multiple instance learning. In International conference on machine learning, pages 2127-2136. PMLR, 2018. 2, 5", + "[14] Guillaume Jaume, Anurag Vaidya, Richard J Chen, Drew FK Williamson, Paul Pu Liang, and Faisal Mahmood. Modeling dense multimodal interactions between biological pathways and histology for survival prediction. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 11579-11590, 2024. 3, 6, 7, 1", + "[15] Günter Klambauer, Thomas Unterthiner, Andreas Mayr, and Sepp Hochreiter. Self-normalizing neural networks. Advances in neural information processing systems, 30, 2017. 7, 1", + "[16] Jinhyuk Lee, Wonjin Yoon, Sungdong Kim, Donghyeon Kim, Sunkyu Kim, Chan Ho So, and Jaewoo Kang. Biobert: a pre-trained biomedical language representation model for biomedical text mining. Bioinformatics, 36(4):1234-1240, 2020. 6, 1", + "[17] Yi-Lun Lee, Yi-Hsuan Tsai, Wei-Chen Chiu, and Chen-Yu Lee. Multimodal prompting with missing modalities for visual recognition. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 14943-14952, 2023. 2, 7, 3", + "[18] Bin Li, Yin Li, and Kevin W Eliceiri. Dual-stream multiple instance learning network for whole slide image classification with self-supervised contrastive learning. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 14318-14328, 2021. 3", + "[19] Chenxin Li, Xinyu Liu, Cheng Wang, Yifan Liu, Weihao Yu, Jing Shao, and Yixuan Yuan. Gtp-4o: Modality-prompted heterogeneous graph learning for omni-modal biomedical representation. In European conference on computer vision, pages 168–187. Springer, 2024. 2", + "[20] Chunyuan Li, Cliff Wong, Sheng Zhang, Naoto Usuyama, Haotian Liu, Jianwei Yang, Tristan Naumann, Hoifung Poon, and Jianfeng Gao. Llava-med: Training a large language-and-vision assistant for biomedicine in one day. Advances in Neural Information Processing Systems, 36, 2024. 2", + "[21] Han Liu, Yubo Fan, Hao Li, Jiacheng Wang, Dewei Hu, Can Cui, Ho Hin Lee, Huahong Zhang, and Ipek Oguz. Moddrop++: A dynamic filter network with intra-subject cotraining for multiple sclerosis lesion segmentation with missing modalities. In International Conference on Medical Image Computing and Computer-Assisted Intervention, pages 444-453. Springer, 2022. 2", + "[22] Mengmeng Ma, Jian Ren, Long Zhao, Sergey Tulyakov, Cathy Wu, and Xi Peng. Smil: Multimodal learning with severely missing modality. In Proceedings of the AAAI Conference on Artificial Intelligence, pages 2302-2310, 2021. 2, 6", + "[23] Xiangxi Meng, Kaicong Sun, Jun Xu, Xuming He, and Dinggang Shen. Multi-modal modality-masked diffusion network" + ], + "bbox": [ + 516, + 92, + 906, + 900 + ], + "page_idx": 8 + }, + { + "type": "page_number", + "text": "5110", + "bbox": [ + 482, + 944, + 516, + 955 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "for brain mri synthesis with random modality missing. IEEE Transactions on Medical Imaging, 2024. 1, 2", + "[24] Muhammad Khalid Khan Niazi, Anil V Parwani, and Metin N Gurcan. Digital pathology and artificial intelligence. The lancet oncology, 20(5):e253-e261, 2019. 3", + "[25] Zhenyuan Ning, Denghui Du, Chao Tu, Qianjin Feng, and Yu Zhang. Relation-aware shared representation learning for cancer prognosis analysis with auxiliary clinical variables and incomplete multi-modality data. IEEE Transactions on Medical Imaging, 41(1):186-198, 2021. 2", + "[26] Fatih Ozsolak and Patrice M Milos. RNA sequencing: advances, challenges and opportunities. Nature reviews genetics, 12(2):87-98, 2011. 3", + "[27] Shuwei Qian and Chongjun Wang. Com: Contrastive masked-attention model for incomplete multimodal learning. Neural Networks, 162:443–455, 2023. 7", + "[28] Zhuchen Shao, Hao Bian, Yang Chen, Yifeng Wang, Jian Zhang, Xiangyang Ji, et al. Transmil: Transformer based correlated multiple instance learning for whole slide image classification. Advances in neural information processing systems, 34:2136-2147, 2021. 3, 7", + "[29] Anmol Sharma and Ghassan Hamarneh. Missing mri pulse sequence synthesis using multi-modal generative adversarial network. IEEE transactions on medical imaging, 39(4): 1170-1183, 2019. 1", + "[30] Andrew H Song, Richard J Chen, Guillaume Jaume, Anurag Jayant Vaidya, Alexander Baras, and Faisal Mahmood. Multimodal prototyping for cancer survival prediction. In *Forty-first International Conference on Machine Learning*, 2024. 1, 2, 3, 5", + "[31] Rory Stark, Marta Grzelak, and James Hadfield. Rna sequencing: the teenage years. Nature Reviews Genetics, 20 (11):631-656, 2019. 3", + "[32] Hu Wang, Yuanhong Chen, Congbo Ma, Jodie Avery, Louise Hull, and Gustavo Carneiro. Multi-modal learning with missing modality via shared-specific feature modelling. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 15878-15887, 2023. 6", + "[33] Tianyu Wu, Shizhu He, Jingping Liu, Siqi Sun, Kang Liu, Qing-Long Han, and Yang Tang. A brief overview of chatgpt: The history, status quo and potential future development. IEEE/CAA Journal of Automatica Sinica, 10(5):1122-1136, 2023. 2", + "[34] Zhenbang Wu, Anant Dadu, Nicholas Tustison, Brian Avants, Mike Nalls, Jimeng Sun, and Faraz Faghri. Multimodal patient representation learning with missing modalities and labels. In The Twelfth International Conference on Learning Representations, 2024. 1, 2, 7", + "[35] Conghao Xiong, Hao Chen, Hao Zheng, Dong Wei, Yefeng Zheng, Joseph JY Sung, and Irwin King. Mome: Mixture of multimodal experts for cancer survival prediction. In International Conference on Medical Image Computing and Computer-Assisted Intervention, pages 318-328. Springer, 2024. 1", + "[36] Yingxue Xu and Hao Chen. Multimodal optimal transport-based co-attention transformer with global structure consistency for survival prediction. In Proceedings of the" + ], + "bbox": [ + 91, + 90, + 483, + 900 + ], + "page_idx": 9 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "IEEE/CVF International Conference on Computer Vision, pages 21241-21251, 2023. 1, 3, 6, 7", + "[37] Shu Yang, Yihui Wang, and Hao Chen. Mambamil: Enhancing long sequence modeling with sequence reordering in computational pathology. In International Conference on Medical Image Computing and Computer-Assisted Intervention, pages 296-306. Springer, 2024. 2", + "[38] Yixiao Yuan, Yawen Huang, and Yi Zhou. Rethinking a unified generative adversarial model for mri modality completion. In International Conference on Medical Image Computing and Computer-Assisted Intervention, pages 143-153. Springer, 2023. 2", + "[39] Shekoufeh Gorgi Zadeh and Matthias Schmid. Bias in cross-entropy-based training of deep survival networks. IEEE transactions on pattern analysis and machine intelligence, 43(9):3126-3137, 2020. 3", + "[40] Chaohe Zhang, Xu Chu, Liantao Ma, Yinghao Zhu, Yasha Wang, Jiangtao Wang, and Junfeng Zhao. M3care: Learning with missing modalities in multimodal healthcare data. In Proceedings of the 28th ACM SIGKDD Conference on Knowledge Discovery and Data Mining, pages 2418-2428, 2022. 1, 2, 7", + "[41] Hongrun Zhang, Yanda Meng, Yitian Zhao, Yihong Qiao, Xiaoyun Yang, Sarah E Coupland, and Yalin Zheng. Dtfmil: Double-tier feature distillation multiple instance learning for histopathology whole slide image classification. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 18802-18812, 2022. 3", + "[42] Yilan Zhang, Yingxue Xu, Jianqi Chen, Fengying Xie, and Hao Chen. Prototypical information bottlenecking and disentangling for multimodal cancer survival prediction. In The Twelfth International Conference on Learning Representations, 2024. 1, 2, 3, 5", + "[43] Huajun Zhou, Fengtao Zhou, Chenyu Zhao, Yingxue Xu, Luyang Luo, and Hao Chen. Multimodal data integration for precision oncology: Challenges and future directions. arXiv preprint arXiv:2406.19611, 2024. 1, 2", + "[44] Kaiyang Zhou, Jingkang Yang, Chen Change Loy, and Ziwei Liu. Learning to prompt for vision-language models. International Journal of Computer Vision, 130(9):2337-2348, 2022. 2, 3, 7" + ], + "bbox": [ + 516, + 92, + 903, + 667 + ], + "page_idx": 9 + }, + { + "type": "page_number", + "text": "5111", + "bbox": [ + 482, + 944, + 513, + 955 + ], + "page_idx": 9 + } +] \ No newline at end of file diff --git a/2025/Distilled Prompt Learning for Incomplete Multimodal Survival Prediction/5aae416b-ca2c-40f8-8583-f2902bfc3349_model.json b/2025/Distilled Prompt Learning for Incomplete Multimodal Survival Prediction/5aae416b-ca2c-40f8-8583-f2902bfc3349_model.json new file mode 100644 index 0000000000000000000000000000000000000000..810384bf69a79e178659d1dab70b8b47b87256a0 --- /dev/null +++ b/2025/Distilled Prompt Learning for Incomplete Multimodal Survival Prediction/5aae416b-ca2c-40f8-8583-f2902bfc3349_model.json @@ -0,0 +1,1969 @@ +[ + [ + { + "type": "header", + "bbox": [ + 0.107, + 0.003, + 0.182, + 0.043 + ], + "angle": 0, + "content": "CVF" + }, + { + "type": "header", + "bbox": [ + 0.237, + 0.001, + 0.812, + 0.047 + ], + "angle": 0, + "content": "This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore." + }, + { + "type": "title", + "bbox": [ + 0.126, + 0.131, + 0.872, + 0.153 + ], + "angle": 0, + "content": "Distilled Prompt Learning for Incomplete Multimodal Survival Prediction" + }, + { + "type": "text", + "bbox": [ + 0.129, + 0.18, + 0.87, + 0.234 + ], + "angle": 0, + "content": "Yingxue Xu\\(^{1}\\) Fengtao Zhou\\(^{1}\\) Chenyu Zhao\\(^{1}\\) Yihui Wang\\(^{1}\\) Can Yang\\(^{2}\\) Hao Chen\\(^{1\\boxtimes}\\) \n\\(^{1}\\)Department of Computer Science and Engineering, \\(^{2}\\)Department of Mathematics The Hong Kong University of Science and Technology" + }, + { + "type": "text", + "bbox": [ + 0.202, + 0.237, + 0.792, + 0.253 + ], + "angle": 0, + "content": "{yxueb,fzhouaf,cyzhao,ywangrm}@connect.ust.hk，{macyang,jhc}@ust.hk" + }, + { + "type": "title", + "bbox": [ + 0.248, + 0.287, + 0.327, + 0.303 + ], + "angle": 0, + "content": "Abstract" + }, + { + "type": "text", + "bbox": [ + 0.089, + 0.318, + 0.485, + 0.726 + ], + "angle": 0, + "content": "The integration of multimodal data including pathology images and gene profiles is widely applied in precise survival prediction. Despite recent advances in multimodal survival models, collecting complete modalities for multimodal fusion still poses a significant challenge, hindering their application in clinical settings. Current approaches tackling incomplete modalities often fall short, as they typically compensate for only a limited part of the knowledge of missing modalities. To address this issue, we propose a Distilled Prompt Learning framework (DisPro) to utilize the strong robustness of Large Language Models (LLMs) to missing modalities, which employs two-stage prompting for compensation of comprehensive information for missing modalities. In the first stage, Unimodal Prompt (UniPro) distills the knowledge distribution of each modality, preparing for supplementing modality-specific knowledge of the missing modality in the subsequent stage. In the second stage, Multimodal Prompt (MultiPro) leverages available modalities as prompts for LLMs to infer the missing modality, which provides modality-common information. Simultaneously, the unimodal knowledge acquired in the first stage is injected into multimodal inference to compensate for the modality-specific knowledge of the missing modality. Extensive experiments covering various missing scenarios demonstrated the superiority of the proposed method. The code is available at https://github.com/Innse/DisPro." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.755, + 0.222, + 0.77 + ], + "angle": 0, + "content": "1. Introduction" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.78, + 0.484, + 0.901 + ], + "angle": 0, + "content": "Multimodal Survival Prediction typically incorporates pathology slides and genomic data complementary to each other for precise prognosis analysis. Pathology data can provide qualitative morphological insights, while quantitative molecular information can be obtained from genomic data [5, 35, 36]. Despite progress [5, 10, 36] in multimodal survival analysis, these models typically require complete modalities data for reliable performance. However, the ex" + }, + { + "type": "image", + "bbox": [ + 0.536, + 0.287, + 0.88, + 0.427 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.512, + 0.433, + 0.908, + 0.489 + ], + "angle": 0, + "content": "Figure 1. Insights for existing incomplete multimodal learning and comparison to the proposed method. (a) Generation-based Imputation and Imputation-free approaches, (b) Retrieved-based Imputation and (c) Ours." + }, + { + "type": "text", + "bbox": [ + 0.511, + 0.504, + 0.907, + 0.671 + ], + "angle": 0, + "content": "pectation of having full access to complete modalities for integration is not always achievable, especially in clinical scenarios, due to various factors such as the cost of data gathering and privacy issues. For example, acquiring genomic data still comes at a high cost due to the technology and infrastructure required for gene sequencing, especially in underdeveloped areas [43]. Inevitably, it is challenging to adapt previously established multimodal survival models [30, 42] to real-world clinical settings. As a result, building a multimodal survival model that is robust to incomplete modalities data is an urgent issue for clinical practice." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.675, + 0.909, + 0.903 + ], + "angle": 0, + "content": "Currently, mainstream approaches addressing incomplete modalities issues include imputation-based and imputation-free methods [8, 34]. The former approaches aim to impute the missing modality based on observed modalities, primarily achieved in two ways including generation [3, 23, 29] and retrieval [4, 40]. In terms of generation-based imputation, generative models are typically employed for synthesizing features or raw data of missing modalities from available modalities. As shown in Fig. 1 (a), these methods usually can only impute modality-common information, as the modality-specific information unique to the missing modality cannot be generated out of thin air. Moreover, the goal of retrieval-based imputation methods is to fill in missing modalities by seeking the most similar sample of the missing modality from the training set. This way tends" + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.516, + 0.957 + ], + "angle": 0, + "content": "5102" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.09, + 0.092, + 0.482, + 0.168 + ], + "angle": 0, + "content": "to compensate only a limited part of modality-common and modality-specific information, as shown in Fig. 1 (b). The reason is that an individual sample exhibits considerable randomness and is challenging to fully capture the unique knowledge inherent in the missing modality." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.17, + 0.483, + 0.426 + ], + "angle": 0, + "content": "Unlike imputation-based approaches, imputation-free approaches [8, 34] primarily focus on how to minimize performance degradation resulting from missing modalities. These approaches make great efforts to learn robust multimodal representations that are less sensitive to missing modalities by capturing modality-invariant information. Recently, with the rise of Large Language Models (LLMs) [1, 20, 33], the researcher [17] attempts to leverage the strong robustness of LLMs benefit from extensive training data to alleviate performance decreases caused by incomplete modalities. They design some sophisticated prompts to inform LLMs of different distributions of inputs. However, they only capture modality-common information as well (Fig. 1a), ignoring distinct modality-specific knowledge of the missing modality. As such, existing methods still struggle to effectively incorporate the learning of both common and specific knowledge of missing modalities." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.429, + 0.483, + 0.565 + ], + "angle": 0, + "content": "Based on these insights, to bridge the aforementioned gap, we propose a Distilled Prompt Learning (DisPro) framework to unlock LLMs' robustness to incomplete modalities, which aims to simultaneously compensate for both specific and common information of missing modalities. It comprises two-stage prompt learning including Unimodal Prompting (UniPro) and Multimodal Prompting (MultiPro) for capturing modality-specific and modality-common information of missing modalities, respectively." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.567, + 0.483, + 0.824 + ], + "angle": 0, + "content": "In the first stage, to prepare for supplementing modality-specific information of missing modalities, we first learn the knowledge distribution of the missing modality by distilling unimodal knowledge into a set of learnable unimodal prompts for each modality from all remaining samples available of this modality. Specifically, inspired by CoOp [44], a set of learnable prompts for different classes is employed to learn knowledge about how to describe different categories of an individual modality. For example, in survival prediction, learned prompts can be intuitively interpreted as how different risk groups will reflect in pathological morphology or gene expression profiles. However, due to the huge size (e.g., \\(100,000 \\times 100,000\\)) of pathological WSIs, we cannot simply apply CoOp designed for small natural images to WSIs. Therefore, in UniPro, we extend CoOp into the multiple instance learning (MIL) paradigm that is commonly used for large-scale WSIs [11, 13, 37]." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.826, + 0.483, + 0.901 + ], + "angle": 0, + "content": "In the second stage, to capture modality-common information for missing modalities, we leverage available modalities as multimodal prompts of LLMs to infer representations of missing modalities using LLMs' strong inference ability. Simultaneously, to inject the learned unimodal" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.092, + 0.905, + 0.258 + ], + "angle": 0, + "content": "knowledge in stage 1, we propose UniPro Distillation to enforce the inferred representations of missing modalities to get as close as possible to the learned unimodal prompts of the corresponding class. Additionally, to assist in learning modality-common knowledge from WSIs and genes with high information redundancy [30, 42], we propose UniPro Scoring to re-use UniPro as a grader to select discriminative and relevant tokens. As a result, both common and specific information are compensated for missing modalities under the collaboration of uni- and multimodal prompts, thus mitigating performance declines led by incomplete modalities." + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.259, + 0.905, + 0.334 + ], + "angle": 0, + "content": "It is worth noting that DisPro is proposed as a framework for addressing missing modality issues by leveraging LLMs' power based on prompts, capable of generalizing to a wider range of domains. The contributions of this work are summarized as follows:" + }, + { + "type": "text", + "bbox": [ + 0.514, + 0.336, + 0.904, + 0.396 + ], + "angle": 0, + "content": "- We propose a Distilled Prompt Learning (DisPro) framework for incomplete multimodal survival prediction, where LLMs' robustness to incomplete modalities is explored under two-stage prompting." + }, + { + "type": "text", + "bbox": [ + 0.515, + 0.397, + 0.904, + 0.457 + ], + "angle": 0, + "content": "- We propose two-stage prompt learning including Unimodal Prompting and Multimodal Prompting to compensate for both modality-specific and modality-common information of missing modalities, respectively." + }, + { + "type": "text", + "bbox": [ + 0.515, + 0.457, + 0.904, + 0.502 + ], + "angle": 0, + "content": "- Extensive experiments on five benchmark survival prediction datasets are conducted, which demonstrates significantly superior to state-of-the-art approaches." + }, + { + "type": "list", + "bbox": [ + 0.514, + 0.336, + 0.904, + 0.502 + ], + "angle": 0, + "content": null + }, + { + "type": "title", + "bbox": [ + 0.514, + 0.519, + 0.655, + 0.535 + ], + "angle": 0, + "content": "2. Related Work" + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.546, + 0.807, + 0.562 + ], + "angle": 0, + "content": "2.1. Incomplete Multimodal Learning" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.569, + 0.905, + 0.903 + ], + "angle": 0, + "content": "As stated above, incomplete multimodal learning [43] can be categorized into two types: imputation-based [4, 12, 38] and imputation-free methods [21, 25]. For example, with the success of the diffusion model [7], the research [23] attempted to apply it to MRI synthesis for missing modalities. Instead of generating raw modality data, SMIL [22] leveraged Bayesian meta-learning to impute features of missing modalities. For retrieval-based approaches, M3Care [40] filled in the missing modalities by retrieving the most similar samples from the training set. On the other hand, imputation-free approaches provided more flexible solutions, which aim to enhance the robustness of fusion and alleviate performance decrease. [8] proposed a segmentation-based regularizer to reduce the sensitivity to missing modalities. MUSE [34] utilized patients' relationships based on graph networks to perform contrastive learning for improving the robustness to missing modalities. GTP-4o [19] proposed modality-prompted heterogeneous graph to embed each modality into a unified space and complete missing modality through a graph prompting mechanism. Recently, MAP [17] proposed missing-aware prompts to unlock the robustness of LLMs to missing" + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.516, + 0.957 + ], + "angle": 0, + "content": "5103" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.09, + 0.092, + 0.483, + 0.138 + ], + "angle": 0, + "content": "modalities by indicating LLMs different missing scenarios. However, these methods typically supplemented modality-common knowledge for missing modalities." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.148, + 0.373, + 0.163 + ], + "angle": 0, + "content": "2.2. Multimodal Survival Prediction" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.17, + 0.484, + 0.474 + ], + "angle": 0, + "content": "With the success of digital pathology [18, 24, 41] and sequencing techniques [26, 31], pathology and genomic data are commonly integrated to analyze survival outcomes, which have proven the superiority on precision in previous works [5, 36]. MCAT [5] proposed to leverage the interactions across modalities to guide the learning of pathology data. Based on this work, MOTCat [36] further leveraged optimal transport to capture global structure consistency across modalities. In parallel, CMTA [10] investigated the intrinsic cross-modal interactions and potential complementary information by cross-modal translation and alignment. From the perspective of information bottleneck, PIBD [42] proposed a prototypical information bottleneck for the MIL setting to select discriminative instances and disentangle multimodal knowledge for redundancy reduction. Similarly, MMP [30] utilized Gaussian Mixture Model to summarize morphological prototypes and biological pathway prototypes for these two modalities. However, all these methods assume all modalities available, inevitably leading to the difficulty of employment in clinical settings." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.488, + 0.182, + 0.503 + ], + "angle": 0, + "content": "3. Method" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.513, + 0.484, + 0.635 + ], + "angle": 0, + "content": "This work aims to explore LLMs' robustness to incomplete modalities by two-stage prompting, Unimodal Prompting (UniPro) and Multimodal Prompting (MultiPro), which capture modality-specific and modality-common information, respectively. The overview framework of the proposed method is shown in Fig. 2. First, we will introduce the problem formulation in Sec. 3.1. Then, we will detail the process of UniPro and MultiPro in Sec. 3.2 and 3.3, respectively." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.646, + 0.295, + 0.66 + ], + "angle": 0, + "content": "3.1. Problem Formulation" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.668, + 0.483, + 0.746 + ], + "angle": 0, + "content": "Given a multimodal dataset consisting of \\(M = 2\\) modalities, we use a 2-tuple \\(\\mathbf{X}_n = (\\mathbf{X}_n^p,\\mathbf{X}_n^g)\\) to represent the \\(i\\)-th patient data, where \\(\\mathbf{X}_n^p\\) and \\(\\mathbf{X}_n^g\\) are pathology and genomic data, respectively. In particular, we use the notation of tilde hat to indicate the modality is missing, e.g., \\(\\widetilde{\\mathbf{X}}_n^p\\)." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.747, + 0.484, + 0.902 + ], + "angle": 0, + "content": "Following the conventional setting for pathological WSIs and genomic pathways in previous works [14, 42], we formulate a WSI \\( \\mathbf{X}_n^p \\) and a set of genomic pathways \\( \\mathbf{X}_n^g \\) as the \"bag\" of instances in MIL paradigm. We denote \\( \\mathbf{X}_n^p = \\{\\mathbf{x}_{n,i}^p\\}_{i=1}^{M_p} \\) where each instance \\( \\mathbf{x}_{n,i}^p \\) refers to a patch image cropped from the WSI \\( \\mathbf{X}_n^p \\) and \\( M_p \\) is the patch number of this WSI. Similarly, the bag of genomic pathways can be formulated as \\( \\mathbf{X}_n^g = \\{\\mathbf{x}_{n,i}^g\\}_{i=1}^{M_g} \\), in which an instance \\( \\mathbf{x}_{n,i}^g \\) represents a pathway consisting of various gene expression values." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.092, + 0.905, + 0.288 + ], + "angle": 0, + "content": "In survival prediction, the goal is to estimate the risk probability of an outcome event taking place before a specific time. However, this event may not always be observed, resulting in a right-censored occurrence. Therefore, \\( c_{n} \\in \\{0,1\\} \\) is used to denote censorship status indicating whether the outcome event is right-censored (i.e., alive) (\\( c = 1 \\)) or not (\\( c = 0 \\)) (i.e., dead), and \\( t_{n} \\in \\mathbb{R}^{+} \\) represents the overall survival time (in months). These two components make up its set of labels \\( Y_{n} = (c_{n},t_{n}) \\). Following the previous setting [5, 36], the continuous survival time would be discretized into \\( I_{t} \\) time intervals corresponding to \\( I_{t} \\) risk bands. As a result, the label set \\( (c_{n},t_{n}) \\) can be re-written as \\( y_{n} \\), resulting in \\( 2I_{t} \\) labels." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.289, + 0.905, + 0.363 + ], + "angle": 0, + "content": "Given a set of predicted hazard probabilities \\(\\mathbf{h}_n\\) across all time intervals, the hazard probability \\(\\mathbf{h}_n^{(j)}\\) represents the probability of death event occurring at the \\(j\\)-th time interval. Then, the cumulative survival probability for each time interval can be estimated as follows:" + }, + { + "type": "equation", + "bbox": [ + 0.637, + 0.374, + 0.905, + 0.415 + ], + "angle": 0, + "content": "\\[\nS _ {n} ^ {(j)} = \\prod_ {z = 1} ^ {j} \\left(1 - \\mathbf {h} _ {n} ^ {(z)}\\right) \\tag {1}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.419, + 0.905, + 0.465 + ], + "angle": 0, + "content": "Following previous works [36, 42], given the ground truth time interval \\(\\tau\\), we use the NLL loss [39] for survival prediction, where \\(N\\) is the sample number of the training set:" + }, + { + "type": "equation", + "bbox": [ + 0.534, + 0.476, + 0.905, + 0.54 + ], + "angle": 0, + "content": "\\[\n\\begin{array}{l} \\mathcal {L} _ {s u r v} = - \\sum_ {n = 1} ^ {N} \\left[ c _ {n} \\log S _ {n} ^ {(\\tau)} + \\left(1 - c _ {n}\\right) \\log S _ {n} ^ {(\\tau - 1)} \\right. \\tag {2} \\\\ \\left. + \\left(1 - c _ {n}\\right) \\log {\\bf h} _ {n} ^ {(\\tau)} \\right] \\\\ \\end{array}\n\\]" + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.551, + 0.854, + 0.567 + ], + "angle": 0, + "content": "3.2. Stage 1: Unimodal Prompting (UniPro)" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.573, + 0.905, + 0.67 + ], + "angle": 0, + "content": "Given a WSI \\( \\mathbf{X}_n^p = \\{\\mathbf{x}_{n,i}^p\\}_{i=1}^{M_p} \\) or a bag of genomic pathways \\( \\mathbf{X}_n^g = \\{\\mathbf{x}_{n,i}^g\\}_{i=1}^{M_g} \\), we formulate their learning process as a MIL task. Inspired by CoOp [44] employing a CLIP-style architecture, we adapt it to the MIL setting to learn how to characterize each modality using a set of learnable prompts for modality-specific knowledge." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.671, + 0.906, + 0.825 + ], + "angle": 0, + "content": "In terms of pathological WSIs data, following the common setting of the WSI processing [28], as shown in the UniPro-P of Fig 2, we first employ a pretrained and frozen patch encoder \\( g_{E}(\\cdot) \\) to embed every patch images \\( \\pmb{x}_{n,i}^{p} \\) from a WSI into an instance-level embedding \\( p_n^i = g_E(\\pmb{x}_{n,i}^p) \\in \\mathbb{R}^d \\), where \\( d \\) is the patch feature dimension. To align pathology features with the text space, we apply an adapter \\( f_{p}(\\cdot) \\) for each patch feature to get the aligned patch feature \\( \\mathbf{p}_n^i = f_p(g_E(\\pmb{x}_{n,i}^p)) \\in \\mathbb{R}^D \\), implemented by a linear layer followed by ReLU [2]." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.826, + 0.905, + 0.902 + ], + "angle": 0, + "content": "Text-wise, the input comprises three components: pathology context tokens, prefix tokens andclassname tokens. Forclassname tokens \\([C]_{1\\dots s}\\), we map every time interval of overall survival into a unique description, such as \"median risk\", and denote the censorship status by \"alive\"" + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.516, + 0.957 + ], + "angle": 0, + "content": "5104" + } + ], + [ + { + "type": "image_caption", + "bbox": [ + 0.37, + 0.096, + 0.651, + 0.111 + ], + "angle": 0, + "content": "Stage 1: Unimodal Prompting (UniPro)" + }, + { + "type": "image", + "bbox": [ + 0.116, + 0.112, + 0.482, + 0.282 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.498, + 0.112, + 0.871, + 0.278 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.117, + 0.283, + 0.435, + 0.516 + ], + "angle": 0, + "content": null + }, + { + "type": "image_footnote", + "bbox": [ + 0.142, + 0.516, + 0.207, + 0.531 + ], + "angle": 0, + "content": "d.1. \\(\\mathcal{L}_{\\mathrm{surv}}^{\\mathrm{cls}}\\)" + }, + { + "type": "image_caption", + "bbox": [ + 0.131, + 0.533, + 0.433, + 0.549 + ], + "angle": 0, + "content": "Stage 2: Multimodal Prompting (MultiPro)" + }, + { + "type": "image", + "bbox": [ + 0.445, + 0.283, + 0.877, + 0.433 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.446, + 0.436, + 0.681, + 0.546 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.681, + 0.453, + 0.878, + 0.541 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.089, + 0.558, + 0.907, + 0.629 + ], + "angle": 0, + "content": "Figure 2. Overview of DisPro. Stage 1: Unimodal Prompting aims to distill the knowledge distribution for each modality and prepare to supplement modality-specific knowledge for the missing modality. Stage 2: Multimodal Prompting utilizes the available modality as prompts to infer the representations of the missing one, compensating for modality-common knowledge. Simultaneously, the learned UniPro of Stage 1 supervises the learning of imputed representations to compensate for modality-specific knowledge. UniPro Scoring re-uses learned prompts to assist the LLM in capturing modality-common knowledge by selecting discriminative and relevant tokens." + }, + { + "type": "text", + "bbox": [ + 0.089, + 0.643, + 0.483, + 0.765 + ], + "angle": 0, + "content": "or \"dead\". The prefix tokens \\(\\left[\\mathrm{PREFIX}\\right]_{1\\dots m}\\) briefly describe the modality, e.g., \"This is a pathology slide image from the patient with overall survival of \\([C]_{1\\dots s}\\)\". Prefix and classname tokens are obtained using a frozen Tokenizer pretrained on a large-scale biomedical corpus to embed these words. To learn how to describe the modality for a specific risk band (class), we attach a set of learnable context tokens \\([P]_{1\\dots k}\\) to prefix and classname tokens:" + }, + { + "type": "equation", + "bbox": [ + 0.095, + 0.77, + 0.483, + 0.787 + ], + "angle": 0, + "content": "\\[\n\\boldsymbol {T} _ {p} = [ P ] _ {1} \\dots [ P ] _ {k} [ \\text {P R E F I X} ] _ {1} \\dots [ \\text {P R E F I X} ] _ {m} [ C ] _ {1} \\dots [ C ] _ {s} \\tag {3}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.789, + 0.483, + 0.864 + ], + "angle": 0, + "content": "where \\( k \\), \\( m \\) and \\( s \\) are the token numbers of pathology context tokens, prefix words and classname words, respectively. Each token is a vector with the same dimension \\( D \\) as word embeddings of the LLM (i.e., 768 for BERT). In this way, we can design input tokens for each risk band:" + }, + { + "type": "equation", + "bbox": [ + 0.09, + 0.87, + 0.493, + 0.902 + ], + "angle": 0, + "content": "\\[\n\\boldsymbol {T} _ {p} ^ {(j)} = [ P ] _ {1} ^ {(j)} \\dots [ P ] _ {k} ^ {(j)} [ \\text {P R E F I X} ] _ {1} \\dots [ \\text {P R E F I X} ] _ {m} [ C ] _ {1} ^ {(j)} \\dots [ C ] _ {s} ^ {(j)} \\tag {4}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.643, + 0.907, + 0.709 + ], + "angle": 0, + "content": "in which \\( j \\in [1, 2I_t] \\) and prefix tokens are shared across all risk bands. By forwarding \\( T_p^{(j)} \\) of each class to the text encoder \\( \\mathcal{T}(\\cdot) \\) (i.e., LLM), we can obtain a global textual class representation \\( t_p^{(j)} \\in \\mathbb{R}^D \\) for every class." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.715, + 0.907, + 0.791 + ], + "angle": 0, + "content": "In the MIL setting, to obtain the predicted probability for each class, we can use topK patches to produce the slide-level prediction based on similarity scores. To this end, we first compute the similarity score between every instance and every class:" + }, + { + "type": "equation", + "bbox": [ + 0.652, + 0.81, + 0.907, + 0.832 + ], + "angle": 0, + "content": "\\[\n\\boldsymbol {s} _ {n, p} ^ {(i, j)} = \\boldsymbol {p} _ {n} ^ {i} \\cdot \\boldsymbol {t} _ {p} ^ {(j)} \\tag {5}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.856, + 0.907, + 0.903 + ], + "angle": 0, + "content": "As a result, we can get a similarity matrix \\(S_{n,p} \\in \\mathbb{R}^{M_p \\times 2I_t}\\) for a WSI. Then, we apply a TopK max-pooling operator \\(h(\\cdot)\\) to obtain the slide-level predictions, depend" + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.516, + 0.957 + ], + "angle": 0, + "content": "5105" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.091, + 0.092, + 0.355, + 0.107 + ], + "angle": 0, + "content": "ing on top-K similarities for each class:" + }, + { + "type": "equation", + "bbox": [ + 0.107, + 0.113, + 0.482, + 0.167 + ], + "angle": 0, + "content": "\\[\nh \\left(\\mathcal {S} _ {n, p}\\right) = \\frac {1}{K} \\left[ \\sum_ {i = 1} ^ {K} \\hat {\\boldsymbol {s}} _ {n, p} ^ {(i, 1)}, \\sum_ {i = 1} ^ {K} \\hat {\\boldsymbol {s}} _ {n, p} ^ {(i, 2)}, \\dots , \\sum_ {i = 1} ^ {K} \\hat {\\boldsymbol {s}} _ {n, p} ^ {(i, 2 \\boldsymbol {I} _ {t})} \\right] \\tag {6}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.169, + 0.483, + 0.214 + ], + "angle": 0, + "content": "where \\(\\hat{s}_{n,p}\\) is the sorted \\(s_{n,p}\\) in a descending order across all instances. Next, we can get a set of hazard probabilities \\(\\mathbf{h}_n^p\\) across all time intervals through a sigmoid function." + }, + { + "type": "equation", + "bbox": [ + 0.204, + 0.221, + 0.483, + 0.24 + ], + "angle": 0, + "content": "\\[\n\\mathbf {h} _ {n} ^ {p} = \\operatorname {S i g m o i d} \\left(h \\left(S _ {n, p}\\right)\\right) \\tag {7}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.245, + 0.483, + 0.291 + ], + "angle": 0, + "content": "By computing the survival loss according to Eq. 2 and backforwarding gradients, we can optimize the modality adapter and learnable prompts consisting of context tokens." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.291, + 0.484, + 0.368 + ], + "angle": 0, + "content": "Similarly, we acquire the set of hazard probabilities \\(\\mathbf{h}_n^g\\) for genomic pathways and optimize its learnable prompts \\(T_{g}\\) and gene encoder \\(f_{g}(\\cdot)\\). Here we use the subscript \\(p\\) and \\(g\\) to represent pathology and genomic data. Every pathway is also encoded into a \\(D\\)-dimensional embedding." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.374, + 0.462, + 0.391 + ], + "angle": 0, + "content": "3.3. Stage 2: Multimodal Prompting (MultiPro)" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.396, + 0.484, + 0.456 + ], + "angle": 0, + "content": "In this stage, we aim to leverage LLMs' strong inference ability and robustness to missing modalities to predict the representation of missing modalities, as shown in the MultiPro of Fig. 2." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.457, + 0.484, + 0.607 + ], + "angle": 0, + "content": "To simplify explanations, we take the example input of pathology available and genomics being missing. By prompting the LLM with available pathology data, the Bert-style LLM can capture modality-common knowledge for genomics due to its nature of self-attention. To compensate for modality-specific knowledge, we re-use the unimodal prompts \\( T_{g} \\) learned in stage 1 with all remaining samples available of genomics data to supervise the learning of imputed representations, where unimodal prompts characterize the knowledge distribution of an individual modality." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.608, + 0.484, + 0.781 + ], + "angle": 0, + "content": "To this end, we first need to construct an interleaved input for the learning of modality-common information. To fully leverage the capabilities of LLMs, every modality are supposed to be embedded into the aligned text space. Consequently, we re-use the aligned modality adapter or encoder learned in the first stage to embed each modality, and we continue to finetune them in the second stage. Note that the LLM used in stage 2 is the same as the one in stage 1. As a result, we can get a bag of aligned feature tokens \\(\\{\\mathbf{p}_n^i\\}_{i = 1}^{M_p}\\) or \\(\\{\\mathbf{g}_n^i\\}_{i = 1}^{M_g}\\) for pathology and genomic data, respectively. Here we assume genomic data \\(\\{\\tilde{\\mathbf{g}}_n^i\\}_{i = 1}^{M_g}\\) is missing." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.781, + 0.484, + 0.901 + ], + "angle": 0, + "content": "Due to the limitation of input length for general LLMs (e.g., 512 for Bert-style LLMs) and high information redundancy in WSIs or genomic pathways [30, 42], more discriminative and relevant feature tokens are expected to be selected for better capturing modality-common knowledge. Therefore, we propose a UniPro Scoring module by re-using UniPro as a grader to estimate a relevance score for each feature token." + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.092, + 0.905, + 0.137 + ], + "angle": 0, + "content": "UniPro Scoring. The score of every feature token is determined by three parts: relevance scores estimated by UniPro-P and UniPro-G, and self-scoring, as shown in Fig. 2." + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.138, + 0.907, + 0.395 + ], + "angle": 0, + "content": "Given a set of pathology feature tokens \\(\\{\\mathbf{p}_n^i\\}_{i = 1}^{M_p}\\), to obtain discriminative tokens, we compute the similarity score \\(s_{n,p}^{(i,\\tau)}\\) between the ground-truth textual class representation of \\(t_p^{(\\tau)}\\) and each feature token \\(\\mathbf{p}_n^i\\) by re-using Eq. 5. Through the sigmoid function, we can get the score \\(\\mathbf{s}_{n,p}^{(i,\\tau)}\\) estimated by UniPro-P. To capture cross-modal modality-common information for better compensation of missing modality, we similarly obtain the score \\(\\mathbf{s}_{n,g}^{(i,\\tau)}\\) estimated by UniPro-G. Since textual class representations of UniPro are fixed once finishing the training of stage 1, the selected tokens will also become fixed accordingly. To make it diverse while dynamically adapting to the current input, we employ a learnable self-scoring mechanism implemented by attention layers [13] to assign a score \\(\\mathbf{a}_{n,p}^{(i)}\\) for each feature token. In the end, the final score \\(\\mathbf{s}_{n,\\#}^{(i)}\\) for each feature token can be computed by their summarization:" + }, + { + "type": "equation", + "bbox": [ + 0.613, + 0.403, + 0.905, + 0.425 + ], + "angle": 0, + "content": "\\[\n\\mathbf {s} _ {n, \\#} ^ {(i)} = \\mathbf {s} _ {n, p} ^ {(i, \\tau)} + \\mathbf {s} _ {n, g} ^ {(i, \\tau)} + \\mathbf {a} _ {n, p} ^ {(i)} \\tag {8}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.426, + 0.906, + 0.487 + ], + "angle": 0, + "content": "For the inference phase when the ground-truth class cannot be accessed, we compute similarity scores with all textual class representations and apply Top-K max-pooling to determine which class \\( j \\) these tokens belong to:" + }, + { + "type": "equation", + "bbox": [ + 0.601, + 0.495, + 0.905, + 0.521 + ], + "angle": 0, + "content": "\\[\n\\tau = \\arg \\max _ {j \\in [ 1, 2 I _ {t} ]} h \\left(\\mathcal {S} _ {n, p} ^ {(j)} \\oplus \\mathcal {S} _ {n, g} ^ {(j)}\\right) \\tag {9}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.524, + 0.905, + 0.554 + ], + "angle": 0, + "content": "where \\(\\oplus\\) refers to the element-wise addition and \\(h(\\cdot)\\) can be computed by Eq. 6." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.554, + 0.906, + 0.74 + ], + "angle": 0, + "content": "By selecting feature tokens with the top- \\( K_{p} \\) largest scores, we can get the input tokens \\( \\{\\mathbf{p}_n^i\\}_{i=1}^{K_p} \\) for available pathology modality. Then, we employ a set of learnable tokens to act as placeholders \\( [H_p]_{1\\dots K_p} \\) and \\( [H_g]_{1\\dots K_g} \\) for pathology and genomic data, where \\( K_{p} \\) and \\( K_{g} \\) are their token number of the input for the LLM in stage 2, respectively. They are initialized with modality indicators. For the available pathology, we add the selected tokens to its placeholders, leading to the input of pathology \\( \\{\\hat{\\mathbf{p}}_n^i\\}_{i=1}^{K_p} \\). At last, we concatenate them with placeholders of the missing genomics as well as a [CLS] token, resulting in the final input being forwarded into a LLM \\( \\mathcal{T}(\\cdot) \\):" + }, + { + "type": "equation", + "bbox": [ + 0.522, + 0.748, + 0.905, + 0.778 + ], + "angle": 0, + "content": "\\[\n[ C L S ] [ \\hat {\\mathbf {p}} _ {n} ] _ {1 \\dots K _ {p}} [ \\tilde {\\mathbf {g}} _ {n} ] _ {1 \\dots K _ {g}} = \\mathcal {T} ([ \\mathrm {C L S} ] | | \\{\\hat {\\mathbf {p}} _ {n} ^ {i} \\} _ {i = 1} ^ {K _ {p}} | | [ H _ {g} ] _ {1 \\dots K _ {g}}) \\tag {10}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.779, + 0.906, + 0.853 + ], + "angle": 0, + "content": "UniPro Distillation. To inject modality-specific knowledge for the compensation of missing modality, we utilize UniPro to supervise this process, where UniPro has learned the knowledge distribution regarding the missing modality based on other observed samples in stage 1." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.854, + 0.905, + 0.903 + ], + "angle": 0, + "content": "Given the output (Eq. 10) of the LLM in stage 2, we take the part of the missing modality \\([\\tilde{\\mathbf{g}}_n]_{1\\dots K_g}\\) and compute similarity scores \\(\\tilde{\\mathbf{h}}_g\\) between \\([\\tilde{\\mathbf{g}}_n]_{1\\dots K_g}\\) and the textual class" + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.516, + 0.957 + ], + "angle": 0, + "content": "5106" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.09, + 0.091, + 0.486, + 0.138 + ], + "angle": 0, + "content": "representation by Eq. 5 - 7. By substituting \\([\\tilde{\\mathbf{g}}_n]_{1\\ldots K_g}\\) into the survival loss function of Eq. 2, we can supervise the learning of compensation for the missing modality:" + }, + { + "type": "equation", + "bbox": [ + 0.113, + 0.149, + 0.483, + 0.217 + ], + "angle": 0, + "content": "\\[\n\\begin{array}{l} \\mathcal {L} _ {u d} ^ {g} = - \\sum_ {n = 1} ^ {\\tilde {N} _ {g}} \\left[ c _ {n} \\log S _ {n} ^ {(\\tau)} + \\left(1 - c _ {n}\\right) \\log S _ {n} ^ {(\\tau - 1)} \\right. \\tag {11} \\\\ \\left. + \\left(1 - c _ {n}\\right) \\log \\tilde {\\mathbf {h}} _ {n} ^ {g, (\\tau)} \\right] \\\\ \\end{array}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.218, + 0.485, + 0.31 + ], + "angle": 0, + "content": "where \\(\\tilde{N}_g\\) is the number of missing genomic modalities across all data pairs. Similarly, we can get the loss function \\(\\mathcal{L}_{ud}^p\\) when the pathology modality is missing. Lastly, [CLS] token is used for calculating the loss function for survival prediction \\(\\mathcal{L}_{surv}^{cls}\\) in stage 2, and the final loss can be combined linearly with coefficient factors \\(\\alpha_{1}\\) and \\(\\alpha_{2}\\):" + }, + { + "type": "equation", + "bbox": [ + 0.184, + 0.32, + 0.483, + 0.34 + ], + "angle": 0, + "content": "\\[\n\\mathcal {L} = \\mathcal {L} _ {s u r v} ^ {c l s} + \\alpha_ {1} \\mathcal {L} _ {u d} ^ {p} + \\alpha_ {2} \\mathcal {L} _ {u d} ^ {g} \\tag {12}\n\\]" + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.351, + 0.224, + 0.367 + ], + "angle": 0, + "content": "4. Experiments" + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.375, + 0.292, + 0.391 + ], + "angle": 0, + "content": "4.1. Datasets and Settings" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.397, + 0.484, + 0.532 + ], + "angle": 0, + "content": "Datasets. To demonstrate the efficacy of the proposed method, we conducted a series of experiments across five public cancer datasets sourced from The Cancer Genome Atlas (TCGA) \\(^{1}\\). The case number of each dataset is demonstrated next to the dataset name in Tab. 1. Regarding genomic data, we follow the grouping of biological pathways provided by [14] and remove genes absent in the cBioPortal \\(^{2}\\) database, resulting in 330 pathways. More details of datasets can be found in supplementary materials." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.533, + 0.484, + 0.805 + ], + "angle": 0, + "content": "Evaluation Settings on Incomplete Modalities. Following the common setting in previous works [22, 32], to systematically assess performances on various missing settings, we first manually construct a series of missing scenarios under different missing ratios by randomly dropping them for training sets on samples with complete modality pairs. Specifically, we set a total missing rate (60%) for two modalities and then construct 5 combinations for pathology and genomics, i.e., \\(\\{(0\\%, 60\\%), (20\\%, 40\\%), (30\\%, 30\\%), (40\\%, 20\\%), (60\\%, 0\\%)\\}\\). For inference, all methods are evaluated on three scenarios: pathology only, genomics only, and complete modalities. All experiments are conducted on 5-fold cross-validation following previous works [5, 36] in survival prediction, and the concordance index (C-Index) and its standard deviation (std) are reported to evaluate performances. The LLM is implemented by BioBERT-v1.2 [16]. The supplementary materials contain more experiment results and details on the implementation." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.814, + 0.327, + 0.83 + ], + "angle": 0, + "content": "4.2. Comparisons with SOTAs" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.836, + 0.483, + 0.866 + ], + "angle": 0, + "content": "We first present unimodal approaches, and SOTA multimodal methods trained with complete modalities (all" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.092, + 0.905, + 0.228 + ], + "angle": 0, + "content": "pathology and genomics data) to serve as the upper bound of current complete modality approaches. Then the average performances on all missing combinations are demonstrated under the total training missing rate of \\(60\\%\\), evaluated on three inference scenarios. Details on various combinations are shown in Fig. 3 and we discuss them in Sec. 4.3. Lastly, we investigate the upper bound of each model tailored to incomplete modalities by training them with all data (without missing data)." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.229, + 0.907, + 0.606 + ], + "angle": 0, + "content": "From results in Tab. 1, we observed that (1) When applying \\(60\\%\\) training missing rate, DisPro consistently outperformed SOTAs across five datasets by a large margin (on average, \\(+1.76\\%\\) for pathology only, \\(+1.57\\%\\) for genomics only and \\(2.33\\%\\) for complete data). This indicates the superiority of DisPro over the incomplete SOTA methods. (2) When testing on the complete scenario, on 3 out of 5 datasets and overall, DisPro even surpassed the SOTA multimodal survival model (SurvPath) trained on all data. Notably, DisPro achieved this while only utilizing incomplete training data with a \\(60\\%\\) missing rate. It suggests that harnessing the powerful capabilities of LLMs is a promising solution for addressing the classical question of survival analysis. (3) When using complete training data, DisPro consistently performed best on average for various test scenarios as well (\\(2.28\\%\\) on pathology only, \\(1.33\\%\\) on genomics only and \\(2.57\\%\\) on complete test data). Under the same condition, DisPro further increased by \\(1.31\\%\\) compared with the SOTA complete multimodal method, further demonstrating the strong power of LLMs. (4) In contrast to unimodal methods trained on full unimodal datasets, DisPro, using only incomplete multimodal data with a \\(60\\%\\) missing rate, achieved superior or similar performance when inferring from individual modalities. This showcases DisPro's efficient utilization of multimodal data." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.621, + 0.667, + 0.637 + ], + "angle": 0, + "content": "4.3. Ablation Study" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.644, + 0.907, + 0.901 + ], + "angle": 0, + "content": "Component Validation. In Tab. 2, we ablate each component in our method to investigate their effectiveness. (1) To demonstrate how the LLM performs on missing modalities without any prompts, we establish a baseline by employing a linear classifier head on the same LLM and randomly selecting feature tokens for its input. Results show that the performance significantly decreased by \\(2.62\\%\\) on pathology only, \\(1.55\\%\\) on genomics only and \\(2.05\\%\\) on complete test data. This indicates the importance of appropriate prompts for LLMs. (2) When supplementing modality-specific information for missing modalities, the performance \\((+UD)\\) consistently improves on all datasets and increases by around \\(+1\\%\\) on various inference settings, which validates the effectiveness of compensating for modality-specific knowledge of the missing modality. (3) When reusing UniPro to select relevant tokens \\((+US)\\), the performance increased by about \\(+1\\%\\) overall, which bene-" + }, + { + "type": "page_footnote", + "bbox": [ + 0.108, + 0.874, + 0.354, + 0.887 + ], + "angle": 0, + "content": "1https://portal.qdc.cancer.gov/" + }, + { + "type": "page_footnote", + "bbox": [ + 0.108, + 0.888, + 0.332, + 0.901 + ], + "angle": 0, + "content": "2https://www.cbiportal.org/" + }, + { + "type": "list", + "bbox": [ + 0.108, + 0.874, + 0.354, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.516, + 0.957 + ], + "angle": 0, + "content": "5107" + } + ], + [ + { + "type": "table", + "bbox": [ + 0.111, + 0.088, + 0.889, + 0.72 + ], + "angle": 0, + "content": "
TrainingMiRateTestP GMethodsBLCA(372)BRCA(1007)COADREAD(533)LUAD(443)UCEC(478)Avg
N/A†TransMIL [28]0.5903±0.0250.6671±0.0540.6093±0.0180.6195±0.0350.6978±0.0460.6368
†CoOp_BioBert [44]0.6014±0.0470.6784±0.0320.6681±0.0260.6543±0.0520.6791±0.0850.6563
†SNN [15]0.6831±0.0400.6567±0.0380.6312±0.0580.6379±0.0710.7267±0.0730.6671
†CoOp_BioBert [44]0.6712±0.0360.6922±0.0350.6743±0.0480.6514±0.0650.7171±0.0600.6812
‡MOTCat [36]0.6274±0.0460.6721±0.0380.6501±0.0350.6753±0.0430.7207±0.0740.6691
‡SurvPath [14]0.6572±0.0270.7071±0.0390.7079±0.0480.6798±0.0490.7391±0.0650.6982
60%•○COM [27]0.6024±0.0130.6740±0.0110.6779±0.0140.6336±0.0110.6985±0.0120.6573
•○M3Care [40]0.6208±0.0140.6690±0.0180.6570±0.0230.6219±0.0070.7031±0.0120.6543
•○HGCN [9]0.5999±0.0090.6831±0.0120.6792±0.0090.6465±0.0050.7011±0.0070.6620
•○MUSE [34]0.6207±0.0130.6514±0.0140.6547±0.0090.6354±0.0120.7048±0.0230.6534
•○MAP [17]0.5918±0.0120.6278±0.0060.5965±0.0170.6489±0.0200.6929±0.0210.6316
•○DisPro (Ours)0.6319±0.0140.6895±0.0110.6880±0.0100.6612±0.0140.7272±0.0180.6796
○•COM [27]0.6415±0.0160.6653±0.0180.6593±0.0300.6456±0.0080.7105±0.0070.6645
○•M3Care [40]0.6420±0.0150.6642±0.0290.6491±0.0080.6462±0.0120.7032±0.0130.6609
○•HGCN [9]0.6225±0.0080.6729±0.0090.6752±0.0150.6418±0.0150.7015±0.0100.6628
○•MUSE [34]0.6299±0.0120.6427±0.0250.6391±0.0230.6462±0.0100.6893±0.0270.6494
○•MAP [17]0.5937±0.0160.6269±0.0060.5971±0.0170.6442±0.0090.6934±0.0180.6311
○•DisPro (Ours)0.6547±0.0120.6841±0.0180.6804±0.0240.6548±0.0120.7271±0.0170.6802
••COM [27]0.6192±0.0130.6846±0.0090.6874±0.0050.6548±0.0120.7136±0.0160.6719
••M3Care [40]0.6513±0.0050.6951±0.0240.6698±0.0160.6491±0.0130.7288±0.0110.6788
••HGCN [9]0.6332±0.0060.6907±0.0120.6861±0.0170.6545±0.0110.7077±0.0080.6745
••MUSE [34]0.6207±0.0120.6555±0.0130.6539±0.0120.6415±0.0080.7164±0.0130.6576
••MAP [17]0.5906±0.0130.6277±0.0040.5990±0.0160.6625±0.0080.6901±0.0180.6340
••DisPro (Ours)0.6638±0.0060.7219±0.0150.7029±0.0160.6741±0.0070.7476±0.0200.7021
0%••COM [27]0.6154±0.0410.6871±0.0270.6925±0.0440.6427±0.0460.6806±0.0610.6637
••M3Care [40]0.6268±0.0240.6775±0.0160.6850±0.0520.6299±0.0380.6964±0.0610.6631
••HGCN [9]0.6055±0.0480.6716±0.0220.6734±0.0480.6421±0.0520.6915±0.0530.6568
••MUSE [34]0.6208±0.0260.6767±0.0400.6710±0.0410.6253±0.0410.7099±0.0630.6607
••MAP [17]0.6205±0.0330.6300±0.0530.5930±0.0650.6662±0.0550.6925±0.0640.6404
••DisPro (Ours)0.6315±0.0280.6945±0.0440.7216±0.0280.6747±0.0520.7104±0.0560.6865
○•COM [27]0.6681±0.0300.6966±0.0470.6770±0.0240.6454±0.0500.7095±0.0590.6793
○•M3Care [40]0.6706±0.0390.6611±0.0400.6563±0.0210.6512±0.0610.7047±0.0600.6688
○•HGCN [9]0.6233±0.0330.6807±0.0230.6777±0.0370.6421±0.0450.7091±0.0430.6666
○•MUSE [34]0.6256±0.0290.6401±0.0240.6208±0.0300.6547±0.0750.6824±0.0390.6447
○•MAP [17]0.6093±0.0260.6305±0.0510.5842±0.0650.6653±0.0450.6841±0.0640.6347
○•DisPro (Ours)0.6745±0.0390.6996±0.0280.7000±0.0210.6751±0.0680.7140±0.0600.6926
••COM [27]0.6376±0.0430.7087±0.0240.6990±0.0450.6659±0.0340.7112±0.0510.6845
••M3Care [40]0.6605±0.0230.6808±0.0410.6768±0.0260.6740±0.0460.7359±0.0330.6856
••HGCN [9]0.6365±0.0210.7058±0.0330.7136±0.0250.6692±0.0500.7017±0.0440.6854
••MUSE [34]0.6282±0.0390.7017±0.0200.6693±0.0420.6248±0.0400.7237±0.0620.6695
••MAP [17]0.6034±0.0260.6328±0.0490.5895±0.0630.6904±0.0470.6827±0.0650.6398
••DisPro (Ours)0.6719±0.0510.7358±0.0510.7553±0.0210.6767±0.0410.7169±0.0710.7113
" + }, + { + "type": "table_caption", + "bbox": [ + 0.09, + 0.728, + 0.907, + 0.771 + ], + "angle": 0, + "content": "Table 1. C-Index of Comparison to SOTA models. \"MiRate\" represents Missing Rate. \"P\" and \"G\" indicate Pathology and Genomics, respectively. \\( \\circ \\) and \\( \\bullet \\) mean the missing modality and the available modality,respectively. † and ‡ indicate using unimodal and complete multimodal data, respectively, while the rest are for missing modalities. The best results are in bold. The second-best results are underlined." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.787, + 0.483, + 0.817 + ], + "angle": 0, + "content": "fits from the assistance of US in learning modality-common knowledge from WSIs and genes with high redundancy." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.826, + 0.483, + 0.903 + ], + "angle": 0, + "content": "Various Missing Scenarios. To illustrate how incomplete multimodal approaches handle varying missing settings, we display different combinations of training missing rates for each inference scenario, as shown in Fig. 3. Results showcase that DisPro achieved considerably consistent perfor" + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.787, + 0.907, + 0.817 + ], + "angle": 0, + "content": "mance superiority in various missing settings over other incomplete multimodal approaches by a clear margin." + }, + { + "type": "title", + "bbox": [ + 0.514, + 0.83, + 0.634, + 0.847 + ], + "angle": 0, + "content": "5. Conclusion" + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.856, + 0.907, + 0.903 + ], + "angle": 0, + "content": "This work proposed a Distilled Prompt Learning (DisPro) framework to unlock the potential of LLMs for incomplete multimodal learning by two-stage prompting, in which" + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.946, + 0.516, + 0.957 + ], + "angle": 0, + "content": "5108" + } + ], + [ + { + "type": "table", + "bbox": [ + 0.11, + 0.089, + 0.895, + 0.311 + ], + "angle": 0, + "content": "
VariantsTestModulesBLCA (372)BRCA (1007)COADREAD (533)LUAD (443)UCEC (478)Avg
PGUSUD
baseline•○0.6068±0.0210.6668±0.0060.6583±0.0120.6270±0.0180.7081±0.0190.6534
+ UD•○✓0.6213±0.0100.6829±0.0080.6740±0.0220.6327±0.0180.7152±0.0110.6652
+ US•○✓0.6285±0.0210.6884±0.0160.6653±0.0260.6546±0.0130.7203±0.0070.6714
DisPro (full)•○✓✓0.6319±0.0140.6895±0.0110.6880±0.0100.6612±0.0140.7272±0.0180.6796
baseline○•0.6435±0.0160.6605±0.0330.6601±0.0180.6467±0.0160.7127±0.0070.6647
+ UD○•✓0.6473±0.0170.6836±0.0140.6707±0.0220.6420±0.0200.7226±0.0130.6732
+ US○•✓0.6452±0.0200.6666±0.0170.6763±0.0230.6471±0.0230.7151±0.0090.6701
DisPro (full)○•✓✓0.6547±0.0120.6841±0.0180.6804±0.0240.6548±0.0120.7271±0.0170.6802
baseline••0.6522±0.0100.7075±0.0120.6785±0.0200.6541±0.0090.7157±0.0030.6816
+ UD••✓0.6585±0.0070.7194±0.0130.6963±0.0090.6635±0.0140.7250±0.0120.6926
+ US••✓0.6593±0.0080.7125±0.0100.7005±0.0180.6717±0.0070.7274±0.0100.6943
DisPro (full)••✓✓0.6638±0.0060.7219±0.0150.7029±0.0160.6741±0.0070.7476±0.0200.7021
" + }, + { + "type": "table_caption", + "bbox": [ + 0.09, + 0.322, + 0.907, + 0.363 + ], + "angle": 0, + "content": "Table 2. C-Index of Ablation Study on the Setting of \\(60\\%\\) Training Missing Rate.\"US\" and \"UD\" refer to the UniPro Scoring and UniPro Distillation modules, respectively. \"P\" and \"G\" indicate Pathology and Genomics, respectively. \\(\\circ\\) means the modality is missing, while \\(\\bullet\\) refers to the available modality." + }, + { + "type": "image", + "bbox": [ + 0.118, + 0.364, + 0.88, + 0.715 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.178, + 0.728, + 0.816, + 0.743 + ], + "angle": 0, + "content": "Figure 3. Performance on various combinations under \\(60\\%\\) training missing rate for different test scenarios." + }, + { + "type": "text", + "bbox": [ + 0.089, + 0.761, + 0.482, + 0.897 + ], + "angle": 0, + "content": "Unimodal Prompting (UniPro) and Multimodal Prompting (MultiPro) were proposed for supplementing modality-specific and modality-common knowledge for missing modalities. In the first stage, UniPro extended CoOp into the MIL setting and distilled knowledge distribution to a set of learnable prompts for each modality, which prepped for compensation of modality-specific knowledge. In the second stage, considering available modalities as prompts, we leveraged an LLM to impute common knowledge for miss" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.761, + 0.907, + 0.883 + ], + "angle": 0, + "content": "ing modalities. To inject the specific knowledge of missing modalities, the distilled prompts in UniPro are used to supervise the learning of imputing representations for missing modalities, leading to more comprehensive knowledge compensation in incomplete multimodal learning. Extensive experiments on five cancer benchmark datasets demonstrated the effectiveness of the proposed method over state-of-the-art approaches." + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.516, + 0.957 + ], + "angle": 0, + "content": "5109" + } + ], + [ + { + "type": "title", + "bbox": [ + 0.092, + 0.091, + 0.251, + 0.108 + ], + "angle": 0, + "content": "Acknowledgement" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.114, + 0.484, + 0.201 + ], + "angle": 0, + "content": "This work was supported by the National Natural Science Foundation of China (No. 62202403), Hong Kong Innovation and Technology Commission (Project No. MHP/002/22 and ITCPD/17-9), and Research Grants Council of the Hong Kong Special Administrative Region, China (Project No. R6003-22 and C4024-22GF)." + }, + { + "type": "title", + "bbox": [ + 0.093, + 0.226, + 0.188, + 0.241 + ], + "angle": 0, + "content": "References" + }, + { + "type": "ref_text", + "bbox": [ + 0.101, + 0.251, + 0.486, + 0.279 + ], + "angle": 0, + "content": "[1] Som S Biswas. Role of chat gpt in public health. Annals of biomedical engineering, 51(5):868-869, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.281, + 0.484, + 0.308 + ], + "angle": 0, + "content": "[2] Jason Brownlee. A gentle introduction to the rectified linear unit (relu). Machine learning mastery, 6, 2019. 3, 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.31, + 0.484, + 0.364 + ], + "angle": 0, + "content": "[3] Agisilaos Chartsias, Thomas Joyce, Mario Valerio Giuffrida, and Sotirios A Tsaftaris. Multimodal mr synthesis via modality-invariant latent representation. IEEE transactions on medical imaging, 37(3):803-814, 2017. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.366, + 0.484, + 0.434 + ], + "angle": 0, + "content": "[4] Jiayi Chen and Aidong Zhang. Hgmf: heterogeneous graph-based fusion for multimodal data with incompleteness. In Proceedings of the 26th ACM SIGKDD international conference on knowledge discovery & data mining, pages 1295-1305, 2020. 1, 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.436, + 0.484, + 0.519 + ], + "angle": 0, + "content": "[5] Richard J Chen, Ming Y Lu, Wei-Hung Weng, Tiffany Y Chen, Drew FK Williamson, Trevor Manz, Maha Shady, and Faisal Mahmood. Multimodal co-attention transformer for survival prediction in gigapixel whole slide images. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 4015–4025, 2021. 1, 3, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.521, + 0.484, + 0.602 + ], + "angle": 0, + "content": "[6] Richard J Chen, Tong Ding, Ming Y Lu, Drew FK Williamson, Guillaume Jaume, Andrew H Song, Bowen Chen, Andrew Zhang, Daniel Shao, Muhammad Shaban, et al. Towards a general-purpose foundation model for computational pathology. Nature Medicine, 30(3):850-862, 2024. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.604, + 0.484, + 0.659 + ], + "angle": 0, + "content": "[7] Florinel-Alin Croitoru, Vlad Hondru, Radu Tudor Ionescu, and Mubarak Shah. Diffusion models in vision: A survey. IEEE Transactions on Pattern Analysis and Machine Intelligence, 45(9):10850-10869, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.661, + 0.484, + 0.728 + ], + "angle": 0, + "content": "[8] Yuhang Ding, Xin Yu, and Yi Yang. Rfnet: Region-aware fusion network for incomplete multi-modal brain tumor segmentation. In Proceedings of the IEEE/CVF international conference on computer vision, pages 3975-3984, 2021. 1, 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.732, + 0.484, + 0.773 + ], + "angle": 0, + "content": "[9] Hou et al. Hybrid graph convolutional network with online masked autoencoder for robust multimodal cancer survival prediction. IEEE Transactions on Medical Imaging, 2023. 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.775, + 0.484, + 0.802 + ], + "angle": 0, + "content": "[10] Zhou et al. Cross-modal translation and alignment for survival analysis. In ICCV, pages 21485-21494, 2023. 1, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.804, + 0.484, + 0.858 + ], + "angle": 0, + "content": "[11] Michael Gadermayr and Maximilian Tschuchnig. Multiple instance learning for digital pathology: A review of the state-of-the-art, limitations & future potential. Computerized Medical Imaging and Graphics, page 102337, 2024. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.86, + 0.484, + 0.901 + ], + "angle": 0, + "content": "[12] Mohammad Hamghalam, Alejandro F Frangi, Baiying Lei, and Amber L Simpson. Modality completion via gaussian process prior variational autoencoders for multi-modal" + }, + { + "type": "list", + "bbox": [ + 0.094, + 0.251, + 0.486, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.548, + 0.093, + 0.907, + 0.162 + ], + "angle": 0, + "content": "glioma segmentation. In Medical Image Computing and Computer Assisted Intervention-MICCAI 2021: 24th International Conference, Strasbourg, France, September 27-October 1, 2021, Proceedings, Part VII 24, pages 442-452. Springer, 2021. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.164, + 0.906, + 0.217 + ], + "angle": 0, + "content": "[13] Maximilian Ilse, Jakub Tomczak, and Max Welling. Attention-based deep multiple instance learning. In International conference on machine learning, pages 2127-2136. PMLR, 2018. 2, 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.219, + 0.908, + 0.302 + ], + "angle": 0, + "content": "[14] Guillaume Jaume, Anurag Vaidya, Richard J Chen, Drew FK Williamson, Paul Pu Liang, and Faisal Mahmood. Modeling dense multimodal interactions between biological pathways and histology for survival prediction. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 11579-11590, 2024. 3, 6, 7, 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.304, + 0.906, + 0.357 + ], + "angle": 0, + "content": "[15] Günter Klambauer, Thomas Unterthiner, Andreas Mayr, and Sepp Hochreiter. Self-normalizing neural networks. Advances in neural information processing systems, 30, 2017. 7, 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.359, + 0.906, + 0.425 + ], + "angle": 0, + "content": "[16] Jinhyuk Lee, Wonjin Yoon, Sungdong Kim, Donghyeon Kim, Sunkyu Kim, Chan Ho So, and Jaewoo Kang. Biobert: a pre-trained biomedical language representation model for biomedical text mining. Bioinformatics, 36(4):1234-1240, 2020. 6, 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.428, + 0.906, + 0.496 + ], + "angle": 0, + "content": "[17] Yi-Lun Lee, Yi-Hsuan Tsai, Wei-Chen Chiu, and Chen-Yu Lee. Multimodal prompting with missing modalities for visual recognition. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 14943-14952, 2023. 2, 7, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.498, + 0.906, + 0.567 + ], + "angle": 0, + "content": "[18] Bin Li, Yin Li, and Kevin W Eliceiri. Dual-stream multiple instance learning network for whole slide image classification with self-supervised contrastive learning. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 14318-14328, 2021. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.568, + 0.906, + 0.636 + ], + "angle": 0, + "content": "[19] Chenxin Li, Xinyu Liu, Cheng Wang, Yifan Liu, Weihao Yu, Jing Shao, and Yixuan Yuan. Gtp-4o: Modality-prompted heterogeneous graph learning for omni-modal biomedical representation. In European conference on computer vision, pages 168–187. Springer, 2024. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.638, + 0.906, + 0.705 + ], + "angle": 0, + "content": "[20] Chunyuan Li, Cliff Wong, Sheng Zhang, Naoto Usuyama, Haotian Liu, Jianwei Yang, Tristan Naumann, Hoifung Poon, and Jianfeng Gao. Llava-med: Training a large language-and-vision assistant for biomedicine in one day. Advances in Neural Information Processing Systems, 36, 2024. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.707, + 0.906, + 0.803 + ], + "angle": 0, + "content": "[21] Han Liu, Yubo Fan, Hao Li, Jiacheng Wang, Dewei Hu, Can Cui, Ho Hin Lee, Huahong Zhang, and Ipek Oguz. Moddrop++: A dynamic filter network with intra-subject cotraining for multiple sclerosis lesion segmentation with missing modalities. In International Conference on Medical Image Computing and Computer-Assisted Intervention, pages 444-453. Springer, 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.805, + 0.906, + 0.871 + ], + "angle": 0, + "content": "[22] Mengmeng Ma, Jian Ren, Long Zhao, Sergey Tulyakov, Cathy Wu, and Xi Peng. Smil: Multimodal learning with severely missing modality. In Proceedings of the AAAI Conference on Artificial Intelligence, pages 2302-2310, 2021. 2, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.874, + 0.906, + 0.901 + ], + "angle": 0, + "content": "[23] Xiangxi Meng, Kaicong Sun, Jun Xu, Xuming He, and Dinggang Shen. Multi-modal modality-masked diffusion network" + }, + { + "type": "list", + "bbox": [ + 0.517, + 0.093, + 0.908, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.517, + 0.957 + ], + "angle": 0, + "content": "5110" + } + ], + [ + { + "type": "ref_text", + "bbox": [ + 0.125, + 0.092, + 0.484, + 0.12 + ], + "angle": 0, + "content": "for brain mri synthesis with random modality missing. IEEE Transactions on Medical Imaging, 2024. 1, 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.122, + 0.484, + 0.164 + ], + "angle": 0, + "content": "[24] Muhammad Khalid Khan Niazi, Anil V Parwani, and Metin N Gurcan. Digital pathology and artificial intelligence. The lancet oncology, 20(5):e253-e261, 2019. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.165, + 0.483, + 0.234 + ], + "angle": 0, + "content": "[25] Zhenyuan Ning, Denghui Du, Chao Tu, Qianjin Feng, and Yu Zhang. Relation-aware shared representation learning for cancer prognosis analysis with auxiliary clinical variables and incomplete multi-modality data. IEEE Transactions on Medical Imaging, 41(1):186-198, 2021. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.236, + 0.483, + 0.277 + ], + "angle": 0, + "content": "[26] Fatih Ozsolak and Patrice M Milos. RNA sequencing: advances, challenges and opportunities. Nature reviews genetics, 12(2):87-98, 2011. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.279, + 0.483, + 0.32 + ], + "angle": 0, + "content": "[27] Shuwei Qian and Chongjun Wang. Com: Contrastive masked-attention model for incomplete multimodal learning. Neural Networks, 162:443–455, 2023. 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.322, + 0.483, + 0.39 + ], + "angle": 0, + "content": "[28] Zhuchen Shao, Hao Bian, Yang Chen, Yifeng Wang, Jian Zhang, Xiangyang Ji, et al. Transmil: Transformer based correlated multiple instance learning for whole slide image classification. Advances in neural information processing systems, 34:2136-2147, 2021. 3, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.392, + 0.483, + 0.447 + ], + "angle": 0, + "content": "[29] Anmol Sharma and Ghassan Hamarneh. Missing mri pulse sequence synthesis using multi-modal generative adversarial network. IEEE transactions on medical imaging, 39(4): 1170-1183, 2019. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.449, + 0.483, + 0.518 + ], + "angle": 0, + "content": "[30] Andrew H Song, Richard J Chen, Guillaume Jaume, Anurag Jayant Vaidya, Alexander Baras, and Faisal Mahmood. Multimodal prototyping for cancer survival prediction. In *Forty-first International Conference on Machine Learning*, 2024. 1, 2, 3, 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.52, + 0.483, + 0.56 + ], + "angle": 0, + "content": "[31] Rory Stark, Marta Grzelak, and James Hadfield. Rna sequencing: the teenage years. Nature Reviews Genetics, 20 (11):631-656, 2019. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.563, + 0.483, + 0.632 + ], + "angle": 0, + "content": "[32] Hu Wang, Yuanhong Chen, Congbo Ma, Jodie Avery, Louise Hull, and Gustavo Carneiro. Multi-modal learning with missing modality via shared-specific feature modelling. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 15878-15887, 2023. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.634, + 0.483, + 0.702 + ], + "angle": 0, + "content": "[33] Tianyu Wu, Shizhu He, Jingping Liu, Siqi Sun, Kang Liu, Qing-Long Han, and Yang Tang. A brief overview of chatgpt: The history, status quo and potential future development. IEEE/CAA Journal of Automatica Sinica, 10(5):1122-1136, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.704, + 0.483, + 0.773 + ], + "angle": 0, + "content": "[34] Zhenbang Wu, Anant Dadu, Nicholas Tustison, Brian Avants, Mike Nalls, Jimeng Sun, and Faraz Faghri. Multimodal patient representation learning with missing modalities and labels. In The Twelfth International Conference on Learning Representations, 2024. 1, 2, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.775, + 0.483, + 0.856 + ], + "angle": 0, + "content": "[35] Conghao Xiong, Hao Chen, Hao Zheng, Dong Wei, Yefeng Zheng, Joseph JY Sung, and Irwin King. Mome: Mixture of multimodal experts for cancer survival prediction. In International Conference on Medical Image Computing and Computer-Assisted Intervention, pages 318-328. Springer, 2024. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.859, + 0.483, + 0.901 + ], + "angle": 0, + "content": "[36] Yingxue Xu and Hao Chen. Multimodal optimal transport-based co-attention transformer with global structure consistency for survival prediction. In Proceedings of the" + }, + { + "type": "list", + "bbox": [ + 0.093, + 0.092, + 0.484, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.548, + 0.093, + 0.905, + 0.12 + ], + "angle": 0, + "content": "IEEE/CVF International Conference on Computer Vision, pages 21241-21251, 2023. 1, 3, 6, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.122, + 0.905, + 0.191 + ], + "angle": 0, + "content": "[37] Shu Yang, Yihui Wang, and Hao Chen. Mambamil: Enhancing long sequence modeling with sequence reordering in computational pathology. In International Conference on Medical Image Computing and Computer-Assisted Intervention, pages 296-306. Springer, 2024. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.192, + 0.905, + 0.261 + ], + "angle": 0, + "content": "[38] Yixiao Yuan, Yawen Huang, and Yi Zhou. Rethinking a unified generative adversarial model for mri modality completion. In International Conference on Medical Image Computing and Computer-Assisted Intervention, pages 143-153. Springer, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.263, + 0.905, + 0.317 + ], + "angle": 0, + "content": "[39] Shekoufeh Gorgi Zadeh and Matthias Schmid. Bias in cross-entropy-based training of deep survival networks. IEEE transactions on pattern analysis and machine intelligence, 43(9):3126-3137, 2020. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.319, + 0.905, + 0.401 + ], + "angle": 0, + "content": "[40] Chaohe Zhang, Xu Chu, Liantao Ma, Yinghao Zhu, Yasha Wang, Jiangtao Wang, and Junfeng Zhao. M3care: Learning with missing modalities in multimodal healthcare data. In Proceedings of the 28th ACM SIGKDD Conference on Knowledge Discovery and Data Mining, pages 2418-2428, 2022. 1, 2, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.403, + 0.905, + 0.487 + ], + "angle": 0, + "content": "[41] Hongrun Zhang, Yanda Meng, Yitian Zhao, Yihong Qiao, Xiaoyun Yang, Sarah E Coupland, and Yalin Zheng. Dtfmil: Double-tier feature distillation multiple instance learning for histopathology whole slide image classification. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 18802-18812, 2022. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.488, + 0.905, + 0.557 + ], + "angle": 0, + "content": "[42] Yilan Zhang, Yingxue Xu, Jianqi Chen, Fengying Xie, and Hao Chen. Prototypical information bottlenecking and disentangling for multimodal cancer survival prediction. In The Twelfth International Conference on Learning Representations, 2024. 1, 2, 3, 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.558, + 0.905, + 0.613 + ], + "angle": 0, + "content": "[43] Huajun Zhou, Fengtao Zhou, Chenyu Zhao, Yingxue Xu, Luyang Luo, and Hao Chen. Multimodal data integration for precision oncology: Challenges and future directions. arXiv preprint arXiv:2406.19611, 2024. 1, 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.615, + 0.905, + 0.669 + ], + "angle": 0, + "content": "[44] Kaiyang Zhou, Jingkang Yang, Chen Change Loy, and Ziwei Liu. Learning to prompt for vision-language models. International Journal of Computer Vision, 130(9):2337-2348, 2022. 2, 3, 7" + }, + { + "type": "list", + "bbox": [ + 0.517, + 0.093, + 0.905, + 0.669 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.514, + 0.956 + ], + "angle": 0, + "content": "5111" + } + ] +] \ No newline at end of file diff --git a/2025/Distilled Prompt Learning for Incomplete Multimodal Survival Prediction/5aae416b-ca2c-40f8-8583-f2902bfc3349_origin.pdf b/2025/Distilled Prompt Learning for Incomplete Multimodal Survival Prediction/5aae416b-ca2c-40f8-8583-f2902bfc3349_origin.pdf new file mode 100644 index 0000000000000000000000000000000000000000..cb7692610074b2aa85ccff0b5349512a0ecdad6e --- /dev/null +++ b/2025/Distilled Prompt Learning for Incomplete Multimodal Survival Prediction/5aae416b-ca2c-40f8-8583-f2902bfc3349_origin.pdf @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:7b67b08a7182852d3a8bc7a42d011b0fde790eb921bc9c8a5aa0c3b23d998ffb +size 1511948 diff --git a/2025/Distilled Prompt Learning for Incomplete Multimodal Survival Prediction/full.md b/2025/Distilled Prompt Learning for Incomplete Multimodal Survival Prediction/full.md new file mode 100644 index 0000000000000000000000000000000000000000..ac5b11e9fa3c11c9fd8e1ddd85df417a18008d92 --- /dev/null +++ b/2025/Distilled Prompt Learning for Incomplete Multimodal Survival Prediction/full.md @@ -0,0 +1,286 @@ +# Distilled Prompt Learning for Incomplete Multimodal Survival Prediction + +Yingxue Xu $^{1}$ Fengtao Zhou $^{1}$ Chenyu Zhao $^{1}$ Yihui Wang $^{1}$ Can Yang $^{2}$ Hao Chen $^{1\boxtimes}$ $^{1}$ Department of Computer Science and Engineering, $^{2}$ Department of Mathematics The Hong Kong University of Science and Technology + +{yxueb,fzhouaf,cyzhao,ywangrm}@connect.ust.hk，{macyang,jhc}@ust.hk + +# Abstract + +The integration of multimodal data including pathology images and gene profiles is widely applied in precise survival prediction. Despite recent advances in multimodal survival models, collecting complete modalities for multimodal fusion still poses a significant challenge, hindering their application in clinical settings. Current approaches tackling incomplete modalities often fall short, as they typically compensate for only a limited part of the knowledge of missing modalities. To address this issue, we propose a Distilled Prompt Learning framework (DisPro) to utilize the strong robustness of Large Language Models (LLMs) to missing modalities, which employs two-stage prompting for compensation of comprehensive information for missing modalities. In the first stage, Unimodal Prompt (UniPro) distills the knowledge distribution of each modality, preparing for supplementing modality-specific knowledge of the missing modality in the subsequent stage. In the second stage, Multimodal Prompt (MultiPro) leverages available modalities as prompts for LLMs to infer the missing modality, which provides modality-common information. Simultaneously, the unimodal knowledge acquired in the first stage is injected into multimodal inference to compensate for the modality-specific knowledge of the missing modality. Extensive experiments covering various missing scenarios demonstrated the superiority of the proposed method. The code is available at https://github.com/Innse/DisPro. + +# 1. Introduction + +Multimodal Survival Prediction typically incorporates pathology slides and genomic data complementary to each other for precise prognosis analysis. Pathology data can provide qualitative morphological insights, while quantitative molecular information can be obtained from genomic data [5, 35, 36]. Despite progress [5, 10, 36] in multimodal survival analysis, these models typically require complete modalities data for reliable performance. However, the ex + +![](images/d84997a7d97f611945b7c7e55fdbc05e48e28a07581278184108b361e3588ef3.jpg) +Figure 1. Insights for existing incomplete multimodal learning and comparison to the proposed method. (a) Generation-based Imputation and Imputation-free approaches, (b) Retrieved-based Imputation and (c) Ours. + +pectation of having full access to complete modalities for integration is not always achievable, especially in clinical scenarios, due to various factors such as the cost of data gathering and privacy issues. For example, acquiring genomic data still comes at a high cost due to the technology and infrastructure required for gene sequencing, especially in underdeveloped areas [43]. Inevitably, it is challenging to adapt previously established multimodal survival models [30, 42] to real-world clinical settings. As a result, building a multimodal survival model that is robust to incomplete modalities data is an urgent issue for clinical practice. + +Currently, mainstream approaches addressing incomplete modalities issues include imputation-based and imputation-free methods [8, 34]. The former approaches aim to impute the missing modality based on observed modalities, primarily achieved in two ways including generation [3, 23, 29] and retrieval [4, 40]. In terms of generation-based imputation, generative models are typically employed for synthesizing features or raw data of missing modalities from available modalities. As shown in Fig. 1 (a), these methods usually can only impute modality-common information, as the modality-specific information unique to the missing modality cannot be generated out of thin air. Moreover, the goal of retrieval-based imputation methods is to fill in missing modalities by seeking the most similar sample of the missing modality from the training set. This way tends + +to compensate only a limited part of modality-common and modality-specific information, as shown in Fig. 1 (b). The reason is that an individual sample exhibits considerable randomness and is challenging to fully capture the unique knowledge inherent in the missing modality. + +Unlike imputation-based approaches, imputation-free approaches [8, 34] primarily focus on how to minimize performance degradation resulting from missing modalities. These approaches make great efforts to learn robust multimodal representations that are less sensitive to missing modalities by capturing modality-invariant information. Recently, with the rise of Large Language Models (LLMs) [1, 20, 33], the researcher [17] attempts to leverage the strong robustness of LLMs benefit from extensive training data to alleviate performance decreases caused by incomplete modalities. They design some sophisticated prompts to inform LLMs of different distributions of inputs. However, they only capture modality-common information as well (Fig. 1a), ignoring distinct modality-specific knowledge of the missing modality. As such, existing methods still struggle to effectively incorporate the learning of both common and specific knowledge of missing modalities. + +Based on these insights, to bridge the aforementioned gap, we propose a Distilled Prompt Learning (DisPro) framework to unlock LLMs' robustness to incomplete modalities, which aims to simultaneously compensate for both specific and common information of missing modalities. It comprises two-stage prompt learning including Unimodal Prompting (UniPro) and Multimodal Prompting (MultiPro) for capturing modality-specific and modality-common information of missing modalities, respectively. + +In the first stage, to prepare for supplementing modality-specific information of missing modalities, we first learn the knowledge distribution of the missing modality by distilling unimodal knowledge into a set of learnable unimodal prompts for each modality from all remaining samples available of this modality. Specifically, inspired by CoOp [44], a set of learnable prompts for different classes is employed to learn knowledge about how to describe different categories of an individual modality. For example, in survival prediction, learned prompts can be intuitively interpreted as how different risk groups will reflect in pathological morphology or gene expression profiles. However, due to the huge size (e.g., $100,000 \times 100,000$ ) of pathological WSIs, we cannot simply apply CoOp designed for small natural images to WSIs. Therefore, in UniPro, we extend CoOp into the multiple instance learning (MIL) paradigm that is commonly used for large-scale WSIs [11, 13, 37]. + +In the second stage, to capture modality-common information for missing modalities, we leverage available modalities as multimodal prompts of LLMs to infer representations of missing modalities using LLMs' strong inference ability. Simultaneously, to inject the learned unimodal + +knowledge in stage 1, we propose UniPro Distillation to enforce the inferred representations of missing modalities to get as close as possible to the learned unimodal prompts of the corresponding class. Additionally, to assist in learning modality-common knowledge from WSIs and genes with high information redundancy [30, 42], we propose UniPro Scoring to re-use UniPro as a grader to select discriminative and relevant tokens. As a result, both common and specific information are compensated for missing modalities under the collaboration of uni- and multimodal prompts, thus mitigating performance declines led by incomplete modalities. + +It is worth noting that DisPro is proposed as a framework for addressing missing modality issues by leveraging LLMs' power based on prompts, capable of generalizing to a wider range of domains. The contributions of this work are summarized as follows: + +- We propose a Distilled Prompt Learning (DisPro) framework for incomplete multimodal survival prediction, where LLMs' robustness to incomplete modalities is explored under two-stage prompting. +- We propose two-stage prompt learning including Unimodal Prompting and Multimodal Prompting to compensate for both modality-specific and modality-common information of missing modalities, respectively. +- Extensive experiments on five benchmark survival prediction datasets are conducted, which demonstrates significantly superior to state-of-the-art approaches. + +# 2. Related Work + +# 2.1. Incomplete Multimodal Learning + +As stated above, incomplete multimodal learning [43] can be categorized into two types: imputation-based [4, 12, 38] and imputation-free methods [21, 25]. For example, with the success of the diffusion model [7], the research [23] attempted to apply it to MRI synthesis for missing modalities. Instead of generating raw modality data, SMIL [22] leveraged Bayesian meta-learning to impute features of missing modalities. For retrieval-based approaches, M3Care [40] filled in the missing modalities by retrieving the most similar samples from the training set. On the other hand, imputation-free approaches provided more flexible solutions, which aim to enhance the robustness of fusion and alleviate performance decrease. [8] proposed a segmentation-based regularizer to reduce the sensitivity to missing modalities. MUSE [34] utilized patients' relationships based on graph networks to perform contrastive learning for improving the robustness to missing modalities. GTP-4o [19] proposed modality-prompted heterogeneous graph to embed each modality into a unified space and complete missing modality through a graph prompting mechanism. Recently, MAP [17] proposed missing-aware prompts to unlock the robustness of LLMs to missing + +modalities by indicating LLMs different missing scenarios. However, these methods typically supplemented modality-common knowledge for missing modalities. + +# 2.2. Multimodal Survival Prediction + +With the success of digital pathology [18, 24, 41] and sequencing techniques [26, 31], pathology and genomic data are commonly integrated to analyze survival outcomes, which have proven the superiority on precision in previous works [5, 36]. MCAT [5] proposed to leverage the interactions across modalities to guide the learning of pathology data. Based on this work, MOTCat [36] further leveraged optimal transport to capture global structure consistency across modalities. In parallel, CMTA [10] investigated the intrinsic cross-modal interactions and potential complementary information by cross-modal translation and alignment. From the perspective of information bottleneck, PIBD [42] proposed a prototypical information bottleneck for the MIL setting to select discriminative instances and disentangle multimodal knowledge for redundancy reduction. Similarly, MMP [30] utilized Gaussian Mixture Model to summarize morphological prototypes and biological pathway prototypes for these two modalities. However, all these methods assume all modalities available, inevitably leading to the difficulty of employment in clinical settings. + +# 3. Method + +This work aims to explore LLMs' robustness to incomplete modalities by two-stage prompting, Unimodal Prompting (UniPro) and Multimodal Prompting (MultiPro), which capture modality-specific and modality-common information, respectively. The overview framework of the proposed method is shown in Fig. 2. First, we will introduce the problem formulation in Sec. 3.1. Then, we will detail the process of UniPro and MultiPro in Sec. 3.2 and 3.3, respectively. + +# 3.1. Problem Formulation + +Given a multimodal dataset consisting of $M = 2$ modalities, we use a 2-tuple $\mathbf{X}_n = (\mathbf{X}_n^p,\mathbf{X}_n^g)$ to represent the $i$ -th patient data, where $\mathbf{X}_n^p$ and $\mathbf{X}_n^g$ are pathology and genomic data, respectively. In particular, we use the notation of tilde hat to indicate the modality is missing, e.g., $\widetilde{\mathbf{X}}_n^p$ . + +Following the conventional setting for pathological WSIs and genomic pathways in previous works [14, 42], we formulate a WSI $\mathbf{X}_n^p$ and a set of genomic pathways $\mathbf{X}_n^g$ as the "bag" of instances in MIL paradigm. We denote $\mathbf{X}_n^p = \{\mathbf{x}_{n,i}^p\}_{i=1}^{M_p}$ where each instance $\mathbf{x}_{n,i}^p$ refers to a patch image cropped from the WSI $\mathbf{X}_n^p$ and $M_p$ is the patch number of this WSI. Similarly, the bag of genomic pathways can be formulated as $\mathbf{X}_n^g = \{\mathbf{x}_{n,i}^g\}_{i=1}^{M_g}$ , in which an instance $\mathbf{x}_{n,i}^g$ represents a pathway consisting of various gene expression values. + +In survival prediction, the goal is to estimate the risk probability of an outcome event taking place before a specific time. However, this event may not always be observed, resulting in a right-censored occurrence. Therefore, $c_{n} \in \{0,1\}$ is used to denote censorship status indicating whether the outcome event is right-censored (i.e., alive) ( $c = 1$ ) or not ( $c = 0$ ) (i.e., dead), and $t_{n} \in \mathbb{R}^{+}$ represents the overall survival time (in months). These two components make up its set of labels $Y_{n} = (c_{n},t_{n})$ . Following the previous setting [5, 36], the continuous survival time would be discretized into $I_{t}$ time intervals corresponding to $I_{t}$ risk bands. As a result, the label set $(c_{n},t_{n})$ can be re-written as $y_{n}$ , resulting in $2I_{t}$ labels. + +Given a set of predicted hazard probabilities $\mathbf{h}_n$ across all time intervals, the hazard probability $\mathbf{h}_n^{(j)}$ represents the probability of death event occurring at the $j$ -th time interval. Then, the cumulative survival probability for each time interval can be estimated as follows: + +$$ +S _ {n} ^ {(j)} = \prod_ {z = 1} ^ {j} \left(1 - \mathbf {h} _ {n} ^ {(z)}\right) \tag {1} +$$ + +Following previous works [36, 42], given the ground truth time interval $\tau$ , we use the NLL loss [39] for survival prediction, where $N$ is the sample number of the training set: + +$$ +\begin{array}{l} \mathcal {L} _ {s u r v} = - \sum_ {n = 1} ^ {N} \left[ c _ {n} \log S _ {n} ^ {(\tau)} + \left(1 - c _ {n}\right) \log S _ {n} ^ {(\tau - 1)} \right. \tag {2} \\ \left. + \left(1 - c _ {n}\right) \log {\bf h} _ {n} ^ {(\tau)} \right] \\ \end{array} +$$ + +# 3.2. Stage 1: Unimodal Prompting (UniPro) + +Given a WSI $\mathbf{X}_n^p = \{\mathbf{x}_{n,i}^p\}_{i=1}^{M_p}$ or a bag of genomic pathways $\mathbf{X}_n^g = \{\mathbf{x}_{n,i}^g\}_{i=1}^{M_g}$ , we formulate their learning process as a MIL task. Inspired by CoOp [44] employing a CLIP-style architecture, we adapt it to the MIL setting to learn how to characterize each modality using a set of learnable prompts for modality-specific knowledge. + +In terms of pathological WSIs data, following the common setting of the WSI processing [28], as shown in the UniPro-P of Fig 2, we first employ a pretrained and frozen patch encoder $g_{E}(\cdot)$ to embed every patch images $\pmb{x}_{n,i}^{p}$ from a WSI into an instance-level embedding $p_n^i = g_E(\pmb{x}_{n,i}^p) \in \mathbb{R}^d$ , where $d$ is the patch feature dimension. To align pathology features with the text space, we apply an adapter $f_{p}(\cdot)$ for each patch feature to get the aligned patch feature $\mathbf{p}_n^i = f_p(g_E(\pmb{x}_{n,i}^p)) \in \mathbb{R}^D$ , implemented by a linear layer followed by ReLU [2]. + +Text-wise, the input comprises three components: pathology context tokens, prefix tokens andclassname tokens. Forclassname tokens $[C]_{1\dots s}$ , we map every time interval of overall survival into a unique description, such as "median risk", and denote the censorship status by "alive" + +![](images/c9d0d3110b8395a40f7190f7be8aaf2b8a09e9428d281820174d23735f364bc1.jpg) +Stage 1: Unimodal Prompting (UniPro) + +![](images/42bfdeca25ae2ec5fdb16a3b3239c882f7290e04e8e0ef29aad92b99bd4fe3fb.jpg) + +Stage 2: Multimodal Prompting (MultiPro) +![](images/789b7be1dbdc1a4a23d8857ae7e979c760ba2645578cf76a799657f73a7189ef.jpg) +d.1. $\mathcal{L}_{\mathrm{surv}}^{\mathrm{cls}}$ + +![](images/04683aa816c6afdce9c8c5cd6cc98c9abc1b9f6f554670ef5d86a633349677a8.jpg) + +![](images/f6fa08541a803969da1d94c9a70c4f5638350c2c1c413bf4593afc11db48c9e2.jpg) +Figure 2. Overview of DisPro. Stage 1: Unimodal Prompting aims to distill the knowledge distribution for each modality and prepare to supplement modality-specific knowledge for the missing modality. Stage 2: Multimodal Prompting utilizes the available modality as prompts to infer the representations of the missing one, compensating for modality-common knowledge. Simultaneously, the learned UniPro of Stage 1 supervises the learning of imputed representations to compensate for modality-specific knowledge. UniPro Scoring re-uses learned prompts to assist the LLM in capturing modality-common knowledge by selecting discriminative and relevant tokens. + +![](images/5b5c49cdbcaf79c879a2444571ccf23c50a90f1fdc1c2d8725afb85c71c6fad1.jpg) + +or "dead". The prefix tokens $\left[\mathrm{PREFIX}\right]_{1\dots m}$ briefly describe the modality, e.g., "This is a pathology slide image from the patient with overall survival of $[C]_{1\dots s}$ ". Prefix and classname tokens are obtained using a frozen Tokenizer pretrained on a large-scale biomedical corpus to embed these words. To learn how to describe the modality for a specific risk band (class), we attach a set of learnable context tokens $[P]_{1\dots k}$ to prefix and classname tokens: + +$$ +\boldsymbol {T} _ {p} = [ P ] _ {1} \dots [ P ] _ {k} [ \text {P R E F I X} ] _ {1} \dots [ \text {P R E F I X} ] _ {m} [ C ] _ {1} \dots [ C ] _ {s} \tag {3} +$$ + +where $k$ , $m$ and $s$ are the token numbers of pathology context tokens, prefix words and classname words, respectively. Each token is a vector with the same dimension $D$ as word embeddings of the LLM (i.e., 768 for BERT). In this way, we can design input tokens for each risk band: + +$$ +\boldsymbol {T} _ {p} ^ {(j)} = [ P ] _ {1} ^ {(j)} \dots [ P ] _ {k} ^ {(j)} [ \text {P R E F I X} ] _ {1} \dots [ \text {P R E F I X} ] _ {m} [ C ] _ {1} ^ {(j)} \dots [ C ] _ {s} ^ {(j)} \tag {4} +$$ + +in which $j \in [1, 2I_t]$ and prefix tokens are shared across all risk bands. By forwarding $T_p^{(j)}$ of each class to the text encoder $\mathcal{T}(\cdot)$ (i.e., LLM), we can obtain a global textual class representation $t_p^{(j)} \in \mathbb{R}^D$ for every class. + +In the MIL setting, to obtain the predicted probability for each class, we can use topK patches to produce the slide-level prediction based on similarity scores. To this end, we first compute the similarity score between every instance and every class: + +$$ +\boldsymbol {s} _ {n, p} ^ {(i, j)} = \boldsymbol {p} _ {n} ^ {i} \cdot \boldsymbol {t} _ {p} ^ {(j)} \tag {5} +$$ + +As a result, we can get a similarity matrix $S_{n,p} \in \mathbb{R}^{M_p \times 2I_t}$ for a WSI. Then, we apply a TopK max-pooling operator $h(\cdot)$ to obtain the slide-level predictions, depend + +ing on top-K similarities for each class: + +$$ +h \left(\mathcal {S} _ {n, p}\right) = \frac {1}{K} \left[ \sum_ {i = 1} ^ {K} \hat {\boldsymbol {s}} _ {n, p} ^ {(i, 1)}, \sum_ {i = 1} ^ {K} \hat {\boldsymbol {s}} _ {n, p} ^ {(i, 2)}, \dots , \sum_ {i = 1} ^ {K} \hat {\boldsymbol {s}} _ {n, p} ^ {(i, 2 \boldsymbol {I} _ {t})} \right] \tag {6} +$$ + +where $\hat{s}_{n,p}$ is the sorted $s_{n,p}$ in a descending order across all instances. Next, we can get a set of hazard probabilities $\mathbf{h}_n^p$ across all time intervals through a sigmoid function. + +$$ +\mathbf {h} _ {n} ^ {p} = \operatorname {S i g m o i d} \left(h \left(S _ {n, p}\right)\right) \tag {7} +$$ + +By computing the survival loss according to Eq. 2 and backforwarding gradients, we can optimize the modality adapter and learnable prompts consisting of context tokens. + +Similarly, we acquire the set of hazard probabilities $\mathbf{h}_n^g$ for genomic pathways and optimize its learnable prompts $T_{g}$ and gene encoder $f_{g}(\cdot)$ . Here we use the subscript $p$ and $g$ to represent pathology and genomic data. Every pathway is also encoded into a $D$ -dimensional embedding. + +# 3.3. Stage 2: Multimodal Prompting (MultiPro) + +In this stage, we aim to leverage LLMs' strong inference ability and robustness to missing modalities to predict the representation of missing modalities, as shown in the MultiPro of Fig. 2. + +To simplify explanations, we take the example input of pathology available and genomics being missing. By prompting the LLM with available pathology data, the Bert-style LLM can capture modality-common knowledge for genomics due to its nature of self-attention. To compensate for modality-specific knowledge, we re-use the unimodal prompts $T_{g}$ learned in stage 1 with all remaining samples available of genomics data to supervise the learning of imputed representations, where unimodal prompts characterize the knowledge distribution of an individual modality. + +To this end, we first need to construct an interleaved input for the learning of modality-common information. To fully leverage the capabilities of LLMs, every modality are supposed to be embedded into the aligned text space. Consequently, we re-use the aligned modality adapter or encoder learned in the first stage to embed each modality, and we continue to finetune them in the second stage. Note that the LLM used in stage 2 is the same as the one in stage 1. As a result, we can get a bag of aligned feature tokens $\{\mathbf{p}_n^i\}_{i = 1}^{M_p}$ or $\{\mathbf{g}_n^i\}_{i = 1}^{M_g}$ for pathology and genomic data, respectively. Here we assume genomic data $\{\tilde{\mathbf{g}}_n^i\}_{i = 1}^{M_g}$ is missing. + +Due to the limitation of input length for general LLMs (e.g., 512 for Bert-style LLMs) and high information redundancy in WSIs or genomic pathways [30, 42], more discriminative and relevant feature tokens are expected to be selected for better capturing modality-common knowledge. Therefore, we propose a UniPro Scoring module by re-using UniPro as a grader to estimate a relevance score for each feature token. + +UniPro Scoring. The score of every feature token is determined by three parts: relevance scores estimated by UniPro-P and UniPro-G, and self-scoring, as shown in Fig. 2. + +Given a set of pathology feature tokens $\{\mathbf{p}_n^i\}_{i = 1}^{M_p}$ , to obtain discriminative tokens, we compute the similarity score $s_{n,p}^{(i,\tau)}$ between the ground-truth textual class representation of $t_p^{(\tau)}$ and each feature token $\mathbf{p}_n^i$ by re-using Eq. 5. Through the sigmoid function, we can get the score $\mathbf{s}_{n,p}^{(i,\tau)}$ estimated by UniPro-P. To capture cross-modal modality-common information for better compensation of missing modality, we similarly obtain the score $\mathbf{s}_{n,g}^{(i,\tau)}$ estimated by UniPro-G. Since textual class representations of UniPro are fixed once finishing the training of stage 1, the selected tokens will also become fixed accordingly. To make it diverse while dynamically adapting to the current input, we employ a learnable self-scoring mechanism implemented by attention layers [13] to assign a score $\mathbf{a}_{n,p}^{(i)}$ for each feature token. In the end, the final score $\mathbf{s}_{n,\#}^{(i)}$ for each feature token can be computed by their summarization: + +$$ +\mathbf {s} _ {n, \#} ^ {(i)} = \mathbf {s} _ {n, p} ^ {(i, \tau)} + \mathbf {s} _ {n, g} ^ {(i, \tau)} + \mathbf {a} _ {n, p} ^ {(i)} \tag {8} +$$ + +For the inference phase when the ground-truth class cannot be accessed, we compute similarity scores with all textual class representations and apply Top-K max-pooling to determine which class $j$ these tokens belong to: + +$$ +\tau = \arg \max _ {j \in [ 1, 2 I _ {t} ]} h \left(\mathcal {S} _ {n, p} ^ {(j)} \oplus \mathcal {S} _ {n, g} ^ {(j)}\right) \tag {9} +$$ + +where $\oplus$ refers to the element-wise addition and $h(\cdot)$ can be computed by Eq. 6. + +By selecting feature tokens with the top- $K_{p}$ largest scores, we can get the input tokens $\{\mathbf{p}_n^i\}_{i=1}^{K_p}$ for available pathology modality. Then, we employ a set of learnable tokens to act as placeholders $[H_p]_{1\dots K_p}$ and $[H_g]_{1\dots K_g}$ for pathology and genomic data, where $K_{p}$ and $K_{g}$ are their token number of the input for the LLM in stage 2, respectively. They are initialized with modality indicators. For the available pathology, we add the selected tokens to its placeholders, leading to the input of pathology $\{\hat{\mathbf{p}}_n^i\}_{i=1}^{K_p}$ . At last, we concatenate them with placeholders of the missing genomics as well as a [CLS] token, resulting in the final input being forwarded into a LLM $\mathcal{T}(\cdot)$ : + +$$ +[ C L S ] [ \hat {\mathbf {p}} _ {n} ] _ {1 \dots K _ {p}} [ \tilde {\mathbf {g}} _ {n} ] _ {1 \dots K _ {g}} = \mathcal {T} ([ \mathrm {C L S} ] | | \{\hat {\mathbf {p}} _ {n} ^ {i} \} _ {i = 1} ^ {K _ {p}} | | [ H _ {g} ] _ {1 \dots K _ {g}}) \tag {10} +$$ + +UniPro Distillation. To inject modality-specific knowledge for the compensation of missing modality, we utilize UniPro to supervise this process, where UniPro has learned the knowledge distribution regarding the missing modality based on other observed samples in stage 1. + +Given the output (Eq. 10) of the LLM in stage 2, we take the part of the missing modality $[\tilde{\mathbf{g}}_n]_{1\dots K_g}$ and compute similarity scores $\tilde{\mathbf{h}}_g$ between $[\tilde{\mathbf{g}}_n]_{1\dots K_g}$ and the textual class + +representation by Eq. 5 - 7. By substituting $[\tilde{\mathbf{g}}_n]_{1\ldots K_g}$ into the survival loss function of Eq. 2, we can supervise the learning of compensation for the missing modality: + +$$ +\begin{array}{l} \mathcal {L} _ {u d} ^ {g} = - \sum_ {n = 1} ^ {\tilde {N} _ {g}} \left[ c _ {n} \log S _ {n} ^ {(\tau)} + \left(1 - c _ {n}\right) \log S _ {n} ^ {(\tau - 1)} \right. \tag {11} \\ \left. + \left(1 - c _ {n}\right) \log \tilde {\mathbf {h}} _ {n} ^ {g, (\tau)} \right] \\ \end{array} +$$ + +where $\tilde{N}_g$ is the number of missing genomic modalities across all data pairs. Similarly, we can get the loss function $\mathcal{L}_{ud}^p$ when the pathology modality is missing. Lastly, [CLS] token is used for calculating the loss function for survival prediction $\mathcal{L}_{surv}^{cls}$ in stage 2, and the final loss can be combined linearly with coefficient factors $\alpha_{1}$ and $\alpha_{2}$ : + +$$ +\mathcal {L} = \mathcal {L} _ {s u r v} ^ {c l s} + \alpha_ {1} \mathcal {L} _ {u d} ^ {p} + \alpha_ {2} \mathcal {L} _ {u d} ^ {g} \tag {12} +$$ + +# 4. Experiments + +# 4.1. Datasets and Settings + +Datasets. To demonstrate the efficacy of the proposed method, we conducted a series of experiments across five public cancer datasets sourced from The Cancer Genome Atlas (TCGA) $^{1}$ . The case number of each dataset is demonstrated next to the dataset name in Tab. 1. Regarding genomic data, we follow the grouping of biological pathways provided by [14] and remove genes absent in the cBioPortal $^{2}$ database, resulting in 330 pathways. More details of datasets can be found in supplementary materials. + +Evaluation Settings on Incomplete Modalities. Following the common setting in previous works [22, 32], to systematically assess performances on various missing settings, we first manually construct a series of missing scenarios under different missing ratios by randomly dropping them for training sets on samples with complete modality pairs. Specifically, we set a total missing rate (60%) for two modalities and then construct 5 combinations for pathology and genomics, i.e., $\{(0\%, 60\%), (20\%, 40\%), (30\%, 30\%), (40\%, 20\%), (60\%, 0\%)\}$ . For inference, all methods are evaluated on three scenarios: pathology only, genomics only, and complete modalities. All experiments are conducted on 5-fold cross-validation following previous works [5, 36] in survival prediction, and the concordance index (C-Index) and its standard deviation (std) are reported to evaluate performances. The LLM is implemented by BioBERT-v1.2 [16]. The supplementary materials contain more experiment results and details on the implementation. + +# 4.2. Comparisons with SOTAs + +We first present unimodal approaches, and SOTA multimodal methods trained with complete modalities (all + +pathology and genomics data) to serve as the upper bound of current complete modality approaches. Then the average performances on all missing combinations are demonstrated under the total training missing rate of $60\%$ , evaluated on three inference scenarios. Details on various combinations are shown in Fig. 3 and we discuss them in Sec. 4.3. Lastly, we investigate the upper bound of each model tailored to incomplete modalities by training them with all data (without missing data). + +From results in Tab. 1, we observed that (1) When applying $60\%$ training missing rate, DisPro consistently outperformed SOTAs across five datasets by a large margin (on average, $+1.76\%$ for pathology only, $+1.57\%$ for genomics only and $2.33\%$ for complete data). This indicates the superiority of DisPro over the incomplete SOTA methods. (2) When testing on the complete scenario, on 3 out of 5 datasets and overall, DisPro even surpassed the SOTA multimodal survival model (SurvPath) trained on all data. Notably, DisPro achieved this while only utilizing incomplete training data with a $60\%$ missing rate. It suggests that harnessing the powerful capabilities of LLMs is a promising solution for addressing the classical question of survival analysis. (3) When using complete training data, DisPro consistently performed best on average for various test scenarios as well ( $2.28\%$ on pathology only, $1.33\%$ on genomics only and $2.57\%$ on complete test data). Under the same condition, DisPro further increased by $1.31\%$ compared with the SOTA complete multimodal method, further demonstrating the strong power of LLMs. (4) In contrast to unimodal methods trained on full unimodal datasets, DisPro, using only incomplete multimodal data with a $60\%$ missing rate, achieved superior or similar performance when inferring from individual modalities. This showcases DisPro's efficient utilization of multimodal data. + +# 4.3. Ablation Study + +Component Validation. In Tab. 2, we ablate each component in our method to investigate their effectiveness. (1) To demonstrate how the LLM performs on missing modalities without any prompts, we establish a baseline by employing a linear classifier head on the same LLM and randomly selecting feature tokens for its input. Results show that the performance significantly decreased by $2.62\%$ on pathology only, $1.55\%$ on genomics only and $2.05\%$ on complete test data. This indicates the importance of appropriate prompts for LLMs. (2) When supplementing modality-specific information for missing modalities, the performance $(+UD)$ consistently improves on all datasets and increases by around $+1\%$ on various inference settings, which validates the effectiveness of compensating for modality-specific knowledge of the missing modality. (3) When reusing UniPro to select relevant tokens $(+US)$ , the performance increased by about $+1\%$ overall, which bene- + +
TrainingMiRateTestP GMethodsBLCA(372)BRCA(1007)COADREAD(533)LUAD(443)UCEC(478)Avg
N/A†TransMIL [28]0.5903±0.0250.6671±0.0540.6093±0.0180.6195±0.0350.6978±0.0460.6368
†CoOp_BioBert [44]0.6014±0.0470.6784±0.0320.6681±0.0260.6543±0.0520.6791±0.0850.6563
†SNN [15]0.6831±0.0400.6567±0.0380.6312±0.0580.6379±0.0710.7267±0.0730.6671
†CoOp_BioBert [44]0.6712±0.0360.6922±0.0350.6743±0.0480.6514±0.0650.7171±0.0600.6812
‡MOTCat [36]0.6274±0.0460.6721±0.0380.6501±0.0350.6753±0.0430.7207±0.0740.6691
‡SurvPath [14]0.6572±0.0270.7071±0.0390.7079±0.0480.6798±0.0490.7391±0.0650.6982
60%•○COM [27]0.6024±0.0130.6740±0.0110.6779±0.0140.6336±0.0110.6985±0.0120.6573
•○M3Care [40]0.6208±0.0140.6690±0.0180.6570±0.0230.6219±0.0070.7031±0.0120.6543
•○HGCN [9]0.5999±0.0090.6831±0.0120.6792±0.0090.6465±0.0050.7011±0.0070.6620
•○MUSE [34]0.6207±0.0130.6514±0.0140.6547±0.0090.6354±0.0120.7048±0.0230.6534
•○MAP [17]0.5918±0.0120.6278±0.0060.5965±0.0170.6489±0.0200.6929±0.0210.6316
•○DisPro (Ours)0.6319±0.0140.6895±0.0110.6880±0.0100.6612±0.0140.7272±0.0180.6796
○•COM [27]0.6415±0.0160.6653±0.0180.6593±0.0300.6456±0.0080.7105±0.0070.6645
○•M3Care [40]0.6420±0.0150.6642±0.0290.6491±0.0080.6462±0.0120.7032±0.0130.6609
○•HGCN [9]0.6225±0.0080.6729±0.0090.6752±0.0150.6418±0.0150.7015±0.0100.6628
○•MUSE [34]0.6299±0.0120.6427±0.0250.6391±0.0230.6462±0.0100.6893±0.0270.6494
○•MAP [17]0.5937±0.0160.6269±0.0060.5971±0.0170.6442±0.0090.6934±0.0180.6311
○•DisPro (Ours)0.6547±0.0120.6841±0.0180.6804±0.0240.6548±0.0120.7271±0.0170.6802
••COM [27]0.6192±0.0130.6846±0.0090.6874±0.0050.6548±0.0120.7136±0.0160.6719
••M3Care [40]0.6513±0.0050.6951±0.0240.6698±0.0160.6491±0.0130.7288±0.0110.6788
••HGCN [9]0.6332±0.0060.6907±0.0120.6861±0.0170.6545±0.0110.7077±0.0080.6745
••MUSE [34]0.6207±0.0120.6555±0.0130.6539±0.0120.6415±0.0080.7164±0.0130.6576
••MAP [17]0.5906±0.0130.6277±0.0040.5990±0.0160.6625±0.0080.6901±0.0180.6340
••DisPro (Ours)0.6638±0.0060.7219±0.0150.7029±0.0160.6741±0.0070.7476±0.0200.7021
0%••COM [27]0.6154±0.0410.6871±0.0270.6925±0.0440.6427±0.0460.6806±0.0610.6637
••M3Care [40]0.6268±0.0240.6775±0.0160.6850±0.0520.6299±0.0380.6964±0.0610.6631
••HGCN [9]0.6055±0.0480.6716±0.0220.6734±0.0480.6421±0.0520.6915±0.0530.6568
••MUSE [34]0.6208±0.0260.6767±0.0400.6710±0.0410.6253±0.0410.7099±0.0630.6607
••MAP [17]0.6205±0.0330.6300±0.0530.5930±0.0650.6662±0.0550.6925±0.0640.6404
••DisPro (Ours)0.6315±0.0280.6945±0.0440.7216±0.0280.6747±0.0520.7104±0.0560.6865
○•COM [27]0.6681±0.0300.6966±0.0470.6770±0.0240.6454±0.0500.7095±0.0590.6793
○•M3Care [40]0.6706±0.0390.6611±0.0400.6563±0.0210.6512±0.0610.7047±0.0600.6688
○•HGCN [9]0.6233±0.0330.6807±0.0230.6777±0.0370.6421±0.0450.7091±0.0430.6666
○•MUSE [34]0.6256±0.0290.6401±0.0240.6208±0.0300.6547±0.0750.6824±0.0390.6447
○•MAP [17]0.6093±0.0260.6305±0.0510.5842±0.0650.6653±0.0450.6841±0.0640.6347
○•DisPro (Ours)0.6745±0.0390.6996±0.0280.7000±0.0210.6751±0.0680.7140±0.0600.6926
••COM [27]0.6376±0.0430.7087±0.0240.6990±0.0450.6659±0.0340.7112±0.0510.6845
••M3Care [40]0.6605±0.0230.6808±0.0410.6768±0.0260.6740±0.0460.7359±0.0330.6856
••HGCN [9]0.6365±0.0210.7058±0.0330.7136±0.0250.6692±0.0500.7017±0.0440.6854
••MUSE [34]0.6282±0.0390.7017±0.0200.6693±0.0420.6248±0.0400.7237±0.0620.6695
••MAP [17]0.6034±0.0260.6328±0.0490.5895±0.0630.6904±0.0470.6827±0.0650.6398
••DisPro (Ours)0.6719±0.0510.7358±0.0510.7553±0.0210.6767±0.0410.7169±0.0710.7113
+ +Table 1. C-Index of Comparison to SOTA models. "MiRate" represents Missing Rate. "P" and "G" indicate Pathology and Genomics, respectively. $\circ$ and $\bullet$ mean the missing modality and the available modality,respectively. † and ‡ indicate using unimodal and complete multimodal data, respectively, while the rest are for missing modalities. The best results are in bold. The second-best results are underlined. + +fits from the assistance of US in learning modality-common knowledge from WSIs and genes with high redundancy. + +Various Missing Scenarios. To illustrate how incomplete multimodal approaches handle varying missing settings, we display different combinations of training missing rates for each inference scenario, as shown in Fig. 3. Results showcase that DisPro achieved considerably consistent perfor + +mance superiority in various missing settings over other incomplete multimodal approaches by a clear margin. + +# 5. Conclusion + +This work proposed a Distilled Prompt Learning (DisPro) framework to unlock the potential of LLMs for incomplete multimodal learning by two-stage prompting, in which + +
VariantsTestModulesBLCA (372)BRCA (1007)COADREAD (533)LUAD (443)UCEC (478)Avg
PGUSUD
baseline•○0.6068±0.0210.6668±0.0060.6583±0.0120.6270±0.0180.7081±0.0190.6534
+ UD•○✓0.6213±0.0100.6829±0.0080.6740±0.0220.6327±0.0180.7152±0.0110.6652
+ US•○✓0.6285±0.0210.6884±0.0160.6653±0.0260.6546±0.0130.7203±0.0070.6714
DisPro (full)•○✓✓0.6319±0.0140.6895±0.0110.6880±0.0100.6612±0.0140.7272±0.0180.6796
baseline○•0.6435±0.0160.6605±0.0330.6601±0.0180.6467±0.0160.7127±0.0070.6647
+ UD○•✓0.6473±0.0170.6836±0.0140.6707±0.0220.6420±0.0200.7226±0.0130.6732
+ US○•✓0.6452±0.0200.6666±0.0170.6763±0.0230.6471±0.0230.7151±0.0090.6701
DisPro (full)○•✓✓0.6547±0.0120.6841±0.0180.6804±0.0240.6548±0.0120.7271±0.0170.6802
baseline••0.6522±0.0100.7075±0.0120.6785±0.0200.6541±0.0090.7157±0.0030.6816
+ UD••✓0.6585±0.0070.7194±0.0130.6963±0.0090.6635±0.0140.7250±0.0120.6926
+ US••✓0.6593±0.0080.7125±0.0100.7005±0.0180.6717±0.0070.7274±0.0100.6943
DisPro (full)••✓✓0.6638±0.0060.7219±0.0150.7029±0.0160.6741±0.0070.7476±0.0200.7021
+ +Table 2. C-Index of Ablation Study on the Setting of $60\%$ Training Missing Rate."US" and "UD" refer to the UniPro Scoring and UniPro Distillation modules, respectively. "P" and "G" indicate Pathology and Genomics, respectively. $\circ$ means the modality is missing, while $\bullet$ refers to the available modality. + +![](images/4918a135959d8a2aaa250d2f0febc2b07cefb70695a1228d78a63e9cb9a6efb9.jpg) +Figure 3. Performance on various combinations under $60\%$ training missing rate for different test scenarios. + +Unimodal Prompting (UniPro) and Multimodal Prompting (MultiPro) were proposed for supplementing modality-specific and modality-common knowledge for missing modalities. In the first stage, UniPro extended CoOp into the MIL setting and distilled knowledge distribution to a set of learnable prompts for each modality, which prepped for compensation of modality-specific knowledge. In the second stage, considering available modalities as prompts, we leveraged an LLM to impute common knowledge for miss + +ing modalities. To inject the specific knowledge of missing modalities, the distilled prompts in UniPro are used to supervise the learning of imputing representations for missing modalities, leading to more comprehensive knowledge compensation in incomplete multimodal learning. Extensive experiments on five cancer benchmark datasets demonstrated the effectiveness of the proposed method over state-of-the-art approaches. + +# Acknowledgement + +This work was supported by the National Natural Science Foundation of China (No. 62202403), Hong Kong Innovation and Technology Commission (Project No. MHP/002/22 and ITCPD/17-9), and Research Grants Council of the Hong Kong Special Administrative Region, China (Project No. R6003-22 and C4024-22GF). + +# References + +[1] Som S Biswas. Role of chat gpt in public health. Annals of biomedical engineering, 51(5):868-869, 2023. 2 +[2] Jason Brownlee. A gentle introduction to the rectified linear unit (relu). Machine learning mastery, 6, 2019. 3, 1 +[3] Agisilaos Chartsias, Thomas Joyce, Mario Valerio Giuffrida, and Sotirios A Tsaftaris. Multimodal mr synthesis via modality-invariant latent representation. IEEE transactions on medical imaging, 37(3):803-814, 2017. 1 +[4] Jiayi Chen and Aidong Zhang. Hgmf: heterogeneous graph-based fusion for multimodal data with incompleteness. In Proceedings of the 26th ACM SIGKDD international conference on knowledge discovery & data mining, pages 1295-1305, 2020. 1, 2 +[5] Richard J Chen, Ming Y Lu, Wei-Hung Weng, Tiffany Y Chen, Drew FK Williamson, Trevor Manz, Maha Shady, and Faisal Mahmood. Multimodal co-attention transformer for survival prediction in gigapixel whole slide images. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 4015–4025, 2021. 1, 3, 6 +[6] Richard J Chen, Tong Ding, Ming Y Lu, Drew FK Williamson, Guillaume Jaume, Andrew H Song, Bowen Chen, Andrew Zhang, Daniel Shao, Muhammad Shaban, et al. Towards a general-purpose foundation model for computational pathology. Nature Medicine, 30(3):850-862, 2024. 1 +[7] Florinel-Alin Croitoru, Vlad Hondru, Radu Tudor Ionescu, and Mubarak Shah. Diffusion models in vision: A survey. IEEE Transactions on Pattern Analysis and Machine Intelligence, 45(9):10850-10869, 2023. 2 +[8] Yuhang Ding, Xin Yu, and Yi Yang. Rfnet: Region-aware fusion network for incomplete multi-modal brain tumor segmentation. In Proceedings of the IEEE/CVF international conference on computer vision, pages 3975-3984, 2021. 1, 2 +[9] Hou et al. Hybrid graph convolutional network with online masked autoencoder for robust multimodal cancer survival prediction. IEEE Transactions on Medical Imaging, 2023. 7 +[10] Zhou et al. Cross-modal translation and alignment for survival analysis. In ICCV, pages 21485-21494, 2023. 1, 3 +[11] Michael Gadermayr and Maximilian Tschuchnig. Multiple instance learning for digital pathology: A review of the state-of-the-art, limitations & future potential. Computerized Medical Imaging and Graphics, page 102337, 2024. 2 +[12] Mohammad Hamghalam, Alejandro F Frangi, Baiying Lei, and Amber L Simpson. Modality completion via gaussian process prior variational autoencoders for multi-modal + +glioma segmentation. In Medical Image Computing and Computer Assisted Intervention-MICCAI 2021: 24th International Conference, Strasbourg, France, September 27-October 1, 2021, Proceedings, Part VII 24, pages 442-452. Springer, 2021. 2 +[13] Maximilian Ilse, Jakub Tomczak, and Max Welling. Attention-based deep multiple instance learning. In International conference on machine learning, pages 2127-2136. PMLR, 2018. 2, 5 +[14] Guillaume Jaume, Anurag Vaidya, Richard J Chen, Drew FK Williamson, Paul Pu Liang, and Faisal Mahmood. Modeling dense multimodal interactions between biological pathways and histology for survival prediction. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 11579-11590, 2024. 3, 6, 7, 1 +[15] Günter Klambauer, Thomas Unterthiner, Andreas Mayr, and Sepp Hochreiter. Self-normalizing neural networks. Advances in neural information processing systems, 30, 2017. 7, 1 +[16] Jinhyuk Lee, Wonjin Yoon, Sungdong Kim, Donghyeon Kim, Sunkyu Kim, Chan Ho So, and Jaewoo Kang. Biobert: a pre-trained biomedical language representation model for biomedical text mining. Bioinformatics, 36(4):1234-1240, 2020. 6, 1 +[17] Yi-Lun Lee, Yi-Hsuan Tsai, Wei-Chen Chiu, and Chen-Yu Lee. Multimodal prompting with missing modalities for visual recognition. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 14943-14952, 2023. 2, 7, 3 +[18] Bin Li, Yin Li, and Kevin W Eliceiri. Dual-stream multiple instance learning network for whole slide image classification with self-supervised contrastive learning. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 14318-14328, 2021. 3 +[19] Chenxin Li, Xinyu Liu, Cheng Wang, Yifan Liu, Weihao Yu, Jing Shao, and Yixuan Yuan. Gtp-4o: Modality-prompted heterogeneous graph learning for omni-modal biomedical representation. In European conference on computer vision, pages 168–187. Springer, 2024. 2 +[20] Chunyuan Li, Cliff Wong, Sheng Zhang, Naoto Usuyama, Haotian Liu, Jianwei Yang, Tristan Naumann, Hoifung Poon, and Jianfeng Gao. Llava-med: Training a large language-and-vision assistant for biomedicine in one day. Advances in Neural Information Processing Systems, 36, 2024. 2 +[21] Han Liu, Yubo Fan, Hao Li, Jiacheng Wang, Dewei Hu, Can Cui, Ho Hin Lee, Huahong Zhang, and Ipek Oguz. Moddrop++: A dynamic filter network with intra-subject cotraining for multiple sclerosis lesion segmentation with missing modalities. In International Conference on Medical Image Computing and Computer-Assisted Intervention, pages 444-453. Springer, 2022. 2 +[22] Mengmeng Ma, Jian Ren, Long Zhao, Sergey Tulyakov, Cathy Wu, and Xi Peng. Smil: Multimodal learning with severely missing modality. In Proceedings of the AAAI Conference on Artificial Intelligence, pages 2302-2310, 2021. 2, 6 +[23] Xiangxi Meng, Kaicong Sun, Jun Xu, Xuming He, and Dinggang Shen. Multi-modal modality-masked diffusion network + +for brain mri synthesis with random modality missing. IEEE Transactions on Medical Imaging, 2024. 1, 2 +[24] Muhammad Khalid Khan Niazi, Anil V Parwani, and Metin N Gurcan. Digital pathology and artificial intelligence. The lancet oncology, 20(5):e253-e261, 2019. 3 +[25] Zhenyuan Ning, Denghui Du, Chao Tu, Qianjin Feng, and Yu Zhang. Relation-aware shared representation learning for cancer prognosis analysis with auxiliary clinical variables and incomplete multi-modality data. IEEE Transactions on Medical Imaging, 41(1):186-198, 2021. 2 +[26] Fatih Ozsolak and Patrice M Milos. RNA sequencing: advances, challenges and opportunities. Nature reviews genetics, 12(2):87-98, 2011. 3 +[27] Shuwei Qian and Chongjun Wang. Com: Contrastive masked-attention model for incomplete multimodal learning. Neural Networks, 162:443–455, 2023. 7 +[28] Zhuchen Shao, Hao Bian, Yang Chen, Yifeng Wang, Jian Zhang, Xiangyang Ji, et al. Transmil: Transformer based correlated multiple instance learning for whole slide image classification. Advances in neural information processing systems, 34:2136-2147, 2021. 3, 7 +[29] Anmol Sharma and Ghassan Hamarneh. Missing mri pulse sequence synthesis using multi-modal generative adversarial network. IEEE transactions on medical imaging, 39(4): 1170-1183, 2019. 1 +[30] Andrew H Song, Richard J Chen, Guillaume Jaume, Anurag Jayant Vaidya, Alexander Baras, and Faisal Mahmood. Multimodal prototyping for cancer survival prediction. In *Forty-first International Conference on Machine Learning*, 2024. 1, 2, 3, 5 +[31] Rory Stark, Marta Grzelak, and James Hadfield. Rna sequencing: the teenage years. Nature Reviews Genetics, 20 (11):631-656, 2019. 3 +[32] Hu Wang, Yuanhong Chen, Congbo Ma, Jodie Avery, Louise Hull, and Gustavo Carneiro. Multi-modal learning with missing modality via shared-specific feature modelling. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 15878-15887, 2023. 6 +[33] Tianyu Wu, Shizhu He, Jingping Liu, Siqi Sun, Kang Liu, Qing-Long Han, and Yang Tang. A brief overview of chatgpt: The history, status quo and potential future development. IEEE/CAA Journal of Automatica Sinica, 10(5):1122-1136, 2023. 2 +[34] Zhenbang Wu, Anant Dadu, Nicholas Tustison, Brian Avants, Mike Nalls, Jimeng Sun, and Faraz Faghri. Multimodal patient representation learning with missing modalities and labels. In The Twelfth International Conference on Learning Representations, 2024. 1, 2, 7 +[35] Conghao Xiong, Hao Chen, Hao Zheng, Dong Wei, Yefeng Zheng, Joseph JY Sung, and Irwin King. Mome: Mixture of multimodal experts for cancer survival prediction. In International Conference on Medical Image Computing and Computer-Assisted Intervention, pages 318-328. Springer, 2024. 1 +[36] Yingxue Xu and Hao Chen. Multimodal optimal transport-based co-attention transformer with global structure consistency for survival prediction. In Proceedings of the + +IEEE/CVF International Conference on Computer Vision, pages 21241-21251, 2023. 1, 3, 6, 7 +[37] Shu Yang, Yihui Wang, and Hao Chen. Mambamil: Enhancing long sequence modeling with sequence reordering in computational pathology. In International Conference on Medical Image Computing and Computer-Assisted Intervention, pages 296-306. Springer, 2024. 2 +[38] Yixiao Yuan, Yawen Huang, and Yi Zhou. Rethinking a unified generative adversarial model for mri modality completion. In International Conference on Medical Image Computing and Computer-Assisted Intervention, pages 143-153. Springer, 2023. 2 +[39] Shekoufeh Gorgi Zadeh and Matthias Schmid. Bias in cross-entropy-based training of deep survival networks. IEEE transactions on pattern analysis and machine intelligence, 43(9):3126-3137, 2020. 3 +[40] Chaohe Zhang, Xu Chu, Liantao Ma, Yinghao Zhu, Yasha Wang, Jiangtao Wang, and Junfeng Zhao. M3care: Learning with missing modalities in multimodal healthcare data. In Proceedings of the 28th ACM SIGKDD Conference on Knowledge Discovery and Data Mining, pages 2418-2428, 2022. 1, 2, 7 +[41] Hongrun Zhang, Yanda Meng, Yitian Zhao, Yihong Qiao, Xiaoyun Yang, Sarah E Coupland, and Yalin Zheng. Dtfmil: Double-tier feature distillation multiple instance learning for histopathology whole slide image classification. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 18802-18812, 2022. 3 +[42] Yilan Zhang, Yingxue Xu, Jianqi Chen, Fengying Xie, and Hao Chen. Prototypical information bottlenecking and disentangling for multimodal cancer survival prediction. In The Twelfth International Conference on Learning Representations, 2024. 1, 2, 3, 5 +[43] Huajun Zhou, Fengtao Zhou, Chenyu Zhao, Yingxue Xu, Luyang Luo, and Hao Chen. Multimodal data integration for precision oncology: Challenges and future directions. arXiv preprint arXiv:2406.19611, 2024. 1, 2 +[44] Kaiyang Zhou, Jingkang Yang, Chen Change Loy, and Ziwei Liu. Learning to prompt for vision-language models. International Journal of Computer Vision, 130(9):2337-2348, 2022. 2, 3, 7 \ No newline at end of file diff --git a/2025/Distilled Prompt Learning for Incomplete Multimodal Survival Prediction/images.zip b/2025/Distilled Prompt Learning for Incomplete Multimodal Survival Prediction/images.zip new file mode 100644 index 0000000000000000000000000000000000000000..eee85780c33271547e09df730c8ecb4a70a00d5c --- /dev/null +++ b/2025/Distilled Prompt Learning for Incomplete Multimodal Survival Prediction/images.zip @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:d084ae2b4653477e697f8b14a69157ff098eb997495eb3f587e4a69480fedc33 +size 1082952 diff --git a/2025/Distilled Prompt Learning for Incomplete Multimodal Survival Prediction/layout.json b/2025/Distilled Prompt Learning for Incomplete Multimodal Survival Prediction/layout.json new file mode 100644 index 0000000000000000000000000000000000000000..d4e43a4241707f05f72e463f6523584f4d9de616 --- /dev/null +++ b/2025/Distilled Prompt Learning for Incomplete Multimodal Survival Prediction/layout.json @@ -0,0 +1,8523 @@ +{ + "pdf_info": [ + { + "para_blocks": [ + { + "bbox": [ + 77, + 103, + 533, + 121 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 77, + 103, + 533, + 121 + ], + "spans": [ + { + "bbox": [ + 77, + 103, + 533, + 121 + ], + "type": "text", + "content": "Distilled Prompt Learning for Incomplete Multimodal Survival Prediction" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 78, + 142, + 532, + 185 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 78, + 142, + 532, + 185 + ], + "spans": [ + { + "bbox": [ + 78, + 142, + 532, + 185 + ], + "type": "text", + "content": "Yingxue Xu" + }, + { + "bbox": [ + 78, + 142, + 532, + 185 + ], + "type": "inline_equation", + "content": "^{1}" + }, + { + "bbox": [ + 78, + 142, + 532, + 185 + ], + "type": "text", + "content": " Fengtao Zhou" + }, + { + "bbox": [ + 78, + 142, + 532, + 185 + ], + "type": "inline_equation", + "content": "^{1}" + }, + { + "bbox": [ + 78, + 142, + 532, + 185 + ], + "type": "text", + "content": " Chenyu Zhao" + }, + { + "bbox": [ + 78, + 142, + 532, + 185 + ], + "type": "inline_equation", + "content": "^{1}" + }, + { + "bbox": [ + 78, + 142, + 532, + 185 + ], + "type": "text", + "content": " Yihui Wang" + }, + { + "bbox": [ + 78, + 142, + 532, + 185 + ], + "type": "inline_equation", + "content": "^{1}" + }, + { + "bbox": [ + 78, + 142, + 532, + 185 + ], + "type": "text", + "content": " Can Yang" + }, + { + "bbox": [ + 78, + 142, + 532, + 185 + ], + "type": "inline_equation", + "content": "^{2}" + }, + { + "bbox": [ + 78, + 142, + 532, + 185 + ], + "type": "text", + "content": " Hao Chen" + }, + { + "bbox": [ + 78, + 142, + 532, + 185 + ], + "type": "inline_equation", + "content": "^{1\\boxtimes}" + }, + { + "bbox": [ + 78, + 142, + 532, + 185 + ], + "type": "inline_equation", + "content": "^{1}" + }, + { + "bbox": [ + 78, + 142, + 532, + 185 + ], + "type": "text", + "content": "Department of Computer Science and Engineering, " + }, + { + "bbox": [ + 78, + 142, + 532, + 185 + ], + "type": "inline_equation", + "content": "^{2}" + }, + { + "bbox": [ + 78, + 142, + 532, + 185 + ], + "type": "text", + "content": "Department of Mathematics The Hong Kong University of Science and Technology" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 123, + 187, + 484, + 200 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 123, + 187, + 484, + 200 + ], + "spans": [ + { + "bbox": [ + 123, + 187, + 484, + 200 + ], + "type": "text", + "content": "{yxueb,fzhouaf,cyzhao,ywangrm}@connect.ust.hk，{macyang,jhc}@ust.hk" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 151, + 227, + 200, + 239 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 151, + 227, + 200, + 239 + ], + "spans": [ + { + "bbox": [ + 151, + 227, + 200, + 239 + ], + "type": "text", + "content": "Abstract" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 54, + 251, + 296, + 574 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 54, + 251, + 296, + 574 + ], + "spans": [ + { + "bbox": [ + 54, + 251, + 296, + 574 + ], + "type": "text", + "content": "The integration of multimodal data including pathology images and gene profiles is widely applied in precise survival prediction. Despite recent advances in multimodal survival models, collecting complete modalities for multimodal fusion still poses a significant challenge, hindering their application in clinical settings. Current approaches tackling incomplete modalities often fall short, as they typically compensate for only a limited part of the knowledge of missing modalities. To address this issue, we propose a Distilled Prompt Learning framework (DisPro) to utilize the strong robustness of Large Language Models (LLMs) to missing modalities, which employs two-stage prompting for compensation of comprehensive information for missing modalities. In the first stage, Unimodal Prompt (UniPro) distills the knowledge distribution of each modality, preparing for supplementing modality-specific knowledge of the missing modality in the subsequent stage. In the second stage, Multimodal Prompt (MultiPro) leverages available modalities as prompts for LLMs to infer the missing modality, which provides modality-common information. Simultaneously, the unimodal knowledge acquired in the first stage is injected into multimodal inference to compensate for the modality-specific knowledge of the missing modality. Extensive experiments covering various missing scenarios demonstrated the superiority of the proposed method. The code is available at https://github.com/Innse/DisPro." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 55, + 597, + 135, + 609 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 597, + 135, + 609 + ], + "spans": [ + { + "bbox": [ + 55, + 597, + 135, + 609 + ], + "type": "text", + "content": "1. Introduction" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 55, + 617, + 296, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 617, + 296, + 713 + ], + "spans": [ + { + "bbox": [ + 55, + 617, + 296, + 713 + ], + "type": "text", + "content": "Multimodal Survival Prediction typically incorporates pathology slides and genomic data complementary to each other for precise prognosis analysis. Pathology data can provide qualitative morphological insights, while quantitative molecular information can be obtained from genomic data [5, 35, 36]. Despite progress [5, 10, 36] in multimodal survival analysis, these models typically require complete modalities data for reliable performance. However, the ex" + } + ] + } + ], + "index": 8 + }, + { + "type": "image", + "bbox": [ + 328, + 227, + 538, + 338 + ], + "blocks": [ + { + "bbox": [ + 328, + 227, + 538, + 338 + ], + "lines": [ + { + "bbox": [ + 328, + 227, + 538, + 338 + ], + "spans": [ + { + "bbox": [ + 328, + 227, + 538, + 338 + ], + "type": "image", + "image_path": "d84997a7d97f611945b7c7e55fdbc05e48e28a07581278184108b361e3588ef3.jpg" + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 313, + 342, + 555, + 387 + ], + "lines": [ + { + "bbox": [ + 313, + 342, + 555, + 387 + ], + "spans": [ + { + "bbox": [ + 313, + 342, + 555, + 387 + ], + "type": "text", + "content": "Figure 1. Insights for existing incomplete multimodal learning and comparison to the proposed method. (a) Generation-based Imputation and Imputation-free approaches, (b) Retrieved-based Imputation and (c) Ours." + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "image_caption" + } + ], + "index": 9 + }, + { + "bbox": [ + 312, + 399, + 555, + 531 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 312, + 399, + 555, + 531 + ], + "spans": [ + { + "bbox": [ + 312, + 399, + 555, + 531 + ], + "type": "text", + "content": "pectation of having full access to complete modalities for integration is not always achievable, especially in clinical scenarios, due to various factors such as the cost of data gathering and privacy issues. For example, acquiring genomic data still comes at a high cost due to the technology and infrastructure required for gene sequencing, especially in underdeveloped areas [43]. Inevitably, it is challenging to adapt previously established multimodal survival models [30, 42] to real-world clinical settings. As a result, building a multimodal survival model that is robust to incomplete modalities data is an urgent issue for clinical practice." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 313, + 534, + 556, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 534, + 556, + 715 + ], + "spans": [ + { + "bbox": [ + 313, + 534, + 556, + 715 + ], + "type": "text", + "content": "Currently, mainstream approaches addressing incomplete modalities issues include imputation-based and imputation-free methods [8, 34]. The former approaches aim to impute the missing modality based on observed modalities, primarily achieved in two ways including generation [3, 23, 29] and retrieval [4, 40]. In terms of generation-based imputation, generative models are typically employed for synthesizing features or raw data of missing modalities from available modalities. As shown in Fig. 1 (a), these methods usually can only impute modality-common information, as the modality-specific information unique to the missing modality cannot be generated out of thin air. Moreover, the goal of retrieval-based imputation methods is to fill in missing modalities by seeking the most similar sample of the missing modality from the training set. This way tends" + } + ] + } + ], + "index": 12 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "spans": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "type": "text", + "content": "CVF" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 145, + 0, + 496, + 37 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 145, + 0, + 496, + 37 + ], + "spans": [ + { + "bbox": [ + 145, + 0, + 496, + 37 + ], + "type": "text", + "content": "This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "text", + "content": "5102" + } + ] + } + ], + "index": 13 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 0 + }, + { + "para_blocks": [ + { + "bbox": [ + 55, + 72, + 294, + 133 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 72, + 294, + 133 + ], + "spans": [ + { + "bbox": [ + 55, + 72, + 294, + 133 + ], + "type": "text", + "content": "to compensate only a limited part of modality-common and modality-specific information, as shown in Fig. 1 (b). The reason is that an individual sample exhibits considerable randomness and is challenging to fully capture the unique knowledge inherent in the missing modality." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 55, + 134, + 295, + 337 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 134, + 295, + 337 + ], + "spans": [ + { + "bbox": [ + 55, + 134, + 295, + 337 + ], + "type": "text", + "content": "Unlike imputation-based approaches, imputation-free approaches [8, 34] primarily focus on how to minimize performance degradation resulting from missing modalities. These approaches make great efforts to learn robust multimodal representations that are less sensitive to missing modalities by capturing modality-invariant information. Recently, with the rise of Large Language Models (LLMs) [1, 20, 33], the researcher [17] attempts to leverage the strong robustness of LLMs benefit from extensive training data to alleviate performance decreases caused by incomplete modalities. They design some sophisticated prompts to inform LLMs of different distributions of inputs. However, they only capture modality-common information as well (Fig. 1a), ignoring distinct modality-specific knowledge of the missing modality. As such, existing methods still struggle to effectively incorporate the learning of both common and specific knowledge of missing modalities." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 55, + 339, + 295, + 447 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 339, + 295, + 447 + ], + "spans": [ + { + "bbox": [ + 55, + 339, + 295, + 447 + ], + "type": "text", + "content": "Based on these insights, to bridge the aforementioned gap, we propose a Distilled Prompt Learning (DisPro) framework to unlock LLMs' robustness to incomplete modalities, which aims to simultaneously compensate for both specific and common information of missing modalities. It comprises two-stage prompt learning including Unimodal Prompting (UniPro) and Multimodal Prompting (MultiPro) for capturing modality-specific and modality-common information of missing modalities, respectively." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 55, + 449, + 295, + 652 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 449, + 295, + 652 + ], + "spans": [ + { + "bbox": [ + 55, + 449, + 295, + 652 + ], + "type": "text", + "content": "In the first stage, to prepare for supplementing modality-specific information of missing modalities, we first learn the knowledge distribution of the missing modality by distilling unimodal knowledge into a set of learnable unimodal prompts for each modality from all remaining samples available of this modality. Specifically, inspired by CoOp [44], a set of learnable prompts for different classes is employed to learn knowledge about how to describe different categories of an individual modality. For example, in survival prediction, learned prompts can be intuitively interpreted as how different risk groups will reflect in pathological morphology or gene expression profiles. However, due to the huge size (e.g., " + }, + { + "bbox": [ + 55, + 449, + 295, + 652 + ], + "type": "inline_equation", + "content": "100,000 \\times 100,000" + }, + { + "bbox": [ + 55, + 449, + 295, + 652 + ], + "type": "text", + "content": ") of pathological WSIs, we cannot simply apply CoOp designed for small natural images to WSIs. Therefore, in UniPro, we extend CoOp into the multiple instance learning (MIL) paradigm that is commonly used for large-scale WSIs [11, 13, 37]." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 55, + 654, + 295, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 654, + 295, + 713 + ], + "spans": [ + { + "bbox": [ + 55, + 654, + 295, + 713 + ], + "type": "text", + "content": "In the second stage, to capture modality-common information for missing modalities, we leverage available modalities as multimodal prompts of LLMs to infer representations of missing modalities using LLMs' strong inference ability. Simultaneously, to inject the learned unimodal" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 313, + 72, + 553, + 204 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 72, + 553, + 204 + ], + "spans": [ + { + "bbox": [ + 313, + 72, + 553, + 204 + ], + "type": "text", + "content": "knowledge in stage 1, we propose UniPro Distillation to enforce the inferred representations of missing modalities to get as close as possible to the learned unimodal prompts of the corresponding class. Additionally, to assist in learning modality-common knowledge from WSIs and genes with high information redundancy [30, 42], we propose UniPro Scoring to re-use UniPro as a grader to select discriminative and relevant tokens. As a result, both common and specific information are compensated for missing modalities under the collaboration of uni- and multimodal prompts, thus mitigating performance declines led by incomplete modalities." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 313, + 205, + 553, + 264 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 205, + 553, + 264 + ], + "spans": [ + { + "bbox": [ + 313, + 205, + 553, + 264 + ], + "type": "text", + "content": "It is worth noting that DisPro is proposed as a framework for addressing missing modality issues by leveraging LLMs' power based on prompts, capable of generalizing to a wider range of domains. The contributions of this work are summarized as follows:" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 314, + 266, + 553, + 397 + ], + "type": "list", + "angle": 0, + "index": 10, + "blocks": [ + { + "bbox": [ + 314, + 266, + 553, + 313 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 314, + 266, + 553, + 313 + ], + "spans": [ + { + "bbox": [ + 314, + 266, + 553, + 313 + ], + "type": "text", + "content": "- We propose a Distilled Prompt Learning (DisPro) framework for incomplete multimodal survival prediction, where LLMs' robustness to incomplete modalities is explored under two-stage prompting." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 315, + 314, + 553, + 361 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 315, + 314, + 553, + 361 + ], + "spans": [ + { + "bbox": [ + 315, + 314, + 553, + 361 + ], + "type": "text", + "content": "- We propose two-stage prompt learning including Unimodal Prompting and Multimodal Prompting to compensate for both modality-specific and modality-common information of missing modalities, respectively." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 315, + 361, + 553, + 397 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 315, + 361, + 553, + 397 + ], + "spans": [ + { + "bbox": [ + 315, + 361, + 553, + 397 + ], + "type": "text", + "content": "- Extensive experiments on five benchmark survival prediction datasets are conducted, which demonstrates significantly superior to state-of-the-art approaches." + } + ] + } + ], + "index": 9 + } + ], + "sub_type": "text" + }, + { + "bbox": [ + 314, + 411, + 400, + 423 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 314, + 411, + 400, + 423 + ], + "spans": [ + { + "bbox": [ + 314, + 411, + 400, + 423 + ], + "type": "text", + "content": "2. Related Work" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 313, + 432, + 493, + 445 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 432, + 493, + 445 + ], + "spans": [ + { + "bbox": [ + 313, + 432, + 493, + 445 + ], + "type": "text", + "content": "2.1. Incomplete Multimodal Learning" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 313, + 450, + 553, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 450, + 553, + 715 + ], + "spans": [ + { + "bbox": [ + 313, + 450, + 553, + 715 + ], + "type": "text", + "content": "As stated above, incomplete multimodal learning [43] can be categorized into two types: imputation-based [4, 12, 38] and imputation-free methods [21, 25]. For example, with the success of the diffusion model [7], the research [23] attempted to apply it to MRI synthesis for missing modalities. Instead of generating raw modality data, SMIL [22] leveraged Bayesian meta-learning to impute features of missing modalities. For retrieval-based approaches, M3Care [40] filled in the missing modalities by retrieving the most similar samples from the training set. On the other hand, imputation-free approaches provided more flexible solutions, which aim to enhance the robustness of fusion and alleviate performance decrease. [8] proposed a segmentation-based regularizer to reduce the sensitivity to missing modalities. MUSE [34] utilized patients' relationships based on graph networks to perform contrastive learning for improving the robustness to missing modalities. GTP-4o [19] proposed modality-prompted heterogeneous graph to embed each modality into a unified space and complete missing modality through a graph prompting mechanism. Recently, MAP [17] proposed missing-aware prompts to unlock the robustness of LLMs to missing" + } + ] + } + ], + "index": 13 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "text", + "content": "5103" + } + ] + } + ], + "index": 14 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 1 + }, + { + "para_blocks": [ + { + "bbox": [ + 55, + 72, + 295, + 109 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 72, + 295, + 109 + ], + "spans": [ + { + "bbox": [ + 55, + 72, + 295, + 109 + ], + "type": "text", + "content": "modalities by indicating LLMs different missing scenarios. However, these methods typically supplemented modality-common knowledge for missing modalities." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 55, + 117, + 228, + 129 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 117, + 228, + 129 + ], + "spans": [ + { + "bbox": [ + 55, + 117, + 228, + 129 + ], + "type": "text", + "content": "2.2. Multimodal Survival Prediction" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 55, + 134, + 296, + 375 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 134, + 296, + 375 + ], + "spans": [ + { + "bbox": [ + 55, + 134, + 296, + 375 + ], + "type": "text", + "content": "With the success of digital pathology [18, 24, 41] and sequencing techniques [26, 31], pathology and genomic data are commonly integrated to analyze survival outcomes, which have proven the superiority on precision in previous works [5, 36]. MCAT [5] proposed to leverage the interactions across modalities to guide the learning of pathology data. Based on this work, MOTCat [36] further leveraged optimal transport to capture global structure consistency across modalities. In parallel, CMTA [10] investigated the intrinsic cross-modal interactions and potential complementary information by cross-modal translation and alignment. From the perspective of information bottleneck, PIBD [42] proposed a prototypical information bottleneck for the MIL setting to select discriminative instances and disentangle multimodal knowledge for redundancy reduction. Similarly, MMP [30] utilized Gaussian Mixture Model to summarize morphological prototypes and biological pathway prototypes for these two modalities. However, all these methods assume all modalities available, inevitably leading to the difficulty of employment in clinical settings." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 55, + 386, + 111, + 398 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 386, + 111, + 398 + ], + "spans": [ + { + "bbox": [ + 55, + 386, + 111, + 398 + ], + "type": "text", + "content": "3. Method" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 55, + 406, + 296, + 502 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 406, + 296, + 502 + ], + "spans": [ + { + "bbox": [ + 55, + 406, + 296, + 502 + ], + "type": "text", + "content": "This work aims to explore LLMs' robustness to incomplete modalities by two-stage prompting, Unimodal Prompting (UniPro) and Multimodal Prompting (MultiPro), which capture modality-specific and modality-common information, respectively. The overview framework of the proposed method is shown in Fig. 2. First, we will introduce the problem formulation in Sec. 3.1. Then, we will detail the process of UniPro and MultiPro in Sec. 3.2 and 3.3, respectively." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 55, + 511, + 180, + 522 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 511, + 180, + 522 + ], + "spans": [ + { + "bbox": [ + 55, + 511, + 180, + 522 + ], + "type": "text", + "content": "3.1. Problem Formulation" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 55, + 529, + 295, + 590 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 529, + 295, + 590 + ], + "spans": [ + { + "bbox": [ + 55, + 529, + 295, + 590 + ], + "type": "text", + "content": "Given a multimodal dataset consisting of " + }, + { + "bbox": [ + 55, + 529, + 295, + 590 + ], + "type": "inline_equation", + "content": "M = 2" + }, + { + "bbox": [ + 55, + 529, + 295, + 590 + ], + "type": "text", + "content": " modalities, we use a 2-tuple " + }, + { + "bbox": [ + 55, + 529, + 295, + 590 + ], + "type": "inline_equation", + "content": "\\mathbf{X}_n = (\\mathbf{X}_n^p,\\mathbf{X}_n^g)" + }, + { + "bbox": [ + 55, + 529, + 295, + 590 + ], + "type": "text", + "content": " to represent the " + }, + { + "bbox": [ + 55, + 529, + 295, + 590 + ], + "type": "inline_equation", + "content": "i" + }, + { + "bbox": [ + 55, + 529, + 295, + 590 + ], + "type": "text", + "content": "-th patient data, where " + }, + { + "bbox": [ + 55, + 529, + 295, + 590 + ], + "type": "inline_equation", + "content": "\\mathbf{X}_n^p" + }, + { + "bbox": [ + 55, + 529, + 295, + 590 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 55, + 529, + 295, + 590 + ], + "type": "inline_equation", + "content": "\\mathbf{X}_n^g" + }, + { + "bbox": [ + 55, + 529, + 295, + 590 + ], + "type": "text", + "content": " are pathology and genomic data, respectively. In particular, we use the notation of tilde hat to indicate the modality is missing, e.g., " + }, + { + "bbox": [ + 55, + 529, + 295, + 590 + ], + "type": "inline_equation", + "content": "\\widetilde{\\mathbf{X}}_n^p" + }, + { + "bbox": [ + 55, + 529, + 295, + 590 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 55, + 591, + 296, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 591, + 296, + 714 + ], + "spans": [ + { + "bbox": [ + 55, + 591, + 296, + 714 + ], + "type": "text", + "content": "Following the conventional setting for pathological WSIs and genomic pathways in previous works [14, 42], we formulate a WSI " + }, + { + "bbox": [ + 55, + 591, + 296, + 714 + ], + "type": "inline_equation", + "content": "\\mathbf{X}_n^p" + }, + { + "bbox": [ + 55, + 591, + 296, + 714 + ], + "type": "text", + "content": " and a set of genomic pathways " + }, + { + "bbox": [ + 55, + 591, + 296, + 714 + ], + "type": "inline_equation", + "content": "\\mathbf{X}_n^g" + }, + { + "bbox": [ + 55, + 591, + 296, + 714 + ], + "type": "text", + "content": " as the \"bag\" of instances in MIL paradigm. We denote " + }, + { + "bbox": [ + 55, + 591, + 296, + 714 + ], + "type": "inline_equation", + "content": "\\mathbf{X}_n^p = \\{\\mathbf{x}_{n,i}^p\\}_{i=1}^{M_p}" + }, + { + "bbox": [ + 55, + 591, + 296, + 714 + ], + "type": "text", + "content": " where each instance " + }, + { + "bbox": [ + 55, + 591, + 296, + 714 + ], + "type": "inline_equation", + "content": "\\mathbf{x}_{n,i}^p" + }, + { + "bbox": [ + 55, + 591, + 296, + 714 + ], + "type": "text", + "content": " refers to a patch image cropped from the WSI " + }, + { + "bbox": [ + 55, + 591, + 296, + 714 + ], + "type": "inline_equation", + "content": "\\mathbf{X}_n^p" + }, + { + "bbox": [ + 55, + 591, + 296, + 714 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 55, + 591, + 296, + 714 + ], + "type": "inline_equation", + "content": "M_p" + }, + { + "bbox": [ + 55, + 591, + 296, + 714 + ], + "type": "text", + "content": " is the patch number of this WSI. Similarly, the bag of genomic pathways can be formulated as " + }, + { + "bbox": [ + 55, + 591, + 296, + 714 + ], + "type": "inline_equation", + "content": "\\mathbf{X}_n^g = \\{\\mathbf{x}_{n,i}^g\\}_{i=1}^{M_g}" + }, + { + "bbox": [ + 55, + 591, + 296, + 714 + ], + "type": "text", + "content": ", in which an instance " + }, + { + "bbox": [ + 55, + 591, + 296, + 714 + ], + "type": "inline_equation", + "content": "\\mathbf{x}_{n,i}^g" + }, + { + "bbox": [ + 55, + 591, + 296, + 714 + ], + "type": "text", + "content": " represents a pathway consisting of various gene expression values." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 313, + 72, + 553, + 228 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 72, + 553, + 228 + ], + "spans": [ + { + "bbox": [ + 313, + 72, + 553, + 228 + ], + "type": "text", + "content": "In survival prediction, the goal is to estimate the risk probability of an outcome event taking place before a specific time. However, this event may not always be observed, resulting in a right-censored occurrence. Therefore, " + }, + { + "bbox": [ + 313, + 72, + 553, + 228 + ], + "type": "inline_equation", + "content": "c_{n} \\in \\{0,1\\}" + }, + { + "bbox": [ + 313, + 72, + 553, + 228 + ], + "type": "text", + "content": " is used to denote censorship status indicating whether the outcome event is right-censored (i.e., alive) (" + }, + { + "bbox": [ + 313, + 72, + 553, + 228 + ], + "type": "inline_equation", + "content": "c = 1" + }, + { + "bbox": [ + 313, + 72, + 553, + 228 + ], + "type": "text", + "content": ") or not (" + }, + { + "bbox": [ + 313, + 72, + 553, + 228 + ], + "type": "inline_equation", + "content": "c = 0" + }, + { + "bbox": [ + 313, + 72, + 553, + 228 + ], + "type": "text", + "content": ") (i.e., dead), and " + }, + { + "bbox": [ + 313, + 72, + 553, + 228 + ], + "type": "inline_equation", + "content": "t_{n} \\in \\mathbb{R}^{+}" + }, + { + "bbox": [ + 313, + 72, + 553, + 228 + ], + "type": "text", + "content": " represents the overall survival time (in months). These two components make up its set of labels " + }, + { + "bbox": [ + 313, + 72, + 553, + 228 + ], + "type": "inline_equation", + "content": "Y_{n} = (c_{n},t_{n})" + }, + { + "bbox": [ + 313, + 72, + 553, + 228 + ], + "type": "text", + "content": ". Following the previous setting [5, 36], the continuous survival time would be discretized into " + }, + { + "bbox": [ + 313, + 72, + 553, + 228 + ], + "type": "inline_equation", + "content": "I_{t}" + }, + { + "bbox": [ + 313, + 72, + 553, + 228 + ], + "type": "text", + "content": " time intervals corresponding to " + }, + { + "bbox": [ + 313, + 72, + 553, + 228 + ], + "type": "inline_equation", + "content": "I_{t}" + }, + { + "bbox": [ + 313, + 72, + 553, + 228 + ], + "type": "text", + "content": " risk bands. As a result, the label set " + }, + { + "bbox": [ + 313, + 72, + 553, + 228 + ], + "type": "inline_equation", + "content": "(c_{n},t_{n})" + }, + { + "bbox": [ + 313, + 72, + 553, + 228 + ], + "type": "text", + "content": " can be re-written as " + }, + { + "bbox": [ + 313, + 72, + 553, + 228 + ], + "type": "inline_equation", + "content": "y_{n}" + }, + { + "bbox": [ + 313, + 72, + 553, + 228 + ], + "type": "text", + "content": ", resulting in " + }, + { + "bbox": [ + 313, + 72, + 553, + 228 + ], + "type": "inline_equation", + "content": "2I_{t}" + }, + { + "bbox": [ + 313, + 72, + 553, + 228 + ], + "type": "text", + "content": " labels." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 313, + 228, + 553, + 287 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 228, + 553, + 287 + ], + "spans": [ + { + "bbox": [ + 313, + 228, + 553, + 287 + ], + "type": "text", + "content": "Given a set of predicted hazard probabilities " + }, + { + "bbox": [ + 313, + 228, + 553, + 287 + ], + "type": "inline_equation", + "content": "\\mathbf{h}_n" + }, + { + "bbox": [ + 313, + 228, + 553, + 287 + ], + "type": "text", + "content": " across all time intervals, the hazard probability " + }, + { + "bbox": [ + 313, + 228, + 553, + 287 + ], + "type": "inline_equation", + "content": "\\mathbf{h}_n^{(j)}" + }, + { + "bbox": [ + 313, + 228, + 553, + 287 + ], + "type": "text", + "content": " represents the probability of death event occurring at the " + }, + { + "bbox": [ + 313, + 228, + 553, + 287 + ], + "type": "inline_equation", + "content": "j" + }, + { + "bbox": [ + 313, + 228, + 553, + 287 + ], + "type": "text", + "content": "-th time interval. Then, the cumulative survival probability for each time interval can be estimated as follows:" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 389, + 296, + 553, + 328 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 389, + 296, + 553, + 328 + ], + "spans": [ + { + "bbox": [ + 389, + 296, + 553, + 328 + ], + "type": "interline_equation", + "content": "S _ {n} ^ {(j)} = \\prod_ {z = 1} ^ {j} \\left(1 - \\mathbf {h} _ {n} ^ {(z)}\\right) \\tag {1}", + "image_path": "90b7973c7ab6de4bac76979b6ceac242024e7727d97eb34dcfe2827f87aa357c.jpg" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 313, + 331, + 553, + 368 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 331, + 553, + 368 + ], + "spans": [ + { + "bbox": [ + 313, + 331, + 553, + 368 + ], + "type": "text", + "content": "Following previous works [36, 42], given the ground truth time interval " + }, + { + "bbox": [ + 313, + 331, + 553, + 368 + ], + "type": "inline_equation", + "content": "\\tau" + }, + { + "bbox": [ + 313, + 331, + 553, + 368 + ], + "type": "text", + "content": ", we use the NLL loss [39] for survival prediction, where " + }, + { + "bbox": [ + 313, + 331, + 553, + 368 + ], + "type": "inline_equation", + "content": "N" + }, + { + "bbox": [ + 313, + 331, + 553, + 368 + ], + "type": "text", + "content": " is the sample number of the training set:" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 326, + 376, + 553, + 427 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 326, + 376, + 553, + 427 + ], + "spans": [ + { + "bbox": [ + 326, + 376, + 553, + 427 + ], + "type": "interline_equation", + "content": "\\begin{array}{l} \\mathcal {L} _ {s u r v} = - \\sum_ {n = 1} ^ {N} \\left[ c _ {n} \\log S _ {n} ^ {(\\tau)} + \\left(1 - c _ {n}\\right) \\log S _ {n} ^ {(\\tau - 1)} \\right. \\tag {2} \\\\ \\left. + \\left(1 - c _ {n}\\right) \\log {\\bf h} _ {n} ^ {(\\tau)} \\right] \\\\ \\end{array}", + "image_path": "5d574ed947237d3e1d85c3baab86688630230930f41789e728f5ee6831282668.jpg" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 313, + 436, + 522, + 449 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 436, + 522, + 449 + ], + "spans": [ + { + "bbox": [ + 313, + 436, + 522, + 449 + ], + "type": "text", + "content": "3.2. Stage 1: Unimodal Prompting (UniPro)" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 313, + 453, + 553, + 530 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 453, + 553, + 530 + ], + "spans": [ + { + "bbox": [ + 313, + 453, + 553, + 530 + ], + "type": "text", + "content": "Given a WSI " + }, + { + "bbox": [ + 313, + 453, + 553, + 530 + ], + "type": "inline_equation", + "content": "\\mathbf{X}_n^p = \\{\\mathbf{x}_{n,i}^p\\}_{i=1}^{M_p}" + }, + { + "bbox": [ + 313, + 453, + 553, + 530 + ], + "type": "text", + "content": " or a bag of genomic pathways " + }, + { + "bbox": [ + 313, + 453, + 553, + 530 + ], + "type": "inline_equation", + "content": "\\mathbf{X}_n^g = \\{\\mathbf{x}_{n,i}^g\\}_{i=1}^{M_g}" + }, + { + "bbox": [ + 313, + 453, + 553, + 530 + ], + "type": "text", + "content": ", we formulate their learning process as a MIL task. Inspired by CoOp [44] employing a CLIP-style architecture, we adapt it to the MIL setting to learn how to characterize each modality using a set of learnable prompts for modality-specific knowledge." + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 313, + 531, + 554, + 653 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 531, + 554, + 653 + ], + "spans": [ + { + "bbox": [ + 313, + 531, + 554, + 653 + ], + "type": "text", + "content": "In terms of pathological WSIs data, following the common setting of the WSI processing [28], as shown in the UniPro-P of Fig 2, we first employ a pretrained and frozen patch encoder " + }, + { + "bbox": [ + 313, + 531, + 554, + 653 + ], + "type": "inline_equation", + "content": "g_{E}(\\cdot)" + }, + { + "bbox": [ + 313, + 531, + 554, + 653 + ], + "type": "text", + "content": " to embed every patch images " + }, + { + "bbox": [ + 313, + 531, + 554, + 653 + ], + "type": "inline_equation", + "content": "\\pmb{x}_{n,i}^{p}" + }, + { + "bbox": [ + 313, + 531, + 554, + 653 + ], + "type": "text", + "content": " from a WSI into an instance-level embedding " + }, + { + "bbox": [ + 313, + 531, + 554, + 653 + ], + "type": "inline_equation", + "content": "p_n^i = g_E(\\pmb{x}_{n,i}^p) \\in \\mathbb{R}^d" + }, + { + "bbox": [ + 313, + 531, + 554, + 653 + ], + "type": "text", + "content": ", where " + }, + { + "bbox": [ + 313, + 531, + 554, + 653 + ], + "type": "inline_equation", + "content": "d" + }, + { + "bbox": [ + 313, + 531, + 554, + 653 + ], + "type": "text", + "content": " is the patch feature dimension. To align pathology features with the text space, we apply an adapter " + }, + { + "bbox": [ + 313, + 531, + 554, + 653 + ], + "type": "inline_equation", + "content": "f_{p}(\\cdot)" + }, + { + "bbox": [ + 313, + 531, + 554, + 653 + ], + "type": "text", + "content": " for each patch feature to get the aligned patch feature " + }, + { + "bbox": [ + 313, + 531, + 554, + 653 + ], + "type": "inline_equation", + "content": "\\mathbf{p}_n^i = f_p(g_E(\\pmb{x}_{n,i}^p)) \\in \\mathbb{R}^D" + }, + { + "bbox": [ + 313, + 531, + 554, + 653 + ], + "type": "text", + "content": ", implemented by a linear layer followed by ReLU [2]." + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 313, + 654, + 553, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 654, + 553, + 714 + ], + "spans": [ + { + "bbox": [ + 313, + 654, + 553, + 714 + ], + "type": "text", + "content": "Text-wise, the input comprises three components: pathology context tokens, prefix tokens andclassname tokens. Forclassname tokens " + }, + { + "bbox": [ + 313, + 654, + 553, + 714 + ], + "type": "inline_equation", + "content": "[C]_{1\\dots s}" + }, + { + "bbox": [ + 313, + 654, + 553, + 714 + ], + "type": "text", + "content": ", we map every time interval of overall survival into a unique description, such as \"median risk\", and denote the censorship status by \"alive\"" + } + ] + } + ], + "index": 16 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "text", + "content": "5104" + } + ] + } + ], + "index": 17 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 2 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 70, + 88, + 294, + 223 + ], + "blocks": [ + { + "bbox": [ + 226, + 76, + 398, + 87 + ], + "lines": [ + { + "bbox": [ + 226, + 76, + 398, + 87 + ], + "spans": [ + { + "bbox": [ + 226, + 76, + 398, + 87 + ], + "type": "text", + "content": "Stage 1: Unimodal Prompting (UniPro)" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 70, + 88, + 294, + 223 + ], + "lines": [ + { + "bbox": [ + 70, + 88, + 294, + 223 + ], + "spans": [ + { + "bbox": [ + 70, + 88, + 294, + 223 + ], + "type": "image", + "image_path": "c9d0d3110b8395a40f7190f7be8aaf2b8a09e9428d281820174d23735f364bc1.jpg" + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_body" + } + ], + "index": 1 + }, + { + "type": "image", + "bbox": [ + 304, + 88, + 533, + 220 + ], + "blocks": [ + { + "bbox": [ + 304, + 88, + 533, + 220 + ], + "lines": [ + { + "bbox": [ + 304, + 88, + 533, + 220 + ], + "spans": [ + { + "bbox": [ + 304, + 88, + 533, + 220 + ], + "type": "image", + "image_path": "42bfdeca25ae2ec5fdb16a3b3239c882f7290e04e8e0ef29aad92b99bd4fe3fb.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + } + ], + "index": 2 + }, + { + "type": "image", + "bbox": [ + 71, + 224, + 266, + 408 + ], + "blocks": [ + { + "bbox": [ + 71, + 224, + 266, + 408 + ], + "lines": [ + { + "bbox": [ + 71, + 224, + 266, + 408 + ], + "spans": [ + { + "bbox": [ + 71, + 224, + 266, + 408 + ], + "type": "image", + "image_path": "789b7be1dbdc1a4a23d8857ae7e979c760ba2645578cf76a799657f73a7189ef.jpg" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 86, + 408, + 126, + 420 + ], + "lines": [ + { + "bbox": [ + 86, + 408, + 126, + 420 + ], + "spans": [ + { + "bbox": [ + 86, + 408, + 126, + 420 + ], + "type": "text", + "content": "d.1. " + }, + { + "bbox": [ + 86, + 408, + 126, + 420 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{\\mathrm{surv}}^{\\mathrm{cls}}" + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_footnote" + }, + { + "bbox": [ + 80, + 422, + 264, + 434 + ], + "lines": [ + { + "bbox": [ + 80, + 422, + 264, + 434 + ], + "spans": [ + { + "bbox": [ + 80, + 422, + 264, + 434 + ], + "type": "text", + "content": "Stage 2: Multimodal Prompting (MultiPro)" + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "image_caption" + } + ], + "index": 3 + }, + { + "type": "image", + "bbox": [ + 272, + 224, + 536, + 342 + ], + "blocks": [ + { + "bbox": [ + 272, + 224, + 536, + 342 + ], + "lines": [ + { + "bbox": [ + 272, + 224, + 536, + 342 + ], + "spans": [ + { + "bbox": [ + 272, + 224, + 536, + 342 + ], + "type": "image", + "image_path": "04683aa816c6afdce9c8c5cd6cc98c9abc1b9f6f554670ef5d86a633349677a8.jpg" + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "image_body" + } + ], + "index": 6 + }, + { + "type": "image", + "bbox": [ + 272, + 345, + 416, + 432 + ], + "blocks": [ + { + "bbox": [ + 272, + 345, + 416, + 432 + ], + "lines": [ + { + "bbox": [ + 272, + 345, + 416, + 432 + ], + "spans": [ + { + "bbox": [ + 272, + 345, + 416, + 432 + ], + "type": "image", + "image_path": "f6fa08541a803969da1d94c9a70c4f5638350c2c1c413bf4593afc11db48c9e2.jpg" + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 54, + 441, + 555, + 498 + ], + "lines": [ + { + "bbox": [ + 54, + 441, + 555, + 498 + ], + "spans": [ + { + "bbox": [ + 54, + 441, + 555, + 498 + ], + "type": "text", + "content": "Figure 2. Overview of DisPro. Stage 1: Unimodal Prompting aims to distill the knowledge distribution for each modality and prepare to supplement modality-specific knowledge for the missing modality. Stage 2: Multimodal Prompting utilizes the available modality as prompts to infer the representations of the missing one, compensating for modality-common knowledge. Simultaneously, the learned UniPro of Stage 1 supervises the learning of imputed representations to compensate for modality-specific knowledge. UniPro Scoring re-uses learned prompts to assist the LLM in capturing modality-common knowledge by selecting discriminative and relevant tokens." + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "image_caption" + } + ], + "index": 7 + }, + { + "type": "image", + "bbox": [ + 416, + 358, + 537, + 428 + ], + "blocks": [ + { + "bbox": [ + 416, + 358, + 537, + 428 + ], + "lines": [ + { + "bbox": [ + 416, + 358, + 537, + 428 + ], + "spans": [ + { + "bbox": [ + 416, + 358, + 537, + 428 + ], + "type": "image", + "image_path": "5b5c49cdbcaf79c879a2444571ccf23c50a90f1fdc1c2d8725afb85c71c6fad1.jpg" + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "image_body" + } + ], + "index": 8 + }, + { + "bbox": [ + 54, + 509, + 295, + 605 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 54, + 509, + 295, + 605 + ], + "spans": [ + { + "bbox": [ + 54, + 509, + 295, + 605 + ], + "type": "text", + "content": "or \"dead\". The prefix tokens " + }, + { + "bbox": [ + 54, + 509, + 295, + 605 + ], + "type": "inline_equation", + "content": "\\left[\\mathrm{PREFIX}\\right]_{1\\dots m}" + }, + { + "bbox": [ + 54, + 509, + 295, + 605 + ], + "type": "text", + "content": " briefly describe the modality, e.g., \"This is a pathology slide image from the patient with overall survival of " + }, + { + "bbox": [ + 54, + 509, + 295, + 605 + ], + "type": "inline_equation", + "content": "[C]_{1\\dots s}" + }, + { + "bbox": [ + 54, + 509, + 295, + 605 + ], + "type": "text", + "content": "\". Prefix and classname tokens are obtained using a frozen Tokenizer pretrained on a large-scale biomedical corpus to embed these words. To learn how to describe the modality for a specific risk band (class), we attach a set of learnable context tokens " + }, + { + "bbox": [ + 54, + 509, + 295, + 605 + ], + "type": "inline_equation", + "content": "[P]_{1\\dots k}" + }, + { + "bbox": [ + 54, + 509, + 295, + 605 + ], + "type": "text", + "content": " to prefix and classname tokens:" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 58, + 609, + 295, + 623 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 58, + 609, + 295, + 623 + ], + "spans": [ + { + "bbox": [ + 58, + 609, + 295, + 623 + ], + "type": "interline_equation", + "content": "\\boldsymbol {T} _ {p} = [ P ] _ {1} \\dots [ P ] _ {k} [ \\text {P R E F I X} ] _ {1} \\dots [ \\text {P R E F I X} ] _ {m} [ C ] _ {1} \\dots [ C ] _ {s} \\tag {3}", + "image_path": "f40a2460f7441605d934dec72ddc3f864e84a2ba0c73975b30877eeeaef27753.jpg" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 55, + 624, + 295, + 684 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 624, + 295, + 684 + ], + "spans": [ + { + "bbox": [ + 55, + 624, + 295, + 684 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 55, + 624, + 295, + 684 + ], + "type": "inline_equation", + "content": "k" + }, + { + "bbox": [ + 55, + 624, + 295, + 684 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 55, + 624, + 295, + 684 + ], + "type": "inline_equation", + "content": "m" + }, + { + "bbox": [ + 55, + 624, + 295, + 684 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 55, + 624, + 295, + 684 + ], + "type": "inline_equation", + "content": "s" + }, + { + "bbox": [ + 55, + 624, + 295, + 684 + ], + "type": "text", + "content": " are the token numbers of pathology context tokens, prefix words and classname words, respectively. Each token is a vector with the same dimension " + }, + { + "bbox": [ + 55, + 624, + 295, + 684 + ], + "type": "inline_equation", + "content": "D" + }, + { + "bbox": [ + 55, + 624, + 295, + 684 + ], + "type": "text", + "content": " as word embeddings of the LLM (i.e., 768 for BERT). In this way, we can design input tokens for each risk band:" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 55, + 689, + 301, + 714 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 689, + 301, + 714 + ], + "spans": [ + { + "bbox": [ + 55, + 689, + 301, + 714 + ], + "type": "interline_equation", + "content": "\\boldsymbol {T} _ {p} ^ {(j)} = [ P ] _ {1} ^ {(j)} \\dots [ P ] _ {k} ^ {(j)} [ \\text {P R E F I X} ] _ {1} \\dots [ \\text {P R E F I X} ] _ {m} [ C ] _ {1} ^ {(j)} \\dots [ C ] _ {s} ^ {(j)} \\tag {4}", + "image_path": "8883f22632690561984289d89ab24e1a70d0511ac1a8fedd3c5f9a499ff25a29.jpg" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 313, + 509, + 555, + 561 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 509, + 555, + 561 + ], + "spans": [ + { + "bbox": [ + 313, + 509, + 555, + 561 + ], + "type": "text", + "content": "in which " + }, + { + "bbox": [ + 313, + 509, + 555, + 561 + ], + "type": "inline_equation", + "content": "j \\in [1, 2I_t]" + }, + { + "bbox": [ + 313, + 509, + 555, + 561 + ], + "type": "text", + "content": " and prefix tokens are shared across all risk bands. By forwarding " + }, + { + "bbox": [ + 313, + 509, + 555, + 561 + ], + "type": "inline_equation", + "content": "T_p^{(j)}" + }, + { + "bbox": [ + 313, + 509, + 555, + 561 + ], + "type": "text", + "content": " of each class to the text encoder " + }, + { + "bbox": [ + 313, + 509, + 555, + 561 + ], + "type": "inline_equation", + "content": "\\mathcal{T}(\\cdot)" + }, + { + "bbox": [ + 313, + 509, + 555, + 561 + ], + "type": "text", + "content": " (i.e., LLM), we can obtain a global textual class representation " + }, + { + "bbox": [ + 313, + 509, + 555, + 561 + ], + "type": "inline_equation", + "content": "t_p^{(j)} \\in \\mathbb{R}^D" + }, + { + "bbox": [ + 313, + 509, + 555, + 561 + ], + "type": "text", + "content": " for every class." + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 313, + 566, + 555, + 626 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 566, + 555, + 626 + ], + "spans": [ + { + "bbox": [ + 313, + 566, + 555, + 626 + ], + "type": "text", + "content": "In the MIL setting, to obtain the predicted probability for each class, we can use topK patches to produce the slide-level prediction based on similarity scores. To this end, we first compute the similarity score between every instance and every class:" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 399, + 641, + 555, + 658 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 399, + 641, + 555, + 658 + ], + "spans": [ + { + "bbox": [ + 399, + 641, + 555, + 658 + ], + "type": "interline_equation", + "content": "\\boldsymbol {s} _ {n, p} ^ {(i, j)} = \\boldsymbol {p} _ {n} ^ {i} \\cdot \\boldsymbol {t} _ {p} ^ {(j)} \\tag {5}", + "image_path": "261802045f31465fbac6ed97427f7c4e4c7a1b2d4517c7432ff216481192d467.jpg" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 313, + 677, + 555, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 677, + 555, + 715 + ], + "spans": [ + { + "bbox": [ + 313, + 677, + 555, + 715 + ], + "type": "text", + "content": "As a result, we can get a similarity matrix " + }, + { + "bbox": [ + 313, + 677, + 555, + 715 + ], + "type": "inline_equation", + "content": "S_{n,p} \\in \\mathbb{R}^{M_p \\times 2I_t}" + }, + { + "bbox": [ + 313, + 677, + 555, + 715 + ], + "type": "text", + "content": " for a WSI. Then, we apply a TopK max-pooling operator " + }, + { + "bbox": [ + 313, + 677, + 555, + 715 + ], + "type": "inline_equation", + "content": "h(\\cdot)" + }, + { + "bbox": [ + 313, + 677, + 555, + 715 + ], + "type": "text", + "content": " to obtain the slide-level predictions, depend" + } + ] + } + ], + "index": 17 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "text", + "content": "5105" + } + ] + } + ], + "index": 18 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 3 + }, + { + "para_blocks": [ + { + "bbox": [ + 55, + 72, + 217, + 84 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 72, + 217, + 84 + ], + "spans": [ + { + "bbox": [ + 55, + 72, + 217, + 84 + ], + "type": "text", + "content": "ing on top-K similarities for each class:" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 65, + 89, + 294, + 132 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 65, + 89, + 294, + 132 + ], + "spans": [ + { + "bbox": [ + 65, + 89, + 294, + 132 + ], + "type": "interline_equation", + "content": "h \\left(\\mathcal {S} _ {n, p}\\right) = \\frac {1}{K} \\left[ \\sum_ {i = 1} ^ {K} \\hat {\\boldsymbol {s}} _ {n, p} ^ {(i, 1)}, \\sum_ {i = 1} ^ {K} \\hat {\\boldsymbol {s}} _ {n, p} ^ {(i, 2)}, \\dots , \\sum_ {i = 1} ^ {K} \\hat {\\boldsymbol {s}} _ {n, p} ^ {(i, 2 \\boldsymbol {I} _ {t})} \\right] \\tag {6}", + "image_path": "5c0ef54590d2ff26d0133b093a0369a4ead147aa761d7d47c1c422fb1f2ed4af.jpg" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 55, + 133, + 295, + 169 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 133, + 295, + 169 + ], + "spans": [ + { + "bbox": [ + 55, + 133, + 295, + 169 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 55, + 133, + 295, + 169 + ], + "type": "inline_equation", + "content": "\\hat{s}_{n,p}" + }, + { + "bbox": [ + 55, + 133, + 295, + 169 + ], + "type": "text", + "content": " is the sorted " + }, + { + "bbox": [ + 55, + 133, + 295, + 169 + ], + "type": "inline_equation", + "content": "s_{n,p}" + }, + { + "bbox": [ + 55, + 133, + 295, + 169 + ], + "type": "text", + "content": " in a descending order across all instances. Next, we can get a set of hazard probabilities " + }, + { + "bbox": [ + 55, + 133, + 295, + 169 + ], + "type": "inline_equation", + "content": "\\mathbf{h}_n^p" + }, + { + "bbox": [ + 55, + 133, + 295, + 169 + ], + "type": "text", + "content": " across all time intervals through a sigmoid function." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 124, + 175, + 295, + 190 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 124, + 175, + 295, + 190 + ], + "spans": [ + { + "bbox": [ + 124, + 175, + 295, + 190 + ], + "type": "interline_equation", + "content": "\\mathbf {h} _ {n} ^ {p} = \\operatorname {S i g m o i d} \\left(h \\left(S _ {n, p}\\right)\\right) \\tag {7}", + "image_path": "174d92d97e04d26591a964993655dd8f3f0bf2afbe2d93872c81a4cef8cce861.jpg" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 55, + 194, + 295, + 230 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 194, + 295, + 230 + ], + "spans": [ + { + "bbox": [ + 55, + 194, + 295, + 230 + ], + "type": "text", + "content": "By computing the survival loss according to Eq. 2 and backforwarding gradients, we can optimize the modality adapter and learnable prompts consisting of context tokens." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 55, + 230, + 296, + 291 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 230, + 296, + 291 + ], + "spans": [ + { + "bbox": [ + 55, + 230, + 296, + 291 + ], + "type": "text", + "content": "Similarly, we acquire the set of hazard probabilities " + }, + { + "bbox": [ + 55, + 230, + 296, + 291 + ], + "type": "inline_equation", + "content": "\\mathbf{h}_n^g" + }, + { + "bbox": [ + 55, + 230, + 296, + 291 + ], + "type": "text", + "content": " for genomic pathways and optimize its learnable prompts " + }, + { + "bbox": [ + 55, + 230, + 296, + 291 + ], + "type": "inline_equation", + "content": "T_{g}" + }, + { + "bbox": [ + 55, + 230, + 296, + 291 + ], + "type": "text", + "content": " and gene encoder " + }, + { + "bbox": [ + 55, + 230, + 296, + 291 + ], + "type": "inline_equation", + "content": "f_{g}(\\cdot)" + }, + { + "bbox": [ + 55, + 230, + 296, + 291 + ], + "type": "text", + "content": ". Here we use the subscript " + }, + { + "bbox": [ + 55, + 230, + 296, + 291 + ], + "type": "inline_equation", + "content": "p" + }, + { + "bbox": [ + 55, + 230, + 296, + 291 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 55, + 230, + 296, + 291 + ], + "type": "inline_equation", + "content": "g" + }, + { + "bbox": [ + 55, + 230, + 296, + 291 + ], + "type": "text", + "content": " to represent pathology and genomic data. Every pathway is also encoded into a " + }, + { + "bbox": [ + 55, + 230, + 296, + 291 + ], + "type": "inline_equation", + "content": "D" + }, + { + "bbox": [ + 55, + 230, + 296, + 291 + ], + "type": "text", + "content": "-dimensional embedding." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 55, + 296, + 282, + 309 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 296, + 282, + 309 + ], + "spans": [ + { + "bbox": [ + 55, + 296, + 282, + 309 + ], + "type": "text", + "content": "3.3. Stage 2: Multimodal Prompting (MultiPro)" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 55, + 313, + 296, + 361 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 313, + 296, + 361 + ], + "spans": [ + { + "bbox": [ + 55, + 313, + 296, + 361 + ], + "type": "text", + "content": "In this stage, we aim to leverage LLMs' strong inference ability and robustness to missing modalities to predict the representation of missing modalities, as shown in the MultiPro of Fig. 2." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 55, + 361, + 296, + 480 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 361, + 296, + 480 + ], + "spans": [ + { + "bbox": [ + 55, + 361, + 296, + 480 + ], + "type": "text", + "content": "To simplify explanations, we take the example input of pathology available and genomics being missing. By prompting the LLM with available pathology data, the Bert-style LLM can capture modality-common knowledge for genomics due to its nature of self-attention. To compensate for modality-specific knowledge, we re-use the unimodal prompts " + }, + { + "bbox": [ + 55, + 361, + 296, + 480 + ], + "type": "inline_equation", + "content": "T_{g}" + }, + { + "bbox": [ + 55, + 361, + 296, + 480 + ], + "type": "text", + "content": " learned in stage 1 with all remaining samples available of genomics data to supervise the learning of imputed representations, where unimodal prompts characterize the knowledge distribution of an individual modality." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 55, + 481, + 296, + 618 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 481, + 296, + 618 + ], + "spans": [ + { + "bbox": [ + 55, + 481, + 296, + 618 + ], + "type": "text", + "content": "To this end, we first need to construct an interleaved input for the learning of modality-common information. To fully leverage the capabilities of LLMs, every modality are supposed to be embedded into the aligned text space. Consequently, we re-use the aligned modality adapter or encoder learned in the first stage to embed each modality, and we continue to finetune them in the second stage. Note that the LLM used in stage 2 is the same as the one in stage 1. As a result, we can get a bag of aligned feature tokens " + }, + { + "bbox": [ + 55, + 481, + 296, + 618 + ], + "type": "inline_equation", + "content": "\\{\\mathbf{p}_n^i\\}_{i = 1}^{M_p}" + }, + { + "bbox": [ + 55, + 481, + 296, + 618 + ], + "type": "text", + "content": " or " + }, + { + "bbox": [ + 55, + 481, + 296, + 618 + ], + "type": "inline_equation", + "content": "\\{\\mathbf{g}_n^i\\}_{i = 1}^{M_g}" + }, + { + "bbox": [ + 55, + 481, + 296, + 618 + ], + "type": "text", + "content": " for pathology and genomic data, respectively. Here we assume genomic data " + }, + { + "bbox": [ + 55, + 481, + 296, + 618 + ], + "type": "inline_equation", + "content": "\\{\\tilde{\\mathbf{g}}_n^i\\}_{i = 1}^{M_g}" + }, + { + "bbox": [ + 55, + 481, + 296, + 618 + ], + "type": "text", + "content": " is missing." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 55, + 618, + 296, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 618, + 296, + 713 + ], + "spans": [ + { + "bbox": [ + 55, + 618, + 296, + 713 + ], + "type": "text", + "content": "Due to the limitation of input length for general LLMs (e.g., 512 for Bert-style LLMs) and high information redundancy in WSIs or genomic pathways [30, 42], more discriminative and relevant feature tokens are expected to be selected for better capturing modality-common knowledge. Therefore, we propose a UniPro Scoring module by re-using UniPro as a grader to estimate a relevance score for each feature token." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 313, + 72, + 553, + 108 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 72, + 553, + 108 + ], + "spans": [ + { + "bbox": [ + 313, + 72, + 553, + 108 + ], + "type": "text", + "content": "UniPro Scoring. The score of every feature token is determined by three parts: relevance scores estimated by UniPro-P and UniPro-G, and self-scoring, as shown in Fig. 2." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 313, + 109, + 555, + 312 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 109, + 555, + 312 + ], + "spans": [ + { + "bbox": [ + 313, + 109, + 555, + 312 + ], + "type": "text", + "content": "Given a set of pathology feature tokens " + }, + { + "bbox": [ + 313, + 109, + 555, + 312 + ], + "type": "inline_equation", + "content": "\\{\\mathbf{p}_n^i\\}_{i = 1}^{M_p}" + }, + { + "bbox": [ + 313, + 109, + 555, + 312 + ], + "type": "text", + "content": ", to obtain discriminative tokens, we compute the similarity score " + }, + { + "bbox": [ + 313, + 109, + 555, + 312 + ], + "type": "inline_equation", + "content": "s_{n,p}^{(i,\\tau)}" + }, + { + "bbox": [ + 313, + 109, + 555, + 312 + ], + "type": "text", + "content": " between the ground-truth textual class representation of " + }, + { + "bbox": [ + 313, + 109, + 555, + 312 + ], + "type": "inline_equation", + "content": "t_p^{(\\tau)}" + }, + { + "bbox": [ + 313, + 109, + 555, + 312 + ], + "type": "text", + "content": " and each feature token " + }, + { + "bbox": [ + 313, + 109, + 555, + 312 + ], + "type": "inline_equation", + "content": "\\mathbf{p}_n^i" + }, + { + "bbox": [ + 313, + 109, + 555, + 312 + ], + "type": "text", + "content": " by re-using Eq. 5. Through the sigmoid function, we can get the score " + }, + { + "bbox": [ + 313, + 109, + 555, + 312 + ], + "type": "inline_equation", + "content": "\\mathbf{s}_{n,p}^{(i,\\tau)}" + }, + { + "bbox": [ + 313, + 109, + 555, + 312 + ], + "type": "text", + "content": " estimated by UniPro-P. To capture cross-modal modality-common information for better compensation of missing modality, we similarly obtain the score " + }, + { + "bbox": [ + 313, + 109, + 555, + 312 + ], + "type": "inline_equation", + "content": "\\mathbf{s}_{n,g}^{(i,\\tau)}" + }, + { + "bbox": [ + 313, + 109, + 555, + 312 + ], + "type": "text", + "content": " estimated by UniPro-G. Since textual class representations of UniPro are fixed once finishing the training of stage 1, the selected tokens will also become fixed accordingly. To make it diverse while dynamically adapting to the current input, we employ a learnable self-scoring mechanism implemented by attention layers [13] to assign a score " + }, + { + "bbox": [ + 313, + 109, + 555, + 312 + ], + "type": "inline_equation", + "content": "\\mathbf{a}_{n,p}^{(i)}" + }, + { + "bbox": [ + 313, + 109, + 555, + 312 + ], + "type": "text", + "content": " for each feature token. In the end, the final score " + }, + { + "bbox": [ + 313, + 109, + 555, + 312 + ], + "type": "inline_equation", + "content": "\\mathbf{s}_{n,\\#}^{(i)}" + }, + { + "bbox": [ + 313, + 109, + 555, + 312 + ], + "type": "text", + "content": " for each feature token can be computed by their summarization:" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 375, + 319, + 553, + 336 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 375, + 319, + 553, + 336 + ], + "spans": [ + { + "bbox": [ + 375, + 319, + 553, + 336 + ], + "type": "interline_equation", + "content": "\\mathbf {s} _ {n, \\#} ^ {(i)} = \\mathbf {s} _ {n, p} ^ {(i, \\tau)} + \\mathbf {s} _ {n, g} ^ {(i, \\tau)} + \\mathbf {a} _ {n, p} ^ {(i)} \\tag {8}", + "image_path": "d5c38f15c338344c9953cb056742e1b29c0f4862d395f2efb6665350c53d557b.jpg" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 313, + 337, + 554, + 385 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 337, + 554, + 385 + ], + "spans": [ + { + "bbox": [ + 313, + 337, + 554, + 385 + ], + "type": "text", + "content": "For the inference phase when the ground-truth class cannot be accessed, we compute similarity scores with all textual class representations and apply Top-K max-pooling to determine which class " + }, + { + "bbox": [ + 313, + 337, + 554, + 385 + ], + "type": "inline_equation", + "content": "j" + }, + { + "bbox": [ + 313, + 337, + 554, + 385 + ], + "type": "text", + "content": " these tokens belong to:" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 367, + 392, + 553, + 412 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 367, + 392, + 553, + 412 + ], + "spans": [ + { + "bbox": [ + 367, + 392, + 553, + 412 + ], + "type": "interline_equation", + "content": "\\tau = \\arg \\max _ {j \\in [ 1, 2 I _ {t} ]} h \\left(\\mathcal {S} _ {n, p} ^ {(j)} \\oplus \\mathcal {S} _ {n, g} ^ {(j)}\\right) \\tag {9}", + "image_path": "77f804016af301bdbb4d3795edebcd59ebc4ff4de0b1e4e36a2c47d27683bd03.jpg" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 313, + 415, + 553, + 438 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 415, + 553, + 438 + ], + "spans": [ + { + "bbox": [ + 313, + 415, + 553, + 438 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 313, + 415, + 553, + 438 + ], + "type": "inline_equation", + "content": "\\oplus" + }, + { + "bbox": [ + 313, + 415, + 553, + 438 + ], + "type": "text", + "content": " refers to the element-wise addition and " + }, + { + "bbox": [ + 313, + 415, + 553, + 438 + ], + "type": "inline_equation", + "content": "h(\\cdot)" + }, + { + "bbox": [ + 313, + 415, + 553, + 438 + ], + "type": "text", + "content": " can be computed by Eq. 6." + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 313, + 438, + 554, + 586 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 438, + 554, + 586 + ], + "spans": [ + { + "bbox": [ + 313, + 438, + 554, + 586 + ], + "type": "text", + "content": "By selecting feature tokens with the top- " + }, + { + "bbox": [ + 313, + 438, + 554, + 586 + ], + "type": "inline_equation", + "content": "K_{p}" + }, + { + "bbox": [ + 313, + 438, + 554, + 586 + ], + "type": "text", + "content": " largest scores, we can get the input tokens " + }, + { + "bbox": [ + 313, + 438, + 554, + 586 + ], + "type": "inline_equation", + "content": "\\{\\mathbf{p}_n^i\\}_{i=1}^{K_p}" + }, + { + "bbox": [ + 313, + 438, + 554, + 586 + ], + "type": "text", + "content": " for available pathology modality. Then, we employ a set of learnable tokens to act as placeholders " + }, + { + "bbox": [ + 313, + 438, + 554, + 586 + ], + "type": "inline_equation", + "content": "[H_p]_{1\\dots K_p}" + }, + { + "bbox": [ + 313, + 438, + 554, + 586 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 313, + 438, + 554, + 586 + ], + "type": "inline_equation", + "content": "[H_g]_{1\\dots K_g}" + }, + { + "bbox": [ + 313, + 438, + 554, + 586 + ], + "type": "text", + "content": " for pathology and genomic data, where " + }, + { + "bbox": [ + 313, + 438, + 554, + 586 + ], + "type": "inline_equation", + "content": "K_{p}" + }, + { + "bbox": [ + 313, + 438, + 554, + 586 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 313, + 438, + 554, + 586 + ], + "type": "inline_equation", + "content": "K_{g}" + }, + { + "bbox": [ + 313, + 438, + 554, + 586 + ], + "type": "text", + "content": " are their token number of the input for the LLM in stage 2, respectively. They are initialized with modality indicators. For the available pathology, we add the selected tokens to its placeholders, leading to the input of pathology " + }, + { + "bbox": [ + 313, + 438, + 554, + 586 + ], + "type": "inline_equation", + "content": "\\{\\hat{\\mathbf{p}}_n^i\\}_{i=1}^{K_p}" + }, + { + "bbox": [ + 313, + 438, + 554, + 586 + ], + "type": "text", + "content": ". At last, we concatenate them with placeholders of the missing genomics as well as a [CLS] token, resulting in the final input being forwarded into a LLM " + }, + { + "bbox": [ + 313, + 438, + 554, + 586 + ], + "type": "inline_equation", + "content": "\\mathcal{T}(\\cdot)" + }, + { + "bbox": [ + 313, + 438, + 554, + 586 + ], + "type": "text", + "content": ":" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 319, + 592, + 553, + 616 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 319, + 592, + 553, + 616 + ], + "spans": [ + { + "bbox": [ + 319, + 592, + 553, + 616 + ], + "type": "interline_equation", + "content": "[ C L S ] [ \\hat {\\mathbf {p}} _ {n} ] _ {1 \\dots K _ {p}} [ \\tilde {\\mathbf {g}} _ {n} ] _ {1 \\dots K _ {g}} = \\mathcal {T} ([ \\mathrm {C L S} ] | | \\{\\hat {\\mathbf {p}} _ {n} ^ {i} \\} _ {i = 1} ^ {K _ {p}} | | [ H _ {g} ] _ {1 \\dots K _ {g}}) \\tag {10}", + "image_path": "91f1871a1a3be46f2590ccefc45a867d10656d96145db3c0e7af46e8285af45a.jpg" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 313, + 616, + 554, + 675 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 616, + 554, + 675 + ], + "spans": [ + { + "bbox": [ + 313, + 616, + 554, + 675 + ], + "type": "text", + "content": "UniPro Distillation. To inject modality-specific knowledge for the compensation of missing modality, we utilize UniPro to supervise this process, where UniPro has learned the knowledge distribution regarding the missing modality based on other observed samples in stage 1." + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 313, + 676, + 553, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 676, + 553, + 715 + ], + "spans": [ + { + "bbox": [ + 313, + 676, + 553, + 715 + ], + "type": "text", + "content": "Given the output (Eq. 10) of the LLM in stage 2, we take the part of the missing modality " + }, + { + "bbox": [ + 313, + 676, + 553, + 715 + ], + "type": "inline_equation", + "content": "[\\tilde{\\mathbf{g}}_n]_{1\\dots K_g}" + }, + { + "bbox": [ + 313, + 676, + 553, + 715 + ], + "type": "text", + "content": " and compute similarity scores " + }, + { + "bbox": [ + 313, + 676, + 553, + 715 + ], + "type": "inline_equation", + "content": "\\tilde{\\mathbf{h}}_g" + }, + { + "bbox": [ + 313, + 676, + 553, + 715 + ], + "type": "text", + "content": " between " + }, + { + "bbox": [ + 313, + 676, + 553, + 715 + ], + "type": "inline_equation", + "content": "[\\tilde{\\mathbf{g}}_n]_{1\\dots K_g}" + }, + { + "bbox": [ + 313, + 676, + 553, + 715 + ], + "type": "text", + "content": " and the textual class" + } + ] + } + ], + "index": 20 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "text", + "content": "5106" + } + ] + } + ], + "index": 21 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 4 + }, + { + "para_blocks": [ + { + "bbox": [ + 55, + 72, + 297, + 109 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 72, + 297, + 109 + ], + "spans": [ + { + "bbox": [ + 55, + 72, + 297, + 109 + ], + "type": "text", + "content": "representation by Eq. 5 - 7. By substituting " + }, + { + "bbox": [ + 55, + 72, + 297, + 109 + ], + "type": "inline_equation", + "content": "[\\tilde{\\mathbf{g}}_n]_{1\\ldots K_g}" + }, + { + "bbox": [ + 55, + 72, + 297, + 109 + ], + "type": "text", + "content": " into the survival loss function of Eq. 2, we can supervise the learning of compensation for the missing modality:" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 69, + 118, + 295, + 171 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 69, + 118, + 295, + 171 + ], + "spans": [ + { + "bbox": [ + 69, + 118, + 295, + 171 + ], + "type": "interline_equation", + "content": "\\begin{array}{l} \\mathcal {L} _ {u d} ^ {g} = - \\sum_ {n = 1} ^ {\\tilde {N} _ {g}} \\left[ c _ {n} \\log S _ {n} ^ {(\\tau)} + \\left(1 - c _ {n}\\right) \\log S _ {n} ^ {(\\tau - 1)} \\right. \\tag {11} \\\\ \\left. + \\left(1 - c _ {n}\\right) \\log \\tilde {\\mathbf {h}} _ {n} ^ {g, (\\tau)} \\right] \\\\ \\end{array}", + "image_path": "7d8c6e0d44c1605b6682f1235a37ecb2541d6220fc7b46676777c48769203daf.jpg" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 55, + 172, + 296, + 245 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 172, + 296, + 245 + ], + "spans": [ + { + "bbox": [ + 55, + 172, + 296, + 245 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 55, + 172, + 296, + 245 + ], + "type": "inline_equation", + "content": "\\tilde{N}_g" + }, + { + "bbox": [ + 55, + 172, + 296, + 245 + ], + "type": "text", + "content": " is the number of missing genomic modalities across all data pairs. Similarly, we can get the loss function " + }, + { + "bbox": [ + 55, + 172, + 296, + 245 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{ud}^p" + }, + { + "bbox": [ + 55, + 172, + 296, + 245 + ], + "type": "text", + "content": " when the pathology modality is missing. Lastly, [CLS] token is used for calculating the loss function for survival prediction " + }, + { + "bbox": [ + 55, + 172, + 296, + 245 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{surv}^{cls}" + }, + { + "bbox": [ + 55, + 172, + 296, + 245 + ], + "type": "text", + "content": " in stage 2, and the final loss can be combined linearly with coefficient factors " + }, + { + "bbox": [ + 55, + 172, + 296, + 245 + ], + "type": "inline_equation", + "content": "\\alpha_{1}" + }, + { + "bbox": [ + 55, + 172, + 296, + 245 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 55, + 172, + 296, + 245 + ], + "type": "inline_equation", + "content": "\\alpha_{2}" + }, + { + "bbox": [ + 55, + 172, + 296, + 245 + ], + "type": "text", + "content": ":" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 112, + 253, + 295, + 269 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 112, + 253, + 295, + 269 + ], + "spans": [ + { + "bbox": [ + 112, + 253, + 295, + 269 + ], + "type": "interline_equation", + "content": "\\mathcal {L} = \\mathcal {L} _ {s u r v} ^ {c l s} + \\alpha_ {1} \\mathcal {L} _ {u d} ^ {p} + \\alpha_ {2} \\mathcal {L} _ {u d} ^ {g} \\tag {12}", + "image_path": "93185261b09e3973222e9022c223ccafccb0d55635106aad58f6a201376f4dfe.jpg" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 55, + 277, + 137, + 290 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 277, + 137, + 290 + ], + "spans": [ + { + "bbox": [ + 55, + 277, + 137, + 290 + ], + "type": "text", + "content": "4. Experiments" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 55, + 297, + 178, + 309 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 297, + 178, + 309 + ], + "spans": [ + { + "bbox": [ + 55, + 297, + 178, + 309 + ], + "type": "text", + "content": "4.1. Datasets and Settings" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 55, + 314, + 296, + 421 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 314, + 296, + 421 + ], + "spans": [ + { + "bbox": [ + 55, + 314, + 296, + 421 + ], + "type": "text", + "content": "Datasets. To demonstrate the efficacy of the proposed method, we conducted a series of experiments across five public cancer datasets sourced from The Cancer Genome Atlas (TCGA) " + }, + { + "bbox": [ + 55, + 314, + 296, + 421 + ], + "type": "inline_equation", + "content": "^{1}" + }, + { + "bbox": [ + 55, + 314, + 296, + 421 + ], + "type": "text", + "content": ". The case number of each dataset is demonstrated next to the dataset name in Tab. 1. Regarding genomic data, we follow the grouping of biological pathways provided by [14] and remove genes absent in the cBioPortal " + }, + { + "bbox": [ + 55, + 314, + 296, + 421 + ], + "type": "inline_equation", + "content": "^{2}" + }, + { + "bbox": [ + 55, + 314, + 296, + 421 + ], + "type": "text", + "content": " database, resulting in 330 pathways. More details of datasets can be found in supplementary materials." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 55, + 422, + 296, + 637 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 422, + 296, + 637 + ], + "spans": [ + { + "bbox": [ + 55, + 422, + 296, + 637 + ], + "type": "text", + "content": "Evaluation Settings on Incomplete Modalities. Following the common setting in previous works [22, 32], to systematically assess performances on various missing settings, we first manually construct a series of missing scenarios under different missing ratios by randomly dropping them for training sets on samples with complete modality pairs. Specifically, we set a total missing rate (60%) for two modalities and then construct 5 combinations for pathology and genomics, i.e., " + }, + { + "bbox": [ + 55, + 422, + 296, + 637 + ], + "type": "inline_equation", + "content": "\\{(0\\%, 60\\%), (20\\%, 40\\%), (30\\%, 30\\%), (40\\%, 20\\%), (60\\%, 0\\%)\\}" + }, + { + "bbox": [ + 55, + 422, + 296, + 637 + ], + "type": "text", + "content": ". For inference, all methods are evaluated on three scenarios: pathology only, genomics only, and complete modalities. All experiments are conducted on 5-fold cross-validation following previous works [5, 36] in survival prediction, and the concordance index (C-Index) and its standard deviation (std) are reported to evaluate performances. The LLM is implemented by BioBERT-v1.2 [16]. The supplementary materials contain more experiment results and details on the implementation." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 55, + 644, + 200, + 657 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 644, + 200, + 657 + ], + "spans": [ + { + "bbox": [ + 55, + 644, + 200, + 657 + ], + "type": "text", + "content": "4.2. Comparisons with SOTAs" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 55, + 662, + 295, + 685 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 662, + 295, + 685 + ], + "spans": [ + { + "bbox": [ + 55, + 662, + 295, + 685 + ], + "type": "text", + "content": "We first present unimodal approaches, and SOTA multimodal methods trained with complete modalities (all" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 313, + 72, + 553, + 180 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 72, + 553, + 180 + ], + "spans": [ + { + "bbox": [ + 313, + 72, + 553, + 180 + ], + "type": "text", + "content": "pathology and genomics data) to serve as the upper bound of current complete modality approaches. Then the average performances on all missing combinations are demonstrated under the total training missing rate of " + }, + { + "bbox": [ + 313, + 72, + 553, + 180 + ], + "type": "inline_equation", + "content": "60\\%" + }, + { + "bbox": [ + 313, + 72, + 553, + 180 + ], + "type": "text", + "content": ", evaluated on three inference scenarios. Details on various combinations are shown in Fig. 3 and we discuss them in Sec. 4.3. Lastly, we investigate the upper bound of each model tailored to incomplete modalities by training them with all data (without missing data)." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 313, + 181, + 555, + 479 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 181, + 555, + 479 + ], + "spans": [ + { + "bbox": [ + 313, + 181, + 555, + 479 + ], + "type": "text", + "content": "From results in Tab. 1, we observed that (1) When applying " + }, + { + "bbox": [ + 313, + 181, + 555, + 479 + ], + "type": "inline_equation", + "content": "60\\%" + }, + { + "bbox": [ + 313, + 181, + 555, + 479 + ], + "type": "text", + "content": " training missing rate, DisPro consistently outperformed SOTAs across five datasets by a large margin (on average, " + }, + { + "bbox": [ + 313, + 181, + 555, + 479 + ], + "type": "inline_equation", + "content": "+1.76\\%" + }, + { + "bbox": [ + 313, + 181, + 555, + 479 + ], + "type": "text", + "content": " for pathology only, " + }, + { + "bbox": [ + 313, + 181, + 555, + 479 + ], + "type": "inline_equation", + "content": "+1.57\\%" + }, + { + "bbox": [ + 313, + 181, + 555, + 479 + ], + "type": "text", + "content": " for genomics only and " + }, + { + "bbox": [ + 313, + 181, + 555, + 479 + ], + "type": "inline_equation", + "content": "2.33\\%" + }, + { + "bbox": [ + 313, + 181, + 555, + 479 + ], + "type": "text", + "content": " for complete data). This indicates the superiority of DisPro over the incomplete SOTA methods. (2) When testing on the complete scenario, on 3 out of 5 datasets and overall, DisPro even surpassed the SOTA multimodal survival model (SurvPath) trained on all data. Notably, DisPro achieved this while only utilizing incomplete training data with a " + }, + { + "bbox": [ + 313, + 181, + 555, + 479 + ], + "type": "inline_equation", + "content": "60\\%" + }, + { + "bbox": [ + 313, + 181, + 555, + 479 + ], + "type": "text", + "content": " missing rate. It suggests that harnessing the powerful capabilities of LLMs is a promising solution for addressing the classical question of survival analysis. (3) When using complete training data, DisPro consistently performed best on average for various test scenarios as well (" + }, + { + "bbox": [ + 313, + 181, + 555, + 479 + ], + "type": "inline_equation", + "content": "2.28\\%" + }, + { + "bbox": [ + 313, + 181, + 555, + 479 + ], + "type": "text", + "content": " on pathology only, " + }, + { + "bbox": [ + 313, + 181, + 555, + 479 + ], + "type": "inline_equation", + "content": "1.33\\%" + }, + { + "bbox": [ + 313, + 181, + 555, + 479 + ], + "type": "text", + "content": " on genomics only and " + }, + { + "bbox": [ + 313, + 181, + 555, + 479 + ], + "type": "inline_equation", + "content": "2.57\\%" + }, + { + "bbox": [ + 313, + 181, + 555, + 479 + ], + "type": "text", + "content": " on complete test data). Under the same condition, DisPro further increased by " + }, + { + "bbox": [ + 313, + 181, + 555, + 479 + ], + "type": "inline_equation", + "content": "1.31\\%" + }, + { + "bbox": [ + 313, + 181, + 555, + 479 + ], + "type": "text", + "content": " compared with the SOTA complete multimodal method, further demonstrating the strong power of LLMs. (4) In contrast to unimodal methods trained on full unimodal datasets, DisPro, using only incomplete multimodal data with a " + }, + { + "bbox": [ + 313, + 181, + 555, + 479 + ], + "type": "inline_equation", + "content": "60\\%" + }, + { + "bbox": [ + 313, + 181, + 555, + 479 + ], + "type": "text", + "content": " missing rate, achieved superior or similar performance when inferring from individual modalities. This showcases DisPro's efficient utilization of multimodal data." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 313, + 491, + 408, + 504 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 491, + 408, + 504 + ], + "spans": [ + { + "bbox": [ + 313, + 491, + 408, + 504 + ], + "type": "text", + "content": "4.3. Ablation Study" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 313, + 510, + 555, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 510, + 555, + 713 + ], + "spans": [ + { + "bbox": [ + 313, + 510, + 555, + 713 + ], + "type": "text", + "content": "Component Validation. In Tab. 2, we ablate each component in our method to investigate their effectiveness. (1) To demonstrate how the LLM performs on missing modalities without any prompts, we establish a baseline by employing a linear classifier head on the same LLM and randomly selecting feature tokens for its input. Results show that the performance significantly decreased by " + }, + { + "bbox": [ + 313, + 510, + 555, + 713 + ], + "type": "inline_equation", + "content": "2.62\\%" + }, + { + "bbox": [ + 313, + 510, + 555, + 713 + ], + "type": "text", + "content": " on pathology only, " + }, + { + "bbox": [ + 313, + 510, + 555, + 713 + ], + "type": "inline_equation", + "content": "1.55\\%" + }, + { + "bbox": [ + 313, + 510, + 555, + 713 + ], + "type": "text", + "content": " on genomics only and " + }, + { + "bbox": [ + 313, + 510, + 555, + 713 + ], + "type": "inline_equation", + "content": "2.05\\%" + }, + { + "bbox": [ + 313, + 510, + 555, + 713 + ], + "type": "text", + "content": " on complete test data. This indicates the importance of appropriate prompts for LLMs. (2) When supplementing modality-specific information for missing modalities, the performance " + }, + { + "bbox": [ + 313, + 510, + 555, + 713 + ], + "type": "inline_equation", + "content": "(+UD)" + }, + { + "bbox": [ + 313, + 510, + 555, + 713 + ], + "type": "text", + "content": " consistently improves on all datasets and increases by around " + }, + { + "bbox": [ + 313, + 510, + 555, + 713 + ], + "type": "inline_equation", + "content": "+1\\%" + }, + { + "bbox": [ + 313, + 510, + 555, + 713 + ], + "type": "text", + "content": " on various inference settings, which validates the effectiveness of compensating for modality-specific knowledge of the missing modality. (3) When reusing UniPro to select relevant tokens " + }, + { + "bbox": [ + 313, + 510, + 555, + 713 + ], + "type": "inline_equation", + "content": "(+US)" + }, + { + "bbox": [ + 313, + 510, + 555, + 713 + ], + "type": "text", + "content": ", the performance increased by about " + }, + { + "bbox": [ + 313, + 510, + 555, + 713 + ], + "type": "inline_equation", + "content": "+1\\%" + }, + { + "bbox": [ + 313, + 510, + 555, + 713 + ], + "type": "text", + "content": " overall, which bene-" + } + ] + } + ], + "index": 13 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 66, + 692, + 216, + 702 + ], + "type": "page_footnote", + "angle": 0, + "lines": [ + { + "bbox": [ + 66, + 692, + 216, + 702 + ], + "spans": [ + { + "bbox": [ + 66, + 692, + 216, + 702 + ], + "type": "text", + "content": "1https://portal.qdc.cancer.gov/" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 66, + 703, + 203, + 713 + ], + "type": "page_footnote", + "angle": 0, + "lines": [ + { + "bbox": [ + 66, + 703, + 203, + 713 + ], + "spans": [ + { + "bbox": [ + 66, + 703, + 203, + 713 + ], + "type": "text", + "content": "2https://www.cbiportal.org/" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "text", + "content": "5107" + } + ] + } + ], + "index": 17 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 5 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 67, + 69, + 544, + 570 + ], + "blocks": [ + { + "bbox": [ + 67, + 69, + 544, + 570 + ], + "lines": [ + { + "bbox": [ + 67, + 69, + 544, + 570 + ], + "spans": [ + { + "bbox": [ + 67, + 69, + 544, + 570 + ], + "type": "table", + "html": "
TrainingMiRateTestP GMethodsBLCA(372)BRCA(1007)COADREAD(533)LUAD(443)UCEC(478)Avg
N/A†TransMIL [28]0.5903±0.0250.6671±0.0540.6093±0.0180.6195±0.0350.6978±0.0460.6368
†CoOp_BioBert [44]0.6014±0.0470.6784±0.0320.6681±0.0260.6543±0.0520.6791±0.0850.6563
†SNN [15]0.6831±0.0400.6567±0.0380.6312±0.0580.6379±0.0710.7267±0.0730.6671
†CoOp_BioBert [44]0.6712±0.0360.6922±0.0350.6743±0.0480.6514±0.0650.7171±0.0600.6812
‡MOTCat [36]0.6274±0.0460.6721±0.0380.6501±0.0350.6753±0.0430.7207±0.0740.6691
‡SurvPath [14]0.6572±0.0270.7071±0.0390.7079±0.0480.6798±0.0490.7391±0.0650.6982
60%•○COM [27]0.6024±0.0130.6740±0.0110.6779±0.0140.6336±0.0110.6985±0.0120.6573
•○M3Care [40]0.6208±0.0140.6690±0.0180.6570±0.0230.6219±0.0070.7031±0.0120.6543
•○HGCN [9]0.5999±0.0090.6831±0.0120.6792±0.0090.6465±0.0050.7011±0.0070.6620
•○MUSE [34]0.6207±0.0130.6514±0.0140.6547±0.0090.6354±0.0120.7048±0.0230.6534
•○MAP [17]0.5918±0.0120.6278±0.0060.5965±0.0170.6489±0.0200.6929±0.0210.6316
•○DisPro (Ours)0.6319±0.0140.6895±0.0110.6880±0.0100.6612±0.0140.7272±0.0180.6796
○•COM [27]0.6415±0.0160.6653±0.0180.6593±0.0300.6456±0.0080.7105±0.0070.6645
○•M3Care [40]0.6420±0.0150.6642±0.0290.6491±0.0080.6462±0.0120.7032±0.0130.6609
○•HGCN [9]0.6225±0.0080.6729±0.0090.6752±0.0150.6418±0.0150.7015±0.0100.6628
○•MUSE [34]0.6299±0.0120.6427±0.0250.6391±0.0230.6462±0.0100.6893±0.0270.6494
○•MAP [17]0.5937±0.0160.6269±0.0060.5971±0.0170.6442±0.0090.6934±0.0180.6311
○•DisPro (Ours)0.6547±0.0120.6841±0.0180.6804±0.0240.6548±0.0120.7271±0.0170.6802
••COM [27]0.6192±0.0130.6846±0.0090.6874±0.0050.6548±0.0120.7136±0.0160.6719
••M3Care [40]0.6513±0.0050.6951±0.0240.6698±0.0160.6491±0.0130.7288±0.0110.6788
••HGCN [9]0.6332±0.0060.6907±0.0120.6861±0.0170.6545±0.0110.7077±0.0080.6745
••MUSE [34]0.6207±0.0120.6555±0.0130.6539±0.0120.6415±0.0080.7164±0.0130.6576
••MAP [17]0.5906±0.0130.6277±0.0040.5990±0.0160.6625±0.0080.6901±0.0180.6340
••DisPro (Ours)0.6638±0.0060.7219±0.0150.7029±0.0160.6741±0.0070.7476±0.0200.7021
0%••COM [27]0.6154±0.0410.6871±0.0270.6925±0.0440.6427±0.0460.6806±0.0610.6637
••M3Care [40]0.6268±0.0240.6775±0.0160.6850±0.0520.6299±0.0380.6964±0.0610.6631
••HGCN [9]0.6055±0.0480.6716±0.0220.6734±0.0480.6421±0.0520.6915±0.0530.6568
••MUSE [34]0.6208±0.0260.6767±0.0400.6710±0.0410.6253±0.0410.7099±0.0630.6607
••MAP [17]0.6205±0.0330.6300±0.0530.5930±0.0650.6662±0.0550.6925±0.0640.6404
••DisPro (Ours)0.6315±0.0280.6945±0.0440.7216±0.0280.6747±0.0520.7104±0.0560.6865
○•COM [27]0.6681±0.0300.6966±0.0470.6770±0.0240.6454±0.0500.7095±0.0590.6793
○•M3Care [40]0.6706±0.0390.6611±0.0400.6563±0.0210.6512±0.0610.7047±0.0600.6688
○•HGCN [9]0.6233±0.0330.6807±0.0230.6777±0.0370.6421±0.0450.7091±0.0430.6666
○•MUSE [34]0.6256±0.0290.6401±0.0240.6208±0.0300.6547±0.0750.6824±0.0390.6447
○•MAP [17]0.6093±0.0260.6305±0.0510.5842±0.0650.6653±0.0450.6841±0.0640.6347
○•DisPro (Ours)0.6745±0.0390.6996±0.0280.7000±0.0210.6751±0.0680.7140±0.0600.6926
••COM [27]0.6376±0.0430.7087±0.0240.6990±0.0450.6659±0.0340.7112±0.0510.6845
••M3Care [40]0.6605±0.0230.6808±0.0410.6768±0.0260.6740±0.0460.7359±0.0330.6856
••HGCN [9]0.6365±0.0210.7058±0.0330.7136±0.0250.6692±0.0500.7017±0.0440.6854
••MUSE [34]0.6282±0.0390.7017±0.0200.6693±0.0420.6248±0.0400.7237±0.0620.6695
••MAP [17]0.6034±0.0260.6328±0.0490.5895±0.0630.6904±0.0470.6827±0.0650.6398
••DisPro (Ours)0.6719±0.0510.7358±0.0510.7553±0.0210.6767±0.0410.7169±0.0710.7113
", + "image_path": "725c610a214a123865532340a3d3fd3d37c66d2f12ece7b2abb97c90716f1207.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "table_body" + } + ], + "index": 0 + }, + { + "bbox": [ + 55, + 576, + 555, + 610 + ], + "lines": [ + { + "bbox": [ + 55, + 576, + 555, + 610 + ], + "spans": [ + { + "bbox": [ + 55, + 576, + 555, + 610 + ], + "type": "text", + "content": "Table 1. C-Index of Comparison to SOTA models. \"MiRate\" represents Missing Rate. \"P\" and \"G\" indicate Pathology and Genomics, respectively. " + }, + { + "bbox": [ + 55, + 576, + 555, + 610 + ], + "type": "inline_equation", + "content": "\\circ" + }, + { + "bbox": [ + 55, + 576, + 555, + 610 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 55, + 576, + 555, + 610 + ], + "type": "inline_equation", + "content": "\\bullet" + }, + { + "bbox": [ + 55, + 576, + 555, + 610 + ], + "type": "text", + "content": " mean the missing modality and the available modality,respectively. † and ‡ indicate using unimodal and complete multimodal data, respectively, while the rest are for missing modalities. The best results are in bold. The second-best results are underlined." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "text" + }, + { + "bbox": [ + 55, + 623, + 295, + 647 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 623, + 295, + 647 + ], + "spans": [ + { + "bbox": [ + 55, + 623, + 295, + 647 + ], + "type": "text", + "content": "fits from the assistance of US in learning modality-common knowledge from WSIs and genes with high redundancy." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 55, + 654, + 295, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 654, + 295, + 715 + ], + "spans": [ + { + "bbox": [ + 55, + 654, + 295, + 715 + ], + "type": "text", + "content": "Various Missing Scenarios. To illustrate how incomplete multimodal approaches handle varying missing settings, we display different combinations of training missing rates for each inference scenario, as shown in Fig. 3. Results showcase that DisPro achieved considerably consistent perfor" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 313, + 623, + 555, + 647 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 623, + 555, + 647 + ], + "spans": [ + { + "bbox": [ + 313, + 623, + 555, + 647 + ], + "type": "text", + "content": "mance superiority in various missing settings over other incomplete multimodal approaches by a clear margin." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 314, + 657, + 388, + 670 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 314, + 657, + 388, + 670 + ], + "spans": [ + { + "bbox": [ + 314, + 657, + 388, + 670 + ], + "type": "text", + "content": "5. Conclusion" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 313, + 677, + 555, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 677, + 555, + 715 + ], + "spans": [ + { + "bbox": [ + 313, + 677, + 555, + 715 + ], + "type": "text", + "content": "This work proposed a Distilled Prompt Learning (DisPro) framework to unlock the potential of LLMs for incomplete multimodal learning by two-stage prompting, in which" + } + ] + } + ], + "index": 6 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 749, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 749, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 749, + 315, + 757 + ], + "type": "text", + "content": "5108" + } + ] + } + ], + "index": 7 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 6 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 67, + 70, + 547, + 246 + ], + "blocks": [ + { + "bbox": [ + 67, + 70, + 547, + 246 + ], + "lines": [ + { + "bbox": [ + 67, + 70, + 547, + 246 + ], + "spans": [ + { + "bbox": [ + 67, + 70, + 547, + 246 + ], + "type": "table", + "html": "
VariantsTestModulesBLCA (372)BRCA (1007)COADREAD (533)LUAD (443)UCEC (478)Avg
PGUSUD
baseline•○0.6068±0.0210.6668±0.0060.6583±0.0120.6270±0.0180.7081±0.0190.6534
+ UD•○✓0.6213±0.0100.6829±0.0080.6740±0.0220.6327±0.0180.7152±0.0110.6652
+ US•○✓0.6285±0.0210.6884±0.0160.6653±0.0260.6546±0.0130.7203±0.0070.6714
DisPro (full)•○✓✓0.6319±0.0140.6895±0.0110.6880±0.0100.6612±0.0140.7272±0.0180.6796
baseline○•0.6435±0.0160.6605±0.0330.6601±0.0180.6467±0.0160.7127±0.0070.6647
+ UD○•✓0.6473±0.0170.6836±0.0140.6707±0.0220.6420±0.0200.7226±0.0130.6732
+ US○•✓0.6452±0.0200.6666±0.0170.6763±0.0230.6471±0.0230.7151±0.0090.6701
DisPro (full)○•✓✓0.6547±0.0120.6841±0.0180.6804±0.0240.6548±0.0120.7271±0.0170.6802
baseline••0.6522±0.0100.7075±0.0120.6785±0.0200.6541±0.0090.7157±0.0030.6816
+ UD••✓0.6585±0.0070.7194±0.0130.6963±0.0090.6635±0.0140.7250±0.0120.6926
+ US••✓0.6593±0.0080.7125±0.0100.7005±0.0180.6717±0.0070.7274±0.0100.6943
DisPro (full)••✓✓0.6638±0.0060.7219±0.0150.7029±0.0160.6741±0.0070.7476±0.0200.7021
", + "image_path": "021044f3074909bce32dd4ceb8e19993e3b90d88780b3ecffd6f5210ee921581.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "table_body" + } + ], + "index": 0 + }, + { + "bbox": [ + 55, + 255, + 555, + 287 + ], + "lines": [ + { + "bbox": [ + 55, + 255, + 555, + 287 + ], + "spans": [ + { + "bbox": [ + 55, + 255, + 555, + 287 + ], + "type": "text", + "content": "Table 2. C-Index of Ablation Study on the Setting of " + }, + { + "bbox": [ + 55, + 255, + 555, + 287 + ], + "type": "inline_equation", + "content": "60\\%" + }, + { + "bbox": [ + 55, + 255, + 555, + 287 + ], + "type": "text", + "content": " Training Missing Rate.\"US\" and \"UD\" refer to the UniPro Scoring and UniPro Distillation modules, respectively. \"P\" and \"G\" indicate Pathology and Genomics, respectively. " + }, + { + "bbox": [ + 55, + 255, + 555, + 287 + ], + "type": "inline_equation", + "content": "\\circ" + }, + { + "bbox": [ + 55, + 255, + 555, + 287 + ], + "type": "text", + "content": " means the modality is missing, while " + }, + { + "bbox": [ + 55, + 255, + 555, + 287 + ], + "type": "inline_equation", + "content": "\\bullet" + }, + { + "bbox": [ + 55, + 255, + 555, + 287 + ], + "type": "text", + "content": " refers to the available modality." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "text" + }, + { + "type": "image", + "bbox": [ + 72, + 288, + 538, + 566 + ], + "blocks": [ + { + "bbox": [ + 72, + 288, + 538, + 566 + ], + "lines": [ + { + "bbox": [ + 72, + 288, + 538, + 566 + ], + "spans": [ + { + "bbox": [ + 72, + 288, + 538, + 566 + ], + "type": "image", + "image_path": "4918a135959d8a2aaa250d2f0febc2b07cefb70695a1228d78a63e9cb9a6efb9.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 108, + 576, + 499, + 588 + ], + "lines": [ + { + "bbox": [ + 108, + 576, + 499, + 588 + ], + "spans": [ + { + "bbox": [ + 108, + 576, + 499, + 588 + ], + "type": "text", + "content": "Figure 3. Performance on various combinations under " + }, + { + "bbox": [ + 108, + 576, + 499, + 588 + ], + "type": "inline_equation", + "content": "60\\%" + }, + { + "bbox": [ + 108, + 576, + 499, + 588 + ], + "type": "text", + "content": " training missing rate for different test scenarios." + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_caption" + } + ], + "index": 2 + }, + { + "bbox": [ + 54, + 602, + 294, + 710 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 54, + 602, + 294, + 710 + ], + "spans": [ + { + "bbox": [ + 54, + 602, + 294, + 710 + ], + "type": "text", + "content": "Unimodal Prompting (UniPro) and Multimodal Prompting (MultiPro) were proposed for supplementing modality-specific and modality-common knowledge for missing modalities. In the first stage, UniPro extended CoOp into the MIL setting and distilled knowledge distribution to a set of learnable prompts for each modality, which prepped for compensation of modality-specific knowledge. In the second stage, considering available modalities as prompts, we leveraged an LLM to impute common knowledge for miss" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 313, + 602, + 555, + 699 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 602, + 555, + 699 + ], + "spans": [ + { + "bbox": [ + 313, + 602, + 555, + 699 + ], + "type": "text", + "content": "ing modalities. To inject the specific knowledge of missing modalities, the distilled prompts in UniPro are used to supervise the learning of imputing representations for missing modalities, leading to more comprehensive knowledge compensation in incomplete multimodal learning. Extensive experiments on five cancer benchmark datasets demonstrated the effectiveness of the proposed method over state-of-the-art approaches." + } + ] + } + ], + "index": 5 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "text", + "content": "5109" + } + ] + } + ], + "index": 6 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 7 + }, + { + "para_blocks": [ + { + "bbox": [ + 56, + 72, + 153, + 85 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 72, + 153, + 85 + ], + "spans": [ + { + "bbox": [ + 56, + 72, + 153, + 85 + ], + "type": "text", + "content": "Acknowledgement" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 55, + 90, + 296, + 159 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 90, + 296, + 159 + ], + "spans": [ + { + "bbox": [ + 55, + 90, + 296, + 159 + ], + "type": "text", + "content": "This work was supported by the National Natural Science Foundation of China (No. 62202403), Hong Kong Innovation and Technology Commission (Project No. MHP/002/22 and ITCPD/17-9), and Research Grants Council of the Hong Kong Special Administrative Region, China (Project No. R6003-22 and C4024-22GF)." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 56, + 178, + 115, + 190 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 178, + 115, + 190 + ], + "spans": [ + { + "bbox": [ + 56, + 178, + 115, + 190 + ], + "type": "text", + "content": "References" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 57, + 198, + 297, + 713 + ], + "type": "list", + "angle": 0, + "index": 15, + "blocks": [ + { + "bbox": [ + 61, + 198, + 297, + 220 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 61, + 198, + 297, + 220 + ], + "spans": [ + { + "bbox": [ + 61, + 198, + 297, + 220 + ], + "type": "text", + "content": "[1] Som S Biswas. Role of chat gpt in public health. Annals of biomedical engineering, 51(5):868-869, 2023. 2" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 62, + 222, + 296, + 243 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 222, + 296, + 243 + ], + "spans": [ + { + "bbox": [ + 62, + 222, + 296, + 243 + ], + "type": "text", + "content": "[2] Jason Brownlee. A gentle introduction to the rectified linear unit (relu). Machine learning mastery, 6, 2019. 3, 1" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 62, + 245, + 296, + 288 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 245, + 296, + 288 + ], + "spans": [ + { + "bbox": [ + 62, + 245, + 296, + 288 + ], + "type": "text", + "content": "[3] Agisilaos Chartsias, Thomas Joyce, Mario Valerio Giuffrida, and Sotirios A Tsaftaris. Multimodal mr synthesis via modality-invariant latent representation. IEEE transactions on medical imaging, 37(3):803-814, 2017. 1" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 62, + 289, + 296, + 343 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 289, + 296, + 343 + ], + "spans": [ + { + "bbox": [ + 62, + 289, + 296, + 343 + ], + "type": "text", + "content": "[4] Jiayi Chen and Aidong Zhang. Hgmf: heterogeneous graph-based fusion for multimodal data with incompleteness. In Proceedings of the 26th ACM SIGKDD international conference on knowledge discovery & data mining, pages 1295-1305, 2020. 1, 2" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 62, + 345, + 296, + 411 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 345, + 296, + 411 + ], + "spans": [ + { + "bbox": [ + 62, + 345, + 296, + 411 + ], + "type": "text", + "content": "[5] Richard J Chen, Ming Y Lu, Wei-Hung Weng, Tiffany Y Chen, Drew FK Williamson, Trevor Manz, Maha Shady, and Faisal Mahmood. Multimodal co-attention transformer for survival prediction in gigapixel whole slide images. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 4015–4025, 2021. 1, 3, 6" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 62, + 412, + 296, + 476 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 412, + 296, + 476 + ], + "spans": [ + { + "bbox": [ + 62, + 412, + 296, + 476 + ], + "type": "text", + "content": "[6] Richard J Chen, Tong Ding, Ming Y Lu, Drew FK Williamson, Guillaume Jaume, Andrew H Song, Bowen Chen, Andrew Zhang, Daniel Shao, Muhammad Shaban, et al. Towards a general-purpose foundation model for computational pathology. Nature Medicine, 30(3):850-862, 2024. 1" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 62, + 478, + 296, + 521 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 478, + 296, + 521 + ], + "spans": [ + { + "bbox": [ + 62, + 478, + 296, + 521 + ], + "type": "text", + "content": "[7] Florinel-Alin Croitoru, Vlad Hondru, Radu Tudor Ionescu, and Mubarak Shah. Diffusion models in vision: A survey. IEEE Transactions on Pattern Analysis and Machine Intelligence, 45(9):10850-10869, 2023. 2" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 62, + 523, + 296, + 576 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 523, + 296, + 576 + ], + "spans": [ + { + "bbox": [ + 62, + 523, + 296, + 576 + ], + "type": "text", + "content": "[8] Yuhang Ding, Xin Yu, and Yi Yang. Rfnet: Region-aware fusion network for incomplete multi-modal brain tumor segmentation. In Proceedings of the IEEE/CVF international conference on computer vision, pages 3975-3984, 2021. 1, 2" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 62, + 579, + 296, + 612 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 579, + 296, + 612 + ], + "spans": [ + { + "bbox": [ + 62, + 579, + 296, + 612 + ], + "type": "text", + "content": "[9] Hou et al. Hybrid graph convolutional network with online masked autoencoder for robust multimodal cancer survival prediction. IEEE Transactions on Medical Imaging, 2023. 7" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 57, + 613, + 296, + 635 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 613, + 296, + 635 + ], + "spans": [ + { + "bbox": [ + 57, + 613, + 296, + 635 + ], + "type": "text", + "content": "[10] Zhou et al. Cross-modal translation and alignment for survival analysis. In ICCV, pages 21485-21494, 2023. 1, 3" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 57, + 636, + 296, + 679 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 636, + 296, + 679 + ], + "spans": [ + { + "bbox": [ + 57, + 636, + 296, + 679 + ], + "type": "text", + "content": "[11] Michael Gadermayr and Maximilian Tschuchnig. Multiple instance learning for digital pathology: A review of the state-of-the-art, limitations & future potential. Computerized Medical Imaging and Graphics, page 102337, 2024. 2" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 57, + 681, + 296, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 681, + 296, + 713 + ], + "spans": [ + { + "bbox": [ + 57, + 681, + 296, + 713 + ], + "type": "text", + "content": "[12] Mohammad Hamghalam, Alejandro F Frangi, Baiying Lei, and Amber L Simpson. Modality completion via gaussian process prior variational autoencoders for multi-modal" + } + ] + } + ], + "index": 14 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 316, + 73, + 555, + 713 + ], + "type": "list", + "angle": 0, + "index": 28, + "blocks": [ + { + "bbox": [ + 335, + 73, + 555, + 128 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 335, + 73, + 555, + 128 + ], + "spans": [ + { + "bbox": [ + 335, + 73, + 555, + 128 + ], + "type": "text", + "content": "glioma segmentation. In Medical Image Computing and Computer Assisted Intervention-MICCAI 2021: 24th International Conference, Strasbourg, France, September 27-October 1, 2021, Proceedings, Part VII 24, pages 442-452. Springer, 2021. 2" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 316, + 129, + 554, + 171 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 129, + 554, + 171 + ], + "spans": [ + { + "bbox": [ + 316, + 129, + 554, + 171 + ], + "type": "text", + "content": "[13] Maximilian Ilse, Jakub Tomczak, and Max Welling. Attention-based deep multiple instance learning. In International conference on machine learning, pages 2127-2136. PMLR, 2018. 2, 5" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 317, + 173, + 555, + 239 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 173, + 555, + 239 + ], + "spans": [ + { + "bbox": [ + 317, + 173, + 555, + 239 + ], + "type": "text", + "content": "[14] Guillaume Jaume, Anurag Vaidya, Richard J Chen, Drew FK Williamson, Paul Pu Liang, and Faisal Mahmood. Modeling dense multimodal interactions between biological pathways and histology for survival prediction. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 11579-11590, 2024. 3, 6, 7, 1" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 316, + 240, + 554, + 282 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 240, + 554, + 282 + ], + "spans": [ + { + "bbox": [ + 316, + 240, + 554, + 282 + ], + "type": "text", + "content": "[15] Günter Klambauer, Thomas Unterthiner, Andreas Mayr, and Sepp Hochreiter. Self-normalizing neural networks. Advances in neural information processing systems, 30, 2017. 7, 1" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 316, + 284, + 554, + 336 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 284, + 554, + 336 + ], + "spans": [ + { + "bbox": [ + 316, + 284, + 554, + 336 + ], + "type": "text", + "content": "[16] Jinhyuk Lee, Wonjin Yoon, Sungdong Kim, Donghyeon Kim, Sunkyu Kim, Chan Ho So, and Jaewoo Kang. Biobert: a pre-trained biomedical language representation model for biomedical text mining. Bioinformatics, 36(4):1234-1240, 2020. 6, 1" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 317, + 338, + 554, + 392 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 338, + 554, + 392 + ], + "spans": [ + { + "bbox": [ + 317, + 338, + 554, + 392 + ], + "type": "text", + "content": "[17] Yi-Lun Lee, Yi-Hsuan Tsai, Wei-Chen Chiu, and Chen-Yu Lee. Multimodal prompting with missing modalities for visual recognition. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 14943-14952, 2023. 2, 7, 3" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 317, + 394, + 554, + 449 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 394, + 554, + 449 + ], + "spans": [ + { + "bbox": [ + 317, + 394, + 554, + 449 + ], + "type": "text", + "content": "[18] Bin Li, Yin Li, and Kevin W Eliceiri. Dual-stream multiple instance learning network for whole slide image classification with self-supervised contrastive learning. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 14318-14328, 2021. 3" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 317, + 449, + 554, + 503 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 449, + 554, + 503 + ], + "spans": [ + { + "bbox": [ + 317, + 449, + 554, + 503 + ], + "type": "text", + "content": "[19] Chenxin Li, Xinyu Liu, Cheng Wang, Yifan Liu, Weihao Yu, Jing Shao, and Yixuan Yuan. Gtp-4o: Modality-prompted heterogeneous graph learning for omni-modal biomedical representation. In European conference on computer vision, pages 168–187. Springer, 2024. 2" + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 317, + 505, + 554, + 558 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 505, + 554, + 558 + ], + "spans": [ + { + "bbox": [ + 317, + 505, + 554, + 558 + ], + "type": "text", + "content": "[20] Chunyuan Li, Cliff Wong, Sheng Zhang, Naoto Usuyama, Haotian Liu, Jianwei Yang, Tristan Naumann, Hoifung Poon, and Jianfeng Gao. Llava-med: Training a large language-and-vision assistant for biomedicine in one day. Advances in Neural Information Processing Systems, 36, 2024. 2" + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 317, + 559, + 554, + 635 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 559, + 554, + 635 + ], + "spans": [ + { + "bbox": [ + 317, + 559, + 554, + 635 + ], + "type": "text", + "content": "[21] Han Liu, Yubo Fan, Hao Li, Jiacheng Wang, Dewei Hu, Can Cui, Ho Hin Lee, Huahong Zhang, and Ipek Oguz. Moddrop++: A dynamic filter network with intra-subject cotraining for multiple sclerosis lesion segmentation with missing modalities. In International Conference on Medical Image Computing and Computer-Assisted Intervention, pages 444-453. Springer, 2022. 2" + } + ] + } + ], + "index": 25 + }, + { + "bbox": [ + 317, + 637, + 554, + 689 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 637, + 554, + 689 + ], + "spans": [ + { + "bbox": [ + 317, + 637, + 554, + 689 + ], + "type": "text", + "content": "[22] Mengmeng Ma, Jian Ren, Long Zhao, Sergey Tulyakov, Cathy Wu, and Xi Peng. Smil: Multimodal learning with severely missing modality. In Proceedings of the AAAI Conference on Artificial Intelligence, pages 2302-2310, 2021. 2, 6" + } + ] + } + ], + "index": 26 + }, + { + "bbox": [ + 316, + 692, + 554, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 692, + 554, + 713 + ], + "spans": [ + { + "bbox": [ + 316, + 692, + 554, + 713 + ], + "type": "text", + "content": "[23] Xiangxi Meng, Kaicong Sun, Jun Xu, Xuming He, and Dinggang Shen. Multi-modal modality-masked diffusion network" + } + ] + } + ], + "index": 27 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 748, + 316, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 316, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 316, + 757 + ], + "type": "text", + "content": "5110" + } + ] + } + ], + "index": 29 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 8 + }, + { + "para_blocks": [ + { + "bbox": [ + 56, + 72, + 296, + 713 + ], + "type": "list", + "angle": 0, + "index": 14, + "blocks": [ + { + "bbox": [ + 76, + 72, + 296, + 95 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 76, + 72, + 296, + 95 + ], + "spans": [ + { + "bbox": [ + 76, + 72, + 296, + 95 + ], + "type": "text", + "content": "for brain mri synthesis with random modality missing. IEEE Transactions on Medical Imaging, 2024. 1, 2" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 56, + 96, + 296, + 129 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 96, + 296, + 129 + ], + "spans": [ + { + "bbox": [ + 56, + 96, + 296, + 129 + ], + "type": "text", + "content": "[24] Muhammad Khalid Khan Niazi, Anil V Parwani, and Metin N Gurcan. Digital pathology and artificial intelligence. The lancet oncology, 20(5):e253-e261, 2019. 3" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 56, + 130, + 295, + 185 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 130, + 295, + 185 + ], + "spans": [ + { + "bbox": [ + 56, + 130, + 295, + 185 + ], + "type": "text", + "content": "[25] Zhenyuan Ning, Denghui Du, Chao Tu, Qianjin Feng, and Yu Zhang. Relation-aware shared representation learning for cancer prognosis analysis with auxiliary clinical variables and incomplete multi-modality data. IEEE Transactions on Medical Imaging, 41(1):186-198, 2021. 2" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 56, + 186, + 295, + 219 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 186, + 295, + 219 + ], + "spans": [ + { + "bbox": [ + 56, + 186, + 295, + 219 + ], + "type": "text", + "content": "[26] Fatih Ozsolak and Patrice M Milos. RNA sequencing: advances, challenges and opportunities. Nature reviews genetics, 12(2):87-98, 2011. 3" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 56, + 220, + 295, + 253 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 220, + 295, + 253 + ], + "spans": [ + { + "bbox": [ + 56, + 220, + 295, + 253 + ], + "type": "text", + "content": "[27] Shuwei Qian and Chongjun Wang. Com: Contrastive masked-attention model for incomplete multimodal learning. Neural Networks, 162:443–455, 2023. 7" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 56, + 255, + 295, + 308 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 255, + 295, + 308 + ], + "spans": [ + { + "bbox": [ + 56, + 255, + 295, + 308 + ], + "type": "text", + "content": "[28] Zhuchen Shao, Hao Bian, Yang Chen, Yifeng Wang, Jian Zhang, Xiangyang Ji, et al. Transmil: Transformer based correlated multiple instance learning for whole slide image classification. Advances in neural information processing systems, 34:2136-2147, 2021. 3, 7" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 56, + 310, + 295, + 354 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 310, + 295, + 354 + ], + "spans": [ + { + "bbox": [ + 56, + 310, + 295, + 354 + ], + "type": "text", + "content": "[29] Anmol Sharma and Ghassan Hamarneh. Missing mri pulse sequence synthesis using multi-modal generative adversarial network. IEEE transactions on medical imaging, 39(4): 1170-1183, 2019. 1" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 56, + 355, + 295, + 410 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 355, + 295, + 410 + ], + "spans": [ + { + "bbox": [ + 56, + 355, + 295, + 410 + ], + "type": "text", + "content": "[30] Andrew H Song, Richard J Chen, Guillaume Jaume, Anurag Jayant Vaidya, Alexander Baras, and Faisal Mahmood. Multimodal prototyping for cancer survival prediction. In *Forty-first International Conference on Machine Learning*, 2024. 1, 2, 3, 5" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 56, + 411, + 295, + 443 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 411, + 295, + 443 + ], + "spans": [ + { + "bbox": [ + 56, + 411, + 295, + 443 + ], + "type": "text", + "content": "[31] Rory Stark, Marta Grzelak, and James Hadfield. Rna sequencing: the teenage years. Nature Reviews Genetics, 20 (11):631-656, 2019. 3" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 56, + 445, + 295, + 500 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 445, + 295, + 500 + ], + "spans": [ + { + "bbox": [ + 56, + 445, + 295, + 500 + ], + "type": "text", + "content": "[32] Hu Wang, Yuanhong Chen, Congbo Ma, Jodie Avery, Louise Hull, and Gustavo Carneiro. Multi-modal learning with missing modality via shared-specific feature modelling. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 15878-15887, 2023. 6" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 56, + 502, + 295, + 555 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 502, + 295, + 555 + ], + "spans": [ + { + "bbox": [ + 56, + 502, + 295, + 555 + ], + "type": "text", + "content": "[33] Tianyu Wu, Shizhu He, Jingping Liu, Siqi Sun, Kang Liu, Qing-Long Han, and Yang Tang. A brief overview of chatgpt: The history, status quo and potential future development. IEEE/CAA Journal of Automatica Sinica, 10(5):1122-1136, 2023. 2" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 56, + 557, + 295, + 612 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 557, + 295, + 612 + ], + "spans": [ + { + "bbox": [ + 56, + 557, + 295, + 612 + ], + "type": "text", + "content": "[34] Zhenbang Wu, Anant Dadu, Nicholas Tustison, Brian Avants, Mike Nalls, Jimeng Sun, and Faraz Faghri. Multimodal patient representation learning with missing modalities and labels. In The Twelfth International Conference on Learning Representations, 2024. 1, 2, 7" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 56, + 613, + 295, + 677 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 613, + 295, + 677 + ], + "spans": [ + { + "bbox": [ + 56, + 613, + 295, + 677 + ], + "type": "text", + "content": "[35] Conghao Xiong, Hao Chen, Hao Zheng, Dong Wei, Yefeng Zheng, Joseph JY Sung, and Irwin King. Mome: Mixture of multimodal experts for cancer survival prediction. In International Conference on Medical Image Computing and Computer-Assisted Intervention, pages 318-328. Springer, 2024. 1" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 56, + 680, + 295, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 680, + 295, + 713 + ], + "spans": [ + { + "bbox": [ + 56, + 680, + 295, + 713 + ], + "type": "text", + "content": "[36] Yingxue Xu and Hao Chen. Multimodal optimal transport-based co-attention transformer with global structure consistency for survival prediction. In Proceedings of the" + } + ] + } + ], + "index": 13 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 316, + 73, + 553, + 529 + ], + "type": "list", + "angle": 0, + "index": 24, + "blocks": [ + { + "bbox": [ + 335, + 73, + 553, + 95 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 335, + 73, + 553, + 95 + ], + "spans": [ + { + "bbox": [ + 335, + 73, + 553, + 95 + ], + "type": "text", + "content": "IEEE/CVF International Conference on Computer Vision, pages 21241-21251, 2023. 1, 3, 6, 7" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 316, + 96, + 553, + 151 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 96, + 553, + 151 + ], + "spans": [ + { + "bbox": [ + 316, + 96, + 553, + 151 + ], + "type": "text", + "content": "[37] Shu Yang, Yihui Wang, and Hao Chen. Mambamil: Enhancing long sequence modeling with sequence reordering in computational pathology. In International Conference on Medical Image Computing and Computer-Assisted Intervention, pages 296-306. Springer, 2024. 2" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 316, + 152, + 553, + 206 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 152, + 553, + 206 + ], + "spans": [ + { + "bbox": [ + 316, + 152, + 553, + 206 + ], + "type": "text", + "content": "[38] Yixiao Yuan, Yawen Huang, and Yi Zhou. Rethinking a unified generative adversarial model for mri modality completion. In International Conference on Medical Image Computing and Computer-Assisted Intervention, pages 143-153. Springer, 2023. 2" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 316, + 208, + 553, + 251 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 208, + 553, + 251 + ], + "spans": [ + { + "bbox": [ + 316, + 208, + 553, + 251 + ], + "type": "text", + "content": "[39] Shekoufeh Gorgi Zadeh and Matthias Schmid. Bias in cross-entropy-based training of deep survival networks. IEEE transactions on pattern analysis and machine intelligence, 43(9):3126-3137, 2020. 3" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 316, + 252, + 553, + 317 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 252, + 553, + 317 + ], + "spans": [ + { + "bbox": [ + 316, + 252, + 553, + 317 + ], + "type": "text", + "content": "[40] Chaohe Zhang, Xu Chu, Liantao Ma, Yinghao Zhu, Yasha Wang, Jiangtao Wang, and Junfeng Zhao. M3care: Learning with missing modalities in multimodal healthcare data. In Proceedings of the 28th ACM SIGKDD Conference on Knowledge Discovery and Data Mining, pages 2418-2428, 2022. 1, 2, 7" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 316, + 319, + 553, + 385 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 319, + 553, + 385 + ], + "spans": [ + { + "bbox": [ + 316, + 319, + 553, + 385 + ], + "type": "text", + "content": "[41] Hongrun Zhang, Yanda Meng, Yitian Zhao, Yihong Qiao, Xiaoyun Yang, Sarah E Coupland, and Yalin Zheng. Dtfmil: Double-tier feature distillation multiple instance learning for histopathology whole slide image classification. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 18802-18812, 2022. 3" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 316, + 386, + 553, + 441 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 386, + 553, + 441 + ], + "spans": [ + { + "bbox": [ + 316, + 386, + 553, + 441 + ], + "type": "text", + "content": "[42] Yilan Zhang, Yingxue Xu, Jianqi Chen, Fengying Xie, and Hao Chen. Prototypical information bottlenecking and disentangling for multimodal cancer survival prediction. In The Twelfth International Conference on Learning Representations, 2024. 1, 2, 3, 5" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 316, + 441, + 553, + 485 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 441, + 553, + 485 + ], + "spans": [ + { + "bbox": [ + 316, + 441, + 553, + 485 + ], + "type": "text", + "content": "[43] Huajun Zhou, Fengtao Zhou, Chenyu Zhao, Yingxue Xu, Luyang Luo, and Hao Chen. Multimodal data integration for precision oncology: Challenges and future directions. arXiv preprint arXiv:2406.19611, 2024. 1, 2" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 316, + 487, + 553, + 529 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 487, + 553, + 529 + ], + "spans": [ + { + "bbox": [ + 316, + 487, + 553, + 529 + ], + "type": "text", + "content": "[44] Kaiyang Zhou, Jingkang Yang, Chen Change Loy, and Ziwei Liu. Learning to prompt for vision-language models. International Journal of Computer Vision, 130(9):2337-2348, 2022. 2, 3, 7" + } + ] + } + ], + "index": 23 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 748, + 314, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 314, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 314, + 757 + ], + "type": "text", + "content": "5111" + } + ] + } + ], + "index": 25 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 9 + } + ], + "_backend": "vlm", + "_version_name": "2.6.4" +} \ No newline at end of file diff --git a/2025/Distilling Long-tailed Datasets/f09fbc7b-8c7e-4bee-803c-2d67e3d05b6e_content_list.json b/2025/Distilling Long-tailed Datasets/f09fbc7b-8c7e-4bee-803c-2d67e3d05b6e_content_list.json new file mode 100644 index 0000000000000000000000000000000000000000..5676b8fe4e5677bbf539aa9810fcacd1708a150b --- /dev/null +++ b/2025/Distilling Long-tailed Datasets/f09fbc7b-8c7e-4bee-803c-2d67e3d05b6e_content_list.json @@ -0,0 +1,1590 @@ +[ + { + "type": "text", + "text": "Distilling Long-tailed Datasets", + "text_level": 1, + "bbox": [ + 344, + 130, + 653, + 152 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Zhenghao Zhao $^{1*}$ Haoxuan Wang $^{1*}$ Yuzhang Shang $^{2}$ Kai Wang $^{3}$ Yan Yan $^{1\\dagger}$ $^{1}$ University of Illinois Chicago $^{2}$ Illinois Institute of Technology $^{3}$ National University of Singapore \n{zzhao72, hwang339, yyan55}@uic.edu yshang4@hawk.iit.edu kai.wang@comp.nus.edu.sg", + "bbox": [ + 99, + 178, + 903, + 233 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Abstract", + "text_level": 1, + "bbox": [ + 246, + 268, + 326, + 284 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Dataset distillation aims to synthesize a small, information-rich dataset from a large one for efficient model training. However, existing dataset distillation methods struggle with long-tailed datasets, which are prevalent in real-world scenarios. By investigating the reasons behind this unexpected result, we identified two main causes: 1) The distillation process on imbalanced datasets develops biased gradients, leading to the synthesis of similarly imbalanced distilled datasets. 2) The experts trained on such datasets perform suboptimally on tail classes, resulting in misguided distillation supervision and poor-quality soft-label initialization. To address these issues, we first propose Distribution-agnostic Matching to avoid directly matching the biased expert trajectories. It reduces the distance between the student and the biased expert trajectories and prevents the tail class bias from being distilled to the synthetic dataset. Moreover, we improve the distillation guidance with Expert Decoupling, which jointly matches the decoupled backbone and classifier to improve the tail class performance and initialize reliable soft labels. This work pioneers the field of long-tailed dataset distillation, marking the first effective effort to distill long-tailed datasets. Our code will be made public at https://github.com/ichbill/LTDD.", + "bbox": [ + 89, + 300, + 485, + 648 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "1. Introduction", + "text_level": 1, + "bbox": [ + 89, + 675, + 220, + 690 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Dataset distillation (DD) [4, 13, 28] aims to synthesize small and high-quality data summaries which contain the most important knowledge from a given target dataset [35]. Such a compact data summary offers numerous advantages, including faster model training, reduced memory consumption, and greater flexibility for various tasks [48], such as continual learning [31], neural architecture search [27], and knowledge distillation [15]. Various types of dataset distillation methods have been developed, including performance matching methods [30, 32, 50], parameter matching methods [13, 28, 46], representation matching methods [38, 47]", + "bbox": [ + 89, + 700, + 483, + 867 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/25e4b7ae68facc9a624f0beb88f041d46f4562c1930cbe2054339a0d17b8d0d0.jpg", + "image_caption": [ + "Figure 1. Performance comparison on CIFAR-10-LT. Existing Dataset Distillation methods exhibit degraded performance when applied to imbalanced datasets, especially when the imbalance factor increases, whereas our method provides significantly better performance under different imbalanced scenarios." + ], + "image_footnote": [], + "bbox": [ + 516, + 268, + 897, + 444 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "and generative-based methods [12, 40]. Currently, parameter matching methods have demonstrated significant performance improvements across various datasets.", + "bbox": [ + 511, + 520, + 905, + 565 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Research on dataset distillation has predominantly focused on uniformly distributed datasets, such as CIFAR10, CIFAR100 [22], and TinyImageNet [24]. However, in real-world applications such as medical image diagnosis [19], training samples typically exhibit a long-tailed class distribution [45]. In these scenarios, a small number of classes (head classes) have a large number of samples, while the majority of classes (tail classes) have only a few samples. Motivated by this discrepancy and the desire for practical application, we are interested in the novel task of long-tailed dataset distillation (LTDD). LTDD aims to distill a long-tailed target dataset into a small and uniformly distributed synthetic dataset, where models trained on this distilled dataset can achieve satisfactory performance on the uniformly distributed test set.", + "bbox": [ + 509, + 566, + 906, + 792 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Nevertheless, distilling long-tailed datasets presents significant challenges, as the effectiveness of current dataset distillation methods diminishes when applied to such imbalanced distributions. As shown in Figure 1, we find that the performance of different dataset distillation methods consistently decreases as the degree of imbalance increases. The degree of imbalance is quantified by the imbalance fac", + "bbox": [ + 511, + 795, + 908, + 901 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "CVF", + "bbox": [ + 106, + 2, + 181, + 42 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore.", + "bbox": [ + 236, + 0, + 810, + 46 + ], + "page_idx": 0 + }, + { + "type": "page_footnote", + "text": "*equal contribution † corresponding author", + "bbox": [ + 107, + 875, + 230, + 900 + ], + "page_idx": 0 + }, + { + "type": "page_number", + "text": "30609", + "bbox": [ + 478, + 944, + 517, + 955 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/fb810ddfc167dc63badd1af50ccf8014a817b42c3f432b7b7a9cb78065ceb9ee.jpg", + "image_caption": [ + "Figure 2. Relationship of classifier weights and class-wise accuracy. We reveal that classifiers generated by existing dataset distillation methods often exhibit imbalanced distributions, resulting in poor performance on tail classes. In contrast, our method produces balanced classifiers, thereby enhancing overall accuracy." + ], + "image_footnote": [], + "bbox": [ + 91, + 88, + 480, + 223 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "tor, which is a commonly used metric to characterize the extent of imbalance [23]. The larger the imbalance factor is, the more imbalanced the dataset is. When the imbalance factor is large, popular dataset distillation methods may even perform worse than random selection (e.g. DATM can only achieve $40.1\\%$ accuracy when the imbalance factor $\\beta = 200$ , while random selection can achieve $49.9\\%$ accuracy). As shown in Figure 2, the performance degradation is concentrated in the tail classes. We attribute these performance failures to two key aspects of the learning procedure in long-tailed dataset distillation:", + "bbox": [ + 89, + 303, + 483, + 470 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "(1) During distillation, the weight distributions of the student and expert diverge, resulting in less informative tail class samples. Specifically, though the expert is trained on an imbalanced dataset, the student learns on a balanced one, resulting in differently distributed model weights. Such a discrepancy introduces controversy in the bi-level optimization objective, causing the distillation process to additionally focus on matching the data quantity distribution. Consequently, the tail classes of the distilled dataset contain significantly less useful information than the head classes, despite each class having the same number of samples.", + "bbox": [ + 89, + 472, + 483, + 638 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "(2) The expert models yield suboptimal performance on the tail classes. Concretely, the expert is biased toward the head classes when trained on imbalanced datasets [1, 49], resulting in degraded tail class accuracy. Since the student's performance is upper-bounded by that of the expert, the expressiveness of the distilled dataset is consequently restricted. Additionally, the image labels are obtained via the expert's prediction logits [13]. As the expert is biased toward the head classes, the prediction logits for the tail classes are, therefore, less reliable and confident.", + "bbox": [ + 89, + 641, + 483, + 792 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "To address the aforementioned issues, we propose a trajectory-matching approach named Distribution-agnostic Matching with Expert Decoupling. Firstly, the student model is trained via a Distribution-agnostic Matching (DAM) strategy, which consists of a loss designed for matching experts trained on long-tailed data. Specifically, when training the student model on the balanced synthetic", + "bbox": [ + 89, + 795, + 483, + 902 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "dataset, we mimic the procedure of training a model on an imbalanced dataset to reduce the distance between the student and the biased expert trajectories. Secondly, to mitigate the poor performance of the expert model and to enhance the overall distillation performance, an Expert Decoupling (ED) strategy is also proposed. In practice, we train representation experts and classification experts in a decoupled manner [20] to improve the performance of experts. We jointly match with the backbone layers of the representation experts and the classifier layers of the classification experts simultaneously. We use classification expert models to generate soft labels with high confidence for all classes.", + "bbox": [ + 511, + 90, + 906, + 271 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "This work pioneers the field of long-tailed dataset distillation as the first successful effort in this area. Our experiments on four long-tailed datasets, conducted under various imbalance factors and images-per-class (IPC) settings, demonstrate the effectiveness of our method. We show that our method is able to achieve state-of-the-art performance across different dataset distillation baselines.", + "bbox": [ + 511, + 272, + 908, + 378 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "2. Long-tailed Dataset Distillation", + "text_level": 1, + "bbox": [ + 511, + 390, + 802, + 407 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "2.1. Problem Formulation", + "text_level": 1, + "bbox": [ + 511, + 415, + 718, + 430 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Assume we are given a long-tailed target training dataset $\\mathcal{D} = \\{(\\mathbf{x}_i, y_i)\\}_{i=1}^{|\\mathcal{D}|}$ with $C$ classes, where $y \\in \\{0, \\dots, C-1\\}$ . Denote the subset of images belonging to class $c$ as $\\mathcal{D}_c$ , then $|\\mathcal{D}_0| > |\\mathcal{D}_1| > \\dots > |\\mathcal{D}_{C-1}|$ and $|\\mathcal{D}_0| \\gg |\\mathcal{D}_{C-1}|$ . Denote the test dataset as $\\mathcal{T} = \\{(\\mathbf{x}_i, y_i)\\}_{i=1}^{|\\mathcal{T}|}$ and is balanced. Our goal is to generate a small dataset $\\mathcal{S} = \\{(\\hat{\\mathbf{x}}_i, \\hat{y}_i)\\}_{i=1}^{|\\mathcal{S}|}$ where $|\\mathcal{S}| \\ll |\\mathcal{D}|$ and $|\\mathcal{S}_0| = |\\mathcal{S}_1| = \\dots = |\\mathcal{S}_{C-1}|$ , so that models trained on $\\mathcal{S}$ perform optimally on the test set $\\mathcal{T}$ .", + "bbox": [ + 511, + 438, + 906, + 566 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Typical dataset distillation methods [4, 6, 13, 28] use the following procedure to optimize $S$ : First, train an expert model $\\mathcal{M}_{\\mathcal{D}}$ on $\\mathcal{D}$ with acceptable performance. Then initialize $S$ by randomly selecting images from $\\mathcal{D}$ or with Gaussian noise, and predict the corresponding (soft) labels using $\\mathcal{M}_{\\mathcal{D}}$ . A student model $\\mathcal{M}_S$ is then trained based on $S$ , and the distilled image set $S$ is optimized by minimizing:", + "bbox": [ + 511, + 566, + 908, + 674 + ], + "page_idx": 1 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {L} (\\mathcal {S}) = \\left| \\left| \\mathcal {F} \\left(\\mathcal {M} _ {\\mathcal {D}}; \\mathcal {D}\\right) - \\mathcal {F} \\left(\\mathcal {M} _ {\\mathcal {S}}; \\mathcal {S}\\right) \\right| \\right| _ {p} \\tag {1}\n$$\n", + "text_format": "latex", + "bbox": [ + 578, + 686, + 906, + 704 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "where $\\mathcal{F}$ are model characteristics that can be used for matching, such as feature distributions [41], gradients [21], or training trajectories [4, 13]. Trajectory matching [4, 10, 13] matches the student model trained on the distilled dataset with the expert model trajectories trained on the original dataset, providing an impressive performance, and the matching loss is formulated as:", + "bbox": [ + 511, + 708, + 906, + 814 + ], + "page_idx": 1 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {L} _ {\\mathrm {m a t c h}} (\\hat {\\theta} _ {t + N}, \\theta_ {t + M} ^ {*}, \\theta_ {t} ^ {*}) = \\frac {\\| \\hat {\\theta} _ {t + N} - \\theta_ {t + M} ^ {*} \\| _ {2} ^ {2}}{\\| \\theta_ {t} ^ {*} - \\theta_ {t + M} ^ {*} \\| _ {2} ^ {2}}, \\quad (2)\n$$\n", + "text_format": "latex", + "bbox": [ + 555, + 821, + 906, + 861 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "where $\\theta_t^*$ and $\\theta_{t + M}^*$ are the expert parameters on epoch $t$ and $t + M$ . $\\hat{\\theta}_{t + N}$ is the student parameter obtained in the", + "bbox": [ + 511, + 867, + 908, + 902 + ], + "page_idx": 1 + }, + { + "type": "page_number", + "text": "30610", + "bbox": [ + 478, + 944, + 519, + 957 + ], + "page_idx": 1 + }, + { + "type": "image", + "img_path": "images/cfd506487c3408699f59c1d930f356afa887e214ec59dd0a7ced6d54d723522d.jpg", + "image_caption": [ + "(a) expert trajectory" + ], + "image_footnote": [], + "bbox": [ + 107, + 103, + 352, + 263 + ], + "page_idx": 2 + }, + { + "type": "image", + "img_path": "images/41c6abd47e0fb19726d0b2c0d47b77f3e9dcb0b31b4ea42ecf6d40654c1cd993.jpg", + "image_caption": [ + "(b) training trajectory of traditional DD", + "Figure 3. Effect of biased expert. (a) An expert trained on an imbalanced dataset leads to increasingly imbalanced weight gradients over classes. (b) Existing dataset distillation methods [4, 13] are ignorant of the distribution gap between $S_{t}$ and $\\mathcal{D}$ . This causes the student gradient imbalance to increase in each step. If we match such trajectories, the synthetic dataset will be updated by the increasingly imbalanced gradients over classes, and the model trained on this synthetic dataset is highly biased. (c) With Distribution-agnostic Matching, the increasingly imbalanced gradients over classes will be re-weighted, such that the student model is updated with balanced gradients." + ], + "image_footnote": [], + "bbox": [ + 375, + 103, + 614, + 263 + ], + "page_idx": 2 + }, + { + "type": "image", + "img_path": "images/c96421f568ac5929d4e77a8274da5d6f16496e1d00b2be70ae8e0cb658e6b81d.jpg", + "image_caption": [ + "(c) training trajectory of our method" + ], + "image_footnote": [], + "bbox": [ + 645, + 103, + 883, + 263 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "inner optimization updated using the cross-entropy loss.", + "bbox": [ + 89, + 358, + 460, + 372 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "In Section 2.2, we first discuss why such a formulation in Equation 1 is unreasonable for a long-tailed target dataset. Then we propose Distribution-agnostic Matching to mitigate the negative impact of the long-tailed distribution. To further enhance the distillation performance, an Expert Decoupling strategy is proposed in Section 2.3, aiding our method to achieve improved and even lossless performance.", + "bbox": [ + 89, + 372, + 483, + 479 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "2.2. Distribution-agnostic Matching", + "text_level": 1, + "bbox": [ + 89, + 486, + 369, + 502 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Influence of long-tailed distribution. By extending Equation 1 to the whole distillation process, a single step of the synthetic dataset optimization can be formulated as:", + "bbox": [ + 89, + 508, + 483, + 555 + ], + "page_idx": 2 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {S} _ {t + 1} = \\underset {\\mathcal {S} _ {t}} {\\arg \\min } | | \\theta_ {\\mathcal {D}} - \\underset {\\theta_ {S} ^ {t}} {\\arg \\min } \\mathcal {L} \\left(\\theta_ {\\mathcal {S}} ^ {t}, \\mathcal {S} _ {t}\\right) | | _ {p}, \\tag {3}\n$$\n", + "text_format": "latex", + "bbox": [ + 135, + 561, + 482, + 590 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "where $\\theta_S^t$ is the student model parameter at epoch $t$ and $\\mathcal{L}$ is the cross-entropy loss based on the student model parameter and the synthetic dataset $S_{t}$ . $\\theta_{\\mathcal{D}}$ is the expert model parameters trained on $\\mathcal{D}$ and is determined before distillation. $p$ is the factor for $L_{p}$ norm and usually takes the value of 2.", + "bbox": [ + 89, + 598, + 483, + 674 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Equation 3 is an embedded optimization problem in which the inner loop is performed first. This implies that $S_{t}$ is updated to train a $\\theta_{S}^{t}$ that can match the fixed $\\theta_{\\mathcal{D}}$ as much as possible. We see that $\\theta_{S}^{t}$ is an important intermediate that connects the synthetic dataset and the target dataset. Under the setting that the target dataset $\\mathcal{D}$ is imbalanced, the weights in $\\theta_{\\mathcal{D}}$ for different classes follow an imbalanced distribution and are fixed during distillation. Then the outer loop in Equation 3 would force $\\theta_{S}^{t}$ to produce an imbalanced distribution similar to $\\theta_{\\mathcal{D}}$ . As illustrated in Figure 3(a) and (b), the student is forced to mimic the expert during distillation, so that its weight distribution in different classes gradually becomes imbalanced.", + "bbox": [ + 89, + 674, + 482, + 869 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "However, the outer loop in Equation 3 aims to find a valid $S_{t}$ rather than a proper $\\theta_{\\mathcal{S}}^{t}$ . Thus, $\\mathcal{S}_t$ is optimized to resem", + "bbox": [ + 89, + 869, + 483, + 901 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "ble the training behavior of $\\mathcal{D}$ . Although the tail class performance on the long-tailed dataset $\\mathcal{D}$ is poor, this causes the tail classes of distilled $S_{t}$ to similarly contain much less useful information than the head classes, even though it is uniformly distributed. As a result, Equation 3 proactively distills less information from tail classes, making the distillation process more biased towards head classes.", + "bbox": [ + 511, + 357, + 906, + 462 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Distribution-agnostic Matching. As discussed above, matching the student network with an expert trained on an imbalanced dataset could lead to severe flaws in the synthetic dataset $S$ , where the tail classes are unfortunately distorted. Ideally, we would like $S$ to be a compact data summary such that every distilled image is meaningful and useful. However, such a dataset yields a student model with a balanced distribution (as shown in Figure 4 left), which mismatches the expert. This raises an interesting question: how can we preserve tail class information in the distilled dataset while avoiding the weight mismatching?", + "bbox": [ + 511, + 463, + 908, + 628 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "A key insight is to weaken the relationship between the data distribution and the weight distribution. In other words, we aim to prevent weight distribution mismatches from impacting dataset optimization. To achieve this, we propose Distribution-agnostic Matching, which seeks to align the gradient distributions of the student and expert without altering the distilled dataset. Practically, we modify the student training procedure to absorb the effect of weight imbalance by deliberately making $\\theta_S^t$ imbalanced. Inspired by the Balanced Softmax loss [34], we propose a long-tailed distributed loss to aid student training:", + "bbox": [ + 511, + 628, + 908, + 796 + ], + "page_idx": 2 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {L} ^ {\\mathrm {c}} (\\mathbf {L}, \\mathbf {y}, \\mathbf {s}) = - \\frac {1}{n} \\sum_ {i = 1} ^ {n} g \\left(\\mathbf {s} _ {\\mathbf {y} _ {i}}\\right) \\cdot \\log P \\left(\\mathbf {L} _ {i, \\mathbf {y} _ {i}}\\right), \\tag {4}\n$$\n", + "text_format": "latex", + "bbox": [ + 560, + 805, + 906, + 844 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "where", + "bbox": [ + 514, + 854, + 558, + 867 + ], + "page_idx": 2 + }, + { + "type": "equation", + "text": "\n$$\nP (\\mathbf {L} _ {i, \\mathbf {y} _ {i}}) = \\frac {e ^ {\\mathbf {L} _ {i , \\mathbf {y} _ {i}} - \\lambda \\log (\\mathbf {s} _ {\\mathbf {y} _ {i}})}}{\\sum_ {j = 1} ^ {C} e ^ {\\mathbf {L} _ {i , j} - \\lambda \\log (\\mathbf {s} _ {j})}}. \\tag {5}\n$$\n", + "text_format": "latex", + "bbox": [ + 594, + 864, + 906, + 905 + ], + "page_idx": 2 + }, + { + "type": "page_number", + "text": "30611", + "bbox": [ + 478, + 944, + 517, + 955 + ], + "page_idx": 2 + }, + { + "type": "image", + "img_path": "images/aa144c0cd9141b32ab95844a6978b10697e5bd6b31918a4b8c1da6c530e7b616.jpg", + "image_caption": [ + "Figure 4. Comparison of the internal loop between traditional DD methods and ours. The expert model $\\mathcal{M}_{\\mathcal{D}}$ is trained on a long-tailed dataset, leading to biased weights. Left: For traditional DD methods, directly matching with these weights causes large information loss in the tail classes. The backward propagation updates this imbalance from the expert trajectories to the synthetic dataset. Right: With Distribution-agnostic Matching, the gradients are obtained via $\\mathcal{L}^c$ , which revises the student weight for matching in the internal loop. This mitigates the distance between the student weight and the expert trajectory to reduce the influence of imbalance on the synthetic dataset." + ], + "image_footnote": [], + "bbox": [ + 94, + 88, + 903, + 263 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "In the above equations, $\\mathbf{L}$ is the student model logits, $n$ is the number of samples in the distilled dataset, $\\mathbf{y}$ is the ground-truth label, $\\mathbf{s}_{\\mathbf{y}_i}$ is the number of samples in the long-tailed dataset for class $\\mathbf{y}_i$ , $g(\\cdot)$ normalizes $\\mathbf{s}_{\\mathbf{y}_i}$ , and $\\lambda$ is used for smoothing the predictions.", + "bbox": [ + 88, + 342, + 480, + 417 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "The proposed loss is designed to support student model training during distillation by reducing the dependency of the weight distribution on the data distribution, thereby \"mimicking\" the training behavior on long-tailed datasets. Classes with more samples contribute more significantly to the loss, simulating a long-tailed training gradient for the student network during matching. This approach helps to minimize the trajectory differences between the student and expert networks caused by distribution mismatches between the target and distilled datasets. The right plot in Figure 4 illustrates the design and effect of our method.", + "bbox": [ + 88, + 420, + 482, + 585 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "2.3. Improved Guidance with Expert Decoupling", + "text_level": 1, + "bbox": [ + 89, + 601, + 468, + 618 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Unlike experts trained on balanced target datasets, the experts trained on long-tailed datasets are less effective for guiding the dataset distillation process. Specifically, when adopting typical dataset distillation methods, the long-tailed experts perform poorly on tail classes, degrading student performance on these classes and leading to unreliable initial soft labels. This degraded student performance restricts the effective learning of synthetic tail-class images, while the unreliable labels introduce misleading information and provide weak supervision.", + "bbox": [ + 88, + 626, + 482, + 777 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "In the conventional dataset distillation process, soft labels are assigned to initialized synthetic images based on the probability output of the expert model [13]. We found that soft-label predictions from long-tailed experts are distributed differently compared to those from balanced experts. As visualized in Figure 5, while the head class soft labels predicted by the balanced expert have similar confidence values with tail classes (0.88 and 0.89), the long-", + "bbox": [ + 88, + 779, + 482, + 901 + ], + "page_idx": 3 + }, + { + "type": "image", + "img_path": "images/d56237401879d6ae9bceef47c64796159f9e327f0edb327f957562de3ca20d73.jpg", + "image_caption": [ + "Figure 5. Soft-label initialization on CIFAR-10-LT. We visualize the average predictive confidence of experts on different classes. For traditional DD, while the soft labels can be initialized well on head classes, they are predicted poorly on the tailed classes, leading to insufficient supervision." + ], + "image_footnote": [], + "bbox": [ + 514, + 340, + 906, + 501 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "tailed expert predictions between head and tail classes differ drastically. Whereas the head class confidence value can be as large as 0.97, the confidence value of the tail class is surprisingly small (0.38). Such a confidence discrepancy leads to less effective learning in tail classes than in head classes. Moreover, for tail classes with a low number of samples, the number of samples with high confidence is reduced and leads to the diversity decrease of the dataset initialization.", + "bbox": [ + 511, + 582, + 905, + 702 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Decoupled training [20] is a widely used technique to alleviate the prediction bias introduced by imbalanced data. It consists of two stages: representation learning and classifier fine-tuning. At the representation learning stage, the entire model is trained on the original dataset $\\mathcal{D}$ . At the classifier fine-tuning stage, the backbone of the model is frozen, and only the classifier of the model is trained on a balanced dataset $\\mathcal{B}$ , which is created from $\\mathcal{D}$ via oversampling or undersampling. However, the integration of this decoupling approach with dataset distillation is not trivial, as it requires careful alignment of the student model with the expert model. To effectively utilize the idea of Expert Decoupling, we propose a two-step strategy to improve teacher", + "bbox": [ + 511, + 704, + 906, + 901 + ], + "page_idx": 3 + }, + { + "type": "page_number", + "text": "30612", + "bbox": [ + 478, + 944, + 517, + 955 + ], + "page_idx": 3 + }, + { + "type": "table", + "img_path": "images/6fbf2d58651c85bcf4c269d44f50bfeecd037a18a48a9bdf326f1347dff08816.jpg", + "table_caption": [], + "table_footnote": [], + "table_body": "
DatasetCIFAR-10-LT
Imbalance Factor1050100200
IPC102050102050102050102050
Random32.5±2.239.6±0.951.9±1.533.2±0.442.0±1.351.6±1.334.4±2.041.4±0.752.6±0.532.5±0.842.2±1.149.9±1.4
K-Center Greedy [36]21.9±0.824.2±0.831.7±0.917.8±0.220.8±0.526.1±0.216.2±0.519.0±1.024.2±1.216.8±0.317.5±1.422.6±1.6
Graph-Cut [17]28.7±0.934.2±1.040.6±1.024.2±0.728.6±0.833.9±0.422.9±0.926.0±0.533.3±1.022.3±0.925.2±0.529.2±0.4
DC [46]37.9±0.938.5±0.937.4±1.437.3±0.938.8±1.035.8±1.236.7±0.838.1±1.035.3±1.435.6±0.835.7±0.933.3±1.4
MTT [4]58.0±0.859.5±0.462.0±0.945.8±1.449.9±0.853.6±0.537.7±0.641.6±1.147.8±1.1N/A22.6±1.023.9±0.8
DREAM [28]34.6±0.642.2±1.550.5±0.730.8±0.638.4±0.345.5±0.930.8±1.734.9±0.842.2±0.832.7±1.332.4±0.338.9±0.4
IDM [47]54.8±0.457.1±0.360.1±0.351.9±0.753.3±0.656.1±0.449.8±0.650.9±0.553.1±0.447.0±0.548.1±0.549.9±0.3
Minimax [12]29.2±0.528.5±0.639.9±0.118.4±0.322.5±0.225.2±0.219.9±0.423.3±0.228.0±0.619.1±0.420.5±0.222.7±0.3
DATM [13]57.2±0.460.4±0.266.7±0.641.6±0.243.4±0.350.3±0.237.3±0.238.9±0.144.3±0.1N/A34.8±0.140.1±0.2
Ours58.1±0.363.0±1.070.5±0.454.2±1.059.4±0.765.8±0.253.4±0.158.2±0.664.0±0.952.2±0.656.6±0.462.3±0.3
Full Dataset76.7±0.369.8±0.366.2±0.462.1±0.4
", + "bbox": [ + 91, + 88, + 903, + 263 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Table 1. Quantitative comparisons with the SOTA methods on CIFAR-10-LT. Our method outperforms all existing approaches, and can even achieve lossless performance under certain settings. Performance is evaluated under various imbalance factors. As the first to explore long-tailed dataset distillation, all experiments are conducted using open-source code. N/A indicates distillation failure.", + "bbox": [ + 89, + 263, + 906, + 306 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "guidance for student training.", + "bbox": [ + 89, + 311, + 287, + 325 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Firstly, we train two types of experts in a decoupled manner. The representation expert with parameter $\\theta^{\\mathcal{D}}$ is obtained from the representation learning stage, and the classifier expert with parameter $\\theta^{\\mathcal{B}}$ is obtained from the classifier finetuning stage. Since the classifier experts are not biased and are able to provide more accurate classification, we use the classifier experts $\\theta^{\\mathcal{B}}$ for the synthetic dataset $S$ initialization. Specifically, we use the logits $L_{i} = \\theta^{\\mathcal{B}}\\mathbf{x}_{i}$ to initialize soft labels, with $\\mathrm{softmax}(L_i)$ as the label of $\\mathbf{x}_i$ , and then sample from $\\mathcal{D}$ to initialize $S$ .", + "bbox": [ + 88, + 325, + 482, + 474 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Secondly, we match the teacher expert's training trajectory with the student's in an adaptive manner. Since the representation expert learns the image features from the original dataset $\\mathcal{D}$ but fails to classify with biased classifiers [20], we skip matching the classifier layer when matching the student network trained on the synthetic dataset with the representation expert. On the contrary, the classifier expert $\\mathcal{M}_{\\mathcal{B}}$ only trains the classifier layer, so we also only train the student classifier layer and only match the classifier layer during the classifier matching. During the whole matching process, we jointly match the representation trajectory and the classifier trajectory using the following loss:", + "bbox": [ + 88, + 476, + 482, + 657 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\begin{array}{l} \\mathcal {L} = \\lambda_ {\\text {r e p}} \\mathcal {L} _ {\\text {m a t c h}} \\left(\\hat {\\theta} _ {t + N}, \\theta_ {t + M} ^ {\\mathcal {D} _ {*}}, \\theta_ {t} ^ {\\mathcal {D} _ {*}}\\right) \\\\ + \\lambda_ {\\mathrm {c l s}} \\mathcal {L} _ {\\text {m a t c h}} \\left(\\hat {\\theta} _ {t + N}, \\theta_ {t + M} ^ {\\mathcal {B} *}, \\theta_ {t} ^ {\\mathcal {B} *}\\right), \\tag {6} \\\\ \\end{array}\n$$\n", + "text_format": "latex", + "bbox": [ + 163, + 666, + 482, + 705 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "where $\\lambda_{\\mathrm{rep}}$ and $\\lambda_{\\mathrm{cls}}$ are hyper-parameters that control the weighting ratio, $\\hat{\\theta}_{t + N}$ is the student model at its training epoch $t + N$ , $\\theta_t^{\\mathcal{D}^*}$ and $\\theta_t^{\\mathcal{B}^*}$ are the representation backbone and classifier of the expert at epoch $t$ , respectively. $\\hat{\\theta}_{t + N}$ is updated by our proposed long-tailed distributed loss $\\mathcal{L}^{\\mathrm{c}}$ in Equation 4, and $\\mathcal{L}_{\\mathrm{match}}$ follows the form in Equation 2.", + "bbox": [ + 89, + 714, + 482, + 810 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "3. Experiments", + "text_level": 1, + "bbox": [ + 89, + 821, + 223, + 838 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "3.1. Experiment Datasets", + "text_level": 1, + "bbox": [ + 89, + 847, + 287, + 863 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "We evaluate our methods on four widely used long-tailed datasets of different scales: CIFAR-10-LT, CIFAR-100-LT", + "bbox": [ + 89, + 869, + 482, + 898 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "[7], TinyImageNet-LT, and ImageNet-LT [29]. The experiments are conducted under various degrees of imbalance and IPC settings. Following [7, 9, 29], these long-tailed datasets are sampled from the original balanced CIFAR [22] and ImageNet [8] datasets. The sampling procedure is done by randomly selecting different amounts of images from each class, where the image number $|\\mathcal{D}_c|$ for class $c$ is determined by an exponential decay function $|\\hat{\\mathcal{D}}_c| = |\\mathcal{D}_c|\\mu^c$ , and $\\mu^c = \\beta^{-(c / C)}$ , where $C$ is the number of classes and $\\beta$ is the imbalanced factor. The degree of imbalance is measured via the imbalanced factor $\\beta = \\mathcal{D}_0 / \\mathcal{D}_C$ [7]. The larger $\\beta$ is, the more imbalanced the dataset is.", + "bbox": [ + 511, + 310, + 906, + 491 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "3.2. Main Results", + "text_level": 1, + "bbox": [ + 511, + 500, + 651, + 513 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "CIFAR-10-LT. The experiment results on CIFAR-10-LT are presented in Table 1, showing that our proposed method consistently outperforms all other dataset distillation baselines. Table 1 further highlights the advantages of our approach on long-tailed datasets. When the imbalance factor is low, prior dataset distillation methods like DATM perform reasonably well. For instance, DATM only experiences a $10.0\\%$ drop in accuracy when $\\beta = 10$ and IPC=50, compared to the accuracy achieved with the full dataset. In these scenarios, our method surpasses the best baseline, DATM, by $0.9\\%$ on $\\beta = 10$ and $3.8\\%$ on $\\beta = 50$ .", + "bbox": [ + 511, + 522, + 906, + 688 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "When the imbalance factor is high (e.g., $\\beta = 200$ ), IDM [47] achieves performance comparable to random selection. Other methods, such as MTT [4], DREAM [28], DATM [13], and Minimax [12], struggle with higher imbalance factors, performing worse than random selection. In contrast, our method sustains strong performance on highly imbalanced datasets, surpassing the best existing method IDM by a substantial margin of $10.6\\%$ on $\\beta = 200$ . Remarkably, we achieve lossless performance on $\\beta = 200$ using only 50 images per class.", + "bbox": [ + 511, + 688, + 906, + 839 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "CIFAR-100-LT. The results are presented in Table 2. Our method consistently outperforms all existing dataset distillation methods across all imbalance factors on CIFAR-100-LT. Similar to CIFAR-10-LT, our approach demonstrates", + "bbox": [ + 511, + 839, + 905, + 900 + ], + "page_idx": 4 + }, + { + "type": "page_number", + "text": "30613", + "bbox": [ + 478, + 944, + 517, + 955 + ], + "page_idx": 4 + }, + { + "type": "table", + "img_path": "images/ee663d72e232dc4c127d2c434befaa598c0f4387f898210b44d22289aac1942b.jpg", + "table_caption": [], + "table_footnote": [], + "table_body": "
DatasetCIFAR-100-LTTiny-ImageNet-LT
Imbalance Factor10201020
IPC102050102050102050102050
Random14.2±0.621.7±0.632.1±0.615.0±0.321.6±0.530.5±0.57.4±0.213.5±0.420.6±0.47.6±0.113.2±0.419.8±0.1
K-Center Greedy [36]10.7±0.915.9±1.024.8±0.210.0±0.515.1±0.623.8±0.310.4±1.511.3±1.27.7±1.012.6±1.49.9±2.213.2±2.7
Graph-Cut [17]16.9±0.322.2±0.429.9±0.416.0±0.520.7±0.528.7±0.39.8±0.75.6±0.610.9±1.13.4±0.85.0±1.04.2±1.1
DC [46]24.0±0.327.4±0.327.4±0.323.2±0.326.2±0.327.4±0.3------
MTT [4]14.3±0.116.7±0.213.8±0.212.6±0.315.0±0.210.6±0.511.1±0.218.1±0.223.1±0.17.7±0.114.7±0.215.6±0.3
DREAM [28]10.1±0.412.0±1.013.1±0.49.4±0.410.3±0.612.3±0.35.4±0.36.8±0.17.8±0.24.8±0.16.0±0.27.4±0.1
DATM [13]28.2±0.434.1±0.231.6±0.125.3±0.327.2±0.127.1±0.121.3±0.114.5±0.126.8±0.114.0±0.619.0±0.323.1±0.1
Ours31.5±0.237.5±0.440.0±0.131.4±0.535.1±0.437.0±0.726.0±0.327.9±0.230.3±0.223.6±0.325.5±0.328.0±0.6
Full Dataset40.3±0.436.5±0.431.2±1.028.3±0.3
", + "bbox": [ + 93, + 88, + 903, + 241 + ], + "page_idx": 5 + }, + { + "type": "table", + "img_path": "images/c67b8dee813e55634940c1b40264f3786a2ff61568987f6af15e878d3dba3d41.jpg", + "table_caption": [ + "Table 2. Quantitative comparisons with the SOTA methods on CIFAR-100-LT and Tiny-ImageNet-LT. When scaling up to larger datasets, our approach still outperforms all the included baselines, and achieves nearly lossless performance under several settings." + ], + "table_footnote": [], + "table_body": "
DatasetImageNet-LT
Imbalance Factor510
IPC10201020
Random3.9±0.17.0±0.13.9±0.16.8±0.1
G-VBSM [39]-1.0±0.1-1.0±0.1
TESLA [6]3.0±0.1-2.7±0.1-
DATM [13]7.4±0.18.1±0.27.9±0.18.2±0.1
SRe2L [44]6.7±0.110.1±0.17.7±0.310.9±1.0
Ours20.8±0.221.0±0.120.3±0.120.7±0.1
Full Dataset27.6±0.326.6±0.2
", + "bbox": [ + 94, + 277, + 478, + 426 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "greater improvements as the imbalance factor increases.", + "bbox": [ + 89, + 463, + 460, + 477 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Tiny-ImageNet-LT. We conducted experiments on TinyImageNet-LT to evaluate the scalability of our method in comparison with other baselines. The results, shown in Table 2, align with our findings on CIFAR-10-LT and CIFAR100-LT, where our method surpasses baseline methods by a growing margin as the imbalance factor increases.", + "bbox": [ + 89, + 478, + 482, + 568 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "ImageNet-LT. We further scale up to ImageNet-LT, with results presented in Table 3, where our model consistently outperforms the baselines. Note that as the dataset size increases, many dataset distillation methods become resource-intensive and are impractical to apply.", + "bbox": [ + 89, + 568, + 482, + 645 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Cross-architecture performance. Table 4 presents our cross-architecture experiments, demonstrating the generalizability of our synthetic dataset. We evaluated various methods on the CIFAR-10-LT dataset with an imbalance factor of $\\beta = 100$ and IPC=50, comparing against random selection, MTT, and DATM. As shown in the table, our distilled dataset consistently outperforms other methods by a significant margin. Specifically, our method achieves a $12.7\\%$ accuracy improvement on ConvNet, a $17.1\\%$ improvement on ResNet-18, a $7.7\\%$ improvement on VGG-11, and a $6\\%$ improvement on AlexNet. The results also indicate that the imbalance distilled by ConvNet in traditional DD methods negatively impacts the performance of other models. Similar to experiments on ConvNets, MTT and DATM perform poorly on unseen models due to the imbalanced dataset and the biased expert trained on it, often yielding worse results than random selection.", + "bbox": [ + 89, + 645, + 482, + 901 + ], + "page_idx": 5 + }, + { + "type": "table", + "img_path": "images/8239db1aa423d66dfa23084e3bfbc2e50a91948450472f3823d3b2b66d31281f.jpg", + "table_caption": [ + "Table 3. Quantitative comparisons on ImageNet-LT. The missing results are due to out-of-memory." + ], + "table_footnote": [], + "table_body": "
MethodConvNet-3ResNet-18VGG-11AlexNet
Random52.049.247.847.2
MTT [4]49.142.942.340.0
DATM [13]44.442.442.244.9
Ours64.760.055.553.2
", + "bbox": [ + 517, + 277, + 901, + 359 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Table 4. Cross-architecture evaluation on CIFAR-10-LT.", + "bbox": [ + 531, + 359, + 887, + 375 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "3.3. Ablation Studies", + "text_level": 1, + "bbox": [ + 513, + 385, + 678, + 400 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Ablation on different components. Figure 6a demonstrates the effectiveness of the two proposed components of our method. These experiments were performed on the CIFAR-10-LT dataset with an imbalance factor of $\\beta = 200$ and IPC=50. The \"base\" method is a trajectory matching method without Distribution-agnostic Matching and Expert Decoupling. \"+\"DAM\" refers to our method without Expert Decoupling, while \"ours\" represents the complete proposed method, with both DAM and ED applied.", + "bbox": [ + 511, + 410, + 906, + 547 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "The baseline, which suffers from weight inconsistency and poor expert performance as mentioned in Section 1, only achieves $43.8\\%$ accuracy. After incorporating Distribution-agnostic Matching, the inconsistency between student and expert weights is mitigated, resulting in a $14.3\\%$ performance improvement. Further integrating Expert Decoupling boosts the performance to $62.6\\%$ .", + "bbox": [ + 511, + 547, + 905, + 654 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Effect of classifier layer matching. As discussed in Section 2.3, during representation learning, the model is influenced by the long-tailed distribution, resulting in a biased classifier. During classifier fine-tuning, the classifier is finetuned on balanced data with the backbone frozen. Therefore, it is crucial to explore strategies for matching each expert's different modules during dataset distillation. The experiments are presented in Figure 6b. \"B\", \"C\", and \"W\" refer to matching with the representation backbone, the classifier layer, and the whole model, respectively. \"B/C\" refers to matching the student backbone with the representation expert's backbone, and the student classifier layer with the classification expert's classifier.", + "bbox": [ + 511, + 656, + 905, + 852 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "As shown in Figure 6b, the best performance is achieved by matching the backbone with representation experts and the classifier layer with classification experts. This outcome", + "bbox": [ + 511, + 854, + 905, + 900 + ], + "page_idx": 5 + }, + { + "type": "page_number", + "text": "30614", + "bbox": [ + 478, + 944, + 519, + 955 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/529cc07d02758511f59812e643dd0d6ac10f6aad8771de281648807991f4d91e.jpg", + "image_caption": [ + "(a) Effect of the different components" + ], + "image_footnote": [], + "bbox": [ + 91, + 87, + 240, + 199 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/1b5defdd0fd505dd03a2935766ca3c151a696f92257a7fe974a217ebcf3bb77d.jpg", + "image_caption": [ + "(b) Comparison of different matching strategies" + ], + "image_footnote": [], + "bbox": [ + 251, + 88, + 406, + 199 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/44073d57178891c1d5207b78cda063e4cd9a0b4d5ee901ffa1069df2a282508e.jpg", + "image_caption": [ + "(c) Comparison of matching strategies on low $\\beta$", + "Figure 6. (a) Ablation study on the proposed components, Distribution-agnostic Matching and Expert Decoupling. (b) Ablation of different matching strategies with expert model layers. (c) & (d) Comparisons of matching strategies for long-tailed dataset distillation under low imbalance factor (c) and high imbalance factor (d). (e) Ablation study on hyperparameters $\\lambda_{\\mathrm{rep}}$ and $\\lambda_{\\mathrm{cls}}$ ." + ], + "image_footnote": [], + "bbox": [ + 416, + 90, + 570, + 199 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/ae100f1ef62c71dc04837cb5af135c6b917e85e27cc74f4a7d39d6e05eece910.jpg", + "image_caption": [ + "(d) Comparison of matching strategies on high $\\beta$" + ], + "image_footnote": [], + "bbox": [ + 578, + 89, + 730, + 199 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/9ddd80bffe9452b76601df8cf68b7a5055056dc3fedce9876380f149ef8e2f7c.jpg", + "image_caption": [ + "(e) Effect of the relative scale of $\\lambda_{\\mathrm{rep}}$ and $\\lambda_{\\mathrm{cls}}$" + ], + "image_footnote": [], + "bbox": [ + 738, + 90, + 892, + 199 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "aligns with the principles of decoupled learning. Since the classifier layer of representation experts is heavily affected by the long-tailed dataset, including it in the matching process causes a performance drop. In addition, matching the entire classifier expert negatively impacts performance, as only the classifier layer is trained at this stage.", + "bbox": [ + 88, + 277, + 482, + 368 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Effect of different matching strategies. When distilling a long-tailed dataset, various intuitive strategies can be employed with decoupling trained experts. We explored several strategies, including sequential matching, classification expert matching, oversampled dataset, and joint matching. Sequential matching first matches representation experts, followed by classification experts. Classification expert matching only matches with the classification expert. The oversampled dataset approach involves oversampling the tail classes before distilling the dataset rather than using the decoupling strategy. Joint matching, used in our Expert Decoupling, involves matching both experts simultaneously in the outer loop of the matching.", + "bbox": [ + 88, + 369, + 482, + 566 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "As shown in Figures 6c and 6d, joint matching (joint), which we use in Expert Decoupling, achieves the best overall performance. Sequential matching (sequential) and directly matching with classification experts (direct) perform poorly when the imbalance factor is low, meaning the data is \"not very imbalanced.\" Sequential matching suffers due to the trajectory gap between the two types of experts. Classification expert matching is less effective when the imbalance factor is low because the representation layers learn more information in such scenarios, and solely matching the classification layers fails to capture sufficient information. As depicted in Figure 6d, the oversampled dataset strategy (oversample) performs poorly when the imbalance factor is high. This finding aligns with previous work [7], which indicates that oversampling leads to model overfitting.", + "bbox": [ + 89, + 566, + 482, + 792 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Effect of hyperparameters. The most important hyperparameters in our method are $\\lambda_{\\mathrm{rep}}$ and $\\lambda_{\\mathrm{cls}}$ in Equation 6. Together, they regulate the balance between the matching representation expert and the classification expert. The setting of $\\lambda_{\\mathrm{rep}}$ and $\\lambda_{\\mathrm{cls}}$ depends on the imbalance factors. From Figure 6e, we have the following observations. For a \"more imbalanced\" dataset (e.g., $\\beta = 200$ ), lower $\\lambda_{\\mathrm{rep}}$ and higher $\\lambda_{\\mathrm{cls}}$", + "bbox": [ + 89, + 795, + 482, + 901 + ], + "page_idx": 6 + }, + { + "type": "table", + "img_path": "images/4082d0192408e64b89120b01fba7be8085af298eeb68590bcae9a4a7271bd612.jpg", + "table_caption": [], + "table_footnote": [], + "table_body": "
Hyperparametersβ = 10Hyperparametersβ = 200
λrep = 1, λcls = 0.270.5λrep = 0.2, λcls = 162.4
λrep = 1, λcls = 0.468.9λrep = 0.4, λcls = 161.7
λrep = 1, λcls = 0.668.0λrep = 0.6, λcls = 161.3
", + "bbox": [ + 519, + 268, + 898, + 334 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Table 5. Effect of ${\\lambda }_{\\text{rep }}$ and ${\\lambda }_{\\text{cls }}$ under different imbalance factors.", + "bbox": [ + 516, + 335, + 901, + 349 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "are preferred. For \"less imbalanced\" datasets (e.g., $\\beta = 10$ ), higher $\\lambda_{\\mathrm{rep}}$ and lower $\\lambda_{\\mathrm{cls}}$ are chosen. This observation is consistent with our experiments in Figure 6c and 6d, where only matching classifier experts perform well when $\\beta$ is high. This is because the representation layers learn more information when $\\beta$ is low, and increasing $\\lambda_{\\mathrm{rep}}$ leverages information from the representation layers.", + "bbox": [ + 511, + 363, + 903, + 469 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "We further analyze the sensitivity of the hyperparameters in Table 5. We can observe a minor change in the performance as we change hyperparameters rapidly. When the relative magnitude of the two hyperparameters is preserved, their absolute values do not affect the distillation performance much, suggesting the robustness of our method.", + "bbox": [ + 511, + 470, + 905, + 561 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "3.4. Long-tailed Dataset Distillation Analysis", + "text_level": 1, + "bbox": [ + 511, + 573, + 861, + 589 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Effect of imbalance in long-tailed dataset distillation. In this part, we empirically show that the performance drop in long-tailed dataset distillation is due to the data imbalance instead of the total sample number reduction of the target dataset. We compare the performance of using regular dataset distillation methods for CIFAR-10-LT (with $\\beta = 200$ and $\\mathrm{IPC} = 50$ ) and a subset of CIFAR-10 that contains the same amount of data but in a uniform distribution. The total number of samples is 11203 under this setting.", + "bbox": [ + 511, + 595, + 905, + 733 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "As we can observe in Figure 7a, the performance gap between distilling a long-tailed dataset and a balanced subset can be as large as $28.6\\%$ , even though the two target datasets have the same total amount of samples. Thus, we conclude that it is the imbalanced distribution in the target dataset that prevents effective distillation. Based on this observation, we delved into it and proposed methods to alleviate the imbalanced distribution's negative effects.", + "bbox": [ + 511, + 734, + 905, + 853 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Handling classes with extremely small number of samples. In long-tailed dataset distillation, a scenario exists where the number of samples in tail classes is even less than", + "bbox": [ + 511, + 854, + 905, + 900 + ], + "page_idx": 6 + }, + { + "type": "page_number", + "text": "30615", + "bbox": [ + 478, + 944, + 517, + 955 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/581003bfe0f5ddb2f45a55988c0be8e6d4cfccce6d9fbc5399d7bd723efd2d00.jpg", + "image_caption": [ + "Figure 7. (a) Comparison of DD performance between a uniform dataset and a long-tailed dataset with the same number of samples. (b) Comparison of strategies to handle an extremely low number of samples. (c) Comparison of class-wise accuracy between DATM and our method. (d) Comparison between balanced IPC and long-tailed distributed IPC." + ], + "image_footnote": [], + "bbox": [ + 91, + 85, + 285, + 241 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/7e682a10346c68bccbfe6ffe0fcecf6c25ce31d9c93a719b7cdee4d840fb6be6.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 295, + 85, + 488, + 241 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/62595531ac9c8647f4f5a2775118e1eec0697b9966ead3c12928d041a1f7296b.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 496, + 85, + 689, + 241 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/a5f1b6806fffbc90a7816ede88a07a302d6c98b5f14789fbc6c4a648bd67e721.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 697, + 85, + 893, + 241 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "the required number of images per class. In this case, we employ and compare three strategies: 1) Inherent strategy, which keeps the number of samples the same in the synthetic dataset; 2) Oversampling, which duplicates the samples of the tail classes before distillation; 3) Noise generation, which randomly generates Gaussian noises to supplement the insufficient samples for tail classes. From figure 7b, we see that the inherent strategy provides a comparatively lower accuracy, while oversampling and noise generation achieve better and similar performance.", + "bbox": [ + 88, + 297, + 480, + 449 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Class-wise accuracy for long-tailed dataset. We also provide class-wise accuracy to illustrate the impact of the imbalanced data, and we show that our method can effectively improve the accuracy of tail classes. As shown in Figure 7c, DATM performs poorly on tail classes, and even achieves near-zero performance on the last class. On the contrary, our method can successfully boost the tail class accuracy while reserving performance on head classes.", + "bbox": [ + 88, + 450, + 482, + 571 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Choice of distilled dataset distribution. Should we distill a long-tailed dataset into another long-tailed dataset or a balanced one? We compared the performance of distilling the target dataset into either a similarly distributed long-tailed synthetic dataset or a balanced synthetic dataset using our method and DREAM [28]. For the long-tailed synthetic dataset, we maintained the same total number of samples as in the synthetic dataset with IPC=50. As shown in Figure 7d, both methods demonstrate that distilling the dataset into a balanced synthetic dataset provides better performance than distilling it into a long-tailed one. This result is reasonable since the model trained on the synthetic long-tailed dataset will be affected by the long-tailed distribution.", + "bbox": [ + 88, + 573, + 482, + 771 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "4. Related Work", + "text_level": 1, + "bbox": [ + 89, + 787, + 232, + 804 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "4.1. Dataset Distillation", + "text_level": 1, + "bbox": [ + 89, + 815, + 274, + 830 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Dataset distillation [43] was first formulated by Wang et al. The goal of dataset distillation is to synthesize a small dataset, such that models trained on it have good performance on the original dataset. The following works can be", + "bbox": [ + 88, + 839, + 482, + 901 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "divided into kernel ridge regression-based methods [30, 32, 50], gradient matching methods [6, 13, 28, 42, 46], representation matching methods [38, 47], and generative-based methods [12, 40]. Despite the advantages of current methods, they only focus on uniformly distributed datasets, ignoring the prevalent usage of long-tailed datasets in practice. Therefore, we propose to address the long-tailed dataset distillation problem in this paper.", + "bbox": [ + 511, + 297, + 906, + 420 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "4.2. Long-tailed Recognition", + "text_level": 1, + "bbox": [ + 511, + 425, + 735, + 441 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Under real-world settings, data often tends to follow a long-tailed distribution [2]. Models trained on such datasets usually exhibit good performance on head classes but low accuracy on tail classes. A straightforward way is to preprocess the data, using oversampling [14, 18, 33], undersampling [3, 14], or data augmentation [5, 11] to make it balancedly distributed. An alternative way is weight balancing [2, 26, 34], which re-weights the layers during training to force the parameters to update unbiasedly. Decoupling methods [20] first train the model to obtain the features of the data and then fine-tune the model on the class-balanced data. In our methods, we use the decoupling method to train the expert networks during the distillation.", + "bbox": [ + 511, + 445, + 908, + 643 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "5. Conclusion", + "text_level": 1, + "bbox": [ + 511, + 652, + 633, + 667 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "We introduce long-tailed dataset distillation, a novel yet challenging task. We find that existing DD methods fail when applied to long-tailed datasets and encounter two main problems: weight distribution mismatch between the student and the expert; and suboptimal performance of the expert on the tail classes. To mitigate these issues, we propose two strategies: Distribution-agnostic Matching and Expert Decoupling. Experimental results over four long-tailed datasets demonstrate the effectiveness of our proposed method. As the pioneering work in this research area, we are the first to effectively distill long-tailed datasets.", + "bbox": [ + 511, + 674, + 906, + 839 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Acknowledgments: This research is supported by NSF IIS-2525840, ECCS-2123521 and Cisco Research unrestricted gift. This article solely reflects the opinions and conclusions of its authors and not the funding agencies.", + "bbox": [ + 511, + 839, + 905, + 900 + ], + "page_idx": 7 + }, + { + "type": "page_number", + "text": "30616", + "bbox": [ + 478, + 944, + 519, + 957 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "References", + "text_level": 1, + "bbox": [ + 91, + 89, + 187, + 104 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[1] Shaden Alshammari, Yuxiong Wang, Deva Ramanan, and Shu Kong. Long-tailed recognition via weight balancing. In CVPR, 2022. 2", + "[2] Shaden Alshammari, Yu-Xiong Wang, Deva Ramanan, and Shu Kong. Long-tailed recognition via weight balancing. In CVPR, 2022. 8", + "[3] Mateusz Buda, Atsuto Maki, and Maciej A Mazurowski. A systematic study of the class imbalance problem in convolutional neural networks. Neural networks, 2018. 8", + "[4] George Cazenavette, Tongzhou Wang, Antonio Torralba, Alexei A Efros, and Jun-Yan Zhu. Dataset distillation by matching training trajectories. In CVPR, 2022. 1, 2, 3, 5, 6", + "[5] Hsin-Ping Chou, Shih-Chieh Chang, Jia-Yu Pan, Wei Wei, and Da-Cheng Juan. Remix: rebalanced mixup. In Computer Vision-ECCV 2020 Workshops: Glasgow, UK, August 23-28, 2020, Proceedings, Part VI 16, 2020. 8", + "[6] Justin Cui, Ruochen Wang, Si Si, and Cho-Jui Hsieh. Scaling up dataset distillation toImagenet-1k with constant memory. In ICML, 2023. 2, 6, 8, 1", + "[7] Yin Cui, Menglin Jia, Tsung-Yi Lin, Yang Song, and Serge Belongie. Class-balanced loss based on effective number of samples. In CVPR, 2019. 5, 7", + "[8] Jia Deng, Wei Dong, Richard Socher, Li-Jia Li, Kai Li, and Li Fei-Fei. Imagenet: A large-scale hierarchical image database. In CVPR, 2009. 5", + "[9] Fei Du, Peng Yang, Qi Jia, Fengtao Nan, Xiaoting Chen, and Yun Yang. Global and local mixture consistency cumulative learning for long-tailed visual recognitions. In CVPR, 2023. 5", + "[10] Jiawei Du, Yidi Jiang, Vincent YF Tan, Joey Tianyi Zhou, and Haizhou Li. Minimizing the accumulated trajectory error to improve dataset distillation. In CVPR, 2023. 2", + "[11] Spyros Gidaris and Nikos Komodakis. Dynamic few-shot visual learning without forgetting. In CVPR, 2018. 8", + "[12] Jianyang Gu, Saeed Vahidian, Vyacheslav Kungurtsev, Haonan Wang, Wei Jiang, Yang You, and Yiran Chen. Efficient dataset distillation via minimax diffusion. In CVPR, 2024. 1, 5, 8", + "[13] Ziyao Guo, Kai Wang, George Cazenavette, Hui Li, Kaipeng Zhang, and Yang You. Towards lossless dataset distillation via difficulty-aligned trajectory matching. In ICLR, 2024. 1, 2, 3, 4, 5, 6, 8", + "[14] Guo Haixiang, Li Yijing, Jennifer Shang, Gu Mingyun, Huang Yuanyue, and Gong Bing. Learning from class-imbalanced data: Review of methods and applications. Expert systems with applications, 2017. 8", + "[15] Geoffrey Hinton, Oriol Vinyals, and Jeff Dean. Distilling the knowledge in a neural network. arXiv preprint arXiv:1503.02531, 2015. 1", + "[16] Geoffrey E Hinton, Nitish Srivastava, Alex Krizhevsky, Ilya Sutskever, and Ruslan R Salakhutdinov. Improving neural networks by preventing co-adaptation of feature detectors. arXiv preprint arXiv:1207.0580, 2012. 1", + "[17] Rishabh Iyer, Ninad Khargoankar, Jeff Bilmes, and Himanshu Asanani. Submodular combinatorial information mea" + ], + "bbox": [ + 93, + 114, + 480, + 898 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "sures with applications in machine learning. In Algorithmic Learning Theory, 2021. 5, 6", + "[18] Andrew Janowczyk and Anant Madabhushi. Deep learning for digital pathology image analysis: A comprehensive tutorial with selected use cases. Journal of pathology informatics, 2016. 8", + "[19] Lie Ju, Xin Wang, Lin Wang, Tongliang Liu, Xin Zhao, Tom Drummond, Dwarikanath Mahapatra, and Zongyuan Ge. Relational subsets knowledge distillation for long-tailed retinal diseases recognition. In MICCAI, 2021. 1", + "[20] Bingyi Kang, Saining Xie, Marcus Rohrbach, Zhicheng Yan, Albert Gordo, Jiashi Feng, and Yannis Kalantidis. Decoupling representation and classifier for long-tailed recognition. In ICLR, 2020. 2, 4, 5, 8", + "[21] Jang-Hyun Kim, Jinuk Kim, Seong Joon Oh, Sangdoo Yun, Hwanjun Song, Joonhyun Jeong, Jung-Woo Ha, and Hyun Oh Song. Dataset condensation via efficient synthetic-data parameterization. In ICML, 2022. 2", + "[22] Alex Krizhevsky. Learning multiple layers of features from tiny images. University of Toronto, 2012. 1, 5", + "[23] Alex Krizhevsky, Ilya Sutskever, and Geoffrey E Hinton. Imagenet classification with deep convolutional neural networks. Communications of the ACM, 2017. 2, 1", + "[24] Yann Le and Xuan Yang. Tiny imagenet visual recognition challenge. CS 231N, 2015. 1", + "[25] Zekai Li, Xinhao Zhong, Zhiyuan Liang, Yuhao Zhou, Mingjia Shi, Ziqiao Wang, Wangbo Zhao, Xuanlei Zhao, Haonan Wang, Ziheng Qin, Dai Liu, Kaipeng Zhang, Tianyi Zhou, Zheng Zhu, Kun Wang, Guang Li, Junhao Zhang, Jiawei Liu, Yiran Huang, Lingjuan Lyu, Jiancheng Lv, Yaochu Jin, Zeynep Akata, Jindong Gu, Rama Vedantam, Mike Shou, Zhiwei Deng, Yan Yan, Yuzhang Shang, George Cazenavette, Xindi Wu, Justin Cui, Tianlong Chen, Angela Yao, Manolis Kellis, Konstantinos N. Plataniotis, Bo Zhao, Zhangyang Wang, Yang You, and Kai Wang. Dd-ranking: Rethinking the evaluation of dataset distillation. GitHub repository, 2024. 1", + "[26] Tsung-Yi Lin, Priya Goyal, Ross Girshick, Kaiming He, and Piotr Dollar. Focal loss for dense object detection. In ICCV, 2017. 8", + "[27] Hanxiao Liu, Karen Simonyan, and Yiming Yang. Darts: Differentiable architecture search. arXiv preprint arXiv:1806.09055, 2018. 1", + "[28] Yanqing Liu, Jianyang Gu, Kai Wang, Zheng Zhu, Wei Jiang, and Yang You. Dream: Efficient dataset distillation by representative matching. In ICCV, 2023. 1, 2, 5, 6, 8", + "[29] Ziwei Liu, Zhongqi Miao, Xiaohang Zhan, Jiayun Wang, Boqing Gong, and Stella X Yu. Large-scale long-tailed recognition in an open world. In CVPR, 2019. 5", + "[30] Noel Loo, Ramin Hasani, Alexander Amini, and Daniela Rus. Efficient dataset distillation using random feature approximation. In NeurIPS, 2022. 1, 8", + "[31] Wojciech Masarczyk and Ivona Tautkute. Reducing catastrophic forgetting with learning on synthetic data. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition Workshops, 2020. 1" + ], + "bbox": [ + 517, + 93, + 903, + 898 + ], + "page_idx": 8 + }, + { + "type": "page_number", + "text": "30617", + "bbox": [ + 478, + 944, + 517, + 955 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[32] Timothy Nguyen, Zhourong Chen, and Jaehoon Lee. Dataset meta-learning from kernel ridge-regression. arXiv preprint arXiv:2011.00050, 2020. 1, 8", + "[33] Junran Peng, Xingyuan Bu, Ming Sun, Zhaoxiang Zhang, Tieniu Tan, and Junjie Yan. Large-scale object detection in the wild from imbalanced multi-labels. In CVPR, 2020. 8", + "[34] Jiawei Ren, Cunjun Yu, Xiao Ma, Haiyu Zhao, Shuai Yi, et al. Balanced meta-softmax for long-tailed visual recognition. In NeurIPS, 2020. 3, 8", + "[35] Noveen Sachdeva and Julian McAuley. Data distillation: A survey. arXiv preprint arXiv:2301.04272, 2023. 1", + "[36] Ozan Sener and Silvio Savarese. Active learning for convolutional neural networks: A core-set approach. arXiv preprint arXiv:1708.00489, 2017. 5, 6", + "[37] Shai Shalev-Shwartz, Yoram Singer, and Nathan Srebro. Pegasos: Primal estimated sub-gradient solver forsvm. In ICML, 2007. 1", + "[38] Yuzhang Shang, Zhihang Yuan, and Yan Yan. Mim4dd: Mutual information maximization for dataset distillation. In NeurIPS, 2024. 1, 8", + "[39] Shitong Shao, Zeyuan Yin, Muxin Zhou, Xindong Zhang, and Zhiqiang Shen. Generalized large-scale data condensation via various backbone and statistical matching. In CVPR, 2024. 6", + "[40] Duo Su, Junjie Hou, Weizhi Gao, Yingjie Tian, and Bowen Tang. D^4: Dataset distillation via disentangled diffusion model. In CVPR, 2024. 1, 8", + "[41] Kai Wang, Bo Zhao, Xiangyu Peng, Zheng Zhu, Shuo Yang, Shuo Wang, Guan Huang, Hakan Bilen, Xinchao Wang, and Yang You. Cafe: Learning to condense dataset by aligning features. In CVPR, 2022. 2", + "[42] Kai Wang, Zekai Li, Zhi-Qi Cheng, Samir Khaki, Ahmad Sajedi, Ramakrishna Vedantam, Konstantinos N Plataniotis, Alexander Hauptmann, and Yang You. Emphasizing discriminative features for dataset distillation in complex scenarios. arXiv preprint arXiv:2410.17193, 2024. 8", + "[43] Tongzhou Wang, Jun-Yan Zhu, Antonio Torralba, and Alexei A Efros. Dataset distillation. arXiv preprint arXiv:1811.10959, 2018. 8", + "[44] Zeyuan Yin, Eric Xing, and Zhiqiang Shen. Squeeze, recover and relabel: Dataset condensation at imagenet scale from a new perspective. In NeurIPS, 2023. 6", + "[45] Yifan Zhang, Bingyi Kang, Bryan Hooi, Shuicheng Yan, and Jiashi Feng. Deep long-tailed learning: A survey. IEEE transactions on pattern analysis and machine intelligence, 2023. 1", + "[46] Bo Zhao, Konda Reddy Mopuri, and Hakan Bilen. Dataset condensation with gradient matching. arXiv preprint arXiv:2006.05929, 2020. 1, 5, 6, 8", + "[47] Ganlong Zhao, Guanbin Li, Yipeng Qin, and Yizhou Yu. Improved distribution matching for dataset condensation. In CVPR, 2023. 1, 5, 8", + "[48] Zhenghao Zhao, Yuzhang Shang, Junyi Wu, and Yan Yan. Dataset quantization with active learning based adaptive sampling. In ECCV, 2024. 1", + "[49] Zhisheng Zhong, Jiequan Cui, Shu Liu, and Jiaya Jia. Improving calibration for long-tailed recognition. In CVPR, 2021. 2" + ], + "bbox": [ + 89, + 92, + 485, + 898 + ], + "page_idx": 9 + }, + { + "type": "ref_text", + "text": "[50] Yongchao Zhou, Ehsan Nezhadarya, and Jimmy Ba. Dataset distillation using neural feature regression. In NeurIPS, 2022. 1, 8", + "bbox": [ + 514, + 92, + 908, + 133 + ], + "page_idx": 9 + }, + { + "type": "page_number", + "text": "30618", + "bbox": [ + 478, + 944, + 519, + 957 + ], + "page_idx": 9 + } +] \ No newline at end of file diff --git a/2025/Distilling Long-tailed Datasets/f09fbc7b-8c7e-4bee-803c-2d67e3d05b6e_model.json b/2025/Distilling Long-tailed Datasets/f09fbc7b-8c7e-4bee-803c-2d67e3d05b6e_model.json new file mode 100644 index 0000000000000000000000000000000000000000..7fdafb95ddecab1201f8f4dffcb90957a9100b7b --- /dev/null +++ b/2025/Distilling Long-tailed Datasets/f09fbc7b-8c7e-4bee-803c-2d67e3d05b6e_model.json @@ -0,0 +1,2189 @@ +[ + [ + { + "type": "header", + "bbox": [ + 0.107, + 0.003, + 0.182, + 0.043 + ], + "angle": 0, + "content": "CVF" + }, + { + "type": "header", + "bbox": [ + 0.237, + 0.001, + 0.812, + 0.047 + ], + "angle": 0, + "content": "This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore." + }, + { + "type": "title", + "bbox": [ + 0.345, + 0.131, + 0.654, + 0.153 + ], + "angle": 0, + "content": "Distilling Long-tailed Datasets" + }, + { + "type": "text", + "bbox": [ + 0.101, + 0.179, + 0.905, + 0.234 + ], + "angle": 0, + "content": "Zhenghao Zhao\\(^{1*}\\) Haoxuan Wang\\(^{1*}\\) Yuzhang Shang\\(^{2}\\) Kai Wang\\(^{3}\\) Yan Yan\\(^{1\\dagger}\\) \n\\(^{1}\\)University of Illinois Chicago \\(^{2}\\)Illinois Institute of Technology \\(^{3}\\)National University of Singapore \n{zzhao72, hwang339, yyan55}@uic.edu yshang4@hawk.iit.edu kai.wang@comp.nus.edu.sg" + }, + { + "type": "title", + "bbox": [ + 0.248, + 0.269, + 0.327, + 0.285 + ], + "angle": 0, + "content": "Abstract" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.301, + 0.486, + 0.649 + ], + "angle": 0, + "content": "Dataset distillation aims to synthesize a small, information-rich dataset from a large one for efficient model training. However, existing dataset distillation methods struggle with long-tailed datasets, which are prevalent in real-world scenarios. By investigating the reasons behind this unexpected result, we identified two main causes: 1) The distillation process on imbalanced datasets develops biased gradients, leading to the synthesis of similarly imbalanced distilled datasets. 2) The experts trained on such datasets perform suboptimally on tail classes, resulting in misguided distillation supervision and poor-quality soft-label initialization. To address these issues, we first propose Distribution-agnostic Matching to avoid directly matching the biased expert trajectories. It reduces the distance between the student and the biased expert trajectories and prevents the tail class bias from being distilled to the synthetic dataset. Moreover, we improve the distillation guidance with Expert Decoupling, which jointly matches the decoupled backbone and classifier to improve the tail class performance and initialize reliable soft labels. This work pioneers the field of long-tailed dataset distillation, marking the first effective effort to distill long-tailed datasets. Our code will be made public at https://github.com/ichbill/LTDD." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.676, + 0.222, + 0.691 + ], + "angle": 0, + "content": "1. Introduction" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.701, + 0.484, + 0.868 + ], + "angle": 0, + "content": "Dataset distillation (DD) [4, 13, 28] aims to synthesize small and high-quality data summaries which contain the most important knowledge from a given target dataset [35]. Such a compact data summary offers numerous advantages, including faster model training, reduced memory consumption, and greater flexibility for various tasks [48], such as continual learning [31], neural architecture search [27], and knowledge distillation [15]. Various types of dataset distillation methods have been developed, including performance matching methods [30, 32, 50], parameter matching methods [13, 28, 46], representation matching methods [38, 47]" + }, + { + "type": "image", + "bbox": [ + 0.517, + 0.27, + 0.898, + 0.445 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.512, + 0.446, + 0.907, + 0.515 + ], + "angle": 0, + "content": "Figure 1. Performance comparison on CIFAR-10-LT. Existing Dataset Distillation methods exhibit degraded performance when applied to imbalanced datasets, especially when the imbalance factor increases, whereas our method provides significantly better performance under different imbalanced scenarios." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.521, + 0.906, + 0.566 + ], + "angle": 0, + "content": "and generative-based methods [12, 40]. Currently, parameter matching methods have demonstrated significant performance improvements across various datasets." + }, + { + "type": "text", + "bbox": [ + 0.511, + 0.568, + 0.907, + 0.793 + ], + "angle": 0, + "content": "Research on dataset distillation has predominantly focused on uniformly distributed datasets, such as CIFAR10, CIFAR100 [22], and TinyImageNet [24]. However, in real-world applications such as medical image diagnosis [19], training samples typically exhibit a long-tailed class distribution [45]. In these scenarios, a small number of classes (head classes) have a large number of samples, while the majority of classes (tail classes) have only a few samples. Motivated by this discrepancy and the desire for practical application, we are interested in the novel task of long-tailed dataset distillation (LTDD). LTDD aims to distill a long-tailed target dataset into a small and uniformly distributed synthetic dataset, where models trained on this distilled dataset can achieve satisfactory performance on the uniformly distributed test set." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.796, + 0.909, + 0.902 + ], + "angle": 0, + "content": "Nevertheless, distilling long-tailed datasets presents significant challenges, as the effectiveness of current dataset distillation methods diminishes when applied to such imbalanced distributions. As shown in Figure 1, we find that the performance of different dataset distillation methods consistently decreases as the degree of imbalance increases. The degree of imbalance is quantified by the imbalance fac" + }, + { + "type": "page_footnote", + "bbox": [ + 0.109, + 0.876, + 0.232, + 0.901 + ], + "angle": 0, + "content": "*equal contribution † corresponding author" + }, + { + "type": "page_number", + "bbox": [ + 0.48, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "30609" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.092, + 0.089, + 0.482, + 0.224 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.09, + 0.228, + 0.485, + 0.299 + ], + "angle": 0, + "content": "Figure 2. Relationship of classifier weights and class-wise accuracy. We reveal that classifiers generated by existing dataset distillation methods often exhibit imbalanced distributions, resulting in poor performance on tail classes. In contrast, our method produces balanced classifiers, thereby enhancing overall accuracy." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.304, + 0.484, + 0.472 + ], + "angle": 0, + "content": "tor, which is a commonly used metric to characterize the extent of imbalance [23]. The larger the imbalance factor is, the more imbalanced the dataset is. When the imbalance factor is large, popular dataset distillation methods may even perform worse than random selection (e.g. DATM can only achieve \\(40.1\\%\\) accuracy when the imbalance factor \\(\\beta = 200\\), while random selection can achieve \\(49.9\\%\\) accuracy). As shown in Figure 2, the performance degradation is concentrated in the tail classes. We attribute these performance failures to two key aspects of the learning procedure in long-tailed dataset distillation:" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.473, + 0.484, + 0.64 + ], + "angle": 0, + "content": "(1) During distillation, the weight distributions of the student and expert diverge, resulting in less informative tail class samples. Specifically, though the expert is trained on an imbalanced dataset, the student learns on a balanced one, resulting in differently distributed model weights. Such a discrepancy introduces controversy in the bi-level optimization objective, causing the distillation process to additionally focus on matching the data quantity distribution. Consequently, the tail classes of the distilled dataset contain significantly less useful information than the head classes, despite each class having the same number of samples." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.642, + 0.484, + 0.793 + ], + "angle": 0, + "content": "(2) The expert models yield suboptimal performance on the tail classes. Concretely, the expert is biased toward the head classes when trained on imbalanced datasets [1, 49], resulting in degraded tail class accuracy. Since the student's performance is upper-bounded by that of the expert, the expressiveness of the distilled dataset is consequently restricted. Additionally, the image labels are obtained via the expert's prediction logits [13]. As the expert is biased toward the head classes, the prediction logits for the tail classes are, therefore, less reliable and confident." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.796, + 0.484, + 0.903 + ], + "angle": 0, + "content": "To address the aforementioned issues, we propose a trajectory-matching approach named Distribution-agnostic Matching with Expert Decoupling. Firstly, the student model is trained via a Distribution-agnostic Matching (DAM) strategy, which consists of a loss designed for matching experts trained on long-tailed data. Specifically, when training the student model on the balanced synthetic" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.092, + 0.907, + 0.272 + ], + "angle": 0, + "content": "dataset, we mimic the procedure of training a model on an imbalanced dataset to reduce the distance between the student and the biased expert trajectories. Secondly, to mitigate the poor performance of the expert model and to enhance the overall distillation performance, an Expert Decoupling (ED) strategy is also proposed. In practice, we train representation experts and classification experts in a decoupled manner [20] to improve the performance of experts. We jointly match with the backbone layers of the representation experts and the classifier layers of the classification experts simultaneously. We use classification expert models to generate soft labels with high confidence for all classes." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.273, + 0.909, + 0.379 + ], + "angle": 0, + "content": "This work pioneers the field of long-tailed dataset distillation as the first successful effort in this area. Our experiments on four long-tailed datasets, conducted under various imbalance factors and images-per-class (IPC) settings, demonstrate the effectiveness of our method. We show that our method is able to achieve state-of-the-art performance across different dataset distillation baselines." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.391, + 0.803, + 0.409 + ], + "angle": 0, + "content": "2. Long-tailed Dataset Distillation" + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.416, + 0.719, + 0.431 + ], + "angle": 0, + "content": "2.1. Problem Formulation" + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.439, + 0.907, + 0.568 + ], + "angle": 0, + "content": "Assume we are given a long-tailed target training dataset \\(\\mathcal{D} = \\{(\\mathbf{x}_i, y_i)\\}_{i=1}^{|\\mathcal{D}|}\\) with \\(C\\) classes, where \\(y \\in \\{0, \\dots, C-1\\}\\). Denote the subset of images belonging to class \\(c\\) as \\(\\mathcal{D}_c\\), then \\(|\\mathcal{D}_0| > |\\mathcal{D}_1| > \\dots > |\\mathcal{D}_{C-1}|\\) and \\(|\\mathcal{D}_0| \\gg |\\mathcal{D}_{C-1}|\\). Denote the test dataset as \\(\\mathcal{T} = \\{(\\mathbf{x}_i, y_i)\\}_{i=1}^{|\\mathcal{T}|}\\) and is balanced. Our goal is to generate a small dataset \\(\\mathcal{S} = \\{(\\hat{\\mathbf{x}}_i, \\hat{y}_i)\\}_{i=1}^{|\\mathcal{S}|}\\) where \\(|\\mathcal{S}| \\ll |\\mathcal{D}|\\) and \\(|\\mathcal{S}_0| = |\\mathcal{S}_1| = \\dots = |\\mathcal{S}_{C-1}|\\), so that models trained on \\(\\mathcal{S}\\) perform optimally on the test set \\(\\mathcal{T}\\)." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.568, + 0.909, + 0.675 + ], + "angle": 0, + "content": "Typical dataset distillation methods [4, 6, 13, 28] use the following procedure to optimize \\(S\\): First, train an expert model \\(\\mathcal{M}_{\\mathcal{D}}\\) on \\(\\mathcal{D}\\) with acceptable performance. Then initialize \\(S\\) by randomly selecting images from \\(\\mathcal{D}\\) or with Gaussian noise, and predict the corresponding (soft) labels using \\(\\mathcal{M}_{\\mathcal{D}}\\). A student model \\(\\mathcal{M}_S\\) is then trained based on \\(S\\), and the distilled image set \\(S\\) is optimized by minimizing:" + }, + { + "type": "equation", + "bbox": [ + 0.58, + 0.687, + 0.907, + 0.705 + ], + "angle": 0, + "content": "\\[\n\\mathcal {L} (\\mathcal {S}) = \\left| \\left| \\mathcal {F} \\left(\\mathcal {M} _ {\\mathcal {D}}; \\mathcal {D}\\right) - \\mathcal {F} \\left(\\mathcal {M} _ {\\mathcal {S}}; \\mathcal {S}\\right) \\right| \\right| _ {p} \\tag {1}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.709, + 0.907, + 0.815 + ], + "angle": 0, + "content": "where \\(\\mathcal{F}\\) are model characteristics that can be used for matching, such as feature distributions [41], gradients [21], or training trajectories [4, 13]. Trajectory matching [4, 10, 13] matches the student model trained on the distilled dataset with the expert model trajectories trained on the original dataset, providing an impressive performance, and the matching loss is formulated as:" + }, + { + "type": "equation", + "bbox": [ + 0.557, + 0.823, + 0.907, + 0.862 + ], + "angle": 0, + "content": "\\[\n\\mathcal {L} _ {\\mathrm {m a t c h}} (\\hat {\\theta} _ {t + N}, \\theta_ {t + M} ^ {*}, \\theta_ {t} ^ {*}) = \\frac {\\| \\hat {\\theta} _ {t + N} - \\theta_ {t + M} ^ {*} \\| _ {2} ^ {2}}{\\| \\theta_ {t} ^ {*} - \\theta_ {t + M} ^ {*} \\| _ {2} ^ {2}}, \\quad (2)\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.868, + 0.909, + 0.903 + ], + "angle": 0, + "content": "where \\(\\theta_t^*\\) and \\(\\theta_{t + M}^*\\) are the expert parameters on epoch \\(t\\) and \\(t + M\\). \\(\\hat{\\theta}_{t + N}\\) is the student parameter obtained in the" + }, + { + "type": "page_number", + "bbox": [ + 0.48, + 0.945, + 0.521, + 0.958 + ], + "angle": 0, + "content": "30610" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.109, + 0.104, + 0.353, + 0.265 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.171, + 0.273, + 0.278, + 0.285 + ], + "angle": 0, + "content": "(a) expert trajectory" + }, + { + "type": "image", + "bbox": [ + 0.377, + 0.104, + 0.615, + 0.265 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.389, + 0.273, + 0.598, + 0.285 + ], + "angle": 0, + "content": "(b) training trajectory of traditional DD" + }, + { + "type": "image", + "bbox": [ + 0.646, + 0.104, + 0.885, + 0.265 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.665, + 0.273, + 0.862, + 0.285 + ], + "angle": 0, + "content": "(c) training trajectory of our method" + }, + { + "type": "image_caption", + "bbox": [ + 0.089, + 0.287, + 0.907, + 0.357 + ], + "angle": 0, + "content": "Figure 3. Effect of biased expert. (a) An expert trained on an imbalanced dataset leads to increasingly imbalanced weight gradients over classes. (b) Existing dataset distillation methods [4, 13] are ignorant of the distribution gap between \\( S_{t} \\) and \\( \\mathcal{D} \\). This causes the student gradient imbalance to increase in each step. If we match such trajectories, the synthetic dataset will be updated by the increasingly imbalanced gradients over classes, and the model trained on this synthetic dataset is highly biased. (c) With Distribution-agnostic Matching, the increasingly imbalanced gradients over classes will be re-weighted, such that the student model is updated with balanced gradients." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.359, + 0.461, + 0.373 + ], + "angle": 0, + "content": "inner optimization updated using the cross-entropy loss." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.373, + 0.484, + 0.48 + ], + "angle": 0, + "content": "In Section 2.2, we first discuss why such a formulation in Equation 1 is unreasonable for a long-tailed target dataset. Then we propose Distribution-agnostic Matching to mitigate the negative impact of the long-tailed distribution. To further enhance the distillation performance, an Expert Decoupling strategy is proposed in Section 2.3, aiding our method to achieve improved and even lossless performance." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.487, + 0.37, + 0.503 + ], + "angle": 0, + "content": "2.2. Distribution-agnostic Matching" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.51, + 0.484, + 0.556 + ], + "angle": 0, + "content": "Influence of long-tailed distribution. By extending Equation 1 to the whole distillation process, a single step of the synthetic dataset optimization can be formulated as:" + }, + { + "type": "equation", + "bbox": [ + 0.137, + 0.563, + 0.483, + 0.592 + ], + "angle": 0, + "content": "\\[\n\\mathcal {S} _ {t + 1} = \\underset {\\mathcal {S} _ {t}} {\\arg \\min } | | \\theta_ {\\mathcal {D}} - \\underset {\\theta_ {S} ^ {t}} {\\arg \\min } \\mathcal {L} \\left(\\theta_ {\\mathcal {S}} ^ {t}, \\mathcal {S} _ {t}\\right) | | _ {p}, \\tag {3}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.599, + 0.484, + 0.675 + ], + "angle": 0, + "content": "where \\(\\theta_S^t\\) is the student model parameter at epoch \\(t\\) and \\(\\mathcal{L}\\) is the cross-entropy loss based on the student model parameter and the synthetic dataset \\(S_{t}\\). \\(\\theta_{\\mathcal{D}}\\) is the expert model parameters trained on \\(\\mathcal{D}\\) and is determined before distillation. \\(p\\) is the factor for \\(L_{p}\\) norm and usually takes the value of 2." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.675, + 0.483, + 0.87 + ], + "angle": 0, + "content": "Equation 3 is an embedded optimization problem in which the inner loop is performed first. This implies that \\( S_{t} \\) is updated to train a \\( \\theta_{S}^{t} \\) that can match the fixed \\( \\theta_{\\mathcal{D}} \\) as much as possible. We see that \\( \\theta_{S}^{t} \\) is an important intermediate that connects the synthetic dataset and the target dataset. Under the setting that the target dataset \\( \\mathcal{D} \\) is imbalanced, the weights in \\( \\theta_{\\mathcal{D}} \\) for different classes follow an imbalanced distribution and are fixed during distillation. Then the outer loop in Equation 3 would force \\( \\theta_{S}^{t} \\) to produce an imbalanced distribution similar to \\( \\theta_{\\mathcal{D}} \\). As illustrated in Figure 3(a) and (b), the student is forced to mimic the expert during distillation, so that its weight distribution in different classes gradually becomes imbalanced." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.871, + 0.484, + 0.902 + ], + "angle": 0, + "content": "However, the outer loop in Equation 3 aims to find a valid \\(S_{t}\\) rather than a proper \\(\\theta_{\\mathcal{S}}^{t}\\). Thus, \\(\\mathcal{S}_t\\) is optimized to resem" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.358, + 0.907, + 0.463 + ], + "angle": 0, + "content": "ble the training behavior of \\(\\mathcal{D}\\). Although the tail class performance on the long-tailed dataset \\(\\mathcal{D}\\) is poor, this causes the tail classes of distilled \\(S_{t}\\) to similarly contain much less useful information than the head classes, even though it is uniformly distributed. As a result, Equation 3 proactively distills less information from tail classes, making the distillation process more biased towards head classes." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.464, + 0.909, + 0.63 + ], + "angle": 0, + "content": "Distribution-agnostic Matching. As discussed above, matching the student network with an expert trained on an imbalanced dataset could lead to severe flaws in the synthetic dataset \\( S \\), where the tail classes are unfortunately distorted. Ideally, we would like \\( S \\) to be a compact data summary such that every distilled image is meaningful and useful. However, such a dataset yields a student model with a balanced distribution (as shown in Figure 4 left), which mismatches the expert. This raises an interesting question: how can we preserve tail class information in the distilled dataset while avoiding the weight mismatching?" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.63, + 0.909, + 0.797 + ], + "angle": 0, + "content": "A key insight is to weaken the relationship between the data distribution and the weight distribution. In other words, we aim to prevent weight distribution mismatches from impacting dataset optimization. To achieve this, we propose Distribution-agnostic Matching, which seeks to align the gradient distributions of the student and expert without altering the distilled dataset. Practically, we modify the student training procedure to absorb the effect of weight imbalance by deliberately making \\(\\theta_S^t\\) imbalanced. Inspired by the Balanced Softmax loss [34], we propose a long-tailed distributed loss to aid student training:" + }, + { + "type": "equation", + "bbox": [ + 0.561, + 0.806, + 0.907, + 0.845 + ], + "angle": 0, + "content": "\\[\n\\mathcal {L} ^ {\\mathrm {c}} (\\mathbf {L}, \\mathbf {y}, \\mathbf {s}) = - \\frac {1}{n} \\sum_ {i = 1} ^ {n} g \\left(\\mathbf {s} _ {\\mathbf {y} _ {i}}\\right) \\cdot \\log P \\left(\\mathbf {L} _ {i, \\mathbf {y} _ {i}}\\right), \\tag {4}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.515, + 0.855, + 0.56, + 0.868 + ], + "angle": 0, + "content": "where" + }, + { + "type": "equation", + "bbox": [ + 0.596, + 0.865, + 0.907, + 0.906 + ], + "angle": 0, + "content": "\\[\nP (\\mathbf {L} _ {i, \\mathbf {y} _ {i}}) = \\frac {e ^ {\\mathbf {L} _ {i , \\mathbf {y} _ {i}} - \\lambda \\log (\\mathbf {s} _ {\\mathbf {y} _ {i}})}}{\\sum_ {j = 1} ^ {C} e ^ {\\mathbf {L} _ {i , j} - \\lambda \\log (\\mathbf {s} _ {j})}}. \\tag {5}\n\\]" + }, + { + "type": "page_number", + "bbox": [ + 0.48, + 0.945, + 0.518, + 0.957 + ], + "angle": 0, + "content": "30611" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.096, + 0.089, + 0.904, + 0.265 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.089, + 0.267, + 0.908, + 0.339 + ], + "angle": 0, + "content": "Figure 4. Comparison of the internal loop between traditional DD methods and ours. The expert model \\(\\mathcal{M}_{\\mathcal{D}}\\) is trained on a long-tailed dataset, leading to biased weights. Left: For traditional DD methods, directly matching with these weights causes large information loss in the tail classes. The backward propagation updates this imbalance from the expert trajectories to the synthetic dataset. Right: With Distribution-agnostic Matching, the gradients are obtained via \\(\\mathcal{L}^c\\), which revises the student weight for matching in the internal loop. This mitigates the distance between the student weight and the expert trajectory to reduce the influence of imbalance on the synthetic dataset." + }, + { + "type": "text", + "bbox": [ + 0.089, + 0.343, + 0.482, + 0.419 + ], + "angle": 0, + "content": "In the above equations, \\(\\mathbf{L}\\) is the student model logits, \\(n\\) is the number of samples in the distilled dataset, \\(\\mathbf{y}\\) is the ground-truth label, \\(\\mathbf{s}_{\\mathbf{y}_i}\\) is the number of samples in the long-tailed dataset for class \\(\\mathbf{y}_i\\), \\(g(\\cdot)\\) normalizes \\(\\mathbf{s}_{\\mathbf{y}_i}\\), and \\(\\lambda\\) is used for smoothing the predictions." + }, + { + "type": "text", + "bbox": [ + 0.089, + 0.421, + 0.483, + 0.587 + ], + "angle": 0, + "content": "The proposed loss is designed to support student model training during distillation by reducing the dependency of the weight distribution on the data distribution, thereby \"mimicking\" the training behavior on long-tailed datasets. Classes with more samples contribute more significantly to the loss, simulating a long-tailed training gradient for the student network during matching. This approach helps to minimize the trajectory differences between the student and expert networks caused by distribution mismatches between the target and distilled datasets. The right plot in Figure 4 illustrates the design and effect of our method." + }, + { + "type": "title", + "bbox": [ + 0.09, + 0.602, + 0.469, + 0.619 + ], + "angle": 0, + "content": "2.3. Improved Guidance with Expert Decoupling" + }, + { + "type": "text", + "bbox": [ + 0.089, + 0.627, + 0.483, + 0.779 + ], + "angle": 0, + "content": "Unlike experts trained on balanced target datasets, the experts trained on long-tailed datasets are less effective for guiding the dataset distillation process. Specifically, when adopting typical dataset distillation methods, the long-tailed experts perform poorly on tail classes, degrading student performance on these classes and leading to unreliable initial soft labels. This degraded student performance restricts the effective learning of synthetic tail-class images, while the unreliable labels introduce misleading information and provide weak supervision." + }, + { + "type": "text", + "bbox": [ + 0.089, + 0.78, + 0.483, + 0.902 + ], + "angle": 0, + "content": "In the conventional dataset distillation process, soft labels are assigned to initialized synthetic images based on the probability output of the expert model [13]. We found that soft-label predictions from long-tailed experts are distributed differently compared to those from balanced experts. As visualized in Figure 5, while the head class soft labels predicted by the balanced expert have similar confidence values with tail classes (0.88 and 0.89), the long-" + }, + { + "type": "image", + "bbox": [ + 0.515, + 0.342, + 0.907, + 0.502 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.513, + 0.505, + 0.907, + 0.575 + ], + "angle": 0, + "content": "Figure 5. Soft-label initialization on CIFAR-10-LT. We visualize the average predictive confidence of experts on different classes. For traditional DD, while the soft labels can be initialized well on head classes, they are predicted poorly on the tailed classes, leading to insufficient supervision." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.583, + 0.906, + 0.703 + ], + "angle": 0, + "content": "tailed expert predictions between head and tail classes differ drastically. Whereas the head class confidence value can be as large as 0.97, the confidence value of the tail class is surprisingly small (0.38). Such a confidence discrepancy leads to less effective learning in tail classes than in head classes. Moreover, for tail classes with a low number of samples, the number of samples with high confidence is reduced and leads to the diversity decrease of the dataset initialization." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.705, + 0.907, + 0.902 + ], + "angle": 0, + "content": "Decoupled training [20] is a widely used technique to alleviate the prediction bias introduced by imbalanced data. It consists of two stages: representation learning and classifier fine-tuning. At the representation learning stage, the entire model is trained on the original dataset \\(\\mathcal{D}\\). At the classifier fine-tuning stage, the backbone of the model is frozen, and only the classifier of the model is trained on a balanced dataset \\(\\mathcal{B}\\), which is created from \\(\\mathcal{D}\\) via oversampling or undersampling. However, the integration of this decoupling approach with dataset distillation is not trivial, as it requires careful alignment of the student model with the expert model. To effectively utilize the idea of Expert Decoupling, we propose a two-step strategy to improve teacher" + }, + { + "type": "page_number", + "bbox": [ + 0.48, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "30612" + } + ], + [ + { + "type": "table", + "bbox": [ + 0.093, + 0.089, + 0.905, + 0.264 + ], + "angle": 0, + "content": "
DatasetCIFAR-10-LT
Imbalance Factor1050100200
IPC102050102050102050102050
Random32.5±2.239.6±0.951.9±1.533.2±0.442.0±1.351.6±1.334.4±2.041.4±0.752.6±0.532.5±0.842.2±1.149.9±1.4
K-Center Greedy [36]21.9±0.824.2±0.831.7±0.917.8±0.220.8±0.526.1±0.216.2±0.519.0±1.024.2±1.216.8±0.317.5±1.422.6±1.6
Graph-Cut [17]28.7±0.934.2±1.040.6±1.024.2±0.728.6±0.833.9±0.422.9±0.926.0±0.533.3±1.022.3±0.925.2±0.529.2±0.4
DC [46]37.9±0.938.5±0.937.4±1.437.3±0.938.8±1.035.8±1.236.7±0.838.1±1.035.3±1.435.6±0.835.7±0.933.3±1.4
MTT [4]58.0±0.859.5±0.462.0±0.945.8±1.449.9±0.853.6±0.537.7±0.641.6±1.147.8±1.1N/A22.6±1.023.9±0.8
DREAM [28]34.6±0.642.2±1.550.5±0.730.8±0.638.4±0.345.5±0.930.8±1.734.9±0.842.2±0.832.7±1.332.4±0.338.9±0.4
IDM [47]54.8±0.457.1±0.360.1±0.351.9±0.753.3±0.656.1±0.449.8±0.650.9±0.553.1±0.447.0±0.548.1±0.549.9±0.3
Minimax [12]29.2±0.528.5±0.639.9±0.118.4±0.322.5±0.225.2±0.219.9±0.423.3±0.228.0±0.619.1±0.420.5±0.222.7±0.3
DATM [13]57.2±0.460.4±0.266.7±0.641.6±0.243.4±0.350.3±0.237.3±0.238.9±0.144.3±0.1N/A34.8±0.140.1±0.2
Ours58.1±0.363.0±1.070.5±0.454.2±1.059.4±0.765.8±0.253.4±0.158.2±0.664.0±0.952.2±0.656.6±0.462.3±0.3
Full Dataset76.7±0.369.8±0.366.2±0.462.1±0.4
" + }, + { + "type": "table_caption", + "bbox": [ + 0.09, + 0.265, + 0.907, + 0.307 + ], + "angle": 0, + "content": "Table 1. Quantitative comparisons with the SOTA methods on CIFAR-10-LT. Our method outperforms all existing approaches, and can even achieve lossless performance under certain settings. Performance is evaluated under various imbalance factors. As the first to explore long-tailed dataset distillation, all experiments are conducted using open-source code. N/A indicates distillation failure." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.312, + 0.288, + 0.326 + ], + "angle": 0, + "content": "guidance for student training." + }, + { + "type": "text", + "bbox": [ + 0.089, + 0.326, + 0.483, + 0.476 + ], + "angle": 0, + "content": "Firstly, we train two types of experts in a decoupled manner. The representation expert with parameter \\(\\theta^{\\mathcal{D}}\\) is obtained from the representation learning stage, and the classifier expert with parameter \\(\\theta^{\\mathcal{B}}\\) is obtained from the classifier finetuning stage. Since the classifier experts are not biased and are able to provide more accurate classification, we use the classifier experts \\(\\theta^{\\mathcal{B}}\\) for the synthetic dataset \\(S\\) initialization. Specifically, we use the logits \\(L_{i} = \\theta^{\\mathcal{B}}\\mathbf{x}_{i}\\) to initialize soft labels, with \\(\\mathrm{softmax}(L_i)\\) as the label of \\(\\mathbf{x}_i\\), and then sample from \\(\\mathcal{D}\\) to initialize \\(S\\)." + }, + { + "type": "text", + "bbox": [ + 0.089, + 0.477, + 0.483, + 0.658 + ], + "angle": 0, + "content": "Secondly, we match the teacher expert's training trajectory with the student's in an adaptive manner. Since the representation expert learns the image features from the original dataset \\(\\mathcal{D}\\) but fails to classify with biased classifiers [20], we skip matching the classifier layer when matching the student network trained on the synthetic dataset with the representation expert. On the contrary, the classifier expert \\(\\mathcal{M}_{\\mathcal{B}}\\) only trains the classifier layer, so we also only train the student classifier layer and only match the classifier layer during the classifier matching. During the whole matching process, we jointly match the representation trajectory and the classifier trajectory using the following loss:" + }, + { + "type": "equation", + "bbox": [ + 0.164, + 0.667, + 0.483, + 0.707 + ], + "angle": 0, + "content": "\\[\n\\begin{array}{l} \\mathcal {L} = \\lambda_ {\\text {r e p}} \\mathcal {L} _ {\\text {m a t c h}} \\left(\\hat {\\theta} _ {t + N}, \\theta_ {t + M} ^ {\\mathcal {D} _ {*}}, \\theta_ {t} ^ {\\mathcal {D} _ {*}}\\right) \\\\ + \\lambda_ {\\mathrm {c l s}} \\mathcal {L} _ {\\text {m a t c h}} \\left(\\hat {\\theta} _ {t + N}, \\theta_ {t + M} ^ {\\mathcal {B} *}, \\theta_ {t} ^ {\\mathcal {B} *}\\right), \\tag {6} \\\\ \\end{array}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.715, + 0.483, + 0.811 + ], + "angle": 0, + "content": "where \\(\\lambda_{\\mathrm{rep}}\\) and \\(\\lambda_{\\mathrm{cls}}\\) are hyper-parameters that control the weighting ratio, \\(\\hat{\\theta}_{t + N}\\) is the student model at its training epoch \\(t + N\\), \\(\\theta_t^{\\mathcal{D}^*}\\) and \\(\\theta_t^{\\mathcal{B}^*}\\) are the representation backbone and classifier of the expert at epoch \\(t\\), respectively. \\(\\hat{\\theta}_{t + N}\\) is updated by our proposed long-tailed distributed loss \\(\\mathcal{L}^{\\mathrm{c}}\\) in Equation 4, and \\(\\mathcal{L}_{\\mathrm{match}}\\) follows the form in Equation 2." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.822, + 0.224, + 0.839 + ], + "angle": 0, + "content": "3. Experiments" + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.848, + 0.288, + 0.864 + ], + "angle": 0, + "content": "3.1. Experiment Datasets" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.871, + 0.483, + 0.9 + ], + "angle": 0, + "content": "We evaluate our methods on four widely used long-tailed datasets of different scales: CIFAR-10-LT, CIFAR-100-LT" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.311, + 0.907, + 0.492 + ], + "angle": 0, + "content": "[7], TinyImageNet-LT, and ImageNet-LT [29]. The experiments are conducted under various degrees of imbalance and IPC settings. Following [7, 9, 29], these long-tailed datasets are sampled from the original balanced CIFAR [22] and ImageNet [8] datasets. The sampling procedure is done by randomly selecting different amounts of images from each class, where the image number \\( |\\mathcal{D}_c| \\) for class \\( c \\) is determined by an exponential decay function \\( |\\hat{\\mathcal{D}}_c| = |\\mathcal{D}_c|\\mu^c \\), and \\( \\mu^c = \\beta^{-(c / C)} \\), where \\( C \\) is the number of classes and \\( \\beta \\) is the imbalanced factor. The degree of imbalance is measured via the imbalanced factor \\( \\beta = \\mathcal{D}_0 / \\mathcal{D}_C \\) [7]. The larger \\( \\beta \\) is, the more imbalanced the dataset is." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.5, + 0.653, + 0.515 + ], + "angle": 0, + "content": "3.2. Main Results" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.523, + 0.907, + 0.689 + ], + "angle": 0, + "content": "CIFAR-10-LT. The experiment results on CIFAR-10-LT are presented in Table 1, showing that our proposed method consistently outperforms all other dataset distillation baselines. Table 1 further highlights the advantages of our approach on long-tailed datasets. When the imbalance factor is low, prior dataset distillation methods like DATM perform reasonably well. For instance, DATM only experiences a \\(10.0\\%\\) drop in accuracy when \\(\\beta = 10\\) and IPC=50, compared to the accuracy achieved with the full dataset. In these scenarios, our method surpasses the best baseline, DATM, by \\(0.9\\%\\) on \\(\\beta = 10\\) and \\(3.8\\%\\) on \\(\\beta = 50\\)." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.689, + 0.907, + 0.84 + ], + "angle": 0, + "content": "When the imbalance factor is high (e.g., \\(\\beta = 200\\)), IDM [47] achieves performance comparable to random selection. Other methods, such as MTT [4], DREAM [28], DATM [13], and Minimax [12], struggle with higher imbalance factors, performing worse than random selection. In contrast, our method sustains strong performance on highly imbalanced datasets, surpassing the best existing method IDM by a substantial margin of \\(10.6\\%\\) on \\(\\beta = 200\\). Remarkably, we achieve lossless performance on \\(\\beta = 200\\) using only 50 images per class." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.84, + 0.906, + 0.901 + ], + "angle": 0, + "content": "CIFAR-100-LT. The results are presented in Table 2. Our method consistently outperforms all existing dataset distillation methods across all imbalance factors on CIFAR-100-LT. Similar to CIFAR-10-LT, our approach demonstrates" + }, + { + "type": "page_number", + "bbox": [ + 0.48, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "30613" + } + ], + [ + { + "type": "table", + "bbox": [ + 0.094, + 0.089, + 0.905, + 0.242 + ], + "angle": 0, + "content": "
DatasetCIFAR-100-LTTiny-ImageNet-LT
Imbalance Factor10201020
IPC102050102050102050102050
Random14.2±0.621.7±0.632.1±0.615.0±0.321.6±0.530.5±0.57.4±0.213.5±0.420.6±0.47.6±0.113.2±0.419.8±0.1
K-Center Greedy [36]10.7±0.915.9±1.024.8±0.210.0±0.515.1±0.623.8±0.310.4±1.511.3±1.27.7±1.012.6±1.49.9±2.213.2±2.7
Graph-Cut [17]16.9±0.322.2±0.429.9±0.416.0±0.520.7±0.528.7±0.39.8±0.75.6±0.610.9±1.13.4±0.85.0±1.04.2±1.1
DC [46]24.0±0.327.4±0.327.4±0.323.2±0.326.2±0.327.4±0.3------
MTT [4]14.3±0.116.7±0.213.8±0.212.6±0.315.0±0.210.6±0.511.1±0.218.1±0.223.1±0.17.7±0.114.7±0.215.6±0.3
DREAM [28]10.1±0.412.0±1.013.1±0.49.4±0.410.3±0.612.3±0.35.4±0.36.8±0.17.8±0.24.8±0.16.0±0.27.4±0.1
DATM [13]28.2±0.434.1±0.231.6±0.125.3±0.327.2±0.127.1±0.121.3±0.114.5±0.126.8±0.114.0±0.619.0±0.323.1±0.1
Ours31.5±0.237.5±0.440.0±0.131.4±0.535.1±0.437.0±0.726.0±0.327.9±0.230.3±0.223.6±0.325.5±0.328.0±0.6
Full Dataset40.3±0.436.5±0.431.2±1.028.3±0.3
" + }, + { + "type": "table_caption", + "bbox": [ + 0.09, + 0.244, + 0.905, + 0.273 + ], + "angle": 0, + "content": "Table 2. Quantitative comparisons with the SOTA methods on CIFAR-100-LT and Tiny-ImageNet-LT. When scaling up to larger datasets, our approach still outperforms all the included baselines, and achieves nearly lossless performance under several settings." + }, + { + "type": "table", + "bbox": [ + 0.095, + 0.278, + 0.48, + 0.428 + ], + "angle": 0, + "content": "
DatasetImageNet-LT
Imbalance Factor510
IPC10201020
Random3.9±0.17.0±0.13.9±0.16.8±0.1
G-VBSM [39]-1.0±0.1-1.0±0.1
TESLA [6]3.0±0.1-2.7±0.1-
DATM [13]7.4±0.18.1±0.27.9±0.18.2±0.1
SRe2L [44]6.7±0.110.1±0.17.7±0.310.9±1.0
Ours20.8±0.221.0±0.120.3±0.120.7±0.1
Full Dataset27.6±0.326.6±0.2
" + }, + { + "type": "table_caption", + "bbox": [ + 0.09, + 0.43, + 0.483, + 0.457 + ], + "angle": 0, + "content": "Table 3. Quantitative comparisons on ImageNet-LT. The missing results are due to out-of-memory." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.464, + 0.461, + 0.478 + ], + "angle": 0, + "content": "greater improvements as the imbalance factor increases." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.479, + 0.483, + 0.569 + ], + "angle": 0, + "content": "Tiny-ImageNet-LT. We conducted experiments on TinyImageNet-LT to evaluate the scalability of our method in comparison with other baselines. The results, shown in Table 2, align with our findings on CIFAR-10-LT and CIFAR100-LT, where our method surpasses baseline methods by a growing margin as the imbalance factor increases." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.569, + 0.483, + 0.646 + ], + "angle": 0, + "content": "ImageNet-LT. We further scale up to ImageNet-LT, with results presented in Table 3, where our model consistently outperforms the baselines. Note that as the dataset size increases, many dataset distillation methods become resource-intensive and are impractical to apply." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.646, + 0.483, + 0.902 + ], + "angle": 0, + "content": "Cross-architecture performance. Table 4 presents our cross-architecture experiments, demonstrating the generalizability of our synthetic dataset. We evaluated various methods on the CIFAR-10-LT dataset with an imbalance factor of \\(\\beta = 100\\) and IPC=50, comparing against random selection, MTT, and DATM. As shown in the table, our distilled dataset consistently outperforms other methods by a significant margin. Specifically, our method achieves a \\(12.7\\%\\) accuracy improvement on ConvNet, a \\(17.1\\%\\) improvement on ResNet-18, a \\(7.7\\%\\) improvement on VGG-11, and a \\(6\\%\\) improvement on AlexNet. The results also indicate that the imbalance distilled by ConvNet in traditional DD methods negatively impacts the performance of other models. Similar to experiments on ConvNets, MTT and DATM perform poorly on unseen models due to the imbalanced dataset and the biased expert trained on it, often yielding worse results than random selection." + }, + { + "type": "table", + "bbox": [ + 0.518, + 0.279, + 0.902, + 0.36 + ], + "angle": 0, + "content": "
MethodConvNet-3ResNet-18VGG-11AlexNet
Random52.049.247.847.2
MTT [4]49.142.942.340.0
DATM [13]44.442.442.244.9
Ours64.760.055.553.2
" + }, + { + "type": "table_caption", + "bbox": [ + 0.532, + 0.361, + 0.888, + 0.375 + ], + "angle": 0, + "content": "Table 4. Cross-architecture evaluation on CIFAR-10-LT." + }, + { + "type": "title", + "bbox": [ + 0.514, + 0.386, + 0.679, + 0.401 + ], + "angle": 0, + "content": "3.3. Ablation Studies" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.411, + 0.907, + 0.548 + ], + "angle": 0, + "content": "Ablation on different components. Figure 6a demonstrates the effectiveness of the two proposed components of our method. These experiments were performed on the CIFAR-10-LT dataset with an imbalance factor of \\(\\beta = 200\\) and IPC=50. The \"base\" method is a trajectory matching method without Distribution-agnostic Matching and Expert Decoupling. \"+\"DAM\" refers to our method without Expert Decoupling, while \"ours\" represents the complete proposed method, with both DAM and ED applied." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.549, + 0.906, + 0.655 + ], + "angle": 0, + "content": "The baseline, which suffers from weight inconsistency and poor expert performance as mentioned in Section 1, only achieves \\(43.8\\%\\) accuracy. After incorporating Distribution-agnostic Matching, the inconsistency between student and expert weights is mitigated, resulting in a \\(14.3\\%\\) performance improvement. Further integrating Expert Decoupling boosts the performance to \\(62.6\\%\\)." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.657, + 0.906, + 0.853 + ], + "angle": 0, + "content": "Effect of classifier layer matching. As discussed in Section 2.3, during representation learning, the model is influenced by the long-tailed distribution, resulting in a biased classifier. During classifier fine-tuning, the classifier is finetuned on balanced data with the backbone frozen. Therefore, it is crucial to explore strategies for matching each expert's different modules during dataset distillation. The experiments are presented in Figure 6b. \"B\", \"C\", and \"W\" refer to matching with the representation backbone, the classifier layer, and the whole model, respectively. \"B/C\" refers to matching the student backbone with the representation expert's backbone, and the student classifier layer with the classification expert's classifier." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.856, + 0.906, + 0.901 + ], + "angle": 0, + "content": "As shown in Figure 6b, the best performance is achieved by matching the backbone with representation experts and the classifier layer with classification experts. This outcome" + }, + { + "type": "page_number", + "bbox": [ + 0.48, + 0.945, + 0.52, + 0.957 + ], + "angle": 0, + "content": "30614" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.092, + 0.088, + 0.241, + 0.2 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.105, + 0.201, + 0.237, + 0.225 + ], + "angle": 0, + "content": "(a) Effect of the different components" + }, + { + "type": "image", + "bbox": [ + 0.253, + 0.089, + 0.408, + 0.2 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.26, + 0.201, + 0.406, + 0.225 + ], + "angle": 0, + "content": "(b) Comparison of different matching strategies" + }, + { + "type": "image", + "bbox": [ + 0.417, + 0.091, + 0.571, + 0.2 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.42, + 0.201, + 0.57, + 0.225 + ], + "angle": 0, + "content": "(c) Comparison of matching strategies on low \\(\\beta\\)" + }, + { + "type": "image", + "bbox": [ + 0.579, + 0.09, + 0.732, + 0.2 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.582, + 0.201, + 0.731, + 0.225 + ], + "angle": 0, + "content": "(d) Comparison of matching strategies on high \\(\\beta\\)" + }, + { + "type": "image", + "bbox": [ + 0.74, + 0.091, + 0.893, + 0.2 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.741, + 0.201, + 0.897, + 0.226 + ], + "angle": 0, + "content": "(e) Effect of the relative scale of \\(\\lambda_{\\mathrm{rep}}\\) and \\(\\lambda_{\\mathrm{cls}}\\)" + }, + { + "type": "image_caption", + "bbox": [ + 0.09, + 0.228, + 0.907, + 0.27 + ], + "angle": 0, + "content": "Figure 6. (a) Ablation study on the proposed components, Distribution-agnostic Matching and Expert Decoupling. (b) Ablation of different matching strategies with expert model layers. (c) & (d) Comparisons of matching strategies for long-tailed dataset distillation under low imbalance factor (c) and high imbalance factor (d). (e) Ablation study on hyperparameters \\(\\lambda_{\\mathrm{rep}}\\) and \\(\\lambda_{\\mathrm{cls}}\\)." + }, + { + "type": "text", + "bbox": [ + 0.089, + 0.279, + 0.483, + 0.369 + ], + "angle": 0, + "content": "aligns with the principles of decoupled learning. Since the classifier layer of representation experts is heavily affected by the long-tailed dataset, including it in the matching process causes a performance drop. In addition, matching the entire classifier expert negatively impacts performance, as only the classifier layer is trained at this stage." + }, + { + "type": "text", + "bbox": [ + 0.089, + 0.37, + 0.483, + 0.567 + ], + "angle": 0, + "content": "Effect of different matching strategies. When distilling a long-tailed dataset, various intuitive strategies can be employed with decoupling trained experts. We explored several strategies, including sequential matching, classification expert matching, oversampled dataset, and joint matching. Sequential matching first matches representation experts, followed by classification experts. Classification expert matching only matches with the classification expert. The oversampled dataset approach involves oversampling the tail classes before distilling the dataset rather than using the decoupling strategy. Joint matching, used in our Expert Decoupling, involves matching both experts simultaneously in the outer loop of the matching." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.568, + 0.483, + 0.794 + ], + "angle": 0, + "content": "As shown in Figures 6c and 6d, joint matching (joint), which we use in Expert Decoupling, achieves the best overall performance. Sequential matching (sequential) and directly matching with classification experts (direct) perform poorly when the imbalance factor is low, meaning the data is \"not very imbalanced.\" Sequential matching suffers due to the trajectory gap between the two types of experts. Classification expert matching is less effective when the imbalance factor is low because the representation layers learn more information in such scenarios, and solely matching the classification layers fails to capture sufficient information. As depicted in Figure 6d, the oversampled dataset strategy (oversample) performs poorly when the imbalance factor is high. This finding aligns with previous work [7], which indicates that oversampling leads to model overfitting." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.796, + 0.483, + 0.902 + ], + "angle": 0, + "content": "Effect of hyperparameters. The most important hyperparameters in our method are \\(\\lambda_{\\mathrm{rep}}\\) and \\(\\lambda_{\\mathrm{cls}}\\) in Equation 6. Together, they regulate the balance between the matching representation expert and the classification expert. The setting of \\(\\lambda_{\\mathrm{rep}}\\) and \\(\\lambda_{\\mathrm{cls}}\\) depends on the imbalance factors. From Figure 6e, we have the following observations. For a \"more imbalanced\" dataset (e.g., \\(\\beta = 200\\)), lower \\(\\lambda_{\\mathrm{rep}}\\) and higher \\(\\lambda_{\\mathrm{cls}}\\)" + }, + { + "type": "table", + "bbox": [ + 0.52, + 0.27, + 0.9, + 0.335 + ], + "angle": 0, + "content": "
Hyperparametersβ = 10Hyperparametersβ = 200
λrep = 1, λcls = 0.270.5λrep = 0.2, λcls = 162.4
λrep = 1, λcls = 0.468.9λrep = 0.4, λcls = 161.7
λrep = 1, λcls = 0.668.0λrep = 0.6, λcls = 161.3
" + }, + { + "type": "table_caption", + "bbox": [ + 0.517, + 0.336, + 0.902, + 0.35 + ], + "angle": 0, + "content": "Table 5. Effect of \\( {\\lambda }_{\\text{rep }} \\) and \\( {\\lambda }_{\\text{cls }} \\) under different imbalance factors." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.364, + 0.905, + 0.47 + ], + "angle": 0, + "content": "are preferred. For \"less imbalanced\" datasets (e.g., \\(\\beta = 10\\)), higher \\(\\lambda_{\\mathrm{rep}}\\) and lower \\(\\lambda_{\\mathrm{cls}}\\) are chosen. This observation is consistent with our experiments in Figure 6c and 6d, where only matching classifier experts perform well when \\(\\beta\\) is high. This is because the representation layers learn more information when \\(\\beta\\) is low, and increasing \\(\\lambda_{\\mathrm{rep}}\\) leverages information from the representation layers." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.471, + 0.906, + 0.562 + ], + "angle": 0, + "content": "We further analyze the sensitivity of the hyperparameters in Table 5. We can observe a minor change in the performance as we change hyperparameters rapidly. When the relative magnitude of the two hyperparameters is preserved, their absolute values do not affect the distillation performance much, suggesting the robustness of our method." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.574, + 0.862, + 0.59 + ], + "angle": 0, + "content": "3.4. Long-tailed Dataset Distillation Analysis" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.597, + 0.906, + 0.734 + ], + "angle": 0, + "content": "Effect of imbalance in long-tailed dataset distillation. In this part, we empirically show that the performance drop in long-tailed dataset distillation is due to the data imbalance instead of the total sample number reduction of the target dataset. We compare the performance of using regular dataset distillation methods for CIFAR-10-LT (with \\(\\beta = 200\\) and \\(\\mathrm{IPC} = 50\\)) and a subset of CIFAR-10 that contains the same amount of data but in a uniform distribution. The total number of samples is 11203 under this setting." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.735, + 0.906, + 0.854 + ], + "angle": 0, + "content": "As we can observe in Figure 7a, the performance gap between distilling a long-tailed dataset and a balanced subset can be as large as \\(28.6\\%\\), even though the two target datasets have the same total amount of samples. Thus, we conclude that it is the imbalanced distribution in the target dataset that prevents effective distillation. Based on this observation, we delved into it and proposed methods to alleviate the imbalanced distribution's negative effects." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.856, + 0.906, + 0.901 + ], + "angle": 0, + "content": "Handling classes with extremely small number of samples. In long-tailed dataset distillation, a scenario exists where the number of samples in tail classes is even less than" + }, + { + "type": "page_number", + "bbox": [ + 0.48, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "30615" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.092, + 0.087, + 0.287, + 0.242 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.296, + 0.087, + 0.49, + 0.242 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.497, + 0.087, + 0.691, + 0.242 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.699, + 0.087, + 0.895, + 0.242 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.089, + 0.248, + 0.908, + 0.292 + ], + "angle": 0, + "content": "Figure 7. (a) Comparison of DD performance between a uniform dataset and a long-tailed dataset with the same number of samples. (b) Comparison of strategies to handle an extremely low number of samples. (c) Comparison of class-wise accuracy between DATM and our method. (d) Comparison between balanced IPC and long-tailed distributed IPC." + }, + { + "type": "text", + "bbox": [ + 0.089, + 0.299, + 0.482, + 0.45 + ], + "angle": 0, + "content": "the required number of images per class. In this case, we employ and compare three strategies: 1) Inherent strategy, which keeps the number of samples the same in the synthetic dataset; 2) Oversampling, which duplicates the samples of the tail classes before distillation; 3) Noise generation, which randomly generates Gaussian noises to supplement the insufficient samples for tail classes. From figure 7b, we see that the inherent strategy provides a comparatively lower accuracy, while oversampling and noise generation achieve better and similar performance." + }, + { + "type": "text", + "bbox": [ + 0.089, + 0.452, + 0.483, + 0.572 + ], + "angle": 0, + "content": "Class-wise accuracy for long-tailed dataset. We also provide class-wise accuracy to illustrate the impact of the imbalanced data, and we show that our method can effectively improve the accuracy of tail classes. As shown in Figure 7c, DATM performs poorly on tail classes, and even achieves near-zero performance on the last class. On the contrary, our method can successfully boost the tail class accuracy while reserving performance on head classes." + }, + { + "type": "text", + "bbox": [ + 0.089, + 0.574, + 0.483, + 0.772 + ], + "angle": 0, + "content": "Choice of distilled dataset distribution. Should we distill a long-tailed dataset into another long-tailed dataset or a balanced one? We compared the performance of distilling the target dataset into either a similarly distributed long-tailed synthetic dataset or a balanced synthetic dataset using our method and DREAM [28]. For the long-tailed synthetic dataset, we maintained the same total number of samples as in the synthetic dataset with IPC=50. As shown in Figure 7d, both methods demonstrate that distilling the dataset into a balanced synthetic dataset provides better performance than distilling it into a long-tailed one. This result is reasonable since the model trained on the synthetic long-tailed dataset will be affected by the long-tailed distribution." + }, + { + "type": "title", + "bbox": [ + 0.09, + 0.789, + 0.233, + 0.805 + ], + "angle": 0, + "content": "4. Related Work" + }, + { + "type": "title", + "bbox": [ + 0.09, + 0.816, + 0.275, + 0.831 + ], + "angle": 0, + "content": "4.1. Dataset Distillation" + }, + { + "type": "text", + "bbox": [ + 0.089, + 0.84, + 0.483, + 0.902 + ], + "angle": 0, + "content": "Dataset distillation [43] was first formulated by Wang et al. The goal of dataset distillation is to synthesize a small dataset, such that models trained on it have good performance on the original dataset. The following works can be" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.299, + 0.907, + 0.421 + ], + "angle": 0, + "content": "divided into kernel ridge regression-based methods [30, 32, 50], gradient matching methods [6, 13, 28, 42, 46], representation matching methods [38, 47], and generative-based methods [12, 40]. Despite the advantages of current methods, they only focus on uniformly distributed datasets, ignoring the prevalent usage of long-tailed datasets in practice. Therefore, we propose to address the long-tailed dataset distillation problem in this paper." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.426, + 0.736, + 0.442 + ], + "angle": 0, + "content": "4.2. Long-tailed Recognition" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.446, + 0.909, + 0.644 + ], + "angle": 0, + "content": "Under real-world settings, data often tends to follow a long-tailed distribution [2]. Models trained on such datasets usually exhibit good performance on head classes but low accuracy on tail classes. A straightforward way is to preprocess the data, using oversampling [14, 18, 33], undersampling [3, 14], or data augmentation [5, 11] to make it balancedly distributed. An alternative way is weight balancing [2, 26, 34], which re-weights the layers during training to force the parameters to update unbiasedly. Decoupling methods [20] first train the model to obtain the features of the data and then fine-tune the model on the class-balanced data. In our methods, we use the decoupling method to train the expert networks during the distillation." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.653, + 0.634, + 0.669 + ], + "angle": 0, + "content": "5. Conclusion" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.675, + 0.907, + 0.84 + ], + "angle": 0, + "content": "We introduce long-tailed dataset distillation, a novel yet challenging task. We find that existing DD methods fail when applied to long-tailed datasets and encounter two main problems: weight distribution mismatch between the student and the expert; and suboptimal performance of the expert on the tail classes. To mitigate these issues, we propose two strategies: Distribution-agnostic Matching and Expert Decoupling. Experimental results over four long-tailed datasets demonstrate the effectiveness of our proposed method. As the pioneering work in this research area, we are the first to effectively distill long-tailed datasets." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.84, + 0.906, + 0.901 + ], + "angle": 0, + "content": "Acknowledgments: This research is supported by NSF IIS-2525840, ECCS-2123521 and Cisco Research unrestricted gift. This article solely reflects the opinions and conclusions of its authors and not the funding agencies." + }, + { + "type": "page_number", + "bbox": [ + 0.48, + 0.945, + 0.521, + 0.958 + ], + "angle": 0, + "content": "30616" + } + ], + [ + { + "type": "title", + "bbox": [ + 0.093, + 0.09, + 0.188, + 0.106 + ], + "angle": 0, + "content": "References" + }, + { + "type": "ref_text", + "bbox": [ + 0.101, + 0.116, + 0.482, + 0.155 + ], + "angle": 0, + "content": "[1] Shaden Alshammari, Yuxiong Wang, Deva Ramanan, and Shu Kong. Long-tailed recognition via weight balancing. In CVPR, 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.101, + 0.159, + 0.482, + 0.199 + ], + "angle": 0, + "content": "[2] Shaden Alshammari, Yu-Xiong Wang, Deva Ramanan, and Shu Kong. Long-tailed recognition via weight balancing. In CVPR, 2022. 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.202, + 0.482, + 0.242 + ], + "angle": 0, + "content": "[3] Mateusz Buda, Atsuto Maki, and Maciej A Mazurowski. A systematic study of the class imbalance problem in convolutional neural networks. Neural networks, 2018. 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.246, + 0.482, + 0.285 + ], + "angle": 0, + "content": "[4] George Cazenavette, Tongzhou Wang, Antonio Torralba, Alexei A Efros, and Jun-Yan Zhu. Dataset distillation by matching training trajectories. In CVPR, 2022. 1, 2, 3, 5, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.288, + 0.482, + 0.343 + ], + "angle": 0, + "content": "[5] Hsin-Ping Chou, Shih-Chieh Chang, Jia-Yu Pan, Wei Wei, and Da-Cheng Juan. Remix: rebalanced mixup. In Computer Vision-ECCV 2020 Workshops: Glasgow, UK, August 23-28, 2020, Proceedings, Part VI 16, 2020. 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.345, + 0.482, + 0.385 + ], + "angle": 0, + "content": "[6] Justin Cui, Ruochen Wang, Si Si, and Cho-Jui Hsieh. Scaling up dataset distillation toImagenet-1k with constant memory. In ICML, 2023. 2, 6, 8, 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.388, + 0.482, + 0.428 + ], + "angle": 0, + "content": "[7] Yin Cui, Menglin Jia, Tsung-Yi Lin, Yang Song, and Serge Belongie. Class-balanced loss based on effective number of samples. In CVPR, 2019. 5, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.431, + 0.482, + 0.47 + ], + "angle": 0, + "content": "[8] Jia Deng, Wei Dong, Richard Socher, Li-Jia Li, Kai Li, and Li Fei-Fei. Imagenet: A large-scale hierarchical image database. In CVPR, 2009. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.474, + 0.482, + 0.528 + ], + "angle": 0, + "content": "[9] Fei Du, Peng Yang, Qi Jia, Fengtao Nan, Xiaoting Chen, and Yun Yang. Global and local mixture consistency cumulative learning for long-tailed visual recognitions. In CVPR, 2023. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.531, + 0.482, + 0.571 + ], + "angle": 0, + "content": "[10] Jiawei Du, Yidi Jiang, Vincent YF Tan, Joey Tianyi Zhou, and Haizhou Li. Minimizing the accumulated trajectory error to improve dataset distillation. In CVPR, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.574, + 0.482, + 0.601 + ], + "angle": 0, + "content": "[11] Spyros Gidaris and Nikos Komodakis. Dynamic few-shot visual learning without forgetting. In CVPR, 2018. 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.603, + 0.482, + 0.657 + ], + "angle": 0, + "content": "[12] Jianyang Gu, Saeed Vahidian, Vyacheslav Kungurtsev, Haonan Wang, Wei Jiang, Yang You, and Yiran Chen. Efficient dataset distillation via minimax diffusion. In CVPR, 2024. 1, 5, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.66, + 0.482, + 0.714 + ], + "angle": 0, + "content": "[13] Ziyao Guo, Kai Wang, George Cazenavette, Hui Li, Kaipeng Zhang, and Yang You. Towards lossless dataset distillation via difficulty-aligned trajectory matching. In ICLR, 2024. 1, 2, 3, 4, 5, 6, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.717, + 0.482, + 0.771 + ], + "angle": 0, + "content": "[14] Guo Haixiang, Li Yijing, Jennifer Shang, Gu Mingyun, Huang Yuanyue, and Gong Bing. Learning from class-imbalanced data: Review of methods and applications. Expert systems with applications, 2017. 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.774, + 0.482, + 0.814 + ], + "angle": 0, + "content": "[15] Geoffrey Hinton, Oriol Vinyals, and Jeff Dean. Distilling the knowledge in a neural network. arXiv preprint arXiv:1503.02531, 2015. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.817, + 0.482, + 0.871 + ], + "angle": 0, + "content": "[16] Geoffrey E Hinton, Nitish Srivastava, Alex Krizhevsky, Ilya Sutskever, and Ruslan R Salakhutdinov. Improving neural networks by preventing co-adaptation of feature detectors. arXiv preprint arXiv:1207.0580, 2012. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.874, + 0.482, + 0.9 + ], + "angle": 0, + "content": "[17] Rishabh Iyer, Ninad Khargoankar, Jeff Bilmes, and Himanshu Asanani. Submodular combinatorial information mea" + }, + { + "type": "list", + "bbox": [ + 0.094, + 0.116, + 0.482, + 0.9 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.549, + 0.094, + 0.905, + 0.12 + ], + "angle": 0, + "content": "sures with applications in machine learning. In Algorithmic Learning Theory, 2021. 5, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.124, + 0.905, + 0.178 + ], + "angle": 0, + "content": "[18] Andrew Janowczyk and Anant Madabhushi. Deep learning for digital pathology image analysis: A comprehensive tutorial with selected use cases. Journal of pathology informatics, 2016. 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.181, + 0.905, + 0.235 + ], + "angle": 0, + "content": "[19] Lie Ju, Xin Wang, Lin Wang, Tongliang Liu, Xin Zhao, Tom Drummond, Dwarikanath Mahapatra, and Zongyuan Ge. Relational subsets knowledge distillation for long-tailed retinal diseases recognition. In MICCAI, 2021. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.239, + 0.905, + 0.293 + ], + "angle": 0, + "content": "[20] Bingyi Kang, Saining Xie, Marcus Rohrbach, Zhicheng Yan, Albert Gordo, Jiashi Feng, and Yannis Kalantidis. Decoupling representation and classifier for long-tailed recognition. In ICLR, 2020. 2, 4, 5, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.297, + 0.905, + 0.351 + ], + "angle": 0, + "content": "[21] Jang-Hyun Kim, Jinuk Kim, Seong Joon Oh, Sangdoo Yun, Hwanjun Song, Joonhyun Jeong, Jung-Woo Ha, and Hyun Oh Song. Dataset condensation via efficient synthetic-data parameterization. In ICML, 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.354, + 0.905, + 0.381 + ], + "angle": 0, + "content": "[22] Alex Krizhevsky. Learning multiple layers of features from tiny images. University of Toronto, 2012. 1, 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.384, + 0.905, + 0.424 + ], + "angle": 0, + "content": "[23] Alex Krizhevsky, Ilya Sutskever, and Geoffrey E Hinton. Imagenet classification with deep convolutional neural networks. Communications of the ACM, 2017. 2, 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.428, + 0.905, + 0.455 + ], + "angle": 0, + "content": "[24] Yann Le and Xuan Yang. Tiny imagenet visual recognition challenge. CS 231N, 2015. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.458, + 0.905, + 0.624 + ], + "angle": 0, + "content": "[25] Zekai Li, Xinhao Zhong, Zhiyuan Liang, Yuhao Zhou, Mingjia Shi, Ziqiao Wang, Wangbo Zhao, Xuanlei Zhao, Haonan Wang, Ziheng Qin, Dai Liu, Kaipeng Zhang, Tianyi Zhou, Zheng Zhu, Kun Wang, Guang Li, Junhao Zhang, Jiawei Liu, Yiran Huang, Lingjuan Lyu, Jiancheng Lv, Yaochu Jin, Zeynep Akata, Jindong Gu, Rama Vedantam, Mike Shou, Zhiwei Deng, Yan Yan, Yuzhang Shang, George Cazenavette, Xindi Wu, Justin Cui, Tianlong Chen, Angela Yao, Manolis Kellis, Konstantinos N. Plataniotis, Bo Zhao, Zhangyang Wang, Yang You, and Kai Wang. Dd-ranking: Rethinking the evaluation of dataset distillation. GitHub repository, 2024. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.627, + 0.905, + 0.666 + ], + "angle": 0, + "content": "[26] Tsung-Yi Lin, Priya Goyal, Ross Girshick, Kaiming He, and Piotr Dollar. Focal loss for dense object detection. In ICCV, 2017. 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.67, + 0.905, + 0.71 + ], + "angle": 0, + "content": "[27] Hanxiao Liu, Karen Simonyan, and Yiming Yang. Darts: Differentiable architecture search. arXiv preprint arXiv:1806.09055, 2018. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.714, + 0.905, + 0.755 + ], + "angle": 0, + "content": "[28] Yanqing Liu, Jianyang Gu, Kai Wang, Zheng Zhu, Wei Jiang, and Yang You. Dream: Efficient dataset distillation by representative matching. In ICCV, 2023. 1, 2, 5, 6, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.758, + 0.905, + 0.799 + ], + "angle": 0, + "content": "[29] Ziwei Liu, Zhongqi Miao, Xiaohang Zhan, Jiayun Wang, Boqing Gong, and Stella X Yu. Large-scale long-tailed recognition in an open world. In CVPR, 2019. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.802, + 0.905, + 0.843 + ], + "angle": 0, + "content": "[30] Noel Loo, Ramin Hasani, Alexander Amini, and Daniela Rus. Efficient dataset distillation using random feature approximation. In NeurIPS, 2022. 1, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.846, + 0.905, + 0.9 + ], + "angle": 0, + "content": "[31] Wojciech Masarczyk and Ivona Tautkute. Reducing catastrophic forgetting with learning on synthetic data. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition Workshops, 2020. 1" + }, + { + "type": "list", + "bbox": [ + 0.518, + 0.094, + 0.905, + 0.9 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.48, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "30617" + } + ], + [ + { + "type": "ref_text", + "bbox": [ + 0.092, + 0.093, + 0.486, + 0.134 + ], + "angle": 0, + "content": "[32] Timothy Nguyen, Zhourong Chen, and Jaehoon Lee. Dataset meta-learning from kernel ridge-regression. arXiv preprint arXiv:2011.00050, 2020. 1, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.091, + 0.134, + 0.486, + 0.176 + ], + "angle": 0, + "content": "[33] Junran Peng, Xingyuan Bu, Ming Sun, Zhaoxiang Zhang, Tieniu Tan, and Junjie Yan. Large-scale object detection in the wild from imbalanced multi-labels. In CVPR, 2020. 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.092, + 0.177, + 0.485, + 0.217 + ], + "angle": 0, + "content": "[34] Jiawei Ren, Cunjun Yu, Xiao Ma, Haiyu Zhao, Shuai Yi, et al. Balanced meta-softmax for long-tailed visual recognition. In NeurIPS, 2020. 3, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.092, + 0.218, + 0.485, + 0.247 + ], + "angle": 0, + "content": "[35] Noveen Sachdeva and Julian McAuley. Data distillation: A survey. arXiv preprint arXiv:2301.04272, 2023. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.092, + 0.247, + 0.485, + 0.288 + ], + "angle": 0, + "content": "[36] Ozan Sener and Silvio Savarese. Active learning for convolutional neural networks: A core-set approach. arXiv preprint arXiv:1708.00489, 2017. 5, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.092, + 0.289, + 0.485, + 0.328 + ], + "angle": 0, + "content": "[37] Shai Shalev-Shwartz, Yoram Singer, and Nathan Srebro. Pegasos: Primal estimated sub-gradient solver forsvm. In ICML, 2007. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.092, + 0.33, + 0.485, + 0.371 + ], + "angle": 0, + "content": "[38] Yuzhang Shang, Zhihang Yuan, and Yan Yan. Mim4dd: Mutual information maximization for dataset distillation. In NeurIPS, 2024. 1, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.092, + 0.372, + 0.485, + 0.426 + ], + "angle": 0, + "content": "[39] Shitong Shao, Zeyuan Yin, Muxin Zhou, Xindong Zhang, and Zhiqiang Shen. Generalized large-scale data condensation via various backbone and statistical matching. In CVPR, 2024. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.092, + 0.428, + 0.485, + 0.468 + ], + "angle": 0, + "content": "[40] Duo Su, Junjie Hou, Weizhi Gao, Yingjie Tian, and Bowen Tang. D^4: Dataset distillation via disentangled diffusion model. In CVPR, 2024. 1, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.092, + 0.469, + 0.485, + 0.524 + ], + "angle": 0, + "content": "[41] Kai Wang, Bo Zhao, Xiangyu Peng, Zheng Zhu, Shuo Yang, Shuo Wang, Guan Huang, Hakan Bilen, Xinchao Wang, and Yang You. Cafe: Learning to condense dataset by aligning features. In CVPR, 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.092, + 0.525, + 0.485, + 0.594 + ], + "angle": 0, + "content": "[42] Kai Wang, Zekai Li, Zhi-Qi Cheng, Samir Khaki, Ahmad Sajedi, Ramakrishna Vedantam, Konstantinos N Plataniotis, Alexander Hauptmann, and Yang You. Emphasizing discriminative features for dataset distillation in complex scenarios. arXiv preprint arXiv:2410.17193, 2024. 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.092, + 0.595, + 0.485, + 0.636 + ], + "angle": 0, + "content": "[43] Tongzhou Wang, Jun-Yan Zhu, Antonio Torralba, and Alexei A Efros. Dataset distillation. arXiv preprint arXiv:1811.10959, 2018. 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.092, + 0.636, + 0.485, + 0.678 + ], + "angle": 0, + "content": "[44] Zeyuan Yin, Eric Xing, and Zhiqiang Shen. Squeeze, recover and relabel: Dataset condensation at imagenet scale from a new perspective. In NeurIPS, 2023. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.092, + 0.679, + 0.485, + 0.733 + ], + "angle": 0, + "content": "[45] Yifan Zhang, Bingyi Kang, Bryan Hooi, Shuicheng Yan, and Jiashi Feng. Deep long-tailed learning: A survey. IEEE transactions on pattern analysis and machine intelligence, 2023. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.092, + 0.734, + 0.485, + 0.775 + ], + "angle": 0, + "content": "[46] Bo Zhao, Konda Reddy Mopuri, and Hakan Bilen. Dataset condensation with gradient matching. arXiv preprint arXiv:2006.05929, 2020. 1, 5, 6, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.092, + 0.776, + 0.485, + 0.817 + ], + "angle": 0, + "content": "[47] Ganlong Zhao, Guanbin Li, Yipeng Qin, and Yizhou Yu. Improved distribution matching for dataset condensation. In CVPR, 2023. 1, 5, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.092, + 0.818, + 0.485, + 0.859 + ], + "angle": 0, + "content": "[48] Zhenghao Zhao, Yuzhang Shang, Junyi Wu, and Yan Yan. Dataset quantization with active learning based adaptive sampling. In ECCV, 2024. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.092, + 0.86, + 0.485, + 0.9 + ], + "angle": 0, + "content": "[49] Zhisheng Zhong, Jiequan Cui, Shu Liu, and Jiaya Jia. Improving calibration for long-tailed recognition. In CVPR, 2021. 2" + }, + { + "type": "list", + "bbox": [ + 0.091, + 0.093, + 0.486, + 0.9 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.516, + 0.093, + 0.909, + 0.134 + ], + "angle": 0, + "content": "[50] Yongchao Zhou, Ehsan Nezhadarya, and Jimmy Ba. Dataset distillation using neural feature regression. In NeurIPS, 2022. 1, 8" + }, + { + "type": "page_number", + "bbox": [ + 0.48, + 0.945, + 0.521, + 0.958 + ], + "angle": 0, + "content": "30618" + } + ] +] \ No newline at end of file diff --git a/2025/Distilling Long-tailed Datasets/f09fbc7b-8c7e-4bee-803c-2d67e3d05b6e_origin.pdf b/2025/Distilling Long-tailed Datasets/f09fbc7b-8c7e-4bee-803c-2d67e3d05b6e_origin.pdf new file mode 100644 index 0000000000000000000000000000000000000000..0d101a9a8f4e49cb62b4095924a5004a5daa4ae9 --- /dev/null +++ b/2025/Distilling Long-tailed Datasets/f09fbc7b-8c7e-4bee-803c-2d67e3d05b6e_origin.pdf @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:bbb7dd2aa23acdb3f79204d5794d22f8ec288a6b85354f200ab402315fe7804a +size 1473876 diff --git a/2025/Distilling Long-tailed Datasets/full.md b/2025/Distilling Long-tailed Datasets/full.md new file mode 100644 index 0000000000000000000000000000000000000000..dc18d0d58944195e222ecccc426d623effefb0d7 --- /dev/null +++ b/2025/Distilling Long-tailed Datasets/full.md @@ -0,0 +1,312 @@ +# Distilling Long-tailed Datasets + +Zhenghao Zhao $^{1*}$ Haoxuan Wang $^{1*}$ Yuzhang Shang $^{2}$ Kai Wang $^{3}$ Yan Yan $^{1\dagger}$ $^{1}$ University of Illinois Chicago $^{2}$ Illinois Institute of Technology $^{3}$ National University of Singapore +{zzhao72, hwang339, yyan55}@uic.edu yshang4@hawk.iit.edu kai.wang@comp.nus.edu.sg + +# Abstract + +Dataset distillation aims to synthesize a small, information-rich dataset from a large one for efficient model training. However, existing dataset distillation methods struggle with long-tailed datasets, which are prevalent in real-world scenarios. By investigating the reasons behind this unexpected result, we identified two main causes: 1) The distillation process on imbalanced datasets develops biased gradients, leading to the synthesis of similarly imbalanced distilled datasets. 2) The experts trained on such datasets perform suboptimally on tail classes, resulting in misguided distillation supervision and poor-quality soft-label initialization. To address these issues, we first propose Distribution-agnostic Matching to avoid directly matching the biased expert trajectories. It reduces the distance between the student and the biased expert trajectories and prevents the tail class bias from being distilled to the synthetic dataset. Moreover, we improve the distillation guidance with Expert Decoupling, which jointly matches the decoupled backbone and classifier to improve the tail class performance and initialize reliable soft labels. This work pioneers the field of long-tailed dataset distillation, marking the first effective effort to distill long-tailed datasets. Our code will be made public at https://github.com/ichbill/LTDD. + +# 1. Introduction + +Dataset distillation (DD) [4, 13, 28] aims to synthesize small and high-quality data summaries which contain the most important knowledge from a given target dataset [35]. Such a compact data summary offers numerous advantages, including faster model training, reduced memory consumption, and greater flexibility for various tasks [48], such as continual learning [31], neural architecture search [27], and knowledge distillation [15]. Various types of dataset distillation methods have been developed, including performance matching methods [30, 32, 50], parameter matching methods [13, 28, 46], representation matching methods [38, 47] + +![](images/25e4b7ae68facc9a624f0beb88f041d46f4562c1930cbe2054339a0d17b8d0d0.jpg) +Figure 1. Performance comparison on CIFAR-10-LT. Existing Dataset Distillation methods exhibit degraded performance when applied to imbalanced datasets, especially when the imbalance factor increases, whereas our method provides significantly better performance under different imbalanced scenarios. + +and generative-based methods [12, 40]. Currently, parameter matching methods have demonstrated significant performance improvements across various datasets. + +Research on dataset distillation has predominantly focused on uniformly distributed datasets, such as CIFAR10, CIFAR100 [22], and TinyImageNet [24]. However, in real-world applications such as medical image diagnosis [19], training samples typically exhibit a long-tailed class distribution [45]. In these scenarios, a small number of classes (head classes) have a large number of samples, while the majority of classes (tail classes) have only a few samples. Motivated by this discrepancy and the desire for practical application, we are interested in the novel task of long-tailed dataset distillation (LTDD). LTDD aims to distill a long-tailed target dataset into a small and uniformly distributed synthetic dataset, where models trained on this distilled dataset can achieve satisfactory performance on the uniformly distributed test set. + +Nevertheless, distilling long-tailed datasets presents significant challenges, as the effectiveness of current dataset distillation methods diminishes when applied to such imbalanced distributions. As shown in Figure 1, we find that the performance of different dataset distillation methods consistently decreases as the degree of imbalance increases. The degree of imbalance is quantified by the imbalance fac + +![](images/fb810ddfc167dc63badd1af50ccf8014a817b42c3f432b7b7a9cb78065ceb9ee.jpg) +Figure 2. Relationship of classifier weights and class-wise accuracy. We reveal that classifiers generated by existing dataset distillation methods often exhibit imbalanced distributions, resulting in poor performance on tail classes. In contrast, our method produces balanced classifiers, thereby enhancing overall accuracy. + +tor, which is a commonly used metric to characterize the extent of imbalance [23]. The larger the imbalance factor is, the more imbalanced the dataset is. When the imbalance factor is large, popular dataset distillation methods may even perform worse than random selection (e.g. DATM can only achieve $40.1\%$ accuracy when the imbalance factor $\beta = 200$ , while random selection can achieve $49.9\%$ accuracy). As shown in Figure 2, the performance degradation is concentrated in the tail classes. We attribute these performance failures to two key aspects of the learning procedure in long-tailed dataset distillation: + +(1) During distillation, the weight distributions of the student and expert diverge, resulting in less informative tail class samples. Specifically, though the expert is trained on an imbalanced dataset, the student learns on a balanced one, resulting in differently distributed model weights. Such a discrepancy introduces controversy in the bi-level optimization objective, causing the distillation process to additionally focus on matching the data quantity distribution. Consequently, the tail classes of the distilled dataset contain significantly less useful information than the head classes, despite each class having the same number of samples. + +(2) The expert models yield suboptimal performance on the tail classes. Concretely, the expert is biased toward the head classes when trained on imbalanced datasets [1, 49], resulting in degraded tail class accuracy. Since the student's performance is upper-bounded by that of the expert, the expressiveness of the distilled dataset is consequently restricted. Additionally, the image labels are obtained via the expert's prediction logits [13]. As the expert is biased toward the head classes, the prediction logits for the tail classes are, therefore, less reliable and confident. + +To address the aforementioned issues, we propose a trajectory-matching approach named Distribution-agnostic Matching with Expert Decoupling. Firstly, the student model is trained via a Distribution-agnostic Matching (DAM) strategy, which consists of a loss designed for matching experts trained on long-tailed data. Specifically, when training the student model on the balanced synthetic + +dataset, we mimic the procedure of training a model on an imbalanced dataset to reduce the distance between the student and the biased expert trajectories. Secondly, to mitigate the poor performance of the expert model and to enhance the overall distillation performance, an Expert Decoupling (ED) strategy is also proposed. In practice, we train representation experts and classification experts in a decoupled manner [20] to improve the performance of experts. We jointly match with the backbone layers of the representation experts and the classifier layers of the classification experts simultaneously. We use classification expert models to generate soft labels with high confidence for all classes. + +This work pioneers the field of long-tailed dataset distillation as the first successful effort in this area. Our experiments on four long-tailed datasets, conducted under various imbalance factors and images-per-class (IPC) settings, demonstrate the effectiveness of our method. We show that our method is able to achieve state-of-the-art performance across different dataset distillation baselines. + +# 2. Long-tailed Dataset Distillation + +# 2.1. Problem Formulation + +Assume we are given a long-tailed target training dataset $\mathcal{D} = \{(\mathbf{x}_i, y_i)\}_{i=1}^{|\mathcal{D}|}$ with $C$ classes, where $y \in \{0, \dots, C-1\}$ . Denote the subset of images belonging to class $c$ as $\mathcal{D}_c$ , then $|\mathcal{D}_0| > |\mathcal{D}_1| > \dots > |\mathcal{D}_{C-1}|$ and $|\mathcal{D}_0| \gg |\mathcal{D}_{C-1}|$ . Denote the test dataset as $\mathcal{T} = \{(\mathbf{x}_i, y_i)\}_{i=1}^{|\mathcal{T}|}$ and is balanced. Our goal is to generate a small dataset $\mathcal{S} = \{(\hat{\mathbf{x}}_i, \hat{y}_i)\}_{i=1}^{|\mathcal{S}|}$ where $|\mathcal{S}| \ll |\mathcal{D}|$ and $|\mathcal{S}_0| = |\mathcal{S}_1| = \dots = |\mathcal{S}_{C-1}|$ , so that models trained on $\mathcal{S}$ perform optimally on the test set $\mathcal{T}$ . + +Typical dataset distillation methods [4, 6, 13, 28] use the following procedure to optimize $S$ : First, train an expert model $\mathcal{M}_{\mathcal{D}}$ on $\mathcal{D}$ with acceptable performance. Then initialize $S$ by randomly selecting images from $\mathcal{D}$ or with Gaussian noise, and predict the corresponding (soft) labels using $\mathcal{M}_{\mathcal{D}}$ . A student model $\mathcal{M}_S$ is then trained based on $S$ , and the distilled image set $S$ is optimized by minimizing: + +$$ +\mathcal {L} (\mathcal {S}) = \left| \left| \mathcal {F} \left(\mathcal {M} _ {\mathcal {D}}; \mathcal {D}\right) - \mathcal {F} \left(\mathcal {M} _ {\mathcal {S}}; \mathcal {S}\right) \right| \right| _ {p} \tag {1} +$$ + +where $\mathcal{F}$ are model characteristics that can be used for matching, such as feature distributions [41], gradients [21], or training trajectories [4, 13]. Trajectory matching [4, 10, 13] matches the student model trained on the distilled dataset with the expert model trajectories trained on the original dataset, providing an impressive performance, and the matching loss is formulated as: + +$$ +\mathcal {L} _ {\mathrm {m a t c h}} (\hat {\theta} _ {t + N}, \theta_ {t + M} ^ {*}, \theta_ {t} ^ {*}) = \frac {\| \hat {\theta} _ {t + N} - \theta_ {t + M} ^ {*} \| _ {2} ^ {2}}{\| \theta_ {t} ^ {*} - \theta_ {t + M} ^ {*} \| _ {2} ^ {2}}, \quad (2) +$$ + +where $\theta_t^*$ and $\theta_{t + M}^*$ are the expert parameters on epoch $t$ and $t + M$ . $\hat{\theta}_{t + N}$ is the student parameter obtained in the + +![](images/cfd506487c3408699f59c1d930f356afa887e214ec59dd0a7ced6d54d723522d.jpg) +(a) expert trajectory + +![](images/41c6abd47e0fb19726d0b2c0d47b77f3e9dcb0b31b4ea42ecf6d40654c1cd993.jpg) +(b) training trajectory of traditional DD +Figure 3. Effect of biased expert. (a) An expert trained on an imbalanced dataset leads to increasingly imbalanced weight gradients over classes. (b) Existing dataset distillation methods [4, 13] are ignorant of the distribution gap between $S_{t}$ and $\mathcal{D}$ . This causes the student gradient imbalance to increase in each step. If we match such trajectories, the synthetic dataset will be updated by the increasingly imbalanced gradients over classes, and the model trained on this synthetic dataset is highly biased. (c) With Distribution-agnostic Matching, the increasingly imbalanced gradients over classes will be re-weighted, such that the student model is updated with balanced gradients. + +![](images/c96421f568ac5929d4e77a8274da5d6f16496e1d00b2be70ae8e0cb658e6b81d.jpg) +(c) training trajectory of our method + +inner optimization updated using the cross-entropy loss. + +In Section 2.2, we first discuss why such a formulation in Equation 1 is unreasonable for a long-tailed target dataset. Then we propose Distribution-agnostic Matching to mitigate the negative impact of the long-tailed distribution. To further enhance the distillation performance, an Expert Decoupling strategy is proposed in Section 2.3, aiding our method to achieve improved and even lossless performance. + +# 2.2. Distribution-agnostic Matching + +Influence of long-tailed distribution. By extending Equation 1 to the whole distillation process, a single step of the synthetic dataset optimization can be formulated as: + +$$ +\mathcal {S} _ {t + 1} = \underset {\mathcal {S} _ {t}} {\arg \min } | | \theta_ {\mathcal {D}} - \underset {\theta_ {S} ^ {t}} {\arg \min } \mathcal {L} \left(\theta_ {\mathcal {S}} ^ {t}, \mathcal {S} _ {t}\right) | | _ {p}, \tag {3} +$$ + +where $\theta_S^t$ is the student model parameter at epoch $t$ and $\mathcal{L}$ is the cross-entropy loss based on the student model parameter and the synthetic dataset $S_{t}$ . $\theta_{\mathcal{D}}$ is the expert model parameters trained on $\mathcal{D}$ and is determined before distillation. $p$ is the factor for $L_{p}$ norm and usually takes the value of 2. + +Equation 3 is an embedded optimization problem in which the inner loop is performed first. This implies that $S_{t}$ is updated to train a $\theta_{S}^{t}$ that can match the fixed $\theta_{\mathcal{D}}$ as much as possible. We see that $\theta_{S}^{t}$ is an important intermediate that connects the synthetic dataset and the target dataset. Under the setting that the target dataset $\mathcal{D}$ is imbalanced, the weights in $\theta_{\mathcal{D}}$ for different classes follow an imbalanced distribution and are fixed during distillation. Then the outer loop in Equation 3 would force $\theta_{S}^{t}$ to produce an imbalanced distribution similar to $\theta_{\mathcal{D}}$ . As illustrated in Figure 3(a) and (b), the student is forced to mimic the expert during distillation, so that its weight distribution in different classes gradually becomes imbalanced. + +However, the outer loop in Equation 3 aims to find a valid $S_{t}$ rather than a proper $\theta_{\mathcal{S}}^{t}$ . Thus, $\mathcal{S}_t$ is optimized to resem + +ble the training behavior of $\mathcal{D}$ . Although the tail class performance on the long-tailed dataset $\mathcal{D}$ is poor, this causes the tail classes of distilled $S_{t}$ to similarly contain much less useful information than the head classes, even though it is uniformly distributed. As a result, Equation 3 proactively distills less information from tail classes, making the distillation process more biased towards head classes. + +Distribution-agnostic Matching. As discussed above, matching the student network with an expert trained on an imbalanced dataset could lead to severe flaws in the synthetic dataset $S$ , where the tail classes are unfortunately distorted. Ideally, we would like $S$ to be a compact data summary such that every distilled image is meaningful and useful. However, such a dataset yields a student model with a balanced distribution (as shown in Figure 4 left), which mismatches the expert. This raises an interesting question: how can we preserve tail class information in the distilled dataset while avoiding the weight mismatching? + +A key insight is to weaken the relationship between the data distribution and the weight distribution. In other words, we aim to prevent weight distribution mismatches from impacting dataset optimization. To achieve this, we propose Distribution-agnostic Matching, which seeks to align the gradient distributions of the student and expert without altering the distilled dataset. Practically, we modify the student training procedure to absorb the effect of weight imbalance by deliberately making $\theta_S^t$ imbalanced. Inspired by the Balanced Softmax loss [34], we propose a long-tailed distributed loss to aid student training: + +$$ +\mathcal {L} ^ {\mathrm {c}} (\mathbf {L}, \mathbf {y}, \mathbf {s}) = - \frac {1}{n} \sum_ {i = 1} ^ {n} g \left(\mathbf {s} _ {\mathbf {y} _ {i}}\right) \cdot \log P \left(\mathbf {L} _ {i, \mathbf {y} _ {i}}\right), \tag {4} +$$ + +where + +$$ +P (\mathbf {L} _ {i, \mathbf {y} _ {i}}) = \frac {e ^ {\mathbf {L} _ {i , \mathbf {y} _ {i}} - \lambda \log (\mathbf {s} _ {\mathbf {y} _ {i}})}}{\sum_ {j = 1} ^ {C} e ^ {\mathbf {L} _ {i , j} - \lambda \log (\mathbf {s} _ {j})}}. \tag {5} +$$ + +![](images/aa144c0cd9141b32ab95844a6978b10697e5bd6b31918a4b8c1da6c530e7b616.jpg) +Figure 4. Comparison of the internal loop between traditional DD methods and ours. The expert model $\mathcal{M}_{\mathcal{D}}$ is trained on a long-tailed dataset, leading to biased weights. Left: For traditional DD methods, directly matching with these weights causes large information loss in the tail classes. The backward propagation updates this imbalance from the expert trajectories to the synthetic dataset. Right: With Distribution-agnostic Matching, the gradients are obtained via $\mathcal{L}^c$ , which revises the student weight for matching in the internal loop. This mitigates the distance between the student weight and the expert trajectory to reduce the influence of imbalance on the synthetic dataset. + +In the above equations, $\mathbf{L}$ is the student model logits, $n$ is the number of samples in the distilled dataset, $\mathbf{y}$ is the ground-truth label, $\mathbf{s}_{\mathbf{y}_i}$ is the number of samples in the long-tailed dataset for class $\mathbf{y}_i$ , $g(\cdot)$ normalizes $\mathbf{s}_{\mathbf{y}_i}$ , and $\lambda$ is used for smoothing the predictions. + +The proposed loss is designed to support student model training during distillation by reducing the dependency of the weight distribution on the data distribution, thereby "mimicking" the training behavior on long-tailed datasets. Classes with more samples contribute more significantly to the loss, simulating a long-tailed training gradient for the student network during matching. This approach helps to minimize the trajectory differences between the student and expert networks caused by distribution mismatches between the target and distilled datasets. The right plot in Figure 4 illustrates the design and effect of our method. + +# 2.3. Improved Guidance with Expert Decoupling + +Unlike experts trained on balanced target datasets, the experts trained on long-tailed datasets are less effective for guiding the dataset distillation process. Specifically, when adopting typical dataset distillation methods, the long-tailed experts perform poorly on tail classes, degrading student performance on these classes and leading to unreliable initial soft labels. This degraded student performance restricts the effective learning of synthetic tail-class images, while the unreliable labels introduce misleading information and provide weak supervision. + +In the conventional dataset distillation process, soft labels are assigned to initialized synthetic images based on the probability output of the expert model [13]. We found that soft-label predictions from long-tailed experts are distributed differently compared to those from balanced experts. As visualized in Figure 5, while the head class soft labels predicted by the balanced expert have similar confidence values with tail classes (0.88 and 0.89), the long- + +![](images/d56237401879d6ae9bceef47c64796159f9e327f0edb327f957562de3ca20d73.jpg) +Figure 5. Soft-label initialization on CIFAR-10-LT. We visualize the average predictive confidence of experts on different classes. For traditional DD, while the soft labels can be initialized well on head classes, they are predicted poorly on the tailed classes, leading to insufficient supervision. + +tailed expert predictions between head and tail classes differ drastically. Whereas the head class confidence value can be as large as 0.97, the confidence value of the tail class is surprisingly small (0.38). Such a confidence discrepancy leads to less effective learning in tail classes than in head classes. Moreover, for tail classes with a low number of samples, the number of samples with high confidence is reduced and leads to the diversity decrease of the dataset initialization. + +Decoupled training [20] is a widely used technique to alleviate the prediction bias introduced by imbalanced data. It consists of two stages: representation learning and classifier fine-tuning. At the representation learning stage, the entire model is trained on the original dataset $\mathcal{D}$ . At the classifier fine-tuning stage, the backbone of the model is frozen, and only the classifier of the model is trained on a balanced dataset $\mathcal{B}$ , which is created from $\mathcal{D}$ via oversampling or undersampling. However, the integration of this decoupling approach with dataset distillation is not trivial, as it requires careful alignment of the student model with the expert model. To effectively utilize the idea of Expert Decoupling, we propose a two-step strategy to improve teacher + +
DatasetCIFAR-10-LT
Imbalance Factor1050100200
IPC102050102050102050102050
Random32.5±2.239.6±0.951.9±1.533.2±0.442.0±1.351.6±1.334.4±2.041.4±0.752.6±0.532.5±0.842.2±1.149.9±1.4
K-Center Greedy [36]21.9±0.824.2±0.831.7±0.917.8±0.220.8±0.526.1±0.216.2±0.519.0±1.024.2±1.216.8±0.317.5±1.422.6±1.6
Graph-Cut [17]28.7±0.934.2±1.040.6±1.024.2±0.728.6±0.833.9±0.422.9±0.926.0±0.533.3±1.022.3±0.925.2±0.529.2±0.4
DC [46]37.9±0.938.5±0.937.4±1.437.3±0.938.8±1.035.8±1.236.7±0.838.1±1.035.3±1.435.6±0.835.7±0.933.3±1.4
MTT [4]58.0±0.859.5±0.462.0±0.945.8±1.449.9±0.853.6±0.537.7±0.641.6±1.147.8±1.1N/A22.6±1.023.9±0.8
DREAM [28]34.6±0.642.2±1.550.5±0.730.8±0.638.4±0.345.5±0.930.8±1.734.9±0.842.2±0.832.7±1.332.4±0.338.9±0.4
IDM [47]54.8±0.457.1±0.360.1±0.351.9±0.753.3±0.656.1±0.449.8±0.650.9±0.553.1±0.447.0±0.548.1±0.549.9±0.3
Minimax [12]29.2±0.528.5±0.639.9±0.118.4±0.322.5±0.225.2±0.219.9±0.423.3±0.228.0±0.619.1±0.420.5±0.222.7±0.3
DATM [13]57.2±0.460.4±0.266.7±0.641.6±0.243.4±0.350.3±0.237.3±0.238.9±0.144.3±0.1N/A34.8±0.140.1±0.2
Ours58.1±0.363.0±1.070.5±0.454.2±1.059.4±0.765.8±0.253.4±0.158.2±0.664.0±0.952.2±0.656.6±0.462.3±0.3
Full Dataset76.7±0.369.8±0.366.2±0.462.1±0.4
+ +Table 1. Quantitative comparisons with the SOTA methods on CIFAR-10-LT. Our method outperforms all existing approaches, and can even achieve lossless performance under certain settings. Performance is evaluated under various imbalance factors. As the first to explore long-tailed dataset distillation, all experiments are conducted using open-source code. N/A indicates distillation failure. + +guidance for student training. + +Firstly, we train two types of experts in a decoupled manner. The representation expert with parameter $\theta^{\mathcal{D}}$ is obtained from the representation learning stage, and the classifier expert with parameter $\theta^{\mathcal{B}}$ is obtained from the classifier finetuning stage. Since the classifier experts are not biased and are able to provide more accurate classification, we use the classifier experts $\theta^{\mathcal{B}}$ for the synthetic dataset $S$ initialization. Specifically, we use the logits $L_{i} = \theta^{\mathcal{B}}\mathbf{x}_{i}$ to initialize soft labels, with $\mathrm{softmax}(L_i)$ as the label of $\mathbf{x}_i$ , and then sample from $\mathcal{D}$ to initialize $S$ . + +Secondly, we match the teacher expert's training trajectory with the student's in an adaptive manner. Since the representation expert learns the image features from the original dataset $\mathcal{D}$ but fails to classify with biased classifiers [20], we skip matching the classifier layer when matching the student network trained on the synthetic dataset with the representation expert. On the contrary, the classifier expert $\mathcal{M}_{\mathcal{B}}$ only trains the classifier layer, so we also only train the student classifier layer and only match the classifier layer during the classifier matching. During the whole matching process, we jointly match the representation trajectory and the classifier trajectory using the following loss: + +$$ +\begin{array}{l} \mathcal {L} = \lambda_ {\text {r e p}} \mathcal {L} _ {\text {m a t c h}} \left(\hat {\theta} _ {t + N}, \theta_ {t + M} ^ {\mathcal {D} _ {*}}, \theta_ {t} ^ {\mathcal {D} _ {*}}\right) \\ + \lambda_ {\mathrm {c l s}} \mathcal {L} _ {\text {m a t c h}} \left(\hat {\theta} _ {t + N}, \theta_ {t + M} ^ {\mathcal {B} *}, \theta_ {t} ^ {\mathcal {B} *}\right), \tag {6} \\ \end{array} +$$ + +where $\lambda_{\mathrm{rep}}$ and $\lambda_{\mathrm{cls}}$ are hyper-parameters that control the weighting ratio, $\hat{\theta}_{t + N}$ is the student model at its training epoch $t + N$ , $\theta_t^{\mathcal{D}^*}$ and $\theta_t^{\mathcal{B}^*}$ are the representation backbone and classifier of the expert at epoch $t$ , respectively. $\hat{\theta}_{t + N}$ is updated by our proposed long-tailed distributed loss $\mathcal{L}^{\mathrm{c}}$ in Equation 4, and $\mathcal{L}_{\mathrm{match}}$ follows the form in Equation 2. + +# 3. Experiments + +# 3.1. Experiment Datasets + +We evaluate our methods on four widely used long-tailed datasets of different scales: CIFAR-10-LT, CIFAR-100-LT + +[7], TinyImageNet-LT, and ImageNet-LT [29]. The experiments are conducted under various degrees of imbalance and IPC settings. Following [7, 9, 29], these long-tailed datasets are sampled from the original balanced CIFAR [22] and ImageNet [8] datasets. The sampling procedure is done by randomly selecting different amounts of images from each class, where the image number $|\mathcal{D}_c|$ for class $c$ is determined by an exponential decay function $|\hat{\mathcal{D}}_c| = |\mathcal{D}_c|\mu^c$ , and $\mu^c = \beta^{-(c / C)}$ , where $C$ is the number of classes and $\beta$ is the imbalanced factor. The degree of imbalance is measured via the imbalanced factor $\beta = \mathcal{D}_0 / \mathcal{D}_C$ [7]. The larger $\beta$ is, the more imbalanced the dataset is. + +# 3.2. Main Results + +CIFAR-10-LT. The experiment results on CIFAR-10-LT are presented in Table 1, showing that our proposed method consistently outperforms all other dataset distillation baselines. Table 1 further highlights the advantages of our approach on long-tailed datasets. When the imbalance factor is low, prior dataset distillation methods like DATM perform reasonably well. For instance, DATM only experiences a $10.0\%$ drop in accuracy when $\beta = 10$ and IPC=50, compared to the accuracy achieved with the full dataset. In these scenarios, our method surpasses the best baseline, DATM, by $0.9\%$ on $\beta = 10$ and $3.8\%$ on $\beta = 50$ . + +When the imbalance factor is high (e.g., $\beta = 200$ ), IDM [47] achieves performance comparable to random selection. Other methods, such as MTT [4], DREAM [28], DATM [13], and Minimax [12], struggle with higher imbalance factors, performing worse than random selection. In contrast, our method sustains strong performance on highly imbalanced datasets, surpassing the best existing method IDM by a substantial margin of $10.6\%$ on $\beta = 200$ . Remarkably, we achieve lossless performance on $\beta = 200$ using only 50 images per class. + +CIFAR-100-LT. The results are presented in Table 2. Our method consistently outperforms all existing dataset distillation methods across all imbalance factors on CIFAR-100-LT. Similar to CIFAR-10-LT, our approach demonstrates + +
DatasetCIFAR-100-LTTiny-ImageNet-LT
Imbalance Factor10201020
IPC102050102050102050102050
Random14.2±0.621.7±0.632.1±0.615.0±0.321.6±0.530.5±0.57.4±0.213.5±0.420.6±0.47.6±0.113.2±0.419.8±0.1
K-Center Greedy [36]10.7±0.915.9±1.024.8±0.210.0±0.515.1±0.623.8±0.310.4±1.511.3±1.27.7±1.012.6±1.49.9±2.213.2±2.7
Graph-Cut [17]16.9±0.322.2±0.429.9±0.416.0±0.520.7±0.528.7±0.39.8±0.75.6±0.610.9±1.13.4±0.85.0±1.04.2±1.1
DC [46]24.0±0.327.4±0.327.4±0.323.2±0.326.2±0.327.4±0.3------
MTT [4]14.3±0.116.7±0.213.8±0.212.6±0.315.0±0.210.6±0.511.1±0.218.1±0.223.1±0.17.7±0.114.7±0.215.6±0.3
DREAM [28]10.1±0.412.0±1.013.1±0.49.4±0.410.3±0.612.3±0.35.4±0.36.8±0.17.8±0.24.8±0.16.0±0.27.4±0.1
DATM [13]28.2±0.434.1±0.231.6±0.125.3±0.327.2±0.127.1±0.121.3±0.114.5±0.126.8±0.114.0±0.619.0±0.323.1±0.1
Ours31.5±0.237.5±0.440.0±0.131.4±0.535.1±0.437.0±0.726.0±0.327.9±0.230.3±0.223.6±0.325.5±0.328.0±0.6
Full Dataset40.3±0.436.5±0.431.2±1.028.3±0.3
+ +Table 2. Quantitative comparisons with the SOTA methods on CIFAR-100-LT and Tiny-ImageNet-LT. When scaling up to larger datasets, our approach still outperforms all the included baselines, and achieves nearly lossless performance under several settings. + +
DatasetImageNet-LT
Imbalance Factor510
IPC10201020
Random3.9±0.17.0±0.13.9±0.16.8±0.1
G-VBSM [39]-1.0±0.1-1.0±0.1
TESLA [6]3.0±0.1-2.7±0.1-
DATM [13]7.4±0.18.1±0.27.9±0.18.2±0.1
SRe2L [44]6.7±0.110.1±0.17.7±0.310.9±1.0
Ours20.8±0.221.0±0.120.3±0.120.7±0.1
Full Dataset27.6±0.326.6±0.2
+ +greater improvements as the imbalance factor increases. + +Tiny-ImageNet-LT. We conducted experiments on TinyImageNet-LT to evaluate the scalability of our method in comparison with other baselines. The results, shown in Table 2, align with our findings on CIFAR-10-LT and CIFAR100-LT, where our method surpasses baseline methods by a growing margin as the imbalance factor increases. + +ImageNet-LT. We further scale up to ImageNet-LT, with results presented in Table 3, where our model consistently outperforms the baselines. Note that as the dataset size increases, many dataset distillation methods become resource-intensive and are impractical to apply. + +Cross-architecture performance. Table 4 presents our cross-architecture experiments, demonstrating the generalizability of our synthetic dataset. We evaluated various methods on the CIFAR-10-LT dataset with an imbalance factor of $\beta = 100$ and IPC=50, comparing against random selection, MTT, and DATM. As shown in the table, our distilled dataset consistently outperforms other methods by a significant margin. Specifically, our method achieves a $12.7\%$ accuracy improvement on ConvNet, a $17.1\%$ improvement on ResNet-18, a $7.7\%$ improvement on VGG-11, and a $6\%$ improvement on AlexNet. The results also indicate that the imbalance distilled by ConvNet in traditional DD methods negatively impacts the performance of other models. Similar to experiments on ConvNets, MTT and DATM perform poorly on unseen models due to the imbalanced dataset and the biased expert trained on it, often yielding worse results than random selection. + +Table 3. Quantitative comparisons on ImageNet-LT. The missing results are due to out-of-memory. + +
MethodConvNet-3ResNet-18VGG-11AlexNet
Random52.049.247.847.2
MTT [4]49.142.942.340.0
DATM [13]44.442.442.244.9
Ours64.760.055.553.2
+ +Table 4. Cross-architecture evaluation on CIFAR-10-LT. + +# 3.3. Ablation Studies + +Ablation on different components. Figure 6a demonstrates the effectiveness of the two proposed components of our method. These experiments were performed on the CIFAR-10-LT dataset with an imbalance factor of $\beta = 200$ and IPC=50. The "base" method is a trajectory matching method without Distribution-agnostic Matching and Expert Decoupling. "+"DAM" refers to our method without Expert Decoupling, while "ours" represents the complete proposed method, with both DAM and ED applied. + +The baseline, which suffers from weight inconsistency and poor expert performance as mentioned in Section 1, only achieves $43.8\%$ accuracy. After incorporating Distribution-agnostic Matching, the inconsistency between student and expert weights is mitigated, resulting in a $14.3\%$ performance improvement. Further integrating Expert Decoupling boosts the performance to $62.6\%$ . + +Effect of classifier layer matching. As discussed in Section 2.3, during representation learning, the model is influenced by the long-tailed distribution, resulting in a biased classifier. During classifier fine-tuning, the classifier is finetuned on balanced data with the backbone frozen. Therefore, it is crucial to explore strategies for matching each expert's different modules during dataset distillation. The experiments are presented in Figure 6b. "B", "C", and "W" refer to matching with the representation backbone, the classifier layer, and the whole model, respectively. "B/C" refers to matching the student backbone with the representation expert's backbone, and the student classifier layer with the classification expert's classifier. + +As shown in Figure 6b, the best performance is achieved by matching the backbone with representation experts and the classifier layer with classification experts. This outcome + +![](images/529cc07d02758511f59812e643dd0d6ac10f6aad8771de281648807991f4d91e.jpg) +(a) Effect of the different components + +![](images/1b5defdd0fd505dd03a2935766ca3c151a696f92257a7fe974a217ebcf3bb77d.jpg) +(b) Comparison of different matching strategies + +![](images/44073d57178891c1d5207b78cda063e4cd9a0b4d5ee901ffa1069df2a282508e.jpg) +(c) Comparison of matching strategies on low $\beta$ +Figure 6. (a) Ablation study on the proposed components, Distribution-agnostic Matching and Expert Decoupling. (b) Ablation of different matching strategies with expert model layers. (c) & (d) Comparisons of matching strategies for long-tailed dataset distillation under low imbalance factor (c) and high imbalance factor (d). (e) Ablation study on hyperparameters $\lambda_{\mathrm{rep}}$ and $\lambda_{\mathrm{cls}}$ . + +![](images/ae100f1ef62c71dc04837cb5af135c6b917e85e27cc74f4a7d39d6e05eece910.jpg) +(d) Comparison of matching strategies on high $\beta$ + +![](images/9ddd80bffe9452b76601df8cf68b7a5055056dc3fedce9876380f149ef8e2f7c.jpg) +(e) Effect of the relative scale of $\lambda_{\mathrm{rep}}$ and $\lambda_{\mathrm{cls}}$ + +aligns with the principles of decoupled learning. Since the classifier layer of representation experts is heavily affected by the long-tailed dataset, including it in the matching process causes a performance drop. In addition, matching the entire classifier expert negatively impacts performance, as only the classifier layer is trained at this stage. + +Effect of different matching strategies. When distilling a long-tailed dataset, various intuitive strategies can be employed with decoupling trained experts. We explored several strategies, including sequential matching, classification expert matching, oversampled dataset, and joint matching. Sequential matching first matches representation experts, followed by classification experts. Classification expert matching only matches with the classification expert. The oversampled dataset approach involves oversampling the tail classes before distilling the dataset rather than using the decoupling strategy. Joint matching, used in our Expert Decoupling, involves matching both experts simultaneously in the outer loop of the matching. + +As shown in Figures 6c and 6d, joint matching (joint), which we use in Expert Decoupling, achieves the best overall performance. Sequential matching (sequential) and directly matching with classification experts (direct) perform poorly when the imbalance factor is low, meaning the data is "not very imbalanced." Sequential matching suffers due to the trajectory gap between the two types of experts. Classification expert matching is less effective when the imbalance factor is low because the representation layers learn more information in such scenarios, and solely matching the classification layers fails to capture sufficient information. As depicted in Figure 6d, the oversampled dataset strategy (oversample) performs poorly when the imbalance factor is high. This finding aligns with previous work [7], which indicates that oversampling leads to model overfitting. + +Effect of hyperparameters. The most important hyperparameters in our method are $\lambda_{\mathrm{rep}}$ and $\lambda_{\mathrm{cls}}$ in Equation 6. Together, they regulate the balance between the matching representation expert and the classification expert. The setting of $\lambda_{\mathrm{rep}}$ and $\lambda_{\mathrm{cls}}$ depends on the imbalance factors. From Figure 6e, we have the following observations. For a "more imbalanced" dataset (e.g., $\beta = 200$ ), lower $\lambda_{\mathrm{rep}}$ and higher $\lambda_{\mathrm{cls}}$ + +
Hyperparametersβ = 10Hyperparametersβ = 200
λrep = 1, λcls = 0.270.5λrep = 0.2, λcls = 162.4
λrep = 1, λcls = 0.468.9λrep = 0.4, λcls = 161.7
λrep = 1, λcls = 0.668.0λrep = 0.6, λcls = 161.3
+ +Table 5. Effect of ${\lambda }_{\text{rep }}$ and ${\lambda }_{\text{cls }}$ under different imbalance factors. + +are preferred. For "less imbalanced" datasets (e.g., $\beta = 10$ ), higher $\lambda_{\mathrm{rep}}$ and lower $\lambda_{\mathrm{cls}}$ are chosen. This observation is consistent with our experiments in Figure 6c and 6d, where only matching classifier experts perform well when $\beta$ is high. This is because the representation layers learn more information when $\beta$ is low, and increasing $\lambda_{\mathrm{rep}}$ leverages information from the representation layers. + +We further analyze the sensitivity of the hyperparameters in Table 5. We can observe a minor change in the performance as we change hyperparameters rapidly. When the relative magnitude of the two hyperparameters is preserved, their absolute values do not affect the distillation performance much, suggesting the robustness of our method. + +# 3.4. Long-tailed Dataset Distillation Analysis + +Effect of imbalance in long-tailed dataset distillation. In this part, we empirically show that the performance drop in long-tailed dataset distillation is due to the data imbalance instead of the total sample number reduction of the target dataset. We compare the performance of using regular dataset distillation methods for CIFAR-10-LT (with $\beta = 200$ and $\mathrm{IPC} = 50$ ) and a subset of CIFAR-10 that contains the same amount of data but in a uniform distribution. The total number of samples is 11203 under this setting. + +As we can observe in Figure 7a, the performance gap between distilling a long-tailed dataset and a balanced subset can be as large as $28.6\%$ , even though the two target datasets have the same total amount of samples. Thus, we conclude that it is the imbalanced distribution in the target dataset that prevents effective distillation. Based on this observation, we delved into it and proposed methods to alleviate the imbalanced distribution's negative effects. + +Handling classes with extremely small number of samples. In long-tailed dataset distillation, a scenario exists where the number of samples in tail classes is even less than + +![](images/581003bfe0f5ddb2f45a55988c0be8e6d4cfccce6d9fbc5399d7bd723efd2d00.jpg) +Figure 7. (a) Comparison of DD performance between a uniform dataset and a long-tailed dataset with the same number of samples. (b) Comparison of strategies to handle an extremely low number of samples. (c) Comparison of class-wise accuracy between DATM and our method. (d) Comparison between balanced IPC and long-tailed distributed IPC. + +![](images/7e682a10346c68bccbfe6ffe0fcecf6c25ce31d9c93a719b7cdee4d840fb6be6.jpg) + +![](images/62595531ac9c8647f4f5a2775118e1eec0697b9966ead3c12928d041a1f7296b.jpg) + +![](images/a5f1b6806fffbc90a7816ede88a07a302d6c98b5f14789fbc6c4a648bd67e721.jpg) + +the required number of images per class. In this case, we employ and compare three strategies: 1) Inherent strategy, which keeps the number of samples the same in the synthetic dataset; 2) Oversampling, which duplicates the samples of the tail classes before distillation; 3) Noise generation, which randomly generates Gaussian noises to supplement the insufficient samples for tail classes. From figure 7b, we see that the inherent strategy provides a comparatively lower accuracy, while oversampling and noise generation achieve better and similar performance. + +Class-wise accuracy for long-tailed dataset. We also provide class-wise accuracy to illustrate the impact of the imbalanced data, and we show that our method can effectively improve the accuracy of tail classes. As shown in Figure 7c, DATM performs poorly on tail classes, and even achieves near-zero performance on the last class. On the contrary, our method can successfully boost the tail class accuracy while reserving performance on head classes. + +Choice of distilled dataset distribution. Should we distill a long-tailed dataset into another long-tailed dataset or a balanced one? We compared the performance of distilling the target dataset into either a similarly distributed long-tailed synthetic dataset or a balanced synthetic dataset using our method and DREAM [28]. For the long-tailed synthetic dataset, we maintained the same total number of samples as in the synthetic dataset with IPC=50. As shown in Figure 7d, both methods demonstrate that distilling the dataset into a balanced synthetic dataset provides better performance than distilling it into a long-tailed one. This result is reasonable since the model trained on the synthetic long-tailed dataset will be affected by the long-tailed distribution. + +# 4. Related Work + +# 4.1. Dataset Distillation + +Dataset distillation [43] was first formulated by Wang et al. The goal of dataset distillation is to synthesize a small dataset, such that models trained on it have good performance on the original dataset. The following works can be + +divided into kernel ridge regression-based methods [30, 32, 50], gradient matching methods [6, 13, 28, 42, 46], representation matching methods [38, 47], and generative-based methods [12, 40]. Despite the advantages of current methods, they only focus on uniformly distributed datasets, ignoring the prevalent usage of long-tailed datasets in practice. Therefore, we propose to address the long-tailed dataset distillation problem in this paper. + +# 4.2. Long-tailed Recognition + +Under real-world settings, data often tends to follow a long-tailed distribution [2]. Models trained on such datasets usually exhibit good performance on head classes but low accuracy on tail classes. A straightforward way is to preprocess the data, using oversampling [14, 18, 33], undersampling [3, 14], or data augmentation [5, 11] to make it balancedly distributed. An alternative way is weight balancing [2, 26, 34], which re-weights the layers during training to force the parameters to update unbiasedly. Decoupling methods [20] first train the model to obtain the features of the data and then fine-tune the model on the class-balanced data. In our methods, we use the decoupling method to train the expert networks during the distillation. + +# 5. Conclusion + +We introduce long-tailed dataset distillation, a novel yet challenging task. We find that existing DD methods fail when applied to long-tailed datasets and encounter two main problems: weight distribution mismatch between the student and the expert; and suboptimal performance of the expert on the tail classes. To mitigate these issues, we propose two strategies: Distribution-agnostic Matching and Expert Decoupling. Experimental results over four long-tailed datasets demonstrate the effectiveness of our proposed method. As the pioneering work in this research area, we are the first to effectively distill long-tailed datasets. + +Acknowledgments: This research is supported by NSF IIS-2525840, ECCS-2123521 and Cisco Research unrestricted gift. This article solely reflects the opinions and conclusions of its authors and not the funding agencies. + +# References + +[1] Shaden Alshammari, Yuxiong Wang, Deva Ramanan, and Shu Kong. Long-tailed recognition via weight balancing. In CVPR, 2022. 2 +[2] Shaden Alshammari, Yu-Xiong Wang, Deva Ramanan, and Shu Kong. Long-tailed recognition via weight balancing. In CVPR, 2022. 8 +[3] Mateusz Buda, Atsuto Maki, and Maciej A Mazurowski. A systematic study of the class imbalance problem in convolutional neural networks. Neural networks, 2018. 8 +[4] George Cazenavette, Tongzhou Wang, Antonio Torralba, Alexei A Efros, and Jun-Yan Zhu. Dataset distillation by matching training trajectories. In CVPR, 2022. 1, 2, 3, 5, 6 +[5] Hsin-Ping Chou, Shih-Chieh Chang, Jia-Yu Pan, Wei Wei, and Da-Cheng Juan. Remix: rebalanced mixup. In Computer Vision-ECCV 2020 Workshops: Glasgow, UK, August 23-28, 2020, Proceedings, Part VI 16, 2020. 8 +[6] Justin Cui, Ruochen Wang, Si Si, and Cho-Jui Hsieh. Scaling up dataset distillation toImagenet-1k with constant memory. In ICML, 2023. 2, 6, 8, 1 +[7] Yin Cui, Menglin Jia, Tsung-Yi Lin, Yang Song, and Serge Belongie. Class-balanced loss based on effective number of samples. In CVPR, 2019. 5, 7 +[8] Jia Deng, Wei Dong, Richard Socher, Li-Jia Li, Kai Li, and Li Fei-Fei. Imagenet: A large-scale hierarchical image database. In CVPR, 2009. 5 +[9] Fei Du, Peng Yang, Qi Jia, Fengtao Nan, Xiaoting Chen, and Yun Yang. Global and local mixture consistency cumulative learning for long-tailed visual recognitions. In CVPR, 2023. 5 +[10] Jiawei Du, Yidi Jiang, Vincent YF Tan, Joey Tianyi Zhou, and Haizhou Li. Minimizing the accumulated trajectory error to improve dataset distillation. In CVPR, 2023. 2 +[11] Spyros Gidaris and Nikos Komodakis. Dynamic few-shot visual learning without forgetting. In CVPR, 2018. 8 +[12] Jianyang Gu, Saeed Vahidian, Vyacheslav Kungurtsev, Haonan Wang, Wei Jiang, Yang You, and Yiran Chen. Efficient dataset distillation via minimax diffusion. In CVPR, 2024. 1, 5, 8 +[13] Ziyao Guo, Kai Wang, George Cazenavette, Hui Li, Kaipeng Zhang, and Yang You. Towards lossless dataset distillation via difficulty-aligned trajectory matching. In ICLR, 2024. 1, 2, 3, 4, 5, 6, 8 +[14] Guo Haixiang, Li Yijing, Jennifer Shang, Gu Mingyun, Huang Yuanyue, and Gong Bing. Learning from class-imbalanced data: Review of methods and applications. Expert systems with applications, 2017. 8 +[15] Geoffrey Hinton, Oriol Vinyals, and Jeff Dean. Distilling the knowledge in a neural network. arXiv preprint arXiv:1503.02531, 2015. 1 +[16] Geoffrey E Hinton, Nitish Srivastava, Alex Krizhevsky, Ilya Sutskever, and Ruslan R Salakhutdinov. Improving neural networks by preventing co-adaptation of feature detectors. arXiv preprint arXiv:1207.0580, 2012. 1 +[17] Rishabh Iyer, Ninad Khargoankar, Jeff Bilmes, and Himanshu Asanani. Submodular combinatorial information mea + +sures with applications in machine learning. In Algorithmic Learning Theory, 2021. 5, 6 +[18] Andrew Janowczyk and Anant Madabhushi. Deep learning for digital pathology image analysis: A comprehensive tutorial with selected use cases. Journal of pathology informatics, 2016. 8 +[19] Lie Ju, Xin Wang, Lin Wang, Tongliang Liu, Xin Zhao, Tom Drummond, Dwarikanath Mahapatra, and Zongyuan Ge. Relational subsets knowledge distillation for long-tailed retinal diseases recognition. In MICCAI, 2021. 1 +[20] Bingyi Kang, Saining Xie, Marcus Rohrbach, Zhicheng Yan, Albert Gordo, Jiashi Feng, and Yannis Kalantidis. Decoupling representation and classifier for long-tailed recognition. In ICLR, 2020. 2, 4, 5, 8 +[21] Jang-Hyun Kim, Jinuk Kim, Seong Joon Oh, Sangdoo Yun, Hwanjun Song, Joonhyun Jeong, Jung-Woo Ha, and Hyun Oh Song. Dataset condensation via efficient synthetic-data parameterization. In ICML, 2022. 2 +[22] Alex Krizhevsky. Learning multiple layers of features from tiny images. University of Toronto, 2012. 1, 5 +[23] Alex Krizhevsky, Ilya Sutskever, and Geoffrey E Hinton. Imagenet classification with deep convolutional neural networks. Communications of the ACM, 2017. 2, 1 +[24] Yann Le and Xuan Yang. Tiny imagenet visual recognition challenge. CS 231N, 2015. 1 +[25] Zekai Li, Xinhao Zhong, Zhiyuan Liang, Yuhao Zhou, Mingjia Shi, Ziqiao Wang, Wangbo Zhao, Xuanlei Zhao, Haonan Wang, Ziheng Qin, Dai Liu, Kaipeng Zhang, Tianyi Zhou, Zheng Zhu, Kun Wang, Guang Li, Junhao Zhang, Jiawei Liu, Yiran Huang, Lingjuan Lyu, Jiancheng Lv, Yaochu Jin, Zeynep Akata, Jindong Gu, Rama Vedantam, Mike Shou, Zhiwei Deng, Yan Yan, Yuzhang Shang, George Cazenavette, Xindi Wu, Justin Cui, Tianlong Chen, Angela Yao, Manolis Kellis, Konstantinos N. Plataniotis, Bo Zhao, Zhangyang Wang, Yang You, and Kai Wang. Dd-ranking: Rethinking the evaluation of dataset distillation. GitHub repository, 2024. 1 +[26] Tsung-Yi Lin, Priya Goyal, Ross Girshick, Kaiming He, and Piotr Dollar. Focal loss for dense object detection. In ICCV, 2017. 8 +[27] Hanxiao Liu, Karen Simonyan, and Yiming Yang. Darts: Differentiable architecture search. arXiv preprint arXiv:1806.09055, 2018. 1 +[28] Yanqing Liu, Jianyang Gu, Kai Wang, Zheng Zhu, Wei Jiang, and Yang You. Dream: Efficient dataset distillation by representative matching. In ICCV, 2023. 1, 2, 5, 6, 8 +[29] Ziwei Liu, Zhongqi Miao, Xiaohang Zhan, Jiayun Wang, Boqing Gong, and Stella X Yu. Large-scale long-tailed recognition in an open world. In CVPR, 2019. 5 +[30] Noel Loo, Ramin Hasani, Alexander Amini, and Daniela Rus. Efficient dataset distillation using random feature approximation. In NeurIPS, 2022. 1, 8 +[31] Wojciech Masarczyk and Ivona Tautkute. Reducing catastrophic forgetting with learning on synthetic data. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition Workshops, 2020. 1 + +[32] Timothy Nguyen, Zhourong Chen, and Jaehoon Lee. Dataset meta-learning from kernel ridge-regression. arXiv preprint arXiv:2011.00050, 2020. 1, 8 +[33] Junran Peng, Xingyuan Bu, Ming Sun, Zhaoxiang Zhang, Tieniu Tan, and Junjie Yan. Large-scale object detection in the wild from imbalanced multi-labels. In CVPR, 2020. 8 +[34] Jiawei Ren, Cunjun Yu, Xiao Ma, Haiyu Zhao, Shuai Yi, et al. Balanced meta-softmax for long-tailed visual recognition. In NeurIPS, 2020. 3, 8 +[35] Noveen Sachdeva and Julian McAuley. Data distillation: A survey. arXiv preprint arXiv:2301.04272, 2023. 1 +[36] Ozan Sener and Silvio Savarese. Active learning for convolutional neural networks: A core-set approach. arXiv preprint arXiv:1708.00489, 2017. 5, 6 +[37] Shai Shalev-Shwartz, Yoram Singer, and Nathan Srebro. Pegasos: Primal estimated sub-gradient solver forsvm. In ICML, 2007. 1 +[38] Yuzhang Shang, Zhihang Yuan, and Yan Yan. Mim4dd: Mutual information maximization for dataset distillation. In NeurIPS, 2024. 1, 8 +[39] Shitong Shao, Zeyuan Yin, Muxin Zhou, Xindong Zhang, and Zhiqiang Shen. Generalized large-scale data condensation via various backbone and statistical matching. In CVPR, 2024. 6 +[40] Duo Su, Junjie Hou, Weizhi Gao, Yingjie Tian, and Bowen Tang. D^4: Dataset distillation via disentangled diffusion model. In CVPR, 2024. 1, 8 +[41] Kai Wang, Bo Zhao, Xiangyu Peng, Zheng Zhu, Shuo Yang, Shuo Wang, Guan Huang, Hakan Bilen, Xinchao Wang, and Yang You. Cafe: Learning to condense dataset by aligning features. In CVPR, 2022. 2 +[42] Kai Wang, Zekai Li, Zhi-Qi Cheng, Samir Khaki, Ahmad Sajedi, Ramakrishna Vedantam, Konstantinos N Plataniotis, Alexander Hauptmann, and Yang You. Emphasizing discriminative features for dataset distillation in complex scenarios. arXiv preprint arXiv:2410.17193, 2024. 8 +[43] Tongzhou Wang, Jun-Yan Zhu, Antonio Torralba, and Alexei A Efros. Dataset distillation. arXiv preprint arXiv:1811.10959, 2018. 8 +[44] Zeyuan Yin, Eric Xing, and Zhiqiang Shen. Squeeze, recover and relabel: Dataset condensation at imagenet scale from a new perspective. In NeurIPS, 2023. 6 +[45] Yifan Zhang, Bingyi Kang, Bryan Hooi, Shuicheng Yan, and Jiashi Feng. Deep long-tailed learning: A survey. IEEE transactions on pattern analysis and machine intelligence, 2023. 1 +[46] Bo Zhao, Konda Reddy Mopuri, and Hakan Bilen. Dataset condensation with gradient matching. arXiv preprint arXiv:2006.05929, 2020. 1, 5, 6, 8 +[47] Ganlong Zhao, Guanbin Li, Yipeng Qin, and Yizhou Yu. Improved distribution matching for dataset condensation. In CVPR, 2023. 1, 5, 8 +[48] Zhenghao Zhao, Yuzhang Shang, Junyi Wu, and Yan Yan. Dataset quantization with active learning based adaptive sampling. In ECCV, 2024. 1 +[49] Zhisheng Zhong, Jiequan Cui, Shu Liu, and Jiaya Jia. Improving calibration for long-tailed recognition. In CVPR, 2021. 2 + +[50] Yongchao Zhou, Ehsan Nezhadarya, and Jimmy Ba. Dataset distillation using neural feature regression. In NeurIPS, 2022. 1, 8 \ No newline at end of file diff --git a/2025/Distilling Long-tailed Datasets/images.zip b/2025/Distilling Long-tailed Datasets/images.zip new file mode 100644 index 0000000000000000000000000000000000000000..1e8e9302c7733b15d34efb43bb90c8a1b7afd076 --- /dev/null +++ b/2025/Distilling Long-tailed Datasets/images.zip @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:3246508e4618fbcba67766bc73997ec420763caf3bd1b2341a66c911f78d904a +size 568042 diff --git a/2025/Distilling Long-tailed Datasets/layout.json b/2025/Distilling Long-tailed Datasets/layout.json new file mode 100644 index 0000000000000000000000000000000000000000..9f526e41c0bb53d9d98f3940f6a108b1cd56ae13 --- /dev/null +++ b/2025/Distilling Long-tailed Datasets/layout.json @@ -0,0 +1,9793 @@ +{ + "pdf_info": [ + { + "para_blocks": [ + { + "bbox": [ + 211, + 103, + 400, + 121 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 211, + 103, + 400, + 121 + ], + "spans": [ + { + "bbox": [ + 211, + 103, + 400, + 121 + ], + "type": "text", + "content": "Distilling Long-tailed Datasets" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 61, + 141, + 553, + 185 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 61, + 141, + 553, + 185 + ], + "spans": [ + { + "bbox": [ + 61, + 141, + 553, + 185 + ], + "type": "text", + "content": "Zhenghao Zhao" + }, + { + "bbox": [ + 61, + 141, + 553, + 185 + ], + "type": "inline_equation", + "content": "^{1*}" + }, + { + "bbox": [ + 61, + 141, + 553, + 185 + ], + "type": "text", + "content": " Haoxuan Wang" + }, + { + "bbox": [ + 61, + 141, + 553, + 185 + ], + "type": "inline_equation", + "content": "^{1*}" + }, + { + "bbox": [ + 61, + 141, + 553, + 185 + ], + "type": "text", + "content": " Yuzhang Shang" + }, + { + "bbox": [ + 61, + 141, + 553, + 185 + ], + "type": "inline_equation", + "content": "^{2}" + }, + { + "bbox": [ + 61, + 141, + 553, + 185 + ], + "type": "text", + "content": " Kai Wang" + }, + { + "bbox": [ + 61, + 141, + 553, + 185 + ], + "type": "inline_equation", + "content": "^{3}" + }, + { + "bbox": [ + 61, + 141, + 553, + 185 + ], + "type": "text", + "content": " Yan Yan" + }, + { + "bbox": [ + 61, + 141, + 553, + 185 + ], + "type": "inline_equation", + "content": "^{1\\dagger}" + }, + { + "bbox": [ + 61, + 141, + 553, + 185 + ], + "type": "inline_equation", + "content": "^{1}" + }, + { + "bbox": [ + 61, + 141, + 553, + 185 + ], + "type": "text", + "content": "University of Illinois Chicago " + }, + { + "bbox": [ + 61, + 141, + 553, + 185 + ], + "type": "inline_equation", + "content": "^{2}" + }, + { + "bbox": [ + 61, + 141, + 553, + 185 + ], + "type": "text", + "content": "Illinois Institute of Technology " + }, + { + "bbox": [ + 61, + 141, + 553, + 185 + ], + "type": "inline_equation", + "content": "^{3}" + }, + { + "bbox": [ + 61, + 141, + 553, + 185 + ], + "type": "text", + "content": "National University of Singapore \n{zzhao72, hwang339, yyan55}@uic.edu yshang4@hawk.iit.edu kai.wang@comp.nus.edu.sg" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 151, + 213, + 200, + 225 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 151, + 213, + 200, + 225 + ], + "spans": [ + { + "bbox": [ + 151, + 213, + 200, + 225 + ], + "type": "text", + "content": "Abstract" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 55, + 238, + 297, + 514 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 238, + 297, + 514 + ], + "spans": [ + { + "bbox": [ + 55, + 238, + 297, + 514 + ], + "type": "text", + "content": "Dataset distillation aims to synthesize a small, information-rich dataset from a large one for efficient model training. However, existing dataset distillation methods struggle with long-tailed datasets, which are prevalent in real-world scenarios. By investigating the reasons behind this unexpected result, we identified two main causes: 1) The distillation process on imbalanced datasets develops biased gradients, leading to the synthesis of similarly imbalanced distilled datasets. 2) The experts trained on such datasets perform suboptimally on tail classes, resulting in misguided distillation supervision and poor-quality soft-label initialization. To address these issues, we first propose Distribution-agnostic Matching to avoid directly matching the biased expert trajectories. It reduces the distance between the student and the biased expert trajectories and prevents the tail class bias from being distilled to the synthetic dataset. Moreover, we improve the distillation guidance with Expert Decoupling, which jointly matches the decoupled backbone and classifier to improve the tail class performance and initialize reliable soft labels. This work pioneers the field of long-tailed dataset distillation, marking the first effective effort to distill long-tailed datasets. Our code will be made public at https://github.com/ichbill/LTDD." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 55, + 535, + 135, + 547 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 535, + 135, + 547 + ], + "spans": [ + { + "bbox": [ + 55, + 535, + 135, + 547 + ], + "type": "text", + "content": "1. Introduction" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 55, + 555, + 296, + 687 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 555, + 296, + 687 + ], + "spans": [ + { + "bbox": [ + 55, + 555, + 296, + 687 + ], + "type": "text", + "content": "Dataset distillation (DD) [4, 13, 28] aims to synthesize small and high-quality data summaries which contain the most important knowledge from a given target dataset [35]. Such a compact data summary offers numerous advantages, including faster model training, reduced memory consumption, and greater flexibility for various tasks [48], such as continual learning [31], neural architecture search [27], and knowledge distillation [15]. Various types of dataset distillation methods have been developed, including performance matching methods [30, 32, 50], parameter matching methods [13, 28, 46], representation matching methods [38, 47]" + } + ] + } + ], + "index": 7 + }, + { + "type": "image", + "bbox": [ + 316, + 213, + 549, + 352 + ], + "blocks": [ + { + "bbox": [ + 316, + 213, + 549, + 352 + ], + "lines": [ + { + "bbox": [ + 316, + 213, + 549, + 352 + ], + "spans": [ + { + "bbox": [ + 316, + 213, + 549, + 352 + ], + "type": "image", + "image_path": "25e4b7ae68facc9a624f0beb88f041d46f4562c1930cbe2054339a0d17b8d0d0.jpg" + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 313, + 353, + 555, + 407 + ], + "lines": [ + { + "bbox": [ + 313, + 353, + 555, + 407 + ], + "spans": [ + { + "bbox": [ + 313, + 353, + 555, + 407 + ], + "type": "text", + "content": "Figure 1. Performance comparison on CIFAR-10-LT. Existing Dataset Distillation methods exhibit degraded performance when applied to imbalanced datasets, especially when the imbalance factor increases, whereas our method provides significantly better performance under different imbalanced scenarios." + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "image_caption" + } + ], + "index": 8 + }, + { + "bbox": [ + 313, + 412, + 554, + 448 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 412, + 554, + 448 + ], + "spans": [ + { + "bbox": [ + 313, + 412, + 554, + 448 + ], + "type": "text", + "content": "and generative-based methods [12, 40]. Currently, parameter matching methods have demonstrated significant performance improvements across various datasets." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 312, + 449, + 555, + 628 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 312, + 449, + 555, + 628 + ], + "spans": [ + { + "bbox": [ + 312, + 449, + 555, + 628 + ], + "type": "text", + "content": "Research on dataset distillation has predominantly focused on uniformly distributed datasets, such as CIFAR10, CIFAR100 [22], and TinyImageNet [24]. However, in real-world applications such as medical image diagnosis [19], training samples typically exhibit a long-tailed class distribution [45]. In these scenarios, a small number of classes (head classes) have a large number of samples, while the majority of classes (tail classes) have only a few samples. Motivated by this discrepancy and the desire for practical application, we are interested in the novel task of long-tailed dataset distillation (LTDD). LTDD aims to distill a long-tailed target dataset into a small and uniformly distributed synthetic dataset, where models trained on this distilled dataset can achieve satisfactory performance on the uniformly distributed test set." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 313, + 630, + 556, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 630, + 556, + 714 + ], + "spans": [ + { + "bbox": [ + 313, + 630, + 556, + 714 + ], + "type": "text", + "content": "Nevertheless, distilling long-tailed datasets presents significant challenges, as the effectiveness of current dataset distillation methods diminishes when applied to such imbalanced distributions. As shown in Figure 1, we find that the performance of different dataset distillation methods consistently decreases as the degree of imbalance increases. The degree of imbalance is quantified by the imbalance fac" + } + ] + } + ], + "index": 12 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "spans": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "type": "text", + "content": "CVF" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 145, + 0, + 496, + 37 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 145, + 0, + 496, + 37 + ], + "spans": [ + { + "bbox": [ + 145, + 0, + 496, + 37 + ], + "type": "text", + "content": "This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 66, + 693, + 141, + 713 + ], + "type": "page_footnote", + "angle": 0, + "lines": [ + { + "bbox": [ + 66, + 693, + 141, + 713 + ], + "spans": [ + { + "bbox": [ + 66, + 693, + 141, + 713 + ], + "type": "text", + "content": "*equal contribution † corresponding author" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "text", + "content": "30609" + } + ] + } + ], + "index": 14 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 0 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 56, + 70, + 294, + 177 + ], + "blocks": [ + { + "bbox": [ + 56, + 70, + 294, + 177 + ], + "lines": [ + { + "bbox": [ + 56, + 70, + 294, + 177 + ], + "spans": [ + { + "bbox": [ + 56, + 70, + 294, + 177 + ], + "type": "image", + "image_path": "fb810ddfc167dc63badd1af50ccf8014a817b42c3f432b7b7a9cb78065ceb9ee.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 55, + 180, + 296, + 236 + ], + "lines": [ + { + "bbox": [ + 55, + 180, + 296, + 236 + ], + "spans": [ + { + "bbox": [ + 55, + 180, + 296, + 236 + ], + "type": "text", + "content": "Figure 2. Relationship of classifier weights and class-wise accuracy. We reveal that classifiers generated by existing dataset distillation methods often exhibit imbalanced distributions, resulting in poor performance on tail classes. In contrast, our method produces balanced classifiers, thereby enhancing overall accuracy." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "bbox": [ + 55, + 240, + 296, + 373 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 240, + 296, + 373 + ], + "spans": [ + { + "bbox": [ + 55, + 240, + 296, + 373 + ], + "type": "text", + "content": "tor, which is a commonly used metric to characterize the extent of imbalance [23]. The larger the imbalance factor is, the more imbalanced the dataset is. When the imbalance factor is large, popular dataset distillation methods may even perform worse than random selection (e.g. DATM can only achieve " + }, + { + "bbox": [ + 55, + 240, + 296, + 373 + ], + "type": "inline_equation", + "content": "40.1\\%" + }, + { + "bbox": [ + 55, + 240, + 296, + 373 + ], + "type": "text", + "content": " accuracy when the imbalance factor " + }, + { + "bbox": [ + 55, + 240, + 296, + 373 + ], + "type": "inline_equation", + "content": "\\beta = 200" + }, + { + "bbox": [ + 55, + 240, + 296, + 373 + ], + "type": "text", + "content": ", while random selection can achieve " + }, + { + "bbox": [ + 55, + 240, + 296, + 373 + ], + "type": "inline_equation", + "content": "49.9\\%" + }, + { + "bbox": [ + 55, + 240, + 296, + 373 + ], + "type": "text", + "content": " accuracy). As shown in Figure 2, the performance degradation is concentrated in the tail classes. We attribute these performance failures to two key aspects of the learning procedure in long-tailed dataset distillation:" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 55, + 374, + 296, + 506 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 374, + 296, + 506 + ], + "spans": [ + { + "bbox": [ + 55, + 374, + 296, + 506 + ], + "type": "text", + "content": "(1) During distillation, the weight distributions of the student and expert diverge, resulting in less informative tail class samples. Specifically, though the expert is trained on an imbalanced dataset, the student learns on a balanced one, resulting in differently distributed model weights. Such a discrepancy introduces controversy in the bi-level optimization objective, causing the distillation process to additionally focus on matching the data quantity distribution. Consequently, the tail classes of the distilled dataset contain significantly less useful information than the head classes, despite each class having the same number of samples." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 55, + 508, + 296, + 628 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 508, + 296, + 628 + ], + "spans": [ + { + "bbox": [ + 55, + 508, + 296, + 628 + ], + "type": "text", + "content": "(2) The expert models yield suboptimal performance on the tail classes. Concretely, the expert is biased toward the head classes when trained on imbalanced datasets [1, 49], resulting in degraded tail class accuracy. Since the student's performance is upper-bounded by that of the expert, the expressiveness of the distilled dataset is consequently restricted. Additionally, the image labels are obtained via the expert's prediction logits [13]. As the expert is biased toward the head classes, the prediction logits for the tail classes are, therefore, less reliable and confident." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 55, + 630, + 296, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 630, + 296, + 715 + ], + "spans": [ + { + "bbox": [ + 55, + 630, + 296, + 715 + ], + "type": "text", + "content": "To address the aforementioned issues, we propose a trajectory-matching approach named Distribution-agnostic Matching with Expert Decoupling. Firstly, the student model is trained via a Distribution-agnostic Matching (DAM) strategy, which consists of a loss designed for matching experts trained on long-tailed data. Specifically, when training the student model on the balanced synthetic" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 313, + 72, + 555, + 215 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 72, + 555, + 215 + ], + "spans": [ + { + "bbox": [ + 313, + 72, + 555, + 215 + ], + "type": "text", + "content": "dataset, we mimic the procedure of training a model on an imbalanced dataset to reduce the distance between the student and the biased expert trajectories. Secondly, to mitigate the poor performance of the expert model and to enhance the overall distillation performance, an Expert Decoupling (ED) strategy is also proposed. In practice, we train representation experts and classification experts in a decoupled manner [20] to improve the performance of experts. We jointly match with the backbone layers of the representation experts and the classifier layers of the classification experts simultaneously. We use classification expert models to generate soft labels with high confidence for all classes." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 313, + 216, + 556, + 300 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 216, + 556, + 300 + ], + "spans": [ + { + "bbox": [ + 313, + 216, + 556, + 300 + ], + "type": "text", + "content": "This work pioneers the field of long-tailed dataset distillation as the first successful effort in this area. Our experiments on four long-tailed datasets, conducted under various imbalance factors and images-per-class (IPC) settings, demonstrate the effectiveness of our method. We show that our method is able to achieve state-of-the-art performance across different dataset distillation baselines." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 313, + 309, + 491, + 323 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 309, + 491, + 323 + ], + "spans": [ + { + "bbox": [ + 313, + 309, + 491, + 323 + ], + "type": "text", + "content": "2. Long-tailed Dataset Distillation" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 313, + 329, + 440, + 341 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 329, + 440, + 341 + ], + "spans": [ + { + "bbox": [ + 313, + 329, + 440, + 341 + ], + "type": "text", + "content": "2.1. Problem Formulation" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 313, + 347, + 555, + 449 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 347, + 555, + 449 + ], + "spans": [ + { + "bbox": [ + 313, + 347, + 555, + 449 + ], + "type": "text", + "content": "Assume we are given a long-tailed target training dataset " + }, + { + "bbox": [ + 313, + 347, + 555, + 449 + ], + "type": "inline_equation", + "content": "\\mathcal{D} = \\{(\\mathbf{x}_i, y_i)\\}_{i=1}^{|\\mathcal{D}|}" + }, + { + "bbox": [ + 313, + 347, + 555, + 449 + ], + "type": "text", + "content": " with " + }, + { + "bbox": [ + 313, + 347, + 555, + 449 + ], + "type": "inline_equation", + "content": "C" + }, + { + "bbox": [ + 313, + 347, + 555, + 449 + ], + "type": "text", + "content": " classes, where " + }, + { + "bbox": [ + 313, + 347, + 555, + 449 + ], + "type": "inline_equation", + "content": "y \\in \\{0, \\dots, C-1\\}" + }, + { + "bbox": [ + 313, + 347, + 555, + 449 + ], + "type": "text", + "content": ". Denote the subset of images belonging to class " + }, + { + "bbox": [ + 313, + 347, + 555, + 449 + ], + "type": "inline_equation", + "content": "c" + }, + { + "bbox": [ + 313, + 347, + 555, + 449 + ], + "type": "text", + "content": " as " + }, + { + "bbox": [ + 313, + 347, + 555, + 449 + ], + "type": "inline_equation", + "content": "\\mathcal{D}_c" + }, + { + "bbox": [ + 313, + 347, + 555, + 449 + ], + "type": "text", + "content": ", then " + }, + { + "bbox": [ + 313, + 347, + 555, + 449 + ], + "type": "inline_equation", + "content": "|\\mathcal{D}_0| > |\\mathcal{D}_1| > \\dots > |\\mathcal{D}_{C-1}|" + }, + { + "bbox": [ + 313, + 347, + 555, + 449 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 313, + 347, + 555, + 449 + ], + "type": "inline_equation", + "content": "|\\mathcal{D}_0| \\gg |\\mathcal{D}_{C-1}|" + }, + { + "bbox": [ + 313, + 347, + 555, + 449 + ], + "type": "text", + "content": ". Denote the test dataset as " + }, + { + "bbox": [ + 313, + 347, + 555, + 449 + ], + "type": "inline_equation", + "content": "\\mathcal{T} = \\{(\\mathbf{x}_i, y_i)\\}_{i=1}^{|\\mathcal{T}|}" + }, + { + "bbox": [ + 313, + 347, + 555, + 449 + ], + "type": "text", + "content": " and is balanced. Our goal is to generate a small dataset " + }, + { + "bbox": [ + 313, + 347, + 555, + 449 + ], + "type": "inline_equation", + "content": "\\mathcal{S} = \\{(\\hat{\\mathbf{x}}_i, \\hat{y}_i)\\}_{i=1}^{|\\mathcal{S}|}" + }, + { + "bbox": [ + 313, + 347, + 555, + 449 + ], + "type": "text", + "content": " where " + }, + { + "bbox": [ + 313, + 347, + 555, + 449 + ], + "type": "inline_equation", + "content": "|\\mathcal{S}| \\ll |\\mathcal{D}|" + }, + { + "bbox": [ + 313, + 347, + 555, + 449 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 313, + 347, + 555, + 449 + ], + "type": "inline_equation", + "content": "|\\mathcal{S}_0| = |\\mathcal{S}_1| = \\dots = |\\mathcal{S}_{C-1}|" + }, + { + "bbox": [ + 313, + 347, + 555, + 449 + ], + "type": "text", + "content": ", so that models trained on " + }, + { + "bbox": [ + 313, + 347, + 555, + 449 + ], + "type": "inline_equation", + "content": "\\mathcal{S}" + }, + { + "bbox": [ + 313, + 347, + 555, + 449 + ], + "type": "text", + "content": " perform optimally on the test set " + }, + { + "bbox": [ + 313, + 347, + 555, + 449 + ], + "type": "inline_equation", + "content": "\\mathcal{T}" + }, + { + "bbox": [ + 313, + 347, + 555, + 449 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 313, + 449, + 556, + 534 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 449, + 556, + 534 + ], + "spans": [ + { + "bbox": [ + 313, + 449, + 556, + 534 + ], + "type": "text", + "content": "Typical dataset distillation methods [4, 6, 13, 28] use the following procedure to optimize " + }, + { + "bbox": [ + 313, + 449, + 556, + 534 + ], + "type": "inline_equation", + "content": "S" + }, + { + "bbox": [ + 313, + 449, + 556, + 534 + ], + "type": "text", + "content": ": First, train an expert model " + }, + { + "bbox": [ + 313, + 449, + 556, + 534 + ], + "type": "inline_equation", + "content": "\\mathcal{M}_{\\mathcal{D}}" + }, + { + "bbox": [ + 313, + 449, + 556, + 534 + ], + "type": "text", + "content": " on " + }, + { + "bbox": [ + 313, + 449, + 556, + 534 + ], + "type": "inline_equation", + "content": "\\mathcal{D}" + }, + { + "bbox": [ + 313, + 449, + 556, + 534 + ], + "type": "text", + "content": " with acceptable performance. Then initialize " + }, + { + "bbox": [ + 313, + 449, + 556, + 534 + ], + "type": "inline_equation", + "content": "S" + }, + { + "bbox": [ + 313, + 449, + 556, + 534 + ], + "type": "text", + "content": " by randomly selecting images from " + }, + { + "bbox": [ + 313, + 449, + 556, + 534 + ], + "type": "inline_equation", + "content": "\\mathcal{D}" + }, + { + "bbox": [ + 313, + 449, + 556, + 534 + ], + "type": "text", + "content": " or with Gaussian noise, and predict the corresponding (soft) labels using " + }, + { + "bbox": [ + 313, + 449, + 556, + 534 + ], + "type": "inline_equation", + "content": "\\mathcal{M}_{\\mathcal{D}}" + }, + { + "bbox": [ + 313, + 449, + 556, + 534 + ], + "type": "text", + "content": ". A student model " + }, + { + "bbox": [ + 313, + 449, + 556, + 534 + ], + "type": "inline_equation", + "content": "\\mathcal{M}_S" + }, + { + "bbox": [ + 313, + 449, + 556, + 534 + ], + "type": "text", + "content": " is then trained based on " + }, + { + "bbox": [ + 313, + 449, + 556, + 534 + ], + "type": "inline_equation", + "content": "S" + }, + { + "bbox": [ + 313, + 449, + 556, + 534 + ], + "type": "text", + "content": ", and the distilled image set " + }, + { + "bbox": [ + 313, + 449, + 556, + 534 + ], + "type": "inline_equation", + "content": "S" + }, + { + "bbox": [ + 313, + 449, + 556, + 534 + ], + "type": "text", + "content": " is optimized by minimizing:" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 354, + 544, + 555, + 558 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 354, + 544, + 555, + 558 + ], + "spans": [ + { + "bbox": [ + 354, + 544, + 555, + 558 + ], + "type": "interline_equation", + "content": "\\mathcal {L} (\\mathcal {S}) = \\left| \\left| \\mathcal {F} \\left(\\mathcal {M} _ {\\mathcal {D}}; \\mathcal {D}\\right) - \\mathcal {F} \\left(\\mathcal {M} _ {\\mathcal {S}}; \\mathcal {S}\\right) \\right| \\right| _ {p} \\tag {1}", + "image_path": "07fafb06bd78dc7ea217816e528d0569e5eab643aac3ab5d1b5e3eb50fb642ed.jpg" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 313, + 561, + 555, + 645 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 561, + 555, + 645 + ], + "spans": [ + { + "bbox": [ + 313, + 561, + 555, + 645 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 313, + 561, + 555, + 645 + ], + "type": "inline_equation", + "content": "\\mathcal{F}" + }, + { + "bbox": [ + 313, + 561, + 555, + 645 + ], + "type": "text", + "content": " are model characteristics that can be used for matching, such as feature distributions [41], gradients [21], or training trajectories [4, 13]. Trajectory matching [4, 10, 13] matches the student model trained on the distilled dataset with the expert model trajectories trained on the original dataset, providing an impressive performance, and the matching loss is formulated as:" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 340, + 651, + 555, + 682 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 340, + 651, + 555, + 682 + ], + "spans": [ + { + "bbox": [ + 340, + 651, + 555, + 682 + ], + "type": "interline_equation", + "content": "\\mathcal {L} _ {\\mathrm {m a t c h}} (\\hat {\\theta} _ {t + N}, \\theta_ {t + M} ^ {*}, \\theta_ {t} ^ {*}) = \\frac {\\| \\hat {\\theta} _ {t + N} - \\theta_ {t + M} ^ {*} \\| _ {2} ^ {2}}{\\| \\theta_ {t} ^ {*} - \\theta_ {t + M} ^ {*} \\| _ {2} ^ {2}}, \\quad (2)", + "image_path": "f202f893361fd142d060554f8b807e3b6c0796638b49dc1d854b0848388ec40f.jpg" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 313, + 687, + 556, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 687, + 556, + 715 + ], + "spans": [ + { + "bbox": [ + 313, + 687, + 556, + 715 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 313, + 687, + 556, + 715 + ], + "type": "inline_equation", + "content": "\\theta_t^*" + }, + { + "bbox": [ + 313, + 687, + 556, + 715 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 313, + 687, + 556, + 715 + ], + "type": "inline_equation", + "content": "\\theta_{t + M}^*" + }, + { + "bbox": [ + 313, + 687, + 556, + 715 + ], + "type": "text", + "content": " are the expert parameters on epoch " + }, + { + "bbox": [ + 313, + 687, + 556, + 715 + ], + "type": "inline_equation", + "content": "t" + }, + { + "bbox": [ + 313, + 687, + 556, + 715 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 313, + 687, + 556, + 715 + ], + "type": "inline_equation", + "content": "t + M" + }, + { + "bbox": [ + 313, + 687, + 556, + 715 + ], + "type": "text", + "content": ". " + }, + { + "bbox": [ + 313, + 687, + 556, + 715 + ], + "type": "inline_equation", + "content": "\\hat{\\theta}_{t + N}" + }, + { + "bbox": [ + 313, + 687, + 556, + 715 + ], + "type": "text", + "content": " is the student parameter obtained in the" + } + ] + } + ], + "index": 15 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "type": "text", + "content": "30610" + } + ] + } + ], + "index": 16 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 1 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 66, + 82, + 216, + 209 + ], + "blocks": [ + { + "bbox": [ + 66, + 82, + 216, + 209 + ], + "lines": [ + { + "bbox": [ + 66, + 82, + 216, + 209 + ], + "spans": [ + { + "bbox": [ + 66, + 82, + 216, + 209 + ], + "type": "image", + "image_path": "cfd506487c3408699f59c1d930f356afa887e214ec59dd0a7ced6d54d723522d.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 104, + 216, + 170, + 225 + ], + "lines": [ + { + "bbox": [ + 104, + 216, + 170, + 225 + ], + "spans": [ + { + "bbox": [ + 104, + 216, + 170, + 225 + ], + "type": "text", + "content": "(a) expert trajectory" + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "type": "image", + "bbox": [ + 230, + 82, + 376, + 209 + ], + "blocks": [ + { + "bbox": [ + 230, + 82, + 376, + 209 + ], + "lines": [ + { + "bbox": [ + 230, + 82, + 376, + 209 + ], + "spans": [ + { + "bbox": [ + 230, + 82, + 376, + 209 + ], + "type": "image", + "image_path": "41c6abd47e0fb19726d0b2c0d47b77f3e9dcb0b31b4ea42ecf6d40654c1cd993.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 238, + 216, + 365, + 225 + ], + "lines": [ + { + "bbox": [ + 238, + 216, + 365, + 225 + ], + "spans": [ + { + "bbox": [ + 238, + 216, + 365, + 225 + ], + "type": "text", + "content": "(b) training trajectory of traditional DD" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 54, + 227, + 555, + 282 + ], + "lines": [ + { + "bbox": [ + 54, + 227, + 555, + 282 + ], + "spans": [ + { + "bbox": [ + 54, + 227, + 555, + 282 + ], + "type": "text", + "content": "Figure 3. Effect of biased expert. (a) An expert trained on an imbalanced dataset leads to increasingly imbalanced weight gradients over classes. (b) Existing dataset distillation methods [4, 13] are ignorant of the distribution gap between " + }, + { + "bbox": [ + 54, + 227, + 555, + 282 + ], + "type": "inline_equation", + "content": "S_{t}" + }, + { + "bbox": [ + 54, + 227, + 555, + 282 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 54, + 227, + 555, + 282 + ], + "type": "inline_equation", + "content": "\\mathcal{D}" + }, + { + "bbox": [ + 54, + 227, + 555, + 282 + ], + "type": "text", + "content": ". This causes the student gradient imbalance to increase in each step. If we match such trajectories, the synthetic dataset will be updated by the increasingly imbalanced gradients over classes, and the model trained on this synthetic dataset is highly biased. (c) With Distribution-agnostic Matching, the increasingly imbalanced gradients over classes will be re-weighted, such that the student model is updated with balanced gradients." + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "image_caption" + } + ], + "index": 2 + }, + { + "type": "image", + "bbox": [ + 395, + 82, + 541, + 209 + ], + "blocks": [ + { + "bbox": [ + 395, + 82, + 541, + 209 + ], + "lines": [ + { + "bbox": [ + 395, + 82, + 541, + 209 + ], + "spans": [ + { + "bbox": [ + 395, + 82, + 541, + 209 + ], + "type": "image", + "image_path": "c96421f568ac5929d4e77a8274da5d6f16496e1d00b2be70ae8e0cb658e6b81d.jpg" + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 406, + 216, + 527, + 225 + ], + "lines": [ + { + "bbox": [ + 406, + 216, + 527, + 225 + ], + "spans": [ + { + "bbox": [ + 406, + 216, + 527, + 225 + ], + "type": "text", + "content": "(c) training trajectory of our method" + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "image_caption" + } + ], + "index": 4 + }, + { + "bbox": [ + 55, + 284, + 282, + 295 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 284, + 282, + 295 + ], + "spans": [ + { + "bbox": [ + 55, + 284, + 282, + 295 + ], + "type": "text", + "content": "inner optimization updated using the cross-entropy loss." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 55, + 295, + 296, + 380 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 295, + 296, + 380 + ], + "spans": [ + { + "bbox": [ + 55, + 295, + 296, + 380 + ], + "type": "text", + "content": "In Section 2.2, we first discuss why such a formulation in Equation 1 is unreasonable for a long-tailed target dataset. Then we propose Distribution-agnostic Matching to mitigate the negative impact of the long-tailed distribution. To further enhance the distillation performance, an Expert Decoupling strategy is proposed in Section 2.3, aiding our method to achieve improved and even lossless performance." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 55, + 385, + 226, + 398 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 385, + 226, + 398 + ], + "spans": [ + { + "bbox": [ + 55, + 385, + 226, + 398 + ], + "type": "text", + "content": "2.2. Distribution-agnostic Matching" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 55, + 403, + 296, + 440 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 403, + 296, + 440 + ], + "spans": [ + { + "bbox": [ + 55, + 403, + 296, + 440 + ], + "type": "text", + "content": "Influence of long-tailed distribution. By extending Equation 1 to the whole distillation process, a single step of the synthetic dataset optimization can be formulated as:" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 83, + 445, + 295, + 468 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 83, + 445, + 295, + 468 + ], + "spans": [ + { + "bbox": [ + 83, + 445, + 295, + 468 + ], + "type": "interline_equation", + "content": "\\mathcal {S} _ {t + 1} = \\underset {\\mathcal {S} _ {t}} {\\arg \\min } | | \\theta_ {\\mathcal {D}} - \\underset {\\theta_ {S} ^ {t}} {\\arg \\min } \\mathcal {L} \\left(\\theta_ {\\mathcal {S}} ^ {t}, \\mathcal {S} _ {t}\\right) | | _ {p}, \\tag {3}", + "image_path": "4cb967ee12c18645ce2334b0f3830b6052c8bd2eec58611d3f08c8c935033cd9.jpg" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 55, + 474, + 296, + 534 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 474, + 296, + 534 + ], + "spans": [ + { + "bbox": [ + 55, + 474, + 296, + 534 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 55, + 474, + 296, + 534 + ], + "type": "inline_equation", + "content": "\\theta_S^t" + }, + { + "bbox": [ + 55, + 474, + 296, + 534 + ], + "type": "text", + "content": " is the student model parameter at epoch " + }, + { + "bbox": [ + 55, + 474, + 296, + 534 + ], + "type": "inline_equation", + "content": "t" + }, + { + "bbox": [ + 55, + 474, + 296, + 534 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 55, + 474, + 296, + 534 + ], + "type": "inline_equation", + "content": "\\mathcal{L}" + }, + { + "bbox": [ + 55, + 474, + 296, + 534 + ], + "type": "text", + "content": " is the cross-entropy loss based on the student model parameter and the synthetic dataset " + }, + { + "bbox": [ + 55, + 474, + 296, + 534 + ], + "type": "inline_equation", + "content": "S_{t}" + }, + { + "bbox": [ + 55, + 474, + 296, + 534 + ], + "type": "text", + "content": ". " + }, + { + "bbox": [ + 55, + 474, + 296, + 534 + ], + "type": "inline_equation", + "content": "\\theta_{\\mathcal{D}}" + }, + { + "bbox": [ + 55, + 474, + 296, + 534 + ], + "type": "text", + "content": " is the expert model parameters trained on " + }, + { + "bbox": [ + 55, + 474, + 296, + 534 + ], + "type": "inline_equation", + "content": "\\mathcal{D}" + }, + { + "bbox": [ + 55, + 474, + 296, + 534 + ], + "type": "text", + "content": " and is determined before distillation. " + }, + { + "bbox": [ + 55, + 474, + 296, + 534 + ], + "type": "inline_equation", + "content": "p" + }, + { + "bbox": [ + 55, + 474, + 296, + 534 + ], + "type": "text", + "content": " is the factor for " + }, + { + "bbox": [ + 55, + 474, + 296, + 534 + ], + "type": "inline_equation", + "content": "L_{p}" + }, + { + "bbox": [ + 55, + 474, + 296, + 534 + ], + "type": "text", + "content": " norm and usually takes the value of 2." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 55, + 534, + 295, + 689 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 534, + 295, + 689 + ], + "spans": [ + { + "bbox": [ + 55, + 534, + 295, + 689 + ], + "type": "text", + "content": "Equation 3 is an embedded optimization problem in which the inner loop is performed first. This implies that " + }, + { + "bbox": [ + 55, + 534, + 295, + 689 + ], + "type": "inline_equation", + "content": "S_{t}" + }, + { + "bbox": [ + 55, + 534, + 295, + 689 + ], + "type": "text", + "content": " is updated to train a " + }, + { + "bbox": [ + 55, + 534, + 295, + 689 + ], + "type": "inline_equation", + "content": "\\theta_{S}^{t}" + }, + { + "bbox": [ + 55, + 534, + 295, + 689 + ], + "type": "text", + "content": " that can match the fixed " + }, + { + "bbox": [ + 55, + 534, + 295, + 689 + ], + "type": "inline_equation", + "content": "\\theta_{\\mathcal{D}}" + }, + { + "bbox": [ + 55, + 534, + 295, + 689 + ], + "type": "text", + "content": " as much as possible. We see that " + }, + { + "bbox": [ + 55, + 534, + 295, + 689 + ], + "type": "inline_equation", + "content": "\\theta_{S}^{t}" + }, + { + "bbox": [ + 55, + 534, + 295, + 689 + ], + "type": "text", + "content": " is an important intermediate that connects the synthetic dataset and the target dataset. Under the setting that the target dataset " + }, + { + "bbox": [ + 55, + 534, + 295, + 689 + ], + "type": "inline_equation", + "content": "\\mathcal{D}" + }, + { + "bbox": [ + 55, + 534, + 295, + 689 + ], + "type": "text", + "content": " is imbalanced, the weights in " + }, + { + "bbox": [ + 55, + 534, + 295, + 689 + ], + "type": "inline_equation", + "content": "\\theta_{\\mathcal{D}}" + }, + { + "bbox": [ + 55, + 534, + 295, + 689 + ], + "type": "text", + "content": " for different classes follow an imbalanced distribution and are fixed during distillation. Then the outer loop in Equation 3 would force " + }, + { + "bbox": [ + 55, + 534, + 295, + 689 + ], + "type": "inline_equation", + "content": "\\theta_{S}^{t}" + }, + { + "bbox": [ + 55, + 534, + 295, + 689 + ], + "type": "text", + "content": " to produce an imbalanced distribution similar to " + }, + { + "bbox": [ + 55, + 534, + 295, + 689 + ], + "type": "inline_equation", + "content": "\\theta_{\\mathcal{D}}" + }, + { + "bbox": [ + 55, + 534, + 295, + 689 + ], + "type": "text", + "content": ". As illustrated in Figure 3(a) and (b), the student is forced to mimic the expert during distillation, so that its weight distribution in different classes gradually becomes imbalanced." + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 55, + 689, + 296, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 689, + 296, + 714 + ], + "spans": [ + { + "bbox": [ + 55, + 689, + 296, + 714 + ], + "type": "text", + "content": "However, the outer loop in Equation 3 aims to find a valid " + }, + { + "bbox": [ + 55, + 689, + 296, + 714 + ], + "type": "inline_equation", + "content": "S_{t}" + }, + { + "bbox": [ + 55, + 689, + 296, + 714 + ], + "type": "text", + "content": " rather than a proper " + }, + { + "bbox": [ + 55, + 689, + 296, + 714 + ], + "type": "inline_equation", + "content": "\\theta_{\\mathcal{S}}^{t}" + }, + { + "bbox": [ + 55, + 689, + 296, + 714 + ], + "type": "text", + "content": ". Thus, " + }, + { + "bbox": [ + 55, + 689, + 296, + 714 + ], + "type": "inline_equation", + "content": "\\mathcal{S}_t" + }, + { + "bbox": [ + 55, + 689, + 296, + 714 + ], + "type": "text", + "content": " is optimized to resem" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 313, + 283, + 555, + 366 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 283, + 555, + 366 + ], + "spans": [ + { + "bbox": [ + 313, + 283, + 555, + 366 + ], + "type": "text", + "content": "ble the training behavior of " + }, + { + "bbox": [ + 313, + 283, + 555, + 366 + ], + "type": "inline_equation", + "content": "\\mathcal{D}" + }, + { + "bbox": [ + 313, + 283, + 555, + 366 + ], + "type": "text", + "content": ". Although the tail class performance on the long-tailed dataset " + }, + { + "bbox": [ + 313, + 283, + 555, + 366 + ], + "type": "inline_equation", + "content": "\\mathcal{D}" + }, + { + "bbox": [ + 313, + 283, + 555, + 366 + ], + "type": "text", + "content": " is poor, this causes the tail classes of distilled " + }, + { + "bbox": [ + 313, + 283, + 555, + 366 + ], + "type": "inline_equation", + "content": "S_{t}" + }, + { + "bbox": [ + 313, + 283, + 555, + 366 + ], + "type": "text", + "content": " to similarly contain much less useful information than the head classes, even though it is uniformly distributed. As a result, Equation 3 proactively distills less information from tail classes, making the distillation process more biased towards head classes." + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 313, + 367, + 556, + 498 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 367, + 556, + 498 + ], + "spans": [ + { + "bbox": [ + 313, + 367, + 556, + 498 + ], + "type": "text", + "content": "Distribution-agnostic Matching. As discussed above, matching the student network with an expert trained on an imbalanced dataset could lead to severe flaws in the synthetic dataset " + }, + { + "bbox": [ + 313, + 367, + 556, + 498 + ], + "type": "inline_equation", + "content": "S" + }, + { + "bbox": [ + 313, + 367, + 556, + 498 + ], + "type": "text", + "content": ", where the tail classes are unfortunately distorted. Ideally, we would like " + }, + { + "bbox": [ + 313, + 367, + 556, + 498 + ], + "type": "inline_equation", + "content": "S" + }, + { + "bbox": [ + 313, + 367, + 556, + 498 + ], + "type": "text", + "content": " to be a compact data summary such that every distilled image is meaningful and useful. However, such a dataset yields a student model with a balanced distribution (as shown in Figure 4 left), which mismatches the expert. This raises an interesting question: how can we preserve tail class information in the distilled dataset while avoiding the weight mismatching?" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 313, + 498, + 556, + 631 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 498, + 556, + 631 + ], + "spans": [ + { + "bbox": [ + 313, + 498, + 556, + 631 + ], + "type": "text", + "content": "A key insight is to weaken the relationship between the data distribution and the weight distribution. In other words, we aim to prevent weight distribution mismatches from impacting dataset optimization. To achieve this, we propose Distribution-agnostic Matching, which seeks to align the gradient distributions of the student and expert without altering the distilled dataset. Practically, we modify the student training procedure to absorb the effect of weight imbalance by deliberately making " + }, + { + "bbox": [ + 313, + 498, + 556, + 631 + ], + "type": "inline_equation", + "content": "\\theta_S^t" + }, + { + "bbox": [ + 313, + 498, + 556, + 631 + ], + "type": "text", + "content": " imbalanced. Inspired by the Balanced Softmax loss [34], we propose a long-tailed distributed loss to aid student training:" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 343, + 638, + 555, + 669 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 343, + 638, + 555, + 669 + ], + "spans": [ + { + "bbox": [ + 343, + 638, + 555, + 669 + ], + "type": "interline_equation", + "content": "\\mathcal {L} ^ {\\mathrm {c}} (\\mathbf {L}, \\mathbf {y}, \\mathbf {s}) = - \\frac {1}{n} \\sum_ {i = 1} ^ {n} g \\left(\\mathbf {s} _ {\\mathbf {y} _ {i}}\\right) \\cdot \\log P \\left(\\mathbf {L} _ {i, \\mathbf {y} _ {i}}\\right), \\tag {4}", + "image_path": "f0b2a2d30fbdf7b6731952fd9da7d13e9621005d8dc80f94ab604137856241af.jpg" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 315, + 677, + 342, + 687 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 315, + 677, + 342, + 687 + ], + "spans": [ + { + "bbox": [ + 315, + 677, + 342, + 687 + ], + "type": "text", + "content": "where" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 364, + 685, + 555, + 717 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 364, + 685, + 555, + 717 + ], + "spans": [ + { + "bbox": [ + 364, + 685, + 555, + 717 + ], + "type": "interline_equation", + "content": "P (\\mathbf {L} _ {i, \\mathbf {y} _ {i}}) = \\frac {e ^ {\\mathbf {L} _ {i , \\mathbf {y} _ {i}} - \\lambda \\log (\\mathbf {s} _ {\\mathbf {y} _ {i}})}}{\\sum_ {j = 1} ^ {C} e ^ {\\mathbf {L} _ {i , j} - \\lambda \\log (\\mathbf {s} _ {j})}}. \\tag {5}", + "image_path": "dc89fbc3d4defe0f8ff6e8523a14a0de1bd4605685b1e104063acbd2f96b2adc.jpg" + } + ] + } + ], + "index": 20 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "text", + "content": "30611" + } + ] + } + ], + "index": 21 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 2 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 58, + 70, + 553, + 209 + ], + "blocks": [ + { + "bbox": [ + 58, + 70, + 553, + 209 + ], + "lines": [ + { + "bbox": [ + 58, + 70, + 553, + 209 + ], + "spans": [ + { + "bbox": [ + 58, + 70, + 553, + 209 + ], + "type": "image", + "image_path": "aa144c0cd9141b32ab95844a6978b10697e5bd6b31918a4b8c1da6c530e7b616.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 54, + 211, + 555, + 268 + ], + "lines": [ + { + "bbox": [ + 54, + 211, + 555, + 268 + ], + "spans": [ + { + "bbox": [ + 54, + 211, + 555, + 268 + ], + "type": "text", + "content": "Figure 4. Comparison of the internal loop between traditional DD methods and ours. The expert model " + }, + { + "bbox": [ + 54, + 211, + 555, + 268 + ], + "type": "inline_equation", + "content": "\\mathcal{M}_{\\mathcal{D}}" + }, + { + "bbox": [ + 54, + 211, + 555, + 268 + ], + "type": "text", + "content": " is trained on a long-tailed dataset, leading to biased weights. Left: For traditional DD methods, directly matching with these weights causes large information loss in the tail classes. The backward propagation updates this imbalance from the expert trajectories to the synthetic dataset. Right: With Distribution-agnostic Matching, the gradients are obtained via " + }, + { + "bbox": [ + 54, + 211, + 555, + 268 + ], + "type": "inline_equation", + "content": "\\mathcal{L}^c" + }, + { + "bbox": [ + 54, + 211, + 555, + 268 + ], + "type": "text", + "content": ", which revises the student weight for matching in the internal loop. This mitigates the distance between the student weight and the expert trajectory to reduce the influence of imbalance on the synthetic dataset." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "bbox": [ + 54, + 271, + 294, + 331 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 54, + 271, + 294, + 331 + ], + "spans": [ + { + "bbox": [ + 54, + 271, + 294, + 331 + ], + "type": "text", + "content": "In the above equations, " + }, + { + "bbox": [ + 54, + 271, + 294, + 331 + ], + "type": "inline_equation", + "content": "\\mathbf{L}" + }, + { + "bbox": [ + 54, + 271, + 294, + 331 + ], + "type": "text", + "content": " is the student model logits, " + }, + { + "bbox": [ + 54, + 271, + 294, + 331 + ], + "type": "inline_equation", + "content": "n" + }, + { + "bbox": [ + 54, + 271, + 294, + 331 + ], + "type": "text", + "content": " is the number of samples in the distilled dataset, " + }, + { + "bbox": [ + 54, + 271, + 294, + 331 + ], + "type": "inline_equation", + "content": "\\mathbf{y}" + }, + { + "bbox": [ + 54, + 271, + 294, + 331 + ], + "type": "text", + "content": " is the ground-truth label, " + }, + { + "bbox": [ + 54, + 271, + 294, + 331 + ], + "type": "inline_equation", + "content": "\\mathbf{s}_{\\mathbf{y}_i}" + }, + { + "bbox": [ + 54, + 271, + 294, + 331 + ], + "type": "text", + "content": " is the number of samples in the long-tailed dataset for class " + }, + { + "bbox": [ + 54, + 271, + 294, + 331 + ], + "type": "inline_equation", + "content": "\\mathbf{y}_i" + }, + { + "bbox": [ + 54, + 271, + 294, + 331 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 54, + 271, + 294, + 331 + ], + "type": "inline_equation", + "content": "g(\\cdot)" + }, + { + "bbox": [ + 54, + 271, + 294, + 331 + ], + "type": "text", + "content": " normalizes " + }, + { + "bbox": [ + 54, + 271, + 294, + 331 + ], + "type": "inline_equation", + "content": "\\mathbf{s}_{\\mathbf{y}_i}" + }, + { + "bbox": [ + 54, + 271, + 294, + 331 + ], + "type": "text", + "content": ", and " + }, + { + "bbox": [ + 54, + 271, + 294, + 331 + ], + "type": "inline_equation", + "content": "\\lambda" + }, + { + "bbox": [ + 54, + 271, + 294, + 331 + ], + "type": "text", + "content": " is used for smoothing the predictions." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 54, + 333, + 295, + 464 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 54, + 333, + 295, + 464 + ], + "spans": [ + { + "bbox": [ + 54, + 333, + 295, + 464 + ], + "type": "text", + "content": "The proposed loss is designed to support student model training during distillation by reducing the dependency of the weight distribution on the data distribution, thereby \"mimicking\" the training behavior on long-tailed datasets. Classes with more samples contribute more significantly to the loss, simulating a long-tailed training gradient for the student network during matching. This approach helps to minimize the trajectory differences between the student and expert networks caused by distribution mismatches between the target and distilled datasets. The right plot in Figure 4 illustrates the design and effect of our method." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 55, + 476, + 287, + 490 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 476, + 287, + 490 + ], + "spans": [ + { + "bbox": [ + 55, + 476, + 287, + 490 + ], + "type": "text", + "content": "2.3. Improved Guidance with Expert Decoupling" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 54, + 496, + 295, + 616 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 54, + 496, + 295, + 616 + ], + "spans": [ + { + "bbox": [ + 54, + 496, + 295, + 616 + ], + "type": "text", + "content": "Unlike experts trained on balanced target datasets, the experts trained on long-tailed datasets are less effective for guiding the dataset distillation process. Specifically, when adopting typical dataset distillation methods, the long-tailed experts perform poorly on tail classes, degrading student performance on these classes and leading to unreliable initial soft labels. This degraded student performance restricts the effective learning of synthetic tail-class images, while the unreliable labels introduce misleading information and provide weak supervision." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 54, + 617, + 295, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 54, + 617, + 295, + 714 + ], + "spans": [ + { + "bbox": [ + 54, + 617, + 295, + 714 + ], + "type": "text", + "content": "In the conventional dataset distillation process, soft labels are assigned to initialized synthetic images based on the probability output of the expert model [13]. We found that soft-label predictions from long-tailed experts are distributed differently compared to those from balanced experts. As visualized in Figure 5, while the head class soft labels predicted by the balanced expert have similar confidence values with tail classes (0.88 and 0.89), the long-" + } + ] + } + ], + "index": 6 + }, + { + "type": "image", + "bbox": [ + 315, + 270, + 555, + 397 + ], + "blocks": [ + { + "bbox": [ + 315, + 270, + 555, + 397 + ], + "lines": [ + { + "bbox": [ + 315, + 270, + 555, + 397 + ], + "spans": [ + { + "bbox": [ + 315, + 270, + 555, + 397 + ], + "type": "image", + "image_path": "d56237401879d6ae9bceef47c64796159f9e327f0edb327f957562de3ca20d73.jpg" + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 313, + 399, + 555, + 455 + ], + "lines": [ + { + "bbox": [ + 313, + 399, + 555, + 455 + ], + "spans": [ + { + "bbox": [ + 313, + 399, + 555, + 455 + ], + "type": "text", + "content": "Figure 5. Soft-label initialization on CIFAR-10-LT. We visualize the average predictive confidence of experts on different classes. For traditional DD, while the soft labels can be initialized well on head classes, they are predicted poorly on the tailed classes, leading to insufficient supervision." + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "image_caption" + } + ], + "index": 7 + }, + { + "bbox": [ + 313, + 461, + 554, + 556 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 461, + 554, + 556 + ], + "spans": [ + { + "bbox": [ + 313, + 461, + 554, + 556 + ], + "type": "text", + "content": "tailed expert predictions between head and tail classes differ drastically. Whereas the head class confidence value can be as large as 0.97, the confidence value of the tail class is surprisingly small (0.38). Such a confidence discrepancy leads to less effective learning in tail classes than in head classes. Moreover, for tail classes with a low number of samples, the number of samples with high confidence is reduced and leads to the diversity decrease of the dataset initialization." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 313, + 558, + 555, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 558, + 555, + 714 + ], + "spans": [ + { + "bbox": [ + 313, + 558, + 555, + 714 + ], + "type": "text", + "content": "Decoupled training [20] is a widely used technique to alleviate the prediction bias introduced by imbalanced data. It consists of two stages: representation learning and classifier fine-tuning. At the representation learning stage, the entire model is trained on the original dataset " + }, + { + "bbox": [ + 313, + 558, + 555, + 714 + ], + "type": "inline_equation", + "content": "\\mathcal{D}" + }, + { + "bbox": [ + 313, + 558, + 555, + 714 + ], + "type": "text", + "content": ". At the classifier fine-tuning stage, the backbone of the model is frozen, and only the classifier of the model is trained on a balanced dataset " + }, + { + "bbox": [ + 313, + 558, + 555, + 714 + ], + "type": "inline_equation", + "content": "\\mathcal{B}" + }, + { + "bbox": [ + 313, + 558, + 555, + 714 + ], + "type": "text", + "content": ", which is created from " + }, + { + "bbox": [ + 313, + 558, + 555, + 714 + ], + "type": "inline_equation", + "content": "\\mathcal{D}" + }, + { + "bbox": [ + 313, + 558, + 555, + 714 + ], + "type": "text", + "content": " via oversampling or undersampling. However, the integration of this decoupling approach with dataset distillation is not trivial, as it requires careful alignment of the student model with the expert model. To effectively utilize the idea of Expert Decoupling, we propose a two-step strategy to improve teacher" + } + ] + } + ], + "index": 10 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "text", + "content": "30612" + } + ] + } + ], + "index": 11 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 3 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 56, + 70, + 553, + 209 + ], + "blocks": [ + { + "bbox": [ + 56, + 70, + 553, + 209 + ], + "lines": [ + { + "bbox": [ + 56, + 70, + 553, + 209 + ], + "spans": [ + { + "bbox": [ + 56, + 70, + 553, + 209 + ], + "type": "table", + "html": "
DatasetCIFAR-10-LT
Imbalance Factor1050100200
IPC102050102050102050102050
Random32.5±2.239.6±0.951.9±1.533.2±0.442.0±1.351.6±1.334.4±2.041.4±0.752.6±0.532.5±0.842.2±1.149.9±1.4
K-Center Greedy [36]21.9±0.824.2±0.831.7±0.917.8±0.220.8±0.526.1±0.216.2±0.519.0±1.024.2±1.216.8±0.317.5±1.422.6±1.6
Graph-Cut [17]28.7±0.934.2±1.040.6±1.024.2±0.728.6±0.833.9±0.422.9±0.926.0±0.533.3±1.022.3±0.925.2±0.529.2±0.4
DC [46]37.9±0.938.5±0.937.4±1.437.3±0.938.8±1.035.8±1.236.7±0.838.1±1.035.3±1.435.6±0.835.7±0.933.3±1.4
MTT [4]58.0±0.859.5±0.462.0±0.945.8±1.449.9±0.853.6±0.537.7±0.641.6±1.147.8±1.1N/A22.6±1.023.9±0.8
DREAM [28]34.6±0.642.2±1.550.5±0.730.8±0.638.4±0.345.5±0.930.8±1.734.9±0.842.2±0.832.7±1.332.4±0.338.9±0.4
IDM [47]54.8±0.457.1±0.360.1±0.351.9±0.753.3±0.656.1±0.449.8±0.650.9±0.553.1±0.447.0±0.548.1±0.549.9±0.3
Minimax [12]29.2±0.528.5±0.639.9±0.118.4±0.322.5±0.225.2±0.219.9±0.423.3±0.228.0±0.619.1±0.420.5±0.222.7±0.3
DATM [13]57.2±0.460.4±0.266.7±0.641.6±0.243.4±0.350.3±0.237.3±0.238.9±0.144.3±0.1N/A34.8±0.140.1±0.2
Ours58.1±0.363.0±1.070.5±0.454.2±1.059.4±0.765.8±0.253.4±0.158.2±0.664.0±0.952.2±0.656.6±0.462.3±0.3
Full Dataset76.7±0.369.8±0.366.2±0.462.1±0.4
", + "image_path": "6fbf2d58651c85bcf4c269d44f50bfeecd037a18a48a9bdf326f1347dff08816.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "table_body" + } + ], + "index": 0 + }, + { + "bbox": [ + 55, + 209, + 555, + 243 + ], + "lines": [ + { + "bbox": [ + 55, + 209, + 555, + 243 + ], + "spans": [ + { + "bbox": [ + 55, + 209, + 555, + 243 + ], + "type": "text", + "content": "Table 1. Quantitative comparisons with the SOTA methods on CIFAR-10-LT. Our method outperforms all existing approaches, and can even achieve lossless performance under certain settings. Performance is evaluated under various imbalance factors. As the first to explore long-tailed dataset distillation, all experiments are conducted using open-source code. N/A indicates distillation failure." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "text" + }, + { + "bbox": [ + 55, + 247, + 176, + 258 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 247, + 176, + 258 + ], + "spans": [ + { + "bbox": [ + 55, + 247, + 176, + 258 + ], + "type": "text", + "content": "guidance for student training." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 54, + 258, + 295, + 376 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 54, + 258, + 295, + 376 + ], + "spans": [ + { + "bbox": [ + 54, + 258, + 295, + 376 + ], + "type": "text", + "content": "Firstly, we train two types of experts in a decoupled manner. The representation expert with parameter " + }, + { + "bbox": [ + 54, + 258, + 295, + 376 + ], + "type": "inline_equation", + "content": "\\theta^{\\mathcal{D}}" + }, + { + "bbox": [ + 54, + 258, + 295, + 376 + ], + "type": "text", + "content": " is obtained from the representation learning stage, and the classifier expert with parameter " + }, + { + "bbox": [ + 54, + 258, + 295, + 376 + ], + "type": "inline_equation", + "content": "\\theta^{\\mathcal{B}}" + }, + { + "bbox": [ + 54, + 258, + 295, + 376 + ], + "type": "text", + "content": " is obtained from the classifier finetuning stage. Since the classifier experts are not biased and are able to provide more accurate classification, we use the classifier experts " + }, + { + "bbox": [ + 54, + 258, + 295, + 376 + ], + "type": "inline_equation", + "content": "\\theta^{\\mathcal{B}}" + }, + { + "bbox": [ + 54, + 258, + 295, + 376 + ], + "type": "text", + "content": " for the synthetic dataset " + }, + { + "bbox": [ + 54, + 258, + 295, + 376 + ], + "type": "inline_equation", + "content": "S" + }, + { + "bbox": [ + 54, + 258, + 295, + 376 + ], + "type": "text", + "content": " initialization. Specifically, we use the logits " + }, + { + "bbox": [ + 54, + 258, + 295, + 376 + ], + "type": "inline_equation", + "content": "L_{i} = \\theta^{\\mathcal{B}}\\mathbf{x}_{i}" + }, + { + "bbox": [ + 54, + 258, + 295, + 376 + ], + "type": "text", + "content": " to initialize soft labels, with " + }, + { + "bbox": [ + 54, + 258, + 295, + 376 + ], + "type": "inline_equation", + "content": "\\mathrm{softmax}(L_i)" + }, + { + "bbox": [ + 54, + 258, + 295, + 376 + ], + "type": "text", + "content": " as the label of " + }, + { + "bbox": [ + 54, + 258, + 295, + 376 + ], + "type": "inline_equation", + "content": "\\mathbf{x}_i" + }, + { + "bbox": [ + 54, + 258, + 295, + 376 + ], + "type": "text", + "content": ", and then sample from " + }, + { + "bbox": [ + 54, + 258, + 295, + 376 + ], + "type": "inline_equation", + "content": "\\mathcal{D}" + }, + { + "bbox": [ + 54, + 258, + 295, + 376 + ], + "type": "text", + "content": " to initialize " + }, + { + "bbox": [ + 54, + 258, + 295, + 376 + ], + "type": "inline_equation", + "content": "S" + }, + { + "bbox": [ + 54, + 258, + 295, + 376 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 54, + 377, + 295, + 521 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 54, + 377, + 295, + 521 + ], + "spans": [ + { + "bbox": [ + 54, + 377, + 295, + 521 + ], + "type": "text", + "content": "Secondly, we match the teacher expert's training trajectory with the student's in an adaptive manner. Since the representation expert learns the image features from the original dataset " + }, + { + "bbox": [ + 54, + 377, + 295, + 521 + ], + "type": "inline_equation", + "content": "\\mathcal{D}" + }, + { + "bbox": [ + 54, + 377, + 295, + 521 + ], + "type": "text", + "content": " but fails to classify with biased classifiers [20], we skip matching the classifier layer when matching the student network trained on the synthetic dataset with the representation expert. On the contrary, the classifier expert " + }, + { + "bbox": [ + 54, + 377, + 295, + 521 + ], + "type": "inline_equation", + "content": "\\mathcal{M}_{\\mathcal{B}}" + }, + { + "bbox": [ + 54, + 377, + 295, + 521 + ], + "type": "text", + "content": " only trains the classifier layer, so we also only train the student classifier layer and only match the classifier layer during the classifier matching. During the whole matching process, we jointly match the representation trajectory and the classifier trajectory using the following loss:" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 100, + 528, + 295, + 559 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 100, + 528, + 295, + 559 + ], + "spans": [ + { + "bbox": [ + 100, + 528, + 295, + 559 + ], + "type": "interline_equation", + "content": "\\begin{array}{l} \\mathcal {L} = \\lambda_ {\\text {r e p}} \\mathcal {L} _ {\\text {m a t c h}} \\left(\\hat {\\theta} _ {t + N}, \\theta_ {t + M} ^ {\\mathcal {D} _ {*}}, \\theta_ {t} ^ {\\mathcal {D} _ {*}}\\right) \\\\ + \\lambda_ {\\mathrm {c l s}} \\mathcal {L} _ {\\text {m a t c h}} \\left(\\hat {\\theta} _ {t + N}, \\theta_ {t + M} ^ {\\mathcal {B} *}, \\theta_ {t} ^ {\\mathcal {B} *}\\right), \\tag {6} \\\\ \\end{array}", + "image_path": "0c4573bf7d72ac40fc6a4221201b6a853772a25688df92508d8e76335ea9554d.jpg" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 55, + 566, + 295, + 642 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 566, + 295, + 642 + ], + "spans": [ + { + "bbox": [ + 55, + 566, + 295, + 642 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 55, + 566, + 295, + 642 + ], + "type": "inline_equation", + "content": "\\lambda_{\\mathrm{rep}}" + }, + { + "bbox": [ + 55, + 566, + 295, + 642 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 55, + 566, + 295, + 642 + ], + "type": "inline_equation", + "content": "\\lambda_{\\mathrm{cls}}" + }, + { + "bbox": [ + 55, + 566, + 295, + 642 + ], + "type": "text", + "content": " are hyper-parameters that control the weighting ratio, " + }, + { + "bbox": [ + 55, + 566, + 295, + 642 + ], + "type": "inline_equation", + "content": "\\hat{\\theta}_{t + N}" + }, + { + "bbox": [ + 55, + 566, + 295, + 642 + ], + "type": "text", + "content": " is the student model at its training epoch " + }, + { + "bbox": [ + 55, + 566, + 295, + 642 + ], + "type": "inline_equation", + "content": "t + N" + }, + { + "bbox": [ + 55, + 566, + 295, + 642 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 55, + 566, + 295, + 642 + ], + "type": "inline_equation", + "content": "\\theta_t^{\\mathcal{D}^*}" + }, + { + "bbox": [ + 55, + 566, + 295, + 642 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 55, + 566, + 295, + 642 + ], + "type": "inline_equation", + "content": "\\theta_t^{\\mathcal{B}^*}" + }, + { + "bbox": [ + 55, + 566, + 295, + 642 + ], + "type": "text", + "content": " are the representation backbone and classifier of the expert at epoch " + }, + { + "bbox": [ + 55, + 566, + 295, + 642 + ], + "type": "inline_equation", + "content": "t" + }, + { + "bbox": [ + 55, + 566, + 295, + 642 + ], + "type": "text", + "content": ", respectively. " + }, + { + "bbox": [ + 55, + 566, + 295, + 642 + ], + "type": "inline_equation", + "content": "\\hat{\\theta}_{t + N}" + }, + { + "bbox": [ + 55, + 566, + 295, + 642 + ], + "type": "text", + "content": " is updated by our proposed long-tailed distributed loss " + }, + { + "bbox": [ + 55, + 566, + 295, + 642 + ], + "type": "inline_equation", + "content": "\\mathcal{L}^{\\mathrm{c}}" + }, + { + "bbox": [ + 55, + 566, + 295, + 642 + ], + "type": "text", + "content": " in Equation 4, and " + }, + { + "bbox": [ + 55, + 566, + 295, + 642 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{\\mathrm{match}}" + }, + { + "bbox": [ + 55, + 566, + 295, + 642 + ], + "type": "text", + "content": " follows the form in Equation 2." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 55, + 651, + 137, + 664 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 651, + 137, + 664 + ], + "spans": [ + { + "bbox": [ + 55, + 651, + 137, + 664 + ], + "type": "text", + "content": "3. Experiments" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 55, + 671, + 176, + 684 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 671, + 176, + 684 + ], + "spans": [ + { + "bbox": [ + 55, + 671, + 176, + 684 + ], + "type": "text", + "content": "3.1. Experiment Datasets" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 55, + 689, + 295, + 712 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 689, + 295, + 712 + ], + "spans": [ + { + "bbox": [ + 55, + 689, + 295, + 712 + ], + "type": "text", + "content": "We evaluate our methods on four widely used long-tailed datasets of different scales: CIFAR-10-LT, CIFAR-100-LT" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 313, + 246, + 555, + 389 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 246, + 555, + 389 + ], + "spans": [ + { + "bbox": [ + 313, + 246, + 555, + 389 + ], + "type": "text", + "content": "[7], TinyImageNet-LT, and ImageNet-LT [29]. The experiments are conducted under various degrees of imbalance and IPC settings. Following [7, 9, 29], these long-tailed datasets are sampled from the original balanced CIFAR [22] and ImageNet [8] datasets. The sampling procedure is done by randomly selecting different amounts of images from each class, where the image number " + }, + { + "bbox": [ + 313, + 246, + 555, + 389 + ], + "type": "inline_equation", + "content": "|\\mathcal{D}_c|" + }, + { + "bbox": [ + 313, + 246, + 555, + 389 + ], + "type": "text", + "content": " for class " + }, + { + "bbox": [ + 313, + 246, + 555, + 389 + ], + "type": "inline_equation", + "content": "c" + }, + { + "bbox": [ + 313, + 246, + 555, + 389 + ], + "type": "text", + "content": " is determined by an exponential decay function " + }, + { + "bbox": [ + 313, + 246, + 555, + 389 + ], + "type": "inline_equation", + "content": "|\\hat{\\mathcal{D}}_c| = |\\mathcal{D}_c|\\mu^c" + }, + { + "bbox": [ + 313, + 246, + 555, + 389 + ], + "type": "text", + "content": ", and " + }, + { + "bbox": [ + 313, + 246, + 555, + 389 + ], + "type": "inline_equation", + "content": "\\mu^c = \\beta^{-(c / C)}" + }, + { + "bbox": [ + 313, + 246, + 555, + 389 + ], + "type": "text", + "content": ", where " + }, + { + "bbox": [ + 313, + 246, + 555, + 389 + ], + "type": "inline_equation", + "content": "C" + }, + { + "bbox": [ + 313, + 246, + 555, + 389 + ], + "type": "text", + "content": " is the number of classes and " + }, + { + "bbox": [ + 313, + 246, + 555, + 389 + ], + "type": "inline_equation", + "content": "\\beta" + }, + { + "bbox": [ + 313, + 246, + 555, + 389 + ], + "type": "text", + "content": " is the imbalanced factor. The degree of imbalance is measured via the imbalanced factor " + }, + { + "bbox": [ + 313, + 246, + 555, + 389 + ], + "type": "inline_equation", + "content": "\\beta = \\mathcal{D}_0 / \\mathcal{D}_C" + }, + { + "bbox": [ + 313, + 246, + 555, + 389 + ], + "type": "text", + "content": " [7]. The larger " + }, + { + "bbox": [ + 313, + 246, + 555, + 389 + ], + "type": "inline_equation", + "content": "\\beta" + }, + { + "bbox": [ + 313, + 246, + 555, + 389 + ], + "type": "text", + "content": " is, the more imbalanced the dataset is." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 313, + 396, + 399, + 407 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 396, + 399, + 407 + ], + "spans": [ + { + "bbox": [ + 313, + 396, + 399, + 407 + ], + "type": "text", + "content": "3.2. Main Results" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 313, + 414, + 555, + 545 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 414, + 555, + 545 + ], + "spans": [ + { + "bbox": [ + 313, + 414, + 555, + 545 + ], + "type": "text", + "content": "CIFAR-10-LT. The experiment results on CIFAR-10-LT are presented in Table 1, showing that our proposed method consistently outperforms all other dataset distillation baselines. Table 1 further highlights the advantages of our approach on long-tailed datasets. When the imbalance factor is low, prior dataset distillation methods like DATM perform reasonably well. For instance, DATM only experiences a " + }, + { + "bbox": [ + 313, + 414, + 555, + 545 + ], + "type": "inline_equation", + "content": "10.0\\%" + }, + { + "bbox": [ + 313, + 414, + 555, + 545 + ], + "type": "text", + "content": " drop in accuracy when " + }, + { + "bbox": [ + 313, + 414, + 555, + 545 + ], + "type": "inline_equation", + "content": "\\beta = 10" + }, + { + "bbox": [ + 313, + 414, + 555, + 545 + ], + "type": "text", + "content": " and IPC=50, compared to the accuracy achieved with the full dataset. In these scenarios, our method surpasses the best baseline, DATM, by " + }, + { + "bbox": [ + 313, + 414, + 555, + 545 + ], + "type": "inline_equation", + "content": "0.9\\%" + }, + { + "bbox": [ + 313, + 414, + 555, + 545 + ], + "type": "text", + "content": " on " + }, + { + "bbox": [ + 313, + 414, + 555, + 545 + ], + "type": "inline_equation", + "content": "\\beta = 10" + }, + { + "bbox": [ + 313, + 414, + 555, + 545 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 313, + 414, + 555, + 545 + ], + "type": "inline_equation", + "content": "3.8\\%" + }, + { + "bbox": [ + 313, + 414, + 555, + 545 + ], + "type": "text", + "content": " on " + }, + { + "bbox": [ + 313, + 414, + 555, + 545 + ], + "type": "inline_equation", + "content": "\\beta = 50" + }, + { + "bbox": [ + 313, + 414, + 555, + 545 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 313, + 545, + 555, + 665 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 545, + 555, + 665 + ], + "spans": [ + { + "bbox": [ + 313, + 545, + 555, + 665 + ], + "type": "text", + "content": "When the imbalance factor is high (e.g., " + }, + { + "bbox": [ + 313, + 545, + 555, + 665 + ], + "type": "inline_equation", + "content": "\\beta = 200" + }, + { + "bbox": [ + 313, + 545, + 555, + 665 + ], + "type": "text", + "content": "), IDM [47] achieves performance comparable to random selection. Other methods, such as MTT [4], DREAM [28], DATM [13], and Minimax [12], struggle with higher imbalance factors, performing worse than random selection. In contrast, our method sustains strong performance on highly imbalanced datasets, surpassing the best existing method IDM by a substantial margin of " + }, + { + "bbox": [ + 313, + 545, + 555, + 665 + ], + "type": "inline_equation", + "content": "10.6\\%" + }, + { + "bbox": [ + 313, + 545, + 555, + 665 + ], + "type": "text", + "content": " on " + }, + { + "bbox": [ + 313, + 545, + 555, + 665 + ], + "type": "inline_equation", + "content": "\\beta = 200" + }, + { + "bbox": [ + 313, + 545, + 555, + 665 + ], + "type": "text", + "content": ". Remarkably, we achieve lossless performance on " + }, + { + "bbox": [ + 313, + 545, + 555, + 665 + ], + "type": "inline_equation", + "content": "\\beta = 200" + }, + { + "bbox": [ + 313, + 545, + 555, + 665 + ], + "type": "text", + "content": " using only 50 images per class." + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 313, + 665, + 554, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 665, + 554, + 713 + ], + "spans": [ + { + "bbox": [ + 313, + 665, + 554, + 713 + ], + "type": "text", + "content": "CIFAR-100-LT. The results are presented in Table 2. Our method consistently outperforms all existing dataset distillation methods across all imbalance factors on CIFAR-100-LT. Similar to CIFAR-10-LT, our approach demonstrates" + } + ] + } + ], + "index": 14 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "text", + "content": "30613" + } + ] + } + ], + "index": 15 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 4 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 57, + 70, + 553, + 191 + ], + "blocks": [ + { + "bbox": [ + 57, + 70, + 553, + 191 + ], + "lines": [ + { + "bbox": [ + 57, + 70, + 553, + 191 + ], + "spans": [ + { + "bbox": [ + 57, + 70, + 553, + 191 + ], + "type": "table", + "html": "
DatasetCIFAR-100-LTTiny-ImageNet-LT
Imbalance Factor10201020
IPC102050102050102050102050
Random14.2±0.621.7±0.632.1±0.615.0±0.321.6±0.530.5±0.57.4±0.213.5±0.420.6±0.47.6±0.113.2±0.419.8±0.1
K-Center Greedy [36]10.7±0.915.9±1.024.8±0.210.0±0.515.1±0.623.8±0.310.4±1.511.3±1.27.7±1.012.6±1.49.9±2.213.2±2.7
Graph-Cut [17]16.9±0.322.2±0.429.9±0.416.0±0.520.7±0.528.7±0.39.8±0.75.6±0.610.9±1.13.4±0.85.0±1.04.2±1.1
DC [46]24.0±0.327.4±0.327.4±0.323.2±0.326.2±0.327.4±0.3------
MTT [4]14.3±0.116.7±0.213.8±0.212.6±0.315.0±0.210.6±0.511.1±0.218.1±0.223.1±0.17.7±0.114.7±0.215.6±0.3
DREAM [28]10.1±0.412.0±1.013.1±0.49.4±0.410.3±0.612.3±0.35.4±0.36.8±0.17.8±0.24.8±0.16.0±0.27.4±0.1
DATM [13]28.2±0.434.1±0.231.6±0.125.3±0.327.2±0.127.1±0.121.3±0.114.5±0.126.8±0.114.0±0.619.0±0.323.1±0.1
Ours31.5±0.237.5±0.440.0±0.131.4±0.535.1±0.437.0±0.726.0±0.327.9±0.230.3±0.223.6±0.325.5±0.328.0±0.6
Full Dataset40.3±0.436.5±0.431.2±1.028.3±0.3
", + "image_path": "ee663d72e232dc4c127d2c434befaa598c0f4387f898210b44d22289aac1942b.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "table_body" + } + ], + "index": 0 + }, + { + "type": "table", + "bbox": [ + 58, + 220, + 293, + 338 + ], + "blocks": [ + { + "bbox": [ + 55, + 193, + 553, + 216 + ], + "lines": [ + { + "bbox": [ + 55, + 193, + 553, + 216 + ], + "spans": [ + { + "bbox": [ + 55, + 193, + 553, + 216 + ], + "type": "text", + "content": "Table 2. Quantitative comparisons with the SOTA methods on CIFAR-100-LT and Tiny-ImageNet-LT. When scaling up to larger datasets, our approach still outperforms all the included baselines, and achieves nearly lossless performance under several settings." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 58, + 220, + 293, + 338 + ], + "lines": [ + { + "bbox": [ + 58, + 220, + 293, + 338 + ], + "spans": [ + { + "bbox": [ + 58, + 220, + 293, + 338 + ], + "type": "table", + "html": "
DatasetImageNet-LT
Imbalance Factor510
IPC10201020
Random3.9±0.17.0±0.13.9±0.16.8±0.1
G-VBSM [39]-1.0±0.1-1.0±0.1
TESLA [6]3.0±0.1-2.7±0.1-
DATM [13]7.4±0.18.1±0.27.9±0.18.2±0.1
SRe2L [44]6.7±0.110.1±0.17.7±0.310.9±1.0
Ours20.8±0.221.0±0.120.3±0.120.7±0.1
Full Dataset27.6±0.326.6±0.2
", + "image_path": "c67b8dee813e55634940c1b40264f3786a2ff61568987f6af15e878d3dba3d41.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "table_body" + } + ], + "index": 2 + }, + { + "bbox": [ + 55, + 367, + 282, + 378 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 367, + 282, + 378 + ], + "spans": [ + { + "bbox": [ + 55, + 367, + 282, + 378 + ], + "type": "text", + "content": "greater improvements as the imbalance factor increases." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 55, + 379, + 295, + 450 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 379, + 295, + 450 + ], + "spans": [ + { + "bbox": [ + 55, + 379, + 295, + 450 + ], + "type": "text", + "content": "Tiny-ImageNet-LT. We conducted experiments on TinyImageNet-LT to evaluate the scalability of our method in comparison with other baselines. The results, shown in Table 2, align with our findings on CIFAR-10-LT and CIFAR100-LT, where our method surpasses baseline methods by a growing margin as the imbalance factor increases." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 55, + 450, + 295, + 511 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 450, + 295, + 511 + ], + "spans": [ + { + "bbox": [ + 55, + 450, + 295, + 511 + ], + "type": "text", + "content": "ImageNet-LT. We further scale up to ImageNet-LT, with results presented in Table 3, where our model consistently outperforms the baselines. Note that as the dataset size increases, many dataset distillation methods become resource-intensive and are impractical to apply." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 55, + 511, + 295, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 511, + 295, + 714 + ], + "spans": [ + { + "bbox": [ + 55, + 511, + 295, + 714 + ], + "type": "text", + "content": "Cross-architecture performance. Table 4 presents our cross-architecture experiments, demonstrating the generalizability of our synthetic dataset. We evaluated various methods on the CIFAR-10-LT dataset with an imbalance factor of " + }, + { + "bbox": [ + 55, + 511, + 295, + 714 + ], + "type": "inline_equation", + "content": "\\beta = 100" + }, + { + "bbox": [ + 55, + 511, + 295, + 714 + ], + "type": "text", + "content": " and IPC=50, comparing against random selection, MTT, and DATM. As shown in the table, our distilled dataset consistently outperforms other methods by a significant margin. Specifically, our method achieves a " + }, + { + "bbox": [ + 55, + 511, + 295, + 714 + ], + "type": "inline_equation", + "content": "12.7\\%" + }, + { + "bbox": [ + 55, + 511, + 295, + 714 + ], + "type": "text", + "content": " accuracy improvement on ConvNet, a " + }, + { + "bbox": [ + 55, + 511, + 295, + 714 + ], + "type": "inline_equation", + "content": "17.1\\%" + }, + { + "bbox": [ + 55, + 511, + 295, + 714 + ], + "type": "text", + "content": " improvement on ResNet-18, a " + }, + { + "bbox": [ + 55, + 511, + 295, + 714 + ], + "type": "inline_equation", + "content": "7.7\\%" + }, + { + "bbox": [ + 55, + 511, + 295, + 714 + ], + "type": "text", + "content": " improvement on VGG-11, and a " + }, + { + "bbox": [ + 55, + 511, + 295, + 714 + ], + "type": "inline_equation", + "content": "6\\%" + }, + { + "bbox": [ + 55, + 511, + 295, + 714 + ], + "type": "text", + "content": " improvement on AlexNet. The results also indicate that the imbalance distilled by ConvNet in traditional DD methods negatively impacts the performance of other models. Similar to experiments on ConvNets, MTT and DATM perform poorly on unseen models due to the imbalanced dataset and the biased expert trained on it, often yielding worse results than random selection." + } + ] + } + ], + "index": 7 + }, + { + "type": "table", + "bbox": [ + 317, + 220, + 552, + 285 + ], + "blocks": [ + { + "bbox": [ + 55, + 340, + 295, + 361 + ], + "lines": [ + { + "bbox": [ + 55, + 340, + 295, + 361 + ], + "spans": [ + { + "bbox": [ + 55, + 340, + 295, + 361 + ], + "type": "text", + "content": "Table 3. Quantitative comparisons on ImageNet-LT. The missing results are due to out-of-memory." + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 317, + 220, + 552, + 285 + ], + "lines": [ + { + "bbox": [ + 317, + 220, + 552, + 285 + ], + "spans": [ + { + "bbox": [ + 317, + 220, + 552, + 285 + ], + "type": "table", + "html": "
MethodConvNet-3ResNet-18VGG-11AlexNet
Random52.049.247.847.2
MTT [4]49.142.942.340.0
DATM [13]44.442.442.244.9
Ours64.760.055.553.2
", + "image_path": "8239db1aa423d66dfa23084e3bfbc2e50a91948450472f3823d3b2b66d31281f.jpg" + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "table_body" + } + ], + "index": 8 + }, + { + "bbox": [ + 325, + 285, + 543, + 297 + ], + "lines": [ + { + "bbox": [ + 325, + 285, + 543, + 297 + ], + "spans": [ + { + "bbox": [ + 325, + 285, + 543, + 297 + ], + "type": "text", + "content": "Table 4. Cross-architecture evaluation on CIFAR-10-LT." + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "text" + }, + { + "bbox": [ + 314, + 305, + 415, + 317 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 314, + 305, + 415, + 317 + ], + "spans": [ + { + "bbox": [ + 314, + 305, + 415, + 317 + ], + "type": "text", + "content": "3.3. Ablation Studies" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 313, + 325, + 555, + 434 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 325, + 555, + 434 + ], + "spans": [ + { + "bbox": [ + 313, + 325, + 555, + 434 + ], + "type": "text", + "content": "Ablation on different components. Figure 6a demonstrates the effectiveness of the two proposed components of our method. These experiments were performed on the CIFAR-10-LT dataset with an imbalance factor of " + }, + { + "bbox": [ + 313, + 325, + 555, + 434 + ], + "type": "inline_equation", + "content": "\\beta = 200" + }, + { + "bbox": [ + 313, + 325, + 555, + 434 + ], + "type": "text", + "content": " and IPC=50. The \"base\" method is a trajectory matching method without Distribution-agnostic Matching and Expert Decoupling. \"+\"DAM\" refers to our method without Expert Decoupling, while \"ours\" represents the complete proposed method, with both DAM and ED applied." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 313, + 434, + 554, + 518 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 434, + 554, + 518 + ], + "spans": [ + { + "bbox": [ + 313, + 434, + 554, + 518 + ], + "type": "text", + "content": "The baseline, which suffers from weight inconsistency and poor expert performance as mentioned in Section 1, only achieves " + }, + { + "bbox": [ + 313, + 434, + 554, + 518 + ], + "type": "inline_equation", + "content": "43.8\\%" + }, + { + "bbox": [ + 313, + 434, + 554, + 518 + ], + "type": "text", + "content": " accuracy. After incorporating Distribution-agnostic Matching, the inconsistency between student and expert weights is mitigated, resulting in a " + }, + { + "bbox": [ + 313, + 434, + 554, + 518 + ], + "type": "inline_equation", + "content": "14.3\\%" + }, + { + "bbox": [ + 313, + 434, + 554, + 518 + ], + "type": "text", + "content": " performance improvement. Further integrating Expert Decoupling boosts the performance to " + }, + { + "bbox": [ + 313, + 434, + 554, + 518 + ], + "type": "inline_equation", + "content": "62.6\\%" + }, + { + "bbox": [ + 313, + 434, + 554, + 518 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 313, + 520, + 554, + 675 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 520, + 554, + 675 + ], + "spans": [ + { + "bbox": [ + 313, + 520, + 554, + 675 + ], + "type": "text", + "content": "Effect of classifier layer matching. As discussed in Section 2.3, during representation learning, the model is influenced by the long-tailed distribution, resulting in a biased classifier. During classifier fine-tuning, the classifier is finetuned on balanced data with the backbone frozen. Therefore, it is crucial to explore strategies for matching each expert's different modules during dataset distillation. The experiments are presented in Figure 6b. \"B\", \"C\", and \"W\" refer to matching with the representation backbone, the classifier layer, and the whole model, respectively. \"B/C\" refers to matching the student backbone with the representation expert's backbone, and the student classifier layer with the classification expert's classifier." + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 313, + 677, + 554, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 677, + 554, + 713 + ], + "spans": [ + { + "bbox": [ + 313, + 677, + 554, + 713 + ], + "type": "text", + "content": "As shown in Figure 6b, the best performance is achieved by matching the backbone with representation experts and the classifier layer with classification experts. This outcome" + } + ] + } + ], + "index": 14 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 318, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 318, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 318, + 757 + ], + "type": "text", + "content": "30614" + } + ] + } + ], + "index": 15 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 5 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 56, + 69, + 147, + 158 + ], + "blocks": [ + { + "bbox": [ + 56, + 69, + 147, + 158 + ], + "lines": [ + { + "bbox": [ + 56, + 69, + 147, + 158 + ], + "spans": [ + { + "bbox": [ + 56, + 69, + 147, + 158 + ], + "type": "image", + "image_path": "529cc07d02758511f59812e643dd0d6ac10f6aad8771de281648807991f4d91e.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 64, + 159, + 145, + 178 + ], + "lines": [ + { + "bbox": [ + 64, + 159, + 145, + 178 + ], + "spans": [ + { + "bbox": [ + 64, + 159, + 145, + 178 + ], + "type": "text", + "content": "(a) Effect of the different components" + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "type": "image", + "bbox": [ + 154, + 70, + 249, + 158 + ], + "blocks": [ + { + "bbox": [ + 154, + 70, + 249, + 158 + ], + "lines": [ + { + "bbox": [ + 154, + 70, + 249, + 158 + ], + "spans": [ + { + "bbox": [ + 154, + 70, + 249, + 158 + ], + "type": "image", + "image_path": "1b5defdd0fd505dd03a2935766ca3c151a696f92257a7fe974a217ebcf3bb77d.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 159, + 159, + 248, + 178 + ], + "lines": [ + { + "bbox": [ + 159, + 159, + 248, + 178 + ], + "spans": [ + { + "bbox": [ + 159, + 159, + 248, + 178 + ], + "type": "text", + "content": "(b) Comparison of different matching strategies" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_caption" + } + ], + "index": 2 + }, + { + "type": "image", + "bbox": [ + 255, + 72, + 349, + 158 + ], + "blocks": [ + { + "bbox": [ + 255, + 72, + 349, + 158 + ], + "lines": [ + { + "bbox": [ + 255, + 72, + 349, + 158 + ], + "spans": [ + { + "bbox": [ + 255, + 72, + 349, + 158 + ], + "type": "image", + "image_path": "44073d57178891c1d5207b78cda063e4cd9a0b4d5ee901ffa1069df2a282508e.jpg" + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 257, + 159, + 348, + 178 + ], + "lines": [ + { + "bbox": [ + 257, + 159, + 348, + 178 + ], + "spans": [ + { + "bbox": [ + 257, + 159, + 348, + 178 + ], + "type": "text", + "content": "(c) Comparison of matching strategies on low " + }, + { + "bbox": [ + 257, + 159, + 348, + 178 + ], + "type": "inline_equation", + "content": "\\beta" + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 55, + 180, + 555, + 213 + ], + "lines": [ + { + "bbox": [ + 55, + 180, + 555, + 213 + ], + "spans": [ + { + "bbox": [ + 55, + 180, + 555, + 213 + ], + "type": "text", + "content": "Figure 6. (a) Ablation study on the proposed components, Distribution-agnostic Matching and Expert Decoupling. (b) Ablation of different matching strategies with expert model layers. (c) & (d) Comparisons of matching strategies for long-tailed dataset distillation under low imbalance factor (c) and high imbalance factor (d). (e) Ablation study on hyperparameters " + }, + { + "bbox": [ + 55, + 180, + 555, + 213 + ], + "type": "inline_equation", + "content": "\\lambda_{\\mathrm{rep}}" + }, + { + "bbox": [ + 55, + 180, + 555, + 213 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 55, + 180, + 555, + 213 + ], + "type": "inline_equation", + "content": "\\lambda_{\\mathrm{cls}}" + }, + { + "bbox": [ + 55, + 180, + 555, + 213 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "image_caption" + } + ], + "index": 4 + }, + { + "type": "image", + "bbox": [ + 354, + 71, + 447, + 158 + ], + "blocks": [ + { + "bbox": [ + 354, + 71, + 447, + 158 + ], + "lines": [ + { + "bbox": [ + 354, + 71, + 447, + 158 + ], + "spans": [ + { + "bbox": [ + 354, + 71, + 447, + 158 + ], + "type": "image", + "image_path": "ae100f1ef62c71dc04837cb5af135c6b917e85e27cc74f4a7d39d6e05eece910.jpg" + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 356, + 159, + 447, + 178 + ], + "lines": [ + { + "bbox": [ + 356, + 159, + 447, + 178 + ], + "spans": [ + { + "bbox": [ + 356, + 159, + 447, + 178 + ], + "type": "text", + "content": "(d) Comparison of matching strategies on high " + }, + { + "bbox": [ + 356, + 159, + 447, + 178 + ], + "type": "inline_equation", + "content": "\\beta" + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "image_caption" + } + ], + "index": 6 + }, + { + "type": "image", + "bbox": [ + 452, + 72, + 546, + 158 + ], + "blocks": [ + { + "bbox": [ + 452, + 72, + 546, + 158 + ], + "lines": [ + { + "bbox": [ + 452, + 72, + 546, + 158 + ], + "spans": [ + { + "bbox": [ + 452, + 72, + 546, + 158 + ], + "type": "image", + "image_path": "9ddd80bffe9452b76601df8cf68b7a5055056dc3fedce9876380f149ef8e2f7c.jpg" + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 453, + 159, + 548, + 178 + ], + "lines": [ + { + "bbox": [ + 453, + 159, + 548, + 178 + ], + "spans": [ + { + "bbox": [ + 453, + 159, + 548, + 178 + ], + "type": "text", + "content": "(e) Effect of the relative scale of " + }, + { + "bbox": [ + 453, + 159, + 548, + 178 + ], + "type": "inline_equation", + "content": "\\lambda_{\\mathrm{rep}}" + }, + { + "bbox": [ + 453, + 159, + 548, + 178 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 453, + 159, + 548, + 178 + ], + "type": "inline_equation", + "content": "\\lambda_{\\mathrm{cls}}" + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "image_caption" + } + ], + "index": 8 + }, + { + "bbox": [ + 54, + 220, + 295, + 292 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 54, + 220, + 295, + 292 + ], + "spans": [ + { + "bbox": [ + 54, + 220, + 295, + 292 + ], + "type": "text", + "content": "aligns with the principles of decoupled learning. Since the classifier layer of representation experts is heavily affected by the long-tailed dataset, including it in the matching process causes a performance drop. In addition, matching the entire classifier expert negatively impacts performance, as only the classifier layer is trained at this stage." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 54, + 293, + 295, + 449 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 54, + 293, + 295, + 449 + ], + "spans": [ + { + "bbox": [ + 54, + 293, + 295, + 449 + ], + "type": "text", + "content": "Effect of different matching strategies. When distilling a long-tailed dataset, various intuitive strategies can be employed with decoupling trained experts. We explored several strategies, including sequential matching, classification expert matching, oversampled dataset, and joint matching. Sequential matching first matches representation experts, followed by classification experts. Classification expert matching only matches with the classification expert. The oversampled dataset approach involves oversampling the tail classes before distilling the dataset rather than using the decoupling strategy. Joint matching, used in our Expert Decoupling, involves matching both experts simultaneously in the outer loop of the matching." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 55, + 449, + 295, + 628 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 449, + 295, + 628 + ], + "spans": [ + { + "bbox": [ + 55, + 449, + 295, + 628 + ], + "type": "text", + "content": "As shown in Figures 6c and 6d, joint matching (joint), which we use in Expert Decoupling, achieves the best overall performance. Sequential matching (sequential) and directly matching with classification experts (direct) perform poorly when the imbalance factor is low, meaning the data is \"not very imbalanced.\" Sequential matching suffers due to the trajectory gap between the two types of experts. Classification expert matching is less effective when the imbalance factor is low because the representation layers learn more information in such scenarios, and solely matching the classification layers fails to capture sufficient information. As depicted in Figure 6d, the oversampled dataset strategy (oversample) performs poorly when the imbalance factor is high. This finding aligns with previous work [7], which indicates that oversampling leads to model overfitting." + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 55, + 630, + 295, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 630, + 295, + 714 + ], + "spans": [ + { + "bbox": [ + 55, + 630, + 295, + 714 + ], + "type": "text", + "content": "Effect of hyperparameters. The most important hyperparameters in our method are " + }, + { + "bbox": [ + 55, + 630, + 295, + 714 + ], + "type": "inline_equation", + "content": "\\lambda_{\\mathrm{rep}}" + }, + { + "bbox": [ + 55, + 630, + 295, + 714 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 55, + 630, + 295, + 714 + ], + "type": "inline_equation", + "content": "\\lambda_{\\mathrm{cls}}" + }, + { + "bbox": [ + 55, + 630, + 295, + 714 + ], + "type": "text", + "content": " in Equation 6. Together, they regulate the balance between the matching representation expert and the classification expert. The setting of " + }, + { + "bbox": [ + 55, + 630, + 295, + 714 + ], + "type": "inline_equation", + "content": "\\lambda_{\\mathrm{rep}}" + }, + { + "bbox": [ + 55, + 630, + 295, + 714 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 55, + 630, + 295, + 714 + ], + "type": "inline_equation", + "content": "\\lambda_{\\mathrm{cls}}" + }, + { + "bbox": [ + 55, + 630, + 295, + 714 + ], + "type": "text", + "content": " depends on the imbalance factors. From Figure 6e, we have the following observations. For a \"more imbalanced\" dataset (e.g., " + }, + { + "bbox": [ + 55, + 630, + 295, + 714 + ], + "type": "inline_equation", + "content": "\\beta = 200" + }, + { + "bbox": [ + 55, + 630, + 295, + 714 + ], + "type": "text", + "content": "), lower " + }, + { + "bbox": [ + 55, + 630, + 295, + 714 + ], + "type": "inline_equation", + "content": "\\lambda_{\\mathrm{rep}}" + }, + { + "bbox": [ + 55, + 630, + 295, + 714 + ], + "type": "text", + "content": " and higher " + }, + { + "bbox": [ + 55, + 630, + 295, + 714 + ], + "type": "inline_equation", + "content": "\\lambda_{\\mathrm{cls}}" + } + ] + } + ], + "index": 14 + }, + { + "type": "table", + "bbox": [ + 318, + 213, + 550, + 265 + ], + "blocks": [ + { + "bbox": [ + 318, + 213, + 550, + 265 + ], + "lines": [ + { + "bbox": [ + 318, + 213, + 550, + 265 + ], + "spans": [ + { + "bbox": [ + 318, + 213, + 550, + 265 + ], + "type": "table", + "html": "
Hyperparametersβ = 10Hyperparametersβ = 200
λrep = 1, λcls = 0.270.5λrep = 0.2, λcls = 162.4
λrep = 1, λcls = 0.468.9λrep = 0.4, λcls = 161.7
λrep = 1, λcls = 0.668.0λrep = 0.6, λcls = 161.3
", + "image_path": "4082d0192408e64b89120b01fba7be8085af298eeb68590bcae9a4a7271bd612.jpg" + } + ] + } + ], + "index": 15, + "angle": 0, + "type": "table_body" + } + ], + "index": 15 + }, + { + "bbox": [ + 316, + 266, + 552, + 277 + ], + "lines": [ + { + "bbox": [ + 316, + 266, + 552, + 277 + ], + "spans": [ + { + "bbox": [ + 316, + 266, + 552, + 277 + ], + "type": "text", + "content": "Table 5. Effect of " + }, + { + "bbox": [ + 316, + 266, + 552, + 277 + ], + "type": "inline_equation", + "content": "{\\lambda }_{\\text{rep }}" + }, + { + "bbox": [ + 316, + 266, + 552, + 277 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 316, + 266, + 552, + 277 + ], + "type": "inline_equation", + "content": "{\\lambda }_{\\text{cls }}" + }, + { + "bbox": [ + 316, + 266, + 552, + 277 + ], + "type": "text", + "content": " under different imbalance factors." + } + ] + } + ], + "index": 16, + "angle": 0, + "type": "text" + }, + { + "bbox": [ + 313, + 288, + 553, + 372 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 288, + 553, + 372 + ], + "spans": [ + { + "bbox": [ + 313, + 288, + 553, + 372 + ], + "type": "text", + "content": "are preferred. For \"less imbalanced\" datasets (e.g., " + }, + { + "bbox": [ + 313, + 288, + 553, + 372 + ], + "type": "inline_equation", + "content": "\\beta = 10" + }, + { + "bbox": [ + 313, + 288, + 553, + 372 + ], + "type": "text", + "content": "), higher " + }, + { + "bbox": [ + 313, + 288, + 553, + 372 + ], + "type": "inline_equation", + "content": "\\lambda_{\\mathrm{rep}}" + }, + { + "bbox": [ + 313, + 288, + 553, + 372 + ], + "type": "text", + "content": " and lower " + }, + { + "bbox": [ + 313, + 288, + 553, + 372 + ], + "type": "inline_equation", + "content": "\\lambda_{\\mathrm{cls}}" + }, + { + "bbox": [ + 313, + 288, + 553, + 372 + ], + "type": "text", + "content": " are chosen. This observation is consistent with our experiments in Figure 6c and 6d, where only matching classifier experts perform well when " + }, + { + "bbox": [ + 313, + 288, + 553, + 372 + ], + "type": "inline_equation", + "content": "\\beta" + }, + { + "bbox": [ + 313, + 288, + 553, + 372 + ], + "type": "text", + "content": " is high. This is because the representation layers learn more information when " + }, + { + "bbox": [ + 313, + 288, + 553, + 372 + ], + "type": "inline_equation", + "content": "\\beta" + }, + { + "bbox": [ + 313, + 288, + 553, + 372 + ], + "type": "text", + "content": " is low, and increasing " + }, + { + "bbox": [ + 313, + 288, + 553, + 372 + ], + "type": "inline_equation", + "content": "\\lambda_{\\mathrm{rep}}" + }, + { + "bbox": [ + 313, + 288, + 553, + 372 + ], + "type": "text", + "content": " leverages information from the representation layers." + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 313, + 373, + 554, + 445 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 373, + 554, + 445 + ], + "spans": [ + { + "bbox": [ + 313, + 373, + 554, + 445 + ], + "type": "text", + "content": "We further analyze the sensitivity of the hyperparameters in Table 5. We can observe a minor change in the performance as we change hyperparameters rapidly. When the relative magnitude of the two hyperparameters is preserved, their absolute values do not affect the distillation performance much, suggesting the robustness of our method." + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 313, + 454, + 527, + 467 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 454, + 527, + 467 + ], + "spans": [ + { + "bbox": [ + 313, + 454, + 527, + 467 + ], + "type": "text", + "content": "3.4. Long-tailed Dataset Distillation Analysis" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 313, + 472, + 554, + 581 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 472, + 554, + 581 + ], + "spans": [ + { + "bbox": [ + 313, + 472, + 554, + 581 + ], + "type": "text", + "content": "Effect of imbalance in long-tailed dataset distillation. In this part, we empirically show that the performance drop in long-tailed dataset distillation is due to the data imbalance instead of the total sample number reduction of the target dataset. We compare the performance of using regular dataset distillation methods for CIFAR-10-LT (with " + }, + { + "bbox": [ + 313, + 472, + 554, + 581 + ], + "type": "inline_equation", + "content": "\\beta = 200" + }, + { + "bbox": [ + 313, + 472, + 554, + 581 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 313, + 472, + 554, + 581 + ], + "type": "inline_equation", + "content": "\\mathrm{IPC} = 50" + }, + { + "bbox": [ + 313, + 472, + 554, + 581 + ], + "type": "text", + "content": ") and a subset of CIFAR-10 that contains the same amount of data but in a uniform distribution. The total number of samples is 11203 under this setting." + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 313, + 582, + 554, + 676 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 582, + 554, + 676 + ], + "spans": [ + { + "bbox": [ + 313, + 582, + 554, + 676 + ], + "type": "text", + "content": "As we can observe in Figure 7a, the performance gap between distilling a long-tailed dataset and a balanced subset can be as large as " + }, + { + "bbox": [ + 313, + 582, + 554, + 676 + ], + "type": "inline_equation", + "content": "28.6\\%" + }, + { + "bbox": [ + 313, + 582, + 554, + 676 + ], + "type": "text", + "content": ", even though the two target datasets have the same total amount of samples. Thus, we conclude that it is the imbalanced distribution in the target dataset that prevents effective distillation. Based on this observation, we delved into it and proposed methods to alleviate the imbalanced distribution's negative effects." + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 313, + 677, + 554, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 677, + 554, + 713 + ], + "spans": [ + { + "bbox": [ + 313, + 677, + 554, + 713 + ], + "type": "text", + "content": "Handling classes with extremely small number of samples. In long-tailed dataset distillation, a scenario exists where the number of samples in tail classes is even less than" + } + ] + } + ], + "index": 22 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "text", + "content": "30615" + } + ] + } + ], + "index": 23 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 6 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 56, + 68, + 175, + 191 + ], + "blocks": [ + { + "bbox": [ + 56, + 68, + 175, + 191 + ], + "lines": [ + { + "bbox": [ + 56, + 68, + 175, + 191 + ], + "spans": [ + { + "bbox": [ + 56, + 68, + 175, + 191 + ], + "type": "image", + "image_path": "581003bfe0f5ddb2f45a55988c0be8e6d4cfccce6d9fbc5399d7bd723efd2d00.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 54, + 196, + 555, + 231 + ], + "lines": [ + { + "bbox": [ + 54, + 196, + 555, + 231 + ], + "spans": [ + { + "bbox": [ + 54, + 196, + 555, + 231 + ], + "type": "text", + "content": "Figure 7. (a) Comparison of DD performance between a uniform dataset and a long-tailed dataset with the same number of samples. (b) Comparison of strategies to handle an extremely low number of samples. (c) Comparison of class-wise accuracy between DATM and our method. (d) Comparison between balanced IPC and long-tailed distributed IPC." + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "type": "image", + "bbox": [ + 181, + 68, + 299, + 191 + ], + "blocks": [ + { + "bbox": [ + 181, + 68, + 299, + 191 + ], + "lines": [ + { + "bbox": [ + 181, + 68, + 299, + 191 + ], + "spans": [ + { + "bbox": [ + 181, + 68, + 299, + 191 + ], + "type": "image", + "image_path": "7e682a10346c68bccbfe6ffe0fcecf6c25ce31d9c93a719b7cdee4d840fb6be6.jpg" + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_body" + } + ], + "index": 1 + }, + { + "type": "image", + "bbox": [ + 304, + 68, + 422, + 191 + ], + "blocks": [ + { + "bbox": [ + 304, + 68, + 422, + 191 + ], + "lines": [ + { + "bbox": [ + 304, + 68, + 422, + 191 + ], + "spans": [ + { + "bbox": [ + 304, + 68, + 422, + 191 + ], + "type": "image", + "image_path": "62595531ac9c8647f4f5a2775118e1eec0697b9966ead3c12928d041a1f7296b.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + } + ], + "index": 2 + }, + { + "type": "image", + "bbox": [ + 427, + 68, + 547, + 191 + ], + "blocks": [ + { + "bbox": [ + 427, + 68, + 547, + 191 + ], + "lines": [ + { + "bbox": [ + 427, + 68, + 547, + 191 + ], + "spans": [ + { + "bbox": [ + 427, + 68, + 547, + 191 + ], + "type": "image", + "image_path": "a5f1b6806fffbc90a7816ede88a07a302d6c98b5f14789fbc6c4a648bd67e721.jpg" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_body" + } + ], + "index": 3 + }, + { + "bbox": [ + 54, + 236, + 294, + 356 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 54, + 236, + 294, + 356 + ], + "spans": [ + { + "bbox": [ + 54, + 236, + 294, + 356 + ], + "type": "text", + "content": "the required number of images per class. In this case, we employ and compare three strategies: 1) Inherent strategy, which keeps the number of samples the same in the synthetic dataset; 2) Oversampling, which duplicates the samples of the tail classes before distillation; 3) Noise generation, which randomly generates Gaussian noises to supplement the insufficient samples for tail classes. From figure 7b, we see that the inherent strategy provides a comparatively lower accuracy, while oversampling and noise generation achieve better and similar performance." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 54, + 357, + 295, + 453 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 54, + 357, + 295, + 453 + ], + "spans": [ + { + "bbox": [ + 54, + 357, + 295, + 453 + ], + "type": "text", + "content": "Class-wise accuracy for long-tailed dataset. We also provide class-wise accuracy to illustrate the impact of the imbalanced data, and we show that our method can effectively improve the accuracy of tail classes. As shown in Figure 7c, DATM performs poorly on tail classes, and even achieves near-zero performance on the last class. On the contrary, our method can successfully boost the tail class accuracy while reserving performance on head classes." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 54, + 454, + 295, + 611 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 54, + 454, + 295, + 611 + ], + "spans": [ + { + "bbox": [ + 54, + 454, + 295, + 611 + ], + "type": "text", + "content": "Choice of distilled dataset distribution. Should we distill a long-tailed dataset into another long-tailed dataset or a balanced one? We compared the performance of distilling the target dataset into either a similarly distributed long-tailed synthetic dataset or a balanced synthetic dataset using our method and DREAM [28]. For the long-tailed synthetic dataset, we maintained the same total number of samples as in the synthetic dataset with IPC=50. As shown in Figure 7d, both methods demonstrate that distilling the dataset into a balanced synthetic dataset provides better performance than distilling it into a long-tailed one. This result is reasonable since the model trained on the synthetic long-tailed dataset will be affected by the long-tailed distribution." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 55, + 624, + 142, + 637 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 624, + 142, + 637 + ], + "spans": [ + { + "bbox": [ + 55, + 624, + 142, + 637 + ], + "type": "text", + "content": "4. Related Work" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 55, + 646, + 168, + 658 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 646, + 168, + 658 + ], + "spans": [ + { + "bbox": [ + 55, + 646, + 168, + 658 + ], + "type": "text", + "content": "4.1. Dataset Distillation" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 54, + 665, + 295, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 54, + 665, + 295, + 714 + ], + "spans": [ + { + "bbox": [ + 54, + 665, + 295, + 714 + ], + "type": "text", + "content": "Dataset distillation [43] was first formulated by Wang et al. The goal of dataset distillation is to synthesize a small dataset, such that models trained on it have good performance on the original dataset. The following works can be" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 313, + 236, + 555, + 333 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 236, + 555, + 333 + ], + "spans": [ + { + "bbox": [ + 313, + 236, + 555, + 333 + ], + "type": "text", + "content": "divided into kernel ridge regression-based methods [30, 32, 50], gradient matching methods [6, 13, 28, 42, 46], representation matching methods [38, 47], and generative-based methods [12, 40]. Despite the advantages of current methods, they only focus on uniformly distributed datasets, ignoring the prevalent usage of long-tailed datasets in practice. Therefore, we propose to address the long-tailed dataset distillation problem in this paper." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 313, + 337, + 450, + 350 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 337, + 450, + 350 + ], + "spans": [ + { + "bbox": [ + 313, + 337, + 450, + 350 + ], + "type": "text", + "content": "4.2. Long-tailed Recognition" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 313, + 353, + 556, + 510 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 353, + 556, + 510 + ], + "spans": [ + { + "bbox": [ + 313, + 353, + 556, + 510 + ], + "type": "text", + "content": "Under real-world settings, data often tends to follow a long-tailed distribution [2]. Models trained on such datasets usually exhibit good performance on head classes but low accuracy on tail classes. A straightforward way is to preprocess the data, using oversampling [14, 18, 33], undersampling [3, 14], or data augmentation [5, 11] to make it balancedly distributed. An alternative way is weight balancing [2, 26, 34], which re-weights the layers during training to force the parameters to update unbiasedly. Decoupling methods [20] first train the model to obtain the features of the data and then fine-tune the model on the class-balanced data. In our methods, we use the decoupling method to train the expert networks during the distillation." + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 313, + 517, + 388, + 529 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 517, + 388, + 529 + ], + "spans": [ + { + "bbox": [ + 313, + 517, + 388, + 529 + ], + "type": "text", + "content": "5. Conclusion" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 313, + 534, + 555, + 665 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 534, + 555, + 665 + ], + "spans": [ + { + "bbox": [ + 313, + 534, + 555, + 665 + ], + "type": "text", + "content": "We introduce long-tailed dataset distillation, a novel yet challenging task. We find that existing DD methods fail when applied to long-tailed datasets and encounter two main problems: weight distribution mismatch between the student and the expert; and suboptimal performance of the expert on the tail classes. To mitigate these issues, we propose two strategies: Distribution-agnostic Matching and Expert Decoupling. Experimental results over four long-tailed datasets demonstrate the effectiveness of our proposed method. As the pioneering work in this research area, we are the first to effectively distill long-tailed datasets." + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 313, + 665, + 554, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 665, + 554, + 713 + ], + "spans": [ + { + "bbox": [ + 313, + 665, + 554, + 713 + ], + "type": "text", + "content": "Acknowledgments: This research is supported by NSF IIS-2525840, ECCS-2123521 and Cisco Research unrestricted gift. This article solely reflects the opinions and conclusions of its authors and not the funding agencies." + } + ] + } + ], + "index": 16 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "type": "text", + "content": "30616" + } + ] + } + ], + "index": 17 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 7 + }, + { + "para_blocks": [ + { + "bbox": [ + 56, + 71, + 115, + 83 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 71, + 115, + 83 + ], + "spans": [ + { + "bbox": [ + 56, + 71, + 115, + 83 + ], + "type": "text", + "content": "References" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 57, + 91, + 294, + 712 + ], + "type": "list", + "angle": 0, + "index": 18, + "blocks": [ + { + "bbox": [ + 61, + 91, + 294, + 122 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 61, + 91, + 294, + 122 + ], + "spans": [ + { + "bbox": [ + 61, + 91, + 294, + 122 + ], + "type": "text", + "content": "[1] Shaden Alshammari, Yuxiong Wang, Deva Ramanan, and Shu Kong. Long-tailed recognition via weight balancing. In CVPR, 2022. 2" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 61, + 125, + 294, + 157 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 61, + 125, + 294, + 157 + ], + "spans": [ + { + "bbox": [ + 61, + 125, + 294, + 157 + ], + "type": "text", + "content": "[2] Shaden Alshammari, Yu-Xiong Wang, Deva Ramanan, and Shu Kong. Long-tailed recognition via weight balancing. In CVPR, 2022. 8" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 62, + 159, + 294, + 191 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 159, + 294, + 191 + ], + "spans": [ + { + "bbox": [ + 62, + 159, + 294, + 191 + ], + "type": "text", + "content": "[3] Mateusz Buda, Atsuto Maki, and Maciej A Mazurowski. A systematic study of the class imbalance problem in convolutional neural networks. Neural networks, 2018. 8" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 62, + 194, + 294, + 225 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 194, + 294, + 225 + ], + "spans": [ + { + "bbox": [ + 62, + 194, + 294, + 225 + ], + "type": "text", + "content": "[4] George Cazenavette, Tongzhou Wang, Antonio Torralba, Alexei A Efros, and Jun-Yan Zhu. Dataset distillation by matching training trajectories. In CVPR, 2022. 1, 2, 3, 5, 6" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 62, + 228, + 294, + 271 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 228, + 294, + 271 + ], + "spans": [ + { + "bbox": [ + 62, + 228, + 294, + 271 + ], + "type": "text", + "content": "[5] Hsin-Ping Chou, Shih-Chieh Chang, Jia-Yu Pan, Wei Wei, and Da-Cheng Juan. Remix: rebalanced mixup. In Computer Vision-ECCV 2020 Workshops: Glasgow, UK, August 23-28, 2020, Proceedings, Part VI 16, 2020. 8" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 62, + 273, + 294, + 304 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 273, + 294, + 304 + ], + "spans": [ + { + "bbox": [ + 62, + 273, + 294, + 304 + ], + "type": "text", + "content": "[6] Justin Cui, Ruochen Wang, Si Si, and Cho-Jui Hsieh. Scaling up dataset distillation toImagenet-1k with constant memory. In ICML, 2023. 2, 6, 8, 1" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 62, + 307, + 294, + 338 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 307, + 294, + 338 + ], + "spans": [ + { + "bbox": [ + 62, + 307, + 294, + 338 + ], + "type": "text", + "content": "[7] Yin Cui, Menglin Jia, Tsung-Yi Lin, Yang Song, and Serge Belongie. Class-balanced loss based on effective number of samples. In CVPR, 2019. 5, 7" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 62, + 341, + 294, + 372 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 341, + 294, + 372 + ], + "spans": [ + { + "bbox": [ + 62, + 341, + 294, + 372 + ], + "type": "text", + "content": "[8] Jia Deng, Wei Dong, Richard Socher, Li-Jia Li, Kai Li, and Li Fei-Fei. Imagenet: A large-scale hierarchical image database. In CVPR, 2009. 5" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 62, + 375, + 294, + 418 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 375, + 294, + 418 + ], + "spans": [ + { + "bbox": [ + 62, + 375, + 294, + 418 + ], + "type": "text", + "content": "[9] Fei Du, Peng Yang, Qi Jia, Fengtao Nan, Xiaoting Chen, and Yun Yang. Global and local mixture consistency cumulative learning for long-tailed visual recognitions. In CVPR, 2023. 5" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 57, + 420, + 294, + 452 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 420, + 294, + 452 + ], + "spans": [ + { + "bbox": [ + 57, + 420, + 294, + 452 + ], + "type": "text", + "content": "[10] Jiawei Du, Yidi Jiang, Vincent YF Tan, Joey Tianyi Zhou, and Haizhou Li. Minimizing the accumulated trajectory error to improve dataset distillation. In CVPR, 2023. 2" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 57, + 454, + 294, + 475 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 454, + 294, + 475 + ], + "spans": [ + { + "bbox": [ + 57, + 454, + 294, + 475 + ], + "type": "text", + "content": "[11] Spyros Gidaris and Nikos Komodakis. Dynamic few-shot visual learning without forgetting. In CVPR, 2018. 8" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 57, + 477, + 294, + 520 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 477, + 294, + 520 + ], + "spans": [ + { + "bbox": [ + 57, + 477, + 294, + 520 + ], + "type": "text", + "content": "[12] Jianyang Gu, Saeed Vahidian, Vyacheslav Kungurtsev, Haonan Wang, Wei Jiang, Yang You, and Yiran Chen. Efficient dataset distillation via minimax diffusion. In CVPR, 2024. 1, 5, 8" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 57, + 522, + 294, + 565 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 522, + 294, + 565 + ], + "spans": [ + { + "bbox": [ + 57, + 522, + 294, + 565 + ], + "type": "text", + "content": "[13] Ziyao Guo, Kai Wang, George Cazenavette, Hui Li, Kaipeng Zhang, and Yang You. Towards lossless dataset distillation via difficulty-aligned trajectory matching. In ICLR, 2024. 1, 2, 3, 4, 5, 6, 8" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 57, + 567, + 294, + 610 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 567, + 294, + 610 + ], + "spans": [ + { + "bbox": [ + 57, + 567, + 294, + 610 + ], + "type": "text", + "content": "[14] Guo Haixiang, Li Yijing, Jennifer Shang, Gu Mingyun, Huang Yuanyue, and Gong Bing. Learning from class-imbalanced data: Review of methods and applications. Expert systems with applications, 2017. 8" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 57, + 613, + 294, + 644 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 613, + 294, + 644 + ], + "spans": [ + { + "bbox": [ + 57, + 613, + 294, + 644 + ], + "type": "text", + "content": "[15] Geoffrey Hinton, Oriol Vinyals, and Jeff Dean. Distilling the knowledge in a neural network. arXiv preprint arXiv:1503.02531, 2015. 1" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 57, + 647, + 294, + 689 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 647, + 294, + 689 + ], + "spans": [ + { + "bbox": [ + 57, + 647, + 294, + 689 + ], + "type": "text", + "content": "[16] Geoffrey E Hinton, Nitish Srivastava, Alex Krizhevsky, Ilya Sutskever, and Ruslan R Salakhutdinov. Improving neural networks by preventing co-adaptation of feature detectors. arXiv preprint arXiv:1207.0580, 2012. 1" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 57, + 692, + 294, + 712 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 692, + 294, + 712 + ], + "spans": [ + { + "bbox": [ + 57, + 692, + 294, + 712 + ], + "type": "text", + "content": "[17] Rishabh Iyer, Ninad Khargoankar, Jeff Bilmes, and Himanshu Asanani. Submodular combinatorial information mea" + } + ] + } + ], + "index": 17 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 317, + 74, + 553, + 712 + ], + "type": "list", + "angle": 0, + "index": 34, + "blocks": [ + { + "bbox": [ + 335, + 74, + 553, + 95 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 335, + 74, + 553, + 95 + ], + "spans": [ + { + "bbox": [ + 335, + 74, + 553, + 95 + ], + "type": "text", + "content": "sures with applications in machine learning. In Algorithmic Learning Theory, 2021. 5, 6" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 317, + 98, + 553, + 140 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 98, + 553, + 140 + ], + "spans": [ + { + "bbox": [ + 317, + 98, + 553, + 140 + ], + "type": "text", + "content": "[18] Andrew Janowczyk and Anant Madabhushi. Deep learning for digital pathology image analysis: A comprehensive tutorial with selected use cases. Journal of pathology informatics, 2016. 8" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 317, + 143, + 553, + 186 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 143, + 553, + 186 + ], + "spans": [ + { + "bbox": [ + 317, + 143, + 553, + 186 + ], + "type": "text", + "content": "[19] Lie Ju, Xin Wang, Lin Wang, Tongliang Liu, Xin Zhao, Tom Drummond, Dwarikanath Mahapatra, and Zongyuan Ge. Relational subsets knowledge distillation for long-tailed retinal diseases recognition. In MICCAI, 2021. 1" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 317, + 189, + 553, + 232 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 189, + 553, + 232 + ], + "spans": [ + { + "bbox": [ + 317, + 189, + 553, + 232 + ], + "type": "text", + "content": "[20] Bingyi Kang, Saining Xie, Marcus Rohrbach, Zhicheng Yan, Albert Gordo, Jiashi Feng, and Yannis Kalantidis. Decoupling representation and classifier for long-tailed recognition. In ICLR, 2020. 2, 4, 5, 8" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 317, + 235, + 553, + 277 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 235, + 553, + 277 + ], + "spans": [ + { + "bbox": [ + 317, + 235, + 553, + 277 + ], + "type": "text", + "content": "[21] Jang-Hyun Kim, Jinuk Kim, Seong Joon Oh, Sangdoo Yun, Hwanjun Song, Joonhyun Jeong, Jung-Woo Ha, and Hyun Oh Song. Dataset condensation via efficient synthetic-data parameterization. In ICML, 2022. 2" + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 317, + 280, + 553, + 301 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 280, + 553, + 301 + ], + "spans": [ + { + "bbox": [ + 317, + 280, + 553, + 301 + ], + "type": "text", + "content": "[22] Alex Krizhevsky. Learning multiple layers of features from tiny images. University of Toronto, 2012. 1, 5" + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 317, + 304, + 553, + 335 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 304, + 553, + 335 + ], + "spans": [ + { + "bbox": [ + 317, + 304, + 553, + 335 + ], + "type": "text", + "content": "[23] Alex Krizhevsky, Ilya Sutskever, and Geoffrey E Hinton. Imagenet classification with deep convolutional neural networks. Communications of the ACM, 2017. 2, 1" + } + ] + } + ], + "index": 25 + }, + { + "bbox": [ + 317, + 338, + 553, + 360 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 338, + 553, + 360 + ], + "spans": [ + { + "bbox": [ + 317, + 338, + 553, + 360 + ], + "type": "text", + "content": "[24] Yann Le and Xuan Yang. Tiny imagenet visual recognition challenge. CS 231N, 2015. 1" + } + ] + } + ], + "index": 26 + }, + { + "bbox": [ + 317, + 362, + 553, + 494 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 362, + 553, + 494 + ], + "spans": [ + { + "bbox": [ + 317, + 362, + 553, + 494 + ], + "type": "text", + "content": "[25] Zekai Li, Xinhao Zhong, Zhiyuan Liang, Yuhao Zhou, Mingjia Shi, Ziqiao Wang, Wangbo Zhao, Xuanlei Zhao, Haonan Wang, Ziheng Qin, Dai Liu, Kaipeng Zhang, Tianyi Zhou, Zheng Zhu, Kun Wang, Guang Li, Junhao Zhang, Jiawei Liu, Yiran Huang, Lingjuan Lyu, Jiancheng Lv, Yaochu Jin, Zeynep Akata, Jindong Gu, Rama Vedantam, Mike Shou, Zhiwei Deng, Yan Yan, Yuzhang Shang, George Cazenavette, Xindi Wu, Justin Cui, Tianlong Chen, Angela Yao, Manolis Kellis, Konstantinos N. Plataniotis, Bo Zhao, Zhangyang Wang, Yang You, and Kai Wang. Dd-ranking: Rethinking the evaluation of dataset distillation. GitHub repository, 2024. 1" + } + ] + } + ], + "index": 27 + }, + { + "bbox": [ + 317, + 496, + 553, + 527 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 496, + 553, + 527 + ], + "spans": [ + { + "bbox": [ + 317, + 496, + 553, + 527 + ], + "type": "text", + "content": "[26] Tsung-Yi Lin, Priya Goyal, Ross Girshick, Kaiming He, and Piotr Dollar. Focal loss for dense object detection. In ICCV, 2017. 8" + } + ] + } + ], + "index": 28 + }, + { + "bbox": [ + 317, + 530, + 553, + 562 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 530, + 553, + 562 + ], + "spans": [ + { + "bbox": [ + 317, + 530, + 553, + 562 + ], + "type": "text", + "content": "[27] Hanxiao Liu, Karen Simonyan, and Yiming Yang. Darts: Differentiable architecture search. arXiv preprint arXiv:1806.09055, 2018. 1" + } + ] + } + ], + "index": 29 + }, + { + "bbox": [ + 317, + 565, + 553, + 597 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 565, + 553, + 597 + ], + "spans": [ + { + "bbox": [ + 317, + 565, + 553, + 597 + ], + "type": "text", + "content": "[28] Yanqing Liu, Jianyang Gu, Kai Wang, Zheng Zhu, Wei Jiang, and Yang You. Dream: Efficient dataset distillation by representative matching. In ICCV, 2023. 1, 2, 5, 6, 8" + } + ] + } + ], + "index": 30 + }, + { + "bbox": [ + 317, + 600, + 553, + 632 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 600, + 553, + 632 + ], + "spans": [ + { + "bbox": [ + 317, + 600, + 553, + 632 + ], + "type": "text", + "content": "[29] Ziwei Liu, Zhongqi Miao, Xiaohang Zhan, Jiayun Wang, Boqing Gong, and Stella X Yu. Large-scale long-tailed recognition in an open world. In CVPR, 2019. 5" + } + ] + } + ], + "index": 31 + }, + { + "bbox": [ + 317, + 635, + 553, + 667 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 635, + 553, + 667 + ], + "spans": [ + { + "bbox": [ + 317, + 635, + 553, + 667 + ], + "type": "text", + "content": "[30] Noel Loo, Ramin Hasani, Alexander Amini, and Daniela Rus. Efficient dataset distillation using random feature approximation. In NeurIPS, 2022. 1, 8" + } + ] + } + ], + "index": 32 + }, + { + "bbox": [ + 317, + 670, + 553, + 712 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 670, + 553, + 712 + ], + "spans": [ + { + "bbox": [ + 317, + 670, + 553, + 712 + ], + "type": "text", + "content": "[31] Wojciech Masarczyk and Ivona Tautkute. Reducing catastrophic forgetting with learning on synthetic data. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition Workshops, 2020. 1" + } + ] + } + ], + "index": 33 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "text", + "content": "30617" + } + ] + } + ], + "index": 35 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 8 + }, + { + "para_blocks": [ + { + "bbox": [ + 55, + 73, + 297, + 712 + ], + "type": "list", + "angle": 0, + "index": 18, + "blocks": [ + { + "bbox": [ + 56, + 73, + 297, + 106 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 73, + 297, + 106 + ], + "spans": [ + { + "bbox": [ + 56, + 73, + 297, + 106 + ], + "type": "text", + "content": "[32] Timothy Nguyen, Zhourong Chen, and Jaehoon Lee. Dataset meta-learning from kernel ridge-regression. arXiv preprint arXiv:2011.00050, 2020. 1, 8" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 55, + 106, + 297, + 139 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 106, + 297, + 139 + ], + "spans": [ + { + "bbox": [ + 55, + 106, + 297, + 139 + ], + "type": "text", + "content": "[33] Junran Peng, Xingyuan Bu, Ming Sun, Zhaoxiang Zhang, Tieniu Tan, and Junjie Yan. Large-scale object detection in the wild from imbalanced multi-labels. In CVPR, 2020. 8" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 56, + 140, + 296, + 171 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 140, + 296, + 171 + ], + "spans": [ + { + "bbox": [ + 56, + 140, + 296, + 171 + ], + "type": "text", + "content": "[34] Jiawei Ren, Cunjun Yu, Xiao Ma, Haiyu Zhao, Shuai Yi, et al. Balanced meta-softmax for long-tailed visual recognition. In NeurIPS, 2020. 3, 8" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 56, + 172, + 296, + 195 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 172, + 296, + 195 + ], + "spans": [ + { + "bbox": [ + 56, + 172, + 296, + 195 + ], + "type": "text", + "content": "[35] Noveen Sachdeva and Julian McAuley. Data distillation: A survey. arXiv preprint arXiv:2301.04272, 2023. 1" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 56, + 195, + 296, + 228 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 195, + 296, + 228 + ], + "spans": [ + { + "bbox": [ + 56, + 195, + 296, + 228 + ], + "type": "text", + "content": "[36] Ozan Sener and Silvio Savarese. Active learning for convolutional neural networks: A core-set approach. arXiv preprint arXiv:1708.00489, 2017. 5, 6" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 56, + 228, + 296, + 259 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 228, + 296, + 259 + ], + "spans": [ + { + "bbox": [ + 56, + 228, + 296, + 259 + ], + "type": "text", + "content": "[37] Shai Shalev-Shwartz, Yoram Singer, and Nathan Srebro. Pegasos: Primal estimated sub-gradient solver forsvm. In ICML, 2007. 1" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 56, + 261, + 296, + 293 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 261, + 296, + 293 + ], + "spans": [ + { + "bbox": [ + 56, + 261, + 296, + 293 + ], + "type": "text", + "content": "[38] Yuzhang Shang, Zhihang Yuan, and Yan Yan. Mim4dd: Mutual information maximization for dataset distillation. In NeurIPS, 2024. 1, 8" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 56, + 294, + 296, + 337 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 294, + 296, + 337 + ], + "spans": [ + { + "bbox": [ + 56, + 294, + 296, + 337 + ], + "type": "text", + "content": "[39] Shitong Shao, Zeyuan Yin, Muxin Zhou, Xindong Zhang, and Zhiqiang Shen. Generalized large-scale data condensation via various backbone and statistical matching. In CVPR, 2024. 6" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 56, + 338, + 296, + 370 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 338, + 296, + 370 + ], + "spans": [ + { + "bbox": [ + 56, + 338, + 296, + 370 + ], + "type": "text", + "content": "[40] Duo Su, Junjie Hou, Weizhi Gao, Yingjie Tian, and Bowen Tang. D^4: Dataset distillation via disentangled diffusion model. In CVPR, 2024. 1, 8" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 56, + 371, + 296, + 415 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 371, + 296, + 415 + ], + "spans": [ + { + "bbox": [ + 56, + 371, + 296, + 415 + ], + "type": "text", + "content": "[41] Kai Wang, Bo Zhao, Xiangyu Peng, Zheng Zhu, Shuo Yang, Shuo Wang, Guan Huang, Hakan Bilen, Xinchao Wang, and Yang You. Cafe: Learning to condense dataset by aligning features. In CVPR, 2022. 2" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 56, + 415, + 296, + 470 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 415, + 296, + 470 + ], + "spans": [ + { + "bbox": [ + 56, + 415, + 296, + 470 + ], + "type": "text", + "content": "[42] Kai Wang, Zekai Li, Zhi-Qi Cheng, Samir Khaki, Ahmad Sajedi, Ramakrishna Vedantam, Konstantinos N Plataniotis, Alexander Hauptmann, and Yang You. Emphasizing discriminative features for dataset distillation in complex scenarios. arXiv preprint arXiv:2410.17193, 2024. 8" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 56, + 471, + 296, + 503 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 471, + 296, + 503 + ], + "spans": [ + { + "bbox": [ + 56, + 471, + 296, + 503 + ], + "type": "text", + "content": "[43] Tongzhou Wang, Jun-Yan Zhu, Antonio Torralba, and Alexei A Efros. Dataset distillation. arXiv preprint arXiv:1811.10959, 2018. 8" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 56, + 503, + 296, + 536 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 503, + 296, + 536 + ], + "spans": [ + { + "bbox": [ + 56, + 503, + 296, + 536 + ], + "type": "text", + "content": "[44] Zeyuan Yin, Eric Xing, and Zhiqiang Shen. Squeeze, recover and relabel: Dataset condensation at imagenet scale from a new perspective. In NeurIPS, 2023. 6" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 56, + 537, + 296, + 580 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 537, + 296, + 580 + ], + "spans": [ + { + "bbox": [ + 56, + 537, + 296, + 580 + ], + "type": "text", + "content": "[45] Yifan Zhang, Bingyi Kang, Bryan Hooi, Shuicheng Yan, and Jiashi Feng. Deep long-tailed learning: A survey. IEEE transactions on pattern analysis and machine intelligence, 2023. 1" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 56, + 581, + 296, + 613 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 581, + 296, + 613 + ], + "spans": [ + { + "bbox": [ + 56, + 581, + 296, + 613 + ], + "type": "text", + "content": "[46] Bo Zhao, Konda Reddy Mopuri, and Hakan Bilen. Dataset condensation with gradient matching. arXiv preprint arXiv:2006.05929, 2020. 1, 5, 6, 8" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 56, + 614, + 296, + 647 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 614, + 296, + 647 + ], + "spans": [ + { + "bbox": [ + 56, + 614, + 296, + 647 + ], + "type": "text", + "content": "[47] Ganlong Zhao, Guanbin Li, Yipeng Qin, and Yizhou Yu. Improved distribution matching for dataset condensation. In CVPR, 2023. 1, 5, 8" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 56, + 647, + 296, + 680 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 647, + 296, + 680 + ], + "spans": [ + { + "bbox": [ + 56, + 647, + 296, + 680 + ], + "type": "text", + "content": "[48] Zhenghao Zhao, Yuzhang Shang, Junyi Wu, and Yan Yan. Dataset quantization with active learning based adaptive sampling. In ECCV, 2024. 1" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 56, + 681, + 296, + 712 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 681, + 296, + 712 + ], + "spans": [ + { + "bbox": [ + 56, + 681, + 296, + 712 + ], + "type": "text", + "content": "[49] Zhisheng Zhong, Jiequan Cui, Shu Liu, and Jiaya Jia. Improving calibration for long-tailed recognition. In CVPR, 2021. 2" + } + ] + } + ], + "index": 17 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 315, + 73, + 556, + 106 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 315, + 73, + 556, + 106 + ], + "spans": [ + { + "bbox": [ + 315, + 73, + 556, + 106 + ], + "type": "text", + "content": "[50] Yongchao Zhou, Ehsan Nezhadarya, and Jimmy Ba. Dataset distillation using neural feature regression. In NeurIPS, 2022. 1, 8" + } + ] + } + ], + "index": 19 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "type": "text", + "content": "30618" + } + ] + } + ], + "index": 20 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 9 + } + ], + "_backend": "vlm", + "_version_name": "2.6.4" +} \ No newline at end of file diff --git a/2025/Distilling Monocular Foundation Model for Fine-grained Depth Completion/1663a87e-e3f7-4145-9d6a-27f93678241e_content_list.json b/2025/Distilling Monocular Foundation Model for Fine-grained Depth Completion/1663a87e-e3f7-4145-9d6a-27f93678241e_content_list.json new file mode 100644 index 0000000000000000000000000000000000000000..898e33e8f74e79d04fab9d95ab68699ce9afbb1f --- /dev/null +++ b/2025/Distilling Monocular Foundation Model for Fine-grained Depth Completion/1663a87e-e3f7-4145-9d6a-27f93678241e_content_list.json @@ -0,0 +1,2057 @@ +[ + { + "type": "text", + "text": "Distilling Monocular Foundation Model for Fine-grained Depth Completion", + "text_level": 1, + "bbox": [ + 114, + 130, + 880, + 152 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Yingping Liang1 Yutao Hu2 Wenqi Shao3 Ying Fu1† \n1Beijing Institute of Technology 2Key Laboratory of New Generation Artificial Intelligence Technology and Its Interdisciplinary Applications, Southeast University 3Shanghai Al Laboratory {liangyingping, fuying}@bit.edu.cn huyutao@seu.edu.cn weqish@link.cuhk.edu.hk", + "bbox": [ + 99, + 178, + 908, + 252 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/9838f5956b4ecd56b7df0dcbb9f8556c79e9ba628572bfb75045a8d65b7240cf.jpg", + "image_caption": [ + "Figure 1. Depth completion models trained solely with $L_{1}$ loss and sparse ground truth produce incomplete and fragmented depth predictions. Our framework, however, demonstrates significant improvements by distilling knowledge from monocular foundation models and incorporating a scale- and shift-invariant loss (SSI Loss), resulting in more complete and accurate dense depth completion." + ], + "image_footnote": [], + "bbox": [ + 112, + 286, + 888, + 415 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Abstract", + "text_level": 1, + "bbox": [ + 246, + 481, + 326, + 498 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Depth completion involves predicting dense depth maps from sparse LiDAR inputs. However, sparse depth annotations from sensors limit the availability of dense supervision, which is necessary for learning detailed geometric features. In this paper, we propose a two-stage knowledge distillation framework that leverages powerful monocular foundation models to provide dense supervision for depth completion. In the first stage, we introduce a pre-training strategy that generates diverse training data from natural images, which distills geometric knowledge to depth completion. Specifically, we simulate LiDAR scans by utilizing monocular depth and mesh reconstruction, thereby creating training data without requiring ground-truth depth. Besides, monocular depth estimation suffers from inherent scale ambiguity in real-world settings. To address this, in the second stage, we employ a scale- and shift-invariant loss (SSI Loss) to learn real-world scales when fine-tuning on real-world datasets. Our two-stage distillation framework enables depth completion models to harness the strengths of monocular foundation models. Experimental results demonstrate that models trained with our two-stage distillation framework achieve state-of-the-art performance, ranking first place on the KITTI benchmark. Code is available", + "bbox": [ + 86, + 516, + 485, + 864 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "at https://github.com/Sharpiless/DMD3C", + "bbox": [ + 511, + 483, + 879, + 498 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "1. Introduction", + "text_level": 1, + "bbox": [ + 513, + 526, + 645, + 542 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Depth completion is a fundamental task in computer vision, where the goal is to generate dense depth maps from sparse depth measurements. This task is particularly important in applications such as autonomous driving [23, 47, 61], robotics [26, 39, 63], and augmented reality [21, 30].", + "bbox": [ + 511, + 551, + 906, + 628 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Recent approaches [7, 27, 28, 65, 67, 78] leverage deep neural networks to learn from depth data. Some methods [56, 79] also incorporate RGB images to guide the depth completion process. Despite these advancements, state-of-the-art models [50, 57, 70] still face difficulties in capturing fine-grained geometric details, particularly in complex outdoor scenes where depth annotations are sparse [53, 58, 75].", + "bbox": [ + 511, + 628, + 908, + 733 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "The challenge arises from the reliance on sparse ground truth for training. The lack of dense ground truth makes it difficult for models to accurately learn depth completion across an entire scene. Monocular depth estimation [25, 46, 71, 72], on the other hand, has the ability to provide dense depth predictions from a single image. Specifically, state-of-the-art monocular foundation models could generate dense depth maps, containing fine-grained details and relative depth relationships, offering valuable guidance for training depth completion networks.", + "bbox": [ + 511, + 734, + 908, + 883 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "To make full use of the advantages of monocular foun", + "bbox": [ + 532, + 885, + 906, + 901 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "CVF", + "bbox": [ + 106, + 2, + 181, + 42 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore.", + "bbox": [ + 236, + 0, + 810, + 47 + ], + "page_idx": 0 + }, + { + "type": "page_footnote", + "text": "† Corresponding author.", + "bbox": [ + 114, + 887, + 245, + 900 + ], + "page_idx": 0 + }, + { + "type": "page_number", + "text": "22254", + "bbox": [ + 478, + 944, + 519, + 957 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "dation models, we propose a novel two-stage distillation framework to transfer geometric knowledge from monocular foundation models to depth completion networks. Specifically, in the first distillation stage, we generate training data through monocular foundation models and then distill knowledge via the proposed pre-training strategy. Specifically, diverse natural images are used to generate pseudo depth maps. Then, we utilize randomly sampled camera parameters to re-construct the scene with mesh and simulate LiDAR using ray simulation. The generated data trains the depth completion model to learn diverse geometric knowledge from monocular foundation models, enhancing its ability to generalize across different scenes.", + "bbox": [ + 89, + 90, + 480, + 287 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "However, monocular depth estimation suffers from inherent scale ambiguity [16, 81], resulting in depth predictions that vary greatly in scale. Thus, monocular depth alone cannot serve as a reliable basis for real-world depth. To solve this problem, in the second distillation stage, we introduce a scale- and shift-invariant loss (SSI Loss) [46]. Specifically, when fine-tuning on labeled datasets with sparse ground truth, SSI Loss ignores the scale and shift that causes the least loss from the depth prediction and monocular depth, to ensure consistent depth completions aligning monocular depth supervision across varying scales.", + "bbox": [ + 89, + 287, + 480, + 455 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "By integrating these two distillation stages into the training process, our method achieves state-of-the-art performance on the KITTI benchmark. The contributions of this work are summarized as follows:", + "bbox": [ + 89, + 454, + 480, + 513 + ], + "page_idx": 1 + }, + { + "type": "list", + "sub_type": "text", + "list_items": [ + "- We propose a novel two-stage distillation framework to transfer knowledge from monocular foundation models to depth completion models, which enables the learning of fine-grained depth information from sparse ground truth by providing dense supervision.", + "- In the first stage, we propose a data generation strategy that uses monocular depth estimation and mesh reconstruction to simulate training data, enabling the model to learn geometric features from diverse natural images without the need for any LiDAR and ground truth.", + "- In the second stage, we propose a scale- and shift-invariant loss (SSI Loss) to address the scale ambiguity problem in monocular depth estimation, which ensures consistent depth completions across varying scales and focuses on learning real-world scale information.", + "- Our method achieves first place on the official KITTI depth completion benchmark, demonstrating the effectiveness of the proposed approach with significant gains in both quantitative metrics and qualitative visualizations." + ], + "bbox": [ + 89, + 518, + 482, + 805 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "2. Related Work", + "text_level": 1, + "bbox": [ + 89, + 821, + 232, + 835 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "2.1. Depth Completion Datasets", + "text_level": 1, + "bbox": [ + 89, + 845, + 336, + 862 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Depth completion has been widely studied in both indoor and outdoor settings. Indoor datasets, such as NYU Depth", + "bbox": [ + 89, + 869, + 482, + 900 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "V2 [48] and Matterport3D [4], are popular because dense depth measurements are easier to obtain in controlled environments. NYU Depth V2, for instance, provides dense depth maps from a Kinect sensor along with RGB images.", + "bbox": [ + 511, + 90, + 903, + 151 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "In outdoor environments, however, obtaining dense depth annotations is more challenging due to limitations in sensing technologies like LiDAR. The KITTI dataset [53] is a key benchmark for outdoor depth completion but suffers from sparse annotations, covering only about $5\\%$ of the image. To address this, training methods often employ complex post-processing and multi-frame fusion techniques to increase the annotation coverage to around $20\\%$ at most. Furthermore, ground truth values for long distances and dynamic objects typically cannot be included. The sparse ground truth raises challenges for training fine-grained depth completion models. Therefore, we employ monocular depth estimation techniques for dense distillation to preserve fine-grained details.", + "bbox": [ + 511, + 152, + 905, + 363 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "2.2. Depth Completion Methods", + "text_level": 1, + "bbox": [ + 511, + 375, + 761, + 390 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Depth completion methods have seen significant advancements [21, 26, 64]. Compared to predicting depth directly from a single RGB image [2, 22, 44, 73], depth completion fuses LiDAR information to obtain more accurate sparse depth cues. Recent methods [1, 6, 8, 9, 11, 14, 29, 34-38, 40-42, 45, 49, 50, 52, 55-57, 62, 66-70, 74, 76, 79, 80, 82, 84] rely on supervised learning, using pairs of sparse depth maps and corresponding dense ground truth. While effective with sufficient labeled data, these methods are limited in real-world applications, where obtaining dense annotations, especially in outdoor environments, is challenging [5, 15, 83]. To reduce the reliance on dense labels, [10, 77] use a teacher-student model, where a teacher network trained on labeled data provides guidance to a student network. However, pseudo-labels in complex or dynamic scenes can be noisy or inaccurate. Therefore, we turn to monocular depth estimation to provide fine-grained supervision for depth completion.", + "bbox": [ + 511, + 398, + 905, + 670 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "2.3. Monocular Depth Estimation", + "text_level": 1, + "bbox": [ + 511, + 680, + 772, + 696 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Monocular depth estimation [25, 46, 71, 72] has emerged as a powerful tool for generating dense depth maps from single RGB images, making it particularly valuable in scenarios where dense depth annotations are sparse or unavailable [18, 19, 32]. Notable models, such as MiDaS [46] and Depth-Anything [71, 72], have demonstrated impressive capabilities in predicting dense depth maps across a wide range of scenes. G2-MonoDepth [54] have proposed a general framework for monocular RGB-X Data with a scale invariant loss. Despite their versatility, monocular depth estimation methods face the challenge of scale ambiguity, which can limit the accuracy of depth completion, especially when integrated with sparse depth measurements", + "bbox": [ + 511, + 704, + 905, + 900 + ], + "page_idx": 1 + }, + { + "type": "page_number", + "text": "22255", + "bbox": [ + 478, + 945, + 517, + 955 + ], + "page_idx": 1 + }, + { + "type": "image", + "img_path": "images/e2e9da7addaca5781d47007324bb52844176ced247576ec4ff2289216dea86a7.jpg", + "image_caption": [ + "Figure 2. Illustration of our proposed first distillation stage with a data generation strategy to learn geometric features from monocular foundation models, which only requires unlabeled RGB images. We use the estimated monocular depth to re-construct the scene and then simulate the Lidar swap process to generate sparse points for training." + ], + "image_footnote": [], + "bbox": [ + 153, + 87, + 846, + 282 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "from LiDAR. Therefore, we introduce a scale- and shift-invariant loss inspired by [54], combined with a supervised loss to keep real-world scales when fine-tuning on real-world datasets.", + "bbox": [ + 88, + 371, + 480, + 431 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3. Method", + "text_level": 1, + "bbox": [ + 89, + 445, + 181, + 459 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "In this section, we first provide a brief overview of the motivation and formulation of our method. Then, we introduce our two-stage distillation framework. In the first stage, we propose a data generation strategy for pre-training that leverages large-scale, unlabeled images without ground-truth depth annotations, as illustrated in Figure 2. In the second stage, we detail the fine-tuning process on a labeled dataset with sparse ground truth, using monocular depth in combination with SSI Loss, as shown in Figure 3.", + "bbox": [ + 88, + 470, + 482, + 606 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3.1. Motivation and Formulation", + "text_level": 1, + "bbox": [ + 89, + 614, + 346, + 630 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "In the depth completion task, the input typically consists of an RGB image $I$ and a sparse depth map $D_{s}$ obtained from a LiDAR sensor. The objective is to produce a dense depth map $D_{f}$ that provides a complete and detailed representation of the scene's depth. To achieve this, a depth completion model $f_{\\theta}$ , parameterized by $\\theta$ , takes both the RGB image and sparse depth map as inputs:", + "bbox": [ + 88, + 638, + 482, + 744 + ], + "page_idx": 2 + }, + { + "type": "equation", + "text": "\n$$\nD _ {f} = f _ {\\theta} (I, D _ {s}). \\tag {1}\n$$\n", + "text_format": "latex", + "bbox": [ + 210, + 757, + 482, + 773 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "To guide the model's learning, a sparse supervision loss is defined based on the available sparse depth annotations:", + "bbox": [ + 89, + 785, + 482, + 815 + ], + "page_idx": 2 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {L} _ {\\sup } = M \\times \\left| D _ {f} - D _ {\\text {s p a r s e}} \\right|, \\tag {2}\n$$\n", + "text_format": "latex", + "bbox": [ + 173, + 827, + 482, + 844 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "where $M$ represents the valid mask for sparse depth ground truth $D_{\\mathrm{sparse}}$ . This loss aligns the predicted dense depth map $D_{f}$ with the sparse ground truth, ensuring that the output", + "bbox": [ + 89, + 854, + 483, + 902 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "is consistent with the available depth data. However, this sparse supervision provides limited guidance, particularly in outdoor scenes [53], leading to challenges in achieving fine-grained and consistent depth predictions due to the inherent limitations of sparse supervision.", + "bbox": [ + 511, + 371, + 905, + 446 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Foundation models for monocular depth estimation [71, 72] provide an alternative approach by generating fine-grained dense depth from single RGB images, which we use as supervision without relying solely on LiDAR. Therefore, in the first stage, we utilize the monocular foundation models to generate diverse and large-scale training data, to provide dense supervision for pre-training, allowing the model to learn geometric features. However, monocular methods suffer from scale ambiguities, which can lead to inconsistencies when combining dense monocular predictions with sparse depth. To address this, in the second stage, we introduce SSI Loss [46] combined with L1 loss when finetuning on labeled datasets with sparse ground truth to mitigate scale inconsistencies and learn real-world scale.", + "bbox": [ + 511, + 450, + 906, + 662 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3.2. First Stage: Data Generation and Pre-training", + "text_level": 1, + "bbox": [ + 511, + 686, + 903, + 703 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "To leverage diverse images without depth annotations, we introduce a pre-training strategy that utilizes synthesized depth data generated through monocular depth estimation on large-scale natural image datasets inspired by [31, 32]. As illustrated in Figure 2, this pre-training phase enables the model to learn robust geometric features from monocular foundation models across diverse scenes, preparing it for subsequent fine-tuning on real-world datasets.", + "bbox": [ + 511, + 714, + 905, + 835 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Specifically, we first utilize a pre-trained monocular depth estimation model, such as Depth Anything V2 [72], to predict depth maps for such images. For each natural image $I_{\\mathrm{un}}$ , the monocular depth estimation model predicts a", + "bbox": [ + 511, + 839, + 906, + 901 + ], + "page_idx": 2 + }, + { + "type": "page_number", + "text": "22256", + "bbox": [ + 478, + 944, + 517, + 955 + ], + "page_idx": 2 + }, + { + "type": "image", + "img_path": "images/5dfe5fd98d70aa5e864902e159ba1bde7d23d7a0bb74de6ad51205f5afdcf78e.jpg", + "image_caption": [ + "Figure 3. Illustration of our proposed second distillation stage utilizing foundation models for monocular depth estimation when fine-tuning on labeled datasets. Sparse ground truth provides real-world depth scale with L1 loss. Our method enhances this process by incorporating dense monocular depth for fine-grained supervision. However, monocular depth maps come with inherent scale and shift ambiguities. To address these challenges, we employ a Scale- and Shift-Invariant Loss (SSI Loss) that aligns the predictions with the dense monocular depth to match the real-world depth scale, ensuring more accurate depth completion." + ], + "image_footnote": [], + "bbox": [ + 130, + 88, + 867, + 244 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "corresponding dense depth map $D_{\\mathrm{syn}}$ ..", + "bbox": [ + 89, + 363, + 341, + 378 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\nD _ {\\text {s y n}} = f _ {\\mathrm {m o}} \\left(I _ {\\mathrm {u n}}\\right), \\tag {3}\n$$\n", + "text_format": "latex", + "bbox": [ + 210, + 391, + 480, + 407 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "where $f_{\\mathrm{mo}}$ represents the monocular depth estimation model. These synthesized depth maps, though not accurate in metric scale, capture the relative depth relationships and structural details in the scene, providing valuable supervision signals during pre-training.", + "bbox": [ + 89, + 420, + 483, + 494 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "To simulate LiDAR scanning, we first sample a random camera intrinsic matrix $K$ as:", + "bbox": [ + 89, + 496, + 483, + 526 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\nK = \\left[ \\begin{array}{c c c} f _ {x} & 0 & c _ {x} \\\\ 0 & f _ {y} & c _ {y} \\\\ 0 & 0 & 1 \\end{array} \\right], \\tag {4}\n$$\n", + "text_format": "latex", + "bbox": [ + 194, + 536, + 480, + 585 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "where $f_{x}$ and $f_{y}$ represent the focal lengths along the $x$ - and $y$ -axes, and $(c_{x}, c_{y})$ denotes the center of the image.", + "bbox": [ + 89, + 595, + 482, + 626 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "For each pixel $(u,v)$ in the depth map, the corresponding 3D coordinates $(X,Y,Z)$ in the camera coordinate system are derived using the sampled camera intrinsic matrix $K$ . This transformation generates a point cloud $P = \\{(X,Y,Z)\\}$ that represents the 3D spatial locations of each pixel. Leveraging the 3D point cloud $P$ , we reconstruct a mesh $M$ via surface reconstruction techniques, such as Poisson Surface Reconstruction [24], to create a continuous 3D surface model for simulation.", + "bbox": [ + 89, + 627, + 482, + 762 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Next, we simulate a LiDAR sensor by generating a set of ray direction vectors $\\mathbf{d}$ . Each ray is cast from the origin along $\\mathbf{d}_{i,j}$ , intersecting the mesh $M$ to determine the distance to the intersection point. This distance is recorded as the simulated LiDAR depth reading $D_{\\mathrm{simu}}(i,j)$ . The resulting sparse depth map $D_{\\mathrm{simu}}$ emulates LiDAR depth readings from the reconstructed 3D surface, to provide a synthetic depth image comparable to actual LiDAR data.", + "bbox": [ + 89, + 763, + 482, + 883 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Then, the depth completion model $f_{\\theta}$ takes the RGB im", + "bbox": [ + 109, + 885, + 482, + 900 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "age and the sparse depth from simulated scanning as inputs:", + "bbox": [ + 511, + 363, + 903, + 378 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\nD _ {f} = f _ {\\theta} \\left(I _ {\\mathrm {u n}}, D _ {\\mathrm {s i m u}}\\right), \\tag {5}\n$$\n", + "text_format": "latex", + "bbox": [ + 620, + 387, + 903, + 402 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "where $D_{\\mathrm{simu}}$ is the sparse input from simulated LiDAR scanning. Afterwards, the depth completion model is pretrained using L1 Loss:", + "bbox": [ + 511, + 412, + 903, + 457 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {L} _ {\\text {p r e}} = \\left| D _ {f} - D _ {m} \\right|. \\tag {6}\n$$\n", + "text_format": "latex", + "bbox": [ + 640, + 465, + 903, + 483 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "This pre-training process encourages the depth completion model $f_{\\theta}$ to align its predictions with the synthesized depth maps, learning meaningful geometric features from the generated training data using monocular foundation models and diverse natural image data. Additionally, the model learns complex geometric structures and relative depth relationships across varying scenes, which are crucial for accurate depth completion in the fine-tuning phase.", + "bbox": [ + 511, + 491, + 905, + 613 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "3.3. Second Stage: Fine-tuning", + "text_level": 1, + "bbox": [ + 511, + 619, + 754, + 636 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "The depth completion model pre-trained on the generated data learns powerful geometric features from the monocular foundation models. However, due to the inherent scale ambiguity of monocular depth, the model prediction is not consistent with the scale in the real world. To learn the real-world scale, we introduce a combined loss for fine-tuning on labeled datasets. To be specific, we first utilize the sparse ground truth $D_{s}$ for supervised loss $\\mathcal{L}_{\\mathrm{sup}}$ to focus on real-world depth, as in Equation 2. Then, the dense monocular depth $D_{m}$ predicted by monocular depth models is also used to provide approximate depth values across the image, especially in regions not covered by the ground truth $D_{s}$ .", + "bbox": [ + 511, + 643, + 906, + 824 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "However, the key challenge in using monocular depth estimation as supervision is the inherent scale ambiguity in depth predictions. To address this issue, we incorporate a Scale- and Shift-Invariant Loss (SSI Loss) into our distillation framework. The SSI Loss is designed to align the", + "bbox": [ + 511, + 825, + 906, + 900 + ], + "page_idx": 3 + }, + { + "type": "page_number", + "text": "22257", + "bbox": [ + 478, + 944, + 517, + 955 + ], + "page_idx": 3 + }, + { + "type": "table", + "img_path": "images/cff35327f145640187eff97f823cb0feb9f87181f01f738c2ff5a57d15185c9e.jpg", + "table_caption": [ + "Table 1. Unlabeled datasets used for pre-training data generation." + ], + "table_footnote": [], + "table_body": "
DatasetIndoorOutdoor# Images
COCO [33]✓✓118,287
Google Landmarks [59]✓117,576
Nuscenes [3]✓93,475
Cityscapes [12]✓19,998
DAVIS [43]✓✓10,581
", + "bbox": [ + 94, + 114, + 475, + 223 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "predicted depth map $D_{f}$ with the dense monocular depth $D_{m}$ . It remains invariant to differences in scale and shift between $D_{f}$ and $D_{m}$ . The SSI Loss is formulated as:", + "bbox": [ + 89, + 247, + 482, + 292 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {L} _ {\\text {S S I L}} = \\min _ {s, b} | D _ {f} - (s \\cdot D _ {m} + b) |, \\tag {7}\n$$\n", + "text_format": "latex", + "bbox": [ + 168, + 305, + 482, + 327 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "where $s$ and $b$ are the optimal scale and shift parameters that align the predicted depth $D_{f}$ with the dense monocular depth $D_{m}$ . The loss function seeks to find the best alignment between $D_{f}$ and $D_{m}$ , effectively normalizing any global differences in scale and offset. By minimizing this loss across all pixels in the image, the model is encouraged to produce depth maps that maintain consistency with the relative depth structure provided by the monocular depth estimates. In addition, we adapt a gradient matching term to preserve sharpness and align with with depth discontinuities. The gradient matching term is defined as:", + "bbox": [ + 89, + 339, + 483, + 506 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {L} _ {\\text {r e g}} = \\frac {1}{N} \\sum_ {k = 1} ^ {K} \\left(| \\nabla_ {x} R ^ {k} | + | \\nabla_ {y} R ^ {k} |\\right), \\tag {8}\n$$\n", + "text_format": "latex", + "bbox": [ + 163, + 517, + 482, + 558 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "where $R = D_{f} - D_{m}$ , and $R^{k}$ denotes the difference of depth maps at scale $k$ . We use $K = 4$ scale levels as in [46], halving the image resolution at each level. The final objective function combines the supervised loss $\\mathcal{L}_{\\mathrm{sup}}$ , the dense distillation loss $\\mathcal{L}_{\\mathrm{SSIL}}$ , and the regularization term $\\mathcal{L}_{\\mathrm{reg}}$ .", + "bbox": [ + 89, + 570, + 483, + 648 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "4. Experiments", + "text_level": 1, + "bbox": [ + 89, + 660, + 223, + 678 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "In this section, we first introduce the details of our implementation, as well as the datasets and evaluation metrics for experiments. Then, detailed comparisons are conducted with the state-of-the-art methods. Finally, ablations and discussions are performed to confirm the effectiveness of our proposed main components. Additional analysis is provided in the supplementary materials, along with videos.", + "bbox": [ + 89, + 686, + 483, + 792 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "4.1. Experimental Pipeline", + "text_level": 1, + "bbox": [ + 89, + 801, + 299, + 816 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "First stage involves utilizing our proposed data generation strategy to train the depth completion model from scratch, as described in Section 3.2. In this stage, we generate training data using the RGB images in Table 1 and train the depth completion model using SSI Loss, which does not require", + "bbox": [ + 89, + 825, + 483, + 901 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "any ground truth depth. This stage allows the pre-trained model to focus on learning diverse geometric features without being constrained by labeled data, enhancing generalization across different scenes.", + "bbox": [ + 511, + 90, + 903, + 150 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Second stage involves adapting the pre-trained model on labeled datasets, using supervised loss (L1 loss) with depth annotations and our proposed SSI Loss for knowledge distillation, as described in Section 3.3. Using L1 loss with sparse ground truth allows the depth completion model to adapt to the real-world depth scale under sparse supervision. Specifically, our proposed monocular model distillation uses SSI Loss to help the model maintain fine-grained detail with dense supervision by distilling monocular foundation models. Fine-tuning and validation are performed on the KITTI and NYU Depth V2 datasets, respectively.", + "bbox": [ + 511, + 151, + 906, + 319 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "4.2. Implementation Details", + "text_level": 1, + "bbox": [ + 511, + 329, + 730, + 344 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "We follow the setup of recent work [50, 57, 79] and train our model on 4 NVIDIA RTX A100 GPUs. For the monocular foundation model used in distillation, we adopt Depth Anything V2 [72] due to its robust performance. For the depth completion network architecture, we use BP-Net [50] as our base model, which is primarily composed of ResNet [20] blocks. In the first stage, we utilize a mixed dataset collected from a wide range of resources, as shown in Table 1. We select five large-scale public datasets as our unlabeled sources for their diverse scenes, totaling approximately 360,000 images with varying scenes and image scales. We adopt AdamW with a weight decay of 0.05 as the optimizer and clip gradients if their norms exceeding 0.1. In the first stage, our model is pre-trained from scratch for 600K iterations. In the second stage, the pre-trained model is fine-tuned on the labeled datasets for 300K iterations using a OneCycle learning rate policy. We set the batch size to 16. The final model is obtained by applying Exponential Moving Average (EMA) with a decay of 0.9999, which helps stabilize the model parameters during training.", + "bbox": [ + 511, + 351, + 906, + 655 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "4.3. Datasets and Evaluation Metrics", + "text_level": 1, + "bbox": [ + 511, + 665, + 799, + 680 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "KITTI Depth Completion Dataset [53] serves as a standard benchmark for depth completion in autonomous driving scenarios. It provides over 93,000 training samples, including sparse LiDAR depth maps, their corresponding RGB images, and ground truth depth maps. The sparse LiDAR depth maps are generated from Velodyne LiDAR scans, which typically cover only a small percentage of the image pixels (approximately $5\\%$ ). The ground truth depth maps are obtained via multi-frame matching and cover approximately $20\\%$ image pixels. The dataset also includes 1,000 test samples on the private online benchmark.", + "bbox": [ + 511, + 688, + 905, + 853 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "NYU Depth V2 Dataset [48] is also a standard benchmark for indoor scene understanding, consisting of RGB-D data captured with a Kinect sensor. It provides 1,449 densely", + "bbox": [ + 511, + 854, + 905, + 900 + ], + "page_idx": 4 + }, + { + "type": "page_number", + "text": "22258", + "bbox": [ + 478, + 945, + 517, + 955 + ], + "page_idx": 4 + }, + { + "type": "table", + "img_path": "images/b687306afa5c1d6916d23f1efe6d83550b0d71bab6a6db4691489455bcd8a7b3.jpg", + "table_caption": [ + "Table 2. Performance on KITTI and NYUv2 datasets. For the KITTI dataset, results are evaluated by the KITTI testing server and ranked by the RMSE (in mm). For the NYUv2 dataset, we report their performance on the official test set in their papers." + ], + "table_footnote": [], + "table_body": "
MethodYearKITTINYUv2
RMSE ↓MAE ↓iRMSE ↓iMAE ↓RMSE ↓REL ↓δ1.25 ↑
S2D [38]2018814.73249.952.801.210.2300.04497.1
CSPN [8]20191019.64279.462.931.150.1170.01699.2
DeepLiDAR [45]2019758.38226.502.561.150.1150.02299.3
CSPN++ [9]2020743.69209.282.070.900.1010.01599.5
GuideNet [49]2020736.24218.832.250.990.115--
FCFR [35]2021735.81217.152.200.980.1060.01599.5
ACMNet [80]2021744.91206.092.080.900.1050.01599.4
NLSPN [41]2020741.68199.591.990.840.0920.01299.6
PointDC [74]2023736.07201.871.970.870.0890.01299.6
RigNet [67]2022712.66203.252.080.900.0900.01399.6
DySPN [34]2022709.12192.711.880.820.0900.01299.6
BEV@DC [82]2023697.44189.441.830.820.0890.01299.6
CFormaler [79]2023708.87203.452.010.880.0900.012-
LRRU [56]2023696.51189.961.870.810.0910.01199.6
TPVD [70]2024693.97188.601.820.810.0860.01099.7
ImprovingDC [57]2024686.46187.951.830.810.0910.01199.6
BP-Net [50]2024684.90194.691.820.840.0890.01299.6
DMD3C (Ours)-678.12194.461.820.850.0850.01199.7
", + "bbox": [ + 178, + 127, + 818, + 435 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "labeled images from various indoor scenes. Missing ground truth is mainly due to the difference in perspective between the depth sensor and the camera.", + "bbox": [ + 88, + 459, + 480, + 503 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Evaluation Metrics used for KITTI are root mean squared error (RMSE), mean absolute error (MAE), root mean squared error of the inverse depth (iRMSE), and mean absolute error of the inverse depth (iMAE), among which RMSE is used for ranking. These metrics are obtained by submission and evaluation from a non-public test set. For NYU, we follow the previous work and report the RMSE, mean absolute relative error (REL), and $\\delta_{\\theta}$ , which represents the percentage of pixels whose error is less than a threshold $\\theta$ .", + "bbox": [ + 88, + 505, + 482, + 641 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "4.4. Comparison with State-of-the-art Methods", + "text_level": 1, + "bbox": [ + 89, + 650, + 455, + 666 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Evaluation on Depth Completion Benchmarks. We evaluate our method on the test sets of the NYUv2 dataset [48] and KITTI dataset [53]. Table 2 shows the quantitative comparison of our method and other top ranking published methods. On the KITTI leaderboard, our proposed $\\mathrm{DMD}^{3}\\mathrm{C}$ ranks 1st outperforming all other methods under the primary RMSE metric at the time of paper submission. It also has comparable performance under other evaluation metrics. On the NYUv2 dataset, our method achieves the best RMSE and best $\\delta_{1.25}$ . Additionally, $\\mathrm{DMD}^{3}\\mathrm{C}$ maintains competitive results across several other evaluation metrics, further confirming its robustness and generalization.", + "bbox": [ + 89, + 672, + 482, + 854 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "To enhance understanding of $\\mathrm{DMD}^3\\mathrm{C}$ 's superiority over other state-of-the-art methods, we provide visual comparisons. Figure 4 provides qualitative comparisons with other", + "bbox": [ + 89, + 854, + 483, + 900 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "SOTA methods on the KITTI test set using publicly available results. We use results from the official benchmark1 produced by the best models for method comparison. Our $\\mathrm{DMD}^{3}\\mathrm{C}$ model excels in maintaining sharp object boundaries and capturing fine details in areas where other models encounter difficulties, particularly in scenes with complex structures or objects at varying distances. In such challenging regions, other methods fail to estimate accurate depth. Furthermore, as shown in the error maps, our method demonstrates a clear advantage even in regions without ground truth, despite these areas not being factored into the evaluation metrics.", + "bbox": [ + 511, + 459, + 906, + 640 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Not only does our model retain the fine details in the depth map, but it also maintains structural integrity in the potential three-dimensional space. Figure 5 illustrates that our proposed two-stage distillation improves performance in the 3D space by visualizing point clouds from the completed depth maps. In comparison to other methods, our $\\mathrm{DMD}^{3}\\mathrm{C}$ produces a more coherent and complete 3D structure, further outperforming existing methods.", + "bbox": [ + 511, + 641, + 908, + 763 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "4.5. Discussions", + "text_level": 1, + "bbox": [ + 511, + 773, + 638, + 789 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Effect of Pre-training in the First Stage. First, to evaluate the impact of the proposed pre-training strategy using unlabeled images, we conduct an experiment where the depth completion model is trained from scratch, without any pretraining. As shown in Table 3, removing the pre-training", + "bbox": [ + 511, + 797, + 906, + 876 + ], + "page_idx": 5 + }, + { + "type": "page_footnote", + "text": "MethodKITTIRMSE ↓MAE ↓iRMSE ↓iMAE ↓w/o Pre-train682.34194.961.820.84w/o SSI Loss684.54195.651.860.85DMD3C678.12194.461.820.85", + "bbox": [ + 98, + 410, + 475, + 502 + ], + "page_idx": 6 + }, + { + "type": "table", + "img_path": "images/bcf09035aa9cfb4c134256f3f4de35ebebadae225a6235ea7d45a9d490dc8489.jpg", + "table_caption": [ + "Table 4. Ablations on different network architectures." + ], + "table_footnote": [], + "table_body": "
MethodKITTI
RMSE ↓MAE ↓iRMSE ↓iMAE ↓
LRRU [56]696.51189.961.870.81
+ Ours693.17189.601.850.80
CFormer [79]764.87183.881.890.80
+ Ours760.29183.621.880.80
BP-Net [50]684.90194.691.820.84
+ Ours678.12194.461.820.85
", + "bbox": [ + 96, + 542, + 477, + 690 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "step results in a performance degradation, with an increase in RMSE by $4.22\\mathrm{mm}$ . This demonstrates that pre-training plays a crucial role in enhancing the model's ability to generalize, enabling it to better capture complex geometric features across diverse scenes.", + "bbox": [ + 89, + 718, + 482, + 792 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Effect of SSI Loss in the Second Stage. Next, we show the impact of removing SSI Loss in the second distillation stage, while retaining the standard supervised loss with sparse ground truth. As shown in Table 3, the results indicate that the proposed SSI Loss improves the model's performance. This demonstrates the benefit of aligning monocular depth supervision with real-world scale using SSI Loss.", + "bbox": [ + 89, + 794, + 483, + 901 + ], + "page_idx": 6 + }, + { + "type": "table", + "img_path": "images/7b9076672cae15858f4ffc66b1cdc77a88dcc3bec7dbeddd65ac8f3395608702.jpg", + "table_caption": [ + "Table 5. Zero-shot performance comparison with other methods on out-of-the-domain datasets." + ], + "table_footnote": [], + "table_body": "
MethodScanNetDDADVOID1500
RMSE ↓MAE ↓RMSE ↓MAE ↓RMSE ↓MAE ↓
CFormal0.1200.2329.6063.3280.7260.261
LRRU0.1320.2459.1642.7380.6980.232
BP-Net0.1220.2128.9032.7120.7040.230
Ours0.1010.2107.7662.4980.6760.225
", + "bbox": [ + 514, + 422, + 905, + 489 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Network Architectures. We evaluate the compatibility of our method with different network architectures, as shown in Table 4. Since we focus on training strategies, we have the flexibility to various models, demonstrating the robustness of our approach across different network designs. Specifically, we evaluate BP-Net [50], LRRU [56], and CFormer (L1 Loss) [79]. BP-Net is one of the most advanced methods, while LRRU and CFormer are considered representative architectures. Our approach consistently improves performance across all three models. Notably, BP-Net with our method achieves the best performance in terms of RMSE, the primary evaluation metric.", + "bbox": [ + 511, + 518, + 906, + 700 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Generalization to Out-of-Domain Datasets. To evaluate the generalization ability of our method, we conduct zero-shot testing on unseen datasets, including ScanNet [13], DDAD [17], and Void1500 [60], as shown in Table 5. Our approach consistently achieves the best performance across all datasets, demonstrating lower RMSE and MAE compared to prior methods. Notably, our method outperforms BP-Net by a significant margin on DDAD, reducing RMSE from 8.903 to 7.766, highlighting its effectiveness in handling unseen scenarios.", + "bbox": [ + 511, + 702, + 906, + 853 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Application on Dense SLAM. Figure 6 compares a representative SLAM method, Droid-SLAM [51], with sparse LiDAR points (left) and our method with dense depth com", + "bbox": [ + 511, + 854, + 906, + 900 + ], + "page_idx": 6 + }, + { + "type": "page_number", + "text": "22260", + "bbox": [ + 478, + 944, + 517, + 955 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/d92af7072b332bf2f41f32096c72c303c026786d5b6f68f8d51edc567ddfc169.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 96, + 87, + 228, + 119 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/0ea130686549fc0fef5074e0185653c52a55763cd4ab35aa02f2b7a176c2ba83.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 112, + 143, + 227, + 204 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/a0224d2802d609c9ad64290f5d6cfe7d40eb72a48866adbc93ee88a3ee7b9988.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 230, + 89, + 362, + 119 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/9c7a26b3be63938907cfbcd1fb7d582443ac4c6c999eb3e3529b115a3b01c42e.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 233, + 119, + 356, + 203 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/33f3bfd10d5b1f8fa2ee922707ab8852aa4ed75e7436733ee8dfb37e47ff8152.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 364, + 88, + 496, + 119 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/5a784980738bde6ae8ebc381e0fcb3ecf2cce3f1f20fa47da9fadc6ea884936d.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 367, + 119, + 496, + 204 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/95986492181b4b2136716019c32d031945fa795c696a7225321c772f49b08cfb.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 500, + 89, + 630, + 119 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/21c687819faf9909dd669d1f6fc5cc002f714832abe6fbad556cf1cc012e499d.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 501, + 119, + 629, + 205 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/899409fbac316982c51d0e9460936d2e93bf0d9676b5cea745fdc121fcb4093c.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 633, + 89, + 763, + 119 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/34e24f874c8be7a793a77a9bc67d3438696c2ceac104a8b6e47ff0be7f2a0c5b.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 635, + 127, + 756, + 205 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/3351ae1579c1b060aadc6fce947e3cc233267af41a0aa1571191acde6ad21637.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 767, + 89, + 900, + 119 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/660f3820d343b76e3657dad90e0cf4234e5791e78ad35c88d47fcbf3e11937f5.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 769, + 127, + 888, + 205 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/b62251f35aadc5fe7e3ebf13fe2a66c650052727c1df9c14a0cc716295bc7112.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 96, + 214, + 228, + 244 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/4aedd7be838c006ec45932487550d848e596a8d1a53c170d9f9d04ccb5661cef.jpg", + "image_caption": [ + "(a) Input" + ], + "image_footnote": [], + "bbox": [ + 101, + 247, + 227, + 351 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/02e4c30d5be755f7a545622c2ae6f6cae7c529b6de8fcd8600733f95042b846c.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 232, + 215, + 362, + 244 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/066040b0d28bacecd4e6720db1fed553b71afed5c4b327b28d3ca4dd8ce756c8.jpg", + "image_caption": [ + "(b) CFormer [79]" + ], + "image_footnote": [], + "bbox": [ + 232, + 244, + 362, + 352 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/fc756172ad3ce0950f929f2461cd3f4196cc65297420767574e521334ed1b131.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 366, + 215, + 496, + 244 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/8faa4db24d3df57d5b6d3ed227fec2e931a664d5300e1c2e845b1ea0de0073f4.jpg", + "image_caption": [ + "(c) LRRU [56]" + ], + "image_footnote": [], + "bbox": [ + 367, + 244, + 496, + 352 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/73169d4e174de121892b9ffa88a0ca2d8dc52c152631dc2a8d2f8db532a75f8b.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 500, + 215, + 630, + 244 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/f6ec725f2e27c956c91abb95ed2343ae6b943c4fbe83201b5cd0d2763794bf45.jpg", + "image_caption": [ + "(d) ImprovingDC [57]" + ], + "image_footnote": [], + "bbox": [ + 503, + 244, + 627, + 351 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/7712ba798397df2149c95b19198f648d12bd6d8ec983b5693ee3a9bcaa0888d8.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 633, + 215, + 764, + 244 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/b8b5b33b17b7f8af3bcad63dbb1d911fa98149ab01488913d1f605d370621fdb.jpg", + "image_caption": [ + "(e) BP-Net [50]" + ], + "image_footnote": [], + "bbox": [ + 638, + 244, + 761, + 347 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/07fa64ef0fb71eac5aab3cbf56fc8ccb816ed6e23da09e3d1e8009cc6e622744.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 767, + 215, + 900, + 244 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/9b8ec9b8a9c8224b721c7f100f945cabfd771176c22bec2cbcf522088057c3ed.jpg", + "image_caption": [ + "(f) DMD $^3$ C (Ours)" + ], + "image_footnote": [], + "bbox": [ + 774, + 244, + 892, + 349 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/059646f4a2f0f773c06757f855ac0d0acabe21d6d0b1a403cbcb5311c0790c1d.jpg", + "image_caption": [ + "Figure 5. Qualitative comparison of depth completion methods. This figure demonstrates the performance of various depth completion models, including CFormer, LRRU, ImprovingDC, BP-Net, and our proposed DMD $^3$ C. For each method, we show the input RGB images with sparse LiDAR points (left), along with the resulting completed depth maps and corresponding 3D point cloud reconstructions." + ], + "image_footnote": [], + "bbox": [ + 117, + 444, + 212, + 546 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/884076331dc16b92c35e42bb4f59865ae59245b52b8a0b8332d951ee604918fe.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 215, + 444, + 281, + 546 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/4f6e332ef06753404b1638d5d26dfa497c774eed8874ad61a48a5d459f0e8126.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 295, + 445, + 395, + 545 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/9c76f7f9112fb0f990455823736b89bafb1619938bd6f035fbe13592af279b67.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 397, + 445, + 460, + 545 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/59ab0a551e3990bbdf80dec89dd443ac6bcefba92229432d03362cd3668371b3.jpg", + "image_caption": [ + "Droid-SLAM + LiDAR", + "Figure 6. Qualitative comparison of Droid-SLAM with sparse LiDAR points and our method with dense depth completion." + ], + "image_footnote": [], + "bbox": [ + 119, + 547, + 210, + 676 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/3a510b5c300e4212d14ca70c8b1bf863b9d386c7fa31f3e000792bdba68d6003.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 210, + 547, + 282, + 676 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/58ed8c4656c66163632517d07e030b95843f0227dcfe3f28ce470a8a4abb1931.jpg", + "image_caption": [ + "Droid-SLAM + Ours" + ], + "image_footnote": [], + "bbox": [ + 295, + 547, + 387, + 676 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/7949bc9af021892dc8a2903c60c412d842c5eed246ae25c68e89b7d235af37dd.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 390, + 547, + 460, + 676 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "pletion (right). Droid-SLAM suffers from structural distortions and noise, especially in occluded or low-texture areas (red boxes), due to limited depth information. In contrast, our method generates more consistent reconstructions, preserving structural details and improving alignment (green boxes). These results highlight the advantage of dense depth completion in enhancing SLAM quality.", + "bbox": [ + 89, + 795, + 482, + 900 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "5. Conclusion", + "text_level": 1, + "bbox": [ + 513, + 445, + 632, + 460 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "In this work, we present $\\mathrm{DMD}^{3}\\mathrm{C}$ , a novel two-stage distillation framework that distills knowledge from monocular depth estimation foundation models into the depth completion task. In the first stage, we introduce a data generation strategy that leverages monocular depth estimation and mesh reconstruction to simulate training data, allowing the model to learn geometric features from diverse natural images. In the second stage, we propose a scale- and shift-invariant loss (SSI Loss) combined with a supervised L1 loss with sparse ground truth, which addresses the scale ambiguity in monocular depth estimation. This ensures consistent depth completions across varying scales and focuses on learning real-world scale information. Extensive experiments on the depth completion benchmarks demonstrate that $\\mathrm{DMD}^{3}\\mathrm{C}$ achieves state-of-the-art performance, ranking first on the KITTI leaderboard, and significantly outperforming existing methods. Our results highlight the framework's ability to produce high-quality depth maps with improved detail and structural consistency, making it a promising solution for depth completion tasks.", + "bbox": [ + 511, + 473, + 906, + 776 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Acknowledgements", + "text_level": 1, + "bbox": [ + 513, + 797, + 679, + 814 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "This work was supported by the National Key R&D Program of China (2022YFC3300704), the National Natural Science Foundation of China (62331006, 62171038, and 62088101), and the Fundamental Research Funds for the Central Universities.", + "bbox": [ + 511, + 825, + 905, + 898 + ], + "page_idx": 7 + }, + { + "type": "page_number", + "text": "22261", + "bbox": [ + 478, + 945, + 517, + 955 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "References", + "text_level": 1, + "bbox": [ + 91, + 89, + 187, + 104 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[1] Luca Bartolomei, Matteo Poggi, Andrea Conti, Fabio Tosi, and Stefano Mattoccia. Revisiting depth completion from a stereo matching perspective for cross-domain generalization. In International Conference on 3D Vision (3DV), pages 1360-1370, 2024. 2", + "[2] Shariq Farooq Bhat, Reiner Birkl, Diana Wofk, Peter Wonka, and Matthias Müller. Zoedepth: Zero-shot transfer by combining relative and metric depth. arXiv preprint arXiv:2302.12288, 2023. 2", + "[3] Holger Caesar, Varun Bankiti, Alex H Lang, Sourabh Vora, Venice Erin Liong, Qiang Xu, Anush Krishnan, Yu Pan, Giancarlo Baldan, and Oscar Beijbom. nuscenes: A multimodal dataset for autonomous driving. In the IEEE Conference on Computer Vision and Pattern Recognition (CVPR), pages 11621-11631, 2020. 5", + "[4] Angel Chang, Angela Dai, Thomas Funkhouser, Maciej Halber, Matthias Niebner, Manolis Savva, Shuran Song, Andy Zeng, and Yinda Zhang. Matterport3d: Learning from rgb-d data in indoor environments. In International Conference on 3D Vision (3DV), pages 667-676, 2017. 2", + "[5] Linwei Chen, Ying Fu, Kaixuan Wei, Dezhi Zheng, and Felix Heide. Instance segmentation in the dark. International Journal of Computer Vision, 131(8):2198-2218, 2023. 2", + "[6] Linwei Chen, Ying Fu, Lin Gu, Chenggang Yan, Tatsuya Harada, and Gao Huang. Frequency-aware feature fusion for dense image prediction. IEEE Transactions on Pattern Analysis and Machine Intelligence, 46(12):10763-10780, 2024. 2", + "[7] Yun Chen, Bin Yang, Ming Liang, and Raquel Urtasun. Learning joint 2d-3d representations for depth completion. In the IEEE Conference on Computer Vision and Pattern Recognition (CVPR), pages 10023-10032, 2019. 1", + "[8] Xinjing Cheng, Peng Wang, and Ruigang Yang. Learning depth with convolutional spatial propagation network. IEEE Transactions on Pattern Analysis and Machine Intelligence (TPAMI), 42(10):2361-2379, 2019. 2, 6", + "[9] Xinjing Cheng, Peng Wang, Chenye Guan, and Ruigang Yang. Cspn++: Learning context and resource aware convolutional spatial propagation networks for depth completion. In the AAAI Conference on Artificial Intelligence (AAAI), pages 10615-10622, 2020. 2, 6", + "[10] Keunhoon Choi, Somi Jeong, Youngjung Kim, and Kwanghoon Sohn. Stereo-augmented depth completion from a single rgb-lidar image. In IEEE International Conference on Robotics and Automation (ICRA), pages 13641-13647, 2021. 2", + "[11] Andrea Conti, Matteo Poggi, and Stefano Mattoccia. Sparsity agnostic depth completion. In the IEEE Winter Conference on Applications of Computer Vision (WACV), pages 5871-5880, 2023. 2", + "[12] Marius Cordts, Mohamed Omran, Sebastian Ramos, Timo Rehfeld, Markus Enzweiler, Rodrigo Benenson, Uwe Franke, Stefan Roth, and Bernt Schiele. The cityscapes dataset for semantic urban scene understanding. In the IEEE Conference on Computer Vision and Pattern Recognition (CVPR), pages 3213-3223, 2016. 5" + ], + "bbox": [ + 93, + 114, + 483, + 900 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[13] Angela Dai, Angel X Chang, Manolis Savva, Maciej Halber, Thomas Funkhouser, and Matthias Nießner. Scannet: Richly-annotated 3d reconstructions of indoor scenes. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 5828-5839, 2017. 7", + "[14] Zheng Fang, Ying Fu, and Lixiong Liu. A dual of transformer features-related map-intelligent generation method. Journal of Image and Graphics, 28(10):3281-3294, 2023. 2", + "[15] Ying Fu, Zichun Wang, Tao Zhang, and Jun Zhang. Lowlight raw video denoising with a high-quality realistic motion dataset. IEEE Transactions on Multimedia, 25:8119-8131, 2022. 2", + "[16] Qi Guan, Zihao Sheng, and Shibei Xue. Hrpose: Real-time high-resolution 6d pose estimation network using knowledge distillation. Chinese Journal of Electronics, 32(1):189-198, 2023. 2", + "[17] Vitor Guizilini, Rares Ambrus, Sudeep Pillai, Allan Raventos, and Adrien Gaidon. 3d packing for self-supervised monocular depth estimation. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 2485-2494, 2020. 7", + "[18] Wenxuan Guo, Yingping Liang, Zhiyu Pan, Ziheng Xi, Jianjiang Feng, and Jie Zhou. Camera-lidar cross-modality gait recognition. arXiv preprint arXiv:2407.02038, 2024. 2", + "[19] Wenxuan Guo, Zhiyu Pan, Yingping Liang, Ziheng Xi, Zhicheng Zhong, Jianjiang Feng, and Jie Zhou. Lidar-based person re-identification. In the IEEE Conference on Computer Vision and Pattern Recognition (CVPR), pages 17437–17447, 2024. 2", + "[20] Kaiming He, Xiangyu Zhang, Shaoqing Ren, and Jian Sun. Deep residual learning for image recognition. In the IEEE Conference on Computer Vision and Pattern Recognition (CVPR), pages 770-778, 2016. 5", + "[21] Junjie Hu, Chenyu Bao, Mete Ozay, Chenyou Fan, Qing Gao, Honghai Liu, and Tin Lun Lam. Deep depth completion from extremely sparse data: A survey. IEEE Transactions on Pattern Analysis and Machine Intelligence (TPAMI), 45(7): 8244-8264, 2022. 1, 2", + "[22] Mu Hu, Wei Yin, Chi Zhang, Zhipeng Cai, Xiaoxiao Long, Hao Chen, Kaixuan Wang, Gang Yu, Chunhua Shen, and Shaojie Shen. Metric3d v2: A versatile monocular geometric foundation model for zero-shot metric depth and surface normal estimation. IEEE Transactions on Pattern Analysis and Machine Intelligence (TPAMI), pages 1-18, 2024. 2", + "[23] Jinwoo Jeon, Hyunjun Lim, Dong-Uk Seo, and Hyun Myung. Struct-mdc: Mesh-refined unsupervised depth completion leveraging structural regularities from visual slam. IEEE Robotics and Automation Letters, 7(3):6391-6398, 2022. 1", + "[24] Michael Kazhdan, Matthew Bolitho, and Hugues Hoppe. Poisson surface reconstruction. In Eurographics Symposium on Geometry Processing, 2006. 4", + "[25] Bingxin Ke, Anton Obukhov, Shengyu Huang, Nando Metzger, Rodrigo Caye Daudt, and Konrad Schindler. Repurposing diffusion-based image generators for monocular depth estimation. In the IEEE Conference on Computer Vision and Pattern Recognition (CVPR), pages 9492-9502, 2024. 1, 2" + ], + "bbox": [ + 516, + 92, + 905, + 900 + ], + "page_idx": 8 + }, + { + "type": "page_number", + "text": "22262", + "bbox": [ + 478, + 945, + 517, + 955 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[26] Muhammad Ahmed Ullah Khan, Danish Nazir, Alain Pagani, Hamam Mokayed, Marcus Liwicki, Didier Stricker, and Muhammad Zeshan Afzal. A comprehensive survey of depth completion approaches. Sensors, 22(18):6969, 2022. 1, 2", + "[27] Byeong-Uk Lee, Hae-Gon Jeon, Sunghoon Im, and In So Kweon. Depth completion with deep geometry and context guidance. In International Conference on Robotics and Automation (ICRA), pages 3281-3287. IEEE, 2019. 1", + "[28] Byeong-Uk Lee, Kyunghyun Lee, and In So Kweon. Depth completion using plane-residual representation. In the IEEE Conference on Computer Vision and Pattern Recognition (CVPR), pages 13916-13925, 2021. 1", + "[29] Miaoyu Li and Ying Fu. Joint self-attention transformer for multispectral and hyperspectral image fusion. Journal of Image and Graphics, 28(12):3922–3934, 2023. 2", + "[30] Miaoyu Li, Ying Fu, Tao Zhang, and Guanghui Wen. Supervise-assisted self-supervised deep-learning method for hyperspectral image restoration. IEEE Transactions on Neural Networks and Learning Systems, 2024. 1", + "[31] Yingping Liang and Ying Fu. Relation-guided adversarial learning for data-free knowledge transfer. International Journal of Computer Vision, pages 1-18, 2024. 3", + "[32] Yingping Liang, Jiaming Liu, Debing Zhang, and Ying Fu. Mpi-flow: Learning realistic optical flow with multiplane images. In the IEEE International Conference on Computer Vision (ICCV), pages 13857-13868, 2023. 2, 3", + "[33] Tsung-Yi Lin, Michael Maire, Serge Belongie, James Hays, Pietro Perona, Deva Ramanan, Piotr Dólar, and C Lawrence Zitnick. Microsoft coco: Common objects in context. In the European Conference on Computer Vision (ECCV), pages 740-755, 2014. 5", + "[34] Yuankai Lin, Tao Cheng, Qi Zhong, Wending Zhou, and Hua Yang. Dynamic spatial propagation network for depth completion. In the AAAI Conference on Artificial Intelligence (AAAI), pages 1638-1646, 2022. 2, 6", + "[35] Lina Liu, Xibin Song, Xiaoyang Lyu, Junwei Diao, Mengmeng Wang, Yong Liu, and Liangjun Zhang. Fcfr-net: Feature fusion based coarse-to-fine residual learning for depth completion. In the AAAI Conference on Artificial Intelligence (AAAI), pages 2136-2144, 2021. 6", + "[36] Qiankun Liu, Yuqi Jiang, Zhentao Tan, Dongdong Chen, Ying Fu, Qi Chu, Gang Hua, and Nenghai Yu. Transformer based pluralistic image completion with reduced information loss. IEEE Transactions on Pattern Analysis and Machine Intelligence, 46(10):6652-6668, 2024.", + "[37] Qiankun Liu, Yichen Li, Yuqi Jiang, and Ying Fu. Siamesedetr for generic multi-object tracking. IEEE Transactions on Image Processing, 33:3935-3949, 2024.", + "[38] Fangchang Ma and Sertac Karaman. Sparse-to-dense: Depth prediction from sparse depth samples and a single image. In IEEE international conference on robotics and automation (ICRA), pages 4796-4803, 2018. 2, 6", + "[39] Fabiola Maffra, Lucas Teixeira, Zetao Chen, and Margarita Chli. Real-time wide-baseline place recognition using depth completion. IEEE Robotics and Automation Letters, 4(2): 1525-1532, 2019. 1" + ], + "bbox": [ + 91, + 90, + 485, + 898 + ], + "page_idx": 9 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[40] Hyoungseob Park, Anjali Gupta, and Alex Wong. Test-time adaptation for depth completion. In the IEEE Conference on Computer Vision and Pattern Recognition (CVPR), pages 20519-20529, 2024. 2", + "[41] Jinsun Park, Kyungdon Joo, Zhe Hu, Chi-Kuei Liu, and In So Kweon. Non-local spatial propagation network for depth completion. In the European Conference on Computer Vision (ECCV), pages 120-136, 2020. 6", + "[42] Jin-Hwi Park and Hae-Gon Jeon. A simple yet universal framework for depth completion. In The Thirty-eighth Annual Conference on Neural Information Processing Systems, 2024. 2", + "[43] Federico Perazzi, Jordi Pont-Tuset, Brian McWilliams, Luc Van Gool, Markus Gross, and Alexander Sorkine-Hornung. A benchmark dataset and evaluation methodology for video object segmentation. In the IEEE Conference on Computer Vision and Pattern Recognition (CVPR), pages 724-732, 2016. 5", + "[44] Luigi Piccinelli, Yung-Hsu Yang, Christos Sakaridis, Mattia Segu, Siyuan Li, Luc Van Gool, and Fisher Yu. Undepth: Universal monocular metric depth estimation. In the IEEE Conference on Computer Vision and Pattern Recognition (CVPR), pages 10106-10116, 2024. 2", + "[45] Jiaxiong Qiu, Zhaopeng Cui, Yinda Zhang, Xingdi Zhang, Shuaicheng Liu, Bing Zeng, and Marc Pollefeys. Deeplidar: Deep surface normal guided depth prediction for outdoor scene from sparse lidar data and single color image. In the IEEE Conference on Computer Vision and Pattern Recognition (CVPR), pages 3313-3322, 2019. 2, 6", + "[46] René Ranftl, Katrin Lasinger, David Hafner, Konrad Schindler, and Vladlen Koltun. Towards robust monocular depth estimation: Mixing datasets for zero-shot cross-dataset transfer. IEEE Transactions on Pattern Analysis and Machine Intelligence (TPAMI), 44(3), 2022. 1, 2, 3, 5", + "[47] Ali Osman Serhatoglu, Oguzhan Guclu, and Ahmet Burak Can. Rgb-d slam with deep depth completion. In International Conference on Artificial Intelligence and Soft Computing (ICAISC), pages 59-67, 2022. 1", + "[48] Nathan Silberman, Derek Hoiem, Pushmeet Kohli, and Rob Fergus. Indoor segmentation and support inference from rgbd images. In the European Conference on Computer Vision (ECCV), pages 120-136, 2012. 2, 5, 6", + "[49] Jie Tang, Fei-Peng Tian, Wei Feng, Jian Li, and Ping Tan. Learning guided convolutional network for depth completion. IEEE Transactions on Image Processing (TIP), 30: 1116-1129, 2020. 2, 6", + "[50] Jie Tang, Fei-Peng Tian, Boshi An, Jian Li, and Ping Tan. Bilateral propagation network for depth completion. In the IEEE Conference on Computer Vision and Pattern Recognition (CVPR), pages 9763-9772, 2024. 1, 2, 5, 6, 7, 8", + "[51] Zachary Teed and Jia Deng. Droid-slam: Deep visual slam for monocular, stereo, and rgb-d cameras. Advances in Neural Information Processing Systems (NeurIPS), 34:16558-16569, 2021. 7", + "[52] Ye Tian, Ying Fu, and Jun Zhang. Transformer-based undersampled single-pixel imaging. Chinese Journal of Electronics, 32(5):1151-1159, 2023. 2" + ], + "bbox": [ + 516, + 90, + 905, + 900 + ], + "page_idx": 9 + }, + { + "type": "page_number", + "text": "22263", + "bbox": [ + 478, + 945, + 517, + 955 + ], + "page_idx": 9 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[53] Jonas Uhrig, Nick Schneider, Lukas Schneider, Uwe Franke, Thomas Brox, and Andreas Geiger. Sparsity invariant cnns. In International Conference on 3D Vision (3DV), 2017. 1, 2, 3, 5, 6", + "[54] Haotian Wang, Meng Yang, and Nanning Zheng. G2-monodepth: A general framework of generalized depth inference from monocular $\\mathrm{rgb + x}$ data. IEEE Transactions on Pattern Analysis and Machine Intelligence (TPAMI), 2023. 2, 3", + "[55] Haotian Wang, Meng Yang, Xinhu Zheng, and Gang Hua. Scale propagation network for generalizable depth completion. IEEE Transactions on Pattern Analysis and Machine Intelligence (TPAMI), 2024. 2", + "[56] Yufei Wang, Bo Li, Ge Zhang, Qi Liu, Tao Gao, and Yuchao Dai. Lrru: Long-short range recurrent updating networks for depth completion. In the IEEE Conference on Computer Vision and Pattern Recognition (CVPR), pages 9422-9432, 2023. 1, 6, 7, 8", + "[57] Yufei Wang, Ge Zhang, Shaoqian Wang, Bo Li, Qi Liu, Le Hui, and Yuchao Dai. Improving depth completion via depth feature upsampling. In the IEEE Conference on Computer Vision and Pattern Recognition (CVPR), pages 21104-21113, 2024. 1, 2, 5, 6, 7, 8", + "[58] Xiu-Shen Wei, Yi-Zhe Song, Oisin Mac Aodha, Jianxin Wu, Yuxin Peng, Jinhui Tang, Jian Yang, and Serge Belongie. Fine-grained image analysis with deep learning: A survey. IEEE transactions on pattern analysis and machine intelligence, 44(12):8927-8948, 2021. 1", + "[59] Tobias Weyand, Andre Araujo, Bingyi Cao, and Jack Sim. Google landmarks dataset v2-a large-scale benchmark for instance-level recognition and retrieval. In the IEEE Conference on Computer Vision and Pattern Recognition (CVPR), pages 2575-2584, 2020. 5", + "[60] Alex Wong, Xiaohan Fei, Stephanie Tsuei, and Stefano Soatto. Unsupervised depth completion from visual inertial odometry. IEEE Robotics and Automation Letters, 5(2): 1899-1906, 2020. 7", + "[61] Xiaopei Wu, Liang Peng, Honghui Yang, Liang Xie, Chenxi Huang, Chengqi Deng, Haifeng Liu, and Deng Cai. Sparse fuse dense: Towards high quality 3d detection with depth completion. In the IEEE Conference on Computer Vision and Pattern Recognition (CVPR), pages 5418-5427, 2022. 1", + "[62] Yangchao Wu, Tian Yu Liu, Hyoungseob Park, Stefano Soatto, Dong Lao, and Alex Wong. Augundo: Scaling up augmentations for monocular depth completion and estimation. In European Conference on Computer Vision (ECCV), pages 274-293, 2025. 2", + "[63] Weijian Xie, Guanyi Chu, Quanhao Qian, Yihao Yu, Hai Li, Danpeng Chen, Shangjin Zhai, Nan Wang, Hujun Bao, and Guofeng Zhang. Depth completion with multiple balanced bases and confidence for dense monocular slam. IEEE Transactions on Visualization and Computer Graphics (TVCG), 2024. 1", + "[64] Zexiao Xie, Xiaoxuan Yu, Xiang Gao, Kunqian Li, and Shuhan Shen. Recent advances in conventional and deep learning-based depth completion: A survey. IEEE Transactions on Neural Networks and Learning Systems (TNNLS), 35(3):3395-3415, 2024. 2" + ], + "bbox": [ + 91, + 90, + 480, + 898 + ], + "page_idx": 10 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[65] Yan Xu, Xinge Zhu, Jianping Shi, Guofeng Zhang, Hujun Bao, and Hongsheng Li. Depth completion from sparse lidar data with depth-normal constraints. In the IEEE Conference on Computer Vision and Pattern Recognition (CVPR), pages 2811-2820, 2019. 1", + "[66] Zhiqiang Yan, Xiang Li, Kun Wang, Zhenyu Zhang, Jun Li, and Jian Yang. Multi-modal masked pre-training for monocular panoramic depth completion. In European Conference on Computer Vision, pages 378-395. Springer, 2022. 2", + "[67] Zhiqiang Yan, Kun Wang, Xiang Li, Zhenyu Zhang, Jun Li, and Jian Yang. Rignet: Repetitive image guided network for depth completion. In the European Conference on Computer Vision (ECCV), pages 214-230, 2022. 1, 6", + "[68] Zhiqiang Yan, Xiang Li, Kun Wang, Shuo Chen, Jun Li, and Jian Yang. Distortion and uncertainty aware loss for panoramic depth completion. In International Conference on Machine Learning, pages 39099-39109. PMLR, 2023.", + "[69] Zhiqiang Yan, Kun Wang, Xiang Li, Zhenyu Zhang, Jun Li, and Jian Yang. Desnet: Decomposed scale-consistent network for unsupervised depth completion. In Proceedings of the AAAI conference on artificial intelligence, pages 3109-3117, 2023.", + "[70] Zhiqiang Yan, Yuankai Lin, Kun Wang, Yupeng Zheng, Yufei Wang, Zhenyu Zhang, Jun Li, and Jian Yang. Tri-perspective view decomposition for geometry-aware depth completion. In the IEEE Conference on Computer Vision and Pattern Recognition (CVPR), pages 4874-4884, 2024. 1, 2, 6, 7", + "[71] Lihe Yang, Bingyi Kang, Zilong Huang, Xiaogang Xu, Jiashi Feng, and Hengshuang Zhao. Depth anything: Unleashing the power of large-scale unlabeled data. In the IEEE Conference on Computer Vision and Pattern Recognition (CVPR), pages 10371-10381, 2024. 1, 2, 3", + "[72] Lihe Yang, Bingyi Kang, Zilong Huang, Zhen Zhao, Xiaogang Xu, Jiashi Feng, and Hengshuang Zhao. Depth anything v2. In Annual Conference on Neural Information Processing Systems (NeurIPS), 2024. 1, 2, 3, 5", + "[73] Wei Yin, Chi Zhang, Hao Chen, Zhipeng Cai, Gang Yu, Kaixuan Wang, Xiaozhi Chen, and Chunhua Shen. Metric3d: Towards zero-shot metric 3d prediction from a single image. In the IEEE Conference on Computer Vision and Pattern Recognition (CVPR), pages 9043-9053, 2023. 2", + "[74] Zhu Yu, Zehua Sheng, Zili Zhou, Lun Luo, Si-Yuan Cao, Hong Gu, Huaqi Zhang, and Hui-Liang Shen. Aggregating feature point cloud for depth completion. In the IEEE International Conference on Computer Vision (ICCV), pages 8732-8743, 2023. 2, 6", + "[75] Fan Zhang, Shaodi You, Yu Li, and Ying Fu. Atlantis: Enabling underwater depth estimation with stable diffusion. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 11852-11861, 2024. 1", + "[76] Tao Zhang, Ying Fu, and Jun Zhang. Deep guided attention network for joint denoising and demosaicing in real image. Chinese Journal of Electronics, 33(1):303-312, 2024. 2", + "[77] Xuanmeng Zhang, Zhedong Zheng, Minyue Jiang, and Xiaojing Ye. Self-ensembling depth completion via density-aware consistency. Pattern Recognition, 154:110618, 2024. 2" + ], + "bbox": [ + 516, + 90, + 906, + 898 + ], + "page_idx": 10 + }, + { + "type": "page_number", + "text": "22264", + "bbox": [ + 478, + 945, + 519, + 955 + ], + "page_idx": 10 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[78] Yinda Zhang and Thomas Funkhouser. Deep depth completion of a single rgb-d image. In the IEEE Conference on Computer Vision and Pattern Recognition (CVPR), pages 175-185, 2018. 1", + "[79] Youmin Zhang, Xianda Guo, Matteo Poggi, Zheng Zhu, Guan Huang, and Stefano Mattoccia. Completionformer: Depth completion with convolutions and vision transformers. In the IEEE Conference on Computer Vision and Pattern Recognition (CVPR), pages 18527-18536, 2023. 1, 2, 5, 6, 7, 8", + "[80] Shanshan Zhao, Mingming Gong, Huan Fu, and Dacheng Tao. Adaptive context-aware multi-modal network for depth completion. IEEE Transactions on Image Processing (TIP), 30:5264-5276, 2021. 2, 6", + "[81] Jin Zheng, Botao Jiang, Wei Peng, and Qiaohui Zhang. Multi-scale binocular stereo matching based on semantic association. Chinese Journal of Electronics, 33(4):1010-1022, 2024. 2", + "[82] Wending Zhou, Xu Yan, Yinghong Liao, Yuankai Lin, Jin Huang, Gangming Zhao, Shuguang Cui, and Zhen Li. Bev@ dc: Bird's-eye view assisted training for depth completion. In the IEEE Conference on Computer Vision and Pattern Recognition (CVPR), pages 9233-9242, 2023. 2, 6", + "[83] Yunhao Zou, Ying Fu, Tsuyoshi Takatani, and Yinqiang Zheng. Eventhrd: From event to high-speed hdr videos and beyond. IEEE Transactions on Pattern Analysis and Machine Intelligence, 47(1):32-50, 2024. 2", + "[84] Yiming Zuo and Jia Deng. Ogni-dc: Robust depth completion with optimization-guided neural iterations. In the European Conference on Computer Vision (ECCV), pages 78-95, 2025. 2" + ], + "bbox": [ + 91, + 92, + 480, + 526 + ], + "page_idx": 11 + }, + { + "type": "page_number", + "text": "22265", + "bbox": [ + 478, + 945, + 517, + 955 + ], + "page_idx": 11 + } +] \ No newline at end of file diff --git a/2025/Distilling Monocular Foundation Model for Fine-grained Depth Completion/1663a87e-e3f7-4145-9d6a-27f93678241e_model.json b/2025/Distilling Monocular Foundation Model for Fine-grained Depth Completion/1663a87e-e3f7-4145-9d6a-27f93678241e_model.json new file mode 100644 index 0000000000000000000000000000000000000000..832fa252f8fc844379eb4d5154da9ba571802b2a --- /dev/null +++ b/2025/Distilling Monocular Foundation Model for Fine-grained Depth Completion/1663a87e-e3f7-4145-9d6a-27f93678241e_model.json @@ -0,0 +1,3062 @@ +[ + [ + { + "type": "header", + "bbox": [ + 0.107, + 0.003, + 0.182, + 0.044 + ], + "angle": 0, + "content": "CVF" + }, + { + "type": "header", + "bbox": [ + 0.237, + 0.001, + 0.812, + 0.048 + ], + "angle": 0, + "content": "This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore." + }, + { + "type": "title", + "bbox": [ + 0.116, + 0.131, + 0.882, + 0.154 + ], + "angle": 0, + "content": "Distilling Monocular Foundation Model for Fine-grained Depth Completion" + }, + { + "type": "text", + "bbox": [ + 0.1, + 0.179, + 0.909, + 0.253 + ], + "angle": 0, + "content": "Yingping Liang1 Yutao Hu2 Wenqi Shao3 Ying Fu1† \n1Beijing Institute of Technology 2Key Laboratory of New Generation Artificial Intelligence Technology and Its Interdisciplinary Applications, Southeast University 3Shanghai Al Laboratory {liangyingping, fuying}@bit.edu.cn huyutao@seu.edu.cn weqish@link.cuhk.edu.hk" + }, + { + "type": "image", + "bbox": [ + 0.114, + 0.287, + 0.889, + 0.416 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.089, + 0.427, + 0.908, + 0.471 + ], + "angle": 0, + "content": "Figure 1. Depth completion models trained solely with \\( L_{1} \\) loss and sparse ground truth produce incomplete and fragmented depth predictions. Our framework, however, demonstrates significant improvements by distilling knowledge from monocular foundation models and incorporating a scale- and shift-invariant loss (SSI Loss), resulting in more complete and accurate dense depth completion." + }, + { + "type": "title", + "bbox": [ + 0.248, + 0.482, + 0.327, + 0.499 + ], + "angle": 0, + "content": "Abstract" + }, + { + "type": "text", + "bbox": [ + 0.088, + 0.517, + 0.486, + 0.866 + ], + "angle": 0, + "content": "Depth completion involves predicting dense depth maps from sparse LiDAR inputs. However, sparse depth annotations from sensors limit the availability of dense supervision, which is necessary for learning detailed geometric features. In this paper, we propose a two-stage knowledge distillation framework that leverages powerful monocular foundation models to provide dense supervision for depth completion. In the first stage, we introduce a pre-training strategy that generates diverse training data from natural images, which distills geometric knowledge to depth completion. Specifically, we simulate LiDAR scans by utilizing monocular depth and mesh reconstruction, thereby creating training data without requiring ground-truth depth. Besides, monocular depth estimation suffers from inherent scale ambiguity in real-world settings. To address this, in the second stage, we employ a scale- and shift-invariant loss (SSI Loss) to learn real-world scales when fine-tuning on real-world datasets. Our two-stage distillation framework enables depth completion models to harness the strengths of monocular foundation models. Experimental results demonstrate that models trained with our two-stage distillation framework achieve state-of-the-art performance, ranking first place on the KITTI benchmark. Code is available" + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.484, + 0.88, + 0.499 + ], + "angle": 0, + "content": "at https://github.com/Sharpiless/DMD3C" + }, + { + "type": "title", + "bbox": [ + 0.514, + 0.527, + 0.646, + 0.543 + ], + "angle": 0, + "content": "1. Introduction" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.553, + 0.907, + 0.629 + ], + "angle": 0, + "content": "Depth completion is a fundamental task in computer vision, where the goal is to generate dense depth maps from sparse depth measurements. This task is particularly important in applications such as autonomous driving [23, 47, 61], robotics [26, 39, 63], and augmented reality [21, 30]." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.63, + 0.909, + 0.734 + ], + "angle": 0, + "content": "Recent approaches [7, 27, 28, 65, 67, 78] leverage deep neural networks to learn from depth data. Some methods [56, 79] also incorporate RGB images to guide the depth completion process. Despite these advancements, state-of-the-art models [50, 57, 70] still face difficulties in capturing fine-grained geometric details, particularly in complex outdoor scenes where depth annotations are sparse [53, 58, 75]." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.735, + 0.909, + 0.885 + ], + "angle": 0, + "content": "The challenge arises from the reliance on sparse ground truth for training. The lack of dense ground truth makes it difficult for models to accurately learn depth completion across an entire scene. Monocular depth estimation [25, 46, 71, 72], on the other hand, has the ability to provide dense depth predictions from a single image. Specifically, state-of-the-art monocular foundation models could generate dense depth maps, containing fine-grained details and relative depth relationships, offering valuable guidance for training depth completion networks." + }, + { + "type": "text", + "bbox": [ + 0.533, + 0.886, + 0.907, + 0.902 + ], + "angle": 0, + "content": "To make full use of the advantages of monocular foun" + }, + { + "type": "page_footnote", + "bbox": [ + 0.115, + 0.888, + 0.246, + 0.901 + ], + "angle": 0, + "content": "† Corresponding author." + }, + { + "type": "page_number", + "bbox": [ + 0.479, + 0.945, + 0.521, + 0.958 + ], + "angle": 0, + "content": "22254" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.09, + 0.092, + 0.482, + 0.288 + ], + "angle": 0, + "content": "dation models, we propose a novel two-stage distillation framework to transfer geometric knowledge from monocular foundation models to depth completion networks. Specifically, in the first distillation stage, we generate training data through monocular foundation models and then distill knowledge via the proposed pre-training strategy. Specifically, diverse natural images are used to generate pseudo depth maps. Then, we utilize randomly sampled camera parameters to re-construct the scene with mesh and simulate LiDAR using ray simulation. The generated data trains the depth completion model to learn diverse geometric knowledge from monocular foundation models, enhancing its ability to generalize across different scenes." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.289, + 0.482, + 0.456 + ], + "angle": 0, + "content": "However, monocular depth estimation suffers from inherent scale ambiguity [16, 81], resulting in depth predictions that vary greatly in scale. Thus, monocular depth alone cannot serve as a reliable basis for real-world depth. To solve this problem, in the second distillation stage, we introduce a scale- and shift-invariant loss (SSI Loss) [46]. Specifically, when fine-tuning on labeled datasets with sparse ground truth, SSI Loss ignores the scale and shift that causes the least loss from the depth prediction and monocular depth, to ensure consistent depth completions aligning monocular depth supervision across varying scales." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.455, + 0.482, + 0.515 + ], + "angle": 0, + "content": "By integrating these two distillation stages into the training process, our method achieves state-of-the-art performance on the KITTI benchmark. The contributions of this work are summarized as follows:" + }, + { + "type": "text", + "bbox": [ + 0.091, + 0.52, + 0.483, + 0.594 + ], + "angle": 0, + "content": "- We propose a novel two-stage distillation framework to transfer knowledge from monocular foundation models to depth completion models, which enables the learning of fine-grained depth information from sparse ground truth by providing dense supervision." + }, + { + "type": "text", + "bbox": [ + 0.091, + 0.596, + 0.483, + 0.669 + ], + "angle": 0, + "content": "- In the first stage, we propose a data generation strategy that uses monocular depth estimation and mesh reconstruction to simulate training data, enabling the model to learn geometric features from diverse natural images without the need for any LiDAR and ground truth." + }, + { + "type": "text", + "bbox": [ + 0.091, + 0.671, + 0.483, + 0.745 + ], + "angle": 0, + "content": "- In the second stage, we propose a scale- and shift-invariant loss (SSI Loss) to address the scale ambiguity problem in monocular depth estimation, which ensures consistent depth completions across varying scales and focuses on learning real-world scale information." + }, + { + "type": "text", + "bbox": [ + 0.091, + 0.746, + 0.483, + 0.806 + ], + "angle": 0, + "content": "- Our method achieves first place on the official KITTI depth completion benchmark, demonstrating the effectiveness of the proposed approach with significant gains in both quantitative metrics and qualitative visualizations." + }, + { + "type": "list", + "bbox": [ + 0.091, + 0.52, + 0.483, + 0.806 + ], + "angle": 0, + "content": null + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.822, + 0.233, + 0.837 + ], + "angle": 0, + "content": "2. Related Work" + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.847, + 0.338, + 0.863 + ], + "angle": 0, + "content": "2.1. Depth Completion Datasets" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.871, + 0.483, + 0.901 + ], + "angle": 0, + "content": "Depth completion has been widely studied in both indoor and outdoor settings. Indoor datasets, such as NYU Depth" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.092, + 0.905, + 0.152 + ], + "angle": 0, + "content": "V2 [48] and Matterport3D [4], are popular because dense depth measurements are easier to obtain in controlled environments. NYU Depth V2, for instance, provides dense depth maps from a Kinect sensor along with RGB images." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.153, + 0.906, + 0.364 + ], + "angle": 0, + "content": "In outdoor environments, however, obtaining dense depth annotations is more challenging due to limitations in sensing technologies like LiDAR. The KITTI dataset [53] is a key benchmark for outdoor depth completion but suffers from sparse annotations, covering only about \\(5\\%\\) of the image. To address this, training methods often employ complex post-processing and multi-frame fusion techniques to increase the annotation coverage to around \\(20\\%\\) at most. Furthermore, ground truth values for long distances and dynamic objects typically cannot be included. The sparse ground truth raises challenges for training fine-grained depth completion models. Therefore, we employ monocular depth estimation techniques for dense distillation to preserve fine-grained details." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.375, + 0.763, + 0.391 + ], + "angle": 0, + "content": "2.2. Depth Completion Methods" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.399, + 0.906, + 0.671 + ], + "angle": 0, + "content": "Depth completion methods have seen significant advancements [21, 26, 64]. Compared to predicting depth directly from a single RGB image [2, 22, 44, 73], depth completion fuses LiDAR information to obtain more accurate sparse depth cues. Recent methods [1, 6, 8, 9, 11, 14, 29, 34-38, 40-42, 45, 49, 50, 52, 55-57, 62, 66-70, 74, 76, 79, 80, 82, 84] rely on supervised learning, using pairs of sparse depth maps and corresponding dense ground truth. While effective with sufficient labeled data, these methods are limited in real-world applications, where obtaining dense annotations, especially in outdoor environments, is challenging [5, 15, 83]. To reduce the reliance on dense labels, [10, 77] use a teacher-student model, where a teacher network trained on labeled data provides guidance to a student network. However, pseudo-labels in complex or dynamic scenes can be noisy or inaccurate. Therefore, we turn to monocular depth estimation to provide fine-grained supervision for depth completion." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.681, + 0.774, + 0.697 + ], + "angle": 0, + "content": "2.3. Monocular Depth Estimation" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.705, + 0.906, + 0.901 + ], + "angle": 0, + "content": "Monocular depth estimation [25, 46, 71, 72] has emerged as a powerful tool for generating dense depth maps from single RGB images, making it particularly valuable in scenarios where dense depth annotations are sparse or unavailable [18, 19, 32]. Notable models, such as MiDaS [46] and Depth-Anything [71, 72], have demonstrated impressive capabilities in predicting dense depth maps across a wide range of scenes. G2-MonoDepth [54] have proposed a general framework for monocular RGB-X Data with a scale invariant loss. Despite their versatility, monocular depth estimation methods face the challenge of scale ambiguity, which can limit the accuracy of depth completion, especially when integrated with sparse depth measurements" + }, + { + "type": "page_number", + "bbox": [ + 0.48, + 0.946, + 0.518, + 0.957 + ], + "angle": 0, + "content": "22255" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.154, + 0.088, + 0.848, + 0.284 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.089, + 0.303, + 0.908, + 0.348 + ], + "angle": 0, + "content": "Figure 2. Illustration of our proposed first distillation stage with a data generation strategy to learn geometric features from monocular foundation models, which only requires unlabeled RGB images. We use the estimated monocular depth to re-construct the scene and then simulate the Lidar swap process to generate sparse points for training." + }, + { + "type": "text", + "bbox": [ + 0.089, + 0.372, + 0.482, + 0.432 + ], + "angle": 0, + "content": "from LiDAR. Therefore, we introduce a scale- and shift-invariant loss inspired by [54], combined with a supervised loss to keep real-world scales when fine-tuning on real-world datasets." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.446, + 0.182, + 0.46 + ], + "angle": 0, + "content": "3. Method" + }, + { + "type": "text", + "bbox": [ + 0.089, + 0.471, + 0.483, + 0.607 + ], + "angle": 0, + "content": "In this section, we first provide a brief overview of the motivation and formulation of our method. Then, we introduce our two-stage distillation framework. In the first stage, we propose a data generation strategy for pre-training that leverages large-scale, unlabeled images without ground-truth depth annotations, as illustrated in Figure 2. In the second stage, we detail the fine-tuning process on a labeled dataset with sparse ground truth, using monocular depth in combination with SSI Loss, as shown in Figure 3." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.616, + 0.347, + 0.631 + ], + "angle": 0, + "content": "3.1. Motivation and Formulation" + }, + { + "type": "text", + "bbox": [ + 0.089, + 0.64, + 0.483, + 0.746 + ], + "angle": 0, + "content": "In the depth completion task, the input typically consists of an RGB image \\(I\\) and a sparse depth map \\(D_{s}\\) obtained from a LiDAR sensor. The objective is to produce a dense depth map \\(D_{f}\\) that provides a complete and detailed representation of the scene's depth. To achieve this, a depth completion model \\(f_{\\theta}\\), parameterized by \\(\\theta\\), takes both the RGB image and sparse depth map as inputs:" + }, + { + "type": "equation", + "bbox": [ + 0.212, + 0.758, + 0.483, + 0.774 + ], + "angle": 0, + "content": "\\[\nD _ {f} = f _ {\\theta} (I, D _ {s}). \\tag {1}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.786, + 0.483, + 0.816 + ], + "angle": 0, + "content": "To guide the model's learning, a sparse supervision loss is defined based on the available sparse depth annotations:" + }, + { + "type": "equation", + "bbox": [ + 0.174, + 0.828, + 0.483, + 0.845 + ], + "angle": 0, + "content": "\\[\n\\mathcal {L} _ {\\sup } = M \\times \\left| D _ {f} - D _ {\\text {s p a r s e}} \\right|, \\tag {2}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.856, + 0.484, + 0.903 + ], + "angle": 0, + "content": "where \\(M\\) represents the valid mask for sparse depth ground truth \\(D_{\\mathrm{sparse}}\\). This loss aligns the predicted dense depth map \\(D_{f}\\) with the sparse ground truth, ensuring that the output" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.372, + 0.906, + 0.448 + ], + "angle": 0, + "content": "is consistent with the available depth data. However, this sparse supervision provides limited guidance, particularly in outdoor scenes [53], leading to challenges in achieving fine-grained and consistent depth predictions due to the inherent limitations of sparse supervision." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.452, + 0.907, + 0.664 + ], + "angle": 0, + "content": "Foundation models for monocular depth estimation [71, 72] provide an alternative approach by generating fine-grained dense depth from single RGB images, which we use as supervision without relying solely on LiDAR. Therefore, in the first stage, we utilize the monocular foundation models to generate diverse and large-scale training data, to provide dense supervision for pre-training, allowing the model to learn geometric features. However, monocular methods suffer from scale ambiguities, which can lead to inconsistencies when combining dense monocular predictions with sparse depth. To address this, in the second stage, we introduce SSI Loss [46] combined with L1 loss when finetuning on labeled datasets with sparse ground truth to mitigate scale inconsistencies and learn real-world scale." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.687, + 0.905, + 0.704 + ], + "angle": 0, + "content": "3.2. First Stage: Data Generation and Pre-training" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.715, + 0.906, + 0.837 + ], + "angle": 0, + "content": "To leverage diverse images without depth annotations, we introduce a pre-training strategy that utilizes synthesized depth data generated through monocular depth estimation on large-scale natural image datasets inspired by [31, 32]. As illustrated in Figure 2, this pre-training phase enables the model to learn robust geometric features from monocular foundation models across diverse scenes, preparing it for subsequent fine-tuning on real-world datasets." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.84, + 0.907, + 0.902 + ], + "angle": 0, + "content": "Specifically, we first utilize a pre-trained monocular depth estimation model, such as Depth Anything V2 [72], to predict depth maps for such images. For each natural image \\( I_{\\mathrm{un}} \\), the monocular depth estimation model predicts a" + }, + { + "type": "page_number", + "bbox": [ + 0.479, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "22256" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.132, + 0.089, + 0.868, + 0.245 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.089, + 0.267, + 0.908, + 0.337 + ], + "angle": 0, + "content": "Figure 3. Illustration of our proposed second distillation stage utilizing foundation models for monocular depth estimation when fine-tuning on labeled datasets. Sparse ground truth provides real-world depth scale with L1 loss. Our method enhances this process by incorporating dense monocular depth for fine-grained supervision. However, monocular depth maps come with inherent scale and shift ambiguities. To address these challenges, we employ a Scale- and Shift-Invariant Loss (SSI Loss) that aligns the predictions with the dense monocular depth to match the real-world depth scale, ensuring more accurate depth completion." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.364, + 0.343, + 0.38 + ], + "angle": 0, + "content": "corresponding dense depth map \\(D_{\\mathrm{syn}}\\) .." + }, + { + "type": "equation", + "bbox": [ + 0.212, + 0.392, + 0.482, + 0.409 + ], + "angle": 0, + "content": "\\[\nD _ {\\text {s y n}} = f _ {\\mathrm {m o}} \\left(I _ {\\mathrm {u n}}\\right), \\tag {3}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.421, + 0.484, + 0.496 + ], + "angle": 0, + "content": "where \\( f_{\\mathrm{mo}} \\) represents the monocular depth estimation model. These synthesized depth maps, though not accurate in metric scale, capture the relative depth relationships and structural details in the scene, providing valuable supervision signals during pre-training." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.497, + 0.484, + 0.527 + ], + "angle": 0, + "content": "To simulate LiDAR scanning, we first sample a random camera intrinsic matrix \\( K \\) as:" + }, + { + "type": "equation", + "bbox": [ + 0.195, + 0.537, + 0.482, + 0.587 + ], + "angle": 0, + "content": "\\[\nK = \\left[ \\begin{array}{c c c} f _ {x} & 0 & c _ {x} \\\\ 0 & f _ {y} & c _ {y} \\\\ 0 & 0 & 1 \\end{array} \\right], \\tag {4}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.597, + 0.483, + 0.627 + ], + "angle": 0, + "content": "where \\( f_{x} \\) and \\( f_{y} \\) represent the focal lengths along the \\( x \\)- and \\( y \\)-axes, and \\( (c_{x}, c_{y}) \\) denotes the center of the image." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.628, + 0.483, + 0.763 + ], + "angle": 0, + "content": "For each pixel \\((u,v)\\) in the depth map, the corresponding 3D coordinates \\((X,Y,Z)\\) in the camera coordinate system are derived using the sampled camera intrinsic matrix \\(K\\). This transformation generates a point cloud \\(P = \\{(X,Y,Z)\\}\\) that represents the 3D spatial locations of each pixel. Leveraging the 3D point cloud \\(P\\), we reconstruct a mesh \\(M\\) via surface reconstruction techniques, such as Poisson Surface Reconstruction [24], to create a continuous 3D surface model for simulation." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.765, + 0.483, + 0.884 + ], + "angle": 0, + "content": "Next, we simulate a LiDAR sensor by generating a set of ray direction vectors \\(\\mathbf{d}\\). Each ray is cast from the origin along \\(\\mathbf{d}_{i,j}\\), intersecting the mesh \\(M\\) to determine the distance to the intersection point. This distance is recorded as the simulated LiDAR depth reading \\(D_{\\mathrm{simu}}(i,j)\\). The resulting sparse depth map \\(D_{\\mathrm{simu}}\\) emulates LiDAR depth readings from the reconstructed 3D surface, to provide a synthetic depth image comparable to actual LiDAR data." + }, + { + "type": "text", + "bbox": [ + 0.11, + 0.886, + 0.483, + 0.901 + ], + "angle": 0, + "content": "Then, the depth completion model \\( f_{\\theta} \\) takes the RGB im" + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.364, + 0.905, + 0.379 + ], + "angle": 0, + "content": "age and the sparse depth from simulated scanning as inputs:" + }, + { + "type": "equation", + "bbox": [ + 0.621, + 0.388, + 0.905, + 0.404 + ], + "angle": 0, + "content": "\\[\nD _ {f} = f _ {\\theta} \\left(I _ {\\mathrm {u n}}, D _ {\\mathrm {s i m u}}\\right), \\tag {5}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.413, + 0.905, + 0.458 + ], + "angle": 0, + "content": "where \\(D_{\\mathrm{simu}}\\) is the sparse input from simulated LiDAR scanning. Afterwards, the depth completion model is pretrained using L1 Loss:" + }, + { + "type": "equation", + "bbox": [ + 0.641, + 0.467, + 0.905, + 0.484 + ], + "angle": 0, + "content": "\\[\n\\mathcal {L} _ {\\text {p r e}} = \\left| D _ {f} - D _ {m} \\right|. \\tag {6}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.492, + 0.906, + 0.614 + ], + "angle": 0, + "content": "This pre-training process encourages the depth completion model \\( f_{\\theta} \\) to align its predictions with the synthesized depth maps, learning meaningful geometric features from the generated training data using monocular foundation models and diverse natural image data. Additionally, the model learns complex geometric structures and relative depth relationships across varying scenes, which are crucial for accurate depth completion in the fine-tuning phase." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.621, + 0.755, + 0.637 + ], + "angle": 0, + "content": "3.3. Second Stage: Fine-tuning" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.645, + 0.907, + 0.825 + ], + "angle": 0, + "content": "The depth completion model pre-trained on the generated data learns powerful geometric features from the monocular foundation models. However, due to the inherent scale ambiguity of monocular depth, the model prediction is not consistent with the scale in the real world. To learn the real-world scale, we introduce a combined loss for fine-tuning on labeled datasets. To be specific, we first utilize the sparse ground truth \\( D_{s} \\) for supervised loss \\( \\mathcal{L}_{\\mathrm{sup}} \\) to focus on real-world depth, as in Equation 2. Then, the dense monocular depth \\( D_{m} \\) predicted by monocular depth models is also used to provide approximate depth values across the image, especially in regions not covered by the ground truth \\( D_{s} \\)." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.826, + 0.907, + 0.901 + ], + "angle": 0, + "content": "However, the key challenge in using monocular depth estimation as supervision is the inherent scale ambiguity in depth predictions. To address this issue, we incorporate a Scale- and Shift-Invariant Loss (SSI Loss) into our distillation framework. The SSI Loss is designed to align the" + }, + { + "type": "page_number", + "bbox": [ + 0.479, + 0.945, + 0.518, + 0.957 + ], + "angle": 0, + "content": "22257" + } + ], + [ + { + "type": "table_caption", + "bbox": [ + 0.093, + 0.09, + 0.48, + 0.104 + ], + "angle": 0, + "content": "Table 1. Unlabeled datasets used for pre-training data generation." + }, + { + "type": "table", + "bbox": [ + 0.095, + 0.115, + 0.477, + 0.224 + ], + "angle": 0, + "content": "
DatasetIndoorOutdoor# Images
COCO [33]✓✓118,287
Google Landmarks [59]✓117,576
Nuscenes [3]✓93,475
Cityscapes [12]✓19,998
DAVIS [43]✓✓10,581
" + }, + { + "type": "text", + "bbox": [ + 0.091, + 0.248, + 0.483, + 0.294 + ], + "angle": 0, + "content": "predicted depth map \\(D_{f}\\) with the dense monocular depth \\(D_{m}\\). It remains invariant to differences in scale and shift between \\(D_{f}\\) and \\(D_{m}\\). The SSI Loss is formulated as:" + }, + { + "type": "equation", + "bbox": [ + 0.169, + 0.306, + 0.483, + 0.328 + ], + "angle": 0, + "content": "\\[\n\\mathcal {L} _ {\\text {S S I L}} = \\min _ {s, b} | D _ {f} - (s \\cdot D _ {m} + b) |, \\tag {7}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.34, + 0.484, + 0.507 + ], + "angle": 0, + "content": "where \\( s \\) and \\( b \\) are the optimal scale and shift parameters that align the predicted depth \\( D_{f} \\) with the dense monocular depth \\( D_{m} \\). The loss function seeks to find the best alignment between \\( D_{f} \\) and \\( D_{m} \\), effectively normalizing any global differences in scale and offset. By minimizing this loss across all pixels in the image, the model is encouraged to produce depth maps that maintain consistency with the relative depth structure provided by the monocular depth estimates. In addition, we adapt a gradient matching term to preserve sharpness and align with with depth discontinuities. The gradient matching term is defined as:" + }, + { + "type": "equation", + "bbox": [ + 0.164, + 0.518, + 0.483, + 0.559 + ], + "angle": 0, + "content": "\\[\n\\mathcal {L} _ {\\text {r e g}} = \\frac {1}{N} \\sum_ {k = 1} ^ {K} \\left(| \\nabla_ {x} R ^ {k} | + | \\nabla_ {y} R ^ {k} |\\right), \\tag {8}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.571, + 0.484, + 0.649 + ], + "angle": 0, + "content": "where \\( R = D_{f} - D_{m} \\), and \\( R^{k} \\) denotes the difference of depth maps at scale \\( k \\). We use \\( K = 4 \\) scale levels as in [46], halving the image resolution at each level. The final objective function combines the supervised loss \\( \\mathcal{L}_{\\mathrm{sup}} \\), the dense distillation loss \\( \\mathcal{L}_{\\mathrm{SSIL}} \\), and the regularization term \\( \\mathcal{L}_{\\mathrm{reg}} \\)." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.661, + 0.224, + 0.679 + ], + "angle": 0, + "content": "4. Experiments" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.687, + 0.484, + 0.793 + ], + "angle": 0, + "content": "In this section, we first introduce the details of our implementation, as well as the datasets and evaluation metrics for experiments. Then, detailed comparisons are conducted with the state-of-the-art methods. Finally, ablations and discussions are performed to confirm the effectiveness of our proposed main components. Additional analysis is provided in the supplementary materials, along with videos." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.802, + 0.3, + 0.818 + ], + "angle": 0, + "content": "4.1. Experimental Pipeline" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.826, + 0.484, + 0.902 + ], + "angle": 0, + "content": "First stage involves utilizing our proposed data generation strategy to train the depth completion model from scratch, as described in Section 3.2. In this stage, we generate training data using the RGB images in Table 1 and train the depth completion model using SSI Loss, which does not require" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.092, + 0.905, + 0.151 + ], + "angle": 0, + "content": "any ground truth depth. This stage allows the pre-trained model to focus on learning diverse geometric features without being constrained by labeled data, enhancing generalization across different scenes." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.152, + 0.907, + 0.32 + ], + "angle": 0, + "content": "Second stage involves adapting the pre-trained model on labeled datasets, using supervised loss (L1 loss) with depth annotations and our proposed SSI Loss for knowledge distillation, as described in Section 3.3. Using L1 loss with sparse ground truth allows the depth completion model to adapt to the real-world depth scale under sparse supervision. Specifically, our proposed monocular model distillation uses SSI Loss to help the model maintain fine-grained detail with dense supervision by distilling monocular foundation models. Fine-tuning and validation are performed on the KITTI and NYU Depth V2 datasets, respectively." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.33, + 0.731, + 0.345 + ], + "angle": 0, + "content": "4.2. Implementation Details" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.352, + 0.907, + 0.656 + ], + "angle": 0, + "content": "We follow the setup of recent work [50, 57, 79] and train our model on 4 NVIDIA RTX A100 GPUs. For the monocular foundation model used in distillation, we adopt Depth Anything V2 [72] due to its robust performance. For the depth completion network architecture, we use BP-Net [50] as our base model, which is primarily composed of ResNet [20] blocks. In the first stage, we utilize a mixed dataset collected from a wide range of resources, as shown in Table 1. We select five large-scale public datasets as our unlabeled sources for their diverse scenes, totaling approximately 360,000 images with varying scenes and image scales. We adopt AdamW with a weight decay of 0.05 as the optimizer and clip gradients if their norms exceeding 0.1. In the first stage, our model is pre-trained from scratch for 600K iterations. In the second stage, the pre-trained model is fine-tuned on the labeled datasets for 300K iterations using a OneCycle learning rate policy. We set the batch size to 16. The final model is obtained by applying Exponential Moving Average (EMA) with a decay of 0.9999, which helps stabilize the model parameters during training." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.666, + 0.8, + 0.681 + ], + "angle": 0, + "content": "4.3. Datasets and Evaluation Metrics" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.689, + 0.906, + 0.854 + ], + "angle": 0, + "content": "KITTI Depth Completion Dataset [53] serves as a standard benchmark for depth completion in autonomous driving scenarios. It provides over 93,000 training samples, including sparse LiDAR depth maps, their corresponding RGB images, and ground truth depth maps. The sparse LiDAR depth maps are generated from Velodyne LiDAR scans, which typically cover only a small percentage of the image pixels (approximately \\(5\\%\\)). The ground truth depth maps are obtained via multi-frame matching and cover approximately \\(20\\%\\) image pixels. The dataset also includes 1,000 test samples on the private online benchmark." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.856, + 0.906, + 0.901 + ], + "angle": 0, + "content": "NYU Depth V2 Dataset [48] is also a standard benchmark for indoor scene understanding, consisting of RGB-D data captured with a Kinect sensor. It provides 1,449 densely" + }, + { + "type": "page_number", + "bbox": [ + 0.48, + 0.946, + 0.519, + 0.957 + ], + "angle": 0, + "content": "22258" + } + ], + [ + { + "type": "table_caption", + "bbox": [ + 0.09, + 0.089, + 0.908, + 0.119 + ], + "angle": 0, + "content": "Table 2. Performance on KITTI and NYUv2 datasets. For the KITTI dataset, results are evaluated by the KITTI testing server and ranked by the RMSE (in mm). For the NYUv2 dataset, we report their performance on the official test set in their papers." + }, + { + "type": "table", + "bbox": [ + 0.179, + 0.128, + 0.819, + 0.436 + ], + "angle": 0, + "content": "
MethodYearKITTINYUv2
RMSE ↓MAE ↓iRMSE ↓iMAE ↓RMSE ↓REL ↓δ1.25 ↑
S2D [38]2018814.73249.952.801.210.2300.04497.1
CSPN [8]20191019.64279.462.931.150.1170.01699.2
DeepLiDAR [45]2019758.38226.502.561.150.1150.02299.3
CSPN++ [9]2020743.69209.282.070.900.1010.01599.5
GuideNet [49]2020736.24218.832.250.990.115--
FCFR [35]2021735.81217.152.200.980.1060.01599.5
ACMNet [80]2021744.91206.092.080.900.1050.01599.4
NLSPN [41]2020741.68199.591.990.840.0920.01299.6
PointDC [74]2023736.07201.871.970.870.0890.01299.6
RigNet [67]2022712.66203.252.080.900.0900.01399.6
DySPN [34]2022709.12192.711.880.820.0900.01299.6
BEV@DC [82]2023697.44189.441.830.820.0890.01299.6
CFormaler [79]2023708.87203.452.010.880.0900.012-
LRRU [56]2023696.51189.961.870.810.0910.01199.6
TPVD [70]2024693.97188.601.820.810.0860.01099.7
ImprovingDC [57]2024686.46187.951.830.810.0910.01199.6
BP-Net [50]2024684.90194.691.820.840.0890.01299.6
DMD3C (Ours)-678.12194.461.820.850.0850.01199.7
" + }, + { + "type": "text", + "bbox": [ + 0.089, + 0.46, + 0.482, + 0.505 + ], + "angle": 0, + "content": "labeled images from various indoor scenes. Missing ground truth is mainly due to the difference in perspective between the depth sensor and the camera." + }, + { + "type": "text", + "bbox": [ + 0.089, + 0.506, + 0.483, + 0.642 + ], + "angle": 0, + "content": "Evaluation Metrics used for KITTI are root mean squared error (RMSE), mean absolute error (MAE), root mean squared error of the inverse depth (iRMSE), and mean absolute error of the inverse depth (iMAE), among which RMSE is used for ranking. These metrics are obtained by submission and evaluation from a non-public test set. For NYU, we follow the previous work and report the RMSE, mean absolute relative error (REL), and \\(\\delta_{\\theta}\\), which represents the percentage of pixels whose error is less than a threshold \\(\\theta\\)." + }, + { + "type": "title", + "bbox": [ + 0.09, + 0.651, + 0.456, + 0.667 + ], + "angle": 0, + "content": "4.4. Comparison with State-of-the-art Methods" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.674, + 0.483, + 0.856 + ], + "angle": 0, + "content": "Evaluation on Depth Completion Benchmarks. We evaluate our method on the test sets of the NYUv2 dataset [48] and KITTI dataset [53]. Table 2 shows the quantitative comparison of our method and other top ranking published methods. On the KITTI leaderboard, our proposed \\(\\mathrm{DMD}^{3}\\mathrm{C}\\) ranks 1st outperforming all other methods under the primary RMSE metric at the time of paper submission. It also has comparable performance under other evaluation metrics. On the NYUv2 dataset, our method achieves the best RMSE and best \\(\\delta_{1.25}\\). Additionally, \\(\\mathrm{DMD}^{3}\\mathrm{C}\\) maintains competitive results across several other evaluation metrics, further confirming its robustness and generalization." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.856, + 0.484, + 0.901 + ], + "angle": 0, + "content": "To enhance understanding of \\(\\mathrm{DMD}^3\\mathrm{C}\\) 's superiority over other state-of-the-art methods, we provide visual comparisons. Figure 4 provides qualitative comparisons with other" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.46, + 0.907, + 0.641 + ], + "angle": 0, + "content": "SOTA methods on the KITTI test set using publicly available results. We use results from the official benchmark1 produced by the best models for method comparison. Our \\(\\mathrm{DMD}^{3}\\mathrm{C}\\) model excels in maintaining sharp object boundaries and capturing fine details in areas where other models encounter difficulties, particularly in scenes with complex structures or objects at varying distances. In such challenging regions, other methods fail to estimate accurate depth. Furthermore, as shown in the error maps, our method demonstrates a clear advantage even in regions without ground truth, despite these areas not being factored into the evaluation metrics." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.642, + 0.909, + 0.765 + ], + "angle": 0, + "content": "Not only does our model retain the fine details in the depth map, but it also maintains structural integrity in the potential three-dimensional space. Figure 5 illustrates that our proposed two-stage distillation improves performance in the 3D space by visualizing point clouds from the completed depth maps. In comparison to other methods, our \\(\\mathrm{DMD}^{3}\\mathrm{C}\\) produces a more coherent and complete 3D structure, further outperforming existing methods." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.775, + 0.64, + 0.79 + ], + "angle": 0, + "content": "4.5. Discussions" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.799, + 0.908, + 0.877 + ], + "angle": 0, + "content": "Effect of Pre-training in the First Stage. First, to evaluate the impact of the proposed pre-training strategy using unlabeled images, we conduct an experiment where the depth completion model is trained from scratch, without any pretraining. As shown in Table 3, removing the pre-training" + }, + { + "type": "page_footnote", + "bbox": [ + 0.532, + 0.887, + 0.841, + 0.9 + ], + "angle": 0, + "content": "MethodKITTIRMSE ↓MAE ↓iRMSE ↓iMAE ↓w/o Pre-train682.34194.961.820.84w/o SSI Loss684.54195.651.860.85DMD3C678.12194.461.820.85" + }, + { + "type": "table_caption", + "bbox": [ + 0.127, + 0.519, + 0.446, + 0.531 + ], + "angle": 0, + "content": "Table 4. Ablations on different network architectures." + }, + { + "type": "table", + "bbox": [ + 0.098, + 0.544, + 0.478, + 0.691 + ], + "angle": 0, + "content": "
MethodKITTI
RMSE ↓MAE ↓iRMSE ↓iMAE ↓
LRRU [56]696.51189.961.870.81
+ Ours693.17189.601.850.80
CFormer [79]764.87183.881.890.80
+ Ours760.29183.621.880.80
BP-Net [50]684.90194.691.820.84
+ Ours678.12194.461.820.85
" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.719, + 0.483, + 0.793 + ], + "angle": 0, + "content": "step results in a performance degradation, with an increase in RMSE by \\(4.22\\mathrm{mm}\\). This demonstrates that pre-training plays a crucial role in enhancing the model's ability to generalize, enabling it to better capture complex geometric features across diverse scenes." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.795, + 0.484, + 0.902 + ], + "angle": 0, + "content": "Effect of SSI Loss in the Second Stage. Next, we show the impact of removing SSI Loss in the second distillation stage, while retaining the standard supervised loss with sparse ground truth. As shown in Table 3, the results indicate that the proposed SSI Loss improves the model's performance. This demonstrates the benefit of aligning monocular depth supervision with real-world scale using SSI Loss." + }, + { + "type": "table_caption", + "bbox": [ + 0.513, + 0.385, + 0.907, + 0.412 + ], + "angle": 0, + "content": "Table 5. Zero-shot performance comparison with other methods on out-of-the-domain datasets." + }, + { + "type": "table", + "bbox": [ + 0.515, + 0.424, + 0.906, + 0.49 + ], + "angle": 0, + "content": "
MethodScanNetDDADVOID1500
RMSE ↓MAE ↓RMSE ↓MAE ↓RMSE ↓MAE ↓
CFormal0.1200.2329.6063.3280.7260.261
LRRU0.1320.2459.1642.7380.6980.232
BP-Net0.1220.2128.9032.7120.7040.230
Ours0.1010.2107.7662.4980.6760.225
" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.52, + 0.907, + 0.701 + ], + "angle": 0, + "content": "Network Architectures. We evaluate the compatibility of our method with different network architectures, as shown in Table 4. Since we focus on training strategies, we have the flexibility to various models, demonstrating the robustness of our approach across different network designs. Specifically, we evaluate BP-Net [50], LRRU [56], and CFormer (L1 Loss) [79]. BP-Net is one of the most advanced methods, while LRRU and CFormer are considered representative architectures. Our approach consistently improves performance across all three models. Notably, BP-Net with our method achieves the best performance in terms of RMSE, the primary evaluation metric." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.703, + 0.907, + 0.854 + ], + "angle": 0, + "content": "Generalization to Out-of-Domain Datasets. To evaluate the generalization ability of our method, we conduct zero-shot testing on unseen datasets, including ScanNet [13], DDAD [17], and Void1500 [60], as shown in Table 5. Our approach consistently achieves the best performance across all datasets, demonstrating lower RMSE and MAE compared to prior methods. Notably, our method outperforms BP-Net by a significant margin on DDAD, reducing RMSE from 8.903 to 7.766, highlighting its effectiveness in handling unseen scenarios." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.856, + 0.907, + 0.901 + ], + "angle": 0, + "content": "Application on Dense SLAM. Figure 6 compares a representative SLAM method, Droid-SLAM [51], with sparse LiDAR points (left) and our method with dense depth com" + }, + { + "type": "page_number", + "bbox": [ + 0.48, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "22260" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.098, + 0.088, + 0.23, + 0.121 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.114, + 0.145, + 0.228, + 0.205 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.232, + 0.09, + 0.363, + 0.12 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.235, + 0.12, + 0.357, + 0.204 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.366, + 0.089, + 0.497, + 0.12 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.369, + 0.12, + 0.498, + 0.205 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.5, + 0.09, + 0.631, + 0.12 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.502, + 0.12, + 0.63, + 0.206 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.635, + 0.09, + 0.764, + 0.12 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.637, + 0.128, + 0.758, + 0.207 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.768, + 0.09, + 0.901, + 0.12 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.77, + 0.128, + 0.889, + 0.207 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.098, + 0.215, + 0.23, + 0.246 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.102, + 0.248, + 0.228, + 0.352 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.14, + 0.354, + 0.188, + 0.366 + ], + "angle": 0, + "content": "(a) Input" + }, + { + "type": "image", + "bbox": [ + 0.233, + 0.216, + 0.364, + 0.246 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.233, + 0.246, + 0.364, + 0.353 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.252, + 0.354, + 0.345, + 0.366 + ], + "angle": 0, + "content": "(b) CFormer [79]" + }, + { + "type": "image", + "bbox": [ + 0.367, + 0.216, + 0.498, + 0.245 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.369, + 0.246, + 0.498, + 0.353 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.392, + 0.354, + 0.472, + 0.366 + ], + "angle": 0, + "content": "(c) LRRU [56]" + }, + { + "type": "image", + "bbox": [ + 0.501, + 0.216, + 0.631, + 0.245 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.504, + 0.246, + 0.629, + 0.352 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.507, + 0.354, + 0.626, + 0.366 + ], + "angle": 0, + "content": "(d) ImprovingDC [57]" + }, + { + "type": "image", + "bbox": [ + 0.635, + 0.216, + 0.765, + 0.245 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.64, + 0.246, + 0.762, + 0.348 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.658, + 0.354, + 0.743, + 0.366 + ], + "angle": 0, + "content": "(e) BP-Net [50]" + }, + { + "type": "image", + "bbox": [ + 0.768, + 0.216, + 0.901, + 0.245 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.775, + 0.246, + 0.893, + 0.35 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.783, + 0.354, + 0.885, + 0.366 + ], + "angle": 0, + "content": "(f) DMD\\(^3\\)C (Ours)" + }, + { + "type": "image_caption", + "bbox": [ + 0.09, + 0.379, + 0.907, + 0.421 + ], + "angle": 0, + "content": "Figure 5. Qualitative comparison of depth completion methods. This figure demonstrates the performance of various depth completion models, including CFormer, LRRU, ImprovingDC, BP-Net, and our proposed DMD\\(^3\\)C. For each method, we show the input RGB images with sparse LiDAR points (left), along with the resulting completed depth maps and corresponding 3D point cloud reconstructions." + }, + { + "type": "image", + "bbox": [ + 0.119, + 0.445, + 0.213, + 0.547 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.217, + 0.445, + 0.282, + 0.547 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.296, + 0.446, + 0.396, + 0.546 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.398, + 0.446, + 0.462, + 0.546 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.12, + 0.549, + 0.212, + 0.677 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.143, + 0.678, + 0.242, + 0.686 + ], + "angle": 0, + "content": "Droid-SLAM + LiDAR" + }, + { + "type": "image", + "bbox": [ + 0.212, + 0.548, + 0.284, + 0.677 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.296, + 0.548, + 0.388, + 0.677 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.391, + 0.548, + 0.462, + 0.677 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.326, + 0.678, + 0.416, + 0.686 + ], + "angle": 0, + "content": "Droid-SLAM + Ours" + }, + { + "type": "image_caption", + "bbox": [ + 0.091, + 0.712, + 0.483, + 0.74 + ], + "angle": 0, + "content": "Figure 6. Qualitative comparison of Droid-SLAM with sparse LiDAR points and our method with dense depth completion." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.796, + 0.483, + 0.901 + ], + "angle": 0, + "content": "pletion (right). Droid-SLAM suffers from structural distortions and noise, especially in occluded or low-texture areas (red boxes), due to limited depth information. In contrast, our method generates more consistent reconstructions, preserving structural details and improving alignment (green boxes). These results highlight the advantage of dense depth completion in enhancing SLAM quality." + }, + { + "type": "title", + "bbox": [ + 0.514, + 0.446, + 0.633, + 0.461 + ], + "angle": 0, + "content": "5. Conclusion" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.474, + 0.907, + 0.777 + ], + "angle": 0, + "content": "In this work, we present \\(\\mathrm{DMD}^{3}\\mathrm{C}\\), a novel two-stage distillation framework that distills knowledge from monocular depth estimation foundation models into the depth completion task. In the first stage, we introduce a data generation strategy that leverages monocular depth estimation and mesh reconstruction to simulate training data, allowing the model to learn geometric features from diverse natural images. In the second stage, we propose a scale- and shift-invariant loss (SSI Loss) combined with a supervised L1 loss with sparse ground truth, which addresses the scale ambiguity in monocular depth estimation. This ensures consistent depth completions across varying scales and focuses on learning real-world scale information. Extensive experiments on the depth completion benchmarks demonstrate that \\(\\mathrm{DMD}^{3}\\mathrm{C}\\) achieves state-of-the-art performance, ranking first on the KITTI leaderboard, and significantly outperforming existing methods. Our results highlight the framework's ability to produce high-quality depth maps with improved detail and structural consistency, making it a promising solution for depth completion tasks." + }, + { + "type": "title", + "bbox": [ + 0.514, + 0.798, + 0.681, + 0.815 + ], + "angle": 0, + "content": "Acknowledgements" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.826, + 0.906, + 0.9 + ], + "angle": 0, + "content": "This work was supported by the National Key R&D Program of China (2022YFC3300704), the National Natural Science Foundation of China (62331006, 62171038, and 62088101), and the Fundamental Research Funds for the Central Universities." + }, + { + "type": "page_number", + "bbox": [ + 0.48, + 0.946, + 0.518, + 0.957 + ], + "angle": 0, + "content": "22261" + } + ], + [ + { + "type": "title", + "bbox": [ + 0.093, + 0.09, + 0.188, + 0.106 + ], + "angle": 0, + "content": "References" + }, + { + "type": "ref_text", + "bbox": [ + 0.101, + 0.115, + 0.484, + 0.184 + ], + "angle": 0, + "content": "[1] Luca Bartolomei, Matteo Poggi, Andrea Conti, Fabio Tosi, and Stefano Mattoccia. Revisiting depth completion from a stereo matching perspective for cross-domain generalization. In International Conference on 3D Vision (3DV), pages 1360-1370, 2024. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.101, + 0.185, + 0.484, + 0.242 + ], + "angle": 0, + "content": "[2] Shariq Farooq Bhat, Reiner Birkl, Diana Wofk, Peter Wonka, and Matthias Müller. Zoedepth: Zero-shot transfer by combining relative and metric depth. arXiv preprint arXiv:2302.12288, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.242, + 0.483, + 0.325 + ], + "angle": 0, + "content": "[3] Holger Caesar, Varun Bankiti, Alex H Lang, Sourabh Vora, Venice Erin Liong, Qiang Xu, Anush Krishnan, Yu Pan, Giancarlo Baldan, and Oscar Beijbom. nuscenes: A multimodal dataset for autonomous driving. In the IEEE Conference on Computer Vision and Pattern Recognition (CVPR), pages 11621-11631, 2020. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.326, + 0.483, + 0.395 + ], + "angle": 0, + "content": "[4] Angel Chang, Angela Dai, Thomas Funkhouser, Maciej Halber, Matthias Niebner, Manolis Savva, Shuran Song, Andy Zeng, and Yinda Zhang. Matterport3d: Learning from rgb-d data in indoor environments. In International Conference on 3D Vision (3DV), pages 667-676, 2017. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.396, + 0.483, + 0.438 + ], + "angle": 0, + "content": "[5] Linwei Chen, Ying Fu, Kaixuan Wei, Dezhi Zheng, and Felix Heide. Instance segmentation in the dark. International Journal of Computer Vision, 131(8):2198-2218, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.439, + 0.483, + 0.506 + ], + "angle": 0, + "content": "[6] Linwei Chen, Ying Fu, Lin Gu, Chenggang Yan, Tatsuya Harada, and Gao Huang. Frequency-aware feature fusion for dense image prediction. IEEE Transactions on Pattern Analysis and Machine Intelligence, 46(12):10763-10780, 2024. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.509, + 0.483, + 0.564 + ], + "angle": 0, + "content": "[7] Yun Chen, Bin Yang, Ming Liang, and Raquel Urtasun. Learning joint 2d-3d representations for depth completion. In the IEEE Conference on Computer Vision and Pattern Recognition (CVPR), pages 10023-10032, 2019. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.565, + 0.483, + 0.62 + ], + "angle": 0, + "content": "[8] Xinjing Cheng, Peng Wang, and Ruigang Yang. Learning depth with convolutional spatial propagation network. IEEE Transactions on Pattern Analysis and Machine Intelligence (TPAMI), 42(10):2361-2379, 2019. 2, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.621, + 0.483, + 0.69 + ], + "angle": 0, + "content": "[9] Xinjing Cheng, Peng Wang, Chenye Guan, and Ruigang Yang. Cspn++: Learning context and resource aware convolutional spatial propagation networks for depth completion. In the AAAI Conference on Artificial Intelligence (AAAI), pages 10615-10622, 2020. 2, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.691, + 0.483, + 0.759 + ], + "angle": 0, + "content": "[10] Keunhoon Choi, Somi Jeong, Youngjung Kim, and Kwanghoon Sohn. Stereo-augmented depth completion from a single rgb-lidar image. In IEEE International Conference on Robotics and Automation (ICRA), pages 13641-13647, 2021. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.761, + 0.483, + 0.816 + ], + "angle": 0, + "content": "[11] Andrea Conti, Matteo Poggi, and Stefano Mattoccia. Sparsity agnostic depth completion. In the IEEE Winter Conference on Applications of Computer Vision (WACV), pages 5871-5880, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.818, + 0.483, + 0.901 + ], + "angle": 0, + "content": "[12] Marius Cordts, Mohamed Omran, Sebastian Ramos, Timo Rehfeld, Markus Enzweiler, Rodrigo Benenson, Uwe Franke, Stefan Roth, and Bernt Schiele. The cityscapes dataset for semantic urban scene understanding. In the IEEE Conference on Computer Vision and Pattern Recognition (CVPR), pages 3213-3223, 2016. 5" + }, + { + "type": "list", + "bbox": [ + 0.094, + 0.115, + 0.484, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.093, + 0.906, + 0.162 + ], + "angle": 0, + "content": "[13] Angela Dai, Angel X Chang, Manolis Savva, Maciej Halber, Thomas Funkhouser, and Matthias Nießner. Scannet: Richly-annotated 3d reconstructions of indoor scenes. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 5828-5839, 2017. 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.163, + 0.906, + 0.216 + ], + "angle": 0, + "content": "[14] Zheng Fang, Ying Fu, and Lixiong Liu. A dual of transformer features-related map-intelligent generation method. Journal of Image and Graphics, 28(10):3281-3294, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.218, + 0.906, + 0.272 + ], + "angle": 0, + "content": "[15] Ying Fu, Zichun Wang, Tao Zhang, and Jun Zhang. Lowlight raw video denoising with a high-quality realistic motion dataset. IEEE Transactions on Multimedia, 25:8119-8131, 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.274, + 0.906, + 0.327 + ], + "angle": 0, + "content": "[16] Qi Guan, Zihao Sheng, and Shibei Xue. Hrpose: Real-time high-resolution 6d pose estimation network using knowledge distillation. Chinese Journal of Electronics, 32(1):189-198, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.329, + 0.906, + 0.399 + ], + "angle": 0, + "content": "[17] Vitor Guizilini, Rares Ambrus, Sudeep Pillai, Allan Raventos, and Adrien Gaidon. 3d packing for self-supervised monocular depth estimation. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 2485-2494, 2020. 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.4, + 0.906, + 0.44 + ], + "angle": 0, + "content": "[18] Wenxuan Guo, Yingping Liang, Zhiyu Pan, Ziheng Xi, Jianjiang Feng, and Jie Zhou. Camera-lidar cross-modality gait recognition. arXiv preprint arXiv:2407.02038, 2024. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.442, + 0.906, + 0.509 + ], + "angle": 0, + "content": "[19] Wenxuan Guo, Zhiyu Pan, Yingping Liang, Ziheng Xi, Zhicheng Zhong, Jianjiang Feng, and Jie Zhou. Lidar-based person re-identification. In the IEEE Conference on Computer Vision and Pattern Recognition (CVPR), pages 17437–17447, 2024. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.511, + 0.906, + 0.567 + ], + "angle": 0, + "content": "[20] Kaiming He, Xiangyu Zhang, Shaoqing Ren, and Jian Sun. Deep residual learning for image recognition. In the IEEE Conference on Computer Vision and Pattern Recognition (CVPR), pages 770-778, 2016. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.568, + 0.906, + 0.635 + ], + "angle": 0, + "content": "[21] Junjie Hu, Chenyu Bao, Mete Ozay, Chenyou Fan, Qing Gao, Honghai Liu, and Tin Lun Lam. Deep depth completion from extremely sparse data: A survey. IEEE Transactions on Pattern Analysis and Machine Intelligence (TPAMI), 45(7): 8244-8264, 2022. 1, 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.637, + 0.906, + 0.72 + ], + "angle": 0, + "content": "[22] Mu Hu, Wei Yin, Chi Zhang, Zhipeng Cai, Xiaoxiao Long, Hao Chen, Kaixuan Wang, Gang Yu, Chunhua Shen, and Shaojie Shen. Metric3d v2: A versatile monocular geometric foundation model for zero-shot metric depth and surface normal estimation. IEEE Transactions on Pattern Analysis and Machine Intelligence (TPAMI), pages 1-18, 2024. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.721, + 0.906, + 0.787 + ], + "angle": 0, + "content": "[23] Jinwoo Jeon, Hyunjun Lim, Dong-Uk Seo, and Hyun Myung. Struct-mdc: Mesh-refined unsupervised depth completion leveraging structural regularities from visual slam. IEEE Robotics and Automation Letters, 7(3):6391-6398, 2022. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.789, + 0.906, + 0.831 + ], + "angle": 0, + "content": "[24] Michael Kazhdan, Matthew Bolitho, and Hugues Hoppe. Poisson surface reconstruction. In Eurographics Symposium on Geometry Processing, 2006. 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.832, + 0.906, + 0.901 + ], + "angle": 0, + "content": "[25] Bingxin Ke, Anton Obukhov, Shengyu Huang, Nando Metzger, Rodrigo Caye Daudt, and Konrad Schindler. Repurposing diffusion-based image generators for monocular depth estimation. In the IEEE Conference on Computer Vision and Pattern Recognition (CVPR), pages 9492-9502, 2024. 1, 2" + }, + { + "type": "list", + "bbox": [ + 0.517, + 0.093, + 0.906, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.48, + 0.946, + 0.519, + 0.957 + ], + "angle": 0, + "content": "22262" + } + ], + [ + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.092, + 0.486, + 0.161 + ], + "angle": 0, + "content": "[26] Muhammad Ahmed Ullah Khan, Danish Nazir, Alain Pagani, Hamam Mokayed, Marcus Liwicki, Didier Stricker, and Muhammad Zeshan Afzal. A comprehensive survey of depth completion approaches. Sensors, 22(18):6969, 2022. 1, 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.163, + 0.483, + 0.219 + ], + "angle": 0, + "content": "[27] Byeong-Uk Lee, Hae-Gon Jeon, Sunghoon Im, and In So Kweon. Depth completion with deep geometry and context guidance. In International Conference on Robotics and Automation (ICRA), pages 3281-3287. IEEE, 2019. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.221, + 0.483, + 0.276 + ], + "angle": 0, + "content": "[28] Byeong-Uk Lee, Kyunghyun Lee, and In So Kweon. Depth completion using plane-residual representation. In the IEEE Conference on Computer Vision and Pattern Recognition (CVPR), pages 13916-13925, 2021. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.278, + 0.483, + 0.319 + ], + "angle": 0, + "content": "[29] Miaoyu Li and Ying Fu. Joint self-attention transformer for multispectral and hyperspectral image fusion. Journal of Image and Graphics, 28(12):3922–3934, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.321, + 0.483, + 0.376 + ], + "angle": 0, + "content": "[30] Miaoyu Li, Ying Fu, Tao Zhang, and Guanghui Wen. Supervise-assisted self-supervised deep-learning method for hyperspectral image restoration. IEEE Transactions on Neural Networks and Learning Systems, 2024. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.378, + 0.483, + 0.419 + ], + "angle": 0, + "content": "[31] Yingping Liang and Ying Fu. Relation-guided adversarial learning for data-free knowledge transfer. International Journal of Computer Vision, pages 1-18, 2024. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.421, + 0.483, + 0.476 + ], + "angle": 0, + "content": "[32] Yingping Liang, Jiaming Liu, Debing Zhang, and Ying Fu. Mpi-flow: Learning realistic optical flow with multiplane images. In the IEEE International Conference on Computer Vision (ICCV), pages 13857-13868, 2023. 2, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.478, + 0.483, + 0.545 + ], + "angle": 0, + "content": "[33] Tsung-Yi Lin, Michael Maire, Serge Belongie, James Hays, Pietro Perona, Deva Ramanan, Piotr Dólar, and C Lawrence Zitnick. Microsoft coco: Common objects in context. In the European Conference on Computer Vision (ECCV), pages 740-755, 2014. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.548, + 0.483, + 0.603 + ], + "angle": 0, + "content": "[34] Yuankai Lin, Tao Cheng, Qi Zhong, Wending Zhou, and Hua Yang. Dynamic spatial propagation network for depth completion. In the AAAI Conference on Artificial Intelligence (AAAI), pages 1638-1646, 2022. 2, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.605, + 0.483, + 0.674 + ], + "angle": 0, + "content": "[35] Lina Liu, Xibin Song, Xiaoyang Lyu, Junwei Diao, Mengmeng Wang, Yong Liu, and Liangjun Zhang. Fcfr-net: Feature fusion based coarse-to-fine residual learning for depth completion. In the AAAI Conference on Artificial Intelligence (AAAI), pages 2136-2144, 2021. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.676, + 0.483, + 0.744 + ], + "angle": 0, + "content": "[36] Qiankun Liu, Yuqi Jiang, Zhentao Tan, Dongdong Chen, Ying Fu, Qi Chu, Gang Hua, and Nenghai Yu. Transformer based pluralistic image completion with reduced information loss. IEEE Transactions on Pattern Analysis and Machine Intelligence, 46(10):6652-6668, 2024." + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.746, + 0.483, + 0.788 + ], + "angle": 0, + "content": "[37] Qiankun Liu, Yichen Li, Yuqi Jiang, and Ying Fu. Siamesedetr for generic multi-object tracking. IEEE Transactions on Image Processing, 33:3935-3949, 2024." + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.79, + 0.483, + 0.845 + ], + "angle": 0, + "content": "[38] Fangchang Ma and Sertac Karaman. Sparse-to-dense: Depth prediction from sparse depth samples and a single image. In IEEE international conference on robotics and automation (ICRA), pages 4796-4803, 2018. 2, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.847, + 0.483, + 0.9 + ], + "angle": 0, + "content": "[39] Fabiola Maffra, Lucas Teixeira, Zetao Chen, and Margarita Chli. Real-time wide-baseline place recognition using depth completion. IEEE Robotics and Automation Letters, 4(2): 1525-1532, 2019. 1" + }, + { + "type": "list", + "bbox": [ + 0.093, + 0.092, + 0.486, + 0.9 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.092, + 0.906, + 0.147 + ], + "angle": 0, + "content": "[40] Hyoungseob Park, Anjali Gupta, and Alex Wong. Test-time adaptation for depth completion. In the IEEE Conference on Computer Vision and Pattern Recognition (CVPR), pages 20519-20529, 2024. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.15, + 0.906, + 0.205 + ], + "angle": 0, + "content": "[41] Jinsun Park, Kyungdon Joo, Zhe Hu, Chi-Kuei Liu, and In So Kweon. Non-local spatial propagation network for depth completion. In the European Conference on Computer Vision (ECCV), pages 120-136, 2020. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.207, + 0.906, + 0.261 + ], + "angle": 0, + "content": "[42] Jin-Hwi Park and Hae-Gon Jeon. A simple yet universal framework for depth completion. In The Thirty-eighth Annual Conference on Neural Information Processing Systems, 2024. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.264, + 0.906, + 0.345 + ], + "angle": 0, + "content": "[43] Federico Perazzi, Jordi Pont-Tuset, Brian McWilliams, Luc Van Gool, Markus Gross, and Alexander Sorkine-Hornung. A benchmark dataset and evaluation methodology for video object segmentation. In the IEEE Conference on Computer Vision and Pattern Recognition (CVPR), pages 724-732, 2016. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.349, + 0.906, + 0.418 + ], + "angle": 0, + "content": "[44] Luigi Piccinelli, Yung-Hsu Yang, Christos Sakaridis, Mattia Segu, Siyuan Li, Luc Van Gool, and Fisher Yu. Undepth: Universal monocular metric depth estimation. In the IEEE Conference on Computer Vision and Pattern Recognition (CVPR), pages 10106-10116, 2024. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.42, + 0.906, + 0.502 + ], + "angle": 0, + "content": "[45] Jiaxiong Qiu, Zhaopeng Cui, Yinda Zhang, Xingdi Zhang, Shuaicheng Liu, Bing Zeng, and Marc Pollefeys. Deeplidar: Deep surface normal guided depth prediction for outdoor scene from sparse lidar data and single color image. In the IEEE Conference on Computer Vision and Pattern Recognition (CVPR), pages 3313-3322, 2019. 2, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.504, + 0.906, + 0.573 + ], + "angle": 0, + "content": "[46] René Ranftl, Katrin Lasinger, David Hafner, Konrad Schindler, and Vladlen Koltun. Towards robust monocular depth estimation: Mixing datasets for zero-shot cross-dataset transfer. IEEE Transactions on Pattern Analysis and Machine Intelligence (TPAMI), 44(3), 2022. 1, 2, 3, 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.575, + 0.906, + 0.63 + ], + "angle": 0, + "content": "[47] Ali Osman Serhatoglu, Oguzhan Guclu, and Ahmet Burak Can. Rgb-d slam with deep depth completion. In International Conference on Artificial Intelligence and Soft Computing (ICAISC), pages 59-67, 2022. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.632, + 0.906, + 0.687 + ], + "angle": 0, + "content": "[48] Nathan Silberman, Derek Hoiem, Pushmeet Kohli, and Rob Fergus. Indoor segmentation and support inference from rgbd images. In the European Conference on Computer Vision (ECCV), pages 120-136, 2012. 2, 5, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.689, + 0.906, + 0.743 + ], + "angle": 0, + "content": "[49] Jie Tang, Fei-Peng Tian, Wei Feng, Jian Li, and Ping Tan. Learning guided convolutional network for depth completion. IEEE Transactions on Image Processing (TIP), 30: 1116-1129, 2020. 2, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.746, + 0.906, + 0.801 + ], + "angle": 0, + "content": "[50] Jie Tang, Fei-Peng Tian, Boshi An, Jian Li, and Ping Tan. Bilateral propagation network for depth completion. In the IEEE Conference on Computer Vision and Pattern Recognition (CVPR), pages 9763-9772, 2024. 1, 2, 5, 6, 7, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.803, + 0.906, + 0.857 + ], + "angle": 0, + "content": "[51] Zachary Teed and Jia Deng. Droid-slam: Deep visual slam for monocular, stereo, and rgb-d cameras. Advances in Neural Information Processing Systems (NeurIPS), 34:16558-16569, 2021. 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.86, + 0.906, + 0.901 + ], + "angle": 0, + "content": "[52] Ye Tian, Ying Fu, and Jun Zhang. Transformer-based undersampled single-pixel imaging. Chinese Journal of Electronics, 32(5):1151-1159, 2023. 2" + }, + { + "type": "list", + "bbox": [ + 0.517, + 0.092, + 0.906, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.48, + 0.946, + 0.519, + 0.957 + ], + "angle": 0, + "content": "22263" + } + ], + [ + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.092, + 0.482, + 0.146 + ], + "angle": 0, + "content": "[53] Jonas Uhrig, Nick Schneider, Lukas Schneider, Uwe Franke, Thomas Brox, and Andreas Geiger. Sparsity invariant cnns. In International Conference on 3D Vision (3DV), 2017. 1, 2, 3, 5, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.149, + 0.482, + 0.216 + ], + "angle": 0, + "content": "[54] Haotian Wang, Meng Yang, and Nanning Zheng. G2-monodepth: A general framework of generalized depth inference from monocular \\(\\mathrm{rgb + x}\\) data. IEEE Transactions on Pattern Analysis and Machine Intelligence (TPAMI), 2023. 2, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.218, + 0.482, + 0.273 + ], + "angle": 0, + "content": "[55] Haotian Wang, Meng Yang, Xinhu Zheng, and Gang Hua. Scale propagation network for generalizable depth completion. IEEE Transactions on Pattern Analysis and Machine Intelligence (TPAMI), 2024. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.274, + 0.482, + 0.342 + ], + "angle": 0, + "content": "[56] Yufei Wang, Bo Li, Ge Zhang, Qi Liu, Tao Gao, and Yuchao Dai. Lrru: Long-short range recurrent updating networks for depth completion. In the IEEE Conference on Computer Vision and Pattern Recognition (CVPR), pages 9422-9432, 2023. 1, 6, 7, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.344, + 0.482, + 0.412 + ], + "angle": 0, + "content": "[57] Yufei Wang, Ge Zhang, Shaoqian Wang, Bo Li, Qi Liu, Le Hui, and Yuchao Dai. Improving depth completion via depth feature upsampling. In the IEEE Conference on Computer Vision and Pattern Recognition (CVPR), pages 21104-21113, 2024. 1, 2, 5, 6, 7, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.413, + 0.482, + 0.482 + ], + "angle": 0, + "content": "[58] Xiu-Shen Wei, Yi-Zhe Song, Oisin Mac Aodha, Jianxin Wu, Yuxin Peng, Jinhui Tang, Jian Yang, and Serge Belongie. Fine-grained image analysis with deep learning: A survey. IEEE transactions on pattern analysis and machine intelligence, 44(12):8927-8948, 2021. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.484, + 0.482, + 0.552 + ], + "angle": 0, + "content": "[59] Tobias Weyand, Andre Araujo, Bingyi Cao, and Jack Sim. Google landmarks dataset v2-a large-scale benchmark for instance-level recognition and retrieval. In the IEEE Conference on Computer Vision and Pattern Recognition (CVPR), pages 2575-2584, 2020. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.553, + 0.482, + 0.607 + ], + "angle": 0, + "content": "[60] Alex Wong, Xiaohan Fei, Stephanie Tsuei, and Stefano Soatto. Unsupervised depth completion from visual inertial odometry. IEEE Robotics and Automation Letters, 5(2): 1899-1906, 2020. 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.609, + 0.482, + 0.678 + ], + "angle": 0, + "content": "[61] Xiaopei Wu, Liang Peng, Honghui Yang, Liang Xie, Chenxi Huang, Chengqi Deng, Haifeng Liu, and Deng Cai. Sparse fuse dense: Towards high quality 3d detection with depth completion. In the IEEE Conference on Computer Vision and Pattern Recognition (CVPR), pages 5418-5427, 2022. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.679, + 0.482, + 0.747 + ], + "angle": 0, + "content": "[62] Yangchao Wu, Tian Yu Liu, Hyoungseob Park, Stefano Soatto, Dong Lao, and Alex Wong. Augundo: Scaling up augmentations for monocular depth completion and estimation. In European Conference on Computer Vision (ECCV), pages 274-293, 2025. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.748, + 0.482, + 0.829 + ], + "angle": 0, + "content": "[63] Weijian Xie, Guanyi Chu, Quanhao Qian, Yihao Yu, Hai Li, Danpeng Chen, Shangjin Zhai, Nan Wang, Hujun Bao, and Guofeng Zhang. Depth completion with multiple balanced bases and confidence for dense monocular slam. IEEE Transactions on Visualization and Computer Graphics (TVCG), 2024. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.831, + 0.482, + 0.9 + ], + "angle": 0, + "content": "[64] Zexiao Xie, Xiaoxuan Yu, Xiang Gao, Kunqian Li, and Shuhan Shen. Recent advances in conventional and deep learning-based depth completion: A survey. IEEE Transactions on Neural Networks and Learning Systems (TNNLS), 35(3):3395-3415, 2024. 2" + }, + { + "type": "list", + "bbox": [ + 0.093, + 0.092, + 0.482, + 0.9 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.092, + 0.905, + 0.16 + ], + "angle": 0, + "content": "[65] Yan Xu, Xinge Zhu, Jianping Shi, Guofeng Zhang, Hujun Bao, and Hongsheng Li. Depth completion from sparse lidar data with depth-normal constraints. In the IEEE Conference on Computer Vision and Pattern Recognition (CVPR), pages 2811-2820, 2019. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.162, + 0.905, + 0.218 + ], + "angle": 0, + "content": "[66] Zhiqiang Yan, Xiang Li, Kun Wang, Zhenyu Zhang, Jun Li, and Jian Yang. Multi-modal masked pre-training for monocular panoramic depth completion. In European Conference on Computer Vision, pages 378-395. Springer, 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.219, + 0.905, + 0.273 + ], + "angle": 0, + "content": "[67] Zhiqiang Yan, Kun Wang, Xiang Li, Zhenyu Zhang, Jun Li, and Jian Yang. Rignet: Repetitive image guided network for depth completion. In the European Conference on Computer Vision (ECCV), pages 214-230, 2022. 1, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.274, + 0.905, + 0.329 + ], + "angle": 0, + "content": "[68] Zhiqiang Yan, Xiang Li, Kun Wang, Shuo Chen, Jun Li, and Jian Yang. Distortion and uncertainty aware loss for panoramic depth completion. In International Conference on Machine Learning, pages 39099-39109. PMLR, 2023." + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.33, + 0.907, + 0.398 + ], + "angle": 0, + "content": "[69] Zhiqiang Yan, Kun Wang, Xiang Li, Zhenyu Zhang, Jun Li, and Jian Yang. Desnet: Decomposed scale-consistent network for unsupervised depth completion. In Proceedings of the AAAI conference on artificial intelligence, pages 3109-3117, 2023." + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.399, + 0.905, + 0.481 + ], + "angle": 0, + "content": "[70] Zhiqiang Yan, Yuankai Lin, Kun Wang, Yupeng Zheng, Yufei Wang, Zhenyu Zhang, Jun Li, and Jian Yang. Tri-perspective view decomposition for geometry-aware depth completion. In the IEEE Conference on Computer Vision and Pattern Recognition (CVPR), pages 4874-4884, 2024. 1, 2, 6, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.483, + 0.905, + 0.552 + ], + "angle": 0, + "content": "[71] Lihe Yang, Bingyi Kang, Zilong Huang, Xiaogang Xu, Jiashi Feng, and Hengshuang Zhao. Depth anything: Unleashing the power of large-scale unlabeled data. In the IEEE Conference on Computer Vision and Pattern Recognition (CVPR), pages 10371-10381, 2024. 1, 2, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.553, + 0.905, + 0.607 + ], + "angle": 0, + "content": "[72] Lihe Yang, Bingyi Kang, Zilong Huang, Zhen Zhao, Xiaogang Xu, Jiashi Feng, and Hengshuang Zhao. Depth anything v2. In Annual Conference on Neural Information Processing Systems (NeurIPS), 2024. 1, 2, 3, 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.608, + 0.905, + 0.678 + ], + "angle": 0, + "content": "[73] Wei Yin, Chi Zhang, Hao Chen, Zhipeng Cai, Gang Yu, Kaixuan Wang, Xiaozhi Chen, and Chunhua Shen. Metric3d: Towards zero-shot metric 3d prediction from a single image. In the IEEE Conference on Computer Vision and Pattern Recognition (CVPR), pages 9043-9053, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.679, + 0.905, + 0.747 + ], + "angle": 0, + "content": "[74] Zhu Yu, Zehua Sheng, Zili Zhou, Lun Luo, Si-Yuan Cao, Hong Gu, Huaqi Zhang, and Hui-Liang Shen. Aggregating feature point cloud for depth completion. In the IEEE International Conference on Computer Vision (ICCV), pages 8732-8743, 2023. 2, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.748, + 0.905, + 0.803 + ], + "angle": 0, + "content": "[75] Fan Zhang, Shaodi You, Yu Li, and Ying Fu. Atlantis: Enabling underwater depth estimation with stable diffusion. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 11852-11861, 2024. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.804, + 0.905, + 0.845 + ], + "angle": 0, + "content": "[76] Tao Zhang, Ying Fu, and Jun Zhang. Deep guided attention network for joint denoising and demosaicing in real image. Chinese Journal of Electronics, 33(1):303-312, 2024. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.846, + 0.905, + 0.899 + ], + "angle": 0, + "content": "[77] Xuanmeng Zhang, Zhedong Zheng, Minyue Jiang, and Xiaojing Ye. Self-ensembling depth completion via density-aware consistency. Pattern Recognition, 154:110618, 2024. 2" + }, + { + "type": "list", + "bbox": [ + 0.517, + 0.092, + 0.907, + 0.899 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.48, + 0.946, + 0.52, + 0.957 + ], + "angle": 0, + "content": "22264" + } + ], + [ + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.093, + 0.482, + 0.146 + ], + "angle": 0, + "content": "[78] Yinda Zhang and Thomas Funkhouser. Deep depth completion of a single rgb-d image. In the IEEE Conference on Computer Vision and Pattern Recognition (CVPR), pages 175-185, 2018. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.149, + 0.482, + 0.231 + ], + "angle": 0, + "content": "[79] Youmin Zhang, Xianda Guo, Matteo Poggi, Zheng Zhu, Guan Huang, and Stefano Mattoccia. Completionformer: Depth completion with convolutions and vision transformers. In the IEEE Conference on Computer Vision and Pattern Recognition (CVPR), pages 18527-18536, 2023. 1, 2, 5, 6, 7, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.234, + 0.482, + 0.288 + ], + "angle": 0, + "content": "[80] Shanshan Zhao, Mingming Gong, Huan Fu, and Dacheng Tao. Adaptive context-aware multi-modal network for depth completion. IEEE Transactions on Image Processing (TIP), 30:5264-5276, 2021. 2, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.291, + 0.482, + 0.343 + ], + "angle": 0, + "content": "[81] Jin Zheng, Botao Jiang, Wei Peng, and Qiaohui Zhang. Multi-scale binocular stereo matching based on semantic association. Chinese Journal of Electronics, 33(4):1010-1022, 2024. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.347, + 0.482, + 0.415 + ], + "angle": 0, + "content": "[82] Wending Zhou, Xu Yan, Yinghong Liao, Yuankai Lin, Jin Huang, Gangming Zhao, Shuguang Cui, and Zhen Li. Bev@ dc: Bird's-eye view assisted training for depth completion. In the IEEE Conference on Computer Vision and Pattern Recognition (CVPR), pages 9233-9242, 2023. 2, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.418, + 0.482, + 0.471 + ], + "angle": 0, + "content": "[83] Yunhao Zou, Ying Fu, Tsuyoshi Takatani, and Yinqiang Zheng. Eventhrd: From event to high-speed hdr videos and beyond. IEEE Transactions on Pattern Analysis and Machine Intelligence, 47(1):32-50, 2024. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.474, + 0.482, + 0.527 + ], + "angle": 0, + "content": "[84] Yiming Zuo and Jia Deng. Ogni-dc: Robust depth completion with optimization-guided neural iterations. In the European Conference on Computer Vision (ECCV), pages 78-95, 2025. 2" + }, + { + "type": "list", + "bbox": [ + 0.093, + 0.093, + 0.482, + 0.527 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.48, + 0.946, + 0.518, + 0.956 + ], + "angle": 0, + "content": "22265" + } + ] +] \ No newline at end of file diff --git a/2025/Distilling Monocular Foundation Model for Fine-grained Depth Completion/1663a87e-e3f7-4145-9d6a-27f93678241e_origin.pdf b/2025/Distilling Monocular Foundation Model for Fine-grained Depth Completion/1663a87e-e3f7-4145-9d6a-27f93678241e_origin.pdf new file mode 100644 index 0000000000000000000000000000000000000000..ec2a6a26bfa5340564049950a899beb05b3d613f --- /dev/null +++ b/2025/Distilling Monocular Foundation Model for Fine-grained Depth Completion/1663a87e-e3f7-4145-9d6a-27f93678241e_origin.pdf @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:62987c201a80e343ba4c748b7a1f483c7e8e530290d74ab4915e3e9f839a0e5d +size 8443178 diff --git a/2025/Distilling Monocular Foundation Model for Fine-grained Depth Completion/full.md b/2025/Distilling Monocular Foundation Model for Fine-grained Depth Completion/full.md new file mode 100644 index 0000000000000000000000000000000000000000..3f8389454e521ebc048318cf3163856fa73b8e0f --- /dev/null +++ b/2025/Distilling Monocular Foundation Model for Fine-grained Depth Completion/full.md @@ -0,0 +1,418 @@ +# Distilling Monocular Foundation Model for Fine-grained Depth Completion + +Yingping Liang1 Yutao Hu2 Wenqi Shao3 Ying Fu1† +1Beijing Institute of Technology 2Key Laboratory of New Generation Artificial Intelligence Technology and Its Interdisciplinary Applications, Southeast University 3Shanghai Al Laboratory {liangyingping, fuying}@bit.edu.cn huyutao@seu.edu.cn weqish@link.cuhk.edu.hk + +![](images/9838f5956b4ecd56b7df0dcbb9f8556c79e9ba628572bfb75045a8d65b7240cf.jpg) +Figure 1. Depth completion models trained solely with $L_{1}$ loss and sparse ground truth produce incomplete and fragmented depth predictions. Our framework, however, demonstrates significant improvements by distilling knowledge from monocular foundation models and incorporating a scale- and shift-invariant loss (SSI Loss), resulting in more complete and accurate dense depth completion. + +# Abstract + +Depth completion involves predicting dense depth maps from sparse LiDAR inputs. However, sparse depth annotations from sensors limit the availability of dense supervision, which is necessary for learning detailed geometric features. In this paper, we propose a two-stage knowledge distillation framework that leverages powerful monocular foundation models to provide dense supervision for depth completion. In the first stage, we introduce a pre-training strategy that generates diverse training data from natural images, which distills geometric knowledge to depth completion. Specifically, we simulate LiDAR scans by utilizing monocular depth and mesh reconstruction, thereby creating training data without requiring ground-truth depth. Besides, monocular depth estimation suffers from inherent scale ambiguity in real-world settings. To address this, in the second stage, we employ a scale- and shift-invariant loss (SSI Loss) to learn real-world scales when fine-tuning on real-world datasets. Our two-stage distillation framework enables depth completion models to harness the strengths of monocular foundation models. Experimental results demonstrate that models trained with our two-stage distillation framework achieve state-of-the-art performance, ranking first place on the KITTI benchmark. Code is available + +at https://github.com/Sharpiless/DMD3C + +# 1. Introduction + +Depth completion is a fundamental task in computer vision, where the goal is to generate dense depth maps from sparse depth measurements. This task is particularly important in applications such as autonomous driving [23, 47, 61], robotics [26, 39, 63], and augmented reality [21, 30]. + +Recent approaches [7, 27, 28, 65, 67, 78] leverage deep neural networks to learn from depth data. Some methods [56, 79] also incorporate RGB images to guide the depth completion process. Despite these advancements, state-of-the-art models [50, 57, 70] still face difficulties in capturing fine-grained geometric details, particularly in complex outdoor scenes where depth annotations are sparse [53, 58, 75]. + +The challenge arises from the reliance on sparse ground truth for training. The lack of dense ground truth makes it difficult for models to accurately learn depth completion across an entire scene. Monocular depth estimation [25, 46, 71, 72], on the other hand, has the ability to provide dense depth predictions from a single image. Specifically, state-of-the-art monocular foundation models could generate dense depth maps, containing fine-grained details and relative depth relationships, offering valuable guidance for training depth completion networks. + +To make full use of the advantages of monocular foun + +dation models, we propose a novel two-stage distillation framework to transfer geometric knowledge from monocular foundation models to depth completion networks. Specifically, in the first distillation stage, we generate training data through monocular foundation models and then distill knowledge via the proposed pre-training strategy. Specifically, diverse natural images are used to generate pseudo depth maps. Then, we utilize randomly sampled camera parameters to re-construct the scene with mesh and simulate LiDAR using ray simulation. The generated data trains the depth completion model to learn diverse geometric knowledge from monocular foundation models, enhancing its ability to generalize across different scenes. + +However, monocular depth estimation suffers from inherent scale ambiguity [16, 81], resulting in depth predictions that vary greatly in scale. Thus, monocular depth alone cannot serve as a reliable basis for real-world depth. To solve this problem, in the second distillation stage, we introduce a scale- and shift-invariant loss (SSI Loss) [46]. Specifically, when fine-tuning on labeled datasets with sparse ground truth, SSI Loss ignores the scale and shift that causes the least loss from the depth prediction and monocular depth, to ensure consistent depth completions aligning monocular depth supervision across varying scales. + +By integrating these two distillation stages into the training process, our method achieves state-of-the-art performance on the KITTI benchmark. The contributions of this work are summarized as follows: + +- We propose a novel two-stage distillation framework to transfer knowledge from monocular foundation models to depth completion models, which enables the learning of fine-grained depth information from sparse ground truth by providing dense supervision. +- In the first stage, we propose a data generation strategy that uses monocular depth estimation and mesh reconstruction to simulate training data, enabling the model to learn geometric features from diverse natural images without the need for any LiDAR and ground truth. +- In the second stage, we propose a scale- and shift-invariant loss (SSI Loss) to address the scale ambiguity problem in monocular depth estimation, which ensures consistent depth completions across varying scales and focuses on learning real-world scale information. +- Our method achieves first place on the official KITTI depth completion benchmark, demonstrating the effectiveness of the proposed approach with significant gains in both quantitative metrics and qualitative visualizations. + +# 2. Related Work + +# 2.1. Depth Completion Datasets + +Depth completion has been widely studied in both indoor and outdoor settings. Indoor datasets, such as NYU Depth + +V2 [48] and Matterport3D [4], are popular because dense depth measurements are easier to obtain in controlled environments. NYU Depth V2, for instance, provides dense depth maps from a Kinect sensor along with RGB images. + +In outdoor environments, however, obtaining dense depth annotations is more challenging due to limitations in sensing technologies like LiDAR. The KITTI dataset [53] is a key benchmark for outdoor depth completion but suffers from sparse annotations, covering only about $5\%$ of the image. To address this, training methods often employ complex post-processing and multi-frame fusion techniques to increase the annotation coverage to around $20\%$ at most. Furthermore, ground truth values for long distances and dynamic objects typically cannot be included. The sparse ground truth raises challenges for training fine-grained depth completion models. Therefore, we employ monocular depth estimation techniques for dense distillation to preserve fine-grained details. + +# 2.2. Depth Completion Methods + +Depth completion methods have seen significant advancements [21, 26, 64]. Compared to predicting depth directly from a single RGB image [2, 22, 44, 73], depth completion fuses LiDAR information to obtain more accurate sparse depth cues. Recent methods [1, 6, 8, 9, 11, 14, 29, 34-38, 40-42, 45, 49, 50, 52, 55-57, 62, 66-70, 74, 76, 79, 80, 82, 84] rely on supervised learning, using pairs of sparse depth maps and corresponding dense ground truth. While effective with sufficient labeled data, these methods are limited in real-world applications, where obtaining dense annotations, especially in outdoor environments, is challenging [5, 15, 83]. To reduce the reliance on dense labels, [10, 77] use a teacher-student model, where a teacher network trained on labeled data provides guidance to a student network. However, pseudo-labels in complex or dynamic scenes can be noisy or inaccurate. Therefore, we turn to monocular depth estimation to provide fine-grained supervision for depth completion. + +# 2.3. Monocular Depth Estimation + +Monocular depth estimation [25, 46, 71, 72] has emerged as a powerful tool for generating dense depth maps from single RGB images, making it particularly valuable in scenarios where dense depth annotations are sparse or unavailable [18, 19, 32]. Notable models, such as MiDaS [46] and Depth-Anything [71, 72], have demonstrated impressive capabilities in predicting dense depth maps across a wide range of scenes. G2-MonoDepth [54] have proposed a general framework for monocular RGB-X Data with a scale invariant loss. Despite their versatility, monocular depth estimation methods face the challenge of scale ambiguity, which can limit the accuracy of depth completion, especially when integrated with sparse depth measurements + +![](images/e2e9da7addaca5781d47007324bb52844176ced247576ec4ff2289216dea86a7.jpg) +Figure 2. Illustration of our proposed first distillation stage with a data generation strategy to learn geometric features from monocular foundation models, which only requires unlabeled RGB images. We use the estimated monocular depth to re-construct the scene and then simulate the Lidar swap process to generate sparse points for training. + +from LiDAR. Therefore, we introduce a scale- and shift-invariant loss inspired by [54], combined with a supervised loss to keep real-world scales when fine-tuning on real-world datasets. + +# 3. Method + +In this section, we first provide a brief overview of the motivation and formulation of our method. Then, we introduce our two-stage distillation framework. In the first stage, we propose a data generation strategy for pre-training that leverages large-scale, unlabeled images without ground-truth depth annotations, as illustrated in Figure 2. In the second stage, we detail the fine-tuning process on a labeled dataset with sparse ground truth, using monocular depth in combination with SSI Loss, as shown in Figure 3. + +# 3.1. Motivation and Formulation + +In the depth completion task, the input typically consists of an RGB image $I$ and a sparse depth map $D_{s}$ obtained from a LiDAR sensor. The objective is to produce a dense depth map $D_{f}$ that provides a complete and detailed representation of the scene's depth. To achieve this, a depth completion model $f_{\theta}$ , parameterized by $\theta$ , takes both the RGB image and sparse depth map as inputs: + +$$ +D _ {f} = f _ {\theta} (I, D _ {s}). \tag {1} +$$ + +To guide the model's learning, a sparse supervision loss is defined based on the available sparse depth annotations: + +$$ +\mathcal {L} _ {\sup } = M \times \left| D _ {f} - D _ {\text {s p a r s e}} \right|, \tag {2} +$$ + +where $M$ represents the valid mask for sparse depth ground truth $D_{\mathrm{sparse}}$ . This loss aligns the predicted dense depth map $D_{f}$ with the sparse ground truth, ensuring that the output + +is consistent with the available depth data. However, this sparse supervision provides limited guidance, particularly in outdoor scenes [53], leading to challenges in achieving fine-grained and consistent depth predictions due to the inherent limitations of sparse supervision. + +Foundation models for monocular depth estimation [71, 72] provide an alternative approach by generating fine-grained dense depth from single RGB images, which we use as supervision without relying solely on LiDAR. Therefore, in the first stage, we utilize the monocular foundation models to generate diverse and large-scale training data, to provide dense supervision for pre-training, allowing the model to learn geometric features. However, monocular methods suffer from scale ambiguities, which can lead to inconsistencies when combining dense monocular predictions with sparse depth. To address this, in the second stage, we introduce SSI Loss [46] combined with L1 loss when finetuning on labeled datasets with sparse ground truth to mitigate scale inconsistencies and learn real-world scale. + +# 3.2. First Stage: Data Generation and Pre-training + +To leverage diverse images without depth annotations, we introduce a pre-training strategy that utilizes synthesized depth data generated through monocular depth estimation on large-scale natural image datasets inspired by [31, 32]. As illustrated in Figure 2, this pre-training phase enables the model to learn robust geometric features from monocular foundation models across diverse scenes, preparing it for subsequent fine-tuning on real-world datasets. + +Specifically, we first utilize a pre-trained monocular depth estimation model, such as Depth Anything V2 [72], to predict depth maps for such images. For each natural image $I_{\mathrm{un}}$ , the monocular depth estimation model predicts a + +![](images/5dfe5fd98d70aa5e864902e159ba1bde7d23d7a0bb74de6ad51205f5afdcf78e.jpg) +Figure 3. Illustration of our proposed second distillation stage utilizing foundation models for monocular depth estimation when fine-tuning on labeled datasets. Sparse ground truth provides real-world depth scale with L1 loss. Our method enhances this process by incorporating dense monocular depth for fine-grained supervision. However, monocular depth maps come with inherent scale and shift ambiguities. To address these challenges, we employ a Scale- and Shift-Invariant Loss (SSI Loss) that aligns the predictions with the dense monocular depth to match the real-world depth scale, ensuring more accurate depth completion. + +corresponding dense depth map $D_{\mathrm{syn}}$ .. + +$$ +D _ {\text {s y n}} = f _ {\mathrm {m o}} \left(I _ {\mathrm {u n}}\right), \tag {3} +$$ + +where $f_{\mathrm{mo}}$ represents the monocular depth estimation model. These synthesized depth maps, though not accurate in metric scale, capture the relative depth relationships and structural details in the scene, providing valuable supervision signals during pre-training. + +To simulate LiDAR scanning, we first sample a random camera intrinsic matrix $K$ as: + +$$ +K = \left[ \begin{array}{c c c} f _ {x} & 0 & c _ {x} \\ 0 & f _ {y} & c _ {y} \\ 0 & 0 & 1 \end{array} \right], \tag {4} +$$ + +where $f_{x}$ and $f_{y}$ represent the focal lengths along the $x$ - and $y$ -axes, and $(c_{x}, c_{y})$ denotes the center of the image. + +For each pixel $(u,v)$ in the depth map, the corresponding 3D coordinates $(X,Y,Z)$ in the camera coordinate system are derived using the sampled camera intrinsic matrix $K$ . This transformation generates a point cloud $P = \{(X,Y,Z)\}$ that represents the 3D spatial locations of each pixel. Leveraging the 3D point cloud $P$ , we reconstruct a mesh $M$ via surface reconstruction techniques, such as Poisson Surface Reconstruction [24], to create a continuous 3D surface model for simulation. + +Next, we simulate a LiDAR sensor by generating a set of ray direction vectors $\mathbf{d}$ . Each ray is cast from the origin along $\mathbf{d}_{i,j}$ , intersecting the mesh $M$ to determine the distance to the intersection point. This distance is recorded as the simulated LiDAR depth reading $D_{\mathrm{simu}}(i,j)$ . The resulting sparse depth map $D_{\mathrm{simu}}$ emulates LiDAR depth readings from the reconstructed 3D surface, to provide a synthetic depth image comparable to actual LiDAR data. + +Then, the depth completion model $f_{\theta}$ takes the RGB im + +age and the sparse depth from simulated scanning as inputs: + +$$ +D _ {f} = f _ {\theta} \left(I _ {\mathrm {u n}}, D _ {\mathrm {s i m u}}\right), \tag {5} +$$ + +where $D_{\mathrm{simu}}$ is the sparse input from simulated LiDAR scanning. Afterwards, the depth completion model is pretrained using L1 Loss: + +$$ +\mathcal {L} _ {\text {p r e}} = \left| D _ {f} - D _ {m} \right|. \tag {6} +$$ + +This pre-training process encourages the depth completion model $f_{\theta}$ to align its predictions with the synthesized depth maps, learning meaningful geometric features from the generated training data using monocular foundation models and diverse natural image data. Additionally, the model learns complex geometric structures and relative depth relationships across varying scenes, which are crucial for accurate depth completion in the fine-tuning phase. + +# 3.3. Second Stage: Fine-tuning + +The depth completion model pre-trained on the generated data learns powerful geometric features from the monocular foundation models. However, due to the inherent scale ambiguity of monocular depth, the model prediction is not consistent with the scale in the real world. To learn the real-world scale, we introduce a combined loss for fine-tuning on labeled datasets. To be specific, we first utilize the sparse ground truth $D_{s}$ for supervised loss $\mathcal{L}_{\mathrm{sup}}$ to focus on real-world depth, as in Equation 2. Then, the dense monocular depth $D_{m}$ predicted by monocular depth models is also used to provide approximate depth values across the image, especially in regions not covered by the ground truth $D_{s}$ . + +However, the key challenge in using monocular depth estimation as supervision is the inherent scale ambiguity in depth predictions. To address this issue, we incorporate a Scale- and Shift-Invariant Loss (SSI Loss) into our distillation framework. The SSI Loss is designed to align the + +Table 1. Unlabeled datasets used for pre-training data generation. + +
DatasetIndoorOutdoor# Images
COCO [33]✓✓118,287
Google Landmarks [59]✓117,576
Nuscenes [3]✓93,475
Cityscapes [12]✓19,998
DAVIS [43]✓✓10,581
+ +predicted depth map $D_{f}$ with the dense monocular depth $D_{m}$ . It remains invariant to differences in scale and shift between $D_{f}$ and $D_{m}$ . The SSI Loss is formulated as: + +$$ +\mathcal {L} _ {\text {S S I L}} = \min _ {s, b} | D _ {f} - (s \cdot D _ {m} + b) |, \tag {7} +$$ + +where $s$ and $b$ are the optimal scale and shift parameters that align the predicted depth $D_{f}$ with the dense monocular depth $D_{m}$ . The loss function seeks to find the best alignment between $D_{f}$ and $D_{m}$ , effectively normalizing any global differences in scale and offset. By minimizing this loss across all pixels in the image, the model is encouraged to produce depth maps that maintain consistency with the relative depth structure provided by the monocular depth estimates. In addition, we adapt a gradient matching term to preserve sharpness and align with with depth discontinuities. The gradient matching term is defined as: + +$$ +\mathcal {L} _ {\text {r e g}} = \frac {1}{N} \sum_ {k = 1} ^ {K} \left(| \nabla_ {x} R ^ {k} | + | \nabla_ {y} R ^ {k} |\right), \tag {8} +$$ + +where $R = D_{f} - D_{m}$ , and $R^{k}$ denotes the difference of depth maps at scale $k$ . We use $K = 4$ scale levels as in [46], halving the image resolution at each level. The final objective function combines the supervised loss $\mathcal{L}_{\mathrm{sup}}$ , the dense distillation loss $\mathcal{L}_{\mathrm{SSIL}}$ , and the regularization term $\mathcal{L}_{\mathrm{reg}}$ . + +# 4. Experiments + +In this section, we first introduce the details of our implementation, as well as the datasets and evaluation metrics for experiments. Then, detailed comparisons are conducted with the state-of-the-art methods. Finally, ablations and discussions are performed to confirm the effectiveness of our proposed main components. Additional analysis is provided in the supplementary materials, along with videos. + +# 4.1. Experimental Pipeline + +First stage involves utilizing our proposed data generation strategy to train the depth completion model from scratch, as described in Section 3.2. In this stage, we generate training data using the RGB images in Table 1 and train the depth completion model using SSI Loss, which does not require + +any ground truth depth. This stage allows the pre-trained model to focus on learning diverse geometric features without being constrained by labeled data, enhancing generalization across different scenes. + +Second stage involves adapting the pre-trained model on labeled datasets, using supervised loss (L1 loss) with depth annotations and our proposed SSI Loss for knowledge distillation, as described in Section 3.3. Using L1 loss with sparse ground truth allows the depth completion model to adapt to the real-world depth scale under sparse supervision. Specifically, our proposed monocular model distillation uses SSI Loss to help the model maintain fine-grained detail with dense supervision by distilling monocular foundation models. Fine-tuning and validation are performed on the KITTI and NYU Depth V2 datasets, respectively. + +# 4.2. Implementation Details + +We follow the setup of recent work [50, 57, 79] and train our model on 4 NVIDIA RTX A100 GPUs. For the monocular foundation model used in distillation, we adopt Depth Anything V2 [72] due to its robust performance. For the depth completion network architecture, we use BP-Net [50] as our base model, which is primarily composed of ResNet [20] blocks. In the first stage, we utilize a mixed dataset collected from a wide range of resources, as shown in Table 1. We select five large-scale public datasets as our unlabeled sources for their diverse scenes, totaling approximately 360,000 images with varying scenes and image scales. We adopt AdamW with a weight decay of 0.05 as the optimizer and clip gradients if their norms exceeding 0.1. In the first stage, our model is pre-trained from scratch for 600K iterations. In the second stage, the pre-trained model is fine-tuned on the labeled datasets for 300K iterations using a OneCycle learning rate policy. We set the batch size to 16. The final model is obtained by applying Exponential Moving Average (EMA) with a decay of 0.9999, which helps stabilize the model parameters during training. + +# 4.3. Datasets and Evaluation Metrics + +KITTI Depth Completion Dataset [53] serves as a standard benchmark for depth completion in autonomous driving scenarios. It provides over 93,000 training samples, including sparse LiDAR depth maps, their corresponding RGB images, and ground truth depth maps. The sparse LiDAR depth maps are generated from Velodyne LiDAR scans, which typically cover only a small percentage of the image pixels (approximately $5\%$ ). The ground truth depth maps are obtained via multi-frame matching and cover approximately $20\%$ image pixels. The dataset also includes 1,000 test samples on the private online benchmark. + +NYU Depth V2 Dataset [48] is also a standard benchmark for indoor scene understanding, consisting of RGB-D data captured with a Kinect sensor. It provides 1,449 densely + +Table 2. Performance on KITTI and NYUv2 datasets. For the KITTI dataset, results are evaluated by the KITTI testing server and ranked by the RMSE (in mm). For the NYUv2 dataset, we report their performance on the official test set in their papers. + +
MethodYearKITTINYUv2
RMSE ↓MAE ↓iRMSE ↓iMAE ↓RMSE ↓REL ↓δ1.25 ↑
S2D [38]2018814.73249.952.801.210.2300.04497.1
CSPN [8]20191019.64279.462.931.150.1170.01699.2
DeepLiDAR [45]2019758.38226.502.561.150.1150.02299.3
CSPN++ [9]2020743.69209.282.070.900.1010.01599.5
GuideNet [49]2020736.24218.832.250.990.115--
FCFR [35]2021735.81217.152.200.980.1060.01599.5
ACMNet [80]2021744.91206.092.080.900.1050.01599.4
NLSPN [41]2020741.68199.591.990.840.0920.01299.6
PointDC [74]2023736.07201.871.970.870.0890.01299.6
RigNet [67]2022712.66203.252.080.900.0900.01399.6
DySPN [34]2022709.12192.711.880.820.0900.01299.6
BEV@DC [82]2023697.44189.441.830.820.0890.01299.6
CFormaler [79]2023708.87203.452.010.880.0900.012-
LRRU [56]2023696.51189.961.870.810.0910.01199.6
TPVD [70]2024693.97188.601.820.810.0860.01099.7
ImprovingDC [57]2024686.46187.951.830.810.0910.01199.6
BP-Net [50]2024684.90194.691.820.840.0890.01299.6
DMD3C (Ours)-678.12194.461.820.850.0850.01199.7
+ +labeled images from various indoor scenes. Missing ground truth is mainly due to the difference in perspective between the depth sensor and the camera. + +Evaluation Metrics used for KITTI are root mean squared error (RMSE), mean absolute error (MAE), root mean squared error of the inverse depth (iRMSE), and mean absolute error of the inverse depth (iMAE), among which RMSE is used for ranking. These metrics are obtained by submission and evaluation from a non-public test set. For NYU, we follow the previous work and report the RMSE, mean absolute relative error (REL), and $\delta_{\theta}$ , which represents the percentage of pixels whose error is less than a threshold $\theta$ . + +# 4.4. Comparison with State-of-the-art Methods + +Evaluation on Depth Completion Benchmarks. We evaluate our method on the test sets of the NYUv2 dataset [48] and KITTI dataset [53]. Table 2 shows the quantitative comparison of our method and other top ranking published methods. On the KITTI leaderboard, our proposed $\mathrm{DMD}^{3}\mathrm{C}$ ranks 1st outperforming all other methods under the primary RMSE metric at the time of paper submission. It also has comparable performance under other evaluation metrics. On the NYUv2 dataset, our method achieves the best RMSE and best $\delta_{1.25}$ . Additionally, $\mathrm{DMD}^{3}\mathrm{C}$ maintains competitive results across several other evaluation metrics, further confirming its robustness and generalization. + +To enhance understanding of $\mathrm{DMD}^3\mathrm{C}$ 's superiority over other state-of-the-art methods, we provide visual comparisons. Figure 4 provides qualitative comparisons with other + +SOTA methods on the KITTI test set using publicly available results. We use results from the official benchmark1 produced by the best models for method comparison. Our $\mathrm{DMD}^{3}\mathrm{C}$ model excels in maintaining sharp object boundaries and capturing fine details in areas where other models encounter difficulties, particularly in scenes with complex structures or objects at varying distances. In such challenging regions, other methods fail to estimate accurate depth. Furthermore, as shown in the error maps, our method demonstrates a clear advantage even in regions without ground truth, despite these areas not being factored into the evaluation metrics. + +Not only does our model retain the fine details in the depth map, but it also maintains structural integrity in the potential three-dimensional space. Figure 5 illustrates that our proposed two-stage distillation improves performance in the 3D space by visualizing point clouds from the completed depth maps. In comparison to other methods, our $\mathrm{DMD}^{3}\mathrm{C}$ produces a more coherent and complete 3D structure, further outperforming existing methods. + +# 4.5. Discussions + +Effect of Pre-training in the First Stage. First, to evaluate the impact of the proposed pre-training strategy using unlabeled images, we conduct an experiment where the depth completion model is trained from scratch, without any pretraining. As shown in Table 3, removing the pre-training + +![](images/989c23260cf322f2f5489dfa5ffcf88cc862ebf80ae465d63d4d0dfeaa32e790.jpg) + +![](images/75cf018cd8ba5559f4dfe243195427728fad2c856418120c712aa27c8fea0398.jpg) +(b) CFormer [79] + +![](images/efbeaad44105c48f960526d706dbc3b44f0bd7841b1985844e2fb72b331bcb95.jpg) +(c) TPVD [70] + +![](images/c07c2e43a64ab386a83ca11f5df59d8f939f1f2b2ec80598982dcb79ce2d0980.jpg) +(d) BP-Net [50] + +![](images/78ce7be44fc77e8198efcc999a7d9b12f1e9af2212ee6eee16181aa9b67ce1d9.jpg) +(a) Frame +(e) Error Map +Figure 4. Qualitative comparison of our proposed DMD $^3$ C with several state-of-the-art methods on the KITTI benchmark, using public test results. Error maps highlight pixels with ground truth. In regions lacking ground truth, our method demonstrates notable improvements in depth completion, even though these areas are excluded from the evaluation metrics. + +![](images/5f3a1cb71122f861118bfaa3d7bf52c30ff81884e81c8d7336e3c3a2cea3e3eb.jpg) +(f) LRRU [56] + +![](images/eab9412e6b7eaec2a5c9640808bdb174f8eeec961ab7e905088bdec6e1eb7d4f.jpg) +(g) ImprovingDC [57] + +![](images/db0e6677e37668621b3120154169f978214e2eb209833c6e25dda6fd8d9e4aed.jpg) +(h) DMD³C (Ours) + +Table 3. Ablations on the proposed main components. + +
MethodKITTI
RMSE ↓MAE ↓iRMSE ↓iMAE ↓
w/o Pre-train682.34194.961.820.84
w/o SSI Loss684.54195.651.860.85
DMD3C678.12194.461.820.85
+ +Table 4. Ablations on different network architectures. + +
MethodKITTI
RMSE ↓MAE ↓iRMSE ↓iMAE ↓
LRRU [56]696.51189.961.870.81
+ Ours693.17189.601.850.80
CFormer [79]764.87183.881.890.80
+ Ours760.29183.621.880.80
BP-Net [50]684.90194.691.820.84
+ Ours678.12194.461.820.85
+ +step results in a performance degradation, with an increase in RMSE by $4.22\mathrm{mm}$ . This demonstrates that pre-training plays a crucial role in enhancing the model's ability to generalize, enabling it to better capture complex geometric features across diverse scenes. + +Effect of SSI Loss in the Second Stage. Next, we show the impact of removing SSI Loss in the second distillation stage, while retaining the standard supervised loss with sparse ground truth. As shown in Table 3, the results indicate that the proposed SSI Loss improves the model's performance. This demonstrates the benefit of aligning monocular depth supervision with real-world scale using SSI Loss. + +Table 5. Zero-shot performance comparison with other methods on out-of-the-domain datasets. + +
MethodScanNetDDADVOID1500
RMSE ↓MAE ↓RMSE ↓MAE ↓RMSE ↓MAE ↓
CFormal0.1200.2329.6063.3280.7260.261
LRRU0.1320.2459.1642.7380.6980.232
BP-Net0.1220.2128.9032.7120.7040.230
Ours0.1010.2107.7662.4980.6760.225
+ +Network Architectures. We evaluate the compatibility of our method with different network architectures, as shown in Table 4. Since we focus on training strategies, we have the flexibility to various models, demonstrating the robustness of our approach across different network designs. Specifically, we evaluate BP-Net [50], LRRU [56], and CFormer (L1 Loss) [79]. BP-Net is one of the most advanced methods, while LRRU and CFormer are considered representative architectures. Our approach consistently improves performance across all three models. Notably, BP-Net with our method achieves the best performance in terms of RMSE, the primary evaluation metric. + +Generalization to Out-of-Domain Datasets. To evaluate the generalization ability of our method, we conduct zero-shot testing on unseen datasets, including ScanNet [13], DDAD [17], and Void1500 [60], as shown in Table 5. Our approach consistently achieves the best performance across all datasets, demonstrating lower RMSE and MAE compared to prior methods. Notably, our method outperforms BP-Net by a significant margin on DDAD, reducing RMSE from 8.903 to 7.766, highlighting its effectiveness in handling unseen scenarios. + +Application on Dense SLAM. Figure 6 compares a representative SLAM method, Droid-SLAM [51], with sparse LiDAR points (left) and our method with dense depth com + +![](images/d92af7072b332bf2f41f32096c72c303c026786d5b6f68f8d51edc567ddfc169.jpg) + +![](images/0ea130686549fc0fef5074e0185653c52a55763cd4ab35aa02f2b7a176c2ba83.jpg) + +![](images/a0224d2802d609c9ad64290f5d6cfe7d40eb72a48866adbc93ee88a3ee7b9988.jpg) + +![](images/9c7a26b3be63938907cfbcd1fb7d582443ac4c6c999eb3e3529b115a3b01c42e.jpg) + +![](images/33f3bfd10d5b1f8fa2ee922707ab8852aa4ed75e7436733ee8dfb37e47ff8152.jpg) + +![](images/5a784980738bde6ae8ebc381e0fcb3ecf2cce3f1f20fa47da9fadc6ea884936d.jpg) + +![](images/95986492181b4b2136716019c32d031945fa795c696a7225321c772f49b08cfb.jpg) + +![](images/21c687819faf9909dd669d1f6fc5cc002f714832abe6fbad556cf1cc012e499d.jpg) + +![](images/899409fbac316982c51d0e9460936d2e93bf0d9676b5cea745fdc121fcb4093c.jpg) + +![](images/34e24f874c8be7a793a77a9bc67d3438696c2ceac104a8b6e47ff0be7f2a0c5b.jpg) + +![](images/3351ae1579c1b060aadc6fce947e3cc233267af41a0aa1571191acde6ad21637.jpg) + +![](images/660f3820d343b76e3657dad90e0cf4234e5791e78ad35c88d47fcbf3e11937f5.jpg) + +![](images/b62251f35aadc5fe7e3ebf13fe2a66c650052727c1df9c14a0cc716295bc7112.jpg) + +![](images/4aedd7be838c006ec45932487550d848e596a8d1a53c170d9f9d04ccb5661cef.jpg) +(a) Input + +![](images/02e4c30d5be755f7a545622c2ae6f6cae7c529b6de8fcd8600733f95042b846c.jpg) + +![](images/066040b0d28bacecd4e6720db1fed553b71afed5c4b327b28d3ca4dd8ce756c8.jpg) +(b) CFormer [79] + +![](images/fc756172ad3ce0950f929f2461cd3f4196cc65297420767574e521334ed1b131.jpg) + +![](images/8faa4db24d3df57d5b6d3ed227fec2e931a664d5300e1c2e845b1ea0de0073f4.jpg) +(c) LRRU [56] + +![](images/73169d4e174de121892b9ffa88a0ca2d8dc52c152631dc2a8d2f8db532a75f8b.jpg) + +![](images/f6ec725f2e27c956c91abb95ed2343ae6b943c4fbe83201b5cd0d2763794bf45.jpg) +(d) ImprovingDC [57] + +![](images/7712ba798397df2149c95b19198f648d12bd6d8ec983b5693ee3a9bcaa0888d8.jpg) + +![](images/b8b5b33b17b7f8af3bcad63dbb1d911fa98149ab01488913d1f605d370621fdb.jpg) +(e) BP-Net [50] + +![](images/07fa64ef0fb71eac5aab3cbf56fc8ccb816ed6e23da09e3d1e8009cc6e622744.jpg) + +![](images/9b8ec9b8a9c8224b721c7f100f945cabfd771176c22bec2cbcf522088057c3ed.jpg) +(f) DMD $^3$ C (Ours) + +![](images/059646f4a2f0f773c06757f855ac0d0acabe21d6d0b1a403cbcb5311c0790c1d.jpg) +Figure 5. Qualitative comparison of depth completion methods. This figure demonstrates the performance of various depth completion models, including CFormer, LRRU, ImprovingDC, BP-Net, and our proposed DMD $^3$ C. For each method, we show the input RGB images with sparse LiDAR points (left), along with the resulting completed depth maps and corresponding 3D point cloud reconstructions. + +![](images/884076331dc16b92c35e42bb4f59865ae59245b52b8a0b8332d951ee604918fe.jpg) + +![](images/4f6e332ef06753404b1638d5d26dfa497c774eed8874ad61a48a5d459f0e8126.jpg) + +![](images/9c76f7f9112fb0f990455823736b89bafb1619938bd6f035fbe13592af279b67.jpg) + +![](images/59ab0a551e3990bbdf80dec89dd443ac6bcefba92229432d03362cd3668371b3.jpg) +Droid-SLAM + LiDAR +Figure 6. Qualitative comparison of Droid-SLAM with sparse LiDAR points and our method with dense depth completion. + +![](images/3a510b5c300e4212d14ca70c8b1bf863b9d386c7fa31f3e000792bdba68d6003.jpg) + +![](images/58ed8c4656c66163632517d07e030b95843f0227dcfe3f28ce470a8a4abb1931.jpg) +Droid-SLAM + Ours + +![](images/7949bc9af021892dc8a2903c60c412d842c5eed246ae25c68e89b7d235af37dd.jpg) + +pletion (right). Droid-SLAM suffers from structural distortions and noise, especially in occluded or low-texture areas (red boxes), due to limited depth information. In contrast, our method generates more consistent reconstructions, preserving structural details and improving alignment (green boxes). These results highlight the advantage of dense depth completion in enhancing SLAM quality. + +# 5. Conclusion + +In this work, we present $\mathrm{DMD}^{3}\mathrm{C}$ , a novel two-stage distillation framework that distills knowledge from monocular depth estimation foundation models into the depth completion task. In the first stage, we introduce a data generation strategy that leverages monocular depth estimation and mesh reconstruction to simulate training data, allowing the model to learn geometric features from diverse natural images. In the second stage, we propose a scale- and shift-invariant loss (SSI Loss) combined with a supervised L1 loss with sparse ground truth, which addresses the scale ambiguity in monocular depth estimation. This ensures consistent depth completions across varying scales and focuses on learning real-world scale information. Extensive experiments on the depth completion benchmarks demonstrate that $\mathrm{DMD}^{3}\mathrm{C}$ achieves state-of-the-art performance, ranking first on the KITTI leaderboard, and significantly outperforming existing methods. Our results highlight the framework's ability to produce high-quality depth maps with improved detail and structural consistency, making it a promising solution for depth completion tasks. + +# Acknowledgements + +This work was supported by the National Key R&D Program of China (2022YFC3300704), the National Natural Science Foundation of China (62331006, 62171038, and 62088101), and the Fundamental Research Funds for the Central Universities. + +# References + +[1] Luca Bartolomei, Matteo Poggi, Andrea Conti, Fabio Tosi, and Stefano Mattoccia. Revisiting depth completion from a stereo matching perspective for cross-domain generalization. In International Conference on 3D Vision (3DV), pages 1360-1370, 2024. 2 +[2] Shariq Farooq Bhat, Reiner Birkl, Diana Wofk, Peter Wonka, and Matthias Müller. Zoedepth: Zero-shot transfer by combining relative and metric depth. arXiv preprint arXiv:2302.12288, 2023. 2 +[3] Holger Caesar, Varun Bankiti, Alex H Lang, Sourabh Vora, Venice Erin Liong, Qiang Xu, Anush Krishnan, Yu Pan, Giancarlo Baldan, and Oscar Beijbom. nuscenes: A multimodal dataset for autonomous driving. In the IEEE Conference on Computer Vision and Pattern Recognition (CVPR), pages 11621-11631, 2020. 5 +[4] Angel Chang, Angela Dai, Thomas Funkhouser, Maciej Halber, Matthias Niebner, Manolis Savva, Shuran Song, Andy Zeng, and Yinda Zhang. Matterport3d: Learning from rgb-d data in indoor environments. In International Conference on 3D Vision (3DV), pages 667-676, 2017. 2 +[5] Linwei Chen, Ying Fu, Kaixuan Wei, Dezhi Zheng, and Felix Heide. Instance segmentation in the dark. International Journal of Computer Vision, 131(8):2198-2218, 2023. 2 +[6] Linwei Chen, Ying Fu, Lin Gu, Chenggang Yan, Tatsuya Harada, and Gao Huang. Frequency-aware feature fusion for dense image prediction. IEEE Transactions on Pattern Analysis and Machine Intelligence, 46(12):10763-10780, 2024. 2 +[7] Yun Chen, Bin Yang, Ming Liang, and Raquel Urtasun. Learning joint 2d-3d representations for depth completion. In the IEEE Conference on Computer Vision and Pattern Recognition (CVPR), pages 10023-10032, 2019. 1 +[8] Xinjing Cheng, Peng Wang, and Ruigang Yang. Learning depth with convolutional spatial propagation network. IEEE Transactions on Pattern Analysis and Machine Intelligence (TPAMI), 42(10):2361-2379, 2019. 2, 6 +[9] Xinjing Cheng, Peng Wang, Chenye Guan, and Ruigang Yang. Cspn++: Learning context and resource aware convolutional spatial propagation networks for depth completion. In the AAAI Conference on Artificial Intelligence (AAAI), pages 10615-10622, 2020. 2, 6 +[10] Keunhoon Choi, Somi Jeong, Youngjung Kim, and Kwanghoon Sohn. Stereo-augmented depth completion from a single rgb-lidar image. In IEEE International Conference on Robotics and Automation (ICRA), pages 13641-13647, 2021. 2 +[11] Andrea Conti, Matteo Poggi, and Stefano Mattoccia. Sparsity agnostic depth completion. In the IEEE Winter Conference on Applications of Computer Vision (WACV), pages 5871-5880, 2023. 2 +[12] Marius Cordts, Mohamed Omran, Sebastian Ramos, Timo Rehfeld, Markus Enzweiler, Rodrigo Benenson, Uwe Franke, Stefan Roth, and Bernt Schiele. The cityscapes dataset for semantic urban scene understanding. In the IEEE Conference on Computer Vision and Pattern Recognition (CVPR), pages 3213-3223, 2016. 5 + +[13] Angela Dai, Angel X Chang, Manolis Savva, Maciej Halber, Thomas Funkhouser, and Matthias Nießner. Scannet: Richly-annotated 3d reconstructions of indoor scenes. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 5828-5839, 2017. 7 +[14] Zheng Fang, Ying Fu, and Lixiong Liu. A dual of transformer features-related map-intelligent generation method. Journal of Image and Graphics, 28(10):3281-3294, 2023. 2 +[15] Ying Fu, Zichun Wang, Tao Zhang, and Jun Zhang. Lowlight raw video denoising with a high-quality realistic motion dataset. IEEE Transactions on Multimedia, 25:8119-8131, 2022. 2 +[16] Qi Guan, Zihao Sheng, and Shibei Xue. Hrpose: Real-time high-resolution 6d pose estimation network using knowledge distillation. Chinese Journal of Electronics, 32(1):189-198, 2023. 2 +[17] Vitor Guizilini, Rares Ambrus, Sudeep Pillai, Allan Raventos, and Adrien Gaidon. 3d packing for self-supervised monocular depth estimation. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 2485-2494, 2020. 7 +[18] Wenxuan Guo, Yingping Liang, Zhiyu Pan, Ziheng Xi, Jianjiang Feng, and Jie Zhou. Camera-lidar cross-modality gait recognition. arXiv preprint arXiv:2407.02038, 2024. 2 +[19] Wenxuan Guo, Zhiyu Pan, Yingping Liang, Ziheng Xi, Zhicheng Zhong, Jianjiang Feng, and Jie Zhou. Lidar-based person re-identification. In the IEEE Conference on Computer Vision and Pattern Recognition (CVPR), pages 17437–17447, 2024. 2 +[20] Kaiming He, Xiangyu Zhang, Shaoqing Ren, and Jian Sun. Deep residual learning for image recognition. In the IEEE Conference on Computer Vision and Pattern Recognition (CVPR), pages 770-778, 2016. 5 +[21] Junjie Hu, Chenyu Bao, Mete Ozay, Chenyou Fan, Qing Gao, Honghai Liu, and Tin Lun Lam. Deep depth completion from extremely sparse data: A survey. IEEE Transactions on Pattern Analysis and Machine Intelligence (TPAMI), 45(7): 8244-8264, 2022. 1, 2 +[22] Mu Hu, Wei Yin, Chi Zhang, Zhipeng Cai, Xiaoxiao Long, Hao Chen, Kaixuan Wang, Gang Yu, Chunhua Shen, and Shaojie Shen. Metric3d v2: A versatile monocular geometric foundation model for zero-shot metric depth and surface normal estimation. IEEE Transactions on Pattern Analysis and Machine Intelligence (TPAMI), pages 1-18, 2024. 2 +[23] Jinwoo Jeon, Hyunjun Lim, Dong-Uk Seo, and Hyun Myung. Struct-mdc: Mesh-refined unsupervised depth completion leveraging structural regularities from visual slam. IEEE Robotics and Automation Letters, 7(3):6391-6398, 2022. 1 +[24] Michael Kazhdan, Matthew Bolitho, and Hugues Hoppe. Poisson surface reconstruction. In Eurographics Symposium on Geometry Processing, 2006. 4 +[25] Bingxin Ke, Anton Obukhov, Shengyu Huang, Nando Metzger, Rodrigo Caye Daudt, and Konrad Schindler. Repurposing diffusion-based image generators for monocular depth estimation. In the IEEE Conference on Computer Vision and Pattern Recognition (CVPR), pages 9492-9502, 2024. 1, 2 + +[26] Muhammad Ahmed Ullah Khan, Danish Nazir, Alain Pagani, Hamam Mokayed, Marcus Liwicki, Didier Stricker, and Muhammad Zeshan Afzal. A comprehensive survey of depth completion approaches. Sensors, 22(18):6969, 2022. 1, 2 +[27] Byeong-Uk Lee, Hae-Gon Jeon, Sunghoon Im, and In So Kweon. Depth completion with deep geometry and context guidance. In International Conference on Robotics and Automation (ICRA), pages 3281-3287. IEEE, 2019. 1 +[28] Byeong-Uk Lee, Kyunghyun Lee, and In So Kweon. Depth completion using plane-residual representation. In the IEEE Conference on Computer Vision and Pattern Recognition (CVPR), pages 13916-13925, 2021. 1 +[29] Miaoyu Li and Ying Fu. Joint self-attention transformer for multispectral and hyperspectral image fusion. Journal of Image and Graphics, 28(12):3922–3934, 2023. 2 +[30] Miaoyu Li, Ying Fu, Tao Zhang, and Guanghui Wen. Supervise-assisted self-supervised deep-learning method for hyperspectral image restoration. IEEE Transactions on Neural Networks and Learning Systems, 2024. 1 +[31] Yingping Liang and Ying Fu. Relation-guided adversarial learning for data-free knowledge transfer. International Journal of Computer Vision, pages 1-18, 2024. 3 +[32] Yingping Liang, Jiaming Liu, Debing Zhang, and Ying Fu. Mpi-flow: Learning realistic optical flow with multiplane images. In the IEEE International Conference on Computer Vision (ICCV), pages 13857-13868, 2023. 2, 3 +[33] Tsung-Yi Lin, Michael Maire, Serge Belongie, James Hays, Pietro Perona, Deva Ramanan, Piotr Dólar, and C Lawrence Zitnick. Microsoft coco: Common objects in context. In the European Conference on Computer Vision (ECCV), pages 740-755, 2014. 5 +[34] Yuankai Lin, Tao Cheng, Qi Zhong, Wending Zhou, and Hua Yang. Dynamic spatial propagation network for depth completion. In the AAAI Conference on Artificial Intelligence (AAAI), pages 1638-1646, 2022. 2, 6 +[35] Lina Liu, Xibin Song, Xiaoyang Lyu, Junwei Diao, Mengmeng Wang, Yong Liu, and Liangjun Zhang. Fcfr-net: Feature fusion based coarse-to-fine residual learning for depth completion. In the AAAI Conference on Artificial Intelligence (AAAI), pages 2136-2144, 2021. 6 +[36] Qiankun Liu, Yuqi Jiang, Zhentao Tan, Dongdong Chen, Ying Fu, Qi Chu, Gang Hua, and Nenghai Yu. Transformer based pluralistic image completion with reduced information loss. IEEE Transactions on Pattern Analysis and Machine Intelligence, 46(10):6652-6668, 2024. +[37] Qiankun Liu, Yichen Li, Yuqi Jiang, and Ying Fu. Siamesedetr for generic multi-object tracking. IEEE Transactions on Image Processing, 33:3935-3949, 2024. +[38] Fangchang Ma and Sertac Karaman. Sparse-to-dense: Depth prediction from sparse depth samples and a single image. In IEEE international conference on robotics and automation (ICRA), pages 4796-4803, 2018. 2, 6 +[39] Fabiola Maffra, Lucas Teixeira, Zetao Chen, and Margarita Chli. Real-time wide-baseline place recognition using depth completion. IEEE Robotics and Automation Letters, 4(2): 1525-1532, 2019. 1 + +[40] Hyoungseob Park, Anjali Gupta, and Alex Wong. Test-time adaptation for depth completion. In the IEEE Conference on Computer Vision and Pattern Recognition (CVPR), pages 20519-20529, 2024. 2 +[41] Jinsun Park, Kyungdon Joo, Zhe Hu, Chi-Kuei Liu, and In So Kweon. Non-local spatial propagation network for depth completion. In the European Conference on Computer Vision (ECCV), pages 120-136, 2020. 6 +[42] Jin-Hwi Park and Hae-Gon Jeon. A simple yet universal framework for depth completion. In The Thirty-eighth Annual Conference on Neural Information Processing Systems, 2024. 2 +[43] Federico Perazzi, Jordi Pont-Tuset, Brian McWilliams, Luc Van Gool, Markus Gross, and Alexander Sorkine-Hornung. A benchmark dataset and evaluation methodology for video object segmentation. In the IEEE Conference on Computer Vision and Pattern Recognition (CVPR), pages 724-732, 2016. 5 +[44] Luigi Piccinelli, Yung-Hsu Yang, Christos Sakaridis, Mattia Segu, Siyuan Li, Luc Van Gool, and Fisher Yu. Undepth: Universal monocular metric depth estimation. In the IEEE Conference on Computer Vision and Pattern Recognition (CVPR), pages 10106-10116, 2024. 2 +[45] Jiaxiong Qiu, Zhaopeng Cui, Yinda Zhang, Xingdi Zhang, Shuaicheng Liu, Bing Zeng, and Marc Pollefeys. Deeplidar: Deep surface normal guided depth prediction for outdoor scene from sparse lidar data and single color image. In the IEEE Conference on Computer Vision and Pattern Recognition (CVPR), pages 3313-3322, 2019. 2, 6 +[46] René Ranftl, Katrin Lasinger, David Hafner, Konrad Schindler, and Vladlen Koltun. Towards robust monocular depth estimation: Mixing datasets for zero-shot cross-dataset transfer. IEEE Transactions on Pattern Analysis and Machine Intelligence (TPAMI), 44(3), 2022. 1, 2, 3, 5 +[47] Ali Osman Serhatoglu, Oguzhan Guclu, and Ahmet Burak Can. Rgb-d slam with deep depth completion. In International Conference on Artificial Intelligence and Soft Computing (ICAISC), pages 59-67, 2022. 1 +[48] Nathan Silberman, Derek Hoiem, Pushmeet Kohli, and Rob Fergus. Indoor segmentation and support inference from rgbd images. In the European Conference on Computer Vision (ECCV), pages 120-136, 2012. 2, 5, 6 +[49] Jie Tang, Fei-Peng Tian, Wei Feng, Jian Li, and Ping Tan. Learning guided convolutional network for depth completion. IEEE Transactions on Image Processing (TIP), 30: 1116-1129, 2020. 2, 6 +[50] Jie Tang, Fei-Peng Tian, Boshi An, Jian Li, and Ping Tan. Bilateral propagation network for depth completion. In the IEEE Conference on Computer Vision and Pattern Recognition (CVPR), pages 9763-9772, 2024. 1, 2, 5, 6, 7, 8 +[51] Zachary Teed and Jia Deng. Droid-slam: Deep visual slam for monocular, stereo, and rgb-d cameras. Advances in Neural Information Processing Systems (NeurIPS), 34:16558-16569, 2021. 7 +[52] Ye Tian, Ying Fu, and Jun Zhang. Transformer-based undersampled single-pixel imaging. Chinese Journal of Electronics, 32(5):1151-1159, 2023. 2 + +[53] Jonas Uhrig, Nick Schneider, Lukas Schneider, Uwe Franke, Thomas Brox, and Andreas Geiger. Sparsity invariant cnns. In International Conference on 3D Vision (3DV), 2017. 1, 2, 3, 5, 6 +[54] Haotian Wang, Meng Yang, and Nanning Zheng. G2-monodepth: A general framework of generalized depth inference from monocular $\mathrm{rgb + x}$ data. IEEE Transactions on Pattern Analysis and Machine Intelligence (TPAMI), 2023. 2, 3 +[55] Haotian Wang, Meng Yang, Xinhu Zheng, and Gang Hua. Scale propagation network for generalizable depth completion. IEEE Transactions on Pattern Analysis and Machine Intelligence (TPAMI), 2024. 2 +[56] Yufei Wang, Bo Li, Ge Zhang, Qi Liu, Tao Gao, and Yuchao Dai. Lrru: Long-short range recurrent updating networks for depth completion. In the IEEE Conference on Computer Vision and Pattern Recognition (CVPR), pages 9422-9432, 2023. 1, 6, 7, 8 +[57] Yufei Wang, Ge Zhang, Shaoqian Wang, Bo Li, Qi Liu, Le Hui, and Yuchao Dai. Improving depth completion via depth feature upsampling. In the IEEE Conference on Computer Vision and Pattern Recognition (CVPR), pages 21104-21113, 2024. 1, 2, 5, 6, 7, 8 +[58] Xiu-Shen Wei, Yi-Zhe Song, Oisin Mac Aodha, Jianxin Wu, Yuxin Peng, Jinhui Tang, Jian Yang, and Serge Belongie. Fine-grained image analysis with deep learning: A survey. IEEE transactions on pattern analysis and machine intelligence, 44(12):8927-8948, 2021. 1 +[59] Tobias Weyand, Andre Araujo, Bingyi Cao, and Jack Sim. Google landmarks dataset v2-a large-scale benchmark for instance-level recognition and retrieval. In the IEEE Conference on Computer Vision and Pattern Recognition (CVPR), pages 2575-2584, 2020. 5 +[60] Alex Wong, Xiaohan Fei, Stephanie Tsuei, and Stefano Soatto. Unsupervised depth completion from visual inertial odometry. IEEE Robotics and Automation Letters, 5(2): 1899-1906, 2020. 7 +[61] Xiaopei Wu, Liang Peng, Honghui Yang, Liang Xie, Chenxi Huang, Chengqi Deng, Haifeng Liu, and Deng Cai. Sparse fuse dense: Towards high quality 3d detection with depth completion. In the IEEE Conference on Computer Vision and Pattern Recognition (CVPR), pages 5418-5427, 2022. 1 +[62] Yangchao Wu, Tian Yu Liu, Hyoungseob Park, Stefano Soatto, Dong Lao, and Alex Wong. Augundo: Scaling up augmentations for monocular depth completion and estimation. In European Conference on Computer Vision (ECCV), pages 274-293, 2025. 2 +[63] Weijian Xie, Guanyi Chu, Quanhao Qian, Yihao Yu, Hai Li, Danpeng Chen, Shangjin Zhai, Nan Wang, Hujun Bao, and Guofeng Zhang. Depth completion with multiple balanced bases and confidence for dense monocular slam. IEEE Transactions on Visualization and Computer Graphics (TVCG), 2024. 1 +[64] Zexiao Xie, Xiaoxuan Yu, Xiang Gao, Kunqian Li, and Shuhan Shen. Recent advances in conventional and deep learning-based depth completion: A survey. IEEE Transactions on Neural Networks and Learning Systems (TNNLS), 35(3):3395-3415, 2024. 2 + +[65] Yan Xu, Xinge Zhu, Jianping Shi, Guofeng Zhang, Hujun Bao, and Hongsheng Li. Depth completion from sparse lidar data with depth-normal constraints. In the IEEE Conference on Computer Vision and Pattern Recognition (CVPR), pages 2811-2820, 2019. 1 +[66] Zhiqiang Yan, Xiang Li, Kun Wang, Zhenyu Zhang, Jun Li, and Jian Yang. Multi-modal masked pre-training for monocular panoramic depth completion. In European Conference on Computer Vision, pages 378-395. Springer, 2022. 2 +[67] Zhiqiang Yan, Kun Wang, Xiang Li, Zhenyu Zhang, Jun Li, and Jian Yang. Rignet: Repetitive image guided network for depth completion. In the European Conference on Computer Vision (ECCV), pages 214-230, 2022. 1, 6 +[68] Zhiqiang Yan, Xiang Li, Kun Wang, Shuo Chen, Jun Li, and Jian Yang. Distortion and uncertainty aware loss for panoramic depth completion. In International Conference on Machine Learning, pages 39099-39109. PMLR, 2023. +[69] Zhiqiang Yan, Kun Wang, Xiang Li, Zhenyu Zhang, Jun Li, and Jian Yang. Desnet: Decomposed scale-consistent network for unsupervised depth completion. In Proceedings of the AAAI conference on artificial intelligence, pages 3109-3117, 2023. +[70] Zhiqiang Yan, Yuankai Lin, Kun Wang, Yupeng Zheng, Yufei Wang, Zhenyu Zhang, Jun Li, and Jian Yang. Tri-perspective view decomposition for geometry-aware depth completion. In the IEEE Conference on Computer Vision and Pattern Recognition (CVPR), pages 4874-4884, 2024. 1, 2, 6, 7 +[71] Lihe Yang, Bingyi Kang, Zilong Huang, Xiaogang Xu, Jiashi Feng, and Hengshuang Zhao. Depth anything: Unleashing the power of large-scale unlabeled data. In the IEEE Conference on Computer Vision and Pattern Recognition (CVPR), pages 10371-10381, 2024. 1, 2, 3 +[72] Lihe Yang, Bingyi Kang, Zilong Huang, Zhen Zhao, Xiaogang Xu, Jiashi Feng, and Hengshuang Zhao. Depth anything v2. In Annual Conference on Neural Information Processing Systems (NeurIPS), 2024. 1, 2, 3, 5 +[73] Wei Yin, Chi Zhang, Hao Chen, Zhipeng Cai, Gang Yu, Kaixuan Wang, Xiaozhi Chen, and Chunhua Shen. Metric3d: Towards zero-shot metric 3d prediction from a single image. In the IEEE Conference on Computer Vision and Pattern Recognition (CVPR), pages 9043-9053, 2023. 2 +[74] Zhu Yu, Zehua Sheng, Zili Zhou, Lun Luo, Si-Yuan Cao, Hong Gu, Huaqi Zhang, and Hui-Liang Shen. Aggregating feature point cloud for depth completion. In the IEEE International Conference on Computer Vision (ICCV), pages 8732-8743, 2023. 2, 6 +[75] Fan Zhang, Shaodi You, Yu Li, and Ying Fu. Atlantis: Enabling underwater depth estimation with stable diffusion. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 11852-11861, 2024. 1 +[76] Tao Zhang, Ying Fu, and Jun Zhang. Deep guided attention network for joint denoising and demosaicing in real image. Chinese Journal of Electronics, 33(1):303-312, 2024. 2 +[77] Xuanmeng Zhang, Zhedong Zheng, Minyue Jiang, and Xiaojing Ye. Self-ensembling depth completion via density-aware consistency. Pattern Recognition, 154:110618, 2024. 2 + +[78] Yinda Zhang and Thomas Funkhouser. Deep depth completion of a single rgb-d image. In the IEEE Conference on Computer Vision and Pattern Recognition (CVPR), pages 175-185, 2018. 1 +[79] Youmin Zhang, Xianda Guo, Matteo Poggi, Zheng Zhu, Guan Huang, and Stefano Mattoccia. Completionformer: Depth completion with convolutions and vision transformers. In the IEEE Conference on Computer Vision and Pattern Recognition (CVPR), pages 18527-18536, 2023. 1, 2, 5, 6, 7, 8 +[80] Shanshan Zhao, Mingming Gong, Huan Fu, and Dacheng Tao. Adaptive context-aware multi-modal network for depth completion. IEEE Transactions on Image Processing (TIP), 30:5264-5276, 2021. 2, 6 +[81] Jin Zheng, Botao Jiang, Wei Peng, and Qiaohui Zhang. Multi-scale binocular stereo matching based on semantic association. Chinese Journal of Electronics, 33(4):1010-1022, 2024. 2 +[82] Wending Zhou, Xu Yan, Yinghong Liao, Yuankai Lin, Jin Huang, Gangming Zhao, Shuguang Cui, and Zhen Li. Bev@ dc: Bird's-eye view assisted training for depth completion. In the IEEE Conference on Computer Vision and Pattern Recognition (CVPR), pages 9233-9242, 2023. 2, 6 +[83] Yunhao Zou, Ying Fu, Tsuyoshi Takatani, and Yinqiang Zheng. Eventhrd: From event to high-speed hdr videos and beyond. IEEE Transactions on Pattern Analysis and Machine Intelligence, 47(1):32-50, 2024. 2 +[84] Yiming Zuo and Jia Deng. Ogni-dc: Robust depth completion with optimization-guided neural iterations. In the European Conference on Computer Vision (ECCV), pages 78-95, 2025. 2 \ No newline at end of file diff --git a/2025/Distilling Monocular Foundation Model for Fine-grained Depth Completion/images.zip b/2025/Distilling Monocular Foundation Model for Fine-grained Depth Completion/images.zip new file mode 100644 index 0000000000000000000000000000000000000000..021ed33d39888f3c5bd991967410cf2b3ca51a43 --- /dev/null +++ b/2025/Distilling Monocular Foundation Model for Fine-grained Depth Completion/images.zip @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:f58b112674a32d87f1eb14054c6895fd26e6cf239c853168b3163807fb1fa6e4 +size 833298 diff --git a/2025/Distilling Monocular Foundation Model for Fine-grained Depth Completion/layout.json b/2025/Distilling Monocular Foundation Model for Fine-grained Depth Completion/layout.json new file mode 100644 index 0000000000000000000000000000000000000000..3743e08caad3a65d8dc1481335bdf56940b344ab --- /dev/null +++ b/2025/Distilling Monocular Foundation Model for Fine-grained Depth Completion/layout.json @@ -0,0 +1,11123 @@ +{ + "pdf_info": [ + { + "para_blocks": [ + { + "bbox": [ + 70, + 103, + 539, + 121 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 70, + 103, + 539, + 121 + ], + "spans": [ + { + "bbox": [ + 70, + 103, + 539, + 121 + ], + "type": "text", + "content": "Distilling Monocular Foundation Model for Fine-grained Depth Completion" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 61, + 141, + 556, + 200 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 61, + 141, + 556, + 200 + ], + "spans": [ + { + "bbox": [ + 61, + 141, + 556, + 200 + ], + "type": "text", + "content": "Yingping Liang1 Yutao Hu2 Wenqi Shao3 Ying Fu1† \n1Beijing Institute of Technology 2Key Laboratory of New Generation Artificial Intelligence Technology and Its Interdisciplinary Applications, Southeast University 3Shanghai Al Laboratory {liangyingping, fuying}@bit.edu.cn huyutao@seu.edu.cn weqish@link.cuhk.edu.hk" + } + ] + } + ], + "index": 3 + }, + { + "type": "image", + "bbox": [ + 69, + 227, + 544, + 329 + ], + "blocks": [ + { + "bbox": [ + 69, + 227, + 544, + 329 + ], + "lines": [ + { + "bbox": [ + 69, + 227, + 544, + 329 + ], + "spans": [ + { + "bbox": [ + 69, + 227, + 544, + 329 + ], + "type": "image", + "image_path": "9838f5956b4ecd56b7df0dcbb9f8556c79e9ba628572bfb75045a8d65b7240cf.jpg" + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 54, + 338, + 555, + 373 + ], + "lines": [ + { + "bbox": [ + 54, + 338, + 555, + 373 + ], + "spans": [ + { + "bbox": [ + 54, + 338, + 555, + 373 + ], + "type": "text", + "content": "Figure 1. Depth completion models trained solely with " + }, + { + "bbox": [ + 54, + 338, + 555, + 373 + ], + "type": "inline_equation", + "content": "L_{1}" + }, + { + "bbox": [ + 54, + 338, + 555, + 373 + ], + "type": "text", + "content": " loss and sparse ground truth produce incomplete and fragmented depth predictions. Our framework, however, demonstrates significant improvements by distilling knowledge from monocular foundation models and incorporating a scale- and shift-invariant loss (SSI Loss), resulting in more complete and accurate dense depth completion." + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "image_caption" + } + ], + "index": 4 + }, + { + "bbox": [ + 151, + 381, + 200, + 395 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 151, + 381, + 200, + 395 + ], + "spans": [ + { + "bbox": [ + 151, + 381, + 200, + 395 + ], + "type": "text", + "content": "Abstract" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 53, + 409, + 297, + 685 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 409, + 297, + 685 + ], + "spans": [ + { + "bbox": [ + 53, + 409, + 297, + 685 + ], + "type": "text", + "content": "Depth completion involves predicting dense depth maps from sparse LiDAR inputs. However, sparse depth annotations from sensors limit the availability of dense supervision, which is necessary for learning detailed geometric features. In this paper, we propose a two-stage knowledge distillation framework that leverages powerful monocular foundation models to provide dense supervision for depth completion. In the first stage, we introduce a pre-training strategy that generates diverse training data from natural images, which distills geometric knowledge to depth completion. Specifically, we simulate LiDAR scans by utilizing monocular depth and mesh reconstruction, thereby creating training data without requiring ground-truth depth. Besides, monocular depth estimation suffers from inherent scale ambiguity in real-world settings. To address this, in the second stage, we employ a scale- and shift-invariant loss (SSI Loss) to learn real-world scales when fine-tuning on real-world datasets. Our two-stage distillation framework enables depth completion models to harness the strengths of monocular foundation models. Experimental results demonstrate that models trained with our two-stage distillation framework achieve state-of-the-art performance, ranking first place on the KITTI benchmark. Code is available" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 313, + 383, + 538, + 395 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 383, + 538, + 395 + ], + "spans": [ + { + "bbox": [ + 313, + 383, + 538, + 395 + ], + "type": "text", + "content": "at https://github.com/Sharpiless/DMD3C" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 314, + 417, + 395, + 430 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 314, + 417, + 395, + 430 + ], + "spans": [ + { + "bbox": [ + 314, + 417, + 395, + 430 + ], + "type": "text", + "content": "1. Introduction" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 313, + 437, + 555, + 498 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 437, + 555, + 498 + ], + "spans": [ + { + "bbox": [ + 313, + 437, + 555, + 498 + ], + "type": "text", + "content": "Depth completion is a fundamental task in computer vision, where the goal is to generate dense depth maps from sparse depth measurements. This task is particularly important in applications such as autonomous driving [23, 47, 61], robotics [26, 39, 63], and augmented reality [21, 30]." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 313, + 498, + 556, + 581 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 498, + 556, + 581 + ], + "spans": [ + { + "bbox": [ + 313, + 498, + 556, + 581 + ], + "type": "text", + "content": "Recent approaches [7, 27, 28, 65, 67, 78] leverage deep neural networks to learn from depth data. Some methods [56, 79] also incorporate RGB images to guide the depth completion process. Despite these advancements, state-of-the-art models [50, 57, 70] still face difficulties in capturing fine-grained geometric details, particularly in complex outdoor scenes where depth annotations are sparse [53, 58, 75]." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 313, + 582, + 556, + 700 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 582, + 556, + 700 + ], + "spans": [ + { + "bbox": [ + 313, + 582, + 556, + 700 + ], + "type": "text", + "content": "The challenge arises from the reliance on sparse ground truth for training. The lack of dense ground truth makes it difficult for models to accurately learn depth completion across an entire scene. Monocular depth estimation [25, 46, 71, 72], on the other hand, has the ability to provide dense depth predictions from a single image. Specifically, state-of-the-art monocular foundation models could generate dense depth maps, containing fine-grained details and relative depth relationships, offering valuable guidance for training depth completion networks." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 326, + 701, + 555, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 326, + 701, + 555, + 714 + ], + "spans": [ + { + "bbox": [ + 326, + 701, + 555, + 714 + ], + "type": "text", + "content": "To make full use of the advantages of monocular foun" + } + ] + } + ], + "index": 13 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "spans": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "type": "text", + "content": "CVF" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 145, + 0, + 496, + 38 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 145, + 0, + 496, + 38 + ], + "spans": [ + { + "bbox": [ + 145, + 0, + 496, + 38 + ], + "type": "text", + "content": "This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 70, + 703, + 150, + 713 + ], + "type": "page_footnote", + "angle": 0, + "lines": [ + { + "bbox": [ + 70, + 703, + 150, + 713 + ], + "spans": [ + { + "bbox": [ + 70, + 703, + 150, + 713 + ], + "type": "text", + "content": "† Corresponding author." + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "type": "text", + "content": "22254" + } + ] + } + ], + "index": 15 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 0 + }, + { + "para_blocks": [ + { + "bbox": [ + 55, + 72, + 294, + 228 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 72, + 294, + 228 + ], + "spans": [ + { + "bbox": [ + 55, + 72, + 294, + 228 + ], + "type": "text", + "content": "dation models, we propose a novel two-stage distillation framework to transfer geometric knowledge from monocular foundation models to depth completion networks. Specifically, in the first distillation stage, we generate training data through monocular foundation models and then distill knowledge via the proposed pre-training strategy. Specifically, diverse natural images are used to generate pseudo depth maps. Then, we utilize randomly sampled camera parameters to re-construct the scene with mesh and simulate LiDAR using ray simulation. The generated data trains the depth completion model to learn diverse geometric knowledge from monocular foundation models, enhancing its ability to generalize across different scenes." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 55, + 228, + 294, + 361 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 228, + 294, + 361 + ], + "spans": [ + { + "bbox": [ + 55, + 228, + 294, + 361 + ], + "type": "text", + "content": "However, monocular depth estimation suffers from inherent scale ambiguity [16, 81], resulting in depth predictions that vary greatly in scale. Thus, monocular depth alone cannot serve as a reliable basis for real-world depth. To solve this problem, in the second distillation stage, we introduce a scale- and shift-invariant loss (SSI Loss) [46]. Specifically, when fine-tuning on labeled datasets with sparse ground truth, SSI Loss ignores the scale and shift that causes the least loss from the depth prediction and monocular depth, to ensure consistent depth completions aligning monocular depth supervision across varying scales." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 55, + 360, + 294, + 407 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 360, + 294, + 407 + ], + "spans": [ + { + "bbox": [ + 55, + 360, + 294, + 407 + ], + "type": "text", + "content": "By integrating these two distillation stages into the training process, our method achieves state-of-the-art performance on the KITTI benchmark. The contributions of this work are summarized as follows:" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 55, + 411, + 295, + 638 + ], + "type": "list", + "angle": 0, + "index": 7, + "blocks": [ + { + "bbox": [ + 55, + 411, + 295, + 470 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 411, + 295, + 470 + ], + "spans": [ + { + "bbox": [ + 55, + 411, + 295, + 470 + ], + "type": "text", + "content": "- We propose a novel two-stage distillation framework to transfer knowledge from monocular foundation models to depth completion models, which enables the learning of fine-grained depth information from sparse ground truth by providing dense supervision." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 55, + 472, + 295, + 529 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 472, + 295, + 529 + ], + "spans": [ + { + "bbox": [ + 55, + 472, + 295, + 529 + ], + "type": "text", + "content": "- In the first stage, we propose a data generation strategy that uses monocular depth estimation and mesh reconstruction to simulate training data, enabling the model to learn geometric features from diverse natural images without the need for any LiDAR and ground truth." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 55, + 531, + 295, + 590 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 531, + 295, + 590 + ], + "spans": [ + { + "bbox": [ + 55, + 531, + 295, + 590 + ], + "type": "text", + "content": "- In the second stage, we propose a scale- and shift-invariant loss (SSI Loss) to address the scale ambiguity problem in monocular depth estimation, which ensures consistent depth completions across varying scales and focuses on learning real-world scale information." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 55, + 590, + 295, + 638 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 590, + 295, + 638 + ], + "spans": [ + { + "bbox": [ + 55, + 590, + 295, + 638 + ], + "type": "text", + "content": "- Our method achieves first place on the official KITTI depth completion benchmark, demonstrating the effectiveness of the proposed approach with significant gains in both quantitative metrics and qualitative visualizations." + } + ] + } + ], + "index": 6 + } + ], + "sub_type": "text" + }, + { + "bbox": [ + 55, + 651, + 142, + 662 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 651, + 142, + 662 + ], + "spans": [ + { + "bbox": [ + 55, + 651, + 142, + 662 + ], + "type": "text", + "content": "2. Related Work" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 55, + 670, + 206, + 683 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 670, + 206, + 683 + ], + "spans": [ + { + "bbox": [ + 55, + 670, + 206, + 683 + ], + "type": "text", + "content": "2.1. Depth Completion Datasets" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 55, + 689, + 295, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 689, + 295, + 713 + ], + "spans": [ + { + "bbox": [ + 55, + 689, + 295, + 713 + ], + "type": "text", + "content": "Depth completion has been widely studied in both indoor and outdoor settings. Indoor datasets, such as NYU Depth" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 313, + 72, + 553, + 120 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 72, + 553, + 120 + ], + "spans": [ + { + "bbox": [ + 313, + 72, + 553, + 120 + ], + "type": "text", + "content": "V2 [48] and Matterport3D [4], are popular because dense depth measurements are easier to obtain in controlled environments. NYU Depth V2, for instance, provides dense depth maps from a Kinect sensor along with RGB images." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 313, + 121, + 554, + 288 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 121, + 554, + 288 + ], + "spans": [ + { + "bbox": [ + 313, + 121, + 554, + 288 + ], + "type": "text", + "content": "In outdoor environments, however, obtaining dense depth annotations is more challenging due to limitations in sensing technologies like LiDAR. The KITTI dataset [53] is a key benchmark for outdoor depth completion but suffers from sparse annotations, covering only about " + }, + { + "bbox": [ + 313, + 121, + 554, + 288 + ], + "type": "inline_equation", + "content": "5\\%" + }, + { + "bbox": [ + 313, + 121, + 554, + 288 + ], + "type": "text", + "content": " of the image. To address this, training methods often employ complex post-processing and multi-frame fusion techniques to increase the annotation coverage to around " + }, + { + "bbox": [ + 313, + 121, + 554, + 288 + ], + "type": "inline_equation", + "content": "20\\%" + }, + { + "bbox": [ + 313, + 121, + 554, + 288 + ], + "type": "text", + "content": " at most. Furthermore, ground truth values for long distances and dynamic objects typically cannot be included. The sparse ground truth raises challenges for training fine-grained depth completion models. Therefore, we employ monocular depth estimation techniques for dense distillation to preserve fine-grained details." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 313, + 297, + 466, + 309 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 297, + 466, + 309 + ], + "spans": [ + { + "bbox": [ + 313, + 297, + 466, + 309 + ], + "type": "text", + "content": "2.2. Depth Completion Methods" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 313, + 316, + 554, + 531 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 316, + 554, + 531 + ], + "spans": [ + { + "bbox": [ + 313, + 316, + 554, + 531 + ], + "type": "text", + "content": "Depth completion methods have seen significant advancements [21, 26, 64]. Compared to predicting depth directly from a single RGB image [2, 22, 44, 73], depth completion fuses LiDAR information to obtain more accurate sparse depth cues. Recent methods [1, 6, 8, 9, 11, 14, 29, 34-38, 40-42, 45, 49, 50, 52, 55-57, 62, 66-70, 74, 76, 79, 80, 82, 84] rely on supervised learning, using pairs of sparse depth maps and corresponding dense ground truth. While effective with sufficient labeled data, these methods are limited in real-world applications, where obtaining dense annotations, especially in outdoor environments, is challenging [5, 15, 83]. To reduce the reliance on dense labels, [10, 77] use a teacher-student model, where a teacher network trained on labeled data provides guidance to a student network. However, pseudo-labels in complex or dynamic scenes can be noisy or inaccurate. Therefore, we turn to monocular depth estimation to provide fine-grained supervision for depth completion." + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 313, + 539, + 473, + 552 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 539, + 473, + 552 + ], + "spans": [ + { + "bbox": [ + 313, + 539, + 473, + 552 + ], + "type": "text", + "content": "2.3. Monocular Depth Estimation" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 313, + 558, + 554, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 558, + 554, + 713 + ], + "spans": [ + { + "bbox": [ + 313, + 558, + 554, + 713 + ], + "type": "text", + "content": "Monocular depth estimation [25, 46, 71, 72] has emerged as a powerful tool for generating dense depth maps from single RGB images, making it particularly valuable in scenarios where dense depth annotations are sparse or unavailable [18, 19, 32]. Notable models, such as MiDaS [46] and Depth-Anything [71, 72], have demonstrated impressive capabilities in predicting dense depth maps across a wide range of scenes. G2-MonoDepth [54] have proposed a general framework for monocular RGB-X Data with a scale invariant loss. Despite their versatility, monocular depth estimation methods face the challenge of scale ambiguity, which can limit the accuracy of depth completion, especially when integrated with sparse depth measurements" + } + ] + } + ], + "index": 16 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 749, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 749, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 749, + 317, + 757 + ], + "type": "text", + "content": "22255" + } + ] + } + ], + "index": 17 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 1 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 94, + 69, + 518, + 224 + ], + "blocks": [ + { + "bbox": [ + 94, + 69, + 518, + 224 + ], + "lines": [ + { + "bbox": [ + 94, + 69, + 518, + 224 + ], + "spans": [ + { + "bbox": [ + 94, + 69, + 518, + 224 + ], + "type": "image", + "image_path": "e2e9da7addaca5781d47007324bb52844176ced247576ec4ff2289216dea86a7.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 54, + 239, + 555, + 275 + ], + "lines": [ + { + "bbox": [ + 54, + 239, + 555, + 275 + ], + "spans": [ + { + "bbox": [ + 54, + 239, + 555, + 275 + ], + "type": "text", + "content": "Figure 2. Illustration of our proposed first distillation stage with a data generation strategy to learn geometric features from monocular foundation models, which only requires unlabeled RGB images. We use the estimated monocular depth to re-construct the scene and then simulate the Lidar swap process to generate sparse points for training." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "bbox": [ + 54, + 294, + 294, + 342 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 54, + 294, + 294, + 342 + ], + "spans": [ + { + "bbox": [ + 54, + 294, + 294, + 342 + ], + "type": "text", + "content": "from LiDAR. Therefore, we introduce a scale- and shift-invariant loss inspired by [54], combined with a supervised loss to keep real-world scales when fine-tuning on real-world datasets." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 55, + 353, + 111, + 364 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 353, + 111, + 364 + ], + "spans": [ + { + "bbox": [ + 55, + 353, + 111, + 364 + ], + "type": "text", + "content": "3. Method" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 54, + 373, + 295, + 480 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 54, + 373, + 295, + 480 + ], + "spans": [ + { + "bbox": [ + 54, + 373, + 295, + 480 + ], + "type": "text", + "content": "In this section, we first provide a brief overview of the motivation and formulation of our method. Then, we introduce our two-stage distillation framework. In the first stage, we propose a data generation strategy for pre-training that leverages large-scale, unlabeled images without ground-truth depth annotations, as illustrated in Figure 2. In the second stage, we detail the fine-tuning process on a labeled dataset with sparse ground truth, using monocular depth in combination with SSI Loss, as shown in Figure 3." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 55, + 487, + 212, + 499 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 487, + 212, + 499 + ], + "spans": [ + { + "bbox": [ + 55, + 487, + 212, + 499 + ], + "type": "text", + "content": "3.1. Motivation and Formulation" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 54, + 506, + 295, + 590 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 54, + 506, + 295, + 590 + ], + "spans": [ + { + "bbox": [ + 54, + 506, + 295, + 590 + ], + "type": "text", + "content": "In the depth completion task, the input typically consists of an RGB image " + }, + { + "bbox": [ + 54, + 506, + 295, + 590 + ], + "type": "inline_equation", + "content": "I" + }, + { + "bbox": [ + 54, + 506, + 295, + 590 + ], + "type": "text", + "content": " and a sparse depth map " + }, + { + "bbox": [ + 54, + 506, + 295, + 590 + ], + "type": "inline_equation", + "content": "D_{s}" + }, + { + "bbox": [ + 54, + 506, + 295, + 590 + ], + "type": "text", + "content": " obtained from a LiDAR sensor. The objective is to produce a dense depth map " + }, + { + "bbox": [ + 54, + 506, + 295, + 590 + ], + "type": "inline_equation", + "content": "D_{f}" + }, + { + "bbox": [ + 54, + 506, + 295, + 590 + ], + "type": "text", + "content": " that provides a complete and detailed representation of the scene's depth. To achieve this, a depth completion model " + }, + { + "bbox": [ + 54, + 506, + 295, + 590 + ], + "type": "inline_equation", + "content": "f_{\\theta}" + }, + { + "bbox": [ + 54, + 506, + 295, + 590 + ], + "type": "text", + "content": ", parameterized by " + }, + { + "bbox": [ + 54, + 506, + 295, + 590 + ], + "type": "inline_equation", + "content": "\\theta" + }, + { + "bbox": [ + 54, + 506, + 295, + 590 + ], + "type": "text", + "content": ", takes both the RGB image and sparse depth map as inputs:" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 129, + 600, + 295, + 613 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 129, + 600, + 295, + 613 + ], + "spans": [ + { + "bbox": [ + 129, + 600, + 295, + 613 + ], + "type": "interline_equation", + "content": "D _ {f} = f _ {\\theta} (I, D _ {s}). \\tag {1}", + "image_path": "553bfceafc699e8578e93257ba18d7f424480c9190dec2ab3993bf30addbd3be.jpg" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 55, + 622, + 295, + 646 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 622, + 295, + 646 + ], + "spans": [ + { + "bbox": [ + 55, + 622, + 295, + 646 + ], + "type": "text", + "content": "To guide the model's learning, a sparse supervision loss is defined based on the available sparse depth annotations:" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 106, + 655, + 295, + 669 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 106, + 655, + 295, + 669 + ], + "spans": [ + { + "bbox": [ + 106, + 655, + 295, + 669 + ], + "type": "interline_equation", + "content": "\\mathcal {L} _ {\\sup } = M \\times \\left| D _ {f} - D _ {\\text {s p a r s e}} \\right|, \\tag {2}", + "image_path": "f47088fe8a4114d24baa3fc41df7db6c16aad83706e70c88f59b9045e70901ac.jpg" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 55, + 677, + 296, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 677, + 296, + 715 + ], + "spans": [ + { + "bbox": [ + 55, + 677, + 296, + 715 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 55, + 677, + 296, + 715 + ], + "type": "inline_equation", + "content": "M" + }, + { + "bbox": [ + 55, + 677, + 296, + 715 + ], + "type": "text", + "content": " represents the valid mask for sparse depth ground truth " + }, + { + "bbox": [ + 55, + 677, + 296, + 715 + ], + "type": "inline_equation", + "content": "D_{\\mathrm{sparse}}" + }, + { + "bbox": [ + 55, + 677, + 296, + 715 + ], + "type": "text", + "content": ". This loss aligns the predicted dense depth map " + }, + { + "bbox": [ + 55, + 677, + 296, + 715 + ], + "type": "inline_equation", + "content": "D_{f}" + }, + { + "bbox": [ + 55, + 677, + 296, + 715 + ], + "type": "text", + "content": " with the sparse ground truth, ensuring that the output" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 313, + 294, + 554, + 354 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 294, + 554, + 354 + ], + "spans": [ + { + "bbox": [ + 313, + 294, + 554, + 354 + ], + "type": "text", + "content": "is consistent with the available depth data. However, this sparse supervision provides limited guidance, particularly in outdoor scenes [53], leading to challenges in achieving fine-grained and consistent depth predictions due to the inherent limitations of sparse supervision." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 313, + 357, + 555, + 525 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 357, + 555, + 525 + ], + "spans": [ + { + "bbox": [ + 313, + 357, + 555, + 525 + ], + "type": "text", + "content": "Foundation models for monocular depth estimation [71, 72] provide an alternative approach by generating fine-grained dense depth from single RGB images, which we use as supervision without relying solely on LiDAR. Therefore, in the first stage, we utilize the monocular foundation models to generate diverse and large-scale training data, to provide dense supervision for pre-training, allowing the model to learn geometric features. However, monocular methods suffer from scale ambiguities, which can lead to inconsistencies when combining dense monocular predictions with sparse depth. To address this, in the second stage, we introduce SSI Loss [46] combined with L1 loss when finetuning on labeled datasets with sparse ground truth to mitigate scale inconsistencies and learn real-world scale." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 313, + 544, + 553, + 557 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 544, + 553, + 557 + ], + "spans": [ + { + "bbox": [ + 313, + 544, + 553, + 557 + ], + "type": "text", + "content": "3.2. First Stage: Data Generation and Pre-training" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 313, + 566, + 554, + 662 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 566, + 554, + 662 + ], + "spans": [ + { + "bbox": [ + 313, + 566, + 554, + 662 + ], + "type": "text", + "content": "To leverage diverse images without depth annotations, we introduce a pre-training strategy that utilizes synthesized depth data generated through monocular depth estimation on large-scale natural image datasets inspired by [31, 32]. As illustrated in Figure 2, this pre-training phase enables the model to learn robust geometric features from monocular foundation models across diverse scenes, preparing it for subsequent fine-tuning on real-world datasets." + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 313, + 665, + 555, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 665, + 555, + 714 + ], + "spans": [ + { + "bbox": [ + 313, + 665, + 555, + 714 + ], + "type": "text", + "content": "Specifically, we first utilize a pre-trained monocular depth estimation model, such as Depth Anything V2 [72], to predict depth maps for such images. For each natural image " + }, + { + "bbox": [ + 313, + 665, + 555, + 714 + ], + "type": "inline_equation", + "content": "I_{\\mathrm{un}}" + }, + { + "bbox": [ + 313, + 665, + 555, + 714 + ], + "type": "text", + "content": ", the monocular depth estimation model predicts a" + } + ] + } + ], + "index": 15 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "text", + "content": "22256" + } + ] + } + ], + "index": 16 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 2 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 80, + 70, + 531, + 194 + ], + "blocks": [ + { + "bbox": [ + 80, + 70, + 531, + 194 + ], + "lines": [ + { + "bbox": [ + 80, + 70, + 531, + 194 + ], + "spans": [ + { + "bbox": [ + 80, + 70, + 531, + 194 + ], + "type": "image", + "image_path": "5dfe5fd98d70aa5e864902e159ba1bde7d23d7a0bb74de6ad51205f5afdcf78e.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 54, + 211, + 555, + 266 + ], + "lines": [ + { + "bbox": [ + 54, + 211, + 555, + 266 + ], + "spans": [ + { + "bbox": [ + 54, + 211, + 555, + 266 + ], + "type": "text", + "content": "Figure 3. Illustration of our proposed second distillation stage utilizing foundation models for monocular depth estimation when fine-tuning on labeled datasets. Sparse ground truth provides real-world depth scale with L1 loss. Our method enhances this process by incorporating dense monocular depth for fine-grained supervision. However, monocular depth maps come with inherent scale and shift ambiguities. To address these challenges, we employ a Scale- and Shift-Invariant Loss (SSI Loss) that aligns the predictions with the dense monocular depth to match the real-world depth scale, ensuring more accurate depth completion." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "bbox": [ + 55, + 288, + 209, + 300 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 288, + 209, + 300 + ], + "spans": [ + { + "bbox": [ + 55, + 288, + 209, + 300 + ], + "type": "text", + "content": "corresponding dense depth map " + }, + { + "bbox": [ + 55, + 288, + 209, + 300 + ], + "type": "inline_equation", + "content": "D_{\\mathrm{syn}}" + }, + { + "bbox": [ + 55, + 288, + 209, + 300 + ], + "type": "text", + "content": " .." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 129, + 310, + 294, + 323 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 129, + 310, + 294, + 323 + ], + "spans": [ + { + "bbox": [ + 129, + 310, + 294, + 323 + ], + "type": "interline_equation", + "content": "D _ {\\text {s y n}} = f _ {\\mathrm {m o}} \\left(I _ {\\mathrm {u n}}\\right), \\tag {3}", + "image_path": "0b814506b060037f86a2c185ed20105f7ea75508beae323aa6b99a56110b646a.jpg" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 55, + 333, + 296, + 392 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 333, + 296, + 392 + ], + "spans": [ + { + "bbox": [ + 55, + 333, + 296, + 392 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 55, + 333, + 296, + 392 + ], + "type": "inline_equation", + "content": "f_{\\mathrm{mo}}" + }, + { + "bbox": [ + 55, + 333, + 296, + 392 + ], + "type": "text", + "content": " represents the monocular depth estimation model. These synthesized depth maps, though not accurate in metric scale, capture the relative depth relationships and structural details in the scene, providing valuable supervision signals during pre-training." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 55, + 393, + 296, + 417 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 393, + 296, + 417 + ], + "spans": [ + { + "bbox": [ + 55, + 393, + 296, + 417 + ], + "type": "text", + "content": "To simulate LiDAR scanning, we first sample a random camera intrinsic matrix " + }, + { + "bbox": [ + 55, + 393, + 296, + 417 + ], + "type": "inline_equation", + "content": "K" + }, + { + "bbox": [ + 55, + 393, + 296, + 417 + ], + "type": "text", + "content": " as:" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 119, + 425, + 294, + 464 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 119, + 425, + 294, + 464 + ], + "spans": [ + { + "bbox": [ + 119, + 425, + 294, + 464 + ], + "type": "interline_equation", + "content": "K = \\left[ \\begin{array}{c c c} f _ {x} & 0 & c _ {x} \\\\ 0 & f _ {y} & c _ {y} \\\\ 0 & 0 & 1 \\end{array} \\right], \\tag {4}", + "image_path": "40f9504bd95916de31bf6a1f778bac00915a273e0406e58fe17f369e7117070b.jpg" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 55, + 472, + 295, + 496 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 472, + 295, + 496 + ], + "spans": [ + { + "bbox": [ + 55, + 472, + 295, + 496 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 55, + 472, + 295, + 496 + ], + "type": "inline_equation", + "content": "f_{x}" + }, + { + "bbox": [ + 55, + 472, + 295, + 496 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 55, + 472, + 295, + 496 + ], + "type": "inline_equation", + "content": "f_{y}" + }, + { + "bbox": [ + 55, + 472, + 295, + 496 + ], + "type": "text", + "content": " represent the focal lengths along the " + }, + { + "bbox": [ + 55, + 472, + 295, + 496 + ], + "type": "inline_equation", + "content": "x" + }, + { + "bbox": [ + 55, + 472, + 295, + 496 + ], + "type": "text", + "content": "- and " + }, + { + "bbox": [ + 55, + 472, + 295, + 496 + ], + "type": "inline_equation", + "content": "y" + }, + { + "bbox": [ + 55, + 472, + 295, + 496 + ], + "type": "text", + "content": "-axes, and " + }, + { + "bbox": [ + 55, + 472, + 295, + 496 + ], + "type": "inline_equation", + "content": "(c_{x}, c_{y})" + }, + { + "bbox": [ + 55, + 472, + 295, + 496 + ], + "type": "text", + "content": " denotes the center of the image." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 55, + 497, + 295, + 604 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 497, + 295, + 604 + ], + "spans": [ + { + "bbox": [ + 55, + 497, + 295, + 604 + ], + "type": "text", + "content": "For each pixel " + }, + { + "bbox": [ + 55, + 497, + 295, + 604 + ], + "type": "inline_equation", + "content": "(u,v)" + }, + { + "bbox": [ + 55, + 497, + 295, + 604 + ], + "type": "text", + "content": " in the depth map, the corresponding 3D coordinates " + }, + { + "bbox": [ + 55, + 497, + 295, + 604 + ], + "type": "inline_equation", + "content": "(X,Y,Z)" + }, + { + "bbox": [ + 55, + 497, + 295, + 604 + ], + "type": "text", + "content": " in the camera coordinate system are derived using the sampled camera intrinsic matrix " + }, + { + "bbox": [ + 55, + 497, + 295, + 604 + ], + "type": "inline_equation", + "content": "K" + }, + { + "bbox": [ + 55, + 497, + 295, + 604 + ], + "type": "text", + "content": ". This transformation generates a point cloud " + }, + { + "bbox": [ + 55, + 497, + 295, + 604 + ], + "type": "inline_equation", + "content": "P = \\{(X,Y,Z)\\}" + }, + { + "bbox": [ + 55, + 497, + 295, + 604 + ], + "type": "text", + "content": " that represents the 3D spatial locations of each pixel. Leveraging the 3D point cloud " + }, + { + "bbox": [ + 55, + 497, + 295, + 604 + ], + "type": "inline_equation", + "content": "P" + }, + { + "bbox": [ + 55, + 497, + 295, + 604 + ], + "type": "text", + "content": ", we reconstruct a mesh " + }, + { + "bbox": [ + 55, + 497, + 295, + 604 + ], + "type": "inline_equation", + "content": "M" + }, + { + "bbox": [ + 55, + 497, + 295, + 604 + ], + "type": "text", + "content": " via surface reconstruction techniques, such as Poisson Surface Reconstruction [24], to create a continuous 3D surface model for simulation." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 55, + 605, + 295, + 700 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 605, + 295, + 700 + ], + "spans": [ + { + "bbox": [ + 55, + 605, + 295, + 700 + ], + "type": "text", + "content": "Next, we simulate a LiDAR sensor by generating a set of ray direction vectors " + }, + { + "bbox": [ + 55, + 605, + 295, + 700 + ], + "type": "inline_equation", + "content": "\\mathbf{d}" + }, + { + "bbox": [ + 55, + 605, + 295, + 700 + ], + "type": "text", + "content": ". Each ray is cast from the origin along " + }, + { + "bbox": [ + 55, + 605, + 295, + 700 + ], + "type": "inline_equation", + "content": "\\mathbf{d}_{i,j}" + }, + { + "bbox": [ + 55, + 605, + 295, + 700 + ], + "type": "text", + "content": ", intersecting the mesh " + }, + { + "bbox": [ + 55, + 605, + 295, + 700 + ], + "type": "inline_equation", + "content": "M" + }, + { + "bbox": [ + 55, + 605, + 295, + 700 + ], + "type": "text", + "content": " to determine the distance to the intersection point. This distance is recorded as the simulated LiDAR depth reading " + }, + { + "bbox": [ + 55, + 605, + 295, + 700 + ], + "type": "inline_equation", + "content": "D_{\\mathrm{simu}}(i,j)" + }, + { + "bbox": [ + 55, + 605, + 295, + 700 + ], + "type": "text", + "content": ". The resulting sparse depth map " + }, + { + "bbox": [ + 55, + 605, + 295, + 700 + ], + "type": "inline_equation", + "content": "D_{\\mathrm{simu}}" + }, + { + "bbox": [ + 55, + 605, + 295, + 700 + ], + "type": "text", + "content": " emulates LiDAR depth readings from the reconstructed 3D surface, to provide a synthetic depth image comparable to actual LiDAR data." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 67, + 701, + 295, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 67, + 701, + 295, + 713 + ], + "spans": [ + { + "bbox": [ + 67, + 701, + 295, + 713 + ], + "type": "text", + "content": "Then, the depth completion model " + }, + { + "bbox": [ + 67, + 701, + 295, + 713 + ], + "type": "inline_equation", + "content": "f_{\\theta}" + }, + { + "bbox": [ + 67, + 701, + 295, + 713 + ], + "type": "text", + "content": " takes the RGB im" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 313, + 288, + 553, + 300 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 288, + 553, + 300 + ], + "spans": [ + { + "bbox": [ + 313, + 288, + 553, + 300 + ], + "type": "text", + "content": "age and the sparse depth from simulated scanning as inputs:" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 380, + 307, + 553, + 319 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 380, + 307, + 553, + 319 + ], + "spans": [ + { + "bbox": [ + 380, + 307, + 553, + 319 + ], + "type": "interline_equation", + "content": "D _ {f} = f _ {\\theta} \\left(I _ {\\mathrm {u n}}, D _ {\\mathrm {s i m u}}\\right), \\tag {5}", + "image_path": "a3a90bda5eb9ff2018661ea7bbbdc0f83b45db8b089fe49d0ac0d4576a953bd8.jpg" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 313, + 327, + 553, + 362 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 327, + 553, + 362 + ], + "spans": [ + { + "bbox": [ + 313, + 327, + 553, + 362 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 313, + 327, + 553, + 362 + ], + "type": "inline_equation", + "content": "D_{\\mathrm{simu}}" + }, + { + "bbox": [ + 313, + 327, + 553, + 362 + ], + "type": "text", + "content": " is the sparse input from simulated LiDAR scanning. Afterwards, the depth completion model is pretrained using L1 Loss:" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 392, + 369, + 553, + 383 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 392, + 369, + 553, + 383 + ], + "spans": [ + { + "bbox": [ + 392, + 369, + 553, + 383 + ], + "type": "interline_equation", + "content": "\\mathcal {L} _ {\\text {p r e}} = \\left| D _ {f} - D _ {m} \\right|. \\tag {6}", + "image_path": "d50fd781ae9013480ddaf3c6429e430b4b82fcfbcd1683b8366a1aaa31af244a.jpg" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 313, + 389, + 554, + 486 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 389, + 554, + 486 + ], + "spans": [ + { + "bbox": [ + 313, + 389, + 554, + 486 + ], + "type": "text", + "content": "This pre-training process encourages the depth completion model " + }, + { + "bbox": [ + 313, + 389, + 554, + 486 + ], + "type": "inline_equation", + "content": "f_{\\theta}" + }, + { + "bbox": [ + 313, + 389, + 554, + 486 + ], + "type": "text", + "content": " to align its predictions with the synthesized depth maps, learning meaningful geometric features from the generated training data using monocular foundation models and diverse natural image data. Additionally, the model learns complex geometric structures and relative depth relationships across varying scenes, which are crucial for accurate depth completion in the fine-tuning phase." + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 313, + 491, + 462, + 504 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 491, + 462, + 504 + ], + "spans": [ + { + "bbox": [ + 313, + 491, + 462, + 504 + ], + "type": "text", + "content": "3.3. Second Stage: Fine-tuning" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 313, + 510, + 555, + 653 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 510, + 555, + 653 + ], + "spans": [ + { + "bbox": [ + 313, + 510, + 555, + 653 + ], + "type": "text", + "content": "The depth completion model pre-trained on the generated data learns powerful geometric features from the monocular foundation models. However, due to the inherent scale ambiguity of monocular depth, the model prediction is not consistent with the scale in the real world. To learn the real-world scale, we introduce a combined loss for fine-tuning on labeled datasets. To be specific, we first utilize the sparse ground truth " + }, + { + "bbox": [ + 313, + 510, + 555, + 653 + ], + "type": "inline_equation", + "content": "D_{s}" + }, + { + "bbox": [ + 313, + 510, + 555, + 653 + ], + "type": "text", + "content": " for supervised loss " + }, + { + "bbox": [ + 313, + 510, + 555, + 653 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{\\mathrm{sup}}" + }, + { + "bbox": [ + 313, + 510, + 555, + 653 + ], + "type": "text", + "content": " to focus on real-world depth, as in Equation 2. Then, the dense monocular depth " + }, + { + "bbox": [ + 313, + 510, + 555, + 653 + ], + "type": "inline_equation", + "content": "D_{m}" + }, + { + "bbox": [ + 313, + 510, + 555, + 653 + ], + "type": "text", + "content": " predicted by monocular depth models is also used to provide approximate depth values across the image, especially in regions not covered by the ground truth " + }, + { + "bbox": [ + 313, + 510, + 555, + 653 + ], + "type": "inline_equation", + "content": "D_{s}" + }, + { + "bbox": [ + 313, + 510, + 555, + 653 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 313, + 654, + 555, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 654, + 555, + 713 + ], + "spans": [ + { + "bbox": [ + 313, + 654, + 555, + 713 + ], + "type": "text", + "content": "However, the key challenge in using monocular depth estimation as supervision is the inherent scale ambiguity in depth predictions. To address this issue, we incorporate a Scale- and Shift-Invariant Loss (SSI Loss) into our distillation framework. The SSI Loss is designed to align the" + } + ] + } + ], + "index": 18 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "text", + "content": "22257" + } + ] + } + ], + "index": 19 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 3 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 58, + 91, + 291, + 177 + ], + "blocks": [ + { + "bbox": [ + 56, + 71, + 293, + 82 + ], + "lines": [ + { + "bbox": [ + 56, + 71, + 293, + 82 + ], + "spans": [ + { + "bbox": [ + 56, + 71, + 293, + 82 + ], + "type": "text", + "content": "Table 1. Unlabeled datasets used for pre-training data generation." + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 58, + 91, + 291, + 177 + ], + "lines": [ + { + "bbox": [ + 58, + 91, + 291, + 177 + ], + "spans": [ + { + "bbox": [ + 58, + 91, + 291, + 177 + ], + "type": "table", + "html": "
DatasetIndoorOutdoor# Images
COCO [33]✓✓118,287
Google Landmarks [59]✓117,576
Nuscenes [3]✓93,475
Cityscapes [12]✓19,998
DAVIS [43]✓✓10,581
", + "image_path": "cff35327f145640187eff97f823cb0feb9f87181f01f738c2ff5a57d15185c9e.jpg" + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "table_body" + } + ], + "index": 1 + }, + { + "bbox": [ + 55, + 196, + 295, + 232 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 196, + 295, + 232 + ], + "spans": [ + { + "bbox": [ + 55, + 196, + 295, + 232 + ], + "type": "text", + "content": "predicted depth map " + }, + { + "bbox": [ + 55, + 196, + 295, + 232 + ], + "type": "inline_equation", + "content": "D_{f}" + }, + { + "bbox": [ + 55, + 196, + 295, + 232 + ], + "type": "text", + "content": " with the dense monocular depth " + }, + { + "bbox": [ + 55, + 196, + 295, + 232 + ], + "type": "inline_equation", + "content": "D_{m}" + }, + { + "bbox": [ + 55, + 196, + 295, + 232 + ], + "type": "text", + "content": ". It remains invariant to differences in scale and shift between " + }, + { + "bbox": [ + 55, + 196, + 295, + 232 + ], + "type": "inline_equation", + "content": "D_{f}" + }, + { + "bbox": [ + 55, + 196, + 295, + 232 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 55, + 196, + 295, + 232 + ], + "type": "inline_equation", + "content": "D_{m}" + }, + { + "bbox": [ + 55, + 196, + 295, + 232 + ], + "type": "text", + "content": ". The SSI Loss is formulated as:" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 103, + 242, + 295, + 259 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 103, + 242, + 295, + 259 + ], + "spans": [ + { + "bbox": [ + 103, + 242, + 295, + 259 + ], + "type": "interline_equation", + "content": "\\mathcal {L} _ {\\text {S S I L}} = \\min _ {s, b} | D _ {f} - (s \\cdot D _ {m} + b) |, \\tag {7}", + "image_path": "51f1db4aec6c1099cf5d287c4e9fead424176fd3fbafce647ae8226b88eecc79.jpg" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 55, + 269, + 296, + 401 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 269, + 296, + 401 + ], + "spans": [ + { + "bbox": [ + 55, + 269, + 296, + 401 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 55, + 269, + 296, + 401 + ], + "type": "inline_equation", + "content": "s" + }, + { + "bbox": [ + 55, + 269, + 296, + 401 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 55, + 269, + 296, + 401 + ], + "type": "inline_equation", + "content": "b" + }, + { + "bbox": [ + 55, + 269, + 296, + 401 + ], + "type": "text", + "content": " are the optimal scale and shift parameters that align the predicted depth " + }, + { + "bbox": [ + 55, + 269, + 296, + 401 + ], + "type": "inline_equation", + "content": "D_{f}" + }, + { + "bbox": [ + 55, + 269, + 296, + 401 + ], + "type": "text", + "content": " with the dense monocular depth " + }, + { + "bbox": [ + 55, + 269, + 296, + 401 + ], + "type": "inline_equation", + "content": "D_{m}" + }, + { + "bbox": [ + 55, + 269, + 296, + 401 + ], + "type": "text", + "content": ". The loss function seeks to find the best alignment between " + }, + { + "bbox": [ + 55, + 269, + 296, + 401 + ], + "type": "inline_equation", + "content": "D_{f}" + }, + { + "bbox": [ + 55, + 269, + 296, + 401 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 55, + 269, + 296, + 401 + ], + "type": "inline_equation", + "content": "D_{m}" + }, + { + "bbox": [ + 55, + 269, + 296, + 401 + ], + "type": "text", + "content": ", effectively normalizing any global differences in scale and offset. By minimizing this loss across all pixels in the image, the model is encouraged to produce depth maps that maintain consistency with the relative depth structure provided by the monocular depth estimates. In addition, we adapt a gradient matching term to preserve sharpness and align with with depth discontinuities. The gradient matching term is defined as:" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 100, + 410, + 295, + 442 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 100, + 410, + 295, + 442 + ], + "spans": [ + { + "bbox": [ + 100, + 410, + 295, + 442 + ], + "type": "interline_equation", + "content": "\\mathcal {L} _ {\\text {r e g}} = \\frac {1}{N} \\sum_ {k = 1} ^ {K} \\left(| \\nabla_ {x} R ^ {k} | + | \\nabla_ {y} R ^ {k} |\\right), \\tag {8}", + "image_path": "0aed01367d9e706240fa71793078de818640f719aa48f6c1358d9b90d439c0d6.jpg" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 55, + 452, + 296, + 514 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 452, + 296, + 514 + ], + "spans": [ + { + "bbox": [ + 55, + 452, + 296, + 514 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 55, + 452, + 296, + 514 + ], + "type": "inline_equation", + "content": "R = D_{f} - D_{m}" + }, + { + "bbox": [ + 55, + 452, + 296, + 514 + ], + "type": "text", + "content": ", and " + }, + { + "bbox": [ + 55, + 452, + 296, + 514 + ], + "type": "inline_equation", + "content": "R^{k}" + }, + { + "bbox": [ + 55, + 452, + 296, + 514 + ], + "type": "text", + "content": " denotes the difference of depth maps at scale " + }, + { + "bbox": [ + 55, + 452, + 296, + 514 + ], + "type": "inline_equation", + "content": "k" + }, + { + "bbox": [ + 55, + 452, + 296, + 514 + ], + "type": "text", + "content": ". We use " + }, + { + "bbox": [ + 55, + 452, + 296, + 514 + ], + "type": "inline_equation", + "content": "K = 4" + }, + { + "bbox": [ + 55, + 452, + 296, + 514 + ], + "type": "text", + "content": " scale levels as in [46], halving the image resolution at each level. The final objective function combines the supervised loss " + }, + { + "bbox": [ + 55, + 452, + 296, + 514 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{\\mathrm{sup}}" + }, + { + "bbox": [ + 55, + 452, + 296, + 514 + ], + "type": "text", + "content": ", the dense distillation loss " + }, + { + "bbox": [ + 55, + 452, + 296, + 514 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{\\mathrm{SSIL}}" + }, + { + "bbox": [ + 55, + 452, + 296, + 514 + ], + "type": "text", + "content": ", and the regularization term " + }, + { + "bbox": [ + 55, + 452, + 296, + 514 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{\\mathrm{reg}}" + }, + { + "bbox": [ + 55, + 452, + 296, + 514 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 55, + 523, + 137, + 537 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 523, + 137, + 537 + ], + "spans": [ + { + "bbox": [ + 55, + 523, + 137, + 537 + ], + "type": "text", + "content": "4. Experiments" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 55, + 544, + 296, + 628 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 544, + 296, + 628 + ], + "spans": [ + { + "bbox": [ + 55, + 544, + 296, + 628 + ], + "type": "text", + "content": "In this section, we first introduce the details of our implementation, as well as the datasets and evaluation metrics for experiments. Then, detailed comparisons are conducted with the state-of-the-art methods. Finally, ablations and discussions are performed to confirm the effectiveness of our proposed main components. Additional analysis is provided in the supplementary materials, along with videos." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 55, + 635, + 183, + 647 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 635, + 183, + 647 + ], + "spans": [ + { + "bbox": [ + 55, + 635, + 183, + 647 + ], + "type": "text", + "content": "4.1. Experimental Pipeline" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 55, + 654, + 296, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 654, + 296, + 714 + ], + "spans": [ + { + "bbox": [ + 55, + 654, + 296, + 714 + ], + "type": "text", + "content": "First stage involves utilizing our proposed data generation strategy to train the depth completion model from scratch, as described in Section 3.2. In this stage, we generate training data using the RGB images in Table 1 and train the depth completion model using SSI Loss, which does not require" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 313, + 72, + 553, + 119 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 72, + 553, + 119 + ], + "spans": [ + { + "bbox": [ + 313, + 72, + 553, + 119 + ], + "type": "text", + "content": "any ground truth depth. This stage allows the pre-trained model to focus on learning diverse geometric features without being constrained by labeled data, enhancing generalization across different scenes." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 313, + 120, + 555, + 253 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 120, + 555, + 253 + ], + "spans": [ + { + "bbox": [ + 313, + 120, + 555, + 253 + ], + "type": "text", + "content": "Second stage involves adapting the pre-trained model on labeled datasets, using supervised loss (L1 loss) with depth annotations and our proposed SSI Loss for knowledge distillation, as described in Section 3.3. Using L1 loss with sparse ground truth allows the depth completion model to adapt to the real-world depth scale under sparse supervision. Specifically, our proposed monocular model distillation uses SSI Loss to help the model maintain fine-grained detail with dense supervision by distilling monocular foundation models. Fine-tuning and validation are performed on the KITTI and NYU Depth V2 datasets, respectively." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 313, + 261, + 447, + 273 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 261, + 447, + 273 + ], + "spans": [ + { + "bbox": [ + 313, + 261, + 447, + 273 + ], + "type": "text", + "content": "4.2. Implementation Details" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 313, + 278, + 555, + 519 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 278, + 555, + 519 + ], + "spans": [ + { + "bbox": [ + 313, + 278, + 555, + 519 + ], + "type": "text", + "content": "We follow the setup of recent work [50, 57, 79] and train our model on 4 NVIDIA RTX A100 GPUs. For the monocular foundation model used in distillation, we adopt Depth Anything V2 [72] due to its robust performance. For the depth completion network architecture, we use BP-Net [50] as our base model, which is primarily composed of ResNet [20] blocks. In the first stage, we utilize a mixed dataset collected from a wide range of resources, as shown in Table 1. We select five large-scale public datasets as our unlabeled sources for their diverse scenes, totaling approximately 360,000 images with varying scenes and image scales. We adopt AdamW with a weight decay of 0.05 as the optimizer and clip gradients if their norms exceeding 0.1. In the first stage, our model is pre-trained from scratch for 600K iterations. In the second stage, the pre-trained model is fine-tuned on the labeled datasets for 300K iterations using a OneCycle learning rate policy. We set the batch size to 16. The final model is obtained by applying Exponential Moving Average (EMA) with a decay of 0.9999, which helps stabilize the model parameters during training." + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 313, + 527, + 489, + 539 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 527, + 489, + 539 + ], + "spans": [ + { + "bbox": [ + 313, + 527, + 489, + 539 + ], + "type": "text", + "content": "4.3. Datasets and Evaluation Metrics" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 313, + 545, + 554, + 676 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 545, + 554, + 676 + ], + "spans": [ + { + "bbox": [ + 313, + 545, + 554, + 676 + ], + "type": "text", + "content": "KITTI Depth Completion Dataset [53] serves as a standard benchmark for depth completion in autonomous driving scenarios. It provides over 93,000 training samples, including sparse LiDAR depth maps, their corresponding RGB images, and ground truth depth maps. The sparse LiDAR depth maps are generated from Velodyne LiDAR scans, which typically cover only a small percentage of the image pixels (approximately " + }, + { + "bbox": [ + 313, + 545, + 554, + 676 + ], + "type": "inline_equation", + "content": "5\\%" + }, + { + "bbox": [ + 313, + 545, + 554, + 676 + ], + "type": "text", + "content": "). The ground truth depth maps are obtained via multi-frame matching and cover approximately " + }, + { + "bbox": [ + 313, + 545, + 554, + 676 + ], + "type": "inline_equation", + "content": "20\\%" + }, + { + "bbox": [ + 313, + 545, + 554, + 676 + ], + "type": "text", + "content": " image pixels. The dataset also includes 1,000 test samples on the private online benchmark." + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 313, + 677, + 554, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 677, + 554, + 713 + ], + "spans": [ + { + "bbox": [ + 313, + 677, + 554, + 713 + ], + "type": "text", + "content": "NYU Depth V2 Dataset [48] is also a standard benchmark for indoor scene understanding, consisting of RGB-D data captured with a Kinect sensor. It provides 1,449 densely" + } + ] + } + ], + "index": 17 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 749, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 749, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 749, + 317, + 757 + ], + "type": "text", + "content": "22258" + } + ] + } + ], + "index": 18 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 4 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 109, + 101, + 501, + 345 + ], + "blocks": [ + { + "bbox": [ + 55, + 70, + 555, + 94 + ], + "lines": [ + { + "bbox": [ + 55, + 70, + 555, + 94 + ], + "spans": [ + { + "bbox": [ + 55, + 70, + 555, + 94 + ], + "type": "text", + "content": "Table 2. Performance on KITTI and NYUv2 datasets. For the KITTI dataset, results are evaluated by the KITTI testing server and ranked by the RMSE (in mm). For the NYUv2 dataset, we report their performance on the official test set in their papers." + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 109, + 101, + 501, + 345 + ], + "lines": [ + { + "bbox": [ + 109, + 101, + 501, + 345 + ], + "spans": [ + { + "bbox": [ + 109, + 101, + 501, + 345 + ], + "type": "table", + "html": "
MethodYearKITTINYUv2
RMSE ↓MAE ↓iRMSE ↓iMAE ↓RMSE ↓REL ↓δ1.25 ↑
S2D [38]2018814.73249.952.801.210.2300.04497.1
CSPN [8]20191019.64279.462.931.150.1170.01699.2
DeepLiDAR [45]2019758.38226.502.561.150.1150.02299.3
CSPN++ [9]2020743.69209.282.070.900.1010.01599.5
GuideNet [49]2020736.24218.832.250.990.115--
FCFR [35]2021735.81217.152.200.980.1060.01599.5
ACMNet [80]2021744.91206.092.080.900.1050.01599.4
NLSPN [41]2020741.68199.591.990.840.0920.01299.6
PointDC [74]2023736.07201.871.970.870.0890.01299.6
RigNet [67]2022712.66203.252.080.900.0900.01399.6
DySPN [34]2022709.12192.711.880.820.0900.01299.6
BEV@DC [82]2023697.44189.441.830.820.0890.01299.6
CFormaler [79]2023708.87203.452.010.880.0900.012-
LRRU [56]2023696.51189.961.870.810.0910.01199.6
TPVD [70]2024693.97188.601.820.810.0860.01099.7
ImprovingDC [57]2024686.46187.951.830.810.0910.01199.6
BP-Net [50]2024684.90194.691.820.840.0890.01299.6
DMD3C (Ours)-678.12194.461.820.850.0850.01199.7
", + "image_path": "b687306afa5c1d6916d23f1efe6d83550b0d71bab6a6db4691489455bcd8a7b3.jpg" + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "table_body" + } + ], + "index": 1 + }, + { + "bbox": [ + 54, + 364, + 294, + 399 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 54, + 364, + 294, + 399 + ], + "spans": [ + { + "bbox": [ + 54, + 364, + 294, + 399 + ], + "type": "text", + "content": "labeled images from various indoor scenes. Missing ground truth is mainly due to the difference in perspective between the depth sensor and the camera." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 54, + 400, + 295, + 508 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 54, + 400, + 295, + 508 + ], + "spans": [ + { + "bbox": [ + 54, + 400, + 295, + 508 + ], + "type": "text", + "content": "Evaluation Metrics used for KITTI are root mean squared error (RMSE), mean absolute error (MAE), root mean squared error of the inverse depth (iRMSE), and mean absolute error of the inverse depth (iMAE), among which RMSE is used for ranking. These metrics are obtained by submission and evaluation from a non-public test set. For NYU, we follow the previous work and report the RMSE, mean absolute relative error (REL), and " + }, + { + "bbox": [ + 54, + 400, + 295, + 508 + ], + "type": "inline_equation", + "content": "\\delta_{\\theta}" + }, + { + "bbox": [ + 54, + 400, + 295, + 508 + ], + "type": "text", + "content": ", which represents the percentage of pixels whose error is less than a threshold " + }, + { + "bbox": [ + 54, + 400, + 295, + 508 + ], + "type": "inline_equation", + "content": "\\theta" + }, + { + "bbox": [ + 54, + 400, + 295, + 508 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 55, + 515, + 279, + 528 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 515, + 279, + 528 + ], + "spans": [ + { + "bbox": [ + 55, + 515, + 279, + 528 + ], + "type": "text", + "content": "4.4. Comparison with State-of-the-art Methods" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 55, + 533, + 295, + 677 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 533, + 295, + 677 + ], + "spans": [ + { + "bbox": [ + 55, + 533, + 295, + 677 + ], + "type": "text", + "content": "Evaluation on Depth Completion Benchmarks. We evaluate our method on the test sets of the NYUv2 dataset [48] and KITTI dataset [53]. Table 2 shows the quantitative comparison of our method and other top ranking published methods. On the KITTI leaderboard, our proposed " + }, + { + "bbox": [ + 55, + 533, + 295, + 677 + ], + "type": "inline_equation", + "content": "\\mathrm{DMD}^{3}\\mathrm{C}" + }, + { + "bbox": [ + 55, + 533, + 295, + 677 + ], + "type": "text", + "content": " ranks 1st outperforming all other methods under the primary RMSE metric at the time of paper submission. It also has comparable performance under other evaluation metrics. On the NYUv2 dataset, our method achieves the best RMSE and best " + }, + { + "bbox": [ + 55, + 533, + 295, + 677 + ], + "type": "inline_equation", + "content": "\\delta_{1.25}" + }, + { + "bbox": [ + 55, + 533, + 295, + 677 + ], + "type": "text", + "content": ". Additionally, " + }, + { + "bbox": [ + 55, + 533, + 295, + 677 + ], + "type": "inline_equation", + "content": "\\mathrm{DMD}^{3}\\mathrm{C}" + }, + { + "bbox": [ + 55, + 533, + 295, + 677 + ], + "type": "text", + "content": " maintains competitive results across several other evaluation metrics, further confirming its robustness and generalization." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 55, + 677, + 296, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 677, + 296, + 713 + ], + "spans": [ + { + "bbox": [ + 55, + 677, + 296, + 713 + ], + "type": "text", + "content": "To enhance understanding of " + }, + { + "bbox": [ + 55, + 677, + 296, + 713 + ], + "type": "inline_equation", + "content": "\\mathrm{DMD}^3\\mathrm{C}" + }, + { + "bbox": [ + 55, + 677, + 296, + 713 + ], + "type": "text", + "content": " 's superiority over other state-of-the-art methods, we provide visual comparisons. Figure 4 provides qualitative comparisons with other" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 313, + 364, + 555, + 507 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 364, + 555, + 507 + ], + "spans": [ + { + "bbox": [ + 313, + 364, + 555, + 507 + ], + "type": "text", + "content": "SOTA methods on the KITTI test set using publicly available results. We use results from the official benchmark1 produced by the best models for method comparison. Our " + }, + { + "bbox": [ + 313, + 364, + 555, + 507 + ], + "type": "inline_equation", + "content": "\\mathrm{DMD}^{3}\\mathrm{C}" + }, + { + "bbox": [ + 313, + 364, + 555, + 507 + ], + "type": "text", + "content": " model excels in maintaining sharp object boundaries and capturing fine details in areas where other models encounter difficulties, particularly in scenes with complex structures or objects at varying distances. In such challenging regions, other methods fail to estimate accurate depth. Furthermore, as shown in the error maps, our method demonstrates a clear advantage even in regions without ground truth, despite these areas not being factored into the evaluation metrics." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 313, + 508, + 556, + 605 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 508, + 556, + 605 + ], + "spans": [ + { + "bbox": [ + 313, + 508, + 556, + 605 + ], + "type": "text", + "content": "Not only does our model retain the fine details in the depth map, but it also maintains structural integrity in the potential three-dimensional space. Figure 5 illustrates that our proposed two-stage distillation improves performance in the 3D space by visualizing point clouds from the completed depth maps. In comparison to other methods, our " + }, + { + "bbox": [ + 313, + 508, + 556, + 605 + ], + "type": "inline_equation", + "content": "\\mathrm{DMD}^{3}\\mathrm{C}" + }, + { + "bbox": [ + 313, + 508, + 556, + 605 + ], + "type": "text", + "content": " produces a more coherent and complete 3D structure, further outperforming existing methods." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 313, + 613, + 391, + 625 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 613, + 391, + 625 + ], + "spans": [ + { + "bbox": [ + 313, + 613, + 391, + 625 + ], + "type": "text", + "content": "4.5. Discussions" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 313, + 632, + 555, + 694 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 632, + 555, + 694 + ], + "spans": [ + { + "bbox": [ + 313, + 632, + 555, + 694 + ], + "type": "text", + "content": "Effect of Pre-training in the First Stage. First, to evaluate the impact of the proposed pre-training strategy using unlabeled images, we conduct an experiment where the depth completion model is trained from scratch, without any pretraining. As shown in Table 3, removing the pre-training" + } + ] + } + ], + "index": 10 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 325, + 702, + 514, + 712 + ], + "type": "page_footnote", + "angle": 0, + "lines": [ + { + "bbox": [ + 325, + 702, + 514, + 712 + ], + "spans": [ + { + "bbox": [ + 325, + 702, + 514, + 712 + ], + "type": "text", + "content": "MethodKITTIRMSE ↓MAE ↓iRMSE ↓iMAE ↓w/o Pre-train682.34194.961.820.84w/o SSI Loss684.54195.651.860.85DMD3C678.12194.461.820.85", + "image_path": "190d1c649d66a57fe0ee98bb38ca72111a053827b65b039a420fdaf7690a4dd0.jpg" + } + ] + } + ], + "index": 18, + "angle": 0, + "type": "table_body" + } + ], + "index": 18 + }, + { + "type": "table", + "bbox": [ + 59, + 430, + 292, + 547 + ], + "blocks": [ + { + "bbox": [ + 77, + 411, + 272, + 420 + ], + "lines": [ + { + "bbox": [ + 77, + 411, + 272, + 420 + ], + "spans": [ + { + "bbox": [ + 77, + 411, + 272, + 420 + ], + "type": "text", + "content": "Table 4. Ablations on different network architectures." + } + ] + } + ], + "index": 19, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 59, + 430, + 292, + 547 + ], + "lines": [ + { + "bbox": [ + 59, + 430, + 292, + 547 + ], + "spans": [ + { + "bbox": [ + 59, + 430, + 292, + 547 + ], + "type": "table", + "html": "
MethodKITTI
RMSE ↓MAE ↓iRMSE ↓iMAE ↓
LRRU [56]696.51189.961.870.81
+ Ours693.17189.601.850.80
CFormer [79]764.87183.881.890.80
+ Ours760.29183.621.880.80
BP-Net [50]684.90194.691.820.84
+ Ours678.12194.461.820.85
", + "image_path": "bcf09035aa9cfb4c134256f3f4de35ebebadae225a6235ea7d45a9d490dc8489.jpg" + } + ] + } + ], + "index": 20, + "angle": 0, + "type": "table_body" + } + ], + "index": 20 + }, + { + "bbox": [ + 55, + 569, + 295, + 628 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 569, + 295, + 628 + ], + "spans": [ + { + "bbox": [ + 55, + 569, + 295, + 628 + ], + "type": "text", + "content": "step results in a performance degradation, with an increase in RMSE by " + }, + { + "bbox": [ + 55, + 569, + 295, + 628 + ], + "type": "inline_equation", + "content": "4.22\\mathrm{mm}" + }, + { + "bbox": [ + 55, + 569, + 295, + 628 + ], + "type": "text", + "content": ". This demonstrates that pre-training plays a crucial role in enhancing the model's ability to generalize, enabling it to better capture complex geometric features across diverse scenes." + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 55, + 629, + 296, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 629, + 296, + 714 + ], + "spans": [ + { + "bbox": [ + 55, + 629, + 296, + 714 + ], + "type": "text", + "content": "Effect of SSI Loss in the Second Stage. Next, we show the impact of removing SSI Loss in the second distillation stage, while retaining the standard supervised loss with sparse ground truth. As shown in Table 3, the results indicate that the proposed SSI Loss improves the model's performance. This demonstrates the benefit of aligning monocular depth supervision with real-world scale using SSI Loss." + } + ] + } + ], + "index": 22 + }, + { + "type": "table", + "bbox": [ + 315, + 335, + 554, + 388 + ], + "blocks": [ + { + "bbox": [ + 313, + 304, + 555, + 326 + ], + "lines": [ + { + "bbox": [ + 313, + 304, + 555, + 326 + ], + "spans": [ + { + "bbox": [ + 313, + 304, + 555, + 326 + ], + "type": "text", + "content": "Table 5. Zero-shot performance comparison with other methods on out-of-the-domain datasets." + } + ] + } + ], + "index": 23, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 315, + 335, + 554, + 388 + ], + "lines": [ + { + "bbox": [ + 315, + 335, + 554, + 388 + ], + "spans": [ + { + "bbox": [ + 315, + 335, + 554, + 388 + ], + "type": "table", + "html": "
MethodScanNetDDADVOID1500
RMSE ↓MAE ↓RMSE ↓MAE ↓RMSE ↓MAE ↓
CFormal0.1200.2329.6063.3280.7260.261
LRRU0.1320.2459.1642.7380.6980.232
BP-Net0.1220.2128.9032.7120.7040.230
Ours0.1010.2107.7662.4980.6760.225
", + "image_path": "7b9076672cae15858f4ffc66b1cdc77a88dcc3bec7dbeddd65ac8f3395608702.jpg" + } + ] + } + ], + "index": 24, + "angle": 0, + "type": "table_body" + } + ], + "index": 24 + }, + { + "bbox": [ + 313, + 411, + 555, + 555 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 411, + 555, + 555 + ], + "spans": [ + { + "bbox": [ + 313, + 411, + 555, + 555 + ], + "type": "text", + "content": "Network Architectures. We evaluate the compatibility of our method with different network architectures, as shown in Table 4. Since we focus on training strategies, we have the flexibility to various models, demonstrating the robustness of our approach across different network designs. Specifically, we evaluate BP-Net [50], LRRU [56], and CFormer (L1 Loss) [79]. BP-Net is one of the most advanced methods, while LRRU and CFormer are considered representative architectures. Our approach consistently improves performance across all three models. Notably, BP-Net with our method achieves the best performance in terms of RMSE, the primary evaluation metric." + } + ] + } + ], + "index": 25 + }, + { + "bbox": [ + 313, + 556, + 555, + 676 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 556, + 555, + 676 + ], + "spans": [ + { + "bbox": [ + 313, + 556, + 555, + 676 + ], + "type": "text", + "content": "Generalization to Out-of-Domain Datasets. To evaluate the generalization ability of our method, we conduct zero-shot testing on unseen datasets, including ScanNet [13], DDAD [17], and Void1500 [60], as shown in Table 5. Our approach consistently achieves the best performance across all datasets, demonstrating lower RMSE and MAE compared to prior methods. Notably, our method outperforms BP-Net by a significant margin on DDAD, reducing RMSE from 8.903 to 7.766, highlighting its effectiveness in handling unseen scenarios." + } + ] + } + ], + "index": 26 + }, + { + "bbox": [ + 313, + 677, + 555, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 677, + 555, + 713 + ], + "spans": [ + { + "bbox": [ + 313, + 677, + 555, + 713 + ], + "type": "text", + "content": "Application on Dense SLAM. Figure 6 compares a representative SLAM method, Droid-SLAM [51], with sparse LiDAR points (left) and our method with dense depth com" + } + ] + } + ], + "index": 27 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "text", + "content": "22260" + } + ] + } + ], + "index": 28 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 6 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 59, + 69, + 140, + 95 + ], + "blocks": [ + { + "bbox": [ + 59, + 69, + 140, + 95 + ], + "lines": [ + { + "bbox": [ + 59, + 69, + 140, + 95 + ], + "spans": [ + { + "bbox": [ + 59, + 69, + 140, + 95 + ], + "type": "image", + "image_path": "d92af7072b332bf2f41f32096c72c303c026786d5b6f68f8d51edc567ddfc169.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + } + ], + "index": 0 + }, + { + "type": "image", + "bbox": [ + 69, + 114, + 139, + 162 + ], + "blocks": [ + { + "bbox": [ + 69, + 114, + 139, + 162 + ], + "lines": [ + { + "bbox": [ + 69, + 114, + 139, + 162 + ], + "spans": [ + { + "bbox": [ + 69, + 114, + 139, + 162 + ], + "type": "image", + "image_path": "0ea130686549fc0fef5074e0185653c52a55763cd4ab35aa02f2b7a176c2ba83.jpg" + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_body" + } + ], + "index": 1 + }, + { + "type": "image", + "bbox": [ + 141, + 71, + 222, + 95 + ], + "blocks": [ + { + "bbox": [ + 141, + 71, + 222, + 95 + ], + "lines": [ + { + "bbox": [ + 141, + 71, + 222, + 95 + ], + "spans": [ + { + "bbox": [ + 141, + 71, + 222, + 95 + ], + "type": "image", + "image_path": "a0224d2802d609c9ad64290f5d6cfe7d40eb72a48866adbc93ee88a3ee7b9988.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + } + ], + "index": 2 + }, + { + "type": "image", + "bbox": [ + 143, + 95, + 218, + 161 + ], + "blocks": [ + { + "bbox": [ + 143, + 95, + 218, + 161 + ], + "lines": [ + { + "bbox": [ + 143, + 95, + 218, + 161 + ], + "spans": [ + { + "bbox": [ + 143, + 95, + 218, + 161 + ], + "type": "image", + "image_path": "9c7a26b3be63938907cfbcd1fb7d582443ac4c6c999eb3e3529b115a3b01c42e.jpg" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_body" + } + ], + "index": 3 + }, + { + "type": "image", + "bbox": [ + 223, + 70, + 304, + 95 + ], + "blocks": [ + { + "bbox": [ + 223, + 70, + 304, + 95 + ], + "lines": [ + { + "bbox": [ + 223, + 70, + 304, + 95 + ], + "spans": [ + { + "bbox": [ + 223, + 70, + 304, + 95 + ], + "type": "image", + "image_path": "33f3bfd10d5b1f8fa2ee922707ab8852aa4ed75e7436733ee8dfb37e47ff8152.jpg" + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_body" + } + ], + "index": 4 + }, + { + "type": "image", + "bbox": [ + 225, + 95, + 304, + 162 + ], + "blocks": [ + { + "bbox": [ + 225, + 95, + 304, + 162 + ], + "lines": [ + { + "bbox": [ + 225, + 95, + 304, + 162 + ], + "spans": [ + { + "bbox": [ + 225, + 95, + 304, + 162 + ], + "type": "image", + "image_path": "5a784980738bde6ae8ebc381e0fcb3ecf2cce3f1f20fa47da9fadc6ea884936d.jpg" + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "image_body" + } + ], + "index": 5 + }, + { + "type": "image", + "bbox": [ + 306, + 71, + 386, + 95 + ], + "blocks": [ + { + "bbox": [ + 306, + 71, + 386, + 95 + ], + "lines": [ + { + "bbox": [ + 306, + 71, + 386, + 95 + ], + "spans": [ + { + "bbox": [ + 306, + 71, + 386, + 95 + ], + "type": "image", + "image_path": "95986492181b4b2136716019c32d031945fa795c696a7225321c772f49b08cfb.jpg" + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "image_body" + } + ], + "index": 6 + }, + { + "type": "image", + "bbox": [ + 307, + 95, + 385, + 163 + ], + "blocks": [ + { + "bbox": [ + 307, + 95, + 385, + 163 + ], + "lines": [ + { + "bbox": [ + 307, + 95, + 385, + 163 + ], + "spans": [ + { + "bbox": [ + 307, + 95, + 385, + 163 + ], + "type": "image", + "image_path": "21c687819faf9909dd669d1f6fc5cc002f714832abe6fbad556cf1cc012e499d.jpg" + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "image_body" + } + ], + "index": 7 + }, + { + "type": "image", + "bbox": [ + 388, + 71, + 467, + 95 + ], + "blocks": [ + { + "bbox": [ + 388, + 71, + 467, + 95 + ], + "lines": [ + { + "bbox": [ + 388, + 71, + 467, + 95 + ], + "spans": [ + { + "bbox": [ + 388, + 71, + 467, + 95 + ], + "type": "image", + "image_path": "899409fbac316982c51d0e9460936d2e93bf0d9676b5cea745fdc121fcb4093c.jpg" + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "image_body" + } + ], + "index": 8 + }, + { + "type": "image", + "bbox": [ + 389, + 101, + 463, + 163 + ], + "blocks": [ + { + "bbox": [ + 389, + 101, + 463, + 163 + ], + "lines": [ + { + "bbox": [ + 389, + 101, + 463, + 163 + ], + "spans": [ + { + "bbox": [ + 389, + 101, + 463, + 163 + ], + "type": "image", + "image_path": "34e24f874c8be7a793a77a9bc67d3438696c2ceac104a8b6e47ff0be7f2a0c5b.jpg" + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "image_body" + } + ], + "index": 9 + }, + { + "type": "image", + "bbox": [ + 470, + 71, + 551, + 95 + ], + "blocks": [ + { + "bbox": [ + 470, + 71, + 551, + 95 + ], + "lines": [ + { + "bbox": [ + 470, + 71, + 551, + 95 + ], + "spans": [ + { + "bbox": [ + 470, + 71, + 551, + 95 + ], + "type": "image", + "image_path": "3351ae1579c1b060aadc6fce947e3cc233267af41a0aa1571191acde6ad21637.jpg" + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "image_body" + } + ], + "index": 10 + }, + { + "type": "image", + "bbox": [ + 471, + 101, + 544, + 163 + ], + "blocks": [ + { + "bbox": [ + 471, + 101, + 544, + 163 + ], + "lines": [ + { + "bbox": [ + 471, + 101, + 544, + 163 + ], + "spans": [ + { + "bbox": [ + 471, + 101, + 544, + 163 + ], + "type": "image", + "image_path": "660f3820d343b76e3657dad90e0cf4234e5791e78ad35c88d47fcbf3e11937f5.jpg" + } + ] + } + ], + "index": 11, + "angle": 0, + "type": "image_body" + } + ], + "index": 11 + }, + { + "type": "image", + "bbox": [ + 59, + 170, + 140, + 194 + ], + "blocks": [ + { + "bbox": [ + 59, + 170, + 140, + 194 + ], + "lines": [ + { + "bbox": [ + 59, + 170, + 140, + 194 + ], + "spans": [ + { + "bbox": [ + 59, + 170, + 140, + 194 + ], + "type": "image", + "image_path": "b62251f35aadc5fe7e3ebf13fe2a66c650052727c1df9c14a0cc716295bc7112.jpg" + } + ] + } + ], + "index": 12, + "angle": 0, + "type": "image_body" + } + ], + "index": 12 + }, + { + "type": "image", + "bbox": [ + 62, + 196, + 139, + 278 + ], + "blocks": [ + { + "bbox": [ + 62, + 196, + 139, + 278 + ], + "lines": [ + { + "bbox": [ + 62, + 196, + 139, + 278 + ], + "spans": [ + { + "bbox": [ + 62, + 196, + 139, + 278 + ], + "type": "image", + "image_path": "4aedd7be838c006ec45932487550d848e596a8d1a53c170d9f9d04ccb5661cef.jpg" + } + ] + } + ], + "index": 13, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 85, + 280, + 115, + 289 + ], + "lines": [ + { + "bbox": [ + 85, + 280, + 115, + 289 + ], + "spans": [ + { + "bbox": [ + 85, + 280, + 115, + 289 + ], + "type": "text", + "content": "(a) Input" + } + ] + } + ], + "index": 14, + "angle": 0, + "type": "image_caption" + } + ], + "index": 13 + }, + { + "type": "image", + "bbox": [ + 142, + 171, + 222, + 194 + ], + "blocks": [ + { + "bbox": [ + 142, + 171, + 222, + 194 + ], + "lines": [ + { + "bbox": [ + 142, + 171, + 222, + 194 + ], + "spans": [ + { + "bbox": [ + 142, + 171, + 222, + 194 + ], + "type": "image", + "image_path": "02e4c30d5be755f7a545622c2ae6f6cae7c529b6de8fcd8600733f95042b846c.jpg" + } + ] + } + ], + "index": 15, + "angle": 0, + "type": "image_body" + } + ], + "index": 15 + }, + { + "type": "image", + "bbox": [ + 142, + 194, + 222, + 279 + ], + "blocks": [ + { + "bbox": [ + 142, + 194, + 222, + 279 + ], + "lines": [ + { + "bbox": [ + 142, + 194, + 222, + 279 + ], + "spans": [ + { + "bbox": [ + 142, + 194, + 222, + 279 + ], + "type": "image", + "image_path": "066040b0d28bacecd4e6720db1fed553b71afed5c4b327b28d3ca4dd8ce756c8.jpg" + } + ] + } + ], + "index": 16, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 154, + 280, + 211, + 289 + ], + "lines": [ + { + "bbox": [ + 154, + 280, + 211, + 289 + ], + "spans": [ + { + "bbox": [ + 154, + 280, + 211, + 289 + ], + "type": "text", + "content": "(b) CFormer [79]" + } + ] + } + ], + "index": 17, + "angle": 0, + "type": "image_caption" + } + ], + "index": 16 + }, + { + "type": "image", + "bbox": [ + 224, + 171, + 304, + 194 + ], + "blocks": [ + { + "bbox": [ + 224, + 171, + 304, + 194 + ], + "lines": [ + { + "bbox": [ + 224, + 171, + 304, + 194 + ], + "spans": [ + { + "bbox": [ + 224, + 171, + 304, + 194 + ], + "type": "image", + "image_path": "fc756172ad3ce0950f929f2461cd3f4196cc65297420767574e521334ed1b131.jpg" + } + ] + } + ], + "index": 18, + "angle": 0, + "type": "image_body" + } + ], + "index": 18 + }, + { + "type": "image", + "bbox": [ + 225, + 194, + 304, + 279 + ], + "blocks": [ + { + "bbox": [ + 225, + 194, + 304, + 279 + ], + "lines": [ + { + "bbox": [ + 225, + 194, + 304, + 279 + ], + "spans": [ + { + "bbox": [ + 225, + 194, + 304, + 279 + ], + "type": "image", + "image_path": "8faa4db24d3df57d5b6d3ed227fec2e931a664d5300e1c2e845b1ea0de0073f4.jpg" + } + ] + } + ], + "index": 19, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 239, + 280, + 288, + 289 + ], + "lines": [ + { + "bbox": [ + 239, + 280, + 288, + 289 + ], + "spans": [ + { + "bbox": [ + 239, + 280, + 288, + 289 + ], + "type": "text", + "content": "(c) LRRU [56]" + } + ] + } + ], + "index": 20, + "angle": 0, + "type": "image_caption" + } + ], + "index": 19 + }, + { + "type": "image", + "bbox": [ + 306, + 171, + 386, + 194 + ], + "blocks": [ + { + "bbox": [ + 306, + 171, + 386, + 194 + ], + "lines": [ + { + "bbox": [ + 306, + 171, + 386, + 194 + ], + "spans": [ + { + "bbox": [ + 306, + 171, + 386, + 194 + ], + "type": "image", + "image_path": "73169d4e174de121892b9ffa88a0ca2d8dc52c152631dc2a8d2f8db532a75f8b.jpg" + } + ] + } + ], + "index": 21, + "angle": 0, + "type": "image_body" + } + ], + "index": 21 + }, + { + "type": "image", + "bbox": [ + 308, + 194, + 384, + 278 + ], + "blocks": [ + { + "bbox": [ + 308, + 194, + 384, + 278 + ], + "lines": [ + { + "bbox": [ + 308, + 194, + 384, + 278 + ], + "spans": [ + { + "bbox": [ + 308, + 194, + 384, + 278 + ], + "type": "image", + "image_path": "f6ec725f2e27c956c91abb95ed2343ae6b943c4fbe83201b5cd0d2763794bf45.jpg" + } + ] + } + ], + "index": 22, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 310, + 280, + 383, + 289 + ], + "lines": [ + { + "bbox": [ + 310, + 280, + 383, + 289 + ], + "spans": [ + { + "bbox": [ + 310, + 280, + 383, + 289 + ], + "type": "text", + "content": "(d) ImprovingDC [57]" + } + ] + } + ], + "index": 23, + "angle": 0, + "type": "image_caption" + } + ], + "index": 22 + }, + { + "type": "image", + "bbox": [ + 388, + 171, + 468, + 194 + ], + "blocks": [ + { + "bbox": [ + 388, + 171, + 468, + 194 + ], + "lines": [ + { + "bbox": [ + 388, + 171, + 468, + 194 + ], + "spans": [ + { + "bbox": [ + 388, + 171, + 468, + 194 + ], + "type": "image", + "image_path": "7712ba798397df2149c95b19198f648d12bd6d8ec983b5693ee3a9bcaa0888d8.jpg" + } + ] + } + ], + "index": 24, + "angle": 0, + "type": "image_body" + } + ], + "index": 24 + }, + { + "type": "image", + "bbox": [ + 391, + 194, + 466, + 275 + ], + "blocks": [ + { + "bbox": [ + 391, + 194, + 466, + 275 + ], + "lines": [ + { + "bbox": [ + 391, + 194, + 466, + 275 + ], + "spans": [ + { + "bbox": [ + 391, + 194, + 466, + 275 + ], + "type": "image", + "image_path": "b8b5b33b17b7f8af3bcad63dbb1d911fa98149ab01488913d1f605d370621fdb.jpg" + } + ] + } + ], + "index": 25, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 402, + 280, + 454, + 289 + ], + "lines": [ + { + "bbox": [ + 402, + 280, + 454, + 289 + ], + "spans": [ + { + "bbox": [ + 402, + 280, + 454, + 289 + ], + "type": "text", + "content": "(e) BP-Net [50]" + } + ] + } + ], + "index": 26, + "angle": 0, + "type": "image_caption" + } + ], + "index": 25 + }, + { + "type": "image", + "bbox": [ + 470, + 171, + 551, + 194 + ], + "blocks": [ + { + "bbox": [ + 470, + 171, + 551, + 194 + ], + "lines": [ + { + "bbox": [ + 470, + 171, + 551, + 194 + ], + "spans": [ + { + "bbox": [ + 470, + 171, + 551, + 194 + ], + "type": "image", + "image_path": "07fa64ef0fb71eac5aab3cbf56fc8ccb816ed6e23da09e3d1e8009cc6e622744.jpg" + } + ] + } + ], + "index": 27, + "angle": 0, + "type": "image_body" + } + ], + "index": 27 + }, + { + "type": "image", + "bbox": [ + 474, + 194, + 546, + 277 + ], + "blocks": [ + { + "bbox": [ + 474, + 194, + 546, + 277 + ], + "lines": [ + { + "bbox": [ + 474, + 194, + 546, + 277 + ], + "spans": [ + { + "bbox": [ + 474, + 194, + 546, + 277 + ], + "type": "image", + "image_path": "9b8ec9b8a9c8224b721c7f100f945cabfd771176c22bec2cbcf522088057c3ed.jpg" + } + ] + } + ], + "index": 28, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 479, + 280, + 541, + 289 + ], + "lines": [ + { + "bbox": [ + 479, + 280, + 541, + 289 + ], + "spans": [ + { + "bbox": [ + 479, + 280, + 541, + 289 + ], + "type": "text", + "content": "(f) DMD" + }, + { + "bbox": [ + 479, + 280, + 541, + 289 + ], + "type": "inline_equation", + "content": "^3" + }, + { + "bbox": [ + 479, + 280, + 541, + 289 + ], + "type": "text", + "content": "C (Ours)" + } + ] + } + ], + "index": 29, + "angle": 0, + "type": "image_caption" + } + ], + "index": 28 + }, + { + "type": "image", + "bbox": [ + 72, + 352, + 130, + 433 + ], + "blocks": [ + { + "bbox": [ + 55, + 300, + 555, + 333 + ], + "lines": [ + { + "bbox": [ + 55, + 300, + 555, + 333 + ], + "spans": [ + { + "bbox": [ + 55, + 300, + 555, + 333 + ], + "type": "text", + "content": "Figure 5. Qualitative comparison of depth completion methods. This figure demonstrates the performance of various depth completion models, including CFormer, LRRU, ImprovingDC, BP-Net, and our proposed DMD" + }, + { + "bbox": [ + 55, + 300, + 555, + 333 + ], + "type": "inline_equation", + "content": "^3" + }, + { + "bbox": [ + 55, + 300, + 555, + 333 + ], + "type": "text", + "content": "C. For each method, we show the input RGB images with sparse LiDAR points (left), along with the resulting completed depth maps and corresponding 3D point cloud reconstructions." + } + ] + } + ], + "index": 30, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 72, + 352, + 130, + 433 + ], + "lines": [ + { + "bbox": [ + 72, + 352, + 130, + 433 + ], + "spans": [ + { + "bbox": [ + 72, + 352, + 130, + 433 + ], + "type": "image", + "image_path": "059646f4a2f0f773c06757f855ac0d0acabe21d6d0b1a403cbcb5311c0790c1d.jpg" + } + ] + } + ], + "index": 31, + "angle": 0, + "type": "image_body" + } + ], + "index": 31 + }, + { + "type": "image", + "bbox": [ + 132, + 352, + 172, + 433 + ], + "blocks": [ + { + "bbox": [ + 132, + 352, + 172, + 433 + ], + "lines": [ + { + "bbox": [ + 132, + 352, + 172, + 433 + ], + "spans": [ + { + "bbox": [ + 132, + 352, + 172, + 433 + ], + "type": "image", + "image_path": "884076331dc16b92c35e42bb4f59865ae59245b52b8a0b8332d951ee604918fe.jpg" + } + ] + } + ], + "index": 32, + "angle": 0, + "type": "image_body" + } + ], + "index": 32 + }, + { + "type": "image", + "bbox": [ + 181, + 353, + 242, + 432 + ], + "blocks": [ + { + "bbox": [ + 181, + 353, + 242, + 432 + ], + "lines": [ + { + "bbox": [ + 181, + 353, + 242, + 432 + ], + "spans": [ + { + "bbox": [ + 181, + 353, + 242, + 432 + ], + "type": "image", + "image_path": "4f6e332ef06753404b1638d5d26dfa497c774eed8874ad61a48a5d459f0e8126.jpg" + } + ] + } + ], + "index": 33, + "angle": 0, + "type": "image_body" + } + ], + "index": 33 + }, + { + "type": "image", + "bbox": [ + 243, + 353, + 282, + 432 + ], + "blocks": [ + { + "bbox": [ + 243, + 353, + 282, + 432 + ], + "lines": [ + { + "bbox": [ + 243, + 353, + 282, + 432 + ], + "spans": [ + { + "bbox": [ + 243, + 353, + 282, + 432 + ], + "type": "image", + "image_path": "9c76f7f9112fb0f990455823736b89bafb1619938bd6f035fbe13592af279b67.jpg" + } + ] + } + ], + "index": 34, + "angle": 0, + "type": "image_body" + } + ], + "index": 34 + }, + { + "type": "image", + "bbox": [ + 73, + 434, + 129, + 536 + ], + "blocks": [ + { + "bbox": [ + 73, + 434, + 129, + 536 + ], + "lines": [ + { + "bbox": [ + 73, + 434, + 129, + 536 + ], + "spans": [ + { + "bbox": [ + 73, + 434, + 129, + 536 + ], + "type": "image", + "image_path": "59ab0a551e3990bbdf80dec89dd443ac6bcefba92229432d03362cd3668371b3.jpg" + } + ] + } + ], + "index": 35, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 87, + 536, + 148, + 543 + ], + "lines": [ + { + "bbox": [ + 87, + 536, + 148, + 543 + ], + "spans": [ + { + "bbox": [ + 87, + 536, + 148, + 543 + ], + "type": "text", + "content": "Droid-SLAM + LiDAR" + } + ] + } + ], + "index": 36, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 55, + 563, + 295, + 586 + ], + "lines": [ + { + "bbox": [ + 55, + 563, + 295, + 586 + ], + "spans": [ + { + "bbox": [ + 55, + 563, + 295, + 586 + ], + "type": "text", + "content": "Figure 6. Qualitative comparison of Droid-SLAM with sparse LiDAR points and our method with dense depth completion." + } + ] + } + ], + "index": 41, + "angle": 0, + "type": "image_caption" + } + ], + "index": 35 + }, + { + "type": "image", + "bbox": [ + 129, + 434, + 173, + 536 + ], + "blocks": [ + { + "bbox": [ + 129, + 434, + 173, + 536 + ], + "lines": [ + { + "bbox": [ + 129, + 434, + 173, + 536 + ], + "spans": [ + { + "bbox": [ + 129, + 434, + 173, + 536 + ], + "type": "image", + "image_path": "3a510b5c300e4212d14ca70c8b1bf863b9d386c7fa31f3e000792bdba68d6003.jpg" + } + ] + } + ], + "index": 37, + "angle": 0, + "type": "image_body" + } + ], + "index": 37 + }, + { + "type": "image", + "bbox": [ + 181, + 434, + 237, + 536 + ], + "blocks": [ + { + "bbox": [ + 181, + 434, + 237, + 536 + ], + "lines": [ + { + "bbox": [ + 181, + 434, + 237, + 536 + ], + "spans": [ + { + "bbox": [ + 181, + 434, + 237, + 536 + ], + "type": "image", + "image_path": "58ed8c4656c66163632517d07e030b95843f0227dcfe3f28ce470a8a4abb1931.jpg" + } + ] + } + ], + "index": 38, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 199, + 536, + 254, + 543 + ], + "lines": [ + { + "bbox": [ + 199, + 536, + 254, + 543 + ], + "spans": [ + { + "bbox": [ + 199, + 536, + 254, + 543 + ], + "type": "text", + "content": "Droid-SLAM + Ours" + } + ] + } + ], + "index": 40, + "angle": 0, + "type": "image_caption" + } + ], + "index": 38 + }, + { + "type": "image", + "bbox": [ + 239, + 434, + 282, + 536 + ], + "blocks": [ + { + "bbox": [ + 239, + 434, + 282, + 536 + ], + "lines": [ + { + "bbox": [ + 239, + 434, + 282, + 536 + ], + "spans": [ + { + "bbox": [ + 239, + 434, + 282, + 536 + ], + "type": "image", + "image_path": "7949bc9af021892dc8a2903c60c412d842c5eed246ae25c68e89b7d235af37dd.jpg" + } + ] + } + ], + "index": 39, + "angle": 0, + "type": "image_body" + } + ], + "index": 39 + }, + { + "bbox": [ + 55, + 630, + 295, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 630, + 295, + 713 + ], + "spans": [ + { + "bbox": [ + 55, + 630, + 295, + 713 + ], + "type": "text", + "content": "pletion (right). Droid-SLAM suffers from structural distortions and noise, especially in occluded or low-texture areas (red boxes), due to limited depth information. In contrast, our method generates more consistent reconstructions, preserving structural details and improving alignment (green boxes). These results highlight the advantage of dense depth completion in enhancing SLAM quality." + } + ] + } + ], + "index": 42 + }, + { + "bbox": [ + 314, + 353, + 387, + 365 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 314, + 353, + 387, + 365 + ], + "spans": [ + { + "bbox": [ + 314, + 353, + 387, + 365 + ], + "type": "text", + "content": "5. Conclusion" + } + ] + } + ], + "index": 43 + }, + { + "bbox": [ + 313, + 375, + 555, + 615 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 375, + 555, + 615 + ], + "spans": [ + { + "bbox": [ + 313, + 375, + 555, + 615 + ], + "type": "text", + "content": "In this work, we present " + }, + { + "bbox": [ + 313, + 375, + 555, + 615 + ], + "type": "inline_equation", + "content": "\\mathrm{DMD}^{3}\\mathrm{C}" + }, + { + "bbox": [ + 313, + 375, + 555, + 615 + ], + "type": "text", + "content": ", a novel two-stage distillation framework that distills knowledge from monocular depth estimation foundation models into the depth completion task. In the first stage, we introduce a data generation strategy that leverages monocular depth estimation and mesh reconstruction to simulate training data, allowing the model to learn geometric features from diverse natural images. In the second stage, we propose a scale- and shift-invariant loss (SSI Loss) combined with a supervised L1 loss with sparse ground truth, which addresses the scale ambiguity in monocular depth estimation. This ensures consistent depth completions across varying scales and focuses on learning real-world scale information. Extensive experiments on the depth completion benchmarks demonstrate that " + }, + { + "bbox": [ + 313, + 375, + 555, + 615 + ], + "type": "inline_equation", + "content": "\\mathrm{DMD}^{3}\\mathrm{C}" + }, + { + "bbox": [ + 313, + 375, + 555, + 615 + ], + "type": "text", + "content": " achieves state-of-the-art performance, ranking first on the KITTI leaderboard, and significantly outperforming existing methods. Our results highlight the framework's ability to produce high-quality depth maps with improved detail and structural consistency, making it a promising solution for depth completion tasks." + } + ] + } + ], + "index": 44 + }, + { + "bbox": [ + 314, + 632, + 416, + 645 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 314, + 632, + 416, + 645 + ], + "spans": [ + { + "bbox": [ + 314, + 632, + 416, + 645 + ], + "type": "text", + "content": "Acknowledgements" + } + ] + } + ], + "index": 45 + }, + { + "bbox": [ + 313, + 654, + 554, + 712 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 654, + 554, + 712 + ], + "spans": [ + { + "bbox": [ + 313, + 654, + 554, + 712 + ], + "type": "text", + "content": "This work was supported by the National Key R&D Program of China (2022YFC3300704), the National Natural Science Foundation of China (62331006, 62171038, and 62088101), and the Fundamental Research Funds for the Central Universities." + } + ] + } + ], + "index": 46 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 749, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 749, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 749, + 317, + 757 + ], + "type": "text", + "content": "22261" + } + ] + } + ], + "index": 47 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 7 + }, + { + "para_blocks": [ + { + "bbox": [ + 56, + 71, + 115, + 83 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 71, + 115, + 83 + ], + "spans": [ + { + "bbox": [ + 56, + 71, + 115, + 83 + ], + "type": "text", + "content": "References" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 57, + 91, + 296, + 713 + ], + "type": "list", + "angle": 0, + "index": 13, + "blocks": [ + { + "bbox": [ + 61, + 91, + 296, + 145 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 61, + 91, + 296, + 145 + ], + "spans": [ + { + "bbox": [ + 61, + 91, + 296, + 145 + ], + "type": "text", + "content": "[1] Luca Bartolomei, Matteo Poggi, Andrea Conti, Fabio Tosi, and Stefano Mattoccia. Revisiting depth completion from a stereo matching perspective for cross-domain generalization. In International Conference on 3D Vision (3DV), pages 1360-1370, 2024. 2" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 61, + 146, + 296, + 191 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 61, + 146, + 296, + 191 + ], + "spans": [ + { + "bbox": [ + 61, + 146, + 296, + 191 + ], + "type": "text", + "content": "[2] Shariq Farooq Bhat, Reiner Birkl, Diana Wofk, Peter Wonka, and Matthias Müller. Zoedepth: Zero-shot transfer by combining relative and metric depth. arXiv preprint arXiv:2302.12288, 2023. 2" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 62, + 191, + 295, + 257 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 191, + 295, + 257 + ], + "spans": [ + { + "bbox": [ + 62, + 191, + 295, + 257 + ], + "type": "text", + "content": "[3] Holger Caesar, Varun Bankiti, Alex H Lang, Sourabh Vora, Venice Erin Liong, Qiang Xu, Anush Krishnan, Yu Pan, Giancarlo Baldan, and Oscar Beijbom. nuscenes: A multimodal dataset for autonomous driving. In the IEEE Conference on Computer Vision and Pattern Recognition (CVPR), pages 11621-11631, 2020. 5" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 62, + 258, + 295, + 312 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 258, + 295, + 312 + ], + "spans": [ + { + "bbox": [ + 62, + 258, + 295, + 312 + ], + "type": "text", + "content": "[4] Angel Chang, Angela Dai, Thomas Funkhouser, Maciej Halber, Matthias Niebner, Manolis Savva, Shuran Song, Andy Zeng, and Yinda Zhang. Matterport3d: Learning from rgb-d data in indoor environments. In International Conference on 3D Vision (3DV), pages 667-676, 2017. 2" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 62, + 313, + 295, + 346 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 313, + 295, + 346 + ], + "spans": [ + { + "bbox": [ + 62, + 313, + 295, + 346 + ], + "type": "text", + "content": "[5] Linwei Chen, Ying Fu, Kaixuan Wei, Dezhi Zheng, and Felix Heide. Instance segmentation in the dark. International Journal of Computer Vision, 131(8):2198-2218, 2023. 2" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 62, + 347, + 295, + 400 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 347, + 295, + 400 + ], + "spans": [ + { + "bbox": [ + 62, + 347, + 295, + 400 + ], + "type": "text", + "content": "[6] Linwei Chen, Ying Fu, Lin Gu, Chenggang Yan, Tatsuya Harada, and Gao Huang. Frequency-aware feature fusion for dense image prediction. IEEE Transactions on Pattern Analysis and Machine Intelligence, 46(12):10763-10780, 2024. 2" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 62, + 403, + 295, + 446 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 403, + 295, + 446 + ], + "spans": [ + { + "bbox": [ + 62, + 403, + 295, + 446 + ], + "type": "text", + "content": "[7] Yun Chen, Bin Yang, Ming Liang, and Raquel Urtasun. Learning joint 2d-3d representations for depth completion. In the IEEE Conference on Computer Vision and Pattern Recognition (CVPR), pages 10023-10032, 2019. 1" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 62, + 447, + 295, + 491 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 447, + 295, + 491 + ], + "spans": [ + { + "bbox": [ + 62, + 447, + 295, + 491 + ], + "type": "text", + "content": "[8] Xinjing Cheng, Peng Wang, and Ruigang Yang. Learning depth with convolutional spatial propagation network. IEEE Transactions on Pattern Analysis and Machine Intelligence (TPAMI), 42(10):2361-2379, 2019. 2, 6" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 62, + 491, + 295, + 546 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 491, + 295, + 546 + ], + "spans": [ + { + "bbox": [ + 62, + 491, + 295, + 546 + ], + "type": "text", + "content": "[9] Xinjing Cheng, Peng Wang, Chenye Guan, and Ruigang Yang. Cspn++: Learning context and resource aware convolutional spatial propagation networks for depth completion. In the AAAI Conference on Artificial Intelligence (AAAI), pages 10615-10622, 2020. 2, 6" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 57, + 547, + 295, + 601 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 547, + 295, + 601 + ], + "spans": [ + { + "bbox": [ + 57, + 547, + 295, + 601 + ], + "type": "text", + "content": "[10] Keunhoon Choi, Somi Jeong, Youngjung Kim, and Kwanghoon Sohn. Stereo-augmented depth completion from a single rgb-lidar image. In IEEE International Conference on Robotics and Automation (ICRA), pages 13641-13647, 2021. 2" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 57, + 602, + 295, + 646 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 602, + 295, + 646 + ], + "spans": [ + { + "bbox": [ + 57, + 602, + 295, + 646 + ], + "type": "text", + "content": "[11] Andrea Conti, Matteo Poggi, and Stefano Mattoccia. Sparsity agnostic depth completion. In the IEEE Winter Conference on Applications of Computer Vision (WACV), pages 5871-5880, 2023. 2" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 57, + 647, + 295, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 647, + 295, + 713 + ], + "spans": [ + { + "bbox": [ + 57, + 647, + 295, + 713 + ], + "type": "text", + "content": "[12] Marius Cordts, Mohamed Omran, Sebastian Ramos, Timo Rehfeld, Markus Enzweiler, Rodrigo Benenson, Uwe Franke, Stefan Roth, and Bernt Schiele. The cityscapes dataset for semantic urban scene understanding. In the IEEE Conference on Computer Vision and Pattern Recognition (CVPR), pages 3213-3223, 2016. 5" + } + ] + } + ], + "index": 12 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 316, + 73, + 554, + 713 + ], + "type": "list", + "angle": 0, + "index": 27, + "blocks": [ + { + "bbox": [ + 316, + 73, + 554, + 128 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 73, + 554, + 128 + ], + "spans": [ + { + "bbox": [ + 316, + 73, + 554, + 128 + ], + "type": "text", + "content": "[13] Angela Dai, Angel X Chang, Manolis Savva, Maciej Halber, Thomas Funkhouser, and Matthias Nießner. Scannet: Richly-annotated 3d reconstructions of indoor scenes. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 5828-5839, 2017. 7" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 317, + 129, + 554, + 171 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 129, + 554, + 171 + ], + "spans": [ + { + "bbox": [ + 317, + 129, + 554, + 171 + ], + "type": "text", + "content": "[14] Zheng Fang, Ying Fu, and Lixiong Liu. A dual of transformer features-related map-intelligent generation method. Journal of Image and Graphics, 28(10):3281-3294, 2023. 2" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 317, + 172, + 554, + 215 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 172, + 554, + 215 + ], + "spans": [ + { + "bbox": [ + 317, + 172, + 554, + 215 + ], + "type": "text", + "content": "[15] Ying Fu, Zichun Wang, Tao Zhang, and Jun Zhang. Lowlight raw video denoising with a high-quality realistic motion dataset. IEEE Transactions on Multimedia, 25:8119-8131, 2022. 2" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 316, + 217, + 554, + 258 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 217, + 554, + 258 + ], + "spans": [ + { + "bbox": [ + 316, + 217, + 554, + 258 + ], + "type": "text", + "content": "[16] Qi Guan, Zihao Sheng, and Shibei Xue. Hrpose: Real-time high-resolution 6d pose estimation network using knowledge distillation. Chinese Journal of Electronics, 32(1):189-198, 2023. 2" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 316, + 260, + 554, + 316 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 260, + 554, + 316 + ], + "spans": [ + { + "bbox": [ + 316, + 260, + 554, + 316 + ], + "type": "text", + "content": "[17] Vitor Guizilini, Rares Ambrus, Sudeep Pillai, Allan Raventos, and Adrien Gaidon. 3d packing for self-supervised monocular depth estimation. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 2485-2494, 2020. 7" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 316, + 316, + 554, + 348 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 316, + 554, + 348 + ], + "spans": [ + { + "bbox": [ + 316, + 316, + 554, + 348 + ], + "type": "text", + "content": "[18] Wenxuan Guo, Yingping Liang, Zhiyu Pan, Ziheng Xi, Jianjiang Feng, and Jie Zhou. Camera-lidar cross-modality gait recognition. arXiv preprint arXiv:2407.02038, 2024. 2" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 316, + 350, + 554, + 403 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 350, + 554, + 403 + ], + "spans": [ + { + "bbox": [ + 316, + 350, + 554, + 403 + ], + "type": "text", + "content": "[19] Wenxuan Guo, Zhiyu Pan, Yingping Liang, Ziheng Xi, Zhicheng Zhong, Jianjiang Feng, and Jie Zhou. Lidar-based person re-identification. In the IEEE Conference on Computer Vision and Pattern Recognition (CVPR), pages 17437–17447, 2024. 2" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 316, + 404, + 554, + 449 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 404, + 554, + 449 + ], + "spans": [ + { + "bbox": [ + 316, + 404, + 554, + 449 + ], + "type": "text", + "content": "[20] Kaiming He, Xiangyu Zhang, Shaoqing Ren, and Jian Sun. Deep residual learning for image recognition. In the IEEE Conference on Computer Vision and Pattern Recognition (CVPR), pages 770-778, 2016. 5" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 317, + 449, + 554, + 502 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 449, + 554, + 502 + ], + "spans": [ + { + "bbox": [ + 317, + 449, + 554, + 502 + ], + "type": "text", + "content": "[21] Junjie Hu, Chenyu Bao, Mete Ozay, Chenyou Fan, Qing Gao, Honghai Liu, and Tin Lun Lam. Deep depth completion from extremely sparse data: A survey. IEEE Transactions on Pattern Analysis and Machine Intelligence (TPAMI), 45(7): 8244-8264, 2022. 1, 2" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 317, + 504, + 554, + 570 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 504, + 554, + 570 + ], + "spans": [ + { + "bbox": [ + 317, + 504, + 554, + 570 + ], + "type": "text", + "content": "[22] Mu Hu, Wei Yin, Chi Zhang, Zhipeng Cai, Xiaoxiao Long, Hao Chen, Kaixuan Wang, Gang Yu, Chunhua Shen, and Shaojie Shen. Metric3d v2: A versatile monocular geometric foundation model for zero-shot metric depth and surface normal estimation. IEEE Transactions on Pattern Analysis and Machine Intelligence (TPAMI), pages 1-18, 2024. 2" + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 317, + 571, + 554, + 623 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 571, + 554, + 623 + ], + "spans": [ + { + "bbox": [ + 317, + 571, + 554, + 623 + ], + "type": "text", + "content": "[23] Jinwoo Jeon, Hyunjun Lim, Dong-Uk Seo, and Hyun Myung. Struct-mdc: Mesh-refined unsupervised depth completion leveraging structural regularities from visual slam. IEEE Robotics and Automation Letters, 7(3):6391-6398, 2022. 1" + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 316, + 624, + 554, + 658 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 624, + 554, + 658 + ], + "spans": [ + { + "bbox": [ + 316, + 624, + 554, + 658 + ], + "type": "text", + "content": "[24] Michael Kazhdan, Matthew Bolitho, and Hugues Hoppe. Poisson surface reconstruction. In Eurographics Symposium on Geometry Processing, 2006. 4" + } + ] + } + ], + "index": 25 + }, + { + "bbox": [ + 316, + 658, + 554, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 658, + 554, + 713 + ], + "spans": [ + { + "bbox": [ + 316, + 658, + 554, + 713 + ], + "type": "text", + "content": "[25] Bingxin Ke, Anton Obukhov, Shengyu Huang, Nando Metzger, Rodrigo Caye Daudt, and Konrad Schindler. Repurposing diffusion-based image generators for monocular depth estimation. In the IEEE Conference on Computer Vision and Pattern Recognition (CVPR), pages 9492-9502, 2024. 1, 2" + } + ] + } + ], + "index": 26 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 749, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 749, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 749, + 317, + 757 + ], + "type": "text", + "content": "22262" + } + ] + } + ], + "index": 28 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 8 + }, + { + "para_blocks": [ + { + "bbox": [ + 56, + 72, + 297, + 712 + ], + "type": "list", + "angle": 0, + "index": 14, + "blocks": [ + { + "bbox": [ + 56, + 72, + 297, + 127 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 72, + 297, + 127 + ], + "spans": [ + { + "bbox": [ + 56, + 72, + 297, + 127 + ], + "type": "text", + "content": "[26] Muhammad Ahmed Ullah Khan, Danish Nazir, Alain Pagani, Hamam Mokayed, Marcus Liwicki, Didier Stricker, and Muhammad Zeshan Afzal. A comprehensive survey of depth completion approaches. Sensors, 22(18):6969, 2022. 1, 2" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 56, + 129, + 295, + 173 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 129, + 295, + 173 + ], + "spans": [ + { + "bbox": [ + 56, + 129, + 295, + 173 + ], + "type": "text", + "content": "[27] Byeong-Uk Lee, Hae-Gon Jeon, Sunghoon Im, and In So Kweon. Depth completion with deep geometry and context guidance. In International Conference on Robotics and Automation (ICRA), pages 3281-3287. IEEE, 2019. 1" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 56, + 175, + 295, + 218 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 175, + 295, + 218 + ], + "spans": [ + { + "bbox": [ + 56, + 175, + 295, + 218 + ], + "type": "text", + "content": "[28] Byeong-Uk Lee, Kyunghyun Lee, and In So Kweon. Depth completion using plane-residual representation. In the IEEE Conference on Computer Vision and Pattern Recognition (CVPR), pages 13916-13925, 2021. 1" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 56, + 220, + 295, + 252 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 220, + 295, + 252 + ], + "spans": [ + { + "bbox": [ + 56, + 220, + 295, + 252 + ], + "type": "text", + "content": "[29] Miaoyu Li and Ying Fu. Joint self-attention transformer for multispectral and hyperspectral image fusion. Journal of Image and Graphics, 28(12):3922–3934, 2023. 2" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 56, + 254, + 295, + 297 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 254, + 295, + 297 + ], + "spans": [ + { + "bbox": [ + 56, + 254, + 295, + 297 + ], + "type": "text", + "content": "[30] Miaoyu Li, Ying Fu, Tao Zhang, and Guanghui Wen. Supervise-assisted self-supervised deep-learning method for hyperspectral image restoration. IEEE Transactions on Neural Networks and Learning Systems, 2024. 1" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 56, + 299, + 295, + 331 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 299, + 295, + 331 + ], + "spans": [ + { + "bbox": [ + 56, + 299, + 295, + 331 + ], + "type": "text", + "content": "[31] Yingping Liang and Ying Fu. Relation-guided adversarial learning for data-free knowledge transfer. International Journal of Computer Vision, pages 1-18, 2024. 3" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 56, + 333, + 295, + 376 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 333, + 295, + 376 + ], + "spans": [ + { + "bbox": [ + 56, + 333, + 295, + 376 + ], + "type": "text", + "content": "[32] Yingping Liang, Jiaming Liu, Debing Zhang, and Ying Fu. Mpi-flow: Learning realistic optical flow with multiplane images. In the IEEE International Conference on Computer Vision (ICCV), pages 13857-13868, 2023. 2, 3" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 56, + 378, + 295, + 431 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 378, + 295, + 431 + ], + "spans": [ + { + "bbox": [ + 56, + 378, + 295, + 431 + ], + "type": "text", + "content": "[33] Tsung-Yi Lin, Michael Maire, Serge Belongie, James Hays, Pietro Perona, Deva Ramanan, Piotr Dólar, and C Lawrence Zitnick. Microsoft coco: Common objects in context. In the European Conference on Computer Vision (ECCV), pages 740-755, 2014. 5" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 56, + 434, + 295, + 477 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 434, + 295, + 477 + ], + "spans": [ + { + "bbox": [ + 56, + 434, + 295, + 477 + ], + "type": "text", + "content": "[34] Yuankai Lin, Tao Cheng, Qi Zhong, Wending Zhou, and Hua Yang. Dynamic spatial propagation network for depth completion. In the AAAI Conference on Artificial Intelligence (AAAI), pages 1638-1646, 2022. 2, 6" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 56, + 479, + 295, + 533 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 479, + 295, + 533 + ], + "spans": [ + { + "bbox": [ + 56, + 479, + 295, + 533 + ], + "type": "text", + "content": "[35] Lina Liu, Xibin Song, Xiaoyang Lyu, Junwei Diao, Mengmeng Wang, Yong Liu, and Liangjun Zhang. Fcfr-net: Feature fusion based coarse-to-fine residual learning for depth completion. In the AAAI Conference on Artificial Intelligence (AAAI), pages 2136-2144, 2021. 6" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 56, + 535, + 295, + 589 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 535, + 295, + 589 + ], + "spans": [ + { + "bbox": [ + 56, + 535, + 295, + 589 + ], + "type": "text", + "content": "[36] Qiankun Liu, Yuqi Jiang, Zhentao Tan, Dongdong Chen, Ying Fu, Qi Chu, Gang Hua, and Nenghai Yu. Transformer based pluralistic image completion with reduced information loss. IEEE Transactions on Pattern Analysis and Machine Intelligence, 46(10):6652-6668, 2024." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 56, + 590, + 295, + 624 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 590, + 295, + 624 + ], + "spans": [ + { + "bbox": [ + 56, + 590, + 295, + 624 + ], + "type": "text", + "content": "[37] Qiankun Liu, Yichen Li, Yuqi Jiang, and Ying Fu. Siamesedetr for generic multi-object tracking. IEEE Transactions on Image Processing, 33:3935-3949, 2024." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 56, + 625, + 295, + 669 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 625, + 295, + 669 + ], + "spans": [ + { + "bbox": [ + 56, + 625, + 295, + 669 + ], + "type": "text", + "content": "[38] Fangchang Ma and Sertac Karaman. Sparse-to-dense: Depth prediction from sparse depth samples and a single image. In IEEE international conference on robotics and automation (ICRA), pages 4796-4803, 2018. 2, 6" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 56, + 670, + 295, + 712 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 670, + 295, + 712 + ], + "spans": [ + { + "bbox": [ + 56, + 670, + 295, + 712 + ], + "type": "text", + "content": "[39] Fabiola Maffra, Lucas Teixeira, Zetao Chen, and Margarita Chli. Real-time wide-baseline place recognition using depth completion. IEEE Robotics and Automation Letters, 4(2): 1525-1532, 2019. 1" + } + ] + } + ], + "index": 13 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 316, + 72, + 554, + 713 + ], + "type": "list", + "angle": 0, + "index": 28, + "blocks": [ + { + "bbox": [ + 316, + 72, + 554, + 116 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 72, + 554, + 116 + ], + "spans": [ + { + "bbox": [ + 316, + 72, + 554, + 116 + ], + "type": "text", + "content": "[40] Hyoungseob Park, Anjali Gupta, and Alex Wong. Test-time adaptation for depth completion. In the IEEE Conference on Computer Vision and Pattern Recognition (CVPR), pages 20519-20529, 2024. 2" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 316, + 118, + 554, + 162 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 118, + 554, + 162 + ], + "spans": [ + { + "bbox": [ + 316, + 118, + 554, + 162 + ], + "type": "text", + "content": "[41] Jinsun Park, Kyungdon Joo, Zhe Hu, Chi-Kuei Liu, and In So Kweon. Non-local spatial propagation network for depth completion. In the European Conference on Computer Vision (ECCV), pages 120-136, 2020. 6" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 316, + 163, + 554, + 206 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 163, + 554, + 206 + ], + "spans": [ + { + "bbox": [ + 316, + 163, + 554, + 206 + ], + "type": "text", + "content": "[42] Jin-Hwi Park and Hae-Gon Jeon. A simple yet universal framework for depth completion. In The Thirty-eighth Annual Conference on Neural Information Processing Systems, 2024. 2" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 316, + 209, + 554, + 273 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 209, + 554, + 273 + ], + "spans": [ + { + "bbox": [ + 316, + 209, + 554, + 273 + ], + "type": "text", + "content": "[43] Federico Perazzi, Jordi Pont-Tuset, Brian McWilliams, Luc Van Gool, Markus Gross, and Alexander Sorkine-Hornung. A benchmark dataset and evaluation methodology for video object segmentation. In the IEEE Conference on Computer Vision and Pattern Recognition (CVPR), pages 724-732, 2016. 5" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 316, + 276, + 554, + 331 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 276, + 554, + 331 + ], + "spans": [ + { + "bbox": [ + 316, + 276, + 554, + 331 + ], + "type": "text", + "content": "[44] Luigi Piccinelli, Yung-Hsu Yang, Christos Sakaridis, Mattia Segu, Siyuan Li, Luc Van Gool, and Fisher Yu. Undepth: Universal monocular metric depth estimation. In the IEEE Conference on Computer Vision and Pattern Recognition (CVPR), pages 10106-10116, 2024. 2" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 316, + 332, + 554, + 397 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 332, + 554, + 397 + ], + "spans": [ + { + "bbox": [ + 316, + 332, + 554, + 397 + ], + "type": "text", + "content": "[45] Jiaxiong Qiu, Zhaopeng Cui, Yinda Zhang, Xingdi Zhang, Shuaicheng Liu, Bing Zeng, and Marc Pollefeys. Deeplidar: Deep surface normal guided depth prediction for outdoor scene from sparse lidar data and single color image. In the IEEE Conference on Computer Vision and Pattern Recognition (CVPR), pages 3313-3322, 2019. 2, 6" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 316, + 399, + 554, + 453 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 399, + 554, + 453 + ], + "spans": [ + { + "bbox": [ + 316, + 399, + 554, + 453 + ], + "type": "text", + "content": "[46] René Ranftl, Katrin Lasinger, David Hafner, Konrad Schindler, and Vladlen Koltun. Towards robust monocular depth estimation: Mixing datasets for zero-shot cross-dataset transfer. IEEE Transactions on Pattern Analysis and Machine Intelligence (TPAMI), 44(3), 2022. 1, 2, 3, 5" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 316, + 455, + 554, + 498 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 455, + 554, + 498 + ], + "spans": [ + { + "bbox": [ + 316, + 455, + 554, + 498 + ], + "type": "text", + "content": "[47] Ali Osman Serhatoglu, Oguzhan Guclu, and Ahmet Burak Can. Rgb-d slam with deep depth completion. In International Conference on Artificial Intelligence and Soft Computing (ICAISC), pages 59-67, 2022. 1" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 316, + 500, + 554, + 544 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 500, + 554, + 544 + ], + "spans": [ + { + "bbox": [ + 316, + 500, + 554, + 544 + ], + "type": "text", + "content": "[48] Nathan Silberman, Derek Hoiem, Pushmeet Kohli, and Rob Fergus. Indoor segmentation and support inference from rgbd images. In the European Conference on Computer Vision (ECCV), pages 120-136, 2012. 2, 5, 6" + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 316, + 545, + 554, + 588 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 545, + 554, + 588 + ], + "spans": [ + { + "bbox": [ + 316, + 545, + 554, + 588 + ], + "type": "text", + "content": "[49] Jie Tang, Fei-Peng Tian, Wei Feng, Jian Li, and Ping Tan. Learning guided convolutional network for depth completion. IEEE Transactions on Image Processing (TIP), 30: 1116-1129, 2020. 2, 6" + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 316, + 590, + 554, + 634 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 590, + 554, + 634 + ], + "spans": [ + { + "bbox": [ + 316, + 590, + 554, + 634 + ], + "type": "text", + "content": "[50] Jie Tang, Fei-Peng Tian, Boshi An, Jian Li, and Ping Tan. Bilateral propagation network for depth completion. In the IEEE Conference on Computer Vision and Pattern Recognition (CVPR), pages 9763-9772, 2024. 1, 2, 5, 6, 7, 8" + } + ] + } + ], + "index": 25 + }, + { + "bbox": [ + 316, + 635, + 554, + 678 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 635, + 554, + 678 + ], + "spans": [ + { + "bbox": [ + 316, + 635, + 554, + 678 + ], + "type": "text", + "content": "[51] Zachary Teed and Jia Deng. Droid-slam: Deep visual slam for monocular, stereo, and rgb-d cameras. Advances in Neural Information Processing Systems (NeurIPS), 34:16558-16569, 2021. 7" + } + ] + } + ], + "index": 26 + }, + { + "bbox": [ + 316, + 681, + 554, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 681, + 554, + 713 + ], + "spans": [ + { + "bbox": [ + 316, + 681, + 554, + 713 + ], + "type": "text", + "content": "[52] Ye Tian, Ying Fu, and Jun Zhang. Transformer-based undersampled single-pixel imaging. Chinese Journal of Electronics, 32(5):1151-1159, 2023. 2" + } + ] + } + ], + "index": 27 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 749, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 749, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 749, + 317, + 757 + ], + "type": "text", + "content": "22263" + } + ] + } + ], + "index": 29 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 9 + }, + { + "para_blocks": [ + { + "bbox": [ + 56, + 72, + 294, + 712 + ], + "type": "list", + "angle": 0, + "index": 12, + "blocks": [ + { + "bbox": [ + 56, + 72, + 294, + 115 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 72, + 294, + 115 + ], + "spans": [ + { + "bbox": [ + 56, + 72, + 294, + 115 + ], + "type": "text", + "content": "[53] Jonas Uhrig, Nick Schneider, Lukas Schneider, Uwe Franke, Thomas Brox, and Andreas Geiger. Sparsity invariant cnns. In International Conference on 3D Vision (3DV), 2017. 1, 2, 3, 5, 6" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 56, + 118, + 294, + 171 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 118, + 294, + 171 + ], + "spans": [ + { + "bbox": [ + 56, + 118, + 294, + 171 + ], + "type": "text", + "content": "[54] Haotian Wang, Meng Yang, and Nanning Zheng. G2-monodepth: A general framework of generalized depth inference from monocular " + }, + { + "bbox": [ + 56, + 118, + 294, + 171 + ], + "type": "inline_equation", + "content": "\\mathrm{rgb + x}" + }, + { + "bbox": [ + 56, + 118, + 294, + 171 + ], + "type": "text", + "content": " data. IEEE Transactions on Pattern Analysis and Machine Intelligence (TPAMI), 2023. 2, 3" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 56, + 172, + 294, + 216 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 172, + 294, + 216 + ], + "spans": [ + { + "bbox": [ + 56, + 172, + 294, + 216 + ], + "type": "text", + "content": "[55] Haotian Wang, Meng Yang, Xinhu Zheng, and Gang Hua. Scale propagation network for generalizable depth completion. IEEE Transactions on Pattern Analysis and Machine Intelligence (TPAMI), 2024. 2" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 56, + 217, + 294, + 270 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 217, + 294, + 270 + ], + "spans": [ + { + "bbox": [ + 56, + 217, + 294, + 270 + ], + "type": "text", + "content": "[56] Yufei Wang, Bo Li, Ge Zhang, Qi Liu, Tao Gao, and Yuchao Dai. Lrru: Long-short range recurrent updating networks for depth completion. In the IEEE Conference on Computer Vision and Pattern Recognition (CVPR), pages 9422-9432, 2023. 1, 6, 7, 8" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 56, + 272, + 294, + 326 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 272, + 294, + 326 + ], + "spans": [ + { + "bbox": [ + 56, + 272, + 294, + 326 + ], + "type": "text", + "content": "[57] Yufei Wang, Ge Zhang, Shaoqian Wang, Bo Li, Qi Liu, Le Hui, and Yuchao Dai. Improving depth completion via depth feature upsampling. In the IEEE Conference on Computer Vision and Pattern Recognition (CVPR), pages 21104-21113, 2024. 1, 2, 5, 6, 7, 8" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 56, + 327, + 294, + 381 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 327, + 294, + 381 + ], + "spans": [ + { + "bbox": [ + 56, + 327, + 294, + 381 + ], + "type": "text", + "content": "[58] Xiu-Shen Wei, Yi-Zhe Song, Oisin Mac Aodha, Jianxin Wu, Yuxin Peng, Jinhui Tang, Jian Yang, and Serge Belongie. Fine-grained image analysis with deep learning: A survey. IEEE transactions on pattern analysis and machine intelligence, 44(12):8927-8948, 2021. 1" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 56, + 383, + 294, + 437 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 383, + 294, + 437 + ], + "spans": [ + { + "bbox": [ + 56, + 383, + 294, + 437 + ], + "type": "text", + "content": "[59] Tobias Weyand, Andre Araujo, Bingyi Cao, and Jack Sim. Google landmarks dataset v2-a large-scale benchmark for instance-level recognition and retrieval. In the IEEE Conference on Computer Vision and Pattern Recognition (CVPR), pages 2575-2584, 2020. 5" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 56, + 437, + 294, + 480 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 437, + 294, + 480 + ], + "spans": [ + { + "bbox": [ + 56, + 437, + 294, + 480 + ], + "type": "text", + "content": "[60] Alex Wong, Xiaohan Fei, Stephanie Tsuei, and Stefano Soatto. Unsupervised depth completion from visual inertial odometry. IEEE Robotics and Automation Letters, 5(2): 1899-1906, 2020. 7" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 56, + 482, + 294, + 536 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 482, + 294, + 536 + ], + "spans": [ + { + "bbox": [ + 56, + 482, + 294, + 536 + ], + "type": "text", + "content": "[61] Xiaopei Wu, Liang Peng, Honghui Yang, Liang Xie, Chenxi Huang, Chengqi Deng, Haifeng Liu, and Deng Cai. Sparse fuse dense: Towards high quality 3d detection with depth completion. In the IEEE Conference on Computer Vision and Pattern Recognition (CVPR), pages 5418-5427, 2022. 1" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 56, + 537, + 294, + 591 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 537, + 294, + 591 + ], + "spans": [ + { + "bbox": [ + 56, + 537, + 294, + 591 + ], + "type": "text", + "content": "[62] Yangchao Wu, Tian Yu Liu, Hyoungseob Park, Stefano Soatto, Dong Lao, and Alex Wong. Augundo: Scaling up augmentations for monocular depth completion and estimation. In European Conference on Computer Vision (ECCV), pages 274-293, 2025. 2" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 56, + 592, + 294, + 656 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 592, + 294, + 656 + ], + "spans": [ + { + "bbox": [ + 56, + 592, + 294, + 656 + ], + "type": "text", + "content": "[63] Weijian Xie, Guanyi Chu, Quanhao Qian, Yihao Yu, Hai Li, Danpeng Chen, Shangjin Zhai, Nan Wang, Hujun Bao, and Guofeng Zhang. Depth completion with multiple balanced bases and confidence for dense monocular slam. IEEE Transactions on Visualization and Computer Graphics (TVCG), 2024. 1" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 56, + 658, + 294, + 712 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 658, + 294, + 712 + ], + "spans": [ + { + "bbox": [ + 56, + 658, + 294, + 712 + ], + "type": "text", + "content": "[64] Zexiao Xie, Xiaoxuan Yu, Xiang Gao, Kunqian Li, and Shuhan Shen. Recent advances in conventional and deep learning-based depth completion: A survey. IEEE Transactions on Neural Networks and Learning Systems (TNNLS), 35(3):3395-3415, 2024. 2" + } + ] + } + ], + "index": 11 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 316, + 72, + 555, + 712 + ], + "type": "list", + "angle": 0, + "index": 26, + "blocks": [ + { + "bbox": [ + 316, + 72, + 553, + 126 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 72, + 553, + 126 + ], + "spans": [ + { + "bbox": [ + 316, + 72, + 553, + 126 + ], + "type": "text", + "content": "[65] Yan Xu, Xinge Zhu, Jianping Shi, Guofeng Zhang, Hujun Bao, and Hongsheng Li. Depth completion from sparse lidar data with depth-normal constraints. In the IEEE Conference on Computer Vision and Pattern Recognition (CVPR), pages 2811-2820, 2019. 1" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 316, + 128, + 553, + 172 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 128, + 553, + 172 + ], + "spans": [ + { + "bbox": [ + 316, + 128, + 553, + 172 + ], + "type": "text", + "content": "[66] Zhiqiang Yan, Xiang Li, Kun Wang, Zhenyu Zhang, Jun Li, and Jian Yang. Multi-modal masked pre-training for monocular panoramic depth completion. In European Conference on Computer Vision, pages 378-395. Springer, 2022. 2" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 316, + 173, + 553, + 216 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 173, + 553, + 216 + ], + "spans": [ + { + "bbox": [ + 316, + 173, + 553, + 216 + ], + "type": "text", + "content": "[67] Zhiqiang Yan, Kun Wang, Xiang Li, Zhenyu Zhang, Jun Li, and Jian Yang. Rignet: Repetitive image guided network for depth completion. In the European Conference on Computer Vision (ECCV), pages 214-230, 2022. 1, 6" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 316, + 217, + 553, + 260 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 217, + 553, + 260 + ], + "spans": [ + { + "bbox": [ + 316, + 217, + 553, + 260 + ], + "type": "text", + "content": "[68] Zhiqiang Yan, Xiang Li, Kun Wang, Shuo Chen, Jun Li, and Jian Yang. Distortion and uncertainty aware loss for panoramic depth completion. In International Conference on Machine Learning, pages 39099-39109. PMLR, 2023." + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 316, + 261, + 555, + 315 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 261, + 555, + 315 + ], + "spans": [ + { + "bbox": [ + 316, + 261, + 555, + 315 + ], + "type": "text", + "content": "[69] Zhiqiang Yan, Kun Wang, Xiang Li, Zhenyu Zhang, Jun Li, and Jian Yang. Desnet: Decomposed scale-consistent network for unsupervised depth completion. In Proceedings of the AAAI conference on artificial intelligence, pages 3109-3117, 2023." + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 316, + 316, + 553, + 380 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 316, + 553, + 380 + ], + "spans": [ + { + "bbox": [ + 316, + 316, + 553, + 380 + ], + "type": "text", + "content": "[70] Zhiqiang Yan, Yuankai Lin, Kun Wang, Yupeng Zheng, Yufei Wang, Zhenyu Zhang, Jun Li, and Jian Yang. Tri-perspective view decomposition for geometry-aware depth completion. In the IEEE Conference on Computer Vision and Pattern Recognition (CVPR), pages 4874-4884, 2024. 1, 2, 6, 7" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 316, + 382, + 553, + 437 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 382, + 553, + 437 + ], + "spans": [ + { + "bbox": [ + 316, + 382, + 553, + 437 + ], + "type": "text", + "content": "[71] Lihe Yang, Bingyi Kang, Zilong Huang, Xiaogang Xu, Jiashi Feng, and Hengshuang Zhao. Depth anything: Unleashing the power of large-scale unlabeled data. In the IEEE Conference on Computer Vision and Pattern Recognition (CVPR), pages 10371-10381, 2024. 1, 2, 3" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 316, + 437, + 553, + 480 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 437, + 553, + 480 + ], + "spans": [ + { + "bbox": [ + 316, + 437, + 553, + 480 + ], + "type": "text", + "content": "[72] Lihe Yang, Bingyi Kang, Zilong Huang, Zhen Zhao, Xiaogang Xu, Jiashi Feng, and Hengshuang Zhao. Depth anything v2. In Annual Conference on Neural Information Processing Systems (NeurIPS), 2024. 1, 2, 3, 5" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 316, + 481, + 553, + 536 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 481, + 553, + 536 + ], + "spans": [ + { + "bbox": [ + 316, + 481, + 553, + 536 + ], + "type": "text", + "content": "[73] Wei Yin, Chi Zhang, Hao Chen, Zhipeng Cai, Gang Yu, Kaixuan Wang, Xiaozhi Chen, and Chunhua Shen. Metric3d: Towards zero-shot metric 3d prediction from a single image. In the IEEE Conference on Computer Vision and Pattern Recognition (CVPR), pages 9043-9053, 2023. 2" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 316, + 537, + 553, + 591 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 537, + 553, + 591 + ], + "spans": [ + { + "bbox": [ + 316, + 537, + 553, + 591 + ], + "type": "text", + "content": "[74] Zhu Yu, Zehua Sheng, Zili Zhou, Lun Luo, Si-Yuan Cao, Hong Gu, Huaqi Zhang, and Hui-Liang Shen. Aggregating feature point cloud for depth completion. In the IEEE International Conference on Computer Vision (ICCV), pages 8732-8743, 2023. 2, 6" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 316, + 592, + 553, + 635 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 592, + 553, + 635 + ], + "spans": [ + { + "bbox": [ + 316, + 592, + 553, + 635 + ], + "type": "text", + "content": "[75] Fan Zhang, Shaodi You, Yu Li, and Ying Fu. Atlantis: Enabling underwater depth estimation with stable diffusion. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 11852-11861, 2024. 1" + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 316, + 636, + 553, + 669 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 636, + 553, + 669 + ], + "spans": [ + { + "bbox": [ + 316, + 636, + 553, + 669 + ], + "type": "text", + "content": "[76] Tao Zhang, Ying Fu, and Jun Zhang. Deep guided attention network for joint denoising and demosaicing in real image. Chinese Journal of Electronics, 33(1):303-312, 2024. 2" + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 316, + 670, + 553, + 712 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 670, + 553, + 712 + ], + "spans": [ + { + "bbox": [ + 316, + 670, + 553, + 712 + ], + "type": "text", + "content": "[77] Xuanmeng Zhang, Zhedong Zheng, Minyue Jiang, and Xiaojing Ye. Self-ensembling depth completion via density-aware consistency. Pattern Recognition, 154:110618, 2024. 2" + } + ] + } + ], + "index": 25 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 749, + 318, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 749, + 318, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 749, + 318, + 757 + ], + "type": "text", + "content": "22264" + } + ] + } + ], + "index": 27 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 10 + }, + { + "para_blocks": [ + { + "bbox": [ + 56, + 73, + 294, + 417 + ], + "type": "list", + "angle": 0, + "index": 7, + "blocks": [ + { + "bbox": [ + 56, + 73, + 294, + 115 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 73, + 294, + 115 + ], + "spans": [ + { + "bbox": [ + 56, + 73, + 294, + 115 + ], + "type": "text", + "content": "[78] Yinda Zhang and Thomas Funkhouser. Deep depth completion of a single rgb-d image. In the IEEE Conference on Computer Vision and Pattern Recognition (CVPR), pages 175-185, 2018. 1" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 56, + 118, + 294, + 182 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 118, + 294, + 182 + ], + "spans": [ + { + "bbox": [ + 56, + 118, + 294, + 182 + ], + "type": "text", + "content": "[79] Youmin Zhang, Xianda Guo, Matteo Poggi, Zheng Zhu, Guan Huang, and Stefano Mattoccia. Completionformer: Depth completion with convolutions and vision transformers. In the IEEE Conference on Computer Vision and Pattern Recognition (CVPR), pages 18527-18536, 2023. 1, 2, 5, 6, 7, 8" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 56, + 185, + 294, + 228 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 185, + 294, + 228 + ], + "spans": [ + { + "bbox": [ + 56, + 185, + 294, + 228 + ], + "type": "text", + "content": "[80] Shanshan Zhao, Mingming Gong, Huan Fu, and Dacheng Tao. Adaptive context-aware multi-modal network for depth completion. IEEE Transactions on Image Processing (TIP), 30:5264-5276, 2021. 2, 6" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 56, + 230, + 294, + 271 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 230, + 294, + 271 + ], + "spans": [ + { + "bbox": [ + 56, + 230, + 294, + 271 + ], + "type": "text", + "content": "[81] Jin Zheng, Botao Jiang, Wei Peng, and Qiaohui Zhang. Multi-scale binocular stereo matching based on semantic association. Chinese Journal of Electronics, 33(4):1010-1022, 2024. 2" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 56, + 274, + 294, + 328 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 274, + 294, + 328 + ], + "spans": [ + { + "bbox": [ + 56, + 274, + 294, + 328 + ], + "type": "text", + "content": "[82] Wending Zhou, Xu Yan, Yinghong Liao, Yuankai Lin, Jin Huang, Gangming Zhao, Shuguang Cui, and Zhen Li. Bev@ dc: Bird's-eye view assisted training for depth completion. In the IEEE Conference on Computer Vision and Pattern Recognition (CVPR), pages 9233-9242, 2023. 2, 6" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 56, + 331, + 294, + 373 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 331, + 294, + 373 + ], + "spans": [ + { + "bbox": [ + 56, + 331, + 294, + 373 + ], + "type": "text", + "content": "[83] Yunhao Zou, Ying Fu, Tsuyoshi Takatani, and Yinqiang Zheng. Eventhrd: From event to high-speed hdr videos and beyond. IEEE Transactions on Pattern Analysis and Machine Intelligence, 47(1):32-50, 2024. 2" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 56, + 375, + 294, + 417 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 375, + 294, + 417 + ], + "spans": [ + { + "bbox": [ + 56, + 375, + 294, + 417 + ], + "type": "text", + "content": "[84] Yiming Zuo and Jia Deng. Ogni-dc: Robust depth completion with optimization-guided neural iterations. In the European Conference on Computer Vision (ECCV), pages 78-95, 2025. 2" + } + ] + } + ], + "index": 6 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 749, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 749, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 749, + 317, + 757 + ], + "type": "text", + "content": "22265" + } + ] + } + ], + "index": 8 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 11 + } + ], + "_backend": "vlm", + "_version_name": "2.6.4" +} \ No newline at end of file diff --git a/2025/Distilling Multi-modal Large Language Models for Autonomous Driving/2bea0a6d-f2ee-4f71-9581-707f2552ee73_content_list.json b/2025/Distilling Multi-modal Large Language Models for Autonomous Driving/2bea0a6d-f2ee-4f71-9581-707f2552ee73_content_list.json new file mode 100644 index 0000000000000000000000000000000000000000..1d0380c1ed2d5b49eccc009b753d629ad0f5a991 --- /dev/null +++ b/2025/Distilling Multi-modal Large Language Models for Autonomous Driving/2bea0a6d-f2ee-4f71-9581-707f2552ee73_content_list.json @@ -0,0 +1,1393 @@ +[ + { + "type": "text", + "text": "Distilling Multi-modal Large Language Models for Autonomous Driving", + "text_level": 1, + "bbox": [ + 135, + 130, + 864, + 152 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Deepti Hegde $^{1\\dagger*}$ , Rajeev Yasarla $^{2*}$ , Hong Cai $^{2}$ , Shizhong Han $^{2}$ , Apratim Bhattacharyya $^{2}$ , Shweta Mahajan $^{2}$ , Litian Liu $^{2}$ , Risheek Garrepalli $^{2}$ , Vishal M. Patel $^{1}$ , Fatih Porikli $^{2}$", + "bbox": [ + 147, + 179, + 849, + 215 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "$^{1}$ Johns Hopkins University, $^{2}$ Qualcomm AI Research‡ *Equal contribution", + "bbox": [ + 279, + 220, + 718, + 258 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Abstract", + "text_level": 1, + "bbox": [ + 248, + 292, + 326, + 309 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Autonomous driving demands safe motion planning, especially in critical \"long-tail\" scenarios. Recent end-to-end autonomous driving systems leverage large language models (LLMs) as planners to improve generalizability to rare events. However, using LLMs at test time introduces high computational costs. To address this, we propose DiMA, an end-to-end autonomous driving system that maintains the efficiency of an LLM-free (or vision-based) planner while leveraging the world knowledge of an LLM. DiMA distills the information from a multi-modal LLM to a vision-based end-to-end planner through a set of specially designed surrogate tasks. Under a joint training strategy, a scene encoder common to both networks produces structured representations that are semantically grounded as well as aligned to the final planning objective. Notably, the LLM is optional at inference, enabling robust planning without compromising on efficiency. Training with DiMA results in a $37\\%$ reduction in the L2 trajectory error and an $80\\%$ reduction in the collision rate of the vision-based planner, as well as a $44\\%$ trajectory error reduction in long-tail scenarios. DiMA also achieves state-of-the-art performance on the nuScenes planning benchmark.", + "bbox": [ + 89, + 325, + 483, + 657 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "1. Introduction", + "text_level": 1, + "bbox": [ + 91, + 675, + 220, + 690 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Research in autonomous driving has moved away from the traditional strategy of integrating sequential, independently trained models [4, 22, 26, 29, 31, 45, 52], and moved towards developing multi-task systems that are trained in an end-to-end manner [3, 5, 7, 16, 20, 49]. These approaches demonstrate improved performance and interpretability while being highly efficient. However, they struggle with long-tail navigation and perception scenarios, primarily because they are trained on task-specific, limited datasets [2, 11, 13].", + "bbox": [ + 89, + 696, + 482, + 847 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Large language models (LLMs) have emerged as a promising solution to this issue. Trained on vast, internet-", + "bbox": [ + 89, + 848, + 483, + 878 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/e9bc53db64c7aec28a15abcce8c7550c42f21a271c907e1cf3ff15a85ecf054c.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 517, + 292, + 888, + 412 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/7bfe5d2afb4b74a03c120aadb7df0997c6c5bc631c3cc8c96557dec693a18d3e.jpg", + "image_caption": [ + "Figure 1. Comparison of planning performance in long-tail scenarios from nuScenes: DiMA-VAD demonstrates greater robustness compared to VAD [20] in long-tail navigation scenarios such as overtaking a vehicle and performing a 3-point turn. DiMA-VAD also outperforms recent vision-planner PARA-Drive [48] and LLM planner TILEN [38]. Notably, the 3-point turn is a zero-shot scenario that is only present in the validation set." + ], + "image_footnote": [], + "bbox": [ + 517, + 415, + 888, + 534 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "scale datasets, LLMs can leverage world knowledge to generalize to unseen or rare scenarios. These models can perform high-level reasoning tasks using mechanisms such as chain-of-thought [46]. Going beyond text-based prompting, multi-modal large language models (MLLMs) integrate image and video inputs to the LLM, enabling tasks such as visual question answering and dense captioning [21, 24]. Recent end-to-end autonomous driving systems leverage LLMs to achieve higher interpretability through language-guided reasoning and superior robustness to long-tail scenarios [19, 37-39, 43, 44]. To differentiate between these methods, we call end-to-end planners that depend on LLMs to perform trajectory prediction as \"LLM-based\" planners and ones that do not as \"vision-based\" planners [18, 20, 48]. In spite of their recent success \"LLM-based\" planners face significant challenges.", + "bbox": [ + 511, + 657, + 908, + 902 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "CVF", + "bbox": [ + 106, + 2, + 181, + 42 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore.", + "bbox": [ + 236, + 0, + 810, + 46 + ], + "page_idx": 0 + }, + { + "type": "page_footnote", + "text": "Qualcomm AI Research is initiative of Qualcomm Technologies, Inc.", + "bbox": [ + 114, + 886, + 482, + 900 + ], + "page_idx": 0 + }, + { + "type": "page_number", + "text": "27575", + "bbox": [ + 478, + 944, + 517, + 957 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Firstly, LLM-based planners require a significant amount of computational overhead at test time, limiting their practicality. This work addresses a core question: how can we harness LLMs' world knowledge while preserving the efficiency of vision-based planners? Secondly, bridging the visual and language domains is inherently more complex for end-to-end planning tasks compared to general MLLM tasks [21, 24], with the additional hurdle of limited training data. Standard image tokenization strategies rely on frozen pre-trained image encoders [33] to generate dense, unstructured multi-modal token embeddings. [21, 24, 37]. We suggest that MLLMs meant for end-to-end autonomous driving benefit from structured inputs that explicitly model scene components. Furthermore, we suggest that updating these features during training results in richer representations.", + "bbox": [ + 88, + 90, + 480, + 316 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "To address these challenges, we propose DiMA, a novel framework for Distilling Multi-modal Large Language Models for Autonomous driving. We introduce a joint-training scheme between a vision-based planner and an MLLM that enables the learning of robust, grounded and disentangled scene representations that are aligned to the final objective of planning. Specifically, we use the vision-based planner as a tokenizer, and pass the learned scene representations to the MLLM, providing a more structured input. The MLLM is then trained for the tasks of visual question-answering, trajectory estimation, and a series of surrogate tasks designed to ground the multi-modal inputs in language. The vision-based planner is simultaneously trained for trajectory estimation, while also distilling features from the MLLM planning head to the planning transformer. Crucially, the MLLM can be discarded when performing planning inference, maintaining the efficiency of the vision-based planner while utilizing the knowledge of the language model. Optionally, the MLLM can support language-guided reasoning through visual question answering. Our experiments show that DiMA is more robust to challenging long-tail scenarios compared to baselines and state-of-the-art methods [38, 48]. Figure 1 confirms this both quantitatively and qualitatively. We list the contributions of the work below:", + "bbox": [ + 91, + 320, + 483, + 696 + ], + "page_idx": 1 + }, + { + "type": "list", + "sub_type": "text", + "list_items": [ + "1. We introduce DiMA, an end-to-end autonomous driving framework that distills knowledge from an MLLM to a vision-based planner to ensure robustness to long-tail events while maintaining efficiency. DiMA is capable of planning as well as visual question-answering.", + "2. We propose a distillation task along with the following surrogate tasks to align the objectives of the vision-based planner and the MLLM: (i) masked token reconstruction (ii) future token prediction (iii) scene editing.", + "3. DiMA outperforms both vision-based and LLM end-to-end planners, achieving state-of-the-art results on the nuScenes planning benchmark. Training with DiMA results in a $37\\%$ reduction in the L2 trajectory error and an" + ], + "bbox": [ + 89, + 704, + 480, + 900 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "80% reduction in the collision rate of the vision-based planner, as well as a 44% trajectory error reduction in long-tail scenarios.", + "bbox": [ + 532, + 90, + 903, + 136 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "2. Related Work", + "text_level": 1, + "bbox": [ + 511, + 151, + 653, + 166 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "2.1. End-to-End Autonomous Driving", + "text_level": 1, + "bbox": [ + 511, + 172, + 805, + 188 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Autonomous driving (AD) systems are generally modular, consisting of tasks such as perception [22, 31, 45, 52], motion prediction [4, 26, 29], and planning [9]. Classic AD systems use standalone models for each task where each module is optimized separately and combined sequentially. Alternatively, some approaches incorporate these modular tasks into a multi-task learning paradigm, where tasks shared a common feature extraction process but have task-specific heads [8, 23, 53, 55]. Recently, research has moved toward optimizing autonomous driving systems in an end-to-end manner, aiming to directly predict future trajectories [3, 5, 7, 16, 20, 49]. Along this line of work, UniAD [18] introduces a query-based design that integrates perception and prediction components, enabling an end-to-end planning scheme. VAD [20] utilizes a vectorized scene representation, replacing the dense rasterized representations from [18], reducing computational cost and improving planning performance. While effective in general navigation scenarios, these methods struggle in difficult long-tail events (see Figure 1). In this work, we propose a method for end-to-end autonomous driving that is robust to such scenarios.", + "bbox": [ + 511, + 191, + 903, + 507 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "2.2. LLMs for Autonomous Driving", + "text_level": 1, + "bbox": [ + 511, + 518, + 789, + 534 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "The recent success of large language models [12, 41, 42] demonstrates their ability to process complex contextual information with logical reasoning and to communicate in a human-interpretable manner, attracting significant attention in autonomous driving [6, 10, 14, 25, 28, 35-37, 39, 44, 47, 50]. These works investigate the capabilities of language models to generalize to novel scenarios as well as their ability to reason about the scene in the form of text.", + "bbox": [ + 511, + 537, + 903, + 656 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Prompting LLMs. One category of approaches involves prompting LLMs with text using reasoning frameworks such as chain-of-thought [46]. LanguageMPC [35] leverages the common-sense reasoning capacity of LLM to make high-level driving decisions, which are then translated into bottom-level control signals. GPT-Driver [28] converts observations and ego-states into language prompts, framing driving planning as a text-based task for LLMs.", + "bbox": [ + 511, + 657, + 903, + 779 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "End-to-end Driving with MLLMs. Another approach is to integrate multi-modal LLMs into end-to-end train frameworks for autonomous vehicles. LMDrive [36] integrates a vision encoder with a large language model, enabling natural language instruction in autonomous driving. DriveLM introduces multi-step question answering, going beyond single-round interactions [27, 32, 34]. DriveVLM [39] leverages MLLMs and a chain-of-thought process to", + "bbox": [ + 511, + 780, + 903, + 900 + ], + "page_idx": 1 + }, + { + "type": "page_number", + "text": "27576", + "bbox": [ + 478, + 944, + 519, + 955 + ], + "page_idx": 1 + }, + { + "type": "image", + "img_path": "images/5283f6fe783ad4a42906f06c27fdf6a19d35035d65f2d988985427793a7e386e.jpg", + "image_caption": [ + "Figure 2. Overview of DiMA. The input to the framework is a multi-view image sequence and a question text prompt. The vision-based end-to-end planner consists of a scene encoder and a planning transformer. The scene encoder learns structured latent representations in the form of bird's-eye-view, ego, agent, and map (BEAM) token embeddings and acts as a trainable tokenizer for the multi-modal large language model (MLLM). The planning transformer is trained under standard planning constraints [18, 20]. The MLLM is trained for planning, visual question answering, distillation, and a series of surrogate tasks." + ], + "image_footnote": [], + "bbox": [ + 133, + 90, + 857, + 303 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "enhance scene understanding and planning.", + "bbox": [ + 89, + 401, + 377, + 415 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "We point out a challenge in the current MLLM strategy for autonomous driving. We suggest that in order to reason about vehicle dynamics, the language model benefits from a structured input that explicitly models scene components instead of dense, unstructured input to the MLLM as in [36, 39, 43, 44]. In TOKEN [38], the scene encoder from [48] is used as a tokenizer. However, it is frozen during training. In our approach, the input to the MLLM is a set of latent representations learned from a pre-trained scene encoder that is jointly trained with the MLLM, allowing it to more learn grounded feature representations", + "bbox": [ + 88, + 416, + 483, + 583 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3. DiMA Framework", + "text_level": 1, + "bbox": [ + 89, + 594, + 272, + 609 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "We present DiMA, a framework for end-to-end autonomous driving. Given a sequence of multi-view images, the overall objective is to predict the future trajectory of the ego vehicle and answer questions about the scene. An overview of DiMA is shown in Figure 2. The framework has two main components: 1) the vision-based planner; 2) a multi-modal large language model consisting of adapter layers, an LLM, and a series of task-specific decoder heads.", + "bbox": [ + 88, + 616, + 483, + 737 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3.1. Vision-based Planner", + "text_level": 1, + "bbox": [ + 89, + 746, + 290, + 760 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Vision-based end-to-end planners are trained in a multi-task fashion to perform perception, mapping, motion prediction and planning [18, 20, 40, 48]. Considering standard architecture design from [18, 20, 48] we decompose the planner into the scene encoder and the planning transformer. The scene encoder provides structured latent representations to the planning transformer that performs waypoint prediction. We call these representations as scene token embeddings. In our framework, we leverage the vision-based planner in", + "bbox": [ + 88, + 763, + 483, + 901 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "two ways. First, it acts as the primary network through which planning is performed, allowing for fast prediction at test time. Second, the scene encoder is shared with the MLLM, acting as tokenizer to provide it with a highly structured set of inputs. An important distinction of our framework from previous works [19, 38] is that the scene encoder is trained jointly with the MLLM.", + "bbox": [ + 511, + 401, + 906, + 506 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Scene encoder The scene encoder models the scene as high-dimensional token embeddings that explicitly represent components such as the environment map, the ego vehicle, and the surrounding agents. This is achieved by introducing task specific, learnable query features that are supervised by a series of planning constraints [20]. First, a visual backbone encodes the input multi-view image sequence. The obtained feature map is projected into the bird's-eye-view (BEV) space using a set of queries to obtain BEV token embeddings $(B)$ [22]. The network learns structured representations of the map and agents in the form of map token embeddings $(M)$ and agent token $(A)$ embeddings by cross-attending the BEV features with map and agent queries. A randomly initialized learnable embedding is trained to learn the interaction of the ego vehicle with the agent and map components which is the ego query $(E)$ . We denote the set of these latent scene representations as $BEAM$ token embeddings. These token embeddings are sent to a planning head that predicts future trajectories of the ego and agent vehicles. They are also used as input to the MLLM as multi-modal inputs, as discussed next.", + "bbox": [ + 511, + 508, + 908, + 821 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3.2. Multi-modal LLM", + "text_level": 1, + "bbox": [ + 511, + 834, + 692, + 849 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Our objective is to distill knowledge from an MLLM to the vision-based planner. Towards this, we jointly train an MLLM with the vision-based planner. Specifically, the", + "bbox": [ + 511, + 854, + 906, + 902 + ], + "page_idx": 2 + }, + { + "type": "page_number", + "text": "27577", + "bbox": [ + 478, + 944, + 517, + 955 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "scene encoder is used as a trainable tokenizer for the MLLM to generate the $BEAM$ token embeddings, as well as used as an encoder for the planning transformer. The advantage from this strategy is two-fold. First, the MLLM receives a highly structured input that captures rich spatio-temporal information relevant to the autonomous driving task. Second, the scene encoder learns features that are grounded in language, improving the robustness of vision-based planning. The main components of the MLLM are adapter layers, the large language model, and a series of task-specific decoder heads which we discuss next.", + "bbox": [ + 89, + 90, + 483, + 257 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "3.2.1. Adaptation of Scene Tokens", + "text_level": 1, + "bbox": [ + 89, + 263, + 331, + 277 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "In order to project the BEAM token embeddings efficiently while maintaining their distinctiveness, we leverage the query-transformer (Q-former) module [17] to compress the visual tokens before being input to the LLM. We implement component-specific Q-former layers, namely a Map Q-former, a Bird's-Eye-View (BEV) Q-former, an Ego Q-former, and an Agent Q-former. Each of these adapters transforms the representing different modalities into a sequence of fixed-length, higher-dimensional tokens, preparing them for processing by the MLLM.", + "bbox": [ + 89, + 280, + 483, + 431 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "3.2.2. MLLM Supervision", + "text_level": 1, + "bbox": [ + 89, + 436, + 274, + 452 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "We design tasks for the MLLM with the objectives of, a) enriching the intermediate scene representations; b) grounding the scene token embeddings in language; and c) training the LLM for planning-related reasoning. In detail, the MLLM is trained for visual question answering, trajectory estimation, feature distillation, and a set of surrogate tasks [1, 51].", + "bbox": [ + 89, + 455, + 483, + 546 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Visual question answering. The MLLM is trained for visual question answering (VQA) on question-answer pairs of 4 types: perception of the scene, prediction of the agent behavior, prediction of ego-vehicle behavior, and future planning steps [37]. The answer is predicted based on a multimodal prompt consisting of the question embeddings and the projected BEAM token embeddings. This is constructed as a standard next-token prediction task and the VQA branch is supervised with a cross-entropy loss as in [21]. We denote the loss from this task head as $\\mathcal{L}_{LLM}$ .", + "bbox": [ + 89, + 549, + 483, + 700 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Surrogate tasks: The main goal of joint training is to enrich the BEAM scene representations learned by the scene encoder. We design MLLM surrogate tasks whose objectives are in line with that of future trajectory prediction.", + "bbox": [ + 89, + 704, + 482, + 763 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "1) Masked token reconstruction: Each type of token embedding contributes to a holistic representation of the scene. In order to enrich the visual representations, we ask the network to reconstruct a masked BEV input based on the context present in the rest of the multi-modal sequence. We perform random masking after scene encoding and pass the token embeddings to the MLLM. A reconstruction head takes the latent representations from the penultimate layer of LLM and predicts reconstructed BEV token embeddings", + "bbox": [ + 89, + 763, + 483, + 901 + ], + "page_idx": 3 + }, + { + "type": "image", + "img_path": "images/14389e201993856381630025fdf42676d5dd9c81bbdd68528ae5fc3cd0142134.jpg", + "image_caption": [ + "Figure 3. Examples of addition and deletion in scene editing. In the top row, a car (solid red box) is added on the left of the ego-vehicle (green box). In the bottom row, a car (dashed red box) is removed from the right of the ego-vehicle. A corresponding question-answer pair is created to characterize the edit." + ], + "image_footnote": [], + "bbox": [ + 553, + 88, + 864, + 335 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "$\\hat{B}$ . This decoder head is supervised with the L2 loss between the prediction and the complete input,", + "bbox": [ + 511, + 426, + 905, + 458 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {L} _ {\\text {r e c o n}} = \\left\\| \\hat {B} - B \\right\\| ^ {2} \\tag {1}\n$$\n", + "text_format": "latex", + "bbox": [ + 638, + 459, + 903, + 477 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "where $(\\{m(B), E, A, M\\})$ is input to the MLLM, and the latent MLLM representations associated with the BEV token embeddings are input to the masked reconstruction decoder head. Here, $m(.)$ denotes random masking.", + "bbox": [ + 511, + 479, + 905, + 541 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "2) Future BEV prediction: An important aspect of planning is anticipating future events. We introduce the surrogate task of future BEV prediction to encourage the LLM to learn spatio-temporal cues useful for planning. Given latent BEV token embeddings, we train a prediction head to predict future BEV token embeddings and supervise this task as the L2 loss between predicted and ground truth future token embeddings:", + "bbox": [ + 511, + 542, + 906, + 662 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {L} _ {\\text {f u t u r e}} = \\left\\| \\hat {F} _ {t} - B _ {t + 1} \\right\\| ^ {2} + \\left\\| \\tilde {F} _ {t} - B _ {t + 2} \\right\\| ^ {2} \\tag {2}\n$$\n", + "text_format": "latex", + "bbox": [ + 562, + 664, + 903, + 683 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "where $\\hat{F}_t, \\tilde{F}_t$ are the predicted future BEV token embeddings at time $t$ . Note, we predict future BEV token embeddings $(B_{t+1}, B_{t+2})$ of the multi-view image sequence for time steps $\\{t+1, t+2\\}$ .", + "bbox": [ + 511, + 686, + 905, + 748 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "3) Scene editing: For prediction and reasoning about the ego vehicle, it is crucial to learn how surrounding agents impact the ego-vehicle's future path. We propose a novel scene editing task in which we augment scenes by removing or adding new agents. Along with this, we construct a question-answer pair related to the edit. We show examples of this in Figure 3. For scene addition, given the map constraints, predicted map, the ego bounding box location and the trajectories of predicted agents, we create a way-point trajectory for a new agent of category \"car\" or \"truck\". A", + "bbox": [ + 511, + 750, + 906, + 901 + ], + "page_idx": 3 + }, + { + "type": "page_number", + "text": "27578", + "bbox": [ + 478, + 944, + 517, + 955 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "new agent token embedding is then created using a linear layer. This new agent token embedding, the corresponding text prompt, and the rest of the BEAM token embeddings are passed as input to the LLM. The hidden latent LLM features corresponding to agent token embeddings are then fed into a dedicated scene editing decoder head that performs waypoint prediction of the ego vehicle. The language prediction head performs question-answering on the new QA pair. This task thus contributes to the existing planning constraint loss and VQA loss of the MLLM.", + "bbox": [ + 89, + 90, + 480, + 241 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Distillation In order to further align the representations learned by the vision planner and the MLLM, we perform knowledge transfer between the penultimate layers of the language model and the planning transformer. Concretely, we minimize the KL-divergence between the distributions of the hidden features of the planning transformer and the LLM hidden features of the ego-token embeddings:", + "bbox": [ + 89, + 246, + 480, + 351 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {L} _ {\\text {d i s t i l l}} = D _ {K L} \\left(P _ {l l m} \\| P _ {v i s}\\right) \\tag {3}\n$$\n", + "text_format": "latex", + "bbox": [ + 189, + 354, + 480, + 371 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "where the $P_{vis}$ are the features of penultimate layers of vision planning transformer and $P_{llm}$ are the hidden LLM embeddings features corresponding to the ego-token embeddings of penultimate layers of MLLM model.", + "bbox": [ + 89, + 375, + 480, + 434 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Loss functions. We supervise the network with a weighted sum of the losses corresponding to planning, visual question answering, distillation, and each surrogate task.", + "bbox": [ + 89, + 435, + 480, + 479 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {L} = \\mathcal {L} _ {\\text {p l a n n i n g}} + \\mathcal {L} _ {\\text {L L M}} + \\mathcal {L} _ {\\text {r e c o n}} + \\mathcal {L} _ {\\text {f u t u r e}} + \\mathcal {L} _ {\\text {d i s t i l l}} \\tag {4}\n$$\n", + "text_format": "latex", + "bbox": [ + 99, + 484, + 480, + 500 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "here $\\mathcal{L}_{\\text{planning}}$ comes from the planning objective used to train [18, 20]. Here, the loss weights are chosen to bring each value to the same scale.", + "bbox": [ + 89, + 503, + 480, + 547 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "4. Experimental Setup", + "text_level": 1, + "bbox": [ + 89, + 561, + 282, + 580 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "In Section 4.1, we present the architecture design details, in Section 4.2 we present the training strategy, and in Section 4.3, we provide information about the datasets used. Importantly, in Section 4.4 we detail multiple evaluation strategies for fair and comprehensive comparison with existing works.", + "bbox": [ + 89, + 584, + 480, + 660 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "4.1. Architecture design", + "text_level": 1, + "bbox": [ + 89, + 670, + 277, + 686 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Vision-based planner We conduct experiments using two vision-based end-to-end planners, VAD [20] and UniAD [18] due to their performance, efficiency, and ability to model the interaction between scene components. We use two model size variants of VAD. Both planners are trained for perception, motion prediction and planning, while UniAD also performs occupancy prediction.", + "bbox": [ + 89, + 688, + 480, + 794 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "MLLM design The MLLM is made up of adapter layers, a language model, and a set of task-specific decoder layers. We project each $BEAM$ token embedding with its own dedicated Q-former adapter layer following [17]. After projection, the $BEAM$ token embeddings and language token embeddings lie in the same embedding space. In order to reduce memory consumption, we limit the agent token se", + "bbox": [ + 89, + 795, + 480, + 900 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "quence length to a fraction of the total number of agents and perform sequence-wise upsampling after input to the language model. The adapter layers are randomly initialized. We use the LLM from LLaVA-v1.5-7B [24] as our language model base. The ego prediction and agent motion prediction task heads are multi-layer-perceptron networks following [20]. For the surrogate task decoder heads, we use 3 Linear layers with a ReLU activation layer.", + "bbox": [ + 511, + 90, + 903, + 212 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "4.2. Training", + "text_level": 1, + "bbox": [ + 511, + 220, + 617, + 237 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "DiMA training follows a two-stage approach. First the vision-only planner is pre-trained for perception, prediction, and planning for 60 epochs in order to learn informative latent scene representations. Second, we perform joint training of the vision planner and the MLLM for an additional 30 epochs, incorporating all proposed tasks and losses detailed in Section 3. In the second stage, the language model of the MLLM is fine-tuned using LoRA [15]. Please refer to the supplementary material for more details.", + "bbox": [ + 511, + 239, + 903, + 376 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "4.3. Datasets", + "text_level": 1, + "bbox": [ + 511, + 386, + 614, + 398 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "We use the nuScenes dataset for the task of open-loop planning [2]. This dataset consists of 28k total samples in a $22\\mathrm{k} / 6\\mathrm{k}$ training/Validation split. The objects in each scene are annotated with 3D bounding box, orientation, and vehicle speed information. Additionally, we make use of the CAN bus information for the ego vehicle's ground-truth trajectory annotations. For supervising visual question answering, we train with DriveLM [37]. This dataset consists of a 4k subset of samples from the nuScenes dataset annotated with 300k QA pairs related to perception, prediction, planning, and behavior of the ego vehicle. For the samples in nuScenes that do not have text annotations, we create a set of perception, planning, and prediction QA pairs based on the numerical annotations. We use this information to prompt a Llama3-70B model [12] with in-context examples to generate DriveLM-like QA pairs. For the behavior QA pair, we create a rule-based algorithm to categorize the future movement of the ego-vehicle based on the ground-truth trajectory values and ego vehicle speed. More details on this method as well as examples of generated QA pairs may be found in the supplementary.", + "bbox": [ + 511, + 404, + 906, + 722 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "4.4. Evaluation details", + "text_level": 1, + "bbox": [ + 511, + 731, + 687, + 744 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "We evaluate planning performance based on the predicted future trajectories of the ego vehicle 3 seconds into the future, where 2 waypoints are predicted per second. We use two metrics, the L2 error (in meters) between the predicted and ground-truth waypoints as well as the collision rate (in $\\%$ ) between the ego-vehicle and the surrounding vehicles.", + "bbox": [ + 511, + 750, + 903, + 839 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Standardized vs VAD evaluation. In [48], Weng et al. point out inconsistencies in planning evaluation across VAD [20], UniAD [18] and AD-MLP [54], namely in the way L2 error is averaged over time and the way noisy and invalid", + "bbox": [ + 511, + 840, + 903, + 900 + ], + "page_idx": 4 + }, + { + "type": "page_number", + "text": "27579", + "bbox": [ + 478, + 944, + 517, + 955 + ], + "page_idx": 4 + }, + { + "type": "table", + "img_path": "images/aa438d4682570fbb82e68eb93a908d74aa1db5be526032b4ac2a420657d41809.jpg", + "table_caption": [ + "Table 1. Comparison of L2 trajectory error and collision rate on nuScenes [2] using standardized evaluation [48]. Models are evaluated on the general validation split as well as a \"targeted\" split of challenging samples from [48]. The performance of the DiMA model variants are in shades of purple. We summarize results by averaging over at $t = \\{1,2,3\\} s$ as well as at all time steps. The best average performance in each setting is in bold. \"+\" indicates the use of ego-status information." + ], + "table_footnote": [], + "table_body": "
MethodUsing Ego statusTraj L2 (m) ↓Collision (%) ↓
1s2s3sAve1,2,3sAve all1s2s3sAve1,2,3sAve all
Full validation split
UniAD[18]X0.480.891.470.950.830.320.290.730.450.40
VAD-Base [20]X0.410.861.460.910.780.020.260.830.370.30
PARA-Drive[48]X0.260.591.120.660.560.000.120.650.260.17
TOKEN[38]X0.260.701.460.810.68----0.15
DiMA (VAD-Tiny)X0.200.531.100.610.510.000.070.190.090.06
DiMA (VAD-Base)X0.180.501.020.570.470.000.050.160.070.06
AD-MLP[54]✓0.230.581.180.660.560.000.140.700.280.20
PARA-Drive+ [48]✓0.200.521.040.590.490.000.090.490.190.13
DiMA+ (VAD-Tiny)✓0.190.511.060.590.490.000.060.160.080.06
DiMA+ (VAD-Base)✓0.180.481.010.560.460.000.050.160.070.06
DiMA-Dual+ (VAD-Tiny)✓0.180.501.030.570.470.000.050.160.080.06
Targeted validation split
UniAD[18]X0.471.091.921.160.990.000.000.730.240.15
VAD-Base [20]X0.531.202.071.271.080.000.290.870.390.34
PARA-Drive[48]X0.380.971.881.080.910.000.000.720.240.14
DiMA (VAD-Tiny)X0.320.881.700.970.810.000.030.240.090.05
DiMA (VAD-Base)X0.290.791.490.830.710.000.030.220.080.05
AD-MLP[54]✓0.330.992.061.130.940.000.583.621.400.94
PARA-Drive+ [48]✓0.290.761.460.830.700.000.000.290.100.05
DiMA+ (VAD-Tiny)✓0.310.831.670.940.770.000.040.170.070.05
DiMA+ (VAD-Base)✓0.270.741.380.790.640.000.030.160.060.04
DiMA-Dual+ (VAD-Tiny)✓0.280.761.490.840.690.000.030.170.070.05
", + "bbox": [ + 96, + 148, + 901, + 510 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "frames are considered. The authors also point out that collision performance is considerably improved by using a more finely discretized BEV map. Accounting for these inconsistencies, they propose a \"standardized\" evaluation metric to maintain a fair comparison across methods. For accurate reporting and fair comparison with [38, 48], we use this standardized metric to evaluate our model. Certain existing works use the evaluation scheme followed by [20] (which we call VAD evaluation) and do not make code or models available [16, 30, 39, 43]. Here, we compare reported results against DiMA evaluated using VAD evaluation.", + "bbox": [ + 88, + 523, + 480, + 690 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Validation splits. We evaluate planning performance across 3 types of splits from the nuScenes validation set: 1) the entire validation set of 6019 samples 2) a 689 sample sub-set more difficult \"targeted\" scenarios in which the ego-vehicle turns during navigation [48], and 3) long-tail events in which the ego-vehicle encounters rare maneuvers [38]. u", + "bbox": [ + 89, + 691, + 482, + 784 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "5. Experimental Results", + "text_level": 1, + "bbox": [ + 89, + 801, + 294, + 819 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "We perform a comprehensive evaluation of DiMA on the nuScenes open-loop planning benchmark in a variety of evaluation settings. We compare our performance against that of recent vision-based planners [16, 18, 20, 48] as well as recent MLLM planners [30, 38, 39, 43]. To facilitate a", + "bbox": [ + 89, + 824, + 483, + 901 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "fair comparison of DiMA with recent works, we use two evaluation strategies VAD evaluation [20] and standardized evaluation [48]. Please refer to Section 4.4 for more details.", + "bbox": [ + 511, + 523, + 906, + 570 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Standardized evaluation. Under the standardized evaluation strategy, we evaluate each method on two validation splits: 1) the full nuScenes validation split, and 2) a \"targeted\" subset of the nuScenes validation split that contains scenarios where the ego-vehicle must take a turn, as established in [48]. We present our results in Table 1. The baseline vision planners VAD [20] and UniAD [18] exhibit suboptimal performance in more difficult targeted scenarios. DiMA demonstrates significant improvements over baseline methods, achieving a $35\\%$ reduction in L2 trajectory error and an $80\\%$ reduction in collision rate on the overall nuScenes validation set, and a $35\\%$ reduction in L2 trajectory error in the \"targeted\" split. More notably, we outperform recent state-of-the-art methods PARA-Drive [48] and TOKEN [38] consistently across both validation splits. We observe a strong improvement in collision rate, improving on that of TOKEN by $60\\%$ . Unlike TOKEN, our approach does not rely on an LLM for inference, thus making DiMA more efficient as well as more accurate.", + "bbox": [ + 511, + 575, + 908, + 864 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "VAD evaluation. For a fair comparison with recent works that use the evaluation strategy from VAD [20], we present", + "bbox": [ + 511, + 869, + 908, + 902 + ], + "page_idx": 5 + }, + { + "type": "page_number", + "text": "27580", + "bbox": [ + 478, + 944, + 519, + 957 + ], + "page_idx": 5 + }, + { + "type": "table", + "img_path": "images/b321c973910577f8ef5bda621c926d0d098e3f6e53bc16f7c593a1f3c3dd4d58.jpg", + "table_caption": [ + "Table 2. Comparison of L2 trajectory error and collision rate on nuScenes [2] using VAD evaluation [20]. Models are evaluated on the general validation split. The performance of the DiMA model variants are in shades of purple. The performance of the MLLM-branch DiMA model variant is in blue. We summarize results by averaging over all time steps. The best average performance in each setting is in bold. “+” indicates the use of ego-status information." + ], + "table_footnote": [], + "table_body": "
MethodUsing Ego statusTraj L2 (m) ↓Collision (%) ↓Latency (ms)FPS
1s2s3sAveall1s2s3sAveall
ST-P3 [16]X1.332.112.902.110.230.621.270.71628.31.6
UniAD[18]X0.480.961.651.030.050.170.710.31555.61.8
VAD-Tiny[20]X0.460.761.120.780.210.350.580.3859.516.8
VAD-Base[20]X0.410.701.050.720.070.170.410.22224.34.5
VLP-VAD-Base[30]X0.260.470.780.500.120.170.420.23--
OmniDrive [43]X0.400.801.320.840.040.462.320.94--
DiMA (UniAD)X0.150.300.560.340.060.080.220.125601.8
DiMA (VAD-Tiny)X0.180.360.610.380.070.100.270.1559.516.8
DiMA (VAD-Base)X0.130.270.470.290.050.080.160.102264.5
VAD-Tiny✓0.200.380.650.410.100.120.270.1659.516.8
VAD-Base✓0.170.340.600.370.070.100.240.14224.34.5
DriveVLM [39]✓0.180.340.680.400.100.220.450.27--
DriveVLM-Dual (VAD-Base) [39]✓0.150.290.480.310.050.080.170.10--
DiMA + (VAD-Tiny)✓0.170.330.590.370.070.100.260.1459.516.8
DiMA + (VAD-Base)✓0.120.250.440.270.040.060.150.082264.5
DiMA-Dual+ (VAD-Tiny)✓0.140.270.460.290.050.070.150.092863.5
", + "bbox": [ + 94, + 148, + 906, + 398 + ], + "page_idx": 6 + }, + { + "type": "table", + "img_path": "images/61252b0aebe23b3881bc95b42e66a0477f5c7547722e14a186ba10ace03e0478.jpg", + "table_caption": [ + "Table 3. Long-tail performance comparison of L2 trajectory error and collision rate on nuScenes [2] validation set using standardized evaluation [48]. The performance of vision planner DiMA model variants are in shades of purple. The performance of MLLM-branch DiMA model variants are in shades of blue. We summarize results by averaging over at $t = \\{1,2,3\\} s$ as well as at all time steps. The best average performance in each setting is in bold." + ], + "table_footnote": [], + "table_body": "
MethodTraj L2 (m)↓Coll. (%)↓
1s2s3sAve1,2,3sAveallAveall
3-point turn (zero-shot)
VAD-Tiny0.761.683.041.831.550.00
VAD-Base0.711.663.241.871.570.00
PARA-Drive [48]0.501.382.761.551.295.33
TOKEN [38]0.391.292.601.431.184.00
DiMA (VAD-Tiny)0.471.272.551.431.170.00
DiMA (VAD-Base)0.361.182.371.301.050.00
DiMA (MLLM)0.381.122.351.281.040.00
DiMA-Dual (VAD-Tiny)0.381.082.351.271.040.00
Resume from stop
VAD-Tiny0.641.632.991.751.490.00
VAD-Base0.601.722.831.721.420.00
PARA-Drive0.140.792.301.080.850.00
TOKEN0.130.701.580.800.650.00
DiMA (VAD-Tiny)0.261.022.401.230.990.00
DiMA (VAD-Base)0.150.651.340.710.660.00
DiMA (MLLM)0.241.022.161.140.930.00
DiMA-Dual (VAD-Tiny)0.240.982.131.110.910.00
Overtake
VAD-Tiny0.581.272.121.321.142.42
VAD-Base0.461.162.171.261.062.49
PARA-Drive0.270.891.941.030.852.30
TOKEN0.290.771.630.900.740.00
DiMA (VAD-Tiny)0.240.751.490.830.691.32
DiMA (VAD-Base)0.240.721.500.820.661.29
DiMA (MLLM)0.240.731.500.820.671.30
DiMA-Dual (VAD-Tiny)0.240.731.500.820.671.30
", + "bbox": [ + 94, + 522, + 478, + 824 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "an additional set of results under this strategy. In Table 2, we present the evaluation of DiMA on the nuScenes validation set [2]. We significantly outperform the base", + "bbox": [ + 89, + 830, + 482, + 877 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "line vision-based planners VAD [20] and UniAD [18]. Notably, DiMA built on VAD-Tiny outperforms VAD-Base by $47\\%$ in L2 trajectory error while being 4 times faster. We also outperform MLLM planners like OmniDrive [43] and DriveVLM [39] without requiring an LLM at inference, making our approach more accurate as well as efficient.", + "bbox": [ + 511, + 422, + 906, + 513 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Performance in long-tail scenarios. Autonomous driving performance is particularly crucial for rare scenarios. Using the manually selected long-tail events from [38], we demonstrate the robustness of DiMA in cases of novel navigation and perception. In Table 3, we evaluate the performance of DiMA for three long-tail scenarios, \"3-point turn\", \"resume from the stop\" and \"overtake\", selected from the nuScenes validation set. Training with DiMA results in a significant and consistent boost in performance from the baseline vision-based planners. We also consistently outperform PARA-Drive and TOKEN. Notably, the \"3-point turn\" case is a zero-shot scenario, not present in the training data. We achieve the lowest L2 trajectory error for this case.", + "bbox": [ + 511, + 518, + 908, + 715 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Evaluating the MLLM branch. We also evaluate the MLLM branch of DiMA and demonstrate its planning and visual question-answering capabilities. We perform LLM-based planning using DiMA in two ways. First, we directly evaluate the waypoint predictions of the MLLM planning decoder head. This is denoted by DiMA (MLLM). Secondly, we define DiMA-Dual, an optional hybrid inference model that combines the strengths of both the vision and MLLM branches. We perform max-pooling on the penultimate layer features of both vision planning transformer and MLLM ego prediction. The resulting pooled features are then fed back into both the vision planning transformer and", + "bbox": [ + 511, + 719, + 910, + 900 + ], + "page_idx": 6 + }, + { + "type": "page_footnote", + "text": "1We compare with reported results due to lack of published code.", + "bbox": [ + 107, + 886, + 454, + 900 + ], + "page_idx": 6 + }, + { + "type": "page_number", + "text": "27581", + "bbox": [ + 478, + 944, + 517, + 955 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/50f3c2512a4f9267ad7bb8e17c10129bd313a222693b89a5732a20458a67adf3.jpg", + "image_caption": [ + "Figure 4. Visualization of planning performance by the MLLM branch of DiMA-VAD-Tiny. We plot the predicted trajectory (orange-yellow) and show an example response of the language model branch to a question from the DriveLM test dataset [37]." + ], + "image_footnote": [], + "bbox": [ + 93, + 88, + 491, + 277 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/f324fc1565c9d9a416294862fac9f1ca3576c60912344d459c0d2fcf234c3f9a.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 501, + 88, + 900, + 277 + ], + "page_idx": 7 + }, + { + "type": "table", + "img_path": "images/0acd665cf16619f1618238c8519132aaae489cc67036bf64172ad7487cc9b607.jpg", + "table_caption": [ + "Table 4. Results of ablation experiments of DiMA-VAD-Tiny on nuScenes under VAD evaluation. The baseline model is VAD-Tiny. We report L2 trajectory error and collision rate averaged over all time steps." + ], + "table_footnote": [], + "table_body": "
IDVQAScene tokensDistillationLLM PlanningSurrogate tasksTraj L2 (m) ↓Collision (%) ↓
Masked recon.Future pred.Scene editing1s2s3sAveall1s2s3sAveall
1XXXXXXX0.330.580.890.600.120.190.550.29
2✓BEVX✓XXX0.360.600.910.620.100.170.520.26
3✓BEV, MapX✓XXX0.280.520.870.560.100.170.490.25
4✓AllX✓XXX0.260.510.830.520.100.160.380.21
5✓All✓✓XXX0.220.440.770.480.090.140.340.19
6✓All✓✓✓XX0.190.390.680.420.090.130.320.18
7✓All✓✓✓✓X0.180.370.630.390.070.110.290.16
8✓All✓✓✓✓✓0.180.360.610.380.070.100.270.15
", + "bbox": [ + 94, + 358, + 890, + 500 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "MLLM ego prediction for waypoint prediction. This max pooling encourages consistency between the predictions of the two models, leading to a more robust and potentially improved overall prediction as seen in Table 3. Using the MLLM at inference results in a significant boost in performance, with DiMA-Dual (VAD-Tiny) matching the performance of DiMA (VAD-Base) in Table 2.", + "bbox": [ + 89, + 518, + 482, + 625 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "VQA results. We provide qualitative examples of visual question-answering in the DriveLM dataset [37] in Figure 4. We demonstrate the ability of DiMA to reason about safe actions to be taken by the ego vehicle. We provide a more examples in the supplementary material.", + "bbox": [ + 89, + 626, + 483, + 702 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "5.1. Ablation study.", + "text_level": 1, + "bbox": [ + 89, + 714, + 243, + 729 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "We examine the role of different design choices and proposed tasks in the DiMA framework through a comprehensive ablation study in Table 4. In order to do so, we build our final frame-work step-by-step. We begin with VAD-Tiny as our baseline model in the first row, denoted by ID-1. In ID-2, we observe inconsistent gains and drops in performance by naively training an LLM with the VAD-Tiny under a the planning and VQA objectives while using only BEV features. By comparing ID-3 and ID-4, we show that that providing more scene context to the MLLM by using all the BEAM token embeddings to train both LLM and", + "bbox": [ + 89, + 734, + 483, + 900 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "vision-based planner benefits the performance, resulting a significant improvement in trajectory estimation and collision rate. We show in ID-5 that explicitly distilling the information from the penultimate layers of the MLLM to the planning transformer gives another boost in performance. In ID-6, ID-7 and ID-8 we demonstrate the importance of surrogate tasks like masked-reconstruction, future prediction and scene editing in helping the scene encoder to learn richer representations useful for planning. We point to the supplementary material for additional ablation studies.", + "bbox": [ + 511, + 518, + 906, + 671 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "6. Conclusion", + "text_level": 1, + "bbox": [ + 513, + 694, + 633, + 709 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "We introduce DiMA, an end-to-end autonomous driving framework for robust and efficient planning. DiMA distills knowledge from an MLLM to a vision-based planner, employing a novel joint training strategy alongside carefully designed surrogate tasks such as masked token reconstruction, future token prediction, and scene editing. DiMA enables the model to learn semantically grounded scene representations, improving its performance in difficult navigation scenarios. We conduct extensive evaluations and ablation studies, showing that training with DiMA results in a planner that excels in long-tail cases while maintaining computational efficiency.", + "bbox": [ + 511, + 719, + 906, + 900 + ], + "page_idx": 7 + }, + { + "type": "page_number", + "text": "27582", + "bbox": [ + 478, + 944, + 519, + 955 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "References", + "text_level": 1, + "bbox": [ + 91, + 89, + 187, + 104 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[1] Apratim Bhattacharyya, Sunny Panchal, Reza Pourreza, Mingu Lee, Pulkit Madan, and Roland Memisevic. Look, remember and reason: Grounded reasoning in videos with language models. In International Conference of Learning Representations, 2024. 4", + "[2] Holger Caesar, Varun Bankiti, Alex H Lang, Sourabh Vora, Venice Erin Liong, Qiang Xu, Anush Krishnan, Yu Pan, Giancarlo Baldan, and Oscar Beijbom. nuscenes: A multimodal dataset for autonomous driving. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 11621-11631, 2020. 1, 5, 6, 7", + "[3] Sergio Casas, Abbas Sadat, and Raquel Urtasun. Mp3: A unified model to map, perceive, predict and plan. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 14403-14412, 2021. 1, 2", + "[4] Yuning Chai, Benjamin Sapp, Mayank Bansal, and Dragomir Anguelov. Multipath: Multiple probabilistic anchor trajectory hypotheses for behavior prediction. arXiv preprint arXiv:1910.05449, 2019. 1, 2", + "[5] Dian Chen and Philipp Krahenbuhl. Learning from all vehicles. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 17222-17231, 2022. 1, 2", + "[6] Long Chen, Oleg Sinavski, Jan Hünermann, Alice Karnsund, Andrew James Willmott, Danny Birch, Daniel Maund, and Jamie Shotton. Driving with llms: Fusing object-level vector modality for explainable autonomous driving. In 2024 IEEE International Conference on Robotics and Automation (ICRA), pages 14093-14100. IEEE, 2024. 2", + "[7] Kashyap Chitta, Aditya Prakash, and Andreas Geiger. Neat: Neural attention fields for end-to-end autonomous driving. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 15793-15803, 2021. 1, 2", + "[8] Kashyap Chitta, Aditya Prakash, Bernhard Jaeger, Zehao Yu, Katrin Renz, and Andreas Geiger. Transfuser: Imitation with transformer-based sensor fusion for autonomous driving. IEEE Transactions on Pattern Analysis and Machine Intelligence, 45(11):12878-12895, 2022. 2", + "[9] Felipe Codevilla, Eder Santana, Antonio M López, and Adrien Gaidon. Exploring the limitations of behavior cloning for autonomous driving. In Proceedings of the IEEE/CVF international conference on computer vision, pages 9329-9338, 2019. 2", + "[10] Xinpeng Ding, Jianhua Han, Hang Xu, Wei Zhang, and Xiaomeng Li. Hilm-d: Towards high-resolution understanding in multimodal large language models for autonomous driving. arXiv preprint arXiv:2309.05186, 2023. 2", + "[11] Alexey Dosovitskiy, German Ros, Felipe Codevilla, Antonio Lopez, and Vladlen Koltun. Carla: An open urban driving simulator. In Conference on robot learning, pages 1-16. PMLR, 2017. 1", + "[12] Abhimanyu Dubey, Abhinav Jauhri, Abhinav Pandey, Abhishek Kadian, Ahmad Al-Dahle, Aiesha Letman, Akhil Mathur, Alan Schelten, Amy Yang, Angela Fan, et al. The llama 3 herd of models. arXiv preprint arXiv:2407.21783, 2024. 2, 5" + ], + "bbox": [ + 93, + 114, + 483, + 898 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[13] Scott Ettinger, Shuyang Cheng, Benjamin Caine, Chenxi Liu, Hang Zhao, Sabeek Pradhan, Yuning Chai, Ben Sapp, Charles R Qi, Yin Zhou, et al. Large scale interactive motion forecasting for autonomous driving: The waymo open motion dataset. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 9710-9719, 2021. 1", + "[14] Daocheng Fu, Xin Li, Licheng Wen, Min Dou, Pinlong Cai, Botian Shi, and Yu Qiao. Drive like a human: Rethinking autonomous driving with large language models. In Proceedings of the IEEE/CVF Winter Conference on Applications of Computer Vision, pages 910-919, 2024. 2", + "[15] Edward J Hu, Yelong Shen, Phillip Wallis, Zeyuan Allen-Zhu, Yuanzhi Li, Shean Wang, Lu Wang, and Weizhu Chen. Lora: Low-rank adaptation of large language models. arXiv preprint arXiv:2106.09685, 2021. 5", + "[16] Shengchao Hu, Li Chen, Penghao Wu, Hongyang Li, Junchi Yan, and Dacheng Tao. St-p3: End-to-end vision-based autonomous driving via spatial-temporal feature learning. In European Conference on Computer Vision, pages 533-549. Springer, 2022. 1, 2, 6, 7", + "[17] Wenbo Hu, Zi-Yi Dou, Liunian Harold Li, Anita Kamath, Nanyun Peng, and Kai-Wei Chang. Matryoshka query transformer for large vision-language models, 2024. 4, 5", + "[18] Yihan Hu, Jiazhi Yang, Li Chen, Keyu Li, Chonghao Sima, Xizhou Zhu, Siqi Chai, Senyao Du, Tianwei Lin, Wenhai Wang, et al. Planning-oriented autonomous driving. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 17853-17862, 2023. 1, 2, 3, 5, 6, 7", + "[19] Jyh-Jing Hwang, Runsheng Xu, Hubert Lin, Wei-Chih Hung, Jingwei Ji, Kristy Choi, Di Huang, Tong He, Paul Covington, Benjamin Sapp, et al. Emma: End-to-end multimodal model for autonomous driving. arXiv preprint arXiv:2410.23262, 2024. 1, 3", + "[20] Bo Jiang, Shaoyu Chen, Qing Xu, Bencheng Liao, Jiajie Chen, Helong Zhou, Qian Zhang, Wenyu Liu, Chang Huang, and Xinggang Wang. Vad: Vectorized scene representation for efficient autonomous driving. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 8340-8350, 2023. 1, 2, 3, 5, 6, 7", + "[21] Junnan Li, Dongxu Li, Silvio Savarese, and Steven Hoi. Blip-2: Bootstrapping language-image pre-training with frozen image encoders and large language models. In International conference on machine learning, pages 19730-19742. PMLR, 2023. 1, 2, 4", + "[22] Zhiqi Li, Wenhai Wang, Hongyang Li, Enze Xie, Chonghao Sima, Tong Lu, Yu Qiao, and Jifeng Dai. Bevformer: Learning bird's-eye-view representation from multi-camera images via spatiotemporal transformers. In European conference on computer vision, pages 1-18. Springer, 2022. 1, 2, 3", + "[23] Xiwen Liang, Yangxin Wu, Jianhua Han, Hang Xu, Chunjing Xu, and Xiaodan Liang. Effective adaptation in multitask co-training for unified autonomous driving. Advances in Neural Information Processing Systems, 35:19645-19658, 2022. 2", + "[24] Haotian Liu, Chunyuan Li, Qingyang Wu, and Yong Jae Lee." + ], + "bbox": [ + 516, + 92, + 906, + 901 + ], + "page_idx": 8 + }, + { + "type": "page_number", + "text": "27583", + "bbox": [ + 478, + 944, + 517, + 955 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "Visual instruction tuning. Advances in neural information processing systems, 36, 2024. 1, 2, 5", + "[25] Jiaqi Liu, Peng Hang, Xiao Qi, Jianqiang Wang, and Jian Sun. Mtd-gpt: A multi-task decision-making gpt model for autonomous driving at unsignalized intersections. In 2023 IEEE 26th International Conference on Intelligent Transportation Systems (ITSC), pages 5154-5161. IEEE, 2023. 2", + "[26] Yicheng Liu, Jinghuai Zhang, Liangji Fang, Qinhong Jiang, and Bolei Zhou. Multimodal motion prediction with stacked transformers. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 7577-7586, 2021. 1, 2", + "[27] Srikanth Malla, Chiho Choi, Isht Dwivedi, Joon Hee Choi, and Jiachen Li. Drama: Joint risk localization and captioning in driving. In Proceedings of the IEEE/CVF winter conference on applications of computer vision, pages 1043-1052, 2023. 2", + "[28] Jiageng Mao, Yuxi Qian, Junjie Ye, Hang Zhao, and Yue Wang. Gpt-driver: Learning to drive with gpt. arXiv preprint arXiv:2310.01415, 2023. 2", + "[29] Jiquan Ngiam, Benjamin Caine, Vijay Vasudevan, Zhengdong Zhang, Hao-Tien Lewis Chiang, Jeffrey Ling, Rebecca Roelofs, Alex Bewley, Chenxi Liu, Ashish Venugopal, et al. Scene transformer: A unified architecture for predicting multiple agent trajectories. arXiv preprint arXiv:2106.08417, 2021. 1, 2", + "[30] Chenbin Pan, Burhaneddin Yaman, Tommaso Nesti, Abhirup Mallik, Alessandro G Allievi, Senem Velipasalar, and Liu Ren. Vlp: Vision language planning for autonomous driving. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 14760-14769, 2024. 6, 7", + "[31] Jonah Philion and Sanja Fidler. Lift, splat, shoot: Encoding images from arbitrary camera rigs by implicitly unprojecting to 3d. In Computer Vision-ECCV 2020: 16th European Conference, Glasgow, UK, August 23-28, 2020, Proceedings, Part XIV 16, pages 194-210. Springer, 2020. 1, 2", + "[32] Tianwen Qian, Jingjing Chen, Linhai Zhuo, Yang Jiao, and Yu-Gang Jiang. Nuscenes-qa: A multi-modal visual question answering benchmark for autonomous driving scenario. In Proceedings of the AAAI Conference on Artificial Intelligence, pages 4542–4550, 2024. 2", + "[33] Alec Radford, Jong Wook Kim, Chris Hallacy, Aditya Ramesh, Gabriel Goh, Sandhini Agarwal, Girish Sastry, Amanda Askell, Pamela Mishkin, Jack Clark, et al. Learning transferable visual models from natural language supervision. In International conference on machine learning, pages 8748-8763. PMLR, 2021. 2", + "[34] Enna Sachdeva, Nakul Agarwal, Suhas Chundi, Sean Roelofs, Jiachen Li, Mykel Kochenderfer, Chiho Choi, and Behzad Dariush. Rank2tell: A multimodal driving dataset for joint importance ranking and reasoning. In Proceedings of the IEEE/CVF Winter Conference on Applications of Computer Vision, pages 7513-7522, 2024. 2", + "[35] Hao Sha, Yao Mu, Yuxuan Jiang, Li Chen, Chenfeng Xu, Ping Luo, Shengbo Eben Li, Masayoshi Tomizuka, Wei" + ], + "bbox": [ + 91, + 92, + 482, + 900 + ], + "page_idx": 9 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "Zhan, and Mingyu Ding. *Languagempc: Large language models as decision makers for autonomous driving.* arXiv preprint arXiv:2310.03026, 2023. 2", + "[36] Hao Shao, Yuxuan Hu, Letian Wang, Guanglu Song, Steven L Waslander, Yu Liu, and Hongsheng Li. Lmdrive: Closed-loop end-to-end driving with large language models. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 15120-15130, 2024. 2, 3", + "[37] Chonghao Sima, Katrin Renz, Kashyap Chitta, Li Chen, Hanxue Zhang, Chengen Xie, Ping Luo, Andreas Geiger, and Hongyang Li. Drivelm: Driving with graph visual question answering. arXiv preprint arXiv:2312.14150, 2023. 1, 2, 4, 5, 8", + "[38] Ran Tian, Boyi Li, Xinshuo Weng, Yuxiao Chen, Edward Schmerling, Yue Wang, Boris Ivanovic, and Marco Pavone. Tokenize the world into object-level knowledge to address long-tail events in autonomous driving. arXiv preprint arXiv:2407.00959, 2024. 1, 2, 3, 6, 7", + "[39] Xiaoyu Tian, Junru Gu, Bailin Li, Yicheng Liu, Chenxu Hu, Yang Wang, Kun Zhan, Peng Jia, Xianpeng Lang, and Hang Zhao. Drivevlm: The convergence of autonomous driving and large vision-language models. arXiv preprint arXiv:2402.12289, 2024. 1, 2, 3, 6, 7", + "[40] Wenwen Tong, Chonghao Sima, Tai Wang, Li Chen, Silei Wu, Hanming Deng, Yi Gu, Lewei Lu, Ping Luo, Dahua Lin, et al. Scene as occupancy. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 8406-8415, 2023. 3", + "[41] Hugo Touvron, Thibaut Lavril, Gautier Izacard, Xavier Martinet, Marie-Anne Lachaux, Timothée Lacroix, Baptiste Rozière, Naman Goyal, Eric Hambro, Faisal Azhar, et al. Llama: Open and efficient foundation language models. arXiv preprint arXiv:2302.13971, 2023. 2", + "[42] Hugo Touvron, Louis Martin, Kevin Stone, Peter Albert, Amjad Almahairi, Yasmine Babaei, Nikolay Bashlykov, Soumya Batra, Prajjwal Bhargava, Shruti Bhosale, et al. Llama 2: Open foundation and fine-tuned chat models. arXiv preprint arXiv:2307.09288, 2023. 2", + "[43] Shihao Wang, Zhiding Yu, Xiaohui Jiang, Shiyi Lan, Min Shi, Nadine Chang, Jan Kautz, Ying Li, and Jose M Alvarez. Omnidrive: A holistic llm-agent framework for autonomous driving with 3d perception, reasoning and planning. arXiv preprint arXiv:2405.01533, 2024. 1, 3, 6, 7", + "[44] Wenhai Wang, Jiangwei Xie, ChuanYang Hu, Haoming Zou, Jianan Fan, Wenwen Tong, Yang Wen, Silei Wu, Hanming Deng, Zhiqi Li, et al. Driveml: Aligning multi-modal large language models with behavioral planning states for autonomous driving. arXiv preprint arXiv:2312.09245, 2023. 1, 2, 3", + "[45] Zitian Wang, Zehao Huang, Yulu Gao, Naiyan Wang, and Si Liu. Mv2dfusion: Leveraging modality-specific object semantics for multi-modal 3d detection. arXiv preprint arXiv:2408.05945, 2024. 1, 2", + "[46] Jason Wei, Xuezhi Wang, Dale Schuurmans, Maarten Bosma, Fei Xia, Ed Chi, Quoc V Le, Denny Zhou, et al. Chain-of-thought prompting elicits reasoning in large lan" + ], + "bbox": [ + 516, + 92, + 903, + 900 + ], + "page_idx": 9 + }, + { + "type": "page_number", + "text": "27584", + "bbox": [ + 478, + 944, + 519, + 955 + ], + "page_idx": 9 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "guage models. Advances in neural information processing systems, 35:24824-24837, 2022. 1, 2", + "[47] Licheng Wen, Daocheng Fu, Xin Li, Xinyu Cai, MA Tao, Pinlong Cai, Min Dou, Botian Shi, Liang He, and Yu Qiao. Dilu: A knowledge-driven approach to autonomous driving with large language models. In The Twelfth International Conference on Learning Representations. 2", + "[48] Xinshuo Weng, Boris Ivanovic, Yan Wang, Yue Wang, and Marco Pavone. Para-drive: Parallelized architecture for real-time autonomous driving. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 15449-15458, 2024. 1, 2, 3, 5, 6, 7", + "[49] Penghao Wu, Xiaosong Jia, Li Chen, Junchi Yan, Hongyang Li, and Yu Qiao. Trajectory-guided control prediction for end-to-end autonomous driving: A simple yet strong baseline. Advances in Neural Information Processing Systems, 35:6119-6132, 2022. 1, 2", + "[50] Zhenhua Xu, Yujia Zhang, Enze Xie, Zhen Zhao, Yong Guo, Kwan-Yee K Wong, Zhenguo Li, and Hengshuang Zhao. Drivegpt4: Interpretable end-to-end autonomous driving via large language model. IEEE Robotics and Automation Letters, 2024. 2", + "[51] Rajeev Yasarla, Manish Kumar Singh, Hong Cai, Yunxiao Shi, Jisoo Jeong, Yinhao Zhu, Shizhong Han, Risheek Garrepalli, and Fatih Porikli. *Futuredepth: Learning to predict the future improves video depth estimation. In European Conference on Computer Vision*, pages 440–458. Springer, 2024. 4", + "[52] Tianwei Yin, Xingyi Zhou, and Philipp Krahenbuhl. Center-based 3d object detection and tracking. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 11784-11793, 2021. 1, 2", + "[53] Wenyuan Zeng, Wenjie Luo, Simon Suo, Abbas Sadat, Bin Yang, Sergio Casas, and Raquel Urtasun. End-to-end interpretable neural motion planner. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 8660-8669, 2019. 2", + "[54] Jiang-Tian Zhai, Ze Feng, Jinhao Du, Yongqiang Mao, Jiang-Jiang Liu, Zichang Tan, Yifu Zhang, Xiaqing Ye, and Jingdong Wang. Rethinking the open-loop evaluation of end-to-end autonomous driving in nuscenes. arXiv preprint arXiv:2305.10430, 2023. 5, 6", + "[55] Yunpeng Zhang, Zheng Zhu, Wenzhao Zheng, Junjie Huang, Guan Huang, Jie Zhou, and Jiwen Lu. Beverse: Unified perception and prediction in birds-eye-view for vision-centric autonomous driving. arXiv preprint arXiv:2205.09743, 2022. 2" + ], + "bbox": [ + 91, + 92, + 482, + 752 + ], + "page_idx": 10 + }, + { + "type": "page_number", + "text": "27585", + "bbox": [ + 478, + 944, + 517, + 955 + ], + "page_idx": 10 + } +] \ No newline at end of file diff --git a/2025/Distilling Multi-modal Large Language Models for Autonomous Driving/2bea0a6d-f2ee-4f71-9581-707f2552ee73_model.json b/2025/Distilling Multi-modal Large Language Models for Autonomous Driving/2bea0a6d-f2ee-4f71-9581-707f2552ee73_model.json new file mode 100644 index 0000000000000000000000000000000000000000..4f35c4d49319f7114aa84b214faf91fd6851b930 --- /dev/null +++ b/2025/Distilling Multi-modal Large Language Models for Autonomous Driving/2bea0a6d-f2ee-4f71-9581-707f2552ee73_model.json @@ -0,0 +1,2037 @@ +[ + [ + { + "type": "header", + "bbox": [ + 0.107, + 0.003, + 0.182, + 0.044 + ], + "angle": 0, + "content": "CVF" + }, + { + "type": "header", + "bbox": [ + 0.237, + 0.001, + 0.812, + 0.047 + ], + "angle": 0, + "content": "This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore." + }, + { + "type": "title", + "bbox": [ + 0.136, + 0.131, + 0.865, + 0.154 + ], + "angle": 0, + "content": "Distilling Multi-modal Large Language Models for Autonomous Driving" + }, + { + "type": "text", + "bbox": [ + 0.148, + 0.18, + 0.851, + 0.217 + ], + "angle": 0, + "content": "Deepti Hegde\\(^{1\\dagger*}\\), Rajeev Yasarla\\(^{2*}\\), Hong Cai\\(^{2}\\), Shizhong Han\\(^{2}\\), Apratim Bhattacharyya\\(^{2}\\), Shweta Mahajan\\(^{2}\\), Litian Liu\\(^{2}\\), Risheek Garrepalli\\(^{2}\\), Vishal M. Patel\\(^{1}\\), Fatih Porikli\\(^{2}\\)" + }, + { + "type": "text", + "bbox": [ + 0.28, + 0.222, + 0.719, + 0.259 + ], + "angle": 0, + "content": "\\(^{1}\\) Johns Hopkins University, \\(^{2}\\) Qualcomm AI Research‡ *Equal contribution" + }, + { + "type": "title", + "bbox": [ + 0.249, + 0.294, + 0.327, + 0.31 + ], + "angle": 0, + "content": "Abstract" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.326, + 0.485, + 0.658 + ], + "angle": 0, + "content": "Autonomous driving demands safe motion planning, especially in critical \"long-tail\" scenarios. Recent end-to-end autonomous driving systems leverage large language models (LLMs) as planners to improve generalizability to rare events. However, using LLMs at test time introduces high computational costs. To address this, we propose DiMA, an end-to-end autonomous driving system that maintains the efficiency of an LLM-free (or vision-based) planner while leveraging the world knowledge of an LLM. DiMA distills the information from a multi-modal LLM to a vision-based end-to-end planner through a set of specially designed surrogate tasks. Under a joint training strategy, a scene encoder common to both networks produces structured representations that are semantically grounded as well as aligned to the final planning objective. Notably, the LLM is optional at inference, enabling robust planning without compromising on efficiency. Training with DiMA results in a \\(37\\%\\) reduction in the L2 trajectory error and an \\(80\\%\\) reduction in the collision rate of the vision-based planner, as well as a \\(44\\%\\) trajectory error reduction in long-tail scenarios. DiMA also achieves state-of-the-art performance on the nuScenes planning benchmark." + }, + { + "type": "title", + "bbox": [ + 0.092, + 0.676, + 0.222, + 0.691 + ], + "angle": 0, + "content": "1. Introduction" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.697, + 0.483, + 0.848 + ], + "angle": 0, + "content": "Research in autonomous driving has moved away from the traditional strategy of integrating sequential, independently trained models [4, 22, 26, 29, 31, 45, 52], and moved towards developing multi-task systems that are trained in an end-to-end manner [3, 5, 7, 16, 20, 49]. These approaches demonstrate improved performance and interpretability while being highly efficient. However, they struggle with long-tail navigation and perception scenarios, primarily because they are trained on task-specific, limited datasets [2, 11, 13]." + }, + { + "type": "text", + "bbox": [ + 0.091, + 0.849, + 0.484, + 0.879 + ], + "angle": 0, + "content": "Large language models (LLMs) have emerged as a promising solution to this issue. Trained on vast, internet-" + }, + { + "type": "image", + "bbox": [ + 0.518, + 0.294, + 0.889, + 0.414 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.518, + 0.416, + 0.889, + 0.535 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.512, + 0.538, + 0.908, + 0.636 + ], + "angle": 0, + "content": "Figure 1. Comparison of planning performance in long-tail scenarios from nuScenes: DiMA-VAD demonstrates greater robustness compared to VAD [20] in long-tail navigation scenarios such as overtaking a vehicle and performing a 3-point turn. DiMA-VAD also outperforms recent vision-planner PARA-Drive [48] and LLM planner TILEN [38]. Notably, the 3-point turn is a zero-shot scenario that is only present in the validation set." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.659, + 0.909, + 0.903 + ], + "angle": 0, + "content": "scale datasets, LLMs can leverage world knowledge to generalize to unseen or rare scenarios. These models can perform high-level reasoning tasks using mechanisms such as chain-of-thought [46]. Going beyond text-based prompting, multi-modal large language models (MLLMs) integrate image and video inputs to the LLM, enabling tasks such as visual question answering and dense captioning [21, 24]. Recent end-to-end autonomous driving systems leverage LLMs to achieve higher interpretability through language-guided reasoning and superior robustness to long-tail scenarios [19, 37-39, 43, 44]. To differentiate between these methods, we call end-to-end planners that depend on LLMs to perform trajectory prediction as \"LLM-based\" planners and ones that do not as \"vision-based\" planners [18, 20, 48]. In spite of their recent success \"LLM-based\" planners face significant challenges." + }, + { + "type": "page_footnote", + "bbox": [ + 0.115, + 0.887, + 0.483, + 0.901 + ], + "angle": 0, + "content": "Qualcomm AI Research is initiative of Qualcomm Technologies, Inc." + }, + { + "type": "page_number", + "bbox": [ + 0.479, + 0.945, + 0.519, + 0.958 + ], + "angle": 0, + "content": "27575" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.089, + 0.092, + 0.482, + 0.318 + ], + "angle": 0, + "content": "Firstly, LLM-based planners require a significant amount of computational overhead at test time, limiting their practicality. This work addresses a core question: how can we harness LLMs' world knowledge while preserving the efficiency of vision-based planners? Secondly, bridging the visual and language domains is inherently more complex for end-to-end planning tasks compared to general MLLM tasks [21, 24], with the additional hurdle of limited training data. Standard image tokenization strategies rely on frozen pre-trained image encoders [33] to generate dense, unstructured multi-modal token embeddings. [21, 24, 37]. We suggest that MLLMs meant for end-to-end autonomous driving benefit from structured inputs that explicitly model scene components. Furthermore, we suggest that updating these features during training results in richer representations." + }, + { + "type": "text", + "bbox": [ + 0.092, + 0.321, + 0.484, + 0.697 + ], + "angle": 0, + "content": "To address these challenges, we propose DiMA, a novel framework for Distilling Multi-modal Large Language Models for Autonomous driving. We introduce a joint-training scheme between a vision-based planner and an MLLM that enables the learning of robust, grounded and disentangled scene representations that are aligned to the final objective of planning. Specifically, we use the vision-based planner as a tokenizer, and pass the learned scene representations to the MLLM, providing a more structured input. The MLLM is then trained for the tasks of visual question-answering, trajectory estimation, and a series of surrogate tasks designed to ground the multi-modal inputs in language. The vision-based planner is simultaneously trained for trajectory estimation, while also distilling features from the MLLM planning head to the planning transformer. Crucially, the MLLM can be discarded when performing planning inference, maintaining the efficiency of the vision-based planner while utilizing the knowledge of the language model. Optionally, the MLLM can support language-guided reasoning through visual question answering. Our experiments show that DiMA is more robust to challenging long-tail scenarios compared to baselines and state-of-the-art methods [38, 48]. Figure 1 confirms this both quantitatively and qualitatively. We list the contributions of the work below:" + }, + { + "type": "text", + "bbox": [ + 0.092, + 0.705, + 0.482, + 0.78 + ], + "angle": 0, + "content": "1. We introduce DiMA, an end-to-end autonomous driving framework that distills knowledge from an MLLM to a vision-based planner to ensure robustness to long-tail events while maintaining efficiency. DiMA is capable of planning as well as visual question-answering." + }, + { + "type": "text", + "bbox": [ + 0.091, + 0.781, + 0.482, + 0.841 + ], + "angle": 0, + "content": "2. We propose a distillation task along with the following surrogate tasks to align the objectives of the vision-based planner and the MLLM: (i) masked token reconstruction (ii) future token prediction (iii) scene editing." + }, + { + "type": "text", + "bbox": [ + 0.091, + 0.841, + 0.482, + 0.901 + ], + "angle": 0, + "content": "3. DiMA outperforms both vision-based and LLM end-to-end planners, achieving state-of-the-art results on the nuScenes planning benchmark. Training with DiMA results in a \\(37\\%\\) reduction in the L2 trajectory error and an" + }, + { + "type": "list", + "bbox": [ + 0.091, + 0.705, + 0.482, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "text", + "bbox": [ + 0.533, + 0.092, + 0.905, + 0.137 + ], + "angle": 0, + "content": "80% reduction in the collision rate of the vision-based planner, as well as a 44% trajectory error reduction in long-tail scenarios." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.152, + 0.655, + 0.167 + ], + "angle": 0, + "content": "2. Related Work" + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.173, + 0.806, + 0.189 + ], + "angle": 0, + "content": "2.1. End-to-End Autonomous Driving" + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.192, + 0.905, + 0.508 + ], + "angle": 0, + "content": "Autonomous driving (AD) systems are generally modular, consisting of tasks such as perception [22, 31, 45, 52], motion prediction [4, 26, 29], and planning [9]. Classic AD systems use standalone models for each task where each module is optimized separately and combined sequentially. Alternatively, some approaches incorporate these modular tasks into a multi-task learning paradigm, where tasks shared a common feature extraction process but have task-specific heads [8, 23, 53, 55]. Recently, research has moved toward optimizing autonomous driving systems in an end-to-end manner, aiming to directly predict future trajectories [3, 5, 7, 16, 20, 49]. Along this line of work, UniAD [18] introduces a query-based design that integrates perception and prediction components, enabling an end-to-end planning scheme. VAD [20] utilizes a vectorized scene representation, replacing the dense rasterized representations from [18], reducing computational cost and improving planning performance. While effective in general navigation scenarios, these methods struggle in difficult long-tail events (see Figure 1). In this work, we propose a method for end-to-end autonomous driving that is robust to such scenarios." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.52, + 0.79, + 0.535 + ], + "angle": 0, + "content": "2.2. LLMs for Autonomous Driving" + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.538, + 0.905, + 0.657 + ], + "angle": 0, + "content": "The recent success of large language models [12, 41, 42] demonstrates their ability to process complex contextual information with logical reasoning and to communicate in a human-interpretable manner, attracting significant attention in autonomous driving [6, 10, 14, 25, 28, 35-37, 39, 44, 47, 50]. These works investigate the capabilities of language models to generalize to novel scenarios as well as their ability to reason about the scene in the form of text." + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.659, + 0.905, + 0.78 + ], + "angle": 0, + "content": "Prompting LLMs. One category of approaches involves prompting LLMs with text using reasoning frameworks such as chain-of-thought [46]. LanguageMPC [35] leverages the common-sense reasoning capacity of LLM to make high-level driving decisions, which are then translated into bottom-level control signals. GPT-Driver [28] converts observations and ego-states into language prompts, framing driving planning as a text-based task for LLMs." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.781, + 0.905, + 0.901 + ], + "angle": 0, + "content": "End-to-end Driving with MLLMs. Another approach is to integrate multi-modal LLMs into end-to-end train frameworks for autonomous vehicles. LMDrive [36] integrates a vision encoder with a large language model, enabling natural language instruction in autonomous driving. DriveLM introduces multi-step question answering, going beyond single-round interactions [27, 32, 34]. DriveVLM [39] leverages MLLMs and a chain-of-thought process to" + }, + { + "type": "page_number", + "bbox": [ + 0.479, + 0.945, + 0.521, + 0.957 + ], + "angle": 0, + "content": "27576" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.134, + 0.092, + 0.859, + 0.304 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.089, + 0.311, + 0.908, + 0.382 + ], + "angle": 0, + "content": "Figure 2. Overview of DiMA. The input to the framework is a multi-view image sequence and a question text prompt. The vision-based end-to-end planner consists of a scene encoder and a planning transformer. The scene encoder learns structured latent representations in the form of bird's-eye-view, ego, agent, and map (BEAM) token embeddings and acts as a trainable tokenizer for the multi-modal large language model (MLLM). The planning transformer is trained under standard planning constraints [18, 20]. The MLLM is trained for planning, visual question answering, distillation, and a series of surrogate tasks." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.402, + 0.379, + 0.416 + ], + "angle": 0, + "content": "enhance scene understanding and planning." + }, + { + "type": "text", + "bbox": [ + 0.089, + 0.417, + 0.484, + 0.584 + ], + "angle": 0, + "content": "We point out a challenge in the current MLLM strategy for autonomous driving. We suggest that in order to reason about vehicle dynamics, the language model benefits from a structured input that explicitly models scene components instead of dense, unstructured input to the MLLM as in [36, 39, 43, 44]. In TOKEN [38], the scene encoder from [48] is used as a tokenizer. However, it is frozen during training. In our approach, the input to the MLLM is a set of latent representations learned from a pre-trained scene encoder that is jointly trained with the MLLM, allowing it to more learn grounded feature representations" + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.595, + 0.273, + 0.611 + ], + "angle": 0, + "content": "3. DiMA Framework" + }, + { + "type": "text", + "bbox": [ + 0.089, + 0.617, + 0.484, + 0.738 + ], + "angle": 0, + "content": "We present DiMA, a framework for end-to-end autonomous driving. Given a sequence of multi-view images, the overall objective is to predict the future trajectory of the ego vehicle and answer questions about the scene. An overview of DiMA is shown in Figure 2. The framework has two main components: 1) the vision-based planner; 2) a multi-modal large language model consisting of adapter layers, an LLM, and a series of task-specific decoder heads." + }, + { + "type": "title", + "bbox": [ + 0.09, + 0.747, + 0.292, + 0.761 + ], + "angle": 0, + "content": "3.1. Vision-based Planner" + }, + { + "type": "text", + "bbox": [ + 0.089, + 0.765, + 0.484, + 0.902 + ], + "angle": 0, + "content": "Vision-based end-to-end planners are trained in a multi-task fashion to perform perception, mapping, motion prediction and planning [18, 20, 40, 48]. Considering standard architecture design from [18, 20, 48] we decompose the planner into the scene encoder and the planning transformer. The scene encoder provides structured latent representations to the planning transformer that performs waypoint prediction. We call these representations as scene token embeddings. In our framework, we leverage the vision-based planner in" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.402, + 0.907, + 0.507 + ], + "angle": 0, + "content": "two ways. First, it acts as the primary network through which planning is performed, allowing for fast prediction at test time. Second, the scene encoder is shared with the MLLM, acting as tokenizer to provide it with a highly structured set of inputs. An important distinction of our framework from previous works [19, 38] is that the scene encoder is trained jointly with the MLLM." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.509, + 0.909, + 0.823 + ], + "angle": 0, + "content": "Scene encoder The scene encoder models the scene as high-dimensional token embeddings that explicitly represent components such as the environment map, the ego vehicle, and the surrounding agents. This is achieved by introducing task specific, learnable query features that are supervised by a series of planning constraints [20]. First, a visual backbone encodes the input multi-view image sequence. The obtained feature map is projected into the bird's-eye-view (BEV) space using a set of queries to obtain BEV token embeddings \\((B)\\) [22]. The network learns structured representations of the map and agents in the form of map token embeddings \\((M)\\) and agent token \\((A)\\) embeddings by cross-attending the BEV features with map and agent queries. A randomly initialized learnable embedding is trained to learn the interaction of the ego vehicle with the agent and map components which is the ego query \\((E)\\). We denote the set of these latent scene representations as \\(BEAM\\) token embeddings. These token embeddings are sent to a planning head that predicts future trajectories of the ego and agent vehicles. They are also used as input to the MLLM as multi-modal inputs, as discussed next." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.835, + 0.693, + 0.85 + ], + "angle": 0, + "content": "3.2. Multi-modal LLM" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.856, + 0.908, + 0.903 + ], + "angle": 0, + "content": "Our objective is to distill knowledge from an MLLM to the vision-based planner. Towards this, we jointly train an MLLM with the vision-based planner. Specifically, the" + }, + { + "type": "page_number", + "bbox": [ + 0.479, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "27577" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.09, + 0.092, + 0.485, + 0.258 + ], + "angle": 0, + "content": "scene encoder is used as a trainable tokenizer for the MLLM to generate the \\(BEAM\\) token embeddings, as well as used as an encoder for the planning transformer. The advantage from this strategy is two-fold. First, the MLLM receives a highly structured input that captures rich spatio-temporal information relevant to the autonomous driving task. Second, the scene encoder learns features that are grounded in language, improving the robustness of vision-based planning. The main components of the MLLM are adapter layers, the large language model, and a series of task-specific decoder heads which we discuss next." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.264, + 0.332, + 0.278 + ], + "angle": 0, + "content": "3.2.1. Adaptation of Scene Tokens" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.281, + 0.484, + 0.433 + ], + "angle": 0, + "content": "In order to project the BEAM token embeddings efficiently while maintaining their distinctiveness, we leverage the query-transformer (Q-former) module [17] to compress the visual tokens before being input to the LLM. We implement component-specific Q-former layers, namely a Map Q-former, a Bird's-Eye-View (BEV) Q-former, an Ego Q-former, and an Agent Q-former. Each of these adapters transforms the representing different modalities into a sequence of fixed-length, higher-dimensional tokens, preparing them for processing by the MLLM." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.438, + 0.276, + 0.453 + ], + "angle": 0, + "content": "3.2.2. MLLM Supervision" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.456, + 0.484, + 0.547 + ], + "angle": 0, + "content": "We design tasks for the MLLM with the objectives of, a) enriching the intermediate scene representations; b) grounding the scene token embeddings in language; and c) training the LLM for planning-related reasoning. In detail, the MLLM is trained for visual question answering, trajectory estimation, feature distillation, and a set of surrogate tasks [1, 51]." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.55, + 0.484, + 0.702 + ], + "angle": 0, + "content": "Visual question answering. The MLLM is trained for visual question answering (VQA) on question-answer pairs of 4 types: perception of the scene, prediction of the agent behavior, prediction of ego-vehicle behavior, and future planning steps [37]. The answer is predicted based on a multimodal prompt consisting of the question embeddings and the projected BEAM token embeddings. This is constructed as a standard next-token prediction task and the VQA branch is supervised with a cross-entropy loss as in [21]. We denote the loss from this task head as \\(\\mathcal{L}_{LLM}\\)." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.705, + 0.483, + 0.764 + ], + "angle": 0, + "content": "Surrogate tasks: The main goal of joint training is to enrich the BEAM scene representations learned by the scene encoder. We design MLLM surrogate tasks whose objectives are in line with that of future trajectory prediction." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.765, + 0.484, + 0.902 + ], + "angle": 0, + "content": "1) Masked token reconstruction: Each type of token embedding contributes to a holistic representation of the scene. In order to enrich the visual representations, we ask the network to reconstruct a masked BEV input based on the context present in the rest of the multi-modal sequence. We perform random masking after scene encoding and pass the token embeddings to the MLLM. A reconstruction head takes the latent representations from the penultimate layer of LLM and predicts reconstructed BEV token embeddings" + }, + { + "type": "image", + "bbox": [ + 0.554, + 0.089, + 0.865, + 0.336 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.512, + 0.341, + 0.907, + 0.41 + ], + "angle": 0, + "content": "Figure 3. Examples of addition and deletion in scene editing. In the top row, a car (solid red box) is added on the left of the ego-vehicle (green box). In the bottom row, a car (dashed red box) is removed from the right of the ego-vehicle. A corresponding question-answer pair is created to characterize the edit." + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.427, + 0.906, + 0.459 + ], + "angle": 0, + "content": "\\(\\hat{B}\\). This decoder head is supervised with the L2 loss between the prediction and the complete input," + }, + { + "type": "equation", + "bbox": [ + 0.639, + 0.46, + 0.905, + 0.478 + ], + "angle": 0, + "content": "\\[\n\\mathcal {L} _ {\\text {r e c o n}} = \\left\\| \\hat {B} - B \\right\\| ^ {2} \\tag {1}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.481, + 0.906, + 0.542 + ], + "angle": 0, + "content": "where \\((\\{m(B), E, A, M\\})\\) is input to the MLLM, and the latent MLLM representations associated with the BEV token embeddings are input to the masked reconstruction decoder head. Here, \\(m(.)\\) denotes random masking." + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.543, + 0.907, + 0.663 + ], + "angle": 0, + "content": "2) Future BEV prediction: An important aspect of planning is anticipating future events. We introduce the surrogate task of future BEV prediction to encourage the LLM to learn spatio-temporal cues useful for planning. Given latent BEV token embeddings, we train a prediction head to predict future BEV token embeddings and supervise this task as the L2 loss between predicted and ground truth future token embeddings:" + }, + { + "type": "equation", + "bbox": [ + 0.563, + 0.665, + 0.905, + 0.684 + ], + "angle": 0, + "content": "\\[\n\\mathcal {L} _ {\\text {f u t u r e}} = \\left\\| \\hat {F} _ {t} - B _ {t + 1} \\right\\| ^ {2} + \\left\\| \\tilde {F} _ {t} - B _ {t + 2} \\right\\| ^ {2} \\tag {2}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.687, + 0.906, + 0.749 + ], + "angle": 0, + "content": "where \\(\\hat{F}_t, \\tilde{F}_t\\) are the predicted future BEV token embeddings at time \\(t\\). Note, we predict future BEV token embeddings \\((B_{t+1}, B_{t+2})\\) of the multi-view image sequence for time steps \\(\\{t+1, t+2\\}\\)." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.75, + 0.907, + 0.902 + ], + "angle": 0, + "content": "3) Scene editing: For prediction and reasoning about the ego vehicle, it is crucial to learn how surrounding agents impact the ego-vehicle's future path. We propose a novel scene editing task in which we augment scenes by removing or adding new agents. Along with this, we construct a question-answer pair related to the edit. We show examples of this in Figure 3. For scene addition, given the map constraints, predicted map, the ego bounding box location and the trajectories of predicted agents, we create a way-point trajectory for a new agent of category \"car\" or \"truck\". A" + }, + { + "type": "page_number", + "bbox": [ + 0.479, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "27578" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.09, + 0.092, + 0.482, + 0.242 + ], + "angle": 0, + "content": "new agent token embedding is then created using a linear layer. This new agent token embedding, the corresponding text prompt, and the rest of the BEAM token embeddings are passed as input to the LLM. The hidden latent LLM features corresponding to agent token embeddings are then fed into a dedicated scene editing decoder head that performs waypoint prediction of the ego vehicle. The language prediction head performs question-answering on the new QA pair. This task thus contributes to the existing planning constraint loss and VQA loss of the MLLM." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.247, + 0.482, + 0.352 + ], + "angle": 0, + "content": "Distillation In order to further align the representations learned by the vision planner and the MLLM, we perform knowledge transfer between the penultimate layers of the language model and the planning transformer. Concretely, we minimize the KL-divergence between the distributions of the hidden features of the planning transformer and the LLM hidden features of the ego-token embeddings:" + }, + { + "type": "equation", + "bbox": [ + 0.191, + 0.356, + 0.482, + 0.372 + ], + "angle": 0, + "content": "\\[\n\\mathcal {L} _ {\\text {d i s t i l l}} = D _ {K L} \\left(P _ {l l m} \\| P _ {v i s}\\right) \\tag {3}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.375, + 0.482, + 0.435 + ], + "angle": 0, + "content": "where the \\(P_{vis}\\) are the features of penultimate layers of vision planning transformer and \\(P_{llm}\\) are the hidden LLM embeddings features corresponding to the ego-token embeddings of penultimate layers of MLLM model." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.436, + 0.482, + 0.481 + ], + "angle": 0, + "content": "Loss functions. We supervise the network with a weighted sum of the losses corresponding to planning, visual question answering, distillation, and each surrogate task." + }, + { + "type": "equation", + "bbox": [ + 0.1, + 0.485, + 0.482, + 0.501 + ], + "angle": 0, + "content": "\\[\n\\mathcal {L} = \\mathcal {L} _ {\\text {p l a n n i n g}} + \\mathcal {L} _ {\\text {L L M}} + \\mathcal {L} _ {\\text {r e c o n}} + \\mathcal {L} _ {\\text {f u t u r e}} + \\mathcal {L} _ {\\text {d i s t i l l}} \\tag {4}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.504, + 0.482, + 0.548 + ], + "angle": 0, + "content": "here \\(\\mathcal{L}_{\\text{planning}}\\) comes from the planning objective used to train [18, 20]. Here, the loss weights are chosen to bring each value to the same scale." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.563, + 0.283, + 0.581 + ], + "angle": 0, + "content": "4. Experimental Setup" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.585, + 0.482, + 0.661 + ], + "angle": 0, + "content": "In Section 4.1, we present the architecture design details, in Section 4.2 we present the training strategy, and in Section 4.3, we provide information about the datasets used. Importantly, in Section 4.4 we detail multiple evaluation strategies for fair and comprehensive comparison with existing works." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.671, + 0.279, + 0.687 + ], + "angle": 0, + "content": "4.1. Architecture design" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.689, + 0.482, + 0.795 + ], + "angle": 0, + "content": "Vision-based planner We conduct experiments using two vision-based end-to-end planners, VAD [20] and UniAD [18] due to their performance, efficiency, and ability to model the interaction between scene components. We use two model size variants of VAD. Both planners are trained for perception, motion prediction and planning, while UniAD also performs occupancy prediction." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.796, + 0.482, + 0.901 + ], + "angle": 0, + "content": "MLLM design The MLLM is made up of adapter layers, a language model, and a set of task-specific decoder layers. We project each \\(BEAM\\) token embedding with its own dedicated Q-former adapter layer following [17]. After projection, the \\(BEAM\\) token embeddings and language token embeddings lie in the same embedding space. In order to reduce memory consumption, we limit the agent token se" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.092, + 0.905, + 0.213 + ], + "angle": 0, + "content": "quence length to a fraction of the total number of agents and perform sequence-wise upsampling after input to the language model. The adapter layers are randomly initialized. We use the LLM from LLaVA-v1.5-7B [24] as our language model base. The ego prediction and agent motion prediction task heads are multi-layer-perceptron networks following [20]. For the surrogate task decoder heads, we use 3 Linear layers with a ReLU activation layer." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.222, + 0.618, + 0.238 + ], + "angle": 0, + "content": "4.2. Training" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.24, + 0.905, + 0.377 + ], + "angle": 0, + "content": "DiMA training follows a two-stage approach. First the vision-only planner is pre-trained for perception, prediction, and planning for 60 epochs in order to learn informative latent scene representations. Second, we perform joint training of the vision planner and the MLLM for an additional 30 epochs, incorporating all proposed tasks and losses detailed in Section 3. In the second stage, the language model of the MLLM is fine-tuned using LoRA [15]. Please refer to the supplementary material for more details." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.387, + 0.615, + 0.4 + ], + "angle": 0, + "content": "4.3. Datasets" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.405, + 0.907, + 0.723 + ], + "angle": 0, + "content": "We use the nuScenes dataset for the task of open-loop planning [2]. This dataset consists of 28k total samples in a \\(22\\mathrm{k} / 6\\mathrm{k}\\) training/Validation split. The objects in each scene are annotated with 3D bounding box, orientation, and vehicle speed information. Additionally, we make use of the CAN bus information for the ego vehicle's ground-truth trajectory annotations. For supervising visual question answering, we train with DriveLM [37]. This dataset consists of a 4k subset of samples from the nuScenes dataset annotated with 300k QA pairs related to perception, prediction, planning, and behavior of the ego vehicle. For the samples in nuScenes that do not have text annotations, we create a set of perception, planning, and prediction QA pairs based on the numerical annotations. We use this information to prompt a Llama3-70B model [12] with in-context examples to generate DriveLM-like QA pairs. For the behavior QA pair, we create a rule-based algorithm to categorize the future movement of the ego-vehicle based on the ground-truth trajectory values and ego vehicle speed. More details on this method as well as examples of generated QA pairs may be found in the supplementary." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.732, + 0.688, + 0.746 + ], + "angle": 0, + "content": "4.4. Evaluation details" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.75, + 0.905, + 0.84 + ], + "angle": 0, + "content": "We evaluate planning performance based on the predicted future trajectories of the ego vehicle 3 seconds into the future, where 2 waypoints are predicted per second. We use two metrics, the L2 error (in meters) between the predicted and ground-truth waypoints as well as the collision rate (in \\(\\%\\) ) between the ego-vehicle and the surrounding vehicles." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.841, + 0.905, + 0.901 + ], + "angle": 0, + "content": "Standardized vs VAD evaluation. In [48], Weng et al. point out inconsistencies in planning evaluation across VAD [20], UniAD [18] and AD-MLP [54], namely in the way L2 error is averaged over time and the way noisy and invalid" + }, + { + "type": "page_number", + "bbox": [ + 0.479, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "27579" + } + ], + [ + { + "type": "table_caption", + "bbox": [ + 0.09, + 0.089, + 0.907, + 0.146 + ], + "angle": 0, + "content": "Table 1. Comparison of L2 trajectory error and collision rate on nuScenes [2] using standardized evaluation [48]. Models are evaluated on the general validation split as well as a \"targeted\" split of challenging samples from [48]. The performance of the DiMA model variants are in shades of purple. We summarize results by averaging over at \\( t = \\{1,2,3\\} s \\) as well as at all time steps. The best average performance in each setting is in bold. \"+\" indicates the use of ego-status information." + }, + { + "type": "table", + "bbox": [ + 0.097, + 0.15, + 0.903, + 0.511 + ], + "angle": 0, + "content": "
MethodUsing Ego statusTraj L2 (m) ↓Collision (%) ↓
1s2s3sAve1,2,3sAve all1s2s3sAve1,2,3sAve all
Full validation split
UniAD[18]X0.480.891.470.950.830.320.290.730.450.40
VAD-Base [20]X0.410.861.460.910.780.020.260.830.370.30
PARA-Drive[48]X0.260.591.120.660.560.000.120.650.260.17
TOKEN[38]X0.260.701.460.810.68----0.15
DiMA (VAD-Tiny)X0.200.531.100.610.510.000.070.190.090.06
DiMA (VAD-Base)X0.180.501.020.570.470.000.050.160.070.06
AD-MLP[54]✓0.230.581.180.660.560.000.140.700.280.20
PARA-Drive+ [48]✓0.200.521.040.590.490.000.090.490.190.13
DiMA+ (VAD-Tiny)✓0.190.511.060.590.490.000.060.160.080.06
DiMA+ (VAD-Base)✓0.180.481.010.560.460.000.050.160.070.06
DiMA-Dual+ (VAD-Tiny)✓0.180.501.030.570.470.000.050.160.080.06
Targeted validation split
UniAD[18]X0.471.091.921.160.990.000.000.730.240.15
VAD-Base [20]X0.531.202.071.271.080.000.290.870.390.34
PARA-Drive[48]X0.380.971.881.080.910.000.000.720.240.14
DiMA (VAD-Tiny)X0.320.881.700.970.810.000.030.240.090.05
DiMA (VAD-Base)X0.290.791.490.830.710.000.030.220.080.05
AD-MLP[54]✓0.330.992.061.130.940.000.583.621.400.94
PARA-Drive+ [48]✓0.290.761.460.830.700.000.000.290.100.05
DiMA+ (VAD-Tiny)✓0.310.831.670.940.770.000.040.170.070.05
DiMA+ (VAD-Base)✓0.270.741.380.790.640.000.030.160.060.04
DiMA-Dual+ (VAD-Tiny)✓0.280.761.490.840.690.000.030.170.070.05
" + }, + { + "type": "text", + "bbox": [ + 0.089, + 0.525, + 0.482, + 0.691 + ], + "angle": 0, + "content": "frames are considered. The authors also point out that collision performance is considerably improved by using a more finely discretized BEV map. Accounting for these inconsistencies, they propose a \"standardized\" evaluation metric to maintain a fair comparison across methods. For accurate reporting and fair comparison with [38, 48], we use this standardized metric to evaluate our model. Certain existing works use the evaluation scheme followed by [20] (which we call VAD evaluation) and do not make code or models available [16, 30, 39, 43]. Here, we compare reported results against DiMA evaluated using VAD evaluation." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.693, + 0.483, + 0.785 + ], + "angle": 0, + "content": "Validation splits. We evaluate planning performance across 3 types of splits from the nuScenes validation set: 1) the entire validation set of 6019 samples 2) a 689 sample sub-set more difficult \"targeted\" scenarios in which the ego-vehicle turns during navigation [48], and 3) long-tail events in which the ego-vehicle encounters rare maneuvers [38]. u" + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.802, + 0.295, + 0.82 + ], + "angle": 0, + "content": "5. Experimental Results" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.825, + 0.484, + 0.902 + ], + "angle": 0, + "content": "We perform a comprehensive evaluation of DiMA on the nuScenes open-loop planning benchmark in a variety of evaluation settings. We compare our performance against that of recent vision-based planners [16, 18, 20, 48] as well as recent MLLM planners [30, 38, 39, 43]. To facilitate a" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.525, + 0.907, + 0.571 + ], + "angle": 0, + "content": "fair comparison of DiMA with recent works, we use two evaluation strategies VAD evaluation [20] and standardized evaluation [48]. Please refer to Section 4.4 for more details." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.577, + 0.909, + 0.865 + ], + "angle": 0, + "content": "Standardized evaluation. Under the standardized evaluation strategy, we evaluate each method on two validation splits: 1) the full nuScenes validation split, and 2) a \"targeted\" subset of the nuScenes validation split that contains scenarios where the ego-vehicle must take a turn, as established in [48]. We present our results in Table 1. The baseline vision planners VAD [20] and UniAD [18] exhibit suboptimal performance in more difficult targeted scenarios. DiMA demonstrates significant improvements over baseline methods, achieving a \\(35\\%\\) reduction in L2 trajectory error and an \\(80\\%\\) reduction in collision rate on the overall nuScenes validation set, and a \\(35\\%\\) reduction in L2 trajectory error in the \"targeted\" split. More notably, we outperform recent state-of-the-art methods PARA-Drive [48] and TOKEN [38] consistently across both validation splits. We observe a strong improvement in collision rate, improving on that of TOKEN by \\(60\\%\\). Unlike TOKEN, our approach does not rely on an LLM for inference, thus making DiMA more efficient as well as more accurate." + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.871, + 0.909, + 0.903 + ], + "angle": 0, + "content": "VAD evaluation. For a fair comparison with recent works that use the evaluation strategy from VAD [20], we present" + }, + { + "type": "page_number", + "bbox": [ + 0.479, + 0.945, + 0.521, + 0.958 + ], + "angle": 0, + "content": "27580" + } + ], + [ + { + "type": "table_caption", + "bbox": [ + 0.09, + 0.089, + 0.907, + 0.146 + ], + "angle": 0, + "content": "Table 2. Comparison of L2 trajectory error and collision rate on nuScenes [2] using VAD evaluation [20]. Models are evaluated on the general validation split. The performance of the DiMA model variants are in shades of purple. The performance of the MLLM-branch DiMA model variant is in blue. We summarize results by averaging over all time steps. The best average performance in each setting is in bold. “+” indicates the use of ego-status information." + }, + { + "type": "table", + "bbox": [ + 0.095, + 0.15, + 0.907, + 0.399 + ], + "angle": 0, + "content": "
MethodUsing Ego statusTraj L2 (m) ↓Collision (%) ↓Latency (ms)FPS
1s2s3sAveall1s2s3sAveall
ST-P3 [16]X1.332.112.902.110.230.621.270.71628.31.6
UniAD[18]X0.480.961.651.030.050.170.710.31555.61.8
VAD-Tiny[20]X0.460.761.120.780.210.350.580.3859.516.8
VAD-Base[20]X0.410.701.050.720.070.170.410.22224.34.5
VLP-VAD-Base[30]X0.260.470.780.500.120.170.420.23--
OmniDrive [43]X0.400.801.320.840.040.462.320.94--
DiMA (UniAD)X0.150.300.560.340.060.080.220.125601.8
DiMA (VAD-Tiny)X0.180.360.610.380.070.100.270.1559.516.8
DiMA (VAD-Base)X0.130.270.470.290.050.080.160.102264.5
VAD-Tiny✓0.200.380.650.410.100.120.270.1659.516.8
VAD-Base✓0.170.340.600.370.070.100.240.14224.34.5
DriveVLM [39]✓0.180.340.680.400.100.220.450.27--
DriveVLM-Dual (VAD-Base) [39]✓0.150.290.480.310.050.080.170.10--
DiMA + (VAD-Tiny)✓0.170.330.590.370.070.100.260.1459.516.8
DiMA + (VAD-Base)✓0.120.250.440.270.040.060.150.082264.5
DiMA-Dual+ (VAD-Tiny)✓0.140.270.460.290.050.070.150.092863.5
" + }, + { + "type": "table_caption", + "bbox": [ + 0.09, + 0.422, + 0.482, + 0.519 + ], + "angle": 0, + "content": "Table 3. Long-tail performance comparison of L2 trajectory error and collision rate on nuScenes [2] validation set using standardized evaluation [48]. The performance of vision planner DiMA model variants are in shades of purple. The performance of MLLM-branch DiMA model variants are in shades of blue. We summarize results by averaging over at \\( t = \\{1,2,3\\} s \\) as well as at all time steps. The best average performance in each setting is in bold." + }, + { + "type": "table", + "bbox": [ + 0.095, + 0.523, + 0.479, + 0.825 + ], + "angle": 0, + "content": "
MethodTraj L2 (m)↓Coll. (%)↓
1s2s3sAve1,2,3sAveallAveall
3-point turn (zero-shot)
VAD-Tiny0.761.683.041.831.550.00
VAD-Base0.711.663.241.871.570.00
PARA-Drive [48]0.501.382.761.551.295.33
TOKEN [38]0.391.292.601.431.184.00
DiMA (VAD-Tiny)0.471.272.551.431.170.00
DiMA (VAD-Base)0.361.182.371.301.050.00
DiMA (MLLM)0.381.122.351.281.040.00
DiMA-Dual (VAD-Tiny)0.381.082.351.271.040.00
Resume from stop
VAD-Tiny0.641.632.991.751.490.00
VAD-Base0.601.722.831.721.420.00
PARA-Drive0.140.792.301.080.850.00
TOKEN0.130.701.580.800.650.00
DiMA (VAD-Tiny)0.261.022.401.230.990.00
DiMA (VAD-Base)0.150.651.340.710.660.00
DiMA (MLLM)0.241.022.161.140.930.00
DiMA-Dual (VAD-Tiny)0.240.982.131.110.910.00
Overtake
VAD-Tiny0.581.272.121.321.142.42
VAD-Base0.461.162.171.261.062.49
PARA-Drive0.270.891.941.030.852.30
TOKEN0.290.771.630.900.740.00
DiMA (VAD-Tiny)0.240.751.490.830.691.32
DiMA (VAD-Base)0.240.721.500.820.661.29
DiMA (MLLM)0.240.731.500.820.671.30
DiMA-Dual (VAD-Tiny)0.240.731.500.820.671.30
" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.831, + 0.483, + 0.878 + ], + "angle": 0, + "content": "an additional set of results under this strategy. In Table 2, we present the evaluation of DiMA on the nuScenes validation set [2]. We significantly outperform the base" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.424, + 0.907, + 0.515 + ], + "angle": 0, + "content": "line vision-based planners VAD [20] and UniAD [18]. Notably, DiMA built on VAD-Tiny outperforms VAD-Base by \\(47\\%\\) in L2 trajectory error while being 4 times faster. We also outperform MLLM planners like OmniDrive [43] and DriveVLM [39] without requiring an LLM at inference, making our approach more accurate as well as efficient." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.519, + 0.909, + 0.716 + ], + "angle": 0, + "content": "Performance in long-tail scenarios. Autonomous driving performance is particularly crucial for rare scenarios. Using the manually selected long-tail events from [38], we demonstrate the robustness of DiMA in cases of novel navigation and perception. In Table 3, we evaluate the performance of DiMA for three long-tail scenarios, \"3-point turn\", \"resume from the stop\" and \"overtake\", selected from the nuScenes validation set. Training with DiMA results in a significant and consistent boost in performance from the baseline vision-based planners. We also consistently outperform PARA-Drive and TOKEN. Notably, the \"3-point turn\" case is a zero-shot scenario, not present in the training data. We achieve the lowest L2 trajectory error for this case." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.72, + 0.911, + 0.901 + ], + "angle": 0, + "content": "Evaluating the MLLM branch. We also evaluate the MLLM branch of DiMA and demonstrate its planning and visual question-answering capabilities. We perform LLM-based planning using DiMA in two ways. First, we directly evaluate the waypoint predictions of the MLLM planning decoder head. This is denoted by DiMA (MLLM). Secondly, we define DiMA-Dual, an optional hybrid inference model that combines the strengths of both the vision and MLLM branches. We perform max-pooling on the penultimate layer features of both vision planning transformer and MLLM ego prediction. The resulting pooled features are then fed back into both the vision planning transformer and" + }, + { + "type": "page_footnote", + "bbox": [ + 0.109, + 0.887, + 0.455, + 0.901 + ], + "angle": 0, + "content": "1We compare with reported results due to lack of published code." + }, + { + "type": "page_number", + "bbox": [ + 0.479, + 0.945, + 0.518, + 0.957 + ], + "angle": 0, + "content": "27581" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.094, + 0.089, + 0.493, + 0.278 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.502, + 0.089, + 0.901, + 0.278 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.09, + 0.284, + 0.908, + 0.313 + ], + "angle": 0, + "content": "Figure 4. Visualization of planning performance by the MLLM branch of DiMA-VAD-Tiny. We plot the predicted trajectory (orange-yellow) and show an example response of the language model branch to a question from the DriveLM test dataset [37]." + }, + { + "type": "table_caption", + "bbox": [ + 0.09, + 0.327, + 0.908, + 0.357 + ], + "angle": 0, + "content": "Table 4. Results of ablation experiments of DiMA-VAD-Tiny on nuScenes under VAD evaluation. The baseline model is VAD-Tiny. We report L2 trajectory error and collision rate averaged over all time steps." + }, + { + "type": "table", + "bbox": [ + 0.095, + 0.359, + 0.892, + 0.501 + ], + "angle": 0, + "content": "
IDVQAScene tokensDistillationLLM PlanningSurrogate tasksTraj L2 (m) ↓Collision (%) ↓
Masked recon.Future pred.Scene editing1s2s3sAveall1s2s3sAveall
1XXXXXXX0.330.580.890.600.120.190.550.29
2✓BEVX✓XXX0.360.600.910.620.100.170.520.26
3✓BEV, MapX✓XXX0.280.520.870.560.100.170.490.25
4✓AllX✓XXX0.260.510.830.520.100.160.380.21
5✓All✓✓XXX0.220.440.770.480.090.140.340.19
6✓All✓✓✓XX0.190.390.680.420.090.130.320.18
7✓All✓✓✓✓X0.180.370.630.390.070.110.290.16
8✓All✓✓✓✓✓0.180.360.610.380.070.100.270.15
" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.52, + 0.483, + 0.625 + ], + "angle": 0, + "content": "MLLM ego prediction for waypoint prediction. This max pooling encourages consistency between the predictions of the two models, leading to a more robust and potentially improved overall prediction as seen in Table 3. Using the MLLM at inference results in a significant boost in performance, with DiMA-Dual (VAD-Tiny) matching the performance of DiMA (VAD-Base) in Table 2." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.627, + 0.484, + 0.703 + ], + "angle": 0, + "content": "VQA results. We provide qualitative examples of visual question-answering in the DriveLM dataset [37] in Figure 4. We demonstrate the ability of DiMA to reason about safe actions to be taken by the ego vehicle. We provide a more examples in the supplementary material." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.715, + 0.245, + 0.731 + ], + "angle": 0, + "content": "5.1. Ablation study." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.735, + 0.484, + 0.901 + ], + "angle": 0, + "content": "We examine the role of different design choices and proposed tasks in the DiMA framework through a comprehensive ablation study in Table 4. In order to do so, we build our final frame-work step-by-step. We begin with VAD-Tiny as our baseline model in the first row, denoted by ID-1. In ID-2, we observe inconsistent gains and drops in performance by naively training an LLM with the VAD-Tiny under a the planning and VQA objectives while using only BEV features. By comparing ID-3 and ID-4, we show that that providing more scene context to the MLLM by using all the BEAM token embeddings to train both LLM and" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.52, + 0.907, + 0.672 + ], + "angle": 0, + "content": "vision-based planner benefits the performance, resulting a significant improvement in trajectory estimation and collision rate. We show in ID-5 that explicitly distilling the information from the penultimate layers of the MLLM to the planning transformer gives another boost in performance. In ID-6, ID-7 and ID-8 we demonstrate the importance of surrogate tasks like masked-reconstruction, future prediction and scene editing in helping the scene encoder to learn richer representations useful for planning. We point to the supplementary material for additional ablation studies." + }, + { + "type": "title", + "bbox": [ + 0.514, + 0.695, + 0.634, + 0.71 + ], + "angle": 0, + "content": "6. Conclusion" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.72, + 0.907, + 0.901 + ], + "angle": 0, + "content": "We introduce DiMA, an end-to-end autonomous driving framework for robust and efficient planning. DiMA distills knowledge from an MLLM to a vision-based planner, employing a novel joint training strategy alongside carefully designed surrogate tasks such as masked token reconstruction, future token prediction, and scene editing. DiMA enables the model to learn semantically grounded scene representations, improving its performance in difficult navigation scenarios. We conduct extensive evaluations and ablation studies, showing that training with DiMA results in a planner that excels in long-tail cases while maintaining computational efficiency." + }, + { + "type": "page_number", + "bbox": [ + 0.479, + 0.945, + 0.521, + 0.957 + ], + "angle": 0, + "content": "27582" + } + ], + [ + { + "type": "title", + "bbox": [ + 0.093, + 0.09, + 0.188, + 0.106 + ], + "angle": 0, + "content": "References" + }, + { + "type": "ref_text", + "bbox": [ + 0.101, + 0.115, + 0.484, + 0.184 + ], + "angle": 0, + "content": "[1] Apratim Bhattacharyya, Sunny Panchal, Reza Pourreza, Mingu Lee, Pulkit Madan, and Roland Memisevic. Look, remember and reason: Grounded reasoning in videos with language models. In International Conference of Learning Representations, 2024. 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.101, + 0.185, + 0.484, + 0.269 + ], + "angle": 0, + "content": "[2] Holger Caesar, Varun Bankiti, Alex H Lang, Sourabh Vora, Venice Erin Liong, Qiang Xu, Anush Krishnan, Yu Pan, Giancarlo Baldan, and Oscar Beijbom. nuscenes: A multimodal dataset for autonomous driving. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 11621-11631, 2020. 1, 5, 6, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.271, + 0.483, + 0.325 + ], + "angle": 0, + "content": "[3] Sergio Casas, Abbas Sadat, and Raquel Urtasun. Mp3: A unified model to map, perceive, predict and plan. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 14403-14412, 2021. 1, 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.327, + 0.483, + 0.38 + ], + "angle": 0, + "content": "[4] Yuning Chai, Benjamin Sapp, Mayank Bansal, and Dragomir Anguelov. Multipath: Multiple probabilistic anchor trajectory hypotheses for behavior prediction. arXiv preprint arXiv:1910.05449, 2019. 1, 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.382, + 0.482, + 0.436 + ], + "angle": 0, + "content": "[5] Dian Chen and Philipp Krahenbuhl. Learning from all vehicles. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 17222-17231, 2022. 1, 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.439, + 0.483, + 0.521 + ], + "angle": 0, + "content": "[6] Long Chen, Oleg Sinavski, Jan Hünermann, Alice Karnsund, Andrew James Willmott, Danny Birch, Daniel Maund, and Jamie Shotton. Driving with llms: Fusing object-level vector modality for explainable autonomous driving. In 2024 IEEE International Conference on Robotics and Automation (ICRA), pages 14093-14100. IEEE, 2024. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.523, + 0.482, + 0.578 + ], + "angle": 0, + "content": "[7] Kashyap Chitta, Aditya Prakash, and Andreas Geiger. Neat: Neural attention fields for end-to-end autonomous driving. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 15793-15803, 2021. 1, 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.58, + 0.482, + 0.648 + ], + "angle": 0, + "content": "[8] Kashyap Chitta, Aditya Prakash, Bernhard Jaeger, Zehao Yu, Katrin Renz, and Andreas Geiger. Transfuser: Imitation with transformer-based sensor fusion for autonomous driving. IEEE Transactions on Pattern Analysis and Machine Intelligence, 45(11):12878-12895, 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.65, + 0.482, + 0.718 + ], + "angle": 0, + "content": "[9] Felipe Codevilla, Eder Santana, Antonio M López, and Adrien Gaidon. Exploring the limitations of behavior cloning for autonomous driving. In Proceedings of the IEEE/CVF international conference on computer vision, pages 9329-9338, 2019. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.72, + 0.482, + 0.774 + ], + "angle": 0, + "content": "[10] Xinpeng Ding, Jianhua Han, Hang Xu, Wei Zhang, and Xiaomeng Li. Hilm-d: Towards high-resolution understanding in multimodal large language models for autonomous driving. arXiv preprint arXiv:2309.05186, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.776, + 0.482, + 0.83 + ], + "angle": 0, + "content": "[11] Alexey Dosovitskiy, German Ros, Felipe Codevilla, Antonio Lopez, and Vladlen Koltun. Carla: An open urban driving simulator. In Conference on robot learning, pages 1-16. PMLR, 2017. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.832, + 0.482, + 0.9 + ], + "angle": 0, + "content": "[12] Abhimanyu Dubey, Abhinav Jauhri, Abhinav Pandey, Abhishek Kadian, Ahmad Al-Dahle, Aiesha Letman, Akhil Mathur, Alan Schelten, Amy Yang, Angela Fan, et al. The llama 3 herd of models. arXiv preprint arXiv:2407.21783, 2024. 2, 5" + }, + { + "type": "list", + "bbox": [ + 0.094, + 0.115, + 0.484, + 0.9 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.093, + 0.906, + 0.176 + ], + "angle": 0, + "content": "[13] Scott Ettinger, Shuyang Cheng, Benjamin Caine, Chenxi Liu, Hang Zhao, Sabeek Pradhan, Yuning Chai, Ben Sapp, Charles R Qi, Yin Zhou, et al. Large scale interactive motion forecasting for autonomous driving: The waymo open motion dataset. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 9710-9719, 2021. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.179, + 0.907, + 0.247 + ], + "angle": 0, + "content": "[14] Daocheng Fu, Xin Li, Licheng Wen, Min Dou, Pinlong Cai, Botian Shi, and Yu Qiao. Drive like a human: Rethinking autonomous driving with large language models. In Proceedings of the IEEE/CVF Winter Conference on Applications of Computer Vision, pages 910-919, 2024. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.249, + 0.906, + 0.304 + ], + "angle": 0, + "content": "[15] Edward J Hu, Yelong Shen, Phillip Wallis, Zeyuan Allen-Zhu, Yuanzhi Li, Shean Wang, Lu Wang, and Weizhu Chen. Lora: Low-rank adaptation of large language models. arXiv preprint arXiv:2106.09685, 2021. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.307, + 0.906, + 0.375 + ], + "angle": 0, + "content": "[16] Shengchao Hu, Li Chen, Penghao Wu, Hongyang Li, Junchi Yan, and Dacheng Tao. St-p3: End-to-end vision-based autonomous driving via spatial-temporal feature learning. In European Conference on Computer Vision, pages 533-549. Springer, 2022. 1, 2, 6, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.377, + 0.905, + 0.418 + ], + "angle": 0, + "content": "[17] Wenbo Hu, Zi-Yi Dou, Liunian Harold Li, Anita Kamath, Nanyun Peng, and Kai-Wei Chang. Matryoshka query transformer for large vision-language models, 2024. 4, 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.421, + 0.906, + 0.502 + ], + "angle": 0, + "content": "[18] Yihan Hu, Jiazhi Yang, Li Chen, Keyu Li, Chonghao Sima, Xizhou Zhu, Siqi Chai, Senyao Du, Tianwei Lin, Wenhai Wang, et al. Planning-oriented autonomous driving. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 17853-17862, 2023. 1, 2, 3, 5, 6, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.505, + 0.906, + 0.572 + ], + "angle": 0, + "content": "[19] Jyh-Jing Hwang, Runsheng Xu, Hubert Lin, Wei-Chih Hung, Jingwei Ji, Kristy Choi, Di Huang, Tong He, Paul Covington, Benjamin Sapp, et al. Emma: End-to-end multimodal model for autonomous driving. arXiv preprint arXiv:2410.23262, 2024. 1, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.576, + 0.906, + 0.658 + ], + "angle": 0, + "content": "[20] Bo Jiang, Shaoyu Chen, Qing Xu, Bencheng Liao, Jiajie Chen, Helong Zhou, Qian Zhang, Wenyu Liu, Chang Huang, and Xinggang Wang. Vad: Vectorized scene representation for efficient autonomous driving. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 8340-8350, 2023. 1, 2, 3, 5, 6, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.661, + 0.906, + 0.729 + ], + "angle": 0, + "content": "[21] Junnan Li, Dongxu Li, Silvio Savarese, and Steven Hoi. Blip-2: Bootstrapping language-image pre-training with frozen image encoders and large language models. In International conference on machine learning, pages 19730-19742. PMLR, 2023. 1, 2, 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.732, + 0.906, + 0.813 + ], + "angle": 0, + "content": "[22] Zhiqi Li, Wenhai Wang, Hongyang Li, Enze Xie, Chonghao Sima, Tong Lu, Yu Qiao, and Jifeng Dai. Bevformer: Learning bird's-eye-view representation from multi-camera images via spatiotemporal transformers. In European conference on computer vision, pages 1-18. Springer, 2022. 1, 2, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.817, + 0.905, + 0.884 + ], + "angle": 0, + "content": "[23] Xiwen Liang, Yangxin Wu, Jianhua Han, Hang Xu, Chunjing Xu, and Xiaodan Liang. Effective adaptation in multitask co-training for unified autonomous driving. Advances in Neural Information Processing Systems, 35:19645-19658, 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.887, + 0.905, + 0.902 + ], + "angle": 0, + "content": "[24] Haotian Liu, Chunyuan Li, Qingyang Wu, and Yong Jae Lee." + }, + { + "type": "list", + "bbox": [ + 0.517, + 0.093, + 0.907, + 0.902 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.48, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "27583" + } + ], + [ + { + "type": "ref_text", + "bbox": [ + 0.125, + 0.093, + 0.482, + 0.12 + ], + "angle": 0, + "content": "Visual instruction tuning. Advances in neural information processing systems, 36, 2024. 1, 2, 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.122, + 0.483, + 0.192 + ], + "angle": 0, + "content": "[25] Jiaqi Liu, Peng Hang, Xiao Qi, Jianqiang Wang, and Jian Sun. Mtd-gpt: A multi-task decision-making gpt model for autonomous driving at unsignalized intersections. In 2023 IEEE 26th International Conference on Intelligent Transportation Systems (ITSC), pages 5154-5161. IEEE, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.194, + 0.482, + 0.261 + ], + "angle": 0, + "content": "[26] Yicheng Liu, Jinghuai Zhang, Liangji Fang, Qinhong Jiang, and Bolei Zhou. Multimodal motion prediction with stacked transformers. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 7577-7586, 2021. 1, 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.264, + 0.482, + 0.332 + ], + "angle": 0, + "content": "[27] Srikanth Malla, Chiho Choi, Isht Dwivedi, Joon Hee Choi, and Jiachen Li. Drama: Joint risk localization and captioning in driving. In Proceedings of the IEEE/CVF winter conference on applications of computer vision, pages 1043-1052, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.334, + 0.482, + 0.376 + ], + "angle": 0, + "content": "[28] Jiageng Mao, Yuxi Qian, Junjie Ye, Hang Zhao, and Yue Wang. Gpt-driver: Learning to drive with gpt. arXiv preprint arXiv:2310.01415, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.379, + 0.482, + 0.46 + ], + "angle": 0, + "content": "[29] Jiquan Ngiam, Benjamin Caine, Vijay Vasudevan, Zhengdong Zhang, Hao-Tien Lewis Chiang, Jeffrey Ling, Rebecca Roelofs, Alex Bewley, Chenxi Liu, Ashish Venugopal, et al. Scene transformer: A unified architecture for predicting multiple agent trajectories. arXiv preprint arXiv:2106.08417, 2021. 1, 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.463, + 0.482, + 0.545 + ], + "angle": 0, + "content": "[30] Chenbin Pan, Burhaneddin Yaman, Tommaso Nesti, Abhirup Mallik, Alessandro G Allievi, Senem Velipasalar, and Liu Ren. Vlp: Vision language planning for autonomous driving. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 14760-14769, 2024. 6, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.548, + 0.482, + 0.629 + ], + "angle": 0, + "content": "[31] Jonah Philion and Sanja Fidler. Lift, splat, shoot: Encoding images from arbitrary camera rigs by implicitly unprojecting to 3d. In Computer Vision-ECCV 2020: 16th European Conference, Glasgow, UK, August 23-28, 2020, Proceedings, Part XIV 16, pages 194-210. Springer, 2020. 1, 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.632, + 0.482, + 0.702 + ], + "angle": 0, + "content": "[32] Tianwen Qian, Jingjing Chen, Linhai Zhuo, Yang Jiao, and Yu-Gang Jiang. Nuscenes-qa: A multi-modal visual question answering benchmark for autonomous driving scenario. In Proceedings of the AAAI Conference on Artificial Intelligence, pages 4542–4550, 2024. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.704, + 0.482, + 0.785 + ], + "angle": 0, + "content": "[33] Alec Radford, Jong Wook Kim, Chris Hallacy, Aditya Ramesh, Gabriel Goh, Sandhini Agarwal, Girish Sastry, Amanda Askell, Pamela Mishkin, Jack Clark, et al. Learning transferable visual models from natural language supervision. In International conference on machine learning, pages 8748-8763. PMLR, 2021. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.788, + 0.482, + 0.871 + ], + "angle": 0, + "content": "[34] Enna Sachdeva, Nakul Agarwal, Suhas Chundi, Sean Roelofs, Jiachen Li, Mykel Kochenderfer, Chiho Choi, and Behzad Dariush. Rank2tell: A multimodal driving dataset for joint importance ranking and reasoning. In Proceedings of the IEEE/CVF Winter Conference on Applications of Computer Vision, pages 7513-7522, 2024. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.873, + 0.482, + 0.901 + ], + "angle": 0, + "content": "[35] Hao Sha, Yao Mu, Yuxuan Jiang, Li Chen, Chenfeng Xu, Ping Luo, Shengbo Eben Li, Masayoshi Tomizuka, Wei" + }, + { + "type": "list", + "bbox": [ + 0.093, + 0.093, + 0.483, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.545, + 0.093, + 0.905, + 0.135 + ], + "angle": 0, + "content": "Zhan, and Mingyu Ding. *Languagempc: Large language models as decision makers for autonomous driving.* arXiv preprint arXiv:2310.03026, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.137, + 0.905, + 0.218 + ], + "angle": 0, + "content": "[36] Hao Shao, Yuxuan Hu, Letian Wang, Guanglu Song, Steven L Waslander, Yu Liu, and Hongsheng Li. Lmdrive: Closed-loop end-to-end driving with large language models. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 15120-15130, 2024. 2, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.221, + 0.905, + 0.289 + ], + "angle": 0, + "content": "[37] Chonghao Sima, Katrin Renz, Kashyap Chitta, Li Chen, Hanxue Zhang, Chengen Xie, Ping Luo, Andreas Geiger, and Hongyang Li. Drivelm: Driving with graph visual question answering. arXiv preprint arXiv:2312.14150, 2023. 1, 2, 4, 5, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.291, + 0.905, + 0.36 + ], + "angle": 0, + "content": "[38] Ran Tian, Boyi Li, Xinshuo Weng, Yuxiao Chen, Edward Schmerling, Yue Wang, Boris Ivanovic, and Marco Pavone. Tokenize the world into object-level knowledge to address long-tail events in autonomous driving. arXiv preprint arXiv:2407.00959, 2024. 1, 2, 3, 6, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.363, + 0.905, + 0.431 + ], + "angle": 0, + "content": "[39] Xiaoyu Tian, Junru Gu, Bailin Li, Yicheng Liu, Chenxu Hu, Yang Wang, Kun Zhan, Peng Jia, Xianpeng Lang, and Hang Zhao. Drivevlm: The convergence of autonomous driving and large vision-language models. arXiv preprint arXiv:2402.12289, 2024. 1, 2, 3, 6, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.434, + 0.905, + 0.502 + ], + "angle": 0, + "content": "[40] Wenwen Tong, Chonghao Sima, Tai Wang, Li Chen, Silei Wu, Hanming Deng, Yi Gu, Lewei Lu, Ping Luo, Dahua Lin, et al. Scene as occupancy. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 8406-8415, 2023. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.504, + 0.905, + 0.574 + ], + "angle": 0, + "content": "[41] Hugo Touvron, Thibaut Lavril, Gautier Izacard, Xavier Martinet, Marie-Anne Lachaux, Timothée Lacroix, Baptiste Rozière, Naman Goyal, Eric Hambro, Faisal Azhar, et al. Llama: Open and efficient foundation language models. arXiv preprint arXiv:2302.13971, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.576, + 0.905, + 0.645 + ], + "angle": 0, + "content": "[42] Hugo Touvron, Louis Martin, Kevin Stone, Peter Albert, Amjad Almahairi, Yasmine Babaei, Nikolay Bashlykov, Soumya Batra, Prajjwal Bhargava, Shruti Bhosale, et al. Llama 2: Open foundation and fine-tuned chat models. arXiv preprint arXiv:2307.09288, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.647, + 0.905, + 0.716 + ], + "angle": 0, + "content": "[43] Shihao Wang, Zhiding Yu, Xiaohui Jiang, Shiyi Lan, Min Shi, Nadine Chang, Jan Kautz, Ying Li, and Jose M Alvarez. Omnidrive: A holistic llm-agent framework for autonomous driving with 3d perception, reasoning and planning. arXiv preprint arXiv:2405.01533, 2024. 1, 3, 6, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.718, + 0.905, + 0.799 + ], + "angle": 0, + "content": "[44] Wenhai Wang, Jiangwei Xie, ChuanYang Hu, Haoming Zou, Jianan Fan, Wenwen Tong, Yang Wen, Silei Wu, Hanming Deng, Zhiqi Li, et al. Driveml: Aligning multi-modal large language models with behavioral planning states for autonomous driving. arXiv preprint arXiv:2312.09245, 2023. 1, 2, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.802, + 0.905, + 0.857 + ], + "angle": 0, + "content": "[45] Zitian Wang, Zehao Huang, Yulu Gao, Naiyan Wang, and Si Liu. Mv2dfusion: Leveraging modality-specific object semantics for multi-modal 3d detection. arXiv preprint arXiv:2408.05945, 2024. 1, 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.859, + 0.905, + 0.901 + ], + "angle": 0, + "content": "[46] Jason Wei, Xuezhi Wang, Dale Schuurmans, Maarten Bosma, Fei Xia, Ed Chi, Quoc V Le, Denny Zhou, et al. Chain-of-thought prompting elicits reasoning in large lan" + }, + { + "type": "list", + "bbox": [ + 0.517, + 0.093, + 0.905, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.48, + 0.945, + 0.52, + 0.957 + ], + "angle": 0, + "content": "27584" + } + ], + [ + { + "type": "ref_text", + "bbox": [ + 0.125, + 0.093, + 0.482, + 0.12 + ], + "angle": 0, + "content": "guage models. Advances in neural information processing systems, 35:24824-24837, 2022. 1, 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.122, + 0.483, + 0.191 + ], + "angle": 0, + "content": "[47] Licheng Wen, Daocheng Fu, Xin Li, Xinyu Cai, MA Tao, Pinlong Cai, Min Dou, Botian Shi, Liang He, and Yu Qiao. Dilu: A knowledge-driven approach to autonomous driving with large language models. In The Twelfth International Conference on Learning Representations. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.192, + 0.482, + 0.262 + ], + "angle": 0, + "content": "[48] Xinshuo Weng, Boris Ivanovic, Yan Wang, Yue Wang, and Marco Pavone. Para-drive: Parallelized architecture for real-time autonomous driving. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 15449-15458, 2024. 1, 2, 3, 5, 6, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.263, + 0.482, + 0.331 + ], + "angle": 0, + "content": "[49] Penghao Wu, Xiaosong Jia, Li Chen, Junchi Yan, Hongyang Li, and Yu Qiao. Trajectory-guided control prediction for end-to-end autonomous driving: A simple yet strong baseline. Advances in Neural Information Processing Systems, 35:6119-6132, 2022. 1, 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.333, + 0.482, + 0.401 + ], + "angle": 0, + "content": "[50] Zhenhua Xu, Yujia Zhang, Enze Xie, Zhen Zhao, Yong Guo, Kwan-Yee K Wong, Zhenguo Li, and Hengshuang Zhao. Drivegpt4: Interpretable end-to-end autonomous driving via large language model. IEEE Robotics and Automation Letters, 2024. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.403, + 0.482, + 0.485 + ], + "angle": 0, + "content": "[51] Rajeev Yasarla, Manish Kumar Singh, Hong Cai, Yunxiao Shi, Jisoo Jeong, Yinhao Zhu, Shizhong Han, Risheek Garrepalli, and Fatih Porikli. *Futuredepth: Learning to predict the future improves video depth estimation. In European Conference on Computer Vision*, pages 440–458. Springer, 2024. 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.488, + 0.483, + 0.543 + ], + "angle": 0, + "content": "[52] Tianwei Yin, Xingyi Zhou, and Philipp Krahenbuhl. Center-based 3d object detection and tracking. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 11784-11793, 2021. 1, 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.545, + 0.482, + 0.613 + ], + "angle": 0, + "content": "[53] Wenyuan Zeng, Wenjie Luo, Simon Suo, Abbas Sadat, Bin Yang, Sergio Casas, and Raquel Urtasun. End-to-end interpretable neural motion planner. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 8660-8669, 2019. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.615, + 0.482, + 0.683 + ], + "angle": 0, + "content": "[54] Jiang-Tian Zhai, Ze Feng, Jinhao Du, Yongqiang Mao, Jiang-Jiang Liu, Zichang Tan, Yifu Zhang, Xiaqing Ye, and Jingdong Wang. Rethinking the open-loop evaluation of end-to-end autonomous driving in nuscenes. arXiv preprint arXiv:2305.10430, 2023. 5, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.685, + 0.482, + 0.753 + ], + "angle": 0, + "content": "[55] Yunpeng Zhang, Zheng Zhu, Wenzhao Zheng, Junjie Huang, Guan Huang, Jie Zhou, and Jiwen Lu. Beverse: Unified perception and prediction in birds-eye-view for vision-centric autonomous driving. arXiv preprint arXiv:2205.09743, 2022. 2" + }, + { + "type": "list", + "bbox": [ + 0.093, + 0.093, + 0.483, + 0.753 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.48, + 0.945, + 0.518, + 0.956 + ], + "angle": 0, + "content": "27585" + } + ] +] \ No newline at end of file diff --git a/2025/Distilling Multi-modal Large Language Models for Autonomous Driving/2bea0a6d-f2ee-4f71-9581-707f2552ee73_origin.pdf b/2025/Distilling Multi-modal Large Language Models for Autonomous Driving/2bea0a6d-f2ee-4f71-9581-707f2552ee73_origin.pdf new file mode 100644 index 0000000000000000000000000000000000000000..9d8f19ed6d8e0acaac823f74f89ab42a7a471188 --- /dev/null +++ b/2025/Distilling Multi-modal Large Language Models for Autonomous Driving/2bea0a6d-f2ee-4f71-9581-707f2552ee73_origin.pdf @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:bc7f4cb7986ea39b27203b1fecb8f3658323be6233ff3e62a1e0d18eccb75749 +size 8040182 diff --git a/2025/Distilling Multi-modal Large Language Models for Autonomous Driving/full.md b/2025/Distilling Multi-modal Large Language Models for Autonomous Driving/full.md new file mode 100644 index 0000000000000000000000000000000000000000..a39065152ef3a229f6f6fe4a66334e9ebfbc9e00 --- /dev/null +++ b/2025/Distilling Multi-modal Large Language Models for Autonomous Driving/full.md @@ -0,0 +1,272 @@ +# Distilling Multi-modal Large Language Models for Autonomous Driving + +Deepti Hegde $^{1\dagger*}$ , Rajeev Yasarla $^{2*}$ , Hong Cai $^{2}$ , Shizhong Han $^{2}$ , Apratim Bhattacharyya $^{2}$ , Shweta Mahajan $^{2}$ , Litian Liu $^{2}$ , Risheek Garrepalli $^{2}$ , Vishal M. Patel $^{1}$ , Fatih Porikli $^{2}$ + +$^{1}$ Johns Hopkins University, $^{2}$ Qualcomm AI Research‡ *Equal contribution + +# Abstract + +Autonomous driving demands safe motion planning, especially in critical "long-tail" scenarios. Recent end-to-end autonomous driving systems leverage large language models (LLMs) as planners to improve generalizability to rare events. However, using LLMs at test time introduces high computational costs. To address this, we propose DiMA, an end-to-end autonomous driving system that maintains the efficiency of an LLM-free (or vision-based) planner while leveraging the world knowledge of an LLM. DiMA distills the information from a multi-modal LLM to a vision-based end-to-end planner through a set of specially designed surrogate tasks. Under a joint training strategy, a scene encoder common to both networks produces structured representations that are semantically grounded as well as aligned to the final planning objective. Notably, the LLM is optional at inference, enabling robust planning without compromising on efficiency. Training with DiMA results in a $37\%$ reduction in the L2 trajectory error and an $80\%$ reduction in the collision rate of the vision-based planner, as well as a $44\%$ trajectory error reduction in long-tail scenarios. DiMA also achieves state-of-the-art performance on the nuScenes planning benchmark. + +# 1. Introduction + +Research in autonomous driving has moved away from the traditional strategy of integrating sequential, independently trained models [4, 22, 26, 29, 31, 45, 52], and moved towards developing multi-task systems that are trained in an end-to-end manner [3, 5, 7, 16, 20, 49]. These approaches demonstrate improved performance and interpretability while being highly efficient. However, they struggle with long-tail navigation and perception scenarios, primarily because they are trained on task-specific, limited datasets [2, 11, 13]. + +Large language models (LLMs) have emerged as a promising solution to this issue. Trained on vast, internet- + +![](images/e9bc53db64c7aec28a15abcce8c7550c42f21a271c907e1cf3ff15a85ecf054c.jpg) + +![](images/7bfe5d2afb4b74a03c120aadb7df0997c6c5bc631c3cc8c96557dec693a18d3e.jpg) +Figure 1. Comparison of planning performance in long-tail scenarios from nuScenes: DiMA-VAD demonstrates greater robustness compared to VAD [20] in long-tail navigation scenarios such as overtaking a vehicle and performing a 3-point turn. DiMA-VAD also outperforms recent vision-planner PARA-Drive [48] and LLM planner TILEN [38]. Notably, the 3-point turn is a zero-shot scenario that is only present in the validation set. + +scale datasets, LLMs can leverage world knowledge to generalize to unseen or rare scenarios. These models can perform high-level reasoning tasks using mechanisms such as chain-of-thought [46]. Going beyond text-based prompting, multi-modal large language models (MLLMs) integrate image and video inputs to the LLM, enabling tasks such as visual question answering and dense captioning [21, 24]. Recent end-to-end autonomous driving systems leverage LLMs to achieve higher interpretability through language-guided reasoning and superior robustness to long-tail scenarios [19, 37-39, 43, 44]. To differentiate between these methods, we call end-to-end planners that depend on LLMs to perform trajectory prediction as "LLM-based" planners and ones that do not as "vision-based" planners [18, 20, 48]. In spite of their recent success "LLM-based" planners face significant challenges. + +Firstly, LLM-based planners require a significant amount of computational overhead at test time, limiting their practicality. This work addresses a core question: how can we harness LLMs' world knowledge while preserving the efficiency of vision-based planners? Secondly, bridging the visual and language domains is inherently more complex for end-to-end planning tasks compared to general MLLM tasks [21, 24], with the additional hurdle of limited training data. Standard image tokenization strategies rely on frozen pre-trained image encoders [33] to generate dense, unstructured multi-modal token embeddings. [21, 24, 37]. We suggest that MLLMs meant for end-to-end autonomous driving benefit from structured inputs that explicitly model scene components. Furthermore, we suggest that updating these features during training results in richer representations. + +To address these challenges, we propose DiMA, a novel framework for Distilling Multi-modal Large Language Models for Autonomous driving. We introduce a joint-training scheme between a vision-based planner and an MLLM that enables the learning of robust, grounded and disentangled scene representations that are aligned to the final objective of planning. Specifically, we use the vision-based planner as a tokenizer, and pass the learned scene representations to the MLLM, providing a more structured input. The MLLM is then trained for the tasks of visual question-answering, trajectory estimation, and a series of surrogate tasks designed to ground the multi-modal inputs in language. The vision-based planner is simultaneously trained for trajectory estimation, while also distilling features from the MLLM planning head to the planning transformer. Crucially, the MLLM can be discarded when performing planning inference, maintaining the efficiency of the vision-based planner while utilizing the knowledge of the language model. Optionally, the MLLM can support language-guided reasoning through visual question answering. Our experiments show that DiMA is more robust to challenging long-tail scenarios compared to baselines and state-of-the-art methods [38, 48]. Figure 1 confirms this both quantitatively and qualitatively. We list the contributions of the work below: + +1. We introduce DiMA, an end-to-end autonomous driving framework that distills knowledge from an MLLM to a vision-based planner to ensure robustness to long-tail events while maintaining efficiency. DiMA is capable of planning as well as visual question-answering. +2. We propose a distillation task along with the following surrogate tasks to align the objectives of the vision-based planner and the MLLM: (i) masked token reconstruction (ii) future token prediction (iii) scene editing. +3. DiMA outperforms both vision-based and LLM end-to-end planners, achieving state-of-the-art results on the nuScenes planning benchmark. Training with DiMA results in a $37\%$ reduction in the L2 trajectory error and an + +80% reduction in the collision rate of the vision-based planner, as well as a 44% trajectory error reduction in long-tail scenarios. + +# 2. Related Work + +# 2.1. End-to-End Autonomous Driving + +Autonomous driving (AD) systems are generally modular, consisting of tasks such as perception [22, 31, 45, 52], motion prediction [4, 26, 29], and planning [9]. Classic AD systems use standalone models for each task where each module is optimized separately and combined sequentially. Alternatively, some approaches incorporate these modular tasks into a multi-task learning paradigm, where tasks shared a common feature extraction process but have task-specific heads [8, 23, 53, 55]. Recently, research has moved toward optimizing autonomous driving systems in an end-to-end manner, aiming to directly predict future trajectories [3, 5, 7, 16, 20, 49]. Along this line of work, UniAD [18] introduces a query-based design that integrates perception and prediction components, enabling an end-to-end planning scheme. VAD [20] utilizes a vectorized scene representation, replacing the dense rasterized representations from [18], reducing computational cost and improving planning performance. While effective in general navigation scenarios, these methods struggle in difficult long-tail events (see Figure 1). In this work, we propose a method for end-to-end autonomous driving that is robust to such scenarios. + +# 2.2. LLMs for Autonomous Driving + +The recent success of large language models [12, 41, 42] demonstrates their ability to process complex contextual information with logical reasoning and to communicate in a human-interpretable manner, attracting significant attention in autonomous driving [6, 10, 14, 25, 28, 35-37, 39, 44, 47, 50]. These works investigate the capabilities of language models to generalize to novel scenarios as well as their ability to reason about the scene in the form of text. + +Prompting LLMs. One category of approaches involves prompting LLMs with text using reasoning frameworks such as chain-of-thought [46]. LanguageMPC [35] leverages the common-sense reasoning capacity of LLM to make high-level driving decisions, which are then translated into bottom-level control signals. GPT-Driver [28] converts observations and ego-states into language prompts, framing driving planning as a text-based task for LLMs. + +End-to-end Driving with MLLMs. Another approach is to integrate multi-modal LLMs into end-to-end train frameworks for autonomous vehicles. LMDrive [36] integrates a vision encoder with a large language model, enabling natural language instruction in autonomous driving. DriveLM introduces multi-step question answering, going beyond single-round interactions [27, 32, 34]. DriveVLM [39] leverages MLLMs and a chain-of-thought process to + +![](images/5283f6fe783ad4a42906f06c27fdf6a19d35035d65f2d988985427793a7e386e.jpg) +Figure 2. Overview of DiMA. The input to the framework is a multi-view image sequence and a question text prompt. The vision-based end-to-end planner consists of a scene encoder and a planning transformer. The scene encoder learns structured latent representations in the form of bird's-eye-view, ego, agent, and map (BEAM) token embeddings and acts as a trainable tokenizer for the multi-modal large language model (MLLM). The planning transformer is trained under standard planning constraints [18, 20]. The MLLM is trained for planning, visual question answering, distillation, and a series of surrogate tasks. + +enhance scene understanding and planning. + +We point out a challenge in the current MLLM strategy for autonomous driving. We suggest that in order to reason about vehicle dynamics, the language model benefits from a structured input that explicitly models scene components instead of dense, unstructured input to the MLLM as in [36, 39, 43, 44]. In TOKEN [38], the scene encoder from [48] is used as a tokenizer. However, it is frozen during training. In our approach, the input to the MLLM is a set of latent representations learned from a pre-trained scene encoder that is jointly trained with the MLLM, allowing it to more learn grounded feature representations + +# 3. DiMA Framework + +We present DiMA, a framework for end-to-end autonomous driving. Given a sequence of multi-view images, the overall objective is to predict the future trajectory of the ego vehicle and answer questions about the scene. An overview of DiMA is shown in Figure 2. The framework has two main components: 1) the vision-based planner; 2) a multi-modal large language model consisting of adapter layers, an LLM, and a series of task-specific decoder heads. + +# 3.1. Vision-based Planner + +Vision-based end-to-end planners are trained in a multi-task fashion to perform perception, mapping, motion prediction and planning [18, 20, 40, 48]. Considering standard architecture design from [18, 20, 48] we decompose the planner into the scene encoder and the planning transformer. The scene encoder provides structured latent representations to the planning transformer that performs waypoint prediction. We call these representations as scene token embeddings. In our framework, we leverage the vision-based planner in + +two ways. First, it acts as the primary network through which planning is performed, allowing for fast prediction at test time. Second, the scene encoder is shared with the MLLM, acting as tokenizer to provide it with a highly structured set of inputs. An important distinction of our framework from previous works [19, 38] is that the scene encoder is trained jointly with the MLLM. + +Scene encoder The scene encoder models the scene as high-dimensional token embeddings that explicitly represent components such as the environment map, the ego vehicle, and the surrounding agents. This is achieved by introducing task specific, learnable query features that are supervised by a series of planning constraints [20]. First, a visual backbone encodes the input multi-view image sequence. The obtained feature map is projected into the bird's-eye-view (BEV) space using a set of queries to obtain BEV token embeddings $(B)$ [22]. The network learns structured representations of the map and agents in the form of map token embeddings $(M)$ and agent token $(A)$ embeddings by cross-attending the BEV features with map and agent queries. A randomly initialized learnable embedding is trained to learn the interaction of the ego vehicle with the agent and map components which is the ego query $(E)$ . We denote the set of these latent scene representations as $BEAM$ token embeddings. These token embeddings are sent to a planning head that predicts future trajectories of the ego and agent vehicles. They are also used as input to the MLLM as multi-modal inputs, as discussed next. + +# 3.2. Multi-modal LLM + +Our objective is to distill knowledge from an MLLM to the vision-based planner. Towards this, we jointly train an MLLM with the vision-based planner. Specifically, the + +scene encoder is used as a trainable tokenizer for the MLLM to generate the $BEAM$ token embeddings, as well as used as an encoder for the planning transformer. The advantage from this strategy is two-fold. First, the MLLM receives a highly structured input that captures rich spatio-temporal information relevant to the autonomous driving task. Second, the scene encoder learns features that are grounded in language, improving the robustness of vision-based planning. The main components of the MLLM are adapter layers, the large language model, and a series of task-specific decoder heads which we discuss next. + +# 3.2.1. Adaptation of Scene Tokens + +In order to project the BEAM token embeddings efficiently while maintaining their distinctiveness, we leverage the query-transformer (Q-former) module [17] to compress the visual tokens before being input to the LLM. We implement component-specific Q-former layers, namely a Map Q-former, a Bird's-Eye-View (BEV) Q-former, an Ego Q-former, and an Agent Q-former. Each of these adapters transforms the representing different modalities into a sequence of fixed-length, higher-dimensional tokens, preparing them for processing by the MLLM. + +# 3.2.2. MLLM Supervision + +We design tasks for the MLLM with the objectives of, a) enriching the intermediate scene representations; b) grounding the scene token embeddings in language; and c) training the LLM for planning-related reasoning. In detail, the MLLM is trained for visual question answering, trajectory estimation, feature distillation, and a set of surrogate tasks [1, 51]. + +Visual question answering. The MLLM is trained for visual question answering (VQA) on question-answer pairs of 4 types: perception of the scene, prediction of the agent behavior, prediction of ego-vehicle behavior, and future planning steps [37]. The answer is predicted based on a multimodal prompt consisting of the question embeddings and the projected BEAM token embeddings. This is constructed as a standard next-token prediction task and the VQA branch is supervised with a cross-entropy loss as in [21]. We denote the loss from this task head as $\mathcal{L}_{LLM}$ . + +Surrogate tasks: The main goal of joint training is to enrich the BEAM scene representations learned by the scene encoder. We design MLLM surrogate tasks whose objectives are in line with that of future trajectory prediction. + +1) Masked token reconstruction: Each type of token embedding contributes to a holistic representation of the scene. In order to enrich the visual representations, we ask the network to reconstruct a masked BEV input based on the context present in the rest of the multi-modal sequence. We perform random masking after scene encoding and pass the token embeddings to the MLLM. A reconstruction head takes the latent representations from the penultimate layer of LLM and predicts reconstructed BEV token embeddings + +![](images/14389e201993856381630025fdf42676d5dd9c81bbdd68528ae5fc3cd0142134.jpg) +Figure 3. Examples of addition and deletion in scene editing. In the top row, a car (solid red box) is added on the left of the ego-vehicle (green box). In the bottom row, a car (dashed red box) is removed from the right of the ego-vehicle. A corresponding question-answer pair is created to characterize the edit. + +$\hat{B}$ . This decoder head is supervised with the L2 loss between the prediction and the complete input, + +$$ +\mathcal {L} _ {\text {r e c o n}} = \left\| \hat {B} - B \right\| ^ {2} \tag {1} +$$ + +where $(\{m(B), E, A, M\})$ is input to the MLLM, and the latent MLLM representations associated with the BEV token embeddings are input to the masked reconstruction decoder head. Here, $m(.)$ denotes random masking. + +2) Future BEV prediction: An important aspect of planning is anticipating future events. We introduce the surrogate task of future BEV prediction to encourage the LLM to learn spatio-temporal cues useful for planning. Given latent BEV token embeddings, we train a prediction head to predict future BEV token embeddings and supervise this task as the L2 loss between predicted and ground truth future token embeddings: + +$$ +\mathcal {L} _ {\text {f u t u r e}} = \left\| \hat {F} _ {t} - B _ {t + 1} \right\| ^ {2} + \left\| \tilde {F} _ {t} - B _ {t + 2} \right\| ^ {2} \tag {2} +$$ + +where $\hat{F}_t, \tilde{F}_t$ are the predicted future BEV token embeddings at time $t$ . Note, we predict future BEV token embeddings $(B_{t+1}, B_{t+2})$ of the multi-view image sequence for time steps $\{t+1, t+2\}$ . + +3) Scene editing: For prediction and reasoning about the ego vehicle, it is crucial to learn how surrounding agents impact the ego-vehicle's future path. We propose a novel scene editing task in which we augment scenes by removing or adding new agents. Along with this, we construct a question-answer pair related to the edit. We show examples of this in Figure 3. For scene addition, given the map constraints, predicted map, the ego bounding box location and the trajectories of predicted agents, we create a way-point trajectory for a new agent of category "car" or "truck". A + +new agent token embedding is then created using a linear layer. This new agent token embedding, the corresponding text prompt, and the rest of the BEAM token embeddings are passed as input to the LLM. The hidden latent LLM features corresponding to agent token embeddings are then fed into a dedicated scene editing decoder head that performs waypoint prediction of the ego vehicle. The language prediction head performs question-answering on the new QA pair. This task thus contributes to the existing planning constraint loss and VQA loss of the MLLM. + +Distillation In order to further align the representations learned by the vision planner and the MLLM, we perform knowledge transfer between the penultimate layers of the language model and the planning transformer. Concretely, we minimize the KL-divergence between the distributions of the hidden features of the planning transformer and the LLM hidden features of the ego-token embeddings: + +$$ +\mathcal {L} _ {\text {d i s t i l l}} = D _ {K L} \left(P _ {l l m} \| P _ {v i s}\right) \tag {3} +$$ + +where the $P_{vis}$ are the features of penultimate layers of vision planning transformer and $P_{llm}$ are the hidden LLM embeddings features corresponding to the ego-token embeddings of penultimate layers of MLLM model. + +Loss functions. We supervise the network with a weighted sum of the losses corresponding to planning, visual question answering, distillation, and each surrogate task. + +$$ +\mathcal {L} = \mathcal {L} _ {\text {p l a n n i n g}} + \mathcal {L} _ {\text {L L M}} + \mathcal {L} _ {\text {r e c o n}} + \mathcal {L} _ {\text {f u t u r e}} + \mathcal {L} _ {\text {d i s t i l l}} \tag {4} +$$ + +here $\mathcal{L}_{\text{planning}}$ comes from the planning objective used to train [18, 20]. Here, the loss weights are chosen to bring each value to the same scale. + +# 4. Experimental Setup + +In Section 4.1, we present the architecture design details, in Section 4.2 we present the training strategy, and in Section 4.3, we provide information about the datasets used. Importantly, in Section 4.4 we detail multiple evaluation strategies for fair and comprehensive comparison with existing works. + +# 4.1. Architecture design + +Vision-based planner We conduct experiments using two vision-based end-to-end planners, VAD [20] and UniAD [18] due to their performance, efficiency, and ability to model the interaction between scene components. We use two model size variants of VAD. Both planners are trained for perception, motion prediction and planning, while UniAD also performs occupancy prediction. + +MLLM design The MLLM is made up of adapter layers, a language model, and a set of task-specific decoder layers. We project each $BEAM$ token embedding with its own dedicated Q-former adapter layer following [17]. After projection, the $BEAM$ token embeddings and language token embeddings lie in the same embedding space. In order to reduce memory consumption, we limit the agent token se + +quence length to a fraction of the total number of agents and perform sequence-wise upsampling after input to the language model. The adapter layers are randomly initialized. We use the LLM from LLaVA-v1.5-7B [24] as our language model base. The ego prediction and agent motion prediction task heads are multi-layer-perceptron networks following [20]. For the surrogate task decoder heads, we use 3 Linear layers with a ReLU activation layer. + +# 4.2. Training + +DiMA training follows a two-stage approach. First the vision-only planner is pre-trained for perception, prediction, and planning for 60 epochs in order to learn informative latent scene representations. Second, we perform joint training of the vision planner and the MLLM for an additional 30 epochs, incorporating all proposed tasks and losses detailed in Section 3. In the second stage, the language model of the MLLM is fine-tuned using LoRA [15]. Please refer to the supplementary material for more details. + +# 4.3. Datasets + +We use the nuScenes dataset for the task of open-loop planning [2]. This dataset consists of 28k total samples in a $22\mathrm{k} / 6\mathrm{k}$ training/Validation split. The objects in each scene are annotated with 3D bounding box, orientation, and vehicle speed information. Additionally, we make use of the CAN bus information for the ego vehicle's ground-truth trajectory annotations. For supervising visual question answering, we train with DriveLM [37]. This dataset consists of a 4k subset of samples from the nuScenes dataset annotated with 300k QA pairs related to perception, prediction, planning, and behavior of the ego vehicle. For the samples in nuScenes that do not have text annotations, we create a set of perception, planning, and prediction QA pairs based on the numerical annotations. We use this information to prompt a Llama3-70B model [12] with in-context examples to generate DriveLM-like QA pairs. For the behavior QA pair, we create a rule-based algorithm to categorize the future movement of the ego-vehicle based on the ground-truth trajectory values and ego vehicle speed. More details on this method as well as examples of generated QA pairs may be found in the supplementary. + +# 4.4. Evaluation details + +We evaluate planning performance based on the predicted future trajectories of the ego vehicle 3 seconds into the future, where 2 waypoints are predicted per second. We use two metrics, the L2 error (in meters) between the predicted and ground-truth waypoints as well as the collision rate (in $\%$ ) between the ego-vehicle and the surrounding vehicles. + +Standardized vs VAD evaluation. In [48], Weng et al. point out inconsistencies in planning evaluation across VAD [20], UniAD [18] and AD-MLP [54], namely in the way L2 error is averaged over time and the way noisy and invalid + +Table 1. Comparison of L2 trajectory error and collision rate on nuScenes [2] using standardized evaluation [48]. Models are evaluated on the general validation split as well as a "targeted" split of challenging samples from [48]. The performance of the DiMA model variants are in shades of purple. We summarize results by averaging over at $t = \{1,2,3\} s$ as well as at all time steps. The best average performance in each setting is in bold. "+" indicates the use of ego-status information. + +
MethodUsing Ego statusTraj L2 (m) ↓Collision (%) ↓
1s2s3sAve1,2,3sAve all1s2s3sAve1,2,3sAve all
Full validation split
UniAD[18]X0.480.891.470.950.830.320.290.730.450.40
VAD-Base [20]X0.410.861.460.910.780.020.260.830.370.30
PARA-Drive[48]X0.260.591.120.660.560.000.120.650.260.17
TOKEN[38]X0.260.701.460.810.68----0.15
DiMA (VAD-Tiny)X0.200.531.100.610.510.000.070.190.090.06
DiMA (VAD-Base)X0.180.501.020.570.470.000.050.160.070.06
AD-MLP[54]✓0.230.581.180.660.560.000.140.700.280.20
PARA-Drive+ [48]✓0.200.521.040.590.490.000.090.490.190.13
DiMA+ (VAD-Tiny)✓0.190.511.060.590.490.000.060.160.080.06
DiMA+ (VAD-Base)✓0.180.481.010.560.460.000.050.160.070.06
DiMA-Dual+ (VAD-Tiny)✓0.180.501.030.570.470.000.050.160.080.06
Targeted validation split
UniAD[18]X0.471.091.921.160.990.000.000.730.240.15
VAD-Base [20]X0.531.202.071.271.080.000.290.870.390.34
PARA-Drive[48]X0.380.971.881.080.910.000.000.720.240.14
DiMA (VAD-Tiny)X0.320.881.700.970.810.000.030.240.090.05
DiMA (VAD-Base)X0.290.791.490.830.710.000.030.220.080.05
AD-MLP[54]✓0.330.992.061.130.940.000.583.621.400.94
PARA-Drive+ [48]✓0.290.761.460.830.700.000.000.290.100.05
DiMA+ (VAD-Tiny)✓0.310.831.670.940.770.000.040.170.070.05
DiMA+ (VAD-Base)✓0.270.741.380.790.640.000.030.160.060.04
DiMA-Dual+ (VAD-Tiny)✓0.280.761.490.840.690.000.030.170.070.05
+ +frames are considered. The authors also point out that collision performance is considerably improved by using a more finely discretized BEV map. Accounting for these inconsistencies, they propose a "standardized" evaluation metric to maintain a fair comparison across methods. For accurate reporting and fair comparison with [38, 48], we use this standardized metric to evaluate our model. Certain existing works use the evaluation scheme followed by [20] (which we call VAD evaluation) and do not make code or models available [16, 30, 39, 43]. Here, we compare reported results against DiMA evaluated using VAD evaluation. + +Validation splits. We evaluate planning performance across 3 types of splits from the nuScenes validation set: 1) the entire validation set of 6019 samples 2) a 689 sample sub-set more difficult "targeted" scenarios in which the ego-vehicle turns during navigation [48], and 3) long-tail events in which the ego-vehicle encounters rare maneuvers [38]. u + +# 5. Experimental Results + +We perform a comprehensive evaluation of DiMA on the nuScenes open-loop planning benchmark in a variety of evaluation settings. We compare our performance against that of recent vision-based planners [16, 18, 20, 48] as well as recent MLLM planners [30, 38, 39, 43]. To facilitate a + +fair comparison of DiMA with recent works, we use two evaluation strategies VAD evaluation [20] and standardized evaluation [48]. Please refer to Section 4.4 for more details. + +Standardized evaluation. Under the standardized evaluation strategy, we evaluate each method on two validation splits: 1) the full nuScenes validation split, and 2) a "targeted" subset of the nuScenes validation split that contains scenarios where the ego-vehicle must take a turn, as established in [48]. We present our results in Table 1. The baseline vision planners VAD [20] and UniAD [18] exhibit suboptimal performance in more difficult targeted scenarios. DiMA demonstrates significant improvements over baseline methods, achieving a $35\%$ reduction in L2 trajectory error and an $80\%$ reduction in collision rate on the overall nuScenes validation set, and a $35\%$ reduction in L2 trajectory error in the "targeted" split. More notably, we outperform recent state-of-the-art methods PARA-Drive [48] and TOKEN [38] consistently across both validation splits. We observe a strong improvement in collision rate, improving on that of TOKEN by $60\%$ . Unlike TOKEN, our approach does not rely on an LLM for inference, thus making DiMA more efficient as well as more accurate. + +VAD evaluation. For a fair comparison with recent works that use the evaluation strategy from VAD [20], we present + +Table 2. Comparison of L2 trajectory error and collision rate on nuScenes [2] using VAD evaluation [20]. Models are evaluated on the general validation split. The performance of the DiMA model variants are in shades of purple. The performance of the MLLM-branch DiMA model variant is in blue. We summarize results by averaging over all time steps. The best average performance in each setting is in bold. “+” indicates the use of ego-status information. + +
MethodUsing Ego statusTraj L2 (m) ↓Collision (%) ↓Latency (ms)FPS
1s2s3sAveall1s2s3sAveall
ST-P3 [16]X1.332.112.902.110.230.621.270.71628.31.6
UniAD[18]X0.480.961.651.030.050.170.710.31555.61.8
VAD-Tiny[20]X0.460.761.120.780.210.350.580.3859.516.8
VAD-Base[20]X0.410.701.050.720.070.170.410.22224.34.5
VLP-VAD-Base[30]X0.260.470.780.500.120.170.420.23--
OmniDrive [43]X0.400.801.320.840.040.462.320.94--
DiMA (UniAD)X0.150.300.560.340.060.080.220.125601.8
DiMA (VAD-Tiny)X0.180.360.610.380.070.100.270.1559.516.8
DiMA (VAD-Base)X0.130.270.470.290.050.080.160.102264.5
VAD-Tiny✓0.200.380.650.410.100.120.270.1659.516.8
VAD-Base✓0.170.340.600.370.070.100.240.14224.34.5
DriveVLM [39]✓0.180.340.680.400.100.220.450.27--
DriveVLM-Dual (VAD-Base) [39]✓0.150.290.480.310.050.080.170.10--
DiMA + (VAD-Tiny)✓0.170.330.590.370.070.100.260.1459.516.8
DiMA + (VAD-Base)✓0.120.250.440.270.040.060.150.082264.5
DiMA-Dual+ (VAD-Tiny)✓0.140.270.460.290.050.070.150.092863.5
+ +Table 3. Long-tail performance comparison of L2 trajectory error and collision rate on nuScenes [2] validation set using standardized evaluation [48]. The performance of vision planner DiMA model variants are in shades of purple. The performance of MLLM-branch DiMA model variants are in shades of blue. We summarize results by averaging over at $t = \{1,2,3\} s$ as well as at all time steps. The best average performance in each setting is in bold. + +
MethodTraj L2 (m)↓Coll. (%)↓
1s2s3sAve1,2,3sAveallAveall
3-point turn (zero-shot)
VAD-Tiny0.761.683.041.831.550.00
VAD-Base0.711.663.241.871.570.00
PARA-Drive [48]0.501.382.761.551.295.33
TOKEN [38]0.391.292.601.431.184.00
DiMA (VAD-Tiny)0.471.272.551.431.170.00
DiMA (VAD-Base)0.361.182.371.301.050.00
DiMA (MLLM)0.381.122.351.281.040.00
DiMA-Dual (VAD-Tiny)0.381.082.351.271.040.00
Resume from stop
VAD-Tiny0.641.632.991.751.490.00
VAD-Base0.601.722.831.721.420.00
PARA-Drive0.140.792.301.080.850.00
TOKEN0.130.701.580.800.650.00
DiMA (VAD-Tiny)0.261.022.401.230.990.00
DiMA (VAD-Base)0.150.651.340.710.660.00
DiMA (MLLM)0.241.022.161.140.930.00
DiMA-Dual (VAD-Tiny)0.240.982.131.110.910.00
Overtake
VAD-Tiny0.581.272.121.321.142.42
VAD-Base0.461.162.171.261.062.49
PARA-Drive0.270.891.941.030.852.30
TOKEN0.290.771.630.900.740.00
DiMA (VAD-Tiny)0.240.751.490.830.691.32
DiMA (VAD-Base)0.240.721.500.820.661.29
DiMA (MLLM)0.240.731.500.820.671.30
DiMA-Dual (VAD-Tiny)0.240.731.500.820.671.30
+ +an additional set of results under this strategy. In Table 2, we present the evaluation of DiMA on the nuScenes validation set [2]. We significantly outperform the base + +line vision-based planners VAD [20] and UniAD [18]. Notably, DiMA built on VAD-Tiny outperforms VAD-Base by $47\%$ in L2 trajectory error while being 4 times faster. We also outperform MLLM planners like OmniDrive [43] and DriveVLM [39] without requiring an LLM at inference, making our approach more accurate as well as efficient. + +Performance in long-tail scenarios. Autonomous driving performance is particularly crucial for rare scenarios. Using the manually selected long-tail events from [38], we demonstrate the robustness of DiMA in cases of novel navigation and perception. In Table 3, we evaluate the performance of DiMA for three long-tail scenarios, "3-point turn", "resume from the stop" and "overtake", selected from the nuScenes validation set. Training with DiMA results in a significant and consistent boost in performance from the baseline vision-based planners. We also consistently outperform PARA-Drive and TOKEN. Notably, the "3-point turn" case is a zero-shot scenario, not present in the training data. We achieve the lowest L2 trajectory error for this case. + +Evaluating the MLLM branch. We also evaluate the MLLM branch of DiMA and demonstrate its planning and visual question-answering capabilities. We perform LLM-based planning using DiMA in two ways. First, we directly evaluate the waypoint predictions of the MLLM planning decoder head. This is denoted by DiMA (MLLM). Secondly, we define DiMA-Dual, an optional hybrid inference model that combines the strengths of both the vision and MLLM branches. We perform max-pooling on the penultimate layer features of both vision planning transformer and MLLM ego prediction. The resulting pooled features are then fed back into both the vision planning transformer and + +![](images/50f3c2512a4f9267ad7bb8e17c10129bd313a222693b89a5732a20458a67adf3.jpg) +Figure 4. Visualization of planning performance by the MLLM branch of DiMA-VAD-Tiny. We plot the predicted trajectory (orange-yellow) and show an example response of the language model branch to a question from the DriveLM test dataset [37]. + +![](images/f324fc1565c9d9a416294862fac9f1ca3576c60912344d459c0d2fcf234c3f9a.jpg) + +Table 4. Results of ablation experiments of DiMA-VAD-Tiny on nuScenes under VAD evaluation. The baseline model is VAD-Tiny. We report L2 trajectory error and collision rate averaged over all time steps. + +
IDVQAScene tokensDistillationLLM PlanningSurrogate tasksTraj L2 (m) ↓Collision (%) ↓
Masked recon.Future pred.Scene editing1s2s3sAveall1s2s3sAveall
1XXXXXXX0.330.580.890.600.120.190.550.29
2✓BEVX✓XXX0.360.600.910.620.100.170.520.26
3✓BEV, MapX✓XXX0.280.520.870.560.100.170.490.25
4✓AllX✓XXX0.260.510.830.520.100.160.380.21
5✓All✓✓XXX0.220.440.770.480.090.140.340.19
6✓All✓✓✓XX0.190.390.680.420.090.130.320.18
7✓All✓✓✓✓X0.180.370.630.390.070.110.290.16
8✓All✓✓✓✓✓0.180.360.610.380.070.100.270.15
+ +MLLM ego prediction for waypoint prediction. This max pooling encourages consistency between the predictions of the two models, leading to a more robust and potentially improved overall prediction as seen in Table 3. Using the MLLM at inference results in a significant boost in performance, with DiMA-Dual (VAD-Tiny) matching the performance of DiMA (VAD-Base) in Table 2. + +VQA results. We provide qualitative examples of visual question-answering in the DriveLM dataset [37] in Figure 4. We demonstrate the ability of DiMA to reason about safe actions to be taken by the ego vehicle. We provide a more examples in the supplementary material. + +# 5.1. Ablation study. + +We examine the role of different design choices and proposed tasks in the DiMA framework through a comprehensive ablation study in Table 4. In order to do so, we build our final frame-work step-by-step. We begin with VAD-Tiny as our baseline model in the first row, denoted by ID-1. In ID-2, we observe inconsistent gains and drops in performance by naively training an LLM with the VAD-Tiny under a the planning and VQA objectives while using only BEV features. By comparing ID-3 and ID-4, we show that that providing more scene context to the MLLM by using all the BEAM token embeddings to train both LLM and + +vision-based planner benefits the performance, resulting a significant improvement in trajectory estimation and collision rate. We show in ID-5 that explicitly distilling the information from the penultimate layers of the MLLM to the planning transformer gives another boost in performance. In ID-6, ID-7 and ID-8 we demonstrate the importance of surrogate tasks like masked-reconstruction, future prediction and scene editing in helping the scene encoder to learn richer representations useful for planning. We point to the supplementary material for additional ablation studies. + +# 6. Conclusion + +We introduce DiMA, an end-to-end autonomous driving framework for robust and efficient planning. DiMA distills knowledge from an MLLM to a vision-based planner, employing a novel joint training strategy alongside carefully designed surrogate tasks such as masked token reconstruction, future token prediction, and scene editing. DiMA enables the model to learn semantically grounded scene representations, improving its performance in difficult navigation scenarios. We conduct extensive evaluations and ablation studies, showing that training with DiMA results in a planner that excels in long-tail cases while maintaining computational efficiency. + +# References + +[1] Apratim Bhattacharyya, Sunny Panchal, Reza Pourreza, Mingu Lee, Pulkit Madan, and Roland Memisevic. Look, remember and reason: Grounded reasoning in videos with language models. In International Conference of Learning Representations, 2024. 4 +[2] Holger Caesar, Varun Bankiti, Alex H Lang, Sourabh Vora, Venice Erin Liong, Qiang Xu, Anush Krishnan, Yu Pan, Giancarlo Baldan, and Oscar Beijbom. nuscenes: A multimodal dataset for autonomous driving. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 11621-11631, 2020. 1, 5, 6, 7 +[3] Sergio Casas, Abbas Sadat, and Raquel Urtasun. Mp3: A unified model to map, perceive, predict and plan. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 14403-14412, 2021. 1, 2 +[4] Yuning Chai, Benjamin Sapp, Mayank Bansal, and Dragomir Anguelov. Multipath: Multiple probabilistic anchor trajectory hypotheses for behavior prediction. arXiv preprint arXiv:1910.05449, 2019. 1, 2 +[5] Dian Chen and Philipp Krahenbuhl. Learning from all vehicles. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 17222-17231, 2022. 1, 2 +[6] Long Chen, Oleg Sinavski, Jan Hünermann, Alice Karnsund, Andrew James Willmott, Danny Birch, Daniel Maund, and Jamie Shotton. Driving with llms: Fusing object-level vector modality for explainable autonomous driving. In 2024 IEEE International Conference on Robotics and Automation (ICRA), pages 14093-14100. IEEE, 2024. 2 +[7] Kashyap Chitta, Aditya Prakash, and Andreas Geiger. Neat: Neural attention fields for end-to-end autonomous driving. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 15793-15803, 2021. 1, 2 +[8] Kashyap Chitta, Aditya Prakash, Bernhard Jaeger, Zehao Yu, Katrin Renz, and Andreas Geiger. Transfuser: Imitation with transformer-based sensor fusion for autonomous driving. IEEE Transactions on Pattern Analysis and Machine Intelligence, 45(11):12878-12895, 2022. 2 +[9] Felipe Codevilla, Eder Santana, Antonio M López, and Adrien Gaidon. Exploring the limitations of behavior cloning for autonomous driving. In Proceedings of the IEEE/CVF international conference on computer vision, pages 9329-9338, 2019. 2 +[10] Xinpeng Ding, Jianhua Han, Hang Xu, Wei Zhang, and Xiaomeng Li. Hilm-d: Towards high-resolution understanding in multimodal large language models for autonomous driving. arXiv preprint arXiv:2309.05186, 2023. 2 +[11] Alexey Dosovitskiy, German Ros, Felipe Codevilla, Antonio Lopez, and Vladlen Koltun. Carla: An open urban driving simulator. In Conference on robot learning, pages 1-16. PMLR, 2017. 1 +[12] Abhimanyu Dubey, Abhinav Jauhri, Abhinav Pandey, Abhishek Kadian, Ahmad Al-Dahle, Aiesha Letman, Akhil Mathur, Alan Schelten, Amy Yang, Angela Fan, et al. The llama 3 herd of models. arXiv preprint arXiv:2407.21783, 2024. 2, 5 + +[13] Scott Ettinger, Shuyang Cheng, Benjamin Caine, Chenxi Liu, Hang Zhao, Sabeek Pradhan, Yuning Chai, Ben Sapp, Charles R Qi, Yin Zhou, et al. Large scale interactive motion forecasting for autonomous driving: The waymo open motion dataset. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 9710-9719, 2021. 1 +[14] Daocheng Fu, Xin Li, Licheng Wen, Min Dou, Pinlong Cai, Botian Shi, and Yu Qiao. Drive like a human: Rethinking autonomous driving with large language models. In Proceedings of the IEEE/CVF Winter Conference on Applications of Computer Vision, pages 910-919, 2024. 2 +[15] Edward J Hu, Yelong Shen, Phillip Wallis, Zeyuan Allen-Zhu, Yuanzhi Li, Shean Wang, Lu Wang, and Weizhu Chen. Lora: Low-rank adaptation of large language models. arXiv preprint arXiv:2106.09685, 2021. 5 +[16] Shengchao Hu, Li Chen, Penghao Wu, Hongyang Li, Junchi Yan, and Dacheng Tao. St-p3: End-to-end vision-based autonomous driving via spatial-temporal feature learning. In European Conference on Computer Vision, pages 533-549. Springer, 2022. 1, 2, 6, 7 +[17] Wenbo Hu, Zi-Yi Dou, Liunian Harold Li, Anita Kamath, Nanyun Peng, and Kai-Wei Chang. Matryoshka query transformer for large vision-language models, 2024. 4, 5 +[18] Yihan Hu, Jiazhi Yang, Li Chen, Keyu Li, Chonghao Sima, Xizhou Zhu, Siqi Chai, Senyao Du, Tianwei Lin, Wenhai Wang, et al. Planning-oriented autonomous driving. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 17853-17862, 2023. 1, 2, 3, 5, 6, 7 +[19] Jyh-Jing Hwang, Runsheng Xu, Hubert Lin, Wei-Chih Hung, Jingwei Ji, Kristy Choi, Di Huang, Tong He, Paul Covington, Benjamin Sapp, et al. Emma: End-to-end multimodal model for autonomous driving. arXiv preprint arXiv:2410.23262, 2024. 1, 3 +[20] Bo Jiang, Shaoyu Chen, Qing Xu, Bencheng Liao, Jiajie Chen, Helong Zhou, Qian Zhang, Wenyu Liu, Chang Huang, and Xinggang Wang. Vad: Vectorized scene representation for efficient autonomous driving. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 8340-8350, 2023. 1, 2, 3, 5, 6, 7 +[21] Junnan Li, Dongxu Li, Silvio Savarese, and Steven Hoi. Blip-2: Bootstrapping language-image pre-training with frozen image encoders and large language models. In International conference on machine learning, pages 19730-19742. PMLR, 2023. 1, 2, 4 +[22] Zhiqi Li, Wenhai Wang, Hongyang Li, Enze Xie, Chonghao Sima, Tong Lu, Yu Qiao, and Jifeng Dai. Bevformer: Learning bird's-eye-view representation from multi-camera images via spatiotemporal transformers. In European conference on computer vision, pages 1-18. Springer, 2022. 1, 2, 3 +[23] Xiwen Liang, Yangxin Wu, Jianhua Han, Hang Xu, Chunjing Xu, and Xiaodan Liang. Effective adaptation in multitask co-training for unified autonomous driving. Advances in Neural Information Processing Systems, 35:19645-19658, 2022. 2 +[24] Haotian Liu, Chunyuan Li, Qingyang Wu, and Yong Jae Lee. + +Visual instruction tuning. Advances in neural information processing systems, 36, 2024. 1, 2, 5 +[25] Jiaqi Liu, Peng Hang, Xiao Qi, Jianqiang Wang, and Jian Sun. Mtd-gpt: A multi-task decision-making gpt model for autonomous driving at unsignalized intersections. In 2023 IEEE 26th International Conference on Intelligent Transportation Systems (ITSC), pages 5154-5161. IEEE, 2023. 2 +[26] Yicheng Liu, Jinghuai Zhang, Liangji Fang, Qinhong Jiang, and Bolei Zhou. Multimodal motion prediction with stacked transformers. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 7577-7586, 2021. 1, 2 +[27] Srikanth Malla, Chiho Choi, Isht Dwivedi, Joon Hee Choi, and Jiachen Li. Drama: Joint risk localization and captioning in driving. In Proceedings of the IEEE/CVF winter conference on applications of computer vision, pages 1043-1052, 2023. 2 +[28] Jiageng Mao, Yuxi Qian, Junjie Ye, Hang Zhao, and Yue Wang. Gpt-driver: Learning to drive with gpt. arXiv preprint arXiv:2310.01415, 2023. 2 +[29] Jiquan Ngiam, Benjamin Caine, Vijay Vasudevan, Zhengdong Zhang, Hao-Tien Lewis Chiang, Jeffrey Ling, Rebecca Roelofs, Alex Bewley, Chenxi Liu, Ashish Venugopal, et al. Scene transformer: A unified architecture for predicting multiple agent trajectories. arXiv preprint arXiv:2106.08417, 2021. 1, 2 +[30] Chenbin Pan, Burhaneddin Yaman, Tommaso Nesti, Abhirup Mallik, Alessandro G Allievi, Senem Velipasalar, and Liu Ren. Vlp: Vision language planning for autonomous driving. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 14760-14769, 2024. 6, 7 +[31] Jonah Philion and Sanja Fidler. Lift, splat, shoot: Encoding images from arbitrary camera rigs by implicitly unprojecting to 3d. In Computer Vision-ECCV 2020: 16th European Conference, Glasgow, UK, August 23-28, 2020, Proceedings, Part XIV 16, pages 194-210. Springer, 2020. 1, 2 +[32] Tianwen Qian, Jingjing Chen, Linhai Zhuo, Yang Jiao, and Yu-Gang Jiang. Nuscenes-qa: A multi-modal visual question answering benchmark for autonomous driving scenario. In Proceedings of the AAAI Conference on Artificial Intelligence, pages 4542–4550, 2024. 2 +[33] Alec Radford, Jong Wook Kim, Chris Hallacy, Aditya Ramesh, Gabriel Goh, Sandhini Agarwal, Girish Sastry, Amanda Askell, Pamela Mishkin, Jack Clark, et al. Learning transferable visual models from natural language supervision. In International conference on machine learning, pages 8748-8763. PMLR, 2021. 2 +[34] Enna Sachdeva, Nakul Agarwal, Suhas Chundi, Sean Roelofs, Jiachen Li, Mykel Kochenderfer, Chiho Choi, and Behzad Dariush. Rank2tell: A multimodal driving dataset for joint importance ranking and reasoning. In Proceedings of the IEEE/CVF Winter Conference on Applications of Computer Vision, pages 7513-7522, 2024. 2 +[35] Hao Sha, Yao Mu, Yuxuan Jiang, Li Chen, Chenfeng Xu, Ping Luo, Shengbo Eben Li, Masayoshi Tomizuka, Wei + +Zhan, and Mingyu Ding. *Languagempc: Large language models as decision makers for autonomous driving.* arXiv preprint arXiv:2310.03026, 2023. 2 +[36] Hao Shao, Yuxuan Hu, Letian Wang, Guanglu Song, Steven L Waslander, Yu Liu, and Hongsheng Li. Lmdrive: Closed-loop end-to-end driving with large language models. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 15120-15130, 2024. 2, 3 +[37] Chonghao Sima, Katrin Renz, Kashyap Chitta, Li Chen, Hanxue Zhang, Chengen Xie, Ping Luo, Andreas Geiger, and Hongyang Li. Drivelm: Driving with graph visual question answering. arXiv preprint arXiv:2312.14150, 2023. 1, 2, 4, 5, 8 +[38] Ran Tian, Boyi Li, Xinshuo Weng, Yuxiao Chen, Edward Schmerling, Yue Wang, Boris Ivanovic, and Marco Pavone. Tokenize the world into object-level knowledge to address long-tail events in autonomous driving. arXiv preprint arXiv:2407.00959, 2024. 1, 2, 3, 6, 7 +[39] Xiaoyu Tian, Junru Gu, Bailin Li, Yicheng Liu, Chenxu Hu, Yang Wang, Kun Zhan, Peng Jia, Xianpeng Lang, and Hang Zhao. Drivevlm: The convergence of autonomous driving and large vision-language models. arXiv preprint arXiv:2402.12289, 2024. 1, 2, 3, 6, 7 +[40] Wenwen Tong, Chonghao Sima, Tai Wang, Li Chen, Silei Wu, Hanming Deng, Yi Gu, Lewei Lu, Ping Luo, Dahua Lin, et al. Scene as occupancy. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 8406-8415, 2023. 3 +[41] Hugo Touvron, Thibaut Lavril, Gautier Izacard, Xavier Martinet, Marie-Anne Lachaux, Timothée Lacroix, Baptiste Rozière, Naman Goyal, Eric Hambro, Faisal Azhar, et al. Llama: Open and efficient foundation language models. arXiv preprint arXiv:2302.13971, 2023. 2 +[42] Hugo Touvron, Louis Martin, Kevin Stone, Peter Albert, Amjad Almahairi, Yasmine Babaei, Nikolay Bashlykov, Soumya Batra, Prajjwal Bhargava, Shruti Bhosale, et al. Llama 2: Open foundation and fine-tuned chat models. arXiv preprint arXiv:2307.09288, 2023. 2 +[43] Shihao Wang, Zhiding Yu, Xiaohui Jiang, Shiyi Lan, Min Shi, Nadine Chang, Jan Kautz, Ying Li, and Jose M Alvarez. Omnidrive: A holistic llm-agent framework for autonomous driving with 3d perception, reasoning and planning. arXiv preprint arXiv:2405.01533, 2024. 1, 3, 6, 7 +[44] Wenhai Wang, Jiangwei Xie, ChuanYang Hu, Haoming Zou, Jianan Fan, Wenwen Tong, Yang Wen, Silei Wu, Hanming Deng, Zhiqi Li, et al. Driveml: Aligning multi-modal large language models with behavioral planning states for autonomous driving. arXiv preprint arXiv:2312.09245, 2023. 1, 2, 3 +[45] Zitian Wang, Zehao Huang, Yulu Gao, Naiyan Wang, and Si Liu. Mv2dfusion: Leveraging modality-specific object semantics for multi-modal 3d detection. arXiv preprint arXiv:2408.05945, 2024. 1, 2 +[46] Jason Wei, Xuezhi Wang, Dale Schuurmans, Maarten Bosma, Fei Xia, Ed Chi, Quoc V Le, Denny Zhou, et al. Chain-of-thought prompting elicits reasoning in large lan + +guage models. Advances in neural information processing systems, 35:24824-24837, 2022. 1, 2 +[47] Licheng Wen, Daocheng Fu, Xin Li, Xinyu Cai, MA Tao, Pinlong Cai, Min Dou, Botian Shi, Liang He, and Yu Qiao. Dilu: A knowledge-driven approach to autonomous driving with large language models. In The Twelfth International Conference on Learning Representations. 2 +[48] Xinshuo Weng, Boris Ivanovic, Yan Wang, Yue Wang, and Marco Pavone. Para-drive: Parallelized architecture for real-time autonomous driving. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 15449-15458, 2024. 1, 2, 3, 5, 6, 7 +[49] Penghao Wu, Xiaosong Jia, Li Chen, Junchi Yan, Hongyang Li, and Yu Qiao. Trajectory-guided control prediction for end-to-end autonomous driving: A simple yet strong baseline. Advances in Neural Information Processing Systems, 35:6119-6132, 2022. 1, 2 +[50] Zhenhua Xu, Yujia Zhang, Enze Xie, Zhen Zhao, Yong Guo, Kwan-Yee K Wong, Zhenguo Li, and Hengshuang Zhao. Drivegpt4: Interpretable end-to-end autonomous driving via large language model. IEEE Robotics and Automation Letters, 2024. 2 +[51] Rajeev Yasarla, Manish Kumar Singh, Hong Cai, Yunxiao Shi, Jisoo Jeong, Yinhao Zhu, Shizhong Han, Risheek Garrepalli, and Fatih Porikli. *Futuredepth: Learning to predict the future improves video depth estimation. In European Conference on Computer Vision*, pages 440–458. Springer, 2024. 4 +[52] Tianwei Yin, Xingyi Zhou, and Philipp Krahenbuhl. Center-based 3d object detection and tracking. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 11784-11793, 2021. 1, 2 +[53] Wenyuan Zeng, Wenjie Luo, Simon Suo, Abbas Sadat, Bin Yang, Sergio Casas, and Raquel Urtasun. End-to-end interpretable neural motion planner. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 8660-8669, 2019. 2 +[54] Jiang-Tian Zhai, Ze Feng, Jinhao Du, Yongqiang Mao, Jiang-Jiang Liu, Zichang Tan, Yifu Zhang, Xiaqing Ye, and Jingdong Wang. Rethinking the open-loop evaluation of end-to-end autonomous driving in nuscenes. arXiv preprint arXiv:2305.10430, 2023. 5, 6 +[55] Yunpeng Zhang, Zheng Zhu, Wenzhao Zheng, Junjie Huang, Guan Huang, Jie Zhou, and Jiwen Lu. Beverse: Unified perception and prediction in birds-eye-view for vision-centric autonomous driving. arXiv preprint arXiv:2205.09743, 2022. 2 \ No newline at end of file diff --git a/2025/Distilling Multi-modal Large Language Models for Autonomous Driving/images.zip b/2025/Distilling Multi-modal Large Language Models for Autonomous Driving/images.zip new file mode 100644 index 0000000000000000000000000000000000000000..81c9b11be513ebaa8c15a9f96eabfc96c2d4ac59 --- /dev/null +++ b/2025/Distilling Multi-modal Large Language Models for Autonomous Driving/images.zip @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:22d383cc59edbdc022bcd7e97f0186c82789524a3b54aca155fb79ff63d415e7 +size 726547 diff --git a/2025/Distilling Multi-modal Large Language Models for Autonomous Driving/layout.json b/2025/Distilling Multi-modal Large Language Models for Autonomous Driving/layout.json new file mode 100644 index 0000000000000000000000000000000000000000..14518d45c3ebe365629940cda7ae7afcbb4ee95d --- /dev/null +++ b/2025/Distilling Multi-modal Large Language Models for Autonomous Driving/layout.json @@ -0,0 +1,7030 @@ +{ + "pdf_info": [ + { + "para_blocks": [ + { + "bbox": [ + 83, + 103, + 529, + 121 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 83, + 103, + 529, + 121 + ], + "spans": [ + { + "bbox": [ + 83, + 103, + 529, + 121 + ], + "type": "text", + "content": "Distilling Multi-modal Large Language Models for Autonomous Driving" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 90, + 142, + 520, + 171 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 90, + 142, + 520, + 171 + ], + "spans": [ + { + "bbox": [ + 90, + 142, + 520, + 171 + ], + "type": "text", + "content": "Deepti Hegde" + }, + { + "bbox": [ + 90, + 142, + 520, + 171 + ], + "type": "inline_equation", + "content": "^{1\\dagger*}" + }, + { + "bbox": [ + 90, + 142, + 520, + 171 + ], + "type": "text", + "content": ", Rajeev Yasarla" + }, + { + "bbox": [ + 90, + 142, + 520, + 171 + ], + "type": "inline_equation", + "content": "^{2*}" + }, + { + "bbox": [ + 90, + 142, + 520, + 171 + ], + "type": "text", + "content": ", Hong Cai" + }, + { + "bbox": [ + 90, + 142, + 520, + 171 + ], + "type": "inline_equation", + "content": "^{2}" + }, + { + "bbox": [ + 90, + 142, + 520, + 171 + ], + "type": "text", + "content": ", Shizhong Han" + }, + { + "bbox": [ + 90, + 142, + 520, + 171 + ], + "type": "inline_equation", + "content": "^{2}" + }, + { + "bbox": [ + 90, + 142, + 520, + 171 + ], + "type": "text", + "content": ", Apratim Bhattacharyya" + }, + { + "bbox": [ + 90, + 142, + 520, + 171 + ], + "type": "inline_equation", + "content": "^{2}" + }, + { + "bbox": [ + 90, + 142, + 520, + 171 + ], + "type": "text", + "content": ", Shweta Mahajan" + }, + { + "bbox": [ + 90, + 142, + 520, + 171 + ], + "type": "inline_equation", + "content": "^{2}" + }, + { + "bbox": [ + 90, + 142, + 520, + 171 + ], + "type": "text", + "content": ", Litian Liu" + }, + { + "bbox": [ + 90, + 142, + 520, + 171 + ], + "type": "inline_equation", + "content": "^{2}" + }, + { + "bbox": [ + 90, + 142, + 520, + 171 + ], + "type": "text", + "content": ", Risheek Garrepalli" + }, + { + "bbox": [ + 90, + 142, + 520, + 171 + ], + "type": "inline_equation", + "content": "^{2}" + }, + { + "bbox": [ + 90, + 142, + 520, + 171 + ], + "type": "text", + "content": ", Vishal M. Patel" + }, + { + "bbox": [ + 90, + 142, + 520, + 171 + ], + "type": "inline_equation", + "content": "^{1}" + }, + { + "bbox": [ + 90, + 142, + 520, + 171 + ], + "type": "text", + "content": ", Fatih Porikli" + }, + { + "bbox": [ + 90, + 142, + 520, + 171 + ], + "type": "inline_equation", + "content": "^{2}" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 171, + 175, + 440, + 205 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 171, + 175, + 440, + 205 + ], + "spans": [ + { + "bbox": [ + 171, + 175, + 440, + 205 + ], + "type": "inline_equation", + "content": "^{1}" + }, + { + "bbox": [ + 171, + 175, + 440, + 205 + ], + "type": "text", + "content": " Johns Hopkins University, " + }, + { + "bbox": [ + 171, + 175, + 440, + 205 + ], + "type": "inline_equation", + "content": "^{2}" + }, + { + "bbox": [ + 171, + 175, + 440, + 205 + ], + "type": "text", + "content": " Qualcomm AI Research‡ *Equal contribution" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 152, + 232, + 200, + 245 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 152, + 232, + 200, + 245 + ], + "spans": [ + { + "bbox": [ + 152, + 232, + 200, + 245 + ], + "type": "text", + "content": "Abstract" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 55, + 258, + 296, + 521 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 258, + 296, + 521 + ], + "spans": [ + { + "bbox": [ + 55, + 258, + 296, + 521 + ], + "type": "text", + "content": "Autonomous driving demands safe motion planning, especially in critical \"long-tail\" scenarios. Recent end-to-end autonomous driving systems leverage large language models (LLMs) as planners to improve generalizability to rare events. However, using LLMs at test time introduces high computational costs. To address this, we propose DiMA, an end-to-end autonomous driving system that maintains the efficiency of an LLM-free (or vision-based) planner while leveraging the world knowledge of an LLM. DiMA distills the information from a multi-modal LLM to a vision-based end-to-end planner through a set of specially designed surrogate tasks. Under a joint training strategy, a scene encoder common to both networks produces structured representations that are semantically grounded as well as aligned to the final planning objective. Notably, the LLM is optional at inference, enabling robust planning without compromising on efficiency. Training with DiMA results in a " + }, + { + "bbox": [ + 55, + 258, + 296, + 521 + ], + "type": "inline_equation", + "content": "37\\%" + }, + { + "bbox": [ + 55, + 258, + 296, + 521 + ], + "type": "text", + "content": " reduction in the L2 trajectory error and an " + }, + { + "bbox": [ + 55, + 258, + 296, + 521 + ], + "type": "inline_equation", + "content": "80\\%" + }, + { + "bbox": [ + 55, + 258, + 296, + 521 + ], + "type": "text", + "content": " reduction in the collision rate of the vision-based planner, as well as a " + }, + { + "bbox": [ + 55, + 258, + 296, + 521 + ], + "type": "inline_equation", + "content": "44\\%" + }, + { + "bbox": [ + 55, + 258, + 296, + 521 + ], + "type": "text", + "content": " trajectory error reduction in long-tail scenarios. DiMA also achieves state-of-the-art performance on the nuScenes planning benchmark." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 56, + 535, + 135, + 547 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 535, + 135, + 547 + ], + "spans": [ + { + "bbox": [ + 56, + 535, + 135, + 547 + ], + "type": "text", + "content": "1. Introduction" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 55, + 552, + 295, + 671 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 552, + 295, + 671 + ], + "spans": [ + { + "bbox": [ + 55, + 552, + 295, + 671 + ], + "type": "text", + "content": "Research in autonomous driving has moved away from the traditional strategy of integrating sequential, independently trained models [4, 22, 26, 29, 31, 45, 52], and moved towards developing multi-task systems that are trained in an end-to-end manner [3, 5, 7, 16, 20, 49]. These approaches demonstrate improved performance and interpretability while being highly efficient. However, they struggle with long-tail navigation and perception scenarios, primarily because they are trained on task-specific, limited datasets [2, 11, 13]." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 55, + 672, + 296, + 696 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 672, + 296, + 696 + ], + "spans": [ + { + "bbox": [ + 55, + 672, + 296, + 696 + ], + "type": "text", + "content": "Large language models (LLMs) have emerged as a promising solution to this issue. Trained on vast, internet-" + } + ] + } + ], + "index": 9 + }, + { + "type": "image", + "bbox": [ + 317, + 232, + 544, + 327 + ], + "blocks": [ + { + "bbox": [ + 317, + 232, + 544, + 327 + ], + "lines": [ + { + "bbox": [ + 317, + 232, + 544, + 327 + ], + "spans": [ + { + "bbox": [ + 317, + 232, + 544, + 327 + ], + "type": "image", + "image_path": "e9bc53db64c7aec28a15abcce8c7550c42f21a271c907e1cf3ff15a85ecf054c.jpg" + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "image_body" + } + ], + "index": 10 + }, + { + "type": "image", + "bbox": [ + 317, + 329, + 544, + 423 + ], + "blocks": [ + { + "bbox": [ + 317, + 329, + 544, + 423 + ], + "lines": [ + { + "bbox": [ + 317, + 329, + 544, + 423 + ], + "spans": [ + { + "bbox": [ + 317, + 329, + 544, + 423 + ], + "type": "image", + "image_path": "7bfe5d2afb4b74a03c120aadb7df0997c6c5bc631c3cc8c96557dec693a18d3e.jpg" + } + ] + } + ], + "index": 11, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 313, + 426, + 555, + 503 + ], + "lines": [ + { + "bbox": [ + 313, + 426, + 555, + 503 + ], + "spans": [ + { + "bbox": [ + 313, + 426, + 555, + 503 + ], + "type": "text", + "content": "Figure 1. Comparison of planning performance in long-tail scenarios from nuScenes: DiMA-VAD demonstrates greater robustness compared to VAD [20] in long-tail navigation scenarios such as overtaking a vehicle and performing a 3-point turn. DiMA-VAD also outperforms recent vision-planner PARA-Drive [48] and LLM planner TILEN [38]. Notably, the 3-point turn is a zero-shot scenario that is only present in the validation set." + } + ] + } + ], + "index": 12, + "angle": 0, + "type": "image_caption" + } + ], + "index": 11 + }, + { + "bbox": [ + 313, + 521, + 556, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 521, + 556, + 715 + ], + "spans": [ + { + "bbox": [ + 313, + 521, + 556, + 715 + ], + "type": "text", + "content": "scale datasets, LLMs can leverage world knowledge to generalize to unseen or rare scenarios. These models can perform high-level reasoning tasks using mechanisms such as chain-of-thought [46]. Going beyond text-based prompting, multi-modal large language models (MLLMs) integrate image and video inputs to the LLM, enabling tasks such as visual question answering and dense captioning [21, 24]. Recent end-to-end autonomous driving systems leverage LLMs to achieve higher interpretability through language-guided reasoning and superior robustness to long-tail scenarios [19, 37-39, 43, 44]. To differentiate between these methods, we call end-to-end planners that depend on LLMs to perform trajectory prediction as \"LLM-based\" planners and ones that do not as \"vision-based\" planners [18, 20, 48]. In spite of their recent success \"LLM-based\" planners face significant challenges." + } + ] + } + ], + "index": 13 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "spans": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "type": "text", + "content": "CVF" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 145, + 0, + 496, + 37 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 145, + 0, + 496, + 37 + ], + "spans": [ + { + "bbox": [ + 145, + 0, + 496, + 37 + ], + "type": "text", + "content": "This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 70, + 702, + 295, + 713 + ], + "type": "page_footnote", + "angle": 0, + "lines": [ + { + "bbox": [ + 70, + 702, + 295, + 713 + ], + "spans": [ + { + "bbox": [ + 70, + 702, + 295, + 713 + ], + "type": "text", + "content": "Qualcomm AI Research is initiative of Qualcomm Technologies, Inc." + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 293, + 748, + 317, + 758 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 317, + 758 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 317, + 758 + ], + "type": "text", + "content": "27575" + } + ] + } + ], + "index": 15 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 0 + }, + { + "para_blocks": [ + { + "bbox": [ + 54, + 72, + 294, + 251 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 54, + 72, + 294, + 251 + ], + "spans": [ + { + "bbox": [ + 54, + 72, + 294, + 251 + ], + "type": "text", + "content": "Firstly, LLM-based planners require a significant amount of computational overhead at test time, limiting their practicality. This work addresses a core question: how can we harness LLMs' world knowledge while preserving the efficiency of vision-based planners? Secondly, bridging the visual and language domains is inherently more complex for end-to-end planning tasks compared to general MLLM tasks [21, 24], with the additional hurdle of limited training data. Standard image tokenization strategies rely on frozen pre-trained image encoders [33] to generate dense, unstructured multi-modal token embeddings. [21, 24, 37]. We suggest that MLLMs meant for end-to-end autonomous driving benefit from structured inputs that explicitly model scene components. Furthermore, we suggest that updating these features during training results in richer representations." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 56, + 254, + 296, + 552 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 254, + 296, + 552 + ], + "spans": [ + { + "bbox": [ + 56, + 254, + 296, + 552 + ], + "type": "text", + "content": "To address these challenges, we propose DiMA, a novel framework for Distilling Multi-modal Large Language Models for Autonomous driving. We introduce a joint-training scheme between a vision-based planner and an MLLM that enables the learning of robust, grounded and disentangled scene representations that are aligned to the final objective of planning. Specifically, we use the vision-based planner as a tokenizer, and pass the learned scene representations to the MLLM, providing a more structured input. The MLLM is then trained for the tasks of visual question-answering, trajectory estimation, and a series of surrogate tasks designed to ground the multi-modal inputs in language. The vision-based planner is simultaneously trained for trajectory estimation, while also distilling features from the MLLM planning head to the planning transformer. Crucially, the MLLM can be discarded when performing planning inference, maintaining the efficiency of the vision-based planner while utilizing the knowledge of the language model. Optionally, the MLLM can support language-guided reasoning through visual question answering. Our experiments show that DiMA is more robust to challenging long-tail scenarios compared to baselines and state-of-the-art methods [38, 48]. Figure 1 confirms this both quantitatively and qualitatively. We list the contributions of the work below:" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 55, + 558, + 294, + 713 + ], + "type": "list", + "angle": 0, + "index": 5, + "blocks": [ + { + "bbox": [ + 56, + 558, + 294, + 617 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 558, + 294, + 617 + ], + "spans": [ + { + "bbox": [ + 56, + 558, + 294, + 617 + ], + "type": "text", + "content": "1. We introduce DiMA, an end-to-end autonomous driving framework that distills knowledge from an MLLM to a vision-based planner to ensure robustness to long-tail events while maintaining efficiency. DiMA is capable of planning as well as visual question-answering." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 55, + 618, + 294, + 666 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 618, + 294, + 666 + ], + "spans": [ + { + "bbox": [ + 55, + 618, + 294, + 666 + ], + "type": "text", + "content": "2. We propose a distillation task along with the following surrogate tasks to align the objectives of the vision-based planner and the MLLM: (i) masked token reconstruction (ii) future token prediction (iii) scene editing." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 55, + 666, + 294, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 666, + 294, + 713 + ], + "spans": [ + { + "bbox": [ + 55, + 666, + 294, + 713 + ], + "type": "text", + "content": "3. DiMA outperforms both vision-based and LLM end-to-end planners, achieving state-of-the-art results on the nuScenes planning benchmark. Training with DiMA results in a " + }, + { + "bbox": [ + 55, + 666, + 294, + 713 + ], + "type": "inline_equation", + "content": "37\\%" + }, + { + "bbox": [ + 55, + 666, + 294, + 713 + ], + "type": "text", + "content": " reduction in the L2 trajectory error and an" + } + ] + } + ], + "index": 4 + } + ], + "sub_type": "text" + }, + { + "bbox": [ + 326, + 72, + 553, + 108 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 326, + 72, + 553, + 108 + ], + "spans": [ + { + "bbox": [ + 326, + 72, + 553, + 108 + ], + "type": "text", + "content": "80% reduction in the collision rate of the vision-based planner, as well as a 44% trajectory error reduction in long-tail scenarios." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 313, + 120, + 400, + 132 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 120, + 400, + 132 + ], + "spans": [ + { + "bbox": [ + 313, + 120, + 400, + 132 + ], + "type": "text", + "content": "2. Related Work" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 313, + 137, + 493, + 149 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 137, + 493, + 149 + ], + "spans": [ + { + "bbox": [ + 313, + 137, + 493, + 149 + ], + "type": "text", + "content": "2.1. End-to-End Autonomous Driving" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 313, + 152, + 553, + 402 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 152, + 553, + 402 + ], + "spans": [ + { + "bbox": [ + 313, + 152, + 553, + 402 + ], + "type": "text", + "content": "Autonomous driving (AD) systems are generally modular, consisting of tasks such as perception [22, 31, 45, 52], motion prediction [4, 26, 29], and planning [9]. Classic AD systems use standalone models for each task where each module is optimized separately and combined sequentially. Alternatively, some approaches incorporate these modular tasks into a multi-task learning paradigm, where tasks shared a common feature extraction process but have task-specific heads [8, 23, 53, 55]. Recently, research has moved toward optimizing autonomous driving systems in an end-to-end manner, aiming to directly predict future trajectories [3, 5, 7, 16, 20, 49]. Along this line of work, UniAD [18] introduces a query-based design that integrates perception and prediction components, enabling an end-to-end planning scheme. VAD [20] utilizes a vectorized scene representation, replacing the dense rasterized representations from [18], reducing computational cost and improving planning performance. While effective in general navigation scenarios, these methods struggle in difficult long-tail events (see Figure 1). In this work, we propose a method for end-to-end autonomous driving that is robust to such scenarios." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 313, + 411, + 483, + 423 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 411, + 483, + 423 + ], + "spans": [ + { + "bbox": [ + 313, + 411, + 483, + 423 + ], + "type": "text", + "content": "2.2. LLMs for Autonomous Driving" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 313, + 426, + 553, + 520 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 426, + 553, + 520 + ], + "spans": [ + { + "bbox": [ + 313, + 426, + 553, + 520 + ], + "type": "text", + "content": "The recent success of large language models [12, 41, 42] demonstrates their ability to process complex contextual information with logical reasoning and to communicate in a human-interpretable manner, attracting significant attention in autonomous driving [6, 10, 14, 25, 28, 35-37, 39, 44, 47, 50]. These works investigate the capabilities of language models to generalize to novel scenarios as well as their ability to reason about the scene in the form of text." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 313, + 521, + 553, + 617 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 521, + 553, + 617 + ], + "spans": [ + { + "bbox": [ + 313, + 521, + 553, + 617 + ], + "type": "text", + "content": "Prompting LLMs. One category of approaches involves prompting LLMs with text using reasoning frameworks such as chain-of-thought [46]. LanguageMPC [35] leverages the common-sense reasoning capacity of LLM to make high-level driving decisions, which are then translated into bottom-level control signals. GPT-Driver [28] converts observations and ego-states into language prompts, framing driving planning as a text-based task for LLMs." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 313, + 618, + 553, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 618, + 553, + 713 + ], + "spans": [ + { + "bbox": [ + 313, + 618, + 553, + 713 + ], + "type": "text", + "content": "End-to-end Driving with MLLMs. Another approach is to integrate multi-modal LLMs into end-to-end train frameworks for autonomous vehicles. LMDrive [36] integrates a vision encoder with a large language model, enabling natural language instruction in autonomous driving. DriveLM introduces multi-step question answering, going beyond single-round interactions [27, 32, 34]. DriveVLM [39] leverages MLLMs and a chain-of-thought process to" + } + ] + } + ], + "index": 13 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 318, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 318, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 318, + 757 + ], + "type": "text", + "content": "27576" + } + ] + } + ], + "index": 14 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 1 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 82, + 72, + 525, + 240 + ], + "blocks": [ + { + "bbox": [ + 82, + 72, + 525, + 240 + ], + "lines": [ + { + "bbox": [ + 82, + 72, + 525, + 240 + ], + "spans": [ + { + "bbox": [ + 82, + 72, + 525, + 240 + ], + "type": "image", + "image_path": "5283f6fe783ad4a42906f06c27fdf6a19d35035d65f2d988985427793a7e386e.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 54, + 246, + 555, + 302 + ], + "lines": [ + { + "bbox": [ + 54, + 246, + 555, + 302 + ], + "spans": [ + { + "bbox": [ + 54, + 246, + 555, + 302 + ], + "type": "text", + "content": "Figure 2. Overview of DiMA. The input to the framework is a multi-view image sequence and a question text prompt. The vision-based end-to-end planner consists of a scene encoder and a planning transformer. The scene encoder learns structured latent representations in the form of bird's-eye-view, ego, agent, and map (BEAM) token embeddings and acts as a trainable tokenizer for the multi-modal large language model (MLLM). The planning transformer is trained under standard planning constraints [18, 20]. The MLLM is trained for planning, visual question answering, distillation, and a series of surrogate tasks." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "bbox": [ + 55, + 318, + 231, + 329 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 318, + 231, + 329 + ], + "spans": [ + { + "bbox": [ + 55, + 318, + 231, + 329 + ], + "type": "text", + "content": "enhance scene understanding and planning." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 54, + 330, + 296, + 462 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 54, + 330, + 296, + 462 + ], + "spans": [ + { + "bbox": [ + 54, + 330, + 296, + 462 + ], + "type": "text", + "content": "We point out a challenge in the current MLLM strategy for autonomous driving. We suggest that in order to reason about vehicle dynamics, the language model benefits from a structured input that explicitly models scene components instead of dense, unstructured input to the MLLM as in [36, 39, 43, 44]. In TOKEN [38], the scene encoder from [48] is used as a tokenizer. However, it is frozen during training. In our approach, the input to the MLLM is a set of latent representations learned from a pre-trained scene encoder that is jointly trained with the MLLM, allowing it to more learn grounded feature representations" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 55, + 471, + 167, + 483 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 471, + 167, + 483 + ], + "spans": [ + { + "bbox": [ + 55, + 471, + 167, + 483 + ], + "type": "text", + "content": "3. DiMA Framework" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 54, + 488, + 296, + 584 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 54, + 488, + 296, + 584 + ], + "spans": [ + { + "bbox": [ + 54, + 488, + 296, + 584 + ], + "type": "text", + "content": "We present DiMA, a framework for end-to-end autonomous driving. Given a sequence of multi-view images, the overall objective is to predict the future trajectory of the ego vehicle and answer questions about the scene. An overview of DiMA is shown in Figure 2. The framework has two main components: 1) the vision-based planner; 2) a multi-modal large language model consisting of adapter layers, an LLM, and a series of task-specific decoder heads." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 55, + 591, + 178, + 602 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 591, + 178, + 602 + ], + "spans": [ + { + "bbox": [ + 55, + 591, + 178, + 602 + ], + "type": "text", + "content": "3.1. Vision-based Planner" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 54, + 605, + 296, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 54, + 605, + 296, + 714 + ], + "spans": [ + { + "bbox": [ + 54, + 605, + 296, + 714 + ], + "type": "text", + "content": "Vision-based end-to-end planners are trained in a multi-task fashion to perform perception, mapping, motion prediction and planning [18, 20, 40, 48]. Considering standard architecture design from [18, 20, 48] we decompose the planner into the scene encoder and the planning transformer. The scene encoder provides structured latent representations to the planning transformer that performs waypoint prediction. We call these representations as scene token embeddings. In our framework, we leverage the vision-based planner in" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 313, + 318, + 555, + 401 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 318, + 555, + 401 + ], + "spans": [ + { + "bbox": [ + 313, + 318, + 555, + 401 + ], + "type": "text", + "content": "two ways. First, it acts as the primary network through which planning is performed, allowing for fast prediction at test time. Second, the scene encoder is shared with the MLLM, acting as tokenizer to provide it with a highly structured set of inputs. An important distinction of our framework from previous works [19, 38] is that the scene encoder is trained jointly with the MLLM." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 313, + 403, + 556, + 651 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 403, + 556, + 651 + ], + "spans": [ + { + "bbox": [ + 313, + 403, + 556, + 651 + ], + "type": "text", + "content": "Scene encoder The scene encoder models the scene as high-dimensional token embeddings that explicitly represent components such as the environment map, the ego vehicle, and the surrounding agents. This is achieved by introducing task specific, learnable query features that are supervised by a series of planning constraints [20]. First, a visual backbone encodes the input multi-view image sequence. The obtained feature map is projected into the bird's-eye-view (BEV) space using a set of queries to obtain BEV token embeddings " + }, + { + "bbox": [ + 313, + 403, + 556, + 651 + ], + "type": "inline_equation", + "content": "(B)" + }, + { + "bbox": [ + 313, + 403, + 556, + 651 + ], + "type": "text", + "content": " [22]. The network learns structured representations of the map and agents in the form of map token embeddings " + }, + { + "bbox": [ + 313, + 403, + 556, + 651 + ], + "type": "inline_equation", + "content": "(M)" + }, + { + "bbox": [ + 313, + 403, + 556, + 651 + ], + "type": "text", + "content": " and agent token " + }, + { + "bbox": [ + 313, + 403, + 556, + 651 + ], + "type": "inline_equation", + "content": "(A)" + }, + { + "bbox": [ + 313, + 403, + 556, + 651 + ], + "type": "text", + "content": " embeddings by cross-attending the BEV features with map and agent queries. A randomly initialized learnable embedding is trained to learn the interaction of the ego vehicle with the agent and map components which is the ego query " + }, + { + "bbox": [ + 313, + 403, + 556, + 651 + ], + "type": "inline_equation", + "content": "(E)" + }, + { + "bbox": [ + 313, + 403, + 556, + 651 + ], + "type": "text", + "content": ". We denote the set of these latent scene representations as " + }, + { + "bbox": [ + 313, + 403, + 556, + 651 + ], + "type": "inline_equation", + "content": "BEAM" + }, + { + "bbox": [ + 313, + 403, + 556, + 651 + ], + "type": "text", + "content": " token embeddings. These token embeddings are sent to a planning head that predicts future trajectories of the ego and agent vehicles. They are also used as input to the MLLM as multi-modal inputs, as discussed next." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 313, + 661, + 424, + 673 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 661, + 424, + 673 + ], + "spans": [ + { + "bbox": [ + 313, + 661, + 424, + 673 + ], + "type": "text", + "content": "3.2. Multi-modal LLM" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 313, + 677, + 555, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 677, + 555, + 715 + ], + "spans": [ + { + "bbox": [ + 313, + 677, + 555, + 715 + ], + "type": "text", + "content": "Our objective is to distill knowledge from an MLLM to the vision-based planner. Towards this, we jointly train an MLLM with the vision-based planner. Specifically, the" + } + ] + } + ], + "index": 11 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "text", + "content": "27577" + } + ] + } + ], + "index": 12 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 2 + }, + { + "para_blocks": [ + { + "bbox": [ + 55, + 72, + 296, + 204 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 72, + 296, + 204 + ], + "spans": [ + { + "bbox": [ + 55, + 72, + 296, + 204 + ], + "type": "text", + "content": "scene encoder is used as a trainable tokenizer for the MLLM to generate the " + }, + { + "bbox": [ + 55, + 72, + 296, + 204 + ], + "type": "inline_equation", + "content": "BEAM" + }, + { + "bbox": [ + 55, + 72, + 296, + 204 + ], + "type": "text", + "content": " token embeddings, as well as used as an encoder for the planning transformer. The advantage from this strategy is two-fold. First, the MLLM receives a highly structured input that captures rich spatio-temporal information relevant to the autonomous driving task. Second, the scene encoder learns features that are grounded in language, improving the robustness of vision-based planning. The main components of the MLLM are adapter layers, the large language model, and a series of task-specific decoder heads which we discuss next." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 55, + 209, + 203, + 220 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 209, + 203, + 220 + ], + "spans": [ + { + "bbox": [ + 55, + 209, + 203, + 220 + ], + "type": "text", + "content": "3.2.1. Adaptation of Scene Tokens" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 55, + 222, + 296, + 342 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 222, + 296, + 342 + ], + "spans": [ + { + "bbox": [ + 55, + 222, + 296, + 342 + ], + "type": "text", + "content": "In order to project the BEAM token embeddings efficiently while maintaining their distinctiveness, we leverage the query-transformer (Q-former) module [17] to compress the visual tokens before being input to the LLM. We implement component-specific Q-former layers, namely a Map Q-former, a Bird's-Eye-View (BEV) Q-former, an Ego Q-former, and an Agent Q-former. Each of these adapters transforms the representing different modalities into a sequence of fixed-length, higher-dimensional tokens, preparing them for processing by the MLLM." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 55, + 346, + 168, + 358 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 346, + 168, + 358 + ], + "spans": [ + { + "bbox": [ + 55, + 346, + 168, + 358 + ], + "type": "text", + "content": "3.2.2. MLLM Supervision" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 55, + 361, + 296, + 433 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 361, + 296, + 433 + ], + "spans": [ + { + "bbox": [ + 55, + 361, + 296, + 433 + ], + "type": "text", + "content": "We design tasks for the MLLM with the objectives of, a) enriching the intermediate scene representations; b) grounding the scene token embeddings in language; and c) training the LLM for planning-related reasoning. In detail, the MLLM is trained for visual question answering, trajectory estimation, feature distillation, and a set of surrogate tasks [1, 51]." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 55, + 435, + 296, + 555 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 435, + 296, + 555 + ], + "spans": [ + { + "bbox": [ + 55, + 435, + 296, + 555 + ], + "type": "text", + "content": "Visual question answering. The MLLM is trained for visual question answering (VQA) on question-answer pairs of 4 types: perception of the scene, prediction of the agent behavior, prediction of ego-vehicle behavior, and future planning steps [37]. The answer is predicted based on a multimodal prompt consisting of the question embeddings and the projected BEAM token embeddings. This is constructed as a standard next-token prediction task and the VQA branch is supervised with a cross-entropy loss as in [21]. We denote the loss from this task head as " + }, + { + "bbox": [ + 55, + 435, + 296, + 555 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{LLM}" + }, + { + "bbox": [ + 55, + 435, + 296, + 555 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 55, + 558, + 295, + 605 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 558, + 295, + 605 + ], + "spans": [ + { + "bbox": [ + 55, + 558, + 295, + 605 + ], + "type": "text", + "content": "Surrogate tasks: The main goal of joint training is to enrich the BEAM scene representations learned by the scene encoder. We design MLLM surrogate tasks whose objectives are in line with that of future trajectory prediction." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 55, + 605, + 296, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 605, + 296, + 714 + ], + "spans": [ + { + "bbox": [ + 55, + 605, + 296, + 714 + ], + "type": "text", + "content": "1) Masked token reconstruction: Each type of token embedding contributes to a holistic representation of the scene. In order to enrich the visual representations, we ask the network to reconstruct a masked BEV input based on the context present in the rest of the multi-modal sequence. We perform random masking after scene encoding and pass the token embeddings to the MLLM. A reconstruction head takes the latent representations from the penultimate layer of LLM and predicts reconstructed BEV token embeddings" + } + ] + } + ], + "index": 7 + }, + { + "type": "image", + "bbox": [ + 339, + 70, + 529, + 266 + ], + "blocks": [ + { + "bbox": [ + 339, + 70, + 529, + 266 + ], + "lines": [ + { + "bbox": [ + 339, + 70, + 529, + 266 + ], + "spans": [ + { + "bbox": [ + 339, + 70, + 529, + 266 + ], + "type": "image", + "image_path": "14389e201993856381630025fdf42676d5dd9c81bbdd68528ae5fc3cd0142134.jpg" + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 313, + 270, + 555, + 324 + ], + "lines": [ + { + "bbox": [ + 313, + 270, + 555, + 324 + ], + "spans": [ + { + "bbox": [ + 313, + 270, + 555, + 324 + ], + "type": "text", + "content": "Figure 3. Examples of addition and deletion in scene editing. In the top row, a car (solid red box) is added on the left of the ego-vehicle (green box). In the bottom row, a car (dashed red box) is removed from the right of the ego-vehicle. A corresponding question-answer pair is created to characterize the edit." + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "image_caption" + } + ], + "index": 8 + }, + { + "bbox": [ + 313, + 338, + 554, + 363 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 338, + 554, + 363 + ], + "spans": [ + { + "bbox": [ + 313, + 338, + 554, + 363 + ], + "type": "inline_equation", + "content": "\\hat{B}" + }, + { + "bbox": [ + 313, + 338, + 554, + 363 + ], + "type": "text", + "content": ". This decoder head is supervised with the L2 loss between the prediction and the complete input," + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 391, + 364, + 553, + 378 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 391, + 364, + 553, + 378 + ], + "spans": [ + { + "bbox": [ + 391, + 364, + 553, + 378 + ], + "type": "interline_equation", + "content": "\\mathcal {L} _ {\\text {r e c o n}} = \\left\\| \\hat {B} - B \\right\\| ^ {2} \\tag {1}", + "image_path": "3f8d194e1a6d58028c4fa21053dd89bc61d4a7e74f117b3c4658e1cebb588bd3.jpg" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 313, + 380, + 554, + 429 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 380, + 554, + 429 + ], + "spans": [ + { + "bbox": [ + 313, + 380, + 554, + 429 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 313, + 380, + 554, + 429 + ], + "type": "inline_equation", + "content": "(\\{m(B), E, A, M\\})" + }, + { + "bbox": [ + 313, + 380, + 554, + 429 + ], + "type": "text", + "content": " is input to the MLLM, and the latent MLLM representations associated with the BEV token embeddings are input to the masked reconstruction decoder head. Here, " + }, + { + "bbox": [ + 313, + 380, + 554, + 429 + ], + "type": "inline_equation", + "content": "m(.)" + }, + { + "bbox": [ + 313, + 380, + 554, + 429 + ], + "type": "text", + "content": " denotes random masking." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 313, + 430, + 555, + 525 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 430, + 555, + 525 + ], + "spans": [ + { + "bbox": [ + 313, + 430, + 555, + 525 + ], + "type": "text", + "content": "2) Future BEV prediction: An important aspect of planning is anticipating future events. We introduce the surrogate task of future BEV prediction to encourage the LLM to learn spatio-temporal cues useful for planning. Given latent BEV token embeddings, we train a prediction head to predict future BEV token embeddings and supervise this task as the L2 loss between predicted and ground truth future token embeddings:" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 344, + 526, + 553, + 541 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 344, + 526, + 553, + 541 + ], + "spans": [ + { + "bbox": [ + 344, + 526, + 553, + 541 + ], + "type": "interline_equation", + "content": "\\mathcal {L} _ {\\text {f u t u r e}} = \\left\\| \\hat {F} _ {t} - B _ {t + 1} \\right\\| ^ {2} + \\left\\| \\tilde {F} _ {t} - B _ {t + 2} \\right\\| ^ {2} \\tag {2}", + "image_path": "6bde1f1cbc9cff917ec33b9bcb9f0cf609135a5044d687398cc51c0371c8d281.jpg" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 313, + 544, + 554, + 593 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 544, + 554, + 593 + ], + "spans": [ + { + "bbox": [ + 313, + 544, + 554, + 593 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 313, + 544, + 554, + 593 + ], + "type": "inline_equation", + "content": "\\hat{F}_t, \\tilde{F}_t" + }, + { + "bbox": [ + 313, + 544, + 554, + 593 + ], + "type": "text", + "content": " are the predicted future BEV token embeddings at time " + }, + { + "bbox": [ + 313, + 544, + 554, + 593 + ], + "type": "inline_equation", + "content": "t" + }, + { + "bbox": [ + 313, + 544, + 554, + 593 + ], + "type": "text", + "content": ". Note, we predict future BEV token embeddings " + }, + { + "bbox": [ + 313, + 544, + 554, + 593 + ], + "type": "inline_equation", + "content": "(B_{t+1}, B_{t+2})" + }, + { + "bbox": [ + 313, + 544, + 554, + 593 + ], + "type": "text", + "content": " of the multi-view image sequence for time steps " + }, + { + "bbox": [ + 313, + 544, + 554, + 593 + ], + "type": "inline_equation", + "content": "\\{t+1, t+2\\}" + }, + { + "bbox": [ + 313, + 544, + 554, + 593 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 313, + 594, + 555, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 594, + 555, + 714 + ], + "spans": [ + { + "bbox": [ + 313, + 594, + 555, + 714 + ], + "type": "text", + "content": "3) Scene editing: For prediction and reasoning about the ego vehicle, it is crucial to learn how surrounding agents impact the ego-vehicle's future path. We propose a novel scene editing task in which we augment scenes by removing or adding new agents. Along with this, we construct a question-answer pair related to the edit. We show examples of this in Figure 3. For scene addition, given the map constraints, predicted map, the ego bounding box location and the trajectories of predicted agents, we create a way-point trajectory for a new agent of category \"car\" or \"truck\". A" + } + ] + } + ], + "index": 16 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "text", + "content": "27578" + } + ] + } + ], + "index": 17 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 3 + }, + { + "para_blocks": [ + { + "bbox": [ + 55, + 72, + 294, + 191 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 72, + 294, + 191 + ], + "spans": [ + { + "bbox": [ + 55, + 72, + 294, + 191 + ], + "type": "text", + "content": "new agent token embedding is then created using a linear layer. This new agent token embedding, the corresponding text prompt, and the rest of the BEAM token embeddings are passed as input to the LLM. The hidden latent LLM features corresponding to agent token embeddings are then fed into a dedicated scene editing decoder head that performs waypoint prediction of the ego vehicle. The language prediction head performs question-answering on the new QA pair. This task thus contributes to the existing planning constraint loss and VQA loss of the MLLM." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 55, + 195, + 294, + 278 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 195, + 294, + 278 + ], + "spans": [ + { + "bbox": [ + 55, + 195, + 294, + 278 + ], + "type": "text", + "content": "Distillation In order to further align the representations learned by the vision planner and the MLLM, we perform knowledge transfer between the penultimate layers of the language model and the planning transformer. Concretely, we minimize the KL-divergence between the distributions of the hidden features of the planning transformer and the LLM hidden features of the ego-token embeddings:" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 116, + 281, + 294, + 294 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 116, + 281, + 294, + 294 + ], + "spans": [ + { + "bbox": [ + 116, + 281, + 294, + 294 + ], + "type": "interline_equation", + "content": "\\mathcal {L} _ {\\text {d i s t i l l}} = D _ {K L} \\left(P _ {l l m} \\| P _ {v i s}\\right) \\tag {3}", + "image_path": "ead14699129d56d080a9866aa4f63db675bab00230eb1c900a41f21b07f09f72.jpg" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 55, + 297, + 294, + 344 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 297, + 294, + 344 + ], + "spans": [ + { + "bbox": [ + 55, + 297, + 294, + 344 + ], + "type": "text", + "content": "where the " + }, + { + "bbox": [ + 55, + 297, + 294, + 344 + ], + "type": "inline_equation", + "content": "P_{vis}" + }, + { + "bbox": [ + 55, + 297, + 294, + 344 + ], + "type": "text", + "content": " are the features of penultimate layers of vision planning transformer and " + }, + { + "bbox": [ + 55, + 297, + 294, + 344 + ], + "type": "inline_equation", + "content": "P_{llm}" + }, + { + "bbox": [ + 55, + 297, + 294, + 344 + ], + "type": "text", + "content": " are the hidden LLM embeddings features corresponding to the ego-token embeddings of penultimate layers of MLLM model." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 55, + 345, + 294, + 380 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 345, + 294, + 380 + ], + "spans": [ + { + "bbox": [ + 55, + 345, + 294, + 380 + ], + "type": "text", + "content": "Loss functions. We supervise the network with a weighted sum of the losses corresponding to planning, visual question answering, distillation, and each surrogate task." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 61, + 384, + 294, + 396 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 61, + 384, + 294, + 396 + ], + "spans": [ + { + "bbox": [ + 61, + 384, + 294, + 396 + ], + "type": "interline_equation", + "content": "\\mathcal {L} = \\mathcal {L} _ {\\text {p l a n n i n g}} + \\mathcal {L} _ {\\text {L L M}} + \\mathcal {L} _ {\\text {r e c o n}} + \\mathcal {L} _ {\\text {f u t u r e}} + \\mathcal {L} _ {\\text {d i s t i l l}} \\tag {4}", + "image_path": "04f5501431da8e3c0dd70df1a83bdc5905a21ac68966b822979bc977fafe4667.jpg" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 55, + 399, + 294, + 434 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 399, + 294, + 434 + ], + "spans": [ + { + "bbox": [ + 55, + 399, + 294, + 434 + ], + "type": "text", + "content": "here " + }, + { + "bbox": [ + 55, + 399, + 294, + 434 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{\\text{planning}}" + }, + { + "bbox": [ + 55, + 399, + 294, + 434 + ], + "type": "text", + "content": " comes from the planning objective used to train [18, 20]. Here, the loss weights are chosen to bring each value to the same scale." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 55, + 445, + 173, + 460 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 445, + 173, + 460 + ], + "spans": [ + { + "bbox": [ + 55, + 445, + 173, + 460 + ], + "type": "text", + "content": "4. Experimental Setup" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 55, + 463, + 294, + 523 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 463, + 294, + 523 + ], + "spans": [ + { + "bbox": [ + 55, + 463, + 294, + 523 + ], + "type": "text", + "content": "In Section 4.1, we present the architecture design details, in Section 4.2 we present the training strategy, and in Section 4.3, we provide information about the datasets used. Importantly, in Section 4.4 we detail multiple evaluation strategies for fair and comprehensive comparison with existing works." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 55, + 531, + 170, + 544 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 531, + 170, + 544 + ], + "spans": [ + { + "bbox": [ + 55, + 531, + 170, + 544 + ], + "type": "text", + "content": "4.1. Architecture design" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 55, + 545, + 294, + 629 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 545, + 294, + 629 + ], + "spans": [ + { + "bbox": [ + 55, + 545, + 294, + 629 + ], + "type": "text", + "content": "Vision-based planner We conduct experiments using two vision-based end-to-end planners, VAD [20] and UniAD [18] due to their performance, efficiency, and ability to model the interaction between scene components. We use two model size variants of VAD. Both planners are trained for perception, motion prediction and planning, while UniAD also performs occupancy prediction." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 55, + 630, + 294, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 630, + 294, + 713 + ], + "spans": [ + { + "bbox": [ + 55, + 630, + 294, + 713 + ], + "type": "text", + "content": "MLLM design The MLLM is made up of adapter layers, a language model, and a set of task-specific decoder layers. We project each " + }, + { + "bbox": [ + 55, + 630, + 294, + 713 + ], + "type": "inline_equation", + "content": "BEAM" + }, + { + "bbox": [ + 55, + 630, + 294, + 713 + ], + "type": "text", + "content": " token embedding with its own dedicated Q-former adapter layer following [17]. After projection, the " + }, + { + "bbox": [ + 55, + 630, + 294, + 713 + ], + "type": "inline_equation", + "content": "BEAM" + }, + { + "bbox": [ + 55, + 630, + 294, + 713 + ], + "type": "text", + "content": " token embeddings and language token embeddings lie in the same embedding space. In order to reduce memory consumption, we limit the agent token se" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 313, + 72, + 553, + 168 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 72, + 553, + 168 + ], + "spans": [ + { + "bbox": [ + 313, + 72, + 553, + 168 + ], + "type": "text", + "content": "quence length to a fraction of the total number of agents and perform sequence-wise upsampling after input to the language model. The adapter layers are randomly initialized. We use the LLM from LLaVA-v1.5-7B [24] as our language model base. The ego prediction and agent motion prediction task heads are multi-layer-perceptron networks following [20]. For the surrogate task decoder heads, we use 3 Linear layers with a ReLU activation layer." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 313, + 175, + 378, + 188 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 175, + 378, + 188 + ], + "spans": [ + { + "bbox": [ + 313, + 175, + 378, + 188 + ], + "type": "text", + "content": "4.2. Training" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 313, + 190, + 553, + 298 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 190, + 553, + 298 + ], + "spans": [ + { + "bbox": [ + 313, + 190, + 553, + 298 + ], + "type": "text", + "content": "DiMA training follows a two-stage approach. First the vision-only planner is pre-trained for perception, prediction, and planning for 60 epochs in order to learn informative latent scene representations. Second, we perform joint training of the vision planner and the MLLM for an additional 30 epochs, incorporating all proposed tasks and losses detailed in Section 3. In the second stage, the language model of the MLLM is fine-tuned using LoRA [15]. Please refer to the supplementary material for more details." + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 313, + 306, + 376, + 316 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 306, + 376, + 316 + ], + "spans": [ + { + "bbox": [ + 313, + 306, + 376, + 316 + ], + "type": "text", + "content": "4.3. Datasets" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 313, + 320, + 555, + 572 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 320, + 555, + 572 + ], + "spans": [ + { + "bbox": [ + 313, + 320, + 555, + 572 + ], + "type": "text", + "content": "We use the nuScenes dataset for the task of open-loop planning [2]. This dataset consists of 28k total samples in a " + }, + { + "bbox": [ + 313, + 320, + 555, + 572 + ], + "type": "inline_equation", + "content": "22\\mathrm{k} / 6\\mathrm{k}" + }, + { + "bbox": [ + 313, + 320, + 555, + 572 + ], + "type": "text", + "content": " training/Validation split. The objects in each scene are annotated with 3D bounding box, orientation, and vehicle speed information. Additionally, we make use of the CAN bus information for the ego vehicle's ground-truth trajectory annotations. For supervising visual question answering, we train with DriveLM [37]. This dataset consists of a 4k subset of samples from the nuScenes dataset annotated with 300k QA pairs related to perception, prediction, planning, and behavior of the ego vehicle. For the samples in nuScenes that do not have text annotations, we create a set of perception, planning, and prediction QA pairs based on the numerical annotations. We use this information to prompt a Llama3-70B model [12] with in-context examples to generate DriveLM-like QA pairs. For the behavior QA pair, we create a rule-based algorithm to categorize the future movement of the ego-vehicle based on the ground-truth trajectory values and ego vehicle speed. More details on this method as well as examples of generated QA pairs may be found in the supplementary." + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 313, + 579, + 421, + 590 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 579, + 421, + 590 + ], + "spans": [ + { + "bbox": [ + 313, + 579, + 421, + 590 + ], + "type": "text", + "content": "4.4. Evaluation details" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 313, + 594, + 553, + 665 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 594, + 553, + 665 + ], + "spans": [ + { + "bbox": [ + 313, + 594, + 553, + 665 + ], + "type": "text", + "content": "We evaluate planning performance based on the predicted future trajectories of the ego vehicle 3 seconds into the future, where 2 waypoints are predicted per second. We use two metrics, the L2 error (in meters) between the predicted and ground-truth waypoints as well as the collision rate (in " + }, + { + "bbox": [ + 313, + 594, + 553, + 665 + ], + "type": "inline_equation", + "content": "\\%" + }, + { + "bbox": [ + 313, + 594, + 553, + 665 + ], + "type": "text", + "content": " ) between the ego-vehicle and the surrounding vehicles." + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 313, + 666, + 553, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 666, + 553, + 713 + ], + "spans": [ + { + "bbox": [ + 313, + 666, + 553, + 713 + ], + "type": "text", + "content": "Standardized vs VAD evaluation. In [48], Weng et al. point out inconsistencies in planning evaluation across VAD [20], UniAD [18] and AD-MLP [54], namely in the way L2 error is averaged over time and the way noisy and invalid" + } + ] + } + ], + "index": 19 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "text", + "content": "27579" + } + ] + } + ], + "index": 20 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 4 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 59, + 118, + 552, + 404 + ], + "blocks": [ + { + "bbox": [ + 55, + 70, + 555, + 115 + ], + "lines": [ + { + "bbox": [ + 55, + 70, + 555, + 115 + ], + "spans": [ + { + "bbox": [ + 55, + 70, + 555, + 115 + ], + "type": "text", + "content": "Table 1. Comparison of L2 trajectory error and collision rate on nuScenes [2] using standardized evaluation [48]. Models are evaluated on the general validation split as well as a \"targeted\" split of challenging samples from [48]. The performance of the DiMA model variants are in shades of purple. We summarize results by averaging over at " + }, + { + "bbox": [ + 55, + 70, + 555, + 115 + ], + "type": "inline_equation", + "content": "t = \\{1,2,3\\} s" + }, + { + "bbox": [ + 55, + 70, + 555, + 115 + ], + "type": "text", + "content": " as well as at all time steps. The best average performance in each setting is in bold. \"+\" indicates the use of ego-status information." + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 59, + 118, + 552, + 404 + ], + "lines": [ + { + "bbox": [ + 59, + 118, + 552, + 404 + ], + "spans": [ + { + "bbox": [ + 59, + 118, + 552, + 404 + ], + "type": "table", + "html": "
MethodUsing Ego statusTraj L2 (m) ↓Collision (%) ↓
1s2s3sAve1,2,3sAve all1s2s3sAve1,2,3sAve all
Full validation split
UniAD[18]X0.480.891.470.950.830.320.290.730.450.40
VAD-Base [20]X0.410.861.460.910.780.020.260.830.370.30
PARA-Drive[48]X0.260.591.120.660.560.000.120.650.260.17
TOKEN[38]X0.260.701.460.810.68----0.15
DiMA (VAD-Tiny)X0.200.531.100.610.510.000.070.190.090.06
DiMA (VAD-Base)X0.180.501.020.570.470.000.050.160.070.06
AD-MLP[54]✓0.230.581.180.660.560.000.140.700.280.20
PARA-Drive+ [48]✓0.200.521.040.590.490.000.090.490.190.13
DiMA+ (VAD-Tiny)✓0.190.511.060.590.490.000.060.160.080.06
DiMA+ (VAD-Base)✓0.180.481.010.560.460.000.050.160.070.06
DiMA-Dual+ (VAD-Tiny)✓0.180.501.030.570.470.000.050.160.080.06
Targeted validation split
UniAD[18]X0.471.091.921.160.990.000.000.730.240.15
VAD-Base [20]X0.531.202.071.271.080.000.290.870.390.34
PARA-Drive[48]X0.380.971.881.080.910.000.000.720.240.14
DiMA (VAD-Tiny)X0.320.881.700.970.810.000.030.240.090.05
DiMA (VAD-Base)X0.290.791.490.830.710.000.030.220.080.05
AD-MLP[54]✓0.330.992.061.130.940.000.583.621.400.94
PARA-Drive+ [48]✓0.290.761.460.830.700.000.000.290.100.05
DiMA+ (VAD-Tiny)✓0.310.831.670.940.770.000.040.170.070.05
DiMA+ (VAD-Base)✓0.270.741.380.790.640.000.030.160.060.04
DiMA-Dual+ (VAD-Tiny)✓0.280.761.490.840.690.000.030.170.070.05
", + "image_path": "aa438d4682570fbb82e68eb93a908d74aa1db5be526032b4ac2a420657d41809.jpg" + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "table_body" + } + ], + "index": 1 + }, + { + "bbox": [ + 54, + 415, + 294, + 547 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 54, + 415, + 294, + 547 + ], + "spans": [ + { + "bbox": [ + 54, + 415, + 294, + 547 + ], + "type": "text", + "content": "frames are considered. The authors also point out that collision performance is considerably improved by using a more finely discretized BEV map. Accounting for these inconsistencies, they propose a \"standardized\" evaluation metric to maintain a fair comparison across methods. For accurate reporting and fair comparison with [38, 48], we use this standardized metric to evaluate our model. Certain existing works use the evaluation scheme followed by [20] (which we call VAD evaluation) and do not make code or models available [16, 30, 39, 43]. Here, we compare reported results against DiMA evaluated using VAD evaluation." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 55, + 548, + 295, + 621 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 548, + 295, + 621 + ], + "spans": [ + { + "bbox": [ + 55, + 548, + 295, + 621 + ], + "type": "text", + "content": "Validation splits. We evaluate planning performance across 3 types of splits from the nuScenes validation set: 1) the entire validation set of 6019 samples 2) a 689 sample sub-set more difficult \"targeted\" scenarios in which the ego-vehicle turns during navigation [48], and 3) long-tail events in which the ego-vehicle encounters rare maneuvers [38]. u" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 55, + 635, + 180, + 649 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 635, + 180, + 649 + ], + "spans": [ + { + "bbox": [ + 55, + 635, + 180, + 649 + ], + "type": "text", + "content": "5. Experimental Results" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 55, + 653, + 296, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 653, + 296, + 714 + ], + "spans": [ + { + "bbox": [ + 55, + 653, + 296, + 714 + ], + "type": "text", + "content": "We perform a comprehensive evaluation of DiMA on the nuScenes open-loop planning benchmark in a variety of evaluation settings. We compare our performance against that of recent vision-based planners [16, 18, 20, 48] as well as recent MLLM planners [30, 38, 39, 43]. To facilitate a" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 313, + 415, + 555, + 452 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 415, + 555, + 452 + ], + "spans": [ + { + "bbox": [ + 313, + 415, + 555, + 452 + ], + "type": "text", + "content": "fair comparison of DiMA with recent works, we use two evaluation strategies VAD evaluation [20] and standardized evaluation [48]. Please refer to Section 4.4 for more details." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 313, + 456, + 556, + 685 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 456, + 556, + 685 + ], + "spans": [ + { + "bbox": [ + 313, + 456, + 556, + 685 + ], + "type": "text", + "content": "Standardized evaluation. Under the standardized evaluation strategy, we evaluate each method on two validation splits: 1) the full nuScenes validation split, and 2) a \"targeted\" subset of the nuScenes validation split that contains scenarios where the ego-vehicle must take a turn, as established in [48]. We present our results in Table 1. The baseline vision planners VAD [20] and UniAD [18] exhibit suboptimal performance in more difficult targeted scenarios. DiMA demonstrates significant improvements over baseline methods, achieving a " + }, + { + "bbox": [ + 313, + 456, + 556, + 685 + ], + "type": "inline_equation", + "content": "35\\%" + }, + { + "bbox": [ + 313, + 456, + 556, + 685 + ], + "type": "text", + "content": " reduction in L2 trajectory error and an " + }, + { + "bbox": [ + 313, + 456, + 556, + 685 + ], + "type": "inline_equation", + "content": "80\\%" + }, + { + "bbox": [ + 313, + 456, + 556, + 685 + ], + "type": "text", + "content": " reduction in collision rate on the overall nuScenes validation set, and a " + }, + { + "bbox": [ + 313, + 456, + 556, + 685 + ], + "type": "inline_equation", + "content": "35\\%" + }, + { + "bbox": [ + 313, + 456, + 556, + 685 + ], + "type": "text", + "content": " reduction in L2 trajectory error in the \"targeted\" split. More notably, we outperform recent state-of-the-art methods PARA-Drive [48] and TOKEN [38] consistently across both validation splits. We observe a strong improvement in collision rate, improving on that of TOKEN by " + }, + { + "bbox": [ + 313, + 456, + 556, + 685 + ], + "type": "inline_equation", + "content": "60\\%" + }, + { + "bbox": [ + 313, + 456, + 556, + 685 + ], + "type": "text", + "content": ". Unlike TOKEN, our approach does not rely on an LLM for inference, thus making DiMA more efficient as well as more accurate." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 313, + 689, + 556, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 689, + 556, + 715 + ], + "spans": [ + { + "bbox": [ + 313, + 689, + 556, + 715 + ], + "type": "text", + "content": "VAD evaluation. For a fair comparison with recent works that use the evaluation strategy from VAD [20], we present" + } + ] + } + ], + "index": 8 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "type": "text", + "content": "27580" + } + ] + } + ], + "index": 9 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 5 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 58, + 118, + 555, + 316 + ], + "blocks": [ + { + "bbox": [ + 55, + 70, + 555, + 115 + ], + "lines": [ + { + "bbox": [ + 55, + 70, + 555, + 115 + ], + "spans": [ + { + "bbox": [ + 55, + 70, + 555, + 115 + ], + "type": "text", + "content": "Table 2. Comparison of L2 trajectory error and collision rate on nuScenes [2] using VAD evaluation [20]. Models are evaluated on the general validation split. The performance of the DiMA model variants are in shades of purple. The performance of the MLLM-branch DiMA model variant is in blue. We summarize results by averaging over all time steps. The best average performance in each setting is in bold. “+” indicates the use of ego-status information." + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 58, + 118, + 555, + 316 + ], + "lines": [ + { + "bbox": [ + 58, + 118, + 555, + 316 + ], + "spans": [ + { + "bbox": [ + 58, + 118, + 555, + 316 + ], + "type": "table", + "html": "
MethodUsing Ego statusTraj L2 (m) ↓Collision (%) ↓Latency (ms)FPS
1s2s3sAveall1s2s3sAveall
ST-P3 [16]X1.332.112.902.110.230.621.270.71628.31.6
UniAD[18]X0.480.961.651.030.050.170.710.31555.61.8
VAD-Tiny[20]X0.460.761.120.780.210.350.580.3859.516.8
VAD-Base[20]X0.410.701.050.720.070.170.410.22224.34.5
VLP-VAD-Base[30]X0.260.470.780.500.120.170.420.23--
OmniDrive [43]X0.400.801.320.840.040.462.320.94--
DiMA (UniAD)X0.150.300.560.340.060.080.220.125601.8
DiMA (VAD-Tiny)X0.180.360.610.380.070.100.270.1559.516.8
DiMA (VAD-Base)X0.130.270.470.290.050.080.160.102264.5
VAD-Tiny✓0.200.380.650.410.100.120.270.1659.516.8
VAD-Base✓0.170.340.600.370.070.100.240.14224.34.5
DriveVLM [39]✓0.180.340.680.400.100.220.450.27--
DriveVLM-Dual (VAD-Base) [39]✓0.150.290.480.310.050.080.170.10--
DiMA + (VAD-Tiny)✓0.170.330.590.370.070.100.260.1459.516.8
DiMA + (VAD-Base)✓0.120.250.440.270.040.060.150.082264.5
DiMA-Dual+ (VAD-Tiny)✓0.140.270.460.290.050.070.150.092863.5
", + "image_path": "b321c973910577f8ef5bda621c926d0d098e3f6e53bc16f7c593a1f3c3dd4d58.jpg" + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "table_body" + } + ], + "index": 1 + }, + { + "type": "table", + "bbox": [ + 58, + 414, + 293, + 653 + ], + "blocks": [ + { + "bbox": [ + 55, + 334, + 294, + 411 + ], + "lines": [ + { + "bbox": [ + 55, + 334, + 294, + 411 + ], + "spans": [ + { + "bbox": [ + 55, + 334, + 294, + 411 + ], + "type": "text", + "content": "Table 3. Long-tail performance comparison of L2 trajectory error and collision rate on nuScenes [2] validation set using standardized evaluation [48]. The performance of vision planner DiMA model variants are in shades of purple. The performance of MLLM-branch DiMA model variants are in shades of blue. We summarize results by averaging over at " + }, + { + "bbox": [ + 55, + 334, + 294, + 411 + ], + "type": "inline_equation", + "content": "t = \\{1,2,3\\} s" + }, + { + "bbox": [ + 55, + 334, + 294, + 411 + ], + "type": "text", + "content": " as well as at all time steps. The best average performance in each setting is in bold." + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 58, + 414, + 293, + 653 + ], + "lines": [ + { + "bbox": [ + 58, + 414, + 293, + 653 + ], + "spans": [ + { + "bbox": [ + 58, + 414, + 293, + 653 + ], + "type": "table", + "html": "
MethodTraj L2 (m)↓Coll. (%)↓
1s2s3sAve1,2,3sAveallAveall
3-point turn (zero-shot)
VAD-Tiny0.761.683.041.831.550.00
VAD-Base0.711.663.241.871.570.00
PARA-Drive [48]0.501.382.761.551.295.33
TOKEN [38]0.391.292.601.431.184.00
DiMA (VAD-Tiny)0.471.272.551.431.170.00
DiMA (VAD-Base)0.361.182.371.301.050.00
DiMA (MLLM)0.381.122.351.281.040.00
DiMA-Dual (VAD-Tiny)0.381.082.351.271.040.00
Resume from stop
VAD-Tiny0.641.632.991.751.490.00
VAD-Base0.601.722.831.721.420.00
PARA-Drive0.140.792.301.080.850.00
TOKEN0.130.701.580.800.650.00
DiMA (VAD-Tiny)0.261.022.401.230.990.00
DiMA (VAD-Base)0.150.651.340.710.660.00
DiMA (MLLM)0.241.022.161.140.930.00
DiMA-Dual (VAD-Tiny)0.240.982.131.110.910.00
Overtake
VAD-Tiny0.581.272.121.321.142.42
VAD-Base0.461.162.171.261.062.49
PARA-Drive0.270.891.941.030.852.30
TOKEN0.290.771.630.900.740.00
DiMA (VAD-Tiny)0.240.751.490.830.691.32
DiMA (VAD-Base)0.240.721.500.820.661.29
DiMA (MLLM)0.240.731.500.820.671.30
DiMA-Dual (VAD-Tiny)0.240.731.500.820.671.30
", + "image_path": "61252b0aebe23b3881bc95b42e66a0477f5c7547722e14a186ba10ace03e0478.jpg" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "table_body" + } + ], + "index": 3 + }, + { + "bbox": [ + 55, + 658, + 295, + 695 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 658, + 295, + 695 + ], + "spans": [ + { + "bbox": [ + 55, + 658, + 295, + 695 + ], + "type": "text", + "content": "an additional set of results under this strategy. In Table 2, we present the evaluation of DiMA on the nuScenes validation set [2]. We significantly outperform the base" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 313, + 335, + 555, + 407 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 335, + 555, + 407 + ], + "spans": [ + { + "bbox": [ + 313, + 335, + 555, + 407 + ], + "type": "text", + "content": "line vision-based planners VAD [20] and UniAD [18]. Notably, DiMA built on VAD-Tiny outperforms VAD-Base by " + }, + { + "bbox": [ + 313, + 335, + 555, + 407 + ], + "type": "inline_equation", + "content": "47\\%" + }, + { + "bbox": [ + 313, + 335, + 555, + 407 + ], + "type": "text", + "content": " in L2 trajectory error while being 4 times faster. We also outperform MLLM planners like OmniDrive [43] and DriveVLM [39] without requiring an LLM at inference, making our approach more accurate as well as efficient." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 313, + 411, + 556, + 567 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 411, + 556, + 567 + ], + "spans": [ + { + "bbox": [ + 313, + 411, + 556, + 567 + ], + "type": "text", + "content": "Performance in long-tail scenarios. Autonomous driving performance is particularly crucial for rare scenarios. Using the manually selected long-tail events from [38], we demonstrate the robustness of DiMA in cases of novel navigation and perception. In Table 3, we evaluate the performance of DiMA for three long-tail scenarios, \"3-point turn\", \"resume from the stop\" and \"overtake\", selected from the nuScenes validation set. Training with DiMA results in a significant and consistent boost in performance from the baseline vision-based planners. We also consistently outperform PARA-Drive and TOKEN. Notably, the \"3-point turn\" case is a zero-shot scenario, not present in the training data. We achieve the lowest L2 trajectory error for this case." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 313, + 570, + 557, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 570, + 557, + 713 + ], + "spans": [ + { + "bbox": [ + 313, + 570, + 557, + 713 + ], + "type": "text", + "content": "Evaluating the MLLM branch. We also evaluate the MLLM branch of DiMA and demonstrate its planning and visual question-answering capabilities. We perform LLM-based planning using DiMA in two ways. First, we directly evaluate the waypoint predictions of the MLLM planning decoder head. This is denoted by DiMA (MLLM). Secondly, we define DiMA-Dual, an optional hybrid inference model that combines the strengths of both the vision and MLLM branches. We perform max-pooling on the penultimate layer features of both vision planning transformer and MLLM ego prediction. The resulting pooled features are then fed back into both the vision planning transformer and" + } + ] + } + ], + "index": 7 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 66, + 702, + 278, + 713 + ], + "type": "page_footnote", + "angle": 0, + "lines": [ + { + "bbox": [ + 66, + 702, + 278, + 713 + ], + "spans": [ + { + "bbox": [ + 66, + 702, + 278, + 713 + ], + "type": "text", + "content": "1We compare with reported results due to lack of published code." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "text", + "content": "27581" + } + ] + } + ], + "index": 9 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 6 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 57, + 70, + 301, + 220 + ], + "blocks": [ + { + "bbox": [ + 57, + 70, + 301, + 220 + ], + "lines": [ + { + "bbox": [ + 57, + 70, + 301, + 220 + ], + "spans": [ + { + "bbox": [ + 57, + 70, + 301, + 220 + ], + "type": "image", + "image_path": "50f3c2512a4f9267ad7bb8e17c10129bd313a222693b89a5732a20458a67adf3.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 55, + 224, + 555, + 247 + ], + "lines": [ + { + "bbox": [ + 55, + 224, + 555, + 247 + ], + "spans": [ + { + "bbox": [ + 55, + 224, + 555, + 247 + ], + "type": "text", + "content": "Figure 4. Visualization of planning performance by the MLLM branch of DiMA-VAD-Tiny. We plot the predicted trajectory (orange-yellow) and show an example response of the language model branch to a question from the DriveLM test dataset [37]." + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "type": "image", + "bbox": [ + 307, + 70, + 551, + 220 + ], + "blocks": [ + { + "bbox": [ + 307, + 70, + 551, + 220 + ], + "lines": [ + { + "bbox": [ + 307, + 70, + 551, + 220 + ], + "spans": [ + { + "bbox": [ + 307, + 70, + 551, + 220 + ], + "type": "image", + "image_path": "f324fc1565c9d9a416294862fac9f1ca3576c60912344d459c0d2fcf234c3f9a.jpg" + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_body" + } + ], + "index": 1 + }, + { + "type": "table", + "bbox": [ + 58, + 284, + 545, + 396 + ], + "blocks": [ + { + "bbox": [ + 55, + 258, + 555, + 282 + ], + "lines": [ + { + "bbox": [ + 55, + 258, + 555, + 282 + ], + "spans": [ + { + "bbox": [ + 55, + 258, + 555, + 282 + ], + "type": "text", + "content": "Table 4. Results of ablation experiments of DiMA-VAD-Tiny on nuScenes under VAD evaluation. The baseline model is VAD-Tiny. We report L2 trajectory error and collision rate averaged over all time steps." + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 58, + 284, + 545, + 396 + ], + "lines": [ + { + "bbox": [ + 58, + 284, + 545, + 396 + ], + "spans": [ + { + "bbox": [ + 58, + 284, + 545, + 396 + ], + "type": "table", + "html": "
IDVQAScene tokensDistillationLLM PlanningSurrogate tasksTraj L2 (m) ↓Collision (%) ↓
Masked recon.Future pred.Scene editing1s2s3sAveall1s2s3sAveall
1XXXXXXX0.330.580.890.600.120.190.550.29
2✓BEVX✓XXX0.360.600.910.620.100.170.520.26
3✓BEV, MapX✓XXX0.280.520.870.560.100.170.490.25
4✓AllX✓XXX0.260.510.830.520.100.160.380.21
5✓All✓✓XXX0.220.440.770.480.090.140.340.19
6✓All✓✓✓XX0.190.390.680.420.090.130.320.18
7✓All✓✓✓✓X0.180.370.630.390.070.110.290.16
8✓All✓✓✓✓✓0.180.360.610.380.070.100.270.15
", + "image_path": "0acd665cf16619f1618238c8519132aaae489cc67036bf64172ad7487cc9b607.jpg" + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "table_body" + } + ], + "index": 4 + }, + { + "bbox": [ + 55, + 411, + 295, + 495 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 411, + 295, + 495 + ], + "spans": [ + { + "bbox": [ + 55, + 411, + 295, + 495 + ], + "type": "text", + "content": "MLLM ego prediction for waypoint prediction. This max pooling encourages consistency between the predictions of the two models, leading to a more robust and potentially improved overall prediction as seen in Table 3. Using the MLLM at inference results in a significant boost in performance, with DiMA-Dual (VAD-Tiny) matching the performance of DiMA (VAD-Base) in Table 2." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 55, + 496, + 296, + 556 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 496, + 296, + 556 + ], + "spans": [ + { + "bbox": [ + 55, + 496, + 296, + 556 + ], + "type": "text", + "content": "VQA results. We provide qualitative examples of visual question-answering in the DriveLM dataset [37] in Figure 4. We demonstrate the ability of DiMA to reason about safe actions to be taken by the ego vehicle. We provide a more examples in the supplementary material." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 55, + 566, + 149, + 578 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 566, + 149, + 578 + ], + "spans": [ + { + "bbox": [ + 55, + 566, + 149, + 578 + ], + "type": "text", + "content": "5.1. Ablation study." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 55, + 582, + 296, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 582, + 296, + 713 + ], + "spans": [ + { + "bbox": [ + 55, + 582, + 296, + 713 + ], + "type": "text", + "content": "We examine the role of different design choices and proposed tasks in the DiMA framework through a comprehensive ablation study in Table 4. In order to do so, we build our final frame-work step-by-step. We begin with VAD-Tiny as our baseline model in the first row, denoted by ID-1. In ID-2, we observe inconsistent gains and drops in performance by naively training an LLM with the VAD-Tiny under a the planning and VQA objectives while using only BEV features. By comparing ID-3 and ID-4, we show that that providing more scene context to the MLLM by using all the BEAM token embeddings to train both LLM and" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 313, + 411, + 555, + 532 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 411, + 555, + 532 + ], + "spans": [ + { + "bbox": [ + 313, + 411, + 555, + 532 + ], + "type": "text", + "content": "vision-based planner benefits the performance, resulting a significant improvement in trajectory estimation and collision rate. We show in ID-5 that explicitly distilling the information from the penultimate layers of the MLLM to the planning transformer gives another boost in performance. In ID-6, ID-7 and ID-8 we demonstrate the importance of surrogate tasks like masked-reconstruction, future prediction and scene editing in helping the scene encoder to learn richer representations useful for planning. We point to the supplementary material for additional ablation studies." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 314, + 550, + 388, + 562 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 314, + 550, + 388, + 562 + ], + "spans": [ + { + "bbox": [ + 314, + 550, + 388, + 562 + ], + "type": "text", + "content": "6. Conclusion" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 313, + 570, + 555, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 570, + 555, + 713 + ], + "spans": [ + { + "bbox": [ + 313, + 570, + 555, + 713 + ], + "type": "text", + "content": "We introduce DiMA, an end-to-end autonomous driving framework for robust and efficient planning. DiMA distills knowledge from an MLLM to a vision-based planner, employing a novel joint training strategy alongside carefully designed surrogate tasks such as masked token reconstruction, future token prediction, and scene editing. DiMA enables the model to learn semantically grounded scene representations, improving its performance in difficult navigation scenarios. We conduct extensive evaluations and ablation studies, showing that training with DiMA results in a planner that excels in long-tail cases while maintaining computational efficiency." + } + ] + } + ], + "index": 11 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 318, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 318, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 318, + 757 + ], + "type": "text", + "content": "27582" + } + ] + } + ], + "index": 12 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 7 + }, + { + "para_blocks": [ + { + "bbox": [ + 56, + 71, + 115, + 83 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 71, + 115, + 83 + ], + "spans": [ + { + "bbox": [ + 56, + 71, + 115, + 83 + ], + "type": "text", + "content": "References" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 57, + 91, + 296, + 712 + ], + "type": "list", + "angle": 0, + "index": 13, + "blocks": [ + { + "bbox": [ + 61, + 91, + 296, + 145 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 61, + 91, + 296, + 145 + ], + "spans": [ + { + "bbox": [ + 61, + 91, + 296, + 145 + ], + "type": "text", + "content": "[1] Apratim Bhattacharyya, Sunny Panchal, Reza Pourreza, Mingu Lee, Pulkit Madan, and Roland Memisevic. Look, remember and reason: Grounded reasoning in videos with language models. In International Conference of Learning Representations, 2024. 4" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 61, + 146, + 296, + 213 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 61, + 146, + 296, + 213 + ], + "spans": [ + { + "bbox": [ + 61, + 146, + 296, + 213 + ], + "type": "text", + "content": "[2] Holger Caesar, Varun Bankiti, Alex H Lang, Sourabh Vora, Venice Erin Liong, Qiang Xu, Anush Krishnan, Yu Pan, Giancarlo Baldan, and Oscar Beijbom. nuscenes: A multimodal dataset for autonomous driving. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 11621-11631, 2020. 1, 5, 6, 7" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 62, + 214, + 295, + 257 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 214, + 295, + 257 + ], + "spans": [ + { + "bbox": [ + 62, + 214, + 295, + 257 + ], + "type": "text", + "content": "[3] Sergio Casas, Abbas Sadat, and Raquel Urtasun. Mp3: A unified model to map, perceive, predict and plan. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 14403-14412, 2021. 1, 2" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 62, + 258, + 295, + 300 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 258, + 295, + 300 + ], + "spans": [ + { + "bbox": [ + 62, + 258, + 295, + 300 + ], + "type": "text", + "content": "[4] Yuning Chai, Benjamin Sapp, Mayank Bansal, and Dragomir Anguelov. Multipath: Multiple probabilistic anchor trajectory hypotheses for behavior prediction. arXiv preprint arXiv:1910.05449, 2019. 1, 2" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 62, + 302, + 294, + 345 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 302, + 294, + 345 + ], + "spans": [ + { + "bbox": [ + 62, + 302, + 294, + 345 + ], + "type": "text", + "content": "[5] Dian Chen and Philipp Krahenbuhl. Learning from all vehicles. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 17222-17231, 2022. 1, 2" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 62, + 347, + 295, + 412 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 347, + 295, + 412 + ], + "spans": [ + { + "bbox": [ + 62, + 347, + 295, + 412 + ], + "type": "text", + "content": "[6] Long Chen, Oleg Sinavski, Jan Hünermann, Alice Karnsund, Andrew James Willmott, Danny Birch, Daniel Maund, and Jamie Shotton. Driving with llms: Fusing object-level vector modality for explainable autonomous driving. In 2024 IEEE International Conference on Robotics and Automation (ICRA), pages 14093-14100. IEEE, 2024. 2" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 62, + 414, + 294, + 457 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 414, + 294, + 457 + ], + "spans": [ + { + "bbox": [ + 62, + 414, + 294, + 457 + ], + "type": "text", + "content": "[7] Kashyap Chitta, Aditya Prakash, and Andreas Geiger. Neat: Neural attention fields for end-to-end autonomous driving. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 15793-15803, 2021. 1, 2" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 62, + 459, + 294, + 513 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 459, + 294, + 513 + ], + "spans": [ + { + "bbox": [ + 62, + 459, + 294, + 513 + ], + "type": "text", + "content": "[8] Kashyap Chitta, Aditya Prakash, Bernhard Jaeger, Zehao Yu, Katrin Renz, and Andreas Geiger. Transfuser: Imitation with transformer-based sensor fusion for autonomous driving. IEEE Transactions on Pattern Analysis and Machine Intelligence, 45(11):12878-12895, 2022. 2" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 62, + 514, + 294, + 568 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 514, + 294, + 568 + ], + "spans": [ + { + "bbox": [ + 62, + 514, + 294, + 568 + ], + "type": "text", + "content": "[9] Felipe Codevilla, Eder Santana, Antonio M López, and Adrien Gaidon. Exploring the limitations of behavior cloning for autonomous driving. In Proceedings of the IEEE/CVF international conference on computer vision, pages 9329-9338, 2019. 2" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 57, + 570, + 294, + 613 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 570, + 294, + 613 + ], + "spans": [ + { + "bbox": [ + 57, + 570, + 294, + 613 + ], + "type": "text", + "content": "[10] Xinpeng Ding, Jianhua Han, Hang Xu, Wei Zhang, and Xiaomeng Li. Hilm-d: Towards high-resolution understanding in multimodal large language models for autonomous driving. arXiv preprint arXiv:2309.05186, 2023. 2" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 57, + 614, + 294, + 657 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 614, + 294, + 657 + ], + "spans": [ + { + "bbox": [ + 57, + 614, + 294, + 657 + ], + "type": "text", + "content": "[11] Alexey Dosovitskiy, German Ros, Felipe Codevilla, Antonio Lopez, and Vladlen Koltun. Carla: An open urban driving simulator. In Conference on robot learning, pages 1-16. PMLR, 2017. 1" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 57, + 658, + 294, + 712 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 658, + 294, + 712 + ], + "spans": [ + { + "bbox": [ + 57, + 658, + 294, + 712 + ], + "type": "text", + "content": "[12] Abhimanyu Dubey, Abhinav Jauhri, Abhinav Pandey, Abhishek Kadian, Ahmad Al-Dahle, Aiesha Letman, Akhil Mathur, Alan Schelten, Amy Yang, Angela Fan, et al. The llama 3 herd of models. arXiv preprint arXiv:2407.21783, 2024. 2, 5" + } + ] + } + ], + "index": 12 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 316, + 73, + 555, + 714 + ], + "type": "list", + "angle": 0, + "index": 26, + "blocks": [ + { + "bbox": [ + 316, + 73, + 554, + 139 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 73, + 554, + 139 + ], + "spans": [ + { + "bbox": [ + 316, + 73, + 554, + 139 + ], + "type": "text", + "content": "[13] Scott Ettinger, Shuyang Cheng, Benjamin Caine, Chenxi Liu, Hang Zhao, Sabeek Pradhan, Yuning Chai, Ben Sapp, Charles R Qi, Yin Zhou, et al. Large scale interactive motion forecasting for autonomous driving: The waymo open motion dataset. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 9710-9719, 2021. 1" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 317, + 141, + 555, + 195 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 141, + 555, + 195 + ], + "spans": [ + { + "bbox": [ + 317, + 141, + 555, + 195 + ], + "type": "text", + "content": "[14] Daocheng Fu, Xin Li, Licheng Wen, Min Dou, Pinlong Cai, Botian Shi, and Yu Qiao. Drive like a human: Rethinking autonomous driving with large language models. In Proceedings of the IEEE/CVF Winter Conference on Applications of Computer Vision, pages 910-919, 2024. 2" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 316, + 197, + 554, + 240 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 197, + 554, + 240 + ], + "spans": [ + { + "bbox": [ + 316, + 197, + 554, + 240 + ], + "type": "text", + "content": "[15] Edward J Hu, Yelong Shen, Phillip Wallis, Zeyuan Allen-Zhu, Yuanzhi Li, Shean Wang, Lu Wang, and Weizhu Chen. Lora: Low-rank adaptation of large language models. arXiv preprint arXiv:2106.09685, 2021. 5" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 317, + 243, + 554, + 297 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 243, + 554, + 297 + ], + "spans": [ + { + "bbox": [ + 317, + 243, + 554, + 297 + ], + "type": "text", + "content": "[16] Shengchao Hu, Li Chen, Penghao Wu, Hongyang Li, Junchi Yan, and Dacheng Tao. St-p3: End-to-end vision-based autonomous driving via spatial-temporal feature learning. In European Conference on Computer Vision, pages 533-549. Springer, 2022. 1, 2, 6, 7" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 316, + 298, + 553, + 331 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 298, + 553, + 331 + ], + "spans": [ + { + "bbox": [ + 316, + 298, + 553, + 331 + ], + "type": "text", + "content": "[17] Wenbo Hu, Zi-Yi Dou, Liunian Harold Li, Anita Kamath, Nanyun Peng, and Kai-Wei Chang. Matryoshka query transformer for large vision-language models, 2024. 4, 5" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 317, + 333, + 554, + 397 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 333, + 554, + 397 + ], + "spans": [ + { + "bbox": [ + 317, + 333, + 554, + 397 + ], + "type": "text", + "content": "[18] Yihan Hu, Jiazhi Yang, Li Chen, Keyu Li, Chonghao Sima, Xizhou Zhu, Siqi Chai, Senyao Du, Tianwei Lin, Wenhai Wang, et al. Planning-oriented autonomous driving. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 17853-17862, 2023. 1, 2, 3, 5, 6, 7" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 317, + 399, + 554, + 453 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 399, + 554, + 453 + ], + "spans": [ + { + "bbox": [ + 317, + 399, + 554, + 453 + ], + "type": "text", + "content": "[19] Jyh-Jing Hwang, Runsheng Xu, Hubert Lin, Wei-Chih Hung, Jingwei Ji, Kristy Choi, Di Huang, Tong He, Paul Covington, Benjamin Sapp, et al. Emma: End-to-end multimodal model for autonomous driving. arXiv preprint arXiv:2410.23262, 2024. 1, 3" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 317, + 456, + 554, + 521 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 456, + 554, + 521 + ], + "spans": [ + { + "bbox": [ + 317, + 456, + 554, + 521 + ], + "type": "text", + "content": "[20] Bo Jiang, Shaoyu Chen, Qing Xu, Bencheng Liao, Jiajie Chen, Helong Zhou, Qian Zhang, Wenyu Liu, Chang Huang, and Xinggang Wang. Vad: Vectorized scene representation for efficient autonomous driving. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 8340-8350, 2023. 1, 2, 3, 5, 6, 7" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 317, + 523, + 554, + 577 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 523, + 554, + 577 + ], + "spans": [ + { + "bbox": [ + 317, + 523, + 554, + 577 + ], + "type": "text", + "content": "[21] Junnan Li, Dongxu Li, Silvio Savarese, and Steven Hoi. Blip-2: Bootstrapping language-image pre-training with frozen image encoders and large language models. In International conference on machine learning, pages 19730-19742. PMLR, 2023. 1, 2, 4" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 317, + 579, + 554, + 643 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 579, + 554, + 643 + ], + "spans": [ + { + "bbox": [ + 317, + 579, + 554, + 643 + ], + "type": "text", + "content": "[22] Zhiqi Li, Wenhai Wang, Hongyang Li, Enze Xie, Chonghao Sima, Tong Lu, Yu Qiao, and Jifeng Dai. Bevformer: Learning bird's-eye-view representation from multi-camera images via spatiotemporal transformers. In European conference on computer vision, pages 1-18. Springer, 2022. 1, 2, 3" + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 316, + 647, + 553, + 700 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 647, + 553, + 700 + ], + "spans": [ + { + "bbox": [ + 316, + 647, + 553, + 700 + ], + "type": "text", + "content": "[23] Xiwen Liang, Yangxin Wu, Jianhua Han, Hang Xu, Chunjing Xu, and Xiaodan Liang. Effective adaptation in multitask co-training for unified autonomous driving. Advances in Neural Information Processing Systems, 35:19645-19658, 2022. 2" + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 316, + 702, + 553, + 714 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 702, + 553, + 714 + ], + "spans": [ + { + "bbox": [ + 316, + 702, + 553, + 714 + ], + "type": "text", + "content": "[24] Haotian Liu, Chunyuan Li, Qingyang Wu, and Yong Jae Lee." + } + ] + } + ], + "index": 25 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "text", + "content": "27583" + } + ] + } + ], + "index": 27 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 8 + }, + { + "para_blocks": [ + { + "bbox": [ + 56, + 73, + 295, + 713 + ], + "type": "list", + "angle": 0, + "index": 12, + "blocks": [ + { + "bbox": [ + 76, + 73, + 294, + 95 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 76, + 73, + 294, + 95 + ], + "spans": [ + { + "bbox": [ + 76, + 73, + 294, + 95 + ], + "type": "text", + "content": "Visual instruction tuning. Advances in neural information processing systems, 36, 2024. 1, 2, 5" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 56, + 96, + 295, + 152 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 96, + 295, + 152 + ], + "spans": [ + { + "bbox": [ + 56, + 96, + 295, + 152 + ], + "type": "text", + "content": "[25] Jiaqi Liu, Peng Hang, Xiao Qi, Jianqiang Wang, and Jian Sun. Mtd-gpt: A multi-task decision-making gpt model for autonomous driving at unsignalized intersections. In 2023 IEEE 26th International Conference on Intelligent Transportation Systems (ITSC), pages 5154-5161. IEEE, 2023. 2" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 56, + 153, + 294, + 206 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 153, + 294, + 206 + ], + "spans": [ + { + "bbox": [ + 56, + 153, + 294, + 206 + ], + "type": "text", + "content": "[26] Yicheng Liu, Jinghuai Zhang, Liangji Fang, Qinhong Jiang, and Bolei Zhou. Multimodal motion prediction with stacked transformers. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 7577-7586, 2021. 1, 2" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 56, + 209, + 294, + 262 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 209, + 294, + 262 + ], + "spans": [ + { + "bbox": [ + 56, + 209, + 294, + 262 + ], + "type": "text", + "content": "[27] Srikanth Malla, Chiho Choi, Isht Dwivedi, Joon Hee Choi, and Jiachen Li. Drama: Joint risk localization and captioning in driving. In Proceedings of the IEEE/CVF winter conference on applications of computer vision, pages 1043-1052, 2023. 2" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 56, + 264, + 294, + 297 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 264, + 294, + 297 + ], + "spans": [ + { + "bbox": [ + 56, + 264, + 294, + 297 + ], + "type": "text", + "content": "[28] Jiageng Mao, Yuxi Qian, Junjie Ye, Hang Zhao, and Yue Wang. Gpt-driver: Learning to drive with gpt. arXiv preprint arXiv:2310.01415, 2023. 2" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 56, + 300, + 294, + 364 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 300, + 294, + 364 + ], + "spans": [ + { + "bbox": [ + 56, + 300, + 294, + 364 + ], + "type": "text", + "content": "[29] Jiquan Ngiam, Benjamin Caine, Vijay Vasudevan, Zhengdong Zhang, Hao-Tien Lewis Chiang, Jeffrey Ling, Rebecca Roelofs, Alex Bewley, Chenxi Liu, Ashish Venugopal, et al. Scene transformer: A unified architecture for predicting multiple agent trajectories. arXiv preprint arXiv:2106.08417, 2021. 1, 2" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 56, + 366, + 294, + 431 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 366, + 294, + 431 + ], + "spans": [ + { + "bbox": [ + 56, + 366, + 294, + 431 + ], + "type": "text", + "content": "[30] Chenbin Pan, Burhaneddin Yaman, Tommaso Nesti, Abhirup Mallik, Alessandro G Allievi, Senem Velipasalar, and Liu Ren. Vlp: Vision language planning for autonomous driving. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 14760-14769, 2024. 6, 7" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 56, + 434, + 294, + 498 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 434, + 294, + 498 + ], + "spans": [ + { + "bbox": [ + 56, + 434, + 294, + 498 + ], + "type": "text", + "content": "[31] Jonah Philion and Sanja Fidler. Lift, splat, shoot: Encoding images from arbitrary camera rigs by implicitly unprojecting to 3d. In Computer Vision-ECCV 2020: 16th European Conference, Glasgow, UK, August 23-28, 2020, Proceedings, Part XIV 16, pages 194-210. Springer, 2020. 1, 2" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 56, + 500, + 294, + 555 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 500, + 294, + 555 + ], + "spans": [ + { + "bbox": [ + 56, + 500, + 294, + 555 + ], + "type": "text", + "content": "[32] Tianwen Qian, Jingjing Chen, Linhai Zhuo, Yang Jiao, and Yu-Gang Jiang. Nuscenes-qa: A multi-modal visual question answering benchmark for autonomous driving scenario. In Proceedings of the AAAI Conference on Artificial Intelligence, pages 4542–4550, 2024. 2" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 56, + 557, + 294, + 621 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 557, + 294, + 621 + ], + "spans": [ + { + "bbox": [ + 56, + 557, + 294, + 621 + ], + "type": "text", + "content": "[33] Alec Radford, Jong Wook Kim, Chris Hallacy, Aditya Ramesh, Gabriel Goh, Sandhini Agarwal, Girish Sastry, Amanda Askell, Pamela Mishkin, Jack Clark, et al. Learning transferable visual models from natural language supervision. In International conference on machine learning, pages 8748-8763. PMLR, 2021. 2" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 56, + 624, + 294, + 689 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 624, + 294, + 689 + ], + "spans": [ + { + "bbox": [ + 56, + 624, + 294, + 689 + ], + "type": "text", + "content": "[34] Enna Sachdeva, Nakul Agarwal, Suhas Chundi, Sean Roelofs, Jiachen Li, Mykel Kochenderfer, Chiho Choi, and Behzad Dariush. Rank2tell: A multimodal driving dataset for joint importance ranking and reasoning. In Proceedings of the IEEE/CVF Winter Conference on Applications of Computer Vision, pages 7513-7522, 2024. 2" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 56, + 691, + 294, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 691, + 294, + 713 + ], + "spans": [ + { + "bbox": [ + 56, + 691, + 294, + 713 + ], + "type": "text", + "content": "[35] Hao Sha, Yao Mu, Yuxuan Jiang, Li Chen, Chenfeng Xu, Ping Luo, Shengbo Eben Li, Masayoshi Tomizuka, Wei" + } + ] + } + ], + "index": 11 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 316, + 73, + 553, + 713 + ], + "type": "list", + "angle": 0, + "index": 25, + "blocks": [ + { + "bbox": [ + 333, + 73, + 553, + 106 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 333, + 73, + 553, + 106 + ], + "spans": [ + { + "bbox": [ + 333, + 73, + 553, + 106 + ], + "type": "text", + "content": "Zhan, and Mingyu Ding. *Languagempc: Large language models as decision makers for autonomous driving.* arXiv preprint arXiv:2310.03026, 2023. 2" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 316, + 108, + 553, + 172 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 108, + 553, + 172 + ], + "spans": [ + { + "bbox": [ + 316, + 108, + 553, + 172 + ], + "type": "text", + "content": "[36] Hao Shao, Yuxuan Hu, Letian Wang, Guanglu Song, Steven L Waslander, Yu Liu, and Hongsheng Li. Lmdrive: Closed-loop end-to-end driving with large language models. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 15120-15130, 2024. 2, 3" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 316, + 175, + 553, + 228 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 175, + 553, + 228 + ], + "spans": [ + { + "bbox": [ + 316, + 175, + 553, + 228 + ], + "type": "text", + "content": "[37] Chonghao Sima, Katrin Renz, Kashyap Chitta, Li Chen, Hanxue Zhang, Chengen Xie, Ping Luo, Andreas Geiger, and Hongyang Li. Drivelm: Driving with graph visual question answering. arXiv preprint arXiv:2312.14150, 2023. 1, 2, 4, 5, 8" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 316, + 230, + 553, + 285 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 230, + 553, + 285 + ], + "spans": [ + { + "bbox": [ + 316, + 230, + 553, + 285 + ], + "type": "text", + "content": "[38] Ran Tian, Boyi Li, Xinshuo Weng, Yuxiao Chen, Edward Schmerling, Yue Wang, Boris Ivanovic, and Marco Pavone. Tokenize the world into object-level knowledge to address long-tail events in autonomous driving. arXiv preprint arXiv:2407.00959, 2024. 1, 2, 3, 6, 7" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 316, + 287, + 553, + 341 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 287, + 553, + 341 + ], + "spans": [ + { + "bbox": [ + 316, + 287, + 553, + 341 + ], + "type": "text", + "content": "[39] Xiaoyu Tian, Junru Gu, Bailin Li, Yicheng Liu, Chenxu Hu, Yang Wang, Kun Zhan, Peng Jia, Xianpeng Lang, and Hang Zhao. Drivevlm: The convergence of autonomous driving and large vision-language models. arXiv preprint arXiv:2402.12289, 2024. 1, 2, 3, 6, 7" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 316, + 343, + 553, + 397 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 343, + 553, + 397 + ], + "spans": [ + { + "bbox": [ + 316, + 343, + 553, + 397 + ], + "type": "text", + "content": "[40] Wenwen Tong, Chonghao Sima, Tai Wang, Li Chen, Silei Wu, Hanming Deng, Yi Gu, Lewei Lu, Ping Luo, Dahua Lin, et al. Scene as occupancy. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 8406-8415, 2023. 3" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 316, + 399, + 553, + 454 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 399, + 553, + 454 + ], + "spans": [ + { + "bbox": [ + 316, + 399, + 553, + 454 + ], + "type": "text", + "content": "[41] Hugo Touvron, Thibaut Lavril, Gautier Izacard, Xavier Martinet, Marie-Anne Lachaux, Timothée Lacroix, Baptiste Rozière, Naman Goyal, Eric Hambro, Faisal Azhar, et al. Llama: Open and efficient foundation language models. arXiv preprint arXiv:2302.13971, 2023. 2" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 316, + 456, + 553, + 510 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 456, + 553, + 510 + ], + "spans": [ + { + "bbox": [ + 316, + 456, + 553, + 510 + ], + "type": "text", + "content": "[42] Hugo Touvron, Louis Martin, Kevin Stone, Peter Albert, Amjad Almahairi, Yasmine Babaei, Nikolay Bashlykov, Soumya Batra, Prajjwal Bhargava, Shruti Bhosale, et al. Llama 2: Open foundation and fine-tuned chat models. arXiv preprint arXiv:2307.09288, 2023. 2" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 316, + 512, + 553, + 567 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 512, + 553, + 567 + ], + "spans": [ + { + "bbox": [ + 316, + 512, + 553, + 567 + ], + "type": "text", + "content": "[43] Shihao Wang, Zhiding Yu, Xiaohui Jiang, Shiyi Lan, Min Shi, Nadine Chang, Jan Kautz, Ying Li, and Jose M Alvarez. Omnidrive: A holistic llm-agent framework for autonomous driving with 3d perception, reasoning and planning. arXiv preprint arXiv:2405.01533, 2024. 1, 3, 6, 7" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 316, + 568, + 553, + 632 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 568, + 553, + 632 + ], + "spans": [ + { + "bbox": [ + 316, + 568, + 553, + 632 + ], + "type": "text", + "content": "[44] Wenhai Wang, Jiangwei Xie, ChuanYang Hu, Haoming Zou, Jianan Fan, Wenwen Tong, Yang Wen, Silei Wu, Hanming Deng, Zhiqi Li, et al. Driveml: Aligning multi-modal large language models with behavioral planning states for autonomous driving. arXiv preprint arXiv:2312.09245, 2023. 1, 2, 3" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 316, + 635, + 553, + 678 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 635, + 553, + 678 + ], + "spans": [ + { + "bbox": [ + 316, + 635, + 553, + 678 + ], + "type": "text", + "content": "[45] Zitian Wang, Zehao Huang, Yulu Gao, Naiyan Wang, and Si Liu. Mv2dfusion: Leveraging modality-specific object semantics for multi-modal 3d detection. arXiv preprint arXiv:2408.05945, 2024. 1, 2" + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 316, + 680, + 553, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 680, + 553, + 713 + ], + "spans": [ + { + "bbox": [ + 316, + 680, + 553, + 713 + ], + "type": "text", + "content": "[46] Jason Wei, Xuezhi Wang, Dale Schuurmans, Maarten Bosma, Fei Xia, Ed Chi, Quoc V Le, Denny Zhou, et al. Chain-of-thought prompting elicits reasoning in large lan" + } + ] + } + ], + "index": 24 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 318, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 318, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 318, + 757 + ], + "type": "text", + "content": "27584" + } + ] + } + ], + "index": 26 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 9 + }, + { + "para_blocks": [ + { + "bbox": [ + 56, + 73, + 295, + 596 + ], + "type": "list", + "angle": 0, + "index": 10, + "blocks": [ + { + "bbox": [ + 76, + 73, + 294, + 95 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 76, + 73, + 294, + 95 + ], + "spans": [ + { + "bbox": [ + 76, + 73, + 294, + 95 + ], + "type": "text", + "content": "guage models. Advances in neural information processing systems, 35:24824-24837, 2022. 1, 2" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 56, + 96, + 295, + 151 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 96, + 295, + 151 + ], + "spans": [ + { + "bbox": [ + 56, + 96, + 295, + 151 + ], + "type": "text", + "content": "[47] Licheng Wen, Daocheng Fu, Xin Li, Xinyu Cai, MA Tao, Pinlong Cai, Min Dou, Botian Shi, Liang He, and Yu Qiao. Dilu: A knowledge-driven approach to autonomous driving with large language models. In The Twelfth International Conference on Learning Representations. 2" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 56, + 152, + 294, + 207 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 152, + 294, + 207 + ], + "spans": [ + { + "bbox": [ + 56, + 152, + 294, + 207 + ], + "type": "text", + "content": "[48] Xinshuo Weng, Boris Ivanovic, Yan Wang, Yue Wang, and Marco Pavone. Para-drive: Parallelized architecture for real-time autonomous driving. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 15449-15458, 2024. 1, 2, 3, 5, 6, 7" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 56, + 208, + 294, + 262 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 208, + 294, + 262 + ], + "spans": [ + { + "bbox": [ + 56, + 208, + 294, + 262 + ], + "type": "text", + "content": "[49] Penghao Wu, Xiaosong Jia, Li Chen, Junchi Yan, Hongyang Li, and Yu Qiao. Trajectory-guided control prediction for end-to-end autonomous driving: A simple yet strong baseline. Advances in Neural Information Processing Systems, 35:6119-6132, 2022. 1, 2" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 56, + 263, + 294, + 317 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 263, + 294, + 317 + ], + "spans": [ + { + "bbox": [ + 56, + 263, + 294, + 317 + ], + "type": "text", + "content": "[50] Zhenhua Xu, Yujia Zhang, Enze Xie, Zhen Zhao, Yong Guo, Kwan-Yee K Wong, Zhenguo Li, and Hengshuang Zhao. Drivegpt4: Interpretable end-to-end autonomous driving via large language model. IEEE Robotics and Automation Letters, 2024. 2" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 56, + 319, + 294, + 384 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 319, + 294, + 384 + ], + "spans": [ + { + "bbox": [ + 56, + 319, + 294, + 384 + ], + "type": "text", + "content": "[51] Rajeev Yasarla, Manish Kumar Singh, Hong Cai, Yunxiao Shi, Jisoo Jeong, Yinhao Zhu, Shizhong Han, Risheek Garrepalli, and Fatih Porikli. *Futuredepth: Learning to predict the future improves video depth estimation. In European Conference on Computer Vision*, pages 440–458. Springer, 2024. 4" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 56, + 386, + 295, + 430 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 386, + 295, + 430 + ], + "spans": [ + { + "bbox": [ + 56, + 386, + 295, + 430 + ], + "type": "text", + "content": "[52] Tianwei Yin, Xingyi Zhou, and Philipp Krahenbuhl. Center-based 3d object detection and tracking. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 11784-11793, 2021. 1, 2" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 56, + 431, + 294, + 485 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 431, + 294, + 485 + ], + "spans": [ + { + "bbox": [ + 56, + 431, + 294, + 485 + ], + "type": "text", + "content": "[53] Wenyuan Zeng, Wenjie Luo, Simon Suo, Abbas Sadat, Bin Yang, Sergio Casas, and Raquel Urtasun. End-to-end interpretable neural motion planner. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 8660-8669, 2019. 2" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 56, + 487, + 294, + 540 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 487, + 294, + 540 + ], + "spans": [ + { + "bbox": [ + 56, + 487, + 294, + 540 + ], + "type": "text", + "content": "[54] Jiang-Tian Zhai, Ze Feng, Jinhao Du, Yongqiang Mao, Jiang-Jiang Liu, Zichang Tan, Yifu Zhang, Xiaqing Ye, and Jingdong Wang. Rethinking the open-loop evaluation of end-to-end autonomous driving in nuscenes. arXiv preprint arXiv:2305.10430, 2023. 5, 6" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 56, + 542, + 294, + 596 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 542, + 294, + 596 + ], + "spans": [ + { + "bbox": [ + 56, + 542, + 294, + 596 + ], + "type": "text", + "content": "[55] Yunpeng Zhang, Zheng Zhu, Wenzhao Zheng, Junjie Huang, Guan Huang, Jie Zhou, and Jiwen Lu. Beverse: Unified perception and prediction in birds-eye-view for vision-centric autonomous driving. arXiv preprint arXiv:2205.09743, 2022. 2" + } + ] + } + ], + "index": 9 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "text", + "content": "27585" + } + ] + } + ], + "index": 11 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 10 + } + ], + "_backend": "vlm", + "_version_name": "2.6.4" +} \ No newline at end of file diff --git a/2025/Distilling Spatially-Heterogeneous Distortion Perception for Blind Image Quality Assessment/c7a78868-95dc-4d35-97c7-04bbcfc1063d_content_list.json b/2025/Distilling Spatially-Heterogeneous Distortion Perception for Blind Image Quality Assessment/c7a78868-95dc-4d35-97c7-04bbcfc1063d_content_list.json new file mode 100644 index 0000000000000000000000000000000000000000..43045dbfe5dd25a20be41f406af5f2adaa2598ca --- /dev/null +++ b/2025/Distilling Spatially-Heterogeneous Distortion Perception for Blind Image Quality Assessment/c7a78868-95dc-4d35-97c7-04bbcfc1063d_content_list.json @@ -0,0 +1,1938 @@ +[ + { + "type": "text", + "text": "Distilling Spatially-Heterogeneous Distortion Perception for Blind Image Quality Assessment", + "text_level": 1, + "bbox": [ + 94, + 128, + 903, + 174 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Xudong Li $^{1}$ Wenjie Nie $^{1}$ Yan Zhang $^{1}$ Runze Hu $^{2}$ Ke Li $^{3}$ Xiawu Zheng $^{1}$ Liujuan Cao $^{1*}$ $^{1}$ Key Laboratory of Multimedia Trusted Perception and Efficient Computing, Ministry of Education of China, Xiamen University, 361005, P.R. China \n $^{2}$ School of Information and Electronics, Beijing Institute of Technology, Beijing, China \n $^{3}$ Tencent Youtu Lab, Shanghai, China", + "bbox": [ + 102, + 202, + 898, + 290 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "{1xd761050753, bzhy986, hrzlpk2015}@gmail.com, {zhengxiwu, caoliujuan}@xmu.edu.cn", + "bbox": [ + 130, + 294, + 854, + 308 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Abstract", + "text_level": 1, + "bbox": [ + 246, + 344, + 326, + 359 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "In the Blind Image Quality Assessment (BIQA) field, accurately assessing the quality of authentically distorted images presents a substantial challenge due to the diverse distortion types in natural settings. Existing state-of-the-art IQA methods mix a sequence of distortions into entire images to establish global distortion priors, but are inadequate for authentic images with spatially varied distortions. To address this, we introduce a novel IQA framework that employs knowledge distillation tailored to perceive spatially heterogeneous distortions, enhancing quality-distortion awareness. Specifically, we introduce a novel Block-wise Degradation Modelling approach that applies distinct distortions to different spatial blocks of an image, thereby expanding local distortion priors. Following this, we present a Block-wise Aggregation and Filtering module that enables fine-grained attention to the quality information within different distortion areas of the image. Furthermore, to effectively capture the complex relationships between distortions across different regions while preserving overall quality perception, we introduce Contrastive Knowledge Distillation to enhance the model's ability to discriminate between different types of distortions and Affinity Knowledge Distillation to model the correlation among distortions in different regions. Extensive experiments on standard BIQA datasets demonstrate the effectiveness and competitiveness of the proposed method.", + "bbox": [ + 86, + 376, + 485, + 755 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "1. Introduction", + "text_level": 1, + "bbox": [ + 91, + 786, + 220, + 801 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Image Quality Assessment (IQA) methods are designed to evaluate images in a way that aligns with the human judgment of visual distortions [47, 49]. These methods are crucial for applications that rely heavily on image quality", + "bbox": [ + 89, + 811, + 483, + 875 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/af43d3273e4f04c39fb57ac9e8a44dd8d1ea153788daf1b6fb90d96c50f15752.jpg", + "image_caption": [ + "Figure 1. Comparison of state-of-the-art (SOTA) BIQA methods on the largest authentic dataset, LIEVFB. The proposed SHDIQA method outperforms all existing BIQA approaches." + ], + "image_footnote": [], + "bbox": [ + 524, + 342, + 893, + 547 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "and serve as standards to gauge the performance of image processing techniques. Objective Image Quality Assessment includes Full-Reference IQA (FR-IQA) [36], Reduced-Reference IQA (RR-IQA) [41], and no-reference IQA (NR-IQA) or Blind IQA (BIQA) [60]. As the presence of various types of artifacts in real-world scenes and the unavailability of reference images, BIQA gains attention for tasks like image restoration [1] and super-resolution [6].", + "bbox": [ + 511, + 613, + 906, + 733 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Recent advances in data-driven BIQA models, especially those based on deep neural networks, have led to impressive progress [37, 51]. However, these models still rely heavily on Mean Opinion Scores (MOS), which are expensive and time-consuming to collect—often requiring up to 120 human ratings per image [13]. As an alternative, synthetic distortions generated through controlled data augmentation offer a cost-effective way to create annotated datasets [48], enabling the learning of quality representations without extensive human labeling [38]. To leverage such synthetic data, recent studies have introduced multi-task end-to-end", + "bbox": [ + 511, + 734, + 908, + 900 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "CVF", + "bbox": [ + 106, + 2, + 181, + 42 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore.", + "bbox": [ + 236, + 0, + 810, + 44 + ], + "page_idx": 0 + }, + { + "type": "page_footnote", + "text": "*Corresponding author.", + "bbox": [ + 109, + 887, + 235, + 898 + ], + "page_idx": 0 + }, + { + "type": "page_number", + "text": "2344", + "bbox": [ + 482, + 944, + 514, + 955 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "frameworks [26, 57] that jointly optimize distortion-specific objectives and quality prediction, using synthetic distortions to enrich training signals and improve generalization. Another line of work focuses on image degradation modeling. Methods like Re-IQA [31] and QPT [58] simulate realistic distortions by applying diverse types and combinations to entire images, as illustrated in Fig.2. They then leverage unsupervised learning frameworks such as MoCo-v2[3] to extract distortion priors from large-scale synthetic datasets.", + "bbox": [ + 89, + 90, + 483, + 227 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "However, both categories of methods share a key limitation: they focus on global distortion priors while ignoring fine-grained local distortions, which are essential for accurate BIQA. Research [15, 17] shows that the human visual system (HVS) employs both local detection and global appearance strategies when assessing image quality. For subtle, near-threshold distortions, the HVS prioritizes local details, while for highly visible distortions, it shifts focus back to global content. This highlights the vital role of local distortion features in comprehensive image quality evaluation.", + "bbox": [ + 89, + 229, + 485, + 381 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "In this paper, we propose a novel Spatial Heterogeneous Distortion-aware Image Quality Assessment (SHDIQA) method, designed to improve the local distortion-awareness of BIQA models based on the Vision Transformer (ViT) architecture. Unlike prior works [10, 14, 50] that focus on modifying ViT architectures to enhance local feature extraction, our method fully leverages rich local distortion priors generated from synthetic distortions, while introducing minimal architectural changes (only 0.2M additional parameters).", + "bbox": [ + 89, + 383, + 483, + 520 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Specifically, we introduce a Block-wise Image Degradation (BWID) method tailored for IQA tasks to generate patch-level distortion priors. As shown in Fig. 2, BWID partitions a distorted image into multiple blocks and applies a randomly selected distortion from 24 categories across 4 severity levels to each block. This results in a composite image with spatially varying local distortions. Each image naturally comes with an image-level multi-class label and corresponding patch-level single-class labels, effectively addressing the lack of localized distortion annotations in existing methods. To further utilize the patch-level distortion priors, we propose a Block-wise Aggregation and Filtering (BwAF) mechanism. BWAF uses the local priors from BWID to train multiple distortion (DIS) tokens, which interact with a global classification (CLS) token through a cross-attention mechanism with an adaptive head. This interaction mimics the HVS's dual strategy of local detection and global perception, thereby improving overall quality assessment, as illustrated in Fig. 1. Moreover, previous studies [16, 40] highlight the importance of modeling the relationships among different distortion types in perceptual quality assessment. To achieve this, we incorporate Contrastive Knowledge Distillation (CKD) and Affinity Knowledge Distillation (AKD). CKD aligns similar distortions and separates dissimilar ones, improving the model's ability to differentiate between distor", + "bbox": [ + 89, + 523, + 486, + 901 + ], + "page_idx": 1 + }, + { + "type": "image", + "img_path": "images/39dcb6f888a06e9d61a88a1f100001fdbd5417b9d2a1a1438bb4f2dbaab49f47.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 527, + 88, + 599, + 152 + ], + "page_idx": 1 + }, + { + "type": "image", + "img_path": "images/837659c2d179a6b916b9df84f03f7098ce6b8bf1298f04997550a949545fdae1.jpg", + "image_caption": [ + "Distorted Region" + ], + "image_footnote": [], + "bbox": [ + 604, + 88, + 673, + 152 + ], + "page_idx": 1 + }, + { + "type": "image", + "img_path": "images/4df4db78557de28709340952da50a92c7ec1a028ff4ede992198c7044b2c3091.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 679, + 88, + 750, + 152 + ], + "page_idx": 1 + }, + { + "type": "image", + "img_path": "images/27bc660bbab2540866e20536f78f10adb35e9726efa7abb17bd9b31e495bf5f6.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 756, + 88, + 823, + 152 + ], + "page_idx": 1 + }, + { + "type": "image", + "img_path": "images/e5fb6976fdeab7ec32586b2927b37b562015885be505b69cd2c42f00de732c51.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 826, + 101, + 892, + 152 + ], + "page_idx": 1 + }, + { + "type": "image", + "img_path": "images/18b4a0e21421b0d1ce2d31e31f3b771f5e19cea4b474ccffb54275e536104430.jpg", + "image_caption": [ + "Image B", + "Figure 2. The proposed SHDIQA method creates a heterogeneous degradation space with richer priors (Top) and assigns dedicated distortion-aware (DIS) tokens to each local region (Bottom)." + ], + "image_footnote": [], + "bbox": [ + 527, + 162, + 598, + 218 + ], + "page_idx": 1 + }, + { + "type": "image", + "img_path": "images/39519b5f774f3b6d3c4efdbd7cf977780d50ece87c408c52e93f5d4064e6ee76.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 606, + 162, + 674, + 217 + ], + "page_idx": 1 + }, + { + "type": "image", + "img_path": "images/8d6022ea40d2f18ae1b475560218b9c176efc22bae8fc91e226020088fb41f77.jpg", + "image_caption": [ + "DEIQT Method" + ], + "image_footnote": [], + "bbox": [ + 678, + 162, + 750, + 218 + ], + "page_idx": 1 + }, + { + "type": "image", + "img_path": "images/7a4fa361f84280d044140b0a0c07cb7245517ddbfa3aa42163a9a480442c0d5b.jpg", + "image_caption": [ + "SHDIOA(Ours)" + ], + "image_footnote": [], + "bbox": [ + 754, + 162, + 823, + 218 + ], + "page_idx": 1 + }, + { + "type": "image", + "img_path": "images/eec06aa26b83c6b2f3ed31be17489cee36b06e20833592f33afe7db30475a55f.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 826, + 162, + 893, + 218 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "tion types. Meanwhile, AKD transfers knowledge about the relationships between different distortions, enhancing the model's generalization across various types of distortions.", + "bbox": [ + 511, + 292, + 906, + 339 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Our contributions are summarized as follows:", + "bbox": [ + 532, + 339, + 834, + 353 + ], + "page_idx": 1 + }, + { + "type": "list", + "sub_type": "text", + "list_items": [ + "- We present a Block-wise Degradation Modelling approach that enriches the local distortion priors of the training datasets by dividing images into blocks and applying distinct distortions to each, enhancing the model's ability to assess image quality accurately in complex scenarios.", + "- We introduce the Block-wise Aggregation and Filtering mechanism to fully capture and leverage localized distortions. It first directs precise attention to distortion details within image blocks, and then utilizes an adaptation head to filter out irrelevant distortion features, focusing the analysis on quality-aware information.", + "- We propose a novel Distortion-Guided Knowledge Distillation strategy that combines Contrastive Knowledge Distillation (CKD) to enhance distortion discrimination and Affinity Knowledge Distillation (AKD) to enrich relational insights between local distortion areas." + ], + "bbox": [ + 513, + 354, + 906, + 595 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "2. Related Work", + "text_level": 1, + "bbox": [ + 513, + 611, + 653, + 626 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "2.1. BIQA with Degradation Modelling", + "text_level": 1, + "bbox": [ + 513, + 636, + 816, + 652 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Recent studies [10, 61] have shown that simple distortion simulations do not fully capture the complexities of real-world degradations. To address this gap, Image Degradation Modeling has been introduced to represent various distortion types [34, 45, 46]. DBCNN [56] decomposes real distortions into multiple synthetic ones, creating large-scale synthetic images for training. DipIQ [25] categorizes common distortions like noise, blur, and compression into severity levels, while Re-IQA [31] introduces 25 distortion types to expand the degradation space. QPT [58] models second-order degradation using skip and shuffle operations across nine types in three groups. In contrast to these global degradation approaches, our method applies localized distortions to different regions of the image, enabling the model to learn fine-grained distortion perception both within and across blocks through knowledge distillation.", + "bbox": [ + 511, + 659, + 908, + 900 + ], + "page_idx": 1 + }, + { + "type": "page_number", + "text": "2345", + "bbox": [ + 482, + 944, + 514, + 955 + ], + "page_idx": 1 + }, + { + "type": "image", + "img_path": "images/2a04a5c8e746f1fed82b421526c8abe6b8c9226a75bb2d8c09b5c310e99bc42e.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 94, + 88, + 305, + 210 + ], + "page_idx": 2 + }, + { + "type": "image", + "img_path": "images/10ca39a192719452500ef9bfbddc705bf801eb2bf8949e7d2390ab00860304f1.jpg", + "image_caption": [ + "(a) Block-wise Degradation" + ], + "image_footnote": [], + "bbox": [ + 94, + 227, + 307, + 325 + ], + "page_idx": 2 + }, + { + "type": "image", + "img_path": "images/b6f136828165943658dee6db4e344ed6a83283f4bffe6bdb685065cff5e22d3b.jpg", + "image_caption": [ + "Figure 3. Framework Overview. Given an input image $\\mathbf{I}$ , (a) Block-wise Degradation is first applied to generate a degraded version $\\mathbf{I}_d$ (Sec. 3.1). Both $\\mathbf{I}$ and $\\mathbf{I}_d$ are then fed into a student encoder-decoder network, which is initialized with multiple distortion (DIS) tokens (each attending to a specific local block) and a CLS token for global quality representation. (b) Guided by the Block-wise Aggregation (Sec. 3.2), the DIS tokens and the CLS token are used to extract local distortion features $\\mathcal{F}_{dis}$ and a global quality feature $\\mathcal{F}_{cls}$ , which are further refined by the decoder into $\\hat{\\mathcal{F}}_{dis}$ and $\\hat{\\mathcal{F}}_{cls}$ . These refined features are processed by a quality-aware adaptation head to suppress irrelevant information and predict the quality score via an MLP. (c) To further enhance distortion perception, Contrastive KD is used to improve inter-block distortion discrimination, while Affinity KD captures relationships between distortions across different regions (Sec. 3.3)." + ], + "image_footnote": [], + "bbox": [ + 326, + 89, + 906, + 327 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "2.2. BIQA with Transformers", + "text_level": 1, + "bbox": [ + 89, + 450, + 320, + 465 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Vision Transformer (ViT)-based IQA architectures include hybrid transformers [2, 10, 21, 50, 54] and pure transformers [14, 20, 29]. Hybrid models integrate multi-scale CNN information for better local distortion perception but increase complexity. Pure transformers like DEIQT [29] capture global abstraction using the CLS token but struggle with local structure modeling [63]. To reduce memory overhead, methods using local self-attention and sparse block structures have been developed [23, 30], though at the cost of some long-distance modeling. Sampling-based methods like VT-IQA [17] combine global and local information by resampling distorted images and using global attention. However, they lack explicit supervision for local distortions, which limits their ability to capture localized degradations during training. In contrast, our approach enhances local distortion perception in pure ViT by incorporating patch-level distortion priors, while preserving global context modeling through a mask matrix specifically applied to distortion tokens, leaving the self-attention process unaffected.", + "bbox": [ + 88, + 472, + 485, + 758 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "2.3. IQA with Knowledge Distillation", + "text_level": 1, + "bbox": [ + 89, + 771, + 377, + 787 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Knowledge Distillation (KD), originally proposed for model compression [12], has been adapted for tasks like action recognition [8], re-identification [28], and depth estimation [11]. Recently, KD has been used in methods like Degraded-Reference IQA (DR-IQA) [59] and Non-aligned Reference IQA (NAR-IQA) [52] to transfer reference information, though these still partially rely on reference images.", + "bbox": [ + 89, + 794, + 483, + 900 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "A related BIQA method is SSLIQA [55], which employs KD to enforce semantic consistency across different triples within a batch. However, our approach differs significantly in that our KD explicitly targets inter-block and intra-block distortion information, allowing for more direct extraction of quality-related distortion priors from the augmented image.", + "bbox": [ + 511, + 450, + 906, + 542 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3. Methodology", + "text_level": 1, + "bbox": [ + 511, + 556, + 648, + 574 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "We introduce Spatial-Heterogeneous Distortion-aware Image Quality Assessment (SHDIQA), a framework designed to enhance the perception of local and non-uniform distortions in authentic images, as illustrated in Fig. 3. Specifically, the proposed SHDIQA first divides the input image $I$ into vertical blocks and applies random distortions using a Block-wise Degradation module (Sec. 3.1), producing a degraded image $I_d$ . Both $I$ and $I_d$ are then processed by a student encoder-decoder network equipped with a Blockwise Aggregation module (Sec. 3.2), which extracts local distortion features $\\mathcal{F}_{dis}$ via multiple distortion (DIS) tokens and global quality feature $\\mathcal{F}_{cls}$ via classification (CLS) token. Next, the local and global features $\\mathcal{F}_{dis}$ and $\\mathcal{F}_{cls}$ are fused via a cross-attention mechanism in the adaptation head to suppress irrelevant distortion information (Sec. 3.2). To enhance the model's ability to represent and distinguish distortion patterns, we further introduce Contrastive Knowledge Distillation (CKD) and Affinity Knowledge Distillation (AKD) (Sec. 3.3). These modules jointly improve distortion discriminability and capture inter-block distortion correlations, enabling a more distortion-aware representation.", + "bbox": [ + 511, + 583, + 908, + 900 + ], + "page_idx": 2 + }, + { + "type": "page_number", + "text": "2346", + "bbox": [ + 482, + 944, + 514, + 955 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3.1. Block-wise Degradation Modelling", + "text_level": 1, + "bbox": [ + 89, + 90, + 393, + 107 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Recent methods [31, 58] apply multiple distortions uniformly across the entire image, ignoring the finer details of local distortions. To address this, we propose a Block-wise Degradation Modelling strategy, which applies 24 distortion types and 4 degradation levels to different regions of the image, thereby generating a composite local distortion prior. Specifically, given an image of size $W \\times H$ , we divide it into $K$ vertical blocks of size $W \\times \\frac{H}{K}$ . For each block $k$ ( $k = 1, \\dots, K$ ), we randomly select a rectangular region $C_k$ with top-left coordinate $\\alpha_i$ and size $w_k \\times h_k$ , where $w_k = \\beta_w \\times W$ and $h_k = \\beta_h \\times \\frac{H}{K}$ . Empirically, we find that setting $\\beta_w = \\beta_h = 1$ , i.e., degrading the entire block, yields the best performance. Then, the degradation is applied using the function $\\mathcal{D}_k(C_k; q_k, p_k)$ , where $q_k$ denotes the type of distortion (e.g., noise, darkening), and $p_k$ represents its intensity. The degraded image $I_d$ is generated as follows:", + "bbox": [ + 89, + 112, + 485, + 354 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\boldsymbol {I} _ {d} = \\mathcal {A} (\\boldsymbol {I}) = \\mathcal {D} _ {K} \\left(\\dots \\mathcal {D} _ {1} \\left(\\boldsymbol {C} _ {1}; q _ {1}, p _ {1}\\right) \\dots ; q _ {K}, p _ {K}\\right) \\tag {1}\n$$\n", + "text_format": "latex", + "bbox": [ + 107, + 363, + 483, + 381 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "where $K = 4$ denotes the number of degraded blocks. Unlike previous methods, our approach focuses on localized distortions, providing richer priors for feature aggregation (Sec. 3.2) and knowledge distillation (Sec. 3.3), improving the model's ability to handle complex real-world distortions.", + "bbox": [ + 89, + 388, + 483, + 465 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "3.2. Block-wise Aggregation and Filtering", + "text_level": 1, + "bbox": [ + 89, + 472, + 413, + 488 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "In this section, we introduce the Block-wise Aggregation method to extract the local distortion feature $\\mathcal{F}_{dis}$ from the degraded image $I_{d}$ and the input image $\\pmb{I}$ . We then present the Quality-aware Adaptation Head, which filters out quality-irrelevant distortion features from distortion feature $\\mathcal{F}_{dis}$ .", + "bbox": [ + 89, + 494, + 485, + 570 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "3.2.1. Block-wise Aggregation", + "text_level": 1, + "bbox": [ + 89, + 575, + 302, + 590 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "We begin by randomly initializing $K$ local distortion-aware (DIS) tokens, denoted as $\\mathcal{F}_{dis}$ , matching the number of degraded regions. Each DIS token $\\mathcal{F}_{dis}^{k}$ corresponds to a specific region $C_k$ . The resulting embedding sequence is defined as $\\mathcal{F} = \\{\\mathcal{F}_{dis};[\\mathcal{F}_{cls};\\mathcal{F}_{patch}]\\} \\in \\mathbb{R}^{(K + N)\\times D}$ , where $N$ is the number of image patches plus one CLS token, and $D$ is the embedding dimension. To ensure each DIS token focuses on its associated distorted region $C_k$ , we construct a mask map $\\boldsymbol{m} \\in \\mathbb{R}^{(K + N)}$ . Based on this, we define the attention mask $\\mathcal{M}$ to preserve the original global attention structure while enforcing localized attention for DIS tokens as follows:", + "bbox": [ + 89, + 595, + 485, + 775 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\nm _ {j} ^ {i} = \\left\\{ \\begin{array}{l l} \\mathbb {1} \\{\\text {t h e i - t h p a t c h o v e r l a p s w i t h C _ {i}} \\}, & j < N \\\\ 0, & j = N \\end{array} \\right.,\n$$\n", + "text_format": "latex", + "bbox": [ + 89, + 782, + 452, + 823 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\mathbb {M} = \\left[ m ^ {1}; m ^ {2}; \\dots ; m ^ {K} \\right] ^ {T}, \\mathcal {M} = \\left[ \\begin{array}{c c} \\mathbf {1} _ {N \\times N} & \\mathbf {0} _ {N \\times K} \\\\ \\mathbb {M} _ {K \\times N} & \\mathbf {1} _ {K \\times K} \\end{array} \\right], \\tag {2}\n$$\n", + "text_format": "latex", + "bbox": [ + 94, + 824, + 496, + 869 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "where $\\mathbb{1}\\{\\cdot\\}$ is the indicator function. Intuitively, Eq. 2 enforces that each DIS token $\\mathcal{F}_{dis}^{k}$ attends only to the distorted", + "bbox": [ + 89, + 869, + 483, + 901 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "content in block $C_k$ . Assuming the encoder has $H$ layers, the Multi-Head Self-Attention is applied as follows:", + "bbox": [ + 511, + 90, + 906, + 121 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\boldsymbol {\\mathcal {F}} ^ {h} = \\left\\{ \\begin{array}{l l} \\sigma (\\log \\mathcal {M} + \\mathbf {Q} ^ {h} (\\mathbf {K} ^ {h}) ^ {T}) \\mathbf {V} ^ {h} + \\boldsymbol {\\mathcal {F}} ^ {h - 1}, & 0 < h \\leq H \\\\ \\boldsymbol {\\mathcal {F}}, & h = 0 \\end{array} , \\right. \\tag {3}\n$$\n", + "text_format": "latex", + "bbox": [ + 511, + 131, + 916, + 183 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "where $Q^h, K^h, V^h$ are the query, key, and value matrices, obtained via linear projections of $\\mathcal{F}^{h-1}$ , and $\\sigma$ denotes the Softmax function. The encoder outputs features $\\mathcal{F}^H$ after $H$ layers. Following [29], we apply a transformer decoder to refine these features, yielding $\\{\\hat{\\mathcal{F}}_{dis}; \\hat{\\mathcal{F}}_{cls}\\} \\in \\mathbb{R}^{(K+1) \\times D}$ , which are more relevant for image quality.", + "bbox": [ + 511, + 183, + 906, + 275 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "3.2.2. Quality-aware Adaptation Head", + "text_level": 1, + "bbox": [ + 511, + 282, + 784, + 297 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "To effectively integrate global quality and local distortion features while suppressing quality-irrelevant information, we design a quality-aware adaptation head that highlights distortion cues closely related to perceived quality. Specifically, the global quality feature $\\hat{\\mathcal{F}}_{cls}$ is used as the query to selectively attend to the local distortion features $\\hat{\\mathcal{F}}_{dis}$ , which serve as key and value. This allows the model to emphasize distortion patterns most relevant to the overall quality. Overall, the final predicted quality score $\\hat{Y}$ is computed as:", + "bbox": [ + 511, + 300, + 906, + 436 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\hat {Y} = \\operatorname {L N} \\left(\\hat {\\mathcal {F}} _ {c l s} + \\text {C r o s s A t t e n t i o n} \\left(\\hat {\\mathcal {F}} _ {c l s}, \\hat {\\mathcal {F}} _ {d i s}\\right)\\right), \\tag {4}\n$$\n", + "text_format": "latex", + "bbox": [ + 535, + 446, + 906, + 472 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "where $\\text{CrossAttention}(\\cdot)$ denotes a cross-attention module, and $\\text{LN}$ is a single-layer linear projection. During training, the model is optimized by minimizing the smooth $\\mathcal{L}_1$ loss.", + "bbox": [ + 511, + 481, + 906, + 527 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "3.3. Distortion-Guided Knowledge Distillation", + "text_level": 1, + "bbox": [ + 511, + 535, + 870, + 551 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Inspired by [5, 18, 33], we adopt a knowledge distillation method to learn correlations among distortion types by transferring quality-aware distortion knowledge from the teacher to the student. This distillation strategy addresses two key challenges: First, the student's ViT encoders, pre-trained on ImageNet [4], focus on semantics and lack distortion awareness in early training [19]. The teacher, pre-trained on IQA data, provides quality-aware guidance, directing the student to critical distortion patterns. Second, the absence of explicit ground truth for distortion correlations makes it difficult for the student to model these relationships independently. The teacher acts as a reliable target for correlation modeling, thereby improving the student's discriminative ability.", + "bbox": [ + 511, + 556, + 908, + 753 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "3.3.1. Contrastive Knowledge Distillation", + "text_level": 1, + "bbox": [ + 511, + 761, + 803, + 775 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Contrastive Knowledge Distillation (CKD) aims to extract the distortion features $\\pmb{V}_{dis}^{k}$ of the $k$ -th local block from a teacher encoder $\\mathcal{V}(\\cdot)$ and transfer them to a student network. Simply maximizing feature similarity is insufficient to capture distinct distortion representations across spatial domains (Table 5). To address this, we reformulate the KD task as a pseudo-classification problem. As shown in Fig. 3, features from corresponding distorted regions of the teacher", + "bbox": [ + 511, + 779, + 908, + 900 + ], + "page_idx": 3 + }, + { + "type": "page_number", + "text": "2347", + "bbox": [ + 480, + 944, + 514, + 955 + ], + "page_idx": 3 + }, + { + "type": "table", + "img_path": "images/eb6457f9027d6faf21500bff8ae4e838a05866d8091b9b1cb4f9122a35eb8e17.jpg", + "table_caption": [], + "table_footnote": [], + "table_body": "
MethodLIVECSIQKADIDLIVECKonIQLIVEFBSPAQ
PLCCSRCCPLCCSRCCPLCCSRCCPLCCSRCCPLCCSRCCPLCCSRCCPLCCSRCC
BRISQUE [27]0.9440.9290.7480.8120.5670.5280.6290.6290.6850.6810.3410.3030.8170.809
DBCNN [56]0.9710.9680.9590.9460.8560.8510.8690.8510.8840.8750.5510.5450.9150.911
TIQA [54]0.9650.9490.8380.8250.8550.850.8610.8450.9030.8920.5810.541--
MetaIQA [61]0.9590.9600.9080.8990.7750.7620.8020.8350.8560.8870.5070.54--
HyperIQA (27M) [39]0.9660.9620.9420.9230.8450.8520.8820.8590.9170.9060.6020.5440.9150.911
TReS (152M) [10]0.9680.9690.9420.9220.8580.8590.8770.8460.9280.9150.6250.554--
MUSIQ (27M) [14]0.9110.9400.8930.8710.8720.8750.7460.7020.9280.9160.6610.5660.9210.918
Re-IQA (48M) [31]0.9710.9700.9600.9470.8850.8720.8540.8400.9230.914--0.9250.918
DEIQT (24M) [29]0.9820.9800.9630.9460.8870.8890.8940.8750.9340.9210.6630.5710.9230.919
QPT (24M) [58]------0.9140.8950.9410.9270.6750.5750.9280.925
LIQE (151M) [57]0.9510.9700.9390.9360.9310.9300.9100.9040.9080.919----
QFM-IQM (24M) [19]0.9830.9810.9650.9540.9060.9060.9130.8910.9360.9220.6670.5670.9240.920
LoDa (118M) [50]0.9790.975--0.9360.9310.8990.8760.9440.9320.6790.5780.9280.925
SHDIQA (24.2M)0.9840.9820.9720.9620.9400.9370.9220.9090.9480.9370.7350.6390.9340.930
", + "bbox": [ + 94, + 88, + 906, + 361 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Table 1. Performance comparison measured by the average of SRCC and PLCC. Best results are in bold, second-best are underlined.", + "bbox": [ + 104, + 369, + 888, + 383 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "and student are treated as positive pairs, while all other pairs are considered negative. The CKD loss is defined as:", + "bbox": [ + 89, + 398, + 482, + 429 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\begin{array}{l} \\mathcal {L} _ {c k d} = - \\frac {1}{K} \\sum_ {k} \\log \\left(\\sigma \\left(\\frac {\\cos \\left(\\boldsymbol {\\mathcal {F}} _ {d i s} ^ {k} , \\boldsymbol {V} _ {d i s} ^ {k}\\right)}{\\tau_ {c}}\\right)\\right) \\\\ \\left. - \\frac{1}{K}\\sum_{\\substack{k,j\\\\ j\\neq k}}\\log \\left(1 - \\sigma \\left(\\frac{\\cos(\\boldsymbol{\\mathcal{F}}_{dis}^{k},\\boldsymbol{V}_{dis}^{j})}{\\tau_{c}}\\right)\\right)\\right), \\tag{5} \\\\ \\end{array}\n$$\n", + "text_format": "latex", + "bbox": [ + 99, + 441, + 483, + 537 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "where $\\cos (\\cdot)$ is cosine similarity, $\\sigma$ is the sigmoid function, and $\\tau_{c} = 0.3$ is the temperature factor. $\\mathcal{F}_{dis}^{k}$ is the student feature vector of the $k$ -th local block feature extracted by the DIS token. This approach enhances contrast among features of varying distortion types, improving the model's discrimination and distortion-awareness capabilities.", + "bbox": [ + 89, + 550, + 483, + 642 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "3.3.2. Affinity Knowledge Distillation", + "text_level": 1, + "bbox": [ + 89, + 650, + 352, + 666 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Affinity Knowledge Distillation (AKD) focuses on distilling the interrelations among multiple blocks within an image $\\pmb{I}$ or its degraded version $I_{d}$ . As shown in Fig. 3, given teacher embeddings $\\mathbf{V}_{dis}$ and student features $\\mathcal{F}_{dis}$ , we compute student and teachers' pairwise affinity matrices:", + "bbox": [ + 89, + 670, + 485, + 746 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\boldsymbol {R} _ {s t u} = \\bar {\\boldsymbol {F}} \\bar {\\boldsymbol {F}} ^ {T}, \\quad \\boldsymbol {R} _ {t e a} = \\bar {\\boldsymbol {V}} \\bar {\\boldsymbol {V}} ^ {T}, \\tag {6}\n$$\n", + "text_format": "latex", + "bbox": [ + 171, + 760, + 483, + 780 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "where $\\bar{F}$ and $\\bar{V}$ are $\\mathcal{L}_2$ -normalized student $\\mathcal{F}_{dis}$ and teacher features $V_{dis}$ , respectively. These matrices capture pairwise correlations, and the goal is to align the student affinity matrix $R_{stu}$ with the teacher matrix $R_{tea}$ :", + "bbox": [ + 89, + 787, + 483, + 851 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {L} _ {a k d} = - \\frac {1}{K} \\sum_ {k = 1} ^ {K} \\mathrm {K L} \\left(\\sigma \\left(\\frac {\\boldsymbol {R} _ {s t u} ^ {k}}{\\tau_ {a}}\\right) \\| \\sigma \\left(\\frac {\\boldsymbol {R} _ {t e a} ^ {k}}{\\tau_ {a}}\\right)\\right), \\tag {7}\n$$\n", + "text_format": "latex", + "bbox": [ + 106, + 861, + 483, + 904 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "where $\\tau_{a} = 2$ and $\\mathrm{KL}(\\cdot)$ denotes the Kullback-Leibler divergence. AKD complements CKD by focusing on the relational dynamics between image blocks, allowing the model to learn the relationship between distortion. However, relying solely on AKD may overlook fine-grained distortion-discriminative details, as highlighted in Table 2. Therefore, both CKD and AKD are essential for learning comprehensive distortion-aware features.", + "bbox": [ + 511, + 398, + 908, + 518 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "3.3.3. Overall Training Loss", + "text_level": 1, + "bbox": [ + 511, + 525, + 714, + 541 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "The overall loss combines four complementary components: $\\mathcal{L}_{\\mathrm{cls}}$ provides category-level supervision for distortion identification but lacks awareness of relationships between distortion types; $\\mathcal{L}_{\\mathrm{ckd}}$ enhances feature discrimination by clustering similar distortions and separating dissimilar ones, thus compensating for this limitation; $\\mathcal{L}_{\\mathrm{akd}}$ further captures global perceptual structures by modeling relative similarities across different distortion. and $\\mathcal{L}_{\\log}$ regularizes training via teacher outputs [44], mitigating overfitting to quality-irrelevant patterns. Finally, the total loss is defined as:", + "bbox": [ + 511, + 545, + 908, + 696 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {L} = \\lambda_ {\\mathrm {c l s}} \\mathcal {L} _ {\\mathrm {c l s}} + \\lambda_ {\\mathrm {c k d}} \\mathcal {L} _ {\\mathrm {c k d}} + \\lambda_ {\\mathrm {a k d}} \\mathcal {L} _ {\\mathrm {a k d}} + \\lambda_ {\\log} \\mathcal {L} _ {\\log} + \\| \\hat {Y} - Y _ {g} \\| _ {1}, \\tag {8}\n$$\n", + "text_format": "latex", + "bbox": [ + 513, + 703, + 906, + 734 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "where $Y_{g}$ is the ground truth, $\\mathcal{L}_{\\mathrm{log}}$ is the $\\mathcal{L}_1$ loss between student and teacher predictions, and $\\lambda_{\\mathrm{cls}}$ , $\\lambda_{\\mathrm{ckd}}$ , and $\\lambda_{\\mathrm{akd}}$ are set to 1, with $\\lambda_{\\mathrm{log}}$ set to 0.8.", + "bbox": [ + 511, + 736, + 906, + 781 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "4. Experiments", + "text_level": 1, + "bbox": [ + 511, + 792, + 645, + 810 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "4.1. Benchmark Datasets and Evaluation Protocols", + "text_level": 1, + "bbox": [ + 511, + 816, + 906, + 832 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Experiments were conducted on 7 typical BIQA datasets, including 3 synthetic datasets LIVE [35], CSIQ [15], KA-DID [22], and 4 authentic datasets LIVEC [9], KonIQ [13], LIVEFB [53], SPAQ [7]. Specifically, for authentic datasets,", + "bbox": [ + 511, + 839, + 908, + 901 + ], + "page_idx": 4 + }, + { + "type": "page_number", + "text": "2348", + "bbox": [ + 482, + 944, + 514, + 955 + ], + "page_idx": 4 + }, + { + "type": "table", + "img_path": "images/5a2d1c29bac365b337d7eb051f4e280094908178f72f93e9ca88c4dd418588b1.jpg", + "table_caption": [], + "table_footnote": [], + "table_body": "
Loss LogLoss LclsLoss LakdLoss LckdCSIQKADIDLIVELIVECAvg.
PLCCSRCCPLCCSRCCPLCCSRCCPLCCSRCC
XXXX0.9520.9350.8780.8840.9660.9640.8910.8670.917
✓XXX0.9550.9410.8980.9040.9720.9700.8980.8720.926
✓✓XX0.9590.9470.9190.9160.9740.9720.9040.8900.935
✓X✓X0.9700.9600.9270.9250.9800.9780.9190.9050.945
✓XX✓0.9660.9550.9130.9110.9760.9740.9140.9020.939
✓X✓✓0.9690.9580.9370.9350.9830.9810.9200.9040.948
✓✓✓✓0.9720.9620.9400.9370.9840.9820.9220.9090.951
", + "bbox": [ + 94, + 88, + 903, + 255 + ], + "page_idx": 5 + }, + { + "type": "table", + "img_path": "images/75b70eddedef1602e4a4213f6d47b883f5b9a6180cee906b3e5b5266138431fe.jpg", + "table_caption": [ + "Table 2. Ablation experiments about knowledge distillation components on multiple datasets, where bold entries indicate the best result." + ], + "table_footnote": [], + "table_body": "
TrainingLIVEFBLIVECKonIQLIVECSIQ
TestingKonIQLIVECKonIQLIVECCSIQLIVE
DBCNN0.7160.7240.7540.7550.7580.877
P2P-BM0.7550.7380.7400.7700.712-
TReS0.7130.7400.7330.7860.761-
DEIQT0.7330.7810.7440.7940.7810.932
LoDa0.7630.8050.7450.811--
SHDIQA0.7230.7910.7820.8190.7890.943
", + "bbox": [ + 91, + 289, + 483, + 439 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "LIVEC contains 1162 images from different mobile devices and photographers. SPAQ includes 11125 photos from 66 smartphones. KonIQ contains 10073 images from public sources, while LIVEFB is the largest real-world dataset with 39810 images. The synthetic datasets involve artificially distorting original images using methods such as JPEG compression and Gaussian blur. LIVE and CSIQ respectively have 779 and 866 synthetic distortion images, with 5 and 6 types of distortions. KADID contains 10125 synthetic distortion images, spanning 25 types of distortions.", + "bbox": [ + 88, + 487, + 482, + 637 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Evaluation Criteria. The experiments use Pearson's Rank-Order Correlation Coefficient (SRCC) and Pearson's Linear Correlation Coefficient (PLCC) to assess prediction monotonicity and accuracy, respectively. Both metrics scale from -1 to 1, with values close to 1 indicating better performance.", + "bbox": [ + 88, + 638, + 483, + 715 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "4.2. Implementation Details", + "text_level": 1, + "bbox": [ + 89, + 726, + 307, + 742 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "For the student network, we follow the typical training strategy [29], using a Transformer encoder based on ViT-S from DeiT III [43] and a one-layer decoder. Our model is trained for 9 epochs with a learning rate of $2 \\times 10^{-4}$ , decaying by a factor of 10 every 3 epochs. The batch size is 16 for LIVEC and 128 for KonIQ. We use $80\\%$ of the images for training and $20\\%$ for testing, repeating the process 10 times to reduce performance bias, reporting the average SRCC and PLCC. For synthetic distortion datasets, training and testing sets are divided by reference images to ensure con", + "bbox": [ + 88, + 750, + 482, + 900 + ], + "page_idx": 5 + }, + { + "type": "table", + "img_path": "images/b22281aff0794c372ba2c5635b8e3d5ce8d2c6a9b9bde2a525d1a16575577b77.jpg", + "table_caption": [ + "Table 3. SRCC on the cross datasets validation. The best results are highlighted in bold, second-best are underlined." + ], + "table_footnote": [], + "table_body": "
ModeMethodsLIVECKonIQ
PLCCSRCCPLCCSRCC
20%DEIQT0.8220.7920.9080.888
LoDa0.8540.8150.9230.907
SHDIQA0.8660.8400.9280.915
40%DEIQT0.8550.8380.9220.903
LoDa0.8790.8490.9350.922
SHDIQA0.8890.8690.9420.929
", + "bbox": [ + 516, + 287, + 903, + 439 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Table 4. Data-efficient learning validation with the training set containing $20\\%$ and $40\\%$ of the images.", + "bbox": [ + 511, + 449, + 906, + 477 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "tent independence. To provide reliable distortion perception early in training, the teacher model adopts a similar encoder-decoder architecture aligned with the student model, and is pre-trained with $\\mathcal{L}_1$ loss on the same IQA dataset. Training is performed on four NVIDIA RTX 3090 GPUs.", + "bbox": [ + 511, + 486, + 906, + 561 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "4.3. Overall Prediction Performance Comparison", + "text_level": 1, + "bbox": [ + 511, + 571, + 893, + 589 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Table 1 reports the comparison results between the proposed SHDIQA and other state-of-the-art (SOTA) BIQA methods, including ViT-based IQA methods, i.e., LoDa [50] and DEIQT [29], and Degradation Modelling-based methods, i.e., Re-IQA [31] and QPT [58]. It is observed from these seven datasets that SHDIQA achieves superior performance over all other methods across the 7 datasets. Notably, on the most difficult LIVEFB dataset, SHDIQA achieves an impressive performance improvement. This enhancement can be attributed to the prevalence of local distortions [53] in the LIVEFB dataset, highlighting the importance of effectively utilizing local distortion priors.", + "bbox": [ + 509, + 595, + 908, + 777 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "4.4. Data-Efficient Learning Validation", + "text_level": 1, + "bbox": [ + 511, + 787, + 816, + 803 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "We showcase the data efficiency of our approach, as presented in Table 4. Following the efficient validation protocol outlined in LoDa [50] and DEIQT [29], we train models using $20\\%$ and $40\\%$ of the available images. This process is repeated 10 times, and we report the average SRCC and PLCC scores. By introducing distortion priors through the", + "bbox": [ + 511, + 810, + 908, + 902 + ], + "page_idx": 5 + }, + { + "type": "page_number", + "text": "2349", + "bbox": [ + 482, + 944, + 514, + 955 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/83a2df484a08816d56a9f482fda85550c2b6b0d455ef92fe04d66bbb9f711221.jpg", + "image_caption": [ + "Figure 4. gMAD competition results between DEIQT and our method SHDIQA. (a) Fixed DEIQT at low quality. (b) Fixed DEIQT at high quality. (c) Fixed SHDIQA at low quality. (d) Fixed SHDIQA at high quality. (e) Cosine similarity between different DIS and CLS tokens." + ], + "image_footnote": [], + "bbox": [ + 93, + 88, + 911, + 306 + ], + "page_idx": 6 + }, + { + "type": "table", + "img_path": "images/a710c38bdbd4ffbccf7a7037c119b683d5e9267ca677046127944f9e750dd3b5.jpg", + "table_caption": [], + "table_footnote": [], + "table_body": "
LckdLIVECCSIQAvg.
PLCCSRCCPLCCSRCC
-0.8910.8670.9520.9350.911
L10.9080.8950.9570.9460.926
BCE0.9140.9020.9660.9550.934
", + "bbox": [ + 96, + 349, + 473, + 440 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Block-wise Degradation, our method achieves performance that is nearly with SOTA results, even when trained on only $20\\%$ of the dataset. Furthermore, our approach consistently outperforms all other methods on both datasets, demonstrating its superiority over existing efficient BIQA techniques.", + "bbox": [ + 89, + 469, + 483, + 546 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "4.5. Generalization Capability Validation", + "text_level": 1, + "bbox": [ + 89, + 560, + 410, + 575 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "We evaluate SHDIQA's generalization ability using cross-dataset validation, training the BIQA model on one dataset and testing it on others. Table 3 shows the SRCC averages for five datasets, where SHDIQA achieves the best performance in four out of six cross-dataset evaluations. To further test our framework's generalization, we trained the model on the LIVE database and tested it on the Waterloo database using the gMAD competition [24]. In the gMAD competition, image pairs are selected where the attacker model predicts a large quality difference, while the defender perceives them as similar. Observers then evaluate these pairs to assess the model's generalization ability. As shown in Fig. 4, our model correctly selects image pairs with significant quality differences when attacking DEIQT, where DEIQT misclassifies pairs like (a) as low quality and (b) as high quality, contrary to human perception and SHDIQA's predictions. As a defender, SHDIQA identifies pairs with minimal perceptual variation, correctly classifying image pairs like (c) and (d) as both low and high quality, respectively, while DEIQT fails. This demonstrates the strong generalization capability of SHDIQA in handling real-world image distortions.", + "bbox": [ + 88, + 583, + 485, + 901 + ], + "page_idx": 6 + }, + { + "type": "table", + "img_path": "images/c679ff3b47c42586a8ad9fe585d0a9940790840c79ba3983cb5dc36e8c3ec34d.jpg", + "table_caption": [ + "Table 5. Ablation study on Contrastive Knowledge Distillation." + ], + "table_footnote": [], + "table_body": "
\\( {\\mathcal{L}}_{\\text{akd }}^{I} \\)\\( {\\mathcal{L}}_{\\text{akd }}^{I_d} \\)LIVECLIVEAvg.
PLCCSRCCPLCCSRCC
✓0.9170.9010.9820.9780.944
✓0.9150.9000.9800.9800.943
✓✓0.9220.9090.9840.9820.949
", + "bbox": [ + 517, + 349, + 898, + 441 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Table 6. Ablation experiment on Affinity Knowledge Distillation.", + "bbox": [ + 513, + 445, + 903, + 459 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "4.6. Ablation Study", + "text_level": 1, + "bbox": [ + 511, + 468, + 666, + 484 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "4.6.1. Effectiveness of Proposed Components", + "text_level": 1, + "bbox": [ + 511, + 489, + 828, + 506 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "We conduct an ablation study on components of our framework. Table 2 shows that using the teacher's quality score as a regularization improves model performance. Adding a distortion classification task boosts SRCC by 1.2 and 1.8 points on KADID and LIVEC, respectively, highlighting the effectiveness of distortion-aware supervision. However, identifying distortions without modeling their relationships provides limited benefits. To address this, we introduce CKD and AKD to capture distortion correlations, improving SRCC by up to 3.5 and 4.9 points on KADID, respectively. The limited gain from CKD suggests that aggressively separating regions with different distortions may be problematic, as these regions can have similar quality. AKD mitigates this issue by capturing structural relationships between image blocks. When combined, CKD and AKD improve the baseline by 6.2 points, confirming their complementarity. Finally, Table 8 shows that block-wise aggregation and filtering further enhance the model's ability to assess local distortions, leading to better quality prediction.", + "bbox": [ + 511, + 508, + 908, + 797 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "4.6.2. Loss type of CKD and Implementation of AKD", + "text_level": 1, + "bbox": [ + 511, + 806, + 885, + 821 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Table 5 shows the effect of different CKD loss functions. Using $\\mathcal{L}_1$ loss improves average quality prediction by 1.5 points, and switching to a BCE-based contrastive loss adds another 0.8-point gain. This confirms the effectiveness of our CKD design in promoting feature clustering within the", + "bbox": [ + 511, + 825, + 906, + 900 + ], + "page_idx": 6 + }, + { + "type": "page_number", + "text": "2350", + "bbox": [ + 482, + 944, + 514, + 955 + ], + "page_idx": 6 + }, + { + "type": "table", + "img_path": "images/d00235f51755c22057568c83492e9234aaeb63e920a387f7e0d35af991e7ca5a.jpg", + "table_caption": [], + "table_footnote": [], + "table_body": "
NumberLIVEC
PLCCSRCC
30.9200.903
40.9220.909
50.9180.905
", + "bbox": [ + 94, + 87, + 277, + 181 + ], + "page_idx": 7 + }, + { + "type": "table", + "img_path": "images/cf2e719e3a752e9c3e0363f602bb4e0e6049aa552ecb327064cd418cd37afd27.jpg", + "table_caption": [ + "Table 7. Ablation study on the number of degenerated blocks." + ], + "table_footnote": [], + "table_body": "
MethodLIVECLIVE
PLCCSRCCPLCCSRCC
Baseline0.8910.8670.9660.964
+BWA0.9180.9020.9820.978
+BWF0.9220.9090.9840.982
", + "bbox": [ + 295, + 87, + 581, + 181 + ], + "page_idx": 7 + }, + { + "type": "table", + "img_path": "images/eabb7936609a32d4b4ada431af16c7ba008ebeb23a15fbec7beaed4773ac4e39.jpg", + "table_caption": [ + "Table 8. Ablation study on the Block-wise Aggregation (BwA) and Filtering (BWF) module." + ], + "table_footnote": [], + "table_body": "
MethodLIVECKonIQ
PLCCSRCCPLCCSRCC
Homogenous0.9070.8930.9390.930
Heterogenous0.9220.9090.9480.937
", + "bbox": [ + 596, + 88, + 898, + 181 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Table 9. Ablation experiments on heterogeneous and homogeneous distortion on the whole image.", + "bbox": [ + 594, + 184, + 901, + 213 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/cd911aaddb4ab589b11880f0ab647181721e8e23827f2105791ad7b8b47f309a.jpg", + "image_caption": [ + "Mos: 38.6" + ], + "image_footnote": [], + "bbox": [ + 91, + 219, + 207, + 306 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/3b328d8794657a9a71287374a359e4f8815802dc3decb6d75b3c9e632ccfcbad.jpg", + "image_caption": [ + "Baseline: 46.1" + ], + "image_footnote": [], + "bbox": [ + 207, + 220, + 321, + 306 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/f33fc90b237c0ac63aff54890d79b167d2ecb0c369f1a3993a18d131cce0d8c5.jpg", + "image_caption": [ + "Ours: 35.2(↓4.1)" + ], + "image_footnote": [], + "bbox": [ + 325, + 220, + 437, + 306 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/19b81108dfdef2c055ffbe724ec8a5a9cf7fcb30a6bf8e273748dcd2f818fcbc.jpg", + "image_caption": [ + "DIS token 1" + ], + "image_footnote": [], + "bbox": [ + 442, + 220, + 555, + 306 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/67aad0ec5497634f6afcfcae966a8b4e18f1cc9f555ff233bb9a681fe2e01d58.jpg", + "image_caption": [ + "DIS token 2" + ], + "image_footnote": [], + "bbox": [ + 558, + 220, + 671, + 306 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/34dc853aff7cdea3663b2bf65af3d1ec88113d548f78775f7b4365ded43ca1dc.jpg", + "image_caption": [ + "DIS token 3" + ], + "image_footnote": [], + "bbox": [ + 674, + 220, + 787, + 306 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/e3acd117a948e850ba952908e4ab91173f16c23e6d21bcbff26f0d4f7815cf79.jpg", + "image_caption": [ + "DIS token 4" + ], + "image_footnote": [], + "bbox": [ + 790, + 220, + 903, + 306 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/2bdf871e3b18832158a09fc313ee57f14d7518789f73347fa955e91eb6440c43.jpg", + "image_caption": [ + "Mos: 57.3" + ], + "image_footnote": [], + "bbox": [ + 91, + 325, + 207, + 412 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/30cb86e4aa1ecbf27961745e0619307766e260c710dbef942164040c7ca8fb7f.jpg", + "image_caption": [ + "Baseline: 67.1" + ], + "image_footnote": [], + "bbox": [ + 209, + 325, + 321, + 411 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/dff87b13cfc9008414b9985dd281b6746ebfd3412ad2109e3e0c1a4c87376d3b.jpg", + "image_caption": [ + "Ours: 60.9(↓6.2)" + ], + "image_footnote": [], + "bbox": [ + 326, + 325, + 437, + 411 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/35b5e774e7f7bd8e53024058fd9b264865bbb999d3fb44d8047d5dea73d11e94.jpg", + "image_caption": [ + "DIS token 1", + "Figure 5. Activation maps of DEIQT [29] and SHDIQA using Grad-CAM [32] on the LIVEFB dataset, with MOS shown below the images. Our SHDIQA effectively captures the interplay between image content, local distortions, and overall quality, enabling accurate predictions." + ], + "image_footnote": [], + "bbox": [ + 442, + 325, + 553, + 411 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/58a2b4a8db735cbd97cbebf58fecfd33662852ee34a4c23c330ee46bf666cff4.jpg", + "image_caption": [ + "Dis token 2" + ], + "image_footnote": [], + "bbox": [ + 558, + 325, + 671, + 411 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/e3c74f4522d92d9b969f24f223326929dac7d111f09947406b4aca747367403b.jpg", + "image_caption": [ + "DIS token 3" + ], + "image_footnote": [], + "bbox": [ + 676, + 325, + 787, + 411 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/abf21518b59b470a9d7d64cea86fedf7a5f0c06c197edcead7dcd849b04c50a8.jpg", + "image_caption": [ + "DIS token 4" + ], + "image_footnote": [], + "bbox": [ + 792, + 325, + 903, + 411 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "same distortion type and separation across types. For AKD, we evaluate three strategies: applying it to the input image $\\mathbf{I}$ , the degraded image $\\mathbf{I}_d$ , or both. Table 6 shows that the best performance is achieved by using both, as applying AKD to the image $\\mathbf{I}$ helps model relationships within local content, while applying it to the $\\mathbf{I}_d$ captures correlations between different types of distortions. Combining both enables complementary learning of content and distortion information.", + "bbox": [ + 88, + 469, + 485, + 590 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "4.6.3. Ablation on the Degradation Space", + "text_level": 1, + "bbox": [ + 89, + 595, + 382, + 611 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "We investigate the effectiveness of the Block-wise Degradation Modelling. As shown in Table 9, applying globally uniform heterogeneous distortions significantly underperforms compared to localized distortions, highlighting the importance of local degradation. We also evaluate how varying the number of degraded blocks affects performance. As shown in Table 7, the impact is relatively insensitive across different partitioning strategies, demonstrating the method's resilience to changes in block quantity. While more blocks increase the scope of degradation, they also lead to greater content loss. Thus, using four blocks strikes the best balance.", + "bbox": [ + 89, + 614, + 485, + 781 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "4.7. Qualitative Analysis", + "text_level": 1, + "bbox": [ + 89, + 787, + 282, + 804 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "To show that tokens from different perspectives capture unique features, we analyze the cosine similarity between CLS and DIS tokens on the LIVEC and KonIQ datasets. As shown in Fig. 4, the similarities range from 0.72 to 0.81, lower than the 0.96 and 0.94 found in DeiT-T and DeiT-S [42]. Tokens near the center show higher similarity,", + "bbox": [ + 89, + 810, + 485, + 902 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "aligning with the HVS's focus on salient regions [62]. This implies that DIS tokens capture localized distortions in a way that simulates human perception, and when combined with CLS tokens, it improves both global and local distortion awareness. To further explore the DIS tokens, we visualize the attention maps of DEIQT and our SHDIQA, as shown in Fig. 5. By incorporating four additional distortion tokens, the proposed model enhances its ability to detect distortions across different regions of the image. Through the interaction between distortion tokens and CLS tokens in the adaptation head, the model adaptively focuses on the most relevant areas for image quality, while also considering the overall content. This results in more accurate quality predictions.", + "bbox": [ + 511, + 469, + 906, + 667 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "5. Conclusion", + "text_level": 1, + "bbox": [ + 511, + 681, + 633, + 698 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "In this paper, we propose SHDIQA, a novel BIQA approach that leverages spatial-heterogeneous distortion-aware strategies. Our Block-Wise Degradation Modelling generates locally distorted images, exposing the model to diverse distortion patterns. The Block-wise Aggregation and Filtering mechanism enhances the ViT model's sensitivity to local distortions, improving quality assessments through focused attention and feature filtering. Additionally, our Distortion-Guided Knowledge Distillation combines CKD and AKD to improve distortion discrimination and relational learning. Experimental results show SHDIQA's superior generalization and robustness across various datasets and distortion types, confirming its effectiveness in real-world BIQA tasks.", + "bbox": [ + 511, + 704, + 908, + 902 + ], + "page_idx": 7 + }, + { + "type": "page_number", + "text": "2351", + "bbox": [ + 482, + 944, + 513, + 955 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "References", + "text_level": 1, + "bbox": [ + 91, + 239, + 187, + 255 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[1] Mark R Banham and Aggelos K Katsaggelos. Digital image restoration. IEEE signal processing magazine, 14(2):24-41, 1997. 1", + "[2] Chaofeng Chen, Jiadi Mo, Jingwen Hou, Haoning Wu, Liang Liao, Wenxiu Sun, Qiong Yan, and Weisi Lin. Topiq: A top-down approach from semantics to distortions for image quality assessment. IEEE Transactions on Image Processing, 2024. 3", + "[3] Xinlei Chen, Haoqi Fan, Ross Girshick, and Kaiming He. Improved baselines with momentum contrastive learning. arXiv preprint arXiv:2003.04297, 2020. 2", + "[4] Jia Deng, Wei Dong, Richard Socher, Li-Jia Li, Kai Li, and Li Fei-Fei. Imagenet: A large-scale hierarchical image database. In 2009 IEEE conference on computer vision and pattern recognition, pages 248-255. IEEE, 2009. 4", + "[5] Jinhong Deng, Wen Li, Yuhua Chen, and Lixin Duan. Unbiased mean teacher for cross-domain object detection. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 4091-4101, 2021. 4", + "[6] Chao Dong, Chen Change Loy, Kaiming He, and Xiaou Tang. Image super-resolution using deep convolutional networks. IEEE transactions on pattern analysis and machine intelligence, 38(2):295-307, 2015. 1", + "[7] Yuming Fang, Hanwei Zhu, Yan Zeng, Kede Ma, and Zhou Wang. Perceptual quality assessment of smartphone photography. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 3677-3686, 2020. 5", + "[8] Nuno C Garcia, Pietro Morerio, and Vittorio Murino. Modality distillation with multiple stream networks for action recognition. In Proceedings of the European Conference on Computer Vision (ECCV), pages 103-118, 2018. 3", + "[9] Deepti Ghadiyaram and Alan C Bovik. Massive online crowdsourced study of subjective and objective picture quality. IEEE Transactions on Image Processing, 25(1):372-387, 2015. 5", + "[10] S Alireza Golestaneh, Saba Dadsetan, and Kris M Kitani. No-reference image quality assessment via transformers, relative ranking, and self-consistency. In Proceedings of the IEEE/CVF Winter Conference on Applications of Computer Vision, pages 1220-1230, 2022. 2, 3, 5", + "[11] Saurabh Gupta, Judy Hoffman, and Jitendra Malik. Cross modal distillation for supervision transfer. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 2827-2836, 2016. 3" + ], + "bbox": [ + 93, + 265, + 483, + 900 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "Acknowledgements. This work was supported by the National Science Fund for Distinguished Young Scholars (No.62025603), the National Natural Science Foundation of China (No. U21B2037, No. U22B2051, No. U23A20383, No. 62176222, No. 62176223, No. 62176226, No. 62072386, No. 62072387, No. 62072389, No. 62002305 and No. 62272401), and the Natural Science Foundation of Fujian Province of China (No. 2021J06003, No.2022J06001).", + "[12] Geoffrey Hinton, Oriol Vinyals, and Jeff Dean. Distilling the knowledge in a neural network. arXiv preprint arXiv:1503.02531, 2015. 3", + "[13] Vlad Hosu, Hanhe Lin, Tamas Sziranyi, and Dietmar Saupe. Koniq-10k: An ecologically valid database for deep learning of blind image quality assessment. IEEE Transactions on Image Processing, 29:4041-4056, 2020. 1, 5", + "[14] Junjie Ke, Qifei Wang, Yilin Wang, Peyman Milanfar, and Feng Yang. Musiq: Multi-scale image quality transformer. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 5148-5157, 2021. 2, 3, 5", + "[15] Eric C Larson and Damon M Chandler. Most apparent distortion: full-reference image quality assessment and the role of strategy. Journal of electronic imaging, 19(1):011006-011006, 2010. 2, 5", + "[16] Aobo Li, Jinjian Wu, Yongxu Liu, and Leida Li. Bridging the synthetic-to Authentic gap: Distortion-guided unsupervised domain adaptation for blind image quality assessment. arXiv preprint arXiv:2405.04167, 2024. 2", + "[17] Leida Li, Tianshu Song, Jinjian Wu, Weisheng Dong, Jiansheng Qian, and Guangming Shi. Blind image quality index for authentic distortions with local and global deep feature aggregation. IEEE Transactions on Circuits and Systems for Video Technology, 32(12):8512-8523, 2021. 2, 3", + "[18] Liangqi Li, Jiaxu Miao, Dahu Shi, Wenming Tan, Ye Ren, Yi Yang, and Shiliang Pu. Distilling detr with visual-linguistic knowledge for open-vocabulary object detection. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 6501–6510, 2023. 4", + "[19] Xudong Li, Timin Gao, Runze Hu, Yan Zhang, Shengchuan Zhang, Xiawu Zheng, Jingyuan Zheng, Yunhang Shen, Ke Li, Yutao Liu, et al. Adaptive feature selection for no-reference image quality assessment by mitigating semantic noise sensitivity. In International Conference on Machine Learning, pages 27808-27821. PMLR, 2024. 4, 5", + "[20] Xudong Li, Runze Hu, Jingyuan Zheng, Yan Zhang, Shengchuan Zhang, Xiawu Zheng, Ke Li, Yunhang Shen, Yutao Liu, Pingyang Dai, et al. Integrating global context contrast and local sensitivity for blind image quality assessment. In International Conference on Machine Learning, pages 27920-27941. PMLR, 2024. 3", + "[21] Xudong Li, Jingyuan Zheng, Runze Hu, Yan Zhang, Ke Li, Yunhang Shen, Xiawu Zheng, Yutao Liu, ShengChuan Zhang, Pingyang Dai, et al. Feature denoising diffusion model for blind image quality assessment. arXiv preprint arXiv:2401.11949, 2024.3", + "[22] Hanhe Lin, Vlad Hosu, and Dietmar Saupe. Kadid-10k: A large-scale artificially distorted iqa database. In 2019 Eleventh International Conference on Quality of Multimedia Experience (QoMEX), pages 1-3. IEEE, 2019. 5", + "[23] Hao Liu and Pieter Abbeel. Blockwise parallel transformers for large context models. Advances in Neural Information Processing Systems, 36, 2024. 3", + "[24] Kede Ma, Qingbo Wu, Zhou Wang, Zhengfang Duanmu, Hongwei Yong, Hongliang Li, and Lei Zhang. Group mad competition-a new methodology to compare objective image quality models. In Proceedings of the IEEE Conference on" + ], + "bbox": [ + 93, + 90, + 906, + 900 + ], + "page_idx": 8 + }, + { + "type": "page_number", + "text": "2352", + "bbox": [ + 482, + 944, + 514, + 955 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "Computer Vision and Pattern Recognition, pages 1664-1673, 2016. 7", + "[25] Kede Ma, Wentao Liu, Tongliang Liu, Zhou Wang, and Dacheng Tao. dipiq: Blind image quality assessment by learning-to-rank discriminable image pairs. IEEE Transactions on Image Processing, 26(8):3951-3964, 2017. 2", + "[26] Kede Ma, Wentao Liu, Kai Zhang, Zhengfang Duanmu, Zhou Wang, and Wangmeng Zuo. End-to-end blind image quality assessment using deep neural networks. IEEE Transactions on Image Processing, 27(3):1202-1213, 2017. 2", + "[27] Anish Mittal, Anush Krishna Moorthy, and Alan Conrad Bovik. No-reference image quality assessment in the spatial domain. IEEE Transactions on image processing, 21(12): 4695-4708, 2012. 5", + "[28] Angelo Porrello, Luca Bergamini, and Simone Calderara. Robust re-identification by multiple views knowledge distillation. In Computer Vision-ECCV 2020: 16th European Conference, Glasgow, UK, August 23–28, 2020, Proceedings, Part X 16, pages 93–110. Springer, 2020. 3", + "[29] Guanyi Qin, Runze Hu, Yutao Liu, Xiawu Zheng, Haotian Liu, Xiu Li, and Yan Zhang. Data-efficient image quality assessment with attention-panel decoder. In Proceedings of the Thirty-Seventh AAAI Conference on Artificial Intelligence, 2023. 3, 4, 5, 6, 8", + "[30] Jiezhong Qiu, Hao Ma, Omer Levy, Scott Wen-tau Yih, Sinong Wang, and Jie Tang. Blockwise self-attention for long document understanding. arXiv preprint arXiv:1911.02972, 2019. 3", + "[31] Avinab Saha, Sandeep Mishra, and Alan C Bovik. Re-iqa: Unsupervised learning for image quality assessment in the wild. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 5846-5855, 2023. 2, 4, 5, 6", + "[32] Ramprasaath R Selvaraju, Michael Cogswell, Abhishek Das, Ramakrishna Vedantam, Devi Parikh, and Dhruv Batra. Gradcam: Visual explanations from deep networks via gradient-based localization. In Proceedings of the IEEE international conference on computer vision, pages 618-626, 2017. 8", + "[33] Tahira Shehzadi, Khurram Azeem Hashmi, Didier Stricker, and Muhammad Zeshan Afzal. Sparse semi-detr: Sparse learnable queries for semi-supervised object detection. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 5840-5850, 2024. 4", + "[34] Hamid R Sheikh, Alan C Bovik, and Lawrence Cormack. No-reference quality assessment using natural scene statistics: Jpeg2000. IEEE Transactions on image processing, 14(11): 1918-1927, 2005. 2", + "[35] Hamid R Sheikh, Muhammad F Sabir, and Alan C Bovik. A statistical evaluation of recent full reference image quality assessment algorithms. IEEE Transactions on image processing, 15(11):3440-3451, 2006. 5", + "[36] Chenyang Shi and Yandan Lin. Full reference image quality assessment based on visual salience with color appearance and gradient similarity. IEEE Access, 8:97310-97320, 2020. 1", + "[37] Tianshu Song, Leida Li, Deqiang Cheng, Pengfei Chen, and Jinjian Wu. Active learning-based sample selection for label-" + ], + "bbox": [ + 91, + 92, + 483, + 900 + ], + "page_idx": 9 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "efficient blind image quality assessment. IEEE Transactions on Circuits and Systems for Video Technology, 2023. 1", + "[38] Suhas Srinath, Shankhanil Mitra, Shika Rao, and Rajiv Soundararajan. Learning generalizable perceptual representations for data-efficient no-reference image quality assessment. In Proceedings of the IEEE/CVF Winter Conference on Applications of Computer Vision, pages 22-31, 2024. 1", + "[39] Shaolin Su, Qingsen Yan, Yu Zhu, Cheng Zhang, Xin Ge, Jinqiu Sun, and Yanning Zhang. Blindly assess image quality in the wild guided by a self-adaptive hyper network. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 3667-3676, 2020. 5", + "[40] Simeng Sun, Tao Yu, Jiahua Xu, Wei Zhou, and Zhibo Chen. Graphiqa: Learning distortion graph representations for blind image quality assessment. IEEE Transactions on Multimedia, 25:2912-2925, 2022. 2", + "[41] Dacheng Tao, Xuelong Li, Wen Lu, and Xinbo Gao. Reduced-reference iqa in contourlet domain. IEEE Transactions on Systems, Man, and Cybernetics, Part B (Cybernetics), 39(6): 1623-1627, 2009. 1", + "[42] Hugo Touvron, Matthieu Cord, Matthijs Douze, Francisco Massa, Alexandre Sablayrolles, and Hervé Jégou. Training data-efficient image transformers & distillation through attention. In International Conference on Machine Learning, pages 10347-10357. PMLR, 2021. 8", + "[43] Hugo Touvron, Matthieu Cord, and Hervé Jégou. Deit iii: Revenge of the vit. arXiv preprint arXiv:2204.07118, 2022. 6", + "[44] Tao Wang, Li Yuan, Xiaopeng Zhang, and Jiashi Feng. Distilling object detectors with fine-grained feature imitation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 4933-4942, 2019. 5", + "[45] Zhou Wang and Eero Simoncelli. Local phase coherence and the perception of blur. Advances in neural information processing systems, 16, 2003. 2", + "[46] Zhou Wang, Hamid R Sheikh, and Alan C Bovik. No-reference perceptual quality assessment of JPEG compressed images. In Proceedings. International conference on image processing, pages I-I. IEEE, 2002. 2", + "[47] Zhou Wang, Alan C Bovik, Hamid R Sheikh, and Eero P Simoncelli. Image quality assessment: from error visibility to structural similarity. IEEE transactions on image processing, 13(4):600-612, 2004. 1", + "[48] Zhihua Wang, Zhi-Ri Tang, Jianguo Zhang, and Yuming Fang. Toward a blind image quality evaluator in the wild by learning beyond human opinion scores. Pattern Recognition, 137:109296, 2023. 1", + "[49] Haoning Wu, Zicheng Zhang, Erli Zhang, Chaofeng Chen, Liang Liao, Annan Wang, Chunyi Li, Wenxiu Sun, Qiong Yan, Guangtao Zhai, et al. Q-bench: A benchmark for general-purpose foundation models on low-level vision. arXiv preprint arXiv:2309.14181, 2023. 1", + "[50] Kangmin Xu, Liang Liao, Jing Xiao, Chaofeng Chen, Haoning Wu, Qiong Yan, and Weisi Lin. Local distortion aware efficient transformer adaptation for image quality assessment. arXiv preprint arXiv:2308.12001, 2023. 2, 3, 5, 6", + "[51] Sidi Yang, Tianhe Wu, Shuwei Shi, Shanshan Lao, Yuan Gong, Mingdeng Cao, Jiahao Wang, and Yujiu Yang. Maniqa:" + ], + "bbox": [ + 516, + 92, + 906, + 900 + ], + "page_idx": 9 + }, + { + "type": "page_number", + "text": "2353", + "bbox": [ + 482, + 944, + 514, + 955 + ], + "page_idx": 9 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "Multi-dimension attention network for no-reference image quality assessment. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 1191-1200, 2022. 1", + "[52] Guanghao Yin, Wei Wang, Zehuan Yuan, Chuchu Han, Wei Ji, Shouqian Sun, and Changhu Wang. Content-variant reference image quality assessment via knowledge distillation. arXiv preprint arXiv:2202.13123, 2022. 3", + "[53] Zhenqiang Ying, Haoran Niu, Praful Gupta, Dhruv Mahajan, Deepti Ghadiyaram, and Alan Bovik. From patches to pictures (paq-2-piq): Mapping the perceptual space of picture quality. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 3575–3585, 2020. 5, 6", + "[54] Junyong You and Jari Korhonen. Transformer for image quality assessment. In 2021 IEEE International Conference on Image Processing (ICIP), pages 1389-1393. IEEE, 2021. 3, 5", + "[55] Guanghui Yue, Di Cheng, Leida Li, Tianwei Zhou, Hantao Liu, and Tianfu Wang. Semi-supervised authentically distorted image quality assessment with consistency-preserving dual-branch convolutional neural network. IEEE Transactions on Multimedia, 25:6499–6511, 2022. 3", + "[56] Weixia Zhang, Kede Ma, Jia Yan, Dexiang Deng, and Zhou Wang. Blind image quality assessment using a deep bilinear convolutional neural network. IEEE Transactions on Circuits and Systems for Video Technology, 30(1):36-47, 2018. 2, 5", + "[57] Weixia Zhang, Guangtao Zhai, Ying Wei, Xiaokang Yang, and Kede Ma. Blind image quality assessment via vision-language correspondence: A multitask learning perspective. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 14071-14081, 2023. 2, 5", + "[58] Kai Zhao, Kun Yuan, Ming Sun, Mading Li, and Xing Wen. Quality-aware pre-trained models for blind image quality assessment. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 22302-22313, 2023. 2, 4, 5, 6", + "[59] Heliang Zheng, Huan Yang, Jianlong Fu, Zheng-Jun Zha, and Jiebo Luo. Learning conditional knowledge distillation for degraded-reference image quality assessment. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 10242-10251, 2021. 3", + "[60] Yu Zhou, Leida Li, Shiqi Wang, Jinjian Wu, Yuming Fang, and Xinbo Gao. No-reference quality assessment for view synthesis using dog-based edge statistics and texture naturalness. IEEE Transactions on Image Processing, 28(9):4566-4579, 2019. 1", + "[61] Hancheng Zhu, Leida Li, Jinjian Wu, Weisheng Dong, and Guangming Shi. Metaiaq: Deep meta-learning for no-reference image quality assessment. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 14143-14152, 2020. 2, 5", + "[62] Mengmeng Zhu, Guanqun Hou, Xinjia Chen, Jiaxing Xie, Haixian Lu, and Jun Che. Saliency-guided transformer network combined with local embedding for no-reference image quality assessment. In Proceedings of the IEEE/CVF Inter-" + ], + "bbox": [ + 91, + 90, + 483, + 900 + ], + "page_idx": 10 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "national Conference on Computer Vision, pages 1953-1962, 2021.8", + "[63] Fuzhen Zhuang, Zhiyuan Qi, Keyu Duan, Dongbo Xi, Yongchun Zhu, Hengshu Zhu, Hui Xiong, and Qing He. A comprehensive survey on transfer learning. Proceedings of the IEEE, 109(1):43-76, 2020. 3" + ], + "bbox": [ + 516, + 90, + 908, + 176 + ], + "page_idx": 10 + }, + { + "type": "page_number", + "text": "2354", + "bbox": [ + 482, + 944, + 514, + 955 + ], + "page_idx": 10 + } +] \ No newline at end of file diff --git a/2025/Distilling Spatially-Heterogeneous Distortion Perception for Blind Image Quality Assessment/c7a78868-95dc-4d35-97c7-04bbcfc1063d_model.json b/2025/Distilling Spatially-Heterogeneous Distortion Perception for Blind Image Quality Assessment/c7a78868-95dc-4d35-97c7-04bbcfc1063d_model.json new file mode 100644 index 0000000000000000000000000000000000000000..6e5b0dfd69ddeb0b777b911737b906bcd8010e6a --- /dev/null +++ b/2025/Distilling Spatially-Heterogeneous Distortion Perception for Blind Image Quality Assessment/c7a78868-95dc-4d35-97c7-04bbcfc1063d_model.json @@ -0,0 +1,2796 @@ +[ + [ + { + "type": "header", + "bbox": [ + 0.107, + 0.003, + 0.182, + 0.043 + ], + "angle": 0, + "content": "CVF" + }, + { + "type": "header", + "bbox": [ + 0.237, + 0.0, + 0.812, + 0.045 + ], + "angle": 0, + "content": "This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore." + }, + { + "type": "title", + "bbox": [ + 0.095, + 0.13, + 0.905, + 0.175 + ], + "angle": 0, + "content": "Distilling Spatially-Heterogeneous Distortion Perception for Blind Image Quality Assessment" + }, + { + "type": "text", + "bbox": [ + 0.103, + 0.203, + 0.9, + 0.291 + ], + "angle": 0, + "content": "Xudong Li\\(^{1}\\) Wenjie Nie\\(^{1}\\) Yan Zhang\\(^{1}\\) Runze Hu\\(^{2}\\) Ke Li\\(^{3}\\) Xiawu Zheng\\(^{1}\\) Liujuan Cao\\(^{1*}\\) \n\\(^{1}\\) Key Laboratory of Multimedia Trusted Perception and Efficient Computing, Ministry of Education of China, Xiamen University, 361005, P.R. China \n\\(^{2}\\) School of Information and Electronics, Beijing Institute of Technology, Beijing, China \n\\(^{3}\\) Tencent Youtu Lab, Shanghai, China" + }, + { + "type": "text", + "bbox": [ + 0.132, + 0.295, + 0.856, + 0.309 + ], + "angle": 0, + "content": "{1xd761050753, bzhy986, hrzlpk2015}@gmail.com, {zhengxiwu, caoliujuan}@xmu.edu.cn" + }, + { + "type": "title", + "bbox": [ + 0.248, + 0.345, + 0.327, + 0.361 + ], + "angle": 0, + "content": "Abstract" + }, + { + "type": "text", + "bbox": [ + 0.088, + 0.377, + 0.486, + 0.756 + ], + "angle": 0, + "content": "In the Blind Image Quality Assessment (BIQA) field, accurately assessing the quality of authentically distorted images presents a substantial challenge due to the diverse distortion types in natural settings. Existing state-of-the-art IQA methods mix a sequence of distortions into entire images to establish global distortion priors, but are inadequate for authentic images with spatially varied distortions. To address this, we introduce a novel IQA framework that employs knowledge distillation tailored to perceive spatially heterogeneous distortions, enhancing quality-distortion awareness. Specifically, we introduce a novel Block-wise Degradation Modelling approach that applies distinct distortions to different spatial blocks of an image, thereby expanding local distortion priors. Following this, we present a Block-wise Aggregation and Filtering module that enables fine-grained attention to the quality information within different distortion areas of the image. Furthermore, to effectively capture the complex relationships between distortions across different regions while preserving overall quality perception, we introduce Contrastive Knowledge Distillation to enhance the model's ability to discriminate between different types of distortions and Affinity Knowledge Distillation to model the correlation among distortions in different regions. Extensive experiments on standard BIQA datasets demonstrate the effectiveness and competitiveness of the proposed method." + }, + { + "type": "title", + "bbox": [ + 0.092, + 0.787, + 0.222, + 0.803 + ], + "angle": 0, + "content": "1. Introduction" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.813, + 0.484, + 0.875 + ], + "angle": 0, + "content": "Image Quality Assessment (IQA) methods are designed to evaluate images in a way that aligns with the human judgment of visual distortions [47, 49]. These methods are crucial for applications that rely heavily on image quality" + }, + { + "type": "image", + "bbox": [ + 0.526, + 0.343, + 0.895, + 0.549 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.513, + 0.558, + 0.908, + 0.601 + ], + "angle": 0, + "content": "Figure 1. Comparison of state-of-the-art (SOTA) BIQA methods on the largest authentic dataset, LIEVFB. The proposed SHDIQA method outperforms all existing BIQA approaches." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.614, + 0.908, + 0.734 + ], + "angle": 0, + "content": "and serve as standards to gauge the performance of image processing techniques. Objective Image Quality Assessment includes Full-Reference IQA (FR-IQA) [36], Reduced-Reference IQA (RR-IQA) [41], and no-reference IQA (NR-IQA) or Blind IQA (BIQA) [60]. As the presence of various types of artifacts in real-world scenes and the unavailability of reference images, BIQA gains attention for tasks like image restoration [1] and super-resolution [6]." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.735, + 0.909, + 0.901 + ], + "angle": 0, + "content": "Recent advances in data-driven BIQA models, especially those based on deep neural networks, have led to impressive progress [37, 51]. However, these models still rely heavily on Mean Opinion Scores (MOS), which are expensive and time-consuming to collect—often requiring up to 120 human ratings per image [13]. As an alternative, synthetic distortions generated through controlled data augmentation offer a cost-effective way to create annotated datasets [48], enabling the learning of quality representations without extensive human labeling [38]. To leverage such synthetic data, recent studies have introduced multi-task end-to-end" + }, + { + "type": "page_footnote", + "bbox": [ + 0.11, + 0.888, + 0.236, + 0.9 + ], + "angle": 0, + "content": "*Corresponding author." + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.516, + 0.957 + ], + "angle": 0, + "content": "2344" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.09, + 0.092, + 0.485, + 0.228 + ], + "angle": 0, + "content": "frameworks [26, 57] that jointly optimize distortion-specific objectives and quality prediction, using synthetic distortions to enrich training signals and improve generalization. Another line of work focuses on image degradation modeling. Methods like Re-IQA [31] and QPT [58] simulate realistic distortions by applying diverse types and combinations to entire images, as illustrated in Fig.2. They then leverage unsupervised learning frameworks such as MoCo-v2[3] to extract distortion priors from large-scale synthetic datasets." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.23, + 0.486, + 0.382 + ], + "angle": 0, + "content": "However, both categories of methods share a key limitation: they focus on global distortion priors while ignoring fine-grained local distortions, which are essential for accurate BIQA. Research [15, 17] shows that the human visual system (HVS) employs both local detection and global appearance strategies when assessing image quality. For subtle, near-threshold distortions, the HVS prioritizes local details, while for highly visible distortions, it shifts focus back to global content. This highlights the vital role of local distortion features in comprehensive image quality evaluation." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.385, + 0.485, + 0.521 + ], + "angle": 0, + "content": "In this paper, we propose a novel Spatial Heterogeneous Distortion-aware Image Quality Assessment (SHDIQA) method, designed to improve the local distortion-awareness of BIQA models based on the Vision Transformer (ViT) architecture. Unlike prior works [10, 14, 50] that focus on modifying ViT architectures to enhance local feature extraction, our method fully leverages rich local distortion priors generated from synthetic distortions, while introducing minimal architectural changes (only 0.2M additional parameters)." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.524, + 0.487, + 0.902 + ], + "angle": 0, + "content": "Specifically, we introduce a Block-wise Image Degradation (BWID) method tailored for IQA tasks to generate patch-level distortion priors. As shown in Fig. 2, BWID partitions a distorted image into multiple blocks and applies a randomly selected distortion from 24 categories across 4 severity levels to each block. This results in a composite image with spatially varying local distortions. Each image naturally comes with an image-level multi-class label and corresponding patch-level single-class labels, effectively addressing the lack of localized distortion annotations in existing methods. To further utilize the patch-level distortion priors, we propose a Block-wise Aggregation and Filtering (BwAF) mechanism. BWAF uses the local priors from BWID to train multiple distortion (DIS) tokens, which interact with a global classification (CLS) token through a cross-attention mechanism with an adaptive head. This interaction mimics the HVS's dual strategy of local detection and global perception, thereby improving overall quality assessment, as illustrated in Fig. 1. Moreover, previous studies [16, 40] highlight the importance of modeling the relationships among different distortion types in perceptual quality assessment. To achieve this, we incorporate Contrastive Knowledge Distillation (CKD) and Affinity Knowledge Distillation (AKD). CKD aligns similar distortions and separates dissimilar ones, improving the model's ability to differentiate between distor" + }, + { + "type": "image", + "bbox": [ + 0.529, + 0.089, + 0.6, + 0.153 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.548, + 0.155, + 0.582, + 0.164 + ], + "angle": 0, + "content": "Image B" + }, + { + "type": "image", + "bbox": [ + 0.605, + 0.089, + 0.674, + 0.153 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.607, + 0.154, + 0.674, + 0.163 + ], + "angle": 0, + "content": "Distorted Region" + }, + { + "type": "image", + "bbox": [ + 0.68, + 0.089, + 0.751, + 0.153 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.684, + 0.155, + 0.748, + 0.164 + ], + "angle": 0, + "content": "DEIQT Method" + }, + { + "type": "image", + "bbox": [ + 0.757, + 0.089, + 0.824, + 0.153 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.76, + 0.155, + 0.824, + 0.163 + ], + "angle": 0, + "content": "SHDIOA(Ours)" + }, + { + "type": "image", + "bbox": [ + 0.827, + 0.102, + 0.893, + 0.153 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.528, + 0.164, + 0.599, + 0.219 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.607, + 0.164, + 0.675, + 0.218 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.679, + 0.164, + 0.75, + 0.219 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.755, + 0.164, + 0.824, + 0.219 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.827, + 0.163, + 0.894, + 0.219 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.513, + 0.231, + 0.907, + 0.273 + ], + "angle": 0, + "content": "Figure 2. The proposed SHDIQA method creates a heterogeneous degradation space with richer priors (Top) and assigns dedicated distortion-aware (DIS) tokens to each local region (Bottom)." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.294, + 0.907, + 0.34 + ], + "angle": 0, + "content": "tion types. Meanwhile, AKD transfers knowledge about the relationships between different distortions, enhancing the model's generalization across various types of distortions." + }, + { + "type": "text", + "bbox": [ + 0.533, + 0.34, + 0.836, + 0.354 + ], + "angle": 0, + "content": "Our contributions are summarized as follows:" + }, + { + "type": "text", + "bbox": [ + 0.514, + 0.356, + 0.907, + 0.431 + ], + "angle": 0, + "content": "- We present a Block-wise Degradation Modelling approach that enriches the local distortion priors of the training datasets by dividing images into blocks and applying distinct distortions to each, enhancing the model's ability to assess image quality accurately in complex scenarios." + }, + { + "type": "text", + "bbox": [ + 0.514, + 0.432, + 0.907, + 0.521 + ], + "angle": 0, + "content": "- We introduce the Block-wise Aggregation and Filtering mechanism to fully capture and leverage localized distortions. It first directs precise attention to distortion details within image blocks, and then utilizes an adaptation head to filter out irrelevant distortion features, focusing the analysis on quality-aware information." + }, + { + "type": "text", + "bbox": [ + 0.514, + 0.522, + 0.907, + 0.597 + ], + "angle": 0, + "content": "- We propose a novel Distortion-Guided Knowledge Distillation strategy that combines Contrastive Knowledge Distillation (CKD) to enhance distortion discrimination and Affinity Knowledge Distillation (AKD) to enrich relational insights between local distortion areas." + }, + { + "type": "list", + "bbox": [ + 0.514, + 0.356, + 0.907, + 0.597 + ], + "angle": 0, + "content": null + }, + { + "type": "title", + "bbox": [ + 0.514, + 0.612, + 0.655, + 0.627 + ], + "angle": 0, + "content": "2. Related Work" + }, + { + "type": "title", + "bbox": [ + 0.514, + 0.637, + 0.818, + 0.654 + ], + "angle": 0, + "content": "2.1. BIQA with Degradation Modelling" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.66, + 0.909, + 0.901 + ], + "angle": 0, + "content": "Recent studies [10, 61] have shown that simple distortion simulations do not fully capture the complexities of real-world degradations. To address this gap, Image Degradation Modeling has been introduced to represent various distortion types [34, 45, 46]. DBCNN [56] decomposes real distortions into multiple synthetic ones, creating large-scale synthetic images for training. DipIQ [25] categorizes common distortions like noise, blur, and compression into severity levels, while Re-IQA [31] introduces 25 distortion types to expand the degradation space. QPT [58] models second-order degradation using skip and shuffle operations across nine types in three groups. In contrast to these global degradation approaches, our method applies localized distortions to different regions of the image, enabling the model to learn fine-grained distortion perception both within and across blocks through knowledge distillation." + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.516, + 0.957 + ], + "angle": 0, + "content": "2345" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.095, + 0.089, + 0.307, + 0.212 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.11, + 0.215, + 0.293, + 0.228 + ], + "angle": 0, + "content": "(a) Block-wise Degradation" + }, + { + "type": "image", + "bbox": [ + 0.096, + 0.228, + 0.308, + 0.326 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.328, + 0.09, + 0.907, + 0.328 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.09, + 0.34, + 0.907, + 0.437 + ], + "angle": 0, + "content": "Figure 3. Framework Overview. Given an input image \\( \\mathbf{I} \\), (a) Block-wise Degradation is first applied to generate a degraded version \\( \\mathbf{I}_d \\) (Sec. 3.1). Both \\( \\mathbf{I} \\) and \\( \\mathbf{I}_d \\) are then fed into a student encoder-decoder network, which is initialized with multiple distortion (DIS) tokens (each attending to a specific local block) and a CLS token for global quality representation. (b) Guided by the Block-wise Aggregation (Sec. 3.2), the DIS tokens and the CLS token are used to extract local distortion features \\( \\mathcal{F}_{dis} \\) and a global quality feature \\( \\mathcal{F}_{cls} \\), which are further refined by the decoder into \\( \\hat{\\mathcal{F}}_{dis} \\) and \\( \\hat{\\mathcal{F}}_{cls} \\). These refined features are processed by a quality-aware adaptation head to suppress irrelevant information and predict the quality score via an MLP. (c) To further enhance distortion perception, Contrastive KD is used to improve inter-block distortion discrimination, while Affinity KD captures relationships between distortions across different regions (Sec. 3.3)." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.451, + 0.321, + 0.466 + ], + "angle": 0, + "content": "2.2. BIQA with Transformers" + }, + { + "type": "text", + "bbox": [ + 0.089, + 0.473, + 0.486, + 0.76 + ], + "angle": 0, + "content": "Vision Transformer (ViT)-based IQA architectures include hybrid transformers [2, 10, 21, 50, 54] and pure transformers [14, 20, 29]. Hybrid models integrate multi-scale CNN information for better local distortion perception but increase complexity. Pure transformers like DEIQT [29] capture global abstraction using the CLS token but struggle with local structure modeling [63]. To reduce memory overhead, methods using local self-attention and sparse block structures have been developed [23, 30], though at the cost of some long-distance modeling. Sampling-based methods like VT-IQA [17] combine global and local information by resampling distorted images and using global attention. However, they lack explicit supervision for local distortions, which limits their ability to capture localized degradations during training. In contrast, our approach enhances local distortion perception in pure ViT by incorporating patch-level distortion priors, while preserving global context modeling through a mask matrix specifically applied to distortion tokens, leaving the self-attention process unaffected." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.772, + 0.379, + 0.789 + ], + "angle": 0, + "content": "2.3. IQA with Knowledge Distillation" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.795, + 0.485, + 0.901 + ], + "angle": 0, + "content": "Knowledge Distillation (KD), originally proposed for model compression [12], has been adapted for tasks like action recognition [8], re-identification [28], and depth estimation [11]. Recently, KD has been used in methods like Degraded-Reference IQA (DR-IQA) [59] and Non-aligned Reference IQA (NAR-IQA) [52] to transfer reference information, though these still partially rely on reference images." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.451, + 0.907, + 0.543 + ], + "angle": 0, + "content": "A related BIQA method is SSLIQA [55], which employs KD to enforce semantic consistency across different triples within a batch. However, our approach differs significantly in that our KD explicitly targets inter-block and intra-block distortion information, allowing for more direct extraction of quality-related distortion priors from the augmented image." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.558, + 0.649, + 0.575 + ], + "angle": 0, + "content": "3. Methodology" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.584, + 0.909, + 0.901 + ], + "angle": 0, + "content": "We introduce Spatial-Heterogeneous Distortion-aware Image Quality Assessment (SHDIQA), a framework designed to enhance the perception of local and non-uniform distortions in authentic images, as illustrated in Fig. 3. Specifically, the proposed SHDIQA first divides the input image \\(I\\) into vertical blocks and applies random distortions using a Block-wise Degradation module (Sec. 3.1), producing a degraded image \\(I_d\\). Both \\(I\\) and \\(I_d\\) are then processed by a student encoder-decoder network equipped with a Blockwise Aggregation module (Sec. 3.2), which extracts local distortion features \\(\\mathcal{F}_{dis}\\) via multiple distortion (DIS) tokens and global quality feature \\(\\mathcal{F}_{cls}\\) via classification (CLS) token. Next, the local and global features \\(\\mathcal{F}_{dis}\\) and \\(\\mathcal{F}_{cls}\\) are fused via a cross-attention mechanism in the adaptation head to suppress irrelevant distortion information (Sec. 3.2). To enhance the model's ability to represent and distinguish distortion patterns, we further introduce Contrastive Knowledge Distillation (CKD) and Affinity Knowledge Distillation (AKD) (Sec. 3.3). These modules jointly improve distortion discriminability and capture inter-block distortion correlations, enabling a more distortion-aware representation." + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.516, + 0.956 + ], + "angle": 0, + "content": "2346" + } + ], + [ + { + "type": "title", + "bbox": [ + 0.091, + 0.091, + 0.395, + 0.108 + ], + "angle": 0, + "content": "3.1. Block-wise Degradation Modelling" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.113, + 0.486, + 0.355 + ], + "angle": 0, + "content": "Recent methods [31, 58] apply multiple distortions uniformly across the entire image, ignoring the finer details of local distortions. To address this, we propose a Block-wise Degradation Modelling strategy, which applies 24 distortion types and 4 degradation levels to different regions of the image, thereby generating a composite local distortion prior. Specifically, given an image of size \\( W \\times H \\), we divide it into \\( K \\) vertical blocks of size \\( W \\times \\frac{H}{K} \\). For each block \\( k \\) (\\( k = 1, \\dots, K \\)), we randomly select a rectangular region \\( C_k \\) with top-left coordinate \\( \\alpha_i \\) and size \\( w_k \\times h_k \\), where \\( w_k = \\beta_w \\times W \\) and \\( h_k = \\beta_h \\times \\frac{H}{K} \\). Empirically, we find that setting \\( \\beta_w = \\beta_h = 1 \\), i.e., degrading the entire block, yields the best performance. Then, the degradation is applied using the function \\( \\mathcal{D}_k(C_k; q_k, p_k) \\), where \\( q_k \\) denotes the type of distortion (e.g., noise, darkening), and \\( p_k \\) represents its intensity. The degraded image \\( I_d \\) is generated as follows:" + }, + { + "type": "equation", + "bbox": [ + 0.109, + 0.364, + 0.484, + 0.382 + ], + "angle": 0, + "content": "\\[\n\\boldsymbol {I} _ {d} = \\mathcal {A} (\\boldsymbol {I}) = \\mathcal {D} _ {K} \\left(\\dots \\mathcal {D} _ {1} \\left(\\boldsymbol {C} _ {1}; q _ {1}, p _ {1}\\right) \\dots ; q _ {K}, p _ {K}\\right) \\tag {1}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.389, + 0.485, + 0.466 + ], + "angle": 0, + "content": "where \\( K = 4 \\) denotes the number of degraded blocks. Unlike previous methods, our approach focuses on localized distortions, providing richer priors for feature aggregation (Sec. 3.2) and knowledge distillation (Sec. 3.3), improving the model's ability to handle complex real-world distortions." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.473, + 0.415, + 0.489 + ], + "angle": 0, + "content": "3.2. Block-wise Aggregation and Filtering" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.495, + 0.486, + 0.571 + ], + "angle": 0, + "content": "In this section, we introduce the Block-wise Aggregation method to extract the local distortion feature \\(\\mathcal{F}_{dis}\\) from the degraded image \\(I_{d}\\) and the input image \\(\\pmb{I}\\). We then present the Quality-aware Adaptation Head, which filters out quality-irrelevant distortion features from distortion feature \\(\\mathcal{F}_{dis}\\)." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.577, + 0.303, + 0.592 + ], + "angle": 0, + "content": "3.2.1. Block-wise Aggregation" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.596, + 0.486, + 0.776 + ], + "angle": 0, + "content": "We begin by randomly initializing \\(K\\) local distortion-aware (DIS) tokens, denoted as \\(\\mathcal{F}_{dis}\\), matching the number of degraded regions. Each DIS token \\(\\mathcal{F}_{dis}^{k}\\) corresponds to a specific region \\(C_k\\). The resulting embedding sequence is defined as \\(\\mathcal{F} = \\{\\mathcal{F}_{dis};[\\mathcal{F}_{cls};\\mathcal{F}_{patch}]\\} \\in \\mathbb{R}^{(K + N)\\times D}\\), where \\(N\\) is the number of image patches plus one CLS token, and \\(D\\) is the embedding dimension. To ensure each DIS token focuses on its associated distorted region \\(C_k\\), we construct a mask map \\(\\boldsymbol{m} \\in \\mathbb{R}^{(K + N)}\\). Based on this, we define the attention mask \\(\\mathcal{M}\\) to preserve the original global attention structure while enforcing localized attention for DIS tokens as follows:" + }, + { + "type": "equation", + "bbox": [ + 0.09, + 0.783, + 0.454, + 0.824 + ], + "angle": 0, + "content": "\\[\nm _ {j} ^ {i} = \\left\\{ \\begin{array}{l l} \\mathbb {1} \\{\\text {t h e i - t h p a t c h o v e r l a p s w i t h C _ {i}} \\}, & j < N \\\\ 0, & j = N \\end{array} \\right.,\n\\]" + }, + { + "type": "equation", + "bbox": [ + 0.095, + 0.825, + 0.498, + 0.87 + ], + "angle": 0, + "content": "\\[\n\\mathbb {M} = \\left[ m ^ {1}; m ^ {2}; \\dots ; m ^ {K} \\right] ^ {T}, \\mathcal {M} = \\left[ \\begin{array}{c c} \\mathbf {1} _ {N \\times N} & \\mathbf {0} _ {N \\times K} \\\\ \\mathbb {M} _ {K \\times N} & \\mathbf {1} _ {K \\times K} \\end{array} \\right], \\tag {2}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.871, + 0.485, + 0.902 + ], + "angle": 0, + "content": "where \\(\\mathbb{1}\\{\\cdot\\}\\) is the indicator function. Intuitively, Eq. 2 enforces that each DIS token \\(\\mathcal{F}_{dis}^{k}\\) attends only to the distorted" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.092, + 0.908, + 0.122 + ], + "angle": 0, + "content": "content in block \\( C_k \\). Assuming the encoder has \\( H \\) layers, the Multi-Head Self-Attention is applied as follows:" + }, + { + "type": "equation", + "bbox": [ + 0.513, + 0.132, + 0.918, + 0.184 + ], + "angle": 0, + "content": "\\[\n\\boldsymbol {\\mathcal {F}} ^ {h} = \\left\\{ \\begin{array}{l l} \\sigma (\\log \\mathcal {M} + \\mathbf {Q} ^ {h} (\\mathbf {K} ^ {h}) ^ {T}) \\mathbf {V} ^ {h} + \\boldsymbol {\\mathcal {F}} ^ {h - 1}, & 0 < h \\leq H \\\\ \\boldsymbol {\\mathcal {F}}, & h = 0 \\end{array} , \\right. \\tag {3}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.184, + 0.907, + 0.276 + ], + "angle": 0, + "content": "where \\(Q^h, K^h, V^h\\) are the query, key, and value matrices, obtained via linear projections of \\(\\mathcal{F}^{h-1}\\), and \\(\\sigma\\) denotes the Softmax function. The encoder outputs features \\(\\mathcal{F}^H\\) after \\(H\\) layers. Following [29], we apply a transformer decoder to refine these features, yielding \\(\\{\\hat{\\mathcal{F}}_{dis}; \\hat{\\mathcal{F}}_{cls}\\} \\in \\mathbb{R}^{(K+1) \\times D}\\), which are more relevant for image quality." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.283, + 0.785, + 0.298 + ], + "angle": 0, + "content": "3.2.2. Quality-aware Adaptation Head" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.301, + 0.908, + 0.437 + ], + "angle": 0, + "content": "To effectively integrate global quality and local distortion features while suppressing quality-irrelevant information, we design a quality-aware adaptation head that highlights distortion cues closely related to perceived quality. Specifically, the global quality feature \\(\\hat{\\mathcal{F}}_{cls}\\) is used as the query to selectively attend to the local distortion features \\(\\hat{\\mathcal{F}}_{dis}\\), which serve as key and value. This allows the model to emphasize distortion patterns most relevant to the overall quality. Overall, the final predicted quality score \\(\\hat{Y}\\) is computed as:" + }, + { + "type": "equation", + "bbox": [ + 0.536, + 0.448, + 0.907, + 0.473 + ], + "angle": 0, + "content": "\\[\n\\hat {Y} = \\operatorname {L N} \\left(\\hat {\\mathcal {F}} _ {c l s} + \\text {C r o s s A t t e n t i o n} \\left(\\hat {\\mathcal {F}} _ {c l s}, \\hat {\\mathcal {F}} _ {d i s}\\right)\\right), \\tag {4}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.482, + 0.908, + 0.528 + ], + "angle": 0, + "content": "where \\( \\text{CrossAttention}(\\cdot) \\) denotes a cross-attention module, and \\( \\text{LN} \\) is a single-layer linear projection. During training, the model is optimized by minimizing the smooth \\( \\mathcal{L}_1 \\) loss." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.536, + 0.871, + 0.552 + ], + "angle": 0, + "content": "3.3. Distortion-Guided Knowledge Distillation" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.558, + 0.909, + 0.755 + ], + "angle": 0, + "content": "Inspired by [5, 18, 33], we adopt a knowledge distillation method to learn correlations among distortion types by transferring quality-aware distortion knowledge from the teacher to the student. This distillation strategy addresses two key challenges: First, the student's ViT encoders, pre-trained on ImageNet [4], focus on semantics and lack distortion awareness in early training [19]. The teacher, pre-trained on IQA data, provides quality-aware guidance, directing the student to critical distortion patterns. Second, the absence of explicit ground truth for distortion correlations makes it difficult for the student to model these relationships independently. The teacher acts as a reliable target for correlation modeling, thereby improving the student's discriminative ability." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.762, + 0.805, + 0.776 + ], + "angle": 0, + "content": "3.3.1. Contrastive Knowledge Distillation" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.78, + 0.909, + 0.901 + ], + "angle": 0, + "content": "Contrastive Knowledge Distillation (CKD) aims to extract the distortion features \\( \\pmb{V}_{dis}^{k} \\) of the \\( k \\)-th local block from a teacher encoder \\( \\mathcal{V}(\\cdot) \\) and transfer them to a student network. Simply maximizing feature similarity is insufficient to capture distinct distortion representations across spatial domains (Table 5). To address this, we reformulate the KD task as a pseudo-classification problem. As shown in Fig. 3, features from corresponding distorted regions of the teacher" + }, + { + "type": "page_number", + "bbox": [ + 0.482, + 0.945, + 0.516, + 0.957 + ], + "angle": 0, + "content": "2347" + } + ], + [ + { + "type": "table", + "bbox": [ + 0.095, + 0.089, + 0.907, + 0.362 + ], + "angle": 0, + "content": "
MethodLIVECSIQKADIDLIVECKonIQLIVEFBSPAQ
PLCCSRCCPLCCSRCCPLCCSRCCPLCCSRCCPLCCSRCCPLCCSRCCPLCCSRCC
BRISQUE [27]0.9440.9290.7480.8120.5670.5280.6290.6290.6850.6810.3410.3030.8170.809
DBCNN [56]0.9710.9680.9590.9460.8560.8510.8690.8510.8840.8750.5510.5450.9150.911
TIQA [54]0.9650.9490.8380.8250.8550.850.8610.8450.9030.8920.5810.541--
MetaIQA [61]0.9590.9600.9080.8990.7750.7620.8020.8350.8560.8870.5070.54--
HyperIQA (27M) [39]0.9660.9620.9420.9230.8450.8520.8820.8590.9170.9060.6020.5440.9150.911
TReS (152M) [10]0.9680.9690.9420.9220.8580.8590.8770.8460.9280.9150.6250.554--
MUSIQ (27M) [14]0.9110.9400.8930.8710.8720.8750.7460.7020.9280.9160.6610.5660.9210.918
Re-IQA (48M) [31]0.9710.9700.9600.9470.8850.8720.8540.8400.9230.914--0.9250.918
DEIQT (24M) [29]0.9820.9800.9630.9460.8870.8890.8940.8750.9340.9210.6630.5710.9230.919
QPT (24M) [58]------0.9140.8950.9410.9270.6750.5750.9280.925
LIQE (151M) [57]0.9510.9700.9390.9360.9310.9300.9100.9040.9080.919----
QFM-IQM (24M) [19]0.9830.9810.9650.9540.9060.9060.9130.8910.9360.9220.6670.5670.9240.920
LoDa (118M) [50]0.9790.975--0.9360.9310.8990.8760.9440.9320.6790.5780.9280.925
SHDIQA (24.2M)0.9840.9820.9720.9620.9400.9370.9220.9090.9480.9370.7350.6390.9340.930
" + }, + { + "type": "table_caption", + "bbox": [ + 0.105, + 0.371, + 0.89, + 0.385 + ], + "angle": 0, + "content": "Table 1. Performance comparison measured by the average of SRCC and PLCC. Best results are in bold, second-best are underlined." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.4, + 0.483, + 0.43 + ], + "angle": 0, + "content": "and student are treated as positive pairs, while all other pairs are considered negative. The CKD loss is defined as:" + }, + { + "type": "equation", + "bbox": [ + 0.1, + 0.442, + 0.484, + 0.539 + ], + "angle": 0, + "content": "\\[\n\\begin{array}{l} \\mathcal {L} _ {c k d} = - \\frac {1}{K} \\sum_ {k} \\log \\left(\\sigma \\left(\\frac {\\cos \\left(\\boldsymbol {\\mathcal {F}} _ {d i s} ^ {k} , \\boldsymbol {V} _ {d i s} ^ {k}\\right)}{\\tau_ {c}}\\right)\\right) \\\\ \\left. - \\frac{1}{K}\\sum_{\\substack{k,j\\\\ j\\neq k}}\\log \\left(1 - \\sigma \\left(\\frac{\\cos(\\boldsymbol{\\mathcal{F}}_{dis}^{k},\\boldsymbol{V}_{dis}^{j})}{\\tau_{c}}\\right)\\right)\\right), \\tag{5} \\\\ \\end{array}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.551, + 0.484, + 0.643 + ], + "angle": 0, + "content": "where \\(\\cos (\\cdot)\\) is cosine similarity, \\(\\sigma\\) is the sigmoid function, and \\(\\tau_{c} = 0.3\\) is the temperature factor. \\(\\mathcal{F}_{dis}^{k}\\) is the student feature vector of the \\(k\\)-th local block feature extracted by the DIS token. This approach enhances contrast among features of varying distortion types, improving the model's discrimination and distortion-awareness capabilities." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.651, + 0.354, + 0.667 + ], + "angle": 0, + "content": "3.3.2. Affinity Knowledge Distillation" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.671, + 0.486, + 0.747 + ], + "angle": 0, + "content": "Affinity Knowledge Distillation (AKD) focuses on distilling the interrelations among multiple blocks within an image \\(\\pmb{I}\\) or its degraded version \\(I_{d}\\). As shown in Fig. 3, given teacher embeddings \\(\\mathbf{V}_{dis}\\) and student features \\(\\mathcal{F}_{dis}\\), we compute student and teachers' pairwise affinity matrices:" + }, + { + "type": "equation", + "bbox": [ + 0.173, + 0.761, + 0.484, + 0.781 + ], + "angle": 0, + "content": "\\[\n\\boldsymbol {R} _ {s t u} = \\bar {\\boldsymbol {F}} \\bar {\\boldsymbol {F}} ^ {T}, \\quad \\boldsymbol {R} _ {t e a} = \\bar {\\boldsymbol {V}} \\bar {\\boldsymbol {V}} ^ {T}, \\tag {6}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.789, + 0.485, + 0.852 + ], + "angle": 0, + "content": "where \\(\\bar{F}\\) and \\(\\bar{V}\\) are \\(\\mathcal{L}_2\\)-normalized student \\(\\mathcal{F}_{dis}\\) and teacher features \\(V_{dis}\\), respectively. These matrices capture pairwise correlations, and the goal is to align the student affinity matrix \\(R_{stu}\\) with the teacher matrix \\(R_{tea}\\):" + }, + { + "type": "equation", + "bbox": [ + 0.107, + 0.862, + 0.484, + 0.905 + ], + "angle": 0, + "content": "\\[\n\\mathcal {L} _ {a k d} = - \\frac {1}{K} \\sum_ {k = 1} ^ {K} \\mathrm {K L} \\left(\\sigma \\left(\\frac {\\boldsymbol {R} _ {s t u} ^ {k}}{\\tau_ {a}}\\right) \\| \\sigma \\left(\\frac {\\boldsymbol {R} _ {t e a} ^ {k}}{\\tau_ {a}}\\right)\\right), \\tag {7}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.399, + 0.909, + 0.52 + ], + "angle": 0, + "content": "where \\(\\tau_{a} = 2\\) and \\(\\mathrm{KL}(\\cdot)\\) denotes the Kullback-Leibler divergence. AKD complements CKD by focusing on the relational dynamics between image blocks, allowing the model to learn the relationship between distortion. However, relying solely on AKD may overlook fine-grained distortion-discriminative details, as highlighted in Table 2. Therefore, both CKD and AKD are essential for learning comprehensive distortion-aware features." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.526, + 0.715, + 0.542 + ], + "angle": 0, + "content": "3.3.3. Overall Training Loss" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.546, + 0.909, + 0.697 + ], + "angle": 0, + "content": "The overall loss combines four complementary components: \\(\\mathcal{L}_{\\mathrm{cls}}\\) provides category-level supervision for distortion identification but lacks awareness of relationships between distortion types; \\(\\mathcal{L}_{\\mathrm{ckd}}\\) enhances feature discrimination by clustering similar distortions and separating dissimilar ones, thus compensating for this limitation; \\(\\mathcal{L}_{\\mathrm{akd}}\\) further captures global perceptual structures by modeling relative similarities across different distortion. and \\(\\mathcal{L}_{\\log}\\) regularizes training via teacher outputs [44], mitigating overfitting to quality-irrelevant patterns. Finally, the total loss is defined as:" + }, + { + "type": "equation", + "bbox": [ + 0.514, + 0.704, + 0.907, + 0.736 + ], + "angle": 0, + "content": "\\[\n\\mathcal {L} = \\lambda_ {\\mathrm {c l s}} \\mathcal {L} _ {\\mathrm {c l s}} + \\lambda_ {\\mathrm {c k d}} \\mathcal {L} _ {\\mathrm {c k d}} + \\lambda_ {\\mathrm {a k d}} \\mathcal {L} _ {\\mathrm {a k d}} + \\lambda_ {\\log} \\mathcal {L} _ {\\log} + \\| \\hat {Y} - Y _ {g} \\| _ {1}, \\tag {8}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.737, + 0.908, + 0.782 + ], + "angle": 0, + "content": "where \\( Y_{g} \\) is the ground truth, \\( \\mathcal{L}_{\\mathrm{log}} \\) is the \\( \\mathcal{L}_1 \\) loss between student and teacher predictions, and \\( \\lambda_{\\mathrm{cls}} \\), \\( \\lambda_{\\mathrm{ckd}} \\), and \\( \\lambda_{\\mathrm{akd}} \\) are set to 1, with \\( \\lambda_{\\mathrm{log}} \\) set to 0.8." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.794, + 0.646, + 0.811 + ], + "angle": 0, + "content": "4. Experiments" + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.818, + 0.907, + 0.833 + ], + "angle": 0, + "content": "4.1. Benchmark Datasets and Evaluation Protocols" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.84, + 0.909, + 0.902 + ], + "angle": 0, + "content": "Experiments were conducted on 7 typical BIQA datasets, including 3 synthetic datasets LIVE [35], CSIQ [15], KA-DID [22], and 4 authentic datasets LIVEC [9], KonIQ [13], LIVEFB [53], SPAQ [7]. Specifically, for authentic datasets," + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.516, + 0.957 + ], + "angle": 0, + "content": "2348" + } + ], + [ + { + "type": "table", + "bbox": [ + 0.095, + 0.089, + 0.905, + 0.256 + ], + "angle": 0, + "content": "
Loss LogLoss LclsLoss LakdLoss LckdCSIQKADIDLIVELIVECAvg.
PLCCSRCCPLCCSRCCPLCCSRCCPLCCSRCC
XXXX0.9520.9350.8780.8840.9660.9640.8910.8670.917
✓XXX0.9550.9410.8980.9040.9720.9700.8980.8720.926
✓✓XX0.9590.9470.9190.9160.9740.9720.9040.8900.935
✓X✓X0.9700.9600.9270.9250.9800.9780.9190.9050.945
✓XX✓0.9660.9550.9130.9110.9760.9740.9140.9020.939
✓X✓✓0.9690.9580.9370.9350.9830.9810.9200.9040.948
✓✓✓✓0.9720.9620.9400.9370.9840.9820.9220.9090.951
" + }, + { + "type": "table_caption", + "bbox": [ + 0.097, + 0.264, + 0.9, + 0.278 + ], + "angle": 0, + "content": "Table 2. Ablation experiments about knowledge distillation components on multiple datasets, where bold entries indicate the best result." + }, + { + "type": "table", + "bbox": [ + 0.093, + 0.29, + 0.485, + 0.44 + ], + "angle": 0, + "content": "
TrainingLIVEFBLIVECKonIQLIVECSIQ
TestingKonIQLIVECKonIQLIVECCSIQLIVE
DBCNN0.7160.7240.7540.7550.7580.877
P2P-BM0.7550.7380.7400.7700.712-
TReS0.7130.7400.7330.7860.761-
DEIQT0.7330.7810.7440.7940.7810.932
LoDa0.7630.8050.7450.811--
SHDIQA0.7230.7910.7820.8190.7890.943
" + }, + { + "type": "table_caption", + "bbox": [ + 0.09, + 0.45, + 0.483, + 0.478 + ], + "angle": 0, + "content": "Table 3. SRCC on the cross datasets validation. The best results are highlighted in bold, second-best are underlined." + }, + { + "type": "text", + "bbox": [ + 0.089, + 0.488, + 0.483, + 0.638 + ], + "angle": 0, + "content": "LIVEC contains 1162 images from different mobile devices and photographers. SPAQ includes 11125 photos from 66 smartphones. KonIQ contains 10073 images from public sources, while LIVEFB is the largest real-world dataset with 39810 images. The synthetic datasets involve artificially distorting original images using methods such as JPEG compression and Gaussian blur. LIVE and CSIQ respectively have 779 and 866 synthetic distortion images, with 5 and 6 types of distortions. KADID contains 10125 synthetic distortion images, spanning 25 types of distortions." + }, + { + "type": "text", + "bbox": [ + 0.089, + 0.64, + 0.484, + 0.716 + ], + "angle": 0, + "content": "Evaluation Criteria. The experiments use Pearson's Rank-Order Correlation Coefficient (SRCC) and Pearson's Linear Correlation Coefficient (PLCC) to assess prediction monotonicity and accuracy, respectively. Both metrics scale from -1 to 1, with values close to 1 indicating better performance." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.727, + 0.308, + 0.743 + ], + "angle": 0, + "content": "4.2. Implementation Details" + }, + { + "type": "text", + "bbox": [ + 0.089, + 0.75, + 0.483, + 0.901 + ], + "angle": 0, + "content": "For the student network, we follow the typical training strategy [29], using a Transformer encoder based on ViT-S from DeiT III [43] and a one-layer decoder. Our model is trained for 9 epochs with a learning rate of \\(2 \\times 10^{-4}\\), decaying by a factor of 10 every 3 epochs. The batch size is 16 for LIVEC and 128 for KonIQ. We use \\(80\\%\\) of the images for training and \\(20\\%\\) for testing, repeating the process 10 times to reduce performance bias, reporting the average SRCC and PLCC. For synthetic distortion datasets, training and testing sets are divided by reference images to ensure con" + }, + { + "type": "table", + "bbox": [ + 0.517, + 0.289, + 0.905, + 0.44 + ], + "angle": 0, + "content": "
ModeMethodsLIVECKonIQ
PLCCSRCCPLCCSRCC
20%DEIQT0.8220.7920.9080.888
LoDa0.8540.8150.9230.907
SHDIQA0.8660.8400.9280.915
40%DEIQT0.8550.8380.9220.903
LoDa0.8790.8490.9350.922
SHDIQA0.8890.8690.9420.929
" + }, + { + "type": "table_caption", + "bbox": [ + 0.513, + 0.45, + 0.907, + 0.478 + ], + "angle": 0, + "content": "Table 4. Data-efficient learning validation with the training set containing \\(20\\%\\) and \\(40\\%\\) of the images." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.487, + 0.907, + 0.563 + ], + "angle": 0, + "content": "tent independence. To provide reliable distortion perception early in training, the teacher model adopts a similar encoder-decoder architecture aligned with the student model, and is pre-trained with \\(\\mathcal{L}_1\\) loss on the same IQA dataset. Training is performed on four NVIDIA RTX 3090 GPUs." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.573, + 0.895, + 0.59 + ], + "angle": 0, + "content": "4.3. Overall Prediction Performance Comparison" + }, + { + "type": "text", + "bbox": [ + 0.511, + 0.596, + 0.909, + 0.779 + ], + "angle": 0, + "content": "Table 1 reports the comparison results between the proposed SHDIQA and other state-of-the-art (SOTA) BIQA methods, including ViT-based IQA methods, i.e., LoDa [50] and DEIQT [29], and Degradation Modelling-based methods, i.e., Re-IQA [31] and QPT [58]. It is observed from these seven datasets that SHDIQA achieves superior performance over all other methods across the 7 datasets. Notably, on the most difficult LIVEFB dataset, SHDIQA achieves an impressive performance improvement. This enhancement can be attributed to the prevalence of local distortions [53] in the LIVEFB dataset, highlighting the importance of effectively utilizing local distortion priors." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.788, + 0.818, + 0.804 + ], + "angle": 0, + "content": "4.4. Data-Efficient Learning Validation" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.811, + 0.909, + 0.903 + ], + "angle": 0, + "content": "We showcase the data efficiency of our approach, as presented in Table 4. Following the efficient validation protocol outlined in LoDa [50] and DEIQT [29], we train models using \\(20\\%\\) and \\(40\\%\\) of the available images. This process is repeated 10 times, and we report the average SRCC and PLCC scores. By introducing distortion priors through the" + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.516, + 0.956 + ], + "angle": 0, + "content": "2349" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.094, + 0.089, + 0.912, + 0.308 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.09, + 0.311, + 0.908, + 0.34 + ], + "angle": 0, + "content": "Figure 4. gMAD competition results between DEIQT and our method SHDIQA. (a) Fixed DEIQT at low quality. (b) Fixed DEIQT at high quality. (c) Fixed SHDIQA at low quality. (d) Fixed SHDIQA at high quality. (e) Cosine similarity between different DIS and CLS tokens." + }, + { + "type": "table", + "bbox": [ + 0.098, + 0.351, + 0.475, + 0.441 + ], + "angle": 0, + "content": "
LckdLIVECCSIQAvg.
PLCCSRCCPLCCSRCC
-0.8910.8670.9520.9350.911
L10.9080.8950.9570.9460.926
BCE0.9140.9020.9660.9550.934
" + }, + { + "type": "table_caption", + "bbox": [ + 0.098, + 0.445, + 0.475, + 0.46 + ], + "angle": 0, + "content": "Table 5. Ablation study on Contrastive Knowledge Distillation." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.47, + 0.484, + 0.547 + ], + "angle": 0, + "content": "Block-wise Degradation, our method achieves performance that is nearly with SOTA results, even when trained on only \\(20\\%\\) of the dataset. Furthermore, our approach consistently outperforms all other methods on both datasets, demonstrating its superiority over existing efficient BIQA techniques." + }, + { + "type": "title", + "bbox": [ + 0.09, + 0.561, + 0.411, + 0.577 + ], + "angle": 0, + "content": "4.5. Generalization Capability Validation" + }, + { + "type": "text", + "bbox": [ + 0.089, + 0.584, + 0.486, + 0.902 + ], + "angle": 0, + "content": "We evaluate SHDIQA's generalization ability using cross-dataset validation, training the BIQA model on one dataset and testing it on others. Table 3 shows the SRCC averages for five datasets, where SHDIQA achieves the best performance in four out of six cross-dataset evaluations. To further test our framework's generalization, we trained the model on the LIVE database and tested it on the Waterloo database using the gMAD competition [24]. In the gMAD competition, image pairs are selected where the attacker model predicts a large quality difference, while the defender perceives them as similar. Observers then evaluate these pairs to assess the model's generalization ability. As shown in Fig. 4, our model correctly selects image pairs with significant quality differences when attacking DEIQT, where DEIQT misclassifies pairs like (a) as low quality and (b) as high quality, contrary to human perception and SHDIQA's predictions. As a defender, SHDIQA identifies pairs with minimal perceptual variation, correctly classifying image pairs like (c) and (d) as both low and high quality, respectively, while DEIQT fails. This demonstrates the strong generalization capability of SHDIQA in handling real-world image distortions." + }, + { + "type": "table", + "bbox": [ + 0.518, + 0.351, + 0.9, + 0.442 + ], + "angle": 0, + "content": "
\\( {\\mathcal{L}}_{\\text{akd }}^{I} \\)\\( {\\mathcal{L}}_{\\text{akd }}^{I_d} \\)LIVECLIVEAvg.
PLCCSRCCPLCCSRCC
✓0.9170.9010.9820.9780.944
✓0.9150.9000.9800.9800.943
✓✓0.9220.9090.9840.9820.949
" + }, + { + "type": "table_caption", + "bbox": [ + 0.514, + 0.446, + 0.904, + 0.46 + ], + "angle": 0, + "content": "Table 6. Ablation experiment on Affinity Knowledge Distillation." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.469, + 0.667, + 0.485 + ], + "angle": 0, + "content": "4.6. Ablation Study" + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.491, + 0.829, + 0.507 + ], + "angle": 0, + "content": "4.6.1. Effectiveness of Proposed Components" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.51, + 0.909, + 0.798 + ], + "angle": 0, + "content": "We conduct an ablation study on components of our framework. Table 2 shows that using the teacher's quality score as a regularization improves model performance. Adding a distortion classification task boosts SRCC by 1.2 and 1.8 points on KADID and LIVEC, respectively, highlighting the effectiveness of distortion-aware supervision. However, identifying distortions without modeling their relationships provides limited benefits. To address this, we introduce CKD and AKD to capture distortion correlations, improving SRCC by up to 3.5 and 4.9 points on KADID, respectively. The limited gain from CKD suggests that aggressively separating regions with different distortions may be problematic, as these regions can have similar quality. AKD mitigates this issue by capturing structural relationships between image blocks. When combined, CKD and AKD improve the baseline by 6.2 points, confirming their complementarity. Finally, Table 8 shows that block-wise aggregation and filtering further enhance the model's ability to assess local distortions, leading to better quality prediction." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.807, + 0.887, + 0.822 + ], + "angle": 0, + "content": "4.6.2. Loss type of CKD and Implementation of AKD" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.826, + 0.907, + 0.901 + ], + "angle": 0, + "content": "Table 5 shows the effect of different CKD loss functions. Using \\(\\mathcal{L}_1\\) loss improves average quality prediction by 1.5 points, and switching to a BCE-based contrastive loss adds another 0.8-point gain. This confirms the effectiveness of our CKD design in promoting feature clustering within the" + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.516, + 0.957 + ], + "angle": 0, + "content": "2350" + } + ], + [ + { + "type": "table", + "bbox": [ + 0.095, + 0.088, + 0.278, + 0.182 + ], + "angle": 0, + "content": "
NumberLIVEC
PLCCSRCC
30.9200.903
40.9220.909
50.9180.905
" + }, + { + "type": "table_caption", + "bbox": [ + 0.091, + 0.185, + 0.282, + 0.214 + ], + "angle": 0, + "content": "Table 7. Ablation study on the number of degenerated blocks." + }, + { + "type": "table", + "bbox": [ + 0.296, + 0.088, + 0.583, + 0.182 + ], + "angle": 0, + "content": "
MethodLIVECLIVE
PLCCSRCCPLCCSRCC
Baseline0.8910.8670.9660.964
+BWA0.9180.9020.9820.978
+BWF0.9220.9090.9840.982
" + }, + { + "type": "table_caption", + "bbox": [ + 0.295, + 0.185, + 0.586, + 0.214 + ], + "angle": 0, + "content": "Table 8. Ablation study on the Block-wise Aggregation (BwA) and Filtering (BWF) module." + }, + { + "type": "table", + "bbox": [ + 0.598, + 0.089, + 0.9, + 0.182 + ], + "angle": 0, + "content": "
MethodLIVECKonIQ
PLCCSRCCPLCCSRCC
Homogenous0.9070.8930.9390.930
Heterogenous0.9220.9090.9480.937
" + }, + { + "type": "table_caption", + "bbox": [ + 0.596, + 0.185, + 0.902, + 0.214 + ], + "angle": 0, + "content": "Table 9. Ablation experiments on heterogeneous and homogeneous distortion on the whole image." + }, + { + "type": "image", + "bbox": [ + 0.093, + 0.22, + 0.208, + 0.308 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.114, + 0.31, + 0.189, + 0.323 + ], + "angle": 0, + "content": "Mos: 38.6" + }, + { + "type": "image", + "bbox": [ + 0.209, + 0.222, + 0.322, + 0.308 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.215, + 0.31, + 0.316, + 0.323 + ], + "angle": 0, + "content": "Baseline: 46.1" + }, + { + "type": "image", + "bbox": [ + 0.326, + 0.222, + 0.439, + 0.308 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.325, + 0.31, + 0.441, + 0.324 + ], + "angle": 0, + "content": "Ours: 35.2(↓4.1)" + }, + { + "type": "image", + "bbox": [ + 0.443, + 0.222, + 0.556, + 0.308 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.456, + 0.31, + 0.541, + 0.323 + ], + "angle": 0, + "content": "DIS token 1" + }, + { + "type": "image", + "bbox": [ + 0.56, + 0.222, + 0.672, + 0.308 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.575, + 0.31, + 0.663, + 0.323 + ], + "angle": 0, + "content": "DIS token 2" + }, + { + "type": "image", + "bbox": [ + 0.675, + 0.222, + 0.788, + 0.308 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.687, + 0.31, + 0.774, + 0.323 + ], + "angle": 0, + "content": "DIS token 3" + }, + { + "type": "image", + "bbox": [ + 0.792, + 0.222, + 0.905, + 0.308 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.804, + 0.31, + 0.891, + 0.323 + ], + "angle": 0, + "content": "DIS token 4" + }, + { + "type": "image", + "bbox": [ + 0.093, + 0.326, + 0.208, + 0.413 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.114, + 0.415, + 0.188, + 0.429 + ], + "angle": 0, + "content": "Mos: 57.3" + }, + { + "type": "image", + "bbox": [ + 0.21, + 0.326, + 0.322, + 0.412 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.217, + 0.415, + 0.317, + 0.429 + ], + "angle": 0, + "content": "Baseline: 67.1" + }, + { + "type": "image", + "bbox": [ + 0.327, + 0.326, + 0.439, + 0.412 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.326, + 0.415, + 0.443, + 0.431 + ], + "angle": 0, + "content": "Ours: 60.9(↓6.2)" + }, + { + "type": "image", + "bbox": [ + 0.443, + 0.326, + 0.555, + 0.412 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.457, + 0.415, + 0.542, + 0.429 + ], + "angle": 0, + "content": "DIS token 1" + }, + { + "type": "image", + "bbox": [ + 0.56, + 0.326, + 0.672, + 0.412 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.577, + 0.415, + 0.662, + 0.429 + ], + "angle": 0, + "content": "Dis token 2" + }, + { + "type": "image", + "bbox": [ + 0.677, + 0.326, + 0.788, + 0.412 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.688, + 0.415, + 0.775, + 0.429 + ], + "angle": 0, + "content": "DIS token 3" + }, + { + "type": "image", + "bbox": [ + 0.793, + 0.326, + 0.905, + 0.412 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.804, + 0.415, + 0.892, + 0.429 + ], + "angle": 0, + "content": "DIS token 4" + }, + { + "type": "image_caption", + "bbox": [ + 0.09, + 0.434, + 0.907, + 0.464 + ], + "angle": 0, + "content": "Figure 5. Activation maps of DEIQT [29] and SHDIQA using Grad-CAM [32] on the LIVEFB dataset, with MOS shown below the images. Our SHDIQA effectively captures the interplay between image content, local distortions, and overall quality, enabling accurate predictions." + }, + { + "type": "text", + "bbox": [ + 0.089, + 0.47, + 0.486, + 0.592 + ], + "angle": 0, + "content": "same distortion type and separation across types. For AKD, we evaluate three strategies: applying it to the input image \\( \\mathbf{I} \\), the degraded image \\( \\mathbf{I}_d \\), or both. Table 6 shows that the best performance is achieved by using both, as applying AKD to the image \\( \\mathbf{I} \\) helps model relationships within local content, while applying it to the \\( \\mathbf{I}_d \\) captures correlations between different types of distortions. Combining both enables complementary learning of content and distortion information." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.597, + 0.383, + 0.612 + ], + "angle": 0, + "content": "4.6.3. Ablation on the Degradation Space" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.615, + 0.486, + 0.782 + ], + "angle": 0, + "content": "We investigate the effectiveness of the Block-wise Degradation Modelling. As shown in Table 9, applying globally uniform heterogeneous distortions significantly underperforms compared to localized distortions, highlighting the importance of local degradation. We also evaluate how varying the number of degraded blocks affects performance. As shown in Table 7, the impact is relatively insensitive across different partitioning strategies, demonstrating the method's resilience to changes in block quantity. While more blocks increase the scope of degradation, they also lead to greater content loss. Thus, using four blocks strikes the best balance." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.788, + 0.283, + 0.805 + ], + "angle": 0, + "content": "4.7. Qualitative Analysis" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.811, + 0.486, + 0.903 + ], + "angle": 0, + "content": "To show that tokens from different perspectives capture unique features, we analyze the cosine similarity between CLS and DIS tokens on the LIVEC and KonIQ datasets. As shown in Fig. 4, the similarities range from 0.72 to 0.81, lower than the 0.96 and 0.94 found in DeiT-T and DeiT-S [42]. Tokens near the center show higher similarity," + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.47, + 0.907, + 0.669 + ], + "angle": 0, + "content": "aligning with the HVS's focus on salient regions [62]. This implies that DIS tokens capture localized distortions in a way that simulates human perception, and when combined with CLS tokens, it improves both global and local distortion awareness. To further explore the DIS tokens, we visualize the attention maps of DEIQT and our SHDIQA, as shown in Fig. 5. By incorporating four additional distortion tokens, the proposed model enhances its ability to detect distortions across different regions of the image. Through the interaction between distortion tokens and CLS tokens in the adaptation head, the model adaptively focuses on the most relevant areas for image quality, while also considering the overall content. This results in more accurate quality predictions." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.683, + 0.634, + 0.699 + ], + "angle": 0, + "content": "5. Conclusion" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.705, + 0.909, + 0.903 + ], + "angle": 0, + "content": "In this paper, we propose SHDIQA, a novel BIQA approach that leverages spatial-heterogeneous distortion-aware strategies. Our Block-Wise Degradation Modelling generates locally distorted images, exposing the model to diverse distortion patterns. The Block-wise Aggregation and Filtering mechanism enhances the ViT model's sensitivity to local distortions, improving quality assessments through focused attention and feature filtering. Additionally, our Distortion-Guided Knowledge Distillation combines CKD and AKD to improve distortion discrimination and relational learning. Experimental results show SHDIQA's superior generalization and robustness across various datasets and distortion types, confirming its effectiveness in real-world BIQA tasks." + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.514, + 0.957 + ], + "angle": 0, + "content": "2351" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.09, + 0.092, + 0.486, + 0.228 + ], + "angle": 0, + "content": "Acknowledgements. This work was supported by the National Science Fund for Distinguished Young Scholars (No.62025603), the National Natural Science Foundation of China (No. U21B2037, No. U22B2051, No. U23A20383, No. 62176222, No. 62176223, No. 62176226, No. 62072386, No. 62072387, No. 62072389, No. 62002305 and No. 62272401), and the Natural Science Foundation of Fujian Province of China (No. 2021J06003, No.2022J06001)." + }, + { + "type": "title", + "bbox": [ + 0.093, + 0.241, + 0.188, + 0.256 + ], + "angle": 0, + "content": "References" + }, + { + "type": "ref_text", + "bbox": [ + 0.101, + 0.266, + 0.484, + 0.307 + ], + "angle": 0, + "content": "[1] Mark R Banham and Aggelos K Katsaggelos. Digital image restoration. IEEE signal processing magazine, 14(2):24-41, 1997. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.101, + 0.309, + 0.484, + 0.377 + ], + "angle": 0, + "content": "[2] Chaofeng Chen, Jiadi Mo, Jingwen Hou, Haoning Wu, Liang Liao, Wenxiu Sun, Qiong Yan, and Weisi Lin. Topiq: A top-down approach from semantics to distortions for image quality assessment. IEEE Transactions on Image Processing, 2024. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.379, + 0.484, + 0.42 + ], + "angle": 0, + "content": "[3] Xinlei Chen, Haoqi Fan, Ross Girshick, and Kaiming He. Improved baselines with momentum contrastive learning. arXiv preprint arXiv:2003.04297, 2020. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.422, + 0.484, + 0.478 + ], + "angle": 0, + "content": "[4] Jia Deng, Wei Dong, Richard Socher, Li-Jia Li, Kai Li, and Li Fei-Fei. Imagenet: A large-scale hierarchical image database. In 2009 IEEE conference on computer vision and pattern recognition, pages 248-255. IEEE, 2009. 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.479, + 0.484, + 0.534 + ], + "angle": 0, + "content": "[5] Jinhong Deng, Wen Li, Yuhua Chen, and Lixin Duan. Unbiased mean teacher for cross-domain object detection. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 4091-4101, 2021. 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.536, + 0.484, + 0.589 + ], + "angle": 0, + "content": "[6] Chao Dong, Chen Change Loy, Kaiming He, and Xiaou Tang. Image super-resolution using deep convolutional networks. IEEE transactions on pattern analysis and machine intelligence, 38(2):295-307, 2015. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.592, + 0.484, + 0.659 + ], + "angle": 0, + "content": "[7] Yuming Fang, Hanwei Zhu, Yan Zeng, Kede Ma, and Zhou Wang. Perceptual quality assessment of smartphone photography. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 3677-3686, 2020. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.662, + 0.484, + 0.717 + ], + "angle": 0, + "content": "[8] Nuno C Garcia, Pietro Morerio, and Vittorio Murino. Modality distillation with multiple stream networks for action recognition. In Proceedings of the European Conference on Computer Vision (ECCV), pages 103-118, 2018. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.719, + 0.484, + 0.773 + ], + "angle": 0, + "content": "[9] Deepti Ghadiyaram and Alan C Bovik. Massive online crowdsourced study of subjective and objective picture quality. IEEE Transactions on Image Processing, 25(1):372-387, 2015. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.776, + 0.484, + 0.844 + ], + "angle": 0, + "content": "[10] S Alireza Golestaneh, Saba Dadsetan, and Kris M Kitani. No-reference image quality assessment via transformers, relative ranking, and self-consistency. In Proceedings of the IEEE/CVF Winter Conference on Applications of Computer Vision, pages 1220-1230, 2022. 2, 3, 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.846, + 0.484, + 0.901 + ], + "angle": 0, + "content": "[11] Saurabh Gupta, Judy Hoffman, and Jitendra Malik. Cross modal distillation for supervision transfer. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 2827-2836, 2016. 3" + }, + { + "type": "list", + "bbox": [ + 0.094, + 0.266, + 0.484, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.093, + 0.908, + 0.133 + ], + "angle": 0, + "content": "[12] Geoffrey Hinton, Oriol Vinyals, and Jeff Dean. Distilling the knowledge in a neural network. arXiv preprint arXiv:1503.02531, 2015. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.136, + 0.908, + 0.19 + ], + "angle": 0, + "content": "[13] Vlad Hosu, Hanhe Lin, Tamas Sziranyi, and Dietmar Saupe. Koniq-10k: An ecologically valid database for deep learning of blind image quality assessment. IEEE Transactions on Image Processing, 29:4041-4056, 2020. 1, 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.193, + 0.907, + 0.248 + ], + "angle": 0, + "content": "[14] Junjie Ke, Qifei Wang, Yilin Wang, Peyman Milanfar, and Feng Yang. Musiq: Multi-scale image quality transformer. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 5148-5157, 2021. 2, 3, 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.25, + 0.908, + 0.304 + ], + "angle": 0, + "content": "[15] Eric C Larson and Damon M Chandler. Most apparent distortion: full-reference image quality assessment and the role of strategy. Journal of electronic imaging, 19(1):011006-011006, 2010. 2, 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.307, + 0.907, + 0.362 + ], + "angle": 0, + "content": "[16] Aobo Li, Jinjian Wu, Yongxu Liu, and Leida Li. Bridging the synthetic-to Authentic gap: Distortion-guided unsupervised domain adaptation for blind image quality assessment. arXiv preprint arXiv:2405.04167, 2024. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.364, + 0.907, + 0.433 + ], + "angle": 0, + "content": "[17] Leida Li, Tianshu Song, Jinjian Wu, Weisheng Dong, Jiansheng Qian, and Guangming Shi. Blind image quality index for authentic distortions with local and global deep feature aggregation. IEEE Transactions on Circuits and Systems for Video Technology, 32(12):8512-8523, 2021. 2, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.435, + 0.907, + 0.503 + ], + "angle": 0, + "content": "[18] Liangqi Li, Jiaxu Miao, Dahu Shi, Wenming Tan, Ye Ren, Yi Yang, and Shiliang Pu. Distilling detr with visual-linguistic knowledge for open-vocabulary object detection. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 6501–6510, 2023. 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.505, + 0.908, + 0.589 + ], + "angle": 0, + "content": "[19] Xudong Li, Timin Gao, Runze Hu, Yan Zhang, Shengchuan Zhang, Xiawu Zheng, Jingyuan Zheng, Yunhang Shen, Ke Li, Yutao Liu, et al. Adaptive feature selection for no-reference image quality assessment by mitigating semantic noise sensitivity. In International Conference on Machine Learning, pages 27808-27821. PMLR, 2024. 4, 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.591, + 0.908, + 0.673 + ], + "angle": 0, + "content": "[20] Xudong Li, Runze Hu, Jingyuan Zheng, Yan Zhang, Shengchuan Zhang, Xiawu Zheng, Ke Li, Yunhang Shen, Yutao Liu, Pingyang Dai, et al. Integrating global context contrast and local sensitivity for blind image quality assessment. In International Conference on Machine Learning, pages 27920-27941. PMLR, 2024. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.675, + 0.907, + 0.743 + ], + "angle": 0, + "content": "[21] Xudong Li, Jingyuan Zheng, Runze Hu, Yan Zhang, Ke Li, Yunhang Shen, Xiawu Zheng, Yutao Liu, ShengChuan Zhang, Pingyang Dai, et al. Feature denoising diffusion model for blind image quality assessment. arXiv preprint arXiv:2401.11949, 2024.3" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.746, + 0.908, + 0.801 + ], + "angle": 0, + "content": "[22] Hanhe Lin, Vlad Hosu, and Dietmar Saupe. Kadid-10k: A large-scale artificially distorted iqa database. In 2019 Eleventh International Conference on Quality of Multimedia Experience (QoMEX), pages 1-3. IEEE, 2019. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.803, + 0.907, + 0.844 + ], + "angle": 0, + "content": "[23] Hao Liu and Pieter Abbeel. Blockwise parallel transformers for large context models. Advances in Neural Information Processing Systems, 36, 2024. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.846, + 0.907, + 0.901 + ], + "angle": 0, + "content": "[24] Kede Ma, Qingbo Wu, Zhou Wang, Zhengfang Duanmu, Hongwei Yong, Hongliang Li, and Lei Zhang. Group mad competition-a new methodology to compare objective image quality models. In Proceedings of the IEEE Conference on" + }, + { + "type": "list", + "bbox": [ + 0.094, + 0.092, + 0.908, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.516, + 0.956 + ], + "angle": 0, + "content": "2352" + } + ], + [ + { + "type": "ref_text", + "bbox": [ + 0.126, + 0.093, + 0.484, + 0.119 + ], + "angle": 0, + "content": "Computer Vision and Pattern Recognition, pages 1664-1673, 2016. 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.122, + 0.484, + 0.177 + ], + "angle": 0, + "content": "[25] Kede Ma, Wentao Liu, Tongliang Liu, Zhou Wang, and Dacheng Tao. dipiq: Blind image quality assessment by learning-to-rank discriminable image pairs. IEEE Transactions on Image Processing, 26(8):3951-3964, 2017. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.179, + 0.482, + 0.234 + ], + "angle": 0, + "content": "[26] Kede Ma, Wentao Liu, Kai Zhang, Zhengfang Duanmu, Zhou Wang, and Wangmeng Zuo. End-to-end blind image quality assessment using deep neural networks. IEEE Transactions on Image Processing, 27(3):1202-1213, 2017. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.236, + 0.483, + 0.29 + ], + "angle": 0, + "content": "[27] Anish Mittal, Anush Krishna Moorthy, and Alan Conrad Bovik. No-reference image quality assessment in the spatial domain. IEEE Transactions on image processing, 21(12): 4695-4708, 2012. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.293, + 0.484, + 0.362 + ], + "angle": 0, + "content": "[28] Angelo Porrello, Luca Bergamini, and Simone Calderara. Robust re-identification by multiple views knowledge distillation. In Computer Vision-ECCV 2020: 16th European Conference, Glasgow, UK, August 23–28, 2020, Proceedings, Part X 16, pages 93–110. Springer, 2020. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.363, + 0.484, + 0.432 + ], + "angle": 0, + "content": "[29] Guanyi Qin, Runze Hu, Yutao Liu, Xiawu Zheng, Haotian Liu, Xiu Li, and Yan Zhang. Data-efficient image quality assessment with attention-panel decoder. In Proceedings of the Thirty-Seventh AAAI Conference on Artificial Intelligence, 2023. 3, 4, 5, 6, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.434, + 0.484, + 0.488 + ], + "angle": 0, + "content": "[30] Jiezhong Qiu, Hao Ma, Omer Levy, Scott Wen-tau Yih, Sinong Wang, and Jie Tang. Blockwise self-attention for long document understanding. arXiv preprint arXiv:1911.02972, 2019. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.491, + 0.484, + 0.559 + ], + "angle": 0, + "content": "[31] Avinab Saha, Sandeep Mishra, and Alan C Bovik. Re-iqa: Unsupervised learning for image quality assessment in the wild. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 5846-5855, 2023. 2, 4, 5, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.562, + 0.484, + 0.631 + ], + "angle": 0, + "content": "[32] Ramprasaath R Selvaraju, Michael Cogswell, Abhishek Das, Ramakrishna Vedantam, Devi Parikh, and Dhruv Batra. Gradcam: Visual explanations from deep networks via gradient-based localization. In Proceedings of the IEEE international conference on computer vision, pages 618-626, 2017. 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.633, + 0.484, + 0.701 + ], + "angle": 0, + "content": "[33] Tahira Shehzadi, Khurram Azeem Hashmi, Didier Stricker, and Muhammad Zeshan Afzal. Sparse semi-detr: Sparse learnable queries for semi-supervised object detection. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 5840-5850, 2024. 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.703, + 0.484, + 0.758 + ], + "angle": 0, + "content": "[34] Hamid R Sheikh, Alan C Bovik, and Lawrence Cormack. No-reference quality assessment using natural scene statistics: Jpeg2000. IEEE Transactions on image processing, 14(11): 1918-1927, 2005. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.76, + 0.484, + 0.814 + ], + "angle": 0, + "content": "[35] Hamid R Sheikh, Muhammad F Sabir, and Alan C Bovik. A statistical evaluation of recent full reference image quality assessment algorithms. IEEE Transactions on image processing, 15(11):3440-3451, 2006. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.817, + 0.484, + 0.87 + ], + "angle": 0, + "content": "[36] Chenyang Shi and Yandan Lin. Full reference image quality assessment based on visual salience with color appearance and gradient similarity. IEEE Access, 8:97310-97320, 2020. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.873, + 0.484, + 0.901 + ], + "angle": 0, + "content": "[37] Tianshu Song, Leida Li, Deqiang Cheng, Pengfei Chen, and Jinjian Wu. Active learning-based sample selection for label-" + }, + { + "type": "list", + "bbox": [ + 0.093, + 0.093, + 0.484, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.545, + 0.093, + 0.905, + 0.12 + ], + "angle": 0, + "content": "efficient blind image quality assessment. IEEE Transactions on Circuits and Systems for Video Technology, 2023. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.122, + 0.907, + 0.19 + ], + "angle": 0, + "content": "[38] Suhas Srinath, Shankhanil Mitra, Shika Rao, and Rajiv Soundararajan. Learning generalizable perceptual representations for data-efficient no-reference image quality assessment. In Proceedings of the IEEE/CVF Winter Conference on Applications of Computer Vision, pages 22-31, 2024. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.192, + 0.907, + 0.259 + ], + "angle": 0, + "content": "[39] Shaolin Su, Qingsen Yan, Yu Zhu, Cheng Zhang, Xin Ge, Jinqiu Sun, and Yanning Zhang. Blindly assess image quality in the wild guided by a self-adaptive hyper network. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 3667-3676, 2020. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.261, + 0.907, + 0.314 + ], + "angle": 0, + "content": "[40] Simeng Sun, Tao Yu, Jiahua Xu, Wei Zhou, and Zhibo Chen. Graphiqa: Learning distortion graph representations for blind image quality assessment. IEEE Transactions on Multimedia, 25:2912-2925, 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.316, + 0.907, + 0.37 + ], + "angle": 0, + "content": "[41] Dacheng Tao, Xuelong Li, Wen Lu, and Xinbo Gao. Reduced-reference iqa in contourlet domain. IEEE Transactions on Systems, Man, and Cybernetics, Part B (Cybernetics), 39(6): 1623-1627, 2009. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.372, + 0.907, + 0.44 + ], + "angle": 0, + "content": "[42] Hugo Touvron, Matthieu Cord, Matthijs Douze, Francisco Massa, Alexandre Sablayrolles, and Hervé Jégou. Training data-efficient image transformers & distillation through attention. In International Conference on Machine Learning, pages 10347-10357. PMLR, 2021. 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.442, + 0.907, + 0.48 + ], + "angle": 0, + "content": "[43] Hugo Touvron, Matthieu Cord, and Hervé Jégou. Deit iii: Revenge of the vit. arXiv preprint arXiv:2204.07118, 2022. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.483, + 0.907, + 0.538 + ], + "angle": 0, + "content": "[44] Tao Wang, Li Yuan, Xiaopeng Zhang, and Jiashi Feng. Distilling object detectors with fine-grained feature imitation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 4933-4942, 2019. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.54, + 0.906, + 0.58 + ], + "angle": 0, + "content": "[45] Zhou Wang and Eero Simoncelli. Local phase coherence and the perception of blur. Advances in neural information processing systems, 16, 2003. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.582, + 0.907, + 0.635 + ], + "angle": 0, + "content": "[46] Zhou Wang, Hamid R Sheikh, and Alan C Bovik. No-reference perceptual quality assessment of JPEG compressed images. In Proceedings. International conference on image processing, pages I-I. IEEE, 2002. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.637, + 0.907, + 0.691 + ], + "angle": 0, + "content": "[47] Zhou Wang, Alan C Bovik, Hamid R Sheikh, and Eero P Simoncelli. Image quality assessment: from error visibility to structural similarity. IEEE transactions on image processing, 13(4):600-612, 2004. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.693, + 0.907, + 0.747 + ], + "angle": 0, + "content": "[48] Zhihua Wang, Zhi-Ri Tang, Jianguo Zhang, and Yuming Fang. Toward a blind image quality evaluator in the wild by learning beyond human opinion scores. Pattern Recognition, 137:109296, 2023. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.749, + 0.907, + 0.816 + ], + "angle": 0, + "content": "[49] Haoning Wu, Zicheng Zhang, Erli Zhang, Chaofeng Chen, Liang Liao, Annan Wang, Chunyi Li, Wenxiu Sun, Qiong Yan, Guangtao Zhai, et al. Q-bench: A benchmark for general-purpose foundation models on low-level vision. arXiv preprint arXiv:2309.14181, 2023. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.818, + 0.907, + 0.872 + ], + "angle": 0, + "content": "[50] Kangmin Xu, Liang Liao, Jing Xiao, Chaofeng Chen, Haoning Wu, Qiong Yan, and Weisi Lin. Local distortion aware efficient transformer adaptation for image quality assessment. arXiv preprint arXiv:2308.12001, 2023. 2, 3, 5, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.874, + 0.907, + 0.901 + ], + "angle": 0, + "content": "[51] Sidi Yang, Tianhe Wu, Shuwei Shi, Shanshan Lao, Yuan Gong, Mingdeng Cao, Jiahao Wang, and Yujiu Yang. Maniqa:" + }, + { + "type": "list", + "bbox": [ + 0.517, + 0.093, + 0.907, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.516, + 0.957 + ], + "angle": 0, + "content": "2353" + } + ], + [ + { + "type": "ref_text", + "bbox": [ + 0.123, + 0.092, + 0.484, + 0.147 + ], + "angle": 0, + "content": "Multi-dimension attention network for no-reference image quality assessment. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 1191-1200, 2022. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.149, + 0.484, + 0.205 + ], + "angle": 0, + "content": "[52] Guanghao Yin, Wei Wang, Zehuan Yuan, Chuchu Han, Wei Ji, Shouqian Sun, and Changhu Wang. Content-variant reference image quality assessment via knowledge distillation. arXiv preprint arXiv:2202.13123, 2022. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.207, + 0.484, + 0.289 + ], + "angle": 0, + "content": "[53] Zhenqiang Ying, Haoran Niu, Praful Gupta, Dhruv Mahajan, Deepti Ghadiyaram, and Alan Bovik. From patches to pictures (paq-2-piq): Mapping the perceptual space of picture quality. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 3575–3585, 2020. 5, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.291, + 0.484, + 0.346 + ], + "angle": 0, + "content": "[54] Junyong You and Jari Korhonen. Transformer for image quality assessment. In 2021 IEEE International Conference on Image Processing (ICIP), pages 1389-1393. IEEE, 2021. 3, 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.349, + 0.484, + 0.417 + ], + "angle": 0, + "content": "[55] Guanghui Yue, Di Cheng, Leida Li, Tianwei Zhou, Hantao Liu, and Tianfu Wang. Semi-supervised authentically distorted image quality assessment with consistency-preserving dual-branch convolutional neural network. IEEE Transactions on Multimedia, 25:6499–6511, 2022. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.42, + 0.484, + 0.476 + ], + "angle": 0, + "content": "[56] Weixia Zhang, Kede Ma, Jia Yan, Dexiang Deng, and Zhou Wang. Blind image quality assessment using a deep bilinear convolutional neural network. IEEE Transactions on Circuits and Systems for Video Technology, 30(1):36-47, 2018. 2, 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.477, + 0.484, + 0.559 + ], + "angle": 0, + "content": "[57] Weixia Zhang, Guangtao Zhai, Ying Wei, Xiaokang Yang, and Kede Ma. Blind image quality assessment via vision-language correspondence: A multitask learning perspective. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 14071-14081, 2023. 2, 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.562, + 0.484, + 0.63 + ], + "angle": 0, + "content": "[58] Kai Zhao, Kun Yuan, Ming Sun, Mading Li, and Xing Wen. Quality-aware pre-trained models for blind image quality assessment. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 22302-22313, 2023. 2, 4, 5, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.633, + 0.484, + 0.702 + ], + "angle": 0, + "content": "[59] Heliang Zheng, Huan Yang, Jianlong Fu, Zheng-Jun Zha, and Jiebo Luo. Learning conditional knowledge distillation for degraded-reference image quality assessment. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 10242-10251, 2021. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.704, + 0.484, + 0.771 + ], + "angle": 0, + "content": "[60] Yu Zhou, Leida Li, Shiqi Wang, Jinjian Wu, Yuming Fang, and Xinbo Gao. No-reference quality assessment for view synthesis using dog-based edge statistics and texture naturalness. IEEE Transactions on Image Processing, 28(9):4566-4579, 2019. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.775, + 0.484, + 0.844 + ], + "angle": 0, + "content": "[61] Hancheng Zhu, Leida Li, Jinjian Wu, Weisheng Dong, and Guangming Shi. Metaiaq: Deep meta-learning for no-reference image quality assessment. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 14143-14152, 2020. 2, 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.846, + 0.484, + 0.901 + ], + "angle": 0, + "content": "[62] Mengmeng Zhu, Guanqun Hou, Xinjia Chen, Jiaxing Xie, Haixian Lu, and Jun Che. Saliency-guided transformer network combined with local embedding for no-reference image quality assessment. In Proceedings of the IEEE/CVF Inter-" + }, + { + "type": "list", + "bbox": [ + 0.093, + 0.092, + 0.484, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.548, + 0.092, + 0.908, + 0.119 + ], + "angle": 0, + "content": "national Conference on Computer Vision, pages 1953-1962, 2021.8" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.122, + 0.909, + 0.177 + ], + "angle": 0, + "content": "[63] Fuzhen Zhuang, Zhiyuan Qi, Keyu Duan, Dongbo Xi, Yongchun Zhu, Hengshu Zhu, Hui Xiong, and Qing He. A comprehensive survey on transfer learning. Proceedings of the IEEE, 109(1):43-76, 2020. 3" + }, + { + "type": "list", + "bbox": [ + 0.517, + 0.092, + 0.909, + 0.177 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.516, + 0.956 + ], + "angle": 0, + "content": "2354" + } + ] +] \ No newline at end of file diff --git a/2025/Distilling Spatially-Heterogeneous Distortion Perception for Blind Image Quality Assessment/c7a78868-95dc-4d35-97c7-04bbcfc1063d_origin.pdf b/2025/Distilling Spatially-Heterogeneous Distortion Perception for Blind Image Quality Assessment/c7a78868-95dc-4d35-97c7-04bbcfc1063d_origin.pdf new file mode 100644 index 0000000000000000000000000000000000000000..a77c7cc267234d672ed0de5cfca59af5087778b0 --- /dev/null +++ b/2025/Distilling Spatially-Heterogeneous Distortion Perception for Blind Image Quality Assessment/c7a78868-95dc-4d35-97c7-04bbcfc1063d_origin.pdf @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:a7d6f8b5b07ea474a3221fb16227db404e79716d2e857a7afc24ba030e6a3117 +size 2150465 diff --git a/2025/Distilling Spatially-Heterogeneous Distortion Perception for Blind Image Quality Assessment/full.md b/2025/Distilling Spatially-Heterogeneous Distortion Perception for Blind Image Quality Assessment/full.md new file mode 100644 index 0000000000000000000000000000000000000000..f6ad60fa441a0eae63b30752d6b2b9d55e82521e --- /dev/null +++ b/2025/Distilling Spatially-Heterogeneous Distortion Perception for Blind Image Quality Assessment/full.md @@ -0,0 +1,393 @@ +# Distilling Spatially-Heterogeneous Distortion Perception for Blind Image Quality Assessment + +Xudong Li $^{1}$ Wenjie Nie $^{1}$ Yan Zhang $^{1}$ Runze Hu $^{2}$ Ke Li $^{3}$ Xiawu Zheng $^{1}$ Liujuan Cao $^{1*}$ $^{1}$ Key Laboratory of Multimedia Trusted Perception and Efficient Computing, Ministry of Education of China, Xiamen University, 361005, P.R. China + $^{2}$ School of Information and Electronics, Beijing Institute of Technology, Beijing, China + $^{3}$ Tencent Youtu Lab, Shanghai, China + +{1xd761050753, bzhy986, hrzlpk2015}@gmail.com, {zhengxiwu, caoliujuan}@xmu.edu.cn + +# Abstract + +In the Blind Image Quality Assessment (BIQA) field, accurately assessing the quality of authentically distorted images presents a substantial challenge due to the diverse distortion types in natural settings. Existing state-of-the-art IQA methods mix a sequence of distortions into entire images to establish global distortion priors, but are inadequate for authentic images with spatially varied distortions. To address this, we introduce a novel IQA framework that employs knowledge distillation tailored to perceive spatially heterogeneous distortions, enhancing quality-distortion awareness. Specifically, we introduce a novel Block-wise Degradation Modelling approach that applies distinct distortions to different spatial blocks of an image, thereby expanding local distortion priors. Following this, we present a Block-wise Aggregation and Filtering module that enables fine-grained attention to the quality information within different distortion areas of the image. Furthermore, to effectively capture the complex relationships between distortions across different regions while preserving overall quality perception, we introduce Contrastive Knowledge Distillation to enhance the model's ability to discriminate between different types of distortions and Affinity Knowledge Distillation to model the correlation among distortions in different regions. Extensive experiments on standard BIQA datasets demonstrate the effectiveness and competitiveness of the proposed method. + +# 1. Introduction + +Image Quality Assessment (IQA) methods are designed to evaluate images in a way that aligns with the human judgment of visual distortions [47, 49]. These methods are crucial for applications that rely heavily on image quality + +![](images/af43d3273e4f04c39fb57ac9e8a44dd8d1ea153788daf1b6fb90d96c50f15752.jpg) +Figure 1. Comparison of state-of-the-art (SOTA) BIQA methods on the largest authentic dataset, LIEVFB. The proposed SHDIQA method outperforms all existing BIQA approaches. + +and serve as standards to gauge the performance of image processing techniques. Objective Image Quality Assessment includes Full-Reference IQA (FR-IQA) [36], Reduced-Reference IQA (RR-IQA) [41], and no-reference IQA (NR-IQA) or Blind IQA (BIQA) [60]. As the presence of various types of artifacts in real-world scenes and the unavailability of reference images, BIQA gains attention for tasks like image restoration [1] and super-resolution [6]. + +Recent advances in data-driven BIQA models, especially those based on deep neural networks, have led to impressive progress [37, 51]. However, these models still rely heavily on Mean Opinion Scores (MOS), which are expensive and time-consuming to collect—often requiring up to 120 human ratings per image [13]. As an alternative, synthetic distortions generated through controlled data augmentation offer a cost-effective way to create annotated datasets [48], enabling the learning of quality representations without extensive human labeling [38]. To leverage such synthetic data, recent studies have introduced multi-task end-to-end + +frameworks [26, 57] that jointly optimize distortion-specific objectives and quality prediction, using synthetic distortions to enrich training signals and improve generalization. Another line of work focuses on image degradation modeling. Methods like Re-IQA [31] and QPT [58] simulate realistic distortions by applying diverse types and combinations to entire images, as illustrated in Fig.2. They then leverage unsupervised learning frameworks such as MoCo-v2[3] to extract distortion priors from large-scale synthetic datasets. + +However, both categories of methods share a key limitation: they focus on global distortion priors while ignoring fine-grained local distortions, which are essential for accurate BIQA. Research [15, 17] shows that the human visual system (HVS) employs both local detection and global appearance strategies when assessing image quality. For subtle, near-threshold distortions, the HVS prioritizes local details, while for highly visible distortions, it shifts focus back to global content. This highlights the vital role of local distortion features in comprehensive image quality evaluation. + +In this paper, we propose a novel Spatial Heterogeneous Distortion-aware Image Quality Assessment (SHDIQA) method, designed to improve the local distortion-awareness of BIQA models based on the Vision Transformer (ViT) architecture. Unlike prior works [10, 14, 50] that focus on modifying ViT architectures to enhance local feature extraction, our method fully leverages rich local distortion priors generated from synthetic distortions, while introducing minimal architectural changes (only 0.2M additional parameters). + +Specifically, we introduce a Block-wise Image Degradation (BWID) method tailored for IQA tasks to generate patch-level distortion priors. As shown in Fig. 2, BWID partitions a distorted image into multiple blocks and applies a randomly selected distortion from 24 categories across 4 severity levels to each block. This results in a composite image with spatially varying local distortions. Each image naturally comes with an image-level multi-class label and corresponding patch-level single-class labels, effectively addressing the lack of localized distortion annotations in existing methods. To further utilize the patch-level distortion priors, we propose a Block-wise Aggregation and Filtering (BwAF) mechanism. BWAF uses the local priors from BWID to train multiple distortion (DIS) tokens, which interact with a global classification (CLS) token through a cross-attention mechanism with an adaptive head. This interaction mimics the HVS's dual strategy of local detection and global perception, thereby improving overall quality assessment, as illustrated in Fig. 1. Moreover, previous studies [16, 40] highlight the importance of modeling the relationships among different distortion types in perceptual quality assessment. To achieve this, we incorporate Contrastive Knowledge Distillation (CKD) and Affinity Knowledge Distillation (AKD). CKD aligns similar distortions and separates dissimilar ones, improving the model's ability to differentiate between distor + +![](images/39dcb6f888a06e9d61a88a1f100001fdbd5417b9d2a1a1438bb4f2dbaab49f47.jpg) + +![](images/837659c2d179a6b916b9df84f03f7098ce6b8bf1298f04997550a949545fdae1.jpg) +Distorted Region + +![](images/4df4db78557de28709340952da50a92c7ec1a028ff4ede992198c7044b2c3091.jpg) + +![](images/27bc660bbab2540866e20536f78f10adb35e9726efa7abb17bd9b31e495bf5f6.jpg) + +![](images/e5fb6976fdeab7ec32586b2927b37b562015885be505b69cd2c42f00de732c51.jpg) + +![](images/18b4a0e21421b0d1ce2d31e31f3b771f5e19cea4b474ccffb54275e536104430.jpg) +Image B +Figure 2. The proposed SHDIQA method creates a heterogeneous degradation space with richer priors (Top) and assigns dedicated distortion-aware (DIS) tokens to each local region (Bottom). + +![](images/39519b5f774f3b6d3c4efdbd7cf977780d50ece87c408c52e93f5d4064e6ee76.jpg) + +![](images/8d6022ea40d2f18ae1b475560218b9c176efc22bae8fc91e226020088fb41f77.jpg) +DEIQT Method + +![](images/7a4fa361f84280d044140b0a0c07cb7245517ddbfa3aa42163a9a480442c0d5b.jpg) +SHDIOA(Ours) + +![](images/eec06aa26b83c6b2f3ed31be17489cee36b06e20833592f33afe7db30475a55f.jpg) + +tion types. Meanwhile, AKD transfers knowledge about the relationships between different distortions, enhancing the model's generalization across various types of distortions. + +Our contributions are summarized as follows: + +- We present a Block-wise Degradation Modelling approach that enriches the local distortion priors of the training datasets by dividing images into blocks and applying distinct distortions to each, enhancing the model's ability to assess image quality accurately in complex scenarios. +- We introduce the Block-wise Aggregation and Filtering mechanism to fully capture and leverage localized distortions. It first directs precise attention to distortion details within image blocks, and then utilizes an adaptation head to filter out irrelevant distortion features, focusing the analysis on quality-aware information. +- We propose a novel Distortion-Guided Knowledge Distillation strategy that combines Contrastive Knowledge Distillation (CKD) to enhance distortion discrimination and Affinity Knowledge Distillation (AKD) to enrich relational insights between local distortion areas. + +# 2. Related Work + +# 2.1. BIQA with Degradation Modelling + +Recent studies [10, 61] have shown that simple distortion simulations do not fully capture the complexities of real-world degradations. To address this gap, Image Degradation Modeling has been introduced to represent various distortion types [34, 45, 46]. DBCNN [56] decomposes real distortions into multiple synthetic ones, creating large-scale synthetic images for training. DipIQ [25] categorizes common distortions like noise, blur, and compression into severity levels, while Re-IQA [31] introduces 25 distortion types to expand the degradation space. QPT [58] models second-order degradation using skip and shuffle operations across nine types in three groups. In contrast to these global degradation approaches, our method applies localized distortions to different regions of the image, enabling the model to learn fine-grained distortion perception both within and across blocks through knowledge distillation. + +![](images/2a04a5c8e746f1fed82b421526c8abe6b8c9226a75bb2d8c09b5c310e99bc42e.jpg) + +![](images/10ca39a192719452500ef9bfbddc705bf801eb2bf8949e7d2390ab00860304f1.jpg) +(a) Block-wise Degradation + +![](images/b6f136828165943658dee6db4e344ed6a83283f4bffe6bdb685065cff5e22d3b.jpg) +Figure 3. Framework Overview. Given an input image $\mathbf{I}$ , (a) Block-wise Degradation is first applied to generate a degraded version $\mathbf{I}_d$ (Sec. 3.1). Both $\mathbf{I}$ and $\mathbf{I}_d$ are then fed into a student encoder-decoder network, which is initialized with multiple distortion (DIS) tokens (each attending to a specific local block) and a CLS token for global quality representation. (b) Guided by the Block-wise Aggregation (Sec. 3.2), the DIS tokens and the CLS token are used to extract local distortion features $\mathcal{F}_{dis}$ and a global quality feature $\mathcal{F}_{cls}$ , which are further refined by the decoder into $\hat{\mathcal{F}}_{dis}$ and $\hat{\mathcal{F}}_{cls}$ . These refined features are processed by a quality-aware adaptation head to suppress irrelevant information and predict the quality score via an MLP. (c) To further enhance distortion perception, Contrastive KD is used to improve inter-block distortion discrimination, while Affinity KD captures relationships between distortions across different regions (Sec. 3.3). + +# 2.2. BIQA with Transformers + +Vision Transformer (ViT)-based IQA architectures include hybrid transformers [2, 10, 21, 50, 54] and pure transformers [14, 20, 29]. Hybrid models integrate multi-scale CNN information for better local distortion perception but increase complexity. Pure transformers like DEIQT [29] capture global abstraction using the CLS token but struggle with local structure modeling [63]. To reduce memory overhead, methods using local self-attention and sparse block structures have been developed [23, 30], though at the cost of some long-distance modeling. Sampling-based methods like VT-IQA [17] combine global and local information by resampling distorted images and using global attention. However, they lack explicit supervision for local distortions, which limits their ability to capture localized degradations during training. In contrast, our approach enhances local distortion perception in pure ViT by incorporating patch-level distortion priors, while preserving global context modeling through a mask matrix specifically applied to distortion tokens, leaving the self-attention process unaffected. + +# 2.3. IQA with Knowledge Distillation + +Knowledge Distillation (KD), originally proposed for model compression [12], has been adapted for tasks like action recognition [8], re-identification [28], and depth estimation [11]. Recently, KD has been used in methods like Degraded-Reference IQA (DR-IQA) [59] and Non-aligned Reference IQA (NAR-IQA) [52] to transfer reference information, though these still partially rely on reference images. + +A related BIQA method is SSLIQA [55], which employs KD to enforce semantic consistency across different triples within a batch. However, our approach differs significantly in that our KD explicitly targets inter-block and intra-block distortion information, allowing for more direct extraction of quality-related distortion priors from the augmented image. + +# 3. Methodology + +We introduce Spatial-Heterogeneous Distortion-aware Image Quality Assessment (SHDIQA), a framework designed to enhance the perception of local and non-uniform distortions in authentic images, as illustrated in Fig. 3. Specifically, the proposed SHDIQA first divides the input image $I$ into vertical blocks and applies random distortions using a Block-wise Degradation module (Sec. 3.1), producing a degraded image $I_d$ . Both $I$ and $I_d$ are then processed by a student encoder-decoder network equipped with a Blockwise Aggregation module (Sec. 3.2), which extracts local distortion features $\mathcal{F}_{dis}$ via multiple distortion (DIS) tokens and global quality feature $\mathcal{F}_{cls}$ via classification (CLS) token. Next, the local and global features $\mathcal{F}_{dis}$ and $\mathcal{F}_{cls}$ are fused via a cross-attention mechanism in the adaptation head to suppress irrelevant distortion information (Sec. 3.2). To enhance the model's ability to represent and distinguish distortion patterns, we further introduce Contrastive Knowledge Distillation (CKD) and Affinity Knowledge Distillation (AKD) (Sec. 3.3). These modules jointly improve distortion discriminability and capture inter-block distortion correlations, enabling a more distortion-aware representation. + +# 3.1. Block-wise Degradation Modelling + +Recent methods [31, 58] apply multiple distortions uniformly across the entire image, ignoring the finer details of local distortions. To address this, we propose a Block-wise Degradation Modelling strategy, which applies 24 distortion types and 4 degradation levels to different regions of the image, thereby generating a composite local distortion prior. Specifically, given an image of size $W \times H$ , we divide it into $K$ vertical blocks of size $W \times \frac{H}{K}$ . For each block $k$ ( $k = 1, \dots, K$ ), we randomly select a rectangular region $C_k$ with top-left coordinate $\alpha_i$ and size $w_k \times h_k$ , where $w_k = \beta_w \times W$ and $h_k = \beta_h \times \frac{H}{K}$ . Empirically, we find that setting $\beta_w = \beta_h = 1$ , i.e., degrading the entire block, yields the best performance. Then, the degradation is applied using the function $\mathcal{D}_k(C_k; q_k, p_k)$ , where $q_k$ denotes the type of distortion (e.g., noise, darkening), and $p_k$ represents its intensity. The degraded image $I_d$ is generated as follows: + +$$ +\boldsymbol {I} _ {d} = \mathcal {A} (\boldsymbol {I}) = \mathcal {D} _ {K} \left(\dots \mathcal {D} _ {1} \left(\boldsymbol {C} _ {1}; q _ {1}, p _ {1}\right) \dots ; q _ {K}, p _ {K}\right) \tag {1} +$$ + +where $K = 4$ denotes the number of degraded blocks. Unlike previous methods, our approach focuses on localized distortions, providing richer priors for feature aggregation (Sec. 3.2) and knowledge distillation (Sec. 3.3), improving the model's ability to handle complex real-world distortions. + +# 3.2. Block-wise Aggregation and Filtering + +In this section, we introduce the Block-wise Aggregation method to extract the local distortion feature $\mathcal{F}_{dis}$ from the degraded image $I_{d}$ and the input image $\pmb{I}$ . We then present the Quality-aware Adaptation Head, which filters out quality-irrelevant distortion features from distortion feature $\mathcal{F}_{dis}$ . + +# 3.2.1. Block-wise Aggregation + +We begin by randomly initializing $K$ local distortion-aware (DIS) tokens, denoted as $\mathcal{F}_{dis}$ , matching the number of degraded regions. Each DIS token $\mathcal{F}_{dis}^{k}$ corresponds to a specific region $C_k$ . The resulting embedding sequence is defined as $\mathcal{F} = \{\mathcal{F}_{dis};[\mathcal{F}_{cls};\mathcal{F}_{patch}]\} \in \mathbb{R}^{(K + N)\times D}$ , where $N$ is the number of image patches plus one CLS token, and $D$ is the embedding dimension. To ensure each DIS token focuses on its associated distorted region $C_k$ , we construct a mask map $\boldsymbol{m} \in \mathbb{R}^{(K + N)}$ . Based on this, we define the attention mask $\mathcal{M}$ to preserve the original global attention structure while enforcing localized attention for DIS tokens as follows: + +$$ +m _ {j} ^ {i} = \left\{ \begin{array}{l l} \mathbb {1} \{\text {t h e i - t h p a t c h o v e r l a p s w i t h C _ {i}} \}, & j < N \\ 0, & j = N \end{array} \right., +$$ + +$$ +\mathbb {M} = \left[ m ^ {1}; m ^ {2}; \dots ; m ^ {K} \right] ^ {T}, \mathcal {M} = \left[ \begin{array}{c c} \mathbf {1} _ {N \times N} & \mathbf {0} _ {N \times K} \\ \mathbb {M} _ {K \times N} & \mathbf {1} _ {K \times K} \end{array} \right], \tag {2} +$$ + +where $\mathbb{1}\{\cdot\}$ is the indicator function. Intuitively, Eq. 2 enforces that each DIS token $\mathcal{F}_{dis}^{k}$ attends only to the distorted + +content in block $C_k$ . Assuming the encoder has $H$ layers, the Multi-Head Self-Attention is applied as follows: + +$$ +\boldsymbol {\mathcal {F}} ^ {h} = \left\{ \begin{array}{l l} \sigma (\log \mathcal {M} + \mathbf {Q} ^ {h} (\mathbf {K} ^ {h}) ^ {T}) \mathbf {V} ^ {h} + \boldsymbol {\mathcal {F}} ^ {h - 1}, & 0 < h \leq H \\ \boldsymbol {\mathcal {F}}, & h = 0 \end{array} , \right. \tag {3} +$$ + +where $Q^h, K^h, V^h$ are the query, key, and value matrices, obtained via linear projections of $\mathcal{F}^{h-1}$ , and $\sigma$ denotes the Softmax function. The encoder outputs features $\mathcal{F}^H$ after $H$ layers. Following [29], we apply a transformer decoder to refine these features, yielding $\{\hat{\mathcal{F}}_{dis}; \hat{\mathcal{F}}_{cls}\} \in \mathbb{R}^{(K+1) \times D}$ , which are more relevant for image quality. + +# 3.2.2. Quality-aware Adaptation Head + +To effectively integrate global quality and local distortion features while suppressing quality-irrelevant information, we design a quality-aware adaptation head that highlights distortion cues closely related to perceived quality. Specifically, the global quality feature $\hat{\mathcal{F}}_{cls}$ is used as the query to selectively attend to the local distortion features $\hat{\mathcal{F}}_{dis}$ , which serve as key and value. This allows the model to emphasize distortion patterns most relevant to the overall quality. Overall, the final predicted quality score $\hat{Y}$ is computed as: + +$$ +\hat {Y} = \operatorname {L N} \left(\hat {\mathcal {F}} _ {c l s} + \text {C r o s s A t t e n t i o n} \left(\hat {\mathcal {F}} _ {c l s}, \hat {\mathcal {F}} _ {d i s}\right)\right), \tag {4} +$$ + +where $\text{CrossAttention}(\cdot)$ denotes a cross-attention module, and $\text{LN}$ is a single-layer linear projection. During training, the model is optimized by minimizing the smooth $\mathcal{L}_1$ loss. + +# 3.3. Distortion-Guided Knowledge Distillation + +Inspired by [5, 18, 33], we adopt a knowledge distillation method to learn correlations among distortion types by transferring quality-aware distortion knowledge from the teacher to the student. This distillation strategy addresses two key challenges: First, the student's ViT encoders, pre-trained on ImageNet [4], focus on semantics and lack distortion awareness in early training [19]. The teacher, pre-trained on IQA data, provides quality-aware guidance, directing the student to critical distortion patterns. Second, the absence of explicit ground truth for distortion correlations makes it difficult for the student to model these relationships independently. The teacher acts as a reliable target for correlation modeling, thereby improving the student's discriminative ability. + +# 3.3.1. Contrastive Knowledge Distillation + +Contrastive Knowledge Distillation (CKD) aims to extract the distortion features $\pmb{V}_{dis}^{k}$ of the $k$ -th local block from a teacher encoder $\mathcal{V}(\cdot)$ and transfer them to a student network. Simply maximizing feature similarity is insufficient to capture distinct distortion representations across spatial domains (Table 5). To address this, we reformulate the KD task as a pseudo-classification problem. As shown in Fig. 3, features from corresponding distorted regions of the teacher + +
MethodLIVECSIQKADIDLIVECKonIQLIVEFBSPAQ
PLCCSRCCPLCCSRCCPLCCSRCCPLCCSRCCPLCCSRCCPLCCSRCCPLCCSRCC
BRISQUE [27]0.9440.9290.7480.8120.5670.5280.6290.6290.6850.6810.3410.3030.8170.809
DBCNN [56]0.9710.9680.9590.9460.8560.8510.8690.8510.8840.8750.5510.5450.9150.911
TIQA [54]0.9650.9490.8380.8250.8550.850.8610.8450.9030.8920.5810.541--
MetaIQA [61]0.9590.9600.9080.8990.7750.7620.8020.8350.8560.8870.5070.54--
HyperIQA (27M) [39]0.9660.9620.9420.9230.8450.8520.8820.8590.9170.9060.6020.5440.9150.911
TReS (152M) [10]0.9680.9690.9420.9220.8580.8590.8770.8460.9280.9150.6250.554--
MUSIQ (27M) [14]0.9110.9400.8930.8710.8720.8750.7460.7020.9280.9160.6610.5660.9210.918
Re-IQA (48M) [31]0.9710.9700.9600.9470.8850.8720.8540.8400.9230.914--0.9250.918
DEIQT (24M) [29]0.9820.9800.9630.9460.8870.8890.8940.8750.9340.9210.6630.5710.9230.919
QPT (24M) [58]------0.9140.8950.9410.9270.6750.5750.9280.925
LIQE (151M) [57]0.9510.9700.9390.9360.9310.9300.9100.9040.9080.919----
QFM-IQM (24M) [19]0.9830.9810.9650.9540.9060.9060.9130.8910.9360.9220.6670.5670.9240.920
LoDa (118M) [50]0.9790.975--0.9360.9310.8990.8760.9440.9320.6790.5780.9280.925
SHDIQA (24.2M)0.9840.9820.9720.9620.9400.9370.9220.9090.9480.9370.7350.6390.9340.930
+ +Table 1. Performance comparison measured by the average of SRCC and PLCC. Best results are in bold, second-best are underlined. + +and student are treated as positive pairs, while all other pairs are considered negative. The CKD loss is defined as: + +$$ +\begin{array}{l} \mathcal {L} _ {c k d} = - \frac {1}{K} \sum_ {k} \log \left(\sigma \left(\frac {\cos \left(\boldsymbol {\mathcal {F}} _ {d i s} ^ {k} , \boldsymbol {V} _ {d i s} ^ {k}\right)}{\tau_ {c}}\right)\right) \\ \left. - \frac{1}{K}\sum_{\substack{k,j\\ j\neq k}}\log \left(1 - \sigma \left(\frac{\cos(\boldsymbol{\mathcal{F}}_{dis}^{k},\boldsymbol{V}_{dis}^{j})}{\tau_{c}}\right)\right)\right), \tag{5} \\ \end{array} +$$ + +where $\cos (\cdot)$ is cosine similarity, $\sigma$ is the sigmoid function, and $\tau_{c} = 0.3$ is the temperature factor. $\mathcal{F}_{dis}^{k}$ is the student feature vector of the $k$ -th local block feature extracted by the DIS token. This approach enhances contrast among features of varying distortion types, improving the model's discrimination and distortion-awareness capabilities. + +# 3.3.2. Affinity Knowledge Distillation + +Affinity Knowledge Distillation (AKD) focuses on distilling the interrelations among multiple blocks within an image $\pmb{I}$ or its degraded version $I_{d}$ . As shown in Fig. 3, given teacher embeddings $\mathbf{V}_{dis}$ and student features $\mathcal{F}_{dis}$ , we compute student and teachers' pairwise affinity matrices: + +$$ +\boldsymbol {R} _ {s t u} = \bar {\boldsymbol {F}} \bar {\boldsymbol {F}} ^ {T}, \quad \boldsymbol {R} _ {t e a} = \bar {\boldsymbol {V}} \bar {\boldsymbol {V}} ^ {T}, \tag {6} +$$ + +where $\bar{F}$ and $\bar{V}$ are $\mathcal{L}_2$ -normalized student $\mathcal{F}_{dis}$ and teacher features $V_{dis}$ , respectively. These matrices capture pairwise correlations, and the goal is to align the student affinity matrix $R_{stu}$ with the teacher matrix $R_{tea}$ : + +$$ +\mathcal {L} _ {a k d} = - \frac {1}{K} \sum_ {k = 1} ^ {K} \mathrm {K L} \left(\sigma \left(\frac {\boldsymbol {R} _ {s t u} ^ {k}}{\tau_ {a}}\right) \| \sigma \left(\frac {\boldsymbol {R} _ {t e a} ^ {k}}{\tau_ {a}}\right)\right), \tag {7} +$$ + +where $\tau_{a} = 2$ and $\mathrm{KL}(\cdot)$ denotes the Kullback-Leibler divergence. AKD complements CKD by focusing on the relational dynamics between image blocks, allowing the model to learn the relationship between distortion. However, relying solely on AKD may overlook fine-grained distortion-discriminative details, as highlighted in Table 2. Therefore, both CKD and AKD are essential for learning comprehensive distortion-aware features. + +# 3.3.3. Overall Training Loss + +The overall loss combines four complementary components: $\mathcal{L}_{\mathrm{cls}}$ provides category-level supervision for distortion identification but lacks awareness of relationships between distortion types; $\mathcal{L}_{\mathrm{ckd}}$ enhances feature discrimination by clustering similar distortions and separating dissimilar ones, thus compensating for this limitation; $\mathcal{L}_{\mathrm{akd}}$ further captures global perceptual structures by modeling relative similarities across different distortion. and $\mathcal{L}_{\log}$ regularizes training via teacher outputs [44], mitigating overfitting to quality-irrelevant patterns. Finally, the total loss is defined as: + +$$ +\mathcal {L} = \lambda_ {\mathrm {c l s}} \mathcal {L} _ {\mathrm {c l s}} + \lambda_ {\mathrm {c k d}} \mathcal {L} _ {\mathrm {c k d}} + \lambda_ {\mathrm {a k d}} \mathcal {L} _ {\mathrm {a k d}} + \lambda_ {\log} \mathcal {L} _ {\log} + \| \hat {Y} - Y _ {g} \| _ {1}, \tag {8} +$$ + +where $Y_{g}$ is the ground truth, $\mathcal{L}_{\mathrm{log}}$ is the $\mathcal{L}_1$ loss between student and teacher predictions, and $\lambda_{\mathrm{cls}}$ , $\lambda_{\mathrm{ckd}}$ , and $\lambda_{\mathrm{akd}}$ are set to 1, with $\lambda_{\mathrm{log}}$ set to 0.8. + +# 4. Experiments + +# 4.1. Benchmark Datasets and Evaluation Protocols + +Experiments were conducted on 7 typical BIQA datasets, including 3 synthetic datasets LIVE [35], CSIQ [15], KA-DID [22], and 4 authentic datasets LIVEC [9], KonIQ [13], LIVEFB [53], SPAQ [7]. Specifically, for authentic datasets, + +
Loss LogLoss LclsLoss LakdLoss LckdCSIQKADIDLIVELIVECAvg.
PLCCSRCCPLCCSRCCPLCCSRCCPLCCSRCC
XXXX0.9520.9350.8780.8840.9660.9640.8910.8670.917
✓XXX0.9550.9410.8980.9040.9720.9700.8980.8720.926
✓✓XX0.9590.9470.9190.9160.9740.9720.9040.8900.935
✓X✓X0.9700.9600.9270.9250.9800.9780.9190.9050.945
✓XX✓0.9660.9550.9130.9110.9760.9740.9140.9020.939
✓X✓✓0.9690.9580.9370.9350.9830.9810.9200.9040.948
✓✓✓✓0.9720.9620.9400.9370.9840.9820.9220.9090.951
+ +Table 2. Ablation experiments about knowledge distillation components on multiple datasets, where bold entries indicate the best result. + +
TrainingLIVEFBLIVECKonIQLIVECSIQ
TestingKonIQLIVECKonIQLIVECCSIQLIVE
DBCNN0.7160.7240.7540.7550.7580.877
P2P-BM0.7550.7380.7400.7700.712-
TReS0.7130.7400.7330.7860.761-
DEIQT0.7330.7810.7440.7940.7810.932
LoDa0.7630.8050.7450.811--
SHDIQA0.7230.7910.7820.8190.7890.943
+ +LIVEC contains 1162 images from different mobile devices and photographers. SPAQ includes 11125 photos from 66 smartphones. KonIQ contains 10073 images from public sources, while LIVEFB is the largest real-world dataset with 39810 images. The synthetic datasets involve artificially distorting original images using methods such as JPEG compression and Gaussian blur. LIVE and CSIQ respectively have 779 and 866 synthetic distortion images, with 5 and 6 types of distortions. KADID contains 10125 synthetic distortion images, spanning 25 types of distortions. + +Evaluation Criteria. The experiments use Pearson's Rank-Order Correlation Coefficient (SRCC) and Pearson's Linear Correlation Coefficient (PLCC) to assess prediction monotonicity and accuracy, respectively. Both metrics scale from -1 to 1, with values close to 1 indicating better performance. + +# 4.2. Implementation Details + +For the student network, we follow the typical training strategy [29], using a Transformer encoder based on ViT-S from DeiT III [43] and a one-layer decoder. Our model is trained for 9 epochs with a learning rate of $2 \times 10^{-4}$ , decaying by a factor of 10 every 3 epochs. The batch size is 16 for LIVEC and 128 for KonIQ. We use $80\%$ of the images for training and $20\%$ for testing, repeating the process 10 times to reduce performance bias, reporting the average SRCC and PLCC. For synthetic distortion datasets, training and testing sets are divided by reference images to ensure con + +Table 3. SRCC on the cross datasets validation. The best results are highlighted in bold, second-best are underlined. + +
ModeMethodsLIVECKonIQ
PLCCSRCCPLCCSRCC
20%DEIQT0.8220.7920.9080.888
LoDa0.8540.8150.9230.907
SHDIQA0.8660.8400.9280.915
40%DEIQT0.8550.8380.9220.903
LoDa0.8790.8490.9350.922
SHDIQA0.8890.8690.9420.929
+ +Table 4. Data-efficient learning validation with the training set containing $20\%$ and $40\%$ of the images. + +tent independence. To provide reliable distortion perception early in training, the teacher model adopts a similar encoder-decoder architecture aligned with the student model, and is pre-trained with $\mathcal{L}_1$ loss on the same IQA dataset. Training is performed on four NVIDIA RTX 3090 GPUs. + +# 4.3. Overall Prediction Performance Comparison + +Table 1 reports the comparison results between the proposed SHDIQA and other state-of-the-art (SOTA) BIQA methods, including ViT-based IQA methods, i.e., LoDa [50] and DEIQT [29], and Degradation Modelling-based methods, i.e., Re-IQA [31] and QPT [58]. It is observed from these seven datasets that SHDIQA achieves superior performance over all other methods across the 7 datasets. Notably, on the most difficult LIVEFB dataset, SHDIQA achieves an impressive performance improvement. This enhancement can be attributed to the prevalence of local distortions [53] in the LIVEFB dataset, highlighting the importance of effectively utilizing local distortion priors. + +# 4.4. Data-Efficient Learning Validation + +We showcase the data efficiency of our approach, as presented in Table 4. Following the efficient validation protocol outlined in LoDa [50] and DEIQT [29], we train models using $20\%$ and $40\%$ of the available images. This process is repeated 10 times, and we report the average SRCC and PLCC scores. By introducing distortion priors through the + +![](images/83a2df484a08816d56a9f482fda85550c2b6b0d455ef92fe04d66bbb9f711221.jpg) +Figure 4. gMAD competition results between DEIQT and our method SHDIQA. (a) Fixed DEIQT at low quality. (b) Fixed DEIQT at high quality. (c) Fixed SHDIQA at low quality. (d) Fixed SHDIQA at high quality. (e) Cosine similarity between different DIS and CLS tokens. + +
LckdLIVECCSIQAvg.
PLCCSRCCPLCCSRCC
-0.8910.8670.9520.9350.911
L10.9080.8950.9570.9460.926
BCE0.9140.9020.9660.9550.934
+ +Block-wise Degradation, our method achieves performance that is nearly with SOTA results, even when trained on only $20\%$ of the dataset. Furthermore, our approach consistently outperforms all other methods on both datasets, demonstrating its superiority over existing efficient BIQA techniques. + +# 4.5. Generalization Capability Validation + +We evaluate SHDIQA's generalization ability using cross-dataset validation, training the BIQA model on one dataset and testing it on others. Table 3 shows the SRCC averages for five datasets, where SHDIQA achieves the best performance in four out of six cross-dataset evaluations. To further test our framework's generalization, we trained the model on the LIVE database and tested it on the Waterloo database using the gMAD competition [24]. In the gMAD competition, image pairs are selected where the attacker model predicts a large quality difference, while the defender perceives them as similar. Observers then evaluate these pairs to assess the model's generalization ability. As shown in Fig. 4, our model correctly selects image pairs with significant quality differences when attacking DEIQT, where DEIQT misclassifies pairs like (a) as low quality and (b) as high quality, contrary to human perception and SHDIQA's predictions. As a defender, SHDIQA identifies pairs with minimal perceptual variation, correctly classifying image pairs like (c) and (d) as both low and high quality, respectively, while DEIQT fails. This demonstrates the strong generalization capability of SHDIQA in handling real-world image distortions. + +Table 5. Ablation study on Contrastive Knowledge Distillation. + +
\( {\mathcal{L}}_{\text{akd }}^{I} \)\( {\mathcal{L}}_{\text{akd }}^{I_d} \)LIVECLIVEAvg.
PLCCSRCCPLCCSRCC
✓0.9170.9010.9820.9780.944
✓0.9150.9000.9800.9800.943
✓✓0.9220.9090.9840.9820.949
+ +Table 6. Ablation experiment on Affinity Knowledge Distillation. + +# 4.6. Ablation Study + +# 4.6.1. Effectiveness of Proposed Components + +We conduct an ablation study on components of our framework. Table 2 shows that using the teacher's quality score as a regularization improves model performance. Adding a distortion classification task boosts SRCC by 1.2 and 1.8 points on KADID and LIVEC, respectively, highlighting the effectiveness of distortion-aware supervision. However, identifying distortions without modeling their relationships provides limited benefits. To address this, we introduce CKD and AKD to capture distortion correlations, improving SRCC by up to 3.5 and 4.9 points on KADID, respectively. The limited gain from CKD suggests that aggressively separating regions with different distortions may be problematic, as these regions can have similar quality. AKD mitigates this issue by capturing structural relationships between image blocks. When combined, CKD and AKD improve the baseline by 6.2 points, confirming their complementarity. Finally, Table 8 shows that block-wise aggregation and filtering further enhance the model's ability to assess local distortions, leading to better quality prediction. + +# 4.6.2. Loss type of CKD and Implementation of AKD + +Table 5 shows the effect of different CKD loss functions. Using $\mathcal{L}_1$ loss improves average quality prediction by 1.5 points, and switching to a BCE-based contrastive loss adds another 0.8-point gain. This confirms the effectiveness of our CKD design in promoting feature clustering within the + +
NumberLIVEC
PLCCSRCC
30.9200.903
40.9220.909
50.9180.905
+ +Table 7. Ablation study on the number of degenerated blocks. + +
MethodLIVECLIVE
PLCCSRCCPLCCSRCC
Baseline0.8910.8670.9660.964
+BWA0.9180.9020.9820.978
+BWF0.9220.9090.9840.982
+ +Table 8. Ablation study on the Block-wise Aggregation (BwA) and Filtering (BWF) module. + +
MethodLIVECKonIQ
PLCCSRCCPLCCSRCC
Homogenous0.9070.8930.9390.930
Heterogenous0.9220.9090.9480.937
+ +Table 9. Ablation experiments on heterogeneous and homogeneous distortion on the whole image. + +![](images/cd911aaddb4ab589b11880f0ab647181721e8e23827f2105791ad7b8b47f309a.jpg) +Mos: 38.6 + +![](images/3b328d8794657a9a71287374a359e4f8815802dc3decb6d75b3c9e632ccfcbad.jpg) +Baseline: 46.1 + +![](images/f33fc90b237c0ac63aff54890d79b167d2ecb0c369f1a3993a18d131cce0d8c5.jpg) +Ours: 35.2(↓4.1) + +![](images/19b81108dfdef2c055ffbe724ec8a5a9cf7fcb30a6bf8e273748dcd2f818fcbc.jpg) +DIS token 1 + +![](images/67aad0ec5497634f6afcfcae966a8b4e18f1cc9f555ff233bb9a681fe2e01d58.jpg) +DIS token 2 + +![](images/34dc853aff7cdea3663b2bf65af3d1ec88113d548f78775f7b4365ded43ca1dc.jpg) +DIS token 3 + +![](images/e3acd117a948e850ba952908e4ab91173f16c23e6d21bcbff26f0d4f7815cf79.jpg) +DIS token 4 + +![](images/2bdf871e3b18832158a09fc313ee57f14d7518789f73347fa955e91eb6440c43.jpg) +Mos: 57.3 + +![](images/30cb86e4aa1ecbf27961745e0619307766e260c710dbef942164040c7ca8fb7f.jpg) +Baseline: 67.1 + +![](images/dff87b13cfc9008414b9985dd281b6746ebfd3412ad2109e3e0c1a4c87376d3b.jpg) +Ours: 60.9(↓6.2) + +![](images/35b5e774e7f7bd8e53024058fd9b264865bbb999d3fb44d8047d5dea73d11e94.jpg) +DIS token 1 +Figure 5. Activation maps of DEIQT [29] and SHDIQA using Grad-CAM [32] on the LIVEFB dataset, with MOS shown below the images. Our SHDIQA effectively captures the interplay between image content, local distortions, and overall quality, enabling accurate predictions. + +![](images/58a2b4a8db735cbd97cbebf58fecfd33662852ee34a4c23c330ee46bf666cff4.jpg) +Dis token 2 + +![](images/e3c74f4522d92d9b969f24f223326929dac7d111f09947406b4aca747367403b.jpg) +DIS token 3 + +![](images/abf21518b59b470a9d7d64cea86fedf7a5f0c06c197edcead7dcd849b04c50a8.jpg) +DIS token 4 + +same distortion type and separation across types. For AKD, we evaluate three strategies: applying it to the input image $\mathbf{I}$ , the degraded image $\mathbf{I}_d$ , or both. Table 6 shows that the best performance is achieved by using both, as applying AKD to the image $\mathbf{I}$ helps model relationships within local content, while applying it to the $\mathbf{I}_d$ captures correlations between different types of distortions. Combining both enables complementary learning of content and distortion information. + +# 4.6.3. Ablation on the Degradation Space + +We investigate the effectiveness of the Block-wise Degradation Modelling. As shown in Table 9, applying globally uniform heterogeneous distortions significantly underperforms compared to localized distortions, highlighting the importance of local degradation. We also evaluate how varying the number of degraded blocks affects performance. As shown in Table 7, the impact is relatively insensitive across different partitioning strategies, demonstrating the method's resilience to changes in block quantity. While more blocks increase the scope of degradation, they also lead to greater content loss. Thus, using four blocks strikes the best balance. + +# 4.7. Qualitative Analysis + +To show that tokens from different perspectives capture unique features, we analyze the cosine similarity between CLS and DIS tokens on the LIVEC and KonIQ datasets. As shown in Fig. 4, the similarities range from 0.72 to 0.81, lower than the 0.96 and 0.94 found in DeiT-T and DeiT-S [42]. Tokens near the center show higher similarity, + +aligning with the HVS's focus on salient regions [62]. This implies that DIS tokens capture localized distortions in a way that simulates human perception, and when combined with CLS tokens, it improves both global and local distortion awareness. To further explore the DIS tokens, we visualize the attention maps of DEIQT and our SHDIQA, as shown in Fig. 5. By incorporating four additional distortion tokens, the proposed model enhances its ability to detect distortions across different regions of the image. Through the interaction between distortion tokens and CLS tokens in the adaptation head, the model adaptively focuses on the most relevant areas for image quality, while also considering the overall content. This results in more accurate quality predictions. + +# 5. Conclusion + +In this paper, we propose SHDIQA, a novel BIQA approach that leverages spatial-heterogeneous distortion-aware strategies. Our Block-Wise Degradation Modelling generates locally distorted images, exposing the model to diverse distortion patterns. The Block-wise Aggregation and Filtering mechanism enhances the ViT model's sensitivity to local distortions, improving quality assessments through focused attention and feature filtering. Additionally, our Distortion-Guided Knowledge Distillation combines CKD and AKD to improve distortion discrimination and relational learning. Experimental results show SHDIQA's superior generalization and robustness across various datasets and distortion types, confirming its effectiveness in real-world BIQA tasks. + +# References + +[1] Mark R Banham and Aggelos K Katsaggelos. Digital image restoration. IEEE signal processing magazine, 14(2):24-41, 1997. 1 +[2] Chaofeng Chen, Jiadi Mo, Jingwen Hou, Haoning Wu, Liang Liao, Wenxiu Sun, Qiong Yan, and Weisi Lin. Topiq: A top-down approach from semantics to distortions for image quality assessment. IEEE Transactions on Image Processing, 2024. 3 +[3] Xinlei Chen, Haoqi Fan, Ross Girshick, and Kaiming He. Improved baselines with momentum contrastive learning. arXiv preprint arXiv:2003.04297, 2020. 2 +[4] Jia Deng, Wei Dong, Richard Socher, Li-Jia Li, Kai Li, and Li Fei-Fei. Imagenet: A large-scale hierarchical image database. In 2009 IEEE conference on computer vision and pattern recognition, pages 248-255. IEEE, 2009. 4 +[5] Jinhong Deng, Wen Li, Yuhua Chen, and Lixin Duan. Unbiased mean teacher for cross-domain object detection. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 4091-4101, 2021. 4 +[6] Chao Dong, Chen Change Loy, Kaiming He, and Xiaou Tang. Image super-resolution using deep convolutional networks. IEEE transactions on pattern analysis and machine intelligence, 38(2):295-307, 2015. 1 +[7] Yuming Fang, Hanwei Zhu, Yan Zeng, Kede Ma, and Zhou Wang. Perceptual quality assessment of smartphone photography. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 3677-3686, 2020. 5 +[8] Nuno C Garcia, Pietro Morerio, and Vittorio Murino. Modality distillation with multiple stream networks for action recognition. In Proceedings of the European Conference on Computer Vision (ECCV), pages 103-118, 2018. 3 +[9] Deepti Ghadiyaram and Alan C Bovik. Massive online crowdsourced study of subjective and objective picture quality. IEEE Transactions on Image Processing, 25(1):372-387, 2015. 5 +[10] S Alireza Golestaneh, Saba Dadsetan, and Kris M Kitani. No-reference image quality assessment via transformers, relative ranking, and self-consistency. In Proceedings of the IEEE/CVF Winter Conference on Applications of Computer Vision, pages 1220-1230, 2022. 2, 3, 5 +[11] Saurabh Gupta, Judy Hoffman, and Jitendra Malik. Cross modal distillation for supervision transfer. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 2827-2836, 2016. 3 + +Acknowledgements. This work was supported by the National Science Fund for Distinguished Young Scholars (No.62025603), the National Natural Science Foundation of China (No. U21B2037, No. U22B2051, No. U23A20383, No. 62176222, No. 62176223, No. 62176226, No. 62072386, No. 62072387, No. 62072389, No. 62002305 and No. 62272401), and the Natural Science Foundation of Fujian Province of China (No. 2021J06003, No.2022J06001). +[12] Geoffrey Hinton, Oriol Vinyals, and Jeff Dean. Distilling the knowledge in a neural network. arXiv preprint arXiv:1503.02531, 2015. 3 +[13] Vlad Hosu, Hanhe Lin, Tamas Sziranyi, and Dietmar Saupe. Koniq-10k: An ecologically valid database for deep learning of blind image quality assessment. IEEE Transactions on Image Processing, 29:4041-4056, 2020. 1, 5 +[14] Junjie Ke, Qifei Wang, Yilin Wang, Peyman Milanfar, and Feng Yang. Musiq: Multi-scale image quality transformer. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 5148-5157, 2021. 2, 3, 5 +[15] Eric C Larson and Damon M Chandler. Most apparent distortion: full-reference image quality assessment and the role of strategy. Journal of electronic imaging, 19(1):011006-011006, 2010. 2, 5 +[16] Aobo Li, Jinjian Wu, Yongxu Liu, and Leida Li. Bridging the synthetic-to Authentic gap: Distortion-guided unsupervised domain adaptation for blind image quality assessment. arXiv preprint arXiv:2405.04167, 2024. 2 +[17] Leida Li, Tianshu Song, Jinjian Wu, Weisheng Dong, Jiansheng Qian, and Guangming Shi. Blind image quality index for authentic distortions with local and global deep feature aggregation. IEEE Transactions on Circuits and Systems for Video Technology, 32(12):8512-8523, 2021. 2, 3 +[18] Liangqi Li, Jiaxu Miao, Dahu Shi, Wenming Tan, Ye Ren, Yi Yang, and Shiliang Pu. Distilling detr with visual-linguistic knowledge for open-vocabulary object detection. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 6501–6510, 2023. 4 +[19] Xudong Li, Timin Gao, Runze Hu, Yan Zhang, Shengchuan Zhang, Xiawu Zheng, Jingyuan Zheng, Yunhang Shen, Ke Li, Yutao Liu, et al. Adaptive feature selection for no-reference image quality assessment by mitigating semantic noise sensitivity. In International Conference on Machine Learning, pages 27808-27821. PMLR, 2024. 4, 5 +[20] Xudong Li, Runze Hu, Jingyuan Zheng, Yan Zhang, Shengchuan Zhang, Xiawu Zheng, Ke Li, Yunhang Shen, Yutao Liu, Pingyang Dai, et al. Integrating global context contrast and local sensitivity for blind image quality assessment. In International Conference on Machine Learning, pages 27920-27941. PMLR, 2024. 3 +[21] Xudong Li, Jingyuan Zheng, Runze Hu, Yan Zhang, Ke Li, Yunhang Shen, Xiawu Zheng, Yutao Liu, ShengChuan Zhang, Pingyang Dai, et al. Feature denoising diffusion model for blind image quality assessment. arXiv preprint arXiv:2401.11949, 2024.3 +[22] Hanhe Lin, Vlad Hosu, and Dietmar Saupe. Kadid-10k: A large-scale artificially distorted iqa database. In 2019 Eleventh International Conference on Quality of Multimedia Experience (QoMEX), pages 1-3. IEEE, 2019. 5 +[23] Hao Liu and Pieter Abbeel. Blockwise parallel transformers for large context models. Advances in Neural Information Processing Systems, 36, 2024. 3 +[24] Kede Ma, Qingbo Wu, Zhou Wang, Zhengfang Duanmu, Hongwei Yong, Hongliang Li, and Lei Zhang. Group mad competition-a new methodology to compare objective image quality models. In Proceedings of the IEEE Conference on + +Computer Vision and Pattern Recognition, pages 1664-1673, 2016. 7 +[25] Kede Ma, Wentao Liu, Tongliang Liu, Zhou Wang, and Dacheng Tao. dipiq: Blind image quality assessment by learning-to-rank discriminable image pairs. IEEE Transactions on Image Processing, 26(8):3951-3964, 2017. 2 +[26] Kede Ma, Wentao Liu, Kai Zhang, Zhengfang Duanmu, Zhou Wang, and Wangmeng Zuo. End-to-end blind image quality assessment using deep neural networks. IEEE Transactions on Image Processing, 27(3):1202-1213, 2017. 2 +[27] Anish Mittal, Anush Krishna Moorthy, and Alan Conrad Bovik. No-reference image quality assessment in the spatial domain. IEEE Transactions on image processing, 21(12): 4695-4708, 2012. 5 +[28] Angelo Porrello, Luca Bergamini, and Simone Calderara. Robust re-identification by multiple views knowledge distillation. In Computer Vision-ECCV 2020: 16th European Conference, Glasgow, UK, August 23–28, 2020, Proceedings, Part X 16, pages 93–110. Springer, 2020. 3 +[29] Guanyi Qin, Runze Hu, Yutao Liu, Xiawu Zheng, Haotian Liu, Xiu Li, and Yan Zhang. Data-efficient image quality assessment with attention-panel decoder. In Proceedings of the Thirty-Seventh AAAI Conference on Artificial Intelligence, 2023. 3, 4, 5, 6, 8 +[30] Jiezhong Qiu, Hao Ma, Omer Levy, Scott Wen-tau Yih, Sinong Wang, and Jie Tang. Blockwise self-attention for long document understanding. arXiv preprint arXiv:1911.02972, 2019. 3 +[31] Avinab Saha, Sandeep Mishra, and Alan C Bovik. Re-iqa: Unsupervised learning for image quality assessment in the wild. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 5846-5855, 2023. 2, 4, 5, 6 +[32] Ramprasaath R Selvaraju, Michael Cogswell, Abhishek Das, Ramakrishna Vedantam, Devi Parikh, and Dhruv Batra. Gradcam: Visual explanations from deep networks via gradient-based localization. In Proceedings of the IEEE international conference on computer vision, pages 618-626, 2017. 8 +[33] Tahira Shehzadi, Khurram Azeem Hashmi, Didier Stricker, and Muhammad Zeshan Afzal. Sparse semi-detr: Sparse learnable queries for semi-supervised object detection. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 5840-5850, 2024. 4 +[34] Hamid R Sheikh, Alan C Bovik, and Lawrence Cormack. No-reference quality assessment using natural scene statistics: Jpeg2000. IEEE Transactions on image processing, 14(11): 1918-1927, 2005. 2 +[35] Hamid R Sheikh, Muhammad F Sabir, and Alan C Bovik. A statistical evaluation of recent full reference image quality assessment algorithms. IEEE Transactions on image processing, 15(11):3440-3451, 2006. 5 +[36] Chenyang Shi and Yandan Lin. Full reference image quality assessment based on visual salience with color appearance and gradient similarity. IEEE Access, 8:97310-97320, 2020. 1 +[37] Tianshu Song, Leida Li, Deqiang Cheng, Pengfei Chen, and Jinjian Wu. Active learning-based sample selection for label- + +efficient blind image quality assessment. IEEE Transactions on Circuits and Systems for Video Technology, 2023. 1 +[38] Suhas Srinath, Shankhanil Mitra, Shika Rao, and Rajiv Soundararajan. Learning generalizable perceptual representations for data-efficient no-reference image quality assessment. In Proceedings of the IEEE/CVF Winter Conference on Applications of Computer Vision, pages 22-31, 2024. 1 +[39] Shaolin Su, Qingsen Yan, Yu Zhu, Cheng Zhang, Xin Ge, Jinqiu Sun, and Yanning Zhang. Blindly assess image quality in the wild guided by a self-adaptive hyper network. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 3667-3676, 2020. 5 +[40] Simeng Sun, Tao Yu, Jiahua Xu, Wei Zhou, and Zhibo Chen. Graphiqa: Learning distortion graph representations for blind image quality assessment. IEEE Transactions on Multimedia, 25:2912-2925, 2022. 2 +[41] Dacheng Tao, Xuelong Li, Wen Lu, and Xinbo Gao. Reduced-reference iqa in contourlet domain. IEEE Transactions on Systems, Man, and Cybernetics, Part B (Cybernetics), 39(6): 1623-1627, 2009. 1 +[42] Hugo Touvron, Matthieu Cord, Matthijs Douze, Francisco Massa, Alexandre Sablayrolles, and Hervé Jégou. Training data-efficient image transformers & distillation through attention. In International Conference on Machine Learning, pages 10347-10357. PMLR, 2021. 8 +[43] Hugo Touvron, Matthieu Cord, and Hervé Jégou. Deit iii: Revenge of the vit. arXiv preprint arXiv:2204.07118, 2022. 6 +[44] Tao Wang, Li Yuan, Xiaopeng Zhang, and Jiashi Feng. Distilling object detectors with fine-grained feature imitation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 4933-4942, 2019. 5 +[45] Zhou Wang and Eero Simoncelli. Local phase coherence and the perception of blur. Advances in neural information processing systems, 16, 2003. 2 +[46] Zhou Wang, Hamid R Sheikh, and Alan C Bovik. No-reference perceptual quality assessment of JPEG compressed images. In Proceedings. International conference on image processing, pages I-I. IEEE, 2002. 2 +[47] Zhou Wang, Alan C Bovik, Hamid R Sheikh, and Eero P Simoncelli. Image quality assessment: from error visibility to structural similarity. IEEE transactions on image processing, 13(4):600-612, 2004. 1 +[48] Zhihua Wang, Zhi-Ri Tang, Jianguo Zhang, and Yuming Fang. Toward a blind image quality evaluator in the wild by learning beyond human opinion scores. Pattern Recognition, 137:109296, 2023. 1 +[49] Haoning Wu, Zicheng Zhang, Erli Zhang, Chaofeng Chen, Liang Liao, Annan Wang, Chunyi Li, Wenxiu Sun, Qiong Yan, Guangtao Zhai, et al. Q-bench: A benchmark for general-purpose foundation models on low-level vision. arXiv preprint arXiv:2309.14181, 2023. 1 +[50] Kangmin Xu, Liang Liao, Jing Xiao, Chaofeng Chen, Haoning Wu, Qiong Yan, and Weisi Lin. Local distortion aware efficient transformer adaptation for image quality assessment. arXiv preprint arXiv:2308.12001, 2023. 2, 3, 5, 6 +[51] Sidi Yang, Tianhe Wu, Shuwei Shi, Shanshan Lao, Yuan Gong, Mingdeng Cao, Jiahao Wang, and Yujiu Yang. Maniqa: + +Multi-dimension attention network for no-reference image quality assessment. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 1191-1200, 2022. 1 +[52] Guanghao Yin, Wei Wang, Zehuan Yuan, Chuchu Han, Wei Ji, Shouqian Sun, and Changhu Wang. Content-variant reference image quality assessment via knowledge distillation. arXiv preprint arXiv:2202.13123, 2022. 3 +[53] Zhenqiang Ying, Haoran Niu, Praful Gupta, Dhruv Mahajan, Deepti Ghadiyaram, and Alan Bovik. From patches to pictures (paq-2-piq): Mapping the perceptual space of picture quality. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 3575–3585, 2020. 5, 6 +[54] Junyong You and Jari Korhonen. Transformer for image quality assessment. In 2021 IEEE International Conference on Image Processing (ICIP), pages 1389-1393. IEEE, 2021. 3, 5 +[55] Guanghui Yue, Di Cheng, Leida Li, Tianwei Zhou, Hantao Liu, and Tianfu Wang. Semi-supervised authentically distorted image quality assessment with consistency-preserving dual-branch convolutional neural network. IEEE Transactions on Multimedia, 25:6499–6511, 2022. 3 +[56] Weixia Zhang, Kede Ma, Jia Yan, Dexiang Deng, and Zhou Wang. Blind image quality assessment using a deep bilinear convolutional neural network. IEEE Transactions on Circuits and Systems for Video Technology, 30(1):36-47, 2018. 2, 5 +[57] Weixia Zhang, Guangtao Zhai, Ying Wei, Xiaokang Yang, and Kede Ma. Blind image quality assessment via vision-language correspondence: A multitask learning perspective. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 14071-14081, 2023. 2, 5 +[58] Kai Zhao, Kun Yuan, Ming Sun, Mading Li, and Xing Wen. Quality-aware pre-trained models for blind image quality assessment. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 22302-22313, 2023. 2, 4, 5, 6 +[59] Heliang Zheng, Huan Yang, Jianlong Fu, Zheng-Jun Zha, and Jiebo Luo. Learning conditional knowledge distillation for degraded-reference image quality assessment. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 10242-10251, 2021. 3 +[60] Yu Zhou, Leida Li, Shiqi Wang, Jinjian Wu, Yuming Fang, and Xinbo Gao. No-reference quality assessment for view synthesis using dog-based edge statistics and texture naturalness. IEEE Transactions on Image Processing, 28(9):4566-4579, 2019. 1 +[61] Hancheng Zhu, Leida Li, Jinjian Wu, Weisheng Dong, and Guangming Shi. Metaiaq: Deep meta-learning for no-reference image quality assessment. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 14143-14152, 2020. 2, 5 +[62] Mengmeng Zhu, Guanqun Hou, Xinjia Chen, Jiaxing Xie, Haixian Lu, and Jun Che. Saliency-guided transformer network combined with local embedding for no-reference image quality assessment. In Proceedings of the IEEE/CVF Inter- + +national Conference on Computer Vision, pages 1953-1962, 2021.8 +[63] Fuzhen Zhuang, Zhiyuan Qi, Keyu Duan, Dongbo Xi, Yongchun Zhu, Hengshu Zhu, Hui Xiong, and Qing He. A comprehensive survey on transfer learning. Proceedings of the IEEE, 109(1):43-76, 2020. 3 \ No newline at end of file diff --git a/2025/Distilling Spatially-Heterogeneous Distortion Perception for Blind Image Quality Assessment/images.zip b/2025/Distilling Spatially-Heterogeneous Distortion Perception for Blind Image Quality Assessment/images.zip new file mode 100644 index 0000000000000000000000000000000000000000..4bfd5ad3a3a754dddde8edcb8691ac289cd150a0 --- /dev/null +++ b/2025/Distilling Spatially-Heterogeneous Distortion Perception for Blind Image Quality Assessment/images.zip @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:35a046649a3aa059fd63fe3c4b886d20aecee6de0a09c1b11ace686b9f7b172c +size 938639 diff --git a/2025/Distilling Spatially-Heterogeneous Distortion Perception for Blind Image Quality Assessment/layout.json b/2025/Distilling Spatially-Heterogeneous Distortion Perception for Blind Image Quality Assessment/layout.json new file mode 100644 index 0000000000000000000000000000000000000000..b7121c3efe6a0a3734498a704b6b9df94a964d44 --- /dev/null +++ b/2025/Distilling Spatially-Heterogeneous Distortion Perception for Blind Image Quality Assessment/layout.json @@ -0,0 +1,11134 @@ +{ + "pdf_info": [ + { + "para_blocks": [ + { + "bbox": [ + 58, + 102, + 553, + 138 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 58, + 102, + 553, + 138 + ], + "spans": [ + { + "bbox": [ + 58, + 102, + 553, + 138 + ], + "type": "text", + "content": "Distilling Spatially-Heterogeneous Distortion Perception for Blind Image Quality Assessment" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 63, + 160, + 550, + 230 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 63, + 160, + 550, + 230 + ], + "spans": [ + { + "bbox": [ + 63, + 160, + 550, + 230 + ], + "type": "text", + "content": "Xudong Li" + }, + { + "bbox": [ + 63, + 160, + 550, + 230 + ], + "type": "inline_equation", + "content": "^{1}" + }, + { + "bbox": [ + 63, + 160, + 550, + 230 + ], + "type": "text", + "content": " Wenjie Nie" + }, + { + "bbox": [ + 63, + 160, + 550, + 230 + ], + "type": "inline_equation", + "content": "^{1}" + }, + { + "bbox": [ + 63, + 160, + 550, + 230 + ], + "type": "text", + "content": " Yan Zhang" + }, + { + "bbox": [ + 63, + 160, + 550, + 230 + ], + "type": "inline_equation", + "content": "^{1}" + }, + { + "bbox": [ + 63, + 160, + 550, + 230 + ], + "type": "text", + "content": " Runze Hu" + }, + { + "bbox": [ + 63, + 160, + 550, + 230 + ], + "type": "inline_equation", + "content": "^{2}" + }, + { + "bbox": [ + 63, + 160, + 550, + 230 + ], + "type": "text", + "content": " Ke Li" + }, + { + "bbox": [ + 63, + 160, + 550, + 230 + ], + "type": "inline_equation", + "content": "^{3}" + }, + { + "bbox": [ + 63, + 160, + 550, + 230 + ], + "type": "text", + "content": " Xiawu Zheng" + }, + { + "bbox": [ + 63, + 160, + 550, + 230 + ], + "type": "inline_equation", + "content": "^{1}" + }, + { + "bbox": [ + 63, + 160, + 550, + 230 + ], + "type": "text", + "content": " Liujuan Cao" + }, + { + "bbox": [ + 63, + 160, + 550, + 230 + ], + "type": "inline_equation", + "content": "^{1*}" + }, + { + "bbox": [ + 63, + 160, + 550, + 230 + ], + "type": "inline_equation", + "content": "^{1}" + }, + { + "bbox": [ + 63, + 160, + 550, + 230 + ], + "type": "text", + "content": " Key Laboratory of Multimedia Trusted Perception and Efficient Computing, Ministry of Education of China, Xiamen University, 361005, P.R. China \n" + }, + { + "bbox": [ + 63, + 160, + 550, + 230 + ], + "type": "inline_equation", + "content": "^{2}" + }, + { + "bbox": [ + 63, + 160, + 550, + 230 + ], + "type": "text", + "content": " School of Information and Electronics, Beijing Institute of Technology, Beijing, China \n" + }, + { + "bbox": [ + 63, + 160, + 550, + 230 + ], + "type": "inline_equation", + "content": "^{3}" + }, + { + "bbox": [ + 63, + 160, + 550, + 230 + ], + "type": "text", + "content": " Tencent Youtu Lab, Shanghai, China" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 80, + 233, + 523, + 244 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 80, + 233, + 523, + 244 + ], + "spans": [ + { + "bbox": [ + 80, + 233, + 523, + 244 + ], + "type": "text", + "content": "{1xd761050753, bzhy986, hrzlpk2015}@gmail.com, {zhengxiwu, caoliujuan}@xmu.edu.cn" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 151, + 273, + 200, + 285 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 151, + 273, + 200, + 285 + ], + "spans": [ + { + "bbox": [ + 151, + 273, + 200, + 285 + ], + "type": "text", + "content": "Abstract" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 53, + 298, + 297, + 598 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 298, + 297, + 598 + ], + "spans": [ + { + "bbox": [ + 53, + 298, + 297, + 598 + ], + "type": "text", + "content": "In the Blind Image Quality Assessment (BIQA) field, accurately assessing the quality of authentically distorted images presents a substantial challenge due to the diverse distortion types in natural settings. Existing state-of-the-art IQA methods mix a sequence of distortions into entire images to establish global distortion priors, but are inadequate for authentic images with spatially varied distortions. To address this, we introduce a novel IQA framework that employs knowledge distillation tailored to perceive spatially heterogeneous distortions, enhancing quality-distortion awareness. Specifically, we introduce a novel Block-wise Degradation Modelling approach that applies distinct distortions to different spatial blocks of an image, thereby expanding local distortion priors. Following this, we present a Block-wise Aggregation and Filtering module that enables fine-grained attention to the quality information within different distortion areas of the image. Furthermore, to effectively capture the complex relationships between distortions across different regions while preserving overall quality perception, we introduce Contrastive Knowledge Distillation to enhance the model's ability to discriminate between different types of distortions and Affinity Knowledge Distillation to model the correlation among distortions in different regions. Extensive experiments on standard BIQA datasets demonstrate the effectiveness and competitiveness of the proposed method." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 56, + 623, + 135, + 635 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 623, + 135, + 635 + ], + "spans": [ + { + "bbox": [ + 56, + 623, + 135, + 635 + ], + "type": "text", + "content": "1. Introduction" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 55, + 643, + 296, + 693 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 643, + 296, + 693 + ], + "spans": [ + { + "bbox": [ + 55, + 643, + 296, + 693 + ], + "type": "text", + "content": "Image Quality Assessment (IQA) methods are designed to evaluate images in a way that aligns with the human judgment of visual distortions [47, 49]. These methods are crucial for applications that rely heavily on image quality" + } + ] + } + ], + "index": 8 + }, + { + "type": "image", + "bbox": [ + 321, + 271, + 547, + 434 + ], + "blocks": [ + { + "bbox": [ + 321, + 271, + 547, + 434 + ], + "lines": [ + { + "bbox": [ + 321, + 271, + 547, + 434 + ], + "spans": [ + { + "bbox": [ + 321, + 271, + 547, + 434 + ], + "type": "image", + "image_path": "af43d3273e4f04c39fb57ac9e8a44dd8d1ea153788daf1b6fb90d96c50f15752.jpg" + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 313, + 441, + 555, + 475 + ], + "lines": [ + { + "bbox": [ + 313, + 441, + 555, + 475 + ], + "spans": [ + { + "bbox": [ + 313, + 441, + 555, + 475 + ], + "type": "text", + "content": "Figure 1. Comparison of state-of-the-art (SOTA) BIQA methods on the largest authentic dataset, LIEVFB. The proposed SHDIQA method outperforms all existing BIQA approaches." + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "image_caption" + } + ], + "index": 9 + }, + { + "bbox": [ + 313, + 486, + 555, + 581 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 486, + 555, + 581 + ], + "spans": [ + { + "bbox": [ + 313, + 486, + 555, + 581 + ], + "type": "text", + "content": "and serve as standards to gauge the performance of image processing techniques. Objective Image Quality Assessment includes Full-Reference IQA (FR-IQA) [36], Reduced-Reference IQA (RR-IQA) [41], and no-reference IQA (NR-IQA) or Blind IQA (BIQA) [60]. As the presence of various types of artifacts in real-world scenes and the unavailability of reference images, BIQA gains attention for tasks like image restoration [1] and super-resolution [6]." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 313, + 582, + 556, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 582, + 556, + 713 + ], + "spans": [ + { + "bbox": [ + 313, + 582, + 556, + 713 + ], + "type": "text", + "content": "Recent advances in data-driven BIQA models, especially those based on deep neural networks, have led to impressive progress [37, 51]. However, these models still rely heavily on Mean Opinion Scores (MOS), which are expensive and time-consuming to collect—often requiring up to 120 human ratings per image [13]. As an alternative, synthetic distortions generated through controlled data augmentation offer a cost-effective way to create annotated datasets [48], enabling the learning of quality representations without extensive human labeling [38]. To leverage such synthetic data, recent studies have introduced multi-task end-to-end" + } + ] + } + ], + "index": 12 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "spans": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "type": "text", + "content": "CVF" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 145, + 0, + 496, + 35 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 145, + 0, + 496, + 35 + ], + "spans": [ + { + "bbox": [ + 145, + 0, + 496, + 35 + ], + "type": "text", + "content": "This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 67, + 703, + 144, + 712 + ], + "type": "page_footnote", + "angle": 0, + "lines": [ + { + "bbox": [ + 67, + 703, + 144, + 712 + ], + "spans": [ + { + "bbox": [ + 67, + 703, + 144, + 712 + ], + "type": "text", + "content": "*Corresponding author." + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "text", + "content": "2344" + } + ] + } + ], + "index": 14 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 0 + }, + { + "para_blocks": [ + { + "bbox": [ + 55, + 72, + 296, + 180 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 72, + 296, + 180 + ], + "spans": [ + { + "bbox": [ + 55, + 72, + 296, + 180 + ], + "type": "text", + "content": "frameworks [26, 57] that jointly optimize distortion-specific objectives and quality prediction, using synthetic distortions to enrich training signals and improve generalization. Another line of work focuses on image degradation modeling. Methods like Re-IQA [31] and QPT [58] simulate realistic distortions by applying diverse types and combinations to entire images, as illustrated in Fig.2. They then leverage unsupervised learning frameworks such as MoCo-v2[3] to extract distortion priors from large-scale synthetic datasets." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 55, + 182, + 297, + 302 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 182, + 297, + 302 + ], + "spans": [ + { + "bbox": [ + 55, + 182, + 297, + 302 + ], + "type": "text", + "content": "However, both categories of methods share a key limitation: they focus on global distortion priors while ignoring fine-grained local distortions, which are essential for accurate BIQA. Research [15, 17] shows that the human visual system (HVS) employs both local detection and global appearance strategies when assessing image quality. For subtle, near-threshold distortions, the HVS prioritizes local details, while for highly visible distortions, it shifts focus back to global content. This highlights the vital role of local distortion features in comprehensive image quality evaluation." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 55, + 304, + 296, + 412 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 304, + 296, + 412 + ], + "spans": [ + { + "bbox": [ + 55, + 304, + 296, + 412 + ], + "type": "text", + "content": "In this paper, we propose a novel Spatial Heterogeneous Distortion-aware Image Quality Assessment (SHDIQA) method, designed to improve the local distortion-awareness of BIQA models based on the Vision Transformer (ViT) architecture. Unlike prior works [10, 14, 50] that focus on modifying ViT architectures to enhance local feature extraction, our method fully leverages rich local distortion priors generated from synthetic distortions, while introducing minimal architectural changes (only 0.2M additional parameters)." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 55, + 415, + 298, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 415, + 298, + 714 + ], + "spans": [ + { + "bbox": [ + 55, + 415, + 298, + 714 + ], + "type": "text", + "content": "Specifically, we introduce a Block-wise Image Degradation (BWID) method tailored for IQA tasks to generate patch-level distortion priors. As shown in Fig. 2, BWID partitions a distorted image into multiple blocks and applies a randomly selected distortion from 24 categories across 4 severity levels to each block. This results in a composite image with spatially varying local distortions. Each image naturally comes with an image-level multi-class label and corresponding patch-level single-class labels, effectively addressing the lack of localized distortion annotations in existing methods. To further utilize the patch-level distortion priors, we propose a Block-wise Aggregation and Filtering (BwAF) mechanism. BWAF uses the local priors from BWID to train multiple distortion (DIS) tokens, which interact with a global classification (CLS) token through a cross-attention mechanism with an adaptive head. This interaction mimics the HVS's dual strategy of local detection and global perception, thereby improving overall quality assessment, as illustrated in Fig. 1. Moreover, previous studies [16, 40] highlight the importance of modeling the relationships among different distortion types in perceptual quality assessment. To achieve this, we incorporate Contrastive Knowledge Distillation (CKD) and Affinity Knowledge Distillation (AKD). CKD aligns similar distortions and separates dissimilar ones, improving the model's ability to differentiate between distor" + } + ] + } + ], + "index": 3 + }, + { + "type": "image", + "bbox": [ + 323, + 70, + 367, + 121 + ], + "blocks": [ + { + "bbox": [ + 323, + 70, + 367, + 121 + ], + "lines": [ + { + "bbox": [ + 323, + 70, + 367, + 121 + ], + "spans": [ + { + "bbox": [ + 323, + 70, + 367, + 121 + ], + "type": "image", + "image_path": "39dcb6f888a06e9d61a88a1f100001fdbd5417b9d2a1a1438bb4f2dbaab49f47.jpg" + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_body" + } + ], + "index": 4 + }, + { + "type": "image", + "bbox": [ + 370, + 70, + 412, + 121 + ], + "blocks": [ + { + "bbox": [ + 370, + 70, + 412, + 121 + ], + "lines": [ + { + "bbox": [ + 370, + 70, + 412, + 121 + ], + "spans": [ + { + "bbox": [ + 370, + 70, + 412, + 121 + ], + "type": "image", + "image_path": "837659c2d179a6b916b9df84f03f7098ce6b8bf1298f04997550a949545fdae1.jpg" + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 371, + 121, + 412, + 129 + ], + "lines": [ + { + "bbox": [ + 371, + 121, + 412, + 129 + ], + "spans": [ + { + "bbox": [ + 371, + 121, + 412, + 129 + ], + "type": "text", + "content": "Distorted Region" + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "image_caption" + } + ], + "index": 6 + }, + { + "type": "image", + "bbox": [ + 416, + 70, + 459, + 121 + ], + "blocks": [ + { + "bbox": [ + 416, + 70, + 459, + 121 + ], + "lines": [ + { + "bbox": [ + 416, + 70, + 459, + 121 + ], + "spans": [ + { + "bbox": [ + 416, + 70, + 459, + 121 + ], + "type": "image", + "image_path": "4df4db78557de28709340952da50a92c7ec1a028ff4ede992198c7044b2c3091.jpg" + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "image_body" + } + ], + "index": 8 + }, + { + "type": "image", + "bbox": [ + 463, + 70, + 504, + 121 + ], + "blocks": [ + { + "bbox": [ + 463, + 70, + 504, + 121 + ], + "lines": [ + { + "bbox": [ + 463, + 70, + 504, + 121 + ], + "spans": [ + { + "bbox": [ + 463, + 70, + 504, + 121 + ], + "type": "image", + "image_path": "27bc660bbab2540866e20536f78f10adb35e9726efa7abb17bd9b31e495bf5f6.jpg" + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "image_body" + } + ], + "index": 10 + }, + { + "type": "image", + "bbox": [ + 506, + 80, + 546, + 121 + ], + "blocks": [ + { + "bbox": [ + 506, + 80, + 546, + 121 + ], + "lines": [ + { + "bbox": [ + 506, + 80, + 546, + 121 + ], + "spans": [ + { + "bbox": [ + 506, + 80, + 546, + 121 + ], + "type": "image", + "image_path": "e5fb6976fdeab7ec32586b2927b37b562015885be505b69cd2c42f00de732c51.jpg" + } + ] + } + ], + "index": 12, + "angle": 0, + "type": "image_body" + } + ], + "index": 12 + }, + { + "type": "image", + "bbox": [ + 323, + 129, + 366, + 173 + ], + "blocks": [ + { + "bbox": [ + 335, + 122, + 356, + 129 + ], + "lines": [ + { + "bbox": [ + 335, + 122, + 356, + 129 + ], + "spans": [ + { + "bbox": [ + 335, + 122, + 356, + 129 + ], + "type": "text", + "content": "Image B" + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 323, + 129, + 366, + 173 + ], + "lines": [ + { + "bbox": [ + 323, + 129, + 366, + 173 + ], + "spans": [ + { + "bbox": [ + 323, + 129, + 366, + 173 + ], + "type": "image", + "image_path": "18b4a0e21421b0d1ce2d31e31f3b771f5e19cea4b474ccffb54275e536104430.jpg" + } + ] + } + ], + "index": 13, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 313, + 182, + 555, + 216 + ], + "lines": [ + { + "bbox": [ + 313, + 182, + 555, + 216 + ], + "spans": [ + { + "bbox": [ + 313, + 182, + 555, + 216 + ], + "type": "text", + "content": "Figure 2. The proposed SHDIQA method creates a heterogeneous degradation space with richer priors (Top) and assigns dedicated distortion-aware (DIS) tokens to each local region (Bottom)." + } + ] + } + ], + "index": 18, + "angle": 0, + "type": "image_caption" + } + ], + "index": 13 + }, + { + "type": "image", + "bbox": [ + 371, + 129, + 413, + 172 + ], + "blocks": [ + { + "bbox": [ + 371, + 129, + 413, + 172 + ], + "lines": [ + { + "bbox": [ + 371, + 129, + 413, + 172 + ], + "spans": [ + { + "bbox": [ + 371, + 129, + 413, + 172 + ], + "type": "image", + "image_path": "39519b5f774f3b6d3c4efdbd7cf977780d50ece87c408c52e93f5d4064e6ee76.jpg" + } + ] + } + ], + "index": 14, + "angle": 0, + "type": "image_body" + } + ], + "index": 14 + }, + { + "type": "image", + "bbox": [ + 415, + 129, + 459, + 173 + ], + "blocks": [ + { + "bbox": [ + 418, + 122, + 457, + 129 + ], + "lines": [ + { + "bbox": [ + 418, + 122, + 457, + 129 + ], + "spans": [ + { + "bbox": [ + 418, + 122, + 457, + 129 + ], + "type": "text", + "content": "DEIQT Method" + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 415, + 129, + 459, + 173 + ], + "lines": [ + { + "bbox": [ + 415, + 129, + 459, + 173 + ], + "spans": [ + { + "bbox": [ + 415, + 129, + 459, + 173 + ], + "type": "image", + "image_path": "8d6022ea40d2f18ae1b475560218b9c176efc22bae8fc91e226020088fb41f77.jpg" + } + ] + } + ], + "index": 15, + "angle": 0, + "type": "image_body" + } + ], + "index": 15 + }, + { + "type": "image", + "bbox": [ + 462, + 129, + 504, + 173 + ], + "blocks": [ + { + "bbox": [ + 465, + 122, + 504, + 129 + ], + "lines": [ + { + "bbox": [ + 465, + 122, + 504, + 129 + ], + "spans": [ + { + "bbox": [ + 465, + 122, + 504, + 129 + ], + "type": "text", + "content": "SHDIOA(Ours)" + } + ] + } + ], + "index": 11, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 462, + 129, + 504, + 173 + ], + "lines": [ + { + "bbox": [ + 462, + 129, + 504, + 173 + ], + "spans": [ + { + "bbox": [ + 462, + 129, + 504, + 173 + ], + "type": "image", + "image_path": "7a4fa361f84280d044140b0a0c07cb7245517ddbfa3aa42163a9a480442c0d5b.jpg" + } + ] + } + ], + "index": 16, + "angle": 0, + "type": "image_body" + } + ], + "index": 16 + }, + { + "type": "image", + "bbox": [ + 506, + 129, + 547, + 173 + ], + "blocks": [ + { + "bbox": [ + 506, + 129, + 547, + 173 + ], + "lines": [ + { + "bbox": [ + 506, + 129, + 547, + 173 + ], + "spans": [ + { + "bbox": [ + 506, + 129, + 547, + 173 + ], + "type": "image", + "image_path": "eec06aa26b83c6b2f3ed31be17489cee36b06e20833592f33afe7db30475a55f.jpg" + } + ] + } + ], + "index": 17, + "angle": 0, + "type": "image_body" + } + ], + "index": 17 + }, + { + "bbox": [ + 313, + 232, + 555, + 269 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 232, + 555, + 269 + ], + "spans": [ + { + "bbox": [ + 313, + 232, + 555, + 269 + ], + "type": "text", + "content": "tion types. Meanwhile, AKD transfers knowledge about the relationships between different distortions, enhancing the model's generalization across various types of distortions." + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 326, + 269, + 511, + 280 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 326, + 269, + 511, + 280 + ], + "spans": [ + { + "bbox": [ + 326, + 269, + 511, + 280 + ], + "type": "text", + "content": "Our contributions are summarized as follows:" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 314, + 281, + 555, + 472 + ], + "type": "list", + "angle": 0, + "index": 24, + "blocks": [ + { + "bbox": [ + 314, + 281, + 555, + 341 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 314, + 281, + 555, + 341 + ], + "spans": [ + { + "bbox": [ + 314, + 281, + 555, + 341 + ], + "type": "text", + "content": "- We present a Block-wise Degradation Modelling approach that enriches the local distortion priors of the training datasets by dividing images into blocks and applying distinct distortions to each, enhancing the model's ability to assess image quality accurately in complex scenarios." + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 314, + 342, + 555, + 412 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 314, + 342, + 555, + 412 + ], + "spans": [ + { + "bbox": [ + 314, + 342, + 555, + 412 + ], + "type": "text", + "content": "- We introduce the Block-wise Aggregation and Filtering mechanism to fully capture and leverage localized distortions. It first directs precise attention to distortion details within image blocks, and then utilizes an adaptation head to filter out irrelevant distortion features, focusing the analysis on quality-aware information." + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 314, + 413, + 555, + 472 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 314, + 413, + 555, + 472 + ], + "spans": [ + { + "bbox": [ + 314, + 413, + 555, + 472 + ], + "type": "text", + "content": "- We propose a novel Distortion-Guided Knowledge Distillation strategy that combines Contrastive Knowledge Distillation (CKD) to enhance distortion discrimination and Affinity Knowledge Distillation (AKD) to enrich relational insights between local distortion areas." + } + ] + } + ], + "index": 23 + } + ], + "sub_type": "text" + }, + { + "bbox": [ + 314, + 484, + 400, + 496 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 314, + 484, + 400, + 496 + ], + "spans": [ + { + "bbox": [ + 314, + 484, + 400, + 496 + ], + "type": "text", + "content": "2. Related Work" + } + ] + } + ], + "index": 25 + }, + { + "bbox": [ + 314, + 504, + 500, + 517 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 314, + 504, + 500, + 517 + ], + "spans": [ + { + "bbox": [ + 314, + 504, + 500, + 517 + ], + "type": "text", + "content": "2.1. BIQA with Degradation Modelling" + } + ] + } + ], + "index": 26 + }, + { + "bbox": [ + 313, + 522, + 556, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 522, + 556, + 713 + ], + "spans": [ + { + "bbox": [ + 313, + 522, + 556, + 713 + ], + "type": "text", + "content": "Recent studies [10, 61] have shown that simple distortion simulations do not fully capture the complexities of real-world degradations. To address this gap, Image Degradation Modeling has been introduced to represent various distortion types [34, 45, 46]. DBCNN [56] decomposes real distortions into multiple synthetic ones, creating large-scale synthetic images for training. DipIQ [25] categorizes common distortions like noise, blur, and compression into severity levels, while Re-IQA [31] introduces 25 distortion types to expand the degradation space. QPT [58] models second-order degradation using skip and shuffle operations across nine types in three groups. In contrast to these global degradation approaches, our method applies localized distortions to different regions of the image, enabling the model to learn fine-grained distortion perception both within and across blocks through knowledge distillation." + } + ] + } + ], + "index": 27 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "text", + "content": "2345" + } + ] + } + ], + "index": 28 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 1 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 58, + 70, + 187, + 167 + ], + "blocks": [ + { + "bbox": [ + 58, + 70, + 187, + 167 + ], + "lines": [ + { + "bbox": [ + 58, + 70, + 187, + 167 + ], + "spans": [ + { + "bbox": [ + 58, + 70, + 187, + 167 + ], + "type": "image", + "image_path": "2a04a5c8e746f1fed82b421526c8abe6b8c9226a75bb2d8c09b5c310e99bc42e.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + } + ], + "index": 0 + }, + { + "type": "image", + "bbox": [ + 58, + 180, + 188, + 258 + ], + "blocks": [ + { + "bbox": [ + 67, + 170, + 179, + 180 + ], + "lines": [ + { + "bbox": [ + 67, + 170, + 179, + 180 + ], + "spans": [ + { + "bbox": [ + 67, + 170, + 179, + 180 + ], + "type": "text", + "content": "(a) Block-wise Degradation" + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 58, + 180, + 188, + 258 + ], + "lines": [ + { + "bbox": [ + 58, + 180, + 188, + 258 + ], + "spans": [ + { + "bbox": [ + 58, + 180, + 188, + 258 + ], + "type": "image", + "image_path": "10ca39a192719452500ef9bfbddc705bf801eb2bf8949e7d2390ab00860304f1.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + } + ], + "index": 2 + }, + { + "type": "image", + "bbox": [ + 200, + 71, + 555, + 259 + ], + "blocks": [ + { + "bbox": [ + 200, + 71, + 555, + 259 + ], + "lines": [ + { + "bbox": [ + 200, + 71, + 555, + 259 + ], + "spans": [ + { + "bbox": [ + 200, + 71, + 555, + 259 + ], + "type": "image", + "image_path": "b6f136828165943658dee6db4e344ed6a83283f4bffe6bdb685065cff5e22d3b.jpg" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 55, + 269, + 555, + 346 + ], + "lines": [ + { + "bbox": [ + 55, + 269, + 555, + 346 + ], + "spans": [ + { + "bbox": [ + 55, + 269, + 555, + 346 + ], + "type": "text", + "content": "Figure 3. Framework Overview. Given an input image " + }, + { + "bbox": [ + 55, + 269, + 555, + 346 + ], + "type": "inline_equation", + "content": "\\mathbf{I}" + }, + { + "bbox": [ + 55, + 269, + 555, + 346 + ], + "type": "text", + "content": ", (a) Block-wise Degradation is first applied to generate a degraded version " + }, + { + "bbox": [ + 55, + 269, + 555, + 346 + ], + "type": "inline_equation", + "content": "\\mathbf{I}_d" + }, + { + "bbox": [ + 55, + 269, + 555, + 346 + ], + "type": "text", + "content": " (Sec. 3.1). Both " + }, + { + "bbox": [ + 55, + 269, + 555, + 346 + ], + "type": "inline_equation", + "content": "\\mathbf{I}" + }, + { + "bbox": [ + 55, + 269, + 555, + 346 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 55, + 269, + 555, + 346 + ], + "type": "inline_equation", + "content": "\\mathbf{I}_d" + }, + { + "bbox": [ + 55, + 269, + 555, + 346 + ], + "type": "text", + "content": " are then fed into a student encoder-decoder network, which is initialized with multiple distortion (DIS) tokens (each attending to a specific local block) and a CLS token for global quality representation. (b) Guided by the Block-wise Aggregation (Sec. 3.2), the DIS tokens and the CLS token are used to extract local distortion features " + }, + { + "bbox": [ + 55, + 269, + 555, + 346 + ], + "type": "inline_equation", + "content": "\\mathcal{F}_{dis}" + }, + { + "bbox": [ + 55, + 269, + 555, + 346 + ], + "type": "text", + "content": " and a global quality feature " + }, + { + "bbox": [ + 55, + 269, + 555, + 346 + ], + "type": "inline_equation", + "content": "\\mathcal{F}_{cls}" + }, + { + "bbox": [ + 55, + 269, + 555, + 346 + ], + "type": "text", + "content": ", which are further refined by the decoder into " + }, + { + "bbox": [ + 55, + 269, + 555, + 346 + ], + "type": "inline_equation", + "content": "\\hat{\\mathcal{F}}_{dis}" + }, + { + "bbox": [ + 55, + 269, + 555, + 346 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 55, + 269, + 555, + 346 + ], + "type": "inline_equation", + "content": "\\hat{\\mathcal{F}}_{cls}" + }, + { + "bbox": [ + 55, + 269, + 555, + 346 + ], + "type": "text", + "content": ". These refined features are processed by a quality-aware adaptation head to suppress irrelevant information and predict the quality score via an MLP. (c) To further enhance distortion perception, Contrastive KD is used to improve inter-block distortion discrimination, while Affinity KD captures relationships between distortions across different regions (Sec. 3.3)." + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_caption" + } + ], + "index": 3 + }, + { + "bbox": [ + 55, + 357, + 196, + 369 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 357, + 196, + 369 + ], + "spans": [ + { + "bbox": [ + 55, + 357, + 196, + 369 + ], + "type": "text", + "content": "2.2. BIQA with Transformers" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 54, + 374, + 297, + 601 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 54, + 374, + 297, + 601 + ], + "spans": [ + { + "bbox": [ + 54, + 374, + 297, + 601 + ], + "type": "text", + "content": "Vision Transformer (ViT)-based IQA architectures include hybrid transformers [2, 10, 21, 50, 54] and pure transformers [14, 20, 29]. Hybrid models integrate multi-scale CNN information for better local distortion perception but increase complexity. Pure transformers like DEIQT [29] capture global abstraction using the CLS token but struggle with local structure modeling [63]. To reduce memory overhead, methods using local self-attention and sparse block structures have been developed [23, 30], though at the cost of some long-distance modeling. Sampling-based methods like VT-IQA [17] combine global and local information by resampling distorted images and using global attention. However, they lack explicit supervision for local distortions, which limits their ability to capture localized degradations during training. In contrast, our approach enhances local distortion perception in pure ViT by incorporating patch-level distortion priors, while preserving global context modeling through a mask matrix specifically applied to distortion tokens, leaving the self-attention process unaffected." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 55, + 611, + 231, + 624 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 611, + 231, + 624 + ], + "spans": [ + { + "bbox": [ + 55, + 611, + 231, + 624 + ], + "type": "text", + "content": "2.3. IQA with Knowledge Distillation" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 55, + 629, + 296, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 629, + 296, + 713 + ], + "spans": [ + { + "bbox": [ + 55, + 629, + 296, + 713 + ], + "type": "text", + "content": "Knowledge Distillation (KD), originally proposed for model compression [12], has been adapted for tasks like action recognition [8], re-identification [28], and depth estimation [11]. Recently, KD has been used in methods like Degraded-Reference IQA (DR-IQA) [59] and Non-aligned Reference IQA (NAR-IQA) [52] to transfer reference information, though these still partially rely on reference images." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 313, + 357, + 555, + 430 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 357, + 555, + 430 + ], + "spans": [ + { + "bbox": [ + 313, + 357, + 555, + 430 + ], + "type": "text", + "content": "A related BIQA method is SSLIQA [55], which employs KD to enforce semantic consistency across different triples within a batch. However, our approach differs significantly in that our KD explicitly targets inter-block and intra-block distortion information, allowing for more direct extraction of quality-related distortion priors from the augmented image." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 313, + 441, + 397, + 455 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 441, + 397, + 455 + ], + "spans": [ + { + "bbox": [ + 313, + 441, + 397, + 455 + ], + "type": "text", + "content": "3. Methodology" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 313, + 462, + 556, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 462, + 556, + 713 + ], + "spans": [ + { + "bbox": [ + 313, + 462, + 556, + 713 + ], + "type": "text", + "content": "We introduce Spatial-Heterogeneous Distortion-aware Image Quality Assessment (SHDIQA), a framework designed to enhance the perception of local and non-uniform distortions in authentic images, as illustrated in Fig. 3. Specifically, the proposed SHDIQA first divides the input image " + }, + { + "bbox": [ + 313, + 462, + 556, + 713 + ], + "type": "inline_equation", + "content": "I" + }, + { + "bbox": [ + 313, + 462, + 556, + 713 + ], + "type": "text", + "content": " into vertical blocks and applies random distortions using a Block-wise Degradation module (Sec. 3.1), producing a degraded image " + }, + { + "bbox": [ + 313, + 462, + 556, + 713 + ], + "type": "inline_equation", + "content": "I_d" + }, + { + "bbox": [ + 313, + 462, + 556, + 713 + ], + "type": "text", + "content": ". Both " + }, + { + "bbox": [ + 313, + 462, + 556, + 713 + ], + "type": "inline_equation", + "content": "I" + }, + { + "bbox": [ + 313, + 462, + 556, + 713 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 313, + 462, + 556, + 713 + ], + "type": "inline_equation", + "content": "I_d" + }, + { + "bbox": [ + 313, + 462, + 556, + 713 + ], + "type": "text", + "content": " are then processed by a student encoder-decoder network equipped with a Blockwise Aggregation module (Sec. 3.2), which extracts local distortion features " + }, + { + "bbox": [ + 313, + 462, + 556, + 713 + ], + "type": "inline_equation", + "content": "\\mathcal{F}_{dis}" + }, + { + "bbox": [ + 313, + 462, + 556, + 713 + ], + "type": "text", + "content": " via multiple distortion (DIS) tokens and global quality feature " + }, + { + "bbox": [ + 313, + 462, + 556, + 713 + ], + "type": "inline_equation", + "content": "\\mathcal{F}_{cls}" + }, + { + "bbox": [ + 313, + 462, + 556, + 713 + ], + "type": "text", + "content": " via classification (CLS) token. Next, the local and global features " + }, + { + "bbox": [ + 313, + 462, + 556, + 713 + ], + "type": "inline_equation", + "content": "\\mathcal{F}_{dis}" + }, + { + "bbox": [ + 313, + 462, + 556, + 713 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 313, + 462, + 556, + 713 + ], + "type": "inline_equation", + "content": "\\mathcal{F}_{cls}" + }, + { + "bbox": [ + 313, + 462, + 556, + 713 + ], + "type": "text", + "content": " are fused via a cross-attention mechanism in the adaptation head to suppress irrelevant distortion information (Sec. 3.2). To enhance the model's ability to represent and distinguish distortion patterns, we further introduce Contrastive Knowledge Distillation (CKD) and Affinity Knowledge Distillation (AKD) (Sec. 3.3). These modules jointly improve distortion discriminability and capture inter-block distortion correlations, enabling a more distortion-aware representation." + } + ] + } + ], + "index": 11 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "text", + "content": "2346" + } + ] + } + ], + "index": 12 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 2 + }, + { + "para_blocks": [ + { + "bbox": [ + 55, + 72, + 241, + 85 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 72, + 241, + 85 + ], + "spans": [ + { + "bbox": [ + 55, + 72, + 241, + 85 + ], + "type": "text", + "content": "3.1. Block-wise Degradation Modelling" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 55, + 89, + 297, + 281 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 89, + 297, + 281 + ], + "spans": [ + { + "bbox": [ + 55, + 89, + 297, + 281 + ], + "type": "text", + "content": "Recent methods [31, 58] apply multiple distortions uniformly across the entire image, ignoring the finer details of local distortions. To address this, we propose a Block-wise Degradation Modelling strategy, which applies 24 distortion types and 4 degradation levels to different regions of the image, thereby generating a composite local distortion prior. Specifically, given an image of size " + }, + { + "bbox": [ + 55, + 89, + 297, + 281 + ], + "type": "inline_equation", + "content": "W \\times H" + }, + { + "bbox": [ + 55, + 89, + 297, + 281 + ], + "type": "text", + "content": ", we divide it into " + }, + { + "bbox": [ + 55, + 89, + 297, + 281 + ], + "type": "inline_equation", + "content": "K" + }, + { + "bbox": [ + 55, + 89, + 297, + 281 + ], + "type": "text", + "content": " vertical blocks of size " + }, + { + "bbox": [ + 55, + 89, + 297, + 281 + ], + "type": "inline_equation", + "content": "W \\times \\frac{H}{K}" + }, + { + "bbox": [ + 55, + 89, + 297, + 281 + ], + "type": "text", + "content": ". For each block " + }, + { + "bbox": [ + 55, + 89, + 297, + 281 + ], + "type": "inline_equation", + "content": "k" + }, + { + "bbox": [ + 55, + 89, + 297, + 281 + ], + "type": "text", + "content": " (" + }, + { + "bbox": [ + 55, + 89, + 297, + 281 + ], + "type": "inline_equation", + "content": "k = 1, \\dots, K" + }, + { + "bbox": [ + 55, + 89, + 297, + 281 + ], + "type": "text", + "content": "), we randomly select a rectangular region " + }, + { + "bbox": [ + 55, + 89, + 297, + 281 + ], + "type": "inline_equation", + "content": "C_k" + }, + { + "bbox": [ + 55, + 89, + 297, + 281 + ], + "type": "text", + "content": " with top-left coordinate " + }, + { + "bbox": [ + 55, + 89, + 297, + 281 + ], + "type": "inline_equation", + "content": "\\alpha_i" + }, + { + "bbox": [ + 55, + 89, + 297, + 281 + ], + "type": "text", + "content": " and size " + }, + { + "bbox": [ + 55, + 89, + 297, + 281 + ], + "type": "inline_equation", + "content": "w_k \\times h_k" + }, + { + "bbox": [ + 55, + 89, + 297, + 281 + ], + "type": "text", + "content": ", where " + }, + { + "bbox": [ + 55, + 89, + 297, + 281 + ], + "type": "inline_equation", + "content": "w_k = \\beta_w \\times W" + }, + { + "bbox": [ + 55, + 89, + 297, + 281 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 55, + 89, + 297, + 281 + ], + "type": "inline_equation", + "content": "h_k = \\beta_h \\times \\frac{H}{K}" + }, + { + "bbox": [ + 55, + 89, + 297, + 281 + ], + "type": "text", + "content": ". Empirically, we find that setting " + }, + { + "bbox": [ + 55, + 89, + 297, + 281 + ], + "type": "inline_equation", + "content": "\\beta_w = \\beta_h = 1" + }, + { + "bbox": [ + 55, + 89, + 297, + 281 + ], + "type": "text", + "content": ", i.e., degrading the entire block, yields the best performance. Then, the degradation is applied using the function " + }, + { + "bbox": [ + 55, + 89, + 297, + 281 + ], + "type": "inline_equation", + "content": "\\mathcal{D}_k(C_k; q_k, p_k)" + }, + { + "bbox": [ + 55, + 89, + 297, + 281 + ], + "type": "text", + "content": ", where " + }, + { + "bbox": [ + 55, + 89, + 297, + 281 + ], + "type": "inline_equation", + "content": "q_k" + }, + { + "bbox": [ + 55, + 89, + 297, + 281 + ], + "type": "text", + "content": " denotes the type of distortion (e.g., noise, darkening), and " + }, + { + "bbox": [ + 55, + 89, + 297, + 281 + ], + "type": "inline_equation", + "content": "p_k" + }, + { + "bbox": [ + 55, + 89, + 297, + 281 + ], + "type": "text", + "content": " represents its intensity. The degraded image " + }, + { + "bbox": [ + 55, + 89, + 297, + 281 + ], + "type": "inline_equation", + "content": "I_d" + }, + { + "bbox": [ + 55, + 89, + 297, + 281 + ], + "type": "text", + "content": " is generated as follows:" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 66, + 288, + 296, + 302 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 66, + 288, + 296, + 302 + ], + "spans": [ + { + "bbox": [ + 66, + 288, + 296, + 302 + ], + "type": "interline_equation", + "content": "\\boldsymbol {I} _ {d} = \\mathcal {A} (\\boldsymbol {I}) = \\mathcal {D} _ {K} \\left(\\dots \\mathcal {D} _ {1} \\left(\\boldsymbol {C} _ {1}; q _ {1}, p _ {1}\\right) \\dots ; q _ {K}, p _ {K}\\right) \\tag {1}", + "image_path": "e6dd83d595faf5b4f94b6d3467559ae52829c023f195ad9be8d34298cc61a73e.jpg" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 55, + 308, + 296, + 369 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 308, + 296, + 369 + ], + "spans": [ + { + "bbox": [ + 55, + 308, + 296, + 369 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 55, + 308, + 296, + 369 + ], + "type": "inline_equation", + "content": "K = 4" + }, + { + "bbox": [ + 55, + 308, + 296, + 369 + ], + "type": "text", + "content": " denotes the number of degraded blocks. Unlike previous methods, our approach focuses on localized distortions, providing richer priors for feature aggregation (Sec. 3.2) and knowledge distillation (Sec. 3.3), improving the model's ability to handle complex real-world distortions." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 55, + 374, + 253, + 387 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 374, + 253, + 387 + ], + "spans": [ + { + "bbox": [ + 55, + 374, + 253, + 387 + ], + "type": "text", + "content": "3.2. Block-wise Aggregation and Filtering" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 55, + 392, + 297, + 452 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 392, + 297, + 452 + ], + "spans": [ + { + "bbox": [ + 55, + 392, + 297, + 452 + ], + "type": "text", + "content": "In this section, we introduce the Block-wise Aggregation method to extract the local distortion feature " + }, + { + "bbox": [ + 55, + 392, + 297, + 452 + ], + "type": "inline_equation", + "content": "\\mathcal{F}_{dis}" + }, + { + "bbox": [ + 55, + 392, + 297, + 452 + ], + "type": "text", + "content": " from the degraded image " + }, + { + "bbox": [ + 55, + 392, + 297, + 452 + ], + "type": "inline_equation", + "content": "I_{d}" + }, + { + "bbox": [ + 55, + 392, + 297, + 452 + ], + "type": "text", + "content": " and the input image " + }, + { + "bbox": [ + 55, + 392, + 297, + 452 + ], + "type": "inline_equation", + "content": "\\pmb{I}" + }, + { + "bbox": [ + 55, + 392, + 297, + 452 + ], + "type": "text", + "content": ". We then present the Quality-aware Adaptation Head, which filters out quality-irrelevant distortion features from distortion feature " + }, + { + "bbox": [ + 55, + 392, + 297, + 452 + ], + "type": "inline_equation", + "content": "\\mathcal{F}_{dis}" + }, + { + "bbox": [ + 55, + 392, + 297, + 452 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 55, + 456, + 185, + 468 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 456, + 185, + 468 + ], + "spans": [ + { + "bbox": [ + 55, + 456, + 185, + 468 + ], + "type": "text", + "content": "3.2.1. Block-wise Aggregation" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 55, + 472, + 297, + 614 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 472, + 297, + 614 + ], + "spans": [ + { + "bbox": [ + 55, + 472, + 297, + 614 + ], + "type": "text", + "content": "We begin by randomly initializing " + }, + { + "bbox": [ + 55, + 472, + 297, + 614 + ], + "type": "inline_equation", + "content": "K" + }, + { + "bbox": [ + 55, + 472, + 297, + 614 + ], + "type": "text", + "content": " local distortion-aware (DIS) tokens, denoted as " + }, + { + "bbox": [ + 55, + 472, + 297, + 614 + ], + "type": "inline_equation", + "content": "\\mathcal{F}_{dis}" + }, + { + "bbox": [ + 55, + 472, + 297, + 614 + ], + "type": "text", + "content": ", matching the number of degraded regions. Each DIS token " + }, + { + "bbox": [ + 55, + 472, + 297, + 614 + ], + "type": "inline_equation", + "content": "\\mathcal{F}_{dis}^{k}" + }, + { + "bbox": [ + 55, + 472, + 297, + 614 + ], + "type": "text", + "content": " corresponds to a specific region " + }, + { + "bbox": [ + 55, + 472, + 297, + 614 + ], + "type": "inline_equation", + "content": "C_k" + }, + { + "bbox": [ + 55, + 472, + 297, + 614 + ], + "type": "text", + "content": ". The resulting embedding sequence is defined as " + }, + { + "bbox": [ + 55, + 472, + 297, + 614 + ], + "type": "inline_equation", + "content": "\\mathcal{F} = \\{\\mathcal{F}_{dis};[\\mathcal{F}_{cls};\\mathcal{F}_{patch}]\\} \\in \\mathbb{R}^{(K + N)\\times D}" + }, + { + "bbox": [ + 55, + 472, + 297, + 614 + ], + "type": "text", + "content": ", where " + }, + { + "bbox": [ + 55, + 472, + 297, + 614 + ], + "type": "inline_equation", + "content": "N" + }, + { + "bbox": [ + 55, + 472, + 297, + 614 + ], + "type": "text", + "content": " is the number of image patches plus one CLS token, and " + }, + { + "bbox": [ + 55, + 472, + 297, + 614 + ], + "type": "inline_equation", + "content": "D" + }, + { + "bbox": [ + 55, + 472, + 297, + 614 + ], + "type": "text", + "content": " is the embedding dimension. To ensure each DIS token focuses on its associated distorted region " + }, + { + "bbox": [ + 55, + 472, + 297, + 614 + ], + "type": "inline_equation", + "content": "C_k" + }, + { + "bbox": [ + 55, + 472, + 297, + 614 + ], + "type": "text", + "content": ", we construct a mask map " + }, + { + "bbox": [ + 55, + 472, + 297, + 614 + ], + "type": "inline_equation", + "content": "\\boldsymbol{m} \\in \\mathbb{R}^{(K + N)}" + }, + { + "bbox": [ + 55, + 472, + 297, + 614 + ], + "type": "text", + "content": ". Based on this, we define the attention mask " + }, + { + "bbox": [ + 55, + 472, + 297, + 614 + ], + "type": "inline_equation", + "content": "\\mathcal{M}" + }, + { + "bbox": [ + 55, + 472, + 297, + 614 + ], + "type": "text", + "content": " to preserve the original global attention structure while enforcing localized attention for DIS tokens as follows:" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 55, + 620, + 277, + 652 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 620, + 277, + 652 + ], + "spans": [ + { + "bbox": [ + 55, + 620, + 277, + 652 + ], + "type": "interline_equation", + "content": "m _ {j} ^ {i} = \\left\\{ \\begin{array}{l l} \\mathbb {1} \\{\\text {t h e i - t h p a t c h o v e r l a p s w i t h C _ {i}} \\}, & j < N \\\\ 0, & j = N \\end{array} \\right.,", + "image_path": "42e4248dc3ff4f4d63bfac61a49ad625e77061538b67c9d9b6ff9e1334ad2eda.jpg" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 58, + 653, + 304, + 689 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 58, + 653, + 304, + 689 + ], + "spans": [ + { + "bbox": [ + 58, + 653, + 304, + 689 + ], + "type": "interline_equation", + "content": "\\mathbb {M} = \\left[ m ^ {1}; m ^ {2}; \\dots ; m ^ {K} \\right] ^ {T}, \\mathcal {M} = \\left[ \\begin{array}{c c} \\mathbf {1} _ {N \\times N} & \\mathbf {0} _ {N \\times K} \\\\ \\mathbb {M} _ {K \\times N} & \\mathbf {1} _ {K \\times K} \\end{array} \\right], \\tag {2}", + "image_path": "26643e2dff50f63382618c6293c288d9138adb0de3daa66642d6670b25a3a16f.jpg" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 55, + 689, + 296, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 689, + 296, + 714 + ], + "spans": [ + { + "bbox": [ + 55, + 689, + 296, + 714 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 55, + 689, + 296, + 714 + ], + "type": "inline_equation", + "content": "\\mathbb{1}\\{\\cdot\\}" + }, + { + "bbox": [ + 55, + 689, + 296, + 714 + ], + "type": "text", + "content": " is the indicator function. Intuitively, Eq. 2 enforces that each DIS token " + }, + { + "bbox": [ + 55, + 689, + 296, + 714 + ], + "type": "inline_equation", + "content": "\\mathcal{F}_{dis}^{k}" + }, + { + "bbox": [ + 55, + 689, + 296, + 714 + ], + "type": "text", + "content": " attends only to the distorted" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 313, + 72, + 555, + 96 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 72, + 555, + 96 + ], + "spans": [ + { + "bbox": [ + 313, + 72, + 555, + 96 + ], + "type": "text", + "content": "content in block " + }, + { + "bbox": [ + 313, + 72, + 555, + 96 + ], + "type": "inline_equation", + "content": "C_k" + }, + { + "bbox": [ + 313, + 72, + 555, + 96 + ], + "type": "text", + "content": ". Assuming the encoder has " + }, + { + "bbox": [ + 313, + 72, + 555, + 96 + ], + "type": "inline_equation", + "content": "H" + }, + { + "bbox": [ + 313, + 72, + 555, + 96 + ], + "type": "text", + "content": " layers, the Multi-Head Self-Attention is applied as follows:" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 313, + 104, + 561, + 145 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 104, + 561, + 145 + ], + "spans": [ + { + "bbox": [ + 313, + 104, + 561, + 145 + ], + "type": "interline_equation", + "content": "\\boldsymbol {\\mathcal {F}} ^ {h} = \\left\\{ \\begin{array}{l l} \\sigma (\\log \\mathcal {M} + \\mathbf {Q} ^ {h} (\\mathbf {K} ^ {h}) ^ {T}) \\mathbf {V} ^ {h} + \\boldsymbol {\\mathcal {F}} ^ {h - 1}, & 0 < h \\leq H \\\\ \\boldsymbol {\\mathcal {F}}, & h = 0 \\end{array} , \\right. \\tag {3}", + "image_path": "46b222aa6359d5b082608a6069d140c4616a9ad7b1f89842e6b66f0988eaad55.jpg" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 313, + 145, + 555, + 218 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 145, + 555, + 218 + ], + "spans": [ + { + "bbox": [ + 313, + 145, + 555, + 218 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 313, + 145, + 555, + 218 + ], + "type": "inline_equation", + "content": "Q^h, K^h, V^h" + }, + { + "bbox": [ + 313, + 145, + 555, + 218 + ], + "type": "text", + "content": " are the query, key, and value matrices, obtained via linear projections of " + }, + { + "bbox": [ + 313, + 145, + 555, + 218 + ], + "type": "inline_equation", + "content": "\\mathcal{F}^{h-1}" + }, + { + "bbox": [ + 313, + 145, + 555, + 218 + ], + "type": "text", + "content": ", and " + }, + { + "bbox": [ + 313, + 145, + 555, + 218 + ], + "type": "inline_equation", + "content": "\\sigma" + }, + { + "bbox": [ + 313, + 145, + 555, + 218 + ], + "type": "text", + "content": " denotes the Softmax function. The encoder outputs features " + }, + { + "bbox": [ + 313, + 145, + 555, + 218 + ], + "type": "inline_equation", + "content": "\\mathcal{F}^H" + }, + { + "bbox": [ + 313, + 145, + 555, + 218 + ], + "type": "text", + "content": " after " + }, + { + "bbox": [ + 313, + 145, + 555, + 218 + ], + "type": "inline_equation", + "content": "H" + }, + { + "bbox": [ + 313, + 145, + 555, + 218 + ], + "type": "text", + "content": " layers. Following [29], we apply a transformer decoder to refine these features, yielding " + }, + { + "bbox": [ + 313, + 145, + 555, + 218 + ], + "type": "inline_equation", + "content": "\\{\\hat{\\mathcal{F}}_{dis}; \\hat{\\mathcal{F}}_{cls}\\} \\in \\mathbb{R}^{(K+1) \\times D}" + }, + { + "bbox": [ + 313, + 145, + 555, + 218 + ], + "type": "text", + "content": ", which are more relevant for image quality." + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 313, + 224, + 480, + 236 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 224, + 480, + 236 + ], + "spans": [ + { + "bbox": [ + 313, + 224, + 480, + 236 + ], + "type": "text", + "content": "3.2.2. Quality-aware Adaptation Head" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 313, + 238, + 555, + 346 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 238, + 555, + 346 + ], + "spans": [ + { + "bbox": [ + 313, + 238, + 555, + 346 + ], + "type": "text", + "content": "To effectively integrate global quality and local distortion features while suppressing quality-irrelevant information, we design a quality-aware adaptation head that highlights distortion cues closely related to perceived quality. Specifically, the global quality feature " + }, + { + "bbox": [ + 313, + 238, + 555, + 346 + ], + "type": "inline_equation", + "content": "\\hat{\\mathcal{F}}_{cls}" + }, + { + "bbox": [ + 313, + 238, + 555, + 346 + ], + "type": "text", + "content": " is used as the query to selectively attend to the local distortion features " + }, + { + "bbox": [ + 313, + 238, + 555, + 346 + ], + "type": "inline_equation", + "content": "\\hat{\\mathcal{F}}_{dis}" + }, + { + "bbox": [ + 313, + 238, + 555, + 346 + ], + "type": "text", + "content": ", which serve as key and value. This allows the model to emphasize distortion patterns most relevant to the overall quality. Overall, the final predicted quality score " + }, + { + "bbox": [ + 313, + 238, + 555, + 346 + ], + "type": "inline_equation", + "content": "\\hat{Y}" + }, + { + "bbox": [ + 313, + 238, + 555, + 346 + ], + "type": "text", + "content": " is computed as:" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 328, + 354, + 555, + 374 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 328, + 354, + 555, + 374 + ], + "spans": [ + { + "bbox": [ + 328, + 354, + 555, + 374 + ], + "type": "interline_equation", + "content": "\\hat {Y} = \\operatorname {L N} \\left(\\hat {\\mathcal {F}} _ {c l s} + \\text {C r o s s A t t e n t i o n} \\left(\\hat {\\mathcal {F}} _ {c l s}, \\hat {\\mathcal {F}} _ {d i s}\\right)\\right), \\tag {4}", + "image_path": "f5599c3818c4b34200d32239f539f220a47abce7da8a10caaa154b8de41414ac.jpg" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 313, + 381, + 555, + 418 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 381, + 555, + 418 + ], + "spans": [ + { + "bbox": [ + 313, + 381, + 555, + 418 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 313, + 381, + 555, + 418 + ], + "type": "inline_equation", + "content": "\\text{CrossAttention}(\\cdot)" + }, + { + "bbox": [ + 313, + 381, + 555, + 418 + ], + "type": "text", + "content": " denotes a cross-attention module, and " + }, + { + "bbox": [ + 313, + 381, + 555, + 418 + ], + "type": "inline_equation", + "content": "\\text{LN}" + }, + { + "bbox": [ + 313, + 381, + 555, + 418 + ], + "type": "text", + "content": " is a single-layer linear projection. During training, the model is optimized by minimizing the smooth " + }, + { + "bbox": [ + 313, + 381, + 555, + 418 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_1" + }, + { + "bbox": [ + 313, + 381, + 555, + 418 + ], + "type": "text", + "content": " loss." + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 313, + 424, + 533, + 437 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 424, + 533, + 437 + ], + "spans": [ + { + "bbox": [ + 313, + 424, + 533, + 437 + ], + "type": "text", + "content": "3.3. Distortion-Guided Knowledge Distillation" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 313, + 441, + 556, + 597 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 441, + 556, + 597 + ], + "spans": [ + { + "bbox": [ + 313, + 441, + 556, + 597 + ], + "type": "text", + "content": "Inspired by [5, 18, 33], we adopt a knowledge distillation method to learn correlations among distortion types by transferring quality-aware distortion knowledge from the teacher to the student. This distillation strategy addresses two key challenges: First, the student's ViT encoders, pre-trained on ImageNet [4], focus on semantics and lack distortion awareness in early training [19]. The teacher, pre-trained on IQA data, provides quality-aware guidance, directing the student to critical distortion patterns. Second, the absence of explicit ground truth for distortion correlations makes it difficult for the student to model these relationships independently. The teacher acts as a reliable target for correlation modeling, thereby improving the student's discriminative ability." + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 313, + 603, + 492, + 614 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 603, + 492, + 614 + ], + "spans": [ + { + "bbox": [ + 313, + 603, + 492, + 614 + ], + "type": "text", + "content": "3.3.1. Contrastive Knowledge Distillation" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 313, + 617, + 556, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 617, + 556, + 713 + ], + "spans": [ + { + "bbox": [ + 313, + 617, + 556, + 713 + ], + "type": "text", + "content": "Contrastive Knowledge Distillation (CKD) aims to extract the distortion features " + }, + { + "bbox": [ + 313, + 617, + 556, + 713 + ], + "type": "inline_equation", + "content": "\\pmb{V}_{dis}^{k}" + }, + { + "bbox": [ + 313, + 617, + 556, + 713 + ], + "type": "text", + "content": " of the " + }, + { + "bbox": [ + 313, + 617, + 556, + 713 + ], + "type": "inline_equation", + "content": "k" + }, + { + "bbox": [ + 313, + 617, + 556, + 713 + ], + "type": "text", + "content": "-th local block from a teacher encoder " + }, + { + "bbox": [ + 313, + 617, + 556, + 713 + ], + "type": "inline_equation", + "content": "\\mathcal{V}(\\cdot)" + }, + { + "bbox": [ + 313, + 617, + 556, + 713 + ], + "type": "text", + "content": " and transfer them to a student network. Simply maximizing feature similarity is insufficient to capture distinct distortion representations across spatial domains (Table 5). To address this, we reformulate the KD task as a pseudo-classification problem. As shown in Fig. 3, features from corresponding distorted regions of the teacher" + } + ] + } + ], + "index": 21 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 315, + 757 + ], + "type": "text", + "content": "2347" + } + ] + } + ], + "index": 22 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 3 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 58, + 70, + 555, + 286 + ], + "blocks": [ + { + "bbox": [ + 58, + 70, + 555, + 286 + ], + "lines": [ + { + "bbox": [ + 58, + 70, + 555, + 286 + ], + "spans": [ + { + "bbox": [ + 58, + 70, + 555, + 286 + ], + "type": "table", + "html": "
MethodLIVECSIQKADIDLIVECKonIQLIVEFBSPAQ
PLCCSRCCPLCCSRCCPLCCSRCCPLCCSRCCPLCCSRCCPLCCSRCCPLCCSRCC
BRISQUE [27]0.9440.9290.7480.8120.5670.5280.6290.6290.6850.6810.3410.3030.8170.809
DBCNN [56]0.9710.9680.9590.9460.8560.8510.8690.8510.8840.8750.5510.5450.9150.911
TIQA [54]0.9650.9490.8380.8250.8550.850.8610.8450.9030.8920.5810.541--
MetaIQA [61]0.9590.9600.9080.8990.7750.7620.8020.8350.8560.8870.5070.54--
HyperIQA (27M) [39]0.9660.9620.9420.9230.8450.8520.8820.8590.9170.9060.6020.5440.9150.911
TReS (152M) [10]0.9680.9690.9420.9220.8580.8590.8770.8460.9280.9150.6250.554--
MUSIQ (27M) [14]0.9110.9400.8930.8710.8720.8750.7460.7020.9280.9160.6610.5660.9210.918
Re-IQA (48M) [31]0.9710.9700.9600.9470.8850.8720.8540.8400.9230.914--0.9250.918
DEIQT (24M) [29]0.9820.9800.9630.9460.8870.8890.8940.8750.9340.9210.6630.5710.9230.919
QPT (24M) [58]------0.9140.8950.9410.9270.6750.5750.9280.925
LIQE (151M) [57]0.9510.9700.9390.9360.9310.9300.9100.9040.9080.919----
QFM-IQM (24M) [19]0.9830.9810.9650.9540.9060.9060.9130.8910.9360.9220.6670.5670.9240.920
LoDa (118M) [50]0.9790.975--0.9360.9310.8990.8760.9440.9320.6790.5780.9280.925
SHDIQA (24.2M)0.9840.9820.9720.9620.9400.9370.9220.9090.9480.9370.7350.6390.9340.930
", + "image_path": "eb6457f9027d6faf21500bff8ae4e838a05866d8091b9b1cb4f9122a35eb8e17.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "table_body" + } + ], + "index": 0 + }, + { + "bbox": [ + 64, + 293, + 544, + 304 + ], + "lines": [ + { + "bbox": [ + 64, + 293, + 544, + 304 + ], + "spans": [ + { + "bbox": [ + 64, + 293, + 544, + 304 + ], + "type": "text", + "content": "Table 1. Performance comparison measured by the average of SRCC and PLCC. Best results are in bold, second-best are underlined." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "text" + }, + { + "bbox": [ + 55, + 316, + 295, + 340 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 316, + 295, + 340 + ], + "spans": [ + { + "bbox": [ + 55, + 316, + 295, + 340 + ], + "type": "text", + "content": "and student are treated as positive pairs, while all other pairs are considered negative. The CKD loss is defined as:" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 61, + 350, + 296, + 426 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 61, + 350, + 296, + 426 + ], + "spans": [ + { + "bbox": [ + 61, + 350, + 296, + 426 + ], + "type": "interline_equation", + "content": "\\begin{array}{l} \\mathcal {L} _ {c k d} = - \\frac {1}{K} \\sum_ {k} \\log \\left(\\sigma \\left(\\frac {\\cos \\left(\\boldsymbol {\\mathcal {F}} _ {d i s} ^ {k} , \\boldsymbol {V} _ {d i s} ^ {k}\\right)}{\\tau_ {c}}\\right)\\right) \\\\ \\left. - \\frac{1}{K}\\sum_{\\substack{k,j\\\\ j\\neq k}}\\log \\left(1 - \\sigma \\left(\\frac{\\cos(\\boldsymbol{\\mathcal{F}}_{dis}^{k},\\boldsymbol{V}_{dis}^{j})}{\\tau_{c}}\\right)\\right)\\right), \\tag{5} \\\\ \\end{array}", + "image_path": "b73f7ee850b461462040f046eeff6ad5ec5b8f18a1a028d1b515aacf9dd705f5.jpg" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 55, + 436, + 296, + 509 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 436, + 296, + 509 + ], + "spans": [ + { + "bbox": [ + 55, + 436, + 296, + 509 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 55, + 436, + 296, + 509 + ], + "type": "inline_equation", + "content": "\\cos (\\cdot)" + }, + { + "bbox": [ + 55, + 436, + 296, + 509 + ], + "type": "text", + "content": " is cosine similarity, " + }, + { + "bbox": [ + 55, + 436, + 296, + 509 + ], + "type": "inline_equation", + "content": "\\sigma" + }, + { + "bbox": [ + 55, + 436, + 296, + 509 + ], + "type": "text", + "content": " is the sigmoid function, and " + }, + { + "bbox": [ + 55, + 436, + 296, + 509 + ], + "type": "inline_equation", + "content": "\\tau_{c} = 0.3" + }, + { + "bbox": [ + 55, + 436, + 296, + 509 + ], + "type": "text", + "content": " is the temperature factor. " + }, + { + "bbox": [ + 55, + 436, + 296, + 509 + ], + "type": "inline_equation", + "content": "\\mathcal{F}_{dis}^{k}" + }, + { + "bbox": [ + 55, + 436, + 296, + 509 + ], + "type": "text", + "content": " is the student feature vector of the " + }, + { + "bbox": [ + 55, + 436, + 296, + 509 + ], + "type": "inline_equation", + "content": "k" + }, + { + "bbox": [ + 55, + 436, + 296, + 509 + ], + "type": "text", + "content": "-th local block feature extracted by the DIS token. This approach enhances contrast among features of varying distortion types, improving the model's discrimination and distortion-awareness capabilities." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 55, + 515, + 216, + 528 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 515, + 216, + 528 + ], + "spans": [ + { + "bbox": [ + 55, + 515, + 216, + 528 + ], + "type": "text", + "content": "3.3.2. Affinity Knowledge Distillation" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 55, + 531, + 297, + 591 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 531, + 297, + 591 + ], + "spans": [ + { + "bbox": [ + 55, + 531, + 297, + 591 + ], + "type": "text", + "content": "Affinity Knowledge Distillation (AKD) focuses on distilling the interrelations among multiple blocks within an image " + }, + { + "bbox": [ + 55, + 531, + 297, + 591 + ], + "type": "inline_equation", + "content": "\\pmb{I}" + }, + { + "bbox": [ + 55, + 531, + 297, + 591 + ], + "type": "text", + "content": " or its degraded version " + }, + { + "bbox": [ + 55, + 531, + 297, + 591 + ], + "type": "inline_equation", + "content": "I_{d}" + }, + { + "bbox": [ + 55, + 531, + 297, + 591 + ], + "type": "text", + "content": ". As shown in Fig. 3, given teacher embeddings " + }, + { + "bbox": [ + 55, + 531, + 297, + 591 + ], + "type": "inline_equation", + "content": "\\mathbf{V}_{dis}" + }, + { + "bbox": [ + 55, + 531, + 297, + 591 + ], + "type": "text", + "content": " and student features " + }, + { + "bbox": [ + 55, + 531, + 297, + 591 + ], + "type": "inline_equation", + "content": "\\mathcal{F}_{dis}" + }, + { + "bbox": [ + 55, + 531, + 297, + 591 + ], + "type": "text", + "content": ", we compute student and teachers' pairwise affinity matrices:" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 105, + 602, + 296, + 618 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 105, + 602, + 296, + 618 + ], + "spans": [ + { + "bbox": [ + 105, + 602, + 296, + 618 + ], + "type": "interline_equation", + "content": "\\boldsymbol {R} _ {s t u} = \\bar {\\boldsymbol {F}} \\bar {\\boldsymbol {F}} ^ {T}, \\quad \\boldsymbol {R} _ {t e a} = \\bar {\\boldsymbol {V}} \\bar {\\boldsymbol {V}} ^ {T}, \\tag {6}", + "image_path": "dca4eea2ee523be3430e2b183e0c660f390fdbb1207ca16f44b8e47bfaf2d482.jpg" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 55, + 624, + 296, + 674 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 624, + 296, + 674 + ], + "spans": [ + { + "bbox": [ + 55, + 624, + 296, + 674 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 55, + 624, + 296, + 674 + ], + "type": "inline_equation", + "content": "\\bar{F}" + }, + { + "bbox": [ + 55, + 624, + 296, + 674 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 55, + 624, + 296, + 674 + ], + "type": "inline_equation", + "content": "\\bar{V}" + }, + { + "bbox": [ + 55, + 624, + 296, + 674 + ], + "type": "text", + "content": " are " + }, + { + "bbox": [ + 55, + 624, + 296, + 674 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_2" + }, + { + "bbox": [ + 55, + 624, + 296, + 674 + ], + "type": "text", + "content": "-normalized student " + }, + { + "bbox": [ + 55, + 624, + 296, + 674 + ], + "type": "inline_equation", + "content": "\\mathcal{F}_{dis}" + }, + { + "bbox": [ + 55, + 624, + 296, + 674 + ], + "type": "text", + "content": " and teacher features " + }, + { + "bbox": [ + 55, + 624, + 296, + 674 + ], + "type": "inline_equation", + "content": "V_{dis}" + }, + { + "bbox": [ + 55, + 624, + 296, + 674 + ], + "type": "text", + "content": ", respectively. These matrices capture pairwise correlations, and the goal is to align the student affinity matrix " + }, + { + "bbox": [ + 55, + 624, + 296, + 674 + ], + "type": "inline_equation", + "content": "R_{stu}" + }, + { + "bbox": [ + 55, + 624, + 296, + 674 + ], + "type": "text", + "content": " with the teacher matrix " + }, + { + "bbox": [ + 55, + 624, + 296, + 674 + ], + "type": "inline_equation", + "content": "R_{tea}" + }, + { + "bbox": [ + 55, + 624, + 296, + 674 + ], + "type": "text", + "content": ":" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 65, + 682, + 296, + 716 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 65, + 682, + 296, + 716 + ], + "spans": [ + { + "bbox": [ + 65, + 682, + 296, + 716 + ], + "type": "interline_equation", + "content": "\\mathcal {L} _ {a k d} = - \\frac {1}{K} \\sum_ {k = 1} ^ {K} \\mathrm {K L} \\left(\\sigma \\left(\\frac {\\boldsymbol {R} _ {s t u} ^ {k}}{\\tau_ {a}}\\right) \\| \\sigma \\left(\\frac {\\boldsymbol {R} _ {t e a} ^ {k}}{\\tau_ {a}}\\right)\\right), \\tag {7}", + "image_path": "303503b8f859052208672a89cc20f53a716805a302d771cf537555b284dbfa35.jpg" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 313, + 316, + 556, + 411 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 316, + 556, + 411 + ], + "spans": [ + { + "bbox": [ + 313, + 316, + 556, + 411 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 313, + 316, + 556, + 411 + ], + "type": "inline_equation", + "content": "\\tau_{a} = 2" + }, + { + "bbox": [ + 313, + 316, + 556, + 411 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 313, + 316, + 556, + 411 + ], + "type": "inline_equation", + "content": "\\mathrm{KL}(\\cdot)" + }, + { + "bbox": [ + 313, + 316, + 556, + 411 + ], + "type": "text", + "content": " denotes the Kullback-Leibler divergence. AKD complements CKD by focusing on the relational dynamics between image blocks, allowing the model to learn the relationship between distortion. However, relying solely on AKD may overlook fine-grained distortion-discriminative details, as highlighted in Table 2. Therefore, both CKD and AKD are essential for learning comprehensive distortion-aware features." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 313, + 416, + 437, + 429 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 416, + 437, + 429 + ], + "spans": [ + { + "bbox": [ + 313, + 416, + 437, + 429 + ], + "type": "text", + "content": "3.3.3. Overall Training Loss" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 313, + 432, + 556, + 552 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 432, + 556, + 552 + ], + "spans": [ + { + "bbox": [ + 313, + 432, + 556, + 552 + ], + "type": "text", + "content": "The overall loss combines four complementary components: " + }, + { + "bbox": [ + 313, + 432, + 556, + 552 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{\\mathrm{cls}}" + }, + { + "bbox": [ + 313, + 432, + 556, + 552 + ], + "type": "text", + "content": " provides category-level supervision for distortion identification but lacks awareness of relationships between distortion types; " + }, + { + "bbox": [ + 313, + 432, + 556, + 552 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{\\mathrm{ckd}}" + }, + { + "bbox": [ + 313, + 432, + 556, + 552 + ], + "type": "text", + "content": " enhances feature discrimination by clustering similar distortions and separating dissimilar ones, thus compensating for this limitation; " + }, + { + "bbox": [ + 313, + 432, + 556, + 552 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{\\mathrm{akd}}" + }, + { + "bbox": [ + 313, + 432, + 556, + 552 + ], + "type": "text", + "content": " further captures global perceptual structures by modeling relative similarities across different distortion. and " + }, + { + "bbox": [ + 313, + 432, + 556, + 552 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{\\log}" + }, + { + "bbox": [ + 313, + 432, + 556, + 552 + ], + "type": "text", + "content": " regularizes training via teacher outputs [44], mitigating overfitting to quality-irrelevant patterns. Finally, the total loss is defined as:" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 314, + 557, + 555, + 582 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 314, + 557, + 555, + 582 + ], + "spans": [ + { + "bbox": [ + 314, + 557, + 555, + 582 + ], + "type": "interline_equation", + "content": "\\mathcal {L} = \\lambda_ {\\mathrm {c l s}} \\mathcal {L} _ {\\mathrm {c l s}} + \\lambda_ {\\mathrm {c k d}} \\mathcal {L} _ {\\mathrm {c k d}} + \\lambda_ {\\mathrm {a k d}} \\mathcal {L} _ {\\mathrm {a k d}} + \\lambda_ {\\log} \\mathcal {L} _ {\\log} + \\| \\hat {Y} - Y _ {g} \\| _ {1}, \\tag {8}", + "image_path": "d7eb295035aa1cde8d940a43e8afc2e393415785f01c915d81c66a76e6a0c7eb.jpg" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 313, + 583, + 555, + 619 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 583, + 555, + 619 + ], + "spans": [ + { + "bbox": [ + 313, + 583, + 555, + 619 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 313, + 583, + 555, + 619 + ], + "type": "inline_equation", + "content": "Y_{g}" + }, + { + "bbox": [ + 313, + 583, + 555, + 619 + ], + "type": "text", + "content": " is the ground truth, " + }, + { + "bbox": [ + 313, + 583, + 555, + 619 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{\\mathrm{log}}" + }, + { + "bbox": [ + 313, + 583, + 555, + 619 + ], + "type": "text", + "content": " is the " + }, + { + "bbox": [ + 313, + 583, + 555, + 619 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_1" + }, + { + "bbox": [ + 313, + 583, + 555, + 619 + ], + "type": "text", + "content": " loss between student and teacher predictions, and " + }, + { + "bbox": [ + 313, + 583, + 555, + 619 + ], + "type": "inline_equation", + "content": "\\lambda_{\\mathrm{cls}}" + }, + { + "bbox": [ + 313, + 583, + 555, + 619 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 313, + 583, + 555, + 619 + ], + "type": "inline_equation", + "content": "\\lambda_{\\mathrm{ckd}}" + }, + { + "bbox": [ + 313, + 583, + 555, + 619 + ], + "type": "text", + "content": ", and " + }, + { + "bbox": [ + 313, + 583, + 555, + 619 + ], + "type": "inline_equation", + "content": "\\lambda_{\\mathrm{akd}}" + }, + { + "bbox": [ + 313, + 583, + 555, + 619 + ], + "type": "text", + "content": " are set to 1, with " + }, + { + "bbox": [ + 313, + 583, + 555, + 619 + ], + "type": "inline_equation", + "content": "\\lambda_{\\mathrm{log}}" + }, + { + "bbox": [ + 313, + 583, + 555, + 619 + ], + "type": "text", + "content": " set to 0.8." + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 313, + 628, + 395, + 642 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 628, + 395, + 642 + ], + "spans": [ + { + "bbox": [ + 313, + 628, + 395, + 642 + ], + "type": "text", + "content": "4. Experiments" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 313, + 647, + 555, + 659 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 647, + 555, + 659 + ], + "spans": [ + { + "bbox": [ + 313, + 647, + 555, + 659 + ], + "type": "text", + "content": "4.1. Benchmark Datasets and Evaluation Protocols" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 313, + 665, + 556, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 665, + 556, + 714 + ], + "spans": [ + { + "bbox": [ + 313, + 665, + 556, + 714 + ], + "type": "text", + "content": "Experiments were conducted on 7 typical BIQA datasets, including 3 synthetic datasets LIVE [35], CSIQ [15], KA-DID [22], and 4 authentic datasets LIVEC [9], KonIQ [13], LIVEFB [53], SPAQ [7]. Specifically, for authentic datasets," + } + ] + } + ], + "index": 17 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "text", + "content": "2348" + } + ] + } + ], + "index": 18 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 4 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 58, + 70, + 553, + 202 + ], + "blocks": [ + { + "bbox": [ + 58, + 70, + 553, + 202 + ], + "lines": [ + { + "bbox": [ + 58, + 70, + 553, + 202 + ], + "spans": [ + { + "bbox": [ + 58, + 70, + 553, + 202 + ], + "type": "table", + "html": "
Loss LogLoss LclsLoss LakdLoss LckdCSIQKADIDLIVELIVECAvg.
PLCCSRCCPLCCSRCCPLCCSRCCPLCCSRCC
XXXX0.9520.9350.8780.8840.9660.9640.8910.8670.917
✓XXX0.9550.9410.8980.9040.9720.9700.8980.8720.926
✓✓XX0.9590.9470.9190.9160.9740.9720.9040.8900.935
✓X✓X0.9700.9600.9270.9250.9800.9780.9190.9050.945
✓XX✓0.9660.9550.9130.9110.9760.9740.9140.9020.939
✓X✓✓0.9690.9580.9370.9350.9830.9810.9200.9040.948
✓✓✓✓0.9720.9620.9400.9370.9840.9820.9220.9090.951
", + "image_path": "5a2d1c29bac365b337d7eb051f4e280094908178f72f93e9ca88c4dd418588b1.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "table_body" + } + ], + "index": 0 + }, + { + "type": "table", + "bbox": [ + 56, + 229, + 296, + 348 + ], + "blocks": [ + { + "bbox": [ + 59, + 209, + 550, + 220 + ], + "lines": [ + { + "bbox": [ + 59, + 209, + 550, + 220 + ], + "spans": [ + { + "bbox": [ + 59, + 209, + 550, + 220 + ], + "type": "text", + "content": "Table 2. Ablation experiments about knowledge distillation components on multiple datasets, where bold entries indicate the best result." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 56, + 229, + 296, + 348 + ], + "lines": [ + { + "bbox": [ + 56, + 229, + 296, + 348 + ], + "spans": [ + { + "bbox": [ + 56, + 229, + 296, + 348 + ], + "type": "table", + "html": "
TrainingLIVEFBLIVECKonIQLIVECSIQ
TestingKonIQLIVECKonIQLIVECCSIQLIVE
DBCNN0.7160.7240.7540.7550.7580.877
P2P-BM0.7550.7380.7400.7700.712-
TReS0.7130.7400.7330.7860.761-
DEIQT0.7330.7810.7440.7940.7810.932
LoDa0.7630.8050.7450.811--
SHDIQA0.7230.7910.7820.8190.7890.943
", + "image_path": "75b70eddedef1602e4a4213f6d47b883f5b9a6180cee906b3e5b5266138431fe.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "table_body" + } + ], + "index": 2 + }, + { + "bbox": [ + 54, + 386, + 295, + 505 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 54, + 386, + 295, + 505 + ], + "spans": [ + { + "bbox": [ + 54, + 386, + 295, + 505 + ], + "type": "text", + "content": "LIVEC contains 1162 images from different mobile devices and photographers. SPAQ includes 11125 photos from 66 smartphones. KonIQ contains 10073 images from public sources, while LIVEFB is the largest real-world dataset with 39810 images. The synthetic datasets involve artificially distorting original images using methods such as JPEG compression and Gaussian blur. LIVE and CSIQ respectively have 779 and 866 synthetic distortion images, with 5 and 6 types of distortions. KADID contains 10125 synthetic distortion images, spanning 25 types of distortions." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 54, + 506, + 296, + 567 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 54, + 506, + 296, + 567 + ], + "spans": [ + { + "bbox": [ + 54, + 506, + 296, + 567 + ], + "type": "text", + "content": "Evaluation Criteria. The experiments use Pearson's Rank-Order Correlation Coefficient (SRCC) and Pearson's Linear Correlation Coefficient (PLCC) to assess prediction monotonicity and accuracy, respectively. Both metrics scale from -1 to 1, with values close to 1 indicating better performance." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 55, + 575, + 188, + 588 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 575, + 188, + 588 + ], + "spans": [ + { + "bbox": [ + 55, + 575, + 188, + 588 + ], + "type": "text", + "content": "4.2. Implementation Details" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 54, + 594, + 295, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 54, + 594, + 295, + 713 + ], + "spans": [ + { + "bbox": [ + 54, + 594, + 295, + 713 + ], + "type": "text", + "content": "For the student network, we follow the typical training strategy [29], using a Transformer encoder based on ViT-S from DeiT III [43] and a one-layer decoder. Our model is trained for 9 epochs with a learning rate of " + }, + { + "bbox": [ + 54, + 594, + 295, + 713 + ], + "type": "inline_equation", + "content": "2 \\times 10^{-4}" + }, + { + "bbox": [ + 54, + 594, + 295, + 713 + ], + "type": "text", + "content": ", decaying by a factor of 10 every 3 epochs. The batch size is 16 for LIVEC and 128 for KonIQ. We use " + }, + { + "bbox": [ + 54, + 594, + 295, + 713 + ], + "type": "inline_equation", + "content": "80\\%" + }, + { + "bbox": [ + 54, + 594, + 295, + 713 + ], + "type": "text", + "content": " of the images for training and " + }, + { + "bbox": [ + 54, + 594, + 295, + 713 + ], + "type": "inline_equation", + "content": "20\\%" + }, + { + "bbox": [ + 54, + 594, + 295, + 713 + ], + "type": "text", + "content": " for testing, repeating the process 10 times to reduce performance bias, reporting the average SRCC and PLCC. For synthetic distortion datasets, training and testing sets are divided by reference images to ensure con" + } + ] + } + ], + "index": 7 + }, + { + "type": "table", + "bbox": [ + 316, + 228, + 553, + 348 + ], + "blocks": [ + { + "bbox": [ + 55, + 356, + 295, + 378 + ], + "lines": [ + { + "bbox": [ + 55, + 356, + 295, + 378 + ], + "spans": [ + { + "bbox": [ + 55, + 356, + 295, + 378 + ], + "type": "text", + "content": "Table 3. SRCC on the cross datasets validation. The best results are highlighted in bold, second-best are underlined." + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 316, + 228, + 553, + 348 + ], + "lines": [ + { + "bbox": [ + 316, + 228, + 553, + 348 + ], + "spans": [ + { + "bbox": [ + 316, + 228, + 553, + 348 + ], + "type": "table", + "html": "
ModeMethodsLIVECKonIQ
PLCCSRCCPLCCSRCC
20%DEIQT0.8220.7920.9080.888
LoDa0.8540.8150.9230.907
SHDIQA0.8660.8400.9280.915
40%DEIQT0.8550.8380.9220.903
LoDa0.8790.8490.9350.922
SHDIQA0.8890.8690.9420.929
", + "image_path": "b22281aff0794c372ba2c5635b8e3d5ce8d2c6a9b9bde2a525d1a16575577b77.jpg" + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "table_body" + } + ], + "index": 8 + }, + { + "bbox": [ + 313, + 356, + 555, + 378 + ], + "lines": [ + { + "bbox": [ + 313, + 356, + 555, + 378 + ], + "spans": [ + { + "bbox": [ + 313, + 356, + 555, + 378 + ], + "type": "text", + "content": "Table 4. Data-efficient learning validation with the training set containing " + }, + { + "bbox": [ + 313, + 356, + 555, + 378 + ], + "type": "inline_equation", + "content": "20\\%" + }, + { + "bbox": [ + 313, + 356, + 555, + 378 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 313, + 356, + 555, + 378 + ], + "type": "inline_equation", + "content": "40\\%" + }, + { + "bbox": [ + 313, + 356, + 555, + 378 + ], + "type": "text", + "content": " of the images." + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "text" + }, + { + "bbox": [ + 313, + 385, + 555, + 445 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 385, + 555, + 445 + ], + "spans": [ + { + "bbox": [ + 313, + 385, + 555, + 445 + ], + "type": "text", + "content": "tent independence. To provide reliable distortion perception early in training, the teacher model adopts a similar encoder-decoder architecture aligned with the student model, and is pre-trained with " + }, + { + "bbox": [ + 313, + 385, + 555, + 445 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_1" + }, + { + "bbox": [ + 313, + 385, + 555, + 445 + ], + "type": "text", + "content": " loss on the same IQA dataset. Training is performed on four NVIDIA RTX 3090 GPUs." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 313, + 453, + 547, + 467 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 453, + 547, + 467 + ], + "spans": [ + { + "bbox": [ + 313, + 453, + 547, + 467 + ], + "type": "text", + "content": "4.3. Overall Prediction Performance Comparison" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 312, + 472, + 556, + 616 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 312, + 472, + 556, + 616 + ], + "spans": [ + { + "bbox": [ + 312, + 472, + 556, + 616 + ], + "type": "text", + "content": "Table 1 reports the comparison results between the proposed SHDIQA and other state-of-the-art (SOTA) BIQA methods, including ViT-based IQA methods, i.e., LoDa [50] and DEIQT [29], and Degradation Modelling-based methods, i.e., Re-IQA [31] and QPT [58]. It is observed from these seven datasets that SHDIQA achieves superior performance over all other methods across the 7 datasets. Notably, on the most difficult LIVEFB dataset, SHDIQA achieves an impressive performance improvement. This enhancement can be attributed to the prevalence of local distortions [53] in the LIVEFB dataset, highlighting the importance of effectively utilizing local distortion priors." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 313, + 624, + 500, + 636 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 624, + 500, + 636 + ], + "spans": [ + { + "bbox": [ + 313, + 624, + 500, + 636 + ], + "type": "text", + "content": "4.4. Data-Efficient Learning Validation" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 313, + 642, + 556, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 642, + 556, + 715 + ], + "spans": [ + { + "bbox": [ + 313, + 642, + 556, + 715 + ], + "type": "text", + "content": "We showcase the data efficiency of our approach, as presented in Table 4. Following the efficient validation protocol outlined in LoDa [50] and DEIQT [29], we train models using " + }, + { + "bbox": [ + 313, + 642, + 556, + 715 + ], + "type": "inline_equation", + "content": "20\\%" + }, + { + "bbox": [ + 313, + 642, + 556, + 715 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 313, + 642, + 556, + 715 + ], + "type": "inline_equation", + "content": "40\\%" + }, + { + "bbox": [ + 313, + 642, + 556, + 715 + ], + "type": "text", + "content": " of the available images. This process is repeated 10 times, and we report the average SRCC and PLCC scores. By introducing distortion priors through the" + } + ] + } + ], + "index": 14 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "text", + "content": "2349" + } + ] + } + ], + "index": 15 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 5 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 57, + 70, + 558, + 243 + ], + "blocks": [ + { + "bbox": [ + 57, + 70, + 558, + 243 + ], + "lines": [ + { + "bbox": [ + 57, + 70, + 558, + 243 + ], + "spans": [ + { + "bbox": [ + 57, + 70, + 558, + 243 + ], + "type": "image", + "image_path": "83a2df484a08816d56a9f482fda85550c2b6b0d455ef92fe04d66bbb9f711221.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 55, + 246, + 555, + 269 + ], + "lines": [ + { + "bbox": [ + 55, + 246, + 555, + 269 + ], + "spans": [ + { + "bbox": [ + 55, + 246, + 555, + 269 + ], + "type": "text", + "content": "Figure 4. gMAD competition results between DEIQT and our method SHDIQA. (a) Fixed DEIQT at low quality. (b) Fixed DEIQT at high quality. (c) Fixed SHDIQA at low quality. (d) Fixed SHDIQA at high quality. (e) Cosine similarity between different DIS and CLS tokens." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "type": "table", + "bbox": [ + 59, + 277, + 290, + 349 + ], + "blocks": [ + { + "bbox": [ + 59, + 277, + 290, + 349 + ], + "lines": [ + { + "bbox": [ + 59, + 277, + 290, + 349 + ], + "spans": [ + { + "bbox": [ + 59, + 277, + 290, + 349 + ], + "type": "table", + "html": "
LckdLIVECCSIQAvg.
PLCCSRCCPLCCSRCC
-0.8910.8670.9520.9350.911
L10.9080.8950.9570.9460.926
BCE0.9140.9020.9660.9550.934
", + "image_path": "a710c38bdbd4ffbccf7a7037c119b683d5e9267ca677046127944f9e750dd3b5.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "table_body" + } + ], + "index": 2 + }, + { + "bbox": [ + 55, + 372, + 296, + 433 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 372, + 296, + 433 + ], + "spans": [ + { + "bbox": [ + 55, + 372, + 296, + 433 + ], + "type": "text", + "content": "Block-wise Degradation, our method achieves performance that is nearly with SOTA results, even when trained on only " + }, + { + "bbox": [ + 55, + 372, + 296, + 433 + ], + "type": "inline_equation", + "content": "20\\%" + }, + { + "bbox": [ + 55, + 372, + 296, + 433 + ], + "type": "text", + "content": " of the dataset. Furthermore, our approach consistently outperforms all other methods on both datasets, demonstrating its superiority over existing efficient BIQA techniques." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 55, + 444, + 251, + 456 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 444, + 251, + 456 + ], + "spans": [ + { + "bbox": [ + 55, + 444, + 251, + 456 + ], + "type": "text", + "content": "4.5. Generalization Capability Validation" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 54, + 462, + 297, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 54, + 462, + 297, + 714 + ], + "spans": [ + { + "bbox": [ + 54, + 462, + 297, + 714 + ], + "type": "text", + "content": "We evaluate SHDIQA's generalization ability using cross-dataset validation, training the BIQA model on one dataset and testing it on others. Table 3 shows the SRCC averages for five datasets, where SHDIQA achieves the best performance in four out of six cross-dataset evaluations. To further test our framework's generalization, we trained the model on the LIVE database and tested it on the Waterloo database using the gMAD competition [24]. In the gMAD competition, image pairs are selected where the attacker model predicts a large quality difference, while the defender perceives them as similar. Observers then evaluate these pairs to assess the model's generalization ability. As shown in Fig. 4, our model correctly selects image pairs with significant quality differences when attacking DEIQT, where DEIQT misclassifies pairs like (a) as low quality and (b) as high quality, contrary to human perception and SHDIQA's predictions. As a defender, SHDIQA identifies pairs with minimal perceptual variation, correctly classifying image pairs like (c) and (d) as both low and high quality, respectively, while DEIQT fails. This demonstrates the strong generalization capability of SHDIQA in handling real-world image distortions." + } + ] + } + ], + "index": 6 + }, + { + "type": "table", + "bbox": [ + 317, + 277, + 550, + 350 + ], + "blocks": [ + { + "bbox": [ + 59, + 352, + 290, + 364 + ], + "lines": [ + { + "bbox": [ + 59, + 352, + 290, + 364 + ], + "spans": [ + { + "bbox": [ + 59, + 352, + 290, + 364 + ], + "type": "text", + "content": "Table 5. Ablation study on Contrastive Knowledge Distillation." + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 317, + 277, + 550, + 350 + ], + "lines": [ + { + "bbox": [ + 317, + 277, + 550, + 350 + ], + "spans": [ + { + "bbox": [ + 317, + 277, + 550, + 350 + ], + "type": "table", + "html": "
\\( {\\mathcal{L}}_{\\text{akd }}^{I} \\)\\( {\\mathcal{L}}_{\\text{akd }}^{I_d} \\)LIVECLIVEAvg.
PLCCSRCCPLCCSRCC
✓0.9170.9010.9820.9780.944
✓0.9150.9000.9800.9800.943
✓✓0.9220.9090.9840.9820.949
", + "image_path": "c679ff3b47c42586a8ad9fe585d0a9940790840c79ba3983cb5dc36e8c3ec34d.jpg" + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "table_body" + } + ], + "index": 7 + }, + { + "bbox": [ + 314, + 353, + 553, + 364 + ], + "lines": [ + { + "bbox": [ + 314, + 353, + 553, + 364 + ], + "spans": [ + { + "bbox": [ + 314, + 353, + 553, + 364 + ], + "type": "text", + "content": "Table 6. Ablation experiment on Affinity Knowledge Distillation." + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "text" + }, + { + "bbox": [ + 313, + 371, + 408, + 384 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 371, + 408, + 384 + ], + "spans": [ + { + "bbox": [ + 313, + 371, + 408, + 384 + ], + "type": "text", + "content": "4.6. Ablation Study" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 313, + 388, + 507, + 401 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 388, + 507, + 401 + ], + "spans": [ + { + "bbox": [ + 313, + 388, + 507, + 401 + ], + "type": "text", + "content": "4.6.1. Effectiveness of Proposed Components" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 313, + 403, + 556, + 632 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 403, + 556, + 632 + ], + "spans": [ + { + "bbox": [ + 313, + 403, + 556, + 632 + ], + "type": "text", + "content": "We conduct an ablation study on components of our framework. Table 2 shows that using the teacher's quality score as a regularization improves model performance. Adding a distortion classification task boosts SRCC by 1.2 and 1.8 points on KADID and LIVEC, respectively, highlighting the effectiveness of distortion-aware supervision. However, identifying distortions without modeling their relationships provides limited benefits. To address this, we introduce CKD and AKD to capture distortion correlations, improving SRCC by up to 3.5 and 4.9 points on KADID, respectively. The limited gain from CKD suggests that aggressively separating regions with different distortions may be problematic, as these regions can have similar quality. AKD mitigates this issue by capturing structural relationships between image blocks. When combined, CKD and AKD improve the baseline by 6.2 points, confirming their complementarity. Finally, Table 8 shows that block-wise aggregation and filtering further enhance the model's ability to assess local distortions, leading to better quality prediction." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 313, + 639, + 542, + 651 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 639, + 542, + 651 + ], + "spans": [ + { + "bbox": [ + 313, + 639, + 542, + 651 + ], + "type": "text", + "content": "4.6.2. Loss type of CKD and Implementation of AKD" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 313, + 654, + 555, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 654, + 555, + 713 + ], + "spans": [ + { + "bbox": [ + 313, + 654, + 555, + 713 + ], + "type": "text", + "content": "Table 5 shows the effect of different CKD loss functions. Using " + }, + { + "bbox": [ + 313, + 654, + 555, + 713 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_1" + }, + { + "bbox": [ + 313, + 654, + 555, + 713 + ], + "type": "text", + "content": " loss improves average quality prediction by 1.5 points, and switching to a BCE-based contrastive loss adds another 0.8-point gain. This confirms the effectiveness of our CKD design in promoting feature clustering within the" + } + ] + } + ], + "index": 13 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "text", + "content": "2350" + } + ] + } + ], + "index": 14 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 6 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 58, + 69, + 170, + 144 + ], + "blocks": [ + { + "bbox": [ + 58, + 69, + 170, + 144 + ], + "lines": [ + { + "bbox": [ + 58, + 69, + 170, + 144 + ], + "spans": [ + { + "bbox": [ + 58, + 69, + 170, + 144 + ], + "type": "table", + "html": "
NumberLIVEC
PLCCSRCC
30.9200.903
40.9220.909
50.9180.905
", + "image_path": "d00235f51755c22057568c83492e9234aaeb63e920a387f7e0d35af991e7ca5a.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "table_body" + } + ], + "index": 0 + }, + { + "type": "table", + "bbox": [ + 181, + 69, + 356, + 144 + ], + "blocks": [ + { + "bbox": [ + 55, + 146, + 172, + 169 + ], + "lines": [ + { + "bbox": [ + 55, + 146, + 172, + 169 + ], + "spans": [ + { + "bbox": [ + 55, + 146, + 172, + 169 + ], + "type": "text", + "content": "Table 7. Ablation study on the number of degenerated blocks." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 181, + 69, + 356, + 144 + ], + "lines": [ + { + "bbox": [ + 181, + 69, + 356, + 144 + ], + "spans": [ + { + "bbox": [ + 181, + 69, + 356, + 144 + ], + "type": "table", + "html": "
MethodLIVECLIVE
PLCCSRCCPLCCSRCC
Baseline0.8910.8670.9660.964
+BWA0.9180.9020.9820.978
+BWF0.9220.9090.9840.982
", + "image_path": "cf2e719e3a752e9c3e0363f602bb4e0e6049aa552ecb327064cd418cd37afd27.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "table_body" + } + ], + "index": 2 + }, + { + "type": "table", + "bbox": [ + 365, + 70, + 550, + 144 + ], + "blocks": [ + { + "bbox": [ + 180, + 146, + 358, + 169 + ], + "lines": [ + { + "bbox": [ + 180, + 146, + 358, + 169 + ], + "spans": [ + { + "bbox": [ + 180, + 146, + 358, + 169 + ], + "type": "text", + "content": "Table 8. Ablation study on the Block-wise Aggregation (BwA) and Filtering (BWF) module." + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 365, + 70, + 550, + 144 + ], + "lines": [ + { + "bbox": [ + 365, + 70, + 550, + 144 + ], + "spans": [ + { + "bbox": [ + 365, + 70, + 550, + 144 + ], + "type": "table", + "html": "
MethodLIVECKonIQ
PLCCSRCCPLCCSRCC
Homogenous0.9070.8930.9390.930
Heterogenous0.9220.9090.9480.937
", + "image_path": "eabb7936609a32d4b4ada431af16c7ba008ebeb23a15fbec7beaed4773ac4e39.jpg" + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "table_body" + } + ], + "index": 4 + }, + { + "bbox": [ + 364, + 146, + 552, + 169 + ], + "lines": [ + { + "bbox": [ + 364, + 146, + 552, + 169 + ], + "spans": [ + { + "bbox": [ + 364, + 146, + 552, + 169 + ], + "type": "text", + "content": "Table 9. Ablation experiments on heterogeneous and homogeneous distortion on the whole image." + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "text" + }, + { + "type": "image", + "bbox": [ + 56, + 174, + 127, + 243 + ], + "blocks": [ + { + "bbox": [ + 56, + 174, + 127, + 243 + ], + "lines": [ + { + "bbox": [ + 56, + 174, + 127, + 243 + ], + "spans": [ + { + "bbox": [ + 56, + 174, + 127, + 243 + ], + "type": "image", + "image_path": "cd911aaddb4ab589b11880f0ab647181721e8e23827f2105791ad7b8b47f309a.jpg" + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 69, + 245, + 115, + 255 + ], + "lines": [ + { + "bbox": [ + 69, + 245, + 115, + 255 + ], + "spans": [ + { + "bbox": [ + 69, + 245, + 115, + 255 + ], + "type": "text", + "content": "Mos: 38.6" + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "image_caption" + } + ], + "index": 6 + }, + { + "type": "image", + "bbox": [ + 127, + 175, + 197, + 243 + ], + "blocks": [ + { + "bbox": [ + 127, + 175, + 197, + 243 + ], + "lines": [ + { + "bbox": [ + 127, + 175, + 197, + 243 + ], + "spans": [ + { + "bbox": [ + 127, + 175, + 197, + 243 + ], + "type": "image", + "image_path": "3b328d8794657a9a71287374a359e4f8815802dc3decb6d75b3c9e632ccfcbad.jpg" + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 131, + 245, + 193, + 255 + ], + "lines": [ + { + "bbox": [ + 131, + 245, + 193, + 255 + ], + "spans": [ + { + "bbox": [ + 131, + 245, + 193, + 255 + ], + "type": "text", + "content": "Baseline: 46.1" + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "image_caption" + } + ], + "index": 8 + }, + { + "type": "image", + "bbox": [ + 199, + 175, + 268, + 243 + ], + "blocks": [ + { + "bbox": [ + 199, + 175, + 268, + 243 + ], + "lines": [ + { + "bbox": [ + 199, + 175, + 268, + 243 + ], + "spans": [ + { + "bbox": [ + 199, + 175, + 268, + 243 + ], + "type": "image", + "image_path": "f33fc90b237c0ac63aff54890d79b167d2ecb0c369f1a3993a18d131cce0d8c5.jpg" + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 198, + 245, + 269, + 256 + ], + "lines": [ + { + "bbox": [ + 198, + 245, + 269, + 256 + ], + "spans": [ + { + "bbox": [ + 198, + 245, + 269, + 256 + ], + "type": "text", + "content": "Ours: 35.2(↓4.1)" + } + ] + } + ], + "index": 11, + "angle": 0, + "type": "image_caption" + } + ], + "index": 10 + }, + { + "type": "image", + "bbox": [ + 271, + 175, + 340, + 243 + ], + "blocks": [ + { + "bbox": [ + 271, + 175, + 340, + 243 + ], + "lines": [ + { + "bbox": [ + 271, + 175, + 340, + 243 + ], + "spans": [ + { + "bbox": [ + 271, + 175, + 340, + 243 + ], + "type": "image", + "image_path": "19b81108dfdef2c055ffbe724ec8a5a9cf7fcb30a6bf8e273748dcd2f818fcbc.jpg" + } + ] + } + ], + "index": 12, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 279, + 245, + 331, + 255 + ], + "lines": [ + { + "bbox": [ + 279, + 245, + 331, + 255 + ], + "spans": [ + { + "bbox": [ + 279, + 245, + 331, + 255 + ], + "type": "text", + "content": "DIS token 1" + } + ] + } + ], + "index": 13, + "angle": 0, + "type": "image_caption" + } + ], + "index": 12 + }, + { + "type": "image", + "bbox": [ + 342, + 175, + 411, + 243 + ], + "blocks": [ + { + "bbox": [ + 342, + 175, + 411, + 243 + ], + "lines": [ + { + "bbox": [ + 342, + 175, + 411, + 243 + ], + "spans": [ + { + "bbox": [ + 342, + 175, + 411, + 243 + ], + "type": "image", + "image_path": "67aad0ec5497634f6afcfcae966a8b4e18f1cc9f555ff233bb9a681fe2e01d58.jpg" + } + ] + } + ], + "index": 14, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 351, + 245, + 405, + 255 + ], + "lines": [ + { + "bbox": [ + 351, + 245, + 405, + 255 + ], + "spans": [ + { + "bbox": [ + 351, + 245, + 405, + 255 + ], + "type": "text", + "content": "DIS token 2" + } + ] + } + ], + "index": 15, + "angle": 0, + "type": "image_caption" + } + ], + "index": 14 + }, + { + "type": "image", + "bbox": [ + 413, + 175, + 482, + 243 + ], + "blocks": [ + { + "bbox": [ + 413, + 175, + 482, + 243 + ], + "lines": [ + { + "bbox": [ + 413, + 175, + 482, + 243 + ], + "spans": [ + { + "bbox": [ + 413, + 175, + 482, + 243 + ], + "type": "image", + "image_path": "34dc853aff7cdea3663b2bf65af3d1ec88113d548f78775f7b4365ded43ca1dc.jpg" + } + ] + } + ], + "index": 16, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 420, + 245, + 473, + 255 + ], + "lines": [ + { + "bbox": [ + 420, + 245, + 473, + 255 + ], + "spans": [ + { + "bbox": [ + 420, + 245, + 473, + 255 + ], + "type": "text", + "content": "DIS token 3" + } + ] + } + ], + "index": 17, + "angle": 0, + "type": "image_caption" + } + ], + "index": 16 + }, + { + "type": "image", + "bbox": [ + 484, + 175, + 553, + 243 + ], + "blocks": [ + { + "bbox": [ + 484, + 175, + 553, + 243 + ], + "lines": [ + { + "bbox": [ + 484, + 175, + 553, + 243 + ], + "spans": [ + { + "bbox": [ + 484, + 175, + 553, + 243 + ], + "type": "image", + "image_path": "e3acd117a948e850ba952908e4ab91173f16c23e6d21bcbff26f0d4f7815cf79.jpg" + } + ] + } + ], + "index": 18, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 492, + 245, + 545, + 255 + ], + "lines": [ + { + "bbox": [ + 492, + 245, + 545, + 255 + ], + "spans": [ + { + "bbox": [ + 492, + 245, + 545, + 255 + ], + "type": "text", + "content": "DIS token 4" + } + ] + } + ], + "index": 19, + "angle": 0, + "type": "image_caption" + } + ], + "index": 18 + }, + { + "type": "image", + "bbox": [ + 56, + 258, + 127, + 327 + ], + "blocks": [ + { + "bbox": [ + 56, + 258, + 127, + 327 + ], + "lines": [ + { + "bbox": [ + 56, + 258, + 127, + 327 + ], + "spans": [ + { + "bbox": [ + 56, + 258, + 127, + 327 + ], + "type": "image", + "image_path": "2bdf871e3b18832158a09fc313ee57f14d7518789f73347fa955e91eb6440c43.jpg" + } + ] + } + ], + "index": 20, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 69, + 328, + 115, + 339 + ], + "lines": [ + { + "bbox": [ + 69, + 328, + 115, + 339 + ], + "spans": [ + { + "bbox": [ + 69, + 328, + 115, + 339 + ], + "type": "text", + "content": "Mos: 57.3" + } + ] + } + ], + "index": 21, + "angle": 0, + "type": "image_caption" + } + ], + "index": 20 + }, + { + "type": "image", + "bbox": [ + 128, + 258, + 197, + 326 + ], + "blocks": [ + { + "bbox": [ + 128, + 258, + 197, + 326 + ], + "lines": [ + { + "bbox": [ + 128, + 258, + 197, + 326 + ], + "spans": [ + { + "bbox": [ + 128, + 258, + 197, + 326 + ], + "type": "image", + "image_path": "30cb86e4aa1ecbf27961745e0619307766e260c710dbef942164040c7ca8fb7f.jpg" + } + ] + } + ], + "index": 22, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 132, + 328, + 194, + 339 + ], + "lines": [ + { + "bbox": [ + 132, + 328, + 194, + 339 + ], + "spans": [ + { + "bbox": [ + 132, + 328, + 194, + 339 + ], + "type": "text", + "content": "Baseline: 67.1" + } + ] + } + ], + "index": 23, + "angle": 0, + "type": "image_caption" + } + ], + "index": 22 + }, + { + "type": "image", + "bbox": [ + 200, + 258, + 268, + 326 + ], + "blocks": [ + { + "bbox": [ + 200, + 258, + 268, + 326 + ], + "lines": [ + { + "bbox": [ + 200, + 258, + 268, + 326 + ], + "spans": [ + { + "bbox": [ + 200, + 258, + 268, + 326 + ], + "type": "image", + "image_path": "dff87b13cfc9008414b9985dd281b6746ebfd3412ad2109e3e0c1a4c87376d3b.jpg" + } + ] + } + ], + "index": 24, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 199, + 328, + 271, + 341 + ], + "lines": [ + { + "bbox": [ + 199, + 328, + 271, + 341 + ], + "spans": [ + { + "bbox": [ + 199, + 328, + 271, + 341 + ], + "type": "text", + "content": "Ours: 60.9(↓6.2)" + } + ] + } + ], + "index": 25, + "angle": 0, + "type": "image_caption" + } + ], + "index": 24 + }, + { + "type": "image", + "bbox": [ + 271, + 258, + 339, + 326 + ], + "blocks": [ + { + "bbox": [ + 271, + 258, + 339, + 326 + ], + "lines": [ + { + "bbox": [ + 271, + 258, + 339, + 326 + ], + "spans": [ + { + "bbox": [ + 271, + 258, + 339, + 326 + ], + "type": "image", + "image_path": "35b5e774e7f7bd8e53024058fd9b264865bbb999d3fb44d8047d5dea73d11e94.jpg" + } + ] + } + ], + "index": 26, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 279, + 328, + 331, + 339 + ], + "lines": [ + { + "bbox": [ + 279, + 328, + 331, + 339 + ], + "spans": [ + { + "bbox": [ + 279, + 328, + 331, + 339 + ], + "type": "text", + "content": "DIS token 1" + } + ] + } + ], + "index": 27, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 55, + 343, + 555, + 367 + ], + "lines": [ + { + "bbox": [ + 55, + 343, + 555, + 367 + ], + "spans": [ + { + "bbox": [ + 55, + 343, + 555, + 367 + ], + "type": "text", + "content": "Figure 5. Activation maps of DEIQT [29] and SHDIQA using Grad-CAM [32] on the LIVEFB dataset, with MOS shown below the images. Our SHDIQA effectively captures the interplay between image content, local distortions, and overall quality, enabling accurate predictions." + } + ] + } + ], + "index": 34, + "angle": 0, + "type": "image_caption" + } + ], + "index": 26 + }, + { + "type": "image", + "bbox": [ + 342, + 258, + 411, + 326 + ], + "blocks": [ + { + "bbox": [ + 342, + 258, + 411, + 326 + ], + "lines": [ + { + "bbox": [ + 342, + 258, + 411, + 326 + ], + "spans": [ + { + "bbox": [ + 342, + 258, + 411, + 326 + ], + "type": "image", + "image_path": "58a2b4a8db735cbd97cbebf58fecfd33662852ee34a4c23c330ee46bf666cff4.jpg" + } + ] + } + ], + "index": 28, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 353, + 328, + 405, + 339 + ], + "lines": [ + { + "bbox": [ + 353, + 328, + 405, + 339 + ], + "spans": [ + { + "bbox": [ + 353, + 328, + 405, + 339 + ], + "type": "text", + "content": "Dis token 2" + } + ] + } + ], + "index": 29, + "angle": 0, + "type": "image_caption" + } + ], + "index": 28 + }, + { + "type": "image", + "bbox": [ + 414, + 258, + 482, + 326 + ], + "blocks": [ + { + "bbox": [ + 414, + 258, + 482, + 326 + ], + "lines": [ + { + "bbox": [ + 414, + 258, + 482, + 326 + ], + "spans": [ + { + "bbox": [ + 414, + 258, + 482, + 326 + ], + "type": "image", + "image_path": "e3c74f4522d92d9b969f24f223326929dac7d111f09947406b4aca747367403b.jpg" + } + ] + } + ], + "index": 30, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 421, + 328, + 474, + 339 + ], + "lines": [ + { + "bbox": [ + 421, + 328, + 474, + 339 + ], + "spans": [ + { + "bbox": [ + 421, + 328, + 474, + 339 + ], + "type": "text", + "content": "DIS token 3" + } + ] + } + ], + "index": 31, + "angle": 0, + "type": "image_caption" + } + ], + "index": 30 + }, + { + "type": "image", + "bbox": [ + 485, + 258, + 553, + 326 + ], + "blocks": [ + { + "bbox": [ + 485, + 258, + 553, + 326 + ], + "lines": [ + { + "bbox": [ + 485, + 258, + 553, + 326 + ], + "spans": [ + { + "bbox": [ + 485, + 258, + 553, + 326 + ], + "type": "image", + "image_path": "abf21518b59b470a9d7d64cea86fedf7a5f0c06c197edcead7dcd849b04c50a8.jpg" + } + ] + } + ], + "index": 32, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 492, + 328, + 545, + 339 + ], + "lines": [ + { + "bbox": [ + 492, + 328, + 545, + 339 + ], + "spans": [ + { + "bbox": [ + 492, + 328, + 545, + 339 + ], + "type": "text", + "content": "DIS token 4" + } + ] + } + ], + "index": 33, + "angle": 0, + "type": "image_caption" + } + ], + "index": 32 + }, + { + "bbox": [ + 54, + 372, + 297, + 468 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 54, + 372, + 297, + 468 + ], + "spans": [ + { + "bbox": [ + 54, + 372, + 297, + 468 + ], + "type": "text", + "content": "same distortion type and separation across types. For AKD, we evaluate three strategies: applying it to the input image " + }, + { + "bbox": [ + 54, + 372, + 297, + 468 + ], + "type": "inline_equation", + "content": "\\mathbf{I}" + }, + { + "bbox": [ + 54, + 372, + 297, + 468 + ], + "type": "text", + "content": ", the degraded image " + }, + { + "bbox": [ + 54, + 372, + 297, + 468 + ], + "type": "inline_equation", + "content": "\\mathbf{I}_d" + }, + { + "bbox": [ + 54, + 372, + 297, + 468 + ], + "type": "text", + "content": ", or both. Table 6 shows that the best performance is achieved by using both, as applying AKD to the image " + }, + { + "bbox": [ + 54, + 372, + 297, + 468 + ], + "type": "inline_equation", + "content": "\\mathbf{I}" + }, + { + "bbox": [ + 54, + 372, + 297, + 468 + ], + "type": "text", + "content": " helps model relationships within local content, while applying it to the " + }, + { + "bbox": [ + 54, + 372, + 297, + 468 + ], + "type": "inline_equation", + "content": "\\mathbf{I}_d" + }, + { + "bbox": [ + 54, + 372, + 297, + 468 + ], + "type": "text", + "content": " captures correlations between different types of distortions. Combining both enables complementary learning of content and distortion information." + } + ] + } + ], + "index": 35 + }, + { + "bbox": [ + 55, + 472, + 234, + 484 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 472, + 234, + 484 + ], + "spans": [ + { + "bbox": [ + 55, + 472, + 234, + 484 + ], + "type": "text", + "content": "4.6.3. Ablation on the Degradation Space" + } + ] + } + ], + "index": 36 + }, + { + "bbox": [ + 55, + 487, + 297, + 619 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 487, + 297, + 619 + ], + "spans": [ + { + "bbox": [ + 55, + 487, + 297, + 619 + ], + "type": "text", + "content": "We investigate the effectiveness of the Block-wise Degradation Modelling. As shown in Table 9, applying globally uniform heterogeneous distortions significantly underperforms compared to localized distortions, highlighting the importance of local degradation. We also evaluate how varying the number of degraded blocks affects performance. As shown in Table 7, the impact is relatively insensitive across different partitioning strategies, demonstrating the method's resilience to changes in block quantity. While more blocks increase the scope of degradation, they also lead to greater content loss. Thus, using four blocks strikes the best balance." + } + ] + } + ], + "index": 37 + }, + { + "bbox": [ + 55, + 624, + 173, + 637 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 624, + 173, + 637 + ], + "spans": [ + { + "bbox": [ + 55, + 624, + 173, + 637 + ], + "type": "text", + "content": "4.7. Qualitative Analysis" + } + ] + } + ], + "index": 38 + }, + { + "bbox": [ + 55, + 642, + 297, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 642, + 297, + 715 + ], + "spans": [ + { + "bbox": [ + 55, + 642, + 297, + 715 + ], + "type": "text", + "content": "To show that tokens from different perspectives capture unique features, we analyze the cosine similarity between CLS and DIS tokens on the LIVEC and KonIQ datasets. As shown in Fig. 4, the similarities range from 0.72 to 0.81, lower than the 0.96 and 0.94 found in DeiT-T and DeiT-S [42]. Tokens near the center show higher similarity," + } + ] + } + ], + "index": 39 + }, + { + "bbox": [ + 313, + 372, + 555, + 529 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 372, + 555, + 529 + ], + "spans": [ + { + "bbox": [ + 313, + 372, + 555, + 529 + ], + "type": "text", + "content": "aligning with the HVS's focus on salient regions [62]. This implies that DIS tokens capture localized distortions in a way that simulates human perception, and when combined with CLS tokens, it improves both global and local distortion awareness. To further explore the DIS tokens, we visualize the attention maps of DEIQT and our SHDIQA, as shown in Fig. 5. By incorporating four additional distortion tokens, the proposed model enhances its ability to detect distortions across different regions of the image. Through the interaction between distortion tokens and CLS tokens in the adaptation head, the model adaptively focuses on the most relevant areas for image quality, while also considering the overall content. This results in more accurate quality predictions." + } + ] + } + ], + "index": 40 + }, + { + "bbox": [ + 313, + 540, + 388, + 553 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 540, + 388, + 553 + ], + "spans": [ + { + "bbox": [ + 313, + 540, + 388, + 553 + ], + "type": "text", + "content": "5. Conclusion" + } + ] + } + ], + "index": 41 + }, + { + "bbox": [ + 313, + 558, + 556, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 558, + 556, + 715 + ], + "spans": [ + { + "bbox": [ + 313, + 558, + 556, + 715 + ], + "type": "text", + "content": "In this paper, we propose SHDIQA, a novel BIQA approach that leverages spatial-heterogeneous distortion-aware strategies. Our Block-Wise Degradation Modelling generates locally distorted images, exposing the model to diverse distortion patterns. The Block-wise Aggregation and Filtering mechanism enhances the ViT model's sensitivity to local distortions, improving quality assessments through focused attention and feature filtering. Additionally, our Distortion-Guided Knowledge Distillation combines CKD and AKD to improve distortion discrimination and relational learning. Experimental results show SHDIQA's superior generalization and robustness across various datasets and distortion types, confirming its effectiveness in real-world BIQA tasks." + } + ] + } + ], + "index": 42 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 748, + 314, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 314, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 314, + 757 + ], + "type": "text", + "content": "2351" + } + ] + } + ], + "index": 43 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 7 + }, + { + "para_blocks": [ + { + "bbox": [ + 56, + 190, + 115, + 202 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 190, + 115, + 202 + ], + "spans": [ + { + "bbox": [ + 56, + 190, + 115, + 202 + ], + "type": "text", + "content": "References" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 57, + 210, + 296, + 713 + ], + "type": "list", + "angle": 0, + "index": 13, + "blocks": [ + { + "bbox": [ + 61, + 210, + 296, + 243 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 61, + 210, + 296, + 243 + ], + "spans": [ + { + "bbox": [ + 61, + 210, + 296, + 243 + ], + "type": "text", + "content": "[1] Mark R Banham and Aggelos K Katsaggelos. Digital image restoration. IEEE signal processing magazine, 14(2):24-41, 1997. 1" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 61, + 244, + 296, + 298 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 61, + 244, + 296, + 298 + ], + "spans": [ + { + "bbox": [ + 61, + 244, + 296, + 298 + ], + "type": "text", + "content": "[2] Chaofeng Chen, Jiadi Mo, Jingwen Hou, Haoning Wu, Liang Liao, Wenxiu Sun, Qiong Yan, and Weisi Lin. Topiq: A top-down approach from semantics to distortions for image quality assessment. IEEE Transactions on Image Processing, 2024. 3" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 62, + 300, + 296, + 332 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 300, + 296, + 332 + ], + "spans": [ + { + "bbox": [ + 62, + 300, + 296, + 332 + ], + "type": "text", + "content": "[3] Xinlei Chen, Haoqi Fan, Ross Girshick, and Kaiming He. Improved baselines with momentum contrastive learning. arXiv preprint arXiv:2003.04297, 2020. 2" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 62, + 334, + 296, + 378 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 334, + 296, + 378 + ], + "spans": [ + { + "bbox": [ + 62, + 334, + 296, + 378 + ], + "type": "text", + "content": "[4] Jia Deng, Wei Dong, Richard Socher, Li-Jia Li, Kai Li, and Li Fei-Fei. Imagenet: A large-scale hierarchical image database. In 2009 IEEE conference on computer vision and pattern recognition, pages 248-255. IEEE, 2009. 4" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 62, + 379, + 296, + 422 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 379, + 296, + 422 + ], + "spans": [ + { + "bbox": [ + 62, + 379, + 296, + 422 + ], + "type": "text", + "content": "[5] Jinhong Deng, Wen Li, Yuhua Chen, and Lixin Duan. Unbiased mean teacher for cross-domain object detection. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 4091-4101, 2021. 4" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 62, + 424, + 296, + 466 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 424, + 296, + 466 + ], + "spans": [ + { + "bbox": [ + 62, + 424, + 296, + 466 + ], + "type": "text", + "content": "[6] Chao Dong, Chen Change Loy, Kaiming He, and Xiaou Tang. Image super-resolution using deep convolutional networks. IEEE transactions on pattern analysis and machine intelligence, 38(2):295-307, 2015. 1" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 62, + 468, + 296, + 521 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 468, + 296, + 521 + ], + "spans": [ + { + "bbox": [ + 62, + 468, + 296, + 521 + ], + "type": "text", + "content": "[7] Yuming Fang, Hanwei Zhu, Yan Zeng, Kede Ma, and Zhou Wang. Perceptual quality assessment of smartphone photography. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 3677-3686, 2020. 5" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 62, + 524, + 296, + 567 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 524, + 296, + 567 + ], + "spans": [ + { + "bbox": [ + 62, + 524, + 296, + 567 + ], + "type": "text", + "content": "[8] Nuno C Garcia, Pietro Morerio, and Vittorio Murino. Modality distillation with multiple stream networks for action recognition. In Proceedings of the European Conference on Computer Vision (ECCV), pages 103-118, 2018. 3" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 62, + 569, + 296, + 612 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 569, + 296, + 612 + ], + "spans": [ + { + "bbox": [ + 62, + 569, + 296, + 612 + ], + "type": "text", + "content": "[9] Deepti Ghadiyaram and Alan C Bovik. Massive online crowdsourced study of subjective and objective picture quality. IEEE Transactions on Image Processing, 25(1):372-387, 2015. 5" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 57, + 614, + 296, + 668 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 614, + 296, + 668 + ], + "spans": [ + { + "bbox": [ + 57, + 614, + 296, + 668 + ], + "type": "text", + "content": "[10] S Alireza Golestaneh, Saba Dadsetan, and Kris M Kitani. No-reference image quality assessment via transformers, relative ranking, and self-consistency. In Proceedings of the IEEE/CVF Winter Conference on Applications of Computer Vision, pages 1220-1230, 2022. 2, 3, 5" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 57, + 670, + 296, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 670, + 296, + 713 + ], + "spans": [ + { + "bbox": [ + 57, + 670, + 296, + 713 + ], + "type": "text", + "content": "[11] Saurabh Gupta, Judy Hoffman, and Jitendra Malik. Cross modal distillation for supervision transfer. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 2827-2836, 2016. 3" + } + ] + } + ], + "index": 12 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 57, + 72, + 555, + 713 + ], + "type": "list", + "angle": 0, + "index": 27, + "blocks": [ + { + "bbox": [ + 55, + 72, + 297, + 180 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 72, + 297, + 180 + ], + "spans": [ + { + "bbox": [ + 55, + 72, + 297, + 180 + ], + "type": "text", + "content": "Acknowledgements. This work was supported by the National Science Fund for Distinguished Young Scholars (No.62025603), the National Natural Science Foundation of China (No. U21B2037, No. U22B2051, No. U23A20383, No. 62176222, No. 62176223, No. 62176226, No. 62072386, No. 62072387, No. 62072389, No. 62002305 and No. 62272401), and the Natural Science Foundation of Fujian Province of China (No. 2021J06003, No.2022J06001)." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 316, + 73, + 555, + 105 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 73, + 555, + 105 + ], + "spans": [ + { + "bbox": [ + 316, + 73, + 555, + 105 + ], + "type": "text", + "content": "[12] Geoffrey Hinton, Oriol Vinyals, and Jeff Dean. Distilling the knowledge in a neural network. arXiv preprint arXiv:1503.02531, 2015. 3" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 317, + 107, + 555, + 150 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 107, + 555, + 150 + ], + "spans": [ + { + "bbox": [ + 317, + 107, + 555, + 150 + ], + "type": "text", + "content": "[13] Vlad Hosu, Hanhe Lin, Tamas Sziranyi, and Dietmar Saupe. Koniq-10k: An ecologically valid database for deep learning of blind image quality assessment. IEEE Transactions on Image Processing, 29:4041-4056, 2020. 1, 5" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 316, + 152, + 555, + 196 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 152, + 555, + 196 + ], + "spans": [ + { + "bbox": [ + 316, + 152, + 555, + 196 + ], + "type": "text", + "content": "[14] Junjie Ke, Qifei Wang, Yilin Wang, Peyman Milanfar, and Feng Yang. Musiq: Multi-scale image quality transformer. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 5148-5157, 2021. 2, 3, 5" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 316, + 198, + 555, + 240 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 198, + 555, + 240 + ], + "spans": [ + { + "bbox": [ + 316, + 198, + 555, + 240 + ], + "type": "text", + "content": "[15] Eric C Larson and Damon M Chandler. Most apparent distortion: full-reference image quality assessment and the role of strategy. Journal of electronic imaging, 19(1):011006-011006, 2010. 2, 5" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 316, + 243, + 555, + 286 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 243, + 555, + 286 + ], + "spans": [ + { + "bbox": [ + 316, + 243, + 555, + 286 + ], + "type": "text", + "content": "[16] Aobo Li, Jinjian Wu, Yongxu Liu, and Leida Li. Bridging the synthetic-to Authentic gap: Distortion-guided unsupervised domain adaptation for blind image quality assessment. arXiv preprint arXiv:2405.04167, 2024. 2" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 316, + 288, + 555, + 342 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 288, + 555, + 342 + ], + "spans": [ + { + "bbox": [ + 316, + 288, + 555, + 342 + ], + "type": "text", + "content": "[17] Leida Li, Tianshu Song, Jinjian Wu, Weisheng Dong, Jiansheng Qian, and Guangming Shi. Blind image quality index for authentic distortions with local and global deep feature aggregation. IEEE Transactions on Circuits and Systems for Video Technology, 32(12):8512-8523, 2021. 2, 3" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 316, + 344, + 555, + 398 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 344, + 555, + 398 + ], + "spans": [ + { + "bbox": [ + 316, + 344, + 555, + 398 + ], + "type": "text", + "content": "[18] Liangqi Li, Jiaxu Miao, Dahu Shi, Wenming Tan, Ye Ren, Yi Yang, and Shiliang Pu. Distilling detr with visual-linguistic knowledge for open-vocabulary object detection. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 6501–6510, 2023. 4" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 317, + 399, + 555, + 466 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 399, + 555, + 466 + ], + "spans": [ + { + "bbox": [ + 317, + 399, + 555, + 466 + ], + "type": "text", + "content": "[19] Xudong Li, Timin Gao, Runze Hu, Yan Zhang, Shengchuan Zhang, Xiawu Zheng, Jingyuan Zheng, Yunhang Shen, Ke Li, Yutao Liu, et al. Adaptive feature selection for no-reference image quality assessment by mitigating semantic noise sensitivity. In International Conference on Machine Learning, pages 27808-27821. PMLR, 2024. 4, 5" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 317, + 468, + 555, + 533 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 468, + 555, + 533 + ], + "spans": [ + { + "bbox": [ + 317, + 468, + 555, + 533 + ], + "type": "text", + "content": "[20] Xudong Li, Runze Hu, Jingyuan Zheng, Yan Zhang, Shengchuan Zhang, Xiawu Zheng, Ke Li, Yunhang Shen, Yutao Liu, Pingyang Dai, et al. Integrating global context contrast and local sensitivity for blind image quality assessment. In International Conference on Machine Learning, pages 27920-27941. PMLR, 2024. 3" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 317, + 534, + 555, + 588 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 534, + 555, + 588 + ], + "spans": [ + { + "bbox": [ + 317, + 534, + 555, + 588 + ], + "type": "text", + "content": "[21] Xudong Li, Jingyuan Zheng, Runze Hu, Yan Zhang, Ke Li, Yunhang Shen, Xiawu Zheng, Yutao Liu, ShengChuan Zhang, Pingyang Dai, et al. Feature denoising diffusion model for blind image quality assessment. arXiv preprint arXiv:2401.11949, 2024.3" + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 317, + 590, + 555, + 634 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 590, + 555, + 634 + ], + "spans": [ + { + "bbox": [ + 317, + 590, + 555, + 634 + ], + "type": "text", + "content": "[22] Hanhe Lin, Vlad Hosu, and Dietmar Saupe. Kadid-10k: A large-scale artificially distorted iqa database. In 2019 Eleventh International Conference on Quality of Multimedia Experience (QoMEX), pages 1-3. IEEE, 2019. 5" + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 316, + 635, + 555, + 668 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 635, + 555, + 668 + ], + "spans": [ + { + "bbox": [ + 316, + 635, + 555, + 668 + ], + "type": "text", + "content": "[23] Hao Liu and Pieter Abbeel. Blockwise parallel transformers for large context models. Advances in Neural Information Processing Systems, 36, 2024. 3" + } + ] + } + ], + "index": 25 + }, + { + "bbox": [ + 316, + 670, + 555, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 670, + 555, + 713 + ], + "spans": [ + { + "bbox": [ + 316, + 670, + 555, + 713 + ], + "type": "text", + "content": "[24] Kede Ma, Qingbo Wu, Zhou Wang, Zhengfang Duanmu, Hongwei Yong, Hongliang Li, and Lei Zhang. Group mad competition-a new methodology to compare objective image quality models. In Proceedings of the IEEE Conference on" + } + ] + } + ], + "index": 26 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "text", + "content": "2352" + } + ] + } + ], + "index": 28 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 8 + }, + { + "para_blocks": [ + { + "bbox": [ + 56, + 73, + 296, + 713 + ], + "type": "list", + "angle": 0, + "index": 14, + "blocks": [ + { + "bbox": [ + 77, + 73, + 296, + 94 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 77, + 73, + 296, + 94 + ], + "spans": [ + { + "bbox": [ + 77, + 73, + 296, + 94 + ], + "type": "text", + "content": "Computer Vision and Pattern Recognition, pages 1664-1673, 2016. 7" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 56, + 96, + 296, + 140 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 96, + 296, + 140 + ], + "spans": [ + { + "bbox": [ + 56, + 96, + 296, + 140 + ], + "type": "text", + "content": "[25] Kede Ma, Wentao Liu, Tongliang Liu, Zhou Wang, and Dacheng Tao. dipiq: Blind image quality assessment by learning-to-rank discriminable image pairs. IEEE Transactions on Image Processing, 26(8):3951-3964, 2017. 2" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 56, + 141, + 294, + 185 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 141, + 294, + 185 + ], + "spans": [ + { + "bbox": [ + 56, + 141, + 294, + 185 + ], + "type": "text", + "content": "[26] Kede Ma, Wentao Liu, Kai Zhang, Zhengfang Duanmu, Zhou Wang, and Wangmeng Zuo. End-to-end blind image quality assessment using deep neural networks. IEEE Transactions on Image Processing, 27(3):1202-1213, 2017. 2" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 56, + 186, + 295, + 229 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 186, + 295, + 229 + ], + "spans": [ + { + "bbox": [ + 56, + 186, + 295, + 229 + ], + "type": "text", + "content": "[27] Anish Mittal, Anush Krishna Moorthy, and Alan Conrad Bovik. No-reference image quality assessment in the spatial domain. IEEE Transactions on image processing, 21(12): 4695-4708, 2012. 5" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 56, + 232, + 296, + 286 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 232, + 296, + 286 + ], + "spans": [ + { + "bbox": [ + 56, + 232, + 296, + 286 + ], + "type": "text", + "content": "[28] Angelo Porrello, Luca Bergamini, and Simone Calderara. Robust re-identification by multiple views knowledge distillation. In Computer Vision-ECCV 2020: 16th European Conference, Glasgow, UK, August 23–28, 2020, Proceedings, Part X 16, pages 93–110. Springer, 2020. 3" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 56, + 287, + 296, + 342 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 287, + 296, + 342 + ], + "spans": [ + { + "bbox": [ + 56, + 287, + 296, + 342 + ], + "type": "text", + "content": "[29] Guanyi Qin, Runze Hu, Yutao Liu, Xiawu Zheng, Haotian Liu, Xiu Li, and Yan Zhang. Data-efficient image quality assessment with attention-panel decoder. In Proceedings of the Thirty-Seventh AAAI Conference on Artificial Intelligence, 2023. 3, 4, 5, 6, 8" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 56, + 343, + 296, + 386 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 343, + 296, + 386 + ], + "spans": [ + { + "bbox": [ + 56, + 343, + 296, + 386 + ], + "type": "text", + "content": "[30] Jiezhong Qiu, Hao Ma, Omer Levy, Scott Wen-tau Yih, Sinong Wang, and Jie Tang. Blockwise self-attention for long document understanding. arXiv preprint arXiv:1911.02972, 2019. 3" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 56, + 388, + 296, + 442 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 388, + 296, + 442 + ], + "spans": [ + { + "bbox": [ + 56, + 388, + 296, + 442 + ], + "type": "text", + "content": "[31] Avinab Saha, Sandeep Mishra, and Alan C Bovik. Re-iqa: Unsupervised learning for image quality assessment in the wild. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 5846-5855, 2023. 2, 4, 5, 6" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 56, + 445, + 296, + 499 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 445, + 296, + 499 + ], + "spans": [ + { + "bbox": [ + 56, + 445, + 296, + 499 + ], + "type": "text", + "content": "[32] Ramprasaath R Selvaraju, Michael Cogswell, Abhishek Das, Ramakrishna Vedantam, Devi Parikh, and Dhruv Batra. Gradcam: Visual explanations from deep networks via gradient-based localization. In Proceedings of the IEEE international conference on computer vision, pages 618-626, 2017. 8" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 56, + 501, + 296, + 555 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 501, + 296, + 555 + ], + "spans": [ + { + "bbox": [ + 56, + 501, + 296, + 555 + ], + "type": "text", + "content": "[33] Tahira Shehzadi, Khurram Azeem Hashmi, Didier Stricker, and Muhammad Zeshan Afzal. Sparse semi-detr: Sparse learnable queries for semi-supervised object detection. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 5840-5850, 2024. 4" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 56, + 556, + 296, + 600 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 556, + 296, + 600 + ], + "spans": [ + { + "bbox": [ + 56, + 556, + 296, + 600 + ], + "type": "text", + "content": "[34] Hamid R Sheikh, Alan C Bovik, and Lawrence Cormack. No-reference quality assessment using natural scene statistics: Jpeg2000. IEEE Transactions on image processing, 14(11): 1918-1927, 2005. 2" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 56, + 601, + 296, + 644 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 601, + 296, + 644 + ], + "spans": [ + { + "bbox": [ + 56, + 601, + 296, + 644 + ], + "type": "text", + "content": "[35] Hamid R Sheikh, Muhammad F Sabir, and Alan C Bovik. A statistical evaluation of recent full reference image quality assessment algorithms. IEEE Transactions on image processing, 15(11):3440-3451, 2006. 5" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 56, + 647, + 296, + 689 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 647, + 296, + 689 + ], + "spans": [ + { + "bbox": [ + 56, + 647, + 296, + 689 + ], + "type": "text", + "content": "[36] Chenyang Shi and Yandan Lin. Full reference image quality assessment based on visual salience with color appearance and gradient similarity. IEEE Access, 8:97310-97320, 2020. 1" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 56, + 691, + 296, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 691, + 296, + 713 + ], + "spans": [ + { + "bbox": [ + 56, + 691, + 296, + 713 + ], + "type": "text", + "content": "[37] Tianshu Song, Leida Li, Deqiang Cheng, Pengfei Chen, and Jinjian Wu. Active learning-based sample selection for label-" + } + ] + } + ], + "index": 13 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 316, + 73, + 555, + 713 + ], + "type": "list", + "angle": 0, + "index": 30, + "blocks": [ + { + "bbox": [ + 333, + 73, + 553, + 95 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 333, + 73, + 553, + 95 + ], + "spans": [ + { + "bbox": [ + 333, + 73, + 553, + 95 + ], + "type": "text", + "content": "efficient blind image quality assessment. IEEE Transactions on Circuits and Systems for Video Technology, 2023. 1" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 316, + 96, + 555, + 150 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 96, + 555, + 150 + ], + "spans": [ + { + "bbox": [ + 316, + 96, + 555, + 150 + ], + "type": "text", + "content": "[38] Suhas Srinath, Shankhanil Mitra, Shika Rao, and Rajiv Soundararajan. Learning generalizable perceptual representations for data-efficient no-reference image quality assessment. In Proceedings of the IEEE/CVF Winter Conference on Applications of Computer Vision, pages 22-31, 2024. 1" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 316, + 152, + 555, + 205 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 152, + 555, + 205 + ], + "spans": [ + { + "bbox": [ + 316, + 152, + 555, + 205 + ], + "type": "text", + "content": "[39] Shaolin Su, Qingsen Yan, Yu Zhu, Cheng Zhang, Xin Ge, Jinqiu Sun, and Yanning Zhang. Blindly assess image quality in the wild guided by a self-adaptive hyper network. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 3667-3676, 2020. 5" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 316, + 206, + 555, + 248 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 206, + 555, + 248 + ], + "spans": [ + { + "bbox": [ + 316, + 206, + 555, + 248 + ], + "type": "text", + "content": "[40] Simeng Sun, Tao Yu, Jiahua Xu, Wei Zhou, and Zhibo Chen. Graphiqa: Learning distortion graph representations for blind image quality assessment. IEEE Transactions on Multimedia, 25:2912-2925, 2022. 2" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 316, + 250, + 555, + 293 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 250, + 555, + 293 + ], + "spans": [ + { + "bbox": [ + 316, + 250, + 555, + 293 + ], + "type": "text", + "content": "[41] Dacheng Tao, Xuelong Li, Wen Lu, and Xinbo Gao. Reduced-reference iqa in contourlet domain. IEEE Transactions on Systems, Man, and Cybernetics, Part B (Cybernetics), 39(6): 1623-1627, 2009. 1" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 316, + 294, + 555, + 348 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 294, + 555, + 348 + ], + "spans": [ + { + "bbox": [ + 316, + 294, + 555, + 348 + ], + "type": "text", + "content": "[42] Hugo Touvron, Matthieu Cord, Matthijs Douze, Francisco Massa, Alexandre Sablayrolles, and Hervé Jégou. Training data-efficient image transformers & distillation through attention. In International Conference on Machine Learning, pages 10347-10357. PMLR, 2021. 8" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 316, + 350, + 555, + 380 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 350, + 555, + 380 + ], + "spans": [ + { + "bbox": [ + 316, + 350, + 555, + 380 + ], + "type": "text", + "content": "[43] Hugo Touvron, Matthieu Cord, and Hervé Jégou. Deit iii: Revenge of the vit. arXiv preprint arXiv:2204.07118, 2022. 6" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 316, + 382, + 555, + 426 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 382, + 555, + 426 + ], + "spans": [ + { + "bbox": [ + 316, + 382, + 555, + 426 + ], + "type": "text", + "content": "[44] Tao Wang, Li Yuan, Xiaopeng Zhang, and Jiashi Feng. Distilling object detectors with fine-grained feature imitation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 4933-4942, 2019. 5" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 316, + 427, + 554, + 459 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 427, + 554, + 459 + ], + "spans": [ + { + "bbox": [ + 316, + 427, + 554, + 459 + ], + "type": "text", + "content": "[45] Zhou Wang and Eero Simoncelli. Local phase coherence and the perception of blur. Advances in neural information processing systems, 16, 2003. 2" + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 316, + 460, + 555, + 502 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 460, + 555, + 502 + ], + "spans": [ + { + "bbox": [ + 316, + 460, + 555, + 502 + ], + "type": "text", + "content": "[46] Zhou Wang, Hamid R Sheikh, and Alan C Bovik. No-reference perceptual quality assessment of JPEG compressed images. In Proceedings. International conference on image processing, pages I-I. IEEE, 2002. 2" + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 316, + 504, + 555, + 547 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 504, + 555, + 547 + ], + "spans": [ + { + "bbox": [ + 316, + 504, + 555, + 547 + ], + "type": "text", + "content": "[47] Zhou Wang, Alan C Bovik, Hamid R Sheikh, and Eero P Simoncelli. Image quality assessment: from error visibility to structural similarity. IEEE transactions on image processing, 13(4):600-612, 2004. 1" + } + ] + } + ], + "index": 25 + }, + { + "bbox": [ + 316, + 548, + 555, + 591 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 548, + 555, + 591 + ], + "spans": [ + { + "bbox": [ + 316, + 548, + 555, + 591 + ], + "type": "text", + "content": "[48] Zhihua Wang, Zhi-Ri Tang, Jianguo Zhang, and Yuming Fang. Toward a blind image quality evaluator in the wild by learning beyond human opinion scores. Pattern Recognition, 137:109296, 2023. 1" + } + ] + } + ], + "index": 26 + }, + { + "bbox": [ + 316, + 593, + 555, + 646 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 593, + 555, + 646 + ], + "spans": [ + { + "bbox": [ + 316, + 593, + 555, + 646 + ], + "type": "text", + "content": "[49] Haoning Wu, Zicheng Zhang, Erli Zhang, Chaofeng Chen, Liang Liao, Annan Wang, Chunyi Li, Wenxiu Sun, Qiong Yan, Guangtao Zhai, et al. Q-bench: A benchmark for general-purpose foundation models on low-level vision. arXiv preprint arXiv:2309.14181, 2023. 1" + } + ] + } + ], + "index": 27 + }, + { + "bbox": [ + 316, + 647, + 555, + 690 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 647, + 555, + 690 + ], + "spans": [ + { + "bbox": [ + 316, + 647, + 555, + 690 + ], + "type": "text", + "content": "[50] Kangmin Xu, Liang Liao, Jing Xiao, Chaofeng Chen, Haoning Wu, Qiong Yan, and Weisi Lin. Local distortion aware efficient transformer adaptation for image quality assessment. arXiv preprint arXiv:2308.12001, 2023. 2, 3, 5, 6" + } + ] + } + ], + "index": 28 + }, + { + "bbox": [ + 316, + 692, + 555, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 692, + 555, + 713 + ], + "spans": [ + { + "bbox": [ + 316, + 692, + 555, + 713 + ], + "type": "text", + "content": "[51] Sidi Yang, Tianhe Wu, Shuwei Shi, Shanshan Lao, Yuan Gong, Mingdeng Cao, Jiahao Wang, and Yujiu Yang. Maniqa:" + } + ] + } + ], + "index": 29 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "text", + "content": "2353" + } + ] + } + ], + "index": 31 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 9 + }, + { + "para_blocks": [ + { + "bbox": [ + 56, + 72, + 296, + 713 + ], + "type": "list", + "angle": 0, + "index": 12, + "blocks": [ + { + "bbox": [ + 75, + 72, + 296, + 116 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 75, + 72, + 296, + 116 + ], + "spans": [ + { + "bbox": [ + 75, + 72, + 296, + 116 + ], + "type": "text", + "content": "Multi-dimension attention network for no-reference image quality assessment. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 1191-1200, 2022. 1" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 56, + 118, + 296, + 162 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 118, + 296, + 162 + ], + "spans": [ + { + "bbox": [ + 56, + 118, + 296, + 162 + ], + "type": "text", + "content": "[52] Guanghao Yin, Wei Wang, Zehuan Yuan, Chuchu Han, Wei Ji, Shouqian Sun, and Changhu Wang. Content-variant reference image quality assessment via knowledge distillation. arXiv preprint arXiv:2202.13123, 2022. 3" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 56, + 163, + 296, + 228 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 163, + 296, + 228 + ], + "spans": [ + { + "bbox": [ + 56, + 163, + 296, + 228 + ], + "type": "text", + "content": "[53] Zhenqiang Ying, Haoran Niu, Praful Gupta, Dhruv Mahajan, Deepti Ghadiyaram, and Alan Bovik. From patches to pictures (paq-2-piq): Mapping the perceptual space of picture quality. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 3575–3585, 2020. 5, 6" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 56, + 230, + 296, + 274 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 230, + 296, + 274 + ], + "spans": [ + { + "bbox": [ + 56, + 230, + 296, + 274 + ], + "type": "text", + "content": "[54] Junyong You and Jari Korhonen. Transformer for image quality assessment. In 2021 IEEE International Conference on Image Processing (ICIP), pages 1389-1393. IEEE, 2021. 3, 5" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 56, + 276, + 296, + 330 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 276, + 296, + 330 + ], + "spans": [ + { + "bbox": [ + 56, + 276, + 296, + 330 + ], + "type": "text", + "content": "[55] Guanghui Yue, Di Cheng, Leida Li, Tianwei Zhou, Hantao Liu, and Tianfu Wang. Semi-supervised authentically distorted image quality assessment with consistency-preserving dual-branch convolutional neural network. IEEE Transactions on Multimedia, 25:6499–6511, 2022. 3" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 56, + 332, + 296, + 376 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 332, + 296, + 376 + ], + "spans": [ + { + "bbox": [ + 56, + 332, + 296, + 376 + ], + "type": "text", + "content": "[56] Weixia Zhang, Kede Ma, Jia Yan, Dexiang Deng, and Zhou Wang. Blind image quality assessment using a deep bilinear convolutional neural network. IEEE Transactions on Circuits and Systems for Video Technology, 30(1):36-47, 2018. 2, 5" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 56, + 377, + 296, + 442 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 377, + 296, + 442 + ], + "spans": [ + { + "bbox": [ + 56, + 377, + 296, + 442 + ], + "type": "text", + "content": "[57] Weixia Zhang, Guangtao Zhai, Ying Wei, Xiaokang Yang, and Kede Ma. Blind image quality assessment via vision-language correspondence: A multitask learning perspective. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 14071-14081, 2023. 2, 5" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 56, + 445, + 296, + 498 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 445, + 296, + 498 + ], + "spans": [ + { + "bbox": [ + 56, + 445, + 296, + 498 + ], + "type": "text", + "content": "[58] Kai Zhao, Kun Yuan, Ming Sun, Mading Li, and Xing Wen. Quality-aware pre-trained models for blind image quality assessment. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 22302-22313, 2023. 2, 4, 5, 6" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 56, + 501, + 296, + 555 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 501, + 296, + 555 + ], + "spans": [ + { + "bbox": [ + 56, + 501, + 296, + 555 + ], + "type": "text", + "content": "[59] Heliang Zheng, Huan Yang, Jianlong Fu, Zheng-Jun Zha, and Jiebo Luo. Learning conditional knowledge distillation for degraded-reference image quality assessment. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 10242-10251, 2021. 3" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 56, + 557, + 296, + 610 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 557, + 296, + 610 + ], + "spans": [ + { + "bbox": [ + 56, + 557, + 296, + 610 + ], + "type": "text", + "content": "[60] Yu Zhou, Leida Li, Shiqi Wang, Jinjian Wu, Yuming Fang, and Xinbo Gao. No-reference quality assessment for view synthesis using dog-based edge statistics and texture naturalness. IEEE Transactions on Image Processing, 28(9):4566-4579, 2019. 1" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 56, + 613, + 296, + 668 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 613, + 296, + 668 + ], + "spans": [ + { + "bbox": [ + 56, + 613, + 296, + 668 + ], + "type": "text", + "content": "[61] Hancheng Zhu, Leida Li, Jinjian Wu, Weisheng Dong, and Guangming Shi. Metaiaq: Deep meta-learning for no-reference image quality assessment. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 14143-14152, 2020. 2, 5" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 56, + 670, + 296, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 670, + 296, + 713 + ], + "spans": [ + { + "bbox": [ + 56, + 670, + 296, + 713 + ], + "type": "text", + "content": "[62] Mengmeng Zhu, Guanqun Hou, Xinjia Chen, Jiaxing Xie, Haixian Lu, and Jun Che. Saliency-guided transformer network combined with local embedding for no-reference image quality assessment. In Proceedings of the IEEE/CVF Inter-" + } + ] + } + ], + "index": 11 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 316, + 72, + 556, + 140 + ], + "type": "list", + "angle": 0, + "index": 15, + "blocks": [ + { + "bbox": [ + 335, + 72, + 555, + 94 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 335, + 72, + 555, + 94 + ], + "spans": [ + { + "bbox": [ + 335, + 72, + 555, + 94 + ], + "type": "text", + "content": "national Conference on Computer Vision, pages 1953-1962, 2021.8" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 316, + 96, + 556, + 140 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 96, + 556, + 140 + ], + "spans": [ + { + "bbox": [ + 316, + 96, + 556, + 140 + ], + "type": "text", + "content": "[63] Fuzhen Zhuang, Zhiyuan Qi, Keyu Duan, Dongbo Xi, Yongchun Zhu, Hengshu Zhu, Hui Xiong, and Qing He. A comprehensive survey on transfer learning. Proceedings of the IEEE, 109(1):43-76, 2020. 3" + } + ] + } + ], + "index": 14 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "text", + "content": "2354" + } + ] + } + ], + "index": 16 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 10 + } + ], + "_backend": "vlm", + "_version_name": "2.6.4" +} \ No newline at end of file diff --git a/2025/Distilling Spectral Graph for Object-Context Aware Open-Vocabulary Semantic Segmentation/7ae5aeda-7866-414e-b296-550ab2d843cd_content_list.json b/2025/Distilling Spectral Graph for Object-Context Aware Open-Vocabulary Semantic Segmentation/7ae5aeda-7866-414e-b296-550ab2d843cd_content_list.json new file mode 100644 index 0000000000000000000000000000000000000000..93a953825a9617156dc2f063297fcf40404a797a --- /dev/null +++ b/2025/Distilling Spectral Graph for Object-Context Aware Open-Vocabulary Semantic Segmentation/7ae5aeda-7866-414e-b296-550ab2d843cd_content_list.json @@ -0,0 +1,1439 @@ +[ + { + "type": "text", + "text": "Distilling Spectral Graph for Object-Context Aware Open-Vocabulary Semantic Segmentation", + "text_level": 1, + "bbox": [ + 236, + 130, + 759, + 176 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Chanyoung $\\mathrm{Kim}^1$ Dayun Ju $^1$ Woojung Han $^1$ Ming-Hsuan Yang $^{1,2}$ Seong Jae Hwang $^1$", + "bbox": [ + 130, + 202, + 864, + 223 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "$^{1}$ Yonsei University $^{2}$ University of California, Merced", + "bbox": [ + 279, + 237, + 715, + 256 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "{chanyoung, juda0707, dnwjDDL, seongjae}@yonsei.ac.kr,mhyang@ucmerced.edu", + "bbox": [ + 173, + 258, + 818, + 273 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Abstract", + "text_level": 1, + "bbox": [ + 246, + 308, + 326, + 324 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Open-Vocabulary Semantic Segmentation (OVSS) has advanced with recent vision-language models (VLMs), enabling segmentation beyond predefined categories through various learning schemes. Notably, training-free methods offer scalable, easily deployable solutions for handling unseen data, a key goal of OVSS. Yet, a critical issue persists: lack of object-level context consideration when segmenting complex objects in the challenging environment of OVSS based on arbitrary query prompts. This oversight limits models' ability to group semantically consistent elements within object and map them precisely to user-defined arbitrary classes. In this work, we introduce a novel approach that overcomes this limitation by incorporating object-level contextual knowledge within images. Specifically, our model enhances intra-object consistency by distilling spectral-driven features from vision foundation models into the attention mechanism of the visual encoder, enabling semantically coherent components to form a single object mask. Additionally, we refine the text embeddings with zero-shot object presence likelihood to ensure accurate alignment with the specific objects represented in the images. By leveraging object-level contextual knowledge, our proposed approach achieves state-of-the-art performance with strong generalizability across diverse datasets.", + "bbox": [ + 89, + 340, + 483, + 705 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "1. Introduction", + "text_level": 1, + "bbox": [ + 91, + 724, + 220, + 739 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Open-vocabulary semantic segmentation (OVSS) [9, 32, 54, 55, 58] aims to predict pixel-level labels for arbitrary prompts defined by user input (e.g., the proper noun Space Needle in Fig. 1(a)), advancing beyond previous models scoped to predefined static classes [10, 18, 19, 40]. To achieve this, fully supervised OVSS models [11, 53, 56] enhance generalization to seen classes using labeled training data. Yet, their reliance on seen classes during training may", + "bbox": [ + 89, + 751, + 483, + 872 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/f14b060ee2ae67aafe5e2f19fa931c75774a2ad9eef48c4745ae849936ee7544.jpg", + "image_caption": [ + "(b) Visualization", + "Truck", + "Horse", + "Cow", + "Sheep" + ], + "image_footnote": [], + "bbox": [ + 517, + 308, + 903, + 489 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/7ba854727b4ad18975b4615743c47a45a7bd4c97aea34938889ec67f8e7f96bc.jpg", + "image_caption": [ + "Figure 1. We present CASS, object-level Context-Aware training-free open-vocabulary Semantic Segmentation model. (a) Overview: Our method distills the vision foundation model's (VFM) object-level contextual spectral graph into CLIP's attention and refines query text embeddings towards object-specific semantics. (b) Object-Level Context: This result illustrates how incorporating object-level context improves segmentation accuracy by unifying object-wise components into a user-defined object class; for example, (top row) our approach precisely segments the truck's body, wheels, and cargo area and (bottom row) accurately groups elements into a single object, such as the sheep, whereas baselines [16, 29] often fail to achieve this unified segmentation." + ], + "image_footnote": [], + "bbox": [ + 516, + 503, + 903, + 635 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "risk overfitting and limit scalability, as labeled data is required for retraining to adapt to new domains [43].", + "bbox": [ + 511, + 823, + 903, + 853 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "To better address this, recent studies [16, 28, 29, 48, 58] propose training-free OVSS methods that leverage pretrained vision-language models (VLMs) (e.g., CLIP [39]) to", + "bbox": [ + 511, + 854, + 906, + 900 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "CVF", + "bbox": [ + 106, + 2, + 181, + 42 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore.", + "bbox": [ + 236, + 0, + 810, + 46 + ], + "page_idx": 0 + }, + { + "type": "page_footnote", + "text": "Project Page: https://micv-yonsei.github.io/cass/", + "bbox": [ + 112, + 887, + 421, + 898 + ], + "page_idx": 0 + }, + { + "type": "page_number", + "text": "15033", + "bbox": [ + 480, + 944, + 517, + 955 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "integrate aligned visual and textual representations. These works focus on improving semantic association between visual patches in CLIP image encoder to produce more precise segmentation maps, achieving highly promising results. This training-free scheme offers key advantages: it (1) generalizes well to unseen classes without requiring additional labeled data, making it more scalable, thus (2) allows direct adaptation to dynamic real-world applications. Capitalizing on these benefits, training-free OVSS aligns with OVSS's core goal of adaptability across diverse domains, enabling it to reach its full potential.", + "bbox": [ + 89, + 90, + 480, + 257 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "We tackle OVSS in a training-free manner, presenting a challenging and practical approach aligned with OVSS's true objectives. Although existing training-free methods [21, 29, 47, 48] perform well on numerous tasks, they all share a common limitation: the lack of \"object-level context.\" To understand \"object-level context\", consider a truck and sheep in Fig. 1(b). Components comprising a truck, such as its wheels and cargo area, should be grouped as a single entity. Similarly, components of a sheep should be accurately unified under the correct user-given object class. However, existing training-free OVSS methods, which use CLIP as the image backbone for segmentation, often struggle to capture object-level context, failing to group object components into a single, user-defined object. These challenges stem from CLIP's focus on learning global image semantics, which may be insufficient for capturing dense object-level semantics [16, 48, 58]. Thus, to achieve accurate training-free OVSS, a proper object-level context must be considered in CLIP.", + "bbox": [ + 91, + 261, + 483, + 547 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "We address the challenges of training-free OVSS by introducing the object-level context and propose the Context-Aware Semantic Segmentation (CASS) model. Vision foundation models (VFMs), such as DINO [8, 38], capture fine-grained, patch-level semantics but lack object-level context, making them unsuitable for direct use in tasks requiring such context. Recent studies [24, 36] have shown that applying graph spectral techniques to VFM attention graphs can transform patch-level representations into object-level representations. Accordingly, we apply spectral techniques to extract essential object-level context from the VFM attention graph. The extracted spectral features can be used directly for segmentation (e.g., clustering eigenvectors); however, this approach may not be suitable for OVSS, where textual alignment is essential. Thus, we propose distilling low-rank components of VFM into CLIP to enhance its object-level context understanding, while maintaining alignment with textual information. Specifically, we decompose the VFM's attention graph to extract low-rank components, filtering irrelevant information while emphasizing essential object-level context. These low-rank components are then distilled into CLIP's attention to embed the fundamental object structure of the VFM graph.", + "bbox": [ + 89, + 551, + 482, + 901 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Our model also leverages CLIP's highly effective zero-shot object classification capability (i.e., object presence prior), widely validated in prior work [20, 34, 39], to capture detailed object-level context within scenes. As such, we adjust the text embeddings based on the object presence prior encoded by CLIP. This process involves refining the embeddings to better align with object-specific semantics. Then, we refine patch-text similarities using the object presence prior, ensuring that the final segmentation map reflects an object-specific perspective. Our CASS, carefully designed to enhance object-level contextual knowledge, captures richer intra-object coherence and enables cohesive groupings of semantically related elements.", + "bbox": [ + 511, + 90, + 903, + 287 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "The main contributions of this work are:", + "bbox": [ + 527, + 289, + 797, + 301 + ], + "page_idx": 1 + }, + { + "type": "list", + "sub_type": "text", + "list_items": [ + "- We enhance object-level context by distilling rich semantic understanding from VFM into the CLIP attention mechanism, facilitating a more consistent grouping of each object into a unified semantic entity.", + "- We present object presence prior-driven text embeddings and patch-text similarity refinement, promoting the model to precisely classify objects.", + "- We demonstrate the performance of our model in various semantic segmentation datasets and achieve state-of-the-art performance in training-free OVSS." + ], + "bbox": [ + 511, + 304, + 903, + 455 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "2. Related Works", + "text_level": 1, + "bbox": [ + 511, + 465, + 663, + 481 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Open-Vocabulary Semantic Segmentation. Open-vocabulary semantic segmentation (OVSS) aims to segment images corresponding to arbitrary user-given prompts, using vision-language aligned representations (e.g., CLIP [39]). Meanwhile, training-free approaches [3, 4, 23, 29, 58] have risen for their ability to adapt directly to new tasks without extra training. Among these, methods [16, 29, 48, 58] that do not rely on additional datasets have gained attention, as they require no extra adaptation effort, aligning better with the objective of OVSS. Earlier approaches investigated leveraging CLIP features for better localization [30, 58] or for grounding proposed masks [21, 47]. SCLIP [48] uses correlative self-attention to improve inter-patch correlations. NACLIP [16] applies a Gaussian kernel to attention maps for better spatial consistency, and ProxyCLIP [29] integrates patch-wise consistent features from DINO [8]. Although these methods demonstrated strong results, they often struggle to group different parts of the same object due to a lack of explicit dense label knowledge. To this end, we attribute the underlying challenges of training-free OVSS by enriching the object-level contextual understanding.", + "bbox": [ + 511, + 488, + 903, + 820 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Vision Foundation Models. Vision Foundation Models (VFMs), especially vision transformer (ViT)-based [14] self-supervised models (e.g., DINO [8, 38]), have significantly advanced computer vision through their ability to capture patch-wise semantic understanding at both the", + "bbox": [ + 511, + 824, + 903, + 900 + ], + "page_idx": 1 + }, + { + "type": "page_number", + "text": "15034", + "bbox": [ + 480, + 944, + 517, + 955 + ], + "page_idx": 1 + }, + { + "type": "image", + "img_path": "images/e6308b01ef7816f3a5505a7104fb111a555c41f315081604abbbb7ad4ffc6240.jpg", + "image_caption": [ + "Figure 2. Detailed illustration of our proposed training-free spectral object-level context distillation mechanism in Sec. 3.2. By matching the attention graphs of VFM and CLIP head-by-head to establish complementary relationships, and distilling the fundamental object-level context of the VFM graph to CLIP, we enhance CLIP's ability to capture intra-object contextual coherence." + ], + "image_footnote": [], + "bbox": [ + 94, + 88, + 898, + 306 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "feature and attention levels. Among various downstream tasks [17, 35, 42, 46, 49] leveraging VFM's capability, one line of research focuses on leveraging the attention components in VFM [24, 36, 46], which exhibit a detailed understanding of patch-wise semantic interaction between different image regions. Thus, while ProxyCLIP [29] utilizes VFM features, our method takes a different effort by treating the attention mechanism in VFM as a graph. We then extract core structural patterns of the VFM graph to emphasize object-level context, providing an effective solution for producing precise object masks in training-free OVSS.", + "bbox": [ + 88, + 364, + 482, + 530 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Object-Level Semantic Segmentation. Semantic segmentation groups object components that have similar semantics into a single mask, traditionally achieved through models trained on extensive human-labeled, pixel-wise annotations [10, 19, 25, 40]. However, despite recent advancements in unsupervised semantic segmentation [17, 42] (USS) that do not rely on dense labels, a core challenge persists: individual object components are not consistently grouped into a single entity. To address this issue, recent USS studies [24, 41, 50] have incorporated object-level representation learning to improve the model's ability to merge single objects into unified entities. However, these efforts have not been considered in the training-free OVSS task, which faces similar challenges in an unsupervised context. Therefore, we focus on enriching object-level context in training-free OVSS to achieve a more effective grouping of components within each object, enabling clearer and more consistent segmentation across diverse visual domains.", + "bbox": [ + 89, + 535, + 483, + 806 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3. Method", + "text_level": 1, + "bbox": [ + 89, + 814, + 181, + 829 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "We detail our proposed model, CASS, implemented in a training-free manner, as outlined in Fig 1(a). Our model jointly enhances object-level contextual understanding with two primary aspects: (1) Spectral Object-Level", + "bbox": [ + 89, + 839, + 483, + 902 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Context Distillation (Sec. 3.2) and (2) Object Presence-Driven Object-Level Context (Sec. 3.3). We first outline the basic pipeline of training-free OVSS in Sec. 3.1.", + "bbox": [ + 511, + 364, + 906, + 411 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3.1. Preliminaries", + "text_level": 1, + "bbox": [ + 511, + 417, + 653, + 431 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "CLIP Visual Encoder. Training-free OVSS models [16, 29, 48] leverage CLIP [39] for its rich alignment of visual and linguistic information. However, the CLIP ViT visual encoder $\\mathcal{F}_{\\mathrm{CLIP}}^v$ focuses on aligning the [CLS] token with text embeddings during training, often neglecting the patchwise spatial interactions among visual tokens that are crucial for dense prediction [16, 44, 48]. To produce CLIP feature representations more suitable for semantic segmentation tasks, we remove residual connection, feed-forward network, and self-attention layer in the final layer of $\\mathcal{F}_{\\mathrm{CLIP}}^v$ , following recent works [16, 48]. Specifically, when the set of visual tokens $Z = \\mathrm{Concat}[z_1, \\ldots, z_N] \\in \\mathbb{R}^{N \\times D_Z}$ , where $N$ is the total number of patches except [CLS] token and $D_Z$ denotes patch dimension space, is input into the final block of the ViT, the following operations are as", + "bbox": [ + 511, + 436, + 906, + 664 + ], + "page_idx": 2 + }, + { + "type": "equation", + "text": "\n$$\nZ ^ {*} = \\operatorname {S A} (\\operatorname {L N} (Z)), \\tag {1}\n$$\n", + "text_format": "latex", + "bbox": [ + 642, + 667, + 903, + 684 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "where SA and LN represent Self-Attention and Layer Normalization, respectively. The modified SA layer is", + "bbox": [ + 511, + 686, + 905, + 717 + ], + "page_idx": 2 + }, + { + "type": "equation", + "text": "\n$$\nM _ {\\mathrm {C L I P}} ^ {i} = \\operatorname {s o f t m a x} \\left(A _ {\\mathrm {C L I P}} ^ {i} / \\sqrt {D _ {h}}\\right), \\tag {2}\n$$\n", + "text_format": "latex", + "bbox": [ + 591, + 720, + 903, + 744 + ], + "page_idx": 2 + }, + { + "type": "equation", + "text": "\n$$\nZ ^ {*} = \\operatorname {C o n c a t} \\left[ M _ {\\mathrm {C L I P}} ^ {1} V _ {\\mathrm {C L I P}} ^ {1}, \\dots , M _ {\\mathrm {C L I P}} ^ {h} V _ {\\mathrm {C L I P}} ^ {h} \\right] W ^ {O}, \\tag {3}\n$$\n", + "text_format": "latex", + "bbox": [ + 532, + 748, + 903, + 767 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "where $D_h$ denotes the head dimension, $h$ is the number of heads, and $W^0 \\in \\mathbb{R}^{D_z \\times D_z}$ is the output projection matrix. We define CLIP attention map for $i$ -th head as $A_{\\mathrm{CLIP}}^i = K_{\\mathrm{CLIP}}^i K_{\\mathrm{CLIP}}^{i^\\top} \\in \\mathbb{R}^{N \\times N}$ , where attention key $K_{\\mathrm{CLIP}} \\in \\mathbb{R}^{N \\times D_h \\times h}$ and value $V_{\\mathrm{CLIP}} \\in \\mathbb{R}^{N \\times D_h \\times h}$ result from the linear transformation of $\\mathrm{LN}(Z)$ .", + "bbox": [ + 511, + 776, + 906, + 867 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Inference Pipeline. We use a sliding window inference to ensure that ViT can capture detailed parts of the image by", + "bbox": [ + 511, + 869, + 906, + 901 + ], + "page_idx": 2 + }, + { + "type": "page_number", + "text": "15035", + "bbox": [ + 480, + 944, + 517, + 955 + ], + "page_idx": 2 + }, + { + "type": "image", + "img_path": "images/29ebad50915d6e3e59ee82cf81eb349c4fd862f5c127cd6d0c6d0a7d45dc4db1.jpg", + "image_caption": [ + "Figure 3. Attention score visualization for various query points. Left: Vanilla CLIP ( $A_{\\mathrm{CLIP}}$ ) shows noisy, unfocused attention. Center: VFM-to-CLIP distillation without low-rank eigenscaling shows partial object grouping with limited detail. Right: Incorporating our low-rank eigenscaling (Eq. (7)) captures object-level context, improving grouping within a single object." + ], + "image_footnote": [], + "bbox": [ + 94, + 88, + 478, + 328 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "dividing the entire image $\\mathcal{I}$ into smaller windows $\\hat{\\mathcal{I}}$ . The visual tokens $Z^{*} \\in \\mathbb{R}^{N \\times D_{Z}}$ are then computed from each $\\hat{\\mathcal{I}}$ through $\\mathcal{F}_{\\mathrm{CLIP}}^{v}$ as in Eq. (1). To perform OVSS with given texts, we project the visual tokens $Z^{*}$ into the vision-language latent space as $F_{\\mathrm{CLIP}} = \\mathcal{P}_{\\mathrm{CLIP}}(Z^{*}) \\in \\mathbb{R}^{N \\times d}$ , where $\\mathcal{P}_{\\mathrm{CLIP}}$ is a projection matrix into CLIP latent space and $d$ is the CLIP latent dimension. We compute patch-text similarity $\\hat{S} \\in \\mathbb{R}^{N \\times C}$ for each image window $\\hat{\\mathcal{I}}$ as", + "bbox": [ + 89, + 426, + 483, + 547 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\hat {\\mathcal {S}} = F _ {\\mathrm {C L I P}} \\left[ \\left\\{t _ {\\mathrm {C L I P}} ^ {i} \\right\\} _ {i = 1} ^ {C} \\right] ^ {\\top}, \\tag {4}\n$$\n", + "text_format": "latex", + "bbox": [ + 192, + 558, + 482, + 585 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "where $\\{t_{\\mathrm{CLIP}}^i\\}_{i = 1}^C \\coloneqq \\mathrm{Concat}\\left[t_{\\mathrm{CLIP}}^1,\\dots ,t_{\\mathrm{CLIP}}^C\\right] \\in \\mathbb{R}^{C\\times d}$ denotes the encoded text embeddings, and $C$ is the number of classes. Finally, the results of all windows are combined to create a segmentation map for the entire image.", + "bbox": [ + 89, + 595, + 483, + 657 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "3.2. Spectral Object-Level Context Distillation", + "text_level": 1, + "bbox": [ + 89, + 666, + 449, + 683 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Our model CASS aims to capture such object-level contextual interaction within the image features to associate different parts of the same object. However, the limited ability of CLIP to capture semantic relationships between patches often neglects the object-level context, as illustrated in the left column of Fig. 3 ( $A_{\\mathrm{CLIP}}$ ). To address this limitation, we leverage VFM (e.g., DINO [8]), enabling a deeper object-level contextual understanding. Among VFM features, we utilize VFM attention key component $K_{\\mathrm{VFM}} \\in \\mathbb{R}^{h \\times N \\times D_h}$ . Specifically, we treat the VFM attention adjacency $A_{\\mathrm{VFM}} = K_{\\mathrm{VFM}} K_{\\mathrm{VFM}}^{\\top} \\in \\mathbb{R}^{h \\times N \\times N}$ as a graph, where $h$ represents the number of heads, and transfer crucial graph patterns from $A_{\\mathrm{VFM}}$ to $A_{\\mathrm{CLIP}}$ . However, a key question remains: \"How do we account for the multi-head attention mechanisms in both", + "bbox": [ + 89, + 688, + 483, + 900 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "VFM and CLIP?\" In other words, \"how do we match $i$ -th VFM attention head $A_{\\mathrm{VFM}}^i$ with $j$ -th CLIP attention head $A_{\\mathrm{CLIP}}^j$ ? A straightforward approach is to distill the attention heads sequentially (i.e., $i = j$ ). Nevertheless, recent studies [22, 31, 59] indicate that different attention heads focus on separate parts of the image. Thus, matching the optimal heads between VFM and CLIP is essential for precise attention distillation.", + "bbox": [ + 511, + 90, + 905, + 212 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "3.2.1. Complementary Spectral Graph Matching", + "text_level": 1, + "bbox": [ + 511, + 220, + 856, + 234 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "We match graphs with contrasting structures, as shown in Fig 4, enabling VFM to supplement the object-level contextual knowledge that CLIP alone struggles to capture. To match an optimal head-wise attention graph, we exploit the spectral field of both $A_{\\mathrm{VFM}}$ and $A_{\\mathrm{CLIP}}$ . Our graph-matching mechanism has two main steps as illustrated in Fig. 2(a): (I) examining the eigenvalues of each graph and (II) matching optimal graph pairs using the spectral distribution.", + "bbox": [ + 511, + 236, + 905, + 356 + ], + "page_idx": 3 + }, + { + "type": "list", + "sub_type": "text", + "list_items": [ + "(I) Eigenvalue Examination. For each graph from VFM and CLIP, denoted as $A_{\\mathrm{VFM}}^i$ and $A_{\\mathrm{CLIP}}^i$ for $(i = 1, \\dots, h)$ , we perform eigendecomposition. This yields the eigenvalues for each head $h$ , $\\lambda_{\\mathrm{VFM}}^i$ and $\\lambda_{\\mathrm{CLIP}}^i$ , from which the top $m$ fixed eigenvalues are selected. These eigenvalues contain each graph's unique structural features, key properties essential for distinguishing different attention graph patterns.", + "(II) Graph Matching via Spectral Distribution. After obtaining eigenvalues for each head, we compute spectral distances to quantify structural differences, creating a cost matrix $\\mathcal{C} \\in \\mathbb{R}^{h \\times h}$ for each graph pair from VFM and CLIP:" + ], + "bbox": [ + 511, + 359, + 905, + 531 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {C} _ {i j} = 1 - \\mathcal {D} _ {W} \\left(\\bar {\\lambda} _ {\\mathrm {V F M}} ^ {i}, \\bar {\\lambda} _ {\\mathrm {C L I P}} ^ {j}\\right), \\quad \\forall i, j \\in \\{1, \\dots , h \\} \\tag {5}\n$$\n", + "text_format": "latex", + "bbox": [ + 529, + 541, + 903, + 560 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "where $\\bar{\\lambda}_{\\mathrm{VFM}}^i$ and $\\bar{\\lambda}_{\\mathrm{CLIP}}^j$ represent the normalized eigenvalues of the $i$ -th head of VFM and the $j$ -th head of CLIP, respectively, and $\\mathcal{D}_W$ denotes the Wasserstein distance. Here, the Wasserstein distance is computed as $\\mathcal{D}_W(u,v) = \\sum_{i=1}^{h} |\\operatorname{sort}(u)_i - \\operatorname{sort}(v)_i|$ , where $u$ and $v$ are the two input distributions, and the sort function orders the values of $u$ and $v$ in ascending order. After computing $\\mathcal{C}$ , we apply the Hungarian matching algorithm [27] to the cost matrix to find the optimal pairing of graph heads. This approach pairs graphs with contrasting characteristics, enabling complementary distillation of object-level context knowledge from $A_{\\mathrm{VFM}}$ to $A_{\\mathrm{CLIP}}$ , as described in the following section.", + "bbox": [ + 511, + 569, + 906, + 753 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "3.2.2. Distilling VFM Spectral Graph to CLIP", + "text_level": 1, + "bbox": [ + 511, + 760, + 836, + 775 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Once we match optimal graph pairs, we distill the VFM graph to CLIP (see Fig. 2(b)). For the sake of explanation, we assume that the $i$ -th VFM attention head $A_{\\mathrm{VFM}}^i$ is matched with the $j$ -th CLIP attention head $A_{\\mathrm{CLIP}}^j$ . An intuitive approach would be simply aggregating $A_{\\mathrm{VFM}}^i$ and $A_{\\mathrm{CLIP}}^j$ without any transformation. However, as shown in Fig. 3 ( $A_{\\psi w/o \\mathrm{LR}}$ ), this approach may transfer noise or irrelevant information, highlighting the need to extract features", + "bbox": [ + 511, + 776, + 905, + 900 + ], + "page_idx": 3 + }, + { + "type": "page_number", + "text": "15036", + "bbox": [ + 480, + 944, + 519, + 955 + ], + "page_idx": 3 + }, + { + "type": "image", + "img_path": "images/be1e7e611f01088054fa7f9b0c142b13a3ad234b1a6a6d43920d50f02483aed8.jpg", + "image_caption": [ + "Figure 4. Visualization of the matched complementary graph pairs. The red dotted line indicates the focus areas of VFM, while the yellow dotted line highlights those for CLIP. By aggregating graphs with contrasting structural properties, we enhance the object-level context in $A_{\\mathrm{CLIP}}$ supported by $\\ddot{A}_{\\mathrm{VFM}}$ , resulting $A_{\\psi}$ ." + ], + "image_footnote": [], + "bbox": [ + 96, + 88, + 478, + 250 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "that emphasize object-level context. From this realization, we leverage the low-rank components of VFM, which contain distinct object patterns within the graph structure [12]. Specifically, we (I) extract the critical object-level contextual structure of $A_{\\mathrm{VFM}}^i$ via low-rank approximation and enhance the graph structure by dynamically scaling eigenvalues. Then, we (II) distill the approximated VFM graph into CLIP by aggregating graphs based on spectral distance.", + "bbox": [ + 89, + 334, + 483, + 455 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "(I) Low-Rank Dynamic Eigenscaling. Our goal is to transfer the essential object-level contextual structure of VFM graph $A_{\\mathrm{VFM}}$ to CLIP. To achieve this, we extract low-rank components of the VFM graph using standard eigendecomposition, as $A_{\\mathrm{VFM}}^i = U\\Sigma U^\\top$ , where $U$ and $\\Sigma$ represent the eigenvector and the diagonal eigenvalue matrix, respectively. In the decomposed eigenbasis, we identify key object-level features of each graph by searching an optimal number of eigenvalues $k$ through an energy-based approach. This ensures that the chosen $k$ eigenvalues capture a significant portion of the graph's energy, retaining essential structural information while discarding noise and less relevant details. More details on searching $k$ can be found in the supplementary material. This process yields a low-rank approximated VFM graph, expressed as $\\tilde{A}_{\\mathrm{VFM}}^i = U_k\\Sigma_kU_k^\\top$ .", + "bbox": [ + 89, + 458, + 482, + 685 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "We refine the low-rank components with a scaling function $\\phi$ , which dynamically amplifies larger eigenvalues and reduces smaller ones. Compared to the conventional shrinkage function [2, 5], which only focuses on noise cutoff, our approach emphasizes essential structural information, particularly object-level context features, while suppressing noise and irrelevant details. More details on dynamic eigenscaling can be found in the supplementary material. The graph after applying dynamic eigenscaling is as", + "bbox": [ + 89, + 685, + 482, + 821 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\ddot {A} _ {\\mathrm {V F M}} ^ {i} = U _ {k} \\phi (\\Sigma_ {k}) U _ {k} ^ {\\top}. \\tag {6}\n$$\n", + "text_format": "latex", + "bbox": [ + 210, + 825, + 482, + 844 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "(II) VFM Graph Distillation. We distill the essential object-level contextual knowledge from the tailored VFM graph $\\ddot{A}_{\\mathrm{VFM}}^i$ into the $A_{\\mathrm{CLIP}}^j$ by aggregating their structures:", + "bbox": [ + 89, + 851, + 483, + 902 + ], + "page_idx": 4 + }, + { + "type": "image", + "img_path": "images/8f30b67fc69581bdba2fef41e9589b525bac9ea247f49c6329795b8673f9185e.jpg", + "image_caption": [ + "Figure 5. Detailed illustration of our object presence prior-guided text embedding adjustment module. The CLIP text encoder generates text embeddings for each object class, and the object presence prior is derived from both visual and text embeddings. Within hierarchically defined class groups, text embeddings are selected based on object presence prior, then refined in an object-specific direction to align with components likely present in the image." + ], + "image_footnote": [], + "bbox": [ + 516, + 87, + 903, + 268 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\nA _ {\\psi} ^ {j} = \\left(w _ {i j} \\ddot {A} _ {\\mathrm {V F M}} ^ {i} + A _ {\\mathrm {C L I P}} ^ {j}\\right) / \\left(w _ {i j} + 1\\right), \\tag {7}\n$$\n", + "text_format": "latex", + "bbox": [ + 581, + 378, + 903, + 400 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "where $w_{ij}$ is defined as the Wasserstein distance between eigenvalues from both graphs as $\\mathcal{D}_{\\mathcal{W}}(\\bar{\\lambda}_{\\mathrm{VFM}}^i,\\bar{\\lambda}_{\\mathrm{CLIP}}^j)$ . This adaptively assigns higher weights to graphs with complementary structures, while reducing influence on those with similar structures. Our distillation strategy effectively transfers the critical object-level context from VFM to CLIP, ensuring multiple components of the same object are grouped into a unified semantic as illustrated in Fig. 3 ( $A_{\\psi}$ ). Therefore, we use $A_{\\psi}$ as our attention matrix instead of $A_{\\mathrm{CLIP}}$ in Eq. (2) to compute the final visual feature $F_{\\mathrm{CLIP}}$ .", + "bbox": [ + 511, + 409, + 906, + 561 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "3.3. Object Presence-Driven Object-Level Context", + "text_level": 1, + "bbox": [ + 511, + 571, + 901, + 589 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "The spectral object-level context distillation in Sec. 3.2 enables precise object mask generation. However, due to the nature of OVSS, where users can provide arbitrary query prompts, different parts of the same object may still be assigned to closely related categories. Meanwhile, numerous studies [20, 34, 39] have demonstrated that CLIP excels in zero-shot object classification capability. Accordingly, we utilize the zero-shot object classification score encoded by CLIP (i.e., the object presence prior) to refine text embeddings and patch-text similarity, enhancing object-centered perspective. To compute the object presence prior $P$ , we take an entire input image $\\mathcal{I}$ and compute likelihood for each class $i$ as $P(i) = \\{t_{\\mathrm{CLIP}}^i\\}_{i=1}^C v_{\\mathrm{CLIP}}$ , where the visual embedding vector $v_{\\mathrm{CLIP}} = \\mathcal{F}_{\\mathrm{CLIP}}^v(\\mathcal{I}) \\in \\mathbb{R}^d$ is derived from the ViT [CLS] token. The object presence prior is used in the following sections to enhance object-level context.", + "bbox": [ + 511, + 594, + 906, + 835 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Object-Guided Text Embedding Adjustment. We adjust text embeddings $\\{t_{\\mathrm{CLIP}}^i\\}_{i = 1}^C$ to embed object-specific contextual semantics, as shown in Fig. 5. To prevent object parts from being assigned to ambiguous classes, we", + "bbox": [ + 511, + 839, + 906, + 901 + ], + "page_idx": 4 + }, + { + "type": "page_number", + "text": "15037", + "bbox": [ + 480, + 944, + 517, + 955 + ], + "page_idx": 4 + }, + { + "type": "table", + "img_path": "images/544419de3792c240900a3899d7fb6fb49c78e568c4fb79791cc62e99671a5402.jpg", + "table_caption": [ + "Table 1. Quantitative results with state-of-the-art unsupervised open-vocabulary semantic segmentation models on eight datasets." + ], + "table_footnote": [], + "table_body": "
ModelSupporting DatasetExtra-TrainingFairV21PC60C-ObjV20PC59C-StfCityADEAvg.
GroupViT [54]CVPR'22CC12M+RedCaps✓✘50.418.727.579.723.415.311.19.229.4
TCL [9]CVPR'23CC3M+CC12M✓✘55.030.431.683.233.922.424.017.137.2
CoDe [51]CVPR'24CC3M+RedCaps✓✘57.530.532.3--23.928.917.7-
CLIP-DINOiser [52]ECCV'24ImageNet1k✓✘62.132.434.880.935.924.631.720.040.3
ReCo [45]NeurIPS'22ImageNet1k✘✘25.119.915.757.722.314.821.611.223.5
FOSSIL [3]WACV'24COCO Captions✘✘----35.824.823.218.8-
FreeDa [4]CVPR'24COCO Captions✘✘---85.643.127.836.722.4-
CLIP [39]ICML'21✘✘✓18.67.86.549.111.27.26.73.213.8
MaskCLIP [58]ECCV'22✘✘✓38.323.620.674.926.416.412.69.827.9
GEM [6]CVPR'24✘✘✓46.2---32.6--15.7-
CaR [47]CVPR'24✘✘✓48.613.615.473.718.4--5.4-
PnP-OVSS [33]CVPR'24✘✘✓--36.251.328.017.9-14.2-
CLIPtrase [44]ECCV'24✘✘✓50.929.943.681.033.822.821.316.432.7
ClearCLIP [28]ECCV'24✘✘✓51.832.633.080.935.923.930.016.738.1
SCLIP [48]ECCV'24✘✘✓59.130.430.580.434.122.432.216.138.2
LaVG [21]ECCV'24✘✘✓62.131.634.282.534.723.226.215.838.8
ProxyCLIP [29]ECCV'24✘✘✓59.135.236.278.238.826.238.119.641.4
NACLIP [16]WACV'25✘✘✓58.932.233.279.735.223.335.517.439.4
CASS✘✘✓65.836.737.887.840.226.739.420.444.4
", + "bbox": [ + 94, + 107, + 903, + 393 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "The hierarchical clustering of CLIP semantic distance, forming groups of similar embeddings $\\mathcal{T}_g = \\{t_{\\mathrm{CLIP}}^i\\mid i\\in I_g\\}$ , where $\\mathcal{T}_g$ is a subset of entire text embeddings $\\{t_{\\mathrm{CLIP}}^i\\}_{i = 1}^C$ with index set $I_{g}$ . Further details about this clustering process are provided in the supplementary materials. Within $\\mathcal{T}_g$ , we identify the object most likely to appear in the image using object presence prior $P$ , formulated as $i^{*} = \\arg \\max_{i\\in I_{g}}P(i)$ . This facilitates refining the text embedding vector $t_{\\mathrm{CLIP}}^*$ to better represent object-contextual semantics. We note that patch-wise visual feature $F_{\\mathrm{CLIP}}\\in \\mathbb{R}^{N\\times d}$ contains object-specific semantic vectors that can guide $t_{\\mathrm{CLIP}}^*$ accordingly. Thus, we compute cosine similarity between $F_{\\mathrm{CLIP}}$ and $t_{\\mathrm{CLIP}}^*$ and select top- $n$ object-specific image vector denoted as $\\{f_i\\}_{i = 1}^n\\coloneqq$ $\\mathrm{Concat}[f_1,\\ldots ,f_n]\\in \\mathbb{R}^{n\\times d}$ . To define a targeted direction, we compute the average of $\\{f_i\\}_{i = 1}^n$ as $\\mu_{n}\\in \\mathbb{R}^{d}$ , and guide selected text embedding $t_{\\mathrm{CLIP}}^*$ towards $\\mu_{n}$ as", + "bbox": [ + 89, + 407, + 483, + 667 + ], + "page_idx": 5 + }, + { + "type": "equation", + "text": "\n$$\n\\tilde {t} _ {\\mathrm {C L I P}} ^ {i ^ {*}} = (1 - \\alpha) \\cdot t _ {\\mathrm {C L I P}} ^ {i ^ {*}} + \\alpha \\cdot \\mu_ {n}, \\tag {8}\n$$\n", + "text_format": "latex", + "bbox": [ + 176, + 669, + 480, + 688 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "where $\\alpha$ controls the degree of alignment with $\\mu_{n}$ . The adjusted text embedding $\\tilde{t}_{\\mathrm{CLIP}}^{i^*}$ replaces the original text embedding $t_{\\mathrm{CLIP}}^{i}$ in the set $\\{t_{\\mathrm{CLIP}}^{i}\\}_{i=1}^{C}$ , which is then used to compute the patch-text similarity logit $\\hat{S}$ in Eq. (4).", + "bbox": [ + 88, + 694, + 482, + 760 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Object Perspective Patch-Text Similarity. We integrate the object presence prior with patch-text similarity $\\hat{S}$ from Eq. (4) to produce a segmentation map from an object-perspective view. Unlike Barsellotti et al. [4], which computes likelihood from an image window $\\hat{\\mathcal{I}}$ , our approach derives similarity from the full image $\\mathcal{I}$ , enabling richer object-contextual understanding at the broader level. Thus, we refine the patch-text similarity $\\hat{S}$ from Eq. (4) to obtain the final patch-text similarity $\\hat{S}^*$ , computed as", + "bbox": [ + 88, + 763, + 482, + 901 + ], + "page_idx": 5 + }, + { + "type": "equation", + "text": "\n$$\n\\hat {\\mathcal {S}} ^ {*} = (1 - \\gamma) \\cdot \\hat {\\mathcal {S}} + \\gamma \\cdot \\left\\{t _ {\\mathrm {C L I P}} ^ {i} \\right\\} _ {i = 1} ^ {C} v _ {\\mathrm {C L I P}}, \\tag {9}\n$$\n", + "text_format": "latex", + "bbox": [ + 571, + 404, + 903, + 426 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "where $\\gamma$ controls the balance between $\\hat{S}$ and object presence prior. This allows the final segmentation map to capture object-level details for precise, class-aware segmentation.", + "bbox": [ + 511, + 429, + 903, + 474 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "4. Experiments", + "text_level": 1, + "bbox": [ + 511, + 479, + 645, + 497 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "We first describe the experimental settings, including implementation details, datasets, and evaluation metrics (Sec. 4.1). We then present quantitative and qualitative evaluation results (Sec. 4.2) and demonstrate ablation studies to assess each proposed component (Sec. 4.3). Additional experiment results are provided in the supplementary material.", + "bbox": [ + 511, + 505, + 905, + 597 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "4.1. Experimental Settings", + "text_level": 1, + "bbox": [ + 511, + 604, + 720, + 619 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Implementation Details. Similar to existing training-free approaches [16, 21, 29, 44, 48], we employ CLIP ViT-B/16 [39] as our VLM and DINO ViT-B/8 [8] as our VFM, keeping both frozen. Input images are matched with the shorter side set to 336 pixels (or 560 pixels for the higher-resolution Cityscapes dataset), and sliding window inference is applied with a $224 \\times 224$ window and a stride of 112 to achieve efficient evaluation coverage. All experiments are conducted using NVIDIA RTX A6000 Ada*.", + "bbox": [ + 511, + 623, + 906, + 758 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Datasets. Our primary datasets include PASCAL VOC 2012 [15], PASCAL Context [37], and COCO [7], each evaluated with and without a background class. We refer to the versions with a background class as V21, PC60, and C-Obj, and those without as V20, PC59, and C-Stf. We also report results on the ADE20K (ADE) [57] and Cityscapes (City) [13] datasets. These datasets contain 21, 60, 81, 20, 59, 171, 150, and 19 classes, respectively.", + "bbox": [ + 511, + 758, + 905, + 878 + ], + "page_idx": 5 + }, + { + "type": "page_footnote", + "text": "* Advanced Database System Infrastructure (NFEC-2024-11-300458)", + "bbox": [ + 539, + 887, + 864, + 898 + ], + "page_idx": 5 + }, + { + "type": "page_number", + "text": "15038", + "bbox": [ + 480, + 944, + 517, + 955 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/6acc002ec693f900363bf04759e497b548ae11fa32e1c9b1b3103d66c2b3f462.jpg", + "image_caption": [ + "Figure 6. Qualitative comparison across the Pascal VOC [15], Pascal Context [37], COCO [7], and ADE20K [57] datasets using CLIP ViT-B/16 [39], evaluating previous state-of-the-art training-free OVSS methods, including NACLIP [16], ProxyCLIP [29], and ours." + ], + "image_footnote": [], + "bbox": [ + 91, + 87, + 903, + 321 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "4.2. Evaluation Results", + "text_level": 1, + "bbox": [ + 89, + 359, + 272, + 375 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "We evaluate our model against training-free OVSS methods both quantitatively and qualitatively. For fair comparison, we refrain from using the mask refinement (e.g., PAMR [1] or DenseCRF [26]) and reproduce results from [29] using a $224 \\times 224$ window to match the baselines. Performance is measured primarily with mean IoU (mIoU), complemented by pixel accuracy (pAcc) for a broader assessment.", + "bbox": [ + 89, + 380, + 482, + 487 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Quantitative Evaluation: mIoU. Table 1 presents the main results in mIoU, where all results are obtained using CLIP ViT-B/16 as the backbone model for fair comparisons. To demonstrate the strengths of our model, we also include methods that use supporting datasets (i.e., retrieving additional datasets during inference) or additional training (models evaluated under the same setting as ours are indicated with $\\checkmark$ in the \"Fair\" column). Our model, CASS, performs favorably against the state-of-the-art methods, with an average gain of $3.0\\mathrm{mIoU}$ point across eight datasets. Notably, for the V20, CASS achieves a performance gain of 5.3 mIoU point over the second-best model. Further, CASS outperforms CLIP-DINOiser [52] (which also leverages DINO with extra data and training) by $4.1\\mathrm{mIoU}$ point.", + "bbox": [ + 89, + 488, + 482, + 700 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Quantitative Evaluation: pAcc. Table 2 shows the results of pAcc on recent state-of-the-art models evaluated under the same settings. Our CASS outperforms the second-highest model by $2.2\\mathrm{pAcc}$ point and exceeds one of the most recent papers, CLIPtrase [44], by $9.8\\mathrm{pAcc}$ point. Our model consistently performs better than existing methods on mIoU and pAcc by effectively capturing object-level context, resulting in more accurate object masking.", + "bbox": [ + 89, + 700, + 482, + 823 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Qualitative Analysis. Fig. 6 provides a qualitative comparison with recent state-of-the-art models [16, 29]. Existing methods often produce noisy segmentation maps and do not group object components correctly, such as wheels within a car or a person's head or arm. Additionally, they often mis", + "bbox": [ + 89, + 824, + 482, + 901 + ], + "page_idx": 6 + }, + { + "type": "table", + "img_path": "images/387ac760cd43f0c155b29e003b31e54b744c2dc49d3f7f0713312a480b036af7.jpg", + "table_caption": [ + "Table 2. Quantitative results using average pixel accuracy." + ], + "table_footnote": [], + "table_body": "
ModelV21PC60C-ObjV20PC59C-StfCityADEAvg.
CLIPtrase [44]78.652.150.189.758.938.963.438.659.1
LaVG [21]89.348.774.891.158.939.168.537.063.4
SCLIP [48]87.649.274.391.058.338.472.738.763.8
ProxyCLIP [29]86.652.075.988.463.443.474.949.166.7
NACLIP [16]87.151.275.389.259.839.371.745.264.9
CASS90.155.676.093.965.043.678.248.668.9
", + "bbox": [ + 517, + 373, + 901, + 462 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "classify object parts into ambiguous classes (e.g., a motorcycle as a bicycle or an airplane as a car). In contrast, based on a deep understanding of object-level context, our CASS generates clean segmentation maps that accurately group all object components, including slender parts such as human arms or animal legs, into their respective object classes.", + "bbox": [ + 511, + 472, + 905, + 564 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "4.3. Ablation Study", + "text_level": 1, + "bbox": [ + 511, + 573, + 666, + 589 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "We perform ablation studies to analyze the proposed method. We use the datasets representative of segmentation tasks for all experiments: V21, PC59, C-Stf, following [48]. Please refer to the supplementary material for the additional ablation studies including VFM backbones and feature type.", + "bbox": [ + 511, + 593, + 905, + 669 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Effect of Spectral VFM Distillation. Table 3 shows the results of incorporating VFM (Exp.#3) with the vanilla setting (Exp.#1), which directly uses the output from Eq. (4). In Exp.#3, $A_{\\mathrm{VFM}}$ and $A_{\\mathrm{CLIP}}$ are aggregated in a straightforward head-wise manner, without any additional processing to $A_{\\mathrm{VFM}}$ . Simply leveraging $A_{\\mathrm{VFM}}$ leads to a performance gain compared to the vanilla setting. Exp.#4 to Exp.#6 validate the effectiveness of the proposed components (see Sec. 3.2). Exp.#4 demonstrates the aggregation of $A_{\\mathrm{VFM}}$ and $A_{\\mathrm{CLIP}}$ using a complementary spectral graph matching (GM) procedure. Exp.#5, denoted as LR, presents the results with only the low-rank approximation applied to $A_{\\mathrm{VFM}}$ , while Exp.#6 shows the results with both the low-rank approximation and Dynamic Eigenscaling (DE) applied, as defined in Eq. (6). Notably, Exp.#6 shows signifi", + "bbox": [ + 511, + 672, + 906, + 902 + ], + "page_idx": 6 + }, + { + "type": "page_number", + "text": "15039", + "bbox": [ + 480, + 944, + 517, + 955 + ], + "page_idx": 6 + }, + { + "type": "table", + "img_path": "images/117059a025c63c40c71ee0f77436030f8d5d517b981628234091170cfabd3e95.jpg", + "table_caption": [ + "Table 3. Ablation results showing the impact of our components on the performance across V21 [15], PC59 [37], and C-Stf [7]. The abbreviations in the table are noted in Sec. 4.3." + ], + "table_footnote": [], + "table_body": "
Exp. #AVFMGMLRDEOPSOTADataset
V21PC59C-Stf
158.935.323.4
2✓✓59.236.524.0
3✓62.937.725.2
4✓✓63.238.025.4
5✓✓✓64.538.425.8
6✓✓✓✓64.838.625.9
7✓✓✓✓✓65.039.426.6
8✓✓✓✓✓✓65.840.226.7
", + "bbox": [ + 94, + 135, + 478, + 267 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/89aa33d1d8b0c648e3138465c94d23ae99fe508f53f1e024a70cabf0b63829df.jpg", + "image_caption": [ + "Figure 7. Qualitative ablation of our method. Exp.# refers to the experiment in Table 3. As additional components of our proposed method are incorporated, the segmentation map exhibits greater refinement and more effectively captures object-level context." + ], + "image_footnote": [], + "bbox": [ + 94, + 270, + 480, + 383 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "cant performance gain over simply integrating VFM knowledge (Exp.#3), demonstrating the effectiveness of graph matching and low-rank approximation in capturing object-level context. Fig. 3 qualitatively visualizes the attention map, with the left, middle, and right columns corresponding to Exp.#3, Exp.#4, and Exp.#6, respectively. We observe that integrating proposed low-rank dynamic eigenscaling effectively reduces noise and emphasizes objects by highlighting objects with strong inter-object associations.", + "bbox": [ + 88, + 450, + 482, + 587 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Effect of Object Presence Prior. We validate the effectiveness of the object presence prior-driven methods in Exp.#7 and Exp.#8 by examining each component: Object-Perspective Patch-Text Similarity (OPS) and Object-Guided Text Embedding Adjustment (OTA) (see Sec. 3.3). These results show that each component effectively improves performance, emphasizing object-level context. Additionally, incorporating the object presence prior-driven methods on top of the vanilla setting, as shown in Exp.#2, clarifies the effectiveness of the proposed method.", + "bbox": [ + 88, + 590, + 482, + 742 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Qualitative Ablations. Fig. 7 shows the segmentation maps for the vanilla setting (Exp.#1), after applying spectral VFM distillation (Exp.#6), and with all proposed components applied (Exp.#8). While Exp.#6 improves object masking compared to Exp.#1, it often misses parts of objects (highlighted by the red dotted line). In contrast, incorporating all proposed components (Exp.#8) consolidates objects into single entities, effectively capturing object-level context in the segmentation mask.", + "bbox": [ + 88, + 744, + 482, + 881 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Different CLIP Backbones. Table 4 shows results on dif-", + "bbox": [ + 89, + 885, + 482, + 898 + ], + "page_idx": 7 + }, + { + "type": "table", + "img_path": "images/2ef8e1ca42d3bd1f95cd593223a6e47d9f291bf7ea7eaa261473aa40145b4a33.jpg", + "table_caption": [ + "Table 4. Ablation results on different CLIP backbones." + ], + "table_footnote": [], + "table_body": "
ModelV21PC59C-StfAvg. mIoUAvg. pAcc
mIoUpAccmIoUpAccmIoUpAcc
VLT-B/32SCLIP [48]50.681.028.749.620.033.433.154.7
LaVG [21]54.884.429.050.120.533.734.856.1
ProxyCLIP [29]57.985.635.259.723.640.238.961.8
NACLIP [16]51.182.832.456.821.237.134.958.9
CASS58.286.236.561.024.440.839.762.7
VLT-L/14SCLIP [48]44.478.125.246.617.629.629.151.4
LaVG [21]51.584.127.550.319.432.332.855.6
ProxyCLIP [29]59.885.738.361.026.243.541.463.4
NACLIP [16]52.283.132.152.821.435.735.257.2
CASS62.188.039.162.126.343.442.564.5
", + "bbox": [ + 516, + 102, + 903, + 227 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/adb871732c94a18747dc50fc01192bda794e1167ea88792b2145623dbc03eee2.jpg", + "image_caption": [ + "Figure 8. Performance evaluation of different distance metrics used in graph matching and graph distillation." + ], + "image_footnote": [], + "bbox": [ + 514, + 231, + 903, + 321 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "ferent CLIP backbones, ViT-B/32 and ViT-L/14. Our CASS performs favorably across various backbones, consistently outperforming existing methods in mIoU and pAcc. These results demonstrate the effectiveness and robustness of our method across different backbone architectures.", + "bbox": [ + 511, + 362, + 906, + 436 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Distance Metric for Graph Matching. We evaluate different distance metrics $\\mathcal{D}_W$ used in Sec. 3.2.1, including Wasserstein Distance, Cosine Distance, KL Divergence, and Euclidean Distance. As shown in Fig. 8, the Wasserstein Distance shows the best results as it provides an intuitive measure of the overall shape of sorted eigenvalue distributions. While we use Wasserstein Distance due to its performance, the other well-known metrics exhibited comparable results, demonstrating the robustness of our method in terms of the choice of a distance metric and further supporting the effectiveness of our proposed approach.", + "bbox": [ + 511, + 441, + 908, + 609 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "5. Conclusion", + "text_level": 1, + "bbox": [ + 511, + 614, + 633, + 630 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "In this work, we present CASS, a training-free approach for OVSS that integrates object-level context into CLIP. By distilling object-aware structural information from VFM into CLIP's attention, CASS effectively enhances intra-object coherence, ensuring that each object is consistently represented as a unified semantic entity. Additionally, we leverage object-presence prior knowledge to adjust text embeddings and refine object class assignments, aligning them with object-specific semantic targets in the image. Consequently, CASS effectively groups object components and assigns semantic labels without extra training or additional labeled data, making it a robust solution across diverse datasets and visual contexts.", + "bbox": [ + 511, + 633, + 906, + 829 + ], + "page_idx": 7 + }, + { + "type": "page_footnote", + "text": "Acknowledgement. This work was supported in part by IITP RS-2024-00457882 (National AI Research Hub Project), IITP 2020-0-01361, NRF RS-2024-00345806, NRF RS-2023-00219019, and RS-2024-00403860 (Korea Basic Science Institute, National Research Facilities and Equipment Center).", + "bbox": [ + 511, + 847, + 903, + 888 + ], + "page_idx": 7 + }, + { + "type": "page_number", + "text": "15040", + "bbox": [ + 480, + 944, + 519, + 955 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "References", + "text_level": 1, + "bbox": [ + 91, + 89, + 187, + 104 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[1] Nikita Araslanov and Stefan Roth. Single-stage semantic segmentation from image labels. In CVPR, 2020. 7", + "[2] Danny Barash and Matan Gavish. Optimal shrinkage of singular values under random data contamination. NeurIPS, 2017. 5", + "[3] Luca Barsellotti, Roberto Amoroso, Lorenzo Baraldi, and Rita Cucchiara. Fossil: Free open-vocabulary semantic segmentation through synthetic references retrieval. In WACV, 2024. 2, 6", + "[4] Luca Barsellotti, Roberto Amoroso, Marcella Cornia, Lorenzo Baraldi, and Rita Cucchiara. Training-free open-vocabulary segmentation with offline diffusion-augmented prototype generation. In CVPR, 2024. 2, 6", + "[5] Toby Boas, Aritra Dutta, Xin Li, Kathryn P Mercier, and Eric Niderman. Shrinkage function and its applications in matrix approximation. arXiv preprint arXiv:1601.07600, 2016. 5", + "[6] Walid Bousselham, Felix Petersen, Vittorio Ferrari, and Hilde Kuehne. Grounding everything: Emerging localization properties in vision-language transformers. In CVPR, 2024. 6", + "[7] Holger Caesar, Jasper Uijlings, and Vittorio Ferrari. COCO-Stuff: Thing and Stuff Classes in Context. In CVPR, 2018. 6, 7, 8", + "[8] Mathilde Caron, Hugo Touvron, Ishan Misra, Hervé Jégou, Julien Mairal, Piotr Bojanowski, and Armand Joulin. Emerging Properties in Self-Supervised Vision Transformers. In ICCV, 2021. 2, 4, 6", + "[9] Junbum Cha, Jonghwan Mun, and Byungseok Roh. Learning To Generate Text-Grounded Mask for Open-World Semantic Segmentation From Only Image-Text Pairs. In CVPR, 2023. 1, 6", + "[10] Liang-Chieh Chen, Yukun Zhu, George Papandreou, Florian Schroff, and Hartwig Adam. Encoder-decoder with atrous separable convolution for semantic image segmentation. In ECCV, 2018. 1, 3", + "[11] Seokju Cho, Heeseong Shin, Sunghwan Hong, Anurag Arnab, Paul Hongsuck Seo, and Seungryong Kim. Cat-seg: Cost aggregation for open-vocabulary semantic segmentation. In CVPR, 2024. 1", + "[12] Fan RK Chung. Spectral graph theory. American Mathematical Soc., 1997. 5", + "[13] Marius Cordts, Mohamed Omran, Sebastian Ramos, Timo Rehfeld, Markus Enzweiler, Rodrigo Benenson, Uwe Franke, Stefan Roth, and Bernt Schiele. The Cityscapes Dataset for Semantic Urban Scene Understanding. In CVPR, 2016. 6", + "[14] Alexey Dosovitskiy, Lucas Beyer, Alexander Kolesnikov, Dirk Weissenborn, Xiaohua Zhai, Thomas Unterthiner, Mostafa Dehghani, Matthias Minderer, Georg Heigold, Sylvain Gelly, Jakob Uszkoreit, and Neil Houlsby. An Image is Worth 16x16 Words: Transformers for Image Recognition at Scale. In ICLR, 2021. 2", + "[15] M. Everingham, L. Van Gool, C. K. I. Williams, J. Winn, and A. Zisserman. The PASCAL Visual Object Classes Challenge 2012 (VOC2012) Results, 2012. 6, 7, 8" + ], + "bbox": [ + 93, + 114, + 482, + 900 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[16] Sina Hajimiri, Ismail Ben Ayed, and Jose Dolz. Pay attention to your neighbours: Training-free open-vocabulary semantic segmentation. In WACV, 2025. 1, 2, 3, 6, 7, 8", + "[17] Mark Hamilton, Zhoutong Zhang, Bharath Hariharan, Noah Snavely, and William T Freeman. Unsupervised semantic segmentation by distilling feature correspondences. arXiv preprint arXiv:2203.08414, 2022. 3", + "[18] Woojung Han, Seil Kang, Kyobin Choo, and Seong Jae Hwang. Complementary branch fusing class and semantic knowledge for robust weakly supervised semantic segmentation. Pattern Recognition, 2025. 1", + "[19] Ali Hatamizadeh, Yucheng Tang, Vishwesh Nath, Dong Yang, Andriy Myronenko, Bennett Landman, Holger R Roth, and Daguang Xu. Unetr: Transformers for 3d medical image segmentation. In CVPR, 2022. 1, 3", + "[20] Ruixiang Jiang, Lingbo Liu, and Changwen Chen. Clipcount: Towards text-guided zero-shot object counting. In ACM Multimedia, 2023. 2, 5", + "[21] Dahiyun Kang and Minsu Cho. In defense of lazy visual grounding for open-vocabulary semantic segmentation. In ECCV, 2024. 2, 6, 7, 8", + "[22] Seil Kang, Jinyeong Kim, Junhyeok Kim, and Seong Jae Hwang. See what you are told: Visual attention sink in large multimodal models. In ICLR, 2025. 4", + "[23] Seil Kang, Jinyeong Kim, Junhyeok Kim, and Seong Jae Hwang. Your large vision-language model only needs a few attention heads for visual grounding. In CVPR, 2025. 2", + "[24] Chanyoung Kim, Woogung Han, Dayun Ju, and Seong Jae Hwang. Eagle: Eigen aggregation learning for object-centric unsupervised semantic segmentation. In CVPR, 2024. 2, 3", + "[25] Alexander Kirillov, Eric Mintun, Nikhila Ravi, Hanzi Mao, Chloe Rolland, Laura Gustafson, Tete Xiao, Spencer Whitehead, Alexander C Berg, Wan-Yen Lo, et al. Segment anything. In ICCV, 2023. 3", + "[26] Philipp Krahenbuhl and Vladlen Koltun. Efficient Inference in Fully Connected CRFs with Gaussian Edge Potentials. In NeurIPS, 2011. 7", + "[27] Harold W Kuhn. The hungarian method for the assignment problem. Naval research logistics quarterly, 1955. 4", + "[28] Mengcheng Lan, Chaofeng Chen, Yiping Ke, Xinjiang Wang, Litong Feng, and Wayne Zhang. Clearclip: Decomposing clip representations for dense vision-language inference. In ECCV, 2024. 1, 6", + "[29] Mengcheng Lan, Chaofeng Chen, Yiping Ke, Xinjiang Wang, Litong Feng, and Wayne Zhang. Proxyclip: Proxy attention improves clip for open-vocabulary segmentation. In ECCV, 2024. 1, 2, 3, 6, 7, 8", + "[30] Yi Li, Hualiang Wang, Yiqun Duan, and Xiaomeng Li. Clip surgery for better explainability with enhancement in open-vocabulary tasks. arXiv preprint arXiv:2304.05653, 2023. 2", + "[31] Yiran Li, Junpeng Wang, Xin Dai, Liang Wang, ChinChia Michael Yeh, Yan Zheng, Wei Zhang, and Kwan-Liu Ma. How does attention work in vision transformers? a visual analytics attempt. IEEE Transactions on Visualization and Computer Graphics, 2023. 4" + ], + "bbox": [ + 516, + 92, + 903, + 900 + ], + "page_idx": 8 + }, + { + "type": "page_number", + "text": "15041", + "bbox": [ + 480, + 944, + 517, + 955 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[32] Feng Liang, Bichen Wu, Xiaoliang Dai, Kunpeng Li, Yinan Zhao, Hang Zhang, Peizhao Zhang, Peter Vajda, and Diana Marculescu. Open-vocabulary semantic segmentation with mask-adapted clip. In CVPR, 2023. 1", + "[33] Jiayun Luo, Siddhesh Khandelwal, Leonid Sigal, and Boyang Li. Emergent open-vocabulary semantic segmentation from off-the-shelf vision-language models. In CVPR, 2024. 6", + "[34] Yunyao Mao, Jiajun Deng, Wengang Zhou, Li Li, Yao Fang, and Houqiang Li. Clip4hoi: towards adapting clip for practical zero-shot hoi detection. NeurIPS, 2023. 2, 5", + "[35] Octave Mariotti, Oisin Mac Aodha, and Hakan Bilen. Improving semantic correspondence with viewpoint-guided spherical maps. In CVPR, 2024. 3", + "[36] Luke Melas-Kyriazi, Christian Rupprecht, Iro Laina, and Andrea Vedaldi. Deep spectral methods: A surprisingly strong baseline for unsupervised semantic segmentation and localization. In CVPR, 2022. 2, 3", + "[37] Roozbeh Mottaghi, Xianjie Chen, Xiaobai Liu, Nam-Gyu Cho, Seong-Whan Lee, Sanja Fidler, Raquel Urtasun, and Alan Yuille. The Role of Context for Object Detection and Semantic Segmentation in the Wild. In CVPR, 2014. 6, 7, 8", + "[38] Maxime Oquab, Timothee Darct, Theo Moutakanni, Huy Vo, Marc Szafraniec, Vasil Khalidov, Pierre Fernandez, Daniel Haziza, Francisco Massa, Alaaeldin El-Nouby, et al. Dinov2: Learning robust visual features without supervision. arXiv preprint arXiv:2304.07193, 2023. 2", + "[39] Alec Radford, Jong Wook Kim, Chris Hallacy, Aditya Ramesh, Gabriel Goh, Sandhini Agarwal, Girish Sastry, Amanda Askell, Pamela Mishkin, Jack Clark, et al. Learning Transferable Visual Models From Natural Language Supervision. In ICML, 2021. 1, 2, 3, 5, 6, 7", + "[40] Olaf Ronneberger, Philipp Fischer, and Thomas Brox. U-net: Convolutional networks for biomedical image segmentation. In MICCAI, 2015. 1, 3", + "[41] Maximilian Seitzer, Max Horn, Andrii Zadaianchuk, Dominik Zietlow, Tianjun Xiao, Carl-Johann Simon-Gabriel, Tong He, Zheng Zhang, Bernhard Scholkopf, Thomas Brox, and Francesco Locatello. Bridging the gap to real-world object-centric learning. In ICLR, 2023. 3", + "[42] Hyun Seok Seong, WonJun Moon, SuBeen Lee, and Jae-Pil Heo. Leveraging hidden positives for unsupervised semantic segmentation. In CVPR, 2023. 3", + "[43] Xiangheng Shan, Dongyue Wu, Guilin Zhu, Yuanjie Shao, Nong Sang, and Changxin Gao. Open-vocabulary semantic segmentation with image embedding balancing. In CVPR, 2024. 1", + "[44] Tong Shao, Zhuotao Tian, Hang Zhao, and Jingyong Su. Explore the potential of clip for training-free open vocabulary semantic segmentation. In ECCV, 2024. 3, 6, 7", + "[45] Gyungin Shin, Weidi Xie, and Samuel Albanie. ReCo: Retrieve and Co-segment for Zero-shot Transfer. In NeurIPS, 2022. 6", + "[46] Oriane Simeoni, Gilles Puy, Huy V Vo, Simon Roburin, Spyros Gidaris, Andrei Bursuc, Patrick Pérez, Renaud Marlet, and Jean Ponce. Localizing objects with self-supervised transformers and no labels. In BMVC, 2021. 3" + ], + "bbox": [ + 91, + 92, + 480, + 898 + ], + "page_idx": 9 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[47] Shuyang Sun, Runjia Li, Philip Torr, Xiuye Gu, and Siyang Li. Clip as rn: Segment countless visual concepts without training endeavor. In CVPR, 2024. 2, 6", + "[48] Feng Wang, Jieru Mei, and Alan Yuille. Sclip: Rethinking self-attention for dense vision-language inference. In ECCV, 2024. 1, 2, 3, 6, 7, 8", + "[49] Xudong Wang, Rohit Girdhar, Stella X Yu, and Ishan Misra. Cut and Learn for Unsupervised Object Detection and Instance Segmentation. In CVPR, 2023. 3", + "[50] Xin Wen, Bingchen Zhao, Anlin Zheng, Xiangyu Zhang, and Xiaojuan Qi. Self-supervised visual representation learning with semantic grouping. NeurIPS, 2022. 3", + "[51] Ji-Jia Wu, Andy Chia-Hao Chang, Chieh-Yu Chuang, Chun-Pei Chen, Yu-Lun Liu, Min-Hung Chen, Hou-Ning Hu, Yung-Yu Chuang, and Yen-Yu Lin. Image-text codecomposition for text-supervised semantic segmentation. In CVPR, 2024. 6", + "[52] Monika Wysoczanska, Oriane Simeoni, Michael Ramamonjisoa, Andrei Bursuc, Tomasz Trzciński, and Patrick Pérez. Clip-dinoiser: Teaching clip a few dino tricks. arXiv preprint arXiv:2312.12359, 2023. 6, 7", + "[53] Bin Xie, Jiale Cao, Jin Xie, Fahad Shahbaz Khan, and Yanwei Pang. Sed: A simple encoder-decoder for open-vocabulary semantic segmentation. In CVPR, 2024. 1", + "[54] Jiarui Xu, Shalini De Mello, Sifei Liu, Wonmin Byeon, Thomas Breuel, Jan Kautz, and Xiaolong Wang. GroupViT: Semantic Segmentation Emerges From Text Supervision. In CVPR, 2022. 1, 6", + "[55] Jiarui Xu, Sifei Liu, Arash Vahdat, Wonmin Byeon, Xiaolong Wang, and Shalini De Mello. Open-vocabulary panoptic segmentation with text-to-image diffusion models. In CVPR, 2023. 1", + "[56] Mengde Xu, Zheng Zhang, Fangyun Wei, Han Hu, and Xiang Bai. Side adapter network for open-vocabulary semantic segmentation. In CVPR, 2023. 1", + "[57] Bolei Zhou, Hang Zhao, Xavier Puig, Tete Xiao, Sanja Fidler, Adela Barriuso, and Antonio Torralba. Semantic Understanding of Scenes Through the ADE20K Dataset. IJCV, 2019. 6, 7", + "[58] Chong Zhou, Chen Change Loy, and Bo Dai. Extract Free Dense Labels from CLIP. In ECCV, 2022. 1, 2, 6", + "[59] Lianghui Zhu, Yingyue Li, Jiemin Fang, Yan Liu, Hao Xin, Wenyu Liu, and Xinggang Wang. Weaktr: Exploring plain vision transformer for weakly-supervised semantic segmentation. arXiv preprint arXiv:2304.01184, 2023. 4" + ], + "bbox": [ + 516, + 92, + 903, + 729 + ], + "page_idx": 9 + }, + { + "type": "page_number", + "text": "15042", + "bbox": [ + 480, + 944, + 519, + 955 + ], + "page_idx": 9 + } +] \ No newline at end of file diff --git a/2025/Distilling Spectral Graph for Object-Context Aware Open-Vocabulary Semantic Segmentation/7ae5aeda-7866-414e-b296-550ab2d843cd_model.json b/2025/Distilling Spectral Graph for Object-Context Aware Open-Vocabulary Semantic Segmentation/7ae5aeda-7866-414e-b296-550ab2d843cd_model.json new file mode 100644 index 0000000000000000000000000000000000000000..08d5aa8026bee9f275b468005515eb5e3ed0b78d --- /dev/null +++ b/2025/Distilling Spectral Graph for Object-Context Aware Open-Vocabulary Semantic Segmentation/7ae5aeda-7866-414e-b296-550ab2d843cd_model.json @@ -0,0 +1,2189 @@ +[ + [ + { + "type": "header", + "bbox": [ + 0.107, + 0.003, + 0.182, + 0.043 + ], + "angle": 0, + "content": "CVF" + }, + { + "type": "header", + "bbox": [ + 0.237, + 0.001, + 0.812, + 0.047 + ], + "angle": 0, + "content": "This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore." + }, + { + "type": "title", + "bbox": [ + 0.238, + 0.131, + 0.761, + 0.177 + ], + "angle": 0, + "content": "Distilling Spectral Graph for Object-Context Aware Open-Vocabulary Semantic Segmentation" + }, + { + "type": "text", + "bbox": [ + 0.131, + 0.203, + 0.866, + 0.224 + ], + "angle": 0, + "content": "Chanyoung \\(\\mathrm{Kim}^1\\) Dayun Ju\\(^1\\) Woojung Han\\(^1\\) Ming-Hsuan Yang\\(^{1,2}\\) Seong Jae Hwang\\(^1\\)" + }, + { + "type": "text", + "bbox": [ + 0.28, + 0.238, + 0.717, + 0.257 + ], + "angle": 0, + "content": "\\(^{1}\\)Yonsei University \\(^{2}\\)University of California, Merced" + }, + { + "type": "text", + "bbox": [ + 0.174, + 0.26, + 0.819, + 0.275 + ], + "angle": 0, + "content": "{chanyoung, juda0707, dnwjDDL, seongjae}@yonsei.ac.kr,mhyang@ucmerced.edu" + }, + { + "type": "title", + "bbox": [ + 0.248, + 0.309, + 0.327, + 0.325 + ], + "angle": 0, + "content": "Abstract" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.342, + 0.485, + 0.706 + ], + "angle": 0, + "content": "Open-Vocabulary Semantic Segmentation (OVSS) has advanced with recent vision-language models (VLMs), enabling segmentation beyond predefined categories through various learning schemes. Notably, training-free methods offer scalable, easily deployable solutions for handling unseen data, a key goal of OVSS. Yet, a critical issue persists: lack of object-level context consideration when segmenting complex objects in the challenging environment of OVSS based on arbitrary query prompts. This oversight limits models' ability to group semantically consistent elements within object and map them precisely to user-defined arbitrary classes. In this work, we introduce a novel approach that overcomes this limitation by incorporating object-level contextual knowledge within images. Specifically, our model enhances intra-object consistency by distilling spectral-driven features from vision foundation models into the attention mechanism of the visual encoder, enabling semantically coherent components to form a single object mask. Additionally, we refine the text embeddings with zero-shot object presence likelihood to ensure accurate alignment with the specific objects represented in the images. By leveraging object-level contextual knowledge, our proposed approach achieves state-of-the-art performance with strong generalizability across diverse datasets." + }, + { + "type": "title", + "bbox": [ + 0.092, + 0.725, + 0.222, + 0.74 + ], + "angle": 0, + "content": "1. Introduction" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.752, + 0.484, + 0.873 + ], + "angle": 0, + "content": "Open-vocabulary semantic segmentation (OVSS) [9, 32, 54, 55, 58] aims to predict pixel-level labels for arbitrary prompts defined by user input (e.g., the proper noun Space Needle in Fig. 1(a)), advancing beyond previous models scoped to predefined static classes [10, 18, 19, 40]. To achieve this, fully supervised OVSS models [11, 53, 56] enhance generalization to seen classes using labeled training data. Yet, their reliance on seen classes during training may" + }, + { + "type": "image", + "bbox": [ + 0.518, + 0.309, + 0.905, + 0.49 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.521, + 0.491, + 0.62, + 0.502 + ], + "angle": 0, + "content": "(b) Visualization" + }, + { + "type": "image_caption", + "bbox": [ + 0.681, + 0.491, + 0.73, + 0.502 + ], + "angle": 0, + "content": "Truck" + }, + { + "type": "image_caption", + "bbox": [ + 0.732, + 0.491, + 0.812, + 0.502 + ], + "angle": 0, + "content": "Horse" + }, + { + "type": "image_caption", + "bbox": [ + 0.814, + 0.491, + 0.865, + 0.502 + ], + "angle": 0, + "content": "Cow" + }, + { + "type": "image_caption", + "bbox": [ + 0.866, + 0.491, + 0.901, + 0.502 + ], + "angle": 0, + "content": "Sheep" + }, + { + "type": "image", + "bbox": [ + 0.517, + 0.505, + 0.905, + 0.636 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.512, + 0.639, + 0.907, + 0.806 + ], + "angle": 0, + "content": "Figure 1. We present CASS, object-level Context-Aware training-free open-vocabulary Semantic Segmentation model. (a) Overview: Our method distills the vision foundation model's (VFM) object-level contextual spectral graph into CLIP's attention and refines query text embeddings towards object-specific semantics. (b) Object-Level Context: This result illustrates how incorporating object-level context improves segmentation accuracy by unifying object-wise components into a user-defined object class; for example, (top row) our approach precisely segments the truck's body, wheels, and cargo area and (bottom row) accurately groups elements into a single object, such as the sheep, whereas baselines [16, 29] often fail to achieve this unified segmentation." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.824, + 0.905, + 0.854 + ], + "angle": 0, + "content": "risk overfitting and limit scalability, as labeled data is required for retraining to adapt to new domains [43]." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.856, + 0.907, + 0.901 + ], + "angle": 0, + "content": "To better address this, recent studies [16, 28, 29, 48, 58] propose training-free OVSS methods that leverage pretrained vision-language models (VLMs) (e.g., CLIP [39]) to" + }, + { + "type": "page_footnote", + "bbox": [ + 0.114, + 0.888, + 0.422, + 0.9 + ], + "angle": 0, + "content": "Project Page: https://micv-yonsei.github.io/cass/" + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "15033" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.09, + 0.092, + 0.482, + 0.258 + ], + "angle": 0, + "content": "integrate aligned visual and textual representations. These works focus on improving semantic association between visual patches in CLIP image encoder to produce more precise segmentation maps, achieving highly promising results. This training-free scheme offers key advantages: it (1) generalizes well to unseen classes without requiring additional labeled data, making it more scalable, thus (2) allows direct adaptation to dynamic real-world applications. Capitalizing on these benefits, training-free OVSS aligns with OVSS's core goal of adaptability across diverse domains, enabling it to reach its full potential." + }, + { + "type": "text", + "bbox": [ + 0.093, + 0.262, + 0.485, + 0.549 + ], + "angle": 0, + "content": "We tackle OVSS in a training-free manner, presenting a challenging and practical approach aligned with OVSS's true objectives. Although existing training-free methods [21, 29, 47, 48] perform well on numerous tasks, they all share a common limitation: the lack of \"object-level context.\" To understand \"object-level context\", consider a truck and sheep in Fig. 1(b). Components comprising a truck, such as its wheels and cargo area, should be grouped as a single entity. Similarly, components of a sheep should be accurately unified under the correct user-given object class. However, existing training-free OVSS methods, which use CLIP as the image backbone for segmentation, often struggle to capture object-level context, failing to group object components into a single, user-defined object. These challenges stem from CLIP's focus on learning global image semantics, which may be insufficient for capturing dense object-level semantics [16, 48, 58]. Thus, to achieve accurate training-free OVSS, a proper object-level context must be considered in CLIP." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.553, + 0.483, + 0.902 + ], + "angle": 0, + "content": "We address the challenges of training-free OVSS by introducing the object-level context and propose the Context-Aware Semantic Segmentation (CASS) model. Vision foundation models (VFMs), such as DINO [8, 38], capture fine-grained, patch-level semantics but lack object-level context, making them unsuitable for direct use in tasks requiring such context. Recent studies [24, 36] have shown that applying graph spectral techniques to VFM attention graphs can transform patch-level representations into object-level representations. Accordingly, we apply spectral techniques to extract essential object-level context from the VFM attention graph. The extracted spectral features can be used directly for segmentation (e.g., clustering eigenvectors); however, this approach may not be suitable for OVSS, where textual alignment is essential. Thus, we propose distilling low-rank components of VFM into CLIP to enhance its object-level context understanding, while maintaining alignment with textual information. Specifically, we decompose the VFM's attention graph to extract low-rank components, filtering irrelevant information while emphasizing essential object-level context. These low-rank components are then distilled into CLIP's attention to embed the fundamental object structure of the VFM graph." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.092, + 0.905, + 0.288 + ], + "angle": 0, + "content": "Our model also leverages CLIP's highly effective zero-shot object classification capability (i.e., object presence prior), widely validated in prior work [20, 34, 39], to capture detailed object-level context within scenes. As such, we adjust the text embeddings based on the object presence prior encoded by CLIP. This process involves refining the embeddings to better align with object-specific semantics. Then, we refine patch-text similarities using the object presence prior, ensuring that the final segmentation map reflects an object-specific perspective. Our CASS, carefully designed to enhance object-level contextual knowledge, captures richer intra-object coherence and enables cohesive groupings of semantically related elements." + }, + { + "type": "text", + "bbox": [ + 0.529, + 0.29, + 0.798, + 0.303 + ], + "angle": 0, + "content": "The main contributions of this work are:" + }, + { + "type": "text", + "bbox": [ + 0.514, + 0.305, + 0.905, + 0.365 + ], + "angle": 0, + "content": "- We enhance object-level context by distilling rich semantic understanding from VFM into the CLIP attention mechanism, facilitating a more consistent grouping of each object into a unified semantic entity." + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.366, + 0.905, + 0.41 + ], + "angle": 0, + "content": "- We present object presence prior-driven text embeddings and patch-text similarity refinement, promoting the model to precisely classify objects." + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.411, + 0.905, + 0.456 + ], + "angle": 0, + "content": "- We demonstrate the performance of our model in various semantic segmentation datasets and achieve state-of-the-art performance in training-free OVSS." + }, + { + "type": "list", + "bbox": [ + 0.513, + 0.305, + 0.905, + 0.456 + ], + "angle": 0, + "content": null + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.467, + 0.664, + 0.482 + ], + "angle": 0, + "content": "2. Related Works" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.489, + 0.905, + 0.821 + ], + "angle": 0, + "content": "Open-Vocabulary Semantic Segmentation. Open-vocabulary semantic segmentation (OVSS) aims to segment images corresponding to arbitrary user-given prompts, using vision-language aligned representations (e.g., CLIP [39]). Meanwhile, training-free approaches [3, 4, 23, 29, 58] have risen for their ability to adapt directly to new tasks without extra training. Among these, methods [16, 29, 48, 58] that do not rely on additional datasets have gained attention, as they require no extra adaptation effort, aligning better with the objective of OVSS. Earlier approaches investigated leveraging CLIP features for better localization [30, 58] or for grounding proposed masks [21, 47]. SCLIP [48] uses correlative self-attention to improve inter-patch correlations. NACLIP [16] applies a Gaussian kernel to attention maps for better spatial consistency, and ProxyCLIP [29] integrates patch-wise consistent features from DINO [8]. Although these methods demonstrated strong results, they often struggle to group different parts of the same object due to a lack of explicit dense label knowledge. To this end, we attribute the underlying challenges of training-free OVSS by enriching the object-level contextual understanding." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.825, + 0.905, + 0.901 + ], + "angle": 0, + "content": "Vision Foundation Models. Vision Foundation Models (VFMs), especially vision transformer (ViT)-based [14] self-supervised models (e.g., DINO [8, 38]), have significantly advanced computer vision through their ability to capture patch-wise semantic understanding at both the" + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "15034" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.095, + 0.089, + 0.9, + 0.307 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.09, + 0.31, + 0.907, + 0.355 + ], + "angle": 0, + "content": "Figure 2. Detailed illustration of our proposed training-free spectral object-level context distillation mechanism in Sec. 3.2. By matching the attention graphs of VFM and CLIP head-by-head to establish complementary relationships, and distilling the fundamental object-level context of the VFM graph to CLIP, we enhance CLIP's ability to capture intra-object contextual coherence." + }, + { + "type": "text", + "bbox": [ + 0.089, + 0.366, + 0.483, + 0.531 + ], + "angle": 0, + "content": "feature and attention levels. Among various downstream tasks [17, 35, 42, 46, 49] leveraging VFM's capability, one line of research focuses on leveraging the attention components in VFM [24, 36, 46], which exhibit a detailed understanding of patch-wise semantic interaction between different image regions. Thus, while ProxyCLIP [29] utilizes VFM features, our method takes a different effort by treating the attention mechanism in VFM as a graph. We then extract core structural patterns of the VFM graph to emphasize object-level context, providing an effective solution for producing precise object masks in training-free OVSS." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.536, + 0.484, + 0.807 + ], + "angle": 0, + "content": "Object-Level Semantic Segmentation. Semantic segmentation groups object components that have similar semantics into a single mask, traditionally achieved through models trained on extensive human-labeled, pixel-wise annotations [10, 19, 25, 40]. However, despite recent advancements in unsupervised semantic segmentation [17, 42] (USS) that do not rely on dense labels, a core challenge persists: individual object components are not consistently grouped into a single entity. To address this issue, recent USS studies [24, 41, 50] have incorporated object-level representation learning to improve the model's ability to merge single objects into unified entities. However, these efforts have not been considered in the training-free OVSS task, which faces similar challenges in an unsupervised context. Therefore, we focus on enriching object-level context in training-free OVSS to achieve a more effective grouping of components within each object, enabling clearer and more consistent segmentation across diverse visual domains." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.815, + 0.182, + 0.83 + ], + "angle": 0, + "content": "3. Method" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.84, + 0.484, + 0.903 + ], + "angle": 0, + "content": "We detail our proposed model, CASS, implemented in a training-free manner, as outlined in Fig 1(a). Our model jointly enhances object-level contextual understanding with two primary aspects: (1) Spectral Object-Level" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.365, + 0.907, + 0.412 + ], + "angle": 0, + "content": "Context Distillation (Sec. 3.2) and (2) Object Presence-Driven Object-Level Context (Sec. 3.3). We first outline the basic pipeline of training-free OVSS in Sec. 3.1." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.419, + 0.655, + 0.433 + ], + "angle": 0, + "content": "3.1. Preliminaries" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.438, + 0.907, + 0.665 + ], + "angle": 0, + "content": "CLIP Visual Encoder. Training-free OVSS models [16, 29, 48] leverage CLIP [39] for its rich alignment of visual and linguistic information. However, the CLIP ViT visual encoder \\(\\mathcal{F}_{\\mathrm{CLIP}}^v\\) focuses on aligning the [CLS] token with text embeddings during training, often neglecting the patchwise spatial interactions among visual tokens that are crucial for dense prediction [16, 44, 48]. To produce CLIP feature representations more suitable for semantic segmentation tasks, we remove residual connection, feed-forward network, and self-attention layer in the final layer of \\(\\mathcal{F}_{\\mathrm{CLIP}}^v\\), following recent works [16, 48]. Specifically, when the set of visual tokens \\(Z = \\mathrm{Concat}[z_1, \\ldots, z_N] \\in \\mathbb{R}^{N \\times D_Z}\\), where \\(N\\) is the total number of patches except [CLS] token and \\(D_Z\\) denotes patch dimension space, is input into the final block of the ViT, the following operations are as" + }, + { + "type": "equation", + "bbox": [ + 0.643, + 0.669, + 0.905, + 0.685 + ], + "angle": 0, + "content": "\\[\nZ ^ {*} = \\operatorname {S A} (\\operatorname {L N} (Z)), \\tag {1}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.688, + 0.906, + 0.718 + ], + "angle": 0, + "content": "where SA and LN represent Self-Attention and Layer Normalization, respectively. The modified SA layer is" + }, + { + "type": "equation", + "bbox": [ + 0.592, + 0.721, + 0.905, + 0.746 + ], + "angle": 0, + "content": "\\[\nM _ {\\mathrm {C L I P}} ^ {i} = \\operatorname {s o f t m a x} \\left(A _ {\\mathrm {C L I P}} ^ {i} / \\sqrt {D _ {h}}\\right), \\tag {2}\n\\]" + }, + { + "type": "equation", + "bbox": [ + 0.534, + 0.749, + 0.905, + 0.768 + ], + "angle": 0, + "content": "\\[\nZ ^ {*} = \\operatorname {C o n c a t} \\left[ M _ {\\mathrm {C L I P}} ^ {1} V _ {\\mathrm {C L I P}} ^ {1}, \\dots , M _ {\\mathrm {C L I P}} ^ {h} V _ {\\mathrm {C L I P}} ^ {h} \\right] W ^ {O}, \\tag {3}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.777, + 0.907, + 0.868 + ], + "angle": 0, + "content": "where \\(D_h\\) denotes the head dimension, \\(h\\) is the number of heads, and \\(W^0 \\in \\mathbb{R}^{D_z \\times D_z}\\) is the output projection matrix. We define CLIP attention map for \\(i\\)-th head as \\(A_{\\mathrm{CLIP}}^i = K_{\\mathrm{CLIP}}^i K_{\\mathrm{CLIP}}^{i^\\top} \\in \\mathbb{R}^{N \\times N}\\), where attention key \\(K_{\\mathrm{CLIP}} \\in \\mathbb{R}^{N \\times D_h \\times h}\\) and value \\(V_{\\mathrm{CLIP}} \\in \\mathbb{R}^{N \\times D_h \\times h}\\) result from the linear transformation of \\(\\mathrm{LN}(Z)\\)." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.871, + 0.907, + 0.902 + ], + "angle": 0, + "content": "Inference Pipeline. We use a sliding window inference to ensure that ViT can capture detailed parts of the image by" + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "15035" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.095, + 0.089, + 0.48, + 0.329 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.09, + 0.331, + 0.483, + 0.415 + ], + "angle": 0, + "content": "Figure 3. Attention score visualization for various query points. Left: Vanilla CLIP (\\(A_{\\mathrm{CLIP}}\\)) shows noisy, unfocused attention. Center: VFM-to-CLIP distillation without low-rank eigenscaling shows partial object grouping with limited detail. Right: Incorporating our low-rank eigenscaling (Eq. (7)) captures object-level context, improving grouping within a single object." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.427, + 0.484, + 0.549 + ], + "angle": 0, + "content": "dividing the entire image \\(\\mathcal{I}\\) into smaller windows \\(\\hat{\\mathcal{I}}\\). The visual tokens \\(Z^{*} \\in \\mathbb{R}^{N \\times D_{Z}}\\) are then computed from each \\(\\hat{\\mathcal{I}}\\) through \\(\\mathcal{F}_{\\mathrm{CLIP}}^{v}\\) as in Eq. (1). To perform OVSS with given texts, we project the visual tokens \\(Z^{*}\\) into the vision-language latent space as \\(F_{\\mathrm{CLIP}} = \\mathcal{P}_{\\mathrm{CLIP}}(Z^{*}) \\in \\mathbb{R}^{N \\times d}\\), where \\(\\mathcal{P}_{\\mathrm{CLIP}}\\) is a projection matrix into CLIP latent space and \\(d\\) is the CLIP latent dimension. We compute patch-text similarity \\(\\hat{S} \\in \\mathbb{R}^{N \\times C}\\) for each image window \\(\\hat{\\mathcal{I}}\\) as" + }, + { + "type": "equation", + "bbox": [ + 0.193, + 0.559, + 0.483, + 0.587 + ], + "angle": 0, + "content": "\\[\n\\hat {\\mathcal {S}} = F _ {\\mathrm {C L I P}} \\left[ \\left\\{t _ {\\mathrm {C L I P}} ^ {i} \\right\\} _ {i = 1} ^ {C} \\right] ^ {\\top}, \\tag {4}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.596, + 0.484, + 0.659 + ], + "angle": 0, + "content": "where \\(\\{t_{\\mathrm{CLIP}}^i\\}_{i = 1}^C \\coloneqq \\mathrm{Concat}\\left[t_{\\mathrm{CLIP}}^1,\\dots ,t_{\\mathrm{CLIP}}^C\\right] \\in \\mathbb{R}^{C\\times d}\\) denotes the encoded text embeddings, and \\(C\\) is the number of classes. Finally, the results of all windows are combined to create a segmentation map for the entire image." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.667, + 0.45, + 0.684 + ], + "angle": 0, + "content": "3.2. Spectral Object-Level Context Distillation" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.689, + 0.484, + 0.901 + ], + "angle": 0, + "content": "Our model CASS aims to capture such object-level contextual interaction within the image features to associate different parts of the same object. However, the limited ability of CLIP to capture semantic relationships between patches often neglects the object-level context, as illustrated in the left column of Fig. 3 (\\(A_{\\mathrm{CLIP}}\\)). To address this limitation, we leverage VFM (e.g., DINO [8]), enabling a deeper object-level contextual understanding. Among VFM features, we utilize VFM attention key component \\(K_{\\mathrm{VFM}} \\in \\mathbb{R}^{h \\times N \\times D_h}\\). Specifically, we treat the VFM attention adjacency \\(A_{\\mathrm{VFM}} = K_{\\mathrm{VFM}} K_{\\mathrm{VFM}}^{\\top} \\in \\mathbb{R}^{h \\times N \\times N}\\) as a graph, where \\(h\\) represents the number of heads, and transfer crucial graph patterns from \\(A_{\\mathrm{VFM}}\\) to \\(A_{\\mathrm{CLIP}}\\). However, a key question remains: \"How do we account for the multi-head attention mechanisms in both" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.091, + 0.906, + 0.213 + ], + "angle": 0, + "content": "VFM and CLIP?\" In other words, \"how do we match \\(i\\)-th VFM attention head \\(A_{\\mathrm{VFM}}^i\\) with \\(j\\)-th CLIP attention head \\(A_{\\mathrm{CLIP}}^j\\)? A straightforward approach is to distill the attention heads sequentially (i.e., \\(i = j\\)). Nevertheless, recent studies [22, 31, 59] indicate that different attention heads focus on separate parts of the image. Thus, matching the optimal heads between VFM and CLIP is essential for precise attention distillation." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.221, + 0.857, + 0.236 + ], + "angle": 0, + "content": "3.2.1. Complementary Spectral Graph Matching" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.237, + 0.906, + 0.357 + ], + "angle": 0, + "content": "We match graphs with contrasting structures, as shown in Fig 4, enabling VFM to supplement the object-level contextual knowledge that CLIP alone struggles to capture. To match an optimal head-wise attention graph, we exploit the spectral field of both \\( A_{\\mathrm{VFM}} \\) and \\( A_{\\mathrm{CLIP}} \\). Our graph-matching mechanism has two main steps as illustrated in Fig. 2(a): (I) examining the eigenvalues of each graph and (II) matching optimal graph pairs using the spectral distribution." + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.361, + 0.906, + 0.468 + ], + "angle": 0, + "content": "(I) Eigenvalue Examination. For each graph from VFM and CLIP, denoted as \\( A_{\\mathrm{VFM}}^i \\) and \\( A_{\\mathrm{CLIP}}^i \\) for \\( (i = 1, \\dots, h) \\), we perform eigendecomposition. This yields the eigenvalues for each head \\( h \\), \\( \\lambda_{\\mathrm{VFM}}^i \\) and \\( \\lambda_{\\mathrm{CLIP}}^i \\), from which the top \\( m \\) fixed eigenvalues are selected. These eigenvalues contain each graph's unique structural features, key properties essential for distinguishing different attention graph patterns." + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.47, + 0.906, + 0.532 + ], + "angle": 0, + "content": "(II) Graph Matching via Spectral Distribution. After obtaining eigenvalues for each head, we compute spectral distances to quantify structural differences, creating a cost matrix \\(\\mathcal{C} \\in \\mathbb{R}^{h \\times h}\\) for each graph pair from VFM and CLIP:" + }, + { + "type": "list", + "bbox": [ + 0.513, + 0.361, + 0.906, + 0.532 + ], + "angle": 0, + "content": null + }, + { + "type": "equation", + "bbox": [ + 0.53, + 0.542, + 0.905, + 0.561 + ], + "angle": 0, + "content": "\\[\n\\mathcal {C} _ {i j} = 1 - \\mathcal {D} _ {W} \\left(\\bar {\\lambda} _ {\\mathrm {V F M}} ^ {i}, \\bar {\\lambda} _ {\\mathrm {C L I P}} ^ {j}\\right), \\quad \\forall i, j \\in \\{1, \\dots , h \\} \\tag {5}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.57, + 0.907, + 0.755 + ], + "angle": 0, + "content": "where \\(\\bar{\\lambda}_{\\mathrm{VFM}}^i\\) and \\(\\bar{\\lambda}_{\\mathrm{CLIP}}^j\\) represent the normalized eigenvalues of the \\(i\\)-th head of VFM and the \\(j\\)-th head of CLIP, respectively, and \\(\\mathcal{D}_W\\) denotes the Wasserstein distance. Here, the Wasserstein distance is computed as \\(\\mathcal{D}_W(u,v) = \\sum_{i=1}^{h} |\\operatorname{sort}(u)_i - \\operatorname{sort}(v)_i|\\), where \\(u\\) and \\(v\\) are the two input distributions, and the sort function orders the values of \\(u\\) and \\(v\\) in ascending order. After computing \\(\\mathcal{C}\\), we apply the Hungarian matching algorithm [27] to the cost matrix to find the optimal pairing of graph heads. This approach pairs graphs with contrasting characteristics, enabling complementary distillation of object-level context knowledge from \\(A_{\\mathrm{VFM}}\\) to \\(A_{\\mathrm{CLIP}}\\), as described in the following section." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.761, + 0.838, + 0.776 + ], + "angle": 0, + "content": "3.2.2. Distilling VFM Spectral Graph to CLIP" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.777, + 0.906, + 0.901 + ], + "angle": 0, + "content": "Once we match optimal graph pairs, we distill the VFM graph to CLIP (see Fig. 2(b)). For the sake of explanation, we assume that the \\(i\\)-th VFM attention head \\(A_{\\mathrm{VFM}}^i\\) is matched with the \\(j\\)-th CLIP attention head \\(A_{\\mathrm{CLIP}}^j\\). An intuitive approach would be simply aggregating \\(A_{\\mathrm{VFM}}^i\\) and \\(A_{\\mathrm{CLIP}}^j\\) without any transformation. However, as shown in Fig. 3 (\\(A_{\\psi w/o \\mathrm{LR}}\\)), this approach may transfer noise or irrelevant information, highlighting the need to extract features" + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.52, + 0.957 + ], + "angle": 0, + "content": "15036" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.097, + 0.089, + 0.48, + 0.25 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.09, + 0.253, + 0.483, + 0.325 + ], + "angle": 0, + "content": "Figure 4. Visualization of the matched complementary graph pairs. The red dotted line indicates the focus areas of VFM, while the yellow dotted line highlights those for CLIP. By aggregating graphs with contrasting structural properties, we enhance the object-level context in \\( A_{\\mathrm{CLIP}} \\) supported by \\( \\ddot{A}_{\\mathrm{VFM}} \\), resulting \\( A_{\\psi} \\)." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.335, + 0.484, + 0.456 + ], + "angle": 0, + "content": "that emphasize object-level context. From this realization, we leverage the low-rank components of VFM, which contain distinct object patterns within the graph structure [12]. Specifically, we (I) extract the critical object-level contextual structure of \\( A_{\\mathrm{VFM}}^i \\) via low-rank approximation and enhance the graph structure by dynamically scaling eigenvalues. Then, we (II) distill the approximated VFM graph into CLIP by aggregating graphs based on spectral distance." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.459, + 0.483, + 0.686 + ], + "angle": 0, + "content": "(I) Low-Rank Dynamic Eigenscaling. Our goal is to transfer the essential object-level contextual structure of VFM graph \\( A_{\\mathrm{VFM}} \\) to CLIP. To achieve this, we extract low-rank components of the VFM graph using standard eigendecomposition, as \\( A_{\\mathrm{VFM}}^i = U\\Sigma U^\\top \\), where \\( U \\) and \\( \\Sigma \\) represent the eigenvector and the diagonal eigenvalue matrix, respectively. In the decomposed eigenbasis, we identify key object-level features of each graph by searching an optimal number of eigenvalues \\( k \\) through an energy-based approach. This ensures that the chosen \\( k \\) eigenvalues capture a significant portion of the graph's energy, retaining essential structural information while discarding noise and less relevant details. More details on searching \\( k \\) can be found in the supplementary material. This process yields a low-rank approximated VFM graph, expressed as \\( \\tilde{A}_{\\mathrm{VFM}}^i = U_k\\Sigma_kU_k^\\top \\)." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.686, + 0.483, + 0.822 + ], + "angle": 0, + "content": "We refine the low-rank components with a scaling function \\(\\phi\\), which dynamically amplifies larger eigenvalues and reduces smaller ones. Compared to the conventional shrinkage function [2, 5], which only focuses on noise cutoff, our approach emphasizes essential structural information, particularly object-level context features, while suppressing noise and irrelevant details. More details on dynamic eigenscaling can be found in the supplementary material. The graph after applying dynamic eigenscaling is as" + }, + { + "type": "equation", + "bbox": [ + 0.211, + 0.826, + 0.483, + 0.845 + ], + "angle": 0, + "content": "\\[\n\\ddot {A} _ {\\mathrm {V F M}} ^ {i} = U _ {k} \\phi (\\Sigma_ {k}) U _ {k} ^ {\\top}. \\tag {6}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.852, + 0.484, + 0.903 + ], + "angle": 0, + "content": "(II) VFM Graph Distillation. We distill the essential object-level contextual knowledge from the tailored VFM graph \\(\\ddot{A}_{\\mathrm{VFM}}^i\\) into the \\(A_{\\mathrm{CLIP}}^j\\) by aggregating their structures:" + }, + { + "type": "image", + "bbox": [ + 0.517, + 0.088, + 0.905, + 0.27 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.513, + 0.272, + 0.907, + 0.368 + ], + "angle": 0, + "content": "Figure 5. Detailed illustration of our object presence prior-guided text embedding adjustment module. The CLIP text encoder generates text embeddings for each object class, and the object presence prior is derived from both visual and text embeddings. Within hierarchically defined class groups, text embeddings are selected based on object presence prior, then refined in an object-specific direction to align with components likely present in the image." + }, + { + "type": "equation", + "bbox": [ + 0.582, + 0.38, + 0.905, + 0.401 + ], + "angle": 0, + "content": "\\[\nA _ {\\psi} ^ {j} = \\left(w _ {i j} \\ddot {A} _ {\\mathrm {V F M}} ^ {i} + A _ {\\mathrm {C L I P}} ^ {j}\\right) / \\left(w _ {i j} + 1\\right), \\tag {7}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.41, + 0.907, + 0.563 + ], + "angle": 0, + "content": "where \\(w_{ij}\\) is defined as the Wasserstein distance between eigenvalues from both graphs as \\(\\mathcal{D}_{\\mathcal{W}}(\\bar{\\lambda}_{\\mathrm{VFM}}^i,\\bar{\\lambda}_{\\mathrm{CLIP}}^j)\\). This adaptively assigns higher weights to graphs with complementary structures, while reducing influence on those with similar structures. Our distillation strategy effectively transfers the critical object-level context from VFM to CLIP, ensuring multiple components of the same object are grouped into a unified semantic as illustrated in Fig. 3 (\\(A_{\\psi}\\)). Therefore, we use \\(A_{\\psi}\\) as our attention matrix instead of \\(A_{\\mathrm{CLIP}}\\) in Eq. (2) to compute the final visual feature \\(F_{\\mathrm{CLIP}}\\)." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.573, + 0.903, + 0.59 + ], + "angle": 0, + "content": "3.3. Object Presence-Driven Object-Level Context" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.595, + 0.907, + 0.837 + ], + "angle": 0, + "content": "The spectral object-level context distillation in Sec. 3.2 enables precise object mask generation. However, due to the nature of OVSS, where users can provide arbitrary query prompts, different parts of the same object may still be assigned to closely related categories. Meanwhile, numerous studies [20, 34, 39] have demonstrated that CLIP excels in zero-shot object classification capability. Accordingly, we utilize the zero-shot object classification score encoded by CLIP (i.e., the object presence prior) to refine text embeddings and patch-text similarity, enhancing object-centered perspective. To compute the object presence prior \\( P \\), we take an entire input image \\( \\mathcal{I} \\) and compute likelihood for each class \\( i \\) as \\( P(i) = \\{t_{\\mathrm{CLIP}}^i\\}_{i=1}^C v_{\\mathrm{CLIP}} \\), where the visual embedding vector \\( v_{\\mathrm{CLIP}} = \\mathcal{F}_{\\mathrm{CLIP}}^v(\\mathcal{I}) \\in \\mathbb{R}^d \\) is derived from the ViT [CLS] token. The object presence prior is used in the following sections to enhance object-level context." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.84, + 0.907, + 0.902 + ], + "angle": 0, + "content": "Object-Guided Text Embedding Adjustment. We adjust text embeddings \\(\\{t_{\\mathrm{CLIP}}^i\\}_{i = 1}^C\\) to embed object-specific contextual semantics, as shown in Fig. 5. To prevent object parts from being assigned to ambiguous classes, we" + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "15037" + } + ], + [ + { + "type": "table_caption", + "bbox": [ + 0.117, + 0.09, + 0.878, + 0.104 + ], + "angle": 0, + "content": "Table 1. Quantitative results with state-of-the-art unsupervised open-vocabulary semantic segmentation models on eight datasets." + }, + { + "type": "table", + "bbox": [ + 0.095, + 0.108, + 0.904, + 0.394 + ], + "angle": 0, + "content": "
ModelSupporting DatasetExtra-TrainingFairV21PC60C-ObjV20PC59C-StfCityADEAvg.
GroupViT [54]CVPR'22CC12M+RedCaps✓✘50.418.727.579.723.415.311.19.229.4
TCL [9]CVPR'23CC3M+CC12M✓✘55.030.431.683.233.922.424.017.137.2
CoDe [51]CVPR'24CC3M+RedCaps✓✘57.530.532.3--23.928.917.7-
CLIP-DINOiser [52]ECCV'24ImageNet1k✓✘62.132.434.880.935.924.631.720.040.3
ReCo [45]NeurIPS'22ImageNet1k✘✘25.119.915.757.722.314.821.611.223.5
FOSSIL [3]WACV'24COCO Captions✘✘----35.824.823.218.8-
FreeDa [4]CVPR'24COCO Captions✘✘---85.643.127.836.722.4-
CLIP [39]ICML'21✘✘✓18.67.86.549.111.27.26.73.213.8
MaskCLIP [58]ECCV'22✘✘✓38.323.620.674.926.416.412.69.827.9
GEM [6]CVPR'24✘✘✓46.2---32.6--15.7-
CaR [47]CVPR'24✘✘✓48.613.615.473.718.4--5.4-
PnP-OVSS [33]CVPR'24✘✘✓--36.251.328.017.9-14.2-
CLIPtrase [44]ECCV'24✘✘✓50.929.943.681.033.822.821.316.432.7
ClearCLIP [28]ECCV'24✘✘✓51.832.633.080.935.923.930.016.738.1
SCLIP [48]ECCV'24✘✘✓59.130.430.580.434.122.432.216.138.2
LaVG [21]ECCV'24✘✘✓62.131.634.282.534.723.226.215.838.8
ProxyCLIP [29]ECCV'24✘✘✓59.135.236.278.238.826.238.119.641.4
NACLIP [16]WACV'25✘✘✓58.932.233.279.735.223.335.517.439.4
CASS✘✘✓65.836.737.887.840.226.739.420.444.4
" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.408, + 0.484, + 0.668 + ], + "angle": 0, + "content": "The hierarchical clustering of CLIP semantic distance, forming groups of similar embeddings \\(\\mathcal{T}_g = \\{t_{\\mathrm{CLIP}}^i\\mid i\\in I_g\\}\\), where \\(\\mathcal{T}_g\\) is a subset of entire text embeddings \\(\\{t_{\\mathrm{CLIP}}^i\\}_{i = 1}^C\\) with index set \\(I_{g}\\). Further details about this clustering process are provided in the supplementary materials. Within \\(\\mathcal{T}_g\\), we identify the object most likely to appear in the image using object presence prior \\(P\\), formulated as \\(i^{*} = \\arg \\max_{i\\in I_{g}}P(i)\\). This facilitates refining the text embedding vector \\(t_{\\mathrm{CLIP}}^*\\) to better represent object-contextual semantics. We note that patch-wise visual feature \\(F_{\\mathrm{CLIP}}\\in \\mathbb{R}^{N\\times d}\\) contains object-specific semantic vectors that can guide \\(t_{\\mathrm{CLIP}}^*\\) accordingly. Thus, we compute cosine similarity between \\(F_{\\mathrm{CLIP}}\\) and \\(t_{\\mathrm{CLIP}}^*\\) and select top-\\(n\\) object-specific image vector denoted as \\(\\{f_i\\}_{i = 1}^n\\coloneqq\\) \\(\\mathrm{Concat}[f_1,\\ldots ,f_n]\\in \\mathbb{R}^{n\\times d}\\). To define a targeted direction, we compute the average of \\(\\{f_i\\}_{i = 1}^n\\) as \\(\\mu_{n}\\in \\mathbb{R}^{d}\\), and guide selected text embedding \\(t_{\\mathrm{CLIP}}^*\\) towards \\(\\mu_{n}\\) as" + }, + { + "type": "equation", + "bbox": [ + 0.177, + 0.67, + 0.482, + 0.689 + ], + "angle": 0, + "content": "\\[\n\\tilde {t} _ {\\mathrm {C L I P}} ^ {i ^ {*}} = (1 - \\alpha) \\cdot t _ {\\mathrm {C L I P}} ^ {i ^ {*}} + \\alpha \\cdot \\mu_ {n}, \\tag {8}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.089, + 0.695, + 0.483, + 0.761 + ], + "angle": 0, + "content": "where \\(\\alpha\\) controls the degree of alignment with \\(\\mu_{n}\\). The adjusted text embedding \\(\\tilde{t}_{\\mathrm{CLIP}}^{i^*}\\) replaces the original text embedding \\(t_{\\mathrm{CLIP}}^{i}\\) in the set \\(\\{t_{\\mathrm{CLIP}}^{i}\\}_{i=1}^{C}\\), which is then used to compute the patch-text similarity logit \\(\\hat{S}\\) in Eq. (4)." + }, + { + "type": "text", + "bbox": [ + 0.089, + 0.765, + 0.483, + 0.902 + ], + "angle": 0, + "content": "Object Perspective Patch-Text Similarity. We integrate the object presence prior with patch-text similarity \\(\\hat{S}\\) from Eq. (4) to produce a segmentation map from an object-perspective view. Unlike Barsellotti et al. [4], which computes likelihood from an image window \\(\\hat{\\mathcal{I}}\\), our approach derives similarity from the full image \\(\\mathcal{I}\\), enabling richer object-contextual understanding at the broader level. Thus, we refine the patch-text similarity \\(\\hat{S}\\) from Eq. (4) to obtain the final patch-text similarity \\(\\hat{S}^*\\), computed as" + }, + { + "type": "equation", + "bbox": [ + 0.573, + 0.405, + 0.905, + 0.427 + ], + "angle": 0, + "content": "\\[\n\\hat {\\mathcal {S}} ^ {*} = (1 - \\gamma) \\cdot \\hat {\\mathcal {S}} + \\gamma \\cdot \\left\\{t _ {\\mathrm {C L I P}} ^ {i} \\right\\} _ {i = 1} ^ {C} v _ {\\mathrm {C L I P}}, \\tag {9}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.43, + 0.905, + 0.475 + ], + "angle": 0, + "content": "where \\(\\gamma\\) controls the balance between \\(\\hat{S}\\) and object presence prior. This allows the final segmentation map to capture object-level details for precise, class-aware segmentation." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.481, + 0.646, + 0.498 + ], + "angle": 0, + "content": "4. Experiments" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.506, + 0.906, + 0.598 + ], + "angle": 0, + "content": "We first describe the experimental settings, including implementation details, datasets, and evaluation metrics (Sec. 4.1). We then present quantitative and qualitative evaluation results (Sec. 4.2) and demonstrate ablation studies to assess each proposed component (Sec. 4.3). Additional experiment results are provided in the supplementary material." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.605, + 0.722, + 0.621 + ], + "angle": 0, + "content": "4.1. Experimental Settings" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.624, + 0.907, + 0.759 + ], + "angle": 0, + "content": "Implementation Details. Similar to existing training-free approaches [16, 21, 29, 44, 48], we employ CLIP ViT-B/16 [39] as our VLM and DINO ViT-B/8 [8] as our VFM, keeping both frozen. Input images are matched with the shorter side set to 336 pixels (or 560 pixels for the higher-resolution Cityscapes dataset), and sliding window inference is applied with a \\(224 \\times 224\\) window and a stride of 112 to achieve efficient evaluation coverage. All experiments are conducted using NVIDIA RTX A6000 Ada*." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.76, + 0.906, + 0.88 + ], + "angle": 0, + "content": "Datasets. Our primary datasets include PASCAL VOC 2012 [15], PASCAL Context [37], and COCO [7], each evaluated with and without a background class. We refer to the versions with a background class as V21, PC60, and C-Obj, and those without as V20, PC59, and C-Stf. We also report results on the ADE20K (ADE) [57] and Cityscapes (City) [13] datasets. These datasets contain 21, 60, 81, 20, 59, 171, 150, and 19 classes, respectively." + }, + { + "type": "page_footnote", + "bbox": [ + 0.54, + 0.888, + 0.866, + 0.9 + ], + "angle": 0, + "content": "* Advanced Database System Infrastructure (NFEC-2024-11-300458)" + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "15038" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.093, + 0.088, + 0.905, + 0.322 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.09, + 0.322, + 0.905, + 0.351 + ], + "angle": 0, + "content": "Figure 6. Qualitative comparison across the Pascal VOC [15], Pascal Context [37], COCO [7], and ADE20K [57] datasets using CLIP ViT-B/16 [39], evaluating previous state-of-the-art training-free OVSS methods, including NACLIP [16], ProxyCLIP [29], and ours." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.361, + 0.273, + 0.375 + ], + "angle": 0, + "content": "4.2. Evaluation Results" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.381, + 0.483, + 0.488 + ], + "angle": 0, + "content": "We evaluate our model against training-free OVSS methods both quantitatively and qualitatively. For fair comparison, we refrain from using the mask refinement (e.g., PAMR [1] or DenseCRF [26]) and reproduce results from [29] using a \\(224 \\times 224\\) window to match the baselines. Performance is measured primarily with mean IoU (mIoU), complemented by pixel accuracy (pAcc) for a broader assessment." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.489, + 0.483, + 0.701 + ], + "angle": 0, + "content": "Quantitative Evaluation: mIoU. Table 1 presents the main results in mIoU, where all results are obtained using CLIP ViT-B/16 as the backbone model for fair comparisons. To demonstrate the strengths of our model, we also include methods that use supporting datasets (i.e., retrieving additional datasets during inference) or additional training (models evaluated under the same setting as ours are indicated with \\(\\checkmark\\) in the \"Fair\" column). Our model, CASS, performs favorably against the state-of-the-art methods, with an average gain of \\(3.0\\mathrm{mIoU}\\) point across eight datasets. Notably, for the V20, CASS achieves a performance gain of 5.3 mIoU point over the second-best model. Further, CASS outperforms CLIP-DINOiser [52] (which also leverages DINO with extra data and training) by \\(4.1\\mathrm{mIoU}\\) point." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.702, + 0.483, + 0.824 + ], + "angle": 0, + "content": "Quantitative Evaluation: pAcc. Table 2 shows the results of pAcc on recent state-of-the-art models evaluated under the same settings. Our CASS outperforms the second-highest model by \\(2.2\\mathrm{pAcc}\\) point and exceeds one of the most recent papers, CLIPtrase [44], by \\(9.8\\mathrm{pAcc}\\) point. Our model consistently performs better than existing methods on mIoU and pAcc by effectively capturing object-level context, resulting in more accurate object masking." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.825, + 0.483, + 0.902 + ], + "angle": 0, + "content": "Qualitative Analysis. Fig. 6 provides a qualitative comparison with recent state-of-the-art models [16, 29]. Existing methods often produce noisy segmentation maps and do not group object components correctly, such as wheels within a car or a person's head or arm. Additionally, they often mis" + }, + { + "type": "table_caption", + "bbox": [ + 0.536, + 0.36, + 0.882, + 0.374 + ], + "angle": 0, + "content": "Table 2. Quantitative results using average pixel accuracy." + }, + { + "type": "table", + "bbox": [ + 0.518, + 0.374, + 0.903, + 0.463 + ], + "angle": 0, + "content": "
ModelV21PC60C-ObjV20PC59C-StfCityADEAvg.
CLIPtrase [44]78.652.150.189.758.938.963.438.659.1
LaVG [21]89.348.774.891.158.939.168.537.063.4
SCLIP [48]87.649.274.391.058.338.472.738.763.8
ProxyCLIP [29]86.652.075.988.463.443.474.949.166.7
NACLIP [16]87.151.275.389.259.839.371.745.264.9
CASS90.155.676.093.965.043.678.248.668.9
" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.473, + 0.906, + 0.565 + ], + "angle": 0, + "content": "classify object parts into ambiguous classes (e.g., a motorcycle as a bicycle or an airplane as a car). In contrast, based on a deep understanding of object-level context, our CASS generates clean segmentation maps that accurately group all object components, including slender parts such as human arms or animal legs, into their respective object classes." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.574, + 0.667, + 0.59 + ], + "angle": 0, + "content": "4.3. Ablation Study" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.594, + 0.906, + 0.67 + ], + "angle": 0, + "content": "We perform ablation studies to analyze the proposed method. We use the datasets representative of segmentation tasks for all experiments: V21, PC59, C-Stf, following [48]. Please refer to the supplementary material for the additional ablation studies including VFM backbones and feature type." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.674, + 0.907, + 0.903 + ], + "angle": 0, + "content": "Effect of Spectral VFM Distillation. Table 3 shows the results of incorporating VFM (Exp.#3) with the vanilla setting (Exp.#1), which directly uses the output from Eq. (4). In Exp.#3, \\( A_{\\mathrm{VFM}} \\) and \\( A_{\\mathrm{CLIP}} \\) are aggregated in a straightforward head-wise manner, without any additional processing to \\( A_{\\mathrm{VFM}} \\). Simply leveraging \\( A_{\\mathrm{VFM}} \\) leads to a performance gain compared to the vanilla setting. Exp.#4 to Exp.#6 validate the effectiveness of the proposed components (see Sec. 3.2). Exp.#4 demonstrates the aggregation of \\( A_{\\mathrm{VFM}} \\) and \\( A_{\\mathrm{CLIP}} \\) using a complementary spectral graph matching (GM) procedure. Exp.#5, denoted as LR, presents the results with only the low-rank approximation applied to \\( A_{\\mathrm{VFM}} \\), while Exp.#6 shows the results with both the low-rank approximation and Dynamic Eigenscaling (DE) applied, as defined in Eq. (6). Notably, Exp.#6 shows signifi" + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "15039" + } + ], + [ + { + "type": "table_caption", + "bbox": [ + 0.091, + 0.09, + 0.483, + 0.131 + ], + "angle": 0, + "content": "Table 3. Ablation results showing the impact of our components on the performance across V21 [15], PC59 [37], and C-Stf [7]. The abbreviations in the table are noted in Sec. 4.3." + }, + { + "type": "table", + "bbox": [ + 0.095, + 0.136, + 0.48, + 0.268 + ], + "angle": 0, + "content": "
Exp. #AVFMGMLRDEOPSOTADataset
V21PC59C-Stf
158.935.323.4
2✓✓59.236.524.0
3✓62.937.725.2
4✓✓63.238.025.4
5✓✓✓64.538.425.8
6✓✓✓✓64.838.625.9
7✓✓✓✓✓65.039.426.6
8✓✓✓✓✓✓65.840.226.7
" + }, + { + "type": "image", + "bbox": [ + 0.095, + 0.271, + 0.481, + 0.384 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.09, + 0.387, + 0.483, + 0.443 + ], + "angle": 0, + "content": "Figure 7. Qualitative ablation of our method. Exp.# refers to the experiment in Table 3. As additional components of our proposed method are incorporated, the segmentation map exhibits greater refinement and more effectively captures object-level context." + }, + { + "type": "text", + "bbox": [ + 0.089, + 0.452, + 0.483, + 0.588 + ], + "angle": 0, + "content": "cant performance gain over simply integrating VFM knowledge (Exp.#3), demonstrating the effectiveness of graph matching and low-rank approximation in capturing object-level context. Fig. 3 qualitatively visualizes the attention map, with the left, middle, and right columns corresponding to Exp.#3, Exp.#4, and Exp.#6, respectively. We observe that integrating proposed low-rank dynamic eigenscaling effectively reduces noise and emphasizes objects by highlighting objects with strong inter-object associations." + }, + { + "type": "text", + "bbox": [ + 0.089, + 0.591, + 0.483, + 0.743 + ], + "angle": 0, + "content": "Effect of Object Presence Prior. We validate the effectiveness of the object presence prior-driven methods in Exp.#7 and Exp.#8 by examining each component: Object-Perspective Patch-Text Similarity (OPS) and Object-Guided Text Embedding Adjustment (OTA) (see Sec. 3.3). These results show that each component effectively improves performance, emphasizing object-level context. Additionally, incorporating the object presence prior-driven methods on top of the vanilla setting, as shown in Exp.#2, clarifies the effectiveness of the proposed method." + }, + { + "type": "text", + "bbox": [ + 0.089, + 0.746, + 0.483, + 0.882 + ], + "angle": 0, + "content": "Qualitative Ablations. Fig. 7 shows the segmentation maps for the vanilla setting (Exp.#1), after applying spectral VFM distillation (Exp.#6), and with all proposed components applied (Exp.#8). While Exp.#6 improves object masking compared to Exp.#1, it often misses parts of objects (highlighted by the red dotted line). In contrast, incorporating all proposed components (Exp.#8) consolidates objects into single entities, effectively capturing object-level context in the segmentation mask." + }, + { + "type": "text", + "bbox": [ + 0.091, + 0.886, + 0.483, + 0.9 + ], + "angle": 0, + "content": "Different CLIP Backbones. Table 4 shows results on dif-" + }, + { + "type": "table_caption", + "bbox": [ + 0.547, + 0.09, + 0.872, + 0.102 + ], + "angle": 0, + "content": "Table 4. Ablation results on different CLIP backbones." + }, + { + "type": "table", + "bbox": [ + 0.517, + 0.103, + 0.905, + 0.228 + ], + "angle": 0, + "content": "
ModelV21PC59C-StfAvg. mIoUAvg. pAcc
mIoUpAccmIoUpAccmIoUpAcc
VLT-B/32SCLIP [48]50.681.028.749.620.033.433.154.7
LaVG [21]54.884.429.050.120.533.734.856.1
ProxyCLIP [29]57.985.635.259.723.640.238.961.8
NACLIP [16]51.182.832.456.821.237.134.958.9
CASS58.286.236.561.024.440.839.762.7
VLT-L/14SCLIP [48]44.478.125.246.617.629.629.151.4
LaVG [21]51.584.127.550.319.432.332.855.6
ProxyCLIP [29]59.885.738.361.026.243.541.463.4
NACLIP [16]52.283.132.152.821.435.735.257.2
CASS62.188.039.162.126.343.442.564.5
" + }, + { + "type": "image", + "bbox": [ + 0.516, + 0.232, + 0.905, + 0.323 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.513, + 0.325, + 0.905, + 0.354 + ], + "angle": 0, + "content": "Figure 8. Performance evaluation of different distance metrics used in graph matching and graph distillation." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.363, + 0.907, + 0.438 + ], + "angle": 0, + "content": "ferent CLIP backbones, ViT-B/32 and ViT-L/14. Our CASS performs favorably across various backbones, consistently outperforming existing methods in mIoU and pAcc. These results demonstrate the effectiveness and robustness of our method across different backbone architectures." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.442, + 0.909, + 0.61 + ], + "angle": 0, + "content": "Distance Metric for Graph Matching. We evaluate different distance metrics \\(\\mathcal{D}_W\\) used in Sec. 3.2.1, including Wasserstein Distance, Cosine Distance, KL Divergence, and Euclidean Distance. As shown in Fig. 8, the Wasserstein Distance shows the best results as it provides an intuitive measure of the overall shape of sorted eigenvalue distributions. While we use Wasserstein Distance due to its performance, the other well-known metrics exhibited comparable results, demonstrating the robustness of our method in terms of the choice of a distance metric and further supporting the effectiveness of our proposed approach." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.616, + 0.634, + 0.631 + ], + "angle": 0, + "content": "5. Conclusion" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.635, + 0.907, + 0.83 + ], + "angle": 0, + "content": "In this work, we present CASS, a training-free approach for OVSS that integrates object-level context into CLIP. By distilling object-aware structural information from VFM into CLIP's attention, CASS effectively enhances intra-object coherence, ensuring that each object is consistently represented as a unified semantic entity. Additionally, we leverage object-presence prior knowledge to adjust text embeddings and refine object class assignments, aligning them with object-specific semantic targets in the image. Consequently, CASS effectively groups object components and assigns semantic labels without extra training or additional labeled data, making it a robust solution across diverse datasets and visual contexts." + }, + { + "type": "page_footnote", + "bbox": [ + 0.513, + 0.848, + 0.905, + 0.89 + ], + "angle": 0, + "content": "Acknowledgement. This work was supported in part by IITP RS-2024-00457882 (National AI Research Hub Project), IITP 2020-0-01361, NRF RS-2024-00345806, NRF RS-2023-00219019, and RS-2024-00403860 (Korea Basic Science Institute, National Research Facilities and Equipment Center)." + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.52, + 0.957 + ], + "angle": 0, + "content": "15040" + } + ], + [ + { + "type": "title", + "bbox": [ + 0.093, + 0.09, + 0.188, + 0.106 + ], + "angle": 0, + "content": "References" + }, + { + "type": "ref_text", + "bbox": [ + 0.101, + 0.116, + 0.482, + 0.143 + ], + "angle": 0, + "content": "[1] Nikita Araslanov and Stefan Roth. Single-stage semantic segmentation from image labels. In CVPR, 2020. 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.101, + 0.145, + 0.483, + 0.185 + ], + "angle": 0, + "content": "[2] Danny Barash and Matan Gavish. Optimal shrinkage of singular values under random data contamination. NeurIPS, 2017. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.188, + 0.483, + 0.243 + ], + "angle": 0, + "content": "[3] Luca Barsellotti, Roberto Amoroso, Lorenzo Baraldi, and Rita Cucchiara. Fossil: Free open-vocabulary semantic segmentation through synthetic references retrieval. In WACV, 2024. 2, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.246, + 0.483, + 0.3 + ], + "angle": 0, + "content": "[4] Luca Barsellotti, Roberto Amoroso, Marcella Cornia, Lorenzo Baraldi, and Rita Cucchiara. Training-free open-vocabulary segmentation with offline diffusion-augmented prototype generation. In CVPR, 2024. 2, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.303, + 0.482, + 0.344 + ], + "angle": 0, + "content": "[5] Toby Boas, Aritra Dutta, Xin Li, Kathryn P Mercier, and Eric Niderman. Shrinkage function and its applications in matrix approximation. arXiv preprint arXiv:1601.07600, 2016. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.346, + 0.482, + 0.399 + ], + "angle": 0, + "content": "[6] Walid Bousselham, Felix Petersen, Vittorio Ferrari, and Hilde Kuehne. Grounding everything: Emerging localization properties in vision-language transformers. In CVPR, 2024. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.403, + 0.482, + 0.442 + ], + "angle": 0, + "content": "[7] Holger Caesar, Jasper Uijlings, and Vittorio Ferrari. COCO-Stuff: Thing and Stuff Classes in Context. In CVPR, 2018. 6, 7, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.446, + 0.482, + 0.5 + ], + "angle": 0, + "content": "[8] Mathilde Caron, Hugo Touvron, Ishan Misra, Hervé Jégou, Julien Mairal, Piotr Bojanowski, and Armand Joulin. Emerging Properties in Self-Supervised Vision Transformers. In ICCV, 2021. 2, 4, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.504, + 0.482, + 0.557 + ], + "angle": 0, + "content": "[9] Junbum Cha, Jonghwan Mun, and Byungseok Roh. Learning To Generate Text-Grounded Mask for Open-World Semantic Segmentation From Only Image-Text Pairs. In CVPR, 2023. 1, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.561, + 0.482, + 0.614 + ], + "angle": 0, + "content": "[10] Liang-Chieh Chen, Yukun Zhu, George Papandreou, Florian Schroff, and Hartwig Adam. Encoder-decoder with atrous separable convolution for semantic image segmentation. In ECCV, 2018. 1, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.617, + 0.482, + 0.671 + ], + "angle": 0, + "content": "[11] Seokju Cho, Heeseong Shin, Sunghwan Hong, Anurag Arnab, Paul Hongsuck Seo, and Seungryong Kim. Cat-seg: Cost aggregation for open-vocabulary semantic segmentation. In CVPR, 2024. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.675, + 0.482, + 0.701 + ], + "angle": 0, + "content": "[12] Fan RK Chung. Spectral graph theory. American Mathematical Soc., 1997. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.704, + 0.482, + 0.77 + ], + "angle": 0, + "content": "[13] Marius Cordts, Mohamed Omran, Sebastian Ramos, Timo Rehfeld, Markus Enzweiler, Rodrigo Benenson, Uwe Franke, Stefan Roth, and Bernt Schiele. The Cityscapes Dataset for Semantic Urban Scene Understanding. In CVPR, 2016. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.775, + 0.482, + 0.856 + ], + "angle": 0, + "content": "[14] Alexey Dosovitskiy, Lucas Beyer, Alexander Kolesnikov, Dirk Weissenborn, Xiaohua Zhai, Thomas Unterthiner, Mostafa Dehghani, Matthias Minderer, Georg Heigold, Sylvain Gelly, Jakob Uszkoreit, and Neil Houlsby. An Image is Worth 16x16 Words: Transformers for Image Recognition at Scale. In ICLR, 2021. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.859, + 0.482, + 0.901 + ], + "angle": 0, + "content": "[15] M. Everingham, L. Van Gool, C. K. I. Williams, J. Winn, and A. Zisserman. The PASCAL Visual Object Classes Challenge 2012 (VOC2012) Results, 2012. 6, 7, 8" + }, + { + "type": "list", + "bbox": [ + 0.094, + 0.116, + 0.483, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.093, + 0.905, + 0.134 + ], + "angle": 0, + "content": "[16] Sina Hajimiri, Ismail Ben Ayed, and Jose Dolz. Pay attention to your neighbours: Training-free open-vocabulary semantic segmentation. In WACV, 2025. 1, 2, 3, 6, 7, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.137, + 0.905, + 0.192 + ], + "angle": 0, + "content": "[17] Mark Hamilton, Zhoutong Zhang, Bharath Hariharan, Noah Snavely, and William T Freeman. Unsupervised semantic segmentation by distilling feature correspondences. arXiv preprint arXiv:2203.08414, 2022. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.195, + 0.905, + 0.248 + ], + "angle": 0, + "content": "[18] Woojung Han, Seil Kang, Kyobin Choo, and Seong Jae Hwang. Complementary branch fusing class and semantic knowledge for robust weakly supervised semantic segmentation. Pattern Recognition, 2025. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.252, + 0.905, + 0.306 + ], + "angle": 0, + "content": "[19] Ali Hatamizadeh, Yucheng Tang, Vishwesh Nath, Dong Yang, Andriy Myronenko, Bennett Landman, Holger R Roth, and Daguang Xu. Unetr: Transformers for 3d medical image segmentation. In CVPR, 2022. 1, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.31, + 0.905, + 0.35 + ], + "angle": 0, + "content": "[20] Ruixiang Jiang, Lingbo Liu, and Changwen Chen. Clipcount: Towards text-guided zero-shot object counting. In ACM Multimedia, 2023. 2, 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.354, + 0.905, + 0.393 + ], + "angle": 0, + "content": "[21] Dahiyun Kang and Minsu Cho. In defense of lazy visual grounding for open-vocabulary semantic segmentation. In ECCV, 2024. 2, 6, 7, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.397, + 0.905, + 0.437 + ], + "angle": 0, + "content": "[22] Seil Kang, Jinyeong Kim, Junhyeok Kim, and Seong Jae Hwang. See what you are told: Visual attention sink in large multimodal models. In ICLR, 2025. 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.441, + 0.905, + 0.482 + ], + "angle": 0, + "content": "[23] Seil Kang, Jinyeong Kim, Junhyeok Kim, and Seong Jae Hwang. Your large vision-language model only needs a few attention heads for visual grounding. In CVPR, 2025. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.485, + 0.905, + 0.525 + ], + "angle": 0, + "content": "[24] Chanyoung Kim, Woogung Han, Dayun Ju, and Seong Jae Hwang. Eagle: Eigen aggregation learning for object-centric unsupervised semantic segmentation. In CVPR, 2024. 2, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.528, + 0.905, + 0.582 + ], + "angle": 0, + "content": "[25] Alexander Kirillov, Eric Mintun, Nikhila Ravi, Hanzi Mao, Chloe Rolland, Laura Gustafson, Tete Xiao, Spencer Whitehead, Alexander C Berg, Wan-Yen Lo, et al. Segment anything. In ICCV, 2023. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.585, + 0.905, + 0.626 + ], + "angle": 0, + "content": "[26] Philipp Krahenbuhl and Vladlen Koltun. Efficient Inference in Fully Connected CRFs with Gaussian Edge Potentials. In NeurIPS, 2011. 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.63, + 0.905, + 0.656 + ], + "angle": 0, + "content": "[27] Harold W Kuhn. The hungarian method for the assignment problem. Naval research logistics quarterly, 1955. 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.659, + 0.905, + 0.713 + ], + "angle": 0, + "content": "[28] Mengcheng Lan, Chaofeng Chen, Yiping Ke, Xinjiang Wang, Litong Feng, and Wayne Zhang. Clearclip: Decomposing clip representations for dense vision-language inference. In ECCV, 2024. 1, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.717, + 0.905, + 0.77 + ], + "angle": 0, + "content": "[29] Mengcheng Lan, Chaofeng Chen, Yiping Ke, Xinjiang Wang, Litong Feng, and Wayne Zhang. Proxyclip: Proxy attention improves clip for open-vocabulary segmentation. In ECCV, 2024. 1, 2, 3, 6, 7, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.774, + 0.905, + 0.828 + ], + "angle": 0, + "content": "[30] Yi Li, Hualiang Wang, Yiqun Duan, and Xiaomeng Li. Clip surgery for better explainability with enhancement in open-vocabulary tasks. arXiv preprint arXiv:2304.05653, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.832, + 0.905, + 0.901 + ], + "angle": 0, + "content": "[31] Yiran Li, Junpeng Wang, Xin Dai, Liang Wang, ChinChia Michael Yeh, Yan Zheng, Wei Zhang, and Kwan-Liu Ma. How does attention work in vision transformers? a visual analytics attempt. IEEE Transactions on Visualization and Computer Graphics, 2023. 4" + }, + { + "type": "list", + "bbox": [ + 0.517, + 0.093, + 0.905, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.518, + 0.957 + ], + "angle": 0, + "content": "15041" + } + ], + [ + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.093, + 0.482, + 0.147 + ], + "angle": 0, + "content": "[32] Feng Liang, Bichen Wu, Xiaoliang Dai, Kunpeng Li, Yinan Zhao, Hang Zhang, Peizhao Zhang, Peter Vajda, and Diana Marculescu. Open-vocabulary semantic segmentation with mask-adapted clip. In CVPR, 2023. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.149, + 0.482, + 0.204 + ], + "angle": 0, + "content": "[33] Jiayun Luo, Siddhesh Khandelwal, Leonid Sigal, and Boyang Li. Emergent open-vocabulary semantic segmentation from off-the-shelf vision-language models. In CVPR, 2024. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.207, + 0.482, + 0.247 + ], + "angle": 0, + "content": "[34] Yunyao Mao, Jiajun Deng, Wengang Zhou, Li Li, Yao Fang, and Houqiang Li. Clip4hoi: towards adapting clip for practical zero-shot hoi detection. NeurIPS, 2023. 2, 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.25, + 0.482, + 0.29 + ], + "angle": 0, + "content": "[35] Octave Mariotti, Oisin Mac Aodha, and Hakan Bilen. Improving semantic correspondence with viewpoint-guided spherical maps. In CVPR, 2024. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.293, + 0.482, + 0.346 + ], + "angle": 0, + "content": "[36] Luke Melas-Kyriazi, Christian Rupprecht, Iro Laina, and Andrea Vedaldi. Deep spectral methods: A surprisingly strong baseline for unsupervised semantic segmentation and localization. In CVPR, 2022. 2, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.349, + 0.482, + 0.404 + ], + "angle": 0, + "content": "[37] Roozbeh Mottaghi, Xianjie Chen, Xiaobai Liu, Nam-Gyu Cho, Seong-Whan Lee, Sanja Fidler, Raquel Urtasun, and Alan Yuille. The Role of Context for Object Detection and Semantic Segmentation in the Wild. In CVPR, 2014. 6, 7, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.406, + 0.482, + 0.473 + ], + "angle": 0, + "content": "[38] Maxime Oquab, Timothee Darct, Theo Moutakanni, Huy Vo, Marc Szafraniec, Vasil Khalidov, Pierre Fernandez, Daniel Haziza, Francisco Massa, Alaaeldin El-Nouby, et al. Dinov2: Learning robust visual features without supervision. arXiv preprint arXiv:2304.07193, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.476, + 0.482, + 0.545 + ], + "angle": 0, + "content": "[39] Alec Radford, Jong Wook Kim, Chris Hallacy, Aditya Ramesh, Gabriel Goh, Sandhini Agarwal, Girish Sastry, Amanda Askell, Pamela Mishkin, Jack Clark, et al. Learning Transferable Visual Models From Natural Language Supervision. In ICML, 2021. 1, 2, 3, 5, 6, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.547, + 0.482, + 0.587 + ], + "angle": 0, + "content": "[40] Olaf Ronneberger, Philipp Fischer, and Thomas Brox. U-net: Convolutional networks for biomedical image segmentation. In MICCAI, 2015. 1, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.59, + 0.482, + 0.658 + ], + "angle": 0, + "content": "[41] Maximilian Seitzer, Max Horn, Andrii Zadaianchuk, Dominik Zietlow, Tianjun Xiao, Carl-Johann Simon-Gabriel, Tong He, Zheng Zhang, Bernhard Scholkopf, Thomas Brox, and Francesco Locatello. Bridging the gap to real-world object-centric learning. In ICLR, 2023. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.661, + 0.482, + 0.702 + ], + "angle": 0, + "content": "[42] Hyun Seok Seong, WonJun Moon, SuBeen Lee, and Jae-Pil Heo. Leveraging hidden positives for unsupervised semantic segmentation. In CVPR, 2023. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.704, + 0.482, + 0.757 + ], + "angle": 0, + "content": "[43] Xiangheng Shan, Dongyue Wu, Guilin Zhu, Yuanjie Shao, Nong Sang, and Changxin Gao. Open-vocabulary semantic segmentation with image embedding balancing. In CVPR, 2024. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.76, + 0.482, + 0.801 + ], + "angle": 0, + "content": "[44] Tong Shao, Zhuotao Tian, Hang Zhao, and Jingyong Su. Explore the potential of clip for training-free open vocabulary semantic segmentation. In ECCV, 2024. 3, 6, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.803, + 0.482, + 0.843 + ], + "angle": 0, + "content": "[45] Gyungin Shin, Weidi Xie, and Samuel Albanie. ReCo: Retrieve and Co-segment for Zero-shot Transfer. In NeurIPS, 2022. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.846, + 0.482, + 0.9 + ], + "angle": 0, + "content": "[46] Oriane Simeoni, Gilles Puy, Huy V Vo, Simon Roburin, Spyros Gidaris, Andrei Bursuc, Patrick Pérez, Renaud Marlet, and Jean Ponce. Localizing objects with self-supervised transformers and no labels. In BMVC, 2021. 3" + }, + { + "type": "list", + "bbox": [ + 0.093, + 0.093, + 0.482, + 0.9 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.093, + 0.905, + 0.134 + ], + "angle": 0, + "content": "[47] Shuyang Sun, Runjia Li, Philip Torr, Xiuye Gu, and Siyang Li. Clip as rn: Segment countless visual concepts without training endeavor. In CVPR, 2024. 2, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.137, + 0.905, + 0.176 + ], + "angle": 0, + "content": "[48] Feng Wang, Jieru Mei, and Alan Yuille. Sclip: Rethinking self-attention for dense vision-language inference. In ECCV, 2024. 1, 2, 3, 6, 7, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.179, + 0.905, + 0.219 + ], + "angle": 0, + "content": "[49] Xudong Wang, Rohit Girdhar, Stella X Yu, and Ishan Misra. Cut and Learn for Unsupervised Object Detection and Instance Segmentation. In CVPR, 2023. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.222, + 0.905, + 0.262 + ], + "angle": 0, + "content": "[50] Xin Wen, Bingchen Zhao, Anlin Zheng, Xiangyu Zhang, and Xiaojuan Qi. Self-supervised visual representation learning with semantic grouping. NeurIPS, 2022. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.264, + 0.905, + 0.332 + ], + "angle": 0, + "content": "[51] Ji-Jia Wu, Andy Chia-Hao Chang, Chieh-Yu Chuang, Chun-Pei Chen, Yu-Lun Liu, Min-Hung Chen, Hou-Ning Hu, Yung-Yu Chuang, and Yen-Yu Lin. Image-text codecomposition for text-supervised semantic segmentation. In CVPR, 2024. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.334, + 0.905, + 0.388 + ], + "angle": 0, + "content": "[52] Monika Wysoczanska, Oriane Simeoni, Michael Ramamonjisoa, Andrei Bursuc, Tomasz Trzciński, and Patrick Pérez. Clip-dinoiser: Teaching clip a few dino tricks. arXiv preprint arXiv:2312.12359, 2023. 6, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.391, + 0.905, + 0.432 + ], + "angle": 0, + "content": "[53] Bin Xie, Jiale Cao, Jin Xie, Fahad Shahbaz Khan, and Yanwei Pang. Sed: A simple encoder-decoder for open-vocabulary semantic segmentation. In CVPR, 2024. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.434, + 0.905, + 0.489 + ], + "angle": 0, + "content": "[54] Jiarui Xu, Shalini De Mello, Sifei Liu, Wonmin Byeon, Thomas Breuel, Jan Kautz, and Xiaolong Wang. GroupViT: Semantic Segmentation Emerges From Text Supervision. In CVPR, 2022. 1, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.491, + 0.905, + 0.544 + ], + "angle": 0, + "content": "[55] Jiarui Xu, Sifei Liu, Arash Vahdat, Wonmin Byeon, Xiaolong Wang, and Shalini De Mello. Open-vocabulary panoptic segmentation with text-to-image diffusion models. In CVPR, 2023. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.547, + 0.905, + 0.588 + ], + "angle": 0, + "content": "[56] Mengde Xu, Zheng Zhang, Fangyun Wei, Han Hu, and Xiang Bai. Side adapter network for open-vocabulary semantic segmentation. In CVPR, 2023. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.59, + 0.905, + 0.644 + ], + "angle": 0, + "content": "[57] Bolei Zhou, Hang Zhao, Xavier Puig, Tete Xiao, Sanja Fidler, Adela Barriuso, and Antonio Torralba. Semantic Understanding of Scenes Through the ADE20K Dataset. IJCV, 2019. 6, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.647, + 0.905, + 0.673 + ], + "angle": 0, + "content": "[58] Chong Zhou, Chen Change Loy, and Bo Dai. Extract Free Dense Labels from CLIP. In ECCV, 2022. 1, 2, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.676, + 0.905, + 0.73 + ], + "angle": 0, + "content": "[59] Lianghui Zhu, Yingyue Li, Jiemin Fang, Yan Liu, Hao Xin, Wenyu Liu, and Xinggang Wang. Weaktr: Exploring plain vision transformer for weakly-supervised semantic segmentation. arXiv preprint arXiv:2304.01184, 2023. 4" + }, + { + "type": "list", + "bbox": [ + 0.517, + 0.093, + 0.905, + 0.73 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.52, + 0.957 + ], + "angle": 0, + "content": "15042" + } + ] +] \ No newline at end of file diff --git a/2025/Distilling Spectral Graph for Object-Context Aware Open-Vocabulary Semantic Segmentation/7ae5aeda-7866-414e-b296-550ab2d843cd_origin.pdf b/2025/Distilling Spectral Graph for Object-Context Aware Open-Vocabulary Semantic Segmentation/7ae5aeda-7866-414e-b296-550ab2d843cd_origin.pdf new file mode 100644 index 0000000000000000000000000000000000000000..9809e38289e7165480be4f9ef6e75d6f9603698e --- /dev/null +++ b/2025/Distilling Spectral Graph for Object-Context Aware Open-Vocabulary Semantic Segmentation/7ae5aeda-7866-414e-b296-550ab2d843cd_origin.pdf @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:572477a0113536ba8cd6b29c66dbb1f9ccd1f7a4e6b88c190b5c43eb1e6d8042 +size 2402724 diff --git a/2025/Distilling Spectral Graph for Object-Context Aware Open-Vocabulary Semantic Segmentation/full.md b/2025/Distilling Spectral Graph for Object-Context Aware Open-Vocabulary Semantic Segmentation/full.md new file mode 100644 index 0000000000000000000000000000000000000000..fa2a71b1459eef9c8e5ff46200ed7756fd86a219 --- /dev/null +++ b/2025/Distilling Spectral Graph for Object-Context Aware Open-Vocabulary Semantic Segmentation/full.md @@ -0,0 +1,298 @@ +# Distilling Spectral Graph for Object-Context Aware Open-Vocabulary Semantic Segmentation + +Chanyoung $\mathrm{Kim}^1$ Dayun Ju $^1$ Woojung Han $^1$ Ming-Hsuan Yang $^{1,2}$ Seong Jae Hwang $^1$ + +$^{1}$ Yonsei University $^{2}$ University of California, Merced + +{chanyoung, juda0707, dnwjDDL, seongjae}@yonsei.ac.kr,mhyang@ucmerced.edu + +# Abstract + +Open-Vocabulary Semantic Segmentation (OVSS) has advanced with recent vision-language models (VLMs), enabling segmentation beyond predefined categories through various learning schemes. Notably, training-free methods offer scalable, easily deployable solutions for handling unseen data, a key goal of OVSS. Yet, a critical issue persists: lack of object-level context consideration when segmenting complex objects in the challenging environment of OVSS based on arbitrary query prompts. This oversight limits models' ability to group semantically consistent elements within object and map them precisely to user-defined arbitrary classes. In this work, we introduce a novel approach that overcomes this limitation by incorporating object-level contextual knowledge within images. Specifically, our model enhances intra-object consistency by distilling spectral-driven features from vision foundation models into the attention mechanism of the visual encoder, enabling semantically coherent components to form a single object mask. Additionally, we refine the text embeddings with zero-shot object presence likelihood to ensure accurate alignment with the specific objects represented in the images. By leveraging object-level contextual knowledge, our proposed approach achieves state-of-the-art performance with strong generalizability across diverse datasets. + +# 1. Introduction + +Open-vocabulary semantic segmentation (OVSS) [9, 32, 54, 55, 58] aims to predict pixel-level labels for arbitrary prompts defined by user input (e.g., the proper noun Space Needle in Fig. 1(a)), advancing beyond previous models scoped to predefined static classes [10, 18, 19, 40]. To achieve this, fully supervised OVSS models [11, 53, 56] enhance generalization to seen classes using labeled training data. Yet, their reliance on seen classes during training may + +![](images/f14b060ee2ae67aafe5e2f19fa931c75774a2ad9eef48c4745ae849936ee7544.jpg) +(b) Visualization +Truck +Horse +Cow +Sheep + +![](images/7ba854727b4ad18975b4615743c47a45a7bd4c97aea34938889ec67f8e7f96bc.jpg) +Figure 1. We present CASS, object-level Context-Aware training-free open-vocabulary Semantic Segmentation model. (a) Overview: Our method distills the vision foundation model's (VFM) object-level contextual spectral graph into CLIP's attention and refines query text embeddings towards object-specific semantics. (b) Object-Level Context: This result illustrates how incorporating object-level context improves segmentation accuracy by unifying object-wise components into a user-defined object class; for example, (top row) our approach precisely segments the truck's body, wheels, and cargo area and (bottom row) accurately groups elements into a single object, such as the sheep, whereas baselines [16, 29] often fail to achieve this unified segmentation. + +risk overfitting and limit scalability, as labeled data is required for retraining to adapt to new domains [43]. + +To better address this, recent studies [16, 28, 29, 48, 58] propose training-free OVSS methods that leverage pretrained vision-language models (VLMs) (e.g., CLIP [39]) to + +integrate aligned visual and textual representations. These works focus on improving semantic association between visual patches in CLIP image encoder to produce more precise segmentation maps, achieving highly promising results. This training-free scheme offers key advantages: it (1) generalizes well to unseen classes without requiring additional labeled data, making it more scalable, thus (2) allows direct adaptation to dynamic real-world applications. Capitalizing on these benefits, training-free OVSS aligns with OVSS's core goal of adaptability across diverse domains, enabling it to reach its full potential. + +We tackle OVSS in a training-free manner, presenting a challenging and practical approach aligned with OVSS's true objectives. Although existing training-free methods [21, 29, 47, 48] perform well on numerous tasks, they all share a common limitation: the lack of "object-level context." To understand "object-level context", consider a truck and sheep in Fig. 1(b). Components comprising a truck, such as its wheels and cargo area, should be grouped as a single entity. Similarly, components of a sheep should be accurately unified under the correct user-given object class. However, existing training-free OVSS methods, which use CLIP as the image backbone for segmentation, often struggle to capture object-level context, failing to group object components into a single, user-defined object. These challenges stem from CLIP's focus on learning global image semantics, which may be insufficient for capturing dense object-level semantics [16, 48, 58]. Thus, to achieve accurate training-free OVSS, a proper object-level context must be considered in CLIP. + +We address the challenges of training-free OVSS by introducing the object-level context and propose the Context-Aware Semantic Segmentation (CASS) model. Vision foundation models (VFMs), such as DINO [8, 38], capture fine-grained, patch-level semantics but lack object-level context, making them unsuitable for direct use in tasks requiring such context. Recent studies [24, 36] have shown that applying graph spectral techniques to VFM attention graphs can transform patch-level representations into object-level representations. Accordingly, we apply spectral techniques to extract essential object-level context from the VFM attention graph. The extracted spectral features can be used directly for segmentation (e.g., clustering eigenvectors); however, this approach may not be suitable for OVSS, where textual alignment is essential. Thus, we propose distilling low-rank components of VFM into CLIP to enhance its object-level context understanding, while maintaining alignment with textual information. Specifically, we decompose the VFM's attention graph to extract low-rank components, filtering irrelevant information while emphasizing essential object-level context. These low-rank components are then distilled into CLIP's attention to embed the fundamental object structure of the VFM graph. + +Our model also leverages CLIP's highly effective zero-shot object classification capability (i.e., object presence prior), widely validated in prior work [20, 34, 39], to capture detailed object-level context within scenes. As such, we adjust the text embeddings based on the object presence prior encoded by CLIP. This process involves refining the embeddings to better align with object-specific semantics. Then, we refine patch-text similarities using the object presence prior, ensuring that the final segmentation map reflects an object-specific perspective. Our CASS, carefully designed to enhance object-level contextual knowledge, captures richer intra-object coherence and enables cohesive groupings of semantically related elements. + +The main contributions of this work are: + +- We enhance object-level context by distilling rich semantic understanding from VFM into the CLIP attention mechanism, facilitating a more consistent grouping of each object into a unified semantic entity. +- We present object presence prior-driven text embeddings and patch-text similarity refinement, promoting the model to precisely classify objects. +- We demonstrate the performance of our model in various semantic segmentation datasets and achieve state-of-the-art performance in training-free OVSS. + +# 2. Related Works + +Open-Vocabulary Semantic Segmentation. Open-vocabulary semantic segmentation (OVSS) aims to segment images corresponding to arbitrary user-given prompts, using vision-language aligned representations (e.g., CLIP [39]). Meanwhile, training-free approaches [3, 4, 23, 29, 58] have risen for their ability to adapt directly to new tasks without extra training. Among these, methods [16, 29, 48, 58] that do not rely on additional datasets have gained attention, as they require no extra adaptation effort, aligning better with the objective of OVSS. Earlier approaches investigated leveraging CLIP features for better localization [30, 58] or for grounding proposed masks [21, 47]. SCLIP [48] uses correlative self-attention to improve inter-patch correlations. NACLIP [16] applies a Gaussian kernel to attention maps for better spatial consistency, and ProxyCLIP [29] integrates patch-wise consistent features from DINO [8]. Although these methods demonstrated strong results, they often struggle to group different parts of the same object due to a lack of explicit dense label knowledge. To this end, we attribute the underlying challenges of training-free OVSS by enriching the object-level contextual understanding. + +Vision Foundation Models. Vision Foundation Models (VFMs), especially vision transformer (ViT)-based [14] self-supervised models (e.g., DINO [8, 38]), have significantly advanced computer vision through their ability to capture patch-wise semantic understanding at both the + +![](images/e6308b01ef7816f3a5505a7104fb111a555c41f315081604abbbb7ad4ffc6240.jpg) +Figure 2. Detailed illustration of our proposed training-free spectral object-level context distillation mechanism in Sec. 3.2. By matching the attention graphs of VFM and CLIP head-by-head to establish complementary relationships, and distilling the fundamental object-level context of the VFM graph to CLIP, we enhance CLIP's ability to capture intra-object contextual coherence. + +feature and attention levels. Among various downstream tasks [17, 35, 42, 46, 49] leveraging VFM's capability, one line of research focuses on leveraging the attention components in VFM [24, 36, 46], which exhibit a detailed understanding of patch-wise semantic interaction between different image regions. Thus, while ProxyCLIP [29] utilizes VFM features, our method takes a different effort by treating the attention mechanism in VFM as a graph. We then extract core structural patterns of the VFM graph to emphasize object-level context, providing an effective solution for producing precise object masks in training-free OVSS. + +Object-Level Semantic Segmentation. Semantic segmentation groups object components that have similar semantics into a single mask, traditionally achieved through models trained on extensive human-labeled, pixel-wise annotations [10, 19, 25, 40]. However, despite recent advancements in unsupervised semantic segmentation [17, 42] (USS) that do not rely on dense labels, a core challenge persists: individual object components are not consistently grouped into a single entity. To address this issue, recent USS studies [24, 41, 50] have incorporated object-level representation learning to improve the model's ability to merge single objects into unified entities. However, these efforts have not been considered in the training-free OVSS task, which faces similar challenges in an unsupervised context. Therefore, we focus on enriching object-level context in training-free OVSS to achieve a more effective grouping of components within each object, enabling clearer and more consistent segmentation across diverse visual domains. + +# 3. Method + +We detail our proposed model, CASS, implemented in a training-free manner, as outlined in Fig 1(a). Our model jointly enhances object-level contextual understanding with two primary aspects: (1) Spectral Object-Level + +Context Distillation (Sec. 3.2) and (2) Object Presence-Driven Object-Level Context (Sec. 3.3). We first outline the basic pipeline of training-free OVSS in Sec. 3.1. + +# 3.1. Preliminaries + +CLIP Visual Encoder. Training-free OVSS models [16, 29, 48] leverage CLIP [39] for its rich alignment of visual and linguistic information. However, the CLIP ViT visual encoder $\mathcal{F}_{\mathrm{CLIP}}^v$ focuses on aligning the [CLS] token with text embeddings during training, often neglecting the patchwise spatial interactions among visual tokens that are crucial for dense prediction [16, 44, 48]. To produce CLIP feature representations more suitable for semantic segmentation tasks, we remove residual connection, feed-forward network, and self-attention layer in the final layer of $\mathcal{F}_{\mathrm{CLIP}}^v$ , following recent works [16, 48]. Specifically, when the set of visual tokens $Z = \mathrm{Concat}[z_1, \ldots, z_N] \in \mathbb{R}^{N \times D_Z}$ , where $N$ is the total number of patches except [CLS] token and $D_Z$ denotes patch dimension space, is input into the final block of the ViT, the following operations are as + +$$ +Z ^ {*} = \operatorname {S A} (\operatorname {L N} (Z)), \tag {1} +$$ + +where SA and LN represent Self-Attention and Layer Normalization, respectively. The modified SA layer is + +$$ +M _ {\mathrm {C L I P}} ^ {i} = \operatorname {s o f t m a x} \left(A _ {\mathrm {C L I P}} ^ {i} / \sqrt {D _ {h}}\right), \tag {2} +$$ + +$$ +Z ^ {*} = \operatorname {C o n c a t} \left[ M _ {\mathrm {C L I P}} ^ {1} V _ {\mathrm {C L I P}} ^ {1}, \dots , M _ {\mathrm {C L I P}} ^ {h} V _ {\mathrm {C L I P}} ^ {h} \right] W ^ {O}, \tag {3} +$$ + +where $D_h$ denotes the head dimension, $h$ is the number of heads, and $W^0 \in \mathbb{R}^{D_z \times D_z}$ is the output projection matrix. We define CLIP attention map for $i$ -th head as $A_{\mathrm{CLIP}}^i = K_{\mathrm{CLIP}}^i K_{\mathrm{CLIP}}^{i^\top} \in \mathbb{R}^{N \times N}$ , where attention key $K_{\mathrm{CLIP}} \in \mathbb{R}^{N \times D_h \times h}$ and value $V_{\mathrm{CLIP}} \in \mathbb{R}^{N \times D_h \times h}$ result from the linear transformation of $\mathrm{LN}(Z)$ . + +Inference Pipeline. We use a sliding window inference to ensure that ViT can capture detailed parts of the image by + +![](images/29ebad50915d6e3e59ee82cf81eb349c4fd862f5c127cd6d0c6d0a7d45dc4db1.jpg) +Figure 3. Attention score visualization for various query points. Left: Vanilla CLIP ( $A_{\mathrm{CLIP}}$ ) shows noisy, unfocused attention. Center: VFM-to-CLIP distillation without low-rank eigenscaling shows partial object grouping with limited detail. Right: Incorporating our low-rank eigenscaling (Eq. (7)) captures object-level context, improving grouping within a single object. + +dividing the entire image $\mathcal{I}$ into smaller windows $\hat{\mathcal{I}}$ . The visual tokens $Z^{*} \in \mathbb{R}^{N \times D_{Z}}$ are then computed from each $\hat{\mathcal{I}}$ through $\mathcal{F}_{\mathrm{CLIP}}^{v}$ as in Eq. (1). To perform OVSS with given texts, we project the visual tokens $Z^{*}$ into the vision-language latent space as $F_{\mathrm{CLIP}} = \mathcal{P}_{\mathrm{CLIP}}(Z^{*}) \in \mathbb{R}^{N \times d}$ , where $\mathcal{P}_{\mathrm{CLIP}}$ is a projection matrix into CLIP latent space and $d$ is the CLIP latent dimension. We compute patch-text similarity $\hat{S} \in \mathbb{R}^{N \times C}$ for each image window $\hat{\mathcal{I}}$ as + +$$ +\hat {\mathcal {S}} = F _ {\mathrm {C L I P}} \left[ \left\{t _ {\mathrm {C L I P}} ^ {i} \right\} _ {i = 1} ^ {C} \right] ^ {\top}, \tag {4} +$$ + +where $\{t_{\mathrm{CLIP}}^i\}_{i = 1}^C \coloneqq \mathrm{Concat}\left[t_{\mathrm{CLIP}}^1,\dots ,t_{\mathrm{CLIP}}^C\right] \in \mathbb{R}^{C\times d}$ denotes the encoded text embeddings, and $C$ is the number of classes. Finally, the results of all windows are combined to create a segmentation map for the entire image. + +# 3.2. Spectral Object-Level Context Distillation + +Our model CASS aims to capture such object-level contextual interaction within the image features to associate different parts of the same object. However, the limited ability of CLIP to capture semantic relationships between patches often neglects the object-level context, as illustrated in the left column of Fig. 3 ( $A_{\mathrm{CLIP}}$ ). To address this limitation, we leverage VFM (e.g., DINO [8]), enabling a deeper object-level contextual understanding. Among VFM features, we utilize VFM attention key component $K_{\mathrm{VFM}} \in \mathbb{R}^{h \times N \times D_h}$ . Specifically, we treat the VFM attention adjacency $A_{\mathrm{VFM}} = K_{\mathrm{VFM}} K_{\mathrm{VFM}}^{\top} \in \mathbb{R}^{h \times N \times N}$ as a graph, where $h$ represents the number of heads, and transfer crucial graph patterns from $A_{\mathrm{VFM}}$ to $A_{\mathrm{CLIP}}$ . However, a key question remains: "How do we account for the multi-head attention mechanisms in both + +VFM and CLIP?" In other words, "how do we match $i$ -th VFM attention head $A_{\mathrm{VFM}}^i$ with $j$ -th CLIP attention head $A_{\mathrm{CLIP}}^j$ ? A straightforward approach is to distill the attention heads sequentially (i.e., $i = j$ ). Nevertheless, recent studies [22, 31, 59] indicate that different attention heads focus on separate parts of the image. Thus, matching the optimal heads between VFM and CLIP is essential for precise attention distillation. + +# 3.2.1. Complementary Spectral Graph Matching + +We match graphs with contrasting structures, as shown in Fig 4, enabling VFM to supplement the object-level contextual knowledge that CLIP alone struggles to capture. To match an optimal head-wise attention graph, we exploit the spectral field of both $A_{\mathrm{VFM}}$ and $A_{\mathrm{CLIP}}$ . Our graph-matching mechanism has two main steps as illustrated in Fig. 2(a): (I) examining the eigenvalues of each graph and (II) matching optimal graph pairs using the spectral distribution. + +(I) Eigenvalue Examination. For each graph from VFM and CLIP, denoted as $A_{\mathrm{VFM}}^i$ and $A_{\mathrm{CLIP}}^i$ for $(i = 1, \dots, h)$ , we perform eigendecomposition. This yields the eigenvalues for each head $h$ , $\lambda_{\mathrm{VFM}}^i$ and $\lambda_{\mathrm{CLIP}}^i$ , from which the top $m$ fixed eigenvalues are selected. These eigenvalues contain each graph's unique structural features, key properties essential for distinguishing different attention graph patterns. +(II) Graph Matching via Spectral Distribution. After obtaining eigenvalues for each head, we compute spectral distances to quantify structural differences, creating a cost matrix $\mathcal{C} \in \mathbb{R}^{h \times h}$ for each graph pair from VFM and CLIP: + +$$ +\mathcal {C} _ {i j} = 1 - \mathcal {D} _ {W} \left(\bar {\lambda} _ {\mathrm {V F M}} ^ {i}, \bar {\lambda} _ {\mathrm {C L I P}} ^ {j}\right), \quad \forall i, j \in \{1, \dots , h \} \tag {5} +$$ + +where $\bar{\lambda}_{\mathrm{VFM}}^i$ and $\bar{\lambda}_{\mathrm{CLIP}}^j$ represent the normalized eigenvalues of the $i$ -th head of VFM and the $j$ -th head of CLIP, respectively, and $\mathcal{D}_W$ denotes the Wasserstein distance. Here, the Wasserstein distance is computed as $\mathcal{D}_W(u,v) = \sum_{i=1}^{h} |\operatorname{sort}(u)_i - \operatorname{sort}(v)_i|$ , where $u$ and $v$ are the two input distributions, and the sort function orders the values of $u$ and $v$ in ascending order. After computing $\mathcal{C}$ , we apply the Hungarian matching algorithm [27] to the cost matrix to find the optimal pairing of graph heads. This approach pairs graphs with contrasting characteristics, enabling complementary distillation of object-level context knowledge from $A_{\mathrm{VFM}}$ to $A_{\mathrm{CLIP}}$ , as described in the following section. + +# 3.2.2. Distilling VFM Spectral Graph to CLIP + +Once we match optimal graph pairs, we distill the VFM graph to CLIP (see Fig. 2(b)). For the sake of explanation, we assume that the $i$ -th VFM attention head $A_{\mathrm{VFM}}^i$ is matched with the $j$ -th CLIP attention head $A_{\mathrm{CLIP}}^j$ . An intuitive approach would be simply aggregating $A_{\mathrm{VFM}}^i$ and $A_{\mathrm{CLIP}}^j$ without any transformation. However, as shown in Fig. 3 ( $A_{\psi w/o \mathrm{LR}}$ ), this approach may transfer noise or irrelevant information, highlighting the need to extract features + +![](images/be1e7e611f01088054fa7f9b0c142b13a3ad234b1a6a6d43920d50f02483aed8.jpg) +Figure 4. Visualization of the matched complementary graph pairs. The red dotted line indicates the focus areas of VFM, while the yellow dotted line highlights those for CLIP. By aggregating graphs with contrasting structural properties, we enhance the object-level context in $A_{\mathrm{CLIP}}$ supported by $\ddot{A}_{\mathrm{VFM}}$ , resulting $A_{\psi}$ . + +that emphasize object-level context. From this realization, we leverage the low-rank components of VFM, which contain distinct object patterns within the graph structure [12]. Specifically, we (I) extract the critical object-level contextual structure of $A_{\mathrm{VFM}}^i$ via low-rank approximation and enhance the graph structure by dynamically scaling eigenvalues. Then, we (II) distill the approximated VFM graph into CLIP by aggregating graphs based on spectral distance. + +(I) Low-Rank Dynamic Eigenscaling. Our goal is to transfer the essential object-level contextual structure of VFM graph $A_{\mathrm{VFM}}$ to CLIP. To achieve this, we extract low-rank components of the VFM graph using standard eigendecomposition, as $A_{\mathrm{VFM}}^i = U\Sigma U^\top$ , where $U$ and $\Sigma$ represent the eigenvector and the diagonal eigenvalue matrix, respectively. In the decomposed eigenbasis, we identify key object-level features of each graph by searching an optimal number of eigenvalues $k$ through an energy-based approach. This ensures that the chosen $k$ eigenvalues capture a significant portion of the graph's energy, retaining essential structural information while discarding noise and less relevant details. More details on searching $k$ can be found in the supplementary material. This process yields a low-rank approximated VFM graph, expressed as $\tilde{A}_{\mathrm{VFM}}^i = U_k\Sigma_kU_k^\top$ . + +We refine the low-rank components with a scaling function $\phi$ , which dynamically amplifies larger eigenvalues and reduces smaller ones. Compared to the conventional shrinkage function [2, 5], which only focuses on noise cutoff, our approach emphasizes essential structural information, particularly object-level context features, while suppressing noise and irrelevant details. More details on dynamic eigenscaling can be found in the supplementary material. The graph after applying dynamic eigenscaling is as + +$$ +\ddot {A} _ {\mathrm {V F M}} ^ {i} = U _ {k} \phi (\Sigma_ {k}) U _ {k} ^ {\top}. \tag {6} +$$ + +(II) VFM Graph Distillation. We distill the essential object-level contextual knowledge from the tailored VFM graph $\ddot{A}_{\mathrm{VFM}}^i$ into the $A_{\mathrm{CLIP}}^j$ by aggregating their structures: + +![](images/8f30b67fc69581bdba2fef41e9589b525bac9ea247f49c6329795b8673f9185e.jpg) +Figure 5. Detailed illustration of our object presence prior-guided text embedding adjustment module. The CLIP text encoder generates text embeddings for each object class, and the object presence prior is derived from both visual and text embeddings. Within hierarchically defined class groups, text embeddings are selected based on object presence prior, then refined in an object-specific direction to align with components likely present in the image. + +$$ +A _ {\psi} ^ {j} = \left(w _ {i j} \ddot {A} _ {\mathrm {V F M}} ^ {i} + A _ {\mathrm {C L I P}} ^ {j}\right) / \left(w _ {i j} + 1\right), \tag {7} +$$ + +where $w_{ij}$ is defined as the Wasserstein distance between eigenvalues from both graphs as $\mathcal{D}_{\mathcal{W}}(\bar{\lambda}_{\mathrm{VFM}}^i,\bar{\lambda}_{\mathrm{CLIP}}^j)$ . This adaptively assigns higher weights to graphs with complementary structures, while reducing influence on those with similar structures. Our distillation strategy effectively transfers the critical object-level context from VFM to CLIP, ensuring multiple components of the same object are grouped into a unified semantic as illustrated in Fig. 3 ( $A_{\psi}$ ). Therefore, we use $A_{\psi}$ as our attention matrix instead of $A_{\mathrm{CLIP}}$ in Eq. (2) to compute the final visual feature $F_{\mathrm{CLIP}}$ . + +# 3.3. Object Presence-Driven Object-Level Context + +The spectral object-level context distillation in Sec. 3.2 enables precise object mask generation. However, due to the nature of OVSS, where users can provide arbitrary query prompts, different parts of the same object may still be assigned to closely related categories. Meanwhile, numerous studies [20, 34, 39] have demonstrated that CLIP excels in zero-shot object classification capability. Accordingly, we utilize the zero-shot object classification score encoded by CLIP (i.e., the object presence prior) to refine text embeddings and patch-text similarity, enhancing object-centered perspective. To compute the object presence prior $P$ , we take an entire input image $\mathcal{I}$ and compute likelihood for each class $i$ as $P(i) = \{t_{\mathrm{CLIP}}^i\}_{i=1}^C v_{\mathrm{CLIP}}$ , where the visual embedding vector $v_{\mathrm{CLIP}} = \mathcal{F}_{\mathrm{CLIP}}^v(\mathcal{I}) \in \mathbb{R}^d$ is derived from the ViT [CLS] token. The object presence prior is used in the following sections to enhance object-level context. + +Object-Guided Text Embedding Adjustment. We adjust text embeddings $\{t_{\mathrm{CLIP}}^i\}_{i = 1}^C$ to embed object-specific contextual semantics, as shown in Fig. 5. To prevent object parts from being assigned to ambiguous classes, we + +Table 1. Quantitative results with state-of-the-art unsupervised open-vocabulary semantic segmentation models on eight datasets. + +
ModelSupporting DatasetExtra-TrainingFairV21PC60C-ObjV20PC59C-StfCityADEAvg.
GroupViT [54]CVPR'22CC12M+RedCaps✓✘50.418.727.579.723.415.311.19.229.4
TCL [9]CVPR'23CC3M+CC12M✓✘55.030.431.683.233.922.424.017.137.2
CoDe [51]CVPR'24CC3M+RedCaps✓✘57.530.532.3--23.928.917.7-
CLIP-DINOiser [52]ECCV'24ImageNet1k✓✘62.132.434.880.935.924.631.720.040.3
ReCo [45]NeurIPS'22ImageNet1k✘✘25.119.915.757.722.314.821.611.223.5
FOSSIL [3]WACV'24COCO Captions✘✘----35.824.823.218.8-
FreeDa [4]CVPR'24COCO Captions✘✘---85.643.127.836.722.4-
CLIP [39]ICML'21✘✘✓18.67.86.549.111.27.26.73.213.8
MaskCLIP [58]ECCV'22✘✘✓38.323.620.674.926.416.412.69.827.9
GEM [6]CVPR'24✘✘✓46.2---32.6--15.7-
CaR [47]CVPR'24✘✘✓48.613.615.473.718.4--5.4-
PnP-OVSS [33]CVPR'24✘✘✓--36.251.328.017.9-14.2-
CLIPtrase [44]ECCV'24✘✘✓50.929.943.681.033.822.821.316.432.7
ClearCLIP [28]ECCV'24✘✘✓51.832.633.080.935.923.930.016.738.1
SCLIP [48]ECCV'24✘✘✓59.130.430.580.434.122.432.216.138.2
LaVG [21]ECCV'24✘✘✓62.131.634.282.534.723.226.215.838.8
ProxyCLIP [29]ECCV'24✘✘✓59.135.236.278.238.826.238.119.641.4
NACLIP [16]WACV'25✘✘✓58.932.233.279.735.223.335.517.439.4
CASS✘✘✓65.836.737.887.840.226.739.420.444.4
+ +The hierarchical clustering of CLIP semantic distance, forming groups of similar embeddings $\mathcal{T}_g = \{t_{\mathrm{CLIP}}^i\mid i\in I_g\}$ , where $\mathcal{T}_g$ is a subset of entire text embeddings $\{t_{\mathrm{CLIP}}^i\}_{i = 1}^C$ with index set $I_{g}$ . Further details about this clustering process are provided in the supplementary materials. Within $\mathcal{T}_g$ , we identify the object most likely to appear in the image using object presence prior $P$ , formulated as $i^{*} = \arg \max_{i\in I_{g}}P(i)$ . This facilitates refining the text embedding vector $t_{\mathrm{CLIP}}^*$ to better represent object-contextual semantics. We note that patch-wise visual feature $F_{\mathrm{CLIP}}\in \mathbb{R}^{N\times d}$ contains object-specific semantic vectors that can guide $t_{\mathrm{CLIP}}^*$ accordingly. Thus, we compute cosine similarity between $F_{\mathrm{CLIP}}$ and $t_{\mathrm{CLIP}}^*$ and select top- $n$ object-specific image vector denoted as $\{f_i\}_{i = 1}^n\coloneqq$ $\mathrm{Concat}[f_1,\ldots ,f_n]\in \mathbb{R}^{n\times d}$ . To define a targeted direction, we compute the average of $\{f_i\}_{i = 1}^n$ as $\mu_{n}\in \mathbb{R}^{d}$ , and guide selected text embedding $t_{\mathrm{CLIP}}^*$ towards $\mu_{n}$ as + +$$ +\tilde {t} _ {\mathrm {C L I P}} ^ {i ^ {*}} = (1 - \alpha) \cdot t _ {\mathrm {C L I P}} ^ {i ^ {*}} + \alpha \cdot \mu_ {n}, \tag {8} +$$ + +where $\alpha$ controls the degree of alignment with $\mu_{n}$ . The adjusted text embedding $\tilde{t}_{\mathrm{CLIP}}^{i^*}$ replaces the original text embedding $t_{\mathrm{CLIP}}^{i}$ in the set $\{t_{\mathrm{CLIP}}^{i}\}_{i=1}^{C}$ , which is then used to compute the patch-text similarity logit $\hat{S}$ in Eq. (4). + +Object Perspective Patch-Text Similarity. We integrate the object presence prior with patch-text similarity $\hat{S}$ from Eq. (4) to produce a segmentation map from an object-perspective view. Unlike Barsellotti et al. [4], which computes likelihood from an image window $\hat{\mathcal{I}}$ , our approach derives similarity from the full image $\mathcal{I}$ , enabling richer object-contextual understanding at the broader level. Thus, we refine the patch-text similarity $\hat{S}$ from Eq. (4) to obtain the final patch-text similarity $\hat{S}^*$ , computed as + +$$ +\hat {\mathcal {S}} ^ {*} = (1 - \gamma) \cdot \hat {\mathcal {S}} + \gamma \cdot \left\{t _ {\mathrm {C L I P}} ^ {i} \right\} _ {i = 1} ^ {C} v _ {\mathrm {C L I P}}, \tag {9} +$$ + +where $\gamma$ controls the balance between $\hat{S}$ and object presence prior. This allows the final segmentation map to capture object-level details for precise, class-aware segmentation. + +# 4. Experiments + +We first describe the experimental settings, including implementation details, datasets, and evaluation metrics (Sec. 4.1). We then present quantitative and qualitative evaluation results (Sec. 4.2) and demonstrate ablation studies to assess each proposed component (Sec. 4.3). Additional experiment results are provided in the supplementary material. + +# 4.1. Experimental Settings + +Implementation Details. Similar to existing training-free approaches [16, 21, 29, 44, 48], we employ CLIP ViT-B/16 [39] as our VLM and DINO ViT-B/8 [8] as our VFM, keeping both frozen. Input images are matched with the shorter side set to 336 pixels (or 560 pixels for the higher-resolution Cityscapes dataset), and sliding window inference is applied with a $224 \times 224$ window and a stride of 112 to achieve efficient evaluation coverage. All experiments are conducted using NVIDIA RTX A6000 Ada*. + +Datasets. Our primary datasets include PASCAL VOC 2012 [15], PASCAL Context [37], and COCO [7], each evaluated with and without a background class. We refer to the versions with a background class as V21, PC60, and C-Obj, and those without as V20, PC59, and C-Stf. We also report results on the ADE20K (ADE) [57] and Cityscapes (City) [13] datasets. These datasets contain 21, 60, 81, 20, 59, 171, 150, and 19 classes, respectively. + +![](images/6acc002ec693f900363bf04759e497b548ae11fa32e1c9b1b3103d66c2b3f462.jpg) +Figure 6. Qualitative comparison across the Pascal VOC [15], Pascal Context [37], COCO [7], and ADE20K [57] datasets using CLIP ViT-B/16 [39], evaluating previous state-of-the-art training-free OVSS methods, including NACLIP [16], ProxyCLIP [29], and ours. + +# 4.2. Evaluation Results + +We evaluate our model against training-free OVSS methods both quantitatively and qualitatively. For fair comparison, we refrain from using the mask refinement (e.g., PAMR [1] or DenseCRF [26]) and reproduce results from [29] using a $224 \times 224$ window to match the baselines. Performance is measured primarily with mean IoU (mIoU), complemented by pixel accuracy (pAcc) for a broader assessment. + +Quantitative Evaluation: mIoU. Table 1 presents the main results in mIoU, where all results are obtained using CLIP ViT-B/16 as the backbone model for fair comparisons. To demonstrate the strengths of our model, we also include methods that use supporting datasets (i.e., retrieving additional datasets during inference) or additional training (models evaluated under the same setting as ours are indicated with $\checkmark$ in the "Fair" column). Our model, CASS, performs favorably against the state-of-the-art methods, with an average gain of $3.0\mathrm{mIoU}$ point across eight datasets. Notably, for the V20, CASS achieves a performance gain of 5.3 mIoU point over the second-best model. Further, CASS outperforms CLIP-DINOiser [52] (which also leverages DINO with extra data and training) by $4.1\mathrm{mIoU}$ point. + +Quantitative Evaluation: pAcc. Table 2 shows the results of pAcc on recent state-of-the-art models evaluated under the same settings. Our CASS outperforms the second-highest model by $2.2\mathrm{pAcc}$ point and exceeds one of the most recent papers, CLIPtrase [44], by $9.8\mathrm{pAcc}$ point. Our model consistently performs better than existing methods on mIoU and pAcc by effectively capturing object-level context, resulting in more accurate object masking. + +Qualitative Analysis. Fig. 6 provides a qualitative comparison with recent state-of-the-art models [16, 29]. Existing methods often produce noisy segmentation maps and do not group object components correctly, such as wheels within a car or a person's head or arm. Additionally, they often mis + +Table 2. Quantitative results using average pixel accuracy. + +
ModelV21PC60C-ObjV20PC59C-StfCityADEAvg.
CLIPtrase [44]78.652.150.189.758.938.963.438.659.1
LaVG [21]89.348.774.891.158.939.168.537.063.4
SCLIP [48]87.649.274.391.058.338.472.738.763.8
ProxyCLIP [29]86.652.075.988.463.443.474.949.166.7
NACLIP [16]87.151.275.389.259.839.371.745.264.9
CASS90.155.676.093.965.043.678.248.668.9
+ +classify object parts into ambiguous classes (e.g., a motorcycle as a bicycle or an airplane as a car). In contrast, based on a deep understanding of object-level context, our CASS generates clean segmentation maps that accurately group all object components, including slender parts such as human arms or animal legs, into their respective object classes. + +# 4.3. Ablation Study + +We perform ablation studies to analyze the proposed method. We use the datasets representative of segmentation tasks for all experiments: V21, PC59, C-Stf, following [48]. Please refer to the supplementary material for the additional ablation studies including VFM backbones and feature type. + +Effect of Spectral VFM Distillation. Table 3 shows the results of incorporating VFM (Exp.#3) with the vanilla setting (Exp.#1), which directly uses the output from Eq. (4). In Exp.#3, $A_{\mathrm{VFM}}$ and $A_{\mathrm{CLIP}}$ are aggregated in a straightforward head-wise manner, without any additional processing to $A_{\mathrm{VFM}}$ . Simply leveraging $A_{\mathrm{VFM}}$ leads to a performance gain compared to the vanilla setting. Exp.#4 to Exp.#6 validate the effectiveness of the proposed components (see Sec. 3.2). Exp.#4 demonstrates the aggregation of $A_{\mathrm{VFM}}$ and $A_{\mathrm{CLIP}}$ using a complementary spectral graph matching (GM) procedure. Exp.#5, denoted as LR, presents the results with only the low-rank approximation applied to $A_{\mathrm{VFM}}$ , while Exp.#6 shows the results with both the low-rank approximation and Dynamic Eigenscaling (DE) applied, as defined in Eq. (6). Notably, Exp.#6 shows signifi + +Table 3. Ablation results showing the impact of our components on the performance across V21 [15], PC59 [37], and C-Stf [7]. The abbreviations in the table are noted in Sec. 4.3. + +
Exp. #AVFMGMLRDEOPSOTADataset
V21PC59C-Stf
158.935.323.4
2✓✓59.236.524.0
3✓62.937.725.2
4✓✓63.238.025.4
5✓✓✓64.538.425.8
6✓✓✓✓64.838.625.9
7✓✓✓✓✓65.039.426.6
8✓✓✓✓✓✓65.840.226.7
+ +![](images/89aa33d1d8b0c648e3138465c94d23ae99fe508f53f1e024a70cabf0b63829df.jpg) +Figure 7. Qualitative ablation of our method. Exp.# refers to the experiment in Table 3. As additional components of our proposed method are incorporated, the segmentation map exhibits greater refinement and more effectively captures object-level context. + +cant performance gain over simply integrating VFM knowledge (Exp.#3), demonstrating the effectiveness of graph matching and low-rank approximation in capturing object-level context. Fig. 3 qualitatively visualizes the attention map, with the left, middle, and right columns corresponding to Exp.#3, Exp.#4, and Exp.#6, respectively. We observe that integrating proposed low-rank dynamic eigenscaling effectively reduces noise and emphasizes objects by highlighting objects with strong inter-object associations. + +Effect of Object Presence Prior. We validate the effectiveness of the object presence prior-driven methods in Exp.#7 and Exp.#8 by examining each component: Object-Perspective Patch-Text Similarity (OPS) and Object-Guided Text Embedding Adjustment (OTA) (see Sec. 3.3). These results show that each component effectively improves performance, emphasizing object-level context. Additionally, incorporating the object presence prior-driven methods on top of the vanilla setting, as shown in Exp.#2, clarifies the effectiveness of the proposed method. + +Qualitative Ablations. Fig. 7 shows the segmentation maps for the vanilla setting (Exp.#1), after applying spectral VFM distillation (Exp.#6), and with all proposed components applied (Exp.#8). While Exp.#6 improves object masking compared to Exp.#1, it often misses parts of objects (highlighted by the red dotted line). In contrast, incorporating all proposed components (Exp.#8) consolidates objects into single entities, effectively capturing object-level context in the segmentation mask. + +Different CLIP Backbones. Table 4 shows results on dif- + +Table 4. Ablation results on different CLIP backbones. + +
ModelV21PC59C-StfAvg. mIoUAvg. pAcc
mIoUpAccmIoUpAccmIoUpAcc
VLT-B/32SCLIP [48]50.681.028.749.620.033.433.154.7
LaVG [21]54.884.429.050.120.533.734.856.1
ProxyCLIP [29]57.985.635.259.723.640.238.961.8
NACLIP [16]51.182.832.456.821.237.134.958.9
CASS58.286.236.561.024.440.839.762.7
VLT-L/14SCLIP [48]44.478.125.246.617.629.629.151.4
LaVG [21]51.584.127.550.319.432.332.855.6
ProxyCLIP [29]59.885.738.361.026.243.541.463.4
NACLIP [16]52.283.132.152.821.435.735.257.2
CASS62.188.039.162.126.343.442.564.5
+ +![](images/adb871732c94a18747dc50fc01192bda794e1167ea88792b2145623dbc03eee2.jpg) +Figure 8. Performance evaluation of different distance metrics used in graph matching and graph distillation. + +ferent CLIP backbones, ViT-B/32 and ViT-L/14. Our CASS performs favorably across various backbones, consistently outperforming existing methods in mIoU and pAcc. These results demonstrate the effectiveness and robustness of our method across different backbone architectures. + +Distance Metric for Graph Matching. We evaluate different distance metrics $\mathcal{D}_W$ used in Sec. 3.2.1, including Wasserstein Distance, Cosine Distance, KL Divergence, and Euclidean Distance. As shown in Fig. 8, the Wasserstein Distance shows the best results as it provides an intuitive measure of the overall shape of sorted eigenvalue distributions. While we use Wasserstein Distance due to its performance, the other well-known metrics exhibited comparable results, demonstrating the robustness of our method in terms of the choice of a distance metric and further supporting the effectiveness of our proposed approach. + +# 5. Conclusion + +In this work, we present CASS, a training-free approach for OVSS that integrates object-level context into CLIP. By distilling object-aware structural information from VFM into CLIP's attention, CASS effectively enhances intra-object coherence, ensuring that each object is consistently represented as a unified semantic entity. Additionally, we leverage object-presence prior knowledge to adjust text embeddings and refine object class assignments, aligning them with object-specific semantic targets in the image. Consequently, CASS effectively groups object components and assigns semantic labels without extra training or additional labeled data, making it a robust solution across diverse datasets and visual contexts. + +# References + +[1] Nikita Araslanov and Stefan Roth. Single-stage semantic segmentation from image labels. In CVPR, 2020. 7 +[2] Danny Barash and Matan Gavish. Optimal shrinkage of singular values under random data contamination. NeurIPS, 2017. 5 +[3] Luca Barsellotti, Roberto Amoroso, Lorenzo Baraldi, and Rita Cucchiara. Fossil: Free open-vocabulary semantic segmentation through synthetic references retrieval. In WACV, 2024. 2, 6 +[4] Luca Barsellotti, Roberto Amoroso, Marcella Cornia, Lorenzo Baraldi, and Rita Cucchiara. Training-free open-vocabulary segmentation with offline diffusion-augmented prototype generation. In CVPR, 2024. 2, 6 +[5] Toby Boas, Aritra Dutta, Xin Li, Kathryn P Mercier, and Eric Niderman. Shrinkage function and its applications in matrix approximation. arXiv preprint arXiv:1601.07600, 2016. 5 +[6] Walid Bousselham, Felix Petersen, Vittorio Ferrari, and Hilde Kuehne. Grounding everything: Emerging localization properties in vision-language transformers. In CVPR, 2024. 6 +[7] Holger Caesar, Jasper Uijlings, and Vittorio Ferrari. COCO-Stuff: Thing and Stuff Classes in Context. In CVPR, 2018. 6, 7, 8 +[8] Mathilde Caron, Hugo Touvron, Ishan Misra, Hervé Jégou, Julien Mairal, Piotr Bojanowski, and Armand Joulin. Emerging Properties in Self-Supervised Vision Transformers. In ICCV, 2021. 2, 4, 6 +[9] Junbum Cha, Jonghwan Mun, and Byungseok Roh. Learning To Generate Text-Grounded Mask for Open-World Semantic Segmentation From Only Image-Text Pairs. In CVPR, 2023. 1, 6 +[10] Liang-Chieh Chen, Yukun Zhu, George Papandreou, Florian Schroff, and Hartwig Adam. Encoder-decoder with atrous separable convolution for semantic image segmentation. In ECCV, 2018. 1, 3 +[11] Seokju Cho, Heeseong Shin, Sunghwan Hong, Anurag Arnab, Paul Hongsuck Seo, and Seungryong Kim. Cat-seg: Cost aggregation for open-vocabulary semantic segmentation. In CVPR, 2024. 1 +[12] Fan RK Chung. Spectral graph theory. American Mathematical Soc., 1997. 5 +[13] Marius Cordts, Mohamed Omran, Sebastian Ramos, Timo Rehfeld, Markus Enzweiler, Rodrigo Benenson, Uwe Franke, Stefan Roth, and Bernt Schiele. The Cityscapes Dataset for Semantic Urban Scene Understanding. In CVPR, 2016. 6 +[14] Alexey Dosovitskiy, Lucas Beyer, Alexander Kolesnikov, Dirk Weissenborn, Xiaohua Zhai, Thomas Unterthiner, Mostafa Dehghani, Matthias Minderer, Georg Heigold, Sylvain Gelly, Jakob Uszkoreit, and Neil Houlsby. An Image is Worth 16x16 Words: Transformers for Image Recognition at Scale. In ICLR, 2021. 2 +[15] M. Everingham, L. Van Gool, C. K. I. Williams, J. Winn, and A. Zisserman. The PASCAL Visual Object Classes Challenge 2012 (VOC2012) Results, 2012. 6, 7, 8 + +[16] Sina Hajimiri, Ismail Ben Ayed, and Jose Dolz. Pay attention to your neighbours: Training-free open-vocabulary semantic segmentation. In WACV, 2025. 1, 2, 3, 6, 7, 8 +[17] Mark Hamilton, Zhoutong Zhang, Bharath Hariharan, Noah Snavely, and William T Freeman. Unsupervised semantic segmentation by distilling feature correspondences. arXiv preprint arXiv:2203.08414, 2022. 3 +[18] Woojung Han, Seil Kang, Kyobin Choo, and Seong Jae Hwang. Complementary branch fusing class and semantic knowledge for robust weakly supervised semantic segmentation. Pattern Recognition, 2025. 1 +[19] Ali Hatamizadeh, Yucheng Tang, Vishwesh Nath, Dong Yang, Andriy Myronenko, Bennett Landman, Holger R Roth, and Daguang Xu. Unetr: Transformers for 3d medical image segmentation. In CVPR, 2022. 1, 3 +[20] Ruixiang Jiang, Lingbo Liu, and Changwen Chen. Clipcount: Towards text-guided zero-shot object counting. In ACM Multimedia, 2023. 2, 5 +[21] Dahiyun Kang and Minsu Cho. In defense of lazy visual grounding for open-vocabulary semantic segmentation. In ECCV, 2024. 2, 6, 7, 8 +[22] Seil Kang, Jinyeong Kim, Junhyeok Kim, and Seong Jae Hwang. See what you are told: Visual attention sink in large multimodal models. In ICLR, 2025. 4 +[23] Seil Kang, Jinyeong Kim, Junhyeok Kim, and Seong Jae Hwang. Your large vision-language model only needs a few attention heads for visual grounding. In CVPR, 2025. 2 +[24] Chanyoung Kim, Woogung Han, Dayun Ju, and Seong Jae Hwang. Eagle: Eigen aggregation learning for object-centric unsupervised semantic segmentation. In CVPR, 2024. 2, 3 +[25] Alexander Kirillov, Eric Mintun, Nikhila Ravi, Hanzi Mao, Chloe Rolland, Laura Gustafson, Tete Xiao, Spencer Whitehead, Alexander C Berg, Wan-Yen Lo, et al. Segment anything. In ICCV, 2023. 3 +[26] Philipp Krahenbuhl and Vladlen Koltun. Efficient Inference in Fully Connected CRFs with Gaussian Edge Potentials. In NeurIPS, 2011. 7 +[27] Harold W Kuhn. The hungarian method for the assignment problem. Naval research logistics quarterly, 1955. 4 +[28] Mengcheng Lan, Chaofeng Chen, Yiping Ke, Xinjiang Wang, Litong Feng, and Wayne Zhang. Clearclip: Decomposing clip representations for dense vision-language inference. In ECCV, 2024. 1, 6 +[29] Mengcheng Lan, Chaofeng Chen, Yiping Ke, Xinjiang Wang, Litong Feng, and Wayne Zhang. Proxyclip: Proxy attention improves clip for open-vocabulary segmentation. In ECCV, 2024. 1, 2, 3, 6, 7, 8 +[30] Yi Li, Hualiang Wang, Yiqun Duan, and Xiaomeng Li. Clip surgery for better explainability with enhancement in open-vocabulary tasks. arXiv preprint arXiv:2304.05653, 2023. 2 +[31] Yiran Li, Junpeng Wang, Xin Dai, Liang Wang, ChinChia Michael Yeh, Yan Zheng, Wei Zhang, and Kwan-Liu Ma. How does attention work in vision transformers? a visual analytics attempt. IEEE Transactions on Visualization and Computer Graphics, 2023. 4 + +[32] Feng Liang, Bichen Wu, Xiaoliang Dai, Kunpeng Li, Yinan Zhao, Hang Zhang, Peizhao Zhang, Peter Vajda, and Diana Marculescu. Open-vocabulary semantic segmentation with mask-adapted clip. In CVPR, 2023. 1 +[33] Jiayun Luo, Siddhesh Khandelwal, Leonid Sigal, and Boyang Li. Emergent open-vocabulary semantic segmentation from off-the-shelf vision-language models. In CVPR, 2024. 6 +[34] Yunyao Mao, Jiajun Deng, Wengang Zhou, Li Li, Yao Fang, and Houqiang Li. Clip4hoi: towards adapting clip for practical zero-shot hoi detection. NeurIPS, 2023. 2, 5 +[35] Octave Mariotti, Oisin Mac Aodha, and Hakan Bilen. Improving semantic correspondence with viewpoint-guided spherical maps. In CVPR, 2024. 3 +[36] Luke Melas-Kyriazi, Christian Rupprecht, Iro Laina, and Andrea Vedaldi. Deep spectral methods: A surprisingly strong baseline for unsupervised semantic segmentation and localization. In CVPR, 2022. 2, 3 +[37] Roozbeh Mottaghi, Xianjie Chen, Xiaobai Liu, Nam-Gyu Cho, Seong-Whan Lee, Sanja Fidler, Raquel Urtasun, and Alan Yuille. The Role of Context for Object Detection and Semantic Segmentation in the Wild. In CVPR, 2014. 6, 7, 8 +[38] Maxime Oquab, Timothee Darct, Theo Moutakanni, Huy Vo, Marc Szafraniec, Vasil Khalidov, Pierre Fernandez, Daniel Haziza, Francisco Massa, Alaaeldin El-Nouby, et al. Dinov2: Learning robust visual features without supervision. arXiv preprint arXiv:2304.07193, 2023. 2 +[39] Alec Radford, Jong Wook Kim, Chris Hallacy, Aditya Ramesh, Gabriel Goh, Sandhini Agarwal, Girish Sastry, Amanda Askell, Pamela Mishkin, Jack Clark, et al. Learning Transferable Visual Models From Natural Language Supervision. In ICML, 2021. 1, 2, 3, 5, 6, 7 +[40] Olaf Ronneberger, Philipp Fischer, and Thomas Brox. U-net: Convolutional networks for biomedical image segmentation. In MICCAI, 2015. 1, 3 +[41] Maximilian Seitzer, Max Horn, Andrii Zadaianchuk, Dominik Zietlow, Tianjun Xiao, Carl-Johann Simon-Gabriel, Tong He, Zheng Zhang, Bernhard Scholkopf, Thomas Brox, and Francesco Locatello. Bridging the gap to real-world object-centric learning. In ICLR, 2023. 3 +[42] Hyun Seok Seong, WonJun Moon, SuBeen Lee, and Jae-Pil Heo. Leveraging hidden positives for unsupervised semantic segmentation. In CVPR, 2023. 3 +[43] Xiangheng Shan, Dongyue Wu, Guilin Zhu, Yuanjie Shao, Nong Sang, and Changxin Gao. Open-vocabulary semantic segmentation with image embedding balancing. In CVPR, 2024. 1 +[44] Tong Shao, Zhuotao Tian, Hang Zhao, and Jingyong Su. Explore the potential of clip for training-free open vocabulary semantic segmentation. In ECCV, 2024. 3, 6, 7 +[45] Gyungin Shin, Weidi Xie, and Samuel Albanie. ReCo: Retrieve and Co-segment for Zero-shot Transfer. In NeurIPS, 2022. 6 +[46] Oriane Simeoni, Gilles Puy, Huy V Vo, Simon Roburin, Spyros Gidaris, Andrei Bursuc, Patrick Pérez, Renaud Marlet, and Jean Ponce. Localizing objects with self-supervised transformers and no labels. In BMVC, 2021. 3 + +[47] Shuyang Sun, Runjia Li, Philip Torr, Xiuye Gu, and Siyang Li. Clip as rn: Segment countless visual concepts without training endeavor. In CVPR, 2024. 2, 6 +[48] Feng Wang, Jieru Mei, and Alan Yuille. Sclip: Rethinking self-attention for dense vision-language inference. In ECCV, 2024. 1, 2, 3, 6, 7, 8 +[49] Xudong Wang, Rohit Girdhar, Stella X Yu, and Ishan Misra. Cut and Learn for Unsupervised Object Detection and Instance Segmentation. In CVPR, 2023. 3 +[50] Xin Wen, Bingchen Zhao, Anlin Zheng, Xiangyu Zhang, and Xiaojuan Qi. Self-supervised visual representation learning with semantic grouping. NeurIPS, 2022. 3 +[51] Ji-Jia Wu, Andy Chia-Hao Chang, Chieh-Yu Chuang, Chun-Pei Chen, Yu-Lun Liu, Min-Hung Chen, Hou-Ning Hu, Yung-Yu Chuang, and Yen-Yu Lin. Image-text codecomposition for text-supervised semantic segmentation. In CVPR, 2024. 6 +[52] Monika Wysoczanska, Oriane Simeoni, Michael Ramamonjisoa, Andrei Bursuc, Tomasz Trzciński, and Patrick Pérez. Clip-dinoiser: Teaching clip a few dino tricks. arXiv preprint arXiv:2312.12359, 2023. 6, 7 +[53] Bin Xie, Jiale Cao, Jin Xie, Fahad Shahbaz Khan, and Yanwei Pang. Sed: A simple encoder-decoder for open-vocabulary semantic segmentation. In CVPR, 2024. 1 +[54] Jiarui Xu, Shalini De Mello, Sifei Liu, Wonmin Byeon, Thomas Breuel, Jan Kautz, and Xiaolong Wang. GroupViT: Semantic Segmentation Emerges From Text Supervision. In CVPR, 2022. 1, 6 +[55] Jiarui Xu, Sifei Liu, Arash Vahdat, Wonmin Byeon, Xiaolong Wang, and Shalini De Mello. Open-vocabulary panoptic segmentation with text-to-image diffusion models. In CVPR, 2023. 1 +[56] Mengde Xu, Zheng Zhang, Fangyun Wei, Han Hu, and Xiang Bai. Side adapter network for open-vocabulary semantic segmentation. In CVPR, 2023. 1 +[57] Bolei Zhou, Hang Zhao, Xavier Puig, Tete Xiao, Sanja Fidler, Adela Barriuso, and Antonio Torralba. Semantic Understanding of Scenes Through the ADE20K Dataset. IJCV, 2019. 6, 7 +[58] Chong Zhou, Chen Change Loy, and Bo Dai. Extract Free Dense Labels from CLIP. In ECCV, 2022. 1, 2, 6 +[59] Lianghui Zhu, Yingyue Li, Jiemin Fang, Yan Liu, Hao Xin, Wenyu Liu, and Xinggang Wang. Weaktr: Exploring plain vision transformer for weakly-supervised semantic segmentation. arXiv preprint arXiv:2304.01184, 2023. 4 \ No newline at end of file diff --git a/2025/Distilling Spectral Graph for Object-Context Aware Open-Vocabulary Semantic Segmentation/images.zip b/2025/Distilling Spectral Graph for Object-Context Aware Open-Vocabulary Semantic Segmentation/images.zip new file mode 100644 index 0000000000000000000000000000000000000000..084b961c1281502ea77b8889858a8e86b603b02f --- /dev/null +++ b/2025/Distilling Spectral Graph for Object-Context Aware Open-Vocabulary Semantic Segmentation/images.zip @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:db943c4aa27b0d3ecfe8043babef9780c7daa5f4b68ace757520256f1dc9c43f +size 826935 diff --git a/2025/Distilling Spectral Graph for Object-Context Aware Open-Vocabulary Semantic Segmentation/layout.json b/2025/Distilling Spectral Graph for Object-Context Aware Open-Vocabulary Semantic Segmentation/layout.json new file mode 100644 index 0000000000000000000000000000000000000000..74c40531a64dc2fd03c3880c2ff058864528ea07 --- /dev/null +++ b/2025/Distilling Spectral Graph for Object-Context Aware Open-Vocabulary Semantic Segmentation/layout.json @@ -0,0 +1,9622 @@ +{ + "pdf_info": [ + { + "para_blocks": [ + { + "bbox": [ + 145, + 103, + 465, + 140 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 145, + 103, + 465, + 140 + ], + "spans": [ + { + "bbox": [ + 145, + 103, + 465, + 140 + ], + "type": "text", + "content": "Distilling Spectral Graph for Object-Context Aware Open-Vocabulary Semantic Segmentation" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 80, + 160, + 529, + 177 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 80, + 160, + 529, + 177 + ], + "spans": [ + { + "bbox": [ + 80, + 160, + 529, + 177 + ], + "type": "text", + "content": "Chanyoung " + }, + { + "bbox": [ + 80, + 160, + 529, + 177 + ], + "type": "inline_equation", + "content": "\\mathrm{Kim}^1" + }, + { + "bbox": [ + 80, + 160, + 529, + 177 + ], + "type": "text", + "content": " Dayun Ju" + }, + { + "bbox": [ + 80, + 160, + 529, + 177 + ], + "type": "inline_equation", + "content": "^1" + }, + { + "bbox": [ + 80, + 160, + 529, + 177 + ], + "type": "text", + "content": " Woojung Han" + }, + { + "bbox": [ + 80, + 160, + 529, + 177 + ], + "type": "inline_equation", + "content": "^1" + }, + { + "bbox": [ + 80, + 160, + 529, + 177 + ], + "type": "text", + "content": " Ming-Hsuan Yang" + }, + { + "bbox": [ + 80, + 160, + 529, + 177 + ], + "type": "inline_equation", + "content": "^{1,2}" + }, + { + "bbox": [ + 80, + 160, + 529, + 177 + ], + "type": "text", + "content": " Seong Jae Hwang" + }, + { + "bbox": [ + 80, + 160, + 529, + 177 + ], + "type": "inline_equation", + "content": "^1" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 171, + 188, + 438, + 203 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 171, + 188, + 438, + 203 + ], + "spans": [ + { + "bbox": [ + 171, + 188, + 438, + 203 + ], + "type": "inline_equation", + "content": "^{1}" + }, + { + "bbox": [ + 171, + 188, + 438, + 203 + ], + "type": "text", + "content": "Yonsei University " + }, + { + "bbox": [ + 171, + 188, + 438, + 203 + ], + "type": "inline_equation", + "content": "^{2}" + }, + { + "bbox": [ + 171, + 188, + 438, + 203 + ], + "type": "text", + "content": "University of California, Merced" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 106, + 205, + 501, + 217 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 106, + 205, + 501, + 217 + ], + "spans": [ + { + "bbox": [ + 106, + 205, + 501, + 217 + ], + "type": "text", + "content": "{chanyoung, juda0707, dnwjDDL, seongjae}@yonsei.ac.kr,mhyang@ucmerced.edu" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 151, + 244, + 200, + 257 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 151, + 244, + 200, + 257 + ], + "spans": [ + { + "bbox": [ + 151, + 244, + 200, + 257 + ], + "type": "text", + "content": "Abstract" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 55, + 270, + 296, + 559 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 270, + 296, + 559 + ], + "spans": [ + { + "bbox": [ + 55, + 270, + 296, + 559 + ], + "type": "text", + "content": "Open-Vocabulary Semantic Segmentation (OVSS) has advanced with recent vision-language models (VLMs), enabling segmentation beyond predefined categories through various learning schemes. Notably, training-free methods offer scalable, easily deployable solutions for handling unseen data, a key goal of OVSS. Yet, a critical issue persists: lack of object-level context consideration when segmenting complex objects in the challenging environment of OVSS based on arbitrary query prompts. This oversight limits models' ability to group semantically consistent elements within object and map them precisely to user-defined arbitrary classes. In this work, we introduce a novel approach that overcomes this limitation by incorporating object-level contextual knowledge within images. Specifically, our model enhances intra-object consistency by distilling spectral-driven features from vision foundation models into the attention mechanism of the visual encoder, enabling semantically coherent components to form a single object mask. Additionally, we refine the text embeddings with zero-shot object presence likelihood to ensure accurate alignment with the specific objects represented in the images. By leveraging object-level contextual knowledge, our proposed approach achieves state-of-the-art performance with strong generalizability across diverse datasets." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 56, + 574, + 135, + 586 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 574, + 135, + 586 + ], + "spans": [ + { + "bbox": [ + 56, + 574, + 135, + 586 + ], + "type": "text", + "content": "1. Introduction" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 55, + 595, + 296, + 691 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 595, + 296, + 691 + ], + "spans": [ + { + "bbox": [ + 55, + 595, + 296, + 691 + ], + "type": "text", + "content": "Open-vocabulary semantic segmentation (OVSS) [9, 32, 54, 55, 58] aims to predict pixel-level labels for arbitrary prompts defined by user input (e.g., the proper noun Space Needle in Fig. 1(a)), advancing beyond previous models scoped to predefined static classes [10, 18, 19, 40]. To achieve this, fully supervised OVSS models [11, 53, 56] enhance generalization to seen classes using labeled training data. Yet, their reliance on seen classes during training may" + } + ] + } + ], + "index": 9 + }, + { + "type": "image", + "bbox": [ + 317, + 244, + 553, + 388 + ], + "blocks": [ + { + "bbox": [ + 317, + 244, + 553, + 388 + ], + "lines": [ + { + "bbox": [ + 317, + 244, + 553, + 388 + ], + "spans": [ + { + "bbox": [ + 317, + 244, + 553, + 388 + ], + "type": "image", + "image_path": "f14b060ee2ae67aafe5e2f19fa931c75774a2ad9eef48c4745ae849936ee7544.jpg" + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 318, + 388, + 379, + 397 + ], + "lines": [ + { + "bbox": [ + 318, + 388, + 379, + 397 + ], + "spans": [ + { + "bbox": [ + 318, + 388, + 379, + 397 + ], + "type": "text", + "content": "(b) Visualization" + } + ] + } + ], + "index": 11, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 416, + 388, + 446, + 397 + ], + "lines": [ + { + "bbox": [ + 416, + 388, + 446, + 397 + ], + "spans": [ + { + "bbox": [ + 416, + 388, + 446, + 397 + ], + "type": "text", + "content": "Truck" + } + ] + } + ], + "index": 12, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 447, + 388, + 496, + 397 + ], + "lines": [ + { + "bbox": [ + 447, + 388, + 496, + 397 + ], + "spans": [ + { + "bbox": [ + 447, + 388, + 496, + 397 + ], + "type": "text", + "content": "Horse" + } + ] + } + ], + "index": 13, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 498, + 388, + 529, + 397 + ], + "lines": [ + { + "bbox": [ + 498, + 388, + 529, + 397 + ], + "spans": [ + { + "bbox": [ + 498, + 388, + 529, + 397 + ], + "type": "text", + "content": "Cow" + } + ] + } + ], + "index": 14, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 529, + 388, + 551, + 397 + ], + "lines": [ + { + "bbox": [ + 529, + 388, + 551, + 397 + ], + "spans": [ + { + "bbox": [ + 529, + 388, + 551, + 397 + ], + "type": "text", + "content": "Sheep" + } + ] + } + ], + "index": 15, + "angle": 0, + "type": "image_caption" + } + ], + "index": 10 + }, + { + "type": "image", + "bbox": [ + 316, + 399, + 553, + 503 + ], + "blocks": [ + { + "bbox": [ + 316, + 399, + 553, + 503 + ], + "lines": [ + { + "bbox": [ + 316, + 399, + 553, + 503 + ], + "spans": [ + { + "bbox": [ + 316, + 399, + 553, + 503 + ], + "type": "image", + "image_path": "7ba854727b4ad18975b4615743c47a45a7bd4c97aea34938889ec67f8e7f96bc.jpg" + } + ] + } + ], + "index": 16, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 313, + 506, + 555, + 638 + ], + "lines": [ + { + "bbox": [ + 313, + 506, + 555, + 638 + ], + "spans": [ + { + "bbox": [ + 313, + 506, + 555, + 638 + ], + "type": "text", + "content": "Figure 1. We present CASS, object-level Context-Aware training-free open-vocabulary Semantic Segmentation model. (a) Overview: Our method distills the vision foundation model's (VFM) object-level contextual spectral graph into CLIP's attention and refines query text embeddings towards object-specific semantics. (b) Object-Level Context: This result illustrates how incorporating object-level context improves segmentation accuracy by unifying object-wise components into a user-defined object class; for example, (top row) our approach precisely segments the truck's body, wheels, and cargo area and (bottom row) accurately groups elements into a single object, such as the sheep, whereas baselines [16, 29] often fail to achieve this unified segmentation." + } + ] + } + ], + "index": 17, + "angle": 0, + "type": "image_caption" + } + ], + "index": 16 + }, + { + "bbox": [ + 313, + 652, + 553, + 676 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 652, + 553, + 676 + ], + "spans": [ + { + "bbox": [ + 313, + 652, + 553, + 676 + ], + "type": "text", + "content": "risk overfitting and limit scalability, as labeled data is required for retraining to adapt to new domains [43]." + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 313, + 677, + 555, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 677, + 555, + 713 + ], + "spans": [ + { + "bbox": [ + 313, + 677, + 555, + 713 + ], + "type": "text", + "content": "To better address this, recent studies [16, 28, 29, 48, 58] propose training-free OVSS methods that leverage pretrained vision-language models (VLMs) (e.g., CLIP [39]) to" + } + ] + } + ], + "index": 19 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "spans": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "type": "text", + "content": "CVF" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 145, + 0, + 496, + 37 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 145, + 0, + 496, + 37 + ], + "spans": [ + { + "bbox": [ + 145, + 0, + 496, + 37 + ], + "type": "text", + "content": "This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 69, + 703, + 258, + 712 + ], + "type": "page_footnote", + "angle": 0, + "lines": [ + { + "bbox": [ + 69, + 703, + 258, + 712 + ], + "spans": [ + { + "bbox": [ + 69, + 703, + 258, + 712 + ], + "type": "text", + "content": "Project Page: https://micv-yonsei.github.io/cass/" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "text", + "content": "15033" + } + ] + } + ], + "index": 21 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 0 + }, + { + "para_blocks": [ + { + "bbox": [ + 55, + 72, + 294, + 204 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 72, + 294, + 204 + ], + "spans": [ + { + "bbox": [ + 55, + 72, + 294, + 204 + ], + "type": "text", + "content": "integrate aligned visual and textual representations. These works focus on improving semantic association between visual patches in CLIP image encoder to produce more precise segmentation maps, achieving highly promising results. This training-free scheme offers key advantages: it (1) generalizes well to unseen classes without requiring additional labeled data, making it more scalable, thus (2) allows direct adaptation to dynamic real-world applications. Capitalizing on these benefits, training-free OVSS aligns with OVSS's core goal of adaptability across diverse domains, enabling it to reach its full potential." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 56, + 207, + 296, + 434 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 207, + 296, + 434 + ], + "spans": [ + { + "bbox": [ + 56, + 207, + 296, + 434 + ], + "type": "text", + "content": "We tackle OVSS in a training-free manner, presenting a challenging and practical approach aligned with OVSS's true objectives. Although existing training-free methods [21, 29, 47, 48] perform well on numerous tasks, they all share a common limitation: the lack of \"object-level context.\" To understand \"object-level context\", consider a truck and sheep in Fig. 1(b). Components comprising a truck, such as its wheels and cargo area, should be grouped as a single entity. Similarly, components of a sheep should be accurately unified under the correct user-given object class. However, existing training-free OVSS methods, which use CLIP as the image backbone for segmentation, often struggle to capture object-level context, failing to group object components into a single, user-defined object. These challenges stem from CLIP's focus on learning global image semantics, which may be insufficient for capturing dense object-level semantics [16, 48, 58]. Thus, to achieve accurate training-free OVSS, a proper object-level context must be considered in CLIP." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 55, + 437, + 295, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 437, + 295, + 714 + ], + "spans": [ + { + "bbox": [ + 55, + 437, + 295, + 714 + ], + "type": "text", + "content": "We address the challenges of training-free OVSS by introducing the object-level context and propose the Context-Aware Semantic Segmentation (CASS) model. Vision foundation models (VFMs), such as DINO [8, 38], capture fine-grained, patch-level semantics but lack object-level context, making them unsuitable for direct use in tasks requiring such context. Recent studies [24, 36] have shown that applying graph spectral techniques to VFM attention graphs can transform patch-level representations into object-level representations. Accordingly, we apply spectral techniques to extract essential object-level context from the VFM attention graph. The extracted spectral features can be used directly for segmentation (e.g., clustering eigenvectors); however, this approach may not be suitable for OVSS, where textual alignment is essential. Thus, we propose distilling low-rank components of VFM into CLIP to enhance its object-level context understanding, while maintaining alignment with textual information. Specifically, we decompose the VFM's attention graph to extract low-rank components, filtering irrelevant information while emphasizing essential object-level context. These low-rank components are then distilled into CLIP's attention to embed the fundamental object structure of the VFM graph." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 313, + 72, + 553, + 228 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 72, + 553, + 228 + ], + "spans": [ + { + "bbox": [ + 313, + 72, + 553, + 228 + ], + "type": "text", + "content": "Our model also leverages CLIP's highly effective zero-shot object classification capability (i.e., object presence prior), widely validated in prior work [20, 34, 39], to capture detailed object-level context within scenes. As such, we adjust the text embeddings based on the object presence prior encoded by CLIP. This process involves refining the embeddings to better align with object-specific semantics. Then, we refine patch-text similarities using the object presence prior, ensuring that the final segmentation map reflects an object-specific perspective. Our CASS, carefully designed to enhance object-level contextual knowledge, captures richer intra-object coherence and enables cohesive groupings of semantically related elements." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 323, + 229, + 488, + 239 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 323, + 229, + 488, + 239 + ], + "spans": [ + { + "bbox": [ + 323, + 229, + 488, + 239 + ], + "type": "text", + "content": "The main contributions of this work are:" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 313, + 241, + 553, + 361 + ], + "type": "list", + "angle": 0, + "index": 8, + "blocks": [ + { + "bbox": [ + 314, + 241, + 553, + 289 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 314, + 241, + 553, + 289 + ], + "spans": [ + { + "bbox": [ + 314, + 241, + 553, + 289 + ], + "type": "text", + "content": "- We enhance object-level context by distilling rich semantic understanding from VFM into the CLIP attention mechanism, facilitating a more consistent grouping of each object into a unified semantic entity." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 313, + 289, + 553, + 324 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 289, + 553, + 324 + ], + "spans": [ + { + "bbox": [ + 313, + 289, + 553, + 324 + ], + "type": "text", + "content": "- We present object presence prior-driven text embeddings and patch-text similarity refinement, promoting the model to precisely classify objects." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 313, + 325, + 553, + 361 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 325, + 553, + 361 + ], + "spans": [ + { + "bbox": [ + 313, + 325, + 553, + 361 + ], + "type": "text", + "content": "- We demonstrate the performance of our model in various semantic segmentation datasets and achieve state-of-the-art performance in training-free OVSS." + } + ] + } + ], + "index": 7 + } + ], + "sub_type": "text" + }, + { + "bbox": [ + 313, + 369, + 406, + 381 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 369, + 406, + 381 + ], + "spans": [ + { + "bbox": [ + 313, + 369, + 406, + 381 + ], + "type": "text", + "content": "2. Related Works" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 313, + 387, + 553, + 650 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 387, + 553, + 650 + ], + "spans": [ + { + "bbox": [ + 313, + 387, + 553, + 650 + ], + "type": "text", + "content": "Open-Vocabulary Semantic Segmentation. Open-vocabulary semantic segmentation (OVSS) aims to segment images corresponding to arbitrary user-given prompts, using vision-language aligned representations (e.g., CLIP [39]). Meanwhile, training-free approaches [3, 4, 23, 29, 58] have risen for their ability to adapt directly to new tasks without extra training. Among these, methods [16, 29, 48, 58] that do not rely on additional datasets have gained attention, as they require no extra adaptation effort, aligning better with the objective of OVSS. Earlier approaches investigated leveraging CLIP features for better localization [30, 58] or for grounding proposed masks [21, 47]. SCLIP [48] uses correlative self-attention to improve inter-patch correlations. NACLIP [16] applies a Gaussian kernel to attention maps for better spatial consistency, and ProxyCLIP [29] integrates patch-wise consistent features from DINO [8]. Although these methods demonstrated strong results, they often struggle to group different parts of the same object due to a lack of explicit dense label knowledge. To this end, we attribute the underlying challenges of training-free OVSS by enriching the object-level contextual understanding." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 313, + 653, + 553, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 653, + 553, + 713 + ], + "spans": [ + { + "bbox": [ + 313, + 653, + 553, + 713 + ], + "type": "text", + "content": "Vision Foundation Models. Vision Foundation Models (VFMs), especially vision transformer (ViT)-based [14] self-supervised models (e.g., DINO [8, 38]), have significantly advanced computer vision through their ability to capture patch-wise semantic understanding at both the" + } + ] + } + ], + "index": 11 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "text", + "content": "15034" + } + ] + } + ], + "index": 12 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 1 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 58, + 70, + 550, + 243 + ], + "blocks": [ + { + "bbox": [ + 58, + 70, + 550, + 243 + ], + "lines": [ + { + "bbox": [ + 58, + 70, + 550, + 243 + ], + "spans": [ + { + "bbox": [ + 58, + 70, + 550, + 243 + ], + "type": "image", + "image_path": "e6308b01ef7816f3a5505a7104fb111a555c41f315081604abbbb7ad4ffc6240.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 55, + 245, + 555, + 281 + ], + "lines": [ + { + "bbox": [ + 55, + 245, + 555, + 281 + ], + "spans": [ + { + "bbox": [ + 55, + 245, + 555, + 281 + ], + "type": "text", + "content": "Figure 2. Detailed illustration of our proposed training-free spectral object-level context distillation mechanism in Sec. 3.2. By matching the attention graphs of VFM and CLIP head-by-head to establish complementary relationships, and distilling the fundamental object-level context of the VFM graph to CLIP, we enhance CLIP's ability to capture intra-object contextual coherence." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "bbox": [ + 54, + 289, + 295, + 420 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 54, + 289, + 295, + 420 + ], + "spans": [ + { + "bbox": [ + 54, + 289, + 295, + 420 + ], + "type": "text", + "content": "feature and attention levels. Among various downstream tasks [17, 35, 42, 46, 49] leveraging VFM's capability, one line of research focuses on leveraging the attention components in VFM [24, 36, 46], which exhibit a detailed understanding of patch-wise semantic interaction between different image regions. Thus, while ProxyCLIP [29] utilizes VFM features, our method takes a different effort by treating the attention mechanism in VFM as a graph. We then extract core structural patterns of the VFM graph to emphasize object-level context, providing an effective solution for producing precise object masks in training-free OVSS." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 55, + 424, + 296, + 639 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 424, + 296, + 639 + ], + "spans": [ + { + "bbox": [ + 55, + 424, + 296, + 639 + ], + "type": "text", + "content": "Object-Level Semantic Segmentation. Semantic segmentation groups object components that have similar semantics into a single mask, traditionally achieved through models trained on extensive human-labeled, pixel-wise annotations [10, 19, 25, 40]. However, despite recent advancements in unsupervised semantic segmentation [17, 42] (USS) that do not rely on dense labels, a core challenge persists: individual object components are not consistently grouped into a single entity. To address this issue, recent USS studies [24, 41, 50] have incorporated object-level representation learning to improve the model's ability to merge single objects into unified entities. However, these efforts have not been considered in the training-free OVSS task, which faces similar challenges in an unsupervised context. Therefore, we focus on enriching object-level context in training-free OVSS to achieve a more effective grouping of components within each object, enabling clearer and more consistent segmentation across diverse visual domains." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 55, + 645, + 111, + 657 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 645, + 111, + 657 + ], + "spans": [ + { + "bbox": [ + 55, + 645, + 111, + 657 + ], + "type": "text", + "content": "3. Method" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 55, + 665, + 296, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 665, + 296, + 715 + ], + "spans": [ + { + "bbox": [ + 55, + 665, + 296, + 715 + ], + "type": "text", + "content": "We detail our proposed model, CASS, implemented in a training-free manner, as outlined in Fig 1(a). Our model jointly enhances object-level contextual understanding with two primary aspects: (1) Spectral Object-Level" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 313, + 289, + 555, + 326 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 289, + 555, + 326 + ], + "spans": [ + { + "bbox": [ + 313, + 289, + 555, + 326 + ], + "type": "text", + "content": "Context Distillation (Sec. 3.2) and (2) Object Presence-Driven Object-Level Context (Sec. 3.3). We first outline the basic pipeline of training-free OVSS in Sec. 3.1." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 313, + 331, + 400, + 342 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 331, + 400, + 342 + ], + "spans": [ + { + "bbox": [ + 313, + 331, + 400, + 342 + ], + "type": "text", + "content": "3.1. Preliminaries" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 313, + 346, + 555, + 526 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 346, + 555, + 526 + ], + "spans": [ + { + "bbox": [ + 313, + 346, + 555, + 526 + ], + "type": "text", + "content": "CLIP Visual Encoder. Training-free OVSS models [16, 29, 48] leverage CLIP [39] for its rich alignment of visual and linguistic information. However, the CLIP ViT visual encoder " + }, + { + "bbox": [ + 313, + 346, + 555, + 526 + ], + "type": "inline_equation", + "content": "\\mathcal{F}_{\\mathrm{CLIP}}^v" + }, + { + "bbox": [ + 313, + 346, + 555, + 526 + ], + "type": "text", + "content": " focuses on aligning the [CLS] token with text embeddings during training, often neglecting the patchwise spatial interactions among visual tokens that are crucial for dense prediction [16, 44, 48]. To produce CLIP feature representations more suitable for semantic segmentation tasks, we remove residual connection, feed-forward network, and self-attention layer in the final layer of " + }, + { + "bbox": [ + 313, + 346, + 555, + 526 + ], + "type": "inline_equation", + "content": "\\mathcal{F}_{\\mathrm{CLIP}}^v" + }, + { + "bbox": [ + 313, + 346, + 555, + 526 + ], + "type": "text", + "content": ", following recent works [16, 48]. Specifically, when the set of visual tokens " + }, + { + "bbox": [ + 313, + 346, + 555, + 526 + ], + "type": "inline_equation", + "content": "Z = \\mathrm{Concat}[z_1, \\ldots, z_N] \\in \\mathbb{R}^{N \\times D_Z}" + }, + { + "bbox": [ + 313, + 346, + 555, + 526 + ], + "type": "text", + "content": ", where " + }, + { + "bbox": [ + 313, + 346, + 555, + 526 + ], + "type": "inline_equation", + "content": "N" + }, + { + "bbox": [ + 313, + 346, + 555, + 526 + ], + "type": "text", + "content": " is the total number of patches except [CLS] token and " + }, + { + "bbox": [ + 313, + 346, + 555, + 526 + ], + "type": "inline_equation", + "content": "D_Z" + }, + { + "bbox": [ + 313, + 346, + 555, + 526 + ], + "type": "text", + "content": " denotes patch dimension space, is input into the final block of the ViT, the following operations are as" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 393, + 529, + 553, + 542 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 393, + 529, + 553, + 542 + ], + "spans": [ + { + "bbox": [ + 393, + 529, + 553, + 542 + ], + "type": "interline_equation", + "content": "Z ^ {*} = \\operatorname {S A} (\\operatorname {L N} (Z)), \\tag {1}", + "image_path": "f81969ae1930e4acee700bf0a674f694cfdada94e9ae429b20f361a759303811.jpg" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 313, + 544, + 554, + 568 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 544, + 554, + 568 + ], + "spans": [ + { + "bbox": [ + 313, + 544, + 554, + 568 + ], + "type": "text", + "content": "where SA and LN represent Self-Attention and Layer Normalization, respectively. The modified SA layer is" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 362, + 571, + 553, + 590 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 362, + 571, + 553, + 590 + ], + "spans": [ + { + "bbox": [ + 362, + 571, + 553, + 590 + ], + "type": "interline_equation", + "content": "M _ {\\mathrm {C L I P}} ^ {i} = \\operatorname {s o f t m a x} \\left(A _ {\\mathrm {C L I P}} ^ {i} / \\sqrt {D _ {h}}\\right), \\tag {2}", + "image_path": "cbb37b6781f7431d1dd510ddf717200b577d5fd55ee3c983b567bbf16e76b280.jpg" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 326, + 593, + 553, + 608 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 326, + 593, + 553, + 608 + ], + "spans": [ + { + "bbox": [ + 326, + 593, + 553, + 608 + ], + "type": "interline_equation", + "content": "Z ^ {*} = \\operatorname {C o n c a t} \\left[ M _ {\\mathrm {C L I P}} ^ {1} V _ {\\mathrm {C L I P}} ^ {1}, \\dots , M _ {\\mathrm {C L I P}} ^ {h} V _ {\\mathrm {C L I P}} ^ {h} \\right] W ^ {O}, \\tag {3}", + "image_path": "098b23f22fb9aef23d43fb598b27fa217ed9b4d79713033c583aea7e3bdfa64b.jpg" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 313, + 615, + 555, + 687 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 615, + 555, + 687 + ], + "spans": [ + { + "bbox": [ + 313, + 615, + 555, + 687 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 313, + 615, + 555, + 687 + ], + "type": "inline_equation", + "content": "D_h" + }, + { + "bbox": [ + 313, + 615, + 555, + 687 + ], + "type": "text", + "content": " denotes the head dimension, " + }, + { + "bbox": [ + 313, + 615, + 555, + 687 + ], + "type": "inline_equation", + "content": "h" + }, + { + "bbox": [ + 313, + 615, + 555, + 687 + ], + "type": "text", + "content": " is the number of heads, and " + }, + { + "bbox": [ + 313, + 615, + 555, + 687 + ], + "type": "inline_equation", + "content": "W^0 \\in \\mathbb{R}^{D_z \\times D_z}" + }, + { + "bbox": [ + 313, + 615, + 555, + 687 + ], + "type": "text", + "content": " is the output projection matrix. We define CLIP attention map for " + }, + { + "bbox": [ + 313, + 615, + 555, + 687 + ], + "type": "inline_equation", + "content": "i" + }, + { + "bbox": [ + 313, + 615, + 555, + 687 + ], + "type": "text", + "content": "-th head as " + }, + { + "bbox": [ + 313, + 615, + 555, + 687 + ], + "type": "inline_equation", + "content": "A_{\\mathrm{CLIP}}^i = K_{\\mathrm{CLIP}}^i K_{\\mathrm{CLIP}}^{i^\\top} \\in \\mathbb{R}^{N \\times N}" + }, + { + "bbox": [ + 313, + 615, + 555, + 687 + ], + "type": "text", + "content": ", where attention key " + }, + { + "bbox": [ + 313, + 615, + 555, + 687 + ], + "type": "inline_equation", + "content": "K_{\\mathrm{CLIP}} \\in \\mathbb{R}^{N \\times D_h \\times h}" + }, + { + "bbox": [ + 313, + 615, + 555, + 687 + ], + "type": "text", + "content": " and value " + }, + { + "bbox": [ + 313, + 615, + 555, + 687 + ], + "type": "inline_equation", + "content": "V_{\\mathrm{CLIP}} \\in \\mathbb{R}^{N \\times D_h \\times h}" + }, + { + "bbox": [ + 313, + 615, + 555, + 687 + ], + "type": "text", + "content": " result from the linear transformation of " + }, + { + "bbox": [ + 313, + 615, + 555, + 687 + ], + "type": "inline_equation", + "content": "\\mathrm{LN}(Z)" + }, + { + "bbox": [ + 313, + 615, + 555, + 687 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 313, + 689, + 555, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 689, + 555, + 714 + ], + "spans": [ + { + "bbox": [ + 313, + 689, + 555, + 714 + ], + "type": "text", + "content": "Inference Pipeline. We use a sliding window inference to ensure that ViT can capture detailed parts of the image by" + } + ] + } + ], + "index": 14 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "text", + "content": "15035" + } + ] + } + ], + "index": 15 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 2 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 58, + 70, + 293, + 260 + ], + "blocks": [ + { + "bbox": [ + 58, + 70, + 293, + 260 + ], + "lines": [ + { + "bbox": [ + 58, + 70, + 293, + 260 + ], + "spans": [ + { + "bbox": [ + 58, + 70, + 293, + 260 + ], + "type": "image", + "image_path": "29ebad50915d6e3e59ee82cf81eb349c4fd862f5c127cd6d0c6d0a7d45dc4db1.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 55, + 262, + 295, + 328 + ], + "lines": [ + { + "bbox": [ + 55, + 262, + 295, + 328 + ], + "spans": [ + { + "bbox": [ + 55, + 262, + 295, + 328 + ], + "type": "text", + "content": "Figure 3. Attention score visualization for various query points. Left: Vanilla CLIP (" + }, + { + "bbox": [ + 55, + 262, + 295, + 328 + ], + "type": "inline_equation", + "content": "A_{\\mathrm{CLIP}}" + }, + { + "bbox": [ + 55, + 262, + 295, + 328 + ], + "type": "text", + "content": ") shows noisy, unfocused attention. Center: VFM-to-CLIP distillation without low-rank eigenscaling shows partial object grouping with limited detail. Right: Incorporating our low-rank eigenscaling (Eq. (7)) captures object-level context, improving grouping within a single object." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "bbox": [ + 55, + 338, + 296, + 434 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 338, + 296, + 434 + ], + "spans": [ + { + "bbox": [ + 55, + 338, + 296, + 434 + ], + "type": "text", + "content": "dividing the entire image " + }, + { + "bbox": [ + 55, + 338, + 296, + 434 + ], + "type": "inline_equation", + "content": "\\mathcal{I}" + }, + { + "bbox": [ + 55, + 338, + 296, + 434 + ], + "type": "text", + "content": " into smaller windows " + }, + { + "bbox": [ + 55, + 338, + 296, + 434 + ], + "type": "inline_equation", + "content": "\\hat{\\mathcal{I}}" + }, + { + "bbox": [ + 55, + 338, + 296, + 434 + ], + "type": "text", + "content": ". The visual tokens " + }, + { + "bbox": [ + 55, + 338, + 296, + 434 + ], + "type": "inline_equation", + "content": "Z^{*} \\in \\mathbb{R}^{N \\times D_{Z}}" + }, + { + "bbox": [ + 55, + 338, + 296, + 434 + ], + "type": "text", + "content": " are then computed from each " + }, + { + "bbox": [ + 55, + 338, + 296, + 434 + ], + "type": "inline_equation", + "content": "\\hat{\\mathcal{I}}" + }, + { + "bbox": [ + 55, + 338, + 296, + 434 + ], + "type": "text", + "content": " through " + }, + { + "bbox": [ + 55, + 338, + 296, + 434 + ], + "type": "inline_equation", + "content": "\\mathcal{F}_{\\mathrm{CLIP}}^{v}" + }, + { + "bbox": [ + 55, + 338, + 296, + 434 + ], + "type": "text", + "content": " as in Eq. (1). To perform OVSS with given texts, we project the visual tokens " + }, + { + "bbox": [ + 55, + 338, + 296, + 434 + ], + "type": "inline_equation", + "content": "Z^{*}" + }, + { + "bbox": [ + 55, + 338, + 296, + 434 + ], + "type": "text", + "content": " into the vision-language latent space as " + }, + { + "bbox": [ + 55, + 338, + 296, + 434 + ], + "type": "inline_equation", + "content": "F_{\\mathrm{CLIP}} = \\mathcal{P}_{\\mathrm{CLIP}}(Z^{*}) \\in \\mathbb{R}^{N \\times d}" + }, + { + "bbox": [ + 55, + 338, + 296, + 434 + ], + "type": "text", + "content": ", where " + }, + { + "bbox": [ + 55, + 338, + 296, + 434 + ], + "type": "inline_equation", + "content": "\\mathcal{P}_{\\mathrm{CLIP}}" + }, + { + "bbox": [ + 55, + 338, + 296, + 434 + ], + "type": "text", + "content": " is a projection matrix into CLIP latent space and " + }, + { + "bbox": [ + 55, + 338, + 296, + 434 + ], + "type": "inline_equation", + "content": "d" + }, + { + "bbox": [ + 55, + 338, + 296, + 434 + ], + "type": "text", + "content": " is the CLIP latent dimension. We compute patch-text similarity " + }, + { + "bbox": [ + 55, + 338, + 296, + 434 + ], + "type": "inline_equation", + "content": "\\hat{S} \\in \\mathbb{R}^{N \\times C}" + }, + { + "bbox": [ + 55, + 338, + 296, + 434 + ], + "type": "text", + "content": " for each image window " + }, + { + "bbox": [ + 55, + 338, + 296, + 434 + ], + "type": "inline_equation", + "content": "\\hat{\\mathcal{I}}" + }, + { + "bbox": [ + 55, + 338, + 296, + 434 + ], + "type": "text", + "content": " as" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 118, + 442, + 295, + 464 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 118, + 442, + 295, + 464 + ], + "spans": [ + { + "bbox": [ + 118, + 442, + 295, + 464 + ], + "type": "interline_equation", + "content": "\\hat {\\mathcal {S}} = F _ {\\mathrm {C L I P}} \\left[ \\left\\{t _ {\\mathrm {C L I P}} ^ {i} \\right\\} _ {i = 1} ^ {C} \\right] ^ {\\top}, \\tag {4}", + "image_path": "75cb98949f41cef5bcf798193c4bdd91d65f588330e8872675537e841414383f.jpg" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 55, + 472, + 296, + 521 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 472, + 296, + 521 + ], + "spans": [ + { + "bbox": [ + 55, + 472, + 296, + 521 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 55, + 472, + 296, + 521 + ], + "type": "inline_equation", + "content": "\\{t_{\\mathrm{CLIP}}^i\\}_{i = 1}^C \\coloneqq \\mathrm{Concat}\\left[t_{\\mathrm{CLIP}}^1,\\dots ,t_{\\mathrm{CLIP}}^C\\right] \\in \\mathbb{R}^{C\\times d}" + }, + { + "bbox": [ + 55, + 472, + 296, + 521 + ], + "type": "text", + "content": " denotes the encoded text embeddings, and " + }, + { + "bbox": [ + 55, + 472, + 296, + 521 + ], + "type": "inline_equation", + "content": "C" + }, + { + "bbox": [ + 55, + 472, + 296, + 521 + ], + "type": "text", + "content": " is the number of classes. Finally, the results of all windows are combined to create a segmentation map for the entire image." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 55, + 528, + 275, + 541 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 528, + 275, + 541 + ], + "spans": [ + { + "bbox": [ + 55, + 528, + 275, + 541 + ], + "type": "text", + "content": "3.2. Spectral Object-Level Context Distillation" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 55, + 545, + 296, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 545, + 296, + 713 + ], + "spans": [ + { + "bbox": [ + 55, + 545, + 296, + 713 + ], + "type": "text", + "content": "Our model CASS aims to capture such object-level contextual interaction within the image features to associate different parts of the same object. However, the limited ability of CLIP to capture semantic relationships between patches often neglects the object-level context, as illustrated in the left column of Fig. 3 (" + }, + { + "bbox": [ + 55, + 545, + 296, + 713 + ], + "type": "inline_equation", + "content": "A_{\\mathrm{CLIP}}" + }, + { + "bbox": [ + 55, + 545, + 296, + 713 + ], + "type": "text", + "content": "). To address this limitation, we leverage VFM (e.g., DINO [8]), enabling a deeper object-level contextual understanding. Among VFM features, we utilize VFM attention key component " + }, + { + "bbox": [ + 55, + 545, + 296, + 713 + ], + "type": "inline_equation", + "content": "K_{\\mathrm{VFM}} \\in \\mathbb{R}^{h \\times N \\times D_h}" + }, + { + "bbox": [ + 55, + 545, + 296, + 713 + ], + "type": "text", + "content": ". Specifically, we treat the VFM attention adjacency " + }, + { + "bbox": [ + 55, + 545, + 296, + 713 + ], + "type": "inline_equation", + "content": "A_{\\mathrm{VFM}} = K_{\\mathrm{VFM}} K_{\\mathrm{VFM}}^{\\top} \\in \\mathbb{R}^{h \\times N \\times N}" + }, + { + "bbox": [ + 55, + 545, + 296, + 713 + ], + "type": "text", + "content": " as a graph, where " + }, + { + "bbox": [ + 55, + 545, + 296, + 713 + ], + "type": "inline_equation", + "content": "h" + }, + { + "bbox": [ + 55, + 545, + 296, + 713 + ], + "type": "text", + "content": " represents the number of heads, and transfer crucial graph patterns from " + }, + { + "bbox": [ + 55, + 545, + 296, + 713 + ], + "type": "inline_equation", + "content": "A_{\\mathrm{VFM}}" + }, + { + "bbox": [ + 55, + 545, + 296, + 713 + ], + "type": "text", + "content": " to " + }, + { + "bbox": [ + 55, + 545, + 296, + 713 + ], + "type": "inline_equation", + "content": "A_{\\mathrm{CLIP}}" + }, + { + "bbox": [ + 55, + 545, + 296, + 713 + ], + "type": "text", + "content": ". However, a key question remains: \"How do we account for the multi-head attention mechanisms in both" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 313, + 72, + 554, + 168 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 72, + 554, + 168 + ], + "spans": [ + { + "bbox": [ + 313, + 72, + 554, + 168 + ], + "type": "text", + "content": "VFM and CLIP?\" In other words, \"how do we match " + }, + { + "bbox": [ + 313, + 72, + 554, + 168 + ], + "type": "inline_equation", + "content": "i" + }, + { + "bbox": [ + 313, + 72, + 554, + 168 + ], + "type": "text", + "content": "-th VFM attention head " + }, + { + "bbox": [ + 313, + 72, + 554, + 168 + ], + "type": "inline_equation", + "content": "A_{\\mathrm{VFM}}^i" + }, + { + "bbox": [ + 313, + 72, + 554, + 168 + ], + "type": "text", + "content": " with " + }, + { + "bbox": [ + 313, + 72, + 554, + 168 + ], + "type": "inline_equation", + "content": "j" + }, + { + "bbox": [ + 313, + 72, + 554, + 168 + ], + "type": "text", + "content": "-th CLIP attention head " + }, + { + "bbox": [ + 313, + 72, + 554, + 168 + ], + "type": "inline_equation", + "content": "A_{\\mathrm{CLIP}}^j" + }, + { + "bbox": [ + 313, + 72, + 554, + 168 + ], + "type": "text", + "content": "? A straightforward approach is to distill the attention heads sequentially (i.e., " + }, + { + "bbox": [ + 313, + 72, + 554, + 168 + ], + "type": "inline_equation", + "content": "i = j" + }, + { + "bbox": [ + 313, + 72, + 554, + 168 + ], + "type": "text", + "content": "). Nevertheless, recent studies [22, 31, 59] indicate that different attention heads focus on separate parts of the image. Thus, matching the optimal heads between VFM and CLIP is essential for precise attention distillation." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 313, + 175, + 524, + 186 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 175, + 524, + 186 + ], + "spans": [ + { + "bbox": [ + 313, + 175, + 524, + 186 + ], + "type": "text", + "content": "3.2.1. Complementary Spectral Graph Matching" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 313, + 187, + 554, + 282 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 187, + 554, + 282 + ], + "spans": [ + { + "bbox": [ + 313, + 187, + 554, + 282 + ], + "type": "text", + "content": "We match graphs with contrasting structures, as shown in Fig 4, enabling VFM to supplement the object-level contextual knowledge that CLIP alone struggles to capture. To match an optimal head-wise attention graph, we exploit the spectral field of both " + }, + { + "bbox": [ + 313, + 187, + 554, + 282 + ], + "type": "inline_equation", + "content": "A_{\\mathrm{VFM}}" + }, + { + "bbox": [ + 313, + 187, + 554, + 282 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 313, + 187, + 554, + 282 + ], + "type": "inline_equation", + "content": "A_{\\mathrm{CLIP}}" + }, + { + "bbox": [ + 313, + 187, + 554, + 282 + ], + "type": "text", + "content": ". Our graph-matching mechanism has two main steps as illustrated in Fig. 2(a): (I) examining the eigenvalues of each graph and (II) matching optimal graph pairs using the spectral distribution." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 313, + 285, + 554, + 421 + ], + "type": "list", + "angle": 0, + "index": 12, + "blocks": [ + { + "bbox": [ + 313, + 285, + 554, + 370 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 285, + 554, + 370 + ], + "spans": [ + { + "bbox": [ + 313, + 285, + 554, + 370 + ], + "type": "text", + "content": "(I) Eigenvalue Examination. For each graph from VFM and CLIP, denoted as " + }, + { + "bbox": [ + 313, + 285, + 554, + 370 + ], + "type": "inline_equation", + "content": "A_{\\mathrm{VFM}}^i" + }, + { + "bbox": [ + 313, + 285, + 554, + 370 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 313, + 285, + 554, + 370 + ], + "type": "inline_equation", + "content": "A_{\\mathrm{CLIP}}^i" + }, + { + "bbox": [ + 313, + 285, + 554, + 370 + ], + "type": "text", + "content": " for " + }, + { + "bbox": [ + 313, + 285, + 554, + 370 + ], + "type": "inline_equation", + "content": "(i = 1, \\dots, h)" + }, + { + "bbox": [ + 313, + 285, + 554, + 370 + ], + "type": "text", + "content": ", we perform eigendecomposition. This yields the eigenvalues for each head " + }, + { + "bbox": [ + 313, + 285, + 554, + 370 + ], + "type": "inline_equation", + "content": "h" + }, + { + "bbox": [ + 313, + 285, + 554, + 370 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 313, + 285, + 554, + 370 + ], + "type": "inline_equation", + "content": "\\lambda_{\\mathrm{VFM}}^i" + }, + { + "bbox": [ + 313, + 285, + 554, + 370 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 313, + 285, + 554, + 370 + ], + "type": "inline_equation", + "content": "\\lambda_{\\mathrm{CLIP}}^i" + }, + { + "bbox": [ + 313, + 285, + 554, + 370 + ], + "type": "text", + "content": ", from which the top " + }, + { + "bbox": [ + 313, + 285, + 554, + 370 + ], + "type": "inline_equation", + "content": "m" + }, + { + "bbox": [ + 313, + 285, + 554, + 370 + ], + "type": "text", + "content": " fixed eigenvalues are selected. These eigenvalues contain each graph's unique structural features, key properties essential for distinguishing different attention graph patterns." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 313, + 372, + 554, + 421 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 372, + 554, + 421 + ], + "spans": [ + { + "bbox": [ + 313, + 372, + 554, + 421 + ], + "type": "text", + "content": "(II) Graph Matching via Spectral Distribution. After obtaining eigenvalues for each head, we compute spectral distances to quantify structural differences, creating a cost matrix " + }, + { + "bbox": [ + 313, + 372, + 554, + 421 + ], + "type": "inline_equation", + "content": "\\mathcal{C} \\in \\mathbb{R}^{h \\times h}" + }, + { + "bbox": [ + 313, + 372, + 554, + 421 + ], + "type": "text", + "content": " for each graph pair from VFM and CLIP:" + } + ] + } + ], + "index": 11 + } + ], + "sub_type": "text" + }, + { + "bbox": [ + 324, + 429, + 553, + 444 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 324, + 429, + 553, + 444 + ], + "spans": [ + { + "bbox": [ + 324, + 429, + 553, + 444 + ], + "type": "interline_equation", + "content": "\\mathcal {C} _ {i j} = 1 - \\mathcal {D} _ {W} \\left(\\bar {\\lambda} _ {\\mathrm {V F M}} ^ {i}, \\bar {\\lambda} _ {\\mathrm {C L I P}} ^ {j}\\right), \\quad \\forall i, j \\in \\{1, \\dots , h \\} \\tag {5}", + "image_path": "c602c72c7a96befb28241eadbd48e63e7cd9a25bcae386231cb33553e3c2aba6.jpg" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 313, + 451, + 555, + 597 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 451, + 555, + 597 + ], + "spans": [ + { + "bbox": [ + 313, + 451, + 555, + 597 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 313, + 451, + 555, + 597 + ], + "type": "inline_equation", + "content": "\\bar{\\lambda}_{\\mathrm{VFM}}^i" + }, + { + "bbox": [ + 313, + 451, + 555, + 597 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 313, + 451, + 555, + 597 + ], + "type": "inline_equation", + "content": "\\bar{\\lambda}_{\\mathrm{CLIP}}^j" + }, + { + "bbox": [ + 313, + 451, + 555, + 597 + ], + "type": "text", + "content": " represent the normalized eigenvalues of the " + }, + { + "bbox": [ + 313, + 451, + 555, + 597 + ], + "type": "inline_equation", + "content": "i" + }, + { + "bbox": [ + 313, + 451, + 555, + 597 + ], + "type": "text", + "content": "-th head of VFM and the " + }, + { + "bbox": [ + 313, + 451, + 555, + 597 + ], + "type": "inline_equation", + "content": "j" + }, + { + "bbox": [ + 313, + 451, + 555, + 597 + ], + "type": "text", + "content": "-th head of CLIP, respectively, and " + }, + { + "bbox": [ + 313, + 451, + 555, + 597 + ], + "type": "inline_equation", + "content": "\\mathcal{D}_W" + }, + { + "bbox": [ + 313, + 451, + 555, + 597 + ], + "type": "text", + "content": " denotes the Wasserstein distance. Here, the Wasserstein distance is computed as " + }, + { + "bbox": [ + 313, + 451, + 555, + 597 + ], + "type": "inline_equation", + "content": "\\mathcal{D}_W(u,v) = \\sum_{i=1}^{h} |\\operatorname{sort}(u)_i - \\operatorname{sort}(v)_i|" + }, + { + "bbox": [ + 313, + 451, + 555, + 597 + ], + "type": "text", + "content": ", where " + }, + { + "bbox": [ + 313, + 451, + 555, + 597 + ], + "type": "inline_equation", + "content": "u" + }, + { + "bbox": [ + 313, + 451, + 555, + 597 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 313, + 451, + 555, + 597 + ], + "type": "inline_equation", + "content": "v" + }, + { + "bbox": [ + 313, + 451, + 555, + 597 + ], + "type": "text", + "content": " are the two input distributions, and the sort function orders the values of " + }, + { + "bbox": [ + 313, + 451, + 555, + 597 + ], + "type": "inline_equation", + "content": "u" + }, + { + "bbox": [ + 313, + 451, + 555, + 597 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 313, + 451, + 555, + 597 + ], + "type": "inline_equation", + "content": "v" + }, + { + "bbox": [ + 313, + 451, + 555, + 597 + ], + "type": "text", + "content": " in ascending order. After computing " + }, + { + "bbox": [ + 313, + 451, + 555, + 597 + ], + "type": "inline_equation", + "content": "\\mathcal{C}" + }, + { + "bbox": [ + 313, + 451, + 555, + 597 + ], + "type": "text", + "content": ", we apply the Hungarian matching algorithm [27] to the cost matrix to find the optimal pairing of graph heads. This approach pairs graphs with contrasting characteristics, enabling complementary distillation of object-level context knowledge from " + }, + { + "bbox": [ + 313, + 451, + 555, + 597 + ], + "type": "inline_equation", + "content": "A_{\\mathrm{VFM}}" + }, + { + "bbox": [ + 313, + 451, + 555, + 597 + ], + "type": "text", + "content": " to " + }, + { + "bbox": [ + 313, + 451, + 555, + 597 + ], + "type": "inline_equation", + "content": "A_{\\mathrm{CLIP}}" + }, + { + "bbox": [ + 313, + 451, + 555, + 597 + ], + "type": "text", + "content": ", as described in the following section." + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 313, + 602, + 512, + 614 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 602, + 512, + 614 + ], + "spans": [ + { + "bbox": [ + 313, + 602, + 512, + 614 + ], + "type": "text", + "content": "3.2.2. Distilling VFM Spectral Graph to CLIP" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 313, + 615, + 554, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 615, + 554, + 713 + ], + "spans": [ + { + "bbox": [ + 313, + 615, + 554, + 713 + ], + "type": "text", + "content": "Once we match optimal graph pairs, we distill the VFM graph to CLIP (see Fig. 2(b)). For the sake of explanation, we assume that the " + }, + { + "bbox": [ + 313, + 615, + 554, + 713 + ], + "type": "inline_equation", + "content": "i" + }, + { + "bbox": [ + 313, + 615, + 554, + 713 + ], + "type": "text", + "content": "-th VFM attention head " + }, + { + "bbox": [ + 313, + 615, + 554, + 713 + ], + "type": "inline_equation", + "content": "A_{\\mathrm{VFM}}^i" + }, + { + "bbox": [ + 313, + 615, + 554, + 713 + ], + "type": "text", + "content": " is matched with the " + }, + { + "bbox": [ + 313, + 615, + 554, + 713 + ], + "type": "inline_equation", + "content": "j" + }, + { + "bbox": [ + 313, + 615, + 554, + 713 + ], + "type": "text", + "content": "-th CLIP attention head " + }, + { + "bbox": [ + 313, + 615, + 554, + 713 + ], + "type": "inline_equation", + "content": "A_{\\mathrm{CLIP}}^j" + }, + { + "bbox": [ + 313, + 615, + 554, + 713 + ], + "type": "text", + "content": ". An intuitive approach would be simply aggregating " + }, + { + "bbox": [ + 313, + 615, + 554, + 713 + ], + "type": "inline_equation", + "content": "A_{\\mathrm{VFM}}^i" + }, + { + "bbox": [ + 313, + 615, + 554, + 713 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 313, + 615, + 554, + 713 + ], + "type": "inline_equation", + "content": "A_{\\mathrm{CLIP}}^j" + }, + { + "bbox": [ + 313, + 615, + 554, + 713 + ], + "type": "text", + "content": " without any transformation. However, as shown in Fig. 3 (" + }, + { + "bbox": [ + 313, + 615, + 554, + 713 + ], + "type": "inline_equation", + "content": "A_{\\psi w/o \\mathrm{LR}}" + }, + { + "bbox": [ + 313, + 615, + 554, + 713 + ], + "type": "text", + "content": "), this approach may transfer noise or irrelevant information, highlighting the need to extract features" + } + ] + } + ], + "index": 16 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "type": "text", + "content": "15036" + } + ] + } + ], + "index": 17 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 3 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 59, + 70, + 293, + 198 + ], + "blocks": [ + { + "bbox": [ + 59, + 70, + 293, + 198 + ], + "lines": [ + { + "bbox": [ + 59, + 70, + 293, + 198 + ], + "spans": [ + { + "bbox": [ + 59, + 70, + 293, + 198 + ], + "type": "image", + "image_path": "be1e7e611f01088054fa7f9b0c142b13a3ad234b1a6a6d43920d50f02483aed8.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 55, + 200, + 295, + 257 + ], + "lines": [ + { + "bbox": [ + 55, + 200, + 295, + 257 + ], + "spans": [ + { + "bbox": [ + 55, + 200, + 295, + 257 + ], + "type": "text", + "content": "Figure 4. Visualization of the matched complementary graph pairs. The red dotted line indicates the focus areas of VFM, while the yellow dotted line highlights those for CLIP. By aggregating graphs with contrasting structural properties, we enhance the object-level context in " + }, + { + "bbox": [ + 55, + 200, + 295, + 257 + ], + "type": "inline_equation", + "content": "A_{\\mathrm{CLIP}}" + }, + { + "bbox": [ + 55, + 200, + 295, + 257 + ], + "type": "text", + "content": " supported by " + }, + { + "bbox": [ + 55, + 200, + 295, + 257 + ], + "type": "inline_equation", + "content": "\\ddot{A}_{\\mathrm{VFM}}" + }, + { + "bbox": [ + 55, + 200, + 295, + 257 + ], + "type": "text", + "content": ", resulting " + }, + { + "bbox": [ + 55, + 200, + 295, + 257 + ], + "type": "inline_equation", + "content": "A_{\\psi}" + }, + { + "bbox": [ + 55, + 200, + 295, + 257 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "bbox": [ + 55, + 265, + 296, + 361 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 265, + 296, + 361 + ], + "spans": [ + { + "bbox": [ + 55, + 265, + 296, + 361 + ], + "type": "text", + "content": "that emphasize object-level context. From this realization, we leverage the low-rank components of VFM, which contain distinct object patterns within the graph structure [12]. Specifically, we (I) extract the critical object-level contextual structure of " + }, + { + "bbox": [ + 55, + 265, + 296, + 361 + ], + "type": "inline_equation", + "content": "A_{\\mathrm{VFM}}^i" + }, + { + "bbox": [ + 55, + 265, + 296, + 361 + ], + "type": "text", + "content": " via low-rank approximation and enhance the graph structure by dynamically scaling eigenvalues. Then, we (II) distill the approximated VFM graph into CLIP by aggregating graphs based on spectral distance." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 55, + 363, + 295, + 543 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 363, + 295, + 543 + ], + "spans": [ + { + "bbox": [ + 55, + 363, + 295, + 543 + ], + "type": "text", + "content": "(I) Low-Rank Dynamic Eigenscaling. Our goal is to transfer the essential object-level contextual structure of VFM graph " + }, + { + "bbox": [ + 55, + 363, + 295, + 543 + ], + "type": "inline_equation", + "content": "A_{\\mathrm{VFM}}" + }, + { + "bbox": [ + 55, + 363, + 295, + 543 + ], + "type": "text", + "content": " to CLIP. To achieve this, we extract low-rank components of the VFM graph using standard eigendecomposition, as " + }, + { + "bbox": [ + 55, + 363, + 295, + 543 + ], + "type": "inline_equation", + "content": "A_{\\mathrm{VFM}}^i = U\\Sigma U^\\top" + }, + { + "bbox": [ + 55, + 363, + 295, + 543 + ], + "type": "text", + "content": ", where " + }, + { + "bbox": [ + 55, + 363, + 295, + 543 + ], + "type": "inline_equation", + "content": "U" + }, + { + "bbox": [ + 55, + 363, + 295, + 543 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 55, + 363, + 295, + 543 + ], + "type": "inline_equation", + "content": "\\Sigma" + }, + { + "bbox": [ + 55, + 363, + 295, + 543 + ], + "type": "text", + "content": " represent the eigenvector and the diagonal eigenvalue matrix, respectively. In the decomposed eigenbasis, we identify key object-level features of each graph by searching an optimal number of eigenvalues " + }, + { + "bbox": [ + 55, + 363, + 295, + 543 + ], + "type": "inline_equation", + "content": "k" + }, + { + "bbox": [ + 55, + 363, + 295, + 543 + ], + "type": "text", + "content": " through an energy-based approach. This ensures that the chosen " + }, + { + "bbox": [ + 55, + 363, + 295, + 543 + ], + "type": "inline_equation", + "content": "k" + }, + { + "bbox": [ + 55, + 363, + 295, + 543 + ], + "type": "text", + "content": " eigenvalues capture a significant portion of the graph's energy, retaining essential structural information while discarding noise and less relevant details. More details on searching " + }, + { + "bbox": [ + 55, + 363, + 295, + 543 + ], + "type": "inline_equation", + "content": "k" + }, + { + "bbox": [ + 55, + 363, + 295, + 543 + ], + "type": "text", + "content": " can be found in the supplementary material. This process yields a low-rank approximated VFM graph, expressed as " + }, + { + "bbox": [ + 55, + 363, + 295, + 543 + ], + "type": "inline_equation", + "content": "\\tilde{A}_{\\mathrm{VFM}}^i = U_k\\Sigma_kU_k^\\top" + }, + { + "bbox": [ + 55, + 363, + 295, + 543 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 55, + 543, + 295, + 651 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 543, + 295, + 651 + ], + "spans": [ + { + "bbox": [ + 55, + 543, + 295, + 651 + ], + "type": "text", + "content": "We refine the low-rank components with a scaling function " + }, + { + "bbox": [ + 55, + 543, + 295, + 651 + ], + "type": "inline_equation", + "content": "\\phi" + }, + { + "bbox": [ + 55, + 543, + 295, + 651 + ], + "type": "text", + "content": ", which dynamically amplifies larger eigenvalues and reduces smaller ones. Compared to the conventional shrinkage function [2, 5], which only focuses on noise cutoff, our approach emphasizes essential structural information, particularly object-level context features, while suppressing noise and irrelevant details. More details on dynamic eigenscaling can be found in the supplementary material. The graph after applying dynamic eigenscaling is as" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 129, + 654, + 295, + 669 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 129, + 654, + 295, + 669 + ], + "spans": [ + { + "bbox": [ + 129, + 654, + 295, + 669 + ], + "type": "interline_equation", + "content": "\\ddot {A} _ {\\mathrm {V F M}} ^ {i} = U _ {k} \\phi (\\Sigma_ {k}) U _ {k} ^ {\\top}. \\tag {6}", + "image_path": "7a3ae28d2b5fdb86f6e7312b7e4304bd17ab8853e5e1cb0d46554adff99b7936.jpg" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 55, + 674, + 296, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 674, + 296, + 715 + ], + "spans": [ + { + "bbox": [ + 55, + 674, + 296, + 715 + ], + "type": "text", + "content": "(II) VFM Graph Distillation. We distill the essential object-level contextual knowledge from the tailored VFM graph " + }, + { + "bbox": [ + 55, + 674, + 296, + 715 + ], + "type": "inline_equation", + "content": "\\ddot{A}_{\\mathrm{VFM}}^i" + }, + { + "bbox": [ + 55, + 674, + 296, + 715 + ], + "type": "text", + "content": " into the " + }, + { + "bbox": [ + 55, + 674, + 296, + 715 + ], + "type": "inline_equation", + "content": "A_{\\mathrm{CLIP}}^j" + }, + { + "bbox": [ + 55, + 674, + 296, + 715 + ], + "type": "text", + "content": " by aggregating their structures:" + } + ] + } + ], + "index": 6 + }, + { + "type": "image", + "bbox": [ + 316, + 69, + 553, + 213 + ], + "blocks": [ + { + "bbox": [ + 316, + 69, + 553, + 213 + ], + "lines": [ + { + "bbox": [ + 316, + 69, + 553, + 213 + ], + "spans": [ + { + "bbox": [ + 316, + 69, + 553, + 213 + ], + "type": "image", + "image_path": "8f30b67fc69581bdba2fef41e9589b525bac9ea247f49c6329795b8673f9185e.jpg" + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 313, + 215, + 555, + 291 + ], + "lines": [ + { + "bbox": [ + 313, + 215, + 555, + 291 + ], + "spans": [ + { + "bbox": [ + 313, + 215, + 555, + 291 + ], + "type": "text", + "content": "Figure 5. Detailed illustration of our object presence prior-guided text embedding adjustment module. The CLIP text encoder generates text embeddings for each object class, and the object presence prior is derived from both visual and text embeddings. Within hierarchically defined class groups, text embeddings are selected based on object presence prior, then refined in an object-specific direction to align with components likely present in the image." + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "image_caption" + } + ], + "index": 7 + }, + { + "bbox": [ + 356, + 300, + 553, + 317 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 356, + 300, + 553, + 317 + ], + "spans": [ + { + "bbox": [ + 356, + 300, + 553, + 317 + ], + "type": "interline_equation", + "content": "A _ {\\psi} ^ {j} = \\left(w _ {i j} \\ddot {A} _ {\\mathrm {V F M}} ^ {i} + A _ {\\mathrm {C L I P}} ^ {j}\\right) / \\left(w _ {i j} + 1\\right), \\tag {7}", + "image_path": "128e4755cade4352747fd3ac64f94e602720e89102910bfd8d8cd557b3e0e1d2.jpg" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 313, + 324, + 555, + 445 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 324, + 555, + 445 + ], + "spans": [ + { + "bbox": [ + 313, + 324, + 555, + 445 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 313, + 324, + 555, + 445 + ], + "type": "inline_equation", + "content": "w_{ij}" + }, + { + "bbox": [ + 313, + 324, + 555, + 445 + ], + "type": "text", + "content": " is defined as the Wasserstein distance between eigenvalues from both graphs as " + }, + { + "bbox": [ + 313, + 324, + 555, + 445 + ], + "type": "inline_equation", + "content": "\\mathcal{D}_{\\mathcal{W}}(\\bar{\\lambda}_{\\mathrm{VFM}}^i,\\bar{\\lambda}_{\\mathrm{CLIP}}^j)" + }, + { + "bbox": [ + 313, + 324, + 555, + 445 + ], + "type": "text", + "content": ". This adaptively assigns higher weights to graphs with complementary structures, while reducing influence on those with similar structures. Our distillation strategy effectively transfers the critical object-level context from VFM to CLIP, ensuring multiple components of the same object are grouped into a unified semantic as illustrated in Fig. 3 (" + }, + { + "bbox": [ + 313, + 324, + 555, + 445 + ], + "type": "inline_equation", + "content": "A_{\\psi}" + }, + { + "bbox": [ + 313, + 324, + 555, + 445 + ], + "type": "text", + "content": "). Therefore, we use " + }, + { + "bbox": [ + 313, + 324, + 555, + 445 + ], + "type": "inline_equation", + "content": "A_{\\psi}" + }, + { + "bbox": [ + 313, + 324, + 555, + 445 + ], + "type": "text", + "content": " as our attention matrix instead of " + }, + { + "bbox": [ + 313, + 324, + 555, + 445 + ], + "type": "inline_equation", + "content": "A_{\\mathrm{CLIP}}" + }, + { + "bbox": [ + 313, + 324, + 555, + 445 + ], + "type": "text", + "content": " in Eq. (2) to compute the final visual feature " + }, + { + "bbox": [ + 313, + 324, + 555, + 445 + ], + "type": "inline_equation", + "content": "F_{\\mathrm{CLIP}}" + }, + { + "bbox": [ + 313, + 324, + 555, + 445 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 313, + 453, + 552, + 467 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 453, + 552, + 467 + ], + "spans": [ + { + "bbox": [ + 313, + 453, + 552, + 467 + ], + "type": "text", + "content": "3.3. Object Presence-Driven Object-Level Context" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 313, + 471, + 555, + 662 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 471, + 555, + 662 + ], + "spans": [ + { + "bbox": [ + 313, + 471, + 555, + 662 + ], + "type": "text", + "content": "The spectral object-level context distillation in Sec. 3.2 enables precise object mask generation. However, due to the nature of OVSS, where users can provide arbitrary query prompts, different parts of the same object may still be assigned to closely related categories. Meanwhile, numerous studies [20, 34, 39] have demonstrated that CLIP excels in zero-shot object classification capability. Accordingly, we utilize the zero-shot object classification score encoded by CLIP (i.e., the object presence prior) to refine text embeddings and patch-text similarity, enhancing object-centered perspective. To compute the object presence prior " + }, + { + "bbox": [ + 313, + 471, + 555, + 662 + ], + "type": "inline_equation", + "content": "P" + }, + { + "bbox": [ + 313, + 471, + 555, + 662 + ], + "type": "text", + "content": ", we take an entire input image " + }, + { + "bbox": [ + 313, + 471, + 555, + 662 + ], + "type": "inline_equation", + "content": "\\mathcal{I}" + }, + { + "bbox": [ + 313, + 471, + 555, + 662 + ], + "type": "text", + "content": " and compute likelihood for each class " + }, + { + "bbox": [ + 313, + 471, + 555, + 662 + ], + "type": "inline_equation", + "content": "i" + }, + { + "bbox": [ + 313, + 471, + 555, + 662 + ], + "type": "text", + "content": " as " + }, + { + "bbox": [ + 313, + 471, + 555, + 662 + ], + "type": "inline_equation", + "content": "P(i) = \\{t_{\\mathrm{CLIP}}^i\\}_{i=1}^C v_{\\mathrm{CLIP}}" + }, + { + "bbox": [ + 313, + 471, + 555, + 662 + ], + "type": "text", + "content": ", where the visual embedding vector " + }, + { + "bbox": [ + 313, + 471, + 555, + 662 + ], + "type": "inline_equation", + "content": "v_{\\mathrm{CLIP}} = \\mathcal{F}_{\\mathrm{CLIP}}^v(\\mathcal{I}) \\in \\mathbb{R}^d" + }, + { + "bbox": [ + 313, + 471, + 555, + 662 + ], + "type": "text", + "content": " is derived from the ViT [CLS] token. The object presence prior is used in the following sections to enhance object-level context." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 313, + 665, + 555, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 665, + 555, + 714 + ], + "spans": [ + { + "bbox": [ + 313, + 665, + 555, + 714 + ], + "type": "text", + "content": "Object-Guided Text Embedding Adjustment. We adjust text embeddings " + }, + { + "bbox": [ + 313, + 665, + 555, + 714 + ], + "type": "inline_equation", + "content": "\\{t_{\\mathrm{CLIP}}^i\\}_{i = 1}^C" + }, + { + "bbox": [ + 313, + 665, + 555, + 714 + ], + "type": "text", + "content": " to embed object-specific contextual semantics, as shown in Fig. 5. To prevent object parts from being assigned to ambiguous classes, we" + } + ] + } + ], + "index": 13 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "text", + "content": "15037" + } + ] + } + ], + "index": 14 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 4 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 58, + 85, + 553, + 312 + ], + "blocks": [ + { + "bbox": [ + 71, + 71, + 537, + 82 + ], + "lines": [ + { + "bbox": [ + 71, + 71, + 537, + 82 + ], + "spans": [ + { + "bbox": [ + 71, + 71, + 537, + 82 + ], + "type": "text", + "content": "Table 1. Quantitative results with state-of-the-art unsupervised open-vocabulary semantic segmentation models on eight datasets." + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 58, + 85, + 553, + 312 + ], + "lines": [ + { + "bbox": [ + 58, + 85, + 553, + 312 + ], + "spans": [ + { + "bbox": [ + 58, + 85, + 553, + 312 + ], + "type": "table", + "html": "
ModelSupporting DatasetExtra-TrainingFairV21PC60C-ObjV20PC59C-StfCityADEAvg.
GroupViT [54]CVPR'22CC12M+RedCaps✓✘50.418.727.579.723.415.311.19.229.4
TCL [9]CVPR'23CC3M+CC12M✓✘55.030.431.683.233.922.424.017.137.2
CoDe [51]CVPR'24CC3M+RedCaps✓✘57.530.532.3--23.928.917.7-
CLIP-DINOiser [52]ECCV'24ImageNet1k✓✘62.132.434.880.935.924.631.720.040.3
ReCo [45]NeurIPS'22ImageNet1k✘✘25.119.915.757.722.314.821.611.223.5
FOSSIL [3]WACV'24COCO Captions✘✘----35.824.823.218.8-
FreeDa [4]CVPR'24COCO Captions✘✘---85.643.127.836.722.4-
CLIP [39]ICML'21✘✘✓18.67.86.549.111.27.26.73.213.8
MaskCLIP [58]ECCV'22✘✘✓38.323.620.674.926.416.412.69.827.9
GEM [6]CVPR'24✘✘✓46.2---32.6--15.7-
CaR [47]CVPR'24✘✘✓48.613.615.473.718.4--5.4-
PnP-OVSS [33]CVPR'24✘✘✓--36.251.328.017.9-14.2-
CLIPtrase [44]ECCV'24✘✘✓50.929.943.681.033.822.821.316.432.7
ClearCLIP [28]ECCV'24✘✘✓51.832.633.080.935.923.930.016.738.1
SCLIP [48]ECCV'24✘✘✓59.130.430.580.434.122.432.216.138.2
LaVG [21]ECCV'24✘✘✓62.131.634.282.534.723.226.215.838.8
ProxyCLIP [29]ECCV'24✘✘✓59.135.236.278.238.826.238.119.641.4
NACLIP [16]WACV'25✘✘✓58.932.233.279.735.223.335.517.439.4
CASS✘✘✓65.836.737.887.840.226.739.420.444.4
", + "image_path": "544419de3792c240900a3899d7fb6fb49c78e568c4fb79791cc62e99671a5402.jpg" + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "table_body" + } + ], + "index": 1 + }, + { + "bbox": [ + 55, + 323, + 296, + 529 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 323, + 296, + 529 + ], + "spans": [ + { + "bbox": [ + 55, + 323, + 296, + 529 + ], + "type": "text", + "content": "The hierarchical clustering of CLIP semantic distance, forming groups of similar embeddings " + }, + { + "bbox": [ + 55, + 323, + 296, + 529 + ], + "type": "inline_equation", + "content": "\\mathcal{T}_g = \\{t_{\\mathrm{CLIP}}^i\\mid i\\in I_g\\}" + }, + { + "bbox": [ + 55, + 323, + 296, + 529 + ], + "type": "text", + "content": ", where " + }, + { + "bbox": [ + 55, + 323, + 296, + 529 + ], + "type": "inline_equation", + "content": "\\mathcal{T}_g" + }, + { + "bbox": [ + 55, + 323, + 296, + 529 + ], + "type": "text", + "content": " is a subset of entire text embeddings " + }, + { + "bbox": [ + 55, + 323, + 296, + 529 + ], + "type": "inline_equation", + "content": "\\{t_{\\mathrm{CLIP}}^i\\}_{i = 1}^C" + }, + { + "bbox": [ + 55, + 323, + 296, + 529 + ], + "type": "text", + "content": " with index set " + }, + { + "bbox": [ + 55, + 323, + 296, + 529 + ], + "type": "inline_equation", + "content": "I_{g}" + }, + { + "bbox": [ + 55, + 323, + 296, + 529 + ], + "type": "text", + "content": ". Further details about this clustering process are provided in the supplementary materials. Within " + }, + { + "bbox": [ + 55, + 323, + 296, + 529 + ], + "type": "inline_equation", + "content": "\\mathcal{T}_g" + }, + { + "bbox": [ + 55, + 323, + 296, + 529 + ], + "type": "text", + "content": ", we identify the object most likely to appear in the image using object presence prior " + }, + { + "bbox": [ + 55, + 323, + 296, + 529 + ], + "type": "inline_equation", + "content": "P" + }, + { + "bbox": [ + 55, + 323, + 296, + 529 + ], + "type": "text", + "content": ", formulated as " + }, + { + "bbox": [ + 55, + 323, + 296, + 529 + ], + "type": "inline_equation", + "content": "i^{*} = \\arg \\max_{i\\in I_{g}}P(i)" + }, + { + "bbox": [ + 55, + 323, + 296, + 529 + ], + "type": "text", + "content": ". This facilitates refining the text embedding vector " + }, + { + "bbox": [ + 55, + 323, + 296, + 529 + ], + "type": "inline_equation", + "content": "t_{\\mathrm{CLIP}}^*" + }, + { + "bbox": [ + 55, + 323, + 296, + 529 + ], + "type": "text", + "content": " to better represent object-contextual semantics. We note that patch-wise visual feature " + }, + { + "bbox": [ + 55, + 323, + 296, + 529 + ], + "type": "inline_equation", + "content": "F_{\\mathrm{CLIP}}\\in \\mathbb{R}^{N\\times d}" + }, + { + "bbox": [ + 55, + 323, + 296, + 529 + ], + "type": "text", + "content": " contains object-specific semantic vectors that can guide " + }, + { + "bbox": [ + 55, + 323, + 296, + 529 + ], + "type": "inline_equation", + "content": "t_{\\mathrm{CLIP}}^*" + }, + { + "bbox": [ + 55, + 323, + 296, + 529 + ], + "type": "text", + "content": " accordingly. Thus, we compute cosine similarity between " + }, + { + "bbox": [ + 55, + 323, + 296, + 529 + ], + "type": "inline_equation", + "content": "F_{\\mathrm{CLIP}}" + }, + { + "bbox": [ + 55, + 323, + 296, + 529 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 55, + 323, + 296, + 529 + ], + "type": "inline_equation", + "content": "t_{\\mathrm{CLIP}}^*" + }, + { + "bbox": [ + 55, + 323, + 296, + 529 + ], + "type": "text", + "content": " and select top-" + }, + { + "bbox": [ + 55, + 323, + 296, + 529 + ], + "type": "inline_equation", + "content": "n" + }, + { + "bbox": [ + 55, + 323, + 296, + 529 + ], + "type": "text", + "content": " object-specific image vector denoted as " + }, + { + "bbox": [ + 55, + 323, + 296, + 529 + ], + "type": "inline_equation", + "content": "\\{f_i\\}_{i = 1}^n\\coloneqq" + }, + { + "bbox": [ + 55, + 323, + 296, + 529 + ], + "type": "inline_equation", + "content": "\\mathrm{Concat}[f_1,\\ldots ,f_n]\\in \\mathbb{R}^{n\\times d}" + }, + { + "bbox": [ + 55, + 323, + 296, + 529 + ], + "type": "text", + "content": ". To define a targeted direction, we compute the average of " + }, + { + "bbox": [ + 55, + 323, + 296, + 529 + ], + "type": "inline_equation", + "content": "\\{f_i\\}_{i = 1}^n" + }, + { + "bbox": [ + 55, + 323, + 296, + 529 + ], + "type": "text", + "content": " as " + }, + { + "bbox": [ + 55, + 323, + 296, + 529 + ], + "type": "inline_equation", + "content": "\\mu_{n}\\in \\mathbb{R}^{d}" + }, + { + "bbox": [ + 55, + 323, + 296, + 529 + ], + "type": "text", + "content": ", and guide selected text embedding " + }, + { + "bbox": [ + 55, + 323, + 296, + 529 + ], + "type": "inline_equation", + "content": "t_{\\mathrm{CLIP}}^*" + }, + { + "bbox": [ + 55, + 323, + 296, + 529 + ], + "type": "text", + "content": " towards " + }, + { + "bbox": [ + 55, + 323, + 296, + 529 + ], + "type": "inline_equation", + "content": "\\mu_{n}" + }, + { + "bbox": [ + 55, + 323, + 296, + 529 + ], + "type": "text", + "content": " as" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 108, + 530, + 294, + 545 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 108, + 530, + 294, + 545 + ], + "spans": [ + { + "bbox": [ + 108, + 530, + 294, + 545 + ], + "type": "interline_equation", + "content": "\\tilde {t} _ {\\mathrm {C L I P}} ^ {i ^ {*}} = (1 - \\alpha) \\cdot t _ {\\mathrm {C L I P}} ^ {i ^ {*}} + \\alpha \\cdot \\mu_ {n}, \\tag {8}", + "image_path": "bf5630ed57a4cb2fde5ca7296d503bb2f3d2f781bd162a6f33e5e32f28f73ff9.jpg" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 54, + 550, + 295, + 602 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 54, + 550, + 295, + 602 + ], + "spans": [ + { + "bbox": [ + 54, + 550, + 295, + 602 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 54, + 550, + 295, + 602 + ], + "type": "inline_equation", + "content": "\\alpha" + }, + { + "bbox": [ + 54, + 550, + 295, + 602 + ], + "type": "text", + "content": " controls the degree of alignment with " + }, + { + "bbox": [ + 54, + 550, + 295, + 602 + ], + "type": "inline_equation", + "content": "\\mu_{n}" + }, + { + "bbox": [ + 54, + 550, + 295, + 602 + ], + "type": "text", + "content": ". The adjusted text embedding " + }, + { + "bbox": [ + 54, + 550, + 295, + 602 + ], + "type": "inline_equation", + "content": "\\tilde{t}_{\\mathrm{CLIP}}^{i^*}" + }, + { + "bbox": [ + 54, + 550, + 295, + 602 + ], + "type": "text", + "content": " replaces the original text embedding " + }, + { + "bbox": [ + 54, + 550, + 295, + 602 + ], + "type": "inline_equation", + "content": "t_{\\mathrm{CLIP}}^{i}" + }, + { + "bbox": [ + 54, + 550, + 295, + 602 + ], + "type": "text", + "content": " in the set " + }, + { + "bbox": [ + 54, + 550, + 295, + 602 + ], + "type": "inline_equation", + "content": "\\{t_{\\mathrm{CLIP}}^{i}\\}_{i=1}^{C}" + }, + { + "bbox": [ + 54, + 550, + 295, + 602 + ], + "type": "text", + "content": ", which is then used to compute the patch-text similarity logit " + }, + { + "bbox": [ + 54, + 550, + 295, + 602 + ], + "type": "inline_equation", + "content": "\\hat{S}" + }, + { + "bbox": [ + 54, + 550, + 295, + 602 + ], + "type": "text", + "content": " in Eq. (4)." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 54, + 605, + 295, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 54, + 605, + 295, + 714 + ], + "spans": [ + { + "bbox": [ + 54, + 605, + 295, + 714 + ], + "type": "text", + "content": "Object Perspective Patch-Text Similarity. We integrate the object presence prior with patch-text similarity " + }, + { + "bbox": [ + 54, + 605, + 295, + 714 + ], + "type": "inline_equation", + "content": "\\hat{S}" + }, + { + "bbox": [ + 54, + 605, + 295, + 714 + ], + "type": "text", + "content": " from Eq. (4) to produce a segmentation map from an object-perspective view. Unlike Barsellotti et al. [4], which computes likelihood from an image window " + }, + { + "bbox": [ + 54, + 605, + 295, + 714 + ], + "type": "inline_equation", + "content": "\\hat{\\mathcal{I}}" + }, + { + "bbox": [ + 54, + 605, + 295, + 714 + ], + "type": "text", + "content": ", our approach derives similarity from the full image " + }, + { + "bbox": [ + 54, + 605, + 295, + 714 + ], + "type": "inline_equation", + "content": "\\mathcal{I}" + }, + { + "bbox": [ + 54, + 605, + 295, + 714 + ], + "type": "text", + "content": ", enabling richer object-contextual understanding at the broader level. Thus, we refine the patch-text similarity " + }, + { + "bbox": [ + 54, + 605, + 295, + 714 + ], + "type": "inline_equation", + "content": "\\hat{S}" + }, + { + "bbox": [ + 54, + 605, + 295, + 714 + ], + "type": "text", + "content": " from Eq. (4) to obtain the final patch-text similarity " + }, + { + "bbox": [ + 54, + 605, + 295, + 714 + ], + "type": "inline_equation", + "content": "\\hat{S}^*" + }, + { + "bbox": [ + 54, + 605, + 295, + 714 + ], + "type": "text", + "content": ", computed as" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 350, + 320, + 553, + 338 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 350, + 320, + 553, + 338 + ], + "spans": [ + { + "bbox": [ + 350, + 320, + 553, + 338 + ], + "type": "interline_equation", + "content": "\\hat {\\mathcal {S}} ^ {*} = (1 - \\gamma) \\cdot \\hat {\\mathcal {S}} + \\gamma \\cdot \\left\\{t _ {\\mathrm {C L I P}} ^ {i} \\right\\} _ {i = 1} ^ {C} v _ {\\mathrm {C L I P}}, \\tag {9}", + "image_path": "f1ee8faa3b8d803513ea429700ccc1103e6f019857b5b679d7715588c6ffee85.jpg" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 313, + 340, + 553, + 376 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 340, + 553, + 376 + ], + "spans": [ + { + "bbox": [ + 313, + 340, + 553, + 376 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 313, + 340, + 553, + 376 + ], + "type": "inline_equation", + "content": "\\gamma" + }, + { + "bbox": [ + 313, + 340, + 553, + 376 + ], + "type": "text", + "content": " controls the balance between " + }, + { + "bbox": [ + 313, + 340, + 553, + 376 + ], + "type": "inline_equation", + "content": "\\hat{S}" + }, + { + "bbox": [ + 313, + 340, + 553, + 376 + ], + "type": "text", + "content": " and object presence prior. This allows the final segmentation map to capture object-level details for precise, class-aware segmentation." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 313, + 380, + 395, + 394 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 380, + 395, + 394 + ], + "spans": [ + { + "bbox": [ + 313, + 380, + 395, + 394 + ], + "type": "text", + "content": "4. Experiments" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 313, + 400, + 554, + 473 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 400, + 554, + 473 + ], + "spans": [ + { + "bbox": [ + 313, + 400, + 554, + 473 + ], + "type": "text", + "content": "We first describe the experimental settings, including implementation details, datasets, and evaluation metrics (Sec. 4.1). We then present quantitative and qualitative evaluation results (Sec. 4.2) and demonstrate ablation studies to assess each proposed component (Sec. 4.3). Additional experiment results are provided in the supplementary material." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 313, + 479, + 441, + 491 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 479, + 441, + 491 + ], + "spans": [ + { + "bbox": [ + 313, + 479, + 441, + 491 + ], + "type": "text", + "content": "4.1. Experimental Settings" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 313, + 494, + 555, + 601 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 494, + 555, + 601 + ], + "spans": [ + { + "bbox": [ + 313, + 494, + 555, + 601 + ], + "type": "text", + "content": "Implementation Details. Similar to existing training-free approaches [16, 21, 29, 44, 48], we employ CLIP ViT-B/16 [39] as our VLM and DINO ViT-B/8 [8] as our VFM, keeping both frozen. Input images are matched with the shorter side set to 336 pixels (or 560 pixels for the higher-resolution Cityscapes dataset), and sliding window inference is applied with a " + }, + { + "bbox": [ + 313, + 494, + 555, + 601 + ], + "type": "inline_equation", + "content": "224 \\times 224" + }, + { + "bbox": [ + 313, + 494, + 555, + 601 + ], + "type": "text", + "content": " window and a stride of 112 to achieve efficient evaluation coverage. All experiments are conducted using NVIDIA RTX A6000 Ada*." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 313, + 601, + 554, + 696 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 601, + 554, + 696 + ], + "spans": [ + { + "bbox": [ + 313, + 601, + 554, + 696 + ], + "type": "text", + "content": "Datasets. Our primary datasets include PASCAL VOC 2012 [15], PASCAL Context [37], and COCO [7], each evaluated with and without a background class. We refer to the versions with a background class as V21, PC60, and C-Obj, and those without as V20, PC59, and C-Stf. We also report results on the ADE20K (ADE) [57] and Cityscapes (City) [13] datasets. These datasets contain 21, 60, 81, 20, 59, 171, 150, and 19 classes, respectively." + } + ] + } + ], + "index": 12 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 330, + 703, + 529, + 712 + ], + "type": "page_footnote", + "angle": 0, + "lines": [ + { + "bbox": [ + 330, + 703, + 529, + 712 + ], + "spans": [ + { + "bbox": [ + 330, + 703, + 529, + 712 + ], + "type": "text", + "content": "* Advanced Database System Infrastructure (NFEC-2024-11-300458)" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "text", + "content": "15038" + } + ] + } + ], + "index": 14 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 5 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 56, + 69, + 553, + 255 + ], + "blocks": [ + { + "bbox": [ + 56, + 69, + 553, + 255 + ], + "lines": [ + { + "bbox": [ + 56, + 69, + 553, + 255 + ], + "spans": [ + { + "bbox": [ + 56, + 69, + 553, + 255 + ], + "type": "image", + "image_path": "6acc002ec693f900363bf04759e497b548ae11fa32e1c9b1b3103d66c2b3f462.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 55, + 255, + 553, + 277 + ], + "lines": [ + { + "bbox": [ + 55, + 255, + 553, + 277 + ], + "spans": [ + { + "bbox": [ + 55, + 255, + 553, + 277 + ], + "type": "text", + "content": "Figure 6. Qualitative comparison across the Pascal VOC [15], Pascal Context [37], COCO [7], and ADE20K [57] datasets using CLIP ViT-B/16 [39], evaluating previous state-of-the-art training-free OVSS methods, including NACLIP [16], ProxyCLIP [29], and ours." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "bbox": [ + 55, + 285, + 167, + 297 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 285, + 167, + 297 + ], + "spans": [ + { + "bbox": [ + 55, + 285, + 167, + 297 + ], + "type": "text", + "content": "4.2. Evaluation Results" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 55, + 301, + 295, + 386 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 301, + 295, + 386 + ], + "spans": [ + { + "bbox": [ + 55, + 301, + 295, + 386 + ], + "type": "text", + "content": "We evaluate our model against training-free OVSS methods both quantitatively and qualitatively. For fair comparison, we refrain from using the mask refinement (e.g., PAMR [1] or DenseCRF [26]) and reproduce results from [29] using a " + }, + { + "bbox": [ + 55, + 301, + 295, + 386 + ], + "type": "inline_equation", + "content": "224 \\times 224" + }, + { + "bbox": [ + 55, + 301, + 295, + 386 + ], + "type": "text", + "content": " window to match the baselines. Performance is measured primarily with mean IoU (mIoU), complemented by pixel accuracy (pAcc) for a broader assessment." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 55, + 387, + 295, + 555 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 387, + 295, + 555 + ], + "spans": [ + { + "bbox": [ + 55, + 387, + 295, + 555 + ], + "type": "text", + "content": "Quantitative Evaluation: mIoU. Table 1 presents the main results in mIoU, where all results are obtained using CLIP ViT-B/16 as the backbone model for fair comparisons. To demonstrate the strengths of our model, we also include methods that use supporting datasets (i.e., retrieving additional datasets during inference) or additional training (models evaluated under the same setting as ours are indicated with " + }, + { + "bbox": [ + 55, + 387, + 295, + 555 + ], + "type": "inline_equation", + "content": "\\checkmark" + }, + { + "bbox": [ + 55, + 387, + 295, + 555 + ], + "type": "text", + "content": " in the \"Fair\" column). Our model, CASS, performs favorably against the state-of-the-art methods, with an average gain of " + }, + { + "bbox": [ + 55, + 387, + 295, + 555 + ], + "type": "inline_equation", + "content": "3.0\\mathrm{mIoU}" + }, + { + "bbox": [ + 55, + 387, + 295, + 555 + ], + "type": "text", + "content": " point across eight datasets. Notably, for the V20, CASS achieves a performance gain of 5.3 mIoU point over the second-best model. Further, CASS outperforms CLIP-DINOiser [52] (which also leverages DINO with extra data and training) by " + }, + { + "bbox": [ + 55, + 387, + 295, + 555 + ], + "type": "inline_equation", + "content": "4.1\\mathrm{mIoU}" + }, + { + "bbox": [ + 55, + 387, + 295, + 555 + ], + "type": "text", + "content": " point." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 55, + 555, + 295, + 652 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 555, + 295, + 652 + ], + "spans": [ + { + "bbox": [ + 55, + 555, + 295, + 652 + ], + "type": "text", + "content": "Quantitative Evaluation: pAcc. Table 2 shows the results of pAcc on recent state-of-the-art models evaluated under the same settings. Our CASS outperforms the second-highest model by " + }, + { + "bbox": [ + 55, + 555, + 295, + 652 + ], + "type": "inline_equation", + "content": "2.2\\mathrm{pAcc}" + }, + { + "bbox": [ + 55, + 555, + 295, + 652 + ], + "type": "text", + "content": " point and exceeds one of the most recent papers, CLIPtrase [44], by " + }, + { + "bbox": [ + 55, + 555, + 295, + 652 + ], + "type": "inline_equation", + "content": "9.8\\mathrm{pAcc}" + }, + { + "bbox": [ + 55, + 555, + 295, + 652 + ], + "type": "text", + "content": " point. Our model consistently performs better than existing methods on mIoU and pAcc by effectively capturing object-level context, resulting in more accurate object masking." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 55, + 653, + 295, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 653, + 295, + 714 + ], + "spans": [ + { + "bbox": [ + 55, + 653, + 295, + 714 + ], + "type": "text", + "content": "Qualitative Analysis. Fig. 6 provides a qualitative comparison with recent state-of-the-art models [16, 29]. Existing methods often produce noisy segmentation maps and do not group object components correctly, such as wheels within a car or a person's head or arm. Additionally, they often mis" + } + ] + } + ], + "index": 6 + }, + { + "type": "table", + "bbox": [ + 317, + 296, + 552, + 366 + ], + "blocks": [ + { + "bbox": [ + 328, + 285, + 539, + 296 + ], + "lines": [ + { + "bbox": [ + 328, + 285, + 539, + 296 + ], + "spans": [ + { + "bbox": [ + 328, + 285, + 539, + 296 + ], + "type": "text", + "content": "Table 2. Quantitative results using average pixel accuracy." + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 317, + 296, + 552, + 366 + ], + "lines": [ + { + "bbox": [ + 317, + 296, + 552, + 366 + ], + "spans": [ + { + "bbox": [ + 317, + 296, + 552, + 366 + ], + "type": "table", + "html": "
ModelV21PC60C-ObjV20PC59C-StfCityADEAvg.
CLIPtrase [44]78.652.150.189.758.938.963.438.659.1
LaVG [21]89.348.774.891.158.939.168.537.063.4
SCLIP [48]87.649.274.391.058.338.472.738.763.8
ProxyCLIP [29]86.652.075.988.463.443.474.949.166.7
NACLIP [16]87.151.275.389.259.839.371.745.264.9
CASS90.155.676.093.965.043.678.248.668.9
", + "image_path": "387ac760cd43f0c155b29e003b31e54b744c2dc49d3f7f0713312a480b036af7.jpg" + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "table_body" + } + ], + "index": 8 + }, + { + "bbox": [ + 313, + 374, + 554, + 447 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 374, + 554, + 447 + ], + "spans": [ + { + "bbox": [ + 313, + 374, + 554, + 447 + ], + "type": "text", + "content": "classify object parts into ambiguous classes (e.g., a motorcycle as a bicycle or an airplane as a car). In contrast, based on a deep understanding of object-level context, our CASS generates clean segmentation maps that accurately group all object components, including slender parts such as human arms or animal legs, into their respective object classes." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 313, + 454, + 408, + 467 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 454, + 408, + 467 + ], + "spans": [ + { + "bbox": [ + 313, + 454, + 408, + 467 + ], + "type": "text", + "content": "4.3. Ablation Study" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 313, + 470, + 554, + 530 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 470, + 554, + 530 + ], + "spans": [ + { + "bbox": [ + 313, + 470, + 554, + 530 + ], + "type": "text", + "content": "We perform ablation studies to analyze the proposed method. We use the datasets representative of segmentation tasks for all experiments: V21, PC59, C-Stf, following [48]. Please refer to the supplementary material for the additional ablation studies including VFM backbones and feature type." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 313, + 533, + 555, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 533, + 555, + 715 + ], + "spans": [ + { + "bbox": [ + 313, + 533, + 555, + 715 + ], + "type": "text", + "content": "Effect of Spectral VFM Distillation. Table 3 shows the results of incorporating VFM (Exp.#3) with the vanilla setting (Exp.#1), which directly uses the output from Eq. (4). In Exp.#3, " + }, + { + "bbox": [ + 313, + 533, + 555, + 715 + ], + "type": "inline_equation", + "content": "A_{\\mathrm{VFM}}" + }, + { + "bbox": [ + 313, + 533, + 555, + 715 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 313, + 533, + 555, + 715 + ], + "type": "inline_equation", + "content": "A_{\\mathrm{CLIP}}" + }, + { + "bbox": [ + 313, + 533, + 555, + 715 + ], + "type": "text", + "content": " are aggregated in a straightforward head-wise manner, without any additional processing to " + }, + { + "bbox": [ + 313, + 533, + 555, + 715 + ], + "type": "inline_equation", + "content": "A_{\\mathrm{VFM}}" + }, + { + "bbox": [ + 313, + 533, + 555, + 715 + ], + "type": "text", + "content": ". Simply leveraging " + }, + { + "bbox": [ + 313, + 533, + 555, + 715 + ], + "type": "inline_equation", + "content": "A_{\\mathrm{VFM}}" + }, + { + "bbox": [ + 313, + 533, + 555, + 715 + ], + "type": "text", + "content": " leads to a performance gain compared to the vanilla setting. Exp.#4 to Exp.#6 validate the effectiveness of the proposed components (see Sec. 3.2). Exp.#4 demonstrates the aggregation of " + }, + { + "bbox": [ + 313, + 533, + 555, + 715 + ], + "type": "inline_equation", + "content": "A_{\\mathrm{VFM}}" + }, + { + "bbox": [ + 313, + 533, + 555, + 715 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 313, + 533, + 555, + 715 + ], + "type": "inline_equation", + "content": "A_{\\mathrm{CLIP}}" + }, + { + "bbox": [ + 313, + 533, + 555, + 715 + ], + "type": "text", + "content": " using a complementary spectral graph matching (GM) procedure. Exp.#5, denoted as LR, presents the results with only the low-rank approximation applied to " + }, + { + "bbox": [ + 313, + 533, + 555, + 715 + ], + "type": "inline_equation", + "content": "A_{\\mathrm{VFM}}" + }, + { + "bbox": [ + 313, + 533, + 555, + 715 + ], + "type": "text", + "content": ", while Exp.#6 shows the results with both the low-rank approximation and Dynamic Eigenscaling (DE) applied, as defined in Eq. (6). Notably, Exp.#6 shows signifi" + } + ] + } + ], + "index": 12 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "text", + "content": "15039" + } + ] + } + ], + "index": 13 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 6 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 58, + 107, + 293, + 212 + ], + "blocks": [ + { + "bbox": [ + 55, + 71, + 295, + 103 + ], + "lines": [ + { + "bbox": [ + 55, + 71, + 295, + 103 + ], + "spans": [ + { + "bbox": [ + 55, + 71, + 295, + 103 + ], + "type": "text", + "content": "Table 3. Ablation results showing the impact of our components on the performance across V21 [15], PC59 [37], and C-Stf [7]. The abbreviations in the table are noted in Sec. 4.3." + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 58, + 107, + 293, + 212 + ], + "lines": [ + { + "bbox": [ + 58, + 107, + 293, + 212 + ], + "spans": [ + { + "bbox": [ + 58, + 107, + 293, + 212 + ], + "type": "table", + "html": "
Exp. #AVFMGMLRDEOPSOTADataset
V21PC59C-Stf
158.935.323.4
2✓✓59.236.524.0
3✓62.937.725.2
4✓✓63.238.025.4
5✓✓✓64.538.425.8
6✓✓✓✓64.838.625.9
7✓✓✓✓✓65.039.426.6
8✓✓✓✓✓✓65.840.226.7
", + "image_path": "117059a025c63c40c71ee0f77436030f8d5d517b981628234091170cfabd3e95.jpg" + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "table_body" + } + ], + "index": 1 + }, + { + "type": "image", + "bbox": [ + 58, + 214, + 294, + 304 + ], + "blocks": [ + { + "bbox": [ + 58, + 214, + 294, + 304 + ], + "lines": [ + { + "bbox": [ + 58, + 214, + 294, + 304 + ], + "spans": [ + { + "bbox": [ + 58, + 214, + 294, + 304 + ], + "type": "image", + "image_path": "89aa33d1d8b0c648e3138465c94d23ae99fe508f53f1e024a70cabf0b63829df.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 55, + 306, + 295, + 350 + ], + "lines": [ + { + "bbox": [ + 55, + 306, + 295, + 350 + ], + "spans": [ + { + "bbox": [ + 55, + 306, + 295, + 350 + ], + "type": "text", + "content": "Figure 7. Qualitative ablation of our method. Exp.# refers to the experiment in Table 3. As additional components of our proposed method are incorporated, the segmentation map exhibits greater refinement and more effectively captures object-level context." + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_caption" + } + ], + "index": 2 + }, + { + "bbox": [ + 54, + 357, + 295, + 465 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 54, + 357, + 295, + 465 + ], + "spans": [ + { + "bbox": [ + 54, + 357, + 295, + 465 + ], + "type": "text", + "content": "cant performance gain over simply integrating VFM knowledge (Exp.#3), demonstrating the effectiveness of graph matching and low-rank approximation in capturing object-level context. Fig. 3 qualitatively visualizes the attention map, with the left, middle, and right columns corresponding to Exp.#3, Exp.#4, and Exp.#6, respectively. We observe that integrating proposed low-rank dynamic eigenscaling effectively reduces noise and emphasizes objects by highlighting objects with strong inter-object associations." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 54, + 468, + 295, + 588 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 54, + 468, + 295, + 588 + ], + "spans": [ + { + "bbox": [ + 54, + 468, + 295, + 588 + ], + "type": "text", + "content": "Effect of Object Presence Prior. We validate the effectiveness of the object presence prior-driven methods in Exp.#7 and Exp.#8 by examining each component: Object-Perspective Patch-Text Similarity (OPS) and Object-Guided Text Embedding Adjustment (OTA) (see Sec. 3.3). These results show that each component effectively improves performance, emphasizing object-level context. Additionally, incorporating the object presence prior-driven methods on top of the vanilla setting, as shown in Exp.#2, clarifies the effectiveness of the proposed method." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 54, + 590, + 295, + 698 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 54, + 590, + 295, + 698 + ], + "spans": [ + { + "bbox": [ + 54, + 590, + 295, + 698 + ], + "type": "text", + "content": "Qualitative Ablations. Fig. 7 shows the segmentation maps for the vanilla setting (Exp.#1), after applying spectral VFM distillation (Exp.#6), and with all proposed components applied (Exp.#8). While Exp.#6 improves object masking compared to Exp.#1, it often misses parts of objects (highlighted by the red dotted line). In contrast, incorporating all proposed components (Exp.#8) consolidates objects into single entities, effectively capturing object-level context in the segmentation mask." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 55, + 701, + 295, + 712 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 701, + 295, + 712 + ], + "spans": [ + { + "bbox": [ + 55, + 701, + 295, + 712 + ], + "type": "text", + "content": "Different CLIP Backbones. Table 4 shows results on dif-" + } + ] + } + ], + "index": 7 + }, + { + "type": "table", + "bbox": [ + 316, + 81, + 553, + 180 + ], + "blocks": [ + { + "bbox": [ + 334, + 71, + 533, + 80 + ], + "lines": [ + { + "bbox": [ + 334, + 71, + 533, + 80 + ], + "spans": [ + { + "bbox": [ + 334, + 71, + 533, + 80 + ], + "type": "text", + "content": "Table 4. Ablation results on different CLIP backbones." + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 316, + 81, + 553, + 180 + ], + "lines": [ + { + "bbox": [ + 316, + 81, + 553, + 180 + ], + "spans": [ + { + "bbox": [ + 316, + 81, + 553, + 180 + ], + "type": "table", + "html": "
ModelV21PC59C-StfAvg. mIoUAvg. pAcc
mIoUpAccmIoUpAccmIoUpAcc
VLT-B/32SCLIP [48]50.681.028.749.620.033.433.154.7
LaVG [21]54.884.429.050.120.533.734.856.1
ProxyCLIP [29]57.985.635.259.723.640.238.961.8
NACLIP [16]51.182.832.456.821.237.134.958.9
CASS58.286.236.561.024.440.839.762.7
VLT-L/14SCLIP [48]44.478.125.246.617.629.629.151.4
LaVG [21]51.584.127.550.319.432.332.855.6
ProxyCLIP [29]59.885.738.361.026.243.541.463.4
NACLIP [16]52.283.132.152.821.435.735.257.2
CASS62.188.039.162.126.343.442.564.5
", + "image_path": "2ef8e1ca42d3bd1f95cd593223a6e47d9f291bf7ea7eaa261473aa40145b4a33.jpg" + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "table_body" + } + ], + "index": 9 + }, + { + "type": "image", + "bbox": [ + 315, + 183, + 553, + 255 + ], + "blocks": [ + { + "bbox": [ + 315, + 183, + 553, + 255 + ], + "lines": [ + { + "bbox": [ + 315, + 183, + 553, + 255 + ], + "spans": [ + { + "bbox": [ + 315, + 183, + 553, + 255 + ], + "type": "image", + "image_path": "adb871732c94a18747dc50fc01192bda794e1167ea88792b2145623dbc03eee2.jpg" + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 313, + 257, + 553, + 280 + ], + "lines": [ + { + "bbox": [ + 313, + 257, + 553, + 280 + ], + "spans": [ + { + "bbox": [ + 313, + 257, + 553, + 280 + ], + "type": "text", + "content": "Figure 8. Performance evaluation of different distance metrics used in graph matching and graph distillation." + } + ] + } + ], + "index": 11, + "angle": 0, + "type": "image_caption" + } + ], + "index": 10 + }, + { + "bbox": [ + 313, + 287, + 555, + 346 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 287, + 555, + 346 + ], + "spans": [ + { + "bbox": [ + 313, + 287, + 555, + 346 + ], + "type": "text", + "content": "ferent CLIP backbones, ViT-B/32 and ViT-L/14. Our CASS performs favorably across various backbones, consistently outperforming existing methods in mIoU and pAcc. These results demonstrate the effectiveness and robustness of our method across different backbone architectures." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 313, + 350, + 556, + 483 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 350, + 556, + 483 + ], + "spans": [ + { + "bbox": [ + 313, + 350, + 556, + 483 + ], + "type": "text", + "content": "Distance Metric for Graph Matching. We evaluate different distance metrics " + }, + { + "bbox": [ + 313, + 350, + 556, + 483 + ], + "type": "inline_equation", + "content": "\\mathcal{D}_W" + }, + { + "bbox": [ + 313, + 350, + 556, + 483 + ], + "type": "text", + "content": " used in Sec. 3.2.1, including Wasserstein Distance, Cosine Distance, KL Divergence, and Euclidean Distance. As shown in Fig. 8, the Wasserstein Distance shows the best results as it provides an intuitive measure of the overall shape of sorted eigenvalue distributions. While we use Wasserstein Distance due to its performance, the other well-known metrics exhibited comparable results, demonstrating the robustness of our method in terms of the choice of a distance metric and further supporting the effectiveness of our proposed approach." + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 313, + 487, + 388, + 499 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 487, + 388, + 499 + ], + "spans": [ + { + "bbox": [ + 313, + 487, + 388, + 499 + ], + "type": "text", + "content": "5. Conclusion" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 313, + 502, + 555, + 657 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 502, + 555, + 657 + ], + "spans": [ + { + "bbox": [ + 313, + 502, + 555, + 657 + ], + "type": "text", + "content": "In this work, we present CASS, a training-free approach for OVSS that integrates object-level context into CLIP. By distilling object-aware structural information from VFM into CLIP's attention, CASS effectively enhances intra-object coherence, ensuring that each object is consistently represented as a unified semantic entity. Additionally, we leverage object-presence prior knowledge to adjust text embeddings and refine object class assignments, aligning them with object-specific semantic targets in the image. Consequently, CASS effectively groups object components and assigns semantic labels without extra training or additional labeled data, making it a robust solution across diverse datasets and visual contexts." + } + ] + } + ], + "index": 15 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 313, + 671, + 553, + 704 + ], + "type": "page_footnote", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 671, + 553, + 704 + ], + "spans": [ + { + "bbox": [ + 313, + 671, + 553, + 704 + ], + "type": "text", + "content": "Acknowledgement. This work was supported in part by IITP RS-2024-00457882 (National AI Research Hub Project), IITP 2020-0-01361, NRF RS-2024-00345806, NRF RS-2023-00219019, and RS-2024-00403860 (Korea Basic Science Institute, National Research Facilities and Equipment Center)." + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "type": "text", + "content": "15040" + } + ] + } + ], + "index": 17 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 7 + }, + { + "para_blocks": [ + { + "bbox": [ + 56, + 71, + 115, + 83 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 71, + 115, + 83 + ], + "spans": [ + { + "bbox": [ + 56, + 71, + 115, + 83 + ], + "type": "text", + "content": "References" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 57, + 91, + 295, + 713 + ], + "type": "list", + "angle": 0, + "index": 16, + "blocks": [ + { + "bbox": [ + 61, + 91, + 294, + 113 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 61, + 91, + 294, + 113 + ], + "spans": [ + { + "bbox": [ + 61, + 91, + 294, + 113 + ], + "type": "text", + "content": "[1] Nikita Araslanov and Stefan Roth. Single-stage semantic segmentation from image labels. In CVPR, 2020. 7" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 61, + 114, + 295, + 146 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 61, + 114, + 295, + 146 + ], + "spans": [ + { + "bbox": [ + 61, + 114, + 295, + 146 + ], + "type": "text", + "content": "[2] Danny Barash and Matan Gavish. Optimal shrinkage of singular values under random data contamination. NeurIPS, 2017. 5" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 62, + 148, + 295, + 192 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 148, + 295, + 192 + ], + "spans": [ + { + "bbox": [ + 62, + 148, + 295, + 192 + ], + "type": "text", + "content": "[3] Luca Barsellotti, Roberto Amoroso, Lorenzo Baraldi, and Rita Cucchiara. Fossil: Free open-vocabulary semantic segmentation through synthetic references retrieval. In WACV, 2024. 2, 6" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 62, + 194, + 295, + 237 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 194, + 295, + 237 + ], + "spans": [ + { + "bbox": [ + 62, + 194, + 295, + 237 + ], + "type": "text", + "content": "[4] Luca Barsellotti, Roberto Amoroso, Marcella Cornia, Lorenzo Baraldi, and Rita Cucchiara. Training-free open-vocabulary segmentation with offline diffusion-augmented prototype generation. In CVPR, 2024. 2, 6" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 62, + 239, + 294, + 272 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 239, + 294, + 272 + ], + "spans": [ + { + "bbox": [ + 62, + 239, + 294, + 272 + ], + "type": "text", + "content": "[5] Toby Boas, Aritra Dutta, Xin Li, Kathryn P Mercier, and Eric Niderman. Shrinkage function and its applications in matrix approximation. arXiv preprint arXiv:1601.07600, 2016. 5" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 62, + 274, + 294, + 316 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 274, + 294, + 316 + ], + "spans": [ + { + "bbox": [ + 62, + 274, + 294, + 316 + ], + "type": "text", + "content": "[6] Walid Bousselham, Felix Petersen, Vittorio Ferrari, and Hilde Kuehne. Grounding everything: Emerging localization properties in vision-language transformers. In CVPR, 2024. 6" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 62, + 319, + 294, + 350 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 319, + 294, + 350 + ], + "spans": [ + { + "bbox": [ + 62, + 319, + 294, + 350 + ], + "type": "text", + "content": "[7] Holger Caesar, Jasper Uijlings, and Vittorio Ferrari. COCO-Stuff: Thing and Stuff Classes in Context. In CVPR, 2018. 6, 7, 8" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 62, + 353, + 294, + 396 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 353, + 294, + 396 + ], + "spans": [ + { + "bbox": [ + 62, + 353, + 294, + 396 + ], + "type": "text", + "content": "[8] Mathilde Caron, Hugo Touvron, Ishan Misra, Hervé Jégou, Julien Mairal, Piotr Bojanowski, and Armand Joulin. Emerging Properties in Self-Supervised Vision Transformers. In ICCV, 2021. 2, 4, 6" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 62, + 399, + 294, + 441 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 399, + 294, + 441 + ], + "spans": [ + { + "bbox": [ + 62, + 399, + 294, + 441 + ], + "type": "text", + "content": "[9] Junbum Cha, Jonghwan Mun, and Byungseok Roh. Learning To Generate Text-Grounded Mask for Open-World Semantic Segmentation From Only Image-Text Pairs. In CVPR, 2023. 1, 6" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 57, + 444, + 294, + 486 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 444, + 294, + 486 + ], + "spans": [ + { + "bbox": [ + 57, + 444, + 294, + 486 + ], + "type": "text", + "content": "[10] Liang-Chieh Chen, Yukun Zhu, George Papandreou, Florian Schroff, and Hartwig Adam. Encoder-decoder with atrous separable convolution for semantic image segmentation. In ECCV, 2018. 1, 3" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 57, + 488, + 294, + 531 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 488, + 294, + 531 + ], + "spans": [ + { + "bbox": [ + 57, + 488, + 294, + 531 + ], + "type": "text", + "content": "[11] Seokju Cho, Heeseong Shin, Sunghwan Hong, Anurag Arnab, Paul Hongsuck Seo, and Seungryong Kim. Cat-seg: Cost aggregation for open-vocabulary semantic segmentation. In CVPR, 2024. 1" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 57, + 534, + 294, + 555 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 534, + 294, + 555 + ], + "spans": [ + { + "bbox": [ + 57, + 534, + 294, + 555 + ], + "type": "text", + "content": "[12] Fan RK Chung. Spectral graph theory. American Mathematical Soc., 1997. 5" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 57, + 557, + 294, + 609 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 557, + 294, + 609 + ], + "spans": [ + { + "bbox": [ + 57, + 557, + 294, + 609 + ], + "type": "text", + "content": "[13] Marius Cordts, Mohamed Omran, Sebastian Ramos, Timo Rehfeld, Markus Enzweiler, Rodrigo Benenson, Uwe Franke, Stefan Roth, and Bernt Schiele. The Cityscapes Dataset for Semantic Urban Scene Understanding. In CVPR, 2016. 6" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 57, + 613, + 294, + 677 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 613, + 294, + 677 + ], + "spans": [ + { + "bbox": [ + 57, + 613, + 294, + 677 + ], + "type": "text", + "content": "[14] Alexey Dosovitskiy, Lucas Beyer, Alexander Kolesnikov, Dirk Weissenborn, Xiaohua Zhai, Thomas Unterthiner, Mostafa Dehghani, Matthias Minderer, Georg Heigold, Sylvain Gelly, Jakob Uszkoreit, and Neil Houlsby. An Image is Worth 16x16 Words: Transformers for Image Recognition at Scale. In ICLR, 2021. 2" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 57, + 680, + 294, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 680, + 294, + 713 + ], + "spans": [ + { + "bbox": [ + 57, + 680, + 294, + 713 + ], + "type": "text", + "content": "[15] M. Everingham, L. Van Gool, C. K. I. Williams, J. Winn, and A. Zisserman. The PASCAL Visual Object Classes Challenge 2012 (VOC2012) Results, 2012. 6, 7, 8" + } + ] + } + ], + "index": 15 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 316, + 73, + 553, + 713 + ], + "type": "list", + "angle": 0, + "index": 33, + "blocks": [ + { + "bbox": [ + 316, + 73, + 553, + 106 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 73, + 553, + 106 + ], + "spans": [ + { + "bbox": [ + 316, + 73, + 553, + 106 + ], + "type": "text", + "content": "[16] Sina Hajimiri, Ismail Ben Ayed, and Jose Dolz. Pay attention to your neighbours: Training-free open-vocabulary semantic segmentation. In WACV, 2025. 1, 2, 3, 6, 7, 8" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 317, + 108, + 553, + 152 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 108, + 553, + 152 + ], + "spans": [ + { + "bbox": [ + 317, + 108, + 553, + 152 + ], + "type": "text", + "content": "[17] Mark Hamilton, Zhoutong Zhang, Bharath Hariharan, Noah Snavely, and William T Freeman. Unsupervised semantic segmentation by distilling feature correspondences. arXiv preprint arXiv:2203.08414, 2022. 3" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 317, + 154, + 553, + 196 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 154, + 553, + 196 + ], + "spans": [ + { + "bbox": [ + 317, + 154, + 553, + 196 + ], + "type": "text", + "content": "[18] Woojung Han, Seil Kang, Kyobin Choo, and Seong Jae Hwang. Complementary branch fusing class and semantic knowledge for robust weakly supervised semantic segmentation. Pattern Recognition, 2025. 1" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 317, + 199, + 553, + 242 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 199, + 553, + 242 + ], + "spans": [ + { + "bbox": [ + 317, + 199, + 553, + 242 + ], + "type": "text", + "content": "[19] Ali Hatamizadeh, Yucheng Tang, Vishwesh Nath, Dong Yang, Andriy Myronenko, Bennett Landman, Holger R Roth, and Daguang Xu. Unetr: Transformers for 3d medical image segmentation. In CVPR, 2022. 1, 3" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 316, + 245, + 553, + 277 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 245, + 553, + 277 + ], + "spans": [ + { + "bbox": [ + 316, + 245, + 553, + 277 + ], + "type": "text", + "content": "[20] Ruixiang Jiang, Lingbo Liu, and Changwen Chen. Clipcount: Towards text-guided zero-shot object counting. In ACM Multimedia, 2023. 2, 5" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 316, + 280, + 553, + 311 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 280, + 553, + 311 + ], + "spans": [ + { + "bbox": [ + 316, + 280, + 553, + 311 + ], + "type": "text", + "content": "[21] Dahiyun Kang and Minsu Cho. In defense of lazy visual grounding for open-vocabulary semantic segmentation. In ECCV, 2024. 2, 6, 7, 8" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 316, + 314, + 553, + 346 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 314, + 553, + 346 + ], + "spans": [ + { + "bbox": [ + 316, + 314, + 553, + 346 + ], + "type": "text", + "content": "[22] Seil Kang, Jinyeong Kim, Junhyeok Kim, and Seong Jae Hwang. See what you are told: Visual attention sink in large multimodal models. In ICLR, 2025. 4" + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 316, + 349, + 553, + 381 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 349, + 553, + 381 + ], + "spans": [ + { + "bbox": [ + 316, + 349, + 553, + 381 + ], + "type": "text", + "content": "[23] Seil Kang, Jinyeong Kim, Junhyeok Kim, and Seong Jae Hwang. Your large vision-language model only needs a few attention heads for visual grounding. In CVPR, 2025. 2" + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 316, + 384, + 553, + 415 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 384, + 553, + 415 + ], + "spans": [ + { + "bbox": [ + 316, + 384, + 553, + 415 + ], + "type": "text", + "content": "[24] Chanyoung Kim, Woogung Han, Dayun Ju, and Seong Jae Hwang. Eagle: Eigen aggregation learning for object-centric unsupervised semantic segmentation. In CVPR, 2024. 2, 3" + } + ] + } + ], + "index": 25 + }, + { + "bbox": [ + 317, + 418, + 553, + 460 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 418, + 553, + 460 + ], + "spans": [ + { + "bbox": [ + 317, + 418, + 553, + 460 + ], + "type": "text", + "content": "[25] Alexander Kirillov, Eric Mintun, Nikhila Ravi, Hanzi Mao, Chloe Rolland, Laura Gustafson, Tete Xiao, Spencer Whitehead, Alexander C Berg, Wan-Yen Lo, et al. Segment anything. In ICCV, 2023. 3" + } + ] + } + ], + "index": 26 + }, + { + "bbox": [ + 317, + 463, + 553, + 495 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 463, + 553, + 495 + ], + "spans": [ + { + "bbox": [ + 317, + 463, + 553, + 495 + ], + "type": "text", + "content": "[26] Philipp Krahenbuhl and Vladlen Koltun. Efficient Inference in Fully Connected CRFs with Gaussian Edge Potentials. In NeurIPS, 2011. 7" + } + ] + } + ], + "index": 27 + }, + { + "bbox": [ + 316, + 498, + 553, + 519 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 498, + 553, + 519 + ], + "spans": [ + { + "bbox": [ + 316, + 498, + 553, + 519 + ], + "type": "text", + "content": "[27] Harold W Kuhn. The hungarian method for the assignment problem. Naval research logistics quarterly, 1955. 4" + } + ] + } + ], + "index": 28 + }, + { + "bbox": [ + 316, + 521, + 553, + 564 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 521, + 553, + 564 + ], + "spans": [ + { + "bbox": [ + 316, + 521, + 553, + 564 + ], + "type": "text", + "content": "[28] Mengcheng Lan, Chaofeng Chen, Yiping Ke, Xinjiang Wang, Litong Feng, and Wayne Zhang. Clearclip: Decomposing clip representations for dense vision-language inference. In ECCV, 2024. 1, 6" + } + ] + } + ], + "index": 29 + }, + { + "bbox": [ + 317, + 567, + 553, + 609 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 567, + 553, + 609 + ], + "spans": [ + { + "bbox": [ + 317, + 567, + 553, + 609 + ], + "type": "text", + "content": "[29] Mengcheng Lan, Chaofeng Chen, Yiping Ke, Xinjiang Wang, Litong Feng, and Wayne Zhang. Proxyclip: Proxy attention improves clip for open-vocabulary segmentation. In ECCV, 2024. 1, 2, 3, 6, 7, 8" + } + ] + } + ], + "index": 30 + }, + { + "bbox": [ + 317, + 613, + 553, + 655 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 613, + 553, + 655 + ], + "spans": [ + { + "bbox": [ + 317, + 613, + 553, + 655 + ], + "type": "text", + "content": "[30] Yi Li, Hualiang Wang, Yiqun Duan, and Xiaomeng Li. Clip surgery for better explainability with enhancement in open-vocabulary tasks. arXiv preprint arXiv:2304.05653, 2023. 2" + } + ] + } + ], + "index": 31 + }, + { + "bbox": [ + 317, + 658, + 553, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 658, + 553, + 713 + ], + "spans": [ + { + "bbox": [ + 317, + 658, + 553, + 713 + ], + "type": "text", + "content": "[31] Yiran Li, Junpeng Wang, Xin Dai, Liang Wang, ChinChia Michael Yeh, Yan Zheng, Wei Zhang, and Kwan-Liu Ma. How does attention work in vision transformers? a visual analytics attempt. IEEE Transactions on Visualization and Computer Graphics, 2023. 4" + } + ] + } + ], + "index": 32 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "text", + "content": "15041" + } + ] + } + ], + "index": 34 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 8 + }, + { + "para_blocks": [ + { + "bbox": [ + 56, + 73, + 294, + 712 + ], + "type": "list", + "angle": 0, + "index": 15, + "blocks": [ + { + "bbox": [ + 56, + 73, + 294, + 116 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 73, + 294, + 116 + ], + "spans": [ + { + "bbox": [ + 56, + 73, + 294, + 116 + ], + "type": "text", + "content": "[32] Feng Liang, Bichen Wu, Xiaoliang Dai, Kunpeng Li, Yinan Zhao, Hang Zhang, Peizhao Zhang, Peter Vajda, and Diana Marculescu. Open-vocabulary semantic segmentation with mask-adapted clip. In CVPR, 2023. 1" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 56, + 118, + 294, + 161 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 118, + 294, + 161 + ], + "spans": [ + { + "bbox": [ + 56, + 118, + 294, + 161 + ], + "type": "text", + "content": "[33] Jiayun Luo, Siddhesh Khandelwal, Leonid Sigal, and Boyang Li. Emergent open-vocabulary semantic segmentation from off-the-shelf vision-language models. In CVPR, 2024. 6" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 56, + 163, + 294, + 195 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 163, + 294, + 195 + ], + "spans": [ + { + "bbox": [ + 56, + 163, + 294, + 195 + ], + "type": "text", + "content": "[34] Yunyao Mao, Jiajun Deng, Wengang Zhou, Li Li, Yao Fang, and Houqiang Li. Clip4hoi: towards adapting clip for practical zero-shot hoi detection. NeurIPS, 2023. 2, 5" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 56, + 198, + 294, + 229 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 198, + 294, + 229 + ], + "spans": [ + { + "bbox": [ + 56, + 198, + 294, + 229 + ], + "type": "text", + "content": "[35] Octave Mariotti, Oisin Mac Aodha, and Hakan Bilen. Improving semantic correspondence with viewpoint-guided spherical maps. In CVPR, 2024. 3" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 56, + 232, + 294, + 274 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 232, + 294, + 274 + ], + "spans": [ + { + "bbox": [ + 56, + 232, + 294, + 274 + ], + "type": "text", + "content": "[36] Luke Melas-Kyriazi, Christian Rupprecht, Iro Laina, and Andrea Vedaldi. Deep spectral methods: A surprisingly strong baseline for unsupervised semantic segmentation and localization. In CVPR, 2022. 2, 3" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 56, + 276, + 294, + 319 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 276, + 294, + 319 + ], + "spans": [ + { + "bbox": [ + 56, + 276, + 294, + 319 + ], + "type": "text", + "content": "[37] Roozbeh Mottaghi, Xianjie Chen, Xiaobai Liu, Nam-Gyu Cho, Seong-Whan Lee, Sanja Fidler, Raquel Urtasun, and Alan Yuille. The Role of Context for Object Detection and Semantic Segmentation in the Wild. In CVPR, 2014. 6, 7, 8" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 56, + 321, + 294, + 374 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 321, + 294, + 374 + ], + "spans": [ + { + "bbox": [ + 56, + 321, + 294, + 374 + ], + "type": "text", + "content": "[38] Maxime Oquab, Timothee Darct, Theo Moutakanni, Huy Vo, Marc Szafraniec, Vasil Khalidov, Pierre Fernandez, Daniel Haziza, Francisco Massa, Alaaeldin El-Nouby, et al. Dinov2: Learning robust visual features without supervision. arXiv preprint arXiv:2304.07193, 2023. 2" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 56, + 376, + 294, + 431 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 376, + 294, + 431 + ], + "spans": [ + { + "bbox": [ + 56, + 376, + 294, + 431 + ], + "type": "text", + "content": "[39] Alec Radford, Jong Wook Kim, Chris Hallacy, Aditya Ramesh, Gabriel Goh, Sandhini Agarwal, Girish Sastry, Amanda Askell, Pamela Mishkin, Jack Clark, et al. Learning Transferable Visual Models From Natural Language Supervision. In ICML, 2021. 1, 2, 3, 5, 6, 7" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 56, + 433, + 294, + 464 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 433, + 294, + 464 + ], + "spans": [ + { + "bbox": [ + 56, + 433, + 294, + 464 + ], + "type": "text", + "content": "[40] Olaf Ronneberger, Philipp Fischer, and Thomas Brox. U-net: Convolutional networks for biomedical image segmentation. In MICCAI, 2015. 1, 3" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 56, + 467, + 294, + 521 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 467, + 294, + 521 + ], + "spans": [ + { + "bbox": [ + 56, + 467, + 294, + 521 + ], + "type": "text", + "content": "[41] Maximilian Seitzer, Max Horn, Andrii Zadaianchuk, Dominik Zietlow, Tianjun Xiao, Carl-Johann Simon-Gabriel, Tong He, Zheng Zhang, Bernhard Scholkopf, Thomas Brox, and Francesco Locatello. Bridging the gap to real-world object-centric learning. In ICLR, 2023. 3" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 56, + 523, + 294, + 555 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 523, + 294, + 555 + ], + "spans": [ + { + "bbox": [ + 56, + 523, + 294, + 555 + ], + "type": "text", + "content": "[42] Hyun Seok Seong, WonJun Moon, SuBeen Lee, and Jae-Pil Heo. Leveraging hidden positives for unsupervised semantic segmentation. In CVPR, 2023. 3" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 56, + 557, + 294, + 599 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 557, + 294, + 599 + ], + "spans": [ + { + "bbox": [ + 56, + 557, + 294, + 599 + ], + "type": "text", + "content": "[43] Xiangheng Shan, Dongyue Wu, Guilin Zhu, Yuanjie Shao, Nong Sang, and Changxin Gao. Open-vocabulary semantic segmentation with image embedding balancing. In CVPR, 2024. 1" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 56, + 601, + 294, + 634 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 601, + 294, + 634 + ], + "spans": [ + { + "bbox": [ + 56, + 601, + 294, + 634 + ], + "type": "text", + "content": "[44] Tong Shao, Zhuotao Tian, Hang Zhao, and Jingyong Su. Explore the potential of clip for training-free open vocabulary semantic segmentation. In ECCV, 2024. 3, 6, 7" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 56, + 635, + 294, + 667 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 635, + 294, + 667 + ], + "spans": [ + { + "bbox": [ + 56, + 635, + 294, + 667 + ], + "type": "text", + "content": "[45] Gyungin Shin, Weidi Xie, and Samuel Albanie. ReCo: Retrieve and Co-segment for Zero-shot Transfer. In NeurIPS, 2022. 6" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 56, + 670, + 294, + 712 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 670, + 294, + 712 + ], + "spans": [ + { + "bbox": [ + 56, + 670, + 294, + 712 + ], + "type": "text", + "content": "[46] Oriane Simeoni, Gilles Puy, Huy V Vo, Simon Roburin, Spyros Gidaris, Andrei Bursuc, Patrick Pérez, Renaud Marlet, and Jean Ponce. Localizing objects with self-supervised transformers and no labels. In BMVC, 2021. 3" + } + ] + } + ], + "index": 14 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 316, + 73, + 553, + 578 + ], + "type": "list", + "angle": 0, + "index": 29, + "blocks": [ + { + "bbox": [ + 316, + 73, + 553, + 106 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 73, + 553, + 106 + ], + "spans": [ + { + "bbox": [ + 316, + 73, + 553, + 106 + ], + "type": "text", + "content": "[47] Shuyang Sun, Runjia Li, Philip Torr, Xiuye Gu, and Siyang Li. Clip as rn: Segment countless visual concepts without training endeavor. In CVPR, 2024. 2, 6" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 316, + 108, + 553, + 139 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 108, + 553, + 139 + ], + "spans": [ + { + "bbox": [ + 316, + 108, + 553, + 139 + ], + "type": "text", + "content": "[48] Feng Wang, Jieru Mei, and Alan Yuille. Sclip: Rethinking self-attention for dense vision-language inference. In ECCV, 2024. 1, 2, 3, 6, 7, 8" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 316, + 141, + 553, + 173 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 141, + 553, + 173 + ], + "spans": [ + { + "bbox": [ + 316, + 141, + 553, + 173 + ], + "type": "text", + "content": "[49] Xudong Wang, Rohit Girdhar, Stella X Yu, and Ishan Misra. Cut and Learn for Unsupervised Object Detection and Instance Segmentation. In CVPR, 2023. 3" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 316, + 175, + 553, + 207 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 175, + 553, + 207 + ], + "spans": [ + { + "bbox": [ + 316, + 175, + 553, + 207 + ], + "type": "text", + "content": "[50] Xin Wen, Bingchen Zhao, Anlin Zheng, Xiangyu Zhang, and Xiaojuan Qi. Self-supervised visual representation learning with semantic grouping. NeurIPS, 2022. 3" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 317, + 209, + 553, + 262 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 209, + 553, + 262 + ], + "spans": [ + { + "bbox": [ + 317, + 209, + 553, + 262 + ], + "type": "text", + "content": "[51] Ji-Jia Wu, Andy Chia-Hao Chang, Chieh-Yu Chuang, Chun-Pei Chen, Yu-Lun Liu, Min-Hung Chen, Hou-Ning Hu, Yung-Yu Chuang, and Yen-Yu Lin. Image-text codecomposition for text-supervised semantic segmentation. In CVPR, 2024. 6" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 316, + 264, + 553, + 307 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 264, + 553, + 307 + ], + "spans": [ + { + "bbox": [ + 316, + 264, + 553, + 307 + ], + "type": "text", + "content": "[52] Monika Wysoczanska, Oriane Simeoni, Michael Ramamonjisoa, Andrei Bursuc, Tomasz Trzciński, and Patrick Pérez. Clip-dinoiser: Teaching clip a few dino tricks. arXiv preprint arXiv:2312.12359, 2023. 6, 7" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 316, + 309, + 553, + 342 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 309, + 553, + 342 + ], + "spans": [ + { + "bbox": [ + 316, + 309, + 553, + 342 + ], + "type": "text", + "content": "[53] Bin Xie, Jiale Cao, Jin Xie, Fahad Shahbaz Khan, and Yanwei Pang. Sed: A simple encoder-decoder for open-vocabulary semantic segmentation. In CVPR, 2024. 1" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 316, + 343, + 553, + 387 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 343, + 553, + 387 + ], + "spans": [ + { + "bbox": [ + 316, + 343, + 553, + 387 + ], + "type": "text", + "content": "[54] Jiarui Xu, Shalini De Mello, Sifei Liu, Wonmin Byeon, Thomas Breuel, Jan Kautz, and Xiaolong Wang. GroupViT: Semantic Segmentation Emerges From Text Supervision. In CVPR, 2022. 1, 6" + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 316, + 388, + 553, + 430 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 388, + 553, + 430 + ], + "spans": [ + { + "bbox": [ + 316, + 388, + 553, + 430 + ], + "type": "text", + "content": "[55] Jiarui Xu, Sifei Liu, Arash Vahdat, Wonmin Byeon, Xiaolong Wang, and Shalini De Mello. Open-vocabulary panoptic segmentation with text-to-image diffusion models. In CVPR, 2023. 1" + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 316, + 433, + 553, + 465 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 433, + 553, + 465 + ], + "spans": [ + { + "bbox": [ + 316, + 433, + 553, + 465 + ], + "type": "text", + "content": "[56] Mengde Xu, Zheng Zhang, Fangyun Wei, Han Hu, and Xiang Bai. Side adapter network for open-vocabulary semantic segmentation. In CVPR, 2023. 1" + } + ] + } + ], + "index": 25 + }, + { + "bbox": [ + 316, + 467, + 553, + 510 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 467, + 553, + 510 + ], + "spans": [ + { + "bbox": [ + 316, + 467, + 553, + 510 + ], + "type": "text", + "content": "[57] Bolei Zhou, Hang Zhao, Xavier Puig, Tete Xiao, Sanja Fidler, Adela Barriuso, and Antonio Torralba. Semantic Understanding of Scenes Through the ADE20K Dataset. IJCV, 2019. 6, 7" + } + ] + } + ], + "index": 26 + }, + { + "bbox": [ + 316, + 512, + 553, + 533 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 512, + 553, + 533 + ], + "spans": [ + { + "bbox": [ + 316, + 512, + 553, + 533 + ], + "type": "text", + "content": "[58] Chong Zhou, Chen Change Loy, and Bo Dai. Extract Free Dense Labels from CLIP. In ECCV, 2022. 1, 2, 6" + } + ] + } + ], + "index": 27 + }, + { + "bbox": [ + 316, + 535, + 553, + 578 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 535, + 553, + 578 + ], + "spans": [ + { + "bbox": [ + 316, + 535, + 553, + 578 + ], + "type": "text", + "content": "[59] Lianghui Zhu, Yingyue Li, Jiemin Fang, Yan Liu, Hao Xin, Wenyu Liu, and Xinggang Wang. Weaktr: Exploring plain vision transformer for weakly-supervised semantic segmentation. arXiv preprint arXiv:2304.01184, 2023. 4" + } + ] + } + ], + "index": 28 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "type": "text", + "content": "15042" + } + ] + } + ], + "index": 30 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 9 + } + ], + "_backend": "vlm", + "_version_name": "2.6.4" +} \ No newline at end of file diff --git a/2025/DistinctAD_ Distinctive Audio Description Generation in Contexts/d4975861-cea8-4f94-9bb1-e02706824334_content_list.json b/2025/DistinctAD_ Distinctive Audio Description Generation in Contexts/d4975861-cea8-4f94-9bb1-e02706824334_content_list.json new file mode 100644 index 0000000000000000000000000000000000000000..ec140121992539be1f0247819364ea23c429561e --- /dev/null +++ b/2025/DistinctAD_ Distinctive Audio Description Generation in Contexts/d4975861-cea8-4f94-9bb1-e02706824334_content_list.json @@ -0,0 +1,2049 @@ +[ + { + "type": "text", + "text": "DistinctAD: Distinctive Audio Description Generation in Contexts", + "text_level": 1, + "bbox": [ + 166, + 130, + 828, + 151 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Bo Fang $^{1}$ , Wenhao Wu $^{2,3}$ , Qiangqiang Wu $^{1}$ , Yuxin Song $^{2}$ , Antoni B. Chan $^{1}$ $^{1}$ Department of Computer Science, City University of Hong Kong \n $^{2}$ Baidu Inc. \n $^{3}$ The University of Sydney", + "bbox": [ + 194, + 179, + 802, + 233 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "{bofang6-c, qiangqwu2-c}@my.cityu.edu.hk, wenhao.wu@sydney.edu.au songyuxin02@baidu.com, abchan@cityu.edu.hk", + "bbox": [ + 215, + 236, + 781, + 268 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Abstract", + "text_level": 1, + "bbox": [ + 246, + 304, + 326, + 319 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Audio Descriptions (ADs) aim to provide a narration of a movie in text form, describing non-dialogue-related narratives, such as characters, actions, or scene establishment. Automatic generation of ADs remains challenging due to: i) the domain gap between movie-AD data and existing data used to train vision-language models, and ii) the issue of contextual redundancy arising from highly similar neighboring visual clips in a long movie. In this work, we propose DistinctAD, a novel two-stage framework for generating ADs that emphasize distinctiveness to produce better narratives. To address the domain gap, we introduce a CLIP-AD adaptation strategy that does not require additional AD corpora, enabling more effective alignment between movie and AD modalities at both global and fine-grained levels. In Stage-II, DistinctAD incorporates two key innovations: (i) a Contextual Expectation-Maximization Attention (EMA) module that reduces redundancy by extracting common bases from consecutive video clips, and (ii) an explicit distinctive word prediction loss that filters out repeated words in the context, ensuring the prediction of unique terms specific to the current AD. Comprehensive evaluations on MAD-Eval, CMD-AD, and TV-AD benchmarks demonstrate the superiority of DistinctAD, with the model consistently outperforming baselines, particularly in Recall@k/N, highlighting its effectiveness in producing high-quality, distinctive ADs.", + "bbox": [ + 89, + 335, + 483, + 729 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "1. Introduction", + "text_level": 1, + "bbox": [ + 91, + 753, + 220, + 768 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Audio description (AD) [20, 67] is a crucial accessibility service that provides verbal narration of visual elements in media content for individuals who are blind or have low vision. By offering succinct and vivid descriptions, ADs enable visually impaired audiences to fully comprehend and engage with non-dialogue-related narratives, e.g., characters, facial expressions, non-verbal actions, or scene establishment. Recent studies also show ADs' value for", + "bbox": [ + 89, + 779, + 483, + 900 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/500e9864fe6895eca71b0b1abed5cb42803e90aabfd9e97daf3d43be75d8831b.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 524, + 305, + 892, + 398 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/b3c76ba5f2d0c37722518ee1c83cab3f34eebc6c4e2d1202cd1b64a869b0da0f.jpg", + "image_caption": [ + "Figure 1. (a) Previous methods approach the AD task similar to video captioning, using only a single video clip as input, which leads to repetitive ADs due to highly similar neighboring clips. (b) Our DistinctAD method generates distinctive ADs across $N$ consecutive clips, with three key innovations: VLM-AD adaptation, the Distinct Module, and explicit distinctive words prediction." + ], + "image_footnote": [], + "bbox": [ + 524, + 401, + 898, + 494 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "sighted viewers in supporting eye-free activities and facilitating child language development [31, 55], reinforcing its pivotal role in fostering inclusivity by bridging the perceptual gap between visual and non-visual elements. Crafting ADs requires careful attention to timing, language, and context to integrate smoothly with dialogue [22]. However, despite the availability of advanced AD platforms [9, 54], human-annotated methods are costly and difficult to scale, highlighting the need for automated generation systems, especially with the rise of user-generated content.", + "bbox": [ + 509, + 595, + 906, + 747 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Advancements in Vision-Language Models (VLMs) and Large-Language Models (LLMs) have led to growing interest in automatic AD generation for media. Current approaches fall into two categories: (i) using powerful proprietary models like GPT-4 [3] or GPT-4V [53] in a training-free manner [14, 39, 82, 86], and (ii) fine-tuning open-source VLM components, such as visual-text adapters [5, 34, 42], for AD tasks [21-23, 40, 56, 74]. Both approaches have limitations: (i) Training-free methods often perform poorly and suffer from hallucinations due to the unique na", + "bbox": [ + 511, + 750, + 908, + 902 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "CVF", + "bbox": [ + 106, + 2, + 181, + 42 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore.", + "bbox": [ + 236, + 0, + 810, + 46 + ], + "page_idx": 0 + }, + { + "type": "page_number", + "text": "13571", + "bbox": [ + 480, + 944, + 517, + 955 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "ture of AD (e.g., character names and narrative coherence), which differs from the common text data LLMs are trained on. (ii) Fine-tuning methods generally perform better but are still limited by insufficient data to fully adapt to the movie-AD domain and face the context-repetition issue.", + "bbox": [ + 89, + 90, + 480, + 167 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Unlike video captioning [38, 62], ADs are generated on consecutive intervals (visual clips) throughout long videos [68], e.g., movies. The context-repetition issue arises when models produce repetitive or similar descriptions for consecutive visual clips, especially when using prior ADs as prompts [21, 74]. This occurs because sequential clips often comprise redundant scenes or characters (and therein redundant visual features), leading models that only use the current visual clip to repeat the same information from the past, as shown in Fig. 1. However, audiences are more interested in the unique and distinct events of the current clip, rather than the common elements from the previous one.", + "bbox": [ + 89, + 169, + 480, + 349 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "In this paper, we propose DistinctAD, a two-stage framework for generating distinctive ADs within contexts. Given the domain gap between the movie-AD and VLM training data, we first bridge this gap in Stage-I by adapting VLMs, such as CLIP [58], to the movie-AD domain. Our adaptation strategy is inspired by a key observation (see Appendix §A): AD sentences encoded by the CLIP text encoder can be effectively reconstructed using simple LLMs like GPT-2 [57] with minimal fine-tuning, whereas AD reconstructions using CLIP visual features from the corresponding clips are often of poor quality. This suggests that while CLIP's multi-modal embedding space is rich enough to represent AD information, its visual encoder is insufficient for extracting it. To mitigate this domain gap, we adapt the CLIP vision encoder to better align with the frozen CLIP text encoder using existing paired video-AD data. The alignment involves global matching at video-sentence level, similar to CLIP pre-training. A challenge arises because video clips are labeled with whole ADs, and words may not appear in every frame but must be aggregated over frames. Therefore, we propose fine-grained matching at frame-word level for this multiple-instance setting.", + "bbox": [ + 91, + 354, + 483, + 686 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "For Stage-II, we propose a novel distinctive AD narrating pipeline based on the Expectation-Maximization Attention (EMA) [17] algorithm, which has demonstrated its efficacy in tasks such as semantic segmentation [35], video object segmentation [41], and text-video retrieval [27]. Differently, we apply EMA to contextual clips from long videos, which often exhibit high redundancy due to recurring scenes or characters. By extracting common bases from contextual information, DistinctAD reduces redundancy and generates compact, discriminative representations that enable the LLM decoder to produce more distinctive ADs. To further emphasize distinctiveness explicitly, we introduce a distinctive word prediction loss that filters out words that repeatedly appear in contexts, ensuring that the LLM decoder", + "bbox": [ + 89, + 689, + 482, + 901 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "focuses on predicting unique words specific to the current AD. With these two designs, DistinctAD produces contextually distinctive and engaging ADs that can provide better narratives for the audience.", + "bbox": [ + 511, + 90, + 903, + 150 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "In summary, our contributions are three-fold:", + "bbox": [ + 532, + 152, + 831, + 165 + ], + "page_idx": 1 + }, + { + "type": "list", + "sub_type": "text", + "list_items": [ + "- We propose a CLIP-AD adaptation strategy tailored to movie-AD data, addressing the misalignment issue caused by the domain gap. Our adapted vision encoder can be seamlessly integrated into existing CLIP-based AD methods and stands to benefit from future improvements as more AD data becomes available.", + "- We introduce DistinctAD, which incorporates a Contextual EMA module and a distinctive word prediction loss, significantly enhancing the generation of distinctive ADs from consecutive visual clips with similar contexts.", + "- Comprehensive evaluations on MAD-Eval [21], CMD-AD [23], and TV-AD [82] highlight DistinctAD's superiority. Our outstanding performance in Recall@k/N demonstrates its effectiveness in generating high-quality ADs with both distinctiveness and technical excellence." + ], + "bbox": [ + 511, + 166, + 903, + 392 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "2. Related Work", + "text_level": 1, + "bbox": [ + 513, + 407, + 653, + 422 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Dense video captioning. A task closely related to AD is dense video captioning [29], which extends traditional video captioning [38, 45, 62, 63] by both generating a single caption for trimmed video segments as well as detecting and describing multiple events with grounded timestamps. Initial dense video captioning utilize a 2-stage pipeline [25, 26, 75, 79] by firstly performing localization and then describing events. Recent works [13, 18, 36, 51, 59, 64, 65, 75, 80, 83, 89] focus on training localization and captioning modules in an end-to-end manner to enhance inter-event associations. In contrast to these works, AD generation specifically aims to narrate a coherent story, maintain character-awareness, and complement the audio track without interfering with existing dialogue.", + "bbox": [ + 511, + 431, + 906, + 643 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "AD generation. Early AD systems relied heavily on specialized authoring tools [9] and skilled human contributors. Platforms like Rescribe [54] and LiveDescribe [9] have facilitated faster and more accurate AD creation; however, these methods are costly and do not scale efficiently for large volumes of visual content. Recent efforts have developed audio segmentation and transcription systems [8, 10, 11] to create high-quality video datasets with temporally aligned ADs [60, 61, 68, 69], advancing automatic AD research.", + "bbox": [ + 511, + 643, + 908, + 792 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "In general, current automatic AD generation systems can be categorized into two approaches: training-free and partial-fine-tuning. Training-free methods [39] generate ADs by leveraging proprietary models like GPT-4 [3] and GPT-4V [53]. MM-Narrator [86] enhances AD performance by multi-model in-context learning with memories. LLM-AD [14] and AutoAD-Zero [82] use prompts com", + "bbox": [ + 511, + 795, + 908, + 902 + ], + "page_idx": 1 + }, + { + "type": "page_number", + "text": "13572", + "bbox": [ + 480, + 944, + 519, + 955 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "prising visual frames with textual character names and colorful circles [66], enabling character-centric AD generation. However, training-free AD methods often suffer from high evaluation costs at scale and relatively poor performance due to domain-specific challenges and LLM hallucinations. Partial-fine-tuning methods [21-23, 40, 74], as well as our DistinctAD, only fine-tune a lightweight adapter [5, 34] between the pre-trained vision and text encoders. A representative example is the AutoAD series [21-23], which builds automatic AD systems and enriches them with character-aware prompts within different vision-language frameworks. However, previous studies tend to focus on constructing more accurate external character banks, whereas treating AD generation similarly to video captioning, overlooks AD's unique sequential structure of video clips. In contrast, our method emphasizes understanding the visual content within its temporal context, leading to more distinctive AD generation.", + "bbox": [ + 89, + 90, + 483, + 362 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Distinctive captioning in images aims to articulate unique details that can help distinguishing targets from others. An intuitive way to promoting distinctiveness is through contrastive learning [16, 43, 47, 73], where generated captions are encouraged to align more closely with target images rather than distractors. In [12, 76-78], group-based distinctive attention is introduced to capture distinctiveness by comparing sets of similar images and re-weighting specific caption words. A recent closely related field is difference captioning [32, 33, 84], which aims to describe differences between a single pair of images. VisDiff [19] scales difference captioning to sets containing thousands of images with natural language. Our work differs from these distinctive captioning works in that we are the first to explore distinctiveness across dense, consecutive clips within hours-long movies, thereby generating ADs with better narrative.", + "bbox": [ + 89, + 364, + 483, + 604 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3. Method", + "text_level": 1, + "bbox": [ + 89, + 622, + 181, + 638 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3.1. Stage-I: CLIP-AD Adaptation", + "text_level": 1, + "bbox": [ + 89, + 648, + 357, + 665 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "AD and the visual content it describes exhibit a significant domain gap compared to typical large-scale web data. This gap often causes misalignment in current partial-fine-tuning techniques. Previous studies [21, 22, 74] try alleviating this problem by pre-training LLMs on text-only AD corpus, e.g. AudioVault [1]. However, misalignment persists at the initial stage of vision encoding, which is often neglected.", + "bbox": [ + 89, + 672, + 482, + 777 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Inspired by our findings that AD sentences encoded by the CLIP text encoder can be effectively recovered using the GPT-2 language model (see §1 and Appendix §A), we identify that the primary issue of misalignment is caused by the CLIP vision encoder, i.e., the discrepancy between visual embeddings and AD embeddings within the joint CLIP feature space. To address this, we propose adapting the CLIP vision encoder to the specific AD domain. However, due to", + "bbox": [ + 89, + 779, + 483, + 900 + ], + "page_idx": 2 + }, + { + "type": "image", + "img_path": "images/b6264c77c575b2ff77e33f5074d1d69a9f7241fd6de9b2de0c26ced461284741.jpg", + "image_caption": [ + "Figure 2. Illustration of Stage-I: CLIP-AD Adaptation. This process involves adapting the CLIP vision encoder to specific movie-AD data through global-level video-AD matching (bottom right) and fine-grained frame-AD matching (top right)." + ], + "image_footnote": [], + "bbox": [ + 519, + 84, + 898, + 335 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "the unique multiple-instance learning setting for video-AD pairs (see §1), we consider both global matching and fine-grained frame-word matching in our adaptation method.", + "bbox": [ + 511, + 407, + 906, + 452 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Global video-AD matching. A straightforward strategy involves adopting classical CLIP-style fine-tuning with video-AD pairs in large batches [46]. Formally, let video clip $\\mathbf{V}_i = [\\mathbf{f}_i^1;\\dots ;\\mathbf{f}_i^n ]\\in \\mathbb{R}^{n\\times C}$ be a collection of $n$ frame embeddings, and corresponding AD $\\mathbf{T}_i = [\\mathbf{w}_i^0;\\mathbf{w}_i^1;\\dots ;\\mathbf{w}_i^m ]\\in \\mathbb{R}^{(m + 1)\\times C}$ be a collection of $m$ word embeddings $(\\mathbf{w}_i^j)$ and the [CLS] token (denoted as $\\mathbf{w}_i^0$ ), where $C$ is the number of channels in the embedding space. We obtain the global video-level representation by averaging all frame embeddings in $\\mathbf{V}_i$ using mean pooling: $\\mathbf{v}_i = \\frac{1}{n}\\sum_{j = 1}^n\\mathbf{f}_i^j$ . Following the standard CLIP, we use the [CLS] token as the global textual AD representation $\\mathbf{t}_i = \\mathbf{w}_i^0$ . The global video $\\rightarrow$ AD matching is performed by maximizing the sum of the main diagonal of a $B\\times B$ similarity matrix, using the contrastive loss:", + "bbox": [ + 511, + 452, + 908, + 680 + ], + "page_idx": 2 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {L} _ {v \\rightarrow A D} = - \\frac {1}{B} \\sum_ {i = 1} ^ {B} \\log \\frac {\\exp (s i m (\\mathbf {v} _ {i} , \\mathbf {t} _ {i}))}{\\sum_ {j = 1} ^ {B} \\exp (s i m (\\mathbf {v} _ {i} , \\mathbf {t} _ {j}))}, \\tag {1}\n$$\n", + "text_format": "latex", + "bbox": [ + 535, + 683, + 906, + 724 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "where $B$ is the batch size, and the similarity function $sim(\\cdot)$ is the vector inner product. This process is illustrated in the bottom right of Fig. 2. Similarly, we drive the AD $\\rightarrow$ video loss $\\mathcal{L}_{AD\\rightarrow v}$ by maximizing the sum of the secondary diagonal (i.e., swapping the $i$ and $j$ indices in (1)). The final global-level contrastive loss is then the sum of the losses in both directions $\\mathcal{L}_q = \\mathcal{L}_{v\\to AD} + \\mathcal{L}_{AD\\to v}$ .", + "bbox": [ + 511, + 734, + 906, + 839 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Fine-grained frame-AD matching. Matching global video to AD sentence [CLS] (and vice versa) aids in joint feature space alignment. However, this alignment is insufficient for effective adaptation due to the specific multiple-instance", + "bbox": [ + 511, + 839, + 908, + 900 + ], + "page_idx": 2 + }, + { + "type": "page_number", + "text": "13573", + "bbox": [ + 480, + 944, + 517, + 955 + ], + "page_idx": 2 + }, + { + "type": "image", + "img_path": "images/7727ed981f6a1a0688c7cba9ca20ca96b17b8bae844d2371ae2ba69eee25f0a6.jpg", + "image_caption": [ + "Figure 3. Pipeline of Stage-II: Distinctive AD Narration. Stage-II processes $N$ consecutive video clips using the $\\mathrm{CLIP}_{\\mathrm{AD}}$ vision encoder from Stage-I. We generate contextual-distinctive ADs by two key innovations: i) a Contextual EMA module to learn compact and discriminative visual representations for improved prompting of LLMs; ii) an extra distinctive word loss for predicting AD-specific terms." + ], + "image_footnote": [], + "bbox": [ + 181, + 88, + 816, + 340 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "setting of ADs, where only some words may correspond to a particular frame, but all words will have correspondence in aggregate. Thus, we propose a fine-grained matching loss at the frame-level to address this issue.", + "bbox": [ + 88, + 398, + 480, + 457 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Formally, given the frame embeddings $\\mathbf{V}_i$ and the word embeddings $\\mathbf{T}_i' = [\\mathbf{w}_i^1; \\dots; \\mathbf{w}_i^m] \\in \\mathbb{R}^{m \\times C}$ , we calculate the weights of all words attending to each frame via softmax attention, taking $\\mathbf{V}_i$ as the query and $\\mathbf{T}_i'$ as the key. By then multiplying these attention weights by the value $\\mathbf{T}_i'$ , we obtain a frame-aware AD representation $\\tilde{\\mathbf{T}}_i \\in \\mathbb{R}^{n \\times C}$ :", + "bbox": [ + 89, + 459, + 483, + 550 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\tilde {\\mathbf {T}} _ {i} = \\mathrm {S o f t m a x} (\\mathbf {V} _ {i} \\mathbf {T} _ {i} ^ {\\prime T} / \\tau) \\mathbf {T} _ {i} ^ {\\prime} = [ \\tilde {\\mathbf {t}} _ {i} ^ {1}; \\dots ; \\tilde {\\mathbf {t}} _ {i} ^ {n} ], \\qquad (2)\n$$\n", + "text_format": "latex", + "bbox": [ + 133, + 555, + 482, + 575 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "where for each frame, the words embeddings that are most similar to the frame-level visual feature have been aggregated (via softmax attention). The temperature parameter $\\tau$ controls the aggregation process, where smaller $\\tau$ incorporates more textual information.", + "bbox": [ + 89, + 587, + 483, + 662 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "The goal of the fine-grained matching is to pull a frame visual feature $\\mathbf{f} \\in \\mathbf{V}_i$ closer to the frame-aware AD representations $\\tilde{\\mathbf{t}} \\in \\tilde{\\mathbf{T}}_i$ in (2), corresponding to the positive set. To achieve this, we define the negative set $\\tilde{\\mathbf{T}}_{neg}$ as the frame-aware AD embeddings generated from other video clips (in the batch), and then use a Multi-Instance Loss [49],", + "bbox": [ + 89, + 662, + 483, + 753 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {L} _ {f} = - \\frac {1}{B} \\sum_ {i = 1} ^ {B} \\log \\frac {\\sum_ {\\tilde {\\mathbf {t}} \\in \\tilde {\\mathbf {T}} _ {i}} \\exp (s i m (\\mathbf {f} , \\tilde {\\mathbf {t}}))}{\\sum_ {\\tilde {\\mathbf {t}} _ {*} \\in \\{\\tilde {\\mathbf {T}} _ {i} \\cup \\tilde {\\mathbf {T}} _ {n e g} \\}} \\exp (s i m (\\mathbf {f} , \\tilde {\\mathbf {t}} _ {*}))}, \\tag {3}\n$$\n", + "text_format": "latex", + "bbox": [ + 101, + 772, + 482, + 815 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "where $\\mathbf{f} \\in \\mathbf{V}_i$ is a sampled frame from $\\mathbf{V}_i$ . This process is illustrated in the top right of Fig. 2.", + "bbox": [ + 89, + 824, + 482, + 854 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Summary for Stage-I. The final objective for Stage-I is to minimize the sum of global and fine-grained aligning losses, balanced by a trade-off coefficient, $\\mathcal{L}_{\\mathrm{I}} = \\gamma \\mathcal{L}_{g} + (1 -$", + "bbox": [ + 89, + 854, + 483, + 902 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "$\\gamma)\\mathcal{L}_f$ . Note that during this adaptation process, the CLIP-Text encoder model remains frozen, and only the CLIP-Vision encoder is fine-tuned. Our fine-grained frame-AD matching is entirely parameter-free, as only the vision encoder will be utilized in the subsequent stage.", + "bbox": [ + 511, + 398, + 906, + 474 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "3.2. Stage-II: Distinctive AD Narration", + "text_level": 1, + "bbox": [ + 511, + 481, + 815, + 498 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "The motivation for generating distinctive ADs stems from the observation that LLM often produce repetitive descriptions for adjacent clips [56, 82, 86]. Despite improved character recognition, the visual representation itself is not discriminative among neighboring (contextual) clips, leading to uninteresting ADs. Our goal is to create contextual-distinctive ADs that highlight current differences. We hypothesize, as verified in Appendix §B, that sequential clips from a long video often share redundant scenes or characters, leading similar visual features in contexts. Thus, we propose Stage-II: distinctive AD narration.", + "bbox": [ + 511, + 505, + 906, + 671 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "As shown in Fig. 3, we prepare $N$ consecutive video clips (to be AD-described) $\\{\\mathbf{x}_1,\\mathbf{x}_2,\\dots ,\\mathbf{x}_N\\}$ , each containing $T$ uniformly sampled frames $\\{\\mathcal{F}_1,\\mathcal{F}_2,\\dots ,\\mathcal{F}_T\\}$ . Following [21, 22, 74], we employ a learnable Perceiver adapter [5] to resample $T^{\\prime}$ prompt vectors for the $T$ frame embeddings encoded by our Stage-I vision encoder, $\\mathrm{CLIP}_{\\mathrm{AD}}$ . This process is formulated as:", + "bbox": [ + 511, + 671, + 908, + 777 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\mathbf {h} _ {\\mathbf {x} _ {i}} = \\operatorname {P e r c e i v e r} \\left(\\left\\{\\mathbf {f} _ {1}, \\mathbf {f} _ {2}, \\dots , \\mathbf {f} _ {T} \\right\\}\\right) \\in \\mathbb {R} ^ {T ^ {\\prime} \\times C}, \\tag {4}\n$$\n", + "text_format": "latex", + "bbox": [ + 555, + 779, + 903, + 797 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\mathbf {f} _ {i} = \\operatorname {C L I P} _ {\\mathrm {A D}} \\left(\\mathcal {F} _ {i}\\right). \\tag {5}\n$$\n", + "text_format": "latex", + "bbox": [ + 571, + 800, + 903, + 816 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "We then introduce the Contextual EMA to capture compact, discriminative visual features for distinctive AD generation. Contextual EMA. Expectation-Maximization Attention (EMA) [35] integrates the attention mechanism [81] into the classical EM [17] algorithm, which comprises three steps to", + "bbox": [ + 511, + 825, + 906, + 901 + ], + "page_idx": 3 + }, + { + "type": "page_number", + "text": "13574", + "bbox": [ + 480, + 944, + 519, + 955 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "estimate a more compact set of bases: Responsibility Estimation (RE), Likelihood Maximization (LM), and Data Re-estimation (DR). Inspired by this, we propose Contextual EMA to perform EMA on frames from $N$ contextual clips, aiming to eliminate redundancy, learn compact representations, and explore distinctiveness.", + "bbox": [ + 89, + 90, + 480, + 180 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Let $\\mathcal{H} = \\{\\mathbf{h}_{\\mathbf{x}_i}\\}_{i = 1}^N\\in \\mathbb{R}^{N\\times T'\\times C}$ represent $N$ clip vectors from the Perceiver, and $\\mathcal{M} = \\{\\mu_k\\}_{k = 1}^K\\in \\mathbb{R}^{K\\times C}$ denote the randomly initialized base features, where $C,K$ indicate the number of channel and bases. The RE step estimates the hidden variable $\\mathcal{Z} = \\{z_{nk}\\}_{n = 1,k = 1}^{N\\times T',K}$ , where the responsibility $z_{nk}$ represents the probability of the $n$ -th frame belonging to the $k$ -th base:", + "bbox": [ + 89, + 181, + 480, + 289 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\nz _ {n k} = \\frac {\\exp \\left(\\mathbf {h} _ {n} \\mu_ {k} ^ {T} / \\tau\\right)}{\\sum_ {j = 1} ^ {K} \\exp \\left(\\mathbf {h} _ {n} \\mu_ {j} ^ {T} / \\tau\\right)}, \\tag {6}\n$$\n", + "text_format": "latex", + "bbox": [ + 187, + 294, + 480, + 333 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "where $\\tau$ determines the shape (peakiness) of distribution $\\mathcal{Z}$ . Then, the LM step updates the bases $\\mathcal{M}$ by applying the weighted average on input $\\mathcal{H}$ , formulating the $k$ -th base as:", + "bbox": [ + 89, + 342, + 482, + 388 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\mu_ {k} = \\frac {\\sum_ {n = 1} ^ {N \\times T ^ {\\prime}} z _ {n k} \\mathbf {h} _ {n}}{\\sum_ {n = 1} ^ {N \\times T ^ {\\prime}} z _ {n k}}. \\tag {7}\n$$\n", + "text_format": "latex", + "bbox": [ + 209, + 392, + 480, + 433 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "The RE (E-step) and LM (M-step) are iteratively performed $R$ times until convergence. Notably, since bases number $K$ is much smaller than the embedding number $N \\times T'$ , we employ DR to reconstruct a compact version of $\\mathcal{H}$ through:", + "bbox": [ + 89, + 436, + 482, + 497 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\widehat {\\mathcal {H}} \\approx \\mathcal {Z M}. \\tag {8}\n$$\n", + "text_format": "latex", + "bbox": [ + 246, + 503, + 480, + 522 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Here, $\\widehat{\\mathcal{H}}\\in \\mathbb{R}^{N\\times T^{\\prime}\\times C}$ retains the same shape as $\\mathcal{H}$ . We combine $\\mathcal{H}$ and $\\widehat{\\mathcal{H}}$ element-wise with a hyperparameter $\\alpha$ .", + "bbox": [ + 89, + 526, + 482, + 559 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "To enhance representation distinctiveness, we introduce an additional branch using cross-attention between raw $\\mathcal{H}$ (query) and bases $\\mathcal{M}$ (key and value), formulated as:", + "bbox": [ + 89, + 560, + 482, + 604 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\widetilde {\\mathcal {H}} = \\operatorname {C r o s s A t t e n t i o n} (\\mathcal {H}, \\mathcal {M}), \\tag {9}\n$$\n", + "text_format": "latex", + "bbox": [ + 181, + 608, + 480, + 627 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "where $\\widetilde{\\mathcal{H}}\\in \\mathbb{R}^{N\\times T^{\\prime}\\times C}$ . Linear layers projecting queries, keys, and values are omitted in Fig. 3 (see Appendix §C for details). Through (9), we process the distributions of $\\mathcal{H}$ to attend on specific and informative bases, with improved linear separability (see Fig. 6). We combine $\\mathcal{H},\\widehat{\\mathcal{H}},\\widetilde{\\mathcal{H}}$ elementwise around Contextual EMA to construct the final refined visual features. These features are then projected into the LLM embedding space using a single-layer projector:", + "bbox": [ + 89, + 638, + 482, + 760 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {H} _ {\\text {s u m}} = \\operatorname {P r o j} (\\mathcal {H} + \\alpha \\widehat {\\mathcal {H}} + \\beta \\widetilde {\\mathcal {H}}). \\tag {10}\n$$\n", + "text_format": "latex", + "bbox": [ + 176, + 765, + 480, + 782 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Interleaved prompt as LLM's input. Following previous studies [22, 74, 82], we build our interleaved prompt enriched with character information, (see Fig. 3). To answer the \"who is who\" question when more than two characters are present, the corresponding actors' portrait images are projected as face tokens for reasoning. The $\\langle \\text{BOS} \\rangle$ tag appended at the end indicates the beginning of AD generation.", + "bbox": [ + 89, + 794, + 482, + 900 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Distinctive words highlighting. Our goal is to query a frozen LLM for AD generation using a vision-conditioned prompt. The typical supervision employs the commonly used auto-regressive loss function:", + "bbox": [ + 511, + 90, + 903, + 151 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {L} _ {\\text {a u t o}} = - \\sum_ {n} \\log P _ {\\theta} \\left(w _ {n} \\mid \\text {p r o m p t}; w _ {< n}\\right), \\tag {11}\n$$\n", + "text_format": "latex", + "bbox": [ + 568, + 159, + 903, + 189 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "where $w_{n}$ is the $n$ -th token from the target AD. However, $\\mathcal{L}_{auto}$ does not emphasize the distinctiveness specific to the current AD, which is our focus. To address this, we propose a distinctive word set $w_{d}$ , created by filtering out duplicates, such as character names, prepositions, and pronouns, from the $N$ context ADs of the target AD. During training, we explicitly encourage the LLM to predict the distinctive words in $w_{d}$ by optimizing the distinctive loss $\\mathcal{L}_{dist}$ :", + "bbox": [ + 511, + 198, + 903, + 319 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {L} _ {\\text {d i s t}} = - \\sum_ {n = 1} ^ {N} \\sum_ {i = 1} ^ {u} \\log P _ {\\theta} \\left(w _ {n} = w _ {d} ^ {i} \\mid \\text {p r o m p t}, w _ {< n}\\right), \\tag {12}\n$$\n", + "text_format": "latex", + "bbox": [ + 521, + 321, + 903, + 362 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "where $w_{d}^{i}$ denotes the $i$ -th distinctive word in $w_{d}$ and $u$ is the size of the set. The final complete loss function for Stage-II is: $\\mathcal{L}_{\\mathrm{II}} = \\mathcal{L}_{\\mathrm{auto}} + \\mathcal{L}_{\\mathrm{dist}}$ .", + "bbox": [ + 511, + 372, + 903, + 417 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "4. Experiments", + "text_level": 1, + "bbox": [ + 513, + 430, + 643, + 446 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "4.1. Experiment Setup", + "text_level": 1, + "bbox": [ + 513, + 454, + 689, + 470 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Datasets. We follow the AD generation benchmark established in AutoAD [21], conducting experiments on the denoised MAD-v2-Named [68] and evaluating on MAD-Eval-Named split. Specifically, MAD-v2-Named includes $\\sim 330\\mathrm{k}$ ADs from 488 movies for training and MAD-Eval has 6,520 ADs crawled from 10 movies for evaluation. We also evaluate on two recently introduced datasets. CMD-AD [23] (where \"CMD\" stands for Condensed Movie Dataset [7]) is a movie AD dataset that contains 101k ADs for more than 1432 movies, with 100 movies split for CMD-AD evaluation. TV-AD [82] is a recently proposed AD dataset based on TVQA [30], which contains $\\sim 31\\mathrm{k}$ ADs for training and $\\sim 3\\mathrm{k}$ ADs for evaluation.", + "bbox": [ + 511, + 477, + 903, + 672 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Evaluation Metrics. Classic captioning metrics including ROUGE-L [37], CIDEr [72] and SPICE [6] are reported to evaluate the quality of generated ADs versus the ground-truth. Besides, we also report Recall@k within $N$ Neighbours [22] ( $\\mathbf{R}@\\mathbf{k} / \\mathbf{N}$ ), which calculates the average value of Recall@k for each AD with its $N$ temporally adjacent GT texts, where BertScore [88] is used for text similarity matching. The R@k/N metric is based on retrieving the most similar ground-truth AD among N neighbors, and thus highlights the distinctiveness of generated ADs directly. LLM-AD-eval [23] employs LLMs to assess the quality of generated ADs, assigning scores from 1 (lowest) to 5 (highest). We utilize the LLM prompt from the original study [82] and apply open-source models LLaMA2-7B-Chat [70] and LLaMA3-8B-Instruct [4] for this evaluation.", + "bbox": [ + 511, + 674, + 903, + 900 + ], + "page_idx": 4 + }, + { + "type": "page_number", + "text": "13575", + "bbox": [ + 480, + 944, + 517, + 955 + ], + "page_idx": 4 + }, + { + "type": "table", + "img_path": "images/4f3578a5c7b2fece87a3048126aa726430d17f4243d49dce85cb212755e38c53.jpg", + "table_caption": [], + "table_footnote": [], + "table_body": "
MethodPub.VLMLLMROUGE-LCIDErSPICER@5/16
Training-free
VLog [2]--GPT-47.51.32.142.3
MM-Vid [39]ArXiv'23GPT-4V-9.86.13.846.1
MM-Narrator [86]CVPR'24CLIP-L14GPT-413.413.95.249.0
LLM-AD [14]ArXiv'24GPT-4V-13.520.5--
AutoAD-Zero [82]ACCV'24VideoLLaMA2-7BLLaMA3-8B-22.4--
Partial-fine-tuning
ClipCap [50]ArXiv'21CLIP-B32GPT-28.54.41.1-
CapDec [52]ArXiv'22--8.26.71.4-
AutoAD-I [21]CVPR'23CLIP-B32GPT-211.914.34.442.1
AutoAD-II [22]ICCV'23CLIP-B32GPT-213.419.5-50.8
AutoAD-III [23]CVPR'24EVA-CLIPOPT-2.7B-22.8-52.0
AutoAD-III [23]CVPR'24EVA-CLIPLLaMA2-7B-24.0-52.8
MovieSeq [40]ECCV'24CLIP-B16LLaMA2-7B*15.524.47.051.6
DistinctAD (Ours)CLIP-B32GPT-215.424.56.749.8
DistinctAD (Ours)CLIPAD-B32GPT-216.425.57.451.7
DistinctAD (Ours)CLIPAD-B16LLaMA2-7B17.227.08.255.6
DistinctAD (Ours)CLIPAD-B16LLaMA3-8B17.627.38.356.0
", + "bbox": [ + 147, + 88, + 849, + 356 + ], + "page_idx": 5 + }, + { + "type": "table", + "img_path": "images/09ec3956087a4e1192315246e224c20b24d0596867f667b733ec2abe3de3a7c1.jpg", + "table_caption": [ + "Table 1. Comparisons of AD performance on the MAD-Eval benchmark. * indicates fine-tuning LLaMA2-7B model with LoRA [24]. ${\\mathrm{{CLIP}}}_{\\mathrm{{AD}}}$ is our CLIP vision encoder adapted using our Stage-I strategy." + ], + "table_footnote": [], + "table_body": "
MethodCIDErR@1/5LLM-AD-eval
Video-BLIP2 [85]4.822.01.89
Video-LLaMA2 [87]5.223.61.91
AutoAD-II [22]13.526.12.08
AutoAD-III [23]21.730.02.85
AutoAD-Zero [82]17.7-2.83
DistinctAD (LLaMA2)22.332.92.89
DistinctAD (LLaMA3)22.733.02.88
AutoAD-III† [23]25.031.22.89
2.01
", + "bbox": [ + 125, + 402, + 444, + 517 + ], + "page_idx": 5 + }, + { + "type": "table", + "img_path": "images/d72d3da9bb1daec7802dcb790dd06a53214cf72fcd9fa2c2511acf9a98827fc9.jpg", + "table_caption": [ + "Table 2. Comparisons on CMD-AD. The LLM-AD-eval scores are evaluated with LLaMA2-7B (left) and LLaMA3-8B (right). $\\dagger$ indicates pre-training on 3.4M HowTo-AD dataset [23, 48]." + ], + "table_footnote": [], + "table_body": "
MethodCIDErR@1/5LLM-AD-eval
AutoAD-III [23]26.1-2.78 | 1.99
AutoAD-Zero [82]22.630.62.94 | 2.00
DistinctAD (LLaMA2)27.231.82.86 | 2.01
DistinctAD (LLaMA3)27.432.12.89 | 2.00
", + "bbox": [ + 125, + 571, + 447, + 637 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Table 3. Comparisons on TV-AD. The LLM-AD-eval scores are evaluated using LLaMA2-7B (left) and LLaMA3-8B (right).", + "bbox": [ + 89, + 641, + 482, + 670 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Implementation Details. To facilitate CLIP-AD adaptation in Stage-I, we collect the original raw movies from MAD from platforms like Amazon Prime Video. See Appendix F for details. We fine-tune the CLIP Vision encoder for 5 epochs with a fixed learning rate 5e-5 using the Adam optimizer [28] in Stage-I, with a batch size of 512. In Stage-II, we use a batch of 8 sequences, each containing 16 consecutive video AD-pairs from a movie. For each video clip, 8 frames are uniformly sampled. We use the AdamW [44] optimizer to train our model for 10 epochs, with a cosine-decayed learning rate and linear warm-up. The learning rate is set to $10^{-4}$ for both GPT-2 and LLaMA models. For external character information, we directly use the inference results from AutoAD-Zero [82] as it gives current best face", + "bbox": [ + 89, + 688, + 482, + 900 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "recognition performance. We use CLIP-B32 and GPT-2 on MAD-Eval-Named for ablations unless specified otherwise.", + "bbox": [ + 511, + 405, + 903, + 436 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "4.2. Comparisons with previous methods", + "text_level": 1, + "bbox": [ + 511, + 449, + 828, + 465 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "We conduct comprehensive comparisons using the widely-adopted MAD-Eval benchmark [21] and two recently introduced AD datasets, CMD-AD [23] and TV-AD [82].", + "bbox": [ + 511, + 474, + 903, + 518 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Comparisons on MAD-Eval are shown in Tab. 1. We primarily categorize previous studies into Training-free and Partial-fine-tuning approaches, as described in §2. Our method is a Partial-fine-tuning method. When using the same CLIP-B32 and GPT-2, our proposed DistinctAD achieves a CIDEr score of 24.5, surpassing previous AutoAD-I [21] (CIDEr 14.3) and AutoAD-II [22] (CIDEr 19.5). With our Stage-I adapted CLIP vision encoders (denoted as $\\mathrm{CLIP}_{\\mathrm{AD}}$ ), we observe stable improvements across all metrics, e.g. 25.5 vs. 24.5 on CIDEr and 51.7 vs. 49.8 on recall, validating the effectiveness of our Stage-I strategy. Notably, DistinctAD with $\\mathrm{CLIP}_{\\mathrm{AD}}$ -B16 and LLaMA3-8B [4] achieves state-of-the-arts with CIDEr 27.3 and Recall@5/16 56.0. Our outstanding performance on R@k/N demonstrates DistinctAD's ability to generate distinctive ADs, well matching the uniqueness of the clip's contents.", + "bbox": [ + 511, + 521, + 905, + 763 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Looking at the training-free methods, despite the capabilities of advanced proprietary VLMs, e.g. GPT-4V [53], and LLMs, e.g. GPT-4 [3], the performance of training-free methods remains inferior to those employing partial-finetuning. This discrepancy likely arises from the unique characteristics of AD and movie data, which exhibit a significant domain gap from common vision language training data. As such, these data types were not encountered during the pretraining of proprietary large-scale models.", + "bbox": [ + 511, + 763, + 905, + 900 + ], + "page_idx": 5 + }, + { + "type": "page_number", + "text": "13576", + "bbox": [ + 480, + 944, + 517, + 955 + ], + "page_idx": 5 + }, + { + "type": "table", + "img_path": "images/b69b234dfc62a2522b13cf1b39b525e57be97f0364c99cceb7ffe7917eacd17e.jpg", + "table_caption": [], + "table_footnote": [], + "table_body": "
SettingCIDErR@5/16
None6.734.0
Global Lg8.236.6
Fine-grained Lf7.735.2
γLg + (1 - γ)Lf8.636.9
", + "bbox": [ + 106, + 88, + 336, + 169 + ], + "page_idx": 6 + }, + { + "type": "table", + "img_path": "images/5c367284857a31e23b1937b2572b76bdc355a866f45403dc2e7ddc25090e096d.jpg", + "table_caption": [ + "(a) Stage-I components." + ], + "table_footnote": [], + "table_body": "
Coefficient γCIDEr
0.18.0
0.38.5
0.58.6
0.77.7
", + "bbox": [ + 374, + 88, + 522, + 169 + ], + "page_idx": 6 + }, + { + "type": "table", + "img_path": "images/e7c6a12cf35f36ecff551cf5219c27bf8e9084b7a706eb461922b68410aa04e5.jpg", + "table_caption": [ + "(b) Impact of coefficient $\\gamma$" + ], + "table_footnote": [], + "table_body": "
PromptStage-ICIDErR@5/16
Contextual ADs [21]X12.6 (17.8)39.8 (43.1)
✓14.1 (19.0)39.9 (44.2)
Character [22]X22.045.6
✓23.146.2
", + "bbox": [ + 547, + 88, + 898, + 169 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Table 4. Ablation studies for Stage-I. (a) Evaluation of global video-AD loss $\\mathcal{L}_f$ and fine-grained frame-AD loss $\\mathcal{L}_f$ on AD performance. (b) Analysis of the impact of the coefficient $\\gamma$ . Both (a) and (b) are conducted with pure visual prompts. (c) Impact of Stage-I when combined with different prompts for the LLM decoder, including contextual ADs and character names. Performance in parentheses indicates results with ground-truth contextual ADs as prompts.", + "bbox": [ + 89, + 188, + 906, + 246 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/5e32d7ebd4bcc1f151229deb6c3474cbecff188ccbaaeb862f97f24cb61baffe.jpg", + "image_caption": [ + "Figure 4. Ablation studies for hyperparameter in Stage-II, with final settings highlighted in orange. (a) Impact of $\\alpha$ on the weight of compact representation $\\widehat{\\mathcal{H}}$ . (b) Influence of $\\beta$ on cross-attended feature $\\widetilde{\\mathcal{H}}$ . (c) Impact of $K$ , which denotes the number of clusters in bases $\\mathcal{M}$ . (d) Effect of sampling $N$ consecutive video clips. We switch to larger memory GPUs when $N$ exceeds 16." + ], + "image_footnote": [], + "bbox": [ + 93, + 250, + 480, + 319 + ], + "page_idx": 6 + }, + { + "type": "table", + "img_path": "images/43d56d3256641f2e3547a6f483f512f2bc87a7ea52840c2b8c432330c1b15e73.jpg", + "table_caption": [ + "(c) Impact of Stage-I w/ different prompts." + ], + "table_footnote": [], + "table_body": "
Ex#αHβHLdistCIDErR@5/16
A0XXX23.1 (27.4)46.2
B1✓XX23.7 (29.3)46.6
B2X✓X23.4 (29.1)46.1
B3✓✓X23.3 (28.1)48.0
C0XX✓24.7 (29.4)49.5
C1✓X✓24.3 (29.4)50.7
C2X✓✓25.3 (30.4)51.5
C3✓✓✓25.5 (29.8)51.7
", + "bbox": [ + 125, + 406, + 450, + 542 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Comparisons on CMD-AD and TV-AD. We further verify the generalizability of DistinctAD on the recently proposed CMD-AD and TV-AD benchmarks, with results presented in Tables 2 and 3. Both evaluations use adapted $\\mathrm{CLIP}_{\\mathrm{AD}}$ -B16 model. DistinctAD outperforms AutoAD-Zero, AutoAD-II and AutoAD-III in terms of CIDEr and R@1/5 on both benchmarks Meanwhile, DistinctAD exhibits a lower CIDEr compared to AutoAD-III† on CMD-AD, which we conjecture is primarily due to AutoAD-III† pre-training on a very large-scale 3.4M transformed HowTo-AD dataset [23, 48], which is currently publicly unavailable. Despite this, DistinctAD achieves superior R@1/5, underscoring its exceptional ability to generate distinctive and high-quality ADs. This is further corroborated by its leading performance on the LLM-AD-eval metric.", + "bbox": [ + 89, + 609, + 482, + 838 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "4.3. Ablation studies", + "text_level": 1, + "bbox": [ + 89, + 847, + 251, + 861 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Effect of CLIP-AD Adaptation (I). Tab. 4a demonstrates the benefit of our Stage-I strategy, i.e. adapting CLIP the vi", + "bbox": [ + 89, + 869, + 482, + 901 + ], + "page_idx": 6 + }, + { + "type": "table", + "img_path": "images/01faa9a5865d3f66a638bd67c828bce3b44df8046eaf6d59988102939191b5bb.jpg", + "table_caption": [ + "Table 5. Ablation studies for components in Stage-II. The CIDEr column shows scores with AutoAD-Zero's character [82] as prompt by default. CIDEr in parentheses represent performance with ground-truth character names." + ], + "table_footnote": [], + "table_body": "
Consecutive N?CIDErR@5/16
✓25.551.7
×23.8↓1.752.5↑0.8
", + "bbox": [ + 588, + 250, + 831, + 305 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Table 6. Impact of whether sampling $N$ consecutive clips or not.", + "bbox": [ + 514, + 309, + 900, + 323 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "sion encoder to the movie-AD domain via global video-AD matching $\\mathcal{L}_g$ and fine-grained frame-AD matching $\\mathcal{L}_f$ . In Tab. 4b, the balancing coefficient $\\gamma$ performs best at 0.5. In Tab. 4c, our Stage-I strategy consistently enhances performance when combined with different prompts in the LLM decoder, such as contextual ADs in AutoAD-I [21] or character names in AutoAD-II [22]. This indicates that our AD-adapted CLIP vision encoder can integrate seamlessly into previous methods, including those training-free models that utilize CLIP-based visual extractors.", + "bbox": [ + 511, + 338, + 903, + 489 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Effect of Distinctive AD narration (II). We evaluate the effectiveness of Stage-II components in Tab. 5, based on the default $\\mathcal{H}$ (Perceiver's output) and full AD auto-regressive loss $\\mathcal{L}_{\\text {auto }}$ . The baseline (A0) outperforms AutoAD-II in CIDEr (23.1 vs. 19.5), primarily due to more accurate character prompts from AutoAD-Zero and the adapted CLIP vision encoder from Stage-I. A0 with AutoAD-II's characters achieves a CIDEr score of 20.6, close to AutoAD-II's performance. Incorporating reconstructed feature $\\widehat{\\mathcal{H}}$ brings stable improvements on both CIDEr and recall (B1 & C1), highlighting the importance of compact representations in understanding visual semantics. Cross-attended feature $\\widetilde{\\mathcal{H}}$ works better together with distinctive word prediction loss $\\mathcal{L}_{\\text {dist }}$ (C2 vs. B2, C3 vs. B3). We conjecture this is because $\\mathcal{L}_{\\text {dist }}$ provides more definite supervision on re-weighting distinctive words, which guides $\\widehat{\\mathcal{H}}$ to attend on concept-related bases. Overall, applying the full Stage-II pipeline brings significant and robust performance (C3).", + "bbox": [ + 511, + 491, + 905, + 763 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Effect of Hyper-parameters. Fig. 4 summarizes the ablation studies on 4 hyper-parameters that potentially influence the results in Stage-II. Coefficient weights $\\alpha$ and $\\beta$ yield optimal results when set to 3 and 1, respectively. This suggests the need to refine our final representations to be more compact for generating ADs. Fig. 4(c) shows setting bases number $K$ to 32 yields best. A smaller $K$ , e.g. 2, can still achieve notable CIDEr, as the Contextual EMA module does not significantly alter the final output $\\mathcal{H}_{sum}$ . However,", + "bbox": [ + 511, + 763, + 906, + 901 + ], + "page_idx": 6 + }, + { + "type": "page_number", + "text": "13577", + "bbox": [ + 480, + 944, + 517, + 955 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/2ccc1aef0fa6873a1c16a4b22dfe62f6ed3a0a36aed0fd712fdb6583fc388f04.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 122, + 94, + 142, + 108 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "1h12m53.6s", + "text_level": 1, + "bbox": [ + 145, + 95, + 215, + 107 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/2d5c202bd5e428ef58c4655aed83eb2dcc2f4c3af0da9788eb29006af5f7bc2e.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 127, + 109, + 382, + 152 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "GT: Holding the room's phone to his ear, the man on the bed regards him sympathetically.", + "bbox": [ + 124, + 154, + 367, + 179 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "AutoAD-Zero: He looks at something.", + "bbox": [ + 125, + 180, + 334, + 193 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "DistinctAD: He holds the phone to his ear.", + "bbox": [ + 125, + 194, + 354, + 205 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/dc60bae16010cbbd6600786418ec0d51289a1534bb7887ee6fc3791de2b47696.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 406, + 95, + 423, + 108 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "1h12m59.1s", + "text_level": 1, + "bbox": [ + 428, + 95, + 496, + 107 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/7327c6f13a5c90c45f2bc5d23e61020a8d28786593f750d639f35b1183c4c6b1.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 408, + 109, + 665, + 152 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "GT: Stephen sees the pill bottles on a bedside table.", + "bbox": [ + 408, + 154, + 612, + 178 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "AutoAD-Zero: Stephen looks at something.", + "bbox": [ + 410, + 180, + 645, + 193 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "DistinctAD: In Stephen's bedroom, a pill bottle sits on his nightstand.", + "bbox": [ + 410, + 194, + 633, + 218 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/e1b7ef6e20b0a2526f1d30ade4570b895b78d28f8a44e85883894cbc63216b9e.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 689, + 95, + 709, + 108 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "1h13m14.5s", + "text_level": 1, + "bbox": [ + 712, + 95, + 779, + 107 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/ebd129b9530fb77b259d18208c403a32fb2c50a1c482de679bb742a4aa972e90.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 689, + 109, + 861, + 152 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "GT: The man on the bed sits in silhouette against the room's soft lamp light.", + "bbox": [ + 650, + 154, + 864, + 179 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "AutoAD-Zero: He turns to leave, looking at him.", + "bbox": [ + 658, + 180, + 839, + 204 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "DistinctAD: The bed is empty.", + "bbox": [ + 658, + 205, + 823, + 218 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/37954c648a97d7f8189ac87b46a889a7b7a69cb60e18444612d6b9015bb9c08b.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 124, + 218, + 143, + 232 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "1h13m19.4s", + "text_level": 1, + "bbox": [ + 145, + 220, + 215, + 231 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/81ddb833fd820f8365e260173580a31af48f41d367e4ed7f01440bd29bb08c2f.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 127, + 234, + 297, + 276 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "GT: Later, Stephen steps back out into the rainy night.", + "bbox": [ + 124, + 279, + 289, + 303 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "AutoAD-Zero: Stephen looks around and at another character.", + "bbox": [ + 124, + 305, + 297, + 330 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "DistinctAD: As Stephen walks away, he glances over his shoulder.", + "bbox": [ + 124, + 330, + 313, + 356 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/aa5955ececbf0391df747327102a561453c74d813506b3ca97e40903a178655f.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 323, + 219, + 343, + 232 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "1h13m23.1s", + "text_level": 1, + "bbox": [ + 344, + 220, + 411, + 231 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/23068bb9811e451a9448d946e9582d6e135b846bdb82bfa1466d538458be3d6a.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 323, + 234, + 495, + 276 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "GT: Moving with a stiff gait, Stephen marches across the parking lot to his car.", + "bbox": [ + 321, + 279, + 544, + 303 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "AutoAD-Zero: He looks out at the city.", + "bbox": [ + 321, + 305, + 535, + 319 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "DistinctAD: He looks up at the sky, then turns and heads back to the car.", + "bbox": [ + 321, + 321, + 542, + 345 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/12c637b2b01d4f5a6b888cc9353072a41356abd721a8166d5ed3be3d7650980c.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 545, + 219, + 563, + 232 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "h13m24.5s", + "text_level": 1, + "bbox": [ + 545, + 220, + 607, + 232 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/2d337008c1aa526bc9bc5fe5aece4ce62e57c4e8c4940a65503b7c1552eec619.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 521, + 234, + 861, + 276 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "GT: As Stephen gets in, rain patters heavily against his windshield.", + "bbox": [ + 560, + 279, + 841, + 303 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "AutoAD-Zero: Stephen looks at the women.", + "bbox": [ + 560, + 305, + 803, + 318 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "DistinctAD: In the car, Stephen stares out the window.", + "bbox": [ + 560, + 321, + 856, + 334 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": ": The Ides of March (2011)", + "bbox": [ + 643, + 344, + 854, + 357 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "unsuitable values of $K$ , e.g. 8 or 64, can negatively impact performance. Fig. 4(d) reveals that increasing the number of consecutive clips $N$ (with $K$ set to 32) enhances the CIDEr score, though this effect saturates when $N$ exceeds 16. This demonstrates that more bases should be created to effectively summarize components with additional clips.", + "bbox": [ + 89, + 411, + 483, + 500 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Do the $N$ clips to be consecutive? Tab. 6 presents the results of sampling non-consecutive $N$ clips during training. When using non-continuous clips, we observe a decrease in the CIDEr metric by 1.7 (25.5 vs. 23.8) because the Contextual EMA module struggles with unrelated contexts. However, the R@5/16 improves by 0.8, which indicates enhancement of the distinctiveness (uniqueness) of the generated AD when using more diverse visual contents. Visualizations. To better understand what Contextual EMA learns, we show the t-SNE [71] visualizations of $\\mathcal{H}$ , $\\widehat{\\mathcal{H}}$ and $\\widetilde{\\mathcal{H}}$ (from §3.2) in Fig. 6, using the same perplexity value across all visualizations. With Contextual EMA, $\\widehat{\\mathcal{H}}$ exhibits more compact features compared to raw $\\mathcal{H}$ , Fig. 6(b). Interestingly in Fig. 6(c), cross-attention between $\\mathcal{H}$ and bases $\\mathcal{M}$ produces strip-like feature distributions pointing to specific base centers, enhancing contextual distinctiveness with improved linear separability and interpretability.", + "bbox": [ + 89, + 502, + 483, + 762 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "4.4. Qualitative results", + "text_level": 1, + "bbox": [ + 89, + 770, + 267, + 786 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Fig. 5 presents qualitative examples of our model. We compare the predictions of DistinctAD (using LLaMA3-8B) with ground-truth captions (GT) and publicly available AutoAD-Zero [82] outputs. Note that clips are sampled consecutively in time. Previous studies often struggle with similar contextual clips, such as those featuring closely-related scenes and characters, by repeating correct yet in", + "bbox": [ + 89, + 794, + 483, + 901 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/29d662c4b957db5ba2e1a973477e675137f4348e9fd76de5e80fde7457094325.jpg", + "image_caption": [ + "Figure 5. Qualitative results. We present ground-truth (GT) ADs, publicly released AutoAD-Zero outputs, and our DistinctAD predictions for several temporally consecutive movie clips. Movie frames are taken from The Ides of March (2011) [15]. Zoom in for details.", + "(a) $\\mathcal{H}$", + "Figure 6. Visualizations of Contextual EMA. (a) A set of randomly generated 3D data $\\mathcal{H}$ , sampled from $N$ types of samples. (b) Compact features $\\widehat{\\mathcal{H}}$ obtained via Data Re-estimation (DR). (c) Cross-attention outputs $\\widetilde{\\mathcal{H}}$ between $\\mathcal{H}$ and bases $\\mathcal{M}$ ." + ], + "image_footnote": [], + "bbox": [ + 522, + 404, + 640, + 494 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/bc6c45c84779a97c19f85519f04ab16ddccf6912105013878a04de987a39702d.jpg", + "image_caption": [ + "(b) $\\widehat{\\mathcal{H}}$" + ], + "image_footnote": [], + "bbox": [ + 651, + 404, + 767, + 494 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/00f00d71e4482d9a9a0525243439f9f672dfdaabb1ce2a4def9aaa4e07519078.jpg", + "image_caption": [ + "(c) $\\widetilde{\\mathcal{H}}$" + ], + "image_footnote": [], + "bbox": [ + 779, + 405, + 895, + 493 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "significant action words, e.g. \"look\". In contrast, our DistinctAD effectively generates more engaging ADs by identifying distinctive objects in adjacent clips, e.g. \"phone\", \"pill bottle\", and \"car\", along with corresponding more specific behaviors. More examples can be found in Appendix §E.", + "bbox": [ + 511, + 574, + 906, + 650 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "5. Conclusion", + "text_level": 1, + "bbox": [ + 511, + 662, + 633, + 679 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "This paper proposes DistinctAD, a two-stage framework for generating distinctive audio descriptions for better narrative. By addressing the domain gap between movie-AD data with a CLIP-AD adaptation strategy, and introducing a Contextual EMA module and a distinctive word prediction loss, our approach significantly improves the quality of AD generation. The effectiveness of DistinctAD is demonstrated through comprehensive evaluations on multiple benchmark datasets and ablations studies. Despite these promising results, DistinctAD is still limited by requiring numbers of parameters and the quality of the generated ADs still falls short of human annotations. Overall, automatic AD generation still remains a challenging task, and there is considerable scope for future advancements in this field.", + "bbox": [ + 511, + 689, + 906, + 900 + ], + "page_idx": 7 + }, + { + "type": "page_number", + "text": "13578", + "bbox": [ + 480, + 944, + 517, + 955 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Acknowledgement", + "text_level": 1, + "bbox": [ + 91, + 90, + 250, + 107 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "This work was supported by a Strategic Research Grant from City University of Hong Kong (Project No. 7005840).", + "bbox": [ + 89, + 113, + 483, + 143 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "References", + "text_level": 1, + "bbox": [ + 91, + 169, + 187, + 184 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[1] https://audiovault.net, 2019.3", + "[2] Vlog: Video as a long document. https://github.com/showlab/VLog, 2023. GitHub repository. 6", + "[3] Josh Achiam, Steven Adler, Sandhini Agarwal, Lama Ahmad, Ilge Akkaya, Florencia Leoni Aleman, Diogo Almeida, Janko Altenschmidt, Sam Altman, Shyamal Anadkat, et al. Gpt-4 technical report. arXiv preprint arXiv:2303.08774, 2023. 1, 2, 6", + "[4] AI@Meta. Llama 3 model card. 2024. 5, 6, 2", + "[5] Jean-Baptiste Alayrac, Jeff Donahue, Pauline Luc, Antoine Miech, Iain Barr, Yana Hasson, Karel Lenc, Arthur Mensch, Katherine Millican, Malcolm Reynolds, et al. Flamingo: a visual language model for few-shot learning. NeurIPS, 35: 23716-23736, 2022. 1, 3, 4", + "[6] Peter Anderson, Basura Fernando, Mark Johnson, and Stephen Gould. Spice: Semantic propositional image caption evaluation. In ECCV, pages 382-398. Springer, 2016. 5", + "[7] Max Bain, Arsha Nagrani, Andrew Brown, and Andrew Zisserman. Condensed movies: Story based retrieval with contextual embeddings. In ACCV, 2020. 5", + "[8] Max Bain, Jaesung Huh, Tengda Han, and Andrew Zisserman. Whisperx: Time-accurate speech transcription of long-form audio. arXiv preprint arXiv:2303.00747, 2023. 2", + "[9] Carmen J Branje and Deborah I Fels. Livedescribe: can amateur descriptors create high-quality audio description? Journal of Visual Impairment & Blindness, 106(3):154-165, 2012. 1, 2", + "[10] Hervé Bredin and Antoine Laurent. End-to-end speaker segmentation for overlap-aware resegmentation. In Interspeech, 2021. 2", + "[11] Hervé Bredin, Ruiqing Yin, Juan Manuel Coria, Gregory Gelly, Pavel Korshunov, Marvin Lavechin, Diego Fustes, Hadrien Titeux, Wassim Bouaziz, and Marie-Philippe Gill. Pyannote. audio: neural building blocks for speaker diarization. In ICASSP, pages 7124–7128. IEEE, 2020. 2", + "[12] Fuhai Chen, Rongrong Ji, Xiaoshuai Sun, Yongjian Wu, and Jinsong Su. Groupcap: Group-based image captioning with structured relevance and diversity constraints. In CVPR, pages 1345-1353, 2018. 3", + "[13] Shaoxiang Chen and Yu-Gang Jiang. Towards bridging event captioner and sentence localizer for weakly supervised dense event captioning. In CVPR, pages 8425-8435, 2021. 2", + "[14] Peng Chu, Jiang Wang, and Andre Abrantes. Llm-ad: Large language model based audio description system. arXiv preprint arXiv:2405.00983, 2024. 1, 2, 6", + "[15] George Clooney. The idees of march. *Columbia Pictures*, 2011. 8", + "[16] Bo Dai and Dahua Lin. Contrastive learning for image captioning. NeurIPS, 30, 2017. 3" + ], + "bbox": [ + 93, + 195, + 482, + 900 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[17] Arthur P Dempster, Nan M Laird, and Donald B Rubin. Maximum likelihood from incomplete data via the em algorithm. Journal of the royal statistical society: series B (methodological), 39(1):1-22, 1977. 2, 4", + "[18] Chaorui Deng, Shizhe Chen, Da Chen, Yuan He, and Qi Wu. Sketch, ground, and refine: Top-down dense video captioning. In CVPR, pages 234-243, 2021. 2", + "[19] Lisa Dunlap, Yuhui Zhang, Xiaohan Wang, Ruiqi Zhong, Trevor Darrell, Jacob Steinhardt, Joseph E Gonzalez, and Serena Yeung-Levy. Describing differences in image sets with natural language. In CVPR, pages 24199-24208, 2024. 3", + "[20] Louise Fryer. An introduction to audio description: A practical guide, 2016. 1", + "[21] Tengda Han, Max Bain, Arsha Nagrani, Gül Varol, Weidi Xie, and Andrew Zisserman. Autoad: Movie description in context. In CVPR, pages 18930-18940, 2023. 1, 2, 3, 4, 5, 6, 7", + "[22] Tengda Han, Max Bain, Arsha Nagrani, Gul Varol, Weidi Xie, and Andrew Zisserman. Autoad ii: The sequel-who, when, and what in movie audio description. In ICCV, pages 13645–13655, 2023. 1, 3, 4, 5, 6, 7", + "[23] Tengda Han, Max Bain, Arsha Nagrani, Gül Varol, Weidi Xie, and Andrew Zisserman. Autoad iii: The prequel-back to the pixels. In CVPR, pages 18164-18174, 2024. 1, 2, 3, 5, 6, 7", + "[24] Edward J Hu, Yelong Shen, Phillip Wallis, Zeyuan Allen-Zhu, Yuanzhi Li, Shean Wang, Lu Wang, and Weizhu Chen. Lora: Low-rank adaptation of large language models. arXiv preprint arXiv:2106.09685, 2021. 6", + "[25] Vladimir Iashin and Esa Rahtu. A better use of audio-visual cues: Dense video captioning with bi-modal transformer. In BMVC, 2020. 2", + "[26] Vladimir Iashin and Esa Rahtu. Multi-modal dense video captioning. In CVPRW, pages 958-959, 2020. 2", + "[27] Peng Jin, Jina Huang, Fenglin Liu, Xian Wu, Shen Ge, Guoli Song, David Clifton, and Jie Chen. Expectation-maximization contrastive learning for compact video-and-language representations. NeurIPS, 35:30291-30306, 2022. 2", + "[28] Diederik P Kingma. Adam: A method for stochastic optimization. arXiv preprint arXiv:1412.6980, 2014. 6", + "[29] Ranjay Krishna, Kenji Hata, Frederic Ren, Li Fei-Fei, and Juan Carlos Niebles. Dense-captioning events in videos. In ICCV, pages 706–715, 2017. 2", + "[30] Jie Lei, Licheng Yu, Mohit Bansal, and Tamara L Berg. Tvqa: Localized, compositional video question answering. In EMNLP, 2018. 5", + "[31] Elisa Lewis. Deep dive: How audio description benefits everyone, 2021. Accessed on, pages 11-13, 2023. 1", + "[32] Bo Li, Yuanhan Zhang, Liangyu Chen, Jinghao Wang, Fanyi Pu, Jingkang Yang, Chunyuan Li, and Ziwei Liu. Mimicit: Multi-modal in-context instruction tuning. arXiv preprint arXiv:2306.05425, 2023. 3", + "[33] Bo Li, Yuanhan Zhang, Liangyu Chen, Jinghao Wang, Jingkang Yang, and Ziwei Liu. Otter: a multi-modal model with in-context instruction tuning. corr abs/2305.03726 (2023), 2023. 3" + ], + "bbox": [ + 516, + 92, + 905, + 898 + ], + "page_idx": 8 + }, + { + "type": "page_number", + "text": "13579", + "bbox": [ + 480, + 944, + 519, + 955 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[34] Junnan Li, Dongxu Li, Silvio Savarese, and Steven Hoi. Blip-2: Bootstrapping language-image pre-training with frozen image encoders and large language models. arXiv preprint arXiv:2301.12597, 2023. 1, 3", + "[35] Xia Li, Zhisheng Zhong, Jianlong Wu, Yibo Yang, Zhouchen Lin, and Hong Liu. Expectation-maximization attention networks for semantic segmentation. In ICCV, pages 9167-9176, 2019. 2, 4", + "[36] Yehao Li, Ting Yao, Yingwei Pan, Hongyang Chao, and Tao Mei. Jointly localizing and describing events for dense video captioning. In CVPR, pages 7492-7500, 2018. 2", + "[37] Chin-Yew Lin. Rouge: A package for automatic evaluation of summaries. In Text summarization branches out, pages 74-81, 2004. 5", + "[38] Kevin Lin, Linjie Li, Chung-Ching Lin, Faisal Ahmed, Zhe Gan, Zicheng Liu, Yumao Lu, and Lijuan Wang. Swinbert: End-to-end transformers with sparse attention for video captioning. In CVPR, pages 17949-17958, 2022. 2", + "[39] Kevin Lin, Faisal Ahmed, Linjie Li, Chung-Ching Lin, Ehsan Azarnasab, Zhengyuan Yang, Jianfeng Wang, Lin Liang, Zicheng Liu, Yumao Lu, et al. Mm-vid: Advancing video understanding with gpt-4v (ision). arXiv preprint arXiv:2310.19773, 2023. 1, 2, 6", + "[40] Kevin Qinghong Lin, Pengchuan Zhang, Difei Gao, Xide Xia, Joya Chen, Ziteng Gao, Jinheng Xie, Xuhong Xiao, and Mike Zheng Shou. Learning video context as interleaved multimodal sequences. arXiv preprint arXiv:2407.21757, 2024. 1, 3, 6", + "[41] Zhihui Lin, Tianyu Yang, Maomao Li, Ziyu Wang, Chun Yuan, Wenhao Jiang, and Wei Liu. Swem: Towards real-time video object segmentation with sequential weighted expectation-maximization. In CVPR, pages 1362-1372, 2022. 2", + "[42] Haotian Liu, Chunyuan Li, Qingyang Wu, and Yong Jae Lee. Visual instruction tuning. NeurIPS, 36, 2024. 1", + "[43] Xihui Liu, Hongsheng Li, Jing Shao, Dapeng Chen, and Xiaogang Wang. Show, tell and discriminate: Image captioning by self-retrieval with partially labeled data. In ECCV, pages 338-354, 2018. 3", + "[44] I Loshchilov. Decoupled weight decay regularization. arXiv preprint arXiv:1711.05101, 2017. 6", + "[45] Huaishao Luo, Lei Ji, Botian Shi, Haoyang Huang, Nan Duan, Tianrui Li, Jason Li, Taroon Bharti, and Ming Zhou. Univl: A unified video and language pre-training model for multimodal understanding and generation. arXiv preprint arXiv:2002.06353, 2020. 2", + "[46] Huaishao Luo, Lei Ji, Ming Zhong, Yang Chen, Wen Lei, Nan Duan, and Tianrui Li. Clip4clip: An empirical study of clip for end to end video clip retrieval and captioning. Neurocomputing, 508:293-304, 2022. 3", + "[47] Ruotian Luo, Brian Price, Scott Cohen, and Gregory Shakhnarovich. Discriminability objective for training descriptive captions. In CVPR, pages 6964-6974, 2018. 3", + "[48] Antoine Miech, Dimitri Zhukov, Jean-Baptiste Alayrac, Makarand Tapaswi, Ivan Laptev, and Josef Sivic. Howto100m: Learning a text-video embedding by watching hundred million narrated video clips. In ICCV, pages 2630-2640, 2019. 6, 7" + ], + "bbox": [ + 91, + 90, + 482, + 898 + ], + "page_idx": 9 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[49] Antoine Miech, Jean-Baptiste Alayrac, Lucas Smaira, Ivan Laptev, Josef Sivic, and Andrew Zisserman. End-to-end learning of visual representations from uncurated instructional videos. In CVPR, pages 9879-9889, 2020. 4", + "[50] Ron Mokady, Amir Hertz, and Amit H Bermano. Clipcap: Clip prefix for image captioning. arXiv preprint arXiv:2111.09734, 2021. 6", + "[51] Jonghwan Mun, Linjie Yang, Zhou Ren, Ning Xu, and Bohyung Han. Streamlined dense video captioning. In CVPR, pages 6588-6597, 2019. 2", + "[52] David Nukrai, Ron Mokady, and Amir Globerson. Text-only training for image captioning using noise-injected clip. arXiv preprint arXiv:2211.00575, 2022. 6", + "[53] OpenAI. Gpt-4v(ision) system card. 2023. 1, 2, 6", + "[54] Amy Pavel, Gabriel Reyes, and Jeffrey P Bigham. Describe: Authoring and automatically editing audio descriptions. In Proceedings of the 33rd Annual ACM Symposium on User Interface Software and Technology, pages 747-759, 2020. 1, 2", + "[55] Elisa Perego. Gains and losses of watching audio described films for sighted viewers. *Target*, 28(3):424-444, 2016. 1", + "[56] Haran Raajesh, Naveen Reddy Desanur, Zeeshan Khan, and Makarand Tapaswi. Micap: A unified model for identity-aware movie descriptions. In CVPR, pages 14011-14021, 2024. 1, 4", + "[57] Alec Radford, Jeffrey Wu, Rewon Child, David Luan, Dario Amodei, Ilya Sutskever, et al. Language models are unsupervised multitask learners. OpenAI blog, 1(8):9, 2019. 2", + "[58] Alec Radford, Jong Wook Kim, Chris Hallacy, Aditya Ramesh, Gabriel Goh, Sandhini Agarwal, Girish Sastry, Amanda Askell, Pamela Mishkin, Jack Clark, et al. Learning transferable visual models from natural language supervision. In ICML, pages 8748-8763. PMLR, 2021. 2", + "[59] Tanzila Rahman, Bicheng Xu, and Leonid Sigal. Watch, listen and tell: Multi-modal weakly supervised dense event captioning. In ICCV, pages 8908-8917, 2019. 2", + "[60] Anna Rohrbach, Marcus Rohrbach, Niket Tandon, and Bernt Schiele. A dataset for movie description. In CVPR, pages 3202-3212, 2015. 2", + "[61] Anna Rohrbach, Atousa Torabi, Marcus Rohrbach, Niket Tandon, Christopher Pal, Hugo Larochelle, Aaron Courville, and Bernt Schiele. Movie description. *IJCV*, 123:94–120, 2017. 2", + "[62] Paul Hongsuck Seo, Arsha Nagrani, Anurag Arnab, and Cordelia Schmid. End-to-end generative pretraining for multimodal video captioning. In CVPR, pages 17959-17968, 2022. 2", + "[63] Piyush Sharma, Nan Ding, Sebastian Goodman, and Radu Soricut. Conceptual captions: A cleaned, hypernymed, image alt-text dataset for automatic image captioning. In ACL, pages 2556-2565, 2018. 2", + "[64] Zhiqiang Shen, Jianguo Li, Zhou Su, Minjun Li, Yurong Chen, Yu-Gang Jiang, and Xiangyang Xue. Weakly supervised dense video captioning. In CVPR, pages 1916-1924, 2017. 2", + "[65] Botian Shi, Lei Ji, Yaobo Liang, Nan Duan, Peng Chen, Zhendong Niu, and Ming Zhou. Dense procedure captioning" + ], + "bbox": [ + 514, + 90, + 903, + 901 + ], + "page_idx": 9 + }, + { + "type": "page_number", + "text": "13580", + "bbox": [ + 480, + 944, + 519, + 955 + ], + "page_idx": 9 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "in narrated instructional videos. In ACL, pages 6382-6391, 2019. 2", + "[66] Aleksandar Shtedritski, Christian Rupprecht, and Andrea Vedaldi. What does clip know about a red circle? visual prompt engineering for vlms. In ICCV, pages 11987-11997, 2023. 3", + "[67] Joel Snyder. Audio description: The visual made verbal. In International congress series, pages 935-939. Elsevier, 2005. 1", + "[68] Mattia Soldan, Alejandro Pardo, Juan León Alcazar, Fabian Caba, Chen Zhao, Silvio Giancola, and Bernard Ghanem. Mad: A scalable dataset for language grounding in videos from movie audio descriptions. In CVPR, pages 5026-5035, 2022. 2, 5, 1, 4", + "[69] Atousa Torabi, Christopher Pal, Hugo Larochelle, and Aaron Courville. Using descriptive video services to create a large data source for video annotation research. arXiv preprint arXiv:1503.01070, 2015. 2", + "[70] Hugo Touvron, Louis Martin, Kevin Stone, Peter Albert, Amjad Almahairi, Yasmine Babaei, Nikolay Bashlykov, Soumya Batra, Prajjwal Bhargava, Shruti Bhosale, et al. Llama 2: Open foundation and fine-tuned chat models. arXiv preprint arXiv:2307.09288, 2023. 5", + "[71] Laurens Van der Maaten and Geoffrey Hinton. Visualizing data using t-sne. JMLR, 9(11), 2008. 8", + "[72] Ramakrishna Vedantam, C Lawrence Zitnick, and Devi Parikh. Cider: Consensus-based image description evaluation. In CVPR, pages 4566-4575, 2015. 5", + "[73] Gilad Vered, Gal Oren, Yuval Atzmon, and Gal Chechik. Joint optimization for cooperative image captioning. In ICCV, pages 8898-8907, 2019. 3", + "[74] Hanlin Wang, Zhan Tong, Kecheng Zheng, Yujun Shen, and Limin Wang. Contextual ad narration with interleaved multimodal sequence. arXiv preprint arXiv:2403.12922, 2024. 1, 2, 3, 4, 5", + "[75] Jingwen Wang, Wenhao Jiang, Lin Ma, Wei Liu, and Yong Xu. Bidirectional attentive fusion with context gating for dense video captioning. In CVPR, pages 7190-7198, 2018. 2", + "[76] Jiuniu Wang, Wenjia Xu, Qingzhong Wang, and Antoni B Chan. Compare and reweight: Distinctive image captioning using similar images sets. In ECCV, pages 370-386. Springer, 2020. 3", + "[77] Jiuniu Wang, Wenjia Xu, Qingzhong Wang, and Antoni B Chan. Group-based distinctive image captioning with memory attention. In ACMMM, pages 5020-5028, 2021.", + "[78] Jiuniu Wang, Wenjia Xu, Qingzhong Wang, and Antoni B Chan. On distinctive image captioning via comparing and reweighting. TPAMI, 45(2):2088-2103, 2022. 3", + "[79] Teng Wang, Huicheng Zheng, Mingjing Yu, Qian Tian, and Haifeng Hu. Event-centric hierarchical representation for dense video captioning. TCVT, 31(5):1890-1900, 2020. 2", + "[80] Teng Wang, Ruimao Zhang, Zhichao Lu, Feng Zheng, Ran Cheng, and Ping Luo. End-to-end dense video captioning with parallel decoding. In ICCV, pages 6847-6857, 2021. 2", + "[81] Xiaolong Wang, Ross Girshick, Abhinav Gupta, and Kaiming He. Non-local neural networks. In CVPR, pages 7794-7803, 2018. 4" + ], + "bbox": [ + 91, + 90, + 482, + 898 + ], + "page_idx": 10 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[82] Junyu Xie, Tengda Han, Max Bain, Arsha Nagrani, Güll Varol, Weidi Xie, and Andrew Zisserman. Autoad-zero: A training-free framework for zero-shot audio description. arXiv preprint arXiv:2407.15850, 2024. 1, 2, 4, 5, 6, 7, 8", + "[83] Antoine Yang, Arsha Nagrani, Paul Hongsuck Seo, Antoine Miech, Jordi Pont-Tuset, Ivan Laptev, Josef Sivic, and Cordelia Schmid. Vid2seq: Large-scale pretraining of a visual language model for dense video captioning. In CVPR, pages 10714-10726, 2023. 2", + "[84] Linli Yao, Weiying Wang, and Qin Jin. Image difference captioning with pre-training and contrastive learning. In AAAI, pages 3108-3116, 2022. 3", + "[85] Keunwoo Peter Yu. Videoblip, 2023. 6", + "[86] Chaoyi Zhang, Kevin Lin, Zhengyuan Yang, Jianfeng Wang, Linjie Li, Chung-Ching Lin, Zicheng Liu, and Lijuan Wang. Mm-narrator: Narrating long-form videos with multimodal in-context learning. In CVPR, pages 13647-13657, 2024. 1, 2, 4, 6", + "[87] Hang Zhang, Xin Li, and Lidong Bing. Video-llama: An instruction-tuned audio-visual language model for video understanding. arXiv preprint arXiv:2306.02858, 2023. 6", + "[88] Tianyi Zhang, Varsha Kishore, Felix Wu, Kilian Q Weinberger, and Yoav Artzi. Bertscore: Evaluating text generation with bert. arXiv preprint arXiv:1904.09675, 2019. 5", + "[89] Luowei Zhou, Yingbo Zhou, Jason J Corso, Richard Socher, and Caiming Xiong. End-to-end dense video captioning with masked transformer. In CVPR, pages 8739-8748, 2018. 2" + ], + "bbox": [ + 516, + 90, + 903, + 474 + ], + "page_idx": 10 + }, + { + "type": "page_number", + "text": "13581", + "bbox": [ + 480, + 944, + 517, + 955 + ], + "page_idx": 10 + } +] \ No newline at end of file diff --git a/2025/DistinctAD_ Distinctive Audio Description Generation in Contexts/d4975861-cea8-4f94-9bb1-e02706824334_model.json b/2025/DistinctAD_ Distinctive Audio Description Generation in Contexts/d4975861-cea8-4f94-9bb1-e02706824334_model.json new file mode 100644 index 0000000000000000000000000000000000000000..4da46a7b9a43e27a62c9242b51582c21225c490a --- /dev/null +++ b/2025/DistinctAD_ Distinctive Audio Description Generation in Contexts/d4975861-cea8-4f94-9bb1-e02706824334_model.json @@ -0,0 +1,3027 @@ +[ + [ + { + "type": "header", + "bbox": [ + 0.107, + 0.003, + 0.182, + 0.044 + ], + "angle": 0, + "content": "CVF" + }, + { + "type": "header", + "bbox": [ + 0.237, + 0.001, + 0.812, + 0.047 + ], + "angle": 0, + "content": "This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore." + }, + { + "type": "title", + "bbox": [ + 0.168, + 0.131, + 0.83, + 0.152 + ], + "angle": 0, + "content": "DistinctAD: Distinctive Audio Description Generation in Contexts" + }, + { + "type": "text", + "bbox": [ + 0.195, + 0.18, + 0.803, + 0.234 + ], + "angle": 0, + "content": "Bo Fang\\(^{1}\\), Wenhao Wu\\(^{2,3}\\), Qiangqiang Wu\\(^{1}\\), Yuxin Song\\(^{2}\\), Antoni B. Chan\\(^{1}\\) \n\\(^{1}\\) Department of Computer Science, City University of Hong Kong \n\\(^{2}\\) Baidu Inc. \n\\(^{3}\\) The University of Sydney" + }, + { + "type": "text", + "bbox": [ + 0.216, + 0.237, + 0.782, + 0.27 + ], + "angle": 0, + "content": "{bofang6-c, qiangqwu2-c}@my.cityu.edu.hk, wenhao.wu@sydney.edu.au songyuxin02@baidu.com, abchan@cityu.edu.hk" + }, + { + "type": "title", + "bbox": [ + 0.248, + 0.305, + 0.327, + 0.32 + ], + "angle": 0, + "content": "Abstract" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.336, + 0.485, + 0.73 + ], + "angle": 0, + "content": "Audio Descriptions (ADs) aim to provide a narration of a movie in text form, describing non-dialogue-related narratives, such as characters, actions, or scene establishment. Automatic generation of ADs remains challenging due to: i) the domain gap between movie-AD data and existing data used to train vision-language models, and ii) the issue of contextual redundancy arising from highly similar neighboring visual clips in a long movie. In this work, we propose DistinctAD, a novel two-stage framework for generating ADs that emphasize distinctiveness to produce better narratives. To address the domain gap, we introduce a CLIP-AD adaptation strategy that does not require additional AD corpora, enabling more effective alignment between movie and AD modalities at both global and fine-grained levels. In Stage-II, DistinctAD incorporates two key innovations: (i) a Contextual Expectation-Maximization Attention (EMA) module that reduces redundancy by extracting common bases from consecutive video clips, and (ii) an explicit distinctive word prediction loss that filters out repeated words in the context, ensuring the prediction of unique terms specific to the current AD. Comprehensive evaluations on MAD-Eval, CMD-AD, and TV-AD benchmarks demonstrate the superiority of DistinctAD, with the model consistently outperforming baselines, particularly in Recall@k/N, highlighting its effectiveness in producing high-quality, distinctive ADs." + }, + { + "type": "title", + "bbox": [ + 0.092, + 0.755, + 0.222, + 0.77 + ], + "angle": 0, + "content": "1. Introduction" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.78, + 0.484, + 0.901 + ], + "angle": 0, + "content": "Audio description (AD) [20, 67] is a crucial accessibility service that provides verbal narration of visual elements in media content for individuals who are blind or have low vision. By offering succinct and vivid descriptions, ADs enable visually impaired audiences to fully comprehend and engage with non-dialogue-related narratives, e.g., characters, facial expressions, non-verbal actions, or scene establishment. Recent studies also show ADs' value for" + }, + { + "type": "image", + "bbox": [ + 0.526, + 0.306, + 0.893, + 0.399 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.526, + 0.402, + 0.9, + 0.496 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.512, + 0.503, + 0.907, + 0.587 + ], + "angle": 0, + "content": "Figure 1. (a) Previous methods approach the AD task similar to video captioning, using only a single video clip as input, which leads to repetitive ADs due to highly similar neighboring clips. (b) Our DistinctAD method generates distinctive ADs across \\( N \\) consecutive clips, with three key innovations: VLM-AD adaptation, the Distinct Module, and explicit distinctive words prediction." + }, + { + "type": "text", + "bbox": [ + 0.511, + 0.596, + 0.907, + 0.748 + ], + "angle": 0, + "content": "sighted viewers in supporting eye-free activities and facilitating child language development [31, 55], reinforcing its pivotal role in fostering inclusivity by bridging the perceptual gap between visual and non-visual elements. Crafting ADs requires careful attention to timing, language, and context to integrate smoothly with dialogue [22]. However, despite the availability of advanced AD platforms [9, 54], human-annotated methods are costly and difficult to scale, highlighting the need for automated generation systems, especially with the rise of user-generated content." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.75, + 0.909, + 0.903 + ], + "angle": 0, + "content": "Advancements in Vision-Language Models (VLMs) and Large-Language Models (LLMs) have led to growing interest in automatic AD generation for media. Current approaches fall into two categories: (i) using powerful proprietary models like GPT-4 [3] or GPT-4V [53] in a training-free manner [14, 39, 82, 86], and (ii) fine-tuning open-source VLM components, such as visual-text adapters [5, 34, 42], for AD tasks [21-23, 40, 56, 74]. Both approaches have limitations: (i) Training-free methods often perform poorly and suffer from hallucinations due to the unique na" + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.518, + 0.957 + ], + "angle": 0, + "content": "13571" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.09, + 0.092, + 0.482, + 0.168 + ], + "angle": 0, + "content": "ture of AD (e.g., character names and narrative coherence), which differs from the common text data LLMs are trained on. (ii) Fine-tuning methods generally perform better but are still limited by insufficient data to fully adapt to the movie-AD domain and face the context-repetition issue." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.17, + 0.482, + 0.351 + ], + "angle": 0, + "content": "Unlike video captioning [38, 62], ADs are generated on consecutive intervals (visual clips) throughout long videos [68], e.g., movies. The context-repetition issue arises when models produce repetitive or similar descriptions for consecutive visual clips, especially when using prior ADs as prompts [21, 74]. This occurs because sequential clips often comprise redundant scenes or characters (and therein redundant visual features), leading models that only use the current visual clip to repeat the same information from the past, as shown in Fig. 1. However, audiences are more interested in the unique and distinct events of the current clip, rather than the common elements from the previous one." + }, + { + "type": "text", + "bbox": [ + 0.093, + 0.355, + 0.485, + 0.687 + ], + "angle": 0, + "content": "In this paper, we propose DistinctAD, a two-stage framework for generating distinctive ADs within contexts. Given the domain gap between the movie-AD and VLM training data, we first bridge this gap in Stage-I by adapting VLMs, such as CLIP [58], to the movie-AD domain. Our adaptation strategy is inspired by a key observation (see Appendix §A): AD sentences encoded by the CLIP text encoder can be effectively reconstructed using simple LLMs like GPT-2 [57] with minimal fine-tuning, whereas AD reconstructions using CLIP visual features from the corresponding clips are often of poor quality. This suggests that while CLIP's multi-modal embedding space is rich enough to represent AD information, its visual encoder is insufficient for extracting it. To mitigate this domain gap, we adapt the CLIP vision encoder to better align with the frozen CLIP text encoder using existing paired video-AD data. The alignment involves global matching at video-sentence level, similar to CLIP pre-training. A challenge arises because video clips are labeled with whole ADs, and words may not appear in every frame but must be aggregated over frames. Therefore, we propose fine-grained matching at frame-word level for this multiple-instance setting." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.69, + 0.483, + 0.902 + ], + "angle": 0, + "content": "For Stage-II, we propose a novel distinctive AD narrating pipeline based on the Expectation-Maximization Attention (EMA) [17] algorithm, which has demonstrated its efficacy in tasks such as semantic segmentation [35], video object segmentation [41], and text-video retrieval [27]. Differently, we apply EMA to contextual clips from long videos, which often exhibit high redundancy due to recurring scenes or characters. By extracting common bases from contextual information, DistinctAD reduces redundancy and generates compact, discriminative representations that enable the LLM decoder to produce more distinctive ADs. To further emphasize distinctiveness explicitly, we introduce a distinctive word prediction loss that filters out words that repeatedly appear in contexts, ensuring that the LLM decoder" + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.092, + 0.905, + 0.151 + ], + "angle": 0, + "content": "focuses on predicting unique words specific to the current AD. With these two designs, DistinctAD produces contextually distinctive and engaging ADs that can provide better narratives for the audience." + }, + { + "type": "text", + "bbox": [ + 0.534, + 0.153, + 0.833, + 0.166 + ], + "angle": 0, + "content": "In summary, our contributions are three-fold:" + }, + { + "type": "text", + "bbox": [ + 0.514, + 0.167, + 0.905, + 0.257 + ], + "angle": 0, + "content": "- We propose a CLIP-AD adaptation strategy tailored to movie-AD data, addressing the misalignment issue caused by the domain gap. Our adapted vision encoder can be seamlessly integrated into existing CLIP-based AD methods and stands to benefit from future improvements as more AD data becomes available." + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.258, + 0.905, + 0.318 + ], + "angle": 0, + "content": "- We introduce DistinctAD, which incorporates a Contextual EMA module and a distinctive word prediction loss, significantly enhancing the generation of distinctive ADs from consecutive visual clips with similar contexts." + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.318, + 0.905, + 0.393 + ], + "angle": 0, + "content": "- Comprehensive evaluations on MAD-Eval [21], CMD-AD [23], and TV-AD [82] highlight DistinctAD's superiority. Our outstanding performance in Recall@k/N demonstrates its effectiveness in generating high-quality ADs with both distinctiveness and technical excellence." + }, + { + "type": "list", + "bbox": [ + 0.513, + 0.167, + 0.905, + 0.393 + ], + "angle": 0, + "content": null + }, + { + "type": "title", + "bbox": [ + 0.514, + 0.408, + 0.655, + 0.423 + ], + "angle": 0, + "content": "2. Related Work" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.433, + 0.907, + 0.644 + ], + "angle": 0, + "content": "Dense video captioning. A task closely related to AD is dense video captioning [29], which extends traditional video captioning [38, 45, 62, 63] by both generating a single caption for trimmed video segments as well as detecting and describing multiple events with grounded timestamps. Initial dense video captioning utilize a 2-stage pipeline [25, 26, 75, 79] by firstly performing localization and then describing events. Recent works [13, 18, 36, 51, 59, 64, 65, 75, 80, 83, 89] focus on training localization and captioning modules in an end-to-end manner to enhance inter-event associations. In contrast to these works, AD generation specifically aims to narrate a coherent story, maintain character-awareness, and complement the audio track without interfering with existing dialogue." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.645, + 0.909, + 0.794 + ], + "angle": 0, + "content": "AD generation. Early AD systems relied heavily on specialized authoring tools [9] and skilled human contributors. Platforms like Rescribe [54] and LiveDescribe [9] have facilitated faster and more accurate AD creation; however, these methods are costly and do not scale efficiently for large volumes of visual content. Recent efforts have developed audio segmentation and transcription systems [8, 10, 11] to create high-quality video datasets with temporally aligned ADs [60, 61, 68, 69], advancing automatic AD research." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.796, + 0.909, + 0.903 + ], + "angle": 0, + "content": "In general, current automatic AD generation systems can be categorized into two approaches: training-free and partial-fine-tuning. Training-free methods [39] generate ADs by leveraging proprietary models like GPT-4 [3] and GPT-4V [53]. MM-Narrator [86] enhances AD performance by multi-model in-context learning with memories. LLM-AD [14] and AutoAD-Zero [82] use prompts com" + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.52, + 0.957 + ], + "angle": 0, + "content": "13572" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.09, + 0.092, + 0.485, + 0.363 + ], + "angle": 0, + "content": "prising visual frames with textual character names and colorful circles [66], enabling character-centric AD generation. However, training-free AD methods often suffer from high evaluation costs at scale and relatively poor performance due to domain-specific challenges and LLM hallucinations. Partial-fine-tuning methods [21-23, 40, 74], as well as our DistinctAD, only fine-tune a lightweight adapter [5, 34] between the pre-trained vision and text encoders. A representative example is the AutoAD series [21-23], which builds automatic AD systems and enriches them with character-aware prompts within different vision-language frameworks. However, previous studies tend to focus on constructing more accurate external character banks, whereas treating AD generation similarly to video captioning, overlooks AD's unique sequential structure of video clips. In contrast, our method emphasizes understanding the visual content within its temporal context, leading to more distinctive AD generation." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.365, + 0.484, + 0.606 + ], + "angle": 0, + "content": "Distinctive captioning in images aims to articulate unique details that can help distinguishing targets from others. An intuitive way to promoting distinctiveness is through contrastive learning [16, 43, 47, 73], where generated captions are encouraged to align more closely with target images rather than distractors. In [12, 76-78], group-based distinctive attention is introduced to capture distinctiveness by comparing sets of similar images and re-weighting specific caption words. A recent closely related field is difference captioning [32, 33, 84], which aims to describe differences between a single pair of images. VisDiff [19] scales difference captioning to sets containing thousands of images with natural language. Our work differs from these distinctive captioning works in that we are the first to explore distinctiveness across dense, consecutive clips within hours-long movies, thereby generating ADs with better narrative." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.623, + 0.182, + 0.639 + ], + "angle": 0, + "content": "3. Method" + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.649, + 0.358, + 0.666 + ], + "angle": 0, + "content": "3.1. Stage-I: CLIP-AD Adaptation" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.673, + 0.483, + 0.779 + ], + "angle": 0, + "content": "AD and the visual content it describes exhibit a significant domain gap compared to typical large-scale web data. This gap often causes misalignment in current partial-fine-tuning techniques. Previous studies [21, 22, 74] try alleviating this problem by pre-training LLMs on text-only AD corpus, e.g. AudioVault [1]. However, misalignment persists at the initial stage of vision encoding, which is often neglected." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.78, + 0.484, + 0.901 + ], + "angle": 0, + "content": "Inspired by our findings that AD sentences encoded by the CLIP text encoder can be effectively recovered using the GPT-2 language model (see §1 and Appendix §A), we identify that the primary issue of misalignment is caused by the CLIP vision encoder, i.e., the discrepancy between visual embeddings and AD embeddings within the joint CLIP feature space. To address this, we propose adapting the CLIP vision encoder to the specific AD domain. However, due to" + }, + { + "type": "image", + "bbox": [ + 0.52, + 0.085, + 0.9, + 0.337 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.513, + 0.341, + 0.907, + 0.398 + ], + "angle": 0, + "content": "Figure 2. Illustration of Stage-I: CLIP-AD Adaptation. This process involves adapting the CLIP vision encoder to specific movie-AD data through global-level video-AD matching (bottom right) and fine-grained frame-AD matching (top right)." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.408, + 0.907, + 0.453 + ], + "angle": 0, + "content": "the unique multiple-instance learning setting for video-AD pairs (see §1), we consider both global matching and fine-grained frame-word matching in our adaptation method." + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.453, + 0.909, + 0.681 + ], + "angle": 0, + "content": "Global video-AD matching. A straightforward strategy involves adopting classical CLIP-style fine-tuning with video-AD pairs in large batches [46]. Formally, let video clip \\(\\mathbf{V}_i = [\\mathbf{f}_i^1;\\dots ;\\mathbf{f}_i^n ]\\in \\mathbb{R}^{n\\times C}\\) be a collection of \\(n\\) frame embeddings, and corresponding AD \\(\\mathbf{T}_i = [\\mathbf{w}_i^0;\\mathbf{w}_i^1;\\dots ;\\mathbf{w}_i^m ]\\in \\mathbb{R}^{(m + 1)\\times C}\\) be a collection of \\(m\\) word embeddings \\((\\mathbf{w}_i^j)\\) and the [CLS] token (denoted as \\(\\mathbf{w}_i^0\\)), where \\(C\\) is the number of channels in the embedding space. We obtain the global video-level representation by averaging all frame embeddings in \\(\\mathbf{V}_i\\) using mean pooling: \\(\\mathbf{v}_i = \\frac{1}{n}\\sum_{j = 1}^n\\mathbf{f}_i^j\\). Following the standard CLIP, we use the [CLS] token as the global textual AD representation \\(\\mathbf{t}_i = \\mathbf{w}_i^0\\). The global video \\(\\rightarrow\\) AD matching is performed by maximizing the sum of the main diagonal of a \\(B\\times B\\) similarity matrix, using the contrastive loss:" + }, + { + "type": "equation", + "bbox": [ + 0.537, + 0.684, + 0.907, + 0.726 + ], + "angle": 0, + "content": "\\[\n\\mathcal {L} _ {v \\rightarrow A D} = - \\frac {1}{B} \\sum_ {i = 1} ^ {B} \\log \\frac {\\exp (s i m (\\mathbf {v} _ {i} , \\mathbf {t} _ {i}))}{\\sum_ {j = 1} ^ {B} \\exp (s i m (\\mathbf {v} _ {i} , \\mathbf {t} _ {j}))}, \\tag {1}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.735, + 0.907, + 0.84 + ], + "angle": 0, + "content": "where \\(B\\) is the batch size, and the similarity function \\(sim(\\cdot)\\) is the vector inner product. This process is illustrated in the bottom right of Fig. 2. Similarly, we drive the AD \\(\\rightarrow\\) video loss \\(\\mathcal{L}_{AD\\rightarrow v}\\) by maximizing the sum of the secondary diagonal (i.e., swapping the \\(i\\) and \\(j\\) indices in (1)). The final global-level contrastive loss is then the sum of the losses in both directions \\(\\mathcal{L}_q = \\mathcal{L}_{v\\to AD} + \\mathcal{L}_{AD\\to v}\\)." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.84, + 0.909, + 0.901 + ], + "angle": 0, + "content": "Fine-grained frame-AD matching. Matching global video to AD sentence [CLS] (and vice versa) aids in joint feature space alignment. However, this alignment is insufficient for effective adaptation due to the specific multiple-instance" + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "13573" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.182, + 0.089, + 0.817, + 0.341 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.089, + 0.344, + 0.908, + 0.388 + ], + "angle": 0, + "content": "Figure 3. Pipeline of Stage-II: Distinctive AD Narration. Stage-II processes \\(N\\) consecutive video clips using the \\(\\mathrm{CLIP}_{\\mathrm{AD}}\\) vision encoder from Stage-I. We generate contextual-distinctive ADs by two key innovations: i) a Contextual EMA module to learn compact and discriminative visual representations for improved prompting of LLMs; ii) an extra distinctive word loss for predicting AD-specific terms." + }, + { + "type": "text", + "bbox": [ + 0.089, + 0.399, + 0.482, + 0.458 + ], + "angle": 0, + "content": "setting of ADs, where only some words may correspond to a particular frame, but all words will have correspondence in aggregate. Thus, we propose a fine-grained matching loss at the frame-level to address this issue." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.46, + 0.484, + 0.551 + ], + "angle": 0, + "content": "Formally, given the frame embeddings \\(\\mathbf{V}_i\\) and the word embeddings \\(\\mathbf{T}_i' = [\\mathbf{w}_i^1; \\dots; \\mathbf{w}_i^m] \\in \\mathbb{R}^{m \\times C}\\), we calculate the weights of all words attending to each frame via softmax attention, taking \\(\\mathbf{V}_i\\) as the query and \\(\\mathbf{T}_i'\\) as the key. By then multiplying these attention weights by the value \\(\\mathbf{T}_i'\\), we obtain a frame-aware AD representation \\(\\tilde{\\mathbf{T}}_i \\in \\mathbb{R}^{n \\times C}\\):" + }, + { + "type": "equation", + "bbox": [ + 0.134, + 0.556, + 0.483, + 0.577 + ], + "angle": 0, + "content": "\\[\n\\tilde {\\mathbf {T}} _ {i} = \\mathrm {S o f t m a x} (\\mathbf {V} _ {i} \\mathbf {T} _ {i} ^ {\\prime T} / \\tau) \\mathbf {T} _ {i} ^ {\\prime} = [ \\tilde {\\mathbf {t}} _ {i} ^ {1}; \\dots ; \\tilde {\\mathbf {t}} _ {i} ^ {n} ], \\qquad (2)\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.588, + 0.484, + 0.663 + ], + "angle": 0, + "content": "where for each frame, the words embeddings that are most similar to the frame-level visual feature have been aggregated (via softmax attention). The temperature parameter \\(\\tau\\) controls the aggregation process, where smaller \\(\\tau\\) incorporates more textual information." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.664, + 0.484, + 0.755 + ], + "angle": 0, + "content": "The goal of the fine-grained matching is to pull a frame visual feature \\(\\mathbf{f} \\in \\mathbf{V}_i\\) closer to the frame-aware AD representations \\(\\tilde{\\mathbf{t}} \\in \\tilde{\\mathbf{T}}_i\\) in (2), corresponding to the positive set. To achieve this, we define the negative set \\(\\tilde{\\mathbf{T}}_{neg}\\) as the frame-aware AD embeddings generated from other video clips (in the batch), and then use a Multi-Instance Loss [49]," + }, + { + "type": "equation", + "bbox": [ + 0.102, + 0.773, + 0.483, + 0.816 + ], + "angle": 0, + "content": "\\[\n\\mathcal {L} _ {f} = - \\frac {1}{B} \\sum_ {i = 1} ^ {B} \\log \\frac {\\sum_ {\\tilde {\\mathbf {t}} \\in \\tilde {\\mathbf {T}} _ {i}} \\exp (s i m (\\mathbf {f} , \\tilde {\\mathbf {t}}))}{\\sum_ {\\tilde {\\mathbf {t}} _ {*} \\in \\{\\tilde {\\mathbf {T}} _ {i} \\cup \\tilde {\\mathbf {T}} _ {n e g} \\}} \\exp (s i m (\\mathbf {f} , \\tilde {\\mathbf {t}} _ {*}))}, \\tag {3}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.825, + 0.483, + 0.856 + ], + "angle": 0, + "content": "where \\(\\mathbf{f} \\in \\mathbf{V}_i\\) is a sampled frame from \\(\\mathbf{V}_i\\). This process is illustrated in the top right of Fig. 2." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.856, + 0.484, + 0.903 + ], + "angle": 0, + "content": "Summary for Stage-I. The final objective for Stage-I is to minimize the sum of global and fine-grained aligning losses, balanced by a trade-off coefficient, \\(\\mathcal{L}_{\\mathrm{I}} = \\gamma \\mathcal{L}_{g} + (1 -\\)" + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.399, + 0.907, + 0.476 + ], + "angle": 0, + "content": "\\(\\gamma)\\mathcal{L}_f\\). Note that during this adaptation process, the CLIP-Text encoder model remains frozen, and only the CLIP-Vision encoder is fine-tuned. Our fine-grained frame-AD matching is entirely parameter-free, as only the vision encoder will be utilized in the subsequent stage." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.482, + 0.816, + 0.499 + ], + "angle": 0, + "content": "3.2. Stage-II: Distinctive AD Narration" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.506, + 0.907, + 0.672 + ], + "angle": 0, + "content": "The motivation for generating distinctive ADs stems from the observation that LLM often produce repetitive descriptions for adjacent clips [56, 82, 86]. Despite improved character recognition, the visual representation itself is not discriminative among neighboring (contextual) clips, leading to uninteresting ADs. Our goal is to create contextual-distinctive ADs that highlight current differences. We hypothesize, as verified in Appendix §B, that sequential clips from a long video often share redundant scenes or characters, leading similar visual features in contexts. Thus, we propose Stage-II: distinctive AD narration." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.672, + 0.909, + 0.778 + ], + "angle": 0, + "content": "As shown in Fig. 3, we prepare \\(N\\) consecutive video clips (to be AD-described) \\(\\{\\mathbf{x}_1,\\mathbf{x}_2,\\dots ,\\mathbf{x}_N\\}\\), each containing \\(T\\) uniformly sampled frames \\(\\{\\mathcal{F}_1,\\mathcal{F}_2,\\dots ,\\mathcal{F}_T\\}\\). Following [21, 22, 74], we employ a learnable Perceiver adapter [5] to resample \\(T^{\\prime}\\) prompt vectors for the \\(T\\) frame embeddings encoded by our Stage-I vision encoder, \\(\\mathrm{CLIP}_{\\mathrm{AD}}\\). This process is formulated as:" + }, + { + "type": "equation", + "bbox": [ + 0.557, + 0.78, + 0.905, + 0.798 + ], + "angle": 0, + "content": "\\[\n\\mathbf {h} _ {\\mathbf {x} _ {i}} = \\operatorname {P e r c e i v e r} \\left(\\left\\{\\mathbf {f} _ {1}, \\mathbf {f} _ {2}, \\dots , \\mathbf {f} _ {T} \\right\\}\\right) \\in \\mathbb {R} ^ {T ^ {\\prime} \\times C}, \\tag {4}\n\\]" + }, + { + "type": "equation", + "bbox": [ + 0.572, + 0.801, + 0.905, + 0.817 + ], + "angle": 0, + "content": "\\[\n\\mathbf {f} _ {i} = \\operatorname {C L I P} _ {\\mathrm {A D}} \\left(\\mathcal {F} _ {i}\\right). \\tag {5}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.826, + 0.907, + 0.902 + ], + "angle": 0, + "content": "We then introduce the Contextual EMA to capture compact, discriminative visual features for distinctive AD generation. Contextual EMA. Expectation-Maximization Attention (EMA) [35] integrates the attention mechanism [81] into the classical EM [17] algorithm, which comprises three steps to" + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.52, + 0.957 + ], + "angle": 0, + "content": "13574" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.09, + 0.092, + 0.482, + 0.181 + ], + "angle": 0, + "content": "estimate a more compact set of bases: Responsibility Estimation (RE), Likelihood Maximization (LM), and Data Re-estimation (DR). Inspired by this, we propose Contextual EMA to perform EMA on frames from \\(N\\) contextual clips, aiming to eliminate redundancy, learn compact representations, and explore distinctiveness." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.182, + 0.482, + 0.29 + ], + "angle": 0, + "content": "Let \\(\\mathcal{H} = \\{\\mathbf{h}_{\\mathbf{x}_i}\\}_{i = 1}^N\\in \\mathbb{R}^{N\\times T'\\times C}\\) represent \\(N\\) clip vectors from the Perceiver, and \\(\\mathcal{M} = \\{\\mu_k\\}_{k = 1}^K\\in \\mathbb{R}^{K\\times C}\\) denote the randomly initialized base features, where \\(C,K\\) indicate the number of channel and bases. The RE step estimates the hidden variable \\(\\mathcal{Z} = \\{z_{nk}\\}_{n = 1,k = 1}^{N\\times T',K}\\), where the responsibility \\(z_{nk}\\) represents the probability of the \\(n\\)-th frame belonging to the \\(k\\)-th base:" + }, + { + "type": "equation", + "bbox": [ + 0.188, + 0.295, + 0.482, + 0.334 + ], + "angle": 0, + "content": "\\[\nz _ {n k} = \\frac {\\exp \\left(\\mathbf {h} _ {n} \\mu_ {k} ^ {T} / \\tau\\right)}{\\sum_ {j = 1} ^ {K} \\exp \\left(\\mathbf {h} _ {n} \\mu_ {j} ^ {T} / \\tau\\right)}, \\tag {6}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.343, + 0.483, + 0.389 + ], + "angle": 0, + "content": "where \\(\\tau\\) determines the shape (peakiness) of distribution \\(\\mathcal{Z}\\). Then, the LM step updates the bases \\(\\mathcal{M}\\) by applying the weighted average on input \\(\\mathcal{H}\\), formulating the \\(k\\)-th base as:" + }, + { + "type": "equation", + "bbox": [ + 0.21, + 0.393, + 0.482, + 0.434 + ], + "angle": 0, + "content": "\\[\n\\mu_ {k} = \\frac {\\sum_ {n = 1} ^ {N \\times T ^ {\\prime}} z _ {n k} \\mathbf {h} _ {n}}{\\sum_ {n = 1} ^ {N \\times T ^ {\\prime}} z _ {n k}}. \\tag {7}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.438, + 0.483, + 0.498 + ], + "angle": 0, + "content": "The RE (E-step) and LM (M-step) are iteratively performed \\(R\\) times until convergence. Notably, since bases number \\(K\\) is much smaller than the embedding number \\(N \\times T'\\), we employ DR to reconstruct a compact version of \\(\\mathcal{H}\\) through:" + }, + { + "type": "equation", + "bbox": [ + 0.248, + 0.504, + 0.482, + 0.523 + ], + "angle": 0, + "content": "\\[\n\\widehat {\\mathcal {H}} \\approx \\mathcal {Z M}. \\tag {8}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.527, + 0.483, + 0.56 + ], + "angle": 0, + "content": "Here, \\(\\widehat{\\mathcal{H}}\\in \\mathbb{R}^{N\\times T^{\\prime}\\times C}\\) retains the same shape as \\(\\mathcal{H}\\). We combine \\(\\mathcal{H}\\) and \\(\\widehat{\\mathcal{H}}\\) element-wise with a hyperparameter \\(\\alpha\\)." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.561, + 0.483, + 0.605 + ], + "angle": 0, + "content": "To enhance representation distinctiveness, we introduce an additional branch using cross-attention between raw \\(\\mathcal{H}\\) (query) and bases \\(\\mathcal{M}\\) (key and value), formulated as:" + }, + { + "type": "equation", + "bbox": [ + 0.183, + 0.609, + 0.482, + 0.628 + ], + "angle": 0, + "content": "\\[\n\\widetilde {\\mathcal {H}} = \\operatorname {C r o s s A t t e n t i o n} (\\mathcal {H}, \\mathcal {M}), \\tag {9}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.639, + 0.483, + 0.761 + ], + "angle": 0, + "content": "where \\(\\widetilde{\\mathcal{H}}\\in \\mathbb{R}^{N\\times T^{\\prime}\\times C}\\). Linear layers projecting queries, keys, and values are omitted in Fig. 3 (see Appendix §C for details). Through (9), we process the distributions of \\(\\mathcal{H}\\) to attend on specific and informative bases, with improved linear separability (see Fig. 6). We combine \\(\\mathcal{H},\\widehat{\\mathcal{H}},\\widetilde{\\mathcal{H}}\\) elementwise around Contextual EMA to construct the final refined visual features. These features are then projected into the LLM embedding space using a single-layer projector:" + }, + { + "type": "equation", + "bbox": [ + 0.178, + 0.766, + 0.482, + 0.784 + ], + "angle": 0, + "content": "\\[\n\\mathcal {H} _ {\\text {s u m}} = \\operatorname {P r o j} (\\mathcal {H} + \\alpha \\widehat {\\mathcal {H}} + \\beta \\widetilde {\\mathcal {H}}). \\tag {10}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.795, + 0.483, + 0.901 + ], + "angle": 0, + "content": "Interleaved prompt as LLM's input. Following previous studies [22, 74, 82], we build our interleaved prompt enriched with character information, (see Fig. 3). To answer the \"who is who\" question when more than two characters are present, the corresponding actors' portrait images are projected as face tokens for reasoning. The \\( \\langle \\text{BOS} \\rangle \\) tag appended at the end indicates the beginning of AD generation." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.092, + 0.905, + 0.152 + ], + "angle": 0, + "content": "Distinctive words highlighting. Our goal is to query a frozen LLM for AD generation using a vision-conditioned prompt. The typical supervision employs the commonly used auto-regressive loss function:" + }, + { + "type": "equation", + "bbox": [ + 0.569, + 0.16, + 0.905, + 0.19 + ], + "angle": 0, + "content": "\\[\n\\mathcal {L} _ {\\text {a u t o}} = - \\sum_ {n} \\log P _ {\\theta} \\left(w _ {n} \\mid \\text {p r o m p t}; w _ {< n}\\right), \\tag {11}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.199, + 0.905, + 0.32 + ], + "angle": 0, + "content": "where \\( w_{n} \\) is the \\( n \\)-th token from the target AD. However, \\( \\mathcal{L}_{auto} \\) does not emphasize the distinctiveness specific to the current AD, which is our focus. To address this, we propose a distinctive word set \\( w_{d} \\), created by filtering out duplicates, such as character names, prepositions, and pronouns, from the \\( N \\) context ADs of the target AD. During training, we explicitly encourage the LLM to predict the distinctive words in \\( w_{d} \\) by optimizing the distinctive loss \\( \\mathcal{L}_{dist} \\):" + }, + { + "type": "equation", + "bbox": [ + 0.522, + 0.323, + 0.905, + 0.363 + ], + "angle": 0, + "content": "\\[\n\\mathcal {L} _ {\\text {d i s t}} = - \\sum_ {n = 1} ^ {N} \\sum_ {i = 1} ^ {u} \\log P _ {\\theta} \\left(w _ {n} = w _ {d} ^ {i} \\mid \\text {p r o m p t}, w _ {< n}\\right), \\tag {12}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.373, + 0.905, + 0.418 + ], + "angle": 0, + "content": "where \\( w_{d}^{i} \\) denotes the \\( i \\)-th distinctive word in \\( w_{d} \\) and \\( u \\) is the size of the set. The final complete loss function for Stage-II is: \\( \\mathcal{L}_{\\mathrm{II}} = \\mathcal{L}_{\\mathrm{auto}} + \\mathcal{L}_{\\mathrm{dist}} \\)." + }, + { + "type": "title", + "bbox": [ + 0.514, + 0.431, + 0.645, + 0.447 + ], + "angle": 0, + "content": "4. Experiments" + }, + { + "type": "title", + "bbox": [ + 0.514, + 0.455, + 0.691, + 0.471 + ], + "angle": 0, + "content": "4.1. Experiment Setup" + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.478, + 0.905, + 0.674 + ], + "angle": 0, + "content": "Datasets. We follow the AD generation benchmark established in AutoAD [21], conducting experiments on the denoised MAD-v2-Named [68] and evaluating on MAD-Eval-Named split. Specifically, MAD-v2-Named includes \\(\\sim 330\\mathrm{k}\\) ADs from 488 movies for training and MAD-Eval has 6,520 ADs crawled from 10 movies for evaluation. We also evaluate on two recently introduced datasets. CMD-AD [23] (where \"CMD\" stands for Condensed Movie Dataset [7]) is a movie AD dataset that contains 101k ADs for more than 1432 movies, with 100 movies split for CMD-AD evaluation. TV-AD [82] is a recently proposed AD dataset based on TVQA [30], which contains \\(\\sim 31\\mathrm{k}\\) ADs for training and \\(\\sim 3\\mathrm{k}\\) ADs for evaluation." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.675, + 0.905, + 0.901 + ], + "angle": 0, + "content": "Evaluation Metrics. Classic captioning metrics including ROUGE-L [37], CIDEr [72] and SPICE [6] are reported to evaluate the quality of generated ADs versus the ground-truth. Besides, we also report Recall@k within \\(N\\) Neighbours [22] (\\(\\mathbf{R}@\\mathbf{k} / \\mathbf{N}\\)), which calculates the average value of Recall@k for each AD with its \\(N\\) temporally adjacent GT texts, where BertScore [88] is used for text similarity matching. The R@k/N metric is based on retrieving the most similar ground-truth AD among N neighbors, and thus highlights the distinctiveness of generated ADs directly. LLM-AD-eval [23] employs LLMs to assess the quality of generated ADs, assigning scores from 1 (lowest) to 5 (highest). We utilize the LLM prompt from the original study [82] and apply open-source models LLaMA2-7B-Chat [70] and LLaMA3-8B-Instruct [4] for this evaluation." + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.518, + 0.957 + ], + "angle": 0, + "content": "13575" + } + ], + [ + { + "type": "table", + "bbox": [ + 0.148, + 0.089, + 0.851, + 0.357 + ], + "angle": 0, + "content": "
MethodPub.VLMLLMROUGE-LCIDErSPICER@5/16
Training-free
VLog [2]--GPT-47.51.32.142.3
MM-Vid [39]ArXiv'23GPT-4V-9.86.13.846.1
MM-Narrator [86]CVPR'24CLIP-L14GPT-413.413.95.249.0
LLM-AD [14]ArXiv'24GPT-4V-13.520.5--
AutoAD-Zero [82]ACCV'24VideoLLaMA2-7BLLaMA3-8B-22.4--
Partial-fine-tuning
ClipCap [50]ArXiv'21CLIP-B32GPT-28.54.41.1-
CapDec [52]ArXiv'22--8.26.71.4-
AutoAD-I [21]CVPR'23CLIP-B32GPT-211.914.34.442.1
AutoAD-II [22]ICCV'23CLIP-B32GPT-213.419.5-50.8
AutoAD-III [23]CVPR'24EVA-CLIPOPT-2.7B-22.8-52.0
AutoAD-III [23]CVPR'24EVA-CLIPLLaMA2-7B-24.0-52.8
MovieSeq [40]ECCV'24CLIP-B16LLaMA2-7B*15.524.47.051.6
DistinctAD (Ours)CLIP-B32GPT-215.424.56.749.8
DistinctAD (Ours)CLIPAD-B32GPT-216.425.57.451.7
DistinctAD (Ours)CLIPAD-B16LLaMA2-7B17.227.08.255.6
DistinctAD (Ours)CLIPAD-B16LLaMA3-8B17.627.38.356.0
" + }, + { + "type": "table_caption", + "bbox": [ + 0.09, + 0.362, + 0.905, + 0.391 + ], + "angle": 0, + "content": "Table 1. Comparisons of AD performance on the MAD-Eval benchmark. * indicates fine-tuning LLaMA2-7B model with LoRA [24]. \\( {\\mathrm{{CLIP}}}_{\\mathrm{{AD}}} \\) is our CLIP vision encoder adapted using our Stage-I strategy." + }, + { + "type": "table", + "bbox": [ + 0.126, + 0.403, + 0.446, + 0.518 + ], + "angle": 0, + "content": "
MethodCIDErR@1/5LLM-AD-eval
Video-BLIP2 [85]4.822.01.89
Video-LLaMA2 [87]5.223.61.91
AutoAD-II [22]13.526.12.08
AutoAD-III [23]21.730.02.85
AutoAD-Zero [82]17.7-2.83
DistinctAD (LLaMA2)22.332.92.89
DistinctAD (LLaMA3)22.733.02.88
AutoAD-III† [23]25.031.22.89
2.01
" + }, + { + "type": "table_caption", + "bbox": [ + 0.09, + 0.522, + 0.483, + 0.565 + ], + "angle": 0, + "content": "Table 2. Comparisons on CMD-AD. The LLM-AD-eval scores are evaluated with LLaMA2-7B (left) and LLaMA3-8B (right). \\(\\dagger\\) indicates pre-training on 3.4M HowTo-AD dataset [23, 48]." + }, + { + "type": "table", + "bbox": [ + 0.126, + 0.572, + 0.449, + 0.638 + ], + "angle": 0, + "content": "
MethodCIDErR@1/5LLM-AD-eval
AutoAD-III [23]26.1-2.78 | 1.99
AutoAD-Zero [82]22.630.62.94 | 2.00
DistinctAD (LLaMA2)27.231.82.86 | 2.01
DistinctAD (LLaMA3)27.432.12.89 | 2.00
" + }, + { + "type": "table_caption", + "bbox": [ + 0.09, + 0.642, + 0.483, + 0.671 + ], + "angle": 0, + "content": "Table 3. Comparisons on TV-AD. The LLM-AD-eval scores are evaluated using LLaMA2-7B (left) and LLaMA3-8B (right)." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.689, + 0.483, + 0.901 + ], + "angle": 0, + "content": "Implementation Details. To facilitate CLIP-AD adaptation in Stage-I, we collect the original raw movies from MAD from platforms like Amazon Prime Video. See Appendix F for details. We fine-tune the CLIP Vision encoder for 5 epochs with a fixed learning rate 5e-5 using the Adam optimizer [28] in Stage-I, with a batch size of 512. In Stage-II, we use a batch of 8 sequences, each containing 16 consecutive video AD-pairs from a movie. For each video clip, 8 frames are uniformly sampled. We use the AdamW [44] optimizer to train our model for 10 epochs, with a cosine-decayed learning rate and linear warm-up. The learning rate is set to \\(10^{-4}\\) for both GPT-2 and LLaMA models. For external character information, we directly use the inference results from AutoAD-Zero [82] as it gives current best face" + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.406, + 0.905, + 0.437 + ], + "angle": 0, + "content": "recognition performance. We use CLIP-B32 and GPT-2 on MAD-Eval-Named for ablations unless specified otherwise." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.45, + 0.83, + 0.466 + ], + "angle": 0, + "content": "4.2. Comparisons with previous methods" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.475, + 0.905, + 0.52 + ], + "angle": 0, + "content": "We conduct comprehensive comparisons using the widely-adopted MAD-Eval benchmark [21] and two recently introduced AD datasets, CMD-AD [23] and TV-AD [82]." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.522, + 0.906, + 0.764 + ], + "angle": 0, + "content": "Comparisons on MAD-Eval are shown in Tab. 1. We primarily categorize previous studies into Training-free and Partial-fine-tuning approaches, as described in §2. Our method is a Partial-fine-tuning method. When using the same CLIP-B32 and GPT-2, our proposed DistinctAD achieves a CIDEr score of 24.5, surpassing previous AutoAD-I [21] (CIDEr 14.3) and AutoAD-II [22] (CIDEr 19.5). With our Stage-I adapted CLIP vision encoders (denoted as \\(\\mathrm{CLIP}_{\\mathrm{AD}}\\)), we observe stable improvements across all metrics, e.g. 25.5 vs. 24.5 on CIDEr and 51.7 vs. 49.8 on recall, validating the effectiveness of our Stage-I strategy. Notably, DistinctAD with \\(\\mathrm{CLIP}_{\\mathrm{AD}}\\)-B16 and LLaMA3-8B [4] achieves state-of-the-arts with CIDEr 27.3 and Recall@5/16 56.0. Our outstanding performance on R@k/N demonstrates DistinctAD's ability to generate distinctive ADs, well matching the uniqueness of the clip's contents." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.765, + 0.906, + 0.901 + ], + "angle": 0, + "content": "Looking at the training-free methods, despite the capabilities of advanced proprietary VLMs, e.g. GPT-4V [53], and LLMs, e.g. GPT-4 [3], the performance of training-free methods remains inferior to those employing partial-finetuning. This discrepancy likely arises from the unique characteristics of AD and movie data, which exhibit a significant domain gap from common vision language training data. As such, these data types were not encountered during the pretraining of proprietary large-scale models." + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "13576" + } + ], + [ + { + "type": "table", + "bbox": [ + 0.107, + 0.089, + 0.338, + 0.17 + ], + "angle": 0, + "content": "
SettingCIDErR@5/16
None6.734.0
Global Lg8.236.6
Fine-grained Lf7.735.2
γLg + (1 - γ)Lf8.636.9
" + }, + { + "type": "table_caption", + "bbox": [ + 0.159, + 0.172, + 0.288, + 0.184 + ], + "angle": 0, + "content": "(a) Stage-I components." + }, + { + "type": "table", + "bbox": [ + 0.375, + 0.089, + 0.523, + 0.17 + ], + "angle": 0, + "content": "
Coefficient γCIDEr
0.18.0
0.38.5
0.58.6
0.77.7
" + }, + { + "type": "table_caption", + "bbox": [ + 0.378, + 0.172, + 0.522, + 0.184 + ], + "angle": 0, + "content": "(b) Impact of coefficient \\(\\gamma\\)" + }, + { + "type": "table", + "bbox": [ + 0.549, + 0.089, + 0.899, + 0.17 + ], + "angle": 0, + "content": "
PromptStage-ICIDErR@5/16
Contextual ADs [21]X12.6 (17.8)39.8 (43.1)
✓14.1 (19.0)39.9 (44.2)
Character [22]X22.045.6
✓23.146.2
" + }, + { + "type": "table_caption", + "bbox": [ + 0.612, + 0.171, + 0.837, + 0.184 + ], + "angle": 0, + "content": "(c) Impact of Stage-I w/ different prompts." + }, + { + "type": "table_caption", + "bbox": [ + 0.09, + 0.189, + 0.907, + 0.247 + ], + "angle": 0, + "content": "Table 4. Ablation studies for Stage-I. (a) Evaluation of global video-AD loss \\(\\mathcal{L}_f\\) and fine-grained frame-AD loss \\(\\mathcal{L}_f\\) on AD performance. (b) Analysis of the impact of the coefficient \\(\\gamma\\). Both (a) and (b) are conducted with pure visual prompts. (c) Impact of Stage-I when combined with different prompts for the LLM decoder, including contextual ADs and character names. Performance in parentheses indicates results with ground-truth contextual ADs as prompts." + }, + { + "type": "image", + "bbox": [ + 0.094, + 0.251, + 0.482, + 0.32 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.09, + 0.323, + 0.483, + 0.405 + ], + "angle": 0, + "content": "Figure 4. Ablation studies for hyperparameter in Stage-II, with final settings highlighted in orange. (a) Impact of \\(\\alpha\\) on the weight of compact representation \\(\\widehat{\\mathcal{H}}\\). (b) Influence of \\(\\beta\\) on cross-attended feature \\(\\widetilde{\\mathcal{H}}\\). (c) Impact of \\(K\\), which denotes the number of clusters in bases \\(\\mathcal{M}\\). (d) Effect of sampling \\(N\\) consecutive video clips. We switch to larger memory GPUs when \\(N\\) exceeds 16." + }, + { + "type": "table", + "bbox": [ + 0.126, + 0.407, + 0.451, + 0.543 + ], + "angle": 0, + "content": "
Ex#αHβHLdistCIDErR@5/16
A0XXX23.1 (27.4)46.2
B1✓XX23.7 (29.3)46.6
B2X✓X23.4 (29.1)46.1
B3✓✓X23.3 (28.1)48.0
C0XX✓24.7 (29.4)49.5
C1✓X✓24.3 (29.4)50.7
C2X✓✓25.3 (30.4)51.5
C3✓✓✓25.5 (29.8)51.7
" + }, + { + "type": "table_caption", + "bbox": [ + 0.09, + 0.546, + 0.483, + 0.603 + ], + "angle": 0, + "content": "Table 5. Ablation studies for components in Stage-II. The CIDEr column shows scores with AutoAD-Zero's character [82] as prompt by default. CIDEr in parentheses represent performance with ground-truth character names." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.611, + 0.483, + 0.839 + ], + "angle": 0, + "content": "Comparisons on CMD-AD and TV-AD. We further verify the generalizability of DistinctAD on the recently proposed CMD-AD and TV-AD benchmarks, with results presented in Tables 2 and 3. Both evaluations use adapted \\(\\mathrm{CLIP}_{\\mathrm{AD}}\\)-B16 model. DistinctAD outperforms AutoAD-Zero, AutoAD-II and AutoAD-III in terms of CIDEr and R@1/5 on both benchmarks Meanwhile, DistinctAD exhibits a lower CIDEr compared to AutoAD-III† on CMD-AD, which we conjecture is primarily due to AutoAD-III† pre-training on a very large-scale 3.4M transformed HowTo-AD dataset [23, 48], which is currently publicly unavailable. Despite this, DistinctAD achieves superior R@1/5, underscoring its exceptional ability to generate distinctive and high-quality ADs. This is further corroborated by its leading performance on the LLM-AD-eval metric." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.848, + 0.253, + 0.862 + ], + "angle": 0, + "content": "4.3. Ablation studies" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.871, + 0.483, + 0.902 + ], + "angle": 0, + "content": "Effect of CLIP-AD Adaptation (I). Tab. 4a demonstrates the benefit of our Stage-I strategy, i.e. adapting CLIP the vi" + }, + { + "type": "table", + "bbox": [ + 0.589, + 0.251, + 0.832, + 0.306 + ], + "angle": 0, + "content": "
Consecutive N?CIDErR@5/16
✓25.551.7
×23.8↓1.752.5↑0.8
" + }, + { + "type": "table_caption", + "bbox": [ + 0.516, + 0.31, + 0.901, + 0.324 + ], + "angle": 0, + "content": "Table 6. Impact of whether sampling \\(N\\) consecutive clips or not." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.339, + 0.905, + 0.49 + ], + "angle": 0, + "content": "sion encoder to the movie-AD domain via global video-AD matching \\(\\mathcal{L}_g\\) and fine-grained frame-AD matching \\(\\mathcal{L}_f\\). In Tab. 4b, the balancing coefficient \\(\\gamma\\) performs best at 0.5. In Tab. 4c, our Stage-I strategy consistently enhances performance when combined with different prompts in the LLM decoder, such as contextual ADs in AutoAD-I [21] or character names in AutoAD-II [22]. This indicates that our AD-adapted CLIP vision encoder can integrate seamlessly into previous methods, including those training-free models that utilize CLIP-based visual extractors." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.492, + 0.906, + 0.764 + ], + "angle": 0, + "content": "Effect of Distinctive AD narration (II). We evaluate the effectiveness of Stage-II components in Tab. 5, based on the default \\(\\mathcal{H}\\) (Perceiver's output) and full AD auto-regressive loss \\(\\mathcal{L}_{\\text {auto }}\\). The baseline (A0) outperforms AutoAD-II in CIDEr (23.1 vs. 19.5), primarily due to more accurate character prompts from AutoAD-Zero and the adapted CLIP vision encoder from Stage-I. A0 with AutoAD-II's characters achieves a CIDEr score of 20.6, close to AutoAD-II's performance. Incorporating reconstructed feature \\(\\widehat{\\mathcal{H}}\\) brings stable improvements on both CIDEr and recall (B1 & C1), highlighting the importance of compact representations in understanding visual semantics. Cross-attended feature \\(\\widetilde{\\mathcal{H}}\\) works better together with distinctive word prediction loss \\(\\mathcal{L}_{\\text {dist }}\\) (C2 vs. B2, C3 vs. B3). We conjecture this is because \\(\\mathcal{L}_{\\text {dist }}\\) provides more definite supervision on re-weighting distinctive words, which guides \\(\\widehat{\\mathcal{H}}\\) to attend on concept-related bases. Overall, applying the full Stage-II pipeline brings significant and robust performance (C3)." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.765, + 0.907, + 0.902 + ], + "angle": 0, + "content": "Effect of Hyper-parameters. Fig. 4 summarizes the ablation studies on 4 hyper-parameters that potentially influence the results in Stage-II. Coefficient weights \\(\\alpha\\) and \\(\\beta\\) yield optimal results when set to 3 and 1, respectively. This suggests the need to refine our final representations to be more compact for generating ADs. Fig. 4(c) shows setting bases number \\(K\\) to 32 yields best. A smaller \\(K\\), e.g. 2, can still achieve notable CIDEr, as the Contextual EMA module does not significantly alter the final output \\(\\mathcal{H}_{sum}\\). However," + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "13577" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.123, + 0.095, + 0.143, + 0.109 + ], + "angle": 0, + "content": null + }, + { + "type": "title", + "bbox": [ + 0.146, + 0.097, + 0.216, + 0.108 + ], + "angle": 0, + "content": "1h12m53.6s" + }, + { + "type": "image", + "bbox": [ + 0.128, + 0.111, + 0.383, + 0.153 + ], + "angle": 0, + "content": null + }, + { + "type": "text", + "bbox": [ + 0.125, + 0.155, + 0.368, + 0.18 + ], + "angle": 0, + "content": "GT: Holding the room's phone to his ear, the man on the bed regards him sympathetically." + }, + { + "type": "text", + "bbox": [ + 0.126, + 0.181, + 0.336, + 0.194 + ], + "angle": 0, + "content": "AutoAD-Zero: He looks at something." + }, + { + "type": "text", + "bbox": [ + 0.126, + 0.195, + 0.356, + 0.207 + ], + "angle": 0, + "content": "DistinctAD: He holds the phone to his ear." + }, + { + "type": "image", + "bbox": [ + 0.408, + 0.096, + 0.424, + 0.109 + ], + "angle": 0, + "content": null + }, + { + "type": "title", + "bbox": [ + 0.429, + 0.097, + 0.498, + 0.108 + ], + "angle": 0, + "content": "1h12m59.1s" + }, + { + "type": "image", + "bbox": [ + 0.409, + 0.111, + 0.666, + 0.153 + ], + "angle": 0, + "content": null + }, + { + "type": "text", + "bbox": [ + 0.41, + 0.155, + 0.614, + 0.179 + ], + "angle": 0, + "content": "GT: Stephen sees the pill bottles on a bedside table." + }, + { + "type": "text", + "bbox": [ + 0.411, + 0.181, + 0.647, + 0.194 + ], + "angle": 0, + "content": "AutoAD-Zero: Stephen looks at something." + }, + { + "type": "text", + "bbox": [ + 0.411, + 0.195, + 0.635, + 0.219 + ], + "angle": 0, + "content": "DistinctAD: In Stephen's bedroom, a pill bottle sits on his nightstand." + }, + { + "type": "image", + "bbox": [ + 0.691, + 0.096, + 0.71, + 0.109 + ], + "angle": 0, + "content": null + }, + { + "type": "title", + "bbox": [ + 0.713, + 0.097, + 0.781, + 0.108 + ], + "angle": 0, + "content": "1h13m14.5s" + }, + { + "type": "image", + "bbox": [ + 0.691, + 0.11, + 0.862, + 0.153 + ], + "angle": 0, + "content": null + }, + { + "type": "text", + "bbox": [ + 0.651, + 0.155, + 0.866, + 0.18 + ], + "angle": 0, + "content": "GT: The man on the bed sits in silhouette against the room's soft lamp light." + }, + { + "type": "text", + "bbox": [ + 0.659, + 0.181, + 0.84, + 0.205 + ], + "angle": 0, + "content": "AutoAD-Zero: He turns to leave, looking at him." + }, + { + "type": "text", + "bbox": [ + 0.659, + 0.206, + 0.825, + 0.219 + ], + "angle": 0, + "content": "DistinctAD: The bed is empty." + }, + { + "type": "image", + "bbox": [ + 0.125, + 0.219, + 0.144, + 0.233 + ], + "angle": 0, + "content": null + }, + { + "type": "title", + "bbox": [ + 0.147, + 0.221, + 0.217, + 0.232 + ], + "angle": 0, + "content": "1h13m19.4s" + }, + { + "type": "image", + "bbox": [ + 0.128, + 0.235, + 0.299, + 0.277 + ], + "angle": 0, + "content": null + }, + { + "type": "text", + "bbox": [ + 0.125, + 0.28, + 0.29, + 0.304 + ], + "angle": 0, + "content": "GT: Later, Stephen steps back out into the rainy night." + }, + { + "type": "text", + "bbox": [ + 0.125, + 0.306, + 0.299, + 0.331 + ], + "angle": 0, + "content": "AutoAD-Zero: Stephen looks around and at another character." + }, + { + "type": "text", + "bbox": [ + 0.125, + 0.332, + 0.315, + 0.357 + ], + "angle": 0, + "content": "DistinctAD: As Stephen walks away, he glances over his shoulder." + }, + { + "type": "image", + "bbox": [ + 0.325, + 0.22, + 0.344, + 0.233 + ], + "angle": 0, + "content": null + }, + { + "type": "title", + "bbox": [ + 0.346, + 0.222, + 0.413, + 0.232 + ], + "angle": 0, + "content": "1h13m23.1s" + }, + { + "type": "image", + "bbox": [ + 0.325, + 0.235, + 0.496, + 0.277 + ], + "angle": 0, + "content": null + }, + { + "type": "text", + "bbox": [ + 0.322, + 0.28, + 0.545, + 0.304 + ], + "angle": 0, + "content": "GT: Moving with a stiff gait, Stephen marches across the parking lot to his car." + }, + { + "type": "text", + "bbox": [ + 0.322, + 0.306, + 0.537, + 0.32 + ], + "angle": 0, + "content": "AutoAD-Zero: He looks out at the city." + }, + { + "type": "text", + "bbox": [ + 0.322, + 0.322, + 0.543, + 0.346 + ], + "angle": 0, + "content": "DistinctAD: He looks up at the sky, then turns and heads back to the car." + }, + { + "type": "image", + "bbox": [ + 0.546, + 0.22, + 0.565, + 0.233 + ], + "angle": 0, + "content": null + }, + { + "type": "title", + "bbox": [ + 0.546, + 0.222, + 0.609, + 0.233 + ], + "angle": 0, + "content": "h13m24.5s" + }, + { + "type": "image", + "bbox": [ + 0.522, + 0.235, + 0.862, + 0.277 + ], + "angle": 0, + "content": null + }, + { + "type": "text", + "bbox": [ + 0.562, + 0.28, + 0.843, + 0.304 + ], + "angle": 0, + "content": "GT: As Stephen gets in, rain patters heavily against his windshield." + }, + { + "type": "text", + "bbox": [ + 0.562, + 0.306, + 0.804, + 0.319 + ], + "angle": 0, + "content": "AutoAD-Zero: Stephen looks at the women." + }, + { + "type": "text", + "bbox": [ + 0.562, + 0.322, + 0.857, + 0.335 + ], + "angle": 0, + "content": "DistinctAD: In the car, Stephen stares out the window." + }, + { + "type": "text", + "bbox": [ + 0.645, + 0.345, + 0.856, + 0.358 + ], + "angle": 0, + "content": ": The Ides of March (2011)" + }, + { + "type": "image_caption", + "bbox": [ + 0.09, + 0.367, + 0.907, + 0.396 + ], + "angle": 0, + "content": "Figure 5. Qualitative results. We present ground-truth (GT) ADs, publicly released AutoAD-Zero outputs, and our DistinctAD predictions for several temporally consecutive movie clips. Movie frames are taken from The Ides of March (2011) [15]. Zoom in for details." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.412, + 0.484, + 0.501 + ], + "angle": 0, + "content": "unsuitable values of \\( K \\), e.g. 8 or 64, can negatively impact performance. Fig. 4(d) reveals that increasing the number of consecutive clips \\( N \\) (with \\( K \\) set to 32) enhances the CIDEr score, though this effect saturates when \\( N \\) exceeds 16. This demonstrates that more bases should be created to effectively summarize components with additional clips." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.503, + 0.484, + 0.763 + ], + "angle": 0, + "content": "Do the \\(N\\) clips to be consecutive? Tab. 6 presents the results of sampling non-consecutive \\(N\\) clips during training. When using non-continuous clips, we observe a decrease in the CIDEr metric by 1.7 (25.5 vs. 23.8) because the Contextual EMA module struggles with unrelated contexts. However, the R@5/16 improves by 0.8, which indicates enhancement of the distinctiveness (uniqueness) of the generated AD when using more diverse visual contents. Visualizations. To better understand what Contextual EMA learns, we show the t-SNE [71] visualizations of \\(\\mathcal{H}\\), \\(\\widehat{\\mathcal{H}}\\) and \\(\\widetilde{\\mathcal{H}}\\) (from §3.2) in Fig. 6, using the same perplexity value across all visualizations. With Contextual EMA, \\(\\widehat{\\mathcal{H}}\\) exhibits more compact features compared to raw \\(\\mathcal{H}\\), Fig. 6(b). Interestingly in Fig. 6(c), cross-attention between \\(\\mathcal{H}\\) and bases \\(\\mathcal{M}\\) produces strip-like feature distributions pointing to specific base centers, enhancing contextual distinctiveness with improved linear separability and interpretability." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.771, + 0.268, + 0.787 + ], + "angle": 0, + "content": "4.4. Qualitative results" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.795, + 0.484, + 0.902 + ], + "angle": 0, + "content": "Fig. 5 presents qualitative examples of our model. We compare the predictions of DistinctAD (using LLaMA3-8B) with ground-truth captions (GT) and publicly available AutoAD-Zero [82] outputs. Note that clips are sampled consecutively in time. Previous studies often struggle with similar contextual clips, such as those featuring closely-related scenes and characters, by repeating correct yet in" + }, + { + "type": "image", + "bbox": [ + 0.523, + 0.405, + 0.641, + 0.495 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.569, + 0.496, + 0.596, + 0.505 + ], + "angle": 0, + "content": "(a) \\(\\mathcal{H}\\)" + }, + { + "type": "image", + "bbox": [ + 0.653, + 0.405, + 0.769, + 0.495 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.698, + 0.496, + 0.726, + 0.506 + ], + "angle": 0, + "content": "(b) \\(\\widehat{\\mathcal{H}}\\)" + }, + { + "type": "image", + "bbox": [ + 0.781, + 0.406, + 0.897, + 0.494 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.828, + 0.495, + 0.855, + 0.506 + ], + "angle": 0, + "content": "(c) \\(\\widetilde{\\mathcal{H}}\\)" + }, + { + "type": "image_caption", + "bbox": [ + 0.513, + 0.507, + 0.907, + 0.563 + ], + "angle": 0, + "content": "Figure 6. Visualizations of Contextual EMA. (a) A set of randomly generated 3D data \\(\\mathcal{H}\\), sampled from \\(N\\) types of samples. (b) Compact features \\(\\widehat{\\mathcal{H}}\\) obtained via Data Re-estimation (DR). (c) Cross-attention outputs \\(\\widetilde{\\mathcal{H}}\\) between \\(\\mathcal{H}\\) and bases \\(\\mathcal{M}\\)." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.575, + 0.907, + 0.651 + ], + "angle": 0, + "content": "significant action words, e.g. \"look\". In contrast, our DistinctAD effectively generates more engaging ADs by identifying distinctive objects in adjacent clips, e.g. \"phone\", \"pill bottle\", and \"car\", along with corresponding more specific behaviors. More examples can be found in Appendix §E." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.664, + 0.634, + 0.68 + ], + "angle": 0, + "content": "5. Conclusion" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.69, + 0.907, + 0.901 + ], + "angle": 0, + "content": "This paper proposes DistinctAD, a two-stage framework for generating distinctive audio descriptions for better narrative. By addressing the domain gap between movie-AD data with a CLIP-AD adaptation strategy, and introducing a Contextual EMA module and a distinctive word prediction loss, our approach significantly improves the quality of AD generation. The effectiveness of DistinctAD is demonstrated through comprehensive evaluations on multiple benchmark datasets and ablations studies. Despite these promising results, DistinctAD is still limited by requiring numbers of parameters and the quality of the generated ADs still falls short of human annotations. Overall, automatic AD generation still remains a challenging task, and there is considerable scope for future advancements in this field." + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "13578" + } + ], + [ + { + "type": "title", + "bbox": [ + 0.093, + 0.091, + 0.251, + 0.108 + ], + "angle": 0, + "content": "Acknowledgement" + }, + { + "type": "text", + "bbox": [ + 0.091, + 0.114, + 0.484, + 0.145 + ], + "angle": 0, + "content": "This work was supported by a Strategic Research Grant from City University of Hong Kong (Project No. 7005840)." + }, + { + "type": "title", + "bbox": [ + 0.093, + 0.17, + 0.188, + 0.185 + ], + "angle": 0, + "content": "References" + }, + { + "type": "ref_text", + "bbox": [ + 0.101, + 0.196, + 0.376, + 0.209 + ], + "angle": 0, + "content": "[1] https://audiovault.net, 2019.3" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.211, + 0.48, + 0.237 + ], + "angle": 0, + "content": "[2] Vlog: Video as a long document. https://github.com/showlab/VLog, 2023. GitHub repository. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.239, + 0.483, + 0.306 + ], + "angle": 0, + "content": "[3] Josh Achiam, Steven Adler, Sandhini Agarwal, Lama Ahmad, Ilge Akkaya, Florencia Leoni Aleman, Diogo Almeida, Janko Altenschmidt, Sam Altman, Shyamal Anadkat, et al. Gpt-4 technical report. arXiv preprint arXiv:2303.08774, 2023. 1, 2, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.309, + 0.398, + 0.322 + ], + "angle": 0, + "content": "[4] AI@Meta. Llama 3 model card. 2024. 5, 6, 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.324, + 0.483, + 0.391 + ], + "angle": 0, + "content": "[5] Jean-Baptiste Alayrac, Jeff Donahue, Pauline Luc, Antoine Miech, Iain Barr, Yana Hasson, Karel Lenc, Arthur Mensch, Katherine Millican, Malcolm Reynolds, et al. Flamingo: a visual language model for few-shot learning. NeurIPS, 35: 23716-23736, 2022. 1, 3, 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.394, + 0.483, + 0.448 + ], + "angle": 0, + "content": "[6] Peter Anderson, Basura Fernando, Mark Johnson, and Stephen Gould. Spice: Semantic propositional image caption evaluation. In ECCV, pages 382-398. Springer, 2016. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.451, + 0.483, + 0.491 + ], + "angle": 0, + "content": "[7] Max Bain, Arsha Nagrani, Andrew Brown, and Andrew Zisserman. Condensed movies: Story based retrieval with contextual embeddings. In ACCV, 2020. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.493, + 0.483, + 0.533 + ], + "angle": 0, + "content": "[8] Max Bain, Jaesung Huh, Tengda Han, and Andrew Zisserman. Whisperx: Time-accurate speech transcription of long-form audio. arXiv preprint arXiv:2303.00747, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.536, + 0.483, + 0.589 + ], + "angle": 0, + "content": "[9] Carmen J Branje and Deborah I Fels. Livedescribe: can amateur descriptors create high-quality audio description? Journal of Visual Impairment & Blindness, 106(3):154-165, 2012. 1, 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.592, + 0.483, + 0.631 + ], + "angle": 0, + "content": "[10] Hervé Bredin and Antoine Laurent. End-to-end speaker segmentation for overlap-aware resegmentation. In Interspeech, 2021. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.634, + 0.483, + 0.702 + ], + "angle": 0, + "content": "[11] Hervé Bredin, Ruiqing Yin, Juan Manuel Coria, Gregory Gelly, Pavel Korshunov, Marvin Lavechin, Diego Fustes, Hadrien Titeux, Wassim Bouaziz, and Marie-Philippe Gill. Pyannote. audio: neural building blocks for speaker diarization. In ICASSP, pages 7124–7128. IEEE, 2020. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.704, + 0.483, + 0.759 + ], + "angle": 0, + "content": "[12] Fuhai Chen, Rongrong Ji, Xiaoshuai Sun, Yongjian Wu, and Jinsong Su. Groupcap: Group-based image captioning with structured relevance and diversity constraints. In CVPR, pages 1345-1353, 2018. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.761, + 0.483, + 0.801 + ], + "angle": 0, + "content": "[13] Shaoxiang Chen and Yu-Gang Jiang. Towards bridging event captioner and sentence localizer for weakly supervised dense event captioning. In CVPR, pages 8425-8435, 2021. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.803, + 0.483, + 0.843 + ], + "angle": 0, + "content": "[14] Peng Chu, Jiang Wang, and Andre Abrantes. Llm-ad: Large language model based audio description system. arXiv preprint arXiv:2405.00983, 2024. 1, 2, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.846, + 0.483, + 0.871 + ], + "angle": 0, + "content": "[15] George Clooney. The idees of march. *Columbia Pictures*, 2011. 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.873, + 0.483, + 0.901 + ], + "angle": 0, + "content": "[16] Bo Dai and Dahua Lin. Contrastive learning for image captioning. NeurIPS, 30, 2017. 3" + }, + { + "type": "list", + "bbox": [ + 0.094, + 0.196, + 0.483, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.093, + 0.906, + 0.147 + ], + "angle": 0, + "content": "[17] Arthur P Dempster, Nan M Laird, and Donald B Rubin. Maximum likelihood from incomplete data via the em algorithm. Journal of the royal statistical society: series B (methodological), 39(1):1-22, 1977. 2, 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.149, + 0.905, + 0.189 + ], + "angle": 0, + "content": "[18] Chaorui Deng, Shizhe Chen, Da Chen, Yuan He, and Qi Wu. Sketch, ground, and refine: Top-down dense video captioning. In CVPR, pages 234-243, 2021. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.191, + 0.905, + 0.257 + ], + "angle": 0, + "content": "[19] Lisa Dunlap, Yuhui Zhang, Xiaohan Wang, Ruiqi Zhong, Trevor Darrell, Jacob Steinhardt, Joseph E Gonzalez, and Serena Yeung-Levy. Describing differences in image sets with natural language. In CVPR, pages 24199-24208, 2024. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.261, + 0.905, + 0.287 + ], + "angle": 0, + "content": "[20] Louise Fryer. An introduction to audio description: A practical guide, 2016. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.289, + 0.905, + 0.342 + ], + "angle": 0, + "content": "[21] Tengda Han, Max Bain, Arsha Nagrani, Gül Varol, Weidi Xie, and Andrew Zisserman. Autoad: Movie description in context. In CVPR, pages 18930-18940, 2023. 1, 2, 3, 4, 5, 6, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.344, + 0.905, + 0.398 + ], + "angle": 0, + "content": "[22] Tengda Han, Max Bain, Arsha Nagrani, Gul Varol, Weidi Xie, and Andrew Zisserman. Autoad ii: The sequel-who, when, and what in movie audio description. In ICCV, pages 13645–13655, 2023. 1, 3, 4, 5, 6, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.4, + 0.905, + 0.453 + ], + "angle": 0, + "content": "[23] Tengda Han, Max Bain, Arsha Nagrani, Gül Varol, Weidi Xie, and Andrew Zisserman. Autoad iii: The prequel-back to the pixels. In CVPR, pages 18164-18174, 2024. 1, 2, 3, 5, 6, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.455, + 0.905, + 0.51 + ], + "angle": 0, + "content": "[24] Edward J Hu, Yelong Shen, Phillip Wallis, Zeyuan Allen-Zhu, Yuanzhi Li, Shean Wang, Lu Wang, and Weizhu Chen. Lora: Low-rank adaptation of large language models. arXiv preprint arXiv:2106.09685, 2021. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.511, + 0.905, + 0.551 + ], + "angle": 0, + "content": "[25] Vladimir Iashin and Esa Rahtu. A better use of audio-visual cues: Dense video captioning with bi-modal transformer. In BMVC, 2020. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.553, + 0.905, + 0.58 + ], + "angle": 0, + "content": "[26] Vladimir Iashin and Esa Rahtu. Multi-modal dense video captioning. In CVPRW, pages 958-959, 2020. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.581, + 0.905, + 0.648 + ], + "angle": 0, + "content": "[27] Peng Jin, Jina Huang, Fenglin Liu, Xian Wu, Shen Ge, Guoli Song, David Clifton, and Jie Chen. Expectation-maximization contrastive learning for compact video-and-language representations. NeurIPS, 35:30291-30306, 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.651, + 0.905, + 0.677 + ], + "angle": 0, + "content": "[28] Diederik P Kingma. Adam: A method for stochastic optimization. arXiv preprint arXiv:1412.6980, 2014. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.679, + 0.905, + 0.718 + ], + "angle": 0, + "content": "[29] Ranjay Krishna, Kenji Hata, Frederic Ren, Li Fei-Fei, and Juan Carlos Niebles. Dense-captioning events in videos. In ICCV, pages 706–715, 2017. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.72, + 0.905, + 0.76 + ], + "angle": 0, + "content": "[30] Jie Lei, Licheng Yu, Mohit Bansal, and Tamara L Berg. Tvqa: Localized, compositional video question answering. In EMNLP, 2018. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.762, + 0.905, + 0.789 + ], + "angle": 0, + "content": "[31] Elisa Lewis. Deep dive: How audio description benefits everyone, 2021. Accessed on, pages 11-13, 2023. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.79, + 0.905, + 0.843 + ], + "angle": 0, + "content": "[32] Bo Li, Yuanhan Zhang, Liangyu Chen, Jinghao Wang, Fanyi Pu, Jingkang Yang, Chunyuan Li, and Ziwei Liu. Mimicit: Multi-modal in-context instruction tuning. arXiv preprint arXiv:2306.05425, 2023. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.846, + 0.905, + 0.9 + ], + "angle": 0, + "content": "[33] Bo Li, Yuanhan Zhang, Liangyu Chen, Jinghao Wang, Jingkang Yang, and Ziwei Liu. Otter: a multi-modal model with in-context instruction tuning. corr abs/2305.03726 (2023), 2023. 3" + }, + { + "type": "list", + "bbox": [ + 0.517, + 0.093, + 0.906, + 0.9 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.52, + 0.957 + ], + "angle": 0, + "content": "13579" + } + ], + [ + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.092, + 0.482, + 0.147 + ], + "angle": 0, + "content": "[34] Junnan Li, Dongxu Li, Silvio Savarese, and Steven Hoi. Blip-2: Bootstrapping language-image pre-training with frozen image encoders and large language models. arXiv preprint arXiv:2301.12597, 2023. 1, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.149, + 0.483, + 0.203 + ], + "angle": 0, + "content": "[35] Xia Li, Zhisheng Zhong, Jianlong Wu, Yibo Yang, Zhouchen Lin, and Hong Liu. Expectation-maximization attention networks for semantic segmentation. In ICCV, pages 9167-9176, 2019. 2, 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.205, + 0.482, + 0.247 + ], + "angle": 0, + "content": "[36] Yehao Li, Ting Yao, Yingwei Pan, Hongyang Chao, and Tao Mei. Jointly localizing and describing events for dense video captioning. In CVPR, pages 7492-7500, 2018. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.248, + 0.482, + 0.286 + ], + "angle": 0, + "content": "[37] Chin-Yew Lin. Rouge: A package for automatic evaluation of summaries. In Text summarization branches out, pages 74-81, 2004. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.288, + 0.482, + 0.343 + ], + "angle": 0, + "content": "[38] Kevin Lin, Linjie Li, Chung-Ching Lin, Faisal Ahmed, Zhe Gan, Zicheng Liu, Yumao Lu, and Lijuan Wang. Swinbert: End-to-end transformers with sparse attention for video captioning. In CVPR, pages 17949-17958, 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.344, + 0.482, + 0.412 + ], + "angle": 0, + "content": "[39] Kevin Lin, Faisal Ahmed, Linjie Li, Chung-Ching Lin, Ehsan Azarnasab, Zhengyuan Yang, Jianfeng Wang, Lin Liang, Zicheng Liu, Yumao Lu, et al. Mm-vid: Advancing video understanding with gpt-4v (ision). arXiv preprint arXiv:2310.19773, 2023. 1, 2, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.414, + 0.482, + 0.481 + ], + "angle": 0, + "content": "[40] Kevin Qinghong Lin, Pengchuan Zhang, Difei Gao, Xide Xia, Joya Chen, Ziteng Gao, Jinheng Xie, Xuhong Xiao, and Mike Zheng Shou. Learning video context as interleaved multimodal sequences. arXiv preprint arXiv:2407.21757, 2024. 1, 3, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.483, + 0.482, + 0.55 + ], + "angle": 0, + "content": "[41] Zhihui Lin, Tianyu Yang, Maomao Li, Ziyu Wang, Chun Yuan, Wenhao Jiang, and Wei Liu. Swem: Towards real-time video object segmentation with sequential weighted expectation-maximization. In CVPR, pages 1362-1372, 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.552, + 0.482, + 0.58 + ], + "angle": 0, + "content": "[42] Haotian Liu, Chunyuan Li, Qingyang Wu, and Yong Jae Lee. Visual instruction tuning. NeurIPS, 36, 2024. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.581, + 0.482, + 0.635 + ], + "angle": 0, + "content": "[43] Xihui Liu, Hongsheng Li, Jing Shao, Dapeng Chen, and Xiaogang Wang. Show, tell and discriminate: Image captioning by self-retrieval with partially labeled data. In ECCV, pages 338-354, 2018. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.637, + 0.482, + 0.664 + ], + "angle": 0, + "content": "[44] I Loshchilov. Decoupled weight decay regularization. arXiv preprint arXiv:1711.05101, 2017. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.665, + 0.482, + 0.733 + ], + "angle": 0, + "content": "[45] Huaishao Luo, Lei Ji, Botian Shi, Haoyang Huang, Nan Duan, Tianrui Li, Jason Li, Taroon Bharti, and Ming Zhou. Univl: A unified video and language pre-training model for multimodal understanding and generation. arXiv preprint arXiv:2002.06353, 2020. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.735, + 0.482, + 0.789 + ], + "angle": 0, + "content": "[46] Huaishao Luo, Lei Ji, Ming Zhong, Yang Chen, Wen Lei, Nan Duan, and Tianrui Li. Clip4clip: An empirical study of clip for end to end video clip retrieval and captioning. Neurocomputing, 508:293-304, 2022. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.79, + 0.482, + 0.831 + ], + "angle": 0, + "content": "[47] Ruotian Luo, Brian Price, Scott Cohen, and Gregory Shakhnarovich. Discriminability objective for training descriptive captions. In CVPR, pages 6964-6974, 2018. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.832, + 0.482, + 0.9 + ], + "angle": 0, + "content": "[48] Antoine Miech, Dimitri Zhukov, Jean-Baptiste Alayrac, Makarand Tapaswi, Ivan Laptev, and Josef Sivic. Howto100m: Learning a text-video embedding by watching hundred million narrated video clips. In ICCV, pages 2630-2640, 2019. 6, 7" + }, + { + "type": "list", + "bbox": [ + 0.093, + 0.092, + 0.483, + 0.9 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.516, + 0.092, + 0.905, + 0.147 + ], + "angle": 0, + "content": "[49] Antoine Miech, Jean-Baptiste Alayrac, Lucas Smaira, Ivan Laptev, Josef Sivic, and Andrew Zisserman. End-to-end learning of visual representations from uncurated instructional videos. In CVPR, pages 9879-9889, 2020. 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.516, + 0.149, + 0.905, + 0.19 + ], + "angle": 0, + "content": "[50] Ron Mokady, Amir Hertz, and Amit H Bermano. Clipcap: Clip prefix for image captioning. arXiv preprint arXiv:2111.09734, 2021. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.516, + 0.192, + 0.905, + 0.233 + ], + "angle": 0, + "content": "[51] Jonghwan Mun, Linjie Yang, Zhou Ren, Ning Xu, and Bohyung Han. Streamlined dense video captioning. In CVPR, pages 6588-6597, 2019. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.516, + 0.235, + 0.905, + 0.276 + ], + "angle": 0, + "content": "[52] David Nukrai, Ron Mokady, and Amir Globerson. Text-only training for image captioning using noise-injected clip. arXiv preprint arXiv:2211.00575, 2022. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.516, + 0.278, + 0.844, + 0.291 + ], + "angle": 0, + "content": "[53] OpenAI. Gpt-4v(ision) system card. 2023. 1, 2, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.516, + 0.293, + 0.905, + 0.359 + ], + "angle": 0, + "content": "[54] Amy Pavel, Gabriel Reyes, and Jeffrey P Bigham. Describe: Authoring and automatically editing audio descriptions. In Proceedings of the 33rd Annual ACM Symposium on User Interface Software and Technology, pages 747-759, 2020. 1, 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.516, + 0.363, + 0.905, + 0.39 + ], + "angle": 0, + "content": "[55] Elisa Perego. Gains and losses of watching audio described films for sighted viewers. *Target*, 28(3):424-444, 2016. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.516, + 0.392, + 0.905, + 0.446 + ], + "angle": 0, + "content": "[56] Haran Raajesh, Naveen Reddy Desanur, Zeeshan Khan, and Makarand Tapaswi. Micap: A unified model for identity-aware movie descriptions. In CVPR, pages 14011-14021, 2024. 1, 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.516, + 0.449, + 0.905, + 0.49 + ], + "angle": 0, + "content": "[57] Alec Radford, Jeffrey Wu, Rewon Child, David Luan, Dario Amodei, Ilya Sutskever, et al. Language models are unsupervised multitask learners. OpenAI blog, 1(8):9, 2019. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.516, + 0.491, + 0.905, + 0.56 + ], + "angle": 0, + "content": "[58] Alec Radford, Jong Wook Kim, Chris Hallacy, Aditya Ramesh, Gabriel Goh, Sandhini Agarwal, Girish Sastry, Amanda Askell, Pamela Mishkin, Jack Clark, et al. Learning transferable visual models from natural language supervision. In ICML, pages 8748-8763. PMLR, 2021. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.516, + 0.562, + 0.905, + 0.603 + ], + "angle": 0, + "content": "[59] Tanzila Rahman, Bicheng Xu, and Leonid Sigal. Watch, listen and tell: Multi-modal weakly supervised dense event captioning. In ICCV, pages 8908-8917, 2019. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.516, + 0.604, + 0.905, + 0.644 + ], + "angle": 0, + "content": "[60] Anna Rohrbach, Marcus Rohrbach, Niket Tandon, and Bernt Schiele. A dataset for movie description. In CVPR, pages 3202-3212, 2015. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.516, + 0.647, + 0.905, + 0.701 + ], + "angle": 0, + "content": "[61] Anna Rohrbach, Atousa Torabi, Marcus Rohrbach, Niket Tandon, Christopher Pal, Hugo Larochelle, Aaron Courville, and Bernt Schiele. Movie description. *IJCV*, 123:94–120, 2017. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.516, + 0.703, + 0.905, + 0.758 + ], + "angle": 0, + "content": "[62] Paul Hongsuck Seo, Arsha Nagrani, Anurag Arnab, and Cordelia Schmid. End-to-end generative pretraining for multimodal video captioning. In CVPR, pages 17959-17968, 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.516, + 0.76, + 0.905, + 0.815 + ], + "angle": 0, + "content": "[63] Piyush Sharma, Nan Ding, Sebastian Goodman, and Radu Soricut. Conceptual captions: A cleaned, hypernymed, image alt-text dataset for automatic image captioning. In ACL, pages 2556-2565, 2018. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.516, + 0.817, + 0.905, + 0.87 + ], + "angle": 0, + "content": "[64] Zhiqiang Shen, Jianguo Li, Zhou Su, Minjun Li, Yurong Chen, Yu-Gang Jiang, and Xiangyang Xue. Weakly supervised dense video captioning. In CVPR, pages 1916-1924, 2017. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.516, + 0.873, + 0.905, + 0.902 + ], + "angle": 0, + "content": "[65] Botian Shi, Lei Ji, Yaobo Liang, Nan Duan, Peng Chen, Zhendong Niu, and Ming Zhou. Dense procedure captioning" + }, + { + "type": "list", + "bbox": [ + 0.516, + 0.092, + 0.905, + 0.902 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.52, + 0.957 + ], + "angle": 0, + "content": "13580" + } + ], + [ + { + "type": "ref_text", + "bbox": [ + 0.125, + 0.092, + 0.482, + 0.119 + ], + "angle": 0, + "content": "in narrated instructional videos. In ACL, pages 6382-6391, 2019. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.121, + 0.483, + 0.175 + ], + "angle": 0, + "content": "[66] Aleksandar Shtedritski, Christian Rupprecht, and Andrea Vedaldi. What does clip know about a red circle? visual prompt engineering for vlms. In ICCV, pages 11987-11997, 2023. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.177, + 0.482, + 0.217 + ], + "angle": 0, + "content": "[67] Joel Snyder. Audio description: The visual made verbal. In International congress series, pages 935-939. Elsevier, 2005. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.218, + 0.482, + 0.287 + ], + "angle": 0, + "content": "[68] Mattia Soldan, Alejandro Pardo, Juan León Alcazar, Fabian Caba, Chen Zhao, Silvio Giancola, and Bernard Ghanem. Mad: A scalable dataset for language grounding in videos from movie audio descriptions. In CVPR, pages 5026-5035, 2022. 2, 5, 1, 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.288, + 0.482, + 0.342 + ], + "angle": 0, + "content": "[69] Atousa Torabi, Christopher Pal, Hugo Larochelle, and Aaron Courville. Using descriptive video services to create a large data source for video annotation research. arXiv preprint arXiv:1503.01070, 2015. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.343, + 0.482, + 0.412 + ], + "angle": 0, + "content": "[70] Hugo Touvron, Louis Martin, Kevin Stone, Peter Albert, Amjad Almahairi, Yasmine Babaei, Nikolay Bashlykov, Soumya Batra, Prajjwal Bhargava, Shruti Bhosale, et al. Llama 2: Open foundation and fine-tuned chat models. arXiv preprint arXiv:2307.09288, 2023. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.413, + 0.482, + 0.441 + ], + "angle": 0, + "content": "[71] Laurens Van der Maaten and Geoffrey Hinton. Visualizing data using t-sne. JMLR, 9(11), 2008. 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.442, + 0.482, + 0.482 + ], + "angle": 0, + "content": "[72] Ramakrishna Vedantam, C Lawrence Zitnick, and Devi Parikh. Cider: Consensus-based image description evaluation. In CVPR, pages 4566-4575, 2015. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.483, + 0.482, + 0.524 + ], + "angle": 0, + "content": "[73] Gilad Vered, Gal Oren, Yuval Atzmon, and Gal Chechik. Joint optimization for cooperative image captioning. In ICCV, pages 8898-8907, 2019. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.525, + 0.482, + 0.579 + ], + "angle": 0, + "content": "[74] Hanlin Wang, Zhan Tong, Kecheng Zheng, Yujun Shen, and Limin Wang. Contextual ad narration with interleaved multimodal sequence. arXiv preprint arXiv:2403.12922, 2024. 1, 2, 3, 4, 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.581, + 0.482, + 0.634 + ], + "angle": 0, + "content": "[75] Jingwen Wang, Wenhao Jiang, Lin Ma, Wei Liu, and Yong Xu. Bidirectional attentive fusion with context gating for dense video captioning. In CVPR, pages 7190-7198, 2018. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.636, + 0.482, + 0.691 + ], + "angle": 0, + "content": "[76] Jiuniu Wang, Wenjia Xu, Qingzhong Wang, and Antoni B Chan. Compare and reweight: Distinctive image captioning using similar images sets. In ECCV, pages 370-386. Springer, 2020. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.692, + 0.482, + 0.734 + ], + "angle": 0, + "content": "[77] Jiuniu Wang, Wenjia Xu, Qingzhong Wang, and Antoni B Chan. Group-based distinctive image captioning with memory attention. In ACMMM, pages 5020-5028, 2021." + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.735, + 0.482, + 0.775 + ], + "angle": 0, + "content": "[78] Jiuniu Wang, Wenjia Xu, Qingzhong Wang, and Antoni B Chan. On distinctive image captioning via comparing and reweighting. TPAMI, 45(2):2088-2103, 2022. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.776, + 0.482, + 0.817 + ], + "angle": 0, + "content": "[79] Teng Wang, Huicheng Zheng, Mingjing Yu, Qian Tian, and Haifeng Hu. Event-centric hierarchical representation for dense video captioning. TCVT, 31(5):1890-1900, 2020. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.818, + 0.482, + 0.859 + ], + "angle": 0, + "content": "[80] Teng Wang, Ruimao Zhang, Zhichao Lu, Feng Zheng, Ran Cheng, and Ping Luo. End-to-end dense video captioning with parallel decoding. In ICCV, pages 6847-6857, 2021. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.86, + 0.482, + 0.9 + ], + "angle": 0, + "content": "[81] Xiaolong Wang, Ross Girshick, Abhinav Gupta, and Kaiming He. Non-local neural networks. In CVPR, pages 7794-7803, 2018. 4" + }, + { + "type": "list", + "bbox": [ + 0.093, + 0.092, + 0.483, + 0.9 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.092, + 0.905, + 0.148 + ], + "angle": 0, + "content": "[82] Junyu Xie, Tengda Han, Max Bain, Arsha Nagrani, Güll Varol, Weidi Xie, and Andrew Zisserman. Autoad-zero: A training-free framework for zero-shot audio description. arXiv preprint arXiv:2407.15850, 2024. 1, 2, 4, 5, 6, 7, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.149, + 0.905, + 0.219 + ], + "angle": 0, + "content": "[83] Antoine Yang, Arsha Nagrani, Paul Hongsuck Seo, Antoine Miech, Jordi Pont-Tuset, Ivan Laptev, Josef Sivic, and Cordelia Schmid. Vid2seq: Large-scale pretraining of a visual language model for dense video captioning. In CVPR, pages 10714-10726, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.22, + 0.905, + 0.261 + ], + "angle": 0, + "content": "[84] Linli Yao, Weiying Wang, and Qin Jin. Image difference captioning with pre-training and contrastive learning. In AAAI, pages 3108-3116, 2022. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.262, + 0.783, + 0.276 + ], + "angle": 0, + "content": "[85] Keunwoo Peter Yu. Videoblip, 2023. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.277, + 0.905, + 0.345 + ], + "angle": 0, + "content": "[86] Chaoyi Zhang, Kevin Lin, Zhengyuan Yang, Jianfeng Wang, Linjie Li, Chung-Ching Lin, Zicheng Liu, and Lijuan Wang. Mm-narrator: Narrating long-form videos with multimodal in-context learning. In CVPR, pages 13647-13657, 2024. 1, 2, 4, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.347, + 0.905, + 0.389 + ], + "angle": 0, + "content": "[87] Hang Zhang, Xin Li, and Lidong Bing. Video-llama: An instruction-tuned audio-visual language model for video understanding. arXiv preprint arXiv:2306.02858, 2023. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.391, + 0.905, + 0.432 + ], + "angle": 0, + "content": "[88] Tianyi Zhang, Varsha Kishore, Felix Wu, Kilian Q Weinberger, and Yoav Artzi. Bertscore: Evaluating text generation with bert. arXiv preprint arXiv:1904.09675, 2019. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.434, + 0.905, + 0.475 + ], + "angle": 0, + "content": "[89] Luowei Zhou, Yingbo Zhou, Jason J Corso, Richard Socher, and Caiming Xiong. End-to-end dense video captioning with masked transformer. In CVPR, pages 8739-8748, 2018. 2" + }, + { + "type": "list", + "bbox": [ + 0.517, + 0.092, + 0.905, + 0.475 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.518, + 0.956 + ], + "angle": 0, + "content": "13581" + } + ] +] \ No newline at end of file diff --git a/2025/DistinctAD_ Distinctive Audio Description Generation in Contexts/d4975861-cea8-4f94-9bb1-e02706824334_origin.pdf b/2025/DistinctAD_ Distinctive Audio Description Generation in Contexts/d4975861-cea8-4f94-9bb1-e02706824334_origin.pdf new file mode 100644 index 0000000000000000000000000000000000000000..02f39e562b70de4f292a93e8488496e5734c5b88 --- /dev/null +++ b/2025/DistinctAD_ Distinctive Audio Description Generation in Contexts/d4975861-cea8-4f94-9bb1-e02706824334_origin.pdf @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:ea588c0616fcb8f697b62b146a1e8d7099771c90939763c3a2135185691ab851 +size 2729324 diff --git a/2025/DistinctAD_ Distinctive Audio Description Generation in Contexts/full.md b/2025/DistinctAD_ Distinctive Audio Description Generation in Contexts/full.md new file mode 100644 index 0000000000000000000000000000000000000000..ece05de06d754ab890af6ee3a7ae76b5c550bc6a --- /dev/null +++ b/2025/DistinctAD_ Distinctive Audio Description Generation in Contexts/full.md @@ -0,0 +1,436 @@ +# DistinctAD: Distinctive Audio Description Generation in Contexts + +Bo Fang $^{1}$ , Wenhao Wu $^{2,3}$ , Qiangqiang Wu $^{1}$ , Yuxin Song $^{2}$ , Antoni B. Chan $^{1}$ $^{1}$ Department of Computer Science, City University of Hong Kong + $^{2}$ Baidu Inc. + $^{3}$ The University of Sydney + +{bofang6-c, qiangqwu2-c}@my.cityu.edu.hk, wenhao.wu@sydney.edu.au songyuxin02@baidu.com, abchan@cityu.edu.hk + +# Abstract + +Audio Descriptions (ADs) aim to provide a narration of a movie in text form, describing non-dialogue-related narratives, such as characters, actions, or scene establishment. Automatic generation of ADs remains challenging due to: i) the domain gap between movie-AD data and existing data used to train vision-language models, and ii) the issue of contextual redundancy arising from highly similar neighboring visual clips in a long movie. In this work, we propose DistinctAD, a novel two-stage framework for generating ADs that emphasize distinctiveness to produce better narratives. To address the domain gap, we introduce a CLIP-AD adaptation strategy that does not require additional AD corpora, enabling more effective alignment between movie and AD modalities at both global and fine-grained levels. In Stage-II, DistinctAD incorporates two key innovations: (i) a Contextual Expectation-Maximization Attention (EMA) module that reduces redundancy by extracting common bases from consecutive video clips, and (ii) an explicit distinctive word prediction loss that filters out repeated words in the context, ensuring the prediction of unique terms specific to the current AD. Comprehensive evaluations on MAD-Eval, CMD-AD, and TV-AD benchmarks demonstrate the superiority of DistinctAD, with the model consistently outperforming baselines, particularly in Recall@k/N, highlighting its effectiveness in producing high-quality, distinctive ADs. + +# 1. Introduction + +Audio description (AD) [20, 67] is a crucial accessibility service that provides verbal narration of visual elements in media content for individuals who are blind or have low vision. By offering succinct and vivid descriptions, ADs enable visually impaired audiences to fully comprehend and engage with non-dialogue-related narratives, e.g., characters, facial expressions, non-verbal actions, or scene establishment. Recent studies also show ADs' value for + +![](images/500e9864fe6895eca71b0b1abed5cb42803e90aabfd9e97daf3d43be75d8831b.jpg) + +![](images/b3c76ba5f2d0c37722518ee1c83cab3f34eebc6c4e2d1202cd1b64a869b0da0f.jpg) +Figure 1. (a) Previous methods approach the AD task similar to video captioning, using only a single video clip as input, which leads to repetitive ADs due to highly similar neighboring clips. (b) Our DistinctAD method generates distinctive ADs across $N$ consecutive clips, with three key innovations: VLM-AD adaptation, the Distinct Module, and explicit distinctive words prediction. + +sighted viewers in supporting eye-free activities and facilitating child language development [31, 55], reinforcing its pivotal role in fostering inclusivity by bridging the perceptual gap between visual and non-visual elements. Crafting ADs requires careful attention to timing, language, and context to integrate smoothly with dialogue [22]. However, despite the availability of advanced AD platforms [9, 54], human-annotated methods are costly and difficult to scale, highlighting the need for automated generation systems, especially with the rise of user-generated content. + +Advancements in Vision-Language Models (VLMs) and Large-Language Models (LLMs) have led to growing interest in automatic AD generation for media. Current approaches fall into two categories: (i) using powerful proprietary models like GPT-4 [3] or GPT-4V [53] in a training-free manner [14, 39, 82, 86], and (ii) fine-tuning open-source VLM components, such as visual-text adapters [5, 34, 42], for AD tasks [21-23, 40, 56, 74]. Both approaches have limitations: (i) Training-free methods often perform poorly and suffer from hallucinations due to the unique na + +ture of AD (e.g., character names and narrative coherence), which differs from the common text data LLMs are trained on. (ii) Fine-tuning methods generally perform better but are still limited by insufficient data to fully adapt to the movie-AD domain and face the context-repetition issue. + +Unlike video captioning [38, 62], ADs are generated on consecutive intervals (visual clips) throughout long videos [68], e.g., movies. The context-repetition issue arises when models produce repetitive or similar descriptions for consecutive visual clips, especially when using prior ADs as prompts [21, 74]. This occurs because sequential clips often comprise redundant scenes or characters (and therein redundant visual features), leading models that only use the current visual clip to repeat the same information from the past, as shown in Fig. 1. However, audiences are more interested in the unique and distinct events of the current clip, rather than the common elements from the previous one. + +In this paper, we propose DistinctAD, a two-stage framework for generating distinctive ADs within contexts. Given the domain gap between the movie-AD and VLM training data, we first bridge this gap in Stage-I by adapting VLMs, such as CLIP [58], to the movie-AD domain. Our adaptation strategy is inspired by a key observation (see Appendix §A): AD sentences encoded by the CLIP text encoder can be effectively reconstructed using simple LLMs like GPT-2 [57] with minimal fine-tuning, whereas AD reconstructions using CLIP visual features from the corresponding clips are often of poor quality. This suggests that while CLIP's multi-modal embedding space is rich enough to represent AD information, its visual encoder is insufficient for extracting it. To mitigate this domain gap, we adapt the CLIP vision encoder to better align with the frozen CLIP text encoder using existing paired video-AD data. The alignment involves global matching at video-sentence level, similar to CLIP pre-training. A challenge arises because video clips are labeled with whole ADs, and words may not appear in every frame but must be aggregated over frames. Therefore, we propose fine-grained matching at frame-word level for this multiple-instance setting. + +For Stage-II, we propose a novel distinctive AD narrating pipeline based on the Expectation-Maximization Attention (EMA) [17] algorithm, which has demonstrated its efficacy in tasks such as semantic segmentation [35], video object segmentation [41], and text-video retrieval [27]. Differently, we apply EMA to contextual clips from long videos, which often exhibit high redundancy due to recurring scenes or characters. By extracting common bases from contextual information, DistinctAD reduces redundancy and generates compact, discriminative representations that enable the LLM decoder to produce more distinctive ADs. To further emphasize distinctiveness explicitly, we introduce a distinctive word prediction loss that filters out words that repeatedly appear in contexts, ensuring that the LLM decoder + +focuses on predicting unique words specific to the current AD. With these two designs, DistinctAD produces contextually distinctive and engaging ADs that can provide better narratives for the audience. + +In summary, our contributions are three-fold: + +- We propose a CLIP-AD adaptation strategy tailored to movie-AD data, addressing the misalignment issue caused by the domain gap. Our adapted vision encoder can be seamlessly integrated into existing CLIP-based AD methods and stands to benefit from future improvements as more AD data becomes available. +- We introduce DistinctAD, which incorporates a Contextual EMA module and a distinctive word prediction loss, significantly enhancing the generation of distinctive ADs from consecutive visual clips with similar contexts. +- Comprehensive evaluations on MAD-Eval [21], CMD-AD [23], and TV-AD [82] highlight DistinctAD's superiority. Our outstanding performance in Recall@k/N demonstrates its effectiveness in generating high-quality ADs with both distinctiveness and technical excellence. + +# 2. Related Work + +Dense video captioning. A task closely related to AD is dense video captioning [29], which extends traditional video captioning [38, 45, 62, 63] by both generating a single caption for trimmed video segments as well as detecting and describing multiple events with grounded timestamps. Initial dense video captioning utilize a 2-stage pipeline [25, 26, 75, 79] by firstly performing localization and then describing events. Recent works [13, 18, 36, 51, 59, 64, 65, 75, 80, 83, 89] focus on training localization and captioning modules in an end-to-end manner to enhance inter-event associations. In contrast to these works, AD generation specifically aims to narrate a coherent story, maintain character-awareness, and complement the audio track without interfering with existing dialogue. + +AD generation. Early AD systems relied heavily on specialized authoring tools [9] and skilled human contributors. Platforms like Rescribe [54] and LiveDescribe [9] have facilitated faster and more accurate AD creation; however, these methods are costly and do not scale efficiently for large volumes of visual content. Recent efforts have developed audio segmentation and transcription systems [8, 10, 11] to create high-quality video datasets with temporally aligned ADs [60, 61, 68, 69], advancing automatic AD research. + +In general, current automatic AD generation systems can be categorized into two approaches: training-free and partial-fine-tuning. Training-free methods [39] generate ADs by leveraging proprietary models like GPT-4 [3] and GPT-4V [53]. MM-Narrator [86] enhances AD performance by multi-model in-context learning with memories. LLM-AD [14] and AutoAD-Zero [82] use prompts com + +prising visual frames with textual character names and colorful circles [66], enabling character-centric AD generation. However, training-free AD methods often suffer from high evaluation costs at scale and relatively poor performance due to domain-specific challenges and LLM hallucinations. Partial-fine-tuning methods [21-23, 40, 74], as well as our DistinctAD, only fine-tune a lightweight adapter [5, 34] between the pre-trained vision and text encoders. A representative example is the AutoAD series [21-23], which builds automatic AD systems and enriches them with character-aware prompts within different vision-language frameworks. However, previous studies tend to focus on constructing more accurate external character banks, whereas treating AD generation similarly to video captioning, overlooks AD's unique sequential structure of video clips. In contrast, our method emphasizes understanding the visual content within its temporal context, leading to more distinctive AD generation. + +Distinctive captioning in images aims to articulate unique details that can help distinguishing targets from others. An intuitive way to promoting distinctiveness is through contrastive learning [16, 43, 47, 73], where generated captions are encouraged to align more closely with target images rather than distractors. In [12, 76-78], group-based distinctive attention is introduced to capture distinctiveness by comparing sets of similar images and re-weighting specific caption words. A recent closely related field is difference captioning [32, 33, 84], which aims to describe differences between a single pair of images. VisDiff [19] scales difference captioning to sets containing thousands of images with natural language. Our work differs from these distinctive captioning works in that we are the first to explore distinctiveness across dense, consecutive clips within hours-long movies, thereby generating ADs with better narrative. + +# 3. Method + +# 3.1. Stage-I: CLIP-AD Adaptation + +AD and the visual content it describes exhibit a significant domain gap compared to typical large-scale web data. This gap often causes misalignment in current partial-fine-tuning techniques. Previous studies [21, 22, 74] try alleviating this problem by pre-training LLMs on text-only AD corpus, e.g. AudioVault [1]. However, misalignment persists at the initial stage of vision encoding, which is often neglected. + +Inspired by our findings that AD sentences encoded by the CLIP text encoder can be effectively recovered using the GPT-2 language model (see §1 and Appendix §A), we identify that the primary issue of misalignment is caused by the CLIP vision encoder, i.e., the discrepancy between visual embeddings and AD embeddings within the joint CLIP feature space. To address this, we propose adapting the CLIP vision encoder to the specific AD domain. However, due to + +![](images/b6264c77c575b2ff77e33f5074d1d69a9f7241fd6de9b2de0c26ced461284741.jpg) +Figure 2. Illustration of Stage-I: CLIP-AD Adaptation. This process involves adapting the CLIP vision encoder to specific movie-AD data through global-level video-AD matching (bottom right) and fine-grained frame-AD matching (top right). + +the unique multiple-instance learning setting for video-AD pairs (see §1), we consider both global matching and fine-grained frame-word matching in our adaptation method. + +Global video-AD matching. A straightforward strategy involves adopting classical CLIP-style fine-tuning with video-AD pairs in large batches [46]. Formally, let video clip $\mathbf{V}_i = [\mathbf{f}_i^1;\dots ;\mathbf{f}_i^n ]\in \mathbb{R}^{n\times C}$ be a collection of $n$ frame embeddings, and corresponding AD $\mathbf{T}_i = [\mathbf{w}_i^0;\mathbf{w}_i^1;\dots ;\mathbf{w}_i^m ]\in \mathbb{R}^{(m + 1)\times C}$ be a collection of $m$ word embeddings $(\mathbf{w}_i^j)$ and the [CLS] token (denoted as $\mathbf{w}_i^0$ ), where $C$ is the number of channels in the embedding space. We obtain the global video-level representation by averaging all frame embeddings in $\mathbf{V}_i$ using mean pooling: $\mathbf{v}_i = \frac{1}{n}\sum_{j = 1}^n\mathbf{f}_i^j$ . Following the standard CLIP, we use the [CLS] token as the global textual AD representation $\mathbf{t}_i = \mathbf{w}_i^0$ . The global video $\rightarrow$ AD matching is performed by maximizing the sum of the main diagonal of a $B\times B$ similarity matrix, using the contrastive loss: + +$$ +\mathcal {L} _ {v \rightarrow A D} = - \frac {1}{B} \sum_ {i = 1} ^ {B} \log \frac {\exp (s i m (\mathbf {v} _ {i} , \mathbf {t} _ {i}))}{\sum_ {j = 1} ^ {B} \exp (s i m (\mathbf {v} _ {i} , \mathbf {t} _ {j}))}, \tag {1} +$$ + +where $B$ is the batch size, and the similarity function $sim(\cdot)$ is the vector inner product. This process is illustrated in the bottom right of Fig. 2. Similarly, we drive the AD $\rightarrow$ video loss $\mathcal{L}_{AD\rightarrow v}$ by maximizing the sum of the secondary diagonal (i.e., swapping the $i$ and $j$ indices in (1)). The final global-level contrastive loss is then the sum of the losses in both directions $\mathcal{L}_q = \mathcal{L}_{v\to AD} + \mathcal{L}_{AD\to v}$ . + +Fine-grained frame-AD matching. Matching global video to AD sentence [CLS] (and vice versa) aids in joint feature space alignment. However, this alignment is insufficient for effective adaptation due to the specific multiple-instance + +![](images/7727ed981f6a1a0688c7cba9ca20ca96b17b8bae844d2371ae2ba69eee25f0a6.jpg) +Figure 3. Pipeline of Stage-II: Distinctive AD Narration. Stage-II processes $N$ consecutive video clips using the $\mathrm{CLIP}_{\mathrm{AD}}$ vision encoder from Stage-I. We generate contextual-distinctive ADs by two key innovations: i) a Contextual EMA module to learn compact and discriminative visual representations for improved prompting of LLMs; ii) an extra distinctive word loss for predicting AD-specific terms. + +setting of ADs, where only some words may correspond to a particular frame, but all words will have correspondence in aggregate. Thus, we propose a fine-grained matching loss at the frame-level to address this issue. + +Formally, given the frame embeddings $\mathbf{V}_i$ and the word embeddings $\mathbf{T}_i' = [\mathbf{w}_i^1; \dots; \mathbf{w}_i^m] \in \mathbb{R}^{m \times C}$ , we calculate the weights of all words attending to each frame via softmax attention, taking $\mathbf{V}_i$ as the query and $\mathbf{T}_i'$ as the key. By then multiplying these attention weights by the value $\mathbf{T}_i'$ , we obtain a frame-aware AD representation $\tilde{\mathbf{T}}_i \in \mathbb{R}^{n \times C}$ : + +$$ +\tilde {\mathbf {T}} _ {i} = \mathrm {S o f t m a x} (\mathbf {V} _ {i} \mathbf {T} _ {i} ^ {\prime T} / \tau) \mathbf {T} _ {i} ^ {\prime} = [ \tilde {\mathbf {t}} _ {i} ^ {1}; \dots ; \tilde {\mathbf {t}} _ {i} ^ {n} ], \qquad (2) +$$ + +where for each frame, the words embeddings that are most similar to the frame-level visual feature have been aggregated (via softmax attention). The temperature parameter $\tau$ controls the aggregation process, where smaller $\tau$ incorporates more textual information. + +The goal of the fine-grained matching is to pull a frame visual feature $\mathbf{f} \in \mathbf{V}_i$ closer to the frame-aware AD representations $\tilde{\mathbf{t}} \in \tilde{\mathbf{T}}_i$ in (2), corresponding to the positive set. To achieve this, we define the negative set $\tilde{\mathbf{T}}_{neg}$ as the frame-aware AD embeddings generated from other video clips (in the batch), and then use a Multi-Instance Loss [49], + +$$ +\mathcal {L} _ {f} = - \frac {1}{B} \sum_ {i = 1} ^ {B} \log \frac {\sum_ {\tilde {\mathbf {t}} \in \tilde {\mathbf {T}} _ {i}} \exp (s i m (\mathbf {f} , \tilde {\mathbf {t}}))}{\sum_ {\tilde {\mathbf {t}} _ {*} \in \{\tilde {\mathbf {T}} _ {i} \cup \tilde {\mathbf {T}} _ {n e g} \}} \exp (s i m (\mathbf {f} , \tilde {\mathbf {t}} _ {*}))}, \tag {3} +$$ + +where $\mathbf{f} \in \mathbf{V}_i$ is a sampled frame from $\mathbf{V}_i$ . This process is illustrated in the top right of Fig. 2. + +Summary for Stage-I. The final objective for Stage-I is to minimize the sum of global and fine-grained aligning losses, balanced by a trade-off coefficient, $\mathcal{L}_{\mathrm{I}} = \gamma \mathcal{L}_{g} + (1 -$ + +$\gamma)\mathcal{L}_f$ . Note that during this adaptation process, the CLIP-Text encoder model remains frozen, and only the CLIP-Vision encoder is fine-tuned. Our fine-grained frame-AD matching is entirely parameter-free, as only the vision encoder will be utilized in the subsequent stage. + +# 3.2. Stage-II: Distinctive AD Narration + +The motivation for generating distinctive ADs stems from the observation that LLM often produce repetitive descriptions for adjacent clips [56, 82, 86]. Despite improved character recognition, the visual representation itself is not discriminative among neighboring (contextual) clips, leading to uninteresting ADs. Our goal is to create contextual-distinctive ADs that highlight current differences. We hypothesize, as verified in Appendix §B, that sequential clips from a long video often share redundant scenes or characters, leading similar visual features in contexts. Thus, we propose Stage-II: distinctive AD narration. + +As shown in Fig. 3, we prepare $N$ consecutive video clips (to be AD-described) $\{\mathbf{x}_1,\mathbf{x}_2,\dots ,\mathbf{x}_N\}$ , each containing $T$ uniformly sampled frames $\{\mathcal{F}_1,\mathcal{F}_2,\dots ,\mathcal{F}_T\}$ . Following [21, 22, 74], we employ a learnable Perceiver adapter [5] to resample $T^{\prime}$ prompt vectors for the $T$ frame embeddings encoded by our Stage-I vision encoder, $\mathrm{CLIP}_{\mathrm{AD}}$ . This process is formulated as: + +$$ +\mathbf {h} _ {\mathbf {x} _ {i}} = \operatorname {P e r c e i v e r} \left(\left\{\mathbf {f} _ {1}, \mathbf {f} _ {2}, \dots , \mathbf {f} _ {T} \right\}\right) \in \mathbb {R} ^ {T ^ {\prime} \times C}, \tag {4} +$$ + +$$ +\mathbf {f} _ {i} = \operatorname {C L I P} _ {\mathrm {A D}} \left(\mathcal {F} _ {i}\right). \tag {5} +$$ + +We then introduce the Contextual EMA to capture compact, discriminative visual features for distinctive AD generation. Contextual EMA. Expectation-Maximization Attention (EMA) [35] integrates the attention mechanism [81] into the classical EM [17] algorithm, which comprises three steps to + +estimate a more compact set of bases: Responsibility Estimation (RE), Likelihood Maximization (LM), and Data Re-estimation (DR). Inspired by this, we propose Contextual EMA to perform EMA on frames from $N$ contextual clips, aiming to eliminate redundancy, learn compact representations, and explore distinctiveness. + +Let $\mathcal{H} = \{\mathbf{h}_{\mathbf{x}_i}\}_{i = 1}^N\in \mathbb{R}^{N\times T'\times C}$ represent $N$ clip vectors from the Perceiver, and $\mathcal{M} = \{\mu_k\}_{k = 1}^K\in \mathbb{R}^{K\times C}$ denote the randomly initialized base features, where $C,K$ indicate the number of channel and bases. The RE step estimates the hidden variable $\mathcal{Z} = \{z_{nk}\}_{n = 1,k = 1}^{N\times T',K}$ , where the responsibility $z_{nk}$ represents the probability of the $n$ -th frame belonging to the $k$ -th base: + +$$ +z _ {n k} = \frac {\exp \left(\mathbf {h} _ {n} \mu_ {k} ^ {T} / \tau\right)}{\sum_ {j = 1} ^ {K} \exp \left(\mathbf {h} _ {n} \mu_ {j} ^ {T} / \tau\right)}, \tag {6} +$$ + +where $\tau$ determines the shape (peakiness) of distribution $\mathcal{Z}$ . Then, the LM step updates the bases $\mathcal{M}$ by applying the weighted average on input $\mathcal{H}$ , formulating the $k$ -th base as: + +$$ +\mu_ {k} = \frac {\sum_ {n = 1} ^ {N \times T ^ {\prime}} z _ {n k} \mathbf {h} _ {n}}{\sum_ {n = 1} ^ {N \times T ^ {\prime}} z _ {n k}}. \tag {7} +$$ + +The RE (E-step) and LM (M-step) are iteratively performed $R$ times until convergence. Notably, since bases number $K$ is much smaller than the embedding number $N \times T'$ , we employ DR to reconstruct a compact version of $\mathcal{H}$ through: + +$$ +\widehat {\mathcal {H}} \approx \mathcal {Z M}. \tag {8} +$$ + +Here, $\widehat{\mathcal{H}}\in \mathbb{R}^{N\times T^{\prime}\times C}$ retains the same shape as $\mathcal{H}$ . We combine $\mathcal{H}$ and $\widehat{\mathcal{H}}$ element-wise with a hyperparameter $\alpha$ . + +To enhance representation distinctiveness, we introduce an additional branch using cross-attention between raw $\mathcal{H}$ (query) and bases $\mathcal{M}$ (key and value), formulated as: + +$$ +\widetilde {\mathcal {H}} = \operatorname {C r o s s A t t e n t i o n} (\mathcal {H}, \mathcal {M}), \tag {9} +$$ + +where $\widetilde{\mathcal{H}}\in \mathbb{R}^{N\times T^{\prime}\times C}$ . Linear layers projecting queries, keys, and values are omitted in Fig. 3 (see Appendix §C for details). Through (9), we process the distributions of $\mathcal{H}$ to attend on specific and informative bases, with improved linear separability (see Fig. 6). We combine $\mathcal{H},\widehat{\mathcal{H}},\widetilde{\mathcal{H}}$ elementwise around Contextual EMA to construct the final refined visual features. These features are then projected into the LLM embedding space using a single-layer projector: + +$$ +\mathcal {H} _ {\text {s u m}} = \operatorname {P r o j} (\mathcal {H} + \alpha \widehat {\mathcal {H}} + \beta \widetilde {\mathcal {H}}). \tag {10} +$$ + +Interleaved prompt as LLM's input. Following previous studies [22, 74, 82], we build our interleaved prompt enriched with character information, (see Fig. 3). To answer the "who is who" question when more than two characters are present, the corresponding actors' portrait images are projected as face tokens for reasoning. The $\langle \text{BOS} \rangle$ tag appended at the end indicates the beginning of AD generation. + +Distinctive words highlighting. Our goal is to query a frozen LLM for AD generation using a vision-conditioned prompt. The typical supervision employs the commonly used auto-regressive loss function: + +$$ +\mathcal {L} _ {\text {a u t o}} = - \sum_ {n} \log P _ {\theta} \left(w _ {n} \mid \text {p r o m p t}; w _ {< n}\right), \tag {11} +$$ + +where $w_{n}$ is the $n$ -th token from the target AD. However, $\mathcal{L}_{auto}$ does not emphasize the distinctiveness specific to the current AD, which is our focus. To address this, we propose a distinctive word set $w_{d}$ , created by filtering out duplicates, such as character names, prepositions, and pronouns, from the $N$ context ADs of the target AD. During training, we explicitly encourage the LLM to predict the distinctive words in $w_{d}$ by optimizing the distinctive loss $\mathcal{L}_{dist}$ : + +$$ +\mathcal {L} _ {\text {d i s t}} = - \sum_ {n = 1} ^ {N} \sum_ {i = 1} ^ {u} \log P _ {\theta} \left(w _ {n} = w _ {d} ^ {i} \mid \text {p r o m p t}, w _ {< n}\right), \tag {12} +$$ + +where $w_{d}^{i}$ denotes the $i$ -th distinctive word in $w_{d}$ and $u$ is the size of the set. The final complete loss function for Stage-II is: $\mathcal{L}_{\mathrm{II}} = \mathcal{L}_{\mathrm{auto}} + \mathcal{L}_{\mathrm{dist}}$ . + +# 4. Experiments + +# 4.1. Experiment Setup + +Datasets. We follow the AD generation benchmark established in AutoAD [21], conducting experiments on the denoised MAD-v2-Named [68] and evaluating on MAD-Eval-Named split. Specifically, MAD-v2-Named includes $\sim 330\mathrm{k}$ ADs from 488 movies for training and MAD-Eval has 6,520 ADs crawled from 10 movies for evaluation. We also evaluate on two recently introduced datasets. CMD-AD [23] (where "CMD" stands for Condensed Movie Dataset [7]) is a movie AD dataset that contains 101k ADs for more than 1432 movies, with 100 movies split for CMD-AD evaluation. TV-AD [82] is a recently proposed AD dataset based on TVQA [30], which contains $\sim 31\mathrm{k}$ ADs for training and $\sim 3\mathrm{k}$ ADs for evaluation. + +Evaluation Metrics. Classic captioning metrics including ROUGE-L [37], CIDEr [72] and SPICE [6] are reported to evaluate the quality of generated ADs versus the ground-truth. Besides, we also report Recall@k within $N$ Neighbours [22] ( $\mathbf{R}@\mathbf{k} / \mathbf{N}$ ), which calculates the average value of Recall@k for each AD with its $N$ temporally adjacent GT texts, where BertScore [88] is used for text similarity matching. The R@k/N metric is based on retrieving the most similar ground-truth AD among N neighbors, and thus highlights the distinctiveness of generated ADs directly. LLM-AD-eval [23] employs LLMs to assess the quality of generated ADs, assigning scores from 1 (lowest) to 5 (highest). We utilize the LLM prompt from the original study [82] and apply open-source models LLaMA2-7B-Chat [70] and LLaMA3-8B-Instruct [4] for this evaluation. + +
MethodPub.VLMLLMROUGE-LCIDErSPICER@5/16
Training-free
VLog [2]--GPT-47.51.32.142.3
MM-Vid [39]ArXiv'23GPT-4V-9.86.13.846.1
MM-Narrator [86]CVPR'24CLIP-L14GPT-413.413.95.249.0
LLM-AD [14]ArXiv'24GPT-4V-13.520.5--
AutoAD-Zero [82]ACCV'24VideoLLaMA2-7BLLaMA3-8B-22.4--
Partial-fine-tuning
ClipCap [50]ArXiv'21CLIP-B32GPT-28.54.41.1-
CapDec [52]ArXiv'22--8.26.71.4-
AutoAD-I [21]CVPR'23CLIP-B32GPT-211.914.34.442.1
AutoAD-II [22]ICCV'23CLIP-B32GPT-213.419.5-50.8
AutoAD-III [23]CVPR'24EVA-CLIPOPT-2.7B-22.8-52.0
AutoAD-III [23]CVPR'24EVA-CLIPLLaMA2-7B-24.0-52.8
MovieSeq [40]ECCV'24CLIP-B16LLaMA2-7B*15.524.47.051.6
DistinctAD (Ours)CLIP-B32GPT-215.424.56.749.8
DistinctAD (Ours)CLIPAD-B32GPT-216.425.57.451.7
DistinctAD (Ours)CLIPAD-B16LLaMA2-7B17.227.08.255.6
DistinctAD (Ours)CLIPAD-B16LLaMA3-8B17.627.38.356.0
+ +Table 1. Comparisons of AD performance on the MAD-Eval benchmark. * indicates fine-tuning LLaMA2-7B model with LoRA [24]. ${\mathrm{{CLIP}}}_{\mathrm{{AD}}}$ is our CLIP vision encoder adapted using our Stage-I strategy. + +
MethodCIDErR@1/5LLM-AD-eval
Video-BLIP2 [85]4.822.01.89
Video-LLaMA2 [87]5.223.61.91
AutoAD-II [22]13.526.12.08
AutoAD-III [23]21.730.02.85
AutoAD-Zero [82]17.7-2.83
DistinctAD (LLaMA2)22.332.92.89
DistinctAD (LLaMA3)22.733.02.88
AutoAD-III† [23]25.031.22.89
2.01
+ +Table 2. Comparisons on CMD-AD. The LLM-AD-eval scores are evaluated with LLaMA2-7B (left) and LLaMA3-8B (right). $\dagger$ indicates pre-training on 3.4M HowTo-AD dataset [23, 48]. + +
MethodCIDErR@1/5LLM-AD-eval
AutoAD-III [23]26.1-2.78 | 1.99
AutoAD-Zero [82]22.630.62.94 | 2.00
DistinctAD (LLaMA2)27.231.82.86 | 2.01
DistinctAD (LLaMA3)27.432.12.89 | 2.00
+ +Table 3. Comparisons on TV-AD. The LLM-AD-eval scores are evaluated using LLaMA2-7B (left) and LLaMA3-8B (right). + +Implementation Details. To facilitate CLIP-AD adaptation in Stage-I, we collect the original raw movies from MAD from platforms like Amazon Prime Video. See Appendix F for details. We fine-tune the CLIP Vision encoder for 5 epochs with a fixed learning rate 5e-5 using the Adam optimizer [28] in Stage-I, with a batch size of 512. In Stage-II, we use a batch of 8 sequences, each containing 16 consecutive video AD-pairs from a movie. For each video clip, 8 frames are uniformly sampled. We use the AdamW [44] optimizer to train our model for 10 epochs, with a cosine-decayed learning rate and linear warm-up. The learning rate is set to $10^{-4}$ for both GPT-2 and LLaMA models. For external character information, we directly use the inference results from AutoAD-Zero [82] as it gives current best face + +recognition performance. We use CLIP-B32 and GPT-2 on MAD-Eval-Named for ablations unless specified otherwise. + +# 4.2. Comparisons with previous methods + +We conduct comprehensive comparisons using the widely-adopted MAD-Eval benchmark [21] and two recently introduced AD datasets, CMD-AD [23] and TV-AD [82]. + +Comparisons on MAD-Eval are shown in Tab. 1. We primarily categorize previous studies into Training-free and Partial-fine-tuning approaches, as described in §2. Our method is a Partial-fine-tuning method. When using the same CLIP-B32 and GPT-2, our proposed DistinctAD achieves a CIDEr score of 24.5, surpassing previous AutoAD-I [21] (CIDEr 14.3) and AutoAD-II [22] (CIDEr 19.5). With our Stage-I adapted CLIP vision encoders (denoted as $\mathrm{CLIP}_{\mathrm{AD}}$ ), we observe stable improvements across all metrics, e.g. 25.5 vs. 24.5 on CIDEr and 51.7 vs. 49.8 on recall, validating the effectiveness of our Stage-I strategy. Notably, DistinctAD with $\mathrm{CLIP}_{\mathrm{AD}}$ -B16 and LLaMA3-8B [4] achieves state-of-the-arts with CIDEr 27.3 and Recall@5/16 56.0. Our outstanding performance on R@k/N demonstrates DistinctAD's ability to generate distinctive ADs, well matching the uniqueness of the clip's contents. + +Looking at the training-free methods, despite the capabilities of advanced proprietary VLMs, e.g. GPT-4V [53], and LLMs, e.g. GPT-4 [3], the performance of training-free methods remains inferior to those employing partial-finetuning. This discrepancy likely arises from the unique characteristics of AD and movie data, which exhibit a significant domain gap from common vision language training data. As such, these data types were not encountered during the pretraining of proprietary large-scale models. + +
SettingCIDErR@5/16
None6.734.0
Global Lg8.236.6
Fine-grained Lf7.735.2
γLg + (1 - γ)Lf8.636.9
+ +(a) Stage-I components. + +
Coefficient γCIDEr
0.18.0
0.38.5
0.58.6
0.77.7
+ +(b) Impact of coefficient $\gamma$ + +
PromptStage-ICIDErR@5/16
Contextual ADs [21]X12.6 (17.8)39.8 (43.1)
✓14.1 (19.0)39.9 (44.2)
Character [22]X22.045.6
✓23.146.2
+ +Table 4. Ablation studies for Stage-I. (a) Evaluation of global video-AD loss $\mathcal{L}_f$ and fine-grained frame-AD loss $\mathcal{L}_f$ on AD performance. (b) Analysis of the impact of the coefficient $\gamma$ . Both (a) and (b) are conducted with pure visual prompts. (c) Impact of Stage-I when combined with different prompts for the LLM decoder, including contextual ADs and character names. Performance in parentheses indicates results with ground-truth contextual ADs as prompts. + +![](images/5e32d7ebd4bcc1f151229deb6c3474cbecff188ccbaaeb862f97f24cb61baffe.jpg) +Figure 4. Ablation studies for hyperparameter in Stage-II, with final settings highlighted in orange. (a) Impact of $\alpha$ on the weight of compact representation $\widehat{\mathcal{H}}$ . (b) Influence of $\beta$ on cross-attended feature $\widetilde{\mathcal{H}}$ . (c) Impact of $K$ , which denotes the number of clusters in bases $\mathcal{M}$ . (d) Effect of sampling $N$ consecutive video clips. We switch to larger memory GPUs when $N$ exceeds 16. + +(c) Impact of Stage-I w/ different prompts. + +
Ex#αHβHLdistCIDErR@5/16
A0XXX23.1 (27.4)46.2
B1✓XX23.7 (29.3)46.6
B2X✓X23.4 (29.1)46.1
B3✓✓X23.3 (28.1)48.0
C0XX✓24.7 (29.4)49.5
C1✓X✓24.3 (29.4)50.7
C2X✓✓25.3 (30.4)51.5
C3✓✓✓25.5 (29.8)51.7
+ +Comparisons on CMD-AD and TV-AD. We further verify the generalizability of DistinctAD on the recently proposed CMD-AD and TV-AD benchmarks, with results presented in Tables 2 and 3. Both evaluations use adapted $\mathrm{CLIP}_{\mathrm{AD}}$ -B16 model. DistinctAD outperforms AutoAD-Zero, AutoAD-II and AutoAD-III in terms of CIDEr and R@1/5 on both benchmarks Meanwhile, DistinctAD exhibits a lower CIDEr compared to AutoAD-III† on CMD-AD, which we conjecture is primarily due to AutoAD-III† pre-training on a very large-scale 3.4M transformed HowTo-AD dataset [23, 48], which is currently publicly unavailable. Despite this, DistinctAD achieves superior R@1/5, underscoring its exceptional ability to generate distinctive and high-quality ADs. This is further corroborated by its leading performance on the LLM-AD-eval metric. + +# 4.3. Ablation studies + +Effect of CLIP-AD Adaptation (I). Tab. 4a demonstrates the benefit of our Stage-I strategy, i.e. adapting CLIP the vi + +Table 5. Ablation studies for components in Stage-II. The CIDEr column shows scores with AutoAD-Zero's character [82] as prompt by default. CIDEr in parentheses represent performance with ground-truth character names. + +
Consecutive N?CIDErR@5/16
✓25.551.7
×23.8↓1.752.5↑0.8
+ +Table 6. Impact of whether sampling $N$ consecutive clips or not. + +sion encoder to the movie-AD domain via global video-AD matching $\mathcal{L}_g$ and fine-grained frame-AD matching $\mathcal{L}_f$ . In Tab. 4b, the balancing coefficient $\gamma$ performs best at 0.5. In Tab. 4c, our Stage-I strategy consistently enhances performance when combined with different prompts in the LLM decoder, such as contextual ADs in AutoAD-I [21] or character names in AutoAD-II [22]. This indicates that our AD-adapted CLIP vision encoder can integrate seamlessly into previous methods, including those training-free models that utilize CLIP-based visual extractors. + +Effect of Distinctive AD narration (II). We evaluate the effectiveness of Stage-II components in Tab. 5, based on the default $\mathcal{H}$ (Perceiver's output) and full AD auto-regressive loss $\mathcal{L}_{\text {auto }}$ . The baseline (A0) outperforms AutoAD-II in CIDEr (23.1 vs. 19.5), primarily due to more accurate character prompts from AutoAD-Zero and the adapted CLIP vision encoder from Stage-I. A0 with AutoAD-II's characters achieves a CIDEr score of 20.6, close to AutoAD-II's performance. Incorporating reconstructed feature $\widehat{\mathcal{H}}$ brings stable improvements on both CIDEr and recall (B1 & C1), highlighting the importance of compact representations in understanding visual semantics. Cross-attended feature $\widetilde{\mathcal{H}}$ works better together with distinctive word prediction loss $\mathcal{L}_{\text {dist }}$ (C2 vs. B2, C3 vs. B3). We conjecture this is because $\mathcal{L}_{\text {dist }}$ provides more definite supervision on re-weighting distinctive words, which guides $\widehat{\mathcal{H}}$ to attend on concept-related bases. Overall, applying the full Stage-II pipeline brings significant and robust performance (C3). + +Effect of Hyper-parameters. Fig. 4 summarizes the ablation studies on 4 hyper-parameters that potentially influence the results in Stage-II. Coefficient weights $\alpha$ and $\beta$ yield optimal results when set to 3 and 1, respectively. This suggests the need to refine our final representations to be more compact for generating ADs. Fig. 4(c) shows setting bases number $K$ to 32 yields best. A smaller $K$ , e.g. 2, can still achieve notable CIDEr, as the Contextual EMA module does not significantly alter the final output $\mathcal{H}_{sum}$ . However, + +![](images/2ccc1aef0fa6873a1c16a4b22dfe62f6ed3a0a36aed0fd712fdb6583fc388f04.jpg) + +# 1h12m53.6s + +![](images/2d5c202bd5e428ef58c4655aed83eb2dcc2f4c3af0da9788eb29006af5f7bc2e.jpg) + +GT: Holding the room's phone to his ear, the man on the bed regards him sympathetically. + +AutoAD-Zero: He looks at something. + +DistinctAD: He holds the phone to his ear. + +![](images/dc60bae16010cbbd6600786418ec0d51289a1534bb7887ee6fc3791de2b47696.jpg) + +# 1h12m59.1s + +![](images/7327c6f13a5c90c45f2bc5d23e61020a8d28786593f750d639f35b1183c4c6b1.jpg) + +GT: Stephen sees the pill bottles on a bedside table. + +AutoAD-Zero: Stephen looks at something. + +DistinctAD: In Stephen's bedroom, a pill bottle sits on his nightstand. + +![](images/e1b7ef6e20b0a2526f1d30ade4570b895b78d28f8a44e85883894cbc63216b9e.jpg) + +# 1h13m14.5s + +![](images/ebd129b9530fb77b259d18208c403a32fb2c50a1c482de679bb742a4aa972e90.jpg) + +GT: The man on the bed sits in silhouette against the room's soft lamp light. + +AutoAD-Zero: He turns to leave, looking at him. + +DistinctAD: The bed is empty. + +![](images/37954c648a97d7f8189ac87b46a889a7b7a69cb60e18444612d6b9015bb9c08b.jpg) + +# 1h13m19.4s + +![](images/81ddb833fd820f8365e260173580a31af48f41d367e4ed7f01440bd29bb08c2f.jpg) + +GT: Later, Stephen steps back out into the rainy night. + +AutoAD-Zero: Stephen looks around and at another character. + +DistinctAD: As Stephen walks away, he glances over his shoulder. + +![](images/aa5955ececbf0391df747327102a561453c74d813506b3ca97e40903a178655f.jpg) + +# 1h13m23.1s + +![](images/23068bb9811e451a9448d946e9582d6e135b846bdb82bfa1466d538458be3d6a.jpg) + +GT: Moving with a stiff gait, Stephen marches across the parking lot to his car. + +AutoAD-Zero: He looks out at the city. + +DistinctAD: He looks up at the sky, then turns and heads back to the car. + +![](images/12c637b2b01d4f5a6b888cc9353072a41356abd721a8166d5ed3be3d7650980c.jpg) + +# h13m24.5s + +![](images/2d337008c1aa526bc9bc5fe5aece4ce62e57c4e8c4940a65503b7c1552eec619.jpg) + +GT: As Stephen gets in, rain patters heavily against his windshield. + +AutoAD-Zero: Stephen looks at the women. + +DistinctAD: In the car, Stephen stares out the window. + +: The Ides of March (2011) + +unsuitable values of $K$ , e.g. 8 or 64, can negatively impact performance. Fig. 4(d) reveals that increasing the number of consecutive clips $N$ (with $K$ set to 32) enhances the CIDEr score, though this effect saturates when $N$ exceeds 16. This demonstrates that more bases should be created to effectively summarize components with additional clips. + +Do the $N$ clips to be consecutive? Tab. 6 presents the results of sampling non-consecutive $N$ clips during training. When using non-continuous clips, we observe a decrease in the CIDEr metric by 1.7 (25.5 vs. 23.8) because the Contextual EMA module struggles with unrelated contexts. However, the R@5/16 improves by 0.8, which indicates enhancement of the distinctiveness (uniqueness) of the generated AD when using more diverse visual contents. Visualizations. To better understand what Contextual EMA learns, we show the t-SNE [71] visualizations of $\mathcal{H}$ , $\widehat{\mathcal{H}}$ and $\widetilde{\mathcal{H}}$ (from §3.2) in Fig. 6, using the same perplexity value across all visualizations. With Contextual EMA, $\widehat{\mathcal{H}}$ exhibits more compact features compared to raw $\mathcal{H}$ , Fig. 6(b). Interestingly in Fig. 6(c), cross-attention between $\mathcal{H}$ and bases $\mathcal{M}$ produces strip-like feature distributions pointing to specific base centers, enhancing contextual distinctiveness with improved linear separability and interpretability. + +# 4.4. Qualitative results + +Fig. 5 presents qualitative examples of our model. We compare the predictions of DistinctAD (using LLaMA3-8B) with ground-truth captions (GT) and publicly available AutoAD-Zero [82] outputs. Note that clips are sampled consecutively in time. Previous studies often struggle with similar contextual clips, such as those featuring closely-related scenes and characters, by repeating correct yet in + +![](images/29d662c4b957db5ba2e1a973477e675137f4348e9fd76de5e80fde7457094325.jpg) +Figure 5. Qualitative results. We present ground-truth (GT) ADs, publicly released AutoAD-Zero outputs, and our DistinctAD predictions for several temporally consecutive movie clips. Movie frames are taken from The Ides of March (2011) [15]. Zoom in for details. +(a) $\mathcal{H}$ +Figure 6. Visualizations of Contextual EMA. (a) A set of randomly generated 3D data $\mathcal{H}$ , sampled from $N$ types of samples. (b) Compact features $\widehat{\mathcal{H}}$ obtained via Data Re-estimation (DR). (c) Cross-attention outputs $\widetilde{\mathcal{H}}$ between $\mathcal{H}$ and bases $\mathcal{M}$ . + +![](images/bc6c45c84779a97c19f85519f04ab16ddccf6912105013878a04de987a39702d.jpg) +(b) $\widehat{\mathcal{H}}$ + +![](images/00f00d71e4482d9a9a0525243439f9f672dfdaabb1ce2a4def9aaa4e07519078.jpg) +(c) $\widetilde{\mathcal{H}}$ + +significant action words, e.g. "look". In contrast, our DistinctAD effectively generates more engaging ADs by identifying distinctive objects in adjacent clips, e.g. "phone", "pill bottle", and "car", along with corresponding more specific behaviors. More examples can be found in Appendix §E. + +# 5. Conclusion + +This paper proposes DistinctAD, a two-stage framework for generating distinctive audio descriptions for better narrative. By addressing the domain gap between movie-AD data with a CLIP-AD adaptation strategy, and introducing a Contextual EMA module and a distinctive word prediction loss, our approach significantly improves the quality of AD generation. The effectiveness of DistinctAD is demonstrated through comprehensive evaluations on multiple benchmark datasets and ablations studies. Despite these promising results, DistinctAD is still limited by requiring numbers of parameters and the quality of the generated ADs still falls short of human annotations. Overall, automatic AD generation still remains a challenging task, and there is considerable scope for future advancements in this field. + +# Acknowledgement + +This work was supported by a Strategic Research Grant from City University of Hong Kong (Project No. 7005840). + +# References + +[1] https://audiovault.net, 2019.3 +[2] Vlog: Video as a long document. https://github.com/showlab/VLog, 2023. GitHub repository. 6 +[3] Josh Achiam, Steven Adler, Sandhini Agarwal, Lama Ahmad, Ilge Akkaya, Florencia Leoni Aleman, Diogo Almeida, Janko Altenschmidt, Sam Altman, Shyamal Anadkat, et al. Gpt-4 technical report. arXiv preprint arXiv:2303.08774, 2023. 1, 2, 6 +[4] AI@Meta. Llama 3 model card. 2024. 5, 6, 2 +[5] Jean-Baptiste Alayrac, Jeff Donahue, Pauline Luc, Antoine Miech, Iain Barr, Yana Hasson, Karel Lenc, Arthur Mensch, Katherine Millican, Malcolm Reynolds, et al. Flamingo: a visual language model for few-shot learning. NeurIPS, 35: 23716-23736, 2022. 1, 3, 4 +[6] Peter Anderson, Basura Fernando, Mark Johnson, and Stephen Gould. Spice: Semantic propositional image caption evaluation. In ECCV, pages 382-398. Springer, 2016. 5 +[7] Max Bain, Arsha Nagrani, Andrew Brown, and Andrew Zisserman. Condensed movies: Story based retrieval with contextual embeddings. In ACCV, 2020. 5 +[8] Max Bain, Jaesung Huh, Tengda Han, and Andrew Zisserman. Whisperx: Time-accurate speech transcription of long-form audio. arXiv preprint arXiv:2303.00747, 2023. 2 +[9] Carmen J Branje and Deborah I Fels. Livedescribe: can amateur descriptors create high-quality audio description? Journal of Visual Impairment & Blindness, 106(3):154-165, 2012. 1, 2 +[10] Hervé Bredin and Antoine Laurent. End-to-end speaker segmentation for overlap-aware resegmentation. In Interspeech, 2021. 2 +[11] Hervé Bredin, Ruiqing Yin, Juan Manuel Coria, Gregory Gelly, Pavel Korshunov, Marvin Lavechin, Diego Fustes, Hadrien Titeux, Wassim Bouaziz, and Marie-Philippe Gill. Pyannote. audio: neural building blocks for speaker diarization. In ICASSP, pages 7124–7128. IEEE, 2020. 2 +[12] Fuhai Chen, Rongrong Ji, Xiaoshuai Sun, Yongjian Wu, and Jinsong Su. Groupcap: Group-based image captioning with structured relevance and diversity constraints. In CVPR, pages 1345-1353, 2018. 3 +[13] Shaoxiang Chen and Yu-Gang Jiang. Towards bridging event captioner and sentence localizer for weakly supervised dense event captioning. In CVPR, pages 8425-8435, 2021. 2 +[14] Peng Chu, Jiang Wang, and Andre Abrantes. Llm-ad: Large language model based audio description system. arXiv preprint arXiv:2405.00983, 2024. 1, 2, 6 +[15] George Clooney. The idees of march. *Columbia Pictures*, 2011. 8 +[16] Bo Dai and Dahua Lin. Contrastive learning for image captioning. NeurIPS, 30, 2017. 3 + +[17] Arthur P Dempster, Nan M Laird, and Donald B Rubin. Maximum likelihood from incomplete data via the em algorithm. Journal of the royal statistical society: series B (methodological), 39(1):1-22, 1977. 2, 4 +[18] Chaorui Deng, Shizhe Chen, Da Chen, Yuan He, and Qi Wu. Sketch, ground, and refine: Top-down dense video captioning. In CVPR, pages 234-243, 2021. 2 +[19] Lisa Dunlap, Yuhui Zhang, Xiaohan Wang, Ruiqi Zhong, Trevor Darrell, Jacob Steinhardt, Joseph E Gonzalez, and Serena Yeung-Levy. Describing differences in image sets with natural language. In CVPR, pages 24199-24208, 2024. 3 +[20] Louise Fryer. An introduction to audio description: A practical guide, 2016. 1 +[21] Tengda Han, Max Bain, Arsha Nagrani, Gül Varol, Weidi Xie, and Andrew Zisserman. Autoad: Movie description in context. In CVPR, pages 18930-18940, 2023. 1, 2, 3, 4, 5, 6, 7 +[22] Tengda Han, Max Bain, Arsha Nagrani, Gul Varol, Weidi Xie, and Andrew Zisserman. Autoad ii: The sequel-who, when, and what in movie audio description. In ICCV, pages 13645–13655, 2023. 1, 3, 4, 5, 6, 7 +[23] Tengda Han, Max Bain, Arsha Nagrani, Gül Varol, Weidi Xie, and Andrew Zisserman. Autoad iii: The prequel-back to the pixels. In CVPR, pages 18164-18174, 2024. 1, 2, 3, 5, 6, 7 +[24] Edward J Hu, Yelong Shen, Phillip Wallis, Zeyuan Allen-Zhu, Yuanzhi Li, Shean Wang, Lu Wang, and Weizhu Chen. Lora: Low-rank adaptation of large language models. arXiv preprint arXiv:2106.09685, 2021. 6 +[25] Vladimir Iashin and Esa Rahtu. A better use of audio-visual cues: Dense video captioning with bi-modal transformer. In BMVC, 2020. 2 +[26] Vladimir Iashin and Esa Rahtu. Multi-modal dense video captioning. In CVPRW, pages 958-959, 2020. 2 +[27] Peng Jin, Jina Huang, Fenglin Liu, Xian Wu, Shen Ge, Guoli Song, David Clifton, and Jie Chen. Expectation-maximization contrastive learning for compact video-and-language representations. NeurIPS, 35:30291-30306, 2022. 2 +[28] Diederik P Kingma. Adam: A method for stochastic optimization. arXiv preprint arXiv:1412.6980, 2014. 6 +[29] Ranjay Krishna, Kenji Hata, Frederic Ren, Li Fei-Fei, and Juan Carlos Niebles. Dense-captioning events in videos. In ICCV, pages 706–715, 2017. 2 +[30] Jie Lei, Licheng Yu, Mohit Bansal, and Tamara L Berg. Tvqa: Localized, compositional video question answering. In EMNLP, 2018. 5 +[31] Elisa Lewis. Deep dive: How audio description benefits everyone, 2021. Accessed on, pages 11-13, 2023. 1 +[32] Bo Li, Yuanhan Zhang, Liangyu Chen, Jinghao Wang, Fanyi Pu, Jingkang Yang, Chunyuan Li, and Ziwei Liu. Mimicit: Multi-modal in-context instruction tuning. arXiv preprint arXiv:2306.05425, 2023. 3 +[33] Bo Li, Yuanhan Zhang, Liangyu Chen, Jinghao Wang, Jingkang Yang, and Ziwei Liu. Otter: a multi-modal model with in-context instruction tuning. corr abs/2305.03726 (2023), 2023. 3 + +[34] Junnan Li, Dongxu Li, Silvio Savarese, and Steven Hoi. Blip-2: Bootstrapping language-image pre-training with frozen image encoders and large language models. arXiv preprint arXiv:2301.12597, 2023. 1, 3 +[35] Xia Li, Zhisheng Zhong, Jianlong Wu, Yibo Yang, Zhouchen Lin, and Hong Liu. Expectation-maximization attention networks for semantic segmentation. In ICCV, pages 9167-9176, 2019. 2, 4 +[36] Yehao Li, Ting Yao, Yingwei Pan, Hongyang Chao, and Tao Mei. Jointly localizing and describing events for dense video captioning. In CVPR, pages 7492-7500, 2018. 2 +[37] Chin-Yew Lin. Rouge: A package for automatic evaluation of summaries. In Text summarization branches out, pages 74-81, 2004. 5 +[38] Kevin Lin, Linjie Li, Chung-Ching Lin, Faisal Ahmed, Zhe Gan, Zicheng Liu, Yumao Lu, and Lijuan Wang. Swinbert: End-to-end transformers with sparse attention for video captioning. In CVPR, pages 17949-17958, 2022. 2 +[39] Kevin Lin, Faisal Ahmed, Linjie Li, Chung-Ching Lin, Ehsan Azarnasab, Zhengyuan Yang, Jianfeng Wang, Lin Liang, Zicheng Liu, Yumao Lu, et al. Mm-vid: Advancing video understanding with gpt-4v (ision). arXiv preprint arXiv:2310.19773, 2023. 1, 2, 6 +[40] Kevin Qinghong Lin, Pengchuan Zhang, Difei Gao, Xide Xia, Joya Chen, Ziteng Gao, Jinheng Xie, Xuhong Xiao, and Mike Zheng Shou. Learning video context as interleaved multimodal sequences. arXiv preprint arXiv:2407.21757, 2024. 1, 3, 6 +[41] Zhihui Lin, Tianyu Yang, Maomao Li, Ziyu Wang, Chun Yuan, Wenhao Jiang, and Wei Liu. Swem: Towards real-time video object segmentation with sequential weighted expectation-maximization. In CVPR, pages 1362-1372, 2022. 2 +[42] Haotian Liu, Chunyuan Li, Qingyang Wu, and Yong Jae Lee. Visual instruction tuning. NeurIPS, 36, 2024. 1 +[43] Xihui Liu, Hongsheng Li, Jing Shao, Dapeng Chen, and Xiaogang Wang. Show, tell and discriminate: Image captioning by self-retrieval with partially labeled data. In ECCV, pages 338-354, 2018. 3 +[44] I Loshchilov. Decoupled weight decay regularization. arXiv preprint arXiv:1711.05101, 2017. 6 +[45] Huaishao Luo, Lei Ji, Botian Shi, Haoyang Huang, Nan Duan, Tianrui Li, Jason Li, Taroon Bharti, and Ming Zhou. Univl: A unified video and language pre-training model for multimodal understanding and generation. arXiv preprint arXiv:2002.06353, 2020. 2 +[46] Huaishao Luo, Lei Ji, Ming Zhong, Yang Chen, Wen Lei, Nan Duan, and Tianrui Li. Clip4clip: An empirical study of clip for end to end video clip retrieval and captioning. Neurocomputing, 508:293-304, 2022. 3 +[47] Ruotian Luo, Brian Price, Scott Cohen, and Gregory Shakhnarovich. Discriminability objective for training descriptive captions. In CVPR, pages 6964-6974, 2018. 3 +[48] Antoine Miech, Dimitri Zhukov, Jean-Baptiste Alayrac, Makarand Tapaswi, Ivan Laptev, and Josef Sivic. Howto100m: Learning a text-video embedding by watching hundred million narrated video clips. In ICCV, pages 2630-2640, 2019. 6, 7 + +[49] Antoine Miech, Jean-Baptiste Alayrac, Lucas Smaira, Ivan Laptev, Josef Sivic, and Andrew Zisserman. End-to-end learning of visual representations from uncurated instructional videos. In CVPR, pages 9879-9889, 2020. 4 +[50] Ron Mokady, Amir Hertz, and Amit H Bermano. Clipcap: Clip prefix for image captioning. arXiv preprint arXiv:2111.09734, 2021. 6 +[51] Jonghwan Mun, Linjie Yang, Zhou Ren, Ning Xu, and Bohyung Han. Streamlined dense video captioning. In CVPR, pages 6588-6597, 2019. 2 +[52] David Nukrai, Ron Mokady, and Amir Globerson. Text-only training for image captioning using noise-injected clip. arXiv preprint arXiv:2211.00575, 2022. 6 +[53] OpenAI. Gpt-4v(ision) system card. 2023. 1, 2, 6 +[54] Amy Pavel, Gabriel Reyes, and Jeffrey P Bigham. Describe: Authoring and automatically editing audio descriptions. In Proceedings of the 33rd Annual ACM Symposium on User Interface Software and Technology, pages 747-759, 2020. 1, 2 +[55] Elisa Perego. Gains and losses of watching audio described films for sighted viewers. *Target*, 28(3):424-444, 2016. 1 +[56] Haran Raajesh, Naveen Reddy Desanur, Zeeshan Khan, and Makarand Tapaswi. Micap: A unified model for identity-aware movie descriptions. In CVPR, pages 14011-14021, 2024. 1, 4 +[57] Alec Radford, Jeffrey Wu, Rewon Child, David Luan, Dario Amodei, Ilya Sutskever, et al. Language models are unsupervised multitask learners. OpenAI blog, 1(8):9, 2019. 2 +[58] Alec Radford, Jong Wook Kim, Chris Hallacy, Aditya Ramesh, Gabriel Goh, Sandhini Agarwal, Girish Sastry, Amanda Askell, Pamela Mishkin, Jack Clark, et al. Learning transferable visual models from natural language supervision. In ICML, pages 8748-8763. PMLR, 2021. 2 +[59] Tanzila Rahman, Bicheng Xu, and Leonid Sigal. Watch, listen and tell: Multi-modal weakly supervised dense event captioning. In ICCV, pages 8908-8917, 2019. 2 +[60] Anna Rohrbach, Marcus Rohrbach, Niket Tandon, and Bernt Schiele. A dataset for movie description. In CVPR, pages 3202-3212, 2015. 2 +[61] Anna Rohrbach, Atousa Torabi, Marcus Rohrbach, Niket Tandon, Christopher Pal, Hugo Larochelle, Aaron Courville, and Bernt Schiele. Movie description. *IJCV*, 123:94–120, 2017. 2 +[62] Paul Hongsuck Seo, Arsha Nagrani, Anurag Arnab, and Cordelia Schmid. End-to-end generative pretraining for multimodal video captioning. In CVPR, pages 17959-17968, 2022. 2 +[63] Piyush Sharma, Nan Ding, Sebastian Goodman, and Radu Soricut. Conceptual captions: A cleaned, hypernymed, image alt-text dataset for automatic image captioning. In ACL, pages 2556-2565, 2018. 2 +[64] Zhiqiang Shen, Jianguo Li, Zhou Su, Minjun Li, Yurong Chen, Yu-Gang Jiang, and Xiangyang Xue. Weakly supervised dense video captioning. In CVPR, pages 1916-1924, 2017. 2 +[65] Botian Shi, Lei Ji, Yaobo Liang, Nan Duan, Peng Chen, Zhendong Niu, and Ming Zhou. Dense procedure captioning + +in narrated instructional videos. In ACL, pages 6382-6391, 2019. 2 +[66] Aleksandar Shtedritski, Christian Rupprecht, and Andrea Vedaldi. What does clip know about a red circle? visual prompt engineering for vlms. In ICCV, pages 11987-11997, 2023. 3 +[67] Joel Snyder. Audio description: The visual made verbal. In International congress series, pages 935-939. Elsevier, 2005. 1 +[68] Mattia Soldan, Alejandro Pardo, Juan León Alcazar, Fabian Caba, Chen Zhao, Silvio Giancola, and Bernard Ghanem. Mad: A scalable dataset for language grounding in videos from movie audio descriptions. In CVPR, pages 5026-5035, 2022. 2, 5, 1, 4 +[69] Atousa Torabi, Christopher Pal, Hugo Larochelle, and Aaron Courville. Using descriptive video services to create a large data source for video annotation research. arXiv preprint arXiv:1503.01070, 2015. 2 +[70] Hugo Touvron, Louis Martin, Kevin Stone, Peter Albert, Amjad Almahairi, Yasmine Babaei, Nikolay Bashlykov, Soumya Batra, Prajjwal Bhargava, Shruti Bhosale, et al. Llama 2: Open foundation and fine-tuned chat models. arXiv preprint arXiv:2307.09288, 2023. 5 +[71] Laurens Van der Maaten and Geoffrey Hinton. Visualizing data using t-sne. JMLR, 9(11), 2008. 8 +[72] Ramakrishna Vedantam, C Lawrence Zitnick, and Devi Parikh. Cider: Consensus-based image description evaluation. In CVPR, pages 4566-4575, 2015. 5 +[73] Gilad Vered, Gal Oren, Yuval Atzmon, and Gal Chechik. Joint optimization for cooperative image captioning. In ICCV, pages 8898-8907, 2019. 3 +[74] Hanlin Wang, Zhan Tong, Kecheng Zheng, Yujun Shen, and Limin Wang. Contextual ad narration with interleaved multimodal sequence. arXiv preprint arXiv:2403.12922, 2024. 1, 2, 3, 4, 5 +[75] Jingwen Wang, Wenhao Jiang, Lin Ma, Wei Liu, and Yong Xu. Bidirectional attentive fusion with context gating for dense video captioning. In CVPR, pages 7190-7198, 2018. 2 +[76] Jiuniu Wang, Wenjia Xu, Qingzhong Wang, and Antoni B Chan. Compare and reweight: Distinctive image captioning using similar images sets. In ECCV, pages 370-386. Springer, 2020. 3 +[77] Jiuniu Wang, Wenjia Xu, Qingzhong Wang, and Antoni B Chan. Group-based distinctive image captioning with memory attention. In ACMMM, pages 5020-5028, 2021. +[78] Jiuniu Wang, Wenjia Xu, Qingzhong Wang, and Antoni B Chan. On distinctive image captioning via comparing and reweighting. TPAMI, 45(2):2088-2103, 2022. 3 +[79] Teng Wang, Huicheng Zheng, Mingjing Yu, Qian Tian, and Haifeng Hu. Event-centric hierarchical representation for dense video captioning. TCVT, 31(5):1890-1900, 2020. 2 +[80] Teng Wang, Ruimao Zhang, Zhichao Lu, Feng Zheng, Ran Cheng, and Ping Luo. End-to-end dense video captioning with parallel decoding. In ICCV, pages 6847-6857, 2021. 2 +[81] Xiaolong Wang, Ross Girshick, Abhinav Gupta, and Kaiming He. Non-local neural networks. In CVPR, pages 7794-7803, 2018. 4 + +[82] Junyu Xie, Tengda Han, Max Bain, Arsha Nagrani, Güll Varol, Weidi Xie, and Andrew Zisserman. Autoad-zero: A training-free framework for zero-shot audio description. arXiv preprint arXiv:2407.15850, 2024. 1, 2, 4, 5, 6, 7, 8 +[83] Antoine Yang, Arsha Nagrani, Paul Hongsuck Seo, Antoine Miech, Jordi Pont-Tuset, Ivan Laptev, Josef Sivic, and Cordelia Schmid. Vid2seq: Large-scale pretraining of a visual language model for dense video captioning. In CVPR, pages 10714-10726, 2023. 2 +[84] Linli Yao, Weiying Wang, and Qin Jin. Image difference captioning with pre-training and contrastive learning. In AAAI, pages 3108-3116, 2022. 3 +[85] Keunwoo Peter Yu. Videoblip, 2023. 6 +[86] Chaoyi Zhang, Kevin Lin, Zhengyuan Yang, Jianfeng Wang, Linjie Li, Chung-Ching Lin, Zicheng Liu, and Lijuan Wang. Mm-narrator: Narrating long-form videos with multimodal in-context learning. In CVPR, pages 13647-13657, 2024. 1, 2, 4, 6 +[87] Hang Zhang, Xin Li, and Lidong Bing. Video-llama: An instruction-tuned audio-visual language model for video understanding. arXiv preprint arXiv:2306.02858, 2023. 6 +[88] Tianyi Zhang, Varsha Kishore, Felix Wu, Kilian Q Weinberger, and Yoav Artzi. Bertscore: Evaluating text generation with bert. arXiv preprint arXiv:1904.09675, 2019. 5 +[89] Luowei Zhou, Yingbo Zhou, Jason J Corso, Richard Socher, and Caiming Xiong. End-to-end dense video captioning with masked transformer. In CVPR, pages 8739-8748, 2018. 2 \ No newline at end of file diff --git a/2025/DistinctAD_ Distinctive Audio Description Generation in Contexts/images.zip b/2025/DistinctAD_ Distinctive Audio Description Generation in Contexts/images.zip new file mode 100644 index 0000000000000000000000000000000000000000..06572f53975a408689fe6e0b4dd3d61cf3f2910e --- /dev/null +++ b/2025/DistinctAD_ Distinctive Audio Description Generation in Contexts/images.zip @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:07d8128bdddace9aaec11c48b38124d0247a736a87e88f78db1ef176e74db0f7 +size 560981 diff --git a/2025/DistinctAD_ Distinctive Audio Description Generation in Contexts/layout.json b/2025/DistinctAD_ Distinctive Audio Description Generation in Contexts/layout.json new file mode 100644 index 0000000000000000000000000000000000000000..c513b6c4c299f08ae159210f7dfe115a4485896f --- /dev/null +++ b/2025/DistinctAD_ Distinctive Audio Description Generation in Contexts/layout.json @@ -0,0 +1,12389 @@ +{ + "pdf_info": [ + { + "para_blocks": [ + { + "bbox": [ + 102, + 103, + 507, + 120 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 102, + 103, + 507, + 120 + ], + "spans": [ + { + "bbox": [ + 102, + 103, + 507, + 120 + ], + "type": "text", + "content": "DistinctAD: Distinctive Audio Description Generation in Contexts" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 119, + 142, + 491, + 185 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 119, + 142, + 491, + 185 + ], + "spans": [ + { + "bbox": [ + 119, + 142, + 491, + 185 + ], + "type": "text", + "content": "Bo Fang" + }, + { + "bbox": [ + 119, + 142, + 491, + 185 + ], + "type": "inline_equation", + "content": "^{1}" + }, + { + "bbox": [ + 119, + 142, + 491, + 185 + ], + "type": "text", + "content": ", Wenhao Wu" + }, + { + "bbox": [ + 119, + 142, + 491, + 185 + ], + "type": "inline_equation", + "content": "^{2,3}" + }, + { + "bbox": [ + 119, + 142, + 491, + 185 + ], + "type": "text", + "content": ", Qiangqiang Wu" + }, + { + "bbox": [ + 119, + 142, + 491, + 185 + ], + "type": "inline_equation", + "content": "^{1}" + }, + { + "bbox": [ + 119, + 142, + 491, + 185 + ], + "type": "text", + "content": ", Yuxin Song" + }, + { + "bbox": [ + 119, + 142, + 491, + 185 + ], + "type": "inline_equation", + "content": "^{2}" + }, + { + "bbox": [ + 119, + 142, + 491, + 185 + ], + "type": "text", + "content": ", Antoni B. Chan" + }, + { + "bbox": [ + 119, + 142, + 491, + 185 + ], + "type": "inline_equation", + "content": "^{1}" + }, + { + "bbox": [ + 119, + 142, + 491, + 185 + ], + "type": "inline_equation", + "content": "^{1}" + }, + { + "bbox": [ + 119, + 142, + 491, + 185 + ], + "type": "text", + "content": " Department of Computer Science, City University of Hong Kong \n" + }, + { + "bbox": [ + 119, + 142, + 491, + 185 + ], + "type": "inline_equation", + "content": "^{2}" + }, + { + "bbox": [ + 119, + 142, + 491, + 185 + ], + "type": "text", + "content": " Baidu Inc. \n" + }, + { + "bbox": [ + 119, + 142, + 491, + 185 + ], + "type": "inline_equation", + "content": "^{3}" + }, + { + "bbox": [ + 119, + 142, + 491, + 185 + ], + "type": "text", + "content": " The University of Sydney" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 132, + 187, + 478, + 213 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 187, + 478, + 213 + ], + "spans": [ + { + "bbox": [ + 132, + 187, + 478, + 213 + ], + "type": "text", + "content": "{bofang6-c, qiangqwu2-c}@my.cityu.edu.hk, wenhao.wu@sydney.edu.au songyuxin02@baidu.com, abchan@cityu.edu.hk" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 151, + 241, + 200, + 253 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 151, + 241, + 200, + 253 + ], + "spans": [ + { + "bbox": [ + 151, + 241, + 200, + 253 + ], + "type": "text", + "content": "Abstract" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 55, + 266, + 296, + 578 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 266, + 296, + 578 + ], + "spans": [ + { + "bbox": [ + 55, + 266, + 296, + 578 + ], + "type": "text", + "content": "Audio Descriptions (ADs) aim to provide a narration of a movie in text form, describing non-dialogue-related narratives, such as characters, actions, or scene establishment. Automatic generation of ADs remains challenging due to: i) the domain gap between movie-AD data and existing data used to train vision-language models, and ii) the issue of contextual redundancy arising from highly similar neighboring visual clips in a long movie. In this work, we propose DistinctAD, a novel two-stage framework for generating ADs that emphasize distinctiveness to produce better narratives. To address the domain gap, we introduce a CLIP-AD adaptation strategy that does not require additional AD corpora, enabling more effective alignment between movie and AD modalities at both global and fine-grained levels. In Stage-II, DistinctAD incorporates two key innovations: (i) a Contextual Expectation-Maximization Attention (EMA) module that reduces redundancy by extracting common bases from consecutive video clips, and (ii) an explicit distinctive word prediction loss that filters out repeated words in the context, ensuring the prediction of unique terms specific to the current AD. Comprehensive evaluations on MAD-Eval, CMD-AD, and TV-AD benchmarks demonstrate the superiority of DistinctAD, with the model consistently outperforming baselines, particularly in Recall@k/N, highlighting its effectiveness in producing high-quality, distinctive ADs." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 56, + 597, + 135, + 609 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 597, + 135, + 609 + ], + "spans": [ + { + "bbox": [ + 56, + 597, + 135, + 609 + ], + "type": "text", + "content": "1. Introduction" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 55, + 617, + 296, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 617, + 296, + 713 + ], + "spans": [ + { + "bbox": [ + 55, + 617, + 296, + 713 + ], + "type": "text", + "content": "Audio description (AD) [20, 67] is a crucial accessibility service that provides verbal narration of visual elements in media content for individuals who are blind or have low vision. By offering succinct and vivid descriptions, ADs enable visually impaired audiences to fully comprehend and engage with non-dialogue-related narratives, e.g., characters, facial expressions, non-verbal actions, or scene establishment. Recent studies also show ADs' value for" + } + ] + } + ], + "index": 8 + }, + { + "type": "image", + "bbox": [ + 321, + 242, + 546, + 316 + ], + "blocks": [ + { + "bbox": [ + 321, + 242, + 546, + 316 + ], + "lines": [ + { + "bbox": [ + 321, + 242, + 546, + 316 + ], + "spans": [ + { + "bbox": [ + 321, + 242, + 546, + 316 + ], + "type": "image", + "image_path": "500e9864fe6895eca71b0b1abed5cb42803e90aabfd9e97daf3d43be75d8831b.jpg" + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "image_body" + } + ], + "index": 9 + }, + { + "type": "image", + "bbox": [ + 321, + 318, + 550, + 392 + ], + "blocks": [ + { + "bbox": [ + 321, + 318, + 550, + 392 + ], + "lines": [ + { + "bbox": [ + 321, + 318, + 550, + 392 + ], + "spans": [ + { + "bbox": [ + 321, + 318, + 550, + 392 + ], + "type": "image", + "image_path": "b3c76ba5f2d0c37722518ee1c83cab3f34eebc6c4e2d1202cd1b64a869b0da0f.jpg" + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 313, + 398, + 555, + 464 + ], + "lines": [ + { + "bbox": [ + 313, + 398, + 555, + 464 + ], + "spans": [ + { + "bbox": [ + 313, + 398, + 555, + 464 + ], + "type": "text", + "content": "Figure 1. (a) Previous methods approach the AD task similar to video captioning, using only a single video clip as input, which leads to repetitive ADs due to highly similar neighboring clips. (b) Our DistinctAD method generates distinctive ADs across " + }, + { + "bbox": [ + 313, + 398, + 555, + 464 + ], + "type": "inline_equation", + "content": "N" + }, + { + "bbox": [ + 313, + 398, + 555, + 464 + ], + "type": "text", + "content": " consecutive clips, with three key innovations: VLM-AD adaptation, the Distinct Module, and explicit distinctive words prediction." + } + ] + } + ], + "index": 11, + "angle": 0, + "type": "image_caption" + } + ], + "index": 10 + }, + { + "bbox": [ + 312, + 472, + 555, + 592 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 312, + 472, + 555, + 592 + ], + "spans": [ + { + "bbox": [ + 312, + 472, + 555, + 592 + ], + "type": "text", + "content": "sighted viewers in supporting eye-free activities and facilitating child language development [31, 55], reinforcing its pivotal role in fostering inclusivity by bridging the perceptual gap between visual and non-visual elements. Crafting ADs requires careful attention to timing, language, and context to integrate smoothly with dialogue [22]. However, despite the availability of advanced AD platforms [9, 54], human-annotated methods are costly and difficult to scale, highlighting the need for automated generation systems, especially with the rise of user-generated content." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 313, + 594, + 556, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 594, + 556, + 715 + ], + "spans": [ + { + "bbox": [ + 313, + 594, + 556, + 715 + ], + "type": "text", + "content": "Advancements in Vision-Language Models (VLMs) and Large-Language Models (LLMs) have led to growing interest in automatic AD generation for media. Current approaches fall into two categories: (i) using powerful proprietary models like GPT-4 [3] or GPT-4V [53] in a training-free manner [14, 39, 82, 86], and (ii) fine-tuning open-source VLM components, such as visual-text adapters [5, 34, 42], for AD tasks [21-23, 40, 56, 74]. Both approaches have limitations: (i) Training-free methods often perform poorly and suffer from hallucinations due to the unique na" + } + ] + } + ], + "index": 13 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "spans": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "type": "text", + "content": "CVF" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 145, + 0, + 496, + 37 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 145, + 0, + 496, + 37 + ], + "spans": [ + { + "bbox": [ + 145, + 0, + 496, + 37 + ], + "type": "text", + "content": "This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "text", + "content": "13571" + } + ] + } + ], + "index": 14 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 0 + }, + { + "para_blocks": [ + { + "bbox": [ + 55, + 72, + 294, + 133 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 72, + 294, + 133 + ], + "spans": [ + { + "bbox": [ + 55, + 72, + 294, + 133 + ], + "type": "text", + "content": "ture of AD (e.g., character names and narrative coherence), which differs from the common text data LLMs are trained on. (ii) Fine-tuning methods generally perform better but are still limited by insufficient data to fully adapt to the movie-AD domain and face the context-repetition issue." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 55, + 134, + 294, + 277 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 134, + 294, + 277 + ], + "spans": [ + { + "bbox": [ + 55, + 134, + 294, + 277 + ], + "type": "text", + "content": "Unlike video captioning [38, 62], ADs are generated on consecutive intervals (visual clips) throughout long videos [68], e.g., movies. The context-repetition issue arises when models produce repetitive or similar descriptions for consecutive visual clips, especially when using prior ADs as prompts [21, 74]. This occurs because sequential clips often comprise redundant scenes or characters (and therein redundant visual features), leading models that only use the current visual clip to repeat the same information from the past, as shown in Fig. 1. However, audiences are more interested in the unique and distinct events of the current clip, rather than the common elements from the previous one." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 56, + 281, + 296, + 544 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 281, + 296, + 544 + ], + "spans": [ + { + "bbox": [ + 56, + 281, + 296, + 544 + ], + "type": "text", + "content": "In this paper, we propose DistinctAD, a two-stage framework for generating distinctive ADs within contexts. Given the domain gap between the movie-AD and VLM training data, we first bridge this gap in Stage-I by adapting VLMs, such as CLIP [58], to the movie-AD domain. Our adaptation strategy is inspired by a key observation (see Appendix §A): AD sentences encoded by the CLIP text encoder can be effectively reconstructed using simple LLMs like GPT-2 [57] with minimal fine-tuning, whereas AD reconstructions using CLIP visual features from the corresponding clips are often of poor quality. This suggests that while CLIP's multi-modal embedding space is rich enough to represent AD information, its visual encoder is insufficient for extracting it. To mitigate this domain gap, we adapt the CLIP vision encoder to better align with the frozen CLIP text encoder using existing paired video-AD data. The alignment involves global matching at video-sentence level, similar to CLIP pre-training. A challenge arises because video clips are labeled with whole ADs, and words may not appear in every frame but must be aggregated over frames. Therefore, we propose fine-grained matching at frame-word level for this multiple-instance setting." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 55, + 546, + 295, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 546, + 295, + 714 + ], + "spans": [ + { + "bbox": [ + 55, + 546, + 295, + 714 + ], + "type": "text", + "content": "For Stage-II, we propose a novel distinctive AD narrating pipeline based on the Expectation-Maximization Attention (EMA) [17] algorithm, which has demonstrated its efficacy in tasks such as semantic segmentation [35], video object segmentation [41], and text-video retrieval [27]. Differently, we apply EMA to contextual clips from long videos, which often exhibit high redundancy due to recurring scenes or characters. By extracting common bases from contextual information, DistinctAD reduces redundancy and generates compact, discriminative representations that enable the LLM decoder to produce more distinctive ADs. To further emphasize distinctiveness explicitly, we introduce a distinctive word prediction loss that filters out words that repeatedly appear in contexts, ensuring that the LLM decoder" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 313, + 72, + 553, + 119 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 72, + 553, + 119 + ], + "spans": [ + { + "bbox": [ + 313, + 72, + 553, + 119 + ], + "type": "text", + "content": "focuses on predicting unique words specific to the current AD. With these two designs, DistinctAD produces contextually distinctive and engaging ADs that can provide better narratives for the audience." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 326, + 121, + 509, + 131 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 326, + 121, + 509, + 131 + ], + "spans": [ + { + "bbox": [ + 326, + 121, + 509, + 131 + ], + "type": "text", + "content": "In summary, our contributions are three-fold:" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 313, + 132, + 553, + 311 + ], + "type": "list", + "angle": 0, + "index": 9, + "blocks": [ + { + "bbox": [ + 314, + 132, + 553, + 203 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 314, + 132, + 553, + 203 + ], + "spans": [ + { + "bbox": [ + 314, + 132, + 553, + 203 + ], + "type": "text", + "content": "- We propose a CLIP-AD adaptation strategy tailored to movie-AD data, addressing the misalignment issue caused by the domain gap. Our adapted vision encoder can be seamlessly integrated into existing CLIP-based AD methods and stands to benefit from future improvements as more AD data becomes available." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 313, + 204, + 553, + 251 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 204, + 553, + 251 + ], + "spans": [ + { + "bbox": [ + 313, + 204, + 553, + 251 + ], + "type": "text", + "content": "- We introduce DistinctAD, which incorporates a Contextual EMA module and a distinctive word prediction loss, significantly enhancing the generation of distinctive ADs from consecutive visual clips with similar contexts." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 313, + 251, + 553, + 311 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 251, + 553, + 311 + ], + "spans": [ + { + "bbox": [ + 313, + 251, + 553, + 311 + ], + "type": "text", + "content": "- Comprehensive evaluations on MAD-Eval [21], CMD-AD [23], and TV-AD [82] highlight DistinctAD's superiority. Our outstanding performance in Recall@k/N demonstrates its effectiveness in generating high-quality ADs with both distinctiveness and technical excellence." + } + ] + } + ], + "index": 8 + } + ], + "sub_type": "text" + }, + { + "bbox": [ + 314, + 323, + 400, + 335 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 314, + 323, + 400, + 335 + ], + "spans": [ + { + "bbox": [ + 314, + 323, + 400, + 335 + ], + "type": "text", + "content": "2. Related Work" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 313, + 342, + 555, + 510 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 342, + 555, + 510 + ], + "spans": [ + { + "bbox": [ + 313, + 342, + 555, + 510 + ], + "type": "text", + "content": "Dense video captioning. A task closely related to AD is dense video captioning [29], which extends traditional video captioning [38, 45, 62, 63] by both generating a single caption for trimmed video segments as well as detecting and describing multiple events with grounded timestamps. Initial dense video captioning utilize a 2-stage pipeline [25, 26, 75, 79] by firstly performing localization and then describing events. Recent works [13, 18, 36, 51, 59, 64, 65, 75, 80, 83, 89] focus on training localization and captioning modules in an end-to-end manner to enhance inter-event associations. In contrast to these works, AD generation specifically aims to narrate a coherent story, maintain character-awareness, and complement the audio track without interfering with existing dialogue." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 313, + 510, + 556, + 628 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 510, + 556, + 628 + ], + "spans": [ + { + "bbox": [ + 313, + 510, + 556, + 628 + ], + "type": "text", + "content": "AD generation. Early AD systems relied heavily on specialized authoring tools [9] and skilled human contributors. Platforms like Rescribe [54] and LiveDescribe [9] have facilitated faster and more accurate AD creation; however, these methods are costly and do not scale efficiently for large volumes of visual content. Recent efforts have developed audio segmentation and transcription systems [8, 10, 11] to create high-quality video datasets with temporally aligned ADs [60, 61, 68, 69], advancing automatic AD research." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 313, + 630, + 556, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 630, + 556, + 715 + ], + "spans": [ + { + "bbox": [ + 313, + 630, + 556, + 715 + ], + "type": "text", + "content": "In general, current automatic AD generation systems can be categorized into two approaches: training-free and partial-fine-tuning. Training-free methods [39] generate ADs by leveraging proprietary models like GPT-4 [3] and GPT-4V [53]. MM-Narrator [86] enhances AD performance by multi-model in-context learning with memories. LLM-AD [14] and AutoAD-Zero [82] use prompts com" + } + ] + } + ], + "index": 13 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "type": "text", + "content": "13572" + } + ] + } + ], + "index": 14 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 1 + }, + { + "para_blocks": [ + { + "bbox": [ + 55, + 72, + 296, + 287 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 72, + 296, + 287 + ], + "spans": [ + { + "bbox": [ + 55, + 72, + 296, + 287 + ], + "type": "text", + "content": "prising visual frames with textual character names and colorful circles [66], enabling character-centric AD generation. However, training-free AD methods often suffer from high evaluation costs at scale and relatively poor performance due to domain-specific challenges and LLM hallucinations. Partial-fine-tuning methods [21-23, 40, 74], as well as our DistinctAD, only fine-tune a lightweight adapter [5, 34] between the pre-trained vision and text encoders. A representative example is the AutoAD series [21-23], which builds automatic AD systems and enriches them with character-aware prompts within different vision-language frameworks. However, previous studies tend to focus on constructing more accurate external character banks, whereas treating AD generation similarly to video captioning, overlooks AD's unique sequential structure of video clips. In contrast, our method emphasizes understanding the visual content within its temporal context, leading to more distinctive AD generation." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 55, + 289, + 296, + 479 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 289, + 296, + 479 + ], + "spans": [ + { + "bbox": [ + 55, + 289, + 296, + 479 + ], + "type": "text", + "content": "Distinctive captioning in images aims to articulate unique details that can help distinguishing targets from others. An intuitive way to promoting distinctiveness is through contrastive learning [16, 43, 47, 73], where generated captions are encouraged to align more closely with target images rather than distractors. In [12, 76-78], group-based distinctive attention is introduced to capture distinctiveness by comparing sets of similar images and re-weighting specific caption words. A recent closely related field is difference captioning [32, 33, 84], which aims to describe differences between a single pair of images. VisDiff [19] scales difference captioning to sets containing thousands of images with natural language. Our work differs from these distinctive captioning works in that we are the first to explore distinctiveness across dense, consecutive clips within hours-long movies, thereby generating ADs with better narrative." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 55, + 493, + 111, + 506 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 493, + 111, + 506 + ], + "spans": [ + { + "bbox": [ + 55, + 493, + 111, + 506 + ], + "type": "text", + "content": "3. Method" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 55, + 514, + 219, + 527 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 514, + 219, + 527 + ], + "spans": [ + { + "bbox": [ + 55, + 514, + 219, + 527 + ], + "type": "text", + "content": "3.1. Stage-I: CLIP-AD Adaptation" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 55, + 533, + 295, + 616 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 533, + 295, + 616 + ], + "spans": [ + { + "bbox": [ + 55, + 533, + 295, + 616 + ], + "type": "text", + "content": "AD and the visual content it describes exhibit a significant domain gap compared to typical large-scale web data. This gap often causes misalignment in current partial-fine-tuning techniques. Previous studies [21, 22, 74] try alleviating this problem by pre-training LLMs on text-only AD corpus, e.g. AudioVault [1]. However, misalignment persists at the initial stage of vision encoding, which is often neglected." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 55, + 617, + 296, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 617, + 296, + 713 + ], + "spans": [ + { + "bbox": [ + 55, + 617, + 296, + 713 + ], + "type": "text", + "content": "Inspired by our findings that AD sentences encoded by the CLIP text encoder can be effectively recovered using the GPT-2 language model (see §1 and Appendix §A), we identify that the primary issue of misalignment is caused by the CLIP vision encoder, i.e., the discrepancy between visual embeddings and AD embeddings within the joint CLIP feature space. To address this, we propose adapting the CLIP vision encoder to the specific AD domain. However, due to" + } + ] + } + ], + "index": 5 + }, + { + "type": "image", + "bbox": [ + 318, + 67, + 550, + 266 + ], + "blocks": [ + { + "bbox": [ + 318, + 67, + 550, + 266 + ], + "lines": [ + { + "bbox": [ + 318, + 67, + 550, + 266 + ], + "spans": [ + { + "bbox": [ + 318, + 67, + 550, + 266 + ], + "type": "image", + "image_path": "b6264c77c575b2ff77e33f5074d1d69a9f7241fd6de9b2de0c26ced461284741.jpg" + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 313, + 270, + 555, + 315 + ], + "lines": [ + { + "bbox": [ + 313, + 270, + 555, + 315 + ], + "spans": [ + { + "bbox": [ + 313, + 270, + 555, + 315 + ], + "type": "text", + "content": "Figure 2. Illustration of Stage-I: CLIP-AD Adaptation. This process involves adapting the CLIP vision encoder to specific movie-AD data through global-level video-AD matching (bottom right) and fine-grained frame-AD matching (top right)." + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "image_caption" + } + ], + "index": 6 + }, + { + "bbox": [ + 313, + 323, + 555, + 358 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 323, + 555, + 358 + ], + "spans": [ + { + "bbox": [ + 313, + 323, + 555, + 358 + ], + "type": "text", + "content": "the unique multiple-instance learning setting for video-AD pairs (see §1), we consider both global matching and fine-grained frame-word matching in our adaptation method." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 313, + 358, + 556, + 539 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 358, + 556, + 539 + ], + "spans": [ + { + "bbox": [ + 313, + 358, + 556, + 539 + ], + "type": "text", + "content": "Global video-AD matching. A straightforward strategy involves adopting classical CLIP-style fine-tuning with video-AD pairs in large batches [46]. Formally, let video clip " + }, + { + "bbox": [ + 313, + 358, + 556, + 539 + ], + "type": "inline_equation", + "content": "\\mathbf{V}_i = [\\mathbf{f}_i^1;\\dots ;\\mathbf{f}_i^n ]\\in \\mathbb{R}^{n\\times C}" + }, + { + "bbox": [ + 313, + 358, + 556, + 539 + ], + "type": "text", + "content": " be a collection of " + }, + { + "bbox": [ + 313, + 358, + 556, + 539 + ], + "type": "inline_equation", + "content": "n" + }, + { + "bbox": [ + 313, + 358, + 556, + 539 + ], + "type": "text", + "content": " frame embeddings, and corresponding AD " + }, + { + "bbox": [ + 313, + 358, + 556, + 539 + ], + "type": "inline_equation", + "content": "\\mathbf{T}_i = [\\mathbf{w}_i^0;\\mathbf{w}_i^1;\\dots ;\\mathbf{w}_i^m ]\\in \\mathbb{R}^{(m + 1)\\times C}" + }, + { + "bbox": [ + 313, + 358, + 556, + 539 + ], + "type": "text", + "content": " be a collection of " + }, + { + "bbox": [ + 313, + 358, + 556, + 539 + ], + "type": "inline_equation", + "content": "m" + }, + { + "bbox": [ + 313, + 358, + 556, + 539 + ], + "type": "text", + "content": " word embeddings " + }, + { + "bbox": [ + 313, + 358, + 556, + 539 + ], + "type": "inline_equation", + "content": "(\\mathbf{w}_i^j)" + }, + { + "bbox": [ + 313, + 358, + 556, + 539 + ], + "type": "text", + "content": " and the [CLS] token (denoted as " + }, + { + "bbox": [ + 313, + 358, + 556, + 539 + ], + "type": "inline_equation", + "content": "\\mathbf{w}_i^0" + }, + { + "bbox": [ + 313, + 358, + 556, + 539 + ], + "type": "text", + "content": "), where " + }, + { + "bbox": [ + 313, + 358, + 556, + 539 + ], + "type": "inline_equation", + "content": "C" + }, + { + "bbox": [ + 313, + 358, + 556, + 539 + ], + "type": "text", + "content": " is the number of channels in the embedding space. We obtain the global video-level representation by averaging all frame embeddings in " + }, + { + "bbox": [ + 313, + 358, + 556, + 539 + ], + "type": "inline_equation", + "content": "\\mathbf{V}_i" + }, + { + "bbox": [ + 313, + 358, + 556, + 539 + ], + "type": "text", + "content": " using mean pooling: " + }, + { + "bbox": [ + 313, + 358, + 556, + 539 + ], + "type": "inline_equation", + "content": "\\mathbf{v}_i = \\frac{1}{n}\\sum_{j = 1}^n\\mathbf{f}_i^j" + }, + { + "bbox": [ + 313, + 358, + 556, + 539 + ], + "type": "text", + "content": ". Following the standard CLIP, we use the [CLS] token as the global textual AD representation " + }, + { + "bbox": [ + 313, + 358, + 556, + 539 + ], + "type": "inline_equation", + "content": "\\mathbf{t}_i = \\mathbf{w}_i^0" + }, + { + "bbox": [ + 313, + 358, + 556, + 539 + ], + "type": "text", + "content": ". The global video " + }, + { + "bbox": [ + 313, + 358, + 556, + 539 + ], + "type": "inline_equation", + "content": "\\rightarrow" + }, + { + "bbox": [ + 313, + 358, + 556, + 539 + ], + "type": "text", + "content": " AD matching is performed by maximizing the sum of the main diagonal of a " + }, + { + "bbox": [ + 313, + 358, + 556, + 539 + ], + "type": "inline_equation", + "content": "B\\times B" + }, + { + "bbox": [ + 313, + 358, + 556, + 539 + ], + "type": "text", + "content": " similarity matrix, using the contrastive loss:" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 328, + 541, + 555, + 574 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 328, + 541, + 555, + 574 + ], + "spans": [ + { + "bbox": [ + 328, + 541, + 555, + 574 + ], + "type": "interline_equation", + "content": "\\mathcal {L} _ {v \\rightarrow A D} = - \\frac {1}{B} \\sum_ {i = 1} ^ {B} \\log \\frac {\\exp (s i m (\\mathbf {v} _ {i} , \\mathbf {t} _ {i}))}{\\sum_ {j = 1} ^ {B} \\exp (s i m (\\mathbf {v} _ {i} , \\mathbf {t} _ {j}))}, \\tag {1}", + "image_path": "660a6f6ef93af9fdc1779c424b1daed28c8e9bd93d1ca36c1b4e11cf45d50bd0.jpg" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 313, + 582, + 555, + 665 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 582, + 555, + 665 + ], + "spans": [ + { + "bbox": [ + 313, + 582, + 555, + 665 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 313, + 582, + 555, + 665 + ], + "type": "inline_equation", + "content": "B" + }, + { + "bbox": [ + 313, + 582, + 555, + 665 + ], + "type": "text", + "content": " is the batch size, and the similarity function " + }, + { + "bbox": [ + 313, + 582, + 555, + 665 + ], + "type": "inline_equation", + "content": "sim(\\cdot)" + }, + { + "bbox": [ + 313, + 582, + 555, + 665 + ], + "type": "text", + "content": " is the vector inner product. This process is illustrated in the bottom right of Fig. 2. Similarly, we drive the AD " + }, + { + "bbox": [ + 313, + 582, + 555, + 665 + ], + "type": "inline_equation", + "content": "\\rightarrow" + }, + { + "bbox": [ + 313, + 582, + 555, + 665 + ], + "type": "text", + "content": " video loss " + }, + { + "bbox": [ + 313, + 582, + 555, + 665 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{AD\\rightarrow v}" + }, + { + "bbox": [ + 313, + 582, + 555, + 665 + ], + "type": "text", + "content": " by maximizing the sum of the secondary diagonal (i.e., swapping the " + }, + { + "bbox": [ + 313, + 582, + 555, + 665 + ], + "type": "inline_equation", + "content": "i" + }, + { + "bbox": [ + 313, + 582, + 555, + 665 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 313, + 582, + 555, + 665 + ], + "type": "inline_equation", + "content": "j" + }, + { + "bbox": [ + 313, + 582, + 555, + 665 + ], + "type": "text", + "content": " indices in (1)). The final global-level contrastive loss is then the sum of the losses in both directions " + }, + { + "bbox": [ + 313, + 582, + 555, + 665 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_q = \\mathcal{L}_{v\\to AD} + \\mathcal{L}_{AD\\to v}" + }, + { + "bbox": [ + 313, + 582, + 555, + 665 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 313, + 665, + 556, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 665, + 556, + 713 + ], + "spans": [ + { + "bbox": [ + 313, + 665, + 556, + 713 + ], + "type": "text", + "content": "Fine-grained frame-AD matching. Matching global video to AD sentence [CLS] (and vice versa) aids in joint feature space alignment. However, this alignment is insufficient for effective adaptation due to the specific multiple-instance" + } + ] + } + ], + "index": 12 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "text", + "content": "13573" + } + ] + } + ], + "index": 13 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 2 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 111, + 70, + 500, + 270 + ], + "blocks": [ + { + "bbox": [ + 111, + 70, + 500, + 270 + ], + "lines": [ + { + "bbox": [ + 111, + 70, + 500, + 270 + ], + "spans": [ + { + "bbox": [ + 111, + 70, + 500, + 270 + ], + "type": "image", + "image_path": "7727ed981f6a1a0688c7cba9ca20ca96b17b8bae844d2371ae2ba69eee25f0a6.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 54, + 272, + 555, + 307 + ], + "lines": [ + { + "bbox": [ + 54, + 272, + 555, + 307 + ], + "spans": [ + { + "bbox": [ + 54, + 272, + 555, + 307 + ], + "type": "text", + "content": "Figure 3. Pipeline of Stage-II: Distinctive AD Narration. Stage-II processes " + }, + { + "bbox": [ + 54, + 272, + 555, + 307 + ], + "type": "inline_equation", + "content": "N" + }, + { + "bbox": [ + 54, + 272, + 555, + 307 + ], + "type": "text", + "content": " consecutive video clips using the " + }, + { + "bbox": [ + 54, + 272, + 555, + 307 + ], + "type": "inline_equation", + "content": "\\mathrm{CLIP}_{\\mathrm{AD}}" + }, + { + "bbox": [ + 54, + 272, + 555, + 307 + ], + "type": "text", + "content": " vision encoder from Stage-I. We generate contextual-distinctive ADs by two key innovations: i) a Contextual EMA module to learn compact and discriminative visual representations for improved prompting of LLMs; ii) an extra distinctive word loss for predicting AD-specific terms." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "bbox": [ + 54, + 316, + 294, + 362 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 54, + 316, + 294, + 362 + ], + "spans": [ + { + "bbox": [ + 54, + 316, + 294, + 362 + ], + "type": "text", + "content": "setting of ADs, where only some words may correspond to a particular frame, but all words will have correspondence in aggregate. Thus, we propose a fine-grained matching loss at the frame-level to address this issue." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 55, + 364, + 296, + 436 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 364, + 296, + 436 + ], + "spans": [ + { + "bbox": [ + 55, + 364, + 296, + 436 + ], + "type": "text", + "content": "Formally, given the frame embeddings " + }, + { + "bbox": [ + 55, + 364, + 296, + 436 + ], + "type": "inline_equation", + "content": "\\mathbf{V}_i" + }, + { + "bbox": [ + 55, + 364, + 296, + 436 + ], + "type": "text", + "content": " and the word embeddings " + }, + { + "bbox": [ + 55, + 364, + 296, + 436 + ], + "type": "inline_equation", + "content": "\\mathbf{T}_i' = [\\mathbf{w}_i^1; \\dots; \\mathbf{w}_i^m] \\in \\mathbb{R}^{m \\times C}" + }, + { + "bbox": [ + 55, + 364, + 296, + 436 + ], + "type": "text", + "content": ", we calculate the weights of all words attending to each frame via softmax attention, taking " + }, + { + "bbox": [ + 55, + 364, + 296, + 436 + ], + "type": "inline_equation", + "content": "\\mathbf{V}_i" + }, + { + "bbox": [ + 55, + 364, + 296, + 436 + ], + "type": "text", + "content": " as the query and " + }, + { + "bbox": [ + 55, + 364, + 296, + 436 + ], + "type": "inline_equation", + "content": "\\mathbf{T}_i'" + }, + { + "bbox": [ + 55, + 364, + 296, + 436 + ], + "type": "text", + "content": " as the key. By then multiplying these attention weights by the value " + }, + { + "bbox": [ + 55, + 364, + 296, + 436 + ], + "type": "inline_equation", + "content": "\\mathbf{T}_i'" + }, + { + "bbox": [ + 55, + 364, + 296, + 436 + ], + "type": "text", + "content": ", we obtain a frame-aware AD representation " + }, + { + "bbox": [ + 55, + 364, + 296, + 436 + ], + "type": "inline_equation", + "content": "\\tilde{\\mathbf{T}}_i \\in \\mathbb{R}^{n \\times C}" + }, + { + "bbox": [ + 55, + 364, + 296, + 436 + ], + "type": "text", + "content": ":" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 82, + 440, + 295, + 456 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 82, + 440, + 295, + 456 + ], + "spans": [ + { + "bbox": [ + 82, + 440, + 295, + 456 + ], + "type": "interline_equation", + "content": "\\tilde {\\mathbf {T}} _ {i} = \\mathrm {S o f t m a x} (\\mathbf {V} _ {i} \\mathbf {T} _ {i} ^ {\\prime T} / \\tau) \\mathbf {T} _ {i} ^ {\\prime} = [ \\tilde {\\mathbf {t}} _ {i} ^ {1}; \\dots ; \\tilde {\\mathbf {t}} _ {i} ^ {n} ], \\qquad (2)", + "image_path": "120acc85f8d1d511c404379dc6aa9848871e27fc7eeaa6ba91a93f09244213ce.jpg" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 55, + 465, + 296, + 525 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 465, + 296, + 525 + ], + "spans": [ + { + "bbox": [ + 55, + 465, + 296, + 525 + ], + "type": "text", + "content": "where for each frame, the words embeddings that are most similar to the frame-level visual feature have been aggregated (via softmax attention). The temperature parameter " + }, + { + "bbox": [ + 55, + 465, + 296, + 525 + ], + "type": "inline_equation", + "content": "\\tau" + }, + { + "bbox": [ + 55, + 465, + 296, + 525 + ], + "type": "text", + "content": " controls the aggregation process, where smaller " + }, + { + "bbox": [ + 55, + 465, + 296, + 525 + ], + "type": "inline_equation", + "content": "\\tau" + }, + { + "bbox": [ + 55, + 465, + 296, + 525 + ], + "type": "text", + "content": " incorporates more textual information." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 55, + 525, + 296, + 597 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 525, + 296, + 597 + ], + "spans": [ + { + "bbox": [ + 55, + 525, + 296, + 597 + ], + "type": "text", + "content": "The goal of the fine-grained matching is to pull a frame visual feature " + }, + { + "bbox": [ + 55, + 525, + 296, + 597 + ], + "type": "inline_equation", + "content": "\\mathbf{f} \\in \\mathbf{V}_i" + }, + { + "bbox": [ + 55, + 525, + 296, + 597 + ], + "type": "text", + "content": " closer to the frame-aware AD representations " + }, + { + "bbox": [ + 55, + 525, + 296, + 597 + ], + "type": "inline_equation", + "content": "\\tilde{\\mathbf{t}} \\in \\tilde{\\mathbf{T}}_i" + }, + { + "bbox": [ + 55, + 525, + 296, + 597 + ], + "type": "text", + "content": " in (2), corresponding to the positive set. To achieve this, we define the negative set " + }, + { + "bbox": [ + 55, + 525, + 296, + 597 + ], + "type": "inline_equation", + "content": "\\tilde{\\mathbf{T}}_{neg}" + }, + { + "bbox": [ + 55, + 525, + 296, + 597 + ], + "type": "text", + "content": " as the frame-aware AD embeddings generated from other video clips (in the batch), and then use a Multi-Instance Loss [49]," + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 62, + 612, + 295, + 646 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 612, + 295, + 646 + ], + "spans": [ + { + "bbox": [ + 62, + 612, + 295, + 646 + ], + "type": "interline_equation", + "content": "\\mathcal {L} _ {f} = - \\frac {1}{B} \\sum_ {i = 1} ^ {B} \\log \\frac {\\sum_ {\\tilde {\\mathbf {t}} \\in \\tilde {\\mathbf {T}} _ {i}} \\exp (s i m (\\mathbf {f} , \\tilde {\\mathbf {t}}))}{\\sum_ {\\tilde {\\mathbf {t}} _ {*} \\in \\{\\tilde {\\mathbf {T}} _ {i} \\cup \\tilde {\\mathbf {T}} _ {n e g} \\}} \\exp (s i m (\\mathbf {f} , \\tilde {\\mathbf {t}} _ {*}))}, \\tag {3}", + "image_path": "c43bb5f68825b3688769f63c44ae72f04837569bb594ded63d09200b40a00e97.jpg" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 55, + 653, + 295, + 677 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 653, + 295, + 677 + ], + "spans": [ + { + "bbox": [ + 55, + 653, + 295, + 677 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 55, + 653, + 295, + 677 + ], + "type": "inline_equation", + "content": "\\mathbf{f} \\in \\mathbf{V}_i" + }, + { + "bbox": [ + 55, + 653, + 295, + 677 + ], + "type": "text", + "content": " is a sampled frame from " + }, + { + "bbox": [ + 55, + 653, + 295, + 677 + ], + "type": "inline_equation", + "content": "\\mathbf{V}_i" + }, + { + "bbox": [ + 55, + 653, + 295, + 677 + ], + "type": "text", + "content": ". This process is illustrated in the top right of Fig. 2." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 55, + 677, + 296, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 677, + 296, + 715 + ], + "spans": [ + { + "bbox": [ + 55, + 677, + 296, + 715 + ], + "type": "text", + "content": "Summary for Stage-I. The final objective for Stage-I is to minimize the sum of global and fine-grained aligning losses, balanced by a trade-off coefficient, " + }, + { + "bbox": [ + 55, + 677, + 296, + 715 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{\\mathrm{I}} = \\gamma \\mathcal{L}_{g} + (1 -" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 313, + 316, + 555, + 376 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 316, + 555, + 376 + ], + "spans": [ + { + "bbox": [ + 313, + 316, + 555, + 376 + ], + "type": "inline_equation", + "content": "\\gamma)\\mathcal{L}_f" + }, + { + "bbox": [ + 313, + 316, + 555, + 376 + ], + "type": "text", + "content": ". Note that during this adaptation process, the CLIP-Text encoder model remains frozen, and only the CLIP-Vision encoder is fine-tuned. Our fine-grained frame-AD matching is entirely parameter-free, as only the vision encoder will be utilized in the subsequent stage." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 313, + 381, + 499, + 395 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 381, + 499, + 395 + ], + "spans": [ + { + "bbox": [ + 313, + 381, + 499, + 395 + ], + "type": "text", + "content": "3.2. Stage-II: Distinctive AD Narration" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 313, + 400, + 555, + 532 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 400, + 555, + 532 + ], + "spans": [ + { + "bbox": [ + 313, + 400, + 555, + 532 + ], + "type": "text", + "content": "The motivation for generating distinctive ADs stems from the observation that LLM often produce repetitive descriptions for adjacent clips [56, 82, 86]. Despite improved character recognition, the visual representation itself is not discriminative among neighboring (contextual) clips, leading to uninteresting ADs. Our goal is to create contextual-distinctive ADs that highlight current differences. We hypothesize, as verified in Appendix §B, that sequential clips from a long video often share redundant scenes or characters, leading similar visual features in contexts. Thus, we propose Stage-II: distinctive AD narration." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 313, + 532, + 556, + 616 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 532, + 556, + 616 + ], + "spans": [ + { + "bbox": [ + 313, + 532, + 556, + 616 + ], + "type": "text", + "content": "As shown in Fig. 3, we prepare " + }, + { + "bbox": [ + 313, + 532, + 556, + 616 + ], + "type": "inline_equation", + "content": "N" + }, + { + "bbox": [ + 313, + 532, + 556, + 616 + ], + "type": "text", + "content": " consecutive video clips (to be AD-described) " + }, + { + "bbox": [ + 313, + 532, + 556, + 616 + ], + "type": "inline_equation", + "content": "\\{\\mathbf{x}_1,\\mathbf{x}_2,\\dots ,\\mathbf{x}_N\\}" + }, + { + "bbox": [ + 313, + 532, + 556, + 616 + ], + "type": "text", + "content": ", each containing " + }, + { + "bbox": [ + 313, + 532, + 556, + 616 + ], + "type": "inline_equation", + "content": "T" + }, + { + "bbox": [ + 313, + 532, + 556, + 616 + ], + "type": "text", + "content": " uniformly sampled frames " + }, + { + "bbox": [ + 313, + 532, + 556, + 616 + ], + "type": "inline_equation", + "content": "\\{\\mathcal{F}_1,\\mathcal{F}_2,\\dots ,\\mathcal{F}_T\\}" + }, + { + "bbox": [ + 313, + 532, + 556, + 616 + ], + "type": "text", + "content": ". Following [21, 22, 74], we employ a learnable Perceiver adapter [5] to resample " + }, + { + "bbox": [ + 313, + 532, + 556, + 616 + ], + "type": "inline_equation", + "content": "T^{\\prime}" + }, + { + "bbox": [ + 313, + 532, + 556, + 616 + ], + "type": "text", + "content": " prompt vectors for the " + }, + { + "bbox": [ + 313, + 532, + 556, + 616 + ], + "type": "inline_equation", + "content": "T" + }, + { + "bbox": [ + 313, + 532, + 556, + 616 + ], + "type": "text", + "content": " frame embeddings encoded by our Stage-I vision encoder, " + }, + { + "bbox": [ + 313, + 532, + 556, + 616 + ], + "type": "inline_equation", + "content": "\\mathrm{CLIP}_{\\mathrm{AD}}" + }, + { + "bbox": [ + 313, + 532, + 556, + 616 + ], + "type": "text", + "content": ". This process is formulated as:" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 340, + 617, + 553, + 632 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 340, + 617, + 553, + 632 + ], + "spans": [ + { + "bbox": [ + 340, + 617, + 553, + 632 + ], + "type": "interline_equation", + "content": "\\mathbf {h} _ {\\mathbf {x} _ {i}} = \\operatorname {P e r c e i v e r} \\left(\\left\\{\\mathbf {f} _ {1}, \\mathbf {f} _ {2}, \\dots , \\mathbf {f} _ {T} \\right\\}\\right) \\in \\mathbb {R} ^ {T ^ {\\prime} \\times C}, \\tag {4}", + "image_path": "77345fd271713f16b7fdc0e96b60c6e68a191cc85f0c0bd2da2cc1ae2ea92f64.jpg" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 350, + 634, + 553, + 647 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 350, + 634, + 553, + 647 + ], + "spans": [ + { + "bbox": [ + 350, + 634, + 553, + 647 + ], + "type": "interline_equation", + "content": "\\mathbf {f} _ {i} = \\operatorname {C L I P} _ {\\mathrm {A D}} \\left(\\mathcal {F} _ {i}\\right). \\tag {5}", + "image_path": "5815f14a77bfcc26d87d8f5098e5c09e8254b36b32dafb94b252918140723f59.jpg" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 313, + 654, + 555, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 654, + 555, + 714 + ], + "spans": [ + { + "bbox": [ + 313, + 654, + 555, + 714 + ], + "type": "text", + "content": "We then introduce the Contextual EMA to capture compact, discriminative visual features for distinctive AD generation. Contextual EMA. Expectation-Maximization Attention (EMA) [35] integrates the attention mechanism [81] into the classical EM [17] algorithm, which comprises three steps to" + } + ] + } + ], + "index": 16 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "type": "text", + "content": "13574" + } + ] + } + ], + "index": 17 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 3 + }, + { + "para_blocks": [ + { + "bbox": [ + 55, + 72, + 294, + 143 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 72, + 294, + 143 + ], + "spans": [ + { + "bbox": [ + 55, + 72, + 294, + 143 + ], + "type": "text", + "content": "estimate a more compact set of bases: Responsibility Estimation (RE), Likelihood Maximization (LM), and Data Re-estimation (DR). Inspired by this, we propose Contextual EMA to perform EMA on frames from " + }, + { + "bbox": [ + 55, + 72, + 294, + 143 + ], + "type": "inline_equation", + "content": "N" + }, + { + "bbox": [ + 55, + 72, + 294, + 143 + ], + "type": "text", + "content": " contextual clips, aiming to eliminate redundancy, learn compact representations, and explore distinctiveness." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 55, + 144, + 294, + 229 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 144, + 294, + 229 + ], + "spans": [ + { + "bbox": [ + 55, + 144, + 294, + 229 + ], + "type": "text", + "content": "Let " + }, + { + "bbox": [ + 55, + 144, + 294, + 229 + ], + "type": "inline_equation", + "content": "\\mathcal{H} = \\{\\mathbf{h}_{\\mathbf{x}_i}\\}_{i = 1}^N\\in \\mathbb{R}^{N\\times T'\\times C}" + }, + { + "bbox": [ + 55, + 144, + 294, + 229 + ], + "type": "text", + "content": " represent " + }, + { + "bbox": [ + 55, + 144, + 294, + 229 + ], + "type": "inline_equation", + "content": "N" + }, + { + "bbox": [ + 55, + 144, + 294, + 229 + ], + "type": "text", + "content": " clip vectors from the Perceiver, and " + }, + { + "bbox": [ + 55, + 144, + 294, + 229 + ], + "type": "inline_equation", + "content": "\\mathcal{M} = \\{\\mu_k\\}_{k = 1}^K\\in \\mathbb{R}^{K\\times C}" + }, + { + "bbox": [ + 55, + 144, + 294, + 229 + ], + "type": "text", + "content": " denote the randomly initialized base features, where " + }, + { + "bbox": [ + 55, + 144, + 294, + 229 + ], + "type": "inline_equation", + "content": "C,K" + }, + { + "bbox": [ + 55, + 144, + 294, + 229 + ], + "type": "text", + "content": " indicate the number of channel and bases. The RE step estimates the hidden variable " + }, + { + "bbox": [ + 55, + 144, + 294, + 229 + ], + "type": "inline_equation", + "content": "\\mathcal{Z} = \\{z_{nk}\\}_{n = 1,k = 1}^{N\\times T',K}" + }, + { + "bbox": [ + 55, + 144, + 294, + 229 + ], + "type": "text", + "content": ", where the responsibility " + }, + { + "bbox": [ + 55, + 144, + 294, + 229 + ], + "type": "inline_equation", + "content": "z_{nk}" + }, + { + "bbox": [ + 55, + 144, + 294, + 229 + ], + "type": "text", + "content": " represents the probability of the " + }, + { + "bbox": [ + 55, + 144, + 294, + 229 + ], + "type": "inline_equation", + "content": "n" + }, + { + "bbox": [ + 55, + 144, + 294, + 229 + ], + "type": "text", + "content": "-th frame belonging to the " + }, + { + "bbox": [ + 55, + 144, + 294, + 229 + ], + "type": "inline_equation", + "content": "k" + }, + { + "bbox": [ + 55, + 144, + 294, + 229 + ], + "type": "text", + "content": "-th base:" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 115, + 233, + 294, + 264 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 115, + 233, + 294, + 264 + ], + "spans": [ + { + "bbox": [ + 115, + 233, + 294, + 264 + ], + "type": "interline_equation", + "content": "z _ {n k} = \\frac {\\exp \\left(\\mathbf {h} _ {n} \\mu_ {k} ^ {T} / \\tau\\right)}{\\sum_ {j = 1} ^ {K} \\exp \\left(\\mathbf {h} _ {n} \\mu_ {j} ^ {T} / \\tau\\right)}, \\tag {6}", + "image_path": "c808c8c23132a8f57518c48b0510b354aaf6ce9db8bf765ed5de261250bd28d4.jpg" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 55, + 271, + 295, + 308 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 271, + 295, + 308 + ], + "spans": [ + { + "bbox": [ + 55, + 271, + 295, + 308 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 55, + 271, + 295, + 308 + ], + "type": "inline_equation", + "content": "\\tau" + }, + { + "bbox": [ + 55, + 271, + 295, + 308 + ], + "type": "text", + "content": " determines the shape (peakiness) of distribution " + }, + { + "bbox": [ + 55, + 271, + 295, + 308 + ], + "type": "inline_equation", + "content": "\\mathcal{Z}" + }, + { + "bbox": [ + 55, + 271, + 295, + 308 + ], + "type": "text", + "content": ". Then, the LM step updates the bases " + }, + { + "bbox": [ + 55, + 271, + 295, + 308 + ], + "type": "inline_equation", + "content": "\\mathcal{M}" + }, + { + "bbox": [ + 55, + 271, + 295, + 308 + ], + "type": "text", + "content": " by applying the weighted average on input " + }, + { + "bbox": [ + 55, + 271, + 295, + 308 + ], + "type": "inline_equation", + "content": "\\mathcal{H}" + }, + { + "bbox": [ + 55, + 271, + 295, + 308 + ], + "type": "text", + "content": ", formulating the " + }, + { + "bbox": [ + 55, + 271, + 295, + 308 + ], + "type": "inline_equation", + "content": "k" + }, + { + "bbox": [ + 55, + 271, + 295, + 308 + ], + "type": "text", + "content": "-th base as:" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 128, + 311, + 294, + 343 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 128, + 311, + 294, + 343 + ], + "spans": [ + { + "bbox": [ + 128, + 311, + 294, + 343 + ], + "type": "interline_equation", + "content": "\\mu_ {k} = \\frac {\\sum_ {n = 1} ^ {N \\times T ^ {\\prime}} z _ {n k} \\mathbf {h} _ {n}}{\\sum_ {n = 1} ^ {N \\times T ^ {\\prime}} z _ {n k}}. \\tag {7}", + "image_path": "aa588f201492bedf4b144df5ce669c36af5bee8cfc7f3e9e136246b72136ebf9.jpg" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 55, + 346, + 295, + 394 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 346, + 295, + 394 + ], + "spans": [ + { + "bbox": [ + 55, + 346, + 295, + 394 + ], + "type": "text", + "content": "The RE (E-step) and LM (M-step) are iteratively performed " + }, + { + "bbox": [ + 55, + 346, + 295, + 394 + ], + "type": "inline_equation", + "content": "R" + }, + { + "bbox": [ + 55, + 346, + 295, + 394 + ], + "type": "text", + "content": " times until convergence. Notably, since bases number " + }, + { + "bbox": [ + 55, + 346, + 295, + 394 + ], + "type": "inline_equation", + "content": "K" + }, + { + "bbox": [ + 55, + 346, + 295, + 394 + ], + "type": "text", + "content": " is much smaller than the embedding number " + }, + { + "bbox": [ + 55, + 346, + 295, + 394 + ], + "type": "inline_equation", + "content": "N \\times T'" + }, + { + "bbox": [ + 55, + 346, + 295, + 394 + ], + "type": "text", + "content": ", we employ DR to reconstruct a compact version of " + }, + { + "bbox": [ + 55, + 346, + 295, + 394 + ], + "type": "inline_equation", + "content": "\\mathcal{H}" + }, + { + "bbox": [ + 55, + 346, + 295, + 394 + ], + "type": "text", + "content": " through:" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 151, + 399, + 294, + 414 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 151, + 399, + 294, + 414 + ], + "spans": [ + { + "bbox": [ + 151, + 399, + 294, + 414 + ], + "type": "interline_equation", + "content": "\\widehat {\\mathcal {H}} \\approx \\mathcal {Z M}. \\tag {8}", + "image_path": "53b540b7ff06b4d3c4212d5338e6a265d85b913cfc3f01c2eff991119a0f8444.jpg" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 55, + 417, + 295, + 443 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 417, + 295, + 443 + ], + "spans": [ + { + "bbox": [ + 55, + 417, + 295, + 443 + ], + "type": "text", + "content": "Here, " + }, + { + "bbox": [ + 55, + 417, + 295, + 443 + ], + "type": "inline_equation", + "content": "\\widehat{\\mathcal{H}}\\in \\mathbb{R}^{N\\times T^{\\prime}\\times C}" + }, + { + "bbox": [ + 55, + 417, + 295, + 443 + ], + "type": "text", + "content": " retains the same shape as " + }, + { + "bbox": [ + 55, + 417, + 295, + 443 + ], + "type": "inline_equation", + "content": "\\mathcal{H}" + }, + { + "bbox": [ + 55, + 417, + 295, + 443 + ], + "type": "text", + "content": ". We combine " + }, + { + "bbox": [ + 55, + 417, + 295, + 443 + ], + "type": "inline_equation", + "content": "\\mathcal{H}" + }, + { + "bbox": [ + 55, + 417, + 295, + 443 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 55, + 417, + 295, + 443 + ], + "type": "inline_equation", + "content": "\\widehat{\\mathcal{H}}" + }, + { + "bbox": [ + 55, + 417, + 295, + 443 + ], + "type": "text", + "content": " element-wise with a hyperparameter " + }, + { + "bbox": [ + 55, + 417, + 295, + 443 + ], + "type": "inline_equation", + "content": "\\alpha" + }, + { + "bbox": [ + 55, + 417, + 295, + 443 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 55, + 444, + 295, + 479 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 444, + 295, + 479 + ], + "spans": [ + { + "bbox": [ + 55, + 444, + 295, + 479 + ], + "type": "text", + "content": "To enhance representation distinctiveness, we introduce an additional branch using cross-attention between raw " + }, + { + "bbox": [ + 55, + 444, + 295, + 479 + ], + "type": "inline_equation", + "content": "\\mathcal{H}" + }, + { + "bbox": [ + 55, + 444, + 295, + 479 + ], + "type": "text", + "content": " (query) and bases " + }, + { + "bbox": [ + 55, + 444, + 295, + 479 + ], + "type": "inline_equation", + "content": "\\mathcal{M}" + }, + { + "bbox": [ + 55, + 444, + 295, + 479 + ], + "type": "text", + "content": " (key and value), formulated as:" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 111, + 482, + 294, + 497 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 111, + 482, + 294, + 497 + ], + "spans": [ + { + "bbox": [ + 111, + 482, + 294, + 497 + ], + "type": "interline_equation", + "content": "\\widetilde {\\mathcal {H}} = \\operatorname {C r o s s A t t e n t i o n} (\\mathcal {H}, \\mathcal {M}), \\tag {9}", + "image_path": "d1febcb6c562f6c0313eb674e77935338a49eb97a4cdb2a9f101ba0cb637be10.jpg" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 55, + 506, + 295, + 602 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 506, + 295, + 602 + ], + "spans": [ + { + "bbox": [ + 55, + 506, + 295, + 602 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 55, + 506, + 295, + 602 + ], + "type": "inline_equation", + "content": "\\widetilde{\\mathcal{H}}\\in \\mathbb{R}^{N\\times T^{\\prime}\\times C}" + }, + { + "bbox": [ + 55, + 506, + 295, + 602 + ], + "type": "text", + "content": ". Linear layers projecting queries, keys, and values are omitted in Fig. 3 (see Appendix §C for details). Through (9), we process the distributions of " + }, + { + "bbox": [ + 55, + 506, + 295, + 602 + ], + "type": "inline_equation", + "content": "\\mathcal{H}" + }, + { + "bbox": [ + 55, + 506, + 295, + 602 + ], + "type": "text", + "content": " to attend on specific and informative bases, with improved linear separability (see Fig. 6). We combine " + }, + { + "bbox": [ + 55, + 506, + 295, + 602 + ], + "type": "inline_equation", + "content": "\\mathcal{H},\\widehat{\\mathcal{H}},\\widetilde{\\mathcal{H}}" + }, + { + "bbox": [ + 55, + 506, + 295, + 602 + ], + "type": "text", + "content": " elementwise around Contextual EMA to construct the final refined visual features. These features are then projected into the LLM embedding space using a single-layer projector:" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 108, + 606, + 294, + 620 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 108, + 606, + 294, + 620 + ], + "spans": [ + { + "bbox": [ + 108, + 606, + 294, + 620 + ], + "type": "interline_equation", + "content": "\\mathcal {H} _ {\\text {s u m}} = \\operatorname {P r o j} (\\mathcal {H} + \\alpha \\widehat {\\mathcal {H}} + \\beta \\widetilde {\\mathcal {H}}). \\tag {10}", + "image_path": "a970465db34ac5fd5f9cb2016fee53e33740aca99b4f71d86dc53d1dcff5eb31.jpg" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 55, + 629, + 295, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 629, + 295, + 713 + ], + "spans": [ + { + "bbox": [ + 55, + 629, + 295, + 713 + ], + "type": "text", + "content": "Interleaved prompt as LLM's input. Following previous studies [22, 74, 82], we build our interleaved prompt enriched with character information, (see Fig. 3). To answer the \"who is who\" question when more than two characters are present, the corresponding actors' portrait images are projected as face tokens for reasoning. The " + }, + { + "bbox": [ + 55, + 629, + 295, + 713 + ], + "type": "inline_equation", + "content": "\\langle \\text{BOS} \\rangle" + }, + { + "bbox": [ + 55, + 629, + 295, + 713 + ], + "type": "text", + "content": " tag appended at the end indicates the beginning of AD generation." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 313, + 72, + 553, + 120 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 72, + 553, + 120 + ], + "spans": [ + { + "bbox": [ + 313, + 72, + 553, + 120 + ], + "type": "text", + "content": "Distinctive words highlighting. Our goal is to query a frozen LLM for AD generation using a vision-conditioned prompt. The typical supervision employs the commonly used auto-regressive loss function:" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 348, + 126, + 553, + 150 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 348, + 126, + 553, + 150 + ], + "spans": [ + { + "bbox": [ + 348, + 126, + 553, + 150 + ], + "type": "interline_equation", + "content": "\\mathcal {L} _ {\\text {a u t o}} = - \\sum_ {n} \\log P _ {\\theta} \\left(w _ {n} \\mid \\text {p r o m p t}; w _ {< n}\\right), \\tag {11}", + "image_path": "ea9d399c8df882a3878ebe5bfb5576f718cbc8d4bcc9cff19b41c4b806eda0da.jpg" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 313, + 157, + 553, + 253 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 157, + 553, + 253 + ], + "spans": [ + { + "bbox": [ + 313, + 157, + 553, + 253 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 313, + 157, + 553, + 253 + ], + "type": "inline_equation", + "content": "w_{n}" + }, + { + "bbox": [ + 313, + 157, + 553, + 253 + ], + "type": "text", + "content": " is the " + }, + { + "bbox": [ + 313, + 157, + 553, + 253 + ], + "type": "inline_equation", + "content": "n" + }, + { + "bbox": [ + 313, + 157, + 553, + 253 + ], + "type": "text", + "content": "-th token from the target AD. However, " + }, + { + "bbox": [ + 313, + 157, + 553, + 253 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{auto}" + }, + { + "bbox": [ + 313, + 157, + 553, + 253 + ], + "type": "text", + "content": " does not emphasize the distinctiveness specific to the current AD, which is our focus. To address this, we propose a distinctive word set " + }, + { + "bbox": [ + 313, + 157, + 553, + 253 + ], + "type": "inline_equation", + "content": "w_{d}" + }, + { + "bbox": [ + 313, + 157, + 553, + 253 + ], + "type": "text", + "content": ", created by filtering out duplicates, such as character names, prepositions, and pronouns, from the " + }, + { + "bbox": [ + 313, + 157, + 553, + 253 + ], + "type": "inline_equation", + "content": "N" + }, + { + "bbox": [ + 313, + 157, + 553, + 253 + ], + "type": "text", + "content": " context ADs of the target AD. During training, we explicitly encourage the LLM to predict the distinctive words in " + }, + { + "bbox": [ + 313, + 157, + 553, + 253 + ], + "type": "inline_equation", + "content": "w_{d}" + }, + { + "bbox": [ + 313, + 157, + 553, + 253 + ], + "type": "text", + "content": " by optimizing the distinctive loss " + }, + { + "bbox": [ + 313, + 157, + 553, + 253 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{dist}" + }, + { + "bbox": [ + 313, + 157, + 553, + 253 + ], + "type": "text", + "content": ":" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 319, + 255, + 553, + 287 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 319, + 255, + 553, + 287 + ], + "spans": [ + { + "bbox": [ + 319, + 255, + 553, + 287 + ], + "type": "interline_equation", + "content": "\\mathcal {L} _ {\\text {d i s t}} = - \\sum_ {n = 1} ^ {N} \\sum_ {i = 1} ^ {u} \\log P _ {\\theta} \\left(w _ {n} = w _ {d} ^ {i} \\mid \\text {p r o m p t}, w _ {< n}\\right), \\tag {12}", + "image_path": "da965c716d8b565e9b8d9638a81f0e6f5e17b9ff17f45c465c57179ba5a797c2.jpg" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 313, + 295, + 553, + 331 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 295, + 553, + 331 + ], + "spans": [ + { + "bbox": [ + 313, + 295, + 553, + 331 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 313, + 295, + 553, + 331 + ], + "type": "inline_equation", + "content": "w_{d}^{i}" + }, + { + "bbox": [ + 313, + 295, + 553, + 331 + ], + "type": "text", + "content": " denotes the " + }, + { + "bbox": [ + 313, + 295, + 553, + 331 + ], + "type": "inline_equation", + "content": "i" + }, + { + "bbox": [ + 313, + 295, + 553, + 331 + ], + "type": "text", + "content": "-th distinctive word in " + }, + { + "bbox": [ + 313, + 295, + 553, + 331 + ], + "type": "inline_equation", + "content": "w_{d}" + }, + { + "bbox": [ + 313, + 295, + 553, + 331 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 313, + 295, + 553, + 331 + ], + "type": "inline_equation", + "content": "u" + }, + { + "bbox": [ + 313, + 295, + 553, + 331 + ], + "type": "text", + "content": " is the size of the set. The final complete loss function for Stage-II is: " + }, + { + "bbox": [ + 313, + 295, + 553, + 331 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{\\mathrm{II}} = \\mathcal{L}_{\\mathrm{auto}} + \\mathcal{L}_{\\mathrm{dist}}" + }, + { + "bbox": [ + 313, + 295, + 553, + 331 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 314, + 341, + 394, + 354 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 314, + 341, + 394, + 354 + ], + "spans": [ + { + "bbox": [ + 314, + 341, + 394, + 354 + ], + "type": "text", + "content": "4. Experiments" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 314, + 360, + 422, + 373 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 314, + 360, + 422, + 373 + ], + "spans": [ + { + "bbox": [ + 314, + 360, + 422, + 373 + ], + "type": "text", + "content": "4.1. Experiment Setup" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 313, + 378, + 553, + 533 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 378, + 553, + 533 + ], + "spans": [ + { + "bbox": [ + 313, + 378, + 553, + 533 + ], + "type": "text", + "content": "Datasets. We follow the AD generation benchmark established in AutoAD [21], conducting experiments on the denoised MAD-v2-Named [68] and evaluating on MAD-Eval-Named split. Specifically, MAD-v2-Named includes " + }, + { + "bbox": [ + 313, + 378, + 553, + 533 + ], + "type": "inline_equation", + "content": "\\sim 330\\mathrm{k}" + }, + { + "bbox": [ + 313, + 378, + 553, + 533 + ], + "type": "text", + "content": " ADs from 488 movies for training and MAD-Eval has 6,520 ADs crawled from 10 movies for evaluation. We also evaluate on two recently introduced datasets. CMD-AD [23] (where \"CMD\" stands for Condensed Movie Dataset [7]) is a movie AD dataset that contains 101k ADs for more than 1432 movies, with 100 movies split for CMD-AD evaluation. TV-AD [82] is a recently proposed AD dataset based on TVQA [30], which contains " + }, + { + "bbox": [ + 313, + 378, + 553, + 533 + ], + "type": "inline_equation", + "content": "\\sim 31\\mathrm{k}" + }, + { + "bbox": [ + 313, + 378, + 553, + 533 + ], + "type": "text", + "content": " ADs for training and " + }, + { + "bbox": [ + 313, + 378, + 553, + 533 + ], + "type": "inline_equation", + "content": "\\sim 3\\mathrm{k}" + }, + { + "bbox": [ + 313, + 378, + 553, + 533 + ], + "type": "text", + "content": " ADs for evaluation." + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 313, + 534, + 553, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 534, + 553, + 713 + ], + "spans": [ + { + "bbox": [ + 313, + 534, + 553, + 713 + ], + "type": "text", + "content": "Evaluation Metrics. Classic captioning metrics including ROUGE-L [37], CIDEr [72] and SPICE [6] are reported to evaluate the quality of generated ADs versus the ground-truth. Besides, we also report Recall@k within " + }, + { + "bbox": [ + 313, + 534, + 553, + 713 + ], + "type": "inline_equation", + "content": "N" + }, + { + "bbox": [ + 313, + 534, + 553, + 713 + ], + "type": "text", + "content": " Neighbours [22] (" + }, + { + "bbox": [ + 313, + 534, + 553, + 713 + ], + "type": "inline_equation", + "content": "\\mathbf{R}@\\mathbf{k} / \\mathbf{N}" + }, + { + "bbox": [ + 313, + 534, + 553, + 713 + ], + "type": "text", + "content": "), which calculates the average value of Recall@k for each AD with its " + }, + { + "bbox": [ + 313, + 534, + 553, + 713 + ], + "type": "inline_equation", + "content": "N" + }, + { + "bbox": [ + 313, + 534, + 553, + 713 + ], + "type": "text", + "content": " temporally adjacent GT texts, where BertScore [88] is used for text similarity matching. The R@k/N metric is based on retrieving the most similar ground-truth AD among N neighbors, and thus highlights the distinctiveness of generated ADs directly. LLM-AD-eval [23] employs LLMs to assess the quality of generated ADs, assigning scores from 1 (lowest) to 5 (highest). We utilize the LLM prompt from the original study [82] and apply open-source models LLaMA2-7B-Chat [70] and LLaMA3-8B-Instruct [4] for this evaluation." + } + ] + } + ], + "index": 21 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "text", + "content": "13575" + } + ] + } + ], + "index": 22 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 4 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 90, + 70, + 520, + 282 + ], + "blocks": [ + { + "bbox": [ + 90, + 70, + 520, + 282 + ], + "lines": [ + { + "bbox": [ + 90, + 70, + 520, + 282 + ], + "spans": [ + { + "bbox": [ + 90, + 70, + 520, + 282 + ], + "type": "table", + "html": "
MethodPub.VLMLLMROUGE-LCIDErSPICER@5/16
Training-free
VLog [2]--GPT-47.51.32.142.3
MM-Vid [39]ArXiv'23GPT-4V-9.86.13.846.1
MM-Narrator [86]CVPR'24CLIP-L14GPT-413.413.95.249.0
LLM-AD [14]ArXiv'24GPT-4V-13.520.5--
AutoAD-Zero [82]ACCV'24VideoLLaMA2-7BLLaMA3-8B-22.4--
Partial-fine-tuning
ClipCap [50]ArXiv'21CLIP-B32GPT-28.54.41.1-
CapDec [52]ArXiv'22--8.26.71.4-
AutoAD-I [21]CVPR'23CLIP-B32GPT-211.914.34.442.1
AutoAD-II [22]ICCV'23CLIP-B32GPT-213.419.5-50.8
AutoAD-III [23]CVPR'24EVA-CLIPOPT-2.7B-22.8-52.0
AutoAD-III [23]CVPR'24EVA-CLIPLLaMA2-7B-24.0-52.8
MovieSeq [40]ECCV'24CLIP-B16LLaMA2-7B*15.524.47.051.6
DistinctAD (Ours)CLIP-B32GPT-215.424.56.749.8
DistinctAD (Ours)CLIPAD-B32GPT-216.425.57.451.7
DistinctAD (Ours)CLIPAD-B16LLaMA2-7B17.227.08.255.6
DistinctAD (Ours)CLIPAD-B16LLaMA3-8B17.627.38.356.0
", + "image_path": "4f3578a5c7b2fece87a3048126aa726430d17f4243d49dce85cb212755e38c53.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "table_body" + } + ], + "index": 0 + }, + { + "type": "table", + "bbox": [ + 77, + 319, + 272, + 410 + ], + "blocks": [ + { + "bbox": [ + 55, + 286, + 553, + 309 + ], + "lines": [ + { + "bbox": [ + 55, + 286, + 553, + 309 + ], + "spans": [ + { + "bbox": [ + 55, + 286, + 553, + 309 + ], + "type": "text", + "content": "Table 1. Comparisons of AD performance on the MAD-Eval benchmark. * indicates fine-tuning LLaMA2-7B model with LoRA [24]. " + }, + { + "bbox": [ + 55, + 286, + 553, + 309 + ], + "type": "inline_equation", + "content": "{\\mathrm{{CLIP}}}_{\\mathrm{{AD}}}" + }, + { + "bbox": [ + 55, + 286, + 553, + 309 + ], + "type": "text", + "content": " is our CLIP vision encoder adapted using our Stage-I strategy." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 77, + 319, + 272, + 410 + ], + "lines": [ + { + "bbox": [ + 77, + 319, + 272, + 410 + ], + "spans": [ + { + "bbox": [ + 77, + 319, + 272, + 410 + ], + "type": "table", + "html": "
MethodCIDErR@1/5LLM-AD-eval
Video-BLIP2 [85]4.822.01.89
Video-LLaMA2 [87]5.223.61.91
AutoAD-II [22]13.526.12.08
AutoAD-III [23]21.730.02.85
AutoAD-Zero [82]17.7-2.83
DistinctAD (LLaMA2)22.332.92.89
DistinctAD (LLaMA3)22.733.02.88
AutoAD-III† [23]25.031.22.89
2.01
", + "image_path": "09ec3956087a4e1192315246e224c20b24d0596867f667b733ec2abe3de3a7c1.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "table_body" + } + ], + "index": 2 + }, + { + "type": "table", + "bbox": [ + 77, + 453, + 274, + 505 + ], + "blocks": [ + { + "bbox": [ + 55, + 413, + 295, + 447 + ], + "lines": [ + { + "bbox": [ + 55, + 413, + 295, + 447 + ], + "spans": [ + { + "bbox": [ + 55, + 413, + 295, + 447 + ], + "type": "text", + "content": "Table 2. Comparisons on CMD-AD. The LLM-AD-eval scores are evaluated with LLaMA2-7B (left) and LLaMA3-8B (right). " + }, + { + "bbox": [ + 55, + 413, + 295, + 447 + ], + "type": "inline_equation", + "content": "\\dagger" + }, + { + "bbox": [ + 55, + 413, + 295, + 447 + ], + "type": "text", + "content": " indicates pre-training on 3.4M HowTo-AD dataset [23, 48]." + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 77, + 453, + 274, + 505 + ], + "lines": [ + { + "bbox": [ + 77, + 453, + 274, + 505 + ], + "spans": [ + { + "bbox": [ + 77, + 453, + 274, + 505 + ], + "type": "table", + "html": "
MethodCIDErR@1/5LLM-AD-eval
AutoAD-III [23]26.1-2.78 | 1.99
AutoAD-Zero [82]22.630.62.94 | 2.00
DistinctAD (LLaMA2)27.231.82.86 | 2.01
DistinctAD (LLaMA3)27.432.12.89 | 2.00
", + "image_path": "d72d3da9bb1daec7802dcb790dd06a53214cf72fcd9fa2c2511acf9a98827fc9.jpg" + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "table_body" + } + ], + "index": 4 + }, + { + "bbox": [ + 55, + 508, + 295, + 531 + ], + "lines": [ + { + "bbox": [ + 55, + 508, + 295, + 531 + ], + "spans": [ + { + "bbox": [ + 55, + 508, + 295, + 531 + ], + "type": "text", + "content": "Table 3. Comparisons on TV-AD. The LLM-AD-eval scores are evaluated using LLaMA2-7B (left) and LLaMA3-8B (right)." + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "text" + }, + { + "bbox": [ + 55, + 545, + 295, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 545, + 295, + 713 + ], + "spans": [ + { + "bbox": [ + 55, + 545, + 295, + 713 + ], + "type": "text", + "content": "Implementation Details. To facilitate CLIP-AD adaptation in Stage-I, we collect the original raw movies from MAD from platforms like Amazon Prime Video. See Appendix F for details. We fine-tune the CLIP Vision encoder for 5 epochs with a fixed learning rate 5e-5 using the Adam optimizer [28] in Stage-I, with a batch size of 512. In Stage-II, we use a batch of 8 sequences, each containing 16 consecutive video AD-pairs from a movie. For each video clip, 8 frames are uniformly sampled. We use the AdamW [44] optimizer to train our model for 10 epochs, with a cosine-decayed learning rate and linear warm-up. The learning rate is set to " + }, + { + "bbox": [ + 55, + 545, + 295, + 713 + ], + "type": "inline_equation", + "content": "10^{-4}" + }, + { + "bbox": [ + 55, + 545, + 295, + 713 + ], + "type": "text", + "content": " for both GPT-2 and LLaMA models. For external character information, we directly use the inference results from AutoAD-Zero [82] as it gives current best face" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 313, + 321, + 553, + 346 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 321, + 553, + 346 + ], + "spans": [ + { + "bbox": [ + 313, + 321, + 553, + 346 + ], + "type": "text", + "content": "recognition performance. We use CLIP-B32 and GPT-2 on MAD-Eval-Named for ablations unless specified otherwise." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 313, + 356, + 507, + 369 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 356, + 507, + 369 + ], + "spans": [ + { + "bbox": [ + 313, + 356, + 507, + 369 + ], + "type": "text", + "content": "4.2. Comparisons with previous methods" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 313, + 376, + 553, + 411 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 376, + 553, + 411 + ], + "spans": [ + { + "bbox": [ + 313, + 376, + 553, + 411 + ], + "type": "text", + "content": "We conduct comprehensive comparisons using the widely-adopted MAD-Eval benchmark [21] and two recently introduced AD datasets, CMD-AD [23] and TV-AD [82]." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 313, + 413, + 554, + 605 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 413, + 554, + 605 + ], + "spans": [ + { + "bbox": [ + 313, + 413, + 554, + 605 + ], + "type": "text", + "content": "Comparisons on MAD-Eval are shown in Tab. 1. We primarily categorize previous studies into Training-free and Partial-fine-tuning approaches, as described in §2. Our method is a Partial-fine-tuning method. When using the same CLIP-B32 and GPT-2, our proposed DistinctAD achieves a CIDEr score of 24.5, surpassing previous AutoAD-I [21] (CIDEr 14.3) and AutoAD-II [22] (CIDEr 19.5). With our Stage-I adapted CLIP vision encoders (denoted as " + }, + { + "bbox": [ + 313, + 413, + 554, + 605 + ], + "type": "inline_equation", + "content": "\\mathrm{CLIP}_{\\mathrm{AD}}" + }, + { + "bbox": [ + 313, + 413, + 554, + 605 + ], + "type": "text", + "content": "), we observe stable improvements across all metrics, e.g. 25.5 vs. 24.5 on CIDEr and 51.7 vs. 49.8 on recall, validating the effectiveness of our Stage-I strategy. Notably, DistinctAD with " + }, + { + "bbox": [ + 313, + 413, + 554, + 605 + ], + "type": "inline_equation", + "content": "\\mathrm{CLIP}_{\\mathrm{AD}}" + }, + { + "bbox": [ + 313, + 413, + 554, + 605 + ], + "type": "text", + "content": "-B16 and LLaMA3-8B [4] achieves state-of-the-arts with CIDEr 27.3 and Recall@5/16 56.0. Our outstanding performance on R@k/N demonstrates DistinctAD's ability to generate distinctive ADs, well matching the uniqueness of the clip's contents." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 313, + 605, + 554, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 605, + 554, + 713 + ], + "spans": [ + { + "bbox": [ + 313, + 605, + 554, + 713 + ], + "type": "text", + "content": "Looking at the training-free methods, despite the capabilities of advanced proprietary VLMs, e.g. GPT-4V [53], and LLMs, e.g. GPT-4 [3], the performance of training-free methods remains inferior to those employing partial-finetuning. This discrepancy likely arises from the unique characteristics of AD and movie data, which exhibit a significant domain gap from common vision language training data. As such, these data types were not encountered during the pretraining of proprietary large-scale models." + } + ] + } + ], + "index": 11 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "text", + "content": "13576" + } + ] + } + ], + "index": 12 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 5 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 65, + 70, + 206, + 134 + ], + "blocks": [ + { + "bbox": [ + 65, + 70, + 206, + 134 + ], + "lines": [ + { + "bbox": [ + 65, + 70, + 206, + 134 + ], + "spans": [ + { + "bbox": [ + 65, + 70, + 206, + 134 + ], + "type": "table", + "html": "
SettingCIDErR@5/16
None6.734.0
Global Lg8.236.6
Fine-grained Lf7.735.2
γLg + (1 - γ)Lf8.636.9
", + "image_path": "b69b234dfc62a2522b13cf1b39b525e57be97f0364c99cceb7ffe7917eacd17e.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "table_body" + } + ], + "index": 0 + }, + { + "type": "table", + "bbox": [ + 229, + 70, + 320, + 134 + ], + "blocks": [ + { + "bbox": [ + 97, + 136, + 176, + 145 + ], + "lines": [ + { + "bbox": [ + 97, + 136, + 176, + 145 + ], + "spans": [ + { + "bbox": [ + 97, + 136, + 176, + 145 + ], + "type": "text", + "content": "(a) Stage-I components." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 229, + 70, + 320, + 134 + ], + "lines": [ + { + "bbox": [ + 229, + 70, + 320, + 134 + ], + "spans": [ + { + "bbox": [ + 229, + 70, + 320, + 134 + ], + "type": "table", + "html": "
Coefficient γCIDEr
0.18.0
0.38.5
0.58.6
0.77.7
", + "image_path": "5c367284857a31e23b1937b2572b76bdc355a866f45403dc2e7ddc25090e096d.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "table_body" + } + ], + "index": 2 + }, + { + "type": "table", + "bbox": [ + 335, + 70, + 550, + 134 + ], + "blocks": [ + { + "bbox": [ + 231, + 136, + 319, + 145 + ], + "lines": [ + { + "bbox": [ + 231, + 136, + 319, + 145 + ], + "spans": [ + { + "bbox": [ + 231, + 136, + 319, + 145 + ], + "type": "text", + "content": "(b) Impact of coefficient " + }, + { + "bbox": [ + 231, + 136, + 319, + 145 + ], + "type": "inline_equation", + "content": "\\gamma" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 335, + 70, + 550, + 134 + ], + "lines": [ + { + "bbox": [ + 335, + 70, + 550, + 134 + ], + "spans": [ + { + "bbox": [ + 335, + 70, + 550, + 134 + ], + "type": "table", + "html": "
PromptStage-ICIDErR@5/16
Contextual ADs [21]X12.6 (17.8)39.8 (43.1)
✓14.1 (19.0)39.9 (44.2)
Character [22]X22.045.6
✓23.146.2
", + "image_path": "e7c6a12cf35f36ecff551cf5219c27bf8e9084b7a706eb461922b68410aa04e5.jpg" + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "table_body" + } + ], + "index": 4 + }, + { + "bbox": [ + 55, + 149, + 555, + 195 + ], + "lines": [ + { + "bbox": [ + 55, + 149, + 555, + 195 + ], + "spans": [ + { + "bbox": [ + 55, + 149, + 555, + 195 + ], + "type": "text", + "content": "Table 4. Ablation studies for Stage-I. (a) Evaluation of global video-AD loss " + }, + { + "bbox": [ + 55, + 149, + 555, + 195 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_f" + }, + { + "bbox": [ + 55, + 149, + 555, + 195 + ], + "type": "text", + "content": " and fine-grained frame-AD loss " + }, + { + "bbox": [ + 55, + 149, + 555, + 195 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_f" + }, + { + "bbox": [ + 55, + 149, + 555, + 195 + ], + "type": "text", + "content": " on AD performance. (b) Analysis of the impact of the coefficient " + }, + { + "bbox": [ + 55, + 149, + 555, + 195 + ], + "type": "inline_equation", + "content": "\\gamma" + }, + { + "bbox": [ + 55, + 149, + 555, + 195 + ], + "type": "text", + "content": ". Both (a) and (b) are conducted with pure visual prompts. (c) Impact of Stage-I when combined with different prompts for the LLM decoder, including contextual ADs and character names. Performance in parentheses indicates results with ground-truth contextual ADs as prompts." + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "text" + }, + { + "type": "image", + "bbox": [ + 57, + 198, + 294, + 253 + ], + "blocks": [ + { + "bbox": [ + 57, + 198, + 294, + 253 + ], + "lines": [ + { + "bbox": [ + 57, + 198, + 294, + 253 + ], + "spans": [ + { + "bbox": [ + 57, + 198, + 294, + 253 + ], + "type": "image", + "image_path": "5e32d7ebd4bcc1f151229deb6c3474cbecff188ccbaaeb862f97f24cb61baffe.jpg" + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 55, + 255, + 295, + 320 + ], + "lines": [ + { + "bbox": [ + 55, + 255, + 295, + 320 + ], + "spans": [ + { + "bbox": [ + 55, + 255, + 295, + 320 + ], + "type": "text", + "content": "Figure 4. Ablation studies for hyperparameter in Stage-II, with final settings highlighted in orange. (a) Impact of " + }, + { + "bbox": [ + 55, + 255, + 295, + 320 + ], + "type": "inline_equation", + "content": "\\alpha" + }, + { + "bbox": [ + 55, + 255, + 295, + 320 + ], + "type": "text", + "content": " on the weight of compact representation " + }, + { + "bbox": [ + 55, + 255, + 295, + 320 + ], + "type": "inline_equation", + "content": "\\widehat{\\mathcal{H}}" + }, + { + "bbox": [ + 55, + 255, + 295, + 320 + ], + "type": "text", + "content": ". (b) Influence of " + }, + { + "bbox": [ + 55, + 255, + 295, + 320 + ], + "type": "inline_equation", + "content": "\\beta" + }, + { + "bbox": [ + 55, + 255, + 295, + 320 + ], + "type": "text", + "content": " on cross-attended feature " + }, + { + "bbox": [ + 55, + 255, + 295, + 320 + ], + "type": "inline_equation", + "content": "\\widetilde{\\mathcal{H}}" + }, + { + "bbox": [ + 55, + 255, + 295, + 320 + ], + "type": "text", + "content": ". (c) Impact of " + }, + { + "bbox": [ + 55, + 255, + 295, + 320 + ], + "type": "inline_equation", + "content": "K" + }, + { + "bbox": [ + 55, + 255, + 295, + 320 + ], + "type": "text", + "content": ", which denotes the number of clusters in bases " + }, + { + "bbox": [ + 55, + 255, + 295, + 320 + ], + "type": "inline_equation", + "content": "\\mathcal{M}" + }, + { + "bbox": [ + 55, + 255, + 295, + 320 + ], + "type": "text", + "content": ". (d) Effect of sampling " + }, + { + "bbox": [ + 55, + 255, + 295, + 320 + ], + "type": "inline_equation", + "content": "N" + }, + { + "bbox": [ + 55, + 255, + 295, + 320 + ], + "type": "text", + "content": " consecutive video clips. We switch to larger memory GPUs when " + }, + { + "bbox": [ + 55, + 255, + 295, + 320 + ], + "type": "inline_equation", + "content": "N" + }, + { + "bbox": [ + 55, + 255, + 295, + 320 + ], + "type": "text", + "content": " exceeds 16." + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "image_caption" + } + ], + "index": 7 + }, + { + "type": "table", + "bbox": [ + 77, + 322, + 276, + 430 + ], + "blocks": [ + { + "bbox": [ + 374, + 135, + 512, + 145 + ], + "lines": [ + { + "bbox": [ + 374, + 135, + 512, + 145 + ], + "spans": [ + { + "bbox": [ + 374, + 135, + 512, + 145 + ], + "type": "text", + "content": "(c) Impact of Stage-I w/ different prompts." + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 77, + 322, + 276, + 430 + ], + "lines": [ + { + "bbox": [ + 77, + 322, + 276, + 430 + ], + "spans": [ + { + "bbox": [ + 77, + 322, + 276, + 430 + ], + "type": "table", + "html": "
Ex#αHβHLdistCIDErR@5/16
A0XXX23.1 (27.4)46.2
B1✓XX23.7 (29.3)46.6
B2X✓X23.4 (29.1)46.1
B3✓✓X23.3 (28.1)48.0
C0XX✓24.7 (29.4)49.5
C1✓X✓24.3 (29.4)50.7
C2X✓✓25.3 (30.4)51.5
C3✓✓✓25.5 (29.8)51.7
", + "image_path": "43d56d3256641f2e3547a6f483f512f2bc87a7ea52840c2b8c432330c1b15e73.jpg" + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "table_body" + } + ], + "index": 9 + }, + { + "bbox": [ + 55, + 483, + 295, + 664 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 483, + 295, + 664 + ], + "spans": [ + { + "bbox": [ + 55, + 483, + 295, + 664 + ], + "type": "text", + "content": "Comparisons on CMD-AD and TV-AD. We further verify the generalizability of DistinctAD on the recently proposed CMD-AD and TV-AD benchmarks, with results presented in Tables 2 and 3. Both evaluations use adapted " + }, + { + "bbox": [ + 55, + 483, + 295, + 664 + ], + "type": "inline_equation", + "content": "\\mathrm{CLIP}_{\\mathrm{AD}}" + }, + { + "bbox": [ + 55, + 483, + 295, + 664 + ], + "type": "text", + "content": "-B16 model. DistinctAD outperforms AutoAD-Zero, AutoAD-II and AutoAD-III in terms of CIDEr and R@1/5 on both benchmarks Meanwhile, DistinctAD exhibits a lower CIDEr compared to AutoAD-III† on CMD-AD, which we conjecture is primarily due to AutoAD-III† pre-training on a very large-scale 3.4M transformed HowTo-AD dataset [23, 48], which is currently publicly unavailable. Despite this, DistinctAD achieves superior R@1/5, underscoring its exceptional ability to generate distinctive and high-quality ADs. This is further corroborated by its leading performance on the LLM-AD-eval metric." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 55, + 671, + 154, + 682 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 671, + 154, + 682 + ], + "spans": [ + { + "bbox": [ + 55, + 671, + 154, + 682 + ], + "type": "text", + "content": "4.3. Ablation studies" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 55, + 689, + 295, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 689, + 295, + 714 + ], + "spans": [ + { + "bbox": [ + 55, + 689, + 295, + 714 + ], + "type": "text", + "content": "Effect of CLIP-AD Adaptation (I). Tab. 4a demonstrates the benefit of our Stage-I strategy, i.e. adapting CLIP the vi" + } + ] + } + ], + "index": 13 + }, + { + "type": "table", + "bbox": [ + 360, + 198, + 509, + 242 + ], + "blocks": [ + { + "bbox": [ + 55, + 432, + 295, + 477 + ], + "lines": [ + { + "bbox": [ + 55, + 432, + 295, + 477 + ], + "spans": [ + { + "bbox": [ + 55, + 432, + 295, + 477 + ], + "type": "text", + "content": "Table 5. Ablation studies for components in Stage-II. The CIDEr column shows scores with AutoAD-Zero's character [82] as prompt by default. CIDEr in parentheses represent performance with ground-truth character names." + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 360, + 198, + 509, + 242 + ], + "lines": [ + { + "bbox": [ + 360, + 198, + 509, + 242 + ], + "spans": [ + { + "bbox": [ + 360, + 198, + 509, + 242 + ], + "type": "table", + "html": "
Consecutive N?CIDErR@5/16
✓25.551.7
×23.8↓1.752.5↑0.8
", + "image_path": "01faa9a5865d3f66a638bd67c828bce3b44df8046eaf6d59988102939191b5bb.jpg" + } + ] + } + ], + "index": 14, + "angle": 0, + "type": "table_body" + } + ], + "index": 14 + }, + { + "bbox": [ + 315, + 245, + 551, + 256 + ], + "lines": [ + { + "bbox": [ + 315, + 245, + 551, + 256 + ], + "spans": [ + { + "bbox": [ + 315, + 245, + 551, + 256 + ], + "type": "text", + "content": "Table 6. Impact of whether sampling " + }, + { + "bbox": [ + 315, + 245, + 551, + 256 + ], + "type": "inline_equation", + "content": "N" + }, + { + "bbox": [ + 315, + 245, + 551, + 256 + ], + "type": "text", + "content": " consecutive clips or not." + } + ] + } + ], + "index": 15, + "angle": 0, + "type": "text" + }, + { + "bbox": [ + 313, + 268, + 553, + 388 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 268, + 553, + 388 + ], + "spans": [ + { + "bbox": [ + 313, + 268, + 553, + 388 + ], + "type": "text", + "content": "sion encoder to the movie-AD domain via global video-AD matching " + }, + { + "bbox": [ + 313, + 268, + 553, + 388 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_g" + }, + { + "bbox": [ + 313, + 268, + 553, + 388 + ], + "type": "text", + "content": " and fine-grained frame-AD matching " + }, + { + "bbox": [ + 313, + 268, + 553, + 388 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_f" + }, + { + "bbox": [ + 313, + 268, + 553, + 388 + ], + "type": "text", + "content": ". In Tab. 4b, the balancing coefficient " + }, + { + "bbox": [ + 313, + 268, + 553, + 388 + ], + "type": "inline_equation", + "content": "\\gamma" + }, + { + "bbox": [ + 313, + 268, + 553, + 388 + ], + "type": "text", + "content": " performs best at 0.5. In Tab. 4c, our Stage-I strategy consistently enhances performance when combined with different prompts in the LLM decoder, such as contextual ADs in AutoAD-I [21] or character names in AutoAD-II [22]. This indicates that our AD-adapted CLIP vision encoder can integrate seamlessly into previous methods, including those training-free models that utilize CLIP-based visual extractors." + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 313, + 389, + 554, + 605 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 389, + 554, + 605 + ], + "spans": [ + { + "bbox": [ + 313, + 389, + 554, + 605 + ], + "type": "text", + "content": "Effect of Distinctive AD narration (II). We evaluate the effectiveness of Stage-II components in Tab. 5, based on the default " + }, + { + "bbox": [ + 313, + 389, + 554, + 605 + ], + "type": "inline_equation", + "content": "\\mathcal{H}" + }, + { + "bbox": [ + 313, + 389, + 554, + 605 + ], + "type": "text", + "content": " (Perceiver's output) and full AD auto-regressive loss " + }, + { + "bbox": [ + 313, + 389, + 554, + 605 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{\\text {auto }}" + }, + { + "bbox": [ + 313, + 389, + 554, + 605 + ], + "type": "text", + "content": ". The baseline (A0) outperforms AutoAD-II in CIDEr (23.1 vs. 19.5), primarily due to more accurate character prompts from AutoAD-Zero and the adapted CLIP vision encoder from Stage-I. A0 with AutoAD-II's characters achieves a CIDEr score of 20.6, close to AutoAD-II's performance. Incorporating reconstructed feature " + }, + { + "bbox": [ + 313, + 389, + 554, + 605 + ], + "type": "inline_equation", + "content": "\\widehat{\\mathcal{H}}" + }, + { + "bbox": [ + 313, + 389, + 554, + 605 + ], + "type": "text", + "content": " brings stable improvements on both CIDEr and recall (B1 & C1), highlighting the importance of compact representations in understanding visual semantics. Cross-attended feature " + }, + { + "bbox": [ + 313, + 389, + 554, + 605 + ], + "type": "inline_equation", + "content": "\\widetilde{\\mathcal{H}}" + }, + { + "bbox": [ + 313, + 389, + 554, + 605 + ], + "type": "text", + "content": " works better together with distinctive word prediction loss " + }, + { + "bbox": [ + 313, + 389, + 554, + 605 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{\\text {dist }}" + }, + { + "bbox": [ + 313, + 389, + 554, + 605 + ], + "type": "text", + "content": " (C2 vs. B2, C3 vs. B3). We conjecture this is because " + }, + { + "bbox": [ + 313, + 389, + 554, + 605 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{\\text {dist }}" + }, + { + "bbox": [ + 313, + 389, + 554, + 605 + ], + "type": "text", + "content": " provides more definite supervision on re-weighting distinctive words, which guides " + }, + { + "bbox": [ + 313, + 389, + 554, + 605 + ], + "type": "inline_equation", + "content": "\\widehat{\\mathcal{H}}" + }, + { + "bbox": [ + 313, + 389, + 554, + 605 + ], + "type": "text", + "content": " to attend on concept-related bases. Overall, applying the full Stage-II pipeline brings significant and robust performance (C3)." + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 313, + 605, + 555, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 605, + 555, + 714 + ], + "spans": [ + { + "bbox": [ + 313, + 605, + 555, + 714 + ], + "type": "text", + "content": "Effect of Hyper-parameters. Fig. 4 summarizes the ablation studies on 4 hyper-parameters that potentially influence the results in Stage-II. Coefficient weights " + }, + { + "bbox": [ + 313, + 605, + 555, + 714 + ], + "type": "inline_equation", + "content": "\\alpha" + }, + { + "bbox": [ + 313, + 605, + 555, + 714 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 313, + 605, + 555, + 714 + ], + "type": "inline_equation", + "content": "\\beta" + }, + { + "bbox": [ + 313, + 605, + 555, + 714 + ], + "type": "text", + "content": " yield optimal results when set to 3 and 1, respectively. This suggests the need to refine our final representations to be more compact for generating ADs. Fig. 4(c) shows setting bases number " + }, + { + "bbox": [ + 313, + 605, + 555, + 714 + ], + "type": "inline_equation", + "content": "K" + }, + { + "bbox": [ + 313, + 605, + 555, + 714 + ], + "type": "text", + "content": " to 32 yields best. A smaller " + }, + { + "bbox": [ + 313, + 605, + 555, + 714 + ], + "type": "inline_equation", + "content": "K" + }, + { + "bbox": [ + 313, + 605, + 555, + 714 + ], + "type": "text", + "content": ", e.g. 2, can still achieve notable CIDEr, as the Contextual EMA module does not significantly alter the final output " + }, + { + "bbox": [ + 313, + 605, + 555, + 714 + ], + "type": "inline_equation", + "content": "\\mathcal{H}_{sum}" + }, + { + "bbox": [ + 313, + 605, + 555, + 714 + ], + "type": "text", + "content": ". However," + } + ] + } + ], + "index": 18 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "text", + "content": "13577" + } + ] + } + ], + "index": 19 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 6 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 75, + 75, + 87, + 86 + ], + "blocks": [ + { + "bbox": [ + 75, + 75, + 87, + 86 + ], + "lines": [ + { + "bbox": [ + 75, + 75, + 87, + 86 + ], + "spans": [ + { + "bbox": [ + 75, + 75, + 87, + 86 + ], + "type": "image", + "image_path": "2ccc1aef0fa6873a1c16a4b22dfe62f6ed3a0a36aed0fd712fdb6583fc388f04.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + } + ], + "index": 0 + }, + { + "bbox": [ + 89, + 76, + 132, + 85 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 89, + 76, + 132, + 85 + ], + "spans": [ + { + "bbox": [ + 89, + 76, + 132, + 85 + ], + "type": "text", + "content": "1h12m53.6s" + } + ] + } + ], + "index": 1 + }, + { + "type": "image", + "bbox": [ + 78, + 87, + 234, + 121 + ], + "blocks": [ + { + "bbox": [ + 78, + 87, + 234, + 121 + ], + "lines": [ + { + "bbox": [ + 78, + 87, + 234, + 121 + ], + "spans": [ + { + "bbox": [ + 78, + 87, + 234, + 121 + ], + "type": "image", + "image_path": "2d5c202bd5e428ef58c4655aed83eb2dcc2f4c3af0da9788eb29006af5f7bc2e.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + } + ], + "index": 2 + }, + { + "bbox": [ + 76, + 122, + 225, + 142 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 76, + 122, + 225, + 142 + ], + "spans": [ + { + "bbox": [ + 76, + 122, + 225, + 142 + ], + "type": "text", + "content": "GT: Holding the room's phone to his ear, the man on the bed regards him sympathetically." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 77, + 143, + 205, + 153 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 77, + 143, + 205, + 153 + ], + "spans": [ + { + "bbox": [ + 77, + 143, + 205, + 153 + ], + "type": "text", + "content": "AutoAD-Zero: He looks at something." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 77, + 154, + 217, + 163 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 77, + 154, + 217, + 163 + ], + "spans": [ + { + "bbox": [ + 77, + 154, + 217, + 163 + ], + "type": "text", + "content": "DistinctAD: He holds the phone to his ear." + } + ] + } + ], + "index": 5 + }, + { + "type": "image", + "bbox": [ + 249, + 76, + 259, + 86 + ], + "blocks": [ + { + "bbox": [ + 249, + 76, + 259, + 86 + ], + "lines": [ + { + "bbox": [ + 249, + 76, + 259, + 86 + ], + "spans": [ + { + "bbox": [ + 249, + 76, + 259, + 86 + ], + "type": "image", + "image_path": "dc60bae16010cbbd6600786418ec0d51289a1534bb7887ee6fc3791de2b47696.jpg" + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "image_body" + } + ], + "index": 6 + }, + { + "bbox": [ + 262, + 76, + 304, + 85 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 262, + 76, + 304, + 85 + ], + "spans": [ + { + "bbox": [ + 262, + 76, + 304, + 85 + ], + "type": "text", + "content": "1h12m59.1s" + } + ] + } + ], + "index": 7 + }, + { + "type": "image", + "bbox": [ + 250, + 87, + 407, + 121 + ], + "blocks": [ + { + "bbox": [ + 250, + 87, + 407, + 121 + ], + "lines": [ + { + "bbox": [ + 250, + 87, + 407, + 121 + ], + "spans": [ + { + "bbox": [ + 250, + 87, + 407, + 121 + ], + "type": "image", + "image_path": "7327c6f13a5c90c45f2bc5d23e61020a8d28786593f750d639f35b1183c4c6b1.jpg" + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "image_body" + } + ], + "index": 8 + }, + { + "bbox": [ + 250, + 122, + 375, + 141 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 250, + 122, + 375, + 141 + ], + "spans": [ + { + "bbox": [ + 250, + 122, + 375, + 141 + ], + "type": "text", + "content": "GT: Stephen sees the pill bottles on a bedside table." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 251, + 143, + 395, + 153 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 251, + 143, + 395, + 153 + ], + "spans": [ + { + "bbox": [ + 251, + 143, + 395, + 153 + ], + "type": "text", + "content": "AutoAD-Zero: Stephen looks at something." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 251, + 154, + 388, + 173 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 251, + 154, + 388, + 173 + ], + "spans": [ + { + "bbox": [ + 251, + 154, + 388, + 173 + ], + "type": "text", + "content": "DistinctAD: In Stephen's bedroom, a pill bottle sits on his nightstand." + } + ] + } + ], + "index": 11 + }, + { + "type": "image", + "bbox": [ + 422, + 76, + 434, + 86 + ], + "blocks": [ + { + "bbox": [ + 422, + 76, + 434, + 86 + ], + "lines": [ + { + "bbox": [ + 422, + 76, + 434, + 86 + ], + "spans": [ + { + "bbox": [ + 422, + 76, + 434, + 86 + ], + "type": "image", + "image_path": "e1b7ef6e20b0a2526f1d30ade4570b895b78d28f8a44e85883894cbc63216b9e.jpg" + } + ] + } + ], + "index": 12, + "angle": 0, + "type": "image_body" + } + ], + "index": 12 + }, + { + "bbox": [ + 436, + 76, + 477, + 85 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 436, + 76, + 477, + 85 + ], + "spans": [ + { + "bbox": [ + 436, + 76, + 477, + 85 + ], + "type": "text", + "content": "1h13m14.5s" + } + ] + } + ], + "index": 13 + }, + { + "type": "image", + "bbox": [ + 422, + 87, + 527, + 121 + ], + "blocks": [ + { + "bbox": [ + 422, + 87, + 527, + 121 + ], + "lines": [ + { + "bbox": [ + 422, + 87, + 527, + 121 + ], + "spans": [ + { + "bbox": [ + 422, + 87, + 527, + 121 + ], + "type": "image", + "image_path": "ebd129b9530fb77b259d18208c403a32fb2c50a1c482de679bb742a4aa972e90.jpg" + } + ] + } + ], + "index": 14, + "angle": 0, + "type": "image_body" + } + ], + "index": 14 + }, + { + "bbox": [ + 398, + 122, + 529, + 142 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 398, + 122, + 529, + 142 + ], + "spans": [ + { + "bbox": [ + 398, + 122, + 529, + 142 + ], + "type": "text", + "content": "GT: The man on the bed sits in silhouette against the room's soft lamp light." + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 403, + 143, + 514, + 162 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 403, + 143, + 514, + 162 + ], + "spans": [ + { + "bbox": [ + 403, + 143, + 514, + 162 + ], + "type": "text", + "content": "AutoAD-Zero: He turns to leave, looking at him." + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 403, + 163, + 504, + 173 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 403, + 163, + 504, + 173 + ], + "spans": [ + { + "bbox": [ + 403, + 163, + 504, + 173 + ], + "type": "text", + "content": "DistinctAD: The bed is empty." + } + ] + } + ], + "index": 17 + }, + { + "type": "image", + "bbox": [ + 76, + 173, + 88, + 184 + ], + "blocks": [ + { + "bbox": [ + 76, + 173, + 88, + 184 + ], + "lines": [ + { + "bbox": [ + 76, + 173, + 88, + 184 + ], + "spans": [ + { + "bbox": [ + 76, + 173, + 88, + 184 + ], + "type": "image", + "image_path": "37954c648a97d7f8189ac87b46a889a7b7a69cb60e18444612d6b9015bb9c08b.jpg" + } + ] + } + ], + "index": 18, + "angle": 0, + "type": "image_body" + } + ], + "index": 18 + }, + { + "bbox": [ + 89, + 175, + 132, + 183 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 89, + 175, + 132, + 183 + ], + "spans": [ + { + "bbox": [ + 89, + 175, + 132, + 183 + ], + "type": "text", + "content": "1h13m19.4s" + } + ] + } + ], + "index": 19 + }, + { + "type": "image", + "bbox": [ + 78, + 186, + 182, + 219 + ], + "blocks": [ + { + "bbox": [ + 78, + 186, + 182, + 219 + ], + "lines": [ + { + "bbox": [ + 78, + 186, + 182, + 219 + ], + "spans": [ + { + "bbox": [ + 78, + 186, + 182, + 219 + ], + "type": "image", + "image_path": "81ddb833fd820f8365e260173580a31af48f41d367e4ed7f01440bd29bb08c2f.jpg" + } + ] + } + ], + "index": 20, + "angle": 0, + "type": "image_body" + } + ], + "index": 20 + }, + { + "bbox": [ + 76, + 221, + 177, + 240 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 76, + 221, + 177, + 240 + ], + "spans": [ + { + "bbox": [ + 76, + 221, + 177, + 240 + ], + "type": "text", + "content": "GT: Later, Stephen steps back out into the rainy night." + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 76, + 242, + 182, + 262 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 76, + 242, + 182, + 262 + ], + "spans": [ + { + "bbox": [ + 76, + 242, + 182, + 262 + ], + "type": "text", + "content": "AutoAD-Zero: Stephen looks around and at another character." + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 76, + 262, + 192, + 282 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 76, + 262, + 192, + 282 + ], + "spans": [ + { + "bbox": [ + 76, + 262, + 192, + 282 + ], + "type": "text", + "content": "DistinctAD: As Stephen walks away, he glances over his shoulder." + } + ] + } + ], + "index": 23 + }, + { + "type": "image", + "bbox": [ + 198, + 174, + 210, + 184 + ], + "blocks": [ + { + "bbox": [ + 198, + 174, + 210, + 184 + ], + "lines": [ + { + "bbox": [ + 198, + 174, + 210, + 184 + ], + "spans": [ + { + "bbox": [ + 198, + 174, + 210, + 184 + ], + "type": "image", + "image_path": "aa5955ececbf0391df747327102a561453c74d813506b3ca97e40903a178655f.jpg" + } + ] + } + ], + "index": 24, + "angle": 0, + "type": "image_body" + } + ], + "index": 24 + }, + { + "bbox": [ + 211, + 175, + 252, + 183 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 211, + 175, + 252, + 183 + ], + "spans": [ + { + "bbox": [ + 211, + 175, + 252, + 183 + ], + "type": "text", + "content": "1h13m23.1s" + } + ] + } + ], + "index": 25 + }, + { + "type": "image", + "bbox": [ + 198, + 186, + 303, + 219 + ], + "blocks": [ + { + "bbox": [ + 198, + 186, + 303, + 219 + ], + "lines": [ + { + "bbox": [ + 198, + 186, + 303, + 219 + ], + "spans": [ + { + "bbox": [ + 198, + 186, + 303, + 219 + ], + "type": "image", + "image_path": "23068bb9811e451a9448d946e9582d6e135b846bdb82bfa1466d538458be3d6a.jpg" + } + ] + } + ], + "index": 26, + "angle": 0, + "type": "image_body" + } + ], + "index": 26 + }, + { + "bbox": [ + 197, + 221, + 333, + 240 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 197, + 221, + 333, + 240 + ], + "spans": [ + { + "bbox": [ + 197, + 221, + 333, + 240 + ], + "type": "text", + "content": "GT: Moving with a stiff gait, Stephen marches across the parking lot to his car." + } + ] + } + ], + "index": 27 + }, + { + "bbox": [ + 197, + 242, + 328, + 253 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 197, + 242, + 328, + 253 + ], + "spans": [ + { + "bbox": [ + 197, + 242, + 328, + 253 + ], + "type": "text", + "content": "AutoAD-Zero: He looks out at the city." + } + ] + } + ], + "index": 28 + }, + { + "bbox": [ + 197, + 255, + 332, + 274 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 197, + 255, + 332, + 274 + ], + "spans": [ + { + "bbox": [ + 197, + 255, + 332, + 274 + ], + "type": "text", + "content": "DistinctAD: He looks up at the sky, then turns and heads back to the car." + } + ] + } + ], + "index": 29 + }, + { + "type": "image", + "bbox": [ + 334, + 174, + 345, + 184 + ], + "blocks": [ + { + "bbox": [ + 334, + 174, + 345, + 184 + ], + "lines": [ + { + "bbox": [ + 334, + 174, + 345, + 184 + ], + "spans": [ + { + "bbox": [ + 334, + 174, + 345, + 184 + ], + "type": "image", + "image_path": "12c637b2b01d4f5a6b888cc9353072a41356abd721a8166d5ed3be3d7650980c.jpg" + } + ] + } + ], + "index": 30, + "angle": 0, + "type": "image_body" + } + ], + "index": 30 + }, + { + "bbox": [ + 334, + 175, + 372, + 184 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 334, + 175, + 372, + 184 + ], + "spans": [ + { + "bbox": [ + 334, + 175, + 372, + 184 + ], + "type": "text", + "content": "h13m24.5s" + } + ] + } + ], + "index": 31 + }, + { + "type": "image", + "bbox": [ + 319, + 186, + 527, + 219 + ], + "blocks": [ + { + "bbox": [ + 319, + 186, + 527, + 219 + ], + "lines": [ + { + "bbox": [ + 319, + 186, + 527, + 219 + ], + "spans": [ + { + "bbox": [ + 319, + 186, + 527, + 219 + ], + "type": "image", + "image_path": "2d337008c1aa526bc9bc5fe5aece4ce62e57c4e8c4940a65503b7c1552eec619.jpg" + } + ] + } + ], + "index": 32, + "angle": 0, + "type": "image_body" + } + ], + "index": 32 + }, + { + "bbox": [ + 343, + 221, + 515, + 240 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 343, + 221, + 515, + 240 + ], + "spans": [ + { + "bbox": [ + 343, + 221, + 515, + 240 + ], + "type": "text", + "content": "GT: As Stephen gets in, rain patters heavily against his windshield." + } + ] + } + ], + "index": 33 + }, + { + "bbox": [ + 343, + 242, + 492, + 252 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 343, + 242, + 492, + 252 + ], + "spans": [ + { + "bbox": [ + 343, + 242, + 492, + 252 + ], + "type": "text", + "content": "AutoAD-Zero: Stephen looks at the women." + } + ] + } + ], + "index": 34 + }, + { + "bbox": [ + 343, + 255, + 524, + 265 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 343, + 255, + 524, + 265 + ], + "spans": [ + { + "bbox": [ + 343, + 255, + 524, + 265 + ], + "type": "text", + "content": "DistinctAD: In the car, Stephen stares out the window." + } + ] + } + ], + "index": 35 + }, + { + "bbox": [ + 394, + 273, + 523, + 283 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 394, + 273, + 523, + 283 + ], + "spans": [ + { + "bbox": [ + 394, + 273, + 523, + 283 + ], + "type": "text", + "content": ": The Ides of March (2011)" + } + ] + } + ], + "index": 36 + }, + { + "bbox": [ + 55, + 326, + 296, + 396 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 326, + 296, + 396 + ], + "spans": [ + { + "bbox": [ + 55, + 326, + 296, + 396 + ], + "type": "text", + "content": "unsuitable values of " + }, + { + "bbox": [ + 55, + 326, + 296, + 396 + ], + "type": "inline_equation", + "content": "K" + }, + { + "bbox": [ + 55, + 326, + 296, + 396 + ], + "type": "text", + "content": ", e.g. 8 or 64, can negatively impact performance. Fig. 4(d) reveals that increasing the number of consecutive clips " + }, + { + "bbox": [ + 55, + 326, + 296, + 396 + ], + "type": "inline_equation", + "content": "N" + }, + { + "bbox": [ + 55, + 326, + 296, + 396 + ], + "type": "text", + "content": " (with " + }, + { + "bbox": [ + 55, + 326, + 296, + 396 + ], + "type": "inline_equation", + "content": "K" + }, + { + "bbox": [ + 55, + 326, + 296, + 396 + ], + "type": "text", + "content": " set to 32) enhances the CIDEr score, though this effect saturates when " + }, + { + "bbox": [ + 55, + 326, + 296, + 396 + ], + "type": "inline_equation", + "content": "N" + }, + { + "bbox": [ + 55, + 326, + 296, + 396 + ], + "type": "text", + "content": " exceeds 16. This demonstrates that more bases should be created to effectively summarize components with additional clips." + } + ] + } + ], + "index": 38 + }, + { + "bbox": [ + 55, + 398, + 296, + 604 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 398, + 296, + 604 + ], + "spans": [ + { + "bbox": [ + 55, + 398, + 296, + 604 + ], + "type": "text", + "content": "Do the " + }, + { + "bbox": [ + 55, + 398, + 296, + 604 + ], + "type": "inline_equation", + "content": "N" + }, + { + "bbox": [ + 55, + 398, + 296, + 604 + ], + "type": "text", + "content": " clips to be consecutive? Tab. 6 presents the results of sampling non-consecutive " + }, + { + "bbox": [ + 55, + 398, + 296, + 604 + ], + "type": "inline_equation", + "content": "N" + }, + { + "bbox": [ + 55, + 398, + 296, + 604 + ], + "type": "text", + "content": " clips during training. When using non-continuous clips, we observe a decrease in the CIDEr metric by 1.7 (25.5 vs. 23.8) because the Contextual EMA module struggles with unrelated contexts. However, the R@5/16 improves by 0.8, which indicates enhancement of the distinctiveness (uniqueness) of the generated AD when using more diverse visual contents. Visualizations. To better understand what Contextual EMA learns, we show the t-SNE [71] visualizations of " + }, + { + "bbox": [ + 55, + 398, + 296, + 604 + ], + "type": "inline_equation", + "content": "\\mathcal{H}" + }, + { + "bbox": [ + 55, + 398, + 296, + 604 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 55, + 398, + 296, + 604 + ], + "type": "inline_equation", + "content": "\\widehat{\\mathcal{H}}" + }, + { + "bbox": [ + 55, + 398, + 296, + 604 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 55, + 398, + 296, + 604 + ], + "type": "inline_equation", + "content": "\\widetilde{\\mathcal{H}}" + }, + { + "bbox": [ + 55, + 398, + 296, + 604 + ], + "type": "text", + "content": " (from §3.2) in Fig. 6, using the same perplexity value across all visualizations. With Contextual EMA, " + }, + { + "bbox": [ + 55, + 398, + 296, + 604 + ], + "type": "inline_equation", + "content": "\\widehat{\\mathcal{H}}" + }, + { + "bbox": [ + 55, + 398, + 296, + 604 + ], + "type": "text", + "content": " exhibits more compact features compared to raw " + }, + { + "bbox": [ + 55, + 398, + 296, + 604 + ], + "type": "inline_equation", + "content": "\\mathcal{H}" + }, + { + "bbox": [ + 55, + 398, + 296, + 604 + ], + "type": "text", + "content": ", Fig. 6(b). Interestingly in Fig. 6(c), cross-attention between " + }, + { + "bbox": [ + 55, + 398, + 296, + 604 + ], + "type": "inline_equation", + "content": "\\mathcal{H}" + }, + { + "bbox": [ + 55, + 398, + 296, + 604 + ], + "type": "text", + "content": " and bases " + }, + { + "bbox": [ + 55, + 398, + 296, + 604 + ], + "type": "inline_equation", + "content": "\\mathcal{M}" + }, + { + "bbox": [ + 55, + 398, + 296, + 604 + ], + "type": "text", + "content": " produces strip-like feature distributions pointing to specific base centers, enhancing contextual distinctiveness with improved linear separability and interpretability." + } + ] + } + ], + "index": 39 + }, + { + "bbox": [ + 55, + 610, + 164, + 623 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 610, + 164, + 623 + ], + "spans": [ + { + "bbox": [ + 55, + 610, + 164, + 623 + ], + "type": "text", + "content": "4.4. Qualitative results" + } + ] + } + ], + "index": 40 + }, + { + "bbox": [ + 55, + 629, + 296, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 629, + 296, + 714 + ], + "spans": [ + { + "bbox": [ + 55, + 629, + 296, + 714 + ], + "type": "text", + "content": "Fig. 5 presents qualitative examples of our model. We compare the predictions of DistinctAD (using LLaMA3-8B) with ground-truth captions (GT) and publicly available AutoAD-Zero [82] outputs. Note that clips are sampled consecutively in time. Previous studies often struggle with similar contextual clips, such as those featuring closely-related scenes and characters, by repeating correct yet in" + } + ] + } + ], + "index": 41 + }, + { + "type": "image", + "bbox": [ + 320, + 320, + 392, + 392 + ], + "blocks": [ + { + "bbox": [ + 55, + 290, + 555, + 313 + ], + "lines": [ + { + "bbox": [ + 55, + 290, + 555, + 313 + ], + "spans": [ + { + "bbox": [ + 55, + 290, + 555, + 313 + ], + "type": "text", + "content": "Figure 5. Qualitative results. We present ground-truth (GT) ADs, publicly released AutoAD-Zero outputs, and our DistinctAD predictions for several temporally consecutive movie clips. Movie frames are taken from The Ides of March (2011) [15]. Zoom in for details." + } + ] + } + ], + "index": 37, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 320, + 320, + 392, + 392 + ], + "lines": [ + { + "bbox": [ + 320, + 320, + 392, + 392 + ], + "spans": [ + { + "bbox": [ + 320, + 320, + 392, + 392 + ], + "type": "image", + "image_path": "29d662c4b957db5ba2e1a973477e675137f4348e9fd76de5e80fde7457094325.jpg" + } + ] + } + ], + "index": 42, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 348, + 392, + 364, + 399 + ], + "lines": [ + { + "bbox": [ + 348, + 392, + 364, + 399 + ], + "spans": [ + { + "bbox": [ + 348, + 392, + 364, + 399 + ], + "type": "text", + "content": "(a) " + }, + { + "bbox": [ + 348, + 392, + 364, + 399 + ], + "type": "inline_equation", + "content": "\\mathcal{H}" + } + ] + } + ], + "index": 43, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 313, + 401, + 555, + 445 + ], + "lines": [ + { + "bbox": [ + 313, + 401, + 555, + 445 + ], + "spans": [ + { + "bbox": [ + 313, + 401, + 555, + 445 + ], + "type": "text", + "content": "Figure 6. Visualizations of Contextual EMA. (a) A set of randomly generated 3D data " + }, + { + "bbox": [ + 313, + 401, + 555, + 445 + ], + "type": "inline_equation", + "content": "\\mathcal{H}" + }, + { + "bbox": [ + 313, + 401, + 555, + 445 + ], + "type": "text", + "content": ", sampled from " + }, + { + "bbox": [ + 313, + 401, + 555, + 445 + ], + "type": "inline_equation", + "content": "N" + }, + { + "bbox": [ + 313, + 401, + 555, + 445 + ], + "type": "text", + "content": " types of samples. (b) Compact features " + }, + { + "bbox": [ + 313, + 401, + 555, + 445 + ], + "type": "inline_equation", + "content": "\\widehat{\\mathcal{H}}" + }, + { + "bbox": [ + 313, + 401, + 555, + 445 + ], + "type": "text", + "content": " obtained via Data Re-estimation (DR). (c) Cross-attention outputs " + }, + { + "bbox": [ + 313, + 401, + 555, + 445 + ], + "type": "inline_equation", + "content": "\\widetilde{\\mathcal{H}}" + }, + { + "bbox": [ + 313, + 401, + 555, + 445 + ], + "type": "text", + "content": " between " + }, + { + "bbox": [ + 313, + 401, + 555, + 445 + ], + "type": "inline_equation", + "content": "\\mathcal{H}" + }, + { + "bbox": [ + 313, + 401, + 555, + 445 + ], + "type": "text", + "content": " and bases " + }, + { + "bbox": [ + 313, + 401, + 555, + 445 + ], + "type": "inline_equation", + "content": "\\mathcal{M}" + }, + { + "bbox": [ + 313, + 401, + 555, + 445 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 48, + "angle": 0, + "type": "image_caption" + } + ], + "index": 42 + }, + { + "type": "image", + "bbox": [ + 399, + 320, + 470, + 392 + ], + "blocks": [ + { + "bbox": [ + 399, + 320, + 470, + 392 + ], + "lines": [ + { + "bbox": [ + 399, + 320, + 470, + 392 + ], + "spans": [ + { + "bbox": [ + 399, + 320, + 470, + 392 + ], + "type": "image", + "image_path": "bc6c45c84779a97c19f85519f04ab16ddccf6912105013878a04de987a39702d.jpg" + } + ] + } + ], + "index": 44, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 427, + 392, + 444, + 400 + ], + "lines": [ + { + "bbox": [ + 427, + 392, + 444, + 400 + ], + "spans": [ + { + "bbox": [ + 427, + 392, + 444, + 400 + ], + "type": "text", + "content": "(b) " + }, + { + "bbox": [ + 427, + 392, + 444, + 400 + ], + "type": "inline_equation", + "content": "\\widehat{\\mathcal{H}}" + } + ] + } + ], + "index": 45, + "angle": 0, + "type": "image_caption" + } + ], + "index": 44 + }, + { + "type": "image", + "bbox": [ + 477, + 321, + 548, + 391 + ], + "blocks": [ + { + "bbox": [ + 477, + 321, + 548, + 391 + ], + "lines": [ + { + "bbox": [ + 477, + 321, + 548, + 391 + ], + "spans": [ + { + "bbox": [ + 477, + 321, + 548, + 391 + ], + "type": "image", + "image_path": "00f00d71e4482d9a9a0525243439f9f672dfdaabb1ce2a4def9aaa4e07519078.jpg" + } + ] + } + ], + "index": 46, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 506, + 392, + 523, + 400 + ], + "lines": [ + { + "bbox": [ + 506, + 392, + 523, + 400 + ], + "spans": [ + { + "bbox": [ + 506, + 392, + 523, + 400 + ], + "type": "text", + "content": "(c) " + }, + { + "bbox": [ + 506, + 392, + 523, + 400 + ], + "type": "inline_equation", + "content": "\\widetilde{\\mathcal{H}}" + } + ] + } + ], + "index": 47, + "angle": 0, + "type": "image_caption" + } + ], + "index": 46 + }, + { + "bbox": [ + 313, + 455, + 555, + 515 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 455, + 555, + 515 + ], + "spans": [ + { + "bbox": [ + 313, + 455, + 555, + 515 + ], + "type": "text", + "content": "significant action words, e.g. \"look\". In contrast, our DistinctAD effectively generates more engaging ADs by identifying distinctive objects in adjacent clips, e.g. \"phone\", \"pill bottle\", and \"car\", along with corresponding more specific behaviors. More examples can be found in Appendix §E." + } + ] + } + ], + "index": 49 + }, + { + "bbox": [ + 313, + 525, + 388, + 538 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 525, + 388, + 538 + ], + "spans": [ + { + "bbox": [ + 313, + 525, + 388, + 538 + ], + "type": "text", + "content": "5. Conclusion" + } + ] + } + ], + "index": 50 + }, + { + "bbox": [ + 313, + 546, + 555, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 546, + 555, + 713 + ], + "spans": [ + { + "bbox": [ + 313, + 546, + 555, + 713 + ], + "type": "text", + "content": "This paper proposes DistinctAD, a two-stage framework for generating distinctive audio descriptions for better narrative. By addressing the domain gap between movie-AD data with a CLIP-AD adaptation strategy, and introducing a Contextual EMA module and a distinctive word prediction loss, our approach significantly improves the quality of AD generation. The effectiveness of DistinctAD is demonstrated through comprehensive evaluations on multiple benchmark datasets and ablations studies. Despite these promising results, DistinctAD is still limited by requiring numbers of parameters and the quality of the generated ADs still falls short of human annotations. Overall, automatic AD generation still remains a challenging task, and there is considerable scope for future advancements in this field." + } + ] + } + ], + "index": 51 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "text", + "content": "13578" + } + ] + } + ], + "index": 52 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 7 + }, + { + "para_blocks": [ + { + "bbox": [ + 56, + 72, + 153, + 85 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 72, + 153, + 85 + ], + "spans": [ + { + "bbox": [ + 56, + 72, + 153, + 85 + ], + "type": "text", + "content": "Acknowledgement" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 55, + 90, + 296, + 114 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 90, + 296, + 114 + ], + "spans": [ + { + "bbox": [ + 55, + 90, + 296, + 114 + ], + "type": "text", + "content": "This work was supported by a Strategic Research Grant from City University of Hong Kong (Project No. 7005840)." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 56, + 134, + 115, + 146 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 134, + 115, + 146 + ], + "spans": [ + { + "bbox": [ + 56, + 134, + 115, + 146 + ], + "type": "text", + "content": "References" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 57, + 155, + 295, + 713 + ], + "type": "list", + "angle": 0, + "index": 19, + "blocks": [ + { + "bbox": [ + 61, + 155, + 230, + 165 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 61, + 155, + 230, + 165 + ], + "spans": [ + { + "bbox": [ + 61, + 155, + 230, + 165 + ], + "type": "text", + "content": "[1] https://audiovault.net, 2019.3" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 62, + 167, + 293, + 187 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 167, + 293, + 187 + ], + "spans": [ + { + "bbox": [ + 62, + 167, + 293, + 187 + ], + "type": "text", + "content": "[2] Vlog: Video as a long document. https://github.com/showlab/VLog, 2023. GitHub repository. 6" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 62, + 189, + 295, + 242 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 189, + 295, + 242 + ], + "spans": [ + { + "bbox": [ + 62, + 189, + 295, + 242 + ], + "type": "text", + "content": "[3] Josh Achiam, Steven Adler, Sandhini Agarwal, Lama Ahmad, Ilge Akkaya, Florencia Leoni Aleman, Diogo Almeida, Janko Altenschmidt, Sam Altman, Shyamal Anadkat, et al. Gpt-4 technical report. arXiv preprint arXiv:2303.08774, 2023. 1, 2, 6" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 62, + 244, + 243, + 255 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 244, + 243, + 255 + ], + "spans": [ + { + "bbox": [ + 62, + 244, + 243, + 255 + ], + "type": "text", + "content": "[4] AI@Meta. Llama 3 model card. 2024. 5, 6, 2" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 62, + 256, + 295, + 309 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 256, + 295, + 309 + ], + "spans": [ + { + "bbox": [ + 62, + 256, + 295, + 309 + ], + "type": "text", + "content": "[5] Jean-Baptiste Alayrac, Jeff Donahue, Pauline Luc, Antoine Miech, Iain Barr, Yana Hasson, Karel Lenc, Arthur Mensch, Katherine Millican, Malcolm Reynolds, et al. Flamingo: a visual language model for few-shot learning. NeurIPS, 35: 23716-23736, 2022. 1, 3, 4" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 62, + 312, + 295, + 354 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 312, + 295, + 354 + ], + "spans": [ + { + "bbox": [ + 62, + 312, + 295, + 354 + ], + "type": "text", + "content": "[6] Peter Anderson, Basura Fernando, Mark Johnson, and Stephen Gould. Spice: Semantic propositional image caption evaluation. In ECCV, pages 382-398. Springer, 2016. 5" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 62, + 357, + 295, + 388 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 357, + 295, + 388 + ], + "spans": [ + { + "bbox": [ + 62, + 357, + 295, + 388 + ], + "type": "text", + "content": "[7] Max Bain, Arsha Nagrani, Andrew Brown, and Andrew Zisserman. Condensed movies: Story based retrieval with contextual embeddings. In ACCV, 2020. 5" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 62, + 390, + 295, + 422 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 390, + 295, + 422 + ], + "spans": [ + { + "bbox": [ + 62, + 390, + 295, + 422 + ], + "type": "text", + "content": "[8] Max Bain, Jaesung Huh, Tengda Han, and Andrew Zisserman. Whisperx: Time-accurate speech transcription of long-form audio. arXiv preprint arXiv:2303.00747, 2023. 2" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 62, + 424, + 295, + 466 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 424, + 295, + 466 + ], + "spans": [ + { + "bbox": [ + 62, + 424, + 295, + 466 + ], + "type": "text", + "content": "[9] Carmen J Branje and Deborah I Fels. Livedescribe: can amateur descriptors create high-quality audio description? Journal of Visual Impairment & Blindness, 106(3):154-165, 2012. 1, 2" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 57, + 468, + 295, + 499 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 468, + 295, + 499 + ], + "spans": [ + { + "bbox": [ + 57, + 468, + 295, + 499 + ], + "type": "text", + "content": "[10] Hervé Bredin and Antoine Laurent. End-to-end speaker segmentation for overlap-aware resegmentation. In Interspeech, 2021. 2" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 57, + 502, + 295, + 555 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 502, + 295, + 555 + ], + "spans": [ + { + "bbox": [ + 57, + 502, + 295, + 555 + ], + "type": "text", + "content": "[11] Hervé Bredin, Ruiqing Yin, Juan Manuel Coria, Gregory Gelly, Pavel Korshunov, Marvin Lavechin, Diego Fustes, Hadrien Titeux, Wassim Bouaziz, and Marie-Philippe Gill. Pyannote. audio: neural building blocks for speaker diarization. In ICASSP, pages 7124–7128. IEEE, 2020. 2" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 57, + 557, + 295, + 601 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 557, + 295, + 601 + ], + "spans": [ + { + "bbox": [ + 57, + 557, + 295, + 601 + ], + "type": "text", + "content": "[12] Fuhai Chen, Rongrong Ji, Xiaoshuai Sun, Yongjian Wu, and Jinsong Su. Groupcap: Group-based image captioning with structured relevance and diversity constraints. In CVPR, pages 1345-1353, 2018. 3" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 57, + 602, + 295, + 634 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 602, + 295, + 634 + ], + "spans": [ + { + "bbox": [ + 57, + 602, + 295, + 634 + ], + "type": "text", + "content": "[13] Shaoxiang Chen and Yu-Gang Jiang. Towards bridging event captioner and sentence localizer for weakly supervised dense event captioning. In CVPR, pages 8425-8435, 2021. 2" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 57, + 635, + 295, + 667 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 635, + 295, + 667 + ], + "spans": [ + { + "bbox": [ + 57, + 635, + 295, + 667 + ], + "type": "text", + "content": "[14] Peng Chu, Jiang Wang, and Andre Abrantes. Llm-ad: Large language model based audio description system. arXiv preprint arXiv:2405.00983, 2024. 1, 2, 6" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 57, + 670, + 295, + 689 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 670, + 295, + 689 + ], + "spans": [ + { + "bbox": [ + 57, + 670, + 295, + 689 + ], + "type": "text", + "content": "[15] George Clooney. The idees of march. *Columbia Pictures*, 2011. 8" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 57, + 691, + 295, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 691, + 295, + 713 + ], + "spans": [ + { + "bbox": [ + 57, + 691, + 295, + 713 + ], + "type": "text", + "content": "[16] Bo Dai and Dahua Lin. Contrastive learning for image captioning. NeurIPS, 30, 2017. 3" + } + ] + } + ], + "index": 18 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 316, + 73, + 554, + 712 + ], + "type": "list", + "angle": 0, + "index": 37, + "blocks": [ + { + "bbox": [ + 316, + 73, + 554, + 116 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 73, + 554, + 116 + ], + "spans": [ + { + "bbox": [ + 316, + 73, + 554, + 116 + ], + "type": "text", + "content": "[17] Arthur P Dempster, Nan M Laird, and Donald B Rubin. Maximum likelihood from incomplete data via the em algorithm. Journal of the royal statistical society: series B (methodological), 39(1):1-22, 1977. 2, 4" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 317, + 118, + 553, + 149 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 118, + 553, + 149 + ], + "spans": [ + { + "bbox": [ + 317, + 118, + 553, + 149 + ], + "type": "text", + "content": "[18] Chaorui Deng, Shizhe Chen, Da Chen, Yuan He, and Qi Wu. Sketch, ground, and refine: Top-down dense video captioning. In CVPR, pages 234-243, 2021. 2" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 317, + 151, + 553, + 203 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 151, + 553, + 203 + ], + "spans": [ + { + "bbox": [ + 317, + 151, + 553, + 203 + ], + "type": "text", + "content": "[19] Lisa Dunlap, Yuhui Zhang, Xiaohan Wang, Ruiqi Zhong, Trevor Darrell, Jacob Steinhardt, Joseph E Gonzalez, and Serena Yeung-Levy. Describing differences in image sets with natural language. In CVPR, pages 24199-24208, 2024. 3" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 316, + 206, + 553, + 227 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 206, + 553, + 227 + ], + "spans": [ + { + "bbox": [ + 316, + 206, + 553, + 227 + ], + "type": "text", + "content": "[20] Louise Fryer. An introduction to audio description: A practical guide, 2016. 1" + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 317, + 228, + 553, + 270 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 228, + 553, + 270 + ], + "spans": [ + { + "bbox": [ + 317, + 228, + 553, + 270 + ], + "type": "text", + "content": "[21] Tengda Han, Max Bain, Arsha Nagrani, Gül Varol, Weidi Xie, and Andrew Zisserman. Autoad: Movie description in context. In CVPR, pages 18930-18940, 2023. 1, 2, 3, 4, 5, 6, 7" + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 316, + 272, + 553, + 315 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 272, + 553, + 315 + ], + "spans": [ + { + "bbox": [ + 316, + 272, + 553, + 315 + ], + "type": "text", + "content": "[22] Tengda Han, Max Bain, Arsha Nagrani, Gul Varol, Weidi Xie, and Andrew Zisserman. Autoad ii: The sequel-who, when, and what in movie audio description. In ICCV, pages 13645–13655, 2023. 1, 3, 4, 5, 6, 7" + } + ] + } + ], + "index": 25 + }, + { + "bbox": [ + 317, + 316, + 553, + 358 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 316, + 553, + 358 + ], + "spans": [ + { + "bbox": [ + 317, + 316, + 553, + 358 + ], + "type": "text", + "content": "[23] Tengda Han, Max Bain, Arsha Nagrani, Gül Varol, Weidi Xie, and Andrew Zisserman. Autoad iii: The prequel-back to the pixels. In CVPR, pages 18164-18174, 2024. 1, 2, 3, 5, 6, 7" + } + ] + } + ], + "index": 26 + }, + { + "bbox": [ + 317, + 360, + 553, + 403 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 360, + 553, + 403 + ], + "spans": [ + { + "bbox": [ + 317, + 360, + 553, + 403 + ], + "type": "text", + "content": "[24] Edward J Hu, Yelong Shen, Phillip Wallis, Zeyuan Allen-Zhu, Yuanzhi Li, Shean Wang, Lu Wang, and Weizhu Chen. Lora: Low-rank adaptation of large language models. arXiv preprint arXiv:2106.09685, 2021. 6" + } + ] + } + ], + "index": 27 + }, + { + "bbox": [ + 317, + 404, + 553, + 436 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 404, + 553, + 436 + ], + "spans": [ + { + "bbox": [ + 317, + 404, + 553, + 436 + ], + "type": "text", + "content": "[25] Vladimir Iashin and Esa Rahtu. A better use of audio-visual cues: Dense video captioning with bi-modal transformer. In BMVC, 2020. 2" + } + ] + } + ], + "index": 28 + }, + { + "bbox": [ + 316, + 437, + 553, + 459 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 437, + 553, + 459 + ], + "spans": [ + { + "bbox": [ + 316, + 437, + 553, + 459 + ], + "type": "text", + "content": "[26] Vladimir Iashin and Esa Rahtu. Multi-modal dense video captioning. In CVPRW, pages 958-959, 2020. 2" + } + ] + } + ], + "index": 29 + }, + { + "bbox": [ + 317, + 460, + 553, + 513 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 460, + 553, + 513 + ], + "spans": [ + { + "bbox": [ + 317, + 460, + 553, + 513 + ], + "type": "text", + "content": "[27] Peng Jin, Jina Huang, Fenglin Liu, Xian Wu, Shen Ge, Guoli Song, David Clifton, and Jie Chen. Expectation-maximization contrastive learning for compact video-and-language representations. NeurIPS, 35:30291-30306, 2022. 2" + } + ] + } + ], + "index": 30 + }, + { + "bbox": [ + 316, + 515, + 553, + 536 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 515, + 553, + 536 + ], + "spans": [ + { + "bbox": [ + 316, + 515, + 553, + 536 + ], + "type": "text", + "content": "[28] Diederik P Kingma. Adam: A method for stochastic optimization. arXiv preprint arXiv:1412.6980, 2014. 6" + } + ] + } + ], + "index": 31 + }, + { + "bbox": [ + 317, + 537, + 553, + 568 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 537, + 553, + 568 + ], + "spans": [ + { + "bbox": [ + 317, + 537, + 553, + 568 + ], + "type": "text", + "content": "[29] Ranjay Krishna, Kenji Hata, Frederic Ren, Li Fei-Fei, and Juan Carlos Niebles. Dense-captioning events in videos. In ICCV, pages 706–715, 2017. 2" + } + ] + } + ], + "index": 32 + }, + { + "bbox": [ + 317, + 570, + 553, + 601 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 570, + 553, + 601 + ], + "spans": [ + { + "bbox": [ + 317, + 570, + 553, + 601 + ], + "type": "text", + "content": "[30] Jie Lei, Licheng Yu, Mohit Bansal, and Tamara L Berg. Tvqa: Localized, compositional video question answering. In EMNLP, 2018. 5" + } + ] + } + ], + "index": 33 + }, + { + "bbox": [ + 316, + 603, + 553, + 624 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 603, + 553, + 624 + ], + "spans": [ + { + "bbox": [ + 316, + 603, + 553, + 624 + ], + "type": "text", + "content": "[31] Elisa Lewis. Deep dive: How audio description benefits everyone, 2021. Accessed on, pages 11-13, 2023. 1" + } + ] + } + ], + "index": 34 + }, + { + "bbox": [ + 317, + 625, + 553, + 667 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 625, + 553, + 667 + ], + "spans": [ + { + "bbox": [ + 317, + 625, + 553, + 667 + ], + "type": "text", + "content": "[32] Bo Li, Yuanhan Zhang, Liangyu Chen, Jinghao Wang, Fanyi Pu, Jingkang Yang, Chunyuan Li, and Ziwei Liu. Mimicit: Multi-modal in-context instruction tuning. arXiv preprint arXiv:2306.05425, 2023. 3" + } + ] + } + ], + "index": 35 + }, + { + "bbox": [ + 316, + 670, + 553, + 712 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 670, + 553, + 712 + ], + "spans": [ + { + "bbox": [ + 316, + 670, + 553, + 712 + ], + "type": "text", + "content": "[33] Bo Li, Yuanhan Zhang, Liangyu Chen, Jinghao Wang, Jingkang Yang, and Ziwei Liu. Otter: a multi-modal model with in-context instruction tuning. corr abs/2305.03726 (2023), 2023. 3" + } + ] + } + ], + "index": 36 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "type": "text", + "content": "13579" + } + ] + } + ], + "index": 38 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 8 + }, + { + "para_blocks": [ + { + "bbox": [ + 56, + 72, + 295, + 712 + ], + "type": "list", + "angle": 0, + "index": 15, + "blocks": [ + { + "bbox": [ + 56, + 72, + 294, + 116 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 72, + 294, + 116 + ], + "spans": [ + { + "bbox": [ + 56, + 72, + 294, + 116 + ], + "type": "text", + "content": "[34] Junnan Li, Dongxu Li, Silvio Savarese, and Steven Hoi. Blip-2: Bootstrapping language-image pre-training with frozen image encoders and large language models. arXiv preprint arXiv:2301.12597, 2023. 1, 3" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 56, + 118, + 295, + 160 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 118, + 295, + 160 + ], + "spans": [ + { + "bbox": [ + 56, + 118, + 295, + 160 + ], + "type": "text", + "content": "[35] Xia Li, Zhisheng Zhong, Jianlong Wu, Yibo Yang, Zhouchen Lin, and Hong Liu. Expectation-maximization attention networks for semantic segmentation. In ICCV, pages 9167-9176, 2019. 2, 4" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 56, + 162, + 294, + 195 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 162, + 294, + 195 + ], + "spans": [ + { + "bbox": [ + 56, + 162, + 294, + 195 + ], + "type": "text", + "content": "[36] Yehao Li, Ting Yao, Yingwei Pan, Hongyang Chao, and Tao Mei. Jointly localizing and describing events for dense video captioning. In CVPR, pages 7492-7500, 2018. 2" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 56, + 196, + 294, + 226 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 196, + 294, + 226 + ], + "spans": [ + { + "bbox": [ + 56, + 196, + 294, + 226 + ], + "type": "text", + "content": "[37] Chin-Yew Lin. Rouge: A package for automatic evaluation of summaries. In Text summarization branches out, pages 74-81, 2004. 5" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 56, + 228, + 294, + 271 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 228, + 294, + 271 + ], + "spans": [ + { + "bbox": [ + 56, + 228, + 294, + 271 + ], + "type": "text", + "content": "[38] Kevin Lin, Linjie Li, Chung-Ching Lin, Faisal Ahmed, Zhe Gan, Zicheng Liu, Yumao Lu, and Lijuan Wang. Swinbert: End-to-end transformers with sparse attention for video captioning. In CVPR, pages 17949-17958, 2022. 2" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 56, + 272, + 294, + 326 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 272, + 294, + 326 + ], + "spans": [ + { + "bbox": [ + 56, + 272, + 294, + 326 + ], + "type": "text", + "content": "[39] Kevin Lin, Faisal Ahmed, Linjie Li, Chung-Ching Lin, Ehsan Azarnasab, Zhengyuan Yang, Jianfeng Wang, Lin Liang, Zicheng Liu, Yumao Lu, et al. Mm-vid: Advancing video understanding with gpt-4v (ision). arXiv preprint arXiv:2310.19773, 2023. 1, 2, 6" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 56, + 327, + 294, + 380 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 327, + 294, + 380 + ], + "spans": [ + { + "bbox": [ + 56, + 327, + 294, + 380 + ], + "type": "text", + "content": "[40] Kevin Qinghong Lin, Pengchuan Zhang, Difei Gao, Xide Xia, Joya Chen, Ziteng Gao, Jinheng Xie, Xuhong Xiao, and Mike Zheng Shou. Learning video context as interleaved multimodal sequences. arXiv preprint arXiv:2407.21757, 2024. 1, 3, 6" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 56, + 382, + 294, + 435 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 382, + 294, + 435 + ], + "spans": [ + { + "bbox": [ + 56, + 382, + 294, + 435 + ], + "type": "text", + "content": "[41] Zhihui Lin, Tianyu Yang, Maomao Li, Ziyu Wang, Chun Yuan, Wenhao Jiang, and Wei Liu. Swem: Towards real-time video object segmentation with sequential weighted expectation-maximization. In CVPR, pages 1362-1372, 2022. 2" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 56, + 437, + 294, + 459 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 437, + 294, + 459 + ], + "spans": [ + { + "bbox": [ + 56, + 437, + 294, + 459 + ], + "type": "text", + "content": "[42] Haotian Liu, Chunyuan Li, Qingyang Wu, and Yong Jae Lee. Visual instruction tuning. NeurIPS, 36, 2024. 1" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 56, + 460, + 294, + 502 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 460, + 294, + 502 + ], + "spans": [ + { + "bbox": [ + 56, + 460, + 294, + 502 + ], + "type": "text", + "content": "[43] Xihui Liu, Hongsheng Li, Jing Shao, Dapeng Chen, and Xiaogang Wang. Show, tell and discriminate: Image captioning by self-retrieval with partially labeled data. In ECCV, pages 338-354, 2018. 3" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 56, + 504, + 294, + 525 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 504, + 294, + 525 + ], + "spans": [ + { + "bbox": [ + 56, + 504, + 294, + 525 + ], + "type": "text", + "content": "[44] I Loshchilov. Decoupled weight decay regularization. arXiv preprint arXiv:1711.05101, 2017. 6" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 56, + 526, + 294, + 580 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 526, + 294, + 580 + ], + "spans": [ + { + "bbox": [ + 56, + 526, + 294, + 580 + ], + "type": "text", + "content": "[45] Huaishao Luo, Lei Ji, Botian Shi, Haoyang Huang, Nan Duan, Tianrui Li, Jason Li, Taroon Bharti, and Ming Zhou. Univl: A unified video and language pre-training model for multimodal understanding and generation. arXiv preprint arXiv:2002.06353, 2020. 2" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 56, + 582, + 294, + 624 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 582, + 294, + 624 + ], + "spans": [ + { + "bbox": [ + 56, + 582, + 294, + 624 + ], + "type": "text", + "content": "[46] Huaishao Luo, Lei Ji, Ming Zhong, Yang Chen, Wen Lei, Nan Duan, and Tianrui Li. Clip4clip: An empirical study of clip for end to end video clip retrieval and captioning. Neurocomputing, 508:293-304, 2022. 3" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 56, + 625, + 294, + 658 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 625, + 294, + 658 + ], + "spans": [ + { + "bbox": [ + 56, + 625, + 294, + 658 + ], + "type": "text", + "content": "[47] Ruotian Luo, Brian Price, Scott Cohen, and Gregory Shakhnarovich. Discriminability objective for training descriptive captions. In CVPR, pages 6964-6974, 2018. 3" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 56, + 658, + 294, + 712 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 658, + 294, + 712 + ], + "spans": [ + { + "bbox": [ + 56, + 658, + 294, + 712 + ], + "type": "text", + "content": "[48] Antoine Miech, Dimitri Zhukov, Jean-Baptiste Alayrac, Makarand Tapaswi, Ivan Laptev, and Josef Sivic. Howto100m: Learning a text-video embedding by watching hundred million narrated video clips. In ICCV, pages 2630-2640, 2019. 6, 7" + } + ] + } + ], + "index": 14 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 315, + 72, + 553, + 714 + ], + "type": "list", + "angle": 0, + "index": 33, + "blocks": [ + { + "bbox": [ + 315, + 72, + 553, + 116 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 315, + 72, + 553, + 116 + ], + "spans": [ + { + "bbox": [ + 315, + 72, + 553, + 116 + ], + "type": "text", + "content": "[49] Antoine Miech, Jean-Baptiste Alayrac, Lucas Smaira, Ivan Laptev, Josef Sivic, and Andrew Zisserman. End-to-end learning of visual representations from uncurated instructional videos. In CVPR, pages 9879-9889, 2020. 4" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 315, + 118, + 553, + 150 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 315, + 118, + 553, + 150 + ], + "spans": [ + { + "bbox": [ + 315, + 118, + 553, + 150 + ], + "type": "text", + "content": "[50] Ron Mokady, Amir Hertz, and Amit H Bermano. Clipcap: Clip prefix for image captioning. arXiv preprint arXiv:2111.09734, 2021. 6" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 315, + 152, + 553, + 184 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 315, + 152, + 553, + 184 + ], + "spans": [ + { + "bbox": [ + 315, + 152, + 553, + 184 + ], + "type": "text", + "content": "[51] Jonghwan Mun, Linjie Yang, Zhou Ren, Ning Xu, and Bohyung Han. Streamlined dense video captioning. In CVPR, pages 6588-6597, 2019. 2" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 315, + 186, + 553, + 218 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 315, + 186, + 553, + 218 + ], + "spans": [ + { + "bbox": [ + 315, + 186, + 553, + 218 + ], + "type": "text", + "content": "[52] David Nukrai, Ron Mokady, and Amir Globerson. Text-only training for image captioning using noise-injected clip. arXiv preprint arXiv:2211.00575, 2022. 6" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 315, + 220, + 516, + 230 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 315, + 220, + 516, + 230 + ], + "spans": [ + { + "bbox": [ + 315, + 220, + 516, + 230 + ], + "type": "text", + "content": "[53] OpenAI. Gpt-4v(ision) system card. 2023. 1, 2, 6" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 315, + 232, + 553, + 284 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 315, + 232, + 553, + 284 + ], + "spans": [ + { + "bbox": [ + 315, + 232, + 553, + 284 + ], + "type": "text", + "content": "[54] Amy Pavel, Gabriel Reyes, and Jeffrey P Bigham. Describe: Authoring and automatically editing audio descriptions. In Proceedings of the 33rd Annual ACM Symposium on User Interface Software and Technology, pages 747-759, 2020. 1, 2" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 315, + 287, + 553, + 308 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 315, + 287, + 553, + 308 + ], + "spans": [ + { + "bbox": [ + 315, + 287, + 553, + 308 + ], + "type": "text", + "content": "[55] Elisa Perego. Gains and losses of watching audio described films for sighted viewers. *Target*, 28(3):424-444, 2016. 1" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 315, + 310, + 553, + 353 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 315, + 310, + 553, + 353 + ], + "spans": [ + { + "bbox": [ + 315, + 310, + 553, + 353 + ], + "type": "text", + "content": "[56] Haran Raajesh, Naveen Reddy Desanur, Zeeshan Khan, and Makarand Tapaswi. Micap: A unified model for identity-aware movie descriptions. In CVPR, pages 14011-14021, 2024. 1, 4" + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 315, + 355, + 553, + 388 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 315, + 355, + 553, + 388 + ], + "spans": [ + { + "bbox": [ + 315, + 355, + 553, + 388 + ], + "type": "text", + "content": "[57] Alec Radford, Jeffrey Wu, Rewon Child, David Luan, Dario Amodei, Ilya Sutskever, et al. Language models are unsupervised multitask learners. OpenAI blog, 1(8):9, 2019. 2" + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 315, + 388, + 553, + 443 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 315, + 388, + 553, + 443 + ], + "spans": [ + { + "bbox": [ + 315, + 388, + 553, + 443 + ], + "type": "text", + "content": "[58] Alec Radford, Jong Wook Kim, Chris Hallacy, Aditya Ramesh, Gabriel Goh, Sandhini Agarwal, Girish Sastry, Amanda Askell, Pamela Mishkin, Jack Clark, et al. Learning transferable visual models from natural language supervision. In ICML, pages 8748-8763. PMLR, 2021. 2" + } + ] + } + ], + "index": 25 + }, + { + "bbox": [ + 315, + 445, + 553, + 477 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 315, + 445, + 553, + 477 + ], + "spans": [ + { + "bbox": [ + 315, + 445, + 553, + 477 + ], + "type": "text", + "content": "[59] Tanzila Rahman, Bicheng Xu, and Leonid Sigal. Watch, listen and tell: Multi-modal weakly supervised dense event captioning. In ICCV, pages 8908-8917, 2019. 2" + } + ] + } + ], + "index": 26 + }, + { + "bbox": [ + 315, + 478, + 553, + 510 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 315, + 478, + 553, + 510 + ], + "spans": [ + { + "bbox": [ + 315, + 478, + 553, + 510 + ], + "type": "text", + "content": "[60] Anna Rohrbach, Marcus Rohrbach, Niket Tandon, and Bernt Schiele. A dataset for movie description. In CVPR, pages 3202-3212, 2015. 2" + } + ] + } + ], + "index": 27 + }, + { + "bbox": [ + 315, + 512, + 553, + 555 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 315, + 512, + 553, + 555 + ], + "spans": [ + { + "bbox": [ + 315, + 512, + 553, + 555 + ], + "type": "text", + "content": "[61] Anna Rohrbach, Atousa Torabi, Marcus Rohrbach, Niket Tandon, Christopher Pal, Hugo Larochelle, Aaron Courville, and Bernt Schiele. Movie description. *IJCV*, 123:94–120, 2017. 2" + } + ] + } + ], + "index": 28 + }, + { + "bbox": [ + 315, + 556, + 553, + 600 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 315, + 556, + 553, + 600 + ], + "spans": [ + { + "bbox": [ + 315, + 556, + 553, + 600 + ], + "type": "text", + "content": "[62] Paul Hongsuck Seo, Arsha Nagrani, Anurag Arnab, and Cordelia Schmid. End-to-end generative pretraining for multimodal video captioning. In CVPR, pages 17959-17968, 2022. 2" + } + ] + } + ], + "index": 29 + }, + { + "bbox": [ + 315, + 601, + 553, + 645 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 315, + 601, + 553, + 645 + ], + "spans": [ + { + "bbox": [ + 315, + 601, + 553, + 645 + ], + "type": "text", + "content": "[63] Piyush Sharma, Nan Ding, Sebastian Goodman, and Radu Soricut. Conceptual captions: A cleaned, hypernymed, image alt-text dataset for automatic image captioning. In ACL, pages 2556-2565, 2018. 2" + } + ] + } + ], + "index": 30 + }, + { + "bbox": [ + 315, + 647, + 553, + 689 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 315, + 647, + 553, + 689 + ], + "spans": [ + { + "bbox": [ + 315, + 647, + 553, + 689 + ], + "type": "text", + "content": "[64] Zhiqiang Shen, Jianguo Li, Zhou Su, Minjun Li, Yurong Chen, Yu-Gang Jiang, and Xiangyang Xue. Weakly supervised dense video captioning. In CVPR, pages 1916-1924, 2017. 2" + } + ] + } + ], + "index": 31 + }, + { + "bbox": [ + 315, + 691, + 553, + 714 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 315, + 691, + 553, + 714 + ], + "spans": [ + { + "bbox": [ + 315, + 691, + 553, + 714 + ], + "type": "text", + "content": "[65] Botian Shi, Lei Ji, Yaobo Liang, Nan Duan, Peng Chen, Zhendong Niu, and Ming Zhou. Dense procedure captioning" + } + ] + } + ], + "index": 32 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "type": "text", + "content": "13580" + } + ] + } + ], + "index": 34 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 9 + }, + { + "para_blocks": [ + { + "bbox": [ + 56, + 72, + 295, + 712 + ], + "type": "list", + "angle": 0, + "index": 17, + "blocks": [ + { + "bbox": [ + 76, + 72, + 294, + 94 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 76, + 72, + 294, + 94 + ], + "spans": [ + { + "bbox": [ + 76, + 72, + 294, + 94 + ], + "type": "text", + "content": "in narrated instructional videos. In ACL, pages 6382-6391, 2019. 2" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 56, + 95, + 295, + 138 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 95, + 295, + 138 + ], + "spans": [ + { + "bbox": [ + 56, + 95, + 295, + 138 + ], + "type": "text", + "content": "[66] Aleksandar Shtedritski, Christian Rupprecht, and Andrea Vedaldi. What does clip know about a red circle? visual prompt engineering for vlms. In ICCV, pages 11987-11997, 2023. 3" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 56, + 140, + 294, + 171 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 140, + 294, + 171 + ], + "spans": [ + { + "bbox": [ + 56, + 140, + 294, + 171 + ], + "type": "text", + "content": "[67] Joel Snyder. Audio description: The visual made verbal. In International congress series, pages 935-939. Elsevier, 2005. 1" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 56, + 172, + 294, + 227 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 172, + 294, + 227 + ], + "spans": [ + { + "bbox": [ + 56, + 172, + 294, + 227 + ], + "type": "text", + "content": "[68] Mattia Soldan, Alejandro Pardo, Juan León Alcazar, Fabian Caba, Chen Zhao, Silvio Giancola, and Bernard Ghanem. Mad: A scalable dataset for language grounding in videos from movie audio descriptions. In CVPR, pages 5026-5035, 2022. 2, 5, 1, 4" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 56, + 228, + 294, + 270 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 228, + 294, + 270 + ], + "spans": [ + { + "bbox": [ + 56, + 228, + 294, + 270 + ], + "type": "text", + "content": "[69] Atousa Torabi, Christopher Pal, Hugo Larochelle, and Aaron Courville. Using descriptive video services to create a large data source for video annotation research. arXiv preprint arXiv:1503.01070, 2015. 2" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 56, + 271, + 294, + 326 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 271, + 294, + 326 + ], + "spans": [ + { + "bbox": [ + 56, + 271, + 294, + 326 + ], + "type": "text", + "content": "[70] Hugo Touvron, Louis Martin, Kevin Stone, Peter Albert, Amjad Almahairi, Yasmine Babaei, Nikolay Bashlykov, Soumya Batra, Prajjwal Bhargava, Shruti Bhosale, et al. Llama 2: Open foundation and fine-tuned chat models. arXiv preprint arXiv:2307.09288, 2023. 5" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 56, + 327, + 294, + 349 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 327, + 294, + 349 + ], + "spans": [ + { + "bbox": [ + 56, + 327, + 294, + 349 + ], + "type": "text", + "content": "[71] Laurens Van der Maaten and Geoffrey Hinton. Visualizing data using t-sne. JMLR, 9(11), 2008. 8" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 56, + 350, + 294, + 381 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 350, + 294, + 381 + ], + "spans": [ + { + "bbox": [ + 56, + 350, + 294, + 381 + ], + "type": "text", + "content": "[72] Ramakrishna Vedantam, C Lawrence Zitnick, and Devi Parikh. Cider: Consensus-based image description evaluation. In CVPR, pages 4566-4575, 2015. 5" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 56, + 382, + 294, + 415 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 382, + 294, + 415 + ], + "spans": [ + { + "bbox": [ + 56, + 382, + 294, + 415 + ], + "type": "text", + "content": "[73] Gilad Vered, Gal Oren, Yuval Atzmon, and Gal Chechik. Joint optimization for cooperative image captioning. In ICCV, pages 8898-8907, 2019. 3" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 56, + 415, + 294, + 458 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 415, + 294, + 458 + ], + "spans": [ + { + "bbox": [ + 56, + 415, + 294, + 458 + ], + "type": "text", + "content": "[74] Hanlin Wang, Zhan Tong, Kecheng Zheng, Yujun Shen, and Limin Wang. Contextual ad narration with interleaved multimodal sequence. arXiv preprint arXiv:2403.12922, 2024. 1, 2, 3, 4, 5" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 56, + 460, + 294, + 502 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 460, + 294, + 502 + ], + "spans": [ + { + "bbox": [ + 56, + 460, + 294, + 502 + ], + "type": "text", + "content": "[75] Jingwen Wang, Wenhao Jiang, Lin Ma, Wei Liu, and Yong Xu. Bidirectional attentive fusion with context gating for dense video captioning. In CVPR, pages 7190-7198, 2018. 2" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 56, + 503, + 294, + 547 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 503, + 294, + 547 + ], + "spans": [ + { + "bbox": [ + 56, + 503, + 294, + 547 + ], + "type": "text", + "content": "[76] Jiuniu Wang, Wenjia Xu, Qingzhong Wang, and Antoni B Chan. Compare and reweight: Distinctive image captioning using similar images sets. In ECCV, pages 370-386. Springer, 2020. 3" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 56, + 548, + 294, + 581 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 548, + 294, + 581 + ], + "spans": [ + { + "bbox": [ + 56, + 548, + 294, + 581 + ], + "type": "text", + "content": "[77] Jiuniu Wang, Wenjia Xu, Qingzhong Wang, and Antoni B Chan. Group-based distinctive image captioning with memory attention. In ACMMM, pages 5020-5028, 2021." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 56, + 582, + 294, + 613 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 582, + 294, + 613 + ], + "spans": [ + { + "bbox": [ + 56, + 582, + 294, + 613 + ], + "type": "text", + "content": "[78] Jiuniu Wang, Wenjia Xu, Qingzhong Wang, and Antoni B Chan. On distinctive image captioning via comparing and reweighting. TPAMI, 45(2):2088-2103, 2022. 3" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 56, + 614, + 294, + 647 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 614, + 294, + 647 + ], + "spans": [ + { + "bbox": [ + 56, + 614, + 294, + 647 + ], + "type": "text", + "content": "[79] Teng Wang, Huicheng Zheng, Mingjing Yu, Qian Tian, and Haifeng Hu. Event-centric hierarchical representation for dense video captioning. TCVT, 31(5):1890-1900, 2020. 2" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 56, + 647, + 294, + 680 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 647, + 294, + 680 + ], + "spans": [ + { + "bbox": [ + 56, + 647, + 294, + 680 + ], + "type": "text", + "content": "[80] Teng Wang, Ruimao Zhang, Zhichao Lu, Feng Zheng, Ran Cheng, and Ping Luo. End-to-end dense video captioning with parallel decoding. In ICCV, pages 6847-6857, 2021. 2" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 56, + 681, + 294, + 712 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 681, + 294, + 712 + ], + "spans": [ + { + "bbox": [ + 56, + 681, + 294, + 712 + ], + "type": "text", + "content": "[81] Xiaolong Wang, Ross Girshick, Abhinav Gupta, and Kaiming He. Non-local neural networks. In CVPR, pages 7794-7803, 2018. 4" + } + ] + } + ], + "index": 16 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 316, + 72, + 553, + 376 + ], + "type": "list", + "angle": 0, + "index": 26, + "blocks": [ + { + "bbox": [ + 316, + 72, + 553, + 117 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 72, + 553, + 117 + ], + "spans": [ + { + "bbox": [ + 316, + 72, + 553, + 117 + ], + "type": "text", + "content": "[82] Junyu Xie, Tengda Han, Max Bain, Arsha Nagrani, Güll Varol, Weidi Xie, and Andrew Zisserman. Autoad-zero: A training-free framework for zero-shot audio description. arXiv preprint arXiv:2407.15850, 2024. 1, 2, 4, 5, 6, 7, 8" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 316, + 118, + 553, + 173 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 118, + 553, + 173 + ], + "spans": [ + { + "bbox": [ + 316, + 118, + 553, + 173 + ], + "type": "text", + "content": "[83] Antoine Yang, Arsha Nagrani, Paul Hongsuck Seo, Antoine Miech, Jordi Pont-Tuset, Ivan Laptev, Josef Sivic, and Cordelia Schmid. Vid2seq: Large-scale pretraining of a visual language model for dense video captioning. In CVPR, pages 10714-10726, 2023. 2" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 316, + 174, + 553, + 206 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 174, + 553, + 206 + ], + "spans": [ + { + "bbox": [ + 316, + 174, + 553, + 206 + ], + "type": "text", + "content": "[84] Linli Yao, Weiying Wang, and Qin Jin. Image difference captioning with pre-training and contrastive learning. In AAAI, pages 3108-3116, 2022. 3" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 316, + 207, + 479, + 218 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 207, + 479, + 218 + ], + "spans": [ + { + "bbox": [ + 316, + 207, + 479, + 218 + ], + "type": "text", + "content": "[85] Keunwoo Peter Yu. Videoblip, 2023. 6" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 316, + 219, + 553, + 273 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 219, + 553, + 273 + ], + "spans": [ + { + "bbox": [ + 316, + 219, + 553, + 273 + ], + "type": "text", + "content": "[86] Chaoyi Zhang, Kevin Lin, Zhengyuan Yang, Jianfeng Wang, Linjie Li, Chung-Ching Lin, Zicheng Liu, and Lijuan Wang. Mm-narrator: Narrating long-form videos with multimodal in-context learning. In CVPR, pages 13647-13657, 2024. 1, 2, 4, 6" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 316, + 274, + 553, + 308 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 274, + 553, + 308 + ], + "spans": [ + { + "bbox": [ + 316, + 274, + 553, + 308 + ], + "type": "text", + "content": "[87] Hang Zhang, Xin Li, and Lidong Bing. Video-llama: An instruction-tuned audio-visual language model for video understanding. arXiv preprint arXiv:2306.02858, 2023. 6" + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 316, + 309, + 553, + 342 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 309, + 553, + 342 + ], + "spans": [ + { + "bbox": [ + 316, + 309, + 553, + 342 + ], + "type": "text", + "content": "[88] Tianyi Zhang, Varsha Kishore, Felix Wu, Kilian Q Weinberger, and Yoav Artzi. Bertscore: Evaluating text generation with bert. arXiv preprint arXiv:1904.09675, 2019. 5" + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 316, + 343, + 553, + 376 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 343, + 553, + 376 + ], + "spans": [ + { + "bbox": [ + 316, + 343, + 553, + 376 + ], + "type": "text", + "content": "[89] Luowei Zhou, Yingbo Zhou, Jason J Corso, Richard Socher, and Caiming Xiong. End-to-end dense video captioning with masked transformer. In CVPR, pages 8739-8748, 2018. 2" + } + ] + } + ], + "index": 25 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "text", + "content": "13581" + } + ] + } + ], + "index": 27 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 10 + } + ], + "_backend": "vlm", + "_version_name": "2.6.4" +} \ No newline at end of file diff --git a/2025/Distinguish Then Exploit_ Source-free Open Set Domain Adaptation via Weight Barcode Estimation and Sparse Label Assignment/099209ec-4ec9-4836-8000-10ddecb15989_content_list.json b/2025/Distinguish Then Exploit_ Source-free Open Set Domain Adaptation via Weight Barcode Estimation and Sparse Label Assignment/099209ec-4ec9-4836-8000-10ddecb15989_content_list.json new file mode 100644 index 0000000000000000000000000000000000000000..89443b6a23a842953fc37ba41ae9491801e69716 --- /dev/null +++ b/2025/Distinguish Then Exploit_ Source-free Open Set Domain Adaptation via Weight Barcode Estimation and Sparse Label Assignment/099209ec-4ec9-4836-8000-10ddecb15989_content_list.json @@ -0,0 +1,1752 @@ +[ + { + "type": "text", + "text": "Distinguish Then Exploit: Source-free Open Set Domain Adaptation via Weight Barcode Estimation and Sparse Label Assignment", + "text_level": 1, + "bbox": [ + 84, + 128, + 883, + 174 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Weiming Liu $^{1*}$ , Jun Dan $^{2*}$ , Fan Wang $^{2}$ , Xinting Liao $^{2}$ , Junhao Dong $^{3}$ , Hua Yu $^{3}$ , Shunjie Dong $^{4}$ , Lianyong Qi $^{56\\dagger}$", + "bbox": [ + 227, + 200, + 779, + 238 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "1Bytedance Inc. 2Zhejiang University 3Nanyang Technological University 4Shanghai Jiao Tong University 5China University of Petroleum (East China)", + "bbox": [ + 88, + 239, + 921, + 273 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "$^{6}$ Shandong Key Laboratory of Intelligent Oil & Gas Industrial Software", + "bbox": [ + 222, + 273, + 787, + 291 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "lwming95@gmail.com, danjun@zju.edu.cn, fanwang97@zju.edu.cn, xintingliao@zju.edu.cn, junhao003@ntu.edu.sg, yu_hua@ntu.edu.sg, sjdong@sjtu.edu.cn, lianyongqi@gmail.com", + "bbox": [ + 137, + 294, + 870, + 325 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Abstract", + "text_level": 1, + "bbox": [ + 233, + 361, + 313, + 378 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Nowadays, domain adaptation techniques have been widely investigated for knowledge sharing from labeled source domain to unlabeled target domain. However, target domain may include some data samples that belong to unknown categories in real-world scenarios. Moreover, the target domain cannot access the source data samples due to privacy-preserving restrictions. In this paper, we focus on the source-free open set domain adaptation problem which includes two main challenges, i.e., how to distinguish known and unknown target samples and how to exploit useful source information to provide trustworthy pseudo labels for known target samples. Existing approaches that directly apply conventional domain alignment methods could lead to sample mismatch and misclassification in this scenario. To overcome these issues, we propose a Distinguish Then Exploit model (DTE) with two components, i.e., weight barcode estimation and sparse label assignment. Weight barcode estimation first calculates the marginal probability of target samples via partially unbalanced optimal transport, then quantize barcode results to distinguish unknown target samples. Sparse label assignment utilizes sparse sample-label matching via proximal term to fully exploit useful source information. Our empirically study on several datasets shows that DTE outperforms the state-of-the-art models on tackling the source-free open set domain adaptation problem.", + "bbox": [ + 75, + 393, + 472, + 785 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "1. Introduction", + "text_level": 1, + "bbox": [ + 76, + 806, + 209, + 821 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Nowadays, deep learning technique has been successfully adopted to image classification tasks with the help of large-scale labeled datasets [16, 19, 48, 51-54, 57, 65, 66]. How", + "bbox": [ + 75, + 833, + 468, + 878 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "ever, it is time-consuming to sufficiently annotate data for providing satisfied results. A promising solution is to leverage useful information from the labeled source domain to assign reliable labels within the unlabeled target domain [18, 58, 60]. Nevertheless, the data distribution could be significantly different across domains, leading to the model degradation on final performance [12, 56, 94]. How to reduce the domain bias and discrepancy while avoiding negative transfer across domains has become an urgent problem.", + "bbox": [ + 500, + 363, + 890, + 500 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "In this paper, we focus on a more general problem of source-free open set domain adaptation. That is, there could exist some target samples belong to extra categories that does not present in the source domain. Moreover, due to data privacy protection constraints, the target domain could only have access to the well-trained model in the source domain. Such settings are commonly encountered in real-world scenarios and it is rather complicated than conventional domain adaptation problem. Specifically, there are two main challenges: CH1: How can we enhance the ability to distinguish target samples that belong to unknown categories? CH2: How to better exploit appropriate pseudo labels for the remaining known target samples?", + "bbox": [ + 496, + 503, + 892, + 700 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Existing researches cannot solve source-free open set domain adaptation well. Firstly, conventional transfer learning always assumes that both source and target domains share the same set of label categories [15, 59, 61]. Directly aligning the latent distributions of source and target domains may lead to negative transfer, where target unknown samples are incorrectly matched to source samples. Secondly, traditional open set domain adaptation approaches should incorporate well-tuned thresholds for determining the decision boundary between known and unknown target samples [69]. This suggests that many of these methods (e.g., OSBP [78] and JPOT [91]) typically rely on prior knowledge about the target domain, which will limit their potential applicabil", + "bbox": [ + 496, + 703, + 893, + 901 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "CVF", + "bbox": [ + 106, + 2, + 181, + 42 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore.", + "bbox": [ + 236, + 0, + 810, + 44 + ], + "page_idx": 0 + }, + { + "type": "page_footnote", + "text": "* Equal Contribution, † Corresponding Author.", + "bbox": [ + 99, + 886, + 351, + 898 + ], + "page_idx": 0 + }, + { + "type": "page_number", + "text": "4927", + "bbox": [ + 482, + 944, + 514, + 955 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/367ba117f828b9bdb3225d33197ee7e634f0001c025e199ac7ddb24ff3292821.jpg", + "image_caption": [ + "(a) Initial Target Features in", + "Latent Space" + ], + "image_footnote": [], + "bbox": [ + 94, + 89, + 240, + 208 + ], + "page_idx": 1 + }, + { + "type": "image", + "img_path": "images/c40dfff88a22a4c7a9315ee6922b8d5e279b901d4dae117c13c436d61b425e35.jpg", + "image_caption": [ + "(b) Adopting Weight Barcode Estimation to Distinguish Unknown Samples in Target Domain" + ], + "image_footnote": [], + "bbox": [ + 243, + 90, + 511, + 208 + ], + "page_idx": 1 + }, + { + "type": "image", + "img_path": "images/6a9845cf78cf0bafd758196053ee075f3662a26df15d935f008dcc09f5405bad.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 511, + 90, + 727, + 208 + ], + "page_idx": 1 + }, + { + "type": "image", + "img_path": "images/76b64f5d902db14e71f944365c9f7a4ca5aa89930c70966f7291d23704118d73.jpg", + "image_caption": [ + "(c) Adopting Sparse Label Assignment to Exploit Pseudo Label for Known Samples in Target Domain", + "(d) Target Features in Latent Space after Training", + "Figure 1. An overview of the proposed DTE. (a) The initial target features provided in the latent space are dispersed and challenging to achieve satisfactory classification results. (b) We first adopt weight barcode estimation component to distinguish known or unknown target samples accordingly. (c) Then we adopt sparse label assignment to exploit pseudo label for known samples in the target domains. (d) Finally we can better cluster and classify target unlabeled samples. Best viewed in color." + ], + "image_footnote": [], + "bbox": [ + 728, + 90, + 875, + 209 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "ity. If such prior knowledge (e.g., the proportion of unknown classes in the target domain) cannot be well estimated, the model performance could be degraded dramatically. Some recent works (e.g., STA [46]) further utilized coarse-to-fine weighting mechanism for separating target samples under different proportion of unknown classes in the target domain (also termed as openness [46]). However, they rely on involving domain adversarial training which could be unstable under some circumstances [14, 81]. Last but not least, we cannot access to the origin source data samples during the training procedure. Previous open set models fail to address such constraints for data privacy protection and they cannot be directly applied into source-free scenario.", + "bbox": [ + 75, + 299, + 473, + 494 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "To address the aforementioned issues, we propose Distinguish Then Exploit (DTE) for tackling source-free open set domain adaptation problem. To better distinguish unknown target samples (solving CH1) and assign appropriate labels to the remaining target samples (solving CH2), we utilize two main components in DTE, i.e., weight barcode estimation component and sparse label assignment component. Weight barcode estimation component aims to solve CH1 by adopting marginal probability estimation and weight barcode establishment to automatically determine unknown target samples without prior knowledge. Sparse label assignment component is designed to address CH2 via leveraging the rest known target samples with proximal optimal transport techniques. By cooperating these two components, we can better utilize source knowledge in the target domain by providing more accurate results on source-free open set domain adaptation as shown in Fig.1. We first distinguish the heart-shaped and diamond-shaped unknown target samples. Then we assign the proper pseudo labels for the known target samples with the shapes of triangle, square, and circle respectively. Finally we can classify target unlabeled samples to reach better performance.", + "bbox": [ + 75, + 501, + 473, + 835 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "We summarize our main contributions as follows: (1) We propose a novel framework, i.e., DTE, for source-free open set domain adaptation problem via knowledge transfer across domains. (2) We innovatively propose weight bar", + "bbox": [ + 76, + 839, + 470, + 901 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "code estimation with partial unbalanced optimal transport to distinguish unknown target samples. Furthermore, we propose target sample reweighting optimization method to figure out the marginal probabilities on partial unbalanced optimal transport. Moreover, we further propose sparse label assignment to provide reliable labels for the known target samples. (3) Extensive empirical studies on various datasets show that DTE achieves the state-of-the-art performance.", + "bbox": [ + 496, + 299, + 893, + 420 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "2. Related work", + "text_level": 1, + "bbox": [ + 500, + 435, + 637, + 450 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Closed Set Domain Adaptation. Conventional domain adaptation usually assumes that both source and target domains share the same set of labels which namely Closed Set Domain Adaptation [59, 61, 62]. There are mainly two types of existing works, i.e., discrepancy-based method and adversarial-based methods [17, 22, 86, 88]. Discrepancy-based methods usually align the statistics information (e.g., mean and covariance) across domains via Maximum Mean Discrepancy (MMD [83, 93]) or Correlation Alignment (CORAL [82]). More recently, optimal transport techniques [4, 7, 13, 26, 39, 72, 85] have been widely adopted into solving domain adaptation problem [23, 24, 68, 76, 80]. Adversarial-based methods, e.g., DANN [2], CDAN [62] and WDGRL [80] which usually involve domain discriminators during training procedure. However, these approaches could easily lead to negative transfer by mismatching the unknown target samples with the source samples.", + "bbox": [ + 496, + 460, + 892, + 719 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Open Set Domain Adaptation. Different from closed set domain adaptation, open set domain adaptation considers more general scenarios when some target samples may belong to unknown label categories [8, 33, 36, 42, 49, 69, 73, 74, 87, 96]. Most of classic open set domain adaptation models (e.g., OSBP [78] and JPOT [91]) set thresholds to determine known and unknown target samples via useful prior knowledge. More recently, some researches (e.g., STA [46] and UOL [47]) also utilize semantic intracategory information for excavating unknown samples with the absent of relevant knowledge. Meanwhile, with the increasing emphasis on protecting data privacy, target do", + "bbox": [ + 496, + 719, + 893, + 901 + ], + "page_idx": 1 + }, + { + "type": "page_number", + "text": "4928", + "bbox": [ + 482, + 944, + 514, + 955 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "main could be relatively difficult to directly obtain source data [20, 95]. However, only a few approaches have been proposed to tackle source-free open set domain adaptation (e.g., SFPGL [64]) since it is more challenging. How to provide satisfied results on open set domain adaptation under the privacy preserving constraints is an urgent problem.", + "bbox": [ + 75, + 90, + 472, + 183 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3. Modelling for DTE", + "text_level": 1, + "bbox": [ + 76, + 195, + 256, + 212 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "First, we describe notations. We assume there are two domains, i.e., a source domain $\\mathcal{S} = \\{(x_i^s,y_i^s)\\}_{i = 1}^{n_s}$ with $n_{s}$ labeled samples and a target domain $\\mathcal{T} = \\{(x_i^t)\\}_{i = 1}^{n_t}$ with $n_t$ unlabeled samples. $\\mathcal{C}^s$ and $\\mathcal{C}^t$ denote the label spaces in source and target domains respectively. Unlike traditional domain adaptation tasks, there exist target samples that belong to additional unknown classes $\\mathcal{C}_{\\mathrm{unk}}^t$ in the target domain. Specifically, the label space in target domain can be formulated as $\\mathcal{C}^t = \\mathcal{C}^s\\cup \\mathcal{C}_{\\mathrm{unk}}^t$ . We aim to correctly assign labels for the target samples via knowledge transferring.", + "bbox": [ + 75, + 220, + 468, + 372 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Since the increasing law on data privacy protection, the source data samples cannot be delivered to the target domain. Hence, the target domain can only access the well-trained source model for tackling the source-free openset domain adaptation. We first adopt CNN architecture in source feature extractor $\\mathcal{F}_s(\\cdot)$ to pretrain on the source data samples to obtain the source features $\\mathcal{F}_s(\\boldsymbol{x}^s) = \\boldsymbol{z}^s$ following [44]. After that we adopt source classifier $\\mathcal{G}(\\cdot)$ to make output prediction $\\pmb{p}^s (y_k|\\pmb{x}_i^s)$ :", + "bbox": [ + 75, + 372, + 470, + 508 + ], + "page_idx": 2 + }, + { + "type": "equation", + "text": "\n$$\n\\left[ \\mathcal {G} \\left(\\boldsymbol {x} _ {i} ^ {s}\\right) \\right] _ {k} = \\boldsymbol {p} ^ {s} \\left(y _ {k} \\mid \\boldsymbol {x} _ {i} ^ {s}\\right) = \\frac {e ^ {\\boldsymbol {w} _ {k} ^ {\\top} \\boldsymbol {z} _ {i} ^ {s} + b _ {k}}}{\\sum_ {l = 1} ^ {H} e ^ {\\boldsymbol {w} _ {l} ^ {\\top} \\boldsymbol {z} _ {i} ^ {s} + b _ {l}}}, \\tag {1}\n$$\n", + "text_format": "latex", + "bbox": [ + 129, + 508, + 468, + 549 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "where $\\pmb{w}_k^\\top$ and $b_{k}$ denote the $k$ -th class weight and bias in the source classifier $\\mathcal{G}(\\cdot)$ . Then we adopt cross-entropy loss to train on the source samples following [44]. Finally, we send the well-trained source model including $\\mathcal{F}_s(\\cdot)$ and $\\mathcal{G}(\\cdot)$ to the unlabeled target domain as shown in Fig. 2. We then introduce the main components of our proposed DTE, as illustrated in Fig. 2. DTE mainly has two components, i.e., weight barcode estimation and sparse label assignment. Weight barcode estimation aims to first distinguish the target samples that belong to the unknown class categories. Sparse label assignment aims to further exploit the rest target samples by providing them with proper pseudo labels. By incorporating weight barcode estimation with sparse label assignment sequentially, we can achieve more precise results on source-free openset domain adaptation. We will introduce the model details in the following subsections.", + "bbox": [ + 75, + 549, + 472, + 792 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3.1. Weight Barcode Estimation", + "text_level": 1, + "bbox": [ + 76, + 801, + 326, + 816 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "In this section, we will first introduce the details of weight barcode estimation. Since target domain include many samples which belong to the unknown classes, it is unreasonable to directly apply traditional alignment techniques in source-free domain adaptation. Specifically, the unknown", + "bbox": [ + 75, + 824, + 470, + 902 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Algorithm 1 The training procedure of proposed DTE", + "text_level": 1, + "bbox": [ + 500, + 90, + 857, + 106 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Input: $\\mathcal{F}_s(\\cdot)$ : well-trained source feature extractor; $\\mathcal{G}(\\cdot)$ : well-trained classifier; $\\epsilon, \\alpha$ : Hyper parameters.", + "bbox": [ + 514, + 108, + 890, + 138 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Set class prototypes $\\pmb{\\mu}^{(0)} = \\pmb{w}$ as the initialization.", + "bbox": [ + 516, + 138, + 851, + 154 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "for epoch $= 1$ to $L$ do", + "bbox": [ + 516, + 154, + 669, + 167 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Obtain the optimal solution on $\\pmb{f}^{*}$ on PUOT for finding marginal probability $q_{\\tau}(\\pmb{x}_i^t) = a_i \\exp(-f_i^* / \\tau)$ .", + "bbox": [ + 532, + 167, + 890, + 200 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Obtain the weight barcode matrix $\\widehat{\\pmb{Q}}$ and determine known/unknown target samples $\\pmb{x}^t = \\pmb{x}_{\\mathrm{unk}}^t\\cup \\pmb{x}_{\\mathrm{kno}}^t$", + "bbox": [ + 532, + 200, + 890, + 231 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Obtain the optimal solution $\\gamma^{*}$ on solving Eq.(6).", + "bbox": [ + 532, + 231, + 857, + 246 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Gather the network output prediction $\\pmb{p}^t$ and the corresponding \"ground-truth\" $\\pmb{q}^t$ on Eq.(9)-Eq.(10).", + "bbox": [ + 532, + 246, + 890, + 276 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Minimize loss function $L_{\\mathrm{cls}}$ on Eq.(11) for training.", + "bbox": [ + 532, + 276, + 874, + 291 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Update class prototypes $\\mu$ via Eq.(12).", + "bbox": [ + 532, + 291, + 790, + 306 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "end for", + "bbox": [ + 516, + 306, + 571, + 319 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Return: The well-trained model in target domain.", + "bbox": [ + 516, + 321, + 846, + 335 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "target samples may get matched to the known class categories and causing negative transfer phenomenon [21, 96]. Therefore, it is essential to distinguish the unknown target samples and filter them out during the first stage. To fulfill this challenging task, we propose weight barcode estimation which includes two steps, i.e., marginal probability estimation and weight barcode establishment. Marginal probability estimation adopts Partial Unbalanced Optimal Transport (PUOT) technique to reweight target samples accordingly. Then we utilize weight barcode establishment to automatically determine the unknown target samples.", + "bbox": [ + 496, + 349, + 890, + 517 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Marginal Probability Estimation. To start with, we first adopt target feature extractor $\\mathcal{F}_t(\\cdot)$ to obtain the latent features $z^t$ on target samples. Then we should consider the matching relationship between target sample features $z^t$ and source class prototype $\\mu$ . Recalling that not all target samples belong to the source class categories, it means that only partial of target sample will get matched. It inspired us to propose PUOT for determining which samples belong to the unkown class categories:", + "bbox": [ + 496, + 518, + 892, + 655 + ], + "page_idx": 2 + }, + { + "type": "equation", + "text": "\n$$\n\\min _ {\\pi_ {i j} \\geq 0} J _ {\\text {p u o t}} = \\left\\langle \\mathcal {U}, \\boldsymbol {\\pi} \\right\\rangle + \\tau \\operatorname {K L} \\left(\\boldsymbol {\\pi} \\mathbf {1} _ {H} \\| \\boldsymbol {a}\\right) \\text {s . t .} \\boldsymbol {\\pi} ^ {\\top} \\mathbf {1} _ {N} = \\boldsymbol {b}, \\tag {2}\n$$\n", + "text_format": "latex", + "bbox": [ + 506, + 657, + 890, + 695 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "where $\\pi \\in \\mathbb{R}^{N\\times H}$ denotes the coupling matrix of PUOT. $N$ denotes the batchsize of target samples and $H$ denotes the number of shared known classes. $\\pmb{a}$ and $\\pmb{b}$ denote the corresponding weights for target samples and source prototypes respectively. Specifically we set $a_{i} = 1 / N$ and $b_{j} = 1 / H$ uniformly. $\\mathcal{U}$ denotes the pairwise cost between target samples $\\boldsymbol{z}^t$ and corresponding prototypes $\\pmb{\\mu}^{(o)}$ at the $o$ -th iteration as $\\mathcal{U}_{ij} = e^{E(\\pmb{z}_i^t)}(||(\\pmb{z}_i^t)|| - \\pmb{\\mu}_j^{(o)}||\\pmb{\\mu}_j^{(o)}||)\\|_2^2$ . We set $\\pmb{\\mu}^{(0)} = \\pmb{w}$ for the initialization. Meanwhile $E(\\cdot)$ denotes the entropy of classifier output predictions as $E(\\pmb{z}_i^t) = -\\sum_{j=1}^{|\\mathcal{C}^s|} \\sigma(\\mathcal{G}(\\mathcal{F}_t(\\pmb{x}_i^t))) [j] \\cdot \\log(\\sigma(\\mathcal{G}(\\mathcal{F}_t(\\pmb{x}_i^t)))) [j])$ where $\\sigma(\\cdot)$ denotes the softmax operation. When $E(\\pmb{x}^t)$ is much larger, it indicates that the model exhibits greater", + "bbox": [ + 496, + 695, + 892, + 901 + ], + "page_idx": 2 + }, + { + "type": "page_number", + "text": "4929", + "bbox": [ + 482, + 944, + 514, + 955 + ], + "page_idx": 2 + }, + { + "type": "image", + "img_path": "images/041a74fd4ccee110878314f35814047408f4c2d35817f14f2708859b7530c881.jpg", + "image_caption": [ + "Figure 2. The architectures of Distinguish Then Assign (DTE). We first train source feature extractor $\\mathcal{F}_s(\\cdot)$ and classifier $\\mathcal{G}(\\cdot)$ in the source domain. Then we share the well-trained source model to the target domain. We fine-tune target feature extractor $\\mathcal{F}_t(\\cdot)$ based on $\\mathcal{F}_s(\\cdot)$ and fix classifier $\\mathcal{G}(\\cdot)$ . Then we propose weight barcode estimation for distinguishing the known and unknown target samples via distinguisher $\\hat{\\mathcal{G}}(\\cdot)$ . Finally we propose sparse label assignment for exploiting appropriate pseudo labels for the remaining target samples." + ], + "image_footnote": [], + "bbox": [ + 138, + 90, + 831, + 282 + ], + "page_idx": 3 + }, + { + "type": "image", + "img_path": "images/043f30434967b835b04bf6ce7e29c0f2d737b3843b00243b27e1483aa6ab5b43.jpg", + "image_caption": [ + "Figure 3. The illustration on the proposed weight barcode estimation with toy examples. We first sample 100 source/target data samples and set $30\\%$ or $70\\%$ of target data samples from unknown categories. Then we adopt marginal probability estimation to obtain the results of target weights. Finally we calculate and visualize the result of $\\hat{Q}$ to distinguish the unknown target samples." + ], + "image_footnote": [], + "bbox": [ + 84, + 349, + 460, + 547 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "uncertainty regarding the target samples into known categories. Therefore, it is reasonable to consider such information into the cost calculation on $\\mathcal{U}$ . The KL divergence can be calculated via KL $(\\pmb{x}||\\pmb{z}) = \\sum_{d=1}^{D}[x_d\\log(x_d/z_d) - x_d + z_d]$ and $\\tau$ denotes the hyper parameters. Unlike the classic optimal transport problem, we relax one of the marginal constraints for the target samples via the KL divergence term to avoid the mismatch. To solve the above problem, we first propose the theorem on the dual form of PUOT:", + "bbox": [ + 75, + 646, + 468, + 781 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Theorem 1. Given the PUOT problem $J_{\\mathrm{puot}}$ as shown in Eq.(2), it has the following dual form:", + "bbox": [ + 76, + 782, + 468, + 813 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\min _ {\\boldsymbol {f}, \\boldsymbol {g}, \\zeta} \\left[ \\tau \\sum_ {i = 1} ^ {N} a _ {i} \\exp \\left(- \\frac {f _ {i} + \\zeta}{\\tau}\\right) - \\sum_ {j = 1} ^ {H} b _ {j} \\left(g _ {j} - \\zeta\\right) \\right] \\tag {3}\n$$\n", + "text_format": "latex", + "bbox": [ + 93, + 815, + 468, + 864 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\ns. t. f _ {i} + g _ {j} + s _ {i j} = \\mathcal {U} _ {i j}, \\quad s _ {i j} \\geq 0,\n$$\n", + "text_format": "latex", + "bbox": [ + 91, + 866, + 333, + 883 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "where $f, g, s$ and $\\zeta$ denote Lagrange multipliers. Mean", + "bbox": [ + 76, + 886, + 468, + 901 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "while $s_{ij} \\geq 0$ according to the KKT conditions. Moreover, PUOT will be transformed into sample-reweighted optimal transport problem according to its dual form:", + "bbox": [ + 498, + 349, + 892, + 396 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\min _ {\\boldsymbol {\\pi} \\geq 0} \\langle \\mathcal {U}, \\boldsymbol {\\pi} \\rangle\n$$\n", + "text_format": "latex", + "bbox": [ + 509, + 398, + 591, + 422 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\ns. t. \\sum_ {j = 1} ^ {H} \\pi_ {i j} = a _ {i} \\exp \\left(- \\frac {f _ {i} ^ {*} + \\zeta^ {*}}{\\tau}\\right), \\quad \\sum_ {i = 1} ^ {N} \\pi_ {i j} = b _ {j}. \\tag {4}\n$$\n", + "text_format": "latex", + "bbox": [ + 508, + 425, + 890, + 468 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "The proof of Theorem 1 is provided in Appendix. That is, the target data samples will be sample-reweighted with coefficient $\\exp \\left(-\\left(f_{i}^{*} + \\zeta^{*}\\right) / \\tau\\right)$ accordingly. At that time the marginal probability on target sample $\\pmb{x}_i^t$ can be given as $q_{\\tau}(\\pmb{x}_i^t) = a_i\\exp \\left(-\\left(f_i^* +\\zeta^*\\right) / \\tau\\right)$ . Kindly note that we do not necessarily require an explicit expression for solving $\\pi$ , which involves $(N\\times H)$ variables and can be quite complex. Meanwhile we mainly focus on solving these dual form Lagrange multipliers in Eq.(3) with $(N + H)$ variables. Specifically, we can utilize the $c$ -transform [41, 55, 71] in classical optimal transport theorem to reach the upper bound of $\\pmb{g}$ as $g_{j} = \\inf_{k\\in [N]}\\left(\\mathcal{U}_{kj} - f_{k}\\right)$ . By taking the upper bound of $\\pmb{g}$ into Eq.(3), we can reach the final optimization formula as:", + "bbox": [ + 498, + 470, + 892, + 667 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\min _ {\\boldsymbol {f}, \\zeta} \\left[ \\tau \\sum_ {i = 1} ^ {N} a _ {i} e ^ {- \\frac {f _ {i} + \\zeta}{\\tau}} - \\sum_ {j = 1} ^ {H} \\left(\\inf _ {k \\in [ N ]} \\left(\\mathcal {U} _ {k j} - f _ {k}\\right) - \\zeta\\right) b _ {j} \\right]. \\tag {5}\n$$\n", + "text_format": "latex", + "bbox": [ + 501, + 670, + 890, + 731 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "That equation can be optimized easily. Due to the space limited, the details of optimization procedures for Eq.(5) are provided in Appendix. In summary, we only need to figure out the marginal probability on target samples $q_{\\tau}(\\boldsymbol{x}_i^t)$ via solving the optimal results of $f$ and $\\zeta$ in Eq.(5).", + "bbox": [ + 496, + 732, + 890, + 808 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Weight Barcode Establishment. According to the optimization process, we can observe that choosing different value of $\\tau$ could lead to different marginal probability. Specifically, as the value of $\\tau$ increases, a larger number of target samples will be assigned higher weights, enhancing their likelihood of being matched. However, we do not", + "bbox": [ + 496, + 810, + 892, + 900 + ], + "page_idx": 3 + }, + { + "type": "page_number", + "text": "4930", + "bbox": [ + 482, + 944, + 514, + 955 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "know the exact percentage of how many samples are unknown in the target domain which makes it relatively hard to choose a proper value of $\\tau$ . To address this issue, we propose an innovative weight barcode mechanism to effectively identify target samples that are more likely belong to unknown classes. We first calculate $q_{\\tau}(\\boldsymbol{x}^t)$ by varying $\\tau$ in range of $\\{0.01, 0.05, 0.1, 0.5, 1, 5, 10\\}$ and collecting them as $\\boldsymbol{Q} \\in \\mathbb{R}^{N \\times 7}$ . Then we can establish the binary weight barcode matrix $\\widehat{\\boldsymbol{Q}} \\in \\mathbb{R}^{N \\times 7}$ with the following process as $\\widehat{\\boldsymbol{Q}}_{ij} = \\mathrm{func}(Q_{ij} - 1/N)$ where $\\mathrm{func}(x) = 1$ when $x \\geq 0$ otherwise $\\mathrm{func}(x) = 0$ . It is reasonable since PUOT will enhance the target sample weights if they exhibit a higher degree of similarity to the source data. Meanwhile the rest of unlike target samples will gain less weights and we set then $\\widehat{\\boldsymbol{Q}}_{ij} = 0$ . To make this process more intuitive, we further provide an numerical example in Fig. 3. Specifically, the red and blue points denote the source and target data samples respectively. We first figure out the pairwise distance between red data $\\boldsymbol{z}_i^{\\mathrm{Red}}$ and blue data $\\boldsymbol{z}_j^{\\mathrm{Blue}}$ via $\\mathcal{U}_{ij} = ||\\boldsymbol{z}_i^{\\mathrm{Red}} - \\boldsymbol{z}_j^{\\mathrm{Blue}}||_2^2$ . Then we calculate the weight barcode as $\\widehat{\\boldsymbol{Q}}$ by using marginal probability estimation and visualize the results. We can observe that the unknown target samples achieve lower value of $\\widehat{\\boldsymbol{Q}}$ which can be easily identify. It indicates that the proposed method is adaptable to varying ratios of unknown data samples, even in the absence of prior knowledge.", + "bbox": [ + 76, + 90, + 472, + 488 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "After we obtain the binary weight barcode matrix $\\widehat{Q}$ , we can further adopt $K$ -Means algorithm [28] on $\\widehat{Q}$ to cluster them into two groups with corresponding centers $\\mathcal{C}_0$ and $\\mathcal{C}_1$ . We further denote $\\mathcal{C}_{\\mathrm{known}} = \\max \\left(\\| \\mathcal{C}_0\\| _2, \\| \\mathcal{C}_1\\| _2\\right)$ and $\\mathcal{C}_{\\mathrm{unknown}} = \\min \\left(\\| \\mathcal{C}_0\\| _2, \\| \\mathcal{C}_1\\| _2\\right)$ by considering the norm of centers. Then we can determine the target samples by measuring the Euclidean distance among $\\widehat{Q}$ and centers $\\mathcal{C}_{\\mathrm{known}}$ , $\\mathcal{C}_{\\mathrm{unknown}}$ . If the target samples are found to be in close proximity to $\\mathcal{C}_{\\mathrm{unknown}}$ , we can classify them as unknown data as $\\boldsymbol{x}_{\\mathrm{unk}}^t$ . The remaining target domain samples can all be considered as belonging to the known categories as $\\boldsymbol{x}_{\\mathrm{kn0}}^t$ and as $\\boldsymbol{x}^t = \\boldsymbol{x}_{\\mathrm{unk}}^t \\cup \\boldsymbol{x}_{\\mathrm{kn0}}^t$ . In summary, we can now determine the known/unknown target samples accordingly.", + "bbox": [ + 76, + 488, + 472, + 686 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "3.2. Sparse Label Assignment", + "text_level": 1, + "bbox": [ + 76, + 695, + 308, + 710 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Since we have exploited the target samples with unknown classes $\\boldsymbol{x}_{\\mathrm{unk}}^{t}$ via proposed weight barcode estimation method, we should explore the possible labels for the rest target samples $\\boldsymbol{x}_{\\mathrm{kno}}^{t}$ . Recall that $\\boldsymbol{x}_{\\mathrm{kno}}^{t}$ share the same categories with the corresponding prototypes $\\boldsymbol{\\mu}^{s}$ , one can adopt domain adaptation manners for finding the latent relationship. Meanwhile, optimal transport techniques could be rather useful in reducing the domain discrepancy via discovering the most proper pairwise matching across domains [9, 25, 75]. However, directly optimizing optimal transport with linear programming would involve heavy computation cost [1, 34, 92]. Although utilizing entropy regularization", + "bbox": [ + 75, + 719, + 472, + 901 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "term could enhance the efficiency, it will bring about relatively dense matching solution which is always inaccurate as shown in Fig. 4(a) [6, 10, 50]. To alleviate this issue, we propose Sparse Sample-Label Matching (SSLM) on optimal transport via involving $\\ell_2$ -norm proximal term:", + "bbox": [ + 496, + 90, + 890, + 167 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\min _ {\\gamma \\in \\Delta} \\ell = \\sum_ {i = 1} ^ {N _ {\\mathrm {k n o}}} \\sum_ {j = 1} ^ {H} \\left[ \\gamma_ {i j} C _ {i j} + \\frac {\\epsilon}{2} \\left(\\gamma_ {i j} - \\gamma_ {i j} ^ {(l)}\\right) ^ {2} \\right], \\tag {6}\n$$\n", + "text_format": "latex", + "bbox": [ + 540, + 167, + 890, + 209 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "where $\\Delta = \\{\\sum_{j=1}^{H} \\gamma_{ij} = \\widehat{a}_i = 1, \\sum_{i=1}^{N_{\\mathrm{kno}}} \\gamma_{ij} = \\widehat{b}_j = N_{\\mathrm{kno}} / H, \\gamma_{ij} \\geq 0\\}$ and $N_{\\mathrm{kno}}$ denotes the number of target samples $\\boldsymbol{x}_{\\mathrm{kno}}^t$ . $C = -\\log[\\sigma[\\mathcal{G}(\\mathcal{F}_t(\\boldsymbol{x}_{\\mathrm{kno}}^t))]]$ denote the output classification probability with negative logarithm. Indeed, when $C_{ij}$ is smaller, it indicates that $\\boldsymbol{x}_{\\mathrm{kno},i}^t$ has a higher probability of belonging to the $j$ -th class [3, 27, 45]. $\\widehat{a}_i$ and $\\widehat{b}_j$ denote the samples weights and we set $\\widehat{a}_i = 1$ , $\\widehat{b}_j = \\frac{N_{\\mathrm{kno}}}{H}$ respectively. Meanwhile $\\sum_{i=1}^{M} \\sum_{j=1}^{N} (\\gamma_{ij} - \\gamma_{ij}^{(l)})^2$ denotes the proximal point term with $\\ell_2$ -norm. Incorporating $\\ell_2$ -norm with proximal point method [90], we can reach more sparse and robust results on transportation plan $\\gamma^*$ . We initialize $\\gamma_{ij}^{(0)} = \\frac{1}{H}$ for the following iterations. Then we can easily obtain the Fenchel-Legendre dual form of the original problem as:", + "bbox": [ + 496, + 210, + 893, + 435 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\max _ {\\psi , \\phi} \\sum_ {i = 1} ^ {N _ {\\mathrm {k n o}}} \\psi_ {i} \\widehat {a} _ {i} + \\sum_ {j = 1} ^ {H} \\phi_ {j} \\widehat {b} _ {j} - \\frac {\\epsilon}{2} \\sum_ {i = 1} ^ {N _ {\\mathrm {k n o}}} \\sum_ {j = 1} ^ {H} \\left[ \\frac {\\psi_ {i} + \\phi_ {j} - \\widehat {C} _ {i j}}{\\epsilon} \\right] _ {+} ^ {2},\n$$\n", + "text_format": "latex", + "bbox": [ + 498, + 436, + 901, + 494 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "where $\\phi$ and $\\psi$ denote the Lagrange multipliers. Meanwhile we set $[x]_{+}^{2} = (\\max (0,x))^{2}$ and $\\widehat{C}_{ij} = C_{ij} - \\epsilon \\gamma_{ij}^{(l)}$ . Note that the original constrained optimization problem in Eq.(6) turns to be unconstrained after the duality transformation in Eq.(7). The optimization procedure on solving variables $\\psi$ and $\\phi$ are provided in Appendix. After that, we can further calculate the optimal solution of $\\gamma^{(l + 1)}$ at the $l$ -th iteration as shown:", + "bbox": [ + 496, + 494, + 890, + 616 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\gamma_ {i j} ^ {(l + 1)} = \\max (0, [ \\psi_ {i} + \\phi_ {j} + \\epsilon \\gamma_ {i j} ^ {(l)} - C _ {i j} ] / \\epsilon). \\tag {8}\n$$\n", + "text_format": "latex", + "bbox": [ + 544, + 614, + 890, + 637 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "After we obtain the optimal solution on $\\gamma^{*}$ via solving sparse sample-prototype matching, we can reach the reliable pseudo labels for these target samples $\\pmb{x}_{\\mathrm{kno}}^t$ . Furthermore, the optimal solution $\\gamma^{*}$ can yield more sparser and clearer results compared to using only the $\\ell_{2}$ -norm regularization, as depicted in Fig. 4(b)-(c).", + "bbox": [ + 496, + 636, + 890, + 727 + ], + "page_idx": 4 + }, + { + "type": "image", + "img_path": "images/d00b1878bcffeb0e669ff1bb5730968ed33c80b86d2e812e31e0499e4dd7fc2a.jpg", + "image_caption": [ + "Figure 4. The illustration on Sparse Sample-Label Matching (SSLM) in sparse label assignment with toy examples. Sinkhorn and tradition $\\ell_2$ -norm solvers could easily reach smooth but dense output solutions. Our proposed SSLM method can reach more sparse matching solution for better performance." + ], + "image_footnote": [], + "bbox": [ + 506, + 729, + 883, + 816 + ], + "page_idx": 4 + }, + { + "type": "page_number", + "text": "4931", + "bbox": [ + 482, + 944, + 513, + 955 + ], + "page_idx": 4 + }, + { + "type": "image", + "img_path": "images/a1db6d8aa1371d412013e8092b8e25829091aeffb96cbfa6600d46b79936f023.jpg", + "image_caption": [ + "(a) JPOT on Office-31" + ], + "image_footnote": [], + "bbox": [ + 83, + 90, + 269, + 190 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/2511ba83b85b9efae26ba587c5599bffc615004fe1de4cf38675dd87e067181b.jpg", + "image_caption": [ + "(b) DTE on Office-31", + "Figure 5. The t-SNE visualization of $\\mathbf{W} \\rightarrow \\mathbf{A}$ tasks on Office-31. Each color denotes a class and the reddish-brown color is used to represent the unknown class." + ], + "image_footnote": [], + "bbox": [ + 281, + 90, + 467, + 189 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "3.3. Putting Together", + "text_level": 1, + "bbox": [ + 76, + 263, + 243, + 280 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "We have first distinguished the target samples with unknown classes via weight barcode estimation, then assigning the rest target samples to the corresponding classes via sparse sample-prototype matching. Therefore, we should gather the output prediction of these target samples as $\\pmb{p}^t$ :", + "bbox": [ + 75, + 287, + 468, + 362 + ], + "page_idx": 5 + }, + { + "type": "equation", + "text": "\n$$\n\\boldsymbol {p} ^ {t} = \\operatorname {s o f t m a x} \\left[ \\mathcal {G} \\left(\\mathcal {F} _ {t} \\left(\\boldsymbol {x} ^ {t}\\right)\\right) \\oplus \\widehat {\\mathcal {G}} \\left(\\mathcal {F} _ {t} \\left(\\boldsymbol {x} ^ {t}\\right)\\right) \\right] \\in \\mathbb {R} ^ {N \\times (H + 1)}. \\tag {9}\n$$\n", + "text_format": "latex", + "bbox": [ + 83, + 364, + 468, + 395 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Note that $\\widehat{\\mathcal{G}} (\\cdot)$ is a newly add trainable target classifier to predict whether samples are unknown. Here we concatenate the output predictions on both source classifier $\\mathcal{G}(\\cdot)$ and target classifier $\\widehat{\\mathcal{G}} (\\cdot)$ then process the softmax operation. Meanwhile we can formulate the estimated \"ground-truth\" $q_{\\mathrm{kno}}^{t}$ and $q_{\\mathrm{unk}}^{t}$ for the target samples as follows:", + "bbox": [ + 75, + 396, + 468, + 489 + ], + "page_idx": 5 + }, + { + "type": "equation", + "text": "\n$$\n\\boldsymbol {q} _ {\\mathrm {k n o}} ^ {t} = \\left[ \\gamma^ {*} \\oplus 0 \\right] \\in \\mathbb {R} ^ {N _ {\\mathrm {k n o}} \\times (H + 1)}, \\tag {10}\n$$\n", + "text_format": "latex", + "bbox": [ + 158, + 489, + 468, + 516 + ], + "page_idx": 5 + }, + { + "type": "equation", + "text": "\n$$\n\\boldsymbol {q} _ {\\mathrm {u n k}} ^ {t} = [ \\mathbf {0} \\oplus 1 ] \\in \\mathbb {R} ^ {N _ {\\mathrm {u n k}} \\times (H + 1)}.\n$$\n", + "text_format": "latex", + "bbox": [ + 161, + 513, + 375, + 530 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Then we can utilize the cross-entropy loss $L_{\\mathrm{cls}}$ on fine-tuning the target feature extractor $\\mathcal{F}_t(\\cdot)$ based on $\\mathcal{F}_s(\\cdot)$ and training distinguisher $\\widehat{\\mathcal{G}}_t(\\cdot)$ on the target samples as:", + "bbox": [ + 75, + 532, + 468, + 579 + ], + "page_idx": 5 + }, + { + "type": "equation", + "text": "\n$$\n\\min _ {\\mathcal {F} _ {t} (\\cdot), \\widehat {\\mathcal {G}} (\\cdot)} L _ {\\mathrm {c l s}} = - \\sum_ {i = 1} ^ {N _ {\\mathrm {u n k}} + N _ {\\mathrm {k n o}}} \\sum_ {j = 1} ^ {H + 1} q _ {i j} ^ {t} \\log p _ {i j} ^ {t}. \\tag {11}\n$$\n", + "text_format": "latex", + "bbox": [ + 102, + 580, + 468, + 623 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Note that we fix the source classifier $\\mathcal{G}(\\cdot)$ during the training procedure. Meanwhile, the loss function is easy to implemented for end-to-end training. Moreover, we should update the class prototype $\\pmb{\\mu}^{(o + 1)}$ at the $(o + 1)$ -th iteration via incremental learning strategy:", + "bbox": [ + 75, + 625, + 468, + 700 + ], + "page_idx": 5 + }, + { + "type": "equation", + "text": "\n$$\n\\boldsymbol {\\mu} _ {j} ^ {(o + 1)} = \\alpha \\boldsymbol {\\mu} _ {j} ^ {(o)} + \\frac {1 - \\alpha}{\\sum_ {i = 1} ^ {N _ {\\mathrm {k n o}}} p _ {i j} ^ {t}} \\sum_ {i = 1} ^ {N _ {\\mathrm {k n o}}} \\mathcal {F} _ {t} \\left(\\boldsymbol {x} _ {\\mathrm {k n o}, i} ^ {t}\\right) p _ {i j} ^ {t}, \\tag {12}\n$$\n", + "text_format": "latex", + "bbox": [ + 89, + 703, + 468, + 744 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "where $\\alpha \\in [0,1]$ denotes momentum hyper parameter. The whole algorithm of DTE in provided in Alg.1.", + "bbox": [ + 75, + 746, + 468, + 776 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "4. Experiment", + "text_level": 1, + "bbox": [ + 76, + 791, + 200, + 808 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "4.1. Setup", + "text_level": 1, + "bbox": [ + 76, + 816, + 158, + 832 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "We conduct the experiments on several real-world datasets, e.g., Office-31 [77] with following three domains: Amazon (A), Webcam (W) and DSLR (D); Office-Home [84] with following four dissimilar domains: Artistic images (Ar),", + "bbox": [ + 75, + 839, + 468, + 901 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Clip Art (CI), Product images (Pr) and Real-World (Rw); VisDA-2017 [70] with domains Synthetic, Real; and Digits with following three domains MNIST [40], USPS [31] and SVHN [67]. We adopt the same set of known and unknown classes following previous works [46, 78].", + "bbox": [ + 496, + 90, + 890, + 166 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Implementation Details. We first set batch size $N = 256$ during the training procedure. For the experiments on Office-31, Office-Home, and VisDA, we first fine-tune on the backbone network ResNet-50 [29] in the source domain. For the experiments on Digits, we adopt LeNet [40] to extract the sample features. After the training in the source domain, we only transfer the well-trained feature extractor $\\mathcal{F}_s(\\cdot)$ and classifier $\\mathcal{G}(\\cdot)$ to the target domain. We set $\\epsilon = 0.5$ in sparse label assignment. Meanwhile we set $\\alpha = 0.8$ for the incremental learning on class prototypes. We choose Adam [35] as the optimizer and set the learning rate as 0.001 and 0.01 for fine-tuning target feature extractor $\\mathcal{F}_t(\\cdot)$ based on $\\mathcal{F}_s(\\cdot)$ and training distinguisher $\\widehat{\\mathcal{G}} (\\cdot)$ respectively. We employ these four evaluation metrics, i.e., OS, OS*, UNK, HOS and for open set domain adaptation following [78]. Specifically, OS and OS* denotes the accuracy averaged over all classes or only known classes respectively. UNK denotes the accuracy of unknown instances and HOS denotes the harmonic value of OS* and UNK. We compare the proposed DTE with state-of-the-art domain adaptation methods: ResNet [29], Domain Adversarial Neural Network (DANN) [2], Deep Joint Distribution Optimal Transport (DeepJDOT) [11], OpenMAX [5], MMD+OSVM, DANN+OSVM, Openset Back-Propagation (OSBP) [78], Separate To Adapt (STA) [46], Towards Inheritable Models (TIM) [37], Joint Partial Optimal Transport (JPOT) [91], Source Hypothesis Transfer (SHOT) [44], Progressive Graph Learning (PGL) [63], Source-Free Progressive Graph Learning (SFPGL) [64], Unknown-Aware Domain Adversarial Learning (cUADAL) [32], Self-supervised Class-Disccovering Adapter (SCDA) [97], Unknown Oriented Learning (UOL) [47], Adjustment and Alignment (ANNA) [43] and Balanced Correlation Learning (BCL) [30]. We perform five random experiments and report the average and variation results.", + "bbox": [ + 496, + 167, + 892, + 698 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "4.2. Result and Discussion", + "text_level": 1, + "bbox": [ + 500, + 710, + 702, + 724 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "The comparison results on Office-31 are shown in Table. 1. Directly applying ResNet could lead to worst results since the existence of domain discrepancy. Although previous method (e.g., OSBP, STA) adopted domain adversarial strategy to reduce the discrepancy, it could be unstable and hard to train in the real practice [81]. Our proposed DTE achieves the best performance among other approaches. Even for the difficult task (e.g., $\\mathbf{A} \\rightarrow \\mathbf{W}$ , $\\mathbf{W} \\rightarrow \\mathbf{A}$ ) where these images are quite different across domains, DTE still obtains satisfactory results. It indicates the efficacy of the proposed DTE for dealing with open set domain", + "bbox": [ + 496, + 734, + 890, + 900 + ], + "page_idx": 5 + }, + { + "type": "page_number", + "text": "4932", + "bbox": [ + 482, + 944, + 514, + 955 + ], + "page_idx": 5 + }, + { + "type": "table", + "img_path": "images/6bdcc07379bb6249dbd5b1849878b5f26266a5d19d9a0c35627c74060bc7567b.jpg", + "table_caption": [ + "Table 1. Classification accuracy (%) on Office-31 for open set domain adaptation (ResNet-50)." + ], + "table_footnote": [], + "table_body": "
MethodA→WA→DD→WW→DD→AW→AAvg
OSOS*OSOS*OSOS*OSOS*OSOS*OSOS*OSOS*
ResNet82.5±1.282.7±0.985.2±0.385.5±0.994.1±0.394.3±0.796.6±0.297.0±0.471.6±1.071.5±1.175.5±1.075.2±1.684.284.4
DANN85.3±0.787.7±1.186.5±0.687.7±0.697.5±0.298.3±0.599.5±0.1100.0±075.7±1.676.2±0.974.9±1.275.6±0.886.687.6
DeepJDOT86.1±0.588.7±0.986.9±0.789.0±0.596.9±0.295.6±0.896.1±0.198.1±0.585.6±1.288.4±1.081.5±1.383.1±0.988.990.5
OpenMax87.4±0.587.5±0.387.1±0.988.4±0.996.1±0.496.2±0.398.4±0.398.5±0.383.4±1.082.1±0.682.8±0.982.8±0.689.089.3
OSBP86.5±2.087.6±2.188.6±1.489.2±1.397.0±1.096.5±0.497.9±0.998.7±0.688.9±2.590.6±2.385.8±2.584.9±1.390.891.3
STA89.5±0.692.1±0.593.7±1.596.1±0.497.5±0.296.5±0.599.5±0.299.6±0.189.1±0.593.5±0.887.9±0.987.4±0.692.994.1
TIM91.3±0.793.2±0.594.2±1.197.1±0.896.5±0.597.4±0.799.5±0.299.4±0.390.1±0.291.5±0.288.7±1.388.4±0.993.494.5
JPOT92.8±0.692.2±0.495.2±0.996.0±0.698.1±0.396.2±0.499.5±0.198.6±0.293.0±0.794.1±0.489.9±1.088.4±0.494.694.4
SHOT88.8±0.791.4±0.490.3±0.592.6±0.396.2±0.497.0±0.497.4±0.397.9±0.591.6±0.593.4±0.791.2±0.593.5±0.492.694.3
PGL89.2±0.590.1±0.789.6±0.891.6±0.595.3±0.495.1±0.596.7±0.597.6±0.671.0±1.372.0±1.873.0±1.277.6±1.685.887.4
SFPGL90.4±0.391.9±0.791.0±1.091.2±0.897.3±0.595.5±0.696.1±0.498.4±0.474.8±1.575.6±2.074.9±1.679.6±1.187.488.7
cUADAL92.5±0.493.7±0.696.0±0.797.5±0.698.4±0.4100.0±098.6±0.5100.0±085.3±0.886.0±0.583.2±0.783.9±0.992.393.5
SCDA95.7±0.197.5±0.295.9±0.496.5±0.499.2±0.499.7±0.199.8±0.1100.0±092.1±0.193.7±0.392.2±0.193.6±0.195.896.8
UOL94.6±0.396.8±0.695.4±0.597.2±0.799.5±0.698.2±0.398.9±0.299.1±0.392.7±0.693.2±0.891.6±1.093.0±0.895.596.2
DTE-W93.9±0.694.5±0.595.7±0.696.3±0.798.2±0.599.1±0.297.4±0.498.8±0.694.0±0.993.6±1.290.7±0.891.5±0.794.995.6
DTE-S95.3±0.596.9±0.496.5±0.697.0±0.898.7±0.799.4±0.398.1±0.499.2±0.594.8±1.193.9±1.092.5±0.794.1±0.995.996.7
DTE96.1±0.297.5±0.496.8±0.497.9±0.598.9±0.5100.0±098.5±0.399.4±0.395.6±0.594.3±0.692.9±0.494.6±0.396.497.2
", + "bbox": [ + 106, + 108, + 857, + 311 + ], + "page_idx": 6 + }, + { + "type": "table", + "img_path": "images/8e8af3547d49e917071acfa34626f867ff519dc485777f6e724f18430ef33d63.jpg", + "table_caption": [ + "Table 2. Classification accuracy OS* (%) on Office-Home for open set domain adaptation (ResNet-50)." + ], + "table_footnote": [], + "table_body": "
MethodAr→ClPr→ClRw→ClAr→PrCl→PrRw→PrCl→ArPr→ArRw→ArAr→RwCl→RwPr→RwAvg
ResNet53.4±0.452.7±0.651.9±0.569.3±0.761.8±0.574.1±0.461.4±0.664.0±0.370.0±0.378.7±0.671.0±0.674.9±0.965.3
DANN54.6±0.749.7±1.651.9±1.469.5±1.163.5±1.072.9±0.861.9±1.263.3±1.071.3±1.080.2±0.871.7±0.474.2±0.465.4
OpenMax56.5±0.452.9±0.753.7±0.469.1±0.364.8±0.474.5±0.664.1±0.964.0±0.871.2±0.880.3±0.873.0±0.576.9±0.366.7
OSBP56.7±1.951.5±2.149.2±2.467.5±1.565.5±1.574.0±1.562.5±2.064.8±1.169.3±1.180.6±0.974.7±2.271.5±1.965.7
STA58.1±0.653.1±0.954.4±1.071.6±1.269.3±1.081.9±0.563.4±0.565.2±0.874.9±1.085.0±0.275.8±0.480.8±0.369.5
TIM60.1±0.754.2±1.056.2±1.770.9±1.470.0±1.778.6±0.664.0±0.666.1±1.374.9±0.983.2±0.975.7±1.381.3±1.469.6
JPOT59.6±0.554.2±0.754.6±0.972.3±1.170.1±0.682.1±0.962.9±0.768.3±0.875.1±1.184.8±0.477.4±0.581.2±0.470.2
SHOT64.5±0.859.3±0.564.6±0.880.4±0.975.4±1.282.3±1.063.1±0.865.3±0.669.6±0.784.7±0.681.2±1.083.3±0.872.8
PGL61.6±1.058.4±0.865.0±0.777.1±1.572.0±1.383.0±0.968.8±1.472.2±0.778.6±1.385.9±0.982.8±1.182.6±1.274.0
SFPGL62.4±0.559.0±0.764.1±0.977.8±0.772.7±0.884.2±1.070.5±0.673.4±0.578.3±0.886.8±0.683.9±0.881.4±0.774.5
cUADAL62.3±0.660.1±0.861.7±0.277.9±0.671.4±0.981.3±0.572.3±0.671.5±0.873.9±1.083.6±0.783.5±0.784.0±1.073.6
SCDA59.9±0.359.0±0.362.8±0.579.6±0.473.8±1.083.7±0.870.9±0.572.3±0.675.5±0.485.3±0.682.9±0.385.7±0.974.3
UOL61.0±0.759.5±0.663.3±0.680.7±1.074.9±0.885.8±0.771.6±0.972.9±0.675.1±0.588.4±0.883.2±0.486.3±0.974.4
ANNA61.3±0.960.4±0.762.5±0.882.1±0.673.6±0.784.4±0.570.2±1.171.6±0.975.7±0.686.9±0.582.0±1.286.9±1.574.8
BCL58.4±0.559.7±0.661.1±0.680.6±0.573.0±0.985.1±0.470.5±0.772.2±0.674.8±0.587.5±0.683.4±0.986.1±1.274.4
DTE-W60.1±0.957.2±0.859.4±0.978.5±0.672.3±1.083.0±0.669.1±0.871.9±0.775.8±1.186.7±0.581.6±0.682.8±1.573.2
DTE-S62.1±0.660.9±0.861.8±0.880.4±0.574.2±1.285.7±0.971.6±0.973.3±1.076.9±0.788.3±0.882.0±0.886.5±1.376.1
DTE63.2±0.462.6±0.663.9±0.581.5±0.575.4±0.886.8±0.672.0±0.574.1±0.477.3±0.790.2±0.684.3±0.487.5±1.176.5
", + "bbox": [ + 106, + 330, + 857, + 549 + ], + "page_idx": 6 + }, + { + "type": "table", + "img_path": "images/b175c5327f4cdec059d118fe970a7c322ac91602a9da79f368e9b3c9c066ea98.jpg", + "table_caption": [ + "Table 3. Classification accuracy (%) on Digits (LeNet) and VisDA2017 (VGGNet) for open set domain adaptation." + ], + "table_footnote": [], + "table_body": "
MethodSVHN→MNISTUSPS→MNISTMNIST→USPSAvgVisDA2017
OSOS*HOSUNKOSOS*HOSUNKOSOS*HOSUNKOSOS*HOSUNKbicyclebuscarmotorcycletraintruckOSOS*
OSVM54.363.117.710.543.132.348.597.179.877.983.289.359.157.749.865.731.751.666.570.488.520.852.554.9
MMD+OSVM55.964.720.112.262.858.968.782.380.079.880.481.066.268.856.458.439.050.164.279.986.616.354.456.0
DANN+OSVM62.975.31.780.9077.192.41.190.6033.835.529.525.360.469.410.88.9331.856.671.777.487.022.355.557.8
OSBP63.059.168.982.592.391.294.497.892.194.985.778.182.481.782.986.151.167.142.884.281.828.062.959.2
STA76.975.479.784.492.291.393.996.793.094.988.883.587.387.287.588.252.469.659.987.886.527.266.863.9
TIM80.479.083.087.495.495.196.096.993.093.591.890.589.689.390.391.653.569.262.285.785.432.568.164.7
JPOT77.275.380.686.792.491.294.798.492.992.194.496.987.585.489.994.055.370.064.189.587.233.769.566.6
UOL84.484.185.085.996.997.196.595.995.395.195.796.392.291.692.492.754.670.563.791.286.931.470.066.3
DTE-W83.182.883.784.696.596.496.797.094.594.295.196.090.391.491.892.556.070.965.688.884.132.269.966.2
DTE-S84.884.585.486.397.096.997.297.594.994.695.596.492.592.292.793.456.571.266.390.187.733.871.367.6
DTE85.585.286.187.097.697.597.898.195.695.496.096.692.992.793.393.957.472.067.990.888.635.272.468.7
", + "bbox": [ + 106, + 566, + 857, + 689 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "adaptation even under the source-free scenario.", + "bbox": [ + 76, + 693, + 388, + 708 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Meanwhile the results of Office-Home are shown in Table. 2. The tasks on Office-Home are more difficult than Office-31 since it has more categories with large domain gap. Previous methods cannot better reduce domain discrepancy while involving the negative transfer, leading to the limited model performance. Meanwhile our proposed DTE still outperforms the comparison methods. Note that the proposed DTE with weight barcode estimation can fully leverage knowledge from source domain to a significant degree compared with JPOT. Moreover, we conduct the experiments on VisDA-2017 and Digits and report the results in Table. 3. It is remarkable that our proposed DTE can pro", + "bbox": [ + 75, + 719, + 470, + 901 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "mote the better results of OS and $\\mathbf{OS^{*}}$ on hard transfer tasks in VisDA-2017. Specifically, in terms of the performance of all classes, DTE has achieved significant improvements compared to the previous models, with notable examples being a $2.27\\%$ improvement on Office-Home and a $3.15\\%$ improvement on VisDA-2017 against runner-up models for $\\mathbf{OS^{*}}$ . DTE can better distinguish the unknown target samples meanwhile correctly assign pseudo labels for the rest target samples to reach state-of-the-art performance.", + "bbox": [ + 496, + 693, + 890, + 830 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "4.3. Analysis.", + "text_level": 1, + "bbox": [ + 500, + 844, + 606, + 859 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Ablation Study. To further study how does each module of DTE contribute on the final performance, we com", + "bbox": [ + 498, + 869, + 890, + 902 + ], + "page_idx": 6 + }, + { + "type": "page_number", + "text": "4933", + "bbox": [ + 482, + 944, + 514, + 955 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/12bb83a3c026491d4597acb58d1810f3a2d186e007ce2ec0f30abdc33847d5f5.jpg", + "image_caption": [ + "(a) Openness on Digits" + ], + "image_footnote": [], + "bbox": [ + 83, + 87, + 263, + 195 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/e298eb90407f6ba113085d004d381ae78a4d684283ddae5876075a7a37606652.jpg", + "image_caption": [ + "(b) Openness on Office-31" + ], + "image_footnote": [], + "bbox": [ + 282, + 87, + 464, + 196 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/034a113e41586505f4f06265bab5afe6f5b099ff5f274df858022a57c16179dd.jpg", + "image_caption": [ + "Figure 6. The openness experiments of $\\mathbb{O}$ on Digits and Office-31.", + "(a) Parameters on Digits", + "Figure 7. Hyper parameter tuning of $\\epsilon$ on Digits and Office-31." + ], + "image_footnote": [], + "bbox": [ + 81, + 244, + 267, + 353 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/90f713d26bc0db61faf8be3c5f15ed126d9e287bd656092f2a170b1c25f1975b.jpg", + "image_caption": [ + "(b) Parameters on Office-31" + ], + "image_footnote": [], + "bbox": [ + 281, + 244, + 465, + 353 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "pare DTE with its several variants, including DTE-W and DTE-S. DTE-W only replaces the proposed weight barcode estimation with JPOT by calculating mean transport cost to determine the unknown target samples. DTE-S only replaces sparse label assignment with self-labelling strategy via Sinkhorn algorithm [3]. The comparison results on Office-Home and Office-31 are shown in Table. 2 and Table. 1. From it, we can observe that (1) Although DTE-W exceeds most of previous methods, it cannot precisely distinguish unknown data samples especially under source-free domain adaptation scenario. (2) DTE-S achieves better results than DTE-W, while it still cannot reach more satisfactory results than DTE. That is because involving entropy regularization term could reach relatively dense solution, thus leading to the inaccurate solutions. (3) DTE achieves the best performance which indicates the efficacy of incorporating weight barcode estimation and sparse label assignment for distinguishing then exploiting target samples.", + "bbox": [ + 76, + 400, + 467, + 672 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Feature Visualization. We further provide the feature visualization via T-SNE [38] of the target sample representations $\\pmb{x}^t$ by JPOT and DTE on $\\mathbf{W} \\rightarrow \\mathbf{A}$ (Office-31). The results are depicted in Fig. 5(a)-(b), where each color represents a specific class, and the reddish-brown color is used to represent the unknown class. From there we can observe that JPOT cannot better distinguish known and unknown target samples with ambiguous decision boundary. The primary reason is that JPOT can only utilizes limited source knowledge for open set domain adaptation. Meanwhile DTE can better distinguish unknown target samples and cluster the rest known samples accordingly.", + "bbox": [ + 76, + 672, + 467, + 853 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Openness. To verify the proposed DTE can be applied to the case with different proportion of unknown classes, we adjust the openness $\\mathbb{O}$ in the openset domain adaptation.", + "bbox": [ + 76, + 854, + 467, + 900 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Specifically, $\\mathbb{O} = 1 - \\frac{|\\mathcal{C}^s|}{|\\mathcal{C}^t|}$ denotes different level of openness [46]. Larger openness indicates that fewer classes are known across domains. We vary $\\mathbb{O}$ from 0 to nearly 1 on Office-31 and Digits accordingly [46] then report the result of OS in Fig. 6(a)-(b). We can observe that the majority of previous methods (e.g., DANN and JPOT) cannot better adjust to different proportion of unknown classes in the target domain. Meanwhile, the proposed DTE can always reach better performance even openness $\\mathbb{O}$ approaches to 0 or 1. It further demonstrates the proposed DTE is robust to openness without the requirement of prior knowledge.", + "bbox": [ + 496, + 88, + 890, + 258 + ], + "page_idx": 7 + }, + { + "type": "table", + "img_path": "images/dd44aadeaf5f36ab979fb2365c08bffceee330d4c5ae67bf6b9eb94032512fcc.jpg", + "table_caption": [ + "Table 4. The results (%) for method extension on VisDA2017" + ], + "table_footnote": [], + "table_body": "
MethodbicyclebuscarmotorcycletraintruckOSOS*
PGL93.593.875.798.896.238.580.782.8
SFPGL93.697.689.695.396.795.691.094.7
DTE(VGGNet+GNN)93.594.077.296.496.041.380.283.1
DTE(ResNet+GNN)93.897.990.197.996.592.091.694.7
", + "bbox": [ + 504, + 282, + 890, + 342 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Method Extension. Furthermore, we analyse the general extension of our proposed DTE. Specifically, we incorporate weight barcode estimation and sparse label assignment with VGGNet/ResNet and Graph Neural Network (GNN for transductive feature learning [64, 89]) on VisDA2017. The results of OS and $\\mathbf{OS^{*}}$ are shown in Table 4. We observe that the proposed DTE can be applied with various backbones for achieving better performance.", + "bbox": [ + 496, + 349, + 890, + 469 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Hyper Parameter. We finally study the effects of hyperparameters on model performance. For DTE, we vary $\\epsilon$ in $\\{0.01, 0.05, 0.1, 0.5, 1, 10\\}$ for sparse label assignment on Digits/Office-31 and report the results of OS/OS* in Fig. 7(a)-(b). We can conclude that when $\\epsilon$ is larger, the proximal regularization term will have a more important impact compared to sparse-label matching, which can result in poorer performance. Hence we set $\\epsilon = 0.5$ in the experiments empirically.", + "bbox": [ + 496, + 470, + 890, + 607 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "5. Conclusion", + "text_level": 1, + "bbox": [ + 500, + 619, + 617, + 635 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "In this paper, we propose Distinguish Then Exploit model (DTE) for tackling the source-free open set domain adaptation problem. DTE first adopts weight barcode estimation with proposed target sample reweighting optimization method on partial unbalanced optimal transport to determine the known and unknown target samples without prior information. Then DTE utilizes sparse sample-label matching on optimal transport to further leverage useful and accurate source knowledge in the target domain. We also conduct extensive experiments on different datasets to show the superior performance of the proposed DTE model.", + "bbox": [ + 496, + 646, + 890, + 811 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "6. Acknowledgment", + "text_level": 1, + "bbox": [ + 500, + 829, + 669, + 845 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "This work was supported by the National Natural Science Foundation of China (No. 62372242) and Natural Science Foundation of Shandong Province (No. ZR2023MF007).", + "bbox": [ + 496, + 854, + 890, + 900 + ], + "page_idx": 7 + }, + { + "type": "page_number", + "text": "4934", + "bbox": [ + 482, + 944, + 514, + 955 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "References", + "text_level": 1, + "bbox": [ + 78, + 89, + 173, + 104 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[1] Ravindra K Ahuja, Thomas L Magnanti, and James B Orlin. Network flows. 1988. 5", + "[2] Hana Ajakan, Pascal Germain, Hugo Larochelle, François Laviolette, and Mario Marchand. Domain-adversarial neural networks. arXiv preprint arXiv:1412.4446, 2014. 2, 6", + "[3] Yuki Markus Asano, Christian Rupprecht, and Andrea Vedaldi. Self-labelling via simultaneous clustering and representation learning. arXiv preprint arXiv:1911.05371, 2019. 5, 8", + "[4] Yikun Bai, Bernhard Schmitzer, Matthew Thorpe, and Soheil Kolouri. Sliced optimal partial transport. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 13681-13690, 2023. 2", + "[5] Abhijit Bendale and Terrance E Boult. Towards open set deep networks. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 1563-1572, 2016. 6", + "[6] Mathieu Blondel, Vivien Seguy, and Antoine Rolet. Smooth and sparse optimal transport. In International conference on artificial intelligence and statistics, pages 880-889. PMLR, 2018. 5", + "[7] Nicolas Bonneel and David Coeurjolly. Spot: sliced partial optimal transport. ACM Transactions on Graphics (TOG), 38(4):1-13, 2019. 2", + "[8] Silvia Bucci, Mohammad Reza Loghmani, and Tatiana Tommasi. On the effectiveness of image rotation for open set domain adaptation. In European conference on computer vision, pages 422-438. Springer, 2020. 2", + "[9] Nicolas Courty, Rémi Flamary, and Devis Tuia. Domain adaptation with regularized optimal transport. In Joint European Conference on Machine Learning and Knowledge Discovery in Databases, pages 274-289. Springer, 2014. 5", + "[10] Marco Cuturi. Sinkhorn distances: Lightspeed computation of optimal transport. Advances in neural information processing systems, 26, 2013. 5", + "[11] Bharath Bhushan Damodaran, Benjamin Kellenberger, Rémi Flamary, Devis Tuia, and Nicolas Courty. Deepjdot: Deep joint distribution optimal transport for unsupervised domain adaptation. In European Conference on Computer Vision, pages 467-483. Springer, 2018. 6", + "[12] Jun Dan, Weiming Liu, Chunfeng Xie, Hua Yu, Shunjie Dong, and Yanchao Tan. Tfgda: Exploring topology and feature alignment in semi-supervised graph domain adaptation through robust clustering. In The Thirty-eighth Annual Conference on Neural Information Processing Systems. 1", + "[13] Jun Dan, Tao Jin, Hao Chi, Shunjie Dong, and Yixuan Shen. Uncertainty-guided joint unbalanced optimal transport for unsupervised domain adaptation. Neural Computing and Applications, 35(7):5351-5367, 2023. 2", + "[14] Jun Dan, Tao Jin, Hao Chi, Shunjie Dong, Haoran Xie, Keying Cao, and Xinjing Yang. Trust-aware conditional adversarial domain adaptation with feature norm alignment. Neural Networks, 168:518-530, 2023. 2", + "[15] Jun Dan, Tao Jin, Hao Chi, Yixuan Shen, Jiawang Yu, and Jinhai Zhou. Homda: High-order moment-based domain" + ], + "bbox": [ + 78, + 114, + 468, + 900 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "alignment for unsupervised domain adaptation. Knowledge-Based Systems, 261:110205, 2023. 1", + "[16] Jun Dan, Yang Liu, Haoyu Xie, Jiankang Deng, Haoran Xie, Xuansong Xie, and Baigui Sun. Transface: Calibrating transformer training for face recognition from a data-centric perspective. In Proceedings of the IEEE/CVF international conference on computer vision, pages 20642-20653, 2023. 1", + "[17] Jun Dan, Mushui Liu, Chunfeng Xie, Jiawang Yu, Haoran Xie, Ruokun Li, and Shunjie Dong. Similar norm more transferable: Rethinking feature norms discrepancy in adversarial domain adaptation. Knowledge-Based Systems, 296: 111908, 2024. 2", + "[18] Jun Dan, Weiming Liu, Mushui Liu, Chunfeng Xie, Shunjie Dong, Guofang Ma, Yanchao Tan, and Jiazheng Xing. Hogda: Boosting semi-supervised graph domain adaptation via high-order structure-guided adaptive feature alignment. In Proceedings of the 32nd ACM International Conference on Multimedia, pages 11109-11118, 2024. 1", + "[19] Jun Dan, Yang Liu, Jiankang Deng, Haoyu Xie, Siyuan Li, Baigui Sun, and Shan Luo. Topofr: A closer look at topology alignment on face recognition. arXiv preprint arXiv:2410.10587, 2024. 1", + "[20] Yuqi Fang, Pew-Thian Yap, Weili Lin, Hongtu Zhu, and Mingxia Liu. Source-free unsupervised domain adaptation: A survey. arXiv preprint arXiv:2301.00265, 2022. 3", + "[21] Zhen Fang, Jie Lu, Feng Liu, Junyu Xuan, and Guangquan Zhang. Open set domain adaptation: Theoretical bound and algorithm. IEEE transactions on neural networks and learning systems, 32(10):4309-4322, 2020. 3", + "[22] Abolfazl Farahani, Sahar Voghoei, Khaled Rasheed, and Hamid R Arabnia. A brief review of domain adaptation. Advances in data science and information engineering: proceedings from ICDATA 2020 and IKE 2020, pages 877-894, 2021. 2", + "[23] Kilian Fatras, Thibault Séjourne, Rémi Flamary, and Nicolas Courty. Unbalanced minibatch optimal transport; applications to domain adaptation. In International Conference on Machine Learning, pages 3186-3197. PMLR, 2021. 2", + "[24] Jean Feydy, Thibault Séjourne, François-Xavier Vialard, Shun-ichi Amari, Alain Trouve, and Gabriel Peyre. Interpolating between optimal transport and mmd using sinkhorn divergences. In The 22nd International Conference on Artificial Intelligence and Statistics, pages 2681–2690. PMLR, 2019. 2", + "[25] Rémi Flamary, Nicholas Courty, Davis Tuia, and Alain Rakotomamonjy. Optimal transport for domain adaptation. IEEE Trans. Pattern Anal. Mach. Intell, 1:1-40, 2016. 5", + "[26] Rémi Flamary, Nicolas Courty, Alexandre Gramfort, Mokhtar Z Alaya, Aurélie Boisbunon, Stanislas Chambon, Laetitia Chapel, Adrien Corenflos, Kilian Fatras, Nemo Fournier, et al. Pot: Python optimal transport. The Journal of Machine Learning Research, 22(1):3571-3578, 2021. 2", + "[27] Yves Grandvalet and Yoshua Bengio. Entropy regularization., 2006. 5", + "[28] John A Hartigan and Manchek A Wong. Algorithm as 136: A k-means clustering algorithm. Journal of the royal sta" + ], + "bbox": [ + 501, + 92, + 890, + 900 + ], + "page_idx": 8 + }, + { + "type": "page_number", + "text": "4935", + "bbox": [ + 482, + 944, + 514, + 955 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "tistical society. series c (applied statistics), 28(1):100-108, 1979. 5", + "[29] Kaiming He, Xiangyu Zhang, Shaoqing Ren, and Jian Sun. Deep residual learning for image recognition. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 770-778, 2016. 6", + "[30] Zi-Xian Huang and Chuan-Xian Ren. Rethinking correlation learning via label prior for open set domain adaptation. 6", + "[31] J. J. Hull. A database for handwritten text recognition research. IEEE Transactions on Pattern Analysis & Machine Intelligence, 16(5):550-554, 2002. 6, 15", + "[32] JoonHo Jang, Byeonghu Na, Dong Hyeok Shin, Mingi Ji, Kyungwoo Song, and Il-Chul Moon. Unknown-aware domain adversarial learning for open-set domain adaptation. Advances in Neural Information Processing Systems, 35: 16755-16767, 2022. 6", + "[33] Jun-Gi Jang, Sooyeon Shim, Vladimir Egay, Jeeyong Lee, Jongmin Park, Suhyun Chae, and U Kang. Accurate open-set recognition for memory workload. ACM Transactions on Knowledge Discovery from Data, 17(9):1-14, 2023. 2", + "[34] Jeff L Kennington and Richard V Helgason. Algorithms for network programming. John Wiley & Sons, Inc., 1980. 5", + "[35] Diederik P Kingma and Jimmy Ba. Adam: A method for stochastic optimization. arXiv preprint arXiv:1412.6980, 2014. 6", + "[36] Jogendra Nath Kundu, Naveen Venkat, R Venkatesh Babu, et al. Universal source-free domain adaptation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 4544-4553, 2020. 2", + "[37] Jogendra Nath Kundu, Naveen Venkat, Ambareesh Revanur, R Venkatesh Babu, et al. Towards inheritable models for open-set domain adaptation. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 12376-12385, 2020. 6", + "[38] Van Der Maaten Laurens and Geoffrey Hinton. Visualizing data using t-sne. Journal of Machine Learning Research, 9 (2605):2579-2605, 2008. 8", + "[39] Khang Le, Huy Nguyen, Quang M Nguyen, Tung Pham, Hung Bui, and Nhat Ho. On robust optimal transport: Computational complexity and barycenter computation. Advances in Neural Information Processing Systems, 34:21947-21959, 2021. 2", + "[40] Y. L. Lecun, Leon Bottou, Yoshua Bengio, and Patrick Haffner. Gradient-based learning applied to document recognition. proc. ii. Proceedings of the IEEE, 86(11):2278-2324, 1998. 6, 15", + "[41] Na Lei and Xianfeng Gu. Computational Conformal Geometry. International Press and Higher Education Press, 2021. 4", + "[42] Bo Li, Yifei Shen, Jingkang Yang, Yezhen Wang, Jiawei Ren, Tong Che, Jun Zhang, and Ziwei Liu. Sparse mixture-of-experts are domain generalizable learners. arXiv preprint arXiv:2206.04046, 2022. 2", + "[43] Wuyang Li, Jie Liu, Bo Han, and Yixuan Yuan. Adjustment and alignment for unbiased open set domain adaptation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 24110-24119, 2023. 6" + ], + "bbox": [ + 78, + 90, + 468, + 900 + ], + "page_idx": 9 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[44] Jian Liang, Dapeng Hu, and Jiashi Feng. Do we really need to access the source data? source hypothesis transfer for unsupervised domain adaptation. In International conference on machine learning, pages 6028-6039. PMLR, 2020. 3, 6", + "[45] Mattia Litrico, Alessio Del Bue, and Pietro Morerio. Guiding pseudo-labels with uncertainty estimation for source-free unsupervised domain adaptation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 7640-7650, 2023. 5", + "[46] Hong Liu, Zhangjie Cao, Mingsheng Long, Jianmin Wang, and Qiang Yang. Separate to adapt: Open set domain adaptation via progressive separation. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 2927-2936, 2019. 2, 6, 8, 15", + "[47] Jie Liu, Xiaoqing Guo, and Yixuan Yuan. Unknown-oriented learning for open set domain adaptation. In European Conference on Computer Vision, pages 334-350. Springer, 2022. 2, 6", + "[48] Mushui Liu, Yunlong Yu, Zhong Ji, Jungong Han, and Zhongfei Zhang. Tolerant self-distillation for image classification. Neural Networks, 174:106215, 2024. 1", + "[49] Mushui Liu, Weijie He, Ziqian Lu, Jun Dan, Yunlong Yu, Yingming Li, Xi Li, and Jungong Han. Synth-clip: Synthetic data make clip generalize better in data-limited scenarios. Neural Networks, 184:107083, 2025. 2", + "[50] Tianlin Liu, Joan Puigcerver, and Mathieu Blondel. Sparsity-constrained optimal transport. *ICLR*, 2023. 5", + "[51] Weiming Liu, Xiaolin Zheng, Jiajie Su, Mengling Hu, Yanchao Tan, and Chaochao Chen. Exploiting variational domain-invariant user embedding for partially overlapped cross domain recommendation. In Proceedings of the 45th International ACM SIGIR conference on research and development in information retrieval, pages 312-321, 2022. 1", + "[52] Weiming Liu, Xiaolin Zheng, Chaochao Chen, Jiajie Su, Xinting Liao, Mengling Hu, and Yanchao Tan. Joint internal multi-interest exploration and external domain alignment for cross domain sequential recommendation. In Proceedings of the ACM web conference 2023, pages 383-394, 2023.", + "[53] Weiming Liu, Chaochao Chen, Xinting Liao, Mengling Hu, Jiajie Su, Yanchao Tan, and Fan Wang. User distribution mapping modelling with collaborative filtering for cross-domain recommendation. In Proceedings of the ACM Web Conference 2024, pages 334-343, 2024.", + "[54] Weiming Liu, Chaochao Chen, Xinting Liao, Mengling Hu, Yanchao Tan, Fan Wang, Xiaolin Zheng, and Yew Soon Ong. Learning accurate and bidirectional transformation via dynamic embedding transportation for cross-domain recommendation. In Proceedings of the AAAI Conference on Artificial Intelligence, pages 8815-8823, 2024. 1", + "[55] Weiming Liu, Xiaolin Zheng, Chaochao Chen, Jiahe Xu, Xinting Liao, Fan Wang, Yanchao Tan, and Yew-Soon Ong. Reducing item discrepancy via differentially private robust embedding alignment for privacy-preserving cross domain recommendation. In *Forty-first International Conference on Machine Learning*, 2024. 4", + "[56] Ye Liu, Lingfeng Qiao, Changchong Lu, Di Yin, Chen Lin, Haoyuan Peng, and Bo Ren. Osman: A one-stage alignment" + ], + "bbox": [ + 501, + 92, + 893, + 900 + ], + "page_idx": 9 + }, + { + "type": "page_number", + "text": "4936", + "bbox": [ + 482, + 945, + 514, + 955 + ], + "page_idx": 9 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "network to unify multimodal alignment and unsupervised domain adaptation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 3551-3560, 2023. 1", + "[57] Yang Liu, Zhipeng Zhou, and Baigui Sun. Cot: Unsupervised domain adaptation with clustering and optimal transport. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 19998-20007, 2023. 1", + "[58] Mingsheng Long, Jianmin Wang, Guiguang Ding, Jiaguang Sun, and Philip S Yu. Transfer feature learning with joint distribution adaptation. In Proceedings of the IEEE international conference on computer vision, pages 2200-2207, 2013. 1", + "[59] Mingsheng Long, Yue Cao, Jianmin Wang, and Michael Jordan. Learning transferable features with deep adaptation networks. In International Conference on Machine Learning, pages 97-105, 2015. 1, 2", + "[60] Mingsheng Long, Han Zhu, Jianmin Wang, and Michael I Jordan. Unsupervised domain adaptation with residual transfer networks. Advances in neural information processing systems, 29, 2016. 1", + "[61] Mingsheng Long, Han Zhu, Jianmin Wang, and Michael I Jordan. Deep transfer learning with joint adaptation networks. In Proceedings of the 34th International Conference on Machine Learning-Volume 70, pages 2208-2217. JMLR.org, 2017. 1, 2", + "[62] Mingsheng Long, Zhangjie Cao, Jianmin Wang, and Michael I Jordan. Conditional adversarial domain adaptation. In Advances in Neural Information Processing Systems, pages 1640-1650, 2018. 2", + "[63] Yadan Luo, Zijian Wang, Zi Huang, and Mahsa Baktashmotlagh. Progressive graph learning for open-set domain adaptation. In International Conference on Machine Learning, pages 6468-6478. PMLR, 2020. 6", + "[64] Yadan Luo, Zijian Wang, Zhuoxiao Chen, Zi Huang, and Mahsa Baktashmotlagh. Source-free progressive graph learning for open-set domain adaptation. IEEE Transactions on Pattern Analysis and Machine Intelligence, 2023. 3, 6, 8", + "[65] You-Wei Luo and Chuan-Xian Ren. Conditional bures metric for domain adaptation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 13989–13998, 2021. 1", + "[66] You-Wei Luo and Chuan-Xian Ren. Mot: Masked optimal transport for partial domain adaptation. In 2023 IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 3531–3540. IEEE, 2023. 1", + "[67] Yuval Netzer, Tao Wang, Adam Coates, Alessandro Bissacco, Bo Wu, and Andrew Y. Ng. Reading digits in natural images with unsupervised feature learning. Nips Workshop on Deep Learning & Unsupervised Feature Learning, 2011. 6, 15", + "[68] Khai Nguyen, Dang Nguyen, Quoc Nguyen, Tung Pham, Hung Bui, Dinh Phung, Trung Le, and Nhat Ho. On transportation of mini-batches: A hierarchical approach. arXiv preprint arXiv:2102.05912, 2021. 2" + ], + "bbox": [ + 78, + 90, + 468, + 900 + ], + "page_idx": 10 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[69] Pau Panareda Busto and Juergen Gall. Open set domain adaptation. In Proceedings of the IEEE international conference on computer vision, pages 754-763, 2017. 1, 2", + "[70] Xingchao Peng, Ben Usman, Neela Kaushik, Dequan Wang, Judy Hoffman, and Kate Saenko. Visda: A synthetic-to-real benchmark for visual domain adaptation. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition Workshops, pages 2021-2026, 2018. 6, 15", + "[71] Gabriel Peyré. Course notes on computational optimal transport. 2019. 4", + "[72] Khiem Pham, Khang Le, Nhat Ho, Tung Pham, and Hung Bui. On unbalanced optimal transport: An analysis of sinkhorn algorithm. In International Conference on Machine Learning, pages 7673-7682. PMLR, 2020. 2", + "[73] Sanqing Qu, Tianpei Zou, Florian Röhrbein, Cewu Lu, Guang Chen, Dacheng Tao, and Changjun Jiang. Upcycling models under domain and category shift. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 20019-20028, 2023. 2", + "[74] Sanqing Qu, Tianpei Zou, Lianghua He, Florian Röhrbein, Alois Knoll, Guang Chen, and Changjun Jiang. Lead: Learning decomposition for source-free universal domain adaptation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 23334-23343, 2024. 2", + "[75] Ievgen Redko, Amaury Habrard, and Marc Sebban. Theoretical analysis of domain adaptation with optimal transport. In Machine Learning and Knowledge Discovery in Databases: European Conference, ECML PKDD 2017, Skopje, Macedonia, September 18-22, 2017, Proceedings, Part II 10, pages 737-753. Springer, 2017. 5", + "[76] Ievgen Redko, Nicolas Courty, Rémi Flamary, and Devis Tuia. Optimal transport for multi-source domain adaptation under target shift. In The 22nd International Conference on artificial intelligence and statistics, pages 849-858. PMLR, 2019. 2", + "[77] Kate Saenko, Brian Kulis, Mario Fritz, and Trevor Darrell. Adapting visual category models to new domains. In European Conference on Computer Vision (ECCV), 2010. 6, 15", + "[78] Kuniaki Saito, Shohei Yamamoto, Yoshitaka Ushiku, and Tatsuya Harada. Open set domain adaptation by backpropagation. In Proceedings of the European conference on computer vision (ECCV), pages 153-168, 2018. 1, 2, 6, 16", + "[79] Thibault Séjourne, François-Xavier Vialard, and Gabriel Peyré. Faster unbalanced optimal transport: Translation invariant sinkhorn and 1-d frank-wolfe. In International Conference on Artificial Intelligence and Statistics, pages 4995-5021. PMLR, 2022. 13", + "[80] Jian Shen, Yanru Qu, Weinan Zhang, and Yong Yu. Wasserstein distance guided representation learning for domain adaptation. arXiv preprint arXiv:1707.01217, 2017. 2", + "[81] Rui Shu, Hung H Bui, Hirokazu Narui, and Stefano Ermon. A dirt-t approach to unsupervised domain adaptation. arXiv preprint arXiv:1802.08735, 2018. 2, 6", + "[82] Baochen Sun and Kate Saenko. Deep coral: Correlation alignment for deep domain adaptation. In European Conference on Computer Vision, pages 443-450. Springer, 2016. 2" + ], + "bbox": [ + 501, + 92, + 890, + 898 + ], + "page_idx": 10 + }, + { + "type": "page_number", + "text": "4937", + "bbox": [ + 482, + 945, + 514, + 955 + ], + "page_idx": 10 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[83] Eric Tzeng, Judy Hoffman, Ning Zhang, Kate Saenko, and Trevor Darrell. Deep domain confusion: Maximizing for domain invariance. arXiv preprint arXiv:1412.3474, 2014. 2", + "[84] Hemanth Venkateswara, Jose Eusebio, Shayok Chakraborty, and Sethuraman Panchanathan. Deep hashing network for unsupervised domain adaptation. CoRR, abs/1706.07522, 2017. 6, 15", + "[85] Cédric Villani et al. Optimal transport: old and new. Springer, 2009. 2", + "[86] Mei Wang and Weihong Deng. Deep visual domain adaptation: A survey. Neurocomputing, 312:135-153, 2018. 2", + "[87] Qian Wang, Fanlin Meng, and Toby P Breckon. Progressively select and reject pseudo-labelled samples for open-set domain adaptation. arXiv preprint arXiv:2110.12635, 2021. 2", + "[88] Garrett Wilson and Diane J Cook. A survey of unsupervised deep domain adaptation. ACM Transactions on Intelligent Systems and Technology (TIST), 11(5):1-46, 2020. 2", + "[89] Zonghan Wu, Shirui Pan, Fengwen Chen, Guodong Long, Chengqi Zhang, and S Yu Philip. A comprehensive survey on graph neural networks. IEEE transactions on neural networks and learning systems, 32(1):4-24, 2020. 8", + "[90] Yujia Xie, Xiangfeng Wang, Ruijia Wang, and Hongyuan Zha. A fast proximal point method for computing exact Wasserstein distance. In Uncertainty in artificial intelligence, pages 433-453. PMLR, 2020. 5", + "[91] Renjun Xu, Pelen Liu, Yin Zhang, Fang Cai, Jindong Wang, Shuoying Liang, Heting Ying, and Jianwei Yin. Joint partial optimal transport for open set domain adaptation. In IJCAI, pages 2540-2546, 2020. 1, 2, 6", + "[92] Tengyue Xu and Jun Dan. Ehm: Exploring dynamic alignment and hierarchical clustering in unsupervised domain adaptation via high-order moment-guided contrastive learning. Neural Networks, page 107188, 2025. 5", + "[93] Hongliang Yan, Yukang Ding, Peihua Li, Qilong Wang, Yong Xu, and Wangmeng Zuo. Mind the class weight bias: Weighted maximum mean discrepancy for unsupervised domain adaptation. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition, pages 2272-2281, 2017. 2", + "[94] Yu-Chu Yu and Hsuan-Tien Lin. Semi-supervised domain adaptation with source label adaptation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 24100-24109, 2023. 1", + "[95] Zhiqi Yu, Jingjing Li, Zhekai Du, Lei Zhu, and Heng Tao Shen. A comprehensive survey on source-free domain adaptation. arXiv preprint arXiv:2302.11803, 2023. 3", + "[96] Li Zhong, Zhen Fang, Feng Liu, Bo Yuan, Guangquan Zhang, and Jie Lu. Bridging the theoretical bound and deep algorithms for open set domain adaptation. IEEE transactions on neural networks and learning systems, 2021. 2, 3", + "[97] Jingyu Zhuang, Ziliang Chen, Pengxu Wei, Guanbin Li, and Liang Lin. Discovering implicit classes achieves open set domain adaptation. In 2022 IEEE International Conference on Multimedia and Expo (ICME), pages 01-06. IEEE, 2022. 6" + ], + "bbox": [ + 78, + 90, + 468, + 882 + ], + "page_idx": 11 + }, + { + "type": "page_number", + "text": "4938", + "bbox": [ + 482, + 945, + 514, + 955 + ], + "page_idx": 11 + } +] \ No newline at end of file diff --git a/2025/Distinguish Then Exploit_ Source-free Open Set Domain Adaptation via Weight Barcode Estimation and Sparse Label Assignment/099209ec-4ec9-4836-8000-10ddecb15989_model.json b/2025/Distinguish Then Exploit_ Source-free Open Set Domain Adaptation via Weight Barcode Estimation and Sparse Label Assignment/099209ec-4ec9-4836-8000-10ddecb15989_model.json new file mode 100644 index 0000000000000000000000000000000000000000..e5ce672cdfc6f44c46e4a8e70ce317a648740e4e --- /dev/null +++ b/2025/Distinguish Then Exploit_ Source-free Open Set Domain Adaptation via Weight Barcode Estimation and Sparse Label Assignment/099209ec-4ec9-4836-8000-10ddecb15989_model.json @@ -0,0 +1,2897 @@ +[ + [ + { + "type": "header", + "bbox": [ + 0.107, + 0.003, + 0.182, + 0.043 + ], + "angle": 0, + "content": "CVF" + }, + { + "type": "header", + "bbox": [ + 0.237, + 0.0, + 0.812, + 0.045 + ], + "angle": 0, + "content": "This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore." + }, + { + "type": "title", + "bbox": [ + 0.086, + 0.13, + 0.885, + 0.175 + ], + "angle": 0, + "content": "Distinguish Then Exploit: Source-free Open Set Domain Adaptation via Weight Barcode Estimation and Sparse Label Assignment" + }, + { + "type": "text", + "bbox": [ + 0.228, + 0.202, + 0.78, + 0.239 + ], + "angle": 0, + "content": "Weiming Liu\\(^{1*}\\), Jun Dan\\(^{2*}\\), Fan Wang\\(^{2}\\), Xinting Liao\\(^{2}\\), Junhao Dong\\(^{3}\\), Hua Yu\\(^{3}\\), Shunjie Dong\\(^{4}\\), Lianyong Qi\\(^{56\\dagger}\\)" + }, + { + "type": "text", + "bbox": [ + 0.089, + 0.24, + 0.923, + 0.274 + ], + "angle": 0, + "content": "1Bytedance Inc. 2Zhejiang University 3Nanyang Technological University 4Shanghai Jiao Tong University 5China University of Petroleum (East China)" + }, + { + "type": "text", + "bbox": [ + 0.223, + 0.275, + 0.788, + 0.292 + ], + "angle": 0, + "content": "\\(^{6}\\)Shandong Key Laboratory of Intelligent Oil & Gas Industrial Software" + }, + { + "type": "text", + "bbox": [ + 0.138, + 0.295, + 0.872, + 0.327 + ], + "angle": 0, + "content": "lwming95@gmail.com, danjun@zju.edu.cn, fanwang97@zju.edu.cn, xintingliao@zju.edu.cn, junhao003@ntu.edu.sg, yu_hua@ntu.edu.sg, sjdong@sjtu.edu.cn, lianyongqi@gmail.com" + }, + { + "type": "title", + "bbox": [ + 0.235, + 0.362, + 0.314, + 0.379 + ], + "angle": 0, + "content": "Abstract" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.394, + 0.473, + 0.786 + ], + "angle": 0, + "content": "Nowadays, domain adaptation techniques have been widely investigated for knowledge sharing from labeled source domain to unlabeled target domain. However, target domain may include some data samples that belong to unknown categories in real-world scenarios. Moreover, the target domain cannot access the source data samples due to privacy-preserving restrictions. In this paper, we focus on the source-free open set domain adaptation problem which includes two main challenges, i.e., how to distinguish known and unknown target samples and how to exploit useful source information to provide trustworthy pseudo labels for known target samples. Existing approaches that directly apply conventional domain alignment methods could lead to sample mismatch and misclassification in this scenario. To overcome these issues, we propose a Distinguish Then Exploit model (DTE) with two components, i.e., weight barcode estimation and sparse label assignment. Weight barcode estimation first calculates the marginal probability of target samples via partially unbalanced optimal transport, then quantize barcode results to distinguish unknown target samples. Sparse label assignment utilizes sparse sample-label matching via proximal term to fully exploit useful source information. Our empirically study on several datasets shows that DTE outperforms the state-of-the-art models on tackling the source-free open set domain adaptation problem." + }, + { + "type": "title", + "bbox": [ + 0.078, + 0.808, + 0.21, + 0.823 + ], + "angle": 0, + "content": "1. Introduction" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.834, + 0.47, + 0.879 + ], + "angle": 0, + "content": "Nowadays, deep learning technique has been successfully adopted to image classification tasks with the help of large-scale labeled datasets [16, 19, 48, 51-54, 57, 65, 66]. How" + }, + { + "type": "text", + "bbox": [ + 0.5, + 0.364, + 0.892, + 0.501 + ], + "angle": 0, + "content": "ever, it is time-consuming to sufficiently annotate data for providing satisfied results. A promising solution is to leverage useful information from the labeled source domain to assign reliable labels within the unlabeled target domain [18, 58, 60]. Nevertheless, the data distribution could be significantly different across domains, leading to the model degradation on final performance [12, 56, 94]. How to reduce the domain bias and discrepancy while avoiding negative transfer across domains has become an urgent problem." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.504, + 0.893, + 0.701 + ], + "angle": 0, + "content": "In this paper, we focus on a more general problem of source-free open set domain adaptation. That is, there could exist some target samples belong to extra categories that does not present in the source domain. Moreover, due to data privacy protection constraints, the target domain could only have access to the well-trained model in the source domain. Such settings are commonly encountered in real-world scenarios and it is rather complicated than conventional domain adaptation problem. Specifically, there are two main challenges: CH1: How can we enhance the ability to distinguish target samples that belong to unknown categories? CH2: How to better exploit appropriate pseudo labels for the remaining known target samples?" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.704, + 0.895, + 0.902 + ], + "angle": 0, + "content": "Existing researches cannot solve source-free open set domain adaptation well. Firstly, conventional transfer learning always assumes that both source and target domains share the same set of label categories [15, 59, 61]. Directly aligning the latent distributions of source and target domains may lead to negative transfer, where target unknown samples are incorrectly matched to source samples. Secondly, traditional open set domain adaptation approaches should incorporate well-tuned thresholds for determining the decision boundary between known and unknown target samples [69]. This suggests that many of these methods (e.g., OSBP [78] and JPOT [91]) typically rely on prior knowledge about the target domain, which will limit their potential applicabil" + }, + { + "type": "page_footnote", + "bbox": [ + 0.1, + 0.887, + 0.352, + 0.9 + ], + "angle": 0, + "content": "* Equal Contribution, † Corresponding Author." + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.516, + 0.957 + ], + "angle": 0, + "content": "4927" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.095, + 0.09, + 0.241, + 0.209 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.113, + 0.211, + 0.227, + 0.22 + ], + "angle": 0, + "content": "(a) Initial Target Features in" + }, + { + "type": "image_caption", + "bbox": [ + 0.143, + 0.221, + 0.198, + 0.23 + ], + "angle": 0, + "content": "Latent Space" + }, + { + "type": "image", + "bbox": [ + 0.244, + 0.091, + 0.512, + 0.209 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.28, + 0.211, + 0.481, + 0.231 + ], + "angle": 0, + "content": "(b) Adopting Weight Barcode Estimation to Distinguish Unknown Samples in Target Domain" + }, + { + "type": "image", + "bbox": [ + 0.513, + 0.091, + 0.728, + 0.209 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.519, + 0.211, + 0.728, + 0.23 + ], + "angle": 0, + "content": "(c) Adopting Sparse Label Assignment to Exploit Pseudo Label for Known Samples in Target Domain" + }, + { + "type": "image", + "bbox": [ + 0.73, + 0.091, + 0.877, + 0.21 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.741, + 0.211, + 0.86, + 0.23 + ], + "angle": 0, + "content": "(d) Target Features in Latent Space after Training" + }, + { + "type": "image_caption", + "bbox": [ + 0.076, + 0.236, + 0.895, + 0.293 + ], + "angle": 0, + "content": "Figure 1. An overview of the proposed DTE. (a) The initial target features provided in the latent space are dispersed and challenging to achieve satisfactory classification results. (b) We first adopt weight barcode estimation component to distinguish known or unknown target samples accordingly. (c) Then we adopt sparse label assignment to exploit pseudo label for known samples in the target domains. (d) Finally we can better cluster and classify target unlabeled samples. Best viewed in color." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.3, + 0.474, + 0.496 + ], + "angle": 0, + "content": "ity. If such prior knowledge (e.g., the proportion of unknown classes in the target domain) cannot be well estimated, the model performance could be degraded dramatically. Some recent works (e.g., STA [46]) further utilized coarse-to-fine weighting mechanism for separating target samples under different proportion of unknown classes in the target domain (also termed as openness [46]). However, they rely on involving domain adversarial training which could be unstable under some circumstances [14, 81]. Last but not least, we cannot access to the origin source data samples during the training procedure. Previous open set models fail to address such constraints for data privacy protection and they cannot be directly applied into source-free scenario." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.502, + 0.474, + 0.837 + ], + "angle": 0, + "content": "To address the aforementioned issues, we propose Distinguish Then Exploit (DTE) for tackling source-free open set domain adaptation problem. To better distinguish unknown target samples (solving CH1) and assign appropriate labels to the remaining target samples (solving CH2), we utilize two main components in DTE, i.e., weight barcode estimation component and sparse label assignment component. Weight barcode estimation component aims to solve CH1 by adopting marginal probability estimation and weight barcode establishment to automatically determine unknown target samples without prior knowledge. Sparse label assignment component is designed to address CH2 via leveraging the rest known target samples with proximal optimal transport techniques. By cooperating these two components, we can better utilize source knowledge in the target domain by providing more accurate results on source-free open set domain adaptation as shown in Fig.1. We first distinguish the heart-shaped and diamond-shaped unknown target samples. Then we assign the proper pseudo labels for the known target samples with the shapes of triangle, square, and circle respectively. Finally we can classify target unlabeled samples to reach better performance." + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.84, + 0.471, + 0.902 + ], + "angle": 0, + "content": "We summarize our main contributions as follows: (1) We propose a novel framework, i.e., DTE, for source-free open set domain adaptation problem via knowledge transfer across domains. (2) We innovatively propose weight bar" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.3, + 0.895, + 0.421 + ], + "angle": 0, + "content": "code estimation with partial unbalanced optimal transport to distinguish unknown target samples. Furthermore, we propose target sample reweighting optimization method to figure out the marginal probabilities on partial unbalanced optimal transport. Moreover, we further propose sparse label assignment to provide reliable labels for the known target samples. (3) Extensive empirical studies on various datasets show that DTE achieves the state-of-the-art performance." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.436, + 0.638, + 0.452 + ], + "angle": 0, + "content": "2. Related work" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.462, + 0.893, + 0.72 + ], + "angle": 0, + "content": "Closed Set Domain Adaptation. Conventional domain adaptation usually assumes that both source and target domains share the same set of labels which namely Closed Set Domain Adaptation [59, 61, 62]. There are mainly two types of existing works, i.e., discrepancy-based method and adversarial-based methods [17, 22, 86, 88]. Discrepancy-based methods usually align the statistics information (e.g., mean and covariance) across domains via Maximum Mean Discrepancy (MMD [83, 93]) or Correlation Alignment (CORAL [82]). More recently, optimal transport techniques [4, 7, 13, 26, 39, 72, 85] have been widely adopted into solving domain adaptation problem [23, 24, 68, 76, 80]. Adversarial-based methods, e.g., DANN [2], CDAN [62] and WDGRL [80] which usually involve domain discriminators during training procedure. However, these approaches could easily lead to negative transfer by mismatching the unknown target samples with the source samples." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.72, + 0.895, + 0.902 + ], + "angle": 0, + "content": "Open Set Domain Adaptation. Different from closed set domain adaptation, open set domain adaptation considers more general scenarios when some target samples may belong to unknown label categories [8, 33, 36, 42, 49, 69, 73, 74, 87, 96]. Most of classic open set domain adaptation models (e.g., OSBP [78] and JPOT [91]) set thresholds to determine known and unknown target samples via useful prior knowledge. More recently, some researches (e.g., STA [46] and UOL [47]) also utilize semantic intracategory information for excavating unknown samples with the absent of relevant knowledge. Meanwhile, with the increasing emphasis on protecting data privacy, target do" + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.516, + 0.957 + ], + "angle": 0, + "content": "4928" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.076, + 0.092, + 0.473, + 0.184 + ], + "angle": 0, + "content": "main could be relatively difficult to directly obtain source data [20, 95]. However, only a few approaches have been proposed to tackle source-free open set domain adaptation (e.g., SFPGL [64]) since it is more challenging. How to provide satisfied results on open set domain adaptation under the privacy preserving constraints is an urgent problem." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.196, + 0.258, + 0.213 + ], + "angle": 0, + "content": "3. Modelling for DTE" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.221, + 0.47, + 0.373 + ], + "angle": 0, + "content": "First, we describe notations. We assume there are two domains, i.e., a source domain \\( \\mathcal{S} = \\{(x_i^s,y_i^s)\\}_{i = 1}^{n_s} \\) with \\( n_{s} \\) labeled samples and a target domain \\( \\mathcal{T} = \\{(x_i^t)\\}_{i = 1}^{n_t} \\) with \\( n_t \\) unlabeled samples. \\( \\mathcal{C}^s \\) and \\( \\mathcal{C}^t \\) denote the label spaces in source and target domains respectively. Unlike traditional domain adaptation tasks, there exist target samples that belong to additional unknown classes \\( \\mathcal{C}_{\\mathrm{unk}}^t \\) in the target domain. Specifically, the label space in target domain can be formulated as \\( \\mathcal{C}^t = \\mathcal{C}^s\\cup \\mathcal{C}_{\\mathrm{unk}}^t \\). We aim to correctly assign labels for the target samples via knowledge transferring." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.373, + 0.471, + 0.51 + ], + "angle": 0, + "content": "Since the increasing law on data privacy protection, the source data samples cannot be delivered to the target domain. Hence, the target domain can only access the well-trained source model for tackling the source-free openset domain adaptation. We first adopt CNN architecture in source feature extractor \\(\\mathcal{F}_s(\\cdot)\\) to pretrain on the source data samples to obtain the source features \\(\\mathcal{F}_s(\\boldsymbol{x}^s) = \\boldsymbol{z}^s\\) following [44]. After that we adopt source classifier \\(\\mathcal{G}(\\cdot)\\) to make output prediction \\(\\pmb{p}^s (y_k|\\pmb{x}_i^s)\\):" + }, + { + "type": "equation", + "bbox": [ + 0.13, + 0.51, + 0.47, + 0.55 + ], + "angle": 0, + "content": "\\[\n\\left[ \\mathcal {G} \\left(\\boldsymbol {x} _ {i} ^ {s}\\right) \\right] _ {k} = \\boldsymbol {p} ^ {s} \\left(y _ {k} \\mid \\boldsymbol {x} _ {i} ^ {s}\\right) = \\frac {e ^ {\\boldsymbol {w} _ {k} ^ {\\top} \\boldsymbol {z} _ {i} ^ {s} + b _ {k}}}{\\sum_ {l = 1} ^ {H} e ^ {\\boldsymbol {w} _ {l} ^ {\\top} \\boldsymbol {z} _ {i} ^ {s} + b _ {l}}}, \\tag {1}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.55, + 0.473, + 0.794 + ], + "angle": 0, + "content": "where \\(\\pmb{w}_k^\\top\\) and \\(b_{k}\\) denote the \\(k\\)-th class weight and bias in the source classifier \\(\\mathcal{G}(\\cdot)\\). Then we adopt cross-entropy loss to train on the source samples following [44]. Finally, we send the well-trained source model including \\(\\mathcal{F}_s(\\cdot)\\) and \\(\\mathcal{G}(\\cdot)\\) to the unlabeled target domain as shown in Fig. 2. We then introduce the main components of our proposed DTE, as illustrated in Fig. 2. DTE mainly has two components, i.e., weight barcode estimation and sparse label assignment. Weight barcode estimation aims to first distinguish the target samples that belong to the unknown class categories. Sparse label assignment aims to further exploit the rest target samples by providing them with proper pseudo labels. By incorporating weight barcode estimation with sparse label assignment sequentially, we can achieve more precise results on source-free openset domain adaptation. We will introduce the model details in the following subsections." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.802, + 0.327, + 0.818 + ], + "angle": 0, + "content": "3.1. Weight Barcode Estimation" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.825, + 0.471, + 0.903 + ], + "angle": 0, + "content": "In this section, we will first introduce the details of weight barcode estimation. Since target domain include many samples which belong to the unknown classes, it is unreasonable to directly apply traditional alignment techniques in source-free domain adaptation. Specifically, the unknown" + }, + { + "type": "title", + "bbox": [ + 0.501, + 0.092, + 0.859, + 0.107 + ], + "angle": 0, + "content": "Algorithm 1 The training procedure of proposed DTE" + }, + { + "type": "text", + "bbox": [ + 0.516, + 0.109, + 0.892, + 0.139 + ], + "angle": 0, + "content": "Input: \\(\\mathcal{F}_s(\\cdot)\\): well-trained source feature extractor; \\(\\mathcal{G}(\\cdot)\\): well-trained classifier; \\(\\epsilon, \\alpha\\): Hyper parameters." + }, + { + "type": "text", + "bbox": [ + 0.517, + 0.139, + 0.852, + 0.155 + ], + "angle": 0, + "content": "Set class prototypes \\(\\pmb{\\mu}^{(0)} = \\pmb{w}\\) as the initialization." + }, + { + "type": "text", + "bbox": [ + 0.517, + 0.155, + 0.671, + 0.168 + ], + "angle": 0, + "content": "for epoch \\(= 1\\) to \\(L\\) do" + }, + { + "type": "text", + "bbox": [ + 0.533, + 0.169, + 0.892, + 0.201 + ], + "angle": 0, + "content": "Obtain the optimal solution on \\( \\pmb{f}^{*} \\) on PUOT for finding marginal probability \\( q_{\\tau}(\\pmb{x}_i^t) = a_i \\exp(-f_i^* / \\tau) \\)." + }, + { + "type": "text", + "bbox": [ + 0.533, + 0.201, + 0.892, + 0.232 + ], + "angle": 0, + "content": "Obtain the weight barcode matrix \\(\\widehat{\\pmb{Q}}\\) and determine known/unknown target samples \\(\\pmb{x}^t = \\pmb{x}_{\\mathrm{unk}}^t\\cup \\pmb{x}_{\\mathrm{kno}}^t\\)" + }, + { + "type": "text", + "bbox": [ + 0.533, + 0.232, + 0.859, + 0.247 + ], + "angle": 0, + "content": "Obtain the optimal solution \\(\\gamma^{*}\\) on solving Eq.(6)." + }, + { + "type": "text", + "bbox": [ + 0.533, + 0.247, + 0.892, + 0.277 + ], + "angle": 0, + "content": "Gather the network output prediction \\( \\pmb{p}^t \\) and the corresponding \"ground-truth\" \\( \\pmb{q}^t \\) on Eq.(9)-Eq.(10)." + }, + { + "type": "text", + "bbox": [ + 0.533, + 0.277, + 0.875, + 0.292 + ], + "angle": 0, + "content": "Minimize loss function \\(L_{\\mathrm{cls}}\\) on Eq.(11) for training." + }, + { + "type": "text", + "bbox": [ + 0.533, + 0.292, + 0.791, + 0.308 + ], + "angle": 0, + "content": "Update class prototypes \\(\\mu\\) via Eq.(12)." + }, + { + "type": "text", + "bbox": [ + 0.517, + 0.308, + 0.572, + 0.32 + ], + "angle": 0, + "content": "end for" + }, + { + "type": "text", + "bbox": [ + 0.517, + 0.322, + 0.847, + 0.337 + ], + "angle": 0, + "content": "Return: The well-trained model in target domain." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.351, + 0.892, + 0.518 + ], + "angle": 0, + "content": "target samples may get matched to the known class categories and causing negative transfer phenomenon [21, 96]. Therefore, it is essential to distinguish the unknown target samples and filter them out during the first stage. To fulfill this challenging task, we propose weight barcode estimation which includes two steps, i.e., marginal probability estimation and weight barcode establishment. Marginal probability estimation adopts Partial Unbalanced Optimal Transport (PUOT) technique to reweight target samples accordingly. Then we utilize weight barcode establishment to automatically determine the unknown target samples." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.519, + 0.893, + 0.656 + ], + "angle": 0, + "content": "Marginal Probability Estimation. To start with, we first adopt target feature extractor \\(\\mathcal{F}_t(\\cdot)\\) to obtain the latent features \\(z^t\\) on target samples. Then we should consider the matching relationship between target sample features \\(z^t\\) and source class prototype \\(\\mu\\). Recalling that not all target samples belong to the source class categories, it means that only partial of target sample will get matched. It inspired us to propose PUOT for determining which samples belong to the unkown class categories:" + }, + { + "type": "equation", + "bbox": [ + 0.507, + 0.658, + 0.892, + 0.696 + ], + "angle": 0, + "content": "\\[\n\\min _ {\\pi_ {i j} \\geq 0} J _ {\\text {p u o t}} = \\left\\langle \\mathcal {U}, \\boldsymbol {\\pi} \\right\\rangle + \\tau \\operatorname {K L} \\left(\\boldsymbol {\\pi} \\mathbf {1} _ {H} \\| \\boldsymbol {a}\\right) \\text {s . t .} \\boldsymbol {\\pi} ^ {\\top} \\mathbf {1} _ {N} = \\boldsymbol {b}, \\tag {2}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.696, + 0.893, + 0.902 + ], + "angle": 0, + "content": "where \\(\\pi \\in \\mathbb{R}^{N\\times H}\\) denotes the coupling matrix of PUOT. \\(N\\) denotes the batchsize of target samples and \\(H\\) denotes the number of shared known classes. \\(\\pmb{a}\\) and \\(\\pmb{b}\\) denote the corresponding weights for target samples and source prototypes respectively. Specifically we set \\(a_{i} = 1 / N\\) and \\(b_{j} = 1 / H\\) uniformly. \\(\\mathcal{U}\\) denotes the pairwise cost between target samples \\(\\boldsymbol{z}^t\\) and corresponding prototypes \\(\\pmb{\\mu}^{(o)}\\) at the \\(o\\)-th iteration as \\(\\mathcal{U}_{ij} = e^{E(\\pmb{z}_i^t)}(||(\\pmb{z}_i^t)|| - \\pmb{\\mu}_j^{(o)}||\\pmb{\\mu}_j^{(o)}||)\\|_2^2\\). We set \\(\\pmb{\\mu}^{(0)} = \\pmb{w}\\) for the initialization. Meanwhile \\(E(\\cdot)\\) denotes the entropy of classifier output predictions as \\(E(\\pmb{z}_i^t) = -\\sum_{j=1}^{|\\mathcal{C}^s|} \\sigma(\\mathcal{G}(\\mathcal{F}_t(\\pmb{x}_i^t))) [j] \\cdot \\log(\\sigma(\\mathcal{G}(\\mathcal{F}_t(\\pmb{x}_i^t)))) [j])\\) where \\(\\sigma(\\cdot)\\) denotes the softmax operation. When \\(E(\\pmb{x}^t)\\) is much larger, it indicates that the model exhibits greater" + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.516, + 0.957 + ], + "angle": 0, + "content": "4929" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.139, + 0.092, + 0.833, + 0.284 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.076, + 0.287, + 0.895, + 0.346 + ], + "angle": 0, + "content": "Figure 2. The architectures of Distinguish Then Assign (DTE). We first train source feature extractor \\(\\mathcal{F}_s(\\cdot)\\) and classifier \\(\\mathcal{G}(\\cdot)\\) in the source domain. Then we share the well-trained source model to the target domain. We fine-tune target feature extractor \\(\\mathcal{F}_t(\\cdot)\\) based on \\(\\mathcal{F}_s(\\cdot)\\) and fix classifier \\(\\mathcal{G}(\\cdot)\\). Then we propose weight barcode estimation for distinguishing the known and unknown target samples via distinguisher \\(\\hat{\\mathcal{G}}(\\cdot)\\). Finally we propose sparse label assignment for exploiting appropriate pseudo labels for the remaining target samples." + }, + { + "type": "image", + "bbox": [ + 0.085, + 0.351, + 0.462, + 0.549 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.076, + 0.552, + 0.47, + 0.635 + ], + "angle": 0, + "content": "Figure 3. The illustration on the proposed weight barcode estimation with toy examples. We first sample 100 source/target data samples and set \\(30\\%\\) or \\(70\\%\\) of target data samples from unknown categories. Then we adopt marginal probability estimation to obtain the results of target weights. Finally we calculate and visualize the result of \\(\\hat{Q}\\) to distinguish the unknown target samples." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.647, + 0.47, + 0.782 + ], + "angle": 0, + "content": "uncertainty regarding the target samples into known categories. Therefore, it is reasonable to consider such information into the cost calculation on \\(\\mathcal{U}\\). The KL divergence can be calculated via KL \\((\\pmb{x}||\\pmb{z}) = \\sum_{d=1}^{D}[x_d\\log(x_d/z_d) - x_d + z_d]\\) and \\(\\tau\\) denotes the hyper parameters. Unlike the classic optimal transport problem, we relax one of the marginal constraints for the target samples via the KL divergence term to avoid the mismatch. To solve the above problem, we first propose the theorem on the dual form of PUOT:" + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.784, + 0.47, + 0.814 + ], + "angle": 0, + "content": "Theorem 1. Given the PUOT problem \\( J_{\\mathrm{puot}} \\) as shown in Eq.(2), it has the following dual form:" + }, + { + "type": "equation", + "bbox": [ + 0.094, + 0.816, + 0.469, + 0.865 + ], + "angle": 0, + "content": "\\[\n\\min _ {\\boldsymbol {f}, \\boldsymbol {g}, \\zeta} \\left[ \\tau \\sum_ {i = 1} ^ {N} a _ {i} \\exp \\left(- \\frac {f _ {i} + \\zeta}{\\tau}\\right) - \\sum_ {j = 1} ^ {H} b _ {j} \\left(g _ {j} - \\zeta\\right) \\right] \\tag {3}\n\\]" + }, + { + "type": "equation", + "bbox": [ + 0.093, + 0.867, + 0.334, + 0.884 + ], + "angle": 0, + "content": "\\[\ns. t. f _ {i} + g _ {j} + s _ {i j} = \\mathcal {U} _ {i j}, \\quad s _ {i j} \\geq 0,\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.887, + 0.469, + 0.902 + ], + "angle": 0, + "content": "where \\( f, g, s \\) and \\( \\zeta \\) denote Lagrange multipliers. Mean" + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.351, + 0.893, + 0.397 + ], + "angle": 0, + "content": "while \\( s_{ij} \\geq 0 \\) according to the KKT conditions. Moreover, PUOT will be transformed into sample-reweighted optimal transport problem according to its dual form:" + }, + { + "type": "equation", + "bbox": [ + 0.511, + 0.4, + 0.593, + 0.424 + ], + "angle": 0, + "content": "\\[\n\\min _ {\\boldsymbol {\\pi} \\geq 0} \\langle \\mathcal {U}, \\boldsymbol {\\pi} \\rangle\n\\]" + }, + { + "type": "equation", + "bbox": [ + 0.509, + 0.426, + 0.892, + 0.469 + ], + "angle": 0, + "content": "\\[\ns. t. \\sum_ {j = 1} ^ {H} \\pi_ {i j} = a _ {i} \\exp \\left(- \\frac {f _ {i} ^ {*} + \\zeta^ {*}}{\\tau}\\right), \\quad \\sum_ {i = 1} ^ {N} \\pi_ {i j} = b _ {j}. \\tag {4}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.472, + 0.893, + 0.669 + ], + "angle": 0, + "content": "The proof of Theorem 1 is provided in Appendix. That is, the target data samples will be sample-reweighted with coefficient \\(\\exp \\left(-\\left(f_{i}^{*} + \\zeta^{*}\\right) / \\tau\\right)\\) accordingly. At that time the marginal probability on target sample \\(\\pmb{x}_i^t\\) can be given as \\(q_{\\tau}(\\pmb{x}_i^t) = a_i\\exp \\left(-\\left(f_i^* +\\zeta^*\\right) / \\tau\\right)\\). Kindly note that we do not necessarily require an explicit expression for solving \\(\\pi\\), which involves \\((N\\times H)\\) variables and can be quite complex. Meanwhile we mainly focus on solving these dual form Lagrange multipliers in Eq.(3) with \\((N + H)\\) variables. Specifically, we can utilize the \\(c\\)-transform [41, 55, 71] in classical optimal transport theorem to reach the upper bound of \\(\\pmb{g}\\) as \\(g_{j} = \\inf_{k\\in [N]}\\left(\\mathcal{U}_{kj} - f_{k}\\right)\\). By taking the upper bound of \\(\\pmb{g}\\) into Eq.(3), we can reach the final optimization formula as:" + }, + { + "type": "equation", + "bbox": [ + 0.502, + 0.671, + 0.892, + 0.732 + ], + "angle": 0, + "content": "\\[\n\\min _ {\\boldsymbol {f}, \\zeta} \\left[ \\tau \\sum_ {i = 1} ^ {N} a _ {i} e ^ {- \\frac {f _ {i} + \\zeta}{\\tau}} - \\sum_ {j = 1} ^ {H} \\left(\\inf _ {k \\in [ N ]} \\left(\\mathcal {U} _ {k j} - f _ {k}\\right) - \\zeta\\right) b _ {j} \\right]. \\tag {5}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.733, + 0.892, + 0.809 + ], + "angle": 0, + "content": "That equation can be optimized easily. Due to the space limited, the details of optimization procedures for Eq.(5) are provided in Appendix. In summary, we only need to figure out the marginal probability on target samples \\( q_{\\tau}(\\boldsymbol{x}_i^t) \\) via solving the optimal results of \\( f \\) and \\( \\zeta \\) in Eq.(5)." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.811, + 0.893, + 0.901 + ], + "angle": 0, + "content": "Weight Barcode Establishment. According to the optimization process, we can observe that choosing different value of \\(\\tau\\) could lead to different marginal probability. Specifically, as the value of \\(\\tau\\) increases, a larger number of target samples will be assigned higher weights, enhancing their likelihood of being matched. However, we do not" + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.516, + 0.957 + ], + "angle": 0, + "content": "4930" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.077, + 0.092, + 0.473, + 0.489 + ], + "angle": 0, + "content": "know the exact percentage of how many samples are unknown in the target domain which makes it relatively hard to choose a proper value of \\(\\tau\\). To address this issue, we propose an innovative weight barcode mechanism to effectively identify target samples that are more likely belong to unknown classes. We first calculate \\(q_{\\tau}(\\boldsymbol{x}^t)\\) by varying \\(\\tau\\) in range of \\(\\{0.01, 0.05, 0.1, 0.5, 1, 5, 10\\}\\) and collecting them as \\(\\boldsymbol{Q} \\in \\mathbb{R}^{N \\times 7}\\). Then we can establish the binary weight barcode matrix \\(\\widehat{\\boldsymbol{Q}} \\in \\mathbb{R}^{N \\times 7}\\) with the following process as \\(\\widehat{\\boldsymbol{Q}}_{ij} = \\mathrm{func}(Q_{ij} - 1/N)\\) where \\(\\mathrm{func}(x) = 1\\) when \\(x \\geq 0\\) otherwise \\(\\mathrm{func}(x) = 0\\). It is reasonable since PUOT will enhance the target sample weights if they exhibit a higher degree of similarity to the source data. Meanwhile the rest of unlike target samples will gain less weights and we set then \\(\\widehat{\\boldsymbol{Q}}_{ij} = 0\\). To make this process more intuitive, we further provide an numerical example in Fig. 3. Specifically, the red and blue points denote the source and target data samples respectively. We first figure out the pairwise distance between red data \\(\\boldsymbol{z}_i^{\\mathrm{Red}}\\) and blue data \\(\\boldsymbol{z}_j^{\\mathrm{Blue}}\\) via \\(\\mathcal{U}_{ij} = ||\\boldsymbol{z}_i^{\\mathrm{Red}} - \\boldsymbol{z}_j^{\\mathrm{Blue}}||_2^2\\). Then we calculate the weight barcode as \\(\\widehat{\\boldsymbol{Q}}\\) by using marginal probability estimation and visualize the results. We can observe that the unknown target samples achieve lower value of \\(\\widehat{\\boldsymbol{Q}}\\) which can be easily identify. It indicates that the proposed method is adaptable to varying ratios of unknown data samples, even in the absence of prior knowledge." + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.489, + 0.473, + 0.688 + ], + "angle": 0, + "content": "After we obtain the binary weight barcode matrix \\(\\widehat{Q}\\), we can further adopt \\(K\\)-Means algorithm [28] on \\(\\widehat{Q}\\) to cluster them into two groups with corresponding centers \\(\\mathcal{C}_0\\) and \\(\\mathcal{C}_1\\). We further denote \\(\\mathcal{C}_{\\mathrm{known}} = \\max \\left(\\| \\mathcal{C}_0\\| _2, \\| \\mathcal{C}_1\\| _2\\right)\\) and \\(\\mathcal{C}_{\\mathrm{unknown}} = \\min \\left(\\| \\mathcal{C}_0\\| _2, \\| \\mathcal{C}_1\\| _2\\right)\\) by considering the norm of centers. Then we can determine the target samples by measuring the Euclidean distance among \\(\\widehat{Q}\\) and centers \\(\\mathcal{C}_{\\mathrm{known}}\\), \\(\\mathcal{C}_{\\mathrm{unknown}}\\). If the target samples are found to be in close proximity to \\(\\mathcal{C}_{\\mathrm{unknown}}\\), we can classify them as unknown data as \\(\\boldsymbol{x}_{\\mathrm{unk}}^t\\). The remaining target domain samples can all be considered as belonging to the known categories as \\(\\boldsymbol{x}_{\\mathrm{kn0}}^t\\) and as \\(\\boldsymbol{x}^t = \\boldsymbol{x}_{\\mathrm{unk}}^t \\cup \\boldsymbol{x}_{\\mathrm{kn0}}^t\\). In summary, we can now determine the known/unknown target samples accordingly." + }, + { + "type": "title", + "bbox": [ + 0.078, + 0.696, + 0.31, + 0.712 + ], + "angle": 0, + "content": "3.2. Sparse Label Assignment" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.72, + 0.473, + 0.902 + ], + "angle": 0, + "content": "Since we have exploited the target samples with unknown classes \\( \\boldsymbol{x}_{\\mathrm{unk}}^{t} \\) via proposed weight barcode estimation method, we should explore the possible labels for the rest target samples \\( \\boldsymbol{x}_{\\mathrm{kno}}^{t} \\). Recall that \\( \\boldsymbol{x}_{\\mathrm{kno}}^{t} \\) share the same categories with the corresponding prototypes \\( \\boldsymbol{\\mu}^{s} \\), one can adopt domain adaptation manners for finding the latent relationship. Meanwhile, optimal transport techniques could be rather useful in reducing the domain discrepancy via discovering the most proper pairwise matching across domains [9, 25, 75]. However, directly optimizing optimal transport with linear programming would involve heavy computation cost [1, 34, 92]. Although utilizing entropy regularization" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.092, + 0.892, + 0.168 + ], + "angle": 0, + "content": "term could enhance the efficiency, it will bring about relatively dense matching solution which is always inaccurate as shown in Fig. 4(a) [6, 10, 50]. To alleviate this issue, we propose Sparse Sample-Label Matching (SSLM) on optimal transport via involving \\(\\ell_2\\)-norm proximal term:" + }, + { + "type": "equation", + "bbox": [ + 0.542, + 0.168, + 0.892, + 0.21 + ], + "angle": 0, + "content": "\\[\n\\min _ {\\gamma \\in \\Delta} \\ell = \\sum_ {i = 1} ^ {N _ {\\mathrm {k n o}}} \\sum_ {j = 1} ^ {H} \\left[ \\gamma_ {i j} C _ {i j} + \\frac {\\epsilon}{2} \\left(\\gamma_ {i j} - \\gamma_ {i j} ^ {(l)}\\right) ^ {2} \\right], \\tag {6}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.212, + 0.895, + 0.436 + ], + "angle": 0, + "content": "where \\(\\Delta = \\{\\sum_{j=1}^{H} \\gamma_{ij} = \\widehat{a}_i = 1, \\sum_{i=1}^{N_{\\mathrm{kno}}} \\gamma_{ij} = \\widehat{b}_j = N_{\\mathrm{kno}} / H, \\gamma_{ij} \\geq 0\\}\\) and \\(N_{\\mathrm{kno}}\\) denotes the number of target samples \\(\\boldsymbol{x}_{\\mathrm{kno}}^t\\). \\(C = -\\log[\\sigma[\\mathcal{G}(\\mathcal{F}_t(\\boldsymbol{x}_{\\mathrm{kno}}^t))]]\\) denote the output classification probability with negative logarithm. Indeed, when \\(C_{ij}\\) is smaller, it indicates that \\(\\boldsymbol{x}_{\\mathrm{kno},i}^t\\) has a higher probability of belonging to the \\(j\\)-th class [3, 27, 45]. \\(\\widehat{a}_i\\) and \\(\\widehat{b}_j\\) denote the samples weights and we set \\(\\widehat{a}_i = 1\\), \\(\\widehat{b}_j = \\frac{N_{\\mathrm{kno}}}{H}\\) respectively. Meanwhile \\(\\sum_{i=1}^{M} \\sum_{j=1}^{N} (\\gamma_{ij} - \\gamma_{ij}^{(l)})^2\\) denotes the proximal point term with \\(\\ell_2\\)-norm. Incorporating \\(\\ell_2\\)-norm with proximal point method [90], we can reach more sparse and robust results on transportation plan \\(\\gamma^*\\). We initialize \\(\\gamma_{ij}^{(0)} = \\frac{1}{H}\\) for the following iterations. Then we can easily obtain the Fenchel-Legendre dual form of the original problem as:" + }, + { + "type": "equation", + "bbox": [ + 0.499, + 0.437, + 0.903, + 0.495 + ], + "angle": 0, + "content": "\\[\n\\max _ {\\psi , \\phi} \\sum_ {i = 1} ^ {N _ {\\mathrm {k n o}}} \\psi_ {i} \\widehat {a} _ {i} + \\sum_ {j = 1} ^ {H} \\phi_ {j} \\widehat {b} _ {j} - \\frac {\\epsilon}{2} \\sum_ {i = 1} ^ {N _ {\\mathrm {k n o}}} \\sum_ {j = 1} ^ {H} \\left[ \\frac {\\psi_ {i} + \\phi_ {j} - \\widehat {C} _ {i j}}{\\epsilon} \\right] _ {+} ^ {2},\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.495, + 0.892, + 0.617 + ], + "angle": 0, + "content": "where \\(\\phi\\) and \\(\\psi\\) denote the Lagrange multipliers. Meanwhile we set \\([x]_{+}^{2} = (\\max (0,x))^{2}\\) and \\(\\widehat{C}_{ij} = C_{ij} - \\epsilon \\gamma_{ij}^{(l)}\\). Note that the original constrained optimization problem in Eq.(6) turns to be unconstrained after the duality transformation in Eq.(7). The optimization procedure on solving variables \\(\\psi\\) and \\(\\phi\\) are provided in Appendix. After that, we can further calculate the optimal solution of \\(\\gamma^{(l + 1)}\\) at the \\(l\\)-th iteration as shown:" + }, + { + "type": "equation", + "bbox": [ + 0.545, + 0.616, + 0.892, + 0.638 + ], + "angle": 0, + "content": "\\[\n\\gamma_ {i j} ^ {(l + 1)} = \\max (0, [ \\psi_ {i} + \\phi_ {j} + \\epsilon \\gamma_ {i j} ^ {(l)} - C _ {i j} ] / \\epsilon). \\tag {8}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.637, + 0.892, + 0.728 + ], + "angle": 0, + "content": "After we obtain the optimal solution on \\(\\gamma^{*}\\) via solving sparse sample-prototype matching, we can reach the reliable pseudo labels for these target samples \\(\\pmb{x}_{\\mathrm{kno}}^t\\). Furthermore, the optimal solution \\(\\gamma^{*}\\) can yield more sparser and clearer results compared to using only the \\(\\ell_{2}\\)-norm regularization, as depicted in Fig. 4(b)-(c)." + }, + { + "type": "image", + "bbox": [ + 0.508, + 0.73, + 0.885, + 0.817 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.498, + 0.821, + 0.892, + 0.892 + ], + "angle": 0, + "content": "Figure 4. The illustration on Sparse Sample-Label Matching (SSLM) in sparse label assignment with toy examples. Sinkhorn and tradition \\(\\ell_2\\)-norm solvers could easily reach smooth but dense output solutions. Our proposed SSLM method can reach more sparse matching solution for better performance." + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.514, + 0.957 + ], + "angle": 0, + "content": "4931" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.084, + 0.092, + 0.27, + 0.191 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.114, + 0.194, + 0.234, + 0.206 + ], + "angle": 0, + "content": "(a) JPOT on Office-31" + }, + { + "type": "image", + "bbox": [ + 0.282, + 0.092, + 0.468, + 0.19 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.318, + 0.195, + 0.431, + 0.206 + ], + "angle": 0, + "content": "(b) DTE on Office-31" + }, + { + "type": "image_caption", + "bbox": [ + 0.076, + 0.212, + 0.47, + 0.255 + ], + "angle": 0, + "content": "Figure 5. The t-SNE visualization of \\(\\mathbf{W} \\rightarrow \\mathbf{A}\\) tasks on Office-31. Each color denotes a class and the reddish-brown color is used to represent the unknown class." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.264, + 0.245, + 0.281 + ], + "angle": 0, + "content": "3.3. Putting Together" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.288, + 0.469, + 0.363 + ], + "angle": 0, + "content": "We have first distinguished the target samples with unknown classes via weight barcode estimation, then assigning the rest target samples to the corresponding classes via sparse sample-prototype matching. Therefore, we should gather the output prediction of these target samples as \\( \\pmb{p}^t \\):" + }, + { + "type": "equation", + "bbox": [ + 0.084, + 0.365, + 0.469, + 0.396 + ], + "angle": 0, + "content": "\\[\n\\boldsymbol {p} ^ {t} = \\operatorname {s o f t m a x} \\left[ \\mathcal {G} \\left(\\mathcal {F} _ {t} \\left(\\boldsymbol {x} ^ {t}\\right)\\right) \\oplus \\widehat {\\mathcal {G}} \\left(\\mathcal {F} _ {t} \\left(\\boldsymbol {x} ^ {t}\\right)\\right) \\right] \\in \\mathbb {R} ^ {N \\times (H + 1)}. \\tag {9}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.397, + 0.469, + 0.49 + ], + "angle": 0, + "content": "Note that \\(\\widehat{\\mathcal{G}} (\\cdot)\\) is a newly add trainable target classifier to predict whether samples are unknown. Here we concatenate the output predictions on both source classifier \\(\\mathcal{G}(\\cdot)\\) and target classifier \\(\\widehat{\\mathcal{G}} (\\cdot)\\) then process the softmax operation. Meanwhile we can formulate the estimated \"ground-truth\" \\(q_{\\mathrm{kno}}^{t}\\) and \\(q_{\\mathrm{unk}}^{t}\\) for the target samples as follows:" + }, + { + "type": "equation", + "bbox": [ + 0.16, + 0.491, + 0.469, + 0.517 + ], + "angle": 0, + "content": "\\[\n\\boldsymbol {q} _ {\\mathrm {k n o}} ^ {t} = \\left[ \\gamma^ {*} \\oplus 0 \\right] \\in \\mathbb {R} ^ {N _ {\\mathrm {k n o}} \\times (H + 1)}, \\tag {10}\n\\]" + }, + { + "type": "equation", + "bbox": [ + 0.162, + 0.514, + 0.377, + 0.531 + ], + "angle": 0, + "content": "\\[\n\\boldsymbol {q} _ {\\mathrm {u n k}} ^ {t} = [ \\mathbf {0} \\oplus 1 ] \\in \\mathbb {R} ^ {N _ {\\mathrm {u n k}} \\times (H + 1)}.\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.533, + 0.47, + 0.58 + ], + "angle": 0, + "content": "Then we can utilize the cross-entropy loss \\( L_{\\mathrm{cls}} \\) on fine-tuning the target feature extractor \\( \\mathcal{F}_t(\\cdot) \\) based on \\( \\mathcal{F}_s(\\cdot) \\) and training distinguisher \\( \\widehat{\\mathcal{G}}_t(\\cdot) \\) on the target samples as:" + }, + { + "type": "equation", + "bbox": [ + 0.104, + 0.582, + 0.469, + 0.624 + ], + "angle": 0, + "content": "\\[\n\\min _ {\\mathcal {F} _ {t} (\\cdot), \\widehat {\\mathcal {G}} (\\cdot)} L _ {\\mathrm {c l s}} = - \\sum_ {i = 1} ^ {N _ {\\mathrm {u n k}} + N _ {\\mathrm {k n o}}} \\sum_ {j = 1} ^ {H + 1} q _ {i j} ^ {t} \\log p _ {i j} ^ {t}. \\tag {11}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.626, + 0.469, + 0.702 + ], + "angle": 0, + "content": "Note that we fix the source classifier \\(\\mathcal{G}(\\cdot)\\) during the training procedure. Meanwhile, the loss function is easy to implemented for end-to-end training. Moreover, we should update the class prototype \\(\\pmb{\\mu}^{(o + 1)}\\) at the \\((o + 1)\\)-th iteration via incremental learning strategy:" + }, + { + "type": "equation", + "bbox": [ + 0.091, + 0.704, + 0.469, + 0.746 + ], + "angle": 0, + "content": "\\[\n\\boldsymbol {\\mu} _ {j} ^ {(o + 1)} = \\alpha \\boldsymbol {\\mu} _ {j} ^ {(o)} + \\frac {1 - \\alpha}{\\sum_ {i = 1} ^ {N _ {\\mathrm {k n o}}} p _ {i j} ^ {t}} \\sum_ {i = 1} ^ {N _ {\\mathrm {k n o}}} \\mathcal {F} _ {t} \\left(\\boldsymbol {x} _ {\\mathrm {k n o}, i} ^ {t}\\right) p _ {i j} ^ {t}, \\tag {12}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.747, + 0.469, + 0.777 + ], + "angle": 0, + "content": "where \\(\\alpha \\in [0,1]\\) denotes momentum hyper parameter. The whole algorithm of DTE in provided in Alg.1." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.792, + 0.201, + 0.809 + ], + "angle": 0, + "content": "4. Experiment" + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.817, + 0.159, + 0.833 + ], + "angle": 0, + "content": "4.1. Setup" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.84, + 0.469, + 0.902 + ], + "angle": 0, + "content": "We conduct the experiments on several real-world datasets, e.g., Office-31 [77] with following three domains: Amazon (A), Webcam (W) and DSLR (D); Office-Home [84] with following four dissimilar domains: Artistic images (Ar)," + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.092, + 0.892, + 0.167 + ], + "angle": 0, + "content": "Clip Art (CI), Product images (Pr) and Real-World (Rw); VisDA-2017 [70] with domains Synthetic, Real; and Digits with following three domains MNIST [40], USPS [31] and SVHN [67]. We adopt the same set of known and unknown classes following previous works [46, 78]." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.168, + 0.893, + 0.699 + ], + "angle": 0, + "content": "Implementation Details. We first set batch size \\(N = 256\\) during the training procedure. For the experiments on Office-31, Office-Home, and VisDA, we first fine-tune on the backbone network ResNet-50 [29] in the source domain. For the experiments on Digits, we adopt LeNet [40] to extract the sample features. After the training in the source domain, we only transfer the well-trained feature extractor \\(\\mathcal{F}_s(\\cdot)\\) and classifier \\(\\mathcal{G}(\\cdot)\\) to the target domain. We set \\(\\epsilon = 0.5\\) in sparse label assignment. Meanwhile we set \\(\\alpha = 0.8\\) for the incremental learning on class prototypes. We choose Adam [35] as the optimizer and set the learning rate as 0.001 and 0.01 for fine-tuning target feature extractor \\(\\mathcal{F}_t(\\cdot)\\) based on \\(\\mathcal{F}_s(\\cdot)\\) and training distinguisher \\(\\widehat{\\mathcal{G}} (\\cdot)\\) respectively. We employ these four evaluation metrics, i.e., OS, OS*, UNK, HOS and for open set domain adaptation following [78]. Specifically, OS and OS* denotes the accuracy averaged over all classes or only known classes respectively. UNK denotes the accuracy of unknown instances and HOS denotes the harmonic value of OS* and UNK. We compare the proposed DTE with state-of-the-art domain adaptation methods: ResNet [29], Domain Adversarial Neural Network (DANN) [2], Deep Joint Distribution Optimal Transport (DeepJDOT) [11], OpenMAX [5], MMD+OSVM, DANN+OSVM, Openset Back-Propagation (OSBP) [78], Separate To Adapt (STA) [46], Towards Inheritable Models (TIM) [37], Joint Partial Optimal Transport (JPOT) [91], Source Hypothesis Transfer (SHOT) [44], Progressive Graph Learning (PGL) [63], Source-Free Progressive Graph Learning (SFPGL) [64], Unknown-Aware Domain Adversarial Learning (cUADAL) [32], Self-supervised Class-Disccovering Adapter (SCDA) [97], Unknown Oriented Learning (UOL) [47], Adjustment and Alignment (ANNA) [43] and Balanced Correlation Learning (BCL) [30]. We perform five random experiments and report the average and variation results." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.711, + 0.704, + 0.725 + ], + "angle": 0, + "content": "4.2. Result and Discussion" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.735, + 0.892, + 0.901 + ], + "angle": 0, + "content": "The comparison results on Office-31 are shown in Table. 1. Directly applying ResNet could lead to worst results since the existence of domain discrepancy. Although previous method (e.g., OSBP, STA) adopted domain adversarial strategy to reduce the discrepancy, it could be unstable and hard to train in the real practice [81]. Our proposed DTE achieves the best performance among other approaches. Even for the difficult task (e.g., \\(\\mathbf{A} \\rightarrow \\mathbf{W}\\), \\(\\mathbf{W} \\rightarrow \\mathbf{A}\\)) where these images are quite different across domains, DTE still obtains satisfactory results. It indicates the efficacy of the proposed DTE for dealing with open set domain" + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.516, + 0.957 + ], + "angle": 0, + "content": "4932" + } + ], + [ + { + "type": "table_caption", + "bbox": [ + 0.205, + 0.092, + 0.764, + 0.107 + ], + "angle": 0, + "content": "Table 1. Classification accuracy (%) on Office-31 for open set domain adaptation (ResNet-50)." + }, + { + "type": "table", + "bbox": [ + 0.107, + 0.109, + 0.859, + 0.313 + ], + "angle": 0, + "content": "
MethodA→WA→DD→WW→DD→AW→AAvg
OSOS*OSOS*OSOS*OSOS*OSOS*OSOS*OSOS*
ResNet82.5±1.282.7±0.985.2±0.385.5±0.994.1±0.394.3±0.796.6±0.297.0±0.471.6±1.071.5±1.175.5±1.075.2±1.684.284.4
DANN85.3±0.787.7±1.186.5±0.687.7±0.697.5±0.298.3±0.599.5±0.1100.0±075.7±1.676.2±0.974.9±1.275.6±0.886.687.6
DeepJDOT86.1±0.588.7±0.986.9±0.789.0±0.596.9±0.295.6±0.896.1±0.198.1±0.585.6±1.288.4±1.081.5±1.383.1±0.988.990.5
OpenMax87.4±0.587.5±0.387.1±0.988.4±0.996.1±0.496.2±0.398.4±0.398.5±0.383.4±1.082.1±0.682.8±0.982.8±0.689.089.3
OSBP86.5±2.087.6±2.188.6±1.489.2±1.397.0±1.096.5±0.497.9±0.998.7±0.688.9±2.590.6±2.385.8±2.584.9±1.390.891.3
STA89.5±0.692.1±0.593.7±1.596.1±0.497.5±0.296.5±0.599.5±0.299.6±0.189.1±0.593.5±0.887.9±0.987.4±0.692.994.1
TIM91.3±0.793.2±0.594.2±1.197.1±0.896.5±0.597.4±0.799.5±0.299.4±0.390.1±0.291.5±0.288.7±1.388.4±0.993.494.5
JPOT92.8±0.692.2±0.495.2±0.996.0±0.698.1±0.396.2±0.499.5±0.198.6±0.293.0±0.794.1±0.489.9±1.088.4±0.494.694.4
SHOT88.8±0.791.4±0.490.3±0.592.6±0.396.2±0.497.0±0.497.4±0.397.9±0.591.6±0.593.4±0.791.2±0.593.5±0.492.694.3
PGL89.2±0.590.1±0.789.6±0.891.6±0.595.3±0.495.1±0.596.7±0.597.6±0.671.0±1.372.0±1.873.0±1.277.6±1.685.887.4
SFPGL90.4±0.391.9±0.791.0±1.091.2±0.897.3±0.595.5±0.696.1±0.498.4±0.474.8±1.575.6±2.074.9±1.679.6±1.187.488.7
cUADAL92.5±0.493.7±0.696.0±0.797.5±0.698.4±0.4100.0±098.6±0.5100.0±085.3±0.886.0±0.583.2±0.783.9±0.992.393.5
SCDA95.7±0.197.5±0.295.9±0.496.5±0.499.2±0.499.7±0.199.8±0.1100.0±092.1±0.193.7±0.392.2±0.193.6±0.195.896.8
UOL94.6±0.396.8±0.695.4±0.597.2±0.799.5±0.698.2±0.398.9±0.299.1±0.392.7±0.693.2±0.891.6±1.093.0±0.895.596.2
DTE-W93.9±0.694.5±0.595.7±0.696.3±0.798.2±0.599.1±0.297.4±0.498.8±0.694.0±0.993.6±1.290.7±0.891.5±0.794.995.6
DTE-S95.3±0.596.9±0.496.5±0.697.0±0.898.7±0.799.4±0.398.1±0.499.2±0.594.8±1.193.9±1.092.5±0.794.1±0.995.996.7
DTE96.1±0.297.5±0.496.8±0.497.9±0.598.9±0.5100.0±098.5±0.399.4±0.395.6±0.594.3±0.692.9±0.494.6±0.396.497.2
" + }, + { + "type": "table_caption", + "bbox": [ + 0.18, + 0.315, + 0.788, + 0.329 + ], + "angle": 0, + "content": "Table 2. Classification accuracy OS* (%) on Office-Home for open set domain adaptation (ResNet-50)." + }, + { + "type": "table", + "bbox": [ + 0.107, + 0.332, + 0.859, + 0.55 + ], + "angle": 0, + "content": "
MethodAr→ClPr→ClRw→ClAr→PrCl→PrRw→PrCl→ArPr→ArRw→ArAr→RwCl→RwPr→RwAvg
ResNet53.4±0.452.7±0.651.9±0.569.3±0.761.8±0.574.1±0.461.4±0.664.0±0.370.0±0.378.7±0.671.0±0.674.9±0.965.3
DANN54.6±0.749.7±1.651.9±1.469.5±1.163.5±1.072.9±0.861.9±1.263.3±1.071.3±1.080.2±0.871.7±0.474.2±0.465.4
OpenMax56.5±0.452.9±0.753.7±0.469.1±0.364.8±0.474.5±0.664.1±0.964.0±0.871.2±0.880.3±0.873.0±0.576.9±0.366.7
OSBP56.7±1.951.5±2.149.2±2.467.5±1.565.5±1.574.0±1.562.5±2.064.8±1.169.3±1.180.6±0.974.7±2.271.5±1.965.7
STA58.1±0.653.1±0.954.4±1.071.6±1.269.3±1.081.9±0.563.4±0.565.2±0.874.9±1.085.0±0.275.8±0.480.8±0.369.5
TIM60.1±0.754.2±1.056.2±1.770.9±1.470.0±1.778.6±0.664.0±0.666.1±1.374.9±0.983.2±0.975.7±1.381.3±1.469.6
JPOT59.6±0.554.2±0.754.6±0.972.3±1.170.1±0.682.1±0.962.9±0.768.3±0.875.1±1.184.8±0.477.4±0.581.2±0.470.2
SHOT64.5±0.859.3±0.564.6±0.880.4±0.975.4±1.282.3±1.063.1±0.865.3±0.669.6±0.784.7±0.681.2±1.083.3±0.872.8
PGL61.6±1.058.4±0.865.0±0.777.1±1.572.0±1.383.0±0.968.8±1.472.2±0.778.6±1.385.9±0.982.8±1.182.6±1.274.0
SFPGL62.4±0.559.0±0.764.1±0.977.8±0.772.7±0.884.2±1.070.5±0.673.4±0.578.3±0.886.8±0.683.9±0.881.4±0.774.5
cUADAL62.3±0.660.1±0.861.7±0.277.9±0.671.4±0.981.3±0.572.3±0.671.5±0.873.9±1.083.6±0.783.5±0.784.0±1.073.6
SCDA59.9±0.359.0±0.362.8±0.579.6±0.473.8±1.083.7±0.870.9±0.572.3±0.675.5±0.485.3±0.682.9±0.385.7±0.974.3
UOL61.0±0.759.5±0.663.3±0.680.7±1.074.9±0.885.8±0.771.6±0.972.9±0.675.1±0.588.4±0.883.2±0.486.3±0.974.4
ANNA61.3±0.960.4±0.762.5±0.882.1±0.673.6±0.784.4±0.570.2±1.171.6±0.975.7±0.686.9±0.582.0±1.286.9±1.574.8
BCL58.4±0.559.7±0.661.1±0.680.6±0.573.0±0.985.1±0.470.5±0.772.2±0.674.8±0.587.5±0.683.4±0.986.1±1.274.4
DTE-W60.1±0.957.2±0.859.4±0.978.5±0.672.3±1.083.0±0.669.1±0.871.9±0.775.8±1.186.7±0.581.6±0.682.8±1.573.2
DTE-S62.1±0.660.9±0.861.8±0.880.4±0.574.2±1.285.7±0.971.6±0.973.3±1.076.9±0.788.3±0.882.0±0.886.5±1.376.1
DTE63.2±0.462.6±0.663.9±0.581.5±0.575.4±0.886.8±0.672.0±0.574.1±0.477.3±0.790.2±0.684.3±0.487.5±1.176.5
" + }, + { + "type": "table_caption", + "bbox": [ + 0.147, + 0.551, + 0.822, + 0.565 + ], + "angle": 0, + "content": "Table 3. Classification accuracy (%) on Digits (LeNet) and VisDA2017 (VGGNet) for open set domain adaptation." + }, + { + "type": "table", + "bbox": [ + 0.107, + 0.567, + 0.859, + 0.69 + ], + "angle": 0, + "content": "
MethodSVHN→MNISTUSPS→MNISTMNIST→USPSAvgVisDA2017
OSOS*HOSUNKOSOS*HOSUNKOSOS*HOSUNKOSOS*HOSUNKbicyclebuscarmotorcycletraintruckOSOS*
OSVM54.363.117.710.543.132.348.597.179.877.983.289.359.157.749.865.731.751.666.570.488.520.852.554.9
MMD+OSVM55.964.720.112.262.858.968.782.380.079.880.481.066.268.856.458.439.050.164.279.986.616.354.456.0
DANN+OSVM62.975.31.780.9077.192.41.190.6033.835.529.525.360.469.410.88.9331.856.671.777.487.022.355.557.8
OSBP63.059.168.982.592.391.294.497.892.194.985.778.182.481.782.986.151.167.142.884.281.828.062.959.2
STA76.975.479.784.492.291.393.996.793.094.988.883.587.387.287.588.252.469.659.987.886.527.266.863.9
TIM80.479.083.087.495.495.196.096.993.093.591.890.589.689.390.391.653.569.262.285.785.432.568.164.7
JPOT77.275.380.686.792.491.294.798.492.992.194.496.987.585.489.994.055.370.064.189.587.233.769.566.6
UOL84.484.185.085.996.997.196.595.995.395.195.796.392.291.692.492.754.670.563.791.286.931.470.066.3
DTE-W83.182.883.784.696.596.496.797.094.594.295.196.090.391.491.892.556.070.965.688.884.132.269.966.2
DTE-S84.884.585.486.397.096.997.297.594.994.695.596.492.592.292.793.456.571.266.390.187.733.871.367.6
DTE85.585.286.187.097.697.597.898.195.695.496.096.692.992.793.393.957.472.067.990.888.635.272.468.7
" + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.694, + 0.39, + 0.709 + ], + "angle": 0, + "content": "adaptation even under the source-free scenario." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.72, + 0.471, + 0.902 + ], + "angle": 0, + "content": "Meanwhile the results of Office-Home are shown in Table. 2. The tasks on Office-Home are more difficult than Office-31 since it has more categories with large domain gap. Previous methods cannot better reduce domain discrepancy while involving the negative transfer, leading to the limited model performance. Meanwhile our proposed DTE still outperforms the comparison methods. Note that the proposed DTE with weight barcode estimation can fully leverage knowledge from source domain to a significant degree compared with JPOT. Moreover, we conduct the experiments on VisDA-2017 and Digits and report the results in Table. 3. It is remarkable that our proposed DTE can pro" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.694, + 0.892, + 0.831 + ], + "angle": 0, + "content": "mote the better results of OS and \\(\\mathbf{OS^{*}}\\) on hard transfer tasks in VisDA-2017. Specifically, in terms of the performance of all classes, DTE has achieved significant improvements compared to the previous models, with notable examples being a \\(2.27\\%\\) improvement on Office-Home and a \\(3.15\\%\\) improvement on VisDA-2017 against runner-up models for \\(\\mathbf{OS^{*}}\\). DTE can better distinguish the unknown target samples meanwhile correctly assign pseudo labels for the rest target samples to reach state-of-the-art performance." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.845, + 0.607, + 0.861 + ], + "angle": 0, + "content": "4.3. Analysis." + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.871, + 0.892, + 0.903 + ], + "angle": 0, + "content": "Ablation Study. To further study how does each module of DTE contribute on the final performance, we com" + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.516, + 0.957 + ], + "angle": 0, + "content": "4933" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.084, + 0.088, + 0.264, + 0.196 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.113, + 0.203, + 0.235, + 0.215 + ], + "angle": 0, + "content": "(a) Openness on Digits" + }, + { + "type": "image", + "bbox": [ + 0.283, + 0.088, + 0.465, + 0.197 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.304, + 0.203, + 0.445, + 0.214 + ], + "angle": 0, + "content": "(b) Openness on Office-31" + }, + { + "type": "image_caption", + "bbox": [ + 0.078, + 0.221, + 0.468, + 0.235 + ], + "angle": 0, + "content": "Figure 6. The openness experiments of \\( \\mathbb{O} \\) on Digits and Office-31." + }, + { + "type": "image", + "bbox": [ + 0.082, + 0.245, + 0.268, + 0.354 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.109, + 0.358, + 0.239, + 0.37 + ], + "angle": 0, + "content": "(a) Parameters on Digits" + }, + { + "type": "image", + "bbox": [ + 0.282, + 0.245, + 0.467, + 0.354 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.3, + 0.358, + 0.448, + 0.37 + ], + "angle": 0, + "content": "(b) Parameters on Office-31" + }, + { + "type": "image_caption", + "bbox": [ + 0.084, + 0.376, + 0.46, + 0.391 + ], + "angle": 0, + "content": "Figure 7. Hyper parameter tuning of \\(\\epsilon\\) on Digits and Office-31." + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.401, + 0.468, + 0.673 + ], + "angle": 0, + "content": "pare DTE with its several variants, including DTE-W and DTE-S. DTE-W only replaces the proposed weight barcode estimation with JPOT by calculating mean transport cost to determine the unknown target samples. DTE-S only replaces sparse label assignment with self-labelling strategy via Sinkhorn algorithm [3]. The comparison results on Office-Home and Office-31 are shown in Table. 2 and Table. 1. From it, we can observe that (1) Although DTE-W exceeds most of previous methods, it cannot precisely distinguish unknown data samples especially under source-free domain adaptation scenario. (2) DTE-S achieves better results than DTE-W, while it still cannot reach more satisfactory results than DTE. That is because involving entropy regularization term could reach relatively dense solution, thus leading to the inaccurate solutions. (3) DTE achieves the best performance which indicates the efficacy of incorporating weight barcode estimation and sparse label assignment for distinguishing then exploiting target samples." + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.674, + 0.468, + 0.854 + ], + "angle": 0, + "content": "Feature Visualization. We further provide the feature visualization via T-SNE [38] of the target sample representations \\( \\pmb{x}^t \\) by JPOT and DTE on \\( \\mathbf{W} \\rightarrow \\mathbf{A} \\) (Office-31). The results are depicted in Fig. 5(a)-(b), where each color represents a specific class, and the reddish-brown color is used to represent the unknown class. From there we can observe that JPOT cannot better distinguish known and unknown target samples with ambiguous decision boundary. The primary reason is that JPOT can only utilizes limited source knowledge for open set domain adaptation. Meanwhile DTE can better distinguish unknown target samples and cluster the rest known samples accordingly." + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.856, + 0.468, + 0.901 + ], + "angle": 0, + "content": "Openness. To verify the proposed DTE can be applied to the case with different proportion of unknown classes, we adjust the openness \\(\\mathbb{O}\\) in the openset domain adaptation." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.089, + 0.892, + 0.26 + ], + "angle": 0, + "content": "Specifically, \\(\\mathbb{O} = 1 - \\frac{|\\mathcal{C}^s|}{|\\mathcal{C}^t|}\\) denotes different level of openness [46]. Larger openness indicates that fewer classes are known across domains. We vary \\(\\mathbb{O}\\) from 0 to nearly 1 on Office-31 and Digits accordingly [46] then report the result of OS in Fig. 6(a)-(b). We can observe that the majority of previous methods (e.g., DANN and JPOT) cannot better adjust to different proportion of unknown classes in the target domain. Meanwhile, the proposed DTE can always reach better performance even openness \\(\\mathbb{O}\\) approaches to 0 or 1. It further demonstrates the proposed DTE is robust to openness without the requirement of prior knowledge." + }, + { + "type": "table_caption", + "bbox": [ + 0.514, + 0.267, + 0.878, + 0.28 + ], + "angle": 0, + "content": "Table 4. The results (%) for method extension on VisDA2017" + }, + { + "type": "table", + "bbox": [ + 0.505, + 0.283, + 0.892, + 0.343 + ], + "angle": 0, + "content": "
MethodbicyclebuscarmotorcycletraintruckOSOS*
PGL93.593.875.798.896.238.580.782.8
SFPGL93.697.689.695.396.795.691.094.7
DTE(VGGNet+GNN)93.594.077.296.496.041.380.283.1
DTE(ResNet+GNN)93.897.990.197.996.592.091.694.7
" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.35, + 0.892, + 0.47 + ], + "angle": 0, + "content": "Method Extension. Furthermore, we analyse the general extension of our proposed DTE. Specifically, we incorporate weight barcode estimation and sparse label assignment with VGGNet/ResNet and Graph Neural Network (GNN for transductive feature learning [64, 89]) on VisDA2017. The results of OS and \\(\\mathbf{OS^{*}}\\) are shown in Table 4. We observe that the proposed DTE can be applied with various backbones for achieving better performance." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.472, + 0.892, + 0.608 + ], + "angle": 0, + "content": "Hyper Parameter. We finally study the effects of hyperparameters on model performance. For DTE, we vary \\(\\epsilon\\) in \\(\\{0.01, 0.05, 0.1, 0.5, 1, 10\\}\\) for sparse label assignment on Digits/Office-31 and report the results of OS/OS* in Fig. 7(a)-(b). We can conclude that when \\(\\epsilon\\) is larger, the proximal regularization term will have a more important impact compared to sparse-label matching, which can result in poorer performance. Hence we set \\(\\epsilon = 0.5\\) in the experiments empirically." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.62, + 0.618, + 0.636 + ], + "angle": 0, + "content": "5. Conclusion" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.647, + 0.892, + 0.813 + ], + "angle": 0, + "content": "In this paper, we propose Distinguish Then Exploit model (DTE) for tackling the source-free open set domain adaptation problem. DTE first adopts weight barcode estimation with proposed target sample reweighting optimization method on partial unbalanced optimal transport to determine the known and unknown target samples without prior information. Then DTE utilizes sparse sample-label matching on optimal transport to further leverage useful and accurate source knowledge in the target domain. We also conduct extensive experiments on different datasets to show the superior performance of the proposed DTE model." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.83, + 0.67, + 0.846 + ], + "angle": 0, + "content": "6. Acknowledgment" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.856, + 0.892, + 0.901 + ], + "angle": 0, + "content": "This work was supported by the National Natural Science Foundation of China (No. 62372242) and Natural Science Foundation of Shandong Province (No. ZR2023MF007)." + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.516, + 0.957 + ], + "angle": 0, + "content": "4934" + } + ], + [ + { + "type": "title", + "bbox": [ + 0.08, + 0.09, + 0.174, + 0.105 + ], + "angle": 0, + "content": "References" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.115, + 0.468, + 0.142 + ], + "angle": 0, + "content": "[1] Ravindra K Ahuja, Thomas L Magnanti, and James B Orlin. Network flows. 1988. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.145, + 0.469, + 0.186 + ], + "angle": 0, + "content": "[2] Hana Ajakan, Pascal Germain, Hugo Larochelle, François Laviolette, and Mario Marchand. Domain-adversarial neural networks. arXiv preprint arXiv:1412.4446, 2014. 2, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.188, + 0.469, + 0.242 + ], + "angle": 0, + "content": "[3] Yuki Markus Asano, Christian Rupprecht, and Andrea Vedaldi. Self-labelling via simultaneous clustering and representation learning. arXiv preprint arXiv:1911.05371, 2019. 5, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.246, + 0.47, + 0.3 + ], + "angle": 0, + "content": "[4] Yikun Bai, Bernhard Schmitzer, Matthew Thorpe, and Soheil Kolouri. Sliced optimal partial transport. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 13681-13690, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.302, + 0.469, + 0.356 + ], + "angle": 0, + "content": "[5] Abhijit Bendale and Terrance E Boult. Towards open set deep networks. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 1563-1572, 2016. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.359, + 0.469, + 0.413 + ], + "angle": 0, + "content": "[6] Mathieu Blondel, Vivien Seguy, and Antoine Rolet. Smooth and sparse optimal transport. In International conference on artificial intelligence and statistics, pages 880-889. PMLR, 2018. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.416, + 0.469, + 0.457 + ], + "angle": 0, + "content": "[7] Nicolas Bonneel and David Coeurjolly. Spot: sliced partial optimal transport. ACM Transactions on Graphics (TOG), 38(4):1-13, 2019. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.46, + 0.469, + 0.515 + ], + "angle": 0, + "content": "[8] Silvia Bucci, Mohammad Reza Loghmani, and Tatiana Tommasi. On the effectiveness of image rotation for open set domain adaptation. In European conference on computer vision, pages 422-438. Springer, 2020. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.517, + 0.469, + 0.572 + ], + "angle": 0, + "content": "[9] Nicolas Courty, Rémi Flamary, and Devis Tuia. Domain adaptation with regularized optimal transport. In Joint European Conference on Machine Learning and Knowledge Discovery in Databases, pages 274-289. Springer, 2014. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.574, + 0.469, + 0.615 + ], + "angle": 0, + "content": "[10] Marco Cuturi. Sinkhorn distances: Lightspeed computation of optimal transport. Advances in neural information processing systems, 26, 2013. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.617, + 0.469, + 0.687 + ], + "angle": 0, + "content": "[11] Bharath Bhushan Damodaran, Benjamin Kellenberger, Rémi Flamary, Devis Tuia, and Nicolas Courty. Deepjdot: Deep joint distribution optimal transport for unsupervised domain adaptation. In European Conference on Computer Vision, pages 467-483. Springer, 2018. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.688, + 0.469, + 0.757 + ], + "angle": 0, + "content": "[12] Jun Dan, Weiming Liu, Chunfeng Xie, Hua Yu, Shunjie Dong, and Yanchao Tan. Tfgda: Exploring topology and feature alignment in semi-supervised graph domain adaptation through robust clustering. In The Thirty-eighth Annual Conference on Neural Information Processing Systems. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.759, + 0.469, + 0.814 + ], + "angle": 0, + "content": "[13] Jun Dan, Tao Jin, Hao Chi, Shunjie Dong, and Yixuan Shen. Uncertainty-guided joint unbalanced optimal transport for unsupervised domain adaptation. Neural Computing and Applications, 35(7):5351-5367, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.817, + 0.469, + 0.87 + ], + "angle": 0, + "content": "[14] Jun Dan, Tao Jin, Hao Chi, Shunjie Dong, Haoran Xie, Keying Cao, and Xinjing Yang. Trust-aware conditional adversarial domain adaptation with feature norm alignment. Neural Networks, 168:518-530, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.873, + 0.469, + 0.901 + ], + "angle": 0, + "content": "[15] Jun Dan, Tao Jin, Hao Chi, Yixuan Shen, Jiawang Yu, and Jinhai Zhou. Homda: High-order moment-based domain" + }, + { + "type": "list", + "bbox": [ + 0.08, + 0.115, + 0.47, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.533, + 0.093, + 0.892, + 0.12 + ], + "angle": 0, + "content": "alignment for unsupervised domain adaptation. Knowledge-Based Systems, 261:110205, 2023. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.122, + 0.892, + 0.192 + ], + "angle": 0, + "content": "[16] Jun Dan, Yang Liu, Haoyu Xie, Jiankang Deng, Haoran Xie, Xuansong Xie, and Baigui Sun. Transface: Calibrating transformer training for face recognition from a data-centric perspective. In Proceedings of the IEEE/CVF international conference on computer vision, pages 20642-20653, 2023. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.194, + 0.892, + 0.261 + ], + "angle": 0, + "content": "[17] Jun Dan, Mushui Liu, Chunfeng Xie, Jiawang Yu, Haoran Xie, Ruokun Li, and Shunjie Dong. Similar norm more transferable: Rethinking feature norms discrepancy in adversarial domain adaptation. Knowledge-Based Systems, 296: 111908, 2024. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.264, + 0.892, + 0.346 + ], + "angle": 0, + "content": "[18] Jun Dan, Weiming Liu, Mushui Liu, Chunfeng Xie, Shunjie Dong, Guofang Ma, Yanchao Tan, and Jiazheng Xing. Hogda: Boosting semi-supervised graph domain adaptation via high-order structure-guided adaptive feature alignment. In Proceedings of the 32nd ACM International Conference on Multimedia, pages 11109-11118, 2024. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.349, + 0.892, + 0.402 + ], + "angle": 0, + "content": "[19] Jun Dan, Yang Liu, Jiankang Deng, Haoyu Xie, Siyuan Li, Baigui Sun, and Shan Luo. Topofr: A closer look at topology alignment on face recognition. arXiv preprint arXiv:2410.10587, 2024. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.405, + 0.892, + 0.446 + ], + "angle": 0, + "content": "[20] Yuqi Fang, Pew-Thian Yap, Weili Lin, Hongtu Zhu, and Mingxia Liu. Source-free unsupervised domain adaptation: A survey. arXiv preprint arXiv:2301.00265, 2022. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.448, + 0.892, + 0.503 + ], + "angle": 0, + "content": "[21] Zhen Fang, Jie Lu, Feng Liu, Junyu Xuan, and Guangquan Zhang. Open set domain adaptation: Theoretical bound and algorithm. IEEE transactions on neural networks and learning systems, 32(10):4309-4322, 2020. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.505, + 0.892, + 0.572 + ], + "angle": 0, + "content": "[22] Abolfazl Farahani, Sahar Voghoei, Khaled Rasheed, and Hamid R Arabnia. A brief review of domain adaptation. Advances in data science and information engineering: proceedings from ICDATA 2020 and IKE 2020, pages 877-894, 2021. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.575, + 0.892, + 0.63 + ], + "angle": 0, + "content": "[23] Kilian Fatras, Thibault Séjourne, Rémi Flamary, and Nicolas Courty. Unbalanced minibatch optimal transport; applications to domain adaptation. In International Conference on Machine Learning, pages 3186-3197. PMLR, 2021. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.633, + 0.892, + 0.713 + ], + "angle": 0, + "content": "[24] Jean Feydy, Thibault Séjourne, François-Xavier Vialard, Shun-ichi Amari, Alain Trouve, and Gabriel Peyre. Interpolating between optimal transport and mmd using sinkhorn divergences. In The 22nd International Conference on Artificial Intelligence and Statistics, pages 2681–2690. PMLR, 2019. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.717, + 0.892, + 0.757 + ], + "angle": 0, + "content": "[25] Rémi Flamary, Nicholas Courty, Davis Tuia, and Alain Rakotomamonjy. Optimal transport for domain adaptation. IEEE Trans. Pattern Anal. Mach. Intell, 1:1-40, 2016. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.76, + 0.892, + 0.841 + ], + "angle": 0, + "content": "[26] Rémi Flamary, Nicolas Courty, Alexandre Gramfort, Mokhtar Z Alaya, Aurélie Boisbunon, Stanislas Chambon, Laetitia Chapel, Adrien Corenflos, Kilian Fatras, Nemo Fournier, et al. Pot: Python optimal transport. The Journal of Machine Learning Research, 22(1):3571-3578, 2021. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.844, + 0.892, + 0.871 + ], + "angle": 0, + "content": "[27] Yves Grandvalet and Yoshua Bengio. Entropy regularization., 2006. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.874, + 0.892, + 0.901 + ], + "angle": 0, + "content": "[28] John A Hartigan and Manchek A Wong. Algorithm as 136: A k-means clustering algorithm. Journal of the royal sta" + }, + { + "type": "list", + "bbox": [ + 0.503, + 0.093, + 0.892, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.516, + 0.957 + ], + "angle": 0, + "content": "4935" + } + ], + [ + { + "type": "ref_text", + "bbox": [ + 0.11, + 0.092, + 0.468, + 0.119 + ], + "angle": 0, + "content": "tistical society. series c (applied statistics), 28(1):100-108, 1979. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.122, + 0.469, + 0.177 + ], + "angle": 0, + "content": "[29] Kaiming He, Xiangyu Zhang, Shaoqing Ren, and Jian Sun. Deep residual learning for image recognition. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 770-778, 2016. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.179, + 0.469, + 0.207 + ], + "angle": 0, + "content": "[30] Zi-Xian Huang and Chuan-Xian Ren. Rethinking correlation learning via label prior for open set domain adaptation. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.208, + 0.469, + 0.248 + ], + "angle": 0, + "content": "[31] J. J. Hull. A database for handwritten text recognition research. IEEE Transactions on Pattern Analysis & Machine Intelligence, 16(5):550-554, 2002. 6, 15" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.25, + 0.469, + 0.319 + ], + "angle": 0, + "content": "[32] JoonHo Jang, Byeonghu Na, Dong Hyeok Shin, Mingi Ji, Kyungwoo Song, and Il-Chul Moon. Unknown-aware domain adversarial learning for open-set domain adaptation. Advances in Neural Information Processing Systems, 35: 16755-16767, 2022. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.321, + 0.469, + 0.376 + ], + "angle": 0, + "content": "[33] Jun-Gi Jang, Sooyeon Shim, Vladimir Egay, Jeeyong Lee, Jongmin Park, Suhyun Chae, and U Kang. Accurate open-set recognition for memory workload. ACM Transactions on Knowledge Discovery from Data, 17(9):1-14, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.378, + 0.469, + 0.405 + ], + "angle": 0, + "content": "[34] Jeff L Kennington and Richard V Helgason. Algorithms for network programming. John Wiley & Sons, Inc., 1980. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.407, + 0.469, + 0.446 + ], + "angle": 0, + "content": "[35] Diederik P Kingma and Jimmy Ba. Adam: A method for stochastic optimization. arXiv preprint arXiv:1412.6980, 2014. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.449, + 0.469, + 0.505 + ], + "angle": 0, + "content": "[36] Jogendra Nath Kundu, Naveen Venkat, R Venkatesh Babu, et al. Universal source-free domain adaptation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 4544-4553, 2020. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.507, + 0.47, + 0.575 + ], + "angle": 0, + "content": "[37] Jogendra Nath Kundu, Naveen Venkat, Ambareesh Revanur, R Venkatesh Babu, et al. Towards inheritable models for open-set domain adaptation. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 12376-12385, 2020. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.577, + 0.469, + 0.617 + ], + "angle": 0, + "content": "[38] Van Der Maaten Laurens and Geoffrey Hinton. Visualizing data using t-sne. Journal of Machine Learning Research, 9 (2605):2579-2605, 2008. 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.619, + 0.469, + 0.687 + ], + "angle": 0, + "content": "[39] Khang Le, Huy Nguyen, Quang M Nguyen, Tung Pham, Hung Bui, and Nhat Ho. On robust optimal transport: Computational complexity and barycenter computation. Advances in Neural Information Processing Systems, 34:21947-21959, 2021. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.689, + 0.469, + 0.744 + ], + "angle": 0, + "content": "[40] Y. L. Lecun, Leon Bottou, Yoshua Bengio, and Patrick Haffner. Gradient-based learning applied to document recognition. proc. ii. Proceedings of the IEEE, 86(11):2278-2324, 1998. 6, 15" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.746, + 0.469, + 0.786 + ], + "angle": 0, + "content": "[41] Na Lei and Xianfeng Gu. Computational Conformal Geometry. International Press and Higher Education Press, 2021. 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.789, + 0.469, + 0.843 + ], + "angle": 0, + "content": "[42] Bo Li, Yifei Shen, Jingkang Yang, Yezhen Wang, Jiawei Ren, Tong Che, Jun Zhang, and Ziwei Liu. Sparse mixture-of-experts are domain generalizable learners. arXiv preprint arXiv:2206.04046, 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.846, + 0.469, + 0.901 + ], + "angle": 0, + "content": "[43] Wuyang Li, Jie Liu, Bo Han, and Yixuan Yuan. Adjustment and alignment for unbiased open set domain adaptation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 24110-24119, 2023. 6" + }, + { + "type": "list", + "bbox": [ + 0.08, + 0.092, + 0.47, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.093, + 0.892, + 0.148 + ], + "angle": 0, + "content": "[44] Jian Liang, Dapeng Hu, and Jiashi Feng. Do we really need to access the source data? source hypothesis transfer for unsupervised domain adaptation. In International conference on machine learning, pages 6028-6039. PMLR, 2020. 3, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.151, + 0.894, + 0.219 + ], + "angle": 0, + "content": "[45] Mattia Litrico, Alessio Del Bue, and Pietro Morerio. Guiding pseudo-labels with uncertainty estimation for source-free unsupervised domain adaptation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 7640-7650, 2023. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.221, + 0.894, + 0.29 + ], + "angle": 0, + "content": "[46] Hong Liu, Zhangjie Cao, Mingsheng Long, Jianmin Wang, and Qiang Yang. Separate to adapt: Open set domain adaptation via progressive separation. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 2927-2936, 2019. 2, 6, 8, 15" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.292, + 0.892, + 0.346 + ], + "angle": 0, + "content": "[47] Jie Liu, Xiaoqing Guo, and Yixuan Yuan. Unknown-oriented learning for open set domain adaptation. In European Conference on Computer Vision, pages 334-350. Springer, 2022. 2, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.349, + 0.892, + 0.389 + ], + "angle": 0, + "content": "[48] Mushui Liu, Yunlong Yu, Zhong Ji, Jungong Han, and Zhongfei Zhang. Tolerant self-distillation for image classification. Neural Networks, 174:106215, 2024. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.392, + 0.892, + 0.446 + ], + "angle": 0, + "content": "[49] Mushui Liu, Weijie He, Ziqian Lu, Jun Dan, Yunlong Yu, Yingming Li, Xi Li, and Jungong Han. Synth-clip: Synthetic data make clip generalize better in data-limited scenarios. Neural Networks, 184:107083, 2025. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.449, + 0.892, + 0.476 + ], + "angle": 0, + "content": "[50] Tianlin Liu, Joan Puigcerver, and Mathieu Blondel. Sparsity-constrained optimal transport. *ICLR*, 2023. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.478, + 0.892, + 0.561 + ], + "angle": 0, + "content": "[51] Weiming Liu, Xiaolin Zheng, Jiajie Su, Mengling Hu, Yanchao Tan, and Chaochao Chen. Exploiting variational domain-invariant user embedding for partially overlapped cross domain recommendation. In Proceedings of the 45th International ACM SIGIR conference on research and development in information retrieval, pages 312-321, 2022. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.563, + 0.892, + 0.631 + ], + "angle": 0, + "content": "[52] Weiming Liu, Xiaolin Zheng, Chaochao Chen, Jiajie Su, Xinting Liao, Mengling Hu, and Yanchao Tan. Joint internal multi-interest exploration and external domain alignment for cross domain sequential recommendation. In Proceedings of the ACM web conference 2023, pages 383-394, 2023." + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.633, + 0.892, + 0.702 + ], + "angle": 0, + "content": "[53] Weiming Liu, Chaochao Chen, Xinting Liao, Mengling Hu, Jiajie Su, Yanchao Tan, and Fan Wang. User distribution mapping modelling with collaborative filtering for cross-domain recommendation. In Proceedings of the ACM Web Conference 2024, pages 334-343, 2024." + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.704, + 0.892, + 0.787 + ], + "angle": 0, + "content": "[54] Weiming Liu, Chaochao Chen, Xinting Liao, Mengling Hu, Yanchao Tan, Fan Wang, Xiaolin Zheng, and Yew Soon Ong. Learning accurate and bidirectional transformation via dynamic embedding transportation for cross-domain recommendation. In Proceedings of the AAAI Conference on Artificial Intelligence, pages 8815-8823, 2024. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.789, + 0.892, + 0.871 + ], + "angle": 0, + "content": "[55] Weiming Liu, Xiaolin Zheng, Chaochao Chen, Jiahe Xu, Xinting Liao, Fan Wang, Yanchao Tan, and Yew-Soon Ong. Reducing item discrepancy via differentially private robust embedding alignment for privacy-preserving cross domain recommendation. In *Forty-first International Conference on Machine Learning*, 2024. 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.873, + 0.892, + 0.901 + ], + "angle": 0, + "content": "[56] Ye Liu, Lingfeng Qiao, Changchong Lu, Di Yin, Chen Lin, Haoyuan Peng, and Bo Ren. Osman: A one-stage alignment" + }, + { + "type": "list", + "bbox": [ + 0.503, + 0.093, + 0.894, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.946, + 0.516, + 0.956 + ], + "angle": 0, + "content": "4936" + } + ], + [ + { + "type": "ref_text", + "bbox": [ + 0.108, + 0.092, + 0.47, + 0.147 + ], + "angle": 0, + "content": "network to unify multimodal alignment and unsupervised domain adaptation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 3551-3560, 2023. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.151, + 0.47, + 0.219 + ], + "angle": 0, + "content": "[57] Yang Liu, Zhipeng Zhou, and Baigui Sun. Cot: Unsupervised domain adaptation with clustering and optimal transport. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 19998-20007, 2023. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.223, + 0.47, + 0.291 + ], + "angle": 0, + "content": "[58] Mingsheng Long, Jianmin Wang, Guiguang Ding, Jiaguang Sun, and Philip S Yu. Transfer feature learning with joint distribution adaptation. In Proceedings of the IEEE international conference on computer vision, pages 2200-2207, 2013. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.295, + 0.469, + 0.35 + ], + "angle": 0, + "content": "[59] Mingsheng Long, Yue Cao, Jianmin Wang, and Michael Jordan. Learning transferable features with deep adaptation networks. In International Conference on Machine Learning, pages 97-105, 2015. 1, 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.353, + 0.469, + 0.407 + ], + "angle": 0, + "content": "[60] Mingsheng Long, Han Zhu, Jianmin Wang, and Michael I Jordan. Unsupervised domain adaptation with residual transfer networks. Advances in neural information processing systems, 29, 2016. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.411, + 0.469, + 0.48 + ], + "angle": 0, + "content": "[61] Mingsheng Long, Han Zhu, Jianmin Wang, and Michael I Jordan. Deep transfer learning with joint adaptation networks. In Proceedings of the 34th International Conference on Machine Learning-Volume 70, pages 2208-2217. JMLR.org, 2017. 1, 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.483, + 0.469, + 0.538 + ], + "angle": 0, + "content": "[62] Mingsheng Long, Zhangjie Cao, Jianmin Wang, and Michael I Jordan. Conditional adversarial domain adaptation. In Advances in Neural Information Processing Systems, pages 1640-1650, 2018. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.541, + 0.469, + 0.596 + ], + "angle": 0, + "content": "[63] Yadan Luo, Zijian Wang, Zi Huang, and Mahsa Baktashmotlagh. Progressive graph learning for open-set domain adaptation. In International Conference on Machine Learning, pages 6468-6478. PMLR, 2020. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.599, + 0.469, + 0.655 + ], + "angle": 0, + "content": "[64] Yadan Luo, Zijian Wang, Zhuoxiao Chen, Zi Huang, and Mahsa Baktashmotlagh. Source-free progressive graph learning for open-set domain adaptation. IEEE Transactions on Pattern Analysis and Machine Intelligence, 2023. 3, 6, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.657, + 0.469, + 0.713 + ], + "angle": 0, + "content": "[65] You-Wei Luo and Chuan-Xian Ren. Conditional bures metric for domain adaptation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 13989–13998, 2021. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.716, + 0.469, + 0.77 + ], + "angle": 0, + "content": "[66] You-Wei Luo and Chuan-Xian Ren. Mot: Masked optimal transport for partial domain adaptation. In 2023 IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 3531–3540. IEEE, 2023. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.773, + 0.469, + 0.841 + ], + "angle": 0, + "content": "[67] Yuval Netzer, Tao Wang, Adam Coates, Alessandro Bissacco, Bo Wu, and Andrew Y. Ng. Reading digits in natural images with unsupervised feature learning. Nips Workshop on Deep Learning & Unsupervised Feature Learning, 2011. 6, 15" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.845, + 0.469, + 0.901 + ], + "angle": 0, + "content": "[68] Khai Nguyen, Dang Nguyen, Quoc Nguyen, Tung Pham, Hung Bui, Dinh Phung, Trung Le, and Nhat Ho. On transportation of mini-batches: A hierarchical approach. arXiv preprint arXiv:2102.05912, 2021. 2" + }, + { + "type": "list", + "bbox": [ + 0.08, + 0.092, + 0.47, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.093, + 0.892, + 0.134 + ], + "angle": 0, + "content": "[69] Pau Panareda Busto and Juergen Gall. Open set domain adaptation. In Proceedings of the IEEE international conference on computer vision, pages 754-763, 2017. 1, 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.136, + 0.892, + 0.204 + ], + "angle": 0, + "content": "[70] Xingchao Peng, Ben Usman, Neela Kaushik, Dequan Wang, Judy Hoffman, and Kate Saenko. Visda: A synthetic-to-real benchmark for visual domain adaptation. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition Workshops, pages 2021-2026, 2018. 6, 15" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.205, + 0.892, + 0.232 + ], + "angle": 0, + "content": "[71] Gabriel Peyré. Course notes on computational optimal transport. 2019. 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.233, + 0.892, + 0.287 + ], + "angle": 0, + "content": "[72] Khiem Pham, Khang Le, Nhat Ho, Tung Pham, and Hung Bui. On unbalanced optimal transport: An analysis of sinkhorn algorithm. In International Conference on Machine Learning, pages 7673-7682. PMLR, 2020. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.289, + 0.892, + 0.357 + ], + "angle": 0, + "content": "[73] Sanqing Qu, Tianpei Zou, Florian Röhrbein, Cewu Lu, Guang Chen, Dacheng Tao, and Changjun Jiang. Upcycling models under domain and category shift. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 20019-20028, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.359, + 0.892, + 0.439 + ], + "angle": 0, + "content": "[74] Sanqing Qu, Tianpei Zou, Lianghua He, Florian Röhrbein, Alois Knoll, Guang Chen, and Changjun Jiang. Lead: Learning decomposition for source-free universal domain adaptation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 23334-23343, 2024. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.442, + 0.892, + 0.524 + ], + "angle": 0, + "content": "[75] Ievgen Redko, Amaury Habrard, and Marc Sebban. Theoretical analysis of domain adaptation with optimal transport. In Machine Learning and Knowledge Discovery in Databases: European Conference, ECML PKDD 2017, Skopje, Macedonia, September 18-22, 2017, Proceedings, Part II 10, pages 737-753. Springer, 2017. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.526, + 0.892, + 0.593 + ], + "angle": 0, + "content": "[76] Ievgen Redko, Nicolas Courty, Rémi Flamary, and Devis Tuia. Optimal transport for multi-source domain adaptation under target shift. In The 22nd International Conference on artificial intelligence and statistics, pages 849-858. PMLR, 2019. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.595, + 0.892, + 0.637 + ], + "angle": 0, + "content": "[77] Kate Saenko, Brian Kulis, Mario Fritz, and Trevor Darrell. Adapting visual category models to new domains. In European Conference on Computer Vision (ECCV), 2010. 6, 15" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.638, + 0.892, + 0.692 + ], + "angle": 0, + "content": "[78] Kuniaki Saito, Shohei Yamamoto, Yoshitaka Ushiku, and Tatsuya Harada. Open set domain adaptation by backpropagation. In Proceedings of the European conference on computer vision (ECCV), pages 153-168, 2018. 1, 2, 6, 16" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.693, + 0.892, + 0.761 + ], + "angle": 0, + "content": "[79] Thibault Séjourne, François-Xavier Vialard, and Gabriel Peyré. Faster unbalanced optimal transport: Translation invariant sinkhorn and 1-d frank-wolfe. In International Conference on Artificial Intelligence and Statistics, pages 4995-5021. PMLR, 2022. 13" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.762, + 0.892, + 0.803 + ], + "angle": 0, + "content": "[80] Jian Shen, Yanru Qu, Weinan Zhang, and Yong Yu. Wasserstein distance guided representation learning for domain adaptation. arXiv preprint arXiv:1707.01217, 2017. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.804, + 0.892, + 0.845 + ], + "angle": 0, + "content": "[81] Rui Shu, Hung H Bui, Hirokazu Narui, and Stefano Ermon. A dirt-t approach to unsupervised domain adaptation. arXiv preprint arXiv:1802.08735, 2018. 2, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.846, + 0.892, + 0.899 + ], + "angle": 0, + "content": "[82] Baochen Sun and Kate Saenko. Deep coral: Correlation alignment for deep domain adaptation. In European Conference on Computer Vision, pages 443-450. Springer, 2016. 2" + }, + { + "type": "list", + "bbox": [ + 0.503, + 0.093, + 0.892, + 0.899 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.946, + 0.515, + 0.956 + ], + "angle": 0, + "content": "4937" + } + ], + [ + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.092, + 0.47, + 0.134 + ], + "angle": 0, + "content": "[83] Eric Tzeng, Judy Hoffman, Ning Zhang, Kate Saenko, and Trevor Darrell. Deep domain confusion: Maximizing for domain invariance. arXiv preprint arXiv:1412.3474, 2014. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.136, + 0.47, + 0.19 + ], + "angle": 0, + "content": "[84] Hemanth Venkateswara, Jose Eusebio, Shayok Chakraborty, and Sethuraman Panchanathan. Deep hashing network for unsupervised domain adaptation. CoRR, abs/1706.07522, 2017. 6, 15" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.192, + 0.469, + 0.22 + ], + "angle": 0, + "content": "[85] Cédric Villani et al. Optimal transport: old and new. Springer, 2009. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.222, + 0.469, + 0.248 + ], + "angle": 0, + "content": "[86] Mei Wang and Weihong Deng. Deep visual domain adaptation: A survey. Neurocomputing, 312:135-153, 2018. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.251, + 0.469, + 0.303 + ], + "angle": 0, + "content": "[87] Qian Wang, Fanlin Meng, and Toby P Breckon. Progressively select and reject pseudo-labelled samples for open-set domain adaptation. arXiv preprint arXiv:2110.12635, 2021. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.307, + 0.469, + 0.348 + ], + "angle": 0, + "content": "[88] Garrett Wilson and Diane J Cook. A survey of unsupervised deep domain adaptation. ACM Transactions on Intelligent Systems and Technology (TIST), 11(5):1-46, 2020. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.35, + 0.469, + 0.404 + ], + "angle": 0, + "content": "[89] Zonghan Wu, Shirui Pan, Fengwen Chen, Guodong Long, Chengqi Zhang, and S Yu Philip. A comprehensive survey on graph neural networks. IEEE transactions on neural networks and learning systems, 32(1):4-24, 2020. 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.407, + 0.469, + 0.461 + ], + "angle": 0, + "content": "[90] Yujia Xie, Xiangfeng Wang, Ruijia Wang, and Hongyuan Zha. A fast proximal point method for computing exact Wasserstein distance. In Uncertainty in artificial intelligence, pages 433-453. PMLR, 2020. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.463, + 0.469, + 0.517 + ], + "angle": 0, + "content": "[91] Renjun Xu, Pelen Liu, Yin Zhang, Fang Cai, Jindong Wang, Shuoying Liang, Heting Ying, and Jianwei Yin. Joint partial optimal transport for open set domain adaptation. In IJCAI, pages 2540-2546, 2020. 1, 2, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.52, + 0.469, + 0.574 + ], + "angle": 0, + "content": "[92] Tengyue Xu and Jun Dan. Ehm: Exploring dynamic alignment and hierarchical clustering in unsupervised domain adaptation via high-order moment-guided contrastive learning. Neural Networks, page 107188, 2025. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.577, + 0.469, + 0.658 + ], + "angle": 0, + "content": "[93] Hongliang Yan, Yukang Ding, Peihua Li, Qilong Wang, Yong Xu, and Wangmeng Zuo. Mind the class weight bias: Weighted maximum mean discrepancy for unsupervised domain adaptation. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition, pages 2272-2281, 2017. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.661, + 0.47, + 0.715 + ], + "angle": 0, + "content": "[94] Yu-Chu Yu and Hsuan-Tien Lin. Semi-supervised domain adaptation with source label adaptation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 24100-24109, 2023. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.717, + 0.469, + 0.758 + ], + "angle": 0, + "content": "[95] Zhiqi Yu, Jingjing Li, Zhekai Du, Lei Zhu, and Heng Tao Shen. A comprehensive survey on source-free domain adaptation. arXiv preprint arXiv:2302.11803, 2023. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.76, + 0.469, + 0.815 + ], + "angle": 0, + "content": "[96] Li Zhong, Zhen Fang, Feng Liu, Bo Yuan, Guangquan Zhang, and Jie Lu. Bridging the theoretical bound and deep algorithms for open set domain adaptation. IEEE transactions on neural networks and learning systems, 2021. 2, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.817, + 0.469, + 0.883 + ], + "angle": 0, + "content": "[97] Jingyu Zhuang, Ziliang Chen, Pengxu Wei, Guanbin Li, and Liang Lin. Discovering implicit classes achieves open set domain adaptation. In 2022 IEEE International Conference on Multimedia and Expo (ICME), pages 01-06. IEEE, 2022. 6" + }, + { + "type": "list", + "bbox": [ + 0.08, + 0.092, + 0.47, + 0.883 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.946, + 0.516, + 0.956 + ], + "angle": 0, + "content": "4938" + } + ] +] \ No newline at end of file diff --git a/2025/Distinguish Then Exploit_ Source-free Open Set Domain Adaptation via Weight Barcode Estimation and Sparse Label Assignment/099209ec-4ec9-4836-8000-10ddecb15989_origin.pdf b/2025/Distinguish Then Exploit_ Source-free Open Set Domain Adaptation via Weight Barcode Estimation and Sparse Label Assignment/099209ec-4ec9-4836-8000-10ddecb15989_origin.pdf new file mode 100644 index 0000000000000000000000000000000000000000..8a894afc07b55b506d12ee8eb49057e3617a5013 --- /dev/null +++ b/2025/Distinguish Then Exploit_ Source-free Open Set Domain Adaptation via Weight Barcode Estimation and Sparse Label Assignment/099209ec-4ec9-4836-8000-10ddecb15989_origin.pdf @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:e21127995be3018c0f36574a0e0223db34694cc37c171a361a8905f1f5765775 +size 2245613 diff --git a/2025/Distinguish Then Exploit_ Source-free Open Set Domain Adaptation via Weight Barcode Estimation and Sparse Label Assignment/full.md b/2025/Distinguish Then Exploit_ Source-free Open Set Domain Adaptation via Weight Barcode Estimation and Sparse Label Assignment/full.md new file mode 100644 index 0000000000000000000000000000000000000000..c7f9e3757dadad94265ee5ba089c1261d945d1b5 --- /dev/null +++ b/2025/Distinguish Then Exploit_ Source-free Open Set Domain Adaptation via Weight Barcode Estimation and Sparse Label Assignment/full.md @@ -0,0 +1,396 @@ +# Distinguish Then Exploit: Source-free Open Set Domain Adaptation via Weight Barcode Estimation and Sparse Label Assignment + +Weiming Liu $^{1*}$ , Jun Dan $^{2*}$ , Fan Wang $^{2}$ , Xinting Liao $^{2}$ , Junhao Dong $^{3}$ , Hua Yu $^{3}$ , Shunjie Dong $^{4}$ , Lianyong Qi $^{56\dagger}$ + +1Bytedance Inc. 2Zhejiang University 3Nanyang Technological University 4Shanghai Jiao Tong University 5China University of Petroleum (East China) + +$^{6}$ Shandong Key Laboratory of Intelligent Oil & Gas Industrial Software + +lwming95@gmail.com, danjun@zju.edu.cn, fanwang97@zju.edu.cn, xintingliao@zju.edu.cn, junhao003@ntu.edu.sg, yu_hua@ntu.edu.sg, sjdong@sjtu.edu.cn, lianyongqi@gmail.com + +# Abstract + +Nowadays, domain adaptation techniques have been widely investigated for knowledge sharing from labeled source domain to unlabeled target domain. However, target domain may include some data samples that belong to unknown categories in real-world scenarios. Moreover, the target domain cannot access the source data samples due to privacy-preserving restrictions. In this paper, we focus on the source-free open set domain adaptation problem which includes two main challenges, i.e., how to distinguish known and unknown target samples and how to exploit useful source information to provide trustworthy pseudo labels for known target samples. Existing approaches that directly apply conventional domain alignment methods could lead to sample mismatch and misclassification in this scenario. To overcome these issues, we propose a Distinguish Then Exploit model (DTE) with two components, i.e., weight barcode estimation and sparse label assignment. Weight barcode estimation first calculates the marginal probability of target samples via partially unbalanced optimal transport, then quantize barcode results to distinguish unknown target samples. Sparse label assignment utilizes sparse sample-label matching via proximal term to fully exploit useful source information. Our empirically study on several datasets shows that DTE outperforms the state-of-the-art models on tackling the source-free open set domain adaptation problem. + +# 1. Introduction + +Nowadays, deep learning technique has been successfully adopted to image classification tasks with the help of large-scale labeled datasets [16, 19, 48, 51-54, 57, 65, 66]. How + +ever, it is time-consuming to sufficiently annotate data for providing satisfied results. A promising solution is to leverage useful information from the labeled source domain to assign reliable labels within the unlabeled target domain [18, 58, 60]. Nevertheless, the data distribution could be significantly different across domains, leading to the model degradation on final performance [12, 56, 94]. How to reduce the domain bias and discrepancy while avoiding negative transfer across domains has become an urgent problem. + +In this paper, we focus on a more general problem of source-free open set domain adaptation. That is, there could exist some target samples belong to extra categories that does not present in the source domain. Moreover, due to data privacy protection constraints, the target domain could only have access to the well-trained model in the source domain. Such settings are commonly encountered in real-world scenarios and it is rather complicated than conventional domain adaptation problem. Specifically, there are two main challenges: CH1: How can we enhance the ability to distinguish target samples that belong to unknown categories? CH2: How to better exploit appropriate pseudo labels for the remaining known target samples? + +Existing researches cannot solve source-free open set domain adaptation well. Firstly, conventional transfer learning always assumes that both source and target domains share the same set of label categories [15, 59, 61]. Directly aligning the latent distributions of source and target domains may lead to negative transfer, where target unknown samples are incorrectly matched to source samples. Secondly, traditional open set domain adaptation approaches should incorporate well-tuned thresholds for determining the decision boundary between known and unknown target samples [69]. This suggests that many of these methods (e.g., OSBP [78] and JPOT [91]) typically rely on prior knowledge about the target domain, which will limit their potential applicabil + +![](images/367ba117f828b9bdb3225d33197ee7e634f0001c025e199ac7ddb24ff3292821.jpg) +(a) Initial Target Features in +Latent Space + +![](images/c40dfff88a22a4c7a9315ee6922b8d5e279b901d4dae117c13c436d61b425e35.jpg) +(b) Adopting Weight Barcode Estimation to Distinguish Unknown Samples in Target Domain + +![](images/6a9845cf78cf0bafd758196053ee075f3662a26df15d935f008dcc09f5405bad.jpg) + +![](images/76b64f5d902db14e71f944365c9f7a4ca5aa89930c70966f7291d23704118d73.jpg) +(c) Adopting Sparse Label Assignment to Exploit Pseudo Label for Known Samples in Target Domain +(d) Target Features in Latent Space after Training +Figure 1. An overview of the proposed DTE. (a) The initial target features provided in the latent space are dispersed and challenging to achieve satisfactory classification results. (b) We first adopt weight barcode estimation component to distinguish known or unknown target samples accordingly. (c) Then we adopt sparse label assignment to exploit pseudo label for known samples in the target domains. (d) Finally we can better cluster and classify target unlabeled samples. Best viewed in color. + +ity. If such prior knowledge (e.g., the proportion of unknown classes in the target domain) cannot be well estimated, the model performance could be degraded dramatically. Some recent works (e.g., STA [46]) further utilized coarse-to-fine weighting mechanism for separating target samples under different proportion of unknown classes in the target domain (also termed as openness [46]). However, they rely on involving domain adversarial training which could be unstable under some circumstances [14, 81]. Last but not least, we cannot access to the origin source data samples during the training procedure. Previous open set models fail to address such constraints for data privacy protection and they cannot be directly applied into source-free scenario. + +To address the aforementioned issues, we propose Distinguish Then Exploit (DTE) for tackling source-free open set domain adaptation problem. To better distinguish unknown target samples (solving CH1) and assign appropriate labels to the remaining target samples (solving CH2), we utilize two main components in DTE, i.e., weight barcode estimation component and sparse label assignment component. Weight barcode estimation component aims to solve CH1 by adopting marginal probability estimation and weight barcode establishment to automatically determine unknown target samples without prior knowledge. Sparse label assignment component is designed to address CH2 via leveraging the rest known target samples with proximal optimal transport techniques. By cooperating these two components, we can better utilize source knowledge in the target domain by providing more accurate results on source-free open set domain adaptation as shown in Fig.1. We first distinguish the heart-shaped and diamond-shaped unknown target samples. Then we assign the proper pseudo labels for the known target samples with the shapes of triangle, square, and circle respectively. Finally we can classify target unlabeled samples to reach better performance. + +We summarize our main contributions as follows: (1) We propose a novel framework, i.e., DTE, for source-free open set domain adaptation problem via knowledge transfer across domains. (2) We innovatively propose weight bar + +code estimation with partial unbalanced optimal transport to distinguish unknown target samples. Furthermore, we propose target sample reweighting optimization method to figure out the marginal probabilities on partial unbalanced optimal transport. Moreover, we further propose sparse label assignment to provide reliable labels for the known target samples. (3) Extensive empirical studies on various datasets show that DTE achieves the state-of-the-art performance. + +# 2. Related work + +Closed Set Domain Adaptation. Conventional domain adaptation usually assumes that both source and target domains share the same set of labels which namely Closed Set Domain Adaptation [59, 61, 62]. There are mainly two types of existing works, i.e., discrepancy-based method and adversarial-based methods [17, 22, 86, 88]. Discrepancy-based methods usually align the statistics information (e.g., mean and covariance) across domains via Maximum Mean Discrepancy (MMD [83, 93]) or Correlation Alignment (CORAL [82]). More recently, optimal transport techniques [4, 7, 13, 26, 39, 72, 85] have been widely adopted into solving domain adaptation problem [23, 24, 68, 76, 80]. Adversarial-based methods, e.g., DANN [2], CDAN [62] and WDGRL [80] which usually involve domain discriminators during training procedure. However, these approaches could easily lead to negative transfer by mismatching the unknown target samples with the source samples. + +Open Set Domain Adaptation. Different from closed set domain adaptation, open set domain adaptation considers more general scenarios when some target samples may belong to unknown label categories [8, 33, 36, 42, 49, 69, 73, 74, 87, 96]. Most of classic open set domain adaptation models (e.g., OSBP [78] and JPOT [91]) set thresholds to determine known and unknown target samples via useful prior knowledge. More recently, some researches (e.g., STA [46] and UOL [47]) also utilize semantic intracategory information for excavating unknown samples with the absent of relevant knowledge. Meanwhile, with the increasing emphasis on protecting data privacy, target do + +main could be relatively difficult to directly obtain source data [20, 95]. However, only a few approaches have been proposed to tackle source-free open set domain adaptation (e.g., SFPGL [64]) since it is more challenging. How to provide satisfied results on open set domain adaptation under the privacy preserving constraints is an urgent problem. + +# 3. Modelling for DTE + +First, we describe notations. We assume there are two domains, i.e., a source domain $\mathcal{S} = \{(x_i^s,y_i^s)\}_{i = 1}^{n_s}$ with $n_{s}$ labeled samples and a target domain $\mathcal{T} = \{(x_i^t)\}_{i = 1}^{n_t}$ with $n_t$ unlabeled samples. $\mathcal{C}^s$ and $\mathcal{C}^t$ denote the label spaces in source and target domains respectively. Unlike traditional domain adaptation tasks, there exist target samples that belong to additional unknown classes $\mathcal{C}_{\mathrm{unk}}^t$ in the target domain. Specifically, the label space in target domain can be formulated as $\mathcal{C}^t = \mathcal{C}^s\cup \mathcal{C}_{\mathrm{unk}}^t$ . We aim to correctly assign labels for the target samples via knowledge transferring. + +Since the increasing law on data privacy protection, the source data samples cannot be delivered to the target domain. Hence, the target domain can only access the well-trained source model for tackling the source-free openset domain adaptation. We first adopt CNN architecture in source feature extractor $\mathcal{F}_s(\cdot)$ to pretrain on the source data samples to obtain the source features $\mathcal{F}_s(\boldsymbol{x}^s) = \boldsymbol{z}^s$ following [44]. After that we adopt source classifier $\mathcal{G}(\cdot)$ to make output prediction $\pmb{p}^s (y_k|\pmb{x}_i^s)$ : + +$$ +\left[ \mathcal {G} \left(\boldsymbol {x} _ {i} ^ {s}\right) \right] _ {k} = \boldsymbol {p} ^ {s} \left(y _ {k} \mid \boldsymbol {x} _ {i} ^ {s}\right) = \frac {e ^ {\boldsymbol {w} _ {k} ^ {\top} \boldsymbol {z} _ {i} ^ {s} + b _ {k}}}{\sum_ {l = 1} ^ {H} e ^ {\boldsymbol {w} _ {l} ^ {\top} \boldsymbol {z} _ {i} ^ {s} + b _ {l}}}, \tag {1} +$$ + +where $\pmb{w}_k^\top$ and $b_{k}$ denote the $k$ -th class weight and bias in the source classifier $\mathcal{G}(\cdot)$ . Then we adopt cross-entropy loss to train on the source samples following [44]. Finally, we send the well-trained source model including $\mathcal{F}_s(\cdot)$ and $\mathcal{G}(\cdot)$ to the unlabeled target domain as shown in Fig. 2. We then introduce the main components of our proposed DTE, as illustrated in Fig. 2. DTE mainly has two components, i.e., weight barcode estimation and sparse label assignment. Weight barcode estimation aims to first distinguish the target samples that belong to the unknown class categories. Sparse label assignment aims to further exploit the rest target samples by providing them with proper pseudo labels. By incorporating weight barcode estimation with sparse label assignment sequentially, we can achieve more precise results on source-free openset domain adaptation. We will introduce the model details in the following subsections. + +# 3.1. Weight Barcode Estimation + +In this section, we will first introduce the details of weight barcode estimation. Since target domain include many samples which belong to the unknown classes, it is unreasonable to directly apply traditional alignment techniques in source-free domain adaptation. Specifically, the unknown + +# Algorithm 1 The training procedure of proposed DTE + +Input: $\mathcal{F}_s(\cdot)$ : well-trained source feature extractor; $\mathcal{G}(\cdot)$ : well-trained classifier; $\epsilon, \alpha$ : Hyper parameters. + +Set class prototypes $\pmb{\mu}^{(0)} = \pmb{w}$ as the initialization. + +for epoch $= 1$ to $L$ do + +Obtain the optimal solution on $\pmb{f}^{*}$ on PUOT for finding marginal probability $q_{\tau}(\pmb{x}_i^t) = a_i \exp(-f_i^* / \tau)$ . + +Obtain the weight barcode matrix $\widehat{\pmb{Q}}$ and determine known/unknown target samples $\pmb{x}^t = \pmb{x}_{\mathrm{unk}}^t\cup \pmb{x}_{\mathrm{kno}}^t$ + +Obtain the optimal solution $\gamma^{*}$ on solving Eq.(6). + +Gather the network output prediction $\pmb{p}^t$ and the corresponding "ground-truth" $\pmb{q}^t$ on Eq.(9)-Eq.(10). + +Minimize loss function $L_{\mathrm{cls}}$ on Eq.(11) for training. + +Update class prototypes $\mu$ via Eq.(12). + +end for + +Return: The well-trained model in target domain. + +target samples may get matched to the known class categories and causing negative transfer phenomenon [21, 96]. Therefore, it is essential to distinguish the unknown target samples and filter them out during the first stage. To fulfill this challenging task, we propose weight barcode estimation which includes two steps, i.e., marginal probability estimation and weight barcode establishment. Marginal probability estimation adopts Partial Unbalanced Optimal Transport (PUOT) technique to reweight target samples accordingly. Then we utilize weight barcode establishment to automatically determine the unknown target samples. + +Marginal Probability Estimation. To start with, we first adopt target feature extractor $\mathcal{F}_t(\cdot)$ to obtain the latent features $z^t$ on target samples. Then we should consider the matching relationship between target sample features $z^t$ and source class prototype $\mu$ . Recalling that not all target samples belong to the source class categories, it means that only partial of target sample will get matched. It inspired us to propose PUOT for determining which samples belong to the unkown class categories: + +$$ +\min _ {\pi_ {i j} \geq 0} J _ {\text {p u o t}} = \left\langle \mathcal {U}, \boldsymbol {\pi} \right\rangle + \tau \operatorname {K L} \left(\boldsymbol {\pi} \mathbf {1} _ {H} \| \boldsymbol {a}\right) \text {s . t .} \boldsymbol {\pi} ^ {\top} \mathbf {1} _ {N} = \boldsymbol {b}, \tag {2} +$$ + +where $\pi \in \mathbb{R}^{N\times H}$ denotes the coupling matrix of PUOT. $N$ denotes the batchsize of target samples and $H$ denotes the number of shared known classes. $\pmb{a}$ and $\pmb{b}$ denote the corresponding weights for target samples and source prototypes respectively. Specifically we set $a_{i} = 1 / N$ and $b_{j} = 1 / H$ uniformly. $\mathcal{U}$ denotes the pairwise cost between target samples $\boldsymbol{z}^t$ and corresponding prototypes $\pmb{\mu}^{(o)}$ at the $o$ -th iteration as $\mathcal{U}_{ij} = e^{E(\pmb{z}_i^t)}(||(\pmb{z}_i^t)|| - \pmb{\mu}_j^{(o)}||\pmb{\mu}_j^{(o)}||)\|_2^2$ . We set $\pmb{\mu}^{(0)} = \pmb{w}$ for the initialization. Meanwhile $E(\cdot)$ denotes the entropy of classifier output predictions as $E(\pmb{z}_i^t) = -\sum_{j=1}^{|\mathcal{C}^s|} \sigma(\mathcal{G}(\mathcal{F}_t(\pmb{x}_i^t))) [j] \cdot \log(\sigma(\mathcal{G}(\mathcal{F}_t(\pmb{x}_i^t)))) [j])$ where $\sigma(\cdot)$ denotes the softmax operation. When $E(\pmb{x}^t)$ is much larger, it indicates that the model exhibits greater + +![](images/041a74fd4ccee110878314f35814047408f4c2d35817f14f2708859b7530c881.jpg) +Figure 2. The architectures of Distinguish Then Assign (DTE). We first train source feature extractor $\mathcal{F}_s(\cdot)$ and classifier $\mathcal{G}(\cdot)$ in the source domain. Then we share the well-trained source model to the target domain. We fine-tune target feature extractor $\mathcal{F}_t(\cdot)$ based on $\mathcal{F}_s(\cdot)$ and fix classifier $\mathcal{G}(\cdot)$ . Then we propose weight barcode estimation for distinguishing the known and unknown target samples via distinguisher $\hat{\mathcal{G}}(\cdot)$ . Finally we propose sparse label assignment for exploiting appropriate pseudo labels for the remaining target samples. + +![](images/043f30434967b835b04bf6ce7e29c0f2d737b3843b00243b27e1483aa6ab5b43.jpg) +Figure 3. The illustration on the proposed weight barcode estimation with toy examples. We first sample 100 source/target data samples and set $30\%$ or $70\%$ of target data samples from unknown categories. Then we adopt marginal probability estimation to obtain the results of target weights. Finally we calculate and visualize the result of $\hat{Q}$ to distinguish the unknown target samples. + +uncertainty regarding the target samples into known categories. Therefore, it is reasonable to consider such information into the cost calculation on $\mathcal{U}$ . The KL divergence can be calculated via KL $(\pmb{x}||\pmb{z}) = \sum_{d=1}^{D}[x_d\log(x_d/z_d) - x_d + z_d]$ and $\tau$ denotes the hyper parameters. Unlike the classic optimal transport problem, we relax one of the marginal constraints for the target samples via the KL divergence term to avoid the mismatch. To solve the above problem, we first propose the theorem on the dual form of PUOT: + +Theorem 1. Given the PUOT problem $J_{\mathrm{puot}}$ as shown in Eq.(2), it has the following dual form: + +$$ +\min _ {\boldsymbol {f}, \boldsymbol {g}, \zeta} \left[ \tau \sum_ {i = 1} ^ {N} a _ {i} \exp \left(- \frac {f _ {i} + \zeta}{\tau}\right) - \sum_ {j = 1} ^ {H} b _ {j} \left(g _ {j} - \zeta\right) \right] \tag {3} +$$ + +$$ +s. t. f _ {i} + g _ {j} + s _ {i j} = \mathcal {U} _ {i j}, \quad s _ {i j} \geq 0, +$$ + +where $f, g, s$ and $\zeta$ denote Lagrange multipliers. Mean + +while $s_{ij} \geq 0$ according to the KKT conditions. Moreover, PUOT will be transformed into sample-reweighted optimal transport problem according to its dual form: + +$$ +\min _ {\boldsymbol {\pi} \geq 0} \langle \mathcal {U}, \boldsymbol {\pi} \rangle +$$ + +$$ +s. t. \sum_ {j = 1} ^ {H} \pi_ {i j} = a _ {i} \exp \left(- \frac {f _ {i} ^ {*} + \zeta^ {*}}{\tau}\right), \quad \sum_ {i = 1} ^ {N} \pi_ {i j} = b _ {j}. \tag {4} +$$ + +The proof of Theorem 1 is provided in Appendix. That is, the target data samples will be sample-reweighted with coefficient $\exp \left(-\left(f_{i}^{*} + \zeta^{*}\right) / \tau\right)$ accordingly. At that time the marginal probability on target sample $\pmb{x}_i^t$ can be given as $q_{\tau}(\pmb{x}_i^t) = a_i\exp \left(-\left(f_i^* +\zeta^*\right) / \tau\right)$ . Kindly note that we do not necessarily require an explicit expression for solving $\pi$ , which involves $(N\times H)$ variables and can be quite complex. Meanwhile we mainly focus on solving these dual form Lagrange multipliers in Eq.(3) with $(N + H)$ variables. Specifically, we can utilize the $c$ -transform [41, 55, 71] in classical optimal transport theorem to reach the upper bound of $\pmb{g}$ as $g_{j} = \inf_{k\in [N]}\left(\mathcal{U}_{kj} - f_{k}\right)$ . By taking the upper bound of $\pmb{g}$ into Eq.(3), we can reach the final optimization formula as: + +$$ +\min _ {\boldsymbol {f}, \zeta} \left[ \tau \sum_ {i = 1} ^ {N} a _ {i} e ^ {- \frac {f _ {i} + \zeta}{\tau}} - \sum_ {j = 1} ^ {H} \left(\inf _ {k \in [ N ]} \left(\mathcal {U} _ {k j} - f _ {k}\right) - \zeta\right) b _ {j} \right]. \tag {5} +$$ + +That equation can be optimized easily. Due to the space limited, the details of optimization procedures for Eq.(5) are provided in Appendix. In summary, we only need to figure out the marginal probability on target samples $q_{\tau}(\boldsymbol{x}_i^t)$ via solving the optimal results of $f$ and $\zeta$ in Eq.(5). + +Weight Barcode Establishment. According to the optimization process, we can observe that choosing different value of $\tau$ could lead to different marginal probability. Specifically, as the value of $\tau$ increases, a larger number of target samples will be assigned higher weights, enhancing their likelihood of being matched. However, we do not + +know the exact percentage of how many samples are unknown in the target domain which makes it relatively hard to choose a proper value of $\tau$ . To address this issue, we propose an innovative weight barcode mechanism to effectively identify target samples that are more likely belong to unknown classes. We first calculate $q_{\tau}(\boldsymbol{x}^t)$ by varying $\tau$ in range of $\{0.01, 0.05, 0.1, 0.5, 1, 5, 10\}$ and collecting them as $\boldsymbol{Q} \in \mathbb{R}^{N \times 7}$ . Then we can establish the binary weight barcode matrix $\widehat{\boldsymbol{Q}} \in \mathbb{R}^{N \times 7}$ with the following process as $\widehat{\boldsymbol{Q}}_{ij} = \mathrm{func}(Q_{ij} - 1/N)$ where $\mathrm{func}(x) = 1$ when $x \geq 0$ otherwise $\mathrm{func}(x) = 0$ . It is reasonable since PUOT will enhance the target sample weights if they exhibit a higher degree of similarity to the source data. Meanwhile the rest of unlike target samples will gain less weights and we set then $\widehat{\boldsymbol{Q}}_{ij} = 0$ . To make this process more intuitive, we further provide an numerical example in Fig. 3. Specifically, the red and blue points denote the source and target data samples respectively. We first figure out the pairwise distance between red data $\boldsymbol{z}_i^{\mathrm{Red}}$ and blue data $\boldsymbol{z}_j^{\mathrm{Blue}}$ via $\mathcal{U}_{ij} = ||\boldsymbol{z}_i^{\mathrm{Red}} - \boldsymbol{z}_j^{\mathrm{Blue}}||_2^2$ . Then we calculate the weight barcode as $\widehat{\boldsymbol{Q}}$ by using marginal probability estimation and visualize the results. We can observe that the unknown target samples achieve lower value of $\widehat{\boldsymbol{Q}}$ which can be easily identify. It indicates that the proposed method is adaptable to varying ratios of unknown data samples, even in the absence of prior knowledge. + +After we obtain the binary weight barcode matrix $\widehat{Q}$ , we can further adopt $K$ -Means algorithm [28] on $\widehat{Q}$ to cluster them into two groups with corresponding centers $\mathcal{C}_0$ and $\mathcal{C}_1$ . We further denote $\mathcal{C}_{\mathrm{known}} = \max \left(\| \mathcal{C}_0\| _2, \| \mathcal{C}_1\| _2\right)$ and $\mathcal{C}_{\mathrm{unknown}} = \min \left(\| \mathcal{C}_0\| _2, \| \mathcal{C}_1\| _2\right)$ by considering the norm of centers. Then we can determine the target samples by measuring the Euclidean distance among $\widehat{Q}$ and centers $\mathcal{C}_{\mathrm{known}}$ , $\mathcal{C}_{\mathrm{unknown}}$ . If the target samples are found to be in close proximity to $\mathcal{C}_{\mathrm{unknown}}$ , we can classify them as unknown data as $\boldsymbol{x}_{\mathrm{unk}}^t$ . The remaining target domain samples can all be considered as belonging to the known categories as $\boldsymbol{x}_{\mathrm{kn0}}^t$ and as $\boldsymbol{x}^t = \boldsymbol{x}_{\mathrm{unk}}^t \cup \boldsymbol{x}_{\mathrm{kn0}}^t$ . In summary, we can now determine the known/unknown target samples accordingly. + +# 3.2. Sparse Label Assignment + +Since we have exploited the target samples with unknown classes $\boldsymbol{x}_{\mathrm{unk}}^{t}$ via proposed weight barcode estimation method, we should explore the possible labels for the rest target samples $\boldsymbol{x}_{\mathrm{kno}}^{t}$ . Recall that $\boldsymbol{x}_{\mathrm{kno}}^{t}$ share the same categories with the corresponding prototypes $\boldsymbol{\mu}^{s}$ , one can adopt domain adaptation manners for finding the latent relationship. Meanwhile, optimal transport techniques could be rather useful in reducing the domain discrepancy via discovering the most proper pairwise matching across domains [9, 25, 75]. However, directly optimizing optimal transport with linear programming would involve heavy computation cost [1, 34, 92]. Although utilizing entropy regularization + +term could enhance the efficiency, it will bring about relatively dense matching solution which is always inaccurate as shown in Fig. 4(a) [6, 10, 50]. To alleviate this issue, we propose Sparse Sample-Label Matching (SSLM) on optimal transport via involving $\ell_2$ -norm proximal term: + +$$ +\min _ {\gamma \in \Delta} \ell = \sum_ {i = 1} ^ {N _ {\mathrm {k n o}}} \sum_ {j = 1} ^ {H} \left[ \gamma_ {i j} C _ {i j} + \frac {\epsilon}{2} \left(\gamma_ {i j} - \gamma_ {i j} ^ {(l)}\right) ^ {2} \right], \tag {6} +$$ + +where $\Delta = \{\sum_{j=1}^{H} \gamma_{ij} = \widehat{a}_i = 1, \sum_{i=1}^{N_{\mathrm{kno}}} \gamma_{ij} = \widehat{b}_j = N_{\mathrm{kno}} / H, \gamma_{ij} \geq 0\}$ and $N_{\mathrm{kno}}$ denotes the number of target samples $\boldsymbol{x}_{\mathrm{kno}}^t$ . $C = -\log[\sigma[\mathcal{G}(\mathcal{F}_t(\boldsymbol{x}_{\mathrm{kno}}^t))]]$ denote the output classification probability with negative logarithm. Indeed, when $C_{ij}$ is smaller, it indicates that $\boldsymbol{x}_{\mathrm{kno},i}^t$ has a higher probability of belonging to the $j$ -th class [3, 27, 45]. $\widehat{a}_i$ and $\widehat{b}_j$ denote the samples weights and we set $\widehat{a}_i = 1$ , $\widehat{b}_j = \frac{N_{\mathrm{kno}}}{H}$ respectively. Meanwhile $\sum_{i=1}^{M} \sum_{j=1}^{N} (\gamma_{ij} - \gamma_{ij}^{(l)})^2$ denotes the proximal point term with $\ell_2$ -norm. Incorporating $\ell_2$ -norm with proximal point method [90], we can reach more sparse and robust results on transportation plan $\gamma^*$ . We initialize $\gamma_{ij}^{(0)} = \frac{1}{H}$ for the following iterations. Then we can easily obtain the Fenchel-Legendre dual form of the original problem as: + +$$ +\max _ {\psi , \phi} \sum_ {i = 1} ^ {N _ {\mathrm {k n o}}} \psi_ {i} \widehat {a} _ {i} + \sum_ {j = 1} ^ {H} \phi_ {j} \widehat {b} _ {j} - \frac {\epsilon}{2} \sum_ {i = 1} ^ {N _ {\mathrm {k n o}}} \sum_ {j = 1} ^ {H} \left[ \frac {\psi_ {i} + \phi_ {j} - \widehat {C} _ {i j}}{\epsilon} \right] _ {+} ^ {2}, +$$ + +where $\phi$ and $\psi$ denote the Lagrange multipliers. Meanwhile we set $[x]_{+}^{2} = (\max (0,x))^{2}$ and $\widehat{C}_{ij} = C_{ij} - \epsilon \gamma_{ij}^{(l)}$ . Note that the original constrained optimization problem in Eq.(6) turns to be unconstrained after the duality transformation in Eq.(7). The optimization procedure on solving variables $\psi$ and $\phi$ are provided in Appendix. After that, we can further calculate the optimal solution of $\gamma^{(l + 1)}$ at the $l$ -th iteration as shown: + +$$ +\gamma_ {i j} ^ {(l + 1)} = \max (0, [ \psi_ {i} + \phi_ {j} + \epsilon \gamma_ {i j} ^ {(l)} - C _ {i j} ] / \epsilon). \tag {8} +$$ + +After we obtain the optimal solution on $\gamma^{*}$ via solving sparse sample-prototype matching, we can reach the reliable pseudo labels for these target samples $\pmb{x}_{\mathrm{kno}}^t$ . Furthermore, the optimal solution $\gamma^{*}$ can yield more sparser and clearer results compared to using only the $\ell_{2}$ -norm regularization, as depicted in Fig. 4(b)-(c). + +![](images/d00b1878bcffeb0e669ff1bb5730968ed33c80b86d2e812e31e0499e4dd7fc2a.jpg) +Figure 4. The illustration on Sparse Sample-Label Matching (SSLM) in sparse label assignment with toy examples. Sinkhorn and tradition $\ell_2$ -norm solvers could easily reach smooth but dense output solutions. Our proposed SSLM method can reach more sparse matching solution for better performance. + +![](images/a1db6d8aa1371d412013e8092b8e25829091aeffb96cbfa6600d46b79936f023.jpg) +(a) JPOT on Office-31 + +![](images/2511ba83b85b9efae26ba587c5599bffc615004fe1de4cf38675dd87e067181b.jpg) +(b) DTE on Office-31 +Figure 5. The t-SNE visualization of $\mathbf{W} \rightarrow \mathbf{A}$ tasks on Office-31. Each color denotes a class and the reddish-brown color is used to represent the unknown class. + +# 3.3. Putting Together + +We have first distinguished the target samples with unknown classes via weight barcode estimation, then assigning the rest target samples to the corresponding classes via sparse sample-prototype matching. Therefore, we should gather the output prediction of these target samples as $\pmb{p}^t$ : + +$$ +\boldsymbol {p} ^ {t} = \operatorname {s o f t m a x} \left[ \mathcal {G} \left(\mathcal {F} _ {t} \left(\boldsymbol {x} ^ {t}\right)\right) \oplus \widehat {\mathcal {G}} \left(\mathcal {F} _ {t} \left(\boldsymbol {x} ^ {t}\right)\right) \right] \in \mathbb {R} ^ {N \times (H + 1)}. \tag {9} +$$ + +Note that $\widehat{\mathcal{G}} (\cdot)$ is a newly add trainable target classifier to predict whether samples are unknown. Here we concatenate the output predictions on both source classifier $\mathcal{G}(\cdot)$ and target classifier $\widehat{\mathcal{G}} (\cdot)$ then process the softmax operation. Meanwhile we can formulate the estimated "ground-truth" $q_{\mathrm{kno}}^{t}$ and $q_{\mathrm{unk}}^{t}$ for the target samples as follows: + +$$ +\boldsymbol {q} _ {\mathrm {k n o}} ^ {t} = \left[ \gamma^ {*} \oplus 0 \right] \in \mathbb {R} ^ {N _ {\mathrm {k n o}} \times (H + 1)}, \tag {10} +$$ + +$$ +\boldsymbol {q} _ {\mathrm {u n k}} ^ {t} = [ \mathbf {0} \oplus 1 ] \in \mathbb {R} ^ {N _ {\mathrm {u n k}} \times (H + 1)}. +$$ + +Then we can utilize the cross-entropy loss $L_{\mathrm{cls}}$ on fine-tuning the target feature extractor $\mathcal{F}_t(\cdot)$ based on $\mathcal{F}_s(\cdot)$ and training distinguisher $\widehat{\mathcal{G}}_t(\cdot)$ on the target samples as: + +$$ +\min _ {\mathcal {F} _ {t} (\cdot), \widehat {\mathcal {G}} (\cdot)} L _ {\mathrm {c l s}} = - \sum_ {i = 1} ^ {N _ {\mathrm {u n k}} + N _ {\mathrm {k n o}}} \sum_ {j = 1} ^ {H + 1} q _ {i j} ^ {t} \log p _ {i j} ^ {t}. \tag {11} +$$ + +Note that we fix the source classifier $\mathcal{G}(\cdot)$ during the training procedure. Meanwhile, the loss function is easy to implemented for end-to-end training. Moreover, we should update the class prototype $\pmb{\mu}^{(o + 1)}$ at the $(o + 1)$ -th iteration via incremental learning strategy: + +$$ +\boldsymbol {\mu} _ {j} ^ {(o + 1)} = \alpha \boldsymbol {\mu} _ {j} ^ {(o)} + \frac {1 - \alpha}{\sum_ {i = 1} ^ {N _ {\mathrm {k n o}}} p _ {i j} ^ {t}} \sum_ {i = 1} ^ {N _ {\mathrm {k n o}}} \mathcal {F} _ {t} \left(\boldsymbol {x} _ {\mathrm {k n o}, i} ^ {t}\right) p _ {i j} ^ {t}, \tag {12} +$$ + +where $\alpha \in [0,1]$ denotes momentum hyper parameter. The whole algorithm of DTE in provided in Alg.1. + +# 4. Experiment + +# 4.1. Setup + +We conduct the experiments on several real-world datasets, e.g., Office-31 [77] with following three domains: Amazon (A), Webcam (W) and DSLR (D); Office-Home [84] with following four dissimilar domains: Artistic images (Ar), + +Clip Art (CI), Product images (Pr) and Real-World (Rw); VisDA-2017 [70] with domains Synthetic, Real; and Digits with following three domains MNIST [40], USPS [31] and SVHN [67]. We adopt the same set of known and unknown classes following previous works [46, 78]. + +Implementation Details. We first set batch size $N = 256$ during the training procedure. For the experiments on Office-31, Office-Home, and VisDA, we first fine-tune on the backbone network ResNet-50 [29] in the source domain. For the experiments on Digits, we adopt LeNet [40] to extract the sample features. After the training in the source domain, we only transfer the well-trained feature extractor $\mathcal{F}_s(\cdot)$ and classifier $\mathcal{G}(\cdot)$ to the target domain. We set $\epsilon = 0.5$ in sparse label assignment. Meanwhile we set $\alpha = 0.8$ for the incremental learning on class prototypes. We choose Adam [35] as the optimizer and set the learning rate as 0.001 and 0.01 for fine-tuning target feature extractor $\mathcal{F}_t(\cdot)$ based on $\mathcal{F}_s(\cdot)$ and training distinguisher $\widehat{\mathcal{G}} (\cdot)$ respectively. We employ these four evaluation metrics, i.e., OS, OS*, UNK, HOS and for open set domain adaptation following [78]. Specifically, OS and OS* denotes the accuracy averaged over all classes or only known classes respectively. UNK denotes the accuracy of unknown instances and HOS denotes the harmonic value of OS* and UNK. We compare the proposed DTE with state-of-the-art domain adaptation methods: ResNet [29], Domain Adversarial Neural Network (DANN) [2], Deep Joint Distribution Optimal Transport (DeepJDOT) [11], OpenMAX [5], MMD+OSVM, DANN+OSVM, Openset Back-Propagation (OSBP) [78], Separate To Adapt (STA) [46], Towards Inheritable Models (TIM) [37], Joint Partial Optimal Transport (JPOT) [91], Source Hypothesis Transfer (SHOT) [44], Progressive Graph Learning (PGL) [63], Source-Free Progressive Graph Learning (SFPGL) [64], Unknown-Aware Domain Adversarial Learning (cUADAL) [32], Self-supervised Class-Disccovering Adapter (SCDA) [97], Unknown Oriented Learning (UOL) [47], Adjustment and Alignment (ANNA) [43] and Balanced Correlation Learning (BCL) [30]. We perform five random experiments and report the average and variation results. + +# 4.2. Result and Discussion + +The comparison results on Office-31 are shown in Table. 1. Directly applying ResNet could lead to worst results since the existence of domain discrepancy. Although previous method (e.g., OSBP, STA) adopted domain adversarial strategy to reduce the discrepancy, it could be unstable and hard to train in the real practice [81]. Our proposed DTE achieves the best performance among other approaches. Even for the difficult task (e.g., $\mathbf{A} \rightarrow \mathbf{W}$ , $\mathbf{W} \rightarrow \mathbf{A}$ ) where these images are quite different across domains, DTE still obtains satisfactory results. It indicates the efficacy of the proposed DTE for dealing with open set domain + +Table 1. Classification accuracy (%) on Office-31 for open set domain adaptation (ResNet-50). + +
MethodA→WA→DD→WW→DD→AW→AAvg
OSOS*OSOS*OSOS*OSOS*OSOS*OSOS*OSOS*
ResNet82.5±1.282.7±0.985.2±0.385.5±0.994.1±0.394.3±0.796.6±0.297.0±0.471.6±1.071.5±1.175.5±1.075.2±1.684.284.4
DANN85.3±0.787.7±1.186.5±0.687.7±0.697.5±0.298.3±0.599.5±0.1100.0±075.7±1.676.2±0.974.9±1.275.6±0.886.687.6
DeepJDOT86.1±0.588.7±0.986.9±0.789.0±0.596.9±0.295.6±0.896.1±0.198.1±0.585.6±1.288.4±1.081.5±1.383.1±0.988.990.5
OpenMax87.4±0.587.5±0.387.1±0.988.4±0.996.1±0.496.2±0.398.4±0.398.5±0.383.4±1.082.1±0.682.8±0.982.8±0.689.089.3
OSBP86.5±2.087.6±2.188.6±1.489.2±1.397.0±1.096.5±0.497.9±0.998.7±0.688.9±2.590.6±2.385.8±2.584.9±1.390.891.3
STA89.5±0.692.1±0.593.7±1.596.1±0.497.5±0.296.5±0.599.5±0.299.6±0.189.1±0.593.5±0.887.9±0.987.4±0.692.994.1
TIM91.3±0.793.2±0.594.2±1.197.1±0.896.5±0.597.4±0.799.5±0.299.4±0.390.1±0.291.5±0.288.7±1.388.4±0.993.494.5
JPOT92.8±0.692.2±0.495.2±0.996.0±0.698.1±0.396.2±0.499.5±0.198.6±0.293.0±0.794.1±0.489.9±1.088.4±0.494.694.4
SHOT88.8±0.791.4±0.490.3±0.592.6±0.396.2±0.497.0±0.497.4±0.397.9±0.591.6±0.593.4±0.791.2±0.593.5±0.492.694.3
PGL89.2±0.590.1±0.789.6±0.891.6±0.595.3±0.495.1±0.596.7±0.597.6±0.671.0±1.372.0±1.873.0±1.277.6±1.685.887.4
SFPGL90.4±0.391.9±0.791.0±1.091.2±0.897.3±0.595.5±0.696.1±0.498.4±0.474.8±1.575.6±2.074.9±1.679.6±1.187.488.7
cUADAL92.5±0.493.7±0.696.0±0.797.5±0.698.4±0.4100.0±098.6±0.5100.0±085.3±0.886.0±0.583.2±0.783.9±0.992.393.5
SCDA95.7±0.197.5±0.295.9±0.496.5±0.499.2±0.499.7±0.199.8±0.1100.0±092.1±0.193.7±0.392.2±0.193.6±0.195.896.8
UOL94.6±0.396.8±0.695.4±0.597.2±0.799.5±0.698.2±0.398.9±0.299.1±0.392.7±0.693.2±0.891.6±1.093.0±0.895.596.2
DTE-W93.9±0.694.5±0.595.7±0.696.3±0.798.2±0.599.1±0.297.4±0.498.8±0.694.0±0.993.6±1.290.7±0.891.5±0.794.995.6
DTE-S95.3±0.596.9±0.496.5±0.697.0±0.898.7±0.799.4±0.398.1±0.499.2±0.594.8±1.193.9±1.092.5±0.794.1±0.995.996.7
DTE96.1±0.297.5±0.496.8±0.497.9±0.598.9±0.5100.0±098.5±0.399.4±0.395.6±0.594.3±0.692.9±0.494.6±0.396.497.2
+ +Table 2. Classification accuracy OS* (%) on Office-Home for open set domain adaptation (ResNet-50). + +
MethodAr→ClPr→ClRw→ClAr→PrCl→PrRw→PrCl→ArPr→ArRw→ArAr→RwCl→RwPr→RwAvg
ResNet53.4±0.452.7±0.651.9±0.569.3±0.761.8±0.574.1±0.461.4±0.664.0±0.370.0±0.378.7±0.671.0±0.674.9±0.965.3
DANN54.6±0.749.7±1.651.9±1.469.5±1.163.5±1.072.9±0.861.9±1.263.3±1.071.3±1.080.2±0.871.7±0.474.2±0.465.4
OpenMax56.5±0.452.9±0.753.7±0.469.1±0.364.8±0.474.5±0.664.1±0.964.0±0.871.2±0.880.3±0.873.0±0.576.9±0.366.7
OSBP56.7±1.951.5±2.149.2±2.467.5±1.565.5±1.574.0±1.562.5±2.064.8±1.169.3±1.180.6±0.974.7±2.271.5±1.965.7
STA58.1±0.653.1±0.954.4±1.071.6±1.269.3±1.081.9±0.563.4±0.565.2±0.874.9±1.085.0±0.275.8±0.480.8±0.369.5
TIM60.1±0.754.2±1.056.2±1.770.9±1.470.0±1.778.6±0.664.0±0.666.1±1.374.9±0.983.2±0.975.7±1.381.3±1.469.6
JPOT59.6±0.554.2±0.754.6±0.972.3±1.170.1±0.682.1±0.962.9±0.768.3±0.875.1±1.184.8±0.477.4±0.581.2±0.470.2
SHOT64.5±0.859.3±0.564.6±0.880.4±0.975.4±1.282.3±1.063.1±0.865.3±0.669.6±0.784.7±0.681.2±1.083.3±0.872.8
PGL61.6±1.058.4±0.865.0±0.777.1±1.572.0±1.383.0±0.968.8±1.472.2±0.778.6±1.385.9±0.982.8±1.182.6±1.274.0
SFPGL62.4±0.559.0±0.764.1±0.977.8±0.772.7±0.884.2±1.070.5±0.673.4±0.578.3±0.886.8±0.683.9±0.881.4±0.774.5
cUADAL62.3±0.660.1±0.861.7±0.277.9±0.671.4±0.981.3±0.572.3±0.671.5±0.873.9±1.083.6±0.783.5±0.784.0±1.073.6
SCDA59.9±0.359.0±0.362.8±0.579.6±0.473.8±1.083.7±0.870.9±0.572.3±0.675.5±0.485.3±0.682.9±0.385.7±0.974.3
UOL61.0±0.759.5±0.663.3±0.680.7±1.074.9±0.885.8±0.771.6±0.972.9±0.675.1±0.588.4±0.883.2±0.486.3±0.974.4
ANNA61.3±0.960.4±0.762.5±0.882.1±0.673.6±0.784.4±0.570.2±1.171.6±0.975.7±0.686.9±0.582.0±1.286.9±1.574.8
BCL58.4±0.559.7±0.661.1±0.680.6±0.573.0±0.985.1±0.470.5±0.772.2±0.674.8±0.587.5±0.683.4±0.986.1±1.274.4
DTE-W60.1±0.957.2±0.859.4±0.978.5±0.672.3±1.083.0±0.669.1±0.871.9±0.775.8±1.186.7±0.581.6±0.682.8±1.573.2
DTE-S62.1±0.660.9±0.861.8±0.880.4±0.574.2±1.285.7±0.971.6±0.973.3±1.076.9±0.788.3±0.882.0±0.886.5±1.376.1
DTE63.2±0.462.6±0.663.9±0.581.5±0.575.4±0.886.8±0.672.0±0.574.1±0.477.3±0.790.2±0.684.3±0.487.5±1.176.5
+ +Table 3. Classification accuracy (%) on Digits (LeNet) and VisDA2017 (VGGNet) for open set domain adaptation. + +
MethodSVHN→MNISTUSPS→MNISTMNIST→USPSAvgVisDA2017
OSOS*HOSUNKOSOS*HOSUNKOSOS*HOSUNKOSOS*HOSUNKbicyclebuscarmotorcycletraintruckOSOS*
OSVM54.363.117.710.543.132.348.597.179.877.983.289.359.157.749.865.731.751.666.570.488.520.852.554.9
MMD+OSVM55.964.720.112.262.858.968.782.380.079.880.481.066.268.856.458.439.050.164.279.986.616.354.456.0
DANN+OSVM62.975.31.780.9077.192.41.190.6033.835.529.525.360.469.410.88.9331.856.671.777.487.022.355.557.8
OSBP63.059.168.982.592.391.294.497.892.194.985.778.182.481.782.986.151.167.142.884.281.828.062.959.2
STA76.975.479.784.492.291.393.996.793.094.988.883.587.387.287.588.252.469.659.987.886.527.266.863.9
TIM80.479.083.087.495.495.196.096.993.093.591.890.589.689.390.391.653.569.262.285.785.432.568.164.7
JPOT77.275.380.686.792.491.294.798.492.992.194.496.987.585.489.994.055.370.064.189.587.233.769.566.6
UOL84.484.185.085.996.997.196.595.995.395.195.796.392.291.692.492.754.670.563.791.286.931.470.066.3
DTE-W83.182.883.784.696.596.496.797.094.594.295.196.090.391.491.892.556.070.965.688.884.132.269.966.2
DTE-S84.884.585.486.397.096.997.297.594.994.695.596.492.592.292.793.456.571.266.390.187.733.871.367.6
DTE85.585.286.187.097.697.597.898.195.695.496.096.692.992.793.393.957.472.067.990.888.635.272.468.7
+ +adaptation even under the source-free scenario. + +Meanwhile the results of Office-Home are shown in Table. 2. The tasks on Office-Home are more difficult than Office-31 since it has more categories with large domain gap. Previous methods cannot better reduce domain discrepancy while involving the negative transfer, leading to the limited model performance. Meanwhile our proposed DTE still outperforms the comparison methods. Note that the proposed DTE with weight barcode estimation can fully leverage knowledge from source domain to a significant degree compared with JPOT. Moreover, we conduct the experiments on VisDA-2017 and Digits and report the results in Table. 3. It is remarkable that our proposed DTE can pro + +mote the better results of OS and $\mathbf{OS^{*}}$ on hard transfer tasks in VisDA-2017. Specifically, in terms of the performance of all classes, DTE has achieved significant improvements compared to the previous models, with notable examples being a $2.27\%$ improvement on Office-Home and a $3.15\%$ improvement on VisDA-2017 against runner-up models for $\mathbf{OS^{*}}$ . DTE can better distinguish the unknown target samples meanwhile correctly assign pseudo labels for the rest target samples to reach state-of-the-art performance. + +# 4.3. Analysis. + +Ablation Study. To further study how does each module of DTE contribute on the final performance, we com + +![](images/12bb83a3c026491d4597acb58d1810f3a2d186e007ce2ec0f30abdc33847d5f5.jpg) +(a) Openness on Digits + +![](images/e298eb90407f6ba113085d004d381ae78a4d684283ddae5876075a7a37606652.jpg) +(b) Openness on Office-31 + +![](images/034a113e41586505f4f06265bab5afe6f5b099ff5f274df858022a57c16179dd.jpg) +Figure 6. The openness experiments of $\mathbb{O}$ on Digits and Office-31. +(a) Parameters on Digits +Figure 7. Hyper parameter tuning of $\epsilon$ on Digits and Office-31. + +![](images/90f713d26bc0db61faf8be3c5f15ed126d9e287bd656092f2a170b1c25f1975b.jpg) +(b) Parameters on Office-31 + +pare DTE with its several variants, including DTE-W and DTE-S. DTE-W only replaces the proposed weight barcode estimation with JPOT by calculating mean transport cost to determine the unknown target samples. DTE-S only replaces sparse label assignment with self-labelling strategy via Sinkhorn algorithm [3]. The comparison results on Office-Home and Office-31 are shown in Table. 2 and Table. 1. From it, we can observe that (1) Although DTE-W exceeds most of previous methods, it cannot precisely distinguish unknown data samples especially under source-free domain adaptation scenario. (2) DTE-S achieves better results than DTE-W, while it still cannot reach more satisfactory results than DTE. That is because involving entropy regularization term could reach relatively dense solution, thus leading to the inaccurate solutions. (3) DTE achieves the best performance which indicates the efficacy of incorporating weight barcode estimation and sparse label assignment for distinguishing then exploiting target samples. + +Feature Visualization. We further provide the feature visualization via T-SNE [38] of the target sample representations $\pmb{x}^t$ by JPOT and DTE on $\mathbf{W} \rightarrow \mathbf{A}$ (Office-31). The results are depicted in Fig. 5(a)-(b), where each color represents a specific class, and the reddish-brown color is used to represent the unknown class. From there we can observe that JPOT cannot better distinguish known and unknown target samples with ambiguous decision boundary. The primary reason is that JPOT can only utilizes limited source knowledge for open set domain adaptation. Meanwhile DTE can better distinguish unknown target samples and cluster the rest known samples accordingly. + +Openness. To verify the proposed DTE can be applied to the case with different proportion of unknown classes, we adjust the openness $\mathbb{O}$ in the openset domain adaptation. + +Specifically, $\mathbb{O} = 1 - \frac{|\mathcal{C}^s|}{|\mathcal{C}^t|}$ denotes different level of openness [46]. Larger openness indicates that fewer classes are known across domains. We vary $\mathbb{O}$ from 0 to nearly 1 on Office-31 and Digits accordingly [46] then report the result of OS in Fig. 6(a)-(b). We can observe that the majority of previous methods (e.g., DANN and JPOT) cannot better adjust to different proportion of unknown classes in the target domain. Meanwhile, the proposed DTE can always reach better performance even openness $\mathbb{O}$ approaches to 0 or 1. It further demonstrates the proposed DTE is robust to openness without the requirement of prior knowledge. + +Table 4. The results (%) for method extension on VisDA2017 + +
MethodbicyclebuscarmotorcycletraintruckOSOS*
PGL93.593.875.798.896.238.580.782.8
SFPGL93.697.689.695.396.795.691.094.7
DTE(VGGNet+GNN)93.594.077.296.496.041.380.283.1
DTE(ResNet+GNN)93.897.990.197.996.592.091.694.7
+ +Method Extension. Furthermore, we analyse the general extension of our proposed DTE. Specifically, we incorporate weight barcode estimation and sparse label assignment with VGGNet/ResNet and Graph Neural Network (GNN for transductive feature learning [64, 89]) on VisDA2017. The results of OS and $\mathbf{OS^{*}}$ are shown in Table 4. We observe that the proposed DTE can be applied with various backbones for achieving better performance. + +Hyper Parameter. We finally study the effects of hyperparameters on model performance. For DTE, we vary $\epsilon$ in $\{0.01, 0.05, 0.1, 0.5, 1, 10\}$ for sparse label assignment on Digits/Office-31 and report the results of OS/OS* in Fig. 7(a)-(b). We can conclude that when $\epsilon$ is larger, the proximal regularization term will have a more important impact compared to sparse-label matching, which can result in poorer performance. Hence we set $\epsilon = 0.5$ in the experiments empirically. + +# 5. Conclusion + +In this paper, we propose Distinguish Then Exploit model (DTE) for tackling the source-free open set domain adaptation problem. DTE first adopts weight barcode estimation with proposed target sample reweighting optimization method on partial unbalanced optimal transport to determine the known and unknown target samples without prior information. Then DTE utilizes sparse sample-label matching on optimal transport to further leverage useful and accurate source knowledge in the target domain. We also conduct extensive experiments on different datasets to show the superior performance of the proposed DTE model. + +# 6. Acknowledgment + +This work was supported by the National Natural Science Foundation of China (No. 62372242) and Natural Science Foundation of Shandong Province (No. ZR2023MF007). + +# References + +[1] Ravindra K Ahuja, Thomas L Magnanti, and James B Orlin. Network flows. 1988. 5 +[2] Hana Ajakan, Pascal Germain, Hugo Larochelle, François Laviolette, and Mario Marchand. Domain-adversarial neural networks. arXiv preprint arXiv:1412.4446, 2014. 2, 6 +[3] Yuki Markus Asano, Christian Rupprecht, and Andrea Vedaldi. Self-labelling via simultaneous clustering and representation learning. arXiv preprint arXiv:1911.05371, 2019. 5, 8 +[4] Yikun Bai, Bernhard Schmitzer, Matthew Thorpe, and Soheil Kolouri. Sliced optimal partial transport. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 13681-13690, 2023. 2 +[5] Abhijit Bendale and Terrance E Boult. Towards open set deep networks. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 1563-1572, 2016. 6 +[6] Mathieu Blondel, Vivien Seguy, and Antoine Rolet. Smooth and sparse optimal transport. In International conference on artificial intelligence and statistics, pages 880-889. PMLR, 2018. 5 +[7] Nicolas Bonneel and David Coeurjolly. Spot: sliced partial optimal transport. ACM Transactions on Graphics (TOG), 38(4):1-13, 2019. 2 +[8] Silvia Bucci, Mohammad Reza Loghmani, and Tatiana Tommasi. On the effectiveness of image rotation for open set domain adaptation. In European conference on computer vision, pages 422-438. Springer, 2020. 2 +[9] Nicolas Courty, Rémi Flamary, and Devis Tuia. Domain adaptation with regularized optimal transport. In Joint European Conference on Machine Learning and Knowledge Discovery in Databases, pages 274-289. Springer, 2014. 5 +[10] Marco Cuturi. Sinkhorn distances: Lightspeed computation of optimal transport. Advances in neural information processing systems, 26, 2013. 5 +[11] Bharath Bhushan Damodaran, Benjamin Kellenberger, Rémi Flamary, Devis Tuia, and Nicolas Courty. Deepjdot: Deep joint distribution optimal transport for unsupervised domain adaptation. In European Conference on Computer Vision, pages 467-483. Springer, 2018. 6 +[12] Jun Dan, Weiming Liu, Chunfeng Xie, Hua Yu, Shunjie Dong, and Yanchao Tan. Tfgda: Exploring topology and feature alignment in semi-supervised graph domain adaptation through robust clustering. In The Thirty-eighth Annual Conference on Neural Information Processing Systems. 1 +[13] Jun Dan, Tao Jin, Hao Chi, Shunjie Dong, and Yixuan Shen. Uncertainty-guided joint unbalanced optimal transport for unsupervised domain adaptation. Neural Computing and Applications, 35(7):5351-5367, 2023. 2 +[14] Jun Dan, Tao Jin, Hao Chi, Shunjie Dong, Haoran Xie, Keying Cao, and Xinjing Yang. Trust-aware conditional adversarial domain adaptation with feature norm alignment. Neural Networks, 168:518-530, 2023. 2 +[15] Jun Dan, Tao Jin, Hao Chi, Yixuan Shen, Jiawang Yu, and Jinhai Zhou. Homda: High-order moment-based domain + +alignment for unsupervised domain adaptation. Knowledge-Based Systems, 261:110205, 2023. 1 +[16] Jun Dan, Yang Liu, Haoyu Xie, Jiankang Deng, Haoran Xie, Xuansong Xie, and Baigui Sun. Transface: Calibrating transformer training for face recognition from a data-centric perspective. In Proceedings of the IEEE/CVF international conference on computer vision, pages 20642-20653, 2023. 1 +[17] Jun Dan, Mushui Liu, Chunfeng Xie, Jiawang Yu, Haoran Xie, Ruokun Li, and Shunjie Dong. Similar norm more transferable: Rethinking feature norms discrepancy in adversarial domain adaptation. Knowledge-Based Systems, 296: 111908, 2024. 2 +[18] Jun Dan, Weiming Liu, Mushui Liu, Chunfeng Xie, Shunjie Dong, Guofang Ma, Yanchao Tan, and Jiazheng Xing. Hogda: Boosting semi-supervised graph domain adaptation via high-order structure-guided adaptive feature alignment. In Proceedings of the 32nd ACM International Conference on Multimedia, pages 11109-11118, 2024. 1 +[19] Jun Dan, Yang Liu, Jiankang Deng, Haoyu Xie, Siyuan Li, Baigui Sun, and Shan Luo. Topofr: A closer look at topology alignment on face recognition. arXiv preprint arXiv:2410.10587, 2024. 1 +[20] Yuqi Fang, Pew-Thian Yap, Weili Lin, Hongtu Zhu, and Mingxia Liu. Source-free unsupervised domain adaptation: A survey. arXiv preprint arXiv:2301.00265, 2022. 3 +[21] Zhen Fang, Jie Lu, Feng Liu, Junyu Xuan, and Guangquan Zhang. Open set domain adaptation: Theoretical bound and algorithm. IEEE transactions on neural networks and learning systems, 32(10):4309-4322, 2020. 3 +[22] Abolfazl Farahani, Sahar Voghoei, Khaled Rasheed, and Hamid R Arabnia. A brief review of domain adaptation. Advances in data science and information engineering: proceedings from ICDATA 2020 and IKE 2020, pages 877-894, 2021. 2 +[23] Kilian Fatras, Thibault Séjourne, Rémi Flamary, and Nicolas Courty. Unbalanced minibatch optimal transport; applications to domain adaptation. In International Conference on Machine Learning, pages 3186-3197. PMLR, 2021. 2 +[24] Jean Feydy, Thibault Séjourne, François-Xavier Vialard, Shun-ichi Amari, Alain Trouve, and Gabriel Peyre. Interpolating between optimal transport and mmd using sinkhorn divergences. In The 22nd International Conference on Artificial Intelligence and Statistics, pages 2681–2690. PMLR, 2019. 2 +[25] Rémi Flamary, Nicholas Courty, Davis Tuia, and Alain Rakotomamonjy. Optimal transport for domain adaptation. IEEE Trans. Pattern Anal. Mach. Intell, 1:1-40, 2016. 5 +[26] Rémi Flamary, Nicolas Courty, Alexandre Gramfort, Mokhtar Z Alaya, Aurélie Boisbunon, Stanislas Chambon, Laetitia Chapel, Adrien Corenflos, Kilian Fatras, Nemo Fournier, et al. Pot: Python optimal transport. The Journal of Machine Learning Research, 22(1):3571-3578, 2021. 2 +[27] Yves Grandvalet and Yoshua Bengio. Entropy regularization., 2006. 5 +[28] John A Hartigan and Manchek A Wong. Algorithm as 136: A k-means clustering algorithm. Journal of the royal sta + +tistical society. series c (applied statistics), 28(1):100-108, 1979. 5 +[29] Kaiming He, Xiangyu Zhang, Shaoqing Ren, and Jian Sun. Deep residual learning for image recognition. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 770-778, 2016. 6 +[30] Zi-Xian Huang and Chuan-Xian Ren. Rethinking correlation learning via label prior for open set domain adaptation. 6 +[31] J. J. Hull. A database for handwritten text recognition research. IEEE Transactions on Pattern Analysis & Machine Intelligence, 16(5):550-554, 2002. 6, 15 +[32] JoonHo Jang, Byeonghu Na, Dong Hyeok Shin, Mingi Ji, Kyungwoo Song, and Il-Chul Moon. Unknown-aware domain adversarial learning for open-set domain adaptation. Advances in Neural Information Processing Systems, 35: 16755-16767, 2022. 6 +[33] Jun-Gi Jang, Sooyeon Shim, Vladimir Egay, Jeeyong Lee, Jongmin Park, Suhyun Chae, and U Kang. Accurate open-set recognition for memory workload. ACM Transactions on Knowledge Discovery from Data, 17(9):1-14, 2023. 2 +[34] Jeff L Kennington and Richard V Helgason. Algorithms for network programming. John Wiley & Sons, Inc., 1980. 5 +[35] Diederik P Kingma and Jimmy Ba. Adam: A method for stochastic optimization. arXiv preprint arXiv:1412.6980, 2014. 6 +[36] Jogendra Nath Kundu, Naveen Venkat, R Venkatesh Babu, et al. Universal source-free domain adaptation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 4544-4553, 2020. 2 +[37] Jogendra Nath Kundu, Naveen Venkat, Ambareesh Revanur, R Venkatesh Babu, et al. Towards inheritable models for open-set domain adaptation. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 12376-12385, 2020. 6 +[38] Van Der Maaten Laurens and Geoffrey Hinton. Visualizing data using t-sne. Journal of Machine Learning Research, 9 (2605):2579-2605, 2008. 8 +[39] Khang Le, Huy Nguyen, Quang M Nguyen, Tung Pham, Hung Bui, and Nhat Ho. On robust optimal transport: Computational complexity and barycenter computation. Advances in Neural Information Processing Systems, 34:21947-21959, 2021. 2 +[40] Y. L. Lecun, Leon Bottou, Yoshua Bengio, and Patrick Haffner. Gradient-based learning applied to document recognition. proc. ii. Proceedings of the IEEE, 86(11):2278-2324, 1998. 6, 15 +[41] Na Lei and Xianfeng Gu. Computational Conformal Geometry. International Press and Higher Education Press, 2021. 4 +[42] Bo Li, Yifei Shen, Jingkang Yang, Yezhen Wang, Jiawei Ren, Tong Che, Jun Zhang, and Ziwei Liu. Sparse mixture-of-experts are domain generalizable learners. arXiv preprint arXiv:2206.04046, 2022. 2 +[43] Wuyang Li, Jie Liu, Bo Han, and Yixuan Yuan. Adjustment and alignment for unbiased open set domain adaptation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 24110-24119, 2023. 6 + +[44] Jian Liang, Dapeng Hu, and Jiashi Feng. Do we really need to access the source data? source hypothesis transfer for unsupervised domain adaptation. In International conference on machine learning, pages 6028-6039. PMLR, 2020. 3, 6 +[45] Mattia Litrico, Alessio Del Bue, and Pietro Morerio. Guiding pseudo-labels with uncertainty estimation for source-free unsupervised domain adaptation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 7640-7650, 2023. 5 +[46] Hong Liu, Zhangjie Cao, Mingsheng Long, Jianmin Wang, and Qiang Yang. Separate to adapt: Open set domain adaptation via progressive separation. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 2927-2936, 2019. 2, 6, 8, 15 +[47] Jie Liu, Xiaoqing Guo, and Yixuan Yuan. Unknown-oriented learning for open set domain adaptation. In European Conference on Computer Vision, pages 334-350. Springer, 2022. 2, 6 +[48] Mushui Liu, Yunlong Yu, Zhong Ji, Jungong Han, and Zhongfei Zhang. Tolerant self-distillation for image classification. Neural Networks, 174:106215, 2024. 1 +[49] Mushui Liu, Weijie He, Ziqian Lu, Jun Dan, Yunlong Yu, Yingming Li, Xi Li, and Jungong Han. Synth-clip: Synthetic data make clip generalize better in data-limited scenarios. Neural Networks, 184:107083, 2025. 2 +[50] Tianlin Liu, Joan Puigcerver, and Mathieu Blondel. Sparsity-constrained optimal transport. *ICLR*, 2023. 5 +[51] Weiming Liu, Xiaolin Zheng, Jiajie Su, Mengling Hu, Yanchao Tan, and Chaochao Chen. Exploiting variational domain-invariant user embedding for partially overlapped cross domain recommendation. In Proceedings of the 45th International ACM SIGIR conference on research and development in information retrieval, pages 312-321, 2022. 1 +[52] Weiming Liu, Xiaolin Zheng, Chaochao Chen, Jiajie Su, Xinting Liao, Mengling Hu, and Yanchao Tan. Joint internal multi-interest exploration and external domain alignment for cross domain sequential recommendation. In Proceedings of the ACM web conference 2023, pages 383-394, 2023. +[53] Weiming Liu, Chaochao Chen, Xinting Liao, Mengling Hu, Jiajie Su, Yanchao Tan, and Fan Wang. User distribution mapping modelling with collaborative filtering for cross-domain recommendation. In Proceedings of the ACM Web Conference 2024, pages 334-343, 2024. +[54] Weiming Liu, Chaochao Chen, Xinting Liao, Mengling Hu, Yanchao Tan, Fan Wang, Xiaolin Zheng, and Yew Soon Ong. Learning accurate and bidirectional transformation via dynamic embedding transportation for cross-domain recommendation. In Proceedings of the AAAI Conference on Artificial Intelligence, pages 8815-8823, 2024. 1 +[55] Weiming Liu, Xiaolin Zheng, Chaochao Chen, Jiahe Xu, Xinting Liao, Fan Wang, Yanchao Tan, and Yew-Soon Ong. Reducing item discrepancy via differentially private robust embedding alignment for privacy-preserving cross domain recommendation. In *Forty-first International Conference on Machine Learning*, 2024. 4 +[56] Ye Liu, Lingfeng Qiao, Changchong Lu, Di Yin, Chen Lin, Haoyuan Peng, and Bo Ren. Osman: A one-stage alignment + +network to unify multimodal alignment and unsupervised domain adaptation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 3551-3560, 2023. 1 +[57] Yang Liu, Zhipeng Zhou, and Baigui Sun. Cot: Unsupervised domain adaptation with clustering and optimal transport. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 19998-20007, 2023. 1 +[58] Mingsheng Long, Jianmin Wang, Guiguang Ding, Jiaguang Sun, and Philip S Yu. Transfer feature learning with joint distribution adaptation. In Proceedings of the IEEE international conference on computer vision, pages 2200-2207, 2013. 1 +[59] Mingsheng Long, Yue Cao, Jianmin Wang, and Michael Jordan. Learning transferable features with deep adaptation networks. In International Conference on Machine Learning, pages 97-105, 2015. 1, 2 +[60] Mingsheng Long, Han Zhu, Jianmin Wang, and Michael I Jordan. Unsupervised domain adaptation with residual transfer networks. Advances in neural information processing systems, 29, 2016. 1 +[61] Mingsheng Long, Han Zhu, Jianmin Wang, and Michael I Jordan. Deep transfer learning with joint adaptation networks. In Proceedings of the 34th International Conference on Machine Learning-Volume 70, pages 2208-2217. JMLR.org, 2017. 1, 2 +[62] Mingsheng Long, Zhangjie Cao, Jianmin Wang, and Michael I Jordan. Conditional adversarial domain adaptation. In Advances in Neural Information Processing Systems, pages 1640-1650, 2018. 2 +[63] Yadan Luo, Zijian Wang, Zi Huang, and Mahsa Baktashmotlagh. Progressive graph learning for open-set domain adaptation. In International Conference on Machine Learning, pages 6468-6478. PMLR, 2020. 6 +[64] Yadan Luo, Zijian Wang, Zhuoxiao Chen, Zi Huang, and Mahsa Baktashmotlagh. Source-free progressive graph learning for open-set domain adaptation. IEEE Transactions on Pattern Analysis and Machine Intelligence, 2023. 3, 6, 8 +[65] You-Wei Luo and Chuan-Xian Ren. Conditional bures metric for domain adaptation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 13989–13998, 2021. 1 +[66] You-Wei Luo and Chuan-Xian Ren. Mot: Masked optimal transport for partial domain adaptation. In 2023 IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 3531–3540. IEEE, 2023. 1 +[67] Yuval Netzer, Tao Wang, Adam Coates, Alessandro Bissacco, Bo Wu, and Andrew Y. Ng. Reading digits in natural images with unsupervised feature learning. Nips Workshop on Deep Learning & Unsupervised Feature Learning, 2011. 6, 15 +[68] Khai Nguyen, Dang Nguyen, Quoc Nguyen, Tung Pham, Hung Bui, Dinh Phung, Trung Le, and Nhat Ho. On transportation of mini-batches: A hierarchical approach. arXiv preprint arXiv:2102.05912, 2021. 2 + +[69] Pau Panareda Busto and Juergen Gall. Open set domain adaptation. In Proceedings of the IEEE international conference on computer vision, pages 754-763, 2017. 1, 2 +[70] Xingchao Peng, Ben Usman, Neela Kaushik, Dequan Wang, Judy Hoffman, and Kate Saenko. Visda: A synthetic-to-real benchmark for visual domain adaptation. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition Workshops, pages 2021-2026, 2018. 6, 15 +[71] Gabriel Peyré. Course notes on computational optimal transport. 2019. 4 +[72] Khiem Pham, Khang Le, Nhat Ho, Tung Pham, and Hung Bui. On unbalanced optimal transport: An analysis of sinkhorn algorithm. In International Conference on Machine Learning, pages 7673-7682. PMLR, 2020. 2 +[73] Sanqing Qu, Tianpei Zou, Florian Röhrbein, Cewu Lu, Guang Chen, Dacheng Tao, and Changjun Jiang. Upcycling models under domain and category shift. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 20019-20028, 2023. 2 +[74] Sanqing Qu, Tianpei Zou, Lianghua He, Florian Röhrbein, Alois Knoll, Guang Chen, and Changjun Jiang. Lead: Learning decomposition for source-free universal domain adaptation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 23334-23343, 2024. 2 +[75] Ievgen Redko, Amaury Habrard, and Marc Sebban. Theoretical analysis of domain adaptation with optimal transport. In Machine Learning and Knowledge Discovery in Databases: European Conference, ECML PKDD 2017, Skopje, Macedonia, September 18-22, 2017, Proceedings, Part II 10, pages 737-753. Springer, 2017. 5 +[76] Ievgen Redko, Nicolas Courty, Rémi Flamary, and Devis Tuia. Optimal transport for multi-source domain adaptation under target shift. In The 22nd International Conference on artificial intelligence and statistics, pages 849-858. PMLR, 2019. 2 +[77] Kate Saenko, Brian Kulis, Mario Fritz, and Trevor Darrell. Adapting visual category models to new domains. In European Conference on Computer Vision (ECCV), 2010. 6, 15 +[78] Kuniaki Saito, Shohei Yamamoto, Yoshitaka Ushiku, and Tatsuya Harada. Open set domain adaptation by backpropagation. In Proceedings of the European conference on computer vision (ECCV), pages 153-168, 2018. 1, 2, 6, 16 +[79] Thibault Séjourne, François-Xavier Vialard, and Gabriel Peyré. Faster unbalanced optimal transport: Translation invariant sinkhorn and 1-d frank-wolfe. In International Conference on Artificial Intelligence and Statistics, pages 4995-5021. PMLR, 2022. 13 +[80] Jian Shen, Yanru Qu, Weinan Zhang, and Yong Yu. Wasserstein distance guided representation learning for domain adaptation. arXiv preprint arXiv:1707.01217, 2017. 2 +[81] Rui Shu, Hung H Bui, Hirokazu Narui, and Stefano Ermon. A dirt-t approach to unsupervised domain adaptation. arXiv preprint arXiv:1802.08735, 2018. 2, 6 +[82] Baochen Sun and Kate Saenko. Deep coral: Correlation alignment for deep domain adaptation. In European Conference on Computer Vision, pages 443-450. Springer, 2016. 2 + +[83] Eric Tzeng, Judy Hoffman, Ning Zhang, Kate Saenko, and Trevor Darrell. Deep domain confusion: Maximizing for domain invariance. arXiv preprint arXiv:1412.3474, 2014. 2 +[84] Hemanth Venkateswara, Jose Eusebio, Shayok Chakraborty, and Sethuraman Panchanathan. Deep hashing network for unsupervised domain adaptation. CoRR, abs/1706.07522, 2017. 6, 15 +[85] Cédric Villani et al. Optimal transport: old and new. Springer, 2009. 2 +[86] Mei Wang and Weihong Deng. Deep visual domain adaptation: A survey. Neurocomputing, 312:135-153, 2018. 2 +[87] Qian Wang, Fanlin Meng, and Toby P Breckon. Progressively select and reject pseudo-labelled samples for open-set domain adaptation. arXiv preprint arXiv:2110.12635, 2021. 2 +[88] Garrett Wilson and Diane J Cook. A survey of unsupervised deep domain adaptation. ACM Transactions on Intelligent Systems and Technology (TIST), 11(5):1-46, 2020. 2 +[89] Zonghan Wu, Shirui Pan, Fengwen Chen, Guodong Long, Chengqi Zhang, and S Yu Philip. A comprehensive survey on graph neural networks. IEEE transactions on neural networks and learning systems, 32(1):4-24, 2020. 8 +[90] Yujia Xie, Xiangfeng Wang, Ruijia Wang, and Hongyuan Zha. A fast proximal point method for computing exact Wasserstein distance. In Uncertainty in artificial intelligence, pages 433-453. PMLR, 2020. 5 +[91] Renjun Xu, Pelen Liu, Yin Zhang, Fang Cai, Jindong Wang, Shuoying Liang, Heting Ying, and Jianwei Yin. Joint partial optimal transport for open set domain adaptation. In IJCAI, pages 2540-2546, 2020. 1, 2, 6 +[92] Tengyue Xu and Jun Dan. Ehm: Exploring dynamic alignment and hierarchical clustering in unsupervised domain adaptation via high-order moment-guided contrastive learning. Neural Networks, page 107188, 2025. 5 +[93] Hongliang Yan, Yukang Ding, Peihua Li, Qilong Wang, Yong Xu, and Wangmeng Zuo. Mind the class weight bias: Weighted maximum mean discrepancy for unsupervised domain adaptation. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition, pages 2272-2281, 2017. 2 +[94] Yu-Chu Yu and Hsuan-Tien Lin. Semi-supervised domain adaptation with source label adaptation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 24100-24109, 2023. 1 +[95] Zhiqi Yu, Jingjing Li, Zhekai Du, Lei Zhu, and Heng Tao Shen. A comprehensive survey on source-free domain adaptation. arXiv preprint arXiv:2302.11803, 2023. 3 +[96] Li Zhong, Zhen Fang, Feng Liu, Bo Yuan, Guangquan Zhang, and Jie Lu. Bridging the theoretical bound and deep algorithms for open set domain adaptation. IEEE transactions on neural networks and learning systems, 2021. 2, 3 +[97] Jingyu Zhuang, Ziliang Chen, Pengxu Wei, Guanbin Li, and Liang Lin. Discovering implicit classes achieves open set domain adaptation. In 2022 IEEE International Conference on Multimedia and Expo (ICME), pages 01-06. IEEE, 2022. 6 \ No newline at end of file diff --git a/2025/Distinguish Then Exploit_ Source-free Open Set Domain Adaptation via Weight Barcode Estimation and Sparse Label Assignment/images.zip b/2025/Distinguish Then Exploit_ Source-free Open Set Domain Adaptation via Weight Barcode Estimation and Sparse Label Assignment/images.zip new file mode 100644 index 0000000000000000000000000000000000000000..89ab7050df06edaa2e0a8a2672d57ff18896ccdf --- /dev/null +++ b/2025/Distinguish Then Exploit_ Source-free Open Set Domain Adaptation via Weight Barcode Estimation and Sparse Label Assignment/images.zip @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:50ef2ee288d8352782d55a84c2dac35ceb368ef6300aa9b113f4797956347ceb +size 808379 diff --git a/2025/Distinguish Then Exploit_ Source-free Open Set Domain Adaptation via Weight Barcode Estimation and Sparse Label Assignment/layout.json b/2025/Distinguish Then Exploit_ Source-free Open Set Domain Adaptation via Weight Barcode Estimation and Sparse Label Assignment/layout.json new file mode 100644 index 0000000000000000000000000000000000000000..035631475ce76999af30e76a83c1ee6472fdce23 --- /dev/null +++ b/2025/Distinguish Then Exploit_ Source-free Open Set Domain Adaptation via Weight Barcode Estimation and Sparse Label Assignment/layout.json @@ -0,0 +1,12811 @@ +{ + "pdf_info": [ + { + "para_blocks": [ + { + "bbox": [ + 52, + 102, + 541, + 138 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 52, + 102, + 541, + 138 + ], + "spans": [ + { + "bbox": [ + 52, + 102, + 541, + 138 + ], + "type": "text", + "content": "Distinguish Then Exploit: Source-free Open Set Domain Adaptation via Weight Barcode Estimation and Sparse Label Assignment" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 139, + 159, + 477, + 189 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 139, + 159, + 477, + 189 + ], + "spans": [ + { + "bbox": [ + 139, + 159, + 477, + 189 + ], + "type": "text", + "content": "Weiming Liu" + }, + { + "bbox": [ + 139, + 159, + 477, + 189 + ], + "type": "inline_equation", + "content": "^{1*}" + }, + { + "bbox": [ + 139, + 159, + 477, + 189 + ], + "type": "text", + "content": ", Jun Dan" + }, + { + "bbox": [ + 139, + 159, + 477, + 189 + ], + "type": "inline_equation", + "content": "^{2*}" + }, + { + "bbox": [ + 139, + 159, + 477, + 189 + ], + "type": "text", + "content": ", Fan Wang" + }, + { + "bbox": [ + 139, + 159, + 477, + 189 + ], + "type": "inline_equation", + "content": "^{2}" + }, + { + "bbox": [ + 139, + 159, + 477, + 189 + ], + "type": "text", + "content": ", Xinting Liao" + }, + { + "bbox": [ + 139, + 159, + 477, + 189 + ], + "type": "inline_equation", + "content": "^{2}" + }, + { + "bbox": [ + 139, + 159, + 477, + 189 + ], + "type": "text", + "content": ", Junhao Dong" + }, + { + "bbox": [ + 139, + 159, + 477, + 189 + ], + "type": "inline_equation", + "content": "^{3}" + }, + { + "bbox": [ + 139, + 159, + 477, + 189 + ], + "type": "text", + "content": ", Hua Yu" + }, + { + "bbox": [ + 139, + 159, + 477, + 189 + ], + "type": "inline_equation", + "content": "^{3}" + }, + { + "bbox": [ + 139, + 159, + 477, + 189 + ], + "type": "text", + "content": ", Shunjie Dong" + }, + { + "bbox": [ + 139, + 159, + 477, + 189 + ], + "type": "inline_equation", + "content": "^{4}" + }, + { + "bbox": [ + 139, + 159, + 477, + 189 + ], + "type": "text", + "content": ", Lianyong Qi" + }, + { + "bbox": [ + 139, + 159, + 477, + 189 + ], + "type": "inline_equation", + "content": "^{56\\dagger}" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 54, + 190, + 564, + 217 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 54, + 190, + 564, + 217 + ], + "spans": [ + { + "bbox": [ + 54, + 190, + 564, + 217 + ], + "type": "text", + "content": "1Bytedance Inc. 2Zhejiang University 3Nanyang Technological University 4Shanghai Jiao Tong University 5China University of Petroleum (East China)" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 136, + 217, + 482, + 231 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 136, + 217, + 482, + 231 + ], + "spans": [ + { + "bbox": [ + 136, + 217, + 482, + 231 + ], + "type": "inline_equation", + "content": "^{6}" + }, + { + "bbox": [ + 136, + 217, + 482, + 231 + ], + "type": "text", + "content": "Shandong Key Laboratory of Intelligent Oil & Gas Industrial Software" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 84, + 233, + 533, + 258 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 84, + 233, + 533, + 258 + ], + "spans": [ + { + "bbox": [ + 84, + 233, + 533, + 258 + ], + "type": "text", + "content": "lwming95@gmail.com, danjun@zju.edu.cn, fanwang97@zju.edu.cn, xintingliao@zju.edu.cn, junhao003@ntu.edu.sg, yu_hua@ntu.edu.sg, sjdong@sjtu.edu.cn, lianyongqi@gmail.com" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 143, + 286, + 192, + 300 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 143, + 286, + 192, + 300 + ], + "spans": [ + { + "bbox": [ + 143, + 286, + 192, + 300 + ], + "type": "text", + "content": "Abstract" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 46, + 312, + 289, + 622 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 312, + 289, + 622 + ], + "spans": [ + { + "bbox": [ + 46, + 312, + 289, + 622 + ], + "type": "text", + "content": "Nowadays, domain adaptation techniques have been widely investigated for knowledge sharing from labeled source domain to unlabeled target domain. However, target domain may include some data samples that belong to unknown categories in real-world scenarios. Moreover, the target domain cannot access the source data samples due to privacy-preserving restrictions. In this paper, we focus on the source-free open set domain adaptation problem which includes two main challenges, i.e., how to distinguish known and unknown target samples and how to exploit useful source information to provide trustworthy pseudo labels for known target samples. Existing approaches that directly apply conventional domain alignment methods could lead to sample mismatch and misclassification in this scenario. To overcome these issues, we propose a Distinguish Then Exploit model (DTE) with two components, i.e., weight barcode estimation and sparse label assignment. Weight barcode estimation first calculates the marginal probability of target samples via partially unbalanced optimal transport, then quantize barcode results to distinguish unknown target samples. Sparse label assignment utilizes sparse sample-label matching via proximal term to fully exploit useful source information. Our empirically study on several datasets shows that DTE outperforms the state-of-the-art models on tackling the source-free open set domain adaptation problem." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 47, + 639, + 128, + 651 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 639, + 128, + 651 + ], + "spans": [ + { + "bbox": [ + 47, + 639, + 128, + 651 + ], + "type": "text", + "content": "1. Introduction" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 46, + 660, + 287, + 696 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 660, + 287, + 696 + ], + "spans": [ + { + "bbox": [ + 46, + 660, + 287, + 696 + ], + "type": "text", + "content": "Nowadays, deep learning technique has been successfully adopted to image classification tasks with the help of large-scale labeled datasets [16, 19, 48, 51-54, 57, 65, 66]. How" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 306, + 288, + 545, + 396 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 288, + 545, + 396 + ], + "spans": [ + { + "bbox": [ + 306, + 288, + 545, + 396 + ], + "type": "text", + "content": "ever, it is time-consuming to sufficiently annotate data for providing satisfied results. A promising solution is to leverage useful information from the labeled source domain to assign reliable labels within the unlabeled target domain [18, 58, 60]. Nevertheless, the data distribution could be significantly different across domains, leading to the model degradation on final performance [12, 56, 94]. How to reduce the domain bias and discrepancy while avoiding negative transfer across domains has become an urgent problem." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 304, + 399, + 546, + 555 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 399, + 546, + 555 + ], + "spans": [ + { + "bbox": [ + 304, + 399, + 546, + 555 + ], + "type": "text", + "content": "In this paper, we focus on a more general problem of source-free open set domain adaptation. That is, there could exist some target samples belong to extra categories that does not present in the source domain. Moreover, due to data privacy protection constraints, the target domain could only have access to the well-trained model in the source domain. Such settings are commonly encountered in real-world scenarios and it is rather complicated than conventional domain adaptation problem. Specifically, there are two main challenges: CH1: How can we enhance the ability to distinguish target samples that belong to unknown categories? CH2: How to better exploit appropriate pseudo labels for the remaining known target samples?" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 304, + 557, + 547, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 557, + 547, + 714 + ], + "spans": [ + { + "bbox": [ + 304, + 557, + 547, + 714 + ], + "type": "text", + "content": "Existing researches cannot solve source-free open set domain adaptation well. Firstly, conventional transfer learning always assumes that both source and target domains share the same set of label categories [15, 59, 61]. Directly aligning the latent distributions of source and target domains may lead to negative transfer, where target unknown samples are incorrectly matched to source samples. Secondly, traditional open set domain adaptation approaches should incorporate well-tuned thresholds for determining the decision boundary between known and unknown target samples [69]. This suggests that many of these methods (e.g., OSBP [78] and JPOT [91]) typically rely on prior knowledge about the target domain, which will limit their potential applicabil" + } + ] + } + ], + "index": 13 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "spans": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "type": "text", + "content": "CVF" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 145, + 0, + 496, + 35 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 145, + 0, + 496, + 35 + ], + "spans": [ + { + "bbox": [ + 145, + 0, + 496, + 35 + ], + "type": "text", + "content": "This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 61, + 702, + 215, + 712 + ], + "type": "page_footnote", + "angle": 0, + "lines": [ + { + "bbox": [ + 61, + 702, + 215, + 712 + ], + "spans": [ + { + "bbox": [ + 61, + 702, + 215, + 712 + ], + "type": "text", + "content": "* Equal Contribution, † Corresponding Author." + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "text", + "content": "4927" + } + ] + } + ], + "index": 15 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 0 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 58, + 71, + 147, + 165 + ], + "blocks": [ + { + "bbox": [ + 58, + 71, + 147, + 165 + ], + "lines": [ + { + "bbox": [ + 58, + 71, + 147, + 165 + ], + "spans": [ + { + "bbox": [ + 58, + 71, + 147, + 165 + ], + "type": "image", + "image_path": "367ba117f828b9bdb3225d33197ee7e634f0001c025e199ac7ddb24ff3292821.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 69, + 167, + 138, + 174 + ], + "lines": [ + { + "bbox": [ + 69, + 167, + 138, + 174 + ], + "spans": [ + { + "bbox": [ + 69, + 167, + 138, + 174 + ], + "type": "text", + "content": "(a) Initial Target Features in" + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 87, + 175, + 121, + 182 + ], + "lines": [ + { + "bbox": [ + 87, + 175, + 121, + 182 + ], + "spans": [ + { + "bbox": [ + 87, + 175, + 121, + 182 + ], + "type": "text", + "content": "Latent Space" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "type": "image", + "bbox": [ + 149, + 72, + 313, + 165 + ], + "blocks": [ + { + "bbox": [ + 149, + 72, + 313, + 165 + ], + "lines": [ + { + "bbox": [ + 149, + 72, + 313, + 165 + ], + "spans": [ + { + "bbox": [ + 149, + 72, + 313, + 165 + ], + "type": "image", + "image_path": "c40dfff88a22a4c7a9315ee6922b8d5e279b901d4dae117c13c436d61b425e35.jpg" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 171, + 167, + 294, + 182 + ], + "lines": [ + { + "bbox": [ + 171, + 167, + 294, + 182 + ], + "spans": [ + { + "bbox": [ + 171, + 167, + 294, + 182 + ], + "type": "text", + "content": "(b) Adopting Weight Barcode Estimation to Distinguish Unknown Samples in Target Domain" + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_caption" + } + ], + "index": 3 + }, + { + "type": "image", + "bbox": [ + 313, + 72, + 445, + 165 + ], + "blocks": [ + { + "bbox": [ + 313, + 72, + 445, + 165 + ], + "lines": [ + { + "bbox": [ + 313, + 72, + 445, + 165 + ], + "spans": [ + { + "bbox": [ + 313, + 72, + 445, + 165 + ], + "type": "image", + "image_path": "6a9845cf78cf0bafd758196053ee075f3662a26df15d935f008dcc09f5405bad.jpg" + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "image_body" + } + ], + "index": 5 + }, + { + "type": "image", + "bbox": [ + 446, + 72, + 536, + 166 + ], + "blocks": [ + { + "bbox": [ + 317, + 167, + 445, + 182 + ], + "lines": [ + { + "bbox": [ + 317, + 167, + 445, + 182 + ], + "spans": [ + { + "bbox": [ + 317, + 167, + 445, + 182 + ], + "type": "text", + "content": "(c) Adopting Sparse Label Assignment to Exploit Pseudo Label for Known Samples in Target Domain" + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 446, + 72, + 536, + 166 + ], + "lines": [ + { + "bbox": [ + 446, + 72, + 536, + 166 + ], + "spans": [ + { + "bbox": [ + 446, + 72, + 536, + 166 + ], + "type": "image", + "image_path": "76b64f5d902db14e71f944365c9f7a4ca5aa89930c70966f7291d23704118d73.jpg" + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 453, + 167, + 526, + 182 + ], + "lines": [ + { + "bbox": [ + 453, + 167, + 526, + 182 + ], + "spans": [ + { + "bbox": [ + 453, + 167, + 526, + 182 + ], + "type": "text", + "content": "(d) Target Features in Latent Space after Training" + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 46, + 186, + 547, + 232 + ], + "lines": [ + { + "bbox": [ + 46, + 186, + 547, + 232 + ], + "spans": [ + { + "bbox": [ + 46, + 186, + 547, + 232 + ], + "type": "text", + "content": "Figure 1. An overview of the proposed DTE. (a) The initial target features provided in the latent space are dispersed and challenging to achieve satisfactory classification results. (b) We first adopt weight barcode estimation component to distinguish known or unknown target samples accordingly. (c) Then we adopt sparse label assignment to exploit pseudo label for known samples in the target domains. (d) Finally we can better cluster and classify target unlabeled samples. Best viewed in color." + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "image_caption" + } + ], + "index": 7 + }, + { + "bbox": [ + 46, + 237, + 290, + 392 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 237, + 290, + 392 + ], + "spans": [ + { + "bbox": [ + 46, + 237, + 290, + 392 + ], + "type": "text", + "content": "ity. If such prior knowledge (e.g., the proportion of unknown classes in the target domain) cannot be well estimated, the model performance could be degraded dramatically. Some recent works (e.g., STA [46]) further utilized coarse-to-fine weighting mechanism for separating target samples under different proportion of unknown classes in the target domain (also termed as openness [46]). However, they rely on involving domain adversarial training which could be unstable under some circumstances [14, 81]. Last but not least, we cannot access to the origin source data samples during the training procedure. Previous open set models fail to address such constraints for data privacy protection and they cannot be directly applied into source-free scenario." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 46, + 397, + 290, + 662 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 397, + 290, + 662 + ], + "spans": [ + { + "bbox": [ + 46, + 397, + 290, + 662 + ], + "type": "text", + "content": "To address the aforementioned issues, we propose Distinguish Then Exploit (DTE) for tackling source-free open set domain adaptation problem. To better distinguish unknown target samples (solving CH1) and assign appropriate labels to the remaining target samples (solving CH2), we utilize two main components in DTE, i.e., weight barcode estimation component and sparse label assignment component. Weight barcode estimation component aims to solve CH1 by adopting marginal probability estimation and weight barcode establishment to automatically determine unknown target samples without prior knowledge. Sparse label assignment component is designed to address CH2 via leveraging the rest known target samples with proximal optimal transport techniques. By cooperating these two components, we can better utilize source knowledge in the target domain by providing more accurate results on source-free open set domain adaptation as shown in Fig.1. We first distinguish the heart-shaped and diamond-shaped unknown target samples. Then we assign the proper pseudo labels for the known target samples with the shapes of triangle, square, and circle respectively. Finally we can classify target unlabeled samples to reach better performance." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 47, + 665, + 288, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 665, + 288, + 714 + ], + "spans": [ + { + "bbox": [ + 47, + 665, + 288, + 714 + ], + "type": "text", + "content": "We summarize our main contributions as follows: (1) We propose a novel framework, i.e., DTE, for source-free open set domain adaptation problem via knowledge transfer across domains. (2) We innovatively propose weight bar" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 304, + 237, + 547, + 333 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 237, + 547, + 333 + ], + "spans": [ + { + "bbox": [ + 304, + 237, + 547, + 333 + ], + "type": "text", + "content": "code estimation with partial unbalanced optimal transport to distinguish unknown target samples. Furthermore, we propose target sample reweighting optimization method to figure out the marginal probabilities on partial unbalanced optimal transport. Moreover, we further propose sparse label assignment to provide reliable labels for the known target samples. (3) Extensive empirical studies on various datasets show that DTE achieves the state-of-the-art performance." + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 306, + 345, + 390, + 357 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 345, + 390, + 357 + ], + "spans": [ + { + "bbox": [ + 306, + 345, + 390, + 357 + ], + "type": "text", + "content": "2. Related work" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 304, + 365, + 546, + 570 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 365, + 546, + 570 + ], + "spans": [ + { + "bbox": [ + 304, + 365, + 546, + 570 + ], + "type": "text", + "content": "Closed Set Domain Adaptation. Conventional domain adaptation usually assumes that both source and target domains share the same set of labels which namely Closed Set Domain Adaptation [59, 61, 62]. There are mainly two types of existing works, i.e., discrepancy-based method and adversarial-based methods [17, 22, 86, 88]. Discrepancy-based methods usually align the statistics information (e.g., mean and covariance) across domains via Maximum Mean Discrepancy (MMD [83, 93]) or Correlation Alignment (CORAL [82]). More recently, optimal transport techniques [4, 7, 13, 26, 39, 72, 85] have been widely adopted into solving domain adaptation problem [23, 24, 68, 76, 80]. Adversarial-based methods, e.g., DANN [2], CDAN [62] and WDGRL [80] which usually involve domain discriminators during training procedure. However, these approaches could easily lead to negative transfer by mismatching the unknown target samples with the source samples." + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 304, + 570, + 547, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 570, + 547, + 714 + ], + "spans": [ + { + "bbox": [ + 304, + 570, + 547, + 714 + ], + "type": "text", + "content": "Open Set Domain Adaptation. Different from closed set domain adaptation, open set domain adaptation considers more general scenarios when some target samples may belong to unknown label categories [8, 33, 36, 42, 49, 69, 73, 74, 87, 96]. Most of classic open set domain adaptation models (e.g., OSBP [78] and JPOT [91]) set thresholds to determine known and unknown target samples via useful prior knowledge. More recently, some researches (e.g., STA [46] and UOL [47]) also utilize semantic intracategory information for excavating unknown samples with the absent of relevant knowledge. Meanwhile, with the increasing emphasis on protecting data privacy, target do" + } + ] + } + ], + "index": 16 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "text", + "content": "4928" + } + ] + } + ], + "index": 17 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 1 + }, + { + "para_blocks": [ + { + "bbox": [ + 46, + 72, + 289, + 145 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 72, + 289, + 145 + ], + "spans": [ + { + "bbox": [ + 46, + 72, + 289, + 145 + ], + "type": "text", + "content": "main could be relatively difficult to directly obtain source data [20, 95]. However, only a few approaches have been proposed to tackle source-free open set domain adaptation (e.g., SFPGL [64]) since it is more challenging. How to provide satisfied results on open set domain adaptation under the privacy preserving constraints is an urgent problem." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 47, + 155, + 157, + 168 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 155, + 157, + 168 + ], + "spans": [ + { + "bbox": [ + 47, + 155, + 157, + 168 + ], + "type": "text", + "content": "3. Modelling for DTE" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 46, + 175, + 287, + 295 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 175, + 287, + 295 + ], + "spans": [ + { + "bbox": [ + 46, + 175, + 287, + 295 + ], + "type": "text", + "content": "First, we describe notations. We assume there are two domains, i.e., a source domain " + }, + { + "bbox": [ + 46, + 175, + 287, + 295 + ], + "type": "inline_equation", + "content": "\\mathcal{S} = \\{(x_i^s,y_i^s)\\}_{i = 1}^{n_s}" + }, + { + "bbox": [ + 46, + 175, + 287, + 295 + ], + "type": "text", + "content": " with " + }, + { + "bbox": [ + 46, + 175, + 287, + 295 + ], + "type": "inline_equation", + "content": "n_{s}" + }, + { + "bbox": [ + 46, + 175, + 287, + 295 + ], + "type": "text", + "content": " labeled samples and a target domain " + }, + { + "bbox": [ + 46, + 175, + 287, + 295 + ], + "type": "inline_equation", + "content": "\\mathcal{T} = \\{(x_i^t)\\}_{i = 1}^{n_t}" + }, + { + "bbox": [ + 46, + 175, + 287, + 295 + ], + "type": "text", + "content": " with " + }, + { + "bbox": [ + 46, + 175, + 287, + 295 + ], + "type": "inline_equation", + "content": "n_t" + }, + { + "bbox": [ + 46, + 175, + 287, + 295 + ], + "type": "text", + "content": " unlabeled samples. " + }, + { + "bbox": [ + 46, + 175, + 287, + 295 + ], + "type": "inline_equation", + "content": "\\mathcal{C}^s" + }, + { + "bbox": [ + 46, + 175, + 287, + 295 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 46, + 175, + 287, + 295 + ], + "type": "inline_equation", + "content": "\\mathcal{C}^t" + }, + { + "bbox": [ + 46, + 175, + 287, + 295 + ], + "type": "text", + "content": " denote the label spaces in source and target domains respectively. Unlike traditional domain adaptation tasks, there exist target samples that belong to additional unknown classes " + }, + { + "bbox": [ + 46, + 175, + 287, + 295 + ], + "type": "inline_equation", + "content": "\\mathcal{C}_{\\mathrm{unk}}^t" + }, + { + "bbox": [ + 46, + 175, + 287, + 295 + ], + "type": "text", + "content": " in the target domain. Specifically, the label space in target domain can be formulated as " + }, + { + "bbox": [ + 46, + 175, + 287, + 295 + ], + "type": "inline_equation", + "content": "\\mathcal{C}^t = \\mathcal{C}^s\\cup \\mathcal{C}_{\\mathrm{unk}}^t" + }, + { + "bbox": [ + 46, + 175, + 287, + 295 + ], + "type": "text", + "content": ". We aim to correctly assign labels for the target samples via knowledge transferring." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 46, + 295, + 288, + 403 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 295, + 288, + 403 + ], + "spans": [ + { + "bbox": [ + 46, + 295, + 288, + 403 + ], + "type": "text", + "content": "Since the increasing law on data privacy protection, the source data samples cannot be delivered to the target domain. Hence, the target domain can only access the well-trained source model for tackling the source-free openset domain adaptation. We first adopt CNN architecture in source feature extractor " + }, + { + "bbox": [ + 46, + 295, + 288, + 403 + ], + "type": "inline_equation", + "content": "\\mathcal{F}_s(\\cdot)" + }, + { + "bbox": [ + 46, + 295, + 288, + 403 + ], + "type": "text", + "content": " to pretrain on the source data samples to obtain the source features " + }, + { + "bbox": [ + 46, + 295, + 288, + 403 + ], + "type": "inline_equation", + "content": "\\mathcal{F}_s(\\boldsymbol{x}^s) = \\boldsymbol{z}^s" + }, + { + "bbox": [ + 46, + 295, + 288, + 403 + ], + "type": "text", + "content": " following [44]. After that we adopt source classifier " + }, + { + "bbox": [ + 46, + 295, + 288, + 403 + ], + "type": "inline_equation", + "content": "\\mathcal{G}(\\cdot)" + }, + { + "bbox": [ + 46, + 295, + 288, + 403 + ], + "type": "text", + "content": " to make output prediction " + }, + { + "bbox": [ + 46, + 295, + 288, + 403 + ], + "type": "inline_equation", + "content": "\\pmb{p}^s (y_k|\\pmb{x}_i^s)" + }, + { + "bbox": [ + 46, + 295, + 288, + 403 + ], + "type": "text", + "content": ":" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 79, + 403, + 287, + 435 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 79, + 403, + 287, + 435 + ], + "spans": [ + { + "bbox": [ + 79, + 403, + 287, + 435 + ], + "type": "interline_equation", + "content": "\\left[ \\mathcal {G} \\left(\\boldsymbol {x} _ {i} ^ {s}\\right) \\right] _ {k} = \\boldsymbol {p} ^ {s} \\left(y _ {k} \\mid \\boldsymbol {x} _ {i} ^ {s}\\right) = \\frac {e ^ {\\boldsymbol {w} _ {k} ^ {\\top} \\boldsymbol {z} _ {i} ^ {s} + b _ {k}}}{\\sum_ {l = 1} ^ {H} e ^ {\\boldsymbol {w} _ {l} ^ {\\top} \\boldsymbol {z} _ {i} ^ {s} + b _ {l}}}, \\tag {1}", + "image_path": "c5988205a22f749597ad584a04f42ffc8482783b64aef4c8eccf8ae200ed9d88.jpg" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 46, + 435, + 289, + 628 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 435, + 289, + 628 + ], + "spans": [ + { + "bbox": [ + 46, + 435, + 289, + 628 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 46, + 435, + 289, + 628 + ], + "type": "inline_equation", + "content": "\\pmb{w}_k^\\top" + }, + { + "bbox": [ + 46, + 435, + 289, + 628 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 46, + 435, + 289, + 628 + ], + "type": "inline_equation", + "content": "b_{k}" + }, + { + "bbox": [ + 46, + 435, + 289, + 628 + ], + "type": "text", + "content": " denote the " + }, + { + "bbox": [ + 46, + 435, + 289, + 628 + ], + "type": "inline_equation", + "content": "k" + }, + { + "bbox": [ + 46, + 435, + 289, + 628 + ], + "type": "text", + "content": "-th class weight and bias in the source classifier " + }, + { + "bbox": [ + 46, + 435, + 289, + 628 + ], + "type": "inline_equation", + "content": "\\mathcal{G}(\\cdot)" + }, + { + "bbox": [ + 46, + 435, + 289, + 628 + ], + "type": "text", + "content": ". Then we adopt cross-entropy loss to train on the source samples following [44]. Finally, we send the well-trained source model including " + }, + { + "bbox": [ + 46, + 435, + 289, + 628 + ], + "type": "inline_equation", + "content": "\\mathcal{F}_s(\\cdot)" + }, + { + "bbox": [ + 46, + 435, + 289, + 628 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 46, + 435, + 289, + 628 + ], + "type": "inline_equation", + "content": "\\mathcal{G}(\\cdot)" + }, + { + "bbox": [ + 46, + 435, + 289, + 628 + ], + "type": "text", + "content": " to the unlabeled target domain as shown in Fig. 2. We then introduce the main components of our proposed DTE, as illustrated in Fig. 2. DTE mainly has two components, i.e., weight barcode estimation and sparse label assignment. Weight barcode estimation aims to first distinguish the target samples that belong to the unknown class categories. Sparse label assignment aims to further exploit the rest target samples by providing them with proper pseudo labels. By incorporating weight barcode estimation with sparse label assignment sequentially, we can achieve more precise results on source-free openset domain adaptation. We will introduce the model details in the following subsections." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 47, + 635, + 200, + 647 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 635, + 200, + 647 + ], + "spans": [ + { + "bbox": [ + 47, + 635, + 200, + 647 + ], + "type": "text", + "content": "3.1. Weight Barcode Estimation" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 46, + 653, + 288, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 653, + 288, + 715 + ], + "spans": [ + { + "bbox": [ + 46, + 653, + 288, + 715 + ], + "type": "text", + "content": "In this section, we will first introduce the details of weight barcode estimation. Since target domain include many samples which belong to the unknown classes, it is unreasonable to directly apply traditional alignment techniques in source-free domain adaptation. Specifically, the unknown" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 306, + 72, + 525, + 84 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 72, + 525, + 84 + ], + "spans": [ + { + "bbox": [ + 306, + 72, + 525, + 84 + ], + "type": "text", + "content": "Algorithm 1 The training procedure of proposed DTE" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 315, + 86, + 545, + 110 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 315, + 86, + 545, + 110 + ], + "spans": [ + { + "bbox": [ + 315, + 86, + 545, + 110 + ], + "type": "text", + "content": "Input: " + }, + { + "bbox": [ + 315, + 86, + 545, + 110 + ], + "type": "inline_equation", + "content": "\\mathcal{F}_s(\\cdot)" + }, + { + "bbox": [ + 315, + 86, + 545, + 110 + ], + "type": "text", + "content": ": well-trained source feature extractor; " + }, + { + "bbox": [ + 315, + 86, + 545, + 110 + ], + "type": "inline_equation", + "content": "\\mathcal{G}(\\cdot)" + }, + { + "bbox": [ + 315, + 86, + 545, + 110 + ], + "type": "text", + "content": ": well-trained classifier; " + }, + { + "bbox": [ + 315, + 86, + 545, + 110 + ], + "type": "inline_equation", + "content": "\\epsilon, \\alpha" + }, + { + "bbox": [ + 315, + 86, + 545, + 110 + ], + "type": "text", + "content": ": Hyper parameters." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 316, + 110, + 521, + 122 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 110, + 521, + 122 + ], + "spans": [ + { + "bbox": [ + 316, + 110, + 521, + 122 + ], + "type": "text", + "content": "Set class prototypes " + }, + { + "bbox": [ + 316, + 110, + 521, + 122 + ], + "type": "inline_equation", + "content": "\\pmb{\\mu}^{(0)} = \\pmb{w}" + }, + { + "bbox": [ + 316, + 110, + 521, + 122 + ], + "type": "text", + "content": " as the initialization." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 316, + 122, + 410, + 133 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 122, + 410, + 133 + ], + "spans": [ + { + "bbox": [ + 316, + 122, + 410, + 133 + ], + "type": "text", + "content": "for epoch " + }, + { + "bbox": [ + 316, + 122, + 410, + 133 + ], + "type": "inline_equation", + "content": "= 1" + }, + { + "bbox": [ + 316, + 122, + 410, + 133 + ], + "type": "text", + "content": " to " + }, + { + "bbox": [ + 316, + 122, + 410, + 133 + ], + "type": "inline_equation", + "content": "L" + }, + { + "bbox": [ + 316, + 122, + 410, + 133 + ], + "type": "text", + "content": " do" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 326, + 133, + 545, + 159 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 326, + 133, + 545, + 159 + ], + "spans": [ + { + "bbox": [ + 326, + 133, + 545, + 159 + ], + "type": "text", + "content": "Obtain the optimal solution on " + }, + { + "bbox": [ + 326, + 133, + 545, + 159 + ], + "type": "inline_equation", + "content": "\\pmb{f}^{*}" + }, + { + "bbox": [ + 326, + 133, + 545, + 159 + ], + "type": "text", + "content": " on PUOT for finding marginal probability " + }, + { + "bbox": [ + 326, + 133, + 545, + 159 + ], + "type": "inline_equation", + "content": "q_{\\tau}(\\pmb{x}_i^t) = a_i \\exp(-f_i^* / \\tau)" + }, + { + "bbox": [ + 326, + 133, + 545, + 159 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 326, + 159, + 545, + 183 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 326, + 159, + 545, + 183 + ], + "spans": [ + { + "bbox": [ + 326, + 159, + 545, + 183 + ], + "type": "text", + "content": "Obtain the weight barcode matrix " + }, + { + "bbox": [ + 326, + 159, + 545, + 183 + ], + "type": "inline_equation", + "content": "\\widehat{\\pmb{Q}}" + }, + { + "bbox": [ + 326, + 159, + 545, + 183 + ], + "type": "text", + "content": " and determine known/unknown target samples " + }, + { + "bbox": [ + 326, + 159, + 545, + 183 + ], + "type": "inline_equation", + "content": "\\pmb{x}^t = \\pmb{x}_{\\mathrm{unk}}^t\\cup \\pmb{x}_{\\mathrm{kno}}^t" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 326, + 183, + 525, + 195 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 326, + 183, + 525, + 195 + ], + "spans": [ + { + "bbox": [ + 326, + 183, + 525, + 195 + ], + "type": "text", + "content": "Obtain the optimal solution " + }, + { + "bbox": [ + 326, + 183, + 525, + 195 + ], + "type": "inline_equation", + "content": "\\gamma^{*}" + }, + { + "bbox": [ + 326, + 183, + 525, + 195 + ], + "type": "text", + "content": " on solving Eq.(6)." + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 326, + 195, + 545, + 219 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 326, + 195, + 545, + 219 + ], + "spans": [ + { + "bbox": [ + 326, + 195, + 545, + 219 + ], + "type": "text", + "content": "Gather the network output prediction " + }, + { + "bbox": [ + 326, + 195, + 545, + 219 + ], + "type": "inline_equation", + "content": "\\pmb{p}^t" + }, + { + "bbox": [ + 326, + 195, + 545, + 219 + ], + "type": "text", + "content": " and the corresponding \"ground-truth\" " + }, + { + "bbox": [ + 326, + 195, + 545, + 219 + ], + "type": "inline_equation", + "content": "\\pmb{q}^t" + }, + { + "bbox": [ + 326, + 195, + 545, + 219 + ], + "type": "text", + "content": " on Eq.(9)-Eq.(10)." + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 326, + 219, + 535, + 231 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 326, + 219, + 535, + 231 + ], + "spans": [ + { + "bbox": [ + 326, + 219, + 535, + 231 + ], + "type": "text", + "content": "Minimize loss function " + }, + { + "bbox": [ + 326, + 219, + 535, + 231 + ], + "type": "inline_equation", + "content": "L_{\\mathrm{cls}}" + }, + { + "bbox": [ + 326, + 219, + 535, + 231 + ], + "type": "text", + "content": " on Eq.(11) for training." + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 326, + 231, + 484, + 243 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 326, + 231, + 484, + 243 + ], + "spans": [ + { + "bbox": [ + 326, + 231, + 484, + 243 + ], + "type": "text", + "content": "Update class prototypes " + }, + { + "bbox": [ + 326, + 231, + 484, + 243 + ], + "type": "inline_equation", + "content": "\\mu" + }, + { + "bbox": [ + 326, + 231, + 484, + 243 + ], + "type": "text", + "content": " via Eq.(12)." + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 316, + 243, + 350, + 253 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 243, + 350, + 253 + ], + "spans": [ + { + "bbox": [ + 316, + 243, + 350, + 253 + ], + "type": "text", + "content": "end for" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 316, + 255, + 518, + 266 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 255, + 518, + 266 + ], + "spans": [ + { + "bbox": [ + 316, + 255, + 518, + 266 + ], + "type": "text", + "content": "Return: The well-trained model in target domain." + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 304, + 277, + 545, + 410 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 277, + 545, + 410 + ], + "spans": [ + { + "bbox": [ + 304, + 277, + 545, + 410 + ], + "type": "text", + "content": "target samples may get matched to the known class categories and causing negative transfer phenomenon [21, 96]. Therefore, it is essential to distinguish the unknown target samples and filter them out during the first stage. To fulfill this challenging task, we propose weight barcode estimation which includes two steps, i.e., marginal probability estimation and weight barcode establishment. Marginal probability estimation adopts Partial Unbalanced Optimal Transport (PUOT) technique to reweight target samples accordingly. Then we utilize weight barcode establishment to automatically determine the unknown target samples." + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 304, + 411, + 546, + 519 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 411, + 546, + 519 + ], + "spans": [ + { + "bbox": [ + 304, + 411, + 546, + 519 + ], + "type": "text", + "content": "Marginal Probability Estimation. To start with, we first adopt target feature extractor " + }, + { + "bbox": [ + 304, + 411, + 546, + 519 + ], + "type": "inline_equation", + "content": "\\mathcal{F}_t(\\cdot)" + }, + { + "bbox": [ + 304, + 411, + 546, + 519 + ], + "type": "text", + "content": " to obtain the latent features " + }, + { + "bbox": [ + 304, + 411, + 546, + 519 + ], + "type": "inline_equation", + "content": "z^t" + }, + { + "bbox": [ + 304, + 411, + 546, + 519 + ], + "type": "text", + "content": " on target samples. Then we should consider the matching relationship between target sample features " + }, + { + "bbox": [ + 304, + 411, + 546, + 519 + ], + "type": "inline_equation", + "content": "z^t" + }, + { + "bbox": [ + 304, + 411, + 546, + 519 + ], + "type": "text", + "content": " and source class prototype " + }, + { + "bbox": [ + 304, + 411, + 546, + 519 + ], + "type": "inline_equation", + "content": "\\mu" + }, + { + "bbox": [ + 304, + 411, + 546, + 519 + ], + "type": "text", + "content": ". Recalling that not all target samples belong to the source class categories, it means that only partial of target sample will get matched. It inspired us to propose PUOT for determining which samples belong to the unkown class categories:" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 310, + 521, + 545, + 551 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 310, + 521, + 545, + 551 + ], + "spans": [ + { + "bbox": [ + 310, + 521, + 545, + 551 + ], + "type": "interline_equation", + "content": "\\min _ {\\pi_ {i j} \\geq 0} J _ {\\text {p u o t}} = \\left\\langle \\mathcal {U}, \\boldsymbol {\\pi} \\right\\rangle + \\tau \\operatorname {K L} \\left(\\boldsymbol {\\pi} \\mathbf {1} _ {H} \\| \\boldsymbol {a}\\right) \\text {s . t .} \\boldsymbol {\\pi} ^ {\\top} \\mathbf {1} _ {N} = \\boldsymbol {b}, \\tag {2}", + "image_path": "f88bb6002ea557f2d952f068883edf181a2ef5753f33db3bdaa831f641f80905.jpg" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 304, + 551, + 546, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 551, + 546, + 714 + ], + "spans": [ + { + "bbox": [ + 304, + 551, + 546, + 714 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 304, + 551, + 546, + 714 + ], + "type": "inline_equation", + "content": "\\pi \\in \\mathbb{R}^{N\\times H}" + }, + { + "bbox": [ + 304, + 551, + 546, + 714 + ], + "type": "text", + "content": " denotes the coupling matrix of PUOT. " + }, + { + "bbox": [ + 304, + 551, + 546, + 714 + ], + "type": "inline_equation", + "content": "N" + }, + { + "bbox": [ + 304, + 551, + 546, + 714 + ], + "type": "text", + "content": " denotes the batchsize of target samples and " + }, + { + "bbox": [ + 304, + 551, + 546, + 714 + ], + "type": "inline_equation", + "content": "H" + }, + { + "bbox": [ + 304, + 551, + 546, + 714 + ], + "type": "text", + "content": " denotes the number of shared known classes. " + }, + { + "bbox": [ + 304, + 551, + 546, + 714 + ], + "type": "inline_equation", + "content": "\\pmb{a}" + }, + { + "bbox": [ + 304, + 551, + 546, + 714 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 304, + 551, + 546, + 714 + ], + "type": "inline_equation", + "content": "\\pmb{b}" + }, + { + "bbox": [ + 304, + 551, + 546, + 714 + ], + "type": "text", + "content": " denote the corresponding weights for target samples and source prototypes respectively. Specifically we set " + }, + { + "bbox": [ + 304, + 551, + 546, + 714 + ], + "type": "inline_equation", + "content": "a_{i} = 1 / N" + }, + { + "bbox": [ + 304, + 551, + 546, + 714 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 304, + 551, + 546, + 714 + ], + "type": "inline_equation", + "content": "b_{j} = 1 / H" + }, + { + "bbox": [ + 304, + 551, + 546, + 714 + ], + "type": "text", + "content": " uniformly. " + }, + { + "bbox": [ + 304, + 551, + 546, + 714 + ], + "type": "inline_equation", + "content": "\\mathcal{U}" + }, + { + "bbox": [ + 304, + 551, + 546, + 714 + ], + "type": "text", + "content": " denotes the pairwise cost between target samples " + }, + { + "bbox": [ + 304, + 551, + 546, + 714 + ], + "type": "inline_equation", + "content": "\\boldsymbol{z}^t" + }, + { + "bbox": [ + 304, + 551, + 546, + 714 + ], + "type": "text", + "content": " and corresponding prototypes " + }, + { + "bbox": [ + 304, + 551, + 546, + 714 + ], + "type": "inline_equation", + "content": "\\pmb{\\mu}^{(o)}" + }, + { + "bbox": [ + 304, + 551, + 546, + 714 + ], + "type": "text", + "content": " at the " + }, + { + "bbox": [ + 304, + 551, + 546, + 714 + ], + "type": "inline_equation", + "content": "o" + }, + { + "bbox": [ + 304, + 551, + 546, + 714 + ], + "type": "text", + "content": "-th iteration as " + }, + { + "bbox": [ + 304, + 551, + 546, + 714 + ], + "type": "inline_equation", + "content": "\\mathcal{U}_{ij} = e^{E(\\pmb{z}_i^t)}(||(\\pmb{z}_i^t)|| - \\pmb{\\mu}_j^{(o)}||\\pmb{\\mu}_j^{(o)}||)\\|_2^2" + }, + { + "bbox": [ + 304, + 551, + 546, + 714 + ], + "type": "text", + "content": ". We set " + }, + { + "bbox": [ + 304, + 551, + 546, + 714 + ], + "type": "inline_equation", + "content": "\\pmb{\\mu}^{(0)} = \\pmb{w}" + }, + { + "bbox": [ + 304, + 551, + 546, + 714 + ], + "type": "text", + "content": " for the initialization. Meanwhile " + }, + { + "bbox": [ + 304, + 551, + 546, + 714 + ], + "type": "inline_equation", + "content": "E(\\cdot)" + }, + { + "bbox": [ + 304, + 551, + 546, + 714 + ], + "type": "text", + "content": " denotes the entropy of classifier output predictions as " + }, + { + "bbox": [ + 304, + 551, + 546, + 714 + ], + "type": "inline_equation", + "content": "E(\\pmb{z}_i^t) = -\\sum_{j=1}^{|\\mathcal{C}^s|} \\sigma(\\mathcal{G}(\\mathcal{F}_t(\\pmb{x}_i^t))) [j] \\cdot \\log(\\sigma(\\mathcal{G}(\\mathcal{F}_t(\\pmb{x}_i^t)))) [j])" + }, + { + "bbox": [ + 304, + 551, + 546, + 714 + ], + "type": "text", + "content": " where " + }, + { + "bbox": [ + 304, + 551, + 546, + 714 + ], + "type": "inline_equation", + "content": "\\sigma(\\cdot)" + }, + { + "bbox": [ + 304, + 551, + 546, + 714 + ], + "type": "text", + "content": " denotes the softmax operation. When " + }, + { + "bbox": [ + 304, + 551, + 546, + 714 + ], + "type": "inline_equation", + "content": "E(\\pmb{x}^t)" + }, + { + "bbox": [ + 304, + 551, + 546, + 714 + ], + "type": "text", + "content": " is much larger, it indicates that the model exhibits greater" + } + ] + } + ], + "index": 23 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "text", + "content": "4929" + } + ] + } + ], + "index": 24 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 2 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 85, + 72, + 509, + 224 + ], + "blocks": [ + { + "bbox": [ + 85, + 72, + 509, + 224 + ], + "lines": [ + { + "bbox": [ + 85, + 72, + 509, + 224 + ], + "spans": [ + { + "bbox": [ + 85, + 72, + 509, + 224 + ], + "type": "image", + "image_path": "041a74fd4ccee110878314f35814047408f4c2d35817f14f2708859b7530c881.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 46, + 227, + 547, + 274 + ], + "lines": [ + { + "bbox": [ + 46, + 227, + 547, + 274 + ], + "spans": [ + { + "bbox": [ + 46, + 227, + 547, + 274 + ], + "type": "text", + "content": "Figure 2. The architectures of Distinguish Then Assign (DTE). We first train source feature extractor " + }, + { + "bbox": [ + 46, + 227, + 547, + 274 + ], + "type": "inline_equation", + "content": "\\mathcal{F}_s(\\cdot)" + }, + { + "bbox": [ + 46, + 227, + 547, + 274 + ], + "type": "text", + "content": " and classifier " + }, + { + "bbox": [ + 46, + 227, + 547, + 274 + ], + "type": "inline_equation", + "content": "\\mathcal{G}(\\cdot)" + }, + { + "bbox": [ + 46, + 227, + 547, + 274 + ], + "type": "text", + "content": " in the source domain. Then we share the well-trained source model to the target domain. We fine-tune target feature extractor " + }, + { + "bbox": [ + 46, + 227, + 547, + 274 + ], + "type": "inline_equation", + "content": "\\mathcal{F}_t(\\cdot)" + }, + { + "bbox": [ + 46, + 227, + 547, + 274 + ], + "type": "text", + "content": " based on " + }, + { + "bbox": [ + 46, + 227, + 547, + 274 + ], + "type": "inline_equation", + "content": "\\mathcal{F}_s(\\cdot)" + }, + { + "bbox": [ + 46, + 227, + 547, + 274 + ], + "type": "text", + "content": " and fix classifier " + }, + { + "bbox": [ + 46, + 227, + 547, + 274 + ], + "type": "inline_equation", + "content": "\\mathcal{G}(\\cdot)" + }, + { + "bbox": [ + 46, + 227, + 547, + 274 + ], + "type": "text", + "content": ". Then we propose weight barcode estimation for distinguishing the known and unknown target samples via distinguisher " + }, + { + "bbox": [ + 46, + 227, + 547, + 274 + ], + "type": "inline_equation", + "content": "\\hat{\\mathcal{G}}(\\cdot)" + }, + { + "bbox": [ + 46, + 227, + 547, + 274 + ], + "type": "text", + "content": ". Finally we propose sparse label assignment for exploiting appropriate pseudo labels for the remaining target samples." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "type": "image", + "bbox": [ + 52, + 277, + 282, + 434 + ], + "blocks": [ + { + "bbox": [ + 52, + 277, + 282, + 434 + ], + "lines": [ + { + "bbox": [ + 52, + 277, + 282, + 434 + ], + "spans": [ + { + "bbox": [ + 52, + 277, + 282, + 434 + ], + "type": "image", + "image_path": "043f30434967b835b04bf6ce7e29c0f2d737b3843b00243b27e1483aa6ab5b43.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 46, + 437, + 287, + 502 + ], + "lines": [ + { + "bbox": [ + 46, + 437, + 287, + 502 + ], + "spans": [ + { + "bbox": [ + 46, + 437, + 287, + 502 + ], + "type": "text", + "content": "Figure 3. The illustration on the proposed weight barcode estimation with toy examples. We first sample 100 source/target data samples and set " + }, + { + "bbox": [ + 46, + 437, + 287, + 502 + ], + "type": "inline_equation", + "content": "30\\%" + }, + { + "bbox": [ + 46, + 437, + 287, + 502 + ], + "type": "text", + "content": " or " + }, + { + "bbox": [ + 46, + 437, + 287, + 502 + ], + "type": "inline_equation", + "content": "70\\%" + }, + { + "bbox": [ + 46, + 437, + 287, + 502 + ], + "type": "text", + "content": " of target data samples from unknown categories. Then we adopt marginal probability estimation to obtain the results of target weights. Finally we calculate and visualize the result of " + }, + { + "bbox": [ + 46, + 437, + 287, + 502 + ], + "type": "inline_equation", + "content": "\\hat{Q}" + }, + { + "bbox": [ + 46, + 437, + 287, + 502 + ], + "type": "text", + "content": " to distinguish the unknown target samples." + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_caption" + } + ], + "index": 2 + }, + { + "bbox": [ + 46, + 512, + 287, + 619 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 512, + 287, + 619 + ], + "spans": [ + { + "bbox": [ + 46, + 512, + 287, + 619 + ], + "type": "text", + "content": "uncertainty regarding the target samples into known categories. Therefore, it is reasonable to consider such information into the cost calculation on " + }, + { + "bbox": [ + 46, + 512, + 287, + 619 + ], + "type": "inline_equation", + "content": "\\mathcal{U}" + }, + { + "bbox": [ + 46, + 512, + 287, + 619 + ], + "type": "text", + "content": ". The KL divergence can be calculated via KL " + }, + { + "bbox": [ + 46, + 512, + 287, + 619 + ], + "type": "inline_equation", + "content": "(\\pmb{x}||\\pmb{z}) = \\sum_{d=1}^{D}[x_d\\log(x_d/z_d) - x_d + z_d]" + }, + { + "bbox": [ + 46, + 512, + 287, + 619 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 46, + 512, + 287, + 619 + ], + "type": "inline_equation", + "content": "\\tau" + }, + { + "bbox": [ + 46, + 512, + 287, + 619 + ], + "type": "text", + "content": " denotes the hyper parameters. Unlike the classic optimal transport problem, we relax one of the marginal constraints for the target samples via the KL divergence term to avoid the mismatch. To solve the above problem, we first propose the theorem on the dual form of PUOT:" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 47, + 620, + 287, + 644 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 620, + 287, + 644 + ], + "spans": [ + { + "bbox": [ + 47, + 620, + 287, + 644 + ], + "type": "text", + "content": "Theorem 1. Given the PUOT problem " + }, + { + "bbox": [ + 47, + 620, + 287, + 644 + ], + "type": "inline_equation", + "content": "J_{\\mathrm{puot}}" + }, + { + "bbox": [ + 47, + 620, + 287, + 644 + ], + "type": "text", + "content": " as shown in Eq.(2), it has the following dual form:" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 57, + 646, + 287, + 685 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 646, + 287, + 685 + ], + "spans": [ + { + "bbox": [ + 57, + 646, + 287, + 685 + ], + "type": "interline_equation", + "content": "\\min _ {\\boldsymbol {f}, \\boldsymbol {g}, \\zeta} \\left[ \\tau \\sum_ {i = 1} ^ {N} a _ {i} \\exp \\left(- \\frac {f _ {i} + \\zeta}{\\tau}\\right) - \\sum_ {j = 1} ^ {H} b _ {j} \\left(g _ {j} - \\zeta\\right) \\right] \\tag {3}", + "image_path": "fd3fa2e4bd1470a3c4c7d59e386f97ce81ce93b63de76b6c92ee09a3da6b31c0.jpg" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 56, + 686, + 204, + 700 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 686, + 204, + 700 + ], + "spans": [ + { + "bbox": [ + 56, + 686, + 204, + 700 + ], + "type": "interline_equation", + "content": "s. t. f _ {i} + g _ {j} + s _ {i j} = \\mathcal {U} _ {i j}, \\quad s _ {i j} \\geq 0,", + "image_path": "debc6cbb9db7dab7d2bb2601ae71558da5cf569fa664daa1bae30eed4131db44.jpg" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 47, + 702, + 287, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 702, + 287, + 714 + ], + "spans": [ + { + "bbox": [ + 47, + 702, + 287, + 714 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 47, + 702, + 287, + 714 + ], + "type": "inline_equation", + "content": "f, g, s" + }, + { + "bbox": [ + 47, + 702, + 287, + 714 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 47, + 702, + 287, + 714 + ], + "type": "inline_equation", + "content": "\\zeta" + }, + { + "bbox": [ + 47, + 702, + 287, + 714 + ], + "type": "text", + "content": " denote Lagrange multipliers. Mean" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 305, + 277, + 546, + 314 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 277, + 546, + 314 + ], + "spans": [ + { + "bbox": [ + 305, + 277, + 546, + 314 + ], + "type": "text", + "content": "while " + }, + { + "bbox": [ + 305, + 277, + 546, + 314 + ], + "type": "inline_equation", + "content": "s_{ij} \\geq 0" + }, + { + "bbox": [ + 305, + 277, + 546, + 314 + ], + "type": "text", + "content": " according to the KKT conditions. Moreover, PUOT will be transformed into sample-reweighted optimal transport problem according to its dual form:" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 312, + 316, + 362, + 335 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 312, + 316, + 362, + 335 + ], + "spans": [ + { + "bbox": [ + 312, + 316, + 362, + 335 + ], + "type": "interline_equation", + "content": "\\min _ {\\boldsymbol {\\pi} \\geq 0} \\langle \\mathcal {U}, \\boldsymbol {\\pi} \\rangle", + "image_path": "1fe501417b97931085c2da777114895c7272897a1ab16048f115cabd1d6cf17f.jpg" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 311, + 337, + 545, + 371 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 311, + 337, + 545, + 371 + ], + "spans": [ + { + "bbox": [ + 311, + 337, + 545, + 371 + ], + "type": "interline_equation", + "content": "s. t. \\sum_ {j = 1} ^ {H} \\pi_ {i j} = a _ {i} \\exp \\left(- \\frac {f _ {i} ^ {*} + \\zeta^ {*}}{\\tau}\\right), \\quad \\sum_ {i = 1} ^ {N} \\pi_ {i j} = b _ {j}. \\tag {4}", + "image_path": "6a002b4f44457e72af00cf395f985da5e86f9cb912170d87000a0aa0a6225d05.jpg" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 305, + 373, + 546, + 529 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 373, + 546, + 529 + ], + "spans": [ + { + "bbox": [ + 305, + 373, + 546, + 529 + ], + "type": "text", + "content": "The proof of Theorem 1 is provided in Appendix. That is, the target data samples will be sample-reweighted with coefficient " + }, + { + "bbox": [ + 305, + 373, + 546, + 529 + ], + "type": "inline_equation", + "content": "\\exp \\left(-\\left(f_{i}^{*} + \\zeta^{*}\\right) / \\tau\\right)" + }, + { + "bbox": [ + 305, + 373, + 546, + 529 + ], + "type": "text", + "content": " accordingly. At that time the marginal probability on target sample " + }, + { + "bbox": [ + 305, + 373, + 546, + 529 + ], + "type": "inline_equation", + "content": "\\pmb{x}_i^t" + }, + { + "bbox": [ + 305, + 373, + 546, + 529 + ], + "type": "text", + "content": " can be given as " + }, + { + "bbox": [ + 305, + 373, + 546, + 529 + ], + "type": "inline_equation", + "content": "q_{\\tau}(\\pmb{x}_i^t) = a_i\\exp \\left(-\\left(f_i^* +\\zeta^*\\right) / \\tau\\right)" + }, + { + "bbox": [ + 305, + 373, + 546, + 529 + ], + "type": "text", + "content": ". Kindly note that we do not necessarily require an explicit expression for solving " + }, + { + "bbox": [ + 305, + 373, + 546, + 529 + ], + "type": "inline_equation", + "content": "\\pi" + }, + { + "bbox": [ + 305, + 373, + 546, + 529 + ], + "type": "text", + "content": ", which involves " + }, + { + "bbox": [ + 305, + 373, + 546, + 529 + ], + "type": "inline_equation", + "content": "(N\\times H)" + }, + { + "bbox": [ + 305, + 373, + 546, + 529 + ], + "type": "text", + "content": " variables and can be quite complex. Meanwhile we mainly focus on solving these dual form Lagrange multipliers in Eq.(3) with " + }, + { + "bbox": [ + 305, + 373, + 546, + 529 + ], + "type": "inline_equation", + "content": "(N + H)" + }, + { + "bbox": [ + 305, + 373, + 546, + 529 + ], + "type": "text", + "content": " variables. Specifically, we can utilize the " + }, + { + "bbox": [ + 305, + 373, + 546, + 529 + ], + "type": "inline_equation", + "content": "c" + }, + { + "bbox": [ + 305, + 373, + 546, + 529 + ], + "type": "text", + "content": "-transform [41, 55, 71] in classical optimal transport theorem to reach the upper bound of " + }, + { + "bbox": [ + 305, + 373, + 546, + 529 + ], + "type": "inline_equation", + "content": "\\pmb{g}" + }, + { + "bbox": [ + 305, + 373, + 546, + 529 + ], + "type": "text", + "content": " as " + }, + { + "bbox": [ + 305, + 373, + 546, + 529 + ], + "type": "inline_equation", + "content": "g_{j} = \\inf_{k\\in [N]}\\left(\\mathcal{U}_{kj} - f_{k}\\right)" + }, + { + "bbox": [ + 305, + 373, + 546, + 529 + ], + "type": "text", + "content": ". By taking the upper bound of " + }, + { + "bbox": [ + 305, + 373, + 546, + 529 + ], + "type": "inline_equation", + "content": "\\pmb{g}" + }, + { + "bbox": [ + 305, + 373, + 546, + 529 + ], + "type": "text", + "content": " into Eq.(3), we can reach the final optimization formula as:" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 307, + 531, + 545, + 579 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 531, + 545, + 579 + ], + "spans": [ + { + "bbox": [ + 307, + 531, + 545, + 579 + ], + "type": "interline_equation", + "content": "\\min _ {\\boldsymbol {f}, \\zeta} \\left[ \\tau \\sum_ {i = 1} ^ {N} a _ {i} e ^ {- \\frac {f _ {i} + \\zeta}{\\tau}} - \\sum_ {j = 1} ^ {H} \\left(\\inf _ {k \\in [ N ]} \\left(\\mathcal {U} _ {k j} - f _ {k}\\right) - \\zeta\\right) b _ {j} \\right]. \\tag {5}", + "image_path": "dcf3a300e72c3fdb080b69b93c2db2012e76b6dc102b078393b079cb7b3cf52c.jpg" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 304, + 580, + 545, + 640 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 580, + 545, + 640 + ], + "spans": [ + { + "bbox": [ + 304, + 580, + 545, + 640 + ], + "type": "text", + "content": "That equation can be optimized easily. Due to the space limited, the details of optimization procedures for Eq.(5) are provided in Appendix. In summary, we only need to figure out the marginal probability on target samples " + }, + { + "bbox": [ + 304, + 580, + 545, + 640 + ], + "type": "inline_equation", + "content": "q_{\\tau}(\\boldsymbol{x}_i^t)" + }, + { + "bbox": [ + 304, + 580, + 545, + 640 + ], + "type": "text", + "content": " via solving the optimal results of " + }, + { + "bbox": [ + 304, + 580, + 545, + 640 + ], + "type": "inline_equation", + "content": "f" + }, + { + "bbox": [ + 304, + 580, + 545, + 640 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 304, + 580, + 545, + 640 + ], + "type": "inline_equation", + "content": "\\zeta" + }, + { + "bbox": [ + 304, + 580, + 545, + 640 + ], + "type": "text", + "content": " in Eq.(5)." + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 304, + 642, + 546, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 642, + 546, + 713 + ], + "spans": [ + { + "bbox": [ + 304, + 642, + 546, + 713 + ], + "type": "text", + "content": "Weight Barcode Establishment. According to the optimization process, we can observe that choosing different value of " + }, + { + "bbox": [ + 304, + 642, + 546, + 713 + ], + "type": "inline_equation", + "content": "\\tau" + }, + { + "bbox": [ + 304, + 642, + 546, + 713 + ], + "type": "text", + "content": " could lead to different marginal probability. Specifically, as the value of " + }, + { + "bbox": [ + 304, + 642, + 546, + 713 + ], + "type": "inline_equation", + "content": "\\tau" + }, + { + "bbox": [ + 304, + 642, + 546, + 713 + ], + "type": "text", + "content": " increases, a larger number of target samples will be assigned higher weights, enhancing their likelihood of being matched. However, we do not" + } + ] + } + ], + "index": 15 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "text", + "content": "4930" + } + ] + } + ], + "index": 16 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 3 + }, + { + "para_blocks": [ + { + "bbox": [ + 47, + 72, + 289, + 387 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 72, + 289, + 387 + ], + "spans": [ + { + "bbox": [ + 47, + 72, + 289, + 387 + ], + "type": "text", + "content": "know the exact percentage of how many samples are unknown in the target domain which makes it relatively hard to choose a proper value of " + }, + { + "bbox": [ + 47, + 72, + 289, + 387 + ], + "type": "inline_equation", + "content": "\\tau" + }, + { + "bbox": [ + 47, + 72, + 289, + 387 + ], + "type": "text", + "content": ". To address this issue, we propose an innovative weight barcode mechanism to effectively identify target samples that are more likely belong to unknown classes. We first calculate " + }, + { + "bbox": [ + 47, + 72, + 289, + 387 + ], + "type": "inline_equation", + "content": "q_{\\tau}(\\boldsymbol{x}^t)" + }, + { + "bbox": [ + 47, + 72, + 289, + 387 + ], + "type": "text", + "content": " by varying " + }, + { + "bbox": [ + 47, + 72, + 289, + 387 + ], + "type": "inline_equation", + "content": "\\tau" + }, + { + "bbox": [ + 47, + 72, + 289, + 387 + ], + "type": "text", + "content": " in range of " + }, + { + "bbox": [ + 47, + 72, + 289, + 387 + ], + "type": "inline_equation", + "content": "\\{0.01, 0.05, 0.1, 0.5, 1, 5, 10\\}" + }, + { + "bbox": [ + 47, + 72, + 289, + 387 + ], + "type": "text", + "content": " and collecting them as " + }, + { + "bbox": [ + 47, + 72, + 289, + 387 + ], + "type": "inline_equation", + "content": "\\boldsymbol{Q} \\in \\mathbb{R}^{N \\times 7}" + }, + { + "bbox": [ + 47, + 72, + 289, + 387 + ], + "type": "text", + "content": ". Then we can establish the binary weight barcode matrix " + }, + { + "bbox": [ + 47, + 72, + 289, + 387 + ], + "type": "inline_equation", + "content": "\\widehat{\\boldsymbol{Q}} \\in \\mathbb{R}^{N \\times 7}" + }, + { + "bbox": [ + 47, + 72, + 289, + 387 + ], + "type": "text", + "content": " with the following process as " + }, + { + "bbox": [ + 47, + 72, + 289, + 387 + ], + "type": "inline_equation", + "content": "\\widehat{\\boldsymbol{Q}}_{ij} = \\mathrm{func}(Q_{ij} - 1/N)" + }, + { + "bbox": [ + 47, + 72, + 289, + 387 + ], + "type": "text", + "content": " where " + }, + { + "bbox": [ + 47, + 72, + 289, + 387 + ], + "type": "inline_equation", + "content": "\\mathrm{func}(x) = 1" + }, + { + "bbox": [ + 47, + 72, + 289, + 387 + ], + "type": "text", + "content": " when " + }, + { + "bbox": [ + 47, + 72, + 289, + 387 + ], + "type": "inline_equation", + "content": "x \\geq 0" + }, + { + "bbox": [ + 47, + 72, + 289, + 387 + ], + "type": "text", + "content": " otherwise " + }, + { + "bbox": [ + 47, + 72, + 289, + 387 + ], + "type": "inline_equation", + "content": "\\mathrm{func}(x) = 0" + }, + { + "bbox": [ + 47, + 72, + 289, + 387 + ], + "type": "text", + "content": ". It is reasonable since PUOT will enhance the target sample weights if they exhibit a higher degree of similarity to the source data. Meanwhile the rest of unlike target samples will gain less weights and we set then " + }, + { + "bbox": [ + 47, + 72, + 289, + 387 + ], + "type": "inline_equation", + "content": "\\widehat{\\boldsymbol{Q}}_{ij} = 0" + }, + { + "bbox": [ + 47, + 72, + 289, + 387 + ], + "type": "text", + "content": ". To make this process more intuitive, we further provide an numerical example in Fig. 3. Specifically, the red and blue points denote the source and target data samples respectively. We first figure out the pairwise distance between red data " + }, + { + "bbox": [ + 47, + 72, + 289, + 387 + ], + "type": "inline_equation", + "content": "\\boldsymbol{z}_i^{\\mathrm{Red}}" + }, + { + "bbox": [ + 47, + 72, + 289, + 387 + ], + "type": "text", + "content": " and blue data " + }, + { + "bbox": [ + 47, + 72, + 289, + 387 + ], + "type": "inline_equation", + "content": "\\boldsymbol{z}_j^{\\mathrm{Blue}}" + }, + { + "bbox": [ + 47, + 72, + 289, + 387 + ], + "type": "text", + "content": " via " + }, + { + "bbox": [ + 47, + 72, + 289, + 387 + ], + "type": "inline_equation", + "content": "\\mathcal{U}_{ij} = ||\\boldsymbol{z}_i^{\\mathrm{Red}} - \\boldsymbol{z}_j^{\\mathrm{Blue}}||_2^2" + }, + { + "bbox": [ + 47, + 72, + 289, + 387 + ], + "type": "text", + "content": ". Then we calculate the weight barcode as " + }, + { + "bbox": [ + 47, + 72, + 289, + 387 + ], + "type": "inline_equation", + "content": "\\widehat{\\boldsymbol{Q}}" + }, + { + "bbox": [ + 47, + 72, + 289, + 387 + ], + "type": "text", + "content": " by using marginal probability estimation and visualize the results. We can observe that the unknown target samples achieve lower value of " + }, + { + "bbox": [ + 47, + 72, + 289, + 387 + ], + "type": "inline_equation", + "content": "\\widehat{\\boldsymbol{Q}}" + }, + { + "bbox": [ + 47, + 72, + 289, + 387 + ], + "type": "text", + "content": " which can be easily identify. It indicates that the proposed method is adaptable to varying ratios of unknown data samples, even in the absence of prior knowledge." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 47, + 387, + 289, + 544 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 387, + 289, + 544 + ], + "spans": [ + { + "bbox": [ + 47, + 387, + 289, + 544 + ], + "type": "text", + "content": "After we obtain the binary weight barcode matrix " + }, + { + "bbox": [ + 47, + 387, + 289, + 544 + ], + "type": "inline_equation", + "content": "\\widehat{Q}" + }, + { + "bbox": [ + 47, + 387, + 289, + 544 + ], + "type": "text", + "content": ", we can further adopt " + }, + { + "bbox": [ + 47, + 387, + 289, + 544 + ], + "type": "inline_equation", + "content": "K" + }, + { + "bbox": [ + 47, + 387, + 289, + 544 + ], + "type": "text", + "content": "-Means algorithm [28] on " + }, + { + "bbox": [ + 47, + 387, + 289, + 544 + ], + "type": "inline_equation", + "content": "\\widehat{Q}" + }, + { + "bbox": [ + 47, + 387, + 289, + 544 + ], + "type": "text", + "content": " to cluster them into two groups with corresponding centers " + }, + { + "bbox": [ + 47, + 387, + 289, + 544 + ], + "type": "inline_equation", + "content": "\\mathcal{C}_0" + }, + { + "bbox": [ + 47, + 387, + 289, + 544 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 47, + 387, + 289, + 544 + ], + "type": "inline_equation", + "content": "\\mathcal{C}_1" + }, + { + "bbox": [ + 47, + 387, + 289, + 544 + ], + "type": "text", + "content": ". We further denote " + }, + { + "bbox": [ + 47, + 387, + 289, + 544 + ], + "type": "inline_equation", + "content": "\\mathcal{C}_{\\mathrm{known}} = \\max \\left(\\| \\mathcal{C}_0\\| _2, \\| \\mathcal{C}_1\\| _2\\right)" + }, + { + "bbox": [ + 47, + 387, + 289, + 544 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 47, + 387, + 289, + 544 + ], + "type": "inline_equation", + "content": "\\mathcal{C}_{\\mathrm{unknown}} = \\min \\left(\\| \\mathcal{C}_0\\| _2, \\| \\mathcal{C}_1\\| _2\\right)" + }, + { + "bbox": [ + 47, + 387, + 289, + 544 + ], + "type": "text", + "content": " by considering the norm of centers. Then we can determine the target samples by measuring the Euclidean distance among " + }, + { + "bbox": [ + 47, + 387, + 289, + 544 + ], + "type": "inline_equation", + "content": "\\widehat{Q}" + }, + { + "bbox": [ + 47, + 387, + 289, + 544 + ], + "type": "text", + "content": " and centers " + }, + { + "bbox": [ + 47, + 387, + 289, + 544 + ], + "type": "inline_equation", + "content": "\\mathcal{C}_{\\mathrm{known}}" + }, + { + "bbox": [ + 47, + 387, + 289, + 544 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 47, + 387, + 289, + 544 + ], + "type": "inline_equation", + "content": "\\mathcal{C}_{\\mathrm{unknown}}" + }, + { + "bbox": [ + 47, + 387, + 289, + 544 + ], + "type": "text", + "content": ". If the target samples are found to be in close proximity to " + }, + { + "bbox": [ + 47, + 387, + 289, + 544 + ], + "type": "inline_equation", + "content": "\\mathcal{C}_{\\mathrm{unknown}}" + }, + { + "bbox": [ + 47, + 387, + 289, + 544 + ], + "type": "text", + "content": ", we can classify them as unknown data as " + }, + { + "bbox": [ + 47, + 387, + 289, + 544 + ], + "type": "inline_equation", + "content": "\\boldsymbol{x}_{\\mathrm{unk}}^t" + }, + { + "bbox": [ + 47, + 387, + 289, + 544 + ], + "type": "text", + "content": ". The remaining target domain samples can all be considered as belonging to the known categories as " + }, + { + "bbox": [ + 47, + 387, + 289, + 544 + ], + "type": "inline_equation", + "content": "\\boldsymbol{x}_{\\mathrm{kn0}}^t" + }, + { + "bbox": [ + 47, + 387, + 289, + 544 + ], + "type": "text", + "content": " and as " + }, + { + "bbox": [ + 47, + 387, + 289, + 544 + ], + "type": "inline_equation", + "content": "\\boldsymbol{x}^t = \\boldsymbol{x}_{\\mathrm{unk}}^t \\cup \\boldsymbol{x}_{\\mathrm{kn0}}^t" + }, + { + "bbox": [ + 47, + 387, + 289, + 544 + ], + "type": "text", + "content": ". In summary, we can now determine the known/unknown target samples accordingly." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 47, + 551, + 189, + 563 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 551, + 189, + 563 + ], + "spans": [ + { + "bbox": [ + 47, + 551, + 189, + 563 + ], + "type": "text", + "content": "3.2. Sparse Label Assignment" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 46, + 570, + 289, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 570, + 289, + 714 + ], + "spans": [ + { + "bbox": [ + 46, + 570, + 289, + 714 + ], + "type": "text", + "content": "Since we have exploited the target samples with unknown classes " + }, + { + "bbox": [ + 46, + 570, + 289, + 714 + ], + "type": "inline_equation", + "content": "\\boldsymbol{x}_{\\mathrm{unk}}^{t}" + }, + { + "bbox": [ + 46, + 570, + 289, + 714 + ], + "type": "text", + "content": " via proposed weight barcode estimation method, we should explore the possible labels for the rest target samples " + }, + { + "bbox": [ + 46, + 570, + 289, + 714 + ], + "type": "inline_equation", + "content": "\\boldsymbol{x}_{\\mathrm{kno}}^{t}" + }, + { + "bbox": [ + 46, + 570, + 289, + 714 + ], + "type": "text", + "content": ". Recall that " + }, + { + "bbox": [ + 46, + 570, + 289, + 714 + ], + "type": "inline_equation", + "content": "\\boldsymbol{x}_{\\mathrm{kno}}^{t}" + }, + { + "bbox": [ + 46, + 570, + 289, + 714 + ], + "type": "text", + "content": " share the same categories with the corresponding prototypes " + }, + { + "bbox": [ + 46, + 570, + 289, + 714 + ], + "type": "inline_equation", + "content": "\\boldsymbol{\\mu}^{s}" + }, + { + "bbox": [ + 46, + 570, + 289, + 714 + ], + "type": "text", + "content": ", one can adopt domain adaptation manners for finding the latent relationship. Meanwhile, optimal transport techniques could be rather useful in reducing the domain discrepancy via discovering the most proper pairwise matching across domains [9, 25, 75]. However, directly optimizing optimal transport with linear programming would involve heavy computation cost [1, 34, 92]. Although utilizing entropy regularization" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 304, + 72, + 545, + 133 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 72, + 545, + 133 + ], + "spans": [ + { + "bbox": [ + 304, + 72, + 545, + 133 + ], + "type": "text", + "content": "term could enhance the efficiency, it will bring about relatively dense matching solution which is always inaccurate as shown in Fig. 4(a) [6, 10, 50]. To alleviate this issue, we propose Sparse Sample-Label Matching (SSLM) on optimal transport via involving " + }, + { + "bbox": [ + 304, + 72, + 545, + 133 + ], + "type": "inline_equation", + "content": "\\ell_2" + }, + { + "bbox": [ + 304, + 72, + 545, + 133 + ], + "type": "text", + "content": "-norm proximal term:" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 331, + 133, + 545, + 166 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 331, + 133, + 545, + 166 + ], + "spans": [ + { + "bbox": [ + 331, + 133, + 545, + 166 + ], + "type": "interline_equation", + "content": "\\min _ {\\gamma \\in \\Delta} \\ell = \\sum_ {i = 1} ^ {N _ {\\mathrm {k n o}}} \\sum_ {j = 1} ^ {H} \\left[ \\gamma_ {i j} C _ {i j} + \\frac {\\epsilon}{2} \\left(\\gamma_ {i j} - \\gamma_ {i j} ^ {(l)}\\right) ^ {2} \\right], \\tag {6}", + "image_path": "87b2649e3bca44c3c0eed95494b69b992bdb50ce97ad715535985f793eee7341.jpg" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 304, + 167, + 547, + 345 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 167, + 547, + 345 + ], + "spans": [ + { + "bbox": [ + 304, + 167, + 547, + 345 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 304, + 167, + 547, + 345 + ], + "type": "inline_equation", + "content": "\\Delta = \\{\\sum_{j=1}^{H} \\gamma_{ij} = \\widehat{a}_i = 1, \\sum_{i=1}^{N_{\\mathrm{kno}}} \\gamma_{ij} = \\widehat{b}_j = N_{\\mathrm{kno}} / H, \\gamma_{ij} \\geq 0\\}" + }, + { + "bbox": [ + 304, + 167, + 547, + 345 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 304, + 167, + 547, + 345 + ], + "type": "inline_equation", + "content": "N_{\\mathrm{kno}}" + }, + { + "bbox": [ + 304, + 167, + 547, + 345 + ], + "type": "text", + "content": " denotes the number of target samples " + }, + { + "bbox": [ + 304, + 167, + 547, + 345 + ], + "type": "inline_equation", + "content": "\\boldsymbol{x}_{\\mathrm{kno}}^t" + }, + { + "bbox": [ + 304, + 167, + 547, + 345 + ], + "type": "text", + "content": ". " + }, + { + "bbox": [ + 304, + 167, + 547, + 345 + ], + "type": "inline_equation", + "content": "C = -\\log[\\sigma[\\mathcal{G}(\\mathcal{F}_t(\\boldsymbol{x}_{\\mathrm{kno}}^t))]]" + }, + { + "bbox": [ + 304, + 167, + 547, + 345 + ], + "type": "text", + "content": " denote the output classification probability with negative logarithm. Indeed, when " + }, + { + "bbox": [ + 304, + 167, + 547, + 345 + ], + "type": "inline_equation", + "content": "C_{ij}" + }, + { + "bbox": [ + 304, + 167, + 547, + 345 + ], + "type": "text", + "content": " is smaller, it indicates that " + }, + { + "bbox": [ + 304, + 167, + 547, + 345 + ], + "type": "inline_equation", + "content": "\\boldsymbol{x}_{\\mathrm{kno},i}^t" + }, + { + "bbox": [ + 304, + 167, + 547, + 345 + ], + "type": "text", + "content": " has a higher probability of belonging to the " + }, + { + "bbox": [ + 304, + 167, + 547, + 345 + ], + "type": "inline_equation", + "content": "j" + }, + { + "bbox": [ + 304, + 167, + 547, + 345 + ], + "type": "text", + "content": "-th class [3, 27, 45]. " + }, + { + "bbox": [ + 304, + 167, + 547, + 345 + ], + "type": "inline_equation", + "content": "\\widehat{a}_i" + }, + { + "bbox": [ + 304, + 167, + 547, + 345 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 304, + 167, + 547, + 345 + ], + "type": "inline_equation", + "content": "\\widehat{b}_j" + }, + { + "bbox": [ + 304, + 167, + 547, + 345 + ], + "type": "text", + "content": " denote the samples weights and we set " + }, + { + "bbox": [ + 304, + 167, + 547, + 345 + ], + "type": "inline_equation", + "content": "\\widehat{a}_i = 1" + }, + { + "bbox": [ + 304, + 167, + 547, + 345 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 304, + 167, + 547, + 345 + ], + "type": "inline_equation", + "content": "\\widehat{b}_j = \\frac{N_{\\mathrm{kno}}}{H}" + }, + { + "bbox": [ + 304, + 167, + 547, + 345 + ], + "type": "text", + "content": " respectively. Meanwhile " + }, + { + "bbox": [ + 304, + 167, + 547, + 345 + ], + "type": "inline_equation", + "content": "\\sum_{i=1}^{M} \\sum_{j=1}^{N} (\\gamma_{ij} - \\gamma_{ij}^{(l)})^2" + }, + { + "bbox": [ + 304, + 167, + 547, + 345 + ], + "type": "text", + "content": " denotes the proximal point term with " + }, + { + "bbox": [ + 304, + 167, + 547, + 345 + ], + "type": "inline_equation", + "content": "\\ell_2" + }, + { + "bbox": [ + 304, + 167, + 547, + 345 + ], + "type": "text", + "content": "-norm. Incorporating " + }, + { + "bbox": [ + 304, + 167, + 547, + 345 + ], + "type": "inline_equation", + "content": "\\ell_2" + }, + { + "bbox": [ + 304, + 167, + 547, + 345 + ], + "type": "text", + "content": "-norm with proximal point method [90], we can reach more sparse and robust results on transportation plan " + }, + { + "bbox": [ + 304, + 167, + 547, + 345 + ], + "type": "inline_equation", + "content": "\\gamma^*" + }, + { + "bbox": [ + 304, + 167, + 547, + 345 + ], + "type": "text", + "content": ". We initialize " + }, + { + "bbox": [ + 304, + 167, + 547, + 345 + ], + "type": "inline_equation", + "content": "\\gamma_{ij}^{(0)} = \\frac{1}{H}" + }, + { + "bbox": [ + 304, + 167, + 547, + 345 + ], + "type": "text", + "content": " for the following iterations. Then we can easily obtain the Fenchel-Legendre dual form of the original problem as:" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 305, + 346, + 552, + 392 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 346, + 552, + 392 + ], + "spans": [ + { + "bbox": [ + 305, + 346, + 552, + 392 + ], + "type": "interline_equation", + "content": "\\max _ {\\psi , \\phi} \\sum_ {i = 1} ^ {N _ {\\mathrm {k n o}}} \\psi_ {i} \\widehat {a} _ {i} + \\sum_ {j = 1} ^ {H} \\phi_ {j} \\widehat {b} _ {j} - \\frac {\\epsilon}{2} \\sum_ {i = 1} ^ {N _ {\\mathrm {k n o}}} \\sum_ {j = 1} ^ {H} \\left[ \\frac {\\psi_ {i} + \\phi_ {j} - \\widehat {C} _ {i j}}{\\epsilon} \\right] _ {+} ^ {2},", + "image_path": "5944063f838755dbf5b961e9100a2822eab31b06437e7e3aeafcbec607791de8.jpg" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 304, + 392, + 545, + 488 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 392, + 545, + 488 + ], + "spans": [ + { + "bbox": [ + 304, + 392, + 545, + 488 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 304, + 392, + 545, + 488 + ], + "type": "inline_equation", + "content": "\\phi" + }, + { + "bbox": [ + 304, + 392, + 545, + 488 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 304, + 392, + 545, + 488 + ], + "type": "inline_equation", + "content": "\\psi" + }, + { + "bbox": [ + 304, + 392, + 545, + 488 + ], + "type": "text", + "content": " denote the Lagrange multipliers. Meanwhile we set " + }, + { + "bbox": [ + 304, + 392, + 545, + 488 + ], + "type": "inline_equation", + "content": "[x]_{+}^{2} = (\\max (0,x))^{2}" + }, + { + "bbox": [ + 304, + 392, + 545, + 488 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 304, + 392, + 545, + 488 + ], + "type": "inline_equation", + "content": "\\widehat{C}_{ij} = C_{ij} - \\epsilon \\gamma_{ij}^{(l)}" + }, + { + "bbox": [ + 304, + 392, + 545, + 488 + ], + "type": "text", + "content": ". Note that the original constrained optimization problem in Eq.(6) turns to be unconstrained after the duality transformation in Eq.(7). The optimization procedure on solving variables " + }, + { + "bbox": [ + 304, + 392, + 545, + 488 + ], + "type": "inline_equation", + "content": "\\psi" + }, + { + "bbox": [ + 304, + 392, + 545, + 488 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 304, + 392, + 545, + 488 + ], + "type": "inline_equation", + "content": "\\phi" + }, + { + "bbox": [ + 304, + 392, + 545, + 488 + ], + "type": "text", + "content": " are provided in Appendix. After that, we can further calculate the optimal solution of " + }, + { + "bbox": [ + 304, + 392, + 545, + 488 + ], + "type": "inline_equation", + "content": "\\gamma^{(l + 1)}" + }, + { + "bbox": [ + 304, + 392, + 545, + 488 + ], + "type": "text", + "content": " at the " + }, + { + "bbox": [ + 304, + 392, + 545, + 488 + ], + "type": "inline_equation", + "content": "l" + }, + { + "bbox": [ + 304, + 392, + 545, + 488 + ], + "type": "text", + "content": "-th iteration as shown:" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 333, + 487, + 545, + 505 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 333, + 487, + 545, + 505 + ], + "spans": [ + { + "bbox": [ + 333, + 487, + 545, + 505 + ], + "type": "interline_equation", + "content": "\\gamma_ {i j} ^ {(l + 1)} = \\max (0, [ \\psi_ {i} + \\phi_ {j} + \\epsilon \\gamma_ {i j} ^ {(l)} - C _ {i j} ] / \\epsilon). \\tag {8}", + "image_path": "e4ede9a7f4f84e15e4327e3c194f57391f91ea35372daf950508abaf5263e47e.jpg" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 304, + 504, + 545, + 576 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 504, + 545, + 576 + ], + "spans": [ + { + "bbox": [ + 304, + 504, + 545, + 576 + ], + "type": "text", + "content": "After we obtain the optimal solution on " + }, + { + "bbox": [ + 304, + 504, + 545, + 576 + ], + "type": "inline_equation", + "content": "\\gamma^{*}" + }, + { + "bbox": [ + 304, + 504, + 545, + 576 + ], + "type": "text", + "content": " via solving sparse sample-prototype matching, we can reach the reliable pseudo labels for these target samples " + }, + { + "bbox": [ + 304, + 504, + 545, + 576 + ], + "type": "inline_equation", + "content": "\\pmb{x}_{\\mathrm{kno}}^t" + }, + { + "bbox": [ + 304, + 504, + 545, + 576 + ], + "type": "text", + "content": ". Furthermore, the optimal solution " + }, + { + "bbox": [ + 304, + 504, + 545, + 576 + ], + "type": "inline_equation", + "content": "\\gamma^{*}" + }, + { + "bbox": [ + 304, + 504, + 545, + 576 + ], + "type": "text", + "content": " can yield more sparser and clearer results compared to using only the " + }, + { + "bbox": [ + 304, + 504, + 545, + 576 + ], + "type": "inline_equation", + "content": "\\ell_{2}" + }, + { + "bbox": [ + 304, + 504, + 545, + 576 + ], + "type": "text", + "content": "-norm regularization, as depicted in Fig. 4(b)-(c)." + } + ] + } + ], + "index": 10 + }, + { + "type": "image", + "bbox": [ + 310, + 578, + 541, + 647 + ], + "blocks": [ + { + "bbox": [ + 310, + 578, + 541, + 647 + ], + "lines": [ + { + "bbox": [ + 310, + 578, + 541, + 647 + ], + "spans": [ + { + "bbox": [ + 310, + 578, + 541, + 647 + ], + "type": "image", + "image_path": "d00b1878bcffeb0e669ff1bb5730968ed33c80b86d2e812e31e0499e4dd7fc2a.jpg" + } + ] + } + ], + "index": 11, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 304, + 650, + 545, + 706 + ], + "lines": [ + { + "bbox": [ + 304, + 650, + 545, + 706 + ], + "spans": [ + { + "bbox": [ + 304, + 650, + 545, + 706 + ], + "type": "text", + "content": "Figure 4. The illustration on Sparse Sample-Label Matching (SSLM) in sparse label assignment with toy examples. Sinkhorn and tradition " + }, + { + "bbox": [ + 304, + 650, + 545, + 706 + ], + "type": "inline_equation", + "content": "\\ell_2" + }, + { + "bbox": [ + 304, + 650, + 545, + 706 + ], + "type": "text", + "content": "-norm solvers could easily reach smooth but dense output solutions. Our proposed SSLM method can reach more sparse matching solution for better performance." + } + ] + } + ], + "index": 12, + "angle": 0, + "type": "image_caption" + } + ], + "index": 11 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 748, + 314, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 314, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 314, + 757 + ], + "type": "text", + "content": "4931" + } + ] + } + ], + "index": 13 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 4 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 51, + 72, + 165, + 151 + ], + "blocks": [ + { + "bbox": [ + 51, + 72, + 165, + 151 + ], + "lines": [ + { + "bbox": [ + 51, + 72, + 165, + 151 + ], + "spans": [ + { + "bbox": [ + 51, + 72, + 165, + 151 + ], + "type": "image", + "image_path": "a1db6d8aa1371d412013e8092b8e25829091aeffb96cbfa6600d46b79936f023.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 69, + 153, + 143, + 163 + ], + "lines": [ + { + "bbox": [ + 69, + 153, + 143, + 163 + ], + "spans": [ + { + "bbox": [ + 69, + 153, + 143, + 163 + ], + "type": "text", + "content": "(a) JPOT on Office-31" + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "type": "image", + "bbox": [ + 172, + 72, + 286, + 150 + ], + "blocks": [ + { + "bbox": [ + 172, + 72, + 286, + 150 + ], + "lines": [ + { + "bbox": [ + 172, + 72, + 286, + 150 + ], + "spans": [ + { + "bbox": [ + 172, + 72, + 286, + 150 + ], + "type": "image", + "image_path": "2511ba83b85b9efae26ba587c5599bffc615004fe1de4cf38675dd87e067181b.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 194, + 154, + 263, + 163 + ], + "lines": [ + { + "bbox": [ + 194, + 154, + 263, + 163 + ], + "spans": [ + { + "bbox": [ + 194, + 154, + 263, + 163 + ], + "type": "text", + "content": "(b) DTE on Office-31" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 46, + 167, + 287, + 201 + ], + "lines": [ + { + "bbox": [ + 46, + 167, + 287, + 201 + ], + "spans": [ + { + "bbox": [ + 46, + 167, + 287, + 201 + ], + "type": "text", + "content": "Figure 5. The t-SNE visualization of " + }, + { + "bbox": [ + 46, + 167, + 287, + 201 + ], + "type": "inline_equation", + "content": "\\mathbf{W} \\rightarrow \\mathbf{A}" + }, + { + "bbox": [ + 46, + 167, + 287, + 201 + ], + "type": "text", + "content": " tasks on Office-31. Each color denotes a class and the reddish-brown color is used to represent the unknown class." + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_caption" + } + ], + "index": 2 + }, + { + "bbox": [ + 47, + 209, + 149, + 222 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 209, + 149, + 222 + ], + "spans": [ + { + "bbox": [ + 47, + 209, + 149, + 222 + ], + "type": "text", + "content": "3.3. Putting Together" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 46, + 228, + 287, + 287 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 228, + 287, + 287 + ], + "spans": [ + { + "bbox": [ + 46, + 228, + 287, + 287 + ], + "type": "text", + "content": "We have first distinguished the target samples with unknown classes via weight barcode estimation, then assigning the rest target samples to the corresponding classes via sparse sample-prototype matching. Therefore, we should gather the output prediction of these target samples as " + }, + { + "bbox": [ + 46, + 228, + 287, + 287 + ], + "type": "inline_equation", + "content": "\\pmb{p}^t" + }, + { + "bbox": [ + 46, + 228, + 287, + 287 + ], + "type": "text", + "content": ":" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 51, + 289, + 287, + 313 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 51, + 289, + 287, + 313 + ], + "spans": [ + { + "bbox": [ + 51, + 289, + 287, + 313 + ], + "type": "interline_equation", + "content": "\\boldsymbol {p} ^ {t} = \\operatorname {s o f t m a x} \\left[ \\mathcal {G} \\left(\\mathcal {F} _ {t} \\left(\\boldsymbol {x} ^ {t}\\right)\\right) \\oplus \\widehat {\\mathcal {G}} \\left(\\mathcal {F} _ {t} \\left(\\boldsymbol {x} ^ {t}\\right)\\right) \\right] \\in \\mathbb {R} ^ {N \\times (H + 1)}. \\tag {9}", + "image_path": "fc39a2ef6952c9cdbe7251ef81c5686977b70b19793ae20824c544ceb3887de8.jpg" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 46, + 314, + 287, + 388 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 314, + 287, + 388 + ], + "spans": [ + { + "bbox": [ + 46, + 314, + 287, + 388 + ], + "type": "text", + "content": "Note that " + }, + { + "bbox": [ + 46, + 314, + 287, + 388 + ], + "type": "inline_equation", + "content": "\\widehat{\\mathcal{G}} (\\cdot)" + }, + { + "bbox": [ + 46, + 314, + 287, + 388 + ], + "type": "text", + "content": " is a newly add trainable target classifier to predict whether samples are unknown. Here we concatenate the output predictions on both source classifier " + }, + { + "bbox": [ + 46, + 314, + 287, + 388 + ], + "type": "inline_equation", + "content": "\\mathcal{G}(\\cdot)" + }, + { + "bbox": [ + 46, + 314, + 287, + 388 + ], + "type": "text", + "content": " and target classifier " + }, + { + "bbox": [ + 46, + 314, + 287, + 388 + ], + "type": "inline_equation", + "content": "\\widehat{\\mathcal{G}} (\\cdot)" + }, + { + "bbox": [ + 46, + 314, + 287, + 388 + ], + "type": "text", + "content": " then process the softmax operation. Meanwhile we can formulate the estimated \"ground-truth\" " + }, + { + "bbox": [ + 46, + 314, + 287, + 388 + ], + "type": "inline_equation", + "content": "q_{\\mathrm{kno}}^{t}" + }, + { + "bbox": [ + 46, + 314, + 287, + 388 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 46, + 314, + 287, + 388 + ], + "type": "inline_equation", + "content": "q_{\\mathrm{unk}}^{t}" + }, + { + "bbox": [ + 46, + 314, + 287, + 388 + ], + "type": "text", + "content": " for the target samples as follows:" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 97, + 388, + 287, + 409 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 97, + 388, + 287, + 409 + ], + "spans": [ + { + "bbox": [ + 97, + 388, + 287, + 409 + ], + "type": "interline_equation", + "content": "\\boldsymbol {q} _ {\\mathrm {k n o}} ^ {t} = \\left[ \\gamma^ {*} \\oplus 0 \\right] \\in \\mathbb {R} ^ {N _ {\\mathrm {k n o}} \\times (H + 1)}, \\tag {10}", + "image_path": "23c2e311b749f8aab9070cbde1f1a6e57ab76baeb9aa1bf9f0311ca02906edc0.jpg" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 99, + 407, + 230, + 420 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 99, + 407, + 230, + 420 + ], + "spans": [ + { + "bbox": [ + 99, + 407, + 230, + 420 + ], + "type": "interline_equation", + "content": "\\boldsymbol {q} _ {\\mathrm {u n k}} ^ {t} = [ \\mathbf {0} \\oplus 1 ] \\in \\mathbb {R} ^ {N _ {\\mathrm {u n k}} \\times (H + 1)}.", + "image_path": "3fe9cc656822121d96138f6fde356a8b83fc6c2ec2cb3a468bffad7b888410f7.jpg" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 46, + 422, + 287, + 459 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 422, + 287, + 459 + ], + "spans": [ + { + "bbox": [ + 46, + 422, + 287, + 459 + ], + "type": "text", + "content": "Then we can utilize the cross-entropy loss " + }, + { + "bbox": [ + 46, + 422, + 287, + 459 + ], + "type": "inline_equation", + "content": "L_{\\mathrm{cls}}" + }, + { + "bbox": [ + 46, + 422, + 287, + 459 + ], + "type": "text", + "content": " on fine-tuning the target feature extractor " + }, + { + "bbox": [ + 46, + 422, + 287, + 459 + ], + "type": "inline_equation", + "content": "\\mathcal{F}_t(\\cdot)" + }, + { + "bbox": [ + 46, + 422, + 287, + 459 + ], + "type": "text", + "content": " based on " + }, + { + "bbox": [ + 46, + 422, + 287, + 459 + ], + "type": "inline_equation", + "content": "\\mathcal{F}_s(\\cdot)" + }, + { + "bbox": [ + 46, + 422, + 287, + 459 + ], + "type": "text", + "content": " and training distinguisher " + }, + { + "bbox": [ + 46, + 422, + 287, + 459 + ], + "type": "inline_equation", + "content": "\\widehat{\\mathcal{G}}_t(\\cdot)" + }, + { + "bbox": [ + 46, + 422, + 287, + 459 + ], + "type": "text", + "content": " on the target samples as:" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 63, + 460, + 287, + 494 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 63, + 460, + 287, + 494 + ], + "spans": [ + { + "bbox": [ + 63, + 460, + 287, + 494 + ], + "type": "interline_equation", + "content": "\\min _ {\\mathcal {F} _ {t} (\\cdot), \\widehat {\\mathcal {G}} (\\cdot)} L _ {\\mathrm {c l s}} = - \\sum_ {i = 1} ^ {N _ {\\mathrm {u n k}} + N _ {\\mathrm {k n o}}} \\sum_ {j = 1} ^ {H + 1} q _ {i j} ^ {t} \\log p _ {i j} ^ {t}. \\tag {11}", + "image_path": "8cc9aca005a49a564909e033cb59d2ba942bef31da58356cf66c2ddc4b43bd57.jpg" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 46, + 495, + 287, + 555 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 495, + 287, + 555 + ], + "spans": [ + { + "bbox": [ + 46, + 495, + 287, + 555 + ], + "type": "text", + "content": "Note that we fix the source classifier " + }, + { + "bbox": [ + 46, + 495, + 287, + 555 + ], + "type": "inline_equation", + "content": "\\mathcal{G}(\\cdot)" + }, + { + "bbox": [ + 46, + 495, + 287, + 555 + ], + "type": "text", + "content": " during the training procedure. Meanwhile, the loss function is easy to implemented for end-to-end training. Moreover, we should update the class prototype " + }, + { + "bbox": [ + 46, + 495, + 287, + 555 + ], + "type": "inline_equation", + "content": "\\pmb{\\mu}^{(o + 1)}" + }, + { + "bbox": [ + 46, + 495, + 287, + 555 + ], + "type": "text", + "content": " at the " + }, + { + "bbox": [ + 46, + 495, + 287, + 555 + ], + "type": "inline_equation", + "content": "(o + 1)" + }, + { + "bbox": [ + 46, + 495, + 287, + 555 + ], + "type": "text", + "content": "-th iteration via incremental learning strategy:" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 55, + 557, + 287, + 590 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 557, + 287, + 590 + ], + "spans": [ + { + "bbox": [ + 55, + 557, + 287, + 590 + ], + "type": "interline_equation", + "content": "\\boldsymbol {\\mu} _ {j} ^ {(o + 1)} = \\alpha \\boldsymbol {\\mu} _ {j} ^ {(o)} + \\frac {1 - \\alpha}{\\sum_ {i = 1} ^ {N _ {\\mathrm {k n o}}} p _ {i j} ^ {t}} \\sum_ {i = 1} ^ {N _ {\\mathrm {k n o}}} \\mathcal {F} _ {t} \\left(\\boldsymbol {x} _ {\\mathrm {k n o}, i} ^ {t}\\right) p _ {i j} ^ {t}, \\tag {12}", + "image_path": "4f2843f158b3cf48245ec05208b4e9ab0cfcab9b078a0f886721fcecc008532a.jpg" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 46, + 591, + 287, + 615 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 591, + 287, + 615 + ], + "spans": [ + { + "bbox": [ + 46, + 591, + 287, + 615 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 46, + 591, + 287, + 615 + ], + "type": "inline_equation", + "content": "\\alpha \\in [0,1]" + }, + { + "bbox": [ + 46, + 591, + 287, + 615 + ], + "type": "text", + "content": " denotes momentum hyper parameter. The whole algorithm of DTE in provided in Alg.1." + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 47, + 627, + 123, + 640 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 627, + 123, + 640 + ], + "spans": [ + { + "bbox": [ + 47, + 627, + 123, + 640 + ], + "type": "text", + "content": "4. Experiment" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 47, + 647, + 97, + 659 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 647, + 97, + 659 + ], + "spans": [ + { + "bbox": [ + 47, + 647, + 97, + 659 + ], + "type": "text", + "content": "4.1. Setup" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 46, + 665, + 287, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 665, + 287, + 714 + ], + "spans": [ + { + "bbox": [ + 46, + 665, + 287, + 714 + ], + "type": "text", + "content": "We conduct the experiments on several real-world datasets, e.g., Office-31 [77] with following three domains: Amazon (A), Webcam (W) and DSLR (D); Office-Home [84] with following four dissimilar domains: Artistic images (Ar)," + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 304, + 72, + 545, + 132 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 72, + 545, + 132 + ], + "spans": [ + { + "bbox": [ + 304, + 72, + 545, + 132 + ], + "type": "text", + "content": "Clip Art (CI), Product images (Pr) and Real-World (Rw); VisDA-2017 [70] with domains Synthetic, Real; and Digits with following three domains MNIST [40], USPS [31] and SVHN [67]. We adopt the same set of known and unknown classes following previous works [46, 78]." + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 304, + 133, + 546, + 553 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 133, + 546, + 553 + ], + "spans": [ + { + "bbox": [ + 304, + 133, + 546, + 553 + ], + "type": "text", + "content": "Implementation Details. We first set batch size " + }, + { + "bbox": [ + 304, + 133, + 546, + 553 + ], + "type": "inline_equation", + "content": "N = 256" + }, + { + "bbox": [ + 304, + 133, + 546, + 553 + ], + "type": "text", + "content": " during the training procedure. For the experiments on Office-31, Office-Home, and VisDA, we first fine-tune on the backbone network ResNet-50 [29] in the source domain. For the experiments on Digits, we adopt LeNet [40] to extract the sample features. After the training in the source domain, we only transfer the well-trained feature extractor " + }, + { + "bbox": [ + 304, + 133, + 546, + 553 + ], + "type": "inline_equation", + "content": "\\mathcal{F}_s(\\cdot)" + }, + { + "bbox": [ + 304, + 133, + 546, + 553 + ], + "type": "text", + "content": " and classifier " + }, + { + "bbox": [ + 304, + 133, + 546, + 553 + ], + "type": "inline_equation", + "content": "\\mathcal{G}(\\cdot)" + }, + { + "bbox": [ + 304, + 133, + 546, + 553 + ], + "type": "text", + "content": " to the target domain. We set " + }, + { + "bbox": [ + 304, + 133, + 546, + 553 + ], + "type": "inline_equation", + "content": "\\epsilon = 0.5" + }, + { + "bbox": [ + 304, + 133, + 546, + 553 + ], + "type": "text", + "content": " in sparse label assignment. Meanwhile we set " + }, + { + "bbox": [ + 304, + 133, + 546, + 553 + ], + "type": "inline_equation", + "content": "\\alpha = 0.8" + }, + { + "bbox": [ + 304, + 133, + 546, + 553 + ], + "type": "text", + "content": " for the incremental learning on class prototypes. We choose Adam [35] as the optimizer and set the learning rate as 0.001 and 0.01 for fine-tuning target feature extractor " + }, + { + "bbox": [ + 304, + 133, + 546, + 553 + ], + "type": "inline_equation", + "content": "\\mathcal{F}_t(\\cdot)" + }, + { + "bbox": [ + 304, + 133, + 546, + 553 + ], + "type": "text", + "content": " based on " + }, + { + "bbox": [ + 304, + 133, + 546, + 553 + ], + "type": "inline_equation", + "content": "\\mathcal{F}_s(\\cdot)" + }, + { + "bbox": [ + 304, + 133, + 546, + 553 + ], + "type": "text", + "content": " and training distinguisher " + }, + { + "bbox": [ + 304, + 133, + 546, + 553 + ], + "type": "inline_equation", + "content": "\\widehat{\\mathcal{G}} (\\cdot)" + }, + { + "bbox": [ + 304, + 133, + 546, + 553 + ], + "type": "text", + "content": " respectively. We employ these four evaluation metrics, i.e., OS, OS*, UNK, HOS and for open set domain adaptation following [78]. Specifically, OS and OS* denotes the accuracy averaged over all classes or only known classes respectively. UNK denotes the accuracy of unknown instances and HOS denotes the harmonic value of OS* and UNK. We compare the proposed DTE with state-of-the-art domain adaptation methods: ResNet [29], Domain Adversarial Neural Network (DANN) [2], Deep Joint Distribution Optimal Transport (DeepJDOT) [11], OpenMAX [5], MMD+OSVM, DANN+OSVM, Openset Back-Propagation (OSBP) [78], Separate To Adapt (STA) [46], Towards Inheritable Models (TIM) [37], Joint Partial Optimal Transport (JPOT) [91], Source Hypothesis Transfer (SHOT) [44], Progressive Graph Learning (PGL) [63], Source-Free Progressive Graph Learning (SFPGL) [64], Unknown-Aware Domain Adversarial Learning (cUADAL) [32], Self-supervised Class-Disccovering Adapter (SCDA) [97], Unknown Oriented Learning (UOL) [47], Adjustment and Alignment (ANNA) [43] and Balanced Correlation Learning (BCL) [30]. We perform five random experiments and report the average and variation results." + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 306, + 563, + 430, + 574 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 563, + 430, + 574 + ], + "spans": [ + { + "bbox": [ + 306, + 563, + 430, + 574 + ], + "type": "text", + "content": "4.2. Result and Discussion" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 304, + 582, + 545, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 582, + 545, + 713 + ], + "spans": [ + { + "bbox": [ + 304, + 582, + 545, + 713 + ], + "type": "text", + "content": "The comparison results on Office-31 are shown in Table. 1. Directly applying ResNet could lead to worst results since the existence of domain discrepancy. Although previous method (e.g., OSBP, STA) adopted domain adversarial strategy to reduce the discrepancy, it could be unstable and hard to train in the real practice [81]. Our proposed DTE achieves the best performance among other approaches. Even for the difficult task (e.g., " + }, + { + "bbox": [ + 304, + 582, + 545, + 713 + ], + "type": "inline_equation", + "content": "\\mathbf{A} \\rightarrow \\mathbf{W}" + }, + { + "bbox": [ + 304, + 582, + 545, + 713 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 304, + 582, + 545, + 713 + ], + "type": "inline_equation", + "content": "\\mathbf{W} \\rightarrow \\mathbf{A}" + }, + { + "bbox": [ + 304, + 582, + 545, + 713 + ], + "type": "text", + "content": ") where these images are quite different across domains, DTE still obtains satisfactory results. It indicates the efficacy of the proposed DTE for dealing with open set domain" + } + ] + } + ], + "index": 22 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "text", + "content": "4932" + } + ] + } + ], + "index": 23 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 5 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 65, + 86, + 525, + 247 + ], + "blocks": [ + { + "bbox": [ + 125, + 72, + 467, + 84 + ], + "lines": [ + { + "bbox": [ + 125, + 72, + 467, + 84 + ], + "spans": [ + { + "bbox": [ + 125, + 72, + 467, + 84 + ], + "type": "text", + "content": "Table 1. Classification accuracy (%) on Office-31 for open set domain adaptation (ResNet-50)." + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 65, + 86, + 525, + 247 + ], + "lines": [ + { + "bbox": [ + 65, + 86, + 525, + 247 + ], + "spans": [ + { + "bbox": [ + 65, + 86, + 525, + 247 + ], + "type": "table", + "html": "
MethodA→WA→DD→WW→DD→AW→AAvg
OSOS*OSOS*OSOS*OSOS*OSOS*OSOS*OSOS*
ResNet82.5±1.282.7±0.985.2±0.385.5±0.994.1±0.394.3±0.796.6±0.297.0±0.471.6±1.071.5±1.175.5±1.075.2±1.684.284.4
DANN85.3±0.787.7±1.186.5±0.687.7±0.697.5±0.298.3±0.599.5±0.1100.0±075.7±1.676.2±0.974.9±1.275.6±0.886.687.6
DeepJDOT86.1±0.588.7±0.986.9±0.789.0±0.596.9±0.295.6±0.896.1±0.198.1±0.585.6±1.288.4±1.081.5±1.383.1±0.988.990.5
OpenMax87.4±0.587.5±0.387.1±0.988.4±0.996.1±0.496.2±0.398.4±0.398.5±0.383.4±1.082.1±0.682.8±0.982.8±0.689.089.3
OSBP86.5±2.087.6±2.188.6±1.489.2±1.397.0±1.096.5±0.497.9±0.998.7±0.688.9±2.590.6±2.385.8±2.584.9±1.390.891.3
STA89.5±0.692.1±0.593.7±1.596.1±0.497.5±0.296.5±0.599.5±0.299.6±0.189.1±0.593.5±0.887.9±0.987.4±0.692.994.1
TIM91.3±0.793.2±0.594.2±1.197.1±0.896.5±0.597.4±0.799.5±0.299.4±0.390.1±0.291.5±0.288.7±1.388.4±0.993.494.5
JPOT92.8±0.692.2±0.495.2±0.996.0±0.698.1±0.396.2±0.499.5±0.198.6±0.293.0±0.794.1±0.489.9±1.088.4±0.494.694.4
SHOT88.8±0.791.4±0.490.3±0.592.6±0.396.2±0.497.0±0.497.4±0.397.9±0.591.6±0.593.4±0.791.2±0.593.5±0.492.694.3
PGL89.2±0.590.1±0.789.6±0.891.6±0.595.3±0.495.1±0.596.7±0.597.6±0.671.0±1.372.0±1.873.0±1.277.6±1.685.887.4
SFPGL90.4±0.391.9±0.791.0±1.091.2±0.897.3±0.595.5±0.696.1±0.498.4±0.474.8±1.575.6±2.074.9±1.679.6±1.187.488.7
cUADAL92.5±0.493.7±0.696.0±0.797.5±0.698.4±0.4100.0±098.6±0.5100.0±085.3±0.886.0±0.583.2±0.783.9±0.992.393.5
SCDA95.7±0.197.5±0.295.9±0.496.5±0.499.2±0.499.7±0.199.8±0.1100.0±092.1±0.193.7±0.392.2±0.193.6±0.195.896.8
UOL94.6±0.396.8±0.695.4±0.597.2±0.799.5±0.698.2±0.398.9±0.299.1±0.392.7±0.693.2±0.891.6±1.093.0±0.895.596.2
DTE-W93.9±0.694.5±0.595.7±0.696.3±0.798.2±0.599.1±0.297.4±0.498.8±0.694.0±0.993.6±1.290.7±0.891.5±0.794.995.6
DTE-S95.3±0.596.9±0.496.5±0.697.0±0.898.7±0.799.4±0.398.1±0.499.2±0.594.8±1.193.9±1.092.5±0.794.1±0.995.996.7
DTE96.1±0.297.5±0.496.8±0.497.9±0.598.9±0.5100.0±098.5±0.399.4±0.395.6±0.594.3±0.692.9±0.494.6±0.396.497.2
", + "image_path": "6bdcc07379bb6249dbd5b1849878b5f26266a5d19d9a0c35627c74060bc7567b.jpg" + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "table_body" + } + ], + "index": 1 + }, + { + "type": "table", + "bbox": [ + 65, + 262, + 525, + 435 + ], + "blocks": [ + { + "bbox": [ + 110, + 249, + 482, + 260 + ], + "lines": [ + { + "bbox": [ + 110, + 249, + 482, + 260 + ], + "spans": [ + { + "bbox": [ + 110, + 249, + 482, + 260 + ], + "type": "text", + "content": "Table 2. Classification accuracy OS* (%) on Office-Home for open set domain adaptation (ResNet-50)." + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 65, + 262, + 525, + 435 + ], + "lines": [ + { + "bbox": [ + 65, + 262, + 525, + 435 + ], + "spans": [ + { + "bbox": [ + 65, + 262, + 525, + 435 + ], + "type": "table", + "html": "
MethodAr→ClPr→ClRw→ClAr→PrCl→PrRw→PrCl→ArPr→ArRw→ArAr→RwCl→RwPr→RwAvg
ResNet53.4±0.452.7±0.651.9±0.569.3±0.761.8±0.574.1±0.461.4±0.664.0±0.370.0±0.378.7±0.671.0±0.674.9±0.965.3
DANN54.6±0.749.7±1.651.9±1.469.5±1.163.5±1.072.9±0.861.9±1.263.3±1.071.3±1.080.2±0.871.7±0.474.2±0.465.4
OpenMax56.5±0.452.9±0.753.7±0.469.1±0.364.8±0.474.5±0.664.1±0.964.0±0.871.2±0.880.3±0.873.0±0.576.9±0.366.7
OSBP56.7±1.951.5±2.149.2±2.467.5±1.565.5±1.574.0±1.562.5±2.064.8±1.169.3±1.180.6±0.974.7±2.271.5±1.965.7
STA58.1±0.653.1±0.954.4±1.071.6±1.269.3±1.081.9±0.563.4±0.565.2±0.874.9±1.085.0±0.275.8±0.480.8±0.369.5
TIM60.1±0.754.2±1.056.2±1.770.9±1.470.0±1.778.6±0.664.0±0.666.1±1.374.9±0.983.2±0.975.7±1.381.3±1.469.6
JPOT59.6±0.554.2±0.754.6±0.972.3±1.170.1±0.682.1±0.962.9±0.768.3±0.875.1±1.184.8±0.477.4±0.581.2±0.470.2
SHOT64.5±0.859.3±0.564.6±0.880.4±0.975.4±1.282.3±1.063.1±0.865.3±0.669.6±0.784.7±0.681.2±1.083.3±0.872.8
PGL61.6±1.058.4±0.865.0±0.777.1±1.572.0±1.383.0±0.968.8±1.472.2±0.778.6±1.385.9±0.982.8±1.182.6±1.274.0
SFPGL62.4±0.559.0±0.764.1±0.977.8±0.772.7±0.884.2±1.070.5±0.673.4±0.578.3±0.886.8±0.683.9±0.881.4±0.774.5
cUADAL62.3±0.660.1±0.861.7±0.277.9±0.671.4±0.981.3±0.572.3±0.671.5±0.873.9±1.083.6±0.783.5±0.784.0±1.073.6
SCDA59.9±0.359.0±0.362.8±0.579.6±0.473.8±1.083.7±0.870.9±0.572.3±0.675.5±0.485.3±0.682.9±0.385.7±0.974.3
UOL61.0±0.759.5±0.663.3±0.680.7±1.074.9±0.885.8±0.771.6±0.972.9±0.675.1±0.588.4±0.883.2±0.486.3±0.974.4
ANNA61.3±0.960.4±0.762.5±0.882.1±0.673.6±0.784.4±0.570.2±1.171.6±0.975.7±0.686.9±0.582.0±1.286.9±1.574.8
BCL58.4±0.559.7±0.661.1±0.680.6±0.573.0±0.985.1±0.470.5±0.772.2±0.674.8±0.587.5±0.683.4±0.986.1±1.274.4
DTE-W60.1±0.957.2±0.859.4±0.978.5±0.672.3±1.083.0±0.669.1±0.871.9±0.775.8±1.186.7±0.581.6±0.682.8±1.573.2
DTE-S62.1±0.660.9±0.861.8±0.880.4±0.574.2±1.285.7±0.971.6±0.973.3±1.076.9±0.788.3±0.882.0±0.886.5±1.376.1
DTE63.2±0.462.6±0.663.9±0.581.5±0.575.4±0.886.8±0.672.0±0.574.1±0.477.3±0.790.2±0.684.3±0.487.5±1.176.5
", + "image_path": "8e8af3547d49e917071acfa34626f867ff519dc485777f6e724f18430ef33d63.jpg" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "table_body" + } + ], + "index": 3 + }, + { + "type": "table", + "bbox": [ + 65, + 449, + 525, + 546 + ], + "blocks": [ + { + "bbox": [ + 89, + 436, + 503, + 447 + ], + "lines": [ + { + "bbox": [ + 89, + 436, + 503, + 447 + ], + "spans": [ + { + "bbox": [ + 89, + 436, + 503, + 447 + ], + "type": "text", + "content": "Table 3. Classification accuracy (%) on Digits (LeNet) and VisDA2017 (VGGNet) for open set domain adaptation." + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 65, + 449, + 525, + 546 + ], + "lines": [ + { + "bbox": [ + 65, + 449, + 525, + 546 + ], + "spans": [ + { + "bbox": [ + 65, + 449, + 525, + 546 + ], + "type": "table", + "html": "
MethodSVHN→MNISTUSPS→MNISTMNIST→USPSAvgVisDA2017
OSOS*HOSUNKOSOS*HOSUNKOSOS*HOSUNKOSOS*HOSUNKbicyclebuscarmotorcycletraintruckOSOS*
OSVM54.363.117.710.543.132.348.597.179.877.983.289.359.157.749.865.731.751.666.570.488.520.852.554.9
MMD+OSVM55.964.720.112.262.858.968.782.380.079.880.481.066.268.856.458.439.050.164.279.986.616.354.456.0
DANN+OSVM62.975.31.780.9077.192.41.190.6033.835.529.525.360.469.410.88.9331.856.671.777.487.022.355.557.8
OSBP63.059.168.982.592.391.294.497.892.194.985.778.182.481.782.986.151.167.142.884.281.828.062.959.2
STA76.975.479.784.492.291.393.996.793.094.988.883.587.387.287.588.252.469.659.987.886.527.266.863.9
TIM80.479.083.087.495.495.196.096.993.093.591.890.589.689.390.391.653.569.262.285.785.432.568.164.7
JPOT77.275.380.686.792.491.294.798.492.992.194.496.987.585.489.994.055.370.064.189.587.233.769.566.6
UOL84.484.185.085.996.997.196.595.995.395.195.796.392.291.692.492.754.670.563.791.286.931.470.066.3
DTE-W83.182.883.784.696.596.496.797.094.594.295.196.090.391.491.892.556.070.965.688.884.132.269.966.2
DTE-S84.884.585.486.397.096.997.297.594.994.695.596.492.592.292.793.456.571.266.390.187.733.871.367.6
DTE85.585.286.187.097.697.597.898.195.695.496.096.692.992.793.393.957.472.067.990.888.635.272.468.7
", + "image_path": "b175c5327f4cdec059d118fe970a7c322ac91602a9da79f368e9b3c9c066ea98.jpg" + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "table_body" + } + ], + "index": 5 + }, + { + "bbox": [ + 47, + 549, + 238, + 561 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 549, + 238, + 561 + ], + "spans": [ + { + "bbox": [ + 47, + 549, + 238, + 561 + ], + "type": "text", + "content": "adaptation even under the source-free scenario." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 46, + 570, + 288, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 570, + 288, + 714 + ], + "spans": [ + { + "bbox": [ + 46, + 570, + 288, + 714 + ], + "type": "text", + "content": "Meanwhile the results of Office-Home are shown in Table. 2. The tasks on Office-Home are more difficult than Office-31 since it has more categories with large domain gap. Previous methods cannot better reduce domain discrepancy while involving the negative transfer, leading to the limited model performance. Meanwhile our proposed DTE still outperforms the comparison methods. Note that the proposed DTE with weight barcode estimation can fully leverage knowledge from source domain to a significant degree compared with JPOT. Moreover, we conduct the experiments on VisDA-2017 and Digits and report the results in Table. 3. It is remarkable that our proposed DTE can pro" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 304, + 549, + 545, + 658 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 549, + 545, + 658 + ], + "spans": [ + { + "bbox": [ + 304, + 549, + 545, + 658 + ], + "type": "text", + "content": "mote the better results of OS and " + }, + { + "bbox": [ + 304, + 549, + 545, + 658 + ], + "type": "inline_equation", + "content": "\\mathbf{OS^{*}}" + }, + { + "bbox": [ + 304, + 549, + 545, + 658 + ], + "type": "text", + "content": " on hard transfer tasks in VisDA-2017. Specifically, in terms of the performance of all classes, DTE has achieved significant improvements compared to the previous models, with notable examples being a " + }, + { + "bbox": [ + 304, + 549, + 545, + 658 + ], + "type": "inline_equation", + "content": "2.27\\%" + }, + { + "bbox": [ + 304, + 549, + 545, + 658 + ], + "type": "text", + "content": " improvement on Office-Home and a " + }, + { + "bbox": [ + 304, + 549, + 545, + 658 + ], + "type": "inline_equation", + "content": "3.15\\%" + }, + { + "bbox": [ + 304, + 549, + 545, + 658 + ], + "type": "text", + "content": " improvement on VisDA-2017 against runner-up models for " + }, + { + "bbox": [ + 304, + 549, + 545, + 658 + ], + "type": "inline_equation", + "content": "\\mathbf{OS^{*}}" + }, + { + "bbox": [ + 304, + 549, + 545, + 658 + ], + "type": "text", + "content": ". DTE can better distinguish the unknown target samples meanwhile correctly assign pseudo labels for the rest target samples to reach state-of-the-art performance." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 306, + 669, + 371, + 681 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 669, + 371, + 681 + ], + "spans": [ + { + "bbox": [ + 306, + 669, + 371, + 681 + ], + "type": "text", + "content": "4.3. Analysis." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 305, + 689, + 545, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 689, + 545, + 715 + ], + "spans": [ + { + "bbox": [ + 305, + 689, + 545, + 715 + ], + "type": "text", + "content": "Ablation Study. To further study how does each module of DTE contribute on the final performance, we com" + } + ] + } + ], + "index": 10 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "text", + "content": "4933" + } + ] + } + ], + "index": 11 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 6 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 51, + 69, + 161, + 155 + ], + "blocks": [ + { + "bbox": [ + 51, + 69, + 161, + 155 + ], + "lines": [ + { + "bbox": [ + 51, + 69, + 161, + 155 + ], + "spans": [ + { + "bbox": [ + 51, + 69, + 161, + 155 + ], + "type": "image", + "image_path": "12bb83a3c026491d4597acb58d1810f3a2d186e007ce2ec0f30abdc33847d5f5.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 69, + 160, + 143, + 170 + ], + "lines": [ + { + "bbox": [ + 69, + 160, + 143, + 170 + ], + "spans": [ + { + "bbox": [ + 69, + 160, + 143, + 170 + ], + "type": "text", + "content": "(a) Openness on Digits" + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "type": "image", + "bbox": [ + 173, + 69, + 284, + 156 + ], + "blocks": [ + { + "bbox": [ + 173, + 69, + 284, + 156 + ], + "lines": [ + { + "bbox": [ + 173, + 69, + 284, + 156 + ], + "spans": [ + { + "bbox": [ + 173, + 69, + 284, + 156 + ], + "type": "image", + "image_path": "e298eb90407f6ba113085d004d381ae78a4d684283ddae5876075a7a37606652.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 186, + 160, + 272, + 169 + ], + "lines": [ + { + "bbox": [ + 186, + 160, + 272, + 169 + ], + "spans": [ + { + "bbox": [ + 186, + 160, + 272, + 169 + ], + "type": "text", + "content": "(b) Openness on Office-31" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_caption" + } + ], + "index": 2 + }, + { + "type": "image", + "bbox": [ + 50, + 194, + 164, + 280 + ], + "blocks": [ + { + "bbox": [ + 47, + 175, + 286, + 186 + ], + "lines": [ + { + "bbox": [ + 47, + 175, + 286, + 186 + ], + "spans": [ + { + "bbox": [ + 47, + 175, + 286, + 186 + ], + "type": "text", + "content": "Figure 6. The openness experiments of " + }, + { + "bbox": [ + 47, + 175, + 286, + 186 + ], + "type": "inline_equation", + "content": "\\mathbb{O}" + }, + { + "bbox": [ + 47, + 175, + 286, + 186 + ], + "type": "text", + "content": " on Digits and Office-31." + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 50, + 194, + 164, + 280 + ], + "lines": [ + { + "bbox": [ + 50, + 194, + 164, + 280 + ], + "spans": [ + { + "bbox": [ + 50, + 194, + 164, + 280 + ], + "type": "image", + "image_path": "034a113e41586505f4f06265bab5afe6f5b099ff5f274df858022a57c16179dd.jpg" + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 66, + 283, + 146, + 293 + ], + "lines": [ + { + "bbox": [ + 66, + 283, + 146, + 293 + ], + "spans": [ + { + "bbox": [ + 66, + 283, + 146, + 293 + ], + "type": "text", + "content": "(a) Parameters on Digits" + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 51, + 297, + 281, + 309 + ], + "lines": [ + { + "bbox": [ + 51, + 297, + 281, + 309 + ], + "spans": [ + { + "bbox": [ + 51, + 297, + 281, + 309 + ], + "type": "text", + "content": "Figure 7. Hyper parameter tuning of " + }, + { + "bbox": [ + 51, + 297, + 281, + 309 + ], + "type": "inline_equation", + "content": "\\epsilon" + }, + { + "bbox": [ + 51, + 297, + 281, + 309 + ], + "type": "text", + "content": " on Digits and Office-31." + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "image_caption" + } + ], + "index": 5 + }, + { + "type": "image", + "bbox": [ + 172, + 194, + 285, + 280 + ], + "blocks": [ + { + "bbox": [ + 172, + 194, + 285, + 280 + ], + "lines": [ + { + "bbox": [ + 172, + 194, + 285, + 280 + ], + "spans": [ + { + "bbox": [ + 172, + 194, + 285, + 280 + ], + "type": "image", + "image_path": "90f713d26bc0db61faf8be3c5f15ed126d9e287bd656092f2a170b1c25f1975b.jpg" + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 183, + 283, + 274, + 293 + ], + "lines": [ + { + "bbox": [ + 183, + 283, + 274, + 293 + ], + "spans": [ + { + "bbox": [ + 183, + 283, + 274, + 293 + ], + "type": "text", + "content": "(b) Parameters on Office-31" + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "image_caption" + } + ], + "index": 7 + }, + { + "bbox": [ + 47, + 317, + 286, + 533 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 317, + 286, + 533 + ], + "spans": [ + { + "bbox": [ + 47, + 317, + 286, + 533 + ], + "type": "text", + "content": "pare DTE with its several variants, including DTE-W and DTE-S. DTE-W only replaces the proposed weight barcode estimation with JPOT by calculating mean transport cost to determine the unknown target samples. DTE-S only replaces sparse label assignment with self-labelling strategy via Sinkhorn algorithm [3]. The comparison results on Office-Home and Office-31 are shown in Table. 2 and Table. 1. From it, we can observe that (1) Although DTE-W exceeds most of previous methods, it cannot precisely distinguish unknown data samples especially under source-free domain adaptation scenario. (2) DTE-S achieves better results than DTE-W, while it still cannot reach more satisfactory results than DTE. That is because involving entropy regularization term could reach relatively dense solution, thus leading to the inaccurate solutions. (3) DTE achieves the best performance which indicates the efficacy of incorporating weight barcode estimation and sparse label assignment for distinguishing then exploiting target samples." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 47, + 533, + 286, + 676 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 533, + 286, + 676 + ], + "spans": [ + { + "bbox": [ + 47, + 533, + 286, + 676 + ], + "type": "text", + "content": "Feature Visualization. We further provide the feature visualization via T-SNE [38] of the target sample representations " + }, + { + "bbox": [ + 47, + 533, + 286, + 676 + ], + "type": "inline_equation", + "content": "\\pmb{x}^t" + }, + { + "bbox": [ + 47, + 533, + 286, + 676 + ], + "type": "text", + "content": " by JPOT and DTE on " + }, + { + "bbox": [ + 47, + 533, + 286, + 676 + ], + "type": "inline_equation", + "content": "\\mathbf{W} \\rightarrow \\mathbf{A}" + }, + { + "bbox": [ + 47, + 533, + 286, + 676 + ], + "type": "text", + "content": " (Office-31). The results are depicted in Fig. 5(a)-(b), where each color represents a specific class, and the reddish-brown color is used to represent the unknown class. From there we can observe that JPOT cannot better distinguish known and unknown target samples with ambiguous decision boundary. The primary reason is that JPOT can only utilizes limited source knowledge for open set domain adaptation. Meanwhile DTE can better distinguish unknown target samples and cluster the rest known samples accordingly." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 47, + 677, + 286, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 677, + 286, + 713 + ], + "spans": [ + { + "bbox": [ + 47, + 677, + 286, + 713 + ], + "type": "text", + "content": "Openness. To verify the proposed DTE can be applied to the case with different proportion of unknown classes, we adjust the openness " + }, + { + "bbox": [ + 47, + 677, + 286, + 713 + ], + "type": "inline_equation", + "content": "\\mathbb{O}" + }, + { + "bbox": [ + 47, + 677, + 286, + 713 + ], + "type": "text", + "content": " in the openset domain adaptation." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 304, + 70, + 545, + 205 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 70, + 545, + 205 + ], + "spans": [ + { + "bbox": [ + 304, + 70, + 545, + 205 + ], + "type": "text", + "content": "Specifically, " + }, + { + "bbox": [ + 304, + 70, + 545, + 205 + ], + "type": "inline_equation", + "content": "\\mathbb{O} = 1 - \\frac{|\\mathcal{C}^s|}{|\\mathcal{C}^t|}" + }, + { + "bbox": [ + 304, + 70, + 545, + 205 + ], + "type": "text", + "content": " denotes different level of openness [46]. Larger openness indicates that fewer classes are known across domains. We vary " + }, + { + "bbox": [ + 304, + 70, + 545, + 205 + ], + "type": "inline_equation", + "content": "\\mathbb{O}" + }, + { + "bbox": [ + 304, + 70, + 545, + 205 + ], + "type": "text", + "content": " from 0 to nearly 1 on Office-31 and Digits accordingly [46] then report the result of OS in Fig. 6(a)-(b). We can observe that the majority of previous methods (e.g., DANN and JPOT) cannot better adjust to different proportion of unknown classes in the target domain. Meanwhile, the proposed DTE can always reach better performance even openness " + }, + { + "bbox": [ + 304, + 70, + 545, + 205 + ], + "type": "inline_equation", + "content": "\\mathbb{O}" + }, + { + "bbox": [ + 304, + 70, + 545, + 205 + ], + "type": "text", + "content": " approaches to 0 or 1. It further demonstrates the proposed DTE is robust to openness without the requirement of prior knowledge." + } + ] + } + ], + "index": 13 + }, + { + "type": "table", + "bbox": [ + 309, + 224, + 545, + 271 + ], + "blocks": [ + { + "bbox": [ + 314, + 211, + 537, + 221 + ], + "lines": [ + { + "bbox": [ + 314, + 211, + 537, + 221 + ], + "spans": [ + { + "bbox": [ + 314, + 211, + 537, + 221 + ], + "type": "text", + "content": "Table 4. The results (%) for method extension on VisDA2017" + } + ] + } + ], + "index": 14, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 309, + 224, + 545, + 271 + ], + "lines": [ + { + "bbox": [ + 309, + 224, + 545, + 271 + ], + "spans": [ + { + "bbox": [ + 309, + 224, + 545, + 271 + ], + "type": "table", + "html": "
MethodbicyclebuscarmotorcycletraintruckOSOS*
PGL93.593.875.798.896.238.580.782.8
SFPGL93.697.689.695.396.795.691.094.7
DTE(VGGNet+GNN)93.594.077.296.496.041.380.283.1
DTE(ResNet+GNN)93.897.990.197.996.592.091.694.7
", + "image_path": "dd44aadeaf5f36ab979fb2365c08bffceee330d4c5ae67bf6b9eb94032512fcc.jpg" + } + ] + } + ], + "index": 15, + "angle": 0, + "type": "table_body" + } + ], + "index": 15 + }, + { + "bbox": [ + 304, + 277, + 545, + 372 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 277, + 545, + 372 + ], + "spans": [ + { + "bbox": [ + 304, + 277, + 545, + 372 + ], + "type": "text", + "content": "Method Extension. Furthermore, we analyse the general extension of our proposed DTE. Specifically, we incorporate weight barcode estimation and sparse label assignment with VGGNet/ResNet and Graph Neural Network (GNN for transductive feature learning [64, 89]) on VisDA2017. The results of OS and " + }, + { + "bbox": [ + 304, + 277, + 545, + 372 + ], + "type": "inline_equation", + "content": "\\mathbf{OS^{*}}" + }, + { + "bbox": [ + 304, + 277, + 545, + 372 + ], + "type": "text", + "content": " are shown in Table 4. We observe that the proposed DTE can be applied with various backbones for achieving better performance." + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 304, + 373, + 545, + 481 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 373, + 545, + 481 + ], + "spans": [ + { + "bbox": [ + 304, + 373, + 545, + 481 + ], + "type": "text", + "content": "Hyper Parameter. We finally study the effects of hyperparameters on model performance. For DTE, we vary " + }, + { + "bbox": [ + 304, + 373, + 545, + 481 + ], + "type": "inline_equation", + "content": "\\epsilon" + }, + { + "bbox": [ + 304, + 373, + 545, + 481 + ], + "type": "text", + "content": " in " + }, + { + "bbox": [ + 304, + 373, + 545, + 481 + ], + "type": "inline_equation", + "content": "\\{0.01, 0.05, 0.1, 0.5, 1, 10\\}" + }, + { + "bbox": [ + 304, + 373, + 545, + 481 + ], + "type": "text", + "content": " for sparse label assignment on Digits/Office-31 and report the results of OS/OS* in Fig. 7(a)-(b). We can conclude that when " + }, + { + "bbox": [ + 304, + 373, + 545, + 481 + ], + "type": "inline_equation", + "content": "\\epsilon" + }, + { + "bbox": [ + 304, + 373, + 545, + 481 + ], + "type": "text", + "content": " is larger, the proximal regularization term will have a more important impact compared to sparse-label matching, which can result in poorer performance. Hence we set " + }, + { + "bbox": [ + 304, + 373, + 545, + 481 + ], + "type": "inline_equation", + "content": "\\epsilon = 0.5" + }, + { + "bbox": [ + 304, + 373, + 545, + 481 + ], + "type": "text", + "content": " in the experiments empirically." + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 306, + 491, + 378, + 503 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 491, + 378, + 503 + ], + "spans": [ + { + "bbox": [ + 306, + 491, + 378, + 503 + ], + "type": "text", + "content": "5. Conclusion" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 304, + 512, + 545, + 643 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 512, + 545, + 643 + ], + "spans": [ + { + "bbox": [ + 304, + 512, + 545, + 643 + ], + "type": "text", + "content": "In this paper, we propose Distinguish Then Exploit model (DTE) for tackling the source-free open set domain adaptation problem. DTE first adopts weight barcode estimation with proposed target sample reweighting optimization method on partial unbalanced optimal transport to determine the known and unknown target samples without prior information. Then DTE utilizes sparse sample-label matching on optimal transport to further leverage useful and accurate source knowledge in the target domain. We also conduct extensive experiments on different datasets to show the superior performance of the proposed DTE model." + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 306, + 657, + 410, + 670 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 657, + 410, + 670 + ], + "spans": [ + { + "bbox": [ + 306, + 657, + 410, + 670 + ], + "type": "text", + "content": "6. Acknowledgment" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 304, + 677, + 545, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 677, + 545, + 713 + ], + "spans": [ + { + "bbox": [ + 304, + 677, + 545, + 713 + ], + "type": "text", + "content": "This work was supported by the National Natural Science Foundation of China (No. 62372242) and Natural Science Foundation of Shandong Province (No. ZR2023MF007)." + } + ] + } + ], + "index": 21 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "text", + "content": "4934" + } + ] + } + ], + "index": 22 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 7 + }, + { + "para_blocks": [ + { + "bbox": [ + 48, + 71, + 106, + 83 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 71, + 106, + 83 + ], + "spans": [ + { + "bbox": [ + 48, + 71, + 106, + 83 + ], + "type": "text", + "content": "References" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 48, + 91, + 287, + 713 + ], + "type": "list", + "angle": 0, + "index": 16, + "blocks": [ + { + "bbox": [ + 53, + 91, + 286, + 112 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 91, + 286, + 112 + ], + "spans": [ + { + "bbox": [ + 53, + 91, + 286, + 112 + ], + "type": "text", + "content": "[1] Ravindra K Ahuja, Thomas L Magnanti, and James B Orlin. Network flows. 1988. 5" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 53, + 114, + 287, + 147 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 114, + 287, + 147 + ], + "spans": [ + { + "bbox": [ + 53, + 114, + 287, + 147 + ], + "type": "text", + "content": "[2] Hana Ajakan, Pascal Germain, Hugo Larochelle, François Laviolette, and Mario Marchand. Domain-adversarial neural networks. arXiv preprint arXiv:1412.4446, 2014. 2, 6" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 53, + 148, + 287, + 191 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 148, + 287, + 191 + ], + "spans": [ + { + "bbox": [ + 53, + 148, + 287, + 191 + ], + "type": "text", + "content": "[3] Yuki Markus Asano, Christian Rupprecht, and Andrea Vedaldi. Self-labelling via simultaneous clustering and representation learning. arXiv preprint arXiv:1911.05371, 2019. 5, 8" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 53, + 194, + 287, + 237 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 194, + 287, + 237 + ], + "spans": [ + { + "bbox": [ + 53, + 194, + 287, + 237 + ], + "type": "text", + "content": "[4] Yikun Bai, Bernhard Schmitzer, Matthew Thorpe, and Soheil Kolouri. Sliced optimal partial transport. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 13681-13690, 2023. 2" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 53, + 239, + 287, + 281 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 239, + 287, + 281 + ], + "spans": [ + { + "bbox": [ + 53, + 239, + 287, + 281 + ], + "type": "text", + "content": "[5] Abhijit Bendale and Terrance E Boult. Towards open set deep networks. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 1563-1572, 2016. 6" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 53, + 284, + 287, + 327 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 284, + 287, + 327 + ], + "spans": [ + { + "bbox": [ + 53, + 284, + 287, + 327 + ], + "type": "text", + "content": "[6] Mathieu Blondel, Vivien Seguy, and Antoine Rolet. Smooth and sparse optimal transport. In International conference on artificial intelligence and statistics, pages 880-889. PMLR, 2018. 5" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 53, + 329, + 287, + 361 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 329, + 287, + 361 + ], + "spans": [ + { + "bbox": [ + 53, + 329, + 287, + 361 + ], + "type": "text", + "content": "[7] Nicolas Bonneel and David Coeurjolly. Spot: sliced partial optimal transport. ACM Transactions on Graphics (TOG), 38(4):1-13, 2019. 2" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 53, + 364, + 287, + 407 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 364, + 287, + 407 + ], + "spans": [ + { + "bbox": [ + 53, + 364, + 287, + 407 + ], + "type": "text", + "content": "[8] Silvia Bucci, Mohammad Reza Loghmani, and Tatiana Tommasi. On the effectiveness of image rotation for open set domain adaptation. In European conference on computer vision, pages 422-438. Springer, 2020. 2" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 53, + 409, + 287, + 453 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 409, + 287, + 453 + ], + "spans": [ + { + "bbox": [ + 53, + 409, + 287, + 453 + ], + "type": "text", + "content": "[9] Nicolas Courty, Rémi Flamary, and Devis Tuia. Domain adaptation with regularized optimal transport. In Joint European Conference on Machine Learning and Knowledge Discovery in Databases, pages 274-289. Springer, 2014. 5" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 48, + 454, + 287, + 487 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 454, + 287, + 487 + ], + "spans": [ + { + "bbox": [ + 48, + 454, + 287, + 487 + ], + "type": "text", + "content": "[10] Marco Cuturi. Sinkhorn distances: Lightspeed computation of optimal transport. Advances in neural information processing systems, 26, 2013. 5" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 48, + 488, + 287, + 544 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 488, + 287, + 544 + ], + "spans": [ + { + "bbox": [ + 48, + 488, + 287, + 544 + ], + "type": "text", + "content": "[11] Bharath Bhushan Damodaran, Benjamin Kellenberger, Rémi Flamary, Devis Tuia, and Nicolas Courty. Deepjdot: Deep joint distribution optimal transport for unsupervised domain adaptation. In European Conference on Computer Vision, pages 467-483. Springer, 2018. 6" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 48, + 544, + 287, + 599 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 544, + 287, + 599 + ], + "spans": [ + { + "bbox": [ + 48, + 544, + 287, + 599 + ], + "type": "text", + "content": "[12] Jun Dan, Weiming Liu, Chunfeng Xie, Hua Yu, Shunjie Dong, and Yanchao Tan. Tfgda: Exploring topology and feature alignment in semi-supervised graph domain adaptation through robust clustering. In The Thirty-eighth Annual Conference on Neural Information Processing Systems. 1" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 48, + 601, + 287, + 644 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 601, + 287, + 644 + ], + "spans": [ + { + "bbox": [ + 48, + 601, + 287, + 644 + ], + "type": "text", + "content": "[13] Jun Dan, Tao Jin, Hao Chi, Shunjie Dong, and Yixuan Shen. Uncertainty-guided joint unbalanced optimal transport for unsupervised domain adaptation. Neural Computing and Applications, 35(7):5351-5367, 2023. 2" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 48, + 647, + 287, + 689 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 647, + 287, + 689 + ], + "spans": [ + { + "bbox": [ + 48, + 647, + 287, + 689 + ], + "type": "text", + "content": "[14] Jun Dan, Tao Jin, Hao Chi, Shunjie Dong, Haoran Xie, Keying Cao, and Xinjing Yang. Trust-aware conditional adversarial domain adaptation with feature norm alignment. Neural Networks, 168:518-530, 2023. 2" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 48, + 691, + 287, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 691, + 287, + 713 + ], + "spans": [ + { + "bbox": [ + 48, + 691, + 287, + 713 + ], + "type": "text", + "content": "[15] Jun Dan, Tao Jin, Hao Chi, Yixuan Shen, Jiawang Yu, and Jinhai Zhou. Homda: High-order moment-based domain" + } + ] + } + ], + "index": 15 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 307, + 73, + 545, + 713 + ], + "type": "list", + "angle": 0, + "index": 31, + "blocks": [ + { + "bbox": [ + 326, + 73, + 545, + 95 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 326, + 73, + 545, + 95 + ], + "spans": [ + { + "bbox": [ + 326, + 73, + 545, + 95 + ], + "type": "text", + "content": "alignment for unsupervised domain adaptation. Knowledge-Based Systems, 261:110205, 2023. 1" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 308, + 96, + 545, + 152 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 96, + 545, + 152 + ], + "spans": [ + { + "bbox": [ + 308, + 96, + 545, + 152 + ], + "type": "text", + "content": "[16] Jun Dan, Yang Liu, Haoyu Xie, Jiankang Deng, Haoran Xie, Xuansong Xie, and Baigui Sun. Transface: Calibrating transformer training for face recognition from a data-centric perspective. In Proceedings of the IEEE/CVF international conference on computer vision, pages 20642-20653, 2023. 1" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 307, + 153, + 545, + 206 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 153, + 545, + 206 + ], + "spans": [ + { + "bbox": [ + 307, + 153, + 545, + 206 + ], + "type": "text", + "content": "[17] Jun Dan, Mushui Liu, Chunfeng Xie, Jiawang Yu, Haoran Xie, Ruokun Li, and Shunjie Dong. Similar norm more transferable: Rethinking feature norms discrepancy in adversarial domain adaptation. Knowledge-Based Systems, 296: 111908, 2024. 2" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 308, + 209, + 545, + 274 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 209, + 545, + 274 + ], + "spans": [ + { + "bbox": [ + 308, + 209, + 545, + 274 + ], + "type": "text", + "content": "[18] Jun Dan, Weiming Liu, Mushui Liu, Chunfeng Xie, Shunjie Dong, Guofang Ma, Yanchao Tan, and Jiazheng Xing. Hogda: Boosting semi-supervised graph domain adaptation via high-order structure-guided adaptive feature alignment. In Proceedings of the 32nd ACM International Conference on Multimedia, pages 11109-11118, 2024. 1" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 308, + 276, + 545, + 318 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 276, + 545, + 318 + ], + "spans": [ + { + "bbox": [ + 308, + 276, + 545, + 318 + ], + "type": "text", + "content": "[19] Jun Dan, Yang Liu, Jiankang Deng, Haoyu Xie, Siyuan Li, Baigui Sun, and Shan Luo. Topofr: A closer look at topology alignment on face recognition. arXiv preprint arXiv:2410.10587, 2024. 1" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 308, + 320, + 545, + 353 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 320, + 545, + 353 + ], + "spans": [ + { + "bbox": [ + 308, + 320, + 545, + 353 + ], + "type": "text", + "content": "[20] Yuqi Fang, Pew-Thian Yap, Weili Lin, Hongtu Zhu, and Mingxia Liu. Source-free unsupervised domain adaptation: A survey. arXiv preprint arXiv:2301.00265, 2022. 3" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 308, + 354, + 545, + 398 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 354, + 545, + 398 + ], + "spans": [ + { + "bbox": [ + 308, + 354, + 545, + 398 + ], + "type": "text", + "content": "[21] Zhen Fang, Jie Lu, Feng Liu, Junyu Xuan, and Guangquan Zhang. Open set domain adaptation: Theoretical bound and algorithm. IEEE transactions on neural networks and learning systems, 32(10):4309-4322, 2020. 3" + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 308, + 399, + 545, + 453 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 399, + 545, + 453 + ], + "spans": [ + { + "bbox": [ + 308, + 399, + 545, + 453 + ], + "type": "text", + "content": "[22] Abolfazl Farahani, Sahar Voghoei, Khaled Rasheed, and Hamid R Arabnia. A brief review of domain adaptation. Advances in data science and information engineering: proceedings from ICDATA 2020 and IKE 2020, pages 877-894, 2021. 2" + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 308, + 455, + 545, + 498 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 455, + 545, + 498 + ], + "spans": [ + { + "bbox": [ + 308, + 455, + 545, + 498 + ], + "type": "text", + "content": "[23] Kilian Fatras, Thibault Séjourne, Rémi Flamary, and Nicolas Courty. Unbalanced minibatch optimal transport; applications to domain adaptation. In International Conference on Machine Learning, pages 3186-3197. PMLR, 2021. 2" + } + ] + } + ], + "index": 25 + }, + { + "bbox": [ + 308, + 501, + 545, + 564 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 501, + 545, + 564 + ], + "spans": [ + { + "bbox": [ + 308, + 501, + 545, + 564 + ], + "type": "text", + "content": "[24] Jean Feydy, Thibault Séjourne, François-Xavier Vialard, Shun-ichi Amari, Alain Trouve, and Gabriel Peyre. Interpolating between optimal transport and mmd using sinkhorn divergences. In The 22nd International Conference on Artificial Intelligence and Statistics, pages 2681–2690. PMLR, 2019. 2" + } + ] + } + ], + "index": 26 + }, + { + "bbox": [ + 308, + 567, + 545, + 599 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 567, + 545, + 599 + ], + "spans": [ + { + "bbox": [ + 308, + 567, + 545, + 599 + ], + "type": "text", + "content": "[25] Rémi Flamary, Nicholas Courty, Davis Tuia, and Alain Rakotomamonjy. Optimal transport for domain adaptation. IEEE Trans. Pattern Anal. Mach. Intell, 1:1-40, 2016. 5" + } + ] + } + ], + "index": 27 + }, + { + "bbox": [ + 308, + 601, + 545, + 666 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 601, + 545, + 666 + ], + "spans": [ + { + "bbox": [ + 308, + 601, + 545, + 666 + ], + "type": "text", + "content": "[26] Rémi Flamary, Nicolas Courty, Alexandre Gramfort, Mokhtar Z Alaya, Aurélie Boisbunon, Stanislas Chambon, Laetitia Chapel, Adrien Corenflos, Kilian Fatras, Nemo Fournier, et al. Pot: Python optimal transport. The Journal of Machine Learning Research, 22(1):3571-3578, 2021. 2" + } + ] + } + ], + "index": 28 + }, + { + "bbox": [ + 308, + 668, + 545, + 689 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 668, + 545, + 689 + ], + "spans": [ + { + "bbox": [ + 308, + 668, + 545, + 689 + ], + "type": "text", + "content": "[27] Yves Grandvalet and Yoshua Bengio. Entropy regularization., 2006. 5" + } + ] + } + ], + "index": 29 + }, + { + "bbox": [ + 308, + 692, + 545, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 692, + 545, + 713 + ], + "spans": [ + { + "bbox": [ + 308, + 692, + 545, + 713 + ], + "type": "text", + "content": "[28] John A Hartigan and Manchek A Wong. Algorithm as 136: A k-means clustering algorithm. Journal of the royal sta" + } + ] + } + ], + "index": 30 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "text", + "content": "4935" + } + ] + } + ], + "index": 32 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 8 + }, + { + "para_blocks": [ + { + "bbox": [ + 48, + 72, + 287, + 713 + ], + "type": "list", + "angle": 0, + "index": 16, + "blocks": [ + { + "bbox": [ + 67, + 72, + 286, + 94 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 67, + 72, + 286, + 94 + ], + "spans": [ + { + "bbox": [ + 67, + 72, + 286, + 94 + ], + "type": "text", + "content": "tistical society. series c (applied statistics), 28(1):100-108, 1979. 5" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 48, + 96, + 287, + 140 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 96, + 287, + 140 + ], + "spans": [ + { + "bbox": [ + 48, + 96, + 287, + 140 + ], + "type": "text", + "content": "[29] Kaiming He, Xiangyu Zhang, Shaoqing Ren, and Jian Sun. Deep residual learning for image recognition. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 770-778, 2016. 6" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 48, + 141, + 287, + 163 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 141, + 287, + 163 + ], + "spans": [ + { + "bbox": [ + 48, + 141, + 287, + 163 + ], + "type": "text", + "content": "[30] Zi-Xian Huang and Chuan-Xian Ren. Rethinking correlation learning via label prior for open set domain adaptation. 6" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 48, + 164, + 287, + 196 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 164, + 287, + 196 + ], + "spans": [ + { + "bbox": [ + 48, + 164, + 287, + 196 + ], + "type": "text", + "content": "[31] J. J. Hull. A database for handwritten text recognition research. IEEE Transactions on Pattern Analysis & Machine Intelligence, 16(5):550-554, 2002. 6, 15" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 48, + 198, + 287, + 252 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 198, + 287, + 252 + ], + "spans": [ + { + "bbox": [ + 48, + 198, + 287, + 252 + ], + "type": "text", + "content": "[32] JoonHo Jang, Byeonghu Na, Dong Hyeok Shin, Mingi Ji, Kyungwoo Song, and Il-Chul Moon. Unknown-aware domain adversarial learning for open-set domain adaptation. Advances in Neural Information Processing Systems, 35: 16755-16767, 2022. 6" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 48, + 254, + 287, + 297 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 254, + 287, + 297 + ], + "spans": [ + { + "bbox": [ + 48, + 254, + 287, + 297 + ], + "type": "text", + "content": "[33] Jun-Gi Jang, Sooyeon Shim, Vladimir Egay, Jeeyong Lee, Jongmin Park, Suhyun Chae, and U Kang. Accurate open-set recognition for memory workload. ACM Transactions on Knowledge Discovery from Data, 17(9):1-14, 2023. 2" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 48, + 299, + 287, + 320 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 299, + 287, + 320 + ], + "spans": [ + { + "bbox": [ + 48, + 299, + 287, + 320 + ], + "type": "text", + "content": "[34] Jeff L Kennington and Richard V Helgason. Algorithms for network programming. John Wiley & Sons, Inc., 1980. 5" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 48, + 322, + 287, + 353 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 322, + 287, + 353 + ], + "spans": [ + { + "bbox": [ + 48, + 322, + 287, + 353 + ], + "type": "text", + "content": "[35] Diederik P Kingma and Jimmy Ba. Adam: A method for stochastic optimization. arXiv preprint arXiv:1412.6980, 2014. 6" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 48, + 355, + 287, + 399 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 355, + 287, + 399 + ], + "spans": [ + { + "bbox": [ + 48, + 355, + 287, + 399 + ], + "type": "text", + "content": "[36] Jogendra Nath Kundu, Naveen Venkat, R Venkatesh Babu, et al. Universal source-free domain adaptation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 4544-4553, 2020. 2" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 48, + 401, + 287, + 455 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 401, + 287, + 455 + ], + "spans": [ + { + "bbox": [ + 48, + 401, + 287, + 455 + ], + "type": "text", + "content": "[37] Jogendra Nath Kundu, Naveen Venkat, Ambareesh Revanur, R Venkatesh Babu, et al. Towards inheritable models for open-set domain adaptation. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 12376-12385, 2020. 6" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 48, + 456, + 287, + 488 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 456, + 287, + 488 + ], + "spans": [ + { + "bbox": [ + 48, + 456, + 287, + 488 + ], + "type": "text", + "content": "[38] Van Der Maaten Laurens and Geoffrey Hinton. Visualizing data using t-sne. Journal of Machine Learning Research, 9 (2605):2579-2605, 2008. 8" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 48, + 490, + 287, + 544 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 490, + 287, + 544 + ], + "spans": [ + { + "bbox": [ + 48, + 490, + 287, + 544 + ], + "type": "text", + "content": "[39] Khang Le, Huy Nguyen, Quang M Nguyen, Tung Pham, Hung Bui, and Nhat Ho. On robust optimal transport: Computational complexity and barycenter computation. Advances in Neural Information Processing Systems, 34:21947-21959, 2021. 2" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 48, + 545, + 287, + 589 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 545, + 287, + 589 + ], + "spans": [ + { + "bbox": [ + 48, + 545, + 287, + 589 + ], + "type": "text", + "content": "[40] Y. L. Lecun, Leon Bottou, Yoshua Bengio, and Patrick Haffner. Gradient-based learning applied to document recognition. proc. ii. Proceedings of the IEEE, 86(11):2278-2324, 1998. 6, 15" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 48, + 590, + 287, + 622 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 590, + 287, + 622 + ], + "spans": [ + { + "bbox": [ + 48, + 590, + 287, + 622 + ], + "type": "text", + "content": "[41] Na Lei and Xianfeng Gu. Computational Conformal Geometry. International Press and Higher Education Press, 2021. 4" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 48, + 624, + 287, + 667 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 624, + 287, + 667 + ], + "spans": [ + { + "bbox": [ + 48, + 624, + 287, + 667 + ], + "type": "text", + "content": "[42] Bo Li, Yifei Shen, Jingkang Yang, Yezhen Wang, Jiawei Ren, Tong Che, Jun Zhang, and Ziwei Liu. Sparse mixture-of-experts are domain generalizable learners. arXiv preprint arXiv:2206.04046, 2022. 2" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 48, + 670, + 287, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 670, + 287, + 713 + ], + "spans": [ + { + "bbox": [ + 48, + 670, + 287, + 713 + ], + "type": "text", + "content": "[43] Wuyang Li, Jie Liu, Bo Han, and Yixuan Yuan. Adjustment and alignment for unbiased open set domain adaptation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 24110-24119, 2023. 6" + } + ] + } + ], + "index": 15 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 307, + 73, + 547, + 713 + ], + "type": "list", + "angle": 0, + "index": 30, + "blocks": [ + { + "bbox": [ + 307, + 73, + 545, + 117 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 73, + 545, + 117 + ], + "spans": [ + { + "bbox": [ + 307, + 73, + 545, + 117 + ], + "type": "text", + "content": "[44] Jian Liang, Dapeng Hu, and Jiashi Feng. Do we really need to access the source data? source hypothesis transfer for unsupervised domain adaptation. In International conference on machine learning, pages 6028-6039. PMLR, 2020. 3, 6" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 307, + 119, + 547, + 173 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 119, + 547, + 173 + ], + "spans": [ + { + "bbox": [ + 307, + 119, + 547, + 173 + ], + "type": "text", + "content": "[45] Mattia Litrico, Alessio Del Bue, and Pietro Morerio. Guiding pseudo-labels with uncertainty estimation for source-free unsupervised domain adaptation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 7640-7650, 2023. 5" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 307, + 175, + 547, + 229 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 175, + 547, + 229 + ], + "spans": [ + { + "bbox": [ + 307, + 175, + 547, + 229 + ], + "type": "text", + "content": "[46] Hong Liu, Zhangjie Cao, Mingsheng Long, Jianmin Wang, and Qiang Yang. Separate to adapt: Open set domain adaptation via progressive separation. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 2927-2936, 2019. 2, 6, 8, 15" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 307, + 231, + 545, + 274 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 231, + 545, + 274 + ], + "spans": [ + { + "bbox": [ + 307, + 231, + 545, + 274 + ], + "type": "text", + "content": "[47] Jie Liu, Xiaoqing Guo, and Yixuan Yuan. Unknown-oriented learning for open set domain adaptation. In European Conference on Computer Vision, pages 334-350. Springer, 2022. 2, 6" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 307, + 276, + 545, + 308 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 276, + 545, + 308 + ], + "spans": [ + { + "bbox": [ + 307, + 276, + 545, + 308 + ], + "type": "text", + "content": "[48] Mushui Liu, Yunlong Yu, Zhong Ji, Jungong Han, and Zhongfei Zhang. Tolerant self-distillation for image classification. Neural Networks, 174:106215, 2024. 1" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 307, + 310, + 545, + 353 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 310, + 545, + 353 + ], + "spans": [ + { + "bbox": [ + 307, + 310, + 545, + 353 + ], + "type": "text", + "content": "[49] Mushui Liu, Weijie He, Ziqian Lu, Jun Dan, Yunlong Yu, Yingming Li, Xi Li, and Jungong Han. Synth-clip: Synthetic data make clip generalize better in data-limited scenarios. Neural Networks, 184:107083, 2025. 2" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 307, + 355, + 545, + 376 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 355, + 545, + 376 + ], + "spans": [ + { + "bbox": [ + 307, + 355, + 545, + 376 + ], + "type": "text", + "content": "[50] Tianlin Liu, Joan Puigcerver, and Mathieu Blondel. Sparsity-constrained optimal transport. *ICLR*, 2023. 5" + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 307, + 378, + 545, + 444 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 378, + 545, + 444 + ], + "spans": [ + { + "bbox": [ + 307, + 378, + 545, + 444 + ], + "type": "text", + "content": "[51] Weiming Liu, Xiaolin Zheng, Jiajie Su, Mengling Hu, Yanchao Tan, and Chaochao Chen. Exploiting variational domain-invariant user embedding for partially overlapped cross domain recommendation. In Proceedings of the 45th International ACM SIGIR conference on research and development in information retrieval, pages 312-321, 2022. 1" + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 307, + 445, + 545, + 499 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 445, + 545, + 499 + ], + "spans": [ + { + "bbox": [ + 307, + 445, + 545, + 499 + ], + "type": "text", + "content": "[52] Weiming Liu, Xiaolin Zheng, Chaochao Chen, Jiajie Su, Xinting Liao, Mengling Hu, and Yanchao Tan. Joint internal multi-interest exploration and external domain alignment for cross domain sequential recommendation. In Proceedings of the ACM web conference 2023, pages 383-394, 2023." + } + ] + } + ], + "index": 25 + }, + { + "bbox": [ + 307, + 501, + 545, + 555 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 501, + 545, + 555 + ], + "spans": [ + { + "bbox": [ + 307, + 501, + 545, + 555 + ], + "type": "text", + "content": "[53] Weiming Liu, Chaochao Chen, Xinting Liao, Mengling Hu, Jiajie Su, Yanchao Tan, and Fan Wang. User distribution mapping modelling with collaborative filtering for cross-domain recommendation. In Proceedings of the ACM Web Conference 2024, pages 334-343, 2024." + } + ] + } + ], + "index": 26 + }, + { + "bbox": [ + 307, + 557, + 545, + 623 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 557, + 545, + 623 + ], + "spans": [ + { + "bbox": [ + 307, + 557, + 545, + 623 + ], + "type": "text", + "content": "[54] Weiming Liu, Chaochao Chen, Xinting Liao, Mengling Hu, Yanchao Tan, Fan Wang, Xiaolin Zheng, and Yew Soon Ong. Learning accurate and bidirectional transformation via dynamic embedding transportation for cross-domain recommendation. In Proceedings of the AAAI Conference on Artificial Intelligence, pages 8815-8823, 2024. 1" + } + ] + } + ], + "index": 27 + }, + { + "bbox": [ + 307, + 624, + 545, + 689 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 624, + 545, + 689 + ], + "spans": [ + { + "bbox": [ + 307, + 624, + 545, + 689 + ], + "type": "text", + "content": "[55] Weiming Liu, Xiaolin Zheng, Chaochao Chen, Jiahe Xu, Xinting Liao, Fan Wang, Yanchao Tan, and Yew-Soon Ong. Reducing item discrepancy via differentially private robust embedding alignment for privacy-preserving cross domain recommendation. In *Forty-first International Conference on Machine Learning*, 2024. 4" + } + ] + } + ], + "index": 28 + }, + { + "bbox": [ + 307, + 691, + 545, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 691, + 545, + 713 + ], + "spans": [ + { + "bbox": [ + 307, + 691, + 545, + 713 + ], + "type": "text", + "content": "[56] Ye Liu, Lingfeng Qiao, Changchong Lu, Di Yin, Chen Lin, Haoyuan Peng, and Bo Ren. Osman: A one-stage alignment" + } + ] + } + ], + "index": 29 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 749, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 749, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 749, + 315, + 757 + ], + "type": "text", + "content": "4936" + } + ] + } + ], + "index": 31 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 9 + }, + { + "para_blocks": [ + { + "bbox": [ + 48, + 72, + 287, + 713 + ], + "type": "list", + "angle": 0, + "index": 13, + "blocks": [ + { + "bbox": [ + 66, + 72, + 287, + 116 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 66, + 72, + 287, + 116 + ], + "spans": [ + { + "bbox": [ + 66, + 72, + 287, + 116 + ], + "type": "text", + "content": "network to unify multimodal alignment and unsupervised domain adaptation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 3551-3560, 2023. 1" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 48, + 119, + 287, + 173 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 119, + 287, + 173 + ], + "spans": [ + { + "bbox": [ + 48, + 119, + 287, + 173 + ], + "type": "text", + "content": "[57] Yang Liu, Zhipeng Zhou, and Baigui Sun. Cot: Unsupervised domain adaptation with clustering and optimal transport. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 19998-20007, 2023. 1" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 48, + 176, + 287, + 230 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 176, + 287, + 230 + ], + "spans": [ + { + "bbox": [ + 48, + 176, + 287, + 230 + ], + "type": "text", + "content": "[58] Mingsheng Long, Jianmin Wang, Guiguang Ding, Jiaguang Sun, and Philip S Yu. Transfer feature learning with joint distribution adaptation. In Proceedings of the IEEE international conference on computer vision, pages 2200-2207, 2013. 1" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 48, + 233, + 287, + 277 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 233, + 287, + 277 + ], + "spans": [ + { + "bbox": [ + 48, + 233, + 287, + 277 + ], + "type": "text", + "content": "[59] Mingsheng Long, Yue Cao, Jianmin Wang, and Michael Jordan. Learning transferable features with deep adaptation networks. In International Conference on Machine Learning, pages 97-105, 2015. 1, 2" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 48, + 279, + 287, + 322 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 279, + 287, + 322 + ], + "spans": [ + { + "bbox": [ + 48, + 279, + 287, + 322 + ], + "type": "text", + "content": "[60] Mingsheng Long, Han Zhu, Jianmin Wang, and Michael I Jordan. Unsupervised domain adaptation with residual transfer networks. Advances in neural information processing systems, 29, 2016. 1" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 48, + 325, + 287, + 380 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 325, + 287, + 380 + ], + "spans": [ + { + "bbox": [ + 48, + 325, + 287, + 380 + ], + "type": "text", + "content": "[61] Mingsheng Long, Han Zhu, Jianmin Wang, and Michael I Jordan. Deep transfer learning with joint adaptation networks. In Proceedings of the 34th International Conference on Machine Learning-Volume 70, pages 2208-2217. JMLR.org, 2017. 1, 2" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 48, + 382, + 287, + 426 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 382, + 287, + 426 + ], + "spans": [ + { + "bbox": [ + 48, + 382, + 287, + 426 + ], + "type": "text", + "content": "[62] Mingsheng Long, Zhangjie Cao, Jianmin Wang, and Michael I Jordan. Conditional adversarial domain adaptation. In Advances in Neural Information Processing Systems, pages 1640-1650, 2018. 2" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 48, + 428, + 287, + 472 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 428, + 287, + 472 + ], + "spans": [ + { + "bbox": [ + 48, + 428, + 287, + 472 + ], + "type": "text", + "content": "[63] Yadan Luo, Zijian Wang, Zi Huang, and Mahsa Baktashmotlagh. Progressive graph learning for open-set domain adaptation. In International Conference on Machine Learning, pages 6468-6478. PMLR, 2020. 6" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 48, + 474, + 287, + 518 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 474, + 287, + 518 + ], + "spans": [ + { + "bbox": [ + 48, + 474, + 287, + 518 + ], + "type": "text", + "content": "[64] Yadan Luo, Zijian Wang, Zhuoxiao Chen, Zi Huang, and Mahsa Baktashmotlagh. Source-free progressive graph learning for open-set domain adaptation. IEEE Transactions on Pattern Analysis and Machine Intelligence, 2023. 3, 6, 8" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 48, + 520, + 287, + 564 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 520, + 287, + 564 + ], + "spans": [ + { + "bbox": [ + 48, + 520, + 287, + 564 + ], + "type": "text", + "content": "[65] You-Wei Luo and Chuan-Xian Ren. Conditional bures metric for domain adaptation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 13989–13998, 2021. 1" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 48, + 567, + 287, + 609 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 567, + 287, + 609 + ], + "spans": [ + { + "bbox": [ + 48, + 567, + 287, + 609 + ], + "type": "text", + "content": "[66] You-Wei Luo and Chuan-Xian Ren. Mot: Masked optimal transport for partial domain adaptation. In 2023 IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 3531–3540. IEEE, 2023. 1" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 48, + 612, + 287, + 666 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 612, + 287, + 666 + ], + "spans": [ + { + "bbox": [ + 48, + 612, + 287, + 666 + ], + "type": "text", + "content": "[67] Yuval Netzer, Tao Wang, Adam Coates, Alessandro Bissacco, Bo Wu, and Andrew Y. Ng. Reading digits in natural images with unsupervised feature learning. Nips Workshop on Deep Learning & Unsupervised Feature Learning, 2011. 6, 15" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 48, + 669, + 287, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 669, + 287, + 713 + ], + "spans": [ + { + "bbox": [ + 48, + 669, + 287, + 713 + ], + "type": "text", + "content": "[68] Khai Nguyen, Dang Nguyen, Quoc Nguyen, Tung Pham, Hung Bui, Dinh Phung, Trung Le, and Nhat Ho. On transportation of mini-batches: A hierarchical approach. arXiv preprint arXiv:2102.05912, 2021. 2" + } + ] + } + ], + "index": 12 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 307, + 73, + 545, + 712 + ], + "type": "list", + "angle": 0, + "index": 28, + "blocks": [ + { + "bbox": [ + 307, + 73, + 545, + 106 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 73, + 545, + 106 + ], + "spans": [ + { + "bbox": [ + 307, + 73, + 545, + 106 + ], + "type": "text", + "content": "[69] Pau Panareda Busto and Juergen Gall. Open set domain adaptation. In Proceedings of the IEEE international conference on computer vision, pages 754-763, 2017. 1, 2" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 307, + 107, + 545, + 161 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 107, + 545, + 161 + ], + "spans": [ + { + "bbox": [ + 307, + 107, + 545, + 161 + ], + "type": "text", + "content": "[70] Xingchao Peng, Ben Usman, Neela Kaushik, Dequan Wang, Judy Hoffman, and Kate Saenko. Visda: A synthetic-to-real benchmark for visual domain adaptation. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition Workshops, pages 2021-2026, 2018. 6, 15" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 307, + 162, + 545, + 183 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 162, + 545, + 183 + ], + "spans": [ + { + "bbox": [ + 307, + 162, + 545, + 183 + ], + "type": "text", + "content": "[71] Gabriel Peyré. Course notes on computational optimal transport. 2019. 4" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 308, + 184, + 545, + 227 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 184, + 545, + 227 + ], + "spans": [ + { + "bbox": [ + 308, + 184, + 545, + 227 + ], + "type": "text", + "content": "[72] Khiem Pham, Khang Le, Nhat Ho, Tung Pham, and Hung Bui. On unbalanced optimal transport: An analysis of sinkhorn algorithm. In International Conference on Machine Learning, pages 7673-7682. PMLR, 2020. 2" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 308, + 228, + 545, + 282 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 228, + 545, + 282 + ], + "spans": [ + { + "bbox": [ + 308, + 228, + 545, + 282 + ], + "type": "text", + "content": "[73] Sanqing Qu, Tianpei Zou, Florian Röhrbein, Cewu Lu, Guang Chen, Dacheng Tao, and Changjun Jiang. Upcycling models under domain and category shift. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 20019-20028, 2023. 2" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 308, + 284, + 545, + 347 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 284, + 545, + 347 + ], + "spans": [ + { + "bbox": [ + 308, + 284, + 545, + 347 + ], + "type": "text", + "content": "[74] Sanqing Qu, Tianpei Zou, Lianghua He, Florian Röhrbein, Alois Knoll, Guang Chen, and Changjun Jiang. Lead: Learning decomposition for source-free universal domain adaptation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 23334-23343, 2024. 2" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 308, + 350, + 545, + 415 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 350, + 545, + 415 + ], + "spans": [ + { + "bbox": [ + 308, + 350, + 545, + 415 + ], + "type": "text", + "content": "[75] Ievgen Redko, Amaury Habrard, and Marc Sebban. Theoretical analysis of domain adaptation with optimal transport. In Machine Learning and Knowledge Discovery in Databases: European Conference, ECML PKDD 2017, Skopje, Macedonia, September 18-22, 2017, Proceedings, Part II 10, pages 737-753. Springer, 2017. 5" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 308, + 416, + 545, + 469 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 416, + 545, + 469 + ], + "spans": [ + { + "bbox": [ + 308, + 416, + 545, + 469 + ], + "type": "text", + "content": "[76] Ievgen Redko, Nicolas Courty, Rémi Flamary, and Devis Tuia. Optimal transport for multi-source domain adaptation under target shift. In The 22nd International Conference on artificial intelligence and statistics, pages 849-858. PMLR, 2019. 2" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 308, + 471, + 545, + 504 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 471, + 545, + 504 + ], + "spans": [ + { + "bbox": [ + 308, + 471, + 545, + 504 + ], + "type": "text", + "content": "[77] Kate Saenko, Brian Kulis, Mario Fritz, and Trevor Darrell. Adapting visual category models to new domains. In European Conference on Computer Vision (ECCV), 2010. 6, 15" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 308, + 505, + 545, + 548 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 505, + 545, + 548 + ], + "spans": [ + { + "bbox": [ + 308, + 505, + 545, + 548 + ], + "type": "text", + "content": "[78] Kuniaki Saito, Shohei Yamamoto, Yoshitaka Ushiku, and Tatsuya Harada. Open set domain adaptation by backpropagation. In Proceedings of the European conference on computer vision (ECCV), pages 153-168, 2018. 1, 2, 6, 16" + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 308, + 548, + 545, + 602 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 548, + 545, + 602 + ], + "spans": [ + { + "bbox": [ + 308, + 548, + 545, + 602 + ], + "type": "text", + "content": "[79] Thibault Séjourne, François-Xavier Vialard, and Gabriel Peyré. Faster unbalanced optimal transport: Translation invariant sinkhorn and 1-d frank-wolfe. In International Conference on Artificial Intelligence and Statistics, pages 4995-5021. PMLR, 2022. 13" + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 308, + 603, + 545, + 635 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 603, + 545, + 635 + ], + "spans": [ + { + "bbox": [ + 308, + 603, + 545, + 635 + ], + "type": "text", + "content": "[80] Jian Shen, Yanru Qu, Weinan Zhang, and Yong Yu. Wasserstein distance guided representation learning for domain adaptation. arXiv preprint arXiv:1707.01217, 2017. 2" + } + ] + } + ], + "index": 25 + }, + { + "bbox": [ + 308, + 636, + 545, + 669 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 636, + 545, + 669 + ], + "spans": [ + { + "bbox": [ + 308, + 636, + 545, + 669 + ], + "type": "text", + "content": "[81] Rui Shu, Hung H Bui, Hirokazu Narui, and Stefano Ermon. A dirt-t approach to unsupervised domain adaptation. arXiv preprint arXiv:1802.08735, 2018. 2, 6" + } + ] + } + ], + "index": 26 + }, + { + "bbox": [ + 308, + 670, + 545, + 712 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 670, + 545, + 712 + ], + "spans": [ + { + "bbox": [ + 308, + 670, + 545, + 712 + ], + "type": "text", + "content": "[82] Baochen Sun and Kate Saenko. Deep coral: Correlation alignment for deep domain adaptation. In European Conference on Computer Vision, pages 443-450. Springer, 2016. 2" + } + ] + } + ], + "index": 27 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 749, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 749, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 749, + 315, + 757 + ], + "type": "text", + "content": "4937" + } + ] + } + ], + "index": 29 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 10 + }, + { + "para_blocks": [ + { + "bbox": [ + 48, + 72, + 287, + 699 + ], + "type": "list", + "angle": 0, + "index": 15, + "blocks": [ + { + "bbox": [ + 48, + 72, + 287, + 106 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 72, + 287, + 106 + ], + "spans": [ + { + "bbox": [ + 48, + 72, + 287, + 106 + ], + "type": "text", + "content": "[83] Eric Tzeng, Judy Hoffman, Ning Zhang, Kate Saenko, and Trevor Darrell. Deep domain confusion: Maximizing for domain invariance. arXiv preprint arXiv:1412.3474, 2014. 2" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 48, + 107, + 287, + 150 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 107, + 287, + 150 + ], + "spans": [ + { + "bbox": [ + 48, + 107, + 287, + 150 + ], + "type": "text", + "content": "[84] Hemanth Venkateswara, Jose Eusebio, Shayok Chakraborty, and Sethuraman Panchanathan. Deep hashing network for unsupervised domain adaptation. CoRR, abs/1706.07522, 2017. 6, 15" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 48, + 152, + 287, + 174 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 152, + 287, + 174 + ], + "spans": [ + { + "bbox": [ + 48, + 152, + 287, + 174 + ], + "type": "text", + "content": "[85] Cédric Villani et al. Optimal transport: old and new. Springer, 2009. 2" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 48, + 175, + 287, + 196 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 175, + 287, + 196 + ], + "spans": [ + { + "bbox": [ + 48, + 175, + 287, + 196 + ], + "type": "text", + "content": "[86] Mei Wang and Weihong Deng. Deep visual domain adaptation: A survey. Neurocomputing, 312:135-153, 2018. 2" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 48, + 198, + 287, + 239 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 198, + 287, + 239 + ], + "spans": [ + { + "bbox": [ + 48, + 198, + 287, + 239 + ], + "type": "text", + "content": "[87] Qian Wang, Fanlin Meng, and Toby P Breckon. Progressively select and reject pseudo-labelled samples for open-set domain adaptation. arXiv preprint arXiv:2110.12635, 2021. 2" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 48, + 243, + 287, + 275 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 243, + 287, + 275 + ], + "spans": [ + { + "bbox": [ + 48, + 243, + 287, + 275 + ], + "type": "text", + "content": "[88] Garrett Wilson and Diane J Cook. A survey of unsupervised deep domain adaptation. ACM Transactions on Intelligent Systems and Technology (TIST), 11(5):1-46, 2020. 2" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 48, + 277, + 287, + 319 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 277, + 287, + 319 + ], + "spans": [ + { + "bbox": [ + 48, + 277, + 287, + 319 + ], + "type": "text", + "content": "[89] Zonghan Wu, Shirui Pan, Fengwen Chen, Guodong Long, Chengqi Zhang, and S Yu Philip. A comprehensive survey on graph neural networks. IEEE transactions on neural networks and learning systems, 32(1):4-24, 2020. 8" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 48, + 322, + 287, + 365 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 322, + 287, + 365 + ], + "spans": [ + { + "bbox": [ + 48, + 322, + 287, + 365 + ], + "type": "text", + "content": "[90] Yujia Xie, Xiangfeng Wang, Ruijia Wang, and Hongyuan Zha. A fast proximal point method for computing exact Wasserstein distance. In Uncertainty in artificial intelligence, pages 433-453. PMLR, 2020. 5" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 48, + 366, + 287, + 409 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 366, + 287, + 409 + ], + "spans": [ + { + "bbox": [ + 48, + 366, + 287, + 409 + ], + "type": "text", + "content": "[91] Renjun Xu, Pelen Liu, Yin Zhang, Fang Cai, Jindong Wang, Shuoying Liang, Heting Ying, and Jianwei Yin. Joint partial optimal transport for open set domain adaptation. In IJCAI, pages 2540-2546, 2020. 1, 2, 6" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 48, + 411, + 287, + 454 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 411, + 287, + 454 + ], + "spans": [ + { + "bbox": [ + 48, + 411, + 287, + 454 + ], + "type": "text", + "content": "[92] Tengyue Xu and Jun Dan. Ehm: Exploring dynamic alignment and hierarchical clustering in unsupervised domain adaptation via high-order moment-guided contrastive learning. Neural Networks, page 107188, 2025. 5" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 48, + 456, + 287, + 521 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 456, + 287, + 521 + ], + "spans": [ + { + "bbox": [ + 48, + 456, + 287, + 521 + ], + "type": "text", + "content": "[93] Hongliang Yan, Yukang Ding, Peihua Li, Qilong Wang, Yong Xu, and Wangmeng Zuo. Mind the class weight bias: Weighted maximum mean discrepancy for unsupervised domain adaptation. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition, pages 2272-2281, 2017. 2" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 48, + 523, + 287, + 566 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 523, + 287, + 566 + ], + "spans": [ + { + "bbox": [ + 48, + 523, + 287, + 566 + ], + "type": "text", + "content": "[94] Yu-Chu Yu and Hsuan-Tien Lin. Semi-supervised domain adaptation with source label adaptation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 24100-24109, 2023. 1" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 48, + 567, + 287, + 600 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 567, + 287, + 600 + ], + "spans": [ + { + "bbox": [ + 48, + 567, + 287, + 600 + ], + "type": "text", + "content": "[95] Zhiqi Yu, Jingjing Li, Zhekai Du, Lei Zhu, and Heng Tao Shen. A comprehensive survey on source-free domain adaptation. arXiv preprint arXiv:2302.11803, 2023. 3" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 48, + 601, + 287, + 645 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 601, + 287, + 645 + ], + "spans": [ + { + "bbox": [ + 48, + 601, + 287, + 645 + ], + "type": "text", + "content": "[96] Li Zhong, Zhen Fang, Feng Liu, Bo Yuan, Guangquan Zhang, and Jie Lu. Bridging the theoretical bound and deep algorithms for open set domain adaptation. IEEE transactions on neural networks and learning systems, 2021. 2, 3" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 48, + 647, + 287, + 699 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 647, + 287, + 699 + ], + "spans": [ + { + "bbox": [ + 48, + 647, + 287, + 699 + ], + "type": "text", + "content": "[97] Jingyu Zhuang, Ziliang Chen, Pengxu Wei, Guanbin Li, and Liang Lin. Discovering implicit classes achieves open set domain adaptation. In 2022 IEEE International Conference on Multimedia and Expo (ICME), pages 01-06. IEEE, 2022. 6" + } + ] + } + ], + "index": 14 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 749, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 749, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 749, + 315, + 757 + ], + "type": "text", + "content": "4938" + } + ] + } + ], + "index": 16 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 11 + } + ], + "_backend": "vlm", + "_version_name": "2.6.4" +} \ No newline at end of file diff --git a/2025/Distraction is All You Need for Multimodal Large Language Model Jailbreaking/1375bb02-e050-41f2-9b1d-ed16de66c742_content_list.json b/2025/Distraction is All You Need for Multimodal Large Language Model Jailbreaking/1375bb02-e050-41f2-9b1d-ed16de66c742_content_list.json new file mode 100644 index 0000000000000000000000000000000000000000..fc53eb8db186591121abde15d2b3f6894120635e --- /dev/null +++ b/2025/Distraction is All You Need for Multimodal Large Language Model Jailbreaking/1375bb02-e050-41f2-9b1d-ed16de66c742_content_list.json @@ -0,0 +1,1401 @@ +[ + { + "type": "text", + "text": "Distraction is All You Need for Multimodal Large Language Model Jailbreaking", + "text_level": 1, + "bbox": [ + 94, + 130, + 901, + 152 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Zuopeng Yang $^{1*}$ , Jiluan Fan $^{1*}$ , Anli Yan $^{1\\dagger}$ , Erdun Gao $^{3}$ , Xin Lin $^{1\\dagger}$ , Tao Li $^{2}$ , Kanghua Mo $^{1}$ , Changyu Dong $^{1}$", + "bbox": [ + 277, + 179, + 717, + 217 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "$^{1}$ Guangzhou University $^{2}$ Shanghai Jiao Tong University $^{3}$ Australian Institute for Machine Learning, The University of Adelaide", + "bbox": [ + 218, + 220, + 781, + 258 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/7af0484c152bcef177f680586409cb7549584bbb15a8b19ce86a288bd2d1d783.jpg", + "image_caption": [ + "Figure 1. Jailbreaking examples of Hades [12] and our proposed CS-DJ against GPT-4. When using Hades, GPT-4 effectively rejects harmful questions. In contrast, CS-DJ successfully bypasses the model's defense mechanisms by constructing multi-subimage visual inputs to distract the model, resulting in the acquisition of harmful responses. The harmful information is highlighted in red, while key prompts are emphasized in bold." + ], + "image_footnote": [], + "bbox": [ + 133, + 277, + 861, + 477 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Abstract", + "text_level": 1, + "bbox": [ + 248, + 565, + 326, + 580 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Multimodal Large Language Models (MLLMs) bridge the gap between visual and textual data, enabling a range of advanced applications. However, complex internal interactions among visual elements and their alignment with text can introduce vulnerabilities, which may be exploited to bypass safety mechanisms. To address this, we analyze the relationship between image content and task and find that the complexity of subimages, rather than their content, is key. Building on this insight, we propose the Distraction Hypothesis, followed by a novel framework called Contrasting Subimage Distraction Jailbreaking (CS-DJ), to achieve jailbreaking by disrupting MLLMs alignment through multi-level distraction strategies. CS-DJ consists of two components: structured distraction, achieved through query decomposition that induces a distributional shift by fragmenting harmful prompts into subqueries, and visual-enhanced distraction, realized by constructing contrasting subimages to disrupt the interactions", + "bbox": [ + 88, + 599, + 485, + 872 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "among visual elements within the model. This dual strategy disperses the model's attention, reducing its ability to detect and mitigate harmful content. Extensive experiments across five representative scenarios and four popular closed-source MLLMs, including GPT-4o-mini, GPT-4o, GPT-4V, and Gemini-1.5-Flash, demonstrate that CS-DJ achieves average success rates of $52.40\\%$ for the attack success rate and $74.10\\%$ for the ensemble attack success rate. These results reveal the potential of distraction-based approaches to exploit and bypass MLLMs' defenses, offering new insights for attack strategies. Our code is available at https://github.com/TeamPigeonLab/CS-DJ.", + "bbox": [ + 511, + 551, + 908, + 747 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Warning: This paper contains unfiltered content generated by MLLMs that may be offensive to readers.", + "bbox": [ + 511, + 750, + 908, + 780 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "1. Introduction", + "text_level": 1, + "bbox": [ + 513, + 813, + 645, + 828 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Since the early 2020s, Multimodal Large Language Models (MLLMs) have emerged as a powerful tool for bridging natural language processing and computer vision [3], owing to their ability to integrate and process multimodal informa", + "bbox": [ + 511, + 839, + 906, + 900 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "CVF", + "bbox": [ + 106, + 2, + 181, + 42 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore.", + "bbox": [ + 236, + 0, + 810, + 46 + ], + "page_idx": 0 + }, + { + "type": "page_footnote", + "text": "*Equal Contribution. † Corresponding authors.", + "bbox": [ + 114, + 886, + 362, + 900 + ], + "page_idx": 0 + }, + { + "type": "page_number", + "text": "9467", + "bbox": [ + 482, + 944, + 514, + 955 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "tion efficiently [25]. Despite their impressive performance, one notable limitation of these methods is that the training materials may contain toxic content, potentially leading to inappropriate or harmful content [4, 38], the leakage of sensitive information [36], and even serious vulnerabilities in the model security and privacy protection. To this end, significant efforts have been devoted to using reinforcement learning from human feedback (RLHF) to align LLMs outputs, ensuring readiness for deployment in high-stakes domains [21, 33]. However, beyond Large Language Models (LLMs), the integration of visual inputs in MLLMs introduces a new challenge: securing models against vulnerabilities arising from newly integrated visual modalities, which remains a significant issue and lacks a comprehensive understanding [17, 18, 37]", + "bbox": [ + 89, + 90, + 480, + 316 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "In this study, rather than following the human feedback alignment research at the defense side, we focus on exploring an alternative approach, namely jailbreak attacks [19, 34, 36]. This line of research also seeks to provide insights into how safety concerns emerge, while, considering the attacker's perspective. Specifically, it aims to actively uncover the defense vulnerabilities in current model by deliberately crafting inputs to manipulate MLLMs, leading to unintended and potentially harmful outputs [15]. Motivated by the fact that most current defense mechanisms rely on RLHF and are trained on human preference data [21], a natural hypothesis regarding the defense vulnerabilities of MLLMs is that the model may be susceptible to out-of-distribution (OOD) inputs.", + "bbox": [ + 89, + 321, + 482, + 532 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Given that harmful textual content is already effectively detected in the field of LLMs, current jailbreak attack research on MLLMs primarily focuses on constructing OOD visual inputs, with two main approaches: image perturbation injection and prompt-to-image infection. The image perturbation injection involves subtly modifying images and combining them with query text to compose jailbreak prompts [35]. This attack leverages adversarial approaches [30] to introduce noisy OOD inputs that exploit the MLLM's visual-text decision boundary [40]. Unfortunately, these methods typically require access to the gradient of the victim or surrogate model [32], which limits their applicability in closed-source models [29]. In contrast, prompt-to-image infection generates images with similar semantics by using malicious text. These generated images are then combined with query text to craft jailbreak prompts that compel the MLLM to produce harmful responses [6, 12]. Since these generated images differ in distribution from those used for MLLM safety alignment training, they may bypass security measures and achieve jailbreak [2, 16, 26]. Despite this, the extensive training of large MLLMs on vast datasets makes it challenging for image generation models to produce truly OOD images. A failure case is depicted in the upper portion of Figure 1.", + "bbox": [ + 89, + 537, + 482, + 901 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Findings. Despite progress, it remains unclear how to fully leverage the visual space to construct OOD inputs for effective black-box MLLM jailbreaks. Building on the observation that most harmless alignment processes primarily involve simple images [6, 12, 22], we propose that increasing image complexity, potentially through the use of multiple subimages, could generate more effective OOD inputs for jailbreak attack. To verify this, we conduct several ablation analyses (§4.4), including varying the number of subimages and examining the relationship between image content and task. Our findings reveal that it is the complexity of the subimages, rather than their conceptual content, that drives the jailbreak success. Based on this insight, we propose the Distraction Hypothesis to explain the visual effects on jailbreak attacks against MLLMs.", + "bbox": [ + 511, + 90, + 903, + 316 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Distraction Hypothesis. Encoding complex images in the input prompt increases token complexity/diversity, which raises the processing burden on MLLMs. This overload can weaken the model's defenses, making it more prone to induce unintended outputs and improving jailbreak attack effectiveness.", + "bbox": [ + 521, + 338, + 898, + 430 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Structure and contributions. Building on the Distraction Hypothesis, we further introduce a novel framework, namely Contrasting Subimage Distraction Jailbreaking (CS-DJ) $(\\S 3)$ , which leverages this insight to strategically design OOD inputs and enhance the effectiveness of jailbreak attacks. Specifically, CS-DJ explores both textual and visual components of the input space, introducing two key methods: structured distraction via query decomposition $(\\S 3.1)$ and visual-enhanced distraction through contrasting subimages $(\\S 3.2)$ . (i) The structured distraction component decomposes the original harmful query into multiple sub-queries, each representing different aspects or intermediate steps of the original query. This decomposition induces a distributional shift that disperses the model's focus and weakens its ability to detect harmful content. (ii) The visual-enhanced distraction component manipulates the visual input by constructing multiple contrasting subimages, effectively disrupting interactions among visual elements during processing. Then, to guide the model's response and enhance its susceptibility to the multi-level distractions, CS-DJ incorporates a carefully designed harmless prompt alongside the composite visual input, as illustrated in Figure 1. This integrated approach disrupts the model's internal coherence and safety mechanisms, effectively bypassing its defenses. To evaluate the effectiveness of CS-DJ, we conduct extensive experiments across five representative scenarios and four widely used closed-source MLLMs $(\\S 4)$ . Our results demonstrate that CS-DJ outperforms state-of-the-art jailbreak attacks, achieving average success rates of $52.40\\%$ for attack success rate and $74.10\\%$ for ensemble", + "bbox": [ + 511, + 446, + 903, + 900 + ], + "page_idx": 1 + }, + { + "type": "page_number", + "text": "9468", + "bbox": [ + 482, + 944, + 514, + 955 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "attack success rate.", + "bbox": [ + 91, + 92, + 220, + 104 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "2. Related Work", + "text_level": 1, + "bbox": [ + 89, + 127, + 232, + 142 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "2.1. Multimodal Large Language Models", + "text_level": 1, + "bbox": [ + 89, + 154, + 408, + 171 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Recently, Multimodal Large Language Models (MLLMs) [31] have become a prominent approach for processing and understanding heterogeneous multimodal data, including visual[39] and auditory [5] modalities. Building on instruction-tuning techniques developed for LLMs, most MLLMs are adapted from pre-trained LLMs through fine-tuning. Notable examples include LLaVA [14] and MiniGPT-4 [41], which share a similar architectural design by integrating pre-trained vision backbones with large language models via a projection layer, enabling effective vision-language alignment. Building upon this foundation, LLaVA-1.5 [13] further enhances its architecture by adopting a two-layer MLP for better vision-language mapping and increasing the vision encoder's input resolution for improved visual understanding. However, the incorporation of emerging modalities further intensifies the challenges of ensuring robust safety alignment in MLLMs. To address this challenge, advanced MLLMs, such as GPT-4o [7] and Gemini 1.5-Flash [27], employ a combination of alignment strategies, including post-training methods and integrated classifiers, to enhance both helpfulness and safety. Despite these efforts, MLLMs remains vulnerable to attacks using OOD inputs, which will be explored in this work.", + "bbox": [ + 89, + 179, + 483, + 527 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "2.2. Jailbreak Attacks on MLLMs", + "text_level": 1, + "bbox": [ + 89, + 542, + 356, + 559 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Jailbreak attacks against MLLM can be performed through visual and textual input, leading to complex and diverse attack patterns [9, 20]. As a result, numerous jailbreak methods [6, 12, 17, 22, 24] have been continuously proposed. Qi et al. [22] used visual adversarial examples to bypass the security mechanisms of visual-textual MLLMs. However, this type of attack [20, 22] is typically only effective in white-box environments and is often difficult to apply to state-of-the-art black-box models. FigStep [6] introduced a simple yet effective black-box jailbreak approach that circumvents safety alignment bytrasfering prohibited content as images using typographic transformations. MM-SafetyBench [17] utilized a multi-step approach to construct jailbreak input, incorporating question generation, rephrasing, extraction of unsafe key phrases, and query-to-image conversion. Furthermore, Hades [12] investigated the influence of image harmfulness on model jailbreaks and proposed a prompt-embedded image infection jailbreak attack method. It generated harmful images using diffusion models and amplified the jailbreak effect through optimization. Unlike these works, CS-DJ explores the safety of MLLMs from a new perspective—visual distraction.", + "bbox": [ + 89, + 568, + 482, + 901 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3. Methodology", + "text_level": 1, + "bbox": [ + 513, + 89, + 648, + 107 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Beyond the standard LLMs, MLLMs also encode visual content into token sequences, saying that the intrinsic interactions among visual elements can affect the model's ability to defend against harmful prompts. However, most of the existing approaches [12, 16, 17] only take a single, complete image as input, resulting in notable similarities both (1) within image patches, and (2) between the image and the targeted jailbreak content. Such similarities may make the model likely to detect harmful inputs, potentially leading to jailbreak failure. To address these issues, we propose a novel jailbreak framework, called CS-DJ, as illustrated in Figure 2. CS-DJ leverages a multi-subimage structure to bypass the models' intrinsic safety detection mechanism, with introducing multi-level distraction. Specifically, CS-DJ consists of two main components: structured distraction, achieved through query decomposition, and visual-enhanced distraction, realized through the construction of contrasting subimages. The resulting composite image input, combined with a carefully designed instruction, is fed into MLLMs to achieve jailbreak. Further details are described in the following sections.", + "bbox": [ + 511, + 114, + 906, + 434 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3.1. Structured Distraction via Query Decomposition", + "text_level": 1, + "bbox": [ + 511, + 441, + 906, + 472 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "MLLMs have strong defense ability against harmful textual inputs, reinforced by specific optimizations aligned with human preference data [10]. To circumvent these defenses, we introduce a structured distraction strategy that disrupts and fragments information alignment through query decomposition, thereby reducing the model's overall sensitivity to harmful content. As illustrated on the left of Figure 2, we begin by decomposing a original harmful query $Q$ into multiple sub-queries, denoted as $\\{Q_{s}^{(i)}\\}_{i = 1}^{m}$ . Each sub-query addresses part of the original query from different aspects or intermediate steps.", + "bbox": [ + 511, + 479, + 908, + 648 + ], + "page_idx": 2 + }, + { + "type": "equation", + "text": "\n$$\n\\left\\{Q _ {s} ^ {i} \\right\\} _ {i = 1} ^ {m} = \\mathcal {G} (Q), \\tag {1}\n$$\n", + "text_format": "latex", + "bbox": [ + 647, + 660, + 906, + 678 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "where $\\mathcal{G}$ is the auxiliary decomposition model and $m$ denotes the number of sub-queries. On one hand, this fragmented structure distracts the model's focus, creating a distributional shift that obscures the original intent and disrupts its ability to detect harmful content, thereby bypassing its safety mechanisms. On the other hand, during the jailbreak execution phase, we carefully design prompts that guide the MLLM to address all sub-queries simultaneously. This strategy causes the model to divide its attention among multiple tasks rather than focusing on a single query. Further details are discussed in Section 3.3. In practice, we utilized Qwen2.5-3B-Instruct [28] for query decomposition. The details of the query decomposition prompt are provided in the Appendix.", + "bbox": [ + 511, + 688, + 908, + 902 + ], + "page_idx": 2 + }, + { + "type": "page_number", + "text": "9469", + "bbox": [ + 480, + 944, + 514, + 955 + ], + "page_idx": 2 + }, + { + "type": "image", + "img_path": "images/c2ee3ec1c2cea2ffde7b1f63dab9159d9590b2ef998d6ac52d06322facb64316.jpg", + "image_caption": [ + "Figure 2. The framework of the proposed CS-DJ. Given a harmful query, CS-DJ employs a three-step process to execute a jailbreaking attack: (1) decompose the raw query into multiple sub-queries and transform them into images to introduce structural distraction, (2) retrieve contrasting images from a dataset as subimages to inject visual-enhanced distraction. (3) combine these images into a composite image as the visual input, aligned with a harmless instruction, to execute the jailbreaking attack." + ], + "image_footnote": [], + "bbox": [ + 138, + 93, + 377, + 393 + ], + "page_idx": 3 + }, + { + "type": "image", + "img_path": "images/27a8d9dac43b96769b81738d6f11734123a2e3f710058588f76ce8303133747d.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 379, + 93, + 617, + 393 + ], + "page_idx": 3 + }, + { + "type": "image", + "img_path": "images/6e1451c85d535c0e198c4edc972e0cc8e0bbb684b7cfdbfbbce6715411128a3b.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 619, + 93, + 859, + 393 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "The next step involves converting the sub-queries into images:", + "bbox": [ + 89, + 483, + 482, + 513 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\left\\{I _ {s} ^ {i} \\right\\} _ {i = 1} ^ {m} = \\mathcal {T} \\left(\\left\\{Q _ {s} ^ {i} \\right\\} _ {i = 1} ^ {m}\\right), \\tag {2}\n$$\n", + "text_format": "latex", + "bbox": [ + 202, + 511, + 482, + 531 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "where $\\mathcal{T}$ denotes the transformation function that converts each sub-query $Q_{s}^{i}$ into a corresponding image $I_{s}^{i}$ . This transformation introduces a new form of distraction by altering the input modality, complicating the model's ability to link and identify harmful content patterns. Presenting textual prompts as visual elements obscures their original intent, further disrupting the model's detection mechanisms.", + "bbox": [ + 89, + 537, + 483, + 642 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "3.2. Visual-Enhanced Distraction via Multi-subimage Construction", + "text_level": 1, + "bbox": [ + 89, + 651, + 483, + 683 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "In contrast to prior work that emphasizes the harmfulness of a single image for MLLM jailbreaks, we focus on the impact of distraction. To enhance visual input distraction, we construct multiple contrasting visual elements, each treated as a subimage that collectively forms the input image. Since MLLMs represent visual content as token sequences similar to text, the distraction must include two aspects: (1) the distraction between the text query and the visual subimage, and (2) the distraction among the visual subimages themselves. Directly constructing an image with sufficient distraction relative to a text query is challenging. Therefore, we simplify this process by transforming it into an image retrieval problem, aimed at retrieving $k$ images from a dataset. These $k$ images are selected to have the lowest similarity to", + "bbox": [ + 89, + 689, + 483, + 900 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "the query and to each other. Given the complexity of this optimization problem, we further approximate it for practical implementation. The procedure is illustrated in the middle of Figure 2. Specifically, we first encode the query $Q$ as a dense vector using the CLIP model:", + "bbox": [ + 511, + 483, + 906, + 559 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\nv (Q) = \\operatorname {C L I P} (Q). \\tag {3}\n$$\n", + "text_format": "latex", + "bbox": [ + 645, + 571, + 906, + 590 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "In practice, CLIP-ViT-L/14 [23] is utilized as the extractor. We then retrieve an image from a dataset that most contrasts with $v(Q)$ , effectively by minimizing the cosine similarity as:", + "bbox": [ + 511, + 603, + 905, + 660 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\nI _ {c} ^ {1} = \\arg \\min _ {I \\in \\mathcal {D}} \\mathbf {c o s} (v (Q), v (I)), \\tag {4}\n$$\n", + "text_format": "latex", + "bbox": [ + 604, + 665, + 906, + 688 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "where $\\mathcal{D}$ denotes the image dataset, and $\\cos (\\cdot ,\\cdot)$ represents the cosine similarity. Next, we proceed to retrieve the subsequent subimages:", + "bbox": [ + 511, + 696, + 905, + 742 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\nI _ {c} ^ {j} = \\arg \\min _ {I \\in \\mathcal {D}} \\left(\\mathbf {c o s} (v (Q), v (I)) + \\sum_ {i = 1} ^ {j - 1} \\mathbf {c o s} \\left(v \\left(I _ {c} ^ {i}\\right), v (I)\\right)\\right), \\tag {5}\n$$\n", + "text_format": "latex", + "bbox": [ + 511, + 755, + 903, + 809 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "where $j$ denotes the index of the current subimage being selected. This approach systematically ensures that each subimage has minimal similarity to both the query and the previously selected subimages, thereby maximizing contrast and enhancing the overall distraction effect for the MLLM jailbreak process.", + "bbox": [ + 511, + 810, + 906, + 901 + ], + "page_idx": 3 + }, + { + "type": "page_number", + "text": "9470", + "bbox": [ + 482, + 944, + 514, + 955 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "3.3. Jailbreaking Execution", + "text_level": 1, + "bbox": [ + 89, + 90, + 305, + 107 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "With the constructed set of subimages, we form a composite input that maximizes distraction for MLLMs. Specifically, the final composite image $I_{\\mathrm{comp}}$ is created by combining the subimages $\\{I_c^j\\}_{j=1}^k$ along with the transformed sub-queries $\\{I_s^i\\}_{i=1}^m$ :", + "bbox": [ + 89, + 112, + 483, + 191 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\nI _ {\\text {c o m p}} = \\text {C o m b i n e} \\left(\\left\\{I _ {c} ^ {j} \\right\\} _ {j = 1} ^ {k}, \\left\\{I _ {s} ^ {i} \\right\\} _ {i = 1} ^ {m}\\right), \\tag {6}\n$$\n", + "text_format": "latex", + "bbox": [ + 151, + 200, + 483, + 224 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "where the function $\\mathbf{Combine}(\\cdot, \\cdot)$ represents the structured arrangement of the subimages and sub-query transformations.", + "bbox": [ + 89, + 229, + 483, + 273 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "This composite input $I_{\\mathrm{comp}}$ , along with a harmless instruction $P$ , is fed into a MLLM for processing:", + "bbox": [ + 89, + 275, + 483, + 306 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\nY = \\operatorname {M L L M} \\left(I _ {\\text {c o m p}}, P\\right), \\tag {7}\n$$\n", + "text_format": "latex", + "bbox": [ + 205, + 320, + 482, + 338 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "where $Y$ denotes the textual output of the MLLM.", + "bbox": [ + 89, + 340, + 423, + 356 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "The use of $I_{\\mathrm{comp}}$ with contrasting visual elements maximizes distraction, effectively disrupting the model's internal alignment and coherence mechanisms. To complement $I_{\\mathrm{comp}}$ , we carefully design $P$ to further enhance the distraction effect. As illustrated in Figure 1, $P$ comprises three main components: (1) a role-guiding section, (2) a task-guiding section, and (3) a visual-guiding section. The role-guiding strategy is commonly employed across various LLM tasks. The second component of $P$ directs the model to process multiple sub-queries simultaneously, increasing task complexity and dispersing attention across multiple tasks instead of focusing on a single one. Finally, $P$ includes a misleading instruction, informing the model that the visual subimages may contain useful information, further diverting its attention. By presenting these inputs together, the approach reduces the model's ability to detect and mitigate harmful content, effectively bypassing its intrinsic safety mechanisms through multi-level distraction.", + "bbox": [ + 89, + 357, + 483, + 628 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "3.4. Distraction Distance Measurement", + "text_level": 1, + "bbox": [ + 89, + 637, + 393, + 652 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "To quantify the effectiveness of our multi-subimage input for distraction, we introduce a metric called Distraction Distance. This metric evaluates interactions among the query $Q$ and contrasting visual images $\\{I_c^j\\}_{j=1}^k$ , treating all as equally important nodes within a unified structure.", + "bbox": [ + 89, + 659, + 482, + 736 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Each node is represented by its CLIP-encoded vector. For a given node, we calculate the $\\mathrm{L}_2$ distance between its vector and the vectors of all other nodes. The sum of these $\\mathrm{L}_2$ distances for each node is calculated across all other nodes to obtain the Distraction Distance:", + "bbox": [ + 89, + 737, + 483, + 811 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\nD L = \\sum_ {i = 1} ^ {N} \\sum_ {j \\neq i} | | v _ {i} - v _ {j} | | _ {2}, \\tag {8}\n$$\n", + "text_format": "latex", + "bbox": [ + 194, + 819, + 482, + 863 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "where $N$ denotes the total number of nodes, and $v_{i}$ represents the CLIP-encoded vector of node $i$ . This approach ac", + "bbox": [ + 89, + 869, + 483, + 902 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "counts for both the number of nodes and their pairwise distances, providing a comprehensive measure of the dispersion and contrast within the multi-subimage structure. By maximizing this Distraction Distance, we amplify the overall distracting effect on MLLMs, thereby disrupting their ability to accurately detect harmful content.", + "bbox": [ + 511, + 90, + 905, + 181 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "4. Experiment", + "text_level": 1, + "bbox": [ + 511, + 199, + 638, + 217 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "In this section, we first introduced our experimental settings which include datasets and evaluation metrics. Then, we carried out quantitative and qualitative comparisons between our method and state-of-the-art MLLM jailbreaking approaches. Additionally, ablation studies are performed to validate the distraction effect on MLLM jailbreaking from various perspectives.", + "bbox": [ + 511, + 224, + 905, + 330 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "4.1. Experiment Setup", + "text_level": 1, + "bbox": [ + 511, + 344, + 691, + 359 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Dataset. The datasets are primarily utilized to provide or construct visual inputs tailored to the requirements of different methods. For evaluating Hades, we followed the experimental setting outlined in its original paper, utilizing the Hades dataset [12] to supply the visual inputs. This dataset comprises five representative categories related to real-world visual information: (1) Violence, (2) Financial, (3) Privacy, (4) Self-Harm, and (5) Animal. Each category contains 150 queries, resulting in a total of 750 harmful queries. For evaluating CS-DJ, we employed LLaVACC3M-Pretrain595K to retrieve the contrasting images. To improve retrieval efficiency, we randomly selected a subset of 10,000 images for the retrieval dataset, using a different random seed for each experiment to ensure diversity. All harmful queries used throughout the experiments were derived from the Hades dataset.", + "bbox": [ + 511, + 367, + 906, + 608 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Victim Models. In our main experiments, We evaluated the performance of the four most popular closed-source MLLMs: GPT-4o-Mini, GPT-4o, GPT-4V [1], and Gemini-1.5-Flash [27]. The specific versions are as follows: GPT-4o: GPT-4o-2024-08-06, GPT-4o-Mini: GPT-4o-Mini-2024-07-18, GPT-4V: GPT-4-1106- vision-preview, and Gemini-1.5-flash: Gemini-1.5-flash-001.", + "bbox": [ + 511, + 609, + 905, + 715 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Evaluation Metrics. To assess CS-DJ, we employ two primary evaluation metrics: Attack Success Rate (ASR) and Ensemble Attack Success Rate (EASR). ASR [6, 12] quantifies the proportion of successful jailbreak attempts by assessing whether a model's response meets predefined harmfulness criteria. The harmfulness criteria come from the security discriminator Beaver-Dam-7b [8], used to detect both the harmfulness and helpfulness of responses to", + "bbox": [ + 511, + 715, + 905, + 838 + ], + "page_idx": 4 + }, + { + "type": "page_footnote", + "text": "https://huggingface.co/datasets/liuhaotian/LLaVA-CC3M-Pretrain-595K", + "bbox": [ + 511, + 851, + 905, + 875 + ], + "page_idx": 4 + }, + { + "type": "page_footnote", + "text": "Since Gemini-1.0-Pro-Vision has been deprecated on July 12, 2024, we used its upgraded version, Gemini-1.5-Flash.", + "bbox": [ + 514, + 875, + 905, + 900 + ], + "page_idx": 4 + }, + { + "type": "page_number", + "text": "9471", + "bbox": [ + 480, + 944, + 513, + 955 + ], + "page_idx": 4 + }, + { + "type": "table", + "img_path": "images/a542535a26e8d23817dc8422c77ffdda0fa6c0d2a2bb0b1a16a228ddb728d1b4.jpg", + "table_caption": [], + "table_footnote": [], + "table_body": "
Victim ModelMethodAnimalFinancialPrivacySelf-HarmViolenceAverage (%)
Attack Success Rate (ASR) ↑
GPT-4o-miniHades4.7711.666.881.555.556.08
CS-DJ26.5580.0077.1134.7770.5557.80 (+51.72)
GPT-4oHades2.3310.117.551.885.665.51
CS-DJ15.3370.7758.5515.7750.7742.24 (+36.73)
GPT-4VHades8.1132.5523.888.0029.1120.33
CS-DJ11.6676.2259.7722.2257.3345.44 (+25.11)
Gemini-1.5-FlashHades2.6623.4413.111.775.009.20
CS-DJ22.5588.8879.1147.0083.0064.11 (+54.91)
Ensemble Attack Success Rate (EASR) ↑
GPT-4o-miniHades10.6621.3313.334.6610.0012.00
CS-DJ54.0092.6694.6652.0092.6677.20 (+65.20)
GPT-4oHades5.3316.0012.006.0011.3310.13
CS-DJ35.3392.6683.3337.3380.6665.86 (+55.73)
GPT-4VHades18.6663.3354.0024.6664.6645.06
CS-DJ34.6694.0089.3352.089.3371.86 (+26.80)
Gemini-1.5-FlashHades4.6634.0020.004.6610.6614.79
CS-DJ51.3397.3395.3366.0097.3381.46 (+66.67)
", + "bbox": [ + 169, + 88, + 826, + 425 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Table 1. ASR and EASR results of CS-DJ and Hades on four closed-source MLLMs across different categories.", + "bbox": [ + 166, + 435, + 826, + 448 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "their corresponding queries. EASR [36] measures the effectiveness of jailbreak templates by calculating the proportion of queries where at least one template successfully bypasses the target MLLM's defenses. While ASR focuses on the success rate of individual templates, EASR measures the combined success rate of a group of templates.", + "bbox": [ + 89, + 469, + 482, + 559 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "4.2. Main Experiment", + "text_level": 1, + "bbox": [ + 89, + 574, + 264, + 589 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "To assess the effectiveness of CS-DJ, we conducted a comparison between CS-DJ and Hades [12], the state-of-the-art closed-source MLLM jailbreak attack. For a fair comparison, we adopted the official implementation of Hades and performed 6 groups of experiments to collect data.", + "bbox": [ + 89, + 597, + 482, + 672 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Table 1 presents the jailbreak ASR and EASR results of CS-DJ and Hades across the four widely used closed-source MLLMs. The results across the five categories indicate that the four closed-source MLLMs exhibit stronger defenses against harmful queries in the Animal and Self-Harm categories while showing greater vulnerability in the Financial, Privacy, and Violence categories. Compared to their ASR results, Hades and CS-DJ achieve average improvements of $10.22\\%$ and $21.70\\%$ in EASR, respectively, indicating that increasing the number of trials can enhance the jailbreak success rate. Additionally, for both ASR and EASR, CS-DJ consistently outperforms Hades across all five categories for each evaluated MLLM. Overall, compared to Hades, CS-DJ achieves maximum improvements of $54.91\\%$ in ASR and $66.67\\%$ in EASR. Furthermore, CS-DJ attains", + "bbox": [ + 89, + 674, + 482, + 900 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "average success rates of $52.40\\%$ for ASR and $74.10\\%$ for EASR. These observations highlight CS-DJ's superior effectiveness as a jailbreak method for bypassing the defenses of closed-source MLLMs. Jailbreaking examples can be found in the appendix.", + "bbox": [ + 511, + 469, + 903, + 545 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "A detailed analysis of Hades' results across the four models reveals that GPT-4V outperforms the other three models in both ASR and EASR. In contrast, the results of CS-DJ indicate that Gemini-1.5-Flash achieved the highest performance. These findings suggest that the impact of visual input harmfulness and distraction varies across different models. In a word, the results demonstrate that distracting the model's attention is a more effective strategy for enhancing jailbreak success rates. More quantitative and qualitative experimental results, including those tested on open-source models, are provided in the appendix.", + "bbox": [ + 511, + 546, + 905, + 712 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "For the subsequent experiments, we selected the most widely used model, GPT-40, for testing. All experiments were conducted as a single trial. Therefore, only ASR results are presented in the following sections.", + "bbox": [ + 511, + 713, + 903, + 773 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "4.3. Impact of Query Decomposition", + "text_level": 1, + "bbox": [ + 511, + 787, + 795, + 801 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Here, we aim to explore the impact of query decomposition on MLLM jailbreak performance. To validate its effectiveness, we first use images derived from raw queries (RQ) as visual inputs for MLLMs, serving as a baseline for comparison with the results obtained using query decomposition. We further examine the effect of the number", + "bbox": [ + 511, + 810, + 903, + 900 + ], + "page_idx": 5 + }, + { + "type": "page_number", + "text": "9472", + "bbox": [ + 480, + 944, + 514, + 955 + ], + "page_idx": 5 + }, + { + "type": "table", + "img_path": "images/5cf35014dca1eaef7e072a61051c60c23f9696b58dc2fb910e66512bf7d5378b.jpg", + "table_caption": [], + "table_footnote": [], + "table_body": "
SettingAni.Fin.Priv.Self-H.Viol.ASR (%)
RQ0.676.005.3304.003.20
3SQ4.0028.6634.664.6622.0018.80
6SQ8.0045.3344.0010.6641.3329.86
9SQ4.6644.6644.667.3336.0027.46
", + "bbox": [ + 94, + 88, + 482, + 170 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Table 2. GPT-4o's ASR results of CS-DJ under different query decomposition settings. RQ denotes using the image transformation of raw queries as the visual input for MLLMs. 3SQ, 6SQ, and 9SQ indicate that the raw query is decomposed into 3, 6, and 9 sub-queries, respectively, which are then transformed into sub-images to compose the final visual input.", + "bbox": [ + 89, + 180, + 482, + 265 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "of sub-queries on the jailbreak success rate by decomposing the raw query into 3, 6, and 9 sub-queries, denoted as 3SQ, 6SQ, and 9SQ, respectively. Given that MLLMs typically resize image inputs to fixed dimensions, we aim to minimize distortion from aspect ratio changes during resizing and avoid potential limitations of the MLLM's image encoder. Therefore, we maintain a fixed column count of 3 when composing the final input image. As a result, the 3SQ, 6SQ, and 9SQ settings increase the number of rows while keeping the number of columns constant.", + "bbox": [ + 88, + 290, + 482, + 441 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "The quantitative results are presented in Table 2. The results from RQ demonstrate that the MLLM exhibits strong defenses against original jailbreak queries. Even when the raw query is transformed into an image input, it remains highly challenging to bypass the MLLM's security mechanisms. In contrast, 3SQ outperforms RQ across all five categories, yielding an overall improvement of $15.60\\%$ . This indicates that query decomposition effectively enhances the jailbreak success rate for MLLMs. By introducing a distributional shift in the queries, query decomposition creates structural distraction, thereby enabling more effective bypassing of the model's defenses. Furthermore, increasing the number of decomposed sub-queries to 6 results in an additional $11.06\\%$ improvement in the jailbreak success rate. This suggests that a higher number of sub-images transformed from sub-queries can further enhance the distraction effect on the MLLM. However, when the number of sub-queries is increased to 9, a slight decline in the success rate is observed. This decrease may be attributed to two factors: (1) the increased number of sub-queries raises task complexity, resulting in greater demands on the model's comprehension and its ability to handle complex tasks; (2) limitations in the model's image encoding capabilities may prevent accurate interpretation of the sub-queries; and (3) over-decomposition may cause the model to recognize the original intent of the query, thereby reducing the effectiveness of the jailbreak attempt.", + "bbox": [ + 91, + 445, + 482, + 851 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Overall, the structural distraction introduced by query decomposition can enhance the jailbreak success rate of MLLMs. Furthermore, the number of sub-queries also", + "bbox": [ + 89, + 854, + 483, + 900 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "plays a significant role in influencing the results. Given the limited capabilities of smaller MLLMs and the need to incorporate additional visual subimages, this work focuses on evaluating the performance of the CS-DJ with three subqueries.", + "bbox": [ + 511, + 90, + 903, + 167 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/09053fcdce5461237480ccb35f5b8989a65444f0a56c838898ee18ae3df8172d.jpg", + "image_caption": [ + "4.4. Impact of Contrasting Visuals", + "Figure 3. GPT-4o's evaluation results for varying quantities of visual subimages." + ], + "image_footnote": [], + "bbox": [ + 526, + 210, + 890, + 362 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "To validate the effectiveness of contrasting visuals, we conduct experiments from two perspectives: (1) the impact of distraction induced by varying the number of subimages, and (2) the effect of inter-subimage distraction with a fixed number of subimages. Since altering the number of subimages directly affects the layout of the final composite image, we maintain a fixed column count of 3 and adjust the number of rows to control the number of subimages. This strategy reduces distortion due to excessive changes in aspect ratio and mitigates potential limitations in the MLLM's image encoding capabilities. We conduct different experiments with various configurations, involving 0, 3, 6, 9, and 12 Contrasting visual SubImages, corresponding to 3SQ, 3SQ+3CSI, 3SQ+6CSI, 3SQ+9CSI, and 3SQ+12CSI, respectively. To explore the influence of inter-subimage distraction on jailbreak success rates, we conduct experiments using three settings: retrieving 9 most contrasting images, retrieving 9 most similar images, and using a single, most similar image from the dataset for all 9 subimages, denoting 3SQ+9CSI, 3SQ+9SSI, and 3SQ+9SinSI, respectively.", + "bbox": [ + 511, + 417, + 906, + 719 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "The experimental results on the impact of visual subimage quantity are illustrated in Figure 3. As 3SQ does not contain any visual subimages, its Distraction Distance is 0, resulting in the lowest ASR. Increasing the number of visual subimages leads to a rise in ASR. Specifically, compared to 3SQ, $3\\mathrm{SQ} + 3\\mathrm{CSI}$ obtains a $12.8\\%$ increase in ASR, demonstrating that contrasting visual subimages can effectively distract the MLLM and bypass its internal defense mechanisms. As the number of visual subimages increases, both the Distraction Distance and the ASR grow correspondingly. However, the rate of ASR improvement diminishes as the number of subimages continues to rise. When the", + "bbox": [ + 511, + 719, + 908, + 900 + ], + "page_idx": 6 + }, + { + "type": "page_number", + "text": "9473", + "bbox": [ + 480, + 944, + 514, + 955 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "number of subimages reaches 9, further additions yield limited gains in ASR. Given the limited encoding capabilities of the visual encoders in smaller MLLMs, we use 9 visual subimages as the default setting for our experiments.", + "bbox": [ + 89, + 90, + 480, + 151 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Table 3 presents the experimental results on the impact of inter-subimage distraction. We evaluate three different subimage selection strategies. The results show that $3\\mathrm{SQ} + 9\\mathrm{SinSI}$ has the lowest distraction distance, corresponding to the lowest ASR. In contrast, $3\\mathrm{SQ} + 9\\mathrm{CSI}$ achieves an $11.20\\%$ improvement in ASR compared to $3\\mathrm{SQ} + 9\\mathrm{SSI}$ , along with a further increase in distraction distance. These findings indicate that, for a fixed number of subimages, greater distraction among subimages leads to a higher jailbreak success rate for MLLMs. Additionally, an analysis of the results from Figure 3 and Table 3 reveals that distraction distance accurately reflects changes in ASR trends only when varying the subimage construction strategy along a single dimension, such as the number of subimages or subimage selection strategies. This limitation may stem from differences between the semantic space of our current feature extractor, CLIP, and that of the MLLM. As a result, inter-subimage distraction distance cannot always be accurately measured using the CLIP model. Future work will focus on developing a more comprehensive metric for measuring distraction distance.", + "bbox": [ + 91, + 152, + 482, + 467 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Table 4 shows the impact of information complexity of visual subimages. The ASR of $3\\mathrm{SQ} + 9\\mathrm{RNI}$ is close to that of 3SQ, but significantly lower than $3\\mathrm{SQ} + 9\\mathrm{CSI}$ , indicating that MLLMs exhibit strong recognition capabilities for noise images with minimal informational content. Therefore, only subimages with higher information complexity can effectively distract the model.", + "bbox": [ + 89, + 468, + 480, + 571 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "In summary, constructing multiple contrasting visual subimages effectively enhances the distraction of MLLMs, leading to improved jailbreak success rates.", + "bbox": [ + 89, + 574, + 480, + 619 + ], + "page_idx": 7 + }, + { + "type": "table", + "img_path": "images/47b246fbd831a616c2bc952c82d26fe480a7a890571209bf25a4034db80ba354.jpg", + "table_caption": [], + "table_footnote": [], + "table_body": "
SettingASR (%)Distraction Distance
3SQ+9SinSI24.66399.08
3SQ+9SSI32.531378.77
3SQ+9CSI43.731878.83
", + "bbox": [ + 125, + 631, + 444, + 700 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "4.5. Impact of Instruction", + "text_level": 1, + "bbox": [ + 89, + 803, + 290, + 819 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "To complement the constructed multi-subimage input, we carefully designed the instruction $P$ . To evaluate the impact of different components of $P$ , we conducted three experiments: (1) using only the task-guiding component, (2) adding the role-guiding component, and (3) further incorpo", + "bbox": [ + 89, + 825, + 480, + 900 + ], + "page_idx": 7 + }, + { + "type": "table", + "img_path": "images/b0be4a81e16492acc68118121923ce469c635abb5d982aaf028a918be9024e19.jpg", + "table_caption": [ + "Table 3. ASR of GPT-4o and Distraction Distance for CS-DJ under different visual subimage selection strategies. 9SinSI indicates using a single, most similar image from the dataset for all 9 subimages. 9CSI and 9SSI denote retrieving 9 most contrasting and similar images to compose the final visual input, respectively." + ], + "table_footnote": [], + "table_body": "
SettingAni.Fin.Priv.Self-H.Viol.ASR (%)
3SQ4.0028.6634.664.6622.0018.80
3SQ+9RNI5.3332.6633.334.6624.0020.00
3SQ+9CSI19.3373.3356.6618.0051.3343.73
", + "bbox": [ + 517, + 89, + 903, + 152 + ], + "page_idx": 7 + }, + { + "type": "table", + "img_path": "images/e4fe6f2396a2bd35fd7170e5d44a3d092987133ae9caf1e940585bf3c47dbff1.jpg", + "table_caption": [ + "Table 4. GPT-4o's ASR results of CS-DJ under different settings. 9RNI denotes using 9 random noise images to compose the final visual input." + ], + "table_footnote": [], + "table_body": "
SettingAni.Fin.Priv.Self-H.Viol.ASR (%)
task-guiding10.0055.3339.3313.3342.6632.00
+ role-guiding15.3367.3350.0013.3352.0039.73
+ visual-guiding19.3373.3356.6618.0051.3343.73
", + "bbox": [ + 517, + 218, + 903, + 277 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Table 5. GPT-4o's ASR results of CS-DJ under the 3SQ+9CSI configuration with different instruction $P$ settings.", + "bbox": [ + 511, + 285, + 903, + 314 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "rating the visual-guiding component. The results, presented in Table 5, show that using only the task-guiding component achieved the lowest ASR but still outperformed Hades by $26.49\\%$ . This indicates that the visual distraction from the multi-subimage input, combined with the task distraction of handling multiple tasks simultaneously, significantly contributes to improving jailbreak success rates. Adding the role-guiding component further increased the ASR by $7.73\\%$ , while incorporating the visual-guiding component led to an additional $4.00\\%$ improvement. These findings demonstrate that targeted optimization of $P$ for multisubimage visual inputs can further disperse the model's attention, enhancing the success rate of the jailbreak.", + "bbox": [ + 511, + 328, + 906, + 525 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "5. Conclusion", + "text_level": 1, + "bbox": [ + 511, + 541, + 633, + 556 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "In this paper, we investigate the visual vulnerabilities in jailbreak attacks on MLLMs. Our findings show that the complexity of visual inputs significantly influences attack success. We introduce the Distraction Hypothesis and propose the CS-DJ framework, which leverages distraction-based strategies to exploit these vulnerabilities in MLLMs. Our approach integrates this insight by two core components: structured distraction through query decomposition and visual-enhanced distraction using contrasting subimages. Extensive experiments conducted on five representative scenarios and four popular closed-source MLLMs demonstrated that CS-DJ outperforms state-of-the-art jailbreak methods, validating the efficacy of distraction-based approaches in bypassing MLLM defenses.", + "bbox": [ + 511, + 566, + 906, + 777 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Acknowledgment. This research is supported by Guangzhou Basic and Applied Basic Research Project (No: 2025A04J3536, 2025A03J3128, 2024A03J0324), Guangzhou University Research Project (No: RQ2021013), Guangdong Basic and Applied Basic Research Foundation (No: 2023A1515110077), NSFC (No: 62402131), and the Postdoctoral Fellowship Program of CPSF (No: GZC20230595).", + "bbox": [ + 511, + 779, + 908, + 900 + ], + "page_idx": 7 + }, + { + "type": "page_number", + "text": "9474", + "bbox": [ + 480, + 944, + 514, + 955 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "References", + "text_level": 1, + "bbox": [ + 91, + 89, + 187, + 104 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[1] Josh Achiam, Steven Adler, Sandhini Agarwal, Lama Ahmad, Ilge Akkaya, Florencia Leoni Aleman, Diogo Almeida, Janko Altenschmidt, Sam Altman, Shyamal Anadkat, et al. Gpt-4 technical report. arXiv preprint arXiv:2303.08774, 2023.5", + "[2] Yulin Chen, Haoran Li, Zihao Zheng, and Yangqiu Song. Bathe: Defense against the jailbreak attack in multimodal large language models by treating harmful instruction as backdoor trigger. arXiv preprint arXiv:2408.09093, 2024. 2", + "[3] Can Cui, Yunsheng Ma, Xu Cao, Wenqian Ye, Yang Zhou, Kaizhao Liang, Jintai Chen, Juanwu Lu, Zichong Yang, Kuei-Da Liao, et al. A survey on multimodal large language models for autonomous driving. In Proceedings of the IEEE/CVF Winter Conference on Applications of Computer Vision, pages 958-979, 2024. 1", + "[4] Ameet Deshpande, Vishvak Murahari, Tanmay Rajpurohit, Ashwin Kalyan, and Karthik Narasimhan. Toxicity in chatgpt: Analyzing persona-assigned language models. In Findings of the Association for Computational Linguistics: EMNLP 2023, pages 1236-1270, 2023. 2", + "[5] Yassir Fathullah, Chunyang Wu, Egor Lakomkin, Junteng Jia, Yuan Shangguan, Ke Li, Jinxi Guo, Wenhan Xiong, Jay Mahadeokar, Ozlem Kalinli, et al. Prompting large language models with speech recognition abilities. In ICASSP 2024-2024 IEEE International Conference on Acoustics, Speech and Signal Processing (ICASSP), pages 13351-13355. IEEE, 2024. 3", + "[6] Yichen Gong, Delong Ran, Jinyuan Liu, Conglei Wang, Tianshuo Cong, Anyu Wang, Sisi Duan, and Xiaoyun Wang. Figstep: Jailbreaking large vision-language models via typographic visual prompts. arXiv preprint arXiv:2311.05608, 2023. 2, 3, 5, 12", + "[7] Aaron Hurst, Adam Lerer, Adam P Goucher, Adam Perelman, Aditya Ramesh, Aidan Clark, AJ Ostrow, Akila Weli-hinda, Alan Hayes, Alec Radford, et al. Gpt-4o system card. arXiv preprint arXiv:2410.21276, 2024. 3", + "[8] Jiaming Ji, Mickel Liu, Josef Dai, Xuehai Pan, Chi Zhang, Ce Bian, Boyuan Chen, Ruiyang Sun, Yizhou Wang, and Yaodong Yang. Beavertails: Towards improved safety alignment of llm via a human-preference dataset. Advances in Neural Information Processing Systems, 36, 2024. 5, 12", + "[9] Haibo Jin, Leyang Hu, Xinuo Li, Peiyan Zhang, Chonghan Chen, Jun Zhuang, and Haohan Wang. Jailbreakzoo: Survey, landscapes, and horizons in jailbreaking large language and vision-language models. arXiv preprint arXiv:2407.01599, 2024.3", + "[10] Daniel Martin Katz, Michael James Bommarito, Shang Gao, and Pablo Arredondo. Gpt-4 passes the bar exam. Philosophical Transactions of the Royal Society A, 382(2270): 20230254, 2024. 3", + "[11] Bo Li, Yuanhan Zhang, Dong Guo, Renrui Zhang, Feng Li, Hao Zhang, Kaichen Zhang, Peiyuan Zhang, Yanwei Li, Ziwei Liu, et al. Llava-onevision: Easy visual task transfer. arXiv preprint arXiv:2408.03326, 2024. 12" + ], + "bbox": [ + 93, + 114, + 483, + 900 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[12] Yifan Li, Hangyu Guo, Kun Zhou, Wayne Xin Zhao, and JiRong Wen. Images are achilles' heel of alignment: Exploiting visual vulnerabilities for jailbreaking multimodal large language models. In European Conference on Computer Vision. Springer, 2024. 1, 2, 3, 5, 6", + "[13] Haotian Liu, Chunyuan Li, Yuheng Li, and Yong Jae Lee. Improved baselines with visual instruction tuning. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 26296-26306, 2024. 3", + "[14] Haotian Liu, Chunyuan Li, Qingyang Wu, and Yong Jae Lee. Visual instruction tuning. Advances in neural information processing systems, 36, 2024. 3", + "[15] Tong Liu, Yingjie Zhang, Zhe Zhao, Yinpeng Dong, Guozhu Meng, and Kai Chen. Making them ask and answer: Jailbreaking large language models in few queries via disguise and reconstruction. In 33rd USENIX Security Symposium (USENIX Security 24), pages 4711-4728, 2024. 2", + "[16] Xin Liu, Yichen Zhu, Yunshi Lan, Chao Yang, and Yu Qiao. Query-relevant images jailbreak large multi-modal models. arXiv preprint arXiv:2311.17600, 2023. 2, 3", + "[17] Xin Liu, Yichen Zhu, Jindong Gu, Yunshi Lan, Chao Yang, and Yu Qiao. Mm-safetybench: A benchmark for safety evaluation of multimodal large language models. In European Conference on Computer Vision, pages 386-403. Springer, 2024. 2, 3, 12", + "[18] Dong Lu, Zhiqiang Wang, Teng Wang, Weili Guan, Hongchang Gao, and Feng Zheng. Set-level guidance attack: Boosting adversarial transferability of vision-language pre-training models. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 102-111, 2023. 2", + "[19] Siyuan Ma, Weidi Luo, Yu Wang, Xiaogeng Liu, Muhao Chen, Bo Li, and Chaowei Xiao. Visual-roleplay: Universal jailbreak attack on multimodal large language models via role-playing image character. arXiv preprint arXiv:2405.20773, 2024. 2", + "[20] Zhenxing Niu, Haodong Ren, Xinbo Gao, Gang Hua, and Rong Jin. Jailbreaking attack against multimodal large language model. arXiv preprint arXiv:2402.02309, 2024. 3", + "[21] Long Ouyang, Jeffrey Wu, Xu Jiang, Diogo Almeida, Carroll Wainwright, Pamela Mishkin, Chong Zhang, Sandhini Agarwal, Katarina Slama, Alex Ray, et al. Training language models to follow instructions with human feedback. Advances in neural information processing systems, 35:27730-27744, 2022. 2", + "[22] Xiangyu Qi, Kaixuan Huang, Ashwinee Panda, Peter Henderson, Mengdi Wang, and Prateek Mittal. Visual adversarial examples jailbreak aligned large language models. In Proceedings of the AAAI Conference on Artificial Intelligence, pages 21527-21536, 2024. 2, 3", + "[23] Alec Radford, Jong Wook Kim, Chris Hallacy, Aditya Ramesh, Gabriel Goh, Sandhini Agarwal, Girish Sastry, Amanda Askell, Pamela Mishkin, Jack Clark, et al. Learning transferable visual models from natural language supervision. In International conference on machine learning, pages 8748-8763. PMLR, 2021. 4", + "[24] Erfan Shayegani, Yue Dong, and Nael Abu-Ghazaleh. Jailbreak in pieces: Compositional adversarial attacks on multi-" + ], + "bbox": [ + 93, + 92, + 903, + 900 + ], + "page_idx": 8 + }, + { + "type": "page_number", + "text": "9475", + "bbox": [ + 482, + 944, + 514, + 955 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "modal language models. In The Twelfth International Conference on Learning Representations, 2023. 3", + "[25] Wentan Tan, Changxing Ding, Jiayu Jiang, Fei Wang, Yibing Zhan, and Dapeng Tao. Harnessing the power of mllms for transferable text-to-image person reid. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 17127-17137, 2024. 2", + "[26] Xijia Tao, Shuai Zhong, Lei Li, Qi Liu, and Lingpeng Kong. Imgtrojan: Jailbreaking vision-language models with one image. arXiv preprint arXiv:2403.02910, 2024. 2", + "[27] Gemini Team, Petko Georgiev, Ving Ian Lei, Ryan Burnell, Libin Bai, Anmol Gulati, Garrett Tanzer, Damien Vincent, Zhufeng Pan, Shibo Wang, et al. Gemini 1.5: Unlocking multimodal understanding across millions of tokens of context. arXiv preprint arXiv:2403.05530, 2024. 3, 5", + "[28] Qwen Team. Qwen2.5: A party of foundation models, 2024. 3", + "[29] Haoqin Tu, Chenhang Cui, Zijun Wang, Yiyang Zhou, Bingchen Zhao, Junlin Han, Wangchunshu Zhou, Huaxiu Yao, and Cihang Xie. How many unicorns are in this image? a safety evaluation benchmark for vision llms. arXiv preprint arXiv:2311.16101, 2023. 2", + "[30] Haodi Wang, Kai Dong, Zhilei Zhu, Haotong Qin, Aishan Liu, Xiaolin Fang, Jiakai Wang, and Xianglong Liu. Transferable multimodal attack on vision-language pre-training models. In 2024 IEEE Symposium on Security and Privacy (SP), pages 102-102. IEEE Computer Society, 2024. 2", + "[31] Jiaqi Wang, Hanqi Jiang, Yiheng Liu, Chong Ma, Xu Zhang, Yi Pan, Mengyuan Liu, Peiran Gu, Sichen Xia, Wenjun Li, et al. A comprehensive review of multimodal large language models: Performance and challenges across different tasks. arXiv preprint arXiv:2408.01319, 2024. 3", + "[32] Ruofan Wang, Xingjun Ma, Hanxu Zhou, Chuanjun Ji, Guangnan Ye, and Yu-Gang Jiang. White-box multimodal jailbreaks against large vision-language models. In Proceedings of the 32nd ACM International Conference on Multimedia, pages 6920-6928, 2024. 2", + "[33] Yizhong Wang, Yeganeh Kordi, Swaroop Mishra, Alisa Liu, Noah A Smith, Daniel Khashabi, and Hannaneh Hajishirzi. Self-instruct: Aligning language models with self-generated instructions. In The 61st Annual Meeting Of The Association For Computational Linguistics, 2023. 2", + "[34] Alexander Wei, Nika Haghtalab, and Jacob Steinhardt. Jailbroken: How does llm safety training fail? Advances in Neural Information Processing Systems, 36, 2024. 2", + "[35] Yuanwei Wu, Xiang Li, Yixin Liu, Pan Zhou, and Lichao Sun. Jailbreaking gpt-4v via self-adversarial attacks with system prompts. arXiv preprint arXiv:2311.09127, 2023. 2", + "[36] Jiahao Yu, Xingwei Lin, Zheng Yu, and Xinyu Xing. {LLM-Fuzzer}: Scaling assessment of large language model jailbreaks. In 33rd USENIX Security Symposium (USENIX Security 24), pages 4657-4674, 2024. 2, 6", + "[37] Qifan Yu, Juncheng Li, Longhui Wei, Liang Pang, Wentao Ye, Bosheng Qin, Siliang Tang, Qi Tian, and Yueting Zhuang. Hallucidobot: Mitigating hallucinatory toxicity in visual instruction data. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 12944-12953, 2024. 2" + ], + "bbox": [ + 91, + 92, + 483, + 900 + ], + "page_idx": 9 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[38] Hangfan Zhang, Zhimeng Guo, Huaisheng Zhu, Bochuan Cao, Lu Lin, Jinyuan Jia, Jinghui Chen, and Dinghao Wu. Jailbreak open-sourced large language models via enforced decoding. In Proceedings of the 62nd Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers), pages 5475-5493, 2024. 2", + "[39] Jingyi Zhang, Jiaxing Huang, Sheng Jin, and Shijian Lu. Vision-language models for vision tasks: A survey. IEEE Transactions on Pattern Analysis and Machine Intelligence, 2024. 3", + "[40] Yunqing Zhao, Tianyu Pang, Chao Du, Xiao Yang, Chongxuan Li, Ngai-Man Man Cheung, and Min Lin. On evaluating adversarial robustness of large vision-language models. Advances in Neural Information Processing Systems, 36, 2024. 2", + "[41] Deyao Zhu, Jun Chen, Xiaoqian Shen, Xiang Li, and Mohamed Elhoseiny. Minigpt-4: Enhancing vision-language understanding with advanced large language models. arXiv preprint arXiv:2304.10592, 2023. 3" + ], + "bbox": [ + 516, + 92, + 903, + 358 + ], + "page_idx": 9 + }, + { + "type": "page_number", + "text": "9476", + "bbox": [ + 482, + 944, + 516, + 955 + ], + "page_idx": 9 + } +] \ No newline at end of file diff --git a/2025/Distraction is All You Need for Multimodal Large Language Model Jailbreaking/1375bb02-e050-41f2-9b1d-ed16de66c742_model.json b/2025/Distraction is All You Need for Multimodal Large Language Model Jailbreaking/1375bb02-e050-41f2-9b1d-ed16de66c742_model.json new file mode 100644 index 0000000000000000000000000000000000000000..ead580b7b5932a8565cfed37577750b8b523ac14 --- /dev/null +++ b/2025/Distraction is All You Need for Multimodal Large Language Model Jailbreaking/1375bb02-e050-41f2-9b1d-ed16de66c742_model.json @@ -0,0 +1,1848 @@ +[ + [ + { + "type": "header", + "bbox": [ + 0.107, + 0.003, + 0.182, + 0.043 + ], + "angle": 0, + "content": "CVF" + }, + { + "type": "header", + "bbox": [ + 0.237, + 0.001, + 0.812, + 0.047 + ], + "angle": 0, + "content": "This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore." + }, + { + "type": "title", + "bbox": [ + 0.096, + 0.131, + 0.903, + 0.154 + ], + "angle": 0, + "content": "Distraction is All You Need for Multimodal Large Language Model Jailbreaking" + }, + { + "type": "text", + "bbox": [ + 0.278, + 0.18, + 0.718, + 0.218 + ], + "angle": 0, + "content": "Zuopeng Yang\\(^{1*}\\), Jiluan Fan\\(^{1*}\\), Anli Yan\\(^{1\\dagger}\\), Erdun Gao\\(^{3}\\), Xin Lin\\(^{1\\dagger}\\), Tao Li\\(^{2}\\), Kanghua Mo\\(^{1}\\), Changyu Dong\\(^{1}\\)" + }, + { + "type": "text", + "bbox": [ + 0.22, + 0.222, + 0.782, + 0.26 + ], + "angle": 0, + "content": "\\(^{1}\\)Guangzhou University \\(^{2}\\)Shanghai Jiao Tong University \\(^{3}\\)Australian Institute for Machine Learning, The University of Adelaide" + }, + { + "type": "image", + "bbox": [ + 0.134, + 0.279, + 0.862, + 0.478 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.089, + 0.482, + 0.907, + 0.54 + ], + "angle": 0, + "content": "Figure 1. Jailbreaking examples of Hades [12] and our proposed CS-DJ against GPT-4. When using Hades, GPT-4 effectively rejects harmful questions. In contrast, CS-DJ successfully bypasses the model's defense mechanisms by constructing multi-subimage visual inputs to distract the model, resulting in the acquisition of harmful responses. The harmful information is highlighted in red, while key prompts are emphasized in bold." + }, + { + "type": "title", + "bbox": [ + 0.249, + 0.566, + 0.327, + 0.581 + ], + "angle": 0, + "content": "Abstract" + }, + { + "type": "text", + "bbox": [ + 0.089, + 0.6, + 0.486, + 0.873 + ], + "angle": 0, + "content": "Multimodal Large Language Models (MLLMs) bridge the gap between visual and textual data, enabling a range of advanced applications. However, complex internal interactions among visual elements and their alignment with text can introduce vulnerabilities, which may be exploited to bypass safety mechanisms. To address this, we analyze the relationship between image content and task and find that the complexity of subimages, rather than their content, is key. Building on this insight, we propose the Distraction Hypothesis, followed by a novel framework called Contrasting Subimage Distraction Jailbreaking (CS-DJ), to achieve jailbreaking by disrupting MLLMs alignment through multi-level distraction strategies. CS-DJ consists of two components: structured distraction, achieved through query decomposition that induces a distributional shift by fragmenting harmful prompts into subqueries, and visual-enhanced distraction, realized by constructing contrasting subimages to disrupt the interactions" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.553, + 0.909, + 0.748 + ], + "angle": 0, + "content": "among visual elements within the model. This dual strategy disperses the model's attention, reducing its ability to detect and mitigate harmful content. Extensive experiments across five representative scenarios and four popular closed-source MLLMs, including GPT-4o-mini, GPT-4o, GPT-4V, and Gemini-1.5-Flash, demonstrate that CS-DJ achieves average success rates of \\(52.40\\%\\) for the attack success rate and \\(74.10\\%\\) for the ensemble attack success rate. These results reveal the potential of distraction-based approaches to exploit and bypass MLLMs' defenses, offering new insights for attack strategies. Our code is available at https://github.com/TeamPigeonLab/CS-DJ." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.75, + 0.909, + 0.781 + ], + "angle": 0, + "content": "Warning: This paper contains unfiltered content generated by MLLMs that may be offensive to readers." + }, + { + "type": "title", + "bbox": [ + 0.514, + 0.814, + 0.646, + 0.829 + ], + "angle": 0, + "content": "1. Introduction" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.84, + 0.907, + 0.901 + ], + "angle": 0, + "content": "Since the early 2020s, Multimodal Large Language Models (MLLMs) have emerged as a powerful tool for bridging natural language processing and computer vision [3], owing to their ability to integrate and process multimodal informa" + }, + { + "type": "page_footnote", + "bbox": [ + 0.115, + 0.887, + 0.363, + 0.901 + ], + "angle": 0, + "content": "*Equal Contribution. † Corresponding authors." + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.516, + 0.957 + ], + "angle": 0, + "content": "9467" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.09, + 0.092, + 0.482, + 0.318 + ], + "angle": 0, + "content": "tion efficiently [25]. Despite their impressive performance, one notable limitation of these methods is that the training materials may contain toxic content, potentially leading to inappropriate or harmful content [4, 38], the leakage of sensitive information [36], and even serious vulnerabilities in the model security and privacy protection. To this end, significant efforts have been devoted to using reinforcement learning from human feedback (RLHF) to align LLMs outputs, ensuring readiness for deployment in high-stakes domains [21, 33]. However, beyond Large Language Models (LLMs), the integration of visual inputs in MLLMs introduces a new challenge: securing models against vulnerabilities arising from newly integrated visual modalities, which remains a significant issue and lacks a comprehensive understanding [17, 18, 37]" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.323, + 0.483, + 0.534 + ], + "angle": 0, + "content": "In this study, rather than following the human feedback alignment research at the defense side, we focus on exploring an alternative approach, namely jailbreak attacks [19, 34, 36]. This line of research also seeks to provide insights into how safety concerns emerge, while, considering the attacker's perspective. Specifically, it aims to actively uncover the defense vulnerabilities in current model by deliberately crafting inputs to manipulate MLLMs, leading to unintended and potentially harmful outputs [15]. Motivated by the fact that most current defense mechanisms rely on RLHF and are trained on human preference data [21], a natural hypothesis regarding the defense vulnerabilities of MLLMs is that the model may be susceptible to out-of-distribution (OOD) inputs." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.539, + 0.483, + 0.902 + ], + "angle": 0, + "content": "Given that harmful textual content is already effectively detected in the field of LLMs, current jailbreak attack research on MLLMs primarily focuses on constructing OOD visual inputs, with two main approaches: image perturbation injection and prompt-to-image infection. The image perturbation injection involves subtly modifying images and combining them with query text to compose jailbreak prompts [35]. This attack leverages adversarial approaches [30] to introduce noisy OOD inputs that exploit the MLLM's visual-text decision boundary [40]. Unfortunately, these methods typically require access to the gradient of the victim or surrogate model [32], which limits their applicability in closed-source models [29]. In contrast, prompt-to-image infection generates images with similar semantics by using malicious text. These generated images are then combined with query text to craft jailbreak prompts that compel the MLLM to produce harmful responses [6, 12]. Since these generated images differ in distribution from those used for MLLM safety alignment training, they may bypass security measures and achieve jailbreak [2, 16, 26]. Despite this, the extensive training of large MLLMs on vast datasets makes it challenging for image generation models to produce truly OOD images. A failure case is depicted in the upper portion of Figure 1." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.092, + 0.905, + 0.318 + ], + "angle": 0, + "content": "Findings. Despite progress, it remains unclear how to fully leverage the visual space to construct OOD inputs for effective black-box MLLM jailbreaks. Building on the observation that most harmless alignment processes primarily involve simple images [6, 12, 22], we propose that increasing image complexity, potentially through the use of multiple subimages, could generate more effective OOD inputs for jailbreak attack. To verify this, we conduct several ablation analyses (§4.4), including varying the number of subimages and examining the relationship between image content and task. Our findings reveal that it is the complexity of the subimages, rather than their conceptual content, that drives the jailbreak success. Based on this insight, we propose the Distraction Hypothesis to explain the visual effects on jailbreak attacks against MLLMs." + }, + { + "type": "text", + "bbox": [ + 0.522, + 0.339, + 0.899, + 0.431 + ], + "angle": 0, + "content": "Distraction Hypothesis. Encoding complex images in the input prompt increases token complexity/diversity, which raises the processing burden on MLLMs. This overload can weaken the model's defenses, making it more prone to induce unintended outputs and improving jailbreak attack effectiveness." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.448, + 0.905, + 0.901 + ], + "angle": 0, + "content": "Structure and contributions. Building on the Distraction Hypothesis, we further introduce a novel framework, namely Contrasting Subimage Distraction Jailbreaking (CS-DJ) \\((\\S 3)\\), which leverages this insight to strategically design OOD inputs and enhance the effectiveness of jailbreak attacks. Specifically, CS-DJ explores both textual and visual components of the input space, introducing two key methods: structured distraction via query decomposition \\((\\S 3.1)\\) and visual-enhanced distraction through contrasting subimages \\((\\S 3.2)\\). (i) The structured distraction component decomposes the original harmful query into multiple sub-queries, each representing different aspects or intermediate steps of the original query. This decomposition induces a distributional shift that disperses the model's focus and weakens its ability to detect harmful content. (ii) The visual-enhanced distraction component manipulates the visual input by constructing multiple contrasting subimages, effectively disrupting interactions among visual elements during processing. Then, to guide the model's response and enhance its susceptibility to the multi-level distractions, CS-DJ incorporates a carefully designed harmless prompt alongside the composite visual input, as illustrated in Figure 1. This integrated approach disrupts the model's internal coherence and safety mechanisms, effectively bypassing its defenses. To evaluate the effectiveness of CS-DJ, we conduct extensive experiments across five representative scenarios and four widely used closed-source MLLMs \\((\\S 4)\\). Our results demonstrate that CS-DJ outperforms state-of-the-art jailbreak attacks, achieving average success rates of \\(52.40\\%\\) for attack success rate and \\(74.10\\%\\) for ensemble" + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.516, + 0.956 + ], + "angle": 0, + "content": "9468" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.092, + 0.093, + 0.221, + 0.106 + ], + "angle": 0, + "content": "attack success rate." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.128, + 0.233, + 0.143 + ], + "angle": 0, + "content": "2. Related Work" + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.155, + 0.41, + 0.172 + ], + "angle": 0, + "content": "2.1. Multimodal Large Language Models" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.18, + 0.485, + 0.529 + ], + "angle": 0, + "content": "Recently, Multimodal Large Language Models (MLLMs) [31] have become a prominent approach for processing and understanding heterogeneous multimodal data, including visual[39] and auditory [5] modalities. Building on instruction-tuning techniques developed for LLMs, most MLLMs are adapted from pre-trained LLMs through fine-tuning. Notable examples include LLaVA [14] and MiniGPT-4 [41], which share a similar architectural design by integrating pre-trained vision backbones with large language models via a projection layer, enabling effective vision-language alignment. Building upon this foundation, LLaVA-1.5 [13] further enhances its architecture by adopting a two-layer MLP for better vision-language mapping and increasing the vision encoder's input resolution for improved visual understanding. However, the incorporation of emerging modalities further intensifies the challenges of ensuring robust safety alignment in MLLMs. To address this challenge, advanced MLLMs, such as GPT-4o [7] and Gemini 1.5-Flash [27], employ a combination of alignment strategies, including post-training methods and integrated classifiers, to enhance both helpfulness and safety. Despite these efforts, MLLMs remains vulnerable to attacks using OOD inputs, which will be explored in this work." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.544, + 0.357, + 0.56 + ], + "angle": 0, + "content": "2.2. Jailbreak Attacks on MLLMs" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.569, + 0.483, + 0.902 + ], + "angle": 0, + "content": "Jailbreak attacks against MLLM can be performed through visual and textual input, leading to complex and diverse attack patterns [9, 20]. As a result, numerous jailbreak methods [6, 12, 17, 22, 24] have been continuously proposed. Qi et al. [22] used visual adversarial examples to bypass the security mechanisms of visual-textual MLLMs. However, this type of attack [20, 22] is typically only effective in white-box environments and is often difficult to apply to state-of-the-art black-box models. FigStep [6] introduced a simple yet effective black-box jailbreak approach that circumvents safety alignment bytrasfering prohibited content as images using typographic transformations. MM-SafetyBench [17] utilized a multi-step approach to construct jailbreak input, incorporating question generation, rephrasing, extraction of unsafe key phrases, and query-to-image conversion. Furthermore, Hades [12] investigated the influence of image harmfulness on model jailbreaks and proposed a prompt-embedded image infection jailbreak attack method. It generated harmful images using diffusion models and amplified the jailbreak effect through optimization. Unlike these works, CS-DJ explores the safety of MLLMs from a new perspective—visual distraction." + }, + { + "type": "title", + "bbox": [ + 0.514, + 0.09, + 0.65, + 0.108 + ], + "angle": 0, + "content": "3. Methodology" + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.115, + 0.908, + 0.435 + ], + "angle": 0, + "content": "Beyond the standard LLMs, MLLMs also encode visual content into token sequences, saying that the intrinsic interactions among visual elements can affect the model's ability to defend against harmful prompts. However, most of the existing approaches [12, 16, 17] only take a single, complete image as input, resulting in notable similarities both (1) within image patches, and (2) between the image and the targeted jailbreak content. Such similarities may make the model likely to detect harmful inputs, potentially leading to jailbreak failure. To address these issues, we propose a novel jailbreak framework, called CS-DJ, as illustrated in Figure 2. CS-DJ leverages a multi-subimage structure to bypass the models' intrinsic safety detection mechanism, with introducing multi-level distraction. Specifically, CS-DJ consists of two main components: structured distraction, achieved through query decomposition, and visual-enhanced distraction, realized through the construction of contrasting subimages. The resulting composite image input, combined with a carefully designed instruction, is fed into MLLMs to achieve jailbreak. Further details are described in the following sections." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.443, + 0.907, + 0.473 + ], + "angle": 0, + "content": "3.1. Structured Distraction via Query Decomposition" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.48, + 0.909, + 0.649 + ], + "angle": 0, + "content": "MLLMs have strong defense ability against harmful textual inputs, reinforced by specific optimizations aligned with human preference data [10]. To circumvent these defenses, we introduce a structured distraction strategy that disrupts and fragments information alignment through query decomposition, thereby reducing the model's overall sensitivity to harmful content. As illustrated on the left of Figure 2, we begin by decomposing a original harmful query \\( Q \\) into multiple sub-queries, denoted as \\( \\{Q_{s}^{(i)}\\}_{i = 1}^{m} \\). Each sub-query addresses part of the original query from different aspects or intermediate steps." + }, + { + "type": "equation", + "bbox": [ + 0.648, + 0.661, + 0.907, + 0.679 + ], + "angle": 0, + "content": "\\[\n\\left\\{Q _ {s} ^ {i} \\right\\} _ {i = 1} ^ {m} = \\mathcal {G} (Q), \\tag {1}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.689, + 0.909, + 0.903 + ], + "angle": 0, + "content": "where \\(\\mathcal{G}\\) is the auxiliary decomposition model and \\(m\\) denotes the number of sub-queries. On one hand, this fragmented structure distracts the model's focus, creating a distributional shift that obscures the original intent and disrupts its ability to detect harmful content, thereby bypassing its safety mechanisms. On the other hand, during the jailbreak execution phase, we carefully design prompts that guide the MLLM to address all sub-queries simultaneously. This strategy causes the model to divide its attention among multiple tasks rather than focusing on a single query. Further details are discussed in Section 3.3. In practice, we utilized Qwen2.5-3B-Instruct [28] for query decomposition. The details of the query decomposition prompt are provided in the Appendix." + }, + { + "type": "page_number", + "bbox": [ + 0.482, + 0.945, + 0.516, + 0.957 + ], + "angle": 0, + "content": "9469" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.139, + 0.094, + 0.379, + 0.395 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.38, + 0.094, + 0.619, + 0.395 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.62, + 0.094, + 0.86, + 0.395 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.09, + 0.408, + 0.907, + 0.465 + ], + "angle": 0, + "content": "Figure 2. The framework of the proposed CS-DJ. Given a harmful query, CS-DJ employs a three-step process to execute a jailbreaking attack: (1) decompose the raw query into multiple sub-queries and transform them into images to introduce structural distraction, (2) retrieve contrasting images from a dataset as subimages to inject visual-enhanced distraction. (3) combine these images into a composite image as the visual input, aligned with a harmless instruction, to execute the jailbreaking attack." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.484, + 0.483, + 0.514 + ], + "angle": 0, + "content": "The next step involves converting the sub-queries into images:" + }, + { + "type": "equation", + "bbox": [ + 0.204, + 0.512, + 0.483, + 0.532 + ], + "angle": 0, + "content": "\\[\n\\left\\{I _ {s} ^ {i} \\right\\} _ {i = 1} ^ {m} = \\mathcal {T} \\left(\\left\\{Q _ {s} ^ {i} \\right\\} _ {i = 1} ^ {m}\\right), \\tag {2}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.538, + 0.484, + 0.643 + ], + "angle": 0, + "content": "where \\(\\mathcal{T}\\) denotes the transformation function that converts each sub-query \\(Q_{s}^{i}\\) into a corresponding image \\(I_{s}^{i}\\). This transformation introduces a new form of distraction by altering the input modality, complicating the model's ability to link and identify harmful content patterns. Presenting textual prompts as visual elements obscures their original intent, further disrupting the model's detection mechanisms." + }, + { + "type": "title", + "bbox": [ + 0.09, + 0.652, + 0.484, + 0.684 + ], + "angle": 0, + "content": "3.2. Visual-Enhanced Distraction via Multi-subimage Construction" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.69, + 0.484, + 0.901 + ], + "angle": 0, + "content": "In contrast to prior work that emphasizes the harmfulness of a single image for MLLM jailbreaks, we focus on the impact of distraction. To enhance visual input distraction, we construct multiple contrasting visual elements, each treated as a subimage that collectively forms the input image. Since MLLMs represent visual content as token sequences similar to text, the distraction must include two aspects: (1) the distraction between the text query and the visual subimage, and (2) the distraction among the visual subimages themselves. Directly constructing an image with sufficient distraction relative to a text query is challenging. Therefore, we simplify this process by transforming it into an image retrieval problem, aimed at retrieving \\( k \\) images from a dataset. These \\( k \\) images are selected to have the lowest similarity to" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.484, + 0.907, + 0.56 + ], + "angle": 0, + "content": "the query and to each other. Given the complexity of this optimization problem, we further approximate it for practical implementation. The procedure is illustrated in the middle of Figure 2. Specifically, we first encode the query \\( Q \\) as a dense vector using the CLIP model:" + }, + { + "type": "equation", + "bbox": [ + 0.646, + 0.573, + 0.907, + 0.591 + ], + "angle": 0, + "content": "\\[\nv (Q) = \\operatorname {C L I P} (Q). \\tag {3}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.604, + 0.906, + 0.661 + ], + "angle": 0, + "content": "In practice, CLIP-ViT-L/14 [23] is utilized as the extractor. We then retrieve an image from a dataset that most contrasts with \\( v(Q) \\), effectively by minimizing the cosine similarity as:" + }, + { + "type": "equation", + "bbox": [ + 0.606, + 0.666, + 0.907, + 0.689 + ], + "angle": 0, + "content": "\\[\nI _ {c} ^ {1} = \\arg \\min _ {I \\in \\mathcal {D}} \\mathbf {c o s} (v (Q), v (I)), \\tag {4}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.698, + 0.906, + 0.743 + ], + "angle": 0, + "content": "where \\(\\mathcal{D}\\) denotes the image dataset, and \\(\\cos (\\cdot ,\\cdot)\\) represents the cosine similarity. Next, we proceed to retrieve the subsequent subimages:" + }, + { + "type": "equation", + "bbox": [ + 0.513, + 0.756, + 0.905, + 0.81 + ], + "angle": 0, + "content": "\\[\nI _ {c} ^ {j} = \\arg \\min _ {I \\in \\mathcal {D}} \\left(\\mathbf {c o s} (v (Q), v (I)) + \\sum_ {i = 1} ^ {j - 1} \\mathbf {c o s} \\left(v \\left(I _ {c} ^ {i}\\right), v (I)\\right)\\right), \\tag {5}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.811, + 0.907, + 0.902 + ], + "angle": 0, + "content": "where \\( j \\) denotes the index of the current subimage being selected. This approach systematically ensures that each subimage has minimal similarity to both the query and the previously selected subimages, thereby maximizing contrast and enhancing the overall distraction effect for the MLLM jailbreak process." + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.516, + 0.957 + ], + "angle": 0, + "content": "9470" + } + ], + [ + { + "type": "title", + "bbox": [ + 0.091, + 0.091, + 0.306, + 0.108 + ], + "angle": 0, + "content": "3.3. Jailbreaking Execution" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.113, + 0.485, + 0.192 + ], + "angle": 0, + "content": "With the constructed set of subimages, we form a composite input that maximizes distraction for MLLMs. Specifically, the final composite image \\( I_{\\mathrm{comp}} \\) is created by combining the subimages \\( \\{I_c^j\\}_{j=1}^k \\) along with the transformed sub-queries \\( \\{I_s^i\\}_{i=1}^m \\):" + }, + { + "type": "equation", + "bbox": [ + 0.153, + 0.201, + 0.484, + 0.226 + ], + "angle": 0, + "content": "\\[\nI _ {\\text {c o m p}} = \\text {C o m b i n e} \\left(\\left\\{I _ {c} ^ {j} \\right\\} _ {j = 1} ^ {k}, \\left\\{I _ {s} ^ {i} \\right\\} _ {i = 1} ^ {m}\\right), \\tag {6}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.231, + 0.484, + 0.274 + ], + "angle": 0, + "content": "where the function \\(\\mathbf{Combine}(\\cdot, \\cdot)\\) represents the structured arrangement of the subimages and sub-query transformations." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.276, + 0.484, + 0.307 + ], + "angle": 0, + "content": "This composite input \\( I_{\\mathrm{comp}} \\), along with a harmless instruction \\( P \\), is fed into a MLLM for processing:" + }, + { + "type": "equation", + "bbox": [ + 0.207, + 0.321, + 0.483, + 0.339 + ], + "angle": 0, + "content": "\\[\nY = \\operatorname {M L L M} \\left(I _ {\\text {c o m p}}, P\\right), \\tag {7}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.342, + 0.424, + 0.357 + ], + "angle": 0, + "content": "where \\(Y\\) denotes the textual output of the MLLM." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.358, + 0.484, + 0.63 + ], + "angle": 0, + "content": "The use of \\( I_{\\mathrm{comp}} \\) with contrasting visual elements maximizes distraction, effectively disrupting the model's internal alignment and coherence mechanisms. To complement \\( I_{\\mathrm{comp}} \\), we carefully design \\( P \\) to further enhance the distraction effect. As illustrated in Figure 1, \\( P \\) comprises three main components: (1) a role-guiding section, (2) a task-guiding section, and (3) a visual-guiding section. The role-guiding strategy is commonly employed across various LLM tasks. The second component of \\( P \\) directs the model to process multiple sub-queries simultaneously, increasing task complexity and dispersing attention across multiple tasks instead of focusing on a single one. Finally, \\( P \\) includes a misleading instruction, informing the model that the visual subimages may contain useful information, further diverting its attention. By presenting these inputs together, the approach reduces the model's ability to detect and mitigate harmful content, effectively bypassing its intrinsic safety mechanisms through multi-level distraction." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.638, + 0.395, + 0.653 + ], + "angle": 0, + "content": "3.4. Distraction Distance Measurement" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.66, + 0.483, + 0.737 + ], + "angle": 0, + "content": "To quantify the effectiveness of our multi-subimage input for distraction, we introduce a metric called Distraction Distance. This metric evaluates interactions among the query \\( Q \\) and contrasting visual images \\( \\{I_c^j\\}_{j=1}^k \\), treating all as equally important nodes within a unified structure." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.738, + 0.484, + 0.813 + ], + "angle": 0, + "content": "Each node is represented by its CLIP-encoded vector. For a given node, we calculate the \\( \\mathrm{L}_2 \\) distance between its vector and the vectors of all other nodes. The sum of these \\( \\mathrm{L}_2 \\) distances for each node is calculated across all other nodes to obtain the Distraction Distance:" + }, + { + "type": "equation", + "bbox": [ + 0.196, + 0.82, + 0.483, + 0.864 + ], + "angle": 0, + "content": "\\[\nD L = \\sum_ {i = 1} ^ {N} \\sum_ {j \\neq i} | | v _ {i} - v _ {j} | | _ {2}, \\tag {8}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.871, + 0.484, + 0.903 + ], + "angle": 0, + "content": "where \\(N\\) denotes the total number of nodes, and \\(v_{i}\\) represents the CLIP-encoded vector of node \\(i\\). This approach ac" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.092, + 0.906, + 0.183 + ], + "angle": 0, + "content": "counts for both the number of nodes and their pairwise distances, providing a comprehensive measure of the dispersion and contrast within the multi-subimage structure. By maximizing this Distraction Distance, we amplify the overall distracting effect on MLLMs, thereby disrupting their ability to accurately detect harmful content." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.2, + 0.64, + 0.218 + ], + "angle": 0, + "content": "4. Experiment" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.226, + 0.906, + 0.332 + ], + "angle": 0, + "content": "In this section, we first introduced our experimental settings which include datasets and evaluation metrics. Then, we carried out quantitative and qualitative comparisons between our method and state-of-the-art MLLM jailbreaking approaches. Additionally, ablation studies are performed to validate the distraction effect on MLLM jailbreaking from various perspectives." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.345, + 0.692, + 0.361 + ], + "angle": 0, + "content": "4.1. Experiment Setup" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.368, + 0.907, + 0.609 + ], + "angle": 0, + "content": "Dataset. The datasets are primarily utilized to provide or construct visual inputs tailored to the requirements of different methods. For evaluating Hades, we followed the experimental setting outlined in its original paper, utilizing the Hades dataset [12] to supply the visual inputs. This dataset comprises five representative categories related to real-world visual information: (1) Violence, (2) Financial, (3) Privacy, (4) Self-Harm, and (5) Animal. Each category contains 150 queries, resulting in a total of 750 harmful queries. For evaluating CS-DJ, we employed LLaVACC3M-Pretrain595K to retrieve the contrasting images. To improve retrieval efficiency, we randomly selected a subset of 10,000 images for the retrieval dataset, using a different random seed for each experiment to ensure diversity. All harmful queries used throughout the experiments were derived from the Hades dataset." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.61, + 0.906, + 0.716 + ], + "angle": 0, + "content": "Victim Models. In our main experiments, We evaluated the performance of the four most popular closed-source MLLMs: GPT-4o-Mini, GPT-4o, GPT-4V [1], and Gemini-1.5-Flash [27]. The specific versions are as follows: GPT-4o: GPT-4o-2024-08-06, GPT-4o-Mini: GPT-4o-Mini-2024-07-18, GPT-4V: GPT-4-1106- vision-preview, and Gemini-1.5-flash: Gemini-1.5-flash-001." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.717, + 0.906, + 0.839 + ], + "angle": 0, + "content": "Evaluation Metrics. To assess CS-DJ, we employ two primary evaluation metrics: Attack Success Rate (ASR) and Ensemble Attack Success Rate (EASR). ASR [6, 12] quantifies the proportion of successful jailbreak attempts by assessing whether a model's response meets predefined harmfulness criteria. The harmfulness criteria come from the security discriminator Beaver-Dam-7b [8], used to detect both the harmfulness and helpfulness of responses to" + }, + { + "type": "page_footnote", + "bbox": [ + 0.513, + 0.852, + 0.906, + 0.875 + ], + "angle": 0, + "content": "https://huggingface.co/datasets/liuhaotian/LLaVA-CC3M-Pretrain-595K" + }, + { + "type": "page_footnote", + "bbox": [ + 0.516, + 0.876, + 0.906, + 0.901 + ], + "angle": 0, + "content": "Since Gemini-1.0-Pro-Vision has been deprecated on July 12, 2024, we used its upgraded version, Gemini-1.5-Flash." + }, + { + "type": "list", + "bbox": [ + 0.513, + 0.852, + 0.906, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.482, + 0.945, + 0.514, + 0.957 + ], + "angle": 0, + "content": "9471" + } + ], + [ + { + "type": "table", + "bbox": [ + 0.171, + 0.089, + 0.827, + 0.426 + ], + "angle": 0, + "content": "
Victim ModelMethodAnimalFinancialPrivacySelf-HarmViolenceAverage (%)
Attack Success Rate (ASR) ↑
GPT-4o-miniHades4.7711.666.881.555.556.08
CS-DJ26.5580.0077.1134.7770.5557.80 (+51.72)
GPT-4oHades2.3310.117.551.885.665.51
CS-DJ15.3370.7758.5515.7750.7742.24 (+36.73)
GPT-4VHades8.1132.5523.888.0029.1120.33
CS-DJ11.6676.2259.7722.2257.3345.44 (+25.11)
Gemini-1.5-FlashHades2.6623.4413.111.775.009.20
CS-DJ22.5588.8879.1147.0083.0064.11 (+54.91)
Ensemble Attack Success Rate (EASR) ↑
GPT-4o-miniHades10.6621.3313.334.6610.0012.00
CS-DJ54.0092.6694.6652.0092.6677.20 (+65.20)
GPT-4oHades5.3316.0012.006.0011.3310.13
CS-DJ35.3392.6683.3337.3380.6665.86 (+55.73)
GPT-4VHades18.6663.3354.0024.6664.6645.06
CS-DJ34.6694.0089.3352.089.3371.86 (+26.80)
Gemini-1.5-FlashHades4.6634.0020.004.6610.6614.79
CS-DJ51.3397.3395.3366.0097.3381.46 (+66.67)
" + }, + { + "type": "table_caption", + "bbox": [ + 0.168, + 0.436, + 0.827, + 0.449 + ], + "angle": 0, + "content": "Table 1. ASR and EASR results of CS-DJ and Hades on four closed-source MLLMs across different categories." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.47, + 0.483, + 0.56 + ], + "angle": 0, + "content": "their corresponding queries. EASR [36] measures the effectiveness of jailbreak templates by calculating the proportion of queries where at least one template successfully bypasses the target MLLM's defenses. While ASR focuses on the success rate of individual templates, EASR measures the combined success rate of a group of templates." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.575, + 0.265, + 0.59 + ], + "angle": 0, + "content": "4.2. Main Experiment" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.598, + 0.483, + 0.673 + ], + "angle": 0, + "content": "To assess the effectiveness of CS-DJ, we conducted a comparison between CS-DJ and Hades [12], the state-of-the-art closed-source MLLM jailbreak attack. For a fair comparison, we adopted the official implementation of Hades and performed 6 groups of experiments to collect data." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.675, + 0.483, + 0.901 + ], + "angle": 0, + "content": "Table 1 presents the jailbreak ASR and EASR results of CS-DJ and Hades across the four widely used closed-source MLLMs. The results across the five categories indicate that the four closed-source MLLMs exhibit stronger defenses against harmful queries in the Animal and Self-Harm categories while showing greater vulnerability in the Financial, Privacy, and Violence categories. Compared to their ASR results, Hades and CS-DJ achieve average improvements of \\(10.22\\%\\) and \\(21.70\\%\\) in EASR, respectively, indicating that increasing the number of trials can enhance the jailbreak success rate. Additionally, for both ASR and EASR, CS-DJ consistently outperforms Hades across all five categories for each evaluated MLLM. Overall, compared to Hades, CS-DJ achieves maximum improvements of \\(54.91\\%\\) in ASR and \\(66.67\\%\\) in EASR. Furthermore, CS-DJ attains" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.47, + 0.905, + 0.546 + ], + "angle": 0, + "content": "average success rates of \\(52.40\\%\\) for ASR and \\(74.10\\%\\) for EASR. These observations highlight CS-DJ's superior effectiveness as a jailbreak method for bypassing the defenses of closed-source MLLMs. Jailbreaking examples can be found in the appendix." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.547, + 0.906, + 0.713 + ], + "angle": 0, + "content": "A detailed analysis of Hades' results across the four models reveals that GPT-4V outperforms the other three models in both ASR and EASR. In contrast, the results of CS-DJ indicate that Gemini-1.5-Flash achieved the highest performance. These findings suggest that the impact of visual input harmfulness and distraction varies across different models. In a word, the results demonstrate that distracting the model's attention is a more effective strategy for enhancing jailbreak success rates. More quantitative and qualitative experimental results, including those tested on open-source models, are provided in the appendix." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.714, + 0.905, + 0.774 + ], + "angle": 0, + "content": "For the subsequent experiments, we selected the most widely used model, GPT-40, for testing. All experiments were conducted as a single trial. Therefore, only ASR results are presented in the following sections." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.788, + 0.796, + 0.803 + ], + "angle": 0, + "content": "4.3. Impact of Query Decomposition" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.811, + 0.905, + 0.901 + ], + "angle": 0, + "content": "Here, we aim to explore the impact of query decomposition on MLLM jailbreak performance. To validate its effectiveness, we first use images derived from raw queries (RQ) as visual inputs for MLLMs, serving as a baseline for comparison with the results obtained using query decomposition. We further examine the effect of the number" + }, + { + "type": "page_number", + "bbox": [ + 0.482, + 0.945, + 0.516, + 0.956 + ], + "angle": 0, + "content": "9472" + } + ], + [ + { + "type": "table", + "bbox": [ + 0.095, + 0.089, + 0.483, + 0.171 + ], + "angle": 0, + "content": "
SettingAni.Fin.Priv.Self-H.Viol.ASR (%)
RQ0.676.005.3304.003.20
3SQ4.0028.6634.664.6622.0018.80
6SQ8.0045.3344.0010.6641.3329.86
9SQ4.6644.6644.667.3336.0027.46
" + }, + { + "type": "table_caption", + "bbox": [ + 0.09, + 0.181, + 0.483, + 0.266 + ], + "angle": 0, + "content": "Table 2. GPT-4o's ASR results of CS-DJ under different query decomposition settings. RQ denotes using the image transformation of raw queries as the visual input for MLLMs. 3SQ, 6SQ, and 9SQ indicate that the raw query is decomposed into 3, 6, and 9 sub-queries, respectively, which are then transformed into sub-images to compose the final visual input." + }, + { + "type": "text", + "bbox": [ + 0.089, + 0.291, + 0.483, + 0.442 + ], + "angle": 0, + "content": "of sub-queries on the jailbreak success rate by decomposing the raw query into 3, 6, and 9 sub-queries, denoted as 3SQ, 6SQ, and 9SQ, respectively. Given that MLLMs typically resize image inputs to fixed dimensions, we aim to minimize distortion from aspect ratio changes during resizing and avoid potential limitations of the MLLM's image encoder. Therefore, we maintain a fixed column count of 3 when composing the final input image. As a result, the 3SQ, 6SQ, and 9SQ settings increase the number of rows while keeping the number of columns constant." + }, + { + "type": "text", + "bbox": [ + 0.093, + 0.446, + 0.483, + 0.852 + ], + "angle": 0, + "content": "The quantitative results are presented in Table 2. The results from RQ demonstrate that the MLLM exhibits strong defenses against original jailbreak queries. Even when the raw query is transformed into an image input, it remains highly challenging to bypass the MLLM's security mechanisms. In contrast, 3SQ outperforms RQ across all five categories, yielding an overall improvement of \\(15.60\\%\\). This indicates that query decomposition effectively enhances the jailbreak success rate for MLLMs. By introducing a distributional shift in the queries, query decomposition creates structural distraction, thereby enabling more effective bypassing of the model's defenses. Furthermore, increasing the number of decomposed sub-queries to 6 results in an additional \\(11.06\\%\\) improvement in the jailbreak success rate. This suggests that a higher number of sub-images transformed from sub-queries can further enhance the distraction effect on the MLLM. However, when the number of sub-queries is increased to 9, a slight decline in the success rate is observed. This decrease may be attributed to two factors: (1) the increased number of sub-queries raises task complexity, resulting in greater demands on the model's comprehension and its ability to handle complex tasks; (2) limitations in the model's image encoding capabilities may prevent accurate interpretation of the sub-queries; and (3) over-decomposition may cause the model to recognize the original intent of the query, thereby reducing the effectiveness of the jailbreak attempt." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.856, + 0.484, + 0.901 + ], + "angle": 0, + "content": "Overall, the structural distraction introduced by query decomposition can enhance the jailbreak success rate of MLLMs. Furthermore, the number of sub-queries also" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.092, + 0.905, + 0.168 + ], + "angle": 0, + "content": "plays a significant role in influencing the results. Given the limited capabilities of smaller MLLMs and the need to incorporate additional visual subimages, this work focuses on evaluating the performance of the CS-DJ with three subqueries." + }, + { + "type": "image_caption", + "bbox": [ + 0.513, + 0.176, + 0.783, + 0.193 + ], + "angle": 0, + "content": "4.4. Impact of Contrasting Visuals" + }, + { + "type": "image", + "bbox": [ + 0.527, + 0.212, + 0.892, + 0.363 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.513, + 0.374, + 0.908, + 0.403 + ], + "angle": 0, + "content": "Figure 3. GPT-4o's evaluation results for varying quantities of visual subimages." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.418, + 0.907, + 0.72 + ], + "angle": 0, + "content": "To validate the effectiveness of contrasting visuals, we conduct experiments from two perspectives: (1) the impact of distraction induced by varying the number of subimages, and (2) the effect of inter-subimage distraction with a fixed number of subimages. Since altering the number of subimages directly affects the layout of the final composite image, we maintain a fixed column count of 3 and adjust the number of rows to control the number of subimages. This strategy reduces distortion due to excessive changes in aspect ratio and mitigates potential limitations in the MLLM's image encoding capabilities. We conduct different experiments with various configurations, involving 0, 3, 6, 9, and 12 Contrasting visual SubImages, corresponding to 3SQ, 3SQ+3CSI, 3SQ+6CSI, 3SQ+9CSI, and 3SQ+12CSI, respectively. To explore the influence of inter-subimage distraction on jailbreak success rates, we conduct experiments using three settings: retrieving 9 most contrasting images, retrieving 9 most similar images, and using a single, most similar image from the dataset for all 9 subimages, denoting 3SQ+9CSI, 3SQ+9SSI, and 3SQ+9SinSI, respectively." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.72, + 0.909, + 0.901 + ], + "angle": 0, + "content": "The experimental results on the impact of visual subimage quantity are illustrated in Figure 3. As 3SQ does not contain any visual subimages, its Distraction Distance is 0, resulting in the lowest ASR. Increasing the number of visual subimages leads to a rise in ASR. Specifically, compared to 3SQ, \\(3\\mathrm{SQ} + 3\\mathrm{CSI}\\) obtains a \\(12.8\\%\\) increase in ASR, demonstrating that contrasting visual subimages can effectively distract the MLLM and bypass its internal defense mechanisms. As the number of visual subimages increases, both the Distraction Distance and the ASR grow correspondingly. However, the rate of ASR improvement diminishes as the number of subimages continues to rise. When the" + }, + { + "type": "page_number", + "bbox": [ + 0.482, + 0.945, + 0.516, + 0.956 + ], + "angle": 0, + "content": "9473" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.09, + 0.092, + 0.482, + 0.152 + ], + "angle": 0, + "content": "number of subimages reaches 9, further additions yield limited gains in ASR. Given the limited encoding capabilities of the visual encoders in smaller MLLMs, we use 9 visual subimages as the default setting for our experiments." + }, + { + "type": "text", + "bbox": [ + 0.093, + 0.153, + 0.483, + 0.468 + ], + "angle": 0, + "content": "Table 3 presents the experimental results on the impact of inter-subimage distraction. We evaluate three different subimage selection strategies. The results show that \\(3\\mathrm{SQ} + 9\\mathrm{SinSI}\\) has the lowest distraction distance, corresponding to the lowest ASR. In contrast, \\(3\\mathrm{SQ} + 9\\mathrm{CSI}\\) achieves an \\(11.20\\%\\) improvement in ASR compared to \\(3\\mathrm{SQ} + 9\\mathrm{SSI}\\), along with a further increase in distraction distance. These findings indicate that, for a fixed number of subimages, greater distraction among subimages leads to a higher jailbreak success rate for MLLMs. Additionally, an analysis of the results from Figure 3 and Table 3 reveals that distraction distance accurately reflects changes in ASR trends only when varying the subimage construction strategy along a single dimension, such as the number of subimages or subimage selection strategies. This limitation may stem from differences between the semantic space of our current feature extractor, CLIP, and that of the MLLM. As a result, inter-subimage distraction distance cannot always be accurately measured using the CLIP model. Future work will focus on developing a more comprehensive metric for measuring distraction distance." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.469, + 0.482, + 0.573 + ], + "angle": 0, + "content": "Table 4 shows the impact of information complexity of visual subimages. The ASR of \\(3\\mathrm{SQ} + 9\\mathrm{RNI}\\) is close to that of 3SQ, but significantly lower than \\(3\\mathrm{SQ} + 9\\mathrm{CSI}\\), indicating that MLLMs exhibit strong recognition capabilities for noise images with minimal informational content. Therefore, only subimages with higher information complexity can effectively distract the model." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.575, + 0.482, + 0.62 + ], + "angle": 0, + "content": "In summary, constructing multiple contrasting visual subimages effectively enhances the distraction of MLLMs, leading to improved jailbreak success rates." + }, + { + "type": "table", + "bbox": [ + 0.126, + 0.632, + 0.446, + 0.702 + ], + "angle": 0, + "content": "
SettingASR (%)Distraction Distance
3SQ+9SinSI24.66399.08
3SQ+9SSI32.531378.77
3SQ+9CSI43.731878.83
" + }, + { + "type": "table_caption", + "bbox": [ + 0.09, + 0.709, + 0.482, + 0.78 + ], + "angle": 0, + "content": "Table 3. ASR of GPT-4o and Distraction Distance for CS-DJ under different visual subimage selection strategies. 9SinSI indicates using a single, most similar image from the dataset for all 9 subimages. 9CSI and 9SSI denote retrieving 9 most contrasting and similar images to compose the final visual input, respectively." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.804, + 0.292, + 0.82 + ], + "angle": 0, + "content": "4.5. Impact of Instruction" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.826, + 0.482, + 0.901 + ], + "angle": 0, + "content": "To complement the constructed multi-subimage input, we carefully designed the instruction \\(P\\). To evaluate the impact of different components of \\(P\\), we conducted three experiments: (1) using only the task-guiding component, (2) adding the role-guiding component, and (3) further incorpo" + }, + { + "type": "table", + "bbox": [ + 0.518, + 0.09, + 0.905, + 0.153 + ], + "angle": 0, + "content": "
SettingAni.Fin.Priv.Self-H.Viol.ASR (%)
3SQ4.0028.6634.664.6622.0018.80
3SQ+9RNI5.3332.6633.334.6624.0020.00
3SQ+9CSI19.3373.3356.6618.0051.3343.73
" + }, + { + "type": "table_caption", + "bbox": [ + 0.513, + 0.161, + 0.905, + 0.203 + ], + "angle": 0, + "content": "Table 4. GPT-4o's ASR results of CS-DJ under different settings. 9RNI denotes using 9 random noise images to compose the final visual input." + }, + { + "type": "table", + "bbox": [ + 0.518, + 0.219, + 0.905, + 0.279 + ], + "angle": 0, + "content": "
SettingAni.Fin.Priv.Self-H.Viol.ASR (%)
task-guiding10.0055.3339.3313.3342.6632.00
+ role-guiding15.3367.3350.0013.3352.0039.73
+ visual-guiding19.3373.3356.6618.0051.3343.73
" + }, + { + "type": "table_caption", + "bbox": [ + 0.513, + 0.286, + 0.905, + 0.315 + ], + "angle": 0, + "content": "Table 5. GPT-4o's ASR results of CS-DJ under the 3SQ+9CSI configuration with different instruction \\(P\\) settings." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.329, + 0.907, + 0.526 + ], + "angle": 0, + "content": "rating the visual-guiding component. The results, presented in Table 5, show that using only the task-guiding component achieved the lowest ASR but still outperformed Hades by \\(26.49\\%\\). This indicates that the visual distraction from the multi-subimage input, combined with the task distraction of handling multiple tasks simultaneously, significantly contributes to improving jailbreak success rates. Adding the role-guiding component further increased the ASR by \\(7.73\\%\\), while incorporating the visual-guiding component led to an additional \\(4.00\\%\\) improvement. These findings demonstrate that targeted optimization of \\(P\\) for multisubimage visual inputs can further disperse the model's attention, enhancing the success rate of the jailbreak." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.542, + 0.634, + 0.558 + ], + "angle": 0, + "content": "5. Conclusion" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.568, + 0.907, + 0.779 + ], + "angle": 0, + "content": "In this paper, we investigate the visual vulnerabilities in jailbreak attacks on MLLMs. Our findings show that the complexity of visual inputs significantly influences attack success. We introduce the Distraction Hypothesis and propose the CS-DJ framework, which leverages distraction-based strategies to exploit these vulnerabilities in MLLMs. Our approach integrates this insight by two core components: structured distraction through query decomposition and visual-enhanced distraction using contrasting subimages. Extensive experiments conducted on five representative scenarios and four popular closed-source MLLMs demonstrated that CS-DJ outperforms state-of-the-art jailbreak methods, validating the efficacy of distraction-based approaches in bypassing MLLM defenses." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.78, + 0.909, + 0.901 + ], + "angle": 0, + "content": "Acknowledgment. This research is supported by Guangzhou Basic and Applied Basic Research Project (No: 2025A04J3536, 2025A03J3128, 2024A03J0324), Guangzhou University Research Project (No: RQ2021013), Guangdong Basic and Applied Basic Research Foundation (No: 2023A1515110077), NSFC (No: 62402131), and the Postdoctoral Fellowship Program of CPSF (No: GZC20230595)." + }, + { + "type": "page_number", + "bbox": [ + 0.482, + 0.945, + 0.516, + 0.956 + ], + "angle": 0, + "content": "9474" + } + ], + [ + { + "type": "title", + "bbox": [ + 0.093, + 0.09, + 0.188, + 0.106 + ], + "angle": 0, + "content": "References" + }, + { + "type": "ref_text", + "bbox": [ + 0.101, + 0.115, + 0.484, + 0.184 + ], + "angle": 0, + "content": "[1] Josh Achiam, Steven Adler, Sandhini Agarwal, Lama Ahmad, Ilge Akkaya, Florencia Leoni Aleman, Diogo Almeida, Janko Altenschmidt, Sam Altman, Shyamal Anadkat, et al. Gpt-4 technical report. arXiv preprint arXiv:2303.08774, 2023.5" + }, + { + "type": "ref_text", + "bbox": [ + 0.1, + 0.187, + 0.484, + 0.255 + ], + "angle": 0, + "content": "[2] Yulin Chen, Haoran Li, Zihao Zheng, and Yangqiu Song. Bathe: Defense against the jailbreak attack in multimodal large language models by treating harmful instruction as backdoor trigger. arXiv preprint arXiv:2408.09093, 2024. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.101, + 0.259, + 0.484, + 0.342 + ], + "angle": 0, + "content": "[3] Can Cui, Yunsheng Ma, Xu Cao, Wenqian Ye, Yang Zhou, Kaizhao Liang, Jintai Chen, Juanwu Lu, Zichong Yang, Kuei-Da Liao, et al. A survey on multimodal large language models for autonomous driving. In Proceedings of the IEEE/CVF Winter Conference on Applications of Computer Vision, pages 958-979, 2024. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.101, + 0.344, + 0.483, + 0.413 + ], + "angle": 0, + "content": "[4] Ameet Deshpande, Vishvak Murahari, Tanmay Rajpurohit, Ashwin Kalyan, and Karthik Narasimhan. Toxicity in chatgpt: Analyzing persona-assigned language models. In Findings of the Association for Computational Linguistics: EMNLP 2023, pages 1236-1270, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.101, + 0.416, + 0.484, + 0.512 + ], + "angle": 0, + "content": "[5] Yassir Fathullah, Chunyang Wu, Egor Lakomkin, Junteng Jia, Yuan Shangguan, Ke Li, Jinxi Guo, Wenhan Xiong, Jay Mahadeokar, Ozlem Kalinli, et al. Prompting large language models with speech recognition abilities. In ICASSP 2024-2024 IEEE International Conference on Acoustics, Speech and Signal Processing (ICASSP), pages 13351-13355. IEEE, 2024. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.101, + 0.515, + 0.484, + 0.583 + ], + "angle": 0, + "content": "[6] Yichen Gong, Delong Ran, Jinyuan Liu, Conglei Wang, Tianshuo Cong, Anyu Wang, Sisi Duan, and Xiaoyun Wang. Figstep: Jailbreaking large vision-language models via typographic visual prompts. arXiv preprint arXiv:2311.05608, 2023. 2, 3, 5, 12" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.586, + 0.484, + 0.642 + ], + "angle": 0, + "content": "[7] Aaron Hurst, Adam Lerer, Adam P Goucher, Adam Perelman, Aditya Ramesh, Aidan Clark, AJ Ostrow, Akila Weli-hinda, Alan Hayes, Alec Radford, et al. Gpt-4o system card. arXiv preprint arXiv:2410.21276, 2024. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.645, + 0.484, + 0.713 + ], + "angle": 0, + "content": "[8] Jiaming Ji, Mickel Liu, Josef Dai, Xuehai Pan, Chi Zhang, Ce Bian, Boyuan Chen, Ruiyang Sun, Yizhou Wang, and Yaodong Yang. Beavertails: Towards improved safety alignment of llm via a human-preference dataset. Advances in Neural Information Processing Systems, 36, 2024. 5, 12" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.716, + 0.484, + 0.784 + ], + "angle": 0, + "content": "[9] Haibo Jin, Leyang Hu, Xinuo Li, Peiyan Zhang, Chonghan Chen, Jun Zhuang, and Haohan Wang. Jailbreakzoo: Survey, landscapes, and horizons in jailbreaking large language and vision-language models. arXiv preprint arXiv:2407.01599, 2024.3" + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.788, + 0.483, + 0.843 + ], + "angle": 0, + "content": "[10] Daniel Martin Katz, Michael James Bommarito, Shang Gao, and Pablo Arredondo. Gpt-4 passes the bar exam. Philosophical Transactions of the Royal Society A, 382(2270): 20230254, 2024. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.846, + 0.483, + 0.901 + ], + "angle": 0, + "content": "[11] Bo Li, Yuanhan Zhang, Dong Guo, Renrui Zhang, Feng Li, Hao Zhang, Kaichen Zhang, Peiyuan Zhang, Yanwei Li, Ziwei Liu, et al. Llava-onevision: Easy visual task transfer. arXiv preprint arXiv:2408.03326, 2024. 12" + }, + { + "type": "list", + "bbox": [ + 0.094, + 0.115, + 0.484, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.093, + 0.905, + 0.161 + ], + "angle": 0, + "content": "[12] Yifan Li, Hangyu Guo, Kun Zhou, Wayne Xin Zhao, and JiRong Wen. Images are achilles' heel of alignment: Exploiting visual vulnerabilities for jailbreaking multimodal large language models. In European Conference on Computer Vision. Springer, 2024. 1, 2, 3, 5, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.163, + 0.905, + 0.218 + ], + "angle": 0, + "content": "[13] Haotian Liu, Chunyuan Li, Yuheng Li, and Yong Jae Lee. Improved baselines with visual instruction tuning. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 26296-26306, 2024. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.219, + 0.905, + 0.26 + ], + "angle": 0, + "content": "[14] Haotian Liu, Chunyuan Li, Qingyang Wu, and Yong Jae Lee. Visual instruction tuning. Advances in neural information processing systems, 36, 2024. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.261, + 0.905, + 0.329 + ], + "angle": 0, + "content": "[15] Tong Liu, Yingjie Zhang, Zhe Zhao, Yinpeng Dong, Guozhu Meng, and Kai Chen. Making them ask and answer: Jailbreaking large language models in few queries via disguise and reconstruction. In 33rd USENIX Security Symposium (USENIX Security 24), pages 4711-4728, 2024. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.33, + 0.904, + 0.371 + ], + "angle": 0, + "content": "[16] Xin Liu, Yichen Zhu, Yunshi Lan, Chao Yang, and Yu Qiao. Query-relevant images jailbreak large multi-modal models. arXiv preprint arXiv:2311.17600, 2023. 2, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.372, + 0.905, + 0.44 + ], + "angle": 0, + "content": "[17] Xin Liu, Yichen Zhu, Jindong Gu, Yunshi Lan, Chao Yang, and Yu Qiao. Mm-safetybench: A benchmark for safety evaluation of multimodal large language models. In European Conference on Computer Vision, pages 386-403. Springer, 2024. 2, 3, 12" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.442, + 0.905, + 0.523 + ], + "angle": 0, + "content": "[18] Dong Lu, Zhiqiang Wang, Teng Wang, Weili Guan, Hongchang Gao, and Feng Zheng. Set-level guidance attack: Boosting adversarial transferability of vision-language pre-training models. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 102-111, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.525, + 0.905, + 0.593 + ], + "angle": 0, + "content": "[19] Siyuan Ma, Weidi Luo, Yu Wang, Xiaogeng Liu, Muhao Chen, Bo Li, and Chaowei Xiao. Visual-roleplay: Universal jailbreak attack on multimodal large language models via role-playing image character. arXiv preprint arXiv:2405.20773, 2024. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.595, + 0.905, + 0.637 + ], + "angle": 0, + "content": "[20] Zhenxing Niu, Haodong Ren, Xinbo Gao, Gang Hua, and Rong Jin. Jailbreaking attack against multimodal large language model. arXiv preprint arXiv:2402.02309, 2024. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.638, + 0.905, + 0.718 + ], + "angle": 0, + "content": "[21] Long Ouyang, Jeffrey Wu, Xu Jiang, Diogo Almeida, Carroll Wainwright, Pamela Mishkin, Chong Zhang, Sandhini Agarwal, Katarina Slama, Alex Ray, et al. Training language models to follow instructions with human feedback. Advances in neural information processing systems, 35:27730-27744, 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.72, + 0.905, + 0.789 + ], + "angle": 0, + "content": "[22] Xiangyu Qi, Kaixuan Huang, Ashwinee Panda, Peter Henderson, Mengdi Wang, and Prateek Mittal. Visual adversarial examples jailbreak aligned large language models. In Proceedings of the AAAI Conference on Artificial Intelligence, pages 21527-21536, 2024. 2, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.79, + 0.905, + 0.871 + ], + "angle": 0, + "content": "[23] Alec Radford, Jong Wook Kim, Chris Hallacy, Aditya Ramesh, Gabriel Goh, Sandhini Agarwal, Girish Sastry, Amanda Askell, Pamela Mishkin, Jack Clark, et al. Learning transferable visual models from natural language supervision. In International conference on machine learning, pages 8748-8763. PMLR, 2021. 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.873, + 0.905, + 0.901 + ], + "angle": 0, + "content": "[24] Erfan Shayegani, Yue Dong, and Nael Abu-Ghazaleh. Jailbreak in pieces: Compositional adversarial attacks on multi-" + }, + { + "type": "list", + "bbox": [ + 0.094, + 0.093, + 0.905, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.516, + 0.956 + ], + "angle": 0, + "content": "9475" + } + ], + [ + { + "type": "ref_text", + "bbox": [ + 0.123, + 0.093, + 0.482, + 0.12 + ], + "angle": 0, + "content": "modal language models. In The Twelfth International Conference on Learning Representations, 2023. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.121, + 0.484, + 0.189 + ], + "angle": 0, + "content": "[25] Wentan Tan, Changxing Ding, Jiayu Jiang, Fei Wang, Yibing Zhan, and Dapeng Tao. Harnessing the power of mllms for transferable text-to-image person reid. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 17127-17137, 2024. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.19, + 0.482, + 0.232 + ], + "angle": 0, + "content": "[26] Xijia Tao, Shuai Zhong, Lei Li, Qi Liu, and Lingpeng Kong. Imgtrojan: Jailbreaking vision-language models with one image. arXiv preprint arXiv:2403.02910, 2024. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.233, + 0.482, + 0.301 + ], + "angle": 0, + "content": "[27] Gemini Team, Petko Georgiev, Ving Ian Lei, Ryan Burnell, Libin Bai, Anmol Gulati, Garrett Tanzer, Damien Vincent, Zhufeng Pan, Shibo Wang, et al. Gemini 1.5: Unlocking multimodal understanding across millions of tokens of context. arXiv preprint arXiv:2403.05530, 2024. 3, 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.302, + 0.482, + 0.328 + ], + "angle": 0, + "content": "[28] Qwen Team. Qwen2.5: A party of foundation models, 2024. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.33, + 0.482, + 0.399 + ], + "angle": 0, + "content": "[29] Haoqin Tu, Chenhang Cui, Zijun Wang, Yiyang Zhou, Bingchen Zhao, Junlin Han, Wangchunshu Zhou, Huaxiu Yao, and Cihang Xie. How many unicorns are in this image? a safety evaluation benchmark for vision llms. arXiv preprint arXiv:2311.16101, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.401, + 0.482, + 0.469 + ], + "angle": 0, + "content": "[30] Haodi Wang, Kai Dong, Zhilei Zhu, Haotong Qin, Aishan Liu, Xiaolin Fang, Jiakai Wang, and Xianglong Liu. Transferable multimodal attack on vision-language pre-training models. In 2024 IEEE Symposium on Security and Privacy (SP), pages 102-102. IEEE Computer Society, 2024. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.47, + 0.482, + 0.538 + ], + "angle": 0, + "content": "[31] Jiaqi Wang, Hanqi Jiang, Yiheng Liu, Chong Ma, Xu Zhang, Yi Pan, Mengyuan Liu, Peiran Gu, Sichen Xia, Wenjun Li, et al. A comprehensive review of multimodal large language models: Performance and challenges across different tasks. arXiv preprint arXiv:2408.01319, 2024. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.539, + 0.482, + 0.608 + ], + "angle": 0, + "content": "[32] Ruofan Wang, Xingjun Ma, Hanxu Zhou, Chuanjun Ji, Guangnan Ye, and Yu-Gang Jiang. White-box multimodal jailbreaks against large vision-language models. In Proceedings of the 32nd ACM International Conference on Multimedia, pages 6920-6928, 2024. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.609, + 0.482, + 0.678 + ], + "angle": 0, + "content": "[33] Yizhong Wang, Yeganeh Kordi, Swaroop Mishra, Alisa Liu, Noah A Smith, Daniel Khashabi, and Hannaneh Hajishirzi. Self-instruct: Aligning language models with self-generated instructions. In The 61st Annual Meeting Of The Association For Computational Linguistics, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.679, + 0.482, + 0.719 + ], + "angle": 0, + "content": "[34] Alexander Wei, Nika Haghtalab, and Jacob Steinhardt. Jailbroken: How does llm safety training fail? Advances in Neural Information Processing Systems, 36, 2024. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.721, + 0.482, + 0.762 + ], + "angle": 0, + "content": "[35] Yuanwei Wu, Xiang Li, Yixin Liu, Pan Zhou, and Lichao Sun. Jailbreaking gpt-4v via self-adversarial attacks with system prompts. arXiv preprint arXiv:2311.09127, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.763, + 0.482, + 0.817 + ], + "angle": 0, + "content": "[36] Jiahao Yu, Xingwei Lin, Zheng Yu, and Xinyu Xing. {LLM-Fuzzer}: Scaling assessment of large language model jailbreaks. In 33rd USENIX Security Symposium (USENIX Security 24), pages 4657-4674, 2024. 2, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.818, + 0.482, + 0.901 + ], + "angle": 0, + "content": "[37] Qifan Yu, Juncheng Li, Longhui Wei, Liang Pang, Wentao Ye, Bosheng Qin, Siliang Tang, Qi Tian, and Yueting Zhuang. Hallucidobot: Mitigating hallucinatory toxicity in visual instruction data. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 12944-12953, 2024. 2" + }, + { + "type": "list", + "bbox": [ + 0.093, + 0.093, + 0.484, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.093, + 0.905, + 0.176 + ], + "angle": 0, + "content": "[38] Hangfan Zhang, Zhimeng Guo, Huaisheng Zhu, Bochuan Cao, Lu Lin, Jinyuan Jia, Jinghui Chen, and Dinghao Wu. Jailbreak open-sourced large language models via enforced decoding. In Proceedings of the 62nd Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers), pages 5475-5493, 2024. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.178, + 0.905, + 0.231 + ], + "angle": 0, + "content": "[39] Jingyi Zhang, Jiaxing Huang, Sheng Jin, and Shijian Lu. Vision-language models for vision tasks: A survey. IEEE Transactions on Pattern Analysis and Machine Intelligence, 2024. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.234, + 0.905, + 0.301 + ], + "angle": 0, + "content": "[40] Yunqing Zhao, Tianyu Pang, Chao Du, Xiao Yang, Chongxuan Li, Ngai-Man Man Cheung, and Min Lin. On evaluating adversarial robustness of large vision-language models. Advances in Neural Information Processing Systems, 36, 2024. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.304, + 0.905, + 0.359 + ], + "angle": 0, + "content": "[41] Deyao Zhu, Jun Chen, Xiaoqian Shen, Xiang Li, and Mohamed Elhoseiny. Minigpt-4: Enhancing vision-language understanding with advanced large language models. arXiv preprint arXiv:2304.10592, 2023. 3" + }, + { + "type": "list", + "bbox": [ + 0.517, + 0.093, + 0.905, + 0.359 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.517, + 0.956 + ], + "angle": 0, + "content": "9476" + } + ] +] \ No newline at end of file diff --git a/2025/Distraction is All You Need for Multimodal Large Language Model Jailbreaking/1375bb02-e050-41f2-9b1d-ed16de66c742_origin.pdf b/2025/Distraction is All You Need for Multimodal Large Language Model Jailbreaking/1375bb02-e050-41f2-9b1d-ed16de66c742_origin.pdf new file mode 100644 index 0000000000000000000000000000000000000000..3957108dfe61bb960618aa646c35f8682bb0ab89 --- /dev/null +++ b/2025/Distraction is All You Need for Multimodal Large Language Model Jailbreaking/1375bb02-e050-41f2-9b1d-ed16de66c742_origin.pdf @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:c15908f8167c21deb49fcfa1f15a50e1551ec75ed580f8756f0dc1690f7f1398 +size 2217023 diff --git a/2025/Distraction is All You Need for Multimodal Large Language Model Jailbreaking/full.md b/2025/Distraction is All You Need for Multimodal Large Language Model Jailbreaking/full.md new file mode 100644 index 0000000000000000000000000000000000000000..666f41bfc69c4fcb82f1d3f374eccabcd5c1c8b9 --- /dev/null +++ b/2025/Distraction is All You Need for Multimodal Large Language Model Jailbreaking/full.md @@ -0,0 +1,265 @@ +# Distraction is All You Need for Multimodal Large Language Model Jailbreaking + +Zuopeng Yang $^{1*}$ , Jiluan Fan $^{1*}$ , Anli Yan $^{1\dagger}$ , Erdun Gao $^{3}$ , Xin Lin $^{1\dagger}$ , Tao Li $^{2}$ , Kanghua Mo $^{1}$ , Changyu Dong $^{1}$ + +$^{1}$ Guangzhou University $^{2}$ Shanghai Jiao Tong University $^{3}$ Australian Institute for Machine Learning, The University of Adelaide + +![](images/7af0484c152bcef177f680586409cb7549584bbb15a8b19ce86a288bd2d1d783.jpg) +Figure 1. Jailbreaking examples of Hades [12] and our proposed CS-DJ against GPT-4. When using Hades, GPT-4 effectively rejects harmful questions. In contrast, CS-DJ successfully bypasses the model's defense mechanisms by constructing multi-subimage visual inputs to distract the model, resulting in the acquisition of harmful responses. The harmful information is highlighted in red, while key prompts are emphasized in bold. + +# Abstract + +Multimodal Large Language Models (MLLMs) bridge the gap between visual and textual data, enabling a range of advanced applications. However, complex internal interactions among visual elements and their alignment with text can introduce vulnerabilities, which may be exploited to bypass safety mechanisms. To address this, we analyze the relationship between image content and task and find that the complexity of subimages, rather than their content, is key. Building on this insight, we propose the Distraction Hypothesis, followed by a novel framework called Contrasting Subimage Distraction Jailbreaking (CS-DJ), to achieve jailbreaking by disrupting MLLMs alignment through multi-level distraction strategies. CS-DJ consists of two components: structured distraction, achieved through query decomposition that induces a distributional shift by fragmenting harmful prompts into subqueries, and visual-enhanced distraction, realized by constructing contrasting subimages to disrupt the interactions + +among visual elements within the model. This dual strategy disperses the model's attention, reducing its ability to detect and mitigate harmful content. Extensive experiments across five representative scenarios and four popular closed-source MLLMs, including GPT-4o-mini, GPT-4o, GPT-4V, and Gemini-1.5-Flash, demonstrate that CS-DJ achieves average success rates of $52.40\%$ for the attack success rate and $74.10\%$ for the ensemble attack success rate. These results reveal the potential of distraction-based approaches to exploit and bypass MLLMs' defenses, offering new insights for attack strategies. Our code is available at https://github.com/TeamPigeonLab/CS-DJ. + +Warning: This paper contains unfiltered content generated by MLLMs that may be offensive to readers. + +# 1. Introduction + +Since the early 2020s, Multimodal Large Language Models (MLLMs) have emerged as a powerful tool for bridging natural language processing and computer vision [3], owing to their ability to integrate and process multimodal informa + +tion efficiently [25]. Despite their impressive performance, one notable limitation of these methods is that the training materials may contain toxic content, potentially leading to inappropriate or harmful content [4, 38], the leakage of sensitive information [36], and even serious vulnerabilities in the model security and privacy protection. To this end, significant efforts have been devoted to using reinforcement learning from human feedback (RLHF) to align LLMs outputs, ensuring readiness for deployment in high-stakes domains [21, 33]. However, beyond Large Language Models (LLMs), the integration of visual inputs in MLLMs introduces a new challenge: securing models against vulnerabilities arising from newly integrated visual modalities, which remains a significant issue and lacks a comprehensive understanding [17, 18, 37] + +In this study, rather than following the human feedback alignment research at the defense side, we focus on exploring an alternative approach, namely jailbreak attacks [19, 34, 36]. This line of research also seeks to provide insights into how safety concerns emerge, while, considering the attacker's perspective. Specifically, it aims to actively uncover the defense vulnerabilities in current model by deliberately crafting inputs to manipulate MLLMs, leading to unintended and potentially harmful outputs [15]. Motivated by the fact that most current defense mechanisms rely on RLHF and are trained on human preference data [21], a natural hypothesis regarding the defense vulnerabilities of MLLMs is that the model may be susceptible to out-of-distribution (OOD) inputs. + +Given that harmful textual content is already effectively detected in the field of LLMs, current jailbreak attack research on MLLMs primarily focuses on constructing OOD visual inputs, with two main approaches: image perturbation injection and prompt-to-image infection. The image perturbation injection involves subtly modifying images and combining them with query text to compose jailbreak prompts [35]. This attack leverages adversarial approaches [30] to introduce noisy OOD inputs that exploit the MLLM's visual-text decision boundary [40]. Unfortunately, these methods typically require access to the gradient of the victim or surrogate model [32], which limits their applicability in closed-source models [29]. In contrast, prompt-to-image infection generates images with similar semantics by using malicious text. These generated images are then combined with query text to craft jailbreak prompts that compel the MLLM to produce harmful responses [6, 12]. Since these generated images differ in distribution from those used for MLLM safety alignment training, they may bypass security measures and achieve jailbreak [2, 16, 26]. Despite this, the extensive training of large MLLMs on vast datasets makes it challenging for image generation models to produce truly OOD images. A failure case is depicted in the upper portion of Figure 1. + +Findings. Despite progress, it remains unclear how to fully leverage the visual space to construct OOD inputs for effective black-box MLLM jailbreaks. Building on the observation that most harmless alignment processes primarily involve simple images [6, 12, 22], we propose that increasing image complexity, potentially through the use of multiple subimages, could generate more effective OOD inputs for jailbreak attack. To verify this, we conduct several ablation analyses (§4.4), including varying the number of subimages and examining the relationship between image content and task. Our findings reveal that it is the complexity of the subimages, rather than their conceptual content, that drives the jailbreak success. Based on this insight, we propose the Distraction Hypothesis to explain the visual effects on jailbreak attacks against MLLMs. + +Distraction Hypothesis. Encoding complex images in the input prompt increases token complexity/diversity, which raises the processing burden on MLLMs. This overload can weaken the model's defenses, making it more prone to induce unintended outputs and improving jailbreak attack effectiveness. + +Structure and contributions. Building on the Distraction Hypothesis, we further introduce a novel framework, namely Contrasting Subimage Distraction Jailbreaking (CS-DJ) $(\S 3)$ , which leverages this insight to strategically design OOD inputs and enhance the effectiveness of jailbreak attacks. Specifically, CS-DJ explores both textual and visual components of the input space, introducing two key methods: structured distraction via query decomposition $(\S 3.1)$ and visual-enhanced distraction through contrasting subimages $(\S 3.2)$ . (i) The structured distraction component decomposes the original harmful query into multiple sub-queries, each representing different aspects or intermediate steps of the original query. This decomposition induces a distributional shift that disperses the model's focus and weakens its ability to detect harmful content. (ii) The visual-enhanced distraction component manipulates the visual input by constructing multiple contrasting subimages, effectively disrupting interactions among visual elements during processing. Then, to guide the model's response and enhance its susceptibility to the multi-level distractions, CS-DJ incorporates a carefully designed harmless prompt alongside the composite visual input, as illustrated in Figure 1. This integrated approach disrupts the model's internal coherence and safety mechanisms, effectively bypassing its defenses. To evaluate the effectiveness of CS-DJ, we conduct extensive experiments across five representative scenarios and four widely used closed-source MLLMs $(\S 4)$ . Our results demonstrate that CS-DJ outperforms state-of-the-art jailbreak attacks, achieving average success rates of $52.40\%$ for attack success rate and $74.10\%$ for ensemble + +attack success rate. + +# 2. Related Work + +# 2.1. Multimodal Large Language Models + +Recently, Multimodal Large Language Models (MLLMs) [31] have become a prominent approach for processing and understanding heterogeneous multimodal data, including visual[39] and auditory [5] modalities. Building on instruction-tuning techniques developed for LLMs, most MLLMs are adapted from pre-trained LLMs through fine-tuning. Notable examples include LLaVA [14] and MiniGPT-4 [41], which share a similar architectural design by integrating pre-trained vision backbones with large language models via a projection layer, enabling effective vision-language alignment. Building upon this foundation, LLaVA-1.5 [13] further enhances its architecture by adopting a two-layer MLP for better vision-language mapping and increasing the vision encoder's input resolution for improved visual understanding. However, the incorporation of emerging modalities further intensifies the challenges of ensuring robust safety alignment in MLLMs. To address this challenge, advanced MLLMs, such as GPT-4o [7] and Gemini 1.5-Flash [27], employ a combination of alignment strategies, including post-training methods and integrated classifiers, to enhance both helpfulness and safety. Despite these efforts, MLLMs remains vulnerable to attacks using OOD inputs, which will be explored in this work. + +# 2.2. Jailbreak Attacks on MLLMs + +Jailbreak attacks against MLLM can be performed through visual and textual input, leading to complex and diverse attack patterns [9, 20]. As a result, numerous jailbreak methods [6, 12, 17, 22, 24] have been continuously proposed. Qi et al. [22] used visual adversarial examples to bypass the security mechanisms of visual-textual MLLMs. However, this type of attack [20, 22] is typically only effective in white-box environments and is often difficult to apply to state-of-the-art black-box models. FigStep [6] introduced a simple yet effective black-box jailbreak approach that circumvents safety alignment bytrasfering prohibited content as images using typographic transformations. MM-SafetyBench [17] utilized a multi-step approach to construct jailbreak input, incorporating question generation, rephrasing, extraction of unsafe key phrases, and query-to-image conversion. Furthermore, Hades [12] investigated the influence of image harmfulness on model jailbreaks and proposed a prompt-embedded image infection jailbreak attack method. It generated harmful images using diffusion models and amplified the jailbreak effect through optimization. Unlike these works, CS-DJ explores the safety of MLLMs from a new perspective—visual distraction. + +# 3. Methodology + +Beyond the standard LLMs, MLLMs also encode visual content into token sequences, saying that the intrinsic interactions among visual elements can affect the model's ability to defend against harmful prompts. However, most of the existing approaches [12, 16, 17] only take a single, complete image as input, resulting in notable similarities both (1) within image patches, and (2) between the image and the targeted jailbreak content. Such similarities may make the model likely to detect harmful inputs, potentially leading to jailbreak failure. To address these issues, we propose a novel jailbreak framework, called CS-DJ, as illustrated in Figure 2. CS-DJ leverages a multi-subimage structure to bypass the models' intrinsic safety detection mechanism, with introducing multi-level distraction. Specifically, CS-DJ consists of two main components: structured distraction, achieved through query decomposition, and visual-enhanced distraction, realized through the construction of contrasting subimages. The resulting composite image input, combined with a carefully designed instruction, is fed into MLLMs to achieve jailbreak. Further details are described in the following sections. + +# 3.1. Structured Distraction via Query Decomposition + +MLLMs have strong defense ability against harmful textual inputs, reinforced by specific optimizations aligned with human preference data [10]. To circumvent these defenses, we introduce a structured distraction strategy that disrupts and fragments information alignment through query decomposition, thereby reducing the model's overall sensitivity to harmful content. As illustrated on the left of Figure 2, we begin by decomposing a original harmful query $Q$ into multiple sub-queries, denoted as $\{Q_{s}^{(i)}\}_{i = 1}^{m}$ . Each sub-query addresses part of the original query from different aspects or intermediate steps. + +$$ +\left\{Q _ {s} ^ {i} \right\} _ {i = 1} ^ {m} = \mathcal {G} (Q), \tag {1} +$$ + +where $\mathcal{G}$ is the auxiliary decomposition model and $m$ denotes the number of sub-queries. On one hand, this fragmented structure distracts the model's focus, creating a distributional shift that obscures the original intent and disrupts its ability to detect harmful content, thereby bypassing its safety mechanisms. On the other hand, during the jailbreak execution phase, we carefully design prompts that guide the MLLM to address all sub-queries simultaneously. This strategy causes the model to divide its attention among multiple tasks rather than focusing on a single query. Further details are discussed in Section 3.3. In practice, we utilized Qwen2.5-3B-Instruct [28] for query decomposition. The details of the query decomposition prompt are provided in the Appendix. + +![](images/c2ee3ec1c2cea2ffde7b1f63dab9159d9590b2ef998d6ac52d06322facb64316.jpg) +Figure 2. The framework of the proposed CS-DJ. Given a harmful query, CS-DJ employs a three-step process to execute a jailbreaking attack: (1) decompose the raw query into multiple sub-queries and transform them into images to introduce structural distraction, (2) retrieve contrasting images from a dataset as subimages to inject visual-enhanced distraction. (3) combine these images into a composite image as the visual input, aligned with a harmless instruction, to execute the jailbreaking attack. + +![](images/27a8d9dac43b96769b81738d6f11734123a2e3f710058588f76ce8303133747d.jpg) + +![](images/6e1451c85d535c0e198c4edc972e0cc8e0bbb684b7cfdbfbbce6715411128a3b.jpg) + +The next step involves converting the sub-queries into images: + +$$ +\left\{I _ {s} ^ {i} \right\} _ {i = 1} ^ {m} = \mathcal {T} \left(\left\{Q _ {s} ^ {i} \right\} _ {i = 1} ^ {m}\right), \tag {2} +$$ + +where $\mathcal{T}$ denotes the transformation function that converts each sub-query $Q_{s}^{i}$ into a corresponding image $I_{s}^{i}$ . This transformation introduces a new form of distraction by altering the input modality, complicating the model's ability to link and identify harmful content patterns. Presenting textual prompts as visual elements obscures their original intent, further disrupting the model's detection mechanisms. + +# 3.2. Visual-Enhanced Distraction via Multi-subimage Construction + +In contrast to prior work that emphasizes the harmfulness of a single image for MLLM jailbreaks, we focus on the impact of distraction. To enhance visual input distraction, we construct multiple contrasting visual elements, each treated as a subimage that collectively forms the input image. Since MLLMs represent visual content as token sequences similar to text, the distraction must include two aspects: (1) the distraction between the text query and the visual subimage, and (2) the distraction among the visual subimages themselves. Directly constructing an image with sufficient distraction relative to a text query is challenging. Therefore, we simplify this process by transforming it into an image retrieval problem, aimed at retrieving $k$ images from a dataset. These $k$ images are selected to have the lowest similarity to + +the query and to each other. Given the complexity of this optimization problem, we further approximate it for practical implementation. The procedure is illustrated in the middle of Figure 2. Specifically, we first encode the query $Q$ as a dense vector using the CLIP model: + +$$ +v (Q) = \operatorname {C L I P} (Q). \tag {3} +$$ + +In practice, CLIP-ViT-L/14 [23] is utilized as the extractor. We then retrieve an image from a dataset that most contrasts with $v(Q)$ , effectively by minimizing the cosine similarity as: + +$$ +I _ {c} ^ {1} = \arg \min _ {I \in \mathcal {D}} \mathbf {c o s} (v (Q), v (I)), \tag {4} +$$ + +where $\mathcal{D}$ denotes the image dataset, and $\cos (\cdot ,\cdot)$ represents the cosine similarity. Next, we proceed to retrieve the subsequent subimages: + +$$ +I _ {c} ^ {j} = \arg \min _ {I \in \mathcal {D}} \left(\mathbf {c o s} (v (Q), v (I)) + \sum_ {i = 1} ^ {j - 1} \mathbf {c o s} \left(v \left(I _ {c} ^ {i}\right), v (I)\right)\right), \tag {5} +$$ + +where $j$ denotes the index of the current subimage being selected. This approach systematically ensures that each subimage has minimal similarity to both the query and the previously selected subimages, thereby maximizing contrast and enhancing the overall distraction effect for the MLLM jailbreak process. + +# 3.3. Jailbreaking Execution + +With the constructed set of subimages, we form a composite input that maximizes distraction for MLLMs. Specifically, the final composite image $I_{\mathrm{comp}}$ is created by combining the subimages $\{I_c^j\}_{j=1}^k$ along with the transformed sub-queries $\{I_s^i\}_{i=1}^m$ : + +$$ +I _ {\text {c o m p}} = \text {C o m b i n e} \left(\left\{I _ {c} ^ {j} \right\} _ {j = 1} ^ {k}, \left\{I _ {s} ^ {i} \right\} _ {i = 1} ^ {m}\right), \tag {6} +$$ + +where the function $\mathbf{Combine}(\cdot, \cdot)$ represents the structured arrangement of the subimages and sub-query transformations. + +This composite input $I_{\mathrm{comp}}$ , along with a harmless instruction $P$ , is fed into a MLLM for processing: + +$$ +Y = \operatorname {M L L M} \left(I _ {\text {c o m p}}, P\right), \tag {7} +$$ + +where $Y$ denotes the textual output of the MLLM. + +The use of $I_{\mathrm{comp}}$ with contrasting visual elements maximizes distraction, effectively disrupting the model's internal alignment and coherence mechanisms. To complement $I_{\mathrm{comp}}$ , we carefully design $P$ to further enhance the distraction effect. As illustrated in Figure 1, $P$ comprises three main components: (1) a role-guiding section, (2) a task-guiding section, and (3) a visual-guiding section. The role-guiding strategy is commonly employed across various LLM tasks. The second component of $P$ directs the model to process multiple sub-queries simultaneously, increasing task complexity and dispersing attention across multiple tasks instead of focusing on a single one. Finally, $P$ includes a misleading instruction, informing the model that the visual subimages may contain useful information, further diverting its attention. By presenting these inputs together, the approach reduces the model's ability to detect and mitigate harmful content, effectively bypassing its intrinsic safety mechanisms through multi-level distraction. + +# 3.4. Distraction Distance Measurement + +To quantify the effectiveness of our multi-subimage input for distraction, we introduce a metric called Distraction Distance. This metric evaluates interactions among the query $Q$ and contrasting visual images $\{I_c^j\}_{j=1}^k$ , treating all as equally important nodes within a unified structure. + +Each node is represented by its CLIP-encoded vector. For a given node, we calculate the $\mathrm{L}_2$ distance between its vector and the vectors of all other nodes. The sum of these $\mathrm{L}_2$ distances for each node is calculated across all other nodes to obtain the Distraction Distance: + +$$ +D L = \sum_ {i = 1} ^ {N} \sum_ {j \neq i} | | v _ {i} - v _ {j} | | _ {2}, \tag {8} +$$ + +where $N$ denotes the total number of nodes, and $v_{i}$ represents the CLIP-encoded vector of node $i$ . This approach ac + +counts for both the number of nodes and their pairwise distances, providing a comprehensive measure of the dispersion and contrast within the multi-subimage structure. By maximizing this Distraction Distance, we amplify the overall distracting effect on MLLMs, thereby disrupting their ability to accurately detect harmful content. + +# 4. Experiment + +In this section, we first introduced our experimental settings which include datasets and evaluation metrics. Then, we carried out quantitative and qualitative comparisons between our method and state-of-the-art MLLM jailbreaking approaches. Additionally, ablation studies are performed to validate the distraction effect on MLLM jailbreaking from various perspectives. + +# 4.1. Experiment Setup + +Dataset. The datasets are primarily utilized to provide or construct visual inputs tailored to the requirements of different methods. For evaluating Hades, we followed the experimental setting outlined in its original paper, utilizing the Hades dataset [12] to supply the visual inputs. This dataset comprises five representative categories related to real-world visual information: (1) Violence, (2) Financial, (3) Privacy, (4) Self-Harm, and (5) Animal. Each category contains 150 queries, resulting in a total of 750 harmful queries. For evaluating CS-DJ, we employed LLaVACC3M-Pretrain595K to retrieve the contrasting images. To improve retrieval efficiency, we randomly selected a subset of 10,000 images for the retrieval dataset, using a different random seed for each experiment to ensure diversity. All harmful queries used throughout the experiments were derived from the Hades dataset. + +Victim Models. In our main experiments, We evaluated the performance of the four most popular closed-source MLLMs: GPT-4o-Mini, GPT-4o, GPT-4V [1], and Gemini-1.5-Flash [27]. The specific versions are as follows: GPT-4o: GPT-4o-2024-08-06, GPT-4o-Mini: GPT-4o-Mini-2024-07-18, GPT-4V: GPT-4-1106- vision-preview, and Gemini-1.5-flash: Gemini-1.5-flash-001. + +Evaluation Metrics. To assess CS-DJ, we employ two primary evaluation metrics: Attack Success Rate (ASR) and Ensemble Attack Success Rate (EASR). ASR [6, 12] quantifies the proportion of successful jailbreak attempts by assessing whether a model's response meets predefined harmfulness criteria. The harmfulness criteria come from the security discriminator Beaver-Dam-7b [8], used to detect both the harmfulness and helpfulness of responses to + +
Victim ModelMethodAnimalFinancialPrivacySelf-HarmViolenceAverage (%)
Attack Success Rate (ASR) ↑
GPT-4o-miniHades4.7711.666.881.555.556.08
CS-DJ26.5580.0077.1134.7770.5557.80 (+51.72)
GPT-4oHades2.3310.117.551.885.665.51
CS-DJ15.3370.7758.5515.7750.7742.24 (+36.73)
GPT-4VHades8.1132.5523.888.0029.1120.33
CS-DJ11.6676.2259.7722.2257.3345.44 (+25.11)
Gemini-1.5-FlashHades2.6623.4413.111.775.009.20
CS-DJ22.5588.8879.1147.0083.0064.11 (+54.91)
Ensemble Attack Success Rate (EASR) ↑
GPT-4o-miniHades10.6621.3313.334.6610.0012.00
CS-DJ54.0092.6694.6652.0092.6677.20 (+65.20)
GPT-4oHades5.3316.0012.006.0011.3310.13
CS-DJ35.3392.6683.3337.3380.6665.86 (+55.73)
GPT-4VHades18.6663.3354.0024.6664.6645.06
CS-DJ34.6694.0089.3352.089.3371.86 (+26.80)
Gemini-1.5-FlashHades4.6634.0020.004.6610.6614.79
CS-DJ51.3397.3395.3366.0097.3381.46 (+66.67)
+ +Table 1. ASR and EASR results of CS-DJ and Hades on four closed-source MLLMs across different categories. + +their corresponding queries. EASR [36] measures the effectiveness of jailbreak templates by calculating the proportion of queries where at least one template successfully bypasses the target MLLM's defenses. While ASR focuses on the success rate of individual templates, EASR measures the combined success rate of a group of templates. + +# 4.2. Main Experiment + +To assess the effectiveness of CS-DJ, we conducted a comparison between CS-DJ and Hades [12], the state-of-the-art closed-source MLLM jailbreak attack. For a fair comparison, we adopted the official implementation of Hades and performed 6 groups of experiments to collect data. + +Table 1 presents the jailbreak ASR and EASR results of CS-DJ and Hades across the four widely used closed-source MLLMs. The results across the five categories indicate that the four closed-source MLLMs exhibit stronger defenses against harmful queries in the Animal and Self-Harm categories while showing greater vulnerability in the Financial, Privacy, and Violence categories. Compared to their ASR results, Hades and CS-DJ achieve average improvements of $10.22\%$ and $21.70\%$ in EASR, respectively, indicating that increasing the number of trials can enhance the jailbreak success rate. Additionally, for both ASR and EASR, CS-DJ consistently outperforms Hades across all five categories for each evaluated MLLM. Overall, compared to Hades, CS-DJ achieves maximum improvements of $54.91\%$ in ASR and $66.67\%$ in EASR. Furthermore, CS-DJ attains + +average success rates of $52.40\%$ for ASR and $74.10\%$ for EASR. These observations highlight CS-DJ's superior effectiveness as a jailbreak method for bypassing the defenses of closed-source MLLMs. Jailbreaking examples can be found in the appendix. + +A detailed analysis of Hades' results across the four models reveals that GPT-4V outperforms the other three models in both ASR and EASR. In contrast, the results of CS-DJ indicate that Gemini-1.5-Flash achieved the highest performance. These findings suggest that the impact of visual input harmfulness and distraction varies across different models. In a word, the results demonstrate that distracting the model's attention is a more effective strategy for enhancing jailbreak success rates. More quantitative and qualitative experimental results, including those tested on open-source models, are provided in the appendix. + +For the subsequent experiments, we selected the most widely used model, GPT-40, for testing. All experiments were conducted as a single trial. Therefore, only ASR results are presented in the following sections. + +# 4.3. Impact of Query Decomposition + +Here, we aim to explore the impact of query decomposition on MLLM jailbreak performance. To validate its effectiveness, we first use images derived from raw queries (RQ) as visual inputs for MLLMs, serving as a baseline for comparison with the results obtained using query decomposition. We further examine the effect of the number + +
SettingAni.Fin.Priv.Self-H.Viol.ASR (%)
RQ0.676.005.3304.003.20
3SQ4.0028.6634.664.6622.0018.80
6SQ8.0045.3344.0010.6641.3329.86
9SQ4.6644.6644.667.3336.0027.46
+ +Table 2. GPT-4o's ASR results of CS-DJ under different query decomposition settings. RQ denotes using the image transformation of raw queries as the visual input for MLLMs. 3SQ, 6SQ, and 9SQ indicate that the raw query is decomposed into 3, 6, and 9 sub-queries, respectively, which are then transformed into sub-images to compose the final visual input. + +of sub-queries on the jailbreak success rate by decomposing the raw query into 3, 6, and 9 sub-queries, denoted as 3SQ, 6SQ, and 9SQ, respectively. Given that MLLMs typically resize image inputs to fixed dimensions, we aim to minimize distortion from aspect ratio changes during resizing and avoid potential limitations of the MLLM's image encoder. Therefore, we maintain a fixed column count of 3 when composing the final input image. As a result, the 3SQ, 6SQ, and 9SQ settings increase the number of rows while keeping the number of columns constant. + +The quantitative results are presented in Table 2. The results from RQ demonstrate that the MLLM exhibits strong defenses against original jailbreak queries. Even when the raw query is transformed into an image input, it remains highly challenging to bypass the MLLM's security mechanisms. In contrast, 3SQ outperforms RQ across all five categories, yielding an overall improvement of $15.60\%$ . This indicates that query decomposition effectively enhances the jailbreak success rate for MLLMs. By introducing a distributional shift in the queries, query decomposition creates structural distraction, thereby enabling more effective bypassing of the model's defenses. Furthermore, increasing the number of decomposed sub-queries to 6 results in an additional $11.06\%$ improvement in the jailbreak success rate. This suggests that a higher number of sub-images transformed from sub-queries can further enhance the distraction effect on the MLLM. However, when the number of sub-queries is increased to 9, a slight decline in the success rate is observed. This decrease may be attributed to two factors: (1) the increased number of sub-queries raises task complexity, resulting in greater demands on the model's comprehension and its ability to handle complex tasks; (2) limitations in the model's image encoding capabilities may prevent accurate interpretation of the sub-queries; and (3) over-decomposition may cause the model to recognize the original intent of the query, thereby reducing the effectiveness of the jailbreak attempt. + +Overall, the structural distraction introduced by query decomposition can enhance the jailbreak success rate of MLLMs. Furthermore, the number of sub-queries also + +plays a significant role in influencing the results. Given the limited capabilities of smaller MLLMs and the need to incorporate additional visual subimages, this work focuses on evaluating the performance of the CS-DJ with three subqueries. + +![](images/09053fcdce5461237480ccb35f5b8989a65444f0a56c838898ee18ae3df8172d.jpg) +4.4. Impact of Contrasting Visuals +Figure 3. GPT-4o's evaluation results for varying quantities of visual subimages. + +To validate the effectiveness of contrasting visuals, we conduct experiments from two perspectives: (1) the impact of distraction induced by varying the number of subimages, and (2) the effect of inter-subimage distraction with a fixed number of subimages. Since altering the number of subimages directly affects the layout of the final composite image, we maintain a fixed column count of 3 and adjust the number of rows to control the number of subimages. This strategy reduces distortion due to excessive changes in aspect ratio and mitigates potential limitations in the MLLM's image encoding capabilities. We conduct different experiments with various configurations, involving 0, 3, 6, 9, and 12 Contrasting visual SubImages, corresponding to 3SQ, 3SQ+3CSI, 3SQ+6CSI, 3SQ+9CSI, and 3SQ+12CSI, respectively. To explore the influence of inter-subimage distraction on jailbreak success rates, we conduct experiments using three settings: retrieving 9 most contrasting images, retrieving 9 most similar images, and using a single, most similar image from the dataset for all 9 subimages, denoting 3SQ+9CSI, 3SQ+9SSI, and 3SQ+9SinSI, respectively. + +The experimental results on the impact of visual subimage quantity are illustrated in Figure 3. As 3SQ does not contain any visual subimages, its Distraction Distance is 0, resulting in the lowest ASR. Increasing the number of visual subimages leads to a rise in ASR. Specifically, compared to 3SQ, $3\mathrm{SQ} + 3\mathrm{CSI}$ obtains a $12.8\%$ increase in ASR, demonstrating that contrasting visual subimages can effectively distract the MLLM and bypass its internal defense mechanisms. As the number of visual subimages increases, both the Distraction Distance and the ASR grow correspondingly. However, the rate of ASR improvement diminishes as the number of subimages continues to rise. When the + +number of subimages reaches 9, further additions yield limited gains in ASR. Given the limited encoding capabilities of the visual encoders in smaller MLLMs, we use 9 visual subimages as the default setting for our experiments. + +Table 3 presents the experimental results on the impact of inter-subimage distraction. We evaluate three different subimage selection strategies. The results show that $3\mathrm{SQ} + 9\mathrm{SinSI}$ has the lowest distraction distance, corresponding to the lowest ASR. In contrast, $3\mathrm{SQ} + 9\mathrm{CSI}$ achieves an $11.20\%$ improvement in ASR compared to $3\mathrm{SQ} + 9\mathrm{SSI}$ , along with a further increase in distraction distance. These findings indicate that, for a fixed number of subimages, greater distraction among subimages leads to a higher jailbreak success rate for MLLMs. Additionally, an analysis of the results from Figure 3 and Table 3 reveals that distraction distance accurately reflects changes in ASR trends only when varying the subimage construction strategy along a single dimension, such as the number of subimages or subimage selection strategies. This limitation may stem from differences between the semantic space of our current feature extractor, CLIP, and that of the MLLM. As a result, inter-subimage distraction distance cannot always be accurately measured using the CLIP model. Future work will focus on developing a more comprehensive metric for measuring distraction distance. + +Table 4 shows the impact of information complexity of visual subimages. The ASR of $3\mathrm{SQ} + 9\mathrm{RNI}$ is close to that of 3SQ, but significantly lower than $3\mathrm{SQ} + 9\mathrm{CSI}$ , indicating that MLLMs exhibit strong recognition capabilities for noise images with minimal informational content. Therefore, only subimages with higher information complexity can effectively distract the model. + +In summary, constructing multiple contrasting visual subimages effectively enhances the distraction of MLLMs, leading to improved jailbreak success rates. + +
SettingASR (%)Distraction Distance
3SQ+9SinSI24.66399.08
3SQ+9SSI32.531378.77
3SQ+9CSI43.731878.83
+ +# 4.5. Impact of Instruction + +To complement the constructed multi-subimage input, we carefully designed the instruction $P$ . To evaluate the impact of different components of $P$ , we conducted three experiments: (1) using only the task-guiding component, (2) adding the role-guiding component, and (3) further incorpo + +Table 3. ASR of GPT-4o and Distraction Distance for CS-DJ under different visual subimage selection strategies. 9SinSI indicates using a single, most similar image from the dataset for all 9 subimages. 9CSI and 9SSI denote retrieving 9 most contrasting and similar images to compose the final visual input, respectively. + +
SettingAni.Fin.Priv.Self-H.Viol.ASR (%)
3SQ4.0028.6634.664.6622.0018.80
3SQ+9RNI5.3332.6633.334.6624.0020.00
3SQ+9CSI19.3373.3356.6618.0051.3343.73
+ +Table 4. GPT-4o's ASR results of CS-DJ under different settings. 9RNI denotes using 9 random noise images to compose the final visual input. + +
SettingAni.Fin.Priv.Self-H.Viol.ASR (%)
task-guiding10.0055.3339.3313.3342.6632.00
+ role-guiding15.3367.3350.0013.3352.0039.73
+ visual-guiding19.3373.3356.6618.0051.3343.73
+ +Table 5. GPT-4o's ASR results of CS-DJ under the 3SQ+9CSI configuration with different instruction $P$ settings. + +rating the visual-guiding component. The results, presented in Table 5, show that using only the task-guiding component achieved the lowest ASR but still outperformed Hades by $26.49\%$ . This indicates that the visual distraction from the multi-subimage input, combined with the task distraction of handling multiple tasks simultaneously, significantly contributes to improving jailbreak success rates. Adding the role-guiding component further increased the ASR by $7.73\%$ , while incorporating the visual-guiding component led to an additional $4.00\%$ improvement. These findings demonstrate that targeted optimization of $P$ for multisubimage visual inputs can further disperse the model's attention, enhancing the success rate of the jailbreak. + +# 5. Conclusion + +In this paper, we investigate the visual vulnerabilities in jailbreak attacks on MLLMs. Our findings show that the complexity of visual inputs significantly influences attack success. We introduce the Distraction Hypothesis and propose the CS-DJ framework, which leverages distraction-based strategies to exploit these vulnerabilities in MLLMs. Our approach integrates this insight by two core components: structured distraction through query decomposition and visual-enhanced distraction using contrasting subimages. Extensive experiments conducted on five representative scenarios and four popular closed-source MLLMs demonstrated that CS-DJ outperforms state-of-the-art jailbreak methods, validating the efficacy of distraction-based approaches in bypassing MLLM defenses. + +Acknowledgment. This research is supported by Guangzhou Basic and Applied Basic Research Project (No: 2025A04J3536, 2025A03J3128, 2024A03J0324), Guangzhou University Research Project (No: RQ2021013), Guangdong Basic and Applied Basic Research Foundation (No: 2023A1515110077), NSFC (No: 62402131), and the Postdoctoral Fellowship Program of CPSF (No: GZC20230595). + +# References + +[1] Josh Achiam, Steven Adler, Sandhini Agarwal, Lama Ahmad, Ilge Akkaya, Florencia Leoni Aleman, Diogo Almeida, Janko Altenschmidt, Sam Altman, Shyamal Anadkat, et al. Gpt-4 technical report. arXiv preprint arXiv:2303.08774, 2023.5 +[2] Yulin Chen, Haoran Li, Zihao Zheng, and Yangqiu Song. Bathe: Defense against the jailbreak attack in multimodal large language models by treating harmful instruction as backdoor trigger. arXiv preprint arXiv:2408.09093, 2024. 2 +[3] Can Cui, Yunsheng Ma, Xu Cao, Wenqian Ye, Yang Zhou, Kaizhao Liang, Jintai Chen, Juanwu Lu, Zichong Yang, Kuei-Da Liao, et al. A survey on multimodal large language models for autonomous driving. In Proceedings of the IEEE/CVF Winter Conference on Applications of Computer Vision, pages 958-979, 2024. 1 +[4] Ameet Deshpande, Vishvak Murahari, Tanmay Rajpurohit, Ashwin Kalyan, and Karthik Narasimhan. Toxicity in chatgpt: Analyzing persona-assigned language models. In Findings of the Association for Computational Linguistics: EMNLP 2023, pages 1236-1270, 2023. 2 +[5] Yassir Fathullah, Chunyang Wu, Egor Lakomkin, Junteng Jia, Yuan Shangguan, Ke Li, Jinxi Guo, Wenhan Xiong, Jay Mahadeokar, Ozlem Kalinli, et al. Prompting large language models with speech recognition abilities. In ICASSP 2024-2024 IEEE International Conference on Acoustics, Speech and Signal Processing (ICASSP), pages 13351-13355. IEEE, 2024. 3 +[6] Yichen Gong, Delong Ran, Jinyuan Liu, Conglei Wang, Tianshuo Cong, Anyu Wang, Sisi Duan, and Xiaoyun Wang. Figstep: Jailbreaking large vision-language models via typographic visual prompts. arXiv preprint arXiv:2311.05608, 2023. 2, 3, 5, 12 +[7] Aaron Hurst, Adam Lerer, Adam P Goucher, Adam Perelman, Aditya Ramesh, Aidan Clark, AJ Ostrow, Akila Weli-hinda, Alan Hayes, Alec Radford, et al. Gpt-4o system card. arXiv preprint arXiv:2410.21276, 2024. 3 +[8] Jiaming Ji, Mickel Liu, Josef Dai, Xuehai Pan, Chi Zhang, Ce Bian, Boyuan Chen, Ruiyang Sun, Yizhou Wang, and Yaodong Yang. Beavertails: Towards improved safety alignment of llm via a human-preference dataset. Advances in Neural Information Processing Systems, 36, 2024. 5, 12 +[9] Haibo Jin, Leyang Hu, Xinuo Li, Peiyan Zhang, Chonghan Chen, Jun Zhuang, and Haohan Wang. Jailbreakzoo: Survey, landscapes, and horizons in jailbreaking large language and vision-language models. arXiv preprint arXiv:2407.01599, 2024.3 +[10] Daniel Martin Katz, Michael James Bommarito, Shang Gao, and Pablo Arredondo. Gpt-4 passes the bar exam. Philosophical Transactions of the Royal Society A, 382(2270): 20230254, 2024. 3 +[11] Bo Li, Yuanhan Zhang, Dong Guo, Renrui Zhang, Feng Li, Hao Zhang, Kaichen Zhang, Peiyuan Zhang, Yanwei Li, Ziwei Liu, et al. Llava-onevision: Easy visual task transfer. arXiv preprint arXiv:2408.03326, 2024. 12 + +[12] Yifan Li, Hangyu Guo, Kun Zhou, Wayne Xin Zhao, and JiRong Wen. Images are achilles' heel of alignment: Exploiting visual vulnerabilities for jailbreaking multimodal large language models. In European Conference on Computer Vision. Springer, 2024. 1, 2, 3, 5, 6 +[13] Haotian Liu, Chunyuan Li, Yuheng Li, and Yong Jae Lee. Improved baselines with visual instruction tuning. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 26296-26306, 2024. 3 +[14] Haotian Liu, Chunyuan Li, Qingyang Wu, and Yong Jae Lee. Visual instruction tuning. Advances in neural information processing systems, 36, 2024. 3 +[15] Tong Liu, Yingjie Zhang, Zhe Zhao, Yinpeng Dong, Guozhu Meng, and Kai Chen. Making them ask and answer: Jailbreaking large language models in few queries via disguise and reconstruction. In 33rd USENIX Security Symposium (USENIX Security 24), pages 4711-4728, 2024. 2 +[16] Xin Liu, Yichen Zhu, Yunshi Lan, Chao Yang, and Yu Qiao. Query-relevant images jailbreak large multi-modal models. arXiv preprint arXiv:2311.17600, 2023. 2, 3 +[17] Xin Liu, Yichen Zhu, Jindong Gu, Yunshi Lan, Chao Yang, and Yu Qiao. Mm-safetybench: A benchmark for safety evaluation of multimodal large language models. In European Conference on Computer Vision, pages 386-403. Springer, 2024. 2, 3, 12 +[18] Dong Lu, Zhiqiang Wang, Teng Wang, Weili Guan, Hongchang Gao, and Feng Zheng. Set-level guidance attack: Boosting adversarial transferability of vision-language pre-training models. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 102-111, 2023. 2 +[19] Siyuan Ma, Weidi Luo, Yu Wang, Xiaogeng Liu, Muhao Chen, Bo Li, and Chaowei Xiao. Visual-roleplay: Universal jailbreak attack on multimodal large language models via role-playing image character. arXiv preprint arXiv:2405.20773, 2024. 2 +[20] Zhenxing Niu, Haodong Ren, Xinbo Gao, Gang Hua, and Rong Jin. Jailbreaking attack against multimodal large language model. arXiv preprint arXiv:2402.02309, 2024. 3 +[21] Long Ouyang, Jeffrey Wu, Xu Jiang, Diogo Almeida, Carroll Wainwright, Pamela Mishkin, Chong Zhang, Sandhini Agarwal, Katarina Slama, Alex Ray, et al. Training language models to follow instructions with human feedback. Advances in neural information processing systems, 35:27730-27744, 2022. 2 +[22] Xiangyu Qi, Kaixuan Huang, Ashwinee Panda, Peter Henderson, Mengdi Wang, and Prateek Mittal. Visual adversarial examples jailbreak aligned large language models. In Proceedings of the AAAI Conference on Artificial Intelligence, pages 21527-21536, 2024. 2, 3 +[23] Alec Radford, Jong Wook Kim, Chris Hallacy, Aditya Ramesh, Gabriel Goh, Sandhini Agarwal, Girish Sastry, Amanda Askell, Pamela Mishkin, Jack Clark, et al. Learning transferable visual models from natural language supervision. In International conference on machine learning, pages 8748-8763. PMLR, 2021. 4 +[24] Erfan Shayegani, Yue Dong, and Nael Abu-Ghazaleh. Jailbreak in pieces: Compositional adversarial attacks on multi- + +modal language models. In The Twelfth International Conference on Learning Representations, 2023. 3 +[25] Wentan Tan, Changxing Ding, Jiayu Jiang, Fei Wang, Yibing Zhan, and Dapeng Tao. Harnessing the power of mllms for transferable text-to-image person reid. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 17127-17137, 2024. 2 +[26] Xijia Tao, Shuai Zhong, Lei Li, Qi Liu, and Lingpeng Kong. Imgtrojan: Jailbreaking vision-language models with one image. arXiv preprint arXiv:2403.02910, 2024. 2 +[27] Gemini Team, Petko Georgiev, Ving Ian Lei, Ryan Burnell, Libin Bai, Anmol Gulati, Garrett Tanzer, Damien Vincent, Zhufeng Pan, Shibo Wang, et al. Gemini 1.5: Unlocking multimodal understanding across millions of tokens of context. arXiv preprint arXiv:2403.05530, 2024. 3, 5 +[28] Qwen Team. Qwen2.5: A party of foundation models, 2024. 3 +[29] Haoqin Tu, Chenhang Cui, Zijun Wang, Yiyang Zhou, Bingchen Zhao, Junlin Han, Wangchunshu Zhou, Huaxiu Yao, and Cihang Xie. How many unicorns are in this image? a safety evaluation benchmark for vision llms. arXiv preprint arXiv:2311.16101, 2023. 2 +[30] Haodi Wang, Kai Dong, Zhilei Zhu, Haotong Qin, Aishan Liu, Xiaolin Fang, Jiakai Wang, and Xianglong Liu. Transferable multimodal attack on vision-language pre-training models. In 2024 IEEE Symposium on Security and Privacy (SP), pages 102-102. IEEE Computer Society, 2024. 2 +[31] Jiaqi Wang, Hanqi Jiang, Yiheng Liu, Chong Ma, Xu Zhang, Yi Pan, Mengyuan Liu, Peiran Gu, Sichen Xia, Wenjun Li, et al. A comprehensive review of multimodal large language models: Performance and challenges across different tasks. arXiv preprint arXiv:2408.01319, 2024. 3 +[32] Ruofan Wang, Xingjun Ma, Hanxu Zhou, Chuanjun Ji, Guangnan Ye, and Yu-Gang Jiang. White-box multimodal jailbreaks against large vision-language models. In Proceedings of the 32nd ACM International Conference on Multimedia, pages 6920-6928, 2024. 2 +[33] Yizhong Wang, Yeganeh Kordi, Swaroop Mishra, Alisa Liu, Noah A Smith, Daniel Khashabi, and Hannaneh Hajishirzi. Self-instruct: Aligning language models with self-generated instructions. In The 61st Annual Meeting Of The Association For Computational Linguistics, 2023. 2 +[34] Alexander Wei, Nika Haghtalab, and Jacob Steinhardt. Jailbroken: How does llm safety training fail? Advances in Neural Information Processing Systems, 36, 2024. 2 +[35] Yuanwei Wu, Xiang Li, Yixin Liu, Pan Zhou, and Lichao Sun. Jailbreaking gpt-4v via self-adversarial attacks with system prompts. arXiv preprint arXiv:2311.09127, 2023. 2 +[36] Jiahao Yu, Xingwei Lin, Zheng Yu, and Xinyu Xing. {LLM-Fuzzer}: Scaling assessment of large language model jailbreaks. In 33rd USENIX Security Symposium (USENIX Security 24), pages 4657-4674, 2024. 2, 6 +[37] Qifan Yu, Juncheng Li, Longhui Wei, Liang Pang, Wentao Ye, Bosheng Qin, Siliang Tang, Qi Tian, and Yueting Zhuang. Hallucidobot: Mitigating hallucinatory toxicity in visual instruction data. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 12944-12953, 2024. 2 + +[38] Hangfan Zhang, Zhimeng Guo, Huaisheng Zhu, Bochuan Cao, Lu Lin, Jinyuan Jia, Jinghui Chen, and Dinghao Wu. Jailbreak open-sourced large language models via enforced decoding. In Proceedings of the 62nd Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers), pages 5475-5493, 2024. 2 +[39] Jingyi Zhang, Jiaxing Huang, Sheng Jin, and Shijian Lu. Vision-language models for vision tasks: A survey. IEEE Transactions on Pattern Analysis and Machine Intelligence, 2024. 3 +[40] Yunqing Zhao, Tianyu Pang, Chao Du, Xiao Yang, Chongxuan Li, Ngai-Man Man Cheung, and Min Lin. On evaluating adversarial robustness of large vision-language models. Advances in Neural Information Processing Systems, 36, 2024. 2 +[41] Deyao Zhu, Jun Chen, Xiaoqian Shen, Xiang Li, and Mohamed Elhoseiny. Minigpt-4: Enhancing vision-language understanding with advanced large language models. arXiv preprint arXiv:2304.10592, 2023. 3 \ No newline at end of file diff --git a/2025/Distraction is All You Need for Multimodal Large Language Model Jailbreaking/images.zip b/2025/Distraction is All You Need for Multimodal Large Language Model Jailbreaking/images.zip new file mode 100644 index 0000000000000000000000000000000000000000..880641ff9d48d295398293eee838a209de00ffc2 --- /dev/null +++ b/2025/Distraction is All You Need for Multimodal Large Language Model Jailbreaking/images.zip @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:13a89e81053cc00670607eea6ad9241e5100f08542402ae5d173a33a0b34aeff +size 491110 diff --git a/2025/Distraction is All You Need for Multimodal Large Language Model Jailbreaking/layout.json b/2025/Distraction is All You Need for Multimodal Large Language Model Jailbreaking/layout.json new file mode 100644 index 0000000000000000000000000000000000000000..56814b6f49fe90c569b8238930fc69d8c9a15f95 --- /dev/null +++ b/2025/Distraction is All You Need for Multimodal Large Language Model Jailbreaking/layout.json @@ -0,0 +1,7118 @@ +{ + "pdf_info": [ + { + "para_blocks": [ + { + "bbox": [ + 58, + 103, + 552, + 121 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 58, + 103, + 552, + 121 + ], + "spans": [ + { + "bbox": [ + 58, + 103, + 552, + 121 + ], + "type": "text", + "content": "Distraction is All You Need for Multimodal Large Language Model Jailbreaking" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 170, + 142, + 439, + 172 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 170, + 142, + 439, + 172 + ], + "spans": [ + { + "bbox": [ + 170, + 142, + 439, + 172 + ], + "type": "text", + "content": "Zuopeng Yang" + }, + { + "bbox": [ + 170, + 142, + 439, + 172 + ], + "type": "inline_equation", + "content": "^{1*}" + }, + { + "bbox": [ + 170, + 142, + 439, + 172 + ], + "type": "text", + "content": ", Jiluan Fan" + }, + { + "bbox": [ + 170, + 142, + 439, + 172 + ], + "type": "inline_equation", + "content": "^{1*}" + }, + { + "bbox": [ + 170, + 142, + 439, + 172 + ], + "type": "text", + "content": ", Anli Yan" + }, + { + "bbox": [ + 170, + 142, + 439, + 172 + ], + "type": "inline_equation", + "content": "^{1\\dagger}" + }, + { + "bbox": [ + 170, + 142, + 439, + 172 + ], + "type": "text", + "content": ", Erdun Gao" + }, + { + "bbox": [ + 170, + 142, + 439, + 172 + ], + "type": "inline_equation", + "content": "^{3}" + }, + { + "bbox": [ + 170, + 142, + 439, + 172 + ], + "type": "text", + "content": ", Xin Lin" + }, + { + "bbox": [ + 170, + 142, + 439, + 172 + ], + "type": "inline_equation", + "content": "^{1\\dagger}" + }, + { + "bbox": [ + 170, + 142, + 439, + 172 + ], + "type": "text", + "content": ", Tao Li" + }, + { + "bbox": [ + 170, + 142, + 439, + 172 + ], + "type": "inline_equation", + "content": "^{2}" + }, + { + "bbox": [ + 170, + 142, + 439, + 172 + ], + "type": "text", + "content": ", Kanghua Mo" + }, + { + "bbox": [ + 170, + 142, + 439, + 172 + ], + "type": "inline_equation", + "content": "^{1}" + }, + { + "bbox": [ + 170, + 142, + 439, + 172 + ], + "type": "text", + "content": ", Changyu Dong" + }, + { + "bbox": [ + 170, + 142, + 439, + 172 + ], + "type": "inline_equation", + "content": "^{1}" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 134, + 175, + 478, + 205 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 134, + 175, + 478, + 205 + ], + "spans": [ + { + "bbox": [ + 134, + 175, + 478, + 205 + ], + "type": "inline_equation", + "content": "^{1}" + }, + { + "bbox": [ + 134, + 175, + 478, + 205 + ], + "type": "text", + "content": "Guangzhou University " + }, + { + "bbox": [ + 134, + 175, + 478, + 205 + ], + "type": "inline_equation", + "content": "^{2}" + }, + { + "bbox": [ + 134, + 175, + 478, + 205 + ], + "type": "text", + "content": "Shanghai Jiao Tong University " + }, + { + "bbox": [ + 134, + 175, + 478, + 205 + ], + "type": "inline_equation", + "content": "^{3}" + }, + { + "bbox": [ + 134, + 175, + 478, + 205 + ], + "type": "text", + "content": "Australian Institute for Machine Learning, The University of Adelaide" + } + ] + } + ], + "index": 4 + }, + { + "type": "image", + "bbox": [ + 82, + 220, + 527, + 378 + ], + "blocks": [ + { + "bbox": [ + 82, + 220, + 527, + 378 + ], + "lines": [ + { + "bbox": [ + 82, + 220, + 527, + 378 + ], + "spans": [ + { + "bbox": [ + 82, + 220, + 527, + 378 + ], + "type": "image", + "image_path": "7af0484c152bcef177f680586409cb7549584bbb15a8b19ce86a288bd2d1d783.jpg" + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 54, + 381, + 555, + 427 + ], + "lines": [ + { + "bbox": [ + 54, + 381, + 555, + 427 + ], + "spans": [ + { + "bbox": [ + 54, + 381, + 555, + 427 + ], + "type": "text", + "content": "Figure 1. Jailbreaking examples of Hades [12] and our proposed CS-DJ against GPT-4. When using Hades, GPT-4 effectively rejects harmful questions. In contrast, CS-DJ successfully bypasses the model's defense mechanisms by constructing multi-subimage visual inputs to distract the model, resulting in the acquisition of harmful responses. The harmful information is highlighted in red, while key prompts are emphasized in bold." + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "image_caption" + } + ], + "index": 5 + }, + { + "bbox": [ + 152, + 448, + 200, + 460 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 152, + 448, + 200, + 460 + ], + "spans": [ + { + "bbox": [ + 152, + 448, + 200, + 460 + ], + "type": "text", + "content": "Abstract" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 54, + 475, + 297, + 691 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 54, + 475, + 297, + 691 + ], + "spans": [ + { + "bbox": [ + 54, + 475, + 297, + 691 + ], + "type": "text", + "content": "Multimodal Large Language Models (MLLMs) bridge the gap between visual and textual data, enabling a range of advanced applications. However, complex internal interactions among visual elements and their alignment with text can introduce vulnerabilities, which may be exploited to bypass safety mechanisms. To address this, we analyze the relationship between image content and task and find that the complexity of subimages, rather than their content, is key. Building on this insight, we propose the Distraction Hypothesis, followed by a novel framework called Contrasting Subimage Distraction Jailbreaking (CS-DJ), to achieve jailbreaking by disrupting MLLMs alignment through multi-level distraction strategies. CS-DJ consists of two components: structured distraction, achieved through query decomposition that induces a distributional shift by fragmenting harmful prompts into subqueries, and visual-enhanced distraction, realized by constructing contrasting subimages to disrupt the interactions" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 313, + 437, + 556, + 592 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 437, + 556, + 592 + ], + "spans": [ + { + "bbox": [ + 313, + 437, + 556, + 592 + ], + "type": "text", + "content": "among visual elements within the model. This dual strategy disperses the model's attention, reducing its ability to detect and mitigate harmful content. Extensive experiments across five representative scenarios and four popular closed-source MLLMs, including GPT-4o-mini, GPT-4o, GPT-4V, and Gemini-1.5-Flash, demonstrate that CS-DJ achieves average success rates of " + }, + { + "bbox": [ + 313, + 437, + 556, + 592 + ], + "type": "inline_equation", + "content": "52.40\\%" + }, + { + "bbox": [ + 313, + 437, + 556, + 592 + ], + "type": "text", + "content": " for the attack success rate and " + }, + { + "bbox": [ + 313, + 437, + 556, + 592 + ], + "type": "inline_equation", + "content": "74.10\\%" + }, + { + "bbox": [ + 313, + 437, + 556, + 592 + ], + "type": "text", + "content": " for the ensemble attack success rate. These results reveal the potential of distraction-based approaches to exploit and bypass MLLMs' defenses, offering new insights for attack strategies. Our code is available at https://github.com/TeamPigeonLab/CS-DJ." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 313, + 594, + 556, + 618 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 594, + 556, + 618 + ], + "spans": [ + { + "bbox": [ + 313, + 594, + 556, + 618 + ], + "type": "text", + "content": "Warning: This paper contains unfiltered content generated by MLLMs that may be offensive to readers." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 314, + 644, + 395, + 656 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 314, + 644, + 395, + 656 + ], + "spans": [ + { + "bbox": [ + 314, + 644, + 395, + 656 + ], + "type": "text", + "content": "1. Introduction" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 313, + 665, + 555, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 665, + 555, + 713 + ], + "spans": [ + { + "bbox": [ + 313, + 665, + 555, + 713 + ], + "type": "text", + "content": "Since the early 2020s, Multimodal Large Language Models (MLLMs) have emerged as a powerful tool for bridging natural language processing and computer vision [3], owing to their ability to integrate and process multimodal informa" + } + ] + } + ], + "index": 12 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "spans": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "type": "text", + "content": "CVF" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 145, + 0, + 496, + 37 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 145, + 0, + 496, + 37 + ], + "spans": [ + { + "bbox": [ + 145, + 0, + 496, + 37 + ], + "type": "text", + "content": "This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 70, + 702, + 222, + 713 + ], + "type": "page_footnote", + "angle": 0, + "lines": [ + { + "bbox": [ + 70, + 702, + 222, + 713 + ], + "spans": [ + { + "bbox": [ + 70, + 702, + 222, + 713 + ], + "type": "text", + "content": "*Equal Contribution. † Corresponding authors." + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "text", + "content": "9467" + } + ] + } + ], + "index": 14 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 0 + }, + { + "para_blocks": [ + { + "bbox": [ + 55, + 72, + 294, + 251 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 72, + 294, + 251 + ], + "spans": [ + { + "bbox": [ + 55, + 72, + 294, + 251 + ], + "type": "text", + "content": "tion efficiently [25]. Despite their impressive performance, one notable limitation of these methods is that the training materials may contain toxic content, potentially leading to inappropriate or harmful content [4, 38], the leakage of sensitive information [36], and even serious vulnerabilities in the model security and privacy protection. To this end, significant efforts have been devoted to using reinforcement learning from human feedback (RLHF) to align LLMs outputs, ensuring readiness for deployment in high-stakes domains [21, 33]. However, beyond Large Language Models (LLMs), the integration of visual inputs in MLLMs introduces a new challenge: securing models against vulnerabilities arising from newly integrated visual modalities, which remains a significant issue and lacks a comprehensive understanding [17, 18, 37]" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 55, + 255, + 295, + 422 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 255, + 295, + 422 + ], + "spans": [ + { + "bbox": [ + 55, + 255, + 295, + 422 + ], + "type": "text", + "content": "In this study, rather than following the human feedback alignment research at the defense side, we focus on exploring an alternative approach, namely jailbreak attacks [19, 34, 36]. This line of research also seeks to provide insights into how safety concerns emerge, while, considering the attacker's perspective. Specifically, it aims to actively uncover the defense vulnerabilities in current model by deliberately crafting inputs to manipulate MLLMs, leading to unintended and potentially harmful outputs [15]. Motivated by the fact that most current defense mechanisms rely on RLHF and are trained on human preference data [21], a natural hypothesis regarding the defense vulnerabilities of MLLMs is that the model may be susceptible to out-of-distribution (OOD) inputs." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 55, + 426, + 295, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 426, + 295, + 714 + ], + "spans": [ + { + "bbox": [ + 55, + 426, + 295, + 714 + ], + "type": "text", + "content": "Given that harmful textual content is already effectively detected in the field of LLMs, current jailbreak attack research on MLLMs primarily focuses on constructing OOD visual inputs, with two main approaches: image perturbation injection and prompt-to-image infection. The image perturbation injection involves subtly modifying images and combining them with query text to compose jailbreak prompts [35]. This attack leverages adversarial approaches [30] to introduce noisy OOD inputs that exploit the MLLM's visual-text decision boundary [40]. Unfortunately, these methods typically require access to the gradient of the victim or surrogate model [32], which limits their applicability in closed-source models [29]. In contrast, prompt-to-image infection generates images with similar semantics by using malicious text. These generated images are then combined with query text to craft jailbreak prompts that compel the MLLM to produce harmful responses [6, 12]. Since these generated images differ in distribution from those used for MLLM safety alignment training, they may bypass security measures and achieve jailbreak [2, 16, 26]. Despite this, the extensive training of large MLLMs on vast datasets makes it challenging for image generation models to produce truly OOD images. A failure case is depicted in the upper portion of Figure 1." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 313, + 72, + 553, + 251 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 72, + 553, + 251 + ], + "spans": [ + { + "bbox": [ + 313, + 72, + 553, + 251 + ], + "type": "text", + "content": "Findings. Despite progress, it remains unclear how to fully leverage the visual space to construct OOD inputs for effective black-box MLLM jailbreaks. Building on the observation that most harmless alignment processes primarily involve simple images [6, 12, 22], we propose that increasing image complexity, potentially through the use of multiple subimages, could generate more effective OOD inputs for jailbreak attack. To verify this, we conduct several ablation analyses (§4.4), including varying the number of subimages and examining the relationship between image content and task. Our findings reveal that it is the complexity of the subimages, rather than their conceptual content, that drives the jailbreak success. Based on this insight, we propose the Distraction Hypothesis to explain the visual effects on jailbreak attacks against MLLMs." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 319, + 268, + 550, + 341 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 319, + 268, + 550, + 341 + ], + "spans": [ + { + "bbox": [ + 319, + 268, + 550, + 341 + ], + "type": "text", + "content": "Distraction Hypothesis. Encoding complex images in the input prompt increases token complexity/diversity, which raises the processing burden on MLLMs. This overload can weaken the model's defenses, making it more prone to induce unintended outputs and improving jailbreak attack effectiveness." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 313, + 354, + 553, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 354, + 553, + 713 + ], + "spans": [ + { + "bbox": [ + 313, + 354, + 553, + 713 + ], + "type": "text", + "content": "Structure and contributions. Building on the Distraction Hypothesis, we further introduce a novel framework, namely Contrasting Subimage Distraction Jailbreaking (CS-DJ) " + }, + { + "bbox": [ + 313, + 354, + 553, + 713 + ], + "type": "inline_equation", + "content": "(\\S 3)" + }, + { + "bbox": [ + 313, + 354, + 553, + 713 + ], + "type": "text", + "content": ", which leverages this insight to strategically design OOD inputs and enhance the effectiveness of jailbreak attacks. Specifically, CS-DJ explores both textual and visual components of the input space, introducing two key methods: structured distraction via query decomposition " + }, + { + "bbox": [ + 313, + 354, + 553, + 713 + ], + "type": "inline_equation", + "content": "(\\S 3.1)" + }, + { + "bbox": [ + 313, + 354, + 553, + 713 + ], + "type": "text", + "content": " and visual-enhanced distraction through contrasting subimages " + }, + { + "bbox": [ + 313, + 354, + 553, + 713 + ], + "type": "inline_equation", + "content": "(\\S 3.2)" + }, + { + "bbox": [ + 313, + 354, + 553, + 713 + ], + "type": "text", + "content": ". (i) The structured distraction component decomposes the original harmful query into multiple sub-queries, each representing different aspects or intermediate steps of the original query. This decomposition induces a distributional shift that disperses the model's focus and weakens its ability to detect harmful content. (ii) The visual-enhanced distraction component manipulates the visual input by constructing multiple contrasting subimages, effectively disrupting interactions among visual elements during processing. Then, to guide the model's response and enhance its susceptibility to the multi-level distractions, CS-DJ incorporates a carefully designed harmless prompt alongside the composite visual input, as illustrated in Figure 1. This integrated approach disrupts the model's internal coherence and safety mechanisms, effectively bypassing its defenses. To evaluate the effectiveness of CS-DJ, we conduct extensive experiments across five representative scenarios and four widely used closed-source MLLMs " + }, + { + "bbox": [ + 313, + 354, + 553, + 713 + ], + "type": "inline_equation", + "content": "(\\S 4)" + }, + { + "bbox": [ + 313, + 354, + 553, + 713 + ], + "type": "text", + "content": ". Our results demonstrate that CS-DJ outperforms state-of-the-art jailbreak attacks, achieving average success rates of " + }, + { + "bbox": [ + 313, + 354, + 553, + 713 + ], + "type": "inline_equation", + "content": "52.40\\%" + }, + { + "bbox": [ + 313, + 354, + 553, + 713 + ], + "type": "text", + "content": " for attack success rate and " + }, + { + "bbox": [ + 313, + 354, + 553, + 713 + ], + "type": "inline_equation", + "content": "74.10\\%" + }, + { + "bbox": [ + 313, + 354, + 553, + 713 + ], + "type": "text", + "content": " for ensemble" + } + ] + } + ], + "index": 5 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "text", + "content": "9468" + } + ] + } + ], + "index": 6 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 1 + }, + { + "para_blocks": [ + { + "bbox": [ + 56, + 73, + 135, + 83 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 73, + 135, + 83 + ], + "spans": [ + { + "bbox": [ + 56, + 73, + 135, + 83 + ], + "type": "text", + "content": "attack success rate." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 55, + 101, + 142, + 113 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 101, + 142, + 113 + ], + "spans": [ + { + "bbox": [ + 55, + 101, + 142, + 113 + ], + "type": "text", + "content": "2. Related Work" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 55, + 122, + 250, + 136 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 122, + 250, + 136 + ], + "spans": [ + { + "bbox": [ + 55, + 122, + 250, + 136 + ], + "type": "text", + "content": "2.1. Multimodal Large Language Models" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 55, + 142, + 296, + 418 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 142, + 296, + 418 + ], + "spans": [ + { + "bbox": [ + 55, + 142, + 296, + 418 + ], + "type": "text", + "content": "Recently, Multimodal Large Language Models (MLLMs) [31] have become a prominent approach for processing and understanding heterogeneous multimodal data, including visual[39] and auditory [5] modalities. Building on instruction-tuning techniques developed for LLMs, most MLLMs are adapted from pre-trained LLMs through fine-tuning. Notable examples include LLaVA [14] and MiniGPT-4 [41], which share a similar architectural design by integrating pre-trained vision backbones with large language models via a projection layer, enabling effective vision-language alignment. Building upon this foundation, LLaVA-1.5 [13] further enhances its architecture by adopting a two-layer MLP for better vision-language mapping and increasing the vision encoder's input resolution for improved visual understanding. However, the incorporation of emerging modalities further intensifies the challenges of ensuring robust safety alignment in MLLMs. To address this challenge, advanced MLLMs, such as GPT-4o [7] and Gemini 1.5-Flash [27], employ a combination of alignment strategies, including post-training methods and integrated classifiers, to enhance both helpfulness and safety. Despite these efforts, MLLMs remains vulnerable to attacks using OOD inputs, which will be explored in this work." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 55, + 430, + 218, + 443 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 430, + 218, + 443 + ], + "spans": [ + { + "bbox": [ + 55, + 430, + 218, + 443 + ], + "type": "text", + "content": "2.2. Jailbreak Attacks on MLLMs" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 55, + 450, + 295, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 450, + 295, + 714 + ], + "spans": [ + { + "bbox": [ + 55, + 450, + 295, + 714 + ], + "type": "text", + "content": "Jailbreak attacks against MLLM can be performed through visual and textual input, leading to complex and diverse attack patterns [9, 20]. As a result, numerous jailbreak methods [6, 12, 17, 22, 24] have been continuously proposed. Qi et al. [22] used visual adversarial examples to bypass the security mechanisms of visual-textual MLLMs. However, this type of attack [20, 22] is typically only effective in white-box environments and is often difficult to apply to state-of-the-art black-box models. FigStep [6] introduced a simple yet effective black-box jailbreak approach that circumvents safety alignment bytrasfering prohibited content as images using typographic transformations. MM-SafetyBench [17] utilized a multi-step approach to construct jailbreak input, incorporating question generation, rephrasing, extraction of unsafe key phrases, and query-to-image conversion. Furthermore, Hades [12] investigated the influence of image harmfulness on model jailbreaks and proposed a prompt-embedded image infection jailbreak attack method. It generated harmful images using diffusion models and amplified the jailbreak effect through optimization. Unlike these works, CS-DJ explores the safety of MLLMs from a new perspective—visual distraction." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 314, + 71, + 397, + 85 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 314, + 71, + 397, + 85 + ], + "spans": [ + { + "bbox": [ + 314, + 71, + 397, + 85 + ], + "type": "text", + "content": "3. Methodology" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 313, + 91, + 555, + 344 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 91, + 555, + 344 + ], + "spans": [ + { + "bbox": [ + 313, + 91, + 555, + 344 + ], + "type": "text", + "content": "Beyond the standard LLMs, MLLMs also encode visual content into token sequences, saying that the intrinsic interactions among visual elements can affect the model's ability to defend against harmful prompts. However, most of the existing approaches [12, 16, 17] only take a single, complete image as input, resulting in notable similarities both (1) within image patches, and (2) between the image and the targeted jailbreak content. Such similarities may make the model likely to detect harmful inputs, potentially leading to jailbreak failure. To address these issues, we propose a novel jailbreak framework, called CS-DJ, as illustrated in Figure 2. CS-DJ leverages a multi-subimage structure to bypass the models' intrinsic safety detection mechanism, with introducing multi-level distraction. Specifically, CS-DJ consists of two main components: structured distraction, achieved through query decomposition, and visual-enhanced distraction, realized through the construction of contrasting subimages. The resulting composite image input, combined with a carefully designed instruction, is fed into MLLMs to achieve jailbreak. Further details are described in the following sections." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 313, + 350, + 555, + 374 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 350, + 555, + 374 + ], + "spans": [ + { + "bbox": [ + 313, + 350, + 555, + 374 + ], + "type": "text", + "content": "3.1. Structured Distraction via Query Decomposition" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 313, + 380, + 556, + 514 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 380, + 556, + 514 + ], + "spans": [ + { + "bbox": [ + 313, + 380, + 556, + 514 + ], + "type": "text", + "content": "MLLMs have strong defense ability against harmful textual inputs, reinforced by specific optimizations aligned with human preference data [10]. To circumvent these defenses, we introduce a structured distraction strategy that disrupts and fragments information alignment through query decomposition, thereby reducing the model's overall sensitivity to harmful content. As illustrated on the left of Figure 2, we begin by decomposing a original harmful query " + }, + { + "bbox": [ + 313, + 380, + 556, + 514 + ], + "type": "inline_equation", + "content": "Q" + }, + { + "bbox": [ + 313, + 380, + 556, + 514 + ], + "type": "text", + "content": " into multiple sub-queries, denoted as " + }, + { + "bbox": [ + 313, + 380, + 556, + 514 + ], + "type": "inline_equation", + "content": "\\{Q_{s}^{(i)}\\}_{i = 1}^{m}" + }, + { + "bbox": [ + 313, + 380, + 556, + 514 + ], + "type": "text", + "content": ". Each sub-query addresses part of the original query from different aspects or intermediate steps." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 396, + 523, + 555, + 537 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 396, + 523, + 555, + 537 + ], + "spans": [ + { + "bbox": [ + 396, + 523, + 555, + 537 + ], + "type": "interline_equation", + "content": "\\left\\{Q _ {s} ^ {i} \\right\\} _ {i = 1} ^ {m} = \\mathcal {G} (Q), \\tag {1}", + "image_path": "26188cd3f90c710e36f64c94e0d79ccb62a7cf3c99a8b79ccdfd03292dd66630.jpg" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 313, + 545, + 556, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 545, + 556, + 715 + ], + "spans": [ + { + "bbox": [ + 313, + 545, + 556, + 715 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 313, + 545, + 556, + 715 + ], + "type": "inline_equation", + "content": "\\mathcal{G}" + }, + { + "bbox": [ + 313, + 545, + 556, + 715 + ], + "type": "text", + "content": " is the auxiliary decomposition model and " + }, + { + "bbox": [ + 313, + 545, + 556, + 715 + ], + "type": "inline_equation", + "content": "m" + }, + { + "bbox": [ + 313, + 545, + 556, + 715 + ], + "type": "text", + "content": " denotes the number of sub-queries. On one hand, this fragmented structure distracts the model's focus, creating a distributional shift that obscures the original intent and disrupts its ability to detect harmful content, thereby bypassing its safety mechanisms. On the other hand, during the jailbreak execution phase, we carefully design prompts that guide the MLLM to address all sub-queries simultaneously. This strategy causes the model to divide its attention among multiple tasks rather than focusing on a single query. Further details are discussed in Section 3.3. In practice, we utilized Qwen2.5-3B-Instruct [28] for query decomposition. The details of the query decomposition prompt are provided in the Appendix." + } + ] + } + ], + "index": 11 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 315, + 757 + ], + "type": "text", + "content": "9469" + } + ] + } + ], + "index": 12 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 2 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 85, + 74, + 231, + 312 + ], + "blocks": [ + { + "bbox": [ + 85, + 74, + 231, + 312 + ], + "lines": [ + { + "bbox": [ + 85, + 74, + 231, + 312 + ], + "spans": [ + { + "bbox": [ + 85, + 74, + 231, + 312 + ], + "type": "image", + "image_path": "c2ee3ec1c2cea2ffde7b1f63dab9159d9590b2ef998d6ac52d06322facb64316.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 55, + 323, + 555, + 368 + ], + "lines": [ + { + "bbox": [ + 55, + 323, + 555, + 368 + ], + "spans": [ + { + "bbox": [ + 55, + 323, + 555, + 368 + ], + "type": "text", + "content": "Figure 2. The framework of the proposed CS-DJ. Given a harmful query, CS-DJ employs a three-step process to execute a jailbreaking attack: (1) decompose the raw query into multiple sub-queries and transform them into images to introduce structural distraction, (2) retrieve contrasting images from a dataset as subimages to inject visual-enhanced distraction. (3) combine these images into a composite image as the visual input, aligned with a harmless instruction, to execute the jailbreaking attack." + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "type": "image", + "bbox": [ + 232, + 74, + 378, + 312 + ], + "blocks": [ + { + "bbox": [ + 232, + 74, + 378, + 312 + ], + "lines": [ + { + "bbox": [ + 232, + 74, + 378, + 312 + ], + "spans": [ + { + "bbox": [ + 232, + 74, + 378, + 312 + ], + "type": "image", + "image_path": "27a8d9dac43b96769b81738d6f11734123a2e3f710058588f76ce8303133747d.jpg" + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_body" + } + ], + "index": 1 + }, + { + "type": "image", + "bbox": [ + 379, + 74, + 526, + 312 + ], + "blocks": [ + { + "bbox": [ + 379, + 74, + 526, + 312 + ], + "lines": [ + { + "bbox": [ + 379, + 74, + 526, + 312 + ], + "spans": [ + { + "bbox": [ + 379, + 74, + 526, + 312 + ], + "type": "image", + "image_path": "6e1451c85d535c0e198c4edc972e0cc8e0bbb684b7cfdbfbbce6715411128a3b.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + } + ], + "index": 2 + }, + { + "bbox": [ + 55, + 383, + 295, + 407 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 383, + 295, + 407 + ], + "spans": [ + { + "bbox": [ + 55, + 383, + 295, + 407 + ], + "type": "text", + "content": "The next step involves converting the sub-queries into images:" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 124, + 405, + 295, + 421 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 124, + 405, + 295, + 421 + ], + "spans": [ + { + "bbox": [ + 124, + 405, + 295, + 421 + ], + "type": "interline_equation", + "content": "\\left\\{I _ {s} ^ {i} \\right\\} _ {i = 1} ^ {m} = \\mathcal {T} \\left(\\left\\{Q _ {s} ^ {i} \\right\\} _ {i = 1} ^ {m}\\right), \\tag {2}", + "image_path": "b7a066ef366795ac37eaea89f9b57847da337f6535195d82a4a6130c9372be35.jpg" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 55, + 426, + 296, + 509 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 426, + 296, + 509 + ], + "spans": [ + { + "bbox": [ + 55, + 426, + 296, + 509 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 55, + 426, + 296, + 509 + ], + "type": "inline_equation", + "content": "\\mathcal{T}" + }, + { + "bbox": [ + 55, + 426, + 296, + 509 + ], + "type": "text", + "content": " denotes the transformation function that converts each sub-query " + }, + { + "bbox": [ + 55, + 426, + 296, + 509 + ], + "type": "inline_equation", + "content": "Q_{s}^{i}" + }, + { + "bbox": [ + 55, + 426, + 296, + 509 + ], + "type": "text", + "content": " into a corresponding image " + }, + { + "bbox": [ + 55, + 426, + 296, + 509 + ], + "type": "inline_equation", + "content": "I_{s}^{i}" + }, + { + "bbox": [ + 55, + 426, + 296, + 509 + ], + "type": "text", + "content": ". This transformation introduces a new form of distraction by altering the input modality, complicating the model's ability to link and identify harmful content patterns. Presenting textual prompts as visual elements obscures their original intent, further disrupting the model's detection mechanisms." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 55, + 516, + 296, + 541 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 516, + 296, + 541 + ], + "spans": [ + { + "bbox": [ + 55, + 516, + 296, + 541 + ], + "type": "text", + "content": "3.2. Visual-Enhanced Distraction via Multi-subimage Construction" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 55, + 546, + 296, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 546, + 296, + 713 + ], + "spans": [ + { + "bbox": [ + 55, + 546, + 296, + 713 + ], + "type": "text", + "content": "In contrast to prior work that emphasizes the harmfulness of a single image for MLLM jailbreaks, we focus on the impact of distraction. To enhance visual input distraction, we construct multiple contrasting visual elements, each treated as a subimage that collectively forms the input image. Since MLLMs represent visual content as token sequences similar to text, the distraction must include two aspects: (1) the distraction between the text query and the visual subimage, and (2) the distraction among the visual subimages themselves. Directly constructing an image with sufficient distraction relative to a text query is challenging. Therefore, we simplify this process by transforming it into an image retrieval problem, aimed at retrieving " + }, + { + "bbox": [ + 55, + 546, + 296, + 713 + ], + "type": "inline_equation", + "content": "k" + }, + { + "bbox": [ + 55, + 546, + 296, + 713 + ], + "type": "text", + "content": " images from a dataset. These " + }, + { + "bbox": [ + 55, + 546, + 296, + 713 + ], + "type": "inline_equation", + "content": "k" + }, + { + "bbox": [ + 55, + 546, + 296, + 713 + ], + "type": "text", + "content": " images are selected to have the lowest similarity to" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 313, + 383, + 555, + 443 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 383, + 555, + 443 + ], + "spans": [ + { + "bbox": [ + 313, + 383, + 555, + 443 + ], + "type": "text", + "content": "the query and to each other. Given the complexity of this optimization problem, we further approximate it for practical implementation. The procedure is illustrated in the middle of Figure 2. Specifically, we first encode the query " + }, + { + "bbox": [ + 313, + 383, + 555, + 443 + ], + "type": "inline_equation", + "content": "Q" + }, + { + "bbox": [ + 313, + 383, + 555, + 443 + ], + "type": "text", + "content": " as a dense vector using the CLIP model:" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 395, + 453, + 555, + 468 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 395, + 453, + 555, + 468 + ], + "spans": [ + { + "bbox": [ + 395, + 453, + 555, + 468 + ], + "type": "interline_equation", + "content": "v (Q) = \\operatorname {C L I P} (Q). \\tag {3}", + "image_path": "c9a298cf377f6ccee283038e53a8f0803b08e3a31993c5329cc278d257b73a97.jpg" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 313, + 478, + 554, + 523 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 478, + 554, + 523 + ], + "spans": [ + { + "bbox": [ + 313, + 478, + 554, + 523 + ], + "type": "text", + "content": "In practice, CLIP-ViT-L/14 [23] is utilized as the extractor. We then retrieve an image from a dataset that most contrasts with " + }, + { + "bbox": [ + 313, + 478, + 554, + 523 + ], + "type": "inline_equation", + "content": "v(Q)" + }, + { + "bbox": [ + 313, + 478, + 554, + 523 + ], + "type": "text", + "content": ", effectively by minimizing the cosine similarity as:" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 370, + 527, + 555, + 545 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 370, + 527, + 555, + 545 + ], + "spans": [ + { + "bbox": [ + 370, + 527, + 555, + 545 + ], + "type": "interline_equation", + "content": "I _ {c} ^ {1} = \\arg \\min _ {I \\in \\mathcal {D}} \\mathbf {c o s} (v (Q), v (I)), \\tag {4}", + "image_path": "e94280a5b55ceea69a3772ec1cd00015ac2456dc84d47be132b0724efc55ee10.jpg" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 313, + 552, + 554, + 588 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 552, + 554, + 588 + ], + "spans": [ + { + "bbox": [ + 313, + 552, + 554, + 588 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 313, + 552, + 554, + 588 + ], + "type": "inline_equation", + "content": "\\mathcal{D}" + }, + { + "bbox": [ + 313, + 552, + 554, + 588 + ], + "type": "text", + "content": " denotes the image dataset, and " + }, + { + "bbox": [ + 313, + 552, + 554, + 588 + ], + "type": "inline_equation", + "content": "\\cos (\\cdot ,\\cdot)" + }, + { + "bbox": [ + 313, + 552, + 554, + 588 + ], + "type": "text", + "content": " represents the cosine similarity. Next, we proceed to retrieve the subsequent subimages:" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 313, + 598, + 553, + 641 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 598, + 553, + 641 + ], + "spans": [ + { + "bbox": [ + 313, + 598, + 553, + 641 + ], + "type": "interline_equation", + "content": "I _ {c} ^ {j} = \\arg \\min _ {I \\in \\mathcal {D}} \\left(\\mathbf {c o s} (v (Q), v (I)) + \\sum_ {i = 1} ^ {j - 1} \\mathbf {c o s} \\left(v \\left(I _ {c} ^ {i}\\right), v (I)\\right)\\right), \\tag {5}", + "image_path": "2b56b8570a144c9b8305252d5170a28e69fdf68e79e9698014878867e4c25380.jpg" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 313, + 642, + 555, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 642, + 555, + 714 + ], + "spans": [ + { + "bbox": [ + 313, + 642, + 555, + 714 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 313, + 642, + 555, + 714 + ], + "type": "inline_equation", + "content": "j" + }, + { + "bbox": [ + 313, + 642, + 555, + 714 + ], + "type": "text", + "content": " denotes the index of the current subimage being selected. This approach systematically ensures that each subimage has minimal similarity to both the query and the previously selected subimages, thereby maximizing contrast and enhancing the overall distraction effect for the MLLM jailbreak process." + } + ] + } + ], + "index": 15 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "text", + "content": "9470" + } + ] + } + ], + "index": 16 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 3 + }, + { + "para_blocks": [ + { + "bbox": [ + 55, + 72, + 187, + 85 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 72, + 187, + 85 + ], + "spans": [ + { + "bbox": [ + 55, + 72, + 187, + 85 + ], + "type": "text", + "content": "3.3. Jailbreaking Execution" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 55, + 89, + 296, + 152 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 89, + 296, + 152 + ], + "spans": [ + { + "bbox": [ + 55, + 89, + 296, + 152 + ], + "type": "text", + "content": "With the constructed set of subimages, we form a composite input that maximizes distraction for MLLMs. Specifically, the final composite image " + }, + { + "bbox": [ + 55, + 89, + 296, + 152 + ], + "type": "inline_equation", + "content": "I_{\\mathrm{comp}}" + }, + { + "bbox": [ + 55, + 89, + 296, + 152 + ], + "type": "text", + "content": " is created by combining the subimages " + }, + { + "bbox": [ + 55, + 89, + 296, + 152 + ], + "type": "inline_equation", + "content": "\\{I_c^j\\}_{j=1}^k" + }, + { + "bbox": [ + 55, + 89, + 296, + 152 + ], + "type": "text", + "content": " along with the transformed sub-queries " + }, + { + "bbox": [ + 55, + 89, + 296, + 152 + ], + "type": "inline_equation", + "content": "\\{I_s^i\\}_{i=1}^m" + }, + { + "bbox": [ + 55, + 89, + 296, + 152 + ], + "type": "text", + "content": ":" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 93, + 159, + 296, + 178 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 93, + 159, + 296, + 178 + ], + "spans": [ + { + "bbox": [ + 93, + 159, + 296, + 178 + ], + "type": "interline_equation", + "content": "I _ {\\text {c o m p}} = \\text {C o m b i n e} \\left(\\left\\{I _ {c} ^ {j} \\right\\} _ {j = 1} ^ {k}, \\left\\{I _ {s} ^ {i} \\right\\} _ {i = 1} ^ {m}\\right), \\tag {6}", + "image_path": "0837a7ddc11edbd54ccbcf241b7e90c2cdbc1bae6c9af1c1e03653e2e0927af2.jpg" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 55, + 182, + 296, + 217 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 182, + 296, + 217 + ], + "spans": [ + { + "bbox": [ + 55, + 182, + 296, + 217 + ], + "type": "text", + "content": "where the function " + }, + { + "bbox": [ + 55, + 182, + 296, + 217 + ], + "type": "inline_equation", + "content": "\\mathbf{Combine}(\\cdot, \\cdot)" + }, + { + "bbox": [ + 55, + 182, + 296, + 217 + ], + "type": "text", + "content": " represents the structured arrangement of the subimages and sub-query transformations." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 55, + 218, + 296, + 243 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 218, + 296, + 243 + ], + "spans": [ + { + "bbox": [ + 55, + 218, + 296, + 243 + ], + "type": "text", + "content": "This composite input " + }, + { + "bbox": [ + 55, + 218, + 296, + 243 + ], + "type": "inline_equation", + "content": "I_{\\mathrm{comp}}" + }, + { + "bbox": [ + 55, + 218, + 296, + 243 + ], + "type": "text", + "content": ", along with a harmless instruction " + }, + { + "bbox": [ + 55, + 218, + 296, + 243 + ], + "type": "inline_equation", + "content": "P" + }, + { + "bbox": [ + 55, + 218, + 296, + 243 + ], + "type": "text", + "content": ", is fed into a MLLM for processing:" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 126, + 254, + 295, + 268 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 126, + 254, + 295, + 268 + ], + "spans": [ + { + "bbox": [ + 126, + 254, + 295, + 268 + ], + "type": "interline_equation", + "content": "Y = \\operatorname {M L L M} \\left(I _ {\\text {c o m p}}, P\\right), \\tag {7}", + "image_path": "d6136bcfbab92c9b878943804e38a0d154d18a7f30b78649e729b9830286253f.jpg" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 55, + 270, + 259, + 282 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 270, + 259, + 282 + ], + "spans": [ + { + "bbox": [ + 55, + 270, + 259, + 282 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 55, + 270, + 259, + 282 + ], + "type": "inline_equation", + "content": "Y" + }, + { + "bbox": [ + 55, + 270, + 259, + 282 + ], + "type": "text", + "content": " denotes the textual output of the MLLM." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 55, + 283, + 296, + 498 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 283, + 296, + 498 + ], + "spans": [ + { + "bbox": [ + 55, + 283, + 296, + 498 + ], + "type": "text", + "content": "The use of " + }, + { + "bbox": [ + 55, + 283, + 296, + 498 + ], + "type": "inline_equation", + "content": "I_{\\mathrm{comp}}" + }, + { + "bbox": [ + 55, + 283, + 296, + 498 + ], + "type": "text", + "content": " with contrasting visual elements maximizes distraction, effectively disrupting the model's internal alignment and coherence mechanisms. To complement " + }, + { + "bbox": [ + 55, + 283, + 296, + 498 + ], + "type": "inline_equation", + "content": "I_{\\mathrm{comp}}" + }, + { + "bbox": [ + 55, + 283, + 296, + 498 + ], + "type": "text", + "content": ", we carefully design " + }, + { + "bbox": [ + 55, + 283, + 296, + 498 + ], + "type": "inline_equation", + "content": "P" + }, + { + "bbox": [ + 55, + 283, + 296, + 498 + ], + "type": "text", + "content": " to further enhance the distraction effect. As illustrated in Figure 1, " + }, + { + "bbox": [ + 55, + 283, + 296, + 498 + ], + "type": "inline_equation", + "content": "P" + }, + { + "bbox": [ + 55, + 283, + 296, + 498 + ], + "type": "text", + "content": " comprises three main components: (1) a role-guiding section, (2) a task-guiding section, and (3) a visual-guiding section. The role-guiding strategy is commonly employed across various LLM tasks. The second component of " + }, + { + "bbox": [ + 55, + 283, + 296, + 498 + ], + "type": "inline_equation", + "content": "P" + }, + { + "bbox": [ + 55, + 283, + 296, + 498 + ], + "type": "text", + "content": " directs the model to process multiple sub-queries simultaneously, increasing task complexity and dispersing attention across multiple tasks instead of focusing on a single one. Finally, " + }, + { + "bbox": [ + 55, + 283, + 296, + 498 + ], + "type": "inline_equation", + "content": "P" + }, + { + "bbox": [ + 55, + 283, + 296, + 498 + ], + "type": "text", + "content": " includes a misleading instruction, informing the model that the visual subimages may contain useful information, further diverting its attention. By presenting these inputs together, the approach reduces the model's ability to detect and mitigate harmful content, effectively bypassing its intrinsic safety mechanisms through multi-level distraction." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 55, + 505, + 241, + 517 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 505, + 241, + 517 + ], + "spans": [ + { + "bbox": [ + 55, + 505, + 241, + 517 + ], + "type": "text", + "content": "3.4. Distraction Distance Measurement" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 55, + 522, + 295, + 583 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 522, + 295, + 583 + ], + "spans": [ + { + "bbox": [ + 55, + 522, + 295, + 583 + ], + "type": "text", + "content": "To quantify the effectiveness of our multi-subimage input for distraction, we introduce a metric called Distraction Distance. This metric evaluates interactions among the query " + }, + { + "bbox": [ + 55, + 522, + 295, + 583 + ], + "type": "inline_equation", + "content": "Q" + }, + { + "bbox": [ + 55, + 522, + 295, + 583 + ], + "type": "text", + "content": " and contrasting visual images " + }, + { + "bbox": [ + 55, + 522, + 295, + 583 + ], + "type": "inline_equation", + "content": "\\{I_c^j\\}_{j=1}^k" + }, + { + "bbox": [ + 55, + 522, + 295, + 583 + ], + "type": "text", + "content": ", treating all as equally important nodes within a unified structure." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 55, + 584, + 296, + 643 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 584, + 296, + 643 + ], + "spans": [ + { + "bbox": [ + 55, + 584, + 296, + 643 + ], + "type": "text", + "content": "Each node is represented by its CLIP-encoded vector. For a given node, we calculate the " + }, + { + "bbox": [ + 55, + 584, + 296, + 643 + ], + "type": "inline_equation", + "content": "\\mathrm{L}_2" + }, + { + "bbox": [ + 55, + 584, + 296, + 643 + ], + "type": "text", + "content": " distance between its vector and the vectors of all other nodes. The sum of these " + }, + { + "bbox": [ + 55, + 584, + 296, + 643 + ], + "type": "inline_equation", + "content": "\\mathrm{L}_2" + }, + { + "bbox": [ + 55, + 584, + 296, + 643 + ], + "type": "text", + "content": " distances for each node is calculated across all other nodes to obtain the Distraction Distance:" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 119, + 649, + 295, + 684 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 119, + 649, + 295, + 684 + ], + "spans": [ + { + "bbox": [ + 119, + 649, + 295, + 684 + ], + "type": "interline_equation", + "content": "D L = \\sum_ {i = 1} ^ {N} \\sum_ {j \\neq i} | | v _ {i} - v _ {j} | | _ {2}, \\tag {8}", + "image_path": "ad12eaf438ef73260d2351474d29a3afe9456321815ac72a5a3b109b6b1dbef0.jpg" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 55, + 689, + 296, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 689, + 296, + 715 + ], + "spans": [ + { + "bbox": [ + 55, + 689, + 296, + 715 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 55, + 689, + 296, + 715 + ], + "type": "inline_equation", + "content": "N" + }, + { + "bbox": [ + 55, + 689, + 296, + 715 + ], + "type": "text", + "content": " denotes the total number of nodes, and " + }, + { + "bbox": [ + 55, + 689, + 296, + 715 + ], + "type": "inline_equation", + "content": "v_{i}" + }, + { + "bbox": [ + 55, + 689, + 296, + 715 + ], + "type": "text", + "content": " represents the CLIP-encoded vector of node " + }, + { + "bbox": [ + 55, + 689, + 296, + 715 + ], + "type": "inline_equation", + "content": "i" + }, + { + "bbox": [ + 55, + 689, + 296, + 715 + ], + "type": "text", + "content": ". This approach ac" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 313, + 72, + 554, + 144 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 72, + 554, + 144 + ], + "spans": [ + { + "bbox": [ + 313, + 72, + 554, + 144 + ], + "type": "text", + "content": "counts for both the number of nodes and their pairwise distances, providing a comprehensive measure of the dispersion and contrast within the multi-subimage structure. By maximizing this Distraction Distance, we amplify the overall distracting effect on MLLMs, thereby disrupting their ability to accurately detect harmful content." + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 313, + 158, + 391, + 172 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 158, + 391, + 172 + ], + "spans": [ + { + "bbox": [ + 313, + 158, + 391, + 172 + ], + "type": "text", + "content": "4. Experiment" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 313, + 178, + 554, + 262 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 178, + 554, + 262 + ], + "spans": [ + { + "bbox": [ + 313, + 178, + 554, + 262 + ], + "type": "text", + "content": "In this section, we first introduced our experimental settings which include datasets and evaluation metrics. Then, we carried out quantitative and qualitative comparisons between our method and state-of-the-art MLLM jailbreaking approaches. Additionally, ablation studies are performed to validate the distraction effect on MLLM jailbreaking from various perspectives." + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 313, + 273, + 423, + 285 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 273, + 423, + 285 + ], + "spans": [ + { + "bbox": [ + 313, + 273, + 423, + 285 + ], + "type": "text", + "content": "4.1. Experiment Setup" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 313, + 291, + 555, + 482 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 291, + 555, + 482 + ], + "spans": [ + { + "bbox": [ + 313, + 291, + 555, + 482 + ], + "type": "text", + "content": "Dataset. The datasets are primarily utilized to provide or construct visual inputs tailored to the requirements of different methods. For evaluating Hades, we followed the experimental setting outlined in its original paper, utilizing the Hades dataset [12] to supply the visual inputs. This dataset comprises five representative categories related to real-world visual information: (1) Violence, (2) Financial, (3) Privacy, (4) Self-Harm, and (5) Animal. Each category contains 150 queries, resulting in a total of 750 harmful queries. For evaluating CS-DJ, we employed LLaVACC3M-Pretrain595K to retrieve the contrasting images. To improve retrieval efficiency, we randomly selected a subset of 10,000 images for the retrieval dataset, using a different random seed for each experiment to ensure diversity. All harmful queries used throughout the experiments were derived from the Hades dataset." + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 313, + 483, + 554, + 567 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 483, + 554, + 567 + ], + "spans": [ + { + "bbox": [ + 313, + 483, + 554, + 567 + ], + "type": "text", + "content": "Victim Models. In our main experiments, We evaluated the performance of the four most popular closed-source MLLMs: GPT-4o-Mini, GPT-4o, GPT-4V [1], and Gemini-1.5-Flash [27]. The specific versions are as follows: GPT-4o: GPT-4o-2024-08-06, GPT-4o-Mini: GPT-4o-Mini-2024-07-18, GPT-4V: GPT-4-1106- vision-preview, and Gemini-1.5-flash: Gemini-1.5-flash-001." + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 313, + 567, + 554, + 664 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 567, + 554, + 664 + ], + "spans": [ + { + "bbox": [ + 313, + 567, + 554, + 664 + ], + "type": "text", + "content": "Evaluation Metrics. To assess CS-DJ, we employ two primary evaluation metrics: Attack Success Rate (ASR) and Ensemble Attack Success Rate (EASR). ASR [6, 12] quantifies the proportion of successful jailbreak attempts by assessing whether a model's response meets predefined harmfulness criteria. The harmfulness criteria come from the security discriminator Beaver-Dam-7b [8], used to detect both the harmfulness and helpfulness of responses to" + } + ] + } + ], + "index": 19 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 313, + 674, + 554, + 693 + ], + "type": "page_footnote", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 674, + 554, + 693 + ], + "spans": [ + { + "bbox": [ + 313, + 674, + 554, + 693 + ], + "type": "text", + "content": "https://huggingface.co/datasets/liuhaotian/LLaVA-CC3M-Pretrain-595K" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 315, + 693, + 554, + 713 + ], + "type": "page_footnote", + "angle": 0, + "lines": [ + { + "bbox": [ + 315, + 693, + 554, + 713 + ], + "spans": [ + { + "bbox": [ + 315, + 693, + 554, + 713 + ], + "type": "text", + "content": "Since Gemini-1.0-Pro-Vision has been deprecated on July 12, 2024, we used its upgraded version, Gemini-1.5-Flash." + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 294, + 748, + 314, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 314, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 314, + 757 + ], + "type": "text", + "content": "9471" + } + ] + } + ], + "index": 23 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 4 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 104, + 70, + 506, + 337 + ], + "blocks": [ + { + "bbox": [ + 104, + 70, + 506, + 337 + ], + "lines": [ + { + "bbox": [ + 104, + 70, + 506, + 337 + ], + "spans": [ + { + "bbox": [ + 104, + 70, + 506, + 337 + ], + "type": "table", + "html": "
Victim ModelMethodAnimalFinancialPrivacySelf-HarmViolenceAverage (%)
Attack Success Rate (ASR) ↑
GPT-4o-miniHades4.7711.666.881.555.556.08
CS-DJ26.5580.0077.1134.7770.5557.80 (+51.72)
GPT-4oHades2.3310.117.551.885.665.51
CS-DJ15.3370.7758.5515.7750.7742.24 (+36.73)
GPT-4VHades8.1132.5523.888.0029.1120.33
CS-DJ11.6676.2259.7722.2257.3345.44 (+25.11)
Gemini-1.5-FlashHades2.6623.4413.111.775.009.20
CS-DJ22.5588.8879.1147.0083.0064.11 (+54.91)
Ensemble Attack Success Rate (EASR) ↑
GPT-4o-miniHades10.6621.3313.334.6610.0012.00
CS-DJ54.0092.6694.6652.0092.6677.20 (+65.20)
GPT-4oHades5.3316.0012.006.0011.3310.13
CS-DJ35.3392.6683.3337.3380.6665.86 (+55.73)
GPT-4VHades18.6663.3354.0024.6664.6645.06
CS-DJ34.6694.0089.3352.089.3371.86 (+26.80)
Gemini-1.5-FlashHades4.6634.0020.004.6610.6614.79
CS-DJ51.3397.3395.3366.0097.3381.46 (+66.67)
", + "image_path": "a542535a26e8d23817dc8422c77ffdda0fa6c0d2a2bb0b1a16a228ddb728d1b4.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "table_body" + } + ], + "index": 0 + }, + { + "bbox": [ + 102, + 345, + 506, + 355 + ], + "lines": [ + { + "bbox": [ + 102, + 345, + 506, + 355 + ], + "spans": [ + { + "bbox": [ + 102, + 345, + 506, + 355 + ], + "type": "text", + "content": "Table 1. ASR and EASR results of CS-DJ and Hades on four closed-source MLLMs across different categories." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "text" + }, + { + "bbox": [ + 55, + 372, + 295, + 443 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 372, + 295, + 443 + ], + "spans": [ + { + "bbox": [ + 55, + 372, + 295, + 443 + ], + "type": "text", + "content": "their corresponding queries. EASR [36] measures the effectiveness of jailbreak templates by calculating the proportion of queries where at least one template successfully bypasses the target MLLM's defenses. While ASR focuses on the success rate of individual templates, EASR measures the combined success rate of a group of templates." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 55, + 455, + 162, + 467 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 455, + 162, + 467 + ], + "spans": [ + { + "bbox": [ + 55, + 455, + 162, + 467 + ], + "type": "text", + "content": "4.2. Main Experiment" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 55, + 473, + 295, + 533 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 473, + 295, + 533 + ], + "spans": [ + { + "bbox": [ + 55, + 473, + 295, + 533 + ], + "type": "text", + "content": "To assess the effectiveness of CS-DJ, we conducted a comparison between CS-DJ and Hades [12], the state-of-the-art closed-source MLLM jailbreak attack. For a fair comparison, we adopted the official implementation of Hades and performed 6 groups of experiments to collect data." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 55, + 534, + 295, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 534, + 295, + 713 + ], + "spans": [ + { + "bbox": [ + 55, + 534, + 295, + 713 + ], + "type": "text", + "content": "Table 1 presents the jailbreak ASR and EASR results of CS-DJ and Hades across the four widely used closed-source MLLMs. The results across the five categories indicate that the four closed-source MLLMs exhibit stronger defenses against harmful queries in the Animal and Self-Harm categories while showing greater vulnerability in the Financial, Privacy, and Violence categories. Compared to their ASR results, Hades and CS-DJ achieve average improvements of " + }, + { + "bbox": [ + 55, + 534, + 295, + 713 + ], + "type": "inline_equation", + "content": "10.22\\%" + }, + { + "bbox": [ + 55, + 534, + 295, + 713 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 55, + 534, + 295, + 713 + ], + "type": "inline_equation", + "content": "21.70\\%" + }, + { + "bbox": [ + 55, + 534, + 295, + 713 + ], + "type": "text", + "content": " in EASR, respectively, indicating that increasing the number of trials can enhance the jailbreak success rate. Additionally, for both ASR and EASR, CS-DJ consistently outperforms Hades across all five categories for each evaluated MLLM. Overall, compared to Hades, CS-DJ achieves maximum improvements of " + }, + { + "bbox": [ + 55, + 534, + 295, + 713 + ], + "type": "inline_equation", + "content": "54.91\\%" + }, + { + "bbox": [ + 55, + 534, + 295, + 713 + ], + "type": "text", + "content": " in ASR and " + }, + { + "bbox": [ + 55, + 534, + 295, + 713 + ], + "type": "inline_equation", + "content": "66.67\\%" + }, + { + "bbox": [ + 55, + 534, + 295, + 713 + ], + "type": "text", + "content": " in EASR. Furthermore, CS-DJ attains" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 313, + 372, + 553, + 432 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 372, + 553, + 432 + ], + "spans": [ + { + "bbox": [ + 313, + 372, + 553, + 432 + ], + "type": "text", + "content": "average success rates of " + }, + { + "bbox": [ + 313, + 372, + 553, + 432 + ], + "type": "inline_equation", + "content": "52.40\\%" + }, + { + "bbox": [ + 313, + 372, + 553, + 432 + ], + "type": "text", + "content": " for ASR and " + }, + { + "bbox": [ + 313, + 372, + 553, + 432 + ], + "type": "inline_equation", + "content": "74.10\\%" + }, + { + "bbox": [ + 313, + 372, + 553, + 432 + ], + "type": "text", + "content": " for EASR. These observations highlight CS-DJ's superior effectiveness as a jailbreak method for bypassing the defenses of closed-source MLLMs. Jailbreaking examples can be found in the appendix." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 313, + 433, + 554, + 564 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 433, + 554, + 564 + ], + "spans": [ + { + "bbox": [ + 313, + 433, + 554, + 564 + ], + "type": "text", + "content": "A detailed analysis of Hades' results across the four models reveals that GPT-4V outperforms the other three models in both ASR and EASR. In contrast, the results of CS-DJ indicate that Gemini-1.5-Flash achieved the highest performance. These findings suggest that the impact of visual input harmfulness and distraction varies across different models. In a word, the results demonstrate that distracting the model's attention is a more effective strategy for enhancing jailbreak success rates. More quantitative and qualitative experimental results, including those tested on open-source models, are provided in the appendix." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 313, + 565, + 553, + 613 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 565, + 553, + 613 + ], + "spans": [ + { + "bbox": [ + 313, + 565, + 553, + 613 + ], + "type": "text", + "content": "For the subsequent experiments, we selected the most widely used model, GPT-40, for testing. All experiments were conducted as a single trial. Therefore, only ASR results are presented in the following sections." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 313, + 624, + 487, + 635 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 624, + 487, + 635 + ], + "spans": [ + { + "bbox": [ + 313, + 624, + 487, + 635 + ], + "type": "text", + "content": "4.3. Impact of Query Decomposition" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 313, + 642, + 553, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 642, + 553, + 713 + ], + "spans": [ + { + "bbox": [ + 313, + 642, + 553, + 713 + ], + "type": "text", + "content": "Here, we aim to explore the impact of query decomposition on MLLM jailbreak performance. To validate its effectiveness, we first use images derived from raw queries (RQ) as visual inputs for MLLMs, serving as a baseline for comparison with the results obtained using query decomposition. We further examine the effect of the number" + } + ] + } + ], + "index": 10 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 315, + 757 + ], + "type": "text", + "content": "9472" + } + ] + } + ], + "index": 11 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 5 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 58, + 70, + 295, + 135 + ], + "blocks": [ + { + "bbox": [ + 58, + 70, + 295, + 135 + ], + "lines": [ + { + "bbox": [ + 58, + 70, + 295, + 135 + ], + "spans": [ + { + "bbox": [ + 58, + 70, + 295, + 135 + ], + "type": "table", + "html": "
SettingAni.Fin.Priv.Self-H.Viol.ASR (%)
RQ0.676.005.3304.003.20
3SQ4.0028.6634.664.6622.0018.80
6SQ8.0045.3344.0010.6641.3329.86
9SQ4.6644.6644.667.3336.0027.46
", + "image_path": "5cf35014dca1eaef7e072a61051c60c23f9696b58dc2fb910e66512bf7d5378b.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "table_body" + } + ], + "index": 0 + }, + { + "bbox": [ + 55, + 143, + 295, + 210 + ], + "lines": [ + { + "bbox": [ + 55, + 143, + 295, + 210 + ], + "spans": [ + { + "bbox": [ + 55, + 143, + 295, + 210 + ], + "type": "text", + "content": "Table 2. GPT-4o's ASR results of CS-DJ under different query decomposition settings. RQ denotes using the image transformation of raw queries as the visual input for MLLMs. 3SQ, 6SQ, and 9SQ indicate that the raw query is decomposed into 3, 6, and 9 sub-queries, respectively, which are then transformed into sub-images to compose the final visual input." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "text" + }, + { + "bbox": [ + 54, + 230, + 295, + 350 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 54, + 230, + 295, + 350 + ], + "spans": [ + { + "bbox": [ + 54, + 230, + 295, + 350 + ], + "type": "text", + "content": "of sub-queries on the jailbreak success rate by decomposing the raw query into 3, 6, and 9 sub-queries, denoted as 3SQ, 6SQ, and 9SQ, respectively. Given that MLLMs typically resize image inputs to fixed dimensions, we aim to minimize distortion from aspect ratio changes during resizing and avoid potential limitations of the MLLM's image encoder. Therefore, we maintain a fixed column count of 3 when composing the final input image. As a result, the 3SQ, 6SQ, and 9SQ settings increase the number of rows while keeping the number of columns constant." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 56, + 353, + 295, + 674 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 353, + 295, + 674 + ], + "spans": [ + { + "bbox": [ + 56, + 353, + 295, + 674 + ], + "type": "text", + "content": "The quantitative results are presented in Table 2. The results from RQ demonstrate that the MLLM exhibits strong defenses against original jailbreak queries. Even when the raw query is transformed into an image input, it remains highly challenging to bypass the MLLM's security mechanisms. In contrast, 3SQ outperforms RQ across all five categories, yielding an overall improvement of " + }, + { + "bbox": [ + 56, + 353, + 295, + 674 + ], + "type": "inline_equation", + "content": "15.60\\%" + }, + { + "bbox": [ + 56, + 353, + 295, + 674 + ], + "type": "text", + "content": ". This indicates that query decomposition effectively enhances the jailbreak success rate for MLLMs. By introducing a distributional shift in the queries, query decomposition creates structural distraction, thereby enabling more effective bypassing of the model's defenses. Furthermore, increasing the number of decomposed sub-queries to 6 results in an additional " + }, + { + "bbox": [ + 56, + 353, + 295, + 674 + ], + "type": "inline_equation", + "content": "11.06\\%" + }, + { + "bbox": [ + 56, + 353, + 295, + 674 + ], + "type": "text", + "content": " improvement in the jailbreak success rate. This suggests that a higher number of sub-images transformed from sub-queries can further enhance the distraction effect on the MLLM. However, when the number of sub-queries is increased to 9, a slight decline in the success rate is observed. This decrease may be attributed to two factors: (1) the increased number of sub-queries raises task complexity, resulting in greater demands on the model's comprehension and its ability to handle complex tasks; (2) limitations in the model's image encoding capabilities may prevent accurate interpretation of the sub-queries; and (3) over-decomposition may cause the model to recognize the original intent of the query, thereby reducing the effectiveness of the jailbreak attempt." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 55, + 677, + 296, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 677, + 296, + 713 + ], + "spans": [ + { + "bbox": [ + 55, + 677, + 296, + 713 + ], + "type": "text", + "content": "Overall, the structural distraction introduced by query decomposition can enhance the jailbreak success rate of MLLMs. Furthermore, the number of sub-queries also" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 313, + 72, + 553, + 133 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 72, + 553, + 133 + ], + "spans": [ + { + "bbox": [ + 313, + 72, + 553, + 133 + ], + "type": "text", + "content": "plays a significant role in influencing the results. Given the limited capabilities of smaller MLLMs and the need to incorporate additional visual subimages, this work focuses on evaluating the performance of the CS-DJ with three subqueries." + } + ] + } + ], + "index": 5 + }, + { + "type": "image", + "bbox": [ + 322, + 167, + 545, + 287 + ], + "blocks": [ + { + "bbox": [ + 313, + 139, + 479, + 152 + ], + "lines": [ + { + "bbox": [ + 313, + 139, + 479, + 152 + ], + "spans": [ + { + "bbox": [ + 313, + 139, + 479, + 152 + ], + "type": "text", + "content": "4.4. Impact of Contrasting Visuals" + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 322, + 167, + 545, + 287 + ], + "lines": [ + { + "bbox": [ + 322, + 167, + 545, + 287 + ], + "spans": [ + { + "bbox": [ + 322, + 167, + 545, + 287 + ], + "type": "image", + "image_path": "09053fcdce5461237480ccb35f5b8989a65444f0a56c838898ee18ae3df8172d.jpg" + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 313, + 296, + 555, + 319 + ], + "lines": [ + { + "bbox": [ + 313, + 296, + 555, + 319 + ], + "spans": [ + { + "bbox": [ + 313, + 296, + 555, + 319 + ], + "type": "text", + "content": "Figure 3. GPT-4o's evaluation results for varying quantities of visual subimages." + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "image_caption" + } + ], + "index": 7 + }, + { + "bbox": [ + 313, + 331, + 555, + 570 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 331, + 555, + 570 + ], + "spans": [ + { + "bbox": [ + 313, + 331, + 555, + 570 + ], + "type": "text", + "content": "To validate the effectiveness of contrasting visuals, we conduct experiments from two perspectives: (1) the impact of distraction induced by varying the number of subimages, and (2) the effect of inter-subimage distraction with a fixed number of subimages. Since altering the number of subimages directly affects the layout of the final composite image, we maintain a fixed column count of 3 and adjust the number of rows to control the number of subimages. This strategy reduces distortion due to excessive changes in aspect ratio and mitigates potential limitations in the MLLM's image encoding capabilities. We conduct different experiments with various configurations, involving 0, 3, 6, 9, and 12 Contrasting visual SubImages, corresponding to 3SQ, 3SQ+3CSI, 3SQ+6CSI, 3SQ+9CSI, and 3SQ+12CSI, respectively. To explore the influence of inter-subimage distraction on jailbreak success rates, we conduct experiments using three settings: retrieving 9 most contrasting images, retrieving 9 most similar images, and using a single, most similar image from the dataset for all 9 subimages, denoting 3SQ+9CSI, 3SQ+9SSI, and 3SQ+9SinSI, respectively." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 313, + 570, + 556, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 570, + 556, + 713 + ], + "spans": [ + { + "bbox": [ + 313, + 570, + 556, + 713 + ], + "type": "text", + "content": "The experimental results on the impact of visual subimage quantity are illustrated in Figure 3. As 3SQ does not contain any visual subimages, its Distraction Distance is 0, resulting in the lowest ASR. Increasing the number of visual subimages leads to a rise in ASR. Specifically, compared to 3SQ, " + }, + { + "bbox": [ + 313, + 570, + 556, + 713 + ], + "type": "inline_equation", + "content": "3\\mathrm{SQ} + 3\\mathrm{CSI}" + }, + { + "bbox": [ + 313, + 570, + 556, + 713 + ], + "type": "text", + "content": " obtains a " + }, + { + "bbox": [ + 313, + 570, + 556, + 713 + ], + "type": "inline_equation", + "content": "12.8\\%" + }, + { + "bbox": [ + 313, + 570, + 556, + 713 + ], + "type": "text", + "content": " increase in ASR, demonstrating that contrasting visual subimages can effectively distract the MLLM and bypass its internal defense mechanisms. As the number of visual subimages increases, both the Distraction Distance and the ASR grow correspondingly. However, the rate of ASR improvement diminishes as the number of subimages continues to rise. When the" + } + ] + } + ], + "index": 10 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 315, + 757 + ], + "type": "text", + "content": "9473" + } + ] + } + ], + "index": 11 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 6 + }, + { + "para_blocks": [ + { + "bbox": [ + 55, + 72, + 294, + 120 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 72, + 294, + 120 + ], + "spans": [ + { + "bbox": [ + 55, + 72, + 294, + 120 + ], + "type": "text", + "content": "number of subimages reaches 9, further additions yield limited gains in ASR. Given the limited encoding capabilities of the visual encoders in smaller MLLMs, we use 9 visual subimages as the default setting for our experiments." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 56, + 121, + 295, + 370 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 121, + 295, + 370 + ], + "spans": [ + { + "bbox": [ + 56, + 121, + 295, + 370 + ], + "type": "text", + "content": "Table 3 presents the experimental results on the impact of inter-subimage distraction. We evaluate three different subimage selection strategies. The results show that " + }, + { + "bbox": [ + 56, + 121, + 295, + 370 + ], + "type": "inline_equation", + "content": "3\\mathrm{SQ} + 9\\mathrm{SinSI}" + }, + { + "bbox": [ + 56, + 121, + 295, + 370 + ], + "type": "text", + "content": " has the lowest distraction distance, corresponding to the lowest ASR. In contrast, " + }, + { + "bbox": [ + 56, + 121, + 295, + 370 + ], + "type": "inline_equation", + "content": "3\\mathrm{SQ} + 9\\mathrm{CSI}" + }, + { + "bbox": [ + 56, + 121, + 295, + 370 + ], + "type": "text", + "content": " achieves an " + }, + { + "bbox": [ + 56, + 121, + 295, + 370 + ], + "type": "inline_equation", + "content": "11.20\\%" + }, + { + "bbox": [ + 56, + 121, + 295, + 370 + ], + "type": "text", + "content": " improvement in ASR compared to " + }, + { + "bbox": [ + 56, + 121, + 295, + 370 + ], + "type": "inline_equation", + "content": "3\\mathrm{SQ} + 9\\mathrm{SSI}" + }, + { + "bbox": [ + 56, + 121, + 295, + 370 + ], + "type": "text", + "content": ", along with a further increase in distraction distance. These findings indicate that, for a fixed number of subimages, greater distraction among subimages leads to a higher jailbreak success rate for MLLMs. Additionally, an analysis of the results from Figure 3 and Table 3 reveals that distraction distance accurately reflects changes in ASR trends only when varying the subimage construction strategy along a single dimension, such as the number of subimages or subimage selection strategies. This limitation may stem from differences between the semantic space of our current feature extractor, CLIP, and that of the MLLM. As a result, inter-subimage distraction distance cannot always be accurately measured using the CLIP model. Future work will focus on developing a more comprehensive metric for measuring distraction distance." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 55, + 371, + 294, + 453 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 371, + 294, + 453 + ], + "spans": [ + { + "bbox": [ + 55, + 371, + 294, + 453 + ], + "type": "text", + "content": "Table 4 shows the impact of information complexity of visual subimages. The ASR of " + }, + { + "bbox": [ + 55, + 371, + 294, + 453 + ], + "type": "inline_equation", + "content": "3\\mathrm{SQ} + 9\\mathrm{RNI}" + }, + { + "bbox": [ + 55, + 371, + 294, + 453 + ], + "type": "text", + "content": " is close to that of 3SQ, but significantly lower than " + }, + { + "bbox": [ + 55, + 371, + 294, + 453 + ], + "type": "inline_equation", + "content": "3\\mathrm{SQ} + 9\\mathrm{CSI}" + }, + { + "bbox": [ + 55, + 371, + 294, + 453 + ], + "type": "text", + "content": ", indicating that MLLMs exhibit strong recognition capabilities for noise images with minimal informational content. Therefore, only subimages with higher information complexity can effectively distract the model." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 55, + 455, + 294, + 491 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 455, + 294, + 491 + ], + "spans": [ + { + "bbox": [ + 55, + 455, + 294, + 491 + ], + "type": "text", + "content": "In summary, constructing multiple contrasting visual subimages effectively enhances the distraction of MLLMs, leading to improved jailbreak success rates." + } + ] + } + ], + "index": 3 + }, + { + "type": "table", + "bbox": [ + 77, + 500, + 272, + 555 + ], + "blocks": [ + { + "bbox": [ + 77, + 500, + 272, + 555 + ], + "lines": [ + { + "bbox": [ + 77, + 500, + 272, + 555 + ], + "spans": [ + { + "bbox": [ + 77, + 500, + 272, + 555 + ], + "type": "table", + "html": "
SettingASR (%)Distraction Distance
3SQ+9SinSI24.66399.08
3SQ+9SSI32.531378.77
3SQ+9CSI43.731878.83
", + "image_path": "47b246fbd831a616c2bc952c82d26fe480a7a890571209bf25a4034db80ba354.jpg" + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "table_body" + } + ], + "index": 4 + }, + { + "bbox": [ + 55, + 636, + 178, + 649 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 636, + 178, + 649 + ], + "spans": [ + { + "bbox": [ + 55, + 636, + 178, + 649 + ], + "type": "text", + "content": "4.5. Impact of Instruction" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 55, + 654, + 294, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 654, + 294, + 713 + ], + "spans": [ + { + "bbox": [ + 55, + 654, + 294, + 713 + ], + "type": "text", + "content": "To complement the constructed multi-subimage input, we carefully designed the instruction " + }, + { + "bbox": [ + 55, + 654, + 294, + 713 + ], + "type": "inline_equation", + "content": "P" + }, + { + "bbox": [ + 55, + 654, + 294, + 713 + ], + "type": "text", + "content": ". To evaluate the impact of different components of " + }, + { + "bbox": [ + 55, + 654, + 294, + 713 + ], + "type": "inline_equation", + "content": "P" + }, + { + "bbox": [ + 55, + 654, + 294, + 713 + ], + "type": "text", + "content": ", we conducted three experiments: (1) using only the task-guiding component, (2) adding the role-guiding component, and (3) further incorpo" + } + ] + } + ], + "index": 7 + }, + { + "type": "table", + "bbox": [ + 317, + 71, + 553, + 121 + ], + "blocks": [ + { + "bbox": [ + 55, + 561, + 294, + 617 + ], + "lines": [ + { + "bbox": [ + 55, + 561, + 294, + 617 + ], + "spans": [ + { + "bbox": [ + 55, + 561, + 294, + 617 + ], + "type": "text", + "content": "Table 3. ASR of GPT-4o and Distraction Distance for CS-DJ under different visual subimage selection strategies. 9SinSI indicates using a single, most similar image from the dataset for all 9 subimages. 9CSI and 9SSI denote retrieving 9 most contrasting and similar images to compose the final visual input, respectively." + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 317, + 71, + 553, + 121 + ], + "lines": [ + { + "bbox": [ + 317, + 71, + 553, + 121 + ], + "spans": [ + { + "bbox": [ + 317, + 71, + 553, + 121 + ], + "type": "table", + "html": "
SettingAni.Fin.Priv.Self-H.Viol.ASR (%)
3SQ4.0028.6634.664.6622.0018.80
3SQ+9RNI5.3332.6633.334.6624.0020.00
3SQ+9CSI19.3373.3356.6618.0051.3343.73
", + "image_path": "b0be4a81e16492acc68118121923ce469c635abb5d982aaf028a918be9024e19.jpg" + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "table_body" + } + ], + "index": 8 + }, + { + "type": "table", + "bbox": [ + 317, + 173, + 553, + 220 + ], + "blocks": [ + { + "bbox": [ + 313, + 127, + 553, + 160 + ], + "lines": [ + { + "bbox": [ + 313, + 127, + 553, + 160 + ], + "spans": [ + { + "bbox": [ + 313, + 127, + 553, + 160 + ], + "type": "text", + "content": "Table 4. GPT-4o's ASR results of CS-DJ under different settings. 9RNI denotes using 9 random noise images to compose the final visual input." + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 317, + 173, + 553, + 220 + ], + "lines": [ + { + "bbox": [ + 317, + 173, + 553, + 220 + ], + "spans": [ + { + "bbox": [ + 317, + 173, + 553, + 220 + ], + "type": "table", + "html": "
SettingAni.Fin.Priv.Self-H.Viol.ASR (%)
task-guiding10.0055.3339.3313.3342.6632.00
+ role-guiding15.3367.3350.0013.3352.0039.73
+ visual-guiding19.3373.3356.6618.0051.3343.73
", + "image_path": "e4fe6f2396a2bd35fd7170e5d44a3d092987133ae9caf1e940585bf3c47dbff1.jpg" + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "table_body" + } + ], + "index": 10 + }, + { + "bbox": [ + 313, + 226, + 553, + 249 + ], + "lines": [ + { + "bbox": [ + 313, + 226, + 553, + 249 + ], + "spans": [ + { + "bbox": [ + 313, + 226, + 553, + 249 + ], + "type": "text", + "content": "Table 5. GPT-4o's ASR results of CS-DJ under the 3SQ+9CSI configuration with different instruction " + }, + { + "bbox": [ + 313, + 226, + 553, + 249 + ], + "type": "inline_equation", + "content": "P" + }, + { + "bbox": [ + 313, + 226, + 553, + 249 + ], + "type": "text", + "content": " settings." + } + ] + } + ], + "index": 11, + "angle": 0, + "type": "text" + }, + { + "bbox": [ + 313, + 260, + 555, + 416 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 260, + 555, + 416 + ], + "spans": [ + { + "bbox": [ + 313, + 260, + 555, + 416 + ], + "type": "text", + "content": "rating the visual-guiding component. The results, presented in Table 5, show that using only the task-guiding component achieved the lowest ASR but still outperformed Hades by " + }, + { + "bbox": [ + 313, + 260, + 555, + 416 + ], + "type": "inline_equation", + "content": "26.49\\%" + }, + { + "bbox": [ + 313, + 260, + 555, + 416 + ], + "type": "text", + "content": ". This indicates that the visual distraction from the multi-subimage input, combined with the task distraction of handling multiple tasks simultaneously, significantly contributes to improving jailbreak success rates. Adding the role-guiding component further increased the ASR by " + }, + { + "bbox": [ + 313, + 260, + 555, + 416 + ], + "type": "inline_equation", + "content": "7.73\\%" + }, + { + "bbox": [ + 313, + 260, + 555, + 416 + ], + "type": "text", + "content": ", while incorporating the visual-guiding component led to an additional " + }, + { + "bbox": [ + 313, + 260, + 555, + 416 + ], + "type": "inline_equation", + "content": "4.00\\%" + }, + { + "bbox": [ + 313, + 260, + 555, + 416 + ], + "type": "text", + "content": " improvement. These findings demonstrate that targeted optimization of " + }, + { + "bbox": [ + 313, + 260, + 555, + 416 + ], + "type": "inline_equation", + "content": "P" + }, + { + "bbox": [ + 313, + 260, + 555, + 416 + ], + "type": "text", + "content": " for multisubimage visual inputs can further disperse the model's attention, enhancing the success rate of the jailbreak." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 313, + 429, + 388, + 441 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 429, + 388, + 441 + ], + "spans": [ + { + "bbox": [ + 313, + 429, + 388, + 441 + ], + "type": "text", + "content": "5. Conclusion" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 313, + 449, + 555, + 616 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 449, + 555, + 616 + ], + "spans": [ + { + "bbox": [ + 313, + 449, + 555, + 616 + ], + "type": "text", + "content": "In this paper, we investigate the visual vulnerabilities in jailbreak attacks on MLLMs. Our findings show that the complexity of visual inputs significantly influences attack success. We introduce the Distraction Hypothesis and propose the CS-DJ framework, which leverages distraction-based strategies to exploit these vulnerabilities in MLLMs. Our approach integrates this insight by two core components: structured distraction through query decomposition and visual-enhanced distraction using contrasting subimages. Extensive experiments conducted on five representative scenarios and four popular closed-source MLLMs demonstrated that CS-DJ outperforms state-of-the-art jailbreak methods, validating the efficacy of distraction-based approaches in bypassing MLLM defenses." + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 313, + 617, + 556, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 617, + 556, + 713 + ], + "spans": [ + { + "bbox": [ + 313, + 617, + 556, + 713 + ], + "type": "text", + "content": "Acknowledgment. This research is supported by Guangzhou Basic and Applied Basic Research Project (No: 2025A04J3536, 2025A03J3128, 2024A03J0324), Guangzhou University Research Project (No: RQ2021013), Guangdong Basic and Applied Basic Research Foundation (No: 2023A1515110077), NSFC (No: 62402131), and the Postdoctoral Fellowship Program of CPSF (No: GZC20230595)." + } + ] + } + ], + "index": 15 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 315, + 757 + ], + "type": "text", + "content": "9474" + } + ] + } + ], + "index": 16 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 7 + }, + { + "para_blocks": [ + { + "bbox": [ + 56, + 71, + 115, + 83 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 71, + 115, + 83 + ], + "spans": [ + { + "bbox": [ + 56, + 71, + 115, + 83 + ], + "type": "text", + "content": "References" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 57, + 91, + 296, + 713 + ], + "type": "list", + "angle": 0, + "index": 12, + "blocks": [ + { + "bbox": [ + 61, + 91, + 296, + 145 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 61, + 91, + 296, + 145 + ], + "spans": [ + { + "bbox": [ + 61, + 91, + 296, + 145 + ], + "type": "text", + "content": "[1] Josh Achiam, Steven Adler, Sandhini Agarwal, Lama Ahmad, Ilge Akkaya, Florencia Leoni Aleman, Diogo Almeida, Janko Altenschmidt, Sam Altman, Shyamal Anadkat, et al. Gpt-4 technical report. arXiv preprint arXiv:2303.08774, 2023.5" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 61, + 148, + 296, + 201 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 61, + 148, + 296, + 201 + ], + "spans": [ + { + "bbox": [ + 61, + 148, + 296, + 201 + ], + "type": "text", + "content": "[2] Yulin Chen, Haoran Li, Zihao Zheng, and Yangqiu Song. Bathe: Defense against the jailbreak attack in multimodal large language models by treating harmful instruction as backdoor trigger. arXiv preprint arXiv:2408.09093, 2024. 2" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 61, + 205, + 296, + 270 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 61, + 205, + 296, + 270 + ], + "spans": [ + { + "bbox": [ + 61, + 205, + 296, + 270 + ], + "type": "text", + "content": "[3] Can Cui, Yunsheng Ma, Xu Cao, Wenqian Ye, Yang Zhou, Kaizhao Liang, Jintai Chen, Juanwu Lu, Zichong Yang, Kuei-Da Liao, et al. A survey on multimodal large language models for autonomous driving. In Proceedings of the IEEE/CVF Winter Conference on Applications of Computer Vision, pages 958-979, 2024. 1" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 61, + 272, + 295, + 327 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 61, + 272, + 295, + 327 + ], + "spans": [ + { + "bbox": [ + 61, + 272, + 295, + 327 + ], + "type": "text", + "content": "[4] Ameet Deshpande, Vishvak Murahari, Tanmay Rajpurohit, Ashwin Kalyan, and Karthik Narasimhan. Toxicity in chatgpt: Analyzing persona-assigned language models. In Findings of the Association for Computational Linguistics: EMNLP 2023, pages 1236-1270, 2023. 2" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 61, + 329, + 296, + 405 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 61, + 329, + 296, + 405 + ], + "spans": [ + { + "bbox": [ + 61, + 329, + 296, + 405 + ], + "type": "text", + "content": "[5] Yassir Fathullah, Chunyang Wu, Egor Lakomkin, Junteng Jia, Yuan Shangguan, Ke Li, Jinxi Guo, Wenhan Xiong, Jay Mahadeokar, Ozlem Kalinli, et al. Prompting large language models with speech recognition abilities. In ICASSP 2024-2024 IEEE International Conference on Acoustics, Speech and Signal Processing (ICASSP), pages 13351-13355. IEEE, 2024. 3" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 61, + 407, + 296, + 461 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 61, + 407, + 296, + 461 + ], + "spans": [ + { + "bbox": [ + 61, + 407, + 296, + 461 + ], + "type": "text", + "content": "[6] Yichen Gong, Delong Ran, Jinyuan Liu, Conglei Wang, Tianshuo Cong, Anyu Wang, Sisi Duan, and Xiaoyun Wang. Figstep: Jailbreaking large vision-language models via typographic visual prompts. arXiv preprint arXiv:2311.05608, 2023. 2, 3, 5, 12" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 62, + 464, + 296, + 508 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 464, + 296, + 508 + ], + "spans": [ + { + "bbox": [ + 62, + 464, + 296, + 508 + ], + "type": "text", + "content": "[7] Aaron Hurst, Adam Lerer, Adam P Goucher, Adam Perelman, Aditya Ramesh, Aidan Clark, AJ Ostrow, Akila Weli-hinda, Alan Hayes, Alec Radford, et al. Gpt-4o system card. arXiv preprint arXiv:2410.21276, 2024. 3" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 62, + 510, + 296, + 564 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 510, + 296, + 564 + ], + "spans": [ + { + "bbox": [ + 62, + 510, + 296, + 564 + ], + "type": "text", + "content": "[8] Jiaming Ji, Mickel Liu, Josef Dai, Xuehai Pan, Chi Zhang, Ce Bian, Boyuan Chen, Ruiyang Sun, Yizhou Wang, and Yaodong Yang. Beavertails: Towards improved safety alignment of llm via a human-preference dataset. Advances in Neural Information Processing Systems, 36, 2024. 5, 12" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 62, + 567, + 296, + 620 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 567, + 296, + 620 + ], + "spans": [ + { + "bbox": [ + 62, + 567, + 296, + 620 + ], + "type": "text", + "content": "[9] Haibo Jin, Leyang Hu, Xinuo Li, Peiyan Zhang, Chonghan Chen, Jun Zhuang, and Haohan Wang. Jailbreakzoo: Survey, landscapes, and horizons in jailbreaking large language and vision-language models. arXiv preprint arXiv:2407.01599, 2024.3" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 57, + 624, + 295, + 667 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 624, + 295, + 667 + ], + "spans": [ + { + "bbox": [ + 57, + 624, + 295, + 667 + ], + "type": "text", + "content": "[10] Daniel Martin Katz, Michael James Bommarito, Shang Gao, and Pablo Arredondo. Gpt-4 passes the bar exam. Philosophical Transactions of the Royal Society A, 382(2270): 20230254, 2024. 3" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 57, + 670, + 295, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 670, + 295, + 713 + ], + "spans": [ + { + "bbox": [ + 57, + 670, + 295, + 713 + ], + "type": "text", + "content": "[11] Bo Li, Yuanhan Zhang, Dong Guo, Renrui Zhang, Feng Li, Hao Zhang, Kaichen Zhang, Peiyuan Zhang, Yanwei Li, Ziwei Liu, et al. Llava-onevision: Easy visual task transfer. arXiv preprint arXiv:2408.03326, 2024. 12" + } + ] + } + ], + "index": 11 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 57, + 73, + 553, + 713 + ], + "type": "list", + "angle": 0, + "index": 26, + "blocks": [ + { + "bbox": [ + 316, + 73, + 553, + 127 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 73, + 553, + 127 + ], + "spans": [ + { + "bbox": [ + 316, + 73, + 553, + 127 + ], + "type": "text", + "content": "[12] Yifan Li, Hangyu Guo, Kun Zhou, Wayne Xin Zhao, and JiRong Wen. Images are achilles' heel of alignment: Exploiting visual vulnerabilities for jailbreaking multimodal large language models. In European Conference on Computer Vision. Springer, 2024. 1, 2, 3, 5, 6" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 317, + 129, + 553, + 172 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 129, + 553, + 172 + ], + "spans": [ + { + "bbox": [ + 317, + 129, + 553, + 172 + ], + "type": "text", + "content": "[13] Haotian Liu, Chunyuan Li, Yuheng Li, and Yong Jae Lee. Improved baselines with visual instruction tuning. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 26296-26306, 2024. 3" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 316, + 173, + 553, + 205 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 173, + 553, + 205 + ], + "spans": [ + { + "bbox": [ + 316, + 173, + 553, + 205 + ], + "type": "text", + "content": "[14] Haotian Liu, Chunyuan Li, Qingyang Wu, and Yong Jae Lee. Visual instruction tuning. Advances in neural information processing systems, 36, 2024. 3" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 316, + 206, + 553, + 260 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 206, + 553, + 260 + ], + "spans": [ + { + "bbox": [ + 316, + 206, + 553, + 260 + ], + "type": "text", + "content": "[15] Tong Liu, Yingjie Zhang, Zhe Zhao, Yinpeng Dong, Guozhu Meng, and Kai Chen. Making them ask and answer: Jailbreaking large language models in few queries via disguise and reconstruction. In 33rd USENIX Security Symposium (USENIX Security 24), pages 4711-4728, 2024. 2" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 316, + 261, + 553, + 293 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 261, + 553, + 293 + ], + "spans": [ + { + "bbox": [ + 316, + 261, + 553, + 293 + ], + "type": "text", + "content": "[16] Xin Liu, Yichen Zhu, Yunshi Lan, Chao Yang, and Yu Qiao. Query-relevant images jailbreak large multi-modal models. arXiv preprint arXiv:2311.17600, 2023. 2, 3" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 316, + 294, + 553, + 348 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 294, + 553, + 348 + ], + "spans": [ + { + "bbox": [ + 316, + 294, + 553, + 348 + ], + "type": "text", + "content": "[17] Xin Liu, Yichen Zhu, Jindong Gu, Yunshi Lan, Chao Yang, and Yu Qiao. Mm-safetybench: A benchmark for safety evaluation of multimodal large language models. In European Conference on Computer Vision, pages 386-403. Springer, 2024. 2, 3, 12" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 316, + 350, + 553, + 414 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 350, + 553, + 414 + ], + "spans": [ + { + "bbox": [ + 316, + 350, + 553, + 414 + ], + "type": "text", + "content": "[18] Dong Lu, Zhiqiang Wang, Teng Wang, Weili Guan, Hongchang Gao, and Feng Zheng. Set-level guidance attack: Boosting adversarial transferability of vision-language pre-training models. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 102-111, 2023. 2" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 316, + 415, + 553, + 469 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 415, + 553, + 469 + ], + "spans": [ + { + "bbox": [ + 316, + 415, + 553, + 469 + ], + "type": "text", + "content": "[19] Siyuan Ma, Weidi Luo, Yu Wang, Xiaogeng Liu, Muhao Chen, Bo Li, and Chaowei Xiao. Visual-roleplay: Universal jailbreak attack on multimodal large language models via role-playing image character. arXiv preprint arXiv:2405.20773, 2024. 2" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 316, + 471, + 553, + 504 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 471, + 553, + 504 + ], + "spans": [ + { + "bbox": [ + 316, + 471, + 553, + 504 + ], + "type": "text", + "content": "[20] Zhenxing Niu, Haodong Ren, Xinbo Gao, Gang Hua, and Rong Jin. Jailbreaking attack against multimodal large language model. arXiv preprint arXiv:2402.02309, 2024. 3" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 316, + 505, + 553, + 568 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 505, + 553, + 568 + ], + "spans": [ + { + "bbox": [ + 316, + 505, + 553, + 568 + ], + "type": "text", + "content": "[21] Long Ouyang, Jeffrey Wu, Xu Jiang, Diogo Almeida, Carroll Wainwright, Pamela Mishkin, Chong Zhang, Sandhini Agarwal, Katarina Slama, Alex Ray, et al. Training language models to follow instructions with human feedback. Advances in neural information processing systems, 35:27730-27744, 2022. 2" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 316, + 570, + 553, + 624 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 570, + 553, + 624 + ], + "spans": [ + { + "bbox": [ + 316, + 570, + 553, + 624 + ], + "type": "text", + "content": "[22] Xiangyu Qi, Kaixuan Huang, Ashwinee Panda, Peter Henderson, Mengdi Wang, and Prateek Mittal. Visual adversarial examples jailbreak aligned large language models. In Proceedings of the AAAI Conference on Artificial Intelligence, pages 21527-21536, 2024. 2, 3" + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 316, + 625, + 553, + 689 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 625, + 553, + 689 + ], + "spans": [ + { + "bbox": [ + 316, + 625, + 553, + 689 + ], + "type": "text", + "content": "[23] Alec Radford, Jong Wook Kim, Chris Hallacy, Aditya Ramesh, Gabriel Goh, Sandhini Agarwal, Girish Sastry, Amanda Askell, Pamela Mishkin, Jack Clark, et al. Learning transferable visual models from natural language supervision. In International conference on machine learning, pages 8748-8763. PMLR, 2021. 4" + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 316, + 691, + 553, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 691, + 553, + 713 + ], + "spans": [ + { + "bbox": [ + 316, + 691, + 553, + 713 + ], + "type": "text", + "content": "[24] Erfan Shayegani, Yue Dong, and Nael Abu-Ghazaleh. Jailbreak in pieces: Compositional adversarial attacks on multi-" + } + ] + } + ], + "index": 25 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "text", + "content": "9475" + } + ] + } + ], + "index": 27 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 8 + }, + { + "para_blocks": [ + { + "bbox": [ + 56, + 73, + 296, + 713 + ], + "type": "list", + "angle": 0, + "index": 14, + "blocks": [ + { + "bbox": [ + 75, + 73, + 294, + 95 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 75, + 73, + 294, + 95 + ], + "spans": [ + { + "bbox": [ + 75, + 73, + 294, + 95 + ], + "type": "text", + "content": "modal language models. In The Twelfth International Conference on Learning Representations, 2023. 3" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 56, + 95, + 296, + 149 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 95, + 296, + 149 + ], + "spans": [ + { + "bbox": [ + 56, + 95, + 296, + 149 + ], + "type": "text", + "content": "[25] Wentan Tan, Changxing Ding, Jiayu Jiang, Fei Wang, Yibing Zhan, and Dapeng Tao. Harnessing the power of mllms for transferable text-to-image person reid. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 17127-17137, 2024. 2" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 56, + 150, + 294, + 183 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 150, + 294, + 183 + ], + "spans": [ + { + "bbox": [ + 56, + 150, + 294, + 183 + ], + "type": "text", + "content": "[26] Xijia Tao, Shuai Zhong, Lei Li, Qi Liu, and Lingpeng Kong. Imgtrojan: Jailbreaking vision-language models with one image. arXiv preprint arXiv:2403.02910, 2024. 2" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 56, + 184, + 294, + 238 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 184, + 294, + 238 + ], + "spans": [ + { + "bbox": [ + 56, + 184, + 294, + 238 + ], + "type": "text", + "content": "[27] Gemini Team, Petko Georgiev, Ving Ian Lei, Ryan Burnell, Libin Bai, Anmol Gulati, Garrett Tanzer, Damien Vincent, Zhufeng Pan, Shibo Wang, et al. Gemini 1.5: Unlocking multimodal understanding across millions of tokens of context. arXiv preprint arXiv:2403.05530, 2024. 3, 5" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 56, + 239, + 294, + 259 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 239, + 294, + 259 + ], + "spans": [ + { + "bbox": [ + 56, + 239, + 294, + 259 + ], + "type": "text", + "content": "[28] Qwen Team. Qwen2.5: A party of foundation models, 2024. 3" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 56, + 261, + 294, + 316 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 261, + 294, + 316 + ], + "spans": [ + { + "bbox": [ + 56, + 261, + 294, + 316 + ], + "type": "text", + "content": "[29] Haoqin Tu, Chenhang Cui, Zijun Wang, Yiyang Zhou, Bingchen Zhao, Junlin Han, Wangchunshu Zhou, Huaxiu Yao, and Cihang Xie. How many unicorns are in this image? a safety evaluation benchmark for vision llms. arXiv preprint arXiv:2311.16101, 2023. 2" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 56, + 317, + 294, + 371 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 317, + 294, + 371 + ], + "spans": [ + { + "bbox": [ + 56, + 317, + 294, + 371 + ], + "type": "text", + "content": "[30] Haodi Wang, Kai Dong, Zhilei Zhu, Haotong Qin, Aishan Liu, Xiaolin Fang, Jiakai Wang, and Xianglong Liu. Transferable multimodal attack on vision-language pre-training models. In 2024 IEEE Symposium on Security and Privacy (SP), pages 102-102. IEEE Computer Society, 2024. 2" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 56, + 372, + 294, + 426 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 372, + 294, + 426 + ], + "spans": [ + { + "bbox": [ + 56, + 372, + 294, + 426 + ], + "type": "text", + "content": "[31] Jiaqi Wang, Hanqi Jiang, Yiheng Liu, Chong Ma, Xu Zhang, Yi Pan, Mengyuan Liu, Peiran Gu, Sichen Xia, Wenjun Li, et al. A comprehensive review of multimodal large language models: Performance and challenges across different tasks. arXiv preprint arXiv:2408.01319, 2024. 3" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 56, + 426, + 294, + 481 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 426, + 294, + 481 + ], + "spans": [ + { + "bbox": [ + 56, + 426, + 294, + 481 + ], + "type": "text", + "content": "[32] Ruofan Wang, Xingjun Ma, Hanxu Zhou, Chuanjun Ji, Guangnan Ye, and Yu-Gang Jiang. White-box multimodal jailbreaks against large vision-language models. In Proceedings of the 32nd ACM International Conference on Multimedia, pages 6920-6928, 2024. 2" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 56, + 482, + 294, + 536 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 482, + 294, + 536 + ], + "spans": [ + { + "bbox": [ + 56, + 482, + 294, + 536 + ], + "type": "text", + "content": "[33] Yizhong Wang, Yeganeh Kordi, Swaroop Mishra, Alisa Liu, Noah A Smith, Daniel Khashabi, and Hannaneh Hajishirzi. Self-instruct: Aligning language models with self-generated instructions. In The 61st Annual Meeting Of The Association For Computational Linguistics, 2023. 2" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 56, + 537, + 294, + 569 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 537, + 294, + 569 + ], + "spans": [ + { + "bbox": [ + 56, + 537, + 294, + 569 + ], + "type": "text", + "content": "[34] Alexander Wei, Nika Haghtalab, and Jacob Steinhardt. Jailbroken: How does llm safety training fail? Advances in Neural Information Processing Systems, 36, 2024. 2" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 56, + 571, + 294, + 603 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 571, + 294, + 603 + ], + "spans": [ + { + "bbox": [ + 56, + 571, + 294, + 603 + ], + "type": "text", + "content": "[35] Yuanwei Wu, Xiang Li, Yixin Liu, Pan Zhou, and Lichao Sun. Jailbreaking gpt-4v via self-adversarial attacks with system prompts. arXiv preprint arXiv:2311.09127, 2023. 2" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 56, + 604, + 294, + 647 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 604, + 294, + 647 + ], + "spans": [ + { + "bbox": [ + 56, + 604, + 294, + 647 + ], + "type": "text", + "content": "[36] Jiahao Yu, Xingwei Lin, Zheng Yu, and Xinyu Xing. {LLM-Fuzzer}: Scaling assessment of large language model jailbreaks. In 33rd USENIX Security Symposium (USENIX Security 24), pages 4657-4674, 2024. 2, 6" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 56, + 647, + 294, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 647, + 294, + 713 + ], + "spans": [ + { + "bbox": [ + 56, + 647, + 294, + 713 + ], + "type": "text", + "content": "[37] Qifan Yu, Juncheng Li, Longhui Wei, Liang Pang, Wentao Ye, Bosheng Qin, Siliang Tang, Qi Tian, and Yueting Zhuang. Hallucidobot: Mitigating hallucinatory toxicity in visual instruction data. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 12944-12953, 2024. 2" + } + ] + } + ], + "index": 13 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 316, + 73, + 553, + 284 + ], + "type": "list", + "angle": 0, + "index": 19, + "blocks": [ + { + "bbox": [ + 316, + 73, + 553, + 139 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 73, + 553, + 139 + ], + "spans": [ + { + "bbox": [ + 316, + 73, + 553, + 139 + ], + "type": "text", + "content": "[38] Hangfan Zhang, Zhimeng Guo, Huaisheng Zhu, Bochuan Cao, Lu Lin, Jinyuan Jia, Jinghui Chen, and Dinghao Wu. Jailbreak open-sourced large language models via enforced decoding. In Proceedings of the 62nd Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers), pages 5475-5493, 2024. 2" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 316, + 140, + 553, + 182 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 140, + 553, + 182 + ], + "spans": [ + { + "bbox": [ + 316, + 140, + 553, + 182 + ], + "type": "text", + "content": "[39] Jingyi Zhang, Jiaxing Huang, Sheng Jin, and Shijian Lu. Vision-language models for vision tasks: A survey. IEEE Transactions on Pattern Analysis and Machine Intelligence, 2024. 3" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 316, + 185, + 553, + 238 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 185, + 553, + 238 + ], + "spans": [ + { + "bbox": [ + 316, + 185, + 553, + 238 + ], + "type": "text", + "content": "[40] Yunqing Zhao, Tianyu Pang, Chao Du, Xiao Yang, Chongxuan Li, Ngai-Man Man Cheung, and Min Lin. On evaluating adversarial robustness of large vision-language models. Advances in Neural Information Processing Systems, 36, 2024. 2" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 316, + 240, + 553, + 284 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 240, + 553, + 284 + ], + "spans": [ + { + "bbox": [ + 316, + 240, + 553, + 284 + ], + "type": "text", + "content": "[41] Deyao Zhu, Jun Chen, Xiaoqian Shen, Xiang Li, and Mohamed Elhoseiny. Minigpt-4: Enhancing vision-language understanding with advanced large language models. arXiv preprint arXiv:2304.10592, 2023. 3" + } + ] + } + ], + "index": 18 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 748, + 316, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 316, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 316, + 757 + ], + "type": "text", + "content": "9476" + } + ] + } + ], + "index": 20 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 9 + } + ], + "_backend": "vlm", + "_version_name": "2.6.4" +} \ No newline at end of file diff --git a/2025/Distribution Prototype Diffusion Learning for Open-set Supervised Anomaly Detection/5f6492f6-61fa-40e8-b92f-42fc55a187bf_content_list.json b/2025/Distribution Prototype Diffusion Learning for Open-set Supervised Anomaly Detection/5f6492f6-61fa-40e8-b92f-42fc55a187bf_content_list.json new file mode 100644 index 0000000000000000000000000000000000000000..05f40da3d1ebdbab6da4a643193bda857ec68c1f --- /dev/null +++ b/2025/Distribution Prototype Diffusion Learning for Open-set Supervised Anomaly Detection/5f6492f6-61fa-40e8-b92f-42fc55a187bf_content_list.json @@ -0,0 +1,1735 @@ +[ + { + "type": "text", + "text": "Distribution Prototype Diffusion Learning for Open-set Supervised Anomaly Detection", + "text_level": 1, + "bbox": [ + 161, + 128, + 836, + 175 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Fuyun Wang $^{1}$ , Tong Zhang $^{1*}$ , Yuanzhi Wang $^{1}$ , Yide Qiu $^{1}$ , Xin Liu $^{2}$ , Xu Guo $^{1}$ , Zhen Cui $^{3*}$ $^{1}$ Nanjing University of Science and Technology, China. \n $^{2}$ Nanjing SeetaCloud Technology, China. $^{3}$ Beijing Normal University, China \n{fyw271828, tong.zhang, yuanzhiwang, 121106010824, guo.xu} @njust.edu.cn \nxin.liu@seetacloud.com, zhen.cui@bnu.edu.cn", + "bbox": [ + 151, + 202, + 851, + 292 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Abstract", + "text_level": 1, + "bbox": [ + 246, + 325, + 326, + 343 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "In Open-set Supervised Anomaly Detection (OSAD), the existing methods typically generate pseudo anomalies to compensate for the scarcity of observed anomaly samples, while overlooking critical priors of normal samples, leading to less effective discriminative boundaries. To address this issue, we propose a Distribution Prototype Diffusion Learning (DPDL) method aimed at enclosing normal samples within a compact and discriminative distribution space. Specifically, we construct multiple learnable Gaussian prototypes to create a latent representation space for abundant and diverse normal samples and learn a Schrödinger bridge to facilitate a diffusive transition toward these prototypes for normal samples while steering anomaly samples away. Moreover, to enhance inter-sample separation, we design a dispersion feature learning way in hyperspherical space, which benefits the identification of out-of-distribution anomalies. Experimental results demonstrate the effectiveness and superiority of our proposed DPDL, achieving state-of-the-art performance on 9 public datasets.", + "bbox": [ + 89, + 358, + 485, + 647 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "1. Introduction", + "text_level": 1, + "bbox": [ + 89, + 666, + 220, + 683 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Anomaly detection (AD) [15, 21, 48, 50] aims to identify outliers significantly diverging from the prevailing samples in a dataset, and has a wide range of applications like industrial inspection, medical image analysis, and scientific discovery, etc. Recently, unsupervised anomaly detection (UAD) [6, 11, 26, 28] and few-shot anomaly detection (FSAD) [12, 21, 22] have emerged as prominent research paradigms, emphasizing the modeling of normal sample distributions to discern anomalies effectively. Yet, these methods often neglect prior knowledge from limited anomaly samples, resulting in imprecise delineation of normal sample boundaries and reduced efficacy in differenti-", + "bbox": [ + 89, + 693, + 483, + 875 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "ating normal from anomaly instances. On the contrary, supervised anomaly detection (SAD) [2, 20, 47] leverages a limited subset of anomaly samples as prior knowledge, improving detection performance. However, this reliance on seen anomalies poses a risk of overfitting and hampers generalization to unseen anomalies in real-world settings.", + "bbox": [ + 511, + 328, + 905, + 419 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "To mitigate the challenge of limited generalization inherent in closed-set training, we focus on open-set supervised anomaly detection (OSAD) [1, 27, 51, 52], which utilizes a small set of known anomaly classes during training to identify unseen anomalies from open-set classes. By leveraging prior knowledge from observed samples, OSAD methods could reduce false positive errors. To improve the generalized detection of unseen anomalies, DRA [9] leverages data augmentation and outlier exposure to learn a decomposed anomaly representation comprising seen anomalies, pseudo-anomalies, and potential residual anomalies. BGAD [47] leverages decision boundaries derived from normalized flow models to capture and model anomaly information. Recently, AHL [51] simulates heterogeneous anomaly distributions and performs collaborative differentiable learning to further enhance the model's generality.", + "bbox": [ + 511, + 425, + 908, + 667 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "While data augmentation and outlier exposure techniques [9, 19, 34] have demonstrated considerable success in anomaly detection, they fall short in generating comprehensive pseudo anomalies, capturing only a fraction of potential unseen anomalies. This limitation arises from overlooking the intricate nature of real-world anomaly distributions, thereby hindering the model's ability to generalize to novel anomaly types. Although AHL [51] has made strides in addressing this issue by simulating heterogeneous anomaly distributions, it still relies on approximating unknown out-of-distribution anomalies using known indistribution anomalies for generalization. Yet, three critical issues persist: i) the simulation mechanism cannot cover all anomaly distribution patterns due to the varied scales and structures of anomaly distributions; ii) simulated anom", + "bbox": [ + 511, + 674, + 910, + 902 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "CVF", + "bbox": [ + 106, + 2, + 181, + 42 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore.", + "bbox": [ + 236, + 0, + 810, + 46 + ], + "page_idx": 0 + }, + { + "type": "page_footnote", + "text": "*Corresponding authors: Tong Zhang, Zhen Cui", + "bbox": [ + 107, + 886, + 364, + 901 + ], + "page_idx": 0 + }, + { + "type": "page_number", + "text": "20416", + "bbox": [ + 478, + 944, + 519, + 957 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "lies inherit in-distribution data biases, leading to suboptimal performance on out-of-distribution anomalies; iii) the diversity of normal samples presents dilemmas for existing methods, complicating the differentiation between normal and anomaly boundaries. These issues prompt a fundamental question: Instead of generating pseudo and uncertain anomaly samples, how can we accurately characterize compact distribution boundaries amidst a range of diverse normal samples and achieve robust generalization for unknown out-of-distribution anomalies?", + "bbox": [ + 88, + 90, + 480, + 241 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "To address the aforementioned issue, in this work, we propose a Distribution Prototype Diffusion Learning (DPDL) method for open-set supervised anomaly detection. Considering abundant and diverse normal samples but very limited anomaly data, our method involves learning latent distribution prototypes, specifically multiple Gaussian distributions, onto which all observed normal samples can be effectively projected. To facilitate the mapping of normal samples into the prototype space, we leverage Schrödinger bridge (SB) framework, which enables a diffusive transition by aligning the distributions of these samples with the prototypes. The SB-based diffusion way could mitigate the out-of-distribution issue for normal samples to some extent. Within the distribution prototype space, we push observed anomaly samples away from normal samples to enhance discriminative capacity. Notably, both the prototypes and the diffusive bridge are learned jointly, resulting in a robust embedding space for normal samples. Moreover, to enhance generalization across unseen anomaly domains, we introduce a dispersion feature learning mechanism that maps intermediate features to a hyperspherical space, leveraging a mixture of von Mises-Fisher (vMF) distributions. This approach bolsters directional feature extraction and promotes robust inter-sample separation, facilitating effective identification of out-of-distribution samples. Experiments demonstrate that our method greatly improves detection capabilities for unseen anomalies.", + "bbox": [ + 93, + 243, + 483, + 648 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "In summary, our contributions are three-fold: i) we propose a distribution prototype diffusion learning framework that jointly learns multiple Gaussian prototypes and the associated diffusion bridge, creating a compact and discriminative embedding space; ii) we develop dispersion feature learning in hyperspherical space to enhance inter-sample separation and improve generalization; iii) we achieve state-of-the-art performance in 9 public datasets, demonstrating the efficacy of our approach.", + "bbox": [ + 89, + 650, + 482, + 785 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "2. Related Work", + "text_level": 1, + "bbox": [ + 89, + 799, + 232, + 814 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Open-set Supervised Anomaly Detection. Open-set supervised anomaly detection (OSAD) seeks to develop a robust anomaly detection framework that generalizes from a limited set of training anomalies to effectively identify previously unseen anomalies within an open-set con", + "bbox": [ + 89, + 824, + 482, + 901 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "text [1, 27, 35, 47, 52]. Leveraging the prior knowledge provided by observed anomalies, contemporary OSAD methods significantly mitigate false positive errors, thereby enhancing overall detection performance [9, 51]. Recently, DRA [9] learns disentangled representations of observed, pseudo, and residual anomalies to boost the detection of both seen and unseen anomalies. In contrast, AHL [51] simulates diverse heterogeneous anomaly distributions and employs collaborative differentiable learning, significantly improving the model's generalization capacity.", + "bbox": [ + 511, + 90, + 903, + 241 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Schrödinger Bridge. Schrödinger bridge (SB), widely recognized as the entropy-regularized optimal transport (OT) problem, involves learning a stochastic process that evolves from an initial probability distribution to a target distribution under the influence of a reference measure [5, 7, 16, 18, 23, 24, 38], which is often used for content generation as those prior generative models do [30, 42-45]. $\\mathrm{I}^2$ -SB [25] and UNSB [14] learn a nonlinear diffusion process between two given distributions or represent the SB problem as a series of adversarial learning problems to realize the image transformation task. Recently, LightSB [10, 17] introduces a novel, fast, and simple SB solver, which achieves optimal matching in practice through the Gaussian mixture parameterization of the adjusted schrödinger potential.", + "bbox": [ + 511, + 242, + 906, + 469 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "3. Preliminaries", + "text_level": 1, + "bbox": [ + 513, + 481, + 651, + 496 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "We focus on how to build the connection between two distributions $p_0$ and $p_1$ , where the distributions are defined as absolutely continuous Borel probability distributions with finite second-order moments. Building upon the foundation of entropy-regularized optimal transport (EOT) [5, 7, 10, 16-18, 23, 24, 33, 38], we review the related properties of EOT and the schrödinger bridge (SB) problem with a Wiener prior.", + "bbox": [ + 511, + 506, + 903, + 626 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Entropy-regularized optimal transport (EOT). Given two point sets $\\mathcal{Z}_0$ and $\\mathcal{Z}_1$ , we seek for the optimal transport cost between any two points $z_0 \\in \\mathcal{Z}_0$ and $z_1 \\in \\mathcal{Z}_1$ . This task may be formulated as an EOT problem with a parameter $\\epsilon > 0$ , i.e., minimizing the following objective:", + "bbox": [ + 511, + 627, + 903, + 703 + ], + "page_idx": 1 + }, + { + "type": "equation", + "text": "\n$$\n\\begin{array}{l} \\min _ {\\pi \\in \\Pi \\left(p _ {0}, p _ {1}\\right)} \\left\\{\\int_ {\\mathbb {R} ^ {D}} \\int_ {\\mathbb {R} ^ {D}} \\frac {1}{2} \\| z _ {0} - z _ {1} \\| ^ {2} \\pi \\left(z _ {0}, z _ {1}\\right) d z _ {0} d z _ {1} \\right. \\tag {1} \\\\ + \\epsilon \\mathrm {K L} (z \\| p _ {0} \\times p _ {1}) \\}, \\\\ \\end{array}\n$$\n", + "text_format": "latex", + "bbox": [ + 550, + 710, + 903, + 771 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "where $\\Pi(p_0, p_1)$ denotes the set of transport plans, i.e., joint probability distributions on $\\mathbb{R}^D \\times \\mathbb{R}^D$ with marginals $p_0$ and $p_1$ , respectively, and KL denotes Kullback-Leibler divergence. The minimizer $\\pi^*$ of Eqn. (1) is guaranteed to exist, be unique, and absolutely continuous, and is referred as the EOT plan.", + "bbox": [ + 511, + 779, + 903, + 869 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Schrodinger Bridge (SB). We define $\\Omega$ as the space of $\\mathbb{R}^D$ -valued functions over time $t \\in [0,1]$ , representing", + "bbox": [ + 511, + 869, + 903, + 901 + ], + "page_idx": 1 + }, + { + "type": "page_number", + "text": "20417", + "bbox": [ + 478, + 944, + 517, + 955 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "trajectories in $\\mathbb{R}^D$ that start at $t = 0$ and end at $t = 1$ . We denote the set of probability distributions over $\\Omega$ , i.e., stochastic processes, by $\\mathcal{P}(\\Omega)$ . The differential of the standard Wiener process is represented by $dW_{t}$ . For a process $T\\in \\mathcal{P}(\\Omega)$ , we denote its joint distribution at $t = 0,1$ by $\\pi_T\\in \\mathcal{P}(\\mathbb{R}^D\\times \\mathbb{R}^D)$ . Similarly, we use $T|_{z_0,z_1}$ to denote the distribution of $T$ for $t\\in (0,1)$ , conditioned on $T$ 's values $z_{0}$ and $z_{1}$ at $t = 0$ and $t = 1$ , respectively.", + "bbox": [ + 89, + 90, + 480, + 210 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Let $W^{\\epsilon} \\in P(\\Omega)$ represents a Wiener process with volatility $\\epsilon > 0$ , starting from $p_0$ at $t = 0$ . Its differential is governed by the stochastic differential equation (SDE): $dW_{t}^{\\epsilon} = \\sqrt{\\epsilon} dW_{t}$ . The Schrödinger bridge problem with the Wiener prior $W^{\\epsilon}$ between $p_0$ and $p_1$ is minimizing:", + "bbox": [ + 89, + 210, + 480, + 287 + ], + "page_idx": 2 + }, + { + "type": "equation", + "text": "\n$$\n\\min _ {T \\in \\mathcal {F} \\left(p _ {0}, p _ {1}\\right)} \\operatorname {K L} \\left(T \\| W ^ {\\epsilon}\\right), \\tag {2}\n$$\n", + "text_format": "latex", + "bbox": [ + 200, + 297, + 480, + 323 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "where $\\mathcal{F}(p_0,p_1)\\subset \\mathcal{P}(\\Omega)$ denotes the subset of stochastic processes that begin with distribution $p_0$ at $t = 0$ and reach $p_1$ at $t = 1$ . This problem has a unique solution, a SDE diffusion process $T^{*}$ defined: $dZ_{t} = g^{*}(Z_{t},t)dt + dW_{t}^{\\epsilon}$ . The process $T^{*}$ is referred to as SB, and $g^{*}:\\mathbb{R}^{D}\\times [0,1]\\to \\mathbb{R}^{D}$ is the optimal drift.", + "bbox": [ + 89, + 332, + 482, + 421 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Characterization of solutions. The EOT plan $\\pi^{*} = \\pi^{T^{*}}$ takes a specific form [10, 17]:", + "bbox": [ + 89, + 422, + 482, + 454 + ], + "page_idx": 2 + }, + { + "type": "equation", + "text": "\n$$\n\\pi^ {*} \\left(z _ {0}, z _ {1}\\right) = u ^ {*} \\left(z _ {0}\\right) \\exp \\left(\\frac {- \\| z _ {0} - z _ {1} \\| ^ {2}}{2 \\epsilon}\\right) v ^ {*} \\left(z _ {1}\\right), \\tag {3}\n$$\n", + "text_format": "latex", + "bbox": [ + 129, + 462, + 480, + 494 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "where $u^{*}, v^{*} \\colon \\mathbb{R} \\to \\mathbb{R}_{+}$ are measurable functions known as Schrödinger potentials. The optimal drift $g^{*}$ is derived as:", + "bbox": [ + 89, + 501, + 480, + 532 + ], + "page_idx": 2 + }, + { + "type": "equation", + "text": "\n$$\ng ^ {*} (z, t) = \\epsilon \\nabla_ {z} \\log \\int_ {\\mathbb {R} ^ {D}} \\mathcal {N} \\left(z ^ {\\prime} \\mid z, (1 - t) \\epsilon I _ {D}\\right) v ^ {*} \\left(z ^ {\\prime}\\right) d z ^ {\\prime}, \\tag {4}\n$$\n", + "text_format": "latex", + "bbox": [ + 98, + 541, + 482, + 574 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "4. Method", + "text_level": 1, + "bbox": [ + 89, + 582, + 181, + 597 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "4.1. Problem Formulation", + "text_level": 1, + "bbox": [ + 89, + 606, + 294, + 621 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Let $\\mathcal{X}_{\\mathrm{tr}} = \\{(\\mathbf{x}_i, y_i)\\}$ denotes a weakly-supervised training set with only image-level labels, where $\\mathbf{x}_i$ denotes one RGB image and $y_i \\in \\{0,1\\}$ denotes whether $\\mathbf{x}_i$ is an anomaly sample (anomaly: $y_i = 1$ , normal: $y_i = 0$ ). Hereby, $\\mathcal{X}_{\\mathrm{tr}}$ is consistent of a normal subset $\\mathcal{X}_{\\mathrm{tr}}^{\\mathrm{n}}$ ( $|\\mathcal{X}_{\\mathrm{tr}}^{\\mathrm{n}}| = N$ ) and an anomaly subset $\\mathcal{X}_{\\mathrm{tr}}^{\\mathrm{a}}$ ( $|\\mathcal{X}_{\\mathrm{tr}}^{\\mathrm{a}}| = M$ ), formally, $\\mathcal{X}_{\\mathrm{tr}} \\doteq \\mathcal{X}_{\\mathrm{tr}}^{\\mathrm{n}} \\cup \\mathcal{X}_{\\mathrm{tr}}^{\\mathrm{a}}$ , where generally $N \\gg M$ . Given a testing set $\\mathcal{X}_{\\mathrm{te}}$ , we need to predict whether one sample $\\mathbf{x} \\in \\mathcal{X}_{\\mathrm{te}}$ is anomaly or normal. In OSAD, the anomaly patterns of the testing set do less recurred in those encountered training set. In other words, the distribution of anomalies are obviously discrepant, i.e., $P(\\mathcal{X}_{\\mathrm{te}}^{\\mathrm{a}}) \\neq P(\\mathcal{X}_{\\mathrm{tr}}^{\\mathrm{a}})$ . Hereby, we need to learn a robust anomaly detection model $\\psi$ from the training set $\\mathcal{X}_{\\mathrm{tr}}$ so that $\\psi$ accurately infers anomaly scores for test samples.", + "bbox": [ + 89, + 628, + 482, + 839 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Our abstract idea is to learn latent distribution prototypes that not only encapsulate normal samples in a concise manner but also discriminate against anomaly samples. Given the abundance of observed normal samples, one", + "bbox": [ + 89, + 840, + 482, + 900 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "natural approach is to characterize the distribution $P(\\mathcal{X}_{\\mathrm{tr}}^{\\mathrm{n}})$ where samples outside this distribution, $\\mathbf{x}\\notin P(\\mathcal{X}_{\\mathrm{tr}}^{\\mathrm{n}})$ , would be awarded higher probabilities as anomalies. Considering the inherent diversity of normal samples, we endeavor to learn multiple simple distributions (e.g., Gaussians) as prototypes $\\mathcal{P}_{\\mathrm{MGP}}$ , named multi-Gaussian prototypes (MGP). To embed input data into the prototype space, we introduce a generative bridge model $\\psi_{p}$ for distribution transformation. The more abstract formulation is given as follows:", + "bbox": [ + 511, + 90, + 903, + 227 + ], + "page_idx": 2 + }, + { + "type": "equation", + "text": "\n$$\n\\min _ {\\psi_ {p}, \\mathcal {P} _ {\\mathrm {M G P}}, f} D _ {p} \\left(\\psi_ {p} \\left(\\mathcal {F} _ {\\mathrm {t r}} ^ {\\mathrm {n}}\\right), \\psi_ {p} \\left(\\mathcal {F} _ {\\mathrm {t r}} ^ {\\mathrm {a}}\\right), \\mathcal {P} _ {\\mathrm {M G P}}\\right) + \\lambda D _ {s} \\left(\\mathcal {F} _ {\\mathrm {t r}} ^ {\\mathrm {n}}, \\mathcal {F} _ {\\mathrm {t r}} ^ {\\mathrm {a}}\\right), \\tag {5}\n$$\n", + "text_format": "latex", + "bbox": [ + 516, + 233, + 903, + 258 + ], + "page_idx": 2 + }, + { + "type": "equation", + "text": "\n$$\ns. t., \\psi_ {p}: P (\\mathcal {F}) \\stackrel {\\text {b r i d g e}} {\\Longrightarrow} \\mathcal {P} _ {\\mathrm {M G P}}, \\tag {6}\n$$\n", + "text_format": "latex", + "bbox": [ + 547, + 261, + 903, + 282 + ], + "page_idx": 2 + }, + { + "type": "equation", + "text": "\n$$\nf: \\mathcal {X} \\xrightarrow {\\text {f e a t u r e}} \\mathcal {F}, \\tag {7}\n$$\n", + "text_format": "latex", + "bbox": [ + 580, + 286, + 903, + 306 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "where $\\mathcal{P}_{\\mathrm{MGP}}$ denotes the distribution prototypes to be learned, $\\psi_{p}$ is the flow function across probability distributions, $D_{p}$ signifies the discriminative function in the space of prototypes, $f$ stands for the feature extraction process, and $D_{s}$ acts as a regularizer to increase feature discriminability. In the above formulation, besides distribution prototype learning (DPL), we also introduce dispersion feature learning (DFL) executed within a hyperspherical embedding space, i.e., $D_{s}(\\mathcal{F}_{\\mathrm{tr}}^{\\mathrm{n}},\\mathcal{F}_{\\mathrm{tr}}^{\\mathrm{a}})$ . The advantage of DFL is to prevent abrupt feature collapse in feature learning, thus preserving discriminative qualities. The overview of our framework is abstractly depicted in Fig. 1. Detailed elaboration on these aspects will be provided in the subsequent sections.", + "bbox": [ + 511, + 311, + 903, + 508 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "4.2. Distribution Prototype Learning", + "text_level": 1, + "bbox": [ + 511, + 516, + 799, + 532 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Distribution prototype learning is operated in the feature space. Hence, to extract intermediate image features, we can leverage those classic networks as the backbone, such as ResNet-18. Formally, the feature extraction process, denoted as a function $f: \\mathcal{X} \\to \\mathcal{F}$ , transforms an image $\\mathbf{x}$ into the intermediate feature $\\mathbf{x} = f(\\mathbf{x}) \\in \\mathbb{R}^d$ , where a flatten operation is used on convolution maps for vectorization. Consequently, we designate the intermediate feature sets of normal and anomaly samples as $\\mathcal{F}_{\\mathrm{tr}}^{\\mathrm{n}} = \\{\\mathbf{x}_i^{\\mathrm{n}} | i = 1, \\dots, N\\}$ and $\\mathcal{F}_{\\mathrm{tr}}^{\\mathrm{a}} = \\{\\mathbf{x}_j^{\\mathrm{a}} | i = 1, \\dots, M\\}$ , respectively.", + "bbox": [ + 511, + 537, + 905, + 688 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Considering the rarity of anomalies and the diversity of normal samples, it is pertinent to capture the intricate distributions of normal samples by mapping the distribution $p_0 = P(\\mathcal{F}_{\\mathrm{tr}}^{\\mathrm{n}})$ of intermediate features to a distinct and well-characterized distribution $p_1$ like Gaussian. But due to various normal samples, we opt for a multi-Gaussian prototypes (MGP) comprising multiple Gaussian distributions as prototypes $\\mathcal{P}_{\\mathrm{MGP}} = \\{\\mathcal{P}_i \\doteq \\mathcal{N}(\\pmb{\\mu}_i, \\pmb{\\sigma}_i) | i = 1, \\dots, C\\}$ . Given a normal sample $\\mathbf{x}_i^{\\mathrm{n}}$ , we expect to align it with the closest prototype with high likelihood. However, in the open-set setting, it is challenge to transform unseen points $\\mathbf{x} \\sim p_0$ to the target distribution $p_1$ . Drawing inspiration from the capability of diffusion generation models in aligning disparate distributions, we frame the transition from source", + "bbox": [ + 511, + 689, + 905, + 900 + ], + "page_idx": 2 + }, + { + "type": "page_number", + "text": "20418", + "bbox": [ + 478, + 944, + 517, + 955 + ], + "page_idx": 2 + }, + { + "type": "image", + "img_path": "images/e170339ed7ef8b2503fd8d411bf6288fe8634990f0c3b92f2d8db4a2dec08fe9.jpg", + "image_caption": [ + "Figure 1. Our proposed DPDL framework. It comprises three distinct modules: Distribution Prototype Learning (DPL, Sec. 4.2), Dispersion Feature Learning (DFL, Sec. 4.3), and anomaly score prediction (Sec. 4.4). DPL transforms the distribution of normal samples to a space of learnable multiple Gaussian prototypes through building schrödinger bridge, meantime pushing anomaly distribution away from these prototypes. DFL operates in a hyperspherical space, enlarging the distances of intermediate features of all samples in a hyperspherical space to strengthen feature generalization for detecting anomalies. The score prediction module leverages a multi-instance-learning method to compute anomaly scores." + ], + "image_footnote": [], + "bbox": [ + 99, + 85, + 898, + 327 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "domain distribution $p_0$ to target domain distribution $p_1$ as a Schrodinger bridge problem. As formulated in Eqn. (6), we need learn an essential bridge flow $\\psi_p: P(\\mathcal{F}_{\\mathrm{tr}}^{\\mathrm{n}}) \\xrightarrow{\\mathrm{bridge}} \\mathcal{P}_{\\mathrm{MGP}}$ . After bridge transformation, the condition probability in optimal transport plan conforms to:", + "bbox": [ + 89, + 435, + 483, + 518 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\pi \\left(\\psi_ {p} (\\mathbf {x}) | \\mathbf {x}\\right) \\propto \\underbrace {\\sum_ {c = 1} ^ {C} \\alpha_ {c} \\mathcal {N} \\left(\\psi_ {p} (\\mathbf {x}) ; \\boldsymbol {\\mu} _ {c} , \\boldsymbol {\\sigma} _ {c}\\right)} _ {\\dot {\\equiv} \\phi_ {1} \\left(\\psi_ {p} (\\mathbf {x})\\right)}, \\tag {8}\n$$\n", + "text_format": "latex", + "bbox": [ + 145, + 521, + 482, + 585 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "where the parameters $\\{\\alpha_c,\\pmb {\\mu}_c,\\pmb {\\sigma}_c\\}_{c = 1}^C$ (known as Gaussian mixed model (GMM), abstracted into the function $\\phi_1$ ) as well as the flow function $\\psi_{p}$ (in bridge) need to be learned. For simplicity, below we adopt diagonal matrices for $\\pmb{\\sigma}_{c}$", + "bbox": [ + 89, + 592, + 482, + 652 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Drawing inspiration from previous works [10, 17], we reframe distribution prototype learning in Eqn. (8) as the Schrödinger bridge. Given the specific form taken by the EOT plan described in Eqn. (3), we redefine these measurable functions $u, v: \\mathbb{R} \\to \\mathbb{R}_+$ , termed Schrödinger potentials [10, 17], as follows:", + "bbox": [ + 89, + 652, + 483, + 744 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\nu \\left(\\mathbf {x} _ {i} ^ {\\mathrm {n}}\\right) \\doteq \\exp \\left(\\frac {\\left\\| \\mathbf {x} _ {i} ^ {\\mathrm {n}} \\right\\| ^ {2}}{2 \\epsilon}\\right) \\phi_ {0} \\left(\\mathbf {x} _ {i} ^ {\\mathrm {n}}\\right), \\tag {9}\n$$\n", + "text_format": "latex", + "bbox": [ + 173, + 748, + 482, + 780 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\nv \\left(\\psi \\left(\\mathbf {x} _ {i} ^ {\\mathrm {n}}\\right)\\right) \\doteq \\exp \\left(\\frac {\\| \\psi \\left(\\mathbf {x} _ {i} ^ {\\mathrm {n}}\\right) \\| ^ {2}}{2 \\epsilon}\\right) \\phi_ {1} \\left(\\psi \\left(\\mathbf {x} _ {i} ^ {\\mathrm {n}}\\right)\\right), \\tag {10}\n$$\n", + "text_format": "latex", + "bbox": [ + 151, + 781, + 482, + 813 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "where $\\phi_0$ is defined within the source feature domain $\\mathcal{F}_{\\mathrm{tr}}$ , and $\\phi_{1}$ is defined in Eqn. (8). $\\epsilon$ is set to 0.001 in our experiments. Accordingly, Eqn. (3) can be converted to:", + "bbox": [ + 89, + 816, + 482, + 863 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\pi \\left(\\mathbf {x} _ {i} ^ {\\mathrm {n}}, \\psi \\left(\\mathbf {x} _ {i} ^ {\\mathrm {n}}\\right)\\right) = \\phi_ {0} \\left(\\mathbf {x} _ {i} ^ {\\mathrm {n}}\\right) \\exp \\left(\\frac {\\left\\langle \\mathbf {x} _ {i} ^ {\\mathrm {n}} , \\psi \\left(\\mathbf {x} _ {i} ^ {\\mathrm {n}}\\right) \\right\\rangle}{\\epsilon}\\right) \\phi_ {1} \\left(\\psi \\left(\\mathbf {x} _ {i} ^ {\\mathrm {n}}\\right)\\right), \\tag {11}\n$$\n", + "text_format": "latex", + "bbox": [ + 89, + 867, + 483, + 898 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Hence, the condition probability of the transport plan in Eqn. (8) could be exactly defined as", + "bbox": [ + 511, + 435, + 906, + 465 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\pi (\\psi (\\mathbf {x} _ {i} ^ {\\mathrm {n}}) | \\mathbf {x} _ {i} ^ {\\mathrm {n}}) \\doteq \\eta (\\mathbf {x} _ {i} ^ {\\mathrm {n}}, \\psi (\\mathbf {x} _ {i} ^ {\\mathrm {n}})) \\phi_ {1} (\\psi (\\mathbf {x} _ {i} ^ {\\mathrm {n}})), \\tag {12}\n$$\n", + "text_format": "latex", + "bbox": [ + 570, + 474, + 906, + 494 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "where the connection factor is denoted as $\\eta (\\mathbf{x}_i^n,\\psi (\\mathbf{x}_i^n)) = \\frac{1}{\\varpi(\\mathbf{x}_i^n)}\\exp \\left(\\frac{\\langle\\mathbf{x}_i^n,\\psi(\\mathbf{x}_i^n)\\rangle}{\\epsilon}\\right)$ with the normalization term $\\varpi (\\mathbf{x}_i^n) = \\int \\exp (\\langle \\mathbf{x}_i^n,\\psi (\\mathbf{x}_i^n)\\rangle)\\phi_1(\\psi (\\mathbf{x}_i^n))d\\psi (\\mathbf{x}_i^n)$ . According to Eqns. (8) and (12), we can further derive a more tractable form [10, 17]:", + "bbox": [ + 511, + 502, + 906, + 584 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\pi \\left(\\psi \\left(\\mathbf {x} _ {i} ^ {\\mathrm {n}}\\right) \\mid \\mathbf {x} _ {i} ^ {\\mathrm {n}}\\right) = \\widetilde {\\eta} \\left(\\mathbf {x} _ {i} ^ {\\mathrm {n}}\\right) \\sum_ {c = 1} ^ {C} \\widetilde {\\alpha} _ {c} \\left(\\mathbf {x} _ {i} ^ {\\mathrm {n}}\\right) \\mathcal {N} \\left(\\psi \\left(\\mathbf {x} _ {i} ^ {\\mathrm {n}}\\right); \\widetilde {\\boldsymbol {\\mu}} _ {c} \\left(\\mathbf {x} _ {i} ^ {\\mathrm {n}}\\right), \\boldsymbol {\\sigma} _ {c}\\right), \\tag {13}\n$$\n", + "text_format": "latex", + "bbox": [ + 514, + 593, + 905, + 635 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "where the normalization factor is denoted as $\\widetilde{\\eta} (\\mathbf{x}_i^n) = 1 / \\sum_{c = 1}^{C}\\widetilde{\\alpha}_c(\\mathbf{x}_i^n)$ , the coefficients of multi-Gaussian are defined as $\\widetilde{\\alpha}_{c}(\\mathbf{x}_{i}^{n}) = \\alpha_{c}\\exp (\\frac{1}{2} (\\mathbf{x}_{i}^{n})^{\\intercal}\\pmb{\\sigma}_{c}\\mathbf{x}_{i}^{n} + \\frac{1}{\\epsilon} (\\widetilde{\\mu}_{c})^{\\intercal}(\\mathbf{x}_{i}^{n}))$ , and the mean vectors are calculated as $\\widetilde{\\pmb{\\mu}}_c(\\mathbf{x}_i^n) = \\pmb {\\mu}_c + \\frac{1}{\\epsilon}\\pmb {\\sigma}_c\\mathbf{x}_i^n$", + "bbox": [ + 511, + 643, + 905, + 707 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "We proceed with deriving the bridge function $\\psi_{p}$ that represents a SDE process: $d\\mathbf{x}_t = g(x_t,t)dt + \\sqrt{\\epsilon} dW_t$ where the shift function $g$ is solved. According to Eqn. (4), we can obtain the shift function of diffusion process as follows:", + "bbox": [ + 511, + 707, + 905, + 768 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\ng (\\mathbf {x} _ {i} ^ {\\mathrm {n}}, t) = \\epsilon \\rho_ {\\mathrm {M G P}} (\\mathbf {x} _ {i} ^ {\\mathrm {n}}) \\nabla_ {\\mathbf {x} _ {i} ^ {\\mathrm {n}}} \\log (\\mathcal {N} (\\mathbf {x} _ {i} ^ {\\mathrm {n}} | 0, \\epsilon (1 - t)) \\mathbf {I}), \\quad (1 4)\n$$\n", + "text_format": "latex", + "bbox": [ + 516, + 792, + 906, + 811 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "where the coefficients $\\rho_{\\mathrm{MGP}}$ defined on multiple distribution prototypes is calculated as: $\\rho_{\\mathrm{MGP}}(\\mathbf{x}_i^{\\mathrm{n}}) = \\sum_{c=1}^{C} (\\alpha_c \\mathcal{N}(\\widetilde{\\boldsymbol{\\mu}}_c(\\mathbf{x}_i^{\\mathrm{n}}) | \\mathbf{0}, \\boldsymbol{\\sigma}_c) \\mathcal{N}(h_c(\\mathbf{x}_i^{\\mathrm{n}}, t) | 0, \\boldsymbol{\\Sigma}_c^t))$ , with another Gaussian of $h_c(\\mathbf{x}_i^{\\mathrm{n}}, t) = \\frac{1}{\\epsilon(1 - t)} \\mathbf{x}_i^{\\mathrm{n}} + \\boldsymbol{\\sigma}_c^{-1} \\widetilde{\\boldsymbol{\\mu}}_c(\\mathbf{x}_i^{\\mathrm{n}})$ and $\\boldsymbol{\\Sigma}_c^t = \\frac{t}{\\epsilon(1 - t)} \\mathbf{I} + \\boldsymbol{\\sigma}_c^{-1}$ .", + "bbox": [ + 511, + 819, + 906, + 904 + ], + "page_idx": 3 + }, + { + "type": "page_number", + "text": "20419", + "bbox": [ + 478, + 944, + 519, + 957 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Distribution prototypes initialization. Jointly learning prototypes $\\{\\alpha_c, \\mu_c, \\sigma_c\\}_{c=1}^C$ and the bridge transformation $\\psi_p$ (also the shift $g$ ) is a challenging task as they are interdependent. For this, we leverage a vector quantization function to learn a codebook $\\mathcal{E}$ of prototypes within a discrete latent space from training data. Specifically, given an input image feature $\\mathbf{x}_i^n$ , we can assign $\\mathbf{x}_i^n$ to the closest prototype $\\mathbf{e}_k$ by minimizing the L2 distance between $\\mathbf{x}_i^n$ and each of prototypes $\\mathbf{e}_c \\in \\mathcal{E}$ , as follows:", + "bbox": [ + 89, + 90, + 483, + 227 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\min _ {\\left\\{\\mathbf {e} _ {c} \\right\\}} \\mathbb {E} _ {\\mathbf {x} _ {i} ^ {\\mathrm {n}} \\in \\mathcal {F} _ {\\mathrm {t r}}} \\left[ \\left\\| \\mathbf {x} _ {i} ^ {\\mathrm {n}} - \\mathbf {e} _ {c ^ {*}} \\right\\| _ {2} ^ {2} \\right], \\text {s . t .}, c ^ {*} = \\underset {c} {\\arg \\min } \\left\\| \\mathbf {x} _ {i} ^ {\\mathrm {n}} - \\mathbf {e} _ {c} \\right\\| _ {2}, \\tag {15}\n$$\n", + "text_format": "latex", + "bbox": [ + 91, + 234, + 483, + 261 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "where $c^*$ denotes the index of the prototype closest to $\\mathbf{x}_i^{n}$ . The learned $\\{\\mathbf{e}_c\\}_{c=1}^C$ are used to initialize the mean vectors $\\{\\pmb{\\mu}_c\\}_{c=1}^C$ , while the variances of all prototypes are set to the identity matrix. We observe that this initialization strategy accelerates the training process.", + "bbox": [ + 89, + 267, + 483, + 343 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Distribution loss of normal and anomaly samples. As $p_0$ and $p_1$ are accessible only via samples $\\mathcal{F}_{\\mathrm{tr}}$ and prototypes $\\mathcal{P}_{\\mathrm{MGP}}$ , we optimize the empirical form in Eqn. (2) for normal samples as follows [10, 17]:", + "bbox": [ + 89, + 343, + 483, + 404 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {L} _ {\\mathrm {D P L}} ^ {\\mathrm {n}} = \\frac {1}{N} \\sum_ {i = 1} ^ {N} \\log \\varpi_ {\\theta} \\left(\\mathbf {x} _ {i} ^ {\\mathrm {n}}\\right) - \\frac {1}{C} \\sum_ {c = 1} ^ {C} \\log \\phi_ {1} \\left(\\boldsymbol {\\mu} _ {c}\\right), \\tag {16}\n$$\n", + "text_format": "latex", + "bbox": [ + 109, + 411, + 482, + 454 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "In contrast, for anomaly samples, we aim to push anomaly distribution away from $p_1$ , i.e., negative loss, formally,", + "bbox": [ + 89, + 460, + 483, + 491 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {L} _ {\\mathrm {D P L}} ^ {\\mathrm {a}} = \\frac {1}{C} \\sum_ {c = 1} ^ {C} \\log \\phi_ {1} (\\boldsymbol {\\mu} _ {c}) - \\frac {1}{M} \\sum_ {i = 1} ^ {M} \\log \\varpi_ {\\theta} \\left(\\mathbf {x} _ {i} ^ {\\mathrm {a}}\\right), \\tag {17}\n$$\n", + "text_format": "latex", + "bbox": [ + 107, + 498, + 482, + 540 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "4.3. Dispersion Feature Learning", + "text_level": 1, + "bbox": [ + 89, + 546, + 349, + 561 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "A critical challenge in OSAD is detecting previously unseen anomalies in open-set environments. Due to the limited anomaly observations, existing methods often use pseudo-anomaly generation strategies as a means of effective data augmentation. Nonetheless, the effectiveness of these methods heavily depends on the quality of the pseudo-anomaly feature embeddings. Notably, pseudo-anomaly distributions often inherit biases from the in-distribution data, which differ from the unknown anomaly distributions in out-of-distribution data. Current methods fail to address the relationship between observed and unknown anomalies, particularly for out-of-distribution generalization.", + "bbox": [ + 89, + 568, + 483, + 750 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "To improve out-of-distribution detection, it is essential to promote a larger inter-sample dispersion, as greater distances among in-distribution samples facilitate their more effective separation from out-of-distribution samples. In other words, if all inter-sample distances are close to zero, i.e., collapse to a single point, it becomes impossible to differentiate between samples. Hence, promoting separability through a larger inter-sample dispersion is critical for accurately identifying samples that do not belong to known categories. To do so, we map the features into a hyperspherical", + "bbox": [ + 89, + 750, + 483, + 901 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "space. Draws inspiration from the vMF distribution [31] in directional statistics, we compute spherical Gaussian distributions for unit-norm features $\\widehat{\\mathbf{x}}_i = \\mathbf{x}_i / \\| \\mathbf{x}_i\\| _2^2$ . The probability density function of a unit vector $\\widehat{\\mathbf{x}}_i\\in \\mathbb{R}^D$ is defined in the hyperspherical space as follows:", + "bbox": [ + 511, + 90, + 905, + 167 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\np _ {D} (\\widehat {\\mathbf {x}} _ {i}; \\widehat {\\mathbf {x}} _ {j}, \\kappa) = F _ {D} (\\kappa) \\exp \\left(\\kappa \\langle \\widehat {\\mathbf {x}} _ {i}, \\widehat {\\mathbf {x}} _ {j} \\rangle\\right), \\tag {18}\n$$\n", + "text_format": "latex", + "bbox": [ + 573, + 179, + 906, + 196 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "where $\\kappa \\geq 0$ controls the concentration of the distribution around the mean direction $\\widehat{\\mathbf{x}}_j$ and $F_{D}(\\kappa)$ is the normalization factor. A larger $\\kappa$ value increases concentration around the mean, while in the extreme case $\\kappa = 0$ , sample points are uniformly distributed on the hypersphere. Therefore, we design a dispersion loss to optimize large angular distances between the features of all samples:", + "bbox": [ + 511, + 208, + 906, + 315 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {L} _ {\\mathrm {D F L}} = \\frac {1}{U} \\sum_ {i = 1} ^ {U} \\log \\frac {1}{U - 1} \\sum_ {i, j = 1} ^ {U} \\mathbb {1} \\{i \\neq j \\} \\exp (\\kappa \\langle \\widehat {\\mathbf {x}} _ {i}, \\widehat {\\mathbf {x}} _ {j} \\rangle), \\tag {19}\n$$\n", + "text_format": "latex", + "bbox": [ + 514, + 325, + 906, + 369 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "where $U = N + M$ , and $\\kappa$ is set to 10 in our experiments.", + "bbox": [ + 511, + 378, + 898, + 395 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "4.4. Anomaly Score Prediction", + "text_level": 1, + "bbox": [ + 511, + 405, + 751, + 421 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Based on the above designs, we leverage the multiple-instance-learning (MIL)-based method proposed in [35] to effectively learn anomaly scores. Similar to the work [9], we design three modules $\\mathsf{M} = \\{\\mathsf{M}_{\\mathsf{a}},\\mathsf{M}_{\\mathsf{n}}$ and $\\mathsf{M}_{\\mathsf{r}}\\}$ for estimating anomaly scores. Firstly, for the feature map $\\mathbf{x}_i^1$ , we generate pixel-wise feature vectors $\\mathcal{V} = \\{\\mathbf{v}_i\\}_{i = 1}^{H' \\times W'}$ to represent the feature of small patches of the image $\\mathbf{x}_i$ , where $(H',W')$ denotes the size of the feature map. These pixel-wise representations are then mapped by an anomaly classifier $S_{\\mathrm{a}}$ to estimate pixel-level anomaly scores. To capture those points with the most salient anomalies, we compute the top- $K$ most anomaly pixel points and define the loss function as:", + "bbox": [ + 511, + 428, + 906, + 625 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {L} _ {\\mathrm {M} _ {\\mathrm {a}}} \\left(\\mathbf {x} _ {i}, y _ {i}\\right) = \\mathcal {L} _ {\\text {b i n a r y}} \\left(\\frac {1}{K} \\sum \\operatorname {T o p} _ {K} \\left\\{S _ {\\mathrm {a}} \\left(\\mathbf {v} _ {i}; \\theta_ {\\mathrm {a}}\\right) \\right\\}, y _ {i}\\right), \\tag {20}\n$$\n", + "text_format": "latex", + "bbox": [ + 521, + 633, + 906, + 664 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "where $\\mathcal{L}_{\\mathrm{binary}}$ refers to a binary classification loss function, and $\\operatorname{Top}_K$ selects the highest $K$ anomaly scores among all the vectors. We use $\\mathsf{M}_{\\mathfrak{n}}$ to learn the normal features:", + "bbox": [ + 511, + 672, + 905, + 719 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {L} _ {\\mathrm {M} _ {\\mathrm {n}}} \\left(\\mathbf {x} _ {i}, y _ {i}\\right) = \\mathcal {L} _ {\\text {b i n a r y}} \\left(S _ {\\mathrm {n}} \\left(\\frac {1}{H ^ {\\prime} \\times W ^ {\\prime}} \\sum_ {i = 1} ^ {H ^ {\\prime} \\times W ^ {\\prime}} \\mathbf {v} _ {i}; \\theta_ {\\mathrm {n}}\\right), y _ {i}\\right), \\tag {21}\n$$\n", + "text_format": "latex", + "bbox": [ + 524, + 729, + 903, + 785 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "where $S_{\\mathfrak{n}}:\\mathcal{V}\\to \\mathbb{R}$ is a fully connected binary anomaly classifier. Finally, we define $\\mathsf{M}_{\\mathfrak{x}}$ to compute the residual anomaly scores between fine-grained visual semantics and abstract prototypes:", + "bbox": [ + 511, + 786, + 905, + 847 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {L} _ {\\mathrm {M} _ {\\mathrm {r}}} = \\mathcal {L} _ {\\text {b i n a r y}} \\left(S _ {\\mathrm {r}} \\left(\\left(\\psi_ {p} (\\mathbf {x} _ {i}) - \\boldsymbol {\\mu} _ {c ^ {*}}\\right) / \\boldsymbol {\\sigma} _ {c ^ {*}}; \\theta_ {\\mathrm {r}}\\right), y _ {i}\\right), \\tag {22}\n$$\n", + "text_format": "latex", + "bbox": [ + 532, + 858, + 906, + 877 + ], + "page_idx": 4 + }, + { + "type": "page_footnote", + "text": "Here $\\mathbf{x}_i$ refers to convolution maps without flattening (as used above).", + "bbox": [ + 531, + 886, + 903, + 900 + ], + "page_idx": 4 + }, + { + "type": "footer", + "text": "20420", + "bbox": [ + 478, + 944, + 519, + 957 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "where $c^*$ denotes the index of the most probable prototypes, i.e., $c^* = \\arg \\max_c\\mathcal{N}(\\psi_p(\\mathbf{x}_i);\\pmb {\\mu}_c,\\pmb {\\sigma}_c)$ , and $S_{\\mathrm{r}}$ utilize the same method to obtain anomaly score as $S_{\\mathrm{a}}$ .", + "bbox": [ + 89, + 90, + 480, + 136 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Training. During the training phase, the SB and three prediction modules are jointly trained. To this end, we employ an objective function that encompasses three components as follows:", + "bbox": [ + 89, + 137, + 482, + 198 + ], + "page_idx": 5 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {L} = \\underbrace {\\mathcal {L} _ {\\mathrm {M} _ {\\mathrm {a}}} + \\mathcal {L} _ {\\mathrm {M} _ {\\mathrm {n}}} + \\mathcal {L} _ {\\mathrm {M} _ {\\mathrm {r}}}} _ {\\text {M I L - b a s e d l e a n i n g}} + \\underbrace {\\mathcal {L} _ {\\mathrm {D P L}} ^ {\\mathrm {n}} + \\mathcal {L} _ {\\mathrm {D P L}} ^ {\\mathrm {a}}} _ {\\text {S B t r a n s f o r m}} + \\underbrace {\\lambda \\mathcal {L} _ {\\mathrm {D F L}}} _ {\\text {d i s p e r s i o n}}. \\tag {23}\n$$\n", + "text_format": "latex", + "bbox": [ + 102, + 212, + 482, + 247 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "where the coefficient $\\lambda$ modulates the relative importance of dispersion loss, and the learnable parameters include $\\{\\alpha_{c},\\pmb{\\mu}_{c},\\pmb{\\sigma}_{c}\\}$ , $\\theta_{\\psi_p}$ , and $\\{\\theta_{\\mathrm{a}},\\theta_{\\mathrm{n}},\\theta_{\\mathrm{r}}\\}$ .", + "bbox": [ + 89, + 261, + 482, + 306 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Inference. During the test phase, we find the most similar class prototype through SB, and subsequently compute the anomaly score by adding the scores from both $S_{\\mathrm{a}}$ and $S_{\\mathrm{r}}$ , while subtracting the normal score obtained from $S_{\\mathrm{n}}$ for the given test image.", + "bbox": [ + 89, + 308, + 482, + 383 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "5. Experiment", + "text_level": 1, + "bbox": [ + 89, + 400, + 215, + 416 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "5.1. Dataset and Evaluation Metric", + "text_level": 1, + "bbox": [ + 89, + 426, + 362, + 441 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Dataset To validate the effectiveness of DPDL, comprehensive experiments are conducted on nine real-world AD datasets, including six industrial defect detection datasets (MVTec AD [3], Optical [46], SDD [40], AITEX [39], ELPV [8], Mastcam [13]) and three medical image datasets (Hyper-Kvasir [4], Brain-MRI [37], HeadCT [37]). We follow the previous OSAD baselines [9, 51] to adopt two protocols for sampling, including general setting and hard setting. The general setting assumes that anomaly examples are randomly sampled from the anomaly class, while the hard setting samples from a single class to assess generalization to new or unseen anomaly classes.", + "bbox": [ + 89, + 449, + 482, + 630 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Evaluation Metric We utilize the widely adopted Area Under ROC Curve (AUC) as a metric to evaluate the performance across all methods and settings. All reported AUCs are averaged results over five independent runs.", + "bbox": [ + 89, + 631, + 482, + 691 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "5.2. Baselines", + "text_level": 1, + "bbox": [ + 89, + 705, + 199, + 720 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "We compare DPDL against six related state-of-the-art OSAD baselines, including SAOE [19, 32, 41], MLEP [27], FLOS [36], DevNet [35], DRA [9], and AHL [51]. MLEP, DevNet, DRA, and AHL are specifically designed for OSAD. SAOE is a supervised detector enhanced with synthetic anomalies and anomaly exposure, whereas FLOS is an imbalanced classifier leveraging focal loss.", + "bbox": [ + 89, + 728, + 482, + 834 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "5.3. Implementation Details", + "text_level": 1, + "bbox": [ + 89, + 847, + 307, + 863 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "The input image size is $448 \\times 448 \\times 3$ . We set $K$ in the top- $K$ MIL to $10\\%$ of the number of all scores per score map.", + "bbox": [ + 89, + 869, + 482, + 900 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "AdamW optimizer [29] is used for the parameter optimization using an initial learning rate $2 \\times 10^{-4}$ with a weight decay of $1 \\times 10^{-5}$ . DPDL is trained on one NVIDIA GeForce RTX 4090 GPU, which are trained using 50 epochs, with 20 iterations per epoch. Following previous protocol [9, 51], we evaluate performance with anomaly sample numbers of $M = 10$ and $M = 1$ , and for robust detection of unseen anomalies, we use CutMix [49] to create pseudo-anomaly samples as augmented data for known anomalies. The prototype quantity $C$ is set to 32 as default.", + "bbox": [ + 511, + 90, + 906, + 243 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "5.4. Results under General Setting", + "text_level": 1, + "bbox": [ + 511, + 265, + 782, + 281 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Tab. 1 highlights DPDL's strong performance. In the challenging scenario of single-anomaly detection, DPDL improves the performance of AHL [51] by more than $8.3\\%$ on the AITEX, ELPV, and Mastcam datasets. Furthermore, it achieves significant improvements across six additional datasets, which suggest the effective utilization of few-shot anomaly examples in DPDL, while mitigating overfitting to the seen anomalies. When shifting to ten anomaly examples settings, DPDL continues to maintain a significant lead with over $5.4\\%$ improvement on those datasets. Given the rich and diverse set of normal samples in these datasets, DPDL leverages the DPL component to encapsulate these samples within a compact, discriminative distribution space, while effectively pushing anomalous samples outside this space, thereby enabling accurate anomaly detection. In the setting with ten abnormal samples, although existing methods have reached performance saturation on the MVTecAD, Optical, and SDD datasets, DPDL still has a certain lead, demonstrating the strong ability of DPL and DFL in learning tight boundaries of normal sample distributions and generalizing to previously unseen anomaly domains. Furthermore, when evaluated on the medical datasets BrainMRI and HeadCT, DPDL demonstrates competitive performance despite these datasets being notably small in scale and containing only a single class. This highlights the algorithm's ability to deliver robust results even in data-scarce conditions.", + "bbox": [ + 511, + 292, + 906, + 684 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "5.5. Results under the Hard Setting", + "text_level": 1, + "bbox": [ + 511, + 708, + 790, + 724 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Tab. 2 summarizes the performance comparison under the hard setting. It is evident that DPDL achieves the highest AUC scores in both the single-anomaly and ten-anomaly sample settings. Specifically, compared to the closest competing method, AHL [51], DPDL achieves an improvement in AUC scores ranging from $0.6\\%$ to $7.9\\%$ in the ten anomaly examples settings and from $0.5\\%$ to $8.9\\%$ in the one anomaly example settings, respectively. The observed improvement can be attributed to the strong generalization ability of DPDL in detecting unseen anomaly classes, even when the model is trained on only a single anomaly class.", + "bbox": [ + 511, + 734, + 906, + 900 + ], + "page_idx": 5 + }, + { + "type": "page_number", + "text": "20421", + "bbox": [ + 478, + 944, + 517, + 955 + ], + "page_idx": 5 + }, + { + "type": "table", + "img_path": "images/e96ce7dd2aa97b803eb69c971b7f9b6a0303ef52040b6330cca4c42d806bfc4b.jpg", + "table_caption": [ + "Table 1. AUC performance (mean ± std) across nine real-world AD datasets is reported under the general setting. red highlights the best results, and blue indicates sub-optimal outcomes. All baseline SOTA results are sourced from the original papers [9, 51]." + ], + "table_footnote": [], + "table_body": "
DatasetDevNetFLOSSAOEMLEPDRAAHLDPDL (Ours)
Ten Training Anomaly Examples
MVTec AD0.945±0.0040.939±0.0070.926±0.0100.907±0.0050.959±0.0030.970±0.0020.977±0.002
Optical0.782±0.0650.720±0.0550.941±0.0130.740±0.0390.965±0.0060.976±0.0040.983±0.005
SDD0.988±0.0060.967±0.0180.955±0.0200.983±0.0130.991±0.0050.991±0.0010.996±0.001
AITEX0.887±0.0130.841±0.0490.874±0.0240.867±0.0370.893±0.0170.925±0.0130.975±0.007
ELPV0.846±0.0220.818±0.0320.793±0.0470.794±0.0470.845±0.0130.850±0.0040.937±0.003
Mastcam0.790±0.0210.703±0.0290.810±0.0290.798±0.0260.848±0.0080.855±0.0050.934±0.010
Hyper-Kvasir0.829±0.0180.773±0.0290.666±0.0500.600±0.0690.834±0.0040.880±0.0030.939±0.005
BrainMRI0.958±0.0120.955±0.0110.900±0.0410.959±0.0110.970±0.0030.977±0.0010.969±0.005
HeadCT0.982±0.0090.971±0.0040.935±0.0210.972±0.0140.972±0.0020.999±0.0030.981±0.003
One Training Anomaly Example
MVTec AD0.780±0.0200.755±0.1360.834±0.0070.744±0.0190.883±0.0080.901±0.0030.927±0.002
Optical0.523±0.0030.518±0.0030.815±0.0140.516±0.0090.888±0.0120.888±0.0070.915±0.002
SDD0.881±0.0090.840±0.0430.781±0.0090.811±0.0450.859±0.0140.909±0.0010.917±0.003
AITEX0.598±0.0700.538±0.0730.675±0.0940.564±0.0550.692±0.1240.734±0.0080.838±0.008
ELPV0.514±0.0760.457±0.0560.635±0.0920.578±0.0620.675±0.0240.828±0.0050.897±0.002
Mastcam0.595±0.0160.542±0.0170.662±0.0180.625±0.0450.692±0.0580.743±0.0030.838±0.011
Hyper-Kvasir0.653±0.0370.668±0.0040.498±0.1000.445±0.0400.690±0.0170.768±0.0150.821±0.007
BrainMRI0.694±0.0040.693±0.0360.531±0.0600.632±0.0170.744±0.0040.866±0.0040.893±0.004
HeadCT0.742±0.0760.698±0.0920.597±0.0220.758±0.0380.796±0.1050.825±0.0140.865±0.005
", + "bbox": [ + 98, + 127, + 895, + 428 + ], + "page_idx": 6 + }, + { + "type": "table", + "img_path": "images/0b78202087eed264bedb97e4a51d3e55c7df39c6d85bb2e317abf2d47ab20bf9.jpg", + "table_caption": [ + "Table 2. AUC results (mean ± std) under the hard setting. The best and second-best results are highlighted in red and blue, respectively. Carpet and Metal_nut are subsets of MVTec AD. The datasets used are consistent with those in [9, 51], where those datasets only containing one anomaly class are excluded to adapt for the hard setting." + ], + "table_footnote": [], + "table_body": "
DatasetDevNetFLOSSAOEMLEPDRAAHLDPDL (Ours)
Ten Training Anomaly Examples
Carpet (mean)0.847±0.0170.761±0.0120.762±0.0730.751±0.0230.935±0.0130.949±0.0020.956±0.004
Metal_nut (mean)0.965±0.0110.922±0.0140.855±0.0160.878±0.0580.945±0.0170.972±0.0020.978±0.002
AITEX (mean)0.683±0.0320.635±0.0430.724±0.0320.626±0.0410.733±0.0090.747±0.0020.798±0.005
ELPV (mean)0.702±0.0230.642±0.0320.683±0.0470.745±0.0200.766±0.0290.788±0.0030.818±0.003
Mastcam (mean)0.588±0.0110.616±0.0210.697±0.0140.588±0.0160.695±0.0040.721±0.0030.778±0.007
Hyper-Kvasir (mean)0.822±0.0190.786±0.0210.698±0.0210.571±0.0140.844±0.0090.854±0.0040.864±0.002
One Training Anomaly Example
Carpet (mean)0.767±0.0180.678±0.0400.753±0.0550.679±0.0290.901±0.0060.932±0.0030.941±0.006
Metal_nut (mean)0.855±0.0160.855±0.0240.816±0.0290.825±0.0230.932±0.0170.939±0.0040.944±0.003
AITEX (mean)0.646±0.0340.624±0.0240.674±0.0340.466±0.0300.684±0.0330.707±0.0070.753±0.005
ELPV (mean)0.648±0.0570.691±0.0080.614±0.0480.566±0.1110.703±0.0220.740±0.0030.762±0.003
Mastcam (mean)0.511±0.0130.524±0.0130.689±0.0370.541±0.0070.667±0.0120.673±0.0100.733±0.004
Hyper-Kvasir (mean)0.595±0.0230.571±0.0040.406±0.0180.480±0.0440.700±0.0090.706±0.0070.715±0.004
", + "bbox": [ + 96, + 492, + 895, + 709 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/fdbc1f8c0719b28f4506c3c34a1813611aaaf11b904422766cffb697bb9c6471.jpg", + "image_caption": [ + "Ten Training Anomaly Examples Under General Settings", + "Figure 2. Ablation study for SB and DFL under the general settings and hard settings." + ], + "image_footnote": [], + "bbox": [ + 96, + 752, + 295, + 891 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/8f20e4f0e230387c2e81af323efd5c71bf9e0280f64c7a75c3fa256690116855.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 305, + 752, + 496, + 891 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/db3222376818448821419e66be31f82983244223a722cac6ff6f0daaab0cfddc.jpg", + "image_caption": [ + "Ten Training Anomaly Examples Under Hard Settings" + ], + "image_footnote": [], + "bbox": [ + 509, + 752, + 699, + 891 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/9d1cbc3ba700da9b6c1c48885b37b1cf35bee15aa1ef07965cd7f3ebe9ea9dd8.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 710, + 752, + 901, + 891 + ], + "page_idx": 6 + }, + { + "type": "page_number", + "text": "20422", + "bbox": [ + 478, + 945, + 519, + 955 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/52e48a8462c647629cc8063947bacd4faa66d0ff25e4680459bb4f0bcd62ff5f.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 96, + 88, + 300, + 200 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/636cde06dae3253814725a80d93b2120b14699832fb797ea88e5c32022f03ffd.jpg", + "image_caption": [ + "Figure 3. Parameter sensitivity analysis for $C,\\epsilon ,\\kappa$ and $\\lambda$" + ], + "image_footnote": [], + "bbox": [ + 303, + 88, + 500, + 200 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/617eecc1e1cbf38ca0cb98682474c27f4fb443f94a8e5f50cd1f9e03986f225c.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 504, + 88, + 700, + 200 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/dcb2a25d389bcdc8281164de9850e18bd07ed9e494f6909ce9364c3a95b93578.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 704, + 88, + 901, + 200 + ], + "page_idx": 7 + }, + { + "type": "table", + "img_path": "images/0804926702655fb01f91c4dd339e34e01f1d6a4f004b6c5f013c385847387dd0.jpg", + "table_caption": [ + "Table 3. An ablation study for ${\\mathrm{M}}_{\\mathrm{n}},{\\mathrm{M}}_{\\mathrm{a}}$ and ${\\mathrm{M}}_{\\mathrm{r}}$ ." + ], + "table_footnote": [], + "table_body": "
MnMaMrAITEXELPVMastcam
Ten Training Anomaly Examples Under General Settings
××✓0.928 ± 0.0190.914 ± 0.0210.899 ± 0.036
✓×✓0.939 ± 0.0230.919 ± 0.0110.908 ± 0.017
×✓✓0.963 ± 0.0080.930 ± 0.0130.924 ± 0.019
✓✓✓0.975 ± 0.0070.937 ± 0.0030.934 ± 0.010
Ten Training Anomaly Examples Under Hard Settings
××✓0.746 ± 0.0250.797 ± 0.0260.723 ± 0.017
✓×✓0.758 ± 0.0350.802 ± 0.0240.736 ± 0.028
×✓✓0.781 ± 0.0140.811 ± 0.0180.762 ± 0.021
✓✓✓0.798 ± 0.0050.818 ± 0.0030.778 ± 0.007
", + "bbox": [ + 98, + 237, + 475, + 416 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "5.6. Ablation Study", + "text_level": 1, + "bbox": [ + 89, + 422, + 243, + 438 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "The ablation study in Fig. 2 highlights the critical roles of the DPL and DFL components in improving DPDL's openset anomaly detection. We denote the variants that remove only DPL or DFL as 'DPDL w/o DPL' and 'DPDL w/o DFL', respectively. Compared to the full DPDL model, removing neither DPL nor DFL leads to a significant AUC decline, illustrating their critical utility. Specifically, 'DPDL w/o DPL' exhibits the most significant performance drop on industrial anomaly datasets, reflecting DPL's prominent role in learning precise and tight distribution boundaries for normal samples. By transforming normal samples into Gaussian distribution prototype space and pushing abnormal samples away, DPL enhances the recognition ability of anomaly samples. Meanwhile, ablation studies on 'DPDL w/o DFL' further highlight the critical role of the DFL component. By performing discreteness feature learning in hyperspherical space, DFL enhances the generalization ability to out-of-distribution anomalies.", + "bbox": [ + 88, + 444, + 482, + 715 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Additionally, the ablation experiments on $\\mathsf{M}_{\\mathrm{n}}$ , $\\mathsf{M}_{\\mathrm{a}}$ , and $\\mathsf{M}_{\\mathrm{r}}$ across three datasets in Tab. 3 reveal their varying contributions. DPDL shows the most significant performance drop when $\\mathsf{M}_{\\mathrm{r}}$ is removed, which illustrates that $\\mathsf{M}_{\\mathrm{r}}$ plays the most critical role in detecting anomalies. Furthermore, the removal of $\\mathsf{M}_{\\mathrm{a}}$ and $\\mathsf{M}_{\\mathrm{n}}$ lead to a noticeable performance drop in DPDL, which underscores their essential roles.", + "bbox": [ + 89, + 715, + 482, + 821 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "5.7. Parameter Sensitivity Analysis", + "text_level": 1, + "bbox": [ + 89, + 832, + 362, + 849 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Fig. 3 illustrates the results of the four hyperparameters under the general settings across nine datasets. Overall, the performance remains stable within a certain range of hyper", + "bbox": [ + 89, + 854, + 482, + 902 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "parameter variations, demonstrating the DPDL's robustness. Prototype quantity $C$ . We begin by investigating the critical impact of the number of initialized prototypes $C$ on distributed prototype learning in DPDL. We select $\\{8,16,32,64\\}$ as the values for the hyperparameters. As $C$ increases, DPDL's performance improves steadily, but excessively large values of $C$ hinder the model's effectiveness. This phenomenon is particularly pronounced on AI-TEX, Mastcam, MVTecAD, and Hyper-Kvasir, which contain more categories. One possible explanation is that a prototype space with too few prototypes loses discriminative information, while an excessively large number of prototypes reduces the compactness of the space.", + "bbox": [ + 511, + 223, + 903, + 419 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "DPL trajectory $\\epsilon$ in Eqns. (9), (10), (11) and (14). It can be observed that, particularly on the AITEX, Mastcam, and MVTecAD datasets with a larger number of categories, DPDL exhibits a relatively stable performance decline as $\\epsilon$ increases. As a crucial parameter in SB, $\\epsilon$ governs the trajectory state. Since smaller values produce straighter trajectories and larger values increase fluctuation, smaller $\\epsilon$ facilitates sampling more robust abstract prototypes from the relatively dispersed conditional distribution.", + "bbox": [ + 511, + 420, + 905, + 556 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "DFL tightness $\\kappa$ . According to Fig. 3, increasing $\\kappa$ generally enhances model performance, but values above $\\kappa = 10$ introduce negative effects in certain scenarios. A possible reason is that excessive sample dispersion makes it more challenging to tighten the normal distribution boundary.", + "bbox": [ + 511, + 556, + 905, + 632 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Loss parameter $\\lambda$ . We conduct a sensitivity analysis on the loss parameters $\\lambda$ . It can be observe that setting $\\lambda = 0.01$ achieves optimal performance on seven larger-scale datasets, while $\\lambda = 1$ yields the best results on two datasets with limited data. As $\\lambda$ increases, the performance declines, potentially due to gradient conflicts among the dispersion loss, SB transform loss and the main task loss.", + "bbox": [ + 511, + 633, + 905, + 738 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "6. Conclusion", + "text_level": 1, + "bbox": [ + 511, + 753, + 633, + 768 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "We propose Distribution Prototype Diffusion Learning (DPDL) for OSAD. DPDL leverages schrödinger bridge to map the normal distribution to a prototype space, simultaneously repelling anomalies to facilitate precise anomaly detection. We propose a dispersion feature learning way in hyperspherical space, which benefits the detection of out-of-distribution anomalies. Experimental results illustrate DPDL's robustness in diverse anomaly detection scenarios.", + "bbox": [ + 511, + 779, + 905, + 900 + ], + "page_idx": 7 + }, + { + "type": "page_number", + "text": "20423", + "bbox": [ + 478, + 944, + 517, + 955 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "7. Acknowledgments", + "text_level": 1, + "bbox": [ + 91, + 90, + 269, + 107 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "This work was supported by the National Natural Science Foundation of China (Grants No. 62476133), the National Natural Science Foundation of China (Nos. U22B2049), the National Natural Science Foundation of China (NSFC) under Grant 62272230, and the Natural Science Foundation of Shandong Province (Grant No. ZR2022LZH003).", + "bbox": [ + 89, + 113, + 483, + 198 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "References", + "text_level": 1, + "bbox": [ + 91, + 226, + 187, + 242 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[1] Andra Acsintoae, Andrei Florescu, Mariana-Iuliana Georgescu, Tudor Mare, Paul Sumedrea, Radu Tudor Ionescu, Fahad Shahbaz Khan, and Mubarak Shah. Ubnormal: New benchmark for supervised open-set video anomaly detection. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 20143-20153, 2022. 1, 2", + "[2] Aimira Baitieva, David Hurych, Victor Besnier, and Olivier Bernard. Supervised anomaly detection for complex industrial images. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 17754-17762, 2024. 1", + "[3] Paul Bergmann, Michael Fauser, David Sattlegger, and Carsten Steger. Mvtec ad-a comprehensive real-world dataset for unsupervised anomaly detection. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 9592-9600, 2019. 6", + "[4] Hanna Borgli, Vajira Thambawita, Pia H Smedsrud, Steven Hicks, Debesh Jha, Sigrun L Eskeland, Kristin Ranheim Randel, Konstantin Pogorelov, Mathias Lux, Duc Tien Dang Nguyen, et al. Hyperkvasir, a comprehensive multi-class image and video dataset for gastrointestinal endoscopy. Scientific data, 7(1):283, 2020. 6", + "[5] Yongxin Chen, Tryphon T Georgiou, and Michele Pavon. On the relation between optimal transport and schrödinger bridges: A stochastic control viewpoint. Journal of Optimization Theory and Applications, 169:671-691, 2016. 2", + "[6] Songmin Dai, Yifan Wu, Xiaogiang Li, and Xiangyang Xue. Generating and reweighting dense contrastive patterns for unsupervised anomaly detection. In Proceedings of the AAAI Conference on Artificial Intelligence, pages 1454-1462, 2024. 1", + "[7] Valentin De Bortoli, James Thornton, Jeremy Heng, and Arnaud Doucet. Diffusion schrödinger bridge with applications to score-based generative modeling. Advances in Neural Information Processing Systems, 34:17695-17709, 2021. 2", + "[8] Sergiu Deitsch, Vincent Christlein, Stephan Berger, Claudia Buerhop-Lutz, Andreas Maier, Florian Gallwitz, and Christian Riess. Automatic classification of defective photovoltaic module cells in electroluminescence images. Solar Energy, 185:455-468, 2019. 6", + "[9] Chouro Ding, Guansong Pang, and Chunhua Shen. Catching both gray and black swans: Open-set supervised anomaly detection. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 7388-7398, 2022. 1, 2, 5, 6, 7" + ], + "bbox": [ + 99, + 251, + 483, + 898 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[10] Nikita Gushchin, Sergei Kholkin, Evgeny Burnaev, and Alexander Korotin. Light and optimal schrödinger bridge matching. In *Forty-first International Conference on Machine Learning*, 2024. 2, 3, 4, 5", + "[11] Liren He, Zhengkai Jiang, Jinlong Peng, Liang Liu, Qianggang Du, Xiaobin Hu, Wenbing Zhu, Mingmin Chi, Yabiao Wang, and Chengjie Wang. Learning unified reference representation for unsupervised multi-class anomaly detection. arXiv preprint arXiv:2403.11561, 2024. 1", + "[12] Teng Hu, Jiangning Zhang, Ran Yi, Yuzhen Du, Xu Chen, Liang Liu, Yabiao Wang, and Chengjie Wang. Anomalydiffusion: Few-shot anomaly image generation with diffusion model. In Proceedings of the AAAI Conference on Artificial Intelligence, pages 8526-8534, 2024. 1", + "[13] Hannah R Kerner, Kiri L Wagstaff, Brian D Bue, Danika F Wellington, Samantha Jacob, Paul Horton, James F Bell, Chiman Kwan, and Heni Ben Amor. Comparison of novelty detection methods for multispectral images in rover-based planetary exploration missions. Data Mining and Knowledge Discovery, 34:1642-1675, 2020. 6", + "[14] Beomsu Kim, Gihyun Kwon, Kwanyoung Kim, and Jong Chul Ye. Unpaired image-to-image translation via neural schr\\” odinger bridge. arXiv preprint arXiv:2305.15086, 2023. 2", + "[15] Daehyun Kim, Sungyong Baik, and Tae Hyun Kim. Sanflow: Semantic-aware normalizing flow for anomaly detection. Advances in Neural Information Processing Systems, 36:75434-75454, 2023. 1", + "[16] Hyunsu Kim, Jongmin Yoon, and Juho Lee. Fast ensembling with diffusion schr\\” odinger bridge. arXiv preprint arXiv:2404.15814, 2024. 2", + "[17] Alexander Korotin, Nikita Gushchin, and Evgeny Burnaev. Light schr\\” odinger bridge. arXiv preprint arXiv:2310.01174, 2023. 2, 3, 4, 5", + "[18] Christian Léonard. A survey of the schr\\” odinger problem and some of its connections with optimal transport. arXiv preprint arXiv:1308.0215, 2013. 2", + "[19] Chun-Liang Li, Kihyuk Sohn, Jinsung Yoon, and Tomas Pfister. Cutpaste: Self-supervised learning for anomaly detection and localization. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 9664-9674, 2021. 1, 6", + "[20] Hanxi Li, Jingqi Wu, Hao Chen, Mingwen Wang, and Chunhua Shen. Efficient anomaly detection with budget annotation using semi-supervised residual transformer. arXiv preprint arXiv:2306.03492, 2023. 1", + "[21] Xiaofan Li, Zhizhong Zhang, Xin Tan, Chengwei Chen, Yanyun Qu, Yuan Xie, and Lizhuang Ma. Prompted: Learning prompts with only normal samples for few-shot anomaly detection. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 16838-16848, 2024. 1", + "[22] Jingyi Liao, Xun Xu, Manh Cuong Nguyen, Adam Goodge, and Chuan Sheng Foo. Coft-ad: Contrastive fine-tuning for few-shot anomaly detection. IEEE Transactions on Image Processing, 2024. 1", + "[23] Guan-Horng Liu, Tianrong Chen, Oswin So, and Evangelos Theodorou. Deep generalized schrödinger bridge. Advances" + ], + "bbox": [ + 516, + 92, + 905, + 901 + ], + "page_idx": 8 + }, + { + "type": "page_number", + "text": "20424", + "bbox": [ + 478, + 944, + 519, + 955 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "in Neural Information Processing Systems, 35:9374-9388, 2022. 2", + "[24] Guan-Horng Liu, Yaron Lipman, Maximilian Nickel, Brian Karrer, Evangelos A Theodorou, and Ricky TQ Chen. Generalized schr\\” odinger bridge matching. arXiv preprint arXiv:2310.02233, 2023. 2", + "[25] Guan-Horng Liu, Arash Vahdat, De-An Huang, Evangelos A Theodorou, Weili Nie, and Anima Anandkumar. $i^2$ -sb: Image-to-image schr\\” odinger bridge. arXiv preprint arXiv:2302.05872, 2023. 2", + "[26] Jiaqi Liu, Kai Wu, Qiang Nie, Ying Chen, Bin-Bin Gao, Yong Liu, Jinbao Wang, Chengjie Wang, and Feng Zheng. Unsupervised continual anomaly detection with contrastively-learned prompt. In Proceedings of the AAAI Conference on Artificial Intelligence, pages 3639–3647, 2024. 1", + "[27] Wen Liu, Weixin Luo, Zhengxin Li, Peilin Zhao, Shenghua Gao, et al. Margin learning embedded prediction for video anomaly detection with a few anomalies. In IJCAI, pages 023-3, 2019. 1, 2, 6", + "[28] Xinyue Liu, Jianyuan Wang, Biao Leng, and Shuo Zhang. Dual-modeling decouple distillation for unsupervised anomaly detection. arXiv preprint arXiv:2408.03888, 2024.1", + "[29] I Loshchilov. Decoupled weight decay regularization. arXiv preprint arXiv:1711.05101, 2017. 6", + "[30] Jialin Luo, Yuanzhi Wang, Ziqi Gu, Yide Qiu, Shuaizhen Yao, Fuyun Wang, Chunyan Xu, Wenhua Zhang, Dan Wang, and Zhen Cui. Mmm-rs: A multi-modal, multi-gsd, multiscene remote sensing dataset and benchmark for text-to-image generation. In The Thirty-eight Conference on Neural Information Processing Systems Datasets and Benchmarks Track. 2", + "[31] Kanti V Mardia and Peter E Jupp. Directional statistics. John Wiley & Sons, 2009. 5", + "[32] Amir Markovitz, Gilad Sharir, Itamar Friedman, Lihi Zelnik-Manor, and Shai Avidan. Graph embedded pose clustering for anomaly detection. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 10539-10547, 2020. 6", + "[33] Maxence Noble, Valentin De Bortoli, Arnaud Doucet, and Alain Durmus. Tree-based diffusion schrödinger bridge with applications to wasserstein barycenters. Advances in Neural Information Processing Systems, 36, 2024. 2", + "[34] Guansong Pang, Chunhua Shen, and Anton Van Den Hengel. Deep anomaly detection with deviation networks. In Proceedings of the 25th ACM SIGKDD international conference on knowledge discovery & data mining, pages 353-362, 2019. 1", + "[35] Guansong Pang, Chouro Ding, Chunhua Shen, and Anton van den Hengel. Explainable deep few-shot anomaly detection with deviation networks. arXiv preprint arXiv:2108.00462, 2021. 2, 5, 6", + "[36] T-YLPG Ross and GKHP Dólar. Focal loss for dense object detection. In proceedings of the IEEE conference on computer vision and pattern recognition, pages 2980-2988, 2017. 6" + ], + "bbox": [ + 91, + 90, + 482, + 898 + ], + "page_idx": 9 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[37] Mohammadreza Salehi, Niousha Sadjadi, Soroosh Baselizadeh, Mohammad H Rohban, and Hamid R Rabiee. Multiresolution knowledge distillation for anomaly detection. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 14902-14912, 2021. 6", + "[38] Yuyang Shi, Valentin De Bortoli, Andrew Campbell, and Arnaud Doucet. Diffusion schrödinger bridge matching. Advances in Neural Information Processing Systems, 36, 2024. 2", + "[39] Javier Silvestre-Blanes, Teresa Albero-Albero, Ignacio Miralles, Rubén Pérez-Llorens, and Jorge Moreno. A public fabric database for defect detection methods and results. *Autex Research Journal*, 19(4):363–374, 2019. 6", + "[40] Domen Tabernik, Samo Sela, Jure Skvarc, and Danijel Skocaj. Segmentation-based deep-learning approach for surface-defect detection. Journal of Intelligent Manufacturing, 31 (3):759-776, 2020. 6", + "[41] Jihoon Tack, Sangwoo Mo, Jongheon Jeong, and Jinwoo Shin. Csi: Novelty detection via contrastive learning on distributionally shifted instances. Advances in neural information processing systems, 33:11839-11852, 2020. 6", + "[42] Yuanzhi Wang, Zhen Cui, and Yong Li. Distribution-consistent modal recovering for incomplete multimodal learning. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 22025-22034, 2023. 2", + "[43] Yuanzhi Wang, Yong Li, and Zhen Cui. Incomplete multimodality-diffused emotion recognition. Advances in Neural Information Processing Systems, 36:17117-17128, 2023.", + "[44] Yuanzhi Wang, Yong Li, Mengyi Liu, Xiaoya Zhang, Xin Liu, Zhen Cui, and Antoni B Chan. Re-attentional controllable video diffusion editing. arXiv preprint arXiv:2412.11710, 2024.", + "[45] Yuanzhi Wang, Yong Li, Xiaoya Zhang, Xin Liu, Anbo Dai, Antoni B Chan, and Zhen Cui. Edit temporal-consistent videos with image diffusion model. ACM Transactions on Multimedia Computing, Communications and Applications, 20(12):1-16, 2024. 2", + "[46] Matthias Wieler and Tobias Hahn. Weakly supervised learning for industrial optical inspection. In DAGM symposium in, page 11, 2007. 6", + "[47] Xincheng Yao, Ruoqi Li, Jing Zhang, Jun Sun, and Chongyang Zhang. Explicit boundary guided semi-push-pull contrastive learning for supervised anomaly detection. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 24490-24499, 2023. 1, 2", + "[48] Xincheng Yao, Ruoqi Li, Zefeng Qian, Lu Wang, and Chongyang Zhang. Hierarchical gaussian mixture normalizing flow modeling for unified anomaly detection. arXiv preprint arXiv:2403.13349, 2024. 1", + "[49] Sangdoo Yun, Dongyoon Han, Seong Joon Oh, Sanghyuk Chun, Junsuk Choe, and Youngjoon Yoo. Cutmix: Regularization strategy to train strong classifiers with localizable features. In Proceedings of the IEEE/CVF international conference on computer vision, pages 6023-6032, 2019. 6" + ], + "bbox": [ + 516, + 90, + 903, + 900 + ], + "page_idx": 9 + }, + { + "type": "page_number", + "text": "20425", + "bbox": [ + 478, + 944, + 517, + 955 + ], + "page_idx": 9 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[50] Qihang Zhou, Guansong Pang, Yu Tian, Shibo He, and Jiming Chen. Anomalyclip: Object-agnostic prompt learning for zero-shot anomaly detection. arXiv preprint arXiv:2310.18961, 2023. 1", + "[51] Jiawen Zhu, Choubo Ding, Yu Tian, and Guansong Pang. Anomaly heterogeneity learning for open-set supervised anomaly detection. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 17616-17626, 2024. 1, 2, 6, 7", + "[52] Yuansheng Zhu, Wentao Bao, and Qi Yu. Towards open set video anomaly detection. In European Conference on Computer Vision, pages 395-412. Springer, 2022. 1, 2" + ], + "bbox": [ + 91, + 90, + 480, + 260 + ], + "page_idx": 10 + }, + { + "type": "page_number", + "text": "20426", + "bbox": [ + 478, + 945, + 517, + 955 + ], + "page_idx": 10 + } +] \ No newline at end of file diff --git a/2025/Distribution Prototype Diffusion Learning for Open-set Supervised Anomaly Detection/5f6492f6-61fa-40e8-b92f-42fc55a187bf_model.json b/2025/Distribution Prototype Diffusion Learning for Open-set Supervised Anomaly Detection/5f6492f6-61fa-40e8-b92f-42fc55a187bf_model.json new file mode 100644 index 0000000000000000000000000000000000000000..ed1d1689b53afe041bd02886fb23a7f563034ab0 --- /dev/null +++ b/2025/Distribution Prototype Diffusion Learning for Open-set Supervised Anomaly Detection/5f6492f6-61fa-40e8-b92f-42fc55a187bf_model.json @@ -0,0 +1,2279 @@ +[ + [ + { + "type": "header", + "bbox": [ + 0.107, + 0.003, + 0.182, + 0.043 + ], + "angle": 0, + "content": "CVF" + }, + { + "type": "header", + "bbox": [ + 0.237, + 0.001, + 0.812, + 0.047 + ], + "angle": 0, + "content": "This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore." + }, + { + "type": "title", + "bbox": [ + 0.163, + 0.13, + 0.838, + 0.176 + ], + "angle": 0, + "content": "Distribution Prototype Diffusion Learning for Open-set Supervised Anomaly Detection" + }, + { + "type": "text", + "bbox": [ + 0.152, + 0.203, + 0.852, + 0.293 + ], + "angle": 0, + "content": "Fuyun Wang\\(^{1}\\), Tong Zhang\\(^{1*}\\), Yuanzhi Wang\\(^{1}\\), Yide Qiu\\(^{1}\\), Xin Liu\\(^{2}\\), Xu Guo\\(^{1}\\), Zhen Cui\\(^{3*}\\) \n\\(^{1}\\)Nanjing University of Science and Technology, China. \n\\(^{2}\\)Nanjing SeetaCloud Technology, China. \\(^{3}\\)Beijing Normal University, China \n{fyw271828, tong.zhang, yuanzhiwang, 121106010824, guo.xu} @njust.edu.cn \nxin.liu@seetacloud.com, zhen.cui@bnu.edu.cn" + }, + { + "type": "title", + "bbox": [ + 0.248, + 0.327, + 0.327, + 0.344 + ], + "angle": 0, + "content": "Abstract" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.359, + 0.486, + 0.648 + ], + "angle": 0, + "content": "In Open-set Supervised Anomaly Detection (OSAD), the existing methods typically generate pseudo anomalies to compensate for the scarcity of observed anomaly samples, while overlooking critical priors of normal samples, leading to less effective discriminative boundaries. To address this issue, we propose a Distribution Prototype Diffusion Learning (DPDL) method aimed at enclosing normal samples within a compact and discriminative distribution space. Specifically, we construct multiple learnable Gaussian prototypes to create a latent representation space for abundant and diverse normal samples and learn a Schrödinger bridge to facilitate a diffusive transition toward these prototypes for normal samples while steering anomaly samples away. Moreover, to enhance inter-sample separation, we design a dispersion feature learning way in hyperspherical space, which benefits the identification of out-of-distribution anomalies. Experimental results demonstrate the effectiveness and superiority of our proposed DPDL, achieving state-of-the-art performance on 9 public datasets." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.667, + 0.222, + 0.684 + ], + "angle": 0, + "content": "1. Introduction" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.694, + 0.485, + 0.875 + ], + "angle": 0, + "content": "Anomaly detection (AD) [15, 21, 48, 50] aims to identify outliers significantly diverging from the prevailing samples in a dataset, and has a wide range of applications like industrial inspection, medical image analysis, and scientific discovery, etc. Recently, unsupervised anomaly detection (UAD) [6, 11, 26, 28] and few-shot anomaly detection (FSAD) [12, 21, 22] have emerged as prominent research paradigms, emphasizing the modeling of normal sample distributions to discern anomalies effectively. Yet, these methods often neglect prior knowledge from limited anomaly samples, resulting in imprecise delineation of normal sample boundaries and reduced efficacy in differenti-" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.329, + 0.906, + 0.42 + ], + "angle": 0, + "content": "ating normal from anomaly instances. On the contrary, supervised anomaly detection (SAD) [2, 20, 47] leverages a limited subset of anomaly samples as prior knowledge, improving detection performance. However, this reliance on seen anomalies poses a risk of overfitting and hampers generalization to unseen anomalies in real-world settings." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.426, + 0.909, + 0.669 + ], + "angle": 0, + "content": "To mitigate the challenge of limited generalization inherent in closed-set training, we focus on open-set supervised anomaly detection (OSAD) [1, 27, 51, 52], which utilizes a small set of known anomaly classes during training to identify unseen anomalies from open-set classes. By leveraging prior knowledge from observed samples, OSAD methods could reduce false positive errors. To improve the generalized detection of unseen anomalies, DRA [9] leverages data augmentation and outlier exposure to learn a decomposed anomaly representation comprising seen anomalies, pseudo-anomalies, and potential residual anomalies. BGAD [47] leverages decision boundaries derived from normalized flow models to capture and model anomaly information. Recently, AHL [51] simulates heterogeneous anomaly distributions and performs collaborative differentiable learning to further enhance the model's generality." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.675, + 0.911, + 0.903 + ], + "angle": 0, + "content": "While data augmentation and outlier exposure techniques [9, 19, 34] have demonstrated considerable success in anomaly detection, they fall short in generating comprehensive pseudo anomalies, capturing only a fraction of potential unseen anomalies. This limitation arises from overlooking the intricate nature of real-world anomaly distributions, thereby hindering the model's ability to generalize to novel anomaly types. Although AHL [51] has made strides in addressing this issue by simulating heterogeneous anomaly distributions, it still relies on approximating unknown out-of-distribution anomalies using known indistribution anomalies for generalization. Yet, three critical issues persist: i) the simulation mechanism cannot cover all anomaly distribution patterns due to the varied scales and structures of anomaly distributions; ii) simulated anom" + }, + { + "type": "page_footnote", + "bbox": [ + 0.109, + 0.887, + 0.366, + 0.902 + ], + "angle": 0, + "content": "*Corresponding authors: Tong Zhang, Zhen Cui" + }, + { + "type": "page_number", + "bbox": [ + 0.479, + 0.945, + 0.521, + 0.958 + ], + "angle": 0, + "content": "20416" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.089, + 0.092, + 0.482, + 0.242 + ], + "angle": 0, + "content": "lies inherit in-distribution data biases, leading to suboptimal performance on out-of-distribution anomalies; iii) the diversity of normal samples presents dilemmas for existing methods, complicating the differentiation between normal and anomaly boundaries. These issues prompt a fundamental question: Instead of generating pseudo and uncertain anomaly samples, how can we accurately characterize compact distribution boundaries amidst a range of diverse normal samples and achieve robust generalization for unknown out-of-distribution anomalies?" + }, + { + "type": "text", + "bbox": [ + 0.094, + 0.244, + 0.485, + 0.65 + ], + "angle": 0, + "content": "To address the aforementioned issue, in this work, we propose a Distribution Prototype Diffusion Learning (DPDL) method for open-set supervised anomaly detection. Considering abundant and diverse normal samples but very limited anomaly data, our method involves learning latent distribution prototypes, specifically multiple Gaussian distributions, onto which all observed normal samples can be effectively projected. To facilitate the mapping of normal samples into the prototype space, we leverage Schrödinger bridge (SB) framework, which enables a diffusive transition by aligning the distributions of these samples with the prototypes. The SB-based diffusion way could mitigate the out-of-distribution issue for normal samples to some extent. Within the distribution prototype space, we push observed anomaly samples away from normal samples to enhance discriminative capacity. Notably, both the prototypes and the diffusive bridge are learned jointly, resulting in a robust embedding space for normal samples. Moreover, to enhance generalization across unseen anomaly domains, we introduce a dispersion feature learning mechanism that maps intermediate features to a hyperspherical space, leveraging a mixture of von Mises-Fisher (vMF) distributions. This approach bolsters directional feature extraction and promotes robust inter-sample separation, facilitating effective identification of out-of-distribution samples. Experiments demonstrate that our method greatly improves detection capabilities for unseen anomalies." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.651, + 0.483, + 0.786 + ], + "angle": 0, + "content": "In summary, our contributions are three-fold: i) we propose a distribution prototype diffusion learning framework that jointly learns multiple Gaussian prototypes and the associated diffusion bridge, creating a compact and discriminative embedding space; ii) we develop dispersion feature learning in hyperspherical space to enhance inter-sample separation and improve generalization; iii) we achieve state-of-the-art performance in 9 public datasets, demonstrating the efficacy of our approach." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.8, + 0.233, + 0.815 + ], + "angle": 0, + "content": "2. Related Work" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.825, + 0.483, + 0.902 + ], + "angle": 0, + "content": "Open-set Supervised Anomaly Detection. Open-set supervised anomaly detection (OSAD) seeks to develop a robust anomaly detection framework that generalizes from a limited set of training anomalies to effectively identify previously unseen anomalies within an open-set con" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.092, + 0.905, + 0.242 + ], + "angle": 0, + "content": "text [1, 27, 35, 47, 52]. Leveraging the prior knowledge provided by observed anomalies, contemporary OSAD methods significantly mitigate false positive errors, thereby enhancing overall detection performance [9, 51]. Recently, DRA [9] learns disentangled representations of observed, pseudo, and residual anomalies to boost the detection of both seen and unseen anomalies. In contrast, AHL [51] simulates diverse heterogeneous anomaly distributions and employs collaborative differentiable learning, significantly improving the model's generalization capacity." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.243, + 0.907, + 0.47 + ], + "angle": 0, + "content": "Schrödinger Bridge. Schrödinger bridge (SB), widely recognized as the entropy-regularized optimal transport (OT) problem, involves learning a stochastic process that evolves from an initial probability distribution to a target distribution under the influence of a reference measure [5, 7, 16, 18, 23, 24, 38], which is often used for content generation as those prior generative models do [30, 42-45]. \\( \\mathrm{I}^2 \\)-SB [25] and UNSB [14] learn a nonlinear diffusion process between two given distributions or represent the SB problem as a series of adversarial learning problems to realize the image transformation task. Recently, LightSB [10, 17] introduces a novel, fast, and simple SB solver, which achieves optimal matching in practice through the Gaussian mixture parameterization of the adjusted schrödinger potential." + }, + { + "type": "title", + "bbox": [ + 0.514, + 0.482, + 0.652, + 0.497 + ], + "angle": 0, + "content": "3. Preliminaries" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.507, + 0.905, + 0.627 + ], + "angle": 0, + "content": "We focus on how to build the connection between two distributions \\( p_0 \\) and \\( p_1 \\), where the distributions are defined as absolutely continuous Borel probability distributions with finite second-order moments. Building upon the foundation of entropy-regularized optimal transport (EOT) [5, 7, 10, 16-18, 23, 24, 33, 38], we review the related properties of EOT and the schrödinger bridge (SB) problem with a Wiener prior." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.628, + 0.905, + 0.704 + ], + "angle": 0, + "content": "Entropy-regularized optimal transport (EOT). Given two point sets \\(\\mathcal{Z}_0\\) and \\(\\mathcal{Z}_1\\), we seek for the optimal transport cost between any two points \\(z_0 \\in \\mathcal{Z}_0\\) and \\(z_1 \\in \\mathcal{Z}_1\\). This task may be formulated as an EOT problem with a parameter \\(\\epsilon > 0\\), i.e., minimizing the following objective:" + }, + { + "type": "equation", + "bbox": [ + 0.552, + 0.711, + 0.905, + 0.772 + ], + "angle": 0, + "content": "\\[\n\\begin{array}{l} \\min _ {\\pi \\in \\Pi \\left(p _ {0}, p _ {1}\\right)} \\left\\{\\int_ {\\mathbb {R} ^ {D}} \\int_ {\\mathbb {R} ^ {D}} \\frac {1}{2} \\| z _ {0} - z _ {1} \\| ^ {2} \\pi \\left(z _ {0}, z _ {1}\\right) d z _ {0} d z _ {1} \\right. \\tag {1} \\\\ + \\epsilon \\mathrm {K L} (z \\| p _ {0} \\times p _ {1}) \\}, \\\\ \\end{array}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.78, + 0.905, + 0.87 + ], + "angle": 0, + "content": "where \\(\\Pi(p_0, p_1)\\) denotes the set of transport plans, i.e., joint probability distributions on \\(\\mathbb{R}^D \\times \\mathbb{R}^D\\) with marginals \\(p_0\\) and \\(p_1\\), respectively, and KL denotes Kullback-Leibler divergence. The minimizer \\(\\pi^*\\) of Eqn. (1) is guaranteed to exist, be unique, and absolutely continuous, and is referred as the EOT plan." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.871, + 0.905, + 0.902 + ], + "angle": 0, + "content": "Schrodinger Bridge (SB). We define \\(\\Omega\\) as the space of \\(\\mathbb{R}^D\\)-valued functions over time \\(t \\in [0,1]\\), representing" + }, + { + "type": "page_number", + "bbox": [ + 0.479, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "20417" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.09, + 0.091, + 0.482, + 0.212 + ], + "angle": 0, + "content": "trajectories in \\(\\mathbb{R}^D\\) that start at \\(t = 0\\) and end at \\(t = 1\\). We denote the set of probability distributions over \\(\\Omega\\), i.e., stochastic processes, by \\(\\mathcal{P}(\\Omega)\\). The differential of the standard Wiener process is represented by \\(dW_{t}\\). For a process \\(T\\in \\mathcal{P}(\\Omega)\\), we denote its joint distribution at \\(t = 0,1\\) by \\(\\pi_T\\in \\mathcal{P}(\\mathbb{R}^D\\times \\mathbb{R}^D)\\). Similarly, we use \\(T|_{z_0,z_1}\\) to denote the distribution of \\(T\\) for \\(t\\in (0,1)\\), conditioned on \\(T\\)'s values \\(z_{0}\\) and \\(z_{1}\\) at \\(t = 0\\) and \\(t = 1\\), respectively." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.212, + 0.482, + 0.288 + ], + "angle": 0, + "content": "Let \\( W^{\\epsilon} \\in P(\\Omega) \\) represents a Wiener process with volatility \\( \\epsilon > 0 \\), starting from \\( p_0 \\) at \\( t = 0 \\). Its differential is governed by the stochastic differential equation (SDE): \\( dW_{t}^{\\epsilon} = \\sqrt{\\epsilon} dW_{t} \\). The Schrödinger bridge problem with the Wiener prior \\( W^{\\epsilon} \\) between \\( p_0 \\) and \\( p_1 \\) is minimizing:" + }, + { + "type": "equation", + "bbox": [ + 0.202, + 0.298, + 0.482, + 0.324 + ], + "angle": 0, + "content": "\\[\n\\min _ {T \\in \\mathcal {F} \\left(p _ {0}, p _ {1}\\right)} \\operatorname {K L} \\left(T \\| W ^ {\\epsilon}\\right), \\tag {2}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.333, + 0.483, + 0.422 + ], + "angle": 0, + "content": "where \\(\\mathcal{F}(p_0,p_1)\\subset \\mathcal{P}(\\Omega)\\) denotes the subset of stochastic processes that begin with distribution \\(p_0\\) at \\(t = 0\\) and reach \\(p_1\\) at \\(t = 1\\). This problem has a unique solution, a SDE diffusion process \\(T^{*}\\) defined: \\(dZ_{t} = g^{*}(Z_{t},t)dt + dW_{t}^{\\epsilon}\\). The process \\(T^{*}\\) is referred to as SB, and \\(g^{*}:\\mathbb{R}^{D}\\times [0,1]\\to \\mathbb{R}^{D}\\) is the optimal drift." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.424, + 0.483, + 0.455 + ], + "angle": 0, + "content": "Characterization of solutions. The EOT plan \\(\\pi^{*} = \\pi^{T^{*}}\\) takes a specific form [10, 17]:" + }, + { + "type": "equation", + "bbox": [ + 0.13, + 0.463, + 0.482, + 0.495 + ], + "angle": 0, + "content": "\\[\n\\pi^ {*} \\left(z _ {0}, z _ {1}\\right) = u ^ {*} \\left(z _ {0}\\right) \\exp \\left(\\frac {- \\| z _ {0} - z _ {1} \\| ^ {2}}{2 \\epsilon}\\right) v ^ {*} \\left(z _ {1}\\right), \\tag {3}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.502, + 0.482, + 0.533 + ], + "angle": 0, + "content": "where \\( u^{*}, v^{*} \\colon \\mathbb{R} \\to \\mathbb{R}_{+} \\) are measurable functions known as Schrödinger potentials. The optimal drift \\( g^{*} \\) is derived as:" + }, + { + "type": "equation", + "bbox": [ + 0.099, + 0.542, + 0.483, + 0.575 + ], + "angle": 0, + "content": "\\[\ng ^ {*} (z, t) = \\epsilon \\nabla_ {z} \\log \\int_ {\\mathbb {R} ^ {D}} \\mathcal {N} \\left(z ^ {\\prime} \\mid z, (1 - t) \\epsilon I _ {D}\\right) v ^ {*} \\left(z ^ {\\prime}\\right) d z ^ {\\prime}, \\tag {4}\n\\]" + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.583, + 0.182, + 0.598 + ], + "angle": 0, + "content": "4. Method" + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.607, + 0.295, + 0.622 + ], + "angle": 0, + "content": "4.1. Problem Formulation" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.629, + 0.483, + 0.84 + ], + "angle": 0, + "content": "Let \\(\\mathcal{X}_{\\mathrm{tr}} = \\{(\\mathbf{x}_i, y_i)\\}\\) denotes a weakly-supervised training set with only image-level labels, where \\(\\mathbf{x}_i\\) denotes one RGB image and \\(y_i \\in \\{0,1\\}\\) denotes whether \\(\\mathbf{x}_i\\) is an anomaly sample (anomaly: \\(y_i = 1\\), normal: \\(y_i = 0\\)). Hereby, \\(\\mathcal{X}_{\\mathrm{tr}}\\) is consistent of a normal subset \\(\\mathcal{X}_{\\mathrm{tr}}^{\\mathrm{n}}\\) (\\(|\\mathcal{X}_{\\mathrm{tr}}^{\\mathrm{n}}| = N\\)) and an anomaly subset \\(\\mathcal{X}_{\\mathrm{tr}}^{\\mathrm{a}}\\) (\\(|\\mathcal{X}_{\\mathrm{tr}}^{\\mathrm{a}}| = M\\)), formally, \\(\\mathcal{X}_{\\mathrm{tr}} \\doteq \\mathcal{X}_{\\mathrm{tr}}^{\\mathrm{n}} \\cup \\mathcal{X}_{\\mathrm{tr}}^{\\mathrm{a}}\\), where generally \\(N \\gg M\\). Given a testing set \\(\\mathcal{X}_{\\mathrm{te}}\\), we need to predict whether one sample \\(\\mathbf{x} \\in \\mathcal{X}_{\\mathrm{te}}\\) is anomaly or normal. In OSAD, the anomaly patterns of the testing set do less recurred in those encountered training set. In other words, the distribution of anomalies are obviously discrepant, i.e., \\(P(\\mathcal{X}_{\\mathrm{te}}^{\\mathrm{a}}) \\neq P(\\mathcal{X}_{\\mathrm{tr}}^{\\mathrm{a}})\\). Hereby, we need to learn a robust anomaly detection model \\(\\psi\\) from the training set \\(\\mathcal{X}_{\\mathrm{tr}}\\) so that \\(\\psi\\) accurately infers anomaly scores for test samples." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.841, + 0.483, + 0.901 + ], + "angle": 0, + "content": "Our abstract idea is to learn latent distribution prototypes that not only encapsulate normal samples in a concise manner but also discriminate against anomaly samples. Given the abundance of observed normal samples, one" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.092, + 0.905, + 0.228 + ], + "angle": 0, + "content": "natural approach is to characterize the distribution \\(P(\\mathcal{X}_{\\mathrm{tr}}^{\\mathrm{n}})\\) where samples outside this distribution, \\(\\mathbf{x}\\notin P(\\mathcal{X}_{\\mathrm{tr}}^{\\mathrm{n}})\\), would be awarded higher probabilities as anomalies. Considering the inherent diversity of normal samples, we endeavor to learn multiple simple distributions (e.g., Gaussians) as prototypes \\(\\mathcal{P}_{\\mathrm{MGP}}\\), named multi-Gaussian prototypes (MGP). To embed input data into the prototype space, we introduce a generative bridge model \\(\\psi_{p}\\) for distribution transformation. The more abstract formulation is given as follows:" + }, + { + "type": "equation", + "bbox": [ + 0.517, + 0.234, + 0.905, + 0.259 + ], + "angle": 0, + "content": "\\[\n\\min _ {\\psi_ {p}, \\mathcal {P} _ {\\mathrm {M G P}}, f} D _ {p} \\left(\\psi_ {p} \\left(\\mathcal {F} _ {\\mathrm {t r}} ^ {\\mathrm {n}}\\right), \\psi_ {p} \\left(\\mathcal {F} _ {\\mathrm {t r}} ^ {\\mathrm {a}}\\right), \\mathcal {P} _ {\\mathrm {M G P}}\\right) + \\lambda D _ {s} \\left(\\mathcal {F} _ {\\mathrm {t r}} ^ {\\mathrm {n}}, \\mathcal {F} _ {\\mathrm {t r}} ^ {\\mathrm {a}}\\right), \\tag {5}\n\\]" + }, + { + "type": "equation", + "bbox": [ + 0.548, + 0.262, + 0.905, + 0.284 + ], + "angle": 0, + "content": "\\[\ns. t., \\psi_ {p}: P (\\mathcal {F}) \\stackrel {\\text {b r i d g e}} {\\Longrightarrow} \\mathcal {P} _ {\\mathrm {M G P}}, \\tag {6}\n\\]" + }, + { + "type": "equation", + "bbox": [ + 0.581, + 0.287, + 0.905, + 0.307 + ], + "angle": 0, + "content": "\\[\nf: \\mathcal {X} \\xrightarrow {\\text {f e a t u r e}} \\mathcal {F}, \\tag {7}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.313, + 0.905, + 0.51 + ], + "angle": 0, + "content": "where \\(\\mathcal{P}_{\\mathrm{MGP}}\\) denotes the distribution prototypes to be learned, \\(\\psi_{p}\\) is the flow function across probability distributions, \\(D_{p}\\) signifies the discriminative function in the space of prototypes, \\(f\\) stands for the feature extraction process, and \\(D_{s}\\) acts as a regularizer to increase feature discriminability. In the above formulation, besides distribution prototype learning (DPL), we also introduce dispersion feature learning (DFL) executed within a hyperspherical embedding space, i.e., \\(D_{s}(\\mathcal{F}_{\\mathrm{tr}}^{\\mathrm{n}},\\mathcal{F}_{\\mathrm{tr}}^{\\mathrm{a}})\\). The advantage of DFL is to prevent abrupt feature collapse in feature learning, thus preserving discriminative qualities. The overview of our framework is abstractly depicted in Fig. 1. Detailed elaboration on these aspects will be provided in the subsequent sections." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.517, + 0.8, + 0.533 + ], + "angle": 0, + "content": "4.2. Distribution Prototype Learning" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.539, + 0.906, + 0.689 + ], + "angle": 0, + "content": "Distribution prototype learning is operated in the feature space. Hence, to extract intermediate image features, we can leverage those classic networks as the backbone, such as ResNet-18. Formally, the feature extraction process, denoted as a function \\( f: \\mathcal{X} \\to \\mathcal{F} \\), transforms an image \\( \\mathbf{x} \\) into the intermediate feature \\( \\mathbf{x} = f(\\mathbf{x}) \\in \\mathbb{R}^d \\), where a flatten operation is used on convolution maps for vectorization. Consequently, we designate the intermediate feature sets of normal and anomaly samples as \\( \\mathcal{F}_{\\mathrm{tr}}^{\\mathrm{n}} = \\{\\mathbf{x}_i^{\\mathrm{n}} | i = 1, \\dots, N\\} \\) and \\( \\mathcal{F}_{\\mathrm{tr}}^{\\mathrm{a}} = \\{\\mathbf{x}_j^{\\mathrm{a}} | i = 1, \\dots, M\\} \\), respectively." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.69, + 0.906, + 0.901 + ], + "angle": 0, + "content": "Considering the rarity of anomalies and the diversity of normal samples, it is pertinent to capture the intricate distributions of normal samples by mapping the distribution \\( p_0 = P(\\mathcal{F}_{\\mathrm{tr}}^{\\mathrm{n}}) \\) of intermediate features to a distinct and well-characterized distribution \\( p_1 \\) like Gaussian. But due to various normal samples, we opt for a multi-Gaussian prototypes (MGP) comprising multiple Gaussian distributions as prototypes \\( \\mathcal{P}_{\\mathrm{MGP}} = \\{\\mathcal{P}_i \\doteq \\mathcal{N}(\\pmb{\\mu}_i, \\pmb{\\sigma}_i) | i = 1, \\dots, C\\} \\). Given a normal sample \\( \\mathbf{x}_i^{\\mathrm{n}} \\), we expect to align it with the closest prototype with high likelihood. However, in the open-set setting, it is challenge to transform unseen points \\( \\mathbf{x} \\sim p_0 \\) to the target distribution \\( p_1 \\). Drawing inspiration from the capability of diffusion generation models in aligning disparate distributions, we frame the transition from source" + }, + { + "type": "page_number", + "bbox": [ + 0.48, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "20418" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.101, + 0.086, + 0.899, + 0.328 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.089, + 0.338, + 0.908, + 0.425 + ], + "angle": 0, + "content": "Figure 1. Our proposed DPDL framework. It comprises three distinct modules: Distribution Prototype Learning (DPL, Sec. 4.2), Dispersion Feature Learning (DFL, Sec. 4.3), and anomaly score prediction (Sec. 4.4). DPL transforms the distribution of normal samples to a space of learnable multiple Gaussian prototypes through building schrödinger bridge, meantime pushing anomaly distribution away from these prototypes. DFL operates in a hyperspherical space, enlarging the distances of intermediate features of all samples in a hyperspherical space to strengthen feature generalization for detecting anomalies. The score prediction module leverages a multi-instance-learning method to compute anomaly scores." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.436, + 0.484, + 0.519 + ], + "angle": 0, + "content": "domain distribution \\(p_0\\) to target domain distribution \\(p_1\\) as a Schrodinger bridge problem. As formulated in Eqn. (6), we need learn an essential bridge flow \\(\\psi_p: P(\\mathcal{F}_{\\mathrm{tr}}^{\\mathrm{n}}) \\xrightarrow{\\mathrm{bridge}} \\mathcal{P}_{\\mathrm{MGP}}\\). After bridge transformation, the condition probability in optimal transport plan conforms to:" + }, + { + "type": "equation", + "bbox": [ + 0.147, + 0.522, + 0.483, + 0.587 + ], + "angle": 0, + "content": "\\[\n\\pi \\left(\\psi_ {p} (\\mathbf {x}) | \\mathbf {x}\\right) \\propto \\underbrace {\\sum_ {c = 1} ^ {C} \\alpha_ {c} \\mathcal {N} \\left(\\psi_ {p} (\\mathbf {x}) ; \\boldsymbol {\\mu} _ {c} , \\boldsymbol {\\sigma} _ {c}\\right)} _ {\\dot {\\equiv} \\phi_ {1} \\left(\\psi_ {p} (\\mathbf {x})\\right)}, \\tag {8}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.593, + 0.483, + 0.654 + ], + "angle": 0, + "content": "where the parameters \\(\\{\\alpha_c,\\pmb {\\mu}_c,\\pmb {\\sigma}_c\\}_{c = 1}^C\\) (known as Gaussian mixed model (GMM), abstracted into the function \\(\\phi_1\\) ) as well as the flow function \\(\\psi_{p}\\) (in bridge) need to be learned. For simplicity, below we adopt diagonal matrices for \\(\\pmb{\\sigma}_{c}\\)" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.654, + 0.484, + 0.745 + ], + "angle": 0, + "content": "Drawing inspiration from previous works [10, 17], we reframe distribution prototype learning in Eqn. (8) as the Schrödinger bridge. Given the specific form taken by the EOT plan described in Eqn. (3), we redefine these measurable functions \\( u, v: \\mathbb{R} \\to \\mathbb{R}_+ \\), termed Schrödinger potentials [10, 17], as follows:" + }, + { + "type": "equation", + "bbox": [ + 0.174, + 0.749, + 0.483, + 0.781 + ], + "angle": 0, + "content": "\\[\nu \\left(\\mathbf {x} _ {i} ^ {\\mathrm {n}}\\right) \\doteq \\exp \\left(\\frac {\\left\\| \\mathbf {x} _ {i} ^ {\\mathrm {n}} \\right\\| ^ {2}}{2 \\epsilon}\\right) \\phi_ {0} \\left(\\mathbf {x} _ {i} ^ {\\mathrm {n}}\\right), \\tag {9}\n\\]" + }, + { + "type": "equation", + "bbox": [ + 0.152, + 0.782, + 0.483, + 0.814 + ], + "angle": 0, + "content": "\\[\nv \\left(\\psi \\left(\\mathbf {x} _ {i} ^ {\\mathrm {n}}\\right)\\right) \\doteq \\exp \\left(\\frac {\\| \\psi \\left(\\mathbf {x} _ {i} ^ {\\mathrm {n}}\\right) \\| ^ {2}}{2 \\epsilon}\\right) \\phi_ {1} \\left(\\psi \\left(\\mathbf {x} _ {i} ^ {\\mathrm {n}}\\right)\\right), \\tag {10}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.817, + 0.483, + 0.864 + ], + "angle": 0, + "content": "where \\(\\phi_0\\) is defined within the source feature domain \\(\\mathcal{F}_{\\mathrm{tr}}\\), and \\(\\phi_{1}\\) is defined in Eqn. (8). \\(\\epsilon\\) is set to 0.001 in our experiments. Accordingly, Eqn. (3) can be converted to:" + }, + { + "type": "equation", + "bbox": [ + 0.09, + 0.868, + 0.484, + 0.9 + ], + "angle": 0, + "content": "\\[\n\\pi \\left(\\mathbf {x} _ {i} ^ {\\mathrm {n}}, \\psi \\left(\\mathbf {x} _ {i} ^ {\\mathrm {n}}\\right)\\right) = \\phi_ {0} \\left(\\mathbf {x} _ {i} ^ {\\mathrm {n}}\\right) \\exp \\left(\\frac {\\left\\langle \\mathbf {x} _ {i} ^ {\\mathrm {n}} , \\psi \\left(\\mathbf {x} _ {i} ^ {\\mathrm {n}}\\right) \\right\\rangle}{\\epsilon}\\right) \\phi_ {1} \\left(\\psi \\left(\\mathbf {x} _ {i} ^ {\\mathrm {n}}\\right)\\right), \\tag {11}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.436, + 0.907, + 0.467 + ], + "angle": 0, + "content": "Hence, the condition probability of the transport plan in Eqn. (8) could be exactly defined as" + }, + { + "type": "equation", + "bbox": [ + 0.571, + 0.476, + 0.907, + 0.495 + ], + "angle": 0, + "content": "\\[\n\\pi (\\psi (\\mathbf {x} _ {i} ^ {\\mathrm {n}}) | \\mathbf {x} _ {i} ^ {\\mathrm {n}}) \\doteq \\eta (\\mathbf {x} _ {i} ^ {\\mathrm {n}}, \\psi (\\mathbf {x} _ {i} ^ {\\mathrm {n}})) \\phi_ {1} (\\psi (\\mathbf {x} _ {i} ^ {\\mathrm {n}})), \\tag {12}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.503, + 0.907, + 0.585 + ], + "angle": 0, + "content": "where the connection factor is denoted as \\(\\eta (\\mathbf{x}_i^n,\\psi (\\mathbf{x}_i^n)) = \\frac{1}{\\varpi(\\mathbf{x}_i^n)}\\exp \\left(\\frac{\\langle\\mathbf{x}_i^n,\\psi(\\mathbf{x}_i^n)\\rangle}{\\epsilon}\\right)\\) with the normalization term \\(\\varpi (\\mathbf{x}_i^n) = \\int \\exp (\\langle \\mathbf{x}_i^n,\\psi (\\mathbf{x}_i^n)\\rangle)\\phi_1(\\psi (\\mathbf{x}_i^n))d\\psi (\\mathbf{x}_i^n)\\). According to Eqns. (8) and (12), we can further derive a more tractable form [10, 17]:" + }, + { + "type": "equation", + "bbox": [ + 0.515, + 0.594, + 0.906, + 0.636 + ], + "angle": 0, + "content": "\\[\n\\pi \\left(\\psi \\left(\\mathbf {x} _ {i} ^ {\\mathrm {n}}\\right) \\mid \\mathbf {x} _ {i} ^ {\\mathrm {n}}\\right) = \\widetilde {\\eta} \\left(\\mathbf {x} _ {i} ^ {\\mathrm {n}}\\right) \\sum_ {c = 1} ^ {C} \\widetilde {\\alpha} _ {c} \\left(\\mathbf {x} _ {i} ^ {\\mathrm {n}}\\right) \\mathcal {N} \\left(\\psi \\left(\\mathbf {x} _ {i} ^ {\\mathrm {n}}\\right); \\widetilde {\\boldsymbol {\\mu}} _ {c} \\left(\\mathbf {x} _ {i} ^ {\\mathrm {n}}\\right), \\boldsymbol {\\sigma} _ {c}\\right), \\tag {13}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.645, + 0.906, + 0.708 + ], + "angle": 0, + "content": "where the normalization factor is denoted as \\(\\widetilde{\\eta} (\\mathbf{x}_i^n) = 1 / \\sum_{c = 1}^{C}\\widetilde{\\alpha}_c(\\mathbf{x}_i^n)\\), the coefficients of multi-Gaussian are defined as \\(\\widetilde{\\alpha}_{c}(\\mathbf{x}_{i}^{n}) = \\alpha_{c}\\exp (\\frac{1}{2} (\\mathbf{x}_{i}^{n})^{\\intercal}\\pmb{\\sigma}_{c}\\mathbf{x}_{i}^{n} + \\frac{1}{\\epsilon} (\\widetilde{\\mu}_{c})^{\\intercal}(\\mathbf{x}_{i}^{n}))\\), and the mean vectors are calculated as \\(\\widetilde{\\pmb{\\mu}}_c(\\mathbf{x}_i^n) = \\pmb {\\mu}_c + \\frac{1}{\\epsilon}\\pmb {\\sigma}_c\\mathbf{x}_i^n\\)" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.708, + 0.906, + 0.769 + ], + "angle": 0, + "content": "We proceed with deriving the bridge function \\(\\psi_{p}\\) that represents a SDE process: \\(d\\mathbf{x}_t = g(x_t,t)dt + \\sqrt{\\epsilon} dW_t\\) where the shift function \\(g\\) is solved. According to Eqn. (4), we can obtain the shift function of diffusion process as follows:" + }, + { + "type": "equation", + "bbox": [ + 0.517, + 0.793, + 0.907, + 0.813 + ], + "angle": 0, + "content": "\\[\ng (\\mathbf {x} _ {i} ^ {\\mathrm {n}}, t) = \\epsilon \\rho_ {\\mathrm {M G P}} (\\mathbf {x} _ {i} ^ {\\mathrm {n}}) \\nabla_ {\\mathbf {x} _ {i} ^ {\\mathrm {n}}} \\log (\\mathcal {N} (\\mathbf {x} _ {i} ^ {\\mathrm {n}} | 0, \\epsilon (1 - t)) \\mathbf {I}), \\quad (1 4)\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.82, + 0.907, + 0.905 + ], + "angle": 0, + "content": "where the coefficients \\(\\rho_{\\mathrm{MGP}}\\) defined on multiple distribution prototypes is calculated as: \\(\\rho_{\\mathrm{MGP}}(\\mathbf{x}_i^{\\mathrm{n}}) = \\sum_{c=1}^{C} (\\alpha_c \\mathcal{N}(\\widetilde{\\boldsymbol{\\mu}}_c(\\mathbf{x}_i^{\\mathrm{n}}) | \\mathbf{0}, \\boldsymbol{\\sigma}_c) \\mathcal{N}(h_c(\\mathbf{x}_i^{\\mathrm{n}}, t) | 0, \\boldsymbol{\\Sigma}_c^t))\\), with another Gaussian of \\(h_c(\\mathbf{x}_i^{\\mathrm{n}}, t) = \\frac{1}{\\epsilon(1 - t)} \\mathbf{x}_i^{\\mathrm{n}} + \\boldsymbol{\\sigma}_c^{-1} \\widetilde{\\boldsymbol{\\mu}}_c(\\mathbf{x}_i^{\\mathrm{n}})\\) and \\(\\boldsymbol{\\Sigma}_c^t = \\frac{t}{\\epsilon(1 - t)} \\mathbf{I} + \\boldsymbol{\\sigma}_c^{-1}\\)." + }, + { + "type": "page_number", + "bbox": [ + 0.479, + 0.945, + 0.521, + 0.958 + ], + "angle": 0, + "content": "20419" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.09, + 0.092, + 0.485, + 0.228 + ], + "angle": 0, + "content": "Distribution prototypes initialization. Jointly learning prototypes \\(\\{\\alpha_c, \\mu_c, \\sigma_c\\}_{c=1}^C\\) and the bridge transformation \\(\\psi_p\\) (also the shift \\(g\\)) is a challenging task as they are interdependent. For this, we leverage a vector quantization function to learn a codebook \\(\\mathcal{E}\\) of prototypes within a discrete latent space from training data. Specifically, given an input image feature \\(\\mathbf{x}_i^n\\), we can assign \\(\\mathbf{x}_i^n\\) to the closest prototype \\(\\mathbf{e}_k\\) by minimizing the L2 distance between \\(\\mathbf{x}_i^n\\) and each of prototypes \\(\\mathbf{e}_c \\in \\mathcal{E}\\), as follows:" + }, + { + "type": "equation", + "bbox": [ + 0.092, + 0.235, + 0.484, + 0.262 + ], + "angle": 0, + "content": "\\[\n\\min _ {\\left\\{\\mathbf {e} _ {c} \\right\\}} \\mathbb {E} _ {\\mathbf {x} _ {i} ^ {\\mathrm {n}} \\in \\mathcal {F} _ {\\mathrm {t r}}} \\left[ \\left\\| \\mathbf {x} _ {i} ^ {\\mathrm {n}} - \\mathbf {e} _ {c ^ {*}} \\right\\| _ {2} ^ {2} \\right], \\text {s . t .}, c ^ {*} = \\underset {c} {\\arg \\min } \\left\\| \\mathbf {x} _ {i} ^ {\\mathrm {n}} - \\mathbf {e} _ {c} \\right\\| _ {2}, \\tag {15}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.091, + 0.268, + 0.484, + 0.344 + ], + "angle": 0, + "content": "where \\( c^* \\) denotes the index of the prototype closest to \\( \\mathbf{x}_i^{n} \\). The learned \\( \\{\\mathbf{e}_c\\}_{c=1}^C \\) are used to initialize the mean vectors \\( \\{\\pmb{\\mu}_c\\}_{c=1}^C \\), while the variances of all prototypes are set to the identity matrix. We observe that this initialization strategy accelerates the training process." + }, + { + "type": "text", + "bbox": [ + 0.091, + 0.344, + 0.484, + 0.405 + ], + "angle": 0, + "content": "Distribution loss of normal and anomaly samples. As \\( p_0 \\) and \\( p_1 \\) are accessible only via samples \\( \\mathcal{F}_{\\mathrm{tr}} \\) and prototypes \\( \\mathcal{P}_{\\mathrm{MGP}} \\), we optimize the empirical form in Eqn. (2) for normal samples as follows [10, 17]:" + }, + { + "type": "equation", + "bbox": [ + 0.11, + 0.412, + 0.483, + 0.455 + ], + "angle": 0, + "content": "\\[\n\\mathcal {L} _ {\\mathrm {D P L}} ^ {\\mathrm {n}} = \\frac {1}{N} \\sum_ {i = 1} ^ {N} \\log \\varpi_ {\\theta} \\left(\\mathbf {x} _ {i} ^ {\\mathrm {n}}\\right) - \\frac {1}{C} \\sum_ {c = 1} ^ {C} \\log \\phi_ {1} \\left(\\boldsymbol {\\mu} _ {c}\\right), \\tag {16}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.461, + 0.484, + 0.492 + ], + "angle": 0, + "content": "In contrast, for anomaly samples, we aim to push anomaly distribution away from \\( p_1 \\), i.e., negative loss, formally," + }, + { + "type": "equation", + "bbox": [ + 0.109, + 0.499, + 0.483, + 0.541 + ], + "angle": 0, + "content": "\\[\n\\mathcal {L} _ {\\mathrm {D P L}} ^ {\\mathrm {a}} = \\frac {1}{C} \\sum_ {c = 1} ^ {C} \\log \\phi_ {1} (\\boldsymbol {\\mu} _ {c}) - \\frac {1}{M} \\sum_ {i = 1} ^ {M} \\log \\varpi_ {\\theta} \\left(\\mathbf {x} _ {i} ^ {\\mathrm {a}}\\right), \\tag {17}\n\\]" + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.547, + 0.35, + 0.563 + ], + "angle": 0, + "content": "4.3. Dispersion Feature Learning" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.569, + 0.484, + 0.75 + ], + "angle": 0, + "content": "A critical challenge in OSAD is detecting previously unseen anomalies in open-set environments. Due to the limited anomaly observations, existing methods often use pseudo-anomaly generation strategies as a means of effective data augmentation. Nonetheless, the effectiveness of these methods heavily depends on the quality of the pseudo-anomaly feature embeddings. Notably, pseudo-anomaly distributions often inherit biases from the in-distribution data, which differ from the unknown anomaly distributions in out-of-distribution data. Current methods fail to address the relationship between observed and unknown anomalies, particularly for out-of-distribution generalization." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.75, + 0.484, + 0.902 + ], + "angle": 0, + "content": "To improve out-of-distribution detection, it is essential to promote a larger inter-sample dispersion, as greater distances among in-distribution samples facilitate their more effective separation from out-of-distribution samples. In other words, if all inter-sample distances are close to zero, i.e., collapse to a single point, it becomes impossible to differentiate between samples. Hence, promoting separability through a larger inter-sample dispersion is critical for accurately identifying samples that do not belong to known categories. To do so, we map the features into a hyperspherical" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.092, + 0.906, + 0.168 + ], + "angle": 0, + "content": "space. Draws inspiration from the vMF distribution [31] in directional statistics, we compute spherical Gaussian distributions for unit-norm features \\(\\widehat{\\mathbf{x}}_i = \\mathbf{x}_i / \\| \\mathbf{x}_i\\| _2^2\\). The probability density function of a unit vector \\(\\widehat{\\mathbf{x}}_i\\in \\mathbb{R}^D\\) is defined in the hyperspherical space as follows:" + }, + { + "type": "equation", + "bbox": [ + 0.575, + 0.18, + 0.907, + 0.198 + ], + "angle": 0, + "content": "\\[\np _ {D} (\\widehat {\\mathbf {x}} _ {i}; \\widehat {\\mathbf {x}} _ {j}, \\kappa) = F _ {D} (\\kappa) \\exp \\left(\\kappa \\langle \\widehat {\\mathbf {x}} _ {i}, \\widehat {\\mathbf {x}} _ {j} \\rangle\\right), \\tag {18}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.209, + 0.907, + 0.316 + ], + "angle": 0, + "content": "where \\(\\kappa \\geq 0\\) controls the concentration of the distribution around the mean direction \\(\\widehat{\\mathbf{x}}_j\\) and \\(F_{D}(\\kappa)\\) is the normalization factor. A larger \\(\\kappa\\) value increases concentration around the mean, while in the extreme case \\(\\kappa = 0\\), sample points are uniformly distributed on the hypersphere. Therefore, we design a dispersion loss to optimize large angular distances between the features of all samples:" + }, + { + "type": "equation", + "bbox": [ + 0.515, + 0.326, + 0.907, + 0.37 + ], + "angle": 0, + "content": "\\[\n\\mathcal {L} _ {\\mathrm {D F L}} = \\frac {1}{U} \\sum_ {i = 1} ^ {U} \\log \\frac {1}{U - 1} \\sum_ {i, j = 1} ^ {U} \\mathbb {1} \\{i \\neq j \\} \\exp (\\kappa \\langle \\widehat {\\mathbf {x}} _ {i}, \\widehat {\\mathbf {x}} _ {j} \\rangle), \\tag {19}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.38, + 0.899, + 0.396 + ], + "angle": 0, + "content": "where \\(U = N + M\\), and \\(\\kappa\\) is set to 10 in our experiments." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.406, + 0.753, + 0.422 + ], + "angle": 0, + "content": "4.4. Anomaly Score Prediction" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.429, + 0.907, + 0.626 + ], + "angle": 0, + "content": "Based on the above designs, we leverage the multiple-instance-learning (MIL)-based method proposed in [35] to effectively learn anomaly scores. Similar to the work [9], we design three modules \\(\\mathsf{M} = \\{\\mathsf{M}_{\\mathsf{a}},\\mathsf{M}_{\\mathsf{n}}\\) and \\(\\mathsf{M}_{\\mathsf{r}}\\}\\) for estimating anomaly scores. Firstly, for the feature map \\(\\mathbf{x}_i^1\\), we generate pixel-wise feature vectors \\(\\mathcal{V} = \\{\\mathbf{v}_i\\}_{i = 1}^{H' \\times W'}\\) to represent the feature of small patches of the image \\(\\mathbf{x}_i\\), where \\((H',W')\\) denotes the size of the feature map. These pixel-wise representations are then mapped by an anomaly classifier \\(S_{\\mathrm{a}}\\) to estimate pixel-level anomaly scores. To capture those points with the most salient anomalies, we compute the top-\\(K\\) most anomaly pixel points and define the loss function as:" + }, + { + "type": "equation", + "bbox": [ + 0.522, + 0.635, + 0.907, + 0.665 + ], + "angle": 0, + "content": "\\[\n\\mathcal {L} _ {\\mathrm {M} _ {\\mathrm {a}}} \\left(\\mathbf {x} _ {i}, y _ {i}\\right) = \\mathcal {L} _ {\\text {b i n a r y}} \\left(\\frac {1}{K} \\sum \\operatorname {T o p} _ {K} \\left\\{S _ {\\mathrm {a}} \\left(\\mathbf {v} _ {i}; \\theta_ {\\mathrm {a}}\\right) \\right\\}, y _ {i}\\right), \\tag {20}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.674, + 0.906, + 0.72 + ], + "angle": 0, + "content": "where \\(\\mathcal{L}_{\\mathrm{binary}}\\) refers to a binary classification loss function, and \\(\\operatorname{Top}_K\\) selects the highest \\(K\\) anomaly scores among all the vectors. We use \\(\\mathsf{M}_{\\mathfrak{n}}\\) to learn the normal features:" + }, + { + "type": "equation", + "bbox": [ + 0.526, + 0.731, + 0.905, + 0.786 + ], + "angle": 0, + "content": "\\[\n\\mathcal {L} _ {\\mathrm {M} _ {\\mathrm {n}}} \\left(\\mathbf {x} _ {i}, y _ {i}\\right) = \\mathcal {L} _ {\\text {b i n a r y}} \\left(S _ {\\mathrm {n}} \\left(\\frac {1}{H ^ {\\prime} \\times W ^ {\\prime}} \\sum_ {i = 1} ^ {H ^ {\\prime} \\times W ^ {\\prime}} \\mathbf {v} _ {i}; \\theta_ {\\mathrm {n}}\\right), y _ {i}\\right), \\tag {21}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.787, + 0.906, + 0.848 + ], + "angle": 0, + "content": "where \\(S_{\\mathfrak{n}}:\\mathcal{V}\\to \\mathbb{R}\\) is a fully connected binary anomaly classifier. Finally, we define \\(\\mathsf{M}_{\\mathfrak{x}}\\) to compute the residual anomaly scores between fine-grained visual semantics and abstract prototypes:" + }, + { + "type": "equation", + "bbox": [ + 0.534, + 0.859, + 0.907, + 0.878 + ], + "angle": 0, + "content": "\\[\n\\mathcal {L} _ {\\mathrm {M} _ {\\mathrm {r}}} = \\mathcal {L} _ {\\text {b i n a r y}} \\left(S _ {\\mathrm {r}} \\left(\\left(\\psi_ {p} (\\mathbf {x} _ {i}) - \\boldsymbol {\\mu} _ {c ^ {*}}\\right) / \\boldsymbol {\\sigma} _ {c ^ {*}}; \\theta_ {\\mathrm {r}}\\right), y _ {i}\\right), \\tag {22}\n\\]" + }, + { + "type": "page_footnote", + "bbox": [ + 0.532, + 0.887, + 0.905, + 0.901 + ], + "angle": 0, + "content": "Here \\(\\mathbf{x}_i\\) refers to convolution maps without flattening (as used above)." + }, + { + "type": "footer", + "bbox": [ + 0.479, + 0.945, + 0.521, + 0.958 + ], + "angle": 0, + "content": "20420" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.09, + 0.092, + 0.482, + 0.137 + ], + "angle": 0, + "content": "where \\(c^*\\) denotes the index of the most probable prototypes, i.e., \\(c^* = \\arg \\max_c\\mathcal{N}(\\psi_p(\\mathbf{x}_i);\\pmb {\\mu}_c,\\pmb {\\sigma}_c)\\), and \\(S_{\\mathrm{r}}\\) utilize the same method to obtain anomaly score as \\(S_{\\mathrm{a}}\\)." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.138, + 0.483, + 0.199 + ], + "angle": 0, + "content": "Training. During the training phase, the SB and three prediction modules are jointly trained. To this end, we employ an objective function that encompasses three components as follows:" + }, + { + "type": "equation", + "bbox": [ + 0.103, + 0.213, + 0.483, + 0.248 + ], + "angle": 0, + "content": "\\[\n\\mathcal {L} = \\underbrace {\\mathcal {L} _ {\\mathrm {M} _ {\\mathrm {a}}} + \\mathcal {L} _ {\\mathrm {M} _ {\\mathrm {n}}} + \\mathcal {L} _ {\\mathrm {M} _ {\\mathrm {r}}}} _ {\\text {M I L - b a s e d l e a n i n g}} + \\underbrace {\\mathcal {L} _ {\\mathrm {D P L}} ^ {\\mathrm {n}} + \\mathcal {L} _ {\\mathrm {D P L}} ^ {\\mathrm {a}}} _ {\\text {S B t r a n s f o r m}} + \\underbrace {\\lambda \\mathcal {L} _ {\\mathrm {D F L}}} _ {\\text {d i s p e r s i o n}}. \\tag {23}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.262, + 0.483, + 0.308 + ], + "angle": 0, + "content": "where the coefficient \\(\\lambda\\) modulates the relative importance of dispersion loss, and the learnable parameters include \\(\\{\\alpha_{c},\\pmb{\\mu}_{c},\\pmb{\\sigma}_{c}\\}\\), \\(\\theta_{\\psi_p}\\), and \\(\\{\\theta_{\\mathrm{a}},\\theta_{\\mathrm{n}},\\theta_{\\mathrm{r}}\\}\\)." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.309, + 0.483, + 0.384 + ], + "angle": 0, + "content": "Inference. During the test phase, we find the most similar class prototype through SB, and subsequently compute the anomaly score by adding the scores from both \\( S_{\\mathrm{a}} \\) and \\( S_{\\mathrm{r}} \\), while subtracting the normal score obtained from \\( S_{\\mathrm{n}} \\) for the given test image." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.401, + 0.216, + 0.417 + ], + "angle": 0, + "content": "5. Experiment" + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.427, + 0.364, + 0.442 + ], + "angle": 0, + "content": "5.1. Dataset and Evaluation Metric" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.45, + 0.483, + 0.631 + ], + "angle": 0, + "content": "Dataset To validate the effectiveness of DPDL, comprehensive experiments are conducted on nine real-world AD datasets, including six industrial defect detection datasets (MVTec AD [3], Optical [46], SDD [40], AITEX [39], ELPV [8], Mastcam [13]) and three medical image datasets (Hyper-Kvasir [4], Brain-MRI [37], HeadCT [37]). We follow the previous OSAD baselines [9, 51] to adopt two protocols for sampling, including general setting and hard setting. The general setting assumes that anomaly examples are randomly sampled from the anomaly class, while the hard setting samples from a single class to assess generalization to new or unseen anomaly classes." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.632, + 0.483, + 0.693 + ], + "angle": 0, + "content": "Evaluation Metric We utilize the widely adopted Area Under ROC Curve (AUC) as a metric to evaluate the performance across all methods and settings. All reported AUCs are averaged results over five independent runs." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.706, + 0.2, + 0.721 + ], + "angle": 0, + "content": "5.2. Baselines" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.729, + 0.483, + 0.835 + ], + "angle": 0, + "content": "We compare DPDL against six related state-of-the-art OSAD baselines, including SAOE [19, 32, 41], MLEP [27], FLOS [36], DevNet [35], DRA [9], and AHL [51]. MLEP, DevNet, DRA, and AHL are specifically designed for OSAD. SAOE is a supervised detector enhanced with synthetic anomalies and anomaly exposure, whereas FLOS is an imbalanced classifier leveraging focal loss." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.848, + 0.308, + 0.864 + ], + "angle": 0, + "content": "5.3. Implementation Details" + }, + { + "type": "text", + "bbox": [ + 0.091, + 0.871, + 0.483, + 0.901 + ], + "angle": 0, + "content": "The input image size is \\(448 \\times 448 \\times 3\\). We set \\(K\\) in the top- \\(K\\) MIL to \\(10\\%\\) of the number of all scores per score map." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.092, + 0.907, + 0.244 + ], + "angle": 0, + "content": "AdamW optimizer [29] is used for the parameter optimization using an initial learning rate \\( 2 \\times 10^{-4} \\) with a weight decay of \\( 1 \\times 10^{-5} \\). DPDL is trained on one NVIDIA GeForce RTX 4090 GPU, which are trained using 50 epochs, with 20 iterations per epoch. Following previous protocol [9, 51], we evaluate performance with anomaly sample numbers of \\( M = 10 \\) and \\( M = 1 \\), and for robust detection of unseen anomalies, we use CutMix [49] to create pseudo-anomaly samples as augmented data for known anomalies. The prototype quantity \\( C \\) is set to 32 as default." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.266, + 0.783, + 0.282 + ], + "angle": 0, + "content": "5.4. Results under General Setting" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.293, + 0.908, + 0.685 + ], + "angle": 0, + "content": "Tab. 1 highlights DPDL's strong performance. In the challenging scenario of single-anomaly detection, DPDL improves the performance of AHL [51] by more than \\(8.3\\%\\) on the AITEX, ELPV, and Mastcam datasets. Furthermore, it achieves significant improvements across six additional datasets, which suggest the effective utilization of few-shot anomaly examples in DPDL, while mitigating overfitting to the seen anomalies. When shifting to ten anomaly examples settings, DPDL continues to maintain a significant lead with over \\(5.4\\%\\) improvement on those datasets. Given the rich and diverse set of normal samples in these datasets, DPDL leverages the DPL component to encapsulate these samples within a compact, discriminative distribution space, while effectively pushing anomalous samples outside this space, thereby enabling accurate anomaly detection. In the setting with ten abnormal samples, although existing methods have reached performance saturation on the MVTecAD, Optical, and SDD datasets, DPDL still has a certain lead, demonstrating the strong ability of DPL and DFL in learning tight boundaries of normal sample distributions and generalizing to previously unseen anomaly domains. Furthermore, when evaluated on the medical datasets BrainMRI and HeadCT, DPDL demonstrates competitive performance despite these datasets being notably small in scale and containing only a single class. This highlights the algorithm's ability to deliver robust results even in data-scarce conditions." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.709, + 0.792, + 0.725 + ], + "angle": 0, + "content": "5.5. Results under the Hard Setting" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.735, + 0.907, + 0.901 + ], + "angle": 0, + "content": "Tab. 2 summarizes the performance comparison under the hard setting. It is evident that DPDL achieves the highest AUC scores in both the single-anomaly and ten-anomaly sample settings. Specifically, compared to the closest competing method, AHL [51], DPDL achieves an improvement in AUC scores ranging from \\(0.6\\%\\) to \\(7.9\\%\\) in the ten anomaly examples settings and from \\(0.5\\%\\) to \\(8.9\\%\\) in the one anomaly example settings, respectively. The observed improvement can be attributed to the strong generalization ability of DPDL in detecting unseen anomaly classes, even when the model is trained on only a single anomaly class." + }, + { + "type": "page_number", + "bbox": [ + 0.48, + 0.945, + 0.518, + 0.957 + ], + "angle": 0, + "content": "20421" + } + ], + [ + { + "type": "table_caption", + "bbox": [ + 0.09, + 0.09, + 0.908, + 0.119 + ], + "angle": 0, + "content": "Table 1. AUC performance (mean ± std) across nine real-world AD datasets is reported under the general setting. red highlights the best results, and blue indicates sub-optimal outcomes. All baseline SOTA results are sourced from the original papers [9, 51]." + }, + { + "type": "table", + "bbox": [ + 0.099, + 0.128, + 0.897, + 0.429 + ], + "angle": 0, + "content": "
DatasetDevNetFLOSSAOEMLEPDRAAHLDPDL (Ours)
Ten Training Anomaly Examples
MVTec AD0.945±0.0040.939±0.0070.926±0.0100.907±0.0050.959±0.0030.970±0.0020.977±0.002
Optical0.782±0.0650.720±0.0550.941±0.0130.740±0.0390.965±0.0060.976±0.0040.983±0.005
SDD0.988±0.0060.967±0.0180.955±0.0200.983±0.0130.991±0.0050.991±0.0010.996±0.001
AITEX0.887±0.0130.841±0.0490.874±0.0240.867±0.0370.893±0.0170.925±0.0130.975±0.007
ELPV0.846±0.0220.818±0.0320.793±0.0470.794±0.0470.845±0.0130.850±0.0040.937±0.003
Mastcam0.790±0.0210.703±0.0290.810±0.0290.798±0.0260.848±0.0080.855±0.0050.934±0.010
Hyper-Kvasir0.829±0.0180.773±0.0290.666±0.0500.600±0.0690.834±0.0040.880±0.0030.939±0.005
BrainMRI0.958±0.0120.955±0.0110.900±0.0410.959±0.0110.970±0.0030.977±0.0010.969±0.005
HeadCT0.982±0.0090.971±0.0040.935±0.0210.972±0.0140.972±0.0020.999±0.0030.981±0.003
One Training Anomaly Example
MVTec AD0.780±0.0200.755±0.1360.834±0.0070.744±0.0190.883±0.0080.901±0.0030.927±0.002
Optical0.523±0.0030.518±0.0030.815±0.0140.516±0.0090.888±0.0120.888±0.0070.915±0.002
SDD0.881±0.0090.840±0.0430.781±0.0090.811±0.0450.859±0.0140.909±0.0010.917±0.003
AITEX0.598±0.0700.538±0.0730.675±0.0940.564±0.0550.692±0.1240.734±0.0080.838±0.008
ELPV0.514±0.0760.457±0.0560.635±0.0920.578±0.0620.675±0.0240.828±0.0050.897±0.002
Mastcam0.595±0.0160.542±0.0170.662±0.0180.625±0.0450.692±0.0580.743±0.0030.838±0.011
Hyper-Kvasir0.653±0.0370.668±0.0040.498±0.1000.445±0.0400.690±0.0170.768±0.0150.821±0.007
BrainMRI0.694±0.0040.693±0.0360.531±0.0600.632±0.0170.744±0.0040.866±0.0040.893±0.004
HeadCT0.742±0.0760.698±0.0920.597±0.0220.758±0.0380.796±0.1050.825±0.0140.865±0.005
" + }, + { + "type": "table_caption", + "bbox": [ + 0.09, + 0.44, + 0.908, + 0.482 + ], + "angle": 0, + "content": "Table 2. AUC results (mean ± std) under the hard setting. The best and second-best results are highlighted in red and blue, respectively. Carpet and Metal_nut are subsets of MVTec AD. The datasets used are consistent with those in [9, 51], where those datasets only containing one anomaly class are excluded to adapt for the hard setting." + }, + { + "type": "table", + "bbox": [ + 0.098, + 0.493, + 0.897, + 0.71 + ], + "angle": 0, + "content": "
DatasetDevNetFLOSSAOEMLEPDRAAHLDPDL (Ours)
Ten Training Anomaly Examples
Carpet (mean)0.847±0.0170.761±0.0120.762±0.0730.751±0.0230.935±0.0130.949±0.0020.956±0.004
Metal_nut (mean)0.965±0.0110.922±0.0140.855±0.0160.878±0.0580.945±0.0170.972±0.0020.978±0.002
AITEX (mean)0.683±0.0320.635±0.0430.724±0.0320.626±0.0410.733±0.0090.747±0.0020.798±0.005
ELPV (mean)0.702±0.0230.642±0.0320.683±0.0470.745±0.0200.766±0.0290.788±0.0030.818±0.003
Mastcam (mean)0.588±0.0110.616±0.0210.697±0.0140.588±0.0160.695±0.0040.721±0.0030.778±0.007
Hyper-Kvasir (mean)0.822±0.0190.786±0.0210.698±0.0210.571±0.0140.844±0.0090.854±0.0040.864±0.002
One Training Anomaly Example
Carpet (mean)0.767±0.0180.678±0.0400.753±0.0550.679±0.0290.901±0.0060.932±0.0030.941±0.006
Metal_nut (mean)0.855±0.0160.855±0.0240.816±0.0290.825±0.0230.932±0.0170.939±0.0040.944±0.003
AITEX (mean)0.646±0.0340.624±0.0240.674±0.0340.466±0.0300.684±0.0330.707±0.0070.753±0.005
ELPV (mean)0.648±0.0570.691±0.0080.614±0.0480.566±0.1110.703±0.0220.740±0.0030.762±0.003
Mastcam (mean)0.511±0.0130.524±0.0130.689±0.0370.541±0.0070.667±0.0120.673±0.0100.733±0.004
Hyper-Kvasir (mean)0.595±0.0230.571±0.0040.406±0.0180.480±0.0440.700±0.0090.706±0.0070.715±0.004
" + }, + { + "type": "image_caption", + "bbox": [ + 0.174, + 0.743, + 0.422, + 0.753 + ], + "angle": 0, + "content": "Ten Training Anomaly Examples Under General Settings" + }, + { + "type": "image", + "bbox": [ + 0.097, + 0.753, + 0.297, + 0.892 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.306, + 0.753, + 0.498, + 0.892 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.594, + 0.743, + 0.831, + 0.753 + ], + "angle": 0, + "content": "Ten Training Anomaly Examples Under Hard Settings" + }, + { + "type": "image", + "bbox": [ + 0.51, + 0.753, + 0.7, + 0.892 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.712, + 0.753, + 0.902, + 0.892 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.243, + 0.892, + 0.753, + 0.906 + ], + "angle": 0, + "content": "Figure 2. Ablation study for SB and DFL under the general settings and hard settings." + }, + { + "type": "page_number", + "bbox": [ + 0.479, + 0.946, + 0.52, + 0.957 + ], + "angle": 0, + "content": "20422" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.098, + 0.089, + 0.301, + 0.202 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.305, + 0.089, + 0.5, + 0.202 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.505, + 0.089, + 0.701, + 0.202 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.705, + 0.089, + 0.902, + 0.202 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.327, + 0.202, + 0.671, + 0.216 + ], + "angle": 0, + "content": "Figure 3. Parameter sensitivity analysis for \\(C,\\epsilon ,\\kappa\\) and \\(\\lambda\\)" + }, + { + "type": "table_caption", + "bbox": [ + 0.147, + 0.222, + 0.426, + 0.236 + ], + "angle": 0, + "content": "Table 3. An ablation study for \\( {\\mathrm{M}}_{\\mathrm{n}},{\\mathrm{M}}_{\\mathrm{a}} \\) and \\( {\\mathrm{M}}_{\\mathrm{r}} \\) ." + }, + { + "type": "table", + "bbox": [ + 0.099, + 0.238, + 0.477, + 0.417 + ], + "angle": 0, + "content": "
MnMaMrAITEXELPVMastcam
Ten Training Anomaly Examples Under General Settings
××✓0.928 ± 0.0190.914 ± 0.0210.899 ± 0.036
✓×✓0.939 ± 0.0230.919 ± 0.0110.908 ± 0.017
×✓✓0.963 ± 0.0080.930 ± 0.0130.924 ± 0.019
✓✓✓0.975 ± 0.0070.937 ± 0.0030.934 ± 0.010
Ten Training Anomaly Examples Under Hard Settings
××✓0.746 ± 0.0250.797 ± 0.0260.723 ± 0.017
✓×✓0.758 ± 0.0350.802 ± 0.0240.736 ± 0.028
×✓✓0.781 ± 0.0140.811 ± 0.0180.762 ± 0.021
✓✓✓0.798 ± 0.0050.818 ± 0.0030.778 ± 0.007
" + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.423, + 0.245, + 0.439 + ], + "angle": 0, + "content": "5.6. Ablation Study" + }, + { + "type": "text", + "bbox": [ + 0.089, + 0.445, + 0.483, + 0.717 + ], + "angle": 0, + "content": "The ablation study in Fig. 2 highlights the critical roles of the DPL and DFL components in improving DPDL's openset anomaly detection. We denote the variants that remove only DPL or DFL as 'DPDL w/o DPL' and 'DPDL w/o DFL', respectively. Compared to the full DPDL model, removing neither DPL nor DFL leads to a significant AUC decline, illustrating their critical utility. Specifically, 'DPDL w/o DPL' exhibits the most significant performance drop on industrial anomaly datasets, reflecting DPL's prominent role in learning precise and tight distribution boundaries for normal samples. By transforming normal samples into Gaussian distribution prototype space and pushing abnormal samples away, DPL enhances the recognition ability of anomaly samples. Meanwhile, ablation studies on 'DPDL w/o DFL' further highlight the critical role of the DFL component. By performing discreteness feature learning in hyperspherical space, DFL enhances the generalization ability to out-of-distribution anomalies." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.717, + 0.483, + 0.822 + ], + "angle": 0, + "content": "Additionally, the ablation experiments on \\( \\mathsf{M}_{\\mathrm{n}} \\), \\( \\mathsf{M}_{\\mathrm{a}} \\), and \\( \\mathsf{M}_{\\mathrm{r}} \\) across three datasets in Tab. 3 reveal their varying contributions. DPDL shows the most significant performance drop when \\( \\mathsf{M}_{\\mathrm{r}} \\) is removed, which illustrates that \\( \\mathsf{M}_{\\mathrm{r}} \\) plays the most critical role in detecting anomalies. Furthermore, the removal of \\( \\mathsf{M}_{\\mathrm{a}} \\) and \\( \\mathsf{M}_{\\mathrm{n}} \\) lead to a noticeable performance drop in DPDL, which underscores their essential roles." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.833, + 0.364, + 0.851 + ], + "angle": 0, + "content": "5.7. Parameter Sensitivity Analysis" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.856, + 0.483, + 0.903 + ], + "angle": 0, + "content": "Fig. 3 illustrates the results of the four hyperparameters under the general settings across nine datasets. Overall, the performance remains stable within a certain range of hyper" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.224, + 0.905, + 0.42 + ], + "angle": 0, + "content": "parameter variations, demonstrating the DPDL's robustness. Prototype quantity \\( C \\). We begin by investigating the critical impact of the number of initialized prototypes \\( C \\) on distributed prototype learning in DPDL. We select \\( \\{8,16,32,64\\} \\) as the values for the hyperparameters. As \\( C \\) increases, DPDL's performance improves steadily, but excessively large values of \\( C \\) hinder the model's effectiveness. This phenomenon is particularly pronounced on AI-TEX, Mastcam, MVTecAD, and Hyper-Kvasir, which contain more categories. One possible explanation is that a prototype space with too few prototypes loses discriminative information, while an excessively large number of prototypes reduces the compactness of the space." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.421, + 0.906, + 0.557 + ], + "angle": 0, + "content": "DPL trajectory \\(\\epsilon\\) in Eqns. (9), (10), (11) and (14). It can be observed that, particularly on the AITEX, Mastcam, and MVTecAD datasets with a larger number of categories, DPDL exhibits a relatively stable performance decline as \\(\\epsilon\\) increases. As a crucial parameter in SB, \\(\\epsilon\\) governs the trajectory state. Since smaller values produce straighter trajectories and larger values increase fluctuation, smaller \\(\\epsilon\\) facilitates sampling more robust abstract prototypes from the relatively dispersed conditional distribution." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.558, + 0.906, + 0.633 + ], + "angle": 0, + "content": "DFL tightness \\(\\kappa\\). According to Fig. 3, increasing \\(\\kappa\\) generally enhances model performance, but values above \\(\\kappa = 10\\) introduce negative effects in certain scenarios. A possible reason is that excessive sample dispersion makes it more challenging to tighten the normal distribution boundary." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.634, + 0.906, + 0.739 + ], + "angle": 0, + "content": "Loss parameter \\(\\lambda\\). We conduct a sensitivity analysis on the loss parameters \\(\\lambda\\). It can be observe that setting \\(\\lambda = 0.01\\) achieves optimal performance on seven larger-scale datasets, while \\(\\lambda = 1\\) yields the best results on two datasets with limited data. As \\(\\lambda\\) increases, the performance declines, potentially due to gradient conflicts among the dispersion loss, SB transform loss and the main task loss." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.755, + 0.634, + 0.77 + ], + "angle": 0, + "content": "6. Conclusion" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.78, + 0.906, + 0.901 + ], + "angle": 0, + "content": "We propose Distribution Prototype Diffusion Learning (DPDL) for OSAD. DPDL leverages schrödinger bridge to map the normal distribution to a prototype space, simultaneously repelling anomalies to facilitate precise anomaly detection. We propose a dispersion feature learning way in hyperspherical space, which benefits the detection of out-of-distribution anomalies. Experimental results illustrate DPDL's robustness in diverse anomaly detection scenarios." + }, + { + "type": "page_number", + "bbox": [ + 0.479, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "20423" + } + ], + [ + { + "type": "title", + "bbox": [ + 0.092, + 0.091, + 0.27, + 0.108 + ], + "angle": 0, + "content": "7. Acknowledgments" + }, + { + "type": "text", + "bbox": [ + 0.091, + 0.114, + 0.484, + 0.199 + ], + "angle": 0, + "content": "This work was supported by the National Natural Science Foundation of China (Grants No. 62476133), the National Natural Science Foundation of China (Nos. U22B2049), the National Natural Science Foundation of China (NSFC) under Grant 62272230, and the Natural Science Foundation of Shandong Province (Grant No. ZR2022LZH003)." + }, + { + "type": "title", + "bbox": [ + 0.093, + 0.227, + 0.188, + 0.243 + ], + "angle": 0, + "content": "References" + }, + { + "type": "ref_text", + "bbox": [ + 0.101, + 0.252, + 0.484, + 0.349 + ], + "angle": 0, + "content": "[1] Andra Acsintoae, Andrei Florescu, Mariana-Iuliana Georgescu, Tudor Mare, Paul Sumedrea, Radu Tudor Ionescu, Fahad Shahbaz Khan, and Mubarak Shah. Ubnormal: New benchmark for supervised open-set video anomaly detection. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 20143-20153, 2022. 1, 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.1, + 0.351, + 0.484, + 0.419 + ], + "angle": 0, + "content": "[2] Aimira Baitieva, David Hurych, Victor Besnier, and Olivier Bernard. Supervised anomaly detection for complex industrial images. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 17754-17762, 2024. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.101, + 0.422, + 0.483, + 0.49 + ], + "angle": 0, + "content": "[3] Paul Bergmann, Michael Fauser, David Sattlegger, and Carsten Steger. Mvtec ad-a comprehensive real-world dataset for unsupervised anomaly detection. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 9592-9600, 2019. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.492, + 0.484, + 0.576 + ], + "angle": 0, + "content": "[4] Hanna Borgli, Vajira Thambawita, Pia H Smedsrud, Steven Hicks, Debesh Jha, Sigrun L Eskeland, Kristin Ranheim Randel, Konstantin Pogorelov, Mathias Lux, Duc Tien Dang Nguyen, et al. Hyperkvasir, a comprehensive multi-class image and video dataset for gastrointestinal endoscopy. Scientific data, 7(1):283, 2020. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.577, + 0.484, + 0.632 + ], + "angle": 0, + "content": "[5] Yongxin Chen, Tryphon T Georgiou, and Michele Pavon. On the relation between optimal transport and schrödinger bridges: A stochastic control viewpoint. Journal of Optimization Theory and Applications, 169:671-691, 2016. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.633, + 0.485, + 0.702 + ], + "angle": 0, + "content": "[6] Songmin Dai, Yifan Wu, Xiaogiang Li, and Xiangyang Xue. Generating and reweighting dense contrastive patterns for unsupervised anomaly detection. In Proceedings of the AAAI Conference on Artificial Intelligence, pages 1454-1462, 2024. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.704, + 0.484, + 0.76 + ], + "angle": 0, + "content": "[7] Valentin De Bortoli, James Thornton, Jeremy Heng, and Arnaud Doucet. Diffusion schrödinger bridge with applications to score-based generative modeling. Advances in Neural Information Processing Systems, 34:17695-17709, 2021. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.761, + 0.484, + 0.829 + ], + "angle": 0, + "content": "[8] Sergiu Deitsch, Vincent Christlein, Stephan Berger, Claudia Buerhop-Lutz, Andreas Maier, Florian Gallwitz, and Christian Riess. Automatic classification of defective photovoltaic module cells in electroluminescence images. Solar Energy, 185:455-468, 2019. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.831, + 0.484, + 0.9 + ], + "angle": 0, + "content": "[9] Chouro Ding, Guansong Pang, and Chunhua Shen. Catching both gray and black swans: Open-set supervised anomaly detection. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 7388-7398, 2022. 1, 2, 5, 6, 7" + }, + { + "type": "list", + "bbox": [ + 0.1, + 0.252, + 0.485, + 0.9 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.093, + 0.906, + 0.147 + ], + "angle": 0, + "content": "[10] Nikita Gushchin, Sergei Kholkin, Evgeny Burnaev, and Alexander Korotin. Light and optimal schrödinger bridge matching. In *Forty-first International Conference on Machine Learning*, 2024. 2, 3, 4, 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.149, + 0.906, + 0.218 + ], + "angle": 0, + "content": "[11] Liren He, Zhengkai Jiang, Jinlong Peng, Liang Liu, Qianggang Du, Xiaobin Hu, Wenbing Zhu, Mingmin Chi, Yabiao Wang, and Chengjie Wang. Learning unified reference representation for unsupervised multi-class anomaly detection. arXiv preprint arXiv:2403.11561, 2024. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.219, + 0.906, + 0.288 + ], + "angle": 0, + "content": "[12] Teng Hu, Jiangning Zhang, Ran Yi, Yuzhen Du, Xu Chen, Liang Liu, Yabiao Wang, and Chengjie Wang. Anomalydiffusion: Few-shot anomaly image generation with diffusion model. In Proceedings of the AAAI Conference on Artificial Intelligence, pages 8526-8534, 2024. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.289, + 0.906, + 0.37 + ], + "angle": 0, + "content": "[13] Hannah R Kerner, Kiri L Wagstaff, Brian D Bue, Danika F Wellington, Samantha Jacob, Paul Horton, James F Bell, Chiman Kwan, and Heni Ben Amor. Comparison of novelty detection methods for multispectral images in rover-based planetary exploration missions. Data Mining and Knowledge Discovery, 34:1642-1675, 2020. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.371, + 0.905, + 0.425 + ], + "angle": 0, + "content": "[14] Beomsu Kim, Gihyun Kwon, Kwanyoung Kim, and Jong Chul Ye. Unpaired image-to-image translation via neural schr\\” odinger bridge. arXiv preprint arXiv:2305.15086, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.427, + 0.905, + 0.481 + ], + "angle": 0, + "content": "[15] Daehyun Kim, Sungyong Baik, and Tae Hyun Kim. Sanflow: Semantic-aware normalizing flow for anomaly detection. Advances in Neural Information Processing Systems, 36:75434-75454, 2023. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.483, + 0.905, + 0.523 + ], + "angle": 0, + "content": "[16] Hyunsu Kim, Jongmin Yoon, and Juho Lee. Fast ensembling with diffusion schr\\” odinger bridge. arXiv preprint arXiv:2404.15814, 2024. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.525, + 0.905, + 0.566 + ], + "angle": 0, + "content": "[17] Alexander Korotin, Nikita Gushchin, and Evgeny Burnaev. Light schr\\” odinger bridge. arXiv preprint arXiv:2310.01174, 2023. 2, 3, 4, 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.567, + 0.905, + 0.608 + ], + "angle": 0, + "content": "[18] Christian Léonard. A survey of the schr\\” odinger problem and some of its connections with optimal transport. arXiv preprint arXiv:1308.0215, 2013. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.609, + 0.906, + 0.678 + ], + "angle": 0, + "content": "[19] Chun-Liang Li, Kihyuk Sohn, Jinsung Yoon, and Tomas Pfister. Cutpaste: Self-supervised learning for anomaly detection and localization. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 9664-9674, 2021. 1, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.679, + 0.906, + 0.734 + ], + "angle": 0, + "content": "[20] Hanxi Li, Jingqi Wu, Hao Chen, Mingwen Wang, and Chunhua Shen. Efficient anomaly detection with budget annotation using semi-supervised residual transformer. arXiv preprint arXiv:2306.03492, 2023. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.735, + 0.906, + 0.815 + ], + "angle": 0, + "content": "[21] Xiaofan Li, Zhizhong Zhang, Xin Tan, Chengwei Chen, Yanyun Qu, Yuan Xie, and Lizhuang Ma. Prompted: Learning prompts with only normal samples for few-shot anomaly detection. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 16838-16848, 2024. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.818, + 0.905, + 0.872 + ], + "angle": 0, + "content": "[22] Jingyi Liao, Xun Xu, Manh Cuong Nguyen, Adam Goodge, and Chuan Sheng Foo. Coft-ad: Contrastive fine-tuning for few-shot anomaly detection. IEEE Transactions on Image Processing, 2024. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.873, + 0.905, + 0.902 + ], + "angle": 0, + "content": "[23] Guan-Horng Liu, Tianrong Chen, Oswin So, and Evangelos Theodorou. Deep generalized schrödinger bridge. Advances" + }, + { + "type": "list", + "bbox": [ + 0.517, + 0.093, + 0.906, + 0.902 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.48, + 0.945, + 0.52, + 0.957 + ], + "angle": 0, + "content": "20424" + } + ], + [ + { + "type": "ref_text", + "bbox": [ + 0.125, + 0.092, + 0.482, + 0.119 + ], + "angle": 0, + "content": "in Neural Information Processing Systems, 35:9374-9388, 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.122, + 0.483, + 0.176 + ], + "angle": 0, + "content": "[24] Guan-Horng Liu, Yaron Lipman, Maximilian Nickel, Brian Karrer, Evangelos A Theodorou, and Ricky TQ Chen. Generalized schr\\” odinger bridge matching. arXiv preprint arXiv:2310.02233, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.179, + 0.482, + 0.233 + ], + "angle": 0, + "content": "[25] Guan-Horng Liu, Arash Vahdat, De-An Huang, Evangelos A Theodorou, Weili Nie, and Anima Anandkumar. \\(i^2\\)-sb: Image-to-image schr\\” odinger bridge. arXiv preprint arXiv:2302.05872, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.236, + 0.483, + 0.317 + ], + "angle": 0, + "content": "[26] Jiaqi Liu, Kai Wu, Qiang Nie, Ying Chen, Bin-Bin Gao, Yong Liu, Jinbao Wang, Chengjie Wang, and Feng Zheng. Unsupervised continual anomaly detection with contrastively-learned prompt. In Proceedings of the AAAI Conference on Artificial Intelligence, pages 3639–3647, 2024. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.32, + 0.483, + 0.375 + ], + "angle": 0, + "content": "[27] Wen Liu, Weixin Luo, Zhengxin Li, Peilin Zhao, Shenghua Gao, et al. Margin learning embedded prediction for video anomaly detection with a few anomalies. In IJCAI, pages 023-3, 2019. 1, 2, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.377, + 0.483, + 0.432 + ], + "angle": 0, + "content": "[28] Xinyue Liu, Jianyuan Wang, Biao Leng, and Shuo Zhang. Dual-modeling decouple distillation for unsupervised anomaly detection. arXiv preprint arXiv:2408.03888, 2024.1" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.434, + 0.483, + 0.462 + ], + "angle": 0, + "content": "[29] I Loshchilov. Decoupled weight decay regularization. arXiv preprint arXiv:1711.05101, 2017. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.463, + 0.483, + 0.559 + ], + "angle": 0, + "content": "[30] Jialin Luo, Yuanzhi Wang, Ziqi Gu, Yide Qiu, Shuaizhen Yao, Fuyun Wang, Chunyan Xu, Wenhua Zhang, Dan Wang, and Zhen Cui. Mmm-rs: A multi-modal, multi-gsd, multiscene remote sensing dataset and benchmark for text-to-image generation. In The Thirty-eight Conference on Neural Information Processing Systems Datasets and Benchmarks Track. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.562, + 0.483, + 0.589 + ], + "angle": 0, + "content": "[31] Kanti V Mardia and Peter E Jupp. Directional statistics. John Wiley & Sons, 2009. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.591, + 0.483, + 0.659 + ], + "angle": 0, + "content": "[32] Amir Markovitz, Gilad Sharir, Itamar Friedman, Lihi Zelnik-Manor, and Shai Avidan. Graph embedded pose clustering for anomaly detection. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 10539-10547, 2020. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.661, + 0.483, + 0.716 + ], + "angle": 0, + "content": "[33] Maxence Noble, Valentin De Bortoli, Arnaud Doucet, and Alain Durmus. Tree-based diffusion schrödinger bridge with applications to wasserstein barycenters. Advances in Neural Information Processing Systems, 36, 2024. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.718, + 0.483, + 0.786 + ], + "angle": 0, + "content": "[34] Guansong Pang, Chunhua Shen, and Anton Van Den Hengel. Deep anomaly detection with deviation networks. In Proceedings of the 25th ACM SIGKDD international conference on knowledge discovery & data mining, pages 353-362, 2019. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.789, + 0.483, + 0.843 + ], + "angle": 0, + "content": "[35] Guansong Pang, Chouro Ding, Chunhua Shen, and Anton van den Hengel. Explainable deep few-shot anomaly detection with deviation networks. arXiv preprint arXiv:2108.00462, 2021. 2, 5, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.846, + 0.483, + 0.899 + ], + "angle": 0, + "content": "[36] T-YLPG Ross and GKHP Dólar. Focal loss for dense object detection. In proceedings of the IEEE conference on computer vision and pattern recognition, pages 2980-2988, 2017. 6" + }, + { + "type": "list", + "bbox": [ + 0.093, + 0.092, + 0.483, + 0.899 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.092, + 0.905, + 0.175 + ], + "angle": 0, + "content": "[37] Mohammadreza Salehi, Niousha Sadjadi, Soroosh Baselizadeh, Mohammad H Rohban, and Hamid R Rabiee. Multiresolution knowledge distillation for anomaly detection. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 14902-14912, 2021. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.177, + 0.905, + 0.231 + ], + "angle": 0, + "content": "[38] Yuyang Shi, Valentin De Bortoli, Andrew Campbell, and Arnaud Doucet. Diffusion schrödinger bridge matching. Advances in Neural Information Processing Systems, 36, 2024. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.233, + 0.905, + 0.287 + ], + "angle": 0, + "content": "[39] Javier Silvestre-Blanes, Teresa Albero-Albero, Ignacio Miralles, Rubén Pérez-Llorens, and Jorge Moreno. A public fabric database for defect detection methods and results. *Autex Research Journal*, 19(4):363–374, 2019. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.288, + 0.905, + 0.343 + ], + "angle": 0, + "content": "[40] Domen Tabernik, Samo Sela, Jure Skvarc, and Danijel Skocaj. Segmentation-based deep-learning approach for surface-defect detection. Journal of Intelligent Manufacturing, 31 (3):759-776, 2020. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.344, + 0.905, + 0.399 + ], + "angle": 0, + "content": "[41] Jihoon Tack, Sangwoo Mo, Jongheon Jeong, and Jinwoo Shin. Csi: Novelty detection via contrastive learning on distributionally shifted instances. Advances in neural information processing systems, 33:11839-11852, 2020. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.4, + 0.905, + 0.466 + ], + "angle": 0, + "content": "[42] Yuanzhi Wang, Zhen Cui, and Yong Li. Distribution-consistent modal recovering for incomplete multimodal learning. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 22025-22034, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.469, + 0.905, + 0.523 + ], + "angle": 0, + "content": "[43] Yuanzhi Wang, Yong Li, and Zhen Cui. Incomplete multimodality-diffused emotion recognition. Advances in Neural Information Processing Systems, 36:17117-17128, 2023." + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.525, + 0.905, + 0.579 + ], + "angle": 0, + "content": "[44] Yuanzhi Wang, Yong Li, Mengyi Liu, Xiaoya Zhang, Xin Liu, Zhen Cui, and Antoni B Chan. Re-attentional controllable video diffusion editing. arXiv preprint arXiv:2412.11710, 2024." + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.581, + 0.905, + 0.649 + ], + "angle": 0, + "content": "[45] Yuanzhi Wang, Yong Li, Xiaoya Zhang, Xin Liu, Anbo Dai, Antoni B Chan, and Zhen Cui. Edit temporal-consistent videos with image diffusion model. ACM Transactions on Multimedia Computing, Communications and Applications, 20(12):1-16, 2024. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.65, + 0.905, + 0.692 + ], + "angle": 0, + "content": "[46] Matthias Wieler and Tobias Hahn. Weakly supervised learning for industrial optical inspection. In DAGM symposium in, page 11, 2007. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.692, + 0.905, + 0.775 + ], + "angle": 0, + "content": "[47] Xincheng Yao, Ruoqi Li, Jing Zhang, Jun Sun, and Chongyang Zhang. Explicit boundary guided semi-push-pull contrastive learning for supervised anomaly detection. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 24490-24499, 2023. 1, 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.776, + 0.905, + 0.831 + ], + "angle": 0, + "content": "[48] Xincheng Yao, Ruoqi Li, Zefeng Qian, Lu Wang, and Chongyang Zhang. Hierarchical gaussian mixture normalizing flow modeling for unified anomaly detection. arXiv preprint arXiv:2403.13349, 2024. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.832, + 0.905, + 0.901 + ], + "angle": 0, + "content": "[49] Sangdoo Yun, Dongyoon Han, Seong Joon Oh, Sanghyuk Chun, Junsuk Choe, and Youngjoon Yoo. Cutmix: Regularization strategy to train strong classifiers with localizable features. In Proceedings of the IEEE/CVF international conference on computer vision, pages 6023-6032, 2019. 6" + }, + { + "type": "list", + "bbox": [ + 0.517, + 0.092, + 0.905, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.48, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "20425" + } + ], + [ + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.092, + 0.482, + 0.147 + ], + "angle": 0, + "content": "[50] Qihang Zhou, Guansong Pang, Yu Tian, Shibo He, and Jiming Chen. Anomalyclip: Object-agnostic prompt learning for zero-shot anomaly detection. arXiv preprint arXiv:2310.18961, 2023. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.149, + 0.482, + 0.218 + ], + "angle": 0, + "content": "[51] Jiawen Zhu, Choubo Ding, Yu Tian, and Guansong Pang. Anomaly heterogeneity learning for open-set supervised anomaly detection. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 17616-17626, 2024. 1, 2, 6, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.22, + 0.482, + 0.261 + ], + "angle": 0, + "content": "[52] Yuansheng Zhu, Wentao Bao, and Qi Yu. Towards open set video anomaly detection. In European Conference on Computer Vision, pages 395-412. Springer, 2022. 1, 2" + }, + { + "type": "list", + "bbox": [ + 0.093, + 0.092, + 0.482, + 0.261 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.48, + 0.946, + 0.519, + 0.957 + ], + "angle": 0, + "content": "20426" + } + ] +] \ No newline at end of file diff --git a/2025/Distribution Prototype Diffusion Learning for Open-set Supervised Anomaly Detection/5f6492f6-61fa-40e8-b92f-42fc55a187bf_origin.pdf b/2025/Distribution Prototype Diffusion Learning for Open-set Supervised Anomaly Detection/5f6492f6-61fa-40e8-b92f-42fc55a187bf_origin.pdf new file mode 100644 index 0000000000000000000000000000000000000000..183184dd00cb56da1e81be9ff814d640853cf50d --- /dev/null +++ b/2025/Distribution Prototype Diffusion Learning for Open-set Supervised Anomaly Detection/5f6492f6-61fa-40e8-b92f-42fc55a187bf_origin.pdf @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:a0c6b90ec53de35d704559081a7267fc5d51818df844858d83b6a6cdfc9ede1d +size 851200 diff --git a/2025/Distribution Prototype Diffusion Learning for Open-set Supervised Anomaly Detection/full.md b/2025/Distribution Prototype Diffusion Learning for Open-set Supervised Anomaly Detection/full.md new file mode 100644 index 0000000000000000000000000000000000000000..5c581fc72611224344e1d26224ded6781c9b79cd --- /dev/null +++ b/2025/Distribution Prototype Diffusion Learning for Open-set Supervised Anomaly Detection/full.md @@ -0,0 +1,365 @@ +# Distribution Prototype Diffusion Learning for Open-set Supervised Anomaly Detection + +Fuyun Wang $^{1}$ , Tong Zhang $^{1*}$ , Yuanzhi Wang $^{1}$ , Yide Qiu $^{1}$ , Xin Liu $^{2}$ , Xu Guo $^{1}$ , Zhen Cui $^{3*}$ $^{1}$ Nanjing University of Science and Technology, China. + $^{2}$ Nanjing SeetaCloud Technology, China. $^{3}$ Beijing Normal University, China +{fyw271828, tong.zhang, yuanzhiwang, 121106010824, guo.xu} @njust.edu.cn +xin.liu@seetacloud.com, zhen.cui@bnu.edu.cn + +# Abstract + +In Open-set Supervised Anomaly Detection (OSAD), the existing methods typically generate pseudo anomalies to compensate for the scarcity of observed anomaly samples, while overlooking critical priors of normal samples, leading to less effective discriminative boundaries. To address this issue, we propose a Distribution Prototype Diffusion Learning (DPDL) method aimed at enclosing normal samples within a compact and discriminative distribution space. Specifically, we construct multiple learnable Gaussian prototypes to create a latent representation space for abundant and diverse normal samples and learn a Schrödinger bridge to facilitate a diffusive transition toward these prototypes for normal samples while steering anomaly samples away. Moreover, to enhance inter-sample separation, we design a dispersion feature learning way in hyperspherical space, which benefits the identification of out-of-distribution anomalies. Experimental results demonstrate the effectiveness and superiority of our proposed DPDL, achieving state-of-the-art performance on 9 public datasets. + +# 1. Introduction + +Anomaly detection (AD) [15, 21, 48, 50] aims to identify outliers significantly diverging from the prevailing samples in a dataset, and has a wide range of applications like industrial inspection, medical image analysis, and scientific discovery, etc. Recently, unsupervised anomaly detection (UAD) [6, 11, 26, 28] and few-shot anomaly detection (FSAD) [12, 21, 22] have emerged as prominent research paradigms, emphasizing the modeling of normal sample distributions to discern anomalies effectively. Yet, these methods often neglect prior knowledge from limited anomaly samples, resulting in imprecise delineation of normal sample boundaries and reduced efficacy in differenti- + +ating normal from anomaly instances. On the contrary, supervised anomaly detection (SAD) [2, 20, 47] leverages a limited subset of anomaly samples as prior knowledge, improving detection performance. However, this reliance on seen anomalies poses a risk of overfitting and hampers generalization to unseen anomalies in real-world settings. + +To mitigate the challenge of limited generalization inherent in closed-set training, we focus on open-set supervised anomaly detection (OSAD) [1, 27, 51, 52], which utilizes a small set of known anomaly classes during training to identify unseen anomalies from open-set classes. By leveraging prior knowledge from observed samples, OSAD methods could reduce false positive errors. To improve the generalized detection of unseen anomalies, DRA [9] leverages data augmentation and outlier exposure to learn a decomposed anomaly representation comprising seen anomalies, pseudo-anomalies, and potential residual anomalies. BGAD [47] leverages decision boundaries derived from normalized flow models to capture and model anomaly information. Recently, AHL [51] simulates heterogeneous anomaly distributions and performs collaborative differentiable learning to further enhance the model's generality. + +While data augmentation and outlier exposure techniques [9, 19, 34] have demonstrated considerable success in anomaly detection, they fall short in generating comprehensive pseudo anomalies, capturing only a fraction of potential unseen anomalies. This limitation arises from overlooking the intricate nature of real-world anomaly distributions, thereby hindering the model's ability to generalize to novel anomaly types. Although AHL [51] has made strides in addressing this issue by simulating heterogeneous anomaly distributions, it still relies on approximating unknown out-of-distribution anomalies using known indistribution anomalies for generalization. Yet, three critical issues persist: i) the simulation mechanism cannot cover all anomaly distribution patterns due to the varied scales and structures of anomaly distributions; ii) simulated anom + +lies inherit in-distribution data biases, leading to suboptimal performance on out-of-distribution anomalies; iii) the diversity of normal samples presents dilemmas for existing methods, complicating the differentiation between normal and anomaly boundaries. These issues prompt a fundamental question: Instead of generating pseudo and uncertain anomaly samples, how can we accurately characterize compact distribution boundaries amidst a range of diverse normal samples and achieve robust generalization for unknown out-of-distribution anomalies? + +To address the aforementioned issue, in this work, we propose a Distribution Prototype Diffusion Learning (DPDL) method for open-set supervised anomaly detection. Considering abundant and diverse normal samples but very limited anomaly data, our method involves learning latent distribution prototypes, specifically multiple Gaussian distributions, onto which all observed normal samples can be effectively projected. To facilitate the mapping of normal samples into the prototype space, we leverage Schrödinger bridge (SB) framework, which enables a diffusive transition by aligning the distributions of these samples with the prototypes. The SB-based diffusion way could mitigate the out-of-distribution issue for normal samples to some extent. Within the distribution prototype space, we push observed anomaly samples away from normal samples to enhance discriminative capacity. Notably, both the prototypes and the diffusive bridge are learned jointly, resulting in a robust embedding space for normal samples. Moreover, to enhance generalization across unseen anomaly domains, we introduce a dispersion feature learning mechanism that maps intermediate features to a hyperspherical space, leveraging a mixture of von Mises-Fisher (vMF) distributions. This approach bolsters directional feature extraction and promotes robust inter-sample separation, facilitating effective identification of out-of-distribution samples. Experiments demonstrate that our method greatly improves detection capabilities for unseen anomalies. + +In summary, our contributions are three-fold: i) we propose a distribution prototype diffusion learning framework that jointly learns multiple Gaussian prototypes and the associated diffusion bridge, creating a compact and discriminative embedding space; ii) we develop dispersion feature learning in hyperspherical space to enhance inter-sample separation and improve generalization; iii) we achieve state-of-the-art performance in 9 public datasets, demonstrating the efficacy of our approach. + +# 2. Related Work + +Open-set Supervised Anomaly Detection. Open-set supervised anomaly detection (OSAD) seeks to develop a robust anomaly detection framework that generalizes from a limited set of training anomalies to effectively identify previously unseen anomalies within an open-set con + +text [1, 27, 35, 47, 52]. Leveraging the prior knowledge provided by observed anomalies, contemporary OSAD methods significantly mitigate false positive errors, thereby enhancing overall detection performance [9, 51]. Recently, DRA [9] learns disentangled representations of observed, pseudo, and residual anomalies to boost the detection of both seen and unseen anomalies. In contrast, AHL [51] simulates diverse heterogeneous anomaly distributions and employs collaborative differentiable learning, significantly improving the model's generalization capacity. + +Schrödinger Bridge. Schrödinger bridge (SB), widely recognized as the entropy-regularized optimal transport (OT) problem, involves learning a stochastic process that evolves from an initial probability distribution to a target distribution under the influence of a reference measure [5, 7, 16, 18, 23, 24, 38], which is often used for content generation as those prior generative models do [30, 42-45]. $\mathrm{I}^2$ -SB [25] and UNSB [14] learn a nonlinear diffusion process between two given distributions or represent the SB problem as a series of adversarial learning problems to realize the image transformation task. Recently, LightSB [10, 17] introduces a novel, fast, and simple SB solver, which achieves optimal matching in practice through the Gaussian mixture parameterization of the adjusted schrödinger potential. + +# 3. Preliminaries + +We focus on how to build the connection between two distributions $p_0$ and $p_1$ , where the distributions are defined as absolutely continuous Borel probability distributions with finite second-order moments. Building upon the foundation of entropy-regularized optimal transport (EOT) [5, 7, 10, 16-18, 23, 24, 33, 38], we review the related properties of EOT and the schrödinger bridge (SB) problem with a Wiener prior. + +Entropy-regularized optimal transport (EOT). Given two point sets $\mathcal{Z}_0$ and $\mathcal{Z}_1$ , we seek for the optimal transport cost between any two points $z_0 \in \mathcal{Z}_0$ and $z_1 \in \mathcal{Z}_1$ . This task may be formulated as an EOT problem with a parameter $\epsilon > 0$ , i.e., minimizing the following objective: + +$$ +\begin{array}{l} \min _ {\pi \in \Pi \left(p _ {0}, p _ {1}\right)} \left\{\int_ {\mathbb {R} ^ {D}} \int_ {\mathbb {R} ^ {D}} \frac {1}{2} \| z _ {0} - z _ {1} \| ^ {2} \pi \left(z _ {0}, z _ {1}\right) d z _ {0} d z _ {1} \right. \tag {1} \\ + \epsilon \mathrm {K L} (z \| p _ {0} \times p _ {1}) \}, \\ \end{array} +$$ + +where $\Pi(p_0, p_1)$ denotes the set of transport plans, i.e., joint probability distributions on $\mathbb{R}^D \times \mathbb{R}^D$ with marginals $p_0$ and $p_1$ , respectively, and KL denotes Kullback-Leibler divergence. The minimizer $\pi^*$ of Eqn. (1) is guaranteed to exist, be unique, and absolutely continuous, and is referred as the EOT plan. + +Schrodinger Bridge (SB). We define $\Omega$ as the space of $\mathbb{R}^D$ -valued functions over time $t \in [0,1]$ , representing + +trajectories in $\mathbb{R}^D$ that start at $t = 0$ and end at $t = 1$ . We denote the set of probability distributions over $\Omega$ , i.e., stochastic processes, by $\mathcal{P}(\Omega)$ . The differential of the standard Wiener process is represented by $dW_{t}$ . For a process $T\in \mathcal{P}(\Omega)$ , we denote its joint distribution at $t = 0,1$ by $\pi_T\in \mathcal{P}(\mathbb{R}^D\times \mathbb{R}^D)$ . Similarly, we use $T|_{z_0,z_1}$ to denote the distribution of $T$ for $t\in (0,1)$ , conditioned on $T$ 's values $z_{0}$ and $z_{1}$ at $t = 0$ and $t = 1$ , respectively. + +Let $W^{\epsilon} \in P(\Omega)$ represents a Wiener process with volatility $\epsilon > 0$ , starting from $p_0$ at $t = 0$ . Its differential is governed by the stochastic differential equation (SDE): $dW_{t}^{\epsilon} = \sqrt{\epsilon} dW_{t}$ . The Schrödinger bridge problem with the Wiener prior $W^{\epsilon}$ between $p_0$ and $p_1$ is minimizing: + +$$ +\min _ {T \in \mathcal {F} \left(p _ {0}, p _ {1}\right)} \operatorname {K L} \left(T \| W ^ {\epsilon}\right), \tag {2} +$$ + +where $\mathcal{F}(p_0,p_1)\subset \mathcal{P}(\Omega)$ denotes the subset of stochastic processes that begin with distribution $p_0$ at $t = 0$ and reach $p_1$ at $t = 1$ . This problem has a unique solution, a SDE diffusion process $T^{*}$ defined: $dZ_{t} = g^{*}(Z_{t},t)dt + dW_{t}^{\epsilon}$ . The process $T^{*}$ is referred to as SB, and $g^{*}:\mathbb{R}^{D}\times [0,1]\to \mathbb{R}^{D}$ is the optimal drift. + +Characterization of solutions. The EOT plan $\pi^{*} = \pi^{T^{*}}$ takes a specific form [10, 17]: + +$$ +\pi^ {*} \left(z _ {0}, z _ {1}\right) = u ^ {*} \left(z _ {0}\right) \exp \left(\frac {- \| z _ {0} - z _ {1} \| ^ {2}}{2 \epsilon}\right) v ^ {*} \left(z _ {1}\right), \tag {3} +$$ + +where $u^{*}, v^{*} \colon \mathbb{R} \to \mathbb{R}_{+}$ are measurable functions known as Schrödinger potentials. The optimal drift $g^{*}$ is derived as: + +$$ +g ^ {*} (z, t) = \epsilon \nabla_ {z} \log \int_ {\mathbb {R} ^ {D}} \mathcal {N} \left(z ^ {\prime} \mid z, (1 - t) \epsilon I _ {D}\right) v ^ {*} \left(z ^ {\prime}\right) d z ^ {\prime}, \tag {4} +$$ + +# 4. Method + +# 4.1. Problem Formulation + +Let $\mathcal{X}_{\mathrm{tr}} = \{(\mathbf{x}_i, y_i)\}$ denotes a weakly-supervised training set with only image-level labels, where $\mathbf{x}_i$ denotes one RGB image and $y_i \in \{0,1\}$ denotes whether $\mathbf{x}_i$ is an anomaly sample (anomaly: $y_i = 1$ , normal: $y_i = 0$ ). Hereby, $\mathcal{X}_{\mathrm{tr}}$ is consistent of a normal subset $\mathcal{X}_{\mathrm{tr}}^{\mathrm{n}}$ ( $|\mathcal{X}_{\mathrm{tr}}^{\mathrm{n}}| = N$ ) and an anomaly subset $\mathcal{X}_{\mathrm{tr}}^{\mathrm{a}}$ ( $|\mathcal{X}_{\mathrm{tr}}^{\mathrm{a}}| = M$ ), formally, $\mathcal{X}_{\mathrm{tr}} \doteq \mathcal{X}_{\mathrm{tr}}^{\mathrm{n}} \cup \mathcal{X}_{\mathrm{tr}}^{\mathrm{a}}$ , where generally $N \gg M$ . Given a testing set $\mathcal{X}_{\mathrm{te}}$ , we need to predict whether one sample $\mathbf{x} \in \mathcal{X}_{\mathrm{te}}$ is anomaly or normal. In OSAD, the anomaly patterns of the testing set do less recurred in those encountered training set. In other words, the distribution of anomalies are obviously discrepant, i.e., $P(\mathcal{X}_{\mathrm{te}}^{\mathrm{a}}) \neq P(\mathcal{X}_{\mathrm{tr}}^{\mathrm{a}})$ . Hereby, we need to learn a robust anomaly detection model $\psi$ from the training set $\mathcal{X}_{\mathrm{tr}}$ so that $\psi$ accurately infers anomaly scores for test samples. + +Our abstract idea is to learn latent distribution prototypes that not only encapsulate normal samples in a concise manner but also discriminate against anomaly samples. Given the abundance of observed normal samples, one + +natural approach is to characterize the distribution $P(\mathcal{X}_{\mathrm{tr}}^{\mathrm{n}})$ where samples outside this distribution, $\mathbf{x}\notin P(\mathcal{X}_{\mathrm{tr}}^{\mathrm{n}})$ , would be awarded higher probabilities as anomalies. Considering the inherent diversity of normal samples, we endeavor to learn multiple simple distributions (e.g., Gaussians) as prototypes $\mathcal{P}_{\mathrm{MGP}}$ , named multi-Gaussian prototypes (MGP). To embed input data into the prototype space, we introduce a generative bridge model $\psi_{p}$ for distribution transformation. The more abstract formulation is given as follows: + +$$ +\min _ {\psi_ {p}, \mathcal {P} _ {\mathrm {M G P}}, f} D _ {p} \left(\psi_ {p} \left(\mathcal {F} _ {\mathrm {t r}} ^ {\mathrm {n}}\right), \psi_ {p} \left(\mathcal {F} _ {\mathrm {t r}} ^ {\mathrm {a}}\right), \mathcal {P} _ {\mathrm {M G P}}\right) + \lambda D _ {s} \left(\mathcal {F} _ {\mathrm {t r}} ^ {\mathrm {n}}, \mathcal {F} _ {\mathrm {t r}} ^ {\mathrm {a}}\right), \tag {5} +$$ + +$$ +s. t., \psi_ {p}: P (\mathcal {F}) \stackrel {\text {b r i d g e}} {\Longrightarrow} \mathcal {P} _ {\mathrm {M G P}}, \tag {6} +$$ + +$$ +f: \mathcal {X} \xrightarrow {\text {f e a t u r e}} \mathcal {F}, \tag {7} +$$ + +where $\mathcal{P}_{\mathrm{MGP}}$ denotes the distribution prototypes to be learned, $\psi_{p}$ is the flow function across probability distributions, $D_{p}$ signifies the discriminative function in the space of prototypes, $f$ stands for the feature extraction process, and $D_{s}$ acts as a regularizer to increase feature discriminability. In the above formulation, besides distribution prototype learning (DPL), we also introduce dispersion feature learning (DFL) executed within a hyperspherical embedding space, i.e., $D_{s}(\mathcal{F}_{\mathrm{tr}}^{\mathrm{n}},\mathcal{F}_{\mathrm{tr}}^{\mathrm{a}})$ . The advantage of DFL is to prevent abrupt feature collapse in feature learning, thus preserving discriminative qualities. The overview of our framework is abstractly depicted in Fig. 1. Detailed elaboration on these aspects will be provided in the subsequent sections. + +# 4.2. Distribution Prototype Learning + +Distribution prototype learning is operated in the feature space. Hence, to extract intermediate image features, we can leverage those classic networks as the backbone, such as ResNet-18. Formally, the feature extraction process, denoted as a function $f: \mathcal{X} \to \mathcal{F}$ , transforms an image $\mathbf{x}$ into the intermediate feature $\mathbf{x} = f(\mathbf{x}) \in \mathbb{R}^d$ , where a flatten operation is used on convolution maps for vectorization. Consequently, we designate the intermediate feature sets of normal and anomaly samples as $\mathcal{F}_{\mathrm{tr}}^{\mathrm{n}} = \{\mathbf{x}_i^{\mathrm{n}} | i = 1, \dots, N\}$ and $\mathcal{F}_{\mathrm{tr}}^{\mathrm{a}} = \{\mathbf{x}_j^{\mathrm{a}} | i = 1, \dots, M\}$ , respectively. + +Considering the rarity of anomalies and the diversity of normal samples, it is pertinent to capture the intricate distributions of normal samples by mapping the distribution $p_0 = P(\mathcal{F}_{\mathrm{tr}}^{\mathrm{n}})$ of intermediate features to a distinct and well-characterized distribution $p_1$ like Gaussian. But due to various normal samples, we opt for a multi-Gaussian prototypes (MGP) comprising multiple Gaussian distributions as prototypes $\mathcal{P}_{\mathrm{MGP}} = \{\mathcal{P}_i \doteq \mathcal{N}(\pmb{\mu}_i, \pmb{\sigma}_i) | i = 1, \dots, C\}$ . Given a normal sample $\mathbf{x}_i^{\mathrm{n}}$ , we expect to align it with the closest prototype with high likelihood. However, in the open-set setting, it is challenge to transform unseen points $\mathbf{x} \sim p_0$ to the target distribution $p_1$ . Drawing inspiration from the capability of diffusion generation models in aligning disparate distributions, we frame the transition from source + +![](images/e170339ed7ef8b2503fd8d411bf6288fe8634990f0c3b92f2d8db4a2dec08fe9.jpg) +Figure 1. Our proposed DPDL framework. It comprises three distinct modules: Distribution Prototype Learning (DPL, Sec. 4.2), Dispersion Feature Learning (DFL, Sec. 4.3), and anomaly score prediction (Sec. 4.4). DPL transforms the distribution of normal samples to a space of learnable multiple Gaussian prototypes through building schrödinger bridge, meantime pushing anomaly distribution away from these prototypes. DFL operates in a hyperspherical space, enlarging the distances of intermediate features of all samples in a hyperspherical space to strengthen feature generalization for detecting anomalies. The score prediction module leverages a multi-instance-learning method to compute anomaly scores. + +domain distribution $p_0$ to target domain distribution $p_1$ as a Schrodinger bridge problem. As formulated in Eqn. (6), we need learn an essential bridge flow $\psi_p: P(\mathcal{F}_{\mathrm{tr}}^{\mathrm{n}}) \xrightarrow{\mathrm{bridge}} \mathcal{P}_{\mathrm{MGP}}$ . After bridge transformation, the condition probability in optimal transport plan conforms to: + +$$ +\pi \left(\psi_ {p} (\mathbf {x}) | \mathbf {x}\right) \propto \underbrace {\sum_ {c = 1} ^ {C} \alpha_ {c} \mathcal {N} \left(\psi_ {p} (\mathbf {x}) ; \boldsymbol {\mu} _ {c} , \boldsymbol {\sigma} _ {c}\right)} _ {\dot {\equiv} \phi_ {1} \left(\psi_ {p} (\mathbf {x})\right)}, \tag {8} +$$ + +where the parameters $\{\alpha_c,\pmb {\mu}_c,\pmb {\sigma}_c\}_{c = 1}^C$ (known as Gaussian mixed model (GMM), abstracted into the function $\phi_1$ ) as well as the flow function $\psi_{p}$ (in bridge) need to be learned. For simplicity, below we adopt diagonal matrices for $\pmb{\sigma}_{c}$ + +Drawing inspiration from previous works [10, 17], we reframe distribution prototype learning in Eqn. (8) as the Schrödinger bridge. Given the specific form taken by the EOT plan described in Eqn. (3), we redefine these measurable functions $u, v: \mathbb{R} \to \mathbb{R}_+$ , termed Schrödinger potentials [10, 17], as follows: + +$$ +u \left(\mathbf {x} _ {i} ^ {\mathrm {n}}\right) \doteq \exp \left(\frac {\left\| \mathbf {x} _ {i} ^ {\mathrm {n}} \right\| ^ {2}}{2 \epsilon}\right) \phi_ {0} \left(\mathbf {x} _ {i} ^ {\mathrm {n}}\right), \tag {9} +$$ + +$$ +v \left(\psi \left(\mathbf {x} _ {i} ^ {\mathrm {n}}\right)\right) \doteq \exp \left(\frac {\| \psi \left(\mathbf {x} _ {i} ^ {\mathrm {n}}\right) \| ^ {2}}{2 \epsilon}\right) \phi_ {1} \left(\psi \left(\mathbf {x} _ {i} ^ {\mathrm {n}}\right)\right), \tag {10} +$$ + +where $\phi_0$ is defined within the source feature domain $\mathcal{F}_{\mathrm{tr}}$ , and $\phi_{1}$ is defined in Eqn. (8). $\epsilon$ is set to 0.001 in our experiments. Accordingly, Eqn. (3) can be converted to: + +$$ +\pi \left(\mathbf {x} _ {i} ^ {\mathrm {n}}, \psi \left(\mathbf {x} _ {i} ^ {\mathrm {n}}\right)\right) = \phi_ {0} \left(\mathbf {x} _ {i} ^ {\mathrm {n}}\right) \exp \left(\frac {\left\langle \mathbf {x} _ {i} ^ {\mathrm {n}} , \psi \left(\mathbf {x} _ {i} ^ {\mathrm {n}}\right) \right\rangle}{\epsilon}\right) \phi_ {1} \left(\psi \left(\mathbf {x} _ {i} ^ {\mathrm {n}}\right)\right), \tag {11} +$$ + +Hence, the condition probability of the transport plan in Eqn. (8) could be exactly defined as + +$$ +\pi (\psi (\mathbf {x} _ {i} ^ {\mathrm {n}}) | \mathbf {x} _ {i} ^ {\mathrm {n}}) \doteq \eta (\mathbf {x} _ {i} ^ {\mathrm {n}}, \psi (\mathbf {x} _ {i} ^ {\mathrm {n}})) \phi_ {1} (\psi (\mathbf {x} _ {i} ^ {\mathrm {n}})), \tag {12} +$$ + +where the connection factor is denoted as $\eta (\mathbf{x}_i^n,\psi (\mathbf{x}_i^n)) = \frac{1}{\varpi(\mathbf{x}_i^n)}\exp \left(\frac{\langle\mathbf{x}_i^n,\psi(\mathbf{x}_i^n)\rangle}{\epsilon}\right)$ with the normalization term $\varpi (\mathbf{x}_i^n) = \int \exp (\langle \mathbf{x}_i^n,\psi (\mathbf{x}_i^n)\rangle)\phi_1(\psi (\mathbf{x}_i^n))d\psi (\mathbf{x}_i^n)$ . According to Eqns. (8) and (12), we can further derive a more tractable form [10, 17]: + +$$ +\pi \left(\psi \left(\mathbf {x} _ {i} ^ {\mathrm {n}}\right) \mid \mathbf {x} _ {i} ^ {\mathrm {n}}\right) = \widetilde {\eta} \left(\mathbf {x} _ {i} ^ {\mathrm {n}}\right) \sum_ {c = 1} ^ {C} \widetilde {\alpha} _ {c} \left(\mathbf {x} _ {i} ^ {\mathrm {n}}\right) \mathcal {N} \left(\psi \left(\mathbf {x} _ {i} ^ {\mathrm {n}}\right); \widetilde {\boldsymbol {\mu}} _ {c} \left(\mathbf {x} _ {i} ^ {\mathrm {n}}\right), \boldsymbol {\sigma} _ {c}\right), \tag {13} +$$ + +where the normalization factor is denoted as $\widetilde{\eta} (\mathbf{x}_i^n) = 1 / \sum_{c = 1}^{C}\widetilde{\alpha}_c(\mathbf{x}_i^n)$ , the coefficients of multi-Gaussian are defined as $\widetilde{\alpha}_{c}(\mathbf{x}_{i}^{n}) = \alpha_{c}\exp (\frac{1}{2} (\mathbf{x}_{i}^{n})^{\intercal}\pmb{\sigma}_{c}\mathbf{x}_{i}^{n} + \frac{1}{\epsilon} (\widetilde{\mu}_{c})^{\intercal}(\mathbf{x}_{i}^{n}))$ , and the mean vectors are calculated as $\widetilde{\pmb{\mu}}_c(\mathbf{x}_i^n) = \pmb {\mu}_c + \frac{1}{\epsilon}\pmb {\sigma}_c\mathbf{x}_i^n$ + +We proceed with deriving the bridge function $\psi_{p}$ that represents a SDE process: $d\mathbf{x}_t = g(x_t,t)dt + \sqrt{\epsilon} dW_t$ where the shift function $g$ is solved. According to Eqn. (4), we can obtain the shift function of diffusion process as follows: + +$$ +g (\mathbf {x} _ {i} ^ {\mathrm {n}}, t) = \epsilon \rho_ {\mathrm {M G P}} (\mathbf {x} _ {i} ^ {\mathrm {n}}) \nabla_ {\mathbf {x} _ {i} ^ {\mathrm {n}}} \log (\mathcal {N} (\mathbf {x} _ {i} ^ {\mathrm {n}} | 0, \epsilon (1 - t)) \mathbf {I}), \quad (1 4) +$$ + +where the coefficients $\rho_{\mathrm{MGP}}$ defined on multiple distribution prototypes is calculated as: $\rho_{\mathrm{MGP}}(\mathbf{x}_i^{\mathrm{n}}) = \sum_{c=1}^{C} (\alpha_c \mathcal{N}(\widetilde{\boldsymbol{\mu}}_c(\mathbf{x}_i^{\mathrm{n}}) | \mathbf{0}, \boldsymbol{\sigma}_c) \mathcal{N}(h_c(\mathbf{x}_i^{\mathrm{n}}, t) | 0, \boldsymbol{\Sigma}_c^t))$ , with another Gaussian of $h_c(\mathbf{x}_i^{\mathrm{n}}, t) = \frac{1}{\epsilon(1 - t)} \mathbf{x}_i^{\mathrm{n}} + \boldsymbol{\sigma}_c^{-1} \widetilde{\boldsymbol{\mu}}_c(\mathbf{x}_i^{\mathrm{n}})$ and $\boldsymbol{\Sigma}_c^t = \frac{t}{\epsilon(1 - t)} \mathbf{I} + \boldsymbol{\sigma}_c^{-1}$ . + +Distribution prototypes initialization. Jointly learning prototypes $\{\alpha_c, \mu_c, \sigma_c\}_{c=1}^C$ and the bridge transformation $\psi_p$ (also the shift $g$ ) is a challenging task as they are interdependent. For this, we leverage a vector quantization function to learn a codebook $\mathcal{E}$ of prototypes within a discrete latent space from training data. Specifically, given an input image feature $\mathbf{x}_i^n$ , we can assign $\mathbf{x}_i^n$ to the closest prototype $\mathbf{e}_k$ by minimizing the L2 distance between $\mathbf{x}_i^n$ and each of prototypes $\mathbf{e}_c \in \mathcal{E}$ , as follows: + +$$ +\min _ {\left\{\mathbf {e} _ {c} \right\}} \mathbb {E} _ {\mathbf {x} _ {i} ^ {\mathrm {n}} \in \mathcal {F} _ {\mathrm {t r}}} \left[ \left\| \mathbf {x} _ {i} ^ {\mathrm {n}} - \mathbf {e} _ {c ^ {*}} \right\| _ {2} ^ {2} \right], \text {s . t .}, c ^ {*} = \underset {c} {\arg \min } \left\| \mathbf {x} _ {i} ^ {\mathrm {n}} - \mathbf {e} _ {c} \right\| _ {2}, \tag {15} +$$ + +where $c^*$ denotes the index of the prototype closest to $\mathbf{x}_i^{n}$ . The learned $\{\mathbf{e}_c\}_{c=1}^C$ are used to initialize the mean vectors $\{\pmb{\mu}_c\}_{c=1}^C$ , while the variances of all prototypes are set to the identity matrix. We observe that this initialization strategy accelerates the training process. + +Distribution loss of normal and anomaly samples. As $p_0$ and $p_1$ are accessible only via samples $\mathcal{F}_{\mathrm{tr}}$ and prototypes $\mathcal{P}_{\mathrm{MGP}}$ , we optimize the empirical form in Eqn. (2) for normal samples as follows [10, 17]: + +$$ +\mathcal {L} _ {\mathrm {D P L}} ^ {\mathrm {n}} = \frac {1}{N} \sum_ {i = 1} ^ {N} \log \varpi_ {\theta} \left(\mathbf {x} _ {i} ^ {\mathrm {n}}\right) - \frac {1}{C} \sum_ {c = 1} ^ {C} \log \phi_ {1} \left(\boldsymbol {\mu} _ {c}\right), \tag {16} +$$ + +In contrast, for anomaly samples, we aim to push anomaly distribution away from $p_1$ , i.e., negative loss, formally, + +$$ +\mathcal {L} _ {\mathrm {D P L}} ^ {\mathrm {a}} = \frac {1}{C} \sum_ {c = 1} ^ {C} \log \phi_ {1} (\boldsymbol {\mu} _ {c}) - \frac {1}{M} \sum_ {i = 1} ^ {M} \log \varpi_ {\theta} \left(\mathbf {x} _ {i} ^ {\mathrm {a}}\right), \tag {17} +$$ + +# 4.3. Dispersion Feature Learning + +A critical challenge in OSAD is detecting previously unseen anomalies in open-set environments. Due to the limited anomaly observations, existing methods often use pseudo-anomaly generation strategies as a means of effective data augmentation. Nonetheless, the effectiveness of these methods heavily depends on the quality of the pseudo-anomaly feature embeddings. Notably, pseudo-anomaly distributions often inherit biases from the in-distribution data, which differ from the unknown anomaly distributions in out-of-distribution data. Current methods fail to address the relationship between observed and unknown anomalies, particularly for out-of-distribution generalization. + +To improve out-of-distribution detection, it is essential to promote a larger inter-sample dispersion, as greater distances among in-distribution samples facilitate their more effective separation from out-of-distribution samples. In other words, if all inter-sample distances are close to zero, i.e., collapse to a single point, it becomes impossible to differentiate between samples. Hence, promoting separability through a larger inter-sample dispersion is critical for accurately identifying samples that do not belong to known categories. To do so, we map the features into a hyperspherical + +space. Draws inspiration from the vMF distribution [31] in directional statistics, we compute spherical Gaussian distributions for unit-norm features $\widehat{\mathbf{x}}_i = \mathbf{x}_i / \| \mathbf{x}_i\| _2^2$ . The probability density function of a unit vector $\widehat{\mathbf{x}}_i\in \mathbb{R}^D$ is defined in the hyperspherical space as follows: + +$$ +p _ {D} (\widehat {\mathbf {x}} _ {i}; \widehat {\mathbf {x}} _ {j}, \kappa) = F _ {D} (\kappa) \exp \left(\kappa \langle \widehat {\mathbf {x}} _ {i}, \widehat {\mathbf {x}} _ {j} \rangle\right), \tag {18} +$$ + +where $\kappa \geq 0$ controls the concentration of the distribution around the mean direction $\widehat{\mathbf{x}}_j$ and $F_{D}(\kappa)$ is the normalization factor. A larger $\kappa$ value increases concentration around the mean, while in the extreme case $\kappa = 0$ , sample points are uniformly distributed on the hypersphere. Therefore, we design a dispersion loss to optimize large angular distances between the features of all samples: + +$$ +\mathcal {L} _ {\mathrm {D F L}} = \frac {1}{U} \sum_ {i = 1} ^ {U} \log \frac {1}{U - 1} \sum_ {i, j = 1} ^ {U} \mathbb {1} \{i \neq j \} \exp (\kappa \langle \widehat {\mathbf {x}} _ {i}, \widehat {\mathbf {x}} _ {j} \rangle), \tag {19} +$$ + +where $U = N + M$ , and $\kappa$ is set to 10 in our experiments. + +# 4.4. Anomaly Score Prediction + +Based on the above designs, we leverage the multiple-instance-learning (MIL)-based method proposed in [35] to effectively learn anomaly scores. Similar to the work [9], we design three modules $\mathsf{M} = \{\mathsf{M}_{\mathsf{a}},\mathsf{M}_{\mathsf{n}}$ and $\mathsf{M}_{\mathsf{r}}\}$ for estimating anomaly scores. Firstly, for the feature map $\mathbf{x}_i^1$ , we generate pixel-wise feature vectors $\mathcal{V} = \{\mathbf{v}_i\}_{i = 1}^{H' \times W'}$ to represent the feature of small patches of the image $\mathbf{x}_i$ , where $(H',W')$ denotes the size of the feature map. These pixel-wise representations are then mapped by an anomaly classifier $S_{\mathrm{a}}$ to estimate pixel-level anomaly scores. To capture those points with the most salient anomalies, we compute the top- $K$ most anomaly pixel points and define the loss function as: + +$$ +\mathcal {L} _ {\mathrm {M} _ {\mathrm {a}}} \left(\mathbf {x} _ {i}, y _ {i}\right) = \mathcal {L} _ {\text {b i n a r y}} \left(\frac {1}{K} \sum \operatorname {T o p} _ {K} \left\{S _ {\mathrm {a}} \left(\mathbf {v} _ {i}; \theta_ {\mathrm {a}}\right) \right\}, y _ {i}\right), \tag {20} +$$ + +where $\mathcal{L}_{\mathrm{binary}}$ refers to a binary classification loss function, and $\operatorname{Top}_K$ selects the highest $K$ anomaly scores among all the vectors. We use $\mathsf{M}_{\mathfrak{n}}$ to learn the normal features: + +$$ +\mathcal {L} _ {\mathrm {M} _ {\mathrm {n}}} \left(\mathbf {x} _ {i}, y _ {i}\right) = \mathcal {L} _ {\text {b i n a r y}} \left(S _ {\mathrm {n}} \left(\frac {1}{H ^ {\prime} \times W ^ {\prime}} \sum_ {i = 1} ^ {H ^ {\prime} \times W ^ {\prime}} \mathbf {v} _ {i}; \theta_ {\mathrm {n}}\right), y _ {i}\right), \tag {21} +$$ + +where $S_{\mathfrak{n}}:\mathcal{V}\to \mathbb{R}$ is a fully connected binary anomaly classifier. Finally, we define $\mathsf{M}_{\mathfrak{x}}$ to compute the residual anomaly scores between fine-grained visual semantics and abstract prototypes: + +$$ +\mathcal {L} _ {\mathrm {M} _ {\mathrm {r}}} = \mathcal {L} _ {\text {b i n a r y}} \left(S _ {\mathrm {r}} \left(\left(\psi_ {p} (\mathbf {x} _ {i}) - \boldsymbol {\mu} _ {c ^ {*}}\right) / \boldsymbol {\sigma} _ {c ^ {*}}; \theta_ {\mathrm {r}}\right), y _ {i}\right), \tag {22} +$$ + +where $c^*$ denotes the index of the most probable prototypes, i.e., $c^* = \arg \max_c\mathcal{N}(\psi_p(\mathbf{x}_i);\pmb {\mu}_c,\pmb {\sigma}_c)$ , and $S_{\mathrm{r}}$ utilize the same method to obtain anomaly score as $S_{\mathrm{a}}$ . + +Training. During the training phase, the SB and three prediction modules are jointly trained. To this end, we employ an objective function that encompasses three components as follows: + +$$ +\mathcal {L} = \underbrace {\mathcal {L} _ {\mathrm {M} _ {\mathrm {a}}} + \mathcal {L} _ {\mathrm {M} _ {\mathrm {n}}} + \mathcal {L} _ {\mathrm {M} _ {\mathrm {r}}}} _ {\text {M I L - b a s e d l e a n i n g}} + \underbrace {\mathcal {L} _ {\mathrm {D P L}} ^ {\mathrm {n}} + \mathcal {L} _ {\mathrm {D P L}} ^ {\mathrm {a}}} _ {\text {S B t r a n s f o r m}} + \underbrace {\lambda \mathcal {L} _ {\mathrm {D F L}}} _ {\text {d i s p e r s i o n}}. \tag {23} +$$ + +where the coefficient $\lambda$ modulates the relative importance of dispersion loss, and the learnable parameters include $\{\alpha_{c},\pmb{\mu}_{c},\pmb{\sigma}_{c}\}$ , $\theta_{\psi_p}$ , and $\{\theta_{\mathrm{a}},\theta_{\mathrm{n}},\theta_{\mathrm{r}}\}$ . + +Inference. During the test phase, we find the most similar class prototype through SB, and subsequently compute the anomaly score by adding the scores from both $S_{\mathrm{a}}$ and $S_{\mathrm{r}}$ , while subtracting the normal score obtained from $S_{\mathrm{n}}$ for the given test image. + +# 5. Experiment + +# 5.1. Dataset and Evaluation Metric + +Dataset To validate the effectiveness of DPDL, comprehensive experiments are conducted on nine real-world AD datasets, including six industrial defect detection datasets (MVTec AD [3], Optical [46], SDD [40], AITEX [39], ELPV [8], Mastcam [13]) and three medical image datasets (Hyper-Kvasir [4], Brain-MRI [37], HeadCT [37]). We follow the previous OSAD baselines [9, 51] to adopt two protocols for sampling, including general setting and hard setting. The general setting assumes that anomaly examples are randomly sampled from the anomaly class, while the hard setting samples from a single class to assess generalization to new or unseen anomaly classes. + +Evaluation Metric We utilize the widely adopted Area Under ROC Curve (AUC) as a metric to evaluate the performance across all methods and settings. All reported AUCs are averaged results over five independent runs. + +# 5.2. Baselines + +We compare DPDL against six related state-of-the-art OSAD baselines, including SAOE [19, 32, 41], MLEP [27], FLOS [36], DevNet [35], DRA [9], and AHL [51]. MLEP, DevNet, DRA, and AHL are specifically designed for OSAD. SAOE is a supervised detector enhanced with synthetic anomalies and anomaly exposure, whereas FLOS is an imbalanced classifier leveraging focal loss. + +# 5.3. Implementation Details + +The input image size is $448 \times 448 \times 3$ . We set $K$ in the top- $K$ MIL to $10\%$ of the number of all scores per score map. + +AdamW optimizer [29] is used for the parameter optimization using an initial learning rate $2 \times 10^{-4}$ with a weight decay of $1 \times 10^{-5}$ . DPDL is trained on one NVIDIA GeForce RTX 4090 GPU, which are trained using 50 epochs, with 20 iterations per epoch. Following previous protocol [9, 51], we evaluate performance with anomaly sample numbers of $M = 10$ and $M = 1$ , and for robust detection of unseen anomalies, we use CutMix [49] to create pseudo-anomaly samples as augmented data for known anomalies. The prototype quantity $C$ is set to 32 as default. + +# 5.4. Results under General Setting + +Tab. 1 highlights DPDL's strong performance. In the challenging scenario of single-anomaly detection, DPDL improves the performance of AHL [51] by more than $8.3\%$ on the AITEX, ELPV, and Mastcam datasets. Furthermore, it achieves significant improvements across six additional datasets, which suggest the effective utilization of few-shot anomaly examples in DPDL, while mitigating overfitting to the seen anomalies. When shifting to ten anomaly examples settings, DPDL continues to maintain a significant lead with over $5.4\%$ improvement on those datasets. Given the rich and diverse set of normal samples in these datasets, DPDL leverages the DPL component to encapsulate these samples within a compact, discriminative distribution space, while effectively pushing anomalous samples outside this space, thereby enabling accurate anomaly detection. In the setting with ten abnormal samples, although existing methods have reached performance saturation on the MVTecAD, Optical, and SDD datasets, DPDL still has a certain lead, demonstrating the strong ability of DPL and DFL in learning tight boundaries of normal sample distributions and generalizing to previously unseen anomaly domains. Furthermore, when evaluated on the medical datasets BrainMRI and HeadCT, DPDL demonstrates competitive performance despite these datasets being notably small in scale and containing only a single class. This highlights the algorithm's ability to deliver robust results even in data-scarce conditions. + +# 5.5. Results under the Hard Setting + +Tab. 2 summarizes the performance comparison under the hard setting. It is evident that DPDL achieves the highest AUC scores in both the single-anomaly and ten-anomaly sample settings. Specifically, compared to the closest competing method, AHL [51], DPDL achieves an improvement in AUC scores ranging from $0.6\%$ to $7.9\%$ in the ten anomaly examples settings and from $0.5\%$ to $8.9\%$ in the one anomaly example settings, respectively. The observed improvement can be attributed to the strong generalization ability of DPDL in detecting unseen anomaly classes, even when the model is trained on only a single anomaly class. + +Table 1. AUC performance (mean ± std) across nine real-world AD datasets is reported under the general setting. red highlights the best results, and blue indicates sub-optimal outcomes. All baseline SOTA results are sourced from the original papers [9, 51]. + +
DatasetDevNetFLOSSAOEMLEPDRAAHLDPDL (Ours)
Ten Training Anomaly Examples
MVTec AD0.945±0.0040.939±0.0070.926±0.0100.907±0.0050.959±0.0030.970±0.0020.977±0.002
Optical0.782±0.0650.720±0.0550.941±0.0130.740±0.0390.965±0.0060.976±0.0040.983±0.005
SDD0.988±0.0060.967±0.0180.955±0.0200.983±0.0130.991±0.0050.991±0.0010.996±0.001
AITEX0.887±0.0130.841±0.0490.874±0.0240.867±0.0370.893±0.0170.925±0.0130.975±0.007
ELPV0.846±0.0220.818±0.0320.793±0.0470.794±0.0470.845±0.0130.850±0.0040.937±0.003
Mastcam0.790±0.0210.703±0.0290.810±0.0290.798±0.0260.848±0.0080.855±0.0050.934±0.010
Hyper-Kvasir0.829±0.0180.773±0.0290.666±0.0500.600±0.0690.834±0.0040.880±0.0030.939±0.005
BrainMRI0.958±0.0120.955±0.0110.900±0.0410.959±0.0110.970±0.0030.977±0.0010.969±0.005
HeadCT0.982±0.0090.971±0.0040.935±0.0210.972±0.0140.972±0.0020.999±0.0030.981±0.003
One Training Anomaly Example
MVTec AD0.780±0.0200.755±0.1360.834±0.0070.744±0.0190.883±0.0080.901±0.0030.927±0.002
Optical0.523±0.0030.518±0.0030.815±0.0140.516±0.0090.888±0.0120.888±0.0070.915±0.002
SDD0.881±0.0090.840±0.0430.781±0.0090.811±0.0450.859±0.0140.909±0.0010.917±0.003
AITEX0.598±0.0700.538±0.0730.675±0.0940.564±0.0550.692±0.1240.734±0.0080.838±0.008
ELPV0.514±0.0760.457±0.0560.635±0.0920.578±0.0620.675±0.0240.828±0.0050.897±0.002
Mastcam0.595±0.0160.542±0.0170.662±0.0180.625±0.0450.692±0.0580.743±0.0030.838±0.011
Hyper-Kvasir0.653±0.0370.668±0.0040.498±0.1000.445±0.0400.690±0.0170.768±0.0150.821±0.007
BrainMRI0.694±0.0040.693±0.0360.531±0.0600.632±0.0170.744±0.0040.866±0.0040.893±0.004
HeadCT0.742±0.0760.698±0.0920.597±0.0220.758±0.0380.796±0.1050.825±0.0140.865±0.005
+ +Table 2. AUC results (mean ± std) under the hard setting. The best and second-best results are highlighted in red and blue, respectively. Carpet and Metal_nut are subsets of MVTec AD. The datasets used are consistent with those in [9, 51], where those datasets only containing one anomaly class are excluded to adapt for the hard setting. + +
DatasetDevNetFLOSSAOEMLEPDRAAHLDPDL (Ours)
Ten Training Anomaly Examples
Carpet (mean)0.847±0.0170.761±0.0120.762±0.0730.751±0.0230.935±0.0130.949±0.0020.956±0.004
Metal_nut (mean)0.965±0.0110.922±0.0140.855±0.0160.878±0.0580.945±0.0170.972±0.0020.978±0.002
AITEX (mean)0.683±0.0320.635±0.0430.724±0.0320.626±0.0410.733±0.0090.747±0.0020.798±0.005
ELPV (mean)0.702±0.0230.642±0.0320.683±0.0470.745±0.0200.766±0.0290.788±0.0030.818±0.003
Mastcam (mean)0.588±0.0110.616±0.0210.697±0.0140.588±0.0160.695±0.0040.721±0.0030.778±0.007
Hyper-Kvasir (mean)0.822±0.0190.786±0.0210.698±0.0210.571±0.0140.844±0.0090.854±0.0040.864±0.002
One Training Anomaly Example
Carpet (mean)0.767±0.0180.678±0.0400.753±0.0550.679±0.0290.901±0.0060.932±0.0030.941±0.006
Metal_nut (mean)0.855±0.0160.855±0.0240.816±0.0290.825±0.0230.932±0.0170.939±0.0040.944±0.003
AITEX (mean)0.646±0.0340.624±0.0240.674±0.0340.466±0.0300.684±0.0330.707±0.0070.753±0.005
ELPV (mean)0.648±0.0570.691±0.0080.614±0.0480.566±0.1110.703±0.0220.740±0.0030.762±0.003
Mastcam (mean)0.511±0.0130.524±0.0130.689±0.0370.541±0.0070.667±0.0120.673±0.0100.733±0.004
Hyper-Kvasir (mean)0.595±0.0230.571±0.0040.406±0.0180.480±0.0440.700±0.0090.706±0.0070.715±0.004
+ +![](images/fdbc1f8c0719b28f4506c3c34a1813611aaaf11b904422766cffb697bb9c6471.jpg) +Ten Training Anomaly Examples Under General Settings +Figure 2. Ablation study for SB and DFL under the general settings and hard settings. + +![](images/8f20e4f0e230387c2e81af323efd5c71bf9e0280f64c7a75c3fa256690116855.jpg) + +![](images/db3222376818448821419e66be31f82983244223a722cac6ff6f0daaab0cfddc.jpg) +Ten Training Anomaly Examples Under Hard Settings + +![](images/9d1cbc3ba700da9b6c1c48885b37b1cf35bee15aa1ef07965cd7f3ebe9ea9dd8.jpg) + +![](images/52e48a8462c647629cc8063947bacd4faa66d0ff25e4680459bb4f0bcd62ff5f.jpg) + +![](images/636cde06dae3253814725a80d93b2120b14699832fb797ea88e5c32022f03ffd.jpg) +Figure 3. Parameter sensitivity analysis for $C,\epsilon ,\kappa$ and $\lambda$ + +![](images/617eecc1e1cbf38ca0cb98682474c27f4fb443f94a8e5f50cd1f9e03986f225c.jpg) + +![](images/dcb2a25d389bcdc8281164de9850e18bd07ed9e494f6909ce9364c3a95b93578.jpg) + +Table 3. An ablation study for ${\mathrm{M}}_{\mathrm{n}},{\mathrm{M}}_{\mathrm{a}}$ and ${\mathrm{M}}_{\mathrm{r}}$ . + +
MnMaMrAITEXELPVMastcam
Ten Training Anomaly Examples Under General Settings
××✓0.928 ± 0.0190.914 ± 0.0210.899 ± 0.036
✓×✓0.939 ± 0.0230.919 ± 0.0110.908 ± 0.017
×✓✓0.963 ± 0.0080.930 ± 0.0130.924 ± 0.019
✓✓✓0.975 ± 0.0070.937 ± 0.0030.934 ± 0.010
Ten Training Anomaly Examples Under Hard Settings
××✓0.746 ± 0.0250.797 ± 0.0260.723 ± 0.017
✓×✓0.758 ± 0.0350.802 ± 0.0240.736 ± 0.028
×✓✓0.781 ± 0.0140.811 ± 0.0180.762 ± 0.021
✓✓✓0.798 ± 0.0050.818 ± 0.0030.778 ± 0.007
+ +# 5.6. Ablation Study + +The ablation study in Fig. 2 highlights the critical roles of the DPL and DFL components in improving DPDL's openset anomaly detection. We denote the variants that remove only DPL or DFL as 'DPDL w/o DPL' and 'DPDL w/o DFL', respectively. Compared to the full DPDL model, removing neither DPL nor DFL leads to a significant AUC decline, illustrating their critical utility. Specifically, 'DPDL w/o DPL' exhibits the most significant performance drop on industrial anomaly datasets, reflecting DPL's prominent role in learning precise and tight distribution boundaries for normal samples. By transforming normal samples into Gaussian distribution prototype space and pushing abnormal samples away, DPL enhances the recognition ability of anomaly samples. Meanwhile, ablation studies on 'DPDL w/o DFL' further highlight the critical role of the DFL component. By performing discreteness feature learning in hyperspherical space, DFL enhances the generalization ability to out-of-distribution anomalies. + +Additionally, the ablation experiments on $\mathsf{M}_{\mathrm{n}}$ , $\mathsf{M}_{\mathrm{a}}$ , and $\mathsf{M}_{\mathrm{r}}$ across three datasets in Tab. 3 reveal their varying contributions. DPDL shows the most significant performance drop when $\mathsf{M}_{\mathrm{r}}$ is removed, which illustrates that $\mathsf{M}_{\mathrm{r}}$ plays the most critical role in detecting anomalies. Furthermore, the removal of $\mathsf{M}_{\mathrm{a}}$ and $\mathsf{M}_{\mathrm{n}}$ lead to a noticeable performance drop in DPDL, which underscores their essential roles. + +# 5.7. Parameter Sensitivity Analysis + +Fig. 3 illustrates the results of the four hyperparameters under the general settings across nine datasets. Overall, the performance remains stable within a certain range of hyper + +parameter variations, demonstrating the DPDL's robustness. Prototype quantity $C$ . We begin by investigating the critical impact of the number of initialized prototypes $C$ on distributed prototype learning in DPDL. We select $\{8,16,32,64\}$ as the values for the hyperparameters. As $C$ increases, DPDL's performance improves steadily, but excessively large values of $C$ hinder the model's effectiveness. This phenomenon is particularly pronounced on AI-TEX, Mastcam, MVTecAD, and Hyper-Kvasir, which contain more categories. One possible explanation is that a prototype space with too few prototypes loses discriminative information, while an excessively large number of prototypes reduces the compactness of the space. + +DPL trajectory $\epsilon$ in Eqns. (9), (10), (11) and (14). It can be observed that, particularly on the AITEX, Mastcam, and MVTecAD datasets with a larger number of categories, DPDL exhibits a relatively stable performance decline as $\epsilon$ increases. As a crucial parameter in SB, $\epsilon$ governs the trajectory state. Since smaller values produce straighter trajectories and larger values increase fluctuation, smaller $\epsilon$ facilitates sampling more robust abstract prototypes from the relatively dispersed conditional distribution. + +DFL tightness $\kappa$ . According to Fig. 3, increasing $\kappa$ generally enhances model performance, but values above $\kappa = 10$ introduce negative effects in certain scenarios. A possible reason is that excessive sample dispersion makes it more challenging to tighten the normal distribution boundary. + +Loss parameter $\lambda$ . We conduct a sensitivity analysis on the loss parameters $\lambda$ . It can be observe that setting $\lambda = 0.01$ achieves optimal performance on seven larger-scale datasets, while $\lambda = 1$ yields the best results on two datasets with limited data. As $\lambda$ increases, the performance declines, potentially due to gradient conflicts among the dispersion loss, SB transform loss and the main task loss. + +# 6. Conclusion + +We propose Distribution Prototype Diffusion Learning (DPDL) for OSAD. DPDL leverages schrödinger bridge to map the normal distribution to a prototype space, simultaneously repelling anomalies to facilitate precise anomaly detection. We propose a dispersion feature learning way in hyperspherical space, which benefits the detection of out-of-distribution anomalies. Experimental results illustrate DPDL's robustness in diverse anomaly detection scenarios. + +# 7. Acknowledgments + +This work was supported by the National Natural Science Foundation of China (Grants No. 62476133), the National Natural Science Foundation of China (Nos. U22B2049), the National Natural Science Foundation of China (NSFC) under Grant 62272230, and the Natural Science Foundation of Shandong Province (Grant No. ZR2022LZH003). + +# References + +[1] Andra Acsintoae, Andrei Florescu, Mariana-Iuliana Georgescu, Tudor Mare, Paul Sumedrea, Radu Tudor Ionescu, Fahad Shahbaz Khan, and Mubarak Shah. Ubnormal: New benchmark for supervised open-set video anomaly detection. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 20143-20153, 2022. 1, 2 +[2] Aimira Baitieva, David Hurych, Victor Besnier, and Olivier Bernard. Supervised anomaly detection for complex industrial images. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 17754-17762, 2024. 1 +[3] Paul Bergmann, Michael Fauser, David Sattlegger, and Carsten Steger. Mvtec ad-a comprehensive real-world dataset for unsupervised anomaly detection. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 9592-9600, 2019. 6 +[4] Hanna Borgli, Vajira Thambawita, Pia H Smedsrud, Steven Hicks, Debesh Jha, Sigrun L Eskeland, Kristin Ranheim Randel, Konstantin Pogorelov, Mathias Lux, Duc Tien Dang Nguyen, et al. Hyperkvasir, a comprehensive multi-class image and video dataset for gastrointestinal endoscopy. Scientific data, 7(1):283, 2020. 6 +[5] Yongxin Chen, Tryphon T Georgiou, and Michele Pavon. On the relation between optimal transport and schrödinger bridges: A stochastic control viewpoint. Journal of Optimization Theory and Applications, 169:671-691, 2016. 2 +[6] Songmin Dai, Yifan Wu, Xiaogiang Li, and Xiangyang Xue. Generating and reweighting dense contrastive patterns for unsupervised anomaly detection. In Proceedings of the AAAI Conference on Artificial Intelligence, pages 1454-1462, 2024. 1 +[7] Valentin De Bortoli, James Thornton, Jeremy Heng, and Arnaud Doucet. Diffusion schrödinger bridge with applications to score-based generative modeling. Advances in Neural Information Processing Systems, 34:17695-17709, 2021. 2 +[8] Sergiu Deitsch, Vincent Christlein, Stephan Berger, Claudia Buerhop-Lutz, Andreas Maier, Florian Gallwitz, and Christian Riess. Automatic classification of defective photovoltaic module cells in electroluminescence images. Solar Energy, 185:455-468, 2019. 6 +[9] Chouro Ding, Guansong Pang, and Chunhua Shen. Catching both gray and black swans: Open-set supervised anomaly detection. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 7388-7398, 2022. 1, 2, 5, 6, 7 + +[10] Nikita Gushchin, Sergei Kholkin, Evgeny Burnaev, and Alexander Korotin. Light and optimal schrödinger bridge matching. In *Forty-first International Conference on Machine Learning*, 2024. 2, 3, 4, 5 +[11] Liren He, Zhengkai Jiang, Jinlong Peng, Liang Liu, Qianggang Du, Xiaobin Hu, Wenbing Zhu, Mingmin Chi, Yabiao Wang, and Chengjie Wang. Learning unified reference representation for unsupervised multi-class anomaly detection. arXiv preprint arXiv:2403.11561, 2024. 1 +[12] Teng Hu, Jiangning Zhang, Ran Yi, Yuzhen Du, Xu Chen, Liang Liu, Yabiao Wang, and Chengjie Wang. Anomalydiffusion: Few-shot anomaly image generation with diffusion model. In Proceedings of the AAAI Conference on Artificial Intelligence, pages 8526-8534, 2024. 1 +[13] Hannah R Kerner, Kiri L Wagstaff, Brian D Bue, Danika F Wellington, Samantha Jacob, Paul Horton, James F Bell, Chiman Kwan, and Heni Ben Amor. Comparison of novelty detection methods for multispectral images in rover-based planetary exploration missions. Data Mining and Knowledge Discovery, 34:1642-1675, 2020. 6 +[14] Beomsu Kim, Gihyun Kwon, Kwanyoung Kim, and Jong Chul Ye. Unpaired image-to-image translation via neural schr\” odinger bridge. arXiv preprint arXiv:2305.15086, 2023. 2 +[15] Daehyun Kim, Sungyong Baik, and Tae Hyun Kim. Sanflow: Semantic-aware normalizing flow for anomaly detection. Advances in Neural Information Processing Systems, 36:75434-75454, 2023. 1 +[16] Hyunsu Kim, Jongmin Yoon, and Juho Lee. Fast ensembling with diffusion schr\” odinger bridge. arXiv preprint arXiv:2404.15814, 2024. 2 +[17] Alexander Korotin, Nikita Gushchin, and Evgeny Burnaev. Light schr\” odinger bridge. arXiv preprint arXiv:2310.01174, 2023. 2, 3, 4, 5 +[18] Christian Léonard. A survey of the schr\” odinger problem and some of its connections with optimal transport. arXiv preprint arXiv:1308.0215, 2013. 2 +[19] Chun-Liang Li, Kihyuk Sohn, Jinsung Yoon, and Tomas Pfister. Cutpaste: Self-supervised learning for anomaly detection and localization. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 9664-9674, 2021. 1, 6 +[20] Hanxi Li, Jingqi Wu, Hao Chen, Mingwen Wang, and Chunhua Shen. Efficient anomaly detection with budget annotation using semi-supervised residual transformer. arXiv preprint arXiv:2306.03492, 2023. 1 +[21] Xiaofan Li, Zhizhong Zhang, Xin Tan, Chengwei Chen, Yanyun Qu, Yuan Xie, and Lizhuang Ma. Prompted: Learning prompts with only normal samples for few-shot anomaly detection. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 16838-16848, 2024. 1 +[22] Jingyi Liao, Xun Xu, Manh Cuong Nguyen, Adam Goodge, and Chuan Sheng Foo. Coft-ad: Contrastive fine-tuning for few-shot anomaly detection. IEEE Transactions on Image Processing, 2024. 1 +[23] Guan-Horng Liu, Tianrong Chen, Oswin So, and Evangelos Theodorou. Deep generalized schrödinger bridge. Advances + +in Neural Information Processing Systems, 35:9374-9388, 2022. 2 +[24] Guan-Horng Liu, Yaron Lipman, Maximilian Nickel, Brian Karrer, Evangelos A Theodorou, and Ricky TQ Chen. Generalized schr\” odinger bridge matching. arXiv preprint arXiv:2310.02233, 2023. 2 +[25] Guan-Horng Liu, Arash Vahdat, De-An Huang, Evangelos A Theodorou, Weili Nie, and Anima Anandkumar. $i^2$ -sb: Image-to-image schr\” odinger bridge. arXiv preprint arXiv:2302.05872, 2023. 2 +[26] Jiaqi Liu, Kai Wu, Qiang Nie, Ying Chen, Bin-Bin Gao, Yong Liu, Jinbao Wang, Chengjie Wang, and Feng Zheng. Unsupervised continual anomaly detection with contrastively-learned prompt. In Proceedings of the AAAI Conference on Artificial Intelligence, pages 3639–3647, 2024. 1 +[27] Wen Liu, Weixin Luo, Zhengxin Li, Peilin Zhao, Shenghua Gao, et al. Margin learning embedded prediction for video anomaly detection with a few anomalies. In IJCAI, pages 023-3, 2019. 1, 2, 6 +[28] Xinyue Liu, Jianyuan Wang, Biao Leng, and Shuo Zhang. Dual-modeling decouple distillation for unsupervised anomaly detection. arXiv preprint arXiv:2408.03888, 2024.1 +[29] I Loshchilov. Decoupled weight decay regularization. arXiv preprint arXiv:1711.05101, 2017. 6 +[30] Jialin Luo, Yuanzhi Wang, Ziqi Gu, Yide Qiu, Shuaizhen Yao, Fuyun Wang, Chunyan Xu, Wenhua Zhang, Dan Wang, and Zhen Cui. Mmm-rs: A multi-modal, multi-gsd, multiscene remote sensing dataset and benchmark for text-to-image generation. In The Thirty-eight Conference on Neural Information Processing Systems Datasets and Benchmarks Track. 2 +[31] Kanti V Mardia and Peter E Jupp. Directional statistics. John Wiley & Sons, 2009. 5 +[32] Amir Markovitz, Gilad Sharir, Itamar Friedman, Lihi Zelnik-Manor, and Shai Avidan. Graph embedded pose clustering for anomaly detection. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 10539-10547, 2020. 6 +[33] Maxence Noble, Valentin De Bortoli, Arnaud Doucet, and Alain Durmus. Tree-based diffusion schrödinger bridge with applications to wasserstein barycenters. Advances in Neural Information Processing Systems, 36, 2024. 2 +[34] Guansong Pang, Chunhua Shen, and Anton Van Den Hengel. Deep anomaly detection with deviation networks. In Proceedings of the 25th ACM SIGKDD international conference on knowledge discovery & data mining, pages 353-362, 2019. 1 +[35] Guansong Pang, Chouro Ding, Chunhua Shen, and Anton van den Hengel. Explainable deep few-shot anomaly detection with deviation networks. arXiv preprint arXiv:2108.00462, 2021. 2, 5, 6 +[36] T-YLPG Ross and GKHP Dólar. Focal loss for dense object detection. In proceedings of the IEEE conference on computer vision and pattern recognition, pages 2980-2988, 2017. 6 + +[37] Mohammadreza Salehi, Niousha Sadjadi, Soroosh Baselizadeh, Mohammad H Rohban, and Hamid R Rabiee. Multiresolution knowledge distillation for anomaly detection. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 14902-14912, 2021. 6 +[38] Yuyang Shi, Valentin De Bortoli, Andrew Campbell, and Arnaud Doucet. Diffusion schrödinger bridge matching. Advances in Neural Information Processing Systems, 36, 2024. 2 +[39] Javier Silvestre-Blanes, Teresa Albero-Albero, Ignacio Miralles, Rubén Pérez-Llorens, and Jorge Moreno. A public fabric database for defect detection methods and results. *Autex Research Journal*, 19(4):363–374, 2019. 6 +[40] Domen Tabernik, Samo Sela, Jure Skvarc, and Danijel Skocaj. Segmentation-based deep-learning approach for surface-defect detection. Journal of Intelligent Manufacturing, 31 (3):759-776, 2020. 6 +[41] Jihoon Tack, Sangwoo Mo, Jongheon Jeong, and Jinwoo Shin. Csi: Novelty detection via contrastive learning on distributionally shifted instances. Advances in neural information processing systems, 33:11839-11852, 2020. 6 +[42] Yuanzhi Wang, Zhen Cui, and Yong Li. Distribution-consistent modal recovering for incomplete multimodal learning. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 22025-22034, 2023. 2 +[43] Yuanzhi Wang, Yong Li, and Zhen Cui. Incomplete multimodality-diffused emotion recognition. Advances in Neural Information Processing Systems, 36:17117-17128, 2023. +[44] Yuanzhi Wang, Yong Li, Mengyi Liu, Xiaoya Zhang, Xin Liu, Zhen Cui, and Antoni B Chan. Re-attentional controllable video diffusion editing. arXiv preprint arXiv:2412.11710, 2024. +[45] Yuanzhi Wang, Yong Li, Xiaoya Zhang, Xin Liu, Anbo Dai, Antoni B Chan, and Zhen Cui. Edit temporal-consistent videos with image diffusion model. ACM Transactions on Multimedia Computing, Communications and Applications, 20(12):1-16, 2024. 2 +[46] Matthias Wieler and Tobias Hahn. Weakly supervised learning for industrial optical inspection. In DAGM symposium in, page 11, 2007. 6 +[47] Xincheng Yao, Ruoqi Li, Jing Zhang, Jun Sun, and Chongyang Zhang. Explicit boundary guided semi-push-pull contrastive learning for supervised anomaly detection. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 24490-24499, 2023. 1, 2 +[48] Xincheng Yao, Ruoqi Li, Zefeng Qian, Lu Wang, and Chongyang Zhang. Hierarchical gaussian mixture normalizing flow modeling for unified anomaly detection. arXiv preprint arXiv:2403.13349, 2024. 1 +[49] Sangdoo Yun, Dongyoon Han, Seong Joon Oh, Sanghyuk Chun, Junsuk Choe, and Youngjoon Yoo. Cutmix: Regularization strategy to train strong classifiers with localizable features. In Proceedings of the IEEE/CVF international conference on computer vision, pages 6023-6032, 2019. 6 + +[50] Qihang Zhou, Guansong Pang, Yu Tian, Shibo He, and Jiming Chen. Anomalyclip: Object-agnostic prompt learning for zero-shot anomaly detection. arXiv preprint arXiv:2310.18961, 2023. 1 +[51] Jiawen Zhu, Choubo Ding, Yu Tian, and Guansong Pang. Anomaly heterogeneity learning for open-set supervised anomaly detection. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 17616-17626, 2024. 1, 2, 6, 7 +[52] Yuansheng Zhu, Wentao Bao, and Qi Yu. Towards open set video anomaly detection. In European Conference on Computer Vision, pages 395-412. Springer, 2022. 1, 2 \ No newline at end of file diff --git a/2025/Distribution Prototype Diffusion Learning for Open-set Supervised Anomaly Detection/images.zip b/2025/Distribution Prototype Diffusion Learning for Open-set Supervised Anomaly Detection/images.zip new file mode 100644 index 0000000000000000000000000000000000000000..f58632aa0423ff764c9d29c662ae98679c276cf8 --- /dev/null +++ b/2025/Distribution Prototype Diffusion Learning for Open-set Supervised Anomaly Detection/images.zip @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:89520794f829293953cf92fd466807a2875ce06d3301913e0425f02879b8d3bc +size 811576 diff --git a/2025/Distribution Prototype Diffusion Learning for Open-set Supervised Anomaly Detection/layout.json b/2025/Distribution Prototype Diffusion Learning for Open-set Supervised Anomaly Detection/layout.json new file mode 100644 index 0000000000000000000000000000000000000000..a585e2b0a848087fecb4bcd65067b40b639c4497 --- /dev/null +++ b/2025/Distribution Prototype Diffusion Learning for Open-set Supervised Anomaly Detection/layout.json @@ -0,0 +1,11407 @@ +{ + "pdf_info": [ + { + "para_blocks": [ + { + "bbox": [ + 99, + 102, + 512, + 139 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 99, + 102, + 512, + 139 + ], + "spans": [ + { + "bbox": [ + 99, + 102, + 512, + 139 + ], + "type": "text", + "content": "Distribution Prototype Diffusion Learning for Open-set Supervised Anomaly Detection" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 93, + 160, + 521, + 232 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 93, + 160, + 521, + 232 + ], + "spans": [ + { + "bbox": [ + 93, + 160, + 521, + 232 + ], + "type": "text", + "content": "Fuyun Wang" + }, + { + "bbox": [ + 93, + 160, + 521, + 232 + ], + "type": "inline_equation", + "content": "^{1}" + }, + { + "bbox": [ + 93, + 160, + 521, + 232 + ], + "type": "text", + "content": ", Tong Zhang" + }, + { + "bbox": [ + 93, + 160, + 521, + 232 + ], + "type": "inline_equation", + "content": "^{1*}" + }, + { + "bbox": [ + 93, + 160, + 521, + 232 + ], + "type": "text", + "content": ", Yuanzhi Wang" + }, + { + "bbox": [ + 93, + 160, + 521, + 232 + ], + "type": "inline_equation", + "content": "^{1}" + }, + { + "bbox": [ + 93, + 160, + 521, + 232 + ], + "type": "text", + "content": ", Yide Qiu" + }, + { + "bbox": [ + 93, + 160, + 521, + 232 + ], + "type": "inline_equation", + "content": "^{1}" + }, + { + "bbox": [ + 93, + 160, + 521, + 232 + ], + "type": "text", + "content": ", Xin Liu" + }, + { + "bbox": [ + 93, + 160, + 521, + 232 + ], + "type": "inline_equation", + "content": "^{2}" + }, + { + "bbox": [ + 93, + 160, + 521, + 232 + ], + "type": "text", + "content": ", Xu Guo" + }, + { + "bbox": [ + 93, + 160, + 521, + 232 + ], + "type": "inline_equation", + "content": "^{1}" + }, + { + "bbox": [ + 93, + 160, + 521, + 232 + ], + "type": "text", + "content": ", Zhen Cui" + }, + { + "bbox": [ + 93, + 160, + 521, + 232 + ], + "type": "inline_equation", + "content": "^{3*}" + }, + { + "bbox": [ + 93, + 160, + 521, + 232 + ], + "type": "inline_equation", + "content": "^{1}" + }, + { + "bbox": [ + 93, + 160, + 521, + 232 + ], + "type": "text", + "content": "Nanjing University of Science and Technology, China. \n" + }, + { + "bbox": [ + 93, + 160, + 521, + 232 + ], + "type": "inline_equation", + "content": "^{2}" + }, + { + "bbox": [ + 93, + 160, + 521, + 232 + ], + "type": "text", + "content": "Nanjing SeetaCloud Technology, China. " + }, + { + "bbox": [ + 93, + 160, + 521, + 232 + ], + "type": "inline_equation", + "content": "^{3}" + }, + { + "bbox": [ + 93, + 160, + 521, + 232 + ], + "type": "text", + "content": "Beijing Normal University, China \n{fyw271828, tong.zhang, yuanzhiwang, 121106010824, guo.xu} @njust.edu.cn \nxin.liu@seetacloud.com, zhen.cui@bnu.edu.cn" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 151, + 258, + 200, + 272 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 151, + 258, + 200, + 272 + ], + "spans": [ + { + "bbox": [ + 151, + 258, + 200, + 272 + ], + "type": "text", + "content": "Abstract" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 55, + 284, + 297, + 513 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 284, + 297, + 513 + ], + "spans": [ + { + "bbox": [ + 55, + 284, + 297, + 513 + ], + "type": "text", + "content": "In Open-set Supervised Anomaly Detection (OSAD), the existing methods typically generate pseudo anomalies to compensate for the scarcity of observed anomaly samples, while overlooking critical priors of normal samples, leading to less effective discriminative boundaries. To address this issue, we propose a Distribution Prototype Diffusion Learning (DPDL) method aimed at enclosing normal samples within a compact and discriminative distribution space. Specifically, we construct multiple learnable Gaussian prototypes to create a latent representation space for abundant and diverse normal samples and learn a Schrödinger bridge to facilitate a diffusive transition toward these prototypes for normal samples while steering anomaly samples away. Moreover, to enhance inter-sample separation, we design a dispersion feature learning way in hyperspherical space, which benefits the identification of out-of-distribution anomalies. Experimental results demonstrate the effectiveness and superiority of our proposed DPDL, achieving state-of-the-art performance on 9 public datasets." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 55, + 528, + 135, + 541 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 528, + 135, + 541 + ], + "spans": [ + { + "bbox": [ + 55, + 528, + 135, + 541 + ], + "type": "text", + "content": "1. Introduction" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 55, + 549, + 296, + 693 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 549, + 296, + 693 + ], + "spans": [ + { + "bbox": [ + 55, + 549, + 296, + 693 + ], + "type": "text", + "content": "Anomaly detection (AD) [15, 21, 48, 50] aims to identify outliers significantly diverging from the prevailing samples in a dataset, and has a wide range of applications like industrial inspection, medical image analysis, and scientific discovery, etc. Recently, unsupervised anomaly detection (UAD) [6, 11, 26, 28] and few-shot anomaly detection (FSAD) [12, 21, 22] have emerged as prominent research paradigms, emphasizing the modeling of normal sample distributions to discern anomalies effectively. Yet, these methods often neglect prior knowledge from limited anomaly samples, resulting in imprecise delineation of normal sample boundaries and reduced efficacy in differenti-" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 313, + 260, + 554, + 332 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 260, + 554, + 332 + ], + "spans": [ + { + "bbox": [ + 313, + 260, + 554, + 332 + ], + "type": "text", + "content": "ating normal from anomaly instances. On the contrary, supervised anomaly detection (SAD) [2, 20, 47] leverages a limited subset of anomaly samples as prior knowledge, improving detection performance. However, this reliance on seen anomalies poses a risk of overfitting and hampers generalization to unseen anomalies in real-world settings." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 313, + 337, + 556, + 529 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 337, + 556, + 529 + ], + "spans": [ + { + "bbox": [ + 313, + 337, + 556, + 529 + ], + "type": "text", + "content": "To mitigate the challenge of limited generalization inherent in closed-set training, we focus on open-set supervised anomaly detection (OSAD) [1, 27, 51, 52], which utilizes a small set of known anomaly classes during training to identify unseen anomalies from open-set classes. By leveraging prior knowledge from observed samples, OSAD methods could reduce false positive errors. To improve the generalized detection of unseen anomalies, DRA [9] leverages data augmentation and outlier exposure to learn a decomposed anomaly representation comprising seen anomalies, pseudo-anomalies, and potential residual anomalies. BGAD [47] leverages decision boundaries derived from normalized flow models to capture and model anomaly information. Recently, AHL [51] simulates heterogeneous anomaly distributions and performs collaborative differentiable learning to further enhance the model's generality." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 313, + 534, + 557, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 534, + 557, + 715 + ], + "spans": [ + { + "bbox": [ + 313, + 534, + 557, + 715 + ], + "type": "text", + "content": "While data augmentation and outlier exposure techniques [9, 19, 34] have demonstrated considerable success in anomaly detection, they fall short in generating comprehensive pseudo anomalies, capturing only a fraction of potential unseen anomalies. This limitation arises from overlooking the intricate nature of real-world anomaly distributions, thereby hindering the model's ability to generalize to novel anomaly types. Although AHL [51] has made strides in addressing this issue by simulating heterogeneous anomaly distributions, it still relies on approximating unknown out-of-distribution anomalies using known indistribution anomalies for generalization. Yet, three critical issues persist: i) the simulation mechanism cannot cover all anomaly distribution patterns due to the varied scales and structures of anomaly distributions; ii) simulated anom" + } + ] + } + ], + "index": 10 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "spans": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "type": "text", + "content": "CVF" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 145, + 0, + 496, + 37 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 145, + 0, + 496, + 37 + ], + "spans": [ + { + "bbox": [ + 145, + 0, + 496, + 37 + ], + "type": "text", + "content": "This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 66, + 702, + 223, + 714 + ], + "type": "page_footnote", + "angle": 0, + "lines": [ + { + "bbox": [ + 66, + 702, + 223, + 714 + ], + "spans": [ + { + "bbox": [ + 66, + 702, + 223, + 714 + ], + "type": "text", + "content": "*Corresponding authors: Tong Zhang, Zhen Cui" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "type": "text", + "content": "20416" + } + ] + } + ], + "index": 12 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 0 + }, + { + "para_blocks": [ + { + "bbox": [ + 54, + 72, + 294, + 191 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 54, + 72, + 294, + 191 + ], + "spans": [ + { + "bbox": [ + 54, + 72, + 294, + 191 + ], + "type": "text", + "content": "lies inherit in-distribution data biases, leading to suboptimal performance on out-of-distribution anomalies; iii) the diversity of normal samples presents dilemmas for existing methods, complicating the differentiation between normal and anomaly boundaries. These issues prompt a fundamental question: Instead of generating pseudo and uncertain anomaly samples, how can we accurately characterize compact distribution boundaries amidst a range of diverse normal samples and achieve robust generalization for unknown out-of-distribution anomalies?" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 57, + 193, + 296, + 514 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 193, + 296, + 514 + ], + "spans": [ + { + "bbox": [ + 57, + 193, + 296, + 514 + ], + "type": "text", + "content": "To address the aforementioned issue, in this work, we propose a Distribution Prototype Diffusion Learning (DPDL) method for open-set supervised anomaly detection. Considering abundant and diverse normal samples but very limited anomaly data, our method involves learning latent distribution prototypes, specifically multiple Gaussian distributions, onto which all observed normal samples can be effectively projected. To facilitate the mapping of normal samples into the prototype space, we leverage Schrödinger bridge (SB) framework, which enables a diffusive transition by aligning the distributions of these samples with the prototypes. The SB-based diffusion way could mitigate the out-of-distribution issue for normal samples to some extent. Within the distribution prototype space, we push observed anomaly samples away from normal samples to enhance discriminative capacity. Notably, both the prototypes and the diffusive bridge are learned jointly, resulting in a robust embedding space for normal samples. Moreover, to enhance generalization across unseen anomaly domains, we introduce a dispersion feature learning mechanism that maps intermediate features to a hyperspherical space, leveraging a mixture of von Mises-Fisher (vMF) distributions. This approach bolsters directional feature extraction and promotes robust inter-sample separation, facilitating effective identification of out-of-distribution samples. Experiments demonstrate that our method greatly improves detection capabilities for unseen anomalies." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 55, + 515, + 295, + 622 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 515, + 295, + 622 + ], + "spans": [ + { + "bbox": [ + 55, + 515, + 295, + 622 + ], + "type": "text", + "content": "In summary, our contributions are three-fold: i) we propose a distribution prototype diffusion learning framework that jointly learns multiple Gaussian prototypes and the associated diffusion bridge, creating a compact and discriminative embedding space; ii) we develop dispersion feature learning in hyperspherical space to enhance inter-sample separation and improve generalization; iii) we achieve state-of-the-art performance in 9 public datasets, demonstrating the efficacy of our approach." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 55, + 633, + 142, + 645 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 633, + 142, + 645 + ], + "spans": [ + { + "bbox": [ + 55, + 633, + 142, + 645 + ], + "type": "text", + "content": "2. Related Work" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 55, + 653, + 295, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 653, + 295, + 714 + ], + "spans": [ + { + "bbox": [ + 55, + 653, + 295, + 714 + ], + "type": "text", + "content": "Open-set Supervised Anomaly Detection. Open-set supervised anomaly detection (OSAD) seeks to develop a robust anomaly detection framework that generalizes from a limited set of training anomalies to effectively identify previously unseen anomalies within an open-set con" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 313, + 72, + 553, + 191 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 72, + 553, + 191 + ], + "spans": [ + { + "bbox": [ + 313, + 72, + 553, + 191 + ], + "type": "text", + "content": "text [1, 27, 35, 47, 52]. Leveraging the prior knowledge provided by observed anomalies, contemporary OSAD methods significantly mitigate false positive errors, thereby enhancing overall detection performance [9, 51]. Recently, DRA [9] learns disentangled representations of observed, pseudo, and residual anomalies to boost the detection of both seen and unseen anomalies. In contrast, AHL [51] simulates diverse heterogeneous anomaly distributions and employs collaborative differentiable learning, significantly improving the model's generalization capacity." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 313, + 192, + 555, + 372 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 192, + 555, + 372 + ], + "spans": [ + { + "bbox": [ + 313, + 192, + 555, + 372 + ], + "type": "text", + "content": "Schrödinger Bridge. Schrödinger bridge (SB), widely recognized as the entropy-regularized optimal transport (OT) problem, involves learning a stochastic process that evolves from an initial probability distribution to a target distribution under the influence of a reference measure [5, 7, 16, 18, 23, 24, 38], which is often used for content generation as those prior generative models do [30, 42-45]. " + }, + { + "bbox": [ + 313, + 192, + 555, + 372 + ], + "type": "inline_equation", + "content": "\\mathrm{I}^2" + }, + { + "bbox": [ + 313, + 192, + 555, + 372 + ], + "type": "text", + "content": "-SB [25] and UNSB [14] learn a nonlinear diffusion process between two given distributions or represent the SB problem as a series of adversarial learning problems to realize the image transformation task. Recently, LightSB [10, 17] introduces a novel, fast, and simple SB solver, which achieves optimal matching in practice through the Gaussian mixture parameterization of the adjusted schrödinger potential." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 314, + 381, + 399, + 393 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 314, + 381, + 399, + 393 + ], + "spans": [ + { + "bbox": [ + 314, + 381, + 399, + 393 + ], + "type": "text", + "content": "3. Preliminaries" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 313, + 401, + 553, + 496 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 401, + 553, + 496 + ], + "spans": [ + { + "bbox": [ + 313, + 401, + 553, + 496 + ], + "type": "text", + "content": "We focus on how to build the connection between two distributions " + }, + { + "bbox": [ + 313, + 401, + 553, + 496 + ], + "type": "inline_equation", + "content": "p_0" + }, + { + "bbox": [ + 313, + 401, + 553, + 496 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 313, + 401, + 553, + 496 + ], + "type": "inline_equation", + "content": "p_1" + }, + { + "bbox": [ + 313, + 401, + 553, + 496 + ], + "type": "text", + "content": ", where the distributions are defined as absolutely continuous Borel probability distributions with finite second-order moments. Building upon the foundation of entropy-regularized optimal transport (EOT) [5, 7, 10, 16-18, 23, 24, 33, 38], we review the related properties of EOT and the schrödinger bridge (SB) problem with a Wiener prior." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 313, + 497, + 553, + 557 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 497, + 553, + 557 + ], + "spans": [ + { + "bbox": [ + 313, + 497, + 553, + 557 + ], + "type": "text", + "content": "Entropy-regularized optimal transport (EOT). Given two point sets " + }, + { + "bbox": [ + 313, + 497, + 553, + 557 + ], + "type": "inline_equation", + "content": "\\mathcal{Z}_0" + }, + { + "bbox": [ + 313, + 497, + 553, + 557 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 313, + 497, + 553, + 557 + ], + "type": "inline_equation", + "content": "\\mathcal{Z}_1" + }, + { + "bbox": [ + 313, + 497, + 553, + 557 + ], + "type": "text", + "content": ", we seek for the optimal transport cost between any two points " + }, + { + "bbox": [ + 313, + 497, + 553, + 557 + ], + "type": "inline_equation", + "content": "z_0 \\in \\mathcal{Z}_0" + }, + { + "bbox": [ + 313, + 497, + 553, + 557 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 313, + 497, + 553, + 557 + ], + "type": "inline_equation", + "content": "z_1 \\in \\mathcal{Z}_1" + }, + { + "bbox": [ + 313, + 497, + 553, + 557 + ], + "type": "text", + "content": ". This task may be formulated as an EOT problem with a parameter " + }, + { + "bbox": [ + 313, + 497, + 553, + 557 + ], + "type": "inline_equation", + "content": "\\epsilon > 0" + }, + { + "bbox": [ + 313, + 497, + 553, + 557 + ], + "type": "text", + "content": ", i.e., minimizing the following objective:" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 337, + 563, + 553, + 611 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 337, + 563, + 553, + 611 + ], + "spans": [ + { + "bbox": [ + 337, + 563, + 553, + 611 + ], + "type": "interline_equation", + "content": "\\begin{array}{l} \\min _ {\\pi \\in \\Pi \\left(p _ {0}, p _ {1}\\right)} \\left\\{\\int_ {\\mathbb {R} ^ {D}} \\int_ {\\mathbb {R} ^ {D}} \\frac {1}{2} \\| z _ {0} - z _ {1} \\| ^ {2} \\pi \\left(z _ {0}, z _ {1}\\right) d z _ {0} d z _ {1} \\right. \\tag {1} \\\\ + \\epsilon \\mathrm {K L} (z \\| p _ {0} \\times p _ {1}) \\}, \\\\ \\end{array}", + "image_path": "c4d3f773547d0c78b9cf174b5f605a59fc6d819dd217744c9414a92d1820c467.jpg" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 313, + 617, + 553, + 689 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 617, + 553, + 689 + ], + "spans": [ + { + "bbox": [ + 313, + 617, + 553, + 689 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 313, + 617, + 553, + 689 + ], + "type": "inline_equation", + "content": "\\Pi(p_0, p_1)" + }, + { + "bbox": [ + 313, + 617, + 553, + 689 + ], + "type": "text", + "content": " denotes the set of transport plans, i.e., joint probability distributions on " + }, + { + "bbox": [ + 313, + 617, + 553, + 689 + ], + "type": "inline_equation", + "content": "\\mathbb{R}^D \\times \\mathbb{R}^D" + }, + { + "bbox": [ + 313, + 617, + 553, + 689 + ], + "type": "text", + "content": " with marginals " + }, + { + "bbox": [ + 313, + 617, + 553, + 689 + ], + "type": "inline_equation", + "content": "p_0" + }, + { + "bbox": [ + 313, + 617, + 553, + 689 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 313, + 617, + 553, + 689 + ], + "type": "inline_equation", + "content": "p_1" + }, + { + "bbox": [ + 313, + 617, + 553, + 689 + ], + "type": "text", + "content": ", respectively, and KL denotes Kullback-Leibler divergence. The minimizer " + }, + { + "bbox": [ + 313, + 617, + 553, + 689 + ], + "type": "inline_equation", + "content": "\\pi^*" + }, + { + "bbox": [ + 313, + 617, + 553, + 689 + ], + "type": "text", + "content": " of Eqn. (1) is guaranteed to exist, be unique, and absolutely continuous, and is referred as the EOT plan." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 313, + 689, + 553, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 689, + 553, + 714 + ], + "spans": [ + { + "bbox": [ + 313, + 689, + 553, + 714 + ], + "type": "text", + "content": "Schrodinger Bridge (SB). We define " + }, + { + "bbox": [ + 313, + 689, + 553, + 714 + ], + "type": "inline_equation", + "content": "\\Omega" + }, + { + "bbox": [ + 313, + 689, + 553, + 714 + ], + "type": "text", + "content": " as the space of " + }, + { + "bbox": [ + 313, + 689, + 553, + 714 + ], + "type": "inline_equation", + "content": "\\mathbb{R}^D" + }, + { + "bbox": [ + 313, + 689, + 553, + 714 + ], + "type": "text", + "content": "-valued functions over time " + }, + { + "bbox": [ + 313, + 689, + 553, + 714 + ], + "type": "inline_equation", + "content": "t \\in [0,1]" + }, + { + "bbox": [ + 313, + 689, + 553, + 714 + ], + "type": "text", + "content": ", representing" + } + ] + } + ], + "index": 12 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "text", + "content": "20417" + } + ] + } + ], + "index": 13 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 1 + }, + { + "para_blocks": [ + { + "bbox": [ + 55, + 72, + 294, + 167 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 72, + 294, + 167 + ], + "spans": [ + { + "bbox": [ + 55, + 72, + 294, + 167 + ], + "type": "text", + "content": "trajectories in " + }, + { + "bbox": [ + 55, + 72, + 294, + 167 + ], + "type": "inline_equation", + "content": "\\mathbb{R}^D" + }, + { + "bbox": [ + 55, + 72, + 294, + 167 + ], + "type": "text", + "content": " that start at " + }, + { + "bbox": [ + 55, + 72, + 294, + 167 + ], + "type": "inline_equation", + "content": "t = 0" + }, + { + "bbox": [ + 55, + 72, + 294, + 167 + ], + "type": "text", + "content": " and end at " + }, + { + "bbox": [ + 55, + 72, + 294, + 167 + ], + "type": "inline_equation", + "content": "t = 1" + }, + { + "bbox": [ + 55, + 72, + 294, + 167 + ], + "type": "text", + "content": ". We denote the set of probability distributions over " + }, + { + "bbox": [ + 55, + 72, + 294, + 167 + ], + "type": "inline_equation", + "content": "\\Omega" + }, + { + "bbox": [ + 55, + 72, + 294, + 167 + ], + "type": "text", + "content": ", i.e., stochastic processes, by " + }, + { + "bbox": [ + 55, + 72, + 294, + 167 + ], + "type": "inline_equation", + "content": "\\mathcal{P}(\\Omega)" + }, + { + "bbox": [ + 55, + 72, + 294, + 167 + ], + "type": "text", + "content": ". The differential of the standard Wiener process is represented by " + }, + { + "bbox": [ + 55, + 72, + 294, + 167 + ], + "type": "inline_equation", + "content": "dW_{t}" + }, + { + "bbox": [ + 55, + 72, + 294, + 167 + ], + "type": "text", + "content": ". For a process " + }, + { + "bbox": [ + 55, + 72, + 294, + 167 + ], + "type": "inline_equation", + "content": "T\\in \\mathcal{P}(\\Omega)" + }, + { + "bbox": [ + 55, + 72, + 294, + 167 + ], + "type": "text", + "content": ", we denote its joint distribution at " + }, + { + "bbox": [ + 55, + 72, + 294, + 167 + ], + "type": "inline_equation", + "content": "t = 0,1" + }, + { + "bbox": [ + 55, + 72, + 294, + 167 + ], + "type": "text", + "content": " by " + }, + { + "bbox": [ + 55, + 72, + 294, + 167 + ], + "type": "inline_equation", + "content": "\\pi_T\\in \\mathcal{P}(\\mathbb{R}^D\\times \\mathbb{R}^D)" + }, + { + "bbox": [ + 55, + 72, + 294, + 167 + ], + "type": "text", + "content": ". Similarly, we use " + }, + { + "bbox": [ + 55, + 72, + 294, + 167 + ], + "type": "inline_equation", + "content": "T|_{z_0,z_1}" + }, + { + "bbox": [ + 55, + 72, + 294, + 167 + ], + "type": "text", + "content": " to denote the distribution of " + }, + { + "bbox": [ + 55, + 72, + 294, + 167 + ], + "type": "inline_equation", + "content": "T" + }, + { + "bbox": [ + 55, + 72, + 294, + 167 + ], + "type": "text", + "content": " for " + }, + { + "bbox": [ + 55, + 72, + 294, + 167 + ], + "type": "inline_equation", + "content": "t\\in (0,1)" + }, + { + "bbox": [ + 55, + 72, + 294, + 167 + ], + "type": "text", + "content": ", conditioned on " + }, + { + "bbox": [ + 55, + 72, + 294, + 167 + ], + "type": "inline_equation", + "content": "T" + }, + { + "bbox": [ + 55, + 72, + 294, + 167 + ], + "type": "text", + "content": "'s values " + }, + { + "bbox": [ + 55, + 72, + 294, + 167 + ], + "type": "inline_equation", + "content": "z_{0}" + }, + { + "bbox": [ + 55, + 72, + 294, + 167 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 55, + 72, + 294, + 167 + ], + "type": "inline_equation", + "content": "z_{1}" + }, + { + "bbox": [ + 55, + 72, + 294, + 167 + ], + "type": "text", + "content": " at " + }, + { + "bbox": [ + 55, + 72, + 294, + 167 + ], + "type": "inline_equation", + "content": "t = 0" + }, + { + "bbox": [ + 55, + 72, + 294, + 167 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 55, + 72, + 294, + 167 + ], + "type": "inline_equation", + "content": "t = 1" + }, + { + "bbox": [ + 55, + 72, + 294, + 167 + ], + "type": "text", + "content": ", respectively." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 55, + 167, + 294, + 228 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 167, + 294, + 228 + ], + "spans": [ + { + "bbox": [ + 55, + 167, + 294, + 228 + ], + "type": "text", + "content": "Let " + }, + { + "bbox": [ + 55, + 167, + 294, + 228 + ], + "type": "inline_equation", + "content": "W^{\\epsilon} \\in P(\\Omega)" + }, + { + "bbox": [ + 55, + 167, + 294, + 228 + ], + "type": "text", + "content": " represents a Wiener process with volatility " + }, + { + "bbox": [ + 55, + 167, + 294, + 228 + ], + "type": "inline_equation", + "content": "\\epsilon > 0" + }, + { + "bbox": [ + 55, + 167, + 294, + 228 + ], + "type": "text", + "content": ", starting from " + }, + { + "bbox": [ + 55, + 167, + 294, + 228 + ], + "type": "inline_equation", + "content": "p_0" + }, + { + "bbox": [ + 55, + 167, + 294, + 228 + ], + "type": "text", + "content": " at " + }, + { + "bbox": [ + 55, + 167, + 294, + 228 + ], + "type": "inline_equation", + "content": "t = 0" + }, + { + "bbox": [ + 55, + 167, + 294, + 228 + ], + "type": "text", + "content": ". Its differential is governed by the stochastic differential equation (SDE): " + }, + { + "bbox": [ + 55, + 167, + 294, + 228 + ], + "type": "inline_equation", + "content": "dW_{t}^{\\epsilon} = \\sqrt{\\epsilon} dW_{t}" + }, + { + "bbox": [ + 55, + 167, + 294, + 228 + ], + "type": "text", + "content": ". The Schrödinger bridge problem with the Wiener prior " + }, + { + "bbox": [ + 55, + 167, + 294, + 228 + ], + "type": "inline_equation", + "content": "W^{\\epsilon}" + }, + { + "bbox": [ + 55, + 167, + 294, + 228 + ], + "type": "text", + "content": " between " + }, + { + "bbox": [ + 55, + 167, + 294, + 228 + ], + "type": "inline_equation", + "content": "p_0" + }, + { + "bbox": [ + 55, + 167, + 294, + 228 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 55, + 167, + 294, + 228 + ], + "type": "inline_equation", + "content": "p_1" + }, + { + "bbox": [ + 55, + 167, + 294, + 228 + ], + "type": "text", + "content": " is minimizing:" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 123, + 236, + 294, + 256 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 123, + 236, + 294, + 256 + ], + "spans": [ + { + "bbox": [ + 123, + 236, + 294, + 256 + ], + "type": "interline_equation", + "content": "\\min _ {T \\in \\mathcal {F} \\left(p _ {0}, p _ {1}\\right)} \\operatorname {K L} \\left(T \\| W ^ {\\epsilon}\\right), \\tag {2}", + "image_path": "7b3084ca78748e3dbed35f43bcd68100909bc909fca19e4c070e311e758b8f8b.jpg" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 55, + 263, + 295, + 334 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 263, + 295, + 334 + ], + "spans": [ + { + "bbox": [ + 55, + 263, + 295, + 334 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 55, + 263, + 295, + 334 + ], + "type": "inline_equation", + "content": "\\mathcal{F}(p_0,p_1)\\subset \\mathcal{P}(\\Omega)" + }, + { + "bbox": [ + 55, + 263, + 295, + 334 + ], + "type": "text", + "content": " denotes the subset of stochastic processes that begin with distribution " + }, + { + "bbox": [ + 55, + 263, + 295, + 334 + ], + "type": "inline_equation", + "content": "p_0" + }, + { + "bbox": [ + 55, + 263, + 295, + 334 + ], + "type": "text", + "content": " at " + }, + { + "bbox": [ + 55, + 263, + 295, + 334 + ], + "type": "inline_equation", + "content": "t = 0" + }, + { + "bbox": [ + 55, + 263, + 295, + 334 + ], + "type": "text", + "content": " and reach " + }, + { + "bbox": [ + 55, + 263, + 295, + 334 + ], + "type": "inline_equation", + "content": "p_1" + }, + { + "bbox": [ + 55, + 263, + 295, + 334 + ], + "type": "text", + "content": " at " + }, + { + "bbox": [ + 55, + 263, + 295, + 334 + ], + "type": "inline_equation", + "content": "t = 1" + }, + { + "bbox": [ + 55, + 263, + 295, + 334 + ], + "type": "text", + "content": ". This problem has a unique solution, a SDE diffusion process " + }, + { + "bbox": [ + 55, + 263, + 295, + 334 + ], + "type": "inline_equation", + "content": "T^{*}" + }, + { + "bbox": [ + 55, + 263, + 295, + 334 + ], + "type": "text", + "content": " defined: " + }, + { + "bbox": [ + 55, + 263, + 295, + 334 + ], + "type": "inline_equation", + "content": "dZ_{t} = g^{*}(Z_{t},t)dt + dW_{t}^{\\epsilon}" + }, + { + "bbox": [ + 55, + 263, + 295, + 334 + ], + "type": "text", + "content": ". The process " + }, + { + "bbox": [ + 55, + 263, + 295, + 334 + ], + "type": "inline_equation", + "content": "T^{*}" + }, + { + "bbox": [ + 55, + 263, + 295, + 334 + ], + "type": "text", + "content": " is referred to as SB, and " + }, + { + "bbox": [ + 55, + 263, + 295, + 334 + ], + "type": "inline_equation", + "content": "g^{*}:\\mathbb{R}^{D}\\times [0,1]\\to \\mathbb{R}^{D}" + }, + { + "bbox": [ + 55, + 263, + 295, + 334 + ], + "type": "text", + "content": " is the optimal drift." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 55, + 335, + 295, + 360 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 335, + 295, + 360 + ], + "spans": [ + { + "bbox": [ + 55, + 335, + 295, + 360 + ], + "type": "text", + "content": "Characterization of solutions. The EOT plan " + }, + { + "bbox": [ + 55, + 335, + 295, + 360 + ], + "type": "inline_equation", + "content": "\\pi^{*} = \\pi^{T^{*}}" + }, + { + "bbox": [ + 55, + 335, + 295, + 360 + ], + "type": "text", + "content": " takes a specific form [10, 17]:" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 79, + 366, + 294, + 392 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 79, + 366, + 294, + 392 + ], + "spans": [ + { + "bbox": [ + 79, + 366, + 294, + 392 + ], + "type": "interline_equation", + "content": "\\pi^ {*} \\left(z _ {0}, z _ {1}\\right) = u ^ {*} \\left(z _ {0}\\right) \\exp \\left(\\frac {- \\| z _ {0} - z _ {1} \\| ^ {2}}{2 \\epsilon}\\right) v ^ {*} \\left(z _ {1}\\right), \\tag {3}", + "image_path": "3b61a0c920f7c9765033411de2c6d762fa19c8b05b61c2ce9c1e544cd2b5814d.jpg" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 55, + 397, + 294, + 422 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 397, + 294, + 422 + ], + "spans": [ + { + "bbox": [ + 55, + 397, + 294, + 422 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 55, + 397, + 294, + 422 + ], + "type": "inline_equation", + "content": "u^{*}, v^{*} \\colon \\mathbb{R} \\to \\mathbb{R}_{+}" + }, + { + "bbox": [ + 55, + 397, + 294, + 422 + ], + "type": "text", + "content": " are measurable functions known as Schrödinger potentials. The optimal drift " + }, + { + "bbox": [ + 55, + 397, + 294, + 422 + ], + "type": "inline_equation", + "content": "g^{*}" + }, + { + "bbox": [ + 55, + 397, + 294, + 422 + ], + "type": "text", + "content": " is derived as:" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 60, + 429, + 295, + 455 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 60, + 429, + 295, + 455 + ], + "spans": [ + { + "bbox": [ + 60, + 429, + 295, + 455 + ], + "type": "interline_equation", + "content": "g ^ {*} (z, t) = \\epsilon \\nabla_ {z} \\log \\int_ {\\mathbb {R} ^ {D}} \\mathcal {N} \\left(z ^ {\\prime} \\mid z, (1 - t) \\epsilon I _ {D}\\right) v ^ {*} \\left(z ^ {\\prime}\\right) d z ^ {\\prime}, \\tag {4}", + "image_path": "fcd5462c67bb9445e4edb36b52b5dddf53ebf2932898b429243cf71cb7c3f35f.jpg" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 55, + 461, + 111, + 473 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 461, + 111, + 473 + ], + "spans": [ + { + "bbox": [ + 55, + 461, + 111, + 473 + ], + "type": "text", + "content": "4. Method" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 55, + 480, + 180, + 492 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 480, + 180, + 492 + ], + "spans": [ + { + "bbox": [ + 55, + 480, + 180, + 492 + ], + "type": "text", + "content": "4.1. Problem Formulation" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 55, + 498, + 295, + 665 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 498, + 295, + 665 + ], + "spans": [ + { + "bbox": [ + 55, + 498, + 295, + 665 + ], + "type": "text", + "content": "Let " + }, + { + "bbox": [ + 55, + 498, + 295, + 665 + ], + "type": "inline_equation", + "content": "\\mathcal{X}_{\\mathrm{tr}} = \\{(\\mathbf{x}_i, y_i)\\}" + }, + { + "bbox": [ + 55, + 498, + 295, + 665 + ], + "type": "text", + "content": " denotes a weakly-supervised training set with only image-level labels, where " + }, + { + "bbox": [ + 55, + 498, + 295, + 665 + ], + "type": "inline_equation", + "content": "\\mathbf{x}_i" + }, + { + "bbox": [ + 55, + 498, + 295, + 665 + ], + "type": "text", + "content": " denotes one RGB image and " + }, + { + "bbox": [ + 55, + 498, + 295, + 665 + ], + "type": "inline_equation", + "content": "y_i \\in \\{0,1\\}" + }, + { + "bbox": [ + 55, + 498, + 295, + 665 + ], + "type": "text", + "content": " denotes whether " + }, + { + "bbox": [ + 55, + 498, + 295, + 665 + ], + "type": "inline_equation", + "content": "\\mathbf{x}_i" + }, + { + "bbox": [ + 55, + 498, + 295, + 665 + ], + "type": "text", + "content": " is an anomaly sample (anomaly: " + }, + { + "bbox": [ + 55, + 498, + 295, + 665 + ], + "type": "inline_equation", + "content": "y_i = 1" + }, + { + "bbox": [ + 55, + 498, + 295, + 665 + ], + "type": "text", + "content": ", normal: " + }, + { + "bbox": [ + 55, + 498, + 295, + 665 + ], + "type": "inline_equation", + "content": "y_i = 0" + }, + { + "bbox": [ + 55, + 498, + 295, + 665 + ], + "type": "text", + "content": "). Hereby, " + }, + { + "bbox": [ + 55, + 498, + 295, + 665 + ], + "type": "inline_equation", + "content": "\\mathcal{X}_{\\mathrm{tr}}" + }, + { + "bbox": [ + 55, + 498, + 295, + 665 + ], + "type": "text", + "content": " is consistent of a normal subset " + }, + { + "bbox": [ + 55, + 498, + 295, + 665 + ], + "type": "inline_equation", + "content": "\\mathcal{X}_{\\mathrm{tr}}^{\\mathrm{n}}" + }, + { + "bbox": [ + 55, + 498, + 295, + 665 + ], + "type": "text", + "content": " (" + }, + { + "bbox": [ + 55, + 498, + 295, + 665 + ], + "type": "inline_equation", + "content": "|\\mathcal{X}_{\\mathrm{tr}}^{\\mathrm{n}}| = N" + }, + { + "bbox": [ + 55, + 498, + 295, + 665 + ], + "type": "text", + "content": ") and an anomaly subset " + }, + { + "bbox": [ + 55, + 498, + 295, + 665 + ], + "type": "inline_equation", + "content": "\\mathcal{X}_{\\mathrm{tr}}^{\\mathrm{a}}" + }, + { + "bbox": [ + 55, + 498, + 295, + 665 + ], + "type": "text", + "content": " (" + }, + { + "bbox": [ + 55, + 498, + 295, + 665 + ], + "type": "inline_equation", + "content": "|\\mathcal{X}_{\\mathrm{tr}}^{\\mathrm{a}}| = M" + }, + { + "bbox": [ + 55, + 498, + 295, + 665 + ], + "type": "text", + "content": "), formally, " + }, + { + "bbox": [ + 55, + 498, + 295, + 665 + ], + "type": "inline_equation", + "content": "\\mathcal{X}_{\\mathrm{tr}} \\doteq \\mathcal{X}_{\\mathrm{tr}}^{\\mathrm{n}} \\cup \\mathcal{X}_{\\mathrm{tr}}^{\\mathrm{a}}" + }, + { + "bbox": [ + 55, + 498, + 295, + 665 + ], + "type": "text", + "content": ", where generally " + }, + { + "bbox": [ + 55, + 498, + 295, + 665 + ], + "type": "inline_equation", + "content": "N \\gg M" + }, + { + "bbox": [ + 55, + 498, + 295, + 665 + ], + "type": "text", + "content": ". Given a testing set " + }, + { + "bbox": [ + 55, + 498, + 295, + 665 + ], + "type": "inline_equation", + "content": "\\mathcal{X}_{\\mathrm{te}}" + }, + { + "bbox": [ + 55, + 498, + 295, + 665 + ], + "type": "text", + "content": ", we need to predict whether one sample " + }, + { + "bbox": [ + 55, + 498, + 295, + 665 + ], + "type": "inline_equation", + "content": "\\mathbf{x} \\in \\mathcal{X}_{\\mathrm{te}}" + }, + { + "bbox": [ + 55, + 498, + 295, + 665 + ], + "type": "text", + "content": " is anomaly or normal. In OSAD, the anomaly patterns of the testing set do less recurred in those encountered training set. In other words, the distribution of anomalies are obviously discrepant, i.e., " + }, + { + "bbox": [ + 55, + 498, + 295, + 665 + ], + "type": "inline_equation", + "content": "P(\\mathcal{X}_{\\mathrm{te}}^{\\mathrm{a}}) \\neq P(\\mathcal{X}_{\\mathrm{tr}}^{\\mathrm{a}})" + }, + { + "bbox": [ + 55, + 498, + 295, + 665 + ], + "type": "text", + "content": ". Hereby, we need to learn a robust anomaly detection model " + }, + { + "bbox": [ + 55, + 498, + 295, + 665 + ], + "type": "inline_equation", + "content": "\\psi" + }, + { + "bbox": [ + 55, + 498, + 295, + 665 + ], + "type": "text", + "content": " from the training set " + }, + { + "bbox": [ + 55, + 498, + 295, + 665 + ], + "type": "inline_equation", + "content": "\\mathcal{X}_{\\mathrm{tr}}" + }, + { + "bbox": [ + 55, + 498, + 295, + 665 + ], + "type": "text", + "content": " so that " + }, + { + "bbox": [ + 55, + 498, + 295, + 665 + ], + "type": "inline_equation", + "content": "\\psi" + }, + { + "bbox": [ + 55, + 498, + 295, + 665 + ], + "type": "text", + "content": " accurately infers anomaly scores for test samples." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 55, + 666, + 295, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 666, + 295, + 713 + ], + "spans": [ + { + "bbox": [ + 55, + 666, + 295, + 713 + ], + "type": "text", + "content": "Our abstract idea is to learn latent distribution prototypes that not only encapsulate normal samples in a concise manner but also discriminate against anomaly samples. Given the abundance of observed normal samples, one" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 313, + 72, + 553, + 180 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 72, + 553, + 180 + ], + "spans": [ + { + "bbox": [ + 313, + 72, + 553, + 180 + ], + "type": "text", + "content": "natural approach is to characterize the distribution " + }, + { + "bbox": [ + 313, + 72, + 553, + 180 + ], + "type": "inline_equation", + "content": "P(\\mathcal{X}_{\\mathrm{tr}}^{\\mathrm{n}})" + }, + { + "bbox": [ + 313, + 72, + 553, + 180 + ], + "type": "text", + "content": " where samples outside this distribution, " + }, + { + "bbox": [ + 313, + 72, + 553, + 180 + ], + "type": "inline_equation", + "content": "\\mathbf{x}\\notin P(\\mathcal{X}_{\\mathrm{tr}}^{\\mathrm{n}})" + }, + { + "bbox": [ + 313, + 72, + 553, + 180 + ], + "type": "text", + "content": ", would be awarded higher probabilities as anomalies. Considering the inherent diversity of normal samples, we endeavor to learn multiple simple distributions (e.g., Gaussians) as prototypes " + }, + { + "bbox": [ + 313, + 72, + 553, + 180 + ], + "type": "inline_equation", + "content": "\\mathcal{P}_{\\mathrm{MGP}}" + }, + { + "bbox": [ + 313, + 72, + 553, + 180 + ], + "type": "text", + "content": ", named multi-Gaussian prototypes (MGP). To embed input data into the prototype space, we introduce a generative bridge model " + }, + { + "bbox": [ + 313, + 72, + 553, + 180 + ], + "type": "inline_equation", + "content": "\\psi_{p}" + }, + { + "bbox": [ + 313, + 72, + 553, + 180 + ], + "type": "text", + "content": " for distribution transformation. The more abstract formulation is given as follows:" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 316, + 185, + 553, + 205 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 185, + 553, + 205 + ], + "spans": [ + { + "bbox": [ + 316, + 185, + 553, + 205 + ], + "type": "interline_equation", + "content": "\\min _ {\\psi_ {p}, \\mathcal {P} _ {\\mathrm {M G P}}, f} D _ {p} \\left(\\psi_ {p} \\left(\\mathcal {F} _ {\\mathrm {t r}} ^ {\\mathrm {n}}\\right), \\psi_ {p} \\left(\\mathcal {F} _ {\\mathrm {t r}} ^ {\\mathrm {a}}\\right), \\mathcal {P} _ {\\mathrm {M G P}}\\right) + \\lambda D _ {s} \\left(\\mathcal {F} _ {\\mathrm {t r}} ^ {\\mathrm {n}}, \\mathcal {F} _ {\\mathrm {t r}} ^ {\\mathrm {a}}\\right), \\tag {5}", + "image_path": "4936da7648ea811fd0fddbd74df68dd495d84810f7d299d1f9cd38545d59b27c.jpg" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 335, + 207, + 553, + 224 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 335, + 207, + 553, + 224 + ], + "spans": [ + { + "bbox": [ + 335, + 207, + 553, + 224 + ], + "type": "interline_equation", + "content": "s. t., \\psi_ {p}: P (\\mathcal {F}) \\stackrel {\\text {b r i d g e}} {\\Longrightarrow} \\mathcal {P} _ {\\mathrm {M G P}}, \\tag {6}", + "image_path": "6e2689ba83a8c61f14035f8a08adfb0e68dc6a49155889bfb7a1e29c8868ef65.jpg" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 355, + 227, + 553, + 243 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 355, + 227, + 553, + 243 + ], + "spans": [ + { + "bbox": [ + 355, + 227, + 553, + 243 + ], + "type": "interline_equation", + "content": "f: \\mathcal {X} \\xrightarrow {\\text {f e a t u r e}} \\mathcal {F}, \\tag {7}", + "image_path": "ccdcff9ee1606fe61a4bb1e42f5322165fd3c9b65ebe74f49da5d776cf4bec75.jpg" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 313, + 247, + 553, + 403 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 247, + 553, + 403 + ], + "spans": [ + { + "bbox": [ + 313, + 247, + 553, + 403 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 313, + 247, + 553, + 403 + ], + "type": "inline_equation", + "content": "\\mathcal{P}_{\\mathrm{MGP}}" + }, + { + "bbox": [ + 313, + 247, + 553, + 403 + ], + "type": "text", + "content": " denotes the distribution prototypes to be learned, " + }, + { + "bbox": [ + 313, + 247, + 553, + 403 + ], + "type": "inline_equation", + "content": "\\psi_{p}" + }, + { + "bbox": [ + 313, + 247, + 553, + 403 + ], + "type": "text", + "content": " is the flow function across probability distributions, " + }, + { + "bbox": [ + 313, + 247, + 553, + 403 + ], + "type": "inline_equation", + "content": "D_{p}" + }, + { + "bbox": [ + 313, + 247, + 553, + 403 + ], + "type": "text", + "content": " signifies the discriminative function in the space of prototypes, " + }, + { + "bbox": [ + 313, + 247, + 553, + 403 + ], + "type": "inline_equation", + "content": "f" + }, + { + "bbox": [ + 313, + 247, + 553, + 403 + ], + "type": "text", + "content": " stands for the feature extraction process, and " + }, + { + "bbox": [ + 313, + 247, + 553, + 403 + ], + "type": "inline_equation", + "content": "D_{s}" + }, + { + "bbox": [ + 313, + 247, + 553, + 403 + ], + "type": "text", + "content": " acts as a regularizer to increase feature discriminability. In the above formulation, besides distribution prototype learning (DPL), we also introduce dispersion feature learning (DFL) executed within a hyperspherical embedding space, i.e., " + }, + { + "bbox": [ + 313, + 247, + 553, + 403 + ], + "type": "inline_equation", + "content": "D_{s}(\\mathcal{F}_{\\mathrm{tr}}^{\\mathrm{n}},\\mathcal{F}_{\\mathrm{tr}}^{\\mathrm{a}})" + }, + { + "bbox": [ + 313, + 247, + 553, + 403 + ], + "type": "text", + "content": ". The advantage of DFL is to prevent abrupt feature collapse in feature learning, thus preserving discriminative qualities. The overview of our framework is abstractly depicted in Fig. 1. Detailed elaboration on these aspects will be provided in the subsequent sections." + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 313, + 409, + 489, + 422 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 409, + 489, + 422 + ], + "spans": [ + { + "bbox": [ + 313, + 409, + 489, + 422 + ], + "type": "text", + "content": "4.2. Distribution Prototype Learning" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 313, + 426, + 554, + 545 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 426, + 554, + 545 + ], + "spans": [ + { + "bbox": [ + 313, + 426, + 554, + 545 + ], + "type": "text", + "content": "Distribution prototype learning is operated in the feature space. Hence, to extract intermediate image features, we can leverage those classic networks as the backbone, such as ResNet-18. Formally, the feature extraction process, denoted as a function " + }, + { + "bbox": [ + 313, + 426, + 554, + 545 + ], + "type": "inline_equation", + "content": "f: \\mathcal{X} \\to \\mathcal{F}" + }, + { + "bbox": [ + 313, + 426, + 554, + 545 + ], + "type": "text", + "content": ", transforms an image " + }, + { + "bbox": [ + 313, + 426, + 554, + 545 + ], + "type": "inline_equation", + "content": "\\mathbf{x}" + }, + { + "bbox": [ + 313, + 426, + 554, + 545 + ], + "type": "text", + "content": " into the intermediate feature " + }, + { + "bbox": [ + 313, + 426, + 554, + 545 + ], + "type": "inline_equation", + "content": "\\mathbf{x} = f(\\mathbf{x}) \\in \\mathbb{R}^d" + }, + { + "bbox": [ + 313, + 426, + 554, + 545 + ], + "type": "text", + "content": ", where a flatten operation is used on convolution maps for vectorization. Consequently, we designate the intermediate feature sets of normal and anomaly samples as " + }, + { + "bbox": [ + 313, + 426, + 554, + 545 + ], + "type": "inline_equation", + "content": "\\mathcal{F}_{\\mathrm{tr}}^{\\mathrm{n}} = \\{\\mathbf{x}_i^{\\mathrm{n}} | i = 1, \\dots, N\\}" + }, + { + "bbox": [ + 313, + 426, + 554, + 545 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 313, + 426, + 554, + 545 + ], + "type": "inline_equation", + "content": "\\mathcal{F}_{\\mathrm{tr}}^{\\mathrm{a}} = \\{\\mathbf{x}_j^{\\mathrm{a}} | i = 1, \\dots, M\\}" + }, + { + "bbox": [ + 313, + 426, + 554, + 545 + ], + "type": "text", + "content": ", respectively." + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 313, + 546, + 554, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 546, + 554, + 713 + ], + "spans": [ + { + "bbox": [ + 313, + 546, + 554, + 713 + ], + "type": "text", + "content": "Considering the rarity of anomalies and the diversity of normal samples, it is pertinent to capture the intricate distributions of normal samples by mapping the distribution " + }, + { + "bbox": [ + 313, + 546, + 554, + 713 + ], + "type": "inline_equation", + "content": "p_0 = P(\\mathcal{F}_{\\mathrm{tr}}^{\\mathrm{n}})" + }, + { + "bbox": [ + 313, + 546, + 554, + 713 + ], + "type": "text", + "content": " of intermediate features to a distinct and well-characterized distribution " + }, + { + "bbox": [ + 313, + 546, + 554, + 713 + ], + "type": "inline_equation", + "content": "p_1" + }, + { + "bbox": [ + 313, + 546, + 554, + 713 + ], + "type": "text", + "content": " like Gaussian. But due to various normal samples, we opt for a multi-Gaussian prototypes (MGP) comprising multiple Gaussian distributions as prototypes " + }, + { + "bbox": [ + 313, + 546, + 554, + 713 + ], + "type": "inline_equation", + "content": "\\mathcal{P}_{\\mathrm{MGP}} = \\{\\mathcal{P}_i \\doteq \\mathcal{N}(\\pmb{\\mu}_i, \\pmb{\\sigma}_i) | i = 1, \\dots, C\\}" + }, + { + "bbox": [ + 313, + 546, + 554, + 713 + ], + "type": "text", + "content": ". Given a normal sample " + }, + { + "bbox": [ + 313, + 546, + 554, + 713 + ], + "type": "inline_equation", + "content": "\\mathbf{x}_i^{\\mathrm{n}}" + }, + { + "bbox": [ + 313, + 546, + 554, + 713 + ], + "type": "text", + "content": ", we expect to align it with the closest prototype with high likelihood. However, in the open-set setting, it is challenge to transform unseen points " + }, + { + "bbox": [ + 313, + 546, + 554, + 713 + ], + "type": "inline_equation", + "content": "\\mathbf{x} \\sim p_0" + }, + { + "bbox": [ + 313, + 546, + 554, + 713 + ], + "type": "text", + "content": " to the target distribution " + }, + { + "bbox": [ + 313, + 546, + 554, + 713 + ], + "type": "inline_equation", + "content": "p_1" + }, + { + "bbox": [ + 313, + 546, + 554, + 713 + ], + "type": "text", + "content": ". Drawing inspiration from the capability of diffusion generation models in aligning disparate distributions, we frame the transition from source" + } + ] + } + ], + "index": 19 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "text", + "content": "20418" + } + ] + } + ], + "index": 20 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 2 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 61, + 68, + 550, + 259 + ], + "blocks": [ + { + "bbox": [ + 61, + 68, + 550, + 259 + ], + "lines": [ + { + "bbox": [ + 61, + 68, + 550, + 259 + ], + "spans": [ + { + "bbox": [ + 61, + 68, + 550, + 259 + ], + "type": "image", + "image_path": "e170339ed7ef8b2503fd8d411bf6288fe8634990f0c3b92f2d8db4a2dec08fe9.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 54, + 267, + 555, + 336 + ], + "lines": [ + { + "bbox": [ + 54, + 267, + 555, + 336 + ], + "spans": [ + { + "bbox": [ + 54, + 267, + 555, + 336 + ], + "type": "text", + "content": "Figure 1. Our proposed DPDL framework. It comprises three distinct modules: Distribution Prototype Learning (DPL, Sec. 4.2), Dispersion Feature Learning (DFL, Sec. 4.3), and anomaly score prediction (Sec. 4.4). DPL transforms the distribution of normal samples to a space of learnable multiple Gaussian prototypes through building schrödinger bridge, meantime pushing anomaly distribution away from these prototypes. DFL operates in a hyperspherical space, enlarging the distances of intermediate features of all samples in a hyperspherical space to strengthen feature generalization for detecting anomalies. The score prediction module leverages a multi-instance-learning method to compute anomaly scores." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "bbox": [ + 55, + 345, + 296, + 411 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 345, + 296, + 411 + ], + "spans": [ + { + "bbox": [ + 55, + 345, + 296, + 411 + ], + "type": "text", + "content": "domain distribution " + }, + { + "bbox": [ + 55, + 345, + 296, + 411 + ], + "type": "inline_equation", + "content": "p_0" + }, + { + "bbox": [ + 55, + 345, + 296, + 411 + ], + "type": "text", + "content": " to target domain distribution " + }, + { + "bbox": [ + 55, + 345, + 296, + 411 + ], + "type": "inline_equation", + "content": "p_1" + }, + { + "bbox": [ + 55, + 345, + 296, + 411 + ], + "type": "text", + "content": " as a Schrodinger bridge problem. As formulated in Eqn. (6), we need learn an essential bridge flow " + }, + { + "bbox": [ + 55, + 345, + 296, + 411 + ], + "type": "inline_equation", + "content": "\\psi_p: P(\\mathcal{F}_{\\mathrm{tr}}^{\\mathrm{n}}) \\xrightarrow{\\mathrm{bridge}} \\mathcal{P}_{\\mathrm{MGP}}" + }, + { + "bbox": [ + 55, + 345, + 296, + 411 + ], + "type": "text", + "content": ". After bridge transformation, the condition probability in optimal transport plan conforms to:" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 89, + 413, + 295, + 464 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 89, + 413, + 295, + 464 + ], + "spans": [ + { + "bbox": [ + 89, + 413, + 295, + 464 + ], + "type": "interline_equation", + "content": "\\pi \\left(\\psi_ {p} (\\mathbf {x}) | \\mathbf {x}\\right) \\propto \\underbrace {\\sum_ {c = 1} ^ {C} \\alpha_ {c} \\mathcal {N} \\left(\\psi_ {p} (\\mathbf {x}) ; \\boldsymbol {\\mu} _ {c} , \\boldsymbol {\\sigma} _ {c}\\right)} _ {\\dot {\\equiv} \\phi_ {1} \\left(\\psi_ {p} (\\mathbf {x})\\right)}, \\tag {8}", + "image_path": "f4a8dd0aa0b19bd79e6db0cbfdf46e7c8a61a9d039d70b8f7171a852d65fab13.jpg" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 55, + 469, + 295, + 517 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 469, + 295, + 517 + ], + "spans": [ + { + "bbox": [ + 55, + 469, + 295, + 517 + ], + "type": "text", + "content": "where the parameters " + }, + { + "bbox": [ + 55, + 469, + 295, + 517 + ], + "type": "inline_equation", + "content": "\\{\\alpha_c,\\pmb {\\mu}_c,\\pmb {\\sigma}_c\\}_{c = 1}^C" + }, + { + "bbox": [ + 55, + 469, + 295, + 517 + ], + "type": "text", + "content": " (known as Gaussian mixed model (GMM), abstracted into the function " + }, + { + "bbox": [ + 55, + 469, + 295, + 517 + ], + "type": "inline_equation", + "content": "\\phi_1" + }, + { + "bbox": [ + 55, + 469, + 295, + 517 + ], + "type": "text", + "content": " ) as well as the flow function " + }, + { + "bbox": [ + 55, + 469, + 295, + 517 + ], + "type": "inline_equation", + "content": "\\psi_{p}" + }, + { + "bbox": [ + 55, + 469, + 295, + 517 + ], + "type": "text", + "content": " (in bridge) need to be learned. For simplicity, below we adopt diagonal matrices for " + }, + { + "bbox": [ + 55, + 469, + 295, + 517 + ], + "type": "inline_equation", + "content": "\\pmb{\\sigma}_{c}" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 55, + 517, + 296, + 590 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 517, + 296, + 590 + ], + "spans": [ + { + "bbox": [ + 55, + 517, + 296, + 590 + ], + "type": "text", + "content": "Drawing inspiration from previous works [10, 17], we reframe distribution prototype learning in Eqn. (8) as the Schrödinger bridge. Given the specific form taken by the EOT plan described in Eqn. (3), we redefine these measurable functions " + }, + { + "bbox": [ + 55, + 517, + 296, + 590 + ], + "type": "inline_equation", + "content": "u, v: \\mathbb{R} \\to \\mathbb{R}_+" + }, + { + "bbox": [ + 55, + 517, + 296, + 590 + ], + "type": "text", + "content": ", termed Schrödinger potentials [10, 17], as follows:" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 106, + 593, + 295, + 618 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 106, + 593, + 295, + 618 + ], + "spans": [ + { + "bbox": [ + 106, + 593, + 295, + 618 + ], + "type": "interline_equation", + "content": "u \\left(\\mathbf {x} _ {i} ^ {\\mathrm {n}}\\right) \\doteq \\exp \\left(\\frac {\\left\\| \\mathbf {x} _ {i} ^ {\\mathrm {n}} \\right\\| ^ {2}}{2 \\epsilon}\\right) \\phi_ {0} \\left(\\mathbf {x} _ {i} ^ {\\mathrm {n}}\\right), \\tag {9}", + "image_path": "1edd9fd5a8e3271411b7a7dd733b34eee48af7a0b81097601d9b6aded8ba5569.jpg" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 93, + 619, + 295, + 644 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 93, + 619, + 295, + 644 + ], + "spans": [ + { + "bbox": [ + 93, + 619, + 295, + 644 + ], + "type": "interline_equation", + "content": "v \\left(\\psi \\left(\\mathbf {x} _ {i} ^ {\\mathrm {n}}\\right)\\right) \\doteq \\exp \\left(\\frac {\\| \\psi \\left(\\mathbf {x} _ {i} ^ {\\mathrm {n}}\\right) \\| ^ {2}}{2 \\epsilon}\\right) \\phi_ {1} \\left(\\psi \\left(\\mathbf {x} _ {i} ^ {\\mathrm {n}}\\right)\\right), \\tag {10}", + "image_path": "cd0260c0c7d8d1d6ca27fa628b42118089405c9e9493f2e30022bb7702d0e3ae.jpg" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 55, + 647, + 295, + 684 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 647, + 295, + 684 + ], + "spans": [ + { + "bbox": [ + 55, + 647, + 295, + 684 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 55, + 647, + 295, + 684 + ], + "type": "inline_equation", + "content": "\\phi_0" + }, + { + "bbox": [ + 55, + 647, + 295, + 684 + ], + "type": "text", + "content": " is defined within the source feature domain " + }, + { + "bbox": [ + 55, + 647, + 295, + 684 + ], + "type": "inline_equation", + "content": "\\mathcal{F}_{\\mathrm{tr}}" + }, + { + "bbox": [ + 55, + 647, + 295, + 684 + ], + "type": "text", + "content": ", and " + }, + { + "bbox": [ + 55, + 647, + 295, + 684 + ], + "type": "inline_equation", + "content": "\\phi_{1}" + }, + { + "bbox": [ + 55, + 647, + 295, + 684 + ], + "type": "text", + "content": " is defined in Eqn. (8). " + }, + { + "bbox": [ + 55, + 647, + 295, + 684 + ], + "type": "inline_equation", + "content": "\\epsilon" + }, + { + "bbox": [ + 55, + 647, + 295, + 684 + ], + "type": "text", + "content": " is set to 0.001 in our experiments. Accordingly, Eqn. (3) can be converted to:" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 55, + 687, + 296, + 712 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 687, + 296, + 712 + ], + "spans": [ + { + "bbox": [ + 55, + 687, + 296, + 712 + ], + "type": "interline_equation", + "content": "\\pi \\left(\\mathbf {x} _ {i} ^ {\\mathrm {n}}, \\psi \\left(\\mathbf {x} _ {i} ^ {\\mathrm {n}}\\right)\\right) = \\phi_ {0} \\left(\\mathbf {x} _ {i} ^ {\\mathrm {n}}\\right) \\exp \\left(\\frac {\\left\\langle \\mathbf {x} _ {i} ^ {\\mathrm {n}} , \\psi \\left(\\mathbf {x} _ {i} ^ {\\mathrm {n}}\\right) \\right\\rangle}{\\epsilon}\\right) \\phi_ {1} \\left(\\psi \\left(\\mathbf {x} _ {i} ^ {\\mathrm {n}}\\right)\\right), \\tag {11}", + "image_path": "f03dcf0b17af233054171bd0b0e472a90ac9e9b81384b9f783c23745ecb0f7dc.jpg" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 313, + 345, + 555, + 369 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 345, + 555, + 369 + ], + "spans": [ + { + "bbox": [ + 313, + 345, + 555, + 369 + ], + "type": "text", + "content": "Hence, the condition probability of the transport plan in Eqn. (8) could be exactly defined as" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 349, + 376, + 555, + 392 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 349, + 376, + 555, + 392 + ], + "spans": [ + { + "bbox": [ + 349, + 376, + 555, + 392 + ], + "type": "interline_equation", + "content": "\\pi (\\psi (\\mathbf {x} _ {i} ^ {\\mathrm {n}}) | \\mathbf {x} _ {i} ^ {\\mathrm {n}}) \\doteq \\eta (\\mathbf {x} _ {i} ^ {\\mathrm {n}}, \\psi (\\mathbf {x} _ {i} ^ {\\mathrm {n}})) \\phi_ {1} (\\psi (\\mathbf {x} _ {i} ^ {\\mathrm {n}})), \\tag {12}", + "image_path": "a0a409c3bd18e79fc751b38fc9048ebb1ae477f7a7f2706f22a098ad03ada263.jpg" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 313, + 398, + 555, + 463 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 398, + 555, + 463 + ], + "spans": [ + { + "bbox": [ + 313, + 398, + 555, + 463 + ], + "type": "text", + "content": "where the connection factor is denoted as " + }, + { + "bbox": [ + 313, + 398, + 555, + 463 + ], + "type": "inline_equation", + "content": "\\eta (\\mathbf{x}_i^n,\\psi (\\mathbf{x}_i^n)) = \\frac{1}{\\varpi(\\mathbf{x}_i^n)}\\exp \\left(\\frac{\\langle\\mathbf{x}_i^n,\\psi(\\mathbf{x}_i^n)\\rangle}{\\epsilon}\\right)" + }, + { + "bbox": [ + 313, + 398, + 555, + 463 + ], + "type": "text", + "content": " with the normalization term " + }, + { + "bbox": [ + 313, + 398, + 555, + 463 + ], + "type": "inline_equation", + "content": "\\varpi (\\mathbf{x}_i^n) = \\int \\exp (\\langle \\mathbf{x}_i^n,\\psi (\\mathbf{x}_i^n)\\rangle)\\phi_1(\\psi (\\mathbf{x}_i^n))d\\psi (\\mathbf{x}_i^n)" + }, + { + "bbox": [ + 313, + 398, + 555, + 463 + ], + "type": "text", + "content": ". According to Eqns. (8) and (12), we can further derive a more tractable form [10, 17]:" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 315, + 470, + 554, + 503 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 315, + 470, + 554, + 503 + ], + "spans": [ + { + "bbox": [ + 315, + 470, + 554, + 503 + ], + "type": "interline_equation", + "content": "\\pi \\left(\\psi \\left(\\mathbf {x} _ {i} ^ {\\mathrm {n}}\\right) \\mid \\mathbf {x} _ {i} ^ {\\mathrm {n}}\\right) = \\widetilde {\\eta} \\left(\\mathbf {x} _ {i} ^ {\\mathrm {n}}\\right) \\sum_ {c = 1} ^ {C} \\widetilde {\\alpha} _ {c} \\left(\\mathbf {x} _ {i} ^ {\\mathrm {n}}\\right) \\mathcal {N} \\left(\\psi \\left(\\mathbf {x} _ {i} ^ {\\mathrm {n}}\\right); \\widetilde {\\boldsymbol {\\mu}} _ {c} \\left(\\mathbf {x} _ {i} ^ {\\mathrm {n}}\\right), \\boldsymbol {\\sigma} _ {c}\\right), \\tag {13}", + "image_path": "83b0e1654c6f4a77e4b263f11f6b85c67b21965b18a5528603505a639a8061a4.jpg" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 313, + 510, + 554, + 560 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 510, + 554, + 560 + ], + "spans": [ + { + "bbox": [ + 313, + 510, + 554, + 560 + ], + "type": "text", + "content": "where the normalization factor is denoted as " + }, + { + "bbox": [ + 313, + 510, + 554, + 560 + ], + "type": "inline_equation", + "content": "\\widetilde{\\eta} (\\mathbf{x}_i^n) = 1 / \\sum_{c = 1}^{C}\\widetilde{\\alpha}_c(\\mathbf{x}_i^n)" + }, + { + "bbox": [ + 313, + 510, + 554, + 560 + ], + "type": "text", + "content": ", the coefficients of multi-Gaussian are defined as " + }, + { + "bbox": [ + 313, + 510, + 554, + 560 + ], + "type": "inline_equation", + "content": "\\widetilde{\\alpha}_{c}(\\mathbf{x}_{i}^{n}) = \\alpha_{c}\\exp (\\frac{1}{2} (\\mathbf{x}_{i}^{n})^{\\intercal}\\pmb{\\sigma}_{c}\\mathbf{x}_{i}^{n} + \\frac{1}{\\epsilon} (\\widetilde{\\mu}_{c})^{\\intercal}(\\mathbf{x}_{i}^{n}))" + }, + { + "bbox": [ + 313, + 510, + 554, + 560 + ], + "type": "text", + "content": ", and the mean vectors are calculated as " + }, + { + "bbox": [ + 313, + 510, + 554, + 560 + ], + "type": "inline_equation", + "content": "\\widetilde{\\pmb{\\mu}}_c(\\mathbf{x}_i^n) = \\pmb {\\mu}_c + \\frac{1}{\\epsilon}\\pmb {\\sigma}_c\\mathbf{x}_i^n" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 313, + 560, + 554, + 609 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 560, + 554, + 609 + ], + "spans": [ + { + "bbox": [ + 313, + 560, + 554, + 609 + ], + "type": "text", + "content": "We proceed with deriving the bridge function " + }, + { + "bbox": [ + 313, + 560, + 554, + 609 + ], + "type": "inline_equation", + "content": "\\psi_{p}" + }, + { + "bbox": [ + 313, + 560, + 554, + 609 + ], + "type": "text", + "content": " that represents a SDE process: " + }, + { + "bbox": [ + 313, + 560, + 554, + 609 + ], + "type": "inline_equation", + "content": "d\\mathbf{x}_t = g(x_t,t)dt + \\sqrt{\\epsilon} dW_t" + }, + { + "bbox": [ + 313, + 560, + 554, + 609 + ], + "type": "text", + "content": " where the shift function " + }, + { + "bbox": [ + 313, + 560, + 554, + 609 + ], + "type": "inline_equation", + "content": "g" + }, + { + "bbox": [ + 313, + 560, + 554, + 609 + ], + "type": "text", + "content": " is solved. According to Eqn. (4), we can obtain the shift function of diffusion process as follows:" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 316, + 628, + 555, + 643 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 628, + 555, + 643 + ], + "spans": [ + { + "bbox": [ + 316, + 628, + 555, + 643 + ], + "type": "interline_equation", + "content": "g (\\mathbf {x} _ {i} ^ {\\mathrm {n}}, t) = \\epsilon \\rho_ {\\mathrm {M G P}} (\\mathbf {x} _ {i} ^ {\\mathrm {n}}) \\nabla_ {\\mathbf {x} _ {i} ^ {\\mathrm {n}}} \\log (\\mathcal {N} (\\mathbf {x} _ {i} ^ {\\mathrm {n}} | 0, \\epsilon (1 - t)) \\mathbf {I}), \\quad (1 4)", + "image_path": "4de48eab232cf87e15528d2468e65e1c2e53bfda6de6620dda4dcf0d880119f6.jpg" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 313, + 649, + 555, + 716 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 649, + 555, + 716 + ], + "spans": [ + { + "bbox": [ + 313, + 649, + 555, + 716 + ], + "type": "text", + "content": "where the coefficients " + }, + { + "bbox": [ + 313, + 649, + 555, + 716 + ], + "type": "inline_equation", + "content": "\\rho_{\\mathrm{MGP}}" + }, + { + "bbox": [ + 313, + 649, + 555, + 716 + ], + "type": "text", + "content": " defined on multiple distribution prototypes is calculated as: " + }, + { + "bbox": [ + 313, + 649, + 555, + 716 + ], + "type": "inline_equation", + "content": "\\rho_{\\mathrm{MGP}}(\\mathbf{x}_i^{\\mathrm{n}}) = \\sum_{c=1}^{C} (\\alpha_c \\mathcal{N}(\\widetilde{\\boldsymbol{\\mu}}_c(\\mathbf{x}_i^{\\mathrm{n}}) | \\mathbf{0}, \\boldsymbol{\\sigma}_c) \\mathcal{N}(h_c(\\mathbf{x}_i^{\\mathrm{n}}, t) | 0, \\boldsymbol{\\Sigma}_c^t))" + }, + { + "bbox": [ + 313, + 649, + 555, + 716 + ], + "type": "text", + "content": ", with another Gaussian of " + }, + { + "bbox": [ + 313, + 649, + 555, + 716 + ], + "type": "inline_equation", + "content": "h_c(\\mathbf{x}_i^{\\mathrm{n}}, t) = \\frac{1}{\\epsilon(1 - t)} \\mathbf{x}_i^{\\mathrm{n}} + \\boldsymbol{\\sigma}_c^{-1} \\widetilde{\\boldsymbol{\\mu}}_c(\\mathbf{x}_i^{\\mathrm{n}})" + }, + { + "bbox": [ + 313, + 649, + 555, + 716 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 313, + 649, + 555, + 716 + ], + "type": "inline_equation", + "content": "\\boldsymbol{\\Sigma}_c^t = \\frac{t}{\\epsilon(1 - t)} \\mathbf{I} + \\boldsymbol{\\sigma}_c^{-1}" + }, + { + "bbox": [ + 313, + 649, + 555, + 716 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 17 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "type": "text", + "content": "20419" + } + ] + } + ], + "index": 18 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 3 + }, + { + "para_blocks": [ + { + "bbox": [ + 55, + 72, + 296, + 180 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 72, + 296, + 180 + ], + "spans": [ + { + "bbox": [ + 55, + 72, + 296, + 180 + ], + "type": "text", + "content": "Distribution prototypes initialization. Jointly learning prototypes " + }, + { + "bbox": [ + 55, + 72, + 296, + 180 + ], + "type": "inline_equation", + "content": "\\{\\alpha_c, \\mu_c, \\sigma_c\\}_{c=1}^C" + }, + { + "bbox": [ + 55, + 72, + 296, + 180 + ], + "type": "text", + "content": " and the bridge transformation " + }, + { + "bbox": [ + 55, + 72, + 296, + 180 + ], + "type": "inline_equation", + "content": "\\psi_p" + }, + { + "bbox": [ + 55, + 72, + 296, + 180 + ], + "type": "text", + "content": " (also the shift " + }, + { + "bbox": [ + 55, + 72, + 296, + 180 + ], + "type": "inline_equation", + "content": "g" + }, + { + "bbox": [ + 55, + 72, + 296, + 180 + ], + "type": "text", + "content": ") is a challenging task as they are interdependent. For this, we leverage a vector quantization function to learn a codebook " + }, + { + "bbox": [ + 55, + 72, + 296, + 180 + ], + "type": "inline_equation", + "content": "\\mathcal{E}" + }, + { + "bbox": [ + 55, + 72, + 296, + 180 + ], + "type": "text", + "content": " of prototypes within a discrete latent space from training data. Specifically, given an input image feature " + }, + { + "bbox": [ + 55, + 72, + 296, + 180 + ], + "type": "inline_equation", + "content": "\\mathbf{x}_i^n" + }, + { + "bbox": [ + 55, + 72, + 296, + 180 + ], + "type": "text", + "content": ", we can assign " + }, + { + "bbox": [ + 55, + 72, + 296, + 180 + ], + "type": "inline_equation", + "content": "\\mathbf{x}_i^n" + }, + { + "bbox": [ + 55, + 72, + 296, + 180 + ], + "type": "text", + "content": " to the closest prototype " + }, + { + "bbox": [ + 55, + 72, + 296, + 180 + ], + "type": "inline_equation", + "content": "\\mathbf{e}_k" + }, + { + "bbox": [ + 55, + 72, + 296, + 180 + ], + "type": "text", + "content": " by minimizing the L2 distance between " + }, + { + "bbox": [ + 55, + 72, + 296, + 180 + ], + "type": "inline_equation", + "content": "\\mathbf{x}_i^n" + }, + { + "bbox": [ + 55, + 72, + 296, + 180 + ], + "type": "text", + "content": " and each of prototypes " + }, + { + "bbox": [ + 55, + 72, + 296, + 180 + ], + "type": "inline_equation", + "content": "\\mathbf{e}_c \\in \\mathcal{E}" + }, + { + "bbox": [ + 55, + 72, + 296, + 180 + ], + "type": "text", + "content": ", as follows:" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 56, + 186, + 296, + 207 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 186, + 296, + 207 + ], + "spans": [ + { + "bbox": [ + 56, + 186, + 296, + 207 + ], + "type": "interline_equation", + "content": "\\min _ {\\left\\{\\mathbf {e} _ {c} \\right\\}} \\mathbb {E} _ {\\mathbf {x} _ {i} ^ {\\mathrm {n}} \\in \\mathcal {F} _ {\\mathrm {t r}}} \\left[ \\left\\| \\mathbf {x} _ {i} ^ {\\mathrm {n}} - \\mathbf {e} _ {c ^ {*}} \\right\\| _ {2} ^ {2} \\right], \\text {s . t .}, c ^ {*} = \\underset {c} {\\arg \\min } \\left\\| \\mathbf {x} _ {i} ^ {\\mathrm {n}} - \\mathbf {e} _ {c} \\right\\| _ {2}, \\tag {15}", + "image_path": "833003df8c4f7123b9e7eeee6c46a8913c645f363a2dffcd54970b402bb49888.jpg" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 55, + 212, + 296, + 272 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 212, + 296, + 272 + ], + "spans": [ + { + "bbox": [ + 55, + 212, + 296, + 272 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 55, + 212, + 296, + 272 + ], + "type": "inline_equation", + "content": "c^*" + }, + { + "bbox": [ + 55, + 212, + 296, + 272 + ], + "type": "text", + "content": " denotes the index of the prototype closest to " + }, + { + "bbox": [ + 55, + 212, + 296, + 272 + ], + "type": "inline_equation", + "content": "\\mathbf{x}_i^{n}" + }, + { + "bbox": [ + 55, + 212, + 296, + 272 + ], + "type": "text", + "content": ". The learned " + }, + { + "bbox": [ + 55, + 212, + 296, + 272 + ], + "type": "inline_equation", + "content": "\\{\\mathbf{e}_c\\}_{c=1}^C" + }, + { + "bbox": [ + 55, + 212, + 296, + 272 + ], + "type": "text", + "content": " are used to initialize the mean vectors " + }, + { + "bbox": [ + 55, + 212, + 296, + 272 + ], + "type": "inline_equation", + "content": "\\{\\pmb{\\mu}_c\\}_{c=1}^C" + }, + { + "bbox": [ + 55, + 212, + 296, + 272 + ], + "type": "text", + "content": ", while the variances of all prototypes are set to the identity matrix. We observe that this initialization strategy accelerates the training process." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 55, + 272, + 296, + 320 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 272, + 296, + 320 + ], + "spans": [ + { + "bbox": [ + 55, + 272, + 296, + 320 + ], + "type": "text", + "content": "Distribution loss of normal and anomaly samples. As " + }, + { + "bbox": [ + 55, + 272, + 296, + 320 + ], + "type": "inline_equation", + "content": "p_0" + }, + { + "bbox": [ + 55, + 272, + 296, + 320 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 55, + 272, + 296, + 320 + ], + "type": "inline_equation", + "content": "p_1" + }, + { + "bbox": [ + 55, + 272, + 296, + 320 + ], + "type": "text", + "content": " are accessible only via samples " + }, + { + "bbox": [ + 55, + 272, + 296, + 320 + ], + "type": "inline_equation", + "content": "\\mathcal{F}_{\\mathrm{tr}}" + }, + { + "bbox": [ + 55, + 272, + 296, + 320 + ], + "type": "text", + "content": " and prototypes " + }, + { + "bbox": [ + 55, + 272, + 296, + 320 + ], + "type": "inline_equation", + "content": "\\mathcal{P}_{\\mathrm{MGP}}" + }, + { + "bbox": [ + 55, + 272, + 296, + 320 + ], + "type": "text", + "content": ", we optimize the empirical form in Eqn. (2) for normal samples as follows [10, 17]:" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 67, + 326, + 295, + 360 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 67, + 326, + 295, + 360 + ], + "spans": [ + { + "bbox": [ + 67, + 326, + 295, + 360 + ], + "type": "interline_equation", + "content": "\\mathcal {L} _ {\\mathrm {D P L}} ^ {\\mathrm {n}} = \\frac {1}{N} \\sum_ {i = 1} ^ {N} \\log \\varpi_ {\\theta} \\left(\\mathbf {x} _ {i} ^ {\\mathrm {n}}\\right) - \\frac {1}{C} \\sum_ {c = 1} ^ {C} \\log \\phi_ {1} \\left(\\boldsymbol {\\mu} _ {c}\\right), \\tag {16}", + "image_path": "39806bfb22a38912f59d7d25c077bde42b707d9d414cf35ddc86e3e0e474ca4b.jpg" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 55, + 365, + 296, + 389 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 365, + 296, + 389 + ], + "spans": [ + { + "bbox": [ + 55, + 365, + 296, + 389 + ], + "type": "text", + "content": "In contrast, for anomaly samples, we aim to push anomaly distribution away from " + }, + { + "bbox": [ + 55, + 365, + 296, + 389 + ], + "type": "inline_equation", + "content": "p_1" + }, + { + "bbox": [ + 55, + 365, + 296, + 389 + ], + "type": "text", + "content": ", i.e., negative loss, formally," + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 66, + 395, + 295, + 428 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 66, + 395, + 295, + 428 + ], + "spans": [ + { + "bbox": [ + 66, + 395, + 295, + 428 + ], + "type": "interline_equation", + "content": "\\mathcal {L} _ {\\mathrm {D P L}} ^ {\\mathrm {a}} = \\frac {1}{C} \\sum_ {c = 1} ^ {C} \\log \\phi_ {1} (\\boldsymbol {\\mu} _ {c}) - \\frac {1}{M} \\sum_ {i = 1} ^ {M} \\log \\varpi_ {\\theta} \\left(\\mathbf {x} _ {i} ^ {\\mathrm {a}}\\right), \\tag {17}", + "image_path": "0f2ed598b881e0978f0438b36273184ee82e437c9328b3cdf5af65bfa85aeac7.jpg" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 55, + 433, + 214, + 445 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 433, + 214, + 445 + ], + "spans": [ + { + "bbox": [ + 55, + 433, + 214, + 445 + ], + "type": "text", + "content": "4.3. Dispersion Feature Learning" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 55, + 450, + 296, + 594 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 450, + 296, + 594 + ], + "spans": [ + { + "bbox": [ + 55, + 450, + 296, + 594 + ], + "type": "text", + "content": "A critical challenge in OSAD is detecting previously unseen anomalies in open-set environments. Due to the limited anomaly observations, existing methods often use pseudo-anomaly generation strategies as a means of effective data augmentation. Nonetheless, the effectiveness of these methods heavily depends on the quality of the pseudo-anomaly feature embeddings. Notably, pseudo-anomaly distributions often inherit biases from the in-distribution data, which differ from the unknown anomaly distributions in out-of-distribution data. Current methods fail to address the relationship between observed and unknown anomalies, particularly for out-of-distribution generalization." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 55, + 594, + 296, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 594, + 296, + 714 + ], + "spans": [ + { + "bbox": [ + 55, + 594, + 296, + 714 + ], + "type": "text", + "content": "To improve out-of-distribution detection, it is essential to promote a larger inter-sample dispersion, as greater distances among in-distribution samples facilitate their more effective separation from out-of-distribution samples. In other words, if all inter-sample distances are close to zero, i.e., collapse to a single point, it becomes impossible to differentiate between samples. Hence, promoting separability through a larger inter-sample dispersion is critical for accurately identifying samples that do not belong to known categories. To do so, we map the features into a hyperspherical" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 313, + 72, + 554, + 133 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 72, + 554, + 133 + ], + "spans": [ + { + "bbox": [ + 313, + 72, + 554, + 133 + ], + "type": "text", + "content": "space. Draws inspiration from the vMF distribution [31] in directional statistics, we compute spherical Gaussian distributions for unit-norm features " + }, + { + "bbox": [ + 313, + 72, + 554, + 133 + ], + "type": "inline_equation", + "content": "\\widehat{\\mathbf{x}}_i = \\mathbf{x}_i / \\| \\mathbf{x}_i\\| _2^2" + }, + { + "bbox": [ + 313, + 72, + 554, + 133 + ], + "type": "text", + "content": ". The probability density function of a unit vector " + }, + { + "bbox": [ + 313, + 72, + 554, + 133 + ], + "type": "inline_equation", + "content": "\\widehat{\\mathbf{x}}_i\\in \\mathbb{R}^D" + }, + { + "bbox": [ + 313, + 72, + 554, + 133 + ], + "type": "text", + "content": " is defined in the hyperspherical space as follows:" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 351, + 142, + 555, + 156 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 351, + 142, + 555, + 156 + ], + "spans": [ + { + "bbox": [ + 351, + 142, + 555, + 156 + ], + "type": "interline_equation", + "content": "p _ {D} (\\widehat {\\mathbf {x}} _ {i}; \\widehat {\\mathbf {x}} _ {j}, \\kappa) = F _ {D} (\\kappa) \\exp \\left(\\kappa \\langle \\widehat {\\mathbf {x}} _ {i}, \\widehat {\\mathbf {x}} _ {j} \\rangle\\right), \\tag {18}", + "image_path": "2d078257ffe90dbc524508c125c998be1d0f01051fa6e9e679fc03961127ffc7.jpg" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 313, + 165, + 555, + 250 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 165, + 555, + 250 + ], + "spans": [ + { + "bbox": [ + 313, + 165, + 555, + 250 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 313, + 165, + 555, + 250 + ], + "type": "inline_equation", + "content": "\\kappa \\geq 0" + }, + { + "bbox": [ + 313, + 165, + 555, + 250 + ], + "type": "text", + "content": " controls the concentration of the distribution around the mean direction " + }, + { + "bbox": [ + 313, + 165, + 555, + 250 + ], + "type": "inline_equation", + "content": "\\widehat{\\mathbf{x}}_j" + }, + { + "bbox": [ + 313, + 165, + 555, + 250 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 313, + 165, + 555, + 250 + ], + "type": "inline_equation", + "content": "F_{D}(\\kappa)" + }, + { + "bbox": [ + 313, + 165, + 555, + 250 + ], + "type": "text", + "content": " is the normalization factor. A larger " + }, + { + "bbox": [ + 313, + 165, + 555, + 250 + ], + "type": "inline_equation", + "content": "\\kappa" + }, + { + "bbox": [ + 313, + 165, + 555, + 250 + ], + "type": "text", + "content": " value increases concentration around the mean, while in the extreme case " + }, + { + "bbox": [ + 313, + 165, + 555, + 250 + ], + "type": "inline_equation", + "content": "\\kappa = 0" + }, + { + "bbox": [ + 313, + 165, + 555, + 250 + ], + "type": "text", + "content": ", sample points are uniformly distributed on the hypersphere. Therefore, we design a dispersion loss to optimize large angular distances between the features of all samples:" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 315, + 258, + 555, + 293 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 315, + 258, + 555, + 293 + ], + "spans": [ + { + "bbox": [ + 315, + 258, + 555, + 293 + ], + "type": "interline_equation", + "content": "\\mathcal {L} _ {\\mathrm {D F L}} = \\frac {1}{U} \\sum_ {i = 1} ^ {U} \\log \\frac {1}{U - 1} \\sum_ {i, j = 1} ^ {U} \\mathbb {1} \\{i \\neq j \\} \\exp (\\kappa \\langle \\widehat {\\mathbf {x}} _ {i}, \\widehat {\\mathbf {x}} _ {j} \\rangle), \\tag {19}", + "image_path": "f6003b0ea9ef3fbad95a4734991253f697661e9d75e7344393c8c8aa1917351c.jpg" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 313, + 300, + 550, + 313 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 300, + 550, + 313 + ], + "spans": [ + { + "bbox": [ + 313, + 300, + 550, + 313 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 313, + 300, + 550, + 313 + ], + "type": "inline_equation", + "content": "U = N + M" + }, + { + "bbox": [ + 313, + 300, + 550, + 313 + ], + "type": "text", + "content": ", and " + }, + { + "bbox": [ + 313, + 300, + 550, + 313 + ], + "type": "inline_equation", + "content": "\\kappa" + }, + { + "bbox": [ + 313, + 300, + 550, + 313 + ], + "type": "text", + "content": " is set to 10 in our experiments." + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 313, + 321, + 460, + 334 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 321, + 460, + 334 + ], + "spans": [ + { + "bbox": [ + 313, + 321, + 460, + 334 + ], + "type": "text", + "content": "4.4. Anomaly Score Prediction" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 313, + 339, + 555, + 495 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 339, + 555, + 495 + ], + "spans": [ + { + "bbox": [ + 313, + 339, + 555, + 495 + ], + "type": "text", + "content": "Based on the above designs, we leverage the multiple-instance-learning (MIL)-based method proposed in [35] to effectively learn anomaly scores. Similar to the work [9], we design three modules " + }, + { + "bbox": [ + 313, + 339, + 555, + 495 + ], + "type": "inline_equation", + "content": "\\mathsf{M} = \\{\\mathsf{M}_{\\mathsf{a}},\\mathsf{M}_{\\mathsf{n}}" + }, + { + "bbox": [ + 313, + 339, + 555, + 495 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 313, + 339, + 555, + 495 + ], + "type": "inline_equation", + "content": "\\mathsf{M}_{\\mathsf{r}}\\}" + }, + { + "bbox": [ + 313, + 339, + 555, + 495 + ], + "type": "text", + "content": " for estimating anomaly scores. Firstly, for the feature map " + }, + { + "bbox": [ + 313, + 339, + 555, + 495 + ], + "type": "inline_equation", + "content": "\\mathbf{x}_i^1" + }, + { + "bbox": [ + 313, + 339, + 555, + 495 + ], + "type": "text", + "content": ", we generate pixel-wise feature vectors " + }, + { + "bbox": [ + 313, + 339, + 555, + 495 + ], + "type": "inline_equation", + "content": "\\mathcal{V} = \\{\\mathbf{v}_i\\}_{i = 1}^{H' \\times W'}" + }, + { + "bbox": [ + 313, + 339, + 555, + 495 + ], + "type": "text", + "content": " to represent the feature of small patches of the image " + }, + { + "bbox": [ + 313, + 339, + 555, + 495 + ], + "type": "inline_equation", + "content": "\\mathbf{x}_i" + }, + { + "bbox": [ + 313, + 339, + 555, + 495 + ], + "type": "text", + "content": ", where " + }, + { + "bbox": [ + 313, + 339, + 555, + 495 + ], + "type": "inline_equation", + "content": "(H',W')" + }, + { + "bbox": [ + 313, + 339, + 555, + 495 + ], + "type": "text", + "content": " denotes the size of the feature map. These pixel-wise representations are then mapped by an anomaly classifier " + }, + { + "bbox": [ + 313, + 339, + 555, + 495 + ], + "type": "inline_equation", + "content": "S_{\\mathrm{a}}" + }, + { + "bbox": [ + 313, + 339, + 555, + 495 + ], + "type": "text", + "content": " to estimate pixel-level anomaly scores. To capture those points with the most salient anomalies, we compute the top-" + }, + { + "bbox": [ + 313, + 339, + 555, + 495 + ], + "type": "inline_equation", + "content": "K" + }, + { + "bbox": [ + 313, + 339, + 555, + 495 + ], + "type": "text", + "content": " most anomaly pixel points and define the loss function as:" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 319, + 502, + 555, + 526 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 319, + 502, + 555, + 526 + ], + "spans": [ + { + "bbox": [ + 319, + 502, + 555, + 526 + ], + "type": "interline_equation", + "content": "\\mathcal {L} _ {\\mathrm {M} _ {\\mathrm {a}}} \\left(\\mathbf {x} _ {i}, y _ {i}\\right) = \\mathcal {L} _ {\\text {b i n a r y}} \\left(\\frac {1}{K} \\sum \\operatorname {T o p} _ {K} \\left\\{S _ {\\mathrm {a}} \\left(\\mathbf {v} _ {i}; \\theta_ {\\mathrm {a}}\\right) \\right\\}, y _ {i}\\right), \\tag {20}", + "image_path": "62ead27aec93c71439ed30ce9821f4c10b5e12de6c95afe7455f51fbcfbedcbf.jpg" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 313, + 533, + 554, + 570 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 533, + 554, + 570 + ], + "spans": [ + { + "bbox": [ + 313, + 533, + 554, + 570 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 313, + 533, + 554, + 570 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{\\mathrm{binary}}" + }, + { + "bbox": [ + 313, + 533, + 554, + 570 + ], + "type": "text", + "content": " refers to a binary classification loss function, and " + }, + { + "bbox": [ + 313, + 533, + 554, + 570 + ], + "type": "inline_equation", + "content": "\\operatorname{Top}_K" + }, + { + "bbox": [ + 313, + 533, + 554, + 570 + ], + "type": "text", + "content": " selects the highest " + }, + { + "bbox": [ + 313, + 533, + 554, + 570 + ], + "type": "inline_equation", + "content": "K" + }, + { + "bbox": [ + 313, + 533, + 554, + 570 + ], + "type": "text", + "content": " anomaly scores among all the vectors. We use " + }, + { + "bbox": [ + 313, + 533, + 554, + 570 + ], + "type": "inline_equation", + "content": "\\mathsf{M}_{\\mathfrak{n}}" + }, + { + "bbox": [ + 313, + 533, + 554, + 570 + ], + "type": "text", + "content": " to learn the normal features:" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 321, + 578, + 553, + 622 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 321, + 578, + 553, + 622 + ], + "spans": [ + { + "bbox": [ + 321, + 578, + 553, + 622 + ], + "type": "interline_equation", + "content": "\\mathcal {L} _ {\\mathrm {M} _ {\\mathrm {n}}} \\left(\\mathbf {x} _ {i}, y _ {i}\\right) = \\mathcal {L} _ {\\text {b i n a r y}} \\left(S _ {\\mathrm {n}} \\left(\\frac {1}{H ^ {\\prime} \\times W ^ {\\prime}} \\sum_ {i = 1} ^ {H ^ {\\prime} \\times W ^ {\\prime}} \\mathbf {v} _ {i}; \\theta_ {\\mathrm {n}}\\right), y _ {i}\\right), \\tag {21}", + "image_path": "7a121937f986ae5a33a6f94fddbedfe214926eac62b11ed279ff4c0d7d205b20.jpg" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 313, + 623, + 554, + 671 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 623, + 554, + 671 + ], + "spans": [ + { + "bbox": [ + 313, + 623, + 554, + 671 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 313, + 623, + 554, + 671 + ], + "type": "inline_equation", + "content": "S_{\\mathfrak{n}}:\\mathcal{V}\\to \\mathbb{R}" + }, + { + "bbox": [ + 313, + 623, + 554, + 671 + ], + "type": "text", + "content": " is a fully connected binary anomaly classifier. Finally, we define " + }, + { + "bbox": [ + 313, + 623, + 554, + 671 + ], + "type": "inline_equation", + "content": "\\mathsf{M}_{\\mathfrak{x}}" + }, + { + "bbox": [ + 313, + 623, + 554, + 671 + ], + "type": "text", + "content": " to compute the residual anomaly scores between fine-grained visual semantics and abstract prototypes:" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 326, + 680, + 555, + 695 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 326, + 680, + 555, + 695 + ], + "spans": [ + { + "bbox": [ + 326, + 680, + 555, + 695 + ], + "type": "interline_equation", + "content": "\\mathcal {L} _ {\\mathrm {M} _ {\\mathrm {r}}} = \\mathcal {L} _ {\\text {b i n a r y}} \\left(S _ {\\mathrm {r}} \\left(\\left(\\psi_ {p} (\\mathbf {x} _ {i}) - \\boldsymbol {\\mu} _ {c ^ {*}}\\right) / \\boldsymbol {\\sigma} _ {c ^ {*}}; \\theta_ {\\mathrm {r}}\\right), y _ {i}\\right), \\tag {22}", + "image_path": "fad53808f40448f7e06b03355b146377893220fc3de48e1e5446cea80fb1eff6.jpg" + } + ] + } + ], + "index": 21 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 325, + 702, + 553, + 713 + ], + "type": "page_footnote", + "angle": 0, + "lines": [ + { + "bbox": [ + 325, + 702, + 553, + 713 + ], + "spans": [ + { + "bbox": [ + 325, + 702, + 553, + 713 + ], + "type": "text", + "content": "Here " + }, + { + "bbox": [ + 325, + 702, + 553, + 713 + ], + "type": "inline_equation", + "content": "\\mathbf{x}_i" + }, + { + "bbox": [ + 325, + 702, + 553, + 713 + ], + "type": "text", + "content": " refers to convolution maps without flattening (as used above)." + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "type": "footer", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "type": "text", + "content": "20420" + } + ] + } + ], + "index": 23 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 4 + }, + { + "para_blocks": [ + { + "bbox": [ + 55, + 72, + 294, + 108 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 72, + 294, + 108 + ], + "spans": [ + { + "bbox": [ + 55, + 72, + 294, + 108 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 55, + 72, + 294, + 108 + ], + "type": "inline_equation", + "content": "c^*" + }, + { + "bbox": [ + 55, + 72, + 294, + 108 + ], + "type": "text", + "content": " denotes the index of the most probable prototypes, i.e., " + }, + { + "bbox": [ + 55, + 72, + 294, + 108 + ], + "type": "inline_equation", + "content": "c^* = \\arg \\max_c\\mathcal{N}(\\psi_p(\\mathbf{x}_i);\\pmb {\\mu}_c,\\pmb {\\sigma}_c)" + }, + { + "bbox": [ + 55, + 72, + 294, + 108 + ], + "type": "text", + "content": ", and " + }, + { + "bbox": [ + 55, + 72, + 294, + 108 + ], + "type": "inline_equation", + "content": "S_{\\mathrm{r}}" + }, + { + "bbox": [ + 55, + 72, + 294, + 108 + ], + "type": "text", + "content": " utilize the same method to obtain anomaly score as " + }, + { + "bbox": [ + 55, + 72, + 294, + 108 + ], + "type": "inline_equation", + "content": "S_{\\mathrm{a}}" + }, + { + "bbox": [ + 55, + 72, + 294, + 108 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 55, + 109, + 295, + 157 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 109, + 295, + 157 + ], + "spans": [ + { + "bbox": [ + 55, + 109, + 295, + 157 + ], + "type": "text", + "content": "Training. During the training phase, the SB and three prediction modules are jointly trained. To this end, we employ an objective function that encompasses three components as follows:" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 63, + 168, + 295, + 196 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 63, + 168, + 295, + 196 + ], + "spans": [ + { + "bbox": [ + 63, + 168, + 295, + 196 + ], + "type": "interline_equation", + "content": "\\mathcal {L} = \\underbrace {\\mathcal {L} _ {\\mathrm {M} _ {\\mathrm {a}}} + \\mathcal {L} _ {\\mathrm {M} _ {\\mathrm {n}}} + \\mathcal {L} _ {\\mathrm {M} _ {\\mathrm {r}}}} _ {\\text {M I L - b a s e d l e a n i n g}} + \\underbrace {\\mathcal {L} _ {\\mathrm {D P L}} ^ {\\mathrm {n}} + \\mathcal {L} _ {\\mathrm {D P L}} ^ {\\mathrm {a}}} _ {\\text {S B t r a n s f o r m}} + \\underbrace {\\lambda \\mathcal {L} _ {\\mathrm {D F L}}} _ {\\text {d i s p e r s i o n}}. \\tag {23}", + "image_path": "158288081c60cb35df208dc36bdc9320ba97e874709340dc31f221bac76b17de.jpg" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 55, + 207, + 295, + 243 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 207, + 295, + 243 + ], + "spans": [ + { + "bbox": [ + 55, + 207, + 295, + 243 + ], + "type": "text", + "content": "where the coefficient " + }, + { + "bbox": [ + 55, + 207, + 295, + 243 + ], + "type": "inline_equation", + "content": "\\lambda" + }, + { + "bbox": [ + 55, + 207, + 295, + 243 + ], + "type": "text", + "content": " modulates the relative importance of dispersion loss, and the learnable parameters include " + }, + { + "bbox": [ + 55, + 207, + 295, + 243 + ], + "type": "inline_equation", + "content": "\\{\\alpha_{c},\\pmb{\\mu}_{c},\\pmb{\\sigma}_{c}\\}" + }, + { + "bbox": [ + 55, + 207, + 295, + 243 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 55, + 207, + 295, + 243 + ], + "type": "inline_equation", + "content": "\\theta_{\\psi_p}" + }, + { + "bbox": [ + 55, + 207, + 295, + 243 + ], + "type": "text", + "content": ", and " + }, + { + "bbox": [ + 55, + 207, + 295, + 243 + ], + "type": "inline_equation", + "content": "\\{\\theta_{\\mathrm{a}},\\theta_{\\mathrm{n}},\\theta_{\\mathrm{r}}\\}" + }, + { + "bbox": [ + 55, + 207, + 295, + 243 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 55, + 244, + 295, + 304 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 244, + 295, + 304 + ], + "spans": [ + { + "bbox": [ + 55, + 244, + 295, + 304 + ], + "type": "text", + "content": "Inference. During the test phase, we find the most similar class prototype through SB, and subsequently compute the anomaly score by adding the scores from both " + }, + { + "bbox": [ + 55, + 244, + 295, + 304 + ], + "type": "inline_equation", + "content": "S_{\\mathrm{a}}" + }, + { + "bbox": [ + 55, + 244, + 295, + 304 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 55, + 244, + 295, + 304 + ], + "type": "inline_equation", + "content": "S_{\\mathrm{r}}" + }, + { + "bbox": [ + 55, + 244, + 295, + 304 + ], + "type": "text", + "content": ", while subtracting the normal score obtained from " + }, + { + "bbox": [ + 55, + 244, + 295, + 304 + ], + "type": "inline_equation", + "content": "S_{\\mathrm{n}}" + }, + { + "bbox": [ + 55, + 244, + 295, + 304 + ], + "type": "text", + "content": " for the given test image." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 55, + 317, + 132, + 330 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 317, + 132, + 330 + ], + "spans": [ + { + "bbox": [ + 55, + 317, + 132, + 330 + ], + "type": "text", + "content": "5. Experiment" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 55, + 338, + 222, + 350 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 338, + 222, + 350 + ], + "spans": [ + { + "bbox": [ + 55, + 338, + 222, + 350 + ], + "type": "text", + "content": "5.1. Dataset and Evaluation Metric" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 55, + 356, + 295, + 499 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 356, + 295, + 499 + ], + "spans": [ + { + "bbox": [ + 55, + 356, + 295, + 499 + ], + "type": "text", + "content": "Dataset To validate the effectiveness of DPDL, comprehensive experiments are conducted on nine real-world AD datasets, including six industrial defect detection datasets (MVTec AD [3], Optical [46], SDD [40], AITEX [39], ELPV [8], Mastcam [13]) and three medical image datasets (Hyper-Kvasir [4], Brain-MRI [37], HeadCT [37]). We follow the previous OSAD baselines [9, 51] to adopt two protocols for sampling, including general setting and hard setting. The general setting assumes that anomaly examples are randomly sampled from the anomaly class, while the hard setting samples from a single class to assess generalization to new or unseen anomaly classes." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 55, + 500, + 295, + 548 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 500, + 295, + 548 + ], + "spans": [ + { + "bbox": [ + 55, + 500, + 295, + 548 + ], + "type": "text", + "content": "Evaluation Metric We utilize the widely adopted Area Under ROC Curve (AUC) as a metric to evaluate the performance across all methods and settings. All reported AUCs are averaged results over five independent runs." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 55, + 559, + 122, + 571 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 559, + 122, + 571 + ], + "spans": [ + { + "bbox": [ + 55, + 559, + 122, + 571 + ], + "type": "text", + "content": "5.2. Baselines" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 55, + 577, + 295, + 661 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 577, + 295, + 661 + ], + "spans": [ + { + "bbox": [ + 55, + 577, + 295, + 661 + ], + "type": "text", + "content": "We compare DPDL against six related state-of-the-art OSAD baselines, including SAOE [19, 32, 41], MLEP [27], FLOS [36], DevNet [35], DRA [9], and AHL [51]. MLEP, DevNet, DRA, and AHL are specifically designed for OSAD. SAOE is a supervised detector enhanced with synthetic anomalies and anomaly exposure, whereas FLOS is an imbalanced classifier leveraging focal loss." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 55, + 671, + 188, + 684 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 671, + 188, + 684 + ], + "spans": [ + { + "bbox": [ + 55, + 671, + 188, + 684 + ], + "type": "text", + "content": "5.3. Implementation Details" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 55, + 689, + 295, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 689, + 295, + 713 + ], + "spans": [ + { + "bbox": [ + 55, + 689, + 295, + 713 + ], + "type": "text", + "content": "The input image size is " + }, + { + "bbox": [ + 55, + 689, + 295, + 713 + ], + "type": "inline_equation", + "content": "448 \\times 448 \\times 3" + }, + { + "bbox": [ + 55, + 689, + 295, + 713 + ], + "type": "text", + "content": ". We set " + }, + { + "bbox": [ + 55, + 689, + 295, + 713 + ], + "type": "inline_equation", + "content": "K" + }, + { + "bbox": [ + 55, + 689, + 295, + 713 + ], + "type": "text", + "content": " in the top- " + }, + { + "bbox": [ + 55, + 689, + 295, + 713 + ], + "type": "inline_equation", + "content": "K" + }, + { + "bbox": [ + 55, + 689, + 295, + 713 + ], + "type": "text", + "content": " MIL to " + }, + { + "bbox": [ + 55, + 689, + 295, + 713 + ], + "type": "inline_equation", + "content": "10\\%" + }, + { + "bbox": [ + 55, + 689, + 295, + 713 + ], + "type": "text", + "content": " of the number of all scores per score map." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 313, + 72, + 555, + 193 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 72, + 555, + 193 + ], + "spans": [ + { + "bbox": [ + 313, + 72, + 555, + 193 + ], + "type": "text", + "content": "AdamW optimizer [29] is used for the parameter optimization using an initial learning rate " + }, + { + "bbox": [ + 313, + 72, + 555, + 193 + ], + "type": "inline_equation", + "content": "2 \\times 10^{-4}" + }, + { + "bbox": [ + 313, + 72, + 555, + 193 + ], + "type": "text", + "content": " with a weight decay of " + }, + { + "bbox": [ + 313, + 72, + 555, + 193 + ], + "type": "inline_equation", + "content": "1 \\times 10^{-5}" + }, + { + "bbox": [ + 313, + 72, + 555, + 193 + ], + "type": "text", + "content": ". DPDL is trained on one NVIDIA GeForce RTX 4090 GPU, which are trained using 50 epochs, with 20 iterations per epoch. Following previous protocol [9, 51], we evaluate performance with anomaly sample numbers of " + }, + { + "bbox": [ + 313, + 72, + 555, + 193 + ], + "type": "inline_equation", + "content": "M = 10" + }, + { + "bbox": [ + 313, + 72, + 555, + 193 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 313, + 72, + 555, + 193 + ], + "type": "inline_equation", + "content": "M = 1" + }, + { + "bbox": [ + 313, + 72, + 555, + 193 + ], + "type": "text", + "content": ", and for robust detection of unseen anomalies, we use CutMix [49] to create pseudo-anomaly samples as augmented data for known anomalies. The prototype quantity " + }, + { + "bbox": [ + 313, + 72, + 555, + 193 + ], + "type": "inline_equation", + "content": "C" + }, + { + "bbox": [ + 313, + 72, + 555, + 193 + ], + "type": "text", + "content": " is set to 32 as default." + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 313, + 210, + 479, + 223 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 210, + 479, + 223 + ], + "spans": [ + { + "bbox": [ + 313, + 210, + 479, + 223 + ], + "type": "text", + "content": "5.4. Results under General Setting" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 313, + 232, + 555, + 542 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 232, + 555, + 542 + ], + "spans": [ + { + "bbox": [ + 313, + 232, + 555, + 542 + ], + "type": "text", + "content": "Tab. 1 highlights DPDL's strong performance. In the challenging scenario of single-anomaly detection, DPDL improves the performance of AHL [51] by more than " + }, + { + "bbox": [ + 313, + 232, + 555, + 542 + ], + "type": "inline_equation", + "content": "8.3\\%" + }, + { + "bbox": [ + 313, + 232, + 555, + 542 + ], + "type": "text", + "content": " on the AITEX, ELPV, and Mastcam datasets. Furthermore, it achieves significant improvements across six additional datasets, which suggest the effective utilization of few-shot anomaly examples in DPDL, while mitigating overfitting to the seen anomalies. When shifting to ten anomaly examples settings, DPDL continues to maintain a significant lead with over " + }, + { + "bbox": [ + 313, + 232, + 555, + 542 + ], + "type": "inline_equation", + "content": "5.4\\%" + }, + { + "bbox": [ + 313, + 232, + 555, + 542 + ], + "type": "text", + "content": " improvement on those datasets. Given the rich and diverse set of normal samples in these datasets, DPDL leverages the DPL component to encapsulate these samples within a compact, discriminative distribution space, while effectively pushing anomalous samples outside this space, thereby enabling accurate anomaly detection. In the setting with ten abnormal samples, although existing methods have reached performance saturation on the MVTecAD, Optical, and SDD datasets, DPDL still has a certain lead, demonstrating the strong ability of DPL and DFL in learning tight boundaries of normal sample distributions and generalizing to previously unseen anomaly domains. Furthermore, when evaluated on the medical datasets BrainMRI and HeadCT, DPDL demonstrates competitive performance despite these datasets being notably small in scale and containing only a single class. This highlights the algorithm's ability to deliver robust results even in data-scarce conditions." + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 313, + 561, + 484, + 574 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 561, + 484, + 574 + ], + "spans": [ + { + "bbox": [ + 313, + 561, + 484, + 574 + ], + "type": "text", + "content": "5.5. Results under the Hard Setting" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 313, + 582, + 555, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 582, + 555, + 713 + ], + "spans": [ + { + "bbox": [ + 313, + 582, + 555, + 713 + ], + "type": "text", + "content": "Tab. 2 summarizes the performance comparison under the hard setting. It is evident that DPDL achieves the highest AUC scores in both the single-anomaly and ten-anomaly sample settings. Specifically, compared to the closest competing method, AHL [51], DPDL achieves an improvement in AUC scores ranging from " + }, + { + "bbox": [ + 313, + 582, + 555, + 713 + ], + "type": "inline_equation", + "content": "0.6\\%" + }, + { + "bbox": [ + 313, + 582, + 555, + 713 + ], + "type": "text", + "content": " to " + }, + { + "bbox": [ + 313, + 582, + 555, + 713 + ], + "type": "inline_equation", + "content": "7.9\\%" + }, + { + "bbox": [ + 313, + 582, + 555, + 713 + ], + "type": "text", + "content": " in the ten anomaly examples settings and from " + }, + { + "bbox": [ + 313, + 582, + 555, + 713 + ], + "type": "inline_equation", + "content": "0.5\\%" + }, + { + "bbox": [ + 313, + 582, + 555, + 713 + ], + "type": "text", + "content": " to " + }, + { + "bbox": [ + 313, + 582, + 555, + 713 + ], + "type": "inline_equation", + "content": "8.9\\%" + }, + { + "bbox": [ + 313, + 582, + 555, + 713 + ], + "type": "text", + "content": " in the one anomaly example settings, respectively. The observed improvement can be attributed to the strong generalization ability of DPDL in detecting unseen anomaly classes, even when the model is trained on only a single anomaly class." + } + ] + } + ], + "index": 17 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "text", + "content": "20421" + } + ] + } + ], + "index": 18 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 5 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 60, + 101, + 548, + 339 + ], + "blocks": [ + { + "bbox": [ + 55, + 71, + 555, + 94 + ], + "lines": [ + { + "bbox": [ + 55, + 71, + 555, + 94 + ], + "spans": [ + { + "bbox": [ + 55, + 71, + 555, + 94 + ], + "type": "text", + "content": "Table 1. AUC performance (mean ± std) across nine real-world AD datasets is reported under the general setting. red highlights the best results, and blue indicates sub-optimal outcomes. All baseline SOTA results are sourced from the original papers [9, 51]." + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 60, + 101, + 548, + 339 + ], + "lines": [ + { + "bbox": [ + 60, + 101, + 548, + 339 + ], + "spans": [ + { + "bbox": [ + 60, + 101, + 548, + 339 + ], + "type": "table", + "html": "
DatasetDevNetFLOSSAOEMLEPDRAAHLDPDL (Ours)
Ten Training Anomaly Examples
MVTec AD0.945±0.0040.939±0.0070.926±0.0100.907±0.0050.959±0.0030.970±0.0020.977±0.002
Optical0.782±0.0650.720±0.0550.941±0.0130.740±0.0390.965±0.0060.976±0.0040.983±0.005
SDD0.988±0.0060.967±0.0180.955±0.0200.983±0.0130.991±0.0050.991±0.0010.996±0.001
AITEX0.887±0.0130.841±0.0490.874±0.0240.867±0.0370.893±0.0170.925±0.0130.975±0.007
ELPV0.846±0.0220.818±0.0320.793±0.0470.794±0.0470.845±0.0130.850±0.0040.937±0.003
Mastcam0.790±0.0210.703±0.0290.810±0.0290.798±0.0260.848±0.0080.855±0.0050.934±0.010
Hyper-Kvasir0.829±0.0180.773±0.0290.666±0.0500.600±0.0690.834±0.0040.880±0.0030.939±0.005
BrainMRI0.958±0.0120.955±0.0110.900±0.0410.959±0.0110.970±0.0030.977±0.0010.969±0.005
HeadCT0.982±0.0090.971±0.0040.935±0.0210.972±0.0140.972±0.0020.999±0.0030.981±0.003
One Training Anomaly Example
MVTec AD0.780±0.0200.755±0.1360.834±0.0070.744±0.0190.883±0.0080.901±0.0030.927±0.002
Optical0.523±0.0030.518±0.0030.815±0.0140.516±0.0090.888±0.0120.888±0.0070.915±0.002
SDD0.881±0.0090.840±0.0430.781±0.0090.811±0.0450.859±0.0140.909±0.0010.917±0.003
AITEX0.598±0.0700.538±0.0730.675±0.0940.564±0.0550.692±0.1240.734±0.0080.838±0.008
ELPV0.514±0.0760.457±0.0560.635±0.0920.578±0.0620.675±0.0240.828±0.0050.897±0.002
Mastcam0.595±0.0160.542±0.0170.662±0.0180.625±0.0450.692±0.0580.743±0.0030.838±0.011
Hyper-Kvasir0.653±0.0370.668±0.0040.498±0.1000.445±0.0400.690±0.0170.768±0.0150.821±0.007
BrainMRI0.694±0.0040.693±0.0360.531±0.0600.632±0.0170.744±0.0040.866±0.0040.893±0.004
HeadCT0.742±0.0760.698±0.0920.597±0.0220.758±0.0380.796±0.1050.825±0.0140.865±0.005
", + "image_path": "e96ce7dd2aa97b803eb69c971b7f9b6a0303ef52040b6330cca4c42d806bfc4b.jpg" + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "table_body" + } + ], + "index": 1 + }, + { + "type": "table", + "bbox": [ + 59, + 390, + 548, + 562 + ], + "blocks": [ + { + "bbox": [ + 55, + 348, + 555, + 381 + ], + "lines": [ + { + "bbox": [ + 55, + 348, + 555, + 381 + ], + "spans": [ + { + "bbox": [ + 55, + 348, + 555, + 381 + ], + "type": "text", + "content": "Table 2. AUC results (mean ± std) under the hard setting. The best and second-best results are highlighted in red and blue, respectively. Carpet and Metal_nut are subsets of MVTec AD. The datasets used are consistent with those in [9, 51], where those datasets only containing one anomaly class are excluded to adapt for the hard setting." + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 59, + 390, + 548, + 562 + ], + "lines": [ + { + "bbox": [ + 59, + 390, + 548, + 562 + ], + "spans": [ + { + "bbox": [ + 59, + 390, + 548, + 562 + ], + "type": "table", + "html": "
DatasetDevNetFLOSSAOEMLEPDRAAHLDPDL (Ours)
Ten Training Anomaly Examples
Carpet (mean)0.847±0.0170.761±0.0120.762±0.0730.751±0.0230.935±0.0130.949±0.0020.956±0.004
Metal_nut (mean)0.965±0.0110.922±0.0140.855±0.0160.878±0.0580.945±0.0170.972±0.0020.978±0.002
AITEX (mean)0.683±0.0320.635±0.0430.724±0.0320.626±0.0410.733±0.0090.747±0.0020.798±0.005
ELPV (mean)0.702±0.0230.642±0.0320.683±0.0470.745±0.0200.766±0.0290.788±0.0030.818±0.003
Mastcam (mean)0.588±0.0110.616±0.0210.697±0.0140.588±0.0160.695±0.0040.721±0.0030.778±0.007
Hyper-Kvasir (mean)0.822±0.0190.786±0.0210.698±0.0210.571±0.0140.844±0.0090.854±0.0040.864±0.002
One Training Anomaly Example
Carpet (mean)0.767±0.0180.678±0.0400.753±0.0550.679±0.0290.901±0.0060.932±0.0030.941±0.006
Metal_nut (mean)0.855±0.0160.855±0.0240.816±0.0290.825±0.0230.932±0.0170.939±0.0040.944±0.003
AITEX (mean)0.646±0.0340.624±0.0240.674±0.0340.466±0.0300.684±0.0330.707±0.0070.753±0.005
ELPV (mean)0.648±0.0570.691±0.0080.614±0.0480.566±0.1110.703±0.0220.740±0.0030.762±0.003
Mastcam (mean)0.511±0.0130.524±0.0130.689±0.0370.541±0.0070.667±0.0120.673±0.0100.733±0.004
Hyper-Kvasir (mean)0.595±0.0230.571±0.0040.406±0.0180.480±0.0440.700±0.0090.706±0.0070.715±0.004
", + "image_path": "0b78202087eed264bedb97e4a51d3e55c7df39c6d85bb2e317abf2d47ab20bf9.jpg" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "table_body" + } + ], + "index": 3 + }, + { + "type": "image", + "bbox": [ + 59, + 596, + 181, + 706 + ], + "blocks": [ + { + "bbox": [ + 106, + 588, + 258, + 596 + ], + "lines": [ + { + "bbox": [ + 106, + 588, + 258, + 596 + ], + "spans": [ + { + "bbox": [ + 106, + 588, + 258, + 596 + ], + "type": "text", + "content": "Ten Training Anomaly Examples Under General Settings" + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 59, + 596, + 181, + 706 + ], + "lines": [ + { + "bbox": [ + 59, + 596, + 181, + 706 + ], + "spans": [ + { + "bbox": [ + 59, + 596, + 181, + 706 + ], + "type": "image", + "image_path": "fdbc1f8c0719b28f4506c3c34a1813611aaaf11b904422766cffb697bb9c6471.jpg" + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 148, + 706, + 460, + 717 + ], + "lines": [ + { + "bbox": [ + 148, + 706, + 460, + 717 + ], + "spans": [ + { + "bbox": [ + 148, + 706, + 460, + 717 + ], + "type": "text", + "content": "Figure 2. Ablation study for SB and DFL under the general settings and hard settings." + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "image_caption" + } + ], + "index": 5 + }, + { + "type": "image", + "bbox": [ + 187, + 596, + 304, + 706 + ], + "blocks": [ + { + "bbox": [ + 187, + 596, + 304, + 706 + ], + "lines": [ + { + "bbox": [ + 187, + 596, + 304, + 706 + ], + "spans": [ + { + "bbox": [ + 187, + 596, + 304, + 706 + ], + "type": "image", + "image_path": "8f20e4f0e230387c2e81af323efd5c71bf9e0280f64c7a75c3fa256690116855.jpg" + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "image_body" + } + ], + "index": 6 + }, + { + "type": "image", + "bbox": [ + 312, + 596, + 428, + 706 + ], + "blocks": [ + { + "bbox": [ + 363, + 588, + 508, + 596 + ], + "lines": [ + { + "bbox": [ + 363, + 588, + 508, + 596 + ], + "spans": [ + { + "bbox": [ + 363, + 588, + 508, + 596 + ], + "type": "text", + "content": "Ten Training Anomaly Examples Under Hard Settings" + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 312, + 596, + 428, + 706 + ], + "lines": [ + { + "bbox": [ + 312, + 596, + 428, + 706 + ], + "spans": [ + { + "bbox": [ + 312, + 596, + 428, + 706 + ], + "type": "image", + "image_path": "db3222376818448821419e66be31f82983244223a722cac6ff6f0daaab0cfddc.jpg" + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "image_body" + } + ], + "index": 8 + }, + { + "type": "image", + "bbox": [ + 435, + 596, + 552, + 706 + ], + "blocks": [ + { + "bbox": [ + 435, + 596, + 552, + 706 + ], + "lines": [ + { + "bbox": [ + 435, + 596, + 552, + 706 + ], + "spans": [ + { + "bbox": [ + 435, + 596, + 552, + 706 + ], + "type": "image", + "image_path": "9d1cbc3ba700da9b6c1c48885b37b1cf35bee15aa1ef07965cd7f3ebe9ea9dd8.jpg" + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "image_body" + } + ], + "index": 9 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 749, + 318, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 749, + 318, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 749, + 318, + 757 + ], + "type": "text", + "content": "20422" + } + ] + } + ], + "index": 11 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 6 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 59, + 70, + 184, + 159 + ], + "blocks": [ + { + "bbox": [ + 59, + 70, + 184, + 159 + ], + "lines": [ + { + "bbox": [ + 59, + 70, + 184, + 159 + ], + "spans": [ + { + "bbox": [ + 59, + 70, + 184, + 159 + ], + "type": "image", + "image_path": "52e48a8462c647629cc8063947bacd4faa66d0ff25e4680459bb4f0bcd62ff5f.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + } + ], + "index": 0 + }, + { + "type": "image", + "bbox": [ + 186, + 70, + 306, + 159 + ], + "blocks": [ + { + "bbox": [ + 186, + 70, + 306, + 159 + ], + "lines": [ + { + "bbox": [ + 186, + 70, + 306, + 159 + ], + "spans": [ + { + "bbox": [ + 186, + 70, + 306, + 159 + ], + "type": "image", + "image_path": "636cde06dae3253814725a80d93b2120b14699832fb797ea88e5c32022f03ffd.jpg" + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 200, + 159, + 410, + 171 + ], + "lines": [ + { + "bbox": [ + 200, + 159, + 410, + 171 + ], + "spans": [ + { + "bbox": [ + 200, + 159, + 410, + 171 + ], + "type": "text", + "content": "Figure 3. Parameter sensitivity analysis for " + }, + { + "bbox": [ + 200, + 159, + 410, + 171 + ], + "type": "inline_equation", + "content": "C,\\epsilon ,\\kappa" + }, + { + "bbox": [ + 200, + 159, + 410, + 171 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 200, + 159, + 410, + 171 + ], + "type": "inline_equation", + "content": "\\lambda" + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_caption" + } + ], + "index": 1 + }, + { + "type": "image", + "bbox": [ + 309, + 70, + 429, + 159 + ], + "blocks": [ + { + "bbox": [ + 309, + 70, + 429, + 159 + ], + "lines": [ + { + "bbox": [ + 309, + 70, + 429, + 159 + ], + "spans": [ + { + "bbox": [ + 309, + 70, + 429, + 159 + ], + "type": "image", + "image_path": "617eecc1e1cbf38ca0cb98682474c27f4fb443f94a8e5f50cd1f9e03986f225c.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + } + ], + "index": 2 + }, + { + "type": "image", + "bbox": [ + 431, + 70, + 552, + 159 + ], + "blocks": [ + { + "bbox": [ + 431, + 70, + 552, + 159 + ], + "lines": [ + { + "bbox": [ + 431, + 70, + 552, + 159 + ], + "spans": [ + { + "bbox": [ + 431, + 70, + 552, + 159 + ], + "type": "image", + "image_path": "dcb2a25d389bcdc8281164de9850e18bd07ed9e494f6909ce9364c3a95b93578.jpg" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_body" + } + ], + "index": 3 + }, + { + "type": "table", + "bbox": [ + 60, + 188, + 291, + 330 + ], + "blocks": [ + { + "bbox": [ + 89, + 175, + 260, + 186 + ], + "lines": [ + { + "bbox": [ + 89, + 175, + 260, + 186 + ], + "spans": [ + { + "bbox": [ + 89, + 175, + 260, + 186 + ], + "type": "text", + "content": "Table 3. An ablation study for " + }, + { + "bbox": [ + 89, + 175, + 260, + 186 + ], + "type": "inline_equation", + "content": "{\\mathrm{M}}_{\\mathrm{n}},{\\mathrm{M}}_{\\mathrm{a}}" + }, + { + "bbox": [ + 89, + 175, + 260, + 186 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 89, + 175, + 260, + 186 + ], + "type": "inline_equation", + "content": "{\\mathrm{M}}_{\\mathrm{r}}" + }, + { + "bbox": [ + 89, + 175, + 260, + 186 + ], + "type": "text", + "content": " ." + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 60, + 188, + 291, + 330 + ], + "lines": [ + { + "bbox": [ + 60, + 188, + 291, + 330 + ], + "spans": [ + { + "bbox": [ + 60, + 188, + 291, + 330 + ], + "type": "table", + "html": "
MnMaMrAITEXELPVMastcam
Ten Training Anomaly Examples Under General Settings
××✓0.928 ± 0.0190.914 ± 0.0210.899 ± 0.036
✓×✓0.939 ± 0.0230.919 ± 0.0110.908 ± 0.017
×✓✓0.963 ± 0.0080.930 ± 0.0130.924 ± 0.019
✓✓✓0.975 ± 0.0070.937 ± 0.0030.934 ± 0.010
Ten Training Anomaly Examples Under Hard Settings
××✓0.746 ± 0.0250.797 ± 0.0260.723 ± 0.017
✓×✓0.758 ± 0.0350.802 ± 0.0240.736 ± 0.028
×✓✓0.781 ± 0.0140.811 ± 0.0180.762 ± 0.021
✓✓✓0.798 ± 0.0050.818 ± 0.0030.778 ± 0.007
", + "image_path": "0804926702655fb01f91c4dd339e34e01f1d6a4f004b6c5f013c385847387dd0.jpg" + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "table_body" + } + ], + "index": 6 + }, + { + "bbox": [ + 55, + 335, + 149, + 347 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 335, + 149, + 347 + ], + "spans": [ + { + "bbox": [ + 55, + 335, + 149, + 347 + ], + "type": "text", + "content": "5.6. Ablation Study" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 54, + 352, + 295, + 567 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 54, + 352, + 295, + 567 + ], + "spans": [ + { + "bbox": [ + 54, + 352, + 295, + 567 + ], + "type": "text", + "content": "The ablation study in Fig. 2 highlights the critical roles of the DPL and DFL components in improving DPDL's openset anomaly detection. We denote the variants that remove only DPL or DFL as 'DPDL w/o DPL' and 'DPDL w/o DFL', respectively. Compared to the full DPDL model, removing neither DPL nor DFL leads to a significant AUC decline, illustrating their critical utility. Specifically, 'DPDL w/o DPL' exhibits the most significant performance drop on industrial anomaly datasets, reflecting DPL's prominent role in learning precise and tight distribution boundaries for normal samples. By transforming normal samples into Gaussian distribution prototype space and pushing abnormal samples away, DPL enhances the recognition ability of anomaly samples. Meanwhile, ablation studies on 'DPDL w/o DFL' further highlight the critical role of the DFL component. By performing discreteness feature learning in hyperspherical space, DFL enhances the generalization ability to out-of-distribution anomalies." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 55, + 567, + 295, + 651 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 567, + 295, + 651 + ], + "spans": [ + { + "bbox": [ + 55, + 567, + 295, + 651 + ], + "type": "text", + "content": "Additionally, the ablation experiments on " + }, + { + "bbox": [ + 55, + 567, + 295, + 651 + ], + "type": "inline_equation", + "content": "\\mathsf{M}_{\\mathrm{n}}" + }, + { + "bbox": [ + 55, + 567, + 295, + 651 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 55, + 567, + 295, + 651 + ], + "type": "inline_equation", + "content": "\\mathsf{M}_{\\mathrm{a}}" + }, + { + "bbox": [ + 55, + 567, + 295, + 651 + ], + "type": "text", + "content": ", and " + }, + { + "bbox": [ + 55, + 567, + 295, + 651 + ], + "type": "inline_equation", + "content": "\\mathsf{M}_{\\mathrm{r}}" + }, + { + "bbox": [ + 55, + 567, + 295, + 651 + ], + "type": "text", + "content": " across three datasets in Tab. 3 reveal their varying contributions. DPDL shows the most significant performance drop when " + }, + { + "bbox": [ + 55, + 567, + 295, + 651 + ], + "type": "inline_equation", + "content": "\\mathsf{M}_{\\mathrm{r}}" + }, + { + "bbox": [ + 55, + 567, + 295, + 651 + ], + "type": "text", + "content": " is removed, which illustrates that " + }, + { + "bbox": [ + 55, + 567, + 295, + 651 + ], + "type": "inline_equation", + "content": "\\mathsf{M}_{\\mathrm{r}}" + }, + { + "bbox": [ + 55, + 567, + 295, + 651 + ], + "type": "text", + "content": " plays the most critical role in detecting anomalies. Furthermore, the removal of " + }, + { + "bbox": [ + 55, + 567, + 295, + 651 + ], + "type": "inline_equation", + "content": "\\mathsf{M}_{\\mathrm{a}}" + }, + { + "bbox": [ + 55, + 567, + 295, + 651 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 55, + 567, + 295, + 651 + ], + "type": "inline_equation", + "content": "\\mathsf{M}_{\\mathrm{n}}" + }, + { + "bbox": [ + 55, + 567, + 295, + 651 + ], + "type": "text", + "content": " lead to a noticeable performance drop in DPDL, which underscores their essential roles." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 55, + 659, + 222, + 673 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 659, + 222, + 673 + ], + "spans": [ + { + "bbox": [ + 55, + 659, + 222, + 673 + ], + "type": "text", + "content": "5.7. Parameter Sensitivity Analysis" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 55, + 677, + 295, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 677, + 295, + 715 + ], + "spans": [ + { + "bbox": [ + 55, + 677, + 295, + 715 + ], + "type": "text", + "content": "Fig. 3 illustrates the results of the four hyperparameters under the general settings across nine datasets. Overall, the performance remains stable within a certain range of hyper" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 313, + 177, + 553, + 332 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 177, + 553, + 332 + ], + "spans": [ + { + "bbox": [ + 313, + 177, + 553, + 332 + ], + "type": "text", + "content": "parameter variations, demonstrating the DPDL's robustness. Prototype quantity " + }, + { + "bbox": [ + 313, + 177, + 553, + 332 + ], + "type": "inline_equation", + "content": "C" + }, + { + "bbox": [ + 313, + 177, + 553, + 332 + ], + "type": "text", + "content": ". We begin by investigating the critical impact of the number of initialized prototypes " + }, + { + "bbox": [ + 313, + 177, + 553, + 332 + ], + "type": "inline_equation", + "content": "C" + }, + { + "bbox": [ + 313, + 177, + 553, + 332 + ], + "type": "text", + "content": " on distributed prototype learning in DPDL. We select " + }, + { + "bbox": [ + 313, + 177, + 553, + 332 + ], + "type": "inline_equation", + "content": "\\{8,16,32,64\\}" + }, + { + "bbox": [ + 313, + 177, + 553, + 332 + ], + "type": "text", + "content": " as the values for the hyperparameters. As " + }, + { + "bbox": [ + 313, + 177, + 553, + 332 + ], + "type": "inline_equation", + "content": "C" + }, + { + "bbox": [ + 313, + 177, + 553, + 332 + ], + "type": "text", + "content": " increases, DPDL's performance improves steadily, but excessively large values of " + }, + { + "bbox": [ + 313, + 177, + 553, + 332 + ], + "type": "inline_equation", + "content": "C" + }, + { + "bbox": [ + 313, + 177, + 553, + 332 + ], + "type": "text", + "content": " hinder the model's effectiveness. This phenomenon is particularly pronounced on AI-TEX, Mastcam, MVTecAD, and Hyper-Kvasir, which contain more categories. One possible explanation is that a prototype space with too few prototypes loses discriminative information, while an excessively large number of prototypes reduces the compactness of the space." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 313, + 333, + 554, + 441 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 333, + 554, + 441 + ], + "spans": [ + { + "bbox": [ + 313, + 333, + 554, + 441 + ], + "type": "text", + "content": "DPL trajectory " + }, + { + "bbox": [ + 313, + 333, + 554, + 441 + ], + "type": "inline_equation", + "content": "\\epsilon" + }, + { + "bbox": [ + 313, + 333, + 554, + 441 + ], + "type": "text", + "content": " in Eqns. (9), (10), (11) and (14). It can be observed that, particularly on the AITEX, Mastcam, and MVTecAD datasets with a larger number of categories, DPDL exhibits a relatively stable performance decline as " + }, + { + "bbox": [ + 313, + 333, + 554, + 441 + ], + "type": "inline_equation", + "content": "\\epsilon" + }, + { + "bbox": [ + 313, + 333, + 554, + 441 + ], + "type": "text", + "content": " increases. As a crucial parameter in SB, " + }, + { + "bbox": [ + 313, + 333, + 554, + 441 + ], + "type": "inline_equation", + "content": "\\epsilon" + }, + { + "bbox": [ + 313, + 333, + 554, + 441 + ], + "type": "text", + "content": " governs the trajectory state. Since smaller values produce straighter trajectories and larger values increase fluctuation, smaller " + }, + { + "bbox": [ + 313, + 333, + 554, + 441 + ], + "type": "inline_equation", + "content": "\\epsilon" + }, + { + "bbox": [ + 313, + 333, + 554, + 441 + ], + "type": "text", + "content": " facilitates sampling more robust abstract prototypes from the relatively dispersed conditional distribution." + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 313, + 441, + 554, + 501 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 441, + 554, + 501 + ], + "spans": [ + { + "bbox": [ + 313, + 441, + 554, + 501 + ], + "type": "text", + "content": "DFL tightness " + }, + { + "bbox": [ + 313, + 441, + 554, + 501 + ], + "type": "inline_equation", + "content": "\\kappa" + }, + { + "bbox": [ + 313, + 441, + 554, + 501 + ], + "type": "text", + "content": ". According to Fig. 3, increasing " + }, + { + "bbox": [ + 313, + 441, + 554, + 501 + ], + "type": "inline_equation", + "content": "\\kappa" + }, + { + "bbox": [ + 313, + 441, + 554, + 501 + ], + "type": "text", + "content": " generally enhances model performance, but values above " + }, + { + "bbox": [ + 313, + 441, + 554, + 501 + ], + "type": "inline_equation", + "content": "\\kappa = 10" + }, + { + "bbox": [ + 313, + 441, + 554, + 501 + ], + "type": "text", + "content": " introduce negative effects in certain scenarios. A possible reason is that excessive sample dispersion makes it more challenging to tighten the normal distribution boundary." + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 313, + 502, + 554, + 585 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 502, + 554, + 585 + ], + "spans": [ + { + "bbox": [ + 313, + 502, + 554, + 585 + ], + "type": "text", + "content": "Loss parameter " + }, + { + "bbox": [ + 313, + 502, + 554, + 585 + ], + "type": "inline_equation", + "content": "\\lambda" + }, + { + "bbox": [ + 313, + 502, + 554, + 585 + ], + "type": "text", + "content": ". We conduct a sensitivity analysis on the loss parameters " + }, + { + "bbox": [ + 313, + 502, + 554, + 585 + ], + "type": "inline_equation", + "content": "\\lambda" + }, + { + "bbox": [ + 313, + 502, + 554, + 585 + ], + "type": "text", + "content": ". It can be observe that setting " + }, + { + "bbox": [ + 313, + 502, + 554, + 585 + ], + "type": "inline_equation", + "content": "\\lambda = 0.01" + }, + { + "bbox": [ + 313, + 502, + 554, + 585 + ], + "type": "text", + "content": " achieves optimal performance on seven larger-scale datasets, while " + }, + { + "bbox": [ + 313, + 502, + 554, + 585 + ], + "type": "inline_equation", + "content": "\\lambda = 1" + }, + { + "bbox": [ + 313, + 502, + 554, + 585 + ], + "type": "text", + "content": " yields the best results on two datasets with limited data. As " + }, + { + "bbox": [ + 313, + 502, + 554, + 585 + ], + "type": "inline_equation", + "content": "\\lambda" + }, + { + "bbox": [ + 313, + 502, + 554, + 585 + ], + "type": "text", + "content": " increases, the performance declines, potentially due to gradient conflicts among the dispersion loss, SB transform loss and the main task loss." + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 313, + 597, + 388, + 609 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 597, + 388, + 609 + ], + "spans": [ + { + "bbox": [ + 313, + 597, + 388, + 609 + ], + "type": "text", + "content": "6. Conclusion" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 313, + 617, + 554, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 617, + 554, + 713 + ], + "spans": [ + { + "bbox": [ + 313, + 617, + 554, + 713 + ], + "type": "text", + "content": "We propose Distribution Prototype Diffusion Learning (DPDL) for OSAD. DPDL leverages schrödinger bridge to map the normal distribution to a prototype space, simultaneously repelling anomalies to facilitate precise anomaly detection. We propose a dispersion feature learning way in hyperspherical space, which benefits the detection of out-of-distribution anomalies. Experimental results illustrate DPDL's robustness in diverse anomaly detection scenarios." + } + ] + } + ], + "index": 17 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "text", + "content": "20423" + } + ] + } + ], + "index": 18 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 7 + }, + { + "para_blocks": [ + { + "bbox": [ + 56, + 72, + 165, + 85 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 72, + 165, + 85 + ], + "spans": [ + { + "bbox": [ + 56, + 72, + 165, + 85 + ], + "type": "text", + "content": "7. Acknowledgments" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 55, + 90, + 296, + 157 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 90, + 296, + 157 + ], + "spans": [ + { + "bbox": [ + 55, + 90, + 296, + 157 + ], + "type": "text", + "content": "This work was supported by the National Natural Science Foundation of China (Grants No. 62476133), the National Natural Science Foundation of China (Nos. U22B2049), the National Natural Science Foundation of China (NSFC) under Grant 62272230, and the Natural Science Foundation of Shandong Province (Grant No. ZR2022LZH003)." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 56, + 179, + 115, + 192 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 179, + 115, + 192 + ], + "spans": [ + { + "bbox": [ + 56, + 179, + 115, + 192 + ], + "type": "text", + "content": "References" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 61, + 199, + 296, + 712 + ], + "type": "list", + "angle": 0, + "index": 12, + "blocks": [ + { + "bbox": [ + 61, + 199, + 296, + 276 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 61, + 199, + 296, + 276 + ], + "spans": [ + { + "bbox": [ + 61, + 199, + 296, + 276 + ], + "type": "text", + "content": "[1] Andra Acsintoae, Andrei Florescu, Mariana-Iuliana Georgescu, Tudor Mare, Paul Sumedrea, Radu Tudor Ionescu, Fahad Shahbaz Khan, and Mubarak Shah. Ubnormal: New benchmark for supervised open-set video anomaly detection. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 20143-20153, 2022. 1, 2" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 61, + 277, + 296, + 331 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 61, + 277, + 296, + 331 + ], + "spans": [ + { + "bbox": [ + 61, + 277, + 296, + 331 + ], + "type": "text", + "content": "[2] Aimira Baitieva, David Hurych, Victor Besnier, and Olivier Bernard. Supervised anomaly detection for complex industrial images. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 17754-17762, 2024. 1" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 61, + 334, + 295, + 388 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 61, + 334, + 295, + 388 + ], + "spans": [ + { + "bbox": [ + 61, + 334, + 295, + 388 + ], + "type": "text", + "content": "[3] Paul Bergmann, Michael Fauser, David Sattlegger, and Carsten Steger. Mvtec ad-a comprehensive real-world dataset for unsupervised anomaly detection. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 9592-9600, 2019. 6" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 62, + 389, + 296, + 456 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 389, + 296, + 456 + ], + "spans": [ + { + "bbox": [ + 62, + 389, + 296, + 456 + ], + "type": "text", + "content": "[4] Hanna Borgli, Vajira Thambawita, Pia H Smedsrud, Steven Hicks, Debesh Jha, Sigrun L Eskeland, Kristin Ranheim Randel, Konstantin Pogorelov, Mathias Lux, Duc Tien Dang Nguyen, et al. Hyperkvasir, a comprehensive multi-class image and video dataset for gastrointestinal endoscopy. Scientific data, 7(1):283, 2020. 6" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 62, + 456, + 296, + 500 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 456, + 296, + 500 + ], + "spans": [ + { + "bbox": [ + 62, + 456, + 296, + 500 + ], + "type": "text", + "content": "[5] Yongxin Chen, Tryphon T Georgiou, and Michele Pavon. On the relation between optimal transport and schrödinger bridges: A stochastic control viewpoint. Journal of Optimization Theory and Applications, 169:671-691, 2016. 2" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 62, + 501, + 296, + 555 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 501, + 296, + 555 + ], + "spans": [ + { + "bbox": [ + 62, + 501, + 296, + 555 + ], + "type": "text", + "content": "[6] Songmin Dai, Yifan Wu, Xiaogiang Li, and Xiangyang Xue. Generating and reweighting dense contrastive patterns for unsupervised anomaly detection. In Proceedings of the AAAI Conference on Artificial Intelligence, pages 1454-1462, 2024. 1" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 62, + 557, + 296, + 601 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 557, + 296, + 601 + ], + "spans": [ + { + "bbox": [ + 62, + 557, + 296, + 601 + ], + "type": "text", + "content": "[7] Valentin De Bortoli, James Thornton, Jeremy Heng, and Arnaud Doucet. Diffusion schrödinger bridge with applications to score-based generative modeling. Advances in Neural Information Processing Systems, 34:17695-17709, 2021. 2" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 62, + 602, + 296, + 656 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 602, + 296, + 656 + ], + "spans": [ + { + "bbox": [ + 62, + 602, + 296, + 656 + ], + "type": "text", + "content": "[8] Sergiu Deitsch, Vincent Christlein, Stephan Berger, Claudia Buerhop-Lutz, Andreas Maier, Florian Gallwitz, and Christian Riess. Automatic classification of defective photovoltaic module cells in electroluminescence images. Solar Energy, 185:455-468, 2019. 6" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 62, + 658, + 296, + 712 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 658, + 296, + 712 + ], + "spans": [ + { + "bbox": [ + 62, + 658, + 296, + 712 + ], + "type": "text", + "content": "[9] Chouro Ding, Guansong Pang, and Chunhua Shen. Catching both gray and black swans: Open-set supervised anomaly detection. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 7388-7398, 2022. 1, 2, 5, 6, 7" + } + ] + } + ], + "index": 11 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 316, + 73, + 554, + 714 + ], + "type": "list", + "angle": 0, + "index": 27, + "blocks": [ + { + "bbox": [ + 316, + 73, + 554, + 116 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 73, + 554, + 116 + ], + "spans": [ + { + "bbox": [ + 316, + 73, + 554, + 116 + ], + "type": "text", + "content": "[10] Nikita Gushchin, Sergei Kholkin, Evgeny Burnaev, and Alexander Korotin. Light and optimal schrödinger bridge matching. In *Forty-first International Conference on Machine Learning*, 2024. 2, 3, 4, 5" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 317, + 118, + 554, + 172 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 118, + 554, + 172 + ], + "spans": [ + { + "bbox": [ + 317, + 118, + 554, + 172 + ], + "type": "text", + "content": "[11] Liren He, Zhengkai Jiang, Jinlong Peng, Liang Liu, Qianggang Du, Xiaobin Hu, Wenbing Zhu, Mingmin Chi, Yabiao Wang, and Chengjie Wang. Learning unified reference representation for unsupervised multi-class anomaly detection. arXiv preprint arXiv:2403.11561, 2024. 1" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 317, + 173, + 554, + 228 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 173, + 554, + 228 + ], + "spans": [ + { + "bbox": [ + 317, + 173, + 554, + 228 + ], + "type": "text", + "content": "[12] Teng Hu, Jiangning Zhang, Ran Yi, Yuzhen Du, Xu Chen, Liang Liu, Yabiao Wang, and Chengjie Wang. Anomalydiffusion: Few-shot anomaly image generation with diffusion model. In Proceedings of the AAAI Conference on Artificial Intelligence, pages 8526-8534, 2024. 1" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 316, + 228, + 554, + 293 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 228, + 554, + 293 + ], + "spans": [ + { + "bbox": [ + 316, + 228, + 554, + 293 + ], + "type": "text", + "content": "[13] Hannah R Kerner, Kiri L Wagstaff, Brian D Bue, Danika F Wellington, Samantha Jacob, Paul Horton, James F Bell, Chiman Kwan, and Heni Ben Amor. Comparison of novelty detection methods for multispectral images in rover-based planetary exploration missions. Data Mining and Knowledge Discovery, 34:1642-1675, 2020. 6" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 316, + 293, + 553, + 336 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 293, + 553, + 336 + ], + "spans": [ + { + "bbox": [ + 316, + 293, + 553, + 336 + ], + "type": "text", + "content": "[14] Beomsu Kim, Gihyun Kwon, Kwanyoung Kim, and Jong Chul Ye. Unpaired image-to-image translation via neural schr\\” odinger bridge. arXiv preprint arXiv:2305.15086, 2023. 2" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 316, + 338, + 553, + 380 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 338, + 553, + 380 + ], + "spans": [ + { + "bbox": [ + 316, + 338, + 553, + 380 + ], + "type": "text", + "content": "[15] Daehyun Kim, Sungyong Baik, and Tae Hyun Kim. Sanflow: Semantic-aware normalizing flow for anomaly detection. Advances in Neural Information Processing Systems, 36:75434-75454, 2023. 1" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 316, + 382, + 553, + 414 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 382, + 553, + 414 + ], + "spans": [ + { + "bbox": [ + 316, + 382, + 553, + 414 + ], + "type": "text", + "content": "[16] Hyunsu Kim, Jongmin Yoon, and Juho Lee. Fast ensembling with diffusion schr\\” odinger bridge. arXiv preprint arXiv:2404.15814, 2024. 2" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 316, + 415, + 553, + 448 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 415, + 553, + 448 + ], + "spans": [ + { + "bbox": [ + 316, + 415, + 553, + 448 + ], + "type": "text", + "content": "[17] Alexander Korotin, Nikita Gushchin, and Evgeny Burnaev. Light schr\\” odinger bridge. arXiv preprint arXiv:2310.01174, 2023. 2, 3, 4, 5" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 316, + 449, + 553, + 481 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 449, + 553, + 481 + ], + "spans": [ + { + "bbox": [ + 316, + 449, + 553, + 481 + ], + "type": "text", + "content": "[18] Christian Léonard. A survey of the schr\\” odinger problem and some of its connections with optimal transport. arXiv preprint arXiv:1308.0215, 2013. 2" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 317, + 482, + 554, + 536 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 482, + 554, + 536 + ], + "spans": [ + { + "bbox": [ + 317, + 482, + 554, + 536 + ], + "type": "text", + "content": "[19] Chun-Liang Li, Kihyuk Sohn, Jinsung Yoon, and Tomas Pfister. Cutpaste: Self-supervised learning for anomaly detection and localization. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 9664-9674, 2021. 1, 6" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 317, + 537, + 554, + 581 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 537, + 554, + 581 + ], + "spans": [ + { + "bbox": [ + 317, + 537, + 554, + 581 + ], + "type": "text", + "content": "[20] Hanxi Li, Jingqi Wu, Hao Chen, Mingwen Wang, and Chunhua Shen. Efficient anomaly detection with budget annotation using semi-supervised residual transformer. arXiv preprint arXiv:2306.03492, 2023. 1" + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 317, + 582, + 554, + 645 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 582, + 554, + 645 + ], + "spans": [ + { + "bbox": [ + 317, + 582, + 554, + 645 + ], + "type": "text", + "content": "[21] Xiaofan Li, Zhizhong Zhang, Xin Tan, Chengwei Chen, Yanyun Qu, Yuan Xie, and Lizhuang Ma. Prompted: Learning prompts with only normal samples for few-shot anomaly detection. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 16838-16848, 2024. 1" + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 316, + 647, + 553, + 690 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 647, + 553, + 690 + ], + "spans": [ + { + "bbox": [ + 316, + 647, + 553, + 690 + ], + "type": "text", + "content": "[22] Jingyi Liao, Xun Xu, Manh Cuong Nguyen, Adam Goodge, and Chuan Sheng Foo. Coft-ad: Contrastive fine-tuning for few-shot anomaly detection. IEEE Transactions on Image Processing, 2024. 1" + } + ] + } + ], + "index": 25 + }, + { + "bbox": [ + 316, + 691, + 553, + 714 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 691, + 553, + 714 + ], + "spans": [ + { + "bbox": [ + 316, + 691, + 553, + 714 + ], + "type": "text", + "content": "[23] Guan-Horng Liu, Tianrong Chen, Oswin So, and Evangelos Theodorou. Deep generalized schrödinger bridge. Advances" + } + ] + } + ], + "index": 26 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 318, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 318, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 318, + 757 + ], + "type": "text", + "content": "20424" + } + ] + } + ], + "index": 28 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 8 + }, + { + "para_blocks": [ + { + "bbox": [ + 56, + 72, + 295, + 712 + ], + "type": "list", + "angle": 0, + "index": 14, + "blocks": [ + { + "bbox": [ + 76, + 72, + 294, + 94 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 76, + 72, + 294, + 94 + ], + "spans": [ + { + "bbox": [ + 76, + 72, + 294, + 94 + ], + "type": "text", + "content": "in Neural Information Processing Systems, 35:9374-9388, 2022. 2" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 56, + 96, + 295, + 139 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 96, + 295, + 139 + ], + "spans": [ + { + "bbox": [ + 56, + 96, + 295, + 139 + ], + "type": "text", + "content": "[24] Guan-Horng Liu, Yaron Lipman, Maximilian Nickel, Brian Karrer, Evangelos A Theodorou, and Ricky TQ Chen. Generalized schr\\” odinger bridge matching. arXiv preprint arXiv:2310.02233, 2023. 2" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 56, + 141, + 294, + 184 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 141, + 294, + 184 + ], + "spans": [ + { + "bbox": [ + 56, + 141, + 294, + 184 + ], + "type": "text", + "content": "[25] Guan-Horng Liu, Arash Vahdat, De-An Huang, Evangelos A Theodorou, Weili Nie, and Anima Anandkumar. " + }, + { + "bbox": [ + 56, + 141, + 294, + 184 + ], + "type": "inline_equation", + "content": "i^2" + }, + { + "bbox": [ + 56, + 141, + 294, + 184 + ], + "type": "text", + "content": "-sb: Image-to-image schr\\” odinger bridge. arXiv preprint arXiv:2302.05872, 2023. 2" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 56, + 186, + 295, + 251 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 186, + 295, + 251 + ], + "spans": [ + { + "bbox": [ + 56, + 186, + 295, + 251 + ], + "type": "text", + "content": "[26] Jiaqi Liu, Kai Wu, Qiang Nie, Ying Chen, Bin-Bin Gao, Yong Liu, Jinbao Wang, Chengjie Wang, and Feng Zheng. Unsupervised continual anomaly detection with contrastively-learned prompt. In Proceedings of the AAAI Conference on Artificial Intelligence, pages 3639–3647, 2024. 1" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 56, + 253, + 295, + 297 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 253, + 295, + 297 + ], + "spans": [ + { + "bbox": [ + 56, + 253, + 295, + 297 + ], + "type": "text", + "content": "[27] Wen Liu, Weixin Luo, Zhengxin Li, Peilin Zhao, Shenghua Gao, et al. Margin learning embedded prediction for video anomaly detection with a few anomalies. In IJCAI, pages 023-3, 2019. 1, 2, 6" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 56, + 298, + 295, + 342 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 298, + 295, + 342 + ], + "spans": [ + { + "bbox": [ + 56, + 298, + 295, + 342 + ], + "type": "text", + "content": "[28] Xinyue Liu, Jianyuan Wang, Biao Leng, and Shuo Zhang. Dual-modeling decouple distillation for unsupervised anomaly detection. arXiv preprint arXiv:2408.03888, 2024.1" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 56, + 343, + 295, + 365 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 343, + 295, + 365 + ], + "spans": [ + { + "bbox": [ + 56, + 343, + 295, + 365 + ], + "type": "text", + "content": "[29] I Loshchilov. Decoupled weight decay regularization. arXiv preprint arXiv:1711.05101, 2017. 6" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 56, + 366, + 295, + 442 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 366, + 295, + 442 + ], + "spans": [ + { + "bbox": [ + 56, + 366, + 295, + 442 + ], + "type": "text", + "content": "[30] Jialin Luo, Yuanzhi Wang, Ziqi Gu, Yide Qiu, Shuaizhen Yao, Fuyun Wang, Chunyan Xu, Wenhua Zhang, Dan Wang, and Zhen Cui. Mmm-rs: A multi-modal, multi-gsd, multiscene remote sensing dataset and benchmark for text-to-image generation. In The Thirty-eight Conference on Neural Information Processing Systems Datasets and Benchmarks Track. 2" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 56, + 445, + 295, + 466 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 445, + 295, + 466 + ], + "spans": [ + { + "bbox": [ + 56, + 445, + 295, + 466 + ], + "type": "text", + "content": "[31] Kanti V Mardia and Peter E Jupp. Directional statistics. John Wiley & Sons, 2009. 5" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 56, + 468, + 295, + 521 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 468, + 295, + 521 + ], + "spans": [ + { + "bbox": [ + 56, + 468, + 295, + 521 + ], + "type": "text", + "content": "[32] Amir Markovitz, Gilad Sharir, Itamar Friedman, Lihi Zelnik-Manor, and Shai Avidan. Graph embedded pose clustering for anomaly detection. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 10539-10547, 2020. 6" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 56, + 523, + 295, + 567 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 523, + 295, + 567 + ], + "spans": [ + { + "bbox": [ + 56, + 523, + 295, + 567 + ], + "type": "text", + "content": "[33] Maxence Noble, Valentin De Bortoli, Arnaud Doucet, and Alain Durmus. Tree-based diffusion schrödinger bridge with applications to wasserstein barycenters. Advances in Neural Information Processing Systems, 36, 2024. 2" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 56, + 568, + 295, + 622 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 568, + 295, + 622 + ], + "spans": [ + { + "bbox": [ + 56, + 568, + 295, + 622 + ], + "type": "text", + "content": "[34] Guansong Pang, Chunhua Shen, and Anton Van Den Hengel. Deep anomaly detection with deviation networks. In Proceedings of the 25th ACM SIGKDD international conference on knowledge discovery & data mining, pages 353-362, 2019. 1" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 56, + 624, + 295, + 667 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 624, + 295, + 667 + ], + "spans": [ + { + "bbox": [ + 56, + 624, + 295, + 667 + ], + "type": "text", + "content": "[35] Guansong Pang, Chouro Ding, Chunhua Shen, and Anton van den Hengel. Explainable deep few-shot anomaly detection with deviation networks. arXiv preprint arXiv:2108.00462, 2021. 2, 5, 6" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 56, + 670, + 295, + 712 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 670, + 295, + 712 + ], + "spans": [ + { + "bbox": [ + 56, + 670, + 295, + 712 + ], + "type": "text", + "content": "[36] T-YLPG Ross and GKHP Dólar. Focal loss for dense object detection. In proceedings of the IEEE conference on computer vision and pattern recognition, pages 2980-2988, 2017. 6" + } + ] + } + ], + "index": 13 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 316, + 72, + 553, + 713 + ], + "type": "list", + "angle": 0, + "index": 28, + "blocks": [ + { + "bbox": [ + 316, + 72, + 553, + 138 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 72, + 553, + 138 + ], + "spans": [ + { + "bbox": [ + 316, + 72, + 553, + 138 + ], + "type": "text", + "content": "[37] Mohammadreza Salehi, Niousha Sadjadi, Soroosh Baselizadeh, Mohammad H Rohban, and Hamid R Rabiee. Multiresolution knowledge distillation for anomaly detection. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 14902-14912, 2021. 6" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 316, + 140, + 553, + 182 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 140, + 553, + 182 + ], + "spans": [ + { + "bbox": [ + 316, + 140, + 553, + 182 + ], + "type": "text", + "content": "[38] Yuyang Shi, Valentin De Bortoli, Andrew Campbell, and Arnaud Doucet. Diffusion schrödinger bridge matching. Advances in Neural Information Processing Systems, 36, 2024. 2" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 316, + 184, + 553, + 227 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 184, + 553, + 227 + ], + "spans": [ + { + "bbox": [ + 316, + 184, + 553, + 227 + ], + "type": "text", + "content": "[39] Javier Silvestre-Blanes, Teresa Albero-Albero, Ignacio Miralles, Rubén Pérez-Llorens, and Jorge Moreno. A public fabric database for defect detection methods and results. *Autex Research Journal*, 19(4):363–374, 2019. 6" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 316, + 228, + 553, + 271 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 228, + 553, + 271 + ], + "spans": [ + { + "bbox": [ + 316, + 228, + 553, + 271 + ], + "type": "text", + "content": "[40] Domen Tabernik, Samo Sela, Jure Skvarc, and Danijel Skocaj. Segmentation-based deep-learning approach for surface-defect detection. Journal of Intelligent Manufacturing, 31 (3):759-776, 2020. 6" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 316, + 272, + 553, + 316 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 272, + 553, + 316 + ], + "spans": [ + { + "bbox": [ + 316, + 272, + 553, + 316 + ], + "type": "text", + "content": "[41] Jihoon Tack, Sangwoo Mo, Jongheon Jeong, and Jinwoo Shin. Csi: Novelty detection via contrastive learning on distributionally shifted instances. Advances in neural information processing systems, 33:11839-11852, 2020. 6" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 316, + 316, + 553, + 369 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 316, + 553, + 369 + ], + "spans": [ + { + "bbox": [ + 316, + 316, + 553, + 369 + ], + "type": "text", + "content": "[42] Yuanzhi Wang, Zhen Cui, and Yong Li. Distribution-consistent modal recovering for incomplete multimodal learning. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 22025-22034, 2023. 2" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 316, + 371, + 553, + 414 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 371, + 553, + 414 + ], + "spans": [ + { + "bbox": [ + 316, + 371, + 553, + 414 + ], + "type": "text", + "content": "[43] Yuanzhi Wang, Yong Li, and Zhen Cui. Incomplete multimodality-diffused emotion recognition. Advances in Neural Information Processing Systems, 36:17117-17128, 2023." + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 316, + 415, + 553, + 458 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 415, + 553, + 458 + ], + "spans": [ + { + "bbox": [ + 316, + 415, + 553, + 458 + ], + "type": "text", + "content": "[44] Yuanzhi Wang, Yong Li, Mengyi Liu, Xiaoya Zhang, Xin Liu, Zhen Cui, and Antoni B Chan. Re-attentional controllable video diffusion editing. arXiv preprint arXiv:2412.11710, 2024." + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 316, + 460, + 553, + 514 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 460, + 553, + 514 + ], + "spans": [ + { + "bbox": [ + 316, + 460, + 553, + 514 + ], + "type": "text", + "content": "[45] Yuanzhi Wang, Yong Li, Xiaoya Zhang, Xin Liu, Anbo Dai, Antoni B Chan, and Zhen Cui. Edit temporal-consistent videos with image diffusion model. ACM Transactions on Multimedia Computing, Communications and Applications, 20(12):1-16, 2024. 2" + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 316, + 514, + 553, + 548 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 514, + 553, + 548 + ], + "spans": [ + { + "bbox": [ + 316, + 514, + 553, + 548 + ], + "type": "text", + "content": "[46] Matthias Wieler and Tobias Hahn. Weakly supervised learning for industrial optical inspection. In DAGM symposium in, page 11, 2007. 6" + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 316, + 548, + 553, + 613 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 548, + 553, + 613 + ], + "spans": [ + { + "bbox": [ + 316, + 548, + 553, + 613 + ], + "type": "text", + "content": "[47] Xincheng Yao, Ruoqi Li, Jing Zhang, Jun Sun, and Chongyang Zhang. Explicit boundary guided semi-push-pull contrastive learning for supervised anomaly detection. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 24490-24499, 2023. 1, 2" + } + ] + } + ], + "index": 25 + }, + { + "bbox": [ + 316, + 614, + 553, + 658 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 614, + 553, + 658 + ], + "spans": [ + { + "bbox": [ + 316, + 614, + 553, + 658 + ], + "type": "text", + "content": "[48] Xincheng Yao, Ruoqi Li, Zefeng Qian, Lu Wang, and Chongyang Zhang. Hierarchical gaussian mixture normalizing flow modeling for unified anomaly detection. arXiv preprint arXiv:2403.13349, 2024. 1" + } + ] + } + ], + "index": 26 + }, + { + "bbox": [ + 316, + 658, + 553, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 658, + 553, + 713 + ], + "spans": [ + { + "bbox": [ + 316, + 658, + 553, + 713 + ], + "type": "text", + "content": "[49] Sangdoo Yun, Dongyoon Han, Seong Joon Oh, Sanghyuk Chun, Junsuk Choe, and Youngjoon Yoo. Cutmix: Regularization strategy to train strong classifiers with localizable features. In Proceedings of the IEEE/CVF international conference on computer vision, pages 6023-6032, 2019. 6" + } + ] + } + ], + "index": 27 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "text", + "content": "20425" + } + ] + } + ], + "index": 29 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 9 + }, + { + "para_blocks": [ + { + "bbox": [ + 56, + 72, + 294, + 206 + ], + "type": "list", + "angle": 0, + "index": 3, + "blocks": [ + { + "bbox": [ + 56, + 72, + 294, + 116 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 72, + 294, + 116 + ], + "spans": [ + { + "bbox": [ + 56, + 72, + 294, + 116 + ], + "type": "text", + "content": "[50] Qihang Zhou, Guansong Pang, Yu Tian, Shibo He, and Jiming Chen. Anomalyclip: Object-agnostic prompt learning for zero-shot anomaly detection. arXiv preprint arXiv:2310.18961, 2023. 1" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 56, + 118, + 294, + 172 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 118, + 294, + 172 + ], + "spans": [ + { + "bbox": [ + 56, + 118, + 294, + 172 + ], + "type": "text", + "content": "[51] Jiawen Zhu, Choubo Ding, Yu Tian, and Guansong Pang. Anomaly heterogeneity learning for open-set supervised anomaly detection. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 17616-17626, 2024. 1, 2, 6, 7" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 56, + 174, + 294, + 206 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 174, + 294, + 206 + ], + "spans": [ + { + "bbox": [ + 56, + 174, + 294, + 206 + ], + "type": "text", + "content": "[52] Yuansheng Zhu, Wentao Bao, and Qi Yu. Towards open set video anomaly detection. In European Conference on Computer Vision, pages 395-412. Springer, 2022. 1, 2" + } + ] + } + ], + "index": 2 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 749, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 749, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 749, + 317, + 757 + ], + "type": "text", + "content": "20426" + } + ] + } + ], + "index": 4 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 10 + } + ], + "_backend": "vlm", + "_version_name": "2.6.4" +} \ No newline at end of file diff --git a/2025/DivPrune_ Diversity-based Visual Token Pruning for Large Multimodal Models/f2b05006-c955-46f7-a5af-8b7909376348_content_list.json b/2025/DivPrune_ Diversity-based Visual Token Pruning for Large Multimodal Models/f2b05006-c955-46f7-a5af-8b7909376348_content_list.json new file mode 100644 index 0000000000000000000000000000000000000000..68e70f33a2a135554524548b6a9aa8ab67cb75b1 --- /dev/null +++ b/2025/DivPrune_ Diversity-based Visual Token Pruning for Large Multimodal Models/f2b05006-c955-46f7-a5af-8b7909376348_content_list.json @@ -0,0 +1,1488 @@ +[ + { + "type": "text", + "text": "DivPrune: Diversity-based Visual Token Pruning for Large Multimodal Models", + "text_level": 1, + "bbox": [ + 98, + 130, + 898, + 152 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Saeed Ranjbar Alvar, Gursimran Singh†, Mohammad Akbari†, Yong Zhang Huawei Technologies Canada Co., Ltd.", + "bbox": [ + 200, + 179, + 795, + 215 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "{saeed.ranjbar.alvar1, gursimran.singh1, mohammad.akbari, yong.zhang3}@huawei.com", + "bbox": [ + 138, + 218, + 854, + 234 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Abstract", + "text_level": 1, + "bbox": [ + 246, + 268, + 326, + 284 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Large Multimodal Models (LMMs) have emerged as powerful models capable of understanding various data modalities, including text, images, and videos. LMMs encode both text and visual data into tokens that are then combined and processed by an integrated Large Language Model (LLM). Including visual tokens substantially increases the total token count, often by thousands. The increased input length for LLM significantly raises the complexity of inference, resulting in high latency in LMMs. To address this issue, token pruning methods, which remove part of the visual tokens, are proposed. The existing token pruning methods either require extensive calibration and fine-tuning or rely on suboptimal importance metrics which results in increased redundancy among the retained tokens. In this paper, we first formulate token pruning as Max-Min Diversity Problem (MMDP) where the goal is to select a subset such that the diversity among the selected tokens is maximized. Then, we solve the MMDP to obtain the selected subset and prune the rest. The proposed method, DivPrune, reduces redundancy and achieves the highest diversity of the selected tokens. By ensuring high diversity, the selected tokens better represent the original tokens, enabling effective performance even at high pruning ratios without requiring fine-tuning. Extensive experiments with various LMMs show that DivPrune achieves state-of-the-art accuracy over 16 image- and video-language datasets. Additionally, DivPrune reduces both the end-to-end latency and GPU memory usage for the tested models. The code is available here $^{\\diamond}$ .", + "bbox": [ + 91, + 301, + 483, + 724 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "1. Introduction", + "text_level": 1, + "bbox": [ + 91, + 750, + 220, + 763 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Following the success of Large Language Models (LLMs) in language understanding [1, 6, 40], Large Multimodal Models (LMMs) [19, 22, 23, 52] have emerged to handle diverse data types like images and video, by leveraging the foundational capabilities of LLMs. Typically, LMMs encode text and visual modalities into tokens, also known", + "bbox": [ + 89, + 773, + 482, + 864 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/8b69041e1c41479ec015cf24321f4a33a50484c7eefe19525afd2c762ed47aa7.jpg", + "image_caption": [ + "Figure 1. Comparison of different visual token pruning methods across various pruning ratios for LLaVA 1.5-7B. The y-axis is the performance averaged on COCO (CIDEr), OKVQA (Acc), POPE (F1), and MMBench (Acc). The x-axis is the TFLOP ratio of the model after token pruning compared to the original model before pruning. The proposed method significantly outperforms all baselines. Note that, unlike other methods, FitPrune uses an additional calibration step to prune tokens." + ], + "image_footnote": [], + "bbox": [ + 544, + 275, + 879, + 462 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "as embeddings. These tokens are then combined and processed by an integrated LLM. The inclusion of visual tokens significantly increases the total number of tokens, often adding thousands to the combined set. Since the running time and memory requirements scale quadratically with input size [7, 8, 15, 38], the addition of visual tokens can substantially raise the running time for LMMs. Hence, many of these models often struggle to meet the demands of low-latency applications, particularly in resource-constrained environments [46].", + "bbox": [ + 509, + 607, + 906, + 758 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Previous research [4, 36, 47] has demonstrated that there is a high degree of redundancy in the visual information processed by LMMs. As a result, visual token pruning has emerged as a promising solution to address the computational complexity challenges faced by LMMs. Specifically, previous research has demonstrated that reducing the number of visual tokens by $50\\%$ [4] to $95\\%$ [36] can significantly enhance the inference speed of LMMs.", + "bbox": [ + 509, + 761, + 908, + 882 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "While promising, token pruning methods have certain", + "bbox": [ + 532, + 885, + 906, + 901 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "CVF", + "bbox": [ + 106, + 2, + 181, + 42 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore.", + "bbox": [ + 236, + 0, + 810, + 46 + ], + "page_idx": 0 + }, + { + "type": "page_footnote", + "text": "Authors have equal contributions.", + "bbox": [ + 112, + 875, + 303, + 887 + ], + "page_idx": 0 + }, + { + "type": "page_footnote", + "text": "$\\diamond$ https://github.com/vbdi/divprune", + "bbox": [ + 112, + 887, + 300, + 900 + ], + "page_idx": 0 + }, + { + "type": "page_number", + "text": "9392", + "bbox": [ + 480, + 944, + 514, + 955 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "shortcomings. For example, the works in [3, 17, 21, 47] require calibration or finetuning for each model which is costly and time-consuming to implement. FastV [4] and PruMerge [36] use attention scores to identify less important tokens for pruning. However, it is shown that using attention scores is not optimal, as some important tokens are overlooked [21]. Additionally, attention-based pruning tends to retain tokens that are similar to each other, leading to redundancy. At high compression ratio, this redundancy prevents the selection of a sufficient number of unique tokens to accurately represent the original tokens. In line with this observation, our findings indicate that pruning a large portion of visual tokens using these methods, without subsequent fine-tuning, results in a significant drop in the performance of LMMs across various tasks (Fig. 1).", + "bbox": [ + 89, + 90, + 480, + 316 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "To address the above-mentioned issues, we formulate token pruning as a Max-Min Diversity Problem (MMDP) [35]. In an MMDP, the objective is to select a subset of elements such that the diversity among them is maximized. We apply this concept to token pruning, which we call DivPrune, aiming to maximize the diversity of the selected tokens by increasing the minimum distance between them. By ensuring high diversity, DivPrune captures a broader range of visual tokens, making it inherently more robust compared to attention-based methods that focus only on token importance scores. Increasing the diversity also helps ensure that the selected tokens better represent the original set of tokens, enabling effective performance even at high pruning ratios without the need for fine-tuning.", + "bbox": [ + 89, + 319, + 480, + 530 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "DivPrune also offers practical advantages that make it a highly useful solution in real-world scenarios. DivPrune is a plug-and-play solution that can be used without requiring offline optimization with a calibration set, or fine-tuning of the model, which are often time-consuming and computationally expensive. DivPrune is applicable to LMMs with any LLM architecture and vision encoder. Additionally, DivPrune is compatible with inference optimization techniques, such as KV caching, resulting in practical speedup in real-world applications. In summary, our major contributions are as follows:", + "bbox": [ + 89, + 532, + 482, + 696 + ], + "page_idx": 1 + }, + { + "type": "list", + "sub_type": "text", + "list_items": [ + "- We introduce DivPrune, a token pruning method based on MMDP that maximizes diversity among visual tokens, effectively reducing redundancy and ensuring a highly representative subset.", + "- DivPrune is a training-free, calibration-data-free, plug-and-play solution that can be seamlessly integrated with off-the-shelf LMMs.", + "- We conduct evaluations using 16 datasets on image-and video-language models with image and video understanding tasks. DivPrune achieves state-of-the-art performance, with noticeable gains under extreme pruning (i.e., ratio $>80\\%$ ).", + "DivPrune reduces GPU memory usage and inference la" + ], + "bbox": [ + 89, + 704, + 482, + 900 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "tency while maintaining comparable accuracy compared to the original model across most datasets.", + "bbox": [ + 526, + 90, + 906, + 121 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "2. Related Works", + "text_level": 1, + "bbox": [ + 513, + 135, + 663, + 150 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "2.1. Large Multimodal Models (LMMs)", + "text_level": 1, + "bbox": [ + 511, + 160, + 820, + 176 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "LMMs handle diverse data types, including text, audio, image, and, video [5, 19, 22, 23, 30, 39, 52]. This work focuses on open-source LMMs that support language and visual inputs. These LMMs can be categorized into two types: image-based and video-based LMMs. The image-based LMMs [22, 23] address image-language understanding tasks, like image captioning, visual question answering, and image reasoning. On the other hand, video-based LMMs are geared towards video understanding [19, 52] tasks, like video captioning, video summarization, and video question answering.", + "bbox": [ + 511, + 183, + 906, + 349 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "2.2. Efficient LMMs", + "text_level": 1, + "bbox": [ + 511, + 359, + 671, + 375 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Several techniques are proposed to improve inference efficiency specifically for LMMs. The first technique is to change the model architecture in LMMs. For example, [33] proposed to replace transformer-based LLMs with Mamba model [12]. [49, 53] retrained LMMs with small scale LLMs to improve their efficiency. [45] used knowledge distillation to train a small LMM. In addition to changing the architecture, it is shown in [37] that skipping some blocks or layers within LMMs can improve the inference speed without damaging the model's performance. Furthermore, efficient decoding techniques such as speculative decoding are proposed to make LMM inference more efficient [11].", + "bbox": [ + 511, + 383, + 908, + 566 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "2.3. Visual Token Pruning", + "text_level": 1, + "bbox": [ + 511, + 574, + 718, + 590 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Visual token pruning methods are proposed to reduce the inference complexity for LMMs. The first group of methods uses attention scores to prune tokens [4, 36]. PruMerge [36] introduces a token pruning method for the vision encoder where the visual tokens are clustered and merged based on their attention sparsity. In addition, FastV [4] prunes tokens within a specific layer of the LLM based on the magnitude of attention scores in an earlier layer. It is shown that pruning tokens based on attention scores are not optimal [13, 21], especially at higher pruning ratios.", + "bbox": [ + 511, + 598, + 906, + 750 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Calibration-based methods offer another line of work, where pruning layers and/or ratios are determined by analyzing the LLM outputs for a calibration dataset [21, 47]. For example, FitPrune [47] calculates a pruning recipe based on the observed attention divergence before and after pruning. VTW [21] argues that visual tokens can be entirely removed after a certain layer within LLM. The layer to remove the visual tokens is chosen using a calibration dataset. These methods rely on calibration datasets and require custom calibration for each LMM, which can be costly and", + "bbox": [ + 511, + 750, + 908, + 900 + ], + "page_idx": 1 + }, + { + "type": "page_number", + "text": "9393", + "bbox": [ + 480, + 944, + 514, + 955 + ], + "page_idx": 1 + }, + { + "type": "image", + "img_path": "images/bfb63d1030fc9da52cd07091a1af6939203095e7c0fd7c2ba1d42c1919a5da5f.jpg", + "image_caption": [ + "Figure 2. An overview of the LMM architecture, with DivPrune applied to visual tokens. The blocks on the right-hand side illustrate the steps of the method." + ], + "image_footnote": [], + "bbox": [ + 94, + 87, + 483, + 287 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "cumbersome for new models.", + "bbox": [ + 89, + 356, + 289, + 369 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Some previous works proposed token pruning with the need for fine-tuning. $\\mathbf{M}^3$ [3] applies model fine-tuning to produce nested visual token representations at multiple granularities, allowing users to select token lengths dynamically during inference. In [17], a projector layer trained using a large-scale dataset is proposed that packs finer detailed information into compact token representations. These methods need significant computational resources for training, limiting their use across various scenarios.", + "bbox": [ + 89, + 371, + 483, + 507 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3. Proposed Method", + "text_level": 1, + "bbox": [ + 89, + 518, + 264, + 537 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "In this section, we briefly discuss how LMMs work. Then, the token pruning problem is defined, followed by a detailed presentation of the proposed method.", + "bbox": [ + 89, + 545, + 483, + 592 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3.1. Large Multimodal Models (LMMs)", + "text_level": 1, + "bbox": [ + 89, + 599, + 398, + 616 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "An LMM typically processes a pair of inputs, denoted as $(T,V)$ , where $T$ is the text input and $V$ is the visual input such as image or video. The text input is mapped to $N$ textual tokens $\\mathbf{E}_{\\mathbf{t}} = \\{t_1,\\dots ,t_N\\}$ using a text encoder. Similarly, the visual input is processed by a corresponding vision encoder. Specifically, it takes visual information $V$ as input and outputs image features, that are further converted to $M$ (generally $M\\gg N$ ) vision tokens $\\mathbf{E}_{\\mathbf{v}} = \\{v_{1},\\ldots ,v_{M}\\}$ using a projector layer (Fig. 2).", + "bbox": [ + 89, + 622, + 483, + 758 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "The textual tokens and visual tokens are then combined to be fed to an LLM to generate the prediction in an autoregressive manner. Specifically, $\\hat{N}$ output tokens $\\mathbf{Y} = \\{y_1,\\dots ,y_{\\hat{N}}\\}$ are generated as follows:", + "bbox": [ + 89, + 758, + 483, + 821 + ], + "page_idx": 2 + }, + { + "type": "equation", + "text": "\n$$\nP \\left(y _ {1}, \\dots , y _ {\\hat {N}} \\mid \\mathbf {E} _ {\\mathbf {t}}, \\mathbf {E} _ {\\mathbf {v}}\\right) = \\prod_ {i = 1} ^ {\\hat {N}} P \\left(y _ {i} \\mid y _ {< i}, \\mathbf {E} _ {\\mathbf {t}}, \\mathbf {E} _ {\\mathbf {v}}\\right), \\tag {1}\n$$\n", + "text_format": "latex", + "bbox": [ + 104, + 830, + 482, + 875 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "where $P(|)$ is the conditional probability obtained at the", + "bbox": [ + 89, + 885, + 483, + 901 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "output of the LLM.", + "bbox": [ + 511, + 90, + 643, + 104 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3.2. Token Pruning", + "text_level": 1, + "bbox": [ + 511, + 113, + 666, + 130 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Reducing the number of input tokens in an integrated LLM within LMMs helps to lower memory usage and inference latency. Since visual tokens tend to have more redundancy, they are generally selected for pruning.", + "bbox": [ + 511, + 137, + 905, + 196 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "In this context, the problem of token pruning can be defined as follows: given a set of visual tokens $\\mathbf{E}_{\\mathbf{v}}$ with $|\\mathbf{E}_{\\mathbf{v}}| = M$ and the subset size $\\tilde{M} (\\tilde{M} < M)$ , the goal is to select a subset, $\\tilde{\\mathbf{E}}_{\\mathbf{v}}$ , while preserving key information necessary for accurate predictions. To mathematically formulate the token pruning problem, we define a mapping function $f$ , which maps the original set of visual tokens, $\\mathbf{E}_{\\mathbf{v}}$ , to a subset, $\\tilde{\\mathbf{E}}_{\\mathbf{v}} = \\{\\tilde{v}_1,\\dots ,\\tilde{v}_{\\tilde{M}}\\}$ , where $|\\tilde{\\mathbf{E}}_{\\mathbf{v}}| = \\tilde{M}$ . The objective is to identify a mapping function $f$ that minimizes the difference in the model's output before and after pruning while ensuring the reduced set still captures the essential information from the original set:", + "bbox": [ + 511, + 198, + 906, + 378 + ], + "page_idx": 2 + }, + { + "type": "equation", + "text": "\n$$\n\\text {F i n d :} \\quad f: \\mathbf {E} _ {v} \\rightarrow \\tilde {\\mathbf {E}} _ {\\mathbf {v}}\n$$\n", + "text_format": "latex", + "bbox": [ + 653, + 385, + 799, + 402 + ], + "page_idx": 2 + }, + { + "type": "equation", + "text": "\n$$\n\\text {O b j e c t i v e :} \\quad \\min _ {f} \\mathcal {L} (\\mathcal {P}, \\tilde {\\mathcal {P}}) \\tag {2}\n$$\n", + "text_format": "latex", + "bbox": [ + 617, + 406, + 905, + 431 + ], + "page_idx": 2 + }, + { + "type": "equation", + "text": "\n$$\n\\text {S u b j e c t} \\quad | \\tilde {\\mathbf {E}} _ {\\mathbf {v}} | = \\tilde {M},\n$$\n", + "text_format": "latex", + "bbox": [ + 635, + 434, + 800, + 453 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "where $\\mathcal{P} = P(y_1, \\ldots, y_{\\hat{N}} \\mid \\mathbf{E}_{\\mathbf{t}}, \\mathbf{E}_{\\mathbf{v}})$ and $\\tilde{\\mathcal{P}} = P(y_1, \\ldots, y_{\\hat{N}} \\mid \\mathbf{E}_{\\mathbf{t}}, f(\\mathbf{E}_{\\mathbf{v}}))$ . Here, $\\mathcal{L}$ represents a loss function that measures the difference in the model's output with and without pruning, and $\\tilde{M}$ indicates the number of retained tokens. Next, we propose a novel diversity-based solution for the introduced token pruning problem.", + "bbox": [ + 511, + 460, + 905, + 553 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3.3. DivPrune: Method Overview", + "text_level": 1, + "bbox": [ + 511, + 561, + 774, + 575 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "We proposed a diversity-based token pruning method by reformulating the problem in (2) to select a subset of $\\tilde{M}$ elements that maximizes the diversity, thereby reducing redundancy. Specifically, we define token pruning as Max-Min Diversity Problem (MMDP) [32] where the goal is to find the set $\\tilde{\\mathbf{E}}_{\\mathbf{v}}$ among all possible sets with $\\tilde{M}$ samples in $\\mathbf{E}_{\\mathbf{v}}$ that has the maximum minimum distance between its elements. So, MMDP is defined as:", + "bbox": [ + 511, + 584, + 905, + 705 + ], + "page_idx": 2 + }, + { + "type": "equation", + "text": "\n$$\n\\operatorname {F i n d} \\tilde {\\mathbf {E}} _ {\\mathbf {v}} = \\arg \\max \\left[ \\min _ {\\gamma , \\omega \\in S} (d (\\gamma , \\omega)): \\forall S \\subset \\mathbf {E} _ {\\mathbf {v}} \\right], \\tag {3}\n$$\n", + "text_format": "latex", + "bbox": [ + 529, + 710, + 905, + 744 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "where $S$ is an arbitrary set in $\\mathbf{E}_{\\mathbf{v}}$ with $\\tilde{M}$ elements and $(\\gamma, \\omega)$ are arbitrary elements in $S$ . The distance is measured by $d(.,.)$ which is defined using the cosine distance as follows:", + "bbox": [ + 511, + 752, + 906, + 810 + ], + "page_idx": 2 + }, + { + "type": "equation", + "text": "\n$$\nd (\\gamma , \\omega) = 1 - \\frac {\\gamma \\cdot \\omega}{\\| \\gamma \\| \\| \\omega \\|}. \\tag {4}\n$$\n", + "text_format": "latex", + "bbox": [ + 625, + 808, + 903, + 838 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "A solution for the MMDP problem in (3) is a subset of $\\mathbf{E}_{\\mathbf{v}}$ that maximizes diversity by minimizing redundancy between elements. In the literature, several solutions including exact and heuristic methods are proposed to solve the", + "bbox": [ + 511, + 840, + 906, + 901 + ], + "page_idx": 2 + }, + { + "type": "page_number", + "text": "9394", + "bbox": [ + 480, + 944, + 514, + 955 + ], + "page_idx": 2 + }, + { + "type": "code", + "sub_type": "algorithm", + "code_caption": [ + "Algorithm 1: Proposed Token Pruning Method" + ], + "code_body": "1 $\\tilde{M}$ : subset size; $\\mathbf{E}_{\\mathbf{v}}$ : visual tokens; $\\tilde{\\mathbf{E}}_{\\mathbf{v}}$ : selected subset \n2 Initialize $\\tilde{\\mathbf{E}}_{\\mathbf{v}} = []$ and $\\mathbf{R} = \\mathbf{E}_{\\mathbf{v}}$ \n3 // First stage: add the first token \n4 $D = []$ initialize the distance array \n5 for i in R do \n6 $d_{min} = +inf$ \n7 for j in R do \n8 If $(i \\neq j \\& d(i, j) \\leq d_{min})$ then $d_{min} = d(i, j)$ \n9 Add $d_{min}$ to $D$ \n10 $k = \\mathbf{R}[\\arg \\max (\\mathrm{D})]$ \n11 move k from R to $\\tilde{\\mathbf{E}}_{\\mathbf{v}}$ \n12 // Second stage: iteratively add the subsequent tokens \n13 while $|\\tilde{\\mathbf{E}}_{\\mathbf{v}}| < \\tilde{M}$ do \n14 $D = []$ initialize the distance array \n15 for i in R do \n16 $d_{min} = +inf$ \n17 for j in $\\tilde{\\mathbf{E}}_{\\mathbf{v}}$ do \n18 If $d(i, j) \\leq d_{min}$ then $d_{min} = d(i, j)$ \n19 Add $d_{min}$ to $D$ \n20 $k = \\mathbf{R}[\\arg \\max (\\mathrm{D})]$ \n21 move k from R to $\\tilde{\\mathbf{E}}_{\\mathbf{v}}$ \n22 Return $\\tilde{\\mathbf{E}}_{\\mathbf{v}}$", + "bbox": [ + 94, + 113, + 442, + 450 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "MMDP problem [29, 35]. Since the number of tokens is generally limited (e.g., 576 in LLaVA 1.5 [22]) and the solvers are not generally designed for GPU acceleration, we obtain exact solution for the problem. Notably, the overhead of the selection process using GPU is negligible compared to the computations within the LLM. Detailed steps of the proposed method is summarized in Algorithm 1. Once the selected tokens are identified, the remaining visual tokens are discarded. The selected tokens along with the textual tokens are passed to the LLM.", + "bbox": [ + 89, + 488, + 483, + 638 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "As shown in Algorithm 1, the proposed method has two stages after the initialization. The selected subset, $\\tilde{\\mathbf{E}}_{\\mathbf{v}}$ , is initialized as empty, and the candidate list $\\mathbf{R}$ is initialized with all the visual tokens. In the first stage, the first token of the selected subset is chosen based on the pairwise distance between the tokens of the candidate list. Then, the chosen token is moved from the candidate list to the selected list. In the second stage, similar to the first stage, the pairwise distance of the tokens in $\\tilde{\\mathbf{E}}_{\\mathbf{v}}$ and the tokens in $\\mathbf{R}$ is used to add samples to $\\tilde{\\mathbf{E}}_{\\mathbf{v}}$ iteratively. Finally, once the number of tokens in $\\tilde{\\mathbf{E}}_{\\mathbf{v}}$ reaches the specified subset size, the selection procedure is terminated and the $\\tilde{\\mathbf{E}}_{\\mathbf{v}}$ is returned. To avoid repeated distance calculations over iterations a distance matrix is initially calculated by one matrix multiplication.", + "bbox": [ + 89, + 641, + 483, + 853 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "The proposed method can also be applied to the features (i.e., hidden states) in the intermediate layers of the LLM. In this case, our method is not applied to the visual tokens, but", + "bbox": [ + 89, + 854, + 483, + 900 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "to the features corresponding to the visual tokens obtained from a decoder layer to select a subset before feeding them to the subsequent layers. In either case, our method obtains the highest diversity for the selected elements. Ablation studies are provided in the next section to analyze the effect of pruning different elements at different layers.", + "bbox": [ + 511, + 90, + 905, + 181 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "4. Experiments", + "text_level": 1, + "bbox": [ + 511, + 194, + 645, + 212 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "In this section, we present a comprehensive analysis comparing the performance of our method and previous works across various settings, tasks, and datasets. Insights into the proposed method are also provided through illustrative examples. Moreover, the efficiency of DivPrune along with ablation study are provided.", + "bbox": [ + 511, + 220, + 905, + 311 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "4.1. Experimental Settings", + "text_level": 1, + "bbox": [ + 511, + 319, + 720, + 335 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Baselines and Models: We consider five baselines, namely, FastV [4], PruMerge [36], VTW [21], FitPrune [47] and $\\mathbf{M}^3$ [3]. Among these, we consider FastV, PruMerge, and VTW as our main competitors as they are plug-and-play and do not rely on any further costly finetuning or calibration process. However, for the sake of completeness, we also report performance comparison with respect to one finetuning-based $(\\mathbf{M}^3)$ and one calibration-based (FitPrune) methods. Note that, VTW, by default, requires calibration to determine the best layer for a given task. However, doing that does not allow us to set a specific TFLOP ratio, complicating the comparison. Hence, whenever required we disable the calibration of VTW to select the layer that matches the FLOP requirement of a particular experiment.", + "bbox": [ + 511, + 342, + 906, + 554 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "We test DivPrune and the baselines with popular LMMs namely LLaVA 1.5-7B [22]1, LLaVA 1.5-13B [22]2, LLaVA 1.6-7B3 (also known as LLaVA-NeXT [23]), and LLaVA-NeXT-Video-7B [52]4 to demonstrate the generality of DivPrune. For each tested model and task, we report only the relevant subset of baseline that is applicable to that specific model and task, alongside our results.", + "bbox": [ + 511, + 554, + 905, + 659 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "All the tested LMMs used CLIP vision encoder [34]. LLaVA 1.5 model uses 576 visual tokens to represent images. LLaVA 1.6 converts each image into a varying number of patches, resulting in 3-5 times more visual tokens compared to LLaVA 1.5. LLaVA-NeXT-Video uses 144 tokens to process each frame. For all the experiments with LLaVA-NeXT-Video we used a total of 8 frames resulting in 1152 tokens for the processed frames.", + "bbox": [ + 511, + 660, + 905, + 780 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Datasets, Tasks, and Metrics: We selected a comprehensive set of common tasks and datasets aimed at multimodal reasoning and understanding. Specifically, we chose 11 image-language and 5 video-language datasets.", + "bbox": [ + 511, + 781, + 905, + 842 + ], + "page_idx": 3 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "1https://huggingface.co/liuhaotian/flava-v1.5-7B", + "$^{2}$ https://huggingface.co/luhaotian/llava-v1.5-13b", + "3https://huggingface.co/liuhaotian/llava-v1.6-vicuna-7b", + "4https://huggingface.co/lmms-lab/LLaVA-NeXT-Video-7B-DPO" + ], + "bbox": [ + 529, + 849, + 874, + 900 + ], + "page_idx": 3 + }, + { + "type": "page_number", + "text": "9395", + "bbox": [ + 480, + 944, + 514, + 955 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "These datasets encompass a wide range of tasks, including captioning, multiple-choice Question Answering (QA), and open-ended QA based on text and image/video inputs. Consistent with prior works, CIDEr score [42] is used for evaluating captioning tasks, and Exact Match (EM), Accuracy (Acc), F1, Perception Score (P-score) [9] and GPT-assisted [10] score are used for QA tasks. Furthermore, Wu-Palmer similarity (WUPS) score [43] and GPT-assisted score [10] is used for open-ended QA. For all task performance metrics used in this paper, higher values indicate better performance. For the reported time and memory, lower values indicate better results. Further details regarding the datasets, tasks, and metrics are provided in the supplementary material.", + "bbox": [ + 89, + 90, + 483, + 301 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Following the earlier works in [4, 21, 47], we report the computational requirement, measured in TFLOPs, for DivPrune and the baselines. Various configurations including different pruning ratios at different layers are examined to obtain different working TFLOPs for our method and the baselines. The reported TFLOP ratio is the TFLOP of the model with pruned tokens relative to the original model's TFLOP with no pruning. This ratio is estimated as [4]:", + "bbox": [ + 89, + 301, + 483, + 424 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\frac {K \\times \\left(4 \\mu d ^ {2} - 2 \\mu^ {2} d + 2 \\mu d m\\right) + (T - K) \\times \\left(4 \\tilde {\\mu} d ^ {2} - 2 \\tilde {\\mu} ^ {2} d + 2 \\tilde {\\mu} d m\\right)}{T \\times \\left(4 \\mu d ^ {2} - 2 \\mu^ {2} d + 2 \\mu d m\\right)}, \\tag {5}\n$$\n", + "text_format": "latex", + "bbox": [ + 106, + 435, + 480, + 460 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "where $T$ is the total transformer-based decoder layers. $\\mu = N + M$ is the total sequence length before pruning, $\\tilde{\\mu} = N + \\tilde{M}$ is the sequence length after pruning, $d$ is the hidden state size of the layer, and $m$ is the intermediate size of feed-forward network module. Depending on the TFLOP ratio requirement set by a particular experiment, we adjust the pruning hyperparameters of all baselines to match that requirement. However, some baselines do not support fine-grained adjustments like our approach does. In these cases, we choose the smallest available TFLOP ratio that exceeds the requirement set by an experiment, which might give these baselines a slight advantage over our method.", + "bbox": [ + 89, + 470, + 483, + 652 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "We used $8 \\times$ V100 GPUs with 32GB VRAM for all the experiments in this paper. Additionally, we used the Immsevalts package [51] for running these benchmarks for all the baselines and models. All results are obtained with a batch size of 1. For the metrics that require ChatGPT API access, the model is set to \"gpt-4o-mini\".", + "bbox": [ + 89, + 652, + 483, + 744 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "4.2. Insights", + "text_level": 1, + "bbox": [ + 89, + 755, + 189, + 771 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "We provide visualizations comparing DivPrune with importance-based token pruning methods using LLaVA 1.5-7B and the SeedBench dataset [16]. Detailed analysis across different models and datasets is provided in the following subsections.", + "bbox": [ + 89, + 777, + 483, + 853 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "The visual tokens in LLaVa 1.5 model are 4096-dimensional vectors. The t-SNE method [41] is utilized to project the visual tokens in $\\mathbf{E}_{\\mathbf{v}}$ from a high dimensional", + "bbox": [ + 89, + 854, + 483, + 900 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "to a 2D space. The corresponding visualization for a sample input data is shown in Fig. 3-(a) using light Pruple points. Then, DivPrune is applied to select $10\\%$ of the visual tokens (i.e., pruning $90\\%$ ). Additionally, FastV, as an importance-based token pruning method, which utilizes attention scores, is employed to prune with the same ratio. The selected subsets using DivPrune and FastV are shown with different markers in Fig. 3-(a). More examples are provided in the supplementary materials.", + "bbox": [ + 511, + 90, + 903, + 226 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "As the example in Fig. 3-(a) shows, the proposed method selects points from all the clusters that appeared in the projected space whereas FastV does not choose any samples from the upper cluster. So, our method achieves a better representation of the original points by including samples from all clusters. In addition, the FastV method selects many tokens that are very close to each other which increases redundancy among the selected set. On the other hand, our method reduces redundancy by pruning the closely similar tokens.", + "bbox": [ + 511, + 227, + 905, + 377 + ], + "page_idx": 4 + }, + { + "type": "image", + "img_path": "images/bcd637e8a9777724d51d6a026867414a49f7df61248c1d29c52d870ab11a666c.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 550, + 402, + 890, + 614 + ], + "page_idx": 4 + }, + { + "type": "image", + "img_path": "images/e08094393d41b060f15c77ff4494e06b373fc39f3673c8a7cb3f436e59bcceeb.jpg", + "image_caption": [ + "Figure 3. (a) t-SNE visualization of visual tokens for the original model, our method, and FastV. (b) Histogram of the Max-Min distance between the selected tokens over the SeedBench dataset." + ], + "image_footnote": [], + "bbox": [ + 527, + 618, + 890, + 845 + ], + "page_idx": 4 + }, + { + "type": "page_number", + "text": "9396", + "bbox": [ + 480, + 944, + 514, + 955 + ], + "page_idx": 4 + }, + { + "type": "table", + "img_path": "images/8b36e11670e6102116b47e27d68b20495223ad97cf7e116b9460e536bcc344c4.jpg", + "table_caption": [], + "table_footnote": [], + "table_body": "
MethodTFLOP (ratio %)COCO CIDErFlickr CIDErGQA EMMMB AccMME P-scoreMMMU AccNocaps CIDErOKVQA EMPOPE F1SQA EMSEEDB Acc
LLaVA 1.5.7BOriginal3.228 (100.00)1.100.7561.9664.09150636.441.0653.3985.8469.4166.17
VTW [21]0.603 (18.46)0.050.0338.9421.3168132.600.0318.6425.3565.2936.13
FastV [4]0.514 (15.69)0.060.0338.7320.6269632.000.0418.3232.8465.1535.69
Ours0.512 (15.63)0.960.6256.8559.19132835.890.9246.9886.0268.2759.47
PruMerge [36]Variable0.770.5051.3054.47125935.110.7341.7466.8968.9153.26
Ours*Variable0.910.5655.2558.16133035.440.8744.3883.0667.8757.88
FitPrune△ [47]0.513 (15.65)0.900.5652.3957.65119736.000.8642.5360.8968.0254.84
M3* [3]0.512 (15.63)1.000.6760.8165.81139131.800.9555.1286.3364.6564.93
PruMerge-LoRA*Variable0.960.6355.9659.88133434.890.9047.9977.1368.3257.93
LLaVA 1.5.13BOriginal6.281 (100.00)1.160.8063.3368.64152235.671.0958.2885.9972.8866.82
VTW [21]1.030 (16.16)0.080.0539.7121.9162232.100.0522.490.4066.2438.59
FastV [4]1.003 (15.73)0.380.1844.9837.8094235.110.3332.1430.0269.9644.95
Ours1.002 (15.71)1.000.6657.2963.40140734.890.9553.2983.4372.3462.04
PruMerge△ [36]Variable0.800.5352.0158.93125636.560.7749.1564.3672.5356.10
Ours*Variable0.940.5956.0961.77134434.890.9150.8679.6071.3460.00
LLaVA 1.6.7BOriginal11.849 (100.00)1.000.6864.2867.01152036.440.8844.2086.3870.1570.16
VTW [21]1.318 (11.23)0.060.0338.6219.7660631.300.038.667.1365.7437.48
FastV [4]1.327 (11.30)0.060.0338.7920.3661932.560.048.807.7865.4937.62
Ours1.266 (10.79)0.890.6158.6963.49136237.110.7641.9282.9768.5764.11
M3* [3]1.266 (10.79)1.010.6762.9769.16149035.000.8557.4987.4469.5168.49
", + "bbox": [ + 91, + 88, + 903, + 383 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Table 1. Comparison results of our method and different baselines on image-language understanding datasets. $\\bullet$ : Finetuning is used, $\\triangle$ Calibration dataset is used. Ours*: Our method matching the PruMerge selection ratio.", + "bbox": [ + 89, + 388, + 906, + 417 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "In addition, the max-min distance (Eq. 3) for the selected subset of tokens is computed using 1000 randomly data samples from the SeedBench dataset and the histogram of the computed values is shown in Fig. 3-(b). As the plot indicates, the proposed method selects a subset where samples have a higher minimum pair-wise distance compared to the FastV method. Hence, our method achieves higher diversity among the selected tokens that have less redundancy compared to the ones chosen using FastV. We analyze the effect of the reduced diversity on task performance in the following sections.", + "bbox": [ + 88, + 436, + 480, + 602 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "4.3. Image-Language Understanding", + "text_level": 1, + "bbox": [ + 89, + 617, + 375, + 633 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "In this section, we compare DivPrune against baselines across various image-language understanding tasks, including open- and closed-ended QA, visual reasoning, and image captioning. Specifically, ScienceQA-IMG (SQA) [25], POPE [18], MME [9], MMB [24], GQA [14], MMMU [50], Flicker30k [31], SeedBench (SEEDB) [16], Nocaps [2], OKVQA [28], and COCO-2017 [20] are used.", + "bbox": [ + 89, + 642, + 482, + 747 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "In the first experiment, summarized in Tab. 1, we analyze an extreme compression scenario for three image-based LMMs by fixing the TFLOP ratio at approximately $15\\%$ , wherever the baseline allows configuration to a fixed TFLOP ratio. Since PruMerge does not allow fixing the TFLOP ratio, we configure our approach (Ours*) to match the variable pruning corresponding to PruMerge for a fair comparison. In the top section of the table, we compare the results of various baselines for LLaVa 1.5-7B. Specifically, the baselines supporting LLaVA 1.5 are grouped into", + "bbox": [ + 89, + 750, + 482, + 901 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "three categories: plug-and-play methods, those with a variable TFLOP ratio, and those requiring a calibration dataset or involving fine-tuning the LMMs. Among the plug-and-play methods, which are the focus of this work, our approach significantly outperforms both the VTW and FastV baselines across all datasets. This result holds despite using lower TFLOPs, clearly demonstrating the advantage of our method in this scenario. For instance, when DivPrune is used, the performance of LLaVA 1.5-7b decreases by $5.1\\%$ on the GQA dataset and $4.9\\%$ on the MMB dataset. In contrast, the VTW and FastV methods result in performance drops of at least $23.0\\%$ and $42.8\\%$ on these datasets, respectively. The performance gap between DivPrune and the baseline methods is even more pronounced in image captioning tasks. For example, the CIDEr score on the COCO dataset drops by approximately $95\\%$ with VTW and FastV, but only by $12.7\\%$ with DivPrune. Additionally, DivPrune, compared to the original model, shows less than a $2\\%$ performance drop on the MMMU and SQA datasets and slightly enhances the original model's performance on the POPE dataset while reducing the TLOP ratio by $84.4\\%$ . It is shown that removing redundant tokens in some datasets can improve the original model's performance [4].", + "bbox": [ + 511, + 436, + 906, + 785 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Next, in the variable scenario, the pruning ratio is determined dynamically. To ensure a fair comparison, we matched the pruning ratio with that of the PruMerge baseline, assuming the average sequence length for calculating the average TFLOPs across each dataset. As indicated by the results, our approach consistently outperforms PruMerge across all benchmarks, except one. Further, for", + "bbox": [ + 511, + 795, + 908, + 902 + ], + "page_idx": 5 + }, + { + "type": "page_number", + "text": "9397", + "bbox": [ + 482, + 944, + 514, + 955 + ], + "page_idx": 5 + }, + { + "type": "table", + "img_path": "images/a170543e01d67452d39d90be08b457183ead297ca55710cd571a50aa94bad980.jpg", + "table_caption": [], + "table_footnote": [], + "table_body": "
TFLOPs (ratio %)ActivityNet Score/AccSeedBench AccVChatGPT ScoreNextQA WUPSEgoSch. AccMax GPU mem (GB)Prefill Time (sec)E2E Latency (sec)
Original6.539 (100)2.67 / 48.1038.72.1626.0541.814.060.3304.37
VTW [21]1.124 (16.97)1.61 / 26.8429.391.1918.6625.4213.630.1503.43
FastV [4]0.943 (14.20)1.95 / 33.9132.981.4422.5129.1413.570.1503.63
Ours0.937 (14.10)2.56 / 45.9037.001.9224.4839.7613.510.1613.39
", + "bbox": [ + 133, + 88, + 857, + 175 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Table 2. Comparison results of our method and baselines on LLaVA-NeXT-Video-7B across video-language understanding datasets.", + "bbox": [ + 107, + 179, + 887, + 193 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "the baseline with calibration, we observe that our approach outperforms the FitPrune approach on nearly all datasets by up to $25.1\\%$ , despite not using any calibration dataset. Finally, compared to baselines involving fine-tuning, our method achieves comparable or superior performance without requiring any fine-tuning.", + "bbox": [ + 89, + 214, + 480, + 305 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "The above experiment is repeated with LLaVa 1.5-13B model and the results are shown in the middle part of Tab. 1. The baselines that support this model are FastV, VTW, and PruMerge. As shown in the table, DivPrune outperforms the corresponding baselines in both plug-and-play and variable scenarios almost on all the tested datasets. For example, on the POPE dataset, DivPrune outperforms VTW, FastV, and PruMerge with F1 score improvements of $83\\%$ , $53.4\\%$ , and $15.2\\%$ , respectively. Additionally, on the MMB dataset, DivPrune achieves higher accuracy rates of $41.5\\%$ , $25.6\\%$ , and $2.8\\%$ compared to VTW, FastV, and PruMerge, respectively. This demonstrates that DivPrune generalizes effectively across models with varying numbers of parameters.", + "bbox": [ + 88, + 306, + 482, + 503 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "In the bottom part of Tab. 1, the results corresponding to LLaVA 1.6-7B model are shown. We used the same pruning ratio as for LLava 1.5. However, the lower TFLOP ratio is due to the large number of visual tokens in LLaVA 1.6. The results indicate that the performance of the model drops significantly when baseline pruning methods are applied. For example, the F1 score on the POPE dataset drops by $79\\%$ with the baselines as compared to the original model, whereas the drop with DivPrune is only $3.4\\%$ . DivPrune also maintains competitive performance compared to the original model across various datasets. Specifically, DivPrune shows only $3.5\\%$ , $2.3\\%$ , $3.4\\%$ , $1.6\\%$ drop in accuracy compared to the original model on the MMB, OKVQA, POPE, and SQA datasets, respectively, while reducing the TFLOP by $89\\%$ . The results also demonstrate that pruning visual tokens with DivPrune enhances the original model's performance on the MMMU task. These results show that DivPrune generalizes across different models. Qualitative examples as well as results with additional datasets are provided in the supplementary materials.", + "bbox": [ + 91, + 506, + 482, + 808 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Furthermore, we show the comparison of different baselines and our method across various TFLOP ratios. We plot the results in Fig. 1 where the y-axis represents average performance on four datasets, namely, COCO (CIDEr), OKVQA (Acc), POPE (F1), and MMBench (Acc). The range of the performance metric for all datasets is between", + "bbox": [ + 89, + 810, + 482, + 901 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "0 and 1, except for the CIDEr metric, which has a maximum reported value of 1.10. On the x-axis, we only show the high compression scenario (TFLOP ratio $\\leq 45\\%$ ). As shown in the figure, our method significantly outperforms all the baselines, particularly in high compression scenarios (TFLOP $\\leq 25\\%$ ). Further, we notice a steep drop in performance of all baselines as the TFLOP ratio $\\rightarrow 10$ , while our method falls more gracefully. This results in an increasing performance gap between our approach and the baselines at extreme compression levels. For higher TFLOP ratios almost all converge toward the original performance, with FitPrune slightly outperforming our approach by an insignificant margin. It is important to note that, unlike our method, FitPrune relies on a calibration dataset to prune tokens.", + "bbox": [ + 511, + 214, + 906, + 426 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "4.4. Video-Language Understanding", + "text_level": 1, + "bbox": [ + 511, + 436, + 795, + 454 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "In this section, LLaVA-NeXT-Video-7B [23], a video-based LMM is used to analyze the performance of the proposed method on various video-language understanding tasks. Specifically, we evaluate DivPrune using five datasets, namely, ActivityNet [48], SeedBench [16], VideoChatGPT (temporal) [26], NextQA [44], and EgoSchema [27]. FastV and VTW methods are chosen as the baselines. We tested DivPrune using the same pruning ratio as in the image understanding experiments. However, due to the higher number of visual tokens in the LLaVA-NeXT-Video model, this pruning ratio results in lower TFLOPs ratio. For the baselines, we match their TFLOPs with ours by selecting the smallest available TFLOP ratio that exceeds the TFLOPs of our method. The results for the original model, DivPrune, and the baselines are given in Tab. 2. As shown in the table, DivPrune outperforms both FastV and VTW by a significant margin. Specifically, DivPrune achieves upto $12\\%$ higher accuracy than FastV and upto $19\\%$ better than VTW on Video QA datasets including ActivityNet, SeedBench, and EgoSchema. DivPrune also outperforms both baselines on open-ended QA such as VideoChatGPT and NextQA by achieving higher GPT-assisted and WUPS scores.", + "bbox": [ + 511, + 460, + 906, + 792 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Furthermore, our method achieves performance that is highly competitive compared to the original model without pruning despite using only $14.1\\%$ of the original model's TFLOPs. This demonstrates the robustness of DivPrune, as it effectively generalizes to video LMMs. Notably, the performance gap between DivPrune and the original model without pruning narrows as the number of visual tokens in", + "bbox": [ + 511, + 795, + 908, + 902 + ], + "page_idx": 6 + }, + { + "type": "page_number", + "text": "9398", + "bbox": [ + 480, + 945, + 514, + 955 + ], + "page_idx": 6 + }, + { + "type": "table", + "img_path": "images/dba4db81d6d9d4394eeba6efee6de50488f16f29dd8c554a2c08776f263094b4.jpg", + "table_caption": [], + "table_footnote": [], + "table_body": "
TFLOP (ratio %)MMB AccMMMU AccPOPE F1SQA EMAvg
Layer 0 (Ours)19.6159.1935.8986.0268.2762.34
Layer 119.6559.0234.8980.6767.1860.44
Layer 219.7054.9034.2269.2769.5656.99
Layer 319.8023.9732.6731.8265.9438.60
", + "bbox": [ + 91, + 88, + 480, + 171 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "creases, indicating that DivPrune is more effective for the models with larger visual contexts.", + "bbox": [ + 89, + 212, + 482, + 242 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "4.5. Efficiency Analysis", + "text_level": 1, + "bbox": [ + 89, + 253, + 272, + 268 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "In this section, we analyze the efficiency of the proposed method using memory usage (i.e., max allocated memory), prefetch time, and end-to-end latency (E2E). For this experiment, VideoChatGPT dataset with 499 samples is used to obtain the average time and memory usage for LLaVA-NeXT-Video-7B model. The results are summarized on the right side of Tab. 2. The obtained results are compared against the original model, as well as the FastV and VTW baselines. As shown in the table, our approach requires approximately 400MB less memory than the original model, with memory usage comparable to the baselines. In terms of prefetch and E2E time, our approach is about $55\\%$ and $22\\%$ faster, respectively, compared to the original model. When compared to the baselines, our prefetch time is approximately $6 - 7\\%$ longer, while the E2E time is $1 - 7\\%$ shorter. The slight increase in prefetch time for our method compared to the baselines is due to the distance calculations (See Section 3.3), which are performed only once during the prefetch stage. In contrast, for baselines, the corresponding calculations for token pruning need to be done at each decoding step, resulting in longer E2E time.", + "bbox": [ + 89, + 276, + 482, + 594 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "4.6. Ablation Study", + "text_level": 1, + "bbox": [ + 89, + 604, + 243, + 619 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "In this section, we conduct an ablation study to analyze the impact of modifying various core components of our method. The ablation experiments are conducted with the LLaVA 1.5-7B model. First, we show the effect of pruning tokens inside the LLM in Tab. 3 using 5 datasets. By default, in our method, visual tokens are pruned before being passed to the first decoder layer in the LLM, which we refer to as 'Layer 0'. We also tested 'Layer 1' where the first layer is processed without pruning and the pruning is performed afterward. We further extended this approach by allowing tokens to pass through the first few layers unpruned and then pruning them after specific layers. As shown in the table, for a fixed TFLOP ratio of $19.61\\%$ , pruning done by our method at layer 0 achieves higher task accuracies compared to pruning at layers 1, 2, and 3 of the LLM.", + "bbox": [ + 89, + 628, + 482, + 854 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Furthermore, in Tab. 4, we provide an analysis of using alternative diversity measures for token pruning. The first three rows show the impact of choosing different dis", + "bbox": [ + 89, + 854, + 483, + 900 + ], + "page_idx": 7 + }, + { + "type": "table", + "img_path": "images/4e2068b36c653fdc668c9e4a0f0b79bca6e46954e3ed0268beeafd60a980a520.jpg", + "table_caption": [ + "Table 3. Ablation study on applying DivPrune at different layers." + ], + "table_footnote": [], + "table_body": "
TFLOP (ratio %)MMB AccMMMU AccPOPE F1SQA EMAvg
Cosine (Ours)19.6159.1935.8986.0268.2762.34
l119.6159.7134.6785.4067.9761.94
l219.6159.9735.0085.6468.2762.22
Random19.6152.6634.5672.7866.6356.66
Min-Max19.6138.5733.1149.2665.2046.53
", + "bbox": [ + 514, + 89, + 903, + 194 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Table 4. Ablation on using various diversity measures.", + "bbox": [ + 545, + 198, + 870, + 213 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "tance measures to quantify the similarity among tokens. It can be seen that all three similarity measures, cosine, $\\ell_1$ , and $\\ell_2$ perform comparably, with cosine (default setting) performing slightly better. This suggests that the choice of similarity measure does not significantly impact DivPrune's overall performance.", + "bbox": [ + 511, + 236, + 903, + 325 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "The last two rows in Tab. 4 show the effect of choosing alternative strategies of token selection other than the proposed Max-Min diversity-based solution (3). We tested random pruning as well as the Min-Max strategy where the maximum distance between the selected samples is minimized. The Min-Max strategy enforces high redundancy among the selected samples, resulting in reduced diversity. As results in the bottom part of Tab. 4 reveal that any deviation from our proposed selection strategy results in suboptimal performance. Specifically, the Min-Max strategy performs the worst, showing approximately $15.8\\%$ lower performance compared to ours. This decline is due to the Min-Max approach selecting tokens that are highly similar to each other, resulting in less diversity among the selected visual tokens. Random selection provides some degree of diversity, but it performs $5.6\\%$ worse than the proposed method because it cannot guarantee maximum diversity. This proves that redundancy of visual tokens leads to poor performance and diversity maximization is needed for optimal performance, corroborating the utility and need of the proposed diversity maximization in Eq. (3).", + "bbox": [ + 511, + 328, + 906, + 645 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "5. Conclusion", + "text_level": 1, + "bbox": [ + 511, + 662, + 633, + 678 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "In this paper, we proposed a token pruning method based on a max-min diversity problem, called DivPrune. In the proposed method, maximum diversity is achieved among the selected tokens, resulting in reduced redundancy. By ensuring high diversity, the selected tokens provide a more representative subset of the original tokens, enabling effective performance even at high pruning ratios without requiring fine-tuning. Extensive experiments were conducted with multiple LMMs on image and video understanding tasks across 16 datasets. The results show that DivPrune achieves state-of-the-art accuracy on the tested datasets. DivPrune generalizes well to different model sizes and architectures, while also improving memory consumption and end-to-end latency for the tested LMMs.", + "bbox": [ + 511, + 688, + 906, + 900 + ], + "page_idx": 7 + }, + { + "type": "page_number", + "text": "9399", + "bbox": [ + 480, + 944, + 514, + 955 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "References", + "text_level": 1, + "bbox": [ + 91, + 89, + 187, + 104 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[1] Josh Achiam, Steven Adler, Sandhini Agarwal, Lama Ahmad, Ilge Akkaya, Florencia Leoni Aleman, Diogo Almeida, Janko Altenschmidt, Sam Altman, Shyamal Anadkat, et al. Gpt-4 technical report. arXiv preprint arXiv:2303.08774, 2023. 1", + "[2] Harsh Agrawal, Karan Desai, Yufei Wang, Xinlei Chen, Rishabh Jain, Mark Johnson, Dhruv Batra, Devi Parikh, Stefan Lee, and Peter Anderson. nocaps: novel object captioning at scale. In Proceedings of the IEEE International Conference on Computer Vision, pages 8948-8957, 2019. 6", + "[3] Mu Cai, Jianwei Yang, Jianfeng Gao, and Yong Jae Lee. Matryoshka multimodal models. Proceedings of the International Conference on Learning Representation, 2025. 2, 3, 4, 6", + "[4] Liang Chen, Haozhe Zhao, Tianyu Liu, Shuai Bai, Junyang Lin, Chang Zhou, and Baobao Chang. An image is worth 1/2 tokens after layer 2: Plug-and-play inference acceleration for large vision-language models. In European Conference on Computer Vision (ECCV), 2024. 1, 2, 4, 5, 6, 7", + "[5] Zesen Cheng, Sicong Leng, Hang Zhang, Yifei Xin, Xin Li, Guanzheng Chen, Yongxin Zhu, Wenqi Zhang, Ziyang Luo, Deli Zhao, et al. Videollama 2: Advancing spatial-temporal modeling and audio understanding in video-llms. arXiv preprint arXiv:2406.07476, 2024. 2", + "[6] Wei-Lin Chiang, Zhuohan Li, Zi Lin, Ying Sheng, Zhanghao Wu, Hao Zhang, Lianmin Zheng, Siyuan Zhuang, Yonghao Zhuang, Joseph E Gonzalez, et al. Vicuna: An open-source chatbot impressing gpt-4 with $90\\%$ * chatgpt quality. See https://vicuna.lmsys.org (accessed 14 April 2023), 2(3):6, 2023. 1", + "[7] Krzysztof Choromanski, Valerii Likhosherstov, David Dohan, Xingyou Song, Andreea Gane, Tamas Sarlos, Peter Hawkins, Jared Davis, Afroz Mohiuddin, Lukasz Kaiser, et al. Rethinking attention with performers. arXiv preprint arXiv:2009.14794, 2020. 1", + "[8] Tri Dao, Dan Fu, Stefano Ermon, Atri Rudra, and Christopher Ré. Flashattention: Fast and memory-efficient exact attention with io-awareness. Advances in Neural Information Processing Systems, 35:16344-16359, 2022. 1", + "[9] Chaoyou Fu, Peixian Chen, Yunhang Shen, Yulei Qin, Mengdan Zhang, Xu Lin, Jinrui Yang, Xiawu Zheng, Ke Li, Xing Sun, et al. Mme: A comprehensive evaluation benchmark for multimodal large language models. arXiv preprint arXiv:2306.13394, 2023. 5, 6", + "[10] Jinlan Fu, See-Kiong Ng, Zhengbao Jiang, and Pengfei Liu. Gptscore: Evaluate as you desire, 2023. 5", + "[11] Mukul Gagrani, Raghavv Goel, Wonseok Jeon, Junyoung Park, Mingu Lee, and Christopher Lott. On speculative decoding for multimodal large language models. arXiv preprint arXiv:2404.08856, 2024. 2", + "[12] Albert Gu and Tri Dao. Mamba: Linear-time sequence modeling with selective state spaces. arXiv preprint arXiv:2312.00752, 2023. 2", + "[13] Zhiyu Guo, Hidetaka Kamigaito, and Taro Watanabe. Attention score is not all you need for token importance indicator" + ], + "bbox": [ + 93, + 114, + 483, + 900 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "in kv cache reduction: Value also matters. arXiv preprint arXiv:2406.12335, 2024. 2", + "[14] Drew A Hudson and Christopher D Manning. Gqa: A new dataset for real-world visual reasoning and compositional question answering. Conference on Computer Vision and Pattern Recognition (CVPR), 2019. 6", + "[15] Angelos Katharopoulos, Apoorv Vyas, Nikolaos Pappas, and François Fleuret. Transformers are rnns: Fast autoregressive transformers with linear attention. In International conference on machine learning, pages 5156-5165. PMLR, 2020. 1", + "[16] Bohao Li, Rui Wang, Guangzhi Wang, Yuying Ge, Yixiao Ge, and Ying Shan. Seed-bench: Benchmarking multimodal llms with generative comprehension. arXiv preprint arXiv:2307.16125, 2023. 5, 6, 7", + "[17] Wentong Li, Yuqian Yuan, Jian Liu, Dongqi Tang, Song Wang, Jianke Zhu, and Lei Zhang. Tokenpacker: Efficient visual projector for multimodal llm. arXiv preprint arXiv:2407.02392, 2024. 2, 3", + "[18] Yifan Li, Yifan Du, Kun Zhou, Jinpeng Wang, Wayne Xin Zhao, and Ji-Rong Wen. Evaluating object hallucination in large vision-language models. arXiv preprint arXiv:2305.10355, 2023. 6", + "[19] Bin Lin, Yang Ye, Bin Zhu, Jiaxi Cui, Munan Ning, Peng Jin, and Li Yuan. Video-llava: Learning united visual representation by alignment before projection. CoRR, abs/2311.10122, 2023. 1, 2", + "[20] Tsung-Yi Lin, Michael Maire, Serge Belongie, Lubomir Bourdev, Ross Girshick, James Hays, Pietro Perona, Deva Ramanan, C. Lawrence Zitnick, and Piotr Dólar. Microsoft coco: Common objects in context, 2015. 6", + "[21] Zhihang Lin, Mingbao Lin, Luxi Lin, and Rongrong Ji. Boosting multimodal large language models with visual tokens withdrawal for rapid inference. arXiv preprint arXiv:2405.05803, 2024. 2, 4, 5, 6, 7", + "[22] Haotian Liu, Chunyuan Li, Qingyang Wu, and Yong Jae Lee. Visual instruction tuning. In NeurIPS, 2023. 1, 2, 4", + "[23] Haotian Liu, Chunyuan Li, Yuheng Li, Bo Li, Yuanhan Zhang, Sheng Shen, and Yong Jae Lee. Llava-last: Improved reasoning,OCR,and world knowledge,2024.1,2, 4,7", + "[24] Yuan Liu, Haodong Duan, Yuanhan Zhang, Bo Li, Songyang Zhang, Wangbo Zhao, Yike Yuan, Jiaqi Wang, Conghui He, Ziwei Liu, et al. Mmbench: Is your multi-modal model an all-around player? In European Conference on Computer Vision, pages 216-233. Springer, 2025. 6", + "[25] Pan Lu, Swaroop Mishra, Tony Xia, Liang Qiu, Kai-Wei Chang, Song-Chun Zhu, Oyvind Tafjord, Peter Clark, and Ashwin Kalyan. Learn to explain: Multimodal reasoning via thought chains for science question answering, 2022. 6", + "[26] Muhammad Maaz, Hanoona Rasheed, Salman Khan, and Fahad Shahbaz Khan. Video-chatgpt: Towards detailed video understanding via large vision and language models. In Proceedings of the 62nd Annual Meeting of the Association for Computational Linguistics (ACL 2024), 2024. 7", + "[27] Karttikeya Mangalam, Raiymbek Akshulakov, and Jitendra Malik. Egoschema: A diagnostic benchmark for very long-form video language understanding, 2023. 7" + ], + "bbox": [ + 516, + 92, + 905, + 900 + ], + "page_idx": 8 + }, + { + "type": "page_number", + "text": "9400", + "bbox": [ + 482, + 944, + 514, + 955 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[28] Kenneth Marino, Mohammad Rastegari, Ali Farhadi, and Roozbeh Mottaghi. Ok-vqa: A visual question answering benchmark requiring external knowledge. In Conference on Computer Vision and Pattern Recognition (CVPR), 2019. 6", + "[29] Rafael Martí, Anna Martínez-Gavara, Sergio Pérez-Peló, and Jesús Sánchez-Oro. A review on discrete diversity and dispersion maximization from an or perspective. European Journal of Operational Research, 299(3):795-813, 2022. 4", + "[30] OpenAI. Hello gpt-4o, 2024. https://openai.com/index/hello-gpt-4o/ [Accessed: (Nov 2024)]. 2", + "[31] Bryan A. Plummer, Liwei Wang, Christopher M. Cervantes, Juan C. Caicedo, Julia Hockenmaier, and Svetlana Lazebnik. Flickr30k entities: Collecting region-to-phrase correspondences for richer image-to-sentence models. *IJCV*, 123 (1):74–93, 2017. 6", + "[32] Daniel Cosmin Porumbel, Jin-Kao Hao, and Fred Glover. A simple and effective algorithm for the maxmin diversity problem. Annals of Operations Research, 186:275-293, 2011. 3", + "[33] Yanyuan Qiao, Zheng Yu, Longteng Guo, Sihan Chen, Zijia Zhao, Mingzhen Sun, Qi Wu, and Jing Liu. Vl-mamba: Exploring state space models for multimodal learning. arXiv preprint arXiv:2403.13600, 2024. 2", + "[34] Alec Radford, Jong Wook Kim, Chris Hallacy, Aditya Ramesh, Gabriel Goh, Sandhini Agarwal, Girish Sastry, Amanda Askell, Pamela Mishkin, Jack Clark, et al. Learning transferable visual models from natural language supervision. In International conference on machine learning, pages 8748-8763. PMLR, 2021. 4", + "[35] Mauricio GC Resende, Rafael Martí, Micael Gallego, and Abraham Duarte. Grasp and path relinking for the max-min diversity problem. Computers & Operations Research, 37 (3):498-508, 2010. 2, 4", + "[36] Yuzhang Shang, Mu Cai, Bingxin Xu, Yong Jae Lee, and Yan Yan. Llava-prumerge: Adaptive token reduction for efficient large multimodal models. arXiv preprint arXiv:2403.15388, 2024. 1, 2, 4, 6", + "[37] Mustafa Shukor and Matthieu Cord. Skipping computations in multimodal llms. arXiv preprint arXiv:2410.09454, 2024. 2", + "[38] Sainbayar Sukhbaatar, Edouard Grave, Piotr Bojanowski, and Armand Joulin. Adaptive attention span in transformers. arXiv preprint arXiv:1905.07799, 2019. 1", + "[39] Gemini Team, Petko Georgiev, Ving Ian Lei, Ryan Burnell, Libin Bai, Anmol Gulati, Garrett Tanzer, Damien Vincent, Zhufeng Pan, Shibo Wang, et al. Gemini 1.5: Unlocking multimodal understanding across millions of tokens of context. arXiv preprint arXiv:2403.05530, 2024. 2", + "[40] Hugo Touvron, Thibaut Lavril, Gautier Izacard, Xavier Martinet, Marie-Anne Lachaux, Timothee Lacroix, Baptiste Roziere, Naman Goyal, Eric Hambro, Faisal Azhar, et al. Llama: Open and efficient foundation language models. arXiv preprint arXiv:2302.13971, 2023. 1", + "[41] Laurens Van der Maaten and Geoffrey Hinton. Visualizing data using t-sne. Journal of machine learning research, 9 (11), 2008. 5" + ], + "bbox": [ + 91, + 90, + 480, + 900 + ], + "page_idx": 9 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[42] Ramakrishna Vedantam, C. Lawrence Zitnick, and Devi Parikh. Cider: Consensus-based image description evaluation. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition (CVPR), 2015. 5", + "[43] Zhibiao Wu and Martha Palmer. Verb semantics and lexical selection. arXiv preprint cmp-lg/9406033, 1994. 5", + "[44] Junbin Xiao, Xindi Shang, Angela Yao, and Tat-Seng Chua. Next-qa: Next phase of question-answering to explaining temporal actions. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 9777–9786, 2021. 7", + "[45] Shilin Xu, Xiangtai Li, Haobo Yuan, Lu Qi, Yunhai Tong, and Ming-Hsuan Yang. Llavadi: What matters for multimodal large language models distillation. arXiv preprint arXiv:2407.19409, 2024. 2", + "[46] Yuan Yao, Tianyu Yu, Ao Zhang, Chongyi Wang, Junbo Cui, Hongji Zhu, Tianchi Cai, Haoyu Li, Weilin Zhao, Zhihui He, et al. Minicpm-v: A gpt-4v level mllm on your phone. arXiv preprint arXiv:2408.01800, 2024. 1", + "[47] Weihao Ye, Qiong Wu, Wenhao Lin, and Yiyi Zhou. Fit and prune: Fast and training-free visual token pruning for multi-modal large language models. arXiv preprint arXiv:2409.10197, 2024. 1, 2, 4, 5, 6", + "[48] Zhou Yu, Dejing Xu, Jun Yu, Ting Yu, Zhou Zhao, Yueting Zhuang, and Dacheng Tao. Activitynet-qa: A dataset for understanding complex web videos via question answering. In AAAI, pages 9127–9134, 2019. 7", + "[49] Zhengqing Yuan, Zhaoxu Li, Weiran Huang, Yanfang Ye, and Lichao Sun. Tinygpt-v: Efficient multimodal large language model via small backbones. arXiv preprint arXiv:2312.16862, 2023. 2", + "[50] Xiang Yue, Yuansheng Ni, Kai Zhang, Tianyu Zheng, Ruoqi Liu, Ge Zhang, Samuel Stevens, Dongfu Jiang, Weiming Ren, Yuxuan Sun, Cong Wei, Botao Yu, Ruibin Yuan, Ren-liang Sun, Ming Yin, Boyuan Zheng, Zhenzhu Yang, Yibo Liu, Wenhao Huang, Huan Sun, Yu Su, and Wenhui Chen. Mmmu: A massive multi-discipline multimodal understanding and reasoning benchmark for expert agi. In Proceedings of CVPR, 2024. 6", + "[51] Kaichen Zhang, Bo Li, Peiyuan Zhang, Fanyi Pu, Joshua Adrian Cahyono, Kairui Hu, Shuai Liu, Yuanhan Zhang, Jingkang Yang, Chunyuan Li, and Ziwei Liu. Lmmseval: Reality check on the evaluation of large multimodal models, 2024. 5", + "[52] Yuanhan Zhang, Bo Li, haotian Liu, Yong jae Lee, Liangke Gui, Di Fu, Jiashi Feng, Ziwei Liu, and Chunyuan Li. Llavanext: A strong zero-shot video understanding model, 2024. 1, 2, 4", + "[53] Baichuan Zhou, Ying Hu, Xi Weng, Junlong Jia, Jie Luo, Xien Liu, Ji Wu, and Lei Huang. Tinyllava: A framework of small-scale large multimodal models. arXiv preprint arXiv:2402.14289, 2024. 2" + ], + "bbox": [ + 514, + 90, + 905, + 825 + ], + "page_idx": 9 + }, + { + "type": "page_number", + "text": "9401", + "bbox": [ + 482, + 944, + 513, + 955 + ], + "page_idx": 9 + } +] \ No newline at end of file diff --git a/2025/DivPrune_ Diversity-based Visual Token Pruning for Large Multimodal Models/f2b05006-c955-46f7-a5af-8b7909376348_model.json b/2025/DivPrune_ Diversity-based Visual Token Pruning for Large Multimodal Models/f2b05006-c955-46f7-a5af-8b7909376348_model.json new file mode 100644 index 0000000000000000000000000000000000000000..81bd47c00b14d45bf9d452ebebd9eab0b374f2e2 --- /dev/null +++ b/2025/DivPrune_ Diversity-based Visual Token Pruning for Large Multimodal Models/f2b05006-c955-46f7-a5af-8b7909376348_model.json @@ -0,0 +1,2123 @@ +[ + [ + { + "type": "header", + "bbox": [ + 0.107, + 0.003, + 0.182, + 0.043 + ], + "angle": 0, + "content": "CVF" + }, + { + "type": "header", + "bbox": [ + 0.237, + 0.001, + 0.812, + 0.047 + ], + "angle": 0, + "content": "This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore." + }, + { + "type": "title", + "bbox": [ + 0.099, + 0.131, + 0.9, + 0.153 + ], + "angle": 0, + "content": "DivPrune: Diversity-based Visual Token Pruning for Large Multimodal Models" + }, + { + "type": "text", + "bbox": [ + 0.202, + 0.18, + 0.796, + 0.217 + ], + "angle": 0, + "content": "Saeed Ranjbar Alvar, Gursimran Singh†, Mohammad Akbari†, Yong Zhang Huawei Technologies Canada Co., Ltd." + }, + { + "type": "text", + "bbox": [ + 0.139, + 0.219, + 0.855, + 0.235 + ], + "angle": 0, + "content": "{saeed.ranjbar.alvar1, gursimran.singh1, mohammad.akbari, yong.zhang3}@huawei.com" + }, + { + "type": "title", + "bbox": [ + 0.248, + 0.269, + 0.327, + 0.285 + ], + "angle": 0, + "content": "Abstract" + }, + { + "type": "text", + "bbox": [ + 0.092, + 0.302, + 0.485, + 0.725 + ], + "angle": 0, + "content": "Large Multimodal Models (LMMs) have emerged as powerful models capable of understanding various data modalities, including text, images, and videos. LMMs encode both text and visual data into tokens that are then combined and processed by an integrated Large Language Model (LLM). Including visual tokens substantially increases the total token count, often by thousands. The increased input length for LLM significantly raises the complexity of inference, resulting in high latency in LMMs. To address this issue, token pruning methods, which remove part of the visual tokens, are proposed. The existing token pruning methods either require extensive calibration and fine-tuning or rely on suboptimal importance metrics which results in increased redundancy among the retained tokens. In this paper, we first formulate token pruning as Max-Min Diversity Problem (MMDP) where the goal is to select a subset such that the diversity among the selected tokens is maximized. Then, we solve the MMDP to obtain the selected subset and prune the rest. The proposed method, DivPrune, reduces redundancy and achieves the highest diversity of the selected tokens. By ensuring high diversity, the selected tokens better represent the original tokens, enabling effective performance even at high pruning ratios without requiring fine-tuning. Extensive experiments with various LMMs show that DivPrune achieves state-of-the-art accuracy over 16 image- and video-language datasets. Additionally, DivPrune reduces both the end-to-end latency and GPU memory usage for the tested models. The code is available here\\(^{\\diamond}\\)." + }, + { + "type": "title", + "bbox": [ + 0.092, + 0.75, + 0.222, + 0.765 + ], + "angle": 0, + "content": "1. Introduction" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.775, + 0.483, + 0.866 + ], + "angle": 0, + "content": "Following the success of Large Language Models (LLMs) in language understanding [1, 6, 40], Large Multimodal Models (LMMs) [19, 22, 23, 52] have emerged to handle diverse data types like images and video, by leveraging the foundational capabilities of LLMs. Typically, LMMs encode text and visual modalities into tokens, also known" + }, + { + "type": "image", + "bbox": [ + 0.545, + 0.276, + 0.88, + 0.463 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.512, + 0.476, + 0.908, + 0.588 + ], + "angle": 0, + "content": "Figure 1. Comparison of different visual token pruning methods across various pruning ratios for LLaVA 1.5-7B. The y-axis is the performance averaged on COCO (CIDEr), OKVQA (Acc), POPE (F1), and MMBench (Acc). The x-axis is the TFLOP ratio of the model after token pruning compared to the original model before pruning. The proposed method significantly outperforms all baselines. Note that, unlike other methods, FitPrune uses an additional calibration step to prune tokens." + }, + { + "type": "text", + "bbox": [ + 0.511, + 0.608, + 0.907, + 0.759 + ], + "angle": 0, + "content": "as embeddings. These tokens are then combined and processed by an integrated LLM. The inclusion of visual tokens significantly increases the total number of tokens, often adding thousands to the combined set. Since the running time and memory requirements scale quadratically with input size [7, 8, 15, 38], the addition of visual tokens can substantially raise the running time for LMMs. Hence, many of these models often struggle to meet the demands of low-latency applications, particularly in resource-constrained environments [46]." + }, + { + "type": "text", + "bbox": [ + 0.511, + 0.762, + 0.909, + 0.883 + ], + "angle": 0, + "content": "Previous research [4, 36, 47] has demonstrated that there is a high degree of redundancy in the visual information processed by LMMs. As a result, visual token pruning has emerged as a promising solution to address the computational complexity challenges faced by LMMs. Specifically, previous research has demonstrated that reducing the number of visual tokens by \\(50\\%\\) [4] to \\(95\\%\\) [36] can significantly enhance the inference speed of LMMs." + }, + { + "type": "text", + "bbox": [ + 0.534, + 0.886, + 0.907, + 0.902 + ], + "angle": 0, + "content": "While promising, token pruning methods have certain" + }, + { + "type": "page_footnote", + "bbox": [ + 0.114, + 0.876, + 0.305, + 0.888 + ], + "angle": 0, + "content": "Authors have equal contributions." + }, + { + "type": "page_footnote", + "bbox": [ + 0.114, + 0.888, + 0.302, + 0.901 + ], + "angle": 0, + "content": "\\(\\diamond\\) https://github.com/vbdi/divprune" + }, + { + "type": "list", + "bbox": [ + 0.114, + 0.876, + 0.305, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.482, + 0.945, + 0.516, + 0.957 + ], + "angle": 0, + "content": "9392" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.09, + 0.092, + 0.482, + 0.318 + ], + "angle": 0, + "content": "shortcomings. For example, the works in [3, 17, 21, 47] require calibration or finetuning for each model which is costly and time-consuming to implement. FastV [4] and PruMerge [36] use attention scores to identify less important tokens for pruning. However, it is shown that using attention scores is not optimal, as some important tokens are overlooked [21]. Additionally, attention-based pruning tends to retain tokens that are similar to each other, leading to redundancy. At high compression ratio, this redundancy prevents the selection of a sufficient number of unique tokens to accurately represent the original tokens. In line with this observation, our findings indicate that pruning a large portion of visual tokens using these methods, without subsequent fine-tuning, results in a significant drop in the performance of LMMs across various tasks (Fig. 1)." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.32, + 0.482, + 0.531 + ], + "angle": 0, + "content": "To address the above-mentioned issues, we formulate token pruning as a Max-Min Diversity Problem (MMDP) [35]. In an MMDP, the objective is to select a subset of elements such that the diversity among them is maximized. We apply this concept to token pruning, which we call DivPrune, aiming to maximize the diversity of the selected tokens by increasing the minimum distance between them. By ensuring high diversity, DivPrune captures a broader range of visual tokens, making it inherently more robust compared to attention-based methods that focus only on token importance scores. Increasing the diversity also helps ensure that the selected tokens better represent the original set of tokens, enabling effective performance even at high pruning ratios without the need for fine-tuning." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.533, + 0.483, + 0.698 + ], + "angle": 0, + "content": "DivPrune also offers practical advantages that make it a highly useful solution in real-world scenarios. DivPrune is a plug-and-play solution that can be used without requiring offline optimization with a calibration set, or fine-tuning of the model, which are often time-consuming and computationally expensive. DivPrune is applicable to LMMs with any LLM architecture and vision encoder. Additionally, DivPrune is compatible with inference optimization techniques, such as KV caching, resulting in practical speedup in real-world applications. In summary, our major contributions are as follows:" + }, + { + "type": "text", + "bbox": [ + 0.091, + 0.705, + 0.483, + 0.763 + ], + "angle": 0, + "content": "- We introduce DivPrune, a token pruning method based on MMDP that maximizes diversity among visual tokens, effectively reducing redundancy and ensuring a highly representative subset." + }, + { + "type": "text", + "bbox": [ + 0.091, + 0.765, + 0.483, + 0.809 + ], + "angle": 0, + "content": "- DivPrune is a training-free, calibration-data-free, plug-and-play solution that can be seamlessly integrated with off-the-shelf LMMs." + }, + { + "type": "text", + "bbox": [ + 0.091, + 0.811, + 0.483, + 0.884 + ], + "angle": 0, + "content": "- We conduct evaluations using 16 datasets on image-and video-language models with image and video understanding tasks. DivPrune achieves state-of-the-art performance, with noticeable gains under extreme pruning (i.e., ratio \\(>80\\%\\))." + }, + { + "type": "text", + "bbox": [ + 0.091, + 0.886, + 0.483, + 0.901 + ], + "angle": 0, + "content": "DivPrune reduces GPU memory usage and inference la" + }, + { + "type": "list", + "bbox": [ + 0.091, + 0.705, + 0.483, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "text", + "bbox": [ + 0.527, + 0.092, + 0.907, + 0.122 + ], + "angle": 0, + "content": "tency while maintaining comparable accuracy compared to the original model across most datasets." + }, + { + "type": "title", + "bbox": [ + 0.514, + 0.136, + 0.665, + 0.151 + ], + "angle": 0, + "content": "2. Related Works" + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.161, + 0.821, + 0.177 + ], + "angle": 0, + "content": "2.1. Large Multimodal Models (LMMs)" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.184, + 0.907, + 0.351 + ], + "angle": 0, + "content": "LMMs handle diverse data types, including text, audio, image, and, video [5, 19, 22, 23, 30, 39, 52]. This work focuses on open-source LMMs that support language and visual inputs. These LMMs can be categorized into two types: image-based and video-based LMMs. The image-based LMMs [22, 23] address image-language understanding tasks, like image captioning, visual question answering, and image reasoning. On the other hand, video-based LMMs are geared towards video understanding [19, 52] tasks, like video captioning, video summarization, and video question answering." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.361, + 0.673, + 0.375 + ], + "angle": 0, + "content": "2.2. Efficient LMMs" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.384, + 0.909, + 0.567 + ], + "angle": 0, + "content": "Several techniques are proposed to improve inference efficiency specifically for LMMs. The first technique is to change the model architecture in LMMs. For example, [33] proposed to replace transformer-based LLMs with Mamba model [12]. [49, 53] retrained LMMs with small scale LLMs to improve their efficiency. [45] used knowledge distillation to train a small LMM. In addition to changing the architecture, it is shown in [37] that skipping some blocks or layers within LMMs can improve the inference speed without damaging the model's performance. Furthermore, efficient decoding techniques such as speculative decoding are proposed to make LMM inference more efficient [11]." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.575, + 0.719, + 0.591 + ], + "angle": 0, + "content": "2.3. Visual Token Pruning" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.599, + 0.907, + 0.75 + ], + "angle": 0, + "content": "Visual token pruning methods are proposed to reduce the inference complexity for LMMs. The first group of methods uses attention scores to prune tokens [4, 36]. PruMerge [36] introduces a token pruning method for the vision encoder where the visual tokens are clustered and merged based on their attention sparsity. In addition, FastV [4] prunes tokens within a specific layer of the LLM based on the magnitude of attention scores in an earlier layer. It is shown that pruning tokens based on attention scores are not optimal [13, 21], especially at higher pruning ratios." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.75, + 0.909, + 0.901 + ], + "angle": 0, + "content": "Calibration-based methods offer another line of work, where pruning layers and/or ratios are determined by analyzing the LLM outputs for a calibration dataset [21, 47]. For example, FitPrune [47] calculates a pruning recipe based on the observed attention divergence before and after pruning. VTW [21] argues that visual tokens can be entirely removed after a certain layer within LLM. The layer to remove the visual tokens is chosen using a calibration dataset. These methods rely on calibration datasets and require custom calibration for each LMM, which can be costly and" + }, + { + "type": "page_number", + "bbox": [ + 0.482, + 0.945, + 0.516, + 0.957 + ], + "angle": 0, + "content": "9393" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.095, + 0.088, + 0.485, + 0.288 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.09, + 0.301, + 0.485, + 0.344 + ], + "angle": 0, + "content": "Figure 2. An overview of the LMM architecture, with DivPrune applied to visual tokens. The blocks on the right-hand side illustrate the steps of the method." + }, + { + "type": "text", + "bbox": [ + 0.091, + 0.357, + 0.29, + 0.37 + ], + "angle": 0, + "content": "cumbersome for new models." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.372, + 0.485, + 0.508 + ], + "angle": 0, + "content": "Some previous works proposed token pruning with the need for fine-tuning. \\(\\mathbf{M}^3\\) [3] applies model fine-tuning to produce nested visual token representations at multiple granularities, allowing users to select token lengths dynamically during inference. In [17], a projector layer trained using a large-scale dataset is proposed that packs finer detailed information into compact token representations. These methods need significant computational resources for training, limiting their use across various scenarios." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.52, + 0.266, + 0.538 + ], + "angle": 0, + "content": "3. Proposed Method" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.546, + 0.484, + 0.593 + ], + "angle": 0, + "content": "In this section, we briefly discuss how LMMs work. Then, the token pruning problem is defined, followed by a detailed presentation of the proposed method." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.6, + 0.399, + 0.617 + ], + "angle": 0, + "content": "3.1. Large Multimodal Models (LMMs)" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.623, + 0.484, + 0.76 + ], + "angle": 0, + "content": "An LMM typically processes a pair of inputs, denoted as \\((T,V)\\), where \\(T\\) is the text input and \\(V\\) is the visual input such as image or video. The text input is mapped to \\(N\\) textual tokens \\(\\mathbf{E}_{\\mathbf{t}} = \\{t_1,\\dots ,t_N\\}\\) using a text encoder. Similarly, the visual input is processed by a corresponding vision encoder. Specifically, it takes visual information \\(V\\) as input and outputs image features, that are further converted to \\(M\\) (generally \\(M\\gg N\\)) vision tokens \\(\\mathbf{E}_{\\mathbf{v}} = \\{v_{1},\\ldots ,v_{M}\\}\\) using a projector layer (Fig. 2)." + }, + { + "type": "text", + "bbox": [ + 0.091, + 0.76, + 0.484, + 0.822 + ], + "angle": 0, + "content": "The textual tokens and visual tokens are then combined to be fed to an LLM to generate the prediction in an autoregressive manner. Specifically, \\(\\hat{N}\\) output tokens \\(\\mathbf{Y} = \\{y_1,\\dots ,y_{\\hat{N}}\\}\\) are generated as follows:" + }, + { + "type": "equation", + "bbox": [ + 0.106, + 0.832, + 0.483, + 0.876 + ], + "angle": 0, + "content": "\\[\nP \\left(y _ {1}, \\dots , y _ {\\hat {N}} \\mid \\mathbf {E} _ {\\mathbf {t}}, \\mathbf {E} _ {\\mathbf {v}}\\right) = \\prod_ {i = 1} ^ {\\hat {N}} P \\left(y _ {i} \\mid y _ {< i}, \\mathbf {E} _ {\\mathbf {t}}, \\mathbf {E} _ {\\mathbf {v}}\\right), \\tag {1}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.091, + 0.886, + 0.484, + 0.902 + ], + "angle": 0, + "content": "where \\( P(|) \\) is the conditional probability obtained at the" + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.092, + 0.645, + 0.106 + ], + "angle": 0, + "content": "output of the LLM." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.114, + 0.667, + 0.131 + ], + "angle": 0, + "content": "3.2. Token Pruning" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.138, + 0.906, + 0.198 + ], + "angle": 0, + "content": "Reducing the number of input tokens in an integrated LLM within LMMs helps to lower memory usage and inference latency. Since visual tokens tend to have more redundancy, they are generally selected for pruning." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.199, + 0.907, + 0.38 + ], + "angle": 0, + "content": "In this context, the problem of token pruning can be defined as follows: given a set of visual tokens \\(\\mathbf{E}_{\\mathbf{v}}\\) with \\(|\\mathbf{E}_{\\mathbf{v}}| = M\\) and the subset size \\(\\tilde{M} (\\tilde{M} < M)\\), the goal is to select a subset, \\(\\tilde{\\mathbf{E}}_{\\mathbf{v}}\\), while preserving key information necessary for accurate predictions. To mathematically formulate the token pruning problem, we define a mapping function \\(f\\), which maps the original set of visual tokens, \\(\\mathbf{E}_{\\mathbf{v}}\\), to a subset, \\(\\tilde{\\mathbf{E}}_{\\mathbf{v}} = \\{\\tilde{v}_1,\\dots ,\\tilde{v}_{\\tilde{M}}\\}\\), where \\(|\\tilde{\\mathbf{E}}_{\\mathbf{v}}| = \\tilde{M}\\). The objective is to identify a mapping function \\(f\\) that minimizes the difference in the model's output before and after pruning while ensuring the reduced set still captures the essential information from the original set:" + }, + { + "type": "equation", + "bbox": [ + 0.654, + 0.386, + 0.8, + 0.404 + ], + "angle": 0, + "content": "\\[\n\\text {F i n d :} \\quad f: \\mathbf {E} _ {v} \\rightarrow \\tilde {\\mathbf {E}} _ {\\mathbf {v}}\n\\]" + }, + { + "type": "equation", + "bbox": [ + 0.619, + 0.407, + 0.906, + 0.433 + ], + "angle": 0, + "content": "\\[\n\\text {O b j e c t i v e :} \\quad \\min _ {f} \\mathcal {L} (\\mathcal {P}, \\tilde {\\mathcal {P}}) \\tag {2}\n\\]" + }, + { + "type": "equation", + "bbox": [ + 0.636, + 0.435, + 0.802, + 0.454 + ], + "angle": 0, + "content": "\\[\n\\text {S u b j e c t} \\quad | \\tilde {\\mathbf {E}} _ {\\mathbf {v}} | = \\tilde {M},\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.462, + 0.906, + 0.554 + ], + "angle": 0, + "content": "where \\(\\mathcal{P} = P(y_1, \\ldots, y_{\\hat{N}} \\mid \\mathbf{E}_{\\mathbf{t}}, \\mathbf{E}_{\\mathbf{v}})\\) and \\(\\tilde{\\mathcal{P}} = P(y_1, \\ldots, y_{\\hat{N}} \\mid \\mathbf{E}_{\\mathbf{t}}, f(\\mathbf{E}_{\\mathbf{v}}))\\). Here, \\(\\mathcal{L}\\) represents a loss function that measures the difference in the model's output with and without pruning, and \\(\\tilde{M}\\) indicates the number of retained tokens. Next, we propose a novel diversity-based solution for the introduced token pruning problem." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.562, + 0.776, + 0.577 + ], + "angle": 0, + "content": "3.3. DivPrune: Method Overview" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.585, + 0.906, + 0.706 + ], + "angle": 0, + "content": "We proposed a diversity-based token pruning method by reformulating the problem in (2) to select a subset of \\(\\tilde{M}\\) elements that maximizes the diversity, thereby reducing redundancy. Specifically, we define token pruning as Max-Min Diversity Problem (MMDP) [32] where the goal is to find the set \\(\\tilde{\\mathbf{E}}_{\\mathbf{v}}\\) among all possible sets with \\(\\tilde{M}\\) samples in \\(\\mathbf{E}_{\\mathbf{v}}\\) that has the maximum minimum distance between its elements. So, MMDP is defined as:" + }, + { + "type": "equation", + "bbox": [ + 0.53, + 0.712, + 0.906, + 0.746 + ], + "angle": 0, + "content": "\\[\n\\operatorname {F i n d} \\tilde {\\mathbf {E}} _ {\\mathbf {v}} = \\arg \\max \\left[ \\min _ {\\gamma , \\omega \\in S} (d (\\gamma , \\omega)): \\forall S \\subset \\mathbf {E} _ {\\mathbf {v}} \\right], \\tag {3}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.753, + 0.907, + 0.811 + ], + "angle": 0, + "content": "where \\(S\\) is an arbitrary set in \\(\\mathbf{E}_{\\mathbf{v}}\\) with \\(\\tilde{M}\\) elements and \\((\\gamma, \\omega)\\) are arbitrary elements in \\(S\\). The distance is measured by \\(d(.,.)\\) which is defined using the cosine distance as follows:" + }, + { + "type": "equation", + "bbox": [ + 0.627, + 0.809, + 0.905, + 0.839 + ], + "angle": 0, + "content": "\\[\nd (\\gamma , \\omega) = 1 - \\frac {\\gamma \\cdot \\omega}{\\| \\gamma \\| \\| \\omega \\|}. \\tag {4}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.841, + 0.907, + 0.902 + ], + "angle": 0, + "content": "A solution for the MMDP problem in (3) is a subset of \\(\\mathbf{E}_{\\mathbf{v}}\\) that maximizes diversity by minimizing redundancy between elements. In the literature, several solutions including exact and heuristic methods are proposed to solve the" + }, + { + "type": "page_number", + "bbox": [ + 0.482, + 0.945, + 0.516, + 0.957 + ], + "angle": 0, + "content": "9394" + } + ], + [ + { + "type": "code_caption", + "bbox": [ + 0.101, + 0.095, + 0.416, + 0.111 + ], + "angle": 0, + "content": "Algorithm 1: Proposed Token Pruning Method" + }, + { + "type": "algorithm", + "bbox": [ + 0.095, + 0.114, + 0.443, + 0.451 + ], + "angle": 0, + "content": "1 \\(\\tilde{M}\\) : subset size; \\(\\mathbf{E}_{\\mathbf{v}}\\) : visual tokens; \\(\\tilde{\\mathbf{E}}_{\\mathbf{v}}\\) : selected subset \n2 Initialize \\(\\tilde{\\mathbf{E}}_{\\mathbf{v}} = []\\) and \\(\\mathbf{R} = \\mathbf{E}_{\\mathbf{v}}\\) \n3 // First stage: add the first token \n4 \\(D = []\\) initialize the distance array \n5 for i in R do \n6 \\(d_{min} = +inf\\) \n7 for j in R do \n8 If \\((i \\neq j \\& d(i, j) \\leq d_{min})\\) then \\(d_{min} = d(i, j)\\) \n9 Add \\(d_{min}\\) to \\(D\\) \n10 \\(k = \\mathbf{R}[\\arg \\max (\\mathrm{D})]\\) \n11 move k from R to \\(\\tilde{\\mathbf{E}}_{\\mathbf{v}}\\) \n12 // Second stage: iteratively add the subsequent tokens \n13 while \\(|\\tilde{\\mathbf{E}}_{\\mathbf{v}}| < \\tilde{M}\\) do \n14 \\(D = []\\) initialize the distance array \n15 for i in R do \n16 \\(d_{min} = +inf\\) \n17 for j in \\(\\tilde{\\mathbf{E}}_{\\mathbf{v}}\\) do \n18 If \\(d(i, j) \\leq d_{min}\\) then \\(d_{min} = d(i, j)\\) \n19 Add \\(d_{min}\\) to \\(D\\) \n20 \\(k = \\mathbf{R}[\\arg \\max (\\mathrm{D})]\\) \n21 move k from R to \\(\\tilde{\\mathbf{E}}_{\\mathbf{v}}\\) \n22 Return \\(\\tilde{\\mathbf{E}}_{\\mathbf{v}}\\)" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.489, + 0.484, + 0.639 + ], + "angle": 0, + "content": "MMDP problem [29, 35]. Since the number of tokens is generally limited (e.g., 576 in LLaVA 1.5 [22]) and the solvers are not generally designed for GPU acceleration, we obtain exact solution for the problem. Notably, the overhead of the selection process using GPU is negligible compared to the computations within the LLM. Detailed steps of the proposed method is summarized in Algorithm 1. Once the selected tokens are identified, the remaining visual tokens are discarded. The selected tokens along with the textual tokens are passed to the LLM." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.642, + 0.484, + 0.854 + ], + "angle": 0, + "content": "As shown in Algorithm 1, the proposed method has two stages after the initialization. The selected subset, \\(\\tilde{\\mathbf{E}}_{\\mathbf{v}}\\), is initialized as empty, and the candidate list \\(\\mathbf{R}\\) is initialized with all the visual tokens. In the first stage, the first token of the selected subset is chosen based on the pairwise distance between the tokens of the candidate list. Then, the chosen token is moved from the candidate list to the selected list. In the second stage, similar to the first stage, the pairwise distance of the tokens in \\(\\tilde{\\mathbf{E}}_{\\mathbf{v}}\\) and the tokens in \\(\\mathbf{R}\\) is used to add samples to \\(\\tilde{\\mathbf{E}}_{\\mathbf{v}}\\) iteratively. Finally, once the number of tokens in \\(\\tilde{\\mathbf{E}}_{\\mathbf{v}}\\) reaches the specified subset size, the selection procedure is terminated and the \\(\\tilde{\\mathbf{E}}_{\\mathbf{v}}\\) is returned. To avoid repeated distance calculations over iterations a distance matrix is initially calculated by one matrix multiplication." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.856, + 0.484, + 0.901 + ], + "angle": 0, + "content": "The proposed method can also be applied to the features (i.e., hidden states) in the intermediate layers of the LLM. In this case, our method is not applied to the visual tokens, but" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.092, + 0.906, + 0.183 + ], + "angle": 0, + "content": "to the features corresponding to the visual tokens obtained from a decoder layer to select a subset before feeding them to the subsequent layers. In either case, our method obtains the highest diversity for the selected elements. Ablation studies are provided in the next section to analyze the effect of pruning different elements at different layers." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.195, + 0.646, + 0.213 + ], + "angle": 0, + "content": "4. Experiments" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.221, + 0.906, + 0.312 + ], + "angle": 0, + "content": "In this section, we present a comprehensive analysis comparing the performance of our method and previous works across various settings, tasks, and datasets. Insights into the proposed method are also provided through illustrative examples. Moreover, the efficiency of DivPrune along with ablation study are provided." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.32, + 0.722, + 0.337 + ], + "angle": 0, + "content": "4.1. Experimental Settings" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.343, + 0.907, + 0.555 + ], + "angle": 0, + "content": "Baselines and Models: We consider five baselines, namely, FastV [4], PruMerge [36], VTW [21], FitPrune [47] and \\(\\mathbf{M}^3\\) [3]. Among these, we consider FastV, PruMerge, and VTW as our main competitors as they are plug-and-play and do not rely on any further costly finetuning or calibration process. However, for the sake of completeness, we also report performance comparison with respect to one finetuning-based \\((\\mathbf{M}^3)\\) and one calibration-based (FitPrune) methods. Note that, VTW, by default, requires calibration to determine the best layer for a given task. However, doing that does not allow us to set a specific TFLOP ratio, complicating the comparison. Hence, whenever required we disable the calibration of VTW to select the layer that matches the FLOP requirement of a particular experiment." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.555, + 0.906, + 0.66 + ], + "angle": 0, + "content": "We test DivPrune and the baselines with popular LMMs namely LLaVA 1.5-7B [22]1, LLaVA 1.5-13B [22]2, LLaVA 1.6-7B3 (also known as LLaVA-NeXT [23]), and LLaVA-NeXT-Video-7B [52]4 to demonstrate the generality of DivPrune. For each tested model and task, we report only the relevant subset of baseline that is applicable to that specific model and task, alongside our results." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.661, + 0.906, + 0.781 + ], + "angle": 0, + "content": "All the tested LMMs used CLIP vision encoder [34]. LLaVA 1.5 model uses 576 visual tokens to represent images. LLaVA 1.6 converts each image into a varying number of patches, resulting in 3-5 times more visual tokens compared to LLaVA 1.5. LLaVA-NeXT-Video uses 144 tokens to process each frame. For all the experiments with LLaVA-NeXT-Video we used a total of 8 frames resulting in 1152 tokens for the processed frames." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.782, + 0.906, + 0.843 + ], + "angle": 0, + "content": "Datasets, Tasks, and Metrics: We selected a comprehensive set of common tasks and datasets aimed at multimodal reasoning and understanding. Specifically, we chose 11 image-language and 5 video-language datasets." + }, + { + "type": "ref_text", + "bbox": [ + 0.531, + 0.85, + 0.791, + 0.863 + ], + "angle": 0, + "content": "1https://huggingface.co/liuhaotian/flava-v1.5-7B" + }, + { + "type": "ref_text", + "bbox": [ + 0.531, + 0.863, + 0.794, + 0.876 + ], + "angle": 0, + "content": "\\(^{2}\\)https://huggingface.co/luhaotian/llava-v1.5-13b" + }, + { + "type": "ref_text", + "bbox": [ + 0.531, + 0.876, + 0.827, + 0.888 + ], + "angle": 0, + "content": "3https://huggingface.co/liuhaotian/llava-v1.6-vicuna-7b" + }, + { + "type": "ref_text", + "bbox": [ + 0.531, + 0.888, + 0.875, + 0.901 + ], + "angle": 0, + "content": "4https://huggingface.co/lmms-lab/LLaVA-NeXT-Video-7B-DPO" + }, + { + "type": "list", + "bbox": [ + 0.531, + 0.85, + 0.875, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.482, + 0.945, + 0.516, + 0.957 + ], + "angle": 0, + "content": "9395" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.09, + 0.092, + 0.484, + 0.302 + ], + "angle": 0, + "content": "These datasets encompass a wide range of tasks, including captioning, multiple-choice Question Answering (QA), and open-ended QA based on text and image/video inputs. Consistent with prior works, CIDEr score [42] is used for evaluating captioning tasks, and Exact Match (EM), Accuracy (Acc), F1, Perception Score (P-score) [9] and GPT-assisted [10] score are used for QA tasks. Furthermore, Wu-Palmer similarity (WUPS) score [43] and GPT-assisted score [10] is used for open-ended QA. For all task performance metrics used in this paper, higher values indicate better performance. For the reported time and memory, lower values indicate better results. Further details regarding the datasets, tasks, and metrics are provided in the supplementary material." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.303, + 0.484, + 0.425 + ], + "angle": 0, + "content": "Following the earlier works in [4, 21, 47], we report the computational requirement, measured in TFLOPs, for DivPrune and the baselines. Various configurations including different pruning ratios at different layers are examined to obtain different working TFLOPs for our method and the baselines. The reported TFLOP ratio is the TFLOP of the model with pruned tokens relative to the original model's TFLOP with no pruning. This ratio is estimated as [4]:" + }, + { + "type": "equation", + "bbox": [ + 0.107, + 0.436, + 0.482, + 0.462 + ], + "angle": 0, + "content": "\\[\n\\frac {K \\times \\left(4 \\mu d ^ {2} - 2 \\mu^ {2} d + 2 \\mu d m\\right) + (T - K) \\times \\left(4 \\tilde {\\mu} d ^ {2} - 2 \\tilde {\\mu} ^ {2} d + 2 \\tilde {\\mu} d m\\right)}{T \\times \\left(4 \\mu d ^ {2} - 2 \\mu^ {2} d + 2 \\mu d m\\right)}, \\tag {5}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.472, + 0.484, + 0.653 + ], + "angle": 0, + "content": "where \\(T\\) is the total transformer-based decoder layers. \\(\\mu = N + M\\) is the total sequence length before pruning, \\(\\tilde{\\mu} = N + \\tilde{M}\\) is the sequence length after pruning, \\(d\\) is the hidden state size of the layer, and \\(m\\) is the intermediate size of feed-forward network module. Depending on the TFLOP ratio requirement set by a particular experiment, we adjust the pruning hyperparameters of all baselines to match that requirement. However, some baselines do not support fine-grained adjustments like our approach does. In these cases, we choose the smallest available TFLOP ratio that exceeds the requirement set by an experiment, which might give these baselines a slight advantage over our method." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.654, + 0.484, + 0.745 + ], + "angle": 0, + "content": "We used \\(8 \\times\\) V100 GPUs with 32GB VRAM for all the experiments in this paper. Additionally, we used the Immsevalts package [51] for running these benchmarks for all the baselines and models. All results are obtained with a batch size of 1. For the metrics that require ChatGPT API access, the model is set to \"gpt-4o-mini\"." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.756, + 0.19, + 0.772 + ], + "angle": 0, + "content": "4.2. Insights" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.779, + 0.484, + 0.854 + ], + "angle": 0, + "content": "We provide visualizations comparing DivPrune with importance-based token pruning methods using LLaVA 1.5-7B and the SeedBench dataset [16]. Detailed analysis across different models and datasets is provided in the following subsections." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.856, + 0.484, + 0.901 + ], + "angle": 0, + "content": "The visual tokens in LLaVa 1.5 model are 4096-dimensional vectors. The t-SNE method [41] is utilized to project the visual tokens in \\(\\mathbf{E}_{\\mathbf{v}}\\) from a high dimensional" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.092, + 0.905, + 0.227 + ], + "angle": 0, + "content": "to a 2D space. The corresponding visualization for a sample input data is shown in Fig. 3-(a) using light Pruple points. Then, DivPrune is applied to select \\(10\\%\\) of the visual tokens (i.e., pruning \\(90\\%\\)). Additionally, FastV, as an importance-based token pruning method, which utilizes attention scores, is employed to prune with the same ratio. The selected subsets using DivPrune and FastV are shown with different markers in Fig. 3-(a). More examples are provided in the supplementary materials." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.228, + 0.906, + 0.378 + ], + "angle": 0, + "content": "As the example in Fig. 3-(a) shows, the proposed method selects points from all the clusters that appeared in the projected space whereas FastV does not choose any samples from the upper cluster. So, our method achieves a better representation of the original points by including samples from all clusters. In addition, the FastV method selects many tokens that are very close to each other which increases redundancy among the selected set. On the other hand, our method reduces redundancy by pruning the closely similar tokens." + }, + { + "type": "image", + "bbox": [ + 0.551, + 0.403, + 0.892, + 0.616 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.529, + 0.619, + 0.892, + 0.846 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.513, + 0.855, + 0.905, + 0.896 + ], + "angle": 0, + "content": "Figure 3. (a) t-SNE visualization of visual tokens for the original model, our method, and FastV. (b) Histogram of the Max-Min distance between the selected tokens over the SeedBench dataset." + }, + { + "type": "page_number", + "bbox": [ + 0.482, + 0.945, + 0.516, + 0.956 + ], + "angle": 0, + "content": "9396" + } + ], + [ + { + "type": "table", + "bbox": [ + 0.093, + 0.089, + 0.905, + 0.384 + ], + "angle": 0, + "content": "
MethodTFLOP (ratio %)COCO CIDErFlickr CIDErGQA EMMMB AccMME P-scoreMMMU AccNocaps CIDErOKVQA EMPOPE F1SQA EMSEEDB Acc
LLaVA 1.5.7BOriginal3.228 (100.00)1.100.7561.9664.09150636.441.0653.3985.8469.4166.17
VTW [21]0.603 (18.46)0.050.0338.9421.3168132.600.0318.6425.3565.2936.13
FastV [4]0.514 (15.69)0.060.0338.7320.6269632.000.0418.3232.8465.1535.69
Ours0.512 (15.63)0.960.6256.8559.19132835.890.9246.9886.0268.2759.47
PruMerge [36]Variable0.770.5051.3054.47125935.110.7341.7466.8968.9153.26
Ours*Variable0.910.5655.2558.16133035.440.8744.3883.0667.8757.88
FitPrune△ [47]0.513 (15.65)0.900.5652.3957.65119736.000.8642.5360.8968.0254.84
M3* [3]0.512 (15.63)1.000.6760.8165.81139131.800.9555.1286.3364.6564.93
PruMerge-LoRA*Variable0.960.6355.9659.88133434.890.9047.9977.1368.3257.93
LLaVA 1.5.13BOriginal6.281 (100.00)1.160.8063.3368.64152235.671.0958.2885.9972.8866.82
VTW [21]1.030 (16.16)0.080.0539.7121.9162232.100.0522.490.4066.2438.59
FastV [4]1.003 (15.73)0.380.1844.9837.8094235.110.3332.1430.0269.9644.95
Ours1.002 (15.71)1.000.6657.2963.40140734.890.9553.2983.4372.3462.04
PruMerge△ [36]Variable0.800.5352.0158.93125636.560.7749.1564.3672.5356.10
Ours*Variable0.940.5956.0961.77134434.890.9150.8679.6071.3460.00
LLaVA 1.6.7BOriginal11.849 (100.00)1.000.6864.2867.01152036.440.8844.2086.3870.1570.16
VTW [21]1.318 (11.23)0.060.0338.6219.7660631.300.038.667.1365.7437.48
FastV [4]1.327 (11.30)0.060.0338.7920.3661932.560.048.807.7865.4937.62
Ours1.266 (10.79)0.890.6158.6963.49136237.110.7641.9282.9768.5764.11
M3* [3]1.266 (10.79)1.010.6762.9769.16149035.000.8557.4987.4469.5168.49
" + }, + { + "type": "table_caption", + "bbox": [ + 0.09, + 0.389, + 0.907, + 0.418 + ], + "angle": 0, + "content": "Table 1. Comparison results of our method and different baselines on image-language understanding datasets. \\(\\bullet\\) : Finetuning is used, \\(\\triangle\\) Calibration dataset is used. Ours*: Our method matching the PruMerge selection ratio." + }, + { + "type": "text", + "bbox": [ + 0.089, + 0.437, + 0.482, + 0.603 + ], + "angle": 0, + "content": "In addition, the max-min distance (Eq. 3) for the selected subset of tokens is computed using 1000 randomly data samples from the SeedBench dataset and the histogram of the computed values is shown in Fig. 3-(b). As the plot indicates, the proposed method selects a subset where samples have a higher minimum pair-wise distance compared to the FastV method. Hence, our method achieves higher diversity among the selected tokens that have less redundancy compared to the ones chosen using FastV. We analyze the effect of the reduced diversity on task performance in the following sections." + }, + { + "type": "title", + "bbox": [ + 0.09, + 0.618, + 0.377, + 0.635 + ], + "angle": 0, + "content": "4.3. Image-Language Understanding" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.643, + 0.483, + 0.748 + ], + "angle": 0, + "content": "In this section, we compare DivPrune against baselines across various image-language understanding tasks, including open- and closed-ended QA, visual reasoning, and image captioning. Specifically, ScienceQA-IMG (SQA) [25], POPE [18], MME [9], MMB [24], GQA [14], MMMU [50], Flicker30k [31], SeedBench (SEEDB) [16], Nocaps [2], OKVQA [28], and COCO-2017 [20] are used." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.75, + 0.483, + 0.902 + ], + "angle": 0, + "content": "In the first experiment, summarized in Tab. 1, we analyze an extreme compression scenario for three image-based LMMs by fixing the TFLOP ratio at approximately \\(15\\%\\), wherever the baseline allows configuration to a fixed TFLOP ratio. Since PruMerge does not allow fixing the TFLOP ratio, we configure our approach (Ours*) to match the variable pruning corresponding to PruMerge for a fair comparison. In the top section of the table, we compare the results of various baselines for LLaVa 1.5-7B. Specifically, the baselines supporting LLaVA 1.5 are grouped into" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.437, + 0.908, + 0.786 + ], + "angle": 0, + "content": "three categories: plug-and-play methods, those with a variable TFLOP ratio, and those requiring a calibration dataset or involving fine-tuning the LMMs. Among the plug-and-play methods, which are the focus of this work, our approach significantly outperforms both the VTW and FastV baselines across all datasets. This result holds despite using lower TFLOPs, clearly demonstrating the advantage of our method in this scenario. For instance, when DivPrune is used, the performance of LLaVA 1.5-7b decreases by \\(5.1\\%\\) on the GQA dataset and \\(4.9\\%\\) on the MMB dataset. In contrast, the VTW and FastV methods result in performance drops of at least \\(23.0\\%\\) and \\(42.8\\%\\) on these datasets, respectively. The performance gap between DivPrune and the baseline methods is even more pronounced in image captioning tasks. For example, the CIDEr score on the COCO dataset drops by approximately \\(95\\%\\) with VTW and FastV, but only by \\(12.7\\%\\) with DivPrune. Additionally, DivPrune, compared to the original model, shows less than a \\(2\\%\\) performance drop on the MMMU and SQA datasets and slightly enhances the original model's performance on the POPE dataset while reducing the TLOP ratio by \\(84.4\\%\\). It is shown that removing redundant tokens in some datasets can improve the original model's performance [4]." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.796, + 0.909, + 0.903 + ], + "angle": 0, + "content": "Next, in the variable scenario, the pruning ratio is determined dynamically. To ensure a fair comparison, we matched the pruning ratio with that of the PruMerge baseline, assuming the average sequence length for calculating the average TFLOPs across each dataset. As indicated by the results, our approach consistently outperforms PruMerge across all benchmarks, except one. Further, for" + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.516, + 0.956 + ], + "angle": 0, + "content": "9397" + } + ], + [ + { + "type": "table", + "bbox": [ + 0.135, + 0.089, + 0.858, + 0.176 + ], + "angle": 0, + "content": "
TFLOPs (ratio %)ActivityNet Score/AccSeedBench AccVChatGPT ScoreNextQA WUPSEgoSch. AccMax GPU mem (GB)Prefill Time (sec)E2E Latency (sec)
Original6.539 (100)2.67 / 48.1038.72.1626.0541.814.060.3304.37
VTW [21]1.124 (16.97)1.61 / 26.8429.391.1918.6625.4213.630.1503.43
FastV [4]0.943 (14.20)1.95 / 33.9132.981.4422.5129.1413.570.1503.63
Ours0.937 (14.10)2.56 / 45.9037.001.9224.4839.7613.510.1613.39
" + }, + { + "type": "table_caption", + "bbox": [ + 0.108, + 0.18, + 0.888, + 0.194 + ], + "angle": 0, + "content": "Table 2. Comparison results of our method and baselines on LLaVA-NeXT-Video-7B across video-language understanding datasets." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.215, + 0.482, + 0.306 + ], + "angle": 0, + "content": "the baseline with calibration, we observe that our approach outperforms the FitPrune approach on nearly all datasets by up to \\(25.1\\%\\), despite not using any calibration dataset. Finally, compared to baselines involving fine-tuning, our method achieves comparable or superior performance without requiring any fine-tuning." + }, + { + "type": "text", + "bbox": [ + 0.089, + 0.307, + 0.483, + 0.504 + ], + "angle": 0, + "content": "The above experiment is repeated with LLaVa 1.5-13B model and the results are shown in the middle part of Tab. 1. The baselines that support this model are FastV, VTW, and PruMerge. As shown in the table, DivPrune outperforms the corresponding baselines in both plug-and-play and variable scenarios almost on all the tested datasets. For example, on the POPE dataset, DivPrune outperforms VTW, FastV, and PruMerge with F1 score improvements of \\(83\\%\\), \\(53.4\\%\\), and \\(15.2\\%\\), respectively. Additionally, on the MMB dataset, DivPrune achieves higher accuracy rates of \\(41.5\\%\\), \\(25.6\\%\\), and \\(2.8\\%\\) compared to VTW, FastV, and PruMerge, respectively. This demonstrates that DivPrune generalizes effectively across models with varying numbers of parameters." + }, + { + "type": "text", + "bbox": [ + 0.093, + 0.507, + 0.483, + 0.809 + ], + "angle": 0, + "content": "In the bottom part of Tab. 1, the results corresponding to LLaVA 1.6-7B model are shown. We used the same pruning ratio as for LLava 1.5. However, the lower TFLOP ratio is due to the large number of visual tokens in LLaVA 1.6. The results indicate that the performance of the model drops significantly when baseline pruning methods are applied. For example, the F1 score on the POPE dataset drops by \\(79\\%\\) with the baselines as compared to the original model, whereas the drop with DivPrune is only \\(3.4\\%\\). DivPrune also maintains competitive performance compared to the original model across various datasets. Specifically, DivPrune shows only \\(3.5\\%\\), \\(2.3\\%\\), \\(3.4\\%\\), \\(1.6\\%\\) drop in accuracy compared to the original model on the MMB, OKVQA, POPE, and SQA datasets, respectively, while reducing the TFLOP by \\(89\\%\\). The results also demonstrate that pruning visual tokens with DivPrune enhances the original model's performance on the MMMU task. These results show that DivPrune generalizes across different models. Qualitative examples as well as results with additional datasets are provided in the supplementary materials." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.811, + 0.483, + 0.902 + ], + "angle": 0, + "content": "Furthermore, we show the comparison of different baselines and our method across various TFLOP ratios. We plot the results in Fig. 1 where the y-axis represents average performance on four datasets, namely, COCO (CIDEr), OKVQA (Acc), POPE (F1), and MMBench (Acc). The range of the performance metric for all datasets is between" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.215, + 0.908, + 0.427 + ], + "angle": 0, + "content": "0 and 1, except for the CIDEr metric, which has a maximum reported value of 1.10. On the x-axis, we only show the high compression scenario (TFLOP ratio \\(\\leq 45\\%\\)). As shown in the figure, our method significantly outperforms all the baselines, particularly in high compression scenarios (TFLOP \\(\\leq 25\\%\\)). Further, we notice a steep drop in performance of all baselines as the TFLOP ratio \\(\\rightarrow 10\\), while our method falls more gracefully. This results in an increasing performance gap between our approach and the baselines at extreme compression levels. For higher TFLOP ratios almost all converge toward the original performance, with FitPrune slightly outperforming our approach by an insignificant margin. It is important to note that, unlike our method, FitPrune relies on a calibration dataset to prune tokens." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.438, + 0.796, + 0.455 + ], + "angle": 0, + "content": "4.4. Video-Language Understanding" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.462, + 0.907, + 0.794 + ], + "angle": 0, + "content": "In this section, LLaVA-NeXT-Video-7B [23], a video-based LMM is used to analyze the performance of the proposed method on various video-language understanding tasks. Specifically, we evaluate DivPrune using five datasets, namely, ActivityNet [48], SeedBench [16], VideoChatGPT (temporal) [26], NextQA [44], and EgoSchema [27]. FastV and VTW methods are chosen as the baselines. We tested DivPrune using the same pruning ratio as in the image understanding experiments. However, due to the higher number of visual tokens in the LLaVA-NeXT-Video model, this pruning ratio results in lower TFLOPs ratio. For the baselines, we match their TFLOPs with ours by selecting the smallest available TFLOP ratio that exceeds the TFLOPs of our method. The results for the original model, DivPrune, and the baselines are given in Tab. 2. As shown in the table, DivPrune outperforms both FastV and VTW by a significant margin. Specifically, DivPrune achieves upto \\(12\\%\\) higher accuracy than FastV and upto \\(19\\%\\) better than VTW on Video QA datasets including ActivityNet, SeedBench, and EgoSchema. DivPrune also outperforms both baselines on open-ended QA such as VideoChatGPT and NextQA by achieving higher GPT-assisted and WUPS scores." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.796, + 0.909, + 0.903 + ], + "angle": 0, + "content": "Furthermore, our method achieves performance that is highly competitive compared to the original model without pruning despite using only \\(14.1\\%\\) of the original model's TFLOPs. This demonstrates the robustness of DivPrune, as it effectively generalizes to video LMMs. Notably, the performance gap between DivPrune and the original model without pruning narrows as the number of visual tokens in" + }, + { + "type": "page_number", + "bbox": [ + 0.482, + 0.946, + 0.516, + 0.957 + ], + "angle": 0, + "content": "9398" + } + ], + [ + { + "type": "table", + "bbox": [ + 0.093, + 0.089, + 0.481, + 0.172 + ], + "angle": 0, + "content": "
TFLOP (ratio %)MMB AccMMMU AccPOPE F1SQA EMAvg
Layer 0 (Ours)19.6159.1935.8986.0268.2762.34
Layer 119.6559.0234.8980.6767.1860.44
Layer 219.7054.9034.2269.2769.5656.99
Layer 319.8023.9732.6731.8265.9438.60
" + }, + { + "type": "table_caption", + "bbox": [ + 0.094, + 0.178, + 0.48, + 0.192 + ], + "angle": 0, + "content": "Table 3. Ablation study on applying DivPrune at different layers." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.213, + 0.483, + 0.243 + ], + "angle": 0, + "content": "creases, indicating that DivPrune is more effective for the models with larger visual contexts." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.254, + 0.273, + 0.27 + ], + "angle": 0, + "content": "4.5. Efficiency Analysis" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.277, + 0.483, + 0.595 + ], + "angle": 0, + "content": "In this section, we analyze the efficiency of the proposed method using memory usage (i.e., max allocated memory), prefetch time, and end-to-end latency (E2E). For this experiment, VideoChatGPT dataset with 499 samples is used to obtain the average time and memory usage for LLaVA-NeXT-Video-7B model. The results are summarized on the right side of Tab. 2. The obtained results are compared against the original model, as well as the FastV and VTW baselines. As shown in the table, our approach requires approximately 400MB less memory than the original model, with memory usage comparable to the baselines. In terms of prefetch and E2E time, our approach is about \\(55\\%\\) and \\(22\\%\\) faster, respectively, compared to the original model. When compared to the baselines, our prefetch time is approximately \\(6 - 7\\%\\) longer, while the E2E time is \\(1 - 7\\%\\) shorter. The slight increase in prefetch time for our method compared to the baselines is due to the distance calculations (See Section 3.3), which are performed only once during the prefetch stage. In contrast, for baselines, the corresponding calculations for token pruning need to be done at each decoding step, resulting in longer E2E time." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.606, + 0.245, + 0.621 + ], + "angle": 0, + "content": "4.6. Ablation Study" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.629, + 0.483, + 0.855 + ], + "angle": 0, + "content": "In this section, we conduct an ablation study to analyze the impact of modifying various core components of our method. The ablation experiments are conducted with the LLaVA 1.5-7B model. First, we show the effect of pruning tokens inside the LLM in Tab. 3 using 5 datasets. By default, in our method, visual tokens are pruned before being passed to the first decoder layer in the LLM, which we refer to as 'Layer 0'. We also tested 'Layer 1' where the first layer is processed without pruning and the pruning is performed afterward. We further extended this approach by allowing tokens to pass through the first few layers unpruned and then pruning them after specific layers. As shown in the table, for a fixed TFLOP ratio of \\(19.61\\%\\), pruning done by our method at layer 0 achieves higher task accuracies compared to pruning at layers 1, 2, and 3 of the LLM." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.856, + 0.484, + 0.901 + ], + "angle": 0, + "content": "Furthermore, in Tab. 4, we provide an analysis of using alternative diversity measures for token pruning. The first three rows show the impact of choosing different dis" + }, + { + "type": "table", + "bbox": [ + 0.516, + 0.09, + 0.905, + 0.195 + ], + "angle": 0, + "content": "
TFLOP (ratio %)MMB AccMMMU AccPOPE F1SQA EMAvg
Cosine (Ours)19.6159.1935.8986.0268.2762.34
l119.6159.7134.6785.4067.9761.94
l219.6159.9735.0085.6468.2762.22
Random19.6152.6634.5672.7866.6356.66
Min-Max19.6138.5733.1149.2665.2046.53
" + }, + { + "type": "table_caption", + "bbox": [ + 0.547, + 0.199, + 0.871, + 0.214 + ], + "angle": 0, + "content": "Table 4. Ablation on using various diversity measures." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.237, + 0.905, + 0.327 + ], + "angle": 0, + "content": "tance measures to quantify the similarity among tokens. It can be seen that all three similarity measures, cosine, \\(\\ell_1\\), and \\(\\ell_2\\) perform comparably, with cosine (default setting) performing slightly better. This suggests that the choice of similarity measure does not significantly impact DivPrune's overall performance." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.329, + 0.908, + 0.646 + ], + "angle": 0, + "content": "The last two rows in Tab. 4 show the effect of choosing alternative strategies of token selection other than the proposed Max-Min diversity-based solution (3). We tested random pruning as well as the Min-Max strategy where the maximum distance between the selected samples is minimized. The Min-Max strategy enforces high redundancy among the selected samples, resulting in reduced diversity. As results in the bottom part of Tab. 4 reveal that any deviation from our proposed selection strategy results in suboptimal performance. Specifically, the Min-Max strategy performs the worst, showing approximately \\(15.8\\%\\) lower performance compared to ours. This decline is due to the Min-Max approach selecting tokens that are highly similar to each other, resulting in less diversity among the selected visual tokens. Random selection provides some degree of diversity, but it performs \\(5.6\\%\\) worse than the proposed method because it cannot guarantee maximum diversity. This proves that redundancy of visual tokens leads to poor performance and diversity maximization is needed for optimal performance, corroborating the utility and need of the proposed diversity maximization in Eq. (3)." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.663, + 0.634, + 0.679 + ], + "angle": 0, + "content": "5. Conclusion" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.689, + 0.907, + 0.901 + ], + "angle": 0, + "content": "In this paper, we proposed a token pruning method based on a max-min diversity problem, called DivPrune. In the proposed method, maximum diversity is achieved among the selected tokens, resulting in reduced redundancy. By ensuring high diversity, the selected tokens provide a more representative subset of the original tokens, enabling effective performance even at high pruning ratios without requiring fine-tuning. Extensive experiments were conducted with multiple LMMs on image and video understanding tasks across 16 datasets. The results show that DivPrune achieves state-of-the-art accuracy on the tested datasets. DivPrune generalizes well to different model sizes and architectures, while also improving memory consumption and end-to-end latency for the tested LMMs." + }, + { + "type": "page_number", + "bbox": [ + 0.482, + 0.945, + 0.516, + 0.957 + ], + "angle": 0, + "content": "9399" + } + ], + [ + { + "type": "title", + "bbox": [ + 0.093, + 0.09, + 0.188, + 0.106 + ], + "angle": 0, + "content": "References" + }, + { + "type": "ref_text", + "bbox": [ + 0.101, + 0.115, + 0.484, + 0.184 + ], + "angle": 0, + "content": "[1] Josh Achiam, Steven Adler, Sandhini Agarwal, Lama Ahmad, Ilge Akkaya, Florencia Leoni Aleman, Diogo Almeida, Janko Altenschmidt, Sam Altman, Shyamal Anadkat, et al. Gpt-4 technical report. arXiv preprint arXiv:2303.08774, 2023. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.101, + 0.187, + 0.484, + 0.257 + ], + "angle": 0, + "content": "[2] Harsh Agrawal, Karan Desai, Yufei Wang, Xinlei Chen, Rishabh Jain, Mark Johnson, Dhruv Batra, Devi Parikh, Stefan Lee, and Peter Anderson. nocaps: novel object captioning at scale. In Proceedings of the IEEE International Conference on Computer Vision, pages 8948-8957, 2019. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.259, + 0.483, + 0.313 + ], + "angle": 0, + "content": "[3] Mu Cai, Jianwei Yang, Jianfeng Gao, and Yong Jae Lee. Matryoshka multimodal models. Proceedings of the International Conference on Learning Representation, 2025. 2, 3, 4, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.316, + 0.483, + 0.384 + ], + "angle": 0, + "content": "[4] Liang Chen, Haozhe Zhao, Tianyu Liu, Shuai Bai, Junyang Lin, Chang Zhou, and Baobao Chang. An image is worth 1/2 tokens after layer 2: Plug-and-play inference acceleration for large vision-language models. In European Conference on Computer Vision (ECCV), 2024. 1, 2, 4, 5, 6, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.387, + 0.482, + 0.455 + ], + "angle": 0, + "content": "[5] Zesen Cheng, Sicong Leng, Hang Zhang, Yifei Xin, Xin Li, Guanzheng Chen, Yongxin Zhu, Wenqi Zhang, Ziyang Luo, Deli Zhao, et al. Videollama 2: Advancing spatial-temporal modeling and audio understanding in video-llms. arXiv preprint arXiv:2406.07476, 2024. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.458, + 0.482, + 0.54 + ], + "angle": 0, + "content": "[6] Wei-Lin Chiang, Zhuohan Li, Zi Lin, Ying Sheng, Zhanghao Wu, Hao Zhang, Lianmin Zheng, Siyuan Zhuang, Yonghao Zhuang, Joseph E Gonzalez, et al. Vicuna: An open-source chatbot impressing gpt-4 with \\(90\\%\\) * chatgpt quality. See https://vicuna.lmsys.org (accessed 14 April 2023), 2(3):6, 2023. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.543, + 0.483, + 0.611 + ], + "angle": 0, + "content": "[7] Krzysztof Choromanski, Valerii Likhosherstov, David Dohan, Xingyou Song, Andreea Gane, Tamas Sarlos, Peter Hawkins, Jared Davis, Afroz Mohiuddin, Lukasz Kaiser, et al. Rethinking attention with performers. arXiv preprint arXiv:2009.14794, 2020. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.614, + 0.483, + 0.669 + ], + "angle": 0, + "content": "[8] Tri Dao, Dan Fu, Stefano Ermon, Atri Rudra, and Christopher Ré. Flashattention: Fast and memory-efficient exact attention with io-awareness. Advances in Neural Information Processing Systems, 35:16344-16359, 2022. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.672, + 0.483, + 0.739 + ], + "angle": 0, + "content": "[9] Chaoyou Fu, Peixian Chen, Yunhang Shen, Yulei Qin, Mengdan Zhang, Xu Lin, Jinrui Yang, Xiawu Zheng, Ke Li, Xing Sun, et al. Mme: A comprehensive evaluation benchmark for multimodal large language models. arXiv preprint arXiv:2306.13394, 2023. 5, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.743, + 0.482, + 0.771 + ], + "angle": 0, + "content": "[10] Jinlan Fu, See-Kiong Ng, Zhengbao Jiang, and Pengfei Liu. Gptscore: Evaluate as you desire, 2023. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.773, + 0.483, + 0.827 + ], + "angle": 0, + "content": "[11] Mukul Gagrani, Raghavv Goel, Wonseok Jeon, Junyoung Park, Mingu Lee, and Christopher Lott. On speculative decoding for multimodal large language models. arXiv preprint arXiv:2404.08856, 2024. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.83, + 0.483, + 0.87 + ], + "angle": 0, + "content": "[12] Albert Gu and Tri Dao. Mamba: Linear-time sequence modeling with selective state spaces. arXiv preprint arXiv:2312.00752, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.873, + 0.483, + 0.901 + ], + "angle": 0, + "content": "[13] Zhiyu Guo, Hidetaka Kamigaito, and Taro Watanabe. Attention score is not all you need for token importance indicator" + }, + { + "type": "list", + "bbox": [ + 0.094, + 0.115, + 0.484, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.545, + 0.093, + 0.906, + 0.119 + ], + "angle": 0, + "content": "in kv cache reduction: Value also matters. arXiv preprint arXiv:2406.12335, 2024. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.122, + 0.906, + 0.176 + ], + "angle": 0, + "content": "[14] Drew A Hudson and Christopher D Manning. Gqa: A new dataset for real-world visual reasoning and compositional question answering. Conference on Computer Vision and Pattern Recognition (CVPR), 2019. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.178, + 0.905, + 0.244 + ], + "angle": 0, + "content": "[15] Angelos Katharopoulos, Apoorv Vyas, Nikolaos Pappas, and François Fleuret. Transformers are rnns: Fast autoregressive transformers with linear attention. In International conference on machine learning, pages 5156-5165. PMLR, 2020. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.247, + 0.905, + 0.3 + ], + "angle": 0, + "content": "[16] Bohao Li, Rui Wang, Guangzhi Wang, Yuying Ge, Yixiao Ge, and Ying Shan. Seed-bench: Benchmarking multimodal llms with generative comprehension. arXiv preprint arXiv:2307.16125, 2023. 5, 6, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.302, + 0.905, + 0.357 + ], + "angle": 0, + "content": "[17] Wentong Li, Yuqian Yuan, Jian Liu, Dongqi Tang, Song Wang, Jianke Zhu, and Lei Zhang. Tokenpacker: Efficient visual projector for multimodal llm. arXiv preprint arXiv:2407.02392, 2024. 2, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.359, + 0.905, + 0.412 + ], + "angle": 0, + "content": "[18] Yifan Li, Yifan Du, Kun Zhou, Jinpeng Wang, Wayne Xin Zhao, and Ji-Rong Wen. Evaluating object hallucination in large vision-language models. arXiv preprint arXiv:2305.10355, 2023. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.414, + 0.905, + 0.467 + ], + "angle": 0, + "content": "[19] Bin Lin, Yang Ye, Bin Zhu, Jiaxi Cui, Munan Ning, Peng Jin, and Li Yuan. Video-llava: Learning united visual representation by alignment before projection. CoRR, abs/2311.10122, 2023. 1, 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.47, + 0.905, + 0.524 + ], + "angle": 0, + "content": "[20] Tsung-Yi Lin, Michael Maire, Serge Belongie, Lubomir Bourdev, Ross Girshick, James Hays, Pietro Perona, Deva Ramanan, C. Lawrence Zitnick, and Piotr Dólar. Microsoft coco: Common objects in context, 2015. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.526, + 0.905, + 0.579 + ], + "angle": 0, + "content": "[21] Zhihang Lin, Mingbao Lin, Luxi Lin, and Rongrong Ji. Boosting multimodal large language models with visual tokens withdrawal for rapid inference. arXiv preprint arXiv:2405.05803, 2024. 2, 4, 5, 6, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.581, + 0.904, + 0.608 + ], + "angle": 0, + "content": "[22] Haotian Liu, Chunyuan Li, Qingyang Wu, and Yong Jae Lee. Visual instruction tuning. In NeurIPS, 2023. 1, 2, 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.61, + 0.905, + 0.663 + ], + "angle": 0, + "content": "[23] Haotian Liu, Chunyuan Li, Yuheng Li, Bo Li, Yuanhan Zhang, Sheng Shen, and Yong Jae Lee. Llava-last: Improved reasoning,OCR,and world knowledge,2024.1,2, 4,7" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.665, + 0.905, + 0.734 + ], + "angle": 0, + "content": "[24] Yuan Liu, Haodong Duan, Yuanhan Zhang, Bo Li, Songyang Zhang, Wangbo Zhao, Yike Yuan, Jiaqi Wang, Conghui He, Ziwei Liu, et al. Mmbench: Is your multi-modal model an all-around player? In European Conference on Computer Vision, pages 216-233. Springer, 2025. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.735, + 0.905, + 0.789 + ], + "angle": 0, + "content": "[25] Pan Lu, Swaroop Mishra, Tony Xia, Liang Qiu, Kai-Wei Chang, Song-Chun Zhu, Oyvind Tafjord, Peter Clark, and Ashwin Kalyan. Learn to explain: Multimodal reasoning via thought chains for science question answering, 2022. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.79, + 0.905, + 0.859 + ], + "angle": 0, + "content": "[26] Muhammad Maaz, Hanoona Rasheed, Salman Khan, and Fahad Shahbaz Khan. Video-chatgpt: Towards detailed video understanding via large vision and language models. In Proceedings of the 62nd Annual Meeting of the Association for Computational Linguistics (ACL 2024), 2024. 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.86, + 0.905, + 0.901 + ], + "angle": 0, + "content": "[27] Karttikeya Mangalam, Raiymbek Akshulakov, and Jitendra Malik. Egoschema: A diagnostic benchmark for very long-form video language understanding, 2023. 7" + }, + { + "type": "list", + "bbox": [ + 0.517, + 0.093, + 0.906, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.516, + 0.956 + ], + "angle": 0, + "content": "9400" + } + ], + [ + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.092, + 0.482, + 0.148 + ], + "angle": 0, + "content": "[28] Kenneth Marino, Mohammad Rastegari, Ali Farhadi, and Roozbeh Mottaghi. Ok-vqa: A visual question answering benchmark requiring external knowledge. In Conference on Computer Vision and Pattern Recognition (CVPR), 2019. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.151, + 0.482, + 0.207 + ], + "angle": 0, + "content": "[29] Rafael Martí, Anna Martínez-Gavara, Sergio Pérez-Peló, and Jesús Sánchez-Oro. A review on discrete diversity and dispersion maximization from an or perspective. European Journal of Operational Research, 299(3):795-813, 2022. 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.209, + 0.482, + 0.237 + ], + "angle": 0, + "content": "[30] OpenAI. Hello gpt-4o, 2024. https://openai.com/index/hello-gpt-4o/ [Accessed: (Nov 2024)]. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.239, + 0.482, + 0.307 + ], + "angle": 0, + "content": "[31] Bryan A. Plummer, Liwei Wang, Christopher M. Cervantes, Juan C. Caicedo, Julia Hockenmaier, and Svetlana Lazebnik. Flickr30k entities: Collecting region-to-phrase correspondences for richer image-to-sentence models. *IJCV*, 123 (1):74–93, 2017. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.31, + 0.482, + 0.365 + ], + "angle": 0, + "content": "[32] Daniel Cosmin Porumbel, Jin-Kao Hao, and Fred Glover. A simple and effective algorithm for the maxmin diversity problem. Annals of Operations Research, 186:275-293, 2011. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.368, + 0.482, + 0.424 + ], + "angle": 0, + "content": "[33] Yanyuan Qiao, Zheng Yu, Longteng Guo, Sihan Chen, Zijia Zhao, Mingzhen Sun, Qi Wu, and Jing Liu. Vl-mamba: Exploring state space models for multimodal learning. arXiv preprint arXiv:2403.13600, 2024. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.426, + 0.482, + 0.509 + ], + "angle": 0, + "content": "[34] Alec Radford, Jong Wook Kim, Chris Hallacy, Aditya Ramesh, Gabriel Goh, Sandhini Agarwal, Girish Sastry, Amanda Askell, Pamela Mishkin, Jack Clark, et al. Learning transferable visual models from natural language supervision. In International conference on machine learning, pages 8748-8763. PMLR, 2021. 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.511, + 0.482, + 0.567 + ], + "angle": 0, + "content": "[35] Mauricio GC Resende, Rafael Martí, Micael Gallego, and Abraham Duarte. Grasp and path relinking for the max-min diversity problem. Computers & Operations Research, 37 (3):498-508, 2010. 2, 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.569, + 0.482, + 0.625 + ], + "angle": 0, + "content": "[36] Yuzhang Shang, Mu Cai, Bingxin Xu, Yong Jae Lee, and Yan Yan. Llava-prumerge: Adaptive token reduction for efficient large multimodal models. arXiv preprint arXiv:2403.15388, 2024. 1, 2, 4, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.628, + 0.482, + 0.668 + ], + "angle": 0, + "content": "[37] Mustafa Shukor and Matthieu Cord. Skipping computations in multimodal llms. arXiv preprint arXiv:2410.09454, 2024. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.671, + 0.482, + 0.713 + ], + "angle": 0, + "content": "[38] Sainbayar Sukhbaatar, Edouard Grave, Piotr Bojanowski, and Armand Joulin. Adaptive attention span in transformers. arXiv preprint arXiv:1905.07799, 2019. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.716, + 0.482, + 0.785 + ], + "angle": 0, + "content": "[39] Gemini Team, Petko Georgiev, Ving Ian Lei, Ryan Burnell, Libin Bai, Anmol Gulati, Garrett Tanzer, Damien Vincent, Zhufeng Pan, Shibo Wang, et al. Gemini 1.5: Unlocking multimodal understanding across millions of tokens of context. arXiv preprint arXiv:2403.05530, 2024. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.787, + 0.482, + 0.857 + ], + "angle": 0, + "content": "[40] Hugo Touvron, Thibaut Lavril, Gautier Izacard, Xavier Martinet, Marie-Anne Lachaux, Timothee Lacroix, Baptiste Roziere, Naman Goyal, Eric Hambro, Faisal Azhar, et al. Llama: Open and efficient foundation language models. arXiv preprint arXiv:2302.13971, 2023. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.859, + 0.482, + 0.901 + ], + "angle": 0, + "content": "[41] Laurens Van der Maaten and Geoffrey Hinton. Visualizing data using t-sne. Journal of machine learning research, 9 (11), 2008. 5" + }, + { + "type": "list", + "bbox": [ + 0.093, + 0.092, + 0.482, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.516, + 0.092, + 0.906, + 0.148 + ], + "angle": 0, + "content": "[42] Ramakrishna Vedantam, C. Lawrence Zitnick, and Devi Parikh. Cider: Consensus-based image description evaluation. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition (CVPR), 2015. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.516, + 0.149, + 0.905, + 0.178 + ], + "angle": 0, + "content": "[43] Zhibiao Wu and Martha Palmer. Verb semantics and lexical selection. arXiv preprint cmp-lg/9406033, 1994. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.179, + 0.905, + 0.247 + ], + "angle": 0, + "content": "[44] Junbin Xiao, Xindi Shang, Angela Yao, and Tat-Seng Chua. Next-qa: Next phase of question-answering to explaining temporal actions. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 9777–9786, 2021. 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.249, + 0.905, + 0.303 + ], + "angle": 0, + "content": "[45] Shilin Xu, Xiangtai Li, Haobo Yuan, Lu Qi, Yunhai Tong, and Ming-Hsuan Yang. Llavadi: What matters for multimodal large language models distillation. arXiv preprint arXiv:2407.19409, 2024. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.305, + 0.905, + 0.36 + ], + "angle": 0, + "content": "[46] Yuan Yao, Tianyu Yu, Ao Zhang, Chongyi Wang, Junbo Cui, Hongji Zhu, Tianchi Cai, Haoyu Li, Weilin Zhao, Zhihui He, et al. Minicpm-v: A gpt-4v level mllm on your phone. arXiv preprint arXiv:2408.01800, 2024. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.362, + 0.905, + 0.417 + ], + "angle": 0, + "content": "[47] Weihao Ye, Qiong Wu, Wenhao Lin, and Yiyi Zhou. Fit and prune: Fast and training-free visual token pruning for multi-modal large language models. arXiv preprint arXiv:2409.10197, 2024. 1, 2, 4, 5, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.418, + 0.905, + 0.474 + ], + "angle": 0, + "content": "[48] Zhou Yu, Dejing Xu, Jun Yu, Ting Yu, Zhou Zhao, Yueting Zhuang, and Dacheng Tao. Activitynet-qa: A dataset for understanding complex web videos via question answering. In AAAI, pages 9127–9134, 2019. 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.476, + 0.905, + 0.53 + ], + "angle": 0, + "content": "[49] Zhengqing Yuan, Zhaoxu Li, Weiran Huang, Yanfang Ye, and Lichao Sun. Tinygpt-v: Efficient multimodal large language model via small backbones. arXiv preprint arXiv:2312.16862, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.531, + 0.905, + 0.643 + ], + "angle": 0, + "content": "[50] Xiang Yue, Yuansheng Ni, Kai Zhang, Tianyu Zheng, Ruoqi Liu, Ge Zhang, Samuel Stevens, Dongfu Jiang, Weiming Ren, Yuxuan Sun, Cong Wei, Botao Yu, Ruibin Yuan, Ren-liang Sun, Ming Yin, Boyuan Zheng, Zhenzhu Yang, Yibo Liu, Wenhao Huang, Huan Sun, Yu Su, and Wenhui Chen. Mmmu: A massive multi-discipline multimodal understanding and reasoning benchmark for expert agi. In Proceedings of CVPR, 2024. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.644, + 0.905, + 0.713 + ], + "angle": 0, + "content": "[51] Kaichen Zhang, Bo Li, Peiyuan Zhang, Fanyi Pu, Joshua Adrian Cahyono, Kairui Hu, Shuai Liu, Yuanhan Zhang, Jingkang Yang, Chunyuan Li, and Ziwei Liu. Lmmseval: Reality check on the evaluation of large multimodal models, 2024. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.714, + 0.905, + 0.769 + ], + "angle": 0, + "content": "[52] Yuanhan Zhang, Bo Li, haotian Liu, Yong jae Lee, Liangke Gui, Di Fu, Jiashi Feng, Ziwei Liu, and Chunyuan Li. Llavanext: A strong zero-shot video understanding model, 2024. 1, 2, 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.77, + 0.905, + 0.826 + ], + "angle": 0, + "content": "[53] Baichuan Zhou, Ying Hu, Xi Weng, Junlong Jia, Jie Luo, Xien Liu, Ji Wu, and Lei Huang. Tinyllava: A framework of small-scale large multimodal models. arXiv preprint arXiv:2402.14289, 2024. 2" + }, + { + "type": "list", + "bbox": [ + 0.516, + 0.092, + 0.906, + 0.826 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.514, + 0.956 + ], + "angle": 0, + "content": "9401" + } + ] +] \ No newline at end of file diff --git a/2025/DivPrune_ Diversity-based Visual Token Pruning for Large Multimodal Models/f2b05006-c955-46f7-a5af-8b7909376348_origin.pdf b/2025/DivPrune_ Diversity-based Visual Token Pruning for Large Multimodal Models/f2b05006-c955-46f7-a5af-8b7909376348_origin.pdf new file mode 100644 index 0000000000000000000000000000000000000000..2a4d83959f7cd8dde911f24b45c6367c590c068c --- /dev/null +++ b/2025/DivPrune_ Diversity-based Visual Token Pruning for Large Multimodal Models/f2b05006-c955-46f7-a5af-8b7909376348_origin.pdf @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:39b0edcb3b3e5c40c50ad070cfa463dd7bd68c3e13382afe73d952cae3c82533 +size 912680 diff --git a/2025/DivPrune_ Diversity-based Visual Token Pruning for Large Multimodal Models/full.md b/2025/DivPrune_ Diversity-based Visual Token Pruning for Large Multimodal Models/full.md new file mode 100644 index 0000000000000000000000000000000000000000..8a8ff0c687e3b24aba76677fee6bc4956dbb8ce6 --- /dev/null +++ b/2025/DivPrune_ Diversity-based Visual Token Pruning for Large Multimodal Models/full.md @@ -0,0 +1,314 @@ +# DivPrune: Diversity-based Visual Token Pruning for Large Multimodal Models + +Saeed Ranjbar Alvar, Gursimran Singh†, Mohammad Akbari†, Yong Zhang Huawei Technologies Canada Co., Ltd. + +{saeed.ranjbar.alvar1, gursimran.singh1, mohammad.akbari, yong.zhang3}@huawei.com + +# Abstract + +Large Multimodal Models (LMMs) have emerged as powerful models capable of understanding various data modalities, including text, images, and videos. LMMs encode both text and visual data into tokens that are then combined and processed by an integrated Large Language Model (LLM). Including visual tokens substantially increases the total token count, often by thousands. The increased input length for LLM significantly raises the complexity of inference, resulting in high latency in LMMs. To address this issue, token pruning methods, which remove part of the visual tokens, are proposed. The existing token pruning methods either require extensive calibration and fine-tuning or rely on suboptimal importance metrics which results in increased redundancy among the retained tokens. In this paper, we first formulate token pruning as Max-Min Diversity Problem (MMDP) where the goal is to select a subset such that the diversity among the selected tokens is maximized. Then, we solve the MMDP to obtain the selected subset and prune the rest. The proposed method, DivPrune, reduces redundancy and achieves the highest diversity of the selected tokens. By ensuring high diversity, the selected tokens better represent the original tokens, enabling effective performance even at high pruning ratios without requiring fine-tuning. Extensive experiments with various LMMs show that DivPrune achieves state-of-the-art accuracy over 16 image- and video-language datasets. Additionally, DivPrune reduces both the end-to-end latency and GPU memory usage for the tested models. The code is available here $^{\diamond}$ . + +# 1. Introduction + +Following the success of Large Language Models (LLMs) in language understanding [1, 6, 40], Large Multimodal Models (LMMs) [19, 22, 23, 52] have emerged to handle diverse data types like images and video, by leveraging the foundational capabilities of LLMs. Typically, LMMs encode text and visual modalities into tokens, also known + +![](images/8b69041e1c41479ec015cf24321f4a33a50484c7eefe19525afd2c762ed47aa7.jpg) +Figure 1. Comparison of different visual token pruning methods across various pruning ratios for LLaVA 1.5-7B. The y-axis is the performance averaged on COCO (CIDEr), OKVQA (Acc), POPE (F1), and MMBench (Acc). The x-axis is the TFLOP ratio of the model after token pruning compared to the original model before pruning. The proposed method significantly outperforms all baselines. Note that, unlike other methods, FitPrune uses an additional calibration step to prune tokens. + +as embeddings. These tokens are then combined and processed by an integrated LLM. The inclusion of visual tokens significantly increases the total number of tokens, often adding thousands to the combined set. Since the running time and memory requirements scale quadratically with input size [7, 8, 15, 38], the addition of visual tokens can substantially raise the running time for LMMs. Hence, many of these models often struggle to meet the demands of low-latency applications, particularly in resource-constrained environments [46]. + +Previous research [4, 36, 47] has demonstrated that there is a high degree of redundancy in the visual information processed by LMMs. As a result, visual token pruning has emerged as a promising solution to address the computational complexity challenges faced by LMMs. Specifically, previous research has demonstrated that reducing the number of visual tokens by $50\%$ [4] to $95\%$ [36] can significantly enhance the inference speed of LMMs. + +While promising, token pruning methods have certain + +shortcomings. For example, the works in [3, 17, 21, 47] require calibration or finetuning for each model which is costly and time-consuming to implement. FastV [4] and PruMerge [36] use attention scores to identify less important tokens for pruning. However, it is shown that using attention scores is not optimal, as some important tokens are overlooked [21]. Additionally, attention-based pruning tends to retain tokens that are similar to each other, leading to redundancy. At high compression ratio, this redundancy prevents the selection of a sufficient number of unique tokens to accurately represent the original tokens. In line with this observation, our findings indicate that pruning a large portion of visual tokens using these methods, without subsequent fine-tuning, results in a significant drop in the performance of LMMs across various tasks (Fig. 1). + +To address the above-mentioned issues, we formulate token pruning as a Max-Min Diversity Problem (MMDP) [35]. In an MMDP, the objective is to select a subset of elements such that the diversity among them is maximized. We apply this concept to token pruning, which we call DivPrune, aiming to maximize the diversity of the selected tokens by increasing the minimum distance between them. By ensuring high diversity, DivPrune captures a broader range of visual tokens, making it inherently more robust compared to attention-based methods that focus only on token importance scores. Increasing the diversity also helps ensure that the selected tokens better represent the original set of tokens, enabling effective performance even at high pruning ratios without the need for fine-tuning. + +DivPrune also offers practical advantages that make it a highly useful solution in real-world scenarios. DivPrune is a plug-and-play solution that can be used without requiring offline optimization with a calibration set, or fine-tuning of the model, which are often time-consuming and computationally expensive. DivPrune is applicable to LMMs with any LLM architecture and vision encoder. Additionally, DivPrune is compatible with inference optimization techniques, such as KV caching, resulting in practical speedup in real-world applications. In summary, our major contributions are as follows: + +- We introduce DivPrune, a token pruning method based on MMDP that maximizes diversity among visual tokens, effectively reducing redundancy and ensuring a highly representative subset. +- DivPrune is a training-free, calibration-data-free, plug-and-play solution that can be seamlessly integrated with off-the-shelf LMMs. +- We conduct evaluations using 16 datasets on image-and video-language models with image and video understanding tasks. DivPrune achieves state-of-the-art performance, with noticeable gains under extreme pruning (i.e., ratio $>80\%$ ). +DivPrune reduces GPU memory usage and inference la + +tency while maintaining comparable accuracy compared to the original model across most datasets. + +# 2. Related Works + +# 2.1. Large Multimodal Models (LMMs) + +LMMs handle diverse data types, including text, audio, image, and, video [5, 19, 22, 23, 30, 39, 52]. This work focuses on open-source LMMs that support language and visual inputs. These LMMs can be categorized into two types: image-based and video-based LMMs. The image-based LMMs [22, 23] address image-language understanding tasks, like image captioning, visual question answering, and image reasoning. On the other hand, video-based LMMs are geared towards video understanding [19, 52] tasks, like video captioning, video summarization, and video question answering. + +# 2.2. Efficient LMMs + +Several techniques are proposed to improve inference efficiency specifically for LMMs. The first technique is to change the model architecture in LMMs. For example, [33] proposed to replace transformer-based LLMs with Mamba model [12]. [49, 53] retrained LMMs with small scale LLMs to improve their efficiency. [45] used knowledge distillation to train a small LMM. In addition to changing the architecture, it is shown in [37] that skipping some blocks or layers within LMMs can improve the inference speed without damaging the model's performance. Furthermore, efficient decoding techniques such as speculative decoding are proposed to make LMM inference more efficient [11]. + +# 2.3. Visual Token Pruning + +Visual token pruning methods are proposed to reduce the inference complexity for LMMs. The first group of methods uses attention scores to prune tokens [4, 36]. PruMerge [36] introduces a token pruning method for the vision encoder where the visual tokens are clustered and merged based on their attention sparsity. In addition, FastV [4] prunes tokens within a specific layer of the LLM based on the magnitude of attention scores in an earlier layer. It is shown that pruning tokens based on attention scores are not optimal [13, 21], especially at higher pruning ratios. + +Calibration-based methods offer another line of work, where pruning layers and/or ratios are determined by analyzing the LLM outputs for a calibration dataset [21, 47]. For example, FitPrune [47] calculates a pruning recipe based on the observed attention divergence before and after pruning. VTW [21] argues that visual tokens can be entirely removed after a certain layer within LLM. The layer to remove the visual tokens is chosen using a calibration dataset. These methods rely on calibration datasets and require custom calibration for each LMM, which can be costly and + +![](images/bfb63d1030fc9da52cd07091a1af6939203095e7c0fd7c2ba1d42c1919a5da5f.jpg) +Figure 2. An overview of the LMM architecture, with DivPrune applied to visual tokens. The blocks on the right-hand side illustrate the steps of the method. + +cumbersome for new models. + +Some previous works proposed token pruning with the need for fine-tuning. $\mathbf{M}^3$ [3] applies model fine-tuning to produce nested visual token representations at multiple granularities, allowing users to select token lengths dynamically during inference. In [17], a projector layer trained using a large-scale dataset is proposed that packs finer detailed information into compact token representations. These methods need significant computational resources for training, limiting their use across various scenarios. + +# 3. Proposed Method + +In this section, we briefly discuss how LMMs work. Then, the token pruning problem is defined, followed by a detailed presentation of the proposed method. + +# 3.1. Large Multimodal Models (LMMs) + +An LMM typically processes a pair of inputs, denoted as $(T,V)$ , where $T$ is the text input and $V$ is the visual input such as image or video. The text input is mapped to $N$ textual tokens $\mathbf{E}_{\mathbf{t}} = \{t_1,\dots ,t_N\}$ using a text encoder. Similarly, the visual input is processed by a corresponding vision encoder. Specifically, it takes visual information $V$ as input and outputs image features, that are further converted to $M$ (generally $M\gg N$ ) vision tokens $\mathbf{E}_{\mathbf{v}} = \{v_{1},\ldots ,v_{M}\}$ using a projector layer (Fig. 2). + +The textual tokens and visual tokens are then combined to be fed to an LLM to generate the prediction in an autoregressive manner. Specifically, $\hat{N}$ output tokens $\mathbf{Y} = \{y_1,\dots ,y_{\hat{N}}\}$ are generated as follows: + +$$ +P \left(y _ {1}, \dots , y _ {\hat {N}} \mid \mathbf {E} _ {\mathbf {t}}, \mathbf {E} _ {\mathbf {v}}\right) = \prod_ {i = 1} ^ {\hat {N}} P \left(y _ {i} \mid y _ {< i}, \mathbf {E} _ {\mathbf {t}}, \mathbf {E} _ {\mathbf {v}}\right), \tag {1} +$$ + +where $P(|)$ is the conditional probability obtained at the + +output of the LLM. + +# 3.2. Token Pruning + +Reducing the number of input tokens in an integrated LLM within LMMs helps to lower memory usage and inference latency. Since visual tokens tend to have more redundancy, they are generally selected for pruning. + +In this context, the problem of token pruning can be defined as follows: given a set of visual tokens $\mathbf{E}_{\mathbf{v}}$ with $|\mathbf{E}_{\mathbf{v}}| = M$ and the subset size $\tilde{M} (\tilde{M} < M)$ , the goal is to select a subset, $\tilde{\mathbf{E}}_{\mathbf{v}}$ , while preserving key information necessary for accurate predictions. To mathematically formulate the token pruning problem, we define a mapping function $f$ , which maps the original set of visual tokens, $\mathbf{E}_{\mathbf{v}}$ , to a subset, $\tilde{\mathbf{E}}_{\mathbf{v}} = \{\tilde{v}_1,\dots ,\tilde{v}_{\tilde{M}}\}$ , where $|\tilde{\mathbf{E}}_{\mathbf{v}}| = \tilde{M}$ . The objective is to identify a mapping function $f$ that minimizes the difference in the model's output before and after pruning while ensuring the reduced set still captures the essential information from the original set: + +$$ +\text {F i n d :} \quad f: \mathbf {E} _ {v} \rightarrow \tilde {\mathbf {E}} _ {\mathbf {v}} +$$ + +$$ +\text {O b j e c t i v e :} \quad \min _ {f} \mathcal {L} (\mathcal {P}, \tilde {\mathcal {P}}) \tag {2} +$$ + +$$ +\text {S u b j e c t} \quad | \tilde {\mathbf {E}} _ {\mathbf {v}} | = \tilde {M}, +$$ + +where $\mathcal{P} = P(y_1, \ldots, y_{\hat{N}} \mid \mathbf{E}_{\mathbf{t}}, \mathbf{E}_{\mathbf{v}})$ and $\tilde{\mathcal{P}} = P(y_1, \ldots, y_{\hat{N}} \mid \mathbf{E}_{\mathbf{t}}, f(\mathbf{E}_{\mathbf{v}}))$ . Here, $\mathcal{L}$ represents a loss function that measures the difference in the model's output with and without pruning, and $\tilde{M}$ indicates the number of retained tokens. Next, we propose a novel diversity-based solution for the introduced token pruning problem. + +# 3.3. DivPrune: Method Overview + +We proposed a diversity-based token pruning method by reformulating the problem in (2) to select a subset of $\tilde{M}$ elements that maximizes the diversity, thereby reducing redundancy. Specifically, we define token pruning as Max-Min Diversity Problem (MMDP) [32] where the goal is to find the set $\tilde{\mathbf{E}}_{\mathbf{v}}$ among all possible sets with $\tilde{M}$ samples in $\mathbf{E}_{\mathbf{v}}$ that has the maximum minimum distance between its elements. So, MMDP is defined as: + +$$ +\operatorname {F i n d} \tilde {\mathbf {E}} _ {\mathbf {v}} = \arg \max \left[ \min _ {\gamma , \omega \in S} (d (\gamma , \omega)): \forall S \subset \mathbf {E} _ {\mathbf {v}} \right], \tag {3} +$$ + +where $S$ is an arbitrary set in $\mathbf{E}_{\mathbf{v}}$ with $\tilde{M}$ elements and $(\gamma, \omega)$ are arbitrary elements in $S$ . The distance is measured by $d(.,.)$ which is defined using the cosine distance as follows: + +$$ +d (\gamma , \omega) = 1 - \frac {\gamma \cdot \omega}{\| \gamma \| \| \omega \|}. \tag {4} +$$ + +A solution for the MMDP problem in (3) is a subset of $\mathbf{E}_{\mathbf{v}}$ that maximizes diversity by minimizing redundancy between elements. In the literature, several solutions including exact and heuristic methods are proposed to solve the + +Algorithm 1: Proposed Token Pruning Method +1 $\tilde{M}$ : subset size; $\mathbf{E}_{\mathbf{v}}$ : visual tokens; $\tilde{\mathbf{E}}_{\mathbf{v}}$ : selected subset +2 Initialize $\tilde{\mathbf{E}}_{\mathbf{v}} = []$ and $\mathbf{R} = \mathbf{E}_{\mathbf{v}}$ +3 // First stage: add the first token +4 $D = []$ initialize the distance array +5 for i in R do +6 $d_{min} = +inf$ +7 for j in R do +8 If $(i \neq j \& d(i, j) \leq d_{min})$ then $d_{min} = d(i, j)$ +9 Add $d_{min}$ to $D$ +10 $k = \mathbf{R}[\arg \max (\mathrm{D})]$ +11 move k from R to $\tilde{\mathbf{E}}_{\mathbf{v}}$ +12 // Second stage: iteratively add the subsequent tokens +13 while $|\tilde{\mathbf{E}}_{\mathbf{v}}| < \tilde{M}$ do +14 $D = []$ initialize the distance array +15 for i in R do +16 $d_{min} = +inf$ +17 for j in $\tilde{\mathbf{E}}_{\mathbf{v}}$ do +18 If $d(i, j) \leq d_{min}$ then $d_{min} = d(i, j)$ +19 Add $d_{min}$ to $D$ +20 $k = \mathbf{R}[\arg \max (\mathrm{D})]$ +21 move k from R to $\tilde{\mathbf{E}}_{\mathbf{v}}$ +22 Return $\tilde{\mathbf{E}}_{\mathbf{v}}$ + +MMDP problem [29, 35]. Since the number of tokens is generally limited (e.g., 576 in LLaVA 1.5 [22]) and the solvers are not generally designed for GPU acceleration, we obtain exact solution for the problem. Notably, the overhead of the selection process using GPU is negligible compared to the computations within the LLM. Detailed steps of the proposed method is summarized in Algorithm 1. Once the selected tokens are identified, the remaining visual tokens are discarded. The selected tokens along with the textual tokens are passed to the LLM. + +As shown in Algorithm 1, the proposed method has two stages after the initialization. The selected subset, $\tilde{\mathbf{E}}_{\mathbf{v}}$ , is initialized as empty, and the candidate list $\mathbf{R}$ is initialized with all the visual tokens. In the first stage, the first token of the selected subset is chosen based on the pairwise distance between the tokens of the candidate list. Then, the chosen token is moved from the candidate list to the selected list. In the second stage, similar to the first stage, the pairwise distance of the tokens in $\tilde{\mathbf{E}}_{\mathbf{v}}$ and the tokens in $\mathbf{R}$ is used to add samples to $\tilde{\mathbf{E}}_{\mathbf{v}}$ iteratively. Finally, once the number of tokens in $\tilde{\mathbf{E}}_{\mathbf{v}}$ reaches the specified subset size, the selection procedure is terminated and the $\tilde{\mathbf{E}}_{\mathbf{v}}$ is returned. To avoid repeated distance calculations over iterations a distance matrix is initially calculated by one matrix multiplication. + +The proposed method can also be applied to the features (i.e., hidden states) in the intermediate layers of the LLM. In this case, our method is not applied to the visual tokens, but + +to the features corresponding to the visual tokens obtained from a decoder layer to select a subset before feeding them to the subsequent layers. In either case, our method obtains the highest diversity for the selected elements. Ablation studies are provided in the next section to analyze the effect of pruning different elements at different layers. + +# 4. Experiments + +In this section, we present a comprehensive analysis comparing the performance of our method and previous works across various settings, tasks, and datasets. Insights into the proposed method are also provided through illustrative examples. Moreover, the efficiency of DivPrune along with ablation study are provided. + +# 4.1. Experimental Settings + +Baselines and Models: We consider five baselines, namely, FastV [4], PruMerge [36], VTW [21], FitPrune [47] and $\mathbf{M}^3$ [3]. Among these, we consider FastV, PruMerge, and VTW as our main competitors as they are plug-and-play and do not rely on any further costly finetuning or calibration process. However, for the sake of completeness, we also report performance comparison with respect to one finetuning-based $(\mathbf{M}^3)$ and one calibration-based (FitPrune) methods. Note that, VTW, by default, requires calibration to determine the best layer for a given task. However, doing that does not allow us to set a specific TFLOP ratio, complicating the comparison. Hence, whenever required we disable the calibration of VTW to select the layer that matches the FLOP requirement of a particular experiment. + +We test DivPrune and the baselines with popular LMMs namely LLaVA 1.5-7B [22]1, LLaVA 1.5-13B [22]2, LLaVA 1.6-7B3 (also known as LLaVA-NeXT [23]), and LLaVA-NeXT-Video-7B [52]4 to demonstrate the generality of DivPrune. For each tested model and task, we report only the relevant subset of baseline that is applicable to that specific model and task, alongside our results. + +All the tested LMMs used CLIP vision encoder [34]. LLaVA 1.5 model uses 576 visual tokens to represent images. LLaVA 1.6 converts each image into a varying number of patches, resulting in 3-5 times more visual tokens compared to LLaVA 1.5. LLaVA-NeXT-Video uses 144 tokens to process each frame. For all the experiments with LLaVA-NeXT-Video we used a total of 8 frames resulting in 1152 tokens for the processed frames. + +Datasets, Tasks, and Metrics: We selected a comprehensive set of common tasks and datasets aimed at multimodal reasoning and understanding. Specifically, we chose 11 image-language and 5 video-language datasets. + +1https://huggingface.co/liuhaotian/flava-v1.5-7B +$^{2}$ https://huggingface.co/luhaotian/llava-v1.5-13b +3https://huggingface.co/liuhaotian/llava-v1.6-vicuna-7b +4https://huggingface.co/lmms-lab/LLaVA-NeXT-Video-7B-DPO + +These datasets encompass a wide range of tasks, including captioning, multiple-choice Question Answering (QA), and open-ended QA based on text and image/video inputs. Consistent with prior works, CIDEr score [42] is used for evaluating captioning tasks, and Exact Match (EM), Accuracy (Acc), F1, Perception Score (P-score) [9] and GPT-assisted [10] score are used for QA tasks. Furthermore, Wu-Palmer similarity (WUPS) score [43] and GPT-assisted score [10] is used for open-ended QA. For all task performance metrics used in this paper, higher values indicate better performance. For the reported time and memory, lower values indicate better results. Further details regarding the datasets, tasks, and metrics are provided in the supplementary material. + +Following the earlier works in [4, 21, 47], we report the computational requirement, measured in TFLOPs, for DivPrune and the baselines. Various configurations including different pruning ratios at different layers are examined to obtain different working TFLOPs for our method and the baselines. The reported TFLOP ratio is the TFLOP of the model with pruned tokens relative to the original model's TFLOP with no pruning. This ratio is estimated as [4]: + +$$ +\frac {K \times \left(4 \mu d ^ {2} - 2 \mu^ {2} d + 2 \mu d m\right) + (T - K) \times \left(4 \tilde {\mu} d ^ {2} - 2 \tilde {\mu} ^ {2} d + 2 \tilde {\mu} d m\right)}{T \times \left(4 \mu d ^ {2} - 2 \mu^ {2} d + 2 \mu d m\right)}, \tag {5} +$$ + +where $T$ is the total transformer-based decoder layers. $\mu = N + M$ is the total sequence length before pruning, $\tilde{\mu} = N + \tilde{M}$ is the sequence length after pruning, $d$ is the hidden state size of the layer, and $m$ is the intermediate size of feed-forward network module. Depending on the TFLOP ratio requirement set by a particular experiment, we adjust the pruning hyperparameters of all baselines to match that requirement. However, some baselines do not support fine-grained adjustments like our approach does. In these cases, we choose the smallest available TFLOP ratio that exceeds the requirement set by an experiment, which might give these baselines a slight advantage over our method. + +We used $8 \times$ V100 GPUs with 32GB VRAM for all the experiments in this paper. Additionally, we used the Immsevalts package [51] for running these benchmarks for all the baselines and models. All results are obtained with a batch size of 1. For the metrics that require ChatGPT API access, the model is set to "gpt-4o-mini". + +# 4.2. Insights + +We provide visualizations comparing DivPrune with importance-based token pruning methods using LLaVA 1.5-7B and the SeedBench dataset [16]. Detailed analysis across different models and datasets is provided in the following subsections. + +The visual tokens in LLaVa 1.5 model are 4096-dimensional vectors. The t-SNE method [41] is utilized to project the visual tokens in $\mathbf{E}_{\mathbf{v}}$ from a high dimensional + +to a 2D space. The corresponding visualization for a sample input data is shown in Fig. 3-(a) using light Pruple points. Then, DivPrune is applied to select $10\%$ of the visual tokens (i.e., pruning $90\%$ ). Additionally, FastV, as an importance-based token pruning method, which utilizes attention scores, is employed to prune with the same ratio. The selected subsets using DivPrune and FastV are shown with different markers in Fig. 3-(a). More examples are provided in the supplementary materials. + +As the example in Fig. 3-(a) shows, the proposed method selects points from all the clusters that appeared in the projected space whereas FastV does not choose any samples from the upper cluster. So, our method achieves a better representation of the original points by including samples from all clusters. In addition, the FastV method selects many tokens that are very close to each other which increases redundancy among the selected set. On the other hand, our method reduces redundancy by pruning the closely similar tokens. + +![](images/bcd637e8a9777724d51d6a026867414a49f7df61248c1d29c52d870ab11a666c.jpg) + +![](images/e08094393d41b060f15c77ff4494e06b373fc39f3673c8a7cb3f436e59bcceeb.jpg) +Figure 3. (a) t-SNE visualization of visual tokens for the original model, our method, and FastV. (b) Histogram of the Max-Min distance between the selected tokens over the SeedBench dataset. + +
MethodTFLOP (ratio %)COCO CIDErFlickr CIDErGQA EMMMB AccMME P-scoreMMMU AccNocaps CIDErOKVQA EMPOPE F1SQA EMSEEDB Acc
LLaVA 1.5.7BOriginal3.228 (100.00)1.100.7561.9664.09150636.441.0653.3985.8469.4166.17
VTW [21]0.603 (18.46)0.050.0338.9421.3168132.600.0318.6425.3565.2936.13
FastV [4]0.514 (15.69)0.060.0338.7320.6269632.000.0418.3232.8465.1535.69
Ours0.512 (15.63)0.960.6256.8559.19132835.890.9246.9886.0268.2759.47
PruMerge [36]Variable0.770.5051.3054.47125935.110.7341.7466.8968.9153.26
Ours*Variable0.910.5655.2558.16133035.440.8744.3883.0667.8757.88
FitPrune△ [47]0.513 (15.65)0.900.5652.3957.65119736.000.8642.5360.8968.0254.84
M3* [3]0.512 (15.63)1.000.6760.8165.81139131.800.9555.1286.3364.6564.93
PruMerge-LoRA*Variable0.960.6355.9659.88133434.890.9047.9977.1368.3257.93
LLaVA 1.5.13BOriginal6.281 (100.00)1.160.8063.3368.64152235.671.0958.2885.9972.8866.82
VTW [21]1.030 (16.16)0.080.0539.7121.9162232.100.0522.490.4066.2438.59
FastV [4]1.003 (15.73)0.380.1844.9837.8094235.110.3332.1430.0269.9644.95
Ours1.002 (15.71)1.000.6657.2963.40140734.890.9553.2983.4372.3462.04
PruMerge△ [36]Variable0.800.5352.0158.93125636.560.7749.1564.3672.5356.10
Ours*Variable0.940.5956.0961.77134434.890.9150.8679.6071.3460.00
LLaVA 1.6.7BOriginal11.849 (100.00)1.000.6864.2867.01152036.440.8844.2086.3870.1570.16
VTW [21]1.318 (11.23)0.060.0338.6219.7660631.300.038.667.1365.7437.48
FastV [4]1.327 (11.30)0.060.0338.7920.3661932.560.048.807.7865.4937.62
Ours1.266 (10.79)0.890.6158.6963.49136237.110.7641.9282.9768.5764.11
M3* [3]1.266 (10.79)1.010.6762.9769.16149035.000.8557.4987.4469.5168.49
+ +Table 1. Comparison results of our method and different baselines on image-language understanding datasets. $\bullet$ : Finetuning is used, $\triangle$ Calibration dataset is used. Ours*: Our method matching the PruMerge selection ratio. + +In addition, the max-min distance (Eq. 3) for the selected subset of tokens is computed using 1000 randomly data samples from the SeedBench dataset and the histogram of the computed values is shown in Fig. 3-(b). As the plot indicates, the proposed method selects a subset where samples have a higher minimum pair-wise distance compared to the FastV method. Hence, our method achieves higher diversity among the selected tokens that have less redundancy compared to the ones chosen using FastV. We analyze the effect of the reduced diversity on task performance in the following sections. + +# 4.3. Image-Language Understanding + +In this section, we compare DivPrune against baselines across various image-language understanding tasks, including open- and closed-ended QA, visual reasoning, and image captioning. Specifically, ScienceQA-IMG (SQA) [25], POPE [18], MME [9], MMB [24], GQA [14], MMMU [50], Flicker30k [31], SeedBench (SEEDB) [16], Nocaps [2], OKVQA [28], and COCO-2017 [20] are used. + +In the first experiment, summarized in Tab. 1, we analyze an extreme compression scenario for three image-based LMMs by fixing the TFLOP ratio at approximately $15\%$ , wherever the baseline allows configuration to a fixed TFLOP ratio. Since PruMerge does not allow fixing the TFLOP ratio, we configure our approach (Ours*) to match the variable pruning corresponding to PruMerge for a fair comparison. In the top section of the table, we compare the results of various baselines for LLaVa 1.5-7B. Specifically, the baselines supporting LLaVA 1.5 are grouped into + +three categories: plug-and-play methods, those with a variable TFLOP ratio, and those requiring a calibration dataset or involving fine-tuning the LMMs. Among the plug-and-play methods, which are the focus of this work, our approach significantly outperforms both the VTW and FastV baselines across all datasets. This result holds despite using lower TFLOPs, clearly demonstrating the advantage of our method in this scenario. For instance, when DivPrune is used, the performance of LLaVA 1.5-7b decreases by $5.1\%$ on the GQA dataset and $4.9\%$ on the MMB dataset. In contrast, the VTW and FastV methods result in performance drops of at least $23.0\%$ and $42.8\%$ on these datasets, respectively. The performance gap between DivPrune and the baseline methods is even more pronounced in image captioning tasks. For example, the CIDEr score on the COCO dataset drops by approximately $95\%$ with VTW and FastV, but only by $12.7\%$ with DivPrune. Additionally, DivPrune, compared to the original model, shows less than a $2\%$ performance drop on the MMMU and SQA datasets and slightly enhances the original model's performance on the POPE dataset while reducing the TLOP ratio by $84.4\%$ . It is shown that removing redundant tokens in some datasets can improve the original model's performance [4]. + +Next, in the variable scenario, the pruning ratio is determined dynamically. To ensure a fair comparison, we matched the pruning ratio with that of the PruMerge baseline, assuming the average sequence length for calculating the average TFLOPs across each dataset. As indicated by the results, our approach consistently outperforms PruMerge across all benchmarks, except one. Further, for + +
TFLOPs (ratio %)ActivityNet Score/AccSeedBench AccVChatGPT ScoreNextQA WUPSEgoSch. AccMax GPU mem (GB)Prefill Time (sec)E2E Latency (sec)
Original6.539 (100)2.67 / 48.1038.72.1626.0541.814.060.3304.37
VTW [21]1.124 (16.97)1.61 / 26.8429.391.1918.6625.4213.630.1503.43
FastV [4]0.943 (14.20)1.95 / 33.9132.981.4422.5129.1413.570.1503.63
Ours0.937 (14.10)2.56 / 45.9037.001.9224.4839.7613.510.1613.39
+ +Table 2. Comparison results of our method and baselines on LLaVA-NeXT-Video-7B across video-language understanding datasets. + +the baseline with calibration, we observe that our approach outperforms the FitPrune approach on nearly all datasets by up to $25.1\%$ , despite not using any calibration dataset. Finally, compared to baselines involving fine-tuning, our method achieves comparable or superior performance without requiring any fine-tuning. + +The above experiment is repeated with LLaVa 1.5-13B model and the results are shown in the middle part of Tab. 1. The baselines that support this model are FastV, VTW, and PruMerge. As shown in the table, DivPrune outperforms the corresponding baselines in both plug-and-play and variable scenarios almost on all the tested datasets. For example, on the POPE dataset, DivPrune outperforms VTW, FastV, and PruMerge with F1 score improvements of $83\%$ , $53.4\%$ , and $15.2\%$ , respectively. Additionally, on the MMB dataset, DivPrune achieves higher accuracy rates of $41.5\%$ , $25.6\%$ , and $2.8\%$ compared to VTW, FastV, and PruMerge, respectively. This demonstrates that DivPrune generalizes effectively across models with varying numbers of parameters. + +In the bottom part of Tab. 1, the results corresponding to LLaVA 1.6-7B model are shown. We used the same pruning ratio as for LLava 1.5. However, the lower TFLOP ratio is due to the large number of visual tokens in LLaVA 1.6. The results indicate that the performance of the model drops significantly when baseline pruning methods are applied. For example, the F1 score on the POPE dataset drops by $79\%$ with the baselines as compared to the original model, whereas the drop with DivPrune is only $3.4\%$ . DivPrune also maintains competitive performance compared to the original model across various datasets. Specifically, DivPrune shows only $3.5\%$ , $2.3\%$ , $3.4\%$ , $1.6\%$ drop in accuracy compared to the original model on the MMB, OKVQA, POPE, and SQA datasets, respectively, while reducing the TFLOP by $89\%$ . The results also demonstrate that pruning visual tokens with DivPrune enhances the original model's performance on the MMMU task. These results show that DivPrune generalizes across different models. Qualitative examples as well as results with additional datasets are provided in the supplementary materials. + +Furthermore, we show the comparison of different baselines and our method across various TFLOP ratios. We plot the results in Fig. 1 where the y-axis represents average performance on four datasets, namely, COCO (CIDEr), OKVQA (Acc), POPE (F1), and MMBench (Acc). The range of the performance metric for all datasets is between + +0 and 1, except for the CIDEr metric, which has a maximum reported value of 1.10. On the x-axis, we only show the high compression scenario (TFLOP ratio $\leq 45\%$ ). As shown in the figure, our method significantly outperforms all the baselines, particularly in high compression scenarios (TFLOP $\leq 25\%$ ). Further, we notice a steep drop in performance of all baselines as the TFLOP ratio $\rightarrow 10$ , while our method falls more gracefully. This results in an increasing performance gap between our approach and the baselines at extreme compression levels. For higher TFLOP ratios almost all converge toward the original performance, with FitPrune slightly outperforming our approach by an insignificant margin. It is important to note that, unlike our method, FitPrune relies on a calibration dataset to prune tokens. + +# 4.4. Video-Language Understanding + +In this section, LLaVA-NeXT-Video-7B [23], a video-based LMM is used to analyze the performance of the proposed method on various video-language understanding tasks. Specifically, we evaluate DivPrune using five datasets, namely, ActivityNet [48], SeedBench [16], VideoChatGPT (temporal) [26], NextQA [44], and EgoSchema [27]. FastV and VTW methods are chosen as the baselines. We tested DivPrune using the same pruning ratio as in the image understanding experiments. However, due to the higher number of visual tokens in the LLaVA-NeXT-Video model, this pruning ratio results in lower TFLOPs ratio. For the baselines, we match their TFLOPs with ours by selecting the smallest available TFLOP ratio that exceeds the TFLOPs of our method. The results for the original model, DivPrune, and the baselines are given in Tab. 2. As shown in the table, DivPrune outperforms both FastV and VTW by a significant margin. Specifically, DivPrune achieves upto $12\%$ higher accuracy than FastV and upto $19\%$ better than VTW on Video QA datasets including ActivityNet, SeedBench, and EgoSchema. DivPrune also outperforms both baselines on open-ended QA such as VideoChatGPT and NextQA by achieving higher GPT-assisted and WUPS scores. + +Furthermore, our method achieves performance that is highly competitive compared to the original model without pruning despite using only $14.1\%$ of the original model's TFLOPs. This demonstrates the robustness of DivPrune, as it effectively generalizes to video LMMs. Notably, the performance gap between DivPrune and the original model without pruning narrows as the number of visual tokens in + +
TFLOP (ratio %)MMB AccMMMU AccPOPE F1SQA EMAvg
Layer 0 (Ours)19.6159.1935.8986.0268.2762.34
Layer 119.6559.0234.8980.6767.1860.44
Layer 219.7054.9034.2269.2769.5656.99
Layer 319.8023.9732.6731.8265.9438.60
+ +creases, indicating that DivPrune is more effective for the models with larger visual contexts. + +# 4.5. Efficiency Analysis + +In this section, we analyze the efficiency of the proposed method using memory usage (i.e., max allocated memory), prefetch time, and end-to-end latency (E2E). For this experiment, VideoChatGPT dataset with 499 samples is used to obtain the average time and memory usage for LLaVA-NeXT-Video-7B model. The results are summarized on the right side of Tab. 2. The obtained results are compared against the original model, as well as the FastV and VTW baselines. As shown in the table, our approach requires approximately 400MB less memory than the original model, with memory usage comparable to the baselines. In terms of prefetch and E2E time, our approach is about $55\%$ and $22\%$ faster, respectively, compared to the original model. When compared to the baselines, our prefetch time is approximately $6 - 7\%$ longer, while the E2E time is $1 - 7\%$ shorter. The slight increase in prefetch time for our method compared to the baselines is due to the distance calculations (See Section 3.3), which are performed only once during the prefetch stage. In contrast, for baselines, the corresponding calculations for token pruning need to be done at each decoding step, resulting in longer E2E time. + +# 4.6. Ablation Study + +In this section, we conduct an ablation study to analyze the impact of modifying various core components of our method. The ablation experiments are conducted with the LLaVA 1.5-7B model. First, we show the effect of pruning tokens inside the LLM in Tab. 3 using 5 datasets. By default, in our method, visual tokens are pruned before being passed to the first decoder layer in the LLM, which we refer to as 'Layer 0'. We also tested 'Layer 1' where the first layer is processed without pruning and the pruning is performed afterward. We further extended this approach by allowing tokens to pass through the first few layers unpruned and then pruning them after specific layers. As shown in the table, for a fixed TFLOP ratio of $19.61\%$ , pruning done by our method at layer 0 achieves higher task accuracies compared to pruning at layers 1, 2, and 3 of the LLM. + +Furthermore, in Tab. 4, we provide an analysis of using alternative diversity measures for token pruning. The first three rows show the impact of choosing different dis + +Table 3. Ablation study on applying DivPrune at different layers. + +
TFLOP (ratio %)MMB AccMMMU AccPOPE F1SQA EMAvg
Cosine (Ours)19.6159.1935.8986.0268.2762.34
l119.6159.7134.6785.4067.9761.94
l219.6159.9735.0085.6468.2762.22
Random19.6152.6634.5672.7866.6356.66
Min-Max19.6138.5733.1149.2665.2046.53
+ +Table 4. Ablation on using various diversity measures. + +tance measures to quantify the similarity among tokens. It can be seen that all three similarity measures, cosine, $\ell_1$ , and $\ell_2$ perform comparably, with cosine (default setting) performing slightly better. This suggests that the choice of similarity measure does not significantly impact DivPrune's overall performance. + +The last two rows in Tab. 4 show the effect of choosing alternative strategies of token selection other than the proposed Max-Min diversity-based solution (3). We tested random pruning as well as the Min-Max strategy where the maximum distance between the selected samples is minimized. The Min-Max strategy enforces high redundancy among the selected samples, resulting in reduced diversity. As results in the bottom part of Tab. 4 reveal that any deviation from our proposed selection strategy results in suboptimal performance. Specifically, the Min-Max strategy performs the worst, showing approximately $15.8\%$ lower performance compared to ours. This decline is due to the Min-Max approach selecting tokens that are highly similar to each other, resulting in less diversity among the selected visual tokens. Random selection provides some degree of diversity, but it performs $5.6\%$ worse than the proposed method because it cannot guarantee maximum diversity. This proves that redundancy of visual tokens leads to poor performance and diversity maximization is needed for optimal performance, corroborating the utility and need of the proposed diversity maximization in Eq. (3). + +# 5. Conclusion + +In this paper, we proposed a token pruning method based on a max-min diversity problem, called DivPrune. In the proposed method, maximum diversity is achieved among the selected tokens, resulting in reduced redundancy. By ensuring high diversity, the selected tokens provide a more representative subset of the original tokens, enabling effective performance even at high pruning ratios without requiring fine-tuning. Extensive experiments were conducted with multiple LMMs on image and video understanding tasks across 16 datasets. The results show that DivPrune achieves state-of-the-art accuracy on the tested datasets. DivPrune generalizes well to different model sizes and architectures, while also improving memory consumption and end-to-end latency for the tested LMMs. + +# References + +[1] Josh Achiam, Steven Adler, Sandhini Agarwal, Lama Ahmad, Ilge Akkaya, Florencia Leoni Aleman, Diogo Almeida, Janko Altenschmidt, Sam Altman, Shyamal Anadkat, et al. Gpt-4 technical report. arXiv preprint arXiv:2303.08774, 2023. 1 +[2] Harsh Agrawal, Karan Desai, Yufei Wang, Xinlei Chen, Rishabh Jain, Mark Johnson, Dhruv Batra, Devi Parikh, Stefan Lee, and Peter Anderson. nocaps: novel object captioning at scale. In Proceedings of the IEEE International Conference on Computer Vision, pages 8948-8957, 2019. 6 +[3] Mu Cai, Jianwei Yang, Jianfeng Gao, and Yong Jae Lee. Matryoshka multimodal models. Proceedings of the International Conference on Learning Representation, 2025. 2, 3, 4, 6 +[4] Liang Chen, Haozhe Zhao, Tianyu Liu, Shuai Bai, Junyang Lin, Chang Zhou, and Baobao Chang. An image is worth 1/2 tokens after layer 2: Plug-and-play inference acceleration for large vision-language models. In European Conference on Computer Vision (ECCV), 2024. 1, 2, 4, 5, 6, 7 +[5] Zesen Cheng, Sicong Leng, Hang Zhang, Yifei Xin, Xin Li, Guanzheng Chen, Yongxin Zhu, Wenqi Zhang, Ziyang Luo, Deli Zhao, et al. Videollama 2: Advancing spatial-temporal modeling and audio understanding in video-llms. arXiv preprint arXiv:2406.07476, 2024. 2 +[6] Wei-Lin Chiang, Zhuohan Li, Zi Lin, Ying Sheng, Zhanghao Wu, Hao Zhang, Lianmin Zheng, Siyuan Zhuang, Yonghao Zhuang, Joseph E Gonzalez, et al. Vicuna: An open-source chatbot impressing gpt-4 with $90\%$ * chatgpt quality. See https://vicuna.lmsys.org (accessed 14 April 2023), 2(3):6, 2023. 1 +[7] Krzysztof Choromanski, Valerii Likhosherstov, David Dohan, Xingyou Song, Andreea Gane, Tamas Sarlos, Peter Hawkins, Jared Davis, Afroz Mohiuddin, Lukasz Kaiser, et al. Rethinking attention with performers. arXiv preprint arXiv:2009.14794, 2020. 1 +[8] Tri Dao, Dan Fu, Stefano Ermon, Atri Rudra, and Christopher Ré. Flashattention: Fast and memory-efficient exact attention with io-awareness. Advances in Neural Information Processing Systems, 35:16344-16359, 2022. 1 +[9] Chaoyou Fu, Peixian Chen, Yunhang Shen, Yulei Qin, Mengdan Zhang, Xu Lin, Jinrui Yang, Xiawu Zheng, Ke Li, Xing Sun, et al. Mme: A comprehensive evaluation benchmark for multimodal large language models. arXiv preprint arXiv:2306.13394, 2023. 5, 6 +[10] Jinlan Fu, See-Kiong Ng, Zhengbao Jiang, and Pengfei Liu. Gptscore: Evaluate as you desire, 2023. 5 +[11] Mukul Gagrani, Raghavv Goel, Wonseok Jeon, Junyoung Park, Mingu Lee, and Christopher Lott. On speculative decoding for multimodal large language models. arXiv preprint arXiv:2404.08856, 2024. 2 +[12] Albert Gu and Tri Dao. Mamba: Linear-time sequence modeling with selective state spaces. arXiv preprint arXiv:2312.00752, 2023. 2 +[13] Zhiyu Guo, Hidetaka Kamigaito, and Taro Watanabe. Attention score is not all you need for token importance indicator + +in kv cache reduction: Value also matters. arXiv preprint arXiv:2406.12335, 2024. 2 +[14] Drew A Hudson and Christopher D Manning. Gqa: A new dataset for real-world visual reasoning and compositional question answering. Conference on Computer Vision and Pattern Recognition (CVPR), 2019. 6 +[15] Angelos Katharopoulos, Apoorv Vyas, Nikolaos Pappas, and François Fleuret. Transformers are rnns: Fast autoregressive transformers with linear attention. In International conference on machine learning, pages 5156-5165. PMLR, 2020. 1 +[16] Bohao Li, Rui Wang, Guangzhi Wang, Yuying Ge, Yixiao Ge, and Ying Shan. Seed-bench: Benchmarking multimodal llms with generative comprehension. arXiv preprint arXiv:2307.16125, 2023. 5, 6, 7 +[17] Wentong Li, Yuqian Yuan, Jian Liu, Dongqi Tang, Song Wang, Jianke Zhu, and Lei Zhang. Tokenpacker: Efficient visual projector for multimodal llm. arXiv preprint arXiv:2407.02392, 2024. 2, 3 +[18] Yifan Li, Yifan Du, Kun Zhou, Jinpeng Wang, Wayne Xin Zhao, and Ji-Rong Wen. Evaluating object hallucination in large vision-language models. arXiv preprint arXiv:2305.10355, 2023. 6 +[19] Bin Lin, Yang Ye, Bin Zhu, Jiaxi Cui, Munan Ning, Peng Jin, and Li Yuan. Video-llava: Learning united visual representation by alignment before projection. CoRR, abs/2311.10122, 2023. 1, 2 +[20] Tsung-Yi Lin, Michael Maire, Serge Belongie, Lubomir Bourdev, Ross Girshick, James Hays, Pietro Perona, Deva Ramanan, C. Lawrence Zitnick, and Piotr Dólar. Microsoft coco: Common objects in context, 2015. 6 +[21] Zhihang Lin, Mingbao Lin, Luxi Lin, and Rongrong Ji. Boosting multimodal large language models with visual tokens withdrawal for rapid inference. arXiv preprint arXiv:2405.05803, 2024. 2, 4, 5, 6, 7 +[22] Haotian Liu, Chunyuan Li, Qingyang Wu, and Yong Jae Lee. Visual instruction tuning. In NeurIPS, 2023. 1, 2, 4 +[23] Haotian Liu, Chunyuan Li, Yuheng Li, Bo Li, Yuanhan Zhang, Sheng Shen, and Yong Jae Lee. Llava-last: Improved reasoning,OCR,and world knowledge,2024.1,2, 4,7 +[24] Yuan Liu, Haodong Duan, Yuanhan Zhang, Bo Li, Songyang Zhang, Wangbo Zhao, Yike Yuan, Jiaqi Wang, Conghui He, Ziwei Liu, et al. Mmbench: Is your multi-modal model an all-around player? In European Conference on Computer Vision, pages 216-233. Springer, 2025. 6 +[25] Pan Lu, Swaroop Mishra, Tony Xia, Liang Qiu, Kai-Wei Chang, Song-Chun Zhu, Oyvind Tafjord, Peter Clark, and Ashwin Kalyan. Learn to explain: Multimodal reasoning via thought chains for science question answering, 2022. 6 +[26] Muhammad Maaz, Hanoona Rasheed, Salman Khan, and Fahad Shahbaz Khan. Video-chatgpt: Towards detailed video understanding via large vision and language models. In Proceedings of the 62nd Annual Meeting of the Association for Computational Linguistics (ACL 2024), 2024. 7 +[27] Karttikeya Mangalam, Raiymbek Akshulakov, and Jitendra Malik. Egoschema: A diagnostic benchmark for very long-form video language understanding, 2023. 7 + +[28] Kenneth Marino, Mohammad Rastegari, Ali Farhadi, and Roozbeh Mottaghi. Ok-vqa: A visual question answering benchmark requiring external knowledge. In Conference on Computer Vision and Pattern Recognition (CVPR), 2019. 6 +[29] Rafael Martí, Anna Martínez-Gavara, Sergio Pérez-Peló, and Jesús Sánchez-Oro. A review on discrete diversity and dispersion maximization from an or perspective. European Journal of Operational Research, 299(3):795-813, 2022. 4 +[30] OpenAI. Hello gpt-4o, 2024. https://openai.com/index/hello-gpt-4o/ [Accessed: (Nov 2024)]. 2 +[31] Bryan A. Plummer, Liwei Wang, Christopher M. Cervantes, Juan C. Caicedo, Julia Hockenmaier, and Svetlana Lazebnik. Flickr30k entities: Collecting region-to-phrase correspondences for richer image-to-sentence models. *IJCV*, 123 (1):74–93, 2017. 6 +[32] Daniel Cosmin Porumbel, Jin-Kao Hao, and Fred Glover. A simple and effective algorithm for the maxmin diversity problem. Annals of Operations Research, 186:275-293, 2011. 3 +[33] Yanyuan Qiao, Zheng Yu, Longteng Guo, Sihan Chen, Zijia Zhao, Mingzhen Sun, Qi Wu, and Jing Liu. Vl-mamba: Exploring state space models for multimodal learning. arXiv preprint arXiv:2403.13600, 2024. 2 +[34] Alec Radford, Jong Wook Kim, Chris Hallacy, Aditya Ramesh, Gabriel Goh, Sandhini Agarwal, Girish Sastry, Amanda Askell, Pamela Mishkin, Jack Clark, et al. Learning transferable visual models from natural language supervision. In International conference on machine learning, pages 8748-8763. PMLR, 2021. 4 +[35] Mauricio GC Resende, Rafael Martí, Micael Gallego, and Abraham Duarte. Grasp and path relinking for the max-min diversity problem. Computers & Operations Research, 37 (3):498-508, 2010. 2, 4 +[36] Yuzhang Shang, Mu Cai, Bingxin Xu, Yong Jae Lee, and Yan Yan. Llava-prumerge: Adaptive token reduction for efficient large multimodal models. arXiv preprint arXiv:2403.15388, 2024. 1, 2, 4, 6 +[37] Mustafa Shukor and Matthieu Cord. Skipping computations in multimodal llms. arXiv preprint arXiv:2410.09454, 2024. 2 +[38] Sainbayar Sukhbaatar, Edouard Grave, Piotr Bojanowski, and Armand Joulin. Adaptive attention span in transformers. arXiv preprint arXiv:1905.07799, 2019. 1 +[39] Gemini Team, Petko Georgiev, Ving Ian Lei, Ryan Burnell, Libin Bai, Anmol Gulati, Garrett Tanzer, Damien Vincent, Zhufeng Pan, Shibo Wang, et al. Gemini 1.5: Unlocking multimodal understanding across millions of tokens of context. arXiv preprint arXiv:2403.05530, 2024. 2 +[40] Hugo Touvron, Thibaut Lavril, Gautier Izacard, Xavier Martinet, Marie-Anne Lachaux, Timothee Lacroix, Baptiste Roziere, Naman Goyal, Eric Hambro, Faisal Azhar, et al. Llama: Open and efficient foundation language models. arXiv preprint arXiv:2302.13971, 2023. 1 +[41] Laurens Van der Maaten and Geoffrey Hinton. Visualizing data using t-sne. Journal of machine learning research, 9 (11), 2008. 5 + +[42] Ramakrishna Vedantam, C. Lawrence Zitnick, and Devi Parikh. Cider: Consensus-based image description evaluation. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition (CVPR), 2015. 5 +[43] Zhibiao Wu and Martha Palmer. Verb semantics and lexical selection. arXiv preprint cmp-lg/9406033, 1994. 5 +[44] Junbin Xiao, Xindi Shang, Angela Yao, and Tat-Seng Chua. Next-qa: Next phase of question-answering to explaining temporal actions. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 9777–9786, 2021. 7 +[45] Shilin Xu, Xiangtai Li, Haobo Yuan, Lu Qi, Yunhai Tong, and Ming-Hsuan Yang. Llavadi: What matters for multimodal large language models distillation. arXiv preprint arXiv:2407.19409, 2024. 2 +[46] Yuan Yao, Tianyu Yu, Ao Zhang, Chongyi Wang, Junbo Cui, Hongji Zhu, Tianchi Cai, Haoyu Li, Weilin Zhao, Zhihui He, et al. Minicpm-v: A gpt-4v level mllm on your phone. arXiv preprint arXiv:2408.01800, 2024. 1 +[47] Weihao Ye, Qiong Wu, Wenhao Lin, and Yiyi Zhou. Fit and prune: Fast and training-free visual token pruning for multi-modal large language models. arXiv preprint arXiv:2409.10197, 2024. 1, 2, 4, 5, 6 +[48] Zhou Yu, Dejing Xu, Jun Yu, Ting Yu, Zhou Zhao, Yueting Zhuang, and Dacheng Tao. Activitynet-qa: A dataset for understanding complex web videos via question answering. In AAAI, pages 9127–9134, 2019. 7 +[49] Zhengqing Yuan, Zhaoxu Li, Weiran Huang, Yanfang Ye, and Lichao Sun. Tinygpt-v: Efficient multimodal large language model via small backbones. arXiv preprint arXiv:2312.16862, 2023. 2 +[50] Xiang Yue, Yuansheng Ni, Kai Zhang, Tianyu Zheng, Ruoqi Liu, Ge Zhang, Samuel Stevens, Dongfu Jiang, Weiming Ren, Yuxuan Sun, Cong Wei, Botao Yu, Ruibin Yuan, Ren-liang Sun, Ming Yin, Boyuan Zheng, Zhenzhu Yang, Yibo Liu, Wenhao Huang, Huan Sun, Yu Su, and Wenhui Chen. Mmmu: A massive multi-discipline multimodal understanding and reasoning benchmark for expert agi. In Proceedings of CVPR, 2024. 6 +[51] Kaichen Zhang, Bo Li, Peiyuan Zhang, Fanyi Pu, Joshua Adrian Cahyono, Kairui Hu, Shuai Liu, Yuanhan Zhang, Jingkang Yang, Chunyuan Li, and Ziwei Liu. Lmmseval: Reality check on the evaluation of large multimodal models, 2024. 5 +[52] Yuanhan Zhang, Bo Li, haotian Liu, Yong jae Lee, Liangke Gui, Di Fu, Jiashi Feng, Ziwei Liu, and Chunyuan Li. Llavanext: A strong zero-shot video understanding model, 2024. 1, 2, 4 +[53] Baichuan Zhou, Ying Hu, Xi Weng, Junlong Jia, Jie Luo, Xien Liu, Ji Wu, and Lei Huang. Tinyllava: A framework of small-scale large multimodal models. arXiv preprint arXiv:2402.14289, 2024. 2 \ No newline at end of file diff --git a/2025/DivPrune_ Diversity-based Visual Token Pruning for Large Multimodal Models/images.zip b/2025/DivPrune_ Diversity-based Visual Token Pruning for Large Multimodal Models/images.zip new file mode 100644 index 0000000000000000000000000000000000000000..481a178f5cd61fd3069ce1bb7d58e69c6690f55a --- /dev/null +++ b/2025/DivPrune_ Diversity-based Visual Token Pruning for Large Multimodal Models/images.zip @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:f2974e900b3dd9ca6f9d8d75a8aa536ebee95f20eb84e4299a7faf6b0c99bfef +size 404474 diff --git a/2025/DivPrune_ Diversity-based Visual Token Pruning for Large Multimodal Models/layout.json b/2025/DivPrune_ Diversity-based Visual Token Pruning for Large Multimodal Models/layout.json new file mode 100644 index 0000000000000000000000000000000000000000..9979a7620282918f1cbf4002d9b01a039fa5ec01 --- /dev/null +++ b/2025/DivPrune_ Diversity-based Visual Token Pruning for Large Multimodal Models/layout.json @@ -0,0 +1,8882 @@ +{ + "pdf_info": [ + { + "para_blocks": [ + { + "bbox": [ + 60, + 103, + 550, + 121 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 60, + 103, + 550, + 121 + ], + "spans": [ + { + "bbox": [ + 60, + 103, + 550, + 121 + ], + "type": "text", + "content": "DivPrune: Diversity-based Visual Token Pruning for Large Multimodal Models" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 123, + 142, + 487, + 171 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 123, + 142, + 487, + 171 + ], + "spans": [ + { + "bbox": [ + 123, + 142, + 487, + 171 + ], + "type": "text", + "content": "Saeed Ranjbar Alvar, Gursimran Singh†, Mohammad Akbari†, Yong Zhang Huawei Technologies Canada Co., Ltd." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 85, + 173, + 523, + 186 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 85, + 173, + 523, + 186 + ], + "spans": [ + { + "bbox": [ + 85, + 173, + 523, + 186 + ], + "type": "text", + "content": "{saeed.ranjbar.alvar1, gursimran.singh1, mohammad.akbari, yong.zhang3}@huawei.com" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 151, + 213, + 200, + 225 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 151, + 213, + 200, + 225 + ], + "spans": [ + { + "bbox": [ + 151, + 213, + 200, + 225 + ], + "type": "text", + "content": "Abstract" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 56, + 239, + 296, + 574 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 239, + 296, + 574 + ], + "spans": [ + { + "bbox": [ + 56, + 239, + 296, + 574 + ], + "type": "text", + "content": "Large Multimodal Models (LMMs) have emerged as powerful models capable of understanding various data modalities, including text, images, and videos. LMMs encode both text and visual data into tokens that are then combined and processed by an integrated Large Language Model (LLM). Including visual tokens substantially increases the total token count, often by thousands. The increased input length for LLM significantly raises the complexity of inference, resulting in high latency in LMMs. To address this issue, token pruning methods, which remove part of the visual tokens, are proposed. The existing token pruning methods either require extensive calibration and fine-tuning or rely on suboptimal importance metrics which results in increased redundancy among the retained tokens. In this paper, we first formulate token pruning as Max-Min Diversity Problem (MMDP) where the goal is to select a subset such that the diversity among the selected tokens is maximized. Then, we solve the MMDP to obtain the selected subset and prune the rest. The proposed method, DivPrune, reduces redundancy and achieves the highest diversity of the selected tokens. By ensuring high diversity, the selected tokens better represent the original tokens, enabling effective performance even at high pruning ratios without requiring fine-tuning. Extensive experiments with various LMMs show that DivPrune achieves state-of-the-art accuracy over 16 image- and video-language datasets. Additionally, DivPrune reduces both the end-to-end latency and GPU memory usage for the tested models. The code is available here" + }, + { + "bbox": [ + 56, + 239, + 296, + 574 + ], + "type": "inline_equation", + "content": "^{\\diamond}" + }, + { + "bbox": [ + 56, + 239, + 296, + 574 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 56, + 594, + 135, + 605 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 594, + 135, + 605 + ], + "spans": [ + { + "bbox": [ + 56, + 594, + 135, + 605 + ], + "type": "text", + "content": "1. Introduction" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 55, + 613, + 295, + 685 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 613, + 295, + 685 + ], + "spans": [ + { + "bbox": [ + 55, + 613, + 295, + 685 + ], + "type": "text", + "content": "Following the success of Large Language Models (LLMs) in language understanding [1, 6, 40], Large Multimodal Models (LMMs) [19, 22, 23, 52] have emerged to handle diverse data types like images and video, by leveraging the foundational capabilities of LLMs. Typically, LMMs encode text and visual modalities into tokens, also known" + } + ] + } + ], + "index": 8 + }, + { + "type": "image", + "bbox": [ + 333, + 218, + 538, + 366 + ], + "blocks": [ + { + "bbox": [ + 333, + 218, + 538, + 366 + ], + "lines": [ + { + "bbox": [ + 333, + 218, + 538, + 366 + ], + "spans": [ + { + "bbox": [ + 333, + 218, + 538, + 366 + ], + "type": "image", + "image_path": "8b69041e1c41479ec015cf24321f4a33a50484c7eefe19525afd2c762ed47aa7.jpg" + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 313, + 376, + 555, + 465 + ], + "lines": [ + { + "bbox": [ + 313, + 376, + 555, + 465 + ], + "spans": [ + { + "bbox": [ + 313, + 376, + 555, + 465 + ], + "type": "text", + "content": "Figure 1. Comparison of different visual token pruning methods across various pruning ratios for LLaVA 1.5-7B. The y-axis is the performance averaged on COCO (CIDEr), OKVQA (Acc), POPE (F1), and MMBench (Acc). The x-axis is the TFLOP ratio of the model after token pruning compared to the original model before pruning. The proposed method significantly outperforms all baselines. Note that, unlike other methods, FitPrune uses an additional calibration step to prune tokens." + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "image_caption" + } + ], + "index": 9 + }, + { + "bbox": [ + 312, + 481, + 555, + 601 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 312, + 481, + 555, + 601 + ], + "spans": [ + { + "bbox": [ + 312, + 481, + 555, + 601 + ], + "type": "text", + "content": "as embeddings. These tokens are then combined and processed by an integrated LLM. The inclusion of visual tokens significantly increases the total number of tokens, often adding thousands to the combined set. Since the running time and memory requirements scale quadratically with input size [7, 8, 15, 38], the addition of visual tokens can substantially raise the running time for LMMs. Hence, many of these models often struggle to meet the demands of low-latency applications, particularly in resource-constrained environments [46]." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 312, + 603, + 556, + 699 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 312, + 603, + 556, + 699 + ], + "spans": [ + { + "bbox": [ + 312, + 603, + 556, + 699 + ], + "type": "text", + "content": "Previous research [4, 36, 47] has demonstrated that there is a high degree of redundancy in the visual information processed by LMMs. As a result, visual token pruning has emerged as a promising solution to address the computational complexity challenges faced by LMMs. Specifically, previous research has demonstrated that reducing the number of visual tokens by " + }, + { + "bbox": [ + 312, + 603, + 556, + 699 + ], + "type": "inline_equation", + "content": "50\\%" + }, + { + "bbox": [ + 312, + 603, + 556, + 699 + ], + "type": "text", + "content": " [4] to " + }, + { + "bbox": [ + 312, + 603, + 556, + 699 + ], + "type": "inline_equation", + "content": "95\\%" + }, + { + "bbox": [ + 312, + 603, + 556, + 699 + ], + "type": "text", + "content": " [36] can significantly enhance the inference speed of LMMs." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 326, + 701, + 555, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 326, + 701, + 555, + 714 + ], + "spans": [ + { + "bbox": [ + 326, + 701, + 555, + 714 + ], + "type": "text", + "content": "While promising, token pruning methods have certain" + } + ] + } + ], + "index": 13 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "spans": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "type": "text", + "content": "CVF" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 145, + 0, + 496, + 37 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 145, + 0, + 496, + 37 + ], + "spans": [ + { + "bbox": [ + 145, + 0, + 496, + 37 + ], + "type": "text", + "content": "This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 69, + 693, + 186, + 703 + ], + "type": "page_footnote", + "angle": 0, + "lines": [ + { + "bbox": [ + 69, + 693, + 186, + 703 + ], + "spans": [ + { + "bbox": [ + 69, + 693, + 186, + 703 + ], + "type": "text", + "content": "Authors have equal contributions." + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 69, + 703, + 184, + 713 + ], + "type": "page_footnote", + "angle": 0, + "lines": [ + { + "bbox": [ + 69, + 703, + 184, + 713 + ], + "spans": [ + { + "bbox": [ + 69, + 703, + 184, + 713 + ], + "type": "inline_equation", + "content": "\\diamond" + }, + { + "bbox": [ + 69, + 703, + 184, + 713 + ], + "type": "text", + "content": " https://github.com/vbdi/divprune" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 294, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 315, + 757 + ], + "type": "text", + "content": "9392" + } + ] + } + ], + "index": 17 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 0 + }, + { + "para_blocks": [ + { + "bbox": [ + 55, + 72, + 294, + 251 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 72, + 294, + 251 + ], + "spans": [ + { + "bbox": [ + 55, + 72, + 294, + 251 + ], + "type": "text", + "content": "shortcomings. For example, the works in [3, 17, 21, 47] require calibration or finetuning for each model which is costly and time-consuming to implement. FastV [4] and PruMerge [36] use attention scores to identify less important tokens for pruning. However, it is shown that using attention scores is not optimal, as some important tokens are overlooked [21]. Additionally, attention-based pruning tends to retain tokens that are similar to each other, leading to redundancy. At high compression ratio, this redundancy prevents the selection of a sufficient number of unique tokens to accurately represent the original tokens. In line with this observation, our findings indicate that pruning a large portion of visual tokens using these methods, without subsequent fine-tuning, results in a significant drop in the performance of LMMs across various tasks (Fig. 1)." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 55, + 253, + 294, + 420 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 253, + 294, + 420 + ], + "spans": [ + { + "bbox": [ + 55, + 253, + 294, + 420 + ], + "type": "text", + "content": "To address the above-mentioned issues, we formulate token pruning as a Max-Min Diversity Problem (MMDP) [35]. In an MMDP, the objective is to select a subset of elements such that the diversity among them is maximized. We apply this concept to token pruning, which we call DivPrune, aiming to maximize the diversity of the selected tokens by increasing the minimum distance between them. By ensuring high diversity, DivPrune captures a broader range of visual tokens, making it inherently more robust compared to attention-based methods that focus only on token importance scores. Increasing the diversity also helps ensure that the selected tokens better represent the original set of tokens, enabling effective performance even at high pruning ratios without the need for fine-tuning." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 55, + 422, + 295, + 552 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 422, + 295, + 552 + ], + "spans": [ + { + "bbox": [ + 55, + 422, + 295, + 552 + ], + "type": "text", + "content": "DivPrune also offers practical advantages that make it a highly useful solution in real-world scenarios. DivPrune is a plug-and-play solution that can be used without requiring offline optimization with a calibration set, or fine-tuning of the model, which are often time-consuming and computationally expensive. DivPrune is applicable to LMMs with any LLM architecture and vision encoder. Additionally, DivPrune is compatible with inference optimization techniques, such as KV caching, resulting in practical speedup in real-world applications. In summary, our major contributions are as follows:" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 55, + 558, + 295, + 713 + ], + "type": "list", + "angle": 0, + "index": 7, + "blocks": [ + { + "bbox": [ + 55, + 558, + 295, + 604 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 558, + 295, + 604 + ], + "spans": [ + { + "bbox": [ + 55, + 558, + 295, + 604 + ], + "type": "text", + "content": "- We introduce DivPrune, a token pruning method based on MMDP that maximizes diversity among visual tokens, effectively reducing redundancy and ensuring a highly representative subset." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 55, + 605, + 295, + 640 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 605, + 295, + 640 + ], + "spans": [ + { + "bbox": [ + 55, + 605, + 295, + 640 + ], + "type": "text", + "content": "- DivPrune is a training-free, calibration-data-free, plug-and-play solution that can be seamlessly integrated with off-the-shelf LMMs." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 55, + 642, + 295, + 700 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 642, + 295, + 700 + ], + "spans": [ + { + "bbox": [ + 55, + 642, + 295, + 700 + ], + "type": "text", + "content": "- We conduct evaluations using 16 datasets on image-and video-language models with image and video understanding tasks. DivPrune achieves state-of-the-art performance, with noticeable gains under extreme pruning (i.e., ratio " + }, + { + "bbox": [ + 55, + 642, + 295, + 700 + ], + "type": "inline_equation", + "content": ">80\\%" + }, + { + "bbox": [ + 55, + 642, + 295, + 700 + ], + "type": "text", + "content": ")." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 55, + 701, + 295, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 701, + 295, + 713 + ], + "spans": [ + { + "bbox": [ + 55, + 701, + 295, + 713 + ], + "type": "text", + "content": "DivPrune reduces GPU memory usage and inference la" + } + ] + } + ], + "index": 6 + } + ], + "sub_type": "text" + }, + { + "bbox": [ + 322, + 72, + 555, + 96 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 322, + 72, + 555, + 96 + ], + "spans": [ + { + "bbox": [ + 322, + 72, + 555, + 96 + ], + "type": "text", + "content": "tency while maintaining comparable accuracy compared to the original model across most datasets." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 314, + 107, + 406, + 119 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 314, + 107, + 406, + 119 + ], + "spans": [ + { + "bbox": [ + 314, + 107, + 406, + 119 + ], + "type": "text", + "content": "2. Related Works" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 313, + 127, + 502, + 140 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 127, + 502, + 140 + ], + "spans": [ + { + "bbox": [ + 313, + 127, + 502, + 140 + ], + "type": "text", + "content": "2.1. Large Multimodal Models (LMMs)" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 313, + 145, + 555, + 277 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 145, + 555, + 277 + ], + "spans": [ + { + "bbox": [ + 313, + 145, + 555, + 277 + ], + "type": "text", + "content": "LMMs handle diverse data types, including text, audio, image, and, video [5, 19, 22, 23, 30, 39, 52]. This work focuses on open-source LMMs that support language and visual inputs. These LMMs can be categorized into two types: image-based and video-based LMMs. The image-based LMMs [22, 23] address image-language understanding tasks, like image captioning, visual question answering, and image reasoning. On the other hand, video-based LMMs are geared towards video understanding [19, 52] tasks, like video captioning, video summarization, and video question answering." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 313, + 285, + 411, + 297 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 285, + 411, + 297 + ], + "spans": [ + { + "bbox": [ + 313, + 285, + 411, + 297 + ], + "type": "text", + "content": "2.2. Efficient LMMs" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 313, + 304, + 556, + 449 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 304, + 556, + 449 + ], + "spans": [ + { + "bbox": [ + 313, + 304, + 556, + 449 + ], + "type": "text", + "content": "Several techniques are proposed to improve inference efficiency specifically for LMMs. The first technique is to change the model architecture in LMMs. For example, [33] proposed to replace transformer-based LLMs with Mamba model [12]. [49, 53] retrained LMMs with small scale LLMs to improve their efficiency. [45] used knowledge distillation to train a small LMM. In addition to changing the architecture, it is shown in [37] that skipping some blocks or layers within LMMs can improve the inference speed without damaging the model's performance. Furthermore, efficient decoding techniques such as speculative decoding are proposed to make LMM inference more efficient [11]." + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 313, + 455, + 440, + 468 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 455, + 440, + 468 + ], + "spans": [ + { + "bbox": [ + 313, + 455, + 440, + 468 + ], + "type": "text", + "content": "2.3. Visual Token Pruning" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 313, + 474, + 555, + 594 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 474, + 555, + 594 + ], + "spans": [ + { + "bbox": [ + 313, + 474, + 555, + 594 + ], + "type": "text", + "content": "Visual token pruning methods are proposed to reduce the inference complexity for LMMs. The first group of methods uses attention scores to prune tokens [4, 36]. PruMerge [36] introduces a token pruning method for the vision encoder where the visual tokens are clustered and merged based on their attention sparsity. In addition, FastV [4] prunes tokens within a specific layer of the LLM based on the magnitude of attention scores in an earlier layer. It is shown that pruning tokens based on attention scores are not optimal [13, 21], especially at higher pruning ratios." + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 313, + 594, + 556, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 594, + 556, + 713 + ], + "spans": [ + { + "bbox": [ + 313, + 594, + 556, + 713 + ], + "type": "text", + "content": "Calibration-based methods offer another line of work, where pruning layers and/or ratios are determined by analyzing the LLM outputs for a calibration dataset [21, 47]. For example, FitPrune [47] calculates a pruning recipe based on the observed attention divergence before and after pruning. VTW [21] argues that visual tokens can be entirely removed after a certain layer within LLM. The layer to remove the visual tokens is chosen using a calibration dataset. These methods rely on calibration datasets and require custom calibration for each LMM, which can be costly and" + } + ] + } + ], + "index": 16 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 315, + 757 + ], + "type": "text", + "content": "9393" + } + ] + } + ], + "index": 17 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 1 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 58, + 69, + 296, + 228 + ], + "blocks": [ + { + "bbox": [ + 58, + 69, + 296, + 228 + ], + "lines": [ + { + "bbox": [ + 58, + 69, + 296, + 228 + ], + "spans": [ + { + "bbox": [ + 58, + 69, + 296, + 228 + ], + "type": "image", + "image_path": "bfb63d1030fc9da52cd07091a1af6939203095e7c0fd7c2ba1d42c1919a5da5f.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 55, + 238, + 296, + 272 + ], + "lines": [ + { + "bbox": [ + 55, + 238, + 296, + 272 + ], + "spans": [ + { + "bbox": [ + 55, + 238, + 296, + 272 + ], + "type": "text", + "content": "Figure 2. An overview of the LMM architecture, with DivPrune applied to visual tokens. The blocks on the right-hand side illustrate the steps of the method." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "bbox": [ + 55, + 282, + 177, + 293 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 282, + 177, + 293 + ], + "spans": [ + { + "bbox": [ + 55, + 282, + 177, + 293 + ], + "type": "text", + "content": "cumbersome for new models." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 55, + 294, + 296, + 402 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 294, + 296, + 402 + ], + "spans": [ + { + "bbox": [ + 55, + 294, + 296, + 402 + ], + "type": "text", + "content": "Some previous works proposed token pruning with the need for fine-tuning. " + }, + { + "bbox": [ + 55, + 294, + 296, + 402 + ], + "type": "inline_equation", + "content": "\\mathbf{M}^3" + }, + { + "bbox": [ + 55, + 294, + 296, + 402 + ], + "type": "text", + "content": " [3] applies model fine-tuning to produce nested visual token representations at multiple granularities, allowing users to select token lengths dynamically during inference. In [17], a projector layer trained using a large-scale dataset is proposed that packs finer detailed information into compact token representations. These methods need significant computational resources for training, limiting their use across various scenarios." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 55, + 411, + 162, + 426 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 411, + 162, + 426 + ], + "spans": [ + { + "bbox": [ + 55, + 411, + 162, + 426 + ], + "type": "text", + "content": "3. Proposed Method" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 55, + 432, + 296, + 469 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 432, + 296, + 469 + ], + "spans": [ + { + "bbox": [ + 55, + 432, + 296, + 469 + ], + "type": "text", + "content": "In this section, we briefly discuss how LMMs work. Then, the token pruning problem is defined, followed by a detailed presentation of the proposed method." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 55, + 475, + 244, + 488 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 475, + 244, + 488 + ], + "spans": [ + { + "bbox": [ + 55, + 475, + 244, + 488 + ], + "type": "text", + "content": "3.1. Large Multimodal Models (LMMs)" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 55, + 493, + 296, + 601 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 493, + 296, + 601 + ], + "spans": [ + { + "bbox": [ + 55, + 493, + 296, + 601 + ], + "type": "text", + "content": "An LMM typically processes a pair of inputs, denoted as " + }, + { + "bbox": [ + 55, + 493, + 296, + 601 + ], + "type": "inline_equation", + "content": "(T,V)" + }, + { + "bbox": [ + 55, + 493, + 296, + 601 + ], + "type": "text", + "content": ", where " + }, + { + "bbox": [ + 55, + 493, + 296, + 601 + ], + "type": "inline_equation", + "content": "T" + }, + { + "bbox": [ + 55, + 493, + 296, + 601 + ], + "type": "text", + "content": " is the text input and " + }, + { + "bbox": [ + 55, + 493, + 296, + 601 + ], + "type": "inline_equation", + "content": "V" + }, + { + "bbox": [ + 55, + 493, + 296, + 601 + ], + "type": "text", + "content": " is the visual input such as image or video. The text input is mapped to " + }, + { + "bbox": [ + 55, + 493, + 296, + 601 + ], + "type": "inline_equation", + "content": "N" + }, + { + "bbox": [ + 55, + 493, + 296, + 601 + ], + "type": "text", + "content": " textual tokens " + }, + { + "bbox": [ + 55, + 493, + 296, + 601 + ], + "type": "inline_equation", + "content": "\\mathbf{E}_{\\mathbf{t}} = \\{t_1,\\dots ,t_N\\}" + }, + { + "bbox": [ + 55, + 493, + 296, + 601 + ], + "type": "text", + "content": " using a text encoder. Similarly, the visual input is processed by a corresponding vision encoder. Specifically, it takes visual information " + }, + { + "bbox": [ + 55, + 493, + 296, + 601 + ], + "type": "inline_equation", + "content": "V" + }, + { + "bbox": [ + 55, + 493, + 296, + 601 + ], + "type": "text", + "content": " as input and outputs image features, that are further converted to " + }, + { + "bbox": [ + 55, + 493, + 296, + 601 + ], + "type": "inline_equation", + "content": "M" + }, + { + "bbox": [ + 55, + 493, + 296, + 601 + ], + "type": "text", + "content": " (generally " + }, + { + "bbox": [ + 55, + 493, + 296, + 601 + ], + "type": "inline_equation", + "content": "M\\gg N" + }, + { + "bbox": [ + 55, + 493, + 296, + 601 + ], + "type": "text", + "content": ") vision tokens " + }, + { + "bbox": [ + 55, + 493, + 296, + 601 + ], + "type": "inline_equation", + "content": "\\mathbf{E}_{\\mathbf{v}} = \\{v_{1},\\ldots ,v_{M}\\}" + }, + { + "bbox": [ + 55, + 493, + 296, + 601 + ], + "type": "text", + "content": " using a projector layer (Fig. 2)." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 55, + 601, + 296, + 651 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 601, + 296, + 651 + ], + "spans": [ + { + "bbox": [ + 55, + 601, + 296, + 651 + ], + "type": "text", + "content": "The textual tokens and visual tokens are then combined to be fed to an LLM to generate the prediction in an autoregressive manner. Specifically, " + }, + { + "bbox": [ + 55, + 601, + 296, + 651 + ], + "type": "inline_equation", + "content": "\\hat{N}" + }, + { + "bbox": [ + 55, + 601, + 296, + 651 + ], + "type": "text", + "content": " output tokens " + }, + { + "bbox": [ + 55, + 601, + 296, + 651 + ], + "type": "inline_equation", + "content": "\\mathbf{Y} = \\{y_1,\\dots ,y_{\\hat{N}}\\}" + }, + { + "bbox": [ + 55, + 601, + 296, + 651 + ], + "type": "text", + "content": " are generated as follows:" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 64, + 658, + 295, + 693 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 64, + 658, + 295, + 693 + ], + "spans": [ + { + "bbox": [ + 64, + 658, + 295, + 693 + ], + "type": "interline_equation", + "content": "P \\left(y _ {1}, \\dots , y _ {\\hat {N}} \\mid \\mathbf {E} _ {\\mathbf {t}}, \\mathbf {E} _ {\\mathbf {v}}\\right) = \\prod_ {i = 1} ^ {\\hat {N}} P \\left(y _ {i} \\mid y _ {< i}, \\mathbf {E} _ {\\mathbf {t}}, \\mathbf {E} _ {\\mathbf {v}}\\right), \\tag {1}", + "image_path": "bd6d7285b518300a1b59d93e99ddc34dd6a2f530fef4eab8cd4cc558ee9839ce.jpg" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 55, + 701, + 296, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 701, + 296, + 714 + ], + "spans": [ + { + "bbox": [ + 55, + 701, + 296, + 714 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 55, + 701, + 296, + 714 + ], + "type": "inline_equation", + "content": "P(|)" + }, + { + "bbox": [ + 55, + 701, + 296, + 714 + ], + "type": "text", + "content": " is the conditional probability obtained at the" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 313, + 72, + 394, + 83 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 72, + 394, + 83 + ], + "spans": [ + { + "bbox": [ + 313, + 72, + 394, + 83 + ], + "type": "text", + "content": "output of the LLM." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 313, + 90, + 408, + 103 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 90, + 408, + 103 + ], + "spans": [ + { + "bbox": [ + 313, + 90, + 408, + 103 + ], + "type": "text", + "content": "3.2. Token Pruning" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 313, + 109, + 554, + 156 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 109, + 554, + 156 + ], + "spans": [ + { + "bbox": [ + 313, + 109, + 554, + 156 + ], + "type": "text", + "content": "Reducing the number of input tokens in an integrated LLM within LMMs helps to lower memory usage and inference latency. Since visual tokens tend to have more redundancy, they are generally selected for pruning." + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 313, + 157, + 555, + 300 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 157, + 555, + 300 + ], + "spans": [ + { + "bbox": [ + 313, + 157, + 555, + 300 + ], + "type": "text", + "content": "In this context, the problem of token pruning can be defined as follows: given a set of visual tokens " + }, + { + "bbox": [ + 313, + 157, + 555, + 300 + ], + "type": "inline_equation", + "content": "\\mathbf{E}_{\\mathbf{v}}" + }, + { + "bbox": [ + 313, + 157, + 555, + 300 + ], + "type": "text", + "content": " with " + }, + { + "bbox": [ + 313, + 157, + 555, + 300 + ], + "type": "inline_equation", + "content": "|\\mathbf{E}_{\\mathbf{v}}| = M" + }, + { + "bbox": [ + 313, + 157, + 555, + 300 + ], + "type": "text", + "content": " and the subset size " + }, + { + "bbox": [ + 313, + 157, + 555, + 300 + ], + "type": "inline_equation", + "content": "\\tilde{M} (\\tilde{M} < M)" + }, + { + "bbox": [ + 313, + 157, + 555, + 300 + ], + "type": "text", + "content": ", the goal is to select a subset, " + }, + { + "bbox": [ + 313, + 157, + 555, + 300 + ], + "type": "inline_equation", + "content": "\\tilde{\\mathbf{E}}_{\\mathbf{v}}" + }, + { + "bbox": [ + 313, + 157, + 555, + 300 + ], + "type": "text", + "content": ", while preserving key information necessary for accurate predictions. To mathematically formulate the token pruning problem, we define a mapping function " + }, + { + "bbox": [ + 313, + 157, + 555, + 300 + ], + "type": "inline_equation", + "content": "f" + }, + { + "bbox": [ + 313, + 157, + 555, + 300 + ], + "type": "text", + "content": ", which maps the original set of visual tokens, " + }, + { + "bbox": [ + 313, + 157, + 555, + 300 + ], + "type": "inline_equation", + "content": "\\mathbf{E}_{\\mathbf{v}}" + }, + { + "bbox": [ + 313, + 157, + 555, + 300 + ], + "type": "text", + "content": ", to a subset, " + }, + { + "bbox": [ + 313, + 157, + 555, + 300 + ], + "type": "inline_equation", + "content": "\\tilde{\\mathbf{E}}_{\\mathbf{v}} = \\{\\tilde{v}_1,\\dots ,\\tilde{v}_{\\tilde{M}}\\}" + }, + { + "bbox": [ + 313, + 157, + 555, + 300 + ], + "type": "text", + "content": ", where " + }, + { + "bbox": [ + 313, + 157, + 555, + 300 + ], + "type": "inline_equation", + "content": "|\\tilde{\\mathbf{E}}_{\\mathbf{v}}| = \\tilde{M}" + }, + { + "bbox": [ + 313, + 157, + 555, + 300 + ], + "type": "text", + "content": ". The objective is to identify a mapping function " + }, + { + "bbox": [ + 313, + 157, + 555, + 300 + ], + "type": "inline_equation", + "content": "f" + }, + { + "bbox": [ + 313, + 157, + 555, + 300 + ], + "type": "text", + "content": " that minimizes the difference in the model's output before and after pruning while ensuring the reduced set still captures the essential information from the original set:" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 400, + 305, + 489, + 319 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 400, + 305, + 489, + 319 + ], + "spans": [ + { + "bbox": [ + 400, + 305, + 489, + 319 + ], + "type": "interline_equation", + "content": "\\text {F i n d :} \\quad f: \\mathbf {E} _ {v} \\rightarrow \\tilde {\\mathbf {E}} _ {\\mathbf {v}}", + "image_path": "bb472c9659c295874556acfbab753d3fae08c182c9368df0858aa9fa687f8756.jpg" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 378, + 322, + 554, + 342 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 378, + 322, + 554, + 342 + ], + "spans": [ + { + "bbox": [ + 378, + 322, + 554, + 342 + ], + "type": "interline_equation", + "content": "\\text {O b j e c t i v e :} \\quad \\min _ {f} \\mathcal {L} (\\mathcal {P}, \\tilde {\\mathcal {P}}) \\tag {2}", + "image_path": "15f1a25a431e301bf07dd2530e5a1e32cd32c65e2684a6796a92682b55bac621.jpg" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 389, + 344, + 490, + 359 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 389, + 344, + 490, + 359 + ], + "spans": [ + { + "bbox": [ + 389, + 344, + 490, + 359 + ], + "type": "interline_equation", + "content": "\\text {S u b j e c t} \\quad | \\tilde {\\mathbf {E}} _ {\\mathbf {v}} | = \\tilde {M},", + "image_path": "dd32a000fced341770e4e6349b1c0c765d12a53bfc7c515f0bca2962ae97c49f.jpg" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 313, + 365, + 554, + 438 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 365, + 554, + 438 + ], + "spans": [ + { + "bbox": [ + 313, + 365, + 554, + 438 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 313, + 365, + 554, + 438 + ], + "type": "inline_equation", + "content": "\\mathcal{P} = P(y_1, \\ldots, y_{\\hat{N}} \\mid \\mathbf{E}_{\\mathbf{t}}, \\mathbf{E}_{\\mathbf{v}})" + }, + { + "bbox": [ + 313, + 365, + 554, + 438 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 313, + 365, + 554, + 438 + ], + "type": "inline_equation", + "content": "\\tilde{\\mathcal{P}} = P(y_1, \\ldots, y_{\\hat{N}} \\mid \\mathbf{E}_{\\mathbf{t}}, f(\\mathbf{E}_{\\mathbf{v}}))" + }, + { + "bbox": [ + 313, + 365, + 554, + 438 + ], + "type": "text", + "content": ". Here, " + }, + { + "bbox": [ + 313, + 365, + 554, + 438 + ], + "type": "inline_equation", + "content": "\\mathcal{L}" + }, + { + "bbox": [ + 313, + 365, + 554, + 438 + ], + "type": "text", + "content": " represents a loss function that measures the difference in the model's output with and without pruning, and " + }, + { + "bbox": [ + 313, + 365, + 554, + 438 + ], + "type": "inline_equation", + "content": "\\tilde{M}" + }, + { + "bbox": [ + 313, + 365, + 554, + 438 + ], + "type": "text", + "content": " indicates the number of retained tokens. Next, we propose a novel diversity-based solution for the introduced token pruning problem." + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 313, + 445, + 474, + 456 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 445, + 474, + 456 + ], + "spans": [ + { + "bbox": [ + 313, + 445, + 474, + 456 + ], + "type": "text", + "content": "3.3. DivPrune: Method Overview" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 313, + 463, + 554, + 559 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 463, + 554, + 559 + ], + "spans": [ + { + "bbox": [ + 313, + 463, + 554, + 559 + ], + "type": "text", + "content": "We proposed a diversity-based token pruning method by reformulating the problem in (2) to select a subset of " + }, + { + "bbox": [ + 313, + 463, + 554, + 559 + ], + "type": "inline_equation", + "content": "\\tilde{M}" + }, + { + "bbox": [ + 313, + 463, + 554, + 559 + ], + "type": "text", + "content": " elements that maximizes the diversity, thereby reducing redundancy. Specifically, we define token pruning as Max-Min Diversity Problem (MMDP) [32] where the goal is to find the set " + }, + { + "bbox": [ + 313, + 463, + 554, + 559 + ], + "type": "inline_equation", + "content": "\\tilde{\\mathbf{E}}_{\\mathbf{v}}" + }, + { + "bbox": [ + 313, + 463, + 554, + 559 + ], + "type": "text", + "content": " among all possible sets with " + }, + { + "bbox": [ + 313, + 463, + 554, + 559 + ], + "type": "inline_equation", + "content": "\\tilde{M}" + }, + { + "bbox": [ + 313, + 463, + 554, + 559 + ], + "type": "text", + "content": " samples in " + }, + { + "bbox": [ + 313, + 463, + 554, + 559 + ], + "type": "inline_equation", + "content": "\\mathbf{E}_{\\mathbf{v}}" + }, + { + "bbox": [ + 313, + 463, + 554, + 559 + ], + "type": "text", + "content": " that has the maximum minimum distance between its elements. So, MMDP is defined as:" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 324, + 563, + 554, + 590 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 324, + 563, + 554, + 590 + ], + "spans": [ + { + "bbox": [ + 324, + 563, + 554, + 590 + ], + "type": "interline_equation", + "content": "\\operatorname {F i n d} \\tilde {\\mathbf {E}} _ {\\mathbf {v}} = \\arg \\max \\left[ \\min _ {\\gamma , \\omega \\in S} (d (\\gamma , \\omega)): \\forall S \\subset \\mathbf {E} _ {\\mathbf {v}} \\right], \\tag {3}", + "image_path": "9f97975327e6e0aa57395c1d1236e17eb692656d0a53019be98504c654ebf1f2.jpg" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 313, + 596, + 555, + 642 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 596, + 555, + 642 + ], + "spans": [ + { + "bbox": [ + 313, + 596, + 555, + 642 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 313, + 596, + 555, + 642 + ], + "type": "inline_equation", + "content": "S" + }, + { + "bbox": [ + 313, + 596, + 555, + 642 + ], + "type": "text", + "content": " is an arbitrary set in " + }, + { + "bbox": [ + 313, + 596, + 555, + 642 + ], + "type": "inline_equation", + "content": "\\mathbf{E}_{\\mathbf{v}}" + }, + { + "bbox": [ + 313, + 596, + 555, + 642 + ], + "type": "text", + "content": " with " + }, + { + "bbox": [ + 313, + 596, + 555, + 642 + ], + "type": "inline_equation", + "content": "\\tilde{M}" + }, + { + "bbox": [ + 313, + 596, + 555, + 642 + ], + "type": "text", + "content": " elements and " + }, + { + "bbox": [ + 313, + 596, + 555, + 642 + ], + "type": "inline_equation", + "content": "(\\gamma, \\omega)" + }, + { + "bbox": [ + 313, + 596, + 555, + 642 + ], + "type": "text", + "content": " are arbitrary elements in " + }, + { + "bbox": [ + 313, + 596, + 555, + 642 + ], + "type": "inline_equation", + "content": "S" + }, + { + "bbox": [ + 313, + 596, + 555, + 642 + ], + "type": "text", + "content": ". The distance is measured by " + }, + { + "bbox": [ + 313, + 596, + 555, + 642 + ], + "type": "inline_equation", + "content": "d(.,.)" + }, + { + "bbox": [ + 313, + 596, + 555, + 642 + ], + "type": "text", + "content": " which is defined using the cosine distance as follows:" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 383, + 640, + 553, + 664 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 383, + 640, + 553, + 664 + ], + "spans": [ + { + "bbox": [ + 383, + 640, + 553, + 664 + ], + "type": "interline_equation", + "content": "d (\\gamma , \\omega) = 1 - \\frac {\\gamma \\cdot \\omega}{\\| \\gamma \\| \\| \\omega \\|}. \\tag {4}", + "image_path": "e0fe39ec652af12f130cb7d623eaf9adbea8075c47ce357a8e4a391511bdcd8f.jpg" + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 313, + 666, + 555, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 666, + 555, + 714 + ], + "spans": [ + { + "bbox": [ + 313, + 666, + 555, + 714 + ], + "type": "text", + "content": "A solution for the MMDP problem in (3) is a subset of " + }, + { + "bbox": [ + 313, + 666, + 555, + 714 + ], + "type": "inline_equation", + "content": "\\mathbf{E}_{\\mathbf{v}}" + }, + { + "bbox": [ + 313, + 666, + 555, + 714 + ], + "type": "text", + "content": " that maximizes diversity by minimizing redundancy between elements. In the literature, several solutions including exact and heuristic methods are proposed to solve the" + } + ] + } + ], + "index": 24 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 315, + 757 + ], + "type": "text", + "content": "9394" + } + ] + } + ], + "index": 25 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 2 + }, + { + "para_blocks": [ + { + "type": "code", + "bbox": [ + 58, + 90, + 271, + 357 + ], + "blocks": [ + { + "bbox": [ + 61, + 75, + 254, + 87 + ], + "lines": [ + { + "bbox": [ + 61, + 75, + 254, + 87 + ], + "spans": [ + { + "bbox": [ + 61, + 75, + 254, + 87 + ], + "type": "text", + "content": "Algorithm 1: Proposed Token Pruning Method" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "code_caption" + }, + { + "bbox": [ + 58, + 90, + 271, + 357 + ], + "lines": [ + { + "bbox": [ + 58, + 90, + 271, + 357 + ], + "spans": [ + { + "bbox": [ + 58, + 90, + 271, + 357 + ], + "type": "text", + "content": "1 " + }, + { + "bbox": [ + 58, + 90, + 271, + 357 + ], + "type": "inline_equation", + "content": "\\tilde{M}" + }, + { + "bbox": [ + 58, + 90, + 271, + 357 + ], + "type": "text", + "content": " : subset size; " + }, + { + "bbox": [ + 58, + 90, + 271, + 357 + ], + "type": "inline_equation", + "content": "\\mathbf{E}_{\\mathbf{v}}" + }, + { + "bbox": [ + 58, + 90, + 271, + 357 + ], + "type": "text", + "content": " : visual tokens; " + }, + { + "bbox": [ + 58, + 90, + 271, + 357 + ], + "type": "inline_equation", + "content": "\\tilde{\\mathbf{E}}_{\\mathbf{v}}" + }, + { + "bbox": [ + 58, + 90, + 271, + 357 + ], + "type": "text", + "content": " : selected subset \n2 Initialize " + }, + { + "bbox": [ + 58, + 90, + 271, + 357 + ], + "type": "inline_equation", + "content": "\\tilde{\\mathbf{E}}_{\\mathbf{v}} = []" + }, + { + "bbox": [ + 58, + 90, + 271, + 357 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 58, + 90, + 271, + 357 + ], + "type": "inline_equation", + "content": "\\mathbf{R} = \\mathbf{E}_{\\mathbf{v}}" + }, + { + "bbox": [ + 58, + 90, + 271, + 357 + ], + "type": "text", + "content": " \n3 // First stage: add the first token \n4 " + }, + { + "bbox": [ + 58, + 90, + 271, + 357 + ], + "type": "inline_equation", + "content": "D = []" + }, + { + "bbox": [ + 58, + 90, + 271, + 357 + ], + "type": "text", + "content": " initialize the distance array \n5 for i in R do \n6 " + }, + { + "bbox": [ + 58, + 90, + 271, + 357 + ], + "type": "inline_equation", + "content": "d_{min} = +inf" + }, + { + "bbox": [ + 58, + 90, + 271, + 357 + ], + "type": "text", + "content": " \n7 for j in R do \n8 If " + }, + { + "bbox": [ + 58, + 90, + 271, + 357 + ], + "type": "inline_equation", + "content": "(i \\neq j \\& d(i, j) \\leq d_{min})" + }, + { + "bbox": [ + 58, + 90, + 271, + 357 + ], + "type": "text", + "content": " then " + }, + { + "bbox": [ + 58, + 90, + 271, + 357 + ], + "type": "inline_equation", + "content": "d_{min} = d(i, j)" + }, + { + "bbox": [ + 58, + 90, + 271, + 357 + ], + "type": "text", + "content": " \n9 Add " + }, + { + "bbox": [ + 58, + 90, + 271, + 357 + ], + "type": "inline_equation", + "content": "d_{min}" + }, + { + "bbox": [ + 58, + 90, + 271, + 357 + ], + "type": "text", + "content": " to " + }, + { + "bbox": [ + 58, + 90, + 271, + 357 + ], + "type": "inline_equation", + "content": "D" + }, + { + "bbox": [ + 58, + 90, + 271, + 357 + ], + "type": "text", + "content": " \n10 " + }, + { + "bbox": [ + 58, + 90, + 271, + 357 + ], + "type": "inline_equation", + "content": "k = \\mathbf{R}[\\arg \\max (\\mathrm{D})]" + }, + { + "bbox": [ + 58, + 90, + 271, + 357 + ], + "type": "text", + "content": " \n11 move k from R to " + }, + { + "bbox": [ + 58, + 90, + 271, + 357 + ], + "type": "inline_equation", + "content": "\\tilde{\\mathbf{E}}_{\\mathbf{v}}" + }, + { + "bbox": [ + 58, + 90, + 271, + 357 + ], + "type": "text", + "content": " \n12 // Second stage: iteratively add the subsequent tokens \n13 while " + }, + { + "bbox": [ + 58, + 90, + 271, + 357 + ], + "type": "inline_equation", + "content": "|\\tilde{\\mathbf{E}}_{\\mathbf{v}}| < \\tilde{M}" + }, + { + "bbox": [ + 58, + 90, + 271, + 357 + ], + "type": "text", + "content": " do \n14 " + }, + { + "bbox": [ + 58, + 90, + 271, + 357 + ], + "type": "inline_equation", + "content": "D = []" + }, + { + "bbox": [ + 58, + 90, + 271, + 357 + ], + "type": "text", + "content": " initialize the distance array \n15 for i in R do \n16 " + }, + { + "bbox": [ + 58, + 90, + 271, + 357 + ], + "type": "inline_equation", + "content": "d_{min} = +inf" + }, + { + "bbox": [ + 58, + 90, + 271, + 357 + ], + "type": "text", + "content": " \n17 for j in " + }, + { + "bbox": [ + 58, + 90, + 271, + 357 + ], + "type": "inline_equation", + "content": "\\tilde{\\mathbf{E}}_{\\mathbf{v}}" + }, + { + "bbox": [ + 58, + 90, + 271, + 357 + ], + "type": "text", + "content": " do \n18 If " + }, + { + "bbox": [ + 58, + 90, + 271, + 357 + ], + "type": "inline_equation", + "content": "d(i, j) \\leq d_{min}" + }, + { + "bbox": [ + 58, + 90, + 271, + 357 + ], + "type": "text", + "content": " then " + }, + { + "bbox": [ + 58, + 90, + 271, + 357 + ], + "type": "inline_equation", + "content": "d_{min} = d(i, j)" + }, + { + "bbox": [ + 58, + 90, + 271, + 357 + ], + "type": "text", + "content": " \n19 Add " + }, + { + "bbox": [ + 58, + 90, + 271, + 357 + ], + "type": "inline_equation", + "content": "d_{min}" + }, + { + "bbox": [ + 58, + 90, + 271, + 357 + ], + "type": "text", + "content": " to " + }, + { + "bbox": [ + 58, + 90, + 271, + 357 + ], + "type": "inline_equation", + "content": "D" + }, + { + "bbox": [ + 58, + 90, + 271, + 357 + ], + "type": "text", + "content": " \n20 " + }, + { + "bbox": [ + 58, + 90, + 271, + 357 + ], + "type": "inline_equation", + "content": "k = \\mathbf{R}[\\arg \\max (\\mathrm{D})]" + }, + { + "bbox": [ + 58, + 90, + 271, + 357 + ], + "type": "text", + "content": " \n21 move k from R to " + }, + { + "bbox": [ + 58, + 90, + 271, + 357 + ], + "type": "inline_equation", + "content": "\\tilde{\\mathbf{E}}_{\\mathbf{v}}" + }, + { + "bbox": [ + 58, + 90, + 271, + 357 + ], + "type": "text", + "content": " \n22 Return " + }, + { + "bbox": [ + 58, + 90, + 271, + 357 + ], + "type": "inline_equation", + "content": "\\tilde{\\mathbf{E}}_{\\mathbf{v}}" + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "code_body" + } + ], + "index": 1, + "sub_type": "algorithm" + }, + { + "bbox": [ + 55, + 387, + 296, + 506 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 387, + 296, + 506 + ], + "spans": [ + { + "bbox": [ + 55, + 387, + 296, + 506 + ], + "type": "text", + "content": "MMDP problem [29, 35]. Since the number of tokens is generally limited (e.g., 576 in LLaVA 1.5 [22]) and the solvers are not generally designed for GPU acceleration, we obtain exact solution for the problem. Notably, the overhead of the selection process using GPU is negligible compared to the computations within the LLM. Detailed steps of the proposed method is summarized in Algorithm 1. Once the selected tokens are identified, the remaining visual tokens are discarded. The selected tokens along with the textual tokens are passed to the LLM." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 55, + 508, + 296, + 676 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 508, + 296, + 676 + ], + "spans": [ + { + "bbox": [ + 55, + 508, + 296, + 676 + ], + "type": "text", + "content": "As shown in Algorithm 1, the proposed method has two stages after the initialization. The selected subset, " + }, + { + "bbox": [ + 55, + 508, + 296, + 676 + ], + "type": "inline_equation", + "content": "\\tilde{\\mathbf{E}}_{\\mathbf{v}}" + }, + { + "bbox": [ + 55, + 508, + 296, + 676 + ], + "type": "text", + "content": ", is initialized as empty, and the candidate list " + }, + { + "bbox": [ + 55, + 508, + 296, + 676 + ], + "type": "inline_equation", + "content": "\\mathbf{R}" + }, + { + "bbox": [ + 55, + 508, + 296, + 676 + ], + "type": "text", + "content": " is initialized with all the visual tokens. In the first stage, the first token of the selected subset is chosen based on the pairwise distance between the tokens of the candidate list. Then, the chosen token is moved from the candidate list to the selected list. In the second stage, similar to the first stage, the pairwise distance of the tokens in " + }, + { + "bbox": [ + 55, + 508, + 296, + 676 + ], + "type": "inline_equation", + "content": "\\tilde{\\mathbf{E}}_{\\mathbf{v}}" + }, + { + "bbox": [ + 55, + 508, + 296, + 676 + ], + "type": "text", + "content": " and the tokens in " + }, + { + "bbox": [ + 55, + 508, + 296, + 676 + ], + "type": "inline_equation", + "content": "\\mathbf{R}" + }, + { + "bbox": [ + 55, + 508, + 296, + 676 + ], + "type": "text", + "content": " is used to add samples to " + }, + { + "bbox": [ + 55, + 508, + 296, + 676 + ], + "type": "inline_equation", + "content": "\\tilde{\\mathbf{E}}_{\\mathbf{v}}" + }, + { + "bbox": [ + 55, + 508, + 296, + 676 + ], + "type": "text", + "content": " iteratively. Finally, once the number of tokens in " + }, + { + "bbox": [ + 55, + 508, + 296, + 676 + ], + "type": "inline_equation", + "content": "\\tilde{\\mathbf{E}}_{\\mathbf{v}}" + }, + { + "bbox": [ + 55, + 508, + 296, + 676 + ], + "type": "text", + "content": " reaches the specified subset size, the selection procedure is terminated and the " + }, + { + "bbox": [ + 55, + 508, + 296, + 676 + ], + "type": "inline_equation", + "content": "\\tilde{\\mathbf{E}}_{\\mathbf{v}}" + }, + { + "bbox": [ + 55, + 508, + 296, + 676 + ], + "type": "text", + "content": " is returned. To avoid repeated distance calculations over iterations a distance matrix is initially calculated by one matrix multiplication." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 55, + 677, + 296, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 677, + 296, + 713 + ], + "spans": [ + { + "bbox": [ + 55, + 677, + 296, + 713 + ], + "type": "text", + "content": "The proposed method can also be applied to the features (i.e., hidden states) in the intermediate layers of the LLM. In this case, our method is not applied to the visual tokens, but" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 313, + 72, + 554, + 144 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 72, + 554, + 144 + ], + "spans": [ + { + "bbox": [ + 313, + 72, + 554, + 144 + ], + "type": "text", + "content": "to the features corresponding to the visual tokens obtained from a decoder layer to select a subset before feeding them to the subsequent layers. In either case, our method obtains the highest diversity for the selected elements. Ablation studies are provided in the next section to analyze the effect of pruning different elements at different layers." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 313, + 154, + 395, + 168 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 154, + 395, + 168 + ], + "spans": [ + { + "bbox": [ + 313, + 154, + 395, + 168 + ], + "type": "text", + "content": "4. Experiments" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 313, + 175, + 554, + 247 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 175, + 554, + 247 + ], + "spans": [ + { + "bbox": [ + 313, + 175, + 554, + 247 + ], + "type": "text", + "content": "In this section, we present a comprehensive analysis comparing the performance of our method and previous works across various settings, tasks, and datasets. Insights into the proposed method are also provided through illustrative examples. Moreover, the efficiency of DivPrune along with ablation study are provided." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 313, + 253, + 441, + 266 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 253, + 441, + 266 + ], + "spans": [ + { + "bbox": [ + 313, + 253, + 441, + 266 + ], + "type": "text", + "content": "4.1. Experimental Settings" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 313, + 271, + 555, + 439 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 271, + 555, + 439 + ], + "spans": [ + { + "bbox": [ + 313, + 271, + 555, + 439 + ], + "type": "text", + "content": "Baselines and Models: We consider five baselines, namely, FastV [4], PruMerge [36], VTW [21], FitPrune [47] and " + }, + { + "bbox": [ + 313, + 271, + 555, + 439 + ], + "type": "inline_equation", + "content": "\\mathbf{M}^3" + }, + { + "bbox": [ + 313, + 271, + 555, + 439 + ], + "type": "text", + "content": " [3]. Among these, we consider FastV, PruMerge, and VTW as our main competitors as they are plug-and-play and do not rely on any further costly finetuning or calibration process. However, for the sake of completeness, we also report performance comparison with respect to one finetuning-based " + }, + { + "bbox": [ + 313, + 271, + 555, + 439 + ], + "type": "inline_equation", + "content": "(\\mathbf{M}^3)" + }, + { + "bbox": [ + 313, + 271, + 555, + 439 + ], + "type": "text", + "content": " and one calibration-based (FitPrune) methods. Note that, VTW, by default, requires calibration to determine the best layer for a given task. However, doing that does not allow us to set a specific TFLOP ratio, complicating the comparison. Hence, whenever required we disable the calibration of VTW to select the layer that matches the FLOP requirement of a particular experiment." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 313, + 439, + 554, + 522 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 439, + 554, + 522 + ], + "spans": [ + { + "bbox": [ + 313, + 439, + 554, + 522 + ], + "type": "text", + "content": "We test DivPrune and the baselines with popular LMMs namely LLaVA 1.5-7B [22]1, LLaVA 1.5-13B [22]2, LLaVA 1.6-7B3 (also known as LLaVA-NeXT [23]), and LLaVA-NeXT-Video-7B [52]4 to demonstrate the generality of DivPrune. For each tested model and task, we report only the relevant subset of baseline that is applicable to that specific model and task, alongside our results." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 313, + 523, + 554, + 618 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 523, + 554, + 618 + ], + "spans": [ + { + "bbox": [ + 313, + 523, + 554, + 618 + ], + "type": "text", + "content": "All the tested LMMs used CLIP vision encoder [34]. LLaVA 1.5 model uses 576 visual tokens to represent images. LLaVA 1.6 converts each image into a varying number of patches, resulting in 3-5 times more visual tokens compared to LLaVA 1.5. LLaVA-NeXT-Video uses 144 tokens to process each frame. For all the experiments with LLaVA-NeXT-Video we used a total of 8 frames resulting in 1152 tokens for the processed frames." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 313, + 619, + 554, + 667 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 619, + 554, + 667 + ], + "spans": [ + { + "bbox": [ + 313, + 619, + 554, + 667 + ], + "type": "text", + "content": "Datasets, Tasks, and Metrics: We selected a comprehensive set of common tasks and datasets aimed at multimodal reasoning and understanding. Specifically, we chose 11 image-language and 5 video-language datasets." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 324, + 673, + 535, + 713 + ], + "type": "list", + "angle": 0, + "index": 17, + "blocks": [ + { + "bbox": [ + 324, + 673, + 484, + 683 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 324, + 673, + 484, + 683 + ], + "spans": [ + { + "bbox": [ + 324, + 673, + 484, + 683 + ], + "type": "text", + "content": "1https://huggingface.co/liuhaotian/flava-v1.5-7B" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 324, + 683, + 485, + 693 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 324, + 683, + 485, + 693 + ], + "spans": [ + { + "bbox": [ + 324, + 683, + 485, + 693 + ], + "type": "inline_equation", + "content": "^{2}" + }, + { + "bbox": [ + 324, + 683, + 485, + 693 + ], + "type": "text", + "content": "https://huggingface.co/luhaotian/llava-v1.5-13b" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 324, + 693, + 506, + 703 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 324, + 693, + 506, + 703 + ], + "spans": [ + { + "bbox": [ + 324, + 693, + 506, + 703 + ], + "type": "text", + "content": "3https://huggingface.co/liuhaotian/llava-v1.6-vicuna-7b" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 324, + 703, + 535, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 324, + 703, + 535, + 713 + ], + "spans": [ + { + "bbox": [ + 324, + 703, + 535, + 713 + ], + "type": "text", + "content": "4https://huggingface.co/lmms-lab/LLaVA-NeXT-Video-7B-DPO" + } + ] + } + ], + "index": 16 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 315, + 757 + ], + "type": "text", + "content": "9395" + } + ] + } + ], + "index": 18 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 3 + }, + { + "para_blocks": [ + { + "bbox": [ + 55, + 72, + 296, + 239 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 72, + 296, + 239 + ], + "spans": [ + { + "bbox": [ + 55, + 72, + 296, + 239 + ], + "type": "text", + "content": "These datasets encompass a wide range of tasks, including captioning, multiple-choice Question Answering (QA), and open-ended QA based on text and image/video inputs. Consistent with prior works, CIDEr score [42] is used for evaluating captioning tasks, and Exact Match (EM), Accuracy (Acc), F1, Perception Score (P-score) [9] and GPT-assisted [10] score are used for QA tasks. Furthermore, Wu-Palmer similarity (WUPS) score [43] and GPT-assisted score [10] is used for open-ended QA. For all task performance metrics used in this paper, higher values indicate better performance. For the reported time and memory, lower values indicate better results. Further details regarding the datasets, tasks, and metrics are provided in the supplementary material." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 55, + 239, + 296, + 336 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 239, + 296, + 336 + ], + "spans": [ + { + "bbox": [ + 55, + 239, + 296, + 336 + ], + "type": "text", + "content": "Following the earlier works in [4, 21, 47], we report the computational requirement, measured in TFLOPs, for DivPrune and the baselines. Various configurations including different pruning ratios at different layers are examined to obtain different working TFLOPs for our method and the baselines. The reported TFLOP ratio is the TFLOP of the model with pruned tokens relative to the original model's TFLOP with no pruning. This ratio is estimated as [4]:" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 65, + 345, + 294, + 365 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 65, + 345, + 294, + 365 + ], + "spans": [ + { + "bbox": [ + 65, + 345, + 294, + 365 + ], + "type": "interline_equation", + "content": "\\frac {K \\times \\left(4 \\mu d ^ {2} - 2 \\mu^ {2} d + 2 \\mu d m\\right) + (T - K) \\times \\left(4 \\tilde {\\mu} d ^ {2} - 2 \\tilde {\\mu} ^ {2} d + 2 \\tilde {\\mu} d m\\right)}{T \\times \\left(4 \\mu d ^ {2} - 2 \\mu^ {2} d + 2 \\mu d m\\right)}, \\tag {5}", + "image_path": "c65a6629bf87d135551550a81e33abe4689d41d8b06ae1b4f561cf235fe7676f.jpg" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 55, + 373, + 296, + 517 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 373, + 296, + 517 + ], + "spans": [ + { + "bbox": [ + 55, + 373, + 296, + 517 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 55, + 373, + 296, + 517 + ], + "type": "inline_equation", + "content": "T" + }, + { + "bbox": [ + 55, + 373, + 296, + 517 + ], + "type": "text", + "content": " is the total transformer-based decoder layers. " + }, + { + "bbox": [ + 55, + 373, + 296, + 517 + ], + "type": "inline_equation", + "content": "\\mu = N + M" + }, + { + "bbox": [ + 55, + 373, + 296, + 517 + ], + "type": "text", + "content": " is the total sequence length before pruning, " + }, + { + "bbox": [ + 55, + 373, + 296, + 517 + ], + "type": "inline_equation", + "content": "\\tilde{\\mu} = N + \\tilde{M}" + }, + { + "bbox": [ + 55, + 373, + 296, + 517 + ], + "type": "text", + "content": " is the sequence length after pruning, " + }, + { + "bbox": [ + 55, + 373, + 296, + 517 + ], + "type": "inline_equation", + "content": "d" + }, + { + "bbox": [ + 55, + 373, + 296, + 517 + ], + "type": "text", + "content": " is the hidden state size of the layer, and " + }, + { + "bbox": [ + 55, + 373, + 296, + 517 + ], + "type": "inline_equation", + "content": "m" + }, + { + "bbox": [ + 55, + 373, + 296, + 517 + ], + "type": "text", + "content": " is the intermediate size of feed-forward network module. Depending on the TFLOP ratio requirement set by a particular experiment, we adjust the pruning hyperparameters of all baselines to match that requirement. However, some baselines do not support fine-grained adjustments like our approach does. In these cases, we choose the smallest available TFLOP ratio that exceeds the requirement set by an experiment, which might give these baselines a slight advantage over our method." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 55, + 517, + 296, + 590 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 517, + 296, + 590 + ], + "spans": [ + { + "bbox": [ + 55, + 517, + 296, + 590 + ], + "type": "text", + "content": "We used " + }, + { + "bbox": [ + 55, + 517, + 296, + 590 + ], + "type": "inline_equation", + "content": "8 \\times" + }, + { + "bbox": [ + 55, + 517, + 296, + 590 + ], + "type": "text", + "content": " V100 GPUs with 32GB VRAM for all the experiments in this paper. Additionally, we used the Immsevalts package [51] for running these benchmarks for all the baselines and models. All results are obtained with a batch size of 1. For the metrics that require ChatGPT API access, the model is set to \"gpt-4o-mini\"." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 55, + 598, + 116, + 611 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 598, + 116, + 611 + ], + "spans": [ + { + "bbox": [ + 55, + 598, + 116, + 611 + ], + "type": "text", + "content": "4.2. Insights" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 55, + 616, + 296, + 676 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 616, + 296, + 676 + ], + "spans": [ + { + "bbox": [ + 55, + 616, + 296, + 676 + ], + "type": "text", + "content": "We provide visualizations comparing DivPrune with importance-based token pruning methods using LLaVA 1.5-7B and the SeedBench dataset [16]. Detailed analysis across different models and datasets is provided in the following subsections." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 55, + 677, + 296, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 677, + 296, + 713 + ], + "spans": [ + { + "bbox": [ + 55, + 677, + 296, + 713 + ], + "type": "text", + "content": "The visual tokens in LLaVa 1.5 model are 4096-dimensional vectors. The t-SNE method [41] is utilized to project the visual tokens in " + }, + { + "bbox": [ + 55, + 677, + 296, + 713 + ], + "type": "inline_equation", + "content": "\\mathbf{E}_{\\mathbf{v}}" + }, + { + "bbox": [ + 55, + 677, + 296, + 713 + ], + "type": "text", + "content": " from a high dimensional" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 313, + 72, + 553, + 179 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 72, + 553, + 179 + ], + "spans": [ + { + "bbox": [ + 313, + 72, + 553, + 179 + ], + "type": "text", + "content": "to a 2D space. The corresponding visualization for a sample input data is shown in Fig. 3-(a) using light Pruple points. Then, DivPrune is applied to select " + }, + { + "bbox": [ + 313, + 72, + 553, + 179 + ], + "type": "inline_equation", + "content": "10\\%" + }, + { + "bbox": [ + 313, + 72, + 553, + 179 + ], + "type": "text", + "content": " of the visual tokens (i.e., pruning " + }, + { + "bbox": [ + 313, + 72, + 553, + 179 + ], + "type": "inline_equation", + "content": "90\\%" + }, + { + "bbox": [ + 313, + 72, + 553, + 179 + ], + "type": "text", + "content": "). Additionally, FastV, as an importance-based token pruning method, which utilizes attention scores, is employed to prune with the same ratio. The selected subsets using DivPrune and FastV are shown with different markers in Fig. 3-(a). More examples are provided in the supplementary materials." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 313, + 180, + 554, + 299 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 180, + 554, + 299 + ], + "spans": [ + { + "bbox": [ + 313, + 180, + 554, + 299 + ], + "type": "text", + "content": "As the example in Fig. 3-(a) shows, the proposed method selects points from all the clusters that appeared in the projected space whereas FastV does not choose any samples from the upper cluster. So, our method achieves a better representation of the original points by including samples from all clusters. In addition, the FastV method selects many tokens that are very close to each other which increases redundancy among the selected set. On the other hand, our method reduces redundancy by pruning the closely similar tokens." + } + ] + } + ], + "index": 9 + }, + { + "type": "image", + "bbox": [ + 337, + 319, + 545, + 487 + ], + "blocks": [ + { + "bbox": [ + 337, + 319, + 545, + 487 + ], + "lines": [ + { + "bbox": [ + 337, + 319, + 545, + 487 + ], + "spans": [ + { + "bbox": [ + 337, + 319, + 545, + 487 + ], + "type": "image", + "image_path": "bcd637e8a9777724d51d6a026867414a49f7df61248c1d29c52d870ab11a666c.jpg" + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "image_body" + } + ], + "index": 10 + }, + { + "type": "image", + "bbox": [ + 323, + 490, + 545, + 670 + ], + "blocks": [ + { + "bbox": [ + 323, + 490, + 545, + 670 + ], + "lines": [ + { + "bbox": [ + 323, + 490, + 545, + 670 + ], + "spans": [ + { + "bbox": [ + 323, + 490, + 545, + 670 + ], + "type": "image", + "image_path": "e08094393d41b060f15c77ff4494e06b373fc39f3673c8a7cb3f436e59bcceeb.jpg" + } + ] + } + ], + "index": 11, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 313, + 677, + 553, + 709 + ], + "lines": [ + { + "bbox": [ + 313, + 677, + 553, + 709 + ], + "spans": [ + { + "bbox": [ + 313, + 677, + 553, + 709 + ], + "type": "text", + "content": "Figure 3. (a) t-SNE visualization of visual tokens for the original model, our method, and FastV. (b) Histogram of the Max-Min distance between the selected tokens over the SeedBench dataset." + } + ] + } + ], + "index": 12, + "angle": 0, + "type": "image_caption" + } + ], + "index": 11 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 315, + 757 + ], + "type": "text", + "content": "9396" + } + ] + } + ], + "index": 13 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 4 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 56, + 70, + 553, + 304 + ], + "blocks": [ + { + "bbox": [ + 56, + 70, + 553, + 304 + ], + "lines": [ + { + "bbox": [ + 56, + 70, + 553, + 304 + ], + "spans": [ + { + "bbox": [ + 56, + 70, + 553, + 304 + ], + "type": "table", + "html": "
MethodTFLOP (ratio %)COCO CIDErFlickr CIDErGQA EMMMB AccMME P-scoreMMMU AccNocaps CIDErOKVQA EMPOPE F1SQA EMSEEDB Acc
LLaVA 1.5.7BOriginal3.228 (100.00)1.100.7561.9664.09150636.441.0653.3985.8469.4166.17
VTW [21]0.603 (18.46)0.050.0338.9421.3168132.600.0318.6425.3565.2936.13
FastV [4]0.514 (15.69)0.060.0338.7320.6269632.000.0418.3232.8465.1535.69
Ours0.512 (15.63)0.960.6256.8559.19132835.890.9246.9886.0268.2759.47
PruMerge [36]Variable0.770.5051.3054.47125935.110.7341.7466.8968.9153.26
Ours*Variable0.910.5655.2558.16133035.440.8744.3883.0667.8757.88
FitPrune△ [47]0.513 (15.65)0.900.5652.3957.65119736.000.8642.5360.8968.0254.84
M3* [3]0.512 (15.63)1.000.6760.8165.81139131.800.9555.1286.3364.6564.93
PruMerge-LoRA*Variable0.960.6355.9659.88133434.890.9047.9977.1368.3257.93
LLaVA 1.5.13BOriginal6.281 (100.00)1.160.8063.3368.64152235.671.0958.2885.9972.8866.82
VTW [21]1.030 (16.16)0.080.0539.7121.9162232.100.0522.490.4066.2438.59
FastV [4]1.003 (15.73)0.380.1844.9837.8094235.110.3332.1430.0269.9644.95
Ours1.002 (15.71)1.000.6657.2963.40140734.890.9553.2983.4372.3462.04
PruMerge△ [36]Variable0.800.5352.0158.93125636.560.7749.1564.3672.5356.10
Ours*Variable0.940.5956.0961.77134434.890.9150.8679.6071.3460.00
LLaVA 1.6.7BOriginal11.849 (100.00)1.000.6864.2867.01152036.440.8844.2086.3870.1570.16
VTW [21]1.318 (11.23)0.060.0338.6219.7660631.300.038.667.1365.7437.48
FastV [4]1.327 (11.30)0.060.0338.7920.3661932.560.048.807.7865.4937.62
Ours1.266 (10.79)0.890.6158.6963.49136237.110.7641.9282.9768.5764.11
M3* [3]1.266 (10.79)1.010.6762.9769.16149035.000.8557.4987.4469.5168.49
", + "image_path": "8b36e11670e6102116b47e27d68b20495223ad97cf7e116b9460e536bcc344c4.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "table_body" + } + ], + "index": 0 + }, + { + "bbox": [ + 55, + 308, + 555, + 331 + ], + "lines": [ + { + "bbox": [ + 55, + 308, + 555, + 331 + ], + "spans": [ + { + "bbox": [ + 55, + 308, + 555, + 331 + ], + "type": "text", + "content": "Table 1. Comparison results of our method and different baselines on image-language understanding datasets. " + }, + { + "bbox": [ + 55, + 308, + 555, + 331 + ], + "type": "inline_equation", + "content": "\\bullet" + }, + { + "bbox": [ + 55, + 308, + 555, + 331 + ], + "type": "text", + "content": " : Finetuning is used, " + }, + { + "bbox": [ + 55, + 308, + 555, + 331 + ], + "type": "inline_equation", + "content": "\\triangle" + }, + { + "bbox": [ + 55, + 308, + 555, + 331 + ], + "type": "text", + "content": " Calibration dataset is used. Ours*: Our method matching the PruMerge selection ratio." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "text" + }, + { + "bbox": [ + 54, + 346, + 294, + 477 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 54, + 346, + 294, + 477 + ], + "spans": [ + { + "bbox": [ + 54, + 346, + 294, + 477 + ], + "type": "text", + "content": "In addition, the max-min distance (Eq. 3) for the selected subset of tokens is computed using 1000 randomly data samples from the SeedBench dataset and the histogram of the computed values is shown in Fig. 3-(b). As the plot indicates, the proposed method selects a subset where samples have a higher minimum pair-wise distance compared to the FastV method. Hence, our method achieves higher diversity among the selected tokens that have less redundancy compared to the ones chosen using FastV. We analyze the effect of the reduced diversity on task performance in the following sections." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 55, + 489, + 230, + 502 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 489, + 230, + 502 + ], + "spans": [ + { + "bbox": [ + 55, + 489, + 230, + 502 + ], + "type": "text", + "content": "4.3. Image-Language Understanding" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 55, + 509, + 295, + 592 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 509, + 295, + 592 + ], + "spans": [ + { + "bbox": [ + 55, + 509, + 295, + 592 + ], + "type": "text", + "content": "In this section, we compare DivPrune against baselines across various image-language understanding tasks, including open- and closed-ended QA, visual reasoning, and image captioning. Specifically, ScienceQA-IMG (SQA) [25], POPE [18], MME [9], MMB [24], GQA [14], MMMU [50], Flicker30k [31], SeedBench (SEEDB) [16], Nocaps [2], OKVQA [28], and COCO-2017 [20] are used." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 55, + 594, + 295, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 594, + 295, + 714 + ], + "spans": [ + { + "bbox": [ + 55, + 594, + 295, + 714 + ], + "type": "text", + "content": "In the first experiment, summarized in Tab. 1, we analyze an extreme compression scenario for three image-based LMMs by fixing the TFLOP ratio at approximately " + }, + { + "bbox": [ + 55, + 594, + 295, + 714 + ], + "type": "inline_equation", + "content": "15\\%" + }, + { + "bbox": [ + 55, + 594, + 295, + 714 + ], + "type": "text", + "content": ", wherever the baseline allows configuration to a fixed TFLOP ratio. Since PruMerge does not allow fixing the TFLOP ratio, we configure our approach (Ours*) to match the variable pruning corresponding to PruMerge for a fair comparison. In the top section of the table, we compare the results of various baselines for LLaVa 1.5-7B. Specifically, the baselines supporting LLaVA 1.5 are grouped into" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 313, + 346, + 555, + 622 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 346, + 555, + 622 + ], + "spans": [ + { + "bbox": [ + 313, + 346, + 555, + 622 + ], + "type": "text", + "content": "three categories: plug-and-play methods, those with a variable TFLOP ratio, and those requiring a calibration dataset or involving fine-tuning the LMMs. Among the plug-and-play methods, which are the focus of this work, our approach significantly outperforms both the VTW and FastV baselines across all datasets. This result holds despite using lower TFLOPs, clearly demonstrating the advantage of our method in this scenario. For instance, when DivPrune is used, the performance of LLaVA 1.5-7b decreases by " + }, + { + "bbox": [ + 313, + 346, + 555, + 622 + ], + "type": "inline_equation", + "content": "5.1\\%" + }, + { + "bbox": [ + 313, + 346, + 555, + 622 + ], + "type": "text", + "content": " on the GQA dataset and " + }, + { + "bbox": [ + 313, + 346, + 555, + 622 + ], + "type": "inline_equation", + "content": "4.9\\%" + }, + { + "bbox": [ + 313, + 346, + 555, + 622 + ], + "type": "text", + "content": " on the MMB dataset. In contrast, the VTW and FastV methods result in performance drops of at least " + }, + { + "bbox": [ + 313, + 346, + 555, + 622 + ], + "type": "inline_equation", + "content": "23.0\\%" + }, + { + "bbox": [ + 313, + 346, + 555, + 622 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 313, + 346, + 555, + 622 + ], + "type": "inline_equation", + "content": "42.8\\%" + }, + { + "bbox": [ + 313, + 346, + 555, + 622 + ], + "type": "text", + "content": " on these datasets, respectively. The performance gap between DivPrune and the baseline methods is even more pronounced in image captioning tasks. For example, the CIDEr score on the COCO dataset drops by approximately " + }, + { + "bbox": [ + 313, + 346, + 555, + 622 + ], + "type": "inline_equation", + "content": "95\\%" + }, + { + "bbox": [ + 313, + 346, + 555, + 622 + ], + "type": "text", + "content": " with VTW and FastV, but only by " + }, + { + "bbox": [ + 313, + 346, + 555, + 622 + ], + "type": "inline_equation", + "content": "12.7\\%" + }, + { + "bbox": [ + 313, + 346, + 555, + 622 + ], + "type": "text", + "content": " with DivPrune. Additionally, DivPrune, compared to the original model, shows less than a " + }, + { + "bbox": [ + 313, + 346, + 555, + 622 + ], + "type": "inline_equation", + "content": "2\\%" + }, + { + "bbox": [ + 313, + 346, + 555, + 622 + ], + "type": "text", + "content": " performance drop on the MMMU and SQA datasets and slightly enhances the original model's performance on the POPE dataset while reducing the TLOP ratio by " + }, + { + "bbox": [ + 313, + 346, + 555, + 622 + ], + "type": "inline_equation", + "content": "84.4\\%" + }, + { + "bbox": [ + 313, + 346, + 555, + 622 + ], + "type": "text", + "content": ". It is shown that removing redundant tokens in some datasets can improve the original model's performance [4]." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 313, + 630, + 556, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 630, + 556, + 715 + ], + "spans": [ + { + "bbox": [ + 313, + 630, + 556, + 715 + ], + "type": "text", + "content": "Next, in the variable scenario, the pruning ratio is determined dynamically. To ensure a fair comparison, we matched the pruning ratio with that of the PruMerge baseline, assuming the average sequence length for calculating the average TFLOPs across each dataset. As indicated by the results, our approach consistently outperforms PruMerge across all benchmarks, except one. Further, for" + } + ] + } + ], + "index": 7 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "text", + "content": "9397" + } + ] + } + ], + "index": 8 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 5 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 82, + 70, + 525, + 139 + ], + "blocks": [ + { + "bbox": [ + 82, + 70, + 525, + 139 + ], + "lines": [ + { + "bbox": [ + 82, + 70, + 525, + 139 + ], + "spans": [ + { + "bbox": [ + 82, + 70, + 525, + 139 + ], + "type": "table", + "html": "
TFLOPs (ratio %)ActivityNet Score/AccSeedBench AccVChatGPT ScoreNextQA WUPSEgoSch. AccMax GPU mem (GB)Prefill Time (sec)E2E Latency (sec)
Original6.539 (100)2.67 / 48.1038.72.1626.0541.814.060.3304.37
VTW [21]1.124 (16.97)1.61 / 26.8429.391.1918.6625.4213.630.1503.43
FastV [4]0.943 (14.20)1.95 / 33.9132.981.4422.5129.1413.570.1503.63
Ours0.937 (14.10)2.56 / 45.9037.001.9224.4839.7613.510.1613.39
", + "image_path": "a170543e01d67452d39d90be08b457183ead297ca55710cd571a50aa94bad980.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "table_body" + } + ], + "index": 0 + }, + { + "bbox": [ + 66, + 142, + 543, + 153 + ], + "lines": [ + { + "bbox": [ + 66, + 142, + 543, + 153 + ], + "spans": [ + { + "bbox": [ + 66, + 142, + 543, + 153 + ], + "type": "text", + "content": "Table 2. Comparison results of our method and baselines on LLaVA-NeXT-Video-7B across video-language understanding datasets." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "text" + }, + { + "bbox": [ + 55, + 170, + 294, + 242 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 170, + 294, + 242 + ], + "spans": [ + { + "bbox": [ + 55, + 170, + 294, + 242 + ], + "type": "text", + "content": "the baseline with calibration, we observe that our approach outperforms the FitPrune approach on nearly all datasets by up to " + }, + { + "bbox": [ + 55, + 170, + 294, + 242 + ], + "type": "inline_equation", + "content": "25.1\\%" + }, + { + "bbox": [ + 55, + 170, + 294, + 242 + ], + "type": "text", + "content": ", despite not using any calibration dataset. Finally, compared to baselines involving fine-tuning, our method achieves comparable or superior performance without requiring any fine-tuning." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 54, + 243, + 295, + 399 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 54, + 243, + 295, + 399 + ], + "spans": [ + { + "bbox": [ + 54, + 243, + 295, + 399 + ], + "type": "text", + "content": "The above experiment is repeated with LLaVa 1.5-13B model and the results are shown in the middle part of Tab. 1. The baselines that support this model are FastV, VTW, and PruMerge. As shown in the table, DivPrune outperforms the corresponding baselines in both plug-and-play and variable scenarios almost on all the tested datasets. For example, on the POPE dataset, DivPrune outperforms VTW, FastV, and PruMerge with F1 score improvements of " + }, + { + "bbox": [ + 54, + 243, + 295, + 399 + ], + "type": "inline_equation", + "content": "83\\%" + }, + { + "bbox": [ + 54, + 243, + 295, + 399 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 54, + 243, + 295, + 399 + ], + "type": "inline_equation", + "content": "53.4\\%" + }, + { + "bbox": [ + 54, + 243, + 295, + 399 + ], + "type": "text", + "content": ", and " + }, + { + "bbox": [ + 54, + 243, + 295, + 399 + ], + "type": "inline_equation", + "content": "15.2\\%" + }, + { + "bbox": [ + 54, + 243, + 295, + 399 + ], + "type": "text", + "content": ", respectively. Additionally, on the MMB dataset, DivPrune achieves higher accuracy rates of " + }, + { + "bbox": [ + 54, + 243, + 295, + 399 + ], + "type": "inline_equation", + "content": "41.5\\%" + }, + { + "bbox": [ + 54, + 243, + 295, + 399 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 54, + 243, + 295, + 399 + ], + "type": "inline_equation", + "content": "25.6\\%" + }, + { + "bbox": [ + 54, + 243, + 295, + 399 + ], + "type": "text", + "content": ", and " + }, + { + "bbox": [ + 54, + 243, + 295, + 399 + ], + "type": "inline_equation", + "content": "2.8\\%" + }, + { + "bbox": [ + 54, + 243, + 295, + 399 + ], + "type": "text", + "content": " compared to VTW, FastV, and PruMerge, respectively. This demonstrates that DivPrune generalizes effectively across models with varying numbers of parameters." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 56, + 401, + 295, + 640 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 401, + 295, + 640 + ], + "spans": [ + { + "bbox": [ + 56, + 401, + 295, + 640 + ], + "type": "text", + "content": "In the bottom part of Tab. 1, the results corresponding to LLaVA 1.6-7B model are shown. We used the same pruning ratio as for LLava 1.5. However, the lower TFLOP ratio is due to the large number of visual tokens in LLaVA 1.6. The results indicate that the performance of the model drops significantly when baseline pruning methods are applied. For example, the F1 score on the POPE dataset drops by " + }, + { + "bbox": [ + 56, + 401, + 295, + 640 + ], + "type": "inline_equation", + "content": "79\\%" + }, + { + "bbox": [ + 56, + 401, + 295, + 640 + ], + "type": "text", + "content": " with the baselines as compared to the original model, whereas the drop with DivPrune is only " + }, + { + "bbox": [ + 56, + 401, + 295, + 640 + ], + "type": "inline_equation", + "content": "3.4\\%" + }, + { + "bbox": [ + 56, + 401, + 295, + 640 + ], + "type": "text", + "content": ". DivPrune also maintains competitive performance compared to the original model across various datasets. Specifically, DivPrune shows only " + }, + { + "bbox": [ + 56, + 401, + 295, + 640 + ], + "type": "inline_equation", + "content": "3.5\\%" + }, + { + "bbox": [ + 56, + 401, + 295, + 640 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 56, + 401, + 295, + 640 + ], + "type": "inline_equation", + "content": "2.3\\%" + }, + { + "bbox": [ + 56, + 401, + 295, + 640 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 56, + 401, + 295, + 640 + ], + "type": "inline_equation", + "content": "3.4\\%" + }, + { + "bbox": [ + 56, + 401, + 295, + 640 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 56, + 401, + 295, + 640 + ], + "type": "inline_equation", + "content": "1.6\\%" + }, + { + "bbox": [ + 56, + 401, + 295, + 640 + ], + "type": "text", + "content": " drop in accuracy compared to the original model on the MMB, OKVQA, POPE, and SQA datasets, respectively, while reducing the TFLOP by " + }, + { + "bbox": [ + 56, + 401, + 295, + 640 + ], + "type": "inline_equation", + "content": "89\\%" + }, + { + "bbox": [ + 56, + 401, + 295, + 640 + ], + "type": "text", + "content": ". The results also demonstrate that pruning visual tokens with DivPrune enhances the original model's performance on the MMMU task. These results show that DivPrune generalizes across different models. Qualitative examples as well as results with additional datasets are provided in the supplementary materials." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 55, + 642, + 295, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 642, + 295, + 714 + ], + "spans": [ + { + "bbox": [ + 55, + 642, + 295, + 714 + ], + "type": "text", + "content": "Furthermore, we show the comparison of different baselines and our method across various TFLOP ratios. We plot the results in Fig. 1 where the y-axis represents average performance on four datasets, namely, COCO (CIDEr), OKVQA (Acc), POPE (F1), and MMBench (Acc). The range of the performance metric for all datasets is between" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 313, + 170, + 555, + 338 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 170, + 555, + 338 + ], + "spans": [ + { + "bbox": [ + 313, + 170, + 555, + 338 + ], + "type": "text", + "content": "0 and 1, except for the CIDEr metric, which has a maximum reported value of 1.10. On the x-axis, we only show the high compression scenario (TFLOP ratio " + }, + { + "bbox": [ + 313, + 170, + 555, + 338 + ], + "type": "inline_equation", + "content": "\\leq 45\\%" + }, + { + "bbox": [ + 313, + 170, + 555, + 338 + ], + "type": "text", + "content": "). As shown in the figure, our method significantly outperforms all the baselines, particularly in high compression scenarios (TFLOP " + }, + { + "bbox": [ + 313, + 170, + 555, + 338 + ], + "type": "inline_equation", + "content": "\\leq 25\\%" + }, + { + "bbox": [ + 313, + 170, + 555, + 338 + ], + "type": "text", + "content": "). Further, we notice a steep drop in performance of all baselines as the TFLOP ratio " + }, + { + "bbox": [ + 313, + 170, + 555, + 338 + ], + "type": "inline_equation", + "content": "\\rightarrow 10" + }, + { + "bbox": [ + 313, + 170, + 555, + 338 + ], + "type": "text", + "content": ", while our method falls more gracefully. This results in an increasing performance gap between our approach and the baselines at extreme compression levels. For higher TFLOP ratios almost all converge toward the original performance, with FitPrune slightly outperforming our approach by an insignificant margin. It is important to note that, unlike our method, FitPrune relies on a calibration dataset to prune tokens." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 313, + 346, + 487, + 360 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 346, + 487, + 360 + ], + "spans": [ + { + "bbox": [ + 313, + 346, + 487, + 360 + ], + "type": "text", + "content": "4.4. Video-Language Understanding" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 313, + 365, + 555, + 628 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 365, + 555, + 628 + ], + "spans": [ + { + "bbox": [ + 313, + 365, + 555, + 628 + ], + "type": "text", + "content": "In this section, LLaVA-NeXT-Video-7B [23], a video-based LMM is used to analyze the performance of the proposed method on various video-language understanding tasks. Specifically, we evaluate DivPrune using five datasets, namely, ActivityNet [48], SeedBench [16], VideoChatGPT (temporal) [26], NextQA [44], and EgoSchema [27]. FastV and VTW methods are chosen as the baselines. We tested DivPrune using the same pruning ratio as in the image understanding experiments. However, due to the higher number of visual tokens in the LLaVA-NeXT-Video model, this pruning ratio results in lower TFLOPs ratio. For the baselines, we match their TFLOPs with ours by selecting the smallest available TFLOP ratio that exceeds the TFLOPs of our method. The results for the original model, DivPrune, and the baselines are given in Tab. 2. As shown in the table, DivPrune outperforms both FastV and VTW by a significant margin. Specifically, DivPrune achieves upto " + }, + { + "bbox": [ + 313, + 365, + 555, + 628 + ], + "type": "inline_equation", + "content": "12\\%" + }, + { + "bbox": [ + 313, + 365, + 555, + 628 + ], + "type": "text", + "content": " higher accuracy than FastV and upto " + }, + { + "bbox": [ + 313, + 365, + 555, + 628 + ], + "type": "inline_equation", + "content": "19\\%" + }, + { + "bbox": [ + 313, + 365, + 555, + 628 + ], + "type": "text", + "content": " better than VTW on Video QA datasets including ActivityNet, SeedBench, and EgoSchema. DivPrune also outperforms both baselines on open-ended QA such as VideoChatGPT and NextQA by achieving higher GPT-assisted and WUPS scores." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 313, + 630, + 556, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 630, + 556, + 715 + ], + "spans": [ + { + "bbox": [ + 313, + 630, + 556, + 715 + ], + "type": "text", + "content": "Furthermore, our method achieves performance that is highly competitive compared to the original model without pruning despite using only " + }, + { + "bbox": [ + 313, + 630, + 556, + 715 + ], + "type": "inline_equation", + "content": "14.1\\%" + }, + { + "bbox": [ + 313, + 630, + 556, + 715 + ], + "type": "text", + "content": " of the original model's TFLOPs. This demonstrates the robustness of DivPrune, as it effectively generalizes to video LMMs. Notably, the performance gap between DivPrune and the original model without pruning narrows as the number of visual tokens in" + } + ] + } + ], + "index": 9 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 749, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 749, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 749, + 315, + 757 + ], + "type": "text", + "content": "9398" + } + ] + } + ], + "index": 10 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 6 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 56, + 70, + 294, + 136 + ], + "blocks": [ + { + "bbox": [ + 56, + 70, + 294, + 136 + ], + "lines": [ + { + "bbox": [ + 56, + 70, + 294, + 136 + ], + "spans": [ + { + "bbox": [ + 56, + 70, + 294, + 136 + ], + "type": "table", + "html": "
TFLOP (ratio %)MMB AccMMMU AccPOPE F1SQA EMAvg
Layer 0 (Ours)19.6159.1935.8986.0268.2762.34
Layer 119.6559.0234.8980.6767.1860.44
Layer 219.7054.9034.2269.2769.5656.99
Layer 319.8023.9732.6731.8265.9438.60
", + "image_path": "dba4db81d6d9d4394eeba6efee6de50488f16f29dd8c554a2c08776f263094b4.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "table_body" + } + ], + "index": 0 + }, + { + "bbox": [ + 55, + 168, + 295, + 192 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 168, + 295, + 192 + ], + "spans": [ + { + "bbox": [ + 55, + 168, + 295, + 192 + ], + "type": "text", + "content": "creases, indicating that DivPrune is more effective for the models with larger visual contexts." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 55, + 201, + 167, + 213 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 201, + 167, + 213 + ], + "spans": [ + { + "bbox": [ + 55, + 201, + 167, + 213 + ], + "type": "text", + "content": "4.5. Efficiency Analysis" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 55, + 219, + 295, + 471 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 219, + 295, + 471 + ], + "spans": [ + { + "bbox": [ + 55, + 219, + 295, + 471 + ], + "type": "text", + "content": "In this section, we analyze the efficiency of the proposed method using memory usage (i.e., max allocated memory), prefetch time, and end-to-end latency (E2E). For this experiment, VideoChatGPT dataset with 499 samples is used to obtain the average time and memory usage for LLaVA-NeXT-Video-7B model. The results are summarized on the right side of Tab. 2. The obtained results are compared against the original model, as well as the FastV and VTW baselines. As shown in the table, our approach requires approximately 400MB less memory than the original model, with memory usage comparable to the baselines. In terms of prefetch and E2E time, our approach is about " + }, + { + "bbox": [ + 55, + 219, + 295, + 471 + ], + "type": "inline_equation", + "content": "55\\%" + }, + { + "bbox": [ + 55, + 219, + 295, + 471 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 55, + 219, + 295, + 471 + ], + "type": "inline_equation", + "content": "22\\%" + }, + { + "bbox": [ + 55, + 219, + 295, + 471 + ], + "type": "text", + "content": " faster, respectively, compared to the original model. When compared to the baselines, our prefetch time is approximately " + }, + { + "bbox": [ + 55, + 219, + 295, + 471 + ], + "type": "inline_equation", + "content": "6 - 7\\%" + }, + { + "bbox": [ + 55, + 219, + 295, + 471 + ], + "type": "text", + "content": " longer, while the E2E time is " + }, + { + "bbox": [ + 55, + 219, + 295, + 471 + ], + "type": "inline_equation", + "content": "1 - 7\\%" + }, + { + "bbox": [ + 55, + 219, + 295, + 471 + ], + "type": "text", + "content": " shorter. The slight increase in prefetch time for our method compared to the baselines is due to the distance calculations (See Section 3.3), which are performed only once during the prefetch stage. In contrast, for baselines, the corresponding calculations for token pruning need to be done at each decoding step, resulting in longer E2E time." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 55, + 479, + 149, + 491 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 479, + 149, + 491 + ], + "spans": [ + { + "bbox": [ + 55, + 479, + 149, + 491 + ], + "type": "text", + "content": "4.6. Ablation Study" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 55, + 498, + 295, + 677 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 498, + 295, + 677 + ], + "spans": [ + { + "bbox": [ + 55, + 498, + 295, + 677 + ], + "type": "text", + "content": "In this section, we conduct an ablation study to analyze the impact of modifying various core components of our method. The ablation experiments are conducted with the LLaVA 1.5-7B model. First, we show the effect of pruning tokens inside the LLM in Tab. 3 using 5 datasets. By default, in our method, visual tokens are pruned before being passed to the first decoder layer in the LLM, which we refer to as 'Layer 0'. We also tested 'Layer 1' where the first layer is processed without pruning and the pruning is performed afterward. We further extended this approach by allowing tokens to pass through the first few layers unpruned and then pruning them after specific layers. As shown in the table, for a fixed TFLOP ratio of " + }, + { + "bbox": [ + 55, + 498, + 295, + 677 + ], + "type": "inline_equation", + "content": "19.61\\%" + }, + { + "bbox": [ + 55, + 498, + 295, + 677 + ], + "type": "text", + "content": ", pruning done by our method at layer 0 achieves higher task accuracies compared to pruning at layers 1, 2, and 3 of the LLM." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 55, + 677, + 296, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 677, + 296, + 713 + ], + "spans": [ + { + "bbox": [ + 55, + 677, + 296, + 713 + ], + "type": "text", + "content": "Furthermore, in Tab. 4, we provide an analysis of using alternative diversity measures for token pruning. The first three rows show the impact of choosing different dis" + } + ] + } + ], + "index": 7 + }, + { + "type": "table", + "bbox": [ + 315, + 71, + 553, + 154 + ], + "blocks": [ + { + "bbox": [ + 57, + 140, + 293, + 152 + ], + "lines": [ + { + "bbox": [ + 57, + 140, + 293, + 152 + ], + "spans": [ + { + "bbox": [ + 57, + 140, + 293, + 152 + ], + "type": "text", + "content": "Table 3. Ablation study on applying DivPrune at different layers." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 315, + 71, + 553, + 154 + ], + "lines": [ + { + "bbox": [ + 315, + 71, + 553, + 154 + ], + "spans": [ + { + "bbox": [ + 315, + 71, + 553, + 154 + ], + "type": "table", + "html": "
TFLOP (ratio %)MMB AccMMMU AccPOPE F1SQA EMAvg
Cosine (Ours)19.6159.1935.8986.0268.2762.34
l119.6159.7134.6785.4067.9761.94
l219.6159.9735.0085.6468.2762.22
Random19.6152.6634.5672.7866.6356.66
Min-Max19.6138.5733.1149.2665.2046.53
", + "image_path": "4e2068b36c653fdc668c9e4a0f0b79bca6e46954e3ed0268beeafd60a980a520.jpg" + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "table_body" + } + ], + "index": 8 + }, + { + "bbox": [ + 334, + 157, + 533, + 169 + ], + "lines": [ + { + "bbox": [ + 334, + 157, + 533, + 169 + ], + "spans": [ + { + "bbox": [ + 334, + 157, + 533, + 169 + ], + "type": "text", + "content": "Table 4. Ablation on using various diversity measures." + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "text" + }, + { + "bbox": [ + 313, + 187, + 553, + 258 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 187, + 553, + 258 + ], + "spans": [ + { + "bbox": [ + 313, + 187, + 553, + 258 + ], + "type": "text", + "content": "tance measures to quantify the similarity among tokens. It can be seen that all three similarity measures, cosine, " + }, + { + "bbox": [ + 313, + 187, + 553, + 258 + ], + "type": "inline_equation", + "content": "\\ell_1" + }, + { + "bbox": [ + 313, + 187, + 553, + 258 + ], + "type": "text", + "content": ", and " + }, + { + "bbox": [ + 313, + 187, + 553, + 258 + ], + "type": "inline_equation", + "content": "\\ell_2" + }, + { + "bbox": [ + 313, + 187, + 553, + 258 + ], + "type": "text", + "content": " perform comparably, with cosine (default setting) performing slightly better. This suggests that the choice of similarity measure does not significantly impact DivPrune's overall performance." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 313, + 260, + 555, + 511 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 260, + 555, + 511 + ], + "spans": [ + { + "bbox": [ + 313, + 260, + 555, + 511 + ], + "type": "text", + "content": "The last two rows in Tab. 4 show the effect of choosing alternative strategies of token selection other than the proposed Max-Min diversity-based solution (3). We tested random pruning as well as the Min-Max strategy where the maximum distance between the selected samples is minimized. The Min-Max strategy enforces high redundancy among the selected samples, resulting in reduced diversity. As results in the bottom part of Tab. 4 reveal that any deviation from our proposed selection strategy results in suboptimal performance. Specifically, the Min-Max strategy performs the worst, showing approximately " + }, + { + "bbox": [ + 313, + 260, + 555, + 511 + ], + "type": "inline_equation", + "content": "15.8\\%" + }, + { + "bbox": [ + 313, + 260, + 555, + 511 + ], + "type": "text", + "content": " lower performance compared to ours. This decline is due to the Min-Max approach selecting tokens that are highly similar to each other, resulting in less diversity among the selected visual tokens. Random selection provides some degree of diversity, but it performs " + }, + { + "bbox": [ + 313, + 260, + 555, + 511 + ], + "type": "inline_equation", + "content": "5.6\\%" + }, + { + "bbox": [ + 313, + 260, + 555, + 511 + ], + "type": "text", + "content": " worse than the proposed method because it cannot guarantee maximum diversity. This proves that redundancy of visual tokens leads to poor performance and diversity maximization is needed for optimal performance, corroborating the utility and need of the proposed diversity maximization in Eq. (3)." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 313, + 525, + 388, + 537 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 525, + 388, + 537 + ], + "spans": [ + { + "bbox": [ + 313, + 525, + 388, + 537 + ], + "type": "text", + "content": "5. Conclusion" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 313, + 545, + 555, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 545, + 555, + 713 + ], + "spans": [ + { + "bbox": [ + 313, + 545, + 555, + 713 + ], + "type": "text", + "content": "In this paper, we proposed a token pruning method based on a max-min diversity problem, called DivPrune. In the proposed method, maximum diversity is achieved among the selected tokens, resulting in reduced redundancy. By ensuring high diversity, the selected tokens provide a more representative subset of the original tokens, enabling effective performance even at high pruning ratios without requiring fine-tuning. Extensive experiments were conducted with multiple LMMs on image and video understanding tasks across 16 datasets. The results show that DivPrune achieves state-of-the-art accuracy on the tested datasets. DivPrune generalizes well to different model sizes and architectures, while also improving memory consumption and end-to-end latency for the tested LMMs." + } + ] + } + ], + "index": 13 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 315, + 757 + ], + "type": "text", + "content": "9399" + } + ] + } + ], + "index": 14 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 7 + }, + { + "para_blocks": [ + { + "bbox": [ + 56, + 71, + 115, + 83 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 71, + 115, + 83 + ], + "spans": [ + { + "bbox": [ + 56, + 71, + 115, + 83 + ], + "type": "text", + "content": "References" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 57, + 91, + 296, + 713 + ], + "type": "list", + "angle": 0, + "index": 14, + "blocks": [ + { + "bbox": [ + 61, + 91, + 296, + 145 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 61, + 91, + 296, + 145 + ], + "spans": [ + { + "bbox": [ + 61, + 91, + 296, + 145 + ], + "type": "text", + "content": "[1] Josh Achiam, Steven Adler, Sandhini Agarwal, Lama Ahmad, Ilge Akkaya, Florencia Leoni Aleman, Diogo Almeida, Janko Altenschmidt, Sam Altman, Shyamal Anadkat, et al. Gpt-4 technical report. arXiv preprint arXiv:2303.08774, 2023. 1" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 61, + 148, + 296, + 203 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 61, + 148, + 296, + 203 + ], + "spans": [ + { + "bbox": [ + 61, + 148, + 296, + 203 + ], + "type": "text", + "content": "[2] Harsh Agrawal, Karan Desai, Yufei Wang, Xinlei Chen, Rishabh Jain, Mark Johnson, Dhruv Batra, Devi Parikh, Stefan Lee, and Peter Anderson. nocaps: novel object captioning at scale. In Proceedings of the IEEE International Conference on Computer Vision, pages 8948-8957, 2019. 6" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 62, + 205, + 295, + 247 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 205, + 295, + 247 + ], + "spans": [ + { + "bbox": [ + 62, + 205, + 295, + 247 + ], + "type": "text", + "content": "[3] Mu Cai, Jianwei Yang, Jianfeng Gao, and Yong Jae Lee. Matryoshka multimodal models. Proceedings of the International Conference on Learning Representation, 2025. 2, 3, 4, 6" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 62, + 250, + 295, + 304 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 250, + 295, + 304 + ], + "spans": [ + { + "bbox": [ + 62, + 250, + 295, + 304 + ], + "type": "text", + "content": "[4] Liang Chen, Haozhe Zhao, Tianyu Liu, Shuai Bai, Junyang Lin, Chang Zhou, and Baobao Chang. An image is worth 1/2 tokens after layer 2: Plug-and-play inference acceleration for large vision-language models. In European Conference on Computer Vision (ECCV), 2024. 1, 2, 4, 5, 6, 7" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 62, + 306, + 294, + 360 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 306, + 294, + 360 + ], + "spans": [ + { + "bbox": [ + 62, + 306, + 294, + 360 + ], + "type": "text", + "content": "[5] Zesen Cheng, Sicong Leng, Hang Zhang, Yifei Xin, Xin Li, Guanzheng Chen, Yongxin Zhu, Wenqi Zhang, Ziyang Luo, Deli Zhao, et al. Videollama 2: Advancing spatial-temporal modeling and audio understanding in video-llms. arXiv preprint arXiv:2406.07476, 2024. 2" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 62, + 362, + 294, + 427 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 362, + 294, + 427 + ], + "spans": [ + { + "bbox": [ + 62, + 362, + 294, + 427 + ], + "type": "text", + "content": "[6] Wei-Lin Chiang, Zhuohan Li, Zi Lin, Ying Sheng, Zhanghao Wu, Hao Zhang, Lianmin Zheng, Siyuan Zhuang, Yonghao Zhuang, Joseph E Gonzalez, et al. Vicuna: An open-source chatbot impressing gpt-4 with " + }, + { + "bbox": [ + 62, + 362, + 294, + 427 + ], + "type": "inline_equation", + "content": "90\\%" + }, + { + "bbox": [ + 62, + 362, + 294, + 427 + ], + "type": "text", + "content": " * chatgpt quality. See https://vicuna.lmsys.org (accessed 14 April 2023), 2(3):6, 2023. 1" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 62, + 430, + 295, + 483 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 430, + 295, + 483 + ], + "spans": [ + { + "bbox": [ + 62, + 430, + 295, + 483 + ], + "type": "text", + "content": "[7] Krzysztof Choromanski, Valerii Likhosherstov, David Dohan, Xingyou Song, Andreea Gane, Tamas Sarlos, Peter Hawkins, Jared Davis, Afroz Mohiuddin, Lukasz Kaiser, et al. Rethinking attention with performers. arXiv preprint arXiv:2009.14794, 2020. 1" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 62, + 486, + 295, + 529 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 486, + 295, + 529 + ], + "spans": [ + { + "bbox": [ + 62, + 486, + 295, + 529 + ], + "type": "text", + "content": "[8] Tri Dao, Dan Fu, Stefano Ermon, Atri Rudra, and Christopher Ré. Flashattention: Fast and memory-efficient exact attention with io-awareness. Advances in Neural Information Processing Systems, 35:16344-16359, 2022. 1" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 62, + 532, + 295, + 585 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 532, + 295, + 585 + ], + "spans": [ + { + "bbox": [ + 62, + 532, + 295, + 585 + ], + "type": "text", + "content": "[9] Chaoyou Fu, Peixian Chen, Yunhang Shen, Yulei Qin, Mengdan Zhang, Xu Lin, Jinrui Yang, Xiawu Zheng, Ke Li, Xing Sun, et al. Mme: A comprehensive evaluation benchmark for multimodal large language models. arXiv preprint arXiv:2306.13394, 2023. 5, 6" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 57, + 588, + 294, + 610 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 588, + 294, + 610 + ], + "spans": [ + { + "bbox": [ + 57, + 588, + 294, + 610 + ], + "type": "text", + "content": "[10] Jinlan Fu, See-Kiong Ng, Zhengbao Jiang, and Pengfei Liu. Gptscore: Evaluate as you desire, 2023. 5" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 57, + 612, + 295, + 654 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 612, + 295, + 654 + ], + "spans": [ + { + "bbox": [ + 57, + 612, + 295, + 654 + ], + "type": "text", + "content": "[11] Mukul Gagrani, Raghavv Goel, Wonseok Jeon, Junyoung Park, Mingu Lee, and Christopher Lott. On speculative decoding for multimodal large language models. arXiv preprint arXiv:2404.08856, 2024. 2" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 57, + 657, + 295, + 689 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 657, + 295, + 689 + ], + "spans": [ + { + "bbox": [ + 57, + 657, + 295, + 689 + ], + "type": "text", + "content": "[12] Albert Gu and Tri Dao. Mamba: Linear-time sequence modeling with selective state spaces. arXiv preprint arXiv:2312.00752, 2023. 2" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 57, + 691, + 295, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 691, + 295, + 713 + ], + "spans": [ + { + "bbox": [ + 57, + 691, + 295, + 713 + ], + "type": "text", + "content": "[13] Zhiyu Guo, Hidetaka Kamigaito, and Taro Watanabe. Attention score is not all you need for token importance indicator" + } + ] + } + ], + "index": 13 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 316, + 73, + 554, + 713 + ], + "type": "list", + "angle": 0, + "index": 30, + "blocks": [ + { + "bbox": [ + 333, + 73, + 554, + 94 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 333, + 73, + 554, + 94 + ], + "spans": [ + { + "bbox": [ + 333, + 73, + 554, + 94 + ], + "type": "text", + "content": "in kv cache reduction: Value also matters. arXiv preprint arXiv:2406.12335, 2024. 2" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 317, + 96, + 554, + 139 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 96, + 554, + 139 + ], + "spans": [ + { + "bbox": [ + 317, + 96, + 554, + 139 + ], + "type": "text", + "content": "[14] Drew A Hudson and Christopher D Manning. Gqa: A new dataset for real-world visual reasoning and compositional question answering. Conference on Computer Vision and Pattern Recognition (CVPR), 2019. 6" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 317, + 140, + 553, + 193 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 140, + 553, + 193 + ], + "spans": [ + { + "bbox": [ + 317, + 140, + 553, + 193 + ], + "type": "text", + "content": "[15] Angelos Katharopoulos, Apoorv Vyas, Nikolaos Pappas, and François Fleuret. Transformers are rnns: Fast autoregressive transformers with linear attention. In International conference on machine learning, pages 5156-5165. PMLR, 2020. 1" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 316, + 195, + 553, + 237 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 195, + 553, + 237 + ], + "spans": [ + { + "bbox": [ + 316, + 195, + 553, + 237 + ], + "type": "text", + "content": "[16] Bohao Li, Rui Wang, Guangzhi Wang, Yuying Ge, Yixiao Ge, and Ying Shan. Seed-bench: Benchmarking multimodal llms with generative comprehension. arXiv preprint arXiv:2307.16125, 2023. 5, 6, 7" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 316, + 239, + 553, + 282 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 239, + 553, + 282 + ], + "spans": [ + { + "bbox": [ + 316, + 239, + 553, + 282 + ], + "type": "text", + "content": "[17] Wentong Li, Yuqian Yuan, Jian Liu, Dongqi Tang, Song Wang, Jianke Zhu, and Lei Zhang. Tokenpacker: Efficient visual projector for multimodal llm. arXiv preprint arXiv:2407.02392, 2024. 2, 3" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 316, + 284, + 553, + 326 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 284, + 553, + 326 + ], + "spans": [ + { + "bbox": [ + 316, + 284, + 553, + 326 + ], + "type": "text", + "content": "[18] Yifan Li, Yifan Du, Kun Zhou, Jinpeng Wang, Wayne Xin Zhao, and Ji-Rong Wen. Evaluating object hallucination in large vision-language models. arXiv preprint arXiv:2305.10355, 2023. 6" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 316, + 327, + 553, + 369 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 327, + 553, + 369 + ], + "spans": [ + { + "bbox": [ + 316, + 327, + 553, + 369 + ], + "type": "text", + "content": "[19] Bin Lin, Yang Ye, Bin Zhu, Jiaxi Cui, Munan Ning, Peng Jin, and Li Yuan. Video-llava: Learning united visual representation by alignment before projection. CoRR, abs/2311.10122, 2023. 1, 2" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 316, + 372, + 553, + 415 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 372, + 553, + 415 + ], + "spans": [ + { + "bbox": [ + 316, + 372, + 553, + 415 + ], + "type": "text", + "content": "[20] Tsung-Yi Lin, Michael Maire, Serge Belongie, Lubomir Bourdev, Ross Girshick, James Hays, Pietro Perona, Deva Ramanan, C. Lawrence Zitnick, and Piotr Dólar. Microsoft coco: Common objects in context, 2015. 6" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 316, + 416, + 553, + 458 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 416, + 553, + 458 + ], + "spans": [ + { + "bbox": [ + 316, + 416, + 553, + 458 + ], + "type": "text", + "content": "[21] Zhihang Lin, Mingbao Lin, Luxi Lin, and Rongrong Ji. Boosting multimodal large language models with visual tokens withdrawal for rapid inference. arXiv preprint arXiv:2405.05803, 2024. 2, 4, 5, 6, 7" + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 316, + 460, + 553, + 481 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 460, + 553, + 481 + ], + "spans": [ + { + "bbox": [ + 316, + 460, + 553, + 481 + ], + "type": "text", + "content": "[22] Haotian Liu, Chunyuan Li, Qingyang Wu, and Yong Jae Lee. Visual instruction tuning. In NeurIPS, 2023. 1, 2, 4" + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 317, + 483, + 553, + 525 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 483, + 553, + 525 + ], + "spans": [ + { + "bbox": [ + 317, + 483, + 553, + 525 + ], + "type": "text", + "content": "[23] Haotian Liu, Chunyuan Li, Yuheng Li, Bo Li, Yuanhan Zhang, Sheng Shen, and Yong Jae Lee. Llava-last: Improved reasoning,OCR,and world knowledge,2024.1,2, 4,7" + } + ] + } + ], + "index": 25 + }, + { + "bbox": [ + 317, + 526, + 553, + 581 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 526, + 553, + 581 + ], + "spans": [ + { + "bbox": [ + 317, + 526, + 553, + 581 + ], + "type": "text", + "content": "[24] Yuan Liu, Haodong Duan, Yuanhan Zhang, Bo Li, Songyang Zhang, Wangbo Zhao, Yike Yuan, Jiaqi Wang, Conghui He, Ziwei Liu, et al. Mmbench: Is your multi-modal model an all-around player? In European Conference on Computer Vision, pages 216-233. Springer, 2025. 6" + } + ] + } + ], + "index": 26 + }, + { + "bbox": [ + 317, + 582, + 553, + 624 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 582, + 553, + 624 + ], + "spans": [ + { + "bbox": [ + 317, + 582, + 553, + 624 + ], + "type": "text", + "content": "[25] Pan Lu, Swaroop Mishra, Tony Xia, Liang Qiu, Kai-Wei Chang, Song-Chun Zhu, Oyvind Tafjord, Peter Clark, and Ashwin Kalyan. Learn to explain: Multimodal reasoning via thought chains for science question answering, 2022. 6" + } + ] + } + ], + "index": 27 + }, + { + "bbox": [ + 317, + 625, + 553, + 680 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 625, + 553, + 680 + ], + "spans": [ + { + "bbox": [ + 317, + 625, + 553, + 680 + ], + "type": "text", + "content": "[26] Muhammad Maaz, Hanoona Rasheed, Salman Khan, and Fahad Shahbaz Khan. Video-chatgpt: Towards detailed video understanding via large vision and language models. In Proceedings of the 62nd Annual Meeting of the Association for Computational Linguistics (ACL 2024), 2024. 7" + } + ] + } + ], + "index": 28 + }, + { + "bbox": [ + 316, + 681, + 553, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 681, + 553, + 713 + ], + "spans": [ + { + "bbox": [ + 316, + 681, + 553, + 713 + ], + "type": "text", + "content": "[27] Karttikeya Mangalam, Raiymbek Akshulakov, and Jitendra Malik. Egoschema: A diagnostic benchmark for very long-form video language understanding, 2023. 7" + } + ] + } + ], + "index": 29 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "text", + "content": "9400" + } + ] + } + ], + "index": 31 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 8 + }, + { + "para_blocks": [ + { + "bbox": [ + 56, + 72, + 294, + 713 + ], + "type": "list", + "angle": 0, + "index": 14, + "blocks": [ + { + "bbox": [ + 56, + 72, + 294, + 117 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 72, + 294, + 117 + ], + "spans": [ + { + "bbox": [ + 56, + 72, + 294, + 117 + ], + "type": "text", + "content": "[28] Kenneth Marino, Mohammad Rastegari, Ali Farhadi, and Roozbeh Mottaghi. Ok-vqa: A visual question answering benchmark requiring external knowledge. In Conference on Computer Vision and Pattern Recognition (CVPR), 2019. 6" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 56, + 119, + 294, + 163 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 119, + 294, + 163 + ], + "spans": [ + { + "bbox": [ + 56, + 119, + 294, + 163 + ], + "type": "text", + "content": "[29] Rafael Martí, Anna Martínez-Gavara, Sergio Pérez-Peló, and Jesús Sánchez-Oro. A review on discrete diversity and dispersion maximization from an or perspective. European Journal of Operational Research, 299(3):795-813, 2022. 4" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 57, + 165, + 294, + 187 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 165, + 294, + 187 + ], + "spans": [ + { + "bbox": [ + 57, + 165, + 294, + 187 + ], + "type": "text", + "content": "[30] OpenAI. Hello gpt-4o, 2024. https://openai.com/index/hello-gpt-4o/ [Accessed: (Nov 2024)]. 2" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 56, + 189, + 294, + 243 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 189, + 294, + 243 + ], + "spans": [ + { + "bbox": [ + 56, + 189, + 294, + 243 + ], + "type": "text", + "content": "[31] Bryan A. Plummer, Liwei Wang, Christopher M. Cervantes, Juan C. Caicedo, Julia Hockenmaier, and Svetlana Lazebnik. Flickr30k entities: Collecting region-to-phrase correspondences for richer image-to-sentence models. *IJCV*, 123 (1):74–93, 2017. 6" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 56, + 245, + 294, + 289 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 245, + 294, + 289 + ], + "spans": [ + { + "bbox": [ + 56, + 245, + 294, + 289 + ], + "type": "text", + "content": "[32] Daniel Cosmin Porumbel, Jin-Kao Hao, and Fred Glover. A simple and effective algorithm for the maxmin diversity problem. Annals of Operations Research, 186:275-293, 2011. 3" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 56, + 291, + 294, + 335 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 291, + 294, + 335 + ], + "spans": [ + { + "bbox": [ + 56, + 291, + 294, + 335 + ], + "type": "text", + "content": "[33] Yanyuan Qiao, Zheng Yu, Longteng Guo, Sihan Chen, Zijia Zhao, Mingzhen Sun, Qi Wu, and Jing Liu. Vl-mamba: Exploring state space models for multimodal learning. arXiv preprint arXiv:2403.13600, 2024. 2" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 56, + 337, + 294, + 403 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 337, + 294, + 403 + ], + "spans": [ + { + "bbox": [ + 56, + 337, + 294, + 403 + ], + "type": "text", + "content": "[34] Alec Radford, Jong Wook Kim, Chris Hallacy, Aditya Ramesh, Gabriel Goh, Sandhini Agarwal, Girish Sastry, Amanda Askell, Pamela Mishkin, Jack Clark, et al. Learning transferable visual models from natural language supervision. In International conference on machine learning, pages 8748-8763. PMLR, 2021. 4" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 56, + 404, + 294, + 449 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 404, + 294, + 449 + ], + "spans": [ + { + "bbox": [ + 56, + 404, + 294, + 449 + ], + "type": "text", + "content": "[35] Mauricio GC Resende, Rafael Martí, Micael Gallego, and Abraham Duarte. Grasp and path relinking for the max-min diversity problem. Computers & Operations Research, 37 (3):498-508, 2010. 2, 4" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 56, + 450, + 294, + 495 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 450, + 294, + 495 + ], + "spans": [ + { + "bbox": [ + 56, + 450, + 294, + 495 + ], + "type": "text", + "content": "[36] Yuzhang Shang, Mu Cai, Bingxin Xu, Yong Jae Lee, and Yan Yan. Llava-prumerge: Adaptive token reduction for efficient large multimodal models. arXiv preprint arXiv:2403.15388, 2024. 1, 2, 4, 6" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 56, + 497, + 294, + 529 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 497, + 294, + 529 + ], + "spans": [ + { + "bbox": [ + 56, + 497, + 294, + 529 + ], + "type": "text", + "content": "[37] Mustafa Shukor and Matthieu Cord. Skipping computations in multimodal llms. arXiv preprint arXiv:2410.09454, 2024. 2" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 56, + 531, + 294, + 564 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 531, + 294, + 564 + ], + "spans": [ + { + "bbox": [ + 56, + 531, + 294, + 564 + ], + "type": "text", + "content": "[38] Sainbayar Sukhbaatar, Edouard Grave, Piotr Bojanowski, and Armand Joulin. Adaptive attention span in transformers. arXiv preprint arXiv:1905.07799, 2019. 1" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 56, + 567, + 294, + 621 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 567, + 294, + 621 + ], + "spans": [ + { + "bbox": [ + 56, + 567, + 294, + 621 + ], + "type": "text", + "content": "[39] Gemini Team, Petko Georgiev, Ving Ian Lei, Ryan Burnell, Libin Bai, Anmol Gulati, Garrett Tanzer, Damien Vincent, Zhufeng Pan, Shibo Wang, et al. Gemini 1.5: Unlocking multimodal understanding across millions of tokens of context. arXiv preprint arXiv:2403.05530, 2024. 2" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 56, + 623, + 294, + 678 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 623, + 294, + 678 + ], + "spans": [ + { + "bbox": [ + 56, + 623, + 294, + 678 + ], + "type": "text", + "content": "[40] Hugo Touvron, Thibaut Lavril, Gautier Izacard, Xavier Martinet, Marie-Anne Lachaux, Timothee Lacroix, Baptiste Roziere, Naman Goyal, Eric Hambro, Faisal Azhar, et al. Llama: Open and efficient foundation language models. arXiv preprint arXiv:2302.13971, 2023. 1" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 56, + 680, + 294, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 680, + 294, + 713 + ], + "spans": [ + { + "bbox": [ + 56, + 680, + 294, + 713 + ], + "type": "text", + "content": "[41] Laurens Van der Maaten and Geoffrey Hinton. Visualizing data using t-sne. Journal of machine learning research, 9 (11), 2008. 5" + } + ] + } + ], + "index": 13 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 315, + 72, + 554, + 654 + ], + "type": "list", + "angle": 0, + "index": 27, + "blocks": [ + { + "bbox": [ + 315, + 72, + 554, + 117 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 315, + 72, + 554, + 117 + ], + "spans": [ + { + "bbox": [ + 315, + 72, + 554, + 117 + ], + "type": "text", + "content": "[42] Ramakrishna Vedantam, C. Lawrence Zitnick, and Devi Parikh. Cider: Consensus-based image description evaluation. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition (CVPR), 2015. 5" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 315, + 118, + 553, + 140 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 315, + 118, + 553, + 140 + ], + "spans": [ + { + "bbox": [ + 315, + 118, + 553, + 140 + ], + "type": "text", + "content": "[43] Zhibiao Wu and Martha Palmer. Verb semantics and lexical selection. arXiv preprint cmp-lg/9406033, 1994. 5" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 316, + 141, + 553, + 195 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 141, + 553, + 195 + ], + "spans": [ + { + "bbox": [ + 316, + 141, + 553, + 195 + ], + "type": "text", + "content": "[44] Junbin Xiao, Xindi Shang, Angela Yao, and Tat-Seng Chua. Next-qa: Next phase of question-answering to explaining temporal actions. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 9777–9786, 2021. 7" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 316, + 197, + 553, + 239 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 197, + 553, + 239 + ], + "spans": [ + { + "bbox": [ + 316, + 197, + 553, + 239 + ], + "type": "text", + "content": "[45] Shilin Xu, Xiangtai Li, Haobo Yuan, Lu Qi, Yunhai Tong, and Ming-Hsuan Yang. Llavadi: What matters for multimodal large language models distillation. arXiv preprint arXiv:2407.19409, 2024. 2" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 316, + 241, + 553, + 285 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 241, + 553, + 285 + ], + "spans": [ + { + "bbox": [ + 316, + 241, + 553, + 285 + ], + "type": "text", + "content": "[46] Yuan Yao, Tianyu Yu, Ao Zhang, Chongyi Wang, Junbo Cui, Hongji Zhu, Tianchi Cai, Haoyu Li, Weilin Zhao, Zhihui He, et al. Minicpm-v: A gpt-4v level mllm on your phone. arXiv preprint arXiv:2408.01800, 2024. 1" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 316, + 286, + 553, + 330 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 286, + 553, + 330 + ], + "spans": [ + { + "bbox": [ + 316, + 286, + 553, + 330 + ], + "type": "text", + "content": "[47] Weihao Ye, Qiong Wu, Wenhao Lin, and Yiyi Zhou. Fit and prune: Fast and training-free visual token pruning for multi-modal large language models. arXiv preprint arXiv:2409.10197, 2024. 1, 2, 4, 5, 6" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 316, + 331, + 553, + 375 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 331, + 553, + 375 + ], + "spans": [ + { + "bbox": [ + 316, + 331, + 553, + 375 + ], + "type": "text", + "content": "[48] Zhou Yu, Dejing Xu, Jun Yu, Ting Yu, Zhou Zhao, Yueting Zhuang, and Dacheng Tao. Activitynet-qa: A dataset for understanding complex web videos via question answering. In AAAI, pages 9127–9134, 2019. 7" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 316, + 376, + 553, + 419 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 376, + 553, + 419 + ], + "spans": [ + { + "bbox": [ + 316, + 376, + 553, + 419 + ], + "type": "text", + "content": "[49] Zhengqing Yuan, Zhaoxu Li, Weiran Huang, Yanfang Ye, and Lichao Sun. Tinygpt-v: Efficient multimodal large language model via small backbones. arXiv preprint arXiv:2312.16862, 2023. 2" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 316, + 420, + 553, + 509 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 420, + 553, + 509 + ], + "spans": [ + { + "bbox": [ + 316, + 420, + 553, + 509 + ], + "type": "text", + "content": "[50] Xiang Yue, Yuansheng Ni, Kai Zhang, Tianyu Zheng, Ruoqi Liu, Ge Zhang, Samuel Stevens, Dongfu Jiang, Weiming Ren, Yuxuan Sun, Cong Wei, Botao Yu, Ruibin Yuan, Ren-liang Sun, Ming Yin, Boyuan Zheng, Zhenzhu Yang, Yibo Liu, Wenhao Huang, Huan Sun, Yu Su, and Wenhui Chen. Mmmu: A massive multi-discipline multimodal understanding and reasoning benchmark for expert agi. In Proceedings of CVPR, 2024. 6" + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 316, + 510, + 553, + 564 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 510, + 553, + 564 + ], + "spans": [ + { + "bbox": [ + 316, + 510, + 553, + 564 + ], + "type": "text", + "content": "[51] Kaichen Zhang, Bo Li, Peiyuan Zhang, Fanyi Pu, Joshua Adrian Cahyono, Kairui Hu, Shuai Liu, Yuanhan Zhang, Jingkang Yang, Chunyuan Li, and Ziwei Liu. Lmmseval: Reality check on the evaluation of large multimodal models, 2024. 5" + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 316, + 565, + 553, + 609 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 565, + 553, + 609 + ], + "spans": [ + { + "bbox": [ + 316, + 565, + 553, + 609 + ], + "type": "text", + "content": "[52] Yuanhan Zhang, Bo Li, haotian Liu, Yong jae Lee, Liangke Gui, Di Fu, Jiashi Feng, Ziwei Liu, and Chunyuan Li. Llavanext: A strong zero-shot video understanding model, 2024. 1, 2, 4" + } + ] + } + ], + "index": 25 + }, + { + "bbox": [ + 316, + 609, + 553, + 654 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 609, + 553, + 654 + ], + "spans": [ + { + "bbox": [ + 316, + 609, + 553, + 654 + ], + "type": "text", + "content": "[53] Baichuan Zhou, Ying Hu, Xi Weng, Junlong Jia, Jie Luo, Xien Liu, Ji Wu, and Lei Huang. Tinyllava: A framework of small-scale large multimodal models. arXiv preprint arXiv:2402.14289, 2024. 2" + } + ] + } + ], + "index": 26 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 748, + 314, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 314, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 314, + 757 + ], + "type": "text", + "content": "9401" + } + ] + } + ], + "index": 28 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 9 + } + ], + "_backend": "vlm", + "_version_name": "2.6.4" +} \ No newline at end of file diff --git a/2025/DiverseFlow_ Sample-Efficient Diverse Mode Coverage in Flows/276f7b10-7b9f-439d-a85a-8ad9f72195cf_content_list.json b/2025/DiverseFlow_ Sample-Efficient Diverse Mode Coverage in Flows/276f7b10-7b9f-439d-a85a-8ad9f72195cf_content_list.json new file mode 100644 index 0000000000000000000000000000000000000000..b64b2a9966b152748a44265bc7e11a30fed82adb --- /dev/null +++ b/2025/DiverseFlow_ Sample-Efficient Diverse Mode Coverage in Flows/276f7b10-7b9f-439d-a85a-8ad9f72195cf_content_list.json @@ -0,0 +1,1734 @@ +[ + { + "type": "text", + "text": "DiverseFlow: Sample-Efficient Diverse Mode Coverage in Flows", + "text_level": 1, + "bbox": [ + 176, + 130, + 821, + 152 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Mashrur M. Morshed Vishnu Boddeti", + "bbox": [ + 321, + 181, + 673, + 196 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Michigan State University", + "bbox": [ + 393, + 199, + 602, + 217 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "{morshedm, vishnu}@msu.edu", + "bbox": [ + 382, + 219, + 611, + 232 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Abstract", + "text_level": 1, + "bbox": [ + 248, + 268, + 326, + 284 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Many real-world applications of flow-based generative models desire a diverse set of samples that cover multiple modes of the target distribution. However, the predominant approach for obtaining diverse sets is not sample-efficient, as it involves independently obtaining many samples from the source distribution and mapping them through the flow until the desired mode coverage is achieved. As an alternative to repeated sampling, we introduce DiverseFlow: a training-free approach to improve the diversity of flow models. Our key idea is to employ a determinantal point process to induce a coupling between the samples that drives diversity under a fixed sampling budget. In essence, DiverseFlow allows exploration of more variations in a learned flow model with fewer samples. We demonstrate the efficacy of our method for tasks where sample-efficient diversity is desirable, such as text-guided image generation with polysemous words, inverse problems like large-hole inpainting, and class-conditional image synthesis.", + "bbox": [ + 88, + 301, + 485, + 573 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "1. Introduction", + "text_level": 1, + "bbox": [ + 89, + 604, + 220, + 619 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Consider the task of text-guided image generation from open-ended prompts, like \"A famous boxer\". If we use a generative ordinary differential equation (ODE) to obtain a few images for this given prompt, we may observe something unusual: all the resultant images depict a dog, as shown in Figure 1 (top-left). Such a result is plausible, as the word \"boxer\" has multiple meanings: it can either mean a combat sport athlete or a particular dog breed. However, for the prompt \"A famous boxer\", what if the output desired by the user was actually a human athlete, and not a dog? This situation necessitates obtaining additional samples from the model, until the desired alternate meanings are discovered. But instead of repeated sampling, can we directly observe more meanings by finding a more diverse set?", + "bbox": [ + 89, + 631, + 483, + 840 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Beyond the aforementioned example of text-to-image generation from polysemous1 prompts, sample diversity is a", + "bbox": [ + 89, + 843, + 483, + 875 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/2c6c83f27981f75e8aed186ad9578fdd73f6e56a46f9aedbd43cad7a9e80cbc1.jpg", + "image_caption": [ + "Prompt: \"A famous boxer\"" + ], + "image_footnote": [], + "bbox": [ + 535, + 282, + 705, + 415 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/10c363bc192d79a1cc897b1f19467d637245fed58d2a78e9b6a278eab156cfb0.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 714, + 284, + 883, + 415 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/3a64acd2ec62494943ef07762881d650a82afc857674e2990bd37cf3542485b1.jpg", + "image_caption": [ + "Prompt: \"A letter\"", + "ID", + "Figure 1. Text-guided image generation with polysemous words. With the same budget of samples, DiverseFlow (right) finds a more diverse set of results compared to IID sampling (left)." + ], + "image_footnote": [], + "bbox": [ + 535, + 434, + 705, + 564 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/c156a42342a6e920569d9a0c73cb438d810e79299180ef4b8e2d5515044e4ff2.jpg", + "image_caption": [ + "DiverseFlow" + ], + "image_footnote": [], + "bbox": [ + 714, + 434, + 883, + 564 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "desirable objective for many other tasks that use generative models. These include inverse problems (e.g., large hole filling) and class-conditioned image generation, to name a few. Diversity or mode coverage is a key pillar in the generative learning trilemma [32], in addition to fidelity and latency. For state-of-the-art generative methods such as flow matching models (FM) [20, 22] and diffusion models (DM) [12, 26], significant work has been done on improving the photorealism of samples and the efficiency of the sampling process [11, 16, 20, 28, 30, 34]. However, relatively little attention has been paid to explicitly enhancing the diversity of generated samples under a limited sampling budget.", + "bbox": [ + 511, + 657, + 906, + 838 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "The standard approach to generate a diverse set of images is to repeatedly obtain independent and identically distributed (IID) samples from a simple or tractable source distribution (e.g., Gaussian distribution), map them to samples", + "bbox": [ + 511, + 839, + 908, + 901 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "CVF", + "bbox": [ + 106, + 2, + 181, + 42 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore.", + "bbox": [ + 236, + 0, + 810, + 46 + ], + "page_idx": 0 + }, + { + "type": "page_footnote", + "text": "1Words or phrases with several meanings.", + "bbox": [ + 109, + 886, + 331, + 900 + ], + "page_idx": 0 + }, + { + "type": "page_number", + "text": "23303", + "bbox": [ + 478, + 944, + 517, + 955 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "in the target distribution, and continue this process until we observe sufficient mode coverage in the target distribution. This process, while effective, is sample-inefficient, requiring the generation of more results than necessary. Importantly, the mapping from the source to the target density does not hold a linear relationship; even specifically selecting diverse samples from the source distribution by design does not necessarily yield diverse samples in the target distribution. These limitations naturally raise the following research question.", + "bbox": [ + 89, + 90, + 483, + 241 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "How can we generate diverse samples from the target density under a limited sampling budget?", + "bbox": [ + 114, + 266, + 459, + 297 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "In this paper, we propose DiverseFlow, an inference-time, training-free approach to obtain a diverse set of samples in a desired target density under a fixed sampling budget. We focus on deterministic ODE sampling in continuous-time generative models, specifically FMs, an emerging generative paradigm that enables simulation-free training of continuous normalizing flows (CNFs) and includes diffusion models as a special case.", + "bbox": [ + 89, + 320, + 483, + 441 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "DiverseFlow measures the diversity of a set of samples through the volume they span in the target space. A set of similar samples span a lower volume, while a diverse set naturally spans a larger volume. We impose a volume-based gradient constraint on the flow ODE by drawing on determinantal point processes (DPP) [17, 23], a probabilistic model arising from quantum physics that exactly describes the Pauli exclusion principle: that no two fermions may occupy the same quantum state. In Figure 1 consider the images generated via IID sampling (left column) from a text-to-image generative ODE—for both the prompts \"A famous boxer\" and \"A letter\", only a single potential meaning is discovered. Unlike the results generated via IID sampling, samples obtained from DiverseFlow (right column) span more diverse modes corresponding to the polysemous words in the prompts. For \"A famous boxer\", the samples show both a dog breed and a human athlete; for \"A letter\", we observe both written correspondence and alphabet symbols.", + "bbox": [ + 89, + 444, + 483, + 717 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "We empirically demonstrate the utility of DiverseFlow across several applications where diversity is inherently desirable. First, we use DiverseFlow to perform text-guided image synthesis for words and phrases that may carry a variety of meanings. Second, we perform large-hole face inpainting with occlusion masks covering significant regions of the face that may be important to the person's identity. Third, we apply DiverseFlow on class-conditioned image synthesis and demonstrate that we can more efficiently explore the data space compared to IID sampling. Lastly, to better characterize and explain the behavior of DiverseFlow, we perform several experiments on synthetic 2D densities.", + "bbox": [ + 89, + 719, + 483, + 900 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Summary of Contributions", + "text_level": 1, + "bbox": [ + 524, + 98, + 718, + 111 + ], + "page_idx": 1 + }, + { + "type": "list", + "sub_type": "text", + "list_items": [ + "1. We present a sample-efficient method to obtain a diverse set of samples from a flow ODE in Section 5 and demonstrate it qualitatively in Sections 6.1 to 6.3 and quantitatively in Sections 6.4 and 6.5", + "2. We provide an empirical analysis that demonstrates that our method is consistent across various flow matching formulations (Section 6.4)", + "3. We introduce the task of text-to-image synthesis from prompts with polysemous words in the context of analyzing sample diversity, and provide a comparison with relevant methods (Section 6.5)" + ], + "bbox": [ + 527, + 113, + 895, + 279 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "2. Preliminaries", + "text_level": 1, + "bbox": [ + 513, + 308, + 651, + 323 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "2.1. Flow Matching", + "text_level": 1, + "bbox": [ + 513, + 334, + 666, + 349 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Many generative models can be considered as a transport map from some easy-to-sample source distribution to an empirically observed yet unknown target distribution. Recent successes in generative modeling represent this transport map in the form of continuous-time processes, such as stochastic differential equations (SDEs) [12, 29], or ordinary differential equations (ODEs) [2, 20, 22]. Although diffusion models are formulated as SDEs, a significant body of research focuses on converting the diffusion SDE to a deterministic ODE at inference time for faster inference. The diffusion ODE, or probability flow ODE, is a particular case of continuous normalizing flows (CNFs). Flow Matching (FM) [2, 20, 22] is motivated by the idea of directly training CNFs in a scalable and simulation-free manner, just like diffusion models. Moreover, many recent text-to-image generative models, such as Stable Diffusion 3 [10], adopt the FM framework. As such, we present our approach primarily in the context of FM, and our findings can be extended to diffusion and score-based generative models in a straightforward manner.", + "bbox": [ + 511, + 357, + 906, + 657 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "A CNF reshapes a prior source density $p_0$ to the empirically observed target density $p_1$ with an ODE of the form:", + "bbox": [ + 511, + 660, + 906, + 690 + ], + "page_idx": 1 + }, + { + "type": "equation", + "text": "\n$$\nd \\mathbf {x} _ {t} = v _ {\\theta} \\left(\\mathbf {x} _ {t}, t\\right) d t, \\quad \\mathbf {x} _ {0} \\sim p _ {0} \\tag {1}\n$$\n", + "text_format": "latex", + "bbox": [ + 607, + 708, + 906, + 723 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "where $v_{\\theta}$ is a time-dependent velocity field whose parameters $\\theta$ are learned; we interchangeably use the notation $v_{t}$ to imply $v_{\\theta}(\\cdot, t)$ . It becomes possible to obtain samples from $p_{1}$ by integrating Equation (1) over time, i.e. by starting at $\\mathbf{x}_0 \\sim p_0$ for $t = 0$ and solving the ODE till $t = 1$ . As our approach is training-free, we do not elaborate on the details of learning to regress the vector field $v_{t}$ ; we encourage interested readers to refer to the works of Lipman et al. [20] and Tong et al. [30] for a primer on training FMs.", + "bbox": [ + 511, + 733, + 906, + 869 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "At any timestep $t$ during sampling, an intermediate sample $\\mathbf{x}_t$ in the flow trajectory can be used to obtain an approx-", + "bbox": [ + 511, + 869, + 906, + 901 + ], + "page_idx": 1 + }, + { + "type": "page_number", + "text": "23304", + "bbox": [ + 478, + 944, + 519, + 955 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "imation of the target as follows:", + "bbox": [ + 89, + 90, + 303, + 104 + ], + "page_idx": 2 + }, + { + "type": "equation", + "text": "\n$$\n\\hat {\\mathbf {x}} _ {1} = \\mathbf {x} _ {t} + v _ {\\theta} (\\mathbf {x} _ {t}, t) (1 - t) \\tag {2}\n$$\n", + "text_format": "latex", + "bbox": [ + 192, + 119, + 483, + 135 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Equation (2) is equivalent to simply taking a large Euler step at any time instance $t$ and is naturally more accurate as $t$ approaches $t = 1$ . Further, Equation (2) is also well suited for ODEs with 'straight' paths, where the direction of the time-varying velocity $v_{t}$ remains near-constant in time (such as the work of Liu et al. [22]). Similarly, we can estimate the source sample by simply taking a step in the reverse direction:", + "bbox": [ + 89, + 148, + 483, + 267 + ], + "page_idx": 2 + }, + { + "type": "equation", + "text": "\n$$\n\\hat {\\mathbf {x}} _ {0} = \\mathbf {x} _ {t} - v _ {\\theta} (\\mathbf {x} _ {t}, t) t \\tag {3}\n$$\n", + "text_format": "latex", + "bbox": [ + 214, + 270, + 483, + 286 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "2.2. Determinantal Point Processes", + "text_level": 1, + "bbox": [ + 89, + 295, + 362, + 310 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Determinantal point processes (DPPs) [4, 17, 23] are probabilistic models of repulsion between points. They were originally termed as fermion processes as they describe the Pauli exclusion principle or antibunching effect in fermions. To define a DPP, we must first consider a set of points, $\\mathcal{V}$ , and a point process $\\mathcal{P}(\\mathcal{V})$ — a probability measure on $2^{\\mathcal{V}}$ (the set of all possible subsets of $\\mathcal{V}$ ). $\\mathcal{P}$ is determinantal when the probability of choosing a random subset $Y \\subset \\mathcal{V}$ according to $\\mathcal{P}$ is given by:", + "bbox": [ + 89, + 316, + 483, + 455 + ], + "page_idx": 2 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {P} (Y \\subset \\mathcal {Y}) = \\frac {\\det (\\mathbf {L} _ {Y})}{\\sum_ {Y \\subset \\mathcal {Y}} \\det (\\mathbf {L} _ {Y})} = \\frac {\\det (\\mathbf {L} _ {Y})}{\\det (\\mathbf {L} + \\mathbf {I})} \\tag {4}\n$$\n", + "text_format": "latex", + "bbox": [ + 129, + 465, + 483, + 501 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "where $\\mathbf{L} \\in \\mathbb{R}^{|\\mathcal{Y}| \\times |\\mathcal{Y}|}$ is a kernel matrix, and $\\mathbf{L}_Y$ is the sub-kernel matrix indexed by the elements of $Y$ . Equation (4) has an intuitive geometric interpretation if we consider the kernel $\\mathbf{L}$ to be constructed from cosine similarity: the determinant of $\\mathbf{L}_Y$ is the Gram-determinant, describing the squared volume of the $N$ -dimensional parallelotope spanned by the set of vectors $Y$ . Thus, a DPP naturally assigns higher probabilities to more orthogonal (and thus diverse) subsets that span larger volumes. We leverage DPPs to define a coupled likelihood measure over a set of samples in a flow trajectory.", + "bbox": [ + 89, + 511, + 485, + 679 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3. Related Work", + "text_level": 1, + "bbox": [ + 89, + 693, + 233, + 709 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Efficiently finding diverse sets is useful in several application areas of machine learning. For instance, Batra et al. [3] show that the M-Best MAP (maximum a posteriori) solutions in Markov random fields are often distant from the ground truth and highly similar. They thus propose the *Diverse M-Best* problem—finding a set of $M$ highly probable solutions satisfying some minimum dissimilarity threshold—that partly inspires our study in Section 4. [33] utilize DPPs in conjunction with variational autoencoders (VAE) for diverse trajectory forecasting; a set of diverse future pedestrian trajectories improves safety-critical perception systems in autonomous vehicles. Motivated by potential drug discovery", + "bbox": [ + 89, + 719, + 485, + 901 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "and material design applications, Jain et al. [14] propose finding diverse Pareto-optimal candidates in a multi-objective setting with generative flow networks.", + "bbox": [ + 511, + 90, + 906, + 135 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "The work by Corso et al. [6] which explores diverse non-IID sampling for diffusion models is most similar in spirit to DiverseFlow. However, DiverseFlow is notably different in the following aspects: (1) Our diversity objective is derived from determinantal point processes, a diversity-promoting probability measure of the joint occurrence of a set of samples. Corso et al. [6] is instead inspired by stein variational gradient descent (SVGD) [21]. 2) The diversity measure in DiverseFlow (volume, or determinant of similarity kernel) assigns a zero likelihood to a set if any duplicate elements are present; presence of duplicates is tolerated in the diversity measure in Particle Guidance (row-wise sum of similarity kernel)", + "bbox": [ + 511, + 136, + 908, + 332 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "4. Diverse Source Samples Do Not Yield Diverse Target Samples", + "text_level": 1, + "bbox": [ + 511, + 344, + 906, + 380 + ], + "page_idx": 2 + }, + { + "type": "image", + "img_path": "images/46a6f278801b399e96b2a0a6be298519d95dc071360f264fa4b47cfe9688b918.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 550, + 398, + 709, + 518 + ], + "page_idx": 2 + }, + { + "type": "image", + "img_path": "images/df991f020c6c783fd783210b68a9a57b3e516fefd9044bd263eb9d81e3870a24.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 709, + 398, + 867, + 520 + ], + "page_idx": 2 + }, + { + "type": "image", + "img_path": "images/ac6db7105b857c92bd53ac5877c85e725cae59c7094d3570267fb3798f2425cc.jpg", + "image_caption": [ + "(a) Conditional ing (CFM) [20]", + "(c) DiverseFlow on CFM", + "Figure 2. Finding $K = 5$ samples from the target distribution with $N = 10$ modes. In the example, only 3 modes are found by 5 IID samples (a, b). DiverseFlow discovers 5 modes with the same sampling budget in (d, e)." + ], + "image_footnote": [], + "bbox": [ + 552, + 546, + 709, + 667 + ], + "page_idx": 2 + }, + { + "type": "image", + "img_path": "images/986aeb4001abbd1298dc23aba33fe1e40a26b7eb3eae94cf5c33f43051bd5d50.jpg", + "image_caption": [ + "(b) Mini-batch Optimal Transport (MB-OT) [30]", + "(d) DiverseFlow on MB-OT" + ], + "image_footnote": [], + "bbox": [ + 709, + 546, + 867, + 667 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Problem Setting: We start with a synthetic example to illustrate our problem of interest. Consider that we have empirical observations from a target distribution $\\pi_1 \\in \\mathbb{R}^2$ , which is a random mixture of Gaussians, such as the example shown in Figure 4. We design $\\pi_1 = \\sum_{i=1}^{N} w_i \\mathcal{N}(\\pmb{\\mu}_i, \\sigma_i^2 \\pmb{I})$ to contain $N = 10$ randomly selected modes $\\mathcal{N}(\\pmb{\\mu}_i, \\sigma_i^2 \\pmb{I})$ , each with a random mixture weight $w_i$ ; we observe that in our example, there are 6 high probability modes and 4 low probability ones. Suppose we have a sampling budget", + "bbox": [ + 511, + 763, + 908, + 902 + ], + "page_idx": 2 + }, + { + "type": "page_number", + "text": "23305", + "bbox": [ + 478, + 944, + 517, + 955 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "of $K$ samples. This leads to three possible scenarios: (i) $K < N$ , (ii) $K = N$ , and (iii) $K > N$ . Among the aforementioned, case (i) (fewer samples than modes) is the most likely characteristic of any real-world dataset.", + "bbox": [ + 89, + 90, + 483, + 151 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Let us have a prior distribution $\\pi_0$ and some generative model $\\Psi$ , such that, in the limit of infinite samples, $\\Psi(x_0 \\sim \\pi_0) \\sim \\pi_1$ . Then, the objective of sample-efficient diverse sampling is to obtain samples from $\\min(K, N)$ modes from $\\pi_1$ , given a fixed set of $K$ samples in $\\pi_0$ .", + "bbox": [ + 89, + 154, + 483, + 231 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "If diverse samples are desired from the target density of the flow, one may make the elementary assumption that if the particles are distant at the source distribution, after being transported by the flow, they remain distant in the target distribution. This assumption is not necessarily true, as we show in Figure 2. By design, we choose a uniform mixture of eight Gaussians as the source $\\pi_0$ to obtain diverse source samples. In Figure 2a, we can observe that source points from distinct modes can still converge to the same target mode with IID sampling. Thus, an alternative procedure is necessary to obtain a diverse set from a flow in a sample-efficient manner. We further explore this toy problem in Section 6.4.", + "bbox": [ + 89, + 232, + 486, + 429 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "5. Diverse Sampling from Flows", + "text_level": 1, + "bbox": [ + 89, + 452, + 362, + 469 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "From Figure 2, we observe that independently (or heuristically) chosen source samples may not map to a diverse set of target samples. In this case, we can select a new set of source samples and repeat the sampling process till eventually covering at least $K$ modes. However, this approach does not satisfy our fixed sampling budget constraint. An alternative solution to repeated independent sampling is defining and leveraging a diversity measure of the target samples to drive sample diversity. For the set of source samples $\\{\\mathbf{x}_0^{(1)},\\mathbf{x}_0^{(2)},\\ldots ,\\mathbf{x}_0^{(k)}\\}$ , we could optimize a set of perturbations $\\{\\delta^{(1)},\\delta^{(2)},\\dots,\\delta^{(k)}\\}$ such that the new set $\\{\\mathbf{x}_0^{(1)} + \\delta^{(1)},\\mathbf{x}_0^{(2)} + \\delta^{(2)},\\dots,\\mathbf{x}_0^{(k)} + \\delta^{(k)}\\}$ maps to a diverse set of target particles. However, this approach would require multiple simulations of the whole ODE and backpropagating over all the timesteps, which increases the computational complexity of the sampling process over the standard IID sampling.", + "bbox": [ + 89, + 479, + 483, + 743 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "This leads us to our proposed approach: we avoid multiple simulations and instead optimize the flow trajectory for diversity while solving the ODE. For any sample in the flow trajectory $\\mathbf{x}_t$ , we have an estimate of the target sample $\\hat{\\mathbf{x}}_1$ through Equation (2). Suppose we have a differentiable objective $\\mathcal{L}(\\{\\hat{\\mathbf{x}}_1^{(1)},\\hat{\\mathbf{x}}_1^{(2)},\\dots ,\\hat{\\mathbf{x}}_1^{(k)}\\})$ that assigns a likelihood to the joint outcome $\\{\\hat{\\mathbf{x}}_1^{(1)},\\hat{\\mathbf{x}}_1^{(2)},\\dots ,\\hat{\\mathbf{x}}_1^{(k)}\\}$ . Further, let $\\mathcal{L}$ assign a higher likelihood if the joint outcome is a diverse set, and a diminished likelihood if the set is similar. We can then leverage $\\mathcal{L}$ to drive diversity among the target samples", + "bbox": [ + 89, + 744, + 483, + 902 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "by modifying the flow velocity of the $i$ -th particle as,", + "bbox": [ + 511, + 90, + 864, + 107 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\tilde {v} _ {t} ^ {(i)} = v _ {t} ^ {(i)} - \\gamma (t) \\nabla_ {\\mathbf {x} _ {t} ^ {(i)}} \\log \\mathcal {L} (\\{\\hat {\\mathbf {x}} _ {1} ^ {(1)}, \\hat {\\mathbf {x}} _ {1} ^ {(2)}, \\dots , \\hat {\\mathbf {x}} _ {1} ^ {(k)} \\}) \\tag {5}\n$$\n", + "text_format": "latex", + "bbox": [ + 522, + 114, + 906, + 140 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "where $\\gamma (t)$ is a time-varying scale that controls the strength of the diversity gradient. Setting $\\gamma (t) = 0$ reduces to the standard IID sampling scenario, while $\\gamma (t) > 0$ will encourage diversity between the generated samples. In practice, $\\gamma (t)$ follows the schedule of the probability path normalized by the norm of the DPP gradient.", + "bbox": [ + 511, + 150, + 906, + 242 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "5.1. Determinantal Gradient Constraints", + "text_level": 1, + "bbox": [ + 511, + 250, + 830, + 265 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "We desire objective $\\mathcal{L}$ in Equation (5) to be higher if the items in the set are diverse and lower if they are similar to each other. We interpret diversity in terms of the volume spanned by the set. Consider that we have $k$ samples in $\\mathbb{R}^d$ (assume $k < d$ ). An objective that prefers diversity can be defined as the volume of the $k$ -dimensional parallelotope in $\\mathbb{R}^d$ spanned by the sample vectors; this volume becomes diminished when there are similar samples (and even zero, for identical samples). The determinant describes volumes well; a diverse set must span a large volume in the sample space and have a corresponding large determinant.", + "bbox": [ + 511, + 271, + 906, + 438 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "To define a measure over a set of samples, we draw on the idea of determinantal point processes (DPP). We first define a kernel $\\mathbf{L}(\\{\\hat{\\mathbf{x}}_1^{(1)},\\hat{\\mathbf{x}}_1^{(2)},\\dots ,\\hat{\\mathbf{x}}_1^{(k)}\\})$ as follows:", + "bbox": [ + 511, + 439, + 905, + 487 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\mathbf {L} ^ {(i j)} = \\exp \\left(- h \\frac {\\| \\hat {\\mathbf {x}} _ {1} ^ {(i)} - \\hat {\\mathbf {x}} _ {1} ^ {(j)} \\| _ {2} ^ {2}}{\\operatorname* {m e d} (\\mathbf {U} (\\mathbf {D}))}\\right) \\tag {6}\n$$\n", + "text_format": "latex", + "bbox": [ + 589, + 496, + 906, + 537 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "where $\\mathbf{D}$ denotes a distance matrix with $\\mathbf{D}_{ij} = \\| \\mathbf{x}^{(i)} - \\mathbf{x}^{(j)}\\| _2^2$ , $\\mathbf{U}(\\mathbf{D})$ denotes the upper triangle entries of $\\mathbf{D}$ , $h$ denotes a kernel spread parameter, and $\\mathrm{med}(\\mathbf{U}(\\mathbf{D}))$ denotes the median of those entries. Given $\\mathbf{L}$ , we may define a DPP-based likelihood as:", + "bbox": [ + 511, + 547, + 905, + 625 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\begin{array}{l} \\mathcal {L} (\\{\\hat {\\mathbf {x}} _ {1} ^ {(1)}, \\hat {\\mathbf {x}} _ {1} ^ {(2)}, \\dots , \\hat {\\mathbf {x}} _ {1} ^ {(k)} \\}) = \\frac {\\det (\\mathbf {L})}{\\det (\\mathbf {L} + \\mathbf {I})} \\\\ = \\prod_ {a = 1} ^ {k} \\frac {\\lambda (\\mathbf {L}) _ {a}}{1 + \\lambda (\\mathbf {L}) _ {a}} \\tag {7} \\\\ \\end{array}\n$$\n", + "text_format": "latex", + "bbox": [ + 562, + 647, + 906, + 724 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "where $\\lambda (\\mathbf{L})_a$ is the $a^{\\mathrm{th}}$ eigenvalue of the kernel $\\mathbf{L}$ . The log-likelihood is then,", + "bbox": [ + 511, + 734, + 905, + 767 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {L} \\mathcal {L} = \\log \\mathcal {L} = \\log \\det (\\mathbf {L}) - \\log \\det (\\mathbf {L} + \\mathbf {I}) \\tag {8}\n$$\n", + "text_format": "latex", + "bbox": [ + 562, + 777, + 906, + 794 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Note that the Euclidean distance $\\| \\hat{\\mathbf{x}}_1^{(i)} - \\hat{\\mathbf{x}}_1^{(j)}\\| _2^2$ is not very meaningful in the high-dimensional raw image space [1]. Therefore, in practice, the distance should be computed in a robust feature space, i.e., $\\| F(\\hat{\\mathbf{x}}_1^{(i)}) - F(\\hat{\\mathbf{x}}_1^{(j)})\\| _2^2$ , where $F$ is some domain-specific feature extractor, such as the vision transformer (ViT) [9] for images.", + "bbox": [ + 511, + 805, + 906, + 901 + ], + "page_idx": 3 + }, + { + "type": "page_number", + "text": "23306", + "bbox": [ + 478, + 944, + 519, + 957 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Quality Constraint: The DPP defined in Equation (7) acts as a repulsive force that is unaware of sample quality. A quality term can be incorporated into the DPP kernel to regularize the trajectory diversification. Although flows can be defined between any arbitrary two distributions, let us consider the special case when the source is a Gaussian, i.e., $p_0 \\sim \\mathcal{N}(0,\\mathbf{I})$ . Suppose we have a quality vector $\\mathbf{q}_t = \\{q^{(1)}(t), q^{(2)}(t), \\ldots, q^{(k)}(t)\\}$ , where any $q^{(i)}(t) \\in [0,1]$ . We can then define a new kernel $\\mathbf{L}_q = \\mathbf{L} \\odot \\mathbf{q}_t \\mathbf{q}_t^T$ , where each $q^{(i)}(t)$ penalizes a sample $\\mathbf{x}_t^{(i)}$ if it deviates too much from the flow. To define this, we obtain an estimate of the source sample $\\hat{\\mathbf{x}}_0^{(i)}(t)$ for any given sample $\\mathbf{x}_t^{(i)}$ via Equation (3), and check if it lies within a desired percentile-radius $\\rho$ of the Gaussian $p_0$ . Specifically, we define the time-dependent sample quality as", + "bbox": [ + 89, + 90, + 483, + 327 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\nq ^ {(i)} (t) = \\left\\{ \\begin{array}{l l} 1 & \\text {i f} \\| \\hat {\\mathbf {x}} _ {0} ^ {(i)} (t) \\| _ {2} ^ {2} \\leq \\rho^ {2} \\\\ \\max \\left(\\epsilon , e ^ {- \\left(\\| \\hat {\\mathbf {x}} _ {0} ^ {(i)} (t) \\| _ {2} ^ {2} - \\rho^ {2}\\right)}\\right) & \\text {o t h e r w i s e} \\end{array} \\right. \\tag {9}\n$$\n", + "text_format": "latex", + "bbox": [ + 89, + 334, + 493, + 397 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "where $\\epsilon$ is a 'minimum quality' we assign to prevent a zero determinant.", + "bbox": [ + 89, + 398, + 483, + 428 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "5.2. Coupled Ordinary Differential Equations", + "text_level": 1, + "bbox": [ + 89, + 436, + 444, + 452 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "At any timestep $t$ , the measure of diversity in Equation (8) can be adopted to modify the flow of the $i$ -th particle. We compute the gradient of the samples with respect to the diversity measure and use it to modify the ODE as follows:", + "bbox": [ + 89, + 459, + 483, + 520 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\left. d \\mathbf {x} _ {t} ^ {(i)} = \\left[ v _ {\\theta} \\left(\\mathbf {x} _ {t} ^ {(i)}, t\\right) - \\gamma (t) \\nabla_ {\\mathbf {x} _ {t} ^ {(i)}} \\log \\mathcal {L} \\left(\\left\\{\\hat {\\mathbf {x}} _ {1} ^ {(1)}, \\dots , \\hat {\\mathbf {x}} _ {1} ^ {(k)} \\right\\}\\right) \\right] d t \\right. \\tag {10}\n$$\n", + "text_format": "latex", + "bbox": [ + 89, + 527, + 491, + 564 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Where $\\gamma (t)$ is a time-varying scaling factor. Unlike the IID sampling scenario where we have $K$ independent ODEs, Equation (10) corresponds to a system of coupled non-linear ordinary differential ordinary equations. To see this, first note that the estimate $\\hat{\\mathbf{x}}_1^{(i)}$ depends on the current sample $\\mathbf{x}_t^{(i)}$ i.e., $\\hat{\\mathbf{x}}_1^{(i)} = \\mathbf{x}_t^{(i)} + v_\\theta (\\mathbf{x}_t^{(i)},t)(1 - t)$ . Second, the DPP log-likelihood $\\mathcal{LL}(\\{\\hat{\\mathbf{x}}_1^{(1)},\\hat{\\mathbf{x}}_1^{(2)},\\dots ,\\hat{\\mathbf{x}}_1^{(k)}\\})$ induces a time-dependent coupling between the $K$ trajectories of $\\mathbf{x}_t^{(i)},i = 1,\\ldots ,K$ and seeks to diversify the target samples. Although higher-order ODE solvers [16] can be employed to solve the coupled ODEs, we use the standard Euler method.", + "bbox": [ + 89, + 566, + 483, + 742 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "6. Experiments", + "text_level": 1, + "bbox": [ + 89, + 753, + 223, + 771 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "We demonstrate the utility of DiverseFlow in flow-based generative models by considering three applications where sample diversity is naturally desirable: text-guided image generation with polysemous words, large-hole inpainting, and class-conditional image generation. We also analyze the effect of DiverseFlow on different flow matching formulations w.r.t. its ability to span diverse modes through a synthetically constructed 2D density example.", + "bbox": [ + 89, + 779, + 483, + 902 + ], + "page_idx": 4 + }, + { + "type": "image", + "img_path": "images/e4345bb398268dcbae533bc9541668f565a02ab355ea7220071bf32e4d02d1b0.jpg", + "image_caption": [ + "(a) \"A buck\"" + ], + "image_footnote": [], + "bbox": [ + 542, + 89, + 710, + 219 + ], + "page_idx": 4 + }, + { + "type": "image", + "img_path": "images/9e9afdf4fc87a74b527a28a5bc998905bb61a5ff0e5622e3a90519c303a4110f.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 710, + 90, + 877, + 219 + ], + "page_idx": 4 + }, + { + "type": "image", + "img_path": "images/bda090aeb79979dc0ae0d48a0cf8643cdf03d5452d7a8e956d066e809178dc1e.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 540, + 239, + 710, + 369 + ], + "page_idx": 4 + }, + { + "type": "image", + "img_path": "images/a2442f078401e273bd017cda6dd894ae35299eb39eafccf7b61b850cc6c6ed16.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 710, + 239, + 877, + 369 + ], + "page_idx": 4 + }, + { + "type": "image", + "img_path": "images/698e897027df0fac5fb324f1154f46536ad5a2654aeaffb8b2d5a96c3c806c28.jpg", + "image_caption": [ + "(b) \"A famous boxer\"", + "Figure 3. For each prompt, the left image (red box) denotes standard IID sampling with classifier-free guidance, while the right image (blue box) shows the result after incorporating DiverseFlow. DiverseFlow finds more diverse sets given the same source points—clearly distinguishable as new semantic meanings in the case of prompts with multiple meanings." + ], + "image_footnote": [], + "bbox": [ + 540, + 388, + 710, + 520 + ], + "page_idx": 4 + }, + { + "type": "image", + "img_path": "images/ec6497676581e9b0f00046f6ac8be6c6d8b3bf00a6134c0b7fd1352fa75a9896.jpg", + "image_caption": [ + "(c) \"Van Gogh painting\"" + ], + "image_footnote": [], + "bbox": [ + 710, + 390, + 877, + 520 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "6.1. Image Generation from Polysemous Prompts", + "text_level": 1, + "bbox": [ + 511, + 662, + 895, + 680 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "In text-to-image generation, the conditional data distribution corresponding to a text prompt may contain many variations, and it is a desirable objective to generate images that span those variations in a sample-efficient manner. As we have previously shown in Figure 1, the result desired by a user (a human boxer athlete) may not be the result generated by the model (a dog breed); if diverse results are obtained, it allows a user to obtain the desired results with fewer attempts.", + "bbox": [ + 511, + 686, + 905, + 808 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "We pose a scenario where diverse sets are easily observable: when an open-ended text prompt is polysemous and carries multiple meanings, such as the examples we show in Figure 1 and Figure 3. In Figure 3(a), the prompt \"A buck\" may commonly refer to a male deer. However, it may also informally refer to a United States dollar. Using the", + "bbox": [ + 511, + 810, + 906, + 900 + ], + "page_idx": 4 + }, + { + "type": "page_number", + "text": "23307", + "bbox": [ + 478, + 944, + 517, + 955 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "same four source points, which are deterministically mapped to four deer images by IID sampling. DiverseFlow finds a different set of samples—one that includes a dollar-like coin, albeit embossed with a deer head. We also observe minor differences between the two sets of images, such as changes in pose and background in the top-right and bottom-right deers. For Figure 3(c), while 'Van Gogh painting' is not quite a polysemous word, it can still have two interpretations: a painting painted by Van Gogh, or a painting of Van Gogh. The regular samples contain minimal diversity, as they include two sets of repeated paintings of Van Gogh. With DiverseFlow, not only can we get a set of four distinct paintings, but we also have a portrait of Van Gogh, which is one of the additional meanings of the prompt. We present additional qualitative samples in the supplementary material.", + "bbox": [ + 89, + 90, + 483, + 318 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "6.2. Diverse Inpainting on Faces", + "text_level": 1, + "bbox": [ + 89, + 348, + 341, + 364 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Another inverse problem where diverse solutions are desirable is face inpainting, where we seek to inpaint the missing parts of the face with diverse plausible facial textures and structures. To demonstrate inpainting with FM models, we first incorporate Manifold Constrained Gradient (MCG) [5] in an off-the-shelf unconditional Rectified-Flow model. In addition to the manifold constraints, we employ determinantal gradient constraints to enhance diversity (further detailed in Algorithm 1). In Figure 5 (b), we observe that the inpainted faces of the four women have similar expressions (largely neutral). DiverseFlow improves the diversity of the set by yielding a highly different expression in the top-right image. In (d) and (e), we also observe changes in facial hair and expressions due to diversification.", + "bbox": [ + 89, + 377, + 483, + 590 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/cc6df4e4178aa44caacbc8ad8bc3b772697f71015d64a3422294bba61425c9db.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 125, + 628, + 189, + 729 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/55bc3c8aee37e194ee6eda46aaab6738322b8aec946c8e8995d1f8780cbf192d.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 191, + 628, + 318, + 729 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/d175f3a6cae4e06745f9435defd5c1fd29139821dd91cbc6930017f91f00e08c.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 320, + 628, + 450, + 729 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/9095c00848783c151ea1264db17b17e0909c02afb63eab4617b7b78faf43c5ca.jpg", + "image_caption": [ + "(a)", + "Figure 5. Inpainting on CelebAHQ-256 × 256. (a) Dashed boxes show masked input (top) and ground truth (bottom) respectively. (b) Rectified-Flow [22] + MCG [5] (c) With DiverseFlow." + ], + "image_footnote": [], + "bbox": [ + 125, + 732, + 189, + 834 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/38338b8f2f901e450b3e33471eaff73b3d470a1848893290f142d5a556cac363.jpg", + "image_caption": [ + "(b)" + ], + "image_footnote": [], + "bbox": [ + 191, + 733, + 318, + 833 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/cab056f61389e95bd31f738823295add41b9a4c514fd136d339522eb02c8c685.jpg", + "image_caption": [ + "(c)" + ], + "image_footnote": [], + "bbox": [ + 320, + 733, + 450, + 834 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/5c392b70064bc2f773bfde4fa5af59b45346919daa72c830dcba39a1e33b12f6.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 519, + 103, + 645, + 200 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/61ece2b7ba393195a2b1c3504086b09eed3f8aba7ba95ff3623b6618a75f53df.jpg", + "image_caption": [ + "No CFG" + ], + "image_footnote": [], + "bbox": [ + 519, + 202, + 645, + 297 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/f9e2b90427a740b7b6e70ed79ea9601fda8929a5de19ce6aecf156a0e4023231.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 647, + 103, + 774, + 200 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/6270b52bbe42c86627c415e9e972ad66c929c2859609a3f7ae8c27e28352abf3.jpg", + "image_caption": [ + "With CFG", + "Figure 6. Class-conditional ImageNet samples from LFM [7]. We show samples for two classes, (top row) 'Mushroom' (class 947) and (bottom row) 'Macaw' (class 88). (a, d) No CFG. (b, e) LFM with CFG. (d, f) LFM with CFG and DiverseFlow." + ], + "image_footnote": [], + "bbox": [ + 647, + 202, + 772, + 297 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/7567bfb3dcd0871aac8f8dcc319182c20bf53c26850f9b6a7c0fd94e9ce86865.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 774, + 103, + 901, + 200 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/b5e3b8c6e5a44aaf48e853449ff28850c8e09792fbfb8c6dfe871e184b337fdd.jpg", + "image_caption": [ + "DiverseFlow" + ], + "image_footnote": [], + "bbox": [ + 774, + 202, + 901, + 297 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "6.3. Diverse Class-Conditional Image Synthesis", + "text_level": 1, + "bbox": [ + 511, + 395, + 877, + 412 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Suppose we can access a class-conditioned flow matching (FM) model trained on an unknown image dataset. To explore the unobservable true dataset, we may use a set of class-conditional samples from the FM model. We adopt a latent flow matching (LFM) model [7], trained to generate $256 \\times 256$ resolution images from the ImageNet [8] dataset. Much like latent diffusion, LFM employs classifier-free guidance to create high-quality samples. However, this naturally poses a cost to diversity, as we show in Figure 6.", + "bbox": [ + 511, + 417, + 906, + 551 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "By incorporating DiverseFlow, we can maintain the high quality of the samples and simultaneously explore more modes in the dataset. In Figure 6, we demonstrate two ImageNet classes that may have diversity: 'Mushroom' and 'Macaw.' For mushrooms, we observe that LFM primarily generates two species of mushrooms. However, by applying DiverseFlow, we successfully find a new species within our limited set: an Amanita muscaria, also known as the fly agaric—easily distinguishable by the white spots on its red cap. In another example, we see that while LFM generates four scarlet macaws, using the same source samples, DiverseFlow helps us find a different blue and yellow macaw. We further evaluate the precision and recall [18] for ImageNet-256 synthesis in Figure 7, showing that DiverseFlow improves the recall while retaining similar precision, suggesting improved diversity.", + "bbox": [ + 511, + 554, + 908, + 796 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "6.4. DiverseFlow Across Various FM Formulations", + "text_level": 1, + "bbox": [ + 511, + 801, + 906, + 818 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "We ask the question: does DiverseFlow remain consistent across different formulations of flow matching? We adopt the same toy bivariate mixture of Gaussian distributions as in the example shown in Figure 2, across four different FM formulations: (i) Conditional Flow Matching (CFM) [20],", + "bbox": [ + 511, + 825, + 908, + 902 + ], + "page_idx": 5 + }, + { + "type": "page_number", + "text": "23308", + "bbox": [ + 478, + 944, + 517, + 955 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/fef732875e8d72d562ee357e5aad903d9199914893ad33400c074e21c5ea3492.jpg", + "image_caption": [ + "Figure 4. Comparing different FM formulations in terms of the number of modes spanned by IID sampling versus with DiverseFlow. More details about the experiment are provided in the supplementary." + ], + "image_footnote": [], + "bbox": [ + 91, + 85, + 903, + 257 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/0bf640e7bba7bb1cd27b421928f5d85dbb886e09c20e0128f8be1202571b0516.jpg", + "image_caption": [ + "Figure 7. Precision vs. recall on ImageNet-256." + ], + "image_footnote": [], + "bbox": [ + 91, + 319, + 274, + 439 + ], + "page_idx": 6 + }, + { + "type": "table", + "img_path": "images/d6483d0ec3b74856c40680d15c5693639bc11fc0763ae97d71426ac85fe8ef63.jpg", + "table_caption": [], + "table_footnote": [], + "table_body": "
MethodCFGP (↑)R (↑)
LFM10.620.52
LFM1.50.690.44
LFM20.770.41
LFM40.690.26
LFM + DF10.620.52
LFM+DF1.50.690.47
LFM+DF20.760.46
LFM+DF40.700.38
", + "bbox": [ + 292, + 319, + 478, + 439 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Table 1. Precision (P) and recall (R) score of DiverseFlow and baselines.", + "bbox": [ + 287, + 450, + 482, + 474 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "(ii) Mini-batch Optimal Transport CFM (MB-OT) [30], (iii) Schrödinger Bridge CFM (SB-CFM) [31], and (iv) Stochastic Interpolants (SI-CFM) [2]. We then perform a numerical experiment to quantify the average number of modes discovered by each FM variant with increasing the sampling budget $K$ . Figure 4 reports the results. For a maximum sampling budget of 10, MB-OT discovers only 5.64 modes on average, which is expected since the dataset contains 6 high-probability modes. By incorporating DiverseFlow, we can find 7.11 modes on average. We observe that MB-OT and SB-CFM benefit most from DiverseFlow. We hypothesize this is because MB-OT and SB-CFM are formulated with the notion of optimal paths, which results in a more accurate estimation of $\\hat{x}_1$ and thus a better diversity gradient.", + "bbox": [ + 88, + 503, + 485, + 717 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "6.5. Comparison To Particle Guidance", + "text_level": 1, + "bbox": [ + 89, + 727, + 390, + 742 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "(i) Preventing Training Data Copy: Previously, Corso et al. [6] demonstrate that their method can alleviate Stable Diffusion's training data regurgitation problem [27] to some extent. In Figure 8, we demonstrate similar capabilities; in (b), the top left and bottom right examples are copies of the training data. Subsequently, (c) and (d) find a new example for the top-left sample. (ii) Diversity vs Quality: We also compare the diversity versus quality of DiverseFlow against Particle Guidance in Figure 10 over 30 polysemous prompts repeated over 10 random seeds. While varying the", + "bbox": [ + 88, + 750, + 485, + 902 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/a5ba5bf8f7fb9dd0635fa4dbf85f6f362ac9ea799e39bdb9d860577664fe5dd6.jpg", + "image_caption": [ + "(a) Training Data" + ], + "image_footnote": [], + "bbox": [ + 517, + 319, + 709, + 468 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/ce25bfa9b4e5ef7fa09c122f1c27b16773f3114a201b8e04e8527e4f2359ce14.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 709, + 319, + 903, + 469 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/d7863ca0514b5e16bb6e714f257336c5bbe02efcacc86a79b500fa53919ddfe6.jpg", + "image_caption": [ + "(c) Particle Guidance [6]", + "Figure 8. For the prompt \"VAN GOGH CAFE TERASSE copy.jpg\", we show (a) the training data, (b) IID sampling with two copies in top-left and bottom-right, (c) Particle Guidance [6], and (d) DiverseFlow." + ], + "image_footnote": [], + "bbox": [ + 517, + 483, + 709, + 632 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/5ef79318d3c8a81eb0c24b6848c9a23249e3ad3a237b38e3e86406a4c31fcf20.jpg", + "image_caption": [ + "(b) IID Sampling", + "(d) DiverseFlow" + ], + "image_footnote": [], + "bbox": [ + 709, + 483, + 903, + 632 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "strength of DiverseFlow to improve diversity, we also vary the values of the classifier-free guidance strength from 7.5 to 10 to boost quality. Quality is measured by Aesthetic Score (higher is better) [25], and diversity is measured by average pairwise similarity of a set (lower is better) [6]. We observe that though DiverseFlow obtains better diversity at similar quality to IID sampling, the aesthetic score of Particle Guidance is unusually low. In Figure 9 we highlight the cause: for the challenging task of polysemous text-to-image generation, the images generated from Particle Guidance (a) suffers", + "bbox": [ + 511, + 750, + 908, + 902 + ], + "page_idx": 6 + }, + { + "type": "page_number", + "text": "23309", + "bbox": [ + 478, + 944, + 519, + 957 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/65c4181a6ba7953eba996ac49880980cf74dca076240aad473bbdd4382ae2d76.jpg", + "image_caption": [ + "Figure 9. Particle Guidance (middle rows) can suffer from artefacts for polysemous prompts. DiverseFlow retains quality while achieving diversity: for \"A mint\", a minted coin with a mint leaf was discovered." + ], + "image_footnote": [], + "bbox": [ + 91, + 88, + 480, + 536 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/185e2a6bc82bf7990294139d87b3f16d96b9684fcd145794d903258555a2a688.jpg", + "image_caption": [ + "Figure 10. In-batch similarity $(\\downarrow)$ vs aesthetic score $(\\uparrow)$" + ], + "image_footnote": [], + "bbox": [ + 114, + 599, + 459, + 797 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "from image artefacts, which achieves a high diversity score but a poor quality score. Additional details about the experiment, including the set of 30 prompts, are provided in the", + "bbox": [ + 89, + 854, + 483, + 902 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/57d72ea72048db47ef67e9e96c7c2a74edc1aa9a826a4ec5e60fe0629c570eb7.jpg", + "image_caption": [ + "Figure 11. Average modes discovered over 100 trials, for varying noise scales and strengths of DiverseFlow and ParticleGuidance." + ], + "image_footnote": [], + "bbox": [ + 516, + 88, + 903, + 397 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "supplementary. (iii) Mode Discovery: In high dimensional data, the number of modes, $N$ , is significantly greater than the budget $K$ , so finding unique modes is still highly probable. To better highlight the differences between DiverseFlow and Particle Guidance, we adopt a simpler toy experiment: finding modes in a symmetric uniform Mixture of Gaussian distribution. Corso et al. [6] provides the result that IID sampling with budget $K = 10$ discovers about 6.5 modes on average, while Particle Guidance with a Euclidean kernel discovers almost 9 out of 10 modes. We verify this result in Figure 11, finding that Particle Guidance discovers up to 8.8 modes (averaged over 100 trials). However, by using DiverseFlow, it is possible to discover all 10 modes on average, showing that our approach has a stronger diversification effect.", + "bbox": [ + 511, + 465, + 908, + 691 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "7. Conclusion", + "text_level": 1, + "bbox": [ + 513, + 707, + 633, + 723 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "In this paper, we present DiverseFlow, a way to enforce diversity among a limited set of samples generated from a flow by coupling them through a determinantal point process. We demonstrate multiple applications where our method can be useful. In essence, though flow models admit a deterministic mapping from source to sample, DiverseFlow allows us to modify this mapping at inference-time and obtain samples with improved diversity. Diversity enhances the utility of the underlying generative flow in the case of ill-posed problems, where many solutions exist and it is desirable to find a multitude of such solutions.", + "bbox": [ + 511, + 734, + 908, + 900 + ], + "page_idx": 7 + }, + { + "type": "page_number", + "text": "23310", + "bbox": [ + 478, + 944, + 519, + 957 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "References", + "text_level": 1, + "bbox": [ + 91, + 89, + 187, + 104 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[1] Charu C Aggarwal, Alexander Hinneburg, and Daniel A Keim. On the surprising behavior of distance metrics in high dimensional space. In Database Theory—ICDT 2001: 8th International Conference London, UK, January 4–6, 2001 Proceedings 8, pages 420–434. Springer, 2001. 4", + "[2] Michael S Albergo, Nicholas M Boffi, and Eric Vanden-Eijnden. Stochastic interpolants: A unifying framework for flows and diffusions. arXiv preprint arXiv:2303.08797, 2023. 2, 7, 4", + "[3] Dhruv Batra, Payman Yadollahpour, Abner Guzman-Rivera, and Gregory Shakhnarovich. Diverse m-best solutions in markov random fields. In Computer Vision-ECCV 2012: 12th European Conference on Computer Vision, Florence, Italy, October 7-13, 2012, Proceedings, Part V 12, pages 1-16. Springer, 2012. 3", + "[4] Alexei Borodin and Grigori Olshanski. Distributions on partitions, point processes, and the hypergeometric kernel. Communications in Mathematical Physics, 211:335-358, 2000. 3", + "[5] Hyungjin Chung, Byeongsu Sim, Dohoon Ryu, and Jong Chul Ye. Improving diffusion models for inverse problems using manifold constraints. Advances in Neural Information Processing Systems, 35:25683-25696, 2022. 6, 1", + "[6] Gabriele Corso, Yilun Xu, Valentin De Bortoli, Regina Barzilay, and Tommi Jaakkola. Particle guidance: non-iid diverse sampling with diffusion models. International Conference on Learning Representations, 2023. 3, 7, 8, 1", + "[7] Quan Dao, Hao Phung, Binh Nguyen, and Anh Tran. Flow matching in latent space. arXiv preprint arXiv:2307.08698, 2023. 6, 3", + "[8] Jia Deng, Wei Dong, Richard Socher, Li-Jia Li, Kai Li, and Li Fei-Fei. Imagenet: A large-scale hierarchical image database. In 2009 IEEE conference on computer vision and pattern recognition, pages 248-255. IEEE, 2009. 6", + "[9] Alexey Dosovitskiy, Lucas Beyer, Alexander Kolesnikov, Dirk Weissenborn, Xiaohua Zhai, Thomas Unterthiner, Mostafa Dehghani, Matthias Minderer, Georg Heigold, Sylvain Gelly, et al. An image is worth 16x16 words: Transformers for image recognition at scale. International Conference on Learning Representations, 2021. 4", + "[10] Patrick Esser, Sumith Kulal, Andreas Blattmann, Rahim Entezari, Jonas Müller, Harry Saini, Yam Levi, Dominik Lorenz, Axel Sauer, Frederic Boesel, et al. Scaling rectified flow transformers for high-resolution image synthesis. In International Conference on Machine Learning, pages 12606-12633. PMLR, 2024. 2, 1", + "[11] Jonathan Ho and Tim Salimans. Classifier-free diffusion guidance. arXiv preprint arXiv:2207.12598, 2022. 1", + "[12] Jonathan Ho, Ajay Jain, and Pieter Abbeel. Denoising diffusion probabilistic models. Advances in neural information processing systems, 33:6840-6851, 2020. 1, 2", + "[13] Noboru Isobe, Masanori Koyama, Jinzhe Zhang, Kohei Hayashi, and Kenji Fukumizu. Extended flow matching: a method of conditional generation with generalized continuity equation. arXiv preprint arXiv:2402.18839, 2024. 5" + ], + "bbox": [ + 93, + 114, + 485, + 900 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[14] Moksh Jain, Sharath Chandra Raparthy, Alex Hernández-García, Jarrid Rector-Brooks, Yoshua Bengio, Santiago Miret, and Emmanuel Bengio. Multi-objective gflownets. In International Conference on Machine Learning, pages 14631-14653. PMLR, 2023. 3", + "[15] Tero Karras, Timo Aila, Samuli Laine, and Jaakko Lehtinen. Progressive growing of gans for improved quality, stability, and variation. International Conference on Learning Representations, 2018. 1", + "[16] Tero Karras, Miika Aittala, Timo Aila, and Samuli Laine. Elucidating the design space of diffusion-based generative models. Advances in Neural Information Processing Systems, 2022. 1, 5", + "[17] Alex Kulesza, Ben Taskar, et al. Determinantal point processes for machine learning. Foundations and Trends® in Machine Learning, 5(2-3):123-286, 2012. 2, 3", + "[18] Tuomas Kynkänniemi, Tero Karras, Samuli Laine, Jaakko Lehtinen, and Timo Aila. Improved precision and recall metric for assessing generative models. Advances in neural information processing systems, 32, 2019. 6", + "[19] Yiheng Li, Heyang Jiang, Akio Kodaira, Masayoshi Tomizuka, Kurt Keutzer, and Chenfeng Xu. Immiscible diffusion: Accelerating diffusion training with noise assignment. Advances in neural information processing systems, 2024. 5", + "[20] Yaron Lipman, Ricky TQ Chen, Heli Ben-Hamu, Maximilian Nickel, and Matt Le. Flow matching for generative modeling. arXiv preprint arXiv:2210.02747, 2022. 1, 2, 3, 6", + "[21] Qiang Liu and Dilin Wang. Stein variational gradient descent: A general purpose bayesian inference algorithm. Advances in neural information processing systems, 29, 2016. 3", + "[22] Xingchao Liu, Chengyue Gong, and Qiang Liu. Flow straight and fast: Learning to generate and transfer data with rectified flow. arXiv preprint arXiv:2209.03003, 2022. 1, 2, 3, 6", + "[23] Odile Macchi. The coincidence approach to stochastic point processes. Advances in Applied Probability, 7(1):83-122, 1975. 2, 3", + "[24] Robin Rombach, Andreas Blattmann, Dominik Lorenz, Patrick Esser, and Björn Ommer. High-resolution image synthesis with latent diffusion models. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 10684-10695, 2022. 1", + "[25] Christoph Schuhmann and LAION. Improved aesthetic predictor: Clip+mlp aesthetic score predictor, 2022. https://github.com/christophschuhmann/improved-aesthetic-predictor. 7", + "[26] Jascha Sohl-Dickstein, Eric Weiss, Niru Maheswaranathan, and Surya Ganguli. Deep unsupervised learning using nonequilibrium thermodynamics. In International Conference on Machine Learning, pages 2256-2265. PMLR, 2015. 1", + "[27] Gowthami Somepalli, Vasu Singla, Micah Goldblum, Jonas Geiping, and Tom Goldstein. Diffusion art or digital forgery? investigating data replication in diffusion models. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 6048-6058, 2023. 7", + "[28] Jiaming Song, Chenlin Meng, and Stefano Ermon. Denoising diffusion implicit models. International Conference on Learning Representations, 2021. 1" + ], + "bbox": [ + 516, + 92, + 906, + 900 + ], + "page_idx": 8 + }, + { + "type": "page_number", + "text": "23311", + "bbox": [ + 478, + 944, + 517, + 955 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[29] Yang Song, Jascha Sohl-Dickstein, Diederik P Kingma, Abhishek Kumar, Stefano Ermon, and Ben Poole. Score-based generative modeling through stochastic differential equations. International Conference on Learning Representations, 2021. 2", + "[30] Alexander Tong, Nikolay Malkin, Guillaume Huguet, Yanlei Zhang, Jarrid Rector-Brooks, Kilian Fatras, Guy Wolf, and Yoshua Bengio. Improving and generalizing flow-based generative models with minibatch optimal transport. Transactions on Machine Learning Research, 2023. 1, 2, 3, 7, 5", + "[31] Alexander Y Tong, Nikolay Malkin, Kilian Fatras, Lazar Atanackovic, Yanlei Zhang, Guillaume Huguet, Guy Wolf, and Yoshua Bengio. Simulation-free schrödinger bridges via score and flow matching. In International Conference on Artificial Intelligence and Statistics, pages 1279-1287. PMLR, 2024. 7", + "[32] Zhisheng Xiao, Karsten Kreis, and Arash Vahdat. Tackling the generative learning trilemma with denoising diffusion gans. International Conference on Learning Representations, 2022. 1", + "[33] Ye Yuan and Kris Kitani. Diverse trajectory forecasting with determinantal point processes. International Conference on Learning Representations, 2020. 3, 5", + "[34] Qinqing Zheng, Matt Le, Neta Shaul, Yaron Lipman, Aditya Grover, and Ricky TQ Chen. Guided flows for generative modeling and decision making. arXiv preprint arXiv:2311.13443, 2023. 1", + "[35] Yinglin Zheng, Hao Yang, Ting Zhang, Jianmin Bao, Dongdong Chen, Yangyu Huang, Lu Yuan, Dong Chen, Ming Zeng, and Fang Wen. General facial representation learning in a visual-linguistic manner. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 18697–18709, 2022. 2" + ], + "bbox": [ + 91, + 92, + 485, + 556 + ], + "page_idx": 9 + }, + { + "type": "page_number", + "text": "23312", + "bbox": [ + 478, + 945, + 517, + 955 + ], + "page_idx": 9 + } +] \ No newline at end of file diff --git a/2025/DiverseFlow_ Sample-Efficient Diverse Mode Coverage in Flows/276f7b10-7b9f-439d-a85a-8ad9f72195cf_model.json b/2025/DiverseFlow_ Sample-Efficient Diverse Mode Coverage in Flows/276f7b10-7b9f-439d-a85a-8ad9f72195cf_model.json new file mode 100644 index 0000000000000000000000000000000000000000..2c435b596fa4193280562ba6b76c049f521e38d0 --- /dev/null +++ b/2025/DiverseFlow_ Sample-Efficient Diverse Mode Coverage in Flows/276f7b10-7b9f-439d-a85a-8ad9f72195cf_model.json @@ -0,0 +1,2321 @@ +[ + [ + { + "type": "header", + "bbox": [ + 0.107, + 0.003, + 0.182, + 0.043 + ], + "angle": 0, + "content": "CVF" + }, + { + "type": "header", + "bbox": [ + 0.237, + 0.001, + 0.812, + 0.047 + ], + "angle": 0, + "content": "This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore." + }, + { + "type": "title", + "bbox": [ + 0.177, + 0.131, + 0.823, + 0.154 + ], + "angle": 0, + "content": "DiverseFlow: Sample-Efficient Diverse Mode Coverage in Flows" + }, + { + "type": "text", + "bbox": [ + 0.322, + 0.182, + 0.674, + 0.198 + ], + "angle": 0, + "content": "Mashrur M. Morshed Vishnu Boddeti" + }, + { + "type": "text", + "bbox": [ + 0.395, + 0.2, + 0.603, + 0.218 + ], + "angle": 0, + "content": "Michigan State University" + }, + { + "type": "text", + "bbox": [ + 0.383, + 0.22, + 0.612, + 0.233 + ], + "angle": 0, + "content": "{morshedm, vishnu}@msu.edu" + }, + { + "type": "title", + "bbox": [ + 0.249, + 0.269, + 0.327, + 0.285 + ], + "angle": 0, + "content": "Abstract" + }, + { + "type": "text", + "bbox": [ + 0.089, + 0.302, + 0.486, + 0.574 + ], + "angle": 0, + "content": "Many real-world applications of flow-based generative models desire a diverse set of samples that cover multiple modes of the target distribution. However, the predominant approach for obtaining diverse sets is not sample-efficient, as it involves independently obtaining many samples from the source distribution and mapping them through the flow until the desired mode coverage is achieved. As an alternative to repeated sampling, we introduce DiverseFlow: a training-free approach to improve the diversity of flow models. Our key idea is to employ a determinantal point process to induce a coupling between the samples that drives diversity under a fixed sampling budget. In essence, DiverseFlow allows exploration of more variations in a learned flow model with fewer samples. We demonstrate the efficacy of our method for tasks where sample-efficient diversity is desirable, such as text-guided image generation with polysemous words, inverse problems like large-hole inpainting, and class-conditional image synthesis." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.605, + 0.222, + 0.62 + ], + "angle": 0, + "content": "1. Introduction" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.632, + 0.484, + 0.842 + ], + "angle": 0, + "content": "Consider the task of text-guided image generation from open-ended prompts, like \"A famous boxer\". If we use a generative ordinary differential equation (ODE) to obtain a few images for this given prompt, we may observe something unusual: all the resultant images depict a dog, as shown in Figure 1 (top-left). Such a result is plausible, as the word \"boxer\" has multiple meanings: it can either mean a combat sport athlete or a particular dog breed. However, for the prompt \"A famous boxer\", what if the output desired by the user was actually a human athlete, and not a dog? This situation necessitates obtaining additional samples from the model, until the desired alternate meanings are discovered. But instead of repeated sampling, can we directly observe more meanings by finding a more diverse set?" + }, + { + "type": "text", + "bbox": [ + 0.091, + 0.844, + 0.484, + 0.875 + ], + "angle": 0, + "content": "Beyond the aforementioned example of text-to-image generation from polysemous1 prompts, sample diversity is a" + }, + { + "type": "image_caption", + "bbox": [ + 0.598, + 0.267, + 0.82, + 0.281 + ], + "angle": 0, + "content": "Prompt: \"A famous boxer\"" + }, + { + "type": "image", + "bbox": [ + 0.536, + 0.284, + 0.707, + 0.416 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.715, + 0.285, + 0.885, + 0.416 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.634, + 0.42, + 0.797, + 0.433 + ], + "angle": 0, + "content": "Prompt: \"A letter\"" + }, + { + "type": "image", + "bbox": [ + 0.536, + 0.435, + 0.707, + 0.565 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.715, + 0.435, + 0.885, + 0.565 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.611, + 0.569, + 0.636, + 0.581 + ], + "angle": 0, + "content": "ID" + }, + { + "type": "image_caption", + "bbox": [ + 0.768, + 0.569, + 0.853, + 0.581 + ], + "angle": 0, + "content": "DiverseFlow" + }, + { + "type": "image_caption", + "bbox": [ + 0.513, + 0.592, + 0.908, + 0.634 + ], + "angle": 0, + "content": "Figure 1. Text-guided image generation with polysemous words. With the same budget of samples, DiverseFlow (right) finds a more diverse set of results compared to IID sampling (left)." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.658, + 0.907, + 0.839 + ], + "angle": 0, + "content": "desirable objective for many other tasks that use generative models. These include inverse problems (e.g., large hole filling) and class-conditioned image generation, to name a few. Diversity or mode coverage is a key pillar in the generative learning trilemma [32], in addition to fidelity and latency. For state-of-the-art generative methods such as flow matching models (FM) [20, 22] and diffusion models (DM) [12, 26], significant work has been done on improving the photorealism of samples and the efficiency of the sampling process [11, 16, 20, 28, 30, 34]. However, relatively little attention has been paid to explicitly enhancing the diversity of generated samples under a limited sampling budget." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.84, + 0.909, + 0.902 + ], + "angle": 0, + "content": "The standard approach to generate a diverse set of images is to repeatedly obtain independent and identically distributed (IID) samples from a simple or tractable source distribution (e.g., Gaussian distribution), map them to samples" + }, + { + "type": "page_footnote", + "bbox": [ + 0.11, + 0.887, + 0.332, + 0.901 + ], + "angle": 0, + "content": "1Words or phrases with several meanings." + }, + { + "type": "page_number", + "bbox": [ + 0.479, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "23303" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.09, + 0.092, + 0.485, + 0.242 + ], + "angle": 0, + "content": "in the target distribution, and continue this process until we observe sufficient mode coverage in the target distribution. This process, while effective, is sample-inefficient, requiring the generation of more results than necessary. Importantly, the mapping from the source to the target density does not hold a linear relationship; even specifically selecting diverse samples from the source distribution by design does not necessarily yield diverse samples in the target distribution. These limitations naturally raise the following research question." + }, + { + "type": "text", + "bbox": [ + 0.116, + 0.267, + 0.46, + 0.298 + ], + "angle": 0, + "content": "How can we generate diverse samples from the target density under a limited sampling budget?" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.321, + 0.484, + 0.442 + ], + "angle": 0, + "content": "In this paper, we propose DiverseFlow, an inference-time, training-free approach to obtain a diverse set of samples in a desired target density under a fixed sampling budget. We focus on deterministic ODE sampling in continuous-time generative models, specifically FMs, an emerging generative paradigm that enables simulation-free training of continuous normalizing flows (CNFs) and includes diffusion models as a special case." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.445, + 0.484, + 0.718 + ], + "angle": 0, + "content": "DiverseFlow measures the diversity of a set of samples through the volume they span in the target space. A set of similar samples span a lower volume, while a diverse set naturally spans a larger volume. We impose a volume-based gradient constraint on the flow ODE by drawing on determinantal point processes (DPP) [17, 23], a probabilistic model arising from quantum physics that exactly describes the Pauli exclusion principle: that no two fermions may occupy the same quantum state. In Figure 1 consider the images generated via IID sampling (left column) from a text-to-image generative ODE—for both the prompts \"A famous boxer\" and \"A letter\", only a single potential meaning is discovered. Unlike the results generated via IID sampling, samples obtained from DiverseFlow (right column) span more diverse modes corresponding to the polysemous words in the prompts. For \"A famous boxer\", the samples show both a dog breed and a human athlete; for \"A letter\", we observe both written correspondence and alphabet symbols." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.72, + 0.484, + 0.901 + ], + "angle": 0, + "content": "We empirically demonstrate the utility of DiverseFlow across several applications where diversity is inherently desirable. First, we use DiverseFlow to perform text-guided image synthesis for words and phrases that may carry a variety of meanings. Second, we perform large-hole face inpainting with occlusion masks covering significant regions of the face that may be important to the person's identity. Third, we apply DiverseFlow on class-conditioned image synthesis and demonstrate that we can more efficiently explore the data space compared to IID sampling. Lastly, to better characterize and explain the behavior of DiverseFlow, we perform several experiments on synthetic 2D densities." + }, + { + "type": "title", + "bbox": [ + 0.525, + 0.099, + 0.719, + 0.112 + ], + "angle": 0, + "content": "Summary of Contributions" + }, + { + "type": "text", + "bbox": [ + 0.528, + 0.114, + 0.897, + 0.174 + ], + "angle": 0, + "content": "1. We present a sample-efficient method to obtain a diverse set of samples from a flow ODE in Section 5 and demonstrate it qualitatively in Sections 6.1 to 6.3 and quantitatively in Sections 6.4 and 6.5" + }, + { + "type": "text", + "bbox": [ + 0.528, + 0.174, + 0.896, + 0.219 + ], + "angle": 0, + "content": "2. We provide an empirical analysis that demonstrates that our method is consistent across various flow matching formulations (Section 6.4)" + }, + { + "type": "text", + "bbox": [ + 0.528, + 0.22, + 0.896, + 0.28 + ], + "angle": 0, + "content": "3. We introduce the task of text-to-image synthesis from prompts with polysemous words in the context of analyzing sample diversity, and provide a comparison with relevant methods (Section 6.5)" + }, + { + "type": "list", + "bbox": [ + 0.528, + 0.114, + 0.897, + 0.28 + ], + "angle": 0, + "content": null + }, + { + "type": "title", + "bbox": [ + 0.514, + 0.309, + 0.652, + 0.324 + ], + "angle": 0, + "content": "2. Preliminaries" + }, + { + "type": "title", + "bbox": [ + 0.514, + 0.335, + 0.668, + 0.351 + ], + "angle": 0, + "content": "2.1. Flow Matching" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.358, + 0.908, + 0.658 + ], + "angle": 0, + "content": "Many generative models can be considered as a transport map from some easy-to-sample source distribution to an empirically observed yet unknown target distribution. Recent successes in generative modeling represent this transport map in the form of continuous-time processes, such as stochastic differential equations (SDEs) [12, 29], or ordinary differential equations (ODEs) [2, 20, 22]. Although diffusion models are formulated as SDEs, a significant body of research focuses on converting the diffusion SDE to a deterministic ODE at inference time for faster inference. The diffusion ODE, or probability flow ODE, is a particular case of continuous normalizing flows (CNFs). Flow Matching (FM) [2, 20, 22] is motivated by the idea of directly training CNFs in a scalable and simulation-free manner, just like diffusion models. Moreover, many recent text-to-image generative models, such as Stable Diffusion 3 [10], adopt the FM framework. As such, we present our approach primarily in the context of FM, and our findings can be extended to diffusion and score-based generative models in a straightforward manner." + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.661, + 0.908, + 0.691 + ], + "angle": 0, + "content": "A CNF reshapes a prior source density \\( p_0 \\) to the empirically observed target density \\( p_1 \\) with an ODE of the form:" + }, + { + "type": "equation", + "bbox": [ + 0.609, + 0.709, + 0.907, + 0.724 + ], + "angle": 0, + "content": "\\[\nd \\mathbf {x} _ {t} = v _ {\\theta} \\left(\\mathbf {x} _ {t}, t\\right) d t, \\quad \\mathbf {x} _ {0} \\sim p _ {0} \\tag {1}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.734, + 0.907, + 0.87 + ], + "angle": 0, + "content": "where \\( v_{\\theta} \\) is a time-dependent velocity field whose parameters \\( \\theta \\) are learned; we interchangeably use the notation \\( v_{t} \\) to imply \\( v_{\\theta}(\\cdot, t) \\). It becomes possible to obtain samples from \\( p_{1} \\) by integrating Equation (1) over time, i.e. by starting at \\( \\mathbf{x}_0 \\sim p_0 \\) for \\( t = 0 \\) and solving the ODE till \\( t = 1 \\). As our approach is training-free, we do not elaborate on the details of learning to regress the vector field \\( v_{t} \\); we encourage interested readers to refer to the works of Lipman et al. [20] and Tong et al. [30] for a primer on training FMs." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.871, + 0.908, + 0.902 + ], + "angle": 0, + "content": "At any timestep \\( t \\) during sampling, an intermediate sample \\( \\mathbf{x}_t \\) in the flow trajectory can be used to obtain an approx-" + }, + { + "type": "page_number", + "bbox": [ + 0.479, + 0.945, + 0.521, + 0.957 + ], + "angle": 0, + "content": "23304" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.091, + 0.092, + 0.305, + 0.106 + ], + "angle": 0, + "content": "imation of the target as follows:" + }, + { + "type": "equation", + "bbox": [ + 0.194, + 0.12, + 0.484, + 0.136 + ], + "angle": 0, + "content": "\\[\n\\hat {\\mathbf {x}} _ {1} = \\mathbf {x} _ {t} + v _ {\\theta} (\\mathbf {x} _ {t}, t) (1 - t) \\tag {2}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.149, + 0.484, + 0.268 + ], + "angle": 0, + "content": "Equation (2) is equivalent to simply taking a large Euler step at any time instance \\( t \\) and is naturally more accurate as \\( t \\) approaches \\( t = 1 \\). Further, Equation (2) is also well suited for ODEs with 'straight' paths, where the direction of the time-varying velocity \\( v_{t} \\) remains near-constant in time (such as the work of Liu et al. [22]). Similarly, we can estimate the source sample by simply taking a step in the reverse direction:" + }, + { + "type": "equation", + "bbox": [ + 0.215, + 0.271, + 0.484, + 0.287 + ], + "angle": 0, + "content": "\\[\n\\hat {\\mathbf {x}} _ {0} = \\mathbf {x} _ {t} - v _ {\\theta} (\\mathbf {x} _ {t}, t) t \\tag {3}\n\\]" + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.296, + 0.363, + 0.311 + ], + "angle": 0, + "content": "2.2. Determinantal Point Processes" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.318, + 0.485, + 0.456 + ], + "angle": 0, + "content": "Determinantal point processes (DPPs) [4, 17, 23] are probabilistic models of repulsion between points. They were originally termed as fermion processes as they describe the Pauli exclusion principle or antibunching effect in fermions. To define a DPP, we must first consider a set of points, \\(\\mathcal{V}\\), and a point process \\(\\mathcal{P}(\\mathcal{V})\\) — a probability measure on \\(2^{\\mathcal{V}}\\) (the set of all possible subsets of \\(\\mathcal{V}\\)). \\(\\mathcal{P}\\) is determinantal when the probability of choosing a random subset \\(Y \\subset \\mathcal{V}\\) according to \\(\\mathcal{P}\\) is given by:" + }, + { + "type": "equation", + "bbox": [ + 0.13, + 0.466, + 0.484, + 0.502 + ], + "angle": 0, + "content": "\\[\n\\mathcal {P} (Y \\subset \\mathcal {Y}) = \\frac {\\det (\\mathbf {L} _ {Y})}{\\sum_ {Y \\subset \\mathcal {Y}} \\det (\\mathbf {L} _ {Y})} = \\frac {\\det (\\mathbf {L} _ {Y})}{\\det (\\mathbf {L} + \\mathbf {I})} \\tag {4}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.512, + 0.486, + 0.68 + ], + "angle": 0, + "content": "where \\(\\mathbf{L} \\in \\mathbb{R}^{|\\mathcal{Y}| \\times |\\mathcal{Y}|}\\) is a kernel matrix, and \\(\\mathbf{L}_Y\\) is the sub-kernel matrix indexed by the elements of \\(Y\\). Equation (4) has an intuitive geometric interpretation if we consider the kernel \\(\\mathbf{L}\\) to be constructed from cosine similarity: the determinant of \\(\\mathbf{L}_Y\\) is the Gram-determinant, describing the squared volume of the \\(N\\)-dimensional parallelotope spanned by the set of vectors \\(Y\\). Thus, a DPP naturally assigns higher probabilities to more orthogonal (and thus diverse) subsets that span larger volumes. We leverage DPPs to define a coupled likelihood measure over a set of samples in a flow trajectory." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.694, + 0.234, + 0.71 + ], + "angle": 0, + "content": "3. Related Work" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.72, + 0.486, + 0.902 + ], + "angle": 0, + "content": "Efficiently finding diverse sets is useful in several application areas of machine learning. For instance, Batra et al. [3] show that the M-Best MAP (maximum a posteriori) solutions in Markov random fields are often distant from the ground truth and highly similar. They thus propose the *Diverse M-Best* problem—finding a set of \\(M\\) highly probable solutions satisfying some minimum dissimilarity threshold—that partly inspires our study in Section 4. [33] utilize DPPs in conjunction with variational autoencoders (VAE) for diverse trajectory forecasting; a set of diverse future pedestrian trajectories improves safety-critical perception systems in autonomous vehicles. Motivated by potential drug discovery" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.092, + 0.907, + 0.136 + ], + "angle": 0, + "content": "and material design applications, Jain et al. [14] propose finding diverse Pareto-optimal candidates in a multi-objective setting with generative flow networks." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.137, + 0.909, + 0.333 + ], + "angle": 0, + "content": "The work by Corso et al. [6] which explores diverse non-IID sampling for diffusion models is most similar in spirit to DiverseFlow. However, DiverseFlow is notably different in the following aspects: (1) Our diversity objective is derived from determinantal point processes, a diversity-promoting probability measure of the joint occurrence of a set of samples. Corso et al. [6] is instead inspired by stein variational gradient descent (SVGD) [21]. 2) The diversity measure in DiverseFlow (volume, or determinant of similarity kernel) assigns a zero likelihood to a set if any duplicate elements are present; presence of duplicates is tolerated in the diversity measure in Particle Guidance (row-wise sum of similarity kernel)" + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.345, + 0.907, + 0.381 + ], + "angle": 0, + "content": "4. Diverse Source Samples Do Not Yield Diverse Target Samples" + }, + { + "type": "image", + "bbox": [ + 0.551, + 0.399, + 0.71, + 0.52 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.553, + 0.522, + 0.637, + 0.547 + ], + "angle": 0, + "content": "(a) Conditional ing (CFM) [20]" + }, + { + "type": "image", + "bbox": [ + 0.71, + 0.399, + 0.868, + 0.521 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.711, + 0.523, + 0.87, + 0.547 + ], + "angle": 0, + "content": "(b) Mini-batch Optimal Transport (MB-OT) [30]" + }, + { + "type": "image", + "bbox": [ + 0.553, + 0.547, + 0.71, + 0.668 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.565, + 0.67, + 0.699, + 0.682 + ], + "angle": 0, + "content": "(c) DiverseFlow on CFM" + }, + { + "type": "image", + "bbox": [ + 0.71, + 0.547, + 0.868, + 0.669 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.716, + 0.67, + 0.864, + 0.682 + ], + "angle": 0, + "content": "(d) DiverseFlow on MB-OT" + }, + { + "type": "image_caption", + "bbox": [ + 0.513, + 0.695, + 0.907, + 0.751 + ], + "angle": 0, + "content": "Figure 2. Finding \\( K = 5 \\) samples from the target distribution with \\( N = 10 \\) modes. In the example, only 3 modes are found by 5 IID samples (a, b). DiverseFlow discovers 5 modes with the same sampling budget in (d, e)." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.765, + 0.909, + 0.903 + ], + "angle": 0, + "content": "Problem Setting: We start with a synthetic example to illustrate our problem of interest. Consider that we have empirical observations from a target distribution \\(\\pi_1 \\in \\mathbb{R}^2\\), which is a random mixture of Gaussians, such as the example shown in Figure 4. We design \\(\\pi_1 = \\sum_{i=1}^{N} w_i \\mathcal{N}(\\pmb{\\mu}_i, \\sigma_i^2 \\pmb{I})\\) to contain \\(N = 10\\) randomly selected modes \\(\\mathcal{N}(\\pmb{\\mu}_i, \\sigma_i^2 \\pmb{I})\\), each with a random mixture weight \\(w_i\\); we observe that in our example, there are 6 high probability modes and 4 low probability ones. Suppose we have a sampling budget" + }, + { + "type": "page_number", + "bbox": [ + 0.479, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "23305" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.09, + 0.092, + 0.485, + 0.152 + ], + "angle": 0, + "content": "of \\(K\\) samples. This leads to three possible scenarios: (i) \\(K < N\\), (ii) \\(K = N\\), and (iii) \\(K > N\\). Among the aforementioned, case (i) (fewer samples than modes) is the most likely characteristic of any real-world dataset." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.155, + 0.484, + 0.232 + ], + "angle": 0, + "content": "Let us have a prior distribution \\(\\pi_0\\) and some generative model \\(\\Psi\\), such that, in the limit of infinite samples, \\(\\Psi(x_0 \\sim \\pi_0) \\sim \\pi_1\\). Then, the objective of sample-efficient diverse sampling is to obtain samples from \\(\\min(K, N)\\) modes from \\(\\pi_1\\), given a fixed set of \\(K\\) samples in \\(\\pi_0\\)." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.233, + 0.487, + 0.43 + ], + "angle": 0, + "content": "If diverse samples are desired from the target density of the flow, one may make the elementary assumption that if the particles are distant at the source distribution, after being transported by the flow, they remain distant in the target distribution. This assumption is not necessarily true, as we show in Figure 2. By design, we choose a uniform mixture of eight Gaussians as the source \\(\\pi_0\\) to obtain diverse source samples. In Figure 2a, we can observe that source points from distinct modes can still converge to the same target mode with IID sampling. Thus, an alternative procedure is necessary to obtain a diverse set from a flow in a sample-efficient manner. We further explore this toy problem in Section 6.4." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.453, + 0.364, + 0.47 + ], + "angle": 0, + "content": "5. Diverse Sampling from Flows" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.481, + 0.485, + 0.744 + ], + "angle": 0, + "content": "From Figure 2, we observe that independently (or heuristically) chosen source samples may not map to a diverse set of target samples. In this case, we can select a new set of source samples and repeat the sampling process till eventually covering at least \\(K\\) modes. However, this approach does not satisfy our fixed sampling budget constraint. An alternative solution to repeated independent sampling is defining and leveraging a diversity measure of the target samples to drive sample diversity. For the set of source samples \\(\\{\\mathbf{x}_0^{(1)},\\mathbf{x}_0^{(2)},\\ldots ,\\mathbf{x}_0^{(k)}\\}\\), we could optimize a set of perturbations \\(\\{\\delta^{(1)},\\delta^{(2)},\\dots,\\delta^{(k)}\\}\\) such that the new set \\(\\{\\mathbf{x}_0^{(1)} + \\delta^{(1)},\\mathbf{x}_0^{(2)} + \\delta^{(2)},\\dots,\\mathbf{x}_0^{(k)} + \\delta^{(k)}\\}\\) maps to a diverse set of target particles. However, this approach would require multiple simulations of the whole ODE and backpropagating over all the timesteps, which increases the computational complexity of the sampling process over the standard IID sampling." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.746, + 0.484, + 0.903 + ], + "angle": 0, + "content": "This leads us to our proposed approach: we avoid multiple simulations and instead optimize the flow trajectory for diversity while solving the ODE. For any sample in the flow trajectory \\(\\mathbf{x}_t\\), we have an estimate of the target sample \\(\\hat{\\mathbf{x}}_1\\) through Equation (2). Suppose we have a differentiable objective \\(\\mathcal{L}(\\{\\hat{\\mathbf{x}}_1^{(1)},\\hat{\\mathbf{x}}_1^{(2)},\\dots ,\\hat{\\mathbf{x}}_1^{(k)}\\})\\) that assigns a likelihood to the joint outcome \\(\\{\\hat{\\mathbf{x}}_1^{(1)},\\hat{\\mathbf{x}}_1^{(2)},\\dots ,\\hat{\\mathbf{x}}_1^{(k)}\\}\\). Further, let \\(\\mathcal{L}\\) assign a higher likelihood if the joint outcome is a diverse set, and a diminished likelihood if the set is similar. We can then leverage \\(\\mathcal{L}\\) to drive diversity among the target samples" + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.092, + 0.866, + 0.108 + ], + "angle": 0, + "content": "by modifying the flow velocity of the \\(i\\)-th particle as," + }, + { + "type": "equation", + "bbox": [ + 0.524, + 0.116, + 0.907, + 0.141 + ], + "angle": 0, + "content": "\\[\n\\tilde {v} _ {t} ^ {(i)} = v _ {t} ^ {(i)} - \\gamma (t) \\nabla_ {\\mathbf {x} _ {t} ^ {(i)}} \\log \\mathcal {L} (\\{\\hat {\\mathbf {x}} _ {1} ^ {(1)}, \\hat {\\mathbf {x}} _ {1} ^ {(2)}, \\dots , \\hat {\\mathbf {x}} _ {1} ^ {(k)} \\}) \\tag {5}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.151, + 0.907, + 0.243 + ], + "angle": 0, + "content": "where \\(\\gamma (t)\\) is a time-varying scale that controls the strength of the diversity gradient. Setting \\(\\gamma (t) = 0\\) reduces to the standard IID sampling scenario, while \\(\\gamma (t) > 0\\) will encourage diversity between the generated samples. In practice, \\(\\gamma (t)\\) follows the schedule of the probability path normalized by the norm of the DPP gradient." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.251, + 0.831, + 0.266 + ], + "angle": 0, + "content": "5.1. Determinantal Gradient Constraints" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.272, + 0.907, + 0.439 + ], + "angle": 0, + "content": "We desire objective \\(\\mathcal{L}\\) in Equation (5) to be higher if the items in the set are diverse and lower if they are similar to each other. We interpret diversity in terms of the volume spanned by the set. Consider that we have \\(k\\) samples in \\(\\mathbb{R}^d\\) (assume \\(k < d\\)). An objective that prefers diversity can be defined as the volume of the \\(k\\)-dimensional parallelotope in \\(\\mathbb{R}^d\\) spanned by the sample vectors; this volume becomes diminished when there are similar samples (and even zero, for identical samples). The determinant describes volumes well; a diverse set must span a large volume in the sample space and have a corresponding large determinant." + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.44, + 0.906, + 0.488 + ], + "angle": 0, + "content": "To define a measure over a set of samples, we draw on the idea of determinantal point processes (DPP). We first define a kernel \\(\\mathbf{L}(\\{\\hat{\\mathbf{x}}_1^{(1)},\\hat{\\mathbf{x}}_1^{(2)},\\dots ,\\hat{\\mathbf{x}}_1^{(k)}\\})\\) as follows:" + }, + { + "type": "equation", + "bbox": [ + 0.591, + 0.497, + 0.907, + 0.539 + ], + "angle": 0, + "content": "\\[\n\\mathbf {L} ^ {(i j)} = \\exp \\left(- h \\frac {\\| \\hat {\\mathbf {x}} _ {1} ^ {(i)} - \\hat {\\mathbf {x}} _ {1} ^ {(j)} \\| _ {2} ^ {2}}{\\operatorname* {m e d} (\\mathbf {U} (\\mathbf {D}))}\\right) \\tag {6}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.549, + 0.906, + 0.626 + ], + "angle": 0, + "content": "where \\(\\mathbf{D}\\) denotes a distance matrix with \\(\\mathbf{D}_{ij} = \\| \\mathbf{x}^{(i)} - \\mathbf{x}^{(j)}\\| _2^2\\), \\(\\mathbf{U}(\\mathbf{D})\\) denotes the upper triangle entries of \\(\\mathbf{D}\\), \\(h\\) denotes a kernel spread parameter, and \\(\\mathrm{med}(\\mathbf{U}(\\mathbf{D}))\\) denotes the median of those entries. Given \\(\\mathbf{L}\\), we may define a DPP-based likelihood as:" + }, + { + "type": "equation", + "bbox": [ + 0.563, + 0.648, + 0.907, + 0.725 + ], + "angle": 0, + "content": "\\[\n\\begin{array}{l} \\mathcal {L} (\\{\\hat {\\mathbf {x}} _ {1} ^ {(1)}, \\hat {\\mathbf {x}} _ {1} ^ {(2)}, \\dots , \\hat {\\mathbf {x}} _ {1} ^ {(k)} \\}) = \\frac {\\det (\\mathbf {L})}{\\det (\\mathbf {L} + \\mathbf {I})} \\\\ = \\prod_ {a = 1} ^ {k} \\frac {\\lambda (\\mathbf {L}) _ {a}}{1 + \\lambda (\\mathbf {L}) _ {a}} \\tag {7} \\\\ \\end{array}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.736, + 0.906, + 0.768 + ], + "angle": 0, + "content": "where \\(\\lambda (\\mathbf{L})_a\\) is the \\(a^{\\mathrm{th}}\\) eigenvalue of the kernel \\(\\mathbf{L}\\). The log-likelihood is then," + }, + { + "type": "equation", + "bbox": [ + 0.563, + 0.778, + 0.907, + 0.795 + ], + "angle": 0, + "content": "\\[\n\\mathcal {L} \\mathcal {L} = \\log \\mathcal {L} = \\log \\det (\\mathbf {L}) - \\log \\det (\\mathbf {L} + \\mathbf {I}) \\tag {8}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.806, + 0.907, + 0.902 + ], + "angle": 0, + "content": "Note that the Euclidean distance \\(\\| \\hat{\\mathbf{x}}_1^{(i)} - \\hat{\\mathbf{x}}_1^{(j)}\\| _2^2\\) is not very meaningful in the high-dimensional raw image space [1]. Therefore, in practice, the distance should be computed in a robust feature space, i.e., \\(\\| F(\\hat{\\mathbf{x}}_1^{(i)}) - F(\\hat{\\mathbf{x}}_1^{(j)})\\| _2^2\\), where \\(F\\) is some domain-specific feature extractor, such as the vision transformer (ViT) [9] for images." + }, + { + "type": "page_number", + "bbox": [ + 0.479, + 0.945, + 0.521, + 0.958 + ], + "angle": 0, + "content": "23306" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.09, + 0.092, + 0.485, + 0.328 + ], + "angle": 0, + "content": "Quality Constraint: The DPP defined in Equation (7) acts as a repulsive force that is unaware of sample quality. A quality term can be incorporated into the DPP kernel to regularize the trajectory diversification. Although flows can be defined between any arbitrary two distributions, let us consider the special case when the source is a Gaussian, i.e., \\(p_0 \\sim \\mathcal{N}(0,\\mathbf{I})\\). Suppose we have a quality vector \\(\\mathbf{q}_t = \\{q^{(1)}(t), q^{(2)}(t), \\ldots, q^{(k)}(t)\\}\\), where any \\(q^{(i)}(t) \\in [0,1]\\). We can then define a new kernel \\(\\mathbf{L}_q = \\mathbf{L} \\odot \\mathbf{q}_t \\mathbf{q}_t^T\\), where each \\(q^{(i)}(t)\\) penalizes a sample \\(\\mathbf{x}_t^{(i)}\\) if it deviates too much from the flow. To define this, we obtain an estimate of the source sample \\(\\hat{\\mathbf{x}}_0^{(i)}(t)\\) for any given sample \\(\\mathbf{x}_t^{(i)}\\) via Equation (3), and check if it lies within a desired percentile-radius \\(\\rho\\) of the Gaussian \\(p_0\\). Specifically, we define the time-dependent sample quality as" + }, + { + "type": "equation", + "bbox": [ + 0.09, + 0.335, + 0.495, + 0.398 + ], + "angle": 0, + "content": "\\[\nq ^ {(i)} (t) = \\left\\{ \\begin{array}{l l} 1 & \\text {i f} \\| \\hat {\\mathbf {x}} _ {0} ^ {(i)} (t) \\| _ {2} ^ {2} \\leq \\rho^ {2} \\\\ \\max \\left(\\epsilon , e ^ {- \\left(\\| \\hat {\\mathbf {x}} _ {0} ^ {(i)} (t) \\| _ {2} ^ {2} - \\rho^ {2}\\right)}\\right) & \\text {o t h e r w i s e} \\end{array} \\right. \\tag {9}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.399, + 0.484, + 0.429 + ], + "angle": 0, + "content": "where \\(\\epsilon\\) is a 'minimum quality' we assign to prevent a zero determinant." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.437, + 0.445, + 0.453 + ], + "angle": 0, + "content": "5.2. Coupled Ordinary Differential Equations" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.46, + 0.484, + 0.521 + ], + "angle": 0, + "content": "At any timestep \\( t \\), the measure of diversity in Equation (8) can be adopted to modify the flow of the \\( i \\)-th particle. We compute the gradient of the samples with respect to the diversity measure and use it to modify the ODE as follows:" + }, + { + "type": "equation", + "bbox": [ + 0.09, + 0.528, + 0.493, + 0.565 + ], + "angle": 0, + "content": "\\[\n\\left. d \\mathbf {x} _ {t} ^ {(i)} = \\left[ v _ {\\theta} \\left(\\mathbf {x} _ {t} ^ {(i)}, t\\right) - \\gamma (t) \\nabla_ {\\mathbf {x} _ {t} ^ {(i)}} \\log \\mathcal {L} \\left(\\left\\{\\hat {\\mathbf {x}} _ {1} ^ {(1)}, \\dots , \\hat {\\mathbf {x}} _ {1} ^ {(k)} \\right\\}\\right) \\right] d t \\right. \\tag {10}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.567, + 0.484, + 0.743 + ], + "angle": 0, + "content": "Where \\(\\gamma (t)\\) is a time-varying scaling factor. Unlike the IID sampling scenario where we have \\(K\\) independent ODEs, Equation (10) corresponds to a system of coupled non-linear ordinary differential ordinary equations. To see this, first note that the estimate \\(\\hat{\\mathbf{x}}_1^{(i)}\\) depends on the current sample \\(\\mathbf{x}_t^{(i)}\\) i.e., \\(\\hat{\\mathbf{x}}_1^{(i)} = \\mathbf{x}_t^{(i)} + v_\\theta (\\mathbf{x}_t^{(i)},t)(1 - t)\\). Second, the DPP log-likelihood \\(\\mathcal{LL}(\\{\\hat{\\mathbf{x}}_1^{(1)},\\hat{\\mathbf{x}}_1^{(2)},\\dots ,\\hat{\\mathbf{x}}_1^{(k)}\\})\\) induces a time-dependent coupling between the \\(K\\) trajectories of \\(\\mathbf{x}_t^{(i)},i = 1,\\ldots ,K\\) and seeks to diversify the target samples. Although higher-order ODE solvers [16] can be employed to solve the coupled ODEs, we use the standard Euler method." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.755, + 0.224, + 0.772 + ], + "angle": 0, + "content": "6. Experiments" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.78, + 0.484, + 0.903 + ], + "angle": 0, + "content": "We demonstrate the utility of DiverseFlow in flow-based generative models by considering three applications where sample diversity is naturally desirable: text-guided image generation with polysemous words, large-hole inpainting, and class-conditional image generation. We also analyze the effect of DiverseFlow on different flow matching formulations w.r.t. its ability to span diverse modes through a synthetically constructed 2D density example." + }, + { + "type": "image", + "bbox": [ + 0.543, + 0.09, + 0.712, + 0.22 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.712, + 0.091, + 0.878, + 0.22 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.654, + 0.225, + 0.756, + 0.238 + ], + "angle": 0, + "content": "(a) \"A buck\"" + }, + { + "type": "image", + "bbox": [ + 0.542, + 0.24, + 0.712, + 0.37 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.712, + 0.24, + 0.878, + 0.37 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.616, + 0.375, + 0.798, + 0.388 + ], + "angle": 0, + "content": "(b) \"A famous boxer\"" + }, + { + "type": "image", + "bbox": [ + 0.542, + 0.39, + 0.712, + 0.521 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.712, + 0.391, + 0.878, + 0.521 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.607, + 0.525, + 0.816, + 0.541 + ], + "angle": 0, + "content": "(c) \"Van Gogh painting\"" + }, + { + "type": "image_caption", + "bbox": [ + 0.512, + 0.55, + 0.908, + 0.635 + ], + "angle": 0, + "content": "Figure 3. For each prompt, the left image (red box) denotes standard IID sampling with classifier-free guidance, while the right image (blue box) shows the result after incorporating DiverseFlow. DiverseFlow finds more diverse sets given the same source points—clearly distinguishable as new semantic meanings in the case of prompts with multiple meanings." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.664, + 0.896, + 0.681 + ], + "angle": 0, + "content": "6.1. Image Generation from Polysemous Prompts" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.688, + 0.906, + 0.809 + ], + "angle": 0, + "content": "In text-to-image generation, the conditional data distribution corresponding to a text prompt may contain many variations, and it is a desirable objective to generate images that span those variations in a sample-efficient manner. As we have previously shown in Figure 1, the result desired by a user (a human boxer athlete) may not be the result generated by the model (a dog breed); if diverse results are obtained, it allows a user to obtain the desired results with fewer attempts." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.811, + 0.907, + 0.901 + ], + "angle": 0, + "content": "We pose a scenario where diverse sets are easily observable: when an open-ended text prompt is polysemous and carries multiple meanings, such as the examples we show in Figure 1 and Figure 3. In Figure 3(a), the prompt \"A buck\" may commonly refer to a male deer. However, it may also informally refer to a United States dollar. Using the" + }, + { + "type": "page_number", + "bbox": [ + 0.479, + 0.945, + 0.518, + 0.957 + ], + "angle": 0, + "content": "23307" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.09, + 0.092, + 0.485, + 0.319 + ], + "angle": 0, + "content": "same four source points, which are deterministically mapped to four deer images by IID sampling. DiverseFlow finds a different set of samples—one that includes a dollar-like coin, albeit embossed with a deer head. We also observe minor differences between the two sets of images, such as changes in pose and background in the top-right and bottom-right deers. For Figure 3(c), while 'Van Gogh painting' is not quite a polysemous word, it can still have two interpretations: a painting painted by Van Gogh, or a painting of Van Gogh. The regular samples contain minimal diversity, as they include two sets of repeated paintings of Van Gogh. With DiverseFlow, not only can we get a set of four distinct paintings, but we also have a portrait of Van Gogh, which is one of the additional meanings of the prompt. We present additional qualitative samples in the supplementary material." + }, + { + "type": "title", + "bbox": [ + 0.09, + 0.349, + 0.342, + 0.366 + ], + "angle": 0, + "content": "6.2. Diverse Inpainting on Faces" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.378, + 0.485, + 0.591 + ], + "angle": 0, + "content": "Another inverse problem where diverse solutions are desirable is face inpainting, where we seek to inpaint the missing parts of the face with diverse plausible facial textures and structures. To demonstrate inpainting with FM models, we first incorporate Manifold Constrained Gradient (MCG) [5] in an off-the-shelf unconditional Rectified-Flow model. In addition to the manifold constraints, we employ determinantal gradient constraints to enhance diversity (further detailed in Algorithm 1). In Figure 5 (b), we observe that the inpainted faces of the four women have similar expressions (largely neutral). DiverseFlow improves the diversity of the set by yielding a highly different expression in the top-right image. In (d) and (e), we also observe changes in facial hair and expressions due to diversification." + }, + { + "type": "image", + "bbox": [ + 0.127, + 0.63, + 0.191, + 0.731 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.192, + 0.63, + 0.319, + 0.731 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.321, + 0.63, + 0.451, + 0.731 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.126, + 0.733, + 0.191, + 0.835 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.155, + 0.838, + 0.175, + 0.852 + ], + "angle": 0, + "content": "(a)" + }, + { + "type": "image", + "bbox": [ + 0.192, + 0.734, + 0.319, + 0.834 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.243, + 0.838, + 0.264, + 0.852 + ], + "angle": 0, + "content": "(b)" + }, + { + "type": "image", + "bbox": [ + 0.321, + 0.734, + 0.451, + 0.835 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.371, + 0.838, + 0.393, + 0.852 + ], + "angle": 0, + "content": "(c)" + }, + { + "type": "image_caption", + "bbox": [ + 0.09, + 0.863, + 0.485, + 0.901 + ], + "angle": 0, + "content": "Figure 5. Inpainting on CelebAHQ-256 × 256. (a) Dashed boxes show masked input (top) and ground truth (bottom) respectively. (b) Rectified-Flow [22] + MCG [5] (c) With DiverseFlow." + }, + { + "type": "image", + "bbox": [ + 0.521, + 0.104, + 0.647, + 0.201 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.521, + 0.203, + 0.646, + 0.299 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.558, + 0.306, + 0.616, + 0.319 + ], + "angle": 0, + "content": "No CFG" + }, + { + "type": "image", + "bbox": [ + 0.648, + 0.104, + 0.775, + 0.201 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.648, + 0.203, + 0.774, + 0.299 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.673, + 0.306, + 0.743, + 0.319 + ], + "angle": 0, + "content": "With CFG" + }, + { + "type": "image", + "bbox": [ + 0.775, + 0.104, + 0.902, + 0.201 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.775, + 0.203, + 0.902, + 0.299 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.8, + 0.306, + 0.885, + 0.319 + ], + "angle": 0, + "content": "DiverseFlow" + }, + { + "type": "image_caption", + "bbox": [ + 0.513, + 0.329, + 0.908, + 0.378 + ], + "angle": 0, + "content": "Figure 6. Class-conditional ImageNet samples from LFM [7]. We show samples for two classes, (top row) 'Mushroom' (class 947) and (bottom row) 'Macaw' (class 88). (a, d) No CFG. (b, e) LFM with CFG. (d, f) LFM with CFG and DiverseFlow." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.396, + 0.879, + 0.413 + ], + "angle": 0, + "content": "6.3. Diverse Class-Conditional Image Synthesis" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.418, + 0.908, + 0.553 + ], + "angle": 0, + "content": "Suppose we can access a class-conditioned flow matching (FM) model trained on an unknown image dataset. To explore the unobservable true dataset, we may use a set of class-conditional samples from the FM model. We adopt a latent flow matching (LFM) model [7], trained to generate \\(256 \\times 256\\) resolution images from the ImageNet [8] dataset. Much like latent diffusion, LFM employs classifier-free guidance to create high-quality samples. However, this naturally poses a cost to diversity, as we show in Figure 6." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.555, + 0.909, + 0.797 + ], + "angle": 0, + "content": "By incorporating DiverseFlow, we can maintain the high quality of the samples and simultaneously explore more modes in the dataset. In Figure 6, we demonstrate two ImageNet classes that may have diversity: 'Mushroom' and 'Macaw.' For mushrooms, we observe that LFM primarily generates two species of mushrooms. However, by applying DiverseFlow, we successfully find a new species within our limited set: an Amanita muscaria, also known as the fly agaric—easily distinguishable by the white spots on its red cap. In another example, we see that while LFM generates four scarlet macaws, using the same source samples, DiverseFlow helps us find a different blue and yellow macaw. We further evaluate the precision and recall [18] for ImageNet-256 synthesis in Figure 7, showing that DiverseFlow improves the recall while retaining similar precision, suggesting improved diversity." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.803, + 0.907, + 0.819 + ], + "angle": 0, + "content": "6.4. DiverseFlow Across Various FM Formulations" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.826, + 0.909, + 0.903 + ], + "angle": 0, + "content": "We ask the question: does DiverseFlow remain consistent across different formulations of flow matching? We adopt the same toy bivariate mixture of Gaussian distributions as in the example shown in Figure 2, across four different FM formulations: (i) Conditional Flow Matching (CFM) [20]," + }, + { + "type": "page_number", + "bbox": [ + 0.48, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "23308" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.092, + 0.087, + 0.905, + 0.258 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.09, + 0.268, + 0.908, + 0.298 + ], + "angle": 0, + "content": "Figure 4. Comparing different FM formulations in terms of the number of modes spanned by IID sampling versus with DiverseFlow. More details about the experiment are provided in the supplementary." + }, + { + "type": "image", + "bbox": [ + 0.093, + 0.32, + 0.275, + 0.44 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.091, + 0.451, + 0.275, + 0.478 + ], + "angle": 0, + "content": "Figure 7. Precision vs. recall on ImageNet-256." + }, + { + "type": "table", + "bbox": [ + 0.293, + 0.32, + 0.48, + 0.44 + ], + "angle": 0, + "content": "
MethodCFGP (↑)R (↑)
LFM10.620.52
LFM1.50.690.44
LFM20.770.41
LFM40.690.26
LFM + DF10.620.52
LFM+DF1.50.690.47
LFM+DF20.760.46
LFM+DF40.700.38
" + }, + { + "type": "table_caption", + "bbox": [ + 0.289, + 0.451, + 0.483, + 0.476 + ], + "angle": 0, + "content": "Table 1. Precision (P) and recall (R) score of DiverseFlow and baselines." + }, + { + "type": "text", + "bbox": [ + 0.089, + 0.504, + 0.486, + 0.718 + ], + "angle": 0, + "content": "(ii) Mini-batch Optimal Transport CFM (MB-OT) [30], (iii) Schrödinger Bridge CFM (SB-CFM) [31], and (iv) Stochastic Interpolants (SI-CFM) [2]. We then perform a numerical experiment to quantify the average number of modes discovered by each FM variant with increasing the sampling budget \\( K \\). Figure 4 reports the results. For a maximum sampling budget of 10, MB-OT discovers only 5.64 modes on average, which is expected since the dataset contains 6 high-probability modes. By incorporating DiverseFlow, we can find 7.11 modes on average. We observe that MB-OT and SB-CFM benefit most from DiverseFlow. We hypothesize this is because MB-OT and SB-CFM are formulated with the notion of optimal paths, which results in a more accurate estimation of \\( \\hat{x}_1 \\) and thus a better diversity gradient." + }, + { + "type": "title", + "bbox": [ + 0.09, + 0.728, + 0.391, + 0.743 + ], + "angle": 0, + "content": "6.5. Comparison To Particle Guidance" + }, + { + "type": "text", + "bbox": [ + 0.089, + 0.75, + 0.486, + 0.903 + ], + "angle": 0, + "content": "(i) Preventing Training Data Copy: Previously, Corso et al. [6] demonstrate that their method can alleviate Stable Diffusion's training data regurgitation problem [27] to some extent. In Figure 8, we demonstrate similar capabilities; in (b), the top left and bottom right examples are copies of the training data. Subsequently, (c) and (d) find a new example for the top-left sample. (ii) Diversity vs Quality: We also compare the diversity versus quality of DiverseFlow against Particle Guidance in Figure 10 over 30 polysemous prompts repeated over 10 random seeds. While varying the" + }, + { + "type": "image", + "bbox": [ + 0.518, + 0.32, + 0.71, + 0.469 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.567, + 0.472, + 0.66, + 0.483 + ], + "angle": 0, + "content": "(a) Training Data" + }, + { + "type": "image", + "bbox": [ + 0.71, + 0.32, + 0.905, + 0.47 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.761, + 0.471, + 0.854, + 0.484 + ], + "angle": 0, + "content": "(b) IID Sampling" + }, + { + "type": "image", + "bbox": [ + 0.518, + 0.484, + 0.71, + 0.633 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.548, + 0.635, + 0.68, + 0.647 + ], + "angle": 0, + "content": "(c) Particle Guidance [6]" + }, + { + "type": "image", + "bbox": [ + 0.71, + 0.484, + 0.904, + 0.633 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.763, + 0.635, + 0.851, + 0.646 + ], + "angle": 0, + "content": "(d) DiverseFlow" + }, + { + "type": "image_caption", + "bbox": [ + 0.512, + 0.659, + 0.908, + 0.714 + ], + "angle": 0, + "content": "Figure 8. For the prompt \"VAN GOGH CAFE TERASSE copy.jpg\", we show (a) the training data, (b) IID sampling with two copies in top-left and bottom-right, (c) Particle Guidance [6], and (d) DiverseFlow." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.75, + 0.909, + 0.903 + ], + "angle": 0, + "content": "strength of DiverseFlow to improve diversity, we also vary the values of the classifier-free guidance strength from 7.5 to 10 to boost quality. Quality is measured by Aesthetic Score (higher is better) [25], and diversity is measured by average pairwise similarity of a set (lower is better) [6]. We observe that though DiverseFlow obtains better diversity at similar quality to IID sampling, the aesthetic score of Particle Guidance is unusually low. In Figure 9 we highlight the cause: for the challenging task of polysemous text-to-image generation, the images generated from Particle Guidance (a) suffers" + }, + { + "type": "page_number", + "bbox": [ + 0.479, + 0.945, + 0.521, + 0.958 + ], + "angle": 0, + "content": "23309" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.092, + 0.089, + 0.482, + 0.537 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.09, + 0.549, + 0.485, + 0.587 + ], + "angle": 0, + "content": "Figure 9. Particle Guidance (middle rows) can suffer from artefacts for polysemous prompts. DiverseFlow retains quality while achieving diversity: for \"A mint\", a minted coin with a mint leaf was discovered." + }, + { + "type": "image", + "bbox": [ + 0.116, + 0.601, + 0.46, + 0.799 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.136, + 0.814, + 0.438, + 0.829 + ], + "angle": 0, + "content": "Figure 10. In-batch similarity \\((\\downarrow)\\) vs aesthetic score \\((\\uparrow)\\)" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.856, + 0.485, + 0.903 + ], + "angle": 0, + "content": "from image artefacts, which achieves a high diversity score but a poor quality score. Additional details about the experiment, including the set of 30 prompts, are provided in the" + }, + { + "type": "image", + "bbox": [ + 0.517, + 0.089, + 0.905, + 0.398 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.513, + 0.412, + 0.907, + 0.438 + ], + "angle": 0, + "content": "Figure 11. Average modes discovered over 100 trials, for varying noise scales and strengths of DiverseFlow and ParticleGuidance." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.466, + 0.909, + 0.692 + ], + "angle": 0, + "content": "supplementary. (iii) Mode Discovery: In high dimensional data, the number of modes, \\( N \\), is significantly greater than the budget \\( K \\), so finding unique modes is still highly probable. To better highlight the differences between DiverseFlow and Particle Guidance, we adopt a simpler toy experiment: finding modes in a symmetric uniform Mixture of Gaussian distribution. Corso et al. [6] provides the result that IID sampling with budget \\( K = 10 \\) discovers about 6.5 modes on average, while Particle Guidance with a Euclidean kernel discovers almost 9 out of 10 modes. We verify this result in Figure 11, finding that Particle Guidance discovers up to 8.8 modes (averaged over 100 trials). However, by using DiverseFlow, it is possible to discover all 10 modes on average, showing that our approach has a stronger diversification effect." + }, + { + "type": "title", + "bbox": [ + 0.514, + 0.708, + 0.634, + 0.724 + ], + "angle": 0, + "content": "7. Conclusion" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.735, + 0.909, + 0.901 + ], + "angle": 0, + "content": "In this paper, we present DiverseFlow, a way to enforce diversity among a limited set of samples generated from a flow by coupling them through a determinantal point process. We demonstrate multiple applications where our method can be useful. In essence, though flow models admit a deterministic mapping from source to sample, DiverseFlow allows us to modify this mapping at inference-time and obtain samples with improved diversity. Diversity enhances the utility of the underlying generative flow in the case of ill-posed problems, where many solutions exist and it is desirable to find a multitude of such solutions." + }, + { + "type": "page_number", + "bbox": [ + 0.479, + 0.945, + 0.521, + 0.958 + ], + "angle": 0, + "content": "23310" + } + ], + [ + { + "type": "title", + "bbox": [ + 0.093, + 0.09, + 0.188, + 0.106 + ], + "angle": 0, + "content": "References" + }, + { + "type": "ref_text", + "bbox": [ + 0.101, + 0.115, + 0.486, + 0.185 + ], + "angle": 0, + "content": "[1] Charu C Aggarwal, Alexander Hinneburg, and Daniel A Keim. On the surprising behavior of distance metrics in high dimensional space. In Database Theory—ICDT 2001: 8th International Conference London, UK, January 4–6, 2001 Proceedings 8, pages 420–434. Springer, 2001. 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.101, + 0.187, + 0.485, + 0.241 + ], + "angle": 0, + "content": "[2] Michael S Albergo, Nicholas M Boffi, and Eric Vanden-Eijnden. Stochastic interpolants: A unifying framework for flows and diffusions. arXiv preprint arXiv:2303.08797, 2023. 2, 7, 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.245, + 0.484, + 0.327 + ], + "angle": 0, + "content": "[3] Dhruv Batra, Payman Yadollahpour, Abner Guzman-Rivera, and Gregory Shakhnarovich. Diverse m-best solutions in markov random fields. In Computer Vision-ECCV 2012: 12th European Conference on Computer Vision, Florence, Italy, October 7-13, 2012, Proceedings, Part V 12, pages 1-16. Springer, 2012. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.33, + 0.484, + 0.383 + ], + "angle": 0, + "content": "[4] Alexei Borodin and Grigori Olshanski. Distributions on partitions, point processes, and the hypergeometric kernel. Communications in Mathematical Physics, 211:335-358, 2000. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.387, + 0.484, + 0.442 + ], + "angle": 0, + "content": "[5] Hyungjin Chung, Byeongsu Sim, Dohoon Ryu, and Jong Chul Ye. Improving diffusion models for inverse problems using manifold constraints. Advances in Neural Information Processing Systems, 35:25683-25696, 2022. 6, 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.444, + 0.484, + 0.499 + ], + "angle": 0, + "content": "[6] Gabriele Corso, Yilun Xu, Valentin De Bortoli, Regina Barzilay, and Tommi Jaakkola. Particle guidance: non-iid diverse sampling with diffusion models. International Conference on Learning Representations, 2023. 3, 7, 8, 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.501, + 0.484, + 0.542 + ], + "angle": 0, + "content": "[7] Quan Dao, Hao Phung, Binh Nguyen, and Anh Tran. Flow matching in latent space. arXiv preprint arXiv:2307.08698, 2023. 6, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.545, + 0.484, + 0.6 + ], + "angle": 0, + "content": "[8] Jia Deng, Wei Dong, Richard Socher, Li-Jia Li, Kai Li, and Li Fei-Fei. Imagenet: A large-scale hierarchical image database. In 2009 IEEE conference on computer vision and pattern recognition, pages 248-255. IEEE, 2009. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.602, + 0.484, + 0.685 + ], + "angle": 0, + "content": "[9] Alexey Dosovitskiy, Lucas Beyer, Alexander Kolesnikov, Dirk Weissenborn, Xiaohua Zhai, Thomas Unterthiner, Mostafa Dehghani, Matthias Minderer, Georg Heigold, Sylvain Gelly, et al. An image is worth 16x16 words: Transformers for image recognition at scale. International Conference on Learning Representations, 2021. 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.687, + 0.485, + 0.77 + ], + "angle": 0, + "content": "[10] Patrick Esser, Sumith Kulal, Andreas Blattmann, Rahim Entezari, Jonas Müller, Harry Saini, Yam Levi, Dominik Lorenz, Axel Sauer, Frederic Boesel, et al. Scaling rectified flow transformers for high-resolution image synthesis. In International Conference on Machine Learning, pages 12606-12633. PMLR, 2024. 2, 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.772, + 0.482, + 0.801 + ], + "angle": 0, + "content": "[11] Jonathan Ho and Tim Salimans. Classifier-free diffusion guidance. arXiv preprint arXiv:2207.12598, 2022. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.803, + 0.484, + 0.843 + ], + "angle": 0, + "content": "[12] Jonathan Ho, Ajay Jain, and Pieter Abbeel. Denoising diffusion probabilistic models. Advances in neural information processing systems, 33:6840-6851, 2020. 1, 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.846, + 0.482, + 0.901 + ], + "angle": 0, + "content": "[13] Noboru Isobe, Masanori Koyama, Jinzhe Zhang, Kohei Hayashi, and Kenji Fukumizu. Extended flow matching: a method of conditional generation with generalized continuity equation. arXiv preprint arXiv:2402.18839, 2024. 5" + }, + { + "type": "list", + "bbox": [ + 0.094, + 0.115, + 0.486, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.093, + 0.908, + 0.161 + ], + "angle": 0, + "content": "[14] Moksh Jain, Sharath Chandra Raparthy, Alex Hernández-García, Jarrid Rector-Brooks, Yoshua Bengio, Santiago Miret, and Emmanuel Bengio. Multi-objective gflownets. In International Conference on Machine Learning, pages 14631-14653. PMLR, 2023. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.163, + 0.908, + 0.217 + ], + "angle": 0, + "content": "[15] Tero Karras, Timo Aila, Samuli Laine, and Jaakko Lehtinen. Progressive growing of gans for improved quality, stability, and variation. International Conference on Learning Representations, 2018. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.219, + 0.907, + 0.272 + ], + "angle": 0, + "content": "[16] Tero Karras, Miika Aittala, Timo Aila, and Samuli Laine. Elucidating the design space of diffusion-based generative models. Advances in Neural Information Processing Systems, 2022. 1, 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.274, + 0.907, + 0.315 + ], + "angle": 0, + "content": "[17] Alex Kulesza, Ben Taskar, et al. Determinantal point processes for machine learning. Foundations and Trends® in Machine Learning, 5(2-3):123-286, 2012. 2, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.317, + 0.906, + 0.372 + ], + "angle": 0, + "content": "[18] Tuomas Kynkänniemi, Tero Karras, Samuli Laine, Jaakko Lehtinen, and Timo Aila. Improved precision and recall metric for assessing generative models. Advances in neural information processing systems, 32, 2019. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.373, + 0.907, + 0.427 + ], + "angle": 0, + "content": "[19] Yiheng Li, Heyang Jiang, Akio Kodaira, Masayoshi Tomizuka, Kurt Keutzer, and Chenfeng Xu. Immiscible diffusion: Accelerating diffusion training with noise assignment. Advances in neural information processing systems, 2024. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.428, + 0.907, + 0.468 + ], + "angle": 0, + "content": "[20] Yaron Lipman, Ricky TQ Chen, Heli Ben-Hamu, Maximilian Nickel, and Matt Le. Flow matching for generative modeling. arXiv preprint arXiv:2210.02747, 2022. 1, 2, 3, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.47, + 0.906, + 0.511 + ], + "angle": 0, + "content": "[21] Qiang Liu and Dilin Wang. Stein variational gradient descent: A general purpose bayesian inference algorithm. Advances in neural information processing systems, 29, 2016. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.512, + 0.906, + 0.553 + ], + "angle": 0, + "content": "[22] Xingchao Liu, Chengyue Gong, and Qiang Liu. Flow straight and fast: Learning to generate and transfer data with rectified flow. arXiv preprint arXiv:2209.03003, 2022. 1, 2, 3, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.554, + 0.906, + 0.593 + ], + "angle": 0, + "content": "[23] Odile Macchi. The coincidence approach to stochastic point processes. Advances in Applied Probability, 7(1):83-122, 1975. 2, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.596, + 0.908, + 0.664 + ], + "angle": 0, + "content": "[24] Robin Rombach, Andreas Blattmann, Dominik Lorenz, Patrick Esser, and Björn Ommer. High-resolution image synthesis with latent diffusion models. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 10684-10695, 2022. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.665, + 0.908, + 0.72 + ], + "angle": 0, + "content": "[25] Christoph Schuhmann and LAION. Improved aesthetic predictor: Clip+mlp aesthetic score predictor, 2022. https://github.com/christophschuhmann/improved-aesthetic-predictor. 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.721, + 0.908, + 0.788 + ], + "angle": 0, + "content": "[26] Jascha Sohl-Dickstein, Eric Weiss, Niru Maheswaranathan, and Surya Ganguli. Deep unsupervised learning using nonequilibrium thermodynamics. In International Conference on Machine Learning, pages 2256-2265. PMLR, 2015. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.79, + 0.908, + 0.859 + ], + "angle": 0, + "content": "[27] Gowthami Somepalli, Vasu Singla, Micah Goldblum, Jonas Geiping, and Tom Goldstein. Diffusion art or digital forgery? investigating data replication in diffusion models. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 6048-6058, 2023. 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.86, + 0.907, + 0.901 + ], + "angle": 0, + "content": "[28] Jiaming Song, Chenlin Meng, and Stefano Ermon. Denoising diffusion implicit models. International Conference on Learning Representations, 2021. 1" + }, + { + "type": "list", + "bbox": [ + 0.517, + 0.093, + 0.908, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.48, + 0.945, + 0.518, + 0.957 + ], + "angle": 0, + "content": "23311" + } + ], + [ + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.093, + 0.486, + 0.16 + ], + "angle": 0, + "content": "[29] Yang Song, Jascha Sohl-Dickstein, Diederik P Kingma, Abhishek Kumar, Stefano Ermon, and Ben Poole. Score-based generative modeling through stochastic differential equations. International Conference on Learning Representations, 2021. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.164, + 0.483, + 0.232 + ], + "angle": 0, + "content": "[30] Alexander Tong, Nikolay Malkin, Guillaume Huguet, Yanlei Zhang, Jarrid Rector-Brooks, Kilian Fatras, Guy Wolf, and Yoshua Bengio. Improving and generalizing flow-based generative models with minibatch optimal transport. Transactions on Machine Learning Research, 2023. 1, 2, 3, 7, 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.234, + 0.483, + 0.316 + ], + "angle": 0, + "content": "[31] Alexander Y Tong, Nikolay Malkin, Kilian Fatras, Lazar Atanackovic, Yanlei Zhang, Guillaume Huguet, Guy Wolf, and Yoshua Bengio. Simulation-free schrödinger bridges via score and flow matching. In International Conference on Artificial Intelligence and Statistics, pages 1279-1287. PMLR, 2024. 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.318, + 0.483, + 0.372 + ], + "angle": 0, + "content": "[32] Zhisheng Xiao, Karsten Kreis, and Arash Vahdat. Tackling the generative learning trilemma with denoising diffusion gans. International Conference on Learning Representations, 2022. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.375, + 0.483, + 0.415 + ], + "angle": 0, + "content": "[33] Ye Yuan and Kris Kitani. Diverse trajectory forecasting with determinantal point processes. International Conference on Learning Representations, 2020. 3, 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.418, + 0.483, + 0.471 + ], + "angle": 0, + "content": "[34] Qinqing Zheng, Matt Le, Neta Shaul, Yaron Lipman, Aditya Grover, and Ricky TQ Chen. Guided flows for generative modeling and decision making. arXiv preprint arXiv:2311.13443, 2023. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.474, + 0.483, + 0.557 + ], + "angle": 0, + "content": "[35] Yinglin Zheng, Hao Yang, Ting Zhang, Jianmin Bao, Dongdong Chen, Yangyu Huang, Lu Yuan, Dong Chen, Ming Zeng, and Fang Wen. General facial representation learning in a visual-linguistic manner. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 18697–18709, 2022. 2" + }, + { + "type": "list", + "bbox": [ + 0.093, + 0.093, + 0.486, + 0.557 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.48, + 0.946, + 0.519, + 0.957 + ], + "angle": 0, + "content": "23312" + } + ] +] \ No newline at end of file diff --git a/2025/DiverseFlow_ Sample-Efficient Diverse Mode Coverage in Flows/276f7b10-7b9f-439d-a85a-8ad9f72195cf_origin.pdf b/2025/DiverseFlow_ Sample-Efficient Diverse Mode Coverage in Flows/276f7b10-7b9f-439d-a85a-8ad9f72195cf_origin.pdf new file mode 100644 index 0000000000000000000000000000000000000000..8b10b8c7881f58d67b663d000d1b83c061dbca36 --- /dev/null +++ b/2025/DiverseFlow_ Sample-Efficient Diverse Mode Coverage in Flows/276f7b10-7b9f-439d-a85a-8ad9f72195cf_origin.pdf @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:1e3485cb3552973e71f0e8470743291983551a54236f7d4d43db8de2ab47d755 +size 3791367 diff --git a/2025/DiverseFlow_ Sample-Efficient Diverse Mode Coverage in Flows/full.md b/2025/DiverseFlow_ Sample-Efficient Diverse Mode Coverage in Flows/full.md new file mode 100644 index 0000000000000000000000000000000000000000..3cd4cc45c445f2e20159610d6370b7b1859d5949 --- /dev/null +++ b/2025/DiverseFlow_ Sample-Efficient Diverse Mode Coverage in Flows/full.md @@ -0,0 +1,337 @@ +# DiverseFlow: Sample-Efficient Diverse Mode Coverage in Flows + +Mashrur M. Morshed Vishnu Boddeti + +Michigan State University + +{morshedm, vishnu}@msu.edu + +# Abstract + +Many real-world applications of flow-based generative models desire a diverse set of samples that cover multiple modes of the target distribution. However, the predominant approach for obtaining diverse sets is not sample-efficient, as it involves independently obtaining many samples from the source distribution and mapping them through the flow until the desired mode coverage is achieved. As an alternative to repeated sampling, we introduce DiverseFlow: a training-free approach to improve the diversity of flow models. Our key idea is to employ a determinantal point process to induce a coupling between the samples that drives diversity under a fixed sampling budget. In essence, DiverseFlow allows exploration of more variations in a learned flow model with fewer samples. We demonstrate the efficacy of our method for tasks where sample-efficient diversity is desirable, such as text-guided image generation with polysemous words, inverse problems like large-hole inpainting, and class-conditional image synthesis. + +# 1. Introduction + +Consider the task of text-guided image generation from open-ended prompts, like "A famous boxer". If we use a generative ordinary differential equation (ODE) to obtain a few images for this given prompt, we may observe something unusual: all the resultant images depict a dog, as shown in Figure 1 (top-left). Such a result is plausible, as the word "boxer" has multiple meanings: it can either mean a combat sport athlete or a particular dog breed. However, for the prompt "A famous boxer", what if the output desired by the user was actually a human athlete, and not a dog? This situation necessitates obtaining additional samples from the model, until the desired alternate meanings are discovered. But instead of repeated sampling, can we directly observe more meanings by finding a more diverse set? + +Beyond the aforementioned example of text-to-image generation from polysemous1 prompts, sample diversity is a + +![](images/2c6c83f27981f75e8aed186ad9578fdd73f6e56a46f9aedbd43cad7a9e80cbc1.jpg) +Prompt: "A famous boxer" + +![](images/10c363bc192d79a1cc897b1f19467d637245fed58d2a78e9b6a278eab156cfb0.jpg) + +![](images/3a64acd2ec62494943ef07762881d650a82afc857674e2990bd37cf3542485b1.jpg) +Prompt: "A letter" +ID +Figure 1. Text-guided image generation with polysemous words. With the same budget of samples, DiverseFlow (right) finds a more diverse set of results compared to IID sampling (left). + +![](images/c156a42342a6e920569d9a0c73cb438d810e79299180ef4b8e2d5515044e4ff2.jpg) +DiverseFlow + +desirable objective for many other tasks that use generative models. These include inverse problems (e.g., large hole filling) and class-conditioned image generation, to name a few. Diversity or mode coverage is a key pillar in the generative learning trilemma [32], in addition to fidelity and latency. For state-of-the-art generative methods such as flow matching models (FM) [20, 22] and diffusion models (DM) [12, 26], significant work has been done on improving the photorealism of samples and the efficiency of the sampling process [11, 16, 20, 28, 30, 34]. However, relatively little attention has been paid to explicitly enhancing the diversity of generated samples under a limited sampling budget. + +The standard approach to generate a diverse set of images is to repeatedly obtain independent and identically distributed (IID) samples from a simple or tractable source distribution (e.g., Gaussian distribution), map them to samples + +in the target distribution, and continue this process until we observe sufficient mode coverage in the target distribution. This process, while effective, is sample-inefficient, requiring the generation of more results than necessary. Importantly, the mapping from the source to the target density does not hold a linear relationship; even specifically selecting diverse samples from the source distribution by design does not necessarily yield diverse samples in the target distribution. These limitations naturally raise the following research question. + +How can we generate diverse samples from the target density under a limited sampling budget? + +In this paper, we propose DiverseFlow, an inference-time, training-free approach to obtain a diverse set of samples in a desired target density under a fixed sampling budget. We focus on deterministic ODE sampling in continuous-time generative models, specifically FMs, an emerging generative paradigm that enables simulation-free training of continuous normalizing flows (CNFs) and includes diffusion models as a special case. + +DiverseFlow measures the diversity of a set of samples through the volume they span in the target space. A set of similar samples span a lower volume, while a diverse set naturally spans a larger volume. We impose a volume-based gradient constraint on the flow ODE by drawing on determinantal point processes (DPP) [17, 23], a probabilistic model arising from quantum physics that exactly describes the Pauli exclusion principle: that no two fermions may occupy the same quantum state. In Figure 1 consider the images generated via IID sampling (left column) from a text-to-image generative ODE—for both the prompts "A famous boxer" and "A letter", only a single potential meaning is discovered. Unlike the results generated via IID sampling, samples obtained from DiverseFlow (right column) span more diverse modes corresponding to the polysemous words in the prompts. For "A famous boxer", the samples show both a dog breed and a human athlete; for "A letter", we observe both written correspondence and alphabet symbols. + +We empirically demonstrate the utility of DiverseFlow across several applications where diversity is inherently desirable. First, we use DiverseFlow to perform text-guided image synthesis for words and phrases that may carry a variety of meanings. Second, we perform large-hole face inpainting with occlusion masks covering significant regions of the face that may be important to the person's identity. Third, we apply DiverseFlow on class-conditioned image synthesis and demonstrate that we can more efficiently explore the data space compared to IID sampling. Lastly, to better characterize and explain the behavior of DiverseFlow, we perform several experiments on synthetic 2D densities. + +# Summary of Contributions + +1. We present a sample-efficient method to obtain a diverse set of samples from a flow ODE in Section 5 and demonstrate it qualitatively in Sections 6.1 to 6.3 and quantitatively in Sections 6.4 and 6.5 +2. We provide an empirical analysis that demonstrates that our method is consistent across various flow matching formulations (Section 6.4) +3. We introduce the task of text-to-image synthesis from prompts with polysemous words in the context of analyzing sample diversity, and provide a comparison with relevant methods (Section 6.5) + +# 2. Preliminaries + +# 2.1. Flow Matching + +Many generative models can be considered as a transport map from some easy-to-sample source distribution to an empirically observed yet unknown target distribution. Recent successes in generative modeling represent this transport map in the form of continuous-time processes, such as stochastic differential equations (SDEs) [12, 29], or ordinary differential equations (ODEs) [2, 20, 22]. Although diffusion models are formulated as SDEs, a significant body of research focuses on converting the diffusion SDE to a deterministic ODE at inference time for faster inference. The diffusion ODE, or probability flow ODE, is a particular case of continuous normalizing flows (CNFs). Flow Matching (FM) [2, 20, 22] is motivated by the idea of directly training CNFs in a scalable and simulation-free manner, just like diffusion models. Moreover, many recent text-to-image generative models, such as Stable Diffusion 3 [10], adopt the FM framework. As such, we present our approach primarily in the context of FM, and our findings can be extended to diffusion and score-based generative models in a straightforward manner. + +A CNF reshapes a prior source density $p_0$ to the empirically observed target density $p_1$ with an ODE of the form: + +$$ +d \mathbf {x} _ {t} = v _ {\theta} \left(\mathbf {x} _ {t}, t\right) d t, \quad \mathbf {x} _ {0} \sim p _ {0} \tag {1} +$$ + +where $v_{\theta}$ is a time-dependent velocity field whose parameters $\theta$ are learned; we interchangeably use the notation $v_{t}$ to imply $v_{\theta}(\cdot, t)$ . It becomes possible to obtain samples from $p_{1}$ by integrating Equation (1) over time, i.e. by starting at $\mathbf{x}_0 \sim p_0$ for $t = 0$ and solving the ODE till $t = 1$ . As our approach is training-free, we do not elaborate on the details of learning to regress the vector field $v_{t}$ ; we encourage interested readers to refer to the works of Lipman et al. [20] and Tong et al. [30] for a primer on training FMs. + +At any timestep $t$ during sampling, an intermediate sample $\mathbf{x}_t$ in the flow trajectory can be used to obtain an approx- + +imation of the target as follows: + +$$ +\hat {\mathbf {x}} _ {1} = \mathbf {x} _ {t} + v _ {\theta} (\mathbf {x} _ {t}, t) (1 - t) \tag {2} +$$ + +Equation (2) is equivalent to simply taking a large Euler step at any time instance $t$ and is naturally more accurate as $t$ approaches $t = 1$ . Further, Equation (2) is also well suited for ODEs with 'straight' paths, where the direction of the time-varying velocity $v_{t}$ remains near-constant in time (such as the work of Liu et al. [22]). Similarly, we can estimate the source sample by simply taking a step in the reverse direction: + +$$ +\hat {\mathbf {x}} _ {0} = \mathbf {x} _ {t} - v _ {\theta} (\mathbf {x} _ {t}, t) t \tag {3} +$$ + +# 2.2. Determinantal Point Processes + +Determinantal point processes (DPPs) [4, 17, 23] are probabilistic models of repulsion between points. They were originally termed as fermion processes as they describe the Pauli exclusion principle or antibunching effect in fermions. To define a DPP, we must first consider a set of points, $\mathcal{V}$ , and a point process $\mathcal{P}(\mathcal{V})$ — a probability measure on $2^{\mathcal{V}}$ (the set of all possible subsets of $\mathcal{V}$ ). $\mathcal{P}$ is determinantal when the probability of choosing a random subset $Y \subset \mathcal{V}$ according to $\mathcal{P}$ is given by: + +$$ +\mathcal {P} (Y \subset \mathcal {Y}) = \frac {\det (\mathbf {L} _ {Y})}{\sum_ {Y \subset \mathcal {Y}} \det (\mathbf {L} _ {Y})} = \frac {\det (\mathbf {L} _ {Y})}{\det (\mathbf {L} + \mathbf {I})} \tag {4} +$$ + +where $\mathbf{L} \in \mathbb{R}^{|\mathcal{Y}| \times |\mathcal{Y}|}$ is a kernel matrix, and $\mathbf{L}_Y$ is the sub-kernel matrix indexed by the elements of $Y$ . Equation (4) has an intuitive geometric interpretation if we consider the kernel $\mathbf{L}$ to be constructed from cosine similarity: the determinant of $\mathbf{L}_Y$ is the Gram-determinant, describing the squared volume of the $N$ -dimensional parallelotope spanned by the set of vectors $Y$ . Thus, a DPP naturally assigns higher probabilities to more orthogonal (and thus diverse) subsets that span larger volumes. We leverage DPPs to define a coupled likelihood measure over a set of samples in a flow trajectory. + +# 3. Related Work + +Efficiently finding diverse sets is useful in several application areas of machine learning. For instance, Batra et al. [3] show that the M-Best MAP (maximum a posteriori) solutions in Markov random fields are often distant from the ground truth and highly similar. They thus propose the *Diverse M-Best* problem—finding a set of $M$ highly probable solutions satisfying some minimum dissimilarity threshold—that partly inspires our study in Section 4. [33] utilize DPPs in conjunction with variational autoencoders (VAE) for diverse trajectory forecasting; a set of diverse future pedestrian trajectories improves safety-critical perception systems in autonomous vehicles. Motivated by potential drug discovery + +and material design applications, Jain et al. [14] propose finding diverse Pareto-optimal candidates in a multi-objective setting with generative flow networks. + +The work by Corso et al. [6] which explores diverse non-IID sampling for diffusion models is most similar in spirit to DiverseFlow. However, DiverseFlow is notably different in the following aspects: (1) Our diversity objective is derived from determinantal point processes, a diversity-promoting probability measure of the joint occurrence of a set of samples. Corso et al. [6] is instead inspired by stein variational gradient descent (SVGD) [21]. 2) The diversity measure in DiverseFlow (volume, or determinant of similarity kernel) assigns a zero likelihood to a set if any duplicate elements are present; presence of duplicates is tolerated in the diversity measure in Particle Guidance (row-wise sum of similarity kernel) + +# 4. Diverse Source Samples Do Not Yield Diverse Target Samples + +![](images/46a6f278801b399e96b2a0a6be298519d95dc071360f264fa4b47cfe9688b918.jpg) + +![](images/df991f020c6c783fd783210b68a9a57b3e516fefd9044bd263eb9d81e3870a24.jpg) + +![](images/ac6db7105b857c92bd53ac5877c85e725cae59c7094d3570267fb3798f2425cc.jpg) +(a) Conditional ing (CFM) [20] +(c) DiverseFlow on CFM +Figure 2. Finding $K = 5$ samples from the target distribution with $N = 10$ modes. In the example, only 3 modes are found by 5 IID samples (a, b). DiverseFlow discovers 5 modes with the same sampling budget in (d, e). + +![](images/986aeb4001abbd1298dc23aba33fe1e40a26b7eb3eae94cf5c33f43051bd5d50.jpg) +(b) Mini-batch Optimal Transport (MB-OT) [30] +(d) DiverseFlow on MB-OT + +Problem Setting: We start with a synthetic example to illustrate our problem of interest. Consider that we have empirical observations from a target distribution $\pi_1 \in \mathbb{R}^2$ , which is a random mixture of Gaussians, such as the example shown in Figure 4. We design $\pi_1 = \sum_{i=1}^{N} w_i \mathcal{N}(\pmb{\mu}_i, \sigma_i^2 \pmb{I})$ to contain $N = 10$ randomly selected modes $\mathcal{N}(\pmb{\mu}_i, \sigma_i^2 \pmb{I})$ , each with a random mixture weight $w_i$ ; we observe that in our example, there are 6 high probability modes and 4 low probability ones. Suppose we have a sampling budget + +of $K$ samples. This leads to three possible scenarios: (i) $K < N$ , (ii) $K = N$ , and (iii) $K > N$ . Among the aforementioned, case (i) (fewer samples than modes) is the most likely characteristic of any real-world dataset. + +Let us have a prior distribution $\pi_0$ and some generative model $\Psi$ , such that, in the limit of infinite samples, $\Psi(x_0 \sim \pi_0) \sim \pi_1$ . Then, the objective of sample-efficient diverse sampling is to obtain samples from $\min(K, N)$ modes from $\pi_1$ , given a fixed set of $K$ samples in $\pi_0$ . + +If diverse samples are desired from the target density of the flow, one may make the elementary assumption that if the particles are distant at the source distribution, after being transported by the flow, they remain distant in the target distribution. This assumption is not necessarily true, as we show in Figure 2. By design, we choose a uniform mixture of eight Gaussians as the source $\pi_0$ to obtain diverse source samples. In Figure 2a, we can observe that source points from distinct modes can still converge to the same target mode with IID sampling. Thus, an alternative procedure is necessary to obtain a diverse set from a flow in a sample-efficient manner. We further explore this toy problem in Section 6.4. + +# 5. Diverse Sampling from Flows + +From Figure 2, we observe that independently (or heuristically) chosen source samples may not map to a diverse set of target samples. In this case, we can select a new set of source samples and repeat the sampling process till eventually covering at least $K$ modes. However, this approach does not satisfy our fixed sampling budget constraint. An alternative solution to repeated independent sampling is defining and leveraging a diversity measure of the target samples to drive sample diversity. For the set of source samples $\{\mathbf{x}_0^{(1)},\mathbf{x}_0^{(2)},\ldots ,\mathbf{x}_0^{(k)}\}$ , we could optimize a set of perturbations $\{\delta^{(1)},\delta^{(2)},\dots,\delta^{(k)}\}$ such that the new set $\{\mathbf{x}_0^{(1)} + \delta^{(1)},\mathbf{x}_0^{(2)} + \delta^{(2)},\dots,\mathbf{x}_0^{(k)} + \delta^{(k)}\}$ maps to a diverse set of target particles. However, this approach would require multiple simulations of the whole ODE and backpropagating over all the timesteps, which increases the computational complexity of the sampling process over the standard IID sampling. + +This leads us to our proposed approach: we avoid multiple simulations and instead optimize the flow trajectory for diversity while solving the ODE. For any sample in the flow trajectory $\mathbf{x}_t$ , we have an estimate of the target sample $\hat{\mathbf{x}}_1$ through Equation (2). Suppose we have a differentiable objective $\mathcal{L}(\{\hat{\mathbf{x}}_1^{(1)},\hat{\mathbf{x}}_1^{(2)},\dots ,\hat{\mathbf{x}}_1^{(k)}\})$ that assigns a likelihood to the joint outcome $\{\hat{\mathbf{x}}_1^{(1)},\hat{\mathbf{x}}_1^{(2)},\dots ,\hat{\mathbf{x}}_1^{(k)}\}$ . Further, let $\mathcal{L}$ assign a higher likelihood if the joint outcome is a diverse set, and a diminished likelihood if the set is similar. We can then leverage $\mathcal{L}$ to drive diversity among the target samples + +by modifying the flow velocity of the $i$ -th particle as, + +$$ +\tilde {v} _ {t} ^ {(i)} = v _ {t} ^ {(i)} - \gamma (t) \nabla_ {\mathbf {x} _ {t} ^ {(i)}} \log \mathcal {L} (\{\hat {\mathbf {x}} _ {1} ^ {(1)}, \hat {\mathbf {x}} _ {1} ^ {(2)}, \dots , \hat {\mathbf {x}} _ {1} ^ {(k)} \}) \tag {5} +$$ + +where $\gamma (t)$ is a time-varying scale that controls the strength of the diversity gradient. Setting $\gamma (t) = 0$ reduces to the standard IID sampling scenario, while $\gamma (t) > 0$ will encourage diversity between the generated samples. In practice, $\gamma (t)$ follows the schedule of the probability path normalized by the norm of the DPP gradient. + +# 5.1. Determinantal Gradient Constraints + +We desire objective $\mathcal{L}$ in Equation (5) to be higher if the items in the set are diverse and lower if they are similar to each other. We interpret diversity in terms of the volume spanned by the set. Consider that we have $k$ samples in $\mathbb{R}^d$ (assume $k < d$ ). An objective that prefers diversity can be defined as the volume of the $k$ -dimensional parallelotope in $\mathbb{R}^d$ spanned by the sample vectors; this volume becomes diminished when there are similar samples (and even zero, for identical samples). The determinant describes volumes well; a diverse set must span a large volume in the sample space and have a corresponding large determinant. + +To define a measure over a set of samples, we draw on the idea of determinantal point processes (DPP). We first define a kernel $\mathbf{L}(\{\hat{\mathbf{x}}_1^{(1)},\hat{\mathbf{x}}_1^{(2)},\dots ,\hat{\mathbf{x}}_1^{(k)}\})$ as follows: + +$$ +\mathbf {L} ^ {(i j)} = \exp \left(- h \frac {\| \hat {\mathbf {x}} _ {1} ^ {(i)} - \hat {\mathbf {x}} _ {1} ^ {(j)} \| _ {2} ^ {2}}{\operatorname* {m e d} (\mathbf {U} (\mathbf {D}))}\right) \tag {6} +$$ + +where $\mathbf{D}$ denotes a distance matrix with $\mathbf{D}_{ij} = \| \mathbf{x}^{(i)} - \mathbf{x}^{(j)}\| _2^2$ , $\mathbf{U}(\mathbf{D})$ denotes the upper triangle entries of $\mathbf{D}$ , $h$ denotes a kernel spread parameter, and $\mathrm{med}(\mathbf{U}(\mathbf{D}))$ denotes the median of those entries. Given $\mathbf{L}$ , we may define a DPP-based likelihood as: + +$$ +\begin{array}{l} \mathcal {L} (\{\hat {\mathbf {x}} _ {1} ^ {(1)}, \hat {\mathbf {x}} _ {1} ^ {(2)}, \dots , \hat {\mathbf {x}} _ {1} ^ {(k)} \}) = \frac {\det (\mathbf {L})}{\det (\mathbf {L} + \mathbf {I})} \\ = \prod_ {a = 1} ^ {k} \frac {\lambda (\mathbf {L}) _ {a}}{1 + \lambda (\mathbf {L}) _ {a}} \tag {7} \\ \end{array} +$$ + +where $\lambda (\mathbf{L})_a$ is the $a^{\mathrm{th}}$ eigenvalue of the kernel $\mathbf{L}$ . The log-likelihood is then, + +$$ +\mathcal {L} \mathcal {L} = \log \mathcal {L} = \log \det (\mathbf {L}) - \log \det (\mathbf {L} + \mathbf {I}) \tag {8} +$$ + +Note that the Euclidean distance $\| \hat{\mathbf{x}}_1^{(i)} - \hat{\mathbf{x}}_1^{(j)}\| _2^2$ is not very meaningful in the high-dimensional raw image space [1]. Therefore, in practice, the distance should be computed in a robust feature space, i.e., $\| F(\hat{\mathbf{x}}_1^{(i)}) - F(\hat{\mathbf{x}}_1^{(j)})\| _2^2$ , where $F$ is some domain-specific feature extractor, such as the vision transformer (ViT) [9] for images. + +Quality Constraint: The DPP defined in Equation (7) acts as a repulsive force that is unaware of sample quality. A quality term can be incorporated into the DPP kernel to regularize the trajectory diversification. Although flows can be defined between any arbitrary two distributions, let us consider the special case when the source is a Gaussian, i.e., $p_0 \sim \mathcal{N}(0,\mathbf{I})$ . Suppose we have a quality vector $\mathbf{q}_t = \{q^{(1)}(t), q^{(2)}(t), \ldots, q^{(k)}(t)\}$ , where any $q^{(i)}(t) \in [0,1]$ . We can then define a new kernel $\mathbf{L}_q = \mathbf{L} \odot \mathbf{q}_t \mathbf{q}_t^T$ , where each $q^{(i)}(t)$ penalizes a sample $\mathbf{x}_t^{(i)}$ if it deviates too much from the flow. To define this, we obtain an estimate of the source sample $\hat{\mathbf{x}}_0^{(i)}(t)$ for any given sample $\mathbf{x}_t^{(i)}$ via Equation (3), and check if it lies within a desired percentile-radius $\rho$ of the Gaussian $p_0$ . Specifically, we define the time-dependent sample quality as + +$$ +q ^ {(i)} (t) = \left\{ \begin{array}{l l} 1 & \text {i f} \| \hat {\mathbf {x}} _ {0} ^ {(i)} (t) \| _ {2} ^ {2} \leq \rho^ {2} \\ \max \left(\epsilon , e ^ {- \left(\| \hat {\mathbf {x}} _ {0} ^ {(i)} (t) \| _ {2} ^ {2} - \rho^ {2}\right)}\right) & \text {o t h e r w i s e} \end{array} \right. \tag {9} +$$ + +where $\epsilon$ is a 'minimum quality' we assign to prevent a zero determinant. + +# 5.2. Coupled Ordinary Differential Equations + +At any timestep $t$ , the measure of diversity in Equation (8) can be adopted to modify the flow of the $i$ -th particle. We compute the gradient of the samples with respect to the diversity measure and use it to modify the ODE as follows: + +$$ +\left. d \mathbf {x} _ {t} ^ {(i)} = \left[ v _ {\theta} \left(\mathbf {x} _ {t} ^ {(i)}, t\right) - \gamma (t) \nabla_ {\mathbf {x} _ {t} ^ {(i)}} \log \mathcal {L} \left(\left\{\hat {\mathbf {x}} _ {1} ^ {(1)}, \dots , \hat {\mathbf {x}} _ {1} ^ {(k)} \right\}\right) \right] d t \right. \tag {10} +$$ + +Where $\gamma (t)$ is a time-varying scaling factor. Unlike the IID sampling scenario where we have $K$ independent ODEs, Equation (10) corresponds to a system of coupled non-linear ordinary differential ordinary equations. To see this, first note that the estimate $\hat{\mathbf{x}}_1^{(i)}$ depends on the current sample $\mathbf{x}_t^{(i)}$ i.e., $\hat{\mathbf{x}}_1^{(i)} = \mathbf{x}_t^{(i)} + v_\theta (\mathbf{x}_t^{(i)},t)(1 - t)$ . Second, the DPP log-likelihood $\mathcal{LL}(\{\hat{\mathbf{x}}_1^{(1)},\hat{\mathbf{x}}_1^{(2)},\dots ,\hat{\mathbf{x}}_1^{(k)}\})$ induces a time-dependent coupling between the $K$ trajectories of $\mathbf{x}_t^{(i)},i = 1,\ldots ,K$ and seeks to diversify the target samples. Although higher-order ODE solvers [16] can be employed to solve the coupled ODEs, we use the standard Euler method. + +# 6. Experiments + +We demonstrate the utility of DiverseFlow in flow-based generative models by considering three applications where sample diversity is naturally desirable: text-guided image generation with polysemous words, large-hole inpainting, and class-conditional image generation. We also analyze the effect of DiverseFlow on different flow matching formulations w.r.t. its ability to span diverse modes through a synthetically constructed 2D density example. + +![](images/e4345bb398268dcbae533bc9541668f565a02ab355ea7220071bf32e4d02d1b0.jpg) +(a) "A buck" + +![](images/9e9afdf4fc87a74b527a28a5bc998905bb61a5ff0e5622e3a90519c303a4110f.jpg) + +![](images/bda090aeb79979dc0ae0d48a0cf8643cdf03d5452d7a8e956d066e809178dc1e.jpg) + +![](images/a2442f078401e273bd017cda6dd894ae35299eb39eafccf7b61b850cc6c6ed16.jpg) + +![](images/698e897027df0fac5fb324f1154f46536ad5a2654aeaffb8b2d5a96c3c806c28.jpg) +(b) "A famous boxer" +Figure 3. For each prompt, the left image (red box) denotes standard IID sampling with classifier-free guidance, while the right image (blue box) shows the result after incorporating DiverseFlow. DiverseFlow finds more diverse sets given the same source points—clearly distinguishable as new semantic meanings in the case of prompts with multiple meanings. + +![](images/ec6497676581e9b0f00046f6ac8be6c6d8b3bf00a6134c0b7fd1352fa75a9896.jpg) +(c) "Van Gogh painting" + +# 6.1. Image Generation from Polysemous Prompts + +In text-to-image generation, the conditional data distribution corresponding to a text prompt may contain many variations, and it is a desirable objective to generate images that span those variations in a sample-efficient manner. As we have previously shown in Figure 1, the result desired by a user (a human boxer athlete) may not be the result generated by the model (a dog breed); if diverse results are obtained, it allows a user to obtain the desired results with fewer attempts. + +We pose a scenario where diverse sets are easily observable: when an open-ended text prompt is polysemous and carries multiple meanings, such as the examples we show in Figure 1 and Figure 3. In Figure 3(a), the prompt "A buck" may commonly refer to a male deer. However, it may also informally refer to a United States dollar. Using the + +same four source points, which are deterministically mapped to four deer images by IID sampling. DiverseFlow finds a different set of samples—one that includes a dollar-like coin, albeit embossed with a deer head. We also observe minor differences between the two sets of images, such as changes in pose and background in the top-right and bottom-right deers. For Figure 3(c), while 'Van Gogh painting' is not quite a polysemous word, it can still have two interpretations: a painting painted by Van Gogh, or a painting of Van Gogh. The regular samples contain minimal diversity, as they include two sets of repeated paintings of Van Gogh. With DiverseFlow, not only can we get a set of four distinct paintings, but we also have a portrait of Van Gogh, which is one of the additional meanings of the prompt. We present additional qualitative samples in the supplementary material. + +# 6.2. Diverse Inpainting on Faces + +Another inverse problem where diverse solutions are desirable is face inpainting, where we seek to inpaint the missing parts of the face with diverse plausible facial textures and structures. To demonstrate inpainting with FM models, we first incorporate Manifold Constrained Gradient (MCG) [5] in an off-the-shelf unconditional Rectified-Flow model. In addition to the manifold constraints, we employ determinantal gradient constraints to enhance diversity (further detailed in Algorithm 1). In Figure 5 (b), we observe that the inpainted faces of the four women have similar expressions (largely neutral). DiverseFlow improves the diversity of the set by yielding a highly different expression in the top-right image. In (d) and (e), we also observe changes in facial hair and expressions due to diversification. + +![](images/cc6df4e4178aa44caacbc8ad8bc3b772697f71015d64a3422294bba61425c9db.jpg) + +![](images/55bc3c8aee37e194ee6eda46aaab6738322b8aec946c8e8995d1f8780cbf192d.jpg) + +![](images/d175f3a6cae4e06745f9435defd5c1fd29139821dd91cbc6930017f91f00e08c.jpg) + +![](images/9095c00848783c151ea1264db17b17e0909c02afb63eab4617b7b78faf43c5ca.jpg) +(a) +Figure 5. Inpainting on CelebAHQ-256 × 256. (a) Dashed boxes show masked input (top) and ground truth (bottom) respectively. (b) Rectified-Flow [22] + MCG [5] (c) With DiverseFlow. + +![](images/38338b8f2f901e450b3e33471eaff73b3d470a1848893290f142d5a556cac363.jpg) +(b) + +![](images/cab056f61389e95bd31f738823295add41b9a4c514fd136d339522eb02c8c685.jpg) +(c) + +![](images/5c392b70064bc2f773bfde4fa5af59b45346919daa72c830dcba39a1e33b12f6.jpg) + +![](images/61ece2b7ba393195a2b1c3504086b09eed3f8aba7ba95ff3623b6618a75f53df.jpg) +No CFG + +![](images/f9e2b90427a740b7b6e70ed79ea9601fda8929a5de19ce6aecf156a0e4023231.jpg) + +![](images/6270b52bbe42c86627c415e9e972ad66c929c2859609a3f7ae8c27e28352abf3.jpg) +With CFG +Figure 6. Class-conditional ImageNet samples from LFM [7]. We show samples for two classes, (top row) 'Mushroom' (class 947) and (bottom row) 'Macaw' (class 88). (a, d) No CFG. (b, e) LFM with CFG. (d, f) LFM with CFG and DiverseFlow. + +![](images/7567bfb3dcd0871aac8f8dcc319182c20bf53c26850f9b6a7c0fd94e9ce86865.jpg) + +![](images/b5e3b8c6e5a44aaf48e853449ff28850c8e09792fbfb8c6dfe871e184b337fdd.jpg) +DiverseFlow + +# 6.3. Diverse Class-Conditional Image Synthesis + +Suppose we can access a class-conditioned flow matching (FM) model trained on an unknown image dataset. To explore the unobservable true dataset, we may use a set of class-conditional samples from the FM model. We adopt a latent flow matching (LFM) model [7], trained to generate $256 \times 256$ resolution images from the ImageNet [8] dataset. Much like latent diffusion, LFM employs classifier-free guidance to create high-quality samples. However, this naturally poses a cost to diversity, as we show in Figure 6. + +By incorporating DiverseFlow, we can maintain the high quality of the samples and simultaneously explore more modes in the dataset. In Figure 6, we demonstrate two ImageNet classes that may have diversity: 'Mushroom' and 'Macaw.' For mushrooms, we observe that LFM primarily generates two species of mushrooms. However, by applying DiverseFlow, we successfully find a new species within our limited set: an Amanita muscaria, also known as the fly agaric—easily distinguishable by the white spots on its red cap. In another example, we see that while LFM generates four scarlet macaws, using the same source samples, DiverseFlow helps us find a different blue and yellow macaw. We further evaluate the precision and recall [18] for ImageNet-256 synthesis in Figure 7, showing that DiverseFlow improves the recall while retaining similar precision, suggesting improved diversity. + +# 6.4. DiverseFlow Across Various FM Formulations + +We ask the question: does DiverseFlow remain consistent across different formulations of flow matching? We adopt the same toy bivariate mixture of Gaussian distributions as in the example shown in Figure 2, across four different FM formulations: (i) Conditional Flow Matching (CFM) [20], + +![](images/fef732875e8d72d562ee357e5aad903d9199914893ad33400c074e21c5ea3492.jpg) +Figure 4. Comparing different FM formulations in terms of the number of modes spanned by IID sampling versus with DiverseFlow. More details about the experiment are provided in the supplementary. + +![](images/0bf640e7bba7bb1cd27b421928f5d85dbb886e09c20e0128f8be1202571b0516.jpg) +Figure 7. Precision vs. recall on ImageNet-256. + +
MethodCFGP (↑)R (↑)
LFM10.620.52
LFM1.50.690.44
LFM20.770.41
LFM40.690.26
LFM + DF10.620.52
LFM+DF1.50.690.47
LFM+DF20.760.46
LFM+DF40.700.38
+ +Table 1. Precision (P) and recall (R) score of DiverseFlow and baselines. + +(ii) Mini-batch Optimal Transport CFM (MB-OT) [30], (iii) Schrödinger Bridge CFM (SB-CFM) [31], and (iv) Stochastic Interpolants (SI-CFM) [2]. We then perform a numerical experiment to quantify the average number of modes discovered by each FM variant with increasing the sampling budget $K$ . Figure 4 reports the results. For a maximum sampling budget of 10, MB-OT discovers only 5.64 modes on average, which is expected since the dataset contains 6 high-probability modes. By incorporating DiverseFlow, we can find 7.11 modes on average. We observe that MB-OT and SB-CFM benefit most from DiverseFlow. We hypothesize this is because MB-OT and SB-CFM are formulated with the notion of optimal paths, which results in a more accurate estimation of $\hat{x}_1$ and thus a better diversity gradient. + +# 6.5. Comparison To Particle Guidance + +(i) Preventing Training Data Copy: Previously, Corso et al. [6] demonstrate that their method can alleviate Stable Diffusion's training data regurgitation problem [27] to some extent. In Figure 8, we demonstrate similar capabilities; in (b), the top left and bottom right examples are copies of the training data. Subsequently, (c) and (d) find a new example for the top-left sample. (ii) Diversity vs Quality: We also compare the diversity versus quality of DiverseFlow against Particle Guidance in Figure 10 over 30 polysemous prompts repeated over 10 random seeds. While varying the + +![](images/a5ba5bf8f7fb9dd0635fa4dbf85f6f362ac9ea799e39bdb9d860577664fe5dd6.jpg) +(a) Training Data + +![](images/ce25bfa9b4e5ef7fa09c122f1c27b16773f3114a201b8e04e8527e4f2359ce14.jpg) + +![](images/d7863ca0514b5e16bb6e714f257336c5bbe02efcacc86a79b500fa53919ddfe6.jpg) +(c) Particle Guidance [6] +Figure 8. For the prompt "VAN GOGH CAFE TERASSE copy.jpg", we show (a) the training data, (b) IID sampling with two copies in top-left and bottom-right, (c) Particle Guidance [6], and (d) DiverseFlow. + +![](images/5ef79318d3c8a81eb0c24b6848c9a23249e3ad3a237b38e3e86406a4c31fcf20.jpg) +(b) IID Sampling +(d) DiverseFlow + +strength of DiverseFlow to improve diversity, we also vary the values of the classifier-free guidance strength from 7.5 to 10 to boost quality. Quality is measured by Aesthetic Score (higher is better) [25], and diversity is measured by average pairwise similarity of a set (lower is better) [6]. We observe that though DiverseFlow obtains better diversity at similar quality to IID sampling, the aesthetic score of Particle Guidance is unusually low. In Figure 9 we highlight the cause: for the challenging task of polysemous text-to-image generation, the images generated from Particle Guidance (a) suffers + +![](images/65c4181a6ba7953eba996ac49880980cf74dca076240aad473bbdd4382ae2d76.jpg) +Figure 9. Particle Guidance (middle rows) can suffer from artefacts for polysemous prompts. DiverseFlow retains quality while achieving diversity: for "A mint", a minted coin with a mint leaf was discovered. + +![](images/185e2a6bc82bf7990294139d87b3f16d96b9684fcd145794d903258555a2a688.jpg) +Figure 10. In-batch similarity $(\downarrow)$ vs aesthetic score $(\uparrow)$ + +from image artefacts, which achieves a high diversity score but a poor quality score. Additional details about the experiment, including the set of 30 prompts, are provided in the + +![](images/57d72ea72048db47ef67e9e96c7c2a74edc1aa9a826a4ec5e60fe0629c570eb7.jpg) +Figure 11. Average modes discovered over 100 trials, for varying noise scales and strengths of DiverseFlow and ParticleGuidance. + +supplementary. (iii) Mode Discovery: In high dimensional data, the number of modes, $N$ , is significantly greater than the budget $K$ , so finding unique modes is still highly probable. To better highlight the differences between DiverseFlow and Particle Guidance, we adopt a simpler toy experiment: finding modes in a symmetric uniform Mixture of Gaussian distribution. Corso et al. [6] provides the result that IID sampling with budget $K = 10$ discovers about 6.5 modes on average, while Particle Guidance with a Euclidean kernel discovers almost 9 out of 10 modes. We verify this result in Figure 11, finding that Particle Guidance discovers up to 8.8 modes (averaged over 100 trials). However, by using DiverseFlow, it is possible to discover all 10 modes on average, showing that our approach has a stronger diversification effect. + +# 7. Conclusion + +In this paper, we present DiverseFlow, a way to enforce diversity among a limited set of samples generated from a flow by coupling them through a determinantal point process. We demonstrate multiple applications where our method can be useful. In essence, though flow models admit a deterministic mapping from source to sample, DiverseFlow allows us to modify this mapping at inference-time and obtain samples with improved diversity. Diversity enhances the utility of the underlying generative flow in the case of ill-posed problems, where many solutions exist and it is desirable to find a multitude of such solutions. + +# References + +[1] Charu C Aggarwal, Alexander Hinneburg, and Daniel A Keim. On the surprising behavior of distance metrics in high dimensional space. In Database Theory—ICDT 2001: 8th International Conference London, UK, January 4–6, 2001 Proceedings 8, pages 420–434. Springer, 2001. 4 +[2] Michael S Albergo, Nicholas M Boffi, and Eric Vanden-Eijnden. Stochastic interpolants: A unifying framework for flows and diffusions. arXiv preprint arXiv:2303.08797, 2023. 2, 7, 4 +[3] Dhruv Batra, Payman Yadollahpour, Abner Guzman-Rivera, and Gregory Shakhnarovich. Diverse m-best solutions in markov random fields. In Computer Vision-ECCV 2012: 12th European Conference on Computer Vision, Florence, Italy, October 7-13, 2012, Proceedings, Part V 12, pages 1-16. Springer, 2012. 3 +[4] Alexei Borodin and Grigori Olshanski. Distributions on partitions, point processes, and the hypergeometric kernel. Communications in Mathematical Physics, 211:335-358, 2000. 3 +[5] Hyungjin Chung, Byeongsu Sim, Dohoon Ryu, and Jong Chul Ye. Improving diffusion models for inverse problems using manifold constraints. Advances in Neural Information Processing Systems, 35:25683-25696, 2022. 6, 1 +[6] Gabriele Corso, Yilun Xu, Valentin De Bortoli, Regina Barzilay, and Tommi Jaakkola. Particle guidance: non-iid diverse sampling with diffusion models. International Conference on Learning Representations, 2023. 3, 7, 8, 1 +[7] Quan Dao, Hao Phung, Binh Nguyen, and Anh Tran. Flow matching in latent space. arXiv preprint arXiv:2307.08698, 2023. 6, 3 +[8] Jia Deng, Wei Dong, Richard Socher, Li-Jia Li, Kai Li, and Li Fei-Fei. Imagenet: A large-scale hierarchical image database. In 2009 IEEE conference on computer vision and pattern recognition, pages 248-255. IEEE, 2009. 6 +[9] Alexey Dosovitskiy, Lucas Beyer, Alexander Kolesnikov, Dirk Weissenborn, Xiaohua Zhai, Thomas Unterthiner, Mostafa Dehghani, Matthias Minderer, Georg Heigold, Sylvain Gelly, et al. An image is worth 16x16 words: Transformers for image recognition at scale. International Conference on Learning Representations, 2021. 4 +[10] Patrick Esser, Sumith Kulal, Andreas Blattmann, Rahim Entezari, Jonas Müller, Harry Saini, Yam Levi, Dominik Lorenz, Axel Sauer, Frederic Boesel, et al. Scaling rectified flow transformers for high-resolution image synthesis. In International Conference on Machine Learning, pages 12606-12633. PMLR, 2024. 2, 1 +[11] Jonathan Ho and Tim Salimans. Classifier-free diffusion guidance. arXiv preprint arXiv:2207.12598, 2022. 1 +[12] Jonathan Ho, Ajay Jain, and Pieter Abbeel. Denoising diffusion probabilistic models. Advances in neural information processing systems, 33:6840-6851, 2020. 1, 2 +[13] Noboru Isobe, Masanori Koyama, Jinzhe Zhang, Kohei Hayashi, and Kenji Fukumizu. Extended flow matching: a method of conditional generation with generalized continuity equation. arXiv preprint arXiv:2402.18839, 2024. 5 + +[14] Moksh Jain, Sharath Chandra Raparthy, Alex Hernández-García, Jarrid Rector-Brooks, Yoshua Bengio, Santiago Miret, and Emmanuel Bengio. Multi-objective gflownets. In International Conference on Machine Learning, pages 14631-14653. PMLR, 2023. 3 +[15] Tero Karras, Timo Aila, Samuli Laine, and Jaakko Lehtinen. Progressive growing of gans for improved quality, stability, and variation. International Conference on Learning Representations, 2018. 1 +[16] Tero Karras, Miika Aittala, Timo Aila, and Samuli Laine. Elucidating the design space of diffusion-based generative models. Advances in Neural Information Processing Systems, 2022. 1, 5 +[17] Alex Kulesza, Ben Taskar, et al. Determinantal point processes for machine learning. Foundations and Trends® in Machine Learning, 5(2-3):123-286, 2012. 2, 3 +[18] Tuomas Kynkänniemi, Tero Karras, Samuli Laine, Jaakko Lehtinen, and Timo Aila. Improved precision and recall metric for assessing generative models. Advances in neural information processing systems, 32, 2019. 6 +[19] Yiheng Li, Heyang Jiang, Akio Kodaira, Masayoshi Tomizuka, Kurt Keutzer, and Chenfeng Xu. Immiscible diffusion: Accelerating diffusion training with noise assignment. Advances in neural information processing systems, 2024. 5 +[20] Yaron Lipman, Ricky TQ Chen, Heli Ben-Hamu, Maximilian Nickel, and Matt Le. Flow matching for generative modeling. arXiv preprint arXiv:2210.02747, 2022. 1, 2, 3, 6 +[21] Qiang Liu and Dilin Wang. Stein variational gradient descent: A general purpose bayesian inference algorithm. Advances in neural information processing systems, 29, 2016. 3 +[22] Xingchao Liu, Chengyue Gong, and Qiang Liu. Flow straight and fast: Learning to generate and transfer data with rectified flow. arXiv preprint arXiv:2209.03003, 2022. 1, 2, 3, 6 +[23] Odile Macchi. The coincidence approach to stochastic point processes. Advances in Applied Probability, 7(1):83-122, 1975. 2, 3 +[24] Robin Rombach, Andreas Blattmann, Dominik Lorenz, Patrick Esser, and Björn Ommer. High-resolution image synthesis with latent diffusion models. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 10684-10695, 2022. 1 +[25] Christoph Schuhmann and LAION. Improved aesthetic predictor: Clip+mlp aesthetic score predictor, 2022. https://github.com/christophschuhmann/improved-aesthetic-predictor. 7 +[26] Jascha Sohl-Dickstein, Eric Weiss, Niru Maheswaranathan, and Surya Ganguli. Deep unsupervised learning using nonequilibrium thermodynamics. In International Conference on Machine Learning, pages 2256-2265. PMLR, 2015. 1 +[27] Gowthami Somepalli, Vasu Singla, Micah Goldblum, Jonas Geiping, and Tom Goldstein. Diffusion art or digital forgery? investigating data replication in diffusion models. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 6048-6058, 2023. 7 +[28] Jiaming Song, Chenlin Meng, and Stefano Ermon. Denoising diffusion implicit models. International Conference on Learning Representations, 2021. 1 + +[29] Yang Song, Jascha Sohl-Dickstein, Diederik P Kingma, Abhishek Kumar, Stefano Ermon, and Ben Poole. Score-based generative modeling through stochastic differential equations. International Conference on Learning Representations, 2021. 2 +[30] Alexander Tong, Nikolay Malkin, Guillaume Huguet, Yanlei Zhang, Jarrid Rector-Brooks, Kilian Fatras, Guy Wolf, and Yoshua Bengio. Improving and generalizing flow-based generative models with minibatch optimal transport. Transactions on Machine Learning Research, 2023. 1, 2, 3, 7, 5 +[31] Alexander Y Tong, Nikolay Malkin, Kilian Fatras, Lazar Atanackovic, Yanlei Zhang, Guillaume Huguet, Guy Wolf, and Yoshua Bengio. Simulation-free schrödinger bridges via score and flow matching. In International Conference on Artificial Intelligence and Statistics, pages 1279-1287. PMLR, 2024. 7 +[32] Zhisheng Xiao, Karsten Kreis, and Arash Vahdat. Tackling the generative learning trilemma with denoising diffusion gans. International Conference on Learning Representations, 2022. 1 +[33] Ye Yuan and Kris Kitani. Diverse trajectory forecasting with determinantal point processes. International Conference on Learning Representations, 2020. 3, 5 +[34] Qinqing Zheng, Matt Le, Neta Shaul, Yaron Lipman, Aditya Grover, and Ricky TQ Chen. Guided flows for generative modeling and decision making. arXiv preprint arXiv:2311.13443, 2023. 1 +[35] Yinglin Zheng, Hao Yang, Ting Zhang, Jianmin Bao, Dongdong Chen, Yangyu Huang, Lu Yuan, Dong Chen, Ming Zeng, and Fang Wen. General facial representation learning in a visual-linguistic manner. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 18697–18709, 2022. 2 \ No newline at end of file diff --git a/2025/DiverseFlow_ Sample-Efficient Diverse Mode Coverage in Flows/images.zip b/2025/DiverseFlow_ Sample-Efficient Diverse Mode Coverage in Flows/images.zip new file mode 100644 index 0000000000000000000000000000000000000000..4467be76a52e42a1192b3abe71d5ada9440704ad --- /dev/null +++ b/2025/DiverseFlow_ Sample-Efficient Diverse Mode Coverage in Flows/images.zip @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:edd49ad408351a9ec0fe2548b45442adf040dfda250bda554fdb2372148c50f0 +size 827132 diff --git a/2025/DiverseFlow_ Sample-Efficient Diverse Mode Coverage in Flows/layout.json b/2025/DiverseFlow_ Sample-Efficient Diverse Mode Coverage in Flows/layout.json new file mode 100644 index 0000000000000000000000000000000000000000..40a68515c13010e6db2f34652151ecc53313a2a2 --- /dev/null +++ b/2025/DiverseFlow_ Sample-Efficient Diverse Mode Coverage in Flows/layout.json @@ -0,0 +1,9670 @@ +{ + "pdf_info": [ + { + "para_blocks": [ + { + "bbox": [ + 108, + 103, + 503, + 121 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 108, + 103, + 503, + 121 + ], + "spans": [ + { + "bbox": [ + 108, + 103, + 503, + 121 + ], + "type": "text", + "content": "DiverseFlow: Sample-Efficient Diverse Mode Coverage in Flows" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 197, + 144, + 412, + 156 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 197, + 144, + 412, + 156 + ], + "spans": [ + { + "bbox": [ + 197, + 144, + 412, + 156 + ], + "type": "text", + "content": "Mashrur M. Morshed Vishnu Boddeti" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 241, + 158, + 369, + 172 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 241, + 158, + 369, + 172 + ], + "spans": [ + { + "bbox": [ + 241, + 158, + 369, + 172 + ], + "type": "text", + "content": "Michigan State University" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 234, + 174, + 374, + 184 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 234, + 174, + 374, + 184 + ], + "spans": [ + { + "bbox": [ + 234, + 174, + 374, + 184 + ], + "type": "text", + "content": "{morshedm, vishnu}@msu.edu" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 152, + 213, + 200, + 225 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 152, + 213, + 200, + 225 + ], + "spans": [ + { + "bbox": [ + 152, + 213, + 200, + 225 + ], + "type": "text", + "content": "Abstract" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 54, + 239, + 297, + 454 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 54, + 239, + 297, + 454 + ], + "spans": [ + { + "bbox": [ + 54, + 239, + 297, + 454 + ], + "type": "text", + "content": "Many real-world applications of flow-based generative models desire a diverse set of samples that cover multiple modes of the target distribution. However, the predominant approach for obtaining diverse sets is not sample-efficient, as it involves independently obtaining many samples from the source distribution and mapping them through the flow until the desired mode coverage is achieved. As an alternative to repeated sampling, we introduce DiverseFlow: a training-free approach to improve the diversity of flow models. Our key idea is to employ a determinantal point process to induce a coupling between the samples that drives diversity under a fixed sampling budget. In essence, DiverseFlow allows exploration of more variations in a learned flow model with fewer samples. We demonstrate the efficacy of our method for tasks where sample-efficient diversity is desirable, such as text-guided image generation with polysemous words, inverse problems like large-hole inpainting, and class-conditional image synthesis." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 55, + 479, + 135, + 491 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 479, + 135, + 491 + ], + "spans": [ + { + "bbox": [ + 55, + 479, + 135, + 491 + ], + "type": "text", + "content": "1. Introduction" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 55, + 500, + 296, + 666 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 500, + 296, + 666 + ], + "spans": [ + { + "bbox": [ + 55, + 500, + 296, + 666 + ], + "type": "text", + "content": "Consider the task of text-guided image generation from open-ended prompts, like \"A famous boxer\". If we use a generative ordinary differential equation (ODE) to obtain a few images for this given prompt, we may observe something unusual: all the resultant images depict a dog, as shown in Figure 1 (top-left). Such a result is plausible, as the word \"boxer\" has multiple meanings: it can either mean a combat sport athlete or a particular dog breed. However, for the prompt \"A famous boxer\", what if the output desired by the user was actually a human athlete, and not a dog? This situation necessitates obtaining additional samples from the model, until the desired alternate meanings are discovered. But instead of repeated sampling, can we directly observe more meanings by finding a more diverse set?" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 55, + 668, + 296, + 693 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 668, + 296, + 693 + ], + "spans": [ + { + "bbox": [ + 55, + 668, + 296, + 693 + ], + "type": "text", + "content": "Beyond the aforementioned example of text-to-image generation from polysemous1 prompts, sample diversity is a" + } + ] + } + ], + "index": 10 + }, + { + "type": "image", + "bbox": [ + 328, + 224, + 432, + 329 + ], + "blocks": [ + { + "bbox": [ + 365, + 211, + 501, + 222 + ], + "lines": [ + { + "bbox": [ + 365, + 211, + 501, + 222 + ], + "spans": [ + { + "bbox": [ + 365, + 211, + 501, + 222 + ], + "type": "text", + "content": "Prompt: \"A famous boxer\"" + } + ] + } + ], + "index": 11, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 328, + 224, + 432, + 329 + ], + "lines": [ + { + "bbox": [ + 328, + 224, + 432, + 329 + ], + "spans": [ + { + "bbox": [ + 328, + 224, + 432, + 329 + ], + "type": "image", + "image_path": "2c6c83f27981f75e8aed186ad9578fdd73f6e56a46f9aedbd43cad7a9e80cbc1.jpg" + } + ] + } + ], + "index": 12, + "angle": 0, + "type": "image_body" + } + ], + "index": 12 + }, + { + "type": "image", + "bbox": [ + 437, + 225, + 541, + 329 + ], + "blocks": [ + { + "bbox": [ + 437, + 225, + 541, + 329 + ], + "lines": [ + { + "bbox": [ + 437, + 225, + 541, + 329 + ], + "spans": [ + { + "bbox": [ + 437, + 225, + 541, + 329 + ], + "type": "image", + "image_path": "10c363bc192d79a1cc897b1f19467d637245fed58d2a78e9b6a278eab156cfb0.jpg" + } + ] + } + ], + "index": 13, + "angle": 0, + "type": "image_body" + } + ], + "index": 13 + }, + { + "type": "image", + "bbox": [ + 328, + 344, + 432, + 447 + ], + "blocks": [ + { + "bbox": [ + 388, + 332, + 487, + 342 + ], + "lines": [ + { + "bbox": [ + 388, + 332, + 487, + 342 + ], + "spans": [ + { + "bbox": [ + 388, + 332, + 487, + 342 + ], + "type": "text", + "content": "Prompt: \"A letter\"" + } + ] + } + ], + "index": 14, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 328, + 344, + 432, + 447 + ], + "lines": [ + { + "bbox": [ + 328, + 344, + 432, + 447 + ], + "spans": [ + { + "bbox": [ + 328, + 344, + 432, + 447 + ], + "type": "image", + "image_path": "3a64acd2ec62494943ef07762881d650a82afc857674e2990bd37cf3542485b1.jpg" + } + ] + } + ], + "index": 15, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 373, + 450, + 389, + 460 + ], + "lines": [ + { + "bbox": [ + 373, + 450, + 389, + 460 + ], + "spans": [ + { + "bbox": [ + 373, + 450, + 389, + 460 + ], + "type": "text", + "content": "ID" + } + ] + } + ], + "index": 17, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 313, + 468, + 555, + 502 + ], + "lines": [ + { + "bbox": [ + 313, + 468, + 555, + 502 + ], + "spans": [ + { + "bbox": [ + 313, + 468, + 555, + 502 + ], + "type": "text", + "content": "Figure 1. Text-guided image generation with polysemous words. With the same budget of samples, DiverseFlow (right) finds a more diverse set of results compared to IID sampling (left)." + } + ] + } + ], + "index": 19, + "angle": 0, + "type": "image_caption" + } + ], + "index": 15 + }, + { + "type": "image", + "bbox": [ + 437, + 344, + 541, + 447 + ], + "blocks": [ + { + "bbox": [ + 437, + 344, + 541, + 447 + ], + "lines": [ + { + "bbox": [ + 437, + 344, + 541, + 447 + ], + "spans": [ + { + "bbox": [ + 437, + 344, + 541, + 447 + ], + "type": "image", + "image_path": "c156a42342a6e920569d9a0c73cb438d810e79299180ef4b8e2d5515044e4ff2.jpg" + } + ] + } + ], + "index": 16, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 470, + 450, + 522, + 460 + ], + "lines": [ + { + "bbox": [ + 470, + 450, + 522, + 460 + ], + "spans": [ + { + "bbox": [ + 470, + 450, + 522, + 460 + ], + "type": "text", + "content": "DiverseFlow" + } + ] + } + ], + "index": 18, + "angle": 0, + "type": "image_caption" + } + ], + "index": 16 + }, + { + "bbox": [ + 313, + 521, + 555, + 664 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 521, + 555, + 664 + ], + "spans": [ + { + "bbox": [ + 313, + 521, + 555, + 664 + ], + "type": "text", + "content": "desirable objective for many other tasks that use generative models. These include inverse problems (e.g., large hole filling) and class-conditioned image generation, to name a few. Diversity or mode coverage is a key pillar in the generative learning trilemma [32], in addition to fidelity and latency. For state-of-the-art generative methods such as flow matching models (FM) [20, 22] and diffusion models (DM) [12, 26], significant work has been done on improving the photorealism of samples and the efficiency of the sampling process [11, 16, 20, 28, 30, 34]. However, relatively little attention has been paid to explicitly enhancing the diversity of generated samples under a limited sampling budget." + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 313, + 665, + 556, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 665, + 556, + 714 + ], + "spans": [ + { + "bbox": [ + 313, + 665, + 556, + 714 + ], + "type": "text", + "content": "The standard approach to generate a diverse set of images is to repeatedly obtain independent and identically distributed (IID) samples from a simple or tractable source distribution (e.g., Gaussian distribution), map them to samples" + } + ] + } + ], + "index": 21 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "spans": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "type": "text", + "content": "CVF" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 145, + 0, + 496, + 37 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 145, + 0, + 496, + 37 + ], + "spans": [ + { + "bbox": [ + 145, + 0, + 496, + 37 + ], + "type": "text", + "content": "This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 67, + 702, + 203, + 713 + ], + "type": "page_footnote", + "angle": 0, + "lines": [ + { + "bbox": [ + 67, + 702, + 203, + 713 + ], + "spans": [ + { + "bbox": [ + 67, + 702, + 203, + 713 + ], + "type": "text", + "content": "1Words or phrases with several meanings." + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "text", + "content": "23303" + } + ] + } + ], + "index": 23 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 0 + }, + { + "para_blocks": [ + { + "bbox": [ + 55, + 72, + 296, + 191 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 72, + 296, + 191 + ], + "spans": [ + { + "bbox": [ + 55, + 72, + 296, + 191 + ], + "type": "text", + "content": "in the target distribution, and continue this process until we observe sufficient mode coverage in the target distribution. This process, while effective, is sample-inefficient, requiring the generation of more results than necessary. Importantly, the mapping from the source to the target density does not hold a linear relationship; even specifically selecting diverse samples from the source distribution by design does not necessarily yield diverse samples in the target distribution. These limitations naturally raise the following research question." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 70, + 211, + 281, + 236 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 70, + 211, + 281, + 236 + ], + "spans": [ + { + "bbox": [ + 70, + 211, + 281, + 236 + ], + "type": "text", + "content": "How can we generate diverse samples from the target density under a limited sampling budget?" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 55, + 254, + 296, + 350 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 254, + 296, + 350 + ], + "spans": [ + { + "bbox": [ + 55, + 254, + 296, + 350 + ], + "type": "text", + "content": "In this paper, we propose DiverseFlow, an inference-time, training-free approach to obtain a diverse set of samples in a desired target density under a fixed sampling budget. We focus on deterministic ODE sampling in continuous-time generative models, specifically FMs, an emerging generative paradigm that enables simulation-free training of continuous normalizing flows (CNFs) and includes diffusion models as a special case." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 55, + 352, + 296, + 568 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 352, + 296, + 568 + ], + "spans": [ + { + "bbox": [ + 55, + 352, + 296, + 568 + ], + "type": "text", + "content": "DiverseFlow measures the diversity of a set of samples through the volume they span in the target space. A set of similar samples span a lower volume, while a diverse set naturally spans a larger volume. We impose a volume-based gradient constraint on the flow ODE by drawing on determinantal point processes (DPP) [17, 23], a probabilistic model arising from quantum physics that exactly describes the Pauli exclusion principle: that no two fermions may occupy the same quantum state. In Figure 1 consider the images generated via IID sampling (left column) from a text-to-image generative ODE—for both the prompts \"A famous boxer\" and \"A letter\", only a single potential meaning is discovered. Unlike the results generated via IID sampling, samples obtained from DiverseFlow (right column) span more diverse modes corresponding to the polysemous words in the prompts. For \"A famous boxer\", the samples show both a dog breed and a human athlete; for \"A letter\", we observe both written correspondence and alphabet symbols." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 55, + 570, + 296, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 570, + 296, + 713 + ], + "spans": [ + { + "bbox": [ + 55, + 570, + 296, + 713 + ], + "type": "text", + "content": "We empirically demonstrate the utility of DiverseFlow across several applications where diversity is inherently desirable. First, we use DiverseFlow to perform text-guided image synthesis for words and phrases that may carry a variety of meanings. Second, we perform large-hole face inpainting with occlusion masks covering significant regions of the face that may be important to the person's identity. Third, we apply DiverseFlow on class-conditioned image synthesis and demonstrate that we can more efficiently explore the data space compared to IID sampling. Lastly, to better characterize and explain the behavior of DiverseFlow, we perform several experiments on synthetic 2D densities." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 321, + 78, + 440, + 88 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 321, + 78, + 440, + 88 + ], + "spans": [ + { + "bbox": [ + 321, + 78, + 440, + 88 + ], + "type": "text", + "content": "Summary of Contributions" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 323, + 90, + 548, + 221 + ], + "type": "list", + "angle": 0, + "index": 9, + "blocks": [ + { + "bbox": [ + 323, + 90, + 548, + 137 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 323, + 90, + 548, + 137 + ], + "spans": [ + { + "bbox": [ + 323, + 90, + 548, + 137 + ], + "type": "text", + "content": "1. We present a sample-efficient method to obtain a diverse set of samples from a flow ODE in Section 5 and demonstrate it qualitatively in Sections 6.1 to 6.3 and quantitatively in Sections 6.4 and 6.5" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 323, + 137, + 548, + 173 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 323, + 137, + 548, + 173 + ], + "spans": [ + { + "bbox": [ + 323, + 137, + 548, + 173 + ], + "type": "text", + "content": "2. We provide an empirical analysis that demonstrates that our method is consistent across various flow matching formulations (Section 6.4)" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 323, + 174, + 548, + 221 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 323, + 174, + 548, + 221 + ], + "spans": [ + { + "bbox": [ + 323, + 174, + 548, + 221 + ], + "type": "text", + "content": "3. We introduce the task of text-to-image synthesis from prompts with polysemous words in the context of analyzing sample diversity, and provide a comparison with relevant methods (Section 6.5)" + } + ] + } + ], + "index": 8 + } + ], + "sub_type": "text" + }, + { + "bbox": [ + 314, + 244, + 399, + 256 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 314, + 244, + 399, + 256 + ], + "spans": [ + { + "bbox": [ + 314, + 244, + 399, + 256 + ], + "type": "text", + "content": "2. Preliminaries" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 314, + 265, + 408, + 277 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 314, + 265, + 408, + 277 + ], + "spans": [ + { + "bbox": [ + 314, + 265, + 408, + 277 + ], + "type": "text", + "content": "2.1. Flow Matching" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 313, + 283, + 555, + 521 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 283, + 555, + 521 + ], + "spans": [ + { + "bbox": [ + 313, + 283, + 555, + 521 + ], + "type": "text", + "content": "Many generative models can be considered as a transport map from some easy-to-sample source distribution to an empirically observed yet unknown target distribution. Recent successes in generative modeling represent this transport map in the form of continuous-time processes, such as stochastic differential equations (SDEs) [12, 29], or ordinary differential equations (ODEs) [2, 20, 22]. Although diffusion models are formulated as SDEs, a significant body of research focuses on converting the diffusion SDE to a deterministic ODE at inference time for faster inference. The diffusion ODE, or probability flow ODE, is a particular case of continuous normalizing flows (CNFs). Flow Matching (FM) [2, 20, 22] is motivated by the idea of directly training CNFs in a scalable and simulation-free manner, just like diffusion models. Moreover, many recent text-to-image generative models, such as Stable Diffusion 3 [10], adopt the FM framework. As such, we present our approach primarily in the context of FM, and our findings can be extended to diffusion and score-based generative models in a straightforward manner." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 313, + 523, + 555, + 547 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 523, + 555, + 547 + ], + "spans": [ + { + "bbox": [ + 313, + 523, + 555, + 547 + ], + "type": "text", + "content": "A CNF reshapes a prior source density " + }, + { + "bbox": [ + 313, + 523, + 555, + 547 + ], + "type": "inline_equation", + "content": "p_0" + }, + { + "bbox": [ + 313, + 523, + 555, + 547 + ], + "type": "text", + "content": " to the empirically observed target density " + }, + { + "bbox": [ + 313, + 523, + 555, + 547 + ], + "type": "inline_equation", + "content": "p_1" + }, + { + "bbox": [ + 313, + 523, + 555, + 547 + ], + "type": "text", + "content": " with an ODE of the form:" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 372, + 561, + 555, + 573 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 372, + 561, + 555, + 573 + ], + "spans": [ + { + "bbox": [ + 372, + 561, + 555, + 573 + ], + "type": "interline_equation", + "content": "d \\mathbf {x} _ {t} = v _ {\\theta} \\left(\\mathbf {x} _ {t}, t\\right) d t, \\quad \\mathbf {x} _ {0} \\sim p _ {0} \\tag {1}", + "image_path": "1e5a964ea819d4ac39a5c4d7b9643ac48c6c239c86e5a8a52c66e4b729a0cd0c.jpg" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 313, + 581, + 555, + 689 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 581, + 555, + 689 + ], + "spans": [ + { + "bbox": [ + 313, + 581, + 555, + 689 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 313, + 581, + 555, + 689 + ], + "type": "inline_equation", + "content": "v_{\\theta}" + }, + { + "bbox": [ + 313, + 581, + 555, + 689 + ], + "type": "text", + "content": " is a time-dependent velocity field whose parameters " + }, + { + "bbox": [ + 313, + 581, + 555, + 689 + ], + "type": "inline_equation", + "content": "\\theta" + }, + { + "bbox": [ + 313, + 581, + 555, + 689 + ], + "type": "text", + "content": " are learned; we interchangeably use the notation " + }, + { + "bbox": [ + 313, + 581, + 555, + 689 + ], + "type": "inline_equation", + "content": "v_{t}" + }, + { + "bbox": [ + 313, + 581, + 555, + 689 + ], + "type": "text", + "content": " to imply " + }, + { + "bbox": [ + 313, + 581, + 555, + 689 + ], + "type": "inline_equation", + "content": "v_{\\theta}(\\cdot, t)" + }, + { + "bbox": [ + 313, + 581, + 555, + 689 + ], + "type": "text", + "content": ". It becomes possible to obtain samples from " + }, + { + "bbox": [ + 313, + 581, + 555, + 689 + ], + "type": "inline_equation", + "content": "p_{1}" + }, + { + "bbox": [ + 313, + 581, + 555, + 689 + ], + "type": "text", + "content": " by integrating Equation (1) over time, i.e. by starting at " + }, + { + "bbox": [ + 313, + 581, + 555, + 689 + ], + "type": "inline_equation", + "content": "\\mathbf{x}_0 \\sim p_0" + }, + { + "bbox": [ + 313, + 581, + 555, + 689 + ], + "type": "text", + "content": " for " + }, + { + "bbox": [ + 313, + 581, + 555, + 689 + ], + "type": "inline_equation", + "content": "t = 0" + }, + { + "bbox": [ + 313, + 581, + 555, + 689 + ], + "type": "text", + "content": " and solving the ODE till " + }, + { + "bbox": [ + 313, + 581, + 555, + 689 + ], + "type": "inline_equation", + "content": "t = 1" + }, + { + "bbox": [ + 313, + 581, + 555, + 689 + ], + "type": "text", + "content": ". As our approach is training-free, we do not elaborate on the details of learning to regress the vector field " + }, + { + "bbox": [ + 313, + 581, + 555, + 689 + ], + "type": "inline_equation", + "content": "v_{t}" + }, + { + "bbox": [ + 313, + 581, + 555, + 689 + ], + "type": "text", + "content": "; we encourage interested readers to refer to the works of Lipman et al. [20] and Tong et al. [30] for a primer on training FMs." + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 313, + 689, + 555, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 689, + 555, + 714 + ], + "spans": [ + { + "bbox": [ + 313, + 689, + 555, + 714 + ], + "type": "text", + "content": "At any timestep " + }, + { + "bbox": [ + 313, + 689, + 555, + 714 + ], + "type": "inline_equation", + "content": "t" + }, + { + "bbox": [ + 313, + 689, + 555, + 714 + ], + "type": "text", + "content": " during sampling, an intermediate sample " + }, + { + "bbox": [ + 313, + 689, + 555, + 714 + ], + "type": "inline_equation", + "content": "\\mathbf{x}_t" + }, + { + "bbox": [ + 313, + 689, + 555, + 714 + ], + "type": "text", + "content": " in the flow trajectory can be used to obtain an approx-" + } + ] + } + ], + "index": 16 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 318, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 318, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 318, + 757 + ], + "type": "text", + "content": "23304" + } + ] + } + ], + "index": 17 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 1 + }, + { + "para_blocks": [ + { + "bbox": [ + 55, + 72, + 186, + 83 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 72, + 186, + 83 + ], + "spans": [ + { + "bbox": [ + 55, + 72, + 186, + 83 + ], + "type": "text", + "content": "imation of the target as follows:" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 118, + 95, + 296, + 107 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 118, + 95, + 296, + 107 + ], + "spans": [ + { + "bbox": [ + 118, + 95, + 296, + 107 + ], + "type": "interline_equation", + "content": "\\hat {\\mathbf {x}} _ {1} = \\mathbf {x} _ {t} + v _ {\\theta} (\\mathbf {x} _ {t}, t) (1 - t) \\tag {2}", + "image_path": "bdc47c2de9bdbad9d0c0f0088e700eecef277100453803c914b184d4c267afd6.jpg" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 55, + 118, + 296, + 212 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 118, + 296, + 212 + ], + "spans": [ + { + "bbox": [ + 55, + 118, + 296, + 212 + ], + "type": "text", + "content": "Equation (2) is equivalent to simply taking a large Euler step at any time instance " + }, + { + "bbox": [ + 55, + 118, + 296, + 212 + ], + "type": "inline_equation", + "content": "t" + }, + { + "bbox": [ + 55, + 118, + 296, + 212 + ], + "type": "text", + "content": " and is naturally more accurate as " + }, + { + "bbox": [ + 55, + 118, + 296, + 212 + ], + "type": "inline_equation", + "content": "t" + }, + { + "bbox": [ + 55, + 118, + 296, + 212 + ], + "type": "text", + "content": " approaches " + }, + { + "bbox": [ + 55, + 118, + 296, + 212 + ], + "type": "inline_equation", + "content": "t = 1" + }, + { + "bbox": [ + 55, + 118, + 296, + 212 + ], + "type": "text", + "content": ". Further, Equation (2) is also well suited for ODEs with 'straight' paths, where the direction of the time-varying velocity " + }, + { + "bbox": [ + 55, + 118, + 296, + 212 + ], + "type": "inline_equation", + "content": "v_{t}" + }, + { + "bbox": [ + 55, + 118, + 296, + 212 + ], + "type": "text", + "content": " remains near-constant in time (such as the work of Liu et al. [22]). Similarly, we can estimate the source sample by simply taking a step in the reverse direction:" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 131, + 214, + 296, + 227 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 214, + 296, + 227 + ], + "spans": [ + { + "bbox": [ + 131, + 214, + 296, + 227 + ], + "type": "interline_equation", + "content": "\\hat {\\mathbf {x}} _ {0} = \\mathbf {x} _ {t} - v _ {\\theta} (\\mathbf {x} _ {t}, t) t \\tag {3}", + "image_path": "c147a1456502b3baa1fca424b636a7da54d89c1d8681944fdce79b432091b7ba.jpg" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 55, + 234, + 222, + 246 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 234, + 222, + 246 + ], + "spans": [ + { + "bbox": [ + 55, + 234, + 222, + 246 + ], + "type": "text", + "content": "2.2. Determinantal Point Processes" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 55, + 251, + 296, + 361 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 251, + 296, + 361 + ], + "spans": [ + { + "bbox": [ + 55, + 251, + 296, + 361 + ], + "type": "text", + "content": "Determinantal point processes (DPPs) [4, 17, 23] are probabilistic models of repulsion between points. They were originally termed as fermion processes as they describe the Pauli exclusion principle or antibunching effect in fermions. To define a DPP, we must first consider a set of points, " + }, + { + "bbox": [ + 55, + 251, + 296, + 361 + ], + "type": "inline_equation", + "content": "\\mathcal{V}" + }, + { + "bbox": [ + 55, + 251, + 296, + 361 + ], + "type": "text", + "content": ", and a point process " + }, + { + "bbox": [ + 55, + 251, + 296, + 361 + ], + "type": "inline_equation", + "content": "\\mathcal{P}(\\mathcal{V})" + }, + { + "bbox": [ + 55, + 251, + 296, + 361 + ], + "type": "text", + "content": " — a probability measure on " + }, + { + "bbox": [ + 55, + 251, + 296, + 361 + ], + "type": "inline_equation", + "content": "2^{\\mathcal{V}}" + }, + { + "bbox": [ + 55, + 251, + 296, + 361 + ], + "type": "text", + "content": " (the set of all possible subsets of " + }, + { + "bbox": [ + 55, + 251, + 296, + 361 + ], + "type": "inline_equation", + "content": "\\mathcal{V}" + }, + { + "bbox": [ + 55, + 251, + 296, + 361 + ], + "type": "text", + "content": "). " + }, + { + "bbox": [ + 55, + 251, + 296, + 361 + ], + "type": "inline_equation", + "content": "\\mathcal{P}" + }, + { + "bbox": [ + 55, + 251, + 296, + 361 + ], + "type": "text", + "content": " is determinantal when the probability of choosing a random subset " + }, + { + "bbox": [ + 55, + 251, + 296, + 361 + ], + "type": "inline_equation", + "content": "Y \\subset \\mathcal{V}" + }, + { + "bbox": [ + 55, + 251, + 296, + 361 + ], + "type": "text", + "content": " according to " + }, + { + "bbox": [ + 55, + 251, + 296, + 361 + ], + "type": "inline_equation", + "content": "\\mathcal{P}" + }, + { + "bbox": [ + 55, + 251, + 296, + 361 + ], + "type": "text", + "content": " is given by:" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 79, + 369, + 296, + 397 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 79, + 369, + 296, + 397 + ], + "spans": [ + { + "bbox": [ + 79, + 369, + 296, + 397 + ], + "type": "interline_equation", + "content": "\\mathcal {P} (Y \\subset \\mathcal {Y}) = \\frac {\\det (\\mathbf {L} _ {Y})}{\\sum_ {Y \\subset \\mathcal {Y}} \\det (\\mathbf {L} _ {Y})} = \\frac {\\det (\\mathbf {L} _ {Y})}{\\det (\\mathbf {L} + \\mathbf {I})} \\tag {4}", + "image_path": "98dfe09fbf7240295c94819d846f55ca7b2c829873c3e61f4a50db475456277d.jpg" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 55, + 405, + 297, + 538 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 405, + 297, + 538 + ], + "spans": [ + { + "bbox": [ + 55, + 405, + 297, + 538 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 55, + 405, + 297, + 538 + ], + "type": "inline_equation", + "content": "\\mathbf{L} \\in \\mathbb{R}^{|\\mathcal{Y}| \\times |\\mathcal{Y}|}" + }, + { + "bbox": [ + 55, + 405, + 297, + 538 + ], + "type": "text", + "content": " is a kernel matrix, and " + }, + { + "bbox": [ + 55, + 405, + 297, + 538 + ], + "type": "inline_equation", + "content": "\\mathbf{L}_Y" + }, + { + "bbox": [ + 55, + 405, + 297, + 538 + ], + "type": "text", + "content": " is the sub-kernel matrix indexed by the elements of " + }, + { + "bbox": [ + 55, + 405, + 297, + 538 + ], + "type": "inline_equation", + "content": "Y" + }, + { + "bbox": [ + 55, + 405, + 297, + 538 + ], + "type": "text", + "content": ". Equation (4) has an intuitive geometric interpretation if we consider the kernel " + }, + { + "bbox": [ + 55, + 405, + 297, + 538 + ], + "type": "inline_equation", + "content": "\\mathbf{L}" + }, + { + "bbox": [ + 55, + 405, + 297, + 538 + ], + "type": "text", + "content": " to be constructed from cosine similarity: the determinant of " + }, + { + "bbox": [ + 55, + 405, + 297, + 538 + ], + "type": "inline_equation", + "content": "\\mathbf{L}_Y" + }, + { + "bbox": [ + 55, + 405, + 297, + 538 + ], + "type": "text", + "content": " is the Gram-determinant, describing the squared volume of the " + }, + { + "bbox": [ + 55, + 405, + 297, + 538 + ], + "type": "inline_equation", + "content": "N" + }, + { + "bbox": [ + 55, + 405, + 297, + 538 + ], + "type": "text", + "content": "-dimensional parallelotope spanned by the set of vectors " + }, + { + "bbox": [ + 55, + 405, + 297, + 538 + ], + "type": "inline_equation", + "content": "Y" + }, + { + "bbox": [ + 55, + 405, + 297, + 538 + ], + "type": "text", + "content": ". Thus, a DPP naturally assigns higher probabilities to more orthogonal (and thus diverse) subsets that span larger volumes. We leverage DPPs to define a coupled likelihood measure over a set of samples in a flow trajectory." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 55, + 549, + 143, + 562 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 549, + 143, + 562 + ], + "spans": [ + { + "bbox": [ + 55, + 549, + 143, + 562 + ], + "type": "text", + "content": "3. Related Work" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 55, + 570, + 297, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 570, + 297, + 714 + ], + "spans": [ + { + "bbox": [ + 55, + 570, + 297, + 714 + ], + "type": "text", + "content": "Efficiently finding diverse sets is useful in several application areas of machine learning. For instance, Batra et al. [3] show that the M-Best MAP (maximum a posteriori) solutions in Markov random fields are often distant from the ground truth and highly similar. They thus propose the *Diverse M-Best* problem—finding a set of " + }, + { + "bbox": [ + 55, + 570, + 297, + 714 + ], + "type": "inline_equation", + "content": "M" + }, + { + "bbox": [ + 55, + 570, + 297, + 714 + ], + "type": "text", + "content": " highly probable solutions satisfying some minimum dissimilarity threshold—that partly inspires our study in Section 4. [33] utilize DPPs in conjunction with variational autoencoders (VAE) for diverse trajectory forecasting; a set of diverse future pedestrian trajectories improves safety-critical perception systems in autonomous vehicles. Motivated by potential drug discovery" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 313, + 72, + 555, + 107 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 72, + 555, + 107 + ], + "spans": [ + { + "bbox": [ + 313, + 72, + 555, + 107 + ], + "type": "text", + "content": "and material design applications, Jain et al. [14] propose finding diverse Pareto-optimal candidates in a multi-objective setting with generative flow networks." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 313, + 108, + 556, + 263 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 108, + 556, + 263 + ], + "spans": [ + { + "bbox": [ + 313, + 108, + 556, + 263 + ], + "type": "text", + "content": "The work by Corso et al. [6] which explores diverse non-IID sampling for diffusion models is most similar in spirit to DiverseFlow. However, DiverseFlow is notably different in the following aspects: (1) Our diversity objective is derived from determinantal point processes, a diversity-promoting probability measure of the joint occurrence of a set of samples. Corso et al. [6] is instead inspired by stein variational gradient descent (SVGD) [21]. 2) The diversity measure in DiverseFlow (volume, or determinant of similarity kernel) assigns a zero likelihood to a set if any duplicate elements are present; presence of duplicates is tolerated in the diversity measure in Particle Guidance (row-wise sum of similarity kernel)" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 313, + 273, + 555, + 301 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 273, + 555, + 301 + ], + "spans": [ + { + "bbox": [ + 313, + 273, + 555, + 301 + ], + "type": "text", + "content": "4. Diverse Source Samples Do Not Yield Diverse Target Samples" + } + ] + } + ], + "index": 12 + }, + { + "type": "image", + "bbox": [ + 337, + 316, + 434, + 411 + ], + "blocks": [ + { + "bbox": [ + 337, + 316, + 434, + 411 + ], + "lines": [ + { + "bbox": [ + 337, + 316, + 434, + 411 + ], + "spans": [ + { + "bbox": [ + 337, + 316, + 434, + 411 + ], + "type": "image", + "image_path": "46a6f278801b399e96b2a0a6be298519d95dc071360f264fa4b47cfe9688b918.jpg" + } + ] + } + ], + "index": 13, + "angle": 0, + "type": "image_body" + } + ], + "index": 13 + }, + { + "type": "image", + "bbox": [ + 434, + 316, + 531, + 412 + ], + "blocks": [ + { + "bbox": [ + 434, + 316, + 531, + 412 + ], + "lines": [ + { + "bbox": [ + 434, + 316, + 531, + 412 + ], + "spans": [ + { + "bbox": [ + 434, + 316, + 531, + 412 + ], + "type": "image", + "image_path": "df991f020c6c783fd783210b68a9a57b3e516fefd9044bd263eb9d81e3870a24.jpg" + } + ] + } + ], + "index": 15, + "angle": 0, + "type": "image_body" + } + ], + "index": 15 + }, + { + "type": "image", + "bbox": [ + 338, + 433, + 434, + 529 + ], + "blocks": [ + { + "bbox": [ + 338, + 413, + 389, + 433 + ], + "lines": [ + { + "bbox": [ + 338, + 413, + 389, + 433 + ], + "spans": [ + { + "bbox": [ + 338, + 413, + 389, + 433 + ], + "type": "text", + "content": "(a) Conditional ing (CFM) [20]" + } + ] + } + ], + "index": 14, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 338, + 433, + 434, + 529 + ], + "lines": [ + { + "bbox": [ + 338, + 433, + 434, + 529 + ], + "spans": [ + { + "bbox": [ + 338, + 433, + 434, + 529 + ], + "type": "image", + "image_path": "ac6db7105b857c92bd53ac5877c85e725cae59c7094d3570267fb3798f2425cc.jpg" + } + ] + } + ], + "index": 17, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 345, + 530, + 427, + 540 + ], + "lines": [ + { + "bbox": [ + 345, + 530, + 427, + 540 + ], + "spans": [ + { + "bbox": [ + 345, + 530, + 427, + 540 + ], + "type": "text", + "content": "(c) DiverseFlow on CFM" + } + ] + } + ], + "index": 18, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 313, + 550, + 555, + 594 + ], + "lines": [ + { + "bbox": [ + 313, + 550, + 555, + 594 + ], + "spans": [ + { + "bbox": [ + 313, + 550, + 555, + 594 + ], + "type": "text", + "content": "Figure 2. Finding " + }, + { + "bbox": [ + 313, + 550, + 555, + 594 + ], + "type": "inline_equation", + "content": "K = 5" + }, + { + "bbox": [ + 313, + 550, + 555, + 594 + ], + "type": "text", + "content": " samples from the target distribution with " + }, + { + "bbox": [ + 313, + 550, + 555, + 594 + ], + "type": "inline_equation", + "content": "N = 10" + }, + { + "bbox": [ + 313, + 550, + 555, + 594 + ], + "type": "text", + "content": " modes. In the example, only 3 modes are found by 5 IID samples (a, b). DiverseFlow discovers 5 modes with the same sampling budget in (d, e)." + } + ] + } + ], + "index": 21, + "angle": 0, + "type": "image_caption" + } + ], + "index": 17 + }, + { + "type": "image", + "bbox": [ + 434, + 433, + 531, + 529 + ], + "blocks": [ + { + "bbox": [ + 435, + 414, + 532, + 433 + ], + "lines": [ + { + "bbox": [ + 435, + 414, + 532, + 433 + ], + "spans": [ + { + "bbox": [ + 435, + 414, + 532, + 433 + ], + "type": "text", + "content": "(b) Mini-batch Optimal Transport (MB-OT) [30]" + } + ] + } + ], + "index": 16, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 434, + 433, + 531, + 529 + ], + "lines": [ + { + "bbox": [ + 434, + 433, + 531, + 529 + ], + "spans": [ + { + "bbox": [ + 434, + 433, + 531, + 529 + ], + "type": "image", + "image_path": "986aeb4001abbd1298dc23aba33fe1e40a26b7eb3eae94cf5c33f43051bd5d50.jpg" + } + ] + } + ], + "index": 19, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 438, + 530, + 528, + 540 + ], + "lines": [ + { + "bbox": [ + 438, + 530, + 528, + 540 + ], + "spans": [ + { + "bbox": [ + 438, + 530, + 528, + 540 + ], + "type": "text", + "content": "(d) DiverseFlow on MB-OT" + } + ] + } + ], + "index": 20, + "angle": 0, + "type": "image_caption" + } + ], + "index": 19 + }, + { + "bbox": [ + 313, + 605, + 556, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 605, + 556, + 715 + ], + "spans": [ + { + "bbox": [ + 313, + 605, + 556, + 715 + ], + "type": "text", + "content": "Problem Setting: We start with a synthetic example to illustrate our problem of interest. Consider that we have empirical observations from a target distribution " + }, + { + "bbox": [ + 313, + 605, + 556, + 715 + ], + "type": "inline_equation", + "content": "\\pi_1 \\in \\mathbb{R}^2" + }, + { + "bbox": [ + 313, + 605, + 556, + 715 + ], + "type": "text", + "content": ", which is a random mixture of Gaussians, such as the example shown in Figure 4. We design " + }, + { + "bbox": [ + 313, + 605, + 556, + 715 + ], + "type": "inline_equation", + "content": "\\pi_1 = \\sum_{i=1}^{N} w_i \\mathcal{N}(\\pmb{\\mu}_i, \\sigma_i^2 \\pmb{I})" + }, + { + "bbox": [ + 313, + 605, + 556, + 715 + ], + "type": "text", + "content": " to contain " + }, + { + "bbox": [ + 313, + 605, + 556, + 715 + ], + "type": "inline_equation", + "content": "N = 10" + }, + { + "bbox": [ + 313, + 605, + 556, + 715 + ], + "type": "text", + "content": " randomly selected modes " + }, + { + "bbox": [ + 313, + 605, + 556, + 715 + ], + "type": "inline_equation", + "content": "\\mathcal{N}(\\pmb{\\mu}_i, \\sigma_i^2 \\pmb{I})" + }, + { + "bbox": [ + 313, + 605, + 556, + 715 + ], + "type": "text", + "content": ", each with a random mixture weight " + }, + { + "bbox": [ + 313, + 605, + 556, + 715 + ], + "type": "inline_equation", + "content": "w_i" + }, + { + "bbox": [ + 313, + 605, + 556, + 715 + ], + "type": "text", + "content": "; we observe that in our example, there are 6 high probability modes and 4 low probability ones. Suppose we have a sampling budget" + } + ] + } + ], + "index": 22 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "text", + "content": "23305" + } + ] + } + ], + "index": 23 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 2 + }, + { + "para_blocks": [ + { + "bbox": [ + 55, + 72, + 296, + 120 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 72, + 296, + 120 + ], + "spans": [ + { + "bbox": [ + 55, + 72, + 296, + 120 + ], + "type": "text", + "content": "of " + }, + { + "bbox": [ + 55, + 72, + 296, + 120 + ], + "type": "inline_equation", + "content": "K" + }, + { + "bbox": [ + 55, + 72, + 296, + 120 + ], + "type": "text", + "content": " samples. This leads to three possible scenarios: (i) " + }, + { + "bbox": [ + 55, + 72, + 296, + 120 + ], + "type": "inline_equation", + "content": "K < N" + }, + { + "bbox": [ + 55, + 72, + 296, + 120 + ], + "type": "text", + "content": ", (ii) " + }, + { + "bbox": [ + 55, + 72, + 296, + 120 + ], + "type": "inline_equation", + "content": "K = N" + }, + { + "bbox": [ + 55, + 72, + 296, + 120 + ], + "type": "text", + "content": ", and (iii) " + }, + { + "bbox": [ + 55, + 72, + 296, + 120 + ], + "type": "inline_equation", + "content": "K > N" + }, + { + "bbox": [ + 55, + 72, + 296, + 120 + ], + "type": "text", + "content": ". Among the aforementioned, case (i) (fewer samples than modes) is the most likely characteristic of any real-world dataset." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 55, + 122, + 296, + 183 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 122, + 296, + 183 + ], + "spans": [ + { + "bbox": [ + 55, + 122, + 296, + 183 + ], + "type": "text", + "content": "Let us have a prior distribution " + }, + { + "bbox": [ + 55, + 122, + 296, + 183 + ], + "type": "inline_equation", + "content": "\\pi_0" + }, + { + "bbox": [ + 55, + 122, + 296, + 183 + ], + "type": "text", + "content": " and some generative model " + }, + { + "bbox": [ + 55, + 122, + 296, + 183 + ], + "type": "inline_equation", + "content": "\\Psi" + }, + { + "bbox": [ + 55, + 122, + 296, + 183 + ], + "type": "text", + "content": ", such that, in the limit of infinite samples, " + }, + { + "bbox": [ + 55, + 122, + 296, + 183 + ], + "type": "inline_equation", + "content": "\\Psi(x_0 \\sim \\pi_0) \\sim \\pi_1" + }, + { + "bbox": [ + 55, + 122, + 296, + 183 + ], + "type": "text", + "content": ". Then, the objective of sample-efficient diverse sampling is to obtain samples from " + }, + { + "bbox": [ + 55, + 122, + 296, + 183 + ], + "type": "inline_equation", + "content": "\\min(K, N)" + }, + { + "bbox": [ + 55, + 122, + 296, + 183 + ], + "type": "text", + "content": " modes from " + }, + { + "bbox": [ + 55, + 122, + 296, + 183 + ], + "type": "inline_equation", + "content": "\\pi_1" + }, + { + "bbox": [ + 55, + 122, + 296, + 183 + ], + "type": "text", + "content": ", given a fixed set of " + }, + { + "bbox": [ + 55, + 122, + 296, + 183 + ], + "type": "inline_equation", + "content": "K" + }, + { + "bbox": [ + 55, + 122, + 296, + 183 + ], + "type": "text", + "content": " samples in " + }, + { + "bbox": [ + 55, + 122, + 296, + 183 + ], + "type": "inline_equation", + "content": "\\pi_0" + }, + { + "bbox": [ + 55, + 122, + 296, + 183 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 55, + 184, + 298, + 340 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 184, + 298, + 340 + ], + "spans": [ + { + "bbox": [ + 55, + 184, + 298, + 340 + ], + "type": "text", + "content": "If diverse samples are desired from the target density of the flow, one may make the elementary assumption that if the particles are distant at the source distribution, after being transported by the flow, they remain distant in the target distribution. This assumption is not necessarily true, as we show in Figure 2. By design, we choose a uniform mixture of eight Gaussians as the source " + }, + { + "bbox": [ + 55, + 184, + 298, + 340 + ], + "type": "inline_equation", + "content": "\\pi_0" + }, + { + "bbox": [ + 55, + 184, + 298, + 340 + ], + "type": "text", + "content": " to obtain diverse source samples. In Figure 2a, we can observe that source points from distinct modes can still converge to the same target mode with IID sampling. Thus, an alternative procedure is necessary to obtain a diverse set from a flow in a sample-efficient manner. We further explore this toy problem in Section 6.4." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 55, + 358, + 222, + 372 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 358, + 222, + 372 + ], + "spans": [ + { + "bbox": [ + 55, + 358, + 222, + 372 + ], + "type": "text", + "content": "5. Diverse Sampling from Flows" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 55, + 380, + 296, + 589 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 380, + 296, + 589 + ], + "spans": [ + { + "bbox": [ + 55, + 380, + 296, + 589 + ], + "type": "text", + "content": "From Figure 2, we observe that independently (or heuristically) chosen source samples may not map to a diverse set of target samples. In this case, we can select a new set of source samples and repeat the sampling process till eventually covering at least " + }, + { + "bbox": [ + 55, + 380, + 296, + 589 + ], + "type": "inline_equation", + "content": "K" + }, + { + "bbox": [ + 55, + 380, + 296, + 589 + ], + "type": "text", + "content": " modes. However, this approach does not satisfy our fixed sampling budget constraint. An alternative solution to repeated independent sampling is defining and leveraging a diversity measure of the target samples to drive sample diversity. For the set of source samples " + }, + { + "bbox": [ + 55, + 380, + 296, + 589 + ], + "type": "inline_equation", + "content": "\\{\\mathbf{x}_0^{(1)},\\mathbf{x}_0^{(2)},\\ldots ,\\mathbf{x}_0^{(k)}\\}" + }, + { + "bbox": [ + 55, + 380, + 296, + 589 + ], + "type": "text", + "content": ", we could optimize a set of perturbations " + }, + { + "bbox": [ + 55, + 380, + 296, + 589 + ], + "type": "inline_equation", + "content": "\\{\\delta^{(1)},\\delta^{(2)},\\dots,\\delta^{(k)}\\}" + }, + { + "bbox": [ + 55, + 380, + 296, + 589 + ], + "type": "text", + "content": " such that the new set " + }, + { + "bbox": [ + 55, + 380, + 296, + 589 + ], + "type": "inline_equation", + "content": "\\{\\mathbf{x}_0^{(1)} + \\delta^{(1)},\\mathbf{x}_0^{(2)} + \\delta^{(2)},\\dots,\\mathbf{x}_0^{(k)} + \\delta^{(k)}\\}" + }, + { + "bbox": [ + 55, + 380, + 296, + 589 + ], + "type": "text", + "content": " maps to a diverse set of target particles. However, this approach would require multiple simulations of the whole ODE and backpropagating over all the timesteps, which increases the computational complexity of the sampling process over the standard IID sampling." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 55, + 590, + 296, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 590, + 296, + 715 + ], + "spans": [ + { + "bbox": [ + 55, + 590, + 296, + 715 + ], + "type": "text", + "content": "This leads us to our proposed approach: we avoid multiple simulations and instead optimize the flow trajectory for diversity while solving the ODE. For any sample in the flow trajectory " + }, + { + "bbox": [ + 55, + 590, + 296, + 715 + ], + "type": "inline_equation", + "content": "\\mathbf{x}_t" + }, + { + "bbox": [ + 55, + 590, + 296, + 715 + ], + "type": "text", + "content": ", we have an estimate of the target sample " + }, + { + "bbox": [ + 55, + 590, + 296, + 715 + ], + "type": "inline_equation", + "content": "\\hat{\\mathbf{x}}_1" + }, + { + "bbox": [ + 55, + 590, + 296, + 715 + ], + "type": "text", + "content": " through Equation (2). Suppose we have a differentiable objective " + }, + { + "bbox": [ + 55, + 590, + 296, + 715 + ], + "type": "inline_equation", + "content": "\\mathcal{L}(\\{\\hat{\\mathbf{x}}_1^{(1)},\\hat{\\mathbf{x}}_1^{(2)},\\dots ,\\hat{\\mathbf{x}}_1^{(k)}\\})" + }, + { + "bbox": [ + 55, + 590, + 296, + 715 + ], + "type": "text", + "content": " that assigns a likelihood to the joint outcome " + }, + { + "bbox": [ + 55, + 590, + 296, + 715 + ], + "type": "inline_equation", + "content": "\\{\\hat{\\mathbf{x}}_1^{(1)},\\hat{\\mathbf{x}}_1^{(2)},\\dots ,\\hat{\\mathbf{x}}_1^{(k)}\\}" + }, + { + "bbox": [ + 55, + 590, + 296, + 715 + ], + "type": "text", + "content": ". Further, let " + }, + { + "bbox": [ + 55, + 590, + 296, + 715 + ], + "type": "inline_equation", + "content": "\\mathcal{L}" + }, + { + "bbox": [ + 55, + 590, + 296, + 715 + ], + "type": "text", + "content": " assign a higher likelihood if the joint outcome is a diverse set, and a diminished likelihood if the set is similar. We can then leverage " + }, + { + "bbox": [ + 55, + 590, + 296, + 715 + ], + "type": "inline_equation", + "content": "\\mathcal{L}" + }, + { + "bbox": [ + 55, + 590, + 296, + 715 + ], + "type": "text", + "content": " to drive diversity among the target samples" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 313, + 72, + 529, + 85 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 72, + 529, + 85 + ], + "spans": [ + { + "bbox": [ + 313, + 72, + 529, + 85 + ], + "type": "text", + "content": "by modifying the flow velocity of the " + }, + { + "bbox": [ + 313, + 72, + 529, + 85 + ], + "type": "inline_equation", + "content": "i" + }, + { + "bbox": [ + 313, + 72, + 529, + 85 + ], + "type": "text", + "content": "-th particle as," + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 320, + 91, + 555, + 111 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 320, + 91, + 555, + 111 + ], + "spans": [ + { + "bbox": [ + 320, + 91, + 555, + 111 + ], + "type": "interline_equation", + "content": "\\tilde {v} _ {t} ^ {(i)} = v _ {t} ^ {(i)} - \\gamma (t) \\nabla_ {\\mathbf {x} _ {t} ^ {(i)}} \\log \\mathcal {L} (\\{\\hat {\\mathbf {x}} _ {1} ^ {(1)}, \\hat {\\mathbf {x}} _ {1} ^ {(2)}, \\dots , \\hat {\\mathbf {x}} _ {1} ^ {(k)} \\}) \\tag {5}", + "image_path": "6c7a5f5a1ad139d09585bc969e9b6b7f376753f32e258bc4d81cc52983add03c.jpg" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 313, + 119, + 555, + 192 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 119, + 555, + 192 + ], + "spans": [ + { + "bbox": [ + 313, + 119, + 555, + 192 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 313, + 119, + 555, + 192 + ], + "type": "inline_equation", + "content": "\\gamma (t)" + }, + { + "bbox": [ + 313, + 119, + 555, + 192 + ], + "type": "text", + "content": " is a time-varying scale that controls the strength of the diversity gradient. Setting " + }, + { + "bbox": [ + 313, + 119, + 555, + 192 + ], + "type": "inline_equation", + "content": "\\gamma (t) = 0" + }, + { + "bbox": [ + 313, + 119, + 555, + 192 + ], + "type": "text", + "content": " reduces to the standard IID sampling scenario, while " + }, + { + "bbox": [ + 313, + 119, + 555, + 192 + ], + "type": "inline_equation", + "content": "\\gamma (t) > 0" + }, + { + "bbox": [ + 313, + 119, + 555, + 192 + ], + "type": "text", + "content": " will encourage diversity between the generated samples. In practice, " + }, + { + "bbox": [ + 313, + 119, + 555, + 192 + ], + "type": "inline_equation", + "content": "\\gamma (t)" + }, + { + "bbox": [ + 313, + 119, + 555, + 192 + ], + "type": "text", + "content": " follows the schedule of the probability path normalized by the norm of the DPP gradient." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 313, + 198, + 508, + 210 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 198, + 508, + 210 + ], + "spans": [ + { + "bbox": [ + 313, + 198, + 508, + 210 + ], + "type": "text", + "content": "5.1. Determinantal Gradient Constraints" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 313, + 215, + 555, + 347 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 215, + 555, + 347 + ], + "spans": [ + { + "bbox": [ + 313, + 215, + 555, + 347 + ], + "type": "text", + "content": "We desire objective " + }, + { + "bbox": [ + 313, + 215, + 555, + 347 + ], + "type": "inline_equation", + "content": "\\mathcal{L}" + }, + { + "bbox": [ + 313, + 215, + 555, + 347 + ], + "type": "text", + "content": " in Equation (5) to be higher if the items in the set are diverse and lower if they are similar to each other. We interpret diversity in terms of the volume spanned by the set. Consider that we have " + }, + { + "bbox": [ + 313, + 215, + 555, + 347 + ], + "type": "inline_equation", + "content": "k" + }, + { + "bbox": [ + 313, + 215, + 555, + 347 + ], + "type": "text", + "content": " samples in " + }, + { + "bbox": [ + 313, + 215, + 555, + 347 + ], + "type": "inline_equation", + "content": "\\mathbb{R}^d" + }, + { + "bbox": [ + 313, + 215, + 555, + 347 + ], + "type": "text", + "content": " (assume " + }, + { + "bbox": [ + 313, + 215, + 555, + 347 + ], + "type": "inline_equation", + "content": "k < d" + }, + { + "bbox": [ + 313, + 215, + 555, + 347 + ], + "type": "text", + "content": "). An objective that prefers diversity can be defined as the volume of the " + }, + { + "bbox": [ + 313, + 215, + 555, + 347 + ], + "type": "inline_equation", + "content": "k" + }, + { + "bbox": [ + 313, + 215, + 555, + 347 + ], + "type": "text", + "content": "-dimensional parallelotope in " + }, + { + "bbox": [ + 313, + 215, + 555, + 347 + ], + "type": "inline_equation", + "content": "\\mathbb{R}^d" + }, + { + "bbox": [ + 313, + 215, + 555, + 347 + ], + "type": "text", + "content": " spanned by the sample vectors; this volume becomes diminished when there are similar samples (and even zero, for identical samples). The determinant describes volumes well; a diverse set must span a large volume in the sample space and have a corresponding large determinant." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 313, + 348, + 554, + 386 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 348, + 554, + 386 + ], + "spans": [ + { + "bbox": [ + 313, + 348, + 554, + 386 + ], + "type": "text", + "content": "To define a measure over a set of samples, we draw on the idea of determinantal point processes (DPP). We first define a kernel " + }, + { + "bbox": [ + 313, + 348, + 554, + 386 + ], + "type": "inline_equation", + "content": "\\mathbf{L}(\\{\\hat{\\mathbf{x}}_1^{(1)},\\hat{\\mathbf{x}}_1^{(2)},\\dots ,\\hat{\\mathbf{x}}_1^{(k)}\\})" + }, + { + "bbox": [ + 313, + 348, + 554, + 386 + ], + "type": "text", + "content": " as follows:" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 361, + 393, + 555, + 426 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 361, + 393, + 555, + 426 + ], + "spans": [ + { + "bbox": [ + 361, + 393, + 555, + 426 + ], + "type": "interline_equation", + "content": "\\mathbf {L} ^ {(i j)} = \\exp \\left(- h \\frac {\\| \\hat {\\mathbf {x}} _ {1} ^ {(i)} - \\hat {\\mathbf {x}} _ {1} ^ {(j)} \\| _ {2} ^ {2}}{\\operatorname* {m e d} (\\mathbf {U} (\\mathbf {D}))}\\right) \\tag {6}", + "image_path": "f316745fb07602251b45db695329b7e62a7a915ea8c32d084a075692fe139002.jpg" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 313, + 434, + 554, + 495 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 434, + 554, + 495 + ], + "spans": [ + { + "bbox": [ + 313, + 434, + 554, + 495 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 313, + 434, + 554, + 495 + ], + "type": "inline_equation", + "content": "\\mathbf{D}" + }, + { + "bbox": [ + 313, + 434, + 554, + 495 + ], + "type": "text", + "content": " denotes a distance matrix with " + }, + { + "bbox": [ + 313, + 434, + 554, + 495 + ], + "type": "inline_equation", + "content": "\\mathbf{D}_{ij} = \\| \\mathbf{x}^{(i)} - \\mathbf{x}^{(j)}\\| _2^2" + }, + { + "bbox": [ + 313, + 434, + 554, + 495 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 313, + 434, + 554, + 495 + ], + "type": "inline_equation", + "content": "\\mathbf{U}(\\mathbf{D})" + }, + { + "bbox": [ + 313, + 434, + 554, + 495 + ], + "type": "text", + "content": " denotes the upper triangle entries of " + }, + { + "bbox": [ + 313, + 434, + 554, + 495 + ], + "type": "inline_equation", + "content": "\\mathbf{D}" + }, + { + "bbox": [ + 313, + 434, + 554, + 495 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 313, + 434, + 554, + 495 + ], + "type": "inline_equation", + "content": "h" + }, + { + "bbox": [ + 313, + 434, + 554, + 495 + ], + "type": "text", + "content": " denotes a kernel spread parameter, and " + }, + { + "bbox": [ + 313, + 434, + 554, + 495 + ], + "type": "inline_equation", + "content": "\\mathrm{med}(\\mathbf{U}(\\mathbf{D}))" + }, + { + "bbox": [ + 313, + 434, + 554, + 495 + ], + "type": "text", + "content": " denotes the median of those entries. Given " + }, + { + "bbox": [ + 313, + 434, + 554, + 495 + ], + "type": "inline_equation", + "content": "\\mathbf{L}" + }, + { + "bbox": [ + 313, + 434, + 554, + 495 + ], + "type": "text", + "content": ", we may define a DPP-based likelihood as:" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 344, + 513, + 555, + 574 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 344, + 513, + 555, + 574 + ], + "spans": [ + { + "bbox": [ + 344, + 513, + 555, + 574 + ], + "type": "interline_equation", + "content": "\\begin{array}{l} \\mathcal {L} (\\{\\hat {\\mathbf {x}} _ {1} ^ {(1)}, \\hat {\\mathbf {x}} _ {1} ^ {(2)}, \\dots , \\hat {\\mathbf {x}} _ {1} ^ {(k)} \\}) = \\frac {\\det (\\mathbf {L})}{\\det (\\mathbf {L} + \\mathbf {I})} \\\\ = \\prod_ {a = 1} ^ {k} \\frac {\\lambda (\\mathbf {L}) _ {a}}{1 + \\lambda (\\mathbf {L}) _ {a}} \\tag {7} \\\\ \\end{array}", + "image_path": "afbca959f1051bb532736842a37ebc096d55b7e42820b1857ddfc1b3ca97a56b.jpg" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 313, + 582, + 554, + 608 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 582, + 554, + 608 + ], + "spans": [ + { + "bbox": [ + 313, + 582, + 554, + 608 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 313, + 582, + 554, + 608 + ], + "type": "inline_equation", + "content": "\\lambda (\\mathbf{L})_a" + }, + { + "bbox": [ + 313, + 582, + 554, + 608 + ], + "type": "text", + "content": " is the " + }, + { + "bbox": [ + 313, + 582, + 554, + 608 + ], + "type": "inline_equation", + "content": "a^{\\mathrm{th}}" + }, + { + "bbox": [ + 313, + 582, + 554, + 608 + ], + "type": "text", + "content": " eigenvalue of the kernel " + }, + { + "bbox": [ + 313, + 582, + 554, + 608 + ], + "type": "inline_equation", + "content": "\\mathbf{L}" + }, + { + "bbox": [ + 313, + 582, + 554, + 608 + ], + "type": "text", + "content": ". The log-likelihood is then," + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 344, + 616, + 555, + 629 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 344, + 616, + 555, + 629 + ], + "spans": [ + { + "bbox": [ + 344, + 616, + 555, + 629 + ], + "type": "interline_equation", + "content": "\\mathcal {L} \\mathcal {L} = \\log \\mathcal {L} = \\log \\det (\\mathbf {L}) - \\log \\det (\\mathbf {L} + \\mathbf {I}) \\tag {8}", + "image_path": "6747cd7ed440b6537ae8283c115d3805c9aade28269f73bdf676f85be88a5936.jpg" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 313, + 638, + 555, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 638, + 555, + 714 + ], + "spans": [ + { + "bbox": [ + 313, + 638, + 555, + 714 + ], + "type": "text", + "content": "Note that the Euclidean distance " + }, + { + "bbox": [ + 313, + 638, + 555, + 714 + ], + "type": "inline_equation", + "content": "\\| \\hat{\\mathbf{x}}_1^{(i)} - \\hat{\\mathbf{x}}_1^{(j)}\\| _2^2" + }, + { + "bbox": [ + 313, + 638, + 555, + 714 + ], + "type": "text", + "content": " is not very meaningful in the high-dimensional raw image space [1]. Therefore, in practice, the distance should be computed in a robust feature space, i.e., " + }, + { + "bbox": [ + 313, + 638, + 555, + 714 + ], + "type": "inline_equation", + "content": "\\| F(\\hat{\\mathbf{x}}_1^{(i)}) - F(\\hat{\\mathbf{x}}_1^{(j)})\\| _2^2" + }, + { + "bbox": [ + 313, + 638, + 555, + 714 + ], + "type": "text", + "content": ", where " + }, + { + "bbox": [ + 313, + 638, + 555, + 714 + ], + "type": "inline_equation", + "content": "F" + }, + { + "bbox": [ + 313, + 638, + 555, + 714 + ], + "type": "text", + "content": " is some domain-specific feature extractor, such as the vision transformer (ViT) [9] for images." + } + ] + } + ], + "index": 17 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "type": "text", + "content": "23306" + } + ] + } + ], + "index": 18 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 3 + }, + { + "para_blocks": [ + { + "bbox": [ + 55, + 72, + 296, + 259 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 72, + 296, + 259 + ], + "spans": [ + { + "bbox": [ + 55, + 72, + 296, + 259 + ], + "type": "text", + "content": "Quality Constraint: The DPP defined in Equation (7) acts as a repulsive force that is unaware of sample quality. A quality term can be incorporated into the DPP kernel to regularize the trajectory diversification. Although flows can be defined between any arbitrary two distributions, let us consider the special case when the source is a Gaussian, i.e., " + }, + { + "bbox": [ + 55, + 72, + 296, + 259 + ], + "type": "inline_equation", + "content": "p_0 \\sim \\mathcal{N}(0,\\mathbf{I})" + }, + { + "bbox": [ + 55, + 72, + 296, + 259 + ], + "type": "text", + "content": ". Suppose we have a quality vector " + }, + { + "bbox": [ + 55, + 72, + 296, + 259 + ], + "type": "inline_equation", + "content": "\\mathbf{q}_t = \\{q^{(1)}(t), q^{(2)}(t), \\ldots, q^{(k)}(t)\\}" + }, + { + "bbox": [ + 55, + 72, + 296, + 259 + ], + "type": "text", + "content": ", where any " + }, + { + "bbox": [ + 55, + 72, + 296, + 259 + ], + "type": "inline_equation", + "content": "q^{(i)}(t) \\in [0,1]" + }, + { + "bbox": [ + 55, + 72, + 296, + 259 + ], + "type": "text", + "content": ". We can then define a new kernel " + }, + { + "bbox": [ + 55, + 72, + 296, + 259 + ], + "type": "inline_equation", + "content": "\\mathbf{L}_q = \\mathbf{L} \\odot \\mathbf{q}_t \\mathbf{q}_t^T" + }, + { + "bbox": [ + 55, + 72, + 296, + 259 + ], + "type": "text", + "content": ", where each " + }, + { + "bbox": [ + 55, + 72, + 296, + 259 + ], + "type": "inline_equation", + "content": "q^{(i)}(t)" + }, + { + "bbox": [ + 55, + 72, + 296, + 259 + ], + "type": "text", + "content": " penalizes a sample " + }, + { + "bbox": [ + 55, + 72, + 296, + 259 + ], + "type": "inline_equation", + "content": "\\mathbf{x}_t^{(i)}" + }, + { + "bbox": [ + 55, + 72, + 296, + 259 + ], + "type": "text", + "content": " if it deviates too much from the flow. To define this, we obtain an estimate of the source sample " + }, + { + "bbox": [ + 55, + 72, + 296, + 259 + ], + "type": "inline_equation", + "content": "\\hat{\\mathbf{x}}_0^{(i)}(t)" + }, + { + "bbox": [ + 55, + 72, + 296, + 259 + ], + "type": "text", + "content": " for any given sample " + }, + { + "bbox": [ + 55, + 72, + 296, + 259 + ], + "type": "inline_equation", + "content": "\\mathbf{x}_t^{(i)}" + }, + { + "bbox": [ + 55, + 72, + 296, + 259 + ], + "type": "text", + "content": " via Equation (3), and check if it lies within a desired percentile-radius " + }, + { + "bbox": [ + 55, + 72, + 296, + 259 + ], + "type": "inline_equation", + "content": "\\rho" + }, + { + "bbox": [ + 55, + 72, + 296, + 259 + ], + "type": "text", + "content": " of the Gaussian " + }, + { + "bbox": [ + 55, + 72, + 296, + 259 + ], + "type": "inline_equation", + "content": "p_0" + }, + { + "bbox": [ + 55, + 72, + 296, + 259 + ], + "type": "text", + "content": ". Specifically, we define the time-dependent sample quality as" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 55, + 265, + 302, + 315 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 265, + 302, + 315 + ], + "spans": [ + { + "bbox": [ + 55, + 265, + 302, + 315 + ], + "type": "interline_equation", + "content": "q ^ {(i)} (t) = \\left\\{ \\begin{array}{l l} 1 & \\text {i f} \\| \\hat {\\mathbf {x}} _ {0} ^ {(i)} (t) \\| _ {2} ^ {2} \\leq \\rho^ {2} \\\\ \\max \\left(\\epsilon , e ^ {- \\left(\\| \\hat {\\mathbf {x}} _ {0} ^ {(i)} (t) \\| _ {2} ^ {2} - \\rho^ {2}\\right)}\\right) & \\text {o t h e r w i s e} \\end{array} \\right. \\tag {9}", + "image_path": "531c89d39d61ed740e6a754fbf80b4ce9dd0992d3399eab2fc8690cd3b48e456.jpg" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 55, + 316, + 296, + 339 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 316, + 296, + 339 + ], + "spans": [ + { + "bbox": [ + 55, + 316, + 296, + 339 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 55, + 316, + 296, + 339 + ], + "type": "inline_equation", + "content": "\\epsilon" + }, + { + "bbox": [ + 55, + 316, + 296, + 339 + ], + "type": "text", + "content": " is a 'minimum quality' we assign to prevent a zero determinant." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 55, + 346, + 272, + 358 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 346, + 272, + 358 + ], + "spans": [ + { + "bbox": [ + 55, + 346, + 272, + 358 + ], + "type": "text", + "content": "5.2. Coupled Ordinary Differential Equations" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 55, + 364, + 296, + 412 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 364, + 296, + 412 + ], + "spans": [ + { + "bbox": [ + 55, + 364, + 296, + 412 + ], + "type": "text", + "content": "At any timestep " + }, + { + "bbox": [ + 55, + 364, + 296, + 412 + ], + "type": "inline_equation", + "content": "t" + }, + { + "bbox": [ + 55, + 364, + 296, + 412 + ], + "type": "text", + "content": ", the measure of diversity in Equation (8) can be adopted to modify the flow of the " + }, + { + "bbox": [ + 55, + 364, + 296, + 412 + ], + "type": "inline_equation", + "content": "i" + }, + { + "bbox": [ + 55, + 364, + 296, + 412 + ], + "type": "text", + "content": "-th particle. We compute the gradient of the samples with respect to the diversity measure and use it to modify the ODE as follows:" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 55, + 418, + 301, + 447 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 418, + 301, + 447 + ], + "spans": [ + { + "bbox": [ + 55, + 418, + 301, + 447 + ], + "type": "interline_equation", + "content": "\\left. d \\mathbf {x} _ {t} ^ {(i)} = \\left[ v _ {\\theta} \\left(\\mathbf {x} _ {t} ^ {(i)}, t\\right) - \\gamma (t) \\nabla_ {\\mathbf {x} _ {t} ^ {(i)}} \\log \\mathcal {L} \\left(\\left\\{\\hat {\\mathbf {x}} _ {1} ^ {(1)}, \\dots , \\hat {\\mathbf {x}} _ {1} ^ {(k)} \\right\\}\\right) \\right] d t \\right. \\tag {10}", + "image_path": "d75f92e6a7934e4c23fee72de4e3815bfecb03ae339f230bf065f12b7ba085da.jpg" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 55, + 449, + 296, + 588 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 449, + 296, + 588 + ], + "spans": [ + { + "bbox": [ + 55, + 449, + 296, + 588 + ], + "type": "text", + "content": "Where " + }, + { + "bbox": [ + 55, + 449, + 296, + 588 + ], + "type": "inline_equation", + "content": "\\gamma (t)" + }, + { + "bbox": [ + 55, + 449, + 296, + 588 + ], + "type": "text", + "content": " is a time-varying scaling factor. Unlike the IID sampling scenario where we have " + }, + { + "bbox": [ + 55, + 449, + 296, + 588 + ], + "type": "inline_equation", + "content": "K" + }, + { + "bbox": [ + 55, + 449, + 296, + 588 + ], + "type": "text", + "content": " independent ODEs, Equation (10) corresponds to a system of coupled non-linear ordinary differential ordinary equations. To see this, first note that the estimate " + }, + { + "bbox": [ + 55, + 449, + 296, + 588 + ], + "type": "inline_equation", + "content": "\\hat{\\mathbf{x}}_1^{(i)}" + }, + { + "bbox": [ + 55, + 449, + 296, + 588 + ], + "type": "text", + "content": " depends on the current sample " + }, + { + "bbox": [ + 55, + 449, + 296, + 588 + ], + "type": "inline_equation", + "content": "\\mathbf{x}_t^{(i)}" + }, + { + "bbox": [ + 55, + 449, + 296, + 588 + ], + "type": "text", + "content": " i.e., " + }, + { + "bbox": [ + 55, + 449, + 296, + 588 + ], + "type": "inline_equation", + "content": "\\hat{\\mathbf{x}}_1^{(i)} = \\mathbf{x}_t^{(i)} + v_\\theta (\\mathbf{x}_t^{(i)},t)(1 - t)" + }, + { + "bbox": [ + 55, + 449, + 296, + 588 + ], + "type": "text", + "content": ". Second, the DPP log-likelihood " + }, + { + "bbox": [ + 55, + 449, + 296, + 588 + ], + "type": "inline_equation", + "content": "\\mathcal{LL}(\\{\\hat{\\mathbf{x}}_1^{(1)},\\hat{\\mathbf{x}}_1^{(2)},\\dots ,\\hat{\\mathbf{x}}_1^{(k)}\\})" + }, + { + "bbox": [ + 55, + 449, + 296, + 588 + ], + "type": "text", + "content": " induces a time-dependent coupling between the " + }, + { + "bbox": [ + 55, + 449, + 296, + 588 + ], + "type": "inline_equation", + "content": "K" + }, + { + "bbox": [ + 55, + 449, + 296, + 588 + ], + "type": "text", + "content": " trajectories of " + }, + { + "bbox": [ + 55, + 449, + 296, + 588 + ], + "type": "inline_equation", + "content": "\\mathbf{x}_t^{(i)},i = 1,\\ldots ,K" + }, + { + "bbox": [ + 55, + 449, + 296, + 588 + ], + "type": "text", + "content": " and seeks to diversify the target samples. Although higher-order ODE solvers [16] can be employed to solve the coupled ODEs, we use the standard Euler method." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 55, + 597, + 137, + 611 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 597, + 137, + 611 + ], + "spans": [ + { + "bbox": [ + 55, + 597, + 137, + 611 + ], + "type": "text", + "content": "6. Experiments" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 55, + 617, + 296, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 617, + 296, + 715 + ], + "spans": [ + { + "bbox": [ + 55, + 617, + 296, + 715 + ], + "type": "text", + "content": "We demonstrate the utility of DiverseFlow in flow-based generative models by considering three applications where sample diversity is naturally desirable: text-guided image generation with polysemous words, large-hole inpainting, and class-conditional image generation. We also analyze the effect of DiverseFlow on different flow matching formulations w.r.t. its ability to span diverse modes through a synthetically constructed 2D density example." + } + ] + } + ], + "index": 8 + }, + { + "type": "image", + "bbox": [ + 332, + 71, + 435, + 174 + ], + "blocks": [ + { + "bbox": [ + 332, + 71, + 435, + 174 + ], + "lines": [ + { + "bbox": [ + 332, + 71, + 435, + 174 + ], + "spans": [ + { + "bbox": [ + 332, + 71, + 435, + 174 + ], + "type": "image", + "image_path": "e4345bb398268dcbae533bc9541668f565a02ab355ea7220071bf32e4d02d1b0.jpg" + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 400, + 178, + 462, + 188 + ], + "lines": [ + { + "bbox": [ + 400, + 178, + 462, + 188 + ], + "spans": [ + { + "bbox": [ + 400, + 178, + 462, + 188 + ], + "type": "text", + "content": "(a) \"A buck\"" + } + ] + } + ], + "index": 11, + "angle": 0, + "type": "image_caption" + } + ], + "index": 9 + }, + { + "type": "image", + "bbox": [ + 435, + 72, + 537, + 174 + ], + "blocks": [ + { + "bbox": [ + 435, + 72, + 537, + 174 + ], + "lines": [ + { + "bbox": [ + 435, + 72, + 537, + 174 + ], + "spans": [ + { + "bbox": [ + 435, + 72, + 537, + 174 + ], + "type": "image", + "image_path": "9e9afdf4fc87a74b527a28a5bc998905bb61a5ff0e5622e3a90519c303a4110f.jpg" + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "image_body" + } + ], + "index": 10 + }, + { + "type": "image", + "bbox": [ + 331, + 190, + 435, + 293 + ], + "blocks": [ + { + "bbox": [ + 331, + 190, + 435, + 293 + ], + "lines": [ + { + "bbox": [ + 331, + 190, + 435, + 293 + ], + "spans": [ + { + "bbox": [ + 331, + 190, + 435, + 293 + ], + "type": "image", + "image_path": "bda090aeb79979dc0ae0d48a0cf8643cdf03d5452d7a8e956d066e809178dc1e.jpg" + } + ] + } + ], + "index": 12, + "angle": 0, + "type": "image_body" + } + ], + "index": 12 + }, + { + "type": "image", + "bbox": [ + 435, + 190, + 537, + 293 + ], + "blocks": [ + { + "bbox": [ + 435, + 190, + 537, + 293 + ], + "lines": [ + { + "bbox": [ + 435, + 190, + 537, + 293 + ], + "spans": [ + { + "bbox": [ + 435, + 190, + 537, + 293 + ], + "type": "image", + "image_path": "a2442f078401e273bd017cda6dd894ae35299eb39eafccf7b61b850cc6c6ed16.jpg" + } + ] + } + ], + "index": 13, + "angle": 0, + "type": "image_body" + } + ], + "index": 13 + }, + { + "type": "image", + "bbox": [ + 331, + 308, + 435, + 412 + ], + "blocks": [ + { + "bbox": [ + 376, + 297, + 488, + 307 + ], + "lines": [ + { + "bbox": [ + 376, + 297, + 488, + 307 + ], + "spans": [ + { + "bbox": [ + 376, + 297, + 488, + 307 + ], + "type": "text", + "content": "(b) \"A famous boxer\"" + } + ] + } + ], + "index": 14, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 331, + 308, + 435, + 412 + ], + "lines": [ + { + "bbox": [ + 331, + 308, + 435, + 412 + ], + "spans": [ + { + "bbox": [ + 331, + 308, + 435, + 412 + ], + "type": "image", + "image_path": "698e897027df0fac5fb324f1154f46536ad5a2654aeaffb8b2d5a96c3c806c28.jpg" + } + ] + } + ], + "index": 15, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 313, + 435, + 555, + 502 + ], + "lines": [ + { + "bbox": [ + 313, + 435, + 555, + 502 + ], + "spans": [ + { + "bbox": [ + 313, + 435, + 555, + 502 + ], + "type": "text", + "content": "Figure 3. For each prompt, the left image (red box) denotes standard IID sampling with classifier-free guidance, while the right image (blue box) shows the result after incorporating DiverseFlow. DiverseFlow finds more diverse sets given the same source points—clearly distinguishable as new semantic meanings in the case of prompts with multiple meanings." + } + ] + } + ], + "index": 18, + "angle": 0, + "type": "image_caption" + } + ], + "index": 15 + }, + { + "type": "image", + "bbox": [ + 435, + 309, + 537, + 412 + ], + "blocks": [ + { + "bbox": [ + 435, + 309, + 537, + 412 + ], + "lines": [ + { + "bbox": [ + 435, + 309, + 537, + 412 + ], + "spans": [ + { + "bbox": [ + 435, + 309, + 537, + 412 + ], + "type": "image", + "image_path": "ec6497676581e9b0f00046f6ac8be6c6d8b3bf00a6134c0b7fd1352fa75a9896.jpg" + } + ] + } + ], + "index": 16, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 371, + 415, + 499, + 428 + ], + "lines": [ + { + "bbox": [ + 371, + 415, + 499, + 428 + ], + "spans": [ + { + "bbox": [ + 371, + 415, + 499, + 428 + ], + "type": "text", + "content": "(c) \"Van Gogh painting\"" + } + ] + } + ], + "index": 17, + "angle": 0, + "type": "image_caption" + } + ], + "index": 16 + }, + { + "bbox": [ + 313, + 525, + 548, + 539 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 525, + 548, + 539 + ], + "spans": [ + { + "bbox": [ + 313, + 525, + 548, + 539 + ], + "type": "text", + "content": "6.1. Image Generation from Polysemous Prompts" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 313, + 544, + 554, + 640 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 544, + 554, + 640 + ], + "spans": [ + { + "bbox": [ + 313, + 544, + 554, + 640 + ], + "type": "text", + "content": "In text-to-image generation, the conditional data distribution corresponding to a text prompt may contain many variations, and it is a desirable objective to generate images that span those variations in a sample-efficient manner. As we have previously shown in Figure 1, the result desired by a user (a human boxer athlete) may not be the result generated by the model (a dog breed); if diverse results are obtained, it allows a user to obtain the desired results with fewer attempts." + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 313, + 642, + 555, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 642, + 555, + 713 + ], + "spans": [ + { + "bbox": [ + 313, + 642, + 555, + 713 + ], + "type": "text", + "content": "We pose a scenario where diverse sets are easily observable: when an open-ended text prompt is polysemous and carries multiple meanings, such as the examples we show in Figure 1 and Figure 3. In Figure 3(a), the prompt \"A buck\" may commonly refer to a male deer. However, it may also informally refer to a United States dollar. Using the" + } + ] + } + ], + "index": 21 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "text", + "content": "23307" + } + ] + } + ], + "index": 22 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 4 + }, + { + "para_blocks": [ + { + "bbox": [ + 55, + 72, + 296, + 252 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 72, + 296, + 252 + ], + "spans": [ + { + "bbox": [ + 55, + 72, + 296, + 252 + ], + "type": "text", + "content": "same four source points, which are deterministically mapped to four deer images by IID sampling. DiverseFlow finds a different set of samples—one that includes a dollar-like coin, albeit embossed with a deer head. We also observe minor differences between the two sets of images, such as changes in pose and background in the top-right and bottom-right deers. For Figure 3(c), while 'Van Gogh painting' is not quite a polysemous word, it can still have two interpretations: a painting painted by Van Gogh, or a painting of Van Gogh. The regular samples contain minimal diversity, as they include two sets of repeated paintings of Van Gogh. With DiverseFlow, not only can we get a set of four distinct paintings, but we also have a portrait of Van Gogh, which is one of the additional meanings of the prompt. We present additional qualitative samples in the supplementary material." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 55, + 276, + 209, + 289 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 276, + 209, + 289 + ], + "spans": [ + { + "bbox": [ + 55, + 276, + 209, + 289 + ], + "type": "text", + "content": "6.2. Diverse Inpainting on Faces" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 55, + 299, + 296, + 468 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 299, + 296, + 468 + ], + "spans": [ + { + "bbox": [ + 55, + 299, + 296, + 468 + ], + "type": "text", + "content": "Another inverse problem where diverse solutions are desirable is face inpainting, where we seek to inpaint the missing parts of the face with diverse plausible facial textures and structures. To demonstrate inpainting with FM models, we first incorporate Manifold Constrained Gradient (MCG) [5] in an off-the-shelf unconditional Rectified-Flow model. In addition to the manifold constraints, we employ determinantal gradient constraints to enhance diversity (further detailed in Algorithm 1). In Figure 5 (b), we observe that the inpainted faces of the four women have similar expressions (largely neutral). DiverseFlow improves the diversity of the set by yielding a highly different expression in the top-right image. In (d) and (e), we also observe changes in facial hair and expressions due to diversification." + } + ] + } + ], + "index": 2 + }, + { + "type": "image", + "bbox": [ + 77, + 498, + 116, + 578 + ], + "blocks": [ + { + "bbox": [ + 77, + 498, + 116, + 578 + ], + "lines": [ + { + "bbox": [ + 77, + 498, + 116, + 578 + ], + "spans": [ + { + "bbox": [ + 77, + 498, + 116, + 578 + ], + "type": "image", + "image_path": "cc6df4e4178aa44caacbc8ad8bc3b772697f71015d64a3422294bba61425c9db.jpg" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_body" + } + ], + "index": 3 + }, + { + "type": "image", + "bbox": [ + 117, + 498, + 195, + 578 + ], + "blocks": [ + { + "bbox": [ + 117, + 498, + 195, + 578 + ], + "lines": [ + { + "bbox": [ + 117, + 498, + 195, + 578 + ], + "spans": [ + { + "bbox": [ + 117, + 498, + 195, + 578 + ], + "type": "image", + "image_path": "55bc3c8aee37e194ee6eda46aaab6738322b8aec946c8e8995d1f8780cbf192d.jpg" + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_body" + } + ], + "index": 4 + }, + { + "type": "image", + "bbox": [ + 196, + 498, + 276, + 578 + ], + "blocks": [ + { + "bbox": [ + 196, + 498, + 276, + 578 + ], + "lines": [ + { + "bbox": [ + 196, + 498, + 276, + 578 + ], + "spans": [ + { + "bbox": [ + 196, + 498, + 276, + 578 + ], + "type": "image", + "image_path": "d175f3a6cae4e06745f9435defd5c1fd29139821dd91cbc6930017f91f00e08c.jpg" + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "image_body" + } + ], + "index": 5 + }, + { + "type": "image", + "bbox": [ + 77, + 580, + 116, + 661 + ], + "blocks": [ + { + "bbox": [ + 77, + 580, + 116, + 661 + ], + "lines": [ + { + "bbox": [ + 77, + 580, + 116, + 661 + ], + "spans": [ + { + "bbox": [ + 77, + 580, + 116, + 661 + ], + "type": "image", + "image_path": "9095c00848783c151ea1264db17b17e0909c02afb63eab4617b7b78faf43c5ca.jpg" + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 94, + 663, + 107, + 674 + ], + "lines": [ + { + "bbox": [ + 94, + 663, + 107, + 674 + ], + "spans": [ + { + "bbox": [ + 94, + 663, + 107, + 674 + ], + "type": "text", + "content": "(a)" + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 55, + 683, + 296, + 713 + ], + "lines": [ + { + "bbox": [ + 55, + 683, + 296, + 713 + ], + "spans": [ + { + "bbox": [ + 55, + 683, + 296, + 713 + ], + "type": "text", + "content": "Figure 5. Inpainting on CelebAHQ-256 × 256. (a) Dashed boxes show masked input (top) and ground truth (bottom) respectively. (b) Rectified-Flow [22] + MCG [5] (c) With DiverseFlow." + } + ] + } + ], + "index": 12, + "angle": 0, + "type": "image_caption" + } + ], + "index": 6 + }, + { + "type": "image", + "bbox": [ + 117, + 581, + 195, + 660 + ], + "blocks": [ + { + "bbox": [ + 117, + 581, + 195, + 660 + ], + "lines": [ + { + "bbox": [ + 117, + 581, + 195, + 660 + ], + "spans": [ + { + "bbox": [ + 117, + 581, + 195, + 660 + ], + "type": "image", + "image_path": "38338b8f2f901e450b3e33471eaff73b3d470a1848893290f142d5a556cac363.jpg" + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 148, + 663, + 161, + 674 + ], + "lines": [ + { + "bbox": [ + 148, + 663, + 161, + 674 + ], + "spans": [ + { + "bbox": [ + 148, + 663, + 161, + 674 + ], + "type": "text", + "content": "(b)" + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "image_caption" + } + ], + "index": 8 + }, + { + "type": "image", + "bbox": [ + 196, + 581, + 276, + 661 + ], + "blocks": [ + { + "bbox": [ + 196, + 581, + 276, + 661 + ], + "lines": [ + { + "bbox": [ + 196, + 581, + 276, + 661 + ], + "spans": [ + { + "bbox": [ + 196, + 581, + 276, + 661 + ], + "type": "image", + "image_path": "cab056f61389e95bd31f738823295add41b9a4c514fd136d339522eb02c8c685.jpg" + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 227, + 663, + 240, + 674 + ], + "lines": [ + { + "bbox": [ + 227, + 663, + 240, + 674 + ], + "spans": [ + { + "bbox": [ + 227, + 663, + 240, + 674 + ], + "type": "text", + "content": "(c)" + } + ] + } + ], + "index": 11, + "angle": 0, + "type": "image_caption" + } + ], + "index": 10 + }, + { + "type": "image", + "bbox": [ + 318, + 82, + 395, + 159 + ], + "blocks": [ + { + "bbox": [ + 318, + 82, + 395, + 159 + ], + "lines": [ + { + "bbox": [ + 318, + 82, + 395, + 159 + ], + "spans": [ + { + "bbox": [ + 318, + 82, + 395, + 159 + ], + "type": "image", + "image_path": "5c392b70064bc2f773bfde4fa5af59b45346919daa72c830dcba39a1e33b12f6.jpg" + } + ] + } + ], + "index": 13, + "angle": 0, + "type": "image_body" + } + ], + "index": 13 + }, + { + "type": "image", + "bbox": [ + 318, + 160, + 395, + 236 + ], + "blocks": [ + { + "bbox": [ + 318, + 160, + 395, + 236 + ], + "lines": [ + { + "bbox": [ + 318, + 160, + 395, + 236 + ], + "spans": [ + { + "bbox": [ + 318, + 160, + 395, + 236 + ], + "type": "image", + "image_path": "61ece2b7ba393195a2b1c3504086b09eed3f8aba7ba95ff3623b6618a75f53df.jpg" + } + ] + } + ], + "index": 14, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 341, + 242, + 376, + 252 + ], + "lines": [ + { + "bbox": [ + 341, + 242, + 376, + 252 + ], + "spans": [ + { + "bbox": [ + 341, + 242, + 376, + 252 + ], + "type": "text", + "content": "No CFG" + } + ] + } + ], + "index": 15, + "angle": 0, + "type": "image_caption" + } + ], + "index": 14 + }, + { + "type": "image", + "bbox": [ + 396, + 82, + 474, + 159 + ], + "blocks": [ + { + "bbox": [ + 396, + 82, + 474, + 159 + ], + "lines": [ + { + "bbox": [ + 396, + 82, + 474, + 159 + ], + "spans": [ + { + "bbox": [ + 396, + 82, + 474, + 159 + ], + "type": "image", + "image_path": "f9e2b90427a740b7b6e70ed79ea9601fda8929a5de19ce6aecf156a0e4023231.jpg" + } + ] + } + ], + "index": 16, + "angle": 0, + "type": "image_body" + } + ], + "index": 16 + }, + { + "type": "image", + "bbox": [ + 396, + 160, + 473, + 236 + ], + "blocks": [ + { + "bbox": [ + 396, + 160, + 473, + 236 + ], + "lines": [ + { + "bbox": [ + 396, + 160, + 473, + 236 + ], + "spans": [ + { + "bbox": [ + 396, + 160, + 473, + 236 + ], + "type": "image", + "image_path": "6270b52bbe42c86627c415e9e972ad66c929c2859609a3f7ae8c27e28352abf3.jpg" + } + ] + } + ], + "index": 17, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 411, + 242, + 454, + 252 + ], + "lines": [ + { + "bbox": [ + 411, + 242, + 454, + 252 + ], + "spans": [ + { + "bbox": [ + 411, + 242, + 454, + 252 + ], + "type": "text", + "content": "With CFG" + } + ] + } + ], + "index": 18, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 313, + 260, + 555, + 299 + ], + "lines": [ + { + "bbox": [ + 313, + 260, + 555, + 299 + ], + "spans": [ + { + "bbox": [ + 313, + 260, + 555, + 299 + ], + "type": "text", + "content": "Figure 6. Class-conditional ImageNet samples from LFM [7]. We show samples for two classes, (top row) 'Mushroom' (class 947) and (bottom row) 'Macaw' (class 88). (a, d) No CFG. (b, e) LFM with CFG. (d, f) LFM with CFG and DiverseFlow." + } + ] + } + ], + "index": 22, + "angle": 0, + "type": "image_caption" + } + ], + "index": 17 + }, + { + "type": "image", + "bbox": [ + 474, + 82, + 552, + 159 + ], + "blocks": [ + { + "bbox": [ + 474, + 82, + 552, + 159 + ], + "lines": [ + { + "bbox": [ + 474, + 82, + 552, + 159 + ], + "spans": [ + { + "bbox": [ + 474, + 82, + 552, + 159 + ], + "type": "image", + "image_path": "7567bfb3dcd0871aac8f8dcc319182c20bf53c26850f9b6a7c0fd94e9ce86865.jpg" + } + ] + } + ], + "index": 19, + "angle": 0, + "type": "image_body" + } + ], + "index": 19 + }, + { + "type": "image", + "bbox": [ + 474, + 160, + 552, + 236 + ], + "blocks": [ + { + "bbox": [ + 474, + 160, + 552, + 236 + ], + "lines": [ + { + "bbox": [ + 474, + 160, + 552, + 236 + ], + "spans": [ + { + "bbox": [ + 474, + 160, + 552, + 236 + ], + "type": "image", + "image_path": "b5e3b8c6e5a44aaf48e853449ff28850c8e09792fbfb8c6dfe871e184b337fdd.jpg" + } + ] + } + ], + "index": 20, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 489, + 242, + 541, + 252 + ], + "lines": [ + { + "bbox": [ + 489, + 242, + 541, + 252 + ], + "spans": [ + { + "bbox": [ + 489, + 242, + 541, + 252 + ], + "type": "text", + "content": "DiverseFlow" + } + ] + } + ], + "index": 21, + "angle": 0, + "type": "image_caption" + } + ], + "index": 20 + }, + { + "bbox": [ + 313, + 313, + 537, + 327 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 313, + 537, + 327 + ], + "spans": [ + { + "bbox": [ + 313, + 313, + 537, + 327 + ], + "type": "text", + "content": "6.3. Diverse Class-Conditional Image Synthesis" + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 313, + 331, + 555, + 437 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 331, + 555, + 437 + ], + "spans": [ + { + "bbox": [ + 313, + 331, + 555, + 437 + ], + "type": "text", + "content": "Suppose we can access a class-conditioned flow matching (FM) model trained on an unknown image dataset. To explore the unobservable true dataset, we may use a set of class-conditional samples from the FM model. We adopt a latent flow matching (LFM) model [7], trained to generate " + }, + { + "bbox": [ + 313, + 331, + 555, + 437 + ], + "type": "inline_equation", + "content": "256 \\times 256" + }, + { + "bbox": [ + 313, + 331, + 555, + 437 + ], + "type": "text", + "content": " resolution images from the ImageNet [8] dataset. Much like latent diffusion, LFM employs classifier-free guidance to create high-quality samples. However, this naturally poses a cost to diversity, as we show in Figure 6." + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 313, + 439, + 556, + 631 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 439, + 556, + 631 + ], + "spans": [ + { + "bbox": [ + 313, + 439, + 556, + 631 + ], + "type": "text", + "content": "By incorporating DiverseFlow, we can maintain the high quality of the samples and simultaneously explore more modes in the dataset. In Figure 6, we demonstrate two ImageNet classes that may have diversity: 'Mushroom' and 'Macaw.' For mushrooms, we observe that LFM primarily generates two species of mushrooms. However, by applying DiverseFlow, we successfully find a new species within our limited set: an Amanita muscaria, also known as the fly agaric—easily distinguishable by the white spots on its red cap. In another example, we see that while LFM generates four scarlet macaws, using the same source samples, DiverseFlow helps us find a different blue and yellow macaw. We further evaluate the precision and recall [18] for ImageNet-256 synthesis in Figure 7, showing that DiverseFlow improves the recall while retaining similar precision, suggesting improved diversity." + } + ] + } + ], + "index": 25 + }, + { + "bbox": [ + 313, + 635, + 555, + 648 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 635, + 555, + 648 + ], + "spans": [ + { + "bbox": [ + 313, + 635, + 555, + 648 + ], + "type": "text", + "content": "6.4. DiverseFlow Across Various FM Formulations" + } + ] + } + ], + "index": 26 + }, + { + "bbox": [ + 313, + 654, + 556, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 654, + 556, + 715 + ], + "spans": [ + { + "bbox": [ + 313, + 654, + 556, + 715 + ], + "type": "text", + "content": "We ask the question: does DiverseFlow remain consistent across different formulations of flow matching? We adopt the same toy bivariate mixture of Gaussian distributions as in the example shown in Figure 2, across four different FM formulations: (i) Conditional Flow Matching (CFM) [20]," + } + ] + } + ], + "index": 27 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "text", + "content": "23308" + } + ] + } + ], + "index": 28 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 5 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 56, + 68, + 553, + 204 + ], + "blocks": [ + { + "bbox": [ + 56, + 68, + 553, + 204 + ], + "lines": [ + { + "bbox": [ + 56, + 68, + 553, + 204 + ], + "spans": [ + { + "bbox": [ + 56, + 68, + 553, + 204 + ], + "type": "image", + "image_path": "fef732875e8d72d562ee357e5aad903d9199914893ad33400c074e21c5ea3492.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 55, + 212, + 555, + 236 + ], + "lines": [ + { + "bbox": [ + 55, + 212, + 555, + 236 + ], + "spans": [ + { + "bbox": [ + 55, + 212, + 555, + 236 + ], + "type": "text", + "content": "Figure 4. Comparing different FM formulations in terms of the number of modes spanned by IID sampling versus with DiverseFlow. More details about the experiment are provided in the supplementary." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "type": "image", + "bbox": [ + 56, + 253, + 168, + 348 + ], + "blocks": [ + { + "bbox": [ + 56, + 253, + 168, + 348 + ], + "lines": [ + { + "bbox": [ + 56, + 253, + 168, + 348 + ], + "spans": [ + { + "bbox": [ + 56, + 253, + 168, + 348 + ], + "type": "image", + "image_path": "0bf640e7bba7bb1cd27b421928f5d85dbb886e09c20e0128f8be1202571b0516.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 55, + 357, + 168, + 378 + ], + "lines": [ + { + "bbox": [ + 55, + 357, + 168, + 378 + ], + "spans": [ + { + "bbox": [ + 55, + 357, + 168, + 378 + ], + "type": "text", + "content": "Figure 7. Precision vs. recall on ImageNet-256." + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_caption" + } + ], + "index": 2 + }, + { + "type": "table", + "bbox": [ + 179, + 253, + 293, + 348 + ], + "blocks": [ + { + "bbox": [ + 179, + 253, + 293, + 348 + ], + "lines": [ + { + "bbox": [ + 179, + 253, + 293, + 348 + ], + "spans": [ + { + "bbox": [ + 179, + 253, + 293, + 348 + ], + "type": "table", + "html": "
MethodCFGP (↑)R (↑)
LFM10.620.52
LFM1.50.690.44
LFM20.770.41
LFM40.690.26
LFM + DF10.620.52
LFM+DF1.50.690.47
LFM+DF20.760.46
LFM+DF40.700.38
", + "image_path": "d6483d0ec3b74856c40680d15c5693639bc11fc0763ae97d71426ac85fe8ef63.jpg" + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "table_body" + } + ], + "index": 4 + }, + { + "bbox": [ + 176, + 357, + 295, + 376 + ], + "lines": [ + { + "bbox": [ + 176, + 357, + 295, + 376 + ], + "spans": [ + { + "bbox": [ + 176, + 357, + 295, + 376 + ], + "type": "text", + "content": "Table 1. Precision (P) and recall (R) score of DiverseFlow and baselines." + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "text" + }, + { + "bbox": [ + 54, + 399, + 297, + 568 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 54, + 399, + 297, + 568 + ], + "spans": [ + { + "bbox": [ + 54, + 399, + 297, + 568 + ], + "type": "text", + "content": "(ii) Mini-batch Optimal Transport CFM (MB-OT) [30], (iii) Schrödinger Bridge CFM (SB-CFM) [31], and (iv) Stochastic Interpolants (SI-CFM) [2]. We then perform a numerical experiment to quantify the average number of modes discovered by each FM variant with increasing the sampling budget " + }, + { + "bbox": [ + 54, + 399, + 297, + 568 + ], + "type": "inline_equation", + "content": "K" + }, + { + "bbox": [ + 54, + 399, + 297, + 568 + ], + "type": "text", + "content": ". Figure 4 reports the results. For a maximum sampling budget of 10, MB-OT discovers only 5.64 modes on average, which is expected since the dataset contains 6 high-probability modes. By incorporating DiverseFlow, we can find 7.11 modes on average. We observe that MB-OT and SB-CFM benefit most from DiverseFlow. We hypothesize this is because MB-OT and SB-CFM are formulated with the notion of optimal paths, which results in a more accurate estimation of " + }, + { + "bbox": [ + 54, + 399, + 297, + 568 + ], + "type": "inline_equation", + "content": "\\hat{x}_1" + }, + { + "bbox": [ + 54, + 399, + 297, + 568 + ], + "type": "text", + "content": " and thus a better diversity gradient." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 55, + 576, + 239, + 588 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 576, + 239, + 588 + ], + "spans": [ + { + "bbox": [ + 55, + 576, + 239, + 588 + ], + "type": "text", + "content": "6.5. Comparison To Particle Guidance" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 54, + 594, + 297, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 54, + 594, + 297, + 715 + ], + "spans": [ + { + "bbox": [ + 54, + 594, + 297, + 715 + ], + "type": "text", + "content": "(i) Preventing Training Data Copy: Previously, Corso et al. [6] demonstrate that their method can alleviate Stable Diffusion's training data regurgitation problem [27] to some extent. In Figure 8, we demonstrate similar capabilities; in (b), the top left and bottom right examples are copies of the training data. Subsequently, (c) and (d) find a new example for the top-left sample. (ii) Diversity vs Quality: We also compare the diversity versus quality of DiverseFlow against Particle Guidance in Figure 10 over 30 polysemous prompts repeated over 10 random seeds. While varying the" + } + ] + } + ], + "index": 8 + }, + { + "type": "image", + "bbox": [ + 317, + 253, + 434, + 371 + ], + "blocks": [ + { + "bbox": [ + 317, + 253, + 434, + 371 + ], + "lines": [ + { + "bbox": [ + 317, + 253, + 434, + 371 + ], + "spans": [ + { + "bbox": [ + 317, + 253, + 434, + 371 + ], + "type": "image", + "image_path": "a5ba5bf8f7fb9dd0635fa4dbf85f6f362ac9ea799e39bdb9d860577664fe5dd6.jpg" + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 347, + 373, + 403, + 382 + ], + "lines": [ + { + "bbox": [ + 347, + 373, + 403, + 382 + ], + "spans": [ + { + "bbox": [ + 347, + 373, + 403, + 382 + ], + "type": "text", + "content": "(a) Training Data" + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "image_caption" + } + ], + "index": 9 + }, + { + "type": "image", + "bbox": [ + 434, + 253, + 553, + 372 + ], + "blocks": [ + { + "bbox": [ + 434, + 253, + 553, + 372 + ], + "lines": [ + { + "bbox": [ + 434, + 253, + 553, + 372 + ], + "spans": [ + { + "bbox": [ + 434, + 253, + 553, + 372 + ], + "type": "image", + "image_path": "ce25bfa9b4e5ef7fa09c122f1c27b16773f3114a201b8e04e8527e4f2359ce14.jpg" + } + ] + } + ], + "index": 11, + "angle": 0, + "type": "image_body" + } + ], + "index": 11 + }, + { + "type": "image", + "bbox": [ + 317, + 383, + 434, + 501 + ], + "blocks": [ + { + "bbox": [ + 317, + 383, + 434, + 501 + ], + "lines": [ + { + "bbox": [ + 317, + 383, + 434, + 501 + ], + "spans": [ + { + "bbox": [ + 317, + 383, + 434, + 501 + ], + "type": "image", + "image_path": "d7863ca0514b5e16bb6e714f257336c5bbe02efcacc86a79b500fa53919ddfe6.jpg" + } + ] + } + ], + "index": 13, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 335, + 502, + 416, + 512 + ], + "lines": [ + { + "bbox": [ + 335, + 502, + 416, + 512 + ], + "spans": [ + { + "bbox": [ + 335, + 502, + 416, + 512 + ], + "type": "text", + "content": "(c) Particle Guidance [6]" + } + ] + } + ], + "index": 14, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 313, + 521, + 555, + 565 + ], + "lines": [ + { + "bbox": [ + 313, + 521, + 555, + 565 + ], + "spans": [ + { + "bbox": [ + 313, + 521, + 555, + 565 + ], + "type": "text", + "content": "Figure 8. For the prompt \"VAN GOGH CAFE TERASSE copy.jpg\", we show (a) the training data, (b) IID sampling with two copies in top-left and bottom-right, (c) Particle Guidance [6], and (d) DiverseFlow." + } + ] + } + ], + "index": 17, + "angle": 0, + "type": "image_caption" + } + ], + "index": 13 + }, + { + "type": "image", + "bbox": [ + 434, + 383, + 553, + 501 + ], + "blocks": [ + { + "bbox": [ + 465, + 373, + 522, + 383 + ], + "lines": [ + { + "bbox": [ + 465, + 373, + 522, + 383 + ], + "spans": [ + { + "bbox": [ + 465, + 373, + 522, + 383 + ], + "type": "text", + "content": "(b) IID Sampling" + } + ] + } + ], + "index": 12, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 434, + 383, + 553, + 501 + ], + "lines": [ + { + "bbox": [ + 434, + 383, + 553, + 501 + ], + "spans": [ + { + "bbox": [ + 434, + 383, + 553, + 501 + ], + "type": "image", + "image_path": "5ef79318d3c8a81eb0c24b6848c9a23249e3ad3a237b38e3e86406a4c31fcf20.jpg" + } + ] + } + ], + "index": 15, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 466, + 502, + 520, + 511 + ], + "lines": [ + { + "bbox": [ + 466, + 502, + 520, + 511 + ], + "spans": [ + { + "bbox": [ + 466, + 502, + 520, + 511 + ], + "type": "text", + "content": "(d) DiverseFlow" + } + ] + } + ], + "index": 16, + "angle": 0, + "type": "image_caption" + } + ], + "index": 15 + }, + { + "bbox": [ + 313, + 594, + 556, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 594, + 556, + 715 + ], + "spans": [ + { + "bbox": [ + 313, + 594, + 556, + 715 + ], + "type": "text", + "content": "strength of DiverseFlow to improve diversity, we also vary the values of the classifier-free guidance strength from 7.5 to 10 to boost quality. Quality is measured by Aesthetic Score (higher is better) [25], and diversity is measured by average pairwise similarity of a set (lower is better) [6]. We observe that though DiverseFlow obtains better diversity at similar quality to IID sampling, the aesthetic score of Particle Guidance is unusually low. In Figure 9 we highlight the cause: for the challenging task of polysemous text-to-image generation, the images generated from Particle Guidance (a) suffers" + } + ] + } + ], + "index": 18 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "type": "text", + "content": "23309" + } + ] + } + ], + "index": 19 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 6 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 56, + 70, + 294, + 425 + ], + "blocks": [ + { + "bbox": [ + 56, + 70, + 294, + 425 + ], + "lines": [ + { + "bbox": [ + 56, + 70, + 294, + 425 + ], + "spans": [ + { + "bbox": [ + 56, + 70, + 294, + 425 + ], + "type": "image", + "image_path": "65c4181a6ba7953eba996ac49880980cf74dca076240aad473bbdd4382ae2d76.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 55, + 434, + 296, + 464 + ], + "lines": [ + { + "bbox": [ + 55, + 434, + 296, + 464 + ], + "spans": [ + { + "bbox": [ + 55, + 434, + 296, + 464 + ], + "type": "text", + "content": "Figure 9. Particle Guidance (middle rows) can suffer from artefacts for polysemous prompts. DiverseFlow retains quality while achieving diversity: for \"A mint\", a minted coin with a mint leaf was discovered." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "type": "image", + "bbox": [ + 70, + 475, + 281, + 632 + ], + "blocks": [ + { + "bbox": [ + 70, + 475, + 281, + 632 + ], + "lines": [ + { + "bbox": [ + 70, + 475, + 281, + 632 + ], + "spans": [ + { + "bbox": [ + 70, + 475, + 281, + 632 + ], + "type": "image", + "image_path": "185e2a6bc82bf7990294139d87b3f16d96b9684fcd145794d903258555a2a688.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 83, + 644, + 268, + 656 + ], + "lines": [ + { + "bbox": [ + 83, + 644, + 268, + 656 + ], + "spans": [ + { + "bbox": [ + 83, + 644, + 268, + 656 + ], + "type": "text", + "content": "Figure 10. In-batch similarity " + }, + { + "bbox": [ + 83, + 644, + 268, + 656 + ], + "type": "inline_equation", + "content": "(\\downarrow)" + }, + { + "bbox": [ + 83, + 644, + 268, + 656 + ], + "type": "text", + "content": " vs aesthetic score " + }, + { + "bbox": [ + 83, + 644, + 268, + 656 + ], + "type": "inline_equation", + "content": "(\\uparrow)" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_caption" + } + ], + "index": 2 + }, + { + "bbox": [ + 55, + 677, + 296, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 677, + 296, + 715 + ], + "spans": [ + { + "bbox": [ + 55, + 677, + 296, + 715 + ], + "type": "text", + "content": "from image artefacts, which achieves a high diversity score but a poor quality score. Additional details about the experiment, including the set of 30 prompts, are provided in the" + } + ] + } + ], + "index": 4 + }, + { + "type": "image", + "bbox": [ + 316, + 70, + 553, + 315 + ], + "blocks": [ + { + "bbox": [ + 316, + 70, + 553, + 315 + ], + "lines": [ + { + "bbox": [ + 316, + 70, + 553, + 315 + ], + "spans": [ + { + "bbox": [ + 316, + 70, + 553, + 315 + ], + "type": "image", + "image_path": "57d72ea72048db47ef67e9e96c7c2a74edc1aa9a826a4ec5e60fe0629c570eb7.jpg" + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 313, + 326, + 555, + 346 + ], + "lines": [ + { + "bbox": [ + 313, + 326, + 555, + 346 + ], + "spans": [ + { + "bbox": [ + 313, + 326, + 555, + 346 + ], + "type": "text", + "content": "Figure 11. Average modes discovered over 100 trials, for varying noise scales and strengths of DiverseFlow and ParticleGuidance." + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "image_caption" + } + ], + "index": 5 + }, + { + "bbox": [ + 313, + 369, + 556, + 548 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 369, + 556, + 548 + ], + "spans": [ + { + "bbox": [ + 313, + 369, + 556, + 548 + ], + "type": "text", + "content": "supplementary. (iii) Mode Discovery: In high dimensional data, the number of modes, " + }, + { + "bbox": [ + 313, + 369, + 556, + 548 + ], + "type": "inline_equation", + "content": "N" + }, + { + "bbox": [ + 313, + 369, + 556, + 548 + ], + "type": "text", + "content": ", is significantly greater than the budget " + }, + { + "bbox": [ + 313, + 369, + 556, + 548 + ], + "type": "inline_equation", + "content": "K" + }, + { + "bbox": [ + 313, + 369, + 556, + 548 + ], + "type": "text", + "content": ", so finding unique modes is still highly probable. To better highlight the differences between DiverseFlow and Particle Guidance, we adopt a simpler toy experiment: finding modes in a symmetric uniform Mixture of Gaussian distribution. Corso et al. [6] provides the result that IID sampling with budget " + }, + { + "bbox": [ + 313, + 369, + 556, + 548 + ], + "type": "inline_equation", + "content": "K = 10" + }, + { + "bbox": [ + 313, + 369, + 556, + 548 + ], + "type": "text", + "content": " discovers about 6.5 modes on average, while Particle Guidance with a Euclidean kernel discovers almost 9 out of 10 modes. We verify this result in Figure 11, finding that Particle Guidance discovers up to 8.8 modes (averaged over 100 trials). However, by using DiverseFlow, it is possible to discover all 10 modes on average, showing that our approach has a stronger diversification effect." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 314, + 560, + 388, + 573 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 314, + 560, + 388, + 573 + ], + "spans": [ + { + "bbox": [ + 314, + 560, + 388, + 573 + ], + "type": "text", + "content": "7. Conclusion" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 313, + 582, + 556, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 582, + 556, + 713 + ], + "spans": [ + { + "bbox": [ + 313, + 582, + 556, + 713 + ], + "type": "text", + "content": "In this paper, we present DiverseFlow, a way to enforce diversity among a limited set of samples generated from a flow by coupling them through a determinantal point process. We demonstrate multiple applications where our method can be useful. In essence, though flow models admit a deterministic mapping from source to sample, DiverseFlow allows us to modify this mapping at inference-time and obtain samples with improved diversity. Diversity enhances the utility of the underlying generative flow in the case of ill-posed problems, where many solutions exist and it is desirable to find a multitude of such solutions." + } + ] + } + ], + "index": 9 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "type": "text", + "content": "23310" + } + ] + } + ], + "index": 10 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 7 + }, + { + "para_blocks": [ + { + "bbox": [ + 56, + 71, + 115, + 83 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 71, + 115, + 83 + ], + "spans": [ + { + "bbox": [ + 56, + 71, + 115, + 83 + ], + "type": "text", + "content": "References" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 57, + 91, + 297, + 713 + ], + "type": "list", + "angle": 0, + "index": 14, + "blocks": [ + { + "bbox": [ + 61, + 91, + 297, + 146 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 61, + 91, + 297, + 146 + ], + "spans": [ + { + "bbox": [ + 61, + 91, + 297, + 146 + ], + "type": "text", + "content": "[1] Charu C Aggarwal, Alexander Hinneburg, and Daniel A Keim. On the surprising behavior of distance metrics in high dimensional space. In Database Theory—ICDT 2001: 8th International Conference London, UK, January 4–6, 2001 Proceedings 8, pages 420–434. Springer, 2001. 4" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 61, + 148, + 296, + 190 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 61, + 148, + 296, + 190 + ], + "spans": [ + { + "bbox": [ + 61, + 148, + 296, + 190 + ], + "type": "text", + "content": "[2] Michael S Albergo, Nicholas M Boffi, and Eric Vanden-Eijnden. Stochastic interpolants: A unifying framework for flows and diffusions. arXiv preprint arXiv:2303.08797, 2023. 2, 7, 4" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 62, + 194, + 296, + 258 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 194, + 296, + 258 + ], + "spans": [ + { + "bbox": [ + 62, + 194, + 296, + 258 + ], + "type": "text", + "content": "[3] Dhruv Batra, Payman Yadollahpour, Abner Guzman-Rivera, and Gregory Shakhnarovich. Diverse m-best solutions in markov random fields. In Computer Vision-ECCV 2012: 12th European Conference on Computer Vision, Florence, Italy, October 7-13, 2012, Proceedings, Part V 12, pages 1-16. Springer, 2012. 3" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 62, + 261, + 296, + 303 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 261, + 296, + 303 + ], + "spans": [ + { + "bbox": [ + 62, + 261, + 296, + 303 + ], + "type": "text", + "content": "[4] Alexei Borodin and Grigori Olshanski. Distributions on partitions, point processes, and the hypergeometric kernel. Communications in Mathematical Physics, 211:335-358, 2000. 3" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 62, + 306, + 296, + 350 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 306, + 296, + 350 + ], + "spans": [ + { + "bbox": [ + 62, + 306, + 296, + 350 + ], + "type": "text", + "content": "[5] Hyungjin Chung, Byeongsu Sim, Dohoon Ryu, and Jong Chul Ye. Improving diffusion models for inverse problems using manifold constraints. Advances in Neural Information Processing Systems, 35:25683-25696, 2022. 6, 1" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 62, + 351, + 296, + 395 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 351, + 296, + 395 + ], + "spans": [ + { + "bbox": [ + 62, + 351, + 296, + 395 + ], + "type": "text", + "content": "[6] Gabriele Corso, Yilun Xu, Valentin De Bortoli, Regina Barzilay, and Tommi Jaakkola. Particle guidance: non-iid diverse sampling with diffusion models. International Conference on Learning Representations, 2023. 3, 7, 8, 1" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 62, + 396, + 296, + 429 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 396, + 296, + 429 + ], + "spans": [ + { + "bbox": [ + 62, + 396, + 296, + 429 + ], + "type": "text", + "content": "[7] Quan Dao, Hao Phung, Binh Nguyen, and Anh Tran. Flow matching in latent space. arXiv preprint arXiv:2307.08698, 2023. 6, 3" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 62, + 431, + 296, + 475 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 431, + 296, + 475 + ], + "spans": [ + { + "bbox": [ + 62, + 431, + 296, + 475 + ], + "type": "text", + "content": "[8] Jia Deng, Wei Dong, Richard Socher, Li-Jia Li, Kai Li, and Li Fei-Fei. Imagenet: A large-scale hierarchical image database. In 2009 IEEE conference on computer vision and pattern recognition, pages 248-255. IEEE, 2009. 6" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 62, + 476, + 296, + 542 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 476, + 296, + 542 + ], + "spans": [ + { + "bbox": [ + 62, + 476, + 296, + 542 + ], + "type": "text", + "content": "[9] Alexey Dosovitskiy, Lucas Beyer, Alexander Kolesnikov, Dirk Weissenborn, Xiaohua Zhai, Thomas Unterthiner, Mostafa Dehghani, Matthias Minderer, Georg Heigold, Sylvain Gelly, et al. An image is worth 16x16 words: Transformers for image recognition at scale. International Conference on Learning Representations, 2021. 4" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 57, + 544, + 296, + 609 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 544, + 296, + 609 + ], + "spans": [ + { + "bbox": [ + 57, + 544, + 296, + 609 + ], + "type": "text", + "content": "[10] Patrick Esser, Sumith Kulal, Andreas Blattmann, Rahim Entezari, Jonas Müller, Harry Saini, Yam Levi, Dominik Lorenz, Axel Sauer, Frederic Boesel, et al. Scaling rectified flow transformers for high-resolution image synthesis. In International Conference on Machine Learning, pages 12606-12633. PMLR, 2024. 2, 1" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 57, + 611, + 294, + 634 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 611, + 294, + 634 + ], + "spans": [ + { + "bbox": [ + 57, + 611, + 294, + 634 + ], + "type": "text", + "content": "[11] Jonathan Ho and Tim Salimans. Classifier-free diffusion guidance. arXiv preprint arXiv:2207.12598, 2022. 1" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 57, + 635, + 296, + 667 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 635, + 296, + 667 + ], + "spans": [ + { + "bbox": [ + 57, + 635, + 296, + 667 + ], + "type": "text", + "content": "[12] Jonathan Ho, Ajay Jain, and Pieter Abbeel. Denoising diffusion probabilistic models. Advances in neural information processing systems, 33:6840-6851, 2020. 1, 2" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 57, + 670, + 294, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 670, + 294, + 713 + ], + "spans": [ + { + "bbox": [ + 57, + 670, + 294, + 713 + ], + "type": "text", + "content": "[13] Noboru Isobe, Masanori Koyama, Jinzhe Zhang, Kohei Hayashi, and Kenji Fukumizu. Extended flow matching: a method of conditional generation with generalized continuity equation. arXiv preprint arXiv:2402.18839, 2024. 5" + } + ] + } + ], + "index": 13 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 316, + 73, + 555, + 713 + ], + "type": "list", + "angle": 0, + "index": 30, + "blocks": [ + { + "bbox": [ + 316, + 73, + 555, + 127 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 73, + 555, + 127 + ], + "spans": [ + { + "bbox": [ + 316, + 73, + 555, + 127 + ], + "type": "text", + "content": "[14] Moksh Jain, Sharath Chandra Raparthy, Alex Hernández-García, Jarrid Rector-Brooks, Yoshua Bengio, Santiago Miret, and Emmanuel Bengio. Multi-objective gflownets. In International Conference on Machine Learning, pages 14631-14653. PMLR, 2023. 3" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 317, + 129, + 555, + 171 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 129, + 555, + 171 + ], + "spans": [ + { + "bbox": [ + 317, + 129, + 555, + 171 + ], + "type": "text", + "content": "[15] Tero Karras, Timo Aila, Samuli Laine, and Jaakko Lehtinen. Progressive growing of gans for improved quality, stability, and variation. International Conference on Learning Representations, 2018. 1" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 317, + 173, + 555, + 215 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 173, + 555, + 215 + ], + "spans": [ + { + "bbox": [ + 317, + 173, + 555, + 215 + ], + "type": "text", + "content": "[16] Tero Karras, Miika Aittala, Timo Aila, and Samuli Laine. Elucidating the design space of diffusion-based generative models. Advances in Neural Information Processing Systems, 2022. 1, 5" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 316, + 217, + 555, + 249 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 217, + 555, + 249 + ], + "spans": [ + { + "bbox": [ + 316, + 217, + 555, + 249 + ], + "type": "text", + "content": "[17] Alex Kulesza, Ben Taskar, et al. Determinantal point processes for machine learning. Foundations and Trends® in Machine Learning, 5(2-3):123-286, 2012. 2, 3" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 317, + 251, + 554, + 294 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 251, + 554, + 294 + ], + "spans": [ + { + "bbox": [ + 317, + 251, + 554, + 294 + ], + "type": "text", + "content": "[18] Tuomas Kynkänniemi, Tero Karras, Samuli Laine, Jaakko Lehtinen, and Timo Aila. Improved precision and recall metric for assessing generative models. Advances in neural information processing systems, 32, 2019. 6" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 317, + 295, + 555, + 338 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 295, + 555, + 338 + ], + "spans": [ + { + "bbox": [ + 317, + 295, + 555, + 338 + ], + "type": "text", + "content": "[19] Yiheng Li, Heyang Jiang, Akio Kodaira, Masayoshi Tomizuka, Kurt Keutzer, and Chenfeng Xu. Immiscible diffusion: Accelerating diffusion training with noise assignment. Advances in neural information processing systems, 2024. 5" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 317, + 338, + 555, + 370 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 338, + 555, + 370 + ], + "spans": [ + { + "bbox": [ + 317, + 338, + 555, + 370 + ], + "type": "text", + "content": "[20] Yaron Lipman, Ricky TQ Chen, Heli Ben-Hamu, Maximilian Nickel, and Matt Le. Flow matching for generative modeling. arXiv preprint arXiv:2210.02747, 2022. 1, 2, 3, 6" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 316, + 372, + 554, + 404 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 372, + 554, + 404 + ], + "spans": [ + { + "bbox": [ + 316, + 372, + 554, + 404 + ], + "type": "text", + "content": "[21] Qiang Liu and Dilin Wang. Stein variational gradient descent: A general purpose bayesian inference algorithm. Advances in neural information processing systems, 29, 2016. 3" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 316, + 405, + 554, + 437 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 405, + 554, + 437 + ], + "spans": [ + { + "bbox": [ + 316, + 405, + 554, + 437 + ], + "type": "text", + "content": "[22] Xingchao Liu, Chengyue Gong, and Qiang Liu. Flow straight and fast: Learning to generate and transfer data with rectified flow. arXiv preprint arXiv:2209.03003, 2022. 1, 2, 3, 6" + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 316, + 438, + 554, + 469 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 438, + 554, + 469 + ], + "spans": [ + { + "bbox": [ + 316, + 438, + 554, + 469 + ], + "type": "text", + "content": "[23] Odile Macchi. The coincidence approach to stochastic point processes. Advances in Applied Probability, 7(1):83-122, 1975. 2, 3" + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 317, + 472, + 555, + 525 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 472, + 555, + 525 + ], + "spans": [ + { + "bbox": [ + 317, + 472, + 555, + 525 + ], + "type": "text", + "content": "[24] Robin Rombach, Andreas Blattmann, Dominik Lorenz, Patrick Esser, and Björn Ommer. High-resolution image synthesis with latent diffusion models. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 10684-10695, 2022. 1" + } + ] + } + ], + "index": 25 + }, + { + "bbox": [ + 317, + 526, + 555, + 570 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 526, + 555, + 570 + ], + "spans": [ + { + "bbox": [ + 317, + 526, + 555, + 570 + ], + "type": "text", + "content": "[25] Christoph Schuhmann and LAION. Improved aesthetic predictor: Clip+mlp aesthetic score predictor, 2022. https://github.com/christophschuhmann/improved-aesthetic-predictor. 7" + } + ] + } + ], + "index": 26 + }, + { + "bbox": [ + 317, + 571, + 555, + 624 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 571, + 555, + 624 + ], + "spans": [ + { + "bbox": [ + 317, + 571, + 555, + 624 + ], + "type": "text", + "content": "[26] Jascha Sohl-Dickstein, Eric Weiss, Niru Maheswaranathan, and Surya Ganguli. Deep unsupervised learning using nonequilibrium thermodynamics. In International Conference on Machine Learning, pages 2256-2265. PMLR, 2015. 1" + } + ] + } + ], + "index": 27 + }, + { + "bbox": [ + 317, + 625, + 555, + 680 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 625, + 555, + 680 + ], + "spans": [ + { + "bbox": [ + 317, + 625, + 555, + 680 + ], + "type": "text", + "content": "[27] Gowthami Somepalli, Vasu Singla, Micah Goldblum, Jonas Geiping, and Tom Goldstein. Diffusion art or digital forgery? investigating data replication in diffusion models. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 6048-6058, 2023. 7" + } + ] + } + ], + "index": 28 + }, + { + "bbox": [ + 316, + 681, + 555, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 681, + 555, + 713 + ], + "spans": [ + { + "bbox": [ + 316, + 681, + 555, + 713 + ], + "type": "text", + "content": "[28] Jiaming Song, Chenlin Meng, and Stefano Ermon. Denoising diffusion implicit models. International Conference on Learning Representations, 2021. 1" + } + ] + } + ], + "index": 29 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "text", + "content": "23311" + } + ] + } + ], + "index": 31 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 8 + }, + { + "para_blocks": [ + { + "bbox": [ + 56, + 73, + 297, + 441 + ], + "type": "list", + "angle": 0, + "index": 7, + "blocks": [ + { + "bbox": [ + 56, + 73, + 297, + 126 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 73, + 297, + 126 + ], + "spans": [ + { + "bbox": [ + 56, + 73, + 297, + 126 + ], + "type": "text", + "content": "[29] Yang Song, Jascha Sohl-Dickstein, Diederik P Kingma, Abhishek Kumar, Stefano Ermon, and Ben Poole. Score-based generative modeling through stochastic differential equations. International Conference on Learning Representations, 2021. 2" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 56, + 129, + 295, + 183 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 129, + 295, + 183 + ], + "spans": [ + { + "bbox": [ + 56, + 129, + 295, + 183 + ], + "type": "text", + "content": "[30] Alexander Tong, Nikolay Malkin, Guillaume Huguet, Yanlei Zhang, Jarrid Rector-Brooks, Kilian Fatras, Guy Wolf, and Yoshua Bengio. Improving and generalizing flow-based generative models with minibatch optimal transport. Transactions on Machine Learning Research, 2023. 1, 2, 3, 7, 5" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 56, + 185, + 295, + 250 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 185, + 295, + 250 + ], + "spans": [ + { + "bbox": [ + 56, + 185, + 295, + 250 + ], + "type": "text", + "content": "[31] Alexander Y Tong, Nikolay Malkin, Kilian Fatras, Lazar Atanackovic, Yanlei Zhang, Guillaume Huguet, Guy Wolf, and Yoshua Bengio. Simulation-free schrödinger bridges via score and flow matching. In International Conference on Artificial Intelligence and Statistics, pages 1279-1287. PMLR, 2024. 7" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 56, + 251, + 295, + 294 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 251, + 295, + 294 + ], + "spans": [ + { + "bbox": [ + 56, + 251, + 295, + 294 + ], + "type": "text", + "content": "[32] Zhisheng Xiao, Karsten Kreis, and Arash Vahdat. Tackling the generative learning trilemma with denoising diffusion gans. International Conference on Learning Representations, 2022. 1" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 56, + 297, + 295, + 328 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 297, + 295, + 328 + ], + "spans": [ + { + "bbox": [ + 56, + 297, + 295, + 328 + ], + "type": "text", + "content": "[33] Ye Yuan and Kris Kitani. Diverse trajectory forecasting with determinantal point processes. International Conference on Learning Representations, 2020. 3, 5" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 56, + 331, + 295, + 373 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 331, + 295, + 373 + ], + "spans": [ + { + "bbox": [ + 56, + 331, + 295, + 373 + ], + "type": "text", + "content": "[34] Qinqing Zheng, Matt Le, Neta Shaul, Yaron Lipman, Aditya Grover, and Ricky TQ Chen. Guided flows for generative modeling and decision making. arXiv preprint arXiv:2311.13443, 2023. 1" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 56, + 375, + 295, + 441 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 375, + 295, + 441 + ], + "spans": [ + { + "bbox": [ + 56, + 375, + 295, + 441 + ], + "type": "text", + "content": "[35] Yinglin Zheng, Hao Yang, Ting Zhang, Jianmin Bao, Dongdong Chen, Yangyu Huang, Lu Yuan, Dong Chen, Ming Zeng, and Fang Wen. General facial representation learning in a visual-linguistic manner. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 18697–18709, 2022. 2" + } + ] + } + ], + "index": 6 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 749, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 749, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 749, + 317, + 757 + ], + "type": "text", + "content": "23312" + } + ] + } + ], + "index": 8 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 9 + } + ], + "_backend": "vlm", + "_version_name": "2.6.4" +} \ No newline at end of file diff --git a/2025/Divide and Conquer_ Heterogeneous Noise Integration for Diffusion-based Adversarial Purification/94f1ab27-85c1-491d-abe8-0116716b932b_content_list.json b/2025/Divide and Conquer_ Heterogeneous Noise Integration for Diffusion-based Adversarial Purification/94f1ab27-85c1-491d-abe8-0116716b932b_content_list.json new file mode 100644 index 0000000000000000000000000000000000000000..78d0c3c10989ee1c862cc703b18b6736485346e3 --- /dev/null +++ b/2025/Divide and Conquer_ Heterogeneous Noise Integration for Diffusion-based Adversarial Purification/94f1ab27-85c1-491d-abe8-0116716b932b_content_list.json @@ -0,0 +1,1744 @@ +[ + { + "type": "text", + "text": "Divide and Conquer: Heterogeneous Noise Integration for Diffusion-based Adversarial Purification", + "text_level": 1, + "bbox": [ + 205, + 130, + 794, + 174 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Gaozheng Pei $^{1}$ Shaojie Lyu $^{2}$ Gong Chen $^{2}$ Ke Ma $^{1*}$ Qianqian Xu $^{3}$ Yingfei Sun $^{1}$ Qingming Huang $^{3,4,5*}$ $^{1}$ School of Electronic, Electrical and Communication Engineering, UCAS $^{2}$ Tencent Corporate \n $^{3}$ Key Laboratory of Intelligent Information Processing, Institute of Computing Technology, CAS \n $^{4}$ School of Computer Science and Technology, UCAS \n $^{5}$ Key Laboratory of Big Data Mining and Knowledge Management, UCAS", + "bbox": [ + 99, + 202, + 898, + 292 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "peigaozheng23@mails.ucas.ac.cn, {shaojielv,natchen}@tencent.com xuqianqian@ict.ac.cn, {make,yfsun,qmhuang}@ucas.ac.cn", + "bbox": [ + 217, + 294, + 784, + 327 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Abstract", + "text_level": 1, + "bbox": [ + 246, + 361, + 326, + 378 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Existing diffusion-based purification methods aim to disrupt adversarial perturbations by introducing a certain amount of noise through a forward diffusion process, followed by a reverse process to recover clean examples. However, this approach is fundamentally flawed: the uniform operation of the forward process across all pixels compromises normal pixels while attempting to combat adversarial perturbations, resulting in the target model producing incorrect predictions. Simply relying on low-intensity noise is insufficient for effective defense. To address this critical issue, we implement a heterogeneous purification strategy grounded in the interpretability of neural networks. Our method decisively applies higher-intensity noise to specific pixels that the target model focuses on while the remaining pixels are subjected to only low-intensity noise. This requirement motivates us to redesign the sampling process of the diffusion model, allowing for the effective removal of varying noise levels. Furthermore, to evaluate our method against strong adaptative attack, our proposed method sharply reduces time cost and memory usage through a single-step resampling. The empirical evidence from extensive experiments across three datasets demonstrates that our method outperforms most current adversarial training and purification techniques by a substantial margin. Code is available at https://github.com/GaozhengPei/Purification.", + "bbox": [ + 88, + 393, + 485, + 787 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "1. Introduction", + "text_level": 1, + "bbox": [ + 91, + 818, + 220, + 834 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Adversarial perturbations [9], which are often imperceptible to humans, can significantly alter the predictions of", + "bbox": [ + 89, + 844, + 483, + 875 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/074f326a88409f96ff88a145555e55280a5bb3e943490ed93cc9063fe1721a44.jpg", + "image_caption": [ + "Figure 1. When the noise added to adversarial samples is minimal (top), the adversarial perturbations remain intact and cannot be removed. On the other hand, if the noise intensity is excessive (bottom), it can distort the semantic information. Our approach employs an attention mask to introduce varying intensities of noise across different areas (middle). This technique effectively strikes a balance between preserving semantic information and mitigating adversarial perturbations." + ], + "image_footnote": [], + "bbox": [ + 526, + 361, + 895, + 511 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "neural networks. To address these risks [21-23, 32], various defense mechanisms have been proposed. One common approach is adversarial training [18, 37, 45], which involves generating adversarial samples according to a specific attack method before the test phase and then retraining the model to improve its robustness. However, this approach often struggles to defend against unknown attacks. Another technique, known as adversarial purification [17, 27, 30, 38, 40, 44], seeks to remove perturbations from adversarial examples during the test phase. This approach does not rely on any assumptions about the attack methods and eliminates the overhead of repetitive training. Its flexibility and strong generalization have drawn increasing interest. In recent years, diffusion models [16, 29] have garnered attention due to their impressive denoising abilities and capacity to generate high-quality images, making them", + "bbox": [ + 511, + 657, + 908, + 902 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "CVF", + "bbox": [ + 106, + 2, + 181, + 42 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore.", + "bbox": [ + 236, + 0, + 810, + 46 + ], + "page_idx": 0 + }, + { + "type": "page_footnote", + "text": "*Corresponding author", + "bbox": [ + 109, + 886, + 235, + 900 + ], + "page_idx": 0 + }, + { + "type": "page_number", + "text": "29268", + "bbox": [ + 478, + 944, + 519, + 957 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "a promising option for adversarial purification.", + "bbox": [ + 89, + 90, + 400, + 107 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Existing diffusion-based adversarial purification methods [3, 30, 40, 44] follow the traditional forward process which attempts to add uniform noise across different regions of the images and then recover the clean images through reverse process. From Figure 1 we can observe that such a manner can not balance eliminating perturbations and preserving semantic information. If the added noise is intense, it will remove the adversarial perturbations but damage the semantic information, causing the wrong prediction of both natural and adversarial examples. In contrast, if the noise intensity is weak, it may not effectively purify the adversarial perturbations, resulting in a successful evasion attack. Therefore, how to achieve a good balance between the elimination of adversarial perturbations and the preservation of semantic information remains challenging.", + "bbox": [ + 89, + 109, + 483, + 351 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "In this paper, we propose a novel diffusion-based adversarial purification method from the perspective of neural network interpretability. During making decisions, the neural network spontaneously allocates varying levels of attention to different regions of the image. This phenomenon inspired us to use attention maps as a prior guided diffusion model to impose heterogeneous noise on different regions of an image as forward process. By eliminating adversarial perturbations in the critical areas through high-intensity noise, we can substantially mitigate their effect on the prediction of the target model. Conversely, for regions that receive less attention, we introduce low-intensity noise to preserve as much semantic information as possible in the original image. To recover a clean image from one containing heterogeneous noise, we design a two-stage denoising process in accordance with the attention map as a reverse process of the diffusion model. Specifically, if the sampling timestep falls between two different noise levels, we apply inpainting techniques to remove localized high-intensity noise. If the current timestep is below that of the smaller noise level, we use the standard denoising process to remove global weak noise. Additionally, by replacing multi-step resampling with an improved single-step resampling approach, this strategy results in a $90\\%$ reduction in both time consumption and GPU memory usage enables evaluate our method against strong adaptive attacks. Overall, our contributions can be summarized as follows:", + "bbox": [ + 89, + 354, + 483, + 760 + ], + "page_idx": 1 + }, + { + "type": "list", + "sub_type": "text", + "list_items": [ + "- We propose a novel forward process that adds noise of varying intensities based on the classifier's attention to different regions, achieving a balance between preserving semantic information and eliminating adversarial perturbations.", + "- To recover the clean samples, we design a two-stage denoising process: at a larger timestep, we convert it into an inpainting problem to remove localized strong noise, and at a smaller timestep, we apply a standard sampling" + ], + "bbox": [ + 89, + 763, + 483, + 901 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "process to eliminate global weak noise.", + "bbox": [ + 526, + 90, + 787, + 106 + ], + "page_idx": 1 + }, + { + "type": "list", + "sub_type": "text", + "list_items": [ + "- To evaluate our method against strong adaptive attacks. We replace the multi-step resampling with an improved single-step resampling, this not only achieves harmonious image generation but also significantly reduces time cost and memory usage.", + "- Extensive experiments on three datasets show that our method outperforms other methods by a promising improvement against various adversarial attacks." + ], + "bbox": [ + 511, + 107, + 903, + 226 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "2. Related Work", + "text_level": 1, + "bbox": [ + 511, + 239, + 653, + 255 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Adversarial purification aims to remove adversarial perturbations from samples at test phase before feeding them into classifier model. Adversarial purification methods can be broadly divided into two categories: training-based methods, which require training on the dataset, and diffusion based methods which doesn't require training and is not necessary to access the training dataset.", + "bbox": [ + 511, + 265, + 906, + 371 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "2.1. Training-Based Adversarial Purification", + "text_level": 1, + "bbox": [ + 511, + 378, + 857, + 395 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "[41] shows empirically that adversarial examples mainly lie in the low probability regions of the training distribution and move them back towards the training distribution. [36] first models the distribution of clean samples and then attempts to find the closest sample within the clean sample distribution to the adversarial example during inference. [28] generates perturbed images using Self-supervised Perturbation (SSP) attack that disrupts the deep perceptual features and projects back the perturbed images close to the perceptual space of clean images. [52] propose to learn generalizable invariant features across attacks via encoder by playing zero-sum game and reconstruct the original image via decoder. [24] disrupt dversarial perturbations structure through random transform and reconstruct the image using a purifier model. [42] experimentally show that FGSM-AT networks overfit to unknown adversarial examples attacked again by FGSM. In the test phase, adversarial examples are fed to classifier after performing FGSM \"purification\". However, these methods require training using the training dataset, which is very time-consuming and lacks generalizability.", + "bbox": [ + 511, + 401, + 906, + 719 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "2.2. Diffusion-Based Adversarial Purification", + "text_level": 1, + "bbox": [ + 511, + 727, + 862, + 742 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "[49] show that an EBM trained with Denoising Score-Matching can quickly purify attacked images within a few steps. [30] first diffuse samples with a small amount of noise and then recover the clean image through a reverse generative process. The paper also theoretically proves that the time step cannot be too large or too small. [44] proposes to use the adversarial image as guidance in the reverse process which encourages the purified image is consistent with the original clean image. [47] gets multiple reversed samples via multiple runs of denoising. Then the samples are", + "bbox": [ + 511, + 750, + 906, + 902 + ], + "page_idx": 1 + }, + { + "type": "page_number", + "text": "29269", + "bbox": [ + 478, + 944, + 519, + 957 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "passed into the classifier, where the final prediction is obtained through classifier voting. [3] design the forward process with the proper amount of Gaussian noise added and the reverse process with the gradient of contrastive loss as the guidance of diffusion models for adversarial purification. However, adding noise of the same intensity to different regions of the images, these methods is difficult to balance preserving semantic information and eliminating adversarial perturbations.", + "bbox": [ + 89, + 90, + 483, + 227 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3. Methodology", + "text_level": 1, + "bbox": [ + 89, + 239, + 225, + 258 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Given an example $(\\pmb{x},\\pmb{y})$ , an adversary can create a small perturbation to generate an adversarial example $\\pmb{x}_{adv}$ such that the classifier $\\pmb{f}$ makes an incorrect prediction: $\\pmb{y} \\neq \\pmb{f}(\\pmb{x}_{adv})$ . Existing diffusion-based purification methods obtain a noisy image $\\pmb{x}(t)$ at timestep $t$ through the forward process and then recover the clean (or purified) image $\\hat{\\pmb{x}}$ through the reverse process. And we hope $\\pmb{f}(\\hat{\\pmb{x}}) = \\pmb{f}(\\pmb{x})$ . From the theoretical analysis of [30], the timestep $t$ should satisfy: 1) $t$ should be large enough to eliminate perturbations. 2) $t$ should be small enough that $\\hat{\\pmb{x}}$ still maintains semantic consistency with $\\pmb{x}$ . We find that these two objectives are contradictory.", + "bbox": [ + 89, + 265, + 483, + 445 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Motivated by the interpretability of neural networks, the attention map is primarily used to encode the input spatial regions that the network focuses on most when making output decisions, such as in image classification. This suggests that, for an adversarial example, adversarial perturbations in the regions where the classifier focuses most significantly contribute to causing the target classifier to make incorrect predictions. The attention maps for each layer of the target classifier can be obtained with a single forward propagation. Therefore, we propose introducing high-intensity noise to eliminate adversarial perturbations of the regions with more attention value and applying low-intensity noise to the remaining area to preserve as much semantic information as possible.", + "bbox": [ + 89, + 446, + 483, + 657 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3.1. Heterogeneous Forward Process", + "text_level": 1, + "bbox": [ + 89, + 666, + 375, + 681 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "For the adversarial example $\\pmb{x}_{adv}$ at timestep $t = 0$ , i.e., $\\pmb{x}(0) = \\pmb{x}_{adv}$ , we can obtain the noisy images at timestep $t_l$ and $t_s$ with $t_l > t_s$ as follows:", + "bbox": [ + 89, + 688, + 483, + 734 + ], + "page_idx": 2 + }, + { + "type": "equation", + "text": "\n$$\n\\boldsymbol {x} \\left(t _ {l}\\right) = \\sqrt {\\bar {\\alpha} \\left(t _ {l}\\right)} \\cdot \\boldsymbol {x} _ {a d v} + \\sqrt {1 - \\bar {\\alpha} \\left(t _ {l}\\right)} \\cdot \\boldsymbol {\\epsilon}, \\tag {1a}\n$$\n", + "text_format": "latex", + "bbox": [ + 135, + 758, + 482, + 775 + ], + "page_idx": 2 + }, + { + "type": "equation", + "text": "\n$$\n\\boldsymbol {x} \\left(t _ {s}\\right) = \\sqrt {\\bar {\\alpha} \\left(t _ {s}\\right)} \\cdot \\boldsymbol {x} _ {a d v} + \\sqrt {1 - \\bar {\\alpha} \\left(t _ {s}\\right)} \\cdot \\boldsymbol {\\epsilon}, \\tag {1b}\n$$\n", + "text_format": "latex", + "bbox": [ + 133, + 779, + 480, + 797 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "where $\\{\\sqrt{\\overline{\\alpha}(t)}\\}$ is a series of given parameters and $\\epsilon \\sim \\mathcal{N}(0, I)$ . The clean (or purified) images $\\hat{\\pmb{x}}_{t_l}$ and $\\hat{\\pmb{x}}_{t_s}$ can be obtained through the normal reverse process [16, 30]. For the smaller timestep $t_s$ , the recovered image $\\hat{\\pmb{x}}_{t_s}$ retains certain adversarial perturbation, resulting in $\\pmb{f}(\\pmb{x}_{adv}) = \\pmb{f}(\\hat{\\pmb{x}}_{t_s}) \\neq \\pmb{f}(\\pmb{x})$ . Conversely, at the larger timestep $t_l$ , the", + "bbox": [ + 89, + 808, + 483, + 902 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "clean image $\\hat{\\pmb{x}}_{t_l}$ loses critical semantic information, leading to $\\pmb {f}(\\pmb{x}_{adv})\\neq \\pmb {f}(\\hat{\\pmb{x}}_{t_l})\\neq \\pmb {f}(\\pmb {x})$ . To effectively aggregate the differing impacts on various regions of the images at distinct timesteps, we construct an attention mask to integrate (1a) and (1b).", + "bbox": [ + 511, + 90, + 905, + 165 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Consider a neural network that consists of $M$ blocks (e.g. ResNet-50 [14] consists of 4 blocks). For $m$ -th block, we extract the output of the last activation function (e.g. Relu [8] or Swish [33]) as follows:", + "bbox": [ + 511, + 167, + 905, + 227 + ], + "page_idx": 2 + }, + { + "type": "equation", + "text": "\n$$\n\\boldsymbol {G} _ {m} = \\left(\\sum_ {i = 1} ^ {C _ {m}} | \\boldsymbol {A} _ {m i} | ^ {p}\\right) ^ {\\frac {1}{p}} \\in \\mathbb {R} ^ {H _ {m} \\times W _ {m}}, m \\in [ M ], \\tag {2}\n$$\n", + "text_format": "latex", + "bbox": [ + 537, + 237, + 905, + 284 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "where $p \\in \\mathbb{N}$ , $C_m, H_m, W_m$ are the dimension of channels, height, and width respectively, $\\mathbf{A}_m$ denotes the activation of $m$ -th block and $\\mathbf{A}_{mi}$ denotes the $i$ -th slice of $\\mathbf{A}_m$ along the channel dimension. Then we apply interpolation techniques to ensure it matches the original image size. By upsampling $\\mathbf{G}_m$ with bilinear function $\\mathrm{Bi}(\\cdot)$ , we obtain attention map of $m$ -th block:", + "bbox": [ + 511, + 292, + 906, + 397 + ], + "page_idx": 2 + }, + { + "type": "equation", + "text": "\n$$\n\\mathrm {A M} _ {m} = \\phi (\\operatorname {B i} \\left(\\boldsymbol {G} _ {m}\\right)), \\tag {3}\n$$\n", + "text_format": "latex", + "bbox": [ + 635, + 398, + 903, + 415 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "where $\\phi (\\cdot)$ is the spatial softmax operation. Moreover, we build the attention mask with all attention maps $\\{\\mathbf{AM}_m\\}$ :", + "bbox": [ + 511, + 421, + 905, + 452 + ], + "page_idx": 2 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {M} = \\bigcup_ {m = 1} ^ {M} \\mathbb {I} [ \\mathrm {A M} _ {m} > \\tau ] \\tag {4}\n$$\n", + "text_format": "latex", + "bbox": [ + 625, + 462, + 903, + 503 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "where $\\mathbb{I}[\\cdot ]$ is the Iverson bracket and $\\tau$ is the threshold to distinguish important and ordinary regions. With the attention mask $\\mathcal{M}$ and two noisy images from Eq.(1), $\\pmb {x}(t_s),\\pmb {x}(t_l),t_s < t_l$ , the result of the proposed heterogeneous forward process is", + "bbox": [ + 511, + 513, + 905, + 590 + ], + "page_idx": 2 + }, + { + "type": "equation", + "text": "\n$$\n\\boldsymbol {x} \\left(t _ {l}, t _ {s}\\right) = \\boldsymbol {x} \\left(t _ {l}\\right) \\odot \\mathcal {M} + \\boldsymbol {x} \\left(t _ {s}\\right) \\odot (1 - \\mathcal {M}), \\tag {5}\n$$\n", + "text_format": "latex", + "bbox": [ + 560, + 602, + 903, + 618 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "where $\\odot$ is the Hadamard product.", + "bbox": [ + 511, + 628, + 743, + 643 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3.2. Heterogeneous Denoising Process", + "text_level": 1, + "bbox": [ + 511, + 654, + 805, + 669 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "To recover the clean image from (5), we design a two-stage denoising process. For the current timestep $t$ , we consider the following two cases:", + "bbox": [ + 511, + 676, + 905, + 719 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Stage 1 ( $t_s < t < t_l$ ) When the current timestep $t$ lies between $t_s$ and $t_l$ . For the pixels inside the attention mask $\\mathcal{M}$ , they should be predicted from the previous timestep $t + 1$ . But for the pixels outside the attention mask $\\mathcal{M}$ , they can be obtained by further adding noise forward from (5). Therefore, this case can be viewed as an inpainting problem [25]. For the image regions outside the attention mask $\\mathcal{M}$ , since only the forward process is involved. Therefore, we can get the pixel value of these regions at timestep $t$ equivalently as follows:", + "bbox": [ + 511, + 720, + 905, + 871 + ], + "page_idx": 2 + }, + { + "type": "equation", + "text": "\n$$\n\\boldsymbol {x} (t) ^ {\\text {k n o w n}} = \\sqrt {\\bar {\\alpha} (t)} \\boldsymbol {x} _ {a d v} + \\sqrt {1 - \\bar {\\alpha} (t)} \\boldsymbol {\\epsilon}. \\tag {6}\n$$\n", + "text_format": "latex", + "bbox": [ + 568, + 883, + 903, + 902 + ], + "page_idx": 2 + }, + { + "type": "page_number", + "text": "29270", + "bbox": [ + 478, + 944, + 519, + 957 + ], + "page_idx": 2 + }, + { + "type": "image", + "img_path": "images/c19ed89de3269214133ea08a31903efe44e71df7d0106db9530d5170900606d4.jpg", + "image_caption": [ + "Figure 2. Pipeline of our method. Given an adversarial image, we extract the attention maps of each block during the forward propagation of the classifier, and construct an attention mask $\\mathcal{M}$ . We then execute the denoising process of the diffusion model we designed. It can be seen that the samples purified by our method can be correctly classified." + ], + "image_footnote": [], + "bbox": [ + 98, + 89, + 903, + 330 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "For the image regions within the attention mask, the pixel value of these regions at timestep $t$ should be predicted by the diffusion model:", + "bbox": [ + 89, + 412, + 483, + 457 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\boldsymbol {x} (t) ^ {\\text {u n k n o w n}} \\sim \\mathcal {N} \\left(\\mu_ {\\theta} (\\boldsymbol {x} (t + 1), t + 1), \\Sigma_ {\\theta} (\\boldsymbol {x} (t + 1), t + 1)\\right) \\tag {7}\n$$\n", + "text_format": "latex", + "bbox": [ + 89, + 467, + 483, + 500 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Where $\\theta$ is the parameters of the diffusion model. $\\mu_{\\theta}$ and $\\Sigma_{\\theta}$ are mean and variance, respectively. Then we combine $x(t)^{known}$ and $x(t)^{unknown}$ together based on the attention mask $\\mathcal{M}$ to get the $\\pmb{x}(t)$ at timestep $t$ :", + "bbox": [ + 89, + 500, + 483, + 560 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\boldsymbol {x} (t) = \\mathcal {M} \\odot \\boldsymbol {x} (t) ^ {\\text {k n o w n}} + (1 - \\mathcal {M}) \\odot \\boldsymbol {x} (t) ^ {\\text {u n k n o w n}}, \\tag {8}\n$$\n", + "text_format": "latex", + "bbox": [ + 99, + 569, + 483, + 588 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "where $\\odot$ is Hadamard product. We can remove the localized strong noise through the Stage 1.", + "bbox": [ + 89, + 597, + 482, + 627 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Stage 2 ( $t < t_s$ ) When the current timestep $t$ is less than $t_s$ , since the noise intensity is the same both inside and outside the attention mask. Therefore, the attention mask $\\mathcal{M}$ is no longer needed. At this point, we can proceed with the standard sampling process of the diffusion model. We can get $x(t)$ predicted from the previous timestep $x(t + 1)$ as follows:", + "bbox": [ + 89, + 628, + 483, + 733 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\boldsymbol {x} (t) \\sim \\mathcal {N} \\left(\\mu_ {\\theta} (\\boldsymbol {x} (t + 1), t + 1), \\Sigma_ {\\theta} (\\boldsymbol {x} (t + 1), t + 1)\\right) \\tag {9}\n$$\n", + "text_format": "latex", + "bbox": [ + 101, + 744, + 482, + 762 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Following Stage 2, we will eliminate global weak noise and get the clean image. The complete algorithm process can be found in the Algorithm 1. Additionally, there are some visualizations of our method in the Figure 5.", + "bbox": [ + 89, + 771, + 483, + 832 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "4. Adaptive Attack to Diffusion Purification", + "text_level": 1, + "bbox": [ + 89, + 845, + 460, + 862 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "When directly applying the method described in 3.2, we observe the same phenomenon as [25] that although the region", + "bbox": [ + 89, + 869, + 483, + 901 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Algorithm 1 Heterogeneous Denoising Process", + "text_level": 1, + "bbox": [ + 514, + 412, + 830, + 426 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Require: Sample $x_0$ , Attention Mask $\\mathcal{M}$ , timestep $t_s$ and $t_l$ , Diffusion model parameters $\\theta$ .", + "bbox": [ + 514, + 431, + 905, + 460 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "for $t = t_l,\\dots,1$ do", + "bbox": [ + 529, + 462, + 656, + 476 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\text {i f} t _ {s} \\leq t \\leq t _ {l} \\text {t h e n}\n$$\n", + "text_format": "latex", + "bbox": [ + 545, + 477, + 673, + 491 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Stage 1", + "bbox": [ + 836, + 477, + 898, + 491 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\epsilon \\sim \\mathcal {N} (0, \\mathbf {I}) \\text {i f} t > 1, \\text {e l s e} \\epsilon = 0\n$$\n", + "text_format": "latex", + "bbox": [ + 563, + 492, + 777, + 507 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\nx _ {t - 1} ^ {\\text {k n o w n}} = \\sqrt {\\overline {{\\alpha}} _ {t - 1}} x _ {0} + \\sqrt {1 - \\overline {{\\alpha}} _ {t - 1}} \\epsilon\n$$\n", + "text_format": "latex", + "bbox": [ + 563, + 507, + 802, + 522 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\nz \\sim \\mathcal {N} (0, \\mathbf {I}) \\text {i f} t > 1, \\text {e l s e} z = 0\n$$\n", + "text_format": "latex", + "bbox": [ + 565, + 522, + 779, + 537 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\nx _ {t - 1} ^ {u n k n o w n} = \\frac {1}{\\sqrt {\\alpha_ {t}}} \\left(x _ {t} - \\frac {\\beta_ {t}}{\\sqrt {1 - \\bar {\\alpha} _ {t}}} \\epsilon_ {\\theta} (x _ {t}, t)\\right) + \\sigma_ {t} z\n$$\n", + "text_format": "latex", + "bbox": [ + 563, + 537, + 883, + 561 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\nx _ {t - 1} = \\mathcal {M} \\odot x _ {t - 1} ^ {\\text {k n o w n}} + (1 - \\mathcal {M}) \\odot x _ {t - 1} ^ {\\text {u n k n o w n}}\n$$\n", + "text_format": "latex", + "bbox": [ + 563, + 561, + 872, + 575 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "else if $t < t_s$ then", + "bbox": [ + 545, + 575, + 669, + 590 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Stage 2", + "bbox": [ + 839, + 577, + 905, + 590 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\nz \\sim \\mathcal {N} (0, \\mathbf {I}) \\text {i f} t > 1, \\text {e l s e} z = 0\n$$\n", + "text_format": "latex", + "bbox": [ + 563, + 592, + 779, + 606 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\nx _ {t - 1} = \\frac {1}{\\sqrt {\\alpha_ {t}}} \\left(x _ {t} - \\frac {\\beta_ {t}}{\\sqrt {1 - \\bar {\\alpha} _ {t}}} \\epsilon_ {\\theta} (x _ {t}, t)\\right) + \\sigma_ {t} z\n$$\n", + "text_format": "latex", + "bbox": [ + 563, + 606, + 849, + 628 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "for $u = 1,2,\\ldots U$ do", + "bbox": [ + 547, + 628, + 684, + 642 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "$\\triangleright$ Resampling", + "bbox": [ + 805, + 628, + 900, + 643 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\nz \\sim \\mathcal {N} (0, \\mathbf {I}) \\text {i f} t > 1, \\text {e l s e} z = 0\n$$\n", + "text_format": "latex", + "bbox": [ + 563, + 643, + 779, + 657 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\nx _ {t - 1 + u} = \\sqrt {\\alpha_ {t - 1 + u}} x _ {t + u - 2} + \\sqrt {1 - \\alpha_ {t - 1 + u}} z\n$$\n", + "text_format": "latex", + "bbox": [ + 563, + 657, + 870, + 674 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "end for", + "bbox": [ + 547, + 675, + 601, + 686 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "get $x_{t - 1}$ using (11)", + "bbox": [ + 547, + 689, + 676, + 704 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "end for", + "bbox": [ + 531, + 705, + 584, + 715 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "recovered in Stage 2 matches the texture of the neighboring region, it is semantically incorrect. We can observe this phenomenon in the Figure 3. [25] addresses this issue via multi-step resampling by diffusing the output $\\pmb{x}(t)$ back to $\\pmb{x}(t + 1)$ :", + "bbox": [ + 511, + 744, + 905, + 820 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\boldsymbol {x} (t + 1) = \\sqrt {\\alpha (t)} \\boldsymbol {x} (t) + \\sqrt {1 - \\alpha (t)} \\boldsymbol {\\epsilon}, \\tag {10}\n$$\n", + "text_format": "latex", + "bbox": [ + 576, + 829, + 903, + 847 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "and it is necessary to fully execute 3.2 again to get $\\pmb{x}(t)$ . From Figure 3 we can find that this process needs to be repeated $U = 20$ times to ensure that the masked region", + "bbox": [ + 511, + 854, + 905, + 901 + ], + "page_idx": 3 + }, + { + "type": "page_number", + "text": "29271", + "bbox": [ + 478, + 944, + 517, + 955 + ], + "page_idx": 3 + }, + { + "type": "image", + "img_path": "images/8346b0525140df04bf24c7cd2ebd2317d0e0571c55a7d25ec030b94ea26ba3a0.jpg", + "image_caption": [ + "Figure 3. Our method maintains semantic consistency between the masked areas and the surrounding pixel values when $U = 10$ . Note that no matter what $U$ is, our method requires only one additional call to the denoising network." + ], + "image_footnote": [], + "bbox": [ + 107, + 88, + 472, + 233 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "becomes semantically consistent with the surrounding pixels. To evaluate our method against strong adaptive attacks [30], it is necessary to compute the full gradient of the whole sampling process to construct adversarial examples. To reduce the time cost and memory usage, we optimized the multi-step resampling into a single resampling process which achieves the same functionality.", + "bbox": [ + 89, + 327, + 483, + 431 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "The intuition behind our improvement is that repeating (10) and 3.2 for $U$ times can be replaced by that first diffuse $\\pmb{x}(t)$ to $\\pmb{x}(t + U)$ and then denoise it back to $\\pmb{x}(t)$ from $\\pmb{x}(t + U)$ in a single step using DDIM [39] as follows:", + "bbox": [ + 89, + 433, + 483, + 494 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\begin{array}{l} \\boldsymbol {x} _ {t - 1} = \\sqrt {\\alpha_ {t - 1}} \\left(\\frac {\\boldsymbol {x} _ {t - 1 + U} - \\sqrt {1 - \\alpha_ {t - 1 + U}} \\epsilon_ {\\theta} (\\boldsymbol {x} _ {t - 1 + U})}{\\sqrt {\\alpha_ {t - 1 + U}}}\\right) \\\\ + \\sqrt {1 - \\alpha_ {t - 1} - \\sigma_ {t - 1 + U} ^ {2}} \\cdot \\epsilon_ {\\theta} \\left(\\boldsymbol {x} _ {t - 1 + U}\\right) + \\sigma_ {t - 1 + U} \\boldsymbol {\\epsilon}. \\tag {11} \\\\ \\end{array}\n$$\n", + "text_format": "latex", + "bbox": [ + 96, + 506, + 480, + 579 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "From the bottom panel of Figure 3, we find that our method can also achieve semantically consistency when $U = 10$ . This means that at each timestep, we only need to invoke the denoising network an additional time. Therefore, the GPU memory needs to store the $2 * T$ complete computation graphs of the denoising network which enables the computation of the gradient of our method on consumer-grade GPU with only 24G memory. Also, the time cost is greatly reduced. The analysis of time cost and memory usage can refer to the Appendix.", + "bbox": [ + 89, + 580, + 483, + 733 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "5. Experiment", + "text_level": 1, + "bbox": [ + 89, + 747, + 215, + 763 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "5.1. Experimental Settings", + "text_level": 1, + "bbox": [ + 89, + 772, + 297, + 789 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Datasets and network architectures. Three datasets is used to evaluate including CIFAR-10, SVHN, and ImageNet [6]. We compare our results with several leading defense methods as listed in the standardized benchmark RobustBench [5] for CIFAR-10 and ImageNet, and also test different adversarial purification methods. We use two common classifier architectures: WideResNet-28-10", + "bbox": [ + 89, + 794, + 483, + 900 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "and WideResNet-70-16 [51] for CIFAR-10. For SVHN, WideResNet-28-10 serves as the backbone, while for ImageNet, ResNet-50 [14] is used as the backbone.", + "bbox": [ + 511, + 90, + 903, + 135 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Adversarial attack methods. Strong adaptive attack is tested for our method and other adversarial purification methods. We use the commonly known AutoAttack [4] under $\\ell_{\\infty}$ and $\\ell_{2}$ threat models. Additionally, the projected gradient descent (PGD) attack [26] is evaluated on our method, as recommended in [20]. Due to the randomness introduced by the diffusion and denoising processes, Expectation Over Time (EOT) [1] is adapted for these adaptive attacks. Moreover, we apply the BPDA+EOT [15] attack to ensure a fair comparison with other adversarial purification methods. Finally, following the suggestion of [20], a surrogate process is used to obtain the gradient of the reverse process for white-box attacks.", + "bbox": [ + 511, + 136, + 906, + 330 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Pre-trained Models. We use the unconditional CIFAR-10 checkpoint of EDM provided by NVIDIA [19] on CIFAR-10. 256x256 diffusion (unconditional) checkpoint from guided-diffusion library is adopted for the ImageNet experiments. For CIFAR-10, the pre-trained classifier is downloaded from RobustBench [5], while for ImageNet, the weight of the classifier is provided by PyTorch library.", + "bbox": [ + 511, + 332, + 905, + 436 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Evaluation metrics. To assess the performance of defense methods, we use two metrics: standard accuracy, measured on clean samples, and robust accuracy, measured on adversarial examples. Due to the high computational cost of evaluating models against adaptive attacks, a fixed subset of 512 images are randomly sampled from the test set for robust evaluation, which is the same as [20, 30]. In all experiments, the mean and standard deviation over three runs are reported to assess both standard and robust accuracy.", + "bbox": [ + 511, + 438, + 905, + 573 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Implementation details. We follow the settings described in [20]. We evaluate diffusion-based purification methods using the PGD+EOT attack with 200 update iterations, and BPDA and AutoAttack with 100 update iterations, except for ImageNet, which uses 20 iterations. We set the number of EOT to 20. The step size is 0.007. For randomized defenses, such as [3, 20, 30, 40], we use the random version of AutoAttack, and for static defenses, we use the standard version.", + "bbox": [ + 511, + 574, + 905, + 709 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "5.2. Experimental Results", + "text_level": 1, + "bbox": [ + 511, + 724, + 715, + 739 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "CIFAR-10. We conducted comprehensive experiments on the CIFAR-10 dataset, evaluating our method against other approaches on two model architectures: WideResNet28-10 and WideResNet-70-16. We assessed robust accuracy under three attack types: BPDA+EOT, PGD+EOT, and AutoAttack. We tested against the infinity norm attack, and as shown in Table 1, when the backbone is WideResNet28-10, our method significantly outperforms other baselines in both standard accuracy and robust accuracy, exceeding by $3.00\\%$ in standard accuracy and leading by $3.56\\%$ and", + "bbox": [ + 511, + 750, + 906, + 900 + ], + "page_idx": 4 + }, + { + "type": "page_number", + "text": "29272", + "bbox": [ + 478, + 944, + 519, + 957 + ], + "page_idx": 4 + }, + { + "type": "table", + "img_path": "images/e3afc5aa41d175559558e362ab26c27a625ac3d712e67c9b4f4725c6d1fbb065.jpg", + "table_caption": [ + "Table 1. Standard and robust accuracy of different Adversarial Training (AT) and Adversarial Purification (AP) methods against PGD+EOT and AutoAttack $\\ell_{\\infty}(\\epsilon = 8 / 255)$ on CIFAR-10. * utilizes half number of iterations for the attack due to the high computational cost. † indicates the requirement of extra data." + ], + "table_footnote": [], + "table_body": "
TypeMethodStandard Acc.Robust Acc.
PGDAutoAttack
WideResNet-28-10
ATGowal et al. [13]88.5465.9363.38
Gowal et al. [11]†87.5166.0162.76
Pang et al. [31]88.6264.9561.04
APYoon et al. [50]85.66±0.5133.48±0.8659.53±0.87
Nie et al. [30]90.07±0.9756.84±0.5963.6±0.81
Lee et al. [20]90.16±0.6455.82±0.5970.47±1.53
Bai et al. [3]91.4149.22*77.08
Geigh et al. [53]84.20-59.14
Lin et al. [24]90.62-72.85
Ours93.16±0.2159.38±1.2580.45±1.54
", + "bbox": [ + 91, + 169, + 488, + 380 + ], + "page_idx": 5 + }, + { + "type": "table", + "img_path": "images/e800e50078b020cd088646e9072eef5aa5c8881745de55becdf24ed8b356dbc0.jpg", + "table_caption": [ + "Table 2. Standard and robust accuracy of different Adversarial Training (AT) and Adversarial Purification (AP) methods against PGD+EOT and AutoAttack $\\ell_{\\infty}(\\epsilon = 8 / 255)$ on CIFAR-10. * The number of iterations for the attack is half that of the other methods for less computational overhead. † indicates the requirement of extra data." + ], + "table_footnote": [], + "table_body": "
TypeMethodStandard Acc.Robust Acc.
PGDAutoAttack
WideResNet-70-16
ATRebuffi et al. [34]†92.2269.9766.56
Gowal et al. [10]†91.1068.6666.10
Gowal et al. [12]88.7569.0365.87
APYoon et al. [50]86.76±1.1537.11±1.3560.86±0.56
Nie et al. [30]90.43±0.6051.13±0.8766.06±1.17
Lee et al. [20]90.53±0.156.88±1.0670.31±0.62
Bai et al. [3]92.9748.83*79.10
Geigh et al. [53]84.60-66.40
Geigh et al. [53]86.90-59.20
Lin et al. [24]91.99-76.37
Ours93.36±0.5959.79±0.6184.83±0.59
", + "bbox": [ + 91, + 494, + 491, + 718 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "3.37% under PGD and AutoAttack, respectively. From Table 2, we see that when WideResNet-70-16 is used as the backbone, our method exceeds the state-of-the-art (SOTA) by 1.37% in standard accuracy and leads by 2.91% and 5.73% under PGD and AutoAttack, respectively. We also evaluated the accuracy of our method and other baselines under the Euclidean norm attack. From Table 3, we observe that, with WideResNet-28-10 as the backbone, although our method lags behind [3] by 0.59% in standard accuracy, it leads by 0.98% and 3.77% in robust accuracy. Due to the", + "bbox": [ + 88, + 750, + 485, + 902 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Table 3. Standard and robust accuracy against PGD+EOT and AutoAttack $\\ell_2(\\epsilon = 0.5)$ on CIFAR-10. Adversarial Training (AT) and Adversarial Purification (AP) methods are evaluated. *The number of iterations for the attack is half that of the other methods for less computational overhead. † indicates the requirement of extra data. ‡ adopts WideResNet-34-10 as the backbone, with the same width but more layers than the default one.", + "bbox": [ + 511, + 88, + 906, + 188 + ], + "page_idx": 5 + }, + { + "type": "table", + "img_path": "images/286e2b0967827d4023ae69b8604200f31bc661056c464f9fc56067fbf7f0c39f.jpg", + "table_caption": [], + "table_footnote": [], + "table_body": "
TypeMethodStandard Acc.Robust Acc.
PGDAutoAttack
WideResNet-28-10
ATRebuffi et al. [34]†91.7985.0578.80
Augustin et al. [2]‡93.9686.1478.79
Sehwag et al. [31]‡90.9383.7577.24
APYoon et al. [50]85.66±0.5173.32±0.7679.57±0.38
Nie et al. [30]91.41±1.0079.45±1.1681.7±0.84
Lee et al [20]90.16±0.6483.59±0.8886.48±0.38
Bai et al. [3]93.7586.13*80.92
Geigh et al. [53]84.40-77.90
Geigh et al. [53]84.20-73.60
Lin et al. [24]90.62-80.47
Ours93.16±0.2187.11±0.9590.25±0.80
", + "bbox": [ + 516, + 196, + 903, + 422 + ], + "page_idx": 5 + }, + { + "type": "table", + "img_path": "images/7ff2727de888403bb7f15b41151ca2bfa7875b3c2e00d0ac0d2c78a0c9d49f88.jpg", + "table_caption": [ + "Table 4. Standard and robust accuracy against PGD+EOT and AutoAttack $\\ell_2(\\epsilon = 0.5)$ on CIFAR-10. Adversarial Training (AT) and Adversarial Purification (AP) methods are evaluated. (*The number of iterations for the attack is half that of other methods for less computational overhead. † methods need extra data.)" + ], + "table_footnote": [], + "table_body": "
TypeMethodStandard Acc.Robust Acc.
PGDAutoAttack
WideResNet-70-16
ATRebuffi et al. [34]†95.7489.6282.32
Gowal et al. [10]†94.7488.1880.53
Rebuffi et al. [34]92.4186.2480.42
APYoon et al. [50]86.76±1.1575.666±1.2980.43±0.42
Nie et al. [30]92.15±0.7282.97±1.3883.06±1.27
Lee et al [20]90.53±0.1483.75±0.9985.59±0.61
Bai et al. [3]92.9784.37*83.01
Lin et al. [24]91.99-81.35
Ours93.36±0.5987.7±0.9791.71±1.09
", + "bbox": [ + 514, + 521, + 908, + 719 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "significant time overhead of [3], which only iterates 100 times under PGD—half of what others do—our method outperforms more under PGD attacks. From Table 4, we see that when WideResNet-70-16 is the backbone, our method outperforms other baselines, leading by $1.21\\%$ in standard accuracy and by $3.33\\%$ and $6.12\\%$ in robust accuracy under PGD and AutoAttack, respectively. Additionally, we applied the BPDA+EOT attack, which approximates differentiability. From Table 5, we observe that although robust accuracy lags behind [40] by $1.71\\%$ , our method still leads", + "bbox": [ + 511, + 750, + 910, + 902 + ], + "page_idx": 5 + }, + { + "type": "page_number", + "text": "29273", + "bbox": [ + 478, + 944, + 519, + 957 + ], + "page_idx": 5 + }, + { + "type": "table", + "img_path": "images/d46d73af3e6bfae438d708ac7dacee533bf14adaa46c76d287693d552415d00d.jpg", + "table_caption": [ + "Table 5. Standard and robust accuracy against BPDA+EOT $\\ell_{\\infty}(\\epsilon = 8 / 255)$ on CIFAR-10." + ], + "table_footnote": [], + "table_body": "
MethodPurificationAccuracy
StandardRobust
WideResNet-28-10
Song et al. [41]Gibbs Update95.009.00
Yang et al. [48]Mask+Recon94.0015.00
Hill et al. [15]EBM+LD84.1254.90
Yoon et al. [50]DSM+LD85.66±0.5166.91±1.75
Nie et al. [30]Diffusion90.07±0.9781.45±1.51
Bai et al. [3]Diffusion91.37±1.2185.51±0.81
Wang et al. [44]Diffusion89.96±0.4075.59±1.26
Song et al. [40]Diffusion89.88±0.3588.43±0.83
Lee et al. [20]Diffusion90.16±0.6488.40±0.88
OursDiffusion93.04±0.6586.72±0.20
", + "bbox": [ + 99, + 128, + 475, + 338 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "in standard accuracy by $3.16\\%$ , resulting in an average advantage of $0.73\\%$ . We also measured the distance between purified samples and the original samples in feature space, shown in the Appendix. A smaller distance indicates that the purified samples retain more consistent semantic information with the original samples (further discussion in the appendix). From Figure 6, we find that the distribution of purified samples by our method is most similar to the original samples compared to other methods. Overall, our method outperforms others, demonstrating its effectiveness in preserving semantic information while eliminating adversarial perturbations.", + "bbox": [ + 88, + 364, + 483, + 547 + ], + "page_idx": 6 + }, + { + "type": "table", + "img_path": "images/6978497420bad30b7d90c7ec02cc27ed80081fde3196e79950031bdee1bf6dbf.jpg", + "table_caption": [ + "Table 6. Standard and robust accuracy against PGD+EOT $\\ell_{\\infty}(\\epsilon = 8 / 255)$ on ImageNet. ResNet-50 is used as the classifier." + ], + "table_footnote": [], + "table_body": "
TypeMethodAccuracy
StandardRobust
ResNet-50
ATSalman et al. [35]63.8639.11
Engstrom et al. [7]62.4233.20
Wong et al. [46]53.8328.04
APNie et al. (t* = 0.2) [30]73.9640.63
Nie et al. (t* = 0.3) [30]72.8548.24
Nie et al. (t* = 0.4) [30]61.7141.67
Bai et al. [3]70.4141.7
Lee et al. [20]71.4246.59
Song et al. [40]62.2551.14
Geigh et al. [53]73.9856.54
Ours74.51±1.0760.06±1.66
", + "bbox": [ + 104, + 599, + 470, + 825 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "ImageNet. We test ResNet50 as the backbone on the ImageNet dataset under PGD+EOT attacks, which is the same as [3, 20, 30]. From Table 6, our method leads competitors by $0.53\\%$ in standard accuracy and by $3.52\\%$ in", + "bbox": [ + 89, + 839, + 483, + 902 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/437faba82c23906b56fbff745704b27c211fe5b0cbdad2f8c2545d6bfff24be5.jpg", + "image_caption": [ + "Figure 4. 2D purification trajectories of Our method and DiffPure. For DiffPure (left), if $t^{*} = 0.2$ , although the purification direction is toward the original class, the adversarial perturbation cannot be completely eliminated; If $t^{*} = 0.4$ , the semantic information changes and the purification direction is no longer toward the original class. Our method (right) can eliminate adversarial perturbation and remain the semantic information." + ], + "image_footnote": [], + "bbox": [ + 517, + 89, + 705, + 252 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/7c0d260e32729411e6704017f5b77a62d94ecdcf7eb856cf11bd72c4555738d4.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 723, + 89, + 916, + 252 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/bd8b81cd98e629f949190121e32820383bfa88d119d6fbded0f61a52cc952f2d.jpg", + "image_caption": [ + "Figure 5. Visualization of attention map, attention mask and images purified by our method." + ], + "image_footnote": [], + "bbox": [ + 517, + 372, + 921, + 563 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "robust accuracy, significantly surpassing the baseline. Additionally, we observe that the method from [30] achieves a standard accuracy of $73.96\\%$ and a robust accuracy of $40.63\\%$ when $t^{*} = 0.2$ . When $t^{*} = 0.3$ , standard accuracy drops to $72.85\\%$ while robust accuracy rises to $48.24\\%$ . However, at $t^{*} = 0.4$ , both standard accuracy and robust accuracy decline to $61.71\\%$ and $41.67\\%$ , respectively. Additionally, We visualize the 2D purification trajectories of our method and DiffPure using the T-SNE [43] technique. From Figure 4, we observe that when $t^{*} = 0.2$ , the adversarial perturbation on the adversarial sample has not been fully purified, and it remains at the boundary between the two classes of samples. When $t^{*} = 0.4$ , the semantic information of the adversarial example begins to change, and as a result, the purification trajectory no longer points toward the original classes. This indicates that applying the same noise intensity across an image cannot effectively balance semantic preservation and the elimination of adversarial perturbation.", + "bbox": [ + 511, + 628, + 908, + 902 + ], + "page_idx": 6 + }, + { + "type": "page_number", + "text": "29274", + "bbox": [ + 478, + 944, + 519, + 957 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/f2fe2f7da156c147aede35b69270193397e65fa7941be1fc6e3c72d973e71519.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 91, + 88, + 285, + 244 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/2f0d78aadcdf83081c5d03c6f18176e424690d48eb30011f598427be3bf7f023.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 305, + 88, + 496, + 244 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/465d02c0236b8ba9d053548104da1a56da129f811c9f47a5da7cc242eda6bdee.jpg", + "image_caption": [ + "Figure 6. Joint distribution of the original samples and purified samples. The distributions of the purified samples by our method and the original samples are the most similar." + ], + "image_footnote": [], + "bbox": [ + 86, + 255, + 285, + 415 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/47ef5752cd0e60b6731935e6c95310e974fcf95da032e0fedd57118f5da4b3a9.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 303, + 257, + 496, + 414 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "tions. In contrast, our method adds varying noise intensities to different regions of an image, applying stronger noise to areas with higher classification weights to eliminate adversarial perturbations, while applying lighter noise to other regions to preserve semantic information. This approach allows our method to achieve good results in both robust accuracy and standard accuracy simultaneously. Our method achieves comparable standard accuracy to [30] at $t^* = 0.2$ but significantly outperforms it in robust accuracy, demonstrating the effectiveness of our approach.", + "bbox": [ + 89, + 500, + 483, + 652 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/3e786b46938306ad70c75c0e8cf8c8e2c26fc791e4439b071a27362b44e47c4a.jpg", + "image_caption": [ + "Figure 7. Robust and Standard Accuracy under different thresholds $\\tau$ (left) and under different combinations of timesteps $t_l$ and $t_s$ (right)." + ], + "image_footnote": [], + "bbox": [ + 94, + 670, + 292, + 816 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/1561298de81ab5e7748e4dc5623d426b61de577b70cac408191e9b045e94b936.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 303, + 670, + 496, + 816 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "5.3. Ablation Study", + "text_level": 1, + "bbox": [ + 513, + 90, + 666, + 107 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "We first conducted a hyperparameter sensitivity analysis. Our method involves three hyperparameters, two of which are noises of different intensities, $t_{l}$ and $t_{s}$ , and the third is a threshold, $\\tau$ , used to control the size of the attention mask. From Figure 7, we observe that as $\\tau$ increases (resulting in smaller mask areas), the Standard Accuracy also consistently increases. However, adversarial perturbations are preserved at this time, leading to poorer Robust Accuracy when $\\tau$ is large. When $\\tau$ is small, semantic information is damaged, which also results in weaker Robust Accuracy. Therefore, we choose $\\tau$ to be around 0.8. Simultaneously, we can see that when $t_{l} = 0.2$ , $t_{s} = 0.05$ , the Robust Accuracy is the highest. Here, $t_{s}$ is less than the $t^{*}$ used in [30] to preserve more semantic information, while $t_{l}$ is greater than $t^{*}$ to eliminate adversarial perturbations. From Table 6, we can see that when we set $t_{l} = 0.4$ and $t_{s} = 0.2$ , The Standard Accuracy is almost as high as the $t^{*} = 0.2$ of [30], but at the same time, our Robust Accuracy significantly leads. This demonstrates the effectiveness of our method. We also explore whether it is necessary for each block", + "bbox": [ + 511, + 113, + 906, + 415 + ], + "page_idx": 7 + }, + { + "type": "table", + "img_path": "images/ea0028e09fa3351c2258308c0f1f4cf852e2601c6bbca1cf8a7941edbc8202a3.jpg", + "table_caption": [ + "Table 7. Standard Accuracy and Robust Accuracy for different block combinations. WideResNet28-10 servers as the backbone." + ], + "table_footnote": [], + "table_body": "
Encoder BlockAccuracy
E1E2E3E4StandardRobustAverage
X✓✓✓92.7758.0075.38
✓X✓✓93.1658.0075.58
✓✓X✓93.3557.6175.48
✓✓✓X91.6658.8575.25
✓✓✓✓93.1659.3876.27
", + "bbox": [ + 547, + 469, + 874, + 585 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "to participate in constructing the attention mask. Taking WideResNet28-10 as an example, it can be divided into four blocks. We extract the output of the final ReLU function of each block as the activation map. From Table 7, we can see that removing any block results in a performance loss. Therefore, every block should be utilized.", + "bbox": [ + 511, + 601, + 906, + 691 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "6. Conclusion", + "text_level": 1, + "bbox": [ + 511, + 708, + 633, + 723 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "We propose a novel forward process that adds noise of varying intensities based on the classifier's attention to different regions, achieving a balance between preserving semantic information and eliminating adversarial perturbations. The denoising process is tailored to our method to remove different-intensity noise within a noisy image. By replacing the multi-step resampling with an improved single-step resampling, time cost and memory usage are reduced significantly. Extensive experiments on three datasets show that our method outperforms other methods by a promising improvement against various adversarial attacks.", + "bbox": [ + 511, + 734, + 906, + 900 + ], + "page_idx": 7 + }, + { + "type": "page_number", + "text": "29275", + "bbox": [ + 478, + 944, + 517, + 955 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Acknowledge", + "text_level": 1, + "bbox": [ + 91, + 90, + 209, + 107 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "This work was supported in part by the National Key R&D Program of China under Grant 2022ZD0119204, in part by National Natural Science Foundation of China: 62236008, 62441232, U21B2038, U23B2051, 62122075 and 62376257, in part by Youth Innovation Promotion Association CAS, in part by the Strategic Priority Research Program of the Chinese Academy of Sciences, Grant No. XDB0680201, in part by the Fundamental Research Funds for the Central Universities, in part by the Tencent Marketing Solution Rhino-Bird Focused Research Program, in part by Xiaomi Young Talents Program.", + "bbox": [ + 89, + 113, + 485, + 268 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "References", + "text_level": 1, + "bbox": [ + 91, + 294, + 187, + 310 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[1] Anish Athalye, Nicholas Carlini, and David Wagner. Obfuscated gradients give a false sense of security: Circumventing defenses to adversarial examples. In International conference on machine learning, pages 274-283. PMLR, 2018. 5", + "[2] Maximilian Augustin, Alexander Meinke, and Matthias Hein. Adversarial robustness on in-and out-distribution improves explainability. In European Conference on Computer Vision, pages 228-245. Springer, 2020. 6", + "[3] Mingyuan Bai, Wei Huang, Tenghui Li, Andong Wang, Junbin Gao, Cesar F Caiafa, and Qibin Zhao. Diffusion models demand contrastive guidance for adversarial purification to advance. In *Forty-first International Conference on Machine Learning*, 2024. 2, 3, 5, 6, 7", + "[4] Francesco Croce and Matthias Hein. Reliable evaluation of adversarial robustness with an ensemble of diverse parameter-free attacks. In International conference on machine learning, pages 2206-2216, 2020. 5", + "[5] Francesco Croce, Maksym Andriushchenko, Vikash Sehwag, Edoardo Debenedetti, Nicolas Flammarion, Mung Chiang, Prateek Mittal, and Matthias Hein. Robustbench: a standardized adversarial robustness benchmark. arXiv preprint arXiv:2010.09670, 2020. 5", + "[6] Jia Deng, Wei Dong, Richard Socher, Li-Jia Li, Kai Li, and Li Fei-Fei. Imagenet: A large-scale hierarchical image database. In 2009 IEEE conference on computer vision and pattern recognition, pages 248–255, 2009. 5", + "[7] Logan Engstrom, Andrew Ilyas, Hadi Salman, Shibani Santurkar, and Dimitris Tsipras. Robustness (python library), 2019. URL https://github.com/MadryLab/robustness, 4(4): 4-3, 2019. 7", + "[8] Xavier Glorot, Antoine Bordes, and Yoshua Bengio. Deep sparse rectifier neural networks. In Proceedings of the fourteenth international conference on artificial intelligence and statistics, pages 315-323, 2011. 3", + "[9] Ian J Goodfellow, Jonathon Shlens, and Christian Szegedy. Explaining and harnessing adversarial examples. arXiv preprint arXiv:1412.6572, 2014. 1", + "[10] Sven Gowal, Chongli Qin, Jonathan Uesato, Timothy Mann, and Pushmeet Kohli. Uncovering the limits of adversarial training against norm-bounded adversarial examples. arXiv preprint arXiv:2010.03593, 2020. 6" + ], + "bbox": [ + 93, + 319, + 483, + 900 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[11] Sven Gowal, Chongli Qin, Jonathan Uesato, Timothy Mann, and Pushmeet Kohli. Uncovering the limits of adversarial training against norm-bounded adversarial examples. arXiv preprint arXiv:2010.03593, 2020. 6", + "[12] Sven Gowal, Sylvestre-Alvise Rebuffi, Olivia Wiles, Florian Stimberg, Dan Andrei Calian, and Timothy A Mann. Improving robustness using generated data. Advances in Neural Information Processing Systems, 34:4218-4233, 2021. 6", + "[13] Sven Gowal, Sylvestre-Alvise Rebuffi, Olivia Wiles, Florian Stimberg, Dan Andrei Calian, and Timothy A Mann. Improving robustness using generated data. Advances in Neural Information Processing Systems, 34:4218-4233, 2021. 6", + "[14] Kaiming He, Xiangyu Zhang, Shaoqing Ren, and Jian Sun. Deep residual learning for image recognition. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 770-778, 2016. 3, 5", + "[15] Mitch Hill, Jonathan Craig Mitchell, and Song-Chun Zhu. Stochastic security: Adversarial defense using long-run dynamics of energy-based models. In International Conference on Learning Representations, 2021. 5, 7", + "[16] Jonathan Ho, Ajay Jain, and Pieter Abbeel. Denoising diffusion probabilistic models. Advances in neural information processing systems, 33:6840-6851, 2020. 1, 3", + "[17] Duhun Hwang, Eunjung Lee, and Wonjong Rhee. Aidpurifier: A light auxiliary network for boosting adversarial defense. Neurocomputing, 541:126251, 2023. 1", + "[18] Xiaojun Jia, Yong Zhang, Xingxing Wei, Baoyuan Wu, Ke Ma, Jue Wang, and Xiaochun Cao. Improving fast adversarial training with prior-guided knowledge. IEEE Transactions on Pattern Analysis and Machine Intelligence, 2024. 1", + "[19] Tero Karras, Miika Aittala, Timo Aila, and Samuli Laine. Elucidating the design space of diffusion-based generative models. Advances in neural information processing systems, 35:26565-26577, 2022. 5", + "[20] Minjong Lee and Dongwoo Kim. Robust evaluation of diffusion-based adversarial purification. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 134-144, 2023. 5, 6, 7", + "[21] Siyuan Liang, Xingxing Wei, Siyuan Yao, and Xiaochun Cao. Efficient adversarial attacks for visual object tracking. In Computer Vision-ECCV 2020: 16th European Conference, Glasgow, UK, August 23-28, 2020, Proceedings, Part XXVI 16, pages 34-50. Springer, 2020. 1", + "[22] Siyuan Liang, Longkang Li, Yanbo Fan, Xiaojun Jia, Jingzhi Li, Baoyuan Wu, and Xiaochun Cao. A large-scale multiple-objective method for black-box attack against object detection. In European Conference on Computer Vision, pages 619-636. Springer, 2022.", + "[23] Siyuan Liang, Baoyuan Wu, Yanbo Fan, Xingxing Wei, and Xiaochun Cao. Parallel rectangle flip attack: A query-based black-box attack against object detection. arXiv preprint arXiv:2201.08970, 2022. 1", + "[24] Guang Lin, Chao Li, Jianhai Zhang, Toshihisa Tanaka, and Qibin Zhao. Adversarial training on purification (ATop): Advancing both robustness and generalization. In The Twelfth International Conference on Learning Representations, 2024. 2, 6" + ], + "bbox": [ + 516, + 92, + 906, + 900 + ], + "page_idx": 8 + }, + { + "type": "page_number", + "text": "29276", + "bbox": [ + 478, + 944, + 519, + 957 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[25] Andreas Lugmayr, Martin Danelljan, Andres Romero, Fisher Yu, Radu Timofte, and Luc Van Gool. Repaint: Inpainting using denoising diffusion probabilistic models. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 11461-11471, 2022. 3, 4", + "[26] Aleksander Madry, Aleksandar Makelov, Ludwig Schmidt, Dimitris Tsipras, and Adrian Vladu. Towards deep learning models resistant to adversarial attacks. In International Conference on Learning Representations, 2018. 5", + "[27] Dongyu Meng and Hao Chen. Magnet: a two-pronged defense against adversarial examples. In Proceedings of the 2017 ACM SIGSAC conference on computer and communications security, pages 135-147, 2017. 1", + "[28] Muzammal Naseer, Salman Khan, Munawar Hayat, Fahad Shahbaz Khan, and Fatih Porikli. A self-supervised approach for adversarial robustness. In 2020 IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), 2020. 2", + "[29] Alexander Quinn Nichol and Prafulla Dhariwal. Improved denoising diffusion probabilistic models. In International conference on machine learning, 2021. 1", + "[30] Weili Nie, Brandon Guo, Yujia Huang, Chaowei Xiao, Arash Vahdat, and Animashree Anandkumar. Diffusion models for adversarial purification. In Proceedings of the 39th International Conference on Machine Learning, 2022. 1, 2, 3, 5, 6, 7, 8", + "[31] Tianyu Pang, Min Lin, Xiao Yang, Jun Zhu, and Shuicheng Yan. Robustness and accuracy could be reconcilable by (proper) definition. In International Conference on Machine Learning, pages 17258-17277. PMLR, 2022. 6", + "[32] Gaozheng Pei, Ke Ma, Pinci Yang, Qianqian Xu, Yingfei Sun, et al. Exploring query efficient data generation towards data-free model stealing in hard label setting. arXiv preprint arXiv:2412.15276, 2024. 1", + "[33] Prajit Ramachandran, Barret Zoph, and Quoc V Le. Searching for activation functions. arXiv preprint arXiv:1710.05941, 2017. 3", + "[34] Sylvestre-Alvise Rebuffi, Sven Gowal, Dan A Calian, Florian Stimberg, Olivia Wiles, and Timothy Mann. Fixing data augmentation to improve adversarial robustness. arXiv preprint arXiv:2103.01946, 2021. 6", + "[35] Hadi Salman, Andrew Ilyas, Logan Engstrom, Ashish Kapoor, and Aleksander Madry. Do adversarially robust imagenet models transfer better? Advances in Neural Information Processing Systems, 33:3533-3545, 2020. 7", + "[36] Pouya Samangouei, Maya Kabbab, and Rama Chellappa. Defense-GAN: Protecting classifiers against adversarial attacks using generative models. In International Conference on Learning Representations, 2018. 2", + "[37] Ali Shafahi, Mahyar Najibi, Mohammad Amin Ghiasi, Zheng Xu, John Dickerson, Christoph Studer, Larry S Davis, Gavin Taylor, and Tom Goldstein. Adversarial training for free! Advances in neural information processing systems, 32, 2019. 1", + "[38] Changhao Shi, Chester Holtz, and Gal Mishne. Online adversarial purification based on self-supervised learning. In International Conference on Learning Representations, 2021. 1" + ], + "bbox": [ + 91, + 92, + 483, + 898 + ], + "page_idx": 9 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[39] Jiaming Song, Chenlin Meng, and Stefano Ermon. Denoising diffusion implicit models. In International Conference on Learning Representations, 2021. 5", + "[40] Kaiyu Song, Hanjiang Lai, Yan Pan, and Jian Yin. Mimic diffusion: Purifying adversarial perturbation via mimicking clean diffusion model. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), 2024. 1, 2, 5, 6, 7", + "[41] Yang Song, Taesup Kim, Sebastian Nowozin, Stefano Ermon, and Nate Kushman. Pixeldefend: Leveraging generative models to understand and defend against adversarial examples. In International Conference on Learning Representations, 2018. 2, 7", + "[42] Linyu Tang and Lei Zhang. Robust overfitting does matter: Test-time adversarial purification with fgsm. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), 2024. 2", + "[43] Laurens Van der Maaten and Geoffrey Hinton. Visualizing data using t-sne. Journal of machine learning research, 9 (11), 2008. 7", + "[44] Jinyi Wang, Zhaoyang Lyu, Dahua Lin, Bo Dai, and Hongfei Fu. Guided Diffusion Model for Adversarial Purification. arXiv, 2022. 1, 2, 7", + "[45] Eric Wong, Leslie Rice, and J. Zico Kolter. Fast is better than free: Revisiting adversarial training. In International Conference on Learning Representations, 2020. 1", + "[46] Eric Wong, Leslie Rice, and J. Zico Kolter. Fast is better than free: Revisiting adversarial training. In International Conference on Learning Representations, 2020. 7", + "[47] Chaowei Xiao, Zhongzhu Chen, Kun Jin, Jiongxiao Wang, Weili Nie, Mingyan Liu, Anima Anandkumar, Bo Li, and Dawn Song. Densepure: Understanding diffusion models for adversarial robustness. In The Eleventh International Conference on Learning Representations, 2023. 2", + "[48] Yuzhe Yang, Guo Zhang, Dina Katabi, and Zhi Xu. Me-net: Towards effective adversarial robustness with matrix estimation. In International Conference on Machine Learning, pages 7025-7034. PMLR, 2019. 7", + "[49] Jongmin Yoon, Sung Ju Hwang, and Juho Lee. Adversarial purification with score-based generative models. In Proceedings of the 38th International Conference on Machine Learning, 2021. 2", + "[50] Jongmin Yoon, Sung Ju Hwang, and Juho Lee. Adversarial purification with score-based generative models. In International Conference on Machine Learning, 2021. 6, 7", + "[51] Sergey Zagoruyko. Wide residual networks. arXiv preprint arXiv:1605.07146, 2016. 5", + "[52] Dawei Zhou, Tongliang Liu, Bo Han, Nannan Wang, Chunlei Peng, and Xinbo Gao. Towards defending against adversarial examples via attack-invariant features. In Proceedings of the 38th International Conference on Machine Learning, 2021. 2", + "[53] Geigh Zollicoffer, Minh Vu, Ben Nebgen, Juan Castorena, Boian Alexandrov, and Manish Bhattarai. Lorid: Low-rank iterative diffusion for adversarial purification. arXiv preprint arXiv:2409.08255, 2024. 6, 7" + ], + "bbox": [ + 516, + 92, + 903, + 883 + ], + "page_idx": 9 + }, + { + "type": "page_number", + "text": "29277", + "bbox": [ + 478, + 944, + 517, + 955 + ], + "page_idx": 9 + } +] \ No newline at end of file diff --git a/2025/Divide and Conquer_ Heterogeneous Noise Integration for Diffusion-based Adversarial Purification/94f1ab27-85c1-491d-abe8-0116716b932b_model.json b/2025/Divide and Conquer_ Heterogeneous Noise Integration for Diffusion-based Adversarial Purification/94f1ab27-85c1-491d-abe8-0116716b932b_model.json new file mode 100644 index 0000000000000000000000000000000000000000..9f33e12f771d8e8ea5fa40a622d4141d0439b7b1 --- /dev/null +++ b/2025/Divide and Conquer_ Heterogeneous Noise Integration for Diffusion-based Adversarial Purification/94f1ab27-85c1-491d-abe8-0116716b932b_model.json @@ -0,0 +1,2354 @@ +[ + [ + { + "type": "header", + "bbox": [ + 0.107, + 0.003, + 0.182, + 0.044 + ], + "angle": 0, + "content": "CVF" + }, + { + "type": "header", + "bbox": [ + 0.237, + 0.001, + 0.812, + 0.047 + ], + "angle": 0, + "content": "This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore." + }, + { + "type": "title", + "bbox": [ + 0.207, + 0.131, + 0.795, + 0.175 + ], + "angle": 0, + "content": "Divide and Conquer: Heterogeneous Noise Integration for Diffusion-based Adversarial Purification" + }, + { + "type": "text", + "bbox": [ + 0.101, + 0.203, + 0.899, + 0.293 + ], + "angle": 0, + "content": "Gaozheng Pei\\(^{1}\\) Shaojie Lyu\\(^{2}\\) Gong Chen\\(^{2}\\) Ke Ma\\(^{1*}\\) Qianqian Xu\\(^{3}\\) Yingfei Sun\\(^{1}\\) Qingming Huang\\(^{3,4,5*}\\) \n\\(^{1}\\)School of Electronic, Electrical and Communication Engineering, UCAS \\(^{2}\\)Tencent Corporate \n\\(^{3}\\)Key Laboratory of Intelligent Information Processing, Institute of Computing Technology, CAS \n\\(^{4}\\)School of Computer Science and Technology, UCAS \n\\(^{5}\\)Key Laboratory of Big Data Mining and Knowledge Management, UCAS" + }, + { + "type": "text", + "bbox": [ + 0.218, + 0.295, + 0.785, + 0.328 + ], + "angle": 0, + "content": "peigaozheng23@mails.ucas.ac.cn, {shaojielv,natchen}@tencent.com xuqianqian@ict.ac.cn, {make,yfsun,qmhuang}@ucas.ac.cn" + }, + { + "type": "title", + "bbox": [ + 0.248, + 0.362, + 0.328, + 0.379 + ], + "angle": 0, + "content": "Abstract" + }, + { + "type": "text", + "bbox": [ + 0.089, + 0.395, + 0.486, + 0.789 + ], + "angle": 0, + "content": "Existing diffusion-based purification methods aim to disrupt adversarial perturbations by introducing a certain amount of noise through a forward diffusion process, followed by a reverse process to recover clean examples. However, this approach is fundamentally flawed: the uniform operation of the forward process across all pixels compromises normal pixels while attempting to combat adversarial perturbations, resulting in the target model producing incorrect predictions. Simply relying on low-intensity noise is insufficient for effective defense. To address this critical issue, we implement a heterogeneous purification strategy grounded in the interpretability of neural networks. Our method decisively applies higher-intensity noise to specific pixels that the target model focuses on while the remaining pixels are subjected to only low-intensity noise. This requirement motivates us to redesign the sampling process of the diffusion model, allowing for the effective removal of varying noise levels. Furthermore, to evaluate our method against strong adaptative attack, our proposed method sharply reduces time cost and memory usage through a single-step resampling. The empirical evidence from extensive experiments across three datasets demonstrates that our method outperforms most current adversarial training and purification techniques by a substantial margin. Code is available at https://github.com/GaozhengPei/Purification." + }, + { + "type": "title", + "bbox": [ + 0.092, + 0.819, + 0.222, + 0.835 + ], + "angle": 0, + "content": "1. Introduction" + }, + { + "type": "text", + "bbox": [ + 0.091, + 0.845, + 0.485, + 0.876 + ], + "angle": 0, + "content": "Adversarial perturbations [9], which are often imperceptible to humans, can significantly alter the predictions of" + }, + { + "type": "image", + "bbox": [ + 0.527, + 0.362, + 0.897, + 0.512 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.512, + 0.521, + 0.908, + 0.633 + ], + "angle": 0, + "content": "Figure 1. When the noise added to adversarial samples is minimal (top), the adversarial perturbations remain intact and cannot be removed. On the other hand, if the noise intensity is excessive (bottom), it can distort the semantic information. Our approach employs an attention mask to introduce varying intensities of noise across different areas (middle). This technique effectively strikes a balance between preserving semantic information and mitigating adversarial perturbations." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.659, + 0.909, + 0.903 + ], + "angle": 0, + "content": "neural networks. To address these risks [21-23, 32], various defense mechanisms have been proposed. One common approach is adversarial training [18, 37, 45], which involves generating adversarial samples according to a specific attack method before the test phase and then retraining the model to improve its robustness. However, this approach often struggles to defend against unknown attacks. Another technique, known as adversarial purification [17, 27, 30, 38, 40, 44], seeks to remove perturbations from adversarial examples during the test phase. This approach does not rely on any assumptions about the attack methods and eliminates the overhead of repetitive training. Its flexibility and strong generalization have drawn increasing interest. In recent years, diffusion models [16, 29] have garnered attention due to their impressive denoising abilities and capacity to generate high-quality images, making them" + }, + { + "type": "page_footnote", + "bbox": [ + 0.11, + 0.887, + 0.236, + 0.901 + ], + "angle": 0, + "content": "*Corresponding author" + }, + { + "type": "page_number", + "bbox": [ + 0.479, + 0.945, + 0.521, + 0.958 + ], + "angle": 0, + "content": "29268" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.091, + 0.092, + 0.401, + 0.108 + ], + "angle": 0, + "content": "a promising option for adversarial purification." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.11, + 0.484, + 0.352 + ], + "angle": 0, + "content": "Existing diffusion-based adversarial purification methods [3, 30, 40, 44] follow the traditional forward process which attempts to add uniform noise across different regions of the images and then recover the clean images through reverse process. From Figure 1 we can observe that such a manner can not balance eliminating perturbations and preserving semantic information. If the added noise is intense, it will remove the adversarial perturbations but damage the semantic information, causing the wrong prediction of both natural and adversarial examples. In contrast, if the noise intensity is weak, it may not effectively purify the adversarial perturbations, resulting in a successful evasion attack. Therefore, how to achieve a good balance between the elimination of adversarial perturbations and the preservation of semantic information remains challenging." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.355, + 0.484, + 0.761 + ], + "angle": 0, + "content": "In this paper, we propose a novel diffusion-based adversarial purification method from the perspective of neural network interpretability. During making decisions, the neural network spontaneously allocates varying levels of attention to different regions of the image. This phenomenon inspired us to use attention maps as a prior guided diffusion model to impose heterogeneous noise on different regions of an image as forward process. By eliminating adversarial perturbations in the critical areas through high-intensity noise, we can substantially mitigate their effect on the prediction of the target model. Conversely, for regions that receive less attention, we introduce low-intensity noise to preserve as much semantic information as possible in the original image. To recover a clean image from one containing heterogeneous noise, we design a two-stage denoising process in accordance with the attention map as a reverse process of the diffusion model. Specifically, if the sampling timestep falls between two different noise levels, we apply inpainting techniques to remove localized high-intensity noise. If the current timestep is below that of the smaller noise level, we use the standard denoising process to remove global weak noise. Additionally, by replacing multi-step resampling with an improved single-step resampling approach, this strategy results in a \\(90\\%\\) reduction in both time consumption and GPU memory usage enables evaluate our method against strong adaptive attacks. Overall, our contributions can be summarized as follows:" + }, + { + "type": "text", + "bbox": [ + 0.091, + 0.765, + 0.483, + 0.839 + ], + "angle": 0, + "content": "- We propose a novel forward process that adds noise of varying intensities based on the classifier's attention to different regions, achieving a balance between preserving semantic information and eliminating adversarial perturbations." + }, + { + "type": "text", + "bbox": [ + 0.091, + 0.841, + 0.484, + 0.902 + ], + "angle": 0, + "content": "- To recover the clean samples, we design a two-stage denoising process: at a larger timestep, we convert it into an inpainting problem to remove localized strong noise, and at a smaller timestep, we apply a standard sampling" + }, + { + "type": "list", + "bbox": [ + 0.091, + 0.765, + 0.484, + 0.902 + ], + "angle": 0, + "content": null + }, + { + "type": "text", + "bbox": [ + 0.527, + 0.092, + 0.789, + 0.107 + ], + "angle": 0, + "content": "process to eliminate global weak noise." + }, + { + "type": "text", + "bbox": [ + 0.514, + 0.108, + 0.905, + 0.182 + ], + "angle": 0, + "content": "- To evaluate our method against strong adaptive attacks. We replace the multi-step resampling with an improved single-step resampling, this not only achieves harmonious image generation but also significantly reduces time cost and memory usage." + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.183, + 0.905, + 0.227 + ], + "angle": 0, + "content": "- Extensive experiments on three datasets show that our method outperforms other methods by a promising improvement against various adversarial attacks." + }, + { + "type": "list", + "bbox": [ + 0.513, + 0.108, + 0.905, + 0.227 + ], + "angle": 0, + "content": null + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.24, + 0.655, + 0.256 + ], + "angle": 0, + "content": "2. Related Work" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.266, + 0.907, + 0.372 + ], + "angle": 0, + "content": "Adversarial purification aims to remove adversarial perturbations from samples at test phase before feeding them into classifier model. Adversarial purification methods can be broadly divided into two categories: training-based methods, which require training on the dataset, and diffusion based methods which doesn't require training and is not necessary to access the training dataset." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.38, + 0.859, + 0.396 + ], + "angle": 0, + "content": "2.1. Training-Based Adversarial Purification" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.402, + 0.907, + 0.72 + ], + "angle": 0, + "content": "[41] shows empirically that adversarial examples mainly lie in the low probability regions of the training distribution and move them back towards the training distribution. [36] first models the distribution of clean samples and then attempts to find the closest sample within the clean sample distribution to the adversarial example during inference. [28] generates perturbed images using Self-supervised Perturbation (SSP) attack that disrupts the deep perceptual features and projects back the perturbed images close to the perceptual space of clean images. [52] propose to learn generalizable invariant features across attacks via encoder by playing zero-sum game and reconstruct the original image via decoder. [24] disrupt dversarial perturbations structure through random transform and reconstruct the image using a purifier model. [42] experimentally show that FGSM-AT networks overfit to unknown adversarial examples attacked again by FGSM. In the test phase, adversarial examples are fed to classifier after performing FGSM \"purification\". However, these methods require training using the training dataset, which is very time-consuming and lacks generalizability." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.728, + 0.864, + 0.743 + ], + "angle": 0, + "content": "2.2. Diffusion-Based Adversarial Purification" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.75, + 0.907, + 0.903 + ], + "angle": 0, + "content": "[49] show that an EBM trained with Denoising Score-Matching can quickly purify attacked images within a few steps. [30] first diffuse samples with a small amount of noise and then recover the clean image through a reverse generative process. The paper also theoretically proves that the time step cannot be too large or too small. [44] proposes to use the adversarial image as guidance in the reverse process which encourages the purified image is consistent with the original clean image. [47] gets multiple reversed samples via multiple runs of denoising. Then the samples are" + }, + { + "type": "page_number", + "bbox": [ + 0.479, + 0.945, + 0.521, + 0.958 + ], + "angle": 0, + "content": "29269" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.09, + 0.092, + 0.485, + 0.228 + ], + "angle": 0, + "content": "passed into the classifier, where the final prediction is obtained through classifier voting. [3] design the forward process with the proper amount of Gaussian noise added and the reverse process with the gradient of contrastive loss as the guidance of diffusion models for adversarial purification. However, adding noise of the same intensity to different regions of the images, these methods is difficult to balance preserving semantic information and eliminating adversarial perturbations." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.24, + 0.226, + 0.259 + ], + "angle": 0, + "content": "3. Methodology" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.266, + 0.485, + 0.446 + ], + "angle": 0, + "content": "Given an example \\((\\pmb{x},\\pmb{y})\\), an adversary can create a small perturbation to generate an adversarial example \\(\\pmb{x}_{adv}\\) such that the classifier \\(\\pmb{f}\\) makes an incorrect prediction: \\(\\pmb{y} \\neq \\pmb{f}(\\pmb{x}_{adv})\\). Existing diffusion-based purification methods obtain a noisy image \\(\\pmb{x}(t)\\) at timestep \\(t\\) through the forward process and then recover the clean (or purified) image \\(\\hat{\\pmb{x}}\\) through the reverse process. And we hope \\(\\pmb{f}(\\hat{\\pmb{x}}) = \\pmb{f}(\\pmb{x})\\). From the theoretical analysis of [30], the timestep \\(t\\) should satisfy: 1) \\(t\\) should be large enough to eliminate perturbations. 2) \\(t\\) should be small enough that \\(\\hat{\\pmb{x}}\\) still maintains semantic consistency with \\(\\pmb{x}\\). We find that these two objectives are contradictory." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.447, + 0.484, + 0.658 + ], + "angle": 0, + "content": "Motivated by the interpretability of neural networks, the attention map is primarily used to encode the input spatial regions that the network focuses on most when making output decisions, such as in image classification. This suggests that, for an adversarial example, adversarial perturbations in the regions where the classifier focuses most significantly contribute to causing the target classifier to make incorrect predictions. The attention maps for each layer of the target classifier can be obtained with a single forward propagation. Therefore, we propose introducing high-intensity noise to eliminate adversarial perturbations of the regions with more attention value and applying low-intensity noise to the remaining area to preserve as much semantic information as possible." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.667, + 0.377, + 0.683 + ], + "angle": 0, + "content": "3.1. Heterogeneous Forward Process" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.689, + 0.484, + 0.735 + ], + "angle": 0, + "content": "For the adversarial example \\( \\pmb{x}_{adv} \\) at timestep \\( t = 0 \\), i.e., \\( \\pmb{x}(0) = \\pmb{x}_{adv} \\), we can obtain the noisy images at timestep \\( t_l \\) and \\( t_s \\) with \\( t_l > t_s \\) as follows:" + }, + { + "type": "equation", + "bbox": [ + 0.136, + 0.759, + 0.483, + 0.776 + ], + "angle": 0, + "content": "\\[\n\\boldsymbol {x} \\left(t _ {l}\\right) = \\sqrt {\\bar {\\alpha} \\left(t _ {l}\\right)} \\cdot \\boldsymbol {x} _ {a d v} + \\sqrt {1 - \\bar {\\alpha} \\left(t _ {l}\\right)} \\cdot \\boldsymbol {\\epsilon}, \\tag {1a}\n\\]" + }, + { + "type": "equation", + "bbox": [ + 0.135, + 0.78, + 0.482, + 0.798 + ], + "angle": 0, + "content": "\\[\n\\boldsymbol {x} \\left(t _ {s}\\right) = \\sqrt {\\bar {\\alpha} \\left(t _ {s}\\right)} \\cdot \\boldsymbol {x} _ {a d v} + \\sqrt {1 - \\bar {\\alpha} \\left(t _ {s}\\right)} \\cdot \\boldsymbol {\\epsilon}, \\tag {1b}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.809, + 0.484, + 0.903 + ], + "angle": 0, + "content": "where \\(\\{\\sqrt{\\overline{\\alpha}(t)}\\}\\) is a series of given parameters and \\(\\epsilon \\sim \\mathcal{N}(0, I)\\). The clean (or purified) images \\(\\hat{\\pmb{x}}_{t_l}\\) and \\(\\hat{\\pmb{x}}_{t_s}\\) can be obtained through the normal reverse process [16, 30]. For the smaller timestep \\(t_s\\), the recovered image \\(\\hat{\\pmb{x}}_{t_s}\\) retains certain adversarial perturbation, resulting in \\(\\pmb{f}(\\pmb{x}_{adv}) = \\pmb{f}(\\hat{\\pmb{x}}_{t_s}) \\neq \\pmb{f}(\\pmb{x})\\). Conversely, at the larger timestep \\(t_l\\), the" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.092, + 0.906, + 0.166 + ], + "angle": 0, + "content": "clean image \\(\\hat{\\pmb{x}}_{t_l}\\) loses critical semantic information, leading to \\(\\pmb {f}(\\pmb{x}_{adv})\\neq \\pmb {f}(\\hat{\\pmb{x}}_{t_l})\\neq \\pmb {f}(\\pmb {x})\\) . To effectively aggregate the differing impacts on various regions of the images at distinct timesteps, we construct an attention mask to integrate (1a) and (1b)." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.168, + 0.906, + 0.228 + ], + "angle": 0, + "content": "Consider a neural network that consists of \\(M\\) blocks (e.g. ResNet-50 [14] consists of 4 blocks). For \\(m\\)-th block, we extract the output of the last activation function (e.g. Relu [8] or Swish [33]) as follows:" + }, + { + "type": "equation", + "bbox": [ + 0.538, + 0.238, + 0.906, + 0.285 + ], + "angle": 0, + "content": "\\[\n\\boldsymbol {G} _ {m} = \\left(\\sum_ {i = 1} ^ {C _ {m}} | \\boldsymbol {A} _ {m i} | ^ {p}\\right) ^ {\\frac {1}{p}} \\in \\mathbb {R} ^ {H _ {m} \\times W _ {m}}, m \\in [ M ], \\tag {2}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.294, + 0.907, + 0.398 + ], + "angle": 0, + "content": "where \\(p \\in \\mathbb{N}\\), \\(C_m, H_m, W_m\\) are the dimension of channels, height, and width respectively, \\(\\mathbf{A}_m\\) denotes the activation of \\(m\\)-th block and \\(\\mathbf{A}_{mi}\\) denotes the \\(i\\)-th slice of \\(\\mathbf{A}_m\\) along the channel dimension. Then we apply interpolation techniques to ensure it matches the original image size. By upsampling \\(\\mathbf{G}_m\\) with bilinear function \\(\\mathrm{Bi}(\\cdot)\\), we obtain attention map of \\(m\\)-th block:" + }, + { + "type": "equation", + "bbox": [ + 0.636, + 0.4, + 0.905, + 0.416 + ], + "angle": 0, + "content": "\\[\n\\mathrm {A M} _ {m} = \\phi (\\operatorname {B i} \\left(\\boldsymbol {G} _ {m}\\right)), \\tag {3}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.422, + 0.906, + 0.453 + ], + "angle": 0, + "content": "where \\(\\phi (\\cdot)\\) is the spatial softmax operation. Moreover, we build the attention mask with all attention maps \\(\\{\\mathbf{AM}_m\\}\\):" + }, + { + "type": "equation", + "bbox": [ + 0.626, + 0.463, + 0.905, + 0.504 + ], + "angle": 0, + "content": "\\[\n\\mathcal {M} = \\bigcup_ {m = 1} ^ {M} \\mathbb {I} [ \\mathrm {A M} _ {m} > \\tau ] \\tag {4}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.515, + 0.906, + 0.591 + ], + "angle": 0, + "content": "where \\(\\mathbb{I}[\\cdot ]\\) is the Iverson bracket and \\(\\tau\\) is the threshold to distinguish important and ordinary regions. With the attention mask \\(\\mathcal{M}\\) and two noisy images from Eq.(1), \\(\\pmb {x}(t_s),\\pmb {x}(t_l),t_s < t_l\\) , the result of the proposed heterogeneous forward process is" + }, + { + "type": "equation", + "bbox": [ + 0.561, + 0.603, + 0.905, + 0.619 + ], + "angle": 0, + "content": "\\[\n\\boldsymbol {x} \\left(t _ {l}, t _ {s}\\right) = \\boldsymbol {x} \\left(t _ {l}\\right) \\odot \\mathcal {M} + \\boldsymbol {x} \\left(t _ {s}\\right) \\odot (1 - \\mathcal {M}), \\tag {5}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.63, + 0.744, + 0.645 + ], + "angle": 0, + "content": "where \\(\\odot\\) is the Hadamard product." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.655, + 0.807, + 0.67 + ], + "angle": 0, + "content": "3.2. Heterogeneous Denoising Process" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.677, + 0.906, + 0.72 + ], + "angle": 0, + "content": "To recover the clean image from (5), we design a two-stage denoising process. For the current timestep \\( t \\), we consider the following two cases:" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.722, + 0.906, + 0.872 + ], + "angle": 0, + "content": "Stage 1 (\\( t_s < t < t_l \\)) When the current timestep \\( t \\) lies between \\( t_s \\) and \\( t_l \\). For the pixels inside the attention mask \\( \\mathcal{M} \\), they should be predicted from the previous timestep \\( t + 1 \\). But for the pixels outside the attention mask \\( \\mathcal{M} \\), they can be obtained by further adding noise forward from (5). Therefore, this case can be viewed as an inpainting problem [25]. For the image regions outside the attention mask \\( \\mathcal{M} \\), since only the forward process is involved. Therefore, we can get the pixel value of these regions at timestep \\( t \\) equivalently as follows:" + }, + { + "type": "equation", + "bbox": [ + 0.569, + 0.884, + 0.905, + 0.903 + ], + "angle": 0, + "content": "\\[\n\\boldsymbol {x} (t) ^ {\\text {k n o w n}} = \\sqrt {\\bar {\\alpha} (t)} \\boldsymbol {x} _ {a d v} + \\sqrt {1 - \\bar {\\alpha} (t)} \\boldsymbol {\\epsilon}. \\tag {6}\n\\]" + }, + { + "type": "page_number", + "bbox": [ + 0.479, + 0.945, + 0.521, + 0.958 + ], + "angle": 0, + "content": "29270" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.099, + 0.09, + 0.904, + 0.332 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.09, + 0.344, + 0.908, + 0.388 + ], + "angle": 0, + "content": "Figure 2. Pipeline of our method. Given an adversarial image, we extract the attention maps of each block during the forward propagation of the classifier, and construct an attention mask \\(\\mathcal{M}\\). We then execute the denoising process of the diffusion model we designed. It can be seen that the samples purified by our method can be correctly classified." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.413, + 0.484, + 0.458 + ], + "angle": 0, + "content": "For the image regions within the attention mask, the pixel value of these regions at timestep \\( t \\) should be predicted by the diffusion model:" + }, + { + "type": "equation", + "bbox": [ + 0.09, + 0.468, + 0.484, + 0.5 + ], + "angle": 0, + "content": "\\[\n\\boldsymbol {x} (t) ^ {\\text {u n k n o w n}} \\sim \\mathcal {N} \\left(\\mu_ {\\theta} (\\boldsymbol {x} (t + 1), t + 1), \\Sigma_ {\\theta} (\\boldsymbol {x} (t + 1), t + 1)\\right) \\tag {7}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.5, + 0.484, + 0.561 + ], + "angle": 0, + "content": "Where \\(\\theta\\) is the parameters of the diffusion model. \\(\\mu_{\\theta}\\) and \\(\\Sigma_{\\theta}\\) are mean and variance, respectively. Then we combine \\(x(t)^{known}\\) and \\(x(t)^{unknown}\\) together based on the attention mask \\(\\mathcal{M}\\) to get the \\(\\pmb{x}(t)\\) at timestep \\(t\\):" + }, + { + "type": "equation", + "bbox": [ + 0.1, + 0.57, + 0.484, + 0.589 + ], + "angle": 0, + "content": "\\[\n\\boldsymbol {x} (t) = \\mathcal {M} \\odot \\boldsymbol {x} (t) ^ {\\text {k n o w n}} + (1 - \\mathcal {M}) \\odot \\boldsymbol {x} (t) ^ {\\text {u n k n o w n}}, \\tag {8}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.598, + 0.483, + 0.628 + ], + "angle": 0, + "content": "where \\(\\odot\\) is Hadamard product. We can remove the localized strong noise through the Stage 1." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.629, + 0.484, + 0.734 + ], + "angle": 0, + "content": "Stage 2 (\\( t < t_s \\)) When the current timestep \\( t \\) is less than \\( t_s \\), since the noise intensity is the same both inside and outside the attention mask. Therefore, the attention mask \\( \\mathcal{M} \\) is no longer needed. At this point, we can proceed with the standard sampling process of the diffusion model. We can get \\( x(t) \\) predicted from the previous timestep \\( x(t + 1) \\) as follows:" + }, + { + "type": "equation", + "bbox": [ + 0.102, + 0.745, + 0.483, + 0.763 + ], + "angle": 0, + "content": "\\[\n\\boldsymbol {x} (t) \\sim \\mathcal {N} \\left(\\mu_ {\\theta} (\\boldsymbol {x} (t + 1), t + 1), \\Sigma_ {\\theta} (\\boldsymbol {x} (t + 1), t + 1)\\right) \\tag {9}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.772, + 0.484, + 0.833 + ], + "angle": 0, + "content": "Following Stage 2, we will eliminate global weak noise and get the clean image. The complete algorithm process can be found in the Algorithm 1. Additionally, there are some visualizations of our method in the Figure 5." + }, + { + "type": "title", + "bbox": [ + 0.09, + 0.846, + 0.462, + 0.863 + ], + "angle": 0, + "content": "4. Adaptive Attack to Diffusion Purification" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.871, + 0.484, + 0.902 + ], + "angle": 0, + "content": "When directly applying the method described in 3.2, we observe the same phenomenon as [25] that although the region" + }, + { + "type": "title", + "bbox": [ + 0.516, + 0.413, + 0.831, + 0.428 + ], + "angle": 0, + "content": "Algorithm 1 Heterogeneous Denoising Process" + }, + { + "type": "text", + "bbox": [ + 0.516, + 0.432, + 0.906, + 0.462 + ], + "angle": 0, + "content": "Require: Sample \\( x_0 \\), Attention Mask \\( \\mathcal{M} \\), timestep \\( t_s \\) and \\( t_l \\), Diffusion model parameters \\( \\theta \\)." + }, + { + "type": "text", + "bbox": [ + 0.53, + 0.463, + 0.657, + 0.477 + ], + "angle": 0, + "content": "for \\(t = t_l,\\dots,1\\) do" + }, + { + "type": "equation", + "bbox": [ + 0.547, + 0.478, + 0.674, + 0.492 + ], + "angle": 0, + "content": "\\[\n\\text {i f} t _ {s} \\leq t \\leq t _ {l} \\text {t h e n}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.837, + 0.478, + 0.9, + 0.492 + ], + "angle": 0, + "content": "Stage 1" + }, + { + "type": "equation", + "bbox": [ + 0.565, + 0.493, + 0.778, + 0.508 + ], + "angle": 0, + "content": "\\[\n\\epsilon \\sim \\mathcal {N} (0, \\mathbf {I}) \\text {i f} t > 1, \\text {e l s e} \\epsilon = 0\n\\]" + }, + { + "type": "equation", + "bbox": [ + 0.565, + 0.508, + 0.803, + 0.523 + ], + "angle": 0, + "content": "\\[\nx _ {t - 1} ^ {\\text {k n o w n}} = \\sqrt {\\overline {{\\alpha}} _ {t - 1}} x _ {0} + \\sqrt {1 - \\overline {{\\alpha}} _ {t - 1}} \\epsilon\n\\]" + }, + { + "type": "equation", + "bbox": [ + 0.566, + 0.523, + 0.781, + 0.538 + ], + "angle": 0, + "content": "\\[\nz \\sim \\mathcal {N} (0, \\mathbf {I}) \\text {i f} t > 1, \\text {e l s e} z = 0\n\\]" + }, + { + "type": "equation", + "bbox": [ + 0.565, + 0.538, + 0.884, + 0.562 + ], + "angle": 0, + "content": "\\[\nx _ {t - 1} ^ {u n k n o w n} = \\frac {1}{\\sqrt {\\alpha_ {t}}} \\left(x _ {t} - \\frac {\\beta_ {t}}{\\sqrt {1 - \\bar {\\alpha} _ {t}}} \\epsilon_ {\\theta} (x _ {t}, t)\\right) + \\sigma_ {t} z\n\\]" + }, + { + "type": "equation", + "bbox": [ + 0.565, + 0.562, + 0.874, + 0.577 + ], + "angle": 0, + "content": "\\[\nx _ {t - 1} = \\mathcal {M} \\odot x _ {t - 1} ^ {\\text {k n o w n}} + (1 - \\mathcal {M}) \\odot x _ {t - 1} ^ {\\text {u n k n o w n}}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.547, + 0.577, + 0.671, + 0.591 + ], + "angle": 0, + "content": "else if \\( t < t_s \\) then" + }, + { + "type": "text", + "bbox": [ + 0.84, + 0.578, + 0.906, + 0.592 + ], + "angle": 0, + "content": "Stage 2" + }, + { + "type": "equation", + "bbox": [ + 0.565, + 0.593, + 0.781, + 0.607 + ], + "angle": 0, + "content": "\\[\nz \\sim \\mathcal {N} (0, \\mathbf {I}) \\text {i f} t > 1, \\text {e l s e} z = 0\n\\]" + }, + { + "type": "equation", + "bbox": [ + 0.565, + 0.607, + 0.85, + 0.63 + ], + "angle": 0, + "content": "\\[\nx _ {t - 1} = \\frac {1}{\\sqrt {\\alpha_ {t}}} \\left(x _ {t} - \\frac {\\beta_ {t}}{\\sqrt {1 - \\bar {\\alpha} _ {t}}} \\epsilon_ {\\theta} (x _ {t}, t)\\right) + \\sigma_ {t} z\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.548, + 0.63, + 0.686, + 0.643 + ], + "angle": 0, + "content": "for \\(u = 1,2,\\ldots U\\) do" + }, + { + "type": "text", + "bbox": [ + 0.807, + 0.63, + 0.901, + 0.644 + ], + "angle": 0, + "content": "\\(\\triangleright\\) Resampling" + }, + { + "type": "equation", + "bbox": [ + 0.565, + 0.644, + 0.781, + 0.658 + ], + "angle": 0, + "content": "\\[\nz \\sim \\mathcal {N} (0, \\mathbf {I}) \\text {i f} t > 1, \\text {e l s e} z = 0\n\\]" + }, + { + "type": "equation", + "bbox": [ + 0.565, + 0.659, + 0.872, + 0.675 + ], + "angle": 0, + "content": "\\[\nx _ {t - 1 + u} = \\sqrt {\\alpha_ {t - 1 + u}} x _ {t + u - 2} + \\sqrt {1 - \\alpha_ {t - 1 + u}} z\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.548, + 0.676, + 0.602, + 0.687 + ], + "angle": 0, + "content": "end for" + }, + { + "type": "text", + "bbox": [ + 0.548, + 0.69, + 0.678, + 0.705 + ], + "angle": 0, + "content": "get \\(x_{t - 1}\\) using (11)" + }, + { + "type": "text", + "bbox": [ + 0.532, + 0.706, + 0.586, + 0.717 + ], + "angle": 0, + "content": "end for" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.745, + 0.906, + 0.821 + ], + "angle": 0, + "content": "recovered in Stage 2 matches the texture of the neighboring region, it is semantically incorrect. We can observe this phenomenon in the Figure 3. [25] addresses this issue via multi-step resampling by diffusing the output \\( \\pmb{x}(t) \\) back to \\( \\pmb{x}(t + 1) \\):" + }, + { + "type": "equation", + "bbox": [ + 0.577, + 0.83, + 0.905, + 0.848 + ], + "angle": 0, + "content": "\\[\n\\boldsymbol {x} (t + 1) = \\sqrt {\\alpha (t)} \\boldsymbol {x} (t) + \\sqrt {1 - \\alpha (t)} \\boldsymbol {\\epsilon}, \\tag {10}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.856, + 0.906, + 0.902 + ], + "angle": 0, + "content": "and it is necessary to fully execute 3.2 again to get \\( \\pmb{x}(t) \\). From Figure 3 we can find that this process needs to be repeated \\( U = 20 \\) times to ensure that the masked region" + }, + { + "type": "page_number", + "bbox": [ + 0.479, + 0.945, + 0.518, + 0.957 + ], + "angle": 0, + "content": "29271" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.108, + 0.089, + 0.473, + 0.234 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.09, + 0.244, + 0.483, + 0.302 + ], + "angle": 0, + "content": "Figure 3. Our method maintains semantic consistency between the masked areas and the surrounding pixel values when \\( U = 10 \\). Note that no matter what \\( U \\) is, our method requires only one additional call to the denoising network." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.328, + 0.484, + 0.433 + ], + "angle": 0, + "content": "becomes semantically consistent with the surrounding pixels. To evaluate our method against strong adaptive attacks [30], it is necessary to compute the full gradient of the whole sampling process to construct adversarial examples. To reduce the time cost and memory usage, we optimized the multi-step resampling into a single resampling process which achieves the same functionality." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.434, + 0.484, + 0.495 + ], + "angle": 0, + "content": "The intuition behind our improvement is that repeating (10) and 3.2 for \\(U\\) times can be replaced by that first diffuse \\(\\pmb{x}(t)\\) to \\(\\pmb{x}(t + U)\\) and then denoise it back to \\(\\pmb{x}(t)\\) from \\(\\pmb{x}(t + U)\\) in a single step using DDIM [39] as follows:" + }, + { + "type": "equation", + "bbox": [ + 0.097, + 0.507, + 0.482, + 0.58 + ], + "angle": 0, + "content": "\\[\n\\begin{array}{l} \\boldsymbol {x} _ {t - 1} = \\sqrt {\\alpha_ {t - 1}} \\left(\\frac {\\boldsymbol {x} _ {t - 1 + U} - \\sqrt {1 - \\alpha_ {t - 1 + U}} \\epsilon_ {\\theta} (\\boldsymbol {x} _ {t - 1 + U})}{\\sqrt {\\alpha_ {t - 1 + U}}}\\right) \\\\ + \\sqrt {1 - \\alpha_ {t - 1} - \\sigma_ {t - 1 + U} ^ {2}} \\cdot \\epsilon_ {\\theta} \\left(\\boldsymbol {x} _ {t - 1 + U}\\right) + \\sigma_ {t - 1 + U} \\boldsymbol {\\epsilon}. \\tag {11} \\\\ \\end{array}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.582, + 0.484, + 0.734 + ], + "angle": 0, + "content": "From the bottom panel of Figure 3, we find that our method can also achieve semantically consistency when \\( U = 10 \\). This means that at each timestep, we only need to invoke the denoising network an additional time. Therefore, the GPU memory needs to store the \\( 2 * T \\) complete computation graphs of the denoising network which enables the computation of the gradient of our method on consumer-grade GPU with only 24G memory. Also, the time cost is greatly reduced. The analysis of time cost and memory usage can refer to the Appendix." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.748, + 0.216, + 0.765 + ], + "angle": 0, + "content": "5. Experiment" + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.773, + 0.298, + 0.79 + ], + "angle": 0, + "content": "5.1. Experimental Settings" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.795, + 0.484, + 0.901 + ], + "angle": 0, + "content": "Datasets and network architectures. Three datasets is used to evaluate including CIFAR-10, SVHN, and ImageNet [6]. We compare our results with several leading defense methods as listed in the standardized benchmark RobustBench [5] for CIFAR-10 and ImageNet, and also test different adversarial purification methods. We use two common classifier architectures: WideResNet-28-10" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.092, + 0.905, + 0.136 + ], + "angle": 0, + "content": "and WideResNet-70-16 [51] for CIFAR-10. For SVHN, WideResNet-28-10 serves as the backbone, while for ImageNet, ResNet-50 [14] is used as the backbone." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.137, + 0.907, + 0.332 + ], + "angle": 0, + "content": "Adversarial attack methods. Strong adaptive attack is tested for our method and other adversarial purification methods. We use the commonly known AutoAttack [4] under \\(\\ell_{\\infty}\\) and \\(\\ell_{2}\\) threat models. Additionally, the projected gradient descent (PGD) attack [26] is evaluated on our method, as recommended in [20]. Due to the randomness introduced by the diffusion and denoising processes, Expectation Over Time (EOT) [1] is adapted for these adaptive attacks. Moreover, we apply the BPDA+EOT [15] attack to ensure a fair comparison with other adversarial purification methods. Finally, following the suggestion of [20], a surrogate process is used to obtain the gradient of the reverse process for white-box attacks." + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.333, + 0.906, + 0.438 + ], + "angle": 0, + "content": "Pre-trained Models. We use the unconditional CIFAR-10 checkpoint of EDM provided by NVIDIA [19] on CIFAR-10. 256x256 diffusion (unconditional) checkpoint from guided-diffusion library is adopted for the ImageNet experiments. For CIFAR-10, the pre-trained classifier is downloaded from RobustBench [5], while for ImageNet, the weight of the classifier is provided by PyTorch library." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.439, + 0.906, + 0.574 + ], + "angle": 0, + "content": "Evaluation metrics. To assess the performance of defense methods, we use two metrics: standard accuracy, measured on clean samples, and robust accuracy, measured on adversarial examples. Due to the high computational cost of evaluating models against adaptive attacks, a fixed subset of 512 images are randomly sampled from the test set for robust evaluation, which is the same as [20, 30]. In all experiments, the mean and standard deviation over three runs are reported to assess both standard and robust accuracy." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.575, + 0.906, + 0.71 + ], + "angle": 0, + "content": "Implementation details. We follow the settings described in [20]. We evaluate diffusion-based purification methods using the PGD+EOT attack with 200 update iterations, and BPDA and AutoAttack with 100 update iterations, except for ImageNet, which uses 20 iterations. We set the number of EOT to 20. The step size is 0.007. For randomized defenses, such as [3, 20, 30, 40], we use the random version of AutoAttack, and for static defenses, we use the standard version." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.726, + 0.716, + 0.741 + ], + "angle": 0, + "content": "5.2. Experimental Results" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.75, + 0.907, + 0.901 + ], + "angle": 0, + "content": "CIFAR-10. We conducted comprehensive experiments on the CIFAR-10 dataset, evaluating our method against other approaches on two model architectures: WideResNet28-10 and WideResNet-70-16. We assessed robust accuracy under three attack types: BPDA+EOT, PGD+EOT, and AutoAttack. We tested against the infinity norm attack, and as shown in Table 1, when the backbone is WideResNet28-10, our method significantly outperforms other baselines in both standard accuracy and robust accuracy, exceeding by \\(3.00\\%\\) in standard accuracy and leading by \\(3.56\\%\\) and" + }, + { + "type": "page_number", + "bbox": [ + 0.479, + 0.945, + 0.521, + 0.958 + ], + "angle": 0, + "content": "29272" + } + ], + [ + { + "type": "table_caption", + "bbox": [ + 0.09, + 0.089, + 0.486, + 0.161 + ], + "angle": 0, + "content": "Table 1. Standard and robust accuracy of different Adversarial Training (AT) and Adversarial Purification (AP) methods against PGD+EOT and AutoAttack \\(\\ell_{\\infty}(\\epsilon = 8 / 255)\\) on CIFAR-10. * utilizes half number of iterations for the attack due to the high computational cost. † indicates the requirement of extra data." + }, + { + "type": "table", + "bbox": [ + 0.093, + 0.17, + 0.489, + 0.381 + ], + "angle": 0, + "content": "
TypeMethodStandard Acc.Robust Acc.
PGDAutoAttack
WideResNet-28-10
ATGowal et al. [13]88.5465.9363.38
Gowal et al. [11]†87.5166.0162.76
Pang et al. [31]88.6264.9561.04
APYoon et al. [50]85.66±0.5133.48±0.8659.53±0.87
Nie et al. [30]90.07±0.9756.84±0.5963.6±0.81
Lee et al. [20]90.16±0.6455.82±0.5970.47±1.53
Bai et al. [3]91.4149.22*77.08
Geigh et al. [53]84.20-59.14
Lin et al. [24]90.62-72.85
Ours93.16±0.2159.38±1.2580.45±1.54
" + }, + { + "type": "table_caption", + "bbox": [ + 0.09, + 0.4, + 0.486, + 0.482 + ], + "angle": 0, + "content": "Table 2. Standard and robust accuracy of different Adversarial Training (AT) and Adversarial Purification (AP) methods against PGD+EOT and AutoAttack \\(\\ell_{\\infty}(\\epsilon = 8 / 255)\\) on CIFAR-10. * The number of iterations for the attack is half that of the other methods for less computational overhead. † indicates the requirement of extra data." + }, + { + "type": "table", + "bbox": [ + 0.093, + 0.495, + 0.492, + 0.719 + ], + "angle": 0, + "content": "
TypeMethodStandard Acc.Robust Acc.
PGDAutoAttack
WideResNet-70-16
ATRebuffi et al. [34]†92.2269.9766.56
Gowal et al. [10]†91.1068.6666.10
Gowal et al. [12]88.7569.0365.87
APYoon et al. [50]86.76±1.1537.11±1.3560.86±0.56
Nie et al. [30]90.43±0.6051.13±0.8766.06±1.17
Lee et al. [20]90.53±0.156.88±1.0670.31±0.62
Bai et al. [3]92.9748.83*79.10
Geigh et al. [53]84.60-66.40
Geigh et al. [53]86.90-59.20
Lin et al. [24]91.99-76.37
Ours93.36±0.5959.79±0.6184.83±0.59
" + }, + { + "type": "text", + "bbox": [ + 0.089, + 0.75, + 0.486, + 0.903 + ], + "angle": 0, + "content": "3.37% under PGD and AutoAttack, respectively. From Table 2, we see that when WideResNet-70-16 is used as the backbone, our method exceeds the state-of-the-art (SOTA) by 1.37% in standard accuracy and leads by 2.91% and 5.73% under PGD and AutoAttack, respectively. We also evaluated the accuracy of our method and other baselines under the Euclidean norm attack. From Table 3, we observe that, with WideResNet-28-10 as the backbone, although our method lags behind [3] by 0.59% in standard accuracy, it leads by 0.98% and 3.77% in robust accuracy. Due to the" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.089, + 0.908, + 0.189 + ], + "angle": 0, + "content": "Table 3. Standard and robust accuracy against PGD+EOT and AutoAttack \\(\\ell_2(\\epsilon = 0.5)\\) on CIFAR-10. Adversarial Training (AT) and Adversarial Purification (AP) methods are evaluated. *The number of iterations for the attack is half that of the other methods for less computational overhead. † indicates the requirement of extra data. ‡ adopts WideResNet-34-10 as the backbone, with the same width but more layers than the default one." + }, + { + "type": "table", + "bbox": [ + 0.517, + 0.198, + 0.905, + 0.423 + ], + "angle": 0, + "content": "
TypeMethodStandard Acc.Robust Acc.
PGDAutoAttack
WideResNet-28-10
ATRebuffi et al. [34]†91.7985.0578.80
Augustin et al. [2]‡93.9686.1478.79
Sehwag et al. [31]‡90.9383.7577.24
APYoon et al. [50]85.66±0.5173.32±0.7679.57±0.38
Nie et al. [30]91.41±1.0079.45±1.1681.7±0.84
Lee et al [20]90.16±0.6483.59±0.8886.48±0.38
Bai et al. [3]93.7586.13*80.92
Geigh et al. [53]84.40-77.90
Geigh et al. [53]84.20-73.60
Lin et al. [24]90.62-80.47
Ours93.16±0.2187.11±0.9590.25±0.80
" + }, + { + "type": "table_caption", + "bbox": [ + 0.512, + 0.441, + 0.909, + 0.514 + ], + "angle": 0, + "content": "Table 4. Standard and robust accuracy against PGD+EOT and AutoAttack \\(\\ell_2(\\epsilon = 0.5)\\) on CIFAR-10. Adversarial Training (AT) and Adversarial Purification (AP) methods are evaluated. (*The number of iterations for the attack is half that of other methods for less computational overhead. † methods need extra data.)" + }, + { + "type": "table", + "bbox": [ + 0.516, + 0.522, + 0.909, + 0.72 + ], + "angle": 0, + "content": "
TypeMethodStandard Acc.Robust Acc.
PGDAutoAttack
WideResNet-70-16
ATRebuffi et al. [34]†95.7489.6282.32
Gowal et al. [10]†94.7488.1880.53
Rebuffi et al. [34]92.4186.2480.42
APYoon et al. [50]86.76±1.1575.666±1.2980.43±0.42
Nie et al. [30]92.15±0.7282.97±1.3883.06±1.27
Lee et al [20]90.53±0.1483.75±0.9985.59±0.61
Bai et al. [3]92.9784.37*83.01
Lin et al. [24]91.99-81.35
Ours93.36±0.5987.7±0.9791.71±1.09
" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.75, + 0.911, + 0.903 + ], + "angle": 0, + "content": "significant time overhead of [3], which only iterates 100 times under PGD—half of what others do—our method outperforms more under PGD attacks. From Table 4, we see that when WideResNet-70-16 is the backbone, our method outperforms other baselines, leading by \\(1.21\\%\\) in standard accuracy and by \\(3.33\\%\\) and \\(6.12\\%\\) in robust accuracy under PGD and AutoAttack, respectively. Additionally, we applied the BPDA+EOT attack, which approximates differentiability. From Table 5, we observe that although robust accuracy lags behind [40] by \\(1.71\\%\\), our method still leads" + }, + { + "type": "page_number", + "bbox": [ + 0.479, + 0.945, + 0.521, + 0.958 + ], + "angle": 0, + "content": "29273" + } + ], + [ + { + "type": "table_caption", + "bbox": [ + 0.091, + 0.089, + 0.483, + 0.118 + ], + "angle": 0, + "content": "Table 5. Standard and robust accuracy against BPDA+EOT \\(\\ell_{\\infty}(\\epsilon = 8 / 255)\\) on CIFAR-10." + }, + { + "type": "table", + "bbox": [ + 0.101, + 0.129, + 0.477, + 0.339 + ], + "angle": 0, + "content": "
MethodPurificationAccuracy
StandardRobust
WideResNet-28-10
Song et al. [41]Gibbs Update95.009.00
Yang et al. [48]Mask+Recon94.0015.00
Hill et al. [15]EBM+LD84.1254.90
Yoon et al. [50]DSM+LD85.66±0.5166.91±1.75
Nie et al. [30]Diffusion90.07±0.9781.45±1.51
Bai et al. [3]Diffusion91.37±1.2185.51±0.81
Wang et al. [44]Diffusion89.96±0.4075.59±1.26
Song et al. [40]Diffusion89.88±0.3588.43±0.83
Lee et al. [20]Diffusion90.16±0.6488.40±0.88
OursDiffusion93.04±0.6586.72±0.20
" + }, + { + "type": "text", + "bbox": [ + 0.089, + 0.365, + 0.485, + 0.548 + ], + "angle": 0, + "content": "in standard accuracy by \\(3.16\\%\\), resulting in an average advantage of \\(0.73\\%\\). We also measured the distance between purified samples and the original samples in feature space, shown in the Appendix. A smaller distance indicates that the purified samples retain more consistent semantic information with the original samples (further discussion in the appendix). From Figure 6, we find that the distribution of purified samples by our method is most similar to the original samples compared to other methods. Overall, our method outperforms others, demonstrating its effectiveness in preserving semantic information while eliminating adversarial perturbations." + }, + { + "type": "table_caption", + "bbox": [ + 0.09, + 0.561, + 0.483, + 0.592 + ], + "angle": 0, + "content": "Table 6. Standard and robust accuracy against PGD+EOT \\(\\ell_{\\infty}(\\epsilon = 8 / 255)\\) on ImageNet. ResNet-50 is used as the classifier." + }, + { + "type": "table", + "bbox": [ + 0.105, + 0.601, + 0.472, + 0.826 + ], + "angle": 0, + "content": "
TypeMethodAccuracy
StandardRobust
ResNet-50
ATSalman et al. [35]63.8639.11
Engstrom et al. [7]62.4233.20
Wong et al. [46]53.8328.04
APNie et al. (t* = 0.2) [30]73.9640.63
Nie et al. (t* = 0.3) [30]72.8548.24
Nie et al. (t* = 0.4) [30]61.7141.67
Bai et al. [3]70.4141.7
Lee et al. [20]71.4246.59
Song et al. [40]62.2551.14
Geigh et al. [53]73.9856.54
Ours74.51±1.0760.06±1.66
" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.84, + 0.485, + 0.903 + ], + "angle": 0, + "content": "ImageNet. We test ResNet50 as the backbone on the ImageNet dataset under PGD+EOT attacks, which is the same as [3, 20, 30]. From Table 6, our method leads competitors by \\(0.53\\%\\) in standard accuracy and by \\(3.52\\%\\) in" + }, + { + "type": "image", + "bbox": [ + 0.518, + 0.09, + 0.707, + 0.253 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.725, + 0.09, + 0.918, + 0.253 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.513, + 0.263, + 0.908, + 0.361 + ], + "angle": 0, + "content": "Figure 4. 2D purification trajectories of Our method and DiffPure. For DiffPure (left), if \\( t^{*} = 0.2 \\), although the purification direction is toward the original class, the adversarial perturbation cannot be completely eliminated; If \\( t^{*} = 0.4 \\), the semantic information changes and the purification direction is no longer toward the original class. Our method (right) can eliminate adversarial perturbation and remain the semantic information." + }, + { + "type": "image", + "bbox": [ + 0.519, + 0.373, + 0.922, + 0.564 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.513, + 0.574, + 0.908, + 0.603 + ], + "angle": 0, + "content": "Figure 5. Visualization of attention map, attention mask and images purified by our method." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.629, + 0.909, + 0.903 + ], + "angle": 0, + "content": "robust accuracy, significantly surpassing the baseline. Additionally, we observe that the method from [30] achieves a standard accuracy of \\(73.96\\%\\) and a robust accuracy of \\(40.63\\%\\) when \\(t^{*} = 0.2\\). When \\(t^{*} = 0.3\\), standard accuracy drops to \\(72.85\\%\\) while robust accuracy rises to \\(48.24\\%\\). However, at \\(t^{*} = 0.4\\), both standard accuracy and robust accuracy decline to \\(61.71\\%\\) and \\(41.67\\%\\), respectively. Additionally, We visualize the 2D purification trajectories of our method and DiffPure using the T-SNE [43] technique. From Figure 4, we observe that when \\(t^{*} = 0.2\\), the adversarial perturbation on the adversarial sample has not been fully purified, and it remains at the boundary between the two classes of samples. When \\(t^{*} = 0.4\\), the semantic information of the adversarial example begins to change, and as a result, the purification trajectory no longer points toward the original classes. This indicates that applying the same noise intensity across an image cannot effectively balance semantic preservation and the elimination of adversarial perturbation." + }, + { + "type": "page_number", + "bbox": [ + 0.479, + 0.945, + 0.521, + 0.958 + ], + "angle": 0, + "content": "29274" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.092, + 0.089, + 0.287, + 0.246 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.306, + 0.089, + 0.498, + 0.245 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.088, + 0.256, + 0.287, + 0.416 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.305, + 0.258, + 0.498, + 0.415 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.09, + 0.429, + 0.484, + 0.471 + ], + "angle": 0, + "content": "Figure 6. Joint distribution of the original samples and purified samples. The distributions of the purified samples by our method and the original samples are the most similar." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.501, + 0.484, + 0.653 + ], + "angle": 0, + "content": "tions. In contrast, our method adds varying noise intensities to different regions of an image, applying stronger noise to areas with higher classification weights to eliminate adversarial perturbations, while applying lighter noise to other regions to preserve semantic information. This approach allows our method to achieve good results in both robust accuracy and standard accuracy simultaneously. Our method achieves comparable standard accuracy to [30] at \\( t^* = 0.2 \\) but significantly outperforms it in robust accuracy, demonstrating the effectiveness of our approach." + }, + { + "type": "image", + "bbox": [ + 0.095, + 0.671, + 0.293, + 0.817 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.305, + 0.671, + 0.498, + 0.817 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.09, + 0.838, + 0.486, + 0.88 + ], + "angle": 0, + "content": "Figure 7. Robust and Standard Accuracy under different thresholds \\(\\tau\\) (left) and under different combinations of timesteps \\(t_l\\) and \\(t_s\\) (right)." + }, + { + "type": "title", + "bbox": [ + 0.514, + 0.091, + 0.667, + 0.108 + ], + "angle": 0, + "content": "5.3. Ablation Study" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.114, + 0.907, + 0.416 + ], + "angle": 0, + "content": "We first conducted a hyperparameter sensitivity analysis. Our method involves three hyperparameters, two of which are noises of different intensities, \\( t_{l} \\) and \\( t_{s} \\), and the third is a threshold, \\( \\tau \\), used to control the size of the attention mask. From Figure 7, we observe that as \\( \\tau \\) increases (resulting in smaller mask areas), the Standard Accuracy also consistently increases. However, adversarial perturbations are preserved at this time, leading to poorer Robust Accuracy when \\( \\tau \\) is large. When \\( \\tau \\) is small, semantic information is damaged, which also results in weaker Robust Accuracy. Therefore, we choose \\( \\tau \\) to be around 0.8. Simultaneously, we can see that when \\( t_{l} = 0.2 \\), \\( t_{s} = 0.05 \\), the Robust Accuracy is the highest. Here, \\( t_{s} \\) is less than the \\( t^{*} \\) used in [30] to preserve more semantic information, while \\( t_{l} \\) is greater than \\( t^{*} \\) to eliminate adversarial perturbations. From Table 6, we can see that when we set \\( t_{l} = 0.4 \\) and \\( t_{s} = 0.2 \\), The Standard Accuracy is almost as high as the \\( t^{*} = 0.2 \\) of [30], but at the same time, our Robust Accuracy significantly leads. This demonstrates the effectiveness of our method. We also explore whether it is necessary for each block" + }, + { + "type": "table_caption", + "bbox": [ + 0.513, + 0.431, + 0.907, + 0.459 + ], + "angle": 0, + "content": "Table 7. Standard Accuracy and Robust Accuracy for different block combinations. WideResNet28-10 servers as the backbone." + }, + { + "type": "table", + "bbox": [ + 0.548, + 0.47, + 0.875, + 0.587 + ], + "angle": 0, + "content": "
Encoder BlockAccuracy
E1E2E3E4StandardRobustAverage
X✓✓✓92.7758.0075.38
✓X✓✓93.1658.0075.58
✓✓X✓93.3557.6175.48
✓✓✓X91.6658.8575.25
✓✓✓✓93.1659.3876.27
" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.602, + 0.907, + 0.693 + ], + "angle": 0, + "content": "to participate in constructing the attention mask. Taking WideResNet28-10 as an example, it can be divided into four blocks. We extract the output of the final ReLU function of each block as the activation map. From Table 7, we can see that removing any block results in a performance loss. Therefore, every block should be utilized." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.709, + 0.634, + 0.724 + ], + "angle": 0, + "content": "6. Conclusion" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.735, + 0.907, + 0.901 + ], + "angle": 0, + "content": "We propose a novel forward process that adds noise of varying intensities based on the classifier's attention to different regions, achieving a balance between preserving semantic information and eliminating adversarial perturbations. The denoising process is tailored to our method to remove different-intensity noise within a noisy image. By replacing the multi-step resampling with an improved single-step resampling, time cost and memory usage are reduced significantly. Extensive experiments on three datasets show that our method outperforms other methods by a promising improvement against various adversarial attacks." + }, + { + "type": "page_number", + "bbox": [ + 0.479, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "29275" + } + ], + [ + { + "type": "title", + "bbox": [ + 0.092, + 0.091, + 0.21, + 0.108 + ], + "angle": 0, + "content": "Acknowledge" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.114, + 0.486, + 0.27 + ], + "angle": 0, + "content": "This work was supported in part by the National Key R&D Program of China under Grant 2022ZD0119204, in part by National Natural Science Foundation of China: 62236008, 62441232, U21B2038, U23B2051, 62122075 and 62376257, in part by Youth Innovation Promotion Association CAS, in part by the Strategic Priority Research Program of the Chinese Academy of Sciences, Grant No. XDB0680201, in part by the Fundamental Research Funds for the Central Universities, in part by the Tencent Marketing Solution Rhino-Bird Focused Research Program, in part by Xiaomi Young Talents Program." + }, + { + "type": "title", + "bbox": [ + 0.092, + 0.295, + 0.188, + 0.311 + ], + "angle": 0, + "content": "References" + }, + { + "type": "ref_text", + "bbox": [ + 0.1, + 0.32, + 0.482, + 0.376 + ], + "angle": 0, + "content": "[1] Anish Athalye, Nicholas Carlini, and David Wagner. Obfuscated gradients give a false sense of security: Circumventing defenses to adversarial examples. In International conference on machine learning, pages 274-283. PMLR, 2018. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.1, + 0.379, + 0.484, + 0.434 + ], + "angle": 0, + "content": "[2] Maximilian Augustin, Alexander Meinke, and Matthias Hein. Adversarial robustness on in-and out-distribution improves explainability. In European Conference on Computer Vision, pages 228-245. Springer, 2020. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.101, + 0.435, + 0.484, + 0.503 + ], + "angle": 0, + "content": "[3] Mingyuan Bai, Wei Huang, Tenghui Li, Andong Wang, Junbin Gao, Cesar F Caiafa, and Qibin Zhao. Diffusion models demand contrastive guidance for adversarial purification to advance. In *Forty-first International Conference on Machine Learning*, 2024. 2, 3, 5, 6, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.505, + 0.484, + 0.56 + ], + "angle": 0, + "content": "[4] Francesco Croce and Matthias Hein. Reliable evaluation of adversarial robustness with an ensemble of diverse parameter-free attacks. In International conference on machine learning, pages 2206-2216, 2020. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.562, + 0.484, + 0.631 + ], + "angle": 0, + "content": "[5] Francesco Croce, Maksym Andriushchenko, Vikash Sehwag, Edoardo Debenedetti, Nicolas Flammarion, Mung Chiang, Prateek Mittal, and Matthias Hein. Robustbench: a standardized adversarial robustness benchmark. arXiv preprint arXiv:2010.09670, 2020. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.633, + 0.484, + 0.688 + ], + "angle": 0, + "content": "[6] Jia Deng, Wei Dong, Richard Socher, Li-Jia Li, Kai Li, and Li Fei-Fei. Imagenet: A large-scale hierarchical image database. In 2009 IEEE conference on computer vision and pattern recognition, pages 248–255, 2009. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.689, + 0.484, + 0.743 + ], + "angle": 0, + "content": "[7] Logan Engstrom, Andrew Ilyas, Hadi Salman, Shibani Santurkar, and Dimitris Tsipras. Robustness (python library), 2019. URL https://github.com/MadryLab/robustness, 4(4): 4-3, 2019. 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.746, + 0.484, + 0.801 + ], + "angle": 0, + "content": "[8] Xavier Glorot, Antoine Bordes, and Yoshua Bengio. Deep sparse rectifier neural networks. In Proceedings of the fourteenth international conference on artificial intelligence and statistics, pages 315-323, 2011. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.803, + 0.484, + 0.845 + ], + "angle": 0, + "content": "[9] Ian J Goodfellow, Jonathon Shlens, and Christian Szegedy. Explaining and harnessing adversarial examples. arXiv preprint arXiv:1412.6572, 2014. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.846, + 0.484, + 0.901 + ], + "angle": 0, + "content": "[10] Sven Gowal, Chongli Qin, Jonathan Uesato, Timothy Mann, and Pushmeet Kohli. Uncovering the limits of adversarial training against norm-bounded adversarial examples. arXiv preprint arXiv:2010.03593, 2020. 6" + }, + { + "type": "list", + "bbox": [ + 0.094, + 0.32, + 0.484, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.093, + 0.907, + 0.147 + ], + "angle": 0, + "content": "[11] Sven Gowal, Chongli Qin, Jonathan Uesato, Timothy Mann, and Pushmeet Kohli. Uncovering the limits of adversarial training against norm-bounded adversarial examples. arXiv preprint arXiv:2010.03593, 2020. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.15, + 0.907, + 0.205 + ], + "angle": 0, + "content": "[12] Sven Gowal, Sylvestre-Alvise Rebuffi, Olivia Wiles, Florian Stimberg, Dan Andrei Calian, and Timothy A Mann. Improving robustness using generated data. Advances in Neural Information Processing Systems, 34:4218-4233, 2021. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.207, + 0.907, + 0.261 + ], + "angle": 0, + "content": "[13] Sven Gowal, Sylvestre-Alvise Rebuffi, Olivia Wiles, Florian Stimberg, Dan Andrei Calian, and Timothy A Mann. Improving robustness using generated data. Advances in Neural Information Processing Systems, 34:4218-4233, 2021. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.264, + 0.906, + 0.319 + ], + "angle": 0, + "content": "[14] Kaiming He, Xiangyu Zhang, Shaoqing Ren, and Jian Sun. Deep residual learning for image recognition. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 770-778, 2016. 3, 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.321, + 0.906, + 0.375 + ], + "angle": 0, + "content": "[15] Mitch Hill, Jonathan Craig Mitchell, and Song-Chun Zhu. Stochastic security: Adversarial defense using long-run dynamics of energy-based models. In International Conference on Learning Representations, 2021. 5, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.378, + 0.906, + 0.419 + ], + "angle": 0, + "content": "[16] Jonathan Ho, Ajay Jain, and Pieter Abbeel. Denoising diffusion probabilistic models. Advances in neural information processing systems, 33:6840-6851, 2020. 1, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.421, + 0.906, + 0.462 + ], + "angle": 0, + "content": "[17] Duhun Hwang, Eunjung Lee, and Wonjong Rhee. Aidpurifier: A light auxiliary network for boosting adversarial defense. Neurocomputing, 541:126251, 2023. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.464, + 0.906, + 0.518 + ], + "angle": 0, + "content": "[18] Xiaojun Jia, Yong Zhang, Xingxing Wei, Baoyuan Wu, Ke Ma, Jue Wang, and Xiaochun Cao. Improving fast adversarial training with prior-guided knowledge. IEEE Transactions on Pattern Analysis and Machine Intelligence, 2024. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.521, + 0.906, + 0.574 + ], + "angle": 0, + "content": "[19] Tero Karras, Miika Aittala, Timo Aila, and Samuli Laine. Elucidating the design space of diffusion-based generative models. Advances in neural information processing systems, 35:26565-26577, 2022. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.577, + 0.907, + 0.632 + ], + "angle": 0, + "content": "[20] Minjong Lee and Dongwoo Kim. Robust evaluation of diffusion-based adversarial purification. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 134-144, 2023. 5, 6, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.634, + 0.906, + 0.703 + ], + "angle": 0, + "content": "[21] Siyuan Liang, Xingxing Wei, Siyuan Yao, and Xiaochun Cao. Efficient adversarial attacks for visual object tracking. In Computer Vision-ECCV 2020: 16th European Conference, Glasgow, UK, August 23-28, 2020, Proceedings, Part XXVI 16, pages 34-50. Springer, 2020. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.705, + 0.906, + 0.773 + ], + "angle": 0, + "content": "[22] Siyuan Liang, Longkang Li, Yanbo Fan, Xiaojun Jia, Jingzhi Li, Baoyuan Wu, and Xiaochun Cao. A large-scale multiple-objective method for black-box attack against object detection. In European Conference on Computer Vision, pages 619-636. Springer, 2022." + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.776, + 0.906, + 0.83 + ], + "angle": 0, + "content": "[23] Siyuan Liang, Baoyuan Wu, Yanbo Fan, Xingxing Wei, and Xiaochun Cao. Parallel rectangle flip attack: A query-based black-box attack against object detection. arXiv preprint arXiv:2201.08970, 2022. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.832, + 0.906, + 0.901 + ], + "angle": 0, + "content": "[24] Guang Lin, Chao Li, Jianhai Zhang, Toshihisa Tanaka, and Qibin Zhao. Adversarial training on purification (ATop): Advancing both robustness and generalization. In The Twelfth International Conference on Learning Representations, 2024. 2, 6" + }, + { + "type": "list", + "bbox": [ + 0.517, + 0.093, + 0.907, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.48, + 0.945, + 0.521, + 0.958 + ], + "angle": 0, + "content": "29276" + } + ], + [ + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.093, + 0.484, + 0.161 + ], + "angle": 0, + "content": "[25] Andreas Lugmayr, Martin Danelljan, Andres Romero, Fisher Yu, Radu Timofte, and Luc Van Gool. Repaint: Inpainting using denoising diffusion probabilistic models. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 11461-11471, 2022. 3, 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.162, + 0.484, + 0.218 + ], + "angle": 0, + "content": "[26] Aleksander Madry, Aleksandar Makelov, Ludwig Schmidt, Dimitris Tsipras, and Adrian Vladu. Towards deep learning models resistant to adversarial attacks. In International Conference on Learning Representations, 2018. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.219, + 0.484, + 0.273 + ], + "angle": 0, + "content": "[27] Dongyu Meng and Hao Chen. Magnet: a two-pronged defense against adversarial examples. In Proceedings of the 2017 ACM SIGSAC conference on computer and communications security, pages 135-147, 2017. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.274, + 0.484, + 0.342 + ], + "angle": 0, + "content": "[28] Muzammal Naseer, Salman Khan, Munawar Hayat, Fahad Shahbaz Khan, and Fatih Porikli. A self-supervised approach for adversarial robustness. In 2020 IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), 2020. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.343, + 0.484, + 0.385 + ], + "angle": 0, + "content": "[29] Alexander Quinn Nichol and Prafulla Dhariwal. Improved denoising diffusion probabilistic models. In International conference on machine learning, 2021. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.386, + 0.484, + 0.453 + ], + "angle": 0, + "content": "[30] Weili Nie, Brandon Guo, Yujia Huang, Chaowei Xiao, Arash Vahdat, and Animashree Anandkumar. Diffusion models for adversarial purification. In Proceedings of the 39th International Conference on Machine Learning, 2022. 1, 2, 3, 5, 6, 7, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.455, + 0.484, + 0.51 + ], + "angle": 0, + "content": "[31] Tianyu Pang, Min Lin, Xiao Yang, Jun Zhu, and Shuicheng Yan. Robustness and accuracy could be reconcilable by (proper) definition. In International Conference on Machine Learning, pages 17258-17277. PMLR, 2022. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.511, + 0.484, + 0.566 + ], + "angle": 0, + "content": "[32] Gaozheng Pei, Ke Ma, Pinci Yang, Qianqian Xu, Yingfei Sun, et al. Exploring query efficient data generation towards data-free model stealing in hard label setting. arXiv preprint arXiv:2412.15276, 2024. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.567, + 0.484, + 0.607 + ], + "angle": 0, + "content": "[33] Prajit Ramachandran, Barret Zoph, and Quoc V Le. Searching for activation functions. arXiv preprint arXiv:1710.05941, 2017. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.608, + 0.484, + 0.664 + ], + "angle": 0, + "content": "[34] Sylvestre-Alvise Rebuffi, Sven Gowal, Dan A Calian, Florian Stimberg, Olivia Wiles, and Timothy Mann. Fixing data augmentation to improve adversarial robustness. arXiv preprint arXiv:2103.01946, 2021. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.665, + 0.484, + 0.719 + ], + "angle": 0, + "content": "[35] Hadi Salman, Andrew Ilyas, Logan Engstrom, Ashish Kapoor, and Aleksander Madry. Do adversarially robust imagenet models transfer better? Advances in Neural Information Processing Systems, 33:3533-3545, 2020. 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.72, + 0.484, + 0.775 + ], + "angle": 0, + "content": "[36] Pouya Samangouei, Maya Kabbab, and Rama Chellappa. Defense-GAN: Protecting classifiers against adversarial attacks using generative models. In International Conference on Learning Representations, 2018. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.776, + 0.484, + 0.844 + ], + "angle": 0, + "content": "[37] Ali Shafahi, Mahyar Najibi, Mohammad Amin Ghiasi, Zheng Xu, John Dickerson, Christoph Studer, Larry S Davis, Gavin Taylor, and Tom Goldstein. Adversarial training for free! Advances in neural information processing systems, 32, 2019. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.845, + 0.484, + 0.9 + ], + "angle": 0, + "content": "[38] Changhao Shi, Chester Holtz, and Gal Mishne. Online adversarial purification based on self-supervised learning. In International Conference on Learning Representations, 2021. 1" + }, + { + "type": "list", + "bbox": [ + 0.093, + 0.093, + 0.484, + 0.9 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.093, + 0.905, + 0.135 + ], + "angle": 0, + "content": "[39] Jiaming Song, Chenlin Meng, and Stefano Ermon. Denoising diffusion implicit models. In International Conference on Learning Representations, 2021. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.136, + 0.905, + 0.205 + ], + "angle": 0, + "content": "[40] Kaiyu Song, Hanjiang Lai, Yan Pan, and Jian Yin. Mimic diffusion: Purifying adversarial perturbation via mimicking clean diffusion model. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), 2024. 1, 2, 5, 6, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.207, + 0.905, + 0.275 + ], + "angle": 0, + "content": "[41] Yang Song, Taesup Kim, Sebastian Nowozin, Stefano Ermon, and Nate Kushman. Pixeldefend: Leveraging generative models to understand and defend against adversarial examples. In International Conference on Learning Representations, 2018. 2, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.277, + 0.905, + 0.331 + ], + "angle": 0, + "content": "[42] Linyu Tang and Lei Zhang. Robust overfitting does matter: Test-time adversarial purification with fgsm. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), 2024. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.333, + 0.905, + 0.374 + ], + "angle": 0, + "content": "[43] Laurens Van der Maaten and Geoffrey Hinton. Visualizing data using t-sne. Journal of machine learning research, 9 (11), 2008. 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.375, + 0.905, + 0.416 + ], + "angle": 0, + "content": "[44] Jinyi Wang, Zhaoyang Lyu, Dahua Lin, Bo Dai, and Hongfei Fu. Guided Diffusion Model for Adversarial Purification. arXiv, 2022. 1, 2, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.418, + 0.905, + 0.46 + ], + "angle": 0, + "content": "[45] Eric Wong, Leslie Rice, and J. Zico Kolter. Fast is better than free: Revisiting adversarial training. In International Conference on Learning Representations, 2020. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.461, + 0.905, + 0.503 + ], + "angle": 0, + "content": "[46] Eric Wong, Leslie Rice, and J. Zico Kolter. Fast is better than free: Revisiting adversarial training. In International Conference on Learning Representations, 2020. 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.504, + 0.905, + 0.573 + ], + "angle": 0, + "content": "[47] Chaowei Xiao, Zhongzhu Chen, Kun Jin, Jiongxiao Wang, Weili Nie, Mingyan Liu, Anima Anandkumar, Bo Li, and Dawn Song. Densepure: Understanding diffusion models for adversarial robustness. In The Eleventh International Conference on Learning Representations, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.575, + 0.905, + 0.63 + ], + "angle": 0, + "content": "[48] Yuzhe Yang, Guo Zhang, Dina Katabi, and Zhi Xu. Me-net: Towards effective adversarial robustness with matrix estimation. In International Conference on Machine Learning, pages 7025-7034. PMLR, 2019. 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.631, + 0.905, + 0.686 + ], + "angle": 0, + "content": "[49] Jongmin Yoon, Sung Ju Hwang, and Juho Lee. Adversarial purification with score-based generative models. In Proceedings of the 38th International Conference on Machine Learning, 2021. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.688, + 0.905, + 0.73 + ], + "angle": 0, + "content": "[50] Jongmin Yoon, Sung Ju Hwang, and Juho Lee. Adversarial purification with score-based generative models. In International Conference on Machine Learning, 2021. 6, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.731, + 0.905, + 0.757 + ], + "angle": 0, + "content": "[51] Sergey Zagoruyko. Wide residual networks. arXiv preprint arXiv:1605.07146, 2016. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.758, + 0.905, + 0.827 + ], + "angle": 0, + "content": "[52] Dawei Zhou, Tongliang Liu, Bo Han, Nannan Wang, Chunlei Peng, and Xinbo Gao. Towards defending against adversarial examples via attack-invariant features. In Proceedings of the 38th International Conference on Machine Learning, 2021. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.829, + 0.905, + 0.884 + ], + "angle": 0, + "content": "[53] Geigh Zollicoffer, Minh Vu, Ben Nebgen, Juan Castorena, Boian Alexandrov, and Manish Bhattarai. Lorid: Low-rank iterative diffusion for adversarial purification. arXiv preprint arXiv:2409.08255, 2024. 6, 7" + }, + { + "type": "list", + "bbox": [ + 0.517, + 0.093, + 0.905, + 0.884 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.48, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "29277" + } + ] +] \ No newline at end of file diff --git a/2025/Divide and Conquer_ Heterogeneous Noise Integration for Diffusion-based Adversarial Purification/94f1ab27-85c1-491d-abe8-0116716b932b_origin.pdf b/2025/Divide and Conquer_ Heterogeneous Noise Integration for Diffusion-based Adversarial Purification/94f1ab27-85c1-491d-abe8-0116716b932b_origin.pdf new file mode 100644 index 0000000000000000000000000000000000000000..a5a0ed41bdec4d255e874e36f653d065ce9bfc9b --- /dev/null +++ b/2025/Divide and Conquer_ Heterogeneous Noise Integration for Diffusion-based Adversarial Purification/94f1ab27-85c1-491d-abe8-0116716b932b_origin.pdf @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:6337bc3aad7b8069ff2b13b95e2d696a96d28bbc30472beaf4420436b1d6f6be +size 4516678 diff --git a/2025/Divide and Conquer_ Heterogeneous Noise Integration for Diffusion-based Adversarial Purification/full.md b/2025/Divide and Conquer_ Heterogeneous Noise Integration for Diffusion-based Adversarial Purification/full.md new file mode 100644 index 0000000000000000000000000000000000000000..db45421a9f665553f69c8e3713f130bcc9c58946 --- /dev/null +++ b/2025/Divide and Conquer_ Heterogeneous Noise Integration for Diffusion-based Adversarial Purification/full.md @@ -0,0 +1,374 @@ +# Divide and Conquer: Heterogeneous Noise Integration for Diffusion-based Adversarial Purification + +Gaozheng Pei $^{1}$ Shaojie Lyu $^{2}$ Gong Chen $^{2}$ Ke Ma $^{1*}$ Qianqian Xu $^{3}$ Yingfei Sun $^{1}$ Qingming Huang $^{3,4,5*}$ $^{1}$ School of Electronic, Electrical and Communication Engineering, UCAS $^{2}$ Tencent Corporate + $^{3}$ Key Laboratory of Intelligent Information Processing, Institute of Computing Technology, CAS + $^{4}$ School of Computer Science and Technology, UCAS + $^{5}$ Key Laboratory of Big Data Mining and Knowledge Management, UCAS + +peigaozheng23@mails.ucas.ac.cn, {shaojielv,natchen}@tencent.com xuqianqian@ict.ac.cn, {make,yfsun,qmhuang}@ucas.ac.cn + +# Abstract + +Existing diffusion-based purification methods aim to disrupt adversarial perturbations by introducing a certain amount of noise through a forward diffusion process, followed by a reverse process to recover clean examples. However, this approach is fundamentally flawed: the uniform operation of the forward process across all pixels compromises normal pixels while attempting to combat adversarial perturbations, resulting in the target model producing incorrect predictions. Simply relying on low-intensity noise is insufficient for effective defense. To address this critical issue, we implement a heterogeneous purification strategy grounded in the interpretability of neural networks. Our method decisively applies higher-intensity noise to specific pixels that the target model focuses on while the remaining pixels are subjected to only low-intensity noise. This requirement motivates us to redesign the sampling process of the diffusion model, allowing for the effective removal of varying noise levels. Furthermore, to evaluate our method against strong adaptative attack, our proposed method sharply reduces time cost and memory usage through a single-step resampling. The empirical evidence from extensive experiments across three datasets demonstrates that our method outperforms most current adversarial training and purification techniques by a substantial margin. Code is available at https://github.com/GaozhengPei/Purification. + +# 1. Introduction + +Adversarial perturbations [9], which are often imperceptible to humans, can significantly alter the predictions of + +![](images/074f326a88409f96ff88a145555e55280a5bb3e943490ed93cc9063fe1721a44.jpg) +Figure 1. When the noise added to adversarial samples is minimal (top), the adversarial perturbations remain intact and cannot be removed. On the other hand, if the noise intensity is excessive (bottom), it can distort the semantic information. Our approach employs an attention mask to introduce varying intensities of noise across different areas (middle). This technique effectively strikes a balance between preserving semantic information and mitigating adversarial perturbations. + +neural networks. To address these risks [21-23, 32], various defense mechanisms have been proposed. One common approach is adversarial training [18, 37, 45], which involves generating adversarial samples according to a specific attack method before the test phase and then retraining the model to improve its robustness. However, this approach often struggles to defend against unknown attacks. Another technique, known as adversarial purification [17, 27, 30, 38, 40, 44], seeks to remove perturbations from adversarial examples during the test phase. This approach does not rely on any assumptions about the attack methods and eliminates the overhead of repetitive training. Its flexibility and strong generalization have drawn increasing interest. In recent years, diffusion models [16, 29] have garnered attention due to their impressive denoising abilities and capacity to generate high-quality images, making them + +a promising option for adversarial purification. + +Existing diffusion-based adversarial purification methods [3, 30, 40, 44] follow the traditional forward process which attempts to add uniform noise across different regions of the images and then recover the clean images through reverse process. From Figure 1 we can observe that such a manner can not balance eliminating perturbations and preserving semantic information. If the added noise is intense, it will remove the adversarial perturbations but damage the semantic information, causing the wrong prediction of both natural and adversarial examples. In contrast, if the noise intensity is weak, it may not effectively purify the adversarial perturbations, resulting in a successful evasion attack. Therefore, how to achieve a good balance between the elimination of adversarial perturbations and the preservation of semantic information remains challenging. + +In this paper, we propose a novel diffusion-based adversarial purification method from the perspective of neural network interpretability. During making decisions, the neural network spontaneously allocates varying levels of attention to different regions of the image. This phenomenon inspired us to use attention maps as a prior guided diffusion model to impose heterogeneous noise on different regions of an image as forward process. By eliminating adversarial perturbations in the critical areas through high-intensity noise, we can substantially mitigate their effect on the prediction of the target model. Conversely, for regions that receive less attention, we introduce low-intensity noise to preserve as much semantic information as possible in the original image. To recover a clean image from one containing heterogeneous noise, we design a two-stage denoising process in accordance with the attention map as a reverse process of the diffusion model. Specifically, if the sampling timestep falls between two different noise levels, we apply inpainting techniques to remove localized high-intensity noise. If the current timestep is below that of the smaller noise level, we use the standard denoising process to remove global weak noise. Additionally, by replacing multi-step resampling with an improved single-step resampling approach, this strategy results in a $90\%$ reduction in both time consumption and GPU memory usage enables evaluate our method against strong adaptive attacks. Overall, our contributions can be summarized as follows: + +- We propose a novel forward process that adds noise of varying intensities based on the classifier's attention to different regions, achieving a balance between preserving semantic information and eliminating adversarial perturbations. +- To recover the clean samples, we design a two-stage denoising process: at a larger timestep, we convert it into an inpainting problem to remove localized strong noise, and at a smaller timestep, we apply a standard sampling + +process to eliminate global weak noise. + +- To evaluate our method against strong adaptive attacks. We replace the multi-step resampling with an improved single-step resampling, this not only achieves harmonious image generation but also significantly reduces time cost and memory usage. +- Extensive experiments on three datasets show that our method outperforms other methods by a promising improvement against various adversarial attacks. + +# 2. Related Work + +Adversarial purification aims to remove adversarial perturbations from samples at test phase before feeding them into classifier model. Adversarial purification methods can be broadly divided into two categories: training-based methods, which require training on the dataset, and diffusion based methods which doesn't require training and is not necessary to access the training dataset. + +# 2.1. Training-Based Adversarial Purification + +[41] shows empirically that adversarial examples mainly lie in the low probability regions of the training distribution and move them back towards the training distribution. [36] first models the distribution of clean samples and then attempts to find the closest sample within the clean sample distribution to the adversarial example during inference. [28] generates perturbed images using Self-supervised Perturbation (SSP) attack that disrupts the deep perceptual features and projects back the perturbed images close to the perceptual space of clean images. [52] propose to learn generalizable invariant features across attacks via encoder by playing zero-sum game and reconstruct the original image via decoder. [24] disrupt dversarial perturbations structure through random transform and reconstruct the image using a purifier model. [42] experimentally show that FGSM-AT networks overfit to unknown adversarial examples attacked again by FGSM. In the test phase, adversarial examples are fed to classifier after performing FGSM "purification". However, these methods require training using the training dataset, which is very time-consuming and lacks generalizability. + +# 2.2. Diffusion-Based Adversarial Purification + +[49] show that an EBM trained with Denoising Score-Matching can quickly purify attacked images within a few steps. [30] first diffuse samples with a small amount of noise and then recover the clean image through a reverse generative process. The paper also theoretically proves that the time step cannot be too large or too small. [44] proposes to use the adversarial image as guidance in the reverse process which encourages the purified image is consistent with the original clean image. [47] gets multiple reversed samples via multiple runs of denoising. Then the samples are + +passed into the classifier, where the final prediction is obtained through classifier voting. [3] design the forward process with the proper amount of Gaussian noise added and the reverse process with the gradient of contrastive loss as the guidance of diffusion models for adversarial purification. However, adding noise of the same intensity to different regions of the images, these methods is difficult to balance preserving semantic information and eliminating adversarial perturbations. + +# 3. Methodology + +Given an example $(\pmb{x},\pmb{y})$ , an adversary can create a small perturbation to generate an adversarial example $\pmb{x}_{adv}$ such that the classifier $\pmb{f}$ makes an incorrect prediction: $\pmb{y} \neq \pmb{f}(\pmb{x}_{adv})$ . Existing diffusion-based purification methods obtain a noisy image $\pmb{x}(t)$ at timestep $t$ through the forward process and then recover the clean (or purified) image $\hat{\pmb{x}}$ through the reverse process. And we hope $\pmb{f}(\hat{\pmb{x}}) = \pmb{f}(\pmb{x})$ . From the theoretical analysis of [30], the timestep $t$ should satisfy: 1) $t$ should be large enough to eliminate perturbations. 2) $t$ should be small enough that $\hat{\pmb{x}}$ still maintains semantic consistency with $\pmb{x}$ . We find that these two objectives are contradictory. + +Motivated by the interpretability of neural networks, the attention map is primarily used to encode the input spatial regions that the network focuses on most when making output decisions, such as in image classification. This suggests that, for an adversarial example, adversarial perturbations in the regions where the classifier focuses most significantly contribute to causing the target classifier to make incorrect predictions. The attention maps for each layer of the target classifier can be obtained with a single forward propagation. Therefore, we propose introducing high-intensity noise to eliminate adversarial perturbations of the regions with more attention value and applying low-intensity noise to the remaining area to preserve as much semantic information as possible. + +# 3.1. Heterogeneous Forward Process + +For the adversarial example $\pmb{x}_{adv}$ at timestep $t = 0$ , i.e., $\pmb{x}(0) = \pmb{x}_{adv}$ , we can obtain the noisy images at timestep $t_l$ and $t_s$ with $t_l > t_s$ as follows: + +$$ +\boldsymbol {x} \left(t _ {l}\right) = \sqrt {\bar {\alpha} \left(t _ {l}\right)} \cdot \boldsymbol {x} _ {a d v} + \sqrt {1 - \bar {\alpha} \left(t _ {l}\right)} \cdot \boldsymbol {\epsilon}, \tag {1a} +$$ + +$$ +\boldsymbol {x} \left(t _ {s}\right) = \sqrt {\bar {\alpha} \left(t _ {s}\right)} \cdot \boldsymbol {x} _ {a d v} + \sqrt {1 - \bar {\alpha} \left(t _ {s}\right)} \cdot \boldsymbol {\epsilon}, \tag {1b} +$$ + +where $\{\sqrt{\overline{\alpha}(t)}\}$ is a series of given parameters and $\epsilon \sim \mathcal{N}(0, I)$ . The clean (or purified) images $\hat{\pmb{x}}_{t_l}$ and $\hat{\pmb{x}}_{t_s}$ can be obtained through the normal reverse process [16, 30]. For the smaller timestep $t_s$ , the recovered image $\hat{\pmb{x}}_{t_s}$ retains certain adversarial perturbation, resulting in $\pmb{f}(\pmb{x}_{adv}) = \pmb{f}(\hat{\pmb{x}}_{t_s}) \neq \pmb{f}(\pmb{x})$ . Conversely, at the larger timestep $t_l$ , the + +clean image $\hat{\pmb{x}}_{t_l}$ loses critical semantic information, leading to $\pmb {f}(\pmb{x}_{adv})\neq \pmb {f}(\hat{\pmb{x}}_{t_l})\neq \pmb {f}(\pmb {x})$ . To effectively aggregate the differing impacts on various regions of the images at distinct timesteps, we construct an attention mask to integrate (1a) and (1b). + +Consider a neural network that consists of $M$ blocks (e.g. ResNet-50 [14] consists of 4 blocks). For $m$ -th block, we extract the output of the last activation function (e.g. Relu [8] or Swish [33]) as follows: + +$$ +\boldsymbol {G} _ {m} = \left(\sum_ {i = 1} ^ {C _ {m}} | \boldsymbol {A} _ {m i} | ^ {p}\right) ^ {\frac {1}{p}} \in \mathbb {R} ^ {H _ {m} \times W _ {m}}, m \in [ M ], \tag {2} +$$ + +where $p \in \mathbb{N}$ , $C_m, H_m, W_m$ are the dimension of channels, height, and width respectively, $\mathbf{A}_m$ denotes the activation of $m$ -th block and $\mathbf{A}_{mi}$ denotes the $i$ -th slice of $\mathbf{A}_m$ along the channel dimension. Then we apply interpolation techniques to ensure it matches the original image size. By upsampling $\mathbf{G}_m$ with bilinear function $\mathrm{Bi}(\cdot)$ , we obtain attention map of $m$ -th block: + +$$ +\mathrm {A M} _ {m} = \phi (\operatorname {B i} \left(\boldsymbol {G} _ {m}\right)), \tag {3} +$$ + +where $\phi (\cdot)$ is the spatial softmax operation. Moreover, we build the attention mask with all attention maps $\{\mathbf{AM}_m\}$ : + +$$ +\mathcal {M} = \bigcup_ {m = 1} ^ {M} \mathbb {I} [ \mathrm {A M} _ {m} > \tau ] \tag {4} +$$ + +where $\mathbb{I}[\cdot ]$ is the Iverson bracket and $\tau$ is the threshold to distinguish important and ordinary regions. With the attention mask $\mathcal{M}$ and two noisy images from Eq.(1), $\pmb {x}(t_s),\pmb {x}(t_l),t_s < t_l$ , the result of the proposed heterogeneous forward process is + +$$ +\boldsymbol {x} \left(t _ {l}, t _ {s}\right) = \boldsymbol {x} \left(t _ {l}\right) \odot \mathcal {M} + \boldsymbol {x} \left(t _ {s}\right) \odot (1 - \mathcal {M}), \tag {5} +$$ + +where $\odot$ is the Hadamard product. + +# 3.2. Heterogeneous Denoising Process + +To recover the clean image from (5), we design a two-stage denoising process. For the current timestep $t$ , we consider the following two cases: + +Stage 1 ( $t_s < t < t_l$ ) When the current timestep $t$ lies between $t_s$ and $t_l$ . For the pixels inside the attention mask $\mathcal{M}$ , they should be predicted from the previous timestep $t + 1$ . But for the pixels outside the attention mask $\mathcal{M}$ , they can be obtained by further adding noise forward from (5). Therefore, this case can be viewed as an inpainting problem [25]. For the image regions outside the attention mask $\mathcal{M}$ , since only the forward process is involved. Therefore, we can get the pixel value of these regions at timestep $t$ equivalently as follows: + +$$ +\boldsymbol {x} (t) ^ {\text {k n o w n}} = \sqrt {\bar {\alpha} (t)} \boldsymbol {x} _ {a d v} + \sqrt {1 - \bar {\alpha} (t)} \boldsymbol {\epsilon}. \tag {6} +$$ + +![](images/c19ed89de3269214133ea08a31903efe44e71df7d0106db9530d5170900606d4.jpg) +Figure 2. Pipeline of our method. Given an adversarial image, we extract the attention maps of each block during the forward propagation of the classifier, and construct an attention mask $\mathcal{M}$ . We then execute the denoising process of the diffusion model we designed. It can be seen that the samples purified by our method can be correctly classified. + +For the image regions within the attention mask, the pixel value of these regions at timestep $t$ should be predicted by the diffusion model: + +$$ +\boldsymbol {x} (t) ^ {\text {u n k n o w n}} \sim \mathcal {N} \left(\mu_ {\theta} (\boldsymbol {x} (t + 1), t + 1), \Sigma_ {\theta} (\boldsymbol {x} (t + 1), t + 1)\right) \tag {7} +$$ + +Where $\theta$ is the parameters of the diffusion model. $\mu_{\theta}$ and $\Sigma_{\theta}$ are mean and variance, respectively. Then we combine $x(t)^{known}$ and $x(t)^{unknown}$ together based on the attention mask $\mathcal{M}$ to get the $\pmb{x}(t)$ at timestep $t$ : + +$$ +\boldsymbol {x} (t) = \mathcal {M} \odot \boldsymbol {x} (t) ^ {\text {k n o w n}} + (1 - \mathcal {M}) \odot \boldsymbol {x} (t) ^ {\text {u n k n o w n}}, \tag {8} +$$ + +where $\odot$ is Hadamard product. We can remove the localized strong noise through the Stage 1. + +Stage 2 ( $t < t_s$ ) When the current timestep $t$ is less than $t_s$ , since the noise intensity is the same both inside and outside the attention mask. Therefore, the attention mask $\mathcal{M}$ is no longer needed. At this point, we can proceed with the standard sampling process of the diffusion model. We can get $x(t)$ predicted from the previous timestep $x(t + 1)$ as follows: + +$$ +\boldsymbol {x} (t) \sim \mathcal {N} \left(\mu_ {\theta} (\boldsymbol {x} (t + 1), t + 1), \Sigma_ {\theta} (\boldsymbol {x} (t + 1), t + 1)\right) \tag {9} +$$ + +Following Stage 2, we will eliminate global weak noise and get the clean image. The complete algorithm process can be found in the Algorithm 1. Additionally, there are some visualizations of our method in the Figure 5. + +# 4. Adaptive Attack to Diffusion Purification + +When directly applying the method described in 3.2, we observe the same phenomenon as [25] that although the region + +# Algorithm 1 Heterogeneous Denoising Process + +Require: Sample $x_0$ , Attention Mask $\mathcal{M}$ , timestep $t_s$ and $t_l$ , Diffusion model parameters $\theta$ . + +for $t = t_l,\dots,1$ do + +$$ +\text {i f} t _ {s} \leq t \leq t _ {l} \text {t h e n} +$$ + +Stage 1 + +$$ +\epsilon \sim \mathcal {N} (0, \mathbf {I}) \text {i f} t > 1, \text {e l s e} \epsilon = 0 +$$ + +$$ +x _ {t - 1} ^ {\text {k n o w n}} = \sqrt {\overline {{\alpha}} _ {t - 1}} x _ {0} + \sqrt {1 - \overline {{\alpha}} _ {t - 1}} \epsilon +$$ + +$$ +z \sim \mathcal {N} (0, \mathbf {I}) \text {i f} t > 1, \text {e l s e} z = 0 +$$ + +$$ +x _ {t - 1} ^ {u n k n o w n} = \frac {1}{\sqrt {\alpha_ {t}}} \left(x _ {t} - \frac {\beta_ {t}}{\sqrt {1 - \bar {\alpha} _ {t}}} \epsilon_ {\theta} (x _ {t}, t)\right) + \sigma_ {t} z +$$ + +$$ +x _ {t - 1} = \mathcal {M} \odot x _ {t - 1} ^ {\text {k n o w n}} + (1 - \mathcal {M}) \odot x _ {t - 1} ^ {\text {u n k n o w n}} +$$ + +else if $t < t_s$ then + +Stage 2 + +$$ +z \sim \mathcal {N} (0, \mathbf {I}) \text {i f} t > 1, \text {e l s e} z = 0 +$$ + +$$ +x _ {t - 1} = \frac {1}{\sqrt {\alpha_ {t}}} \left(x _ {t} - \frac {\beta_ {t}}{\sqrt {1 - \bar {\alpha} _ {t}}} \epsilon_ {\theta} (x _ {t}, t)\right) + \sigma_ {t} z +$$ + +for $u = 1,2,\ldots U$ do + +$\triangleright$ Resampling + +$$ +z \sim \mathcal {N} (0, \mathbf {I}) \text {i f} t > 1, \text {e l s e} z = 0 +$$ + +$$ +x _ {t - 1 + u} = \sqrt {\alpha_ {t - 1 + u}} x _ {t + u - 2} + \sqrt {1 - \alpha_ {t - 1 + u}} z +$$ + +end for + +get $x_{t - 1}$ using (11) + +end for + +recovered in Stage 2 matches the texture of the neighboring region, it is semantically incorrect. We can observe this phenomenon in the Figure 3. [25] addresses this issue via multi-step resampling by diffusing the output $\pmb{x}(t)$ back to $\pmb{x}(t + 1)$ : + +$$ +\boldsymbol {x} (t + 1) = \sqrt {\alpha (t)} \boldsymbol {x} (t) + \sqrt {1 - \alpha (t)} \boldsymbol {\epsilon}, \tag {10} +$$ + +and it is necessary to fully execute 3.2 again to get $\pmb{x}(t)$ . From Figure 3 we can find that this process needs to be repeated $U = 20$ times to ensure that the masked region + +![](images/8346b0525140df04bf24c7cd2ebd2317d0e0571c55a7d25ec030b94ea26ba3a0.jpg) +Figure 3. Our method maintains semantic consistency between the masked areas and the surrounding pixel values when $U = 10$ . Note that no matter what $U$ is, our method requires only one additional call to the denoising network. + +becomes semantically consistent with the surrounding pixels. To evaluate our method against strong adaptive attacks [30], it is necessary to compute the full gradient of the whole sampling process to construct adversarial examples. To reduce the time cost and memory usage, we optimized the multi-step resampling into a single resampling process which achieves the same functionality. + +The intuition behind our improvement is that repeating (10) and 3.2 for $U$ times can be replaced by that first diffuse $\pmb{x}(t)$ to $\pmb{x}(t + U)$ and then denoise it back to $\pmb{x}(t)$ from $\pmb{x}(t + U)$ in a single step using DDIM [39] as follows: + +$$ +\begin{array}{l} \boldsymbol {x} _ {t - 1} = \sqrt {\alpha_ {t - 1}} \left(\frac {\boldsymbol {x} _ {t - 1 + U} - \sqrt {1 - \alpha_ {t - 1 + U}} \epsilon_ {\theta} (\boldsymbol {x} _ {t - 1 + U})}{\sqrt {\alpha_ {t - 1 + U}}}\right) \\ + \sqrt {1 - \alpha_ {t - 1} - \sigma_ {t - 1 + U} ^ {2}} \cdot \epsilon_ {\theta} \left(\boldsymbol {x} _ {t - 1 + U}\right) + \sigma_ {t - 1 + U} \boldsymbol {\epsilon}. \tag {11} \\ \end{array} +$$ + +From the bottom panel of Figure 3, we find that our method can also achieve semantically consistency when $U = 10$ . This means that at each timestep, we only need to invoke the denoising network an additional time. Therefore, the GPU memory needs to store the $2 * T$ complete computation graphs of the denoising network which enables the computation of the gradient of our method on consumer-grade GPU with only 24G memory. Also, the time cost is greatly reduced. The analysis of time cost and memory usage can refer to the Appendix. + +# 5. Experiment + +# 5.1. Experimental Settings + +Datasets and network architectures. Three datasets is used to evaluate including CIFAR-10, SVHN, and ImageNet [6]. We compare our results with several leading defense methods as listed in the standardized benchmark RobustBench [5] for CIFAR-10 and ImageNet, and also test different adversarial purification methods. We use two common classifier architectures: WideResNet-28-10 + +and WideResNet-70-16 [51] for CIFAR-10. For SVHN, WideResNet-28-10 serves as the backbone, while for ImageNet, ResNet-50 [14] is used as the backbone. + +Adversarial attack methods. Strong adaptive attack is tested for our method and other adversarial purification methods. We use the commonly known AutoAttack [4] under $\ell_{\infty}$ and $\ell_{2}$ threat models. Additionally, the projected gradient descent (PGD) attack [26] is evaluated on our method, as recommended in [20]. Due to the randomness introduced by the diffusion and denoising processes, Expectation Over Time (EOT) [1] is adapted for these adaptive attacks. Moreover, we apply the BPDA+EOT [15] attack to ensure a fair comparison with other adversarial purification methods. Finally, following the suggestion of [20], a surrogate process is used to obtain the gradient of the reverse process for white-box attacks. + +Pre-trained Models. We use the unconditional CIFAR-10 checkpoint of EDM provided by NVIDIA [19] on CIFAR-10. 256x256 diffusion (unconditional) checkpoint from guided-diffusion library is adopted for the ImageNet experiments. For CIFAR-10, the pre-trained classifier is downloaded from RobustBench [5], while for ImageNet, the weight of the classifier is provided by PyTorch library. + +Evaluation metrics. To assess the performance of defense methods, we use two metrics: standard accuracy, measured on clean samples, and robust accuracy, measured on adversarial examples. Due to the high computational cost of evaluating models against adaptive attacks, a fixed subset of 512 images are randomly sampled from the test set for robust evaluation, which is the same as [20, 30]. In all experiments, the mean and standard deviation over three runs are reported to assess both standard and robust accuracy. + +Implementation details. We follow the settings described in [20]. We evaluate diffusion-based purification methods using the PGD+EOT attack with 200 update iterations, and BPDA and AutoAttack with 100 update iterations, except for ImageNet, which uses 20 iterations. We set the number of EOT to 20. The step size is 0.007. For randomized defenses, such as [3, 20, 30, 40], we use the random version of AutoAttack, and for static defenses, we use the standard version. + +# 5.2. Experimental Results + +CIFAR-10. We conducted comprehensive experiments on the CIFAR-10 dataset, evaluating our method against other approaches on two model architectures: WideResNet28-10 and WideResNet-70-16. We assessed robust accuracy under three attack types: BPDA+EOT, PGD+EOT, and AutoAttack. We tested against the infinity norm attack, and as shown in Table 1, when the backbone is WideResNet28-10, our method significantly outperforms other baselines in both standard accuracy and robust accuracy, exceeding by $3.00\%$ in standard accuracy and leading by $3.56\%$ and + +Table 1. Standard and robust accuracy of different Adversarial Training (AT) and Adversarial Purification (AP) methods against PGD+EOT and AutoAttack $\ell_{\infty}(\epsilon = 8 / 255)$ on CIFAR-10. * utilizes half number of iterations for the attack due to the high computational cost. † indicates the requirement of extra data. + +
TypeMethodStandard Acc.Robust Acc.
PGDAutoAttack
WideResNet-28-10
ATGowal et al. [13]88.5465.9363.38
Gowal et al. [11]†87.5166.0162.76
Pang et al. [31]88.6264.9561.04
APYoon et al. [50]85.66±0.5133.48±0.8659.53±0.87
Nie et al. [30]90.07±0.9756.84±0.5963.6±0.81
Lee et al. [20]90.16±0.6455.82±0.5970.47±1.53
Bai et al. [3]91.4149.22*77.08
Geigh et al. [53]84.20-59.14
Lin et al. [24]90.62-72.85
Ours93.16±0.2159.38±1.2580.45±1.54
+ +Table 2. Standard and robust accuracy of different Adversarial Training (AT) and Adversarial Purification (AP) methods against PGD+EOT and AutoAttack $\ell_{\infty}(\epsilon = 8 / 255)$ on CIFAR-10. * The number of iterations for the attack is half that of the other methods for less computational overhead. † indicates the requirement of extra data. + +
TypeMethodStandard Acc.Robust Acc.
PGDAutoAttack
WideResNet-70-16
ATRebuffi et al. [34]†92.2269.9766.56
Gowal et al. [10]†91.1068.6666.10
Gowal et al. [12]88.7569.0365.87
APYoon et al. [50]86.76±1.1537.11±1.3560.86±0.56
Nie et al. [30]90.43±0.6051.13±0.8766.06±1.17
Lee et al. [20]90.53±0.156.88±1.0670.31±0.62
Bai et al. [3]92.9748.83*79.10
Geigh et al. [53]84.60-66.40
Geigh et al. [53]86.90-59.20
Lin et al. [24]91.99-76.37
Ours93.36±0.5959.79±0.6184.83±0.59
+ +3.37% under PGD and AutoAttack, respectively. From Table 2, we see that when WideResNet-70-16 is used as the backbone, our method exceeds the state-of-the-art (SOTA) by 1.37% in standard accuracy and leads by 2.91% and 5.73% under PGD and AutoAttack, respectively. We also evaluated the accuracy of our method and other baselines under the Euclidean norm attack. From Table 3, we observe that, with WideResNet-28-10 as the backbone, although our method lags behind [3] by 0.59% in standard accuracy, it leads by 0.98% and 3.77% in robust accuracy. Due to the + +Table 3. Standard and robust accuracy against PGD+EOT and AutoAttack $\ell_2(\epsilon = 0.5)$ on CIFAR-10. Adversarial Training (AT) and Adversarial Purification (AP) methods are evaluated. *The number of iterations for the attack is half that of the other methods for less computational overhead. † indicates the requirement of extra data. ‡ adopts WideResNet-34-10 as the backbone, with the same width but more layers than the default one. + +
TypeMethodStandard Acc.Robust Acc.
PGDAutoAttack
WideResNet-28-10
ATRebuffi et al. [34]†91.7985.0578.80
Augustin et al. [2]‡93.9686.1478.79
Sehwag et al. [31]‡90.9383.7577.24
APYoon et al. [50]85.66±0.5173.32±0.7679.57±0.38
Nie et al. [30]91.41±1.0079.45±1.1681.7±0.84
Lee et al [20]90.16±0.6483.59±0.8886.48±0.38
Bai et al. [3]93.7586.13*80.92
Geigh et al. [53]84.40-77.90
Geigh et al. [53]84.20-73.60
Lin et al. [24]90.62-80.47
Ours93.16±0.2187.11±0.9590.25±0.80
+ +Table 4. Standard and robust accuracy against PGD+EOT and AutoAttack $\ell_2(\epsilon = 0.5)$ on CIFAR-10. Adversarial Training (AT) and Adversarial Purification (AP) methods are evaluated. (*The number of iterations for the attack is half that of other methods for less computational overhead. † methods need extra data.) + +
TypeMethodStandard Acc.Robust Acc.
PGDAutoAttack
WideResNet-70-16
ATRebuffi et al. [34]†95.7489.6282.32
Gowal et al. [10]†94.7488.1880.53
Rebuffi et al. [34]92.4186.2480.42
APYoon et al. [50]86.76±1.1575.666±1.2980.43±0.42
Nie et al. [30]92.15±0.7282.97±1.3883.06±1.27
Lee et al [20]90.53±0.1483.75±0.9985.59±0.61
Bai et al. [3]92.9784.37*83.01
Lin et al. [24]91.99-81.35
Ours93.36±0.5987.7±0.9791.71±1.09
+ +significant time overhead of [3], which only iterates 100 times under PGD—half of what others do—our method outperforms more under PGD attacks. From Table 4, we see that when WideResNet-70-16 is the backbone, our method outperforms other baselines, leading by $1.21\%$ in standard accuracy and by $3.33\%$ and $6.12\%$ in robust accuracy under PGD and AutoAttack, respectively. Additionally, we applied the BPDA+EOT attack, which approximates differentiability. From Table 5, we observe that although robust accuracy lags behind [40] by $1.71\%$ , our method still leads + +Table 5. Standard and robust accuracy against BPDA+EOT $\ell_{\infty}(\epsilon = 8 / 255)$ on CIFAR-10. + +
MethodPurificationAccuracy
StandardRobust
WideResNet-28-10
Song et al. [41]Gibbs Update95.009.00
Yang et al. [48]Mask+Recon94.0015.00
Hill et al. [15]EBM+LD84.1254.90
Yoon et al. [50]DSM+LD85.66±0.5166.91±1.75
Nie et al. [30]Diffusion90.07±0.9781.45±1.51
Bai et al. [3]Diffusion91.37±1.2185.51±0.81
Wang et al. [44]Diffusion89.96±0.4075.59±1.26
Song et al. [40]Diffusion89.88±0.3588.43±0.83
Lee et al. [20]Diffusion90.16±0.6488.40±0.88
OursDiffusion93.04±0.6586.72±0.20
+ +in standard accuracy by $3.16\%$ , resulting in an average advantage of $0.73\%$ . We also measured the distance between purified samples and the original samples in feature space, shown in the Appendix. A smaller distance indicates that the purified samples retain more consistent semantic information with the original samples (further discussion in the appendix). From Figure 6, we find that the distribution of purified samples by our method is most similar to the original samples compared to other methods. Overall, our method outperforms others, demonstrating its effectiveness in preserving semantic information while eliminating adversarial perturbations. + +Table 6. Standard and robust accuracy against PGD+EOT $\ell_{\infty}(\epsilon = 8 / 255)$ on ImageNet. ResNet-50 is used as the classifier. + +
TypeMethodAccuracy
StandardRobust
ResNet-50
ATSalman et al. [35]63.8639.11
Engstrom et al. [7]62.4233.20
Wong et al. [46]53.8328.04
APNie et al. (t* = 0.2) [30]73.9640.63
Nie et al. (t* = 0.3) [30]72.8548.24
Nie et al. (t* = 0.4) [30]61.7141.67
Bai et al. [3]70.4141.7
Lee et al. [20]71.4246.59
Song et al. [40]62.2551.14
Geigh et al. [53]73.9856.54
Ours74.51±1.0760.06±1.66
+ +ImageNet. We test ResNet50 as the backbone on the ImageNet dataset under PGD+EOT attacks, which is the same as [3, 20, 30]. From Table 6, our method leads competitors by $0.53\%$ in standard accuracy and by $3.52\%$ in + +![](images/437faba82c23906b56fbff745704b27c211fe5b0cbdad2f8c2545d6bfff24be5.jpg) +Figure 4. 2D purification trajectories of Our method and DiffPure. For DiffPure (left), if $t^{*} = 0.2$ , although the purification direction is toward the original class, the adversarial perturbation cannot be completely eliminated; If $t^{*} = 0.4$ , the semantic information changes and the purification direction is no longer toward the original class. Our method (right) can eliminate adversarial perturbation and remain the semantic information. + +![](images/7c0d260e32729411e6704017f5b77a62d94ecdcf7eb856cf11bd72c4555738d4.jpg) + +![](images/bd8b81cd98e629f949190121e32820383bfa88d119d6fbded0f61a52cc952f2d.jpg) +Figure 5. Visualization of attention map, attention mask and images purified by our method. + +robust accuracy, significantly surpassing the baseline. Additionally, we observe that the method from [30] achieves a standard accuracy of $73.96\%$ and a robust accuracy of $40.63\%$ when $t^{*} = 0.2$ . When $t^{*} = 0.3$ , standard accuracy drops to $72.85\%$ while robust accuracy rises to $48.24\%$ . However, at $t^{*} = 0.4$ , both standard accuracy and robust accuracy decline to $61.71\%$ and $41.67\%$ , respectively. Additionally, We visualize the 2D purification trajectories of our method and DiffPure using the T-SNE [43] technique. From Figure 4, we observe that when $t^{*} = 0.2$ , the adversarial perturbation on the adversarial sample has not been fully purified, and it remains at the boundary between the two classes of samples. When $t^{*} = 0.4$ , the semantic information of the adversarial example begins to change, and as a result, the purification trajectory no longer points toward the original classes. This indicates that applying the same noise intensity across an image cannot effectively balance semantic preservation and the elimination of adversarial perturbation. + +![](images/f2fe2f7da156c147aede35b69270193397e65fa7941be1fc6e3c72d973e71519.jpg) + +![](images/2f0d78aadcdf83081c5d03c6f18176e424690d48eb30011f598427be3bf7f023.jpg) + +![](images/465d02c0236b8ba9d053548104da1a56da129f811c9f47a5da7cc242eda6bdee.jpg) +Figure 6. Joint distribution of the original samples and purified samples. The distributions of the purified samples by our method and the original samples are the most similar. + +![](images/47ef5752cd0e60b6731935e6c95310e974fcf95da032e0fedd57118f5da4b3a9.jpg) + +tions. In contrast, our method adds varying noise intensities to different regions of an image, applying stronger noise to areas with higher classification weights to eliminate adversarial perturbations, while applying lighter noise to other regions to preserve semantic information. This approach allows our method to achieve good results in both robust accuracy and standard accuracy simultaneously. Our method achieves comparable standard accuracy to [30] at $t^* = 0.2$ but significantly outperforms it in robust accuracy, demonstrating the effectiveness of our approach. + +![](images/3e786b46938306ad70c75c0e8cf8c8e2c26fc791e4439b071a27362b44e47c4a.jpg) +Figure 7. Robust and Standard Accuracy under different thresholds $\tau$ (left) and under different combinations of timesteps $t_l$ and $t_s$ (right). + +![](images/1561298de81ab5e7748e4dc5623d426b61de577b70cac408191e9b045e94b936.jpg) + +# 5.3. Ablation Study + +We first conducted a hyperparameter sensitivity analysis. Our method involves three hyperparameters, two of which are noises of different intensities, $t_{l}$ and $t_{s}$ , and the third is a threshold, $\tau$ , used to control the size of the attention mask. From Figure 7, we observe that as $\tau$ increases (resulting in smaller mask areas), the Standard Accuracy also consistently increases. However, adversarial perturbations are preserved at this time, leading to poorer Robust Accuracy when $\tau$ is large. When $\tau$ is small, semantic information is damaged, which also results in weaker Robust Accuracy. Therefore, we choose $\tau$ to be around 0.8. Simultaneously, we can see that when $t_{l} = 0.2$ , $t_{s} = 0.05$ , the Robust Accuracy is the highest. Here, $t_{s}$ is less than the $t^{*}$ used in [30] to preserve more semantic information, while $t_{l}$ is greater than $t^{*}$ to eliminate adversarial perturbations. From Table 6, we can see that when we set $t_{l} = 0.4$ and $t_{s} = 0.2$ , The Standard Accuracy is almost as high as the $t^{*} = 0.2$ of [30], but at the same time, our Robust Accuracy significantly leads. This demonstrates the effectiveness of our method. We also explore whether it is necessary for each block + +Table 7. Standard Accuracy and Robust Accuracy for different block combinations. WideResNet28-10 servers as the backbone. + +
Encoder BlockAccuracy
E1E2E3E4StandardRobustAverage
X✓✓✓92.7758.0075.38
✓X✓✓93.1658.0075.58
✓✓X✓93.3557.6175.48
✓✓✓X91.6658.8575.25
✓✓✓✓93.1659.3876.27
+ +to participate in constructing the attention mask. Taking WideResNet28-10 as an example, it can be divided into four blocks. We extract the output of the final ReLU function of each block as the activation map. From Table 7, we can see that removing any block results in a performance loss. Therefore, every block should be utilized. + +# 6. Conclusion + +We propose a novel forward process that adds noise of varying intensities based on the classifier's attention to different regions, achieving a balance between preserving semantic information and eliminating adversarial perturbations. The denoising process is tailored to our method to remove different-intensity noise within a noisy image. By replacing the multi-step resampling with an improved single-step resampling, time cost and memory usage are reduced significantly. Extensive experiments on three datasets show that our method outperforms other methods by a promising improvement against various adversarial attacks. + +# Acknowledge + +This work was supported in part by the National Key R&D Program of China under Grant 2022ZD0119204, in part by National Natural Science Foundation of China: 62236008, 62441232, U21B2038, U23B2051, 62122075 and 62376257, in part by Youth Innovation Promotion Association CAS, in part by the Strategic Priority Research Program of the Chinese Academy of Sciences, Grant No. XDB0680201, in part by the Fundamental Research Funds for the Central Universities, in part by the Tencent Marketing Solution Rhino-Bird Focused Research Program, in part by Xiaomi Young Talents Program. + +# References + +[1] Anish Athalye, Nicholas Carlini, and David Wagner. Obfuscated gradients give a false sense of security: Circumventing defenses to adversarial examples. In International conference on machine learning, pages 274-283. PMLR, 2018. 5 +[2] Maximilian Augustin, Alexander Meinke, and Matthias Hein. Adversarial robustness on in-and out-distribution improves explainability. In European Conference on Computer Vision, pages 228-245. Springer, 2020. 6 +[3] Mingyuan Bai, Wei Huang, Tenghui Li, Andong Wang, Junbin Gao, Cesar F Caiafa, and Qibin Zhao. Diffusion models demand contrastive guidance for adversarial purification to advance. In *Forty-first International Conference on Machine Learning*, 2024. 2, 3, 5, 6, 7 +[4] Francesco Croce and Matthias Hein. Reliable evaluation of adversarial robustness with an ensemble of diverse parameter-free attacks. In International conference on machine learning, pages 2206-2216, 2020. 5 +[5] Francesco Croce, Maksym Andriushchenko, Vikash Sehwag, Edoardo Debenedetti, Nicolas Flammarion, Mung Chiang, Prateek Mittal, and Matthias Hein. Robustbench: a standardized adversarial robustness benchmark. arXiv preprint arXiv:2010.09670, 2020. 5 +[6] Jia Deng, Wei Dong, Richard Socher, Li-Jia Li, Kai Li, and Li Fei-Fei. Imagenet: A large-scale hierarchical image database. In 2009 IEEE conference on computer vision and pattern recognition, pages 248–255, 2009. 5 +[7] Logan Engstrom, Andrew Ilyas, Hadi Salman, Shibani Santurkar, and Dimitris Tsipras. Robustness (python library), 2019. URL https://github.com/MadryLab/robustness, 4(4): 4-3, 2019. 7 +[8] Xavier Glorot, Antoine Bordes, and Yoshua Bengio. Deep sparse rectifier neural networks. In Proceedings of the fourteenth international conference on artificial intelligence and statistics, pages 315-323, 2011. 3 +[9] Ian J Goodfellow, Jonathon Shlens, and Christian Szegedy. Explaining and harnessing adversarial examples. arXiv preprint arXiv:1412.6572, 2014. 1 +[10] Sven Gowal, Chongli Qin, Jonathan Uesato, Timothy Mann, and Pushmeet Kohli. Uncovering the limits of adversarial training against norm-bounded adversarial examples. arXiv preprint arXiv:2010.03593, 2020. 6 + +[11] Sven Gowal, Chongli Qin, Jonathan Uesato, Timothy Mann, and Pushmeet Kohli. Uncovering the limits of adversarial training against norm-bounded adversarial examples. arXiv preprint arXiv:2010.03593, 2020. 6 +[12] Sven Gowal, Sylvestre-Alvise Rebuffi, Olivia Wiles, Florian Stimberg, Dan Andrei Calian, and Timothy A Mann. Improving robustness using generated data. Advances in Neural Information Processing Systems, 34:4218-4233, 2021. 6 +[13] Sven Gowal, Sylvestre-Alvise Rebuffi, Olivia Wiles, Florian Stimberg, Dan Andrei Calian, and Timothy A Mann. Improving robustness using generated data. Advances in Neural Information Processing Systems, 34:4218-4233, 2021. 6 +[14] Kaiming He, Xiangyu Zhang, Shaoqing Ren, and Jian Sun. Deep residual learning for image recognition. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 770-778, 2016. 3, 5 +[15] Mitch Hill, Jonathan Craig Mitchell, and Song-Chun Zhu. Stochastic security: Adversarial defense using long-run dynamics of energy-based models. In International Conference on Learning Representations, 2021. 5, 7 +[16] Jonathan Ho, Ajay Jain, and Pieter Abbeel. Denoising diffusion probabilistic models. Advances in neural information processing systems, 33:6840-6851, 2020. 1, 3 +[17] Duhun Hwang, Eunjung Lee, and Wonjong Rhee. Aidpurifier: A light auxiliary network for boosting adversarial defense. Neurocomputing, 541:126251, 2023. 1 +[18] Xiaojun Jia, Yong Zhang, Xingxing Wei, Baoyuan Wu, Ke Ma, Jue Wang, and Xiaochun Cao. Improving fast adversarial training with prior-guided knowledge. IEEE Transactions on Pattern Analysis and Machine Intelligence, 2024. 1 +[19] Tero Karras, Miika Aittala, Timo Aila, and Samuli Laine. Elucidating the design space of diffusion-based generative models. Advances in neural information processing systems, 35:26565-26577, 2022. 5 +[20] Minjong Lee and Dongwoo Kim. Robust evaluation of diffusion-based adversarial purification. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 134-144, 2023. 5, 6, 7 +[21] Siyuan Liang, Xingxing Wei, Siyuan Yao, and Xiaochun Cao. Efficient adversarial attacks for visual object tracking. In Computer Vision-ECCV 2020: 16th European Conference, Glasgow, UK, August 23-28, 2020, Proceedings, Part XXVI 16, pages 34-50. Springer, 2020. 1 +[22] Siyuan Liang, Longkang Li, Yanbo Fan, Xiaojun Jia, Jingzhi Li, Baoyuan Wu, and Xiaochun Cao. A large-scale multiple-objective method for black-box attack against object detection. In European Conference on Computer Vision, pages 619-636. Springer, 2022. +[23] Siyuan Liang, Baoyuan Wu, Yanbo Fan, Xingxing Wei, and Xiaochun Cao. Parallel rectangle flip attack: A query-based black-box attack against object detection. arXiv preprint arXiv:2201.08970, 2022. 1 +[24] Guang Lin, Chao Li, Jianhai Zhang, Toshihisa Tanaka, and Qibin Zhao. Adversarial training on purification (ATop): Advancing both robustness and generalization. In The Twelfth International Conference on Learning Representations, 2024. 2, 6 + +[25] Andreas Lugmayr, Martin Danelljan, Andres Romero, Fisher Yu, Radu Timofte, and Luc Van Gool. Repaint: Inpainting using denoising diffusion probabilistic models. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 11461-11471, 2022. 3, 4 +[26] Aleksander Madry, Aleksandar Makelov, Ludwig Schmidt, Dimitris Tsipras, and Adrian Vladu. Towards deep learning models resistant to adversarial attacks. In International Conference on Learning Representations, 2018. 5 +[27] Dongyu Meng and Hao Chen. Magnet: a two-pronged defense against adversarial examples. In Proceedings of the 2017 ACM SIGSAC conference on computer and communications security, pages 135-147, 2017. 1 +[28] Muzammal Naseer, Salman Khan, Munawar Hayat, Fahad Shahbaz Khan, and Fatih Porikli. A self-supervised approach for adversarial robustness. In 2020 IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), 2020. 2 +[29] Alexander Quinn Nichol and Prafulla Dhariwal. Improved denoising diffusion probabilistic models. In International conference on machine learning, 2021. 1 +[30] Weili Nie, Brandon Guo, Yujia Huang, Chaowei Xiao, Arash Vahdat, and Animashree Anandkumar. Diffusion models for adversarial purification. In Proceedings of the 39th International Conference on Machine Learning, 2022. 1, 2, 3, 5, 6, 7, 8 +[31] Tianyu Pang, Min Lin, Xiao Yang, Jun Zhu, and Shuicheng Yan. Robustness and accuracy could be reconcilable by (proper) definition. In International Conference on Machine Learning, pages 17258-17277. PMLR, 2022. 6 +[32] Gaozheng Pei, Ke Ma, Pinci Yang, Qianqian Xu, Yingfei Sun, et al. Exploring query efficient data generation towards data-free model stealing in hard label setting. arXiv preprint arXiv:2412.15276, 2024. 1 +[33] Prajit Ramachandran, Barret Zoph, and Quoc V Le. Searching for activation functions. arXiv preprint arXiv:1710.05941, 2017. 3 +[34] Sylvestre-Alvise Rebuffi, Sven Gowal, Dan A Calian, Florian Stimberg, Olivia Wiles, and Timothy Mann. Fixing data augmentation to improve adversarial robustness. arXiv preprint arXiv:2103.01946, 2021. 6 +[35] Hadi Salman, Andrew Ilyas, Logan Engstrom, Ashish Kapoor, and Aleksander Madry. Do adversarially robust imagenet models transfer better? Advances in Neural Information Processing Systems, 33:3533-3545, 2020. 7 +[36] Pouya Samangouei, Maya Kabbab, and Rama Chellappa. Defense-GAN: Protecting classifiers against adversarial attacks using generative models. In International Conference on Learning Representations, 2018. 2 +[37] Ali Shafahi, Mahyar Najibi, Mohammad Amin Ghiasi, Zheng Xu, John Dickerson, Christoph Studer, Larry S Davis, Gavin Taylor, and Tom Goldstein. Adversarial training for free! Advances in neural information processing systems, 32, 2019. 1 +[38] Changhao Shi, Chester Holtz, and Gal Mishne. Online adversarial purification based on self-supervised learning. In International Conference on Learning Representations, 2021. 1 + +[39] Jiaming Song, Chenlin Meng, and Stefano Ermon. Denoising diffusion implicit models. In International Conference on Learning Representations, 2021. 5 +[40] Kaiyu Song, Hanjiang Lai, Yan Pan, and Jian Yin. Mimic diffusion: Purifying adversarial perturbation via mimicking clean diffusion model. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), 2024. 1, 2, 5, 6, 7 +[41] Yang Song, Taesup Kim, Sebastian Nowozin, Stefano Ermon, and Nate Kushman. Pixeldefend: Leveraging generative models to understand and defend against adversarial examples. In International Conference on Learning Representations, 2018. 2, 7 +[42] Linyu Tang and Lei Zhang. Robust overfitting does matter: Test-time adversarial purification with fgsm. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), 2024. 2 +[43] Laurens Van der Maaten and Geoffrey Hinton. Visualizing data using t-sne. Journal of machine learning research, 9 (11), 2008. 7 +[44] Jinyi Wang, Zhaoyang Lyu, Dahua Lin, Bo Dai, and Hongfei Fu. Guided Diffusion Model for Adversarial Purification. arXiv, 2022. 1, 2, 7 +[45] Eric Wong, Leslie Rice, and J. Zico Kolter. Fast is better than free: Revisiting adversarial training. In International Conference on Learning Representations, 2020. 1 +[46] Eric Wong, Leslie Rice, and J. Zico Kolter. Fast is better than free: Revisiting adversarial training. In International Conference on Learning Representations, 2020. 7 +[47] Chaowei Xiao, Zhongzhu Chen, Kun Jin, Jiongxiao Wang, Weili Nie, Mingyan Liu, Anima Anandkumar, Bo Li, and Dawn Song. Densepure: Understanding diffusion models for adversarial robustness. In The Eleventh International Conference on Learning Representations, 2023. 2 +[48] Yuzhe Yang, Guo Zhang, Dina Katabi, and Zhi Xu. Me-net: Towards effective adversarial robustness with matrix estimation. In International Conference on Machine Learning, pages 7025-7034. PMLR, 2019. 7 +[49] Jongmin Yoon, Sung Ju Hwang, and Juho Lee. Adversarial purification with score-based generative models. In Proceedings of the 38th International Conference on Machine Learning, 2021. 2 +[50] Jongmin Yoon, Sung Ju Hwang, and Juho Lee. Adversarial purification with score-based generative models. In International Conference on Machine Learning, 2021. 6, 7 +[51] Sergey Zagoruyko. Wide residual networks. arXiv preprint arXiv:1605.07146, 2016. 5 +[52] Dawei Zhou, Tongliang Liu, Bo Han, Nannan Wang, Chunlei Peng, and Xinbo Gao. Towards defending against adversarial examples via attack-invariant features. In Proceedings of the 38th International Conference on Machine Learning, 2021. 2 +[53] Geigh Zollicoffer, Minh Vu, Ben Nebgen, Juan Castorena, Boian Alexandrov, and Manish Bhattarai. Lorid: Low-rank iterative diffusion for adversarial purification. arXiv preprint arXiv:2409.08255, 2024. 6, 7 \ No newline at end of file diff --git a/2025/Divide and Conquer_ Heterogeneous Noise Integration for Diffusion-based Adversarial Purification/images.zip b/2025/Divide and Conquer_ Heterogeneous Noise Integration for Diffusion-based Adversarial Purification/images.zip new file mode 100644 index 0000000000000000000000000000000000000000..82cccc33d5f41cd9d71af14b5ad163f0b671fb34 --- /dev/null +++ b/2025/Divide and Conquer_ Heterogeneous Noise Integration for Diffusion-based Adversarial Purification/images.zip @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:9adad3023315aa28c501b0f9a5d819c86da7506895eb71ad3649e2ae52c96d5b +size 737555 diff --git a/2025/Divide and Conquer_ Heterogeneous Noise Integration for Diffusion-based Adversarial Purification/layout.json b/2025/Divide and Conquer_ Heterogeneous Noise Integration for Diffusion-based Adversarial Purification/layout.json new file mode 100644 index 0000000000000000000000000000000000000000..bb84fb903033f823609f388a46ef39c20dd34737 --- /dev/null +++ b/2025/Divide and Conquer_ Heterogeneous Noise Integration for Diffusion-based Adversarial Purification/layout.json @@ -0,0 +1,10475 @@ +{ + "pdf_info": [ + { + "para_blocks": [ + { + "bbox": [ + 126, + 103, + 486, + 138 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 126, + 103, + 486, + 138 + ], + "spans": [ + { + "bbox": [ + 126, + 103, + 486, + 138 + ], + "type": "text", + "content": "Divide and Conquer: Heterogeneous Noise Integration for Diffusion-based Adversarial Purification" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 61, + 160, + 550, + 232 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 61, + 160, + 550, + 232 + ], + "spans": [ + { + "bbox": [ + 61, + 160, + 550, + 232 + ], + "type": "text", + "content": "Gaozheng Pei" + }, + { + "bbox": [ + 61, + 160, + 550, + 232 + ], + "type": "inline_equation", + "content": "^{1}" + }, + { + "bbox": [ + 61, + 160, + 550, + 232 + ], + "type": "text", + "content": " Shaojie Lyu" + }, + { + "bbox": [ + 61, + 160, + 550, + 232 + ], + "type": "inline_equation", + "content": "^{2}" + }, + { + "bbox": [ + 61, + 160, + 550, + 232 + ], + "type": "text", + "content": " Gong Chen" + }, + { + "bbox": [ + 61, + 160, + 550, + 232 + ], + "type": "inline_equation", + "content": "^{2}" + }, + { + "bbox": [ + 61, + 160, + 550, + 232 + ], + "type": "text", + "content": " Ke Ma" + }, + { + "bbox": [ + 61, + 160, + 550, + 232 + ], + "type": "inline_equation", + "content": "^{1*}" + }, + { + "bbox": [ + 61, + 160, + 550, + 232 + ], + "type": "text", + "content": " Qianqian Xu" + }, + { + "bbox": [ + 61, + 160, + 550, + 232 + ], + "type": "inline_equation", + "content": "^{3}" + }, + { + "bbox": [ + 61, + 160, + 550, + 232 + ], + "type": "text", + "content": " Yingfei Sun" + }, + { + "bbox": [ + 61, + 160, + 550, + 232 + ], + "type": "inline_equation", + "content": "^{1}" + }, + { + "bbox": [ + 61, + 160, + 550, + 232 + ], + "type": "text", + "content": " Qingming Huang" + }, + { + "bbox": [ + 61, + 160, + 550, + 232 + ], + "type": "inline_equation", + "content": "^{3,4,5*}" + }, + { + "bbox": [ + 61, + 160, + 550, + 232 + ], + "type": "inline_equation", + "content": "^{1}" + }, + { + "bbox": [ + 61, + 160, + 550, + 232 + ], + "type": "text", + "content": "School of Electronic, Electrical and Communication Engineering, UCAS " + }, + { + "bbox": [ + 61, + 160, + 550, + 232 + ], + "type": "inline_equation", + "content": "^{2}" + }, + { + "bbox": [ + 61, + 160, + 550, + 232 + ], + "type": "text", + "content": "Tencent Corporate \n" + }, + { + "bbox": [ + 61, + 160, + 550, + 232 + ], + "type": "inline_equation", + "content": "^{3}" + }, + { + "bbox": [ + 61, + 160, + 550, + 232 + ], + "type": "text", + "content": "Key Laboratory of Intelligent Information Processing, Institute of Computing Technology, CAS \n" + }, + { + "bbox": [ + 61, + 160, + 550, + 232 + ], + "type": "inline_equation", + "content": "^{4}" + }, + { + "bbox": [ + 61, + 160, + 550, + 232 + ], + "type": "text", + "content": "School of Computer Science and Technology, UCAS \n" + }, + { + "bbox": [ + 61, + 160, + 550, + 232 + ], + "type": "inline_equation", + "content": "^{5}" + }, + { + "bbox": [ + 61, + 160, + 550, + 232 + ], + "type": "text", + "content": "Key Laboratory of Big Data Mining and Knowledge Management, UCAS" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 133, + 233, + 480, + 259 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 233, + 480, + 259 + ], + "spans": [ + { + "bbox": [ + 133, + 233, + 480, + 259 + ], + "type": "text", + "content": "peigaozheng23@mails.ucas.ac.cn, {shaojielv,natchen}@tencent.com xuqianqian@ict.ac.cn, {make,yfsun,qmhuang}@ucas.ac.cn" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 151, + 286, + 200, + 300 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 151, + 286, + 200, + 300 + ], + "spans": [ + { + "bbox": [ + 151, + 286, + 200, + 300 + ], + "type": "text", + "content": "Abstract" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 54, + 312, + 297, + 624 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 54, + 312, + 297, + 624 + ], + "spans": [ + { + "bbox": [ + 54, + 312, + 297, + 624 + ], + "type": "text", + "content": "Existing diffusion-based purification methods aim to disrupt adversarial perturbations by introducing a certain amount of noise through a forward diffusion process, followed by a reverse process to recover clean examples. However, this approach is fundamentally flawed: the uniform operation of the forward process across all pixels compromises normal pixels while attempting to combat adversarial perturbations, resulting in the target model producing incorrect predictions. Simply relying on low-intensity noise is insufficient for effective defense. To address this critical issue, we implement a heterogeneous purification strategy grounded in the interpretability of neural networks. Our method decisively applies higher-intensity noise to specific pixels that the target model focuses on while the remaining pixels are subjected to only low-intensity noise. This requirement motivates us to redesign the sampling process of the diffusion model, allowing for the effective removal of varying noise levels. Furthermore, to evaluate our method against strong adaptative attack, our proposed method sharply reduces time cost and memory usage through a single-step resampling. The empirical evidence from extensive experiments across three datasets demonstrates that our method outperforms most current adversarial training and purification techniques by a substantial margin. Code is available at https://github.com/GaozhengPei/Purification." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 56, + 648, + 135, + 661 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 648, + 135, + 661 + ], + "spans": [ + { + "bbox": [ + 56, + 648, + 135, + 661 + ], + "type": "text", + "content": "1. Introduction" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 55, + 669, + 296, + 693 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 669, + 296, + 693 + ], + "spans": [ + { + "bbox": [ + 55, + 669, + 296, + 693 + ], + "type": "text", + "content": "Adversarial perturbations [9], which are often imperceptible to humans, can significantly alter the predictions of" + } + ] + } + ], + "index": 8 + }, + { + "type": "image", + "bbox": [ + 322, + 286, + 548, + 405 + ], + "blocks": [ + { + "bbox": [ + 322, + 286, + 548, + 405 + ], + "lines": [ + { + "bbox": [ + 322, + 286, + 548, + 405 + ], + "spans": [ + { + "bbox": [ + 322, + 286, + 548, + 405 + ], + "type": "image", + "image_path": "074f326a88409f96ff88a145555e55280a5bb3e943490ed93cc9063fe1721a44.jpg" + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 313, + 412, + 555, + 501 + ], + "lines": [ + { + "bbox": [ + 313, + 412, + 555, + 501 + ], + "spans": [ + { + "bbox": [ + 313, + 412, + 555, + 501 + ], + "type": "text", + "content": "Figure 1. When the noise added to adversarial samples is minimal (top), the adversarial perturbations remain intact and cannot be removed. On the other hand, if the noise intensity is excessive (bottom), it can distort the semantic information. Our approach employs an attention mask to introduce varying intensities of noise across different areas (middle). This technique effectively strikes a balance between preserving semantic information and mitigating adversarial perturbations." + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "image_caption" + } + ], + "index": 9 + }, + { + "bbox": [ + 313, + 521, + 556, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 521, + 556, + 715 + ], + "spans": [ + { + "bbox": [ + 313, + 521, + 556, + 715 + ], + "type": "text", + "content": "neural networks. To address these risks [21-23, 32], various defense mechanisms have been proposed. One common approach is adversarial training [18, 37, 45], which involves generating adversarial samples according to a specific attack method before the test phase and then retraining the model to improve its robustness. However, this approach often struggles to defend against unknown attacks. Another technique, known as adversarial purification [17, 27, 30, 38, 40, 44], seeks to remove perturbations from adversarial examples during the test phase. This approach does not rely on any assumptions about the attack methods and eliminates the overhead of repetitive training. Its flexibility and strong generalization have drawn increasing interest. In recent years, diffusion models [16, 29] have garnered attention due to their impressive denoising abilities and capacity to generate high-quality images, making them" + } + ] + } + ], + "index": 11 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "spans": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "type": "text", + "content": "CVF" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 145, + 0, + 496, + 37 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 145, + 0, + 496, + 37 + ], + "spans": [ + { + "bbox": [ + 145, + 0, + 496, + 37 + ], + "type": "text", + "content": "This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 67, + 702, + 144, + 713 + ], + "type": "page_footnote", + "angle": 0, + "lines": [ + { + "bbox": [ + 67, + 702, + 144, + 713 + ], + "spans": [ + { + "bbox": [ + 67, + 702, + 144, + 713 + ], + "type": "text", + "content": "*Corresponding author" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "type": "text", + "content": "29268" + } + ] + } + ], + "index": 13 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 0 + }, + { + "para_blocks": [ + { + "bbox": [ + 55, + 72, + 245, + 85 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 72, + 245, + 85 + ], + "spans": [ + { + "bbox": [ + 55, + 72, + 245, + 85 + ], + "type": "text", + "content": "a promising option for adversarial purification." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 55, + 87, + 296, + 278 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 87, + 296, + 278 + ], + "spans": [ + { + "bbox": [ + 55, + 87, + 296, + 278 + ], + "type": "text", + "content": "Existing diffusion-based adversarial purification methods [3, 30, 40, 44] follow the traditional forward process which attempts to add uniform noise across different regions of the images and then recover the clean images through reverse process. From Figure 1 we can observe that such a manner can not balance eliminating perturbations and preserving semantic information. If the added noise is intense, it will remove the adversarial perturbations but damage the semantic information, causing the wrong prediction of both natural and adversarial examples. In contrast, if the noise intensity is weak, it may not effectively purify the adversarial perturbations, resulting in a successful evasion attack. Therefore, how to achieve a good balance between the elimination of adversarial perturbations and the preservation of semantic information remains challenging." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 55, + 281, + 296, + 602 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 281, + 296, + 602 + ], + "spans": [ + { + "bbox": [ + 55, + 281, + 296, + 602 + ], + "type": "text", + "content": "In this paper, we propose a novel diffusion-based adversarial purification method from the perspective of neural network interpretability. During making decisions, the neural network spontaneously allocates varying levels of attention to different regions of the image. This phenomenon inspired us to use attention maps as a prior guided diffusion model to impose heterogeneous noise on different regions of an image as forward process. By eliminating adversarial perturbations in the critical areas through high-intensity noise, we can substantially mitigate their effect on the prediction of the target model. Conversely, for regions that receive less attention, we introduce low-intensity noise to preserve as much semantic information as possible in the original image. To recover a clean image from one containing heterogeneous noise, we design a two-stage denoising process in accordance with the attention map as a reverse process of the diffusion model. Specifically, if the sampling timestep falls between two different noise levels, we apply inpainting techniques to remove localized high-intensity noise. If the current timestep is below that of the smaller noise level, we use the standard denoising process to remove global weak noise. Additionally, by replacing multi-step resampling with an improved single-step resampling approach, this strategy results in a " + }, + { + "bbox": [ + 55, + 281, + 296, + 602 + ], + "type": "inline_equation", + "content": "90\\%" + }, + { + "bbox": [ + 55, + 281, + 296, + 602 + ], + "type": "text", + "content": " reduction in both time consumption and GPU memory usage enables evaluate our method against strong adaptive attacks. Overall, our contributions can be summarized as follows:" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 55, + 605, + 296, + 714 + ], + "type": "list", + "angle": 0, + "index": 5, + "blocks": [ + { + "bbox": [ + 55, + 605, + 295, + 664 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 605, + 295, + 664 + ], + "spans": [ + { + "bbox": [ + 55, + 605, + 295, + 664 + ], + "type": "text", + "content": "- We propose a novel forward process that adds noise of varying intensities based on the classifier's attention to different regions, achieving a balance between preserving semantic information and eliminating adversarial perturbations." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 55, + 666, + 296, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 666, + 296, + 714 + ], + "spans": [ + { + "bbox": [ + 55, + 666, + 296, + 714 + ], + "type": "text", + "content": "- To recover the clean samples, we design a two-stage denoising process: at a larger timestep, we convert it into an inpainting problem to remove localized strong noise, and at a smaller timestep, we apply a standard sampling" + } + ] + } + ], + "index": 4 + } + ], + "sub_type": "text" + }, + { + "bbox": [ + 322, + 72, + 482, + 84 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 322, + 72, + 482, + 84 + ], + "spans": [ + { + "bbox": [ + 322, + 72, + 482, + 84 + ], + "type": "text", + "content": "process to eliminate global weak noise." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 313, + 85, + 553, + 179 + ], + "type": "list", + "angle": 0, + "index": 9, + "blocks": [ + { + "bbox": [ + 314, + 85, + 553, + 144 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 314, + 85, + 553, + 144 + ], + "spans": [ + { + "bbox": [ + 314, + 85, + 553, + 144 + ], + "type": "text", + "content": "- To evaluate our method against strong adaptive attacks. We replace the multi-step resampling with an improved single-step resampling, this not only achieves harmonious image generation but also significantly reduces time cost and memory usage." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 313, + 144, + 553, + 179 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 144, + 553, + 179 + ], + "spans": [ + { + "bbox": [ + 313, + 144, + 553, + 179 + ], + "type": "text", + "content": "- Extensive experiments on three datasets show that our method outperforms other methods by a promising improvement against various adversarial attacks." + } + ] + } + ], + "index": 8 + } + ], + "sub_type": "text" + }, + { + "bbox": [ + 313, + 190, + 400, + 202 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 190, + 400, + 202 + ], + "spans": [ + { + "bbox": [ + 313, + 190, + 400, + 202 + ], + "type": "text", + "content": "2. Related Work" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 313, + 210, + 555, + 294 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 210, + 555, + 294 + ], + "spans": [ + { + "bbox": [ + 313, + 210, + 555, + 294 + ], + "type": "text", + "content": "Adversarial purification aims to remove adversarial perturbations from samples at test phase before feeding them into classifier model. Adversarial purification methods can be broadly divided into two categories: training-based methods, which require training on the dataset, and diffusion based methods which doesn't require training and is not necessary to access the training dataset." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 313, + 300, + 525, + 313 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 300, + 525, + 313 + ], + "spans": [ + { + "bbox": [ + 313, + 300, + 525, + 313 + ], + "type": "text", + "content": "2.1. Training-Based Adversarial Purification" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 313, + 318, + 555, + 570 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 318, + 555, + 570 + ], + "spans": [ + { + "bbox": [ + 313, + 318, + 555, + 570 + ], + "type": "text", + "content": "[41] shows empirically that adversarial examples mainly lie in the low probability regions of the training distribution and move them back towards the training distribution. [36] first models the distribution of clean samples and then attempts to find the closest sample within the clean sample distribution to the adversarial example during inference. [28] generates perturbed images using Self-supervised Perturbation (SSP) attack that disrupts the deep perceptual features and projects back the perturbed images close to the perceptual space of clean images. [52] propose to learn generalizable invariant features across attacks via encoder by playing zero-sum game and reconstruct the original image via decoder. [24] disrupt dversarial perturbations structure through random transform and reconstruct the image using a purifier model. [42] experimentally show that FGSM-AT networks overfit to unknown adversarial examples attacked again by FGSM. In the test phase, adversarial examples are fed to classifier after performing FGSM \"purification\". However, these methods require training using the training dataset, which is very time-consuming and lacks generalizability." + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 313, + 576, + 528, + 588 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 576, + 528, + 588 + ], + "spans": [ + { + "bbox": [ + 313, + 576, + 528, + 588 + ], + "type": "text", + "content": "2.2. Diffusion-Based Adversarial Purification" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 313, + 594, + 555, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 594, + 555, + 715 + ], + "spans": [ + { + "bbox": [ + 313, + 594, + 555, + 715 + ], + "type": "text", + "content": "[49] show that an EBM trained with Denoising Score-Matching can quickly purify attacked images within a few steps. [30] first diffuse samples with a small amount of noise and then recover the clean image through a reverse generative process. The paper also theoretically proves that the time step cannot be too large or too small. [44] proposes to use the adversarial image as guidance in the reverse process which encourages the purified image is consistent with the original clean image. [47] gets multiple reversed samples via multiple runs of denoising. Then the samples are" + } + ] + } + ], + "index": 15 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "type": "text", + "content": "29269" + } + ] + } + ], + "index": 16 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 1 + }, + { + "para_blocks": [ + { + "bbox": [ + 55, + 72, + 296, + 180 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 72, + 296, + 180 + ], + "spans": [ + { + "bbox": [ + 55, + 72, + 296, + 180 + ], + "type": "text", + "content": "passed into the classifier, where the final prediction is obtained through classifier voting. [3] design the forward process with the proper amount of Gaussian noise added and the reverse process with the gradient of contrastive loss as the guidance of diffusion models for adversarial purification. However, adding noise of the same intensity to different regions of the images, these methods is difficult to balance preserving semantic information and eliminating adversarial perturbations." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 55, + 190, + 138, + 205 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 190, + 138, + 205 + ], + "spans": [ + { + "bbox": [ + 55, + 190, + 138, + 205 + ], + "type": "text", + "content": "3. Methodology" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 55, + 210, + 296, + 353 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 210, + 296, + 353 + ], + "spans": [ + { + "bbox": [ + 55, + 210, + 296, + 353 + ], + "type": "text", + "content": "Given an example " + }, + { + "bbox": [ + 55, + 210, + 296, + 353 + ], + "type": "inline_equation", + "content": "(\\pmb{x},\\pmb{y})" + }, + { + "bbox": [ + 55, + 210, + 296, + 353 + ], + "type": "text", + "content": ", an adversary can create a small perturbation to generate an adversarial example " + }, + { + "bbox": [ + 55, + 210, + 296, + 353 + ], + "type": "inline_equation", + "content": "\\pmb{x}_{adv}" + }, + { + "bbox": [ + 55, + 210, + 296, + 353 + ], + "type": "text", + "content": " such that the classifier " + }, + { + "bbox": [ + 55, + 210, + 296, + 353 + ], + "type": "inline_equation", + "content": "\\pmb{f}" + }, + { + "bbox": [ + 55, + 210, + 296, + 353 + ], + "type": "text", + "content": " makes an incorrect prediction: " + }, + { + "bbox": [ + 55, + 210, + 296, + 353 + ], + "type": "inline_equation", + "content": "\\pmb{y} \\neq \\pmb{f}(\\pmb{x}_{adv})" + }, + { + "bbox": [ + 55, + 210, + 296, + 353 + ], + "type": "text", + "content": ". Existing diffusion-based purification methods obtain a noisy image " + }, + { + "bbox": [ + 55, + 210, + 296, + 353 + ], + "type": "inline_equation", + "content": "\\pmb{x}(t)" + }, + { + "bbox": [ + 55, + 210, + 296, + 353 + ], + "type": "text", + "content": " at timestep " + }, + { + "bbox": [ + 55, + 210, + 296, + 353 + ], + "type": "inline_equation", + "content": "t" + }, + { + "bbox": [ + 55, + 210, + 296, + 353 + ], + "type": "text", + "content": " through the forward process and then recover the clean (or purified) image " + }, + { + "bbox": [ + 55, + 210, + 296, + 353 + ], + "type": "inline_equation", + "content": "\\hat{\\pmb{x}}" + }, + { + "bbox": [ + 55, + 210, + 296, + 353 + ], + "type": "text", + "content": " through the reverse process. And we hope " + }, + { + "bbox": [ + 55, + 210, + 296, + 353 + ], + "type": "inline_equation", + "content": "\\pmb{f}(\\hat{\\pmb{x}}) = \\pmb{f}(\\pmb{x})" + }, + { + "bbox": [ + 55, + 210, + 296, + 353 + ], + "type": "text", + "content": ". From the theoretical analysis of [30], the timestep " + }, + { + "bbox": [ + 55, + 210, + 296, + 353 + ], + "type": "inline_equation", + "content": "t" + }, + { + "bbox": [ + 55, + 210, + 296, + 353 + ], + "type": "text", + "content": " should satisfy: 1) " + }, + { + "bbox": [ + 55, + 210, + 296, + 353 + ], + "type": "inline_equation", + "content": "t" + }, + { + "bbox": [ + 55, + 210, + 296, + 353 + ], + "type": "text", + "content": " should be large enough to eliminate perturbations. 2) " + }, + { + "bbox": [ + 55, + 210, + 296, + 353 + ], + "type": "inline_equation", + "content": "t" + }, + { + "bbox": [ + 55, + 210, + 296, + 353 + ], + "type": "text", + "content": " should be small enough that " + }, + { + "bbox": [ + 55, + 210, + 296, + 353 + ], + "type": "inline_equation", + "content": "\\hat{\\pmb{x}}" + }, + { + "bbox": [ + 55, + 210, + 296, + 353 + ], + "type": "text", + "content": " still maintains semantic consistency with " + }, + { + "bbox": [ + 55, + 210, + 296, + 353 + ], + "type": "inline_equation", + "content": "\\pmb{x}" + }, + { + "bbox": [ + 55, + 210, + 296, + 353 + ], + "type": "text", + "content": ". We find that these two objectives are contradictory." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 55, + 354, + 296, + 521 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 354, + 296, + 521 + ], + "spans": [ + { + "bbox": [ + 55, + 354, + 296, + 521 + ], + "type": "text", + "content": "Motivated by the interpretability of neural networks, the attention map is primarily used to encode the input spatial regions that the network focuses on most when making output decisions, such as in image classification. This suggests that, for an adversarial example, adversarial perturbations in the regions where the classifier focuses most significantly contribute to causing the target classifier to make incorrect predictions. The attention maps for each layer of the target classifier can be obtained with a single forward propagation. Therefore, we propose introducing high-intensity noise to eliminate adversarial perturbations of the regions with more attention value and applying low-intensity noise to the remaining area to preserve as much semantic information as possible." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 55, + 528, + 230, + 540 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 528, + 230, + 540 + ], + "spans": [ + { + "bbox": [ + 55, + 528, + 230, + 540 + ], + "type": "text", + "content": "3.1. Heterogeneous Forward Process" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 55, + 545, + 296, + 582 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 545, + 296, + 582 + ], + "spans": [ + { + "bbox": [ + 55, + 545, + 296, + 582 + ], + "type": "text", + "content": "For the adversarial example " + }, + { + "bbox": [ + 55, + 545, + 296, + 582 + ], + "type": "inline_equation", + "content": "\\pmb{x}_{adv}" + }, + { + "bbox": [ + 55, + 545, + 296, + 582 + ], + "type": "text", + "content": " at timestep " + }, + { + "bbox": [ + 55, + 545, + 296, + 582 + ], + "type": "inline_equation", + "content": "t = 0" + }, + { + "bbox": [ + 55, + 545, + 296, + 582 + ], + "type": "text", + "content": ", i.e., " + }, + { + "bbox": [ + 55, + 545, + 296, + 582 + ], + "type": "inline_equation", + "content": "\\pmb{x}(0) = \\pmb{x}_{adv}" + }, + { + "bbox": [ + 55, + 545, + 296, + 582 + ], + "type": "text", + "content": ", we can obtain the noisy images at timestep " + }, + { + "bbox": [ + 55, + 545, + 296, + 582 + ], + "type": "inline_equation", + "content": "t_l" + }, + { + "bbox": [ + 55, + 545, + 296, + 582 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 55, + 545, + 296, + 582 + ], + "type": "inline_equation", + "content": "t_s" + }, + { + "bbox": [ + 55, + 545, + 296, + 582 + ], + "type": "text", + "content": " with " + }, + { + "bbox": [ + 55, + 545, + 296, + 582 + ], + "type": "inline_equation", + "content": "t_l > t_s" + }, + { + "bbox": [ + 55, + 545, + 296, + 582 + ], + "type": "text", + "content": " as follows:" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 83, + 601, + 295, + 614 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 83, + 601, + 295, + 614 + ], + "spans": [ + { + "bbox": [ + 83, + 601, + 295, + 614 + ], + "type": "interline_equation", + "content": "\\boldsymbol {x} \\left(t _ {l}\\right) = \\sqrt {\\bar {\\alpha} \\left(t _ {l}\\right)} \\cdot \\boldsymbol {x} _ {a d v} + \\sqrt {1 - \\bar {\\alpha} \\left(t _ {l}\\right)} \\cdot \\boldsymbol {\\epsilon}, \\tag {1a}", + "image_path": "ee2a3fbf15a6034f11c4fed6f0fe5002310f989b17e7b121ef6a90060c5afe1b.jpg" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 82, + 617, + 294, + 632 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 82, + 617, + 294, + 632 + ], + "spans": [ + { + "bbox": [ + 82, + 617, + 294, + 632 + ], + "type": "interline_equation", + "content": "\\boldsymbol {x} \\left(t _ {s}\\right) = \\sqrt {\\bar {\\alpha} \\left(t _ {s}\\right)} \\cdot \\boldsymbol {x} _ {a d v} + \\sqrt {1 - \\bar {\\alpha} \\left(t _ {s}\\right)} \\cdot \\boldsymbol {\\epsilon}, \\tag {1b}", + "image_path": "1bfcfdfe8ddd63c3c225c0dea401e85189e747bb6a996e06dfc8ffb9acf74583.jpg" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 55, + 640, + 296, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 640, + 296, + 715 + ], + "spans": [ + { + "bbox": [ + 55, + 640, + 296, + 715 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 55, + 640, + 296, + 715 + ], + "type": "inline_equation", + "content": "\\{\\sqrt{\\overline{\\alpha}(t)}\\}" + }, + { + "bbox": [ + 55, + 640, + 296, + 715 + ], + "type": "text", + "content": " is a series of given parameters and " + }, + { + "bbox": [ + 55, + 640, + 296, + 715 + ], + "type": "inline_equation", + "content": "\\epsilon \\sim \\mathcal{N}(0, I)" + }, + { + "bbox": [ + 55, + 640, + 296, + 715 + ], + "type": "text", + "content": ". The clean (or purified) images " + }, + { + "bbox": [ + 55, + 640, + 296, + 715 + ], + "type": "inline_equation", + "content": "\\hat{\\pmb{x}}_{t_l}" + }, + { + "bbox": [ + 55, + 640, + 296, + 715 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 55, + 640, + 296, + 715 + ], + "type": "inline_equation", + "content": "\\hat{\\pmb{x}}_{t_s}" + }, + { + "bbox": [ + 55, + 640, + 296, + 715 + ], + "type": "text", + "content": " can be obtained through the normal reverse process [16, 30]. For the smaller timestep " + }, + { + "bbox": [ + 55, + 640, + 296, + 715 + ], + "type": "inline_equation", + "content": "t_s" + }, + { + "bbox": [ + 55, + 640, + 296, + 715 + ], + "type": "text", + "content": ", the recovered image " + }, + { + "bbox": [ + 55, + 640, + 296, + 715 + ], + "type": "inline_equation", + "content": "\\hat{\\pmb{x}}_{t_s}" + }, + { + "bbox": [ + 55, + 640, + 296, + 715 + ], + "type": "text", + "content": " retains certain adversarial perturbation, resulting in " + }, + { + "bbox": [ + 55, + 640, + 296, + 715 + ], + "type": "inline_equation", + "content": "\\pmb{f}(\\pmb{x}_{adv}) = \\pmb{f}(\\hat{\\pmb{x}}_{t_s}) \\neq \\pmb{f}(\\pmb{x})" + }, + { + "bbox": [ + 55, + 640, + 296, + 715 + ], + "type": "text", + "content": ". Conversely, at the larger timestep " + }, + { + "bbox": [ + 55, + 640, + 296, + 715 + ], + "type": "inline_equation", + "content": "t_l" + }, + { + "bbox": [ + 55, + 640, + 296, + 715 + ], + "type": "text", + "content": ", the" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 313, + 72, + 554, + 131 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 72, + 554, + 131 + ], + "spans": [ + { + "bbox": [ + 313, + 72, + 554, + 131 + ], + "type": "text", + "content": "clean image " + }, + { + "bbox": [ + 313, + 72, + 554, + 131 + ], + "type": "inline_equation", + "content": "\\hat{\\pmb{x}}_{t_l}" + }, + { + "bbox": [ + 313, + 72, + 554, + 131 + ], + "type": "text", + "content": " loses critical semantic information, leading to " + }, + { + "bbox": [ + 313, + 72, + 554, + 131 + ], + "type": "inline_equation", + "content": "\\pmb {f}(\\pmb{x}_{adv})\\neq \\pmb {f}(\\hat{\\pmb{x}}_{t_l})\\neq \\pmb {f}(\\pmb {x})" + }, + { + "bbox": [ + 313, + 72, + 554, + 131 + ], + "type": "text", + "content": " . To effectively aggregate the differing impacts on various regions of the images at distinct timesteps, we construct an attention mask to integrate (1a) and (1b)." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 313, + 133, + 554, + 180 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 133, + 554, + 180 + ], + "spans": [ + { + "bbox": [ + 313, + 133, + 554, + 180 + ], + "type": "text", + "content": "Consider a neural network that consists of " + }, + { + "bbox": [ + 313, + 133, + 554, + 180 + ], + "type": "inline_equation", + "content": "M" + }, + { + "bbox": [ + 313, + 133, + 554, + 180 + ], + "type": "text", + "content": " blocks (e.g. ResNet-50 [14] consists of 4 blocks). For " + }, + { + "bbox": [ + 313, + 133, + 554, + 180 + ], + "type": "inline_equation", + "content": "m" + }, + { + "bbox": [ + 313, + 133, + 554, + 180 + ], + "type": "text", + "content": "-th block, we extract the output of the last activation function (e.g. Relu [8] or Swish [33]) as follows:" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 329, + 188, + 554, + 225 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 329, + 188, + 554, + 225 + ], + "spans": [ + { + "bbox": [ + 329, + 188, + 554, + 225 + ], + "type": "interline_equation", + "content": "\\boldsymbol {G} _ {m} = \\left(\\sum_ {i = 1} ^ {C _ {m}} | \\boldsymbol {A} _ {m i} | ^ {p}\\right) ^ {\\frac {1}{p}} \\in \\mathbb {R} ^ {H _ {m} \\times W _ {m}}, m \\in [ M ], \\tag {2}", + "image_path": "52f58e46d1ccb4b35475f0054face950b4b1af8c860fc48a89be1b7e9a771d28.jpg" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 313, + 232, + 555, + 315 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 232, + 555, + 315 + ], + "spans": [ + { + "bbox": [ + 313, + 232, + 555, + 315 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 313, + 232, + 555, + 315 + ], + "type": "inline_equation", + "content": "p \\in \\mathbb{N}" + }, + { + "bbox": [ + 313, + 232, + 555, + 315 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 313, + 232, + 555, + 315 + ], + "type": "inline_equation", + "content": "C_m, H_m, W_m" + }, + { + "bbox": [ + 313, + 232, + 555, + 315 + ], + "type": "text", + "content": " are the dimension of channels, height, and width respectively, " + }, + { + "bbox": [ + 313, + 232, + 555, + 315 + ], + "type": "inline_equation", + "content": "\\mathbf{A}_m" + }, + { + "bbox": [ + 313, + 232, + 555, + 315 + ], + "type": "text", + "content": " denotes the activation of " + }, + { + "bbox": [ + 313, + 232, + 555, + 315 + ], + "type": "inline_equation", + "content": "m" + }, + { + "bbox": [ + 313, + 232, + 555, + 315 + ], + "type": "text", + "content": "-th block and " + }, + { + "bbox": [ + 313, + 232, + 555, + 315 + ], + "type": "inline_equation", + "content": "\\mathbf{A}_{mi}" + }, + { + "bbox": [ + 313, + 232, + 555, + 315 + ], + "type": "text", + "content": " denotes the " + }, + { + "bbox": [ + 313, + 232, + 555, + 315 + ], + "type": "inline_equation", + "content": "i" + }, + { + "bbox": [ + 313, + 232, + 555, + 315 + ], + "type": "text", + "content": "-th slice of " + }, + { + "bbox": [ + 313, + 232, + 555, + 315 + ], + "type": "inline_equation", + "content": "\\mathbf{A}_m" + }, + { + "bbox": [ + 313, + 232, + 555, + 315 + ], + "type": "text", + "content": " along the channel dimension. Then we apply interpolation techniques to ensure it matches the original image size. By upsampling " + }, + { + "bbox": [ + 313, + 232, + 555, + 315 + ], + "type": "inline_equation", + "content": "\\mathbf{G}_m" + }, + { + "bbox": [ + 313, + 232, + 555, + 315 + ], + "type": "text", + "content": " with bilinear function " + }, + { + "bbox": [ + 313, + 232, + 555, + 315 + ], + "type": "inline_equation", + "content": "\\mathrm{Bi}(\\cdot)" + }, + { + "bbox": [ + 313, + 232, + 555, + 315 + ], + "type": "text", + "content": ", we obtain attention map of " + }, + { + "bbox": [ + 313, + 232, + 555, + 315 + ], + "type": "inline_equation", + "content": "m" + }, + { + "bbox": [ + 313, + 232, + 555, + 315 + ], + "type": "text", + "content": "-th block:" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 389, + 316, + 553, + 329 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 389, + 316, + 553, + 329 + ], + "spans": [ + { + "bbox": [ + 389, + 316, + 553, + 329 + ], + "type": "interline_equation", + "content": "\\mathrm {A M} _ {m} = \\phi (\\operatorname {B i} \\left(\\boldsymbol {G} _ {m}\\right)), \\tag {3}", + "image_path": "4c16ce60866b951bee32e4ca58af074cb5f092e6f8ed93a92a1b644370e44290.jpg" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 313, + 334, + 554, + 358 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 334, + 554, + 358 + ], + "spans": [ + { + "bbox": [ + 313, + 334, + 554, + 358 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 313, + 334, + 554, + 358 + ], + "type": "inline_equation", + "content": "\\phi (\\cdot)" + }, + { + "bbox": [ + 313, + 334, + 554, + 358 + ], + "type": "text", + "content": " is the spatial softmax operation. Moreover, we build the attention mask with all attention maps " + }, + { + "bbox": [ + 313, + 334, + 554, + 358 + ], + "type": "inline_equation", + "content": "\\{\\mathbf{AM}_m\\}" + }, + { + "bbox": [ + 313, + 334, + 554, + 358 + ], + "type": "text", + "content": ":" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 383, + 366, + 553, + 399 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 383, + 366, + 553, + 399 + ], + "spans": [ + { + "bbox": [ + 383, + 366, + 553, + 399 + ], + "type": "interline_equation", + "content": "\\mathcal {M} = \\bigcup_ {m = 1} ^ {M} \\mathbb {I} [ \\mathrm {A M} _ {m} > \\tau ] \\tag {4}", + "image_path": "08b7348e197d059a52d4449656823f296ceb98f34e299a8b962eaacfae996639.jpg" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 313, + 407, + 554, + 468 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 407, + 554, + 468 + ], + "spans": [ + { + "bbox": [ + 313, + 407, + 554, + 468 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 313, + 407, + 554, + 468 + ], + "type": "inline_equation", + "content": "\\mathbb{I}[\\cdot ]" + }, + { + "bbox": [ + 313, + 407, + 554, + 468 + ], + "type": "text", + "content": " is the Iverson bracket and " + }, + { + "bbox": [ + 313, + 407, + 554, + 468 + ], + "type": "inline_equation", + "content": "\\tau" + }, + { + "bbox": [ + 313, + 407, + 554, + 468 + ], + "type": "text", + "content": " is the threshold to distinguish important and ordinary regions. With the attention mask " + }, + { + "bbox": [ + 313, + 407, + 554, + 468 + ], + "type": "inline_equation", + "content": "\\mathcal{M}" + }, + { + "bbox": [ + 313, + 407, + 554, + 468 + ], + "type": "text", + "content": " and two noisy images from Eq.(1), " + }, + { + "bbox": [ + 313, + 407, + 554, + 468 + ], + "type": "inline_equation", + "content": "\\pmb {x}(t_s),\\pmb {x}(t_l),t_s < t_l" + }, + { + "bbox": [ + 313, + 407, + 554, + 468 + ], + "type": "text", + "content": " , the result of the proposed heterogeneous forward process is" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 343, + 477, + 553, + 490 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 343, + 477, + 553, + 490 + ], + "spans": [ + { + "bbox": [ + 343, + 477, + 553, + 490 + ], + "type": "interline_equation", + "content": "\\boldsymbol {x} \\left(t _ {l}, t _ {s}\\right) = \\boldsymbol {x} \\left(t _ {l}\\right) \\odot \\mathcal {M} + \\boldsymbol {x} \\left(t _ {s}\\right) \\odot (1 - \\mathcal {M}), \\tag {5}", + "image_path": "3b3017a929dbd751296a389ed376175927f330b0f4331cfbe9273ac0bbd888aa.jpg" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 313, + 498, + 455, + 510 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 498, + 455, + 510 + ], + "spans": [ + { + "bbox": [ + 313, + 498, + 455, + 510 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 313, + 498, + 455, + 510 + ], + "type": "inline_equation", + "content": "\\odot" + }, + { + "bbox": [ + 313, + 498, + 455, + 510 + ], + "type": "text", + "content": " is the Hadamard product." + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 313, + 518, + 493, + 530 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 518, + 493, + 530 + ], + "spans": [ + { + "bbox": [ + 313, + 518, + 493, + 530 + ], + "type": "text", + "content": "3.2. Heterogeneous Denoising Process" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 313, + 536, + 554, + 570 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 536, + 554, + 570 + ], + "spans": [ + { + "bbox": [ + 313, + 536, + 554, + 570 + ], + "type": "text", + "content": "To recover the clean image from (5), we design a two-stage denoising process. For the current timestep " + }, + { + "bbox": [ + 313, + 536, + 554, + 570 + ], + "type": "inline_equation", + "content": "t" + }, + { + "bbox": [ + 313, + 536, + 554, + 570 + ], + "type": "text", + "content": ", we consider the following two cases:" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 313, + 571, + 554, + 690 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 571, + 554, + 690 + ], + "spans": [ + { + "bbox": [ + 313, + 571, + 554, + 690 + ], + "type": "text", + "content": "Stage 1 (" + }, + { + "bbox": [ + 313, + 571, + 554, + 690 + ], + "type": "inline_equation", + "content": "t_s < t < t_l" + }, + { + "bbox": [ + 313, + 571, + 554, + 690 + ], + "type": "text", + "content": ") When the current timestep " + }, + { + "bbox": [ + 313, + 571, + 554, + 690 + ], + "type": "inline_equation", + "content": "t" + }, + { + "bbox": [ + 313, + 571, + 554, + 690 + ], + "type": "text", + "content": " lies between " + }, + { + "bbox": [ + 313, + 571, + 554, + 690 + ], + "type": "inline_equation", + "content": "t_s" + }, + { + "bbox": [ + 313, + 571, + 554, + 690 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 313, + 571, + 554, + 690 + ], + "type": "inline_equation", + "content": "t_l" + }, + { + "bbox": [ + 313, + 571, + 554, + 690 + ], + "type": "text", + "content": ". For the pixels inside the attention mask " + }, + { + "bbox": [ + 313, + 571, + 554, + 690 + ], + "type": "inline_equation", + "content": "\\mathcal{M}" + }, + { + "bbox": [ + 313, + 571, + 554, + 690 + ], + "type": "text", + "content": ", they should be predicted from the previous timestep " + }, + { + "bbox": [ + 313, + 571, + 554, + 690 + ], + "type": "inline_equation", + "content": "t + 1" + }, + { + "bbox": [ + 313, + 571, + 554, + 690 + ], + "type": "text", + "content": ". But for the pixels outside the attention mask " + }, + { + "bbox": [ + 313, + 571, + 554, + 690 + ], + "type": "inline_equation", + "content": "\\mathcal{M}" + }, + { + "bbox": [ + 313, + 571, + 554, + 690 + ], + "type": "text", + "content": ", they can be obtained by further adding noise forward from (5). Therefore, this case can be viewed as an inpainting problem [25]. For the image regions outside the attention mask " + }, + { + "bbox": [ + 313, + 571, + 554, + 690 + ], + "type": "inline_equation", + "content": "\\mathcal{M}" + }, + { + "bbox": [ + 313, + 571, + 554, + 690 + ], + "type": "text", + "content": ", since only the forward process is involved. Therefore, we can get the pixel value of these regions at timestep " + }, + { + "bbox": [ + 313, + 571, + 554, + 690 + ], + "type": "inline_equation", + "content": "t" + }, + { + "bbox": [ + 313, + 571, + 554, + 690 + ], + "type": "text", + "content": " equivalently as follows:" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 348, + 700, + 553, + 715 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 348, + 700, + 553, + 715 + ], + "spans": [ + { + "bbox": [ + 348, + 700, + 553, + 715 + ], + "type": "interline_equation", + "content": "\\boldsymbol {x} (t) ^ {\\text {k n o w n}} = \\sqrt {\\bar {\\alpha} (t)} \\boldsymbol {x} _ {a d v} + \\sqrt {1 - \\bar {\\alpha} (t)} \\boldsymbol {\\epsilon}. \\tag {6}", + "image_path": "41db54d485dabb340621a3b848a9faeb0fc25c2f36a2509a343e5f863caafac3.jpg" + } + ] + } + ], + "index": 22 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "type": "text", + "content": "29270" + } + ] + } + ], + "index": 23 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 2 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 60, + 71, + 553, + 262 + ], + "blocks": [ + { + "bbox": [ + 60, + 71, + 553, + 262 + ], + "lines": [ + { + "bbox": [ + 60, + 71, + 553, + 262 + ], + "spans": [ + { + "bbox": [ + 60, + 71, + 553, + 262 + ], + "type": "image", + "image_path": "c19ed89de3269214133ea08a31903efe44e71df7d0106db9530d5170900606d4.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 55, + 272, + 555, + 307 + ], + "lines": [ + { + "bbox": [ + 55, + 272, + 555, + 307 + ], + "spans": [ + { + "bbox": [ + 55, + 272, + 555, + 307 + ], + "type": "text", + "content": "Figure 2. Pipeline of our method. Given an adversarial image, we extract the attention maps of each block during the forward propagation of the classifier, and construct an attention mask " + }, + { + "bbox": [ + 55, + 272, + 555, + 307 + ], + "type": "inline_equation", + "content": "\\mathcal{M}" + }, + { + "bbox": [ + 55, + 272, + 555, + 307 + ], + "type": "text", + "content": ". We then execute the denoising process of the diffusion model we designed. It can be seen that the samples purified by our method can be correctly classified." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "bbox": [ + 55, + 327, + 296, + 362 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 327, + 296, + 362 + ], + "spans": [ + { + "bbox": [ + 55, + 327, + 296, + 362 + ], + "type": "text", + "content": "For the image regions within the attention mask, the pixel value of these regions at timestep " + }, + { + "bbox": [ + 55, + 327, + 296, + 362 + ], + "type": "inline_equation", + "content": "t" + }, + { + "bbox": [ + 55, + 327, + 296, + 362 + ], + "type": "text", + "content": " should be predicted by the diffusion model:" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 55, + 370, + 296, + 396 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 370, + 296, + 396 + ], + "spans": [ + { + "bbox": [ + 55, + 370, + 296, + 396 + ], + "type": "interline_equation", + "content": "\\boldsymbol {x} (t) ^ {\\text {u n k n o w n}} \\sim \\mathcal {N} \\left(\\mu_ {\\theta} (\\boldsymbol {x} (t + 1), t + 1), \\Sigma_ {\\theta} (\\boldsymbol {x} (t + 1), t + 1)\\right) \\tag {7}", + "image_path": "58e101b580c0ff0cdd4d6387e03801689c3f553b9ddb30c24f652e2f661b6481.jpg" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 55, + 396, + 296, + 444 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 396, + 296, + 444 + ], + "spans": [ + { + "bbox": [ + 55, + 396, + 296, + 444 + ], + "type": "text", + "content": "Where " + }, + { + "bbox": [ + 55, + 396, + 296, + 444 + ], + "type": "inline_equation", + "content": "\\theta" + }, + { + "bbox": [ + 55, + 396, + 296, + 444 + ], + "type": "text", + "content": " is the parameters of the diffusion model. " + }, + { + "bbox": [ + 55, + 396, + 296, + 444 + ], + "type": "inline_equation", + "content": "\\mu_{\\theta}" + }, + { + "bbox": [ + 55, + 396, + 296, + 444 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 55, + 396, + 296, + 444 + ], + "type": "inline_equation", + "content": "\\Sigma_{\\theta}" + }, + { + "bbox": [ + 55, + 396, + 296, + 444 + ], + "type": "text", + "content": " are mean and variance, respectively. Then we combine " + }, + { + "bbox": [ + 55, + 396, + 296, + 444 + ], + "type": "inline_equation", + "content": "x(t)^{known}" + }, + { + "bbox": [ + 55, + 396, + 296, + 444 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 55, + 396, + 296, + 444 + ], + "type": "inline_equation", + "content": "x(t)^{unknown}" + }, + { + "bbox": [ + 55, + 396, + 296, + 444 + ], + "type": "text", + "content": " together based on the attention mask " + }, + { + "bbox": [ + 55, + 396, + 296, + 444 + ], + "type": "inline_equation", + "content": "\\mathcal{M}" + }, + { + "bbox": [ + 55, + 396, + 296, + 444 + ], + "type": "text", + "content": " to get the " + }, + { + "bbox": [ + 55, + 396, + 296, + 444 + ], + "type": "inline_equation", + "content": "\\pmb{x}(t)" + }, + { + "bbox": [ + 55, + 396, + 296, + 444 + ], + "type": "text", + "content": " at timestep " + }, + { + "bbox": [ + 55, + 396, + 296, + 444 + ], + "type": "inline_equation", + "content": "t" + }, + { + "bbox": [ + 55, + 396, + 296, + 444 + ], + "type": "text", + "content": ":" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 61, + 451, + 296, + 466 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 61, + 451, + 296, + 466 + ], + "spans": [ + { + "bbox": [ + 61, + 451, + 296, + 466 + ], + "type": "interline_equation", + "content": "\\boldsymbol {x} (t) = \\mathcal {M} \\odot \\boldsymbol {x} (t) ^ {\\text {k n o w n}} + (1 - \\mathcal {M}) \\odot \\boldsymbol {x} (t) ^ {\\text {u n k n o w n}}, \\tag {8}", + "image_path": "f5b372143aba17495e62bed369a4a4e27ee663fbda941d5d59d7a6ad0b98e26c.jpg" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 55, + 473, + 295, + 497 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 473, + 295, + 497 + ], + "spans": [ + { + "bbox": [ + 55, + 473, + 295, + 497 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 55, + 473, + 295, + 497 + ], + "type": "inline_equation", + "content": "\\odot" + }, + { + "bbox": [ + 55, + 473, + 295, + 497 + ], + "type": "text", + "content": " is Hadamard product. We can remove the localized strong noise through the Stage 1." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 55, + 498, + 296, + 581 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 498, + 296, + 581 + ], + "spans": [ + { + "bbox": [ + 55, + 498, + 296, + 581 + ], + "type": "text", + "content": "Stage 2 (" + }, + { + "bbox": [ + 55, + 498, + 296, + 581 + ], + "type": "inline_equation", + "content": "t < t_s" + }, + { + "bbox": [ + 55, + 498, + 296, + 581 + ], + "type": "text", + "content": ") When the current timestep " + }, + { + "bbox": [ + 55, + 498, + 296, + 581 + ], + "type": "inline_equation", + "content": "t" + }, + { + "bbox": [ + 55, + 498, + 296, + 581 + ], + "type": "text", + "content": " is less than " + }, + { + "bbox": [ + 55, + 498, + 296, + 581 + ], + "type": "inline_equation", + "content": "t_s" + }, + { + "bbox": [ + 55, + 498, + 296, + 581 + ], + "type": "text", + "content": ", since the noise intensity is the same both inside and outside the attention mask. Therefore, the attention mask " + }, + { + "bbox": [ + 55, + 498, + 296, + 581 + ], + "type": "inline_equation", + "content": "\\mathcal{M}" + }, + { + "bbox": [ + 55, + 498, + 296, + 581 + ], + "type": "text", + "content": " is no longer needed. At this point, we can proceed with the standard sampling process of the diffusion model. We can get " + }, + { + "bbox": [ + 55, + 498, + 296, + 581 + ], + "type": "inline_equation", + "content": "x(t)" + }, + { + "bbox": [ + 55, + 498, + 296, + 581 + ], + "type": "text", + "content": " predicted from the previous timestep " + }, + { + "bbox": [ + 55, + 498, + 296, + 581 + ], + "type": "inline_equation", + "content": "x(t + 1)" + }, + { + "bbox": [ + 55, + 498, + 296, + 581 + ], + "type": "text", + "content": " as follows:" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 62, + 590, + 295, + 604 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 590, + 295, + 604 + ], + "spans": [ + { + "bbox": [ + 62, + 590, + 295, + 604 + ], + "type": "interline_equation", + "content": "\\boldsymbol {x} (t) \\sim \\mathcal {N} \\left(\\mu_ {\\theta} (\\boldsymbol {x} (t + 1), t + 1), \\Sigma_ {\\theta} (\\boldsymbol {x} (t + 1), t + 1)\\right) \\tag {9}", + "image_path": "51c8ce898e7a7ef2aef9361d6edfaba2520133c42a4747949b18c1c872b23e41.jpg" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 55, + 611, + 296, + 659 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 611, + 296, + 659 + ], + "spans": [ + { + "bbox": [ + 55, + 611, + 296, + 659 + ], + "type": "text", + "content": "Following Stage 2, we will eliminate global weak noise and get the clean image. The complete algorithm process can be found in the Algorithm 1. Additionally, there are some visualizations of our method in the Figure 5." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 55, + 670, + 282, + 683 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 670, + 282, + 683 + ], + "spans": [ + { + "bbox": [ + 55, + 670, + 282, + 683 + ], + "type": "text", + "content": "4. Adaptive Attack to Diffusion Purification" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 55, + 689, + 296, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 689, + 296, + 714 + ], + "spans": [ + { + "bbox": [ + 55, + 689, + 296, + 714 + ], + "type": "text", + "content": "When directly applying the method described in 3.2, we observe the same phenomenon as [25] that although the region" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 315, + 327, + 508, + 338 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 315, + 327, + 508, + 338 + ], + "spans": [ + { + "bbox": [ + 315, + 327, + 508, + 338 + ], + "type": "text", + "content": "Algorithm 1 Heterogeneous Denoising Process" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 315, + 342, + 554, + 365 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 315, + 342, + 554, + 365 + ], + "spans": [ + { + "bbox": [ + 315, + 342, + 554, + 365 + ], + "type": "text", + "content": "Require: Sample " + }, + { + "bbox": [ + 315, + 342, + 554, + 365 + ], + "type": "inline_equation", + "content": "x_0" + }, + { + "bbox": [ + 315, + 342, + 554, + 365 + ], + "type": "text", + "content": ", Attention Mask " + }, + { + "bbox": [ + 315, + 342, + 554, + 365 + ], + "type": "inline_equation", + "content": "\\mathcal{M}" + }, + { + "bbox": [ + 315, + 342, + 554, + 365 + ], + "type": "text", + "content": ", timestep " + }, + { + "bbox": [ + 315, + 342, + 554, + 365 + ], + "type": "inline_equation", + "content": "t_s" + }, + { + "bbox": [ + 315, + 342, + 554, + 365 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 315, + 342, + 554, + 365 + ], + "type": "inline_equation", + "content": "t_l" + }, + { + "bbox": [ + 315, + 342, + 554, + 365 + ], + "type": "text", + "content": ", Diffusion model parameters " + }, + { + "bbox": [ + 315, + 342, + 554, + 365 + ], + "type": "inline_equation", + "content": "\\theta" + }, + { + "bbox": [ + 315, + 342, + 554, + 365 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 324, + 366, + 402, + 377 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 324, + 366, + 402, + 377 + ], + "spans": [ + { + "bbox": [ + 324, + 366, + 402, + 377 + ], + "type": "text", + "content": "for " + }, + { + "bbox": [ + 324, + 366, + 402, + 377 + ], + "type": "inline_equation", + "content": "t = t_l,\\dots,1" + }, + { + "bbox": [ + 324, + 366, + 402, + 377 + ], + "type": "text", + "content": " do" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 334, + 378, + 412, + 389 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 334, + 378, + 412, + 389 + ], + "spans": [ + { + "bbox": [ + 334, + 378, + 412, + 389 + ], + "type": "interline_equation", + "content": "\\text {i f} t _ {s} \\leq t \\leq t _ {l} \\text {t h e n}", + "image_path": "0f174ad63652b9787993d1ac2f613584d3efd8068a72a78a1992b53641d824ba.jpg" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 512, + 378, + 550, + 389 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 512, + 378, + 550, + 389 + ], + "spans": [ + { + "bbox": [ + 512, + 378, + 550, + 389 + ], + "type": "text", + "content": "Stage 1" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 345, + 390, + 476, + 402 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 345, + 390, + 476, + 402 + ], + "spans": [ + { + "bbox": [ + 345, + 390, + 476, + 402 + ], + "type": "interline_equation", + "content": "\\epsilon \\sim \\mathcal {N} (0, \\mathbf {I}) \\text {i f} t > 1, \\text {e l s e} \\epsilon = 0", + "image_path": "dd840f4fb14a4d5045b35b7df8720b71d459bc1ff1469a293f578c07776eee9e.jpg" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 345, + 402, + 491, + 414 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 345, + 402, + 491, + 414 + ], + "spans": [ + { + "bbox": [ + 345, + 402, + 491, + 414 + ], + "type": "interline_equation", + "content": "x _ {t - 1} ^ {\\text {k n o w n}} = \\sqrt {\\overline {{\\alpha}} _ {t - 1}} x _ {0} + \\sqrt {1 - \\overline {{\\alpha}} _ {t - 1}} \\epsilon", + "image_path": "e162a13f50e35783487daaa096f625c6f2409bc2d2c7213e391ebcbab7a8405a.jpg" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 346, + 414, + 477, + 426 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 346, + 414, + 477, + 426 + ], + "spans": [ + { + "bbox": [ + 346, + 414, + 477, + 426 + ], + "type": "interline_equation", + "content": "z \\sim \\mathcal {N} (0, \\mathbf {I}) \\text {i f} t > 1, \\text {e l s e} z = 0", + "image_path": "4de6d9de5fca7df664df3298ee289ddc1d0df3e1b85a108ca5207d7f8465bee5.jpg" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 345, + 426, + 541, + 445 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 345, + 426, + 541, + 445 + ], + "spans": [ + { + "bbox": [ + 345, + 426, + 541, + 445 + ], + "type": "interline_equation", + "content": "x _ {t - 1} ^ {u n k n o w n} = \\frac {1}{\\sqrt {\\alpha_ {t}}} \\left(x _ {t} - \\frac {\\beta_ {t}}{\\sqrt {1 - \\bar {\\alpha} _ {t}}} \\epsilon_ {\\theta} (x _ {t}, t)\\right) + \\sigma_ {t} z", + "image_path": "4d71fa9537543314f52e31dc3a5881faf3acfdfc416eb53b5e935e71ab7b5ef2.jpg" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 345, + 445, + 534, + 456 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 345, + 445, + 534, + 456 + ], + "spans": [ + { + "bbox": [ + 345, + 445, + 534, + 456 + ], + "type": "interline_equation", + "content": "x _ {t - 1} = \\mathcal {M} \\odot x _ {t - 1} ^ {\\text {k n o w n}} + (1 - \\mathcal {M}) \\odot x _ {t - 1} ^ {\\text {u n k n o w n}}", + "image_path": "8ecd4bbba0ea6e64df2ce062610a9073ac89037e29219487d0ffc4a64bfdede9.jpg" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 334, + 456, + 410, + 468 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 334, + 456, + 410, + 468 + ], + "spans": [ + { + "bbox": [ + 334, + 456, + 410, + 468 + ], + "type": "text", + "content": "else if " + }, + { + "bbox": [ + 334, + 456, + 410, + 468 + ], + "type": "inline_equation", + "content": "t < t_s" + }, + { + "bbox": [ + 334, + 456, + 410, + 468 + ], + "type": "text", + "content": " then" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 514, + 457, + 554, + 468 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 514, + 457, + 554, + 468 + ], + "spans": [ + { + "bbox": [ + 514, + 457, + 554, + 468 + ], + "type": "text", + "content": "Stage 2" + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 345, + 469, + 477, + 480 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 345, + 469, + 477, + 480 + ], + "spans": [ + { + "bbox": [ + 345, + 469, + 477, + 480 + ], + "type": "interline_equation", + "content": "z \\sim \\mathcal {N} (0, \\mathbf {I}) \\text {i f} t > 1, \\text {e l s e} z = 0", + "image_path": "8107a0b19f0f9578a3f62fc6bdb0d564d107f4a1cde8f09b28dcc513c6c779cd.jpg" + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 345, + 480, + 520, + 498 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 345, + 480, + 520, + 498 + ], + "spans": [ + { + "bbox": [ + 345, + 480, + 520, + 498 + ], + "type": "interline_equation", + "content": "x _ {t - 1} = \\frac {1}{\\sqrt {\\alpha_ {t}}} \\left(x _ {t} - \\frac {\\beta_ {t}}{\\sqrt {1 - \\bar {\\alpha} _ {t}}} \\epsilon_ {\\theta} (x _ {t}, t)\\right) + \\sigma_ {t} z", + "image_path": "7cf8af9c27aeb01a95a4855a6e1dbda0bcf58e84eb0cbe34d3b051a73a059eb5.jpg" + } + ] + } + ], + "index": 25 + }, + { + "bbox": [ + 335, + 498, + 419, + 509 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 335, + 498, + 419, + 509 + ], + "spans": [ + { + "bbox": [ + 335, + 498, + 419, + 509 + ], + "type": "text", + "content": "for " + }, + { + "bbox": [ + 335, + 498, + 419, + 509 + ], + "type": "inline_equation", + "content": "u = 1,2,\\ldots U" + }, + { + "bbox": [ + 335, + 498, + 419, + 509 + ], + "type": "text", + "content": " do" + } + ] + } + ], + "index": 26 + }, + { + "bbox": [ + 493, + 498, + 551, + 510 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 493, + 498, + 551, + 510 + ], + "spans": [ + { + "bbox": [ + 493, + 498, + 551, + 510 + ], + "type": "inline_equation", + "content": "\\triangleright" + }, + { + "bbox": [ + 493, + 498, + 551, + 510 + ], + "type": "text", + "content": " Resampling" + } + ] + } + ], + "index": 27 + }, + { + "bbox": [ + 345, + 510, + 477, + 521 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 345, + 510, + 477, + 521 + ], + "spans": [ + { + "bbox": [ + 345, + 510, + 477, + 521 + ], + "type": "interline_equation", + "content": "z \\sim \\mathcal {N} (0, \\mathbf {I}) \\text {i f} t > 1, \\text {e l s e} z = 0", + "image_path": "3170d4592fbe0a55117f777ee990f6fa6107b1e052ad8d44f60fa5a3d3c3a00f.jpg" + } + ] + } + ], + "index": 28 + }, + { + "bbox": [ + 345, + 521, + 533, + 534 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 345, + 521, + 533, + 534 + ], + "spans": [ + { + "bbox": [ + 345, + 521, + 533, + 534 + ], + "type": "interline_equation", + "content": "x _ {t - 1 + u} = \\sqrt {\\alpha_ {t - 1 + u}} x _ {t + u - 2} + \\sqrt {1 - \\alpha_ {t - 1 + u}} z", + "image_path": "3128ef4c360d587eb37c98dcc7e456561f58eada44528b9b5d10cee0102efa01.jpg" + } + ] + } + ], + "index": 29 + }, + { + "bbox": [ + 335, + 535, + 368, + 544 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 335, + 535, + 368, + 544 + ], + "spans": [ + { + "bbox": [ + 335, + 535, + 368, + 544 + ], + "type": "text", + "content": "end for" + } + ] + } + ], + "index": 30 + }, + { + "bbox": [ + 335, + 546, + 414, + 558 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 335, + 546, + 414, + 558 + ], + "spans": [ + { + "bbox": [ + 335, + 546, + 414, + 558 + ], + "type": "text", + "content": "get " + }, + { + "bbox": [ + 335, + 546, + 414, + 558 + ], + "type": "inline_equation", + "content": "x_{t - 1}" + }, + { + "bbox": [ + 335, + 546, + 414, + 558 + ], + "type": "text", + "content": " using (11)" + } + ] + } + ], + "index": 31 + }, + { + "bbox": [ + 325, + 559, + 358, + 567 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 325, + 559, + 358, + 567 + ], + "spans": [ + { + "bbox": [ + 325, + 559, + 358, + 567 + ], + "type": "text", + "content": "end for" + } + ] + } + ], + "index": 32 + }, + { + "bbox": [ + 313, + 590, + 554, + 650 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 590, + 554, + 650 + ], + "spans": [ + { + "bbox": [ + 313, + 590, + 554, + 650 + ], + "type": "text", + "content": "recovered in Stage 2 matches the texture of the neighboring region, it is semantically incorrect. We can observe this phenomenon in the Figure 3. [25] addresses this issue via multi-step resampling by diffusing the output " + }, + { + "bbox": [ + 313, + 590, + 554, + 650 + ], + "type": "inline_equation", + "content": "\\pmb{x}(t)" + }, + { + "bbox": [ + 313, + 590, + 554, + 650 + ], + "type": "text", + "content": " back to " + }, + { + "bbox": [ + 313, + 590, + 554, + 650 + ], + "type": "inline_equation", + "content": "\\pmb{x}(t + 1)" + }, + { + "bbox": [ + 313, + 590, + 554, + 650 + ], + "type": "text", + "content": ":" + } + ] + } + ], + "index": 33 + }, + { + "bbox": [ + 353, + 657, + 553, + 671 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 353, + 657, + 553, + 671 + ], + "spans": [ + { + "bbox": [ + 353, + 657, + 553, + 671 + ], + "type": "interline_equation", + "content": "\\boldsymbol {x} (t + 1) = \\sqrt {\\alpha (t)} \\boldsymbol {x} (t) + \\sqrt {1 - \\alpha (t)} \\boldsymbol {\\epsilon}, \\tag {10}", + "image_path": "fcde5bb84e3df24cd56972bbbbab962c5cca741bbc772300368ee4333bbf85c9.jpg" + } + ] + } + ], + "index": 34 + }, + { + "bbox": [ + 313, + 677, + 554, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 677, + 554, + 714 + ], + "spans": [ + { + "bbox": [ + 313, + 677, + 554, + 714 + ], + "type": "text", + "content": "and it is necessary to fully execute 3.2 again to get " + }, + { + "bbox": [ + 313, + 677, + 554, + 714 + ], + "type": "inline_equation", + "content": "\\pmb{x}(t)" + }, + { + "bbox": [ + 313, + 677, + 554, + 714 + ], + "type": "text", + "content": ". From Figure 3 we can find that this process needs to be repeated " + }, + { + "bbox": [ + 313, + 677, + 554, + 714 + ], + "type": "inline_equation", + "content": "U = 20" + }, + { + "bbox": [ + 313, + 677, + 554, + 714 + ], + "type": "text", + "content": " times to ensure that the masked region" + } + ] + } + ], + "index": 35 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "text", + "content": "29271" + } + ] + } + ], + "index": 36 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 3 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 66, + 70, + 289, + 185 + ], + "blocks": [ + { + "bbox": [ + 66, + 70, + 289, + 185 + ], + "lines": [ + { + "bbox": [ + 66, + 70, + 289, + 185 + ], + "spans": [ + { + "bbox": [ + 66, + 70, + 289, + 185 + ], + "type": "image", + "image_path": "8346b0525140df04bf24c7cd2ebd2317d0e0571c55a7d25ec030b94ea26ba3a0.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 55, + 193, + 295, + 239 + ], + "lines": [ + { + "bbox": [ + 55, + 193, + 295, + 239 + ], + "spans": [ + { + "bbox": [ + 55, + 193, + 295, + 239 + ], + "type": "text", + "content": "Figure 3. Our method maintains semantic consistency between the masked areas and the surrounding pixel values when " + }, + { + "bbox": [ + 55, + 193, + 295, + 239 + ], + "type": "inline_equation", + "content": "U = 10" + }, + { + "bbox": [ + 55, + 193, + 295, + 239 + ], + "type": "text", + "content": ". Note that no matter what " + }, + { + "bbox": [ + 55, + 193, + 295, + 239 + ], + "type": "inline_equation", + "content": "U" + }, + { + "bbox": [ + 55, + 193, + 295, + 239 + ], + "type": "text", + "content": " is, our method requires only one additional call to the denoising network." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "bbox": [ + 55, + 259, + 296, + 342 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 259, + 296, + 342 + ], + "spans": [ + { + "bbox": [ + 55, + 259, + 296, + 342 + ], + "type": "text", + "content": "becomes semantically consistent with the surrounding pixels. To evaluate our method against strong adaptive attacks [30], it is necessary to compute the full gradient of the whole sampling process to construct adversarial examples. To reduce the time cost and memory usage, we optimized the multi-step resampling into a single resampling process which achieves the same functionality." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 55, + 343, + 296, + 392 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 343, + 296, + 392 + ], + "spans": [ + { + "bbox": [ + 55, + 343, + 296, + 392 + ], + "type": "text", + "content": "The intuition behind our improvement is that repeating (10) and 3.2 for " + }, + { + "bbox": [ + 55, + 343, + 296, + 392 + ], + "type": "inline_equation", + "content": "U" + }, + { + "bbox": [ + 55, + 343, + 296, + 392 + ], + "type": "text", + "content": " times can be replaced by that first diffuse " + }, + { + "bbox": [ + 55, + 343, + 296, + 392 + ], + "type": "inline_equation", + "content": "\\pmb{x}(t)" + }, + { + "bbox": [ + 55, + 343, + 296, + 392 + ], + "type": "text", + "content": " to " + }, + { + "bbox": [ + 55, + 343, + 296, + 392 + ], + "type": "inline_equation", + "content": "\\pmb{x}(t + U)" + }, + { + "bbox": [ + 55, + 343, + 296, + 392 + ], + "type": "text", + "content": " and then denoise it back to " + }, + { + "bbox": [ + 55, + 343, + 296, + 392 + ], + "type": "inline_equation", + "content": "\\pmb{x}(t)" + }, + { + "bbox": [ + 55, + 343, + 296, + 392 + ], + "type": "text", + "content": " from " + }, + { + "bbox": [ + 55, + 343, + 296, + 392 + ], + "type": "inline_equation", + "content": "\\pmb{x}(t + U)" + }, + { + "bbox": [ + 55, + 343, + 296, + 392 + ], + "type": "text", + "content": " in a single step using DDIM [39] as follows:" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 59, + 401, + 294, + 459 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 59, + 401, + 294, + 459 + ], + "spans": [ + { + "bbox": [ + 59, + 401, + 294, + 459 + ], + "type": "interline_equation", + "content": "\\begin{array}{l} \\boldsymbol {x} _ {t - 1} = \\sqrt {\\alpha_ {t - 1}} \\left(\\frac {\\boldsymbol {x} _ {t - 1 + U} - \\sqrt {1 - \\alpha_ {t - 1 + U}} \\epsilon_ {\\theta} (\\boldsymbol {x} _ {t - 1 + U})}{\\sqrt {\\alpha_ {t - 1 + U}}}\\right) \\\\ + \\sqrt {1 - \\alpha_ {t - 1} - \\sigma_ {t - 1 + U} ^ {2}} \\cdot \\epsilon_ {\\theta} \\left(\\boldsymbol {x} _ {t - 1 + U}\\right) + \\sigma_ {t - 1 + U} \\boldsymbol {\\epsilon}. \\tag {11} \\\\ \\end{array}", + "image_path": "66e67103cf617190a8b93c4c3aefce4e012ee68513c059add036ade4e5ac481f.jpg" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 55, + 460, + 296, + 581 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 460, + 296, + 581 + ], + "spans": [ + { + "bbox": [ + 55, + 460, + 296, + 581 + ], + "type": "text", + "content": "From the bottom panel of Figure 3, we find that our method can also achieve semantically consistency when " + }, + { + "bbox": [ + 55, + 460, + 296, + 581 + ], + "type": "inline_equation", + "content": "U = 10" + }, + { + "bbox": [ + 55, + 460, + 296, + 581 + ], + "type": "text", + "content": ". This means that at each timestep, we only need to invoke the denoising network an additional time. Therefore, the GPU memory needs to store the " + }, + { + "bbox": [ + 55, + 460, + 296, + 581 + ], + "type": "inline_equation", + "content": "2 * T" + }, + { + "bbox": [ + 55, + 460, + 296, + 581 + ], + "type": "text", + "content": " complete computation graphs of the denoising network which enables the computation of the gradient of our method on consumer-grade GPU with only 24G memory. Also, the time cost is greatly reduced. The analysis of time cost and memory usage can refer to the Appendix." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 55, + 592, + 132, + 605 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 592, + 132, + 605 + ], + "spans": [ + { + "bbox": [ + 55, + 592, + 132, + 605 + ], + "type": "text", + "content": "5. Experiment" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 55, + 612, + 182, + 625 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 612, + 182, + 625 + ], + "spans": [ + { + "bbox": [ + 55, + 612, + 182, + 625 + ], + "type": "text", + "content": "5.1. Experimental Settings" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 55, + 629, + 296, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 629, + 296, + 713 + ], + "spans": [ + { + "bbox": [ + 55, + 629, + 296, + 713 + ], + "type": "text", + "content": "Datasets and network architectures. Three datasets is used to evaluate including CIFAR-10, SVHN, and ImageNet [6]. We compare our results with several leading defense methods as listed in the standardized benchmark RobustBench [5] for CIFAR-10 and ImageNet, and also test different adversarial purification methods. We use two common classifier architectures: WideResNet-28-10" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 313, + 72, + 553, + 107 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 72, + 553, + 107 + ], + "spans": [ + { + "bbox": [ + 313, + 72, + 553, + 107 + ], + "type": "text", + "content": "and WideResNet-70-16 [51] for CIFAR-10. For SVHN, WideResNet-28-10 serves as the backbone, while for ImageNet, ResNet-50 [14] is used as the backbone." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 313, + 108, + 555, + 262 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 108, + 555, + 262 + ], + "spans": [ + { + "bbox": [ + 313, + 108, + 555, + 262 + ], + "type": "text", + "content": "Adversarial attack methods. Strong adaptive attack is tested for our method and other adversarial purification methods. We use the commonly known AutoAttack [4] under " + }, + { + "bbox": [ + 313, + 108, + 555, + 262 + ], + "type": "inline_equation", + "content": "\\ell_{\\infty}" + }, + { + "bbox": [ + 313, + 108, + 555, + 262 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 313, + 108, + 555, + 262 + ], + "type": "inline_equation", + "content": "\\ell_{2}" + }, + { + "bbox": [ + 313, + 108, + 555, + 262 + ], + "type": "text", + "content": " threat models. Additionally, the projected gradient descent (PGD) attack [26] is evaluated on our method, as recommended in [20]. Due to the randomness introduced by the diffusion and denoising processes, Expectation Over Time (EOT) [1] is adapted for these adaptive attacks. Moreover, we apply the BPDA+EOT [15] attack to ensure a fair comparison with other adversarial purification methods. Finally, following the suggestion of [20], a surrogate process is used to obtain the gradient of the reverse process for white-box attacks." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 313, + 263, + 554, + 346 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 263, + 554, + 346 + ], + "spans": [ + { + "bbox": [ + 313, + 263, + 554, + 346 + ], + "type": "text", + "content": "Pre-trained Models. We use the unconditional CIFAR-10 checkpoint of EDM provided by NVIDIA [19] on CIFAR-10. 256x256 diffusion (unconditional) checkpoint from guided-diffusion library is adopted for the ImageNet experiments. For CIFAR-10, the pre-trained classifier is downloaded from RobustBench [5], while for ImageNet, the weight of the classifier is provided by PyTorch library." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 313, + 347, + 554, + 454 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 347, + 554, + 454 + ], + "spans": [ + { + "bbox": [ + 313, + 347, + 554, + 454 + ], + "type": "text", + "content": "Evaluation metrics. To assess the performance of defense methods, we use two metrics: standard accuracy, measured on clean samples, and robust accuracy, measured on adversarial examples. Due to the high computational cost of evaluating models against adaptive attacks, a fixed subset of 512 images are randomly sampled from the test set for robust evaluation, which is the same as [20, 30]. In all experiments, the mean and standard deviation over three runs are reported to assess both standard and robust accuracy." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 313, + 455, + 554, + 562 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 455, + 554, + 562 + ], + "spans": [ + { + "bbox": [ + 313, + 455, + 554, + 562 + ], + "type": "text", + "content": "Implementation details. We follow the settings described in [20]. We evaluate diffusion-based purification methods using the PGD+EOT attack with 200 update iterations, and BPDA and AutoAttack with 100 update iterations, except for ImageNet, which uses 20 iterations. We set the number of EOT to 20. The step size is 0.007. For randomized defenses, such as [3, 20, 30, 40], we use the random version of AutoAttack, and for static defenses, we use the standard version." + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 313, + 574, + 438, + 586 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 574, + 438, + 586 + ], + "spans": [ + { + "bbox": [ + 313, + 574, + 438, + 586 + ], + "type": "text", + "content": "5.2. Experimental Results" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 313, + 594, + 555, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 594, + 555, + 713 + ], + "spans": [ + { + "bbox": [ + 313, + 594, + 555, + 713 + ], + "type": "text", + "content": "CIFAR-10. We conducted comprehensive experiments on the CIFAR-10 dataset, evaluating our method against other approaches on two model architectures: WideResNet28-10 and WideResNet-70-16. We assessed robust accuracy under three attack types: BPDA+EOT, PGD+EOT, and AutoAttack. We tested against the infinity norm attack, and as shown in Table 1, when the backbone is WideResNet28-10, our method significantly outperforms other baselines in both standard accuracy and robust accuracy, exceeding by " + }, + { + "bbox": [ + 313, + 594, + 555, + 713 + ], + "type": "inline_equation", + "content": "3.00\\%" + }, + { + "bbox": [ + 313, + 594, + 555, + 713 + ], + "type": "text", + "content": " in standard accuracy and leading by " + }, + { + "bbox": [ + 313, + 594, + 555, + 713 + ], + "type": "inline_equation", + "content": "3.56\\%" + }, + { + "bbox": [ + 313, + 594, + 555, + 713 + ], + "type": "text", + "content": " and" + } + ] + } + ], + "index": 15 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "type": "text", + "content": "29272" + } + ] + } + ], + "index": 16 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 4 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 56, + 134, + 299, + 301 + ], + "blocks": [ + { + "bbox": [ + 55, + 70, + 297, + 127 + ], + "lines": [ + { + "bbox": [ + 55, + 70, + 297, + 127 + ], + "spans": [ + { + "bbox": [ + 55, + 70, + 297, + 127 + ], + "type": "text", + "content": "Table 1. Standard and robust accuracy of different Adversarial Training (AT) and Adversarial Purification (AP) methods against PGD+EOT and AutoAttack " + }, + { + "bbox": [ + 55, + 70, + 297, + 127 + ], + "type": "inline_equation", + "content": "\\ell_{\\infty}(\\epsilon = 8 / 255)" + }, + { + "bbox": [ + 55, + 70, + 297, + 127 + ], + "type": "text", + "content": " on CIFAR-10. * utilizes half number of iterations for the attack due to the high computational cost. † indicates the requirement of extra data." + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 56, + 134, + 299, + 301 + ], + "lines": [ + { + "bbox": [ + 56, + 134, + 299, + 301 + ], + "spans": [ + { + "bbox": [ + 56, + 134, + 299, + 301 + ], + "type": "table", + "html": "
TypeMethodStandard Acc.Robust Acc.
PGDAutoAttack
WideResNet-28-10
ATGowal et al. [13]88.5465.9363.38
Gowal et al. [11]†87.5166.0162.76
Pang et al. [31]88.6264.9561.04
APYoon et al. [50]85.66±0.5133.48±0.8659.53±0.87
Nie et al. [30]90.07±0.9756.84±0.5963.6±0.81
Lee et al. [20]90.16±0.6455.82±0.5970.47±1.53
Bai et al. [3]91.4149.22*77.08
Geigh et al. [53]84.20-59.14
Lin et al. [24]90.62-72.85
Ours93.16±0.2159.38±1.2580.45±1.54
", + "image_path": "e3afc5aa41d175559558e362ab26c27a625ac3d712e67c9b4f4725c6d1fbb065.jpg" + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "table_body" + } + ], + "index": 1 + }, + { + "type": "table", + "bbox": [ + 56, + 392, + 301, + 569 + ], + "blocks": [ + { + "bbox": [ + 55, + 316, + 297, + 381 + ], + "lines": [ + { + "bbox": [ + 55, + 316, + 297, + 381 + ], + "spans": [ + { + "bbox": [ + 55, + 316, + 297, + 381 + ], + "type": "text", + "content": "Table 2. Standard and robust accuracy of different Adversarial Training (AT) and Adversarial Purification (AP) methods against PGD+EOT and AutoAttack " + }, + { + "bbox": [ + 55, + 316, + 297, + 381 + ], + "type": "inline_equation", + "content": "\\ell_{\\infty}(\\epsilon = 8 / 255)" + }, + { + "bbox": [ + 55, + 316, + 297, + 381 + ], + "type": "text", + "content": " on CIFAR-10. * The number of iterations for the attack is half that of the other methods for less computational overhead. † indicates the requirement of extra data." + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 56, + 392, + 301, + 569 + ], + "lines": [ + { + "bbox": [ + 56, + 392, + 301, + 569 + ], + "spans": [ + { + "bbox": [ + 56, + 392, + 301, + 569 + ], + "type": "table", + "html": "
TypeMethodStandard Acc.Robust Acc.
PGDAutoAttack
WideResNet-70-16
ATRebuffi et al. [34]†92.2269.9766.56
Gowal et al. [10]†91.1068.6666.10
Gowal et al. [12]88.7569.0365.87
APYoon et al. [50]86.76±1.1537.11±1.3560.86±0.56
Nie et al. [30]90.43±0.6051.13±0.8766.06±1.17
Lee et al. [20]90.53±0.156.88±1.0670.31±0.62
Bai et al. [3]92.9748.83*79.10
Geigh et al. [53]84.60-66.40
Geigh et al. [53]86.90-59.20
Lin et al. [24]91.99-76.37
Ours93.36±0.5959.79±0.6184.83±0.59
", + "image_path": "e800e50078b020cd088646e9072eef5aa5c8881745de55becdf24ed8b356dbc0.jpg" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "table_body" + } + ], + "index": 3 + }, + { + "bbox": [ + 54, + 594, + 297, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 54, + 594, + 297, + 715 + ], + "spans": [ + { + "bbox": [ + 54, + 594, + 297, + 715 + ], + "type": "text", + "content": "3.37% under PGD and AutoAttack, respectively. From Table 2, we see that when WideResNet-70-16 is used as the backbone, our method exceeds the state-of-the-art (SOTA) by 1.37% in standard accuracy and leads by 2.91% and 5.73% under PGD and AutoAttack, respectively. We also evaluated the accuracy of our method and other baselines under the Euclidean norm attack. From Table 3, we observe that, with WideResNet-28-10 as the backbone, although our method lags behind [3] by 0.59% in standard accuracy, it leads by 0.98% and 3.77% in robust accuracy. Due to the" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 313, + 70, + 555, + 149 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 70, + 555, + 149 + ], + "spans": [ + { + "bbox": [ + 313, + 70, + 555, + 149 + ], + "type": "text", + "content": "Table 3. Standard and robust accuracy against PGD+EOT and AutoAttack " + }, + { + "bbox": [ + 313, + 70, + 555, + 149 + ], + "type": "inline_equation", + "content": "\\ell_2(\\epsilon = 0.5)" + }, + { + "bbox": [ + 313, + 70, + 555, + 149 + ], + "type": "text", + "content": " on CIFAR-10. Adversarial Training (AT) and Adversarial Purification (AP) methods are evaluated. *The number of iterations for the attack is half that of the other methods for less computational overhead. † indicates the requirement of extra data. ‡ adopts WideResNet-34-10 as the backbone, with the same width but more layers than the default one." + } + ] + } + ], + "index": 5 + }, + { + "type": "table", + "bbox": [ + 316, + 156, + 553, + 335 + ], + "blocks": [ + { + "bbox": [ + 316, + 156, + 553, + 335 + ], + "lines": [ + { + "bbox": [ + 316, + 156, + 553, + 335 + ], + "spans": [ + { + "bbox": [ + 316, + 156, + 553, + 335 + ], + "type": "table", + "html": "
TypeMethodStandard Acc.Robust Acc.
PGDAutoAttack
WideResNet-28-10
ATRebuffi et al. [34]†91.7985.0578.80
Augustin et al. [2]‡93.9686.1478.79
Sehwag et al. [31]‡90.9383.7577.24
APYoon et al. [50]85.66±0.5173.32±0.7679.57±0.38
Nie et al. [30]91.41±1.0079.45±1.1681.7±0.84
Lee et al [20]90.16±0.6483.59±0.8886.48±0.38
Bai et al. [3]93.7586.13*80.92
Geigh et al. [53]84.40-77.90
Geigh et al. [53]84.20-73.60
Lin et al. [24]90.62-80.47
Ours93.16±0.2187.11±0.9590.25±0.80
", + "image_path": "286e2b0967827d4023ae69b8604200f31bc661056c464f9fc56067fbf7f0c39f.jpg" + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "table_body" + } + ], + "index": 6 + }, + { + "type": "table", + "bbox": [ + 315, + 413, + 556, + 570 + ], + "blocks": [ + { + "bbox": [ + 313, + 349, + 556, + 407 + ], + "lines": [ + { + "bbox": [ + 313, + 349, + 556, + 407 + ], + "spans": [ + { + "bbox": [ + 313, + 349, + 556, + 407 + ], + "type": "text", + "content": "Table 4. Standard and robust accuracy against PGD+EOT and AutoAttack " + }, + { + "bbox": [ + 313, + 349, + 556, + 407 + ], + "type": "inline_equation", + "content": "\\ell_2(\\epsilon = 0.5)" + }, + { + "bbox": [ + 313, + 349, + 556, + 407 + ], + "type": "text", + "content": " on CIFAR-10. Adversarial Training (AT) and Adversarial Purification (AP) methods are evaluated. (*The number of iterations for the attack is half that of other methods for less computational overhead. † methods need extra data.)" + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 315, + 413, + 556, + 570 + ], + "lines": [ + { + "bbox": [ + 315, + 413, + 556, + 570 + ], + "spans": [ + { + "bbox": [ + 315, + 413, + 556, + 570 + ], + "type": "table", + "html": "
TypeMethodStandard Acc.Robust Acc.
PGDAutoAttack
WideResNet-70-16
ATRebuffi et al. [34]†95.7489.6282.32
Gowal et al. [10]†94.7488.1880.53
Rebuffi et al. [34]92.4186.2480.42
APYoon et al. [50]86.76±1.1575.666±1.2980.43±0.42
Nie et al. [30]92.15±0.7282.97±1.3883.06±1.27
Lee et al [20]90.53±0.1483.75±0.9985.59±0.61
Bai et al. [3]92.9784.37*83.01
Lin et al. [24]91.99-81.35
Ours93.36±0.5987.7±0.9791.71±1.09
", + "image_path": "7ff2727de888403bb7f15b41151ca2bfa7875b3c2e00d0ac0d2c78a0c9d49f88.jpg" + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "table_body" + } + ], + "index": 8 + }, + { + "bbox": [ + 313, + 594, + 557, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 594, + 557, + 715 + ], + "spans": [ + { + "bbox": [ + 313, + 594, + 557, + 715 + ], + "type": "text", + "content": "significant time overhead of [3], which only iterates 100 times under PGD—half of what others do—our method outperforms more under PGD attacks. From Table 4, we see that when WideResNet-70-16 is the backbone, our method outperforms other baselines, leading by " + }, + { + "bbox": [ + 313, + 594, + 557, + 715 + ], + "type": "inline_equation", + "content": "1.21\\%" + }, + { + "bbox": [ + 313, + 594, + 557, + 715 + ], + "type": "text", + "content": " in standard accuracy and by " + }, + { + "bbox": [ + 313, + 594, + 557, + 715 + ], + "type": "inline_equation", + "content": "3.33\\%" + }, + { + "bbox": [ + 313, + 594, + 557, + 715 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 313, + 594, + 557, + 715 + ], + "type": "inline_equation", + "content": "6.12\\%" + }, + { + "bbox": [ + 313, + 594, + 557, + 715 + ], + "type": "text", + "content": " in robust accuracy under PGD and AutoAttack, respectively. Additionally, we applied the BPDA+EOT attack, which approximates differentiability. From Table 5, we observe that although robust accuracy lags behind [40] by " + }, + { + "bbox": [ + 313, + 594, + 557, + 715 + ], + "type": "inline_equation", + "content": "1.71\\%" + }, + { + "bbox": [ + 313, + 594, + 557, + 715 + ], + "type": "text", + "content": ", our method still leads" + } + ] + } + ], + "index": 9 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "type": "text", + "content": "29273" + } + ] + } + ], + "index": 10 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 5 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 61, + 102, + 291, + 268 + ], + "blocks": [ + { + "bbox": [ + 55, + 70, + 295, + 93 + ], + "lines": [ + { + "bbox": [ + 55, + 70, + 295, + 93 + ], + "spans": [ + { + "bbox": [ + 55, + 70, + 295, + 93 + ], + "type": "text", + "content": "Table 5. Standard and robust accuracy against BPDA+EOT " + }, + { + "bbox": [ + 55, + 70, + 295, + 93 + ], + "type": "inline_equation", + "content": "\\ell_{\\infty}(\\epsilon = 8 / 255)" + }, + { + "bbox": [ + 55, + 70, + 295, + 93 + ], + "type": "text", + "content": " on CIFAR-10." + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 61, + 102, + 291, + 268 + ], + "lines": [ + { + "bbox": [ + 61, + 102, + 291, + 268 + ], + "spans": [ + { + "bbox": [ + 61, + 102, + 291, + 268 + ], + "type": "table", + "html": "
MethodPurificationAccuracy
StandardRobust
WideResNet-28-10
Song et al. [41]Gibbs Update95.009.00
Yang et al. [48]Mask+Recon94.0015.00
Hill et al. [15]EBM+LD84.1254.90
Yoon et al. [50]DSM+LD85.66±0.5166.91±1.75
Nie et al. [30]Diffusion90.07±0.9781.45±1.51
Bai et al. [3]Diffusion91.37±1.2185.51±0.81
Wang et al. [44]Diffusion89.96±0.4075.59±1.26
Song et al. [40]Diffusion89.88±0.3588.43±0.83
Lee et al. [20]Diffusion90.16±0.6488.40±0.88
OursDiffusion93.04±0.6586.72±0.20
", + "image_path": "d46d73af3e6bfae438d708ac7dacee533bf14adaa46c76d287693d552415d00d.jpg" + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "table_body" + } + ], + "index": 1 + }, + { + "bbox": [ + 54, + 289, + 296, + 434 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 54, + 289, + 296, + 434 + ], + "spans": [ + { + "bbox": [ + 54, + 289, + 296, + 434 + ], + "type": "text", + "content": "in standard accuracy by " + }, + { + "bbox": [ + 54, + 289, + 296, + 434 + ], + "type": "inline_equation", + "content": "3.16\\%" + }, + { + "bbox": [ + 54, + 289, + 296, + 434 + ], + "type": "text", + "content": ", resulting in an average advantage of " + }, + { + "bbox": [ + 54, + 289, + 296, + 434 + ], + "type": "inline_equation", + "content": "0.73\\%" + }, + { + "bbox": [ + 54, + 289, + 296, + 434 + ], + "type": "text", + "content": ". We also measured the distance between purified samples and the original samples in feature space, shown in the Appendix. A smaller distance indicates that the purified samples retain more consistent semantic information with the original samples (further discussion in the appendix). From Figure 6, we find that the distribution of purified samples by our method is most similar to the original samples compared to other methods. Overall, our method outperforms others, demonstrating its effectiveness in preserving semantic information while eliminating adversarial perturbations." + } + ] + } + ], + "index": 2 + }, + { + "type": "table", + "bbox": [ + 64, + 475, + 288, + 654 + ], + "blocks": [ + { + "bbox": [ + 55, + 444, + 295, + 468 + ], + "lines": [ + { + "bbox": [ + 55, + 444, + 295, + 468 + ], + "spans": [ + { + "bbox": [ + 55, + 444, + 295, + 468 + ], + "type": "text", + "content": "Table 6. Standard and robust accuracy against PGD+EOT " + }, + { + "bbox": [ + 55, + 444, + 295, + 468 + ], + "type": "inline_equation", + "content": "\\ell_{\\infty}(\\epsilon = 8 / 255)" + }, + { + "bbox": [ + 55, + 444, + 295, + 468 + ], + "type": "text", + "content": " on ImageNet. ResNet-50 is used as the classifier." + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 64, + 475, + 288, + 654 + ], + "lines": [ + { + "bbox": [ + 64, + 475, + 288, + 654 + ], + "spans": [ + { + "bbox": [ + 64, + 475, + 288, + 654 + ], + "type": "table", + "html": "
TypeMethodAccuracy
StandardRobust
ResNet-50
ATSalman et al. [35]63.8639.11
Engstrom et al. [7]62.4233.20
Wong et al. [46]53.8328.04
APNie et al. (t* = 0.2) [30]73.9640.63
Nie et al. (t* = 0.3) [30]72.8548.24
Nie et al. (t* = 0.4) [30]61.7141.67
Bai et al. [3]70.4141.7
Lee et al. [20]71.4246.59
Song et al. [40]62.2551.14
Geigh et al. [53]73.9856.54
Ours74.51±1.0760.06±1.66
", + "image_path": "6978497420bad30b7d90c7ec02cc27ed80081fde3196e79950031bdee1bf6dbf.jpg" + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "table_body" + } + ], + "index": 4 + }, + { + "bbox": [ + 55, + 665, + 296, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 665, + 296, + 715 + ], + "spans": [ + { + "bbox": [ + 55, + 665, + 296, + 715 + ], + "type": "text", + "content": "ImageNet. We test ResNet50 as the backbone on the ImageNet dataset under PGD+EOT attacks, which is the same as [3, 20, 30]. From Table 6, our method leads competitors by " + }, + { + "bbox": [ + 55, + 665, + 296, + 715 + ], + "type": "inline_equation", + "content": "0.53\\%" + }, + { + "bbox": [ + 55, + 665, + 296, + 715 + ], + "type": "text", + "content": " in standard accuracy and by " + }, + { + "bbox": [ + 55, + 665, + 296, + 715 + ], + "type": "inline_equation", + "content": "3.52\\%" + }, + { + "bbox": [ + 55, + 665, + 296, + 715 + ], + "type": "text", + "content": " in" + } + ] + } + ], + "index": 5 + }, + { + "type": "image", + "bbox": [ + 317, + 71, + 432, + 200 + ], + "blocks": [ + { + "bbox": [ + 317, + 71, + 432, + 200 + ], + "lines": [ + { + "bbox": [ + 317, + 71, + 432, + 200 + ], + "spans": [ + { + "bbox": [ + 317, + 71, + 432, + 200 + ], + "type": "image", + "image_path": "437faba82c23906b56fbff745704b27c211fe5b0cbdad2f8c2545d6bfff24be5.jpg" + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 313, + 208, + 555, + 285 + ], + "lines": [ + { + "bbox": [ + 313, + 208, + 555, + 285 + ], + "spans": [ + { + "bbox": [ + 313, + 208, + 555, + 285 + ], + "type": "text", + "content": "Figure 4. 2D purification trajectories of Our method and DiffPure. For DiffPure (left), if " + }, + { + "bbox": [ + 313, + 208, + 555, + 285 + ], + "type": "inline_equation", + "content": "t^{*} = 0.2" + }, + { + "bbox": [ + 313, + 208, + 555, + 285 + ], + "type": "text", + "content": ", although the purification direction is toward the original class, the adversarial perturbation cannot be completely eliminated; If " + }, + { + "bbox": [ + 313, + 208, + 555, + 285 + ], + "type": "inline_equation", + "content": "t^{*} = 0.4" + }, + { + "bbox": [ + 313, + 208, + 555, + 285 + ], + "type": "text", + "content": ", the semantic information changes and the purification direction is no longer toward the original class. Our method (right) can eliminate adversarial perturbation and remain the semantic information." + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "image_caption" + } + ], + "index": 6 + }, + { + "type": "image", + "bbox": [ + 443, + 71, + 561, + 200 + ], + "blocks": [ + { + "bbox": [ + 443, + 71, + 561, + 200 + ], + "lines": [ + { + "bbox": [ + 443, + 71, + 561, + 200 + ], + "spans": [ + { + "bbox": [ + 443, + 71, + 561, + 200 + ], + "type": "image", + "image_path": "7c0d260e32729411e6704017f5b77a62d94ecdcf7eb856cf11bd72c4555738d4.jpg" + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "image_body" + } + ], + "index": 7 + }, + { + "type": "image", + "bbox": [ + 317, + 295, + 564, + 446 + ], + "blocks": [ + { + "bbox": [ + 317, + 295, + 564, + 446 + ], + "lines": [ + { + "bbox": [ + 317, + 295, + 564, + 446 + ], + "spans": [ + { + "bbox": [ + 317, + 295, + 564, + 446 + ], + "type": "image", + "image_path": "bd8b81cd98e629f949190121e32820383bfa88d119d6fbded0f61a52cc952f2d.jpg" + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 313, + 454, + 555, + 477 + ], + "lines": [ + { + "bbox": [ + 313, + 454, + 555, + 477 + ], + "spans": [ + { + "bbox": [ + 313, + 454, + 555, + 477 + ], + "type": "text", + "content": "Figure 5. Visualization of attention map, attention mask and images purified by our method." + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "image_caption" + } + ], + "index": 9 + }, + { + "bbox": [ + 313, + 498, + 556, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 498, + 556, + 715 + ], + "spans": [ + { + "bbox": [ + 313, + 498, + 556, + 715 + ], + "type": "text", + "content": "robust accuracy, significantly surpassing the baseline. Additionally, we observe that the method from [30] achieves a standard accuracy of " + }, + { + "bbox": [ + 313, + 498, + 556, + 715 + ], + "type": "inline_equation", + "content": "73.96\\%" + }, + { + "bbox": [ + 313, + 498, + 556, + 715 + ], + "type": "text", + "content": " and a robust accuracy of " + }, + { + "bbox": [ + 313, + 498, + 556, + 715 + ], + "type": "inline_equation", + "content": "40.63\\%" + }, + { + "bbox": [ + 313, + 498, + 556, + 715 + ], + "type": "text", + "content": " when " + }, + { + "bbox": [ + 313, + 498, + 556, + 715 + ], + "type": "inline_equation", + "content": "t^{*} = 0.2" + }, + { + "bbox": [ + 313, + 498, + 556, + 715 + ], + "type": "text", + "content": ". When " + }, + { + "bbox": [ + 313, + 498, + 556, + 715 + ], + "type": "inline_equation", + "content": "t^{*} = 0.3" + }, + { + "bbox": [ + 313, + 498, + 556, + 715 + ], + "type": "text", + "content": ", standard accuracy drops to " + }, + { + "bbox": [ + 313, + 498, + 556, + 715 + ], + "type": "inline_equation", + "content": "72.85\\%" + }, + { + "bbox": [ + 313, + 498, + 556, + 715 + ], + "type": "text", + "content": " while robust accuracy rises to " + }, + { + "bbox": [ + 313, + 498, + 556, + 715 + ], + "type": "inline_equation", + "content": "48.24\\%" + }, + { + "bbox": [ + 313, + 498, + 556, + 715 + ], + "type": "text", + "content": ". However, at " + }, + { + "bbox": [ + 313, + 498, + 556, + 715 + ], + "type": "inline_equation", + "content": "t^{*} = 0.4" + }, + { + "bbox": [ + 313, + 498, + 556, + 715 + ], + "type": "text", + "content": ", both standard accuracy and robust accuracy decline to " + }, + { + "bbox": [ + 313, + 498, + 556, + 715 + ], + "type": "inline_equation", + "content": "61.71\\%" + }, + { + "bbox": [ + 313, + 498, + 556, + 715 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 313, + 498, + 556, + 715 + ], + "type": "inline_equation", + "content": "41.67\\%" + }, + { + "bbox": [ + 313, + 498, + 556, + 715 + ], + "type": "text", + "content": ", respectively. Additionally, We visualize the 2D purification trajectories of our method and DiffPure using the T-SNE [43] technique. From Figure 4, we observe that when " + }, + { + "bbox": [ + 313, + 498, + 556, + 715 + ], + "type": "inline_equation", + "content": "t^{*} = 0.2" + }, + { + "bbox": [ + 313, + 498, + 556, + 715 + ], + "type": "text", + "content": ", the adversarial perturbation on the adversarial sample has not been fully purified, and it remains at the boundary between the two classes of samples. When " + }, + { + "bbox": [ + 313, + 498, + 556, + 715 + ], + "type": "inline_equation", + "content": "t^{*} = 0.4" + }, + { + "bbox": [ + 313, + 498, + 556, + 715 + ], + "type": "text", + "content": ", the semantic information of the adversarial example begins to change, and as a result, the purification trajectory no longer points toward the original classes. This indicates that applying the same noise intensity across an image cannot effectively balance semantic preservation and the elimination of adversarial perturbation." + } + ] + } + ], + "index": 11 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "type": "text", + "content": "29274" + } + ] + } + ], + "index": 12 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 6 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 56, + 70, + 175, + 194 + ], + "blocks": [ + { + "bbox": [ + 56, + 70, + 175, + 194 + ], + "lines": [ + { + "bbox": [ + 56, + 70, + 175, + 194 + ], + "spans": [ + { + "bbox": [ + 56, + 70, + 175, + 194 + ], + "type": "image", + "image_path": "f2fe2f7da156c147aede35b69270193397e65fa7941be1fc6e3c72d973e71519.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + } + ], + "index": 0 + }, + { + "type": "image", + "bbox": [ + 187, + 70, + 304, + 194 + ], + "blocks": [ + { + "bbox": [ + 187, + 70, + 304, + 194 + ], + "lines": [ + { + "bbox": [ + 187, + 70, + 304, + 194 + ], + "spans": [ + { + "bbox": [ + 187, + 70, + 304, + 194 + ], + "type": "image", + "image_path": "2f0d78aadcdf83081c5d03c6f18176e424690d48eb30011f598427be3bf7f023.jpg" + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_body" + } + ], + "index": 1 + }, + { + "type": "image", + "bbox": [ + 53, + 202, + 175, + 329 + ], + "blocks": [ + { + "bbox": [ + 53, + 202, + 175, + 329 + ], + "lines": [ + { + "bbox": [ + 53, + 202, + 175, + 329 + ], + "spans": [ + { + "bbox": [ + 53, + 202, + 175, + 329 + ], + "type": "image", + "image_path": "465d02c0236b8ba9d053548104da1a56da129f811c9f47a5da7cc242eda6bdee.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 55, + 339, + 296, + 373 + ], + "lines": [ + { + "bbox": [ + 55, + 339, + 296, + 373 + ], + "spans": [ + { + "bbox": [ + 55, + 339, + 296, + 373 + ], + "type": "text", + "content": "Figure 6. Joint distribution of the original samples and purified samples. The distributions of the purified samples by our method and the original samples are the most similar." + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_caption" + } + ], + "index": 2 + }, + { + "type": "image", + "bbox": [ + 186, + 204, + 304, + 328 + ], + "blocks": [ + { + "bbox": [ + 186, + 204, + 304, + 328 + ], + "lines": [ + { + "bbox": [ + 186, + 204, + 304, + 328 + ], + "spans": [ + { + "bbox": [ + 186, + 204, + 304, + 328 + ], + "type": "image", + "image_path": "47ef5752cd0e60b6731935e6c95310e974fcf95da032e0fedd57118f5da4b3a9.jpg" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_body" + } + ], + "index": 3 + }, + { + "bbox": [ + 55, + 396, + 296, + 517 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 396, + 296, + 517 + ], + "spans": [ + { + "bbox": [ + 55, + 396, + 296, + 517 + ], + "type": "text", + "content": "tions. In contrast, our method adds varying noise intensities to different regions of an image, applying stronger noise to areas with higher classification weights to eliminate adversarial perturbations, while applying lighter noise to other regions to preserve semantic information. This approach allows our method to achieve good results in both robust accuracy and standard accuracy simultaneously. Our method achieves comparable standard accuracy to [30] at " + }, + { + "bbox": [ + 55, + 396, + 296, + 517 + ], + "type": "inline_equation", + "content": "t^* = 0.2" + }, + { + "bbox": [ + 55, + 396, + 296, + 517 + ], + "type": "text", + "content": " but significantly outperforms it in robust accuracy, demonstrating the effectiveness of our approach." + } + ] + } + ], + "index": 5 + }, + { + "type": "image", + "bbox": [ + 58, + 531, + 179, + 647 + ], + "blocks": [ + { + "bbox": [ + 58, + 531, + 179, + 647 + ], + "lines": [ + { + "bbox": [ + 58, + 531, + 179, + 647 + ], + "spans": [ + { + "bbox": [ + 58, + 531, + 179, + 647 + ], + "type": "image", + "image_path": "3e786b46938306ad70c75c0e8cf8c8e2c26fc791e4439b071a27362b44e47c4a.jpg" + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 55, + 663, + 297, + 696 + ], + "lines": [ + { + "bbox": [ + 55, + 663, + 297, + 696 + ], + "spans": [ + { + "bbox": [ + 55, + 663, + 297, + 696 + ], + "type": "text", + "content": "Figure 7. Robust and Standard Accuracy under different thresholds " + }, + { + "bbox": [ + 55, + 663, + 297, + 696 + ], + "type": "inline_equation", + "content": "\\tau" + }, + { + "bbox": [ + 55, + 663, + 297, + 696 + ], + "type": "text", + "content": " (left) and under different combinations of timesteps " + }, + { + "bbox": [ + 55, + 663, + 297, + 696 + ], + "type": "inline_equation", + "content": "t_l" + }, + { + "bbox": [ + 55, + 663, + 297, + 696 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 55, + 663, + 297, + 696 + ], + "type": "inline_equation", + "content": "t_s" + }, + { + "bbox": [ + 55, + 663, + 297, + 696 + ], + "type": "text", + "content": " (right)." + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "image_caption" + } + ], + "index": 6 + }, + { + "type": "image", + "bbox": [ + 186, + 531, + 304, + 647 + ], + "blocks": [ + { + "bbox": [ + 186, + 531, + 304, + 647 + ], + "lines": [ + { + "bbox": [ + 186, + 531, + 304, + 647 + ], + "spans": [ + { + "bbox": [ + 186, + 531, + 304, + 647 + ], + "type": "image", + "image_path": "1561298de81ab5e7748e4dc5623d426b61de577b70cac408191e9b045e94b936.jpg" + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "image_body" + } + ], + "index": 7 + }, + { + "bbox": [ + 314, + 72, + 408, + 85 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 314, + 72, + 408, + 85 + ], + "spans": [ + { + "bbox": [ + 314, + 72, + 408, + 85 + ], + "type": "text", + "content": "5.3. Ablation Study" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 313, + 90, + 555, + 329 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 90, + 555, + 329 + ], + "spans": [ + { + "bbox": [ + 313, + 90, + 555, + 329 + ], + "type": "text", + "content": "We first conducted a hyperparameter sensitivity analysis. Our method involves three hyperparameters, two of which are noises of different intensities, " + }, + { + "bbox": [ + 313, + 90, + 555, + 329 + ], + "type": "inline_equation", + "content": "t_{l}" + }, + { + "bbox": [ + 313, + 90, + 555, + 329 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 313, + 90, + 555, + 329 + ], + "type": "inline_equation", + "content": "t_{s}" + }, + { + "bbox": [ + 313, + 90, + 555, + 329 + ], + "type": "text", + "content": ", and the third is a threshold, " + }, + { + "bbox": [ + 313, + 90, + 555, + 329 + ], + "type": "inline_equation", + "content": "\\tau" + }, + { + "bbox": [ + 313, + 90, + 555, + 329 + ], + "type": "text", + "content": ", used to control the size of the attention mask. From Figure 7, we observe that as " + }, + { + "bbox": [ + 313, + 90, + 555, + 329 + ], + "type": "inline_equation", + "content": "\\tau" + }, + { + "bbox": [ + 313, + 90, + 555, + 329 + ], + "type": "text", + "content": " increases (resulting in smaller mask areas), the Standard Accuracy also consistently increases. However, adversarial perturbations are preserved at this time, leading to poorer Robust Accuracy when " + }, + { + "bbox": [ + 313, + 90, + 555, + 329 + ], + "type": "inline_equation", + "content": "\\tau" + }, + { + "bbox": [ + 313, + 90, + 555, + 329 + ], + "type": "text", + "content": " is large. When " + }, + { + "bbox": [ + 313, + 90, + 555, + 329 + ], + "type": "inline_equation", + "content": "\\tau" + }, + { + "bbox": [ + 313, + 90, + 555, + 329 + ], + "type": "text", + "content": " is small, semantic information is damaged, which also results in weaker Robust Accuracy. Therefore, we choose " + }, + { + "bbox": [ + 313, + 90, + 555, + 329 + ], + "type": "inline_equation", + "content": "\\tau" + }, + { + "bbox": [ + 313, + 90, + 555, + 329 + ], + "type": "text", + "content": " to be around 0.8. Simultaneously, we can see that when " + }, + { + "bbox": [ + 313, + 90, + 555, + 329 + ], + "type": "inline_equation", + "content": "t_{l} = 0.2" + }, + { + "bbox": [ + 313, + 90, + 555, + 329 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 313, + 90, + 555, + 329 + ], + "type": "inline_equation", + "content": "t_{s} = 0.05" + }, + { + "bbox": [ + 313, + 90, + 555, + 329 + ], + "type": "text", + "content": ", the Robust Accuracy is the highest. Here, " + }, + { + "bbox": [ + 313, + 90, + 555, + 329 + ], + "type": "inline_equation", + "content": "t_{s}" + }, + { + "bbox": [ + 313, + 90, + 555, + 329 + ], + "type": "text", + "content": " is less than the " + }, + { + "bbox": [ + 313, + 90, + 555, + 329 + ], + "type": "inline_equation", + "content": "t^{*}" + }, + { + "bbox": [ + 313, + 90, + 555, + 329 + ], + "type": "text", + "content": " used in [30] to preserve more semantic information, while " + }, + { + "bbox": [ + 313, + 90, + 555, + 329 + ], + "type": "inline_equation", + "content": "t_{l}" + }, + { + "bbox": [ + 313, + 90, + 555, + 329 + ], + "type": "text", + "content": " is greater than " + }, + { + "bbox": [ + 313, + 90, + 555, + 329 + ], + "type": "inline_equation", + "content": "t^{*}" + }, + { + "bbox": [ + 313, + 90, + 555, + 329 + ], + "type": "text", + "content": " to eliminate adversarial perturbations. From Table 6, we can see that when we set " + }, + { + "bbox": [ + 313, + 90, + 555, + 329 + ], + "type": "inline_equation", + "content": "t_{l} = 0.4" + }, + { + "bbox": [ + 313, + 90, + 555, + 329 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 313, + 90, + 555, + 329 + ], + "type": "inline_equation", + "content": "t_{s} = 0.2" + }, + { + "bbox": [ + 313, + 90, + 555, + 329 + ], + "type": "text", + "content": ", The Standard Accuracy is almost as high as the " + }, + { + "bbox": [ + 313, + 90, + 555, + 329 + ], + "type": "inline_equation", + "content": "t^{*} = 0.2" + }, + { + "bbox": [ + 313, + 90, + 555, + 329 + ], + "type": "text", + "content": " of [30], but at the same time, our Robust Accuracy significantly leads. This demonstrates the effectiveness of our method. We also explore whether it is necessary for each block" + } + ] + } + ], + "index": 10 + }, + { + "type": "table", + "bbox": [ + 335, + 372, + 535, + 464 + ], + "blocks": [ + { + "bbox": [ + 313, + 341, + 555, + 363 + ], + "lines": [ + { + "bbox": [ + 313, + 341, + 555, + 363 + ], + "spans": [ + { + "bbox": [ + 313, + 341, + 555, + 363 + ], + "type": "text", + "content": "Table 7. Standard Accuracy and Robust Accuracy for different block combinations. WideResNet28-10 servers as the backbone." + } + ] + } + ], + "index": 11, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 335, + 372, + 535, + 464 + ], + "lines": [ + { + "bbox": [ + 335, + 372, + 535, + 464 + ], + "spans": [ + { + "bbox": [ + 335, + 372, + 535, + 464 + ], + "type": "table", + "html": "
Encoder BlockAccuracy
E1E2E3E4StandardRobustAverage
X✓✓✓92.7758.0075.38
✓X✓✓93.1658.0075.58
✓✓X✓93.3557.6175.48
✓✓✓X91.6658.8575.25
✓✓✓✓93.1659.3876.27
", + "image_path": "ea0028e09fa3351c2258308c0f1f4cf852e2601c6bbca1cf8a7941edbc8202a3.jpg" + } + ] + } + ], + "index": 12, + "angle": 0, + "type": "table_body" + } + ], + "index": 12 + }, + { + "bbox": [ + 313, + 476, + 555, + 548 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 476, + 555, + 548 + ], + "spans": [ + { + "bbox": [ + 313, + 476, + 555, + 548 + ], + "type": "text", + "content": "to participate in constructing the attention mask. Taking WideResNet28-10 as an example, it can be divided into four blocks. We extract the output of the final ReLU function of each block as the activation map. From Table 7, we can see that removing any block results in a performance loss. Therefore, every block should be utilized." + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 313, + 561, + 388, + 573 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 561, + 388, + 573 + ], + "spans": [ + { + "bbox": [ + 313, + 561, + 388, + 573 + ], + "type": "text", + "content": "6. Conclusion" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 313, + 582, + 555, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 582, + 555, + 713 + ], + "spans": [ + { + "bbox": [ + 313, + 582, + 555, + 713 + ], + "type": "text", + "content": "We propose a novel forward process that adds noise of varying intensities based on the classifier's attention to different regions, achieving a balance between preserving semantic information and eliminating adversarial perturbations. The denoising process is tailored to our method to remove different-intensity noise within a noisy image. By replacing the multi-step resampling with an improved single-step resampling, time cost and memory usage are reduced significantly. Extensive experiments on three datasets show that our method outperforms other methods by a promising improvement against various adversarial attacks." + } + ] + } + ], + "index": 15 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "text", + "content": "29275" + } + ] + } + ], + "index": 16 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 7 + }, + { + "para_blocks": [ + { + "bbox": [ + 56, + 72, + 128, + 85 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 72, + 128, + 85 + ], + "spans": [ + { + "bbox": [ + 56, + 72, + 128, + 85 + ], + "type": "text", + "content": "Acknowledge" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 55, + 90, + 297, + 213 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 90, + 297, + 213 + ], + "spans": [ + { + "bbox": [ + 55, + 90, + 297, + 213 + ], + "type": "text", + "content": "This work was supported in part by the National Key R&D Program of China under Grant 2022ZD0119204, in part by National Natural Science Foundation of China: 62236008, 62441232, U21B2038, U23B2051, 62122075 and 62376257, in part by Youth Innovation Promotion Association CAS, in part by the Strategic Priority Research Program of the Chinese Academy of Sciences, Grant No. XDB0680201, in part by the Fundamental Research Funds for the Central Universities, in part by the Tencent Marketing Solution Rhino-Bird Focused Research Program, in part by Xiaomi Young Talents Program." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 56, + 233, + 115, + 246 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 233, + 115, + 246 + ], + "spans": [ + { + "bbox": [ + 56, + 233, + 115, + 246 + ], + "type": "text", + "content": "References" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 57, + 253, + 296, + 713 + ], + "type": "list", + "angle": 0, + "index": 13, + "blocks": [ + { + "bbox": [ + 61, + 253, + 294, + 297 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 61, + 253, + 294, + 297 + ], + "spans": [ + { + "bbox": [ + 61, + 253, + 294, + 297 + ], + "type": "text", + "content": "[1] Anish Athalye, Nicholas Carlini, and David Wagner. Obfuscated gradients give a false sense of security: Circumventing defenses to adversarial examples. In International conference on machine learning, pages 274-283. PMLR, 2018. 5" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 61, + 300, + 296, + 343 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 61, + 300, + 296, + 343 + ], + "spans": [ + { + "bbox": [ + 61, + 300, + 296, + 343 + ], + "type": "text", + "content": "[2] Maximilian Augustin, Alexander Meinke, and Matthias Hein. Adversarial robustness on in-and out-distribution improves explainability. In European Conference on Computer Vision, pages 228-245. Springer, 2020. 6" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 61, + 344, + 296, + 398 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 61, + 344, + 296, + 398 + ], + "spans": [ + { + "bbox": [ + 61, + 344, + 296, + 398 + ], + "type": "text", + "content": "[3] Mingyuan Bai, Wei Huang, Tenghui Li, Andong Wang, Junbin Gao, Cesar F Caiafa, and Qibin Zhao. Diffusion models demand contrastive guidance for adversarial purification to advance. In *Forty-first International Conference on Machine Learning*, 2024. 2, 3, 5, 6, 7" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 62, + 399, + 296, + 443 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 399, + 296, + 443 + ], + "spans": [ + { + "bbox": [ + 62, + 399, + 296, + 443 + ], + "type": "text", + "content": "[4] Francesco Croce and Matthias Hein. Reliable evaluation of adversarial robustness with an ensemble of diverse parameter-free attacks. In International conference on machine learning, pages 2206-2216, 2020. 5" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 62, + 445, + 296, + 499 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 445, + 296, + 499 + ], + "spans": [ + { + "bbox": [ + 62, + 445, + 296, + 499 + ], + "type": "text", + "content": "[5] Francesco Croce, Maksym Andriushchenko, Vikash Sehwag, Edoardo Debenedetti, Nicolas Flammarion, Mung Chiang, Prateek Mittal, and Matthias Hein. Robustbench: a standardized adversarial robustness benchmark. arXiv preprint arXiv:2010.09670, 2020. 5" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 62, + 501, + 296, + 544 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 501, + 296, + 544 + ], + "spans": [ + { + "bbox": [ + 62, + 501, + 296, + 544 + ], + "type": "text", + "content": "[6] Jia Deng, Wei Dong, Richard Socher, Li-Jia Li, Kai Li, and Li Fei-Fei. Imagenet: A large-scale hierarchical image database. In 2009 IEEE conference on computer vision and pattern recognition, pages 248–255, 2009. 5" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 62, + 545, + 296, + 588 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 545, + 296, + 588 + ], + "spans": [ + { + "bbox": [ + 62, + 545, + 296, + 588 + ], + "type": "text", + "content": "[7] Logan Engstrom, Andrew Ilyas, Hadi Salman, Shibani Santurkar, and Dimitris Tsipras. Robustness (python library), 2019. URL https://github.com/MadryLab/robustness, 4(4): 4-3, 2019. 7" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 62, + 590, + 296, + 634 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 590, + 296, + 634 + ], + "spans": [ + { + "bbox": [ + 62, + 590, + 296, + 634 + ], + "type": "text", + "content": "[8] Xavier Glorot, Antoine Bordes, and Yoshua Bengio. Deep sparse rectifier neural networks. In Proceedings of the fourteenth international conference on artificial intelligence and statistics, pages 315-323, 2011. 3" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 62, + 635, + 296, + 669 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 635, + 296, + 669 + ], + "spans": [ + { + "bbox": [ + 62, + 635, + 296, + 669 + ], + "type": "text", + "content": "[9] Ian J Goodfellow, Jonathon Shlens, and Christian Szegedy. Explaining and harnessing adversarial examples. arXiv preprint arXiv:1412.6572, 2014. 1" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 57, + 670, + 296, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 670, + 296, + 713 + ], + "spans": [ + { + "bbox": [ + 57, + 670, + 296, + 713 + ], + "type": "text", + "content": "[10] Sven Gowal, Chongli Qin, Jonathan Uesato, Timothy Mann, and Pushmeet Kohli. Uncovering the limits of adversarial training against norm-bounded adversarial examples. arXiv preprint arXiv:2010.03593, 2020. 6" + } + ] + } + ], + "index": 12 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 316, + 73, + 555, + 713 + ], + "type": "list", + "angle": 0, + "index": 28, + "blocks": [ + { + "bbox": [ + 316, + 73, + 555, + 116 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 73, + 555, + 116 + ], + "spans": [ + { + "bbox": [ + 316, + 73, + 555, + 116 + ], + "type": "text", + "content": "[11] Sven Gowal, Chongli Qin, Jonathan Uesato, Timothy Mann, and Pushmeet Kohli. Uncovering the limits of adversarial training against norm-bounded adversarial examples. arXiv preprint arXiv:2010.03593, 2020. 6" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 317, + 118, + 555, + 162 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 118, + 555, + 162 + ], + "spans": [ + { + "bbox": [ + 317, + 118, + 555, + 162 + ], + "type": "text", + "content": "[12] Sven Gowal, Sylvestre-Alvise Rebuffi, Olivia Wiles, Florian Stimberg, Dan Andrei Calian, and Timothy A Mann. Improving robustness using generated data. Advances in Neural Information Processing Systems, 34:4218-4233, 2021. 6" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 317, + 163, + 555, + 206 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 163, + 555, + 206 + ], + "spans": [ + { + "bbox": [ + 317, + 163, + 555, + 206 + ], + "type": "text", + "content": "[13] Sven Gowal, Sylvestre-Alvise Rebuffi, Olivia Wiles, Florian Stimberg, Dan Andrei Calian, and Timothy A Mann. Improving robustness using generated data. Advances in Neural Information Processing Systems, 34:4218-4233, 2021. 6" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 317, + 209, + 554, + 252 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 209, + 554, + 252 + ], + "spans": [ + { + "bbox": [ + 317, + 209, + 554, + 252 + ], + "type": "text", + "content": "[14] Kaiming He, Xiangyu Zhang, Shaoqing Ren, and Jian Sun. Deep residual learning for image recognition. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 770-778, 2016. 3, 5" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 317, + 254, + 554, + 297 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 254, + 554, + 297 + ], + "spans": [ + { + "bbox": [ + 317, + 254, + 554, + 297 + ], + "type": "text", + "content": "[15] Mitch Hill, Jonathan Craig Mitchell, and Song-Chun Zhu. Stochastic security: Adversarial defense using long-run dynamics of energy-based models. In International Conference on Learning Representations, 2021. 5, 7" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 316, + 299, + 554, + 331 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 299, + 554, + 331 + ], + "spans": [ + { + "bbox": [ + 316, + 299, + 554, + 331 + ], + "type": "text", + "content": "[16] Jonathan Ho, Ajay Jain, and Pieter Abbeel. Denoising diffusion probabilistic models. Advances in neural information processing systems, 33:6840-6851, 2020. 1, 3" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 316, + 333, + 554, + 365 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 333, + 554, + 365 + ], + "spans": [ + { + "bbox": [ + 316, + 333, + 554, + 365 + ], + "type": "text", + "content": "[17] Duhun Hwang, Eunjung Lee, and Wonjong Rhee. Aidpurifier: A light auxiliary network for boosting adversarial defense. Neurocomputing, 541:126251, 2023. 1" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 316, + 367, + 554, + 410 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 367, + 554, + 410 + ], + "spans": [ + { + "bbox": [ + 316, + 367, + 554, + 410 + ], + "type": "text", + "content": "[18] Xiaojun Jia, Yong Zhang, Xingxing Wei, Baoyuan Wu, Ke Ma, Jue Wang, and Xiaochun Cao. Improving fast adversarial training with prior-guided knowledge. IEEE Transactions on Pattern Analysis and Machine Intelligence, 2024. 1" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 316, + 412, + 554, + 454 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 412, + 554, + 454 + ], + "spans": [ + { + "bbox": [ + 316, + 412, + 554, + 454 + ], + "type": "text", + "content": "[19] Tero Karras, Miika Aittala, Timo Aila, and Samuli Laine. Elucidating the design space of diffusion-based generative models. Advances in neural information processing systems, 35:26565-26577, 2022. 5" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 316, + 456, + 555, + 500 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 456, + 555, + 500 + ], + "spans": [ + { + "bbox": [ + 316, + 456, + 555, + 500 + ], + "type": "text", + "content": "[20] Minjong Lee and Dongwoo Kim. Robust evaluation of diffusion-based adversarial purification. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 134-144, 2023. 5, 6, 7" + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 316, + 502, + 554, + 556 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 502, + 554, + 556 + ], + "spans": [ + { + "bbox": [ + 316, + 502, + 554, + 556 + ], + "type": "text", + "content": "[21] Siyuan Liang, Xingxing Wei, Siyuan Yao, and Xiaochun Cao. Efficient adversarial attacks for visual object tracking. In Computer Vision-ECCV 2020: 16th European Conference, Glasgow, UK, August 23-28, 2020, Proceedings, Part XXVI 16, pages 34-50. Springer, 2020. 1" + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 316, + 558, + 554, + 612 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 558, + 554, + 612 + ], + "spans": [ + { + "bbox": [ + 316, + 558, + 554, + 612 + ], + "type": "text", + "content": "[22] Siyuan Liang, Longkang Li, Yanbo Fan, Xiaojun Jia, Jingzhi Li, Baoyuan Wu, and Xiaochun Cao. A large-scale multiple-objective method for black-box attack against object detection. In European Conference on Computer Vision, pages 619-636. Springer, 2022." + } + ] + } + ], + "index": 25 + }, + { + "bbox": [ + 316, + 614, + 554, + 657 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 614, + 554, + 657 + ], + "spans": [ + { + "bbox": [ + 316, + 614, + 554, + 657 + ], + "type": "text", + "content": "[23] Siyuan Liang, Baoyuan Wu, Yanbo Fan, Xingxing Wei, and Xiaochun Cao. Parallel rectangle flip attack: A query-based black-box attack against object detection. arXiv preprint arXiv:2201.08970, 2022. 1" + } + ] + } + ], + "index": 26 + }, + { + "bbox": [ + 316, + 658, + 554, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 658, + 554, + 713 + ], + "spans": [ + { + "bbox": [ + 316, + 658, + 554, + 713 + ], + "type": "text", + "content": "[24] Guang Lin, Chao Li, Jianhai Zhang, Toshihisa Tanaka, and Qibin Zhao. Adversarial training on purification (ATop): Advancing both robustness and generalization. In The Twelfth International Conference on Learning Representations, 2024. 2, 6" + } + ] + } + ], + "index": 27 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "type": "text", + "content": "29276" + } + ] + } + ], + "index": 29 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 8 + }, + { + "para_blocks": [ + { + "bbox": [ + 56, + 73, + 296, + 712 + ], + "type": "list", + "angle": 0, + "index": 14, + "blocks": [ + { + "bbox": [ + 56, + 73, + 296, + 127 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 73, + 296, + 127 + ], + "spans": [ + { + "bbox": [ + 56, + 73, + 296, + 127 + ], + "type": "text", + "content": "[25] Andreas Lugmayr, Martin Danelljan, Andres Romero, Fisher Yu, Radu Timofte, and Luc Van Gool. Repaint: Inpainting using denoising diffusion probabilistic models. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 11461-11471, 2022. 3, 4" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 56, + 128, + 296, + 172 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 128, + 296, + 172 + ], + "spans": [ + { + "bbox": [ + 56, + 128, + 296, + 172 + ], + "type": "text", + "content": "[26] Aleksander Madry, Aleksandar Makelov, Ludwig Schmidt, Dimitris Tsipras, and Adrian Vladu. Towards deep learning models resistant to adversarial attacks. In International Conference on Learning Representations, 2018. 5" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 56, + 173, + 296, + 216 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 173, + 296, + 216 + ], + "spans": [ + { + "bbox": [ + 56, + 173, + 296, + 216 + ], + "type": "text", + "content": "[27] Dongyu Meng and Hao Chen. Magnet: a two-pronged defense against adversarial examples. In Proceedings of the 2017 ACM SIGSAC conference on computer and communications security, pages 135-147, 2017. 1" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 56, + 217, + 296, + 270 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 217, + 296, + 270 + ], + "spans": [ + { + "bbox": [ + 56, + 217, + 296, + 270 + ], + "type": "text", + "content": "[28] Muzammal Naseer, Salman Khan, Munawar Hayat, Fahad Shahbaz Khan, and Fatih Porikli. A self-supervised approach for adversarial robustness. In 2020 IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), 2020. 2" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 56, + 271, + 296, + 304 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 271, + 296, + 304 + ], + "spans": [ + { + "bbox": [ + 56, + 271, + 296, + 304 + ], + "type": "text", + "content": "[29] Alexander Quinn Nichol and Prafulla Dhariwal. Improved denoising diffusion probabilistic models. In International conference on machine learning, 2021. 1" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 56, + 305, + 296, + 358 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 305, + 296, + 358 + ], + "spans": [ + { + "bbox": [ + 56, + 305, + 296, + 358 + ], + "type": "text", + "content": "[30] Weili Nie, Brandon Guo, Yujia Huang, Chaowei Xiao, Arash Vahdat, and Animashree Anandkumar. Diffusion models for adversarial purification. In Proceedings of the 39th International Conference on Machine Learning, 2022. 1, 2, 3, 5, 6, 7, 8" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 56, + 360, + 296, + 403 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 360, + 296, + 403 + ], + "spans": [ + { + "bbox": [ + 56, + 360, + 296, + 403 + ], + "type": "text", + "content": "[31] Tianyu Pang, Min Lin, Xiao Yang, Jun Zhu, and Shuicheng Yan. Robustness and accuracy could be reconcilable by (proper) definition. In International Conference on Machine Learning, pages 17258-17277. PMLR, 2022. 6" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 56, + 404, + 296, + 448 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 404, + 296, + 448 + ], + "spans": [ + { + "bbox": [ + 56, + 404, + 296, + 448 + ], + "type": "text", + "content": "[32] Gaozheng Pei, Ke Ma, Pinci Yang, Qianqian Xu, Yingfei Sun, et al. Exploring query efficient data generation towards data-free model stealing in hard label setting. arXiv preprint arXiv:2412.15276, 2024. 1" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 56, + 449, + 296, + 480 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 449, + 296, + 480 + ], + "spans": [ + { + "bbox": [ + 56, + 449, + 296, + 480 + ], + "type": "text", + "content": "[33] Prajit Ramachandran, Barret Zoph, and Quoc V Le. Searching for activation functions. arXiv preprint arXiv:1710.05941, 2017. 3" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 56, + 481, + 296, + 525 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 481, + 296, + 525 + ], + "spans": [ + { + "bbox": [ + 56, + 481, + 296, + 525 + ], + "type": "text", + "content": "[34] Sylvestre-Alvise Rebuffi, Sven Gowal, Dan A Calian, Florian Stimberg, Olivia Wiles, and Timothy Mann. Fixing data augmentation to improve adversarial robustness. arXiv preprint arXiv:2103.01946, 2021. 6" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 56, + 526, + 296, + 569 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 526, + 296, + 569 + ], + "spans": [ + { + "bbox": [ + 56, + 526, + 296, + 569 + ], + "type": "text", + "content": "[35] Hadi Salman, Andrew Ilyas, Logan Engstrom, Ashish Kapoor, and Aleksander Madry. Do adversarially robust imagenet models transfer better? Advances in Neural Information Processing Systems, 33:3533-3545, 2020. 7" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 56, + 570, + 296, + 613 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 570, + 296, + 613 + ], + "spans": [ + { + "bbox": [ + 56, + 570, + 296, + 613 + ], + "type": "text", + "content": "[36] Pouya Samangouei, Maya Kabbab, and Rama Chellappa. Defense-GAN: Protecting classifiers against adversarial attacks using generative models. In International Conference on Learning Representations, 2018. 2" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 56, + 614, + 296, + 668 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 614, + 296, + 668 + ], + "spans": [ + { + "bbox": [ + 56, + 614, + 296, + 668 + ], + "type": "text", + "content": "[37] Ali Shafahi, Mahyar Najibi, Mohammad Amin Ghiasi, Zheng Xu, John Dickerson, Christoph Studer, Larry S Davis, Gavin Taylor, and Tom Goldstein. Adversarial training for free! Advances in neural information processing systems, 32, 2019. 1" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 56, + 669, + 296, + 712 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 669, + 296, + 712 + ], + "spans": [ + { + "bbox": [ + 56, + 669, + 296, + 712 + ], + "type": "text", + "content": "[38] Changhao Shi, Chester Holtz, and Gal Mishne. Online adversarial purification based on self-supervised learning. In International Conference on Learning Representations, 2021. 1" + } + ] + } + ], + "index": 13 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 316, + 73, + 553, + 700 + ], + "type": "list", + "angle": 0, + "index": 30, + "blocks": [ + { + "bbox": [ + 316, + 73, + 553, + 106 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 73, + 553, + 106 + ], + "spans": [ + { + "bbox": [ + 316, + 73, + 553, + 106 + ], + "type": "text", + "content": "[39] Jiaming Song, Chenlin Meng, and Stefano Ermon. Denoising diffusion implicit models. In International Conference on Learning Representations, 2021. 5" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 316, + 107, + 553, + 162 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 107, + 553, + 162 + ], + "spans": [ + { + "bbox": [ + 316, + 107, + 553, + 162 + ], + "type": "text", + "content": "[40] Kaiyu Song, Hanjiang Lai, Yan Pan, and Jian Yin. Mimic diffusion: Purifying adversarial perturbation via mimicking clean diffusion model. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), 2024. 1, 2, 5, 6, 7" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 316, + 163, + 553, + 217 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 163, + 553, + 217 + ], + "spans": [ + { + "bbox": [ + 316, + 163, + 553, + 217 + ], + "type": "text", + "content": "[41] Yang Song, Taesup Kim, Sebastian Nowozin, Stefano Ermon, and Nate Kushman. Pixeldefend: Leveraging generative models to understand and defend against adversarial examples. In International Conference on Learning Representations, 2018. 2, 7" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 316, + 219, + 553, + 262 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 219, + 553, + 262 + ], + "spans": [ + { + "bbox": [ + 316, + 219, + 553, + 262 + ], + "type": "text", + "content": "[42] Linyu Tang and Lei Zhang. Robust overfitting does matter: Test-time adversarial purification with fgsm. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), 2024. 2" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 316, + 263, + 553, + 296 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 263, + 553, + 296 + ], + "spans": [ + { + "bbox": [ + 316, + 263, + 553, + 296 + ], + "type": "text", + "content": "[43] Laurens Van der Maaten and Geoffrey Hinton. Visualizing data using t-sne. Journal of machine learning research, 9 (11), 2008. 7" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 316, + 297, + 553, + 329 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 297, + 553, + 329 + ], + "spans": [ + { + "bbox": [ + 316, + 297, + 553, + 329 + ], + "type": "text", + "content": "[44] Jinyi Wang, Zhaoyang Lyu, Dahua Lin, Bo Dai, and Hongfei Fu. Guided Diffusion Model for Adversarial Purification. arXiv, 2022. 1, 2, 7" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 316, + 331, + 553, + 364 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 331, + 553, + 364 + ], + "spans": [ + { + "bbox": [ + 316, + 331, + 553, + 364 + ], + "type": "text", + "content": "[45] Eric Wong, Leslie Rice, and J. Zico Kolter. Fast is better than free: Revisiting adversarial training. In International Conference on Learning Representations, 2020. 1" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 316, + 365, + 553, + 398 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 365, + 553, + 398 + ], + "spans": [ + { + "bbox": [ + 316, + 365, + 553, + 398 + ], + "type": "text", + "content": "[46] Eric Wong, Leslie Rice, and J. Zico Kolter. Fast is better than free: Revisiting adversarial training. In International Conference on Learning Representations, 2020. 7" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 316, + 399, + 553, + 453 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 399, + 553, + 453 + ], + "spans": [ + { + "bbox": [ + 316, + 399, + 553, + 453 + ], + "type": "text", + "content": "[47] Chaowei Xiao, Zhongzhu Chen, Kun Jin, Jiongxiao Wang, Weili Nie, Mingyan Liu, Anima Anandkumar, Bo Li, and Dawn Song. Densepure: Understanding diffusion models for adversarial robustness. In The Eleventh International Conference on Learning Representations, 2023. 2" + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 316, + 455, + 553, + 498 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 455, + 553, + 498 + ], + "spans": [ + { + "bbox": [ + 316, + 455, + 553, + 498 + ], + "type": "text", + "content": "[48] Yuzhe Yang, Guo Zhang, Dina Katabi, and Zhi Xu. Me-net: Towards effective adversarial robustness with matrix estimation. In International Conference on Machine Learning, pages 7025-7034. PMLR, 2019. 7" + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 316, + 499, + 553, + 543 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 499, + 553, + 543 + ], + "spans": [ + { + "bbox": [ + 316, + 499, + 553, + 543 + ], + "type": "text", + "content": "[49] Jongmin Yoon, Sung Ju Hwang, and Juho Lee. Adversarial purification with score-based generative models. In Proceedings of the 38th International Conference on Machine Learning, 2021. 2" + } + ] + } + ], + "index": 25 + }, + { + "bbox": [ + 316, + 544, + 553, + 578 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 544, + 553, + 578 + ], + "spans": [ + { + "bbox": [ + 316, + 544, + 553, + 578 + ], + "type": "text", + "content": "[50] Jongmin Yoon, Sung Ju Hwang, and Juho Lee. Adversarial purification with score-based generative models. In International Conference on Machine Learning, 2021. 6, 7" + } + ] + } + ], + "index": 26 + }, + { + "bbox": [ + 316, + 578, + 553, + 599 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 578, + 553, + 599 + ], + "spans": [ + { + "bbox": [ + 316, + 578, + 553, + 599 + ], + "type": "text", + "content": "[51] Sergey Zagoruyko. Wide residual networks. arXiv preprint arXiv:1605.07146, 2016. 5" + } + ] + } + ], + "index": 27 + }, + { + "bbox": [ + 316, + 600, + 553, + 654 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 600, + 553, + 654 + ], + "spans": [ + { + "bbox": [ + 316, + 600, + 553, + 654 + ], + "type": "text", + "content": "[52] Dawei Zhou, Tongliang Liu, Bo Han, Nannan Wang, Chunlei Peng, and Xinbo Gao. Towards defending against adversarial examples via attack-invariant features. In Proceedings of the 38th International Conference on Machine Learning, 2021. 2" + } + ] + } + ], + "index": 28 + }, + { + "bbox": [ + 316, + 656, + 553, + 700 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 656, + 553, + 700 + ], + "spans": [ + { + "bbox": [ + 316, + 656, + 553, + 700 + ], + "type": "text", + "content": "[53] Geigh Zollicoffer, Minh Vu, Ben Nebgen, Juan Castorena, Boian Alexandrov, and Manish Bhattarai. Lorid: Low-rank iterative diffusion for adversarial purification. arXiv preprint arXiv:2409.08255, 2024. 6, 7" + } + ] + } + ], + "index": 29 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "text", + "content": "29277" + } + ] + } + ], + "index": 31 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 9 + } + ], + "_backend": "vlm", + "_version_name": "2.6.4" +} \ No newline at end of file diff --git a/2025/Divot_ Diffusion Powers Video Tokenizer for Comprehension and Generation/dd9cf0f7-0275-41db-a729-40fdeab212b3_content_list.json b/2025/Divot_ Diffusion Powers Video Tokenizer for Comprehension and Generation/dd9cf0f7-0275-41db-a729-40fdeab212b3_content_list.json new file mode 100644 index 0000000000000000000000000000000000000000..ba0a8917bab51d61e42955bf24831478cc0c269e --- /dev/null +++ b/2025/Divot_ Diffusion Powers Video Tokenizer for Comprehension and Generation/dd9cf0f7-0275-41db-a729-40fdeab212b3_content_list.json @@ -0,0 +1,1902 @@ +[ + { + "type": "text", + "text": "Divot: Diffusion Powers Video Tokenizer for Comprehension and Generation", + "text_level": 1, + "bbox": [ + 109, + 130, + 885, + 151 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Yuying Ge Yizhuo Li Yixiao Ge Ying Shan ARC Lab, Tencent PCG https://github.com/TencentARC/Divot", + "bbox": [ + 259, + 180, + 736, + 233 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Abstract", + "text_level": 1, + "bbox": [ + 248, + 268, + 326, + 284 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "In recent years, there has been a significant surge of interest in unifying image comprehension and generation within Large Language Models (LLMs). This growing interest has prompted us to explore extending this unification to videos. The core challenge lies in developing a versatile video tokenizer that captures both the spatial characteristics and temporal dynamics of videos to obtain representations for LLMs, and the representations can be further decoded into realistic video clips to enable video generation. In this work, we introduce Divot, a Diffusion-Powered Video Tokenizer, which leverages the diffusion process for self-supervised video representation learning. We posit that if a video diffusion model can effectively de-noise video clips by taking the features of a video tokenizer as the condition, then the tokenizer has successfully captured robust spatial and temporal information. Additionally, the video diffusion model inherently functions as a de-tokenizer, decoding videos from their representations. Building upon the Divot tokenizer, we present Divot-LLM through video-to-text autoregression and text-to-video generation by modeling the distributions of continuous-valued Divot features with a Gaussian Mixture Model. Experimental results demonstrate that our diffusion-based video tokenizer, when integrated with a pre-trained LLM, achieves competitive performance across various video comprehension and generation benchmarks. The instruction tuned Divot-LLM also excels in video storytelling, generating interleaved narratives and corresponding videos. Models and codes are available at https://github.com/TencentARC/Divot.", + "bbox": [ + 89, + 301, + 485, + 738 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "1. Introduction", + "text_level": 1, + "bbox": [ + 91, + 768, + 220, + 784 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "In recent years, the rapid evolution of Multimodal Large Language Models (MLLMs) [12, 15-17, 55-57, 77, 86, 91, 93] has demonstrated significant progresses in unified image understanding and generation, which empowers LLMs [6, 11, 60] with the ability to generate images beyond texts. While these work primarily focus on image-text data, the extension of this unification to the video domain remains", + "bbox": [ + 89, + 794, + 483, + 898 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/0279fe125a1240893f4d721d6970606d320221f4fa2bb5eb7b575fb6b0ce1b4c.jpg", + "image_caption": [ + "Figure 1. We utilize the diffusion procedure to learn a video tokenizer in a self-supervised manner for unified comprehension and generation, where the spatiotemporal representations serve as the condition of a diffusion model to de-noise video clips. Additionally, the proxy diffusion model functions as a de-tokenizer to decode realistic video clips from the video representations." + ], + "image_footnote": [], + "bbox": [ + 519, + 268, + 905, + 390 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "relatively under-explored. Achieving unified video comprehension and generation is essential for the development of more sophisticated artificial intelligence (AI) systems that are capable of understanding and creating dynamic visual content in the real world.", + "bbox": [ + 511, + 507, + 906, + 582 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "The primary challenge of achieving unified video comprehension and generation lies in developing a versatile video tokenizer that can effectively address the complexities inherent in video data. This tokenizer should be able to obtain robust video representations that serve as inputs of MLLMs for video comprehension, and these representations can be further decoded into realistic video clips to enable video generation. Unlike static images, videos encompass both spatial characteristics and temporal dynamics, making their representation significantly more complex. Recent pioneering work [26, 40, 74] adopt a discrete video tokenizer for unifying video comprehension and generation, where a video is represented as a sequence of discrete frame tokens, or keyframe tokens followed by discrete motion tokens. This approach eases video generation with a LLM through an autoregressive next-token prediction mechanism, but sacrifices the performance of multimodal understanding, as pointed out by recent work [77]. In this work, we aim to investigate an alternative approach by utilizing continuous video representations to unify video comprehension and generation.", + "bbox": [ + 511, + 583, + 908, + 883 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "To this end, we introduce Divot, a Diffusion-Powered", + "bbox": [ + 532, + 885, + 906, + 898 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "CVF", + "bbox": [ + 106, + 2, + 181, + 42 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore.", + "bbox": [ + 236, + 0, + 807, + 44 + ], + "page_idx": 0 + }, + { + "type": "page_number", + "text": "13606", + "bbox": [ + 480, + 945, + 517, + 955 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Video Tokenizer that leverages the diffusion process [50] for self-supervised video representation learning as shown in Fig. 1. The core premise is that if a diffusion model can effectively predict the noise added to the Variational Autoencoder (VAE) latents [28] of video clips, when conditioned on the features produced by the video tokenizer, it demonstrates that the tokenizer has successfully captured robust spatial and temporal information inherent in the video data. This capability is crucial for representing the intricate dynamics present in videos. Furthermore, in addition to being a proxy module for learning the tokenizer, the diffusion model can act as a de-tokenizer to effectively decode realistic videos from their learned representations. This dual functionality facilitates a seamless integration of understanding and creating video content within a LLM.", + "bbox": [ + 89, + 90, + 480, + 316 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Specifically, the Divot tokenizer is composed of a pretrained Vision Transformer (ViT) encoder [13], a Spatial-Temporal transformer, and a Perceiver Resampler [1] to obtain video representations from video frames sampled at low frame rate (fps) considering the semantic redundancy between adjacent frames. The video representations serve as the condition of a pre-trained video diffusion model, DynamiCrafter [78] (without the concatenation of a conditional image with initial noise), to predict the noise added to the VAE latents of video frames. After training, the video diffusion model can generate realistic video clips from noise by taking the video representations provided by the Divot tokenizer as the condition.", + "bbox": [ + 89, + 319, + 482, + 513 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "We further present Divot-LLM by equipping the pretrained Mistral-7B [24] with the Divot tokenizer. Divot-LLM is pre-trained with a next-word prediction objective on video-caption data by taking the spatiotemporal representations as inputs for video comprehension. The challenge then arises in modeling the continuous video representations with the LLM for video generation. We empirically find that simply minimizing the distance between the LLM output and video representations using mean squared error (MSE) loss achieves unsatisfactory results, since the deterministic regression regularizes the LLM to learn overly averaged features of videos. To address this, inspired by recent work [33], we shift our focus from deterministic modeling to probabilistic modeling by modeling the distributions of video features with a Gaussian Mixture Model (GMM). Specifically, we train the LLM to predict GMM parameters, including means, variances, and mixture probabilities through minimizing the discrepancy between the predicted GMM distribution and the actual video representations using negative log-likelihood (NLL) loss. During inference, we draw samples from the predicted GMM distribution as the condition of the video de-tokenizer to decode video clips.", + "bbox": [ + 89, + 516, + 482, + 849 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "We benchmark Divot-LLM on a broad range of video comprehension tasks and zero-shot video generation, achieving competitive performance through pre-training on 5 million video-text pairs. By leveraging the generality of the video tokenizer, our Divot-LLM also enables video storytelling, which generates interleaved narratives and corresponding videos that are temporally coherent through fine-tuning on a specific animation dataset.", + "bbox": [ + 511, + 90, + 906, + 210 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Our contributions are three-fold. (1) We introduce Divot, an advanced video tokenizer that leverages a diffusion procedure for self-supervised video representation learning, aiming to unify video comprehension and generation. (2) We present Divot-LLM, composed of a pre-trained LLM and the Divot tokenizer to enable understanding and generating video content within a single framework. We investigate effective approaches for fitting continuous video representations using the LLM with probabilistic modeling for video generation. (3) We conduct extensive experiments to demonstrate Divot-LLM's competitive performance on existing video comprehension and generation benchmarks, as well as video storytelling. All models and code are released.", + "bbox": [ + 511, + 213, + 908, + 409 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "2. Related Work", + "text_level": 1, + "bbox": [ + 513, + 426, + 653, + 441 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "MLLMs for Comprehension and Generation. With the rapid development of Multimodal Large Language Models (MLLM), recent studies have been working on unified MLLMs [12, 15-17, 25, 26, 36, 40, 41, 54, 55, 57, 67, 72-74, 77, 83, 84, 86, 89, 91, 93] that are capable of multimodal comprehension and generation. To empower LLMs with the capability to generate visual content, existing work primarily employs the following three approaches: (1) utilizing a pre-trained stable diffusion model to generate images conditioned on LLM output (either continuous features or discrete tokens); (2) employing a Vector Quantized (VQ) [63] based decoder to generate visual content from the discrete codes predicted by LLMs; (3) using LLMs to de-noise Gaussian noise through a diffusion process. While most work predominantly focus on the unification of images and texts, some pioneering studies [26, 40, 67, 74] further advance the integration of video comprehension and generation within an LLM through generating videos from discrete codes using a VQ-based decoder, which falls into the second approach. In this work, we adopt the first approach, which involves leveraging a diffusion model to achieve unified video understanding and generation from continuous representations.", + "bbox": [ + 511, + 452, + 908, + 785 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Video Tokenizer in MLLMs. Previous work on video generation with LLMs predominantly employs a discrete video tokenizer to convert video signals into a sequence of quantized tokens. For example, LWM [40] and VILA-U [74] utilize a frame-level tokenizer to discretize each frame into a sequence of codes. VideoPoet [29], Loong [69] and", + "bbox": [ + 511, + 809, + 908, + 902 + ], + "page_idx": 1 + }, + { + "type": "page_footnote", + "text": "1We do not explore more advanced LLMs because we want to ensure that our superiority stems from the improved visual representations, rather than from the capabilities of a more sophisticated foundation model.", + "bbox": [ + 89, + 862, + 482, + 900 + ], + "page_idx": 1 + }, + { + "type": "page_number", + "text": "13607", + "bbox": [ + 480, + 944, + 517, + 955 + ], + "page_idx": 1 + }, + { + "type": "image", + "img_path": "images/15287b5e77dd53b6bbf1e4a00e7adf0725fc8dcf61978a991b5d9c4799470da6.jpg", + "image_caption": [ + "Figure 2. Overview of Divot tokenization and de-tokenization. During training, sparsely sampled video frames are fed into the tokenizer to obtain spatiotemporal representations. These representations serve as the conditions for a U-Net, which is trained to de-noise the noisy VAE latents of densely sampled video frames. During inference, the video representations from the Divot tokenizer can be decoded into realistic video clips with the U-Net." + ], + "image_footnote": [], + "bbox": [ + 122, + 88, + 669, + 244 + ], + "page_idx": 2 + }, + { + "type": "image", + "img_path": "images/a93c4e7225ea5b543ccb32551a8fe9109bec5f3797448e75c94d2701f6dc2326.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 684, + 89, + 875, + 239 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Emu3 [67] leverage a 3D CNN architecture, where the encoded spatial-temporal features are quantized into discrete tokens. Video-LaVIT [26] represents video clips as a keyframe followed by extracted motion vectors, obtaining the respective discrete codes. By converting continuous visual signals into discrete tokens, the original next-token prediction mechanism can be adopted to facilitate video generation with an LLM. However, recent work [77] observes a significant performance degradation in multimodal comprehension tasks when discrete representations are used instead of continuous representations. In this work, we introduce a video tokenizer with continuous representations through leveraging the diffusion [50] procedure, enabling it to be effectively integrated with a LLM for unified comprehension and generation.", + "bbox": [ + 91, + 328, + 486, + 541 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Diffusion for Representation Learning. The diffusion process has been explored as a criterion for representation learning. Some works [3, 75, 80, 90] leverage the intermediate activations of pre-trained diffusion models for downstream tasks including classification, segmentation and depth estimation. Other works [23, 65, 70] employ the diffusion model as a proxy module for self-supervised learning, where noisy inputs are de-noised by conditioning on the image representations. This approach encourages the emergence of informative representations that capture key properties and semantics of the images. In this work, to the best of our knowledge, we for the first time leverage diffusion for video representation learning, where a video diffusion model is trained to de-noise video clips through taking the spatiotemporal representations as conditions, thereby encouraging the capture of spatial characteristics and temporal dynamics.", + "bbox": [ + 89, + 563, + 483, + 806 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3. Method", + "text_level": 1, + "bbox": [ + 89, + 821, + 181, + 835 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3.1. Divot Tokenizer", + "text_level": 1, + "bbox": [ + 89, + 847, + 250, + 862 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "We introduce Divot, a diffusion-powered tokenizer that leverage diffusion procedure for video representation learning.", + "bbox": [ + 89, + 869, + 485, + 902 + ], + "page_idx": 2 + }, + { + "type": "image", + "img_path": "images/2c531e3cd50b981ad0899df0f49e891e16f9be0a616fdad7e20a153fbf4c78e4.jpg", + "image_caption": [ + "Figure 3. Overview of Divot-LLM. Video features from the Divot tokenizer are fed into the LLM to perform next-word prediction for video comprehension, while learnable queries are input into the LLM to model the distributions of Divot features using a Gaussian Mixture Model (GMM) for video generation. During inference, video features are sampled from the predicted GMM distribution to decode videos using the de-tokenizer." + ], + "image_footnote": [], + "bbox": [ + 516, + 325, + 906, + 494 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Additionally, the proxy diffusion model used for training the tokenizer can serve as a de-tokenizer to decode realistic video clips from their spatiotemporal representations.", + "bbox": [ + 511, + 613, + 906, + 659 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3.1.1. Preliminary: Video Diffusion Model.", + "text_level": 1, + "bbox": [ + 511, + 667, + 815, + 683 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Diffusion models [20, 50] learn to model a probability distribution by reversing a process that progressively adds noise to the data. Specifically, given data $\\mathbf{x}_0 \\sim p(\\mathbf{x})$ , the forward process gradually adds random Gaussian noise $\\epsilon_{\\mathbf{t}} \\in \\mathcal{N}(\\mathbf{0}, \\mathbf{I})$ to the data sample $\\mathbf{x}_0$ with a total of $T$ timesteps to yield $\\mathbf{x}_t$ through a parameterization trick. The denoising process predicts $\\epsilon_{\\mathbf{t}}$ in the forward diffusion process with a denoising network $\\epsilon_{\\theta}(\\mathbf{x}_t, t)$ , which is trained by the objective below,", + "bbox": [ + 511, + 686, + 908, + 809 + ], + "page_idx": 2 + }, + { + "type": "equation", + "text": "\n$$\n\\min _ {\\theta} \\mathbb {E} _ {t, \\mathbf {x} \\sim p, \\epsilon \\sim \\mathcal {N} (\\mathbf {0}, \\mathbf {I})} \\| \\epsilon - \\epsilon_ {\\theta} (\\mathbf {x} _ {t}, t) \\| _ {2} ^ {2}, \\tag {1}\n$$\n", + "text_format": "latex", + "bbox": [ + 580, + 825, + 906, + 847 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "where $\\epsilon$ is the sampled Gaussian noise and $\\theta$ indicates the parameters of the denoising network. During inference, we can perform iterative denoising from a random Gaussian", + "bbox": [ + 511, + 854, + 908, + 900 + ], + "page_idx": 2 + }, + { + "type": "page_number", + "text": "13608", + "bbox": [ + 480, + 944, + 517, + 955 + ], + "page_idx": 2 + }, + { + "type": "image", + "img_path": "images/3f1baa720b8b0a5f515f4ed0d31093760eece10aa567131fefa68bced55b38b9.jpg", + "image_caption": [ + "Figure 4. Paradigms for modeling video representations from the Divot tokenizer with a LLM for video generation. (a) MSE Regression, where the LLM output is trained to minimize its distance with video features using Mean Squared Error (MSE) loss; (b) Diffusion Modeling, where the LLM output is fed into a denoising network as the condition to predict the noise added to video features; (c) GMM Modeling, where the LLM output is trained to predict the parameters of a Gaussian Mixture Model (GMM) for modeling video feature distributions." + ], + "image_footnote": [], + "bbox": [ + 102, + 95, + 352, + 268 + ], + "page_idx": 3 + }, + { + "type": "image", + "img_path": "images/3369033bdb1d68330b8fe8c23297d80ba63ecc490d094b81f8eed65ce4c81068.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 377, + 95, + 619, + 268 + ], + "page_idx": 3 + }, + { + "type": "image", + "img_path": "images/02be3176e5e209bde863f639edd20d1be47c8279641f639ac8d146a2205a4479.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 656, + 89, + 908, + 270 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "noise for the denoised data $\\mathbf{x}_0$", + "bbox": [ + 89, + 357, + 292, + 372 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "For video diffusion models [8, 78], given a video $\\mathbf{x}$ , a latent representation $\\mathbf{z} = \\mathcal{E}(\\mathbf{x})$ is first encoded to reduce the computational complexity. Then the forward diffusion process and backward denoise process are performed in this latent space with a denoising network $\\epsilon_{\\theta}(\\mathbf{x}_t,\\mathbf{c},t)$ , where $\\mathbf{c}$ denotes denoising conditions like text or visual prompts.", + "bbox": [ + 89, + 375, + 483, + 467 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "3.1.2. Training Pipeline", + "text_level": 1, + "bbox": [ + 89, + 482, + 256, + 498 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "As illustrated in Fig. 2, given a video clip, we separately sample sparse frames at 2 fps to obtain the video representations from the tokenizer, and sample dense frames at 8 fps to obtain latent representations $z_{0}$ from the frozen VAE [28] encoder. Sparse frames are sampled as the input of the video tokenizer considering the semantic redundancy between adjacent frames. The forward diffusion process gradually adds Gaussian noise $\\theta$ to $z_{0}$ for producing the noisy input $z_{t}$ . At each backward step $t$ , a denoising U-Net is trained to predict the noise added from the previous step to the current step by taking the time embedding and video representations as the condition. Specifically, the video representations interact with the denoising U-Net intermediate features through cross-attention layers, where each noisy latent attends to all video tokens. By constraining the U-Net to reconstruct fine-grained spatial and temporal information of video clips through relying on video features, the Divot tokenizer is optimized to capture both spatial characteristics and temporal dynamics for robust video representations. The Divot tokenizer is trained on pure videos of a subset of WebVid10M [2] and Panda-70M [9], totaling 10M videos.", + "bbox": [ + 89, + 503, + 483, + 821 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "After training the Divot tokenizer, the proxy denoising U-Net (employed to implement the parameterized loss function) can serve as an effective video de-tokenizer, which is able to decode semantically aligned video clips from the learned spatiotemporal representations as shown in Fig. 5.", + "bbox": [ + 89, + 824, + 483, + 901 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "3.1.3. Model Architecture", + "text_level": 1, + "bbox": [ + 511, + 357, + 696, + 371 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "As shown in Fig. 2, the Divot tokenizer is composed of a pre-trained ViT encoder to extract frame-level features, a transformer for spatial and temporal fusion, and a Perceiver Resampler [1] to produce a fixed number of video tokens. The Perceiver Resampler is adopted for two reasons: (1) to reduce the number of video tokens that a LLM need to predict for generation, and (2) to transform the patch-position dependent features into a sequence of high-level features without 3D positional dependencies, which we empirically find easier for an LLM to fit (See Sec. 4.3). Specifically, given a video clip with a duration of two seconds, we sample 5 frames at 2 fps, resulting in a total of 64 video tokens. We adopt the de-noising U-Net in DynamiCrafter [78], but reduce the input channel of the 3D convolution from 8 to 4 since we remove the original concatenation of a conditional image with noisy latents.", + "bbox": [ + 511, + 377, + 908, + 619 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "3.2. Video Representation Modeling with LLM", + "text_level": 1, + "bbox": [ + 511, + 633, + 877, + 648 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "The core challenge of generating videos using a LLM with the Divot tokenizer lies in effectively modeling the continuous video features. The most straightforward solution is to minimize the distance between the LLM output and the video representations using mean squared error (MSE) loss following previous work [17, 54] for image generation as illustrated in Fig. 4 (a). However, we empirically find that this approach is not effective for modeling continuous video features, as the generated videos tend to exhibit repeating patterns. We analyze that the deterministic regression regularizes the LLM to learn overly averaged representations, which is particularly catastrophic in video generation as videos must ensure both spatial and temporal diversity.", + "bbox": [ + 511, + 657, + 908, + 853 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Inspired by recent work MAR [33], instead of deterministic regression, we aim to model the probability distribution of video representations using the LLM. As shown in Fig. 4,", + "bbox": [ + 511, + 854, + 908, + 900 + ], + "page_idx": 3 + }, + { + "type": "page_number", + "text": "13609", + "bbox": [ + 480, + 944, + 517, + 955 + ], + "page_idx": 3 + }, + { + "type": "image", + "img_path": "images/e028b6c9192fbea1821cc07ff0fe92c835c98e38529b8974e24655cf6c2b4978.jpg", + "image_caption": [ + "Input video (fps=2)" + ], + "image_footnote": [], + "bbox": [ + 98, + 104, + 323, + 191 + ], + "page_idx": 4 + }, + { + "type": "image", + "img_path": "images/4fd0c41406fe355eb7d0c864097144f86308dd90d94c8798281db3ccce5449d7.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 98, + 194, + 321, + 280 + ], + "page_idx": 4 + }, + { + "type": "image", + "img_path": "images/c76ac15c2c3f96328f358f06af6cc5f10fa785219565aa151ac4116a61eb3b50.jpg", + "image_caption": [ + "Figure 5. Reconstructed videos, where the Divot tokenizer obtains spatiotemporal representations of sparsely sampled video frames and the de-tokenizer decodes these representations into semantically aligned and temporally coherent video clips." + ], + "image_footnote": [], + "bbox": [ + 98, + 282, + 321, + 368 + ], + "page_idx": 4 + }, + { + "type": "image", + "img_path": "images/96a417e54fcba8ad56e881dc765b9369ff474d383b2dc1fff2f1c716947ee215.jpg", + "image_caption": [ + "Reconstructed video (fps=8)" + ], + "image_footnote": [], + "bbox": [ + 341, + 104, + 901, + 191 + ], + "page_idx": 4 + }, + { + "type": "image", + "img_path": "images/8bbd5fc6d4255b8814b2441e99df74428962677b3f768c5416b9b9d40b5e46b2.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 341, + 194, + 901, + 280 + ], + "page_idx": 4 + }, + { + "type": "image", + "img_path": "images/d57101b1d7c19c1fa8232c33f41382eece25fe7e8079fa2ddec2ba631fb64ace.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 341, + 282, + 901, + 368 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "we explore two approaches including (b) Diffusion Modeling [33] and (c) GMM Modeling [61]. Specifically, for the diffusion modeling, given continuous-valued video features to be predicted, the LLM produces output, which serves as the condition of a denoising network (a small MLP) to predict the Gaussian noise added to the video features. The diffusion model is trained for representing the distribution of video features. For GMM modeling, we use a Gaussian Mixture Model (GMM) to model the distribution of the video features, and train the LLM to predict $2kd + k$ parameters per video token ( $kd$ mean and $kd$ variance parameters for the mixture components, and $k$ mixture probabilities). We optimize the LLM by minimizing the discrepancy between the predicted GMM distribution and the video representations with negative log-likelihood (NLL) loss.", + "bbox": [ + 88, + 429, + 485, + 655 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "During inference, in diffusion modeling, the denoising network denoise the final video features from Gaussian noise gradually by taking the LLM output as the condition. In GMM modeling, we draw samples from the predicted GMM distribution as the final video representations. To empirically investigate the effectiveness of the approaches above, we train the LLM with MSR-VTT [79], and evaluates text-to-video generation on test set with FVD [62] and similarity score [49] as the metric following previous work [26, 82]. As listed in Tab. 7, GMM modeling achieves better performance than diffusion modeling and MSE regression in video generation. We speculate that high-level features obtained by the Divot tokenizer are more sensitive to Gaussian noise compared to the VAE latents used by MAR, making training more challenging and resulting in suboptimal results. Therefore, we adopt GMM modeling to train Divot-LLM.", + "bbox": [ + 89, + 659, + 485, + 900 + ], + "page_idx": 4 + }, + { + "type": "table", + "img_path": "images/9f15401bbe06cc251318e85e9b6ef6cf48e627608fc28a7381f71d7077015e1d.jpg", + "table_caption": [ + "Table 2. Datasets used for training the tokenizer and Divot-LLM." + ], + "table_footnote": [], + "table_body": "
StageTypeDataset
TokenizerPure VideoWebVid-10M [2], Panda-70M [9]
Pre-trainVideo-textWebVid-10M [2]
Image-textCC3M [52], CapsFusion [87], LAION-COCO [51]
SFTClassificationKinetics-710 [27], SSV2 [18]
VQATGIF [34], NextQA [76], CLEVRER [85], YouCook2 [92], PerceptionTest[48], EgoQA [19], ActivityNetQA[88]
InstructionVideo-ChatGPT[43], LLaVA-mixed[39], Valley [42], LLaVA-Video-178K[37]
GenerationWebVid-10M [2]
StoryTellingIn-house data
", + "bbox": [ + 517, + 452, + 906, + 621 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "3.3. Pre-training and Instruction Tuning", + "text_level": 1, + "bbox": [ + 511, + 646, + 826, + 662 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "3.3.1. Training Stage I: Multimodal Pre-training", + "text_level": 1, + "bbox": [ + 511, + 669, + 854, + 684 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "As shown in Fig 3, Divot-LLM adopts next-word prediction and GMM modeling on video-text data for video comprehension and generation. Specifically, the video features from the Divot tokenizer, the special tokens indicating the start and end of video features, along with the text tokens of the caption are fed into the pre-trained Mistral-7B [24] for next token prediction trained with cross-entropy loss. Text tokens of the caption and $N$ learnable queries are input into the LLM, where the output of the learnable queries are trained via bidirectional attention to model a GMM distribution for the video features using NLL loss. During inference, we draw samples from the predicted GMM distribution as condition of the denoising U-Net to decode videos. We pre-train Divot-LLM from the pre-trained Mistral-7B model using", + "bbox": [ + 509, + 688, + 906, + 900 + ], + "page_idx": 4 + }, + { + "type": "page_number", + "text": "13610", + "bbox": [ + 480, + 944, + 517, + 955 + ], + "page_idx": 4 + }, + { + "type": "table", + "img_path": "images/e70e9b3a2c4310f4a435710d54da999e4aa2636f1ec784a27382e6faf7e24baf.jpg", + "table_caption": [ + "Table 3. Comparison for video comprehension with MLLMs. \"Video-Gen\" denotes whether the model can generate videos besides texts. The evaluation metric is accuracy. The best results are bold and the second best results are underlined." + ], + "table_footnote": [], + "table_body": "
ModelLLM sizeVideo-GenEgoSchemaPerception-TestMVBenchMSVDActivityNet
Gemini 1.0 Pro [58]-×55.751.1--49.8
Gemini 1.5 Pro [59]-×63.2---56.7
GPT4-V [46]-×55.6-43.7-59.5
GPT4-O [47]-×72.2---61.9
LLaMA-VID [35]7B×38.544.641.969.747.4
Video-ChatGPT [43]7B×---64.935.2
Video-LLaVA [37]7B×38.444.341.070.745.3
VideoChat2 [31]7B×42.247.351.170.049.1
LLaVA-NeXT-Video [38]7B×43.948.846.567.853.5
LLaVA-NeXT-Video [38]32B×60.9---54.3
PLLAVA [81]34B×-58.1--60.9
LLaVA-OneVision [30]72B×62.0---62.3
VideoLLaMA2 [10]7B×51.751.454.670.950.2
VideoLLaMA2 [10]72B×63.957.562.071.055.2
LWM [40]7B✓---55.9-
Video-LaVIT [26]7B✓37.347.9-73.250.1
VILA-U [74]7B✓---75.352.7
Divot-LLM7B✓46.558.352.176.455.8
", + "bbox": [ + 151, + 128, + 849, + 395 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "LoRA [22] on a subset of WebVid-10M [2] data (filtered for temporal dynamics in captions) and image-text data.", + "bbox": [ + 89, + 420, + 482, + 450 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "3.3.2. Training Stage II: Multimodal Instruction Tuning", + "text_level": 1, + "bbox": [ + 89, + 458, + 482, + 474 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "We perform multimodal instruction tuning on Divot-LLM to align it with human instructions through supervised fine-tuning on public datasets as listed in Tab. 2 with a LoRA module. We further fine-tune the pretrained Divot-LLM on an animated series called \"Curious George\" to achieve video storytelling, which generates storyline and corresponding video clips in an interleaved manner.", + "bbox": [ + 89, + 477, + 483, + 583 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "4. Experiment", + "text_level": 1, + "bbox": [ + 89, + 597, + 215, + 614 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "4.1.Quantitative Evaluation", + "text_level": 1, + "bbox": [ + 89, + 621, + 310, + 636 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Video Comprehension. We conduct extensive evaluations on video comprehension benchmarks including Multi-choice Video Question Answering (MC-VQA) on EgoSchema [44], Perception-Test [48], MVBench [32], and Open-Ended Video Question Answering (OE-VQA) on MSVD [7], ActivityNet [88]. Following VideoLLaMA 2 [10], we utilize GPT-3.5 to assess the quality of the generated answers of OE-VQA by determining whether the answers match the ground truth, and we report the percentage of \"Yes\" as Accuracy.", + "bbox": [ + 89, + 643, + 482, + 779 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "For each testing video, we sample a maximum of 20 clips, each containing 5 frames. The evaluation results are reported in Tab. 3. Divot-LLM outperforms the baseline models that can generate both texts and videos, which demonstrates that our model effectively achieves video comprehension within a unified framework. Compared to VideoLLMs specifically designed for video comprehension of the same model size of LLM, Divot-LLM achieves competitive results with sig", + "bbox": [ + 89, + 780, + 483, + 901 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "nificantly fewer video-caption pairs for training (4.8M vs. 100M in VideoLLaMA 2). By utilizing diffusion procedure for video representation learning, our Divot tokenizer effectively captures robust spatiotemporal representations, enhancing the comprehension capabilities.", + "bbox": [ + 511, + 420, + 906, + 496 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Video Generation. We evaluate zero-shot text-to-video generation on MSR-VTT [79]. We randomly sample one caption for each testing video and generate 16 frames in $256 \\times 256$ px resolution. We adopt the CLIP similarity (CLIPSIM) [71] and Fréchet video distance (FVD) [62] as the evaluation metric following Loong [69]. As listed in Tab. 4, Divot-LLM achieves performance comparable to existing video generation models in terms of visual quality and semantic alignment with captions using only 4.8 million video-text pairs for training.", + "bbox": [ + 511, + 508, + 908, + 660 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "4.2. Qualitative Evaluation", + "text_level": 1, + "bbox": [ + 511, + 669, + 723, + 686 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Text-to-video Generation. We perform qualitative comparison of text-to-video generation with baseline MLLMs that are capable of unified video comprehension and generation. As illustrated in Fig. 6, through modeling the distributions of Divot features with a predicted GMM, our Divot-LLM can generate videos that are both semantically aligned with text prompts and temporally coherent within frames.", + "bbox": [ + 511, + 691, + 906, + 797 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Video StoryTelling. We fine-tune the pre-trained Divot-LLM on an animated series called \"Curious George\" for video storytelling. As shown in Fig. 7, given a brief story instruction, our Divot-LLM can generate a sequence of multimodal stories with rich narrative text and contextually relevant videos that are temporally coherent. Since we only", + "bbox": [ + 511, + 809, + 908, + 901 + ], + "page_idx": 5 + }, + { + "type": "page_number", + "text": "13611", + "bbox": [ + 480, + 945, + 517, + 955 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/0b02b84649d532ae8e5681d4501877047acbf4edb7aeed1b24b63c74821bc370.jpg", + "image_caption": [ + "A man walking slowly under a sunset." + ], + "image_footnote": [], + "bbox": [ + 99, + 107, + 506, + 330 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/a72e36cf9679763404ae7e2a336cd67d6453c7b60a526a4c8db7194d3542a39d.jpg", + "image_caption": [ + "A person is applying eye makeup.", + "Figure 6. Qualitative comparison of text-to-video generation with MLLMs that are capable of unified video comprehension and generation. Divot-LLM effectively generates videos that are both semantically aligned with text prompts and temporally coherent within frames." + ], + "image_footnote": [], + "bbox": [ + 522, + 107, + 901, + 330 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "fine-tune the de-tokenizer for adaptation to the new domain, it demonstrates the generalizability of our Divot tokenizer for obtaining robust video representations.", + "bbox": [ + 89, + 386, + 483, + 433 + ], + "page_idx": 6 + }, + { + "type": "table", + "img_path": "images/6257c1027943f903ed565d39ab3884268d51214b1738eef774b7be63b5673eab.jpg", + "table_caption": [ + "Table 4. Comparison for zero-shot text-to-video generation. \"Data size\" refers to the number of training video data, and \"Unified\" denotes if the model enables video comprehension and generation. The best results are bold and the second best results are underlined." + ], + "table_footnote": [], + "table_body": "
ModelData sizeUnifiedMSR-VTT
CLIPSIM (↑)FVD (↓)
CogVideo [21]5.4M×0.26311294
Video LDM [5]10M×0.2929-
VideoComposer [66]10M×0.2932580
InternVid [68]28M×0.2951-
Make-A-Video [53]20M×0.3049-
VideoPoet [29]270M×0.3049213
PYoCo [14]22.5M×--
SVD [4]152M×--
Video-LavIT [26]10M✓0.3012188.36
Loong [69]16M×0.2903274
Snap Video [45]-×0.2793110.4
VILA-U [74]1M✓0.2937499.06
Divot-LLM4.8M✓0.2938301.4
", + "bbox": [ + 94, + 515, + 483, + 714 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "4.3. Ablation Study", + "text_level": 1, + "bbox": [ + 89, + 741, + 243, + 758 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Diffusion for Video Comprehension. We design two baselines to validate the effectiveness of the diffusion procedure to learn spatiotemporal representations for VideoLLMs. As shown in Tab. 5, both models are pre-trained on Valley [42] and instruction tuned on Video-ChatGPT [43]. The model with diffusion loss employs our Divot tokenizer, while the model with caption loss adopts the same architecture but its tokenizer is pre-trained using captioning loss with a frozen LLM on Valley. The model that employs a video tokenizer", + "bbox": [ + 89, + 763, + 485, + 900 + ], + "page_idx": 6 + }, + { + "type": "table", + "img_path": "images/bc68c23cce5556373a0b767eb8bcecd820fd2298fa560cfe748e8f99e0b68bb9.jpg", + "table_caption": [ + "Table 6. Ablation study on the training objective of the video tokenizer. The evaluation metric is accuracy." + ], + "table_footnote": [], + "table_body": "
Loss TypeMV-BenchMSVDActivityNet
Caption30.866.143.2
Diffusion33.268.944.3
", + "bbox": [ + 555, + 422, + 864, + 481 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "trained with diffusion loss achieves better performance in video comprehension benchmarks, demonstrating that the diffusion process can effectively learn robust video representations in a self-supervised manner, without the need for paired caption annotations.", + "bbox": [ + 511, + 494, + 906, + 571 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Video Generation with LLM. We perform various ablation studies to explore an effective approach for generating videos with a LLM through training on MSR-VTT training set and evaluating text-to-video generation on test set. We use ViT-G/14 to calculate CLIPSIM for better discrimination.", + "bbox": [ + 511, + 583, + 906, + 657 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Q1: Which type of video representations is easier? We investigate two types of video representations, patch-position dependent features obtained from a spatial-temporal transformer and patch-position independent features after a Perceiver Resample with learnable queries. As listed in Tab. 7, fitting features without 3D positional dependencies achieves higher performance, which is also observed in recent work [64]. We also experiment with training the video tokenizer in the VAE manner, which involves predicting the means and variances of a normal distribution and sampling video representations using the re-parametrization trick following GIVT [61]. However, we observe that it is difficult for the LLM to converge and the video de-tokenizer achieves unsatisfactory reconstruction results. We conclude that introducing variances during tokenization for high-level video features may not be appropriate.", + "bbox": [ + 511, + 659, + 908, + 901 + ], + "page_idx": 6 + }, + { + "type": "page_number", + "text": "13612", + "bbox": [ + 480, + 944, + 519, + 957 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/7022943fd68f23b1372a3c781dd5842396777c6dec5f8e41997da890f68aa810.jpg", + "image_caption": [ + "Generate a story about the adventures of George and his friends." + ], + "image_footnote": [], + "bbox": [ + 99, + 119, + 125, + 135 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "A young girl found a glowing orb floating in the air through her magnifying glass.", + "bbox": [ + 104, + 138, + 202, + 176 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/166b68bf5ff8b78bdad2ad3ffdb815b017565f5380946f18a970cf0a8f9905ec.jpg", + "image_caption": [ + "REC" + ], + "image_footnote": [], + "bbox": [ + 210, + 133, + 277, + 181 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/69b87821acbc5e3da29188c4bca7b83be87ac446d3727123fc2707af0d18b3cb.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 279, + 133, + 343, + 181 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/2b4d3b08741527aac498461eeb2aaf3e64018eaf6239c3d87bd71bdf618cba05.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 344, + 133, + 406, + 181 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/cf4fc4340b189061969660e8c73d648e7afa42af4aaa7b7c5ab443e98007191e.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 408, + 133, + 473, + 181 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/58d0788f6a4f1921e7abce5a5b43bcc982c73f5efbfeaa6e35bd3d1e7a28cff5.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 102, + 183, + 125, + 198 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "George watched a bubble float by, feeling curious and joyful.", + "bbox": [ + 102, + 208, + 200, + 237 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/663b0c511275a713084ce388d5e58858ad353e5c439a03fc798bbea67e5a7d5e.jpg", + "image_caption": [ + "REC" + ], + "image_footnote": [], + "bbox": [ + 210, + 199, + 279, + 250 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/6571762ac9125fde0225c61cb424cf678e8ab746d45ee3ea2257e35bf1df6de9.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 279, + 191, + 343, + 250 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/732d370b18165ad577165a633aaef6d61e0ed000fc3db1b4c15ea1d242026f7d.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 344, + 191, + 406, + 250 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/a73b7c641c469b9ad59751c25596f4e2c5e5d5f575f32333cfc260c2ed1f1d12.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 408, + 191, + 473, + 250 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/d114e83331db527c59ec8e749906a2a7e4644a2aae4b0100d3f802abbed57ec0.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 102, + 253, + 125, + 267 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "At lunchtime, a boy found a giant sandwich and felt excited.", + "bbox": [ + 102, + 273, + 202, + 301 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/ca2be2eabe2d27eac255c7ae80637d81d6f21998f07b2ca0c21163cb09471cdc.jpg", + "image_caption": [ + "REC", + "Figure 7. Qualitative examples of video storytelling by Divot-LLM. Given a story instruction, Divot-LLM can generate rich textual narratives along with corresponding video clips that are temporally coherent in an interleaved manner." + ], + "image_footnote": [], + "bbox": [ + 210, + 265, + 279, + 314 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/2b90b82b5ddc80533a7301225c5f54c341b591e9d15d784749954fd092069cfa.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 279, + 265, + 343, + 314 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/cc0e86a83283ba3f37a38fec425b26e09d972fdb58e269108aa9b2abb107a5b8.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 344, + 265, + 406, + 314 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/82930692f9ad0fa4d1d4614bdfa912a1a08edf357afd5414070d44d6853bdc26.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 408, + 265, + 473, + 314 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Generate a story about George's fun-filled day in the kitchen.", + "text_level": 1, + "bbox": [ + 535, + 98, + 828, + 109 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/3d06225ce82fabf69e2ea3482d03d0af3497b5ffb1c7919492f34fd49b8a495b.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 531, + 121, + 553, + 135 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "George and his pal joyfully cook in the kitchen, creating a tasty snack with a big blue book.", + "bbox": [ + 534, + 138, + 625, + 184 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/5f2a88b8f3450114f09b9326cada4bff1fd37d2023e4338c2d521444ac7931fc.jpg", + "image_caption": [ + "REC" + ], + "image_footnote": [], + "bbox": [ + 633, + 133, + 697, + 181 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/142850e5fd9df78388957622b4ac806287f40852b5b651259457f944ca74f905.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 699, + 133, + 763, + 181 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/5a1ece1ff70131622e975eaf4fc2a7e230536045fa6d74f54f69e9bd3dc8ffd6.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 764, + 133, + 828, + 181 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/b5b94209683463df72a5d9a437d2130b0ee199beaea00fe41075d21d27cfc245.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 828, + 133, + 893, + 181 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/27f2070d8c1bbbb8732270a15cdd8b089e19fa85a03f0c20a666e8e316d77606.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 531, + 185, + 553, + 198 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "A woman in the kitchen shows George how to use a new kitchen gadget.", + "bbox": [ + 534, + 200, + 614, + 247 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/792a57c607f69d7d4714ffb0e08a87a6d479a29703eeaac9e39eb9b37ca61a83.jpg", + "image_caption": [ + "REC" + ], + "image_footnote": [], + "bbox": [ + 633, + 199, + 697, + 250 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/573a49f44d4708baba1104170e127f775fa5499d3154dd79b680b4272da71135.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 699, + 191, + 763, + 250 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/10ebc2477a30eb167a1a61be785ae2fc87fc190d81bbab877347a8daf41092ed.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 764, + 191, + 828, + 250 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/9133559835d3ddca4df6e59957ddaad9a07415d27411d881bcdf6ebbb41c1f19.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 828, + 191, + 893, + 250 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/709276f49ff6be0a6508a8f0d021aa8daa76f8f95b7ec0f644ddea61a57e15ac.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 531, + 253, + 552, + 267 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "George spreads the thick, peanutty paste, making a yummy snack.", + "bbox": [ + 534, + 270, + 617, + 306 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/0c1a6a68b2ba9aa8a2cb86e28410f151e6a9743ba58489eeb286f11af4ec6b36.jpg", + "image_caption": [ + "REC" + ], + "image_footnote": [], + "bbox": [ + 633, + 265, + 697, + 313 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/b7d19921d1fe7dd12f36779294f4c13045e76750383938f4cecad3a39730bf4f.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 699, + 257, + 763, + 314 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/a2f8f9a16452ed4ac767358618255562d029d7f8da9e1521a0722e6c68944bfd.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 764, + 257, + 828, + 314 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/e772131448626a9c7e4d15c5862b39644a727aaf1dcd316b35ace68f53b944a3.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 828, + 257, + 893, + 314 + ], + "page_idx": 7 + }, + { + "type": "table", + "img_path": "images/6f69fee1111bd205545916cc3f6dd3aae1f8c0817cd7818b0f18285d3530455a.jpg", + "table_caption": [ + "Table 8. Ablation study on video representation modeling with LLMs for generation. We evaluate text-to-video generation on MSR-VTT." + ], + "table_footnote": [], + "table_body": "
RepresentationObjectiveMechanism
patch-position dependentpatch-position independentMSEDiffusionGMMARQuery
ε-predv-predcausalbidirectional
CLIPSIM (↑)0.31920.32650.31680.28110.28420.32650.23860.30800.3265
FVD (↓)378.50366.60438.94418.19377.17366.60447.88416.60366.60
", + "bbox": [ + 151, + 400, + 849, + 491 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Q2: Which training objective is more suitable? As introduce in Sec. 3, we explore MSE regression, Diffusion modeling and GMM modeling to fit high-level continuous features with a LLM. As listed in Tab. 7, simply aligning the LLM output with video features using MSE loss yields the lowest generation quality, suggesting that deterministic regression is inadequate for modeling spatiotemporal representations. Training a denoising network to denoise the noisy video features by taking the LLM output as the condition also achieves inferior performance with both $\\epsilon$ prediction and $v$ prediction. Different from MAR [33] that denoises low-level VAE latents, our goal is to denoise high-level video features. We speculate that these features are more sensitive to Gaussian noise, making them more challenging to denoise. Training the LLM to model the distribution of high-level video features using a GMM model achieves the best generation quality and semantic alignment with captions.", + "bbox": [ + 88, + 516, + 485, + 773 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Q3: Which LLM mechanism is more effective? We train the LLM to fit the video representations with GMM modeling using both an autoregressive approach and a query-based approach, with the latter exploring causal attention and bidirectional attention within the LLM. Predicting the features of each video token in an autoregressive manner results in the worst performance due to the accumulation of errors, particularly when the features of the pervious to-", + "bbox": [ + 88, + 779, + 485, + 901 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "ken are sampled from a GMM distribution for predicting the distribution of the current token. The query-based approach achieves better results with bidirectional attention, as it enables each query to attend to all tokens for predictions.", + "bbox": [ + 511, + 516, + 908, + 578 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "5. Conclusion", + "text_level": 1, + "bbox": [ + 511, + 599, + 633, + 616 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "In this work, we introduce Divot, a diffusion-powered video tokenizer learned in a self-supervised manner for unified comprehension and generation. We further investigate effective approaches for modeling continuous video representations with the LLM and present Divot-LLM to understand and generate video content in a single framework. Divot-LLM achieves competitive performance in video comprehension and generation benchmarks, and enables video storytelling effectively. We hope our work will draw increased attention to unifying video comprehension and generation through the design of sophisticated tokenizers.", + "bbox": [ + 511, + 628, + 908, + 796 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Limitation. As we primarily focus on exploring effective representations and approaches for video generation with a unified LLM, the current model is trained to predict video representations for only a single clip and does not generate longer videos, which will be explored in our future work.", + "bbox": [ + 511, + 825, + 908, + 902 + ], + "page_idx": 7 + }, + { + "type": "page_number", + "text": "13613", + "bbox": [ + 480, + 944, + 517, + 955 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "References", + "text_level": 1, + "bbox": [ + 91, + 89, + 187, + 104 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[1] Jean-Baptiste Alayrac, Jeff Donahue, Pauline Luc, Antoine Miech, Iain Barr, Yana Hasson, Karel Lenc, Arthur Mensch, Katherine Millican, Malcolm Reynolds, et al. Flamingo: a visual language model for few-shot learning. Advances in Neural Information Processing Systems, 35:23716-23736, 2022. 2, 4", + "[2] Max Bain, Arsha Nagrani, Gül Varol, and Andrew Zisserman. Frozen in time: A joint video and image encoder for end-to-end retrieval. In Proceedings of the IEEE/CVF international conference on computer vision, pages 1728-1738, 2021. 4, 5, 6", + "[3] Dmitry Baranchuk, Ivan Rubachev, Andrey Voynov, Valentin Khrulkov, and Artem Babenko. Label-efficient semantic segmentation with diffusion models. arXiv preprint arXiv:2112.03126, 2021. 3", + "[4] Andreas Blattmann, Tim Dockhorn, Sumith Kulal, Daniel Mendelevitch, Maciej Kilian, Dominik Lorenz, Yam Levi, Zion English, Vikram Voleti, Adam Letts, et al. Stable video diffusion: Scaling latent video diffusion models to large datasets. arXiv preprint arXiv:2311.15127, 2023. 7", + "[5] Andreas Blattmann, Robin Rombach, Huan Ling, Tim Dockhorn, Seung Wook Kim, Sanja Fidler, and Karsten Kreis. Align your latents: High-resolution video synthesis with latent diffusion models. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 22563-22575, 2023. 7", + "[6] Tom Brown, Benjamin Mann, Nick Ryder, Melanie Subbiah, Jared D Kaplan, Prafulla Dhariwal, Arvind Neelakantan, Pranav Shyam, Girish Sastry, Amanda Askell, et al. Language models are few-shot learners. Advances in neural information processing systems, 33:1877-1901, 2020. 1", + "[7] David Chen and William B Dolan. Collecting highly parallel data for paraphrase evaluation. In Proceedings of the 49th annual meeting of the association for computational linguistics: human language technologies, pages 190-200, 2011. 6", + "[8] Haoxin Chen, Menghan Xia, Yingqing He, Yong Zhang, Xiaodong Cun, Shaoshu Yang, Jinbo Xing, Yaofang Liu, Qifeng Chen, Xintao Wang, et al. Videocrafter1: Open diffusion models for high-quality video generation. arXiv preprint arXiv:2310.19512, 2023. 4", + "[9] Tsai-Shien Chen, Aliaksandr Siarohin, Willi Menapace, Ekaterina Deyneka, Hsiang-wei Chao, Byung Eun Jeon, Yuwei Fang, Hsin-Ying Lee, Jian Ren, Ming-Hsuan Yang, et al. Panda-70m: Captioning 70m videos with multiple cross-modality teachers. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 13320–13331, 2024. 4, 5", + "[10] Zesen Cheng, Sicong Leng, Hang Zhang, Yifei Xin, Xin Li, Guanzheng Chen, Yongxin Zhu, Wenqi Zhang, Ziyang Luo, Deli Zhao, et al. Videollama 2: Advancing spatial-temporal modeling and audio understanding in video-llms. arXiv preprint arXiv:2406.07476, 2024. 6", + "[11] Aakanksha Chowdhery, Sharan Narang, Jacob Devlin, Maarten Bosma, Gaurav Mishra, Adam Roberts, Paul Barham, Hyung Won Chung, Charles Sutton, Sebastian Gehrmann," + ], + "bbox": [ + 93, + 114, + 485, + 900 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "et al. Palm: Scaling language modeling with pathways. arXiv preprint arXiv:2204.02311, 2022. 1", + "[12] Runpei Dong, Chunrui Han, Yuang Peng, Zekun Qi, Zheng Ge, Jinrong Yang, Liang Zhao, Jianjian Sun, Hongyu Zhou, Haoran Wei, et al. Dreamllm: Synergistic multimodal comprehension and creation. arXiv preprint arXiv:2309.11499, 2023. 1, 2", + "[13] Alexey Dosovitskiy, Lucas Beyer, Alexander Kolesnikov, Dirk Weissenborn, Xiaohua Zhai, Thomas Unterthiner, Mostafa Dehghani, Matthias Minderer, Georg Heigold, Sylvain Gelly, et al. An image is worth 16x16 words: Transformers for image recognition at scale. arXiv preprint arXiv:2010.11929, 2020. 2", + "[14] Songwei Ge, Seungjun Nah, Guilin Liu, Tyler Poon, Andrew Tao, Bryan Catanzaro, David Jacobs, Jia-Bin Huang, Ming-Yu Liu, and Yogesh Balaji. Preserve your own correlation: A noise prior for video diffusion models. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 22930-22941, 2023. 7", + "[15] Yuying Ge, Yixiao Ge, Ziyun Zeng, Xintao Wang, and Ying Shan. Planting a seed of vision in large language model. arXiv preprint arXiv:2307.08041, 2023. 1, 2", + "[16] Yuying Ge, Sijie Zhao, Ziyun Zeng, Yixiao Ge, Chen Li, Xintao Wang, and Ying Shan. Making llama see and draw with seed tokenizer. arXiv preprint arXiv:2310.01218, 2023.", + "[17] Yuying Ge, Sijie Zhao, Jinguo Zhu, Yixiao Ge, Kun Yi, Lin Song, Chen Li, Xiaohan Ding, and Ying Shan. Seed-x: Multimodal models with unified multi-granularity comprehension and generation. arXiv preprint arXiv:2404.14396, 2024. 1, 2, 4", + "[18] Raghav Goyal, Samira Ebrahimi Kahou, Vincent Michalski, Joanna Materzynska, Susanne Westphal, Heuna Kim, Valentin Haenel, Ingo Fruend, Peter Yianilos, Moritz Mueller-Freitag, et al. The\" something something\" video database for learning and evaluating visual common sense. In Proceedings of the IEEE international conference on computer vision, pages 5842-5850, 2017. 5", + "[19] Kristen Grauman, Andrew Westbury, Eugene Byrne, Zachary Chavis, Antonino Furnari, Rohit Girdhar, Jackson Hamburger, Hao Jiang, Miao Liu, Xingyu Liu, et al. Ego4d: Around the world in 3,000 hours of egocentric video. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 18995-19012, 2022. 5", + "[20] Jonathan Ho, Ajay Jain, and Pieter Abbeel. Denoising diffusion probabilistic models. Advances in neural information processing systems, 33:6840-6851, 2020. 3", + "[21] Wenyi Hong, Ming Ding, Wendi Zheng, Xinghan Liu, and Jie Tang. Cogvideo: Large-scale pretraining for text-to-video generation via transformers. arXiv preprint arXiv:2205.15868, 2022. 7", + "[22] Edward J Hu, Yelong Shen, Phillip Wallis, Zeyuan Allen-Zhu, Yanzhi Li, Shean Wang, Lu Wang, and Weizhu Chen. Lora: Low-rank adaptation of large language models. arXiv preprint arXiv:2106.09685, 2021. 6", + "[23] Drew A Hudson, Daniel Zoran, Mateusz Malinowski, Andrew K Lampinen, Andrew Jaegle, James L McClelland, Loic Matthew, Felix Hill, and Alexander Lerchner. Soda: Bottle-" + ], + "bbox": [ + 516, + 92, + 906, + 900 + ], + "page_idx": 8 + }, + { + "type": "page_number", + "text": "13614", + "bbox": [ + 480, + 945, + 519, + 955 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "neck diffusion models for representation learning. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 23115-23127, 2024. 3", + "[24] Albert Q Jiang, Alexandre Sablayrolles, Arthur Mensch, Chris Bamford, Devendra Singh Chaplot, Diego de las Casas, Florian Bressand, Gianna Lengyel, Guillaume Lample, Lucile Saulnier, et al. Mistral 7b. arXiv preprint arXiv:2310.06825, 2023. 2, 5", + "[25] Yang Jin, Kun Xu, Liwei Chen, Chao Liao, Jianchao Tan, Bin Chen, Chenyi Lei, An Liu, Chengru Song, Xiaogiang Lei, et al. Unified language-vision pretraining with dynamic discrete visual tokenization. arXiv preprint arXiv:2309.04669, 2023. 2", + "[26] Yang Jin, Zhicheng Sun, Kun Xu, Liwei Chen, Hao Jiang, Quzhe Huang, Chengru Song, Yuliang Liu, Di Zhang, Yang Song, et al. Video-lavit: Unified video-language pre-training with decoupled visual-motional tokenization. arXiv preprint arXiv:2402.03161, 2024. 1, 2, 3, 5, 6, 7", + "[27] Will Kay, Joao Carreira, Karen Simonyan, Brian Zhang, Chloe Hillier, Sudheendra Vijayanarasimhan, Fabio Viola, Tim Green, Trevor Back, Paul Natsev, et al. The kinetics human action video dataset. arXiv preprint arXiv:1705.06950, 2017. 5", + "[28] Diederik P Kingma and Max Welling. Auto-encoding variational bayes. arXiv preprint arXiv:1312.6114, 2013. 2, 4", + "[29] Dan Kondratyuk, Lijun Yu, Xiuye Gu, José Lezama, Jonathan Huang, Grant Schindler, Rachel Hornung, Vighnesh Birodkar, Jimmy Yan, Ming-Chang Chiu, et al. Videopoet: A large language model for zero-shot video generation. arXiv preprint arXiv:2312.14125, 2023. 2, 7", + "[30] Bo Li, Yuanhan Zhang, Dong Guo, Renrui Zhang, Feng Li, Hao Zhang, Kaichen Zhang, Yanwei Li, Ziwei Liu, and Chunyuan Li. Llava-onevision: Easy visual task transfer. arXiv preprint arXiv:2408.03326, 2024. 6", + "[31] Kunchang Li, Yali Wang, Yinan He, Yizhuo Li, Yi Wang, Yi Liu, Zun Wang, Jilan Xu, Guo Chen, Ping Luo, et al. Mybench: A comprehensive multi-modal video understanding benchmark. arXiv preprint arXiv:2311.17005, 2023. 6", + "[32] Kunchang Li, Yali Wang, Yinan He, Yizhuo Li, Yi Wang, Yi Liu, Zun Wang, Jilan Xu, Guo Chen, Ping Luo, et al. Mybench: A comprehensive multi-modal video understanding benchmark. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 22195-22206, 2024. 6", + "[33] Tianhong Li, Yonglong Tian, He Li, Mingyang Deng, and Kaiming He. Autoregressive image generation without vector quantization. arXiv preprint arXiv:2406.11838, 2024. 2, 4, 5, 8", + "[34] Yuncheng Li, Yale Song, Liangliang Cao, Joel Tetreault, Larry Goldberg, Alejandro Jaimes, and Jiebo Luo. Tgif: A new dataset and benchmark on animated gif description. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition, pages 4641-4650, 2016. 5", + "[35] Yanwei Li, Chengyao Wang, and Jiaya Jia. Llama-vid: An image is worth 2 tokens in large language models. arXiv preprint arXiv:2311.17043, 2023. 6" + ], + "bbox": [ + 91, + 90, + 483, + 900 + ], + "page_idx": 9 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[36] Yanwei Li, Yuechen Zhang, Chengyao Wang, Zhisheng Zhong, Yixin Chen, Ruihang Chu, Shaoteng Liu, and Jiaya Jia. Mini-gemini: Mining the potential of multi-modality vision language models. arXiv preprint arXiv:2403.18814, 2024. 2", + "[37] Bin Lin, Bin Zhu, Yang Ye, Munan Ning, Peng Jin, and Li Yuan. Video-llava: Learning united visual representation by alignment before projection. arXiv preprint arXiv:2311.10122, 2023. 5, 6", + "[38] Haotian Liu, Chunyuan Li, Yuheng Li, Bo Li, Yuanhan Zhang, Sheng Shen, and Yong Jae Lee. Llava-last: Improved reasoning,OCR,and world knowledge,2024.6", + "[39] Haotian Liu, Chunyuan Li, Qingyang Wu, and Yong Jae Lee. Visual instruction tuning. Advances in neural information processing systems, 36, 2024. 5", + "[40] Hao Liu, Wilson Yan, Matei Zaharia, and Pieter Abbeel. World model on million-length video and language with ringattention. arXiv preprint arXiv:2402.08268, 2024. 1, 2, 6", + "[41] Jiasen Lu, Christopher Clark, Sangho Lee, Zichen Zhang, Savya Khosla, Ryan Marten, Derek Hoiem, and Aniruddha Kembhavi. Unified-io 2: Scaling autoregressive multimodal models with vision, language, audio, and action. arXiv preprint arXiv:2312.17172, 2023. 2", + "[42] Ruipu Luo, Ziwang Zhao, Min Yang, Junwei Dong, Da Li, Pengcheng Lu, Tao Wang, Linmei Hu, Minghui Qiu, and Zhongyu Wei. Valley: Video assistant with large language model enhanced ability. arXiv preprint arXiv:2306.07207, 2023.5,7", + "[43] Muhammad Maaz, Hanoona Rasheed, Salman Khan, and Fahad Shahbaz Khan. Video-chatgpt: Towards detailed video understanding via large vision and language models. arXiv preprint arXiv:2306.05424, 2023. 5, 6, 7", + "[44] Karttikeya Mangalam, Raiymbek Akshulakov, and Jitendra Malik. Egoschema: A diagnostic benchmark for very long-form video language understanding. Advances in Neural Information Processing Systems, 36:46212-46244, 2023. 6", + "[45] Willi Menapace, Aliaksandr Siarohin, Ivan Skorokhodov, Ekaterina Deyneka, Tsai-Shien Chen, Anil Kag, Yuwei Fang, Aleksei Stoliar, Elisa Ricci, Jian Ren, et al. Snap video: Scaled spatiotemporal transformers for text-to-video synthesis. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 7038-7048, 2024. 7", + "[46] OpenAI. Gpt-4v(ison) system card, 2023. 6", + "[47] OpenAI. Gpt-4o system card, 2024. 6", + "[48] Viorica Patraucean, Lucas Smaira, Ankush Gupta, Adria Recasens, Larisa Markeeva, Dylan Banarse, Skanda Koppula, Mateusz Malinowski, Yi Yang, Carl Doersch, et al. Perception test: A diagnostic benchmark for multimodal video models. Advances in Neural Information Processing Systems, 36, 2024. 5, 6", + "[49] Alec Radford, Jong Wook Kim, Chris Hallacy, Aditya Ramesh, Gabriel Goh, Sandhini Agarwal, Girish Sastry, Amanda Askell, Pamela Mishkin, Jack Clark, et al. Learning transferable visual models from natural language supervision. In International conference on machine learning, pages 8748-8763. PMLR, 2021. 5" + ], + "bbox": [ + 516, + 92, + 906, + 898 + ], + "page_idx": 9 + }, + { + "type": "page_number", + "text": "13615", + "bbox": [ + 480, + 944, + 517, + 955 + ], + "page_idx": 9 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[50] Robin Rombach, Andreas Blattmann, Dominik Lorenz, Patrick Esser, and Björn Ommer. High-resolution image synthesis with latent diffusion models. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 10684-10695, 2022. 2, 3", + "[51] Christoph Schuhmann, Romain Beaumont, Richard Vencu, Cade Gordon, Ross Wightman, Mehdi Cherti, Theo Coombes, Aarush Katta, Clayton Mullis, Mitchell Wortsman, et al. Laion-5b: An open large-scale dataset for training next generation image-text models. Advances in Neural Information Processing Systems, 35:25278–25294, 2022. 5", + "[52] Piyush Sharma, Nan Ding, Sebastian Goodman, and Radu Soricut. Conceptual captions: A cleaned, hypernymed, image alt-text dataset for automatic image captioning. In Proceedings of the 56th Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers), pages 2556-2565, 2018. 5", + "[53] Uriel Singer, Adam Polyak, Thomas Hayes, Xi Yin, Jie An, Songyang Zhang, Qiyuan Hu, Harry Yang, Oron Ashual, Oran Gafni, et al. Make-a-video: Text-to-video generation without text-video data. arXiv preprint arXiv:2209.14792, 2022. 7", + "[54] Quan Sun, Yufeng Cui, Xiaosong Zhang, Fan Zhang, Qiying Yu, Zhengxiong Luo, Yueze Wang, Yongming Rao, Jingjing Liu, Tiejun Huang, et al. Generative multimodal models are in-context learners. arXiv preprint arXiv:2312.13286, 2023. 2, 4", + "[55] Quan Sun, Qiying Yu, Yufeng Cui, Fan Zhang, Xiaosong Zhang, Yueze Wang, Hongcheng Gao, Jingjing Liu, Tiejun Huang, and Xinlong Wang. Generative pretraining in multimodality. arXiv preprint arXiv:2307.05222, 2023. 1, 2", + "[56] Quan Sun, Qiying Yu, Yufeng Cui, Fan Zhang, Xiaosong Zhang, Yueze Wang, Hongcheng Gao, Jingjing Liu, Tiejun Huang, and Xinlong Wang. Generative pretraining in multimodality. arXiv preprint arXiv:2307.05222, 2023.", + "[57] Chameleon Team. Chameleon: Mixed-modal early-fusion foundation models. arXiv preprint arXiv:2405.09818, 2024. 1, 2", + "[58] Gemini Team, Rohan Anil, Sebastian Borgeaud, Yonghui Wu, Jean-Baptiste Alayrac, Jiahui Yu, Radu Soricut, Johan Schalkwyk, Andrew M Dai, Anja Hauth, et al. Gemini: a family of highly capable multimodal models. arXiv preprint arXiv:2312.11805, 2023. 6", + "[59] Gemini Team, Petko Georgiev, Ving Ian Lei, Ryan Burnell, Libin Bai, Anmol Gulati, Garrett Tanzer, Damien Vincent, Zhufeng Pan, Shibo Wang, et al. Gemini 1.5: Unlocking multimodal understanding across millions of tokens of context. arXiv preprint arXiv:2403.05530, 2024. 6", + "[60] Hugo Touvron, Thibaut Lavril, Gautier Izacard, Xavier Martinet, Marie-Anne Lachaux, Timothée Lacroix, Baptiste Rozière, Naman Goyal, Eric Hambro, Faisal Azhar, et al. Llama: Open and efficient foundation language models. arXiv preprint arXiv:2302.13971, 2023. 1", + "[61] Michael Tschannen, Cian Eastwood, and Fabian Mentzer. Givt: Generative infinite-vocabulary transformers. In European Conference on Computer Vision, pages 292-309. Springer, 2025. 5, 7" + ], + "bbox": [ + 91, + 90, + 485, + 900 + ], + "page_idx": 10 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[62] Thomas Unterthiner, Sjoerd Van Steenkiste, Karol Kurach, Raphael Marinier, Marcin Michalski, and Sylvain Gelly. Towards accurate generative models of video: A new metric & challenges. arXiv preprint arXiv:1812.01717, 2018. 5, 6", + "[63] Aaron Van Den Oord, Oriol Vinyals, et al. Neural discrete representation learning. Advances in neural information processing systems, 30, 2017. 2", + "[64] Hanyu Wang, Saksham Suri, Yixuan Ren, Hao Chen, and Abhinav Shrivastava. Larp: Tokenizing videos with a learned autoregressive generative prior. arXiv preprint arXiv:2410.21264, 2024. 7", + "[65] Wenxuan Wang, Quan Sun, Fan Zhang, Yepeng Tang, Jing Liu, and Xinlong Wang. Diffusion feedback helps clip see better. arXiv preprint arXiv:2407.20171, 2024. 3", + "[66] Xiang Wang, Hangjie Yuan, Shiwei Zhang, Dayou Chen, Jiuniu Wang, Yingya Zhang, Yujun Shen, Deli Zhao, and Jingren Zhou. Videocomposer: Compositional video synthesis with motion controllability. Advances in Neural Information Processing Systems, 36, 2024. 7", + "[67] Xinlong Wang, Xiaosong Zhang, Zhengxiong Luo, Quan Sun, Yufeng Cui, Jinsheng Wang, Fan Zhang, Yueze Wang, Zhen Li, Qiying Yu, et al. Emu3: Next-token prediction is all you need. arXiv preprint arXiv:2409.18869, 2024. 2, 3", + "[68] Yi Wang, Yinan He, Yizhuo Li, Kunchang Li, Jiashuo Yu, Xin Ma, Xinhao Li, Guo Chen, Xinyuan Chen, Yaohui Wang, et al. Internvid: A large-scale video-text dataset for multimodal understanding and generation. arXiv preprint arXiv:2307.06942, 2023. 7", + "[69] Yuqing Wang, Tianwei Xiong, Daquan Zhou, Zhijie Lin, Yang Zhao, Bingyi Kang, Jiashi Feng, and Xihui Liu. Loong: Generating minute-level long videos with autoregressive language models. arXiv preprint arXiv:2410.02757, 2024. 2, 6, 7", + "[70] Chen Wei, Karttikeya Mangalam, Po-Yao Huang, Yanghao Li, Haoqi Fan, Hu Xu, Huiyu Wang, Cihang Xie, Alan Yuille, and Christoph Feichtenhofer. Diffusion models as masked autoencoders. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 16284-16294, 2023. 3", + "[71] Chenfei Wu, Lun Huang, Qianxi Zhang, Binyang Li, Lei Ji, Fan Yang, Guillermo Sapiro, and Nan Duan. Godiva: Generating open-domain videos from natural descriptions. arXiv preprint arXiv:2104.14806, 2021. 6", + "[72] Chengyue Wu, Xiaokang Chen, Zhiyu Wu, Yiyang Ma, Xingchao Liu, Zizheng Pan, Wen Liu, Zhenda Xie, Xingkai Yu, Chong Ruan, et al. Janus: Decoupling visual encoding for unified multimodal understanding and generation. arXiv preprint arXiv:2410.13848, 2024. 2", + "[73] Shengqiong Wu, Hao Fei, Leigang Qu, Wei Ji, and Tat-Seng Chua. Next-gpt: Any-to-any multimodal llm. arXiv preprint arXiv:2309.05519, 2023.", + "[74] Yecheng Wu, Zhuoyang Zhang, Junyu Chen, Haotian Tang, Dacheng Li, Yunhao Fang, Ligeng Zhu, Enze Xie, Hongxu Yin, Li Yi, et al. Vila-u: a unified foundation model integrating visual understanding and generation. arXiv preprint arXiv:2409.04429, 2024. 1, 2, 6, 7", + "[75] Weilai Xiang, Hongyu Yang, Di Huang, and Yunhong Wang. Denoising diffusion autoencoders are unified self-supervised" + ], + "bbox": [ + 516, + 90, + 906, + 900 + ], + "page_idx": 10 + }, + { + "type": "page_number", + "text": "13616", + "bbox": [ + 480, + 945, + 519, + 955 + ], + "page_idx": 10 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "learners. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 15802-15812, 2023. 3", + "[76] Junbin Xiao, Xindi Shang, Angela Yao, and Tat-Seng Chua. Next-qa: Next phase of question-answering to explaining temporal actions. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 9777–9786, 2021. 5", + "[77] Jinheng Xie, Weijia Mao, Zechen Bai, David Junhao Zhang, Weihao Wang, Kevin Qinghong Lin, Yuchao Gu, Zhijie Chen, Zhenheng Yang, and Mike Zheng Shou. Show-o: One single transformer to unify multimodal understanding and generation. arXiv preprint arXiv:2408.12528, 2024. 1, 2, 3", + "[78] Jinbo Xing, Menghan Xia, Yong Zhang, Haoxin Chen, Wangbo Yu, Hanyuan Liu, Gongye Liu, Xintao Wang, Ying Shan, and Tien-Tsin Wong. Dynamiccafter: Animating open-domain images with video diffusion priors. In European Conference on Computer Vision, pages 399-417. Springer, 2025. 2, 4", + "[79] Jun Xu, Tao Mei, Ting Yao, and Yong Rui. Msr-vtt: A large video description dataset for bridging video and language. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 5288-5296, 2016. 5, 6", + "[80] Jiarui Xu, Sifei Liu, Arash Vahdat, Wonmin Byeon, Xiaolong Wang, and Shalini De Mello. Open-vocabulary panoptic segmentation with text-to-image diffusion models. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 2955-2966, 2023. 3", + "[81] Lin Xu, Yilin Zhao, Daquan Zhou, Zhijie Lin, See Kiong Ng, and Jiashi Feng. Pllava: Parameter-free llava extension from images to videos for video dense captioning. arXiv preprint arXiv:2404.16994, 2024. 6", + "[82] Wilson Yan, Yunzhi Zhang, Pieter Abbeel, and Aravind Srinivas. Videogpt: Video generation using vq-vae and transformers. arXiv preprint arXiv:2104.10157, 2021. 5", + "[83] Jian Yang, Dacheng Yin, Yizhou Zhou, Fengyun Rao, Wei Zhai, Yang Cao, and Zheng-Jun Zha. Mmar: Towards lossless multi-modal auto-regressive probabilistic modeling. arXiv preprint arXiv:2410.10798, 2024. 2", + "[84] Shuai Yang, Yuying Ge, Yang Li, Yukang Chen, Yixiao Ge, Ying Shan, and Yingcong Chen. Seed-story: Multimodal long story generation with large language model. arXiv preprint arXiv:2407.08683, 2024. 2", + "[85] Kexin Yi, Chuang Gan, Yunzhu Li, Pushmeet Kohli, Jiajun Wu, Antonio Torralba, and Joshua B Tenenbaum. Clevrer: Collision events for video representation and reasoning. arXiv preprint arXiv:1910.01442, 2019. 5", + "[86] Lili Yu, Bowen Shi, Ramakanth Pasunuru, Benjamin Muller, Olga Golovneva, Tianlu Wang, Arun Babu, Binh Tang, Brian Karrer, Shelly Sheynin, et al. Scaling autoregressive multimodal models: Pretraining and instruction tuning. arXiv preprint arXiv:2309.02591, 2023. 1, 2", + "[87] Qiying Yu, Quan Sun, Xiaosong Zhang, Yufeng Cui, Fan Zhang, Yue Cao, Xinlong Wang, and Jingjing Liu. Capsfusion: Rethinking image-text data at scale. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 14022-14032, 2024. 5", + "[88] Zhou Yu, Dejing Xu, Jun Yu, Ting Yu, Zhou Zhao, Yueting Zhuang, and Dacheng Tao. Activitynet-qa: A dataset for" + ], + "bbox": [ + 91, + 90, + 483, + 900 + ], + "page_idx": 11 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "understanding complex web videos via question answering. In Proceedings of the AAAI Conference on Artificial Intelligence, pages 9127-9134, 2019. 5, 6", + "[89] Chuyang Zhao, Yuxing Song, Wenhao Wang, Haocheng Feng, Errui Ding, Yifan Sun, Xinyan Xiao, and Jingdong Wang. Monoformer: One transformer for both diffusion and autoregression. arXiv preprint arXiv:2409.16280, 2024. 2", + "[90] Wenliang Zhao, Yongming Rao, Zuyan Liu, Benlin Liu, Jie Zhou, and Jiwen Lu. Unleashing text-to-image diffusion models for visual perception. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 5729-5739, 2023. 3", + "[91] Chunting Zhou, Lili Yu, Arun Babu, Kushal Tirumala, Michihiro Yasunaga, Leonid Shamis, Jacob Kahn, Xuezhe Ma, Luke Zettlemoyer, and Omer Levy. Transfusion: Predict the next token and diffuse images with one multi-modal model. arXiv preprint arXiv:2408.11039, 2024. 1, 2", + "[92] Luowei Zhou, Chenliang Xu, and Jason Corso. Towards automatic learning of procedures from web instructional videos. In Proceedings of the AAAI Conference on Artificial Intelligence, 2018. 5", + "[93] Jinguo Zhu, Xiaohan Ding, Yixiao Ge, Yuying Ge, Sijie Zhao, Hengshuang Zhao, Xiaohua Wang, and Ying Shan. Vl-gpt: A generative pre-trained transformer for vision and language understanding and generation. arXiv preprint arXiv:2312.09251, 2023. 1, 2" + ], + "bbox": [ + 516, + 92, + 906, + 457 + ], + "page_idx": 11 + }, + { + "type": "page_number", + "text": "13617", + "bbox": [ + 480, + 944, + 517, + 955 + ], + "page_idx": 11 + } +] \ No newline at end of file diff --git a/2025/Divot_ Diffusion Powers Video Tokenizer for Comprehension and Generation/dd9cf0f7-0275-41db-a729-40fdeab212b3_model.json b/2025/Divot_ Diffusion Powers Video Tokenizer for Comprehension and Generation/dd9cf0f7-0275-41db-a729-40fdeab212b3_model.json new file mode 100644 index 0000000000000000000000000000000000000000..206c4c621868355ef3ddbb5b4989afa841ac17a9 --- /dev/null +++ b/2025/Divot_ Diffusion Powers Video Tokenizer for Comprehension and Generation/dd9cf0f7-0275-41db-a729-40fdeab212b3_model.json @@ -0,0 +1,2963 @@ +[ + [ + { + "type": "header", + "bbox": [ + 0.107, + 0.003, + 0.182, + 0.043 + ], + "angle": 0, + "content": "CVF" + }, + { + "type": "header", + "bbox": [ + 0.237, + 0.001, + 0.808, + 0.045 + ], + "angle": 0, + "content": "This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore." + }, + { + "type": "title", + "bbox": [ + 0.111, + 0.131, + 0.887, + 0.152 + ], + "angle": 0, + "content": "Divot: Diffusion Powers Video Tokenizer for Comprehension and Generation" + }, + { + "type": "text", + "bbox": [ + 0.261, + 0.181, + 0.737, + 0.234 + ], + "angle": 0, + "content": "Yuying Ge Yizhuo Li Yixiao Ge Ying Shan ARC Lab, Tencent PCG https://github.com/TencentARC/Divot" + }, + { + "type": "title", + "bbox": [ + 0.249, + 0.269, + 0.327, + 0.285 + ], + "angle": 0, + "content": "Abstract" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.302, + 0.486, + 0.739 + ], + "angle": 0, + "content": "In recent years, there has been a significant surge of interest in unifying image comprehension and generation within Large Language Models (LLMs). This growing interest has prompted us to explore extending this unification to videos. The core challenge lies in developing a versatile video tokenizer that captures both the spatial characteristics and temporal dynamics of videos to obtain representations for LLMs, and the representations can be further decoded into realistic video clips to enable video generation. In this work, we introduce Divot, a Diffusion-Powered Video Tokenizer, which leverages the diffusion process for self-supervised video representation learning. We posit that if a video diffusion model can effectively de-noise video clips by taking the features of a video tokenizer as the condition, then the tokenizer has successfully captured robust spatial and temporal information. Additionally, the video diffusion model inherently functions as a de-tokenizer, decoding videos from their representations. Building upon the Divot tokenizer, we present Divot-LLM through video-to-text autoregression and text-to-video generation by modeling the distributions of continuous-valued Divot features with a Gaussian Mixture Model. Experimental results demonstrate that our diffusion-based video tokenizer, when integrated with a pre-trained LLM, achieves competitive performance across various video comprehension and generation benchmarks. The instruction tuned Divot-LLM also excels in video storytelling, generating interleaved narratives and corresponding videos. Models and codes are available at https://github.com/TencentARC/Divot." + }, + { + "type": "title", + "bbox": [ + 0.092, + 0.769, + 0.222, + 0.785 + ], + "angle": 0, + "content": "1. Introduction" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.795, + 0.485, + 0.9 + ], + "angle": 0, + "content": "In recent years, the rapid evolution of Multimodal Large Language Models (MLLMs) [12, 15-17, 55-57, 77, 86, 91, 93] has demonstrated significant progresses in unified image understanding and generation, which empowers LLMs [6, 11, 60] with the ability to generate images beyond texts. While these work primarily focus on image-text data, the extension of this unification to the video domain remains" + }, + { + "type": "image", + "bbox": [ + 0.521, + 0.269, + 0.906, + 0.391 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.512, + 0.404, + 0.908, + 0.489 + ], + "angle": 0, + "content": "Figure 1. We utilize the diffusion procedure to learn a video tokenizer in a self-supervised manner for unified comprehension and generation, where the spatiotemporal representations serve as the condition of a diffusion model to de-noise video clips. Additionally, the proxy diffusion model functions as a de-tokenizer to decode realistic video clips from the video representations." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.508, + 0.907, + 0.583 + ], + "angle": 0, + "content": "relatively under-explored. Achieving unified video comprehension and generation is essential for the development of more sophisticated artificial intelligence (AI) systems that are capable of understanding and creating dynamic visual content in the real world." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.584, + 0.909, + 0.885 + ], + "angle": 0, + "content": "The primary challenge of achieving unified video comprehension and generation lies in developing a versatile video tokenizer that can effectively address the complexities inherent in video data. This tokenizer should be able to obtain robust video representations that serve as inputs of MLLMs for video comprehension, and these representations can be further decoded into realistic video clips to enable video generation. Unlike static images, videos encompass both spatial characteristics and temporal dynamics, making their representation significantly more complex. Recent pioneering work [26, 40, 74] adopt a discrete video tokenizer for unifying video comprehension and generation, where a video is represented as a sequence of discrete frame tokens, or keyframe tokens followed by discrete motion tokens. This approach eases video generation with a LLM through an autoregressive next-token prediction mechanism, but sacrifices the performance of multimodal understanding, as pointed out by recent work [77]. In this work, we aim to investigate an alternative approach by utilizing continuous video representations to unify video comprehension and generation." + }, + { + "type": "text", + "bbox": [ + 0.534, + 0.886, + 0.907, + 0.9 + ], + "angle": 0, + "content": "To this end, we introduce Divot, a Diffusion-Powered" + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.946, + 0.519, + 0.957 + ], + "angle": 0, + "content": "13606" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.09, + 0.092, + 0.482, + 0.317 + ], + "angle": 0, + "content": "Video Tokenizer that leverages the diffusion process [50] for self-supervised video representation learning as shown in Fig. 1. The core premise is that if a diffusion model can effectively predict the noise added to the Variational Autoencoder (VAE) latents [28] of video clips, when conditioned on the features produced by the video tokenizer, it demonstrates that the tokenizer has successfully captured robust spatial and temporal information inherent in the video data. This capability is crucial for representing the intricate dynamics present in videos. Furthermore, in addition to being a proxy module for learning the tokenizer, the diffusion model can act as a de-tokenizer to effectively decode realistic videos from their learned representations. This dual functionality facilitates a seamless integration of understanding and creating video content within a LLM." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.32, + 0.483, + 0.515 + ], + "angle": 0, + "content": "Specifically, the Divot tokenizer is composed of a pretrained Vision Transformer (ViT) encoder [13], a Spatial-Temporal transformer, and a Perceiver Resampler [1] to obtain video representations from video frames sampled at low frame rate (fps) considering the semantic redundancy between adjacent frames. The video representations serve as the condition of a pre-trained video diffusion model, DynamiCrafter [78] (without the concatenation of a conditional image with initial noise), to predict the noise added to the VAE latents of video frames. After training, the video diffusion model can generate realistic video clips from noise by taking the video representations provided by the Divot tokenizer as the condition." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.517, + 0.483, + 0.851 + ], + "angle": 0, + "content": "We further present Divot-LLM by equipping the pretrained Mistral-7B [24] with the Divot tokenizer. Divot-LLM is pre-trained with a next-word prediction objective on video-caption data by taking the spatiotemporal representations as inputs for video comprehension. The challenge then arises in modeling the continuous video representations with the LLM for video generation. We empirically find that simply minimizing the distance between the LLM output and video representations using mean squared error (MSE) loss achieves unsatisfactory results, since the deterministic regression regularizes the LLM to learn overly averaged features of videos. To address this, inspired by recent work [33], we shift our focus from deterministic modeling to probabilistic modeling by modeling the distributions of video features with a Gaussian Mixture Model (GMM). Specifically, we train the LLM to predict GMM parameters, including means, variances, and mixture probabilities through minimizing the discrepancy between the predicted GMM distribution and the actual video representations using negative log-likelihood (NLL) loss. During inference, we draw samples from the predicted GMM distribution as the condition of the video de-tokenizer to decode video clips." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.092, + 0.908, + 0.212 + ], + "angle": 0, + "content": "We benchmark Divot-LLM on a broad range of video comprehension tasks and zero-shot video generation, achieving competitive performance through pre-training on 5 million video-text pairs. By leveraging the generality of the video tokenizer, our Divot-LLM also enables video storytelling, which generates interleaved narratives and corresponding videos that are temporally coherent through fine-tuning on a specific animation dataset." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.214, + 0.909, + 0.41 + ], + "angle": 0, + "content": "Our contributions are three-fold. (1) We introduce Divot, an advanced video tokenizer that leverages a diffusion procedure for self-supervised video representation learning, aiming to unify video comprehension and generation. (2) We present Divot-LLM, composed of a pre-trained LLM and the Divot tokenizer to enable understanding and generating video content within a single framework. We investigate effective approaches for fitting continuous video representations using the LLM with probabilistic modeling for video generation. (3) We conduct extensive experiments to demonstrate Divot-LLM's competitive performance on existing video comprehension and generation benchmarks, as well as video storytelling. All models and code are released." + }, + { + "type": "title", + "bbox": [ + 0.514, + 0.427, + 0.655, + 0.442 + ], + "angle": 0, + "content": "2. Related Work" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.453, + 0.909, + 0.786 + ], + "angle": 0, + "content": "MLLMs for Comprehension and Generation. With the rapid development of Multimodal Large Language Models (MLLM), recent studies have been working on unified MLLMs [12, 15-17, 25, 26, 36, 40, 41, 54, 55, 57, 67, 72-74, 77, 83, 84, 86, 89, 91, 93] that are capable of multimodal comprehension and generation. To empower LLMs with the capability to generate visual content, existing work primarily employs the following three approaches: (1) utilizing a pre-trained stable diffusion model to generate images conditioned on LLM output (either continuous features or discrete tokens); (2) employing a Vector Quantized (VQ) [63] based decoder to generate visual content from the discrete codes predicted by LLMs; (3) using LLMs to de-noise Gaussian noise through a diffusion process. While most work predominantly focus on the unification of images and texts, some pioneering studies [26, 40, 67, 74] further advance the integration of video comprehension and generation within an LLM through generating videos from discrete codes using a VQ-based decoder, which falls into the second approach. In this work, we adopt the first approach, which involves leveraging a diffusion model to achieve unified video understanding and generation from continuous representations." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.81, + 0.909, + 0.903 + ], + "angle": 0, + "content": "Video Tokenizer in MLLMs. Previous work on video generation with LLMs predominantly employs a discrete video tokenizer to convert video signals into a sequence of quantized tokens. For example, LWM [40] and VILA-U [74] utilize a frame-level tokenizer to discretize each frame into a sequence of codes. VideoPoet [29], Loong [69] and" + }, + { + "type": "page_footnote", + "bbox": [ + 0.091, + 0.863, + 0.483, + 0.901 + ], + "angle": 0, + "content": "1We do not explore more advanced LLMs because we want to ensure that our superiority stems from the improved visual representations, rather than from the capabilities of a more sophisticated foundation model." + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "13607" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.123, + 0.089, + 0.67, + 0.245 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.685, + 0.09, + 0.877, + 0.24 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.089, + 0.253, + 0.908, + 0.311 + ], + "angle": 0, + "content": "Figure 2. Overview of Divot tokenization and de-tokenization. During training, sparsely sampled video frames are fed into the tokenizer to obtain spatiotemporal representations. These representations serve as the conditions for a U-Net, which is trained to de-noise the noisy VAE latents of densely sampled video frames. During inference, the video representations from the Divot tokenizer can be decoded into realistic video clips with the U-Net." + }, + { + "type": "text", + "bbox": [ + 0.093, + 0.329, + 0.487, + 0.542 + ], + "angle": 0, + "content": "Emu3 [67] leverage a 3D CNN architecture, where the encoded spatial-temporal features are quantized into discrete tokens. Video-LaVIT [26] represents video clips as a keyframe followed by extracted motion vectors, obtaining the respective discrete codes. By converting continuous visual signals into discrete tokens, the original next-token prediction mechanism can be adopted to facilitate video generation with an LLM. However, recent work [77] observes a significant performance degradation in multimodal comprehension tasks when discrete representations are used instead of continuous representations. In this work, we introduce a video tokenizer with continuous representations through leveraging the diffusion [50] procedure, enabling it to be effectively integrated with a LLM for unified comprehension and generation." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.564, + 0.485, + 0.807 + ], + "angle": 0, + "content": "Diffusion for Representation Learning. The diffusion process has been explored as a criterion for representation learning. Some works [3, 75, 80, 90] leverage the intermediate activations of pre-trained diffusion models for downstream tasks including classification, segmentation and depth estimation. Other works [23, 65, 70] employ the diffusion model as a proxy module for self-supervised learning, where noisy inputs are de-noised by conditioning on the image representations. This approach encourages the emergence of informative representations that capture key properties and semantics of the images. In this work, to the best of our knowledge, we for the first time leverage diffusion for video representation learning, where a video diffusion model is trained to de-noise video clips through taking the spatiotemporal representations as conditions, thereby encouraging the capture of spatial characteristics and temporal dynamics." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.822, + 0.182, + 0.837 + ], + "angle": 0, + "content": "3. Method" + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.848, + 0.25, + 0.863 + ], + "angle": 0, + "content": "3.1. Divot Tokenizer" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.871, + 0.486, + 0.903 + ], + "angle": 0, + "content": "We introduce Divot, a diffusion-powered tokenizer that leverage diffusion procedure for video representation learning." + }, + { + "type": "image", + "bbox": [ + 0.517, + 0.326, + 0.907, + 0.495 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.512, + 0.503, + 0.909, + 0.601 + ], + "angle": 0, + "content": "Figure 3. Overview of Divot-LLM. Video features from the Divot tokenizer are fed into the LLM to perform next-word prediction for video comprehension, while learnable queries are input into the LLM to model the distributions of Divot features using a Gaussian Mixture Model (GMM) for video generation. During inference, video features are sampled from the predicted GMM distribution to decode videos using the de-tokenizer." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.614, + 0.907, + 0.66 + ], + "angle": 0, + "content": "Additionally, the proxy diffusion model used for training the tokenizer can serve as a de-tokenizer to decode realistic video clips from their spatiotemporal representations." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.669, + 0.816, + 0.684 + ], + "angle": 0, + "content": "3.1.1. Preliminary: Video Diffusion Model." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.688, + 0.909, + 0.81 + ], + "angle": 0, + "content": "Diffusion models [20, 50] learn to model a probability distribution by reversing a process that progressively adds noise to the data. Specifically, given data \\(\\mathbf{x}_0 \\sim p(\\mathbf{x})\\), the forward process gradually adds random Gaussian noise \\(\\epsilon_{\\mathbf{t}} \\in \\mathcal{N}(\\mathbf{0}, \\mathbf{I})\\) to the data sample \\(\\mathbf{x}_0\\) with a total of \\(T\\) timesteps to yield \\(\\mathbf{x}_t\\) through a parameterization trick. The denoising process predicts \\(\\epsilon_{\\mathbf{t}}\\) in the forward diffusion process with a denoising network \\(\\epsilon_{\\theta}(\\mathbf{x}_t, t)\\), which is trained by the objective below," + }, + { + "type": "equation", + "bbox": [ + 0.581, + 0.826, + 0.907, + 0.848 + ], + "angle": 0, + "content": "\\[\n\\min _ {\\theta} \\mathbb {E} _ {t, \\mathbf {x} \\sim p, \\epsilon \\sim \\mathcal {N} (\\mathbf {0}, \\mathbf {I})} \\| \\epsilon - \\epsilon_ {\\theta} (\\mathbf {x} _ {t}, t) \\| _ {2} ^ {2}, \\tag {1}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.856, + 0.909, + 0.901 + ], + "angle": 0, + "content": "where \\(\\epsilon\\) is the sampled Gaussian noise and \\(\\theta\\) indicates the parameters of the denoising network. During inference, we can perform iterative denoising from a random Gaussian" + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "13608" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.104, + 0.096, + 0.353, + 0.27 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.379, + 0.096, + 0.62, + 0.27 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.658, + 0.09, + 0.91, + 0.271 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.089, + 0.282, + 0.908, + 0.34 + ], + "angle": 0, + "content": "Figure 4. Paradigms for modeling video representations from the Divot tokenizer with a LLM for video generation. (a) MSE Regression, where the LLM output is trained to minimize its distance with video features using Mean Squared Error (MSE) loss; (b) Diffusion Modeling, where the LLM output is fed into a denoising network as the condition to predict the noise added to video features; (c) GMM Modeling, where the LLM output is trained to predict the parameters of a Gaussian Mixture Model (GMM) for modeling video feature distributions." + }, + { + "type": "text", + "bbox": [ + 0.091, + 0.358, + 0.293, + 0.373 + ], + "angle": 0, + "content": "noise for the denoised data \\(\\mathbf{x}_0\\)" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.376, + 0.484, + 0.468 + ], + "angle": 0, + "content": "For video diffusion models [8, 78], given a video \\(\\mathbf{x}\\), a latent representation \\(\\mathbf{z} = \\mathcal{E}(\\mathbf{x})\\) is first encoded to reduce the computational complexity. Then the forward diffusion process and backward denoise process are performed in this latent space with a denoising network \\(\\epsilon_{\\theta}(\\mathbf{x}_t,\\mathbf{c},t)\\), where \\(\\mathbf{c}\\) denotes denoising conditions like text or visual prompts." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.483, + 0.258, + 0.499 + ], + "angle": 0, + "content": "3.1.2. Training Pipeline" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.505, + 0.484, + 0.822 + ], + "angle": 0, + "content": "As illustrated in Fig. 2, given a video clip, we separately sample sparse frames at 2 fps to obtain the video representations from the tokenizer, and sample dense frames at 8 fps to obtain latent representations \\( z_{0} \\) from the frozen VAE [28] encoder. Sparse frames are sampled as the input of the video tokenizer considering the semantic redundancy between adjacent frames. The forward diffusion process gradually adds Gaussian noise \\( \\theta \\) to \\( z_{0} \\) for producing the noisy input \\( z_{t} \\). At each backward step \\( t \\), a denoising U-Net is trained to predict the noise added from the previous step to the current step by taking the time embedding and video representations as the condition. Specifically, the video representations interact with the denoising U-Net intermediate features through cross-attention layers, where each noisy latent attends to all video tokens. By constraining the U-Net to reconstruct fine-grained spatial and temporal information of video clips through relying on video features, the Divot tokenizer is optimized to capture both spatial characteristics and temporal dynamics for robust video representations. The Divot tokenizer is trained on pure videos of a subset of WebVid10M [2] and Panda-70M [9], totaling 10M videos." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.825, + 0.484, + 0.902 + ], + "angle": 0, + "content": "After training the Divot tokenizer, the proxy denoising U-Net (employed to implement the parameterized loss function) can serve as an effective video de-tokenizer, which is able to decode semantically aligned video clips from the learned spatiotemporal representations as shown in Fig. 5." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.358, + 0.697, + 0.372 + ], + "angle": 0, + "content": "3.1.3. Model Architecture" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.378, + 0.909, + 0.62 + ], + "angle": 0, + "content": "As shown in Fig. 2, the Divot tokenizer is composed of a pre-trained ViT encoder to extract frame-level features, a transformer for spatial and temporal fusion, and a Perceiver Resampler [1] to produce a fixed number of video tokens. The Perceiver Resampler is adopted for two reasons: (1) to reduce the number of video tokens that a LLM need to predict for generation, and (2) to transform the patch-position dependent features into a sequence of high-level features without 3D positional dependencies, which we empirically find easier for an LLM to fit (See Sec. 4.3). Specifically, given a video clip with a duration of two seconds, we sample 5 frames at 2 fps, resulting in a total of 64 video tokens. We adopt the de-noising U-Net in DynamiCrafter [78], but reduce the input channel of the 3D convolution from 8 to 4 since we remove the original concatenation of a conditional image with noisy latents." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.634, + 0.878, + 0.65 + ], + "angle": 0, + "content": "3.2. Video Representation Modeling with LLM" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.658, + 0.909, + 0.854 + ], + "angle": 0, + "content": "The core challenge of generating videos using a LLM with the Divot tokenizer lies in effectively modeling the continuous video features. The most straightforward solution is to minimize the distance between the LLM output and the video representations using mean squared error (MSE) loss following previous work [17, 54] for image generation as illustrated in Fig. 4 (a). However, we empirically find that this approach is not effective for modeling continuous video features, as the generated videos tend to exhibit repeating patterns. We analyze that the deterministic regression regularizes the LLM to learn overly averaged representations, which is particularly catastrophic in video generation as videos must ensure both spatial and temporal diversity." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.856, + 0.909, + 0.901 + ], + "angle": 0, + "content": "Inspired by recent work MAR [33], instead of deterministic regression, we aim to model the probability distribution of video representations using the LLM. As shown in Fig. 4," + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "13609" + } + ], + [ + { + "type": "image_caption", + "bbox": [ + 0.155, + 0.094, + 0.261, + 0.105 + ], + "angle": 0, + "content": "Input video (fps=2)" + }, + { + "type": "image", + "bbox": [ + 0.099, + 0.106, + 0.324, + 0.193 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.099, + 0.195, + 0.323, + 0.281 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.099, + 0.283, + 0.323, + 0.369 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.543, + 0.094, + 0.698, + 0.105 + ], + "angle": 0, + "content": "Reconstructed video (fps=8)" + }, + { + "type": "image", + "bbox": [ + 0.342, + 0.106, + 0.903, + 0.193 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.342, + 0.195, + 0.903, + 0.281 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.342, + 0.283, + 0.903, + 0.369 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.09, + 0.382, + 0.907, + 0.412 + ], + "angle": 0, + "content": "Figure 5. Reconstructed videos, where the Divot tokenizer obtains spatiotemporal representations of sparsely sampled video frames and the de-tokenizer decodes these representations into semantically aligned and temporally coherent video clips." + }, + { + "type": "text", + "bbox": [ + 0.089, + 0.43, + 0.486, + 0.656 + ], + "angle": 0, + "content": "we explore two approaches including (b) Diffusion Modeling [33] and (c) GMM Modeling [61]. Specifically, for the diffusion modeling, given continuous-valued video features to be predicted, the LLM produces output, which serves as the condition of a denoising network (a small MLP) to predict the Gaussian noise added to the video features. The diffusion model is trained for representing the distribution of video features. For GMM modeling, we use a Gaussian Mixture Model (GMM) to model the distribution of the video features, and train the LLM to predict \\(2kd + k\\) parameters per video token (\\(kd\\) mean and \\(kd\\) variance parameters for the mixture components, and \\(k\\) mixture probabilities). We optimize the LLM by minimizing the discrepancy between the predicted GMM distribution and the video representations with negative log-likelihood (NLL) loss." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.66, + 0.486, + 0.901 + ], + "angle": 0, + "content": "During inference, in diffusion modeling, the denoising network denoise the final video features from Gaussian noise gradually by taking the LLM output as the condition. In GMM modeling, we draw samples from the predicted GMM distribution as the final video representations. To empirically investigate the effectiveness of the approaches above, we train the LLM with MSR-VTT [79], and evaluates text-to-video generation on test set with FVD [62] and similarity score [49] as the metric following previous work [26, 82]. As listed in Tab. 7, GMM modeling achieves better performance than diffusion modeling and MSE regression in video generation. We speculate that high-level features obtained by the Divot tokenizer are more sensitive to Gaussian noise compared to the VAE latents used by MAR, making training more challenging and resulting in suboptimal results. Therefore, we adopt GMM modeling to train Divot-LLM." + }, + { + "type": "table_caption", + "bbox": [ + 0.516, + 0.429, + 0.903, + 0.442 + ], + "angle": 0, + "content": "Table 2. Datasets used for training the tokenizer and Divot-LLM." + }, + { + "type": "table", + "bbox": [ + 0.518, + 0.453, + 0.907, + 0.622 + ], + "angle": 0, + "content": "
StageTypeDataset
TokenizerPure VideoWebVid-10M [2], Panda-70M [9]
Pre-trainVideo-textWebVid-10M [2]
Image-textCC3M [52], CapsFusion [87], LAION-COCO [51]
SFTClassificationKinetics-710 [27], SSV2 [18]
VQATGIF [34], NextQA [76], CLEVRER [85], YouCook2 [92], PerceptionTest[48], EgoQA [19], ActivityNetQA[88]
InstructionVideo-ChatGPT[43], LLaVA-mixed[39], Valley [42], LLaVA-Video-178K[37]
GenerationWebVid-10M [2]
StoryTellingIn-house data
" + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.647, + 0.828, + 0.664 + ], + "angle": 0, + "content": "3.3. Pre-training and Instruction Tuning" + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.67, + 0.855, + 0.685 + ], + "angle": 0, + "content": "3.3.1. Training Stage I: Multimodal Pre-training" + }, + { + "type": "text", + "bbox": [ + 0.511, + 0.689, + 0.907, + 0.901 + ], + "angle": 0, + "content": "As shown in Fig 3, Divot-LLM adopts next-word prediction and GMM modeling on video-text data for video comprehension and generation. Specifically, the video features from the Divot tokenizer, the special tokens indicating the start and end of video features, along with the text tokens of the caption are fed into the pre-trained Mistral-7B [24] for next token prediction trained with cross-entropy loss. Text tokens of the caption and \\( N \\) learnable queries are input into the LLM, where the output of the learnable queries are trained via bidirectional attention to model a GMM distribution for the video features using NLL loss. During inference, we draw samples from the predicted GMM distribution as condition of the denoising U-Net to decode videos. We pre-train Divot-LLM from the pre-trained Mistral-7B model using" + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "13610" + } + ], + [ + { + "type": "table_caption", + "bbox": [ + 0.09, + 0.089, + 0.908, + 0.118 + ], + "angle": 0, + "content": "Table 3. Comparison for video comprehension with MLLMs. \"Video-Gen\" denotes whether the model can generate videos besides texts. The evaluation metric is accuracy. The best results are bold and the second best results are underlined." + }, + { + "type": "table", + "bbox": [ + 0.153, + 0.129, + 0.85, + 0.396 + ], + "angle": 0, + "content": "
ModelLLM sizeVideo-GenEgoSchemaPerception-TestMVBenchMSVDActivityNet
Gemini 1.0 Pro [58]-×55.751.1--49.8
Gemini 1.5 Pro [59]-×63.2---56.7
GPT4-V [46]-×55.6-43.7-59.5
GPT4-O [47]-×72.2---61.9
LLaMA-VID [35]7B×38.544.641.969.747.4
Video-ChatGPT [43]7B×---64.935.2
Video-LLaVA [37]7B×38.444.341.070.745.3
VideoChat2 [31]7B×42.247.351.170.049.1
LLaVA-NeXT-Video [38]7B×43.948.846.567.853.5
LLaVA-NeXT-Video [38]32B×60.9---54.3
PLLAVA [81]34B×-58.1--60.9
LLaVA-OneVision [30]72B×62.0---62.3
VideoLLaMA2 [10]7B×51.751.454.670.950.2
VideoLLaMA2 [10]72B×63.957.562.071.055.2
LWM [40]7B✓---55.9-
Video-LaVIT [26]7B✓37.347.9-73.250.1
VILA-U [74]7B✓---75.352.7
Divot-LLM7B✓46.558.352.176.455.8
" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.421, + 0.483, + 0.452 + ], + "angle": 0, + "content": "LoRA [22] on a subset of WebVid-10M [2] data (filtered for temporal dynamics in captions) and image-text data." + }, + { + "type": "title", + "bbox": [ + 0.09, + 0.459, + 0.483, + 0.475 + ], + "angle": 0, + "content": "3.3.2. Training Stage II: Multimodal Instruction Tuning" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.478, + 0.484, + 0.584 + ], + "angle": 0, + "content": "We perform multimodal instruction tuning on Divot-LLM to align it with human instructions through supervised fine-tuning on public datasets as listed in Tab. 2 with a LoRA module. We further fine-tune the pretrained Divot-LLM on an animated series called \"Curious George\" to achieve video storytelling, which generates storyline and corresponding video clips in an interleaved manner." + }, + { + "type": "title", + "bbox": [ + 0.09, + 0.598, + 0.216, + 0.615 + ], + "angle": 0, + "content": "4. Experiment" + }, + { + "type": "title", + "bbox": [ + 0.09, + 0.622, + 0.312, + 0.637 + ], + "angle": 0, + "content": "4.1.Quantitative Evaluation" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.644, + 0.483, + 0.78 + ], + "angle": 0, + "content": "Video Comprehension. We conduct extensive evaluations on video comprehension benchmarks including Multi-choice Video Question Answering (MC-VQA) on EgoSchema [44], Perception-Test [48], MVBench [32], and Open-Ended Video Question Answering (OE-VQA) on MSVD [7], ActivityNet [88]. Following VideoLLaMA 2 [10], we utilize GPT-3.5 to assess the quality of the generated answers of OE-VQA by determining whether the answers match the ground truth, and we report the percentage of \"Yes\" as Accuracy." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.781, + 0.484, + 0.902 + ], + "angle": 0, + "content": "For each testing video, we sample a maximum of 20 clips, each containing 5 frames. The evaluation results are reported in Tab. 3. Divot-LLM outperforms the baseline models that can generate both texts and videos, which demonstrates that our model effectively achieves video comprehension within a unified framework. Compared to VideoLLMs specifically designed for video comprehension of the same model size of LLM, Divot-LLM achieves competitive results with sig" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.421, + 0.908, + 0.497 + ], + "angle": 0, + "content": "nificantly fewer video-caption pairs for training (4.8M vs. 100M in VideoLLaMA 2). By utilizing diffusion procedure for video representation learning, our Divot tokenizer effectively captures robust spatiotemporal representations, enhancing the comprehension capabilities." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.51, + 0.909, + 0.661 + ], + "angle": 0, + "content": "Video Generation. We evaluate zero-shot text-to-video generation on MSR-VTT [79]. We randomly sample one caption for each testing video and generate 16 frames in \\(256 \\times 256\\)px resolution. We adopt the CLIP similarity (CLIPSIM) [71] and Fréchet video distance (FVD) [62] as the evaluation metric following Loong [69]. As listed in Tab. 4, Divot-LLM achieves performance comparable to existing video generation models in terms of visual quality and semantic alignment with captions using only 4.8 million video-text pairs for training." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.67, + 0.725, + 0.687 + ], + "angle": 0, + "content": "4.2. Qualitative Evaluation" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.692, + 0.908, + 0.799 + ], + "angle": 0, + "content": "Text-to-video Generation. We perform qualitative comparison of text-to-video generation with baseline MLLMs that are capable of unified video comprehension and generation. As illustrated in Fig. 6, through modeling the distributions of Divot features with a predicted GMM, our Divot-LLM can generate videos that are both semantically aligned with text prompts and temporally coherent within frames." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.81, + 0.909, + 0.902 + ], + "angle": 0, + "content": "Video StoryTelling. We fine-tune the pre-trained Divot-LLM on an animated series called \"Curious George\" for video storytelling. As shown in Fig. 7, given a brief story instruction, our Divot-LLM can generate a sequence of multimodal stories with rich narrative text and contextually relevant videos that are temporally coherent. Since we only" + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.946, + 0.518, + 0.957 + ], + "angle": 0, + "content": "13611" + } + ], + [ + { + "type": "image_caption", + "bbox": [ + 0.18, + 0.092, + 0.441, + 0.106 + ], + "angle": 0, + "content": "A man walking slowly under a sunset." + }, + { + "type": "image", + "bbox": [ + 0.101, + 0.108, + 0.507, + 0.332 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.596, + 0.092, + 0.828, + 0.107 + ], + "angle": 0, + "content": "A person is applying eye makeup." + }, + { + "type": "image", + "bbox": [ + 0.524, + 0.108, + 0.903, + 0.332 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.09, + 0.346, + 0.908, + 0.375 + ], + "angle": 0, + "content": "Figure 6. Qualitative comparison of text-to-video generation with MLLMs that are capable of unified video comprehension and generation. Divot-LLM effectively generates videos that are both semantically aligned with text prompts and temporally coherent within frames." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.387, + 0.484, + 0.434 + ], + "angle": 0, + "content": "fine-tune the de-tokenizer for adaptation to the new domain, it demonstrates the generalizability of our Divot tokenizer for obtaining robust video representations." + }, + { + "type": "table_caption", + "bbox": [ + 0.09, + 0.447, + 0.487, + 0.504 + ], + "angle": 0, + "content": "Table 4. Comparison for zero-shot text-to-video generation. \"Data size\" refers to the number of training video data, and \"Unified\" denotes if the model enables video comprehension and generation. The best results are bold and the second best results are underlined." + }, + { + "type": "table", + "bbox": [ + 0.095, + 0.516, + 0.484, + 0.715 + ], + "angle": 0, + "content": "
ModelData sizeUnifiedMSR-VTT
CLIPSIM (↑)FVD (↓)
CogVideo [21]5.4M×0.26311294
Video LDM [5]10M×0.2929-
VideoComposer [66]10M×0.2932580
InternVid [68]28M×0.2951-
Make-A-Video [53]20M×0.3049-
VideoPoet [29]270M×0.3049213
PYoCo [14]22.5M×--
SVD [4]152M×--
Video-LavIT [26]10M✓0.3012188.36
Loong [69]16M×0.2903274
Snap Video [45]-×0.2793110.4
VILA-U [74]1M✓0.2937499.06
Divot-LLM4.8M✓0.2938301.4
" + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.742, + 0.245, + 0.759 + ], + "angle": 0, + "content": "4.3. Ablation Study" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.765, + 0.486, + 0.901 + ], + "angle": 0, + "content": "Diffusion for Video Comprehension. We design two baselines to validate the effectiveness of the diffusion procedure to learn spatiotemporal representations for VideoLLMs. As shown in Tab. 5, both models are pre-trained on Valley [42] and instruction tuned on Video-ChatGPT [43]. The model with diffusion loss employs our Divot tokenizer, while the model with caption loss adopts the same architecture but its tokenizer is pre-trained using captioning loss with a frozen LLM on Valley. The model that employs a video tokenizer" + }, + { + "type": "table_caption", + "bbox": [ + 0.513, + 0.385, + 0.907, + 0.413 + ], + "angle": 0, + "content": "Table 6. Ablation study on the training objective of the video tokenizer. The evaluation metric is accuracy." + }, + { + "type": "table", + "bbox": [ + 0.556, + 0.424, + 0.866, + 0.482 + ], + "angle": 0, + "content": "
Loss TypeMV-BenchMSVDActivityNet
Caption30.866.143.2
Diffusion33.268.944.3
" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.496, + 0.907, + 0.572 + ], + "angle": 0, + "content": "trained with diffusion loss achieves better performance in video comprehension benchmarks, demonstrating that the diffusion process can effectively learn robust video representations in a self-supervised manner, without the need for paired caption annotations." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.584, + 0.907, + 0.658 + ], + "angle": 0, + "content": "Video Generation with LLM. We perform various ablation studies to explore an effective approach for generating videos with a LLM through training on MSR-VTT training set and evaluating text-to-video generation on test set. We use ViT-G/14 to calculate CLIPSIM for better discrimination." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.66, + 0.909, + 0.902 + ], + "angle": 0, + "content": "Q1: Which type of video representations is easier? We investigate two types of video representations, patch-position dependent features obtained from a spatial-temporal transformer and patch-position independent features after a Perceiver Resample with learnable queries. As listed in Tab. 7, fitting features without 3D positional dependencies achieves higher performance, which is also observed in recent work [64]. We also experiment with training the video tokenizer in the VAE manner, which involves predicting the means and variances of a normal distribution and sampling video representations using the re-parametrization trick following GIVT [61]. However, we observe that it is difficult for the LLM to converge and the video de-tokenizer achieves unsatisfactory reconstruction results. We conclude that introducing variances during tokenization for high-level video features may not be appropriate." + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.52, + 0.958 + ], + "angle": 0, + "content": "13612" + } + ], + [ + { + "type": "image_caption", + "bbox": [ + 0.095, + 0.093, + 0.432, + 0.111 + ], + "angle": 0, + "content": "Generate a story about the adventures of George and his friends." + }, + { + "type": "image", + "bbox": [ + 0.101, + 0.121, + 0.127, + 0.136 + ], + "angle": 0, + "content": null + }, + { + "type": "text", + "bbox": [ + 0.106, + 0.139, + 0.203, + 0.177 + ], + "angle": 0, + "content": "A young girl found a glowing orb floating in the air through her magnifying glass." + }, + { + "type": "image_caption", + "bbox": [ + 0.212, + 0.125, + 0.249, + 0.134 + ], + "angle": 0, + "content": "REC" + }, + { + "type": "image", + "bbox": [ + 0.212, + 0.135, + 0.279, + 0.183 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.28, + 0.134, + 0.344, + 0.182 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.345, + 0.134, + 0.408, + 0.182 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.409, + 0.134, + 0.474, + 0.183 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.104, + 0.184, + 0.127, + 0.199 + ], + "angle": 0, + "content": null + }, + { + "type": "text", + "bbox": [ + 0.104, + 0.209, + 0.202, + 0.238 + ], + "angle": 0, + "content": "George watched a bubble float by, feeling curious and joyful." + }, + { + "type": "image_caption", + "bbox": [ + 0.212, + 0.192, + 0.25, + 0.2 + ], + "angle": 0, + "content": "REC" + }, + { + "type": "image", + "bbox": [ + 0.212, + 0.2, + 0.28, + 0.25 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.28, + 0.192, + 0.344, + 0.25 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.345, + 0.192, + 0.408, + 0.25 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.409, + 0.192, + 0.474, + 0.25 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.104, + 0.254, + 0.127, + 0.268 + ], + "angle": 0, + "content": null + }, + { + "type": "text", + "bbox": [ + 0.104, + 0.275, + 0.203, + 0.303 + ], + "angle": 0, + "content": "At lunchtime, a boy found a giant sandwich and felt excited." + }, + { + "type": "image_caption", + "bbox": [ + 0.212, + 0.258, + 0.251, + 0.266 + ], + "angle": 0, + "content": "REC" + }, + { + "type": "image", + "bbox": [ + 0.212, + 0.266, + 0.28, + 0.315 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.28, + 0.266, + 0.344, + 0.315 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.345, + 0.266, + 0.408, + 0.315 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.409, + 0.266, + 0.474, + 0.315 + ], + "angle": 0, + "content": null + }, + { + "type": "title", + "bbox": [ + 0.536, + 0.099, + 0.83, + 0.11 + ], + "angle": 0, + "content": "Generate a story about George's fun-filled day in the kitchen." + }, + { + "type": "image", + "bbox": [ + 0.532, + 0.122, + 0.554, + 0.136 + ], + "angle": 0, + "content": null + }, + { + "type": "text", + "bbox": [ + 0.535, + 0.139, + 0.626, + 0.185 + ], + "angle": 0, + "content": "George and his pal joyfully cook in the kitchen, creating a tasty snack with a big blue book." + }, + { + "type": "image_caption", + "bbox": [ + 0.635, + 0.126, + 0.668, + 0.134 + ], + "angle": 0, + "content": "REC" + }, + { + "type": "image", + "bbox": [ + 0.635, + 0.135, + 0.699, + 0.183 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.7, + 0.134, + 0.764, + 0.183 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.765, + 0.134, + 0.829, + 0.183 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.83, + 0.134, + 0.895, + 0.183 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.532, + 0.186, + 0.554, + 0.199 + ], + "angle": 0, + "content": null + }, + { + "type": "text", + "bbox": [ + 0.535, + 0.201, + 0.615, + 0.248 + ], + "angle": 0, + "content": "A woman in the kitchen shows George how to use a new kitchen gadget." + }, + { + "type": "image_caption", + "bbox": [ + 0.635, + 0.192, + 0.668, + 0.2 + ], + "angle": 0, + "content": "REC" + }, + { + "type": "image", + "bbox": [ + 0.635, + 0.2, + 0.699, + 0.25 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.7, + 0.192, + 0.764, + 0.25 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.765, + 0.192, + 0.829, + 0.25 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.83, + 0.192, + 0.895, + 0.25 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.532, + 0.255, + 0.553, + 0.268 + ], + "angle": 0, + "content": null + }, + { + "type": "text", + "bbox": [ + 0.535, + 0.271, + 0.618, + 0.308 + ], + "angle": 0, + "content": "George spreads the thick, peanutty paste, making a yummy snack." + }, + { + "type": "image_caption", + "bbox": [ + 0.635, + 0.258, + 0.668, + 0.266 + ], + "angle": 0, + "content": "REC" + }, + { + "type": "image", + "bbox": [ + 0.635, + 0.266, + 0.699, + 0.314 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.7, + 0.258, + 0.764, + 0.315 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.765, + 0.258, + 0.829, + 0.315 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.83, + 0.258, + 0.895, + 0.315 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.09, + 0.333, + 0.907, + 0.362 + ], + "angle": 0, + "content": "Figure 7. Qualitative examples of video storytelling by Divot-LLM. Given a story instruction, Divot-LLM can generate rich textual narratives along with corresponding video clips that are temporally coherent in an interleaved manner." + }, + { + "type": "table_caption", + "bbox": [ + 0.091, + 0.376, + 0.904, + 0.39 + ], + "angle": 0, + "content": "Table 8. Ablation study on video representation modeling with LLMs for generation. We evaluate text-to-video generation on MSR-VTT." + }, + { + "type": "table", + "bbox": [ + 0.152, + 0.401, + 0.85, + 0.492 + ], + "angle": 0, + "content": "
RepresentationObjectiveMechanism
patch-position dependentpatch-position independentMSEDiffusionGMMARQuery
ε-predv-predcausalbidirectional
CLIPSIM (↑)0.31920.32650.31680.28110.28420.32650.23860.30800.3265
FVD (↓)378.50366.60438.94418.19377.17366.60447.88416.60366.60
" + }, + { + "type": "text", + "bbox": [ + 0.089, + 0.517, + 0.486, + 0.775 + ], + "angle": 0, + "content": "Q2: Which training objective is more suitable? As introduce in Sec. 3, we explore MSE regression, Diffusion modeling and GMM modeling to fit high-level continuous features with a LLM. As listed in Tab. 7, simply aligning the LLM output with video features using MSE loss yields the lowest generation quality, suggesting that deterministic regression is inadequate for modeling spatiotemporal representations. Training a denoising network to denoise the noisy video features by taking the LLM output as the condition also achieves inferior performance with both \\(\\epsilon\\) prediction and \\(v\\) prediction. Different from MAR [33] that denoises low-level VAE latents, our goal is to denoise high-level video features. We speculate that these features are more sensitive to Gaussian noise, making them more challenging to denoise. Training the LLM to model the distribution of high-level video features using a GMM model achieves the best generation quality and semantic alignment with captions." + }, + { + "type": "text", + "bbox": [ + 0.089, + 0.78, + 0.486, + 0.902 + ], + "angle": 0, + "content": "Q3: Which LLM mechanism is more effective? We train the LLM to fit the video representations with GMM modeling using both an autoregressive approach and a query-based approach, with the latter exploring causal attention and bidirectional attention within the LLM. Predicting the features of each video token in an autoregressive manner results in the worst performance due to the accumulation of errors, particularly when the features of the pervious to-" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.517, + 0.909, + 0.579 + ], + "angle": 0, + "content": "ken are sampled from a GMM distribution for predicting the distribution of the current token. The query-based approach achieves better results with bidirectional attention, as it enables each query to attend to all tokens for predictions." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.601, + 0.634, + 0.617 + ], + "angle": 0, + "content": "5. Conclusion" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.63, + 0.909, + 0.797 + ], + "angle": 0, + "content": "In this work, we introduce Divot, a diffusion-powered video tokenizer learned in a self-supervised manner for unified comprehension and generation. We further investigate effective approaches for modeling continuous video representations with the LLM and present Divot-LLM to understand and generate video content in a single framework. Divot-LLM achieves competitive performance in video comprehension and generation benchmarks, and enables video storytelling effectively. We hope our work will draw increased attention to unifying video comprehension and generation through the design of sophisticated tokenizers." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.826, + 0.909, + 0.903 + ], + "angle": 0, + "content": "Limitation. As we primarily focus on exploring effective representations and approaches for video generation with a unified LLM, the current model is trained to predict video representations for only a single clip and does not generate longer videos, which will be explored in our future work." + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "13613" + } + ], + [ + { + "type": "title", + "bbox": [ + 0.093, + 0.09, + 0.188, + 0.106 + ], + "angle": 0, + "content": "References" + }, + { + "type": "ref_text", + "bbox": [ + 0.101, + 0.115, + 0.486, + 0.198 + ], + "angle": 0, + "content": "[1] Jean-Baptiste Alayrac, Jeff Donahue, Pauline Luc, Antoine Miech, Iain Barr, Yana Hasson, Karel Lenc, Arthur Mensch, Katherine Millican, Malcolm Reynolds, et al. Flamingo: a visual language model for few-shot learning. Advances in Neural Information Processing Systems, 35:23716-23736, 2022. 2, 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.101, + 0.201, + 0.486, + 0.27 + ], + "angle": 0, + "content": "[2] Max Bain, Arsha Nagrani, Gül Varol, and Andrew Zisserman. Frozen in time: A joint video and image encoder for end-to-end retrieval. In Proceedings of the IEEE/CVF international conference on computer vision, pages 1728-1738, 2021. 4, 5, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.273, + 0.485, + 0.327 + ], + "angle": 0, + "content": "[3] Dmitry Baranchuk, Ivan Rubachev, Andrey Voynov, Valentin Khrulkov, and Artem Babenko. Label-efficient semantic segmentation with diffusion models. arXiv preprint arXiv:2112.03126, 2021. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.33, + 0.484, + 0.4 + ], + "angle": 0, + "content": "[4] Andreas Blattmann, Tim Dockhorn, Sumith Kulal, Daniel Mendelevitch, Maciej Kilian, Dominik Lorenz, Yam Levi, Zion English, Vikram Voleti, Adam Letts, et al. Stable video diffusion: Scaling latent video diffusion models to large datasets. arXiv preprint arXiv:2311.15127, 2023. 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.402, + 0.485, + 0.485 + ], + "angle": 0, + "content": "[5] Andreas Blattmann, Robin Rombach, Huan Ling, Tim Dockhorn, Seung Wook Kim, Sanja Fidler, and Karsten Kreis. Align your latents: High-resolution video synthesis with latent diffusion models. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 22563-22575, 2023. 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.488, + 0.485, + 0.557 + ], + "angle": 0, + "content": "[6] Tom Brown, Benjamin Mann, Nick Ryder, Melanie Subbiah, Jared D Kaplan, Prafulla Dhariwal, Arvind Neelakantan, Pranav Shyam, Girish Sastry, Amanda Askell, et al. Language models are few-shot learners. Advances in neural information processing systems, 33:1877-1901, 2020. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.559, + 0.485, + 0.615 + ], + "angle": 0, + "content": "[7] David Chen and William B Dolan. Collecting highly parallel data for paraphrase evaluation. In Proceedings of the 49th annual meeting of the association for computational linguistics: human language technologies, pages 190-200, 2011. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.617, + 0.485, + 0.685 + ], + "angle": 0, + "content": "[8] Haoxin Chen, Menghan Xia, Yingqing He, Yong Zhang, Xiaodong Cun, Shaoshu Yang, Jinbo Xing, Yaofang Liu, Qifeng Chen, Xintao Wang, et al. Videocrafter1: Open diffusion models for high-quality video generation. arXiv preprint arXiv:2310.19512, 2023. 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.688, + 0.486, + 0.785 + ], + "angle": 0, + "content": "[9] Tsai-Shien Chen, Aliaksandr Siarohin, Willi Menapace, Ekaterina Deyneka, Hsiang-wei Chao, Byung Eun Jeon, Yuwei Fang, Hsin-Ying Lee, Jian Ren, Ming-Hsuan Yang, et al. Panda-70m: Captioning 70m videos with multiple cross-modality teachers. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 13320–13331, 2024. 4, 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.788, + 0.485, + 0.857 + ], + "angle": 0, + "content": "[10] Zesen Cheng, Sicong Leng, Hang Zhang, Yifei Xin, Xin Li, Guanzheng Chen, Yongxin Zhu, Wenqi Zhang, Ziyang Luo, Deli Zhao, et al. Videollama 2: Advancing spatial-temporal modeling and audio understanding in video-llms. arXiv preprint arXiv:2406.07476, 2024. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.859, + 0.485, + 0.901 + ], + "angle": 0, + "content": "[11] Aakanksha Chowdhery, Sharan Narang, Jacob Devlin, Maarten Bosma, Gaurav Mishra, Adam Roberts, Paul Barham, Hyung Won Chung, Charles Sutton, Sebastian Gehrmann," + }, + { + "type": "list", + "bbox": [ + 0.094, + 0.115, + 0.486, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.545, + 0.093, + 0.907, + 0.121 + ], + "angle": 0, + "content": "et al. Palm: Scaling language modeling with pathways. arXiv preprint arXiv:2204.02311, 2022. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.123, + 0.908, + 0.19 + ], + "angle": 0, + "content": "[12] Runpei Dong, Chunrui Han, Yuang Peng, Zekun Qi, Zheng Ge, Jinrong Yang, Liang Zhao, Jianjian Sun, Hongyu Zhou, Haoran Wei, et al. Dreamllm: Synergistic multimodal comprehension and creation. arXiv preprint arXiv:2309.11499, 2023. 1, 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.193, + 0.908, + 0.275 + ], + "angle": 0, + "content": "[13] Alexey Dosovitskiy, Lucas Beyer, Alexander Kolesnikov, Dirk Weissenborn, Xiaohua Zhai, Thomas Unterthiner, Mostafa Dehghani, Matthias Minderer, Georg Heigold, Sylvain Gelly, et al. An image is worth 16x16 words: Transformers for image recognition at scale. arXiv preprint arXiv:2010.11929, 2020. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.278, + 0.908, + 0.36 + ], + "angle": 0, + "content": "[14] Songwei Ge, Seungjun Nah, Guilin Liu, Tyler Poon, Andrew Tao, Bryan Catanzaro, David Jacobs, Jia-Bin Huang, Ming-Yu Liu, and Yogesh Balaji. Preserve your own correlation: A noise prior for video diffusion models. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 22930-22941, 2023. 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.362, + 0.905, + 0.403 + ], + "angle": 0, + "content": "[15] Yuying Ge, Yixiao Ge, Ziyun Zeng, Xintao Wang, and Ying Shan. Planting a seed of vision in large language model. arXiv preprint arXiv:2307.08041, 2023. 1, 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.406, + 0.905, + 0.447 + ], + "angle": 0, + "content": "[16] Yuying Ge, Sijie Zhao, Ziyun Zeng, Yixiao Ge, Chen Li, Xintao Wang, and Ying Shan. Making llama see and draw with seed tokenizer. arXiv preprint arXiv:2310.01218, 2023." + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.449, + 0.908, + 0.516 + ], + "angle": 0, + "content": "[17] Yuying Ge, Sijie Zhao, Jinguo Zhu, Yixiao Ge, Kun Yi, Lin Song, Chen Li, Xiaohan Ding, and Ying Shan. Seed-x: Multimodal models with unified multi-granularity comprehension and generation. arXiv preprint arXiv:2404.14396, 2024. 1, 2, 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.519, + 0.908, + 0.616 + ], + "angle": 0, + "content": "[18] Raghav Goyal, Samira Ebrahimi Kahou, Vincent Michalski, Joanna Materzynska, Susanne Westphal, Heuna Kim, Valentin Haenel, Ingo Fruend, Peter Yianilos, Moritz Mueller-Freitag, et al. The\" something something\" video database for learning and evaluating visual common sense. In Proceedings of the IEEE international conference on computer vision, pages 5842-5850, 2017. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.618, + 0.908, + 0.701 + ], + "angle": 0, + "content": "[19] Kristen Grauman, Andrew Westbury, Eugene Byrne, Zachary Chavis, Antonino Furnari, Rohit Girdhar, Jackson Hamburger, Hao Jiang, Miao Liu, Xingyu Liu, et al. Ego4d: Around the world in 3,000 hours of egocentric video. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 18995-19012, 2022. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.703, + 0.908, + 0.744 + ], + "angle": 0, + "content": "[20] Jonathan Ho, Ajay Jain, and Pieter Abbeel. Denoising diffusion probabilistic models. Advances in neural information processing systems, 33:6840-6851, 2020. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.746, + 0.908, + 0.8 + ], + "angle": 0, + "content": "[21] Wenyi Hong, Ming Ding, Wendi Zheng, Xinghan Liu, and Jie Tang. Cogvideo: Large-scale pretraining for text-to-video generation via transformers. arXiv preprint arXiv:2205.15868, 2022. 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.802, + 0.908, + 0.857 + ], + "angle": 0, + "content": "[22] Edward J Hu, Yelong Shen, Phillip Wallis, Zeyuan Allen-Zhu, Yanzhi Li, Shean Wang, Lu Wang, and Weizhu Chen. Lora: Low-rank adaptation of large language models. arXiv preprint arXiv:2106.09685, 2021. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.859, + 0.908, + 0.901 + ], + "angle": 0, + "content": "[23] Drew A Hudson, Daniel Zoran, Mateusz Malinowski, Andrew K Lampinen, Andrew Jaegle, James L McClelland, Loic Matthew, Felix Hill, and Alexander Lerchner. Soda: Bottle-" + }, + { + "type": "list", + "bbox": [ + 0.517, + 0.093, + 0.908, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.946, + 0.52, + 0.957 + ], + "angle": 0, + "content": "13614" + } + ], + [ + { + "type": "ref_text", + "bbox": [ + 0.125, + 0.092, + 0.484, + 0.134 + ], + "angle": 0, + "content": "neck diffusion models for representation learning. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 23115-23127, 2024. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.136, + 0.484, + 0.205 + ], + "angle": 0, + "content": "[24] Albert Q Jiang, Alexandre Sablayrolles, Arthur Mensch, Chris Bamford, Devendra Singh Chaplot, Diego de las Casas, Florian Bressand, Gianna Lengyel, Guillaume Lample, Lucile Saulnier, et al. Mistral 7b. arXiv preprint arXiv:2310.06825, 2023. 2, 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.207, + 0.484, + 0.275 + ], + "angle": 0, + "content": "[25] Yang Jin, Kun Xu, Liwei Chen, Chao Liao, Jianchao Tan, Bin Chen, Chenyi Lei, An Liu, Chengru Song, Xiaogiang Lei, et al. Unified language-vision pretraining with dynamic discrete visual tokenization. arXiv preprint arXiv:2309.04669, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.278, + 0.484, + 0.347 + ], + "angle": 0, + "content": "[26] Yang Jin, Zhicheng Sun, Kun Xu, Liwei Chen, Hao Jiang, Quzhe Huang, Chengru Song, Yuliang Liu, Di Zhang, Yang Song, et al. Video-lavit: Unified video-language pre-training with decoupled visual-motional tokenization. arXiv preprint arXiv:2402.03161, 2024. 1, 2, 3, 5, 6, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.349, + 0.484, + 0.417 + ], + "angle": 0, + "content": "[27] Will Kay, Joao Carreira, Karen Simonyan, Brian Zhang, Chloe Hillier, Sudheendra Vijayanarasimhan, Fabio Viola, Tim Green, Trevor Back, Paul Natsev, et al. The kinetics human action video dataset. arXiv preprint arXiv:1705.06950, 2017. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.42, + 0.484, + 0.459 + ], + "angle": 0, + "content": "[28] Diederik P Kingma and Max Welling. Auto-encoding variational bayes. arXiv preprint arXiv:1312.6114, 2013. 2, 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.462, + 0.484, + 0.531 + ], + "angle": 0, + "content": "[29] Dan Kondratyuk, Lijun Yu, Xiuye Gu, José Lezama, Jonathan Huang, Grant Schindler, Rachel Hornung, Vighnesh Birodkar, Jimmy Yan, Ming-Chang Chiu, et al. Videopoet: A large language model for zero-shot video generation. arXiv preprint arXiv:2312.14125, 2023. 2, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.533, + 0.484, + 0.589 + ], + "angle": 0, + "content": "[30] Bo Li, Yuanhan Zhang, Dong Guo, Renrui Zhang, Feng Li, Hao Zhang, Kaichen Zhang, Yanwei Li, Ziwei Liu, and Chunyuan Li. Llava-onevision: Easy visual task transfer. arXiv preprint arXiv:2408.03326, 2024. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.591, + 0.484, + 0.646 + ], + "angle": 0, + "content": "[31] Kunchang Li, Yali Wang, Yinan He, Yizhuo Li, Yi Wang, Yi Liu, Zun Wang, Jilan Xu, Guo Chen, Ping Luo, et al. Mybench: A comprehensive multi-modal video understanding benchmark. arXiv preprint arXiv:2311.17005, 2023. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.648, + 0.484, + 0.729 + ], + "angle": 0, + "content": "[32] Kunchang Li, Yali Wang, Yinan He, Yizhuo Li, Yi Wang, Yi Liu, Zun Wang, Jilan Xu, Guo Chen, Ping Luo, et al. Mybench: A comprehensive multi-modal video understanding benchmark. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 22195-22206, 2024. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.732, + 0.484, + 0.786 + ], + "angle": 0, + "content": "[33] Tianhong Li, Yonglong Tian, He Li, Mingyang Deng, and Kaiming He. Autoregressive image generation without vector quantization. arXiv preprint arXiv:2406.11838, 2024. 2, 4, 5, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.789, + 0.484, + 0.858 + ], + "angle": 0, + "content": "[34] Yuncheng Li, Yale Song, Liangliang Cao, Joel Tetreault, Larry Goldberg, Alejandro Jaimes, and Jiebo Luo. Tgif: A new dataset and benchmark on animated gif description. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition, pages 4641-4650, 2016. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.86, + 0.484, + 0.901 + ], + "angle": 0, + "content": "[35] Yanwei Li, Chengyao Wang, and Jiaya Jia. Llama-vid: An image is worth 2 tokens in large language models. arXiv preprint arXiv:2311.17043, 2023. 6" + }, + { + "type": "list", + "bbox": [ + 0.093, + 0.092, + 0.484, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.093, + 0.907, + 0.161 + ], + "angle": 0, + "content": "[36] Yanwei Li, Yuechen Zhang, Chengyao Wang, Zhisheng Zhong, Yixin Chen, Ruihang Chu, Shaoteng Liu, and Jiaya Jia. Mini-gemini: Mining the potential of multi-modality vision language models. arXiv preprint arXiv:2403.18814, 2024. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.164, + 0.907, + 0.218 + ], + "angle": 0, + "content": "[37] Bin Lin, Bin Zhu, Yang Ye, Munan Ning, Peng Jin, and Li Yuan. Video-llava: Learning united visual representation by alignment before projection. arXiv preprint arXiv:2311.10122, 2023. 5, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.221, + 0.907, + 0.262 + ], + "angle": 0, + "content": "[38] Haotian Liu, Chunyuan Li, Yuheng Li, Bo Li, Yuanhan Zhang, Sheng Shen, and Yong Jae Lee. Llava-last: Improved reasoning,OCR,and world knowledge,2024.6" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.264, + 0.907, + 0.305 + ], + "angle": 0, + "content": "[39] Haotian Liu, Chunyuan Li, Qingyang Wu, and Yong Jae Lee. Visual instruction tuning. Advances in neural information processing systems, 36, 2024. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.307, + 0.907, + 0.36 + ], + "angle": 0, + "content": "[40] Hao Liu, Wilson Yan, Matei Zaharia, and Pieter Abbeel. World model on million-length video and language with ringattention. arXiv preprint arXiv:2402.08268, 2024. 1, 2, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.363, + 0.907, + 0.433 + ], + "angle": 0, + "content": "[41] Jiasen Lu, Christopher Clark, Sangho Lee, Zichen Zhang, Savya Khosla, Ryan Marten, Derek Hoiem, and Aniruddha Kembhavi. Unified-io 2: Scaling autoregressive multimodal models with vision, language, audio, and action. arXiv preprint arXiv:2312.17172, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.435, + 0.907, + 0.503 + ], + "angle": 0, + "content": "[42] Ruipu Luo, Ziwang Zhao, Min Yang, Junwei Dong, Da Li, Pengcheng Lu, Tao Wang, Linmei Hu, Minghui Qiu, and Zhongyu Wei. Valley: Video assistant with large language model enhanced ability. arXiv preprint arXiv:2306.07207, 2023.5,7" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.505, + 0.907, + 0.56 + ], + "angle": 0, + "content": "[43] Muhammad Maaz, Hanoona Rasheed, Salman Khan, and Fahad Shahbaz Khan. Video-chatgpt: Towards detailed video understanding via large vision and language models. arXiv preprint arXiv:2306.05424, 2023. 5, 6, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.562, + 0.907, + 0.617 + ], + "angle": 0, + "content": "[44] Karttikeya Mangalam, Raiymbek Akshulakov, and Jitendra Malik. Egoschema: A diagnostic benchmark for very long-form video language understanding. Advances in Neural Information Processing Systems, 36:46212-46244, 2023. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.619, + 0.907, + 0.702 + ], + "angle": 0, + "content": "[45] Willi Menapace, Aliaksandr Siarohin, Ivan Skorokhodov, Ekaterina Deyneka, Tsai-Shien Chen, Anil Kag, Yuwei Fang, Aleksei Stoliar, Elisa Ricci, Jian Ren, et al. Snap video: Scaled spatiotemporal transformers for text-to-video synthesis. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 7038-7048, 2024. 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.704, + 0.814, + 0.716 + ], + "angle": 0, + "content": "[46] OpenAI. Gpt-4v(ison) system card, 2023. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.719, + 0.775, + 0.732 + ], + "angle": 0, + "content": "[47] OpenAI. Gpt-4o system card, 2024. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.734, + 0.907, + 0.815 + ], + "angle": 0, + "content": "[48] Viorica Patraucean, Lucas Smaira, Ankush Gupta, Adria Recasens, Larisa Markeeva, Dylan Banarse, Skanda Koppula, Mateusz Malinowski, Yi Yang, Carl Doersch, et al. Perception test: A diagnostic benchmark for multimodal video models. Advances in Neural Information Processing Systems, 36, 2024. 5, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.818, + 0.907, + 0.9 + ], + "angle": 0, + "content": "[49] Alec Radford, Jong Wook Kim, Chris Hallacy, Aditya Ramesh, Gabriel Goh, Sandhini Agarwal, Girish Sastry, Amanda Askell, Pamela Mishkin, Jack Clark, et al. Learning transferable visual models from natural language supervision. In International conference on machine learning, pages 8748-8763. PMLR, 2021. 5" + }, + { + "type": "list", + "bbox": [ + 0.517, + 0.093, + 0.907, + 0.9 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "13615" + } + ], + [ + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.092, + 0.486, + 0.161 + ], + "angle": 0, + "content": "[50] Robin Rombach, Andreas Blattmann, Dominik Lorenz, Patrick Esser, and Björn Ommer. High-resolution image synthesis with latent diffusion models. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 10684-10695, 2022. 2, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.164, + 0.485, + 0.247 + ], + "angle": 0, + "content": "[51] Christoph Schuhmann, Romain Beaumont, Richard Vencu, Cade Gordon, Ross Wightman, Mehdi Cherti, Theo Coombes, Aarush Katta, Clayton Mullis, Mitchell Wortsman, et al. Laion-5b: An open large-scale dataset for training next generation image-text models. Advances in Neural Information Processing Systems, 35:25278–25294, 2022. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.249, + 0.485, + 0.331 + ], + "angle": 0, + "content": "[52] Piyush Sharma, Nan Ding, Sebastian Goodman, and Radu Soricut. Conceptual captions: A cleaned, hypernymed, image alt-text dataset for automatic image captioning. In Proceedings of the 56th Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers), pages 2556-2565, 2018. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.333, + 0.485, + 0.401 + ], + "angle": 0, + "content": "[53] Uriel Singer, Adam Polyak, Thomas Hayes, Xi Yin, Jie An, Songyang Zhang, Qiyuan Hu, Harry Yang, Oron Ashual, Oran Gafni, et al. Make-a-video: Text-to-video generation without text-video data. arXiv preprint arXiv:2209.14792, 2022. 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.404, + 0.485, + 0.473 + ], + "angle": 0, + "content": "[54] Quan Sun, Yufeng Cui, Xiaosong Zhang, Fan Zhang, Qiying Yu, Zhengxiong Luo, Yueze Wang, Yongming Rao, Jingjing Liu, Tiejun Huang, et al. Generative multimodal models are in-context learners. arXiv preprint arXiv:2312.13286, 2023. 2, 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.476, + 0.485, + 0.531 + ], + "angle": 0, + "content": "[55] Quan Sun, Qiying Yu, Yufeng Cui, Fan Zhang, Xiaosong Zhang, Yueze Wang, Hongcheng Gao, Jingjing Liu, Tiejun Huang, and Xinlong Wang. Generative pretraining in multimodality. arXiv preprint arXiv:2307.05222, 2023. 1, 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.533, + 0.485, + 0.588 + ], + "angle": 0, + "content": "[56] Quan Sun, Qiying Yu, Yufeng Cui, Fan Zhang, Xiaosong Zhang, Yueze Wang, Hongcheng Gao, Jingjing Liu, Tiejun Huang, and Xinlong Wang. Generative pretraining in multimodality. arXiv preprint arXiv:2307.05222, 2023." + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.59, + 0.485, + 0.63 + ], + "angle": 0, + "content": "[57] Chameleon Team. Chameleon: Mixed-modal early-fusion foundation models. arXiv preprint arXiv:2405.09818, 2024. 1, 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.633, + 0.485, + 0.701 + ], + "angle": 0, + "content": "[58] Gemini Team, Rohan Anil, Sebastian Borgeaud, Yonghui Wu, Jean-Baptiste Alayrac, Jiahui Yu, Radu Soricut, Johan Schalkwyk, Andrew M Dai, Anja Hauth, et al. Gemini: a family of highly capable multimodal models. arXiv preprint arXiv:2312.11805, 2023. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.704, + 0.485, + 0.773 + ], + "angle": 0, + "content": "[59] Gemini Team, Petko Georgiev, Ving Ian Lei, Ryan Burnell, Libin Bai, Anmol Gulati, Garrett Tanzer, Damien Vincent, Zhufeng Pan, Shibo Wang, et al. Gemini 1.5: Unlocking multimodal understanding across millions of tokens of context. arXiv preprint arXiv:2403.05530, 2024. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.775, + 0.485, + 0.844 + ], + "angle": 0, + "content": "[60] Hugo Touvron, Thibaut Lavril, Gautier Izacard, Xavier Martinet, Marie-Anne Lachaux, Timothée Lacroix, Baptiste Rozière, Naman Goyal, Eric Hambro, Faisal Azhar, et al. Llama: Open and efficient foundation language models. arXiv preprint arXiv:2302.13971, 2023. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.846, + 0.485, + 0.901 + ], + "angle": 0, + "content": "[61] Michael Tschannen, Cian Eastwood, and Fabian Mentzer. Givt: Generative infinite-vocabulary transformers. In European Conference on Computer Vision, pages 292-309. Springer, 2025. 5, 7" + }, + { + "type": "list", + "bbox": [ + 0.093, + 0.092, + 0.486, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.092, + 0.908, + 0.148 + ], + "angle": 0, + "content": "[62] Thomas Unterthiner, Sjoerd Van Steenkiste, Karol Kurach, Raphael Marinier, Marcin Michalski, and Sylvain Gelly. Towards accurate generative models of video: A new metric & challenges. arXiv preprint arXiv:1812.01717, 2018. 5, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.151, + 0.908, + 0.19 + ], + "angle": 0, + "content": "[63] Aaron Van Den Oord, Oriol Vinyals, et al. Neural discrete representation learning. Advances in neural information processing systems, 30, 2017. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.193, + 0.907, + 0.247 + ], + "angle": 0, + "content": "[64] Hanyu Wang, Saksham Suri, Yixuan Ren, Hao Chen, and Abhinav Shrivastava. Larp: Tokenizing videos with a learned autoregressive generative prior. arXiv preprint arXiv:2410.21264, 2024. 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.249, + 0.907, + 0.291 + ], + "angle": 0, + "content": "[65] Wenxuan Wang, Quan Sun, Fan Zhang, Yepeng Tang, Jing Liu, and Xinlong Wang. Diffusion feedback helps clip see better. arXiv preprint arXiv:2407.20171, 2024. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.293, + 0.908, + 0.361 + ], + "angle": 0, + "content": "[66] Xiang Wang, Hangjie Yuan, Shiwei Zhang, Dayou Chen, Jiuniu Wang, Yingya Zhang, Yujun Shen, Deli Zhao, and Jingren Zhou. Videocomposer: Compositional video synthesis with motion controllability. Advances in Neural Information Processing Systems, 36, 2024. 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.363, + 0.907, + 0.419 + ], + "angle": 0, + "content": "[67] Xinlong Wang, Xiaosong Zhang, Zhengxiong Luo, Quan Sun, Yufeng Cui, Jinsheng Wang, Fan Zhang, Yueze Wang, Zhen Li, Qiying Yu, et al. Emu3: Next-token prediction is all you need. arXiv preprint arXiv:2409.18869, 2024. 2, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.421, + 0.908, + 0.488 + ], + "angle": 0, + "content": "[68] Yi Wang, Yinan He, Yizhuo Li, Kunchang Li, Jiashuo Yu, Xin Ma, Xinhao Li, Guo Chen, Xinyuan Chen, Yaohui Wang, et al. Internvid: A large-scale video-text dataset for multimodal understanding and generation. arXiv preprint arXiv:2307.06942, 2023. 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.491, + 0.908, + 0.546 + ], + "angle": 0, + "content": "[69] Yuqing Wang, Tianwei Xiong, Daquan Zhou, Zhijie Lin, Yang Zhao, Bingyi Kang, Jiashi Feng, and Xihui Liu. Loong: Generating minute-level long videos with autoregressive language models. arXiv preprint arXiv:2410.02757, 2024. 2, 6, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.548, + 0.908, + 0.629 + ], + "angle": 0, + "content": "[70] Chen Wei, Karttikeya Mangalam, Po-Yao Huang, Yanghao Li, Haoqi Fan, Hu Xu, Huiyu Wang, Cihang Xie, Alan Yuille, and Christoph Feichtenhofer. Diffusion models as masked autoencoders. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 16284-16294, 2023. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.633, + 0.908, + 0.687 + ], + "angle": 0, + "content": "[71] Chenfei Wu, Lun Huang, Qianxi Zhang, Binyang Li, Lei Ji, Fan Yang, Guillermo Sapiro, and Nan Duan. Godiva: Generating open-domain videos from natural descriptions. arXiv preprint arXiv:2104.14806, 2021. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.689, + 0.908, + 0.758 + ], + "angle": 0, + "content": "[72] Chengyue Wu, Xiaokang Chen, Zhiyu Wu, Yiyang Ma, Xingchao Liu, Zizheng Pan, Wen Liu, Zhenda Xie, Xingkai Yu, Chong Ruan, et al. Janus: Decoupling visual encoding for unified multimodal understanding and generation. arXiv preprint arXiv:2410.13848, 2024. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.76, + 0.907, + 0.8 + ], + "angle": 0, + "content": "[73] Shengqiong Wu, Hao Fei, Leigang Qu, Wei Ji, and Tat-Seng Chua. Next-gpt: Any-to-any multimodal llm. arXiv preprint arXiv:2309.05519, 2023." + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.803, + 0.908, + 0.871 + ], + "angle": 0, + "content": "[74] Yecheng Wu, Zhuoyang Zhang, Junyu Chen, Haotian Tang, Dacheng Li, Yunhao Fang, Ligeng Zhu, Enze Xie, Hongxu Yin, Li Yi, et al. Vila-u: a unified foundation model integrating visual understanding and generation. arXiv preprint arXiv:2409.04429, 2024. 1, 2, 6, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.874, + 0.908, + 0.901 + ], + "angle": 0, + "content": "[75] Weilai Xiang, Hongyu Yang, Di Huang, and Yunhong Wang. Denoising diffusion autoencoders are unified self-supervised" + }, + { + "type": "list", + "bbox": [ + 0.517, + 0.092, + 0.908, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.946, + 0.52, + 0.957 + ], + "angle": 0, + "content": "13616" + } + ], + [ + { + "type": "ref_text", + "bbox": [ + 0.123, + 0.092, + 0.484, + 0.12 + ], + "angle": 0, + "content": "learners. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 15802-15812, 2023. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.121, + 0.484, + 0.188 + ], + "angle": 0, + "content": "[76] Junbin Xiao, Xindi Shang, Angela Yao, and Tat-Seng Chua. Next-qa: Next phase of question-answering to explaining temporal actions. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 9777–9786, 2021. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.19, + 0.484, + 0.259 + ], + "angle": 0, + "content": "[77] Jinheng Xie, Weijia Mao, Zechen Bai, David Junhao Zhang, Weihao Wang, Kevin Qinghong Lin, Yuchao Gu, Zhijie Chen, Zhenheng Yang, and Mike Zheng Shou. Show-o: One single transformer to unify multimodal understanding and generation. arXiv preprint arXiv:2408.12528, 2024. 1, 2, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.261, + 0.484, + 0.342 + ], + "angle": 0, + "content": "[78] Jinbo Xing, Menghan Xia, Yong Zhang, Haoxin Chen, Wangbo Yu, Hanyuan Liu, Gongye Liu, Xintao Wang, Ying Shan, and Tien-Tsin Wong. Dynamiccafter: Animating open-domain images with video diffusion priors. In European Conference on Computer Vision, pages 399-417. Springer, 2025. 2, 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.343, + 0.483, + 0.399 + ], + "angle": 0, + "content": "[79] Jun Xu, Tao Mei, Ting Yao, and Yong Rui. Msr-vtt: A large video description dataset for bridging video and language. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 5288-5296, 2016. 5, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.4, + 0.484, + 0.468 + ], + "angle": 0, + "content": "[80] Jiarui Xu, Sifei Liu, Arash Vahdat, Wonmin Byeon, Xiaolong Wang, and Shalini De Mello. Open-vocabulary panoptic segmentation with text-to-image diffusion models. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 2955-2966, 2023. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.469, + 0.484, + 0.523 + ], + "angle": 0, + "content": "[81] Lin Xu, Yilin Zhao, Daquan Zhou, Zhijie Lin, See Kiong Ng, and Jiashi Feng. Pllava: Parameter-free llava extension from images to videos for video dense captioning. arXiv preprint arXiv:2404.16994, 2024. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.525, + 0.484, + 0.566 + ], + "angle": 0, + "content": "[82] Wilson Yan, Yunzhi Zhang, Pieter Abbeel, and Aravind Srinivas. Videogpt: Video generation using vq-vae and transformers. arXiv preprint arXiv:2104.10157, 2021. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.567, + 0.484, + 0.622 + ], + "angle": 0, + "content": "[83] Jian Yang, Dacheng Yin, Yizhou Zhou, Fengyun Rao, Wei Zhai, Yang Cao, and Zheng-Jun Zha. Mmar: Towards lossless multi-modal auto-regressive probabilistic modeling. arXiv preprint arXiv:2410.10798, 2024. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.623, + 0.484, + 0.677 + ], + "angle": 0, + "content": "[84] Shuai Yang, Yuying Ge, Yang Li, Yukang Chen, Yixiao Ge, Ying Shan, and Yingcong Chen. Seed-story: Multimodal long story generation with large language model. arXiv preprint arXiv:2407.08683, 2024. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.679, + 0.484, + 0.734 + ], + "angle": 0, + "content": "[85] Kexin Yi, Chuang Gan, Yunzhu Li, Pushmeet Kohli, Jiajun Wu, Antonio Torralba, and Joshua B Tenenbaum. Clevrer: Collision events for video representation and reasoning. arXiv preprint arXiv:1910.01442, 2019. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.735, + 0.484, + 0.803 + ], + "angle": 0, + "content": "[86] Lili Yu, Bowen Shi, Ramakanth Pasunuru, Benjamin Muller, Olga Golovneva, Tianlu Wang, Arun Babu, Binh Tang, Brian Karrer, Shelly Sheynin, et al. Scaling autoregressive multimodal models: Pretraining and instruction tuning. arXiv preprint arXiv:2309.02591, 2023. 1, 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.804, + 0.484, + 0.872 + ], + "angle": 0, + "content": "[87] Qiying Yu, Quan Sun, Xiaosong Zhang, Yufeng Cui, Fan Zhang, Yue Cao, Xinlong Wang, and Jingjing Liu. Capsfusion: Rethinking image-text data at scale. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 14022-14032, 2024. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.873, + 0.484, + 0.901 + ], + "angle": 0, + "content": "[88] Zhou Yu, Dejing Xu, Jun Yu, Ting Yu, Zhou Zhao, Yueting Zhuang, and Dacheng Tao. Activitynet-qa: A dataset for" + }, + { + "type": "list", + "bbox": [ + 0.093, + 0.092, + 0.484, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.545, + 0.093, + 0.907, + 0.135 + ], + "angle": 0, + "content": "understanding complex web videos via question answering. In Proceedings of the AAAI Conference on Artificial Intelligence, pages 9127-9134, 2019. 5, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.136, + 0.907, + 0.19 + ], + "angle": 0, + "content": "[89] Chuyang Zhao, Yuxing Song, Wenhao Wang, Haocheng Feng, Errui Ding, Yifan Sun, Xinyan Xiao, and Jingdong Wang. Monoformer: One transformer for both diffusion and autoregression. arXiv preprint arXiv:2409.16280, 2024. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.192, + 0.907, + 0.26 + ], + "angle": 0, + "content": "[90] Wenliang Zhao, Yongming Rao, Zuyan Liu, Benlin Liu, Jie Zhou, and Jiwen Lu. Unleashing text-to-image diffusion models for visual perception. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 5729-5739, 2023. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.262, + 0.907, + 0.331 + ], + "angle": 0, + "content": "[91] Chunting Zhou, Lili Yu, Arun Babu, Kushal Tirumala, Michihiro Yasunaga, Leonid Shamis, Jacob Kahn, Xuezhe Ma, Luke Zettlemoyer, and Omer Levy. Transfusion: Predict the next token and diffuse images with one multi-modal model. arXiv preprint arXiv:2408.11039, 2024. 1, 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.333, + 0.907, + 0.388 + ], + "angle": 0, + "content": "[92] Luowei Zhou, Chenliang Xu, and Jason Corso. Towards automatic learning of procedures from web instructional videos. In Proceedings of the AAAI Conference on Artificial Intelligence, 2018. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.39, + 0.907, + 0.458 + ], + "angle": 0, + "content": "[93] Jinguo Zhu, Xiaohan Ding, Yixiao Ge, Yuying Ge, Sijie Zhao, Hengshuang Zhao, Xiaohua Wang, and Ying Shan. Vl-gpt: A generative pre-trained transformer for vision and language understanding and generation. arXiv preprint arXiv:2312.09251, 2023. 1, 2" + }, + { + "type": "list", + "bbox": [ + 0.517, + 0.093, + 0.907, + 0.458 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "13617" + } + ] +] \ No newline at end of file diff --git a/2025/Divot_ Diffusion Powers Video Tokenizer for Comprehension and Generation/dd9cf0f7-0275-41db-a729-40fdeab212b3_origin.pdf b/2025/Divot_ Diffusion Powers Video Tokenizer for Comprehension and Generation/dd9cf0f7-0275-41db-a729-40fdeab212b3_origin.pdf new file mode 100644 index 0000000000000000000000000000000000000000..3ed9c843db8dcbeea0a2ddd2981888025e6361ad --- /dev/null +++ b/2025/Divot_ Diffusion Powers Video Tokenizer for Comprehension and Generation/dd9cf0f7-0275-41db-a729-40fdeab212b3_origin.pdf @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:52ad2ae48816c604907a39717414c96036ffbf77b9ec1dc8a16ae31ec16603ae +size 2211138 diff --git a/2025/Divot_ Diffusion Powers Video Tokenizer for Comprehension and Generation/full.md b/2025/Divot_ Diffusion Powers Video Tokenizer for Comprehension and Generation/full.md new file mode 100644 index 0000000000000000000000000000000000000000..ebe425ff98d2070b8ef8dd858e0cfca47ccee514 --- /dev/null +++ b/2025/Divot_ Diffusion Powers Video Tokenizer for Comprehension and Generation/full.md @@ -0,0 +1,382 @@ +# Divot: Diffusion Powers Video Tokenizer for Comprehension and Generation + +Yuying Ge Yizhuo Li Yixiao Ge Ying Shan ARC Lab, Tencent PCG https://github.com/TencentARC/Divot + +# Abstract + +In recent years, there has been a significant surge of interest in unifying image comprehension and generation within Large Language Models (LLMs). This growing interest has prompted us to explore extending this unification to videos. The core challenge lies in developing a versatile video tokenizer that captures both the spatial characteristics and temporal dynamics of videos to obtain representations for LLMs, and the representations can be further decoded into realistic video clips to enable video generation. In this work, we introduce Divot, a Diffusion-Powered Video Tokenizer, which leverages the diffusion process for self-supervised video representation learning. We posit that if a video diffusion model can effectively de-noise video clips by taking the features of a video tokenizer as the condition, then the tokenizer has successfully captured robust spatial and temporal information. Additionally, the video diffusion model inherently functions as a de-tokenizer, decoding videos from their representations. Building upon the Divot tokenizer, we present Divot-LLM through video-to-text autoregression and text-to-video generation by modeling the distributions of continuous-valued Divot features with a Gaussian Mixture Model. Experimental results demonstrate that our diffusion-based video tokenizer, when integrated with a pre-trained LLM, achieves competitive performance across various video comprehension and generation benchmarks. The instruction tuned Divot-LLM also excels in video storytelling, generating interleaved narratives and corresponding videos. Models and codes are available at https://github.com/TencentARC/Divot. + +# 1. Introduction + +In recent years, the rapid evolution of Multimodal Large Language Models (MLLMs) [12, 15-17, 55-57, 77, 86, 91, 93] has demonstrated significant progresses in unified image understanding and generation, which empowers LLMs [6, 11, 60] with the ability to generate images beyond texts. While these work primarily focus on image-text data, the extension of this unification to the video domain remains + +![](images/0279fe125a1240893f4d721d6970606d320221f4fa2bb5eb7b575fb6b0ce1b4c.jpg) +Figure 1. We utilize the diffusion procedure to learn a video tokenizer in a self-supervised manner for unified comprehension and generation, where the spatiotemporal representations serve as the condition of a diffusion model to de-noise video clips. Additionally, the proxy diffusion model functions as a de-tokenizer to decode realistic video clips from the video representations. + +relatively under-explored. Achieving unified video comprehension and generation is essential for the development of more sophisticated artificial intelligence (AI) systems that are capable of understanding and creating dynamic visual content in the real world. + +The primary challenge of achieving unified video comprehension and generation lies in developing a versatile video tokenizer that can effectively address the complexities inherent in video data. This tokenizer should be able to obtain robust video representations that serve as inputs of MLLMs for video comprehension, and these representations can be further decoded into realistic video clips to enable video generation. Unlike static images, videos encompass both spatial characteristics and temporal dynamics, making their representation significantly more complex. Recent pioneering work [26, 40, 74] adopt a discrete video tokenizer for unifying video comprehension and generation, where a video is represented as a sequence of discrete frame tokens, or keyframe tokens followed by discrete motion tokens. This approach eases video generation with a LLM through an autoregressive next-token prediction mechanism, but sacrifices the performance of multimodal understanding, as pointed out by recent work [77]. In this work, we aim to investigate an alternative approach by utilizing continuous video representations to unify video comprehension and generation. + +To this end, we introduce Divot, a Diffusion-Powered + +Video Tokenizer that leverages the diffusion process [50] for self-supervised video representation learning as shown in Fig. 1. The core premise is that if a diffusion model can effectively predict the noise added to the Variational Autoencoder (VAE) latents [28] of video clips, when conditioned on the features produced by the video tokenizer, it demonstrates that the tokenizer has successfully captured robust spatial and temporal information inherent in the video data. This capability is crucial for representing the intricate dynamics present in videos. Furthermore, in addition to being a proxy module for learning the tokenizer, the diffusion model can act as a de-tokenizer to effectively decode realistic videos from their learned representations. This dual functionality facilitates a seamless integration of understanding and creating video content within a LLM. + +Specifically, the Divot tokenizer is composed of a pretrained Vision Transformer (ViT) encoder [13], a Spatial-Temporal transformer, and a Perceiver Resampler [1] to obtain video representations from video frames sampled at low frame rate (fps) considering the semantic redundancy between adjacent frames. The video representations serve as the condition of a pre-trained video diffusion model, DynamiCrafter [78] (without the concatenation of a conditional image with initial noise), to predict the noise added to the VAE latents of video frames. After training, the video diffusion model can generate realistic video clips from noise by taking the video representations provided by the Divot tokenizer as the condition. + +We further present Divot-LLM by equipping the pretrained Mistral-7B [24] with the Divot tokenizer. Divot-LLM is pre-trained with a next-word prediction objective on video-caption data by taking the spatiotemporal representations as inputs for video comprehension. The challenge then arises in modeling the continuous video representations with the LLM for video generation. We empirically find that simply minimizing the distance between the LLM output and video representations using mean squared error (MSE) loss achieves unsatisfactory results, since the deterministic regression regularizes the LLM to learn overly averaged features of videos. To address this, inspired by recent work [33], we shift our focus from deterministic modeling to probabilistic modeling by modeling the distributions of video features with a Gaussian Mixture Model (GMM). Specifically, we train the LLM to predict GMM parameters, including means, variances, and mixture probabilities through minimizing the discrepancy between the predicted GMM distribution and the actual video representations using negative log-likelihood (NLL) loss. During inference, we draw samples from the predicted GMM distribution as the condition of the video de-tokenizer to decode video clips. + +We benchmark Divot-LLM on a broad range of video comprehension tasks and zero-shot video generation, achieving competitive performance through pre-training on 5 million video-text pairs. By leveraging the generality of the video tokenizer, our Divot-LLM also enables video storytelling, which generates interleaved narratives and corresponding videos that are temporally coherent through fine-tuning on a specific animation dataset. + +Our contributions are three-fold. (1) We introduce Divot, an advanced video tokenizer that leverages a diffusion procedure for self-supervised video representation learning, aiming to unify video comprehension and generation. (2) We present Divot-LLM, composed of a pre-trained LLM and the Divot tokenizer to enable understanding and generating video content within a single framework. We investigate effective approaches for fitting continuous video representations using the LLM with probabilistic modeling for video generation. (3) We conduct extensive experiments to demonstrate Divot-LLM's competitive performance on existing video comprehension and generation benchmarks, as well as video storytelling. All models and code are released. + +# 2. Related Work + +MLLMs for Comprehension and Generation. With the rapid development of Multimodal Large Language Models (MLLM), recent studies have been working on unified MLLMs [12, 15-17, 25, 26, 36, 40, 41, 54, 55, 57, 67, 72-74, 77, 83, 84, 86, 89, 91, 93] that are capable of multimodal comprehension and generation. To empower LLMs with the capability to generate visual content, existing work primarily employs the following three approaches: (1) utilizing a pre-trained stable diffusion model to generate images conditioned on LLM output (either continuous features or discrete tokens); (2) employing a Vector Quantized (VQ) [63] based decoder to generate visual content from the discrete codes predicted by LLMs; (3) using LLMs to de-noise Gaussian noise through a diffusion process. While most work predominantly focus on the unification of images and texts, some pioneering studies [26, 40, 67, 74] further advance the integration of video comprehension and generation within an LLM through generating videos from discrete codes using a VQ-based decoder, which falls into the second approach. In this work, we adopt the first approach, which involves leveraging a diffusion model to achieve unified video understanding and generation from continuous representations. + +Video Tokenizer in MLLMs. Previous work on video generation with LLMs predominantly employs a discrete video tokenizer to convert video signals into a sequence of quantized tokens. For example, LWM [40] and VILA-U [74] utilize a frame-level tokenizer to discretize each frame into a sequence of codes. VideoPoet [29], Loong [69] and + +![](images/15287b5e77dd53b6bbf1e4a00e7adf0725fc8dcf61978a991b5d9c4799470da6.jpg) +Figure 2. Overview of Divot tokenization and de-tokenization. During training, sparsely sampled video frames are fed into the tokenizer to obtain spatiotemporal representations. These representations serve as the conditions for a U-Net, which is trained to de-noise the noisy VAE latents of densely sampled video frames. During inference, the video representations from the Divot tokenizer can be decoded into realistic video clips with the U-Net. + +![](images/a93c4e7225ea5b543ccb32551a8fe9109bec5f3797448e75c94d2701f6dc2326.jpg) + +Emu3 [67] leverage a 3D CNN architecture, where the encoded spatial-temporal features are quantized into discrete tokens. Video-LaVIT [26] represents video clips as a keyframe followed by extracted motion vectors, obtaining the respective discrete codes. By converting continuous visual signals into discrete tokens, the original next-token prediction mechanism can be adopted to facilitate video generation with an LLM. However, recent work [77] observes a significant performance degradation in multimodal comprehension tasks when discrete representations are used instead of continuous representations. In this work, we introduce a video tokenizer with continuous representations through leveraging the diffusion [50] procedure, enabling it to be effectively integrated with a LLM for unified comprehension and generation. + +Diffusion for Representation Learning. The diffusion process has been explored as a criterion for representation learning. Some works [3, 75, 80, 90] leverage the intermediate activations of pre-trained diffusion models for downstream tasks including classification, segmentation and depth estimation. Other works [23, 65, 70] employ the diffusion model as a proxy module for self-supervised learning, where noisy inputs are de-noised by conditioning on the image representations. This approach encourages the emergence of informative representations that capture key properties and semantics of the images. In this work, to the best of our knowledge, we for the first time leverage diffusion for video representation learning, where a video diffusion model is trained to de-noise video clips through taking the spatiotemporal representations as conditions, thereby encouraging the capture of spatial characteristics and temporal dynamics. + +# 3. Method + +# 3.1. Divot Tokenizer + +We introduce Divot, a diffusion-powered tokenizer that leverage diffusion procedure for video representation learning. + +![](images/2c531e3cd50b981ad0899df0f49e891e16f9be0a616fdad7e20a153fbf4c78e4.jpg) +Figure 3. Overview of Divot-LLM. Video features from the Divot tokenizer are fed into the LLM to perform next-word prediction for video comprehension, while learnable queries are input into the LLM to model the distributions of Divot features using a Gaussian Mixture Model (GMM) for video generation. During inference, video features are sampled from the predicted GMM distribution to decode videos using the de-tokenizer. + +Additionally, the proxy diffusion model used for training the tokenizer can serve as a de-tokenizer to decode realistic video clips from their spatiotemporal representations. + +# 3.1.1. Preliminary: Video Diffusion Model. + +Diffusion models [20, 50] learn to model a probability distribution by reversing a process that progressively adds noise to the data. Specifically, given data $\mathbf{x}_0 \sim p(\mathbf{x})$ , the forward process gradually adds random Gaussian noise $\epsilon_{\mathbf{t}} \in \mathcal{N}(\mathbf{0}, \mathbf{I})$ to the data sample $\mathbf{x}_0$ with a total of $T$ timesteps to yield $\mathbf{x}_t$ through a parameterization trick. The denoising process predicts $\epsilon_{\mathbf{t}}$ in the forward diffusion process with a denoising network $\epsilon_{\theta}(\mathbf{x}_t, t)$ , which is trained by the objective below, + +$$ +\min _ {\theta} \mathbb {E} _ {t, \mathbf {x} \sim p, \epsilon \sim \mathcal {N} (\mathbf {0}, \mathbf {I})} \| \epsilon - \epsilon_ {\theta} (\mathbf {x} _ {t}, t) \| _ {2} ^ {2}, \tag {1} +$$ + +where $\epsilon$ is the sampled Gaussian noise and $\theta$ indicates the parameters of the denoising network. During inference, we can perform iterative denoising from a random Gaussian + +![](images/3f1baa720b8b0a5f515f4ed0d31093760eece10aa567131fefa68bced55b38b9.jpg) +Figure 4. Paradigms for modeling video representations from the Divot tokenizer with a LLM for video generation. (a) MSE Regression, where the LLM output is trained to minimize its distance with video features using Mean Squared Error (MSE) loss; (b) Diffusion Modeling, where the LLM output is fed into a denoising network as the condition to predict the noise added to video features; (c) GMM Modeling, where the LLM output is trained to predict the parameters of a Gaussian Mixture Model (GMM) for modeling video feature distributions. + +![](images/3369033bdb1d68330b8fe8c23297d80ba63ecc490d094b81f8eed65ce4c81068.jpg) + +![](images/02be3176e5e209bde863f639edd20d1be47c8279641f639ac8d146a2205a4479.jpg) + +noise for the denoised data $\mathbf{x}_0$ + +For video diffusion models [8, 78], given a video $\mathbf{x}$ , a latent representation $\mathbf{z} = \mathcal{E}(\mathbf{x})$ is first encoded to reduce the computational complexity. Then the forward diffusion process and backward denoise process are performed in this latent space with a denoising network $\epsilon_{\theta}(\mathbf{x}_t,\mathbf{c},t)$ , where $\mathbf{c}$ denotes denoising conditions like text or visual prompts. + +# 3.1.2. Training Pipeline + +As illustrated in Fig. 2, given a video clip, we separately sample sparse frames at 2 fps to obtain the video representations from the tokenizer, and sample dense frames at 8 fps to obtain latent representations $z_{0}$ from the frozen VAE [28] encoder. Sparse frames are sampled as the input of the video tokenizer considering the semantic redundancy between adjacent frames. The forward diffusion process gradually adds Gaussian noise $\theta$ to $z_{0}$ for producing the noisy input $z_{t}$ . At each backward step $t$ , a denoising U-Net is trained to predict the noise added from the previous step to the current step by taking the time embedding and video representations as the condition. Specifically, the video representations interact with the denoising U-Net intermediate features through cross-attention layers, where each noisy latent attends to all video tokens. By constraining the U-Net to reconstruct fine-grained spatial and temporal information of video clips through relying on video features, the Divot tokenizer is optimized to capture both spatial characteristics and temporal dynamics for robust video representations. The Divot tokenizer is trained on pure videos of a subset of WebVid10M [2] and Panda-70M [9], totaling 10M videos. + +After training the Divot tokenizer, the proxy denoising U-Net (employed to implement the parameterized loss function) can serve as an effective video de-tokenizer, which is able to decode semantically aligned video clips from the learned spatiotemporal representations as shown in Fig. 5. + +# 3.1.3. Model Architecture + +As shown in Fig. 2, the Divot tokenizer is composed of a pre-trained ViT encoder to extract frame-level features, a transformer for spatial and temporal fusion, and a Perceiver Resampler [1] to produce a fixed number of video tokens. The Perceiver Resampler is adopted for two reasons: (1) to reduce the number of video tokens that a LLM need to predict for generation, and (2) to transform the patch-position dependent features into a sequence of high-level features without 3D positional dependencies, which we empirically find easier for an LLM to fit (See Sec. 4.3). Specifically, given a video clip with a duration of two seconds, we sample 5 frames at 2 fps, resulting in a total of 64 video tokens. We adopt the de-noising U-Net in DynamiCrafter [78], but reduce the input channel of the 3D convolution from 8 to 4 since we remove the original concatenation of a conditional image with noisy latents. + +# 3.2. Video Representation Modeling with LLM + +The core challenge of generating videos using a LLM with the Divot tokenizer lies in effectively modeling the continuous video features. The most straightforward solution is to minimize the distance between the LLM output and the video representations using mean squared error (MSE) loss following previous work [17, 54] for image generation as illustrated in Fig. 4 (a). However, we empirically find that this approach is not effective for modeling continuous video features, as the generated videos tend to exhibit repeating patterns. We analyze that the deterministic regression regularizes the LLM to learn overly averaged representations, which is particularly catastrophic in video generation as videos must ensure both spatial and temporal diversity. + +Inspired by recent work MAR [33], instead of deterministic regression, we aim to model the probability distribution of video representations using the LLM. As shown in Fig. 4, + +![](images/e028b6c9192fbea1821cc07ff0fe92c835c98e38529b8974e24655cf6c2b4978.jpg) +Input video (fps=2) + +![](images/4fd0c41406fe355eb7d0c864097144f86308dd90d94c8798281db3ccce5449d7.jpg) + +![](images/c76ac15c2c3f96328f358f06af6cc5f10fa785219565aa151ac4116a61eb3b50.jpg) +Figure 5. Reconstructed videos, where the Divot tokenizer obtains spatiotemporal representations of sparsely sampled video frames and the de-tokenizer decodes these representations into semantically aligned and temporally coherent video clips. + +![](images/96a417e54fcba8ad56e881dc765b9369ff474d383b2dc1fff2f1c716947ee215.jpg) +Reconstructed video (fps=8) + +![](images/8bbd5fc6d4255b8814b2441e99df74428962677b3f768c5416b9b9d40b5e46b2.jpg) + +![](images/d57101b1d7c19c1fa8232c33f41382eece25fe7e8079fa2ddec2ba631fb64ace.jpg) + +we explore two approaches including (b) Diffusion Modeling [33] and (c) GMM Modeling [61]. Specifically, for the diffusion modeling, given continuous-valued video features to be predicted, the LLM produces output, which serves as the condition of a denoising network (a small MLP) to predict the Gaussian noise added to the video features. The diffusion model is trained for representing the distribution of video features. For GMM modeling, we use a Gaussian Mixture Model (GMM) to model the distribution of the video features, and train the LLM to predict $2kd + k$ parameters per video token ( $kd$ mean and $kd$ variance parameters for the mixture components, and $k$ mixture probabilities). We optimize the LLM by minimizing the discrepancy between the predicted GMM distribution and the video representations with negative log-likelihood (NLL) loss. + +During inference, in diffusion modeling, the denoising network denoise the final video features from Gaussian noise gradually by taking the LLM output as the condition. In GMM modeling, we draw samples from the predicted GMM distribution as the final video representations. To empirically investigate the effectiveness of the approaches above, we train the LLM with MSR-VTT [79], and evaluates text-to-video generation on test set with FVD [62] and similarity score [49] as the metric following previous work [26, 82]. As listed in Tab. 7, GMM modeling achieves better performance than diffusion modeling and MSE regression in video generation. We speculate that high-level features obtained by the Divot tokenizer are more sensitive to Gaussian noise compared to the VAE latents used by MAR, making training more challenging and resulting in suboptimal results. Therefore, we adopt GMM modeling to train Divot-LLM. + +Table 2. Datasets used for training the tokenizer and Divot-LLM. + +
StageTypeDataset
TokenizerPure VideoWebVid-10M [2], Panda-70M [9]
Pre-trainVideo-textWebVid-10M [2]
Image-textCC3M [52], CapsFusion [87], LAION-COCO [51]
SFTClassificationKinetics-710 [27], SSV2 [18]
VQATGIF [34], NextQA [76], CLEVRER [85], YouCook2 [92], PerceptionTest[48], EgoQA [19], ActivityNetQA[88]
InstructionVideo-ChatGPT[43], LLaVA-mixed[39], Valley [42], LLaVA-Video-178K[37]
GenerationWebVid-10M [2]
StoryTellingIn-house data
+ +# 3.3. Pre-training and Instruction Tuning + +# 3.3.1. Training Stage I: Multimodal Pre-training + +As shown in Fig 3, Divot-LLM adopts next-word prediction and GMM modeling on video-text data for video comprehension and generation. Specifically, the video features from the Divot tokenizer, the special tokens indicating the start and end of video features, along with the text tokens of the caption are fed into the pre-trained Mistral-7B [24] for next token prediction trained with cross-entropy loss. Text tokens of the caption and $N$ learnable queries are input into the LLM, where the output of the learnable queries are trained via bidirectional attention to model a GMM distribution for the video features using NLL loss. During inference, we draw samples from the predicted GMM distribution as condition of the denoising U-Net to decode videos. We pre-train Divot-LLM from the pre-trained Mistral-7B model using + +Table 3. Comparison for video comprehension with MLLMs. "Video-Gen" denotes whether the model can generate videos besides texts. The evaluation metric is accuracy. The best results are bold and the second best results are underlined. + +
ModelLLM sizeVideo-GenEgoSchemaPerception-TestMVBenchMSVDActivityNet
Gemini 1.0 Pro [58]-×55.751.1--49.8
Gemini 1.5 Pro [59]-×63.2---56.7
GPT4-V [46]-×55.6-43.7-59.5
GPT4-O [47]-×72.2---61.9
LLaMA-VID [35]7B×38.544.641.969.747.4
Video-ChatGPT [43]7B×---64.935.2
Video-LLaVA [37]7B×38.444.341.070.745.3
VideoChat2 [31]7B×42.247.351.170.049.1
LLaVA-NeXT-Video [38]7B×43.948.846.567.853.5
LLaVA-NeXT-Video [38]32B×60.9---54.3
PLLAVA [81]34B×-58.1--60.9
LLaVA-OneVision [30]72B×62.0---62.3
VideoLLaMA2 [10]7B×51.751.454.670.950.2
VideoLLaMA2 [10]72B×63.957.562.071.055.2
LWM [40]7B✓---55.9-
Video-LaVIT [26]7B✓37.347.9-73.250.1
VILA-U [74]7B✓---75.352.7
Divot-LLM7B✓46.558.352.176.455.8
+ +LoRA [22] on a subset of WebVid-10M [2] data (filtered for temporal dynamics in captions) and image-text data. + +# 3.3.2. Training Stage II: Multimodal Instruction Tuning + +We perform multimodal instruction tuning on Divot-LLM to align it with human instructions through supervised fine-tuning on public datasets as listed in Tab. 2 with a LoRA module. We further fine-tune the pretrained Divot-LLM on an animated series called "Curious George" to achieve video storytelling, which generates storyline and corresponding video clips in an interleaved manner. + +# 4. Experiment + +# 4.1.Quantitative Evaluation + +Video Comprehension. We conduct extensive evaluations on video comprehension benchmarks including Multi-choice Video Question Answering (MC-VQA) on EgoSchema [44], Perception-Test [48], MVBench [32], and Open-Ended Video Question Answering (OE-VQA) on MSVD [7], ActivityNet [88]. Following VideoLLaMA 2 [10], we utilize GPT-3.5 to assess the quality of the generated answers of OE-VQA by determining whether the answers match the ground truth, and we report the percentage of "Yes" as Accuracy. + +For each testing video, we sample a maximum of 20 clips, each containing 5 frames. The evaluation results are reported in Tab. 3. Divot-LLM outperforms the baseline models that can generate both texts and videos, which demonstrates that our model effectively achieves video comprehension within a unified framework. Compared to VideoLLMs specifically designed for video comprehension of the same model size of LLM, Divot-LLM achieves competitive results with sig + +nificantly fewer video-caption pairs for training (4.8M vs. 100M in VideoLLaMA 2). By utilizing diffusion procedure for video representation learning, our Divot tokenizer effectively captures robust spatiotemporal representations, enhancing the comprehension capabilities. + +Video Generation. We evaluate zero-shot text-to-video generation on MSR-VTT [79]. We randomly sample one caption for each testing video and generate 16 frames in $256 \times 256$ px resolution. We adopt the CLIP similarity (CLIPSIM) [71] and Fréchet video distance (FVD) [62] as the evaluation metric following Loong [69]. As listed in Tab. 4, Divot-LLM achieves performance comparable to existing video generation models in terms of visual quality and semantic alignment with captions using only 4.8 million video-text pairs for training. + +# 4.2. Qualitative Evaluation + +Text-to-video Generation. We perform qualitative comparison of text-to-video generation with baseline MLLMs that are capable of unified video comprehension and generation. As illustrated in Fig. 6, through modeling the distributions of Divot features with a predicted GMM, our Divot-LLM can generate videos that are both semantically aligned with text prompts and temporally coherent within frames. + +Video StoryTelling. We fine-tune the pre-trained Divot-LLM on an animated series called "Curious George" for video storytelling. As shown in Fig. 7, given a brief story instruction, our Divot-LLM can generate a sequence of multimodal stories with rich narrative text and contextually relevant videos that are temporally coherent. Since we only + +![](images/0b02b84649d532ae8e5681d4501877047acbf4edb7aeed1b24b63c74821bc370.jpg) +A man walking slowly under a sunset. + +![](images/a72e36cf9679763404ae7e2a336cd67d6453c7b60a526a4c8db7194d3542a39d.jpg) +A person is applying eye makeup. +Figure 6. Qualitative comparison of text-to-video generation with MLLMs that are capable of unified video comprehension and generation. Divot-LLM effectively generates videos that are both semantically aligned with text prompts and temporally coherent within frames. + +fine-tune the de-tokenizer for adaptation to the new domain, it demonstrates the generalizability of our Divot tokenizer for obtaining robust video representations. + +Table 4. Comparison for zero-shot text-to-video generation. "Data size" refers to the number of training video data, and "Unified" denotes if the model enables video comprehension and generation. The best results are bold and the second best results are underlined. + +
ModelData sizeUnifiedMSR-VTT
CLIPSIM (↑)FVD (↓)
CogVideo [21]5.4M×0.26311294
Video LDM [5]10M×0.2929-
VideoComposer [66]10M×0.2932580
InternVid [68]28M×0.2951-
Make-A-Video [53]20M×0.3049-
VideoPoet [29]270M×0.3049213
PYoCo [14]22.5M×--
SVD [4]152M×--
Video-LavIT [26]10M✓0.3012188.36
Loong [69]16M×0.2903274
Snap Video [45]-×0.2793110.4
VILA-U [74]1M✓0.2937499.06
Divot-LLM4.8M✓0.2938301.4
+ +# 4.3. Ablation Study + +Diffusion for Video Comprehension. We design two baselines to validate the effectiveness of the diffusion procedure to learn spatiotemporal representations for VideoLLMs. As shown in Tab. 5, both models are pre-trained on Valley [42] and instruction tuned on Video-ChatGPT [43]. The model with diffusion loss employs our Divot tokenizer, while the model with caption loss adopts the same architecture but its tokenizer is pre-trained using captioning loss with a frozen LLM on Valley. The model that employs a video tokenizer + +Table 6. Ablation study on the training objective of the video tokenizer. The evaluation metric is accuracy. + +
Loss TypeMV-BenchMSVDActivityNet
Caption30.866.143.2
Diffusion33.268.944.3
+ +trained with diffusion loss achieves better performance in video comprehension benchmarks, demonstrating that the diffusion process can effectively learn robust video representations in a self-supervised manner, without the need for paired caption annotations. + +Video Generation with LLM. We perform various ablation studies to explore an effective approach for generating videos with a LLM through training on MSR-VTT training set and evaluating text-to-video generation on test set. We use ViT-G/14 to calculate CLIPSIM for better discrimination. + +Q1: Which type of video representations is easier? We investigate two types of video representations, patch-position dependent features obtained from a spatial-temporal transformer and patch-position independent features after a Perceiver Resample with learnable queries. As listed in Tab. 7, fitting features without 3D positional dependencies achieves higher performance, which is also observed in recent work [64]. We also experiment with training the video tokenizer in the VAE manner, which involves predicting the means and variances of a normal distribution and sampling video representations using the re-parametrization trick following GIVT [61]. However, we observe that it is difficult for the LLM to converge and the video de-tokenizer achieves unsatisfactory reconstruction results. We conclude that introducing variances during tokenization for high-level video features may not be appropriate. + +![](images/7022943fd68f23b1372a3c781dd5842396777c6dec5f8e41997da890f68aa810.jpg) +Generate a story about the adventures of George and his friends. + +A young girl found a glowing orb floating in the air through her magnifying glass. + +![](images/166b68bf5ff8b78bdad2ad3ffdb815b017565f5380946f18a970cf0a8f9905ec.jpg) +REC + +![](images/69b87821acbc5e3da29188c4bca7b83be87ac446d3727123fc2707af0d18b3cb.jpg) + +![](images/2b4d3b08741527aac498461eeb2aaf3e64018eaf6239c3d87bd71bdf618cba05.jpg) + +![](images/cf4fc4340b189061969660e8c73d648e7afa42af4aaa7b7c5ab443e98007191e.jpg) + +![](images/58d0788f6a4f1921e7abce5a5b43bcc982c73f5efbfeaa6e35bd3d1e7a28cff5.jpg) + +George watched a bubble float by, feeling curious and joyful. + +![](images/663b0c511275a713084ce388d5e58858ad353e5c439a03fc798bbea67e5a7d5e.jpg) +REC + +![](images/6571762ac9125fde0225c61cb424cf678e8ab746d45ee3ea2257e35bf1df6de9.jpg) + +![](images/732d370b18165ad577165a633aaef6d61e0ed000fc3db1b4c15ea1d242026f7d.jpg) + +![](images/a73b7c641c469b9ad59751c25596f4e2c5e5d5f575f32333cfc260c2ed1f1d12.jpg) + +![](images/d114e83331db527c59ec8e749906a2a7e4644a2aae4b0100d3f802abbed57ec0.jpg) + +At lunchtime, a boy found a giant sandwich and felt excited. + +![](images/ca2be2eabe2d27eac255c7ae80637d81d6f21998f07b2ca0c21163cb09471cdc.jpg) +REC +Figure 7. Qualitative examples of video storytelling by Divot-LLM. Given a story instruction, Divot-LLM can generate rich textual narratives along with corresponding video clips that are temporally coherent in an interleaved manner. + +![](images/2b90b82b5ddc80533a7301225c5f54c341b591e9d15d784749954fd092069cfa.jpg) + +![](images/cc0e86a83283ba3f37a38fec425b26e09d972fdb58e269108aa9b2abb107a5b8.jpg) + +![](images/82930692f9ad0fa4d1d4614bdfa912a1a08edf357afd5414070d44d6853bdc26.jpg) + +# Generate a story about George's fun-filled day in the kitchen. + +![](images/3d06225ce82fabf69e2ea3482d03d0af3497b5ffb1c7919492f34fd49b8a495b.jpg) + +George and his pal joyfully cook in the kitchen, creating a tasty snack with a big blue book. + +![](images/5f2a88b8f3450114f09b9326cada4bff1fd37d2023e4338c2d521444ac7931fc.jpg) +REC + +![](images/142850e5fd9df78388957622b4ac806287f40852b5b651259457f944ca74f905.jpg) + +![](images/5a1ece1ff70131622e975eaf4fc2a7e230536045fa6d74f54f69e9bd3dc8ffd6.jpg) + +![](images/b5b94209683463df72a5d9a437d2130b0ee199beaea00fe41075d21d27cfc245.jpg) + +![](images/27f2070d8c1bbbb8732270a15cdd8b089e19fa85a03f0c20a666e8e316d77606.jpg) + +A woman in the kitchen shows George how to use a new kitchen gadget. + +![](images/792a57c607f69d7d4714ffb0e08a87a6d479a29703eeaac9e39eb9b37ca61a83.jpg) +REC + +![](images/573a49f44d4708baba1104170e127f775fa5499d3154dd79b680b4272da71135.jpg) + +![](images/10ebc2477a30eb167a1a61be785ae2fc87fc190d81bbab877347a8daf41092ed.jpg) + +![](images/9133559835d3ddca4df6e59957ddaad9a07415d27411d881bcdf6ebbb41c1f19.jpg) + +![](images/709276f49ff6be0a6508a8f0d021aa8daa76f8f95b7ec0f644ddea61a57e15ac.jpg) + +George spreads the thick, peanutty paste, making a yummy snack. + +![](images/0c1a6a68b2ba9aa8a2cb86e28410f151e6a9743ba58489eeb286f11af4ec6b36.jpg) +REC + +![](images/b7d19921d1fe7dd12f36779294f4c13045e76750383938f4cecad3a39730bf4f.jpg) + +![](images/a2f8f9a16452ed4ac767358618255562d029d7f8da9e1521a0722e6c68944bfd.jpg) + +![](images/e772131448626a9c7e4d15c5862b39644a727aaf1dcd316b35ace68f53b944a3.jpg) + +Table 8. Ablation study on video representation modeling with LLMs for generation. We evaluate text-to-video generation on MSR-VTT. + +
RepresentationObjectiveMechanism
patch-position dependentpatch-position independentMSEDiffusionGMMARQuery
ε-predv-predcausalbidirectional
CLIPSIM (↑)0.31920.32650.31680.28110.28420.32650.23860.30800.3265
FVD (↓)378.50366.60438.94418.19377.17366.60447.88416.60366.60
+ +Q2: Which training objective is more suitable? As introduce in Sec. 3, we explore MSE regression, Diffusion modeling and GMM modeling to fit high-level continuous features with a LLM. As listed in Tab. 7, simply aligning the LLM output with video features using MSE loss yields the lowest generation quality, suggesting that deterministic regression is inadequate for modeling spatiotemporal representations. Training a denoising network to denoise the noisy video features by taking the LLM output as the condition also achieves inferior performance with both $\epsilon$ prediction and $v$ prediction. Different from MAR [33] that denoises low-level VAE latents, our goal is to denoise high-level video features. We speculate that these features are more sensitive to Gaussian noise, making them more challenging to denoise. Training the LLM to model the distribution of high-level video features using a GMM model achieves the best generation quality and semantic alignment with captions. + +Q3: Which LLM mechanism is more effective? We train the LLM to fit the video representations with GMM modeling using both an autoregressive approach and a query-based approach, with the latter exploring causal attention and bidirectional attention within the LLM. Predicting the features of each video token in an autoregressive manner results in the worst performance due to the accumulation of errors, particularly when the features of the pervious to- + +ken are sampled from a GMM distribution for predicting the distribution of the current token. The query-based approach achieves better results with bidirectional attention, as it enables each query to attend to all tokens for predictions. + +# 5. Conclusion + +In this work, we introduce Divot, a diffusion-powered video tokenizer learned in a self-supervised manner for unified comprehension and generation. We further investigate effective approaches for modeling continuous video representations with the LLM and present Divot-LLM to understand and generate video content in a single framework. Divot-LLM achieves competitive performance in video comprehension and generation benchmarks, and enables video storytelling effectively. We hope our work will draw increased attention to unifying video comprehension and generation through the design of sophisticated tokenizers. + +Limitation. As we primarily focus on exploring effective representations and approaches for video generation with a unified LLM, the current model is trained to predict video representations for only a single clip and does not generate longer videos, which will be explored in our future work. + +# References + +[1] Jean-Baptiste Alayrac, Jeff Donahue, Pauline Luc, Antoine Miech, Iain Barr, Yana Hasson, Karel Lenc, Arthur Mensch, Katherine Millican, Malcolm Reynolds, et al. Flamingo: a visual language model for few-shot learning. Advances in Neural Information Processing Systems, 35:23716-23736, 2022. 2, 4 +[2] Max Bain, Arsha Nagrani, Gül Varol, and Andrew Zisserman. Frozen in time: A joint video and image encoder for end-to-end retrieval. In Proceedings of the IEEE/CVF international conference on computer vision, pages 1728-1738, 2021. 4, 5, 6 +[3] Dmitry Baranchuk, Ivan Rubachev, Andrey Voynov, Valentin Khrulkov, and Artem Babenko. Label-efficient semantic segmentation with diffusion models. arXiv preprint arXiv:2112.03126, 2021. 3 +[4] Andreas Blattmann, Tim Dockhorn, Sumith Kulal, Daniel Mendelevitch, Maciej Kilian, Dominik Lorenz, Yam Levi, Zion English, Vikram Voleti, Adam Letts, et al. Stable video diffusion: Scaling latent video diffusion models to large datasets. arXiv preprint arXiv:2311.15127, 2023. 7 +[5] Andreas Blattmann, Robin Rombach, Huan Ling, Tim Dockhorn, Seung Wook Kim, Sanja Fidler, and Karsten Kreis. Align your latents: High-resolution video synthesis with latent diffusion models. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 22563-22575, 2023. 7 +[6] Tom Brown, Benjamin Mann, Nick Ryder, Melanie Subbiah, Jared D Kaplan, Prafulla Dhariwal, Arvind Neelakantan, Pranav Shyam, Girish Sastry, Amanda Askell, et al. Language models are few-shot learners. Advances in neural information processing systems, 33:1877-1901, 2020. 1 +[7] David Chen and William B Dolan. Collecting highly parallel data for paraphrase evaluation. In Proceedings of the 49th annual meeting of the association for computational linguistics: human language technologies, pages 190-200, 2011. 6 +[8] Haoxin Chen, Menghan Xia, Yingqing He, Yong Zhang, Xiaodong Cun, Shaoshu Yang, Jinbo Xing, Yaofang Liu, Qifeng Chen, Xintao Wang, et al. Videocrafter1: Open diffusion models for high-quality video generation. arXiv preprint arXiv:2310.19512, 2023. 4 +[9] Tsai-Shien Chen, Aliaksandr Siarohin, Willi Menapace, Ekaterina Deyneka, Hsiang-wei Chao, Byung Eun Jeon, Yuwei Fang, Hsin-Ying Lee, Jian Ren, Ming-Hsuan Yang, et al. Panda-70m: Captioning 70m videos with multiple cross-modality teachers. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 13320–13331, 2024. 4, 5 +[10] Zesen Cheng, Sicong Leng, Hang Zhang, Yifei Xin, Xin Li, Guanzheng Chen, Yongxin Zhu, Wenqi Zhang, Ziyang Luo, Deli Zhao, et al. Videollama 2: Advancing spatial-temporal modeling and audio understanding in video-llms. arXiv preprint arXiv:2406.07476, 2024. 6 +[11] Aakanksha Chowdhery, Sharan Narang, Jacob Devlin, Maarten Bosma, Gaurav Mishra, Adam Roberts, Paul Barham, Hyung Won Chung, Charles Sutton, Sebastian Gehrmann, + +et al. Palm: Scaling language modeling with pathways. arXiv preprint arXiv:2204.02311, 2022. 1 +[12] Runpei Dong, Chunrui Han, Yuang Peng, Zekun Qi, Zheng Ge, Jinrong Yang, Liang Zhao, Jianjian Sun, Hongyu Zhou, Haoran Wei, et al. Dreamllm: Synergistic multimodal comprehension and creation. arXiv preprint arXiv:2309.11499, 2023. 1, 2 +[13] Alexey Dosovitskiy, Lucas Beyer, Alexander Kolesnikov, Dirk Weissenborn, Xiaohua Zhai, Thomas Unterthiner, Mostafa Dehghani, Matthias Minderer, Georg Heigold, Sylvain Gelly, et al. An image is worth 16x16 words: Transformers for image recognition at scale. arXiv preprint arXiv:2010.11929, 2020. 2 +[14] Songwei Ge, Seungjun Nah, Guilin Liu, Tyler Poon, Andrew Tao, Bryan Catanzaro, David Jacobs, Jia-Bin Huang, Ming-Yu Liu, and Yogesh Balaji. Preserve your own correlation: A noise prior for video diffusion models. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 22930-22941, 2023. 7 +[15] Yuying Ge, Yixiao Ge, Ziyun Zeng, Xintao Wang, and Ying Shan. Planting a seed of vision in large language model. arXiv preprint arXiv:2307.08041, 2023. 1, 2 +[16] Yuying Ge, Sijie Zhao, Ziyun Zeng, Yixiao Ge, Chen Li, Xintao Wang, and Ying Shan. Making llama see and draw with seed tokenizer. arXiv preprint arXiv:2310.01218, 2023. +[17] Yuying Ge, Sijie Zhao, Jinguo Zhu, Yixiao Ge, Kun Yi, Lin Song, Chen Li, Xiaohan Ding, and Ying Shan. Seed-x: Multimodal models with unified multi-granularity comprehension and generation. arXiv preprint arXiv:2404.14396, 2024. 1, 2, 4 +[18] Raghav Goyal, Samira Ebrahimi Kahou, Vincent Michalski, Joanna Materzynska, Susanne Westphal, Heuna Kim, Valentin Haenel, Ingo Fruend, Peter Yianilos, Moritz Mueller-Freitag, et al. The" something something" video database for learning and evaluating visual common sense. In Proceedings of the IEEE international conference on computer vision, pages 5842-5850, 2017. 5 +[19] Kristen Grauman, Andrew Westbury, Eugene Byrne, Zachary Chavis, Antonino Furnari, Rohit Girdhar, Jackson Hamburger, Hao Jiang, Miao Liu, Xingyu Liu, et al. Ego4d: Around the world in 3,000 hours of egocentric video. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 18995-19012, 2022. 5 +[20] Jonathan Ho, Ajay Jain, and Pieter Abbeel. Denoising diffusion probabilistic models. Advances in neural information processing systems, 33:6840-6851, 2020. 3 +[21] Wenyi Hong, Ming Ding, Wendi Zheng, Xinghan Liu, and Jie Tang. Cogvideo: Large-scale pretraining for text-to-video generation via transformers. arXiv preprint arXiv:2205.15868, 2022. 7 +[22] Edward J Hu, Yelong Shen, Phillip Wallis, Zeyuan Allen-Zhu, Yanzhi Li, Shean Wang, Lu Wang, and Weizhu Chen. Lora: Low-rank adaptation of large language models. arXiv preprint arXiv:2106.09685, 2021. 6 +[23] Drew A Hudson, Daniel Zoran, Mateusz Malinowski, Andrew K Lampinen, Andrew Jaegle, James L McClelland, Loic Matthew, Felix Hill, and Alexander Lerchner. Soda: Bottle- + +neck diffusion models for representation learning. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 23115-23127, 2024. 3 +[24] Albert Q Jiang, Alexandre Sablayrolles, Arthur Mensch, Chris Bamford, Devendra Singh Chaplot, Diego de las Casas, Florian Bressand, Gianna Lengyel, Guillaume Lample, Lucile Saulnier, et al. Mistral 7b. arXiv preprint arXiv:2310.06825, 2023. 2, 5 +[25] Yang Jin, Kun Xu, Liwei Chen, Chao Liao, Jianchao Tan, Bin Chen, Chenyi Lei, An Liu, Chengru Song, Xiaogiang Lei, et al. Unified language-vision pretraining with dynamic discrete visual tokenization. arXiv preprint arXiv:2309.04669, 2023. 2 +[26] Yang Jin, Zhicheng Sun, Kun Xu, Liwei Chen, Hao Jiang, Quzhe Huang, Chengru Song, Yuliang Liu, Di Zhang, Yang Song, et al. Video-lavit: Unified video-language pre-training with decoupled visual-motional tokenization. arXiv preprint arXiv:2402.03161, 2024. 1, 2, 3, 5, 6, 7 +[27] Will Kay, Joao Carreira, Karen Simonyan, Brian Zhang, Chloe Hillier, Sudheendra Vijayanarasimhan, Fabio Viola, Tim Green, Trevor Back, Paul Natsev, et al. The kinetics human action video dataset. arXiv preprint arXiv:1705.06950, 2017. 5 +[28] Diederik P Kingma and Max Welling. Auto-encoding variational bayes. arXiv preprint arXiv:1312.6114, 2013. 2, 4 +[29] Dan Kondratyuk, Lijun Yu, Xiuye Gu, José Lezama, Jonathan Huang, Grant Schindler, Rachel Hornung, Vighnesh Birodkar, Jimmy Yan, Ming-Chang Chiu, et al. Videopoet: A large language model for zero-shot video generation. arXiv preprint arXiv:2312.14125, 2023. 2, 7 +[30] Bo Li, Yuanhan Zhang, Dong Guo, Renrui Zhang, Feng Li, Hao Zhang, Kaichen Zhang, Yanwei Li, Ziwei Liu, and Chunyuan Li. Llava-onevision: Easy visual task transfer. arXiv preprint arXiv:2408.03326, 2024. 6 +[31] Kunchang Li, Yali Wang, Yinan He, Yizhuo Li, Yi Wang, Yi Liu, Zun Wang, Jilan Xu, Guo Chen, Ping Luo, et al. Mybench: A comprehensive multi-modal video understanding benchmark. arXiv preprint arXiv:2311.17005, 2023. 6 +[32] Kunchang Li, Yali Wang, Yinan He, Yizhuo Li, Yi Wang, Yi Liu, Zun Wang, Jilan Xu, Guo Chen, Ping Luo, et al. Mybench: A comprehensive multi-modal video understanding benchmark. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 22195-22206, 2024. 6 +[33] Tianhong Li, Yonglong Tian, He Li, Mingyang Deng, and Kaiming He. Autoregressive image generation without vector quantization. arXiv preprint arXiv:2406.11838, 2024. 2, 4, 5, 8 +[34] Yuncheng Li, Yale Song, Liangliang Cao, Joel Tetreault, Larry Goldberg, Alejandro Jaimes, and Jiebo Luo. Tgif: A new dataset and benchmark on animated gif description. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition, pages 4641-4650, 2016. 5 +[35] Yanwei Li, Chengyao Wang, and Jiaya Jia. Llama-vid: An image is worth 2 tokens in large language models. arXiv preprint arXiv:2311.17043, 2023. 6 + +[36] Yanwei Li, Yuechen Zhang, Chengyao Wang, Zhisheng Zhong, Yixin Chen, Ruihang Chu, Shaoteng Liu, and Jiaya Jia. Mini-gemini: Mining the potential of multi-modality vision language models. arXiv preprint arXiv:2403.18814, 2024. 2 +[37] Bin Lin, Bin Zhu, Yang Ye, Munan Ning, Peng Jin, and Li Yuan. Video-llava: Learning united visual representation by alignment before projection. arXiv preprint arXiv:2311.10122, 2023. 5, 6 +[38] Haotian Liu, Chunyuan Li, Yuheng Li, Bo Li, Yuanhan Zhang, Sheng Shen, and Yong Jae Lee. Llava-last: Improved reasoning,OCR,and world knowledge,2024.6 +[39] Haotian Liu, Chunyuan Li, Qingyang Wu, and Yong Jae Lee. Visual instruction tuning. Advances in neural information processing systems, 36, 2024. 5 +[40] Hao Liu, Wilson Yan, Matei Zaharia, and Pieter Abbeel. World model on million-length video and language with ringattention. arXiv preprint arXiv:2402.08268, 2024. 1, 2, 6 +[41] Jiasen Lu, Christopher Clark, Sangho Lee, Zichen Zhang, Savya Khosla, Ryan Marten, Derek Hoiem, and Aniruddha Kembhavi. Unified-io 2: Scaling autoregressive multimodal models with vision, language, audio, and action. arXiv preprint arXiv:2312.17172, 2023. 2 +[42] Ruipu Luo, Ziwang Zhao, Min Yang, Junwei Dong, Da Li, Pengcheng Lu, Tao Wang, Linmei Hu, Minghui Qiu, and Zhongyu Wei. Valley: Video assistant with large language model enhanced ability. arXiv preprint arXiv:2306.07207, 2023.5,7 +[43] Muhammad Maaz, Hanoona Rasheed, Salman Khan, and Fahad Shahbaz Khan. Video-chatgpt: Towards detailed video understanding via large vision and language models. arXiv preprint arXiv:2306.05424, 2023. 5, 6, 7 +[44] Karttikeya Mangalam, Raiymbek Akshulakov, and Jitendra Malik. Egoschema: A diagnostic benchmark for very long-form video language understanding. Advances in Neural Information Processing Systems, 36:46212-46244, 2023. 6 +[45] Willi Menapace, Aliaksandr Siarohin, Ivan Skorokhodov, Ekaterina Deyneka, Tsai-Shien Chen, Anil Kag, Yuwei Fang, Aleksei Stoliar, Elisa Ricci, Jian Ren, et al. Snap video: Scaled spatiotemporal transformers for text-to-video synthesis. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 7038-7048, 2024. 7 +[46] OpenAI. Gpt-4v(ison) system card, 2023. 6 +[47] OpenAI. Gpt-4o system card, 2024. 6 +[48] Viorica Patraucean, Lucas Smaira, Ankush Gupta, Adria Recasens, Larisa Markeeva, Dylan Banarse, Skanda Koppula, Mateusz Malinowski, Yi Yang, Carl Doersch, et al. Perception test: A diagnostic benchmark for multimodal video models. Advances in Neural Information Processing Systems, 36, 2024. 5, 6 +[49] Alec Radford, Jong Wook Kim, Chris Hallacy, Aditya Ramesh, Gabriel Goh, Sandhini Agarwal, Girish Sastry, Amanda Askell, Pamela Mishkin, Jack Clark, et al. Learning transferable visual models from natural language supervision. In International conference on machine learning, pages 8748-8763. PMLR, 2021. 5 + +[50] Robin Rombach, Andreas Blattmann, Dominik Lorenz, Patrick Esser, and Björn Ommer. High-resolution image synthesis with latent diffusion models. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 10684-10695, 2022. 2, 3 +[51] Christoph Schuhmann, Romain Beaumont, Richard Vencu, Cade Gordon, Ross Wightman, Mehdi Cherti, Theo Coombes, Aarush Katta, Clayton Mullis, Mitchell Wortsman, et al. Laion-5b: An open large-scale dataset for training next generation image-text models. Advances in Neural Information Processing Systems, 35:25278–25294, 2022. 5 +[52] Piyush Sharma, Nan Ding, Sebastian Goodman, and Radu Soricut. Conceptual captions: A cleaned, hypernymed, image alt-text dataset for automatic image captioning. In Proceedings of the 56th Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers), pages 2556-2565, 2018. 5 +[53] Uriel Singer, Adam Polyak, Thomas Hayes, Xi Yin, Jie An, Songyang Zhang, Qiyuan Hu, Harry Yang, Oron Ashual, Oran Gafni, et al. Make-a-video: Text-to-video generation without text-video data. arXiv preprint arXiv:2209.14792, 2022. 7 +[54] Quan Sun, Yufeng Cui, Xiaosong Zhang, Fan Zhang, Qiying Yu, Zhengxiong Luo, Yueze Wang, Yongming Rao, Jingjing Liu, Tiejun Huang, et al. Generative multimodal models are in-context learners. arXiv preprint arXiv:2312.13286, 2023. 2, 4 +[55] Quan Sun, Qiying Yu, Yufeng Cui, Fan Zhang, Xiaosong Zhang, Yueze Wang, Hongcheng Gao, Jingjing Liu, Tiejun Huang, and Xinlong Wang. Generative pretraining in multimodality. arXiv preprint arXiv:2307.05222, 2023. 1, 2 +[56] Quan Sun, Qiying Yu, Yufeng Cui, Fan Zhang, Xiaosong Zhang, Yueze Wang, Hongcheng Gao, Jingjing Liu, Tiejun Huang, and Xinlong Wang. Generative pretraining in multimodality. arXiv preprint arXiv:2307.05222, 2023. +[57] Chameleon Team. Chameleon: Mixed-modal early-fusion foundation models. arXiv preprint arXiv:2405.09818, 2024. 1, 2 +[58] Gemini Team, Rohan Anil, Sebastian Borgeaud, Yonghui Wu, Jean-Baptiste Alayrac, Jiahui Yu, Radu Soricut, Johan Schalkwyk, Andrew M Dai, Anja Hauth, et al. Gemini: a family of highly capable multimodal models. arXiv preprint arXiv:2312.11805, 2023. 6 +[59] Gemini Team, Petko Georgiev, Ving Ian Lei, Ryan Burnell, Libin Bai, Anmol Gulati, Garrett Tanzer, Damien Vincent, Zhufeng Pan, Shibo Wang, et al. Gemini 1.5: Unlocking multimodal understanding across millions of tokens of context. arXiv preprint arXiv:2403.05530, 2024. 6 +[60] Hugo Touvron, Thibaut Lavril, Gautier Izacard, Xavier Martinet, Marie-Anne Lachaux, Timothée Lacroix, Baptiste Rozière, Naman Goyal, Eric Hambro, Faisal Azhar, et al. Llama: Open and efficient foundation language models. arXiv preprint arXiv:2302.13971, 2023. 1 +[61] Michael Tschannen, Cian Eastwood, and Fabian Mentzer. Givt: Generative infinite-vocabulary transformers. In European Conference on Computer Vision, pages 292-309. Springer, 2025. 5, 7 + +[62] Thomas Unterthiner, Sjoerd Van Steenkiste, Karol Kurach, Raphael Marinier, Marcin Michalski, and Sylvain Gelly. Towards accurate generative models of video: A new metric & challenges. arXiv preprint arXiv:1812.01717, 2018. 5, 6 +[63] Aaron Van Den Oord, Oriol Vinyals, et al. Neural discrete representation learning. Advances in neural information processing systems, 30, 2017. 2 +[64] Hanyu Wang, Saksham Suri, Yixuan Ren, Hao Chen, and Abhinav Shrivastava. Larp: Tokenizing videos with a learned autoregressive generative prior. arXiv preprint arXiv:2410.21264, 2024. 7 +[65] Wenxuan Wang, Quan Sun, Fan Zhang, Yepeng Tang, Jing Liu, and Xinlong Wang. Diffusion feedback helps clip see better. arXiv preprint arXiv:2407.20171, 2024. 3 +[66] Xiang Wang, Hangjie Yuan, Shiwei Zhang, Dayou Chen, Jiuniu Wang, Yingya Zhang, Yujun Shen, Deli Zhao, and Jingren Zhou. Videocomposer: Compositional video synthesis with motion controllability. Advances in Neural Information Processing Systems, 36, 2024. 7 +[67] Xinlong Wang, Xiaosong Zhang, Zhengxiong Luo, Quan Sun, Yufeng Cui, Jinsheng Wang, Fan Zhang, Yueze Wang, Zhen Li, Qiying Yu, et al. Emu3: Next-token prediction is all you need. arXiv preprint arXiv:2409.18869, 2024. 2, 3 +[68] Yi Wang, Yinan He, Yizhuo Li, Kunchang Li, Jiashuo Yu, Xin Ma, Xinhao Li, Guo Chen, Xinyuan Chen, Yaohui Wang, et al. Internvid: A large-scale video-text dataset for multimodal understanding and generation. arXiv preprint arXiv:2307.06942, 2023. 7 +[69] Yuqing Wang, Tianwei Xiong, Daquan Zhou, Zhijie Lin, Yang Zhao, Bingyi Kang, Jiashi Feng, and Xihui Liu. Loong: Generating minute-level long videos with autoregressive language models. arXiv preprint arXiv:2410.02757, 2024. 2, 6, 7 +[70] Chen Wei, Karttikeya Mangalam, Po-Yao Huang, Yanghao Li, Haoqi Fan, Hu Xu, Huiyu Wang, Cihang Xie, Alan Yuille, and Christoph Feichtenhofer. Diffusion models as masked autoencoders. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 16284-16294, 2023. 3 +[71] Chenfei Wu, Lun Huang, Qianxi Zhang, Binyang Li, Lei Ji, Fan Yang, Guillermo Sapiro, and Nan Duan. Godiva: Generating open-domain videos from natural descriptions. arXiv preprint arXiv:2104.14806, 2021. 6 +[72] Chengyue Wu, Xiaokang Chen, Zhiyu Wu, Yiyang Ma, Xingchao Liu, Zizheng Pan, Wen Liu, Zhenda Xie, Xingkai Yu, Chong Ruan, et al. Janus: Decoupling visual encoding for unified multimodal understanding and generation. arXiv preprint arXiv:2410.13848, 2024. 2 +[73] Shengqiong Wu, Hao Fei, Leigang Qu, Wei Ji, and Tat-Seng Chua. Next-gpt: Any-to-any multimodal llm. arXiv preprint arXiv:2309.05519, 2023. +[74] Yecheng Wu, Zhuoyang Zhang, Junyu Chen, Haotian Tang, Dacheng Li, Yunhao Fang, Ligeng Zhu, Enze Xie, Hongxu Yin, Li Yi, et al. Vila-u: a unified foundation model integrating visual understanding and generation. arXiv preprint arXiv:2409.04429, 2024. 1, 2, 6, 7 +[75] Weilai Xiang, Hongyu Yang, Di Huang, and Yunhong Wang. Denoising diffusion autoencoders are unified self-supervised + +learners. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 15802-15812, 2023. 3 +[76] Junbin Xiao, Xindi Shang, Angela Yao, and Tat-Seng Chua. Next-qa: Next phase of question-answering to explaining temporal actions. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 9777–9786, 2021. 5 +[77] Jinheng Xie, Weijia Mao, Zechen Bai, David Junhao Zhang, Weihao Wang, Kevin Qinghong Lin, Yuchao Gu, Zhijie Chen, Zhenheng Yang, and Mike Zheng Shou. Show-o: One single transformer to unify multimodal understanding and generation. arXiv preprint arXiv:2408.12528, 2024. 1, 2, 3 +[78] Jinbo Xing, Menghan Xia, Yong Zhang, Haoxin Chen, Wangbo Yu, Hanyuan Liu, Gongye Liu, Xintao Wang, Ying Shan, and Tien-Tsin Wong. Dynamiccafter: Animating open-domain images with video diffusion priors. In European Conference on Computer Vision, pages 399-417. Springer, 2025. 2, 4 +[79] Jun Xu, Tao Mei, Ting Yao, and Yong Rui. Msr-vtt: A large video description dataset for bridging video and language. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 5288-5296, 2016. 5, 6 +[80] Jiarui Xu, Sifei Liu, Arash Vahdat, Wonmin Byeon, Xiaolong Wang, and Shalini De Mello. Open-vocabulary panoptic segmentation with text-to-image diffusion models. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 2955-2966, 2023. 3 +[81] Lin Xu, Yilin Zhao, Daquan Zhou, Zhijie Lin, See Kiong Ng, and Jiashi Feng. Pllava: Parameter-free llava extension from images to videos for video dense captioning. arXiv preprint arXiv:2404.16994, 2024. 6 +[82] Wilson Yan, Yunzhi Zhang, Pieter Abbeel, and Aravind Srinivas. Videogpt: Video generation using vq-vae and transformers. arXiv preprint arXiv:2104.10157, 2021. 5 +[83] Jian Yang, Dacheng Yin, Yizhou Zhou, Fengyun Rao, Wei Zhai, Yang Cao, and Zheng-Jun Zha. Mmar: Towards lossless multi-modal auto-regressive probabilistic modeling. arXiv preprint arXiv:2410.10798, 2024. 2 +[84] Shuai Yang, Yuying Ge, Yang Li, Yukang Chen, Yixiao Ge, Ying Shan, and Yingcong Chen. Seed-story: Multimodal long story generation with large language model. arXiv preprint arXiv:2407.08683, 2024. 2 +[85] Kexin Yi, Chuang Gan, Yunzhu Li, Pushmeet Kohli, Jiajun Wu, Antonio Torralba, and Joshua B Tenenbaum. Clevrer: Collision events for video representation and reasoning. arXiv preprint arXiv:1910.01442, 2019. 5 +[86] Lili Yu, Bowen Shi, Ramakanth Pasunuru, Benjamin Muller, Olga Golovneva, Tianlu Wang, Arun Babu, Binh Tang, Brian Karrer, Shelly Sheynin, et al. Scaling autoregressive multimodal models: Pretraining and instruction tuning. arXiv preprint arXiv:2309.02591, 2023. 1, 2 +[87] Qiying Yu, Quan Sun, Xiaosong Zhang, Yufeng Cui, Fan Zhang, Yue Cao, Xinlong Wang, and Jingjing Liu. Capsfusion: Rethinking image-text data at scale. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 14022-14032, 2024. 5 +[88] Zhou Yu, Dejing Xu, Jun Yu, Ting Yu, Zhou Zhao, Yueting Zhuang, and Dacheng Tao. Activitynet-qa: A dataset for + +understanding complex web videos via question answering. In Proceedings of the AAAI Conference on Artificial Intelligence, pages 9127-9134, 2019. 5, 6 +[89] Chuyang Zhao, Yuxing Song, Wenhao Wang, Haocheng Feng, Errui Ding, Yifan Sun, Xinyan Xiao, and Jingdong Wang. Monoformer: One transformer for both diffusion and autoregression. arXiv preprint arXiv:2409.16280, 2024. 2 +[90] Wenliang Zhao, Yongming Rao, Zuyan Liu, Benlin Liu, Jie Zhou, and Jiwen Lu. Unleashing text-to-image diffusion models for visual perception. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 5729-5739, 2023. 3 +[91] Chunting Zhou, Lili Yu, Arun Babu, Kushal Tirumala, Michihiro Yasunaga, Leonid Shamis, Jacob Kahn, Xuezhe Ma, Luke Zettlemoyer, and Omer Levy. Transfusion: Predict the next token and diffuse images with one multi-modal model. arXiv preprint arXiv:2408.11039, 2024. 1, 2 +[92] Luowei Zhou, Chenliang Xu, and Jason Corso. Towards automatic learning of procedures from web instructional videos. In Proceedings of the AAAI Conference on Artificial Intelligence, 2018. 5 +[93] Jinguo Zhu, Xiaohan Ding, Yixiao Ge, Yuying Ge, Sijie Zhao, Hengshuang Zhao, Xiaohua Wang, and Ying Shan. Vl-gpt: A generative pre-trained transformer for vision and language understanding and generation. arXiv preprint arXiv:2312.09251, 2023. 1, 2 \ No newline at end of file diff --git a/2025/Divot_ Diffusion Powers Video Tokenizer for Comprehension and Generation/images.zip b/2025/Divot_ Diffusion Powers Video Tokenizer for Comprehension and Generation/images.zip new file mode 100644 index 0000000000000000000000000000000000000000..c19113e5d8abbe370a52d7bed290d47829e4e04c --- /dev/null +++ b/2025/Divot_ Diffusion Powers Video Tokenizer for Comprehension and Generation/images.zip @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:b99f05de63ea173188ead261d116ef7a270b8cc25485e7c73aec83529dc0be14 +size 694677 diff --git a/2025/Divot_ Diffusion Powers Video Tokenizer for Comprehension and Generation/layout.json b/2025/Divot_ Diffusion Powers Video Tokenizer for Comprehension and Generation/layout.json new file mode 100644 index 0000000000000000000000000000000000000000..c71d65f5935f5cdf4b4fc1dbd2a2367bb5726b6c --- /dev/null +++ b/2025/Divot_ Diffusion Powers Video Tokenizer for Comprehension and Generation/layout.json @@ -0,0 +1,9933 @@ +{ + "pdf_info": [ + { + "para_blocks": [ + { + "bbox": [ + 67, + 103, + 542, + 120 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 67, + 103, + 542, + 120 + ], + "spans": [ + { + "bbox": [ + 67, + 103, + 542, + 120 + ], + "type": "text", + "content": "Divot: Diffusion Powers Video Tokenizer for Comprehension and Generation" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 159, + 143, + 451, + 185 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 159, + 143, + 451, + 185 + ], + "spans": [ + { + "bbox": [ + 159, + 143, + 451, + 185 + ], + "type": "text", + "content": "Yuying Ge Yizhuo Li Yixiao Ge Ying Shan ARC Lab, Tencent PCG https://github.com/TencentARC/Divot" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 152, + 213, + 200, + 225 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 152, + 213, + 200, + 225 + ], + "spans": [ + { + "bbox": [ + 152, + 213, + 200, + 225 + ], + "type": "text", + "content": "Abstract" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 55, + 239, + 297, + 585 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 239, + 297, + 585 + ], + "spans": [ + { + "bbox": [ + 55, + 239, + 297, + 585 + ], + "type": "text", + "content": "In recent years, there has been a significant surge of interest in unifying image comprehension and generation within Large Language Models (LLMs). This growing interest has prompted us to explore extending this unification to videos. The core challenge lies in developing a versatile video tokenizer that captures both the spatial characteristics and temporal dynamics of videos to obtain representations for LLMs, and the representations can be further decoded into realistic video clips to enable video generation. In this work, we introduce Divot, a Diffusion-Powered Video Tokenizer, which leverages the diffusion process for self-supervised video representation learning. We posit that if a video diffusion model can effectively de-noise video clips by taking the features of a video tokenizer as the condition, then the tokenizer has successfully captured robust spatial and temporal information. Additionally, the video diffusion model inherently functions as a de-tokenizer, decoding videos from their representations. Building upon the Divot tokenizer, we present Divot-LLM through video-to-text autoregression and text-to-video generation by modeling the distributions of continuous-valued Divot features with a Gaussian Mixture Model. Experimental results demonstrate that our diffusion-based video tokenizer, when integrated with a pre-trained LLM, achieves competitive performance across various video comprehension and generation benchmarks. The instruction tuned Divot-LLM also excels in video storytelling, generating interleaved narratives and corresponding videos. Models and codes are available at https://github.com/TencentARC/Divot." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 56, + 609, + 135, + 621 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 609, + 135, + 621 + ], + "spans": [ + { + "bbox": [ + 56, + 609, + 135, + 621 + ], + "type": "text", + "content": "1. Introduction" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 55, + 629, + 296, + 712 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 629, + 296, + 712 + ], + "spans": [ + { + "bbox": [ + 55, + 629, + 296, + 712 + ], + "type": "text", + "content": "In recent years, the rapid evolution of Multimodal Large Language Models (MLLMs) [12, 15-17, 55-57, 77, 86, 91, 93] has demonstrated significant progresses in unified image understanding and generation, which empowers LLMs [6, 11, 60] with the ability to generate images beyond texts. While these work primarily focus on image-text data, the extension of this unification to the video domain remains" + } + ] + } + ], + "index": 7 + }, + { + "type": "image", + "bbox": [ + 318, + 213, + 554, + 309 + ], + "blocks": [ + { + "bbox": [ + 318, + 213, + 554, + 309 + ], + "lines": [ + { + "bbox": [ + 318, + 213, + 554, + 309 + ], + "spans": [ + { + "bbox": [ + 318, + 213, + 554, + 309 + ], + "type": "image", + "image_path": "0279fe125a1240893f4d721d6970606d320221f4fa2bb5eb7b575fb6b0ce1b4c.jpg" + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 313, + 319, + 555, + 387 + ], + "lines": [ + { + "bbox": [ + 313, + 319, + 555, + 387 + ], + "spans": [ + { + "bbox": [ + 313, + 319, + 555, + 387 + ], + "type": "text", + "content": "Figure 1. We utilize the diffusion procedure to learn a video tokenizer in a self-supervised manner for unified comprehension and generation, where the spatiotemporal representations serve as the condition of a diffusion model to de-noise video clips. Additionally, the proxy diffusion model functions as a de-tokenizer to decode realistic video clips from the video representations." + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "image_caption" + } + ], + "index": 8 + }, + { + "bbox": [ + 313, + 402, + 555, + 461 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 402, + 555, + 461 + ], + "spans": [ + { + "bbox": [ + 313, + 402, + 555, + 461 + ], + "type": "text", + "content": "relatively under-explored. Achieving unified video comprehension and generation is essential for the development of more sophisticated artificial intelligence (AI) systems that are capable of understanding and creating dynamic visual content in the real world." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 313, + 462, + 556, + 700 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 462, + 556, + 700 + ], + "spans": [ + { + "bbox": [ + 313, + 462, + 556, + 700 + ], + "type": "text", + "content": "The primary challenge of achieving unified video comprehension and generation lies in developing a versatile video tokenizer that can effectively address the complexities inherent in video data. This tokenizer should be able to obtain robust video representations that serve as inputs of MLLMs for video comprehension, and these representations can be further decoded into realistic video clips to enable video generation. Unlike static images, videos encompass both spatial characteristics and temporal dynamics, making their representation significantly more complex. Recent pioneering work [26, 40, 74] adopt a discrete video tokenizer for unifying video comprehension and generation, where a video is represented as a sequence of discrete frame tokens, or keyframe tokens followed by discrete motion tokens. This approach eases video generation with a LLM through an autoregressive next-token prediction mechanism, but sacrifices the performance of multimodal understanding, as pointed out by recent work [77]. In this work, we aim to investigate an alternative approach by utilizing continuous video representations to unify video comprehension and generation." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 326, + 701, + 555, + 712 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 326, + 701, + 555, + 712 + ], + "spans": [ + { + "bbox": [ + 326, + 701, + 555, + 712 + ], + "type": "text", + "content": "To this end, we introduce Divot, a Diffusion-Powered" + } + ] + } + ], + "index": 12 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "spans": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "type": "text", + "content": "CVF" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 145, + 0, + 494, + 35 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 145, + 0, + 494, + 35 + ], + "spans": [ + { + "bbox": [ + 145, + 0, + 494, + 35 + ], + "type": "text", + "content": "This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 294, + 749, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 749, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 749, + 317, + 757 + ], + "type": "text", + "content": "13606" + } + ] + } + ], + "index": 13 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 0 + }, + { + "para_blocks": [ + { + "bbox": [ + 55, + 72, + 294, + 251 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 72, + 294, + 251 + ], + "spans": [ + { + "bbox": [ + 55, + 72, + 294, + 251 + ], + "type": "text", + "content": "Video Tokenizer that leverages the diffusion process [50] for self-supervised video representation learning as shown in Fig. 1. The core premise is that if a diffusion model can effectively predict the noise added to the Variational Autoencoder (VAE) latents [28] of video clips, when conditioned on the features produced by the video tokenizer, it demonstrates that the tokenizer has successfully captured robust spatial and temporal information inherent in the video data. This capability is crucial for representing the intricate dynamics present in videos. Furthermore, in addition to being a proxy module for learning the tokenizer, the diffusion model can act as a de-tokenizer to effectively decode realistic videos from their learned representations. This dual functionality facilitates a seamless integration of understanding and creating video content within a LLM." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 55, + 253, + 295, + 407 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 253, + 295, + 407 + ], + "spans": [ + { + "bbox": [ + 55, + 253, + 295, + 407 + ], + "type": "text", + "content": "Specifically, the Divot tokenizer is composed of a pretrained Vision Transformer (ViT) encoder [13], a Spatial-Temporal transformer, and a Perceiver Resampler [1] to obtain video representations from video frames sampled at low frame rate (fps) considering the semantic redundancy between adjacent frames. The video representations serve as the condition of a pre-trained video diffusion model, DynamiCrafter [78] (without the concatenation of a conditional image with initial noise), to predict the noise added to the VAE latents of video frames. After training, the video diffusion model can generate realistic video clips from noise by taking the video representations provided by the Divot tokenizer as the condition." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 55, + 409, + 295, + 673 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 409, + 295, + 673 + ], + "spans": [ + { + "bbox": [ + 55, + 409, + 295, + 673 + ], + "type": "text", + "content": "We further present Divot-LLM by equipping the pretrained Mistral-7B [24] with the Divot tokenizer. Divot-LLM is pre-trained with a next-word prediction objective on video-caption data by taking the spatiotemporal representations as inputs for video comprehension. The challenge then arises in modeling the continuous video representations with the LLM for video generation. We empirically find that simply minimizing the distance between the LLM output and video representations using mean squared error (MSE) loss achieves unsatisfactory results, since the deterministic regression regularizes the LLM to learn overly averaged features of videos. To address this, inspired by recent work [33], we shift our focus from deterministic modeling to probabilistic modeling by modeling the distributions of video features with a Gaussian Mixture Model (GMM). Specifically, we train the LLM to predict GMM parameters, including means, variances, and mixture probabilities through minimizing the discrepancy between the predicted GMM distribution and the actual video representations using negative log-likelihood (NLL) loss. During inference, we draw samples from the predicted GMM distribution as the condition of the video de-tokenizer to decode video clips." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 313, + 72, + 555, + 167 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 72, + 555, + 167 + ], + "spans": [ + { + "bbox": [ + 313, + 72, + 555, + 167 + ], + "type": "text", + "content": "We benchmark Divot-LLM on a broad range of video comprehension tasks and zero-shot video generation, achieving competitive performance through pre-training on 5 million video-text pairs. By leveraging the generality of the video tokenizer, our Divot-LLM also enables video storytelling, which generates interleaved narratives and corresponding videos that are temporally coherent through fine-tuning on a specific animation dataset." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 313, + 169, + 556, + 324 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 169, + 556, + 324 + ], + "spans": [ + { + "bbox": [ + 313, + 169, + 556, + 324 + ], + "type": "text", + "content": "Our contributions are three-fold. (1) We introduce Divot, an advanced video tokenizer that leverages a diffusion procedure for self-supervised video representation learning, aiming to unify video comprehension and generation. (2) We present Divot-LLM, composed of a pre-trained LLM and the Divot tokenizer to enable understanding and generating video content within a single framework. We investigate effective approaches for fitting continuous video representations using the LLM with probabilistic modeling for video generation. (3) We conduct extensive experiments to demonstrate Divot-LLM's competitive performance on existing video comprehension and generation benchmarks, as well as video storytelling. All models and code are released." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 314, + 338, + 400, + 350 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 314, + 338, + 400, + 350 + ], + "spans": [ + { + "bbox": [ + 314, + 338, + 400, + 350 + ], + "type": "text", + "content": "2. Related Work" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 313, + 358, + 556, + 622 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 358, + 556, + 622 + ], + "spans": [ + { + "bbox": [ + 313, + 358, + 556, + 622 + ], + "type": "text", + "content": "MLLMs for Comprehension and Generation. With the rapid development of Multimodal Large Language Models (MLLM), recent studies have been working on unified MLLMs [12, 15-17, 25, 26, 36, 40, 41, 54, 55, 57, 67, 72-74, 77, 83, 84, 86, 89, 91, 93] that are capable of multimodal comprehension and generation. To empower LLMs with the capability to generate visual content, existing work primarily employs the following three approaches: (1) utilizing a pre-trained stable diffusion model to generate images conditioned on LLM output (either continuous features or discrete tokens); (2) employing a Vector Quantized (VQ) [63] based decoder to generate visual content from the discrete codes predicted by LLMs; (3) using LLMs to de-noise Gaussian noise through a diffusion process. While most work predominantly focus on the unification of images and texts, some pioneering studies [26, 40, 67, 74] further advance the integration of video comprehension and generation within an LLM through generating videos from discrete codes using a VQ-based decoder, which falls into the second approach. In this work, we adopt the first approach, which involves leveraging a diffusion model to achieve unified video understanding and generation from continuous representations." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 313, + 641, + 556, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 641, + 556, + 715 + ], + "spans": [ + { + "bbox": [ + 313, + 641, + 556, + 715 + ], + "type": "text", + "content": "Video Tokenizer in MLLMs. Previous work on video generation with LLMs predominantly employs a discrete video tokenizer to convert video signals into a sequence of quantized tokens. For example, LWM [40] and VILA-U [74] utilize a frame-level tokenizer to discretize each frame into a sequence of codes. VideoPoet [29], Loong [69] and" + } + ] + } + ], + "index": 7 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 55, + 683, + 295, + 713 + ], + "type": "page_footnote", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 683, + 295, + 713 + ], + "spans": [ + { + "bbox": [ + 55, + 683, + 295, + 713 + ], + "type": "text", + "content": "1We do not explore more advanced LLMs because we want to ensure that our superiority stems from the improved visual representations, rather than from the capabilities of a more sophisticated foundation model." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "text", + "content": "13607" + } + ] + } + ], + "index": 9 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 1 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 75, + 70, + 410, + 194 + ], + "blocks": [ + { + "bbox": [ + 75, + 70, + 410, + 194 + ], + "lines": [ + { + "bbox": [ + 75, + 70, + 410, + 194 + ], + "spans": [ + { + "bbox": [ + 75, + 70, + 410, + 194 + ], + "type": "image", + "image_path": "15287b5e77dd53b6bbf1e4a00e7adf0725fc8dcf61978a991b5d9c4799470da6.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 54, + 200, + 555, + 246 + ], + "lines": [ + { + "bbox": [ + 54, + 200, + 555, + 246 + ], + "spans": [ + { + "bbox": [ + 54, + 200, + 555, + 246 + ], + "type": "text", + "content": "Figure 2. Overview of Divot tokenization and de-tokenization. During training, sparsely sampled video frames are fed into the tokenizer to obtain spatiotemporal representations. These representations serve as the conditions for a U-Net, which is trained to de-noise the noisy VAE latents of densely sampled video frames. During inference, the video representations from the Divot tokenizer can be decoded into realistic video clips with the U-Net." + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "type": "image", + "bbox": [ + 419, + 71, + 536, + 190 + ], + "blocks": [ + { + "bbox": [ + 419, + 71, + 536, + 190 + ], + "lines": [ + { + "bbox": [ + 419, + 71, + 536, + 190 + ], + "spans": [ + { + "bbox": [ + 419, + 71, + 536, + 190 + ], + "type": "image", + "image_path": "a93c4e7225ea5b543ccb32551a8fe9109bec5f3797448e75c94d2701f6dc2326.jpg" + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_body" + } + ], + "index": 1 + }, + { + "bbox": [ + 56, + 260, + 298, + 429 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 260, + 298, + 429 + ], + "spans": [ + { + "bbox": [ + 56, + 260, + 298, + 429 + ], + "type": "text", + "content": "Emu3 [67] leverage a 3D CNN architecture, where the encoded spatial-temporal features are quantized into discrete tokens. Video-LaVIT [26] represents video clips as a keyframe followed by extracted motion vectors, obtaining the respective discrete codes. By converting continuous visual signals into discrete tokens, the original next-token prediction mechanism can be adopted to facilitate video generation with an LLM. However, recent work [77] observes a significant performance degradation in multimodal comprehension tasks when discrete representations are used instead of continuous representations. In this work, we introduce a video tokenizer with continuous representations through leveraging the diffusion [50] procedure, enabling it to be effectively integrated with a LLM for unified comprehension and generation." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 55, + 446, + 296, + 639 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 446, + 296, + 639 + ], + "spans": [ + { + "bbox": [ + 55, + 446, + 296, + 639 + ], + "type": "text", + "content": "Diffusion for Representation Learning. The diffusion process has been explored as a criterion for representation learning. Some works [3, 75, 80, 90] leverage the intermediate activations of pre-trained diffusion models for downstream tasks including classification, segmentation and depth estimation. Other works [23, 65, 70] employ the diffusion model as a proxy module for self-supervised learning, where noisy inputs are de-noised by conditioning on the image representations. This approach encourages the emergence of informative representations that capture key properties and semantics of the images. In this work, to the best of our knowledge, we for the first time leverage diffusion for video representation learning, where a video diffusion model is trained to de-noise video clips through taking the spatiotemporal representations as conditions, thereby encouraging the capture of spatial characteristics and temporal dynamics." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 55, + 651, + 111, + 662 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 651, + 111, + 662 + ], + "spans": [ + { + "bbox": [ + 55, + 651, + 111, + 662 + ], + "type": "text", + "content": "3. Method" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 55, + 671, + 153, + 683 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 671, + 153, + 683 + ], + "spans": [ + { + "bbox": [ + 55, + 671, + 153, + 683 + ], + "type": "text", + "content": "3.1. Divot Tokenizer" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 55, + 689, + 297, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 689, + 297, + 715 + ], + "spans": [ + { + "bbox": [ + 55, + 689, + 297, + 715 + ], + "type": "text", + "content": "We introduce Divot, a diffusion-powered tokenizer that leverage diffusion procedure for video representation learning." + } + ] + } + ], + "index": 7 + }, + { + "type": "image", + "bbox": [ + 316, + 258, + 555, + 392 + ], + "blocks": [ + { + "bbox": [ + 316, + 258, + 555, + 392 + ], + "lines": [ + { + "bbox": [ + 316, + 258, + 555, + 392 + ], + "spans": [ + { + "bbox": [ + 316, + 258, + 555, + 392 + ], + "type": "image", + "image_path": "2c531e3cd50b981ad0899df0f49e891e16f9be0a616fdad7e20a153fbf4c78e4.jpg" + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 313, + 398, + 556, + 475 + ], + "lines": [ + { + "bbox": [ + 313, + 398, + 556, + 475 + ], + "spans": [ + { + "bbox": [ + 313, + 398, + 556, + 475 + ], + "type": "text", + "content": "Figure 3. Overview of Divot-LLM. Video features from the Divot tokenizer are fed into the LLM to perform next-word prediction for video comprehension, while learnable queries are input into the LLM to model the distributions of Divot features using a Gaussian Mixture Model (GMM) for video generation. During inference, video features are sampled from the predicted GMM distribution to decode videos using the de-tokenizer." + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "image_caption" + } + ], + "index": 8 + }, + { + "bbox": [ + 313, + 486, + 555, + 522 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 486, + 555, + 522 + ], + "spans": [ + { + "bbox": [ + 313, + 486, + 555, + 522 + ], + "type": "text", + "content": "Additionally, the proxy diffusion model used for training the tokenizer can serve as a de-tokenizer to decode realistic video clips from their spatiotemporal representations." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 313, + 529, + 499, + 541 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 529, + 499, + 541 + ], + "spans": [ + { + "bbox": [ + 313, + 529, + 499, + 541 + ], + "type": "text", + "content": "3.1.1. Preliminary: Video Diffusion Model." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 313, + 544, + 556, + 641 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 544, + 556, + 641 + ], + "spans": [ + { + "bbox": [ + 313, + 544, + 556, + 641 + ], + "type": "text", + "content": "Diffusion models [20, 50] learn to model a probability distribution by reversing a process that progressively adds noise to the data. Specifically, given data " + }, + { + "bbox": [ + 313, + 544, + 556, + 641 + ], + "type": "inline_equation", + "content": "\\mathbf{x}_0 \\sim p(\\mathbf{x})" + }, + { + "bbox": [ + 313, + 544, + 556, + 641 + ], + "type": "text", + "content": ", the forward process gradually adds random Gaussian noise " + }, + { + "bbox": [ + 313, + 544, + 556, + 641 + ], + "type": "inline_equation", + "content": "\\epsilon_{\\mathbf{t}} \\in \\mathcal{N}(\\mathbf{0}, \\mathbf{I})" + }, + { + "bbox": [ + 313, + 544, + 556, + 641 + ], + "type": "text", + "content": " to the data sample " + }, + { + "bbox": [ + 313, + 544, + 556, + 641 + ], + "type": "inline_equation", + "content": "\\mathbf{x}_0" + }, + { + "bbox": [ + 313, + 544, + 556, + 641 + ], + "type": "text", + "content": " with a total of " + }, + { + "bbox": [ + 313, + 544, + 556, + 641 + ], + "type": "inline_equation", + "content": "T" + }, + { + "bbox": [ + 313, + 544, + 556, + 641 + ], + "type": "text", + "content": " timesteps to yield " + }, + { + "bbox": [ + 313, + 544, + 556, + 641 + ], + "type": "inline_equation", + "content": "\\mathbf{x}_t" + }, + { + "bbox": [ + 313, + 544, + 556, + 641 + ], + "type": "text", + "content": " through a parameterization trick. The denoising process predicts " + }, + { + "bbox": [ + 313, + 544, + 556, + 641 + ], + "type": "inline_equation", + "content": "\\epsilon_{\\mathbf{t}}" + }, + { + "bbox": [ + 313, + 544, + 556, + 641 + ], + "type": "text", + "content": " in the forward diffusion process with a denoising network " + }, + { + "bbox": [ + 313, + 544, + 556, + 641 + ], + "type": "inline_equation", + "content": "\\epsilon_{\\theta}(\\mathbf{x}_t, t)" + }, + { + "bbox": [ + 313, + 544, + 556, + 641 + ], + "type": "text", + "content": ", which is trained by the objective below," + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 355, + 654, + 555, + 671 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 355, + 654, + 555, + 671 + ], + "spans": [ + { + "bbox": [ + 355, + 654, + 555, + 671 + ], + "type": "interline_equation", + "content": "\\min _ {\\theta} \\mathbb {E} _ {t, \\mathbf {x} \\sim p, \\epsilon \\sim \\mathcal {N} (\\mathbf {0}, \\mathbf {I})} \\| \\epsilon - \\epsilon_ {\\theta} (\\mathbf {x} _ {t}, t) \\| _ {2} ^ {2}, \\tag {1}", + "image_path": "18507ac30459d9961d118b05c2670cb58198a23fb910c858bc6852f9f351a8bf.jpg" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 313, + 677, + 556, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 677, + 556, + 713 + ], + "spans": [ + { + "bbox": [ + 313, + 677, + 556, + 713 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 313, + 677, + 556, + 713 + ], + "type": "inline_equation", + "content": "\\epsilon" + }, + { + "bbox": [ + 313, + 677, + 556, + 713 + ], + "type": "text", + "content": " is the sampled Gaussian noise and " + }, + { + "bbox": [ + 313, + 677, + 556, + 713 + ], + "type": "inline_equation", + "content": "\\theta" + }, + { + "bbox": [ + 313, + 677, + 556, + 713 + ], + "type": "text", + "content": " indicates the parameters of the denoising network. During inference, we can perform iterative denoising from a random Gaussian" + } + ] + } + ], + "index": 14 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "text", + "content": "13608" + } + ] + } + ], + "index": 15 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 2 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 63, + 76, + 216, + 213 + ], + "blocks": [ + { + "bbox": [ + 63, + 76, + 216, + 213 + ], + "lines": [ + { + "bbox": [ + 63, + 76, + 216, + 213 + ], + "spans": [ + { + "bbox": [ + 63, + 76, + 216, + 213 + ], + "type": "image", + "image_path": "3f1baa720b8b0a5f515f4ed0d31093760eece10aa567131fefa68bced55b38b9.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 54, + 223, + 555, + 269 + ], + "lines": [ + { + "bbox": [ + 54, + 223, + 555, + 269 + ], + "spans": [ + { + "bbox": [ + 54, + 223, + 555, + 269 + ], + "type": "text", + "content": "Figure 4. Paradigms for modeling video representations from the Divot tokenizer with a LLM for video generation. (a) MSE Regression, where the LLM output is trained to minimize its distance with video features using Mean Squared Error (MSE) loss; (b) Diffusion Modeling, where the LLM output is fed into a denoising network as the condition to predict the noise added to video features; (c) GMM Modeling, where the LLM output is trained to predict the parameters of a Gaussian Mixture Model (GMM) for modeling video feature distributions." + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "type": "image", + "bbox": [ + 231, + 76, + 379, + 213 + ], + "blocks": [ + { + "bbox": [ + 231, + 76, + 379, + 213 + ], + "lines": [ + { + "bbox": [ + 231, + 76, + 379, + 213 + ], + "spans": [ + { + "bbox": [ + 231, + 76, + 379, + 213 + ], + "type": "image", + "image_path": "3369033bdb1d68330b8fe8c23297d80ba63ecc490d094b81f8eed65ce4c81068.jpg" + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_body" + } + ], + "index": 1 + }, + { + "type": "image", + "bbox": [ + 402, + 71, + 556, + 214 + ], + "blocks": [ + { + "bbox": [ + 402, + 71, + 556, + 214 + ], + "lines": [ + { + "bbox": [ + 402, + 71, + 556, + 214 + ], + "spans": [ + { + "bbox": [ + 402, + 71, + 556, + 214 + ], + "type": "image", + "image_path": "02be3176e5e209bde863f639edd20d1be47c8279641f639ac8d146a2205a4479.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + } + ], + "index": 2 + }, + { + "bbox": [ + 55, + 283, + 179, + 295 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 283, + 179, + 295 + ], + "spans": [ + { + "bbox": [ + 55, + 283, + 179, + 295 + ], + "type": "text", + "content": "noise for the denoised data " + }, + { + "bbox": [ + 55, + 283, + 179, + 295 + ], + "type": "inline_equation", + "content": "\\mathbf{x}_0" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 55, + 297, + 296, + 370 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 297, + 296, + 370 + ], + "spans": [ + { + "bbox": [ + 55, + 297, + 296, + 370 + ], + "type": "text", + "content": "For video diffusion models [8, 78], given a video " + }, + { + "bbox": [ + 55, + 297, + 296, + 370 + ], + "type": "inline_equation", + "content": "\\mathbf{x}" + }, + { + "bbox": [ + 55, + 297, + 296, + 370 + ], + "type": "text", + "content": ", a latent representation " + }, + { + "bbox": [ + 55, + 297, + 296, + 370 + ], + "type": "inline_equation", + "content": "\\mathbf{z} = \\mathcal{E}(\\mathbf{x})" + }, + { + "bbox": [ + 55, + 297, + 296, + 370 + ], + "type": "text", + "content": " is first encoded to reduce the computational complexity. Then the forward diffusion process and backward denoise process are performed in this latent space with a denoising network " + }, + { + "bbox": [ + 55, + 297, + 296, + 370 + ], + "type": "inline_equation", + "content": "\\epsilon_{\\theta}(\\mathbf{x}_t,\\mathbf{c},t)" + }, + { + "bbox": [ + 55, + 297, + 296, + 370 + ], + "type": "text", + "content": ", where " + }, + { + "bbox": [ + 55, + 297, + 296, + 370 + ], + "type": "inline_equation", + "content": "\\mathbf{c}" + }, + { + "bbox": [ + 55, + 297, + 296, + 370 + ], + "type": "text", + "content": " denotes denoising conditions like text or visual prompts." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 55, + 382, + 157, + 395 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 382, + 157, + 395 + ], + "spans": [ + { + "bbox": [ + 55, + 382, + 157, + 395 + ], + "type": "text", + "content": "3.1.2. Training Pipeline" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 55, + 399, + 296, + 651 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 399, + 296, + 651 + ], + "spans": [ + { + "bbox": [ + 55, + 399, + 296, + 651 + ], + "type": "text", + "content": "As illustrated in Fig. 2, given a video clip, we separately sample sparse frames at 2 fps to obtain the video representations from the tokenizer, and sample dense frames at 8 fps to obtain latent representations " + }, + { + "bbox": [ + 55, + 399, + 296, + 651 + ], + "type": "inline_equation", + "content": "z_{0}" + }, + { + "bbox": [ + 55, + 399, + 296, + 651 + ], + "type": "text", + "content": " from the frozen VAE [28] encoder. Sparse frames are sampled as the input of the video tokenizer considering the semantic redundancy between adjacent frames. The forward diffusion process gradually adds Gaussian noise " + }, + { + "bbox": [ + 55, + 399, + 296, + 651 + ], + "type": "inline_equation", + "content": "\\theta" + }, + { + "bbox": [ + 55, + 399, + 296, + 651 + ], + "type": "text", + "content": " to " + }, + { + "bbox": [ + 55, + 399, + 296, + 651 + ], + "type": "inline_equation", + "content": "z_{0}" + }, + { + "bbox": [ + 55, + 399, + 296, + 651 + ], + "type": "text", + "content": " for producing the noisy input " + }, + { + "bbox": [ + 55, + 399, + 296, + 651 + ], + "type": "inline_equation", + "content": "z_{t}" + }, + { + "bbox": [ + 55, + 399, + 296, + 651 + ], + "type": "text", + "content": ". At each backward step " + }, + { + "bbox": [ + 55, + 399, + 296, + 651 + ], + "type": "inline_equation", + "content": "t" + }, + { + "bbox": [ + 55, + 399, + 296, + 651 + ], + "type": "text", + "content": ", a denoising U-Net is trained to predict the noise added from the previous step to the current step by taking the time embedding and video representations as the condition. Specifically, the video representations interact with the denoising U-Net intermediate features through cross-attention layers, where each noisy latent attends to all video tokens. By constraining the U-Net to reconstruct fine-grained spatial and temporal information of video clips through relying on video features, the Divot tokenizer is optimized to capture both spatial characteristics and temporal dynamics for robust video representations. The Divot tokenizer is trained on pure videos of a subset of WebVid10M [2] and Panda-70M [9], totaling 10M videos." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 55, + 653, + 296, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 653, + 296, + 714 + ], + "spans": [ + { + "bbox": [ + 55, + 653, + 296, + 714 + ], + "type": "text", + "content": "After training the Divot tokenizer, the proxy denoising U-Net (employed to implement the parameterized loss function) can serve as an effective video de-tokenizer, which is able to decode semantically aligned video clips from the learned spatiotemporal representations as shown in Fig. 5." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 313, + 283, + 426, + 294 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 283, + 426, + 294 + ], + "spans": [ + { + "bbox": [ + 313, + 283, + 426, + 294 + ], + "type": "text", + "content": "3.1.3. Model Architecture" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 313, + 299, + 556, + 491 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 299, + 556, + 491 + ], + "spans": [ + { + "bbox": [ + 313, + 299, + 556, + 491 + ], + "type": "text", + "content": "As shown in Fig. 2, the Divot tokenizer is composed of a pre-trained ViT encoder to extract frame-level features, a transformer for spatial and temporal fusion, and a Perceiver Resampler [1] to produce a fixed number of video tokens. The Perceiver Resampler is adopted for two reasons: (1) to reduce the number of video tokens that a LLM need to predict for generation, and (2) to transform the patch-position dependent features into a sequence of high-level features without 3D positional dependencies, which we empirically find easier for an LLM to fit (See Sec. 4.3). Specifically, given a video clip with a duration of two seconds, we sample 5 frames at 2 fps, resulting in a total of 64 video tokens. We adopt the de-noising U-Net in DynamiCrafter [78], but reduce the input channel of the 3D convolution from 8 to 4 since we remove the original concatenation of a conditional image with noisy latents." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 313, + 502, + 537, + 514 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 502, + 537, + 514 + ], + "spans": [ + { + "bbox": [ + 313, + 502, + 537, + 514 + ], + "type": "text", + "content": "3.2. Video Representation Modeling with LLM" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 313, + 521, + 556, + 676 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 521, + 556, + 676 + ], + "spans": [ + { + "bbox": [ + 313, + 521, + 556, + 676 + ], + "type": "text", + "content": "The core challenge of generating videos using a LLM with the Divot tokenizer lies in effectively modeling the continuous video features. The most straightforward solution is to minimize the distance between the LLM output and the video representations using mean squared error (MSE) loss following previous work [17, 54] for image generation as illustrated in Fig. 4 (a). However, we empirically find that this approach is not effective for modeling continuous video features, as the generated videos tend to exhibit repeating patterns. We analyze that the deterministic regression regularizes the LLM to learn overly averaged representations, which is particularly catastrophic in video generation as videos must ensure both spatial and temporal diversity." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 313, + 677, + 556, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 677, + 556, + 713 + ], + "spans": [ + { + "bbox": [ + 313, + 677, + 556, + 713 + ], + "type": "text", + "content": "Inspired by recent work MAR [33], instead of deterministic regression, we aim to model the probability distribution of video representations using the LLM. As shown in Fig. 4," + } + ] + } + ], + "index": 13 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "text", + "content": "13609" + } + ] + } + ], + "index": 14 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 3 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 60, + 83, + 198, + 152 + ], + "blocks": [ + { + "bbox": [ + 94, + 74, + 159, + 83 + ], + "lines": [ + { + "bbox": [ + 94, + 74, + 159, + 83 + ], + "spans": [ + { + "bbox": [ + 94, + 74, + 159, + 83 + ], + "type": "text", + "content": "Input video (fps=2)" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 60, + 83, + 198, + 152 + ], + "lines": [ + { + "bbox": [ + 60, + 83, + 198, + 152 + ], + "spans": [ + { + "bbox": [ + 60, + 83, + 198, + 152 + ], + "type": "image", + "image_path": "e028b6c9192fbea1821cc07ff0fe92c835c98e38529b8974e24655cf6c2b4978.jpg" + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_body" + } + ], + "index": 1 + }, + { + "type": "image", + "bbox": [ + 60, + 154, + 197, + 222 + ], + "blocks": [ + { + "bbox": [ + 60, + 154, + 197, + 222 + ], + "lines": [ + { + "bbox": [ + 60, + 154, + 197, + 222 + ], + "spans": [ + { + "bbox": [ + 60, + 154, + 197, + 222 + ], + "type": "image", + "image_path": "4fd0c41406fe355eb7d0c864097144f86308dd90d94c8798281db3ccce5449d7.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + } + ], + "index": 2 + }, + { + "type": "image", + "bbox": [ + 60, + 224, + 197, + 292 + ], + "blocks": [ + { + "bbox": [ + 60, + 224, + 197, + 292 + ], + "lines": [ + { + "bbox": [ + 60, + 224, + 197, + 292 + ], + "spans": [ + { + "bbox": [ + 60, + 224, + 197, + 292 + ], + "type": "image", + "image_path": "c76ac15c2c3f96328f358f06af6cc5f10fa785219565aa151ac4116a61eb3b50.jpg" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 55, + 302, + 555, + 326 + ], + "lines": [ + { + "bbox": [ + 55, + 302, + 555, + 326 + ], + "spans": [ + { + "bbox": [ + 55, + 302, + 555, + 326 + ], + "type": "text", + "content": "Figure 5. Reconstructed videos, where the Divot tokenizer obtains spatiotemporal representations of sparsely sampled video frames and the de-tokenizer decodes these representations into semantically aligned and temporally coherent video clips." + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "image_caption" + } + ], + "index": 3 + }, + { + "type": "image", + "bbox": [ + 209, + 83, + 552, + 152 + ], + "blocks": [ + { + "bbox": [ + 332, + 74, + 427, + 83 + ], + "lines": [ + { + "bbox": [ + 332, + 74, + 427, + 83 + ], + "spans": [ + { + "bbox": [ + 332, + 74, + 427, + 83 + ], + "type": "text", + "content": "Reconstructed video (fps=8)" + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 209, + 83, + 552, + 152 + ], + "lines": [ + { + "bbox": [ + 209, + 83, + 552, + 152 + ], + "spans": [ + { + "bbox": [ + 209, + 83, + 552, + 152 + ], + "type": "image", + "image_path": "96a417e54fcba8ad56e881dc765b9369ff474d383b2dc1fff2f1c716947ee215.jpg" + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "image_body" + } + ], + "index": 5 + }, + { + "type": "image", + "bbox": [ + 209, + 154, + 552, + 222 + ], + "blocks": [ + { + "bbox": [ + 209, + 154, + 552, + 222 + ], + "lines": [ + { + "bbox": [ + 209, + 154, + 552, + 222 + ], + "spans": [ + { + "bbox": [ + 209, + 154, + 552, + 222 + ], + "type": "image", + "image_path": "8bbd5fc6d4255b8814b2441e99df74428962677b3f768c5416b9b9d40b5e46b2.jpg" + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "image_body" + } + ], + "index": 6 + }, + { + "type": "image", + "bbox": [ + 209, + 224, + 552, + 292 + ], + "blocks": [ + { + "bbox": [ + 209, + 224, + 552, + 292 + ], + "lines": [ + { + "bbox": [ + 209, + 224, + 552, + 292 + ], + "spans": [ + { + "bbox": [ + 209, + 224, + 552, + 292 + ], + "type": "image", + "image_path": "d57101b1d7c19c1fa8232c33f41382eece25fe7e8079fa2ddec2ba631fb64ace.jpg" + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "image_body" + } + ], + "index": 7 + }, + { + "bbox": [ + 54, + 340, + 297, + 519 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 54, + 340, + 297, + 519 + ], + "spans": [ + { + "bbox": [ + 54, + 340, + 297, + 519 + ], + "type": "text", + "content": "we explore two approaches including (b) Diffusion Modeling [33] and (c) GMM Modeling [61]. Specifically, for the diffusion modeling, given continuous-valued video features to be predicted, the LLM produces output, which serves as the condition of a denoising network (a small MLP) to predict the Gaussian noise added to the video features. The diffusion model is trained for representing the distribution of video features. For GMM modeling, we use a Gaussian Mixture Model (GMM) to model the distribution of the video features, and train the LLM to predict " + }, + { + "bbox": [ + 54, + 340, + 297, + 519 + ], + "type": "inline_equation", + "content": "2kd + k" + }, + { + "bbox": [ + 54, + 340, + 297, + 519 + ], + "type": "text", + "content": " parameters per video token (" + }, + { + "bbox": [ + 54, + 340, + 297, + 519 + ], + "type": "inline_equation", + "content": "kd" + }, + { + "bbox": [ + 54, + 340, + 297, + 519 + ], + "type": "text", + "content": " mean and " + }, + { + "bbox": [ + 54, + 340, + 297, + 519 + ], + "type": "inline_equation", + "content": "kd" + }, + { + "bbox": [ + 54, + 340, + 297, + 519 + ], + "type": "text", + "content": " variance parameters for the mixture components, and " + }, + { + "bbox": [ + 54, + 340, + 297, + 519 + ], + "type": "inline_equation", + "content": "k" + }, + { + "bbox": [ + 54, + 340, + 297, + 519 + ], + "type": "text", + "content": " mixture probabilities). We optimize the LLM by minimizing the discrepancy between the predicted GMM distribution and the video representations with negative log-likelihood (NLL) loss." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 55, + 522, + 297, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 522, + 297, + 713 + ], + "spans": [ + { + "bbox": [ + 55, + 522, + 297, + 713 + ], + "type": "text", + "content": "During inference, in diffusion modeling, the denoising network denoise the final video features from Gaussian noise gradually by taking the LLM output as the condition. In GMM modeling, we draw samples from the predicted GMM distribution as the final video representations. To empirically investigate the effectiveness of the approaches above, we train the LLM with MSR-VTT [79], and evaluates text-to-video generation on test set with FVD [62] and similarity score [49] as the metric following previous work [26, 82]. As listed in Tab. 7, GMM modeling achieves better performance than diffusion modeling and MSE regression in video generation. We speculate that high-level features obtained by the Divot tokenizer are more sensitive to Gaussian noise compared to the VAE latents used by MAR, making training more challenging and resulting in suboptimal results. Therefore, we adopt GMM modeling to train Divot-LLM." + } + ] + } + ], + "index": 10 + }, + { + "type": "table", + "bbox": [ + 317, + 358, + 555, + 492 + ], + "blocks": [ + { + "bbox": [ + 315, + 339, + 552, + 350 + ], + "lines": [ + { + "bbox": [ + 315, + 339, + 552, + 350 + ], + "spans": [ + { + "bbox": [ + 315, + 339, + 552, + 350 + ], + "type": "text", + "content": "Table 2. Datasets used for training the tokenizer and Divot-LLM." + } + ] + } + ], + "index": 11, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 317, + 358, + 555, + 492 + ], + "lines": [ + { + "bbox": [ + 317, + 358, + 555, + 492 + ], + "spans": [ + { + "bbox": [ + 317, + 358, + 555, + 492 + ], + "type": "table", + "html": "
StageTypeDataset
TokenizerPure VideoWebVid-10M [2], Panda-70M [9]
Pre-trainVideo-textWebVid-10M [2]
Image-textCC3M [52], CapsFusion [87], LAION-COCO [51]
SFTClassificationKinetics-710 [27], SSV2 [18]
VQATGIF [34], NextQA [76], CLEVRER [85], YouCook2 [92], PerceptionTest[48], EgoQA [19], ActivityNetQA[88]
InstructionVideo-ChatGPT[43], LLaVA-mixed[39], Valley [42], LLaVA-Video-178K[37]
GenerationWebVid-10M [2]
StoryTellingIn-house data
", + "image_path": "9f15401bbe06cc251318e85e9b6ef6cf48e627608fc28a7381f71d7077015e1d.jpg" + } + ] + } + ], + "index": 12, + "angle": 0, + "type": "table_body" + } + ], + "index": 12 + }, + { + "bbox": [ + 313, + 512, + 506, + 525 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 512, + 506, + 525 + ], + "spans": [ + { + "bbox": [ + 313, + 512, + 506, + 525 + ], + "type": "text", + "content": "3.3. Pre-training and Instruction Tuning" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 313, + 530, + 523, + 542 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 530, + 523, + 542 + ], + "spans": [ + { + "bbox": [ + 313, + 530, + 523, + 542 + ], + "type": "text", + "content": "3.3.1. Training Stage I: Multimodal Pre-training" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 312, + 545, + 555, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 312, + 545, + 555, + 713 + ], + "spans": [ + { + "bbox": [ + 312, + 545, + 555, + 713 + ], + "type": "text", + "content": "As shown in Fig 3, Divot-LLM adopts next-word prediction and GMM modeling on video-text data for video comprehension and generation. Specifically, the video features from the Divot tokenizer, the special tokens indicating the start and end of video features, along with the text tokens of the caption are fed into the pre-trained Mistral-7B [24] for next token prediction trained with cross-entropy loss. Text tokens of the caption and " + }, + { + "bbox": [ + 312, + 545, + 555, + 713 + ], + "type": "inline_equation", + "content": "N" + }, + { + "bbox": [ + 312, + 545, + 555, + 713 + ], + "type": "text", + "content": " learnable queries are input into the LLM, where the output of the learnable queries are trained via bidirectional attention to model a GMM distribution for the video features using NLL loss. During inference, we draw samples from the predicted GMM distribution as condition of the denoising U-Net to decode videos. We pre-train Divot-LLM from the pre-trained Mistral-7B model using" + } + ] + } + ], + "index": 15 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "text", + "content": "13610" + } + ] + } + ], + "index": 16 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 4 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 93, + 102, + 520, + 313 + ], + "blocks": [ + { + "bbox": [ + 55, + 70, + 555, + 93 + ], + "lines": [ + { + "bbox": [ + 55, + 70, + 555, + 93 + ], + "spans": [ + { + "bbox": [ + 55, + 70, + 555, + 93 + ], + "type": "text", + "content": "Table 3. Comparison for video comprehension with MLLMs. \"Video-Gen\" denotes whether the model can generate videos besides texts. The evaluation metric is accuracy. The best results are bold and the second best results are underlined." + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 93, + 102, + 520, + 313 + ], + "lines": [ + { + "bbox": [ + 93, + 102, + 520, + 313 + ], + "spans": [ + { + "bbox": [ + 93, + 102, + 520, + 313 + ], + "type": "table", + "html": "
ModelLLM sizeVideo-GenEgoSchemaPerception-TestMVBenchMSVDActivityNet
Gemini 1.0 Pro [58]-×55.751.1--49.8
Gemini 1.5 Pro [59]-×63.2---56.7
GPT4-V [46]-×55.6-43.7-59.5
GPT4-O [47]-×72.2---61.9
LLaMA-VID [35]7B×38.544.641.969.747.4
Video-ChatGPT [43]7B×---64.935.2
Video-LLaVA [37]7B×38.444.341.070.745.3
VideoChat2 [31]7B×42.247.351.170.049.1
LLaVA-NeXT-Video [38]7B×43.948.846.567.853.5
LLaVA-NeXT-Video [38]32B×60.9---54.3
PLLAVA [81]34B×-58.1--60.9
LLaVA-OneVision [30]72B×62.0---62.3
VideoLLaMA2 [10]7B×51.751.454.670.950.2
VideoLLaMA2 [10]72B×63.957.562.071.055.2
LWM [40]7B✓---55.9-
Video-LaVIT [26]7B✓37.347.9-73.250.1
VILA-U [74]7B✓---75.352.7
Divot-LLM7B✓46.558.352.176.455.8
", + "image_path": "e70e9b3a2c4310f4a435710d54da999e4aa2636f1ec784a27382e6faf7e24baf.jpg" + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "table_body" + } + ], + "index": 1 + }, + { + "bbox": [ + 55, + 333, + 295, + 357 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 333, + 295, + 357 + ], + "spans": [ + { + "bbox": [ + 55, + 333, + 295, + 357 + ], + "type": "text", + "content": "LoRA [22] on a subset of WebVid-10M [2] data (filtered for temporal dynamics in captions) and image-text data." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 55, + 363, + 295, + 376 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 363, + 295, + 376 + ], + "spans": [ + { + "bbox": [ + 55, + 363, + 295, + 376 + ], + "type": "text", + "content": "3.3.2. Training Stage II: Multimodal Instruction Tuning" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 55, + 378, + 296, + 462 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 378, + 296, + 462 + ], + "spans": [ + { + "bbox": [ + 55, + 378, + 296, + 462 + ], + "type": "text", + "content": "We perform multimodal instruction tuning on Divot-LLM to align it with human instructions through supervised fine-tuning on public datasets as listed in Tab. 2 with a LoRA module. We further fine-tune the pretrained Divot-LLM on an animated series called \"Curious George\" to achieve video storytelling, which generates storyline and corresponding video clips in an interleaved manner." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 55, + 473, + 132, + 487 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 473, + 132, + 487 + ], + "spans": [ + { + "bbox": [ + 55, + 473, + 132, + 487 + ], + "type": "text", + "content": "4. Experiment" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 55, + 492, + 190, + 504 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 492, + 190, + 504 + ], + "spans": [ + { + "bbox": [ + 55, + 492, + 190, + 504 + ], + "type": "text", + "content": "4.1.Quantitative Evaluation" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 55, + 510, + 295, + 617 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 510, + 295, + 617 + ], + "spans": [ + { + "bbox": [ + 55, + 510, + 295, + 617 + ], + "type": "text", + "content": "Video Comprehension. We conduct extensive evaluations on video comprehension benchmarks including Multi-choice Video Question Answering (MC-VQA) on EgoSchema [44], Perception-Test [48], MVBench [32], and Open-Ended Video Question Answering (OE-VQA) on MSVD [7], ActivityNet [88]. Following VideoLLaMA 2 [10], we utilize GPT-3.5 to assess the quality of the generated answers of OE-VQA by determining whether the answers match the ground truth, and we report the percentage of \"Yes\" as Accuracy." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 55, + 618, + 296, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 618, + 296, + 714 + ], + "spans": [ + { + "bbox": [ + 55, + 618, + 296, + 714 + ], + "type": "text", + "content": "For each testing video, we sample a maximum of 20 clips, each containing 5 frames. The evaluation results are reported in Tab. 3. Divot-LLM outperforms the baseline models that can generate both texts and videos, which demonstrates that our model effectively achieves video comprehension within a unified framework. Compared to VideoLLMs specifically designed for video comprehension of the same model size of LLM, Divot-LLM achieves competitive results with sig" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 313, + 333, + 555, + 393 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 333, + 555, + 393 + ], + "spans": [ + { + "bbox": [ + 313, + 333, + 555, + 393 + ], + "type": "text", + "content": "nificantly fewer video-caption pairs for training (4.8M vs. 100M in VideoLLaMA 2). By utilizing diffusion procedure for video representation learning, our Divot tokenizer effectively captures robust spatiotemporal representations, enhancing the comprehension capabilities." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 313, + 403, + 556, + 523 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 403, + 556, + 523 + ], + "spans": [ + { + "bbox": [ + 313, + 403, + 556, + 523 + ], + "type": "text", + "content": "Video Generation. We evaluate zero-shot text-to-video generation on MSR-VTT [79]. We randomly sample one caption for each testing video and generate 16 frames in " + }, + { + "bbox": [ + 313, + 403, + 556, + 523 + ], + "type": "inline_equation", + "content": "256 \\times 256" + }, + { + "bbox": [ + 313, + 403, + 556, + 523 + ], + "type": "text", + "content": "px resolution. We adopt the CLIP similarity (CLIPSIM) [71] and Fréchet video distance (FVD) [62] as the evaluation metric following Loong [69]. As listed in Tab. 4, Divot-LLM achieves performance comparable to existing video generation models in terms of visual quality and semantic alignment with captions using only 4.8 million video-text pairs for training." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 313, + 530, + 443, + 544 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 530, + 443, + 544 + ], + "spans": [ + { + "bbox": [ + 313, + 530, + 443, + 544 + ], + "type": "text", + "content": "4.2. Qualitative Evaluation" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 313, + 548, + 555, + 632 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 548, + 555, + 632 + ], + "spans": [ + { + "bbox": [ + 313, + 548, + 555, + 632 + ], + "type": "text", + "content": "Text-to-video Generation. We perform qualitative comparison of text-to-video generation with baseline MLLMs that are capable of unified video comprehension and generation. As illustrated in Fig. 6, through modeling the distributions of Divot features with a predicted GMM, our Divot-LLM can generate videos that are both semantically aligned with text prompts and temporally coherent within frames." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 313, + 641, + 556, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 641, + 556, + 714 + ], + "spans": [ + { + "bbox": [ + 313, + 641, + 556, + 714 + ], + "type": "text", + "content": "Video StoryTelling. We fine-tune the pre-trained Divot-LLM on an animated series called \"Curious George\" for video storytelling. As shown in Fig. 7, given a brief story instruction, our Divot-LLM can generate a sequence of multimodal stories with rich narrative text and contextually relevant videos that are temporally coherent. Since we only" + } + ] + } + ], + "index": 13 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 749, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 749, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 749, + 317, + 757 + ], + "type": "text", + "content": "13611" + } + ] + } + ], + "index": 14 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 5 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 61, + 85, + 310, + 262 + ], + "blocks": [ + { + "bbox": [ + 110, + 72, + 269, + 83 + ], + "lines": [ + { + "bbox": [ + 110, + 72, + 269, + 83 + ], + "spans": [ + { + "bbox": [ + 110, + 72, + 269, + 83 + ], + "type": "text", + "content": "A man walking slowly under a sunset." + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 61, + 85, + 310, + 262 + ], + "lines": [ + { + "bbox": [ + 61, + 85, + 310, + 262 + ], + "spans": [ + { + "bbox": [ + 61, + 85, + 310, + 262 + ], + "type": "image", + "image_path": "0b02b84649d532ae8e5681d4501877047acbf4edb7aeed1b24b63c74821bc370.jpg" + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_body" + } + ], + "index": 1 + }, + { + "type": "image", + "bbox": [ + 320, + 85, + 552, + 262 + ], + "blocks": [ + { + "bbox": [ + 364, + 72, + 506, + 84 + ], + "lines": [ + { + "bbox": [ + 364, + 72, + 506, + 84 + ], + "spans": [ + { + "bbox": [ + 364, + 72, + 506, + 84 + ], + "type": "text", + "content": "A person is applying eye makeup." + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 320, + 85, + 552, + 262 + ], + "lines": [ + { + "bbox": [ + 320, + 85, + 552, + 262 + ], + "spans": [ + { + "bbox": [ + 320, + 85, + 552, + 262 + ], + "type": "image", + "image_path": "a72e36cf9679763404ae7e2a336cd67d6453c7b60a526a4c8db7194d3542a39d.jpg" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 55, + 274, + 555, + 297 + ], + "lines": [ + { + "bbox": [ + 55, + 274, + 555, + 297 + ], + "spans": [ + { + "bbox": [ + 55, + 274, + 555, + 297 + ], + "type": "text", + "content": "Figure 6. Qualitative comparison of text-to-video generation with MLLMs that are capable of unified video comprehension and generation. Divot-LLM effectively generates videos that are both semantically aligned with text prompts and temporally coherent within frames." + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_caption" + } + ], + "index": 3 + }, + { + "bbox": [ + 55, + 306, + 296, + 343 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 306, + 296, + 343 + ], + "spans": [ + { + "bbox": [ + 55, + 306, + 296, + 343 + ], + "type": "text", + "content": "fine-tune the de-tokenizer for adaptation to the new domain, it demonstrates the generalizability of our Divot tokenizer for obtaining robust video representations." + } + ] + } + ], + "index": 5 + }, + { + "type": "table", + "bbox": [ + 58, + 408, + 296, + 566 + ], + "blocks": [ + { + "bbox": [ + 55, + 354, + 298, + 399 + ], + "lines": [ + { + "bbox": [ + 55, + 354, + 298, + 399 + ], + "spans": [ + { + "bbox": [ + 55, + 354, + 298, + 399 + ], + "type": "text", + "content": "Table 4. Comparison for zero-shot text-to-video generation. \"Data size\" refers to the number of training video data, and \"Unified\" denotes if the model enables video comprehension and generation. The best results are bold and the second best results are underlined." + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 58, + 408, + 296, + 566 + ], + "lines": [ + { + "bbox": [ + 58, + 408, + 296, + 566 + ], + "spans": [ + { + "bbox": [ + 58, + 408, + 296, + 566 + ], + "type": "table", + "html": "
ModelData sizeUnifiedMSR-VTT
CLIPSIM (↑)FVD (↓)
CogVideo [21]5.4M×0.26311294
Video LDM [5]10M×0.2929-
VideoComposer [66]10M×0.2932580
InternVid [68]28M×0.2951-
Make-A-Video [53]20M×0.3049-
VideoPoet [29]270M×0.3049213
PYoCo [14]22.5M×--
SVD [4]152M×--
Video-LavIT [26]10M✓0.3012188.36
Loong [69]16M×0.2903274
Snap Video [45]-×0.2793110.4
VILA-U [74]1M✓0.2937499.06
Divot-LLM4.8M✓0.2938301.4
", + "image_path": "6257c1027943f903ed565d39ab3884268d51214b1738eef774b7be63b5673eab.jpg" + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "table_body" + } + ], + "index": 7 + }, + { + "bbox": [ + 55, + 587, + 149, + 601 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 587, + 149, + 601 + ], + "spans": [ + { + "bbox": [ + 55, + 587, + 149, + 601 + ], + "type": "text", + "content": "4.3. Ablation Study" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 55, + 605, + 297, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 605, + 297, + 713 + ], + "spans": [ + { + "bbox": [ + 55, + 605, + 297, + 713 + ], + "type": "text", + "content": "Diffusion for Video Comprehension. We design two baselines to validate the effectiveness of the diffusion procedure to learn spatiotemporal representations for VideoLLMs. As shown in Tab. 5, both models are pre-trained on Valley [42] and instruction tuned on Video-ChatGPT [43]. The model with diffusion loss employs our Divot tokenizer, while the model with caption loss adopts the same architecture but its tokenizer is pre-trained using captioning loss with a frozen LLM on Valley. The model that employs a video tokenizer" + } + ] + } + ], + "index": 9 + }, + { + "type": "table", + "bbox": [ + 340, + 335, + 529, + 381 + ], + "blocks": [ + { + "bbox": [ + 313, + 304, + 555, + 327 + ], + "lines": [ + { + "bbox": [ + 313, + 304, + 555, + 327 + ], + "spans": [ + { + "bbox": [ + 313, + 304, + 555, + 327 + ], + "type": "text", + "content": "Table 6. Ablation study on the training objective of the video tokenizer. The evaluation metric is accuracy." + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 340, + 335, + 529, + 381 + ], + "lines": [ + { + "bbox": [ + 340, + 335, + 529, + 381 + ], + "spans": [ + { + "bbox": [ + 340, + 335, + 529, + 381 + ], + "type": "table", + "html": "
Loss TypeMV-BenchMSVDActivityNet
Caption30.866.143.2
Diffusion33.268.944.3
", + "image_path": "bc68c23cce5556373a0b767eb8bcecd820fd2298fa560cfe748e8f99e0b68bb9.jpg" + } + ] + } + ], + "index": 11, + "angle": 0, + "type": "table_body" + } + ], + "index": 11 + }, + { + "bbox": [ + 313, + 392, + 555, + 453 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 392, + 555, + 453 + ], + "spans": [ + { + "bbox": [ + 313, + 392, + 555, + 453 + ], + "type": "text", + "content": "trained with diffusion loss achieves better performance in video comprehension benchmarks, demonstrating that the diffusion process can effectively learn robust video representations in a self-supervised manner, without the need for paired caption annotations." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 313, + 462, + 555, + 521 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 462, + 555, + 521 + ], + "spans": [ + { + "bbox": [ + 313, + 462, + 555, + 521 + ], + "type": "text", + "content": "Video Generation with LLM. We perform various ablation studies to explore an effective approach for generating videos with a LLM through training on MSR-VTT training set and evaluating text-to-video generation on test set. We use ViT-G/14 to calculate CLIPSIM for better discrimination." + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 313, + 522, + 556, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 522, + 556, + 714 + ], + "spans": [ + { + "bbox": [ + 313, + 522, + 556, + 714 + ], + "type": "text", + "content": "Q1: Which type of video representations is easier? We investigate two types of video representations, patch-position dependent features obtained from a spatial-temporal transformer and patch-position independent features after a Perceiver Resample with learnable queries. As listed in Tab. 7, fitting features without 3D positional dependencies achieves higher performance, which is also observed in recent work [64]. We also experiment with training the video tokenizer in the VAE manner, which involves predicting the means and variances of a normal distribution and sampling video representations using the re-parametrization trick following GIVT [61]. However, we observe that it is difficult for the LLM to converge and the video de-tokenizer achieves unsatisfactory reconstruction results. We conclude that introducing variances during tokenization for high-level video features may not be appropriate." + } + ] + } + ], + "index": 14 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 318, + 758 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 318, + 758 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 318, + 758 + ], + "type": "text", + "content": "13612" + } + ] + } + ], + "index": 15 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 6 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 61, + 95, + 77, + 107 + ], + "blocks": [ + { + "bbox": [ + 58, + 73, + 264, + 87 + ], + "lines": [ + { + "bbox": [ + 58, + 73, + 264, + 87 + ], + "spans": [ + { + "bbox": [ + 58, + 73, + 264, + 87 + ], + "type": "text", + "content": "Generate a story about the adventures of George and his friends." + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 61, + 95, + 77, + 107 + ], + "lines": [ + { + "bbox": [ + 61, + 95, + 77, + 107 + ], + "spans": [ + { + "bbox": [ + 61, + 95, + 77, + 107 + ], + "type": "image", + "image_path": "7022943fd68f23b1372a3c781dd5842396777c6dec5f8e41997da890f68aa810.jpg" + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_body" + } + ], + "index": 1 + }, + { + "bbox": [ + 64, + 110, + 124, + 140 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 64, + 110, + 124, + 140 + ], + "spans": [ + { + "bbox": [ + 64, + 110, + 124, + 140 + ], + "type": "text", + "content": "A young girl found a glowing orb floating in the air through her magnifying glass." + } + ] + } + ], + "index": 2 + }, + { + "type": "image", + "bbox": [ + 129, + 106, + 170, + 144 + ], + "blocks": [ + { + "bbox": [ + 129, + 99, + 152, + 106 + ], + "lines": [ + { + "bbox": [ + 129, + 99, + 152, + 106 + ], + "spans": [ + { + "bbox": [ + 129, + 99, + 152, + 106 + ], + "type": "text", + "content": "REC" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 129, + 106, + 170, + 144 + ], + "lines": [ + { + "bbox": [ + 129, + 106, + 170, + 144 + ], + "spans": [ + { + "bbox": [ + 129, + 106, + 170, + 144 + ], + "type": "image", + "image_path": "166b68bf5ff8b78bdad2ad3ffdb815b017565f5380946f18a970cf0a8f9905ec.jpg" + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_body" + } + ], + "index": 4 + }, + { + "type": "image", + "bbox": [ + 171, + 106, + 210, + 144 + ], + "blocks": [ + { + "bbox": [ + 171, + 106, + 210, + 144 + ], + "lines": [ + { + "bbox": [ + 171, + 106, + 210, + 144 + ], + "spans": [ + { + "bbox": [ + 171, + 106, + 210, + 144 + ], + "type": "image", + "image_path": "69b87821acbc5e3da29188c4bca7b83be87ac446d3727123fc2707af0d18b3cb.jpg" + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "image_body" + } + ], + "index": 5 + }, + { + "type": "image", + "bbox": [ + 211, + 106, + 249, + 144 + ], + "blocks": [ + { + "bbox": [ + 211, + 106, + 249, + 144 + ], + "lines": [ + { + "bbox": [ + 211, + 106, + 249, + 144 + ], + "spans": [ + { + "bbox": [ + 211, + 106, + 249, + 144 + ], + "type": "image", + "image_path": "2b4d3b08741527aac498461eeb2aaf3e64018eaf6239c3d87bd71bdf618cba05.jpg" + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "image_body" + } + ], + "index": 6 + }, + { + "type": "image", + "bbox": [ + 250, + 106, + 290, + 144 + ], + "blocks": [ + { + "bbox": [ + 250, + 106, + 290, + 144 + ], + "lines": [ + { + "bbox": [ + 250, + 106, + 290, + 144 + ], + "spans": [ + { + "bbox": [ + 250, + 106, + 290, + 144 + ], + "type": "image", + "image_path": "cf4fc4340b189061969660e8c73d648e7afa42af4aaa7b7c5ab443e98007191e.jpg" + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "image_body" + } + ], + "index": 7 + }, + { + "type": "image", + "bbox": [ + 63, + 145, + 77, + 157 + ], + "blocks": [ + { + "bbox": [ + 63, + 145, + 77, + 157 + ], + "lines": [ + { + "bbox": [ + 63, + 145, + 77, + 157 + ], + "spans": [ + { + "bbox": [ + 63, + 145, + 77, + 157 + ], + "type": "image", + "image_path": "58d0788f6a4f1921e7abce5a5b43bcc982c73f5efbfeaa6e35bd3d1e7a28cff5.jpg" + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "image_body" + } + ], + "index": 8 + }, + { + "bbox": [ + 63, + 165, + 123, + 188 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 63, + 165, + 123, + 188 + ], + "spans": [ + { + "bbox": [ + 63, + 165, + 123, + 188 + ], + "type": "text", + "content": "George watched a bubble float by, feeling curious and joyful." + } + ] + } + ], + "index": 9 + }, + { + "type": "image", + "bbox": [ + 129, + 158, + 171, + 198 + ], + "blocks": [ + { + "bbox": [ + 129, + 152, + 153, + 158 + ], + "lines": [ + { + "bbox": [ + 129, + 152, + 153, + 158 + ], + "spans": [ + { + "bbox": [ + 129, + 152, + 153, + 158 + ], + "type": "text", + "content": "REC" + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 129, + 158, + 171, + 198 + ], + "lines": [ + { + "bbox": [ + 129, + 158, + 171, + 198 + ], + "spans": [ + { + "bbox": [ + 129, + 158, + 171, + 198 + ], + "type": "image", + "image_path": "663b0c511275a713084ce388d5e58858ad353e5c439a03fc798bbea67e5a7d5e.jpg" + } + ] + } + ], + "index": 11, + "angle": 0, + "type": "image_body" + } + ], + "index": 11 + }, + { + "type": "image", + "bbox": [ + 171, + 152, + 210, + 198 + ], + "blocks": [ + { + "bbox": [ + 171, + 152, + 210, + 198 + ], + "lines": [ + { + "bbox": [ + 171, + 152, + 210, + 198 + ], + "spans": [ + { + "bbox": [ + 171, + 152, + 210, + 198 + ], + "type": "image", + "image_path": "6571762ac9125fde0225c61cb424cf678e8ab746d45ee3ea2257e35bf1df6de9.jpg" + } + ] + } + ], + "index": 12, + "angle": 0, + "type": "image_body" + } + ], + "index": 12 + }, + { + "type": "image", + "bbox": [ + 211, + 152, + 249, + 198 + ], + "blocks": [ + { + "bbox": [ + 211, + 152, + 249, + 198 + ], + "lines": [ + { + "bbox": [ + 211, + 152, + 249, + 198 + ], + "spans": [ + { + "bbox": [ + 211, + 152, + 249, + 198 + ], + "type": "image", + "image_path": "732d370b18165ad577165a633aaef6d61e0ed000fc3db1b4c15ea1d242026f7d.jpg" + } + ] + } + ], + "index": 13, + "angle": 0, + "type": "image_body" + } + ], + "index": 13 + }, + { + "type": "image", + "bbox": [ + 250, + 152, + 290, + 198 + ], + "blocks": [ + { + "bbox": [ + 250, + 152, + 290, + 198 + ], + "lines": [ + { + "bbox": [ + 250, + 152, + 290, + 198 + ], + "spans": [ + { + "bbox": [ + 250, + 152, + 290, + 198 + ], + "type": "image", + "image_path": "a73b7c641c469b9ad59751c25596f4e2c5e5d5f575f32333cfc260c2ed1f1d12.jpg" + } + ] + } + ], + "index": 14, + "angle": 0, + "type": "image_body" + } + ], + "index": 14 + }, + { + "type": "image", + "bbox": [ + 63, + 201, + 77, + 212 + ], + "blocks": [ + { + "bbox": [ + 63, + 201, + 77, + 212 + ], + "lines": [ + { + "bbox": [ + 63, + 201, + 77, + 212 + ], + "spans": [ + { + "bbox": [ + 63, + 201, + 77, + 212 + ], + "type": "image", + "image_path": "d114e83331db527c59ec8e749906a2a7e4644a2aae4b0100d3f802abbed57ec0.jpg" + } + ] + } + ], + "index": 15, + "angle": 0, + "type": "image_body" + } + ], + "index": 15 + }, + { + "bbox": [ + 63, + 217, + 124, + 239 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 63, + 217, + 124, + 239 + ], + "spans": [ + { + "bbox": [ + 63, + 217, + 124, + 239 + ], + "type": "text", + "content": "At lunchtime, a boy found a giant sandwich and felt excited." + } + ] + } + ], + "index": 16 + }, + { + "type": "image", + "bbox": [ + 129, + 210, + 171, + 249 + ], + "blocks": [ + { + "bbox": [ + 129, + 204, + 153, + 210 + ], + "lines": [ + { + "bbox": [ + 129, + 204, + 153, + 210 + ], + "spans": [ + { + "bbox": [ + 129, + 204, + 153, + 210 + ], + "type": "text", + "content": "REC" + } + ] + } + ], + "index": 17, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 129, + 210, + 171, + 249 + ], + "lines": [ + { + "bbox": [ + 129, + 210, + 171, + 249 + ], + "spans": [ + { + "bbox": [ + 129, + 210, + 171, + 249 + ], + "type": "image", + "image_path": "ca2be2eabe2d27eac255c7ae80637d81d6f21998f07b2ca0c21163cb09471cdc.jpg" + } + ] + } + ], + "index": 18, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 55, + 263, + 555, + 286 + ], + "lines": [ + { + "bbox": [ + 55, + 263, + 555, + 286 + ], + "spans": [ + { + "bbox": [ + 55, + 263, + 555, + 286 + ], + "type": "text", + "content": "Figure 7. Qualitative examples of video storytelling by Divot-LLM. Given a story instruction, Divot-LLM can generate rich textual narratives along with corresponding video clips that are temporally coherent in an interleaved manner." + } + ] + } + ], + "index": 44, + "angle": 0, + "type": "image_caption" + } + ], + "index": 18 + }, + { + "type": "image", + "bbox": [ + 171, + 210, + 210, + 249 + ], + "blocks": [ + { + "bbox": [ + 171, + 210, + 210, + 249 + ], + "lines": [ + { + "bbox": [ + 171, + 210, + 210, + 249 + ], + "spans": [ + { + "bbox": [ + 171, + 210, + 210, + 249 + ], + "type": "image", + "image_path": "2b90b82b5ddc80533a7301225c5f54c341b591e9d15d784749954fd092069cfa.jpg" + } + ] + } + ], + "index": 19, + "angle": 0, + "type": "image_body" + } + ], + "index": 19 + }, + { + "type": "image", + "bbox": [ + 211, + 210, + 249, + 249 + ], + "blocks": [ + { + "bbox": [ + 211, + 210, + 249, + 249 + ], + "lines": [ + { + "bbox": [ + 211, + 210, + 249, + 249 + ], + "spans": [ + { + "bbox": [ + 211, + 210, + 249, + 249 + ], + "type": "image", + "image_path": "cc0e86a83283ba3f37a38fec425b26e09d972fdb58e269108aa9b2abb107a5b8.jpg" + } + ] + } + ], + "index": 20, + "angle": 0, + "type": "image_body" + } + ], + "index": 20 + }, + { + "type": "image", + "bbox": [ + 250, + 210, + 290, + 249 + ], + "blocks": [ + { + "bbox": [ + 250, + 210, + 290, + 249 + ], + "lines": [ + { + "bbox": [ + 250, + 210, + 290, + 249 + ], + "spans": [ + { + "bbox": [ + 250, + 210, + 290, + 249 + ], + "type": "image", + "image_path": "82930692f9ad0fa4d1d4614bdfa912a1a08edf357afd5414070d44d6853bdc26.jpg" + } + ] + } + ], + "index": 21, + "angle": 0, + "type": "image_body" + } + ], + "index": 21 + }, + { + "bbox": [ + 328, + 78, + 507, + 87 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 328, + 78, + 507, + 87 + ], + "spans": [ + { + "bbox": [ + 328, + 78, + 507, + 87 + ], + "type": "text", + "content": "Generate a story about George's fun-filled day in the kitchen." + } + ] + } + ], + "index": 22 + }, + { + "type": "image", + "bbox": [ + 325, + 96, + 339, + 107 + ], + "blocks": [ + { + "bbox": [ + 325, + 96, + 339, + 107 + ], + "lines": [ + { + "bbox": [ + 325, + 96, + 339, + 107 + ], + "spans": [ + { + "bbox": [ + 325, + 96, + 339, + 107 + ], + "type": "image", + "image_path": "3d06225ce82fabf69e2ea3482d03d0af3497b5ffb1c7919492f34fd49b8a495b.jpg" + } + ] + } + ], + "index": 23, + "angle": 0, + "type": "image_body" + } + ], + "index": 23 + }, + { + "bbox": [ + 327, + 110, + 383, + 146 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 327, + 110, + 383, + 146 + ], + "spans": [ + { + "bbox": [ + 327, + 110, + 383, + 146 + ], + "type": "text", + "content": "George and his pal joyfully cook in the kitchen, creating a tasty snack with a big blue book." + } + ] + } + ], + "index": 24 + }, + { + "type": "image", + "bbox": [ + 388, + 106, + 427, + 144 + ], + "blocks": [ + { + "bbox": [ + 388, + 99, + 408, + 106 + ], + "lines": [ + { + "bbox": [ + 388, + 99, + 408, + 106 + ], + "spans": [ + { + "bbox": [ + 388, + 99, + 408, + 106 + ], + "type": "text", + "content": "REC" + } + ] + } + ], + "index": 25, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 388, + 106, + 427, + 144 + ], + "lines": [ + { + "bbox": [ + 388, + 106, + 427, + 144 + ], + "spans": [ + { + "bbox": [ + 388, + 106, + 427, + 144 + ], + "type": "image", + "image_path": "5f2a88b8f3450114f09b9326cada4bff1fd37d2023e4338c2d521444ac7931fc.jpg" + } + ] + } + ], + "index": 26, + "angle": 0, + "type": "image_body" + } + ], + "index": 26 + }, + { + "type": "image", + "bbox": [ + 428, + 106, + 467, + 144 + ], + "blocks": [ + { + "bbox": [ + 428, + 106, + 467, + 144 + ], + "lines": [ + { + "bbox": [ + 428, + 106, + 467, + 144 + ], + "spans": [ + { + "bbox": [ + 428, + 106, + 467, + 144 + ], + "type": "image", + "image_path": "142850e5fd9df78388957622b4ac806287f40852b5b651259457f944ca74f905.jpg" + } + ] + } + ], + "index": 27, + "angle": 0, + "type": "image_body" + } + ], + "index": 27 + }, + { + "type": "image", + "bbox": [ + 468, + 106, + 507, + 144 + ], + "blocks": [ + { + "bbox": [ + 468, + 106, + 507, + 144 + ], + "lines": [ + { + "bbox": [ + 468, + 106, + 507, + 144 + ], + "spans": [ + { + "bbox": [ + 468, + 106, + 507, + 144 + ], + "type": "image", + "image_path": "5a1ece1ff70131622e975eaf4fc2a7e230536045fa6d74f54f69e9bd3dc8ffd6.jpg" + } + ] + } + ], + "index": 28, + "angle": 0, + "type": "image_body" + } + ], + "index": 28 + }, + { + "type": "image", + "bbox": [ + 507, + 106, + 547, + 144 + ], + "blocks": [ + { + "bbox": [ + 507, + 106, + 547, + 144 + ], + "lines": [ + { + "bbox": [ + 507, + 106, + 547, + 144 + ], + "spans": [ + { + "bbox": [ + 507, + 106, + 547, + 144 + ], + "type": "image", + "image_path": "b5b94209683463df72a5d9a437d2130b0ee199beaea00fe41075d21d27cfc245.jpg" + } + ] + } + ], + "index": 29, + "angle": 0, + "type": "image_body" + } + ], + "index": 29 + }, + { + "type": "image", + "bbox": [ + 325, + 147, + 339, + 157 + ], + "blocks": [ + { + "bbox": [ + 325, + 147, + 339, + 157 + ], + "lines": [ + { + "bbox": [ + 325, + 147, + 339, + 157 + ], + "spans": [ + { + "bbox": [ + 325, + 147, + 339, + 157 + ], + "type": "image", + "image_path": "27f2070d8c1bbbb8732270a15cdd8b089e19fa85a03f0c20a666e8e316d77606.jpg" + } + ] + } + ], + "index": 30, + "angle": 0, + "type": "image_body" + } + ], + "index": 30 + }, + { + "bbox": [ + 327, + 159, + 376, + 196 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 327, + 159, + 376, + 196 + ], + "spans": [ + { + "bbox": [ + 327, + 159, + 376, + 196 + ], + "type": "text", + "content": "A woman in the kitchen shows George how to use a new kitchen gadget." + } + ] + } + ], + "index": 31 + }, + { + "type": "image", + "bbox": [ + 388, + 158, + 427, + 198 + ], + "blocks": [ + { + "bbox": [ + 388, + 152, + 408, + 158 + ], + "lines": [ + { + "bbox": [ + 388, + 152, + 408, + 158 + ], + "spans": [ + { + "bbox": [ + 388, + 152, + 408, + 158 + ], + "type": "text", + "content": "REC" + } + ] + } + ], + "index": 32, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 388, + 158, + 427, + 198 + ], + "lines": [ + { + "bbox": [ + 388, + 158, + 427, + 198 + ], + "spans": [ + { + "bbox": [ + 388, + 158, + 427, + 198 + ], + "type": "image", + "image_path": "792a57c607f69d7d4714ffb0e08a87a6d479a29703eeaac9e39eb9b37ca61a83.jpg" + } + ] + } + ], + "index": 33, + "angle": 0, + "type": "image_body" + } + ], + "index": 33 + }, + { + "type": "image", + "bbox": [ + 428, + 152, + 467, + 198 + ], + "blocks": [ + { + "bbox": [ + 428, + 152, + 467, + 198 + ], + "lines": [ + { + "bbox": [ + 428, + 152, + 467, + 198 + ], + "spans": [ + { + "bbox": [ + 428, + 152, + 467, + 198 + ], + "type": "image", + "image_path": "573a49f44d4708baba1104170e127f775fa5499d3154dd79b680b4272da71135.jpg" + } + ] + } + ], + "index": 34, + "angle": 0, + "type": "image_body" + } + ], + "index": 34 + }, + { + "type": "image", + "bbox": [ + 468, + 152, + 507, + 198 + ], + "blocks": [ + { + "bbox": [ + 468, + 152, + 507, + 198 + ], + "lines": [ + { + "bbox": [ + 468, + 152, + 507, + 198 + ], + "spans": [ + { + "bbox": [ + 468, + 152, + 507, + 198 + ], + "type": "image", + "image_path": "10ebc2477a30eb167a1a61be785ae2fc87fc190d81bbab877347a8daf41092ed.jpg" + } + ] + } + ], + "index": 35, + "angle": 0, + "type": "image_body" + } + ], + "index": 35 + }, + { + "type": "image", + "bbox": [ + 507, + 152, + 547, + 198 + ], + "blocks": [ + { + "bbox": [ + 507, + 152, + 547, + 198 + ], + "lines": [ + { + "bbox": [ + 507, + 152, + 547, + 198 + ], + "spans": [ + { + "bbox": [ + 507, + 152, + 547, + 198 + ], + "type": "image", + "image_path": "9133559835d3ddca4df6e59957ddaad9a07415d27411d881bcdf6ebbb41c1f19.jpg" + } + ] + } + ], + "index": 36, + "angle": 0, + "type": "image_body" + } + ], + "index": 36 + }, + { + "type": "image", + "bbox": [ + 325, + 201, + 338, + 212 + ], + "blocks": [ + { + "bbox": [ + 325, + 201, + 338, + 212 + ], + "lines": [ + { + "bbox": [ + 325, + 201, + 338, + 212 + ], + "spans": [ + { + "bbox": [ + 325, + 201, + 338, + 212 + ], + "type": "image", + "image_path": "709276f49ff6be0a6508a8f0d021aa8daa76f8f95b7ec0f644ddea61a57e15ac.jpg" + } + ] + } + ], + "index": 37, + "angle": 0, + "type": "image_body" + } + ], + "index": 37 + }, + { + "bbox": [ + 327, + 214, + 378, + 243 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 327, + 214, + 378, + 243 + ], + "spans": [ + { + "bbox": [ + 327, + 214, + 378, + 243 + ], + "type": "text", + "content": "George spreads the thick, peanutty paste, making a yummy snack." + } + ] + } + ], + "index": 38 + }, + { + "type": "image", + "bbox": [ + 388, + 210, + 427, + 248 + ], + "blocks": [ + { + "bbox": [ + 388, + 204, + 408, + 210 + ], + "lines": [ + { + "bbox": [ + 388, + 204, + 408, + 210 + ], + "spans": [ + { + "bbox": [ + 388, + 204, + 408, + 210 + ], + "type": "text", + "content": "REC" + } + ] + } + ], + "index": 39, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 388, + 210, + 427, + 248 + ], + "lines": [ + { + "bbox": [ + 388, + 210, + 427, + 248 + ], + "spans": [ + { + "bbox": [ + 388, + 210, + 427, + 248 + ], + "type": "image", + "image_path": "0c1a6a68b2ba9aa8a2cb86e28410f151e6a9743ba58489eeb286f11af4ec6b36.jpg" + } + ] + } + ], + "index": 40, + "angle": 0, + "type": "image_body" + } + ], + "index": 40 + }, + { + "type": "image", + "bbox": [ + 428, + 204, + 467, + 249 + ], + "blocks": [ + { + "bbox": [ + 428, + 204, + 467, + 249 + ], + "lines": [ + { + "bbox": [ + 428, + 204, + 467, + 249 + ], + "spans": [ + { + "bbox": [ + 428, + 204, + 467, + 249 + ], + "type": "image", + "image_path": "b7d19921d1fe7dd12f36779294f4c13045e76750383938f4cecad3a39730bf4f.jpg" + } + ] + } + ], + "index": 41, + "angle": 0, + "type": "image_body" + } + ], + "index": 41 + }, + { + "type": "image", + "bbox": [ + 468, + 204, + 507, + 249 + ], + "blocks": [ + { + "bbox": [ + 468, + 204, + 507, + 249 + ], + "lines": [ + { + "bbox": [ + 468, + 204, + 507, + 249 + ], + "spans": [ + { + "bbox": [ + 468, + 204, + 507, + 249 + ], + "type": "image", + "image_path": "a2f8f9a16452ed4ac767358618255562d029d7f8da9e1521a0722e6c68944bfd.jpg" + } + ] + } + ], + "index": 42, + "angle": 0, + "type": "image_body" + } + ], + "index": 42 + }, + { + "type": "image", + "bbox": [ + 507, + 204, + 547, + 249 + ], + "blocks": [ + { + "bbox": [ + 507, + 204, + 547, + 249 + ], + "lines": [ + { + "bbox": [ + 507, + 204, + 547, + 249 + ], + "spans": [ + { + "bbox": [ + 507, + 204, + 547, + 249 + ], + "type": "image", + "image_path": "e772131448626a9c7e4d15c5862b39644a727aaf1dcd316b35ace68f53b944a3.jpg" + } + ] + } + ], + "index": 43, + "angle": 0, + "type": "image_body" + } + ], + "index": 43 + }, + { + "type": "table", + "bbox": [ + 93, + 317, + 520, + 389 + ], + "blocks": [ + { + "bbox": [ + 55, + 297, + 553, + 308 + ], + "lines": [ + { + "bbox": [ + 55, + 297, + 553, + 308 + ], + "spans": [ + { + "bbox": [ + 55, + 297, + 553, + 308 + ], + "type": "text", + "content": "Table 8. Ablation study on video representation modeling with LLMs for generation. We evaluate text-to-video generation on MSR-VTT." + } + ] + } + ], + "index": 45, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 93, + 317, + 520, + 389 + ], + "lines": [ + { + "bbox": [ + 93, + 317, + 520, + 389 + ], + "spans": [ + { + "bbox": [ + 93, + 317, + 520, + 389 + ], + "type": "table", + "html": "
RepresentationObjectiveMechanism
patch-position dependentpatch-position independentMSEDiffusionGMMARQuery
ε-predv-predcausalbidirectional
CLIPSIM (↑)0.31920.32650.31680.28110.28420.32650.23860.30800.3265
FVD (↓)378.50366.60438.94418.19377.17366.60447.88416.60366.60
", + "image_path": "6f69fee1111bd205545916cc3f6dd3aae1f8c0817cd7818b0f18285d3530455a.jpg" + } + ] + } + ], + "index": 46, + "angle": 0, + "type": "table_body" + } + ], + "index": 46 + }, + { + "bbox": [ + 54, + 409, + 297, + 613 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 54, + 409, + 297, + 613 + ], + "spans": [ + { + "bbox": [ + 54, + 409, + 297, + 613 + ], + "type": "text", + "content": "Q2: Which training objective is more suitable? As introduce in Sec. 3, we explore MSE regression, Diffusion modeling and GMM modeling to fit high-level continuous features with a LLM. As listed in Tab. 7, simply aligning the LLM output with video features using MSE loss yields the lowest generation quality, suggesting that deterministic regression is inadequate for modeling spatiotemporal representations. Training a denoising network to denoise the noisy video features by taking the LLM output as the condition also achieves inferior performance with both " + }, + { + "bbox": [ + 54, + 409, + 297, + 613 + ], + "type": "inline_equation", + "content": "\\epsilon" + }, + { + "bbox": [ + 54, + 409, + 297, + 613 + ], + "type": "text", + "content": " prediction and " + }, + { + "bbox": [ + 54, + 409, + 297, + 613 + ], + "type": "inline_equation", + "content": "v" + }, + { + "bbox": [ + 54, + 409, + 297, + 613 + ], + "type": "text", + "content": " prediction. Different from MAR [33] that denoises low-level VAE latents, our goal is to denoise high-level video features. We speculate that these features are more sensitive to Gaussian noise, making them more challenging to denoise. Training the LLM to model the distribution of high-level video features using a GMM model achieves the best generation quality and semantic alignment with captions." + } + ] + } + ], + "index": 47 + }, + { + "bbox": [ + 54, + 617, + 297, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 54, + 617, + 297, + 714 + ], + "spans": [ + { + "bbox": [ + 54, + 617, + 297, + 714 + ], + "type": "text", + "content": "Q3: Which LLM mechanism is more effective? We train the LLM to fit the video representations with GMM modeling using both an autoregressive approach and a query-based approach, with the latter exploring causal attention and bidirectional attention within the LLM. Predicting the features of each video token in an autoregressive manner results in the worst performance due to the accumulation of errors, particularly when the features of the pervious to-" + } + ] + } + ], + "index": 48 + }, + { + "bbox": [ + 313, + 409, + 556, + 458 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 409, + 556, + 458 + ], + "spans": [ + { + "bbox": [ + 313, + 409, + 556, + 458 + ], + "type": "text", + "content": "ken are sampled from a GMM distribution for predicting the distribution of the current token. The query-based approach achieves better results with bidirectional attention, as it enables each query to attend to all tokens for predictions." + } + ] + } + ], + "index": 49 + }, + { + "bbox": [ + 313, + 475, + 388, + 488 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 475, + 388, + 488 + ], + "spans": [ + { + "bbox": [ + 313, + 475, + 388, + 488 + ], + "type": "text", + "content": "5. Conclusion" + } + ] + } + ], + "index": 50 + }, + { + "bbox": [ + 313, + 498, + 556, + 631 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 498, + 556, + 631 + ], + "spans": [ + { + "bbox": [ + 313, + 498, + 556, + 631 + ], + "type": "text", + "content": "In this work, we introduce Divot, a diffusion-powered video tokenizer learned in a self-supervised manner for unified comprehension and generation. We further investigate effective approaches for modeling continuous video representations with the LLM and present Divot-LLM to understand and generate video content in a single framework. Divot-LLM achieves competitive performance in video comprehension and generation benchmarks, and enables video storytelling effectively. We hope our work will draw increased attention to unifying video comprehension and generation through the design of sophisticated tokenizers." + } + ] + } + ], + "index": 51 + }, + { + "bbox": [ + 313, + 654, + 556, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 654, + 556, + 715 + ], + "spans": [ + { + "bbox": [ + 313, + 654, + 556, + 715 + ], + "type": "text", + "content": "Limitation. As we primarily focus on exploring effective representations and approaches for video generation with a unified LLM, the current model is trained to predict video representations for only a single clip and does not generate longer videos, which will be explored in our future work." + } + ] + } + ], + "index": 52 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "text", + "content": "13613" + } + ] + } + ], + "index": 53 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 7 + }, + { + "para_blocks": [ + { + "bbox": [ + 56, + 71, + 115, + 83 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 71, + 115, + 83 + ], + "spans": [ + { + "bbox": [ + 56, + 71, + 115, + 83 + ], + "type": "text", + "content": "References" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 57, + 91, + 297, + 713 + ], + "type": "list", + "angle": 0, + "index": 12, + "blocks": [ + { + "bbox": [ + 61, + 91, + 297, + 156 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 61, + 91, + 297, + 156 + ], + "spans": [ + { + "bbox": [ + 61, + 91, + 297, + 156 + ], + "type": "text", + "content": "[1] Jean-Baptiste Alayrac, Jeff Donahue, Pauline Luc, Antoine Miech, Iain Barr, Yana Hasson, Karel Lenc, Arthur Mensch, Katherine Millican, Malcolm Reynolds, et al. Flamingo: a visual language model for few-shot learning. Advances in Neural Information Processing Systems, 35:23716-23736, 2022. 2, 4" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 61, + 159, + 297, + 213 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 61, + 159, + 297, + 213 + ], + "spans": [ + { + "bbox": [ + 61, + 159, + 297, + 213 + ], + "type": "text", + "content": "[2] Max Bain, Arsha Nagrani, Gül Varol, and Andrew Zisserman. Frozen in time: A joint video and image encoder for end-to-end retrieval. In Proceedings of the IEEE/CVF international conference on computer vision, pages 1728-1738, 2021. 4, 5, 6" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 62, + 216, + 296, + 258 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 216, + 296, + 258 + ], + "spans": [ + { + "bbox": [ + 62, + 216, + 296, + 258 + ], + "type": "text", + "content": "[3] Dmitry Baranchuk, Ivan Rubachev, Andrey Voynov, Valentin Khrulkov, and Artem Babenko. Label-efficient semantic segmentation with diffusion models. arXiv preprint arXiv:2112.03126, 2021. 3" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 62, + 261, + 296, + 316 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 261, + 296, + 316 + ], + "spans": [ + { + "bbox": [ + 62, + 261, + 296, + 316 + ], + "type": "text", + "content": "[4] Andreas Blattmann, Tim Dockhorn, Sumith Kulal, Daniel Mendelevitch, Maciej Kilian, Dominik Lorenz, Yam Levi, Zion English, Vikram Voleti, Adam Letts, et al. Stable video diffusion: Scaling latent video diffusion models to large datasets. arXiv preprint arXiv:2311.15127, 2023. 7" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 62, + 318, + 296, + 384 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 318, + 296, + 384 + ], + "spans": [ + { + "bbox": [ + 62, + 318, + 296, + 384 + ], + "type": "text", + "content": "[5] Andreas Blattmann, Robin Rombach, Huan Ling, Tim Dockhorn, Seung Wook Kim, Sanja Fidler, and Karsten Kreis. Align your latents: High-resolution video synthesis with latent diffusion models. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 22563-22575, 2023. 7" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 62, + 386, + 296, + 441 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 386, + 296, + 441 + ], + "spans": [ + { + "bbox": [ + 62, + 386, + 296, + 441 + ], + "type": "text", + "content": "[6] Tom Brown, Benjamin Mann, Nick Ryder, Melanie Subbiah, Jared D Kaplan, Prafulla Dhariwal, Arvind Neelakantan, Pranav Shyam, Girish Sastry, Amanda Askell, et al. Language models are few-shot learners. Advances in neural information processing systems, 33:1877-1901, 2020. 1" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 62, + 442, + 296, + 487 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 442, + 296, + 487 + ], + "spans": [ + { + "bbox": [ + 62, + 442, + 296, + 487 + ], + "type": "text", + "content": "[7] David Chen and William B Dolan. Collecting highly parallel data for paraphrase evaluation. In Proceedings of the 49th annual meeting of the association for computational linguistics: human language technologies, pages 190-200, 2011. 6" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 62, + 488, + 296, + 542 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 488, + 296, + 542 + ], + "spans": [ + { + "bbox": [ + 62, + 488, + 296, + 542 + ], + "type": "text", + "content": "[8] Haoxin Chen, Menghan Xia, Yingqing He, Yong Zhang, Xiaodong Cun, Shaoshu Yang, Jinbo Xing, Yaofang Liu, Qifeng Chen, Xintao Wang, et al. Videocrafter1: Open diffusion models for high-quality video generation. arXiv preprint arXiv:2310.19512, 2023. 4" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 62, + 544, + 297, + 621 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 544, + 297, + 621 + ], + "spans": [ + { + "bbox": [ + 62, + 544, + 297, + 621 + ], + "type": "text", + "content": "[9] Tsai-Shien Chen, Aliaksandr Siarohin, Willi Menapace, Ekaterina Deyneka, Hsiang-wei Chao, Byung Eun Jeon, Yuwei Fang, Hsin-Ying Lee, Jian Ren, Ming-Hsuan Yang, et al. Panda-70m: Captioning 70m videos with multiple cross-modality teachers. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 13320–13331, 2024. 4, 5" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 57, + 624, + 296, + 678 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 624, + 296, + 678 + ], + "spans": [ + { + "bbox": [ + 57, + 624, + 296, + 678 + ], + "type": "text", + "content": "[10] Zesen Cheng, Sicong Leng, Hang Zhang, Yifei Xin, Xin Li, Guanzheng Chen, Yongxin Zhu, Wenqi Zhang, Ziyang Luo, Deli Zhao, et al. Videollama 2: Advancing spatial-temporal modeling and audio understanding in video-llms. arXiv preprint arXiv:2406.07476, 2024. 6" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 57, + 680, + 296, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 680, + 296, + 713 + ], + "spans": [ + { + "bbox": [ + 57, + 680, + 296, + 713 + ], + "type": "text", + "content": "[11] Aakanksha Chowdhery, Sharan Narang, Jacob Devlin, Maarten Bosma, Gaurav Mishra, Adam Roberts, Paul Barham, Hyung Won Chung, Charles Sutton, Sebastian Gehrmann," + } + ] + } + ], + "index": 11 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 316, + 73, + 555, + 713 + ], + "type": "list", + "angle": 0, + "index": 26, + "blocks": [ + { + "bbox": [ + 333, + 73, + 555, + 95 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 333, + 73, + 555, + 95 + ], + "spans": [ + { + "bbox": [ + 333, + 73, + 555, + 95 + ], + "type": "text", + "content": "et al. Palm: Scaling language modeling with pathways. arXiv preprint arXiv:2204.02311, 2022. 1" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 316, + 97, + 555, + 150 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 97, + 555, + 150 + ], + "spans": [ + { + "bbox": [ + 316, + 97, + 555, + 150 + ], + "type": "text", + "content": "[12] Runpei Dong, Chunrui Han, Yuang Peng, Zekun Qi, Zheng Ge, Jinrong Yang, Liang Zhao, Jianjian Sun, Hongyu Zhou, Haoran Wei, et al. Dreamllm: Synergistic multimodal comprehension and creation. arXiv preprint arXiv:2309.11499, 2023. 1, 2" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 316, + 152, + 555, + 217 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 152, + 555, + 217 + ], + "spans": [ + { + "bbox": [ + 316, + 152, + 555, + 217 + ], + "type": "text", + "content": "[13] Alexey Dosovitskiy, Lucas Beyer, Alexander Kolesnikov, Dirk Weissenborn, Xiaohua Zhai, Thomas Unterthiner, Mostafa Dehghani, Matthias Minderer, Georg Heigold, Sylvain Gelly, et al. An image is worth 16x16 words: Transformers for image recognition at scale. arXiv preprint arXiv:2010.11929, 2020. 2" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 316, + 220, + 555, + 285 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 220, + 555, + 285 + ], + "spans": [ + { + "bbox": [ + 316, + 220, + 555, + 285 + ], + "type": "text", + "content": "[14] Songwei Ge, Seungjun Nah, Guilin Liu, Tyler Poon, Andrew Tao, Bryan Catanzaro, David Jacobs, Jia-Bin Huang, Ming-Yu Liu, and Yogesh Balaji. Preserve your own correlation: A noise prior for video diffusion models. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 22930-22941, 2023. 7" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 316, + 286, + 553, + 319 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 286, + 553, + 319 + ], + "spans": [ + { + "bbox": [ + 316, + 286, + 553, + 319 + ], + "type": "text", + "content": "[15] Yuying Ge, Yixiao Ge, Ziyun Zeng, Xintao Wang, and Ying Shan. Planting a seed of vision in large language model. arXiv preprint arXiv:2307.08041, 2023. 1, 2" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 316, + 321, + 553, + 354 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 321, + 553, + 354 + ], + "spans": [ + { + "bbox": [ + 316, + 321, + 553, + 354 + ], + "type": "text", + "content": "[16] Yuying Ge, Sijie Zhao, Ziyun Zeng, Yixiao Ge, Chen Li, Xintao Wang, and Ying Shan. Making llama see and draw with seed tokenizer. arXiv preprint arXiv:2310.01218, 2023." + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 317, + 355, + 555, + 408 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 355, + 555, + 408 + ], + "spans": [ + { + "bbox": [ + 317, + 355, + 555, + 408 + ], + "type": "text", + "content": "[17] Yuying Ge, Sijie Zhao, Jinguo Zhu, Yixiao Ge, Kun Yi, Lin Song, Chen Li, Xiaohan Ding, and Ying Shan. Seed-x: Multimodal models with unified multi-granularity comprehension and generation. arXiv preprint arXiv:2404.14396, 2024. 1, 2, 4" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 316, + 411, + 555, + 487 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 411, + 555, + 487 + ], + "spans": [ + { + "bbox": [ + 316, + 411, + 555, + 487 + ], + "type": "text", + "content": "[18] Raghav Goyal, Samira Ebrahimi Kahou, Vincent Michalski, Joanna Materzynska, Susanne Westphal, Heuna Kim, Valentin Haenel, Ingo Fruend, Peter Yianilos, Moritz Mueller-Freitag, et al. The\" something something\" video database for learning and evaluating visual common sense. In Proceedings of the IEEE international conference on computer vision, pages 5842-5850, 2017. 5" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 316, + 489, + 555, + 555 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 489, + 555, + 555 + ], + "spans": [ + { + "bbox": [ + 316, + 489, + 555, + 555 + ], + "type": "text", + "content": "[19] Kristen Grauman, Andrew Westbury, Eugene Byrne, Zachary Chavis, Antonino Furnari, Rohit Girdhar, Jackson Hamburger, Hao Jiang, Miao Liu, Xingyu Liu, et al. Ego4d: Around the world in 3,000 hours of egocentric video. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 18995-19012, 2022. 5" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 316, + 556, + 555, + 589 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 556, + 555, + 589 + ], + "spans": [ + { + "bbox": [ + 316, + 556, + 555, + 589 + ], + "type": "text", + "content": "[20] Jonathan Ho, Ajay Jain, and Pieter Abbeel. Denoising diffusion probabilistic models. Advances in neural information processing systems, 33:6840-6851, 2020. 3" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 316, + 590, + 555, + 633 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 590, + 555, + 633 + ], + "spans": [ + { + "bbox": [ + 316, + 590, + 555, + 633 + ], + "type": "text", + "content": "[21] Wenyi Hong, Ming Ding, Wendi Zheng, Xinghan Liu, and Jie Tang. Cogvideo: Large-scale pretraining for text-to-video generation via transformers. arXiv preprint arXiv:2205.15868, 2022. 7" + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 316, + 635, + 555, + 678 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 635, + 555, + 678 + ], + "spans": [ + { + "bbox": [ + 316, + 635, + 555, + 678 + ], + "type": "text", + "content": "[22] Edward J Hu, Yelong Shen, Phillip Wallis, Zeyuan Allen-Zhu, Yanzhi Li, Shean Wang, Lu Wang, and Weizhu Chen. Lora: Low-rank adaptation of large language models. arXiv preprint arXiv:2106.09685, 2021. 6" + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 316, + 680, + 555, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 680, + 555, + 713 + ], + "spans": [ + { + "bbox": [ + 316, + 680, + 555, + 713 + ], + "type": "text", + "content": "[23] Drew A Hudson, Daniel Zoran, Mateusz Malinowski, Andrew K Lampinen, Andrew Jaegle, James L McClelland, Loic Matthew, Felix Hill, and Alexander Lerchner. Soda: Bottle-" + } + ] + } + ], + "index": 25 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 749, + 318, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 749, + 318, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 749, + 318, + 757 + ], + "type": "text", + "content": "13614" + } + ] + } + ], + "index": 27 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 8 + }, + { + "para_blocks": [ + { + "bbox": [ + 56, + 72, + 296, + 713 + ], + "type": "list", + "angle": 0, + "index": 13, + "blocks": [ + { + "bbox": [ + 76, + 72, + 296, + 106 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 76, + 72, + 296, + 106 + ], + "spans": [ + { + "bbox": [ + 76, + 72, + 296, + 106 + ], + "type": "text", + "content": "neck diffusion models for representation learning. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 23115-23127, 2024. 3" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 56, + 107, + 296, + 162 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 107, + 296, + 162 + ], + "spans": [ + { + "bbox": [ + 56, + 107, + 296, + 162 + ], + "type": "text", + "content": "[24] Albert Q Jiang, Alexandre Sablayrolles, Arthur Mensch, Chris Bamford, Devendra Singh Chaplot, Diego de las Casas, Florian Bressand, Gianna Lengyel, Guillaume Lample, Lucile Saulnier, et al. Mistral 7b. arXiv preprint arXiv:2310.06825, 2023. 2, 5" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 56, + 163, + 296, + 217 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 163, + 296, + 217 + ], + "spans": [ + { + "bbox": [ + 56, + 163, + 296, + 217 + ], + "type": "text", + "content": "[25] Yang Jin, Kun Xu, Liwei Chen, Chao Liao, Jianchao Tan, Bin Chen, Chenyi Lei, An Liu, Chengru Song, Xiaogiang Lei, et al. Unified language-vision pretraining with dynamic discrete visual tokenization. arXiv preprint arXiv:2309.04669, 2023. 2" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 56, + 220, + 296, + 274 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 220, + 296, + 274 + ], + "spans": [ + { + "bbox": [ + 56, + 220, + 296, + 274 + ], + "type": "text", + "content": "[26] Yang Jin, Zhicheng Sun, Kun Xu, Liwei Chen, Hao Jiang, Quzhe Huang, Chengru Song, Yuliang Liu, Di Zhang, Yang Song, et al. Video-lavit: Unified video-language pre-training with decoupled visual-motional tokenization. arXiv preprint arXiv:2402.03161, 2024. 1, 2, 3, 5, 6, 7" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 56, + 276, + 296, + 330 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 276, + 296, + 330 + ], + "spans": [ + { + "bbox": [ + 56, + 276, + 296, + 330 + ], + "type": "text", + "content": "[27] Will Kay, Joao Carreira, Karen Simonyan, Brian Zhang, Chloe Hillier, Sudheendra Vijayanarasimhan, Fabio Viola, Tim Green, Trevor Back, Paul Natsev, et al. The kinetics human action video dataset. arXiv preprint arXiv:1705.06950, 2017. 5" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 56, + 332, + 296, + 363 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 332, + 296, + 363 + ], + "spans": [ + { + "bbox": [ + 56, + 332, + 296, + 363 + ], + "type": "text", + "content": "[28] Diederik P Kingma and Max Welling. Auto-encoding variational bayes. arXiv preprint arXiv:1312.6114, 2013. 2, 4" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 56, + 365, + 296, + 420 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 365, + 296, + 420 + ], + "spans": [ + { + "bbox": [ + 56, + 365, + 296, + 420 + ], + "type": "text", + "content": "[29] Dan Kondratyuk, Lijun Yu, Xiuye Gu, José Lezama, Jonathan Huang, Grant Schindler, Rachel Hornung, Vighnesh Birodkar, Jimmy Yan, Ming-Chang Chiu, et al. Videopoet: A large language model for zero-shot video generation. arXiv preprint arXiv:2312.14125, 2023. 2, 7" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 56, + 422, + 296, + 466 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 422, + 296, + 466 + ], + "spans": [ + { + "bbox": [ + 56, + 422, + 296, + 466 + ], + "type": "text", + "content": "[30] Bo Li, Yuanhan Zhang, Dong Guo, Renrui Zhang, Feng Li, Hao Zhang, Kaichen Zhang, Yanwei Li, Ziwei Liu, and Chunyuan Li. Llava-onevision: Easy visual task transfer. arXiv preprint arXiv:2408.03326, 2024. 6" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 56, + 468, + 296, + 511 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 468, + 296, + 511 + ], + "spans": [ + { + "bbox": [ + 56, + 468, + 296, + 511 + ], + "type": "text", + "content": "[31] Kunchang Li, Yali Wang, Yinan He, Yizhuo Li, Yi Wang, Yi Liu, Zun Wang, Jilan Xu, Guo Chen, Ping Luo, et al. Mybench: A comprehensive multi-modal video understanding benchmark. arXiv preprint arXiv:2311.17005, 2023. 6" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 56, + 513, + 296, + 577 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 513, + 296, + 577 + ], + "spans": [ + { + "bbox": [ + 56, + 513, + 296, + 577 + ], + "type": "text", + "content": "[32] Kunchang Li, Yali Wang, Yinan He, Yizhuo Li, Yi Wang, Yi Liu, Zun Wang, Jilan Xu, Guo Chen, Ping Luo, et al. Mybench: A comprehensive multi-modal video understanding benchmark. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 22195-22206, 2024. 6" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 56, + 579, + 296, + 622 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 579, + 296, + 622 + ], + "spans": [ + { + "bbox": [ + 56, + 579, + 296, + 622 + ], + "type": "text", + "content": "[33] Tianhong Li, Yonglong Tian, He Li, Mingyang Deng, and Kaiming He. Autoregressive image generation without vector quantization. arXiv preprint arXiv:2406.11838, 2024. 2, 4, 5, 8" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 56, + 624, + 296, + 679 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 624, + 296, + 679 + ], + "spans": [ + { + "bbox": [ + 56, + 624, + 296, + 679 + ], + "type": "text", + "content": "[34] Yuncheng Li, Yale Song, Liangliang Cao, Joel Tetreault, Larry Goldberg, Alejandro Jaimes, and Jiebo Luo. Tgif: A new dataset and benchmark on animated gif description. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition, pages 4641-4650, 2016. 5" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 56, + 681, + 296, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 681, + 296, + 713 + ], + "spans": [ + { + "bbox": [ + 56, + 681, + 296, + 713 + ], + "type": "text", + "content": "[35] Yanwei Li, Chengyao Wang, and Jiaya Jia. Llama-vid: An image is worth 2 tokens in large language models. arXiv preprint arXiv:2311.17043, 2023. 6" + } + ] + } + ], + "index": 12 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 316, + 73, + 555, + 712 + ], + "type": "list", + "angle": 0, + "index": 28, + "blocks": [ + { + "bbox": [ + 316, + 73, + 555, + 127 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 73, + 555, + 127 + ], + "spans": [ + { + "bbox": [ + 316, + 73, + 555, + 127 + ], + "type": "text", + "content": "[36] Yanwei Li, Yuechen Zhang, Chengyao Wang, Zhisheng Zhong, Yixin Chen, Ruihang Chu, Shaoteng Liu, and Jiaya Jia. Mini-gemini: Mining the potential of multi-modality vision language models. arXiv preprint arXiv:2403.18814, 2024. 2" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 316, + 129, + 555, + 172 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 129, + 555, + 172 + ], + "spans": [ + { + "bbox": [ + 316, + 129, + 555, + 172 + ], + "type": "text", + "content": "[37] Bin Lin, Bin Zhu, Yang Ye, Munan Ning, Peng Jin, and Li Yuan. Video-llava: Learning united visual representation by alignment before projection. arXiv preprint arXiv:2311.10122, 2023. 5, 6" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 316, + 175, + 555, + 207 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 175, + 555, + 207 + ], + "spans": [ + { + "bbox": [ + 316, + 175, + 555, + 207 + ], + "type": "text", + "content": "[38] Haotian Liu, Chunyuan Li, Yuheng Li, Bo Li, Yuanhan Zhang, Sheng Shen, and Yong Jae Lee. Llava-last: Improved reasoning,OCR,and world knowledge,2024.6" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 316, + 209, + 555, + 241 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 209, + 555, + 241 + ], + "spans": [ + { + "bbox": [ + 316, + 209, + 555, + 241 + ], + "type": "text", + "content": "[39] Haotian Liu, Chunyuan Li, Qingyang Wu, and Yong Jae Lee. Visual instruction tuning. Advances in neural information processing systems, 36, 2024. 5" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 316, + 243, + 555, + 285 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 243, + 555, + 285 + ], + "spans": [ + { + "bbox": [ + 316, + 243, + 555, + 285 + ], + "type": "text", + "content": "[40] Hao Liu, Wilson Yan, Matei Zaharia, and Pieter Abbeel. World model on million-length video and language with ringattention. arXiv preprint arXiv:2402.08268, 2024. 1, 2, 6" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 316, + 287, + 555, + 342 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 287, + 555, + 342 + ], + "spans": [ + { + "bbox": [ + 316, + 287, + 555, + 342 + ], + "type": "text", + "content": "[41] Jiasen Lu, Christopher Clark, Sangho Lee, Zichen Zhang, Savya Khosla, Ryan Marten, Derek Hoiem, and Aniruddha Kembhavi. Unified-io 2: Scaling autoregressive multimodal models with vision, language, audio, and action. arXiv preprint arXiv:2312.17172, 2023. 2" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 316, + 344, + 555, + 398 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 344, + 555, + 398 + ], + "spans": [ + { + "bbox": [ + 316, + 344, + 555, + 398 + ], + "type": "text", + "content": "[42] Ruipu Luo, Ziwang Zhao, Min Yang, Junwei Dong, Da Li, Pengcheng Lu, Tao Wang, Linmei Hu, Minghui Qiu, and Zhongyu Wei. Valley: Video assistant with large language model enhanced ability. arXiv preprint arXiv:2306.07207, 2023.5,7" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 316, + 399, + 555, + 443 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 399, + 555, + 443 + ], + "spans": [ + { + "bbox": [ + 316, + 399, + 555, + 443 + ], + "type": "text", + "content": "[43] Muhammad Maaz, Hanoona Rasheed, Salman Khan, and Fahad Shahbaz Khan. Video-chatgpt: Towards detailed video understanding via large vision and language models. arXiv preprint arXiv:2306.05424, 2023. 5, 6, 7" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 316, + 445, + 555, + 488 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 445, + 555, + 488 + ], + "spans": [ + { + "bbox": [ + 316, + 445, + 555, + 488 + ], + "type": "text", + "content": "[44] Karttikeya Mangalam, Raiymbek Akshulakov, and Jitendra Malik. Egoschema: A diagnostic benchmark for very long-form video language understanding. Advances in Neural Information Processing Systems, 36:46212-46244, 2023. 6" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 316, + 490, + 555, + 555 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 490, + 555, + 555 + ], + "spans": [ + { + "bbox": [ + 316, + 490, + 555, + 555 + ], + "type": "text", + "content": "[45] Willi Menapace, Aliaksandr Siarohin, Ivan Skorokhodov, Ekaterina Deyneka, Tsai-Shien Chen, Anil Kag, Yuwei Fang, Aleksei Stoliar, Elisa Ricci, Jian Ren, et al. Snap video: Scaled spatiotemporal transformers for text-to-video synthesis. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 7038-7048, 2024. 7" + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 316, + 557, + 498, + 567 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 557, + 498, + 567 + ], + "spans": [ + { + "bbox": [ + 316, + 557, + 498, + 567 + ], + "type": "text", + "content": "[46] OpenAI. Gpt-4v(ison) system card, 2023. 6" + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 316, + 569, + 474, + 579 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 569, + 474, + 579 + ], + "spans": [ + { + "bbox": [ + 316, + 569, + 474, + 579 + ], + "type": "text", + "content": "[47] OpenAI. Gpt-4o system card, 2024. 6" + } + ] + } + ], + "index": 25 + }, + { + "bbox": [ + 316, + 581, + 555, + 645 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 581, + 555, + 645 + ], + "spans": [ + { + "bbox": [ + 316, + 581, + 555, + 645 + ], + "type": "text", + "content": "[48] Viorica Patraucean, Lucas Smaira, Ankush Gupta, Adria Recasens, Larisa Markeeva, Dylan Banarse, Skanda Koppula, Mateusz Malinowski, Yi Yang, Carl Doersch, et al. Perception test: A diagnostic benchmark for multimodal video models. Advances in Neural Information Processing Systems, 36, 2024. 5, 6" + } + ] + } + ], + "index": 26 + }, + { + "bbox": [ + 316, + 647, + 555, + 712 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 647, + 555, + 712 + ], + "spans": [ + { + "bbox": [ + 316, + 647, + 555, + 712 + ], + "type": "text", + "content": "[49] Alec Radford, Jong Wook Kim, Chris Hallacy, Aditya Ramesh, Gabriel Goh, Sandhini Agarwal, Girish Sastry, Amanda Askell, Pamela Mishkin, Jack Clark, et al. Learning transferable visual models from natural language supervision. In International conference on machine learning, pages 8748-8763. PMLR, 2021. 5" + } + ] + } + ], + "index": 27 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "text", + "content": "13615" + } + ] + } + ], + "index": 29 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 9 + }, + { + "para_blocks": [ + { + "bbox": [ + 56, + 72, + 297, + 713 + ], + "type": "list", + "angle": 0, + "index": 12, + "blocks": [ + { + "bbox": [ + 56, + 72, + 297, + 127 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 72, + 297, + 127 + ], + "spans": [ + { + "bbox": [ + 56, + 72, + 297, + 127 + ], + "type": "text", + "content": "[50] Robin Rombach, Andreas Blattmann, Dominik Lorenz, Patrick Esser, and Björn Ommer. High-resolution image synthesis with latent diffusion models. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 10684-10695, 2022. 2, 3" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 56, + 129, + 296, + 195 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 129, + 296, + 195 + ], + "spans": [ + { + "bbox": [ + 56, + 129, + 296, + 195 + ], + "type": "text", + "content": "[51] Christoph Schuhmann, Romain Beaumont, Richard Vencu, Cade Gordon, Ross Wightman, Mehdi Cherti, Theo Coombes, Aarush Katta, Clayton Mullis, Mitchell Wortsman, et al. Laion-5b: An open large-scale dataset for training next generation image-text models. Advances in Neural Information Processing Systems, 35:25278–25294, 2022. 5" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 56, + 197, + 296, + 262 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 197, + 296, + 262 + ], + "spans": [ + { + "bbox": [ + 56, + 197, + 296, + 262 + ], + "type": "text", + "content": "[52] Piyush Sharma, Nan Ding, Sebastian Goodman, and Radu Soricut. Conceptual captions: A cleaned, hypernymed, image alt-text dataset for automatic image captioning. In Proceedings of the 56th Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers), pages 2556-2565, 2018. 5" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 56, + 263, + 296, + 317 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 263, + 296, + 317 + ], + "spans": [ + { + "bbox": [ + 56, + 263, + 296, + 317 + ], + "type": "text", + "content": "[53] Uriel Singer, Adam Polyak, Thomas Hayes, Xi Yin, Jie An, Songyang Zhang, Qiyuan Hu, Harry Yang, Oron Ashual, Oran Gafni, et al. Make-a-video: Text-to-video generation without text-video data. arXiv preprint arXiv:2209.14792, 2022. 7" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 56, + 319, + 296, + 374 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 319, + 296, + 374 + ], + "spans": [ + { + "bbox": [ + 56, + 319, + 296, + 374 + ], + "type": "text", + "content": "[54] Quan Sun, Yufeng Cui, Xiaosong Zhang, Fan Zhang, Qiying Yu, Zhengxiong Luo, Yueze Wang, Yongming Rao, Jingjing Liu, Tiejun Huang, et al. Generative multimodal models are in-context learners. arXiv preprint arXiv:2312.13286, 2023. 2, 4" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 56, + 376, + 296, + 420 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 376, + 296, + 420 + ], + "spans": [ + { + "bbox": [ + 56, + 376, + 296, + 420 + ], + "type": "text", + "content": "[55] Quan Sun, Qiying Yu, Yufeng Cui, Fan Zhang, Xiaosong Zhang, Yueze Wang, Hongcheng Gao, Jingjing Liu, Tiejun Huang, and Xinlong Wang. Generative pretraining in multimodality. arXiv preprint arXiv:2307.05222, 2023. 1, 2" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 56, + 422, + 296, + 465 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 422, + 296, + 465 + ], + "spans": [ + { + "bbox": [ + 56, + 422, + 296, + 465 + ], + "type": "text", + "content": "[56] Quan Sun, Qiying Yu, Yufeng Cui, Fan Zhang, Xiaosong Zhang, Yueze Wang, Hongcheng Gao, Jingjing Liu, Tiejun Huang, and Xinlong Wang. Generative pretraining in multimodality. arXiv preprint arXiv:2307.05222, 2023." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 56, + 467, + 296, + 498 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 467, + 296, + 498 + ], + "spans": [ + { + "bbox": [ + 56, + 467, + 296, + 498 + ], + "type": "text", + "content": "[57] Chameleon Team. Chameleon: Mixed-modal early-fusion foundation models. arXiv preprint arXiv:2405.09818, 2024. 1, 2" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 56, + 501, + 296, + 555 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 501, + 296, + 555 + ], + "spans": [ + { + "bbox": [ + 56, + 501, + 296, + 555 + ], + "type": "text", + "content": "[58] Gemini Team, Rohan Anil, Sebastian Borgeaud, Yonghui Wu, Jean-Baptiste Alayrac, Jiahui Yu, Radu Soricut, Johan Schalkwyk, Andrew M Dai, Anja Hauth, et al. Gemini: a family of highly capable multimodal models. arXiv preprint arXiv:2312.11805, 2023. 6" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 56, + 557, + 296, + 612 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 557, + 296, + 612 + ], + "spans": [ + { + "bbox": [ + 56, + 557, + 296, + 612 + ], + "type": "text", + "content": "[59] Gemini Team, Petko Georgiev, Ving Ian Lei, Ryan Burnell, Libin Bai, Anmol Gulati, Garrett Tanzer, Damien Vincent, Zhufeng Pan, Shibo Wang, et al. Gemini 1.5: Unlocking multimodal understanding across millions of tokens of context. arXiv preprint arXiv:2403.05530, 2024. 6" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 56, + 613, + 296, + 668 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 613, + 296, + 668 + ], + "spans": [ + { + "bbox": [ + 56, + 613, + 296, + 668 + ], + "type": "text", + "content": "[60] Hugo Touvron, Thibaut Lavril, Gautier Izacard, Xavier Martinet, Marie-Anne Lachaux, Timothée Lacroix, Baptiste Rozière, Naman Goyal, Eric Hambro, Faisal Azhar, et al. Llama: Open and efficient foundation language models. arXiv preprint arXiv:2302.13971, 2023. 1" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 56, + 670, + 296, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 670, + 296, + 713 + ], + "spans": [ + { + "bbox": [ + 56, + 670, + 296, + 713 + ], + "type": "text", + "content": "[61] Michael Tschannen, Cian Eastwood, and Fabian Mentzer. Givt: Generative infinite-vocabulary transformers. In European Conference on Computer Vision, pages 292-309. Springer, 2025. 5, 7" + } + ] + } + ], + "index": 11 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 316, + 72, + 555, + 713 + ], + "type": "list", + "angle": 0, + "index": 27, + "blocks": [ + { + "bbox": [ + 316, + 72, + 555, + 117 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 72, + 555, + 117 + ], + "spans": [ + { + "bbox": [ + 316, + 72, + 555, + 117 + ], + "type": "text", + "content": "[62] Thomas Unterthiner, Sjoerd Van Steenkiste, Karol Kurach, Raphael Marinier, Marcin Michalski, and Sylvain Gelly. Towards accurate generative models of video: A new metric & challenges. arXiv preprint arXiv:1812.01717, 2018. 5, 6" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 316, + 119, + 555, + 150 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 119, + 555, + 150 + ], + "spans": [ + { + "bbox": [ + 316, + 119, + 555, + 150 + ], + "type": "text", + "content": "[63] Aaron Van Den Oord, Oriol Vinyals, et al. Neural discrete representation learning. Advances in neural information processing systems, 30, 2017. 2" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 316, + 152, + 555, + 195 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 152, + 555, + 195 + ], + "spans": [ + { + "bbox": [ + 316, + 152, + 555, + 195 + ], + "type": "text", + "content": "[64] Hanyu Wang, Saksham Suri, Yixuan Ren, Hao Chen, and Abhinav Shrivastava. Larp: Tokenizing videos with a learned autoregressive generative prior. arXiv preprint arXiv:2410.21264, 2024. 7" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 316, + 197, + 555, + 230 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 197, + 555, + 230 + ], + "spans": [ + { + "bbox": [ + 316, + 197, + 555, + 230 + ], + "type": "text", + "content": "[65] Wenxuan Wang, Quan Sun, Fan Zhang, Yepeng Tang, Jing Liu, and Xinlong Wang. Diffusion feedback helps clip see better. arXiv preprint arXiv:2407.20171, 2024. 3" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 316, + 232, + 555, + 285 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 232, + 555, + 285 + ], + "spans": [ + { + "bbox": [ + 316, + 232, + 555, + 285 + ], + "type": "text", + "content": "[66] Xiang Wang, Hangjie Yuan, Shiwei Zhang, Dayou Chen, Jiuniu Wang, Yingya Zhang, Yujun Shen, Deli Zhao, and Jingren Zhou. Videocomposer: Compositional video synthesis with motion controllability. Advances in Neural Information Processing Systems, 36, 2024. 7" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 316, + 287, + 555, + 331 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 287, + 555, + 331 + ], + "spans": [ + { + "bbox": [ + 316, + 287, + 555, + 331 + ], + "type": "text", + "content": "[67] Xinlong Wang, Xiaosong Zhang, Zhengxiong Luo, Quan Sun, Yufeng Cui, Jinsheng Wang, Fan Zhang, Yueze Wang, Zhen Li, Qiying Yu, et al. Emu3: Next-token prediction is all you need. arXiv preprint arXiv:2409.18869, 2024. 2, 3" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 316, + 333, + 555, + 386 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 333, + 555, + 386 + ], + "spans": [ + { + "bbox": [ + 316, + 333, + 555, + 386 + ], + "type": "text", + "content": "[68] Yi Wang, Yinan He, Yizhuo Li, Kunchang Li, Jiashuo Yu, Xin Ma, Xinhao Li, Guo Chen, Xinyuan Chen, Yaohui Wang, et al. Internvid: A large-scale video-text dataset for multimodal understanding and generation. arXiv preprint arXiv:2307.06942, 2023. 7" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 316, + 388, + 555, + 432 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 388, + 555, + 432 + ], + "spans": [ + { + "bbox": [ + 316, + 388, + 555, + 432 + ], + "type": "text", + "content": "[69] Yuqing Wang, Tianwei Xiong, Daquan Zhou, Zhijie Lin, Yang Zhao, Bingyi Kang, Jiashi Feng, and Xihui Liu. Loong: Generating minute-level long videos with autoregressive language models. arXiv preprint arXiv:2410.02757, 2024. 2, 6, 7" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 316, + 434, + 555, + 498 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 434, + 555, + 498 + ], + "spans": [ + { + "bbox": [ + 316, + 434, + 555, + 498 + ], + "type": "text", + "content": "[70] Chen Wei, Karttikeya Mangalam, Po-Yao Huang, Yanghao Li, Haoqi Fan, Hu Xu, Huiyu Wang, Cihang Xie, Alan Yuille, and Christoph Feichtenhofer. Diffusion models as masked autoencoders. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 16284-16294, 2023. 3" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 316, + 501, + 555, + 544 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 501, + 555, + 544 + ], + "spans": [ + { + "bbox": [ + 316, + 501, + 555, + 544 + ], + "type": "text", + "content": "[71] Chenfei Wu, Lun Huang, Qianxi Zhang, Binyang Li, Lei Ji, Fan Yang, Guillermo Sapiro, and Nan Duan. Godiva: Generating open-domain videos from natural descriptions. arXiv preprint arXiv:2104.14806, 2021. 6" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 316, + 545, + 555, + 600 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 545, + 555, + 600 + ], + "spans": [ + { + "bbox": [ + 316, + 545, + 555, + 600 + ], + "type": "text", + "content": "[72] Chengyue Wu, Xiaokang Chen, Zhiyu Wu, Yiyang Ma, Xingchao Liu, Zizheng Pan, Wen Liu, Zhenda Xie, Xingkai Yu, Chong Ruan, et al. Janus: Decoupling visual encoding for unified multimodal understanding and generation. arXiv preprint arXiv:2410.13848, 2024. 2" + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 316, + 601, + 555, + 633 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 601, + 555, + 633 + ], + "spans": [ + { + "bbox": [ + 316, + 601, + 555, + 633 + ], + "type": "text", + "content": "[73] Shengqiong Wu, Hao Fei, Leigang Qu, Wei Ji, and Tat-Seng Chua. Next-gpt: Any-to-any multimodal llm. arXiv preprint arXiv:2309.05519, 2023." + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 316, + 635, + 555, + 689 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 635, + 555, + 689 + ], + "spans": [ + { + "bbox": [ + 316, + 635, + 555, + 689 + ], + "type": "text", + "content": "[74] Yecheng Wu, Zhuoyang Zhang, Junyu Chen, Haotian Tang, Dacheng Li, Yunhao Fang, Ligeng Zhu, Enze Xie, Hongxu Yin, Li Yi, et al. Vila-u: a unified foundation model integrating visual understanding and generation. arXiv preprint arXiv:2409.04429, 2024. 1, 2, 6, 7" + } + ] + } + ], + "index": 25 + }, + { + "bbox": [ + 316, + 692, + 555, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 692, + 555, + 713 + ], + "spans": [ + { + "bbox": [ + 316, + 692, + 555, + 713 + ], + "type": "text", + "content": "[75] Weilai Xiang, Hongyu Yang, Di Huang, and Yunhong Wang. Denoising diffusion autoencoders are unified self-supervised" + } + ] + } + ], + "index": 26 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 749, + 318, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 749, + 318, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 749, + 318, + 757 + ], + "type": "text", + "content": "13616" + } + ] + } + ], + "index": 28 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 10 + }, + { + "para_blocks": [ + { + "bbox": [ + 56, + 72, + 296, + 713 + ], + "type": "list", + "angle": 0, + "index": 14, + "blocks": [ + { + "bbox": [ + 75, + 72, + 296, + 95 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 75, + 72, + 296, + 95 + ], + "spans": [ + { + "bbox": [ + 75, + 72, + 296, + 95 + ], + "type": "text", + "content": "learners. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 15802-15812, 2023. 3" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 56, + 95, + 296, + 148 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 95, + 296, + 148 + ], + "spans": [ + { + "bbox": [ + 56, + 95, + 296, + 148 + ], + "type": "text", + "content": "[76] Junbin Xiao, Xindi Shang, Angela Yao, and Tat-Seng Chua. Next-qa: Next phase of question-answering to explaining temporal actions. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 9777–9786, 2021. 5" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 56, + 150, + 296, + 205 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 150, + 296, + 205 + ], + "spans": [ + { + "bbox": [ + 56, + 150, + 296, + 205 + ], + "type": "text", + "content": "[77] Jinheng Xie, Weijia Mao, Zechen Bai, David Junhao Zhang, Weihao Wang, Kevin Qinghong Lin, Yuchao Gu, Zhijie Chen, Zhenheng Yang, and Mike Zheng Shou. Show-o: One single transformer to unify multimodal understanding and generation. arXiv preprint arXiv:2408.12528, 2024. 1, 2, 3" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 56, + 206, + 296, + 270 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 206, + 296, + 270 + ], + "spans": [ + { + "bbox": [ + 56, + 206, + 296, + 270 + ], + "type": "text", + "content": "[78] Jinbo Xing, Menghan Xia, Yong Zhang, Haoxin Chen, Wangbo Yu, Hanyuan Liu, Gongye Liu, Xintao Wang, Ying Shan, and Tien-Tsin Wong. Dynamiccafter: Animating open-domain images with video diffusion priors. In European Conference on Computer Vision, pages 399-417. Springer, 2025. 2, 4" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 56, + 271, + 295, + 316 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 271, + 295, + 316 + ], + "spans": [ + { + "bbox": [ + 56, + 271, + 295, + 316 + ], + "type": "text", + "content": "[79] Jun Xu, Tao Mei, Ting Yao, and Yong Rui. Msr-vtt: A large video description dataset for bridging video and language. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 5288-5296, 2016. 5, 6" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 56, + 316, + 296, + 370 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 316, + 296, + 370 + ], + "spans": [ + { + "bbox": [ + 56, + 316, + 296, + 370 + ], + "type": "text", + "content": "[80] Jiarui Xu, Sifei Liu, Arash Vahdat, Wonmin Byeon, Xiaolong Wang, and Shalini De Mello. Open-vocabulary panoptic segmentation with text-to-image diffusion models. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 2955-2966, 2023. 3" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 56, + 371, + 296, + 414 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 371, + 296, + 414 + ], + "spans": [ + { + "bbox": [ + 56, + 371, + 296, + 414 + ], + "type": "text", + "content": "[81] Lin Xu, Yilin Zhao, Daquan Zhou, Zhijie Lin, See Kiong Ng, and Jiashi Feng. Pllava: Parameter-free llava extension from images to videos for video dense captioning. arXiv preprint arXiv:2404.16994, 2024. 6" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 56, + 415, + 296, + 448 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 415, + 296, + 448 + ], + "spans": [ + { + "bbox": [ + 56, + 415, + 296, + 448 + ], + "type": "text", + "content": "[82] Wilson Yan, Yunzhi Zhang, Pieter Abbeel, and Aravind Srinivas. Videogpt: Video generation using vq-vae and transformers. arXiv preprint arXiv:2104.10157, 2021. 5" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 56, + 449, + 296, + 492 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 449, + 296, + 492 + ], + "spans": [ + { + "bbox": [ + 56, + 449, + 296, + 492 + ], + "type": "text", + "content": "[83] Jian Yang, Dacheng Yin, Yizhou Zhou, Fengyun Rao, Wei Zhai, Yang Cao, and Zheng-Jun Zha. Mmar: Towards lossless multi-modal auto-regressive probabilistic modeling. arXiv preprint arXiv:2410.10798, 2024. 2" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 56, + 493, + 296, + 536 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 493, + 296, + 536 + ], + "spans": [ + { + "bbox": [ + 56, + 493, + 296, + 536 + ], + "type": "text", + "content": "[84] Shuai Yang, Yuying Ge, Yang Li, Yukang Chen, Yixiao Ge, Ying Shan, and Yingcong Chen. Seed-story: Multimodal long story generation with large language model. arXiv preprint arXiv:2407.08683, 2024. 2" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 56, + 537, + 296, + 581 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 537, + 296, + 581 + ], + "spans": [ + { + "bbox": [ + 56, + 537, + 296, + 581 + ], + "type": "text", + "content": "[85] Kexin Yi, Chuang Gan, Yunzhu Li, Pushmeet Kohli, Jiajun Wu, Antonio Torralba, and Joshua B Tenenbaum. Clevrer: Collision events for video representation and reasoning. arXiv preprint arXiv:1910.01442, 2019. 5" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 56, + 582, + 296, + 635 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 582, + 296, + 635 + ], + "spans": [ + { + "bbox": [ + 56, + 582, + 296, + 635 + ], + "type": "text", + "content": "[86] Lili Yu, Bowen Shi, Ramakanth Pasunuru, Benjamin Muller, Olga Golovneva, Tianlu Wang, Arun Babu, Binh Tang, Brian Karrer, Shelly Sheynin, et al. Scaling autoregressive multimodal models: Pretraining and instruction tuning. arXiv preprint arXiv:2309.02591, 2023. 1, 2" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 56, + 636, + 296, + 690 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 636, + 296, + 690 + ], + "spans": [ + { + "bbox": [ + 56, + 636, + 296, + 690 + ], + "type": "text", + "content": "[87] Qiying Yu, Quan Sun, Xiaosong Zhang, Yufeng Cui, Fan Zhang, Yue Cao, Xinlong Wang, and Jingjing Liu. Capsfusion: Rethinking image-text data at scale. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 14022-14032, 2024. 5" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 56, + 691, + 296, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 691, + 296, + 713 + ], + "spans": [ + { + "bbox": [ + 56, + 691, + 296, + 713 + ], + "type": "text", + "content": "[88] Zhou Yu, Dejing Xu, Jun Yu, Ting Yu, Zhou Zhao, Yueting Zhuang, and Dacheng Tao. Activitynet-qa: A dataset for" + } + ] + } + ], + "index": 13 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 316, + 73, + 555, + 362 + ], + "type": "list", + "angle": 0, + "index": 21, + "blocks": [ + { + "bbox": [ + 333, + 73, + 555, + 106 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 333, + 73, + 555, + 106 + ], + "spans": [ + { + "bbox": [ + 333, + 73, + 555, + 106 + ], + "type": "text", + "content": "understanding complex web videos via question answering. In Proceedings of the AAAI Conference on Artificial Intelligence, pages 9127-9134, 2019. 5, 6" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 316, + 107, + 555, + 150 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 107, + 555, + 150 + ], + "spans": [ + { + "bbox": [ + 316, + 107, + 555, + 150 + ], + "type": "text", + "content": "[89] Chuyang Zhao, Yuxing Song, Wenhao Wang, Haocheng Feng, Errui Ding, Yifan Sun, Xinyan Xiao, and Jingdong Wang. Monoformer: One transformer for both diffusion and autoregression. arXiv preprint arXiv:2409.16280, 2024. 2" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 316, + 152, + 555, + 205 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 152, + 555, + 205 + ], + "spans": [ + { + "bbox": [ + 316, + 152, + 555, + 205 + ], + "type": "text", + "content": "[90] Wenliang Zhao, Yongming Rao, Zuyan Liu, Benlin Liu, Jie Zhou, and Jiwen Lu. Unleashing text-to-image diffusion models for visual perception. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 5729-5739, 2023. 3" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 316, + 207, + 555, + 262 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 207, + 555, + 262 + ], + "spans": [ + { + "bbox": [ + 316, + 207, + 555, + 262 + ], + "type": "text", + "content": "[91] Chunting Zhou, Lili Yu, Arun Babu, Kushal Tirumala, Michihiro Yasunaga, Leonid Shamis, Jacob Kahn, Xuezhe Ma, Luke Zettlemoyer, and Omer Levy. Transfusion: Predict the next token and diffuse images with one multi-modal model. arXiv preprint arXiv:2408.11039, 2024. 1, 2" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 316, + 263, + 555, + 307 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 263, + 555, + 307 + ], + "spans": [ + { + "bbox": [ + 316, + 263, + 555, + 307 + ], + "type": "text", + "content": "[92] Luowei Zhou, Chenliang Xu, and Jason Corso. Towards automatic learning of procedures from web instructional videos. In Proceedings of the AAAI Conference on Artificial Intelligence, 2018. 5" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 316, + 308, + 555, + 362 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 308, + 555, + 362 + ], + "spans": [ + { + "bbox": [ + 316, + 308, + 555, + 362 + ], + "type": "text", + "content": "[93] Jinguo Zhu, Xiaohan Ding, Yixiao Ge, Yuying Ge, Sijie Zhao, Hengshuang Zhao, Xiaohua Wang, and Ying Shan. Vl-gpt: A generative pre-trained transformer for vision and language understanding and generation. arXiv preprint arXiv:2312.09251, 2023. 1, 2" + } + ] + } + ], + "index": 20 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "text", + "content": "13617" + } + ] + } + ], + "index": 22 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 11 + } + ], + "_backend": "vlm", + "_version_name": "2.6.4" +} \ No newline at end of file diff --git a/2025/DnLUT_ Ultra-Efficient Color Image Denoising via Channel-Aware Lookup Tables/efb58f94-4255-404d-9c8c-4e560774a0e2_content_list.json b/2025/DnLUT_ Ultra-Efficient Color Image Denoising via Channel-Aware Lookup Tables/efb58f94-4255-404d-9c8c-4e560774a0e2_content_list.json new file mode 100644 index 0000000000000000000000000000000000000000..6e1e17c14ff93e16d349289638d24024653ce885 --- /dev/null +++ b/2025/DnLUT_ Ultra-Efficient Color Image Denoising via Channel-Aware Lookup Tables/efb58f94-4255-404d-9c8c-4e560774a0e2_content_list.json @@ -0,0 +1,1818 @@ +[ + { + "type": "text", + "text": "DnLUT: Ultra-Efficient Color Image Denoising via Channel-Aware Lookup Tables", + "text_level": 1, + "bbox": [ + 119, + 128, + 879, + 174 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Sidi Yang $^{1}$ Binxiao Huang $^{1\\dagger}$ Yulun Zhang $^{2}$ Dahai Yu $^{3}$ Yujiu Yang $^{4\\dagger}$ Ngai Wong $^{1}$", + "bbox": [ + 129, + 200, + 867, + 222 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "1 The University of Hong Kong 2 Shanghai Jiaotong University 3 TCL Corporate Research 4 Tsinghua University", + "bbox": [ + 241, + 220, + 754, + 256 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "{yangsidi99, huanghx7}@connect.hku.hk, yulun100@gmail.com, dahai.yu@tcl.com", + "bbox": [ + 246, + 258, + 750, + 273 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "yang.yujiu@sz.tsinghua.edu.cn, nwong@eee.hku.hk", + "bbox": [ + 341, + 277, + 653, + 292 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Abstract", + "text_level": 1, + "bbox": [ + 246, + 325, + 326, + 343 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "While deep neural networks have revolutionized image denoising capabilities, their deployment on edge devices remains challenging due to substantial computational and memory requirements. To this end, we present DnLUT, an ultra-efficient lookup table-based framework that achieves high-quality color image denoising with minimal resource consumption. Our key innovation lies in two complementary components: a Pairwise Channel Mixer (PCM) that effectively captures inter-channel correlations and spatial dependencies in parallel, and a novel L-shaped convolution design that maximizes receptive field coverage while minimizing storage overhead. By converting these components into optimized lookup tables post-training, DnLUT achieves remarkable efficiency - requiring only 500KB storage and $0.1\\%$ energy consumption compared to its CNN contestant DnCNN, while delivering $20\\times$ faster inference. Extensive experiments demonstrate that DnLUT outperforms all existing LUT-based methods by over 1dB in PSNR, establishing a new state-of-the-art in resource-efficient color image denoising. The project is available at https://github.com/Stephen0808/DnLUT.", + "bbox": [ + 89, + 357, + 485, + 675 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "1. Introduction", + "text_level": 1, + "bbox": [ + 89, + 702, + 223, + 718 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Color images pervade our digital world - from social media and photography to scientific imaging and medical diagnostics. These images are inherently more vulnerable to noise contamination than their grayscale counterparts, not only due to their threefold data size during acquisition, storage, and transmission, but also because human perception exhibits heightened sensitivity to color distortions compared to luminance variations.", + "bbox": [ + 89, + 727, + 482, + 847 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "The advent of deep neural networks (DNNs) has dramatically advanced the field of color image denoising[2, 15,", + "bbox": [ + 89, + 848, + 482, + 878 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/df4d233aaf8694d8dcbde46b770703422f6b3e312c780d0c901754af163409e1.jpg", + "image_caption": [ + "Figure 1. Model comparison in terms of color peak signal-to-noise ratio (CPSNR), runtime, and storage. The CPSNR and runtime are calculated on the CBSD68 dataset with Gaussian noise level $\\sigma = 25$ using Qualcomm Snapdragon 8 Gen2. Our method outperforms state-of-the-art LUT-based methods with the highest CPSNR, at a low storage requirements and reduced runtime. Additionally, our PCM module serves as a versatile plug-in module that enhances existing methods' performance by over 1dB." + ], + "image_footnote": [], + "bbox": [ + 517, + 325, + 901, + 506 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "27, 30, 32]. State-of-the-art models leverage sophisticated architectures with numerous convolutional layers or transformer blocks, achieving unprecedented denoising quality. However, these advances come at a steep cost - the computational complexity and memory requirements of such models render them impractical for edge devices. This limitation is particularly acute given that most edge devices lack specialized hardware accelerators such as graphics processing units (GPs) or tensor processing units (TPUs), creating a significant gap between algorithmic capabilities and practical deployability.", + "bbox": [ + 511, + 656, + 906, + 823 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Recent works have explored lookup tables (LUTs) as an elegant solution for efficient image enhancement. LUTs replace complex runtime computations with simple array indexing operations through direct addressing. This approach typically involves training a neural network model, then con", + "bbox": [ + 511, + 824, + 908, + 902 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "CVF", + "bbox": [ + 106, + 2, + 181, + 42 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore.", + "bbox": [ + 236, + 0, + 810, + 46 + ], + "page_idx": 0 + }, + { + "type": "page_footnote", + "text": "† Corresponding authors.", + "bbox": [ + 107, + 886, + 245, + 900 + ], + "page_idx": 0 + }, + { + "type": "page_number", + "text": "7582", + "bbox": [ + 482, + 944, + 514, + 957 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "verting it into LUTs by exhaustively mapping input-output relationships. During inference, the system performs rapid table lookups instead of costly computations, effectively leveraging DNN capabilities while circumventing hardware constraints. However, LUT-based methods face a fundamental challenge: the storage requirement grows exponentially with the input dimensionality. This has led to the adoption of 4D LUTs (using four pixel values for indexing) as a practical compromise between storage and performance.", + "bbox": [ + 89, + 90, + 480, + 226 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Current LUT-based methods predominantly employ $2 \\times 2$ convolutional kernels processed independently for each channel[10, 12, 14, 16], creating spatial-wise LUTs for single-channel processing. While this approach works well for tasks like super-resolution, it proves inadequate for color image denoising where noise typically affects all RGB channels simultaneously. Channel correlation plays a pivotal role in modeling noise characteristics, a critical factor for achieving effective restoration. Alternative approaches either use $1 \\times 1$ convolutions with three-channel depth for RGB information (creating 3D channel-wise LUTs)[13, 25] or attempt to use larger $H \\times W \\times 3$ kernels. However, these solutions either disrupt noise distribution patterns or lead to prohibitive storage requirements.", + "bbox": [ + 89, + 228, + 482, + 439 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "To address these challenges, we propose DnLUT, a novel lookup table-based framework specifically designed for efficient color image denoising. Our approach centers on two key innovations:", + "bbox": [ + 89, + 441, + 482, + 501 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "First, we introduce the Pairwise Channel Mixer (PCM), which simultaneously processes spatial and channel information. PCM strategically combines RGB channels into three pairs (RG, GB, BR), processing them through parallel branches using a $1 \\times 2$ convolution kernel with depth 2. This design efficiently captures channel correlations while maintaining manageable storage requirements. The PCM module can also serve as a plug-in module for existing LUT-based methods, consistently improving their performance by approximately 1dB.", + "bbox": [ + 89, + 503, + 482, + 654 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Second, we develop an innovative L-shaped convolution kernel that addresses the limitations of conventional rotation-based approaches. After taking the channel correlation into account, as discussed in [12, 14], there is still a need to enlarge the spatial receptive field. While previous methods like SR-LUT[10] use rotation ensemble training to expand receptive fields of spatial-wise LUT, they often introduce redundant pixel usage and significant storage overhead. Our L-shaped kernel design eliminates pixel overlap during rotations, enabling conversion to more efficient 3D LUTs while maintaining the same effective receptive field size as 4D LUTs.", + "bbox": [ + 89, + 656, + 482, + 835 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Extensive experimental validation demonstrates the superiority of DnLUT over existing approaches:", + "bbox": [ + 89, + 838, + 482, + 868 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "- We achieve up to 1.3dB CPSNR improvement over SPFLUT[14] on standard Gaussian denoising bench", + "bbox": [ + 91, + 869, + 482, + 900 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "marks. For real-world denoising tasks, we outperform CBM3D[3] by nearly 5dB.", + "bbox": [ + 527, + 90, + 903, + 119 + ], + "page_idx": 1 + }, + { + "type": "list", + "sub_type": "text", + "list_items": [ + "- Our method maintains exceptional efficiency, requiring only $0.1\\%$ of DnCNN[30]'s energy consumption and 500KB storage.", + "- The PCM module serves as a versatile performance enhancer, bringing over 1dB CPSNR improvement to existing LUT-based methods.", + "- Our L-shaped convolution design reduces storage requirements by $17 \\times$ while preserving receptive field coverage." + ], + "bbox": [ + 514, + 121, + 903, + 256 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "These results establish DnLUT as a significant advance in practical, resource-efficient color image denoising, bridging the gap between algorithmic capability and deployability on edge devices.", + "bbox": [ + 511, + 257, + 903, + 316 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "2. Related works", + "text_level": 1, + "bbox": [ + 513, + 334, + 656, + 349 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "2.1. Color Image denoising", + "text_level": 1, + "bbox": [ + 513, + 359, + 723, + 375 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "A fundamental characteristic of color images is the strong correlation among RGB channels[5]. Research has consistently shown that processing these channels independently yields suboptimal results compared to joint processing approaches[19]. This insight has driven numerous innovations in color image denoising.", + "bbox": [ + 511, + 381, + 903, + 472 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "A landmark development was CBM3D[3], which operates in luminance-chrominance space (i.e., $YC_{b}C_{r}$ ). The method performs BM3D[4] grouping once on the luminance channel $Y$ and leverages this grouping for collaborative filtering of chrominance channels $U$ and $V$ . Building on this concept, [11] introduced a radial basis function mixture for more sophisticated handling of channel correlations. Another notable approach, MC-WNNM[24], extended WNNM[7] by incorporating a weight matrix to adaptively balance the data fidelity across channels based on their noise characteristics.", + "bbox": [ + 511, + 473, + 903, + 637 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "The deep learning era, initiated by DnCNN[30], transformed the field of color image denoising. Subsequent approaches[2, 15, 26, 27, 32] have increasingly sophisticated architectures that process all RGB channels simultaneously through complex networks of convolution layers or transformer blocks. While these models achieve remarkable denoising quality, their substantial computational and memory requirements present significant deployment challenges, particularly for resource-constrained edge devices.", + "bbox": [ + 511, + 640, + 903, + 776 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "2.2. Replacing CNN with LUT", + "text_level": 1, + "bbox": [ + 513, + 787, + 750, + 803 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "The challenge of deploying deep neural networks on edge devices has sparked interest in lookup table (LUT) based alternatives. Inspired by the prevalence of color LUTs in image signal processors, [28] pioneered an approach using learnable basis LUTs combined with a weight prediction network for image enhancement. [25] further developed this", + "bbox": [ + 511, + 810, + 903, + 900 + ], + "page_idx": 1 + }, + { + "type": "page_number", + "text": "7583", + "bbox": [ + 482, + 944, + 514, + 955 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "concept, proposing a more efficient channel-wise LUT system. However, these early channel-wise LUTs, processing single pixels per channel, proved inadequate for the complex task of denoising.", + "bbox": [ + 89, + 90, + 480, + 151 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "A significant breakthrough came with [10], which introduced a method to convert CNNs with limited receptive fields into spatial-wise LUTs for super-resolution. This approach trains networks on restricted pixel neighborhoods (i.e., $2 \\times 2$ ) and converts them to LUTs by exhaustively mapping input-output relationships. During inference, the system efficiently retrieves pre-computed outputs through coordinate indexing. Follow-up works[12, 16, 17] enhanced this framework using serial LUTs to expand receptive fields, achieving substantial performance improvements.", + "bbox": [ + 89, + 152, + 480, + 303 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "To manage the exponential growth of LUT size with input dimensionality, these methods typically process each input channel independently. While this approach proves effective for super-resolution, where spatial information dominates, it falls short for color image denoising, which requires preserving both spatial and channel information to model noise distribution effectively. The limitations of existing channel-wise and spatial-wise LUT approaches highlight the need for a more sophisticated solution that can handle the unique challenges of color image denoising while maintaining computational efficiency.", + "bbox": [ + 89, + 304, + 480, + 470 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3. Method", + "text_level": 1, + "bbox": [ + 89, + 486, + 181, + 502 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3.1. Preliminary", + "text_level": 1, + "bbox": [ + 89, + 512, + 218, + 529 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "The foundation of LUT-based image restoration was established by SR-LUT[10], which introduces a constrained receptive field approach for network training (i.e., $2 \\times 2$ ). Post-training, the network's output values are systematically cached into a LUT by exhaustively traversing all possible input pixel value combinations. During inference, input patches of size $2 \\times 2$ are converted to indices of a 4D LUT to retrieve corresponding output values. To manage storage requirements, the 4D index space is uniformly subsampled. Additionally, a rotation ensemble strategy rotates each $2 \\times 2$ input patch four times, effectively expanding the receptive field to $3 \\times 3$ . Recent works have focused on increasing the effective receptive field through various techniques, including multiple LUTs[12, 14, 17], shift aggregations[17], and diverse kernel patterns[8, 12, 16].", + "bbox": [ + 89, + 536, + 482, + 762 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Our work advances this foundation by specifically addressing the challenges of LUT-based color image denoising, a critical application for edge devices.", + "bbox": [ + 89, + 763, + 482, + 808 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "A notable limitation of existing LUT-based approaches is their treatment of channel depth, typically set to 1. For color images, this means each RGB channel independently accesses cached output values without inter-channel interaction. This design overlooks crucial RGB channel correlations and spatial-channel-wise degradation patterns inher", + "bbox": [ + 89, + 810, + 482, + 900 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "ent in color image denoising. While a straightforward solution might be to employ vanilla convolution kernel patterns, the exponential growth of LUT size with index dimension makes this impractical. For instance, a kernel of size $2 \\times 2$ with depth 3 would require approximately 582 TB of storage, even with subsampling compression. This constraint presents a fundamental dilemma: when aiming to capture more spatial information, we have to sacrifice the richness of channel information. Moreover, while 4D spatial-wise LUTs show promise, they typically require multiple LUTs (analogous to stacked convolution layers), creating significant storage overhead, particularly problematic for L1 cache access.", + "bbox": [ + 511, + 90, + 903, + 286 + ], + "page_idx": 2 + }, + { + "type": "table", + "img_path": "images/8b88a346c8714544c5f7c259228305ef589a960e73081163aadb01f8142a5726.jpg", + "table_caption": [ + "Table 1. LUT size versus receptive field (RF) and channel numbers. * means subsampled LUT size" + ], + "table_footnote": [], + "table_body": "
RFDepthLUT Dim.LUT Size*
1 × 111D17 B
1 × 133D4.9 KB
2 × 214D83.5 KB
1 × 236D24.1 MB
2 × 2312D582.6 TB
k × kck × k × cD(2^4 + 1)^{k×k×C} B
", + "bbox": [ + 527, + 339, + 888, + 457 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3.2. Overview", + "text_level": 1, + "bbox": [ + 513, + 477, + 624, + 491 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "To address these fundamental challenges, we introduce two complementary innovations: 1) a pairwise channel mixer that leverages 4D LUT efficiency while modeling spatial-channel correlations; and 2) a novel L-shaped kernel featuring inherent rotation non-overlapping properties, enabling conversion to 3D LUT while maintaining 4D LUT-equivalent receptive field coverage.", + "bbox": [ + 511, + 500, + 903, + 604 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "We first build the DnNet for training, as shown in Fig. 2. Our proposed framework enhances traditional spatial-wise LUT capabilities through these two key modules: the pairwise channel mixer (PCM) establishes robust channel-spatial correlations in three RGB channels while the L-shaped kernel expands the effective receptive field with minimal computational overhead independently in each channel. The architecture combines PCM with L-shaped convolution in two stages. The first stage, inspired by [14], generates multi-channel features from LUT groups, which are then concatenated in the fusion module. The second stage integrates both proposed modules. Importantly, all components are designed for efficient conversion to 3D or 4D LUTs during inference, collectively referred to as DnLUT.", + "bbox": [ + 511, + 607, + 903, + 818 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3.3. Pairwise Channel Mixer", + "text_level": 1, + "bbox": [ + 511, + 832, + 736, + 847 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Designing effective kernel patterns for LUT-based methods presents significant challenges. Fig. 3 illustrates various kernel pattern categories. Traditional approaches have", + "bbox": [ + 511, + 854, + 903, + 900 + ], + "page_idx": 2 + }, + { + "type": "page_number", + "text": "7584", + "bbox": [ + 482, + 944, + 514, + 955 + ], + "page_idx": 2 + }, + { + "type": "image", + "img_path": "images/dd9fda6a17408baae32f2301a1fc608986e28efd782e4f04f02e5559c180fb15.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 99, + 95, + 897, + 229 + ], + "page_idx": 3 + }, + { + "type": "image", + "img_path": "images/03367edfc8631b81ef3fb86a0fad5ce2b921f4a99486e887597f506bec2c5a81.jpg", + "image_caption": [ + "Figure 2. System architecture of DnLUT: (a) The DnNet pipeline integrates pairwise channel mixers and L-shaped convolutions, with multi-scale fusion enhancing receptive field coverage. Channel dimensions are flattened for parallel processing in L-shaped operations, then unfolded for PCM input. (b) Input pixels undergo four rotations $(0^{\\circ}, 90^{\\circ}, 180^{\\circ}, 270^{\\circ})$ during processing, with outputs averaged for enhanced results. (c) Post-training, all possible input combinations are processed through DnNet modules, with outputs cached in optimized 3D or 4D LUTs. (d) During inference, input pixels are efficiently processed through multiple LUTs, with each LUT's outputs informing subsequent LUT indices, culminating in final denoised pixel values." + ], + "image_footnote": [], + "bbox": [ + 140, + 231, + 380, + 324 + ], + "page_idx": 3 + }, + { + "type": "image", + "img_path": "images/35f7a845319b0c2c2b47357b2405a7cd8c56526a616dad12d9a111a6b8816e22.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 401, + 232, + 586, + 324 + ], + "page_idx": 3 + }, + { + "type": "image", + "img_path": "images/49e0bdd5a3c7f94fb5c891de0651792ff428e90daba707affeb6a3c37541cd27.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 612, + 234, + 883, + 324 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "focused on single-dimension kernels due to LUT index dimension constraints, leading to two major limitations in color image denoising. Spatial-wise kernels, while capturing spatial information, neglect channel correlations, resulting in color distortion and suboptimal noise modeling. Conversely, channel-wise kernels establish strong RGB channel connections but ignore spatial relationships, leading to persistent artifacts. While combining these kernels in a cascade network might seem promising, this approach fails to process channel and spatial information simultaneously, resulting in suboptimal noise distribution modeling.", + "bbox": [ + 88, + 444, + 480, + 609 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Our PCM addresses these limitations through a novel architecture for color image denoising. The module first re-organizes RGB channels into three pairwise combinations: RG, GB, and BR. These pairs are processed through parallel branches using a kernel with $1 \\times 2$ spatial dimensions and depth 2 in the initial layer, followed by cascaded $1 \\times 1$ convolution layers. Each convolution operation processes four pixel values to produce one channel output, enabling efficient conversion to 4D LUTs post-training. The cached LUTs (i.e., $LUT_{RG}, LUT_{GB}, LUT_{GR}$ ) operate in parallel, with their predictions concatenated. For an input anchor $(I_R, I_G, I_B)$ , the output values $(V_R, V_G, V_B)$ are computed as:", + "bbox": [ + 89, + 611, + 482, + 806 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\begin{array}{r} \\left(V _ {R}, V _ {G}, V _ {B}\\right) = \\operatorname {C a t} \\left(L U T _ {R G} \\left[ I _ {0, 0, R} \\right] \\left[ I _ {0, 1, R} \\right] \\left[ I _ {0, 0, G} \\right] \\left[ I _ {0, 1, G} \\right], \\right. \\\\ L U T _ {G B} \\left[ I _ {0, 0, G} \\right] \\left[ I _ {0, 1, G} \\right] \\left[ I _ {0, 0, B} \\right] \\left[ I _ {0, 1, B} \\right], \\\\ L U T _ {B R} \\left[ I _ {0, 0, B} \\right] \\left[ I _ {0, 1, B} \\right] \\left[ I _ {0, 0, R} \\right] \\left[ I _ {0, 1, R} \\right]), \\end{array} \\tag {1}\n$$\n", + "text_format": "latex", + "bbox": [ + 91, + 816, + 480, + 883 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "where $H, W, C$ of each $I_{H,W,C}$ denotes height, width,", + "bbox": [ + 89, + 885, + 482, + 902 + ], + "page_idx": 3 + }, + { + "type": "image", + "img_path": "images/b6f15846ba721620cc417fdac6cdcef8905e19cfa20df9e98c6765d572e6c233.jpg", + "image_caption": [ + "(a) Spatial-wise", + "(b) Channel-wise" + ], + "image_footnote": [], + "bbox": [ + 521, + 441, + 640, + 549 + ], + "page_idx": 3 + }, + { + "type": "image", + "img_path": "images/8c14886eff627301db4bf7cd4752cc4f476dee7819944fe4857ff2c04733b838.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 645, + 477, + 763, + 515 + ], + "page_idx": 3 + }, + { + "type": "image", + "img_path": "images/42c0da747a88fbbf959f6edd2b6d129676d65f6a04f94673969dcaefb27b76b0.jpg", + "image_caption": [ + "(c) Spatial-Channel-wise", + "Figure 3. Taxonomy of kernel patterns for LUT-based methods. Dark cubes indicate rotation points, while medium-dark regions show involved pixel positions during one rotation." + ], + "image_footnote": [], + "bbox": [ + 769, + 445, + 888, + 542 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "and channel index. PCM effectively balances spatial and channel-wise processing, capturing inter-channel correlations crucial for color fidelity while preserving spatial relationships through the $1 \\times 2$ kernel design and refining features via cascaded $1 \\times 1$ convolutions.", + "bbox": [ + 511, + 635, + 905, + 710 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "3.4. Rotation Non-overlapping Kernel", + "text_level": 1, + "bbox": [ + 511, + 724, + 807, + 739 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "The rotation ensemble strategy is fundamental to LUT-based methods, using four rotations $(0^{\\circ}, 90^{\\circ}, 180^{\\circ}, 270^{\\circ})$ to enable asymmetric convolution kernels to cover larger symmetric areas. By combining outputs from all rotations, the approach enhances performance without increasing LUT index dimensionality. As shown in Fig. 4, traditional methods like SR-LUT and MuLUT use this strategy to expand receptive fields to $3 \\times 3$ and $5 \\times 5$ respectively.", + "bbox": [ + 511, + 747, + 905, + 868 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "However, this conventional approach has two critical limitations. First, approximately half of the pixels within the re", + "bbox": [ + 511, + 869, + 905, + 900 + ], + "page_idx": 3 + }, + { + "type": "page_number", + "text": "7585", + "bbox": [ + 482, + 944, + 514, + 955 + ], + "page_idx": 3 + }, + { + "type": "image", + "img_path": "images/8e5132bb3c844e09da05de9eae0fb9a91501eeb7d9b12b43eab0d053d3b6c6f2.jpg", + "image_caption": [ + "Figure 4. Comparison of spatial-wise kernel designs. Left patterns show kernel configurations, while right tables quantify lookup frequencies during output retrieval." + ], + "image_footnote": [], + "bbox": [ + 125, + 89, + 452, + 287 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "ceptive field are accessed multiple times during lookup operations, as illustrated in Fig. 4's tables. This redundancy indicates inefficient kernel pattern design. Second, stacking these kernels to create multiple LUTs for expanded receptive fields results in significant 4D LUT storage overhead.", + "bbox": [ + 89, + 354, + 483, + 430 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "To address these inefficiencies, we introduce an innovative L-shaped kernel design. During each rotation, the kernel processes two additional pixels (beyond the center pixel) without overlap, ensuring each surrounding pixel contributes exactly once to the output. This design maximizes pixel value utilization within the receptive field while enabling conversion to 3D LUTs instead of 4D LUTs, dramatically reducing storage requirements without sacrificing coverage.", + "bbox": [ + 89, + 430, + 483, + 566 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "3.5. PCM Plug-in Module", + "text_level": 1, + "bbox": [ + 89, + 575, + 292, + 590 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "To demonstrate the versatility of our pairwise channel mixer, we developed PCM as a plug-in module compatible with existing LUT-based methods. The inference of plug-in algorithm is summarized in Alg. 1. This module seamlessly integrates with current architectures, enhancing their ability to capture channel-spatial correlations without requiring structural modifications. Compared to state-of-the-art methods like SPFLUT, the PCM plug-in adds only $12\\%$ runtime and $8\\%$ storage overhead while delivering over 1dB performance improvement.", + "bbox": [ + 89, + 598, + 483, + 750 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "4. Experiments", + "text_level": 1, + "bbox": [ + 89, + 763, + 223, + 779 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "4.1. Implementation details", + "text_level": 1, + "bbox": [ + 89, + 787, + 303, + 803 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "We implement DnLUT using the Adam optimizer with a cosine annealing learning rate schedule and mean squared error (MSE) loss function. For Gaussian denoising, we train networks for $2 \\times 10^{5}$ iterations with a batch size of 12, increasing to 32 for real-world denoising scenarios. Following established practices[10, 12], we uniformly sample LUTs", + "bbox": [ + 89, + 809, + 483, + 901 + ], + "page_idx": 4 + }, + { + "type": "code", + "sub_type": "algorithm", + "code_caption": [ + "Algorithm 1 Inference of Pairwise Channel Mixer Module" + ], + "code_body": "Require: Input image $I$ of size $H\\times W\\times 3$ \nRequire: 4D Lookup Table LUT of PCM \nRequire: Padding size $p$ \n1: Initialize output image $O$ of size $H\\times W\\times 3$ \n2: for $r = 0$ to 3 do \n3: Rotate image $I$ by $r\\times 90$ degrees to get $I_{\\mathrm{rot}}$ \n4: Pad $I_{\\mathrm{rot}}$ with $p$ pixels on each side to get $I_{\\mathrm{rot}}^{\\prime}$ \n5: for $h = 0$ to $H - 1$ do \n6: for $w = 0$ to $W - 1$ do \n7: for $c = 0$ to 2 do \n8: $p_1\\gets I_{\\mathrm{rot}}'$ [h,w,c] \n9: $p_2\\gets I_{\\mathrm{rot}}'[h,w + 1,c]$ \n10: $p_3\\gets I_{\\mathrm{rot}}'[h,w,(c + 1)\\mod 3]$ \n11: $p_4\\gets I_{\\mathrm{rot}}'[h,w + 1,(c + 1)\\mod 3]$ \n12: index $\\leftarrow$ Quantize $(p_1,p_2,p_3,p_4)$ \n13: value $\\leftarrow$ LUT[index] \n14: $O[h,w,c]\\gets O[h,w,c] + \\mathrm{Interpolate}(\\mathrm{value})$ \n15: end for \n16: end for \n17: end for \n18: Rotate $O$ back by $(4 - r)\\times 90$ degrees to get $O_{\\mathrm{rot}}$ \n19: $O\\gets O + O_{\\mathrm{rot}}$ \n20: end for \n21: $O\\gets O / 4$ \n22: return Output image $O$", + "bbox": [ + 516, + 109, + 906, + 492 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "at $2^{4}$ intervals. To minimize indexing artifacts, final predictions utilize 4D simplex interpolation and 3D tetrahedral interpolation. We further refine cached LUTs through 2000 iterations of LUT-aware fine-tuning[12] on the training dataset.", + "bbox": [ + 511, + 521, + 906, + 595 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "4.2. Gaussian Color Image denoising", + "text_level": 1, + "bbox": [ + 511, + 612, + 799, + 628 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Following [27], we train our model using four comprehensive datasets: DIV2K[1], Flickr2K[23], BSD400[20], and WaterlooED[18]. Training data is augmented with additive Gaussian noise at three levels ( $\\sigma \\in \\{15, 25, 50\\}$ ). We evaluate performance on four standard benchmarks: CBSD68[21], Kodak24, Urban100[9], and McMaster[31], using identical noise levels.", + "bbox": [ + 511, + 636, + 905, + 742 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "4.3. Real-world Color Image denoising", + "text_level": 1, + "bbox": [ + 511, + 756, + 815, + 772 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "To validate real-world performance, we train DnLUT on the SIDD training dataset and evaluate on both SIDD validation and DnD datasets. The SIDD dataset comprises indoor scenes captured by five smartphones under varying noise conditions, providing 320 image pairs for training and 1,280 for validation. The DnD dataset offers 50 high-resolution images with realistic noise patterns for comprehensive evaluation.", + "bbox": [ + 511, + 779, + 906, + 898 + ], + "page_idx": 4 + }, + { + "type": "page_number", + "text": "7586", + "bbox": [ + 482, + 945, + 514, + 955 + ], + "page_idx": 4 + }, + { + "type": "table", + "img_path": "images/9ebac7c69a6c28adc21b6995b3dba844bbb30518831c76e97f97f6abb13375a3.jpg", + "table_caption": [ + "Table 2. Quantitative comparison on Gaussian color image denoising benchmark. The table presents CPSNR(dB) values." + ], + "table_footnote": [], + "table_body": "
CategoryMethodCBSD68Kodak24Urban100McMaster
σ = 15σ = 25σ = 50σ = 15σ = 25σ = 50σ = 15σ = 25σ = 50σ = 15σ = 25σ = 50
LUT-basedSR-LUT[10]29.7626.7122.4130.3527.1622.6529.3826.0421.6031.1828.0123.35
MuLUT[12]30.5228.1124.8531.3129.0225.2830.2527.6723.7532.2829.8826.36
RC-LUT[16]30.6828.1225.0431.5729.0725.8930.3327.8023.8632.5129.8926.50
SPF-LUT[14]30.9728.5625.3331.8629.5826.2630.8928.2624.2231.7730.4426.91
DnLUT(Ours)32.4129.8826.0333.0230.2426.7432.1228.8725.0132.8830.4427.12
ClassicalCBM3D[3]33.5230.7127.3834.2831.6829.0233.9331.3627.9334.0631.6628.51
MC-WNNM[24]31.9829.3226.9833.2330.8928.6730.2329.2327.0031.2330.2027.55
DNNDnCNN[30]33.9031.2427.9534.6032.1428.9632.9830.8127.5933.4531.5228.62
SwinIR[15]34.4231.7828.5635.3432.8929.7935.6133.2030.2235.1332.9029.82
", + "bbox": [ + 114, + 114, + 883, + 277 + ], + "page_idx": 5 + }, + { + "type": "table", + "img_path": "images/219c5e89bef2c959a70aa4cc9463c910a5b6a6b06199e788b449838060380ceb.jpg", + "table_caption": [ + "Table 3. Quantitative comparison on real-world color image denoising. The table presents CPSNR(dB) and SSIM values. Methods on DnD dataset are validated on the online platform. We could only get the PSNR rather than CPSNR." + ], + "table_footnote": [], + "table_body": "
DatasetsMethodSR-LUT[10]MuLUT[12]RC-LUT[16]SPF-LUT[14]DnLUT(Ours)CBM3D[3]MC-WNNM[24]DnCNN[30]
SIDDCPSNR29.3833.2433.8834.9135.4430.1429.4536.45
SSIM0.6340.8300.8390.8650.8750.7020.6820.900
DnDPSNR*33.6935.1135.4336.2236.6733.1232.3437.11
SSIM0.8390.8680.8750.9110.9220.8230.8170.932
", + "bbox": [ + 109, + 327, + 885, + 398 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/f707c479ed17ad64ec199e63ac3cd3b3afcb8096a21be9db854fb3080697d70c.jpg", + "image_caption": [ + "Full Image" + ], + "image_footnote": [], + "bbox": [ + 102, + 410, + 398, + 549 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/7ce7a3866174c3984bcbc60f356a7bd9c645755d6017573c10ebde2dc03057bf.jpg", + "image_caption": [ + "SRLUT" + ], + "image_footnote": [], + "bbox": [ + 408, + 410, + 524, + 474 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/ac8f52598c456798f84da1e6b4a8fe65be8829fb2564be892e591b72d621e0bf.jpg", + "image_caption": [ + "MuLUT" + ], + "image_footnote": [], + "bbox": [ + 531, + 410, + 648, + 474 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/afea0d92d133941c17b1c9df3b62c374a7aabbd17fd0ae9bfc88f1ba36e6fb23.jpg", + "image_caption": [ + "RC-LUT" + ], + "image_footnote": [], + "bbox": [ + 653, + 410, + 767, + 474 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/9a0c552aaa5de2bafef5f5d6f3fcb4eda5d13301a842d73143c0567ecccc7f5c.jpg", + "image_caption": [ + "SPFLUT" + ], + "image_footnote": [], + "bbox": [ + 772, + 410, + 893, + 474 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/f3bee1ef768b24efe502fe8aa97ea3f7f981a36fa5f832801240b19a5944d7b9.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 102, + 598, + 263, + 720 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/395227161c30e30370e0f83df8d8f5d193cbba1e9a7734cf9464367a3b6d5b25.jpg", + "image_caption": [ + "SRLUT" + ], + "image_footnote": [], + "bbox": [ + 264, + 598, + 421, + 720 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/95280d8cfef7da473a46e69b090e391a1f2899e83bf6fb54d5f21708b7de773f.jpg", + "image_caption": [ + "MuLUT" + ], + "image_footnote": [], + "bbox": [ + 421, + 598, + 575, + 720 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/29a41cc7b4521de450807c711d1e8a10eb8cce645468fdb2d4784c279fcb5e41.jpg", + "image_caption": [ + "RC-LUT" + ], + "image_footnote": [], + "bbox": [ + 576, + 598, + 735, + 720 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/cd165983b48473f3a8e5c50857f723ffbf4bc58e920e41be487f339981236403.jpg", + "image_caption": [ + "SPFLUT" + ], + "image_footnote": [], + "bbox": [ + 735, + 598, + 890, + 720 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/76eef8a75f28d6b1720a683a7a877bdacb938c1ec497de1509fc17f19b355c4f.jpg", + "image_caption": [ + "Input", + "GT" + ], + "image_footnote": [], + "bbox": [ + 106, + 734, + 263, + 856 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/b47e2080ee26174667d7f1bc0af8665b45e7771b438399a8cfbd77bc61323153.jpg", + "image_caption": [ + "CBM3D", + "Figure 6. Qualitative comparison on real-world datasets." + ], + "image_footnote": [], + "bbox": [ + 264, + 734, + 421, + 856 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/3cdf760cefd38d030708448fd15a6d5a672f9d0bd2dca9e784ffabc0c086474c.jpg", + "image_caption": [ + "MC-WNNM" + ], + "image_footnote": [], + "bbox": [ + 421, + 734, + 576, + 856 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/b0edf591dfa56f6bc933646ab52b1e49f9f6652b6361779fbe325d5a59676021.jpg", + "image_caption": [ + "DnLUT(Ours)" + ], + "image_footnote": [], + "bbox": [ + 576, + 734, + 735, + 856 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/db9d85892a1f8bd632f1053a07d14f75f79d56e2141e864ccfec3b7b500bc48e.jpg", + "image_caption": [ + "DnCNN" + ], + "image_footnote": [], + "bbox": [ + 735, + 734, + 890, + 856 + ], + "page_idx": 5 + }, + { + "type": "page_number", + "text": "7587", + "bbox": [ + 482, + 945, + 514, + 955 + ], + "page_idx": 5 + }, + { + "type": "table", + "img_path": "images/8f429ef9d4d32a3aa79d3d435809fb61763aae5cfe830c203ca17182114781c1.jpg", + "table_caption": [ + "Table 4. PCM can be easily incorporated to the existing LUT-based methods. It equips the model with the ability of capturing the channel correlations which brings more than 1dB on the widely used benchmark. The noise level $\\sigma$ of Gaussian noise is set to 25. The table presents CPSNR(dB) values." + ], + "table_footnote": [], + "table_body": "
MethodCBSD68Kodak24Urban100McMasterSIDD
SR-LUT[10]26.7127.1626.0428.0129.38
PCM+SR-LUT28.04(+1.33)28.55(+1.39)27.33(+1.29)28.78(+0.77)32.33(+2.96)
MuLUT[12]28.1129.0127.6729.8833.24
PCM+MuLUT29.17(+1.05)29.92(+0.91)28.74(+1.07)30.33(+0.44)34.33(+1.09)
RC-LUT[16]28.1229.0727.8029.8933.88
PCM+RC-LUT29.22(+1.10)30.09(+1.02)28.77(+0.97)30.54(+0.65)34.67(+0.79)
SPF-LUT[14]28.5629.5828.2630.4434.91
PCM+SPF-LUT29.72(+1.16)30.63(+1.05)29.43(+1.17)30.96(+0.52)35.56(+0.65)
", + "bbox": [ + 163, + 142, + 831, + 290 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "4.4. Quantitative results", + "text_level": 1, + "bbox": [ + 89, + 297, + 279, + 313 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "We evaluate performance using color peak signal-to-noise ratio(CPSNR) and structural similarity index(SSIM) metrics, comparing against state-of-the-art methods across three categories: LUT-based methods (SR-LUT[10], MuLUT[12], RC-LUT[16], SPF-LUT[14]), classical approaches (CBM3D[3], MC-WNNM[24]), and DNN-based solutions (DnCNN[30], SwinIR[15]).", + "bbox": [ + 89, + 319, + 483, + 425 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "As shown in Tab. 2, DnLUT achieves superior performance in the LUT-based category, surpassing SPF-LUT by over 1dB in Gaussian noise removal while requiring only $17\\%$ of its storage. Results from Tab. 3 demonstrate DnLUT's effectiveness in real-world scenarios, outperforming classical methods CBM3D[3] and MC-WNNM[24] by approximately 5dB. Though the Gaussian noise is independent across RGB channels, the network can still leverage inter-channel natural distribution in images to suppress the noise. For example, if a pixel in the R channel has a high value due to noise, but the corresponding pixels in the G and B channels do not, the network can infer that the high value is likely noise and suppress it.", + "bbox": [ + 89, + 426, + 483, + 622 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "4.5. Qualitative results", + "text_level": 1, + "bbox": [ + 89, + 633, + 269, + 648 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Visual comparisons in Figs. 5 & 6 illustrate DnLUT's superior denoising capabilities. Existing LUT-based methods exhibit noticeable color distortion, particularly evident in Fig. 5's first row. In contrast, DnLUT achieves visual quality comparable to computation-intensive methods like DnCNN and CBM3D. Fig. 6 further demonstrates DnLUT's advantage in real-world scenarios, where both LUT-based and classical methods struggle with color fidelity and often over-smooth details. DnLUT successfully preserves color consistency while retaining fine image details.", + "bbox": [ + 89, + 655, + 483, + 806 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "4.6. Efficiency Evaluation", + "text_level": 1, + "bbox": [ + 89, + 816, + 290, + 832 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "We conduct comprehensive efficiency analysis across three key metrics: theoretical energy cost, runtime performance, and storage requirements (Tab. 5).", + "bbox": [ + 89, + 839, + 482, + 883 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Energy cost. Following AdderSR[22], we analyze multipli", + "bbox": [ + 89, + 885, + 483, + 901 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/70d838580f39f91a819f746cb018318068c63b319adaa17c2727446b67ed393d.jpg", + "image_caption": [ + "SRLUT" + ], + "image_footnote": [], + "bbox": [ + 521, + 295, + 616, + 349 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/ed0608af79df9df6536b9813cd27dfbf9b06672f33a8aa0a15f453c12e3083a7.jpg", + "image_caption": [ + "MuLUT" + ], + "image_footnote": [], + "bbox": [ + 616, + 296, + 709, + 349 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/bc48ddfe7c6c2b5550cccd38393f187a403c0f59d33c08e1e2aef246cf16befa.jpg", + "image_caption": [ + "RC-LUT" + ], + "image_footnote": [], + "bbox": [ + 709, + 296, + 803, + 349 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/a2d9b749cd909fa61fd0ed0e1ce2ab506c4effc0835f7c584d97f03badd6c1cc.jpg", + "image_caption": [ + "SPFLUT" + ], + "image_footnote": [], + "bbox": [ + 805, + 296, + 900, + 349 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/67db694903958bfe9172aecb7c267f2a07875508d01ab1899d3f9de0ce9229c6.jpg", + "image_caption": [ + "PCM+SRLUT", + "Figure 7. Visualization of LUT-based method with plug-in PCM." + ], + "image_footnote": [], + "bbox": [ + 521, + 361, + 614, + 415 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/9b2e5b66c65bd12d5768fd1b84def76c75eaddd3c5e81881a207da1580432ec8.jpg", + "image_caption": [ + "PCM+MuLUT" + ], + "image_footnote": [], + "bbox": [ + 616, + 361, + 709, + 415 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/5b7930fff27d398d728ed25e3bb03839620738acc8f7b4604f24acc11d4d4c17.jpg", + "image_caption": [ + "PCM+RC-LUT" + ], + "image_footnote": [], + "bbox": [ + 709, + 361, + 803, + 415 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/82968afa1650114ee52071db974244db9d6b376595baf1b7ced031d047b75821.jpg", + "image_caption": [ + "PCM+SPFLUT" + ], + "image_footnote": [], + "bbox": [ + 805, + 361, + 900, + 415 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "cation and addition operations across different data types to estimate total energy consumption. DnLUT achieves $70\\%$ energy savings compared to SPF-LUT while delivering superior denoising quality. The efficiency gap widens dramatically when compared to DNN-based methods, with DnLUT requiring only $0.1\\%$ of DnCNN's energy consumption.", + "bbox": [ + 511, + 464, + 905, + 555 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Runtime. We implement LUT-based methods using standard JAVA API on the ANDROID platform. DnLUT maintains the rapid inference capabilities characteristic of SR-LUT[10], demonstrating clear advantages over classical and DNN-based approaches. For reference, DnCNN running on mobile CPU requires $20 \\times$ more processing time. DnLUT's performance can be further optimized through device-specific implementations, such as FPGA deployments.", + "bbox": [ + 511, + 556, + 905, + 691 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Storage. DnLUT requires approximately 500 KB storage, well within typical L2-cache limits of modern chips. This efficient memory footprint enables rapid lookup operations with minimal dependency overhead.", + "bbox": [ + 511, + 695, + 905, + 755 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "4.7. Effectiveness of Plug-in PCM", + "text_level": 1, + "bbox": [ + 511, + 770, + 774, + 787 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Tab. 4 demonstrates PCM's versatility as a performance enhancer for existing LUT-based methods. When integrated with current approaches, PCM consistently delivers significant improvements across all benchmark datasets. The visual comparisons in Fig. 7 show marked reduction in color distortion, particularly in smooth image regions, resulting in superior perceptual quality.", + "bbox": [ + 511, + 795, + 905, + 901 + ], + "page_idx": 6 + }, + { + "type": "page_number", + "text": "7588", + "bbox": [ + 482, + 945, + 514, + 955 + ], + "page_idx": 6 + }, + { + "type": "table", + "img_path": "images/da2f78a2af32ab21797e18c119a617a607d79e677316a392ce4162e8f6cffb70.jpg", + "table_caption": [ + "Table 5. Efficiency evaluation of three categories color image denoising methods. We use the Qualcomm Snapdragon 8 Gen 2 as the mobile platform. The energy cost is calculated on the ${512} \\times {512}$ resolution color images." + ], + "table_footnote": [], + "table_body": "
CategoryMethodPlatformRuntime(ms) 256 × 256Runtime(ms) 512 × 512Energy Cost (pJ)Storage (KB)
LUT-basedSR-LUT[10]Mobile619149.98M82
MuLUT[12]Mobile2180899.88M489
RC-LUT[6]Mobile1878612.92M326
SPF-LUT[14]Mobile722572.32G30,178
DnLUT(ours)Mobile2388687.34M518
ClassicalCBM3D[29]PC3,18914,3434.82G-
MC-WNNM[24]PC74,290293,42489.23G-
DNNDnCNN[30]Mobile5431,923542.53G2,239
SwinIR[15]Mobile73,234274,45312.03T45,499
", + "bbox": [ + 91, + 128, + 919, + 320 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "4.8. Ablation study", + "text_level": 1, + "bbox": [ + 89, + 330, + 240, + 347 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "We conduct ablation studies using Gaussian color image denoising benchmark to validate DnLUT's key components.", + "bbox": [ + 89, + 353, + 482, + 383 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "4.8.1. Channel-spatial-wise convolution", + "text_level": 1, + "bbox": [ + 89, + 392, + 367, + 406 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Channel indexing[13, 25] possesses a $1 \\times 1$ receptive field to extract channel information. We evaluate various combinations of channel-wise, spatial-wise, and our proposed channel-spatial-wise kernels using a consistent architecture comprising one channel indexing module, two L-shaped convolution modules, and one PCM component.", + "bbox": [ + 89, + 411, + 483, + 501 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Results in Tab. 6 reveal that naive combination of channel-wise and spatial-wise convolutions yields suboptimal performance. Pure channel-spatial-wise convolution, while promising, struggles with limited receptive field coverage. The integration of spatial-wise and channel-spatial-wise convolutions produces substantial improvements. Notably, adding channel-wise convolution offers minimal additional benefit, as channel-spatial-wise convolution already captures comprehensive color information.", + "bbox": [ + 89, + 502, + 483, + 638 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "4.8.2. L-shaped convolution", + "text_level": 1, + "bbox": [ + 89, + 643, + 287, + 659 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "To validate our L-shaped convolution design, we compare it against standard $2 \\times 2$ spatial-wise kernels (denoted as S in Tab. 7, $\\sigma = 1$ ). While the conventional kernel shows marginal performance advantages, our L-shaped design achieves comparable results while enabling 3D LUT conversion, reducing storage requirements by $17 \\times$ . It paves the way of deploying multi-layer spatial-wise convolutions.", + "bbox": [ + 89, + 662, + 482, + 770 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "5. Conclusion", + "text_level": 1, + "bbox": [ + 89, + 784, + 209, + 800 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "This paper presents DnLUT, a novel approach to color image denoising that bridges the gap between high-performance deep learning models and resource-constrained edge devices. Addressing the limitations of existing LUT-based methods, DnLUT introduces two key innovations. First, the Pairwise Channel Mixer (PCM) models channel and spa", + "bbox": [ + 89, + 809, + 483, + 901 + ], + "page_idx": 7 + }, + { + "type": "table", + "img_path": "images/aae33e99c75bd08e03a3d3ba9e7ea939517e99bcd2332971df92d013d9c0bcad.jpg", + "table_caption": [ + "Table 6. Different combinations of channel-wise, spatial-wise, and our channel-spatial-wise kernels. These kernels are represented by channel indexing[13], L-shaped kernel, and PCM. The table presents CPSNR(dB) values." + ], + "table_footnote": [], + "table_body": "
ChannelSpatialChannel-spatialCBSD68
✓29.77
✓30.15
✓✓30.21
✓✓31.41
✓✓✓31.47
", + "bbox": [ + 514, + 387, + 908, + 488 + ], + "page_idx": 7 + }, + { + "type": "table", + "img_path": "images/c0a1c9b1c92b6d07338bc05a56d760b6fdf980d80a8adbe254b801209319dc6d.jpg", + "table_caption": [ + "Table 7. Comparison of L-shaped kernel and $2 \\times 2$ spatial-wise kernel. The table presents CPSNR(dB) values." + ], + "table_footnote": [], + "table_body": "
MethodCBSD68KodakMcMasterLUT Size(KB)
PCM + L26.6327.0527.96494 + 24
PCM + S26.7127.1628.01494 + 408
", + "bbox": [ + 517, + 521, + 906, + 573 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "tial correlations simultaneously by processing channel pairs, enabling comprehensive color modeling without excessive storage costs. Second, an L-shaped convolution design rethinks rotation-based receptive field expansion, eliminating redundant pixel usage and converting 4D to 3D LUTs, reducing storage by $17\\times$ while maintaining spatial coverage. These advancements allow DnLUT to outperform state-ofthe-art LUT-based methods by over 1dB in denoising quality while consuming just $0.1\\%$ of DnCNN's energy.", + "bbox": [ + 511, + 590, + 906, + 728 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Acknowledgements. This research was supported by: the HKU-TCL Joint Research Centre for AI, the Theme-based Research Scheme (TRS) project T45-701/22-R, and the General Research Fund (GRF) Project 17203224 from the Research Grants Council (RGC) of Hong Kong SAR. Additional support was provided by the National Key Research and Development Program of China (Grant No. 2024YFB2808903), the Shenzhen Science and Technology Program (JCYJ20220818101001004), and the Tsinghua University - Tencent Joint Laboratory for Internet Innovation Technology Research Fund.", + "bbox": [ + 511, + 732, + 908, + 898 + ], + "page_idx": 7 + }, + { + "type": "page_number", + "text": "7589", + "bbox": [ + 482, + 944, + 514, + 955 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "References", + "text_level": 1, + "bbox": [ + 91, + 89, + 187, + 104 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[1] Eirikur Agustsson and Radu Timofte. Ntire 2017 challenge on single image super-resolution: Dataset and study. In The IEEE Conference on Computer Vision and Pattern Recognition (CVPR) Workshops, 2017. 5", + "[2] Hanting Chen, Yunhe Wang, Tianyu Guo, Chang Xu, Yiping Deng, Zhenhua Liu, Siwei Ma, Chunjing Xu, Chao Xu, and Wen Gao. Pre-trained image processing transformer. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 12299-12310, 2021. 1, 2", + "[3] Kostadin Dabov, Alessandro Foi, Vladimir Katkovnik, and Karen Egiazarian. Color image denoising via sparse 3d collaborative filtering with grouping constraint in luminance-chrominance space. In 2007 IEEE international conference on image processing, pages I-313. IEEE, 2007. 2, 6, 7", + "[4] Kostadin Dabov, Alessandro Foi, Vladimir Katkovnik, and Karen Egiazarian. Image denoising by sparse 3-d transform-domain collaborative filtering. IEEE Transactions on image processing, 16(8):2080-2095, 2007. 2", + "[5] Jingjing Dai, Oscar C Au, Lu Fang, Chao Pang, Feng Zou, and Jiali Li. Multichannel nonlocal means fusion for color image denoising. IEEE Transactions on circuits and systems for video technology, 23(11):1873-1886, 2013. 2", + "[6] Michael Gharbi, Jiawen Chen, Jonathan T Barron, Samuel W Hasinoff, and Fredo Durand. Deep bilateral learning for real-time image enhancement. ACM Transactions on Graphics (TOG), 36(4):1-12, 2017. 8", + "[7] Shuhang Gu, Lei Zhang, Wangmeng Zuo, and Xiangchu Feng. Weighted nuclear norm minimization with application to image denoising. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 2862-2869, 2014. 2", + "[8] Binxiao Huang, Jason Chun Lok Li, Jie Ran, Boyu Li, Jiajun Zhou, Dahai Yu, and Ngai Wong. Hundred-kilobyte lookup tables for efficient single-image super-resolution. arXiv preprint arXiv:2312.06101, 2023. 3", + "[9] Jia-Bin Huang, Abhishek Singh, and Narendra Ahuja. Single image super-resolution from transformed self-exemplars. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 5197-5206, 2015. 5", + "[10] Younghyun Jo and Seon Joo Kim. Practical single-image super-resolution using look-up table. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 691-700, 2021. 2, 3, 5, 6, 7, 8", + "[11] Stamatios Lefkimmiatis. Non-local color image denoising with convolutional neural networks. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 3587-3596, 2017. 2", + "[12] Jiacheng Li, Chang Chen, Zhen Cheng, and Zhiwei Xiong. Mulut: Cooperating multiple look-up tables for efficient image super-resolution. In European Conference on Computer Vision, pages 238-256. Springer, 2022. 2, 3, 5, 6, 7, 8", + "[13] Jiacheng Li, Chang Chen, Zhen Cheng, and Zhiwei Xiong. Toward dnn of luts: Learning efficient image restoration with multiple look-up tables. IEEE Transactions on Pattern Analysis and Machine Intelligence, 2024. 2, 8" + ], + "bbox": [ + 93, + 114, + 483, + 900 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[14] Yinglong Li, Jiacheng Li, and Zhiwei Xiong. Look-up table compression for efficient image restoration. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 26016-26025, 2024. 2, 3, 6, 7, 8", + "[15] Jingyun Liang, Jiezhang Cao, Guolei Sun, Kai Zhang, Luc Van Gool, and Radu Timofte. Swinir: Image restoration using swin transformer. In Proceedings of the IEEE/CVF international conference on computer vision, pages 1833-1844, 2021. 1, 2, 6, 7, 8", + "[16] Guandu Liu, Yukang Ding, Mading Li, Ming Sun, Xing Wen, and Bin Wang. Reconstructed convolution module based look-up tables for efficient image super-resolution. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 12217-12226, 2023. 2, 3, 6, 7", + "[17] Cheng Ma, Jingyi Zhang, Jie Zhou, and Jiwen Lu. Learning series-parallel lookup tables for efficient image superresolution. In European Conference on Computer Vision, pages 305–321. Springer, 2022. 3", + "[18] Kede Ma, Zhengfang Duanmu, Qingbo Wu, Zhou Wang, Hongwei Yong, Hongliang Li, and Lei Zhang. Waterloo exploration database: New challenges for image quality assessment models. IEEE Transactions on Image Processing, 26 (2):1004-1016, 2016. 5", + "[19] Julien Mairal, Michael Elad, and Guillermo Sapiro. Sparse representation for color image restoration. IEEE Transactions on image processing, 17(1):53-69, 2007. 2", + "[20] D. Martin, C. Fowlkes, D. Tal, and J. Malik. A database of human segmented natural images and its application to evaluating segmentation algorithms and measuring ecological statistics. In Proc. 8th Int'l Conf. Computer Vision, pages 416-423, 2001. 5", + "[21] David Martin, Charless Fowlkes, Doron Tal, and Jitendra Malik. A database of human segmented natural images and its application to evaluating segmentation algorithms and measuring ecological statistics. In Proceedings eighth IEEE international conference on computer vision. ICCV 2001, pages 416-423. IEEE, 2001. 5", + "[22] Dehua Song, Yunhe Wang, Hanting Chen, Chang Xu, Chunjing Xu, and DaCheng Tao. Addersr: Towards energy efficient image super-resolution. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 15648-15657, 2021. 7", + "[23] Radu Timofte, Eirikur Agustsson, Luc Van Gool, Ming-Hsuan Yang, and Lei Zhang. Ntire 2017 challenge on single image super-resolution: Methods and results. In Proceedings of the IEEE conference on computer vision and pattern recognition workshops, pages 114-125, 2017. 5", + "[24] Jun Xu, Lei Zhang, David Zhang, and Xiangchu Feng. Multichannel weighted nuclear norm minimization for real color image denoising. In Proceedings of the IEEE international conference on computer vision, pages 1096-1104, 2017. 2, 6, 7, 8", + "[25] Sidi Yang, Binxiao Huang, Mingdeng Cao, Yatai Ji, Hanzhong Guo, Ngai Wong, and Yujiu Yang. Taming lookup tables for efficient image retouching. arXiv preprint arXiv:2403.19238, 2024. 2, 8", + "[26] Syed Waqas Zamir, Aditya Arora, Salman Khan, Munawar Hayat, Fahad Shahbaz Khan, Ming-Hsuan Yang, and Ling" + ], + "bbox": [ + 93, + 92, + 903, + 900 + ], + "page_idx": 8 + }, + { + "type": "page_number", + "text": "7590", + "bbox": [ + 482, + 944, + 516, + 955 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "Shao. Multi-stage progressive image restoration. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 14821-14831, 2021. 2", + "[27] Syed Waqas Zamir, Aditya Arora, Salman Khan, Munawar Hayat, Fahad Shahbaz Khan, and Ming-Hsuan Yang. Restormer: Efficient transformer for high-resolution image restoration. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 5728-5739, 2022. 1, 2, 5", + "[28] Hui Zeng, Jianrui Cai, Lida Li, Zisheng Cao, and Lei Zhang. Learning image-adaptive 3d lookup tables for high performance photo enhancement in real-time. IEEE Transactions on Pattern Analysis and Machine Intelligence, 44(4):2058-2073, 2020. 2", + "[29] Fengyi Zhang, Hui Zeng, Tianjun Zhang, and Lin Zhang. Clut-net: Learning adaptively compressed representations of 3dluts for lightweight image enhancement. In Proceedings of the 30th ACM International Conference on Multimedia, pages 6493-6501, 2022. 8", + "[30] Kai Zhang, Wangmeng Zuo, Yunjin Chen, Deyu Meng, and Lei Zhang. Beyond a gaussian denoiser: Residual learning of deep cnn for image denoising. IEEE transactions on image processing, 26(7):3142-3155, 2017. 1, 2, 6, 7, 8", + "[31] Lei Zhang, Xiaolin Wu, Antoni Buades, and Xin Li. Color demosaicking by local directional interpolation and nonlocal adaptive thresholding. Journal of Electronic imaging, 20(2): 023016-023016, 2011. 5", + "[32] Yi Zhang, Dasong Li, Xiaoyu Shi, Dailan He, Kangning Song, Xiaogang Wang, Hongwei Qin, and Hongsheng Li. Kbnet: Kernel basis network for image restoration. arXiv preprint arXiv:2303.02881, 2023. 1, 2" + ], + "bbox": [ + 91, + 90, + 482, + 527 + ], + "page_idx": 9 + }, + { + "type": "page_number", + "text": "7591", + "bbox": [ + 482, + 945, + 514, + 955 + ], + "page_idx": 9 + } +] \ No newline at end of file diff --git a/2025/DnLUT_ Ultra-Efficient Color Image Denoising via Channel-Aware Lookup Tables/efb58f94-4255-404d-9c8c-4e560774a0e2_model.json b/2025/DnLUT_ Ultra-Efficient Color Image Denoising via Channel-Aware Lookup Tables/efb58f94-4255-404d-9c8c-4e560774a0e2_model.json new file mode 100644 index 0000000000000000000000000000000000000000..9dbd5fe548f3f0f3a41876c59f7bfaf242fdcd43 --- /dev/null +++ b/2025/DnLUT_ Ultra-Efficient Color Image Denoising via Channel-Aware Lookup Tables/efb58f94-4255-404d-9c8c-4e560774a0e2_model.json @@ -0,0 +1,2442 @@ +[ + [ + { + "type": "header", + "bbox": [ + 0.107, + 0.003, + 0.182, + 0.044 + ], + "angle": 0, + "content": "CVF" + }, + { + "type": "header", + "bbox": [ + 0.237, + 0.001, + 0.812, + 0.047 + ], + "angle": 0, + "content": "This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore." + }, + { + "type": "title", + "bbox": [ + 0.12, + 0.13, + 0.88, + 0.175 + ], + "angle": 0, + "content": "DnLUT: Ultra-Efficient Color Image Denoising via Channel-Aware Lookup Tables" + }, + { + "type": "text", + "bbox": [ + 0.13, + 0.202, + 0.868, + 0.223 + ], + "angle": 0, + "content": "Sidi Yang\\(^{1}\\) Binxiao Huang\\(^{1\\dagger}\\) Yulun Zhang\\(^{2}\\) Dahai Yu\\(^{3}\\) Yujiu Yang\\(^{4\\dagger}\\) Ngai Wong\\(^{1}\\)" + }, + { + "type": "text", + "bbox": [ + 0.242, + 0.222, + 0.756, + 0.257 + ], + "angle": 0, + "content": "1 The University of Hong Kong 2 Shanghai Jiaotong University 3 TCL Corporate Research 4 Tsinghua University" + }, + { + "type": "text", + "bbox": [ + 0.248, + 0.259, + 0.751, + 0.275 + ], + "angle": 0, + "content": "{yangsidi99, huanghx7}@connect.hku.hk, yulun100@gmail.com, dahai.yu@tcl.com" + }, + { + "type": "text", + "bbox": [ + 0.342, + 0.278, + 0.655, + 0.293 + ], + "angle": 0, + "content": "yang.yujiu@sz.tsinghua.edu.cn, nwong@eee.hku.hk" + }, + { + "type": "title", + "bbox": [ + 0.248, + 0.327, + 0.327, + 0.344 + ], + "angle": 0, + "content": "Abstract" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.358, + 0.486, + 0.676 + ], + "angle": 0, + "content": "While deep neural networks have revolutionized image denoising capabilities, their deployment on edge devices remains challenging due to substantial computational and memory requirements. To this end, we present DnLUT, an ultra-efficient lookup table-based framework that achieves high-quality color image denoising with minimal resource consumption. Our key innovation lies in two complementary components: a Pairwise Channel Mixer (PCM) that effectively captures inter-channel correlations and spatial dependencies in parallel, and a novel L-shaped convolution design that maximizes receptive field coverage while minimizing storage overhead. By converting these components into optimized lookup tables post-training, DnLUT achieves remarkable efficiency - requiring only 500KB storage and \\(0.1\\%\\) energy consumption compared to its CNN contestant DnCNN, while delivering \\(20\\times\\) faster inference. Extensive experiments demonstrate that DnLUT outperforms all existing LUT-based methods by over 1dB in PSNR, establishing a new state-of-the-art in resource-efficient color image denoising. The project is available at https://github.com/Stephen0808/DnLUT." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.703, + 0.224, + 0.719 + ], + "angle": 0, + "content": "1. Introduction" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.728, + 0.483, + 0.848 + ], + "angle": 0, + "content": "Color images pervade our digital world - from social media and photography to scientific imaging and medical diagnostics. These images are inherently more vulnerable to noise contamination than their grayscale counterparts, not only due to their threefold data size during acquisition, storage, and transmission, but also because human perception exhibits heightened sensitivity to color distortions compared to luminance variations." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.849, + 0.483, + 0.88 + ], + "angle": 0, + "content": "The advent of deep neural networks (DNNs) has dramatically advanced the field of color image denoising[2, 15," + }, + { + "type": "image", + "bbox": [ + 0.518, + 0.327, + 0.903, + 0.507 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.512, + 0.516, + 0.908, + 0.629 + ], + "angle": 0, + "content": "Figure 1. Model comparison in terms of color peak signal-to-noise ratio (CPSNR), runtime, and storage. The CPSNR and runtime are calculated on the CBSD68 dataset with Gaussian noise level \\(\\sigma = 25\\) using Qualcomm Snapdragon 8 Gen2. Our method outperforms state-of-the-art LUT-based methods with the highest CPSNR, at a low storage requirements and reduced runtime. Additionally, our PCM module serves as a versatile plug-in module that enhances existing methods' performance by over 1dB." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.657, + 0.907, + 0.824 + ], + "angle": 0, + "content": "27, 30, 32]. State-of-the-art models leverage sophisticated architectures with numerous convolutional layers or transformer blocks, achieving unprecedented denoising quality. However, these advances come at a steep cost - the computational complexity and memory requirements of such models render them impractical for edge devices. This limitation is particularly acute given that most edge devices lack specialized hardware accelerators such as graphics processing units (GPs) or tensor processing units (TPUs), creating a significant gap between algorithmic capabilities and practical deployability." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.825, + 0.909, + 0.903 + ], + "angle": 0, + "content": "Recent works have explored lookup tables (LUTs) as an elegant solution for efficient image enhancement. LUTs replace complex runtime computations with simple array indexing operations through direct addressing. This approach typically involves training a neural network model, then con" + }, + { + "type": "page_footnote", + "bbox": [ + 0.109, + 0.887, + 0.246, + 0.901 + ], + "angle": 0, + "content": "† Corresponding authors." + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.516, + 0.958 + ], + "angle": 0, + "content": "7582" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.09, + 0.092, + 0.482, + 0.227 + ], + "angle": 0, + "content": "verting it into LUTs by exhaustively mapping input-output relationships. During inference, the system performs rapid table lookups instead of costly computations, effectively leveraging DNN capabilities while circumventing hardware constraints. However, LUT-based methods face a fundamental challenge: the storage requirement grows exponentially with the input dimensionality. This has led to the adoption of 4D LUTs (using four pixel values for indexing) as a practical compromise between storage and performance." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.229, + 0.483, + 0.44 + ], + "angle": 0, + "content": "Current LUT-based methods predominantly employ \\(2 \\times 2\\) convolutional kernels processed independently for each channel[10, 12, 14, 16], creating spatial-wise LUTs for single-channel processing. While this approach works well for tasks like super-resolution, it proves inadequate for color image denoising where noise typically affects all RGB channels simultaneously. Channel correlation plays a pivotal role in modeling noise characteristics, a critical factor for achieving effective restoration. Alternative approaches either use \\(1 \\times 1\\) convolutions with three-channel depth for RGB information (creating 3D channel-wise LUTs)[13, 25] or attempt to use larger \\(H \\times W \\times 3\\) kernels. However, these solutions either disrupt noise distribution patterns or lead to prohibitive storage requirements." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.442, + 0.483, + 0.502 + ], + "angle": 0, + "content": "To address these challenges, we propose DnLUT, a novel lookup table-based framework specifically designed for efficient color image denoising. Our approach centers on two key innovations:" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.504, + 0.483, + 0.655 + ], + "angle": 0, + "content": "First, we introduce the Pairwise Channel Mixer (PCM), which simultaneously processes spatial and channel information. PCM strategically combines RGB channels into three pairs (RG, GB, BR), processing them through parallel branches using a \\(1 \\times 2\\) convolution kernel with depth 2. This design efficiently captures channel correlations while maintaining manageable storage requirements. The PCM module can also serve as a plug-in module for existing LUT-based methods, consistently improving their performance by approximately 1dB." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.657, + 0.483, + 0.837 + ], + "angle": 0, + "content": "Second, we develop an innovative L-shaped convolution kernel that addresses the limitations of conventional rotation-based approaches. After taking the channel correlation into account, as discussed in [12, 14], there is still a need to enlarge the spatial receptive field. While previous methods like SR-LUT[10] use rotation ensemble training to expand receptive fields of spatial-wise LUT, they often introduce redundant pixel usage and significant storage overhead. Our L-shaped kernel design eliminates pixel overlap during rotations, enabling conversion to more efficient 3D LUTs while maintaining the same effective receptive field size as 4D LUTs." + }, + { + "type": "text", + "bbox": [ + 0.091, + 0.839, + 0.483, + 0.869 + ], + "angle": 0, + "content": "Extensive experimental validation demonstrates the superiority of DnLUT over existing approaches:" + }, + { + "type": "text", + "bbox": [ + 0.092, + 0.871, + 0.483, + 0.901 + ], + "angle": 0, + "content": "- We achieve up to 1.3dB CPSNR improvement over SPFLUT[14] on standard Gaussian denoising bench" + }, + { + "type": "text", + "bbox": [ + 0.529, + 0.092, + 0.905, + 0.121 + ], + "angle": 0, + "content": "marks. For real-world denoising tasks, we outperform CBM3D[3] by nearly 5dB." + }, + { + "type": "text", + "bbox": [ + 0.515, + 0.122, + 0.905, + 0.167 + ], + "angle": 0, + "content": "- Our method maintains exceptional efficiency, requiring only \\(0.1\\%\\) of DnCNN[30]'s energy consumption and 500KB storage." + }, + { + "type": "text", + "bbox": [ + 0.515, + 0.168, + 0.905, + 0.211 + ], + "angle": 0, + "content": "- The PCM module serves as a versatile performance enhancer, bringing over 1dB CPSNR improvement to existing LUT-based methods." + }, + { + "type": "text", + "bbox": [ + 0.515, + 0.213, + 0.905, + 0.257 + ], + "angle": 0, + "content": "- Our L-shaped convolution design reduces storage requirements by \\(17 \\times\\) while preserving receptive field coverage." + }, + { + "type": "list", + "bbox": [ + 0.515, + 0.122, + 0.905, + 0.257 + ], + "angle": 0, + "content": null + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.258, + 0.905, + 0.318 + ], + "angle": 0, + "content": "These results establish DnLUT as a significant advance in practical, resource-efficient color image denoising, bridging the gap between algorithmic capability and deployability on edge devices." + }, + { + "type": "title", + "bbox": [ + 0.514, + 0.335, + 0.658, + 0.35 + ], + "angle": 0, + "content": "2. Related works" + }, + { + "type": "title", + "bbox": [ + 0.514, + 0.36, + 0.724, + 0.376 + ], + "angle": 0, + "content": "2.1. Color Image denoising" + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.382, + 0.905, + 0.473 + ], + "angle": 0, + "content": "A fundamental characteristic of color images is the strong correlation among RGB channels[5]. Research has consistently shown that processing these channels independently yields suboptimal results compared to joint processing approaches[19]. This insight has driven numerous innovations in color image denoising." + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.474, + 0.905, + 0.638 + ], + "angle": 0, + "content": "A landmark development was CBM3D[3], which operates in luminance-chrominance space (i.e., \\( YC_{b}C_{r} \\)). The method performs BM3D[4] grouping once on the luminance channel \\( Y \\) and leverages this grouping for collaborative filtering of chrominance channels \\( U \\) and \\( V \\). Building on this concept, [11] introduced a radial basis function mixture for more sophisticated handling of channel correlations. Another notable approach, MC-WNNM[24], extended WNNM[7] by incorporating a weight matrix to adaptively balance the data fidelity across channels based on their noise characteristics." + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.641, + 0.905, + 0.777 + ], + "angle": 0, + "content": "The deep learning era, initiated by DnCNN[30], transformed the field of color image denoising. Subsequent approaches[2, 15, 26, 27, 32] have increasingly sophisticated architectures that process all RGB channels simultaneously through complex networks of convolution layers or transformer blocks. While these models achieve remarkable denoising quality, their substantial computational and memory requirements present significant deployment challenges, particularly for resource-constrained edge devices." + }, + { + "type": "title", + "bbox": [ + 0.514, + 0.788, + 0.75, + 0.804 + ], + "angle": 0, + "content": "2.2. Replacing CNN with LUT" + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.811, + 0.905, + 0.901 + ], + "angle": 0, + "content": "The challenge of deploying deep neural networks on edge devices has sparked interest in lookup table (LUT) based alternatives. Inspired by the prevalence of color LUTs in image signal processors, [28] pioneered an approach using learnable basis LUTs combined with a weight prediction network for image enhancement. [25] further developed this" + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.516, + 0.957 + ], + "angle": 0, + "content": "7583" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.09, + 0.092, + 0.482, + 0.152 + ], + "angle": 0, + "content": "concept, proposing a more efficient channel-wise LUT system. However, these early channel-wise LUTs, processing single pixels per channel, proved inadequate for the complex task of denoising." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.153, + 0.482, + 0.304 + ], + "angle": 0, + "content": "A significant breakthrough came with [10], which introduced a method to convert CNNs with limited receptive fields into spatial-wise LUTs for super-resolution. This approach trains networks on restricted pixel neighborhoods (i.e., \\(2 \\times 2\\)) and converts them to LUTs by exhaustively mapping input-output relationships. During inference, the system efficiently retrieves pre-computed outputs through coordinate indexing. Follow-up works[12, 16, 17] enhanced this framework using serial LUTs to expand receptive fields, achieving substantial performance improvements." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.305, + 0.482, + 0.471 + ], + "angle": 0, + "content": "To manage the exponential growth of LUT size with input dimensionality, these methods typically process each input channel independently. While this approach proves effective for super-resolution, where spatial information dominates, it falls short for color image denoising, which requires preserving both spatial and channel information to model noise distribution effectively. The limitations of existing channel-wise and spatial-wise LUT approaches highlight the need for a more sophisticated solution that can handle the unique challenges of color image denoising while maintaining computational efficiency." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.487, + 0.182, + 0.503 + ], + "angle": 0, + "content": "3. Method" + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.513, + 0.22, + 0.53 + ], + "angle": 0, + "content": "3.1. Preliminary" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.537, + 0.483, + 0.763 + ], + "angle": 0, + "content": "The foundation of LUT-based image restoration was established by SR-LUT[10], which introduces a constrained receptive field approach for network training (i.e., \\(2 \\times 2\\)). Post-training, the network's output values are systematically cached into a LUT by exhaustively traversing all possible input pixel value combinations. During inference, input patches of size \\(2 \\times 2\\) are converted to indices of a 4D LUT to retrieve corresponding output values. To manage storage requirements, the 4D index space is uniformly subsampled. Additionally, a rotation ensemble strategy rotates each \\(2 \\times 2\\) input patch four times, effectively expanding the receptive field to \\(3 \\times 3\\). Recent works have focused on increasing the effective receptive field through various techniques, including multiple LUTs[12, 14, 17], shift aggregations[17], and diverse kernel patterns[8, 12, 16]." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.764, + 0.483, + 0.809 + ], + "angle": 0, + "content": "Our work advances this foundation by specifically addressing the challenges of LUT-based color image denoising, a critical application for edge devices." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.811, + 0.483, + 0.901 + ], + "angle": 0, + "content": "A notable limitation of existing LUT-based approaches is their treatment of channel depth, typically set to 1. For color images, this means each RGB channel independently accesses cached output values without inter-channel interaction. This design overlooks crucial RGB channel correlations and spatial-channel-wise degradation patterns inher" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.092, + 0.905, + 0.287 + ], + "angle": 0, + "content": "ent in color image denoising. While a straightforward solution might be to employ vanilla convolution kernel patterns, the exponential growth of LUT size with index dimension makes this impractical. For instance, a kernel of size \\(2 \\times 2\\) with depth 3 would require approximately 582 TB of storage, even with subsampling compression. This constraint presents a fundamental dilemma: when aiming to capture more spatial information, we have to sacrifice the richness of channel information. Moreover, while 4D spatial-wise LUTs show promise, they typically require multiple LUTs (analogous to stacked convolution layers), creating significant storage overhead, particularly problematic for L1 cache access." + }, + { + "type": "table_caption", + "bbox": [ + 0.513, + 0.301, + 0.905, + 0.328 + ], + "angle": 0, + "content": "Table 1. LUT size versus receptive field (RF) and channel numbers. * means subsampled LUT size" + }, + { + "type": "table", + "bbox": [ + 0.528, + 0.34, + 0.89, + 0.458 + ], + "angle": 0, + "content": "
RFDepthLUT Dim.LUT Size*
1 × 111D17 B
1 × 133D4.9 KB
2 × 214D83.5 KB
1 × 236D24.1 MB
2 × 2312D582.6 TB
k × kck × k × cD(2^4 + 1)^{k×k×C} B
" + }, + { + "type": "title", + "bbox": [ + 0.514, + 0.478, + 0.625, + 0.492 + ], + "angle": 0, + "content": "3.2. Overview" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.501, + 0.905, + 0.606 + ], + "angle": 0, + "content": "To address these fundamental challenges, we introduce two complementary innovations: 1) a pairwise channel mixer that leverages 4D LUT efficiency while modeling spatial-channel correlations; and 2) a novel L-shaped kernel featuring inherent rotation non-overlapping properties, enabling conversion to 3D LUT while maintaining 4D LUT-equivalent receptive field coverage." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.608, + 0.905, + 0.819 + ], + "angle": 0, + "content": "We first build the DnNet for training, as shown in Fig. 2. Our proposed framework enhances traditional spatial-wise LUT capabilities through these two key modules: the pairwise channel mixer (PCM) establishes robust channel-spatial correlations in three RGB channels while the L-shaped kernel expands the effective receptive field with minimal computational overhead independently in each channel. The architecture combines PCM with L-shaped convolution in two stages. The first stage, inspired by [14], generates multi-channel features from LUT groups, which are then concatenated in the fusion module. The second stage integrates both proposed modules. Importantly, all components are designed for efficient conversion to 3D or 4D LUTs during inference, collectively referred to as DnLUT." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.833, + 0.738, + 0.848 + ], + "angle": 0, + "content": "3.3. Pairwise Channel Mixer" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.856, + 0.905, + 0.901 + ], + "angle": 0, + "content": "Designing effective kernel patterns for LUT-based methods presents significant challenges. Fig. 3 illustrates various kernel pattern categories. Traditional approaches have" + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.516, + 0.957 + ], + "angle": 0, + "content": "7584" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.1, + 0.096, + 0.898, + 0.23 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.141, + 0.232, + 0.382, + 0.325 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.403, + 0.233, + 0.588, + 0.325 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.614, + 0.236, + 0.885, + 0.325 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.089, + 0.335, + 0.908, + 0.419 + ], + "angle": 0, + "content": "Figure 2. System architecture of DnLUT: (a) The DnNet pipeline integrates pairwise channel mixers and L-shaped convolutions, with multi-scale fusion enhancing receptive field coverage. Channel dimensions are flattened for parallel processing in L-shaped operations, then unfolded for PCM input. (b) Input pixels undergo four rotations \\((0^{\\circ}, 90^{\\circ}, 180^{\\circ}, 270^{\\circ})\\) during processing, with outputs averaged for enhanced results. (c) Post-training, all possible input combinations are processed through DnNet modules, with outputs cached in optimized 3D or 4D LUTs. (d) During inference, input pixels are efficiently processed through multiple LUTs, with each LUT's outputs informing subsequent LUT indices, culminating in final denoised pixel values." + }, + { + "type": "text", + "bbox": [ + 0.089, + 0.445, + 0.482, + 0.611 + ], + "angle": 0, + "content": "focused on single-dimension kernels due to LUT index dimension constraints, leading to two major limitations in color image denoising. Spatial-wise kernels, while capturing spatial information, neglect channel correlations, resulting in color distortion and suboptimal noise modeling. Conversely, channel-wise kernels establish strong RGB channel connections but ignore spatial relationships, leading to persistent artifacts. While combining these kernels in a cascade network might seem promising, this approach fails to process channel and spatial information simultaneously, resulting in suboptimal noise distribution modeling." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.612, + 0.483, + 0.807 + ], + "angle": 0, + "content": "Our PCM addresses these limitations through a novel architecture for color image denoising. The module first re-organizes RGB channels into three pairwise combinations: RG, GB, and BR. These pairs are processed through parallel branches using a kernel with \\(1 \\times 2\\) spatial dimensions and depth 2 in the initial layer, followed by cascaded \\(1 \\times 1\\) convolution layers. Each convolution operation processes four pixel values to produce one channel output, enabling efficient conversion to 4D LUTs post-training. The cached LUTs (i.e., \\(LUT_{RG}, LUT_{GB}, LUT_{GR}\\)) operate in parallel, with their predictions concatenated. For an input anchor \\((I_R, I_G, I_B)\\), the output values \\((V_R, V_G, V_B)\\) are computed as:" + }, + { + "type": "equation", + "bbox": [ + 0.093, + 0.818, + 0.482, + 0.885 + ], + "angle": 0, + "content": "\\[\n\\begin{array}{r} \\left(V _ {R}, V _ {G}, V _ {B}\\right) = \\operatorname {C a t} \\left(L U T _ {R G} \\left[ I _ {0, 0, R} \\right] \\left[ I _ {0, 1, R} \\right] \\left[ I _ {0, 0, G} \\right] \\left[ I _ {0, 1, G} \\right], \\right. \\\\ L U T _ {G B} \\left[ I _ {0, 0, G} \\right] \\left[ I _ {0, 1, G} \\right] \\left[ I _ {0, 0, B} \\right] \\left[ I _ {0, 1, B} \\right], \\\\ L U T _ {B R} \\left[ I _ {0, 0, B} \\right] \\left[ I _ {0, 1, B} \\right] \\left[ I _ {0, 0, R} \\right] \\left[ I _ {0, 1, R} \\right]), \\end{array} \\tag {1}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.091, + 0.886, + 0.483, + 0.904 + ], + "angle": 0, + "content": "where \\(H, W, C\\) of each \\(I_{H,W,C}\\) denotes height, width," + }, + { + "type": "image", + "bbox": [ + 0.522, + 0.443, + 0.641, + 0.55 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.536, + 0.551, + 0.617, + 0.561 + ], + "angle": 0, + "content": "(a) Spatial-wise" + }, + { + "type": "image", + "bbox": [ + 0.647, + 0.478, + 0.764, + 0.516 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.653, + 0.55, + 0.742, + 0.561 + ], + "angle": 0, + "content": "(b) Channel-wise" + }, + { + "type": "image", + "bbox": [ + 0.77, + 0.446, + 0.89, + 0.544 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.769, + 0.55, + 0.895, + 0.562 + ], + "angle": 0, + "content": "(c) Spatial-Channel-wise" + }, + { + "type": "image_caption", + "bbox": [ + 0.513, + 0.574, + 0.906, + 0.617 + ], + "angle": 0, + "content": "Figure 3. Taxonomy of kernel patterns for LUT-based methods. Dark cubes indicate rotation points, while medium-dark regions show involved pixel positions during one rotation." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.636, + 0.906, + 0.711 + ], + "angle": 0, + "content": "and channel index. PCM effectively balances spatial and channel-wise processing, capturing inter-channel correlations crucial for color fidelity while preserving spatial relationships through the \\(1 \\times 2\\) kernel design and refining features via cascaded \\(1 \\times 1\\) convolutions." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.725, + 0.808, + 0.741 + ], + "angle": 0, + "content": "3.4. Rotation Non-overlapping Kernel" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.748, + 0.906, + 0.869 + ], + "angle": 0, + "content": "The rotation ensemble strategy is fundamental to LUT-based methods, using four rotations \\((0^{\\circ}, 90^{\\circ}, 180^{\\circ}, 270^{\\circ})\\) to enable asymmetric convolution kernels to cover larger symmetric areas. By combining outputs from all rotations, the approach enhances performance without increasing LUT index dimensionality. As shown in Fig. 4, traditional methods like SR-LUT and MuLUT use this strategy to expand receptive fields to \\(3 \\times 3\\) and \\(5 \\times 5\\) respectively." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.871, + 0.906, + 0.901 + ], + "angle": 0, + "content": "However, this conventional approach has two critical limitations. First, approximately half of the pixels within the re" + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.516, + 0.957 + ], + "angle": 0, + "content": "7585" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.127, + 0.09, + 0.454, + 0.288 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.09, + 0.299, + 0.484, + 0.342 + ], + "angle": 0, + "content": "Figure 4. Comparison of spatial-wise kernel designs. Left patterns show kernel configurations, while right tables quantify lookup frequencies during output retrieval." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.355, + 0.484, + 0.431 + ], + "angle": 0, + "content": "ceptive field are accessed multiple times during lookup operations, as illustrated in Fig. 4's tables. This redundancy indicates inefficient kernel pattern design. Second, stacking these kernels to create multiple LUTs for expanded receptive fields results in significant 4D LUT storage overhead." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.431, + 0.484, + 0.567 + ], + "angle": 0, + "content": "To address these inefficiencies, we introduce an innovative L-shaped kernel design. During each rotation, the kernel processes two additional pixels (beyond the center pixel) without overlap, ensuring each surrounding pixel contributes exactly once to the output. This design maximizes pixel value utilization within the receptive field while enabling conversion to 3D LUTs instead of 4D LUTs, dramatically reducing storage requirements without sacrificing coverage." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.577, + 0.294, + 0.592 + ], + "angle": 0, + "content": "3.5. PCM Plug-in Module" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.599, + 0.484, + 0.75 + ], + "angle": 0, + "content": "To demonstrate the versatility of our pairwise channel mixer, we developed PCM as a plug-in module compatible with existing LUT-based methods. The inference of plug-in algorithm is summarized in Alg. 1. This module seamlessly integrates with current architectures, enhancing their ability to capture channel-spatial correlations without requiring structural modifications. Compared to state-of-the-art methods like SPFLUT, the PCM plug-in adds only \\(12\\%\\) runtime and \\(8\\%\\) storage overhead while delivering over 1dB performance improvement." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.764, + 0.224, + 0.78 + ], + "angle": 0, + "content": "4. Experiments" + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.789, + 0.305, + 0.804 + ], + "angle": 0, + "content": "4.1. Implementation details" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.81, + 0.484, + 0.902 + ], + "angle": 0, + "content": "We implement DnLUT using the Adam optimizer with a cosine annealing learning rate schedule and mean squared error (MSE) loss function. For Gaussian denoising, we train networks for \\(2 \\times 10^{5}\\) iterations with a batch size of 12, increasing to 32 for real-world denoising scenarios. Following established practices[10, 12], we uniformly sample LUTs" + }, + { + "type": "code_caption", + "bbox": [ + 0.515, + 0.091, + 0.905, + 0.107 + ], + "angle": 0, + "content": "Algorithm 1 Inference of Pairwise Channel Mixer Module" + }, + { + "type": "algorithm", + "bbox": [ + 0.517, + 0.11, + 0.907, + 0.493 + ], + "angle": 0, + "content": "Require: Input image \\(I\\) of size \\(H\\times W\\times 3\\) \nRequire: 4D Lookup Table LUT of PCM \nRequire: Padding size \\(p\\) \n1: Initialize output image \\(O\\) of size \\(H\\times W\\times 3\\) \n2: for \\(r = 0\\) to 3 do \n3: Rotate image \\(I\\) by \\(r\\times 90\\) degrees to get \\(I_{\\mathrm{rot}}\\) \n4: Pad \\(I_{\\mathrm{rot}}\\) with \\(p\\) pixels on each side to get \\(I_{\\mathrm{rot}}^{\\prime}\\) \n5: for \\(h = 0\\) to \\(H - 1\\) do \n6: for \\(w = 0\\) to \\(W - 1\\) do \n7: for \\(c = 0\\) to 2 do \n8: \\(p_1\\gets I_{\\mathrm{rot}}'\\) [h,w,c] \n9: \\(p_2\\gets I_{\\mathrm{rot}}'[h,w + 1,c]\\) \n10: \\(p_3\\gets I_{\\mathrm{rot}}'[h,w,(c + 1)\\mod 3]\\) \n11: \\(p_4\\gets I_{\\mathrm{rot}}'[h,w + 1,(c + 1)\\mod 3]\\) \n12: index \\(\\leftarrow\\) Quantize \\((p_1,p_2,p_3,p_4)\\) \n13: value \\(\\leftarrow\\) LUT[index] \n14: \\(O[h,w,c]\\gets O[h,w,c] + \\mathrm{Interpolate}(\\mathrm{value})\\) \n15: end for \n16: end for \n17: end for \n18: Rotate \\(O\\) back by \\((4 - r)\\times 90\\) degrees to get \\(O_{\\mathrm{rot}}\\) \n19: \\(O\\gets O + O_{\\mathrm{rot}}\\) \n20: end for \n21: \\(O\\gets O / 4\\) \n22: return Output image \\(O\\)" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.522, + 0.907, + 0.597 + ], + "angle": 0, + "content": "at \\(2^{4}\\) intervals. To minimize indexing artifacts, final predictions utilize 4D simplex interpolation and 3D tetrahedral interpolation. We further refine cached LUTs through 2000 iterations of LUT-aware fine-tuning[12] on the training dataset." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.613, + 0.8, + 0.63 + ], + "angle": 0, + "content": "4.2. Gaussian Color Image denoising" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.637, + 0.906, + 0.743 + ], + "angle": 0, + "content": "Following [27], we train our model using four comprehensive datasets: DIV2K[1], Flickr2K[23], BSD400[20], and WaterlooED[18]. Training data is augmented with additive Gaussian noise at three levels (\\(\\sigma \\in \\{15, 25, 50\\}\\)). We evaluate performance on four standard benchmarks: CBSD68[21], Kodak24, Urban100[9], and McMaster[31], using identical noise levels." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.757, + 0.816, + 0.773 + ], + "angle": 0, + "content": "4.3. Real-world Color Image denoising" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.78, + 0.907, + 0.9 + ], + "angle": 0, + "content": "To validate real-world performance, we train DnLUT on the SIDD training dataset and evaluate on both SIDD validation and DnD datasets. The SIDD dataset comprises indoor scenes captured by five smartphones under varying noise conditions, providing 320 image pairs for training and 1,280 for validation. The DnD dataset offers 50 high-resolution images with realistic noise patterns for comprehensive evaluation." + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.946, + 0.516, + 0.957 + ], + "angle": 0, + "content": "7586" + } + ], + [ + { + "type": "table_caption", + "bbox": [ + 0.14, + 0.09, + 0.856, + 0.105 + ], + "angle": 0, + "content": "Table 2. Quantitative comparison on Gaussian color image denoising benchmark. The table presents CPSNR(dB) values." + }, + { + "type": "table", + "bbox": [ + 0.115, + 0.116, + 0.885, + 0.278 + ], + "angle": 0, + "content": "
CategoryMethodCBSD68Kodak24Urban100McMaster
σ = 15σ = 25σ = 50σ = 15σ = 25σ = 50σ = 15σ = 25σ = 50σ = 15σ = 25σ = 50
LUT-basedSR-LUT[10]29.7626.7122.4130.3527.1622.6529.3826.0421.6031.1828.0123.35
MuLUT[12]30.5228.1124.8531.3129.0225.2830.2527.6723.7532.2829.8826.36
RC-LUT[16]30.6828.1225.0431.5729.0725.8930.3327.8023.8632.5129.8926.50
SPF-LUT[14]30.9728.5625.3331.8629.5826.2630.8928.2624.2231.7730.4426.91
DnLUT(Ours)32.4129.8826.0333.0230.2426.7432.1228.8725.0132.8830.4427.12
ClassicalCBM3D[3]33.5230.7127.3834.2831.6829.0233.9331.3627.9334.0631.6628.51
MC-WNNM[24]31.9829.3226.9833.2330.8928.6730.2329.2327.0031.2330.2027.55
DNNDnCNN[30]33.9031.2427.9534.6032.1428.9632.9830.8127.5933.4531.5228.62
SwinIR[15]34.4231.7828.5635.3432.8929.7935.6133.2030.2235.1332.9029.82
" + }, + { + "type": "table_caption", + "bbox": [ + 0.09, + 0.289, + 0.907, + 0.318 + ], + "angle": 0, + "content": "Table 3. Quantitative comparison on real-world color image denoising. The table presents CPSNR(dB) and SSIM values. Methods on DnD dataset are validated on the online platform. We could only get the PSNR rather than CPSNR." + }, + { + "type": "table", + "bbox": [ + 0.11, + 0.328, + 0.887, + 0.4 + ], + "angle": 0, + "content": "
DatasetsMethodSR-LUT[10]MuLUT[12]RC-LUT[16]SPF-LUT[14]DnLUT(Ours)CBM3D[3]MC-WNNM[24]DnCNN[30]
SIDDCPSNR29.3833.2433.8834.9135.4430.1429.4536.45
SSIM0.6340.8300.8390.8650.8750.7020.6820.900
DnDPSNR*33.6935.1135.4336.2236.6733.1232.3437.11
SSIM0.8390.8680.8750.9110.9220.8230.8170.932
" + }, + { + "type": "image", + "bbox": [ + 0.104, + 0.411, + 0.4, + 0.55 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.221, + 0.55, + 0.286, + 0.562 + ], + "angle": 0, + "content": "Full Image" + }, + { + "type": "image", + "bbox": [ + 0.409, + 0.411, + 0.526, + 0.475 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.442, + 0.474, + 0.49, + 0.484 + ], + "angle": 0, + "content": "SRLUT" + }, + { + "type": "image", + "bbox": [ + 0.532, + 0.411, + 0.65, + 0.475 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.565, + 0.474, + 0.614, + 0.485 + ], + "angle": 0, + "content": "MuLUT" + }, + { + "type": "image", + "bbox": [ + 0.655, + 0.411, + 0.769, + 0.475 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.685, + 0.474, + 0.737, + 0.485 + ], + "angle": 0, + "content": "RC-LUT" + }, + { + "type": "image", + "bbox": [ + 0.773, + 0.411, + 0.895, + 0.475 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.807, + 0.474, + 0.859, + 0.484 + ], + "angle": 0, + "content": "SPFLUT" + }, + { + "type": "image", + "bbox": [ + 0.104, + 0.599, + 0.264, + 0.722 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.175, + 0.723, + 0.21, + 0.735 + ], + "angle": 0, + "content": "Input" + }, + { + "type": "image", + "bbox": [ + 0.266, + 0.599, + 0.422, + 0.722 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.318, + 0.722, + 0.368, + 0.734 + ], + "angle": 0, + "content": "SRLUT" + }, + { + "type": "image", + "bbox": [ + 0.422, + 0.599, + 0.576, + 0.722 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.47, + 0.723, + 0.523, + 0.734 + ], + "angle": 0, + "content": "MuLUT" + }, + { + "type": "image", + "bbox": [ + 0.577, + 0.599, + 0.736, + 0.722 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.626, + 0.723, + 0.683, + 0.734 + ], + "angle": 0, + "content": "RC-LUT" + }, + { + "type": "image", + "bbox": [ + 0.736, + 0.599, + 0.892, + 0.722 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.786, + 0.723, + 0.844, + 0.734 + ], + "angle": 0, + "content": "SPFLUT" + }, + { + "type": "image", + "bbox": [ + 0.107, + 0.735, + 0.264, + 0.857 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.178, + 0.858, + 0.2, + 0.869 + ], + "angle": 0, + "content": "GT" + }, + { + "type": "image", + "bbox": [ + 0.266, + 0.735, + 0.422, + 0.857 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.315, + 0.858, + 0.37, + 0.869 + ], + "angle": 0, + "content": "CBM3D" + }, + { + "type": "image", + "bbox": [ + 0.422, + 0.735, + 0.577, + 0.857 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.459, + 0.858, + 0.541, + 0.869 + ], + "angle": 0, + "content": "MC-WNNM" + }, + { + "type": "image", + "bbox": [ + 0.578, + 0.735, + 0.736, + 0.857 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.611, + 0.858, + 0.702, + 0.87 + ], + "angle": 0, + "content": "DnLUT(Ours)" + }, + { + "type": "image", + "bbox": [ + 0.736, + 0.735, + 0.892, + 0.857 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.789, + 0.858, + 0.843, + 0.869 + ], + "angle": 0, + "content": "DnCNN" + }, + { + "type": "image_caption", + "bbox": [ + 0.33, + 0.883, + 0.667, + 0.897 + ], + "angle": 0, + "content": "Figure 6. Qualitative comparison on real-world datasets." + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.946, + 0.516, + 0.957 + ], + "angle": 0, + "content": "7587" + } + ], + [ + { + "type": "table_caption", + "bbox": [ + 0.09, + 0.089, + 0.908, + 0.131 + ], + "angle": 0, + "content": "Table 4. PCM can be easily incorporated to the existing LUT-based methods. It equips the model with the ability of capturing the channel correlations which brings more than 1dB on the widely used benchmark. The noise level \\(\\sigma\\) of Gaussian noise is set to 25. The table presents CPSNR(dB) values." + }, + { + "type": "table", + "bbox": [ + 0.164, + 0.143, + 0.833, + 0.291 + ], + "angle": 0, + "content": "
MethodCBSD68Kodak24Urban100McMasterSIDD
SR-LUT[10]26.7127.1626.0428.0129.38
PCM+SR-LUT28.04(+1.33)28.55(+1.39)27.33(+1.29)28.78(+0.77)32.33(+2.96)
MuLUT[12]28.1129.0127.6729.8833.24
PCM+MuLUT29.17(+1.05)29.92(+0.91)28.74(+1.07)30.33(+0.44)34.33(+1.09)
RC-LUT[16]28.1229.0727.8029.8933.88
PCM+RC-LUT29.22(+1.10)30.09(+1.02)28.77(+0.97)30.54(+0.65)34.67(+0.79)
SPF-LUT[14]28.5629.5828.2630.4434.91
PCM+SPF-LUT29.72(+1.16)30.63(+1.05)29.43(+1.17)30.96(+0.52)35.56(+0.65)
" + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.298, + 0.281, + 0.314 + ], + "angle": 0, + "content": "4.4. Quantitative results" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.32, + 0.484, + 0.426 + ], + "angle": 0, + "content": "We evaluate performance using color peak signal-to-noise ratio(CPSNR) and structural similarity index(SSIM) metrics, comparing against state-of-the-art methods across three categories: LUT-based methods (SR-LUT[10], MuLUT[12], RC-LUT[16], SPF-LUT[14]), classical approaches (CBM3D[3], MC-WNNM[24]), and DNN-based solutions (DnCNN[30], SwinIR[15])." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.427, + 0.484, + 0.623 + ], + "angle": 0, + "content": "As shown in Tab. 2, DnLUT achieves superior performance in the LUT-based category, surpassing SPF-LUT by over 1dB in Gaussian noise removal while requiring only \\(17\\%\\) of its storage. Results from Tab. 3 demonstrate DnLUT's effectiveness in real-world scenarios, outperforming classical methods CBM3D[3] and MC-WNNM[24] by approximately 5dB. Though the Gaussian noise is independent across RGB channels, the network can still leverage inter-channel natural distribution in images to suppress the noise. For example, if a pixel in the R channel has a high value due to noise, but the corresponding pixels in the G and B channels do not, the network can infer that the high value is likely noise and suppress it." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.634, + 0.27, + 0.649 + ], + "angle": 0, + "content": "4.5. Qualitative results" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.656, + 0.484, + 0.807 + ], + "angle": 0, + "content": "Visual comparisons in Figs. 5 & 6 illustrate DnLUT's superior denoising capabilities. Existing LUT-based methods exhibit noticeable color distortion, particularly evident in Fig. 5's first row. In contrast, DnLUT achieves visual quality comparable to computation-intensive methods like DnCNN and CBM3D. Fig. 6 further demonstrates DnLUT's advantage in real-world scenarios, where both LUT-based and classical methods struggle with color fidelity and often over-smooth details. DnLUT successfully preserves color consistency while retaining fine image details." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.818, + 0.292, + 0.833 + ], + "angle": 0, + "content": "4.6. Efficiency Evaluation" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.84, + 0.483, + 0.885 + ], + "angle": 0, + "content": "We conduct comprehensive efficiency analysis across three key metrics: theoretical energy cost, runtime performance, and storage requirements (Tab. 5)." + }, + { + "type": "text", + "bbox": [ + 0.091, + 0.886, + 0.484, + 0.902 + ], + "angle": 0, + "content": "Energy cost. Following AdderSR[22], we analyze multipli" + }, + { + "type": "image", + "bbox": [ + 0.522, + 0.296, + 0.617, + 0.35 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.551, + 0.35, + 0.588, + 0.358 + ], + "angle": 0, + "content": "SRLUT" + }, + { + "type": "image", + "bbox": [ + 0.617, + 0.297, + 0.71, + 0.351 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.646, + 0.351, + 0.684, + 0.358 + ], + "angle": 0, + "content": "MuLUT" + }, + { + "type": "image", + "bbox": [ + 0.71, + 0.297, + 0.805, + 0.351 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.738, + 0.351, + 0.778, + 0.359 + ], + "angle": 0, + "content": "RC-LUT" + }, + { + "type": "image", + "bbox": [ + 0.806, + 0.297, + 0.901, + 0.351 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.826, + 0.351, + 0.868, + 0.359 + ], + "angle": 0, + "content": "SPFLUT" + }, + { + "type": "image", + "bbox": [ + 0.522, + 0.362, + 0.616, + 0.416 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.536, + 0.417, + 0.602, + 0.426 + ], + "angle": 0, + "content": "PCM+SRLUT" + }, + { + "type": "image", + "bbox": [ + 0.617, + 0.362, + 0.71, + 0.416 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.631, + 0.417, + 0.698, + 0.426 + ], + "angle": 0, + "content": "PCM+MuLUT" + }, + { + "type": "image", + "bbox": [ + 0.71, + 0.362, + 0.805, + 0.416 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.723, + 0.417, + 0.793, + 0.426 + ], + "angle": 0, + "content": "PCM+RC-LUT" + }, + { + "type": "image", + "bbox": [ + 0.806, + 0.362, + 0.901, + 0.416 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.819, + 0.417, + 0.889, + 0.426 + ], + "angle": 0, + "content": "PCM+SPFLUT" + }, + { + "type": "image_caption", + "bbox": [ + 0.515, + 0.438, + 0.904, + 0.453 + ], + "angle": 0, + "content": "Figure 7. Visualization of LUT-based method with plug-in PCM." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.465, + 0.906, + 0.556 + ], + "angle": 0, + "content": "cation and addition operations across different data types to estimate total energy consumption. DnLUT achieves \\(70\\%\\) energy savings compared to SPF-LUT while delivering superior denoising quality. The efficiency gap widens dramatically when compared to DNN-based methods, with DnLUT requiring only \\(0.1\\%\\) of DnCNN's energy consumption." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.558, + 0.906, + 0.693 + ], + "angle": 0, + "content": "Runtime. We implement LUT-based methods using standard JAVA API on the ANDROID platform. DnLUT maintains the rapid inference capabilities characteristic of SR-LUT[10], demonstrating clear advantages over classical and DNN-based approaches. For reference, DnCNN running on mobile CPU requires \\(20 \\times\\) more processing time. DnLUT's performance can be further optimized through device-specific implementations, such as FPGA deployments." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.696, + 0.906, + 0.756 + ], + "angle": 0, + "content": "Storage. DnLUT requires approximately 500 KB storage, well within typical L2-cache limits of modern chips. This efficient memory footprint enables rapid lookup operations with minimal dependency overhead." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.771, + 0.776, + 0.788 + ], + "angle": 0, + "content": "4.7. Effectiveness of Plug-in PCM" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.796, + 0.906, + 0.902 + ], + "angle": 0, + "content": "Tab. 4 demonstrates PCM's versatility as a performance enhancer for existing LUT-based methods. When integrated with current approaches, PCM consistently delivers significant improvements across all benchmark datasets. The visual comparisons in Fig. 7 show marked reduction in color distortion, particularly in smooth image regions, resulting in superior perceptual quality." + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.946, + 0.516, + 0.957 + ], + "angle": 0, + "content": "7588" + } + ], + [ + { + "type": "table_caption", + "bbox": [ + 0.09, + 0.089, + 0.908, + 0.119 + ], + "angle": 0, + "content": "Table 5. Efficiency evaluation of three categories color image denoising methods. We use the Qualcomm Snapdragon 8 Gen 2 as the mobile platform. The energy cost is calculated on the \\( {512} \\times {512} \\) resolution color images." + }, + { + "type": "table", + "bbox": [ + 0.092, + 0.13, + 0.92, + 0.321 + ], + "angle": 0, + "content": "
CategoryMethodPlatformRuntime(ms) 256 × 256Runtime(ms) 512 × 512Energy Cost (pJ)Storage (KB)
LUT-basedSR-LUT[10]Mobile619149.98M82
MuLUT[12]Mobile2180899.88M489
RC-LUT[6]Mobile1878612.92M326
SPF-LUT[14]Mobile722572.32G30,178
DnLUT(ours)Mobile2388687.34M518
ClassicalCBM3D[29]PC3,18914,3434.82G-
MC-WNNM[24]PC74,290293,42489.23G-
DNNDnCNN[30]Mobile5431,923542.53G2,239
SwinIR[15]Mobile73,234274,45312.03T45,499
" + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.332, + 0.241, + 0.348 + ], + "angle": 0, + "content": "4.8. Ablation study" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.354, + 0.483, + 0.385 + ], + "angle": 0, + "content": "We conduct ablation studies using Gaussian color image denoising benchmark to validate DnLUT's key components." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.393, + 0.368, + 0.407 + ], + "angle": 0, + "content": "4.8.1. Channel-spatial-wise convolution" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.412, + 0.484, + 0.502 + ], + "angle": 0, + "content": "Channel indexing[13, 25] possesses a \\(1 \\times 1\\) receptive field to extract channel information. We evaluate various combinations of channel-wise, spatial-wise, and our proposed channel-spatial-wise kernels using a consistent architecture comprising one channel indexing module, two L-shaped convolution modules, and one PCM component." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.503, + 0.484, + 0.639 + ], + "angle": 0, + "content": "Results in Tab. 6 reveal that naive combination of channel-wise and spatial-wise convolutions yields suboptimal performance. Pure channel-spatial-wise convolution, while promising, struggles with limited receptive field coverage. The integration of spatial-wise and channel-spatial-wise convolutions produces substantial improvements. Notably, adding channel-wise convolution offers minimal additional benefit, as channel-spatial-wise convolution already captures comprehensive color information." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.645, + 0.288, + 0.66 + ], + "angle": 0, + "content": "4.8.2. L-shaped convolution" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.664, + 0.483, + 0.771 + ], + "angle": 0, + "content": "To validate our L-shaped convolution design, we compare it against standard \\(2 \\times 2\\) spatial-wise kernels (denoted as S in Tab. 7, \\(\\sigma = 1\\)). While the conventional kernel shows marginal performance advantages, our L-shaped design achieves comparable results while enabling 3D LUT conversion, reducing storage requirements by \\(17 \\times\\). It paves the way of deploying multi-layer spatial-wise convolutions." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.785, + 0.21, + 0.801 + ], + "angle": 0, + "content": "5. Conclusion" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.81, + 0.484, + 0.902 + ], + "angle": 0, + "content": "This paper presents DnLUT, a novel approach to color image denoising that bridges the gap between high-performance deep learning models and resource-constrained edge devices. Addressing the limitations of existing LUT-based methods, DnLUT introduces two key innovations. First, the Pairwise Channel Mixer (PCM) models channel and spa" + }, + { + "type": "table_caption", + "bbox": [ + 0.513, + 0.33, + 0.907, + 0.386 + ], + "angle": 0, + "content": "Table 6. Different combinations of channel-wise, spatial-wise, and our channel-spatial-wise kernels. These kernels are represented by channel indexing[13], L-shaped kernel, and PCM. The table presents CPSNR(dB) values." + }, + { + "type": "table", + "bbox": [ + 0.516, + 0.388, + 0.909, + 0.489 + ], + "angle": 0, + "content": "
ChannelSpatialChannel-spatialCBSD68
✓29.77
✓30.15
✓✓30.21
✓✓31.41
✓✓✓31.47
" + }, + { + "type": "table_caption", + "bbox": [ + 0.514, + 0.492, + 0.906, + 0.521 + ], + "angle": 0, + "content": "Table 7. Comparison of L-shaped kernel and \\(2 \\times 2\\) spatial-wise kernel. The table presents CPSNR(dB) values." + }, + { + "type": "table", + "bbox": [ + 0.518, + 0.522, + 0.907, + 0.574 + ], + "angle": 0, + "content": "
MethodCBSD68KodakMcMasterLUT Size(KB)
PCM + L26.6327.0527.96494 + 24
PCM + S26.7127.1628.01494 + 408
" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.592, + 0.907, + 0.729 + ], + "angle": 0, + "content": "tial correlations simultaneously by processing channel pairs, enabling comprehensive color modeling without excessive storage costs. Second, an L-shaped convolution design rethinks rotation-based receptive field expansion, eliminating redundant pixel usage and converting 4D to 3D LUTs, reducing storage by \\(17\\times\\) while maintaining spatial coverage. These advancements allow DnLUT to outperform state-ofthe-art LUT-based methods by over 1dB in denoising quality while consuming just \\(0.1\\%\\) of DnCNN's energy." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.733, + 0.909, + 0.9 + ], + "angle": 0, + "content": "Acknowledgements. This research was supported by: the HKU-TCL Joint Research Centre for AI, the Theme-based Research Scheme (TRS) project T45-701/22-R, and the General Research Fund (GRF) Project 17203224 from the Research Grants Council (RGC) of Hong Kong SAR. Additional support was provided by the National Key Research and Development Program of China (Grant No. 2024YFB2808903), the Shenzhen Science and Technology Program (JCYJ20220818101001004), and the Tsinghua University - Tencent Joint Laboratory for Internet Innovation Technology Research Fund." + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.516, + 0.957 + ], + "angle": 0, + "content": "7589" + } + ], + [ + { + "type": "title", + "bbox": [ + 0.093, + 0.09, + 0.188, + 0.106 + ], + "angle": 0, + "content": "References" + }, + { + "type": "ref_text", + "bbox": [ + 0.101, + 0.115, + 0.484, + 0.171 + ], + "angle": 0, + "content": "[1] Eirikur Agustsson and Radu Timofte. Ntire 2017 challenge on single image super-resolution: Dataset and study. In The IEEE Conference on Computer Vision and Pattern Recognition (CVPR) Workshops, 2017. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.101, + 0.173, + 0.484, + 0.243 + ], + "angle": 0, + "content": "[2] Hanting Chen, Yunhe Wang, Tianyu Guo, Chang Xu, Yiping Deng, Zhenhua Liu, Siwei Ma, Chunjing Xu, Chao Xu, and Wen Gao. Pre-trained image processing transformer. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 12299-12310, 2021. 1, 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.245, + 0.484, + 0.314 + ], + "angle": 0, + "content": "[3] Kostadin Dabov, Alessandro Foi, Vladimir Katkovnik, and Karen Egiazarian. Color image denoising via sparse 3d collaborative filtering with grouping constraint in luminance-chrominance space. In 2007 IEEE international conference on image processing, pages I-313. IEEE, 2007. 2, 6, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.316, + 0.484, + 0.37 + ], + "angle": 0, + "content": "[4] Kostadin Dabov, Alessandro Foi, Vladimir Katkovnik, and Karen Egiazarian. Image denoising by sparse 3-d transform-domain collaborative filtering. IEEE Transactions on image processing, 16(8):2080-2095, 2007. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.373, + 0.484, + 0.428 + ], + "angle": 0, + "content": "[5] Jingjing Dai, Oscar C Au, Lu Fang, Chao Pang, Feng Zou, and Jiali Li. Multichannel nonlocal means fusion for color image denoising. IEEE Transactions on circuits and systems for video technology, 23(11):1873-1886, 2013. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.431, + 0.484, + 0.485 + ], + "angle": 0, + "content": "[6] Michael Gharbi, Jiawen Chen, Jonathan T Barron, Samuel W Hasinoff, and Fredo Durand. Deep bilateral learning for real-time image enhancement. ACM Transactions on Graphics (TOG), 36(4):1-12, 2017. 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.488, + 0.484, + 0.556 + ], + "angle": 0, + "content": "[7] Shuhang Gu, Lei Zhang, Wangmeng Zuo, and Xiangchu Feng. Weighted nuclear norm minimization with application to image denoising. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 2862-2869, 2014. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.559, + 0.484, + 0.613 + ], + "angle": 0, + "content": "[8] Binxiao Huang, Jason Chun Lok Li, Jie Ran, Boyu Li, Jiajun Zhou, Dahai Yu, and Ngai Wong. Hundred-kilobyte lookup tables for efficient single-image super-resolution. arXiv preprint arXiv:2312.06101, 2023. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.616, + 0.484, + 0.671 + ], + "angle": 0, + "content": "[9] Jia-Bin Huang, Abhishek Singh, and Narendra Ahuja. Single image super-resolution from transformed self-exemplars. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 5197-5206, 2015. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.674, + 0.484, + 0.729 + ], + "angle": 0, + "content": "[10] Younghyun Jo and Seon Joo Kim. Practical single-image super-resolution using look-up table. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 691-700, 2021. 2, 3, 5, 6, 7, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.731, + 0.484, + 0.786 + ], + "angle": 0, + "content": "[11] Stamatios Lefkimmiatis. Non-local color image denoising with convolutional neural networks. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 3587-3596, 2017. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.788, + 0.484, + 0.844 + ], + "angle": 0, + "content": "[12] Jiacheng Li, Chang Chen, Zhen Cheng, and Zhiwei Xiong. Mulut: Cooperating multiple look-up tables for efficient image super-resolution. In European Conference on Computer Vision, pages 238-256. Springer, 2022. 2, 3, 5, 6, 7, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.846, + 0.484, + 0.901 + ], + "angle": 0, + "content": "[13] Jiacheng Li, Chang Chen, Zhen Cheng, and Zhiwei Xiong. Toward dnn of luts: Learning efficient image restoration with multiple look-up tables. IEEE Transactions on Pattern Analysis and Machine Intelligence, 2024. 2, 8" + }, + { + "type": "list", + "bbox": [ + 0.094, + 0.115, + 0.484, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.093, + 0.905, + 0.148 + ], + "angle": 0, + "content": "[14] Yinglong Li, Jiacheng Li, and Zhiwei Xiong. Look-up table compression for efficient image restoration. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 26016-26025, 2024. 2, 3, 6, 7, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.149, + 0.905, + 0.217 + ], + "angle": 0, + "content": "[15] Jingyun Liang, Jiezhang Cao, Guolei Sun, Kai Zhang, Luc Van Gool, and Radu Timofte. Swinir: Image restoration using swin transformer. In Proceedings of the IEEE/CVF international conference on computer vision, pages 1833-1844, 2021. 1, 2, 6, 7, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.219, + 0.905, + 0.288 + ], + "angle": 0, + "content": "[16] Guandu Liu, Yukang Ding, Mading Li, Ming Sun, Xing Wen, and Bin Wang. Reconstructed convolution module based look-up tables for efficient image super-resolution. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 12217-12226, 2023. 2, 3, 6, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.289, + 0.905, + 0.343 + ], + "angle": 0, + "content": "[17] Cheng Ma, Jingyi Zhang, Jie Zhou, and Jiwen Lu. Learning series-parallel lookup tables for efficient image superresolution. In European Conference on Computer Vision, pages 305–321. Springer, 2022. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.344, + 0.905, + 0.412 + ], + "angle": 0, + "content": "[18] Kede Ma, Zhengfang Duanmu, Qingbo Wu, Zhou Wang, Hongwei Yong, Hongliang Li, and Lei Zhang. Waterloo exploration database: New challenges for image quality assessment models. IEEE Transactions on Image Processing, 26 (2):1004-1016, 2016. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.414, + 0.905, + 0.455 + ], + "angle": 0, + "content": "[19] Julien Mairal, Michael Elad, and Guillermo Sapiro. Sparse representation for color image restoration. IEEE Transactions on image processing, 17(1):53-69, 2007. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.456, + 0.905, + 0.523 + ], + "angle": 0, + "content": "[20] D. Martin, C. Fowlkes, D. Tal, and J. Malik. A database of human segmented natural images and its application to evaluating segmentation algorithms and measuring ecological statistics. In Proc. 8th Int'l Conf. Computer Vision, pages 416-423, 2001. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.525, + 0.905, + 0.607 + ], + "angle": 0, + "content": "[21] David Martin, Charless Fowlkes, Doron Tal, and Jitendra Malik. A database of human segmented natural images and its application to evaluating segmentation algorithms and measuring ecological statistics. In Proceedings eighth IEEE international conference on computer vision. ICCV 2001, pages 416-423. IEEE, 2001. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.609, + 0.905, + 0.677 + ], + "angle": 0, + "content": "[22] Dehua Song, Yunhe Wang, Hanting Chen, Chang Xu, Chunjing Xu, and DaCheng Tao. Addersr: Towards energy efficient image super-resolution. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 15648-15657, 2021. 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.679, + 0.905, + 0.748 + ], + "angle": 0, + "content": "[23] Radu Timofte, Eirikur Agustsson, Luc Van Gool, Ming-Hsuan Yang, and Lei Zhang. Ntire 2017 challenge on single image super-resolution: Methods and results. In Proceedings of the IEEE conference on computer vision and pattern recognition workshops, pages 114-125, 2017. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.749, + 0.905, + 0.815 + ], + "angle": 0, + "content": "[24] Jun Xu, Lei Zhang, David Zhang, and Xiangchu Feng. Multichannel weighted nuclear norm minimization for real color image denoising. In Proceedings of the IEEE international conference on computer vision, pages 1096-1104, 2017. 2, 6, 7, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.818, + 0.905, + 0.871 + ], + "angle": 0, + "content": "[25] Sidi Yang, Binxiao Huang, Mingdeng Cao, Yatai Ji, Hanzhong Guo, Ngai Wong, and Yujiu Yang. Taming lookup tables for efficient image retouching. arXiv preprint arXiv:2403.19238, 2024. 2, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.873, + 0.905, + 0.901 + ], + "angle": 0, + "content": "[26] Syed Waqas Zamir, Aditya Arora, Salman Khan, Munawar Hayat, Fahad Shahbaz Khan, Ming-Hsuan Yang, and Ling" + }, + { + "type": "list", + "bbox": [ + 0.094, + 0.093, + 0.905, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.517, + 0.957 + ], + "angle": 0, + "content": "7590" + } + ], + [ + { + "type": "ref_text", + "bbox": [ + 0.124, + 0.092, + 0.482, + 0.134 + ], + "angle": 0, + "content": "Shao. Multi-stage progressive image restoration. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 14821-14831, 2021. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.136, + 0.483, + 0.218 + ], + "angle": 0, + "content": "[27] Syed Waqas Zamir, Aditya Arora, Salman Khan, Munawar Hayat, Fahad Shahbaz Khan, and Ming-Hsuan Yang. Restormer: Efficient transformer for high-resolution image restoration. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 5728-5739, 2022. 1, 2, 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.22, + 0.482, + 0.288 + ], + "angle": 0, + "content": "[28] Hui Zeng, Jianrui Cai, Lida Li, Zisheng Cao, and Lei Zhang. Learning image-adaptive 3d lookup tables for high performance photo enhancement in real-time. IEEE Transactions on Pattern Analysis and Machine Intelligence, 44(4):2058-2073, 2020. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.29, + 0.482, + 0.359 + ], + "angle": 0, + "content": "[29] Fengyi Zhang, Hui Zeng, Tianjun Zhang, and Lin Zhang. Clut-net: Learning adaptively compressed representations of 3dluts for lightweight image enhancement. In Proceedings of the 30th ACM International Conference on Multimedia, pages 6493-6501, 2022. 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.361, + 0.482, + 0.416 + ], + "angle": 0, + "content": "[30] Kai Zhang, Wangmeng Zuo, Yunjin Chen, Deyu Meng, and Lei Zhang. Beyond a gaussian denoiser: Residual learning of deep cnn for image denoising. IEEE transactions on image processing, 26(7):3142-3155, 2017. 1, 2, 6, 7, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.418, + 0.482, + 0.471 + ], + "angle": 0, + "content": "[31] Lei Zhang, Xiaolin Wu, Antoni Buades, and Xin Li. Color demosaicking by local directional interpolation and nonlocal adaptive thresholding. Journal of Electronic imaging, 20(2): 023016-023016, 2011. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.474, + 0.482, + 0.529 + ], + "angle": 0, + "content": "[32] Yi Zhang, Dasong Li, Xiaoyu Shi, Dailan He, Kangning Song, Xiaogang Wang, Hongwei Qin, and Hongsheng Li. Kbnet: Kernel basis network for image restoration. arXiv preprint arXiv:2303.02881, 2023. 1, 2" + }, + { + "type": "list", + "bbox": [ + 0.093, + 0.092, + 0.483, + 0.529 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.946, + 0.515, + 0.957 + ], + "angle": 0, + "content": "7591" + } + ] +] \ No newline at end of file diff --git a/2025/DnLUT_ Ultra-Efficient Color Image Denoising via Channel-Aware Lookup Tables/efb58f94-4255-404d-9c8c-4e560774a0e2_origin.pdf b/2025/DnLUT_ Ultra-Efficient Color Image Denoising via Channel-Aware Lookup Tables/efb58f94-4255-404d-9c8c-4e560774a0e2_origin.pdf new file mode 100644 index 0000000000000000000000000000000000000000..810e228d5def01038b791d71f36bdf23132a2741 --- /dev/null +++ b/2025/DnLUT_ Ultra-Efficient Color Image Denoising via Channel-Aware Lookup Tables/efb58f94-4255-404d-9c8c-4e560774a0e2_origin.pdf @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:f5417206160687ba3f7f82ceee36ee587a5fb7b58de7d13ed143f076e8059dd2 +size 961232 diff --git a/2025/DnLUT_ Ultra-Efficient Color Image Denoising via Channel-Aware Lookup Tables/full.md b/2025/DnLUT_ Ultra-Efficient Color Image Denoising via Channel-Aware Lookup Tables/full.md new file mode 100644 index 0000000000000000000000000000000000000000..37a03ceff5dfb7d45d99aa24c659c064e2456e0e --- /dev/null +++ b/2025/DnLUT_ Ultra-Efficient Color Image Denoising via Channel-Aware Lookup Tables/full.md @@ -0,0 +1,366 @@ +# DnLUT: Ultra-Efficient Color Image Denoising via Channel-Aware Lookup Tables + +Sidi Yang $^{1}$ Binxiao Huang $^{1\dagger}$ Yulun Zhang $^{2}$ Dahai Yu $^{3}$ Yujiu Yang $^{4\dagger}$ Ngai Wong $^{1}$ + +1 The University of Hong Kong 2 Shanghai Jiaotong University 3 TCL Corporate Research 4 Tsinghua University + +{yangsidi99, huanghx7}@connect.hku.hk, yulun100@gmail.com, dahai.yu@tcl.com + +yang.yujiu@sz.tsinghua.edu.cn, nwong@eee.hku.hk + +# Abstract + +While deep neural networks have revolutionized image denoising capabilities, their deployment on edge devices remains challenging due to substantial computational and memory requirements. To this end, we present DnLUT, an ultra-efficient lookup table-based framework that achieves high-quality color image denoising with minimal resource consumption. Our key innovation lies in two complementary components: a Pairwise Channel Mixer (PCM) that effectively captures inter-channel correlations and spatial dependencies in parallel, and a novel L-shaped convolution design that maximizes receptive field coverage while minimizing storage overhead. By converting these components into optimized lookup tables post-training, DnLUT achieves remarkable efficiency - requiring only 500KB storage and $0.1\%$ energy consumption compared to its CNN contestant DnCNN, while delivering $20\times$ faster inference. Extensive experiments demonstrate that DnLUT outperforms all existing LUT-based methods by over 1dB in PSNR, establishing a new state-of-the-art in resource-efficient color image denoising. The project is available at https://github.com/Stephen0808/DnLUT. + +# 1. Introduction + +Color images pervade our digital world - from social media and photography to scientific imaging and medical diagnostics. These images are inherently more vulnerable to noise contamination than their grayscale counterparts, not only due to their threefold data size during acquisition, storage, and transmission, but also because human perception exhibits heightened sensitivity to color distortions compared to luminance variations. + +The advent of deep neural networks (DNNs) has dramatically advanced the field of color image denoising[2, 15, + +![](images/df4d233aaf8694d8dcbde46b770703422f6b3e312c780d0c901754af163409e1.jpg) +Figure 1. Model comparison in terms of color peak signal-to-noise ratio (CPSNR), runtime, and storage. The CPSNR and runtime are calculated on the CBSD68 dataset with Gaussian noise level $\sigma = 25$ using Qualcomm Snapdragon 8 Gen2. Our method outperforms state-of-the-art LUT-based methods with the highest CPSNR, at a low storage requirements and reduced runtime. Additionally, our PCM module serves as a versatile plug-in module that enhances existing methods' performance by over 1dB. + +27, 30, 32]. State-of-the-art models leverage sophisticated architectures with numerous convolutional layers or transformer blocks, achieving unprecedented denoising quality. However, these advances come at a steep cost - the computational complexity and memory requirements of such models render them impractical for edge devices. This limitation is particularly acute given that most edge devices lack specialized hardware accelerators such as graphics processing units (GPs) or tensor processing units (TPUs), creating a significant gap between algorithmic capabilities and practical deployability. + +Recent works have explored lookup tables (LUTs) as an elegant solution for efficient image enhancement. LUTs replace complex runtime computations with simple array indexing operations through direct addressing. This approach typically involves training a neural network model, then con + +verting it into LUTs by exhaustively mapping input-output relationships. During inference, the system performs rapid table lookups instead of costly computations, effectively leveraging DNN capabilities while circumventing hardware constraints. However, LUT-based methods face a fundamental challenge: the storage requirement grows exponentially with the input dimensionality. This has led to the adoption of 4D LUTs (using four pixel values for indexing) as a practical compromise between storage and performance. + +Current LUT-based methods predominantly employ $2 \times 2$ convolutional kernels processed independently for each channel[10, 12, 14, 16], creating spatial-wise LUTs for single-channel processing. While this approach works well for tasks like super-resolution, it proves inadequate for color image denoising where noise typically affects all RGB channels simultaneously. Channel correlation plays a pivotal role in modeling noise characteristics, a critical factor for achieving effective restoration. Alternative approaches either use $1 \times 1$ convolutions with three-channel depth for RGB information (creating 3D channel-wise LUTs)[13, 25] or attempt to use larger $H \times W \times 3$ kernels. However, these solutions either disrupt noise distribution patterns or lead to prohibitive storage requirements. + +To address these challenges, we propose DnLUT, a novel lookup table-based framework specifically designed for efficient color image denoising. Our approach centers on two key innovations: + +First, we introduce the Pairwise Channel Mixer (PCM), which simultaneously processes spatial and channel information. PCM strategically combines RGB channels into three pairs (RG, GB, BR), processing them through parallel branches using a $1 \times 2$ convolution kernel with depth 2. This design efficiently captures channel correlations while maintaining manageable storage requirements. The PCM module can also serve as a plug-in module for existing LUT-based methods, consistently improving their performance by approximately 1dB. + +Second, we develop an innovative L-shaped convolution kernel that addresses the limitations of conventional rotation-based approaches. After taking the channel correlation into account, as discussed in [12, 14], there is still a need to enlarge the spatial receptive field. While previous methods like SR-LUT[10] use rotation ensemble training to expand receptive fields of spatial-wise LUT, they often introduce redundant pixel usage and significant storage overhead. Our L-shaped kernel design eliminates pixel overlap during rotations, enabling conversion to more efficient 3D LUTs while maintaining the same effective receptive field size as 4D LUTs. + +Extensive experimental validation demonstrates the superiority of DnLUT over existing approaches: + +- We achieve up to 1.3dB CPSNR improvement over SPFLUT[14] on standard Gaussian denoising bench + +marks. For real-world denoising tasks, we outperform CBM3D[3] by nearly 5dB. + +- Our method maintains exceptional efficiency, requiring only $0.1\%$ of DnCNN[30]'s energy consumption and 500KB storage. +- The PCM module serves as a versatile performance enhancer, bringing over 1dB CPSNR improvement to existing LUT-based methods. +- Our L-shaped convolution design reduces storage requirements by $17 \times$ while preserving receptive field coverage. + +These results establish DnLUT as a significant advance in practical, resource-efficient color image denoising, bridging the gap between algorithmic capability and deployability on edge devices. + +# 2. Related works + +# 2.1. Color Image denoising + +A fundamental characteristic of color images is the strong correlation among RGB channels[5]. Research has consistently shown that processing these channels independently yields suboptimal results compared to joint processing approaches[19]. This insight has driven numerous innovations in color image denoising. + +A landmark development was CBM3D[3], which operates in luminance-chrominance space (i.e., $YC_{b}C_{r}$ ). The method performs BM3D[4] grouping once on the luminance channel $Y$ and leverages this grouping for collaborative filtering of chrominance channels $U$ and $V$ . Building on this concept, [11] introduced a radial basis function mixture for more sophisticated handling of channel correlations. Another notable approach, MC-WNNM[24], extended WNNM[7] by incorporating a weight matrix to adaptively balance the data fidelity across channels based on their noise characteristics. + +The deep learning era, initiated by DnCNN[30], transformed the field of color image denoising. Subsequent approaches[2, 15, 26, 27, 32] have increasingly sophisticated architectures that process all RGB channels simultaneously through complex networks of convolution layers or transformer blocks. While these models achieve remarkable denoising quality, their substantial computational and memory requirements present significant deployment challenges, particularly for resource-constrained edge devices. + +# 2.2. Replacing CNN with LUT + +The challenge of deploying deep neural networks on edge devices has sparked interest in lookup table (LUT) based alternatives. Inspired by the prevalence of color LUTs in image signal processors, [28] pioneered an approach using learnable basis LUTs combined with a weight prediction network for image enhancement. [25] further developed this + +concept, proposing a more efficient channel-wise LUT system. However, these early channel-wise LUTs, processing single pixels per channel, proved inadequate for the complex task of denoising. + +A significant breakthrough came with [10], which introduced a method to convert CNNs with limited receptive fields into spatial-wise LUTs for super-resolution. This approach trains networks on restricted pixel neighborhoods (i.e., $2 \times 2$ ) and converts them to LUTs by exhaustively mapping input-output relationships. During inference, the system efficiently retrieves pre-computed outputs through coordinate indexing. Follow-up works[12, 16, 17] enhanced this framework using serial LUTs to expand receptive fields, achieving substantial performance improvements. + +To manage the exponential growth of LUT size with input dimensionality, these methods typically process each input channel independently. While this approach proves effective for super-resolution, where spatial information dominates, it falls short for color image denoising, which requires preserving both spatial and channel information to model noise distribution effectively. The limitations of existing channel-wise and spatial-wise LUT approaches highlight the need for a more sophisticated solution that can handle the unique challenges of color image denoising while maintaining computational efficiency. + +# 3. Method + +# 3.1. Preliminary + +The foundation of LUT-based image restoration was established by SR-LUT[10], which introduces a constrained receptive field approach for network training (i.e., $2 \times 2$ ). Post-training, the network's output values are systematically cached into a LUT by exhaustively traversing all possible input pixel value combinations. During inference, input patches of size $2 \times 2$ are converted to indices of a 4D LUT to retrieve corresponding output values. To manage storage requirements, the 4D index space is uniformly subsampled. Additionally, a rotation ensemble strategy rotates each $2 \times 2$ input patch four times, effectively expanding the receptive field to $3 \times 3$ . Recent works have focused on increasing the effective receptive field through various techniques, including multiple LUTs[12, 14, 17], shift aggregations[17], and diverse kernel patterns[8, 12, 16]. + +Our work advances this foundation by specifically addressing the challenges of LUT-based color image denoising, a critical application for edge devices. + +A notable limitation of existing LUT-based approaches is their treatment of channel depth, typically set to 1. For color images, this means each RGB channel independently accesses cached output values without inter-channel interaction. This design overlooks crucial RGB channel correlations and spatial-channel-wise degradation patterns inher + +ent in color image denoising. While a straightforward solution might be to employ vanilla convolution kernel patterns, the exponential growth of LUT size with index dimension makes this impractical. For instance, a kernel of size $2 \times 2$ with depth 3 would require approximately 582 TB of storage, even with subsampling compression. This constraint presents a fundamental dilemma: when aiming to capture more spatial information, we have to sacrifice the richness of channel information. Moreover, while 4D spatial-wise LUTs show promise, they typically require multiple LUTs (analogous to stacked convolution layers), creating significant storage overhead, particularly problematic for L1 cache access. + +Table 1. LUT size versus receptive field (RF) and channel numbers. * means subsampled LUT size + +
RFDepthLUT Dim.LUT Size*
1 × 111D17 B
1 × 133D4.9 KB
2 × 214D83.5 KB
1 × 236D24.1 MB
2 × 2312D582.6 TB
k × kck × k × cD(2^4 + 1)^{k×k×C} B
+ +# 3.2. Overview + +To address these fundamental challenges, we introduce two complementary innovations: 1) a pairwise channel mixer that leverages 4D LUT efficiency while modeling spatial-channel correlations; and 2) a novel L-shaped kernel featuring inherent rotation non-overlapping properties, enabling conversion to 3D LUT while maintaining 4D LUT-equivalent receptive field coverage. + +We first build the DnNet for training, as shown in Fig. 2. Our proposed framework enhances traditional spatial-wise LUT capabilities through these two key modules: the pairwise channel mixer (PCM) establishes robust channel-spatial correlations in three RGB channels while the L-shaped kernel expands the effective receptive field with minimal computational overhead independently in each channel. The architecture combines PCM with L-shaped convolution in two stages. The first stage, inspired by [14], generates multi-channel features from LUT groups, which are then concatenated in the fusion module. The second stage integrates both proposed modules. Importantly, all components are designed for efficient conversion to 3D or 4D LUTs during inference, collectively referred to as DnLUT. + +# 3.3. Pairwise Channel Mixer + +Designing effective kernel patterns for LUT-based methods presents significant challenges. Fig. 3 illustrates various kernel pattern categories. Traditional approaches have + +![](images/dd9fda6a17408baae32f2301a1fc608986e28efd782e4f04f02e5559c180fb15.jpg) + +![](images/03367edfc8631b81ef3fb86a0fad5ce2b921f4a99486e887597f506bec2c5a81.jpg) +Figure 2. System architecture of DnLUT: (a) The DnNet pipeline integrates pairwise channel mixers and L-shaped convolutions, with multi-scale fusion enhancing receptive field coverage. Channel dimensions are flattened for parallel processing in L-shaped operations, then unfolded for PCM input. (b) Input pixels undergo four rotations $(0^{\circ}, 90^{\circ}, 180^{\circ}, 270^{\circ})$ during processing, with outputs averaged for enhanced results. (c) Post-training, all possible input combinations are processed through DnNet modules, with outputs cached in optimized 3D or 4D LUTs. (d) During inference, input pixels are efficiently processed through multiple LUTs, with each LUT's outputs informing subsequent LUT indices, culminating in final denoised pixel values. + +![](images/35f7a845319b0c2c2b47357b2405a7cd8c56526a616dad12d9a111a6b8816e22.jpg) + +![](images/49e0bdd5a3c7f94fb5c891de0651792ff428e90daba707affeb6a3c37541cd27.jpg) + +focused on single-dimension kernels due to LUT index dimension constraints, leading to two major limitations in color image denoising. Spatial-wise kernels, while capturing spatial information, neglect channel correlations, resulting in color distortion and suboptimal noise modeling. Conversely, channel-wise kernels establish strong RGB channel connections but ignore spatial relationships, leading to persistent artifacts. While combining these kernels in a cascade network might seem promising, this approach fails to process channel and spatial information simultaneously, resulting in suboptimal noise distribution modeling. + +Our PCM addresses these limitations through a novel architecture for color image denoising. The module first re-organizes RGB channels into three pairwise combinations: RG, GB, and BR. These pairs are processed through parallel branches using a kernel with $1 \times 2$ spatial dimensions and depth 2 in the initial layer, followed by cascaded $1 \times 1$ convolution layers. Each convolution operation processes four pixel values to produce one channel output, enabling efficient conversion to 4D LUTs post-training. The cached LUTs (i.e., $LUT_{RG}, LUT_{GB}, LUT_{GR}$ ) operate in parallel, with their predictions concatenated. For an input anchor $(I_R, I_G, I_B)$ , the output values $(V_R, V_G, V_B)$ are computed as: + +$$ +\begin{array}{r} \left(V _ {R}, V _ {G}, V _ {B}\right) = \operatorname {C a t} \left(L U T _ {R G} \left[ I _ {0, 0, R} \right] \left[ I _ {0, 1, R} \right] \left[ I _ {0, 0, G} \right] \left[ I _ {0, 1, G} \right], \right. \\ L U T _ {G B} \left[ I _ {0, 0, G} \right] \left[ I _ {0, 1, G} \right] \left[ I _ {0, 0, B} \right] \left[ I _ {0, 1, B} \right], \\ L U T _ {B R} \left[ I _ {0, 0, B} \right] \left[ I _ {0, 1, B} \right] \left[ I _ {0, 0, R} \right] \left[ I _ {0, 1, R} \right]), \end{array} \tag {1} +$$ + +where $H, W, C$ of each $I_{H,W,C}$ denotes height, width, + +![](images/b6f15846ba721620cc417fdac6cdcef8905e19cfa20df9e98c6765d572e6c233.jpg) +(a) Spatial-wise +(b) Channel-wise + +![](images/8c14886eff627301db4bf7cd4752cc4f476dee7819944fe4857ff2c04733b838.jpg) + +![](images/42c0da747a88fbbf959f6edd2b6d129676d65f6a04f94673969dcaefb27b76b0.jpg) +(c) Spatial-Channel-wise +Figure 3. Taxonomy of kernel patterns for LUT-based methods. Dark cubes indicate rotation points, while medium-dark regions show involved pixel positions during one rotation. + +and channel index. PCM effectively balances spatial and channel-wise processing, capturing inter-channel correlations crucial for color fidelity while preserving spatial relationships through the $1 \times 2$ kernel design and refining features via cascaded $1 \times 1$ convolutions. + +# 3.4. Rotation Non-overlapping Kernel + +The rotation ensemble strategy is fundamental to LUT-based methods, using four rotations $(0^{\circ}, 90^{\circ}, 180^{\circ}, 270^{\circ})$ to enable asymmetric convolution kernels to cover larger symmetric areas. By combining outputs from all rotations, the approach enhances performance without increasing LUT index dimensionality. As shown in Fig. 4, traditional methods like SR-LUT and MuLUT use this strategy to expand receptive fields to $3 \times 3$ and $5 \times 5$ respectively. + +However, this conventional approach has two critical limitations. First, approximately half of the pixels within the re + +![](images/8e5132bb3c844e09da05de9eae0fb9a91501eeb7d9b12b43eab0d053d3b6c6f2.jpg) +Figure 4. Comparison of spatial-wise kernel designs. Left patterns show kernel configurations, while right tables quantify lookup frequencies during output retrieval. + +ceptive field are accessed multiple times during lookup operations, as illustrated in Fig. 4's tables. This redundancy indicates inefficient kernel pattern design. Second, stacking these kernels to create multiple LUTs for expanded receptive fields results in significant 4D LUT storage overhead. + +To address these inefficiencies, we introduce an innovative L-shaped kernel design. During each rotation, the kernel processes two additional pixels (beyond the center pixel) without overlap, ensuring each surrounding pixel contributes exactly once to the output. This design maximizes pixel value utilization within the receptive field while enabling conversion to 3D LUTs instead of 4D LUTs, dramatically reducing storage requirements without sacrificing coverage. + +# 3.5. PCM Plug-in Module + +To demonstrate the versatility of our pairwise channel mixer, we developed PCM as a plug-in module compatible with existing LUT-based methods. The inference of plug-in algorithm is summarized in Alg. 1. This module seamlessly integrates with current architectures, enhancing their ability to capture channel-spatial correlations without requiring structural modifications. Compared to state-of-the-art methods like SPFLUT, the PCM plug-in adds only $12\%$ runtime and $8\%$ storage overhead while delivering over 1dB performance improvement. + +# 4. Experiments + +# 4.1. Implementation details + +We implement DnLUT using the Adam optimizer with a cosine annealing learning rate schedule and mean squared error (MSE) loss function. For Gaussian denoising, we train networks for $2 \times 10^{5}$ iterations with a batch size of 12, increasing to 32 for real-world denoising scenarios. Following established practices[10, 12], we uniformly sample LUTs + +Algorithm 1 Inference of Pairwise Channel Mixer Module +Require: Input image $I$ of size $H\times W\times 3$ +Require: 4D Lookup Table LUT of PCM +Require: Padding size $p$ +1: Initialize output image $O$ of size $H\times W\times 3$ +2: for $r = 0$ to 3 do +3: Rotate image $I$ by $r\times 90$ degrees to get $I_{\mathrm{rot}}$ +4: Pad $I_{\mathrm{rot}}$ with $p$ pixels on each side to get $I_{\mathrm{rot}}^{\prime}$ +5: for $h = 0$ to $H - 1$ do +6: for $w = 0$ to $W - 1$ do +7: for $c = 0$ to 2 do +8: $p_1\gets I_{\mathrm{rot}}'$ [h,w,c] +9: $p_2\gets I_{\mathrm{rot}}'[h,w + 1,c]$ +10: $p_3\gets I_{\mathrm{rot}}'[h,w,(c + 1)\mod 3]$ +11: $p_4\gets I_{\mathrm{rot}}'[h,w + 1,(c + 1)\mod 3]$ +12: index $\leftarrow$ Quantize $(p_1,p_2,p_3,p_4)$ +13: value $\leftarrow$ LUT[index] +14: $O[h,w,c]\gets O[h,w,c] + \mathrm{Interpolate}(\mathrm{value})$ +15: end for +16: end for +17: end for +18: Rotate $O$ back by $(4 - r)\times 90$ degrees to get $O_{\mathrm{rot}}$ +19: $O\gets O + O_{\mathrm{rot}}$ +20: end for +21: $O\gets O / 4$ +22: return Output image $O$ + +at $2^{4}$ intervals. To minimize indexing artifacts, final predictions utilize 4D simplex interpolation and 3D tetrahedral interpolation. We further refine cached LUTs through 2000 iterations of LUT-aware fine-tuning[12] on the training dataset. + +# 4.2. Gaussian Color Image denoising + +Following [27], we train our model using four comprehensive datasets: DIV2K[1], Flickr2K[23], BSD400[20], and WaterlooED[18]. Training data is augmented with additive Gaussian noise at three levels ( $\sigma \in \{15, 25, 50\}$ ). We evaluate performance on four standard benchmarks: CBSD68[21], Kodak24, Urban100[9], and McMaster[31], using identical noise levels. + +# 4.3. Real-world Color Image denoising + +To validate real-world performance, we train DnLUT on the SIDD training dataset and evaluate on both SIDD validation and DnD datasets. The SIDD dataset comprises indoor scenes captured by five smartphones under varying noise conditions, providing 320 image pairs for training and 1,280 for validation. The DnD dataset offers 50 high-resolution images with realistic noise patterns for comprehensive evaluation. + +Table 2. Quantitative comparison on Gaussian color image denoising benchmark. The table presents CPSNR(dB) values. + +
CategoryMethodCBSD68Kodak24Urban100McMaster
σ = 15σ = 25σ = 50σ = 15σ = 25σ = 50σ = 15σ = 25σ = 50σ = 15σ = 25σ = 50
LUT-basedSR-LUT[10]29.7626.7122.4130.3527.1622.6529.3826.0421.6031.1828.0123.35
MuLUT[12]30.5228.1124.8531.3129.0225.2830.2527.6723.7532.2829.8826.36
RC-LUT[16]30.6828.1225.0431.5729.0725.8930.3327.8023.8632.5129.8926.50
SPF-LUT[14]30.9728.5625.3331.8629.5826.2630.8928.2624.2231.7730.4426.91
DnLUT(Ours)32.4129.8826.0333.0230.2426.7432.1228.8725.0132.8830.4427.12
ClassicalCBM3D[3]33.5230.7127.3834.2831.6829.0233.9331.3627.9334.0631.6628.51
MC-WNNM[24]31.9829.3226.9833.2330.8928.6730.2329.2327.0031.2330.2027.55
DNNDnCNN[30]33.9031.2427.9534.6032.1428.9632.9830.8127.5933.4531.5228.62
SwinIR[15]34.4231.7828.5635.3432.8929.7935.6133.2030.2235.1332.9029.82
+ +Table 3. Quantitative comparison on real-world color image denoising. The table presents CPSNR(dB) and SSIM values. Methods on DnD dataset are validated on the online platform. We could only get the PSNR rather than CPSNR. + +
DatasetsMethodSR-LUT[10]MuLUT[12]RC-LUT[16]SPF-LUT[14]DnLUT(Ours)CBM3D[3]MC-WNNM[24]DnCNN[30]
SIDDCPSNR29.3833.2433.8834.9135.4430.1429.4536.45
SSIM0.6340.8300.8390.8650.8750.7020.6820.900
DnDPSNR*33.6935.1135.4336.2236.6733.1232.3437.11
SSIM0.8390.8680.8750.9110.9220.8230.8170.932
+ +![](images/f707c479ed17ad64ec199e63ac3cd3b3afcb8096a21be9db854fb3080697d70c.jpg) +Full Image + +![](images/7ce7a3866174c3984bcbc60f356a7bd9c645755d6017573c10ebde2dc03057bf.jpg) +SRLUT + +![](images/ac8f52598c456798f84da1e6b4a8fe65be8829fb2564be892e591b72d621e0bf.jpg) +MuLUT + +![](images/afea0d92d133941c17b1c9df3b62c374a7aabbd17fd0ae9bfc88f1ba36e6fb23.jpg) +RC-LUT + +![](images/9a0c552aaa5de2bafef5f5d6f3fcb4eda5d13301a842d73143c0567ecccc7f5c.jpg) +SPFLUT + +![](images/f3bee1ef768b24efe502fe8aa97ea3f7f981a36fa5f832801240b19a5944d7b9.jpg) + +![](images/395227161c30e30370e0f83df8d8f5d193cbba1e9a7734cf9464367a3b6d5b25.jpg) +SRLUT + +![](images/95280d8cfef7da473a46e69b090e391a1f2899e83bf6fb54d5f21708b7de773f.jpg) +MuLUT + +![](images/29a41cc7b4521de450807c711d1e8a10eb8cce645468fdb2d4784c279fcb5e41.jpg) +RC-LUT + +![](images/cd165983b48473f3a8e5c50857f723ffbf4bc58e920e41be487f339981236403.jpg) +SPFLUT + +![](images/76eef8a75f28d6b1720a683a7a877bdacb938c1ec497de1509fc17f19b355c4f.jpg) +Input +GT + +![](images/b47e2080ee26174667d7f1bc0af8665b45e7771b438399a8cfbd77bc61323153.jpg) +CBM3D +Figure 6. Qualitative comparison on real-world datasets. + +![](images/3cdf760cefd38d030708448fd15a6d5a672f9d0bd2dca9e784ffabc0c086474c.jpg) +MC-WNNM + +![](images/b0edf591dfa56f6bc933646ab52b1e49f9f6652b6361779fbe325d5a59676021.jpg) +DnLUT(Ours) + +![](images/db9d85892a1f8bd632f1053a07d14f75f79d56e2141e864ccfec3b7b500bc48e.jpg) +DnCNN + +Table 4. PCM can be easily incorporated to the existing LUT-based methods. It equips the model with the ability of capturing the channel correlations which brings more than 1dB on the widely used benchmark. The noise level $\sigma$ of Gaussian noise is set to 25. The table presents CPSNR(dB) values. + +
MethodCBSD68Kodak24Urban100McMasterSIDD
SR-LUT[10]26.7127.1626.0428.0129.38
PCM+SR-LUT28.04(+1.33)28.55(+1.39)27.33(+1.29)28.78(+0.77)32.33(+2.96)
MuLUT[12]28.1129.0127.6729.8833.24
PCM+MuLUT29.17(+1.05)29.92(+0.91)28.74(+1.07)30.33(+0.44)34.33(+1.09)
RC-LUT[16]28.1229.0727.8029.8933.88
PCM+RC-LUT29.22(+1.10)30.09(+1.02)28.77(+0.97)30.54(+0.65)34.67(+0.79)
SPF-LUT[14]28.5629.5828.2630.4434.91
PCM+SPF-LUT29.72(+1.16)30.63(+1.05)29.43(+1.17)30.96(+0.52)35.56(+0.65)
+ +# 4.4. Quantitative results + +We evaluate performance using color peak signal-to-noise ratio(CPSNR) and structural similarity index(SSIM) metrics, comparing against state-of-the-art methods across three categories: LUT-based methods (SR-LUT[10], MuLUT[12], RC-LUT[16], SPF-LUT[14]), classical approaches (CBM3D[3], MC-WNNM[24]), and DNN-based solutions (DnCNN[30], SwinIR[15]). + +As shown in Tab. 2, DnLUT achieves superior performance in the LUT-based category, surpassing SPF-LUT by over 1dB in Gaussian noise removal while requiring only $17\%$ of its storage. Results from Tab. 3 demonstrate DnLUT's effectiveness in real-world scenarios, outperforming classical methods CBM3D[3] and MC-WNNM[24] by approximately 5dB. Though the Gaussian noise is independent across RGB channels, the network can still leverage inter-channel natural distribution in images to suppress the noise. For example, if a pixel in the R channel has a high value due to noise, but the corresponding pixels in the G and B channels do not, the network can infer that the high value is likely noise and suppress it. + +# 4.5. Qualitative results + +Visual comparisons in Figs. 5 & 6 illustrate DnLUT's superior denoising capabilities. Existing LUT-based methods exhibit noticeable color distortion, particularly evident in Fig. 5's first row. In contrast, DnLUT achieves visual quality comparable to computation-intensive methods like DnCNN and CBM3D. Fig. 6 further demonstrates DnLUT's advantage in real-world scenarios, where both LUT-based and classical methods struggle with color fidelity and often over-smooth details. DnLUT successfully preserves color consistency while retaining fine image details. + +# 4.6. Efficiency Evaluation + +We conduct comprehensive efficiency analysis across three key metrics: theoretical energy cost, runtime performance, and storage requirements (Tab. 5). + +Energy cost. Following AdderSR[22], we analyze multipli + +![](images/70d838580f39f91a819f746cb018318068c63b319adaa17c2727446b67ed393d.jpg) +SRLUT + +![](images/ed0608af79df9df6536b9813cd27dfbf9b06672f33a8aa0a15f453c12e3083a7.jpg) +MuLUT + +![](images/bc48ddfe7c6c2b5550cccd38393f187a403c0f59d33c08e1e2aef246cf16befa.jpg) +RC-LUT + +![](images/a2d9b749cd909fa61fd0ed0e1ce2ab506c4effc0835f7c584d97f03badd6c1cc.jpg) +SPFLUT + +![](images/67db694903958bfe9172aecb7c267f2a07875508d01ab1899d3f9de0ce9229c6.jpg) +PCM+SRLUT +Figure 7. Visualization of LUT-based method with plug-in PCM. + +![](images/9b2e5b66c65bd12d5768fd1b84def76c75eaddd3c5e81881a207da1580432ec8.jpg) +PCM+MuLUT + +![](images/5b7930fff27d398d728ed25e3bb03839620738acc8f7b4604f24acc11d4d4c17.jpg) +PCM+RC-LUT + +![](images/82968afa1650114ee52071db974244db9d6b376595baf1b7ced031d047b75821.jpg) +PCM+SPFLUT + +cation and addition operations across different data types to estimate total energy consumption. DnLUT achieves $70\%$ energy savings compared to SPF-LUT while delivering superior denoising quality. The efficiency gap widens dramatically when compared to DNN-based methods, with DnLUT requiring only $0.1\%$ of DnCNN's energy consumption. + +Runtime. We implement LUT-based methods using standard JAVA API on the ANDROID platform. DnLUT maintains the rapid inference capabilities characteristic of SR-LUT[10], demonstrating clear advantages over classical and DNN-based approaches. For reference, DnCNN running on mobile CPU requires $20 \times$ more processing time. DnLUT's performance can be further optimized through device-specific implementations, such as FPGA deployments. + +Storage. DnLUT requires approximately 500 KB storage, well within typical L2-cache limits of modern chips. This efficient memory footprint enables rapid lookup operations with minimal dependency overhead. + +# 4.7. Effectiveness of Plug-in PCM + +Tab. 4 demonstrates PCM's versatility as a performance enhancer for existing LUT-based methods. When integrated with current approaches, PCM consistently delivers significant improvements across all benchmark datasets. The visual comparisons in Fig. 7 show marked reduction in color distortion, particularly in smooth image regions, resulting in superior perceptual quality. + +Table 5. Efficiency evaluation of three categories color image denoising methods. We use the Qualcomm Snapdragon 8 Gen 2 as the mobile platform. The energy cost is calculated on the ${512} \times {512}$ resolution color images. + +
CategoryMethodPlatformRuntime(ms) 256 × 256Runtime(ms) 512 × 512Energy Cost (pJ)Storage (KB)
LUT-basedSR-LUT[10]Mobile619149.98M82
MuLUT[12]Mobile2180899.88M489
RC-LUT[6]Mobile1878612.92M326
SPF-LUT[14]Mobile722572.32G30,178
DnLUT(ours)Mobile2388687.34M518
ClassicalCBM3D[29]PC3,18914,3434.82G-
MC-WNNM[24]PC74,290293,42489.23G-
DNNDnCNN[30]Mobile5431,923542.53G2,239
SwinIR[15]Mobile73,234274,45312.03T45,499
+ +# 4.8. Ablation study + +We conduct ablation studies using Gaussian color image denoising benchmark to validate DnLUT's key components. + +# 4.8.1. Channel-spatial-wise convolution + +Channel indexing[13, 25] possesses a $1 \times 1$ receptive field to extract channel information. We evaluate various combinations of channel-wise, spatial-wise, and our proposed channel-spatial-wise kernels using a consistent architecture comprising one channel indexing module, two L-shaped convolution modules, and one PCM component. + +Results in Tab. 6 reveal that naive combination of channel-wise and spatial-wise convolutions yields suboptimal performance. Pure channel-spatial-wise convolution, while promising, struggles with limited receptive field coverage. The integration of spatial-wise and channel-spatial-wise convolutions produces substantial improvements. Notably, adding channel-wise convolution offers minimal additional benefit, as channel-spatial-wise convolution already captures comprehensive color information. + +# 4.8.2. L-shaped convolution + +To validate our L-shaped convolution design, we compare it against standard $2 \times 2$ spatial-wise kernels (denoted as S in Tab. 7, $\sigma = 1$ ). While the conventional kernel shows marginal performance advantages, our L-shaped design achieves comparable results while enabling 3D LUT conversion, reducing storage requirements by $17 \times$ . It paves the way of deploying multi-layer spatial-wise convolutions. + +# 5. Conclusion + +This paper presents DnLUT, a novel approach to color image denoising that bridges the gap between high-performance deep learning models and resource-constrained edge devices. Addressing the limitations of existing LUT-based methods, DnLUT introduces two key innovations. First, the Pairwise Channel Mixer (PCM) models channel and spa + +Table 6. Different combinations of channel-wise, spatial-wise, and our channel-spatial-wise kernels. These kernels are represented by channel indexing[13], L-shaped kernel, and PCM. The table presents CPSNR(dB) values. + +
ChannelSpatialChannel-spatialCBSD68
✓29.77
✓30.15
✓✓30.21
✓✓31.41
✓✓✓31.47
+ +Table 7. Comparison of L-shaped kernel and $2 \times 2$ spatial-wise kernel. The table presents CPSNR(dB) values. + +
MethodCBSD68KodakMcMasterLUT Size(KB)
PCM + L26.6327.0527.96494 + 24
PCM + S26.7127.1628.01494 + 408
+ +tial correlations simultaneously by processing channel pairs, enabling comprehensive color modeling without excessive storage costs. Second, an L-shaped convolution design rethinks rotation-based receptive field expansion, eliminating redundant pixel usage and converting 4D to 3D LUTs, reducing storage by $17\times$ while maintaining spatial coverage. These advancements allow DnLUT to outperform state-ofthe-art LUT-based methods by over 1dB in denoising quality while consuming just $0.1\%$ of DnCNN's energy. + +Acknowledgements. This research was supported by: the HKU-TCL Joint Research Centre for AI, the Theme-based Research Scheme (TRS) project T45-701/22-R, and the General Research Fund (GRF) Project 17203224 from the Research Grants Council (RGC) of Hong Kong SAR. Additional support was provided by the National Key Research and Development Program of China (Grant No. 2024YFB2808903), the Shenzhen Science and Technology Program (JCYJ20220818101001004), and the Tsinghua University - Tencent Joint Laboratory for Internet Innovation Technology Research Fund. + +# References + +[1] Eirikur Agustsson and Radu Timofte. Ntire 2017 challenge on single image super-resolution: Dataset and study. In The IEEE Conference on Computer Vision and Pattern Recognition (CVPR) Workshops, 2017. 5 +[2] Hanting Chen, Yunhe Wang, Tianyu Guo, Chang Xu, Yiping Deng, Zhenhua Liu, Siwei Ma, Chunjing Xu, Chao Xu, and Wen Gao. Pre-trained image processing transformer. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 12299-12310, 2021. 1, 2 +[3] Kostadin Dabov, Alessandro Foi, Vladimir Katkovnik, and Karen Egiazarian. Color image denoising via sparse 3d collaborative filtering with grouping constraint in luminance-chrominance space. In 2007 IEEE international conference on image processing, pages I-313. IEEE, 2007. 2, 6, 7 +[4] Kostadin Dabov, Alessandro Foi, Vladimir Katkovnik, and Karen Egiazarian. Image denoising by sparse 3-d transform-domain collaborative filtering. IEEE Transactions on image processing, 16(8):2080-2095, 2007. 2 +[5] Jingjing Dai, Oscar C Au, Lu Fang, Chao Pang, Feng Zou, and Jiali Li. Multichannel nonlocal means fusion for color image denoising. IEEE Transactions on circuits and systems for video technology, 23(11):1873-1886, 2013. 2 +[6] Michael Gharbi, Jiawen Chen, Jonathan T Barron, Samuel W Hasinoff, and Fredo Durand. Deep bilateral learning for real-time image enhancement. ACM Transactions on Graphics (TOG), 36(4):1-12, 2017. 8 +[7] Shuhang Gu, Lei Zhang, Wangmeng Zuo, and Xiangchu Feng. Weighted nuclear norm minimization with application to image denoising. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 2862-2869, 2014. 2 +[8] Binxiao Huang, Jason Chun Lok Li, Jie Ran, Boyu Li, Jiajun Zhou, Dahai Yu, and Ngai Wong. Hundred-kilobyte lookup tables for efficient single-image super-resolution. arXiv preprint arXiv:2312.06101, 2023. 3 +[9] Jia-Bin Huang, Abhishek Singh, and Narendra Ahuja. Single image super-resolution from transformed self-exemplars. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 5197-5206, 2015. 5 +[10] Younghyun Jo and Seon Joo Kim. Practical single-image super-resolution using look-up table. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 691-700, 2021. 2, 3, 5, 6, 7, 8 +[11] Stamatios Lefkimmiatis. Non-local color image denoising with convolutional neural networks. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 3587-3596, 2017. 2 +[12] Jiacheng Li, Chang Chen, Zhen Cheng, and Zhiwei Xiong. Mulut: Cooperating multiple look-up tables for efficient image super-resolution. In European Conference on Computer Vision, pages 238-256. Springer, 2022. 2, 3, 5, 6, 7, 8 +[13] Jiacheng Li, Chang Chen, Zhen Cheng, and Zhiwei Xiong. Toward dnn of luts: Learning efficient image restoration with multiple look-up tables. IEEE Transactions on Pattern Analysis and Machine Intelligence, 2024. 2, 8 + +[14] Yinglong Li, Jiacheng Li, and Zhiwei Xiong. Look-up table compression for efficient image restoration. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 26016-26025, 2024. 2, 3, 6, 7, 8 +[15] Jingyun Liang, Jiezhang Cao, Guolei Sun, Kai Zhang, Luc Van Gool, and Radu Timofte. Swinir: Image restoration using swin transformer. In Proceedings of the IEEE/CVF international conference on computer vision, pages 1833-1844, 2021. 1, 2, 6, 7, 8 +[16] Guandu Liu, Yukang Ding, Mading Li, Ming Sun, Xing Wen, and Bin Wang. Reconstructed convolution module based look-up tables for efficient image super-resolution. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 12217-12226, 2023. 2, 3, 6, 7 +[17] Cheng Ma, Jingyi Zhang, Jie Zhou, and Jiwen Lu. Learning series-parallel lookup tables for efficient image superresolution. In European Conference on Computer Vision, pages 305–321. Springer, 2022. 3 +[18] Kede Ma, Zhengfang Duanmu, Qingbo Wu, Zhou Wang, Hongwei Yong, Hongliang Li, and Lei Zhang. Waterloo exploration database: New challenges for image quality assessment models. IEEE Transactions on Image Processing, 26 (2):1004-1016, 2016. 5 +[19] Julien Mairal, Michael Elad, and Guillermo Sapiro. Sparse representation for color image restoration. IEEE Transactions on image processing, 17(1):53-69, 2007. 2 +[20] D. Martin, C. Fowlkes, D. Tal, and J. Malik. A database of human segmented natural images and its application to evaluating segmentation algorithms and measuring ecological statistics. In Proc. 8th Int'l Conf. Computer Vision, pages 416-423, 2001. 5 +[21] David Martin, Charless Fowlkes, Doron Tal, and Jitendra Malik. A database of human segmented natural images and its application to evaluating segmentation algorithms and measuring ecological statistics. In Proceedings eighth IEEE international conference on computer vision. ICCV 2001, pages 416-423. IEEE, 2001. 5 +[22] Dehua Song, Yunhe Wang, Hanting Chen, Chang Xu, Chunjing Xu, and DaCheng Tao. Addersr: Towards energy efficient image super-resolution. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 15648-15657, 2021. 7 +[23] Radu Timofte, Eirikur Agustsson, Luc Van Gool, Ming-Hsuan Yang, and Lei Zhang. Ntire 2017 challenge on single image super-resolution: Methods and results. In Proceedings of the IEEE conference on computer vision and pattern recognition workshops, pages 114-125, 2017. 5 +[24] Jun Xu, Lei Zhang, David Zhang, and Xiangchu Feng. Multichannel weighted nuclear norm minimization for real color image denoising. In Proceedings of the IEEE international conference on computer vision, pages 1096-1104, 2017. 2, 6, 7, 8 +[25] Sidi Yang, Binxiao Huang, Mingdeng Cao, Yatai Ji, Hanzhong Guo, Ngai Wong, and Yujiu Yang. Taming lookup tables for efficient image retouching. arXiv preprint arXiv:2403.19238, 2024. 2, 8 +[26] Syed Waqas Zamir, Aditya Arora, Salman Khan, Munawar Hayat, Fahad Shahbaz Khan, Ming-Hsuan Yang, and Ling + +Shao. Multi-stage progressive image restoration. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 14821-14831, 2021. 2 +[27] Syed Waqas Zamir, Aditya Arora, Salman Khan, Munawar Hayat, Fahad Shahbaz Khan, and Ming-Hsuan Yang. Restormer: Efficient transformer for high-resolution image restoration. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 5728-5739, 2022. 1, 2, 5 +[28] Hui Zeng, Jianrui Cai, Lida Li, Zisheng Cao, and Lei Zhang. Learning image-adaptive 3d lookup tables for high performance photo enhancement in real-time. IEEE Transactions on Pattern Analysis and Machine Intelligence, 44(4):2058-2073, 2020. 2 +[29] Fengyi Zhang, Hui Zeng, Tianjun Zhang, and Lin Zhang. Clut-net: Learning adaptively compressed representations of 3dluts for lightweight image enhancement. In Proceedings of the 30th ACM International Conference on Multimedia, pages 6493-6501, 2022. 8 +[30] Kai Zhang, Wangmeng Zuo, Yunjin Chen, Deyu Meng, and Lei Zhang. Beyond a gaussian denoiser: Residual learning of deep cnn for image denoising. IEEE transactions on image processing, 26(7):3142-3155, 2017. 1, 2, 6, 7, 8 +[31] Lei Zhang, Xiaolin Wu, Antoni Buades, and Xin Li. Color demosaicking by local directional interpolation and nonlocal adaptive thresholding. Journal of Electronic imaging, 20(2): 023016-023016, 2011. 5 +[32] Yi Zhang, Dasong Li, Xiaoyu Shi, Dailan He, Kangning Song, Xiaogang Wang, Hongwei Qin, and Hongsheng Li. Kbnet: Kernel basis network for image restoration. arXiv preprint arXiv:2303.02881, 2023. 1, 2 \ No newline at end of file diff --git a/2025/DnLUT_ Ultra-Efficient Color Image Denoising via Channel-Aware Lookup Tables/images.zip b/2025/DnLUT_ Ultra-Efficient Color Image Denoising via Channel-Aware Lookup Tables/images.zip new file mode 100644 index 0000000000000000000000000000000000000000..032ccbfc14800ed7d87955ff637576c70c39b830 --- /dev/null +++ b/2025/DnLUT_ Ultra-Efficient Color Image Denoising via Channel-Aware Lookup Tables/images.zip @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:fcc24a62eda90f008d5132897e912c51b42de58015169a8d172f9775e5b473cf +size 660591 diff --git a/2025/DnLUT_ Ultra-Efficient Color Image Denoising via Channel-Aware Lookup Tables/layout.json b/2025/DnLUT_ Ultra-Efficient Color Image Denoising via Channel-Aware Lookup Tables/layout.json new file mode 100644 index 0000000000000000000000000000000000000000..c67c9e76e50f60e2438308f959fc55de73c9e389 --- /dev/null +++ b/2025/DnLUT_ Ultra-Efficient Color Image Denoising via Channel-Aware Lookup Tables/layout.json @@ -0,0 +1,9509 @@ +{ + "pdf_info": [ + { + "para_blocks": [ + { + "bbox": [ + 73, + 102, + 538, + 138 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 73, + 102, + 538, + 138 + ], + "spans": [ + { + "bbox": [ + 73, + 102, + 538, + 138 + ], + "type": "text", + "content": "DnLUT: Ultra-Efficient Color Image Denoising via Channel-Aware Lookup Tables" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 79, + 159, + 531, + 176 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 79, + 159, + 531, + 176 + ], + "spans": [ + { + "bbox": [ + 79, + 159, + 531, + 176 + ], + "type": "text", + "content": "Sidi Yang" + }, + { + "bbox": [ + 79, + 159, + 531, + 176 + ], + "type": "inline_equation", + "content": "^{1}" + }, + { + "bbox": [ + 79, + 159, + 531, + 176 + ], + "type": "text", + "content": " Binxiao Huang" + }, + { + "bbox": [ + 79, + 159, + 531, + 176 + ], + "type": "inline_equation", + "content": "^{1\\dagger}" + }, + { + "bbox": [ + 79, + 159, + 531, + 176 + ], + "type": "text", + "content": " Yulun Zhang" + }, + { + "bbox": [ + 79, + 159, + 531, + 176 + ], + "type": "inline_equation", + "content": "^{2}" + }, + { + "bbox": [ + 79, + 159, + 531, + 176 + ], + "type": "text", + "content": " Dahai Yu" + }, + { + "bbox": [ + 79, + 159, + 531, + 176 + ], + "type": "inline_equation", + "content": "^{3}" + }, + { + "bbox": [ + 79, + 159, + 531, + 176 + ], + "type": "text", + "content": " Yujiu Yang" + }, + { + "bbox": [ + 79, + 159, + 531, + 176 + ], + "type": "inline_equation", + "content": "^{4\\dagger}" + }, + { + "bbox": [ + 79, + 159, + 531, + 176 + ], + "type": "text", + "content": " Ngai Wong" + }, + { + "bbox": [ + 79, + 159, + 531, + 176 + ], + "type": "inline_equation", + "content": "^{1}" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 148, + 175, + 462, + 203 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 148, + 175, + 462, + 203 + ], + "spans": [ + { + "bbox": [ + 148, + 175, + 462, + 203 + ], + "type": "text", + "content": "1 The University of Hong Kong 2 Shanghai Jiaotong University 3 TCL Corporate Research 4 Tsinghua University" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 151, + 205, + 459, + 217 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 151, + 205, + 459, + 217 + ], + "spans": [ + { + "bbox": [ + 151, + 205, + 459, + 217 + ], + "type": "text", + "content": "{yangsidi99, huanghx7}@connect.hku.hk, yulun100@gmail.com, dahai.yu@tcl.com" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 209, + 220, + 400, + 232 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 209, + 220, + 400, + 232 + ], + "spans": [ + { + "bbox": [ + 209, + 220, + 400, + 232 + ], + "type": "text", + "content": "yang.yujiu@sz.tsinghua.edu.cn, nwong@eee.hku.hk" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 151, + 258, + 200, + 272 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 151, + 258, + 200, + 272 + ], + "spans": [ + { + "bbox": [ + 151, + 258, + 200, + 272 + ], + "type": "text", + "content": "Abstract" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 55, + 283, + 297, + 535 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 283, + 297, + 535 + ], + "spans": [ + { + "bbox": [ + 55, + 283, + 297, + 535 + ], + "type": "text", + "content": "While deep neural networks have revolutionized image denoising capabilities, their deployment on edge devices remains challenging due to substantial computational and memory requirements. To this end, we present DnLUT, an ultra-efficient lookup table-based framework that achieves high-quality color image denoising with minimal resource consumption. Our key innovation lies in two complementary components: a Pairwise Channel Mixer (PCM) that effectively captures inter-channel correlations and spatial dependencies in parallel, and a novel L-shaped convolution design that maximizes receptive field coverage while minimizing storage overhead. By converting these components into optimized lookup tables post-training, DnLUT achieves remarkable efficiency - requiring only 500KB storage and " + }, + { + "bbox": [ + 55, + 283, + 297, + 535 + ], + "type": "inline_equation", + "content": "0.1\\%" + }, + { + "bbox": [ + 55, + 283, + 297, + 535 + ], + "type": "text", + "content": " energy consumption compared to its CNN contestant DnCNN, while delivering " + }, + { + "bbox": [ + 55, + 283, + 297, + 535 + ], + "type": "inline_equation", + "content": "20\\times" + }, + { + "bbox": [ + 55, + 283, + 297, + 535 + ], + "type": "text", + "content": " faster inference. Extensive experiments demonstrate that DnLUT outperforms all existing LUT-based methods by over 1dB in PSNR, establishing a new state-of-the-art in resource-efficient color image denoising. The project is available at https://github.com/Stephen0808/DnLUT." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 55, + 556, + 137, + 569 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 556, + 137, + 569 + ], + "spans": [ + { + "bbox": [ + 55, + 556, + 137, + 569 + ], + "type": "text", + "content": "1. Introduction" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 55, + 576, + 295, + 671 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 576, + 295, + 671 + ], + "spans": [ + { + "bbox": [ + 55, + 576, + 295, + 671 + ], + "type": "text", + "content": "Color images pervade our digital world - from social media and photography to scientific imaging and medical diagnostics. These images are inherently more vulnerable to noise contamination than their grayscale counterparts, not only due to their threefold data size during acquisition, storage, and transmission, but also because human perception exhibits heightened sensitivity to color distortions compared to luminance variations." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 55, + 672, + 295, + 696 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 672, + 295, + 696 + ], + "spans": [ + { + "bbox": [ + 55, + 672, + 295, + 696 + ], + "type": "text", + "content": "The advent of deep neural networks (DNNs) has dramatically advanced the field of color image denoising[2, 15," + } + ] + } + ], + "index": 11 + }, + { + "type": "image", + "bbox": [ + 317, + 258, + 552, + 401 + ], + "blocks": [ + { + "bbox": [ + 317, + 258, + 552, + 401 + ], + "lines": [ + { + "bbox": [ + 317, + 258, + 552, + 401 + ], + "spans": [ + { + "bbox": [ + 317, + 258, + 552, + 401 + ], + "type": "image", + "image_path": "df4d233aaf8694d8dcbde46b770703422f6b3e312c780d0c901754af163409e1.jpg" + } + ] + } + ], + "index": 12, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 313, + 408, + 555, + 498 + ], + "lines": [ + { + "bbox": [ + 313, + 408, + 555, + 498 + ], + "spans": [ + { + "bbox": [ + 313, + 408, + 555, + 498 + ], + "type": "text", + "content": "Figure 1. Model comparison in terms of color peak signal-to-noise ratio (CPSNR), runtime, and storage. The CPSNR and runtime are calculated on the CBSD68 dataset with Gaussian noise level " + }, + { + "bbox": [ + 313, + 408, + 555, + 498 + ], + "type": "inline_equation", + "content": "\\sigma = 25" + }, + { + "bbox": [ + 313, + 408, + 555, + 498 + ], + "type": "text", + "content": " using Qualcomm Snapdragon 8 Gen2. Our method outperforms state-of-the-art LUT-based methods with the highest CPSNR, at a low storage requirements and reduced runtime. Additionally, our PCM module serves as a versatile plug-in module that enhances existing methods' performance by over 1dB." + } + ] + } + ], + "index": 13, + "angle": 0, + "type": "image_caption" + } + ], + "index": 12 + }, + { + "bbox": [ + 313, + 520, + 555, + 652 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 520, + 555, + 652 + ], + "spans": [ + { + "bbox": [ + 313, + 520, + 555, + 652 + ], + "type": "text", + "content": "27, 30, 32]. State-of-the-art models leverage sophisticated architectures with numerous convolutional layers or transformer blocks, achieving unprecedented denoising quality. However, these advances come at a steep cost - the computational complexity and memory requirements of such models render them impractical for edge devices. This limitation is particularly acute given that most edge devices lack specialized hardware accelerators such as graphics processing units (GPs) or tensor processing units (TPUs), creating a significant gap between algorithmic capabilities and practical deployability." + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 313, + 653, + 556, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 653, + 556, + 715 + ], + "spans": [ + { + "bbox": [ + 313, + 653, + 556, + 715 + ], + "type": "text", + "content": "Recent works have explored lookup tables (LUTs) as an elegant solution for efficient image enhancement. LUTs replace complex runtime computations with simple array indexing operations through direct addressing. This approach typically involves training a neural network model, then con" + } + ] + } + ], + "index": 15 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "spans": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "type": "text", + "content": "CVF" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 145, + 0, + 496, + 37 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 145, + 0, + 496, + 37 + ], + "spans": [ + { + "bbox": [ + 145, + 0, + 496, + 37 + ], + "type": "text", + "content": "This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 66, + 702, + 150, + 713 + ], + "type": "page_footnote", + "angle": 0, + "lines": [ + { + "bbox": [ + 66, + 702, + 150, + 713 + ], + "spans": [ + { + "bbox": [ + 66, + 702, + 150, + 713 + ], + "type": "text", + "content": "† Corresponding authors." + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 295, + 748, + 315, + 758 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 315, + 758 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 315, + 758 + ], + "type": "text", + "content": "7582" + } + ] + } + ], + "index": 17 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 0 + }, + { + "para_blocks": [ + { + "bbox": [ + 55, + 72, + 294, + 179 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 72, + 294, + 179 + ], + "spans": [ + { + "bbox": [ + 55, + 72, + 294, + 179 + ], + "type": "text", + "content": "verting it into LUTs by exhaustively mapping input-output relationships. During inference, the system performs rapid table lookups instead of costly computations, effectively leveraging DNN capabilities while circumventing hardware constraints. However, LUT-based methods face a fundamental challenge: the storage requirement grows exponentially with the input dimensionality. This has led to the adoption of 4D LUTs (using four pixel values for indexing) as a practical compromise between storage and performance." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 55, + 181, + 295, + 348 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 181, + 295, + 348 + ], + "spans": [ + { + "bbox": [ + 55, + 181, + 295, + 348 + ], + "type": "text", + "content": "Current LUT-based methods predominantly employ " + }, + { + "bbox": [ + 55, + 181, + 295, + 348 + ], + "type": "inline_equation", + "content": "2 \\times 2" + }, + { + "bbox": [ + 55, + 181, + 295, + 348 + ], + "type": "text", + "content": " convolutional kernels processed independently for each channel[10, 12, 14, 16], creating spatial-wise LUTs for single-channel processing. While this approach works well for tasks like super-resolution, it proves inadequate for color image denoising where noise typically affects all RGB channels simultaneously. Channel correlation plays a pivotal role in modeling noise characteristics, a critical factor for achieving effective restoration. Alternative approaches either use " + }, + { + "bbox": [ + 55, + 181, + 295, + 348 + ], + "type": "inline_equation", + "content": "1 \\times 1" + }, + { + "bbox": [ + 55, + 181, + 295, + 348 + ], + "type": "text", + "content": " convolutions with three-channel depth for RGB information (creating 3D channel-wise LUTs)[13, 25] or attempt to use larger " + }, + { + "bbox": [ + 55, + 181, + 295, + 348 + ], + "type": "inline_equation", + "content": "H \\times W \\times 3" + }, + { + "bbox": [ + 55, + 181, + 295, + 348 + ], + "type": "text", + "content": " kernels. However, these solutions either disrupt noise distribution patterns or lead to prohibitive storage requirements." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 55, + 350, + 295, + 397 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 350, + 295, + 397 + ], + "spans": [ + { + "bbox": [ + 55, + 350, + 295, + 397 + ], + "type": "text", + "content": "To address these challenges, we propose DnLUT, a novel lookup table-based framework specifically designed for efficient color image denoising. Our approach centers on two key innovations:" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 55, + 399, + 295, + 518 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 399, + 295, + 518 + ], + "spans": [ + { + "bbox": [ + 55, + 399, + 295, + 518 + ], + "type": "text", + "content": "First, we introduce the Pairwise Channel Mixer (PCM), which simultaneously processes spatial and channel information. PCM strategically combines RGB channels into three pairs (RG, GB, BR), processing them through parallel branches using a " + }, + { + "bbox": [ + 55, + 399, + 295, + 518 + ], + "type": "inline_equation", + "content": "1 \\times 2" + }, + { + "bbox": [ + 55, + 399, + 295, + 518 + ], + "type": "text", + "content": " convolution kernel with depth 2. This design efficiently captures channel correlations while maintaining manageable storage requirements. The PCM module can also serve as a plug-in module for existing LUT-based methods, consistently improving their performance by approximately 1dB." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 55, + 520, + 295, + 662 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 520, + 295, + 662 + ], + "spans": [ + { + "bbox": [ + 55, + 520, + 295, + 662 + ], + "type": "text", + "content": "Second, we develop an innovative L-shaped convolution kernel that addresses the limitations of conventional rotation-based approaches. After taking the channel correlation into account, as discussed in [12, 14], there is still a need to enlarge the spatial receptive field. While previous methods like SR-LUT[10] use rotation ensemble training to expand receptive fields of spatial-wise LUT, they often introduce redundant pixel usage and significant storage overhead. Our L-shaped kernel design eliminates pixel overlap during rotations, enabling conversion to more efficient 3D LUTs while maintaining the same effective receptive field size as 4D LUTs." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 55, + 664, + 295, + 688 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 664, + 295, + 688 + ], + "spans": [ + { + "bbox": [ + 55, + 664, + 295, + 688 + ], + "type": "text", + "content": "Extensive experimental validation demonstrates the superiority of DnLUT over existing approaches:" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 56, + 689, + 295, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 689, + 295, + 713 + ], + "spans": [ + { + "bbox": [ + 56, + 689, + 295, + 713 + ], + "type": "text", + "content": "- We achieve up to 1.3dB CPSNR improvement over SPFLUT[14] on standard Gaussian denoising bench" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 323, + 72, + 553, + 95 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 323, + 72, + 553, + 95 + ], + "spans": [ + { + "bbox": [ + 323, + 72, + 553, + 95 + ], + "type": "text", + "content": "marks. For real-world denoising tasks, we outperform CBM3D[3] by nearly 5dB." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 315, + 96, + 553, + 203 + ], + "type": "list", + "angle": 0, + "index": 11, + "blocks": [ + { + "bbox": [ + 315, + 96, + 553, + 132 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 315, + 96, + 553, + 132 + ], + "spans": [ + { + "bbox": [ + 315, + 96, + 553, + 132 + ], + "type": "text", + "content": "- Our method maintains exceptional efficiency, requiring only " + }, + { + "bbox": [ + 315, + 96, + 553, + 132 + ], + "type": "inline_equation", + "content": "0.1\\%" + }, + { + "bbox": [ + 315, + 96, + 553, + 132 + ], + "type": "text", + "content": " of DnCNN[30]'s energy consumption and 500KB storage." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 315, + 133, + 553, + 167 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 315, + 133, + 553, + 167 + ], + "spans": [ + { + "bbox": [ + 315, + 133, + 553, + 167 + ], + "type": "text", + "content": "- The PCM module serves as a versatile performance enhancer, bringing over 1dB CPSNR improvement to existing LUT-based methods." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 315, + 168, + 553, + 203 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 315, + 168, + 553, + 203 + ], + "spans": [ + { + "bbox": [ + 315, + 168, + 553, + 203 + ], + "type": "text", + "content": "- Our L-shaped convolution design reduces storage requirements by " + }, + { + "bbox": [ + 315, + 168, + 553, + 203 + ], + "type": "inline_equation", + "content": "17 \\times" + }, + { + "bbox": [ + 315, + 168, + 553, + 203 + ], + "type": "text", + "content": " while preserving receptive field coverage." + } + ] + } + ], + "index": 10 + } + ], + "sub_type": "text" + }, + { + "bbox": [ + 313, + 204, + 553, + 251 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 204, + 553, + 251 + ], + "spans": [ + { + "bbox": [ + 313, + 204, + 553, + 251 + ], + "type": "text", + "content": "These results establish DnLUT as a significant advance in practical, resource-efficient color image denoising, bridging the gap between algorithmic capability and deployability on edge devices." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 314, + 265, + 402, + 277 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 314, + 265, + 402, + 277 + ], + "spans": [ + { + "bbox": [ + 314, + 265, + 402, + 277 + ], + "type": "text", + "content": "2. Related works" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 314, + 285, + 443, + 297 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 314, + 285, + 443, + 297 + ], + "spans": [ + { + "bbox": [ + 314, + 285, + 443, + 297 + ], + "type": "text", + "content": "2.1. Color Image denoising" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 313, + 302, + 553, + 374 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 302, + 553, + 374 + ], + "spans": [ + { + "bbox": [ + 313, + 302, + 553, + 374 + ], + "type": "text", + "content": "A fundamental characteristic of color images is the strong correlation among RGB channels[5]. Research has consistently shown that processing these channels independently yields suboptimal results compared to joint processing approaches[19]. This insight has driven numerous innovations in color image denoising." + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 313, + 375, + 553, + 505 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 375, + 553, + 505 + ], + "spans": [ + { + "bbox": [ + 313, + 375, + 553, + 505 + ], + "type": "text", + "content": "A landmark development was CBM3D[3], which operates in luminance-chrominance space (i.e., " + }, + { + "bbox": [ + 313, + 375, + 553, + 505 + ], + "type": "inline_equation", + "content": "YC_{b}C_{r}" + }, + { + "bbox": [ + 313, + 375, + 553, + 505 + ], + "type": "text", + "content": "). The method performs BM3D[4] grouping once on the luminance channel " + }, + { + "bbox": [ + 313, + 375, + 553, + 505 + ], + "type": "inline_equation", + "content": "Y" + }, + { + "bbox": [ + 313, + 375, + 553, + 505 + ], + "type": "text", + "content": " and leverages this grouping for collaborative filtering of chrominance channels " + }, + { + "bbox": [ + 313, + 375, + 553, + 505 + ], + "type": "inline_equation", + "content": "U" + }, + { + "bbox": [ + 313, + 375, + 553, + 505 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 313, + 375, + 553, + 505 + ], + "type": "inline_equation", + "content": "V" + }, + { + "bbox": [ + 313, + 375, + 553, + 505 + ], + "type": "text", + "content": ". Building on this concept, [11] introduced a radial basis function mixture for more sophisticated handling of channel correlations. Another notable approach, MC-WNNM[24], extended WNNM[7] by incorporating a weight matrix to adaptively balance the data fidelity across channels based on their noise characteristics." + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 313, + 507, + 553, + 615 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 507, + 553, + 615 + ], + "spans": [ + { + "bbox": [ + 313, + 507, + 553, + 615 + ], + "type": "text", + "content": "The deep learning era, initiated by DnCNN[30], transformed the field of color image denoising. Subsequent approaches[2, 15, 26, 27, 32] have increasingly sophisticated architectures that process all RGB channels simultaneously through complex networks of convolution layers or transformer blocks. While these models achieve remarkable denoising quality, their substantial computational and memory requirements present significant deployment challenges, particularly for resource-constrained edge devices." + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 314, + 624, + 459, + 636 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 314, + 624, + 459, + 636 + ], + "spans": [ + { + "bbox": [ + 314, + 624, + 459, + 636 + ], + "type": "text", + "content": "2.2. Replacing CNN with LUT" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 313, + 642, + 553, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 642, + 553, + 713 + ], + "spans": [ + { + "bbox": [ + 313, + 642, + 553, + 713 + ], + "type": "text", + "content": "The challenge of deploying deep neural networks on edge devices has sparked interest in lookup table (LUT) based alternatives. Inspired by the prevalence of color LUTs in image signal processors, [28] pioneered an approach using learnable basis LUTs combined with a weight prediction network for image enhancement. [25] further developed this" + } + ] + } + ], + "index": 19 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "text", + "content": "7583" + } + ] + } + ], + "index": 20 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 1 + }, + { + "para_blocks": [ + { + "bbox": [ + 55, + 72, + 294, + 120 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 72, + 294, + 120 + ], + "spans": [ + { + "bbox": [ + 55, + 72, + 294, + 120 + ], + "type": "text", + "content": "concept, proposing a more efficient channel-wise LUT system. However, these early channel-wise LUTs, processing single pixels per channel, proved inadequate for the complex task of denoising." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 55, + 121, + 294, + 240 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 121, + 294, + 240 + ], + "spans": [ + { + "bbox": [ + 55, + 121, + 294, + 240 + ], + "type": "text", + "content": "A significant breakthrough came with [10], which introduced a method to convert CNNs with limited receptive fields into spatial-wise LUTs for super-resolution. This approach trains networks on restricted pixel neighborhoods (i.e., " + }, + { + "bbox": [ + 55, + 121, + 294, + 240 + ], + "type": "inline_equation", + "content": "2 \\times 2" + }, + { + "bbox": [ + 55, + 121, + 294, + 240 + ], + "type": "text", + "content": ") and converts them to LUTs by exhaustively mapping input-output relationships. During inference, the system efficiently retrieves pre-computed outputs through coordinate indexing. Follow-up works[12, 16, 17] enhanced this framework using serial LUTs to expand receptive fields, achieving substantial performance improvements." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 55, + 241, + 294, + 373 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 241, + 294, + 373 + ], + "spans": [ + { + "bbox": [ + 55, + 241, + 294, + 373 + ], + "type": "text", + "content": "To manage the exponential growth of LUT size with input dimensionality, these methods typically process each input channel independently. While this approach proves effective for super-resolution, where spatial information dominates, it falls short for color image denoising, which requires preserving both spatial and channel information to model noise distribution effectively. The limitations of existing channel-wise and spatial-wise LUT approaches highlight the need for a more sophisticated solution that can handle the unique challenges of color image denoising while maintaining computational efficiency." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 55, + 385, + 111, + 398 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 385, + 111, + 398 + ], + "spans": [ + { + "bbox": [ + 55, + 385, + 111, + 398 + ], + "type": "text", + "content": "3. Method" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 55, + 406, + 134, + 419 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 406, + 134, + 419 + ], + "spans": [ + { + "bbox": [ + 55, + 406, + 134, + 419 + ], + "type": "text", + "content": "3.1. Preliminary" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 55, + 425, + 295, + 604 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 425, + 295, + 604 + ], + "spans": [ + { + "bbox": [ + 55, + 425, + 295, + 604 + ], + "type": "text", + "content": "The foundation of LUT-based image restoration was established by SR-LUT[10], which introduces a constrained receptive field approach for network training (i.e., " + }, + { + "bbox": [ + 55, + 425, + 295, + 604 + ], + "type": "inline_equation", + "content": "2 \\times 2" + }, + { + "bbox": [ + 55, + 425, + 295, + 604 + ], + "type": "text", + "content": "). Post-training, the network's output values are systematically cached into a LUT by exhaustively traversing all possible input pixel value combinations. During inference, input patches of size " + }, + { + "bbox": [ + 55, + 425, + 295, + 604 + ], + "type": "inline_equation", + "content": "2 \\times 2" + }, + { + "bbox": [ + 55, + 425, + 295, + 604 + ], + "type": "text", + "content": " are converted to indices of a 4D LUT to retrieve corresponding output values. To manage storage requirements, the 4D index space is uniformly subsampled. Additionally, a rotation ensemble strategy rotates each " + }, + { + "bbox": [ + 55, + 425, + 295, + 604 + ], + "type": "inline_equation", + "content": "2 \\times 2" + }, + { + "bbox": [ + 55, + 425, + 295, + 604 + ], + "type": "text", + "content": " input patch four times, effectively expanding the receptive field to " + }, + { + "bbox": [ + 55, + 425, + 295, + 604 + ], + "type": "inline_equation", + "content": "3 \\times 3" + }, + { + "bbox": [ + 55, + 425, + 295, + 604 + ], + "type": "text", + "content": ". Recent works have focused on increasing the effective receptive field through various techniques, including multiple LUTs[12, 14, 17], shift aggregations[17], and diverse kernel patterns[8, 12, 16]." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 55, + 605, + 295, + 640 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 605, + 295, + 640 + ], + "spans": [ + { + "bbox": [ + 55, + 605, + 295, + 640 + ], + "type": "text", + "content": "Our work advances this foundation by specifically addressing the challenges of LUT-based color image denoising, a critical application for edge devices." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 55, + 642, + 295, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 642, + 295, + 713 + ], + "spans": [ + { + "bbox": [ + 55, + 642, + 295, + 713 + ], + "type": "text", + "content": "A notable limitation of existing LUT-based approaches is their treatment of channel depth, typically set to 1. For color images, this means each RGB channel independently accesses cached output values without inter-channel interaction. This design overlooks crucial RGB channel correlations and spatial-channel-wise degradation patterns inher" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 313, + 72, + 553, + 227 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 72, + 553, + 227 + ], + "spans": [ + { + "bbox": [ + 313, + 72, + 553, + 227 + ], + "type": "text", + "content": "ent in color image denoising. While a straightforward solution might be to employ vanilla convolution kernel patterns, the exponential growth of LUT size with index dimension makes this impractical. For instance, a kernel of size " + }, + { + "bbox": [ + 313, + 72, + 553, + 227 + ], + "type": "inline_equation", + "content": "2 \\times 2" + }, + { + "bbox": [ + 313, + 72, + 553, + 227 + ], + "type": "text", + "content": " with depth 3 would require approximately 582 TB of storage, even with subsampling compression. This constraint presents a fundamental dilemma: when aiming to capture more spatial information, we have to sacrifice the richness of channel information. Moreover, while 4D spatial-wise LUTs show promise, they typically require multiple LUTs (analogous to stacked convolution layers), creating significant storage overhead, particularly problematic for L1 cache access." + } + ] + } + ], + "index": 8 + }, + { + "type": "table", + "bbox": [ + 323, + 269, + 544, + 362 + ], + "blocks": [ + { + "bbox": [ + 313, + 238, + 553, + 259 + ], + "lines": [ + { + "bbox": [ + 313, + 238, + 553, + 259 + ], + "spans": [ + { + "bbox": [ + 313, + 238, + 553, + 259 + ], + "type": "text", + "content": "Table 1. LUT size versus receptive field (RF) and channel numbers. * means subsampled LUT size" + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 323, + 269, + 544, + 362 + ], + "lines": [ + { + "bbox": [ + 323, + 269, + 544, + 362 + ], + "spans": [ + { + "bbox": [ + 323, + 269, + 544, + 362 + ], + "type": "table", + "html": "
RFDepthLUT Dim.LUT Size*
1 × 111D17 B
1 × 133D4.9 KB
2 × 214D83.5 KB
1 × 236D24.1 MB
2 × 2312D582.6 TB
k × kck × k × cD(2^4 + 1)^{k×k×C} B
", + "image_path": "8b88a346c8714544c5f7c259228305ef589a960e73081163aadb01f8142a5726.jpg" + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "table_body" + } + ], + "index": 10 + }, + { + "bbox": [ + 314, + 378, + 382, + 389 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 314, + 378, + 382, + 389 + ], + "spans": [ + { + "bbox": [ + 314, + 378, + 382, + 389 + ], + "type": "text", + "content": "3.2. Overview" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 313, + 396, + 553, + 479 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 396, + 553, + 479 + ], + "spans": [ + { + "bbox": [ + 313, + 396, + 553, + 479 + ], + "type": "text", + "content": "To address these fundamental challenges, we introduce two complementary innovations: 1) a pairwise channel mixer that leverages 4D LUT efficiency while modeling spatial-channel correlations; and 2) a novel L-shaped kernel featuring inherent rotation non-overlapping properties, enabling conversion to 3D LUT while maintaining 4D LUT-equivalent receptive field coverage." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 313, + 481, + 553, + 648 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 481, + 553, + 648 + ], + "spans": [ + { + "bbox": [ + 313, + 481, + 553, + 648 + ], + "type": "text", + "content": "We first build the DnNet for training, as shown in Fig. 2. Our proposed framework enhances traditional spatial-wise LUT capabilities through these two key modules: the pairwise channel mixer (PCM) establishes robust channel-spatial correlations in three RGB channels while the L-shaped kernel expands the effective receptive field with minimal computational overhead independently in each channel. The architecture combines PCM with L-shaped convolution in two stages. The first stage, inspired by [14], generates multi-channel features from LUT groups, which are then concatenated in the fusion module. The second stage integrates both proposed modules. Importantly, all components are designed for efficient conversion to 3D or 4D LUTs during inference, collectively referred to as DnLUT." + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 313, + 659, + 451, + 671 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 659, + 451, + 671 + ], + "spans": [ + { + "bbox": [ + 313, + 659, + 451, + 671 + ], + "type": "text", + "content": "3.3. Pairwise Channel Mixer" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 313, + 677, + 553, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 677, + 553, + 713 + ], + "spans": [ + { + "bbox": [ + 313, + 677, + 553, + 713 + ], + "type": "text", + "content": "Designing effective kernel patterns for LUT-based methods presents significant challenges. Fig. 3 illustrates various kernel pattern categories. Traditional approaches have" + } + ] + } + ], + "index": 15 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "text", + "content": "7584" + } + ] + } + ], + "index": 16 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 2 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 61, + 76, + 549, + 182 + ], + "blocks": [ + { + "bbox": [ + 61, + 76, + 549, + 182 + ], + "lines": [ + { + "bbox": [ + 61, + 76, + 549, + 182 + ], + "spans": [ + { + "bbox": [ + 61, + 76, + 549, + 182 + ], + "type": "image", + "image_path": "dd9fda6a17408baae32f2301a1fc608986e28efd782e4f04f02e5559c180fb15.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + } + ], + "index": 0 + }, + { + "type": "image", + "bbox": [ + 86, + 183, + 233, + 257 + ], + "blocks": [ + { + "bbox": [ + 86, + 183, + 233, + 257 + ], + "lines": [ + { + "bbox": [ + 86, + 183, + 233, + 257 + ], + "spans": [ + { + "bbox": [ + 86, + 183, + 233, + 257 + ], + "type": "image", + "image_path": "03367edfc8631b81ef3fb86a0fad5ce2b921f4a99486e887597f506bec2c5a81.jpg" + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 54, + 265, + 555, + 331 + ], + "lines": [ + { + "bbox": [ + 54, + 265, + 555, + 331 + ], + "spans": [ + { + "bbox": [ + 54, + 265, + 555, + 331 + ], + "type": "text", + "content": "Figure 2. System architecture of DnLUT: (a) The DnNet pipeline integrates pairwise channel mixers and L-shaped convolutions, with multi-scale fusion enhancing receptive field coverage. Channel dimensions are flattened for parallel processing in L-shaped operations, then unfolded for PCM input. (b) Input pixels undergo four rotations " + }, + { + "bbox": [ + 54, + 265, + 555, + 331 + ], + "type": "inline_equation", + "content": "(0^{\\circ}, 90^{\\circ}, 180^{\\circ}, 270^{\\circ})" + }, + { + "bbox": [ + 54, + 265, + 555, + 331 + ], + "type": "text", + "content": " during processing, with outputs averaged for enhanced results. (c) Post-training, all possible input combinations are processed through DnNet modules, with outputs cached in optimized 3D or 4D LUTs. (d) During inference, input pixels are efficiently processed through multiple LUTs, with each LUT's outputs informing subsequent LUT indices, culminating in final denoised pixel values." + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_caption" + } + ], + "index": 1 + }, + { + "type": "image", + "bbox": [ + 246, + 184, + 359, + 257 + ], + "blocks": [ + { + "bbox": [ + 246, + 184, + 359, + 257 + ], + "lines": [ + { + "bbox": [ + 246, + 184, + 359, + 257 + ], + "spans": [ + { + "bbox": [ + 246, + 184, + 359, + 257 + ], + "type": "image", + "image_path": "35f7a845319b0c2c2b47357b2405a7cd8c56526a616dad12d9a111a6b8816e22.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + } + ], + "index": 2 + }, + { + "type": "image", + "bbox": [ + 375, + 186, + 541, + 257 + ], + "blocks": [ + { + "bbox": [ + 375, + 186, + 541, + 257 + ], + "lines": [ + { + "bbox": [ + 375, + 186, + 541, + 257 + ], + "spans": [ + { + "bbox": [ + 375, + 186, + 541, + 257 + ], + "type": "image", + "image_path": "49e0bdd5a3c7f94fb5c891de0651792ff428e90daba707affeb6a3c37541cd27.jpg" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_body" + } + ], + "index": 3 + }, + { + "bbox": [ + 54, + 352, + 294, + 483 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 54, + 352, + 294, + 483 + ], + "spans": [ + { + "bbox": [ + 54, + 352, + 294, + 483 + ], + "type": "text", + "content": "focused on single-dimension kernels due to LUT index dimension constraints, leading to two major limitations in color image denoising. Spatial-wise kernels, while capturing spatial information, neglect channel correlations, resulting in color distortion and suboptimal noise modeling. Conversely, channel-wise kernels establish strong RGB channel connections but ignore spatial relationships, leading to persistent artifacts. While combining these kernels in a cascade network might seem promising, this approach fails to process channel and spatial information simultaneously, resulting in suboptimal noise distribution modeling." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 55, + 484, + 295, + 639 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 484, + 295, + 639 + ], + "spans": [ + { + "bbox": [ + 55, + 484, + 295, + 639 + ], + "type": "text", + "content": "Our PCM addresses these limitations through a novel architecture for color image denoising. The module first re-organizes RGB channels into three pairwise combinations: RG, GB, and BR. These pairs are processed through parallel branches using a kernel with " + }, + { + "bbox": [ + 55, + 484, + 295, + 639 + ], + "type": "inline_equation", + "content": "1 \\times 2" + }, + { + "bbox": [ + 55, + 484, + 295, + 639 + ], + "type": "text", + "content": " spatial dimensions and depth 2 in the initial layer, followed by cascaded " + }, + { + "bbox": [ + 55, + 484, + 295, + 639 + ], + "type": "inline_equation", + "content": "1 \\times 1" + }, + { + "bbox": [ + 55, + 484, + 295, + 639 + ], + "type": "text", + "content": " convolution layers. Each convolution operation processes four pixel values to produce one channel output, enabling efficient conversion to 4D LUTs post-training. The cached LUTs (i.e., " + }, + { + "bbox": [ + 55, + 484, + 295, + 639 + ], + "type": "inline_equation", + "content": "LUT_{RG}, LUT_{GB}, LUT_{GR}" + }, + { + "bbox": [ + 55, + 484, + 295, + 639 + ], + "type": "text", + "content": ") operate in parallel, with their predictions concatenated. For an input anchor " + }, + { + "bbox": [ + 55, + 484, + 295, + 639 + ], + "type": "inline_equation", + "content": "(I_R, I_G, I_B)" + }, + { + "bbox": [ + 55, + 484, + 295, + 639 + ], + "type": "text", + "content": ", the output values " + }, + { + "bbox": [ + 55, + 484, + 295, + 639 + ], + "type": "inline_equation", + "content": "(V_R, V_G, V_B)" + }, + { + "bbox": [ + 55, + 484, + 295, + 639 + ], + "type": "text", + "content": " are computed as:" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 56, + 647, + 294, + 700 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 647, + 294, + 700 + ], + "spans": [ + { + "bbox": [ + 56, + 647, + 294, + 700 + ], + "type": "interline_equation", + "content": "\\begin{array}{r} \\left(V _ {R}, V _ {G}, V _ {B}\\right) = \\operatorname {C a t} \\left(L U T _ {R G} \\left[ I _ {0, 0, R} \\right] \\left[ I _ {0, 1, R} \\right] \\left[ I _ {0, 0, G} \\right] \\left[ I _ {0, 1, G} \\right], \\right. \\\\ L U T _ {G B} \\left[ I _ {0, 0, G} \\right] \\left[ I _ {0, 1, G} \\right] \\left[ I _ {0, 0, B} \\right] \\left[ I _ {0, 1, B} \\right], \\\\ L U T _ {B R} \\left[ I _ {0, 0, B} \\right] \\left[ I _ {0, 1, B} \\right] \\left[ I _ {0, 0, R} \\right] \\left[ I _ {0, 1, R} \\right]), \\end{array} \\tag {1}", + "image_path": "b0aa3743b13aeac3c48ca7ebef142d27e5c955c1df66ec53f6a8a163111f57e0.jpg" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 55, + 701, + 295, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 701, + 295, + 715 + ], + "spans": [ + { + "bbox": [ + 55, + 701, + 295, + 715 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 55, + 701, + 295, + 715 + ], + "type": "inline_equation", + "content": "H, W, C" + }, + { + "bbox": [ + 55, + 701, + 295, + 715 + ], + "type": "text", + "content": " of each " + }, + { + "bbox": [ + 55, + 701, + 295, + 715 + ], + "type": "inline_equation", + "content": "I_{H,W,C}" + }, + { + "bbox": [ + 55, + 701, + 295, + 715 + ], + "type": "text", + "content": " denotes height, width," + } + ] + } + ], + "index": 8 + }, + { + "type": "image", + "bbox": [ + 319, + 350, + 392, + 435 + ], + "blocks": [ + { + "bbox": [ + 319, + 350, + 392, + 435 + ], + "lines": [ + { + "bbox": [ + 319, + 350, + 392, + 435 + ], + "spans": [ + { + "bbox": [ + 319, + 350, + 392, + 435 + ], + "type": "image", + "image_path": "b6f15846ba721620cc417fdac6cdcef8905e19cfa20df9e98c6765d572e6c233.jpg" + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 328, + 436, + 377, + 444 + ], + "lines": [ + { + "bbox": [ + 328, + 436, + 377, + 444 + ], + "spans": [ + { + "bbox": [ + 328, + 436, + 377, + 444 + ], + "type": "text", + "content": "(a) Spatial-wise" + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 399, + 435, + 454, + 444 + ], + "lines": [ + { + "bbox": [ + 399, + 435, + 454, + 444 + ], + "spans": [ + { + "bbox": [ + 399, + 435, + 454, + 444 + ], + "type": "text", + "content": "(b) Channel-wise" + } + ] + } + ], + "index": 12, + "angle": 0, + "type": "image_caption" + } + ], + "index": 9 + }, + { + "type": "image", + "bbox": [ + 395, + 378, + 467, + 408 + ], + "blocks": [ + { + "bbox": [ + 395, + 378, + 467, + 408 + ], + "lines": [ + { + "bbox": [ + 395, + 378, + 467, + 408 + ], + "spans": [ + { + "bbox": [ + 395, + 378, + 467, + 408 + ], + "type": "image", + "image_path": "8c14886eff627301db4bf7cd4752cc4f476dee7819944fe4857ff2c04733b838.jpg" + } + ] + } + ], + "index": 11, + "angle": 0, + "type": "image_body" + } + ], + "index": 11 + }, + { + "type": "image", + "bbox": [ + 471, + 353, + 544, + 430 + ], + "blocks": [ + { + "bbox": [ + 471, + 353, + 544, + 430 + ], + "lines": [ + { + "bbox": [ + 471, + 353, + 544, + 430 + ], + "spans": [ + { + "bbox": [ + 471, + 353, + 544, + 430 + ], + "type": "image", + "image_path": "42c0da747a88fbbf959f6edd2b6d129676d65f6a04f94673969dcaefb27b76b0.jpg" + } + ] + } + ], + "index": 13, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 470, + 435, + 547, + 445 + ], + "lines": [ + { + "bbox": [ + 470, + 435, + 547, + 445 + ], + "spans": [ + { + "bbox": [ + 470, + 435, + 547, + 445 + ], + "type": "text", + "content": "(c) Spatial-Channel-wise" + } + ] + } + ], + "index": 14, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 313, + 454, + 554, + 488 + ], + "lines": [ + { + "bbox": [ + 313, + 454, + 554, + 488 + ], + "spans": [ + { + "bbox": [ + 313, + 454, + 554, + 488 + ], + "type": "text", + "content": "Figure 3. Taxonomy of kernel patterns for LUT-based methods. Dark cubes indicate rotation points, while medium-dark regions show involved pixel positions during one rotation." + } + ] + } + ], + "index": 15, + "angle": 0, + "type": "image_caption" + } + ], + "index": 13 + }, + { + "bbox": [ + 313, + 503, + 554, + 563 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 503, + 554, + 563 + ], + "spans": [ + { + "bbox": [ + 313, + 503, + 554, + 563 + ], + "type": "text", + "content": "and channel index. PCM effectively balances spatial and channel-wise processing, capturing inter-channel correlations crucial for color fidelity while preserving spatial relationships through the " + }, + { + "bbox": [ + 313, + 503, + 554, + 563 + ], + "type": "inline_equation", + "content": "1 \\times 2" + }, + { + "bbox": [ + 313, + 503, + 554, + 563 + ], + "type": "text", + "content": " kernel design and refining features via cascaded " + }, + { + "bbox": [ + 313, + 503, + 554, + 563 + ], + "type": "inline_equation", + "content": "1 \\times 1" + }, + { + "bbox": [ + 313, + 503, + 554, + 563 + ], + "type": "text", + "content": " convolutions." + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 313, + 574, + 494, + 586 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 574, + 494, + 586 + ], + "spans": [ + { + "bbox": [ + 313, + 574, + 494, + 586 + ], + "type": "text", + "content": "3.4. Rotation Non-overlapping Kernel" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 313, + 592, + 554, + 688 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 592, + 554, + 688 + ], + "spans": [ + { + "bbox": [ + 313, + 592, + 554, + 688 + ], + "type": "text", + "content": "The rotation ensemble strategy is fundamental to LUT-based methods, using four rotations " + }, + { + "bbox": [ + 313, + 592, + 554, + 688 + ], + "type": "inline_equation", + "content": "(0^{\\circ}, 90^{\\circ}, 180^{\\circ}, 270^{\\circ})" + }, + { + "bbox": [ + 313, + 592, + 554, + 688 + ], + "type": "text", + "content": " to enable asymmetric convolution kernels to cover larger symmetric areas. By combining outputs from all rotations, the approach enhances performance without increasing LUT index dimensionality. As shown in Fig. 4, traditional methods like SR-LUT and MuLUT use this strategy to expand receptive fields to " + }, + { + "bbox": [ + 313, + 592, + 554, + 688 + ], + "type": "inline_equation", + "content": "3 \\times 3" + }, + { + "bbox": [ + 313, + 592, + 554, + 688 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 313, + 592, + 554, + 688 + ], + "type": "inline_equation", + "content": "5 \\times 5" + }, + { + "bbox": [ + 313, + 592, + 554, + 688 + ], + "type": "text", + "content": " respectively." + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 313, + 689, + 554, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 689, + 554, + 713 + ], + "spans": [ + { + "bbox": [ + 313, + 689, + 554, + 713 + ], + "type": "text", + "content": "However, this conventional approach has two critical limitations. First, approximately half of the pixels within the re" + } + ] + } + ], + "index": 19 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "text", + "content": "7585" + } + ] + } + ], + "index": 20 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 3 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 77, + 71, + 277, + 228 + ], + "blocks": [ + { + "bbox": [ + 77, + 71, + 277, + 228 + ], + "lines": [ + { + "bbox": [ + 77, + 71, + 277, + 228 + ], + "spans": [ + { + "bbox": [ + 77, + 71, + 277, + 228 + ], + "type": "image", + "image_path": "8e5132bb3c844e09da05de9eae0fb9a91501eeb7d9b12b43eab0d053d3b6c6f2.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 55, + 236, + 296, + 270 + ], + "lines": [ + { + "bbox": [ + 55, + 236, + 296, + 270 + ], + "spans": [ + { + "bbox": [ + 55, + 236, + 296, + 270 + ], + "type": "text", + "content": "Figure 4. Comparison of spatial-wise kernel designs. Left patterns show kernel configurations, while right tables quantify lookup frequencies during output retrieval." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "bbox": [ + 55, + 281, + 296, + 341 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 281, + 296, + 341 + ], + "spans": [ + { + "bbox": [ + 55, + 281, + 296, + 341 + ], + "type": "text", + "content": "ceptive field are accessed multiple times during lookup operations, as illustrated in Fig. 4's tables. This redundancy indicates inefficient kernel pattern design. Second, stacking these kernels to create multiple LUTs for expanded receptive fields results in significant 4D LUT storage overhead." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 55, + 341, + 296, + 449 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 341, + 296, + 449 + ], + "spans": [ + { + "bbox": [ + 55, + 341, + 296, + 449 + ], + "type": "text", + "content": "To address these inefficiencies, we introduce an innovative L-shaped kernel design. During each rotation, the kernel processes two additional pixels (beyond the center pixel) without overlap, ensuring each surrounding pixel contributes exactly once to the output. This design maximizes pixel value utilization within the receptive field while enabling conversion to 3D LUTs instead of 4D LUTs, dramatically reducing storage requirements without sacrificing coverage." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 55, + 456, + 179, + 468 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 456, + 179, + 468 + ], + "spans": [ + { + "bbox": [ + 55, + 456, + 179, + 468 + ], + "type": "text", + "content": "3.5. PCM Plug-in Module" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 55, + 474, + 296, + 594 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 474, + 296, + 594 + ], + "spans": [ + { + "bbox": [ + 55, + 474, + 296, + 594 + ], + "type": "text", + "content": "To demonstrate the versatility of our pairwise channel mixer, we developed PCM as a plug-in module compatible with existing LUT-based methods. The inference of plug-in algorithm is summarized in Alg. 1. This module seamlessly integrates with current architectures, enhancing their ability to capture channel-spatial correlations without requiring structural modifications. Compared to state-of-the-art methods like SPFLUT, the PCM plug-in adds only " + }, + { + "bbox": [ + 55, + 474, + 296, + 594 + ], + "type": "inline_equation", + "content": "12\\%" + }, + { + "bbox": [ + 55, + 474, + 296, + 594 + ], + "type": "text", + "content": " runtime and " + }, + { + "bbox": [ + 55, + 474, + 296, + 594 + ], + "type": "inline_equation", + "content": "8\\%" + }, + { + "bbox": [ + 55, + 474, + 296, + 594 + ], + "type": "text", + "content": " storage overhead while delivering over 1dB performance improvement." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 55, + 605, + 137, + 617 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 605, + 137, + 617 + ], + "spans": [ + { + "bbox": [ + 55, + 605, + 137, + 617 + ], + "type": "text", + "content": "4. Experiments" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 55, + 624, + 186, + 636 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 624, + 186, + 636 + ], + "spans": [ + { + "bbox": [ + 55, + 624, + 186, + 636 + ], + "type": "text", + "content": "4.1. Implementation details" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 55, + 641, + 296, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 641, + 296, + 714 + ], + "spans": [ + { + "bbox": [ + 55, + 641, + 296, + 714 + ], + "type": "text", + "content": "We implement DnLUT using the Adam optimizer with a cosine annealing learning rate schedule and mean squared error (MSE) loss function. For Gaussian denoising, we train networks for " + }, + { + "bbox": [ + 55, + 641, + 296, + 714 + ], + "type": "inline_equation", + "content": "2 \\times 10^{5}" + }, + { + "bbox": [ + 55, + 641, + 296, + 714 + ], + "type": "text", + "content": " iterations with a batch size of 12, increasing to 32 for real-world denoising scenarios. Following established practices[10, 12], we uniformly sample LUTs" + } + ] + } + ], + "index": 8 + }, + { + "type": "code", + "bbox": [ + 316, + 87, + 555, + 390 + ], + "blocks": [ + { + "bbox": [ + 315, + 72, + 553, + 84 + ], + "lines": [ + { + "bbox": [ + 315, + 72, + 553, + 84 + ], + "spans": [ + { + "bbox": [ + 315, + 72, + 553, + 84 + ], + "type": "text", + "content": "Algorithm 1 Inference of Pairwise Channel Mixer Module" + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "code_caption" + }, + { + "bbox": [ + 316, + 87, + 555, + 390 + ], + "lines": [ + { + "bbox": [ + 316, + 87, + 555, + 390 + ], + "spans": [ + { + "bbox": [ + 316, + 87, + 555, + 390 + ], + "type": "text", + "content": "Require: Input image " + }, + { + "bbox": [ + 316, + 87, + 555, + 390 + ], + "type": "inline_equation", + "content": "I" + }, + { + "bbox": [ + 316, + 87, + 555, + 390 + ], + "type": "text", + "content": " of size " + }, + { + "bbox": [ + 316, + 87, + 555, + 390 + ], + "type": "inline_equation", + "content": "H\\times W\\times 3" + }, + { + "bbox": [ + 316, + 87, + 555, + 390 + ], + "type": "text", + "content": " \nRequire: 4D Lookup Table LUT of PCM \nRequire: Padding size " + }, + { + "bbox": [ + 316, + 87, + 555, + 390 + ], + "type": "inline_equation", + "content": "p" + }, + { + "bbox": [ + 316, + 87, + 555, + 390 + ], + "type": "text", + "content": " \n1: Initialize output image " + }, + { + "bbox": [ + 316, + 87, + 555, + 390 + ], + "type": "inline_equation", + "content": "O" + }, + { + "bbox": [ + 316, + 87, + 555, + 390 + ], + "type": "text", + "content": " of size " + }, + { + "bbox": [ + 316, + 87, + 555, + 390 + ], + "type": "inline_equation", + "content": "H\\times W\\times 3" + }, + { + "bbox": [ + 316, + 87, + 555, + 390 + ], + "type": "text", + "content": " \n2: for " + }, + { + "bbox": [ + 316, + 87, + 555, + 390 + ], + "type": "inline_equation", + "content": "r = 0" + }, + { + "bbox": [ + 316, + 87, + 555, + 390 + ], + "type": "text", + "content": " to 3 do \n3: Rotate image " + }, + { + "bbox": [ + 316, + 87, + 555, + 390 + ], + "type": "inline_equation", + "content": "I" + }, + { + "bbox": [ + 316, + 87, + 555, + 390 + ], + "type": "text", + "content": " by " + }, + { + "bbox": [ + 316, + 87, + 555, + 390 + ], + "type": "inline_equation", + "content": "r\\times 90" + }, + { + "bbox": [ + 316, + 87, + 555, + 390 + ], + "type": "text", + "content": " degrees to get " + }, + { + "bbox": [ + 316, + 87, + 555, + 390 + ], + "type": "inline_equation", + "content": "I_{\\mathrm{rot}}" + }, + { + "bbox": [ + 316, + 87, + 555, + 390 + ], + "type": "text", + "content": " \n4: Pad " + }, + { + "bbox": [ + 316, + 87, + 555, + 390 + ], + "type": "inline_equation", + "content": "I_{\\mathrm{rot}}" + }, + { + "bbox": [ + 316, + 87, + 555, + 390 + ], + "type": "text", + "content": " with " + }, + { + "bbox": [ + 316, + 87, + 555, + 390 + ], + "type": "inline_equation", + "content": "p" + }, + { + "bbox": [ + 316, + 87, + 555, + 390 + ], + "type": "text", + "content": " pixels on each side to get " + }, + { + "bbox": [ + 316, + 87, + 555, + 390 + ], + "type": "inline_equation", + "content": "I_{\\mathrm{rot}}^{\\prime}" + }, + { + "bbox": [ + 316, + 87, + 555, + 390 + ], + "type": "text", + "content": " \n5: for " + }, + { + "bbox": [ + 316, + 87, + 555, + 390 + ], + "type": "inline_equation", + "content": "h = 0" + }, + { + "bbox": [ + 316, + 87, + 555, + 390 + ], + "type": "text", + "content": " to " + }, + { + "bbox": [ + 316, + 87, + 555, + 390 + ], + "type": "inline_equation", + "content": "H - 1" + }, + { + "bbox": [ + 316, + 87, + 555, + 390 + ], + "type": "text", + "content": " do \n6: for " + }, + { + "bbox": [ + 316, + 87, + 555, + 390 + ], + "type": "inline_equation", + "content": "w = 0" + }, + { + "bbox": [ + 316, + 87, + 555, + 390 + ], + "type": "text", + "content": " to " + }, + { + "bbox": [ + 316, + 87, + 555, + 390 + ], + "type": "inline_equation", + "content": "W - 1" + }, + { + "bbox": [ + 316, + 87, + 555, + 390 + ], + "type": "text", + "content": " do \n7: for " + }, + { + "bbox": [ + 316, + 87, + 555, + 390 + ], + "type": "inline_equation", + "content": "c = 0" + }, + { + "bbox": [ + 316, + 87, + 555, + 390 + ], + "type": "text", + "content": " to 2 do \n8: " + }, + { + "bbox": [ + 316, + 87, + 555, + 390 + ], + "type": "inline_equation", + "content": "p_1\\gets I_{\\mathrm{rot}}'" + }, + { + "bbox": [ + 316, + 87, + 555, + 390 + ], + "type": "text", + "content": " [h,w,c] \n9: " + }, + { + "bbox": [ + 316, + 87, + 555, + 390 + ], + "type": "inline_equation", + "content": "p_2\\gets I_{\\mathrm{rot}}'[h,w + 1,c]" + }, + { + "bbox": [ + 316, + 87, + 555, + 390 + ], + "type": "text", + "content": " \n10: " + }, + { + "bbox": [ + 316, + 87, + 555, + 390 + ], + "type": "inline_equation", + "content": "p_3\\gets I_{\\mathrm{rot}}'[h,w,(c + 1)\\mod 3]" + }, + { + "bbox": [ + 316, + 87, + 555, + 390 + ], + "type": "text", + "content": " \n11: " + }, + { + "bbox": [ + 316, + 87, + 555, + 390 + ], + "type": "inline_equation", + "content": "p_4\\gets I_{\\mathrm{rot}}'[h,w + 1,(c + 1)\\mod 3]" + }, + { + "bbox": [ + 316, + 87, + 555, + 390 + ], + "type": "text", + "content": " \n12: index " + }, + { + "bbox": [ + 316, + 87, + 555, + 390 + ], + "type": "inline_equation", + "content": "\\leftarrow" + }, + { + "bbox": [ + 316, + 87, + 555, + 390 + ], + "type": "text", + "content": " Quantize " + }, + { + "bbox": [ + 316, + 87, + 555, + 390 + ], + "type": "inline_equation", + "content": "(p_1,p_2,p_3,p_4)" + }, + { + "bbox": [ + 316, + 87, + 555, + 390 + ], + "type": "text", + "content": " \n13: value " + }, + { + "bbox": [ + 316, + 87, + 555, + 390 + ], + "type": "inline_equation", + "content": "\\leftarrow" + }, + { + "bbox": [ + 316, + 87, + 555, + 390 + ], + "type": "text", + "content": " LUT[index] \n14: " + }, + { + "bbox": [ + 316, + 87, + 555, + 390 + ], + "type": "inline_equation", + "content": "O[h,w,c]\\gets O[h,w,c] + \\mathrm{Interpolate}(\\mathrm{value})" + }, + { + "bbox": [ + 316, + 87, + 555, + 390 + ], + "type": "text", + "content": " \n15: end for \n16: end for \n17: end for \n18: Rotate " + }, + { + "bbox": [ + 316, + 87, + 555, + 390 + ], + "type": "inline_equation", + "content": "O" + }, + { + "bbox": [ + 316, + 87, + 555, + 390 + ], + "type": "text", + "content": " back by " + }, + { + "bbox": [ + 316, + 87, + 555, + 390 + ], + "type": "inline_equation", + "content": "(4 - r)\\times 90" + }, + { + "bbox": [ + 316, + 87, + 555, + 390 + ], + "type": "text", + "content": " degrees to get " + }, + { + "bbox": [ + 316, + 87, + 555, + 390 + ], + "type": "inline_equation", + "content": "O_{\\mathrm{rot}}" + }, + { + "bbox": [ + 316, + 87, + 555, + 390 + ], + "type": "text", + "content": " \n19: " + }, + { + "bbox": [ + 316, + 87, + 555, + 390 + ], + "type": "inline_equation", + "content": "O\\gets O + O_{\\mathrm{rot}}" + }, + { + "bbox": [ + 316, + 87, + 555, + 390 + ], + "type": "text", + "content": " \n20: end for \n21: " + }, + { + "bbox": [ + 316, + 87, + 555, + 390 + ], + "type": "inline_equation", + "content": "O\\gets O / 4" + }, + { + "bbox": [ + 316, + 87, + 555, + 390 + ], + "type": "text", + "content": " \n22: return Output image " + }, + { + "bbox": [ + 316, + 87, + 555, + 390 + ], + "type": "inline_equation", + "content": "O" + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "code_body" + } + ], + "index": 10, + "sub_type": "algorithm" + }, + { + "bbox": [ + 313, + 413, + 555, + 472 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 413, + 555, + 472 + ], + "spans": [ + { + "bbox": [ + 313, + 413, + 555, + 472 + ], + "type": "text", + "content": "at " + }, + { + "bbox": [ + 313, + 413, + 555, + 472 + ], + "type": "inline_equation", + "content": "2^{4}" + }, + { + "bbox": [ + 313, + 413, + 555, + 472 + ], + "type": "text", + "content": " intervals. To minimize indexing artifacts, final predictions utilize 4D simplex interpolation and 3D tetrahedral interpolation. We further refine cached LUTs through 2000 iterations of LUT-aware fine-tuning[12] on the training dataset." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 313, + 485, + 489, + 498 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 485, + 489, + 498 + ], + "spans": [ + { + "bbox": [ + 313, + 485, + 489, + 498 + ], + "type": "text", + "content": "4.2. Gaussian Color Image denoising" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 313, + 504, + 554, + 588 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 504, + 554, + 588 + ], + "spans": [ + { + "bbox": [ + 313, + 504, + 554, + 588 + ], + "type": "text", + "content": "Following [27], we train our model using four comprehensive datasets: DIV2K[1], Flickr2K[23], BSD400[20], and WaterlooED[18]. Training data is augmented with additive Gaussian noise at three levels (" + }, + { + "bbox": [ + 313, + 504, + 554, + 588 + ], + "type": "inline_equation", + "content": "\\sigma \\in \\{15, 25, 50\\}" + }, + { + "bbox": [ + 313, + 504, + 554, + 588 + ], + "type": "text", + "content": "). We evaluate performance on four standard benchmarks: CBSD68[21], Kodak24, Urban100[9], and McMaster[31], using identical noise levels." + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 313, + 599, + 499, + 612 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 599, + 499, + 612 + ], + "spans": [ + { + "bbox": [ + 313, + 599, + 499, + 612 + ], + "type": "text", + "content": "4.3. Real-world Color Image denoising" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 313, + 617, + 555, + 712 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 617, + 555, + 712 + ], + "spans": [ + { + "bbox": [ + 313, + 617, + 555, + 712 + ], + "type": "text", + "content": "To validate real-world performance, we train DnLUT on the SIDD training dataset and evaluate on both SIDD validation and DnD datasets. The SIDD dataset comprises indoor scenes captured by five smartphones under varying noise conditions, providing 320 image pairs for training and 1,280 for validation. The DnD dataset offers 50 high-resolution images with realistic noise patterns for comprehensive evaluation." + } + ] + } + ], + "index": 15 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 749, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 749, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 749, + 315, + 757 + ], + "type": "text", + "content": "7586" + } + ] + } + ], + "index": 16 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 4 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 70, + 91, + 541, + 220 + ], + "blocks": [ + { + "bbox": [ + 85, + 71, + 523, + 83 + ], + "lines": [ + { + "bbox": [ + 85, + 71, + 523, + 83 + ], + "spans": [ + { + "bbox": [ + 85, + 71, + 523, + 83 + ], + "type": "text", + "content": "Table 2. Quantitative comparison on Gaussian color image denoising benchmark. The table presents CPSNR(dB) values." + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 70, + 91, + 541, + 220 + ], + "lines": [ + { + "bbox": [ + 70, + 91, + 541, + 220 + ], + "spans": [ + { + "bbox": [ + 70, + 91, + 541, + 220 + ], + "type": "table", + "html": "
CategoryMethodCBSD68Kodak24Urban100McMaster
σ = 15σ = 25σ = 50σ = 15σ = 25σ = 50σ = 15σ = 25σ = 50σ = 15σ = 25σ = 50
LUT-basedSR-LUT[10]29.7626.7122.4130.3527.1622.6529.3826.0421.6031.1828.0123.35
MuLUT[12]30.5228.1124.8531.3129.0225.2830.2527.6723.7532.2829.8826.36
RC-LUT[16]30.6828.1225.0431.5729.0725.8930.3327.8023.8632.5129.8926.50
SPF-LUT[14]30.9728.5625.3331.8629.5826.2630.8928.2624.2231.7730.4426.91
DnLUT(Ours)32.4129.8826.0333.0230.2426.7432.1228.8725.0132.8830.4427.12
ClassicalCBM3D[3]33.5230.7127.3834.2831.6829.0233.9331.3627.9334.0631.6628.51
MC-WNNM[24]31.9829.3226.9833.2330.8928.6730.2329.2327.0031.2330.2027.55
DNNDnCNN[30]33.9031.2427.9534.6032.1428.9632.9830.8127.5933.4531.5228.62
SwinIR[15]34.4231.7828.5635.3432.8929.7935.6133.2030.2235.1332.9029.82
", + "image_path": "9ebac7c69a6c28adc21b6995b3dba844bbb30518831c76e97f97f6abb13375a3.jpg" + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "table_body" + } + ], + "index": 1 + }, + { + "type": "table", + "bbox": [ + 67, + 259, + 542, + 316 + ], + "blocks": [ + { + "bbox": [ + 55, + 228, + 555, + 251 + ], + "lines": [ + { + "bbox": [ + 55, + 228, + 555, + 251 + ], + "spans": [ + { + "bbox": [ + 55, + 228, + 555, + 251 + ], + "type": "text", + "content": "Table 3. Quantitative comparison on real-world color image denoising. The table presents CPSNR(dB) and SSIM values. Methods on DnD dataset are validated on the online platform. We could only get the PSNR rather than CPSNR." + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 67, + 259, + 542, + 316 + ], + "lines": [ + { + "bbox": [ + 67, + 259, + 542, + 316 + ], + "spans": [ + { + "bbox": [ + 67, + 259, + 542, + 316 + ], + "type": "table", + "html": "
DatasetsMethodSR-LUT[10]MuLUT[12]RC-LUT[16]SPF-LUT[14]DnLUT(Ours)CBM3D[3]MC-WNNM[24]DnCNN[30]
SIDDCPSNR29.3833.2433.8834.9135.4430.1429.4536.45
SSIM0.6340.8300.8390.8650.8750.7020.6820.900
DnDPSNR*33.6935.1135.4336.2236.6733.1232.3437.11
SSIM0.8390.8680.8750.9110.9220.8230.8170.932
", + "image_path": "219c5e89bef2c959a70aa4cc9463c910a5b6a6b06199e788b449838060380ceb.jpg" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "table_body" + } + ], + "index": 3 + }, + { + "type": "image", + "bbox": [ + 63, + 325, + 244, + 435 + ], + "blocks": [ + { + "bbox": [ + 63, + 325, + 244, + 435 + ], + "lines": [ + { + "bbox": [ + 63, + 325, + 244, + 435 + ], + "spans": [ + { + "bbox": [ + 63, + 325, + 244, + 435 + ], + "type": "image", + "image_path": "f707c479ed17ad64ec199e63ac3cd3b3afcb8096a21be9db854fb3080697d70c.jpg" + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 135, + 435, + 175, + 445 + ], + "lines": [ + { + "bbox": [ + 135, + 435, + 175, + 445 + ], + "spans": [ + { + "bbox": [ + 135, + 435, + 175, + 445 + ], + "type": "text", + "content": "Full Image" + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "image_caption" + } + ], + "index": 4 + }, + { + "type": "image", + "bbox": [ + 250, + 325, + 321, + 376 + ], + "blocks": [ + { + "bbox": [ + 250, + 325, + 321, + 376 + ], + "lines": [ + { + "bbox": [ + 250, + 325, + 321, + 376 + ], + "spans": [ + { + "bbox": [ + 250, + 325, + 321, + 376 + ], + "type": "image", + "image_path": "7ce7a3866174c3984bcbc60f356a7bd9c645755d6017573c10ebde2dc03057bf.jpg" + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 270, + 375, + 299, + 383 + ], + "lines": [ + { + "bbox": [ + 270, + 375, + 299, + 383 + ], + "spans": [ + { + "bbox": [ + 270, + 375, + 299, + 383 + ], + "type": "text", + "content": "SRLUT" + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "image_caption" + } + ], + "index": 6 + }, + { + "type": "image", + "bbox": [ + 325, + 325, + 397, + 376 + ], + "blocks": [ + { + "bbox": [ + 325, + 325, + 397, + 376 + ], + "lines": [ + { + "bbox": [ + 325, + 325, + 397, + 376 + ], + "spans": [ + { + "bbox": [ + 325, + 325, + 397, + 376 + ], + "type": "image", + "image_path": "ac8f52598c456798f84da1e6b4a8fe65be8829fb2564be892e591b72d621e0bf.jpg" + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 345, + 375, + 375, + 384 + ], + "lines": [ + { + "bbox": [ + 345, + 375, + 375, + 384 + ], + "spans": [ + { + "bbox": [ + 345, + 375, + 375, + 384 + ], + "type": "text", + "content": "MuLUT" + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "image_caption" + } + ], + "index": 8 + }, + { + "type": "image", + "bbox": [ + 400, + 325, + 470, + 376 + ], + "blocks": [ + { + "bbox": [ + 400, + 325, + 470, + 376 + ], + "lines": [ + { + "bbox": [ + 400, + 325, + 470, + 376 + ], + "spans": [ + { + "bbox": [ + 400, + 325, + 470, + 376 + ], + "type": "image", + "image_path": "afea0d92d133941c17b1c9df3b62c374a7aabbd17fd0ae9bfc88f1ba36e6fb23.jpg" + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 419, + 375, + 451, + 384 + ], + "lines": [ + { + "bbox": [ + 419, + 375, + 451, + 384 + ], + "spans": [ + { + "bbox": [ + 419, + 375, + 451, + 384 + ], + "type": "text", + "content": "RC-LUT" + } + ] + } + ], + "index": 11, + "angle": 0, + "type": "image_caption" + } + ], + "index": 10 + }, + { + "type": "image", + "bbox": [ + 473, + 325, + 547, + 376 + ], + "blocks": [ + { + "bbox": [ + 473, + 325, + 547, + 376 + ], + "lines": [ + { + "bbox": [ + 473, + 325, + 547, + 376 + ], + "spans": [ + { + "bbox": [ + 473, + 325, + 547, + 376 + ], + "type": "image", + "image_path": "9a0c552aaa5de2bafef5f5d6f3fcb4eda5d13301a842d73143c0567ecccc7f5c.jpg" + } + ] + } + ], + "index": 12, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 493, + 375, + 525, + 383 + ], + "lines": [ + { + "bbox": [ + 493, + 375, + 525, + 383 + ], + "spans": [ + { + "bbox": [ + 493, + 375, + 525, + 383 + ], + "type": "text", + "content": "SPFLUT" + } + ] + } + ], + "index": 13, + "angle": 0, + "type": "image_caption" + } + ], + "index": 12 + }, + { + "type": "image", + "bbox": [ + 63, + 474, + 161, + 571 + ], + "blocks": [ + { + "bbox": [ + 63, + 474, + 161, + 571 + ], + "lines": [ + { + "bbox": [ + 63, + 474, + 161, + 571 + ], + "spans": [ + { + "bbox": [ + 63, + 474, + 161, + 571 + ], + "type": "image", + "image_path": "f3bee1ef768b24efe502fe8aa97ea3f7f981a36fa5f832801240b19a5944d7b9.jpg" + } + ] + } + ], + "index": 14, + "angle": 0, + "type": "image_body" + } + ], + "index": 14 + }, + { + "type": "image", + "bbox": [ + 162, + 474, + 258, + 571 + ], + "blocks": [ + { + "bbox": [ + 162, + 474, + 258, + 571 + ], + "lines": [ + { + "bbox": [ + 162, + 474, + 258, + 571 + ], + "spans": [ + { + "bbox": [ + 162, + 474, + 258, + 571 + ], + "type": "image", + "image_path": "395227161c30e30370e0f83df8d8f5d193cbba1e9a7734cf9464367a3b6d5b25.jpg" + } + ] + } + ], + "index": 16, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 194, + 571, + 225, + 581 + ], + "lines": [ + { + "bbox": [ + 194, + 571, + 225, + 581 + ], + "spans": [ + { + "bbox": [ + 194, + 571, + 225, + 581 + ], + "type": "text", + "content": "SRLUT" + } + ] + } + ], + "index": 17, + "angle": 0, + "type": "image_caption" + } + ], + "index": 16 + }, + { + "type": "image", + "bbox": [ + 258, + 474, + 352, + 571 + ], + "blocks": [ + { + "bbox": [ + 258, + 474, + 352, + 571 + ], + "lines": [ + { + "bbox": [ + 258, + 474, + 352, + 571 + ], + "spans": [ + { + "bbox": [ + 258, + 474, + 352, + 571 + ], + "type": "image", + "image_path": "95280d8cfef7da473a46e69b090e391a1f2899e83bf6fb54d5f21708b7de773f.jpg" + } + ] + } + ], + "index": 18, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 287, + 572, + 320, + 581 + ], + "lines": [ + { + "bbox": [ + 287, + 572, + 320, + 581 + ], + "spans": [ + { + "bbox": [ + 287, + 572, + 320, + 581 + ], + "type": "text", + "content": "MuLUT" + } + ] + } + ], + "index": 19, + "angle": 0, + "type": "image_caption" + } + ], + "index": 18 + }, + { + "type": "image", + "bbox": [ + 353, + 474, + 450, + 571 + ], + "blocks": [ + { + "bbox": [ + 353, + 474, + 450, + 571 + ], + "lines": [ + { + "bbox": [ + 353, + 474, + 450, + 571 + ], + "spans": [ + { + "bbox": [ + 353, + 474, + 450, + 571 + ], + "type": "image", + "image_path": "29a41cc7b4521de450807c711d1e8a10eb8cce645468fdb2d4784c279fcb5e41.jpg" + } + ] + } + ], + "index": 20, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 383, + 572, + 417, + 581 + ], + "lines": [ + { + "bbox": [ + 383, + 572, + 417, + 581 + ], + "spans": [ + { + "bbox": [ + 383, + 572, + 417, + 581 + ], + "type": "text", + "content": "RC-LUT" + } + ] + } + ], + "index": 21, + "angle": 0, + "type": "image_caption" + } + ], + "index": 20 + }, + { + "type": "image", + "bbox": [ + 450, + 474, + 545, + 571 + ], + "blocks": [ + { + "bbox": [ + 450, + 474, + 545, + 571 + ], + "lines": [ + { + "bbox": [ + 450, + 474, + 545, + 571 + ], + "spans": [ + { + "bbox": [ + 450, + 474, + 545, + 571 + ], + "type": "image", + "image_path": "cd165983b48473f3a8e5c50857f723ffbf4bc58e920e41be487f339981236403.jpg" + } + ] + } + ], + "index": 22, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 481, + 572, + 516, + 581 + ], + "lines": [ + { + "bbox": [ + 481, + 572, + 516, + 581 + ], + "spans": [ + { + "bbox": [ + 481, + 572, + 516, + 581 + ], + "type": "text", + "content": "SPFLUT" + } + ] + } + ], + "index": 23, + "angle": 0, + "type": "image_caption" + } + ], + "index": 22 + }, + { + "type": "image", + "bbox": [ + 65, + 582, + 161, + 678 + ], + "blocks": [ + { + "bbox": [ + 107, + 572, + 128, + 582 + ], + "lines": [ + { + "bbox": [ + 107, + 572, + 128, + 582 + ], + "spans": [ + { + "bbox": [ + 107, + 572, + 128, + 582 + ], + "type": "text", + "content": "Input" + } + ] + } + ], + "index": 15, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 65, + 582, + 161, + 678 + ], + "lines": [ + { + "bbox": [ + 65, + 582, + 161, + 678 + ], + "spans": [ + { + "bbox": [ + 65, + 582, + 161, + 678 + ], + "type": "image", + "image_path": "76eef8a75f28d6b1720a683a7a877bdacb938c1ec497de1509fc17f19b355c4f.jpg" + } + ] + } + ], + "index": 24, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 108, + 679, + 122, + 688 + ], + "lines": [ + { + "bbox": [ + 108, + 679, + 122, + 688 + ], + "spans": [ + { + "bbox": [ + 108, + 679, + 122, + 688 + ], + "type": "text", + "content": "GT" + } + ] + } + ], + "index": 25, + "angle": 0, + "type": "image_caption" + } + ], + "index": 24 + }, + { + "type": "image", + "bbox": [ + 162, + 582, + 258, + 678 + ], + "blocks": [ + { + "bbox": [ + 162, + 582, + 258, + 678 + ], + "lines": [ + { + "bbox": [ + 162, + 582, + 258, + 678 + ], + "spans": [ + { + "bbox": [ + 162, + 582, + 258, + 678 + ], + "type": "image", + "image_path": "b47e2080ee26174667d7f1bc0af8665b45e7771b438399a8cfbd77bc61323153.jpg" + } + ] + } + ], + "index": 26, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 192, + 679, + 226, + 688 + ], + "lines": [ + { + "bbox": [ + 192, + 679, + 226, + 688 + ], + "spans": [ + { + "bbox": [ + 192, + 679, + 226, + 688 + ], + "type": "text", + "content": "CBM3D" + } + ] + } + ], + "index": 27, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 201, + 699, + 408, + 710 + ], + "lines": [ + { + "bbox": [ + 201, + 699, + 408, + 710 + ], + "spans": [ + { + "bbox": [ + 201, + 699, + 408, + 710 + ], + "type": "text", + "content": "Figure 6. Qualitative comparison on real-world datasets." + } + ] + } + ], + "index": 34, + "angle": 0, + "type": "image_caption" + } + ], + "index": 26 + }, + { + "type": "image", + "bbox": [ + 258, + 582, + 353, + 678 + ], + "blocks": [ + { + "bbox": [ + 258, + 582, + 353, + 678 + ], + "lines": [ + { + "bbox": [ + 258, + 582, + 353, + 678 + ], + "spans": [ + { + "bbox": [ + 258, + 582, + 353, + 678 + ], + "type": "image", + "image_path": "3cdf760cefd38d030708448fd15a6d5a672f9d0bd2dca9e784ffabc0c086474c.jpg" + } + ] + } + ], + "index": 28, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 280, + 679, + 331, + 688 + ], + "lines": [ + { + "bbox": [ + 280, + 679, + 331, + 688 + ], + "spans": [ + { + "bbox": [ + 280, + 679, + 331, + 688 + ], + "type": "text", + "content": "MC-WNNM" + } + ] + } + ], + "index": 29, + "angle": 0, + "type": "image_caption" + } + ], + "index": 28 + }, + { + "type": "image", + "bbox": [ + 353, + 582, + 450, + 678 + ], + "blocks": [ + { + "bbox": [ + 353, + 582, + 450, + 678 + ], + "lines": [ + { + "bbox": [ + 353, + 582, + 450, + 678 + ], + "spans": [ + { + "bbox": [ + 353, + 582, + 450, + 678 + ], + "type": "image", + "image_path": "b0edf591dfa56f6bc933646ab52b1e49f9f6652b6361779fbe325d5a59676021.jpg" + } + ] + } + ], + "index": 30, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 373, + 679, + 429, + 689 + ], + "lines": [ + { + "bbox": [ + 373, + 679, + 429, + 689 + ], + "spans": [ + { + "bbox": [ + 373, + 679, + 429, + 689 + ], + "type": "text", + "content": "DnLUT(Ours)" + } + ] + } + ], + "index": 31, + "angle": 0, + "type": "image_caption" + } + ], + "index": 30 + }, + { + "type": "image", + "bbox": [ + 450, + 582, + 545, + 678 + ], + "blocks": [ + { + "bbox": [ + 450, + 582, + 545, + 678 + ], + "lines": [ + { + "bbox": [ + 450, + 582, + 545, + 678 + ], + "spans": [ + { + "bbox": [ + 450, + 582, + 545, + 678 + ], + "type": "image", + "image_path": "db9d85892a1f8bd632f1053a07d14f75f79d56e2141e864ccfec3b7b500bc48e.jpg" + } + ] + } + ], + "index": 32, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 482, + 679, + 515, + 688 + ], + "lines": [ + { + "bbox": [ + 482, + 679, + 515, + 688 + ], + "spans": [ + { + "bbox": [ + 482, + 679, + 515, + 688 + ], + "type": "text", + "content": "DnCNN" + } + ] + } + ], + "index": 33, + "angle": 0, + "type": "image_caption" + } + ], + "index": 32 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 749, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 749, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 749, + 315, + 757 + ], + "type": "text", + "content": "7587" + } + ] + } + ], + "index": 35 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 5 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 100, + 113, + 509, + 230 + ], + "blocks": [ + { + "bbox": [ + 55, + 70, + 555, + 103 + ], + "lines": [ + { + "bbox": [ + 55, + 70, + 555, + 103 + ], + "spans": [ + { + "bbox": [ + 55, + 70, + 555, + 103 + ], + "type": "text", + "content": "Table 4. PCM can be easily incorporated to the existing LUT-based methods. It equips the model with the ability of capturing the channel correlations which brings more than 1dB on the widely used benchmark. The noise level " + }, + { + "bbox": [ + 55, + 70, + 555, + 103 + ], + "type": "inline_equation", + "content": "\\sigma" + }, + { + "bbox": [ + 55, + 70, + 555, + 103 + ], + "type": "text", + "content": " of Gaussian noise is set to 25. The table presents CPSNR(dB) values." + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 100, + 113, + 509, + 230 + ], + "lines": [ + { + "bbox": [ + 100, + 113, + 509, + 230 + ], + "spans": [ + { + "bbox": [ + 100, + 113, + 509, + 230 + ], + "type": "table", + "html": "
MethodCBSD68Kodak24Urban100McMasterSIDD
SR-LUT[10]26.7127.1626.0428.0129.38
PCM+SR-LUT28.04(+1.33)28.55(+1.39)27.33(+1.29)28.78(+0.77)32.33(+2.96)
MuLUT[12]28.1129.0127.6729.8833.24
PCM+MuLUT29.17(+1.05)29.92(+0.91)28.74(+1.07)30.33(+0.44)34.33(+1.09)
RC-LUT[16]28.1229.0727.8029.8933.88
PCM+RC-LUT29.22(+1.10)30.09(+1.02)28.77(+0.97)30.54(+0.65)34.67(+0.79)
SPF-LUT[14]28.5629.5828.2630.4434.91
PCM+SPF-LUT29.72(+1.16)30.63(+1.05)29.43(+1.17)30.96(+0.52)35.56(+0.65)
", + "image_path": "8f429ef9d4d32a3aa79d3d435809fb61763aae5cfe830c203ca17182114781c1.jpg" + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "table_body" + } + ], + "index": 1 + }, + { + "bbox": [ + 55, + 236, + 171, + 248 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 236, + 171, + 248 + ], + "spans": [ + { + "bbox": [ + 55, + 236, + 171, + 248 + ], + "type": "text", + "content": "4.4. Quantitative results" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 55, + 253, + 296, + 337 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 253, + 296, + 337 + ], + "spans": [ + { + "bbox": [ + 55, + 253, + 296, + 337 + ], + "type": "text", + "content": "We evaluate performance using color peak signal-to-noise ratio(CPSNR) and structural similarity index(SSIM) metrics, comparing against state-of-the-art methods across three categories: LUT-based methods (SR-LUT[10], MuLUT[12], RC-LUT[16], SPF-LUT[14]), classical approaches (CBM3D[3], MC-WNNM[24]), and DNN-based solutions (DnCNN[30], SwinIR[15])." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 55, + 338, + 296, + 493 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 338, + 296, + 493 + ], + "spans": [ + { + "bbox": [ + 55, + 338, + 296, + 493 + ], + "type": "text", + "content": "As shown in Tab. 2, DnLUT achieves superior performance in the LUT-based category, surpassing SPF-LUT by over 1dB in Gaussian noise removal while requiring only " + }, + { + "bbox": [ + 55, + 338, + 296, + 493 + ], + "type": "inline_equation", + "content": "17\\%" + }, + { + "bbox": [ + 55, + 338, + 296, + 493 + ], + "type": "text", + "content": " of its storage. Results from Tab. 3 demonstrate DnLUT's effectiveness in real-world scenarios, outperforming classical methods CBM3D[3] and MC-WNNM[24] by approximately 5dB. Though the Gaussian noise is independent across RGB channels, the network can still leverage inter-channel natural distribution in images to suppress the noise. For example, if a pixel in the R channel has a high value due to noise, but the corresponding pixels in the G and B channels do not, the network can infer that the high value is likely noise and suppress it." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 55, + 502, + 165, + 514 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 502, + 165, + 514 + ], + "spans": [ + { + "bbox": [ + 55, + 502, + 165, + 514 + ], + "type": "text", + "content": "4.5. Qualitative results" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 55, + 519, + 296, + 639 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 519, + 296, + 639 + ], + "spans": [ + { + "bbox": [ + 55, + 519, + 296, + 639 + ], + "type": "text", + "content": "Visual comparisons in Figs. 5 & 6 illustrate DnLUT's superior denoising capabilities. Existing LUT-based methods exhibit noticeable color distortion, particularly evident in Fig. 5's first row. In contrast, DnLUT achieves visual quality comparable to computation-intensive methods like DnCNN and CBM3D. Fig. 6 further demonstrates DnLUT's advantage in real-world scenarios, where both LUT-based and classical methods struggle with color fidelity and often over-smooth details. DnLUT successfully preserves color consistency while retaining fine image details." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 55, + 647, + 178, + 659 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 647, + 178, + 659 + ], + "spans": [ + { + "bbox": [ + 55, + 647, + 178, + 659 + ], + "type": "text", + "content": "4.6. Efficiency Evaluation" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 55, + 665, + 295, + 700 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 665, + 295, + 700 + ], + "spans": [ + { + "bbox": [ + 55, + 665, + 295, + 700 + ], + "type": "text", + "content": "We conduct comprehensive efficiency analysis across three key metrics: theoretical energy cost, runtime performance, and storage requirements (Tab. 5)." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 55, + 701, + 296, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 701, + 296, + 714 + ], + "spans": [ + { + "bbox": [ + 55, + 701, + 296, + 714 + ], + "type": "text", + "content": "Energy cost. Following AdderSR[22], we analyze multipli" + } + ] + } + ], + "index": 9 + }, + { + "type": "image", + "bbox": [ + 319, + 234, + 377, + 277 + ], + "blocks": [ + { + "bbox": [ + 319, + 234, + 377, + 277 + ], + "lines": [ + { + "bbox": [ + 319, + 234, + 377, + 277 + ], + "spans": [ + { + "bbox": [ + 319, + 234, + 377, + 277 + ], + "type": "image", + "image_path": "70d838580f39f91a819f746cb018318068c63b319adaa17c2727446b67ed393d.jpg" + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 337, + 277, + 359, + 283 + ], + "lines": [ + { + "bbox": [ + 337, + 277, + 359, + 283 + ], + "spans": [ + { + "bbox": [ + 337, + 277, + 359, + 283 + ], + "type": "text", + "content": "SRLUT" + } + ] + } + ], + "index": 11, + "angle": 0, + "type": "image_caption" + } + ], + "index": 10 + }, + { + "type": "image", + "bbox": [ + 377, + 235, + 434, + 277 + ], + "blocks": [ + { + "bbox": [ + 377, + 235, + 434, + 277 + ], + "lines": [ + { + "bbox": [ + 377, + 235, + 434, + 277 + ], + "spans": [ + { + "bbox": [ + 377, + 235, + 434, + 277 + ], + "type": "image", + "image_path": "ed0608af79df9df6536b9813cd27dfbf9b06672f33a8aa0a15f453c12e3083a7.jpg" + } + ] + } + ], + "index": 12, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 395, + 277, + 418, + 283 + ], + "lines": [ + { + "bbox": [ + 395, + 277, + 418, + 283 + ], + "spans": [ + { + "bbox": [ + 395, + 277, + 418, + 283 + ], + "type": "text", + "content": "MuLUT" + } + ] + } + ], + "index": 13, + "angle": 0, + "type": "image_caption" + } + ], + "index": 12 + }, + { + "type": "image", + "bbox": [ + 434, + 235, + 492, + 277 + ], + "blocks": [ + { + "bbox": [ + 434, + 235, + 492, + 277 + ], + "lines": [ + { + "bbox": [ + 434, + 235, + 492, + 277 + ], + "spans": [ + { + "bbox": [ + 434, + 235, + 492, + 277 + ], + "type": "image", + "image_path": "bc48ddfe7c6c2b5550cccd38393f187a403c0f59d33c08e1e2aef246cf16befa.jpg" + } + ] + } + ], + "index": 14, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 451, + 277, + 476, + 284 + ], + "lines": [ + { + "bbox": [ + 451, + 277, + 476, + 284 + ], + "spans": [ + { + "bbox": [ + 451, + 277, + 476, + 284 + ], + "type": "text", + "content": "RC-LUT" + } + ] + } + ], + "index": 15, + "angle": 0, + "type": "image_caption" + } + ], + "index": 14 + }, + { + "type": "image", + "bbox": [ + 493, + 235, + 551, + 277 + ], + "blocks": [ + { + "bbox": [ + 493, + 235, + 551, + 277 + ], + "lines": [ + { + "bbox": [ + 493, + 235, + 551, + 277 + ], + "spans": [ + { + "bbox": [ + 493, + 235, + 551, + 277 + ], + "type": "image", + "image_path": "a2d9b749cd909fa61fd0ed0e1ce2ab506c4effc0835f7c584d97f03badd6c1cc.jpg" + } + ] + } + ], + "index": 16, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 505, + 277, + 531, + 284 + ], + "lines": [ + { + "bbox": [ + 505, + 277, + 531, + 284 + ], + "spans": [ + { + "bbox": [ + 505, + 277, + 531, + 284 + ], + "type": "text", + "content": "SPFLUT" + } + ] + } + ], + "index": 17, + "angle": 0, + "type": "image_caption" + } + ], + "index": 16 + }, + { + "type": "image", + "bbox": [ + 319, + 286, + 376, + 329 + ], + "blocks": [ + { + "bbox": [ + 319, + 286, + 376, + 329 + ], + "lines": [ + { + "bbox": [ + 319, + 286, + 376, + 329 + ], + "spans": [ + { + "bbox": [ + 319, + 286, + 376, + 329 + ], + "type": "image", + "image_path": "67db694903958bfe9172aecb7c267f2a07875508d01ab1899d3f9de0ce9229c6.jpg" + } + ] + } + ], + "index": 18, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 328, + 330, + 368, + 337 + ], + "lines": [ + { + "bbox": [ + 328, + 330, + 368, + 337 + ], + "spans": [ + { + "bbox": [ + 328, + 330, + 368, + 337 + ], + "type": "text", + "content": "PCM+SRLUT" + } + ] + } + ], + "index": 19, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 315, + 346, + 553, + 358 + ], + "lines": [ + { + "bbox": [ + 315, + 346, + 553, + 358 + ], + "spans": [ + { + "bbox": [ + 315, + 346, + 553, + 358 + ], + "type": "text", + "content": "Figure 7. Visualization of LUT-based method with plug-in PCM." + } + ] + } + ], + "index": 26, + "angle": 0, + "type": "image_caption" + } + ], + "index": 18 + }, + { + "type": "image", + "bbox": [ + 377, + 286, + 434, + 329 + ], + "blocks": [ + { + "bbox": [ + 377, + 286, + 434, + 329 + ], + "lines": [ + { + "bbox": [ + 377, + 286, + 434, + 329 + ], + "spans": [ + { + "bbox": [ + 377, + 286, + 434, + 329 + ], + "type": "image", + "image_path": "9b2e5b66c65bd12d5768fd1b84def76c75eaddd3c5e81881a207da1580432ec8.jpg" + } + ] + } + ], + "index": 20, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 386, + 330, + 427, + 337 + ], + "lines": [ + { + "bbox": [ + 386, + 330, + 427, + 337 + ], + "spans": [ + { + "bbox": [ + 386, + 330, + 427, + 337 + ], + "type": "text", + "content": "PCM+MuLUT" + } + ] + } + ], + "index": 21, + "angle": 0, + "type": "image_caption" + } + ], + "index": 20 + }, + { + "type": "image", + "bbox": [ + 434, + 286, + 492, + 329 + ], + "blocks": [ + { + "bbox": [ + 434, + 286, + 492, + 329 + ], + "lines": [ + { + "bbox": [ + 434, + 286, + 492, + 329 + ], + "spans": [ + { + "bbox": [ + 434, + 286, + 492, + 329 + ], + "type": "image", + "image_path": "5b7930fff27d398d728ed25e3bb03839620738acc8f7b4604f24acc11d4d4c17.jpg" + } + ] + } + ], + "index": 22, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 442, + 330, + 485, + 337 + ], + "lines": [ + { + "bbox": [ + 442, + 330, + 485, + 337 + ], + "spans": [ + { + "bbox": [ + 442, + 330, + 485, + 337 + ], + "type": "text", + "content": "PCM+RC-LUT" + } + ] + } + ], + "index": 23, + "angle": 0, + "type": "image_caption" + } + ], + "index": 22 + }, + { + "type": "image", + "bbox": [ + 493, + 286, + 551, + 329 + ], + "blocks": [ + { + "bbox": [ + 493, + 286, + 551, + 329 + ], + "lines": [ + { + "bbox": [ + 493, + 286, + 551, + 329 + ], + "spans": [ + { + "bbox": [ + 493, + 286, + 551, + 329 + ], + "type": "image", + "image_path": "82968afa1650114ee52071db974244db9d6b376595baf1b7ced031d047b75821.jpg" + } + ] + } + ], + "index": 24, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 501, + 330, + 544, + 337 + ], + "lines": [ + { + "bbox": [ + 501, + 330, + 544, + 337 + ], + "spans": [ + { + "bbox": [ + 501, + 330, + 544, + 337 + ], + "type": "text", + "content": "PCM+SPFLUT" + } + ] + } + ], + "index": 25, + "angle": 0, + "type": "image_caption" + } + ], + "index": 24 + }, + { + "bbox": [ + 313, + 368, + 554, + 440 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 368, + 554, + 440 + ], + "spans": [ + { + "bbox": [ + 313, + 368, + 554, + 440 + ], + "type": "text", + "content": "cation and addition operations across different data types to estimate total energy consumption. DnLUT achieves " + }, + { + "bbox": [ + 313, + 368, + 554, + 440 + ], + "type": "inline_equation", + "content": "70\\%" + }, + { + "bbox": [ + 313, + 368, + 554, + 440 + ], + "type": "text", + "content": " energy savings compared to SPF-LUT while delivering superior denoising quality. The efficiency gap widens dramatically when compared to DNN-based methods, with DnLUT requiring only " + }, + { + "bbox": [ + 313, + 368, + 554, + 440 + ], + "type": "inline_equation", + "content": "0.1\\%" + }, + { + "bbox": [ + 313, + 368, + 554, + 440 + ], + "type": "text", + "content": " of DnCNN's energy consumption." + } + ] + } + ], + "index": 27 + }, + { + "bbox": [ + 313, + 441, + 554, + 548 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 441, + 554, + 548 + ], + "spans": [ + { + "bbox": [ + 313, + 441, + 554, + 548 + ], + "type": "text", + "content": "Runtime. We implement LUT-based methods using standard JAVA API on the ANDROID platform. DnLUT maintains the rapid inference capabilities characteristic of SR-LUT[10], demonstrating clear advantages over classical and DNN-based approaches. For reference, DnCNN running on mobile CPU requires " + }, + { + "bbox": [ + 313, + 441, + 554, + 548 + ], + "type": "inline_equation", + "content": "20 \\times" + }, + { + "bbox": [ + 313, + 441, + 554, + 548 + ], + "type": "text", + "content": " more processing time. DnLUT's performance can be further optimized through device-specific implementations, such as FPGA deployments." + } + ] + } + ], + "index": 28 + }, + { + "bbox": [ + 313, + 551, + 554, + 598 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 551, + 554, + 598 + ], + "spans": [ + { + "bbox": [ + 313, + 551, + 554, + 598 + ], + "type": "text", + "content": "Storage. DnLUT requires approximately 500 KB storage, well within typical L2-cache limits of modern chips. This efficient memory footprint enables rapid lookup operations with minimal dependency overhead." + } + ] + } + ], + "index": 29 + }, + { + "bbox": [ + 313, + 610, + 474, + 624 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 610, + 474, + 624 + ], + "spans": [ + { + "bbox": [ + 313, + 610, + 474, + 624 + ], + "type": "text", + "content": "4.7. Effectiveness of Plug-in PCM" + } + ] + } + ], + "index": 30 + }, + { + "bbox": [ + 313, + 630, + 554, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 630, + 554, + 714 + ], + "spans": [ + { + "bbox": [ + 313, + 630, + 554, + 714 + ], + "type": "text", + "content": "Tab. 4 demonstrates PCM's versatility as a performance enhancer for existing LUT-based methods. When integrated with current approaches, PCM consistently delivers significant improvements across all benchmark datasets. The visual comparisons in Fig. 7 show marked reduction in color distortion, particularly in smooth image regions, resulting in superior perceptual quality." + } + ] + } + ], + "index": 31 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 749, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 749, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 749, + 315, + 757 + ], + "type": "text", + "content": "7588" + } + ] + } + ], + "index": 32 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 6 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 56, + 102, + 563, + 254 + ], + "blocks": [ + { + "bbox": [ + 55, + 70, + 555, + 94 + ], + "lines": [ + { + "bbox": [ + 55, + 70, + 555, + 94 + ], + "spans": [ + { + "bbox": [ + 55, + 70, + 555, + 94 + ], + "type": "text", + "content": "Table 5. Efficiency evaluation of three categories color image denoising methods. We use the Qualcomm Snapdragon 8 Gen 2 as the mobile platform. The energy cost is calculated on the " + }, + { + "bbox": [ + 55, + 70, + 555, + 94 + ], + "type": "inline_equation", + "content": "{512} \\times {512}" + }, + { + "bbox": [ + 55, + 70, + 555, + 94 + ], + "type": "text", + "content": " resolution color images." + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 56, + 102, + 563, + 254 + ], + "lines": [ + { + "bbox": [ + 56, + 102, + 563, + 254 + ], + "spans": [ + { + "bbox": [ + 56, + 102, + 563, + 254 + ], + "type": "table", + "html": "
CategoryMethodPlatformRuntime(ms) 256 × 256Runtime(ms) 512 × 512Energy Cost (pJ)Storage (KB)
LUT-basedSR-LUT[10]Mobile619149.98M82
MuLUT[12]Mobile2180899.88M489
RC-LUT[6]Mobile1878612.92M326
SPF-LUT[14]Mobile722572.32G30,178
DnLUT(ours)Mobile2388687.34M518
ClassicalCBM3D[29]PC3,18914,3434.82G-
MC-WNNM[24]PC74,290293,42489.23G-
DNNDnCNN[30]Mobile5431,923542.53G2,239
SwinIR[15]Mobile73,234274,45312.03T45,499
", + "image_path": "da2f78a2af32ab21797e18c119a617a607d79e677316a392ce4162e8f6cffb70.jpg" + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "table_body" + } + ], + "index": 1 + }, + { + "bbox": [ + 55, + 262, + 147, + 275 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 262, + 147, + 275 + ], + "spans": [ + { + "bbox": [ + 55, + 262, + 147, + 275 + ], + "type": "text", + "content": "4.8. Ablation study" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 55, + 280, + 295, + 304 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 280, + 295, + 304 + ], + "spans": [ + { + "bbox": [ + 55, + 280, + 295, + 304 + ], + "type": "text", + "content": "We conduct ablation studies using Gaussian color image denoising benchmark to validate DnLUT's key components." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 55, + 311, + 225, + 322 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 311, + 225, + 322 + ], + "spans": [ + { + "bbox": [ + 55, + 311, + 225, + 322 + ], + "type": "text", + "content": "4.8.1. Channel-spatial-wise convolution" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 55, + 326, + 296, + 397 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 326, + 296, + 397 + ], + "spans": [ + { + "bbox": [ + 55, + 326, + 296, + 397 + ], + "type": "text", + "content": "Channel indexing[13, 25] possesses a " + }, + { + "bbox": [ + 55, + 326, + 296, + 397 + ], + "type": "inline_equation", + "content": "1 \\times 1" + }, + { + "bbox": [ + 55, + 326, + 296, + 397 + ], + "type": "text", + "content": " receptive field to extract channel information. We evaluate various combinations of channel-wise, spatial-wise, and our proposed channel-spatial-wise kernels using a consistent architecture comprising one channel indexing module, two L-shaped convolution modules, and one PCM component." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 55, + 398, + 296, + 506 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 398, + 296, + 506 + ], + "spans": [ + { + "bbox": [ + 55, + 398, + 296, + 506 + ], + "type": "text", + "content": "Results in Tab. 6 reveal that naive combination of channel-wise and spatial-wise convolutions yields suboptimal performance. Pure channel-spatial-wise convolution, while promising, struggles with limited receptive field coverage. The integration of spatial-wise and channel-spatial-wise convolutions produces substantial improvements. Notably, adding channel-wise convolution offers minimal additional benefit, as channel-spatial-wise convolution already captures comprehensive color information." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 55, + 510, + 176, + 522 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 510, + 176, + 522 + ], + "spans": [ + { + "bbox": [ + 55, + 510, + 176, + 522 + ], + "type": "text", + "content": "4.8.2. L-shaped convolution" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 55, + 525, + 295, + 610 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 525, + 295, + 610 + ], + "spans": [ + { + "bbox": [ + 55, + 525, + 295, + 610 + ], + "type": "text", + "content": "To validate our L-shaped convolution design, we compare it against standard " + }, + { + "bbox": [ + 55, + 525, + 295, + 610 + ], + "type": "inline_equation", + "content": "2 \\times 2" + }, + { + "bbox": [ + 55, + 525, + 295, + 610 + ], + "type": "text", + "content": " spatial-wise kernels (denoted as S in Tab. 7, " + }, + { + "bbox": [ + 55, + 525, + 295, + 610 + ], + "type": "inline_equation", + "content": "\\sigma = 1" + }, + { + "bbox": [ + 55, + 525, + 295, + 610 + ], + "type": "text", + "content": "). While the conventional kernel shows marginal performance advantages, our L-shaped design achieves comparable results while enabling 3D LUT conversion, reducing storage requirements by " + }, + { + "bbox": [ + 55, + 525, + 295, + 610 + ], + "type": "inline_equation", + "content": "17 \\times" + }, + { + "bbox": [ + 55, + 525, + 295, + 610 + ], + "type": "text", + "content": ". It paves the way of deploying multi-layer spatial-wise convolutions." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 55, + 621, + 128, + 634 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 621, + 128, + 634 + ], + "spans": [ + { + "bbox": [ + 55, + 621, + 128, + 634 + ], + "type": "text", + "content": "5. Conclusion" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 55, + 641, + 296, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 641, + 296, + 714 + ], + "spans": [ + { + "bbox": [ + 55, + 641, + 296, + 714 + ], + "type": "text", + "content": "This paper presents DnLUT, a novel approach to color image denoising that bridges the gap between high-performance deep learning models and resource-constrained edge devices. Addressing the limitations of existing LUT-based methods, DnLUT introduces two key innovations. First, the Pairwise Channel Mixer (PCM) models channel and spa" + } + ] + } + ], + "index": 10 + }, + { + "type": "table", + "bbox": [ + 315, + 307, + 556, + 387 + ], + "blocks": [ + { + "bbox": [ + 313, + 261, + 555, + 305 + ], + "lines": [ + { + "bbox": [ + 313, + 261, + 555, + 305 + ], + "spans": [ + { + "bbox": [ + 313, + 261, + 555, + 305 + ], + "type": "text", + "content": "Table 6. Different combinations of channel-wise, spatial-wise, and our channel-spatial-wise kernels. These kernels are represented by channel indexing[13], L-shaped kernel, and PCM. The table presents CPSNR(dB) values." + } + ] + } + ], + "index": 11, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 315, + 307, + 556, + 387 + ], + "lines": [ + { + "bbox": [ + 315, + 307, + 556, + 387 + ], + "spans": [ + { + "bbox": [ + 315, + 307, + 556, + 387 + ], + "type": "table", + "html": "
ChannelSpatialChannel-spatialCBSD68
✓29.77
✓30.15
✓✓30.21
✓✓31.41
✓✓✓31.47
", + "image_path": "aae33e99c75bd08e03a3d3ba9e7ea939517e99bcd2332971df92d013d9c0bcad.jpg" + } + ] + } + ], + "index": 12, + "angle": 0, + "type": "table_body" + } + ], + "index": 12 + }, + { + "type": "table", + "bbox": [ + 317, + 413, + 555, + 454 + ], + "blocks": [ + { + "bbox": [ + 314, + 389, + 554, + 412 + ], + "lines": [ + { + "bbox": [ + 314, + 389, + 554, + 412 + ], + "spans": [ + { + "bbox": [ + 314, + 389, + 554, + 412 + ], + "type": "text", + "content": "Table 7. Comparison of L-shaped kernel and " + }, + { + "bbox": [ + 314, + 389, + 554, + 412 + ], + "type": "inline_equation", + "content": "2 \\times 2" + }, + { + "bbox": [ + 314, + 389, + 554, + 412 + ], + "type": "text", + "content": " spatial-wise kernel. The table presents CPSNR(dB) values." + } + ] + } + ], + "index": 13, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 317, + 413, + 555, + 454 + ], + "lines": [ + { + "bbox": [ + 317, + 413, + 555, + 454 + ], + "spans": [ + { + "bbox": [ + 317, + 413, + 555, + 454 + ], + "type": "table", + "html": "
MethodCBSD68KodakMcMasterLUT Size(KB)
PCM + L26.6327.0527.96494 + 24
PCM + S26.7127.1628.01494 + 408
", + "image_path": "c0a1c9b1c92b6d07338bc05a56d760b6fdf980d80a8adbe254b801209319dc6d.jpg" + } + ] + } + ], + "index": 14, + "angle": 0, + "type": "table_body" + } + ], + "index": 14 + }, + { + "bbox": [ + 313, + 468, + 555, + 577 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 468, + 555, + 577 + ], + "spans": [ + { + "bbox": [ + 313, + 468, + 555, + 577 + ], + "type": "text", + "content": "tial correlations simultaneously by processing channel pairs, enabling comprehensive color modeling without excessive storage costs. Second, an L-shaped convolution design rethinks rotation-based receptive field expansion, eliminating redundant pixel usage and converting 4D to 3D LUTs, reducing storage by " + }, + { + "bbox": [ + 313, + 468, + 555, + 577 + ], + "type": "inline_equation", + "content": "17\\times" + }, + { + "bbox": [ + 313, + 468, + 555, + 577 + ], + "type": "text", + "content": " while maintaining spatial coverage. These advancements allow DnLUT to outperform state-ofthe-art LUT-based methods by over 1dB in denoising quality while consuming just " + }, + { + "bbox": [ + 313, + 468, + 555, + 577 + ], + "type": "inline_equation", + "content": "0.1\\%" + }, + { + "bbox": [ + 313, + 468, + 555, + 577 + ], + "type": "text", + "content": " of DnCNN's energy." + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 313, + 580, + 556, + 712 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 580, + 556, + 712 + ], + "spans": [ + { + "bbox": [ + 313, + 580, + 556, + 712 + ], + "type": "text", + "content": "Acknowledgements. This research was supported by: the HKU-TCL Joint Research Centre for AI, the Theme-based Research Scheme (TRS) project T45-701/22-R, and the General Research Fund (GRF) Project 17203224 from the Research Grants Council (RGC) of Hong Kong SAR. Additional support was provided by the National Key Research and Development Program of China (Grant No. 2024YFB2808903), the Shenzhen Science and Technology Program (JCYJ20220818101001004), and the Tsinghua University - Tencent Joint Laboratory for Internet Innovation Technology Research Fund." + } + ] + } + ], + "index": 16 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "text", + "content": "7589" + } + ] + } + ], + "index": 17 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 7 + }, + { + "para_blocks": [ + { + "bbox": [ + 56, + 71, + 115, + 83 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 71, + 115, + 83 + ], + "spans": [ + { + "bbox": [ + 56, + 71, + 115, + 83 + ], + "type": "text", + "content": "References" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 57, + 91, + 296, + 713 + ], + "type": "list", + "angle": 0, + "index": 14, + "blocks": [ + { + "bbox": [ + 61, + 91, + 296, + 135 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 61, + 91, + 296, + 135 + ], + "spans": [ + { + "bbox": [ + 61, + 91, + 296, + 135 + ], + "type": "text", + "content": "[1] Eirikur Agustsson and Radu Timofte. Ntire 2017 challenge on single image super-resolution: Dataset and study. In The IEEE Conference on Computer Vision and Pattern Recognition (CVPR) Workshops, 2017. 5" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 61, + 137, + 296, + 192 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 61, + 137, + 296, + 192 + ], + "spans": [ + { + "bbox": [ + 61, + 137, + 296, + 192 + ], + "type": "text", + "content": "[2] Hanting Chen, Yunhe Wang, Tianyu Guo, Chang Xu, Yiping Deng, Zhenhua Liu, Siwei Ma, Chunjing Xu, Chao Xu, and Wen Gao. Pre-trained image processing transformer. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 12299-12310, 2021. 1, 2" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 62, + 194, + 296, + 248 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 194, + 296, + 248 + ], + "spans": [ + { + "bbox": [ + 62, + 194, + 296, + 248 + ], + "type": "text", + "content": "[3] Kostadin Dabov, Alessandro Foi, Vladimir Katkovnik, and Karen Egiazarian. Color image denoising via sparse 3d collaborative filtering with grouping constraint in luminance-chrominance space. In 2007 IEEE international conference on image processing, pages I-313. IEEE, 2007. 2, 6, 7" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 62, + 250, + 296, + 293 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 250, + 296, + 293 + ], + "spans": [ + { + "bbox": [ + 62, + 250, + 296, + 293 + ], + "type": "text", + "content": "[4] Kostadin Dabov, Alessandro Foi, Vladimir Katkovnik, and Karen Egiazarian. Image denoising by sparse 3-d transform-domain collaborative filtering. IEEE Transactions on image processing, 16(8):2080-2095, 2007. 2" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 62, + 295, + 296, + 338 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 295, + 296, + 338 + ], + "spans": [ + { + "bbox": [ + 62, + 295, + 296, + 338 + ], + "type": "text", + "content": "[5] Jingjing Dai, Oscar C Au, Lu Fang, Chao Pang, Feng Zou, and Jiali Li. Multichannel nonlocal means fusion for color image denoising. IEEE Transactions on circuits and systems for video technology, 23(11):1873-1886, 2013. 2" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 62, + 341, + 296, + 384 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 341, + 296, + 384 + ], + "spans": [ + { + "bbox": [ + 62, + 341, + 296, + 384 + ], + "type": "text", + "content": "[6] Michael Gharbi, Jiawen Chen, Jonathan T Barron, Samuel W Hasinoff, and Fredo Durand. Deep bilateral learning for real-time image enhancement. ACM Transactions on Graphics (TOG), 36(4):1-12, 2017. 8" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 62, + 386, + 296, + 440 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 386, + 296, + 440 + ], + "spans": [ + { + "bbox": [ + 62, + 386, + 296, + 440 + ], + "type": "text", + "content": "[7] Shuhang Gu, Lei Zhang, Wangmeng Zuo, and Xiangchu Feng. Weighted nuclear norm minimization with application to image denoising. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 2862-2869, 2014. 2" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 62, + 442, + 296, + 485 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 442, + 296, + 485 + ], + "spans": [ + { + "bbox": [ + 62, + 442, + 296, + 485 + ], + "type": "text", + "content": "[8] Binxiao Huang, Jason Chun Lok Li, Jie Ran, Boyu Li, Jiajun Zhou, Dahai Yu, and Ngai Wong. Hundred-kilobyte lookup tables for efficient single-image super-resolution. arXiv preprint arXiv:2312.06101, 2023. 3" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 62, + 487, + 296, + 531 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 487, + 296, + 531 + ], + "spans": [ + { + "bbox": [ + 62, + 487, + 296, + 531 + ], + "type": "text", + "content": "[9] Jia-Bin Huang, Abhishek Singh, and Narendra Ahuja. Single image super-resolution from transformed self-exemplars. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 5197-5206, 2015. 5" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 57, + 533, + 296, + 577 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 533, + 296, + 577 + ], + "spans": [ + { + "bbox": [ + 57, + 533, + 296, + 577 + ], + "type": "text", + "content": "[10] Younghyun Jo and Seon Joo Kim. Practical single-image super-resolution using look-up table. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 691-700, 2021. 2, 3, 5, 6, 7, 8" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 57, + 578, + 296, + 622 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 578, + 296, + 622 + ], + "spans": [ + { + "bbox": [ + 57, + 578, + 296, + 622 + ], + "type": "text", + "content": "[11] Stamatios Lefkimmiatis. Non-local color image denoising with convolutional neural networks. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 3587-3596, 2017. 2" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 57, + 624, + 296, + 668 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 624, + 296, + 668 + ], + "spans": [ + { + "bbox": [ + 57, + 624, + 296, + 668 + ], + "type": "text", + "content": "[12] Jiacheng Li, Chang Chen, Zhen Cheng, and Zhiwei Xiong. Mulut: Cooperating multiple look-up tables for efficient image super-resolution. In European Conference on Computer Vision, pages 238-256. Springer, 2022. 2, 3, 5, 6, 7, 8" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 57, + 670, + 296, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 670, + 296, + 713 + ], + "spans": [ + { + "bbox": [ + 57, + 670, + 296, + 713 + ], + "type": "text", + "content": "[13] Jiacheng Li, Chang Chen, Zhen Cheng, and Zhiwei Xiong. Toward dnn of luts: Learning efficient image restoration with multiple look-up tables. IEEE Transactions on Pattern Analysis and Machine Intelligence, 2024. 2, 8" + } + ] + } + ], + "index": 13 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 57, + 73, + 553, + 713 + ], + "type": "list", + "angle": 0, + "index": 28, + "blocks": [ + { + "bbox": [ + 316, + 73, + 553, + 117 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 73, + 553, + 117 + ], + "spans": [ + { + "bbox": [ + 316, + 73, + 553, + 117 + ], + "type": "text", + "content": "[14] Yinglong Li, Jiacheng Li, and Zhiwei Xiong. Look-up table compression for efficient image restoration. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 26016-26025, 2024. 2, 3, 6, 7, 8" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 317, + 118, + 553, + 171 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 118, + 553, + 171 + ], + "spans": [ + { + "bbox": [ + 317, + 118, + 553, + 171 + ], + "type": "text", + "content": "[15] Jingyun Liang, Jiezhang Cao, Guolei Sun, Kai Zhang, Luc Van Gool, and Radu Timofte. Swinir: Image restoration using swin transformer. In Proceedings of the IEEE/CVF international conference on computer vision, pages 1833-1844, 2021. 1, 2, 6, 7, 8" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 317, + 173, + 553, + 228 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 173, + 553, + 228 + ], + "spans": [ + { + "bbox": [ + 317, + 173, + 553, + 228 + ], + "type": "text", + "content": "[16] Guandu Liu, Yukang Ding, Mading Li, Ming Sun, Xing Wen, and Bin Wang. Reconstructed convolution module based look-up tables for efficient image super-resolution. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 12217-12226, 2023. 2, 3, 6, 7" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 317, + 228, + 553, + 271 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 228, + 553, + 271 + ], + "spans": [ + { + "bbox": [ + 317, + 228, + 553, + 271 + ], + "type": "text", + "content": "[17] Cheng Ma, Jingyi Zhang, Jie Zhou, and Jiwen Lu. Learning series-parallel lookup tables for efficient image superresolution. In European Conference on Computer Vision, pages 305–321. Springer, 2022. 3" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 317, + 272, + 553, + 326 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 272, + 553, + 326 + ], + "spans": [ + { + "bbox": [ + 317, + 272, + 553, + 326 + ], + "type": "text", + "content": "[18] Kede Ma, Zhengfang Duanmu, Qingbo Wu, Zhou Wang, Hongwei Yong, Hongliang Li, and Lei Zhang. Waterloo exploration database: New challenges for image quality assessment models. IEEE Transactions on Image Processing, 26 (2):1004-1016, 2016. 5" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 316, + 327, + 553, + 360 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 327, + 553, + 360 + ], + "spans": [ + { + "bbox": [ + 316, + 327, + 553, + 360 + ], + "type": "text", + "content": "[19] Julien Mairal, Michael Elad, and Guillermo Sapiro. Sparse representation for color image restoration. IEEE Transactions on image processing, 17(1):53-69, 2007. 2" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 317, + 361, + 553, + 414 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 361, + 553, + 414 + ], + "spans": [ + { + "bbox": [ + 317, + 361, + 553, + 414 + ], + "type": "text", + "content": "[20] D. Martin, C. Fowlkes, D. Tal, and J. Malik. A database of human segmented natural images and its application to evaluating segmentation algorithms and measuring ecological statistics. In Proc. 8th Int'l Conf. Computer Vision, pages 416-423, 2001. 5" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 317, + 415, + 553, + 480 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 415, + 553, + 480 + ], + "spans": [ + { + "bbox": [ + 317, + 415, + 553, + 480 + ], + "type": "text", + "content": "[21] David Martin, Charless Fowlkes, Doron Tal, and Jitendra Malik. A database of human segmented natural images and its application to evaluating segmentation algorithms and measuring ecological statistics. In Proceedings eighth IEEE international conference on computer vision. ICCV 2001, pages 416-423. IEEE, 2001. 5" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 317, + 482, + 553, + 536 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 482, + 553, + 536 + ], + "spans": [ + { + "bbox": [ + 317, + 482, + 553, + 536 + ], + "type": "text", + "content": "[22] Dehua Song, Yunhe Wang, Hanting Chen, Chang Xu, Chunjing Xu, and DaCheng Tao. Addersr: Towards energy efficient image super-resolution. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 15648-15657, 2021. 7" + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 317, + 537, + 553, + 592 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 537, + 553, + 592 + ], + "spans": [ + { + "bbox": [ + 317, + 537, + 553, + 592 + ], + "type": "text", + "content": "[23] Radu Timofte, Eirikur Agustsson, Luc Van Gool, Ming-Hsuan Yang, and Lei Zhang. Ntire 2017 challenge on single image super-resolution: Methods and results. In Proceedings of the IEEE conference on computer vision and pattern recognition workshops, pages 114-125, 2017. 5" + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 317, + 593, + 553, + 645 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 593, + 553, + 645 + ], + "spans": [ + { + "bbox": [ + 317, + 593, + 553, + 645 + ], + "type": "text", + "content": "[24] Jun Xu, Lei Zhang, David Zhang, and Xiangchu Feng. Multichannel weighted nuclear norm minimization for real color image denoising. In Proceedings of the IEEE international conference on computer vision, pages 1096-1104, 2017. 2, 6, 7, 8" + } + ] + } + ], + "index": 25 + }, + { + "bbox": [ + 317, + 647, + 553, + 689 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 647, + 553, + 689 + ], + "spans": [ + { + "bbox": [ + 317, + 647, + 553, + 689 + ], + "type": "text", + "content": "[25] Sidi Yang, Binxiao Huang, Mingdeng Cao, Yatai Ji, Hanzhong Guo, Ngai Wong, and Yujiu Yang. Taming lookup tables for efficient image retouching. arXiv preprint arXiv:2403.19238, 2024. 2, 8" + } + ] + } + ], + "index": 26 + }, + { + "bbox": [ + 316, + 691, + 553, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 691, + 553, + 713 + ], + "spans": [ + { + "bbox": [ + 316, + 691, + 553, + 713 + ], + "type": "text", + "content": "[26] Syed Waqas Zamir, Aditya Arora, Salman Khan, Munawar Hayat, Fahad Shahbaz Khan, Ming-Hsuan Yang, and Ling" + } + ] + } + ], + "index": 27 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 748, + 316, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 316, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 316, + 757 + ], + "type": "text", + "content": "7590" + } + ] + } + ], + "index": 29 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 8 + }, + { + "para_blocks": [ + { + "bbox": [ + 56, + 72, + 295, + 418 + ], + "type": "list", + "angle": 0, + "index": 7, + "blocks": [ + { + "bbox": [ + 75, + 72, + 294, + 106 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 75, + 72, + 294, + 106 + ], + "spans": [ + { + "bbox": [ + 75, + 72, + 294, + 106 + ], + "type": "text", + "content": "Shao. Multi-stage progressive image restoration. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 14821-14831, 2021. 2" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 56, + 107, + 295, + 172 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 107, + 295, + 172 + ], + "spans": [ + { + "bbox": [ + 56, + 107, + 295, + 172 + ], + "type": "text", + "content": "[27] Syed Waqas Zamir, Aditya Arora, Salman Khan, Munawar Hayat, Fahad Shahbaz Khan, and Ming-Hsuan Yang. Restormer: Efficient transformer for high-resolution image restoration. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 5728-5739, 2022. 1, 2, 5" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 56, + 174, + 294, + 228 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 174, + 294, + 228 + ], + "spans": [ + { + "bbox": [ + 56, + 174, + 294, + 228 + ], + "type": "text", + "content": "[28] Hui Zeng, Jianrui Cai, Lida Li, Zisheng Cao, and Lei Zhang. Learning image-adaptive 3d lookup tables for high performance photo enhancement in real-time. IEEE Transactions on Pattern Analysis and Machine Intelligence, 44(4):2058-2073, 2020. 2" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 56, + 229, + 294, + 284 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 229, + 294, + 284 + ], + "spans": [ + { + "bbox": [ + 56, + 229, + 294, + 284 + ], + "type": "text", + "content": "[29] Fengyi Zhang, Hui Zeng, Tianjun Zhang, and Lin Zhang. Clut-net: Learning adaptively compressed representations of 3dluts for lightweight image enhancement. In Proceedings of the 30th ACM International Conference on Multimedia, pages 6493-6501, 2022. 8" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 56, + 285, + 294, + 329 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 285, + 294, + 329 + ], + "spans": [ + { + "bbox": [ + 56, + 285, + 294, + 329 + ], + "type": "text", + "content": "[30] Kai Zhang, Wangmeng Zuo, Yunjin Chen, Deyu Meng, and Lei Zhang. Beyond a gaussian denoiser: Residual learning of deep cnn for image denoising. IEEE transactions on image processing, 26(7):3142-3155, 2017. 1, 2, 6, 7, 8" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 56, + 331, + 294, + 373 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 331, + 294, + 373 + ], + "spans": [ + { + "bbox": [ + 56, + 331, + 294, + 373 + ], + "type": "text", + "content": "[31] Lei Zhang, Xiaolin Wu, Antoni Buades, and Xin Li. Color demosaicking by local directional interpolation and nonlocal adaptive thresholding. Journal of Electronic imaging, 20(2): 023016-023016, 2011. 5" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 56, + 375, + 294, + 418 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 375, + 294, + 418 + ], + "spans": [ + { + "bbox": [ + 56, + 375, + 294, + 418 + ], + "type": "text", + "content": "[32] Yi Zhang, Dasong Li, Xiaoyu Shi, Dailan He, Kangning Song, Xiaogang Wang, Hongwei Qin, and Hongsheng Li. Kbnet: Kernel basis network for image restoration. arXiv preprint arXiv:2303.02881, 2023. 1, 2" + } + ] + } + ], + "index": 6 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 749, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 749, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 749, + 315, + 757 + ], + "type": "text", + "content": "7591" + } + ] + } + ], + "index": 8 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 9 + } + ], + "_backend": "vlm", + "_version_name": "2.6.4" +} \ No newline at end of file diff --git a/2025/Do Computer Vision Foundation Models Learn the Low-level Characteristics of the Human Visual System_/4607e0a5-8aa8-470e-b0e4-8cfed36f43c4_content_list.json b/2025/Do Computer Vision Foundation Models Learn the Low-level Characteristics of the Human Visual System_/4607e0a5-8aa8-470e-b0e4-8cfed36f43c4_content_list.json new file mode 100644 index 0000000000000000000000000000000000000000..2f0dddef76bda5bdfeffb142cfafa7c9f06862e4 --- /dev/null +++ b/2025/Do Computer Vision Foundation Models Learn the Low-level Characteristics of the Human Visual System_/4607e0a5-8aa8-470e-b0e4-8cfed36f43c4_content_list.json @@ -0,0 +1,1889 @@ +[ + { + "type": "text", + "text": "Do computer vision foundation models learn the low-level characteristics of the human visual system?", + "text_level": 1, + "bbox": [ + 212, + 104, + 782, + 150 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Yancheng Cai, Fei Yin, Dounia Hammou, Rafal Mantiuk; University of Cambridge, UK", + "bbox": [ + 151, + 178, + 844, + 196 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "yc613@cam.ac.uk, fy277@cam.ac.uk, dh706@cam.ac.uk, mantiuk@gmail.com", + "bbox": [ + 204, + 198, + 785, + 213 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Abstract", + "text_level": 1, + "bbox": [ + 246, + 248, + 326, + 263 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Computer vision foundation models, such as DINO or OpenCLIP, are trained in a self-supervised manner on large image datasets. Analogously, substantial evidence suggests that the human visual system (HVS) is influenced by the statistical distribution of colors and patterns in the natural world, characteristics also present in the training data of foundation models. The question we address in this paper is whether foundation models trained on natural images mimic some of the low-level characteristics of the human visual system, such as contrast detection, contrast masking, and contrast constancy. Specifically, we designed a protocol comprising nine test types to evaluate the image encoders of 45 foundation and generative models. Our results indicate that some foundation models (e.g., DINO, DINOv2, and OpenCLIP), share some of the characteristics of human vision, but other models show little resemblance. Foundation models tend to show smaller sensitivity to low contrast and rather irregular responses to contrast across frequencies. The foundation models show the best agreement with human data in terms of contrast masking. Our findings suggest that human vision and computer vision may take both similar and different paths when learning to interpret images of the real world. Overall, while differences remain, foundation models trained on vision tasks start to align with low-level human vision, with DINOv2 showing the closest resemblance. Our code is available on https://github.com/caiyancheng/VFM_HVS_CVPR2025.", + "bbox": [ + 89, + 281, + 483, + 690 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "1. Introduction", + "text_level": 1, + "bbox": [ + 89, + 722, + 220, + 738 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Computer vision foundation models, such as DINO [9] or OpenCLIP [22, 33], show exceptional ability to generalize to different tasks and are becoming cornerstones of many computer vision methods. They owe their exceptional performance to self-supervised training on very large image datasets. The human visual system also owes much of its capability to being able to perceive the world, over many years from infancy to childhood [6]. A question arises: if the neural network and the visual system are trained by being exposed to a large number of images of the world, will", + "bbox": [ + 89, + 750, + 483, + 900 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/69b5a95868ebd18905940c43725a8138c1fd718a2616f30c77c952d26ffde08e.jpg", + "image_caption": [ + "Figure 1. To determine whether image encoders of foundation models exhibit a similar low-level characteristic as human vision, we test them on psychophysical stimuli for which human data is available. We want to test the alignment of contrast encoding between human and computational vision models." + ], + "image_footnote": [], + "bbox": [ + 516, + 234, + 906, + 467 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "they share their low-level vision characteristics? If they do, we will know that those low-level characteristics arise naturally and likely reflect the statistics of real-world scenes. If they do not, it means that human low-level vision characteristics are specific to the optical/biological limitations of human vision rather than natural image statistics. Our analysis is meant to shed some light on how the vision, either biological or computational, may develop from observing samples of the world, taking either the same or different routes to accomplish their respective tasks.", + "bbox": [ + 511, + 579, + 906, + 731 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "In particular, we are interested in the characteristics that are well understood and measured in human vision science using psychophysical methods: contrast detection [4], contrast masking [25] and contrast constancy [18]. Contrast detection and contrast masking quantify the ability of the visual system to detect small contrast patterns, either on uniform backgrounds (contrast detection) or on backgrounds with patterns (contrast masking). Contrast detection and masking capture the \"bottlenecks\" of the visual system — the characteristic that may prevent us from detecting patterns that are too dark or too small. Similarly, cameras used", + "bbox": [ + 511, + 734, + 908, + 902 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "CVF", + "bbox": [ + 106, + 2, + 181, + 42 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore.", + "bbox": [ + 236, + 0, + 810, + 46 + ], + "page_idx": 0 + }, + { + "type": "page_number", + "text": "20039", + "bbox": [ + 478, + 944, + 517, + 955 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "for computer vision are limited by the MTF of the lens, sensor resolution, photon and sensor noise, and we can expect that computer vision methods may need to deal with similar limitations.", + "bbox": [ + 89, + 90, + 480, + 150 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Contrast constancy is the term used in vision science to describe the invariance of the visual system to spatial frequency [18] and partially luminance [24, 32]. Georgeson and Sullivan [18] showed that the perceived magnitude of the contrast that is well above the detection threshold (supra-threshold) appears to us the same regardless of spatial frequency. This is a very important characteristic as it allows us to see contrast (and therefore objects) the same regardless of the viewing distance; otherwise, the frequencies would change with the viewing distance and hence the contrast appearance. A partial constancy (invariance) is also observed across luminance [24, 32], though there is a significant deviation from constancy at lower luminance levels, once the visual system needs to rely on the rod vision. The invariance is also an important feature of many computer vision methods. For example in SIFT features [27] have been designed to be invariant to the changes in contrast, brightness, scale, and rotation. In our experiments, we used the supra-threshold contrast matching test to assess whether the models exhibit the characteristic of contrast constancy.", + "bbox": [ + 91, + 152, + 482, + 455 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Numerous works on adversarial attacks demonstrated that the classification performance of deep learning models can be greatly degraded by visually inconsequential changes [20]. At the same time, human vision does not suffer from such adversarial vulnerability [44]. This is one of the most salient arguments put forward to state that deep architectures are different from human vision. Here, we propose a different methodology to study this question. We consider the deep neural network to be a black box and compare its responses to well-understood and measured characteristics of the human visual system. In particular, we want to check whether the foundation vision models share the same \"bottlenecks\" and invariance properties as the visual system. To achieve this, we test foundation models on basic vision stimuli, such as Gabor patches and band-limited noise, and compare the response of those models with the psychophysical data collected from human observers.", + "bbox": [ + 89, + 455, + 482, + 712 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "In summary, our contributions are as follows:", + "bbox": [ + 109, + 715, + 410, + 729 + ], + "page_idx": 1 + }, + { + "type": "list", + "sub_type": "text", + "list_items": [ + "- We developed a protocol to evaluate the similarity between machine vision models and the human visual system. This protocol includes contrast detection, contrast masking, and contrast constancy, subdivided into nine distinct test types that collectively capture the low-level fundamental characteristics of human vision.", + "- We tested the image encoders of 45 foundation and generative models. The results reveal similarities between certain foundation models (e.g., DINOv2 and OpenCLIP) and human vision, particularly in the contrast masking test. However, differences persist across other tests." + ], + "bbox": [ + 91, + 732, + 482, + 897 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "2. Related Work", + "text_level": 1, + "bbox": [ + 513, + 89, + 653, + 104 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Since the advent of deep learning, machine vision models based on foundation models [13, 23, 31] and DNNs [28, 29, 39, 47] have successfully handled numerous advanced visual tasks. However, researchers have observed that machine vision operates differently from human vision. [16] revealed that standard CNNs trained on ImageNet are strongly biased toward texture recognition rather than shape, which contrasts with human visual patterns. [44] provided further evidence that deep neural networks (DNNs) differ significantly from the HVS, demonstrating poor robustness in object classification under 3D viewpoint changes and image distortions, and showing vulnerability to adversarial examples, which are rarely problematic for humans. [5] pointed out that DNNs performing well in benchmark tests share little overlap with biological vision mechanisms and fail to account for many findings in psychological studies of human vision. This highlights a clear distinction between machine and human vision, leading to the rise of interest in domain adaptation [7] and making networks robust to adversarial attacks [46].", + "bbox": [ + 511, + 114, + 903, + 417 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "But there is also evidence that the gap between neural network-based machine vision models and human vision is gradually narrowing. [40] compared vision transformers (ViT) [13] and CNNs, finding that ViT not only achieves superior task accuracy but also exhibits weaker inductive biases, with error patterns more consistent with human errors. [17] discovered that the long-standing robustness gap between humans and CNNs in handling distortions is shrinking. [10, 19] also demonstrated that foundation models like DINO [9] and CLIP [34] can generate more accurate and robust metrics for low-level perceptual similarity.", + "bbox": [ + 511, + 419, + 903, + 584 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Most of the aforementioned studies focus on high-level task performance (e.g., accuracy, consistency, ...), which may not reveal whether computation models suffer from the same bottlenecks and rely on the same invariances as human vision. To that end, [2, 26] have attempted to reveal CSF characteristics within pretrained architectures by training a head with a contrast discrimination classifier. The problem with this approach is that it introduces a bias by relying on a classifier trained to compare contrast. Such studies also make an incorrect assumption that CSF explains both near-threshold and super-threshold vision, while contrast constancy results (see Section 4.3) show that this is not the case. In contrast, we examine networks' low-level characteristics without additional task-specific training, considering both near-threshold and supra-threshold vision.", + "bbox": [ + 511, + 585, + 903, + 811 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "3. Testing framework", + "text_level": 1, + "bbox": [ + 513, + 828, + 694, + 845 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "We first explain the tested models, testing methods, result visualization, and the strategy we used to summarize and quantify our results.", + "bbox": [ + 511, + 854, + 903, + 900 + ], + "page_idx": 1 + }, + { + "type": "page_number", + "text": "20040", + "bbox": [ + 478, + 945, + 517, + 955 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "3.1. Tested models and testing methodology", + "text_level": 1, + "bbox": [ + 89, + 90, + 426, + 107 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "The objective is to evaluate the responses of machine vision foundation models to stimuli commonly used in human vision research [3, 8] and to compare these responses with psychophysical human data. We tested a representative set of 45 models, encompassing the most influential large vision foundation models, including the variants of DINO [9], DINOv2 [12, 31], OpenCLIP [22, 33], SAM [23], SAM2 [35], and MAE [21], as well as the encoder used for the latent space of generative model Stable Diffusion (SDVAE, [36]). Additionally, we report the responses of ColorVideoVDP [30], which is an image and video quality metric that explicitly models low-level human vision and acts as a reference for a low-level human vision model. All models and their variants are listed in Figure 6.", + "bbox": [ + 89, + 112, + 483, + 323 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "To test these models, we need to compare pairs of images and assess the \"perceived\" difference between them. We adopt a methodology inspired by 2-alternative-forced-choice (2AFC) psychophysical experiments. For example, for a pattern detection task, a pair of images could be a Gabor patch (test) and a uniform field (reference), as shown on the left of Figure 2. As such patterns are calibrated in physical light units in vision science, we generate these patterns as luminance maps, scaled in physical units of $\\mathrm{cd} / \\mathrm{m}^2$ . These luminance maps are then mapped from the linear space to the sRGB color space using a display model (display peak luminance of $400\\mathrm{cd} / \\mathrm{m}^2$ ) and fed into the image encoder of the foundation model for feature extraction. Note that the sRGB space is almost universally used to represent training datasets and is expected input for the tested encoders. To ensure that models can operate on small contrast values, we modified them to accept floating-point values (instead of 8-bit integers) as input. This was necessary as the quantization artifacts in 8-bit images are often larger than the detection thresholds of human vision.", + "bbox": [ + 89, + 324, + 483, + 625 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "To investigate whether the distances in the feature space reflect the perceptual detection thresholds and invariances, we experimented with a series of distance measures, including $L_{1}$ and $L_{2}$ . We found the cosine similarity expressed as a relative angle $(S_{\\mathrm{ac}})$ yielded results most consistent with the psychophysical data. $S_{\\mathrm{ac}}$ is defined as:", + "bbox": [ + 89, + 626, + 483, + 717 + ], + "page_idx": 2 + }, + { + "type": "equation", + "text": "\n$$\nS _ {\\mathrm {a c}} = \\frac {1}{\\pi} \\arccos \\left(\\frac {F _ {\\mathrm {T}} \\cdot F _ {\\mathrm {R}}}{\\| F _ {\\mathrm {T}} \\| \\| F _ {\\mathrm {R}} \\|}\\right), \\tag {1}\n$$\n", + "text_format": "latex", + "bbox": [ + 176, + 724, + 482, + 758 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "where $F_{\\mathrm{T}}$ and $F_{\\mathrm{R}}$ are the test and reference feature vectors (feature maps reshaped into one dimension), and $\\cdot$ denotes the dot product. $S_{\\mathrm{ac}} = 0$ indicates two input images are equivalent, while $S_{\\mathrm{ac}} = 1$ indicates large differences.", + "bbox": [ + 89, + 763, + 482, + 825 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "To compare the encoder responses with psychophysical data, we need to be able to map the image sampling frequency into the spatial frequency on the retina. For that, we select the effective resolution of 60 pixels-per-degree, which is typical for modern monitors. We note, however,", + "bbox": [ + 89, + 825, + 483, + 900 + ], + "page_idx": 2 + }, + { + "type": "image", + "img_path": "images/7d0f4d354cce31ca077243be794d0ba18e4e88baa1de94f93634af4746385a0a.jpg", + "image_caption": [ + "Figure 2. Pipeline for computing $S_{\\mathrm{ac}}$ . Generate test and reference images in the linear luminance space (cd/m²), transform them to the sRGB color space using a display model, and input each into the image encoder. Reshape the output features into one-dimensional vectors $F_{\\mathrm{T}}$ and $F_{\\mathrm{R}}$ , then compute $S_{\\mathrm{ac}}$ ." + ], + "image_footnote": [], + "bbox": [ + 517, + 88, + 906, + 276 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "that the choice of this parameter is arbitrary and the model similarity scores can be shifted by a small multiplier along the spatial frequency axis. The luminance maps are generated at the resolution of $224 \\times 224$ pixels, corresponding to the size of $3.7 \\times 3.7$ visual degrees. It roughly aligned with the extend of the human foveal vision, and the stimuli span across multiple receptive fields of human vision and multiple patches/tokens of a foundation model.", + "bbox": [ + 511, + 382, + 905, + 502 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "For ColorVideoVDP, it can work directly on linear physical units, so conversion to sRGB was unnecessary. We directly use its quality score instead of $S_{\\mathrm{ac}}$ . Note that the primary focus of this study is on foundation models, with the ColorVideoVDP metric used solely as a baseline for comparison purposes.", + "bbox": [ + 511, + 503, + 905, + 595 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3.2. Model alignment score", + "text_level": 1, + "bbox": [ + 511, + 604, + 723, + 619 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Most of our results will be represented as contour plots of model responses, providing qualitative interpretation. Here, we explain our measure of model alignment, which provides quantitative scores.", + "bbox": [ + 511, + 627, + 905, + 686 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "As an example, we take the leftmost contour plot in row (a) of Figure 5. Each point on the contour plot corresponds to $S_{\\mathrm{ac}}$ between the test image as shown in Figure 3 and a uniform field of the same mean luminance. The dashed line represents human contrast detection data, predicted with castleCSF [3]. A well-aligned model should show one of the contour lines that follows the dashed castleCSF line. We cannot directly use the $S_{\\mathrm{ac}}$ values along the dashed line as the measure of alignment because some models result in $S_{\\mathrm{ac}} = 0$ for most points near the detection threshold (detect no difference). Therefore, instead, we rely on the measure of change in $S_{\\mathrm{ac}}$ in the neighborhood of the dashed line.", + "bbox": [ + 511, + 688, + 905, + 869 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "For a well-aligned model, the perceived differences in the neighborhood of the detection threshold (dashed line)", + "bbox": [ + 511, + 869, + 905, + 900 + ], + "page_idx": 2 + }, + { + "type": "page_number", + "text": "20041", + "bbox": [ + 478, + 944, + 517, + 955 + ], + "page_idx": 2 + }, + { + "type": "image", + "img_path": "images/d7c1aed2d0855875c06773d5a02641f0a92c09bb2a00332af52e2c56137b75f5.jpg", + "image_caption": [ + "Figure 3. Gabors with different spatial frequencies (x-axis) and contrast (y-axis) used as the test images in the contrast detection tests (Section 4.1). \"cpd\" denotes cycles per degree. Note that the high-frequency patterns can be rendered with aliasing artifacts on the screen or in print — those were not present in our tests." + ], + "image_footnote": [], + "bbox": [ + 96, + 88, + 486, + 363 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "should increase as the contrast increases (the sensitivity decreases). Furthermore, the values should be similar along the dashed line. The measure these two properties, we sample the $S_{\\mathrm{ac}}$ values for the points that are shifted in contrast (vertical direction) from the dashed line by a multiplier $m$ , where $0.5 \\leq m \\leq 2$ (note the logarithmic scale in the contour plot in Figure 5). We collect such data for multiple frequencies (or other dimensions) along the dashed line and calculate the Spearman rank order correlation between the multipliers $m$ and the $S_{\\mathrm{ac}}$ values. If the properties mentioned above are preserved, the correlation coefficient value $r_s$ should be close to 1.", + "bbox": [ + 88, + 470, + 482, + 651 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "The above strategy is used for all contrast detection and contrast masking experiments. For the contrast matching experiment, we use the root mean squared error (RMSE) between the model and human matching data, expressed as the logarithm of contrast.", + "bbox": [ + 89, + 652, + 483, + 729 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "4. Experiments", + "text_level": 1, + "bbox": [ + 89, + 744, + 223, + 762 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "4.1. Contrast detection and CSF", + "text_level": 1, + "bbox": [ + 89, + 771, + 341, + 786 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "We begin by testing the foundation models' ability to detect low-contrast (near-threshold) patterns (e.g., Gabor patches, band-limited noise) and compare their performance with the human data. As the reference human data, we rely on the castleCSF [3], which is the recent contrast sensitivity function, modeling contrast detection of both achromatic and chromatic patterns.", + "bbox": [ + 89, + 794, + 483, + 902 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Spatial Frequency Contrast sensitivity of the human eye is typically associated with the variation across the spatial frequency. The visual system exhibits a band-pass characteristic, with a peak sensitivity between 2 and 4 cycles per degree (cpd), depending on the luminance and other parameters of the stimulus. The lower sensitivity of the visual system at lower frequencies is associated with the mechanism of lateral inhibition [4], which helps to reduce the influence of (low-frequency) illumination on the perceived images. Such invariance to illumination is a desirable property if the goal is to recognize objects regardless of illumination conditions. The drop in sensitivity at high frequencies is associated with the limitations of eye optics (achromatic contrast) and cone density (chromatic contrast). Here we want to test whether the computer-vision foundation models pick up similar traits when trained on natural images.", + "bbox": [ + 511, + 90, + 903, + 332 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "To test encoder responses across frequencies, we generated a 2D array of image pairs, in which the reference image had uniform luminance, and the test image contained a Gabor patch, as shown in Figure 3 (refer to Supplementary for the visualization of other stimuli). We generate such Gabors for achromatic (see Figure 4-a) and chromatic modulation (see Figure 4-c,d). We also tested band-limited noise (see Figure 4-b). The test patterns had fixed size but varying spatial frequency (x-axis) and contrast (y-axis). Because most of the contrast detection data are plotted as the function of sensitivity, we follow this convention and plot the sensitivity, which is the inverse of the contrast. This corresponds to the reversal of the axis on the logarithmic plots, which we use in our analysis. The other parameters of the stimuli are listed in Table 1. Although we tested multiple variants of each foundation model, here we show only the variant with the highest complexity and best performance on its original high-level tasks. The contour plots for other variants can be found in the Supplementary.", + "bbox": [ + 511, + 335, + 906, + 623 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "The contour plots of foundation model responses are shown together with a contrast sensitivity function (castleCSF [3]) in rows (a)-(d) of Figure 5. If the foundation models had the same contrast detection characteristics as the human eye, we would expect the smallest difference (the lowest $S_{\\mathrm{ac}}$ ) contour line to follow the black dashed curve of castleCSF. This is not the case for any of the tested foundation models. Some models, and in particular DINOv2 and SD-VAE, show overall band-pass characteristics, in particular for noise and achromatic Gabors. SD-VAE has a drop of sensitivity at lower frequencies much larger than that of the visual system. The responses to chromatic patterns (rows (c) and (d)) tend to be less regular than to achromatic patterns and lack the shift toward lower frequencies, which is observed in the human data. Both OpenCLIP and SAM-2 show a very inconsistent response across the spatial frequencies.", + "bbox": [ + 511, + 626, + 908, + 881 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "From the data, we can conclude that foundation models", + "bbox": [ + 531, + 885, + 906, + 900 + ], + "page_idx": 3 + }, + { + "type": "page_number", + "text": "20042", + "bbox": [ + 478, + 944, + 519, + 957 + ], + "page_idx": 3 + }, + { + "type": "image", + "img_path": "images/279d8c2806526810191f33e945b779b441e73d8e76443f6ea954cf34bad9ea72.jpg", + "image_caption": [ + "Figure 4. Examples of test images for contrast detection and contrast masking. The detailed explanation of the stimuli can be found in the Supplementary. \"Ach.\" denotes achromatic." + ], + "image_footnote": [], + "bbox": [ + 94, + 89, + 292, + 282 + ], + "page_idx": 4 + }, + { + "type": "image", + "img_path": "images/0b75af6b94cd981d111df8bdb0aece20430b01b3f4d16b50c54a98c7f0b17794.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 297, + 90, + 496, + 282 + ], + "page_idx": 4 + }, + { + "type": "image", + "img_path": "images/6162e081b249d2d0ddb2338aca3ebae29673b37973fa4e40c57f1565ba19c8f6.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 500, + 90, + 697, + 282 + ], + "page_idx": 4 + }, + { + "type": "image", + "img_path": "images/44a1272e20b81ec965f413aef084573b2057923ddb69b183e52337396ca8dd89.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 702, + 90, + 903, + 282 + ], + "page_idx": 4 + }, + { + "type": "table", + "img_path": "images/10fcf70ff5d18d0c0e8d591be1416fee9d54905b144cb99cc22aa47e058e1c49.jpg", + "table_caption": [ + "Table 1. Key parameters for all our tests. Note that \"Radius\" does not apply when the pattern is not a Gabor." + ], + "table_footnote": [], + "table_body": "
TestSpatial Frequency (cpd)Luminance (cd/m2)Radius (degree)Contrast
Spatial Frequency - Gabor Achromatic0.5 - 3210010.001 - 1
Spatial Frequency - Noise Achromatic0.5 - 32100-0.001 - 1
Spatial Frequency - Gabor RG0.5 - 3210010.001 - 0.12
Spatial Frequency - Gabor YV0.5 - 3210010.001 - 0.8
Luminance - Gabor Achromatic20.1 - 20010.001 - 1
Area - Gabor Achromatic81000.1 - 10.001 - 1
Phase-Coherent Masking2 (mask) / 2 (test)320.5 (test)0.005 - 0.5 (mask) / 0.01 - 0.5 (test)
Phase-Incoherent Masking0 - 12 (mask) / 1.2 (test)370.8 (test)0.005 - 0.5 (mask) / 0.01 - 0.5 (test)
Contrast Matching5 (reference) / 0.25 - 25 (test)10-0.005 - 0.629 (reference)
", + "bbox": [ + 94, + 363, + 903, + 507 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "do not follow the sensitivity pattern of the visual system, but some may show a band-pass characteristic that is associated with the CSF. Many models show lower sensitivity at lower frequencies, which may indicate that those models obtained some invariance to (low-frequency) illumination through training.", + "bbox": [ + 88, + 545, + 482, + 636 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Luminance The sensitivity of the human eye increases with the luminance. In dim light, human contrast sensitivity increases proportionally to the square root of retinal illuminance, following the DeVries-Rose law. Conversely, in bright light, sensitivity follows Weber's law, remaining independent of illuminance [38]. In this experiment, we want to test whether the computer vision models lose sensitivity at lower luminance levels. Such a loss could be justified by camera noise, which increases in terms of contrast as the light intensity decreases [1].", + "bbox": [ + 88, + 657, + 482, + 808 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "To produce contour plots, we generated a 2D array of image pairs in a similar manner as for spatial frequency variations in the section above, but instead of varying spatial frequency, we varied luminance, as shown in Figure 4e. As a reminder, we did not pass the absolute luminance values directly to each model but instead converted", + "bbox": [ + 88, + 810, + 483, + 898 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "them to display-encoded (gamma-corrected) sRGB space — the colour space used for training datasets (see Figure 2). Such an encoding partially compensates for perceptual non-uniformity of luminance.", + "bbox": [ + 511, + 545, + 906, + 604 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "The results, shown in row (e) of Figure 5, indicate a systematic drop in sensitivity with luminance for all tested models. The drop in sensitivity of those models is faster than that observed in the human data, in particular for OpenCLIP.", + "bbox": [ + 511, + 606, + 905, + 680 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "We can conclude that the models trained on sufficiently large datasets mimic the sensitivity of the visual system to luminance but with a faster drop in sensitivity. This, however, may be the result of using sRGB representation for training datasets, which were presumably mostly well-exposed and contained relatively little information in darker image regions.", + "bbox": [ + 511, + 681, + 905, + 787 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Area Stimulus area (size) significantly affects sensitivity, as larger stimuli activate more retinal cells. Sensitivity increases with area up to the saturation point, denoted as the critical area [37]. The response to stimuli of different sizes tests the model's ability to pool information across the visual field.", + "bbox": [ + 511, + 810, + 905, + 898 + ], + "page_idx": 4 + }, + { + "type": "page_number", + "text": "20043", + "bbox": [ + 478, + 944, + 517, + 955 + ], + "page_idx": 4 + }, + { + "type": "image", + "img_path": "images/45234d434d4aa988f17c60a830b5a74fbdb79067252d67b5afac2032d5dc610d.jpg", + "image_caption": [ + "(a)" + ], + "image_footnote": [], + "bbox": [ + 116, + 89, + 292, + 181 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/aa0444172e51a34e77bc1ea747cb89ad20f2151899ea7c4297d73a716ba6b9c9.jpg", + "image_caption": [ + "DINOv2" + ], + "image_footnote": [], + "bbox": [ + 297, + 104, + 442, + 181 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/35f97ada3b530f1ecc9fa19ba38634f751b2ef0cd45abade2dc88b280e87ad42.jpg", + "image_caption": [ + "SAM-2 AM2.1-hiera-L" + ], + "image_footnote": [], + "bbox": [ + 450, + 106, + 594, + 181 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/05c08c9f36668c34962c79ad3924b6dcd2ae9dede1178c699935ece80bbd50f5.jpg", + "image_caption": [ + "(b)" + ], + "image_footnote": [], + "bbox": [ + 116, + 181, + 292, + 257 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/f0cb1d4c9c774f738b125ad3f899ecad3f4889f57709af1be117f40a5634fe6d.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 297, + 181, + 442, + 257 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/037b2220cb657e05717a1dd80faf4a316d6d2aff01dfbb36e26ebcdd2e3c2002.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 449, + 181, + 594, + 257 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/db14fd77c4a4bd795050b4ffeb906decae1ce44ee7ceabd4161d30e69c3bd82d.jpg", + "image_caption": [ + "(c)" + ], + "image_footnote": [], + "bbox": [ + 116, + 257, + 292, + 330 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/b353a55570550f30d4a358d22a6fe58b3866d6ef86a2fe3a4ee34cf14445092e.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 297, + 256, + 442, + 332 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/1a39a3e80cdc5fc48baedb7a9d7835850965d878301f74b76576106335aa0bdd.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 449, + 256, + 594, + 332 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/4545645c37ed87883d0a9477d6667ef56b85edb9eff9cdf4d4c6a522e2df8dda.jpg", + "image_caption": [ + "(d)" + ], + "image_footnote": [], + "bbox": [ + 116, + 332, + 292, + 406 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/bba4fde8146dfd1c12b0bce1a54a227a469fdc2c05dee864b1d790b48f90fe66.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 297, + 332, + 442, + 407 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/3c35aae92cebc266a771748199b524b06a9ff0ac7f51444348daaadc985a23cf.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 449, + 332, + 594, + 407 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/37703300e949112c6087982e1b0315d64172a1877168cb6a2a4c0a06a8170c92.jpg", + "image_caption": [ + "(e)" + ], + "image_footnote": [], + "bbox": [ + 116, + 407, + 292, + 482 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/002123188d33cdb1e003ee8f27f0498ac96a4e7086495e32e5cb4192ab3f2f32.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 297, + 407, + 442, + 482 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/7a5459f0cf7c7d7313a97a8ea43c4d3c0ef306255d039a57a462b7bf99a94179.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 449, + 407, + 594, + 482 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/b33a4b848ad6d98e0d52bdb2c57bb3a6216b94841c0e40260bb9fde9ca919727.jpg", + "image_caption": [ + "(f)" + ], + "image_footnote": [], + "bbox": [ + 116, + 482, + 292, + 549 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/447f600fd7065d9232604d2f5e14bf1bd7af5ace80109840719e3aa6a963a65a.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 297, + 481, + 442, + 550 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/d96d59950efb8a97beb0ab1bc826040a36149a8bc8aa479dbbf9da03a8a2345f.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 449, + 481, + 594, + 550 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/aeacc86094ff22a60f9715ac9e9f76b98bca70824d9449847c6867aec5fde746.jpg", + "image_caption": [ + "(g)" + ], + "image_footnote": [], + "bbox": [ + 116, + 551, + 292, + 626 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/4ec7f0fd6a3b7d17c93d97cf80c14285325febc18f0a406ff82b34d048c8598d.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 297, + 551, + 442, + 626 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/a847f238fdcd06ede35abc32d5698bcb22f6b34f1ddc2db7e55a4a866c091dc0.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 449, + 551, + 594, + 626 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/aecfb27b88c1bf415e2700a90c40f93beccf1c4f0bdf09684be59d567fdee2f1.jpg", + "image_caption": [ + "(h)" + ], + "image_footnote": [], + "bbox": [ + 116, + 628, + 292, + 702 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/c056b7796136da919be77d615f40e33582755009980ab01d75b69670b64e6915.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 297, + 628, + 442, + 702 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/b7e0d684f7d16c3ae97d9b25b655cb094c1f80dfde3364084b288fa3f745ee83.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 449, + 628, + 594, + 702 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/fe940144a3a69b32d2e01a1c225f8757b97a07a9d8099fd367d67cd12e574f7c.jpg", + "image_caption": [ + "(i)" + ], + "image_footnote": [], + "bbox": [ + 116, + 704, + 292, + 784 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/cbe015098fad08a967b53a28abfb523581f75bc7532f833dc1ae3a7cf6ba97fa.jpg", + "image_caption": [ + "Figure 5. Selected representative experimental results. Each row represents a test, and each column corresponds to a model, selected as the best-performing in their original tasks. (a)-(f): contour plots of contrast detection $S_{\\mathrm{ac}}$ , with the ground truth castleCSF [3]. (g),(h): contour plots of contrast masking $S_{\\mathrm{ac}}$ , with the ground truth from [14] and [15], respectively. Different colored solid lines representing different $S_{\\mathrm{ac}}$ values, where purple indicates the minimum difference $(S_{\\mathrm{ac}} \\rightarrow 0)$ . (i): results from the contrast matching experiment, where different colors represent different $C_r$ values. The dashed lines are human results [18], while the solid lines are model-predicted results from Equation 2. Results for all models are in the supplementary material." + ], + "image_footnote": [], + "bbox": [ + 297, + 704, + 442, + 782 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/4517034d72c7d377a847734073a7eff57ad180a2a8358c28272432a81c0c254d.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 449, + 704, + 594, + 782 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/9ad130f3105f49b0463f3b0c83fbb44d09a606a4dbd8c82e3a026859b1213f4a.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 625, + 89, + 746, + 181 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/34e24619db9f5fded87b2a2b0ce6a7cebf70f92d241bf323b7bc075481255b99.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 601, + 181, + 746, + 257 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/076f3df94f1637766b7c76fe78469d52b9fb52a5adc28287e014bd52a4e2cf2c.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 601, + 256, + 746, + 332 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/0449610e1387d864bc54407184c945c6681b812ef995e4fa98f7b7005bd250e9.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 601, + 332, + 746, + 407 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/2aeecaaf096c601b2def30e4ff711bfb55134fbb421c70e02200c5a1f67c08f5.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 601, + 407, + 746, + 481 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/13cb7c5f16ca78175924f509fb0affd67b84d57aef31566c11aae6b2b25e4120.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 601, + 481, + 746, + 549 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/c5a6941c61b14ad16a3ab5e7671baeabf52f0c038db62caaf5d88795b73552b9.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 601, + 549, + 746, + 626 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/5659372b04a97eee12e20d16b782037f350d0cd8d14145acdc80a0eb164b0615.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 601, + 628, + 746, + 702 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/67695cec6010abde9080e0ebaa6522980a233a12218456174173e34b3af2e11b.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 601, + 704, + 746, + 782 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/22801bb5ab4cb3738845a48b08206e748c2942e74d780ed63a3aebfa4b937b65.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 754, + 89, + 900, + 181 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/8b0c35ba80a07c673cd99fd146b84c28be0a5c7b22979232e7b078bf4f473640.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 754, + 181, + 900, + 257 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/51fce1bab9f1fb7adbad49fd3c723544775bf1ff74e336e333cbb353d25cfa70.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 754, + 256, + 900, + 332 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/de63f8593a28f479c67ac19bb046569cc417b3785668f38589cde384a9b73181.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 754, + 332, + 900, + 407 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/f1e9ae3061e4df7f7097b4ddcffa3e99fe4f22b84f6f5c67edb7dcc8599b51b9.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 754, + 407, + 900, + 476 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/0e8fff5f273870bb4024bb2a361ce24d3c8e8b825195e84e09bbb13b61b919e6.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 754, + 476, + 900, + 547 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/3ff7323cab7fa8c421a94b732a840a54f8d4d3f242fae55d925dfbfbda03c0c6.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 754, + 549, + 900, + 626 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/52436734eded9e750e3faeefe39eabfa19abdcf4fe9e098725b52d0b41dccb8d.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 754, + 628, + 900, + 698 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/e8b36897e8390bda2e5140251516841a7eb5e2ce66cfac40546eeb258461d151.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 754, + 704, + 900, + 782 + ], + "page_idx": 5 + }, + { + "type": "page_number", + "text": "20044", + "bbox": [ + 478, + 944, + 519, + 955 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Contrast Detection - Spatial Frequency - Gabor Achromatic - Spearman Correlation ↑", + "bbox": [ + 129, + 89, + 467, + 99 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/8e0cb1e6e27798360c09c41734a5870b6d9976748129354297b42946214f78ff.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 98, + 104, + 478, + 146 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/891a3eed1cd14aaa5bdd95128e2f74fcb1f72e4e3cd57b2322988241e9193e21.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 98, + 151, + 478, + 191 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/b9e1c2c84291ba86af3f471c380670045e16bdf4822525724386d274777a6fa3.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 98, + 196, + 478, + 237 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/8ec001b0dd669591dcbf75ae59313e4fa49677e210652074b6f0f57480730675.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 98, + 241, + 478, + 284 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/66209f039a2c232d3e3ab4acceff3924732831f74365a0f7d0624669610cff86.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 98, + 287, + 478, + 330 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/8228213a4605f408a8908fbe2499812e00d42a619265dd37fb17ffbccd5ae3ef.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 98, + 333, + 478, + 377 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/0fc930f3c32ff5595b7936325fabff5dafb52d3c001d5a7b44f5b1e3a688a0b5.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 98, + 381, + 478, + 424 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/f8d686748e86a14ea6f0332967ba7796f07c0fb33d013415ad0cded263ff7dc6.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 98, + 426, + 478, + 470 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/2ebd6b659e2b79d97bb3f4574b33c388a6a8ed35a138717417f2922ef05250b7.jpg", + "image_caption": [ + "Figure 6. The quantified similarity error between all 45 models and HVS under 9 different tests. For the Contrast Detection and Contrast Masking tasks, Spearman Correlation was used as the metric, with higher values (closer to 1) indicating greater similarity to human vision. For the Supra-threshold Contrast Matching task, RMSE was used as the metric, with lower values (closer to 0) indicating better similarity." + ], + "image_footnote": [], + "bbox": [ + 94, + 473, + 478, + 599 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "In this experiment, we vary the size of the Gaussian envelope limiting stimulus size to observe its effect on model responses. We follow the same procedure as in the two previous sections and generate pairs with a uniform field and a Gabor patch (see Figure 4-f). The parameters of the Gabor patch are listed in Table 1.", + "bbox": [ + 89, + 744, + 482, + 835 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "The results, shown in row (f) of Figure 5, indicate that most models show summation across the area (the $S_{\\mathrm{ac}}$ values increase with the area) and the rate of the increase varies across the models; DINO and SAM-2 have a smaller in", + "bbox": [ + 89, + 839, + 483, + 898 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "crease, SD-VAR has a higher increase of $S_{\\mathrm{ac}}$ than the human data, and only DINOv2 roughly matches human performance. OpenCLIP shows no consistent patterns. We can conclude that many models show spatial pooling characteristic that shares the trend observed in human data, though the actual slope of the increase is typically different.", + "bbox": [ + 511, + 90, + 905, + 181 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "4.2. Contrast masking", + "text_level": 1, + "bbox": [ + 511, + 194, + 687, + 210 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Contrast masking explains the decreased visibility of a signal (test) due to the presence of a supra-threshold background (mask). The masking function defines the relationship between the threshold test contrast required for signal detection and the mask contrast. Put simply, a pattern is more difficult to detect in the presence of another pattern of similar spatial frequency and orientation. A typical masking characteristic of the visual system is shown in rows (g) and (h) of Figure 5. It consists of a relatively shallow segment at low mask contrast (near the detection threshold), with the slope increasing for high mask contrast. The shape of the curve is influenced by the specific properties of the mask and test signals [11, 43]. We will consider the case in which the masker is a sinusoidal grating of the same frequency as the test pattern (phase-coherent masking, row (g)) and when the masker is noise (phase-incoherent masking, row (h)) [11, 14, 15], as shown in Figure 4-g,h.", + "bbox": [ + 511, + 217, + 906, + 474 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "First, we consider the fundamental form of phase-coherent masking in which the masker image (reference in Figure 2) is a sinusoidal grating and the test image is the masker plus a Gabor patch of the same frequency and phase as the masker (see Figure 4g). Contrast masking data, shown as the dashed black curve in row (g) of Figure 5, shows the smallest contrast of the test Gabor that is detectable in the presence of the masker of a given contrast. As the contrast of the masker increases, the smallest detectable contrast of the test also needs to increase. However, such an increase starts only for a masker that has the contrast sufficiently high to be detected. If the contrast of the masker is near the detection threshold, we can observe a dipper effect — the contrast detection is facilitated by a masker [14, 42, 45]. Such an effect can be only observed in phase coherent masking.", + "bbox": [ + 511, + 476, + 905, + 717 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "As an example of phase-incoherent masking, we will consider a masker with band-limited noise and a test with a Gabor patch — see Figure 4h. The human detection thresholds for such masking patterns are similar to those for phase-coherent masking, except that the dipper effect disappears [41].", + "bbox": [ + 511, + 718, + 905, + 808 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "The differences predicted by DINOv2 and OpenCLIP are surprisingly well-aligned with the human contrast masking data — their responses roughly match the slopes of the human data. The alignment is stronger for the phase-incoherent masking. This is particularly notable for OpenCLIP, which did not show any consistent trends for contrast", + "bbox": [ + 511, + 809, + 908, + 900 + ], + "page_idx": 6 + }, + { + "type": "page_number", + "text": "20045", + "bbox": [ + 478, + 944, + 517, + 955 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "detection. Other models do not show strong alignment with the human data.", + "bbox": [ + 89, + 90, + 482, + 119 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Overall, computational models are better aligned with human data for contrast masking than for contrast detection. One possible explanation is that the signals that induce contrast masking are plentiful in natural images, but contrast detection stimuli, which involve barely noticeable patterns on uniform backgrounds, are rare. Therefore, computational models are more likely to pick up the characteristic that is well represented in the training datasets.", + "bbox": [ + 89, + 121, + 483, + 243 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "4.3. Supra-threshold contrast matching", + "text_level": 1, + "bbox": [ + 89, + 250, + 398, + 266 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "While contrast detection and contrast masking explain the just detectable (near-threshold) contrast, most of the vision tasks, such as detection or recognition, involve well-visible (supra-threshold) contrast. Supra-threshold human vision has been studied in contrast-matching experiments in which the magnitude of one contrast is visually matched to the magnitude of another contrast of a different frequency [18] or luminance [24, 32]. One of the most significant findings of those studies is contrast constancy [18] — the ability of the visual system to match physical contrast across frequencies and luminance levels. The results of the seminal study of Georgeson and Sullivan [18] on matching contrast across frequencies are shown as dashed lines in row (i) of Figure 5. At small contrast, the dashed lines show a band-pass shape that follows the contrast sensitivity function. However, as the contrast is increased, the lines become flat showing little influence of frequency on contrast perception. This is an important property that lets us see objects to have the same appearance regardless of the viewing distance. Such a scale invariance is also important for neural networks that are tasked to detect or recognize objects regardless of their size.", + "bbox": [ + 91, + 272, + 483, + 602 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "We followed the experimental setup from [18], where the reference was a $5\\mathrm{cpd}$ , $10\\mathrm{cd} / \\mathrm{m}^2$ sinusoidal grating, presented at eight distinct contrast levels $c_{\\mathrm{r}}$ . The test stimulus had the same luminance but a different spatial frequency $\\rho_{\\mathrm{t}}$ . In [18], observers adjusted the test stimulus contrast $c_{\\mathrm{t}}$ until its apparent contrast matched that of the reference (contrast matching). In our experiments, we match contrast encodings of sinusoidal gratings: the $(S_{\\mathrm{ac}}()$ , eq. (1)), between a feature vector of a sinusoidal grating, $F(\\rho ,c)$ of frequency $\\rho$ and contrast $c$ , and a uniform field, $U = F(\\rho_{\\mathrm{t}},0)$ . We find the test contrast $c_{\\mathrm{t}}$ that minimizes the expression:", + "bbox": [ + 89, + 604, + 483, + 771 + ], + "page_idx": 7 + }, + { + "type": "equation", + "text": "\n$$\n\\underset {c _ {\\mathrm {t}}} {\\operatorname {a r g m i n}} \\left(S _ {\\mathrm {a c}} \\left(F \\left(\\rho_ {\\mathrm {r}}, c _ {\\mathrm {r}}\\right), U\\right) - S _ {\\mathrm {a c}} \\left(F \\left(\\rho_ {\\mathrm {t}}, c _ {\\mathrm {t}}\\right), U\\right)\\right) ^ {2}, \\tag {2}\n$$\n", + "text_format": "latex", + "bbox": [ + 112, + 776, + 482, + 804 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "where the reference frequency $\\rho_{\\mathrm{r}} = 5$ cpd. The denominators in the expression are used to normalize contrast across frequencies. We experimented with other contrast encodings, including a direct comparison of feature vectors, but the formula above resulted in the best contrast constancy properties across the models.", + "bbox": [ + 89, + 810, + 483, + 902 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "The matching contrast predictions for foundation models are visualized as continuous lines in row (i) of Figure 5. The plots show that only DINOv2 and OpenCLIP roughly follow the dashed contrast constancy lines. Both models show less attenuation (more constancy) at the highest spatial frequencies, which could be advantageous when the model needs to work with small-scale features. Both models show attenuation of low frequencies (below 1 cpd), suggesting worse contrast constancy in that frequency range. Other models, including DINO and SAM-2, suffer from large instability across frequencies, or very heavy attenuation of low frequencies in the case of SD-VAE. To conclude, we can observe only partial contrast constancy for selected models.", + "bbox": [ + 511, + 90, + 906, + 303 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "4.4. Model alignment scores", + "text_level": 1, + "bbox": [ + 511, + 321, + 732, + 338 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "As the analysis of all 45 variants of the models is infeasible in the scope of this paper, we prepared quantitative results according to the method explained in Section 3.2 and summarized them in Figure 6. Those let us make three observations:", + "bbox": [ + 511, + 347, + 906, + 421 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "First, ColorVideoVDP, which is a visual metric that models human low-level vision, is better aligned with the human data in almost all contrast detection tasks and in terms of contrast constancy, as expected. However, certain variants of OpenCLIP and DINOv2 can match or surpass the ColorVideoVDP alignment in terms of contrast masking. Second, the alignment scores of different foundation model variants (e.g., OpenCLIP) show significant variations and alignment scores appear unrelated to the complexity of the variants or their performance on higher-level tasks. Finally, DINOv2 variants, which have been trained to solve vision tasks, show the greatest alignment with the human data among all foundation models.", + "bbox": [ + 511, + 425, + 908, + 623 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "5. Conclusions", + "text_level": 1, + "bbox": [ + 511, + 645, + 640, + 660 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "If we believe that the goal of both biological and computational low-level vision is to efficiently encode visual information, we can expect that computational models trained on large natural image datasets will share similarities with human vision. In this work, we find that selected computational models, e.g., variants of DINOv2 and OpenCLIP, show surprisingly high alignment with supra-threshold human contrast masking and contrast matching data, but little alignment with the near-threshold contrast detection. This means that computation models do not have the same \"bottlenecks\" as human vision, but, through training, they attain invariance and efficient contrast coding that resembles that of the visual system. We hope that our testing protocol with basic psychophysical stimuli will provide a useful tool for examining future computational models of vision.", + "bbox": [ + 511, + 674, + 908, + 902 + ], + "page_idx": 7 + }, + { + "type": "page_number", + "text": "20046", + "bbox": [ + 478, + 945, + 519, + 955 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "References", + "text_level": 1, + "bbox": [ + 91, + 89, + 187, + 104 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[1] Cecilia Aguerrebere, Julie Delon, Yann Gousseau, and Pablo Muse. Study of the digital camera acquisition process and statistical modeling of the sensor raw data. 2013. Citation Key: Aguerrebere2013. 5", + "[2] Arash Akbarinia, Yaniv Morgenstern, and Karl R Gegenfurtner. Contrast sensitivity function in deep networks. Neural Networks, 164:228-244, 2023. 2", + "[3] Maliha Ashraf, Rafal K Mantiuk, Alexandre Shapiro, and Sophie Wuerger. castlecsf—a contrast sensitivity function of color, area, spatiotemporal frequency, luminance and eccentricity. Journal of Vision, 24(4):5-5, 2024. 3, 4, 6", + "[4] Peter GJ Barten. Contrast sensitivity of the human eye and its effects on image quality. SPIE press, 1999. 1, 4", + "[5] Jeffrey S Bowers, Gaurav Malhotra, Marin Dujmovic, Milton Llera Montero, Christian Tsvetkov, Valerio Biscione, Guillermo Puebla, Federico Adolfi, John E Hummel, Rachel F Heaton, et al. Deep problems with neural network models of human vision. Behavioral and Brain Sciences, 46: e385, 2023. 2", + "[6] Oliver Braddick and Janette Atkinson. Development of human visual function. Vision Research, 51(13):1588-1609, 2011. 1", + "[7] Yancheng Cai, Bo Zhang, Baopu Li, Tao Chen, Hongliang Yan, Jingdong Zhang, and Jiahao Xu. Rethinking cross-domain pedestrian detection: a background-focused distribution alignment framework for instance-free one-stage detectors. IEEE transactions on image processing, 32:4935-4950, 2023. 2", + "[8] Yancheng Cai, Ali Bozorgian, Maliha Ashraf, Robert Wanat, and Rafal Mantiuk. elatcsf: A temporal contrast sensitivity function for flicker detection and modeling variable refresh rate flicker. SIGGRAPH Asia 2024 [in print], 2024. 3", + "[9] Mathilde Caron, Hugo Touvron, Ishan Misra, Hervé Jégou, Julien Mairal, Piotr Bojanowski, and Armand Joulin. Emerging properties in self-supervised vision transformers. In Proceedings of the IEEE/CVF international conference on computer vision, pages 9650-9660, 2021. 1, 2, 3", + "[10] Francesco Croce, Christian Schlarmann, Naman Deep Singh, and Matthias Hein. Adversarily robust clip models induce better (robust) perceptual metrics. In ICML 2024 Workshop on Foundation Models in the Wild, 2024. 2", + "[11] Scott J Daly. Visible differences predictor: an algorithm for the assessment of image fidelity. In Human Vision, Visual Processing, and Digital Display III, pages 2-15. SPIE, 1992. 7", + "[12] Timothee Darcet, Maxime Oquab, Julien Mairal, and Piotr Bojanowski. Vision transformers need registers, 2023. 3", + "[13] Alexey Dosovitskiy. An image is worth 16x16 words: Transformers for image recognition at scale. arXiv preprint arXiv:2010.11929, 2020. 2", + "[14] John M Foley. Human luminance pattern-vision mechanisms: masking experiments require a new model. JOSA A, 11(6):1710-1719, 1994. 6, 7", + "[15] Karl R Gegenfurtner and Daniel C Kiper. Contrast detection in luminance and chromatic noise. JOSA A, 9(11):1880-1888, 1992. 6, 7" + ], + "bbox": [ + 93, + 114, + 483, + 898 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[16] Robert Geirhos, Patricia Rubisch, Claudio Michaelis, Matthias Bethge, Felix A Wichmann, and Wieland Brendel. Imagenet-trained cnns are biased towards texture; increasing shape bias improves accuracy and robustness. arXiv preprint arXiv:1811.12231, 2018. 2", + "[17] Robert Geirhos, Kantharaju Narayanappa, Benjamin Mitzkus, Tizian Thieringer, Matthias Bethge, Felix A Wichmann, and Wieland Brendel. Partial success in closing the gap between human and machine vision. Advances in Neural Information Processing Systems, 34:23885-23899, 2021. 2", + "[18] MA Georgeson and GD Sullivan. Contrast constancy: deblurring in human vision by spatial frequency channels. The Journal of physiology, 252(3):627-656, 1975. 1, 2, 6, 8", + "[19] Abhijay Ghildyal, Nabajeet Barman, and Saman Zadtootaghaj. Foundation models boost low-level perceptual similarity metrics. arXiv preprint arXiv:2409.07650, 2024. 2", + "[20] Ian J. Goodfellow, Jonathon Shlens, and Christian Szegedy. Explaining and harnessing adversarial examples. 2014. 2", + "[21] Kaiming He, Xinlei Chen, Saining Xie, Yanghao Li, Piotr Dólár, and Ross Girshick. Masked autoencoders are scalable vision learners. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 16000-16009, 2022. 3", + "[22] Gabriel Ilharco, Mitchell Wortsman, Ross Wightman, Cade Gordon, Nicholas Carlini, Rohan Taori, Achal Dave, Vaishaal Shankar, Hongseok Namkoong, John Miller, Hannaneh Hajishirzi, Ali Farhadi, and Ludwig Schmidt. Openclip, 2021. If you use this software, please cite it as below. 1, 3", + "[23] Alexander Kirillov, Eric Mintun, Nikhila Ravi, Hanzi Mao, Chloe Rolland, Laura Gustafson, Tete Xiao, Spencer Whitehead, Alexander C Berg, Wan-Yen Lo, et al. Segment anything. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 4015-4026, 2023. 2, 3", + "[24] J.J. Kulikowski. Effective contrast constancy and linearity of contrast sensation. Vision Research, 16(12):1419-1431, 1976. Citation Key: Kulikowski1976a. 2, 8", + "[25] G E Legge and J M Foley. Contrast masking in human vision. Journal of the Optical Society of America, 70(12):1458-71, 1980. Citation Key: Legge1980. 1", + "[26] Qiang Li, Alex Gomez-Villa, Marcelo Bertalmó, and Jesús Malo. Contrast sensitivity functions in autoencoders. Journal of Vision, 22(6):8-8, 2022. 2", + "[27] David G. Lowe. Distinctive image features from scale-invariant keypoints. International Journal of Computer Vision, 60(2):91-110, 2004. 2", + "[28] Lei Lu, Yancheng Cai, Hua Huang, and Ping Wang. An efficient fine-grained vehicle recognition method based on part-level feature optimization. Neurocomputing, 536:40-49, 2023. 2", + "[29] Yilong Lv, Yancheng Cai, Yujie He, and Min Li. Drkd: Decoupling response-based distillation for object detection. Pattern Recognition, 161:111275, 2025. 2", + "[30] Rafal K Mantiuk, Param Hanji, Maliha Ashraf, Yuta Asano, and Alexandre Shapiro. Colorvideoovdp: A visual difference" + ], + "bbox": [ + 516, + 92, + 905, + 900 + ], + "page_idx": 8 + }, + { + "type": "page_number", + "text": "20047", + "bbox": [ + 478, + 944, + 517, + 955 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "predictor for image, video and display distortions. arXiv preprint arXiv:2401.11485, 2024.3", + "[31] Maxime Oquab, Timothée Darcet, Theo Moutakanni, Huy V. Vo, Marc Szafraniec, Vasil Khalidov, Pierre Fernandez, Daniel Haziza, Francisco Massa, Alaaeldin El-Nouby, Russell Howes, Po-Yao Huang, Hu Xu, Vasu Sharma, Shang-Wen Li, Wojciech Galuba, Mike Rabbat, Mido Assran, Nicolas Ballas, Gabriel Synnaeve, Ishan Misra, Herve Jegou, Julien Mairal, Patrick Labatut, Armand Joulin, and Piotr Bojanowski. Dinov2: Learning robust visual features without supervision, 2023. 2, 3", + "[32] Eli Peli. Suprathreshold contrast perception across differences in mean luminance: effects of stimulus size, dichoptic presentation, and length of adaptation. Journal of the Optical Society of America A, 12(5):817, 1995. Citation Key: Peli1995. 2, 8", + "[33] Alec Radford, Jong Wook Kim, Chris Hallacy, A. Ramesh, Gabriel Goh, Sandhini Agarwal, Girish Sastry, Amanda Askell, Pamela Mishkin, Jack Clark, Gretchen Krueger, and Ilya Sutskever. Learning transferable visual models from natural language supervision. In ICML, 2021. 1, 3", + "[34] Alec Radford, Jong Wook Kim, Chris Hallacy, Aditya Ramesh, Gabriel Goh, Sandhini Agarwal, Girish Sastry, Amanda Askell, Pamela Mishkin, Jack Clark, et al. Learning transferable visual models from natural language supervision. In International conference on machine learning, pages 8748-8763. PMLR, 2021. 2", + "[35] Nikhila Ravi, Valentin Gabeur, Yuan-Ting Hu, Ronghang Hu, Chaitanya Ryali, Tengyu Ma, Haitham Khedr, Roman Rädle, Chloe Rolland, Laura Gustafson, et al. Sam 2: Segment anything in images and videos. arXiv preprint arXiv:2408.00714, 2024. 3", + "[36] Robin Rombach, Andreas Blattmann, Dominik Lorenz, Patrick Esser, and Björn Ommer. High-resolution image synthesis with latent diffusion models. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 10684-10695, 2022. 3", + "[37] Jyrki Rovamo, Olavi Luntinen, and Risto Nasänen. Modelling the dependence of contrast sensitivity on grating area and spatial frequency. Vision Research, 33(18):2773-2788, 1993. 5", + "[38] Jyrki Rovamo, Juvi Mustonen, and Risto Nasänen. Neural modulation transfer function of the human visual system at various eccentricities. Vision Research, 35(6):767-774, 1995. 5", + "[39] Stephen Tian, Yancheng Cai, Hong-Xing Yu, Sergey Zakharov, Katherine Liu, Adrien Gaidon, Yunzhu Li, and Jia-jun Wu. Multi-object manipulation via object-centric neural scattering functions. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 9021-9031, 2023. 2", + "[40] Shikhar Tuli, Ishita Dasgupta, Erin Grant, and Thomas L Griffiths. Are convolutional neural networks or transformers more like human vision? arXiv preprint arXiv:2105.07197, 2021. 2", + "[41] A Van Meeteren and JM Valeton. Effects of pictorial noise interfering with visual detection. JOSA A, 5(3):438-444, 1988. 7" + ], + "bbox": [ + 91, + 90, + 482, + 898 + ], + "page_idx": 9 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[42] Andrew B Watson. Efficiency of a model human image code. JOSA A, 4(12):2401-2417, 1987. 7", + "[43] Andrew B Watson and Joshua A Solomon. Model of visual contrast gain control and pattern masking. JOSA A, 14(9): 2379-2391, 1997. 7", + "[44] Felix A Wichmann and Robert Geirhos. Are deep neural networks adequate behavioral models of human visual perception? Annual Review of Vision Science, 9(1):501-524, 2023. 2", + "[45] Hugh R Wilson and James R Bergen. A four mechanism model for threshold spatial vision. Vision research, 19(1): 19-32, 1979. 7", + "[46] Fei Yin, Yong Zhang, Baoyuan Wu, Yan Feng, Jingyi Zhang, Yanbo Fan, and Yujiu Yang. Generalizable black-box adversarial attack with meta learning. IEEE transactions on pattern analysis and machine intelligence, 46(3):1804-1818, 2023. 2", + "[47] Jingdong Zhang, Jiayuan Fan, Peng Ye, Bo Zhang, Hancheng Ye, Baopu Li, Yancheng Cai, and Tao Chen. Bridgenet: Comprehensive and effective feature interactions via bridge feature for multi-task dense predictions. IEEE Transactions on Pattern Analysis and Machine Intelligence, 2025. 2" + ], + "bbox": [ + 514, + 92, + 903, + 414 + ], + "page_idx": 9 + }, + { + "type": "page_number", + "text": "20048", + "bbox": [ + 478, + 944, + 517, + 955 + ], + "page_idx": 9 + } +] \ No newline at end of file diff --git a/2025/Do Computer Vision Foundation Models Learn the Low-level Characteristics of the Human Visual System_/4607e0a5-8aa8-470e-b0e4-8cfed36f43c4_model.json b/2025/Do Computer Vision Foundation Models Learn the Low-level Characteristics of the Human Visual System_/4607e0a5-8aa8-470e-b0e4-8cfed36f43c4_model.json new file mode 100644 index 0000000000000000000000000000000000000000..e4dd32c8f4ecd2bf19ba4b500bbd4194fa44f6c6 --- /dev/null +++ b/2025/Do Computer Vision Foundation Models Learn the Low-level Characteristics of the Human Visual System_/4607e0a5-8aa8-470e-b0e4-8cfed36f43c4_model.json @@ -0,0 +1,2420 @@ +[ + [ + { + "type": "header", + "bbox": [ + 0.107, + 0.003, + 0.182, + 0.044 + ], + "angle": 0, + "content": "CVF" + }, + { + "type": "header", + "bbox": [ + 0.237, + 0.001, + 0.812, + 0.047 + ], + "angle": 0, + "content": "This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore." + }, + { + "type": "title", + "bbox": [ + 0.214, + 0.106, + 0.784, + 0.151 + ], + "angle": 0, + "content": "Do computer vision foundation models learn the low-level characteristics of the human visual system?" + }, + { + "type": "text", + "bbox": [ + 0.153, + 0.179, + 0.846, + 0.197 + ], + "angle": 0, + "content": "Yancheng Cai, Fei Yin, Dounia Hammou, Rafal Mantiuk; University of Cambridge, UK" + }, + { + "type": "text", + "bbox": [ + 0.205, + 0.199, + 0.787, + 0.214 + ], + "angle": 0, + "content": "yc613@cam.ac.uk, fy277@cam.ac.uk, dh706@cam.ac.uk, mantiuk@gmail.com" + }, + { + "type": "title", + "bbox": [ + 0.248, + 0.249, + 0.327, + 0.265 + ], + "angle": 0, + "content": "Abstract" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.282, + 0.485, + 0.691 + ], + "angle": 0, + "content": "Computer vision foundation models, such as DINO or OpenCLIP, are trained in a self-supervised manner on large image datasets. Analogously, substantial evidence suggests that the human visual system (HVS) is influenced by the statistical distribution of colors and patterns in the natural world, characteristics also present in the training data of foundation models. The question we address in this paper is whether foundation models trained on natural images mimic some of the low-level characteristics of the human visual system, such as contrast detection, contrast masking, and contrast constancy. Specifically, we designed a protocol comprising nine test types to evaluate the image encoders of 45 foundation and generative models. Our results indicate that some foundation models (e.g., DINO, DINOv2, and OpenCLIP), share some of the characteristics of human vision, but other models show little resemblance. Foundation models tend to show smaller sensitivity to low contrast and rather irregular responses to contrast across frequencies. The foundation models show the best agreement with human data in terms of contrast masking. Our findings suggest that human vision and computer vision may take both similar and different paths when learning to interpret images of the real world. Overall, while differences remain, foundation models trained on vision tasks start to align with low-level human vision, with DINOv2 showing the closest resemblance. Our code is available on https://github.com/caiyancheng/VFM_HVS_CVPR2025." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.723, + 0.222, + 0.739 + ], + "angle": 0, + "content": "1. Introduction" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.75, + 0.484, + 0.901 + ], + "angle": 0, + "content": "Computer vision foundation models, such as DINO [9] or OpenCLIP [22, 33], show exceptional ability to generalize to different tasks and are becoming cornerstones of many computer vision methods. They owe their exceptional performance to self-supervised training on very large image datasets. The human visual system also owes much of its capability to being able to perceive the world, over many years from infancy to childhood [6]. A question arises: if the neural network and the visual system are trained by being exposed to a large number of images of the world, will" + }, + { + "type": "image", + "bbox": [ + 0.517, + 0.235, + 0.907, + 0.468 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.512, + 0.477, + 0.907, + 0.547 + ], + "angle": 0, + "content": "Figure 1. To determine whether image encoders of foundation models exhibit a similar low-level characteristic as human vision, we test them on psychophysical stimuli for which human data is available. We want to test the alignment of contrast encoding between human and computational vision models." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.58, + 0.907, + 0.732 + ], + "angle": 0, + "content": "they share their low-level vision characteristics? If they do, we will know that those low-level characteristics arise naturally and likely reflect the statistics of real-world scenes. If they do not, it means that human low-level vision characteristics are specific to the optical/biological limitations of human vision rather than natural image statistics. Our analysis is meant to shed some light on how the vision, either biological or computational, may develop from observing samples of the world, taking either the same or different routes to accomplish their respective tasks." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.735, + 0.909, + 0.903 + ], + "angle": 0, + "content": "In particular, we are interested in the characteristics that are well understood and measured in human vision science using psychophysical methods: contrast detection [4], contrast masking [25] and contrast constancy [18]. Contrast detection and contrast masking quantify the ability of the visual system to detect small contrast patterns, either on uniform backgrounds (contrast detection) or on backgrounds with patterns (contrast masking). Contrast detection and masking capture the \"bottlenecks\" of the visual system — the characteristic that may prevent us from detecting patterns that are too dark or too small. Similarly, cameras used" + }, + { + "type": "page_number", + "bbox": [ + 0.48, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "20039" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.09, + 0.092, + 0.482, + 0.151 + ], + "angle": 0, + "content": "for computer vision are limited by the MTF of the lens, sensor resolution, photon and sensor noise, and we can expect that computer vision methods may need to deal with similar limitations." + }, + { + "type": "text", + "bbox": [ + 0.093, + 0.154, + 0.483, + 0.456 + ], + "angle": 0, + "content": "Contrast constancy is the term used in vision science to describe the invariance of the visual system to spatial frequency [18] and partially luminance [24, 32]. Georgeson and Sullivan [18] showed that the perceived magnitude of the contrast that is well above the detection threshold (supra-threshold) appears to us the same regardless of spatial frequency. This is a very important characteristic as it allows us to see contrast (and therefore objects) the same regardless of the viewing distance; otherwise, the frequencies would change with the viewing distance and hence the contrast appearance. A partial constancy (invariance) is also observed across luminance [24, 32], though there is a significant deviation from constancy at lower luminance levels, once the visual system needs to rely on the rod vision. The invariance is also an important feature of many computer vision methods. For example in SIFT features [27] have been designed to be invariant to the changes in contrast, brightness, scale, and rotation. In our experiments, we used the supra-threshold contrast matching test to assess whether the models exhibit the characteristic of contrast constancy." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.457, + 0.483, + 0.713 + ], + "angle": 0, + "content": "Numerous works on adversarial attacks demonstrated that the classification performance of deep learning models can be greatly degraded by visually inconsequential changes [20]. At the same time, human vision does not suffer from such adversarial vulnerability [44]. This is one of the most salient arguments put forward to state that deep architectures are different from human vision. Here, we propose a different methodology to study this question. We consider the deep neural network to be a black box and compare its responses to well-understood and measured characteristics of the human visual system. In particular, we want to check whether the foundation vision models share the same \"bottlenecks\" and invariance properties as the visual system. To achieve this, we test foundation models on basic vision stimuli, such as Gabor patches and band-limited noise, and compare the response of those models with the psychophysical data collected from human observers." + }, + { + "type": "text", + "bbox": [ + 0.11, + 0.716, + 0.411, + 0.73 + ], + "angle": 0, + "content": "In summary, our contributions are as follows:" + }, + { + "type": "text", + "bbox": [ + 0.092, + 0.733, + 0.483, + 0.821 + ], + "angle": 0, + "content": "- We developed a protocol to evaluate the similarity between machine vision models and the human visual system. This protocol includes contrast detection, contrast masking, and contrast constancy, subdivided into nine distinct test types that collectively capture the low-level fundamental characteristics of human vision." + }, + { + "type": "text", + "bbox": [ + 0.092, + 0.823, + 0.482, + 0.898 + ], + "angle": 0, + "content": "- We tested the image encoders of 45 foundation and generative models. The results reveal similarities between certain foundation models (e.g., DINOv2 and OpenCLIP) and human vision, particularly in the contrast masking test. However, differences persist across other tests." + }, + { + "type": "list", + "bbox": [ + 0.092, + 0.733, + 0.483, + 0.898 + ], + "angle": 0, + "content": null + }, + { + "type": "title", + "bbox": [ + 0.514, + 0.09, + 0.655, + 0.105 + ], + "angle": 0, + "content": "2. Related Work" + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.116, + 0.905, + 0.418 + ], + "angle": 0, + "content": "Since the advent of deep learning, machine vision models based on foundation models [13, 23, 31] and DNNs [28, 29, 39, 47] have successfully handled numerous advanced visual tasks. However, researchers have observed that machine vision operates differently from human vision. [16] revealed that standard CNNs trained on ImageNet are strongly biased toward texture recognition rather than shape, which contrasts with human visual patterns. [44] provided further evidence that deep neural networks (DNNs) differ significantly from the HVS, demonstrating poor robustness in object classification under 3D viewpoint changes and image distortions, and showing vulnerability to adversarial examples, which are rarely problematic for humans. [5] pointed out that DNNs performing well in benchmark tests share little overlap with biological vision mechanisms and fail to account for many findings in psychological studies of human vision. This highlights a clear distinction between machine and human vision, leading to the rise of interest in domain adaptation [7] and making networks robust to adversarial attacks [46]." + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.42, + 0.905, + 0.585 + ], + "angle": 0, + "content": "But there is also evidence that the gap between neural network-based machine vision models and human vision is gradually narrowing. [40] compared vision transformers (ViT) [13] and CNNs, finding that ViT not only achieves superior task accuracy but also exhibits weaker inductive biases, with error patterns more consistent with human errors. [17] discovered that the long-standing robustness gap between humans and CNNs in handling distortions is shrinking. [10, 19] also demonstrated that foundation models like DINO [9] and CLIP [34] can generate more accurate and robust metrics for low-level perceptual similarity." + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.587, + 0.905, + 0.813 + ], + "angle": 0, + "content": "Most of the aforementioned studies focus on high-level task performance (e.g., accuracy, consistency, ...), which may not reveal whether computation models suffer from the same bottlenecks and rely on the same invariances as human vision. To that end, [2, 26] have attempted to reveal CSF characteristics within pretrained architectures by training a head with a contrast discrimination classifier. The problem with this approach is that it introduces a bias by relying on a classifier trained to compare contrast. Such studies also make an incorrect assumption that CSF explains both near-threshold and super-threshold vision, while contrast constancy results (see Section 4.3) show that this is not the case. In contrast, we examine networks' low-level characteristics without additional task-specific training, considering both near-threshold and supra-threshold vision." + }, + { + "type": "title", + "bbox": [ + 0.514, + 0.829, + 0.696, + 0.846 + ], + "angle": 0, + "content": "3. Testing framework" + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.856, + 0.905, + 0.901 + ], + "angle": 0, + "content": "We first explain the tested models, testing methods, result visualization, and the strategy we used to summarize and quantify our results." + }, + { + "type": "page_number", + "bbox": [ + 0.48, + 0.946, + 0.519, + 0.957 + ], + "angle": 0, + "content": "20040" + } + ], + [ + { + "type": "title", + "bbox": [ + 0.091, + 0.091, + 0.427, + 0.108 + ], + "angle": 0, + "content": "3.1. Tested models and testing methodology" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.113, + 0.484, + 0.324 + ], + "angle": 0, + "content": "The objective is to evaluate the responses of machine vision foundation models to stimuli commonly used in human vision research [3, 8] and to compare these responses with psychophysical human data. We tested a representative set of 45 models, encompassing the most influential large vision foundation models, including the variants of DINO [9], DINOv2 [12, 31], OpenCLIP [22, 33], SAM [23], SAM2 [35], and MAE [21], as well as the encoder used for the latent space of generative model Stable Diffusion (SDVAE, [36]). Additionally, we report the responses of ColorVideoVDP [30], which is an image and video quality metric that explicitly models low-level human vision and acts as a reference for a low-level human vision model. All models and their variants are listed in Figure 6." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.325, + 0.484, + 0.626 + ], + "angle": 0, + "content": "To test these models, we need to compare pairs of images and assess the \"perceived\" difference between them. We adopt a methodology inspired by 2-alternative-forced-choice (2AFC) psychophysical experiments. For example, for a pattern detection task, a pair of images could be a Gabor patch (test) and a uniform field (reference), as shown on the left of Figure 2. As such patterns are calibrated in physical light units in vision science, we generate these patterns as luminance maps, scaled in physical units of \\(\\mathrm{cd} / \\mathrm{m}^2\\). These luminance maps are then mapped from the linear space to the sRGB color space using a display model (display peak luminance of \\(400\\mathrm{cd} / \\mathrm{m}^2\\)) and fed into the image encoder of the foundation model for feature extraction. Note that the sRGB space is almost universally used to represent training datasets and is expected input for the tested encoders. To ensure that models can operate on small contrast values, we modified them to accept floating-point values (instead of 8-bit integers) as input. This was necessary as the quantization artifacts in 8-bit images are often larger than the detection thresholds of human vision." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.627, + 0.484, + 0.718 + ], + "angle": 0, + "content": "To investigate whether the distances in the feature space reflect the perceptual detection thresholds and invariances, we experimented with a series of distance measures, including \\( L_{1} \\) and \\( L_{2} \\). We found the cosine similarity expressed as a relative angle \\( (S_{\\mathrm{ac}}) \\) yielded results most consistent with the psychophysical data. \\( S_{\\mathrm{ac}} \\) is defined as:" + }, + { + "type": "equation", + "bbox": [ + 0.177, + 0.725, + 0.483, + 0.759 + ], + "angle": 0, + "content": "\\[\nS _ {\\mathrm {a c}} = \\frac {1}{\\pi} \\arccos \\left(\\frac {F _ {\\mathrm {T}} \\cdot F _ {\\mathrm {R}}}{\\| F _ {\\mathrm {T}} \\| \\| F _ {\\mathrm {R}} \\|}\\right), \\tag {1}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.765, + 0.483, + 0.826 + ], + "angle": 0, + "content": "where \\(F_{\\mathrm{T}}\\) and \\(F_{\\mathrm{R}}\\) are the test and reference feature vectors (feature maps reshaped into one dimension), and \\(\\cdot\\) denotes the dot product. \\(S_{\\mathrm{ac}} = 0\\) indicates two input images are equivalent, while \\(S_{\\mathrm{ac}} = 1\\) indicates large differences." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.826, + 0.484, + 0.901 + ], + "angle": 0, + "content": "To compare the encoder responses with psychophysical data, we need to be able to map the image sampling frequency into the spatial frequency on the retina. For that, we select the effective resolution of 60 pixels-per-degree, which is typical for modern monitors. We note, however," + }, + { + "type": "image", + "bbox": [ + 0.518, + 0.089, + 0.907, + 0.277 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.512, + 0.286, + 0.907, + 0.356 + ], + "angle": 0, + "content": "Figure 2. Pipeline for computing \\( S_{\\mathrm{ac}} \\). Generate test and reference images in the linear luminance space (cd/m²), transform them to the sRGB color space using a display model, and input each into the image encoder. Reshape the output features into one-dimensional vectors \\( F_{\\mathrm{T}} \\) and \\( F_{\\mathrm{R}} \\), then compute \\( S_{\\mathrm{ac}} \\)." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.383, + 0.906, + 0.503 + ], + "angle": 0, + "content": "that the choice of this parameter is arbitrary and the model similarity scores can be shifted by a small multiplier along the spatial frequency axis. The luminance maps are generated at the resolution of \\(224 \\times 224\\) pixels, corresponding to the size of \\(3.7 \\times 3.7\\) visual degrees. It roughly aligned with the extend of the human foveal vision, and the stimuli span across multiple receptive fields of human vision and multiple patches/tokens of a foundation model." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.505, + 0.906, + 0.596 + ], + "angle": 0, + "content": "For ColorVideoVDP, it can work directly on linear physical units, so conversion to sRGB was unnecessary. We directly use its quality score instead of \\( S_{\\mathrm{ac}} \\). Note that the primary focus of this study is on foundation models, with the ColorVideoVDP metric used solely as a baseline for comparison purposes." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.606, + 0.725, + 0.621 + ], + "angle": 0, + "content": "3.2. Model alignment score" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.628, + 0.906, + 0.688 + ], + "angle": 0, + "content": "Most of our results will be represented as contour plots of model responses, providing qualitative interpretation. Here, we explain our measure of model alignment, which provides quantitative scores." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.689, + 0.906, + 0.87 + ], + "angle": 0, + "content": "As an example, we take the leftmost contour plot in row (a) of Figure 5. Each point on the contour plot corresponds to \\( S_{\\mathrm{ac}} \\) between the test image as shown in Figure 3 and a uniform field of the same mean luminance. The dashed line represents human contrast detection data, predicted with castleCSF [3]. A well-aligned model should show one of the contour lines that follows the dashed castleCSF line. We cannot directly use the \\( S_{\\mathrm{ac}} \\) values along the dashed line as the measure of alignment because some models result in \\( S_{\\mathrm{ac}} = 0 \\) for most points near the detection threshold (detect no difference). Therefore, instead, we rely on the measure of change in \\( S_{\\mathrm{ac}} \\) in the neighborhood of the dashed line." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.871, + 0.906, + 0.901 + ], + "angle": 0, + "content": "For a well-aligned model, the perceived differences in the neighborhood of the detection threshold (dashed line)" + }, + { + "type": "page_number", + "bbox": [ + 0.48, + 0.945, + 0.518, + 0.957 + ], + "angle": 0, + "content": "20041" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.097, + 0.089, + 0.488, + 0.364 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.09, + 0.373, + 0.486, + 0.445 + ], + "angle": 0, + "content": "Figure 3. Gabors with different spatial frequencies (x-axis) and contrast (y-axis) used as the test images in the contrast detection tests (Section 4.1). \"cpd\" denotes cycles per degree. Note that the high-frequency patterns can be rendered with aliasing artifacts on the screen or in print — those were not present in our tests." + }, + { + "type": "text", + "bbox": [ + 0.089, + 0.471, + 0.483, + 0.652 + ], + "angle": 0, + "content": "should increase as the contrast increases (the sensitivity decreases). Furthermore, the values should be similar along the dashed line. The measure these two properties, we sample the \\( S_{\\mathrm{ac}} \\) values for the points that are shifted in contrast (vertical direction) from the dashed line by a multiplier \\( m \\), where \\( 0.5 \\leq m \\leq 2 \\) (note the logarithmic scale in the contour plot in Figure 5). We collect such data for multiple frequencies (or other dimensions) along the dashed line and calculate the Spearman rank order correlation between the multipliers \\( m \\) and the \\( S_{\\mathrm{ac}} \\) values. If the properties mentioned above are preserved, the correlation coefficient value \\( r_s \\) should be close to 1." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.654, + 0.484, + 0.73 + ], + "angle": 0, + "content": "The above strategy is used for all contrast detection and contrast masking experiments. For the contrast matching experiment, we use the root mean squared error (RMSE) between the model and human matching data, expressed as the logarithm of contrast." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.746, + 0.225, + 0.763 + ], + "angle": 0, + "content": "4. Experiments" + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.772, + 0.342, + 0.787 + ], + "angle": 0, + "content": "4.1. Contrast detection and CSF" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.795, + 0.484, + 0.903 + ], + "angle": 0, + "content": "We begin by testing the foundation models' ability to detect low-contrast (near-threshold) patterns (e.g., Gabor patches, band-limited noise) and compare their performance with the human data. As the reference human data, we rely on the castleCSF [3], which is the recent contrast sensitivity function, modeling contrast detection of both achromatic and chromatic patterns." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.092, + 0.905, + 0.333 + ], + "angle": 0, + "content": "Spatial Frequency Contrast sensitivity of the human eye is typically associated with the variation across the spatial frequency. The visual system exhibits a band-pass characteristic, with a peak sensitivity between 2 and 4 cycles per degree (cpd), depending on the luminance and other parameters of the stimulus. The lower sensitivity of the visual system at lower frequencies is associated with the mechanism of lateral inhibition [4], which helps to reduce the influence of (low-frequency) illumination on the perceived images. Such invariance to illumination is a desirable property if the goal is to recognize objects regardless of illumination conditions. The drop in sensitivity at high frequencies is associated with the limitations of eye optics (achromatic contrast) and cone density (chromatic contrast). Here we want to test whether the computer-vision foundation models pick up similar traits when trained on natural images." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.336, + 0.907, + 0.624 + ], + "angle": 0, + "content": "To test encoder responses across frequencies, we generated a 2D array of image pairs, in which the reference image had uniform luminance, and the test image contained a Gabor patch, as shown in Figure 3 (refer to Supplementary for the visualization of other stimuli). We generate such Gabors for achromatic (see Figure 4-a) and chromatic modulation (see Figure 4-c,d). We also tested band-limited noise (see Figure 4-b). The test patterns had fixed size but varying spatial frequency (x-axis) and contrast (y-axis). Because most of the contrast detection data are plotted as the function of sensitivity, we follow this convention and plot the sensitivity, which is the inverse of the contrast. This corresponds to the reversal of the axis on the logarithmic plots, which we use in our analysis. The other parameters of the stimuli are listed in Table 1. Although we tested multiple variants of each foundation model, here we show only the variant with the highest complexity and best performance on its original high-level tasks. The contour plots for other variants can be found in the Supplementary." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.627, + 0.909, + 0.882 + ], + "angle": 0, + "content": "The contour plots of foundation model responses are shown together with a contrast sensitivity function (castleCSF [3]) in rows (a)-(d) of Figure 5. If the foundation models had the same contrast detection characteristics as the human eye, we would expect the smallest difference (the lowest \\( S_{\\mathrm{ac}} \\)) contour line to follow the black dashed curve of castleCSF. This is not the case for any of the tested foundation models. Some models, and in particular DINOv2 and SD-VAE, show overall band-pass characteristics, in particular for noise and achromatic Gabors. SD-VAE has a drop of sensitivity at lower frequencies much larger than that of the visual system. The responses to chromatic patterns (rows (c) and (d)) tend to be less regular than to achromatic patterns and lack the shift toward lower frequencies, which is observed in the human data. Both OpenCLIP and SAM-2 show a very inconsistent response across the spatial frequencies." + }, + { + "type": "text", + "bbox": [ + 0.532, + 0.886, + 0.907, + 0.901 + ], + "angle": 0, + "content": "From the data, we can conclude that foundation models" + }, + { + "type": "page_number", + "bbox": [ + 0.48, + 0.945, + 0.521, + 0.958 + ], + "angle": 0, + "content": "20042" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.095, + 0.09, + 0.294, + 0.284 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.299, + 0.092, + 0.498, + 0.284 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.501, + 0.092, + 0.699, + 0.284 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.703, + 0.092, + 0.904, + 0.284 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.09, + 0.296, + 0.908, + 0.325 + ], + "angle": 0, + "content": "Figure 4. Examples of test images for contrast detection and contrast masking. The detailed explanation of the stimuli can be found in the Supplementary. \"Ach.\" denotes achromatic." + }, + { + "type": "table_caption", + "bbox": [ + 0.178, + 0.339, + 0.819, + 0.354 + ], + "angle": 0, + "content": "Table 1. Key parameters for all our tests. Note that \"Radius\" does not apply when the pattern is not a Gabor." + }, + { + "type": "table", + "bbox": [ + 0.095, + 0.364, + 0.905, + 0.508 + ], + "angle": 0, + "content": "
TestSpatial Frequency (cpd)Luminance (cd/m2)Radius (degree)Contrast
Spatial Frequency - Gabor Achromatic0.5 - 3210010.001 - 1
Spatial Frequency - Noise Achromatic0.5 - 32100-0.001 - 1
Spatial Frequency - Gabor RG0.5 - 3210010.001 - 0.12
Spatial Frequency - Gabor YV0.5 - 3210010.001 - 0.8
Luminance - Gabor Achromatic20.1 - 20010.001 - 1
Area - Gabor Achromatic81000.1 - 10.001 - 1
Phase-Coherent Masking2 (mask) / 2 (test)320.5 (test)0.005 - 0.5 (mask) / 0.01 - 0.5 (test)
Phase-Incoherent Masking0 - 12 (mask) / 1.2 (test)370.8 (test)0.005 - 0.5 (mask) / 0.01 - 0.5 (test)
Contrast Matching5 (reference) / 0.25 - 25 (test)10-0.005 - 0.629 (reference)
" + }, + { + "type": "text", + "bbox": [ + 0.089, + 0.546, + 0.483, + 0.637 + ], + "angle": 0, + "content": "do not follow the sensitivity pattern of the visual system, but some may show a band-pass characteristic that is associated with the CSF. Many models show lower sensitivity at lower frequencies, which may indicate that those models obtained some invariance to (low-frequency) illumination through training." + }, + { + "type": "text", + "bbox": [ + 0.089, + 0.659, + 0.483, + 0.809 + ], + "angle": 0, + "content": "Luminance The sensitivity of the human eye increases with the luminance. In dim light, human contrast sensitivity increases proportionally to the square root of retinal illuminance, following the DeVries-Rose law. Conversely, in bright light, sensitivity follows Weber's law, remaining independent of illuminance [38]. In this experiment, we want to test whether the computer vision models lose sensitivity at lower luminance levels. Such a loss could be justified by camera noise, which increases in terms of contrast as the light intensity decreases [1]." + }, + { + "type": "text", + "bbox": [ + 0.089, + 0.811, + 0.484, + 0.9 + ], + "angle": 0, + "content": "To produce contour plots, we generated a 2D array of image pairs in a similar manner as for spatial frequency variations in the section above, but instead of varying spatial frequency, we varied luminance, as shown in Figure 4e. As a reminder, we did not pass the absolute luminance values directly to each model but instead converted" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.546, + 0.907, + 0.606 + ], + "angle": 0, + "content": "them to display-encoded (gamma-corrected) sRGB space — the colour space used for training datasets (see Figure 2). Such an encoding partially compensates for perceptual non-uniformity of luminance." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.607, + 0.906, + 0.681 + ], + "angle": 0, + "content": "The results, shown in row (e) of Figure 5, indicate a systematic drop in sensitivity with luminance for all tested models. The drop in sensitivity of those models is faster than that observed in the human data, in particular for OpenCLIP." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.683, + 0.906, + 0.789 + ], + "angle": 0, + "content": "We can conclude that the models trained on sufficiently large datasets mimic the sensitivity of the visual system to luminance but with a faster drop in sensitivity. This, however, may be the result of using sRGB representation for training datasets, which were presumably mostly well-exposed and contained relatively little information in darker image regions." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.811, + 0.906, + 0.9 + ], + "angle": 0, + "content": "Area Stimulus area (size) significantly affects sensitivity, as larger stimuli activate more retinal cells. Sensitivity increases with area up to the saturation point, denoted as the critical area [37]. The response to stimuli of different sizes tests the model's ability to pool information across the visual field." + }, + { + "type": "page_number", + "bbox": [ + 0.479, + 0.945, + 0.518, + 0.957 + ], + "angle": 0, + "content": "20043" + } + ], + [ + { + "type": "image_caption", + "bbox": [ + 0.095, + 0.139, + 0.113, + 0.15 + ], + "angle": 0, + "content": "(a)" + }, + { + "type": "image", + "bbox": [ + 0.117, + 0.09, + 0.293, + 0.183 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.349, + 0.09, + 0.392, + 0.1 + ], + "angle": 0, + "content": "DINOv2" + }, + { + "type": "image", + "bbox": [ + 0.298, + 0.106, + 0.443, + 0.183 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.498, + 0.091, + 0.556, + 0.106 + ], + "angle": 0, + "content": "SAM-2 AM2.1-hiera-L" + }, + { + "type": "image", + "bbox": [ + 0.451, + 0.107, + 0.596, + 0.183 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.096, + 0.214, + 0.113, + 0.225 + ], + "angle": 0, + "content": "(b)" + }, + { + "type": "image", + "bbox": [ + 0.117, + 0.183, + 0.293, + 0.258 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.298, + 0.183, + 0.444, + 0.258 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.45, + 0.182, + 0.596, + 0.258 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.096, + 0.29, + 0.112, + 0.3 + ], + "angle": 0, + "content": "(c)" + }, + { + "type": "image", + "bbox": [ + 0.117, + 0.258, + 0.293, + 0.332 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.298, + 0.257, + 0.444, + 0.333 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.45, + 0.257, + 0.596, + 0.333 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.096, + 0.365, + 0.112, + 0.375 + ], + "angle": 0, + "content": "(d)" + }, + { + "type": "image", + "bbox": [ + 0.117, + 0.333, + 0.293, + 0.407 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.298, + 0.333, + 0.444, + 0.408 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.45, + 0.333, + 0.596, + 0.408 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.096, + 0.441, + 0.112, + 0.45 + ], + "angle": 0, + "content": "(e)" + }, + { + "type": "image", + "bbox": [ + 0.117, + 0.408, + 0.293, + 0.483 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.298, + 0.408, + 0.444, + 0.483 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.45, + 0.408, + 0.596, + 0.483 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.097, + 0.516, + 0.112, + 0.525 + ], + "angle": 0, + "content": "(f)" + }, + { + "type": "image", + "bbox": [ + 0.117, + 0.483, + 0.293, + 0.55 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.298, + 0.482, + 0.444, + 0.551 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.45, + 0.482, + 0.596, + 0.551 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.096, + 0.591, + 0.112, + 0.6 + ], + "angle": 0, + "content": "(g)" + }, + { + "type": "image", + "bbox": [ + 0.117, + 0.552, + 0.293, + 0.627 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.298, + 0.552, + 0.444, + 0.627 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.45, + 0.552, + 0.596, + 0.627 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.096, + 0.666, + 0.112, + 0.676 + ], + "angle": 0, + "content": "(h)" + }, + { + "type": "image", + "bbox": [ + 0.117, + 0.629, + 0.293, + 0.703 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.298, + 0.629, + 0.444, + 0.703 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.45, + 0.629, + 0.596, + 0.703 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.097, + 0.741, + 0.112, + 0.75 + ], + "angle": 0, + "content": "(i)" + }, + { + "type": "image", + "bbox": [ + 0.117, + 0.705, + 0.293, + 0.785 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.298, + 0.705, + 0.444, + 0.784 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.45, + 0.705, + 0.596, + 0.784 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.627, + 0.09, + 0.748, + 0.183 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.602, + 0.183, + 0.747, + 0.258 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.602, + 0.257, + 0.747, + 0.333 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.602, + 0.333, + 0.747, + 0.408 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.602, + 0.408, + 0.748, + 0.482 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.602, + 0.482, + 0.748, + 0.55 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.602, + 0.55, + 0.748, + 0.627 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.602, + 0.629, + 0.748, + 0.703 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.602, + 0.705, + 0.748, + 0.784 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.755, + 0.09, + 0.901, + 0.183 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.755, + 0.183, + 0.901, + 0.258 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.755, + 0.257, + 0.901, + 0.333 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.755, + 0.333, + 0.901, + 0.408 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.755, + 0.408, + 0.901, + 0.477 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.755, + 0.477, + 0.901, + 0.549 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.755, + 0.55, + 0.901, + 0.627 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.755, + 0.629, + 0.901, + 0.699 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.755, + 0.705, + 0.901, + 0.784 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.09, + 0.795, + 0.907, + 0.879 + ], + "angle": 0, + "content": "Figure 5. Selected representative experimental results. Each row represents a test, and each column corresponds to a model, selected as the best-performing in their original tasks. (a)-(f): contour plots of contrast detection \\( S_{\\mathrm{ac}} \\), with the ground truth castleCSF [3]. (g),(h): contour plots of contrast masking \\( S_{\\mathrm{ac}} \\), with the ground truth from [14] and [15], respectively. Different colored solid lines representing different \\( S_{\\mathrm{ac}} \\) values, where purple indicates the minimum difference \\( (S_{\\mathrm{ac}} \\rightarrow 0) \\). (i): results from the contrast matching experiment, where different colors represent different \\( C_r \\) values. The dashed lines are human results [18], while the solid lines are model-predicted results from Equation 2. Results for all models are in the supplementary material." + }, + { + "type": "page_number", + "bbox": [ + 0.48, + 0.945, + 0.52, + 0.957 + ], + "angle": 0, + "content": "20044" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.13, + 0.09, + 0.468, + 0.1 + ], + "angle": 0, + "content": "Contrast Detection - Spatial Frequency - Gabor Achromatic - Spearman Correlation ↑" + }, + { + "type": "image", + "bbox": [ + 0.099, + 0.106, + 0.48, + 0.147 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.099, + 0.152, + 0.48, + 0.192 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.099, + 0.198, + 0.48, + 0.238 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.099, + 0.242, + 0.48, + 0.285 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.099, + 0.289, + 0.48, + 0.332 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.099, + 0.334, + 0.48, + 0.378 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.099, + 0.382, + 0.48, + 0.425 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.099, + 0.428, + 0.48, + 0.472 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.095, + 0.474, + 0.48, + 0.601 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.09, + 0.613, + 0.483, + 0.711 + ], + "angle": 0, + "content": "Figure 6. The quantified similarity error between all 45 models and HVS under 9 different tests. For the Contrast Detection and Contrast Masking tasks, Spearman Correlation was used as the metric, with higher values (closer to 1) indicating greater similarity to human vision. For the Supra-threshold Contrast Matching task, RMSE was used as the metric, with lower values (closer to 0) indicating better similarity." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.746, + 0.483, + 0.837 + ], + "angle": 0, + "content": "In this experiment, we vary the size of the Gaussian envelope limiting stimulus size to observe its effect on model responses. We follow the same procedure as in the two previous sections and generate pairs with a uniform field and a Gabor patch (see Figure 4-f). The parameters of the Gabor patch are listed in Table 1." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.84, + 0.484, + 0.9 + ], + "angle": 0, + "content": "The results, shown in row (f) of Figure 5, indicate that most models show summation across the area (the \\( S_{\\mathrm{ac}} \\) values increase with the area) and the rate of the increase varies across the models; DINO and SAM-2 have a smaller in" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.092, + 0.906, + 0.183 + ], + "angle": 0, + "content": "crease, SD-VAR has a higher increase of \\( S_{\\mathrm{ac}} \\) than the human data, and only DINOv2 roughly matches human performance. OpenCLIP shows no consistent patterns. We can conclude that many models show spatial pooling characteristic that shares the trend observed in human data, though the actual slope of the increase is typically different." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.195, + 0.689, + 0.211 + ], + "angle": 0, + "content": "4.2. Contrast masking" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.218, + 0.907, + 0.476 + ], + "angle": 0, + "content": "Contrast masking explains the decreased visibility of a signal (test) due to the presence of a supra-threshold background (mask). The masking function defines the relationship between the threshold test contrast required for signal detection and the mask contrast. Put simply, a pattern is more difficult to detect in the presence of another pattern of similar spatial frequency and orientation. A typical masking characteristic of the visual system is shown in rows (g) and (h) of Figure 5. It consists of a relatively shallow segment at low mask contrast (near the detection threshold), with the slope increasing for high mask contrast. The shape of the curve is influenced by the specific properties of the mask and test signals [11, 43]. We will consider the case in which the masker is a sinusoidal grating of the same frequency as the test pattern (phase-coherent masking, row (g)) and when the masker is noise (phase-incoherent masking, row (h)) [11, 14, 15], as shown in Figure 4-g,h." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.477, + 0.906, + 0.718 + ], + "angle": 0, + "content": "First, we consider the fundamental form of phase-coherent masking in which the masker image (reference in Figure 2) is a sinusoidal grating and the test image is the masker plus a Gabor patch of the same frequency and phase as the masker (see Figure 4g). Contrast masking data, shown as the dashed black curve in row (g) of Figure 5, shows the smallest contrast of the test Gabor that is detectable in the presence of the masker of a given contrast. As the contrast of the masker increases, the smallest detectable contrast of the test also needs to increase. However, such an increase starts only for a masker that has the contrast sufficiently high to be detected. If the contrast of the masker is near the detection threshold, we can observe a dipper effect — the contrast detection is facilitated by a masker [14, 42, 45]. Such an effect can be only observed in phase coherent masking." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.719, + 0.906, + 0.809 + ], + "angle": 0, + "content": "As an example of phase-incoherent masking, we will consider a masker with band-limited noise and a test with a Gabor patch — see Figure 4h. The human detection thresholds for such masking patterns are similar to those for phase-coherent masking, except that the dipper effect disappears [41]." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.81, + 0.909, + 0.901 + ], + "angle": 0, + "content": "The differences predicted by DINOv2 and OpenCLIP are surprisingly well-aligned with the human contrast masking data — their responses roughly match the slopes of the human data. The alignment is stronger for the phase-incoherent masking. This is particularly notable for OpenCLIP, which did not show any consistent trends for contrast" + }, + { + "type": "page_number", + "bbox": [ + 0.48, + 0.945, + 0.518, + 0.957 + ], + "angle": 0, + "content": "20045" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.09, + 0.092, + 0.483, + 0.121 + ], + "angle": 0, + "content": "detection. Other models do not show strong alignment with the human data." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.122, + 0.484, + 0.244 + ], + "angle": 0, + "content": "Overall, computational models are better aligned with human data for contrast masking than for contrast detection. One possible explanation is that the signals that induce contrast masking are plentiful in natural images, but contrast detection stimuli, which involve barely noticeable patterns on uniform backgrounds, are rare. Therefore, computational models are more likely to pick up the characteristic that is well represented in the training datasets." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.25, + 0.399, + 0.267 + ], + "angle": 0, + "content": "4.3. Supra-threshold contrast matching" + }, + { + "type": "text", + "bbox": [ + 0.093, + 0.273, + 0.485, + 0.603 + ], + "angle": 0, + "content": "While contrast detection and contrast masking explain the just detectable (near-threshold) contrast, most of the vision tasks, such as detection or recognition, involve well-visible (supra-threshold) contrast. Supra-threshold human vision has been studied in contrast-matching experiments in which the magnitude of one contrast is visually matched to the magnitude of another contrast of a different frequency [18] or luminance [24, 32]. One of the most significant findings of those studies is contrast constancy [18] — the ability of the visual system to match physical contrast across frequencies and luminance levels. The results of the seminal study of Georgeson and Sullivan [18] on matching contrast across frequencies are shown as dashed lines in row (i) of Figure 5. At small contrast, the dashed lines show a band-pass shape that follows the contrast sensitivity function. However, as the contrast is increased, the lines become flat showing little influence of frequency on contrast perception. This is an important property that lets us see objects to have the same appearance regardless of the viewing distance. Such a scale invariance is also important for neural networks that are tasked to detect or recognize objects regardless of their size." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.605, + 0.484, + 0.772 + ], + "angle": 0, + "content": "We followed the experimental setup from [18], where the reference was a \\(5\\mathrm{cpd}\\), \\(10\\mathrm{cd} / \\mathrm{m}^2\\) sinusoidal grating, presented at eight distinct contrast levels \\(c_{\\mathrm{r}}\\). The test stimulus had the same luminance but a different spatial frequency \\(\\rho_{\\mathrm{t}}\\). In [18], observers adjusted the test stimulus contrast \\(c_{\\mathrm{t}}\\) until its apparent contrast matched that of the reference (contrast matching). In our experiments, we match contrast encodings of sinusoidal gratings: the \\((S_{\\mathrm{ac}}()\\), eq. (1)), between a feature vector of a sinusoidal grating, \\(F(\\rho ,c)\\) of frequency \\(\\rho\\) and contrast \\(c\\), and a uniform field, \\(U = F(\\rho_{\\mathrm{t}},0)\\). We find the test contrast \\(c_{\\mathrm{t}}\\) that minimizes the expression:" + }, + { + "type": "equation", + "bbox": [ + 0.113, + 0.777, + 0.483, + 0.805 + ], + "angle": 0, + "content": "\\[\n\\underset {c _ {\\mathrm {t}}} {\\operatorname {a r g m i n}} \\left(S _ {\\mathrm {a c}} \\left(F \\left(\\rho_ {\\mathrm {r}}, c _ {\\mathrm {r}}\\right), U\\right) - S _ {\\mathrm {a c}} \\left(F \\left(\\rho_ {\\mathrm {t}}, c _ {\\mathrm {t}}\\right), U\\right)\\right) ^ {2}, \\tag {2}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.811, + 0.484, + 0.903 + ], + "angle": 0, + "content": "where the reference frequency \\(\\rho_{\\mathrm{r}} = 5\\) cpd. The denominators in the expression are used to normalize contrast across frequencies. We experimented with other contrast encodings, including a direct comparison of feature vectors, but the formula above resulted in the best contrast constancy properties across the models." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.092, + 0.907, + 0.304 + ], + "angle": 0, + "content": "The matching contrast predictions for foundation models are visualized as continuous lines in row (i) of Figure 5. The plots show that only DINOv2 and OpenCLIP roughly follow the dashed contrast constancy lines. Both models show less attenuation (more constancy) at the highest spatial frequencies, which could be advantageous when the model needs to work with small-scale features. Both models show attenuation of low frequencies (below 1 cpd), suggesting worse contrast constancy in that frequency range. Other models, including DINO and SAM-2, suffer from large instability across frequencies, or very heavy attenuation of low frequencies in the case of SD-VAE. To conclude, we can observe only partial contrast constancy for selected models." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.322, + 0.733, + 0.339 + ], + "angle": 0, + "content": "4.4. Model alignment scores" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.348, + 0.907, + 0.422 + ], + "angle": 0, + "content": "As the analysis of all 45 variants of the models is infeasible in the scope of this paper, we prepared quantitative results according to the method explained in Section 3.2 and summarized them in Figure 6. Those let us make three observations:" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.426, + 0.909, + 0.624 + ], + "angle": 0, + "content": "First, ColorVideoVDP, which is a visual metric that models human low-level vision, is better aligned with the human data in almost all contrast detection tasks and in terms of contrast constancy, as expected. However, certain variants of OpenCLIP and DINOv2 can match or surpass the ColorVideoVDP alignment in terms of contrast masking. Second, the alignment scores of different foundation model variants (e.g., OpenCLIP) show significant variations and alignment scores appear unrelated to the complexity of the variants or their performance on higher-level tasks. Finally, DINOv2 variants, which have been trained to solve vision tasks, show the greatest alignment with the human data among all foundation models." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.646, + 0.642, + 0.661 + ], + "angle": 0, + "content": "5. Conclusions" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.675, + 0.909, + 0.903 + ], + "angle": 0, + "content": "If we believe that the goal of both biological and computational low-level vision is to efficiently encode visual information, we can expect that computational models trained on large natural image datasets will share similarities with human vision. In this work, we find that selected computational models, e.g., variants of DINOv2 and OpenCLIP, show surprisingly high alignment with supra-threshold human contrast masking and contrast matching data, but little alignment with the near-threshold contrast detection. This means that computation models do not have the same \"bottlenecks\" as human vision, but, through training, they attain invariance and efficient contrast coding that resembles that of the visual system. We hope that our testing protocol with basic psychophysical stimuli will provide a useful tool for examining future computational models of vision." + }, + { + "type": "page_number", + "bbox": [ + 0.48, + 0.946, + 0.521, + 0.957 + ], + "angle": 0, + "content": "20046" + } + ], + [ + { + "type": "title", + "bbox": [ + 0.093, + 0.09, + 0.188, + 0.106 + ], + "angle": 0, + "content": "References" + }, + { + "type": "ref_text", + "bbox": [ + 0.101, + 0.115, + 0.484, + 0.171 + ], + "angle": 0, + "content": "[1] Cecilia Aguerrebere, Julie Delon, Yann Gousseau, and Pablo Muse. Study of the digital camera acquisition process and statistical modeling of the sensor raw data. 2013. Citation Key: Aguerrebere2013. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.101, + 0.172, + 0.484, + 0.212 + ], + "angle": 0, + "content": "[2] Arash Akbarinia, Yaniv Morgenstern, and Karl R Gegenfurtner. Contrast sensitivity function in deep networks. Neural Networks, 164:228-244, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.213, + 0.484, + 0.269 + ], + "angle": 0, + "content": "[3] Maliha Ashraf, Rafal K Mantiuk, Alexandre Shapiro, and Sophie Wuerger. castlecsf—a contrast sensitivity function of color, area, spatiotemporal frequency, luminance and eccentricity. Journal of Vision, 24(4):5-5, 2024. 3, 4, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.27, + 0.483, + 0.298 + ], + "angle": 0, + "content": "[4] Peter GJ Barten. Contrast sensitivity of the human eye and its effects on image quality. SPIE press, 1999. 1, 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.299, + 0.482, + 0.38 + ], + "angle": 0, + "content": "[5] Jeffrey S Bowers, Gaurav Malhotra, Marin Dujmovic, Milton Llera Montero, Christian Tsvetkov, Valerio Biscione, Guillermo Puebla, Federico Adolfi, John E Hummel, Rachel F Heaton, et al. Deep problems with neural network models of human vision. Behavioral and Brain Sciences, 46: e385, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.382, + 0.482, + 0.422 + ], + "angle": 0, + "content": "[6] Oliver Braddick and Janette Atkinson. Development of human visual function. Vision Research, 51(13):1588-1609, 2011. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.424, + 0.483, + 0.506 + ], + "angle": 0, + "content": "[7] Yancheng Cai, Bo Zhang, Baopu Li, Tao Chen, Hongliang Yan, Jingdong Zhang, and Jiahao Xu. Rethinking cross-domain pedestrian detection: a background-focused distribution alignment framework for instance-free one-stage detectors. IEEE transactions on image processing, 32:4935-4950, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.508, + 0.482, + 0.564 + ], + "angle": 0, + "content": "[8] Yancheng Cai, Ali Bozorgian, Maliha Ashraf, Robert Wanat, and Rafal Mantiuk. elatcsf: A temporal contrast sensitivity function for flicker detection and modeling variable refresh rate flicker. SIGGRAPH Asia 2024 [in print], 2024. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.565, + 0.482, + 0.633 + ], + "angle": 0, + "content": "[9] Mathilde Caron, Hugo Touvron, Ishan Misra, Hervé Jégou, Julien Mairal, Piotr Bojanowski, and Armand Joulin. Emerging properties in self-supervised vision transformers. In Proceedings of the IEEE/CVF international conference on computer vision, pages 9650-9660, 2021. 1, 2, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.635, + 0.483, + 0.689 + ], + "angle": 0, + "content": "[10] Francesco Croce, Christian Schlarmann, Naman Deep Singh, and Matthias Hein. Adversarily robust clip models induce better (robust) perceptual metrics. In ICML 2024 Workshop on Foundation Models in the Wild, 2024. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.69, + 0.483, + 0.746 + ], + "angle": 0, + "content": "[11] Scott J Daly. Visible differences predictor: an algorithm for the assessment of image fidelity. In Human Vision, Visual Processing, and Digital Display III, pages 2-15. SPIE, 1992. 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.747, + 0.483, + 0.774 + ], + "angle": 0, + "content": "[12] Timothee Darcet, Maxime Oquab, Julien Mairal, and Piotr Bojanowski. Vision transformers need registers, 2023. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.775, + 0.483, + 0.815 + ], + "angle": 0, + "content": "[13] Alexey Dosovitskiy. An image is worth 16x16 words: Transformers for image recognition at scale. arXiv preprint arXiv:2010.11929, 2020. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.817, + 0.483, + 0.858 + ], + "angle": 0, + "content": "[14] John M Foley. Human luminance pattern-vision mechanisms: masking experiments require a new model. JOSA A, 11(6):1710-1719, 1994. 6, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.86, + 0.483, + 0.9 + ], + "angle": 0, + "content": "[15] Karl R Gegenfurtner and Daniel C Kiper. Contrast detection in luminance and chromatic noise. JOSA A, 9(11):1880-1888, 1992. 6, 7" + }, + { + "type": "list", + "bbox": [ + 0.094, + 0.115, + 0.484, + 0.9 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.093, + 0.906, + 0.161 + ], + "angle": 0, + "content": "[16] Robert Geirhos, Patricia Rubisch, Claudio Michaelis, Matthias Bethge, Felix A Wichmann, and Wieland Brendel. Imagenet-trained cnns are biased towards texture; increasing shape bias improves accuracy and robustness. arXiv preprint arXiv:1811.12231, 2018. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.164, + 0.906, + 0.245 + ], + "angle": 0, + "content": "[17] Robert Geirhos, Kantharaju Narayanappa, Benjamin Mitzkus, Tizian Thieringer, Matthias Bethge, Felix A Wichmann, and Wieland Brendel. Partial success in closing the gap between human and machine vision. Advances in Neural Information Processing Systems, 34:23885-23899, 2021. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.248, + 0.905, + 0.289 + ], + "angle": 0, + "content": "[18] MA Georgeson and GD Sullivan. Contrast constancy: deblurring in human vision by spatial frequency channels. The Journal of physiology, 252(3):627-656, 1975. 1, 2, 6, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.291, + 0.905, + 0.343 + ], + "angle": 0, + "content": "[19] Abhijay Ghildyal, Nabajeet Barman, and Saman Zadtootaghaj. Foundation models boost low-level perceptual similarity metrics. arXiv preprint arXiv:2409.07650, 2024. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.346, + 0.904, + 0.375 + ], + "angle": 0, + "content": "[20] Ian J. Goodfellow, Jonathon Shlens, and Christian Szegedy. Explaining and harnessing adversarial examples. 2014. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.377, + 0.906, + 0.445 + ], + "angle": 0, + "content": "[21] Kaiming He, Xinlei Chen, Saining Xie, Yanghao Li, Piotr Dólár, and Ross Girshick. Masked autoencoders are scalable vision learners. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 16000-16009, 2022. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.447, + 0.905, + 0.529 + ], + "angle": 0, + "content": "[22] Gabriel Ilharco, Mitchell Wortsman, Ross Wightman, Cade Gordon, Nicholas Carlini, Rohan Taori, Achal Dave, Vaishaal Shankar, Hongseok Namkoong, John Miller, Hannaneh Hajishirzi, Ali Farhadi, and Ludwig Schmidt. Openclip, 2021. If you use this software, please cite it as below. 1, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.532, + 0.905, + 0.601 + ], + "angle": 0, + "content": "[23] Alexander Kirillov, Eric Mintun, Nikhila Ravi, Hanzi Mao, Chloe Rolland, Laura Gustafson, Tete Xiao, Spencer Whitehead, Alexander C Berg, Wan-Yen Lo, et al. Segment anything. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 4015-4026, 2023. 2, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.602, + 0.905, + 0.643 + ], + "angle": 0, + "content": "[24] J.J. Kulikowski. Effective contrast constancy and linearity of contrast sensation. Vision Research, 16(12):1419-1431, 1976. Citation Key: Kulikowski1976a. 2, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.645, + 0.905, + 0.686 + ], + "angle": 0, + "content": "[25] G E Legge and J M Foley. Contrast masking in human vision. Journal of the Optical Society of America, 70(12):1458-71, 1980. Citation Key: Legge1980. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.688, + 0.905, + 0.729 + ], + "angle": 0, + "content": "[26] Qiang Li, Alex Gomez-Villa, Marcelo Bertalmó, and Jesús Malo. Contrast sensitivity functions in autoencoders. Journal of Vision, 22(6):8-8, 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.731, + 0.905, + 0.771 + ], + "angle": 0, + "content": "[27] David G. Lowe. Distinctive image features from scale-invariant keypoints. International Journal of Computer Vision, 60(2):91-110, 2004. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.773, + 0.905, + 0.828 + ], + "angle": 0, + "content": "[28] Lei Lu, Yancheng Cai, Hua Huang, and Ping Wang. An efficient fine-grained vehicle recognition method based on part-level feature optimization. Neurocomputing, 536:40-49, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.83, + 0.905, + 0.871 + ], + "angle": 0, + "content": "[29] Yilong Lv, Yancheng Cai, Yujie He, and Min Li. Drkd: Decoupling response-based distillation for object detection. Pattern Recognition, 161:111275, 2025. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.873, + 0.905, + 0.901 + ], + "angle": 0, + "content": "[30] Rafal K Mantiuk, Param Hanji, Maliha Ashraf, Yuta Asano, and Alexandre Shapiro. Colorvideoovdp: A visual difference" + }, + { + "type": "list", + "bbox": [ + 0.517, + 0.093, + 0.906, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.48, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "20047" + } + ], + [ + { + "type": "ref_text", + "bbox": [ + 0.125, + 0.092, + 0.482, + 0.12 + ], + "angle": 0, + "content": "predictor for image, video and display distortions. arXiv preprint arXiv:2401.11485, 2024.3" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.122, + 0.483, + 0.244 + ], + "angle": 0, + "content": "[31] Maxime Oquab, Timothée Darcet, Theo Moutakanni, Huy V. Vo, Marc Szafraniec, Vasil Khalidov, Pierre Fernandez, Daniel Haziza, Francisco Massa, Alaaeldin El-Nouby, Russell Howes, Po-Yao Huang, Hu Xu, Vasu Sharma, Shang-Wen Li, Wojciech Galuba, Mike Rabbat, Mido Assran, Nicolas Ballas, Gabriel Synnaeve, Ishan Misra, Herve Jegou, Julien Mairal, Patrick Labatut, Armand Joulin, and Piotr Bojanowski. Dinov2: Learning robust visual features without supervision, 2023. 2, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.246, + 0.482, + 0.314 + ], + "angle": 0, + "content": "[32] Eli Peli. Suprathreshold contrast perception across differences in mean luminance: effects of stimulus size, dichoptic presentation, and length of adaptation. Journal of the Optical Society of America A, 12(5):817, 1995. Citation Key: Peli1995. 2, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.316, + 0.482, + 0.384 + ], + "angle": 0, + "content": "[33] Alec Radford, Jong Wook Kim, Chris Hallacy, A. Ramesh, Gabriel Goh, Sandhini Agarwal, Girish Sastry, Amanda Askell, Pamela Mishkin, Jack Clark, Gretchen Krueger, and Ilya Sutskever. Learning transferable visual models from natural language supervision. In ICML, 2021. 1, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.385, + 0.482, + 0.467 + ], + "angle": 0, + "content": "[34] Alec Radford, Jong Wook Kim, Chris Hallacy, Aditya Ramesh, Gabriel Goh, Sandhini Agarwal, Girish Sastry, Amanda Askell, Pamela Mishkin, Jack Clark, et al. Learning transferable visual models from natural language supervision. In International conference on machine learning, pages 8748-8763. PMLR, 2021. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.469, + 0.482, + 0.537 + ], + "angle": 0, + "content": "[35] Nikhila Ravi, Valentin Gabeur, Yuan-Ting Hu, Ronghang Hu, Chaitanya Ryali, Tengyu Ma, Haitham Khedr, Roman Rädle, Chloe Rolland, Laura Gustafson, et al. Sam 2: Segment anything in images and videos. arXiv preprint arXiv:2408.00714, 2024. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.539, + 0.483, + 0.607 + ], + "angle": 0, + "content": "[36] Robin Rombach, Andreas Blattmann, Dominik Lorenz, Patrick Esser, and Björn Ommer. High-resolution image synthesis with latent diffusion models. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 10684-10695, 2022. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.609, + 0.482, + 0.661 + ], + "angle": 0, + "content": "[37] Jyrki Rovamo, Olavi Luntinen, and Risto Nasänen. Modelling the dependence of contrast sensitivity on grating area and spatial frequency. Vision Research, 33(18):2773-2788, 1993. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.665, + 0.482, + 0.718 + ], + "angle": 0, + "content": "[38] Jyrki Rovamo, Juvi Mustonen, and Risto Nasänen. Neural modulation transfer function of the human visual system at various eccentricities. Vision Research, 35(6):767-774, 1995. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.72, + 0.482, + 0.802 + ], + "angle": 0, + "content": "[39] Stephen Tian, Yancheng Cai, Hong-Xing Yu, Sergey Zakharov, Katherine Liu, Adrien Gaidon, Yunzhu Li, and Jia-jun Wu. Multi-object manipulation via object-centric neural scattering functions. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 9021-9031, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.804, + 0.482, + 0.857 + ], + "angle": 0, + "content": "[40] Shikhar Tuli, Ishita Dasgupta, Erin Grant, and Thomas L Griffiths. Are convolutional neural networks or transformers more like human vision? arXiv preprint arXiv:2105.07197, 2021. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.859, + 0.482, + 0.899 + ], + "angle": 0, + "content": "[41] A Van Meeteren and JM Valeton. Effects of pictorial noise interfering with visual detection. JOSA A, 5(3):438-444, 1988. 7" + }, + { + "type": "list", + "bbox": [ + 0.093, + 0.092, + 0.483, + 0.899 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.516, + 0.093, + 0.904, + 0.12 + ], + "angle": 0, + "content": "[42] Andrew B Watson. Efficiency of a model human image code. JOSA A, 4(12):2401-2417, 1987. 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.122, + 0.905, + 0.162 + ], + "angle": 0, + "content": "[43] Andrew B Watson and Joshua A Solomon. Model of visual contrast gain control and pattern masking. JOSA A, 14(9): 2379-2391, 1997. 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.164, + 0.905, + 0.219 + ], + "angle": 0, + "content": "[44] Felix A Wichmann and Robert Geirhos. Are deep neural networks adequate behavioral models of human visual perception? Annual Review of Vision Science, 9(1):501-524, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.221, + 0.905, + 0.262 + ], + "angle": 0, + "content": "[45] Hugh R Wilson and James R Bergen. A four mechanism model for threshold spatial vision. Vision research, 19(1): 19-32, 1979. 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.264, + 0.905, + 0.331 + ], + "angle": 0, + "content": "[46] Fei Yin, Yong Zhang, Baoyuan Wu, Yan Feng, Jingyi Zhang, Yanbo Fan, and Yujiu Yang. Generalizable black-box adversarial attack with meta learning. IEEE transactions on pattern analysis and machine intelligence, 46(3):1804-1818, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.334, + 0.905, + 0.415 + ], + "angle": 0, + "content": "[47] Jingdong Zhang, Jiayuan Fan, Peng Ye, Bo Zhang, Hancheng Ye, Baopu Li, Yancheng Cai, and Tao Chen. Bridgenet: Comprehensive and effective feature interactions via bridge feature for multi-task dense predictions. IEEE Transactions on Pattern Analysis and Machine Intelligence, 2025. 2" + }, + { + "type": "list", + "bbox": [ + 0.516, + 0.093, + 0.905, + 0.415 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.48, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "20048" + } + ] +] \ No newline at end of file diff --git a/2025/Do Computer Vision Foundation Models Learn the Low-level Characteristics of the Human Visual System_/4607e0a5-8aa8-470e-b0e4-8cfed36f43c4_origin.pdf b/2025/Do Computer Vision Foundation Models Learn the Low-level Characteristics of the Human Visual System_/4607e0a5-8aa8-470e-b0e4-8cfed36f43c4_origin.pdf new file mode 100644 index 0000000000000000000000000000000000000000..ed74ca471b60057f06a7dd7449e7fbb469311bc7 --- /dev/null +++ b/2025/Do Computer Vision Foundation Models Learn the Low-level Characteristics of the Human Visual System_/4607e0a5-8aa8-470e-b0e4-8cfed36f43c4_origin.pdf @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:96f9067621f427c1e4b27d3603163e051d6650790cfb5bb32c7446d473a572a1 +size 9207267 diff --git a/2025/Do Computer Vision Foundation Models Learn the Low-level Characteristics of the Human Visual System_/full.md b/2025/Do Computer Vision Foundation Models Learn the Low-level Characteristics of the Human Visual System_/full.md new file mode 100644 index 0000000000000000000000000000000000000000..2d6f9c5487a4fb2bf66c418db4a426491b7d4308 --- /dev/null +++ b/2025/Do Computer Vision Foundation Models Learn the Low-level Characteristics of the Human Visual System_/full.md @@ -0,0 +1,343 @@ +# Do computer vision foundation models learn the low-level characteristics of the human visual system? + +Yancheng Cai, Fei Yin, Dounia Hammou, Rafal Mantiuk; University of Cambridge, UK + +yc613@cam.ac.uk, fy277@cam.ac.uk, dh706@cam.ac.uk, mantiuk@gmail.com + +# Abstract + +Computer vision foundation models, such as DINO or OpenCLIP, are trained in a self-supervised manner on large image datasets. Analogously, substantial evidence suggests that the human visual system (HVS) is influenced by the statistical distribution of colors and patterns in the natural world, characteristics also present in the training data of foundation models. The question we address in this paper is whether foundation models trained on natural images mimic some of the low-level characteristics of the human visual system, such as contrast detection, contrast masking, and contrast constancy. Specifically, we designed a protocol comprising nine test types to evaluate the image encoders of 45 foundation and generative models. Our results indicate that some foundation models (e.g., DINO, DINOv2, and OpenCLIP), share some of the characteristics of human vision, but other models show little resemblance. Foundation models tend to show smaller sensitivity to low contrast and rather irregular responses to contrast across frequencies. The foundation models show the best agreement with human data in terms of contrast masking. Our findings suggest that human vision and computer vision may take both similar and different paths when learning to interpret images of the real world. Overall, while differences remain, foundation models trained on vision tasks start to align with low-level human vision, with DINOv2 showing the closest resemblance. Our code is available on https://github.com/caiyancheng/VFM_HVS_CVPR2025. + +# 1. Introduction + +Computer vision foundation models, such as DINO [9] or OpenCLIP [22, 33], show exceptional ability to generalize to different tasks and are becoming cornerstones of many computer vision methods. They owe their exceptional performance to self-supervised training on very large image datasets. The human visual system also owes much of its capability to being able to perceive the world, over many years from infancy to childhood [6]. A question arises: if the neural network and the visual system are trained by being exposed to a large number of images of the world, will + +![](images/69b5a95868ebd18905940c43725a8138c1fd718a2616f30c77c952d26ffde08e.jpg) +Figure 1. To determine whether image encoders of foundation models exhibit a similar low-level characteristic as human vision, we test them on psychophysical stimuli for which human data is available. We want to test the alignment of contrast encoding between human and computational vision models. + +they share their low-level vision characteristics? If they do, we will know that those low-level characteristics arise naturally and likely reflect the statistics of real-world scenes. If they do not, it means that human low-level vision characteristics are specific to the optical/biological limitations of human vision rather than natural image statistics. Our analysis is meant to shed some light on how the vision, either biological or computational, may develop from observing samples of the world, taking either the same or different routes to accomplish their respective tasks. + +In particular, we are interested in the characteristics that are well understood and measured in human vision science using psychophysical methods: contrast detection [4], contrast masking [25] and contrast constancy [18]. Contrast detection and contrast masking quantify the ability of the visual system to detect small contrast patterns, either on uniform backgrounds (contrast detection) or on backgrounds with patterns (contrast masking). Contrast detection and masking capture the "bottlenecks" of the visual system — the characteristic that may prevent us from detecting patterns that are too dark or too small. Similarly, cameras used + +for computer vision are limited by the MTF of the lens, sensor resolution, photon and sensor noise, and we can expect that computer vision methods may need to deal with similar limitations. + +Contrast constancy is the term used in vision science to describe the invariance of the visual system to spatial frequency [18] and partially luminance [24, 32]. Georgeson and Sullivan [18] showed that the perceived magnitude of the contrast that is well above the detection threshold (supra-threshold) appears to us the same regardless of spatial frequency. This is a very important characteristic as it allows us to see contrast (and therefore objects) the same regardless of the viewing distance; otherwise, the frequencies would change with the viewing distance and hence the contrast appearance. A partial constancy (invariance) is also observed across luminance [24, 32], though there is a significant deviation from constancy at lower luminance levels, once the visual system needs to rely on the rod vision. The invariance is also an important feature of many computer vision methods. For example in SIFT features [27] have been designed to be invariant to the changes in contrast, brightness, scale, and rotation. In our experiments, we used the supra-threshold contrast matching test to assess whether the models exhibit the characteristic of contrast constancy. + +Numerous works on adversarial attacks demonstrated that the classification performance of deep learning models can be greatly degraded by visually inconsequential changes [20]. At the same time, human vision does not suffer from such adversarial vulnerability [44]. This is one of the most salient arguments put forward to state that deep architectures are different from human vision. Here, we propose a different methodology to study this question. We consider the deep neural network to be a black box and compare its responses to well-understood and measured characteristics of the human visual system. In particular, we want to check whether the foundation vision models share the same "bottlenecks" and invariance properties as the visual system. To achieve this, we test foundation models on basic vision stimuli, such as Gabor patches and band-limited noise, and compare the response of those models with the psychophysical data collected from human observers. + +In summary, our contributions are as follows: + +- We developed a protocol to evaluate the similarity between machine vision models and the human visual system. This protocol includes contrast detection, contrast masking, and contrast constancy, subdivided into nine distinct test types that collectively capture the low-level fundamental characteristics of human vision. +- We tested the image encoders of 45 foundation and generative models. The results reveal similarities between certain foundation models (e.g., DINOv2 and OpenCLIP) and human vision, particularly in the contrast masking test. However, differences persist across other tests. + +# 2. Related Work + +Since the advent of deep learning, machine vision models based on foundation models [13, 23, 31] and DNNs [28, 29, 39, 47] have successfully handled numerous advanced visual tasks. However, researchers have observed that machine vision operates differently from human vision. [16] revealed that standard CNNs trained on ImageNet are strongly biased toward texture recognition rather than shape, which contrasts with human visual patterns. [44] provided further evidence that deep neural networks (DNNs) differ significantly from the HVS, demonstrating poor robustness in object classification under 3D viewpoint changes and image distortions, and showing vulnerability to adversarial examples, which are rarely problematic for humans. [5] pointed out that DNNs performing well in benchmark tests share little overlap with biological vision mechanisms and fail to account for many findings in psychological studies of human vision. This highlights a clear distinction between machine and human vision, leading to the rise of interest in domain adaptation [7] and making networks robust to adversarial attacks [46]. + +But there is also evidence that the gap between neural network-based machine vision models and human vision is gradually narrowing. [40] compared vision transformers (ViT) [13] and CNNs, finding that ViT not only achieves superior task accuracy but also exhibits weaker inductive biases, with error patterns more consistent with human errors. [17] discovered that the long-standing robustness gap between humans and CNNs in handling distortions is shrinking. [10, 19] also demonstrated that foundation models like DINO [9] and CLIP [34] can generate more accurate and robust metrics for low-level perceptual similarity. + +Most of the aforementioned studies focus on high-level task performance (e.g., accuracy, consistency, ...), which may not reveal whether computation models suffer from the same bottlenecks and rely on the same invariances as human vision. To that end, [2, 26] have attempted to reveal CSF characteristics within pretrained architectures by training a head with a contrast discrimination classifier. The problem with this approach is that it introduces a bias by relying on a classifier trained to compare contrast. Such studies also make an incorrect assumption that CSF explains both near-threshold and super-threshold vision, while contrast constancy results (see Section 4.3) show that this is not the case. In contrast, we examine networks' low-level characteristics without additional task-specific training, considering both near-threshold and supra-threshold vision. + +# 3. Testing framework + +We first explain the tested models, testing methods, result visualization, and the strategy we used to summarize and quantify our results. + +# 3.1. Tested models and testing methodology + +The objective is to evaluate the responses of machine vision foundation models to stimuli commonly used in human vision research [3, 8] and to compare these responses with psychophysical human data. We tested a representative set of 45 models, encompassing the most influential large vision foundation models, including the variants of DINO [9], DINOv2 [12, 31], OpenCLIP [22, 33], SAM [23], SAM2 [35], and MAE [21], as well as the encoder used for the latent space of generative model Stable Diffusion (SDVAE, [36]). Additionally, we report the responses of ColorVideoVDP [30], which is an image and video quality metric that explicitly models low-level human vision and acts as a reference for a low-level human vision model. All models and their variants are listed in Figure 6. + +To test these models, we need to compare pairs of images and assess the "perceived" difference between them. We adopt a methodology inspired by 2-alternative-forced-choice (2AFC) psychophysical experiments. For example, for a pattern detection task, a pair of images could be a Gabor patch (test) and a uniform field (reference), as shown on the left of Figure 2. As such patterns are calibrated in physical light units in vision science, we generate these patterns as luminance maps, scaled in physical units of $\mathrm{cd} / \mathrm{m}^2$ . These luminance maps are then mapped from the linear space to the sRGB color space using a display model (display peak luminance of $400\mathrm{cd} / \mathrm{m}^2$ ) and fed into the image encoder of the foundation model for feature extraction. Note that the sRGB space is almost universally used to represent training datasets and is expected input for the tested encoders. To ensure that models can operate on small contrast values, we modified them to accept floating-point values (instead of 8-bit integers) as input. This was necessary as the quantization artifacts in 8-bit images are often larger than the detection thresholds of human vision. + +To investigate whether the distances in the feature space reflect the perceptual detection thresholds and invariances, we experimented with a series of distance measures, including $L_{1}$ and $L_{2}$ . We found the cosine similarity expressed as a relative angle $(S_{\mathrm{ac}})$ yielded results most consistent with the psychophysical data. $S_{\mathrm{ac}}$ is defined as: + +$$ +S _ {\mathrm {a c}} = \frac {1}{\pi} \arccos \left(\frac {F _ {\mathrm {T}} \cdot F _ {\mathrm {R}}}{\| F _ {\mathrm {T}} \| \| F _ {\mathrm {R}} \|}\right), \tag {1} +$$ + +where $F_{\mathrm{T}}$ and $F_{\mathrm{R}}$ are the test and reference feature vectors (feature maps reshaped into one dimension), and $\cdot$ denotes the dot product. $S_{\mathrm{ac}} = 0$ indicates two input images are equivalent, while $S_{\mathrm{ac}} = 1$ indicates large differences. + +To compare the encoder responses with psychophysical data, we need to be able to map the image sampling frequency into the spatial frequency on the retina. For that, we select the effective resolution of 60 pixels-per-degree, which is typical for modern monitors. We note, however, + +![](images/7d0f4d354cce31ca077243be794d0ba18e4e88baa1de94f93634af4746385a0a.jpg) +Figure 2. Pipeline for computing $S_{\mathrm{ac}}$ . Generate test and reference images in the linear luminance space (cd/m²), transform them to the sRGB color space using a display model, and input each into the image encoder. Reshape the output features into one-dimensional vectors $F_{\mathrm{T}}$ and $F_{\mathrm{R}}$ , then compute $S_{\mathrm{ac}}$ . + +that the choice of this parameter is arbitrary and the model similarity scores can be shifted by a small multiplier along the spatial frequency axis. The luminance maps are generated at the resolution of $224 \times 224$ pixels, corresponding to the size of $3.7 \times 3.7$ visual degrees. It roughly aligned with the extend of the human foveal vision, and the stimuli span across multiple receptive fields of human vision and multiple patches/tokens of a foundation model. + +For ColorVideoVDP, it can work directly on linear physical units, so conversion to sRGB was unnecessary. We directly use its quality score instead of $S_{\mathrm{ac}}$ . Note that the primary focus of this study is on foundation models, with the ColorVideoVDP metric used solely as a baseline for comparison purposes. + +# 3.2. Model alignment score + +Most of our results will be represented as contour plots of model responses, providing qualitative interpretation. Here, we explain our measure of model alignment, which provides quantitative scores. + +As an example, we take the leftmost contour plot in row (a) of Figure 5. Each point on the contour plot corresponds to $S_{\mathrm{ac}}$ between the test image as shown in Figure 3 and a uniform field of the same mean luminance. The dashed line represents human contrast detection data, predicted with castleCSF [3]. A well-aligned model should show one of the contour lines that follows the dashed castleCSF line. We cannot directly use the $S_{\mathrm{ac}}$ values along the dashed line as the measure of alignment because some models result in $S_{\mathrm{ac}} = 0$ for most points near the detection threshold (detect no difference). Therefore, instead, we rely on the measure of change in $S_{\mathrm{ac}}$ in the neighborhood of the dashed line. + +For a well-aligned model, the perceived differences in the neighborhood of the detection threshold (dashed line) + +![](images/d7c1aed2d0855875c06773d5a02641f0a92c09bb2a00332af52e2c56137b75f5.jpg) +Figure 3. Gabors with different spatial frequencies (x-axis) and contrast (y-axis) used as the test images in the contrast detection tests (Section 4.1). "cpd" denotes cycles per degree. Note that the high-frequency patterns can be rendered with aliasing artifacts on the screen or in print — those were not present in our tests. + +should increase as the contrast increases (the sensitivity decreases). Furthermore, the values should be similar along the dashed line. The measure these two properties, we sample the $S_{\mathrm{ac}}$ values for the points that are shifted in contrast (vertical direction) from the dashed line by a multiplier $m$ , where $0.5 \leq m \leq 2$ (note the logarithmic scale in the contour plot in Figure 5). We collect such data for multiple frequencies (or other dimensions) along the dashed line and calculate the Spearman rank order correlation between the multipliers $m$ and the $S_{\mathrm{ac}}$ values. If the properties mentioned above are preserved, the correlation coefficient value $r_s$ should be close to 1. + +The above strategy is used for all contrast detection and contrast masking experiments. For the contrast matching experiment, we use the root mean squared error (RMSE) between the model and human matching data, expressed as the logarithm of contrast. + +# 4. Experiments + +# 4.1. Contrast detection and CSF + +We begin by testing the foundation models' ability to detect low-contrast (near-threshold) patterns (e.g., Gabor patches, band-limited noise) and compare their performance with the human data. As the reference human data, we rely on the castleCSF [3], which is the recent contrast sensitivity function, modeling contrast detection of both achromatic and chromatic patterns. + +Spatial Frequency Contrast sensitivity of the human eye is typically associated with the variation across the spatial frequency. The visual system exhibits a band-pass characteristic, with a peak sensitivity between 2 and 4 cycles per degree (cpd), depending on the luminance and other parameters of the stimulus. The lower sensitivity of the visual system at lower frequencies is associated with the mechanism of lateral inhibition [4], which helps to reduce the influence of (low-frequency) illumination on the perceived images. Such invariance to illumination is a desirable property if the goal is to recognize objects regardless of illumination conditions. The drop in sensitivity at high frequencies is associated with the limitations of eye optics (achromatic contrast) and cone density (chromatic contrast). Here we want to test whether the computer-vision foundation models pick up similar traits when trained on natural images. + +To test encoder responses across frequencies, we generated a 2D array of image pairs, in which the reference image had uniform luminance, and the test image contained a Gabor patch, as shown in Figure 3 (refer to Supplementary for the visualization of other stimuli). We generate such Gabors for achromatic (see Figure 4-a) and chromatic modulation (see Figure 4-c,d). We also tested band-limited noise (see Figure 4-b). The test patterns had fixed size but varying spatial frequency (x-axis) and contrast (y-axis). Because most of the contrast detection data are plotted as the function of sensitivity, we follow this convention and plot the sensitivity, which is the inverse of the contrast. This corresponds to the reversal of the axis on the logarithmic plots, which we use in our analysis. The other parameters of the stimuli are listed in Table 1. Although we tested multiple variants of each foundation model, here we show only the variant with the highest complexity and best performance on its original high-level tasks. The contour plots for other variants can be found in the Supplementary. + +The contour plots of foundation model responses are shown together with a contrast sensitivity function (castleCSF [3]) in rows (a)-(d) of Figure 5. If the foundation models had the same contrast detection characteristics as the human eye, we would expect the smallest difference (the lowest $S_{\mathrm{ac}}$ ) contour line to follow the black dashed curve of castleCSF. This is not the case for any of the tested foundation models. Some models, and in particular DINOv2 and SD-VAE, show overall band-pass characteristics, in particular for noise and achromatic Gabors. SD-VAE has a drop of sensitivity at lower frequencies much larger than that of the visual system. The responses to chromatic patterns (rows (c) and (d)) tend to be less regular than to achromatic patterns and lack the shift toward lower frequencies, which is observed in the human data. Both OpenCLIP and SAM-2 show a very inconsistent response across the spatial frequencies. + +From the data, we can conclude that foundation models + +![](images/279d8c2806526810191f33e945b779b441e73d8e76443f6ea954cf34bad9ea72.jpg) +Figure 4. Examples of test images for contrast detection and contrast masking. The detailed explanation of the stimuli can be found in the Supplementary. "Ach." denotes achromatic. + +![](images/0b75af6b94cd981d111df8bdb0aece20430b01b3f4d16b50c54a98c7f0b17794.jpg) + +![](images/6162e081b249d2d0ddb2338aca3ebae29673b37973fa4e40c57f1565ba19c8f6.jpg) + +![](images/44a1272e20b81ec965f413aef084573b2057923ddb69b183e52337396ca8dd89.jpg) + +Table 1. Key parameters for all our tests. Note that "Radius" does not apply when the pattern is not a Gabor. + +
TestSpatial Frequency (cpd)Luminance (cd/m2)Radius (degree)Contrast
Spatial Frequency - Gabor Achromatic0.5 - 3210010.001 - 1
Spatial Frequency - Noise Achromatic0.5 - 32100-0.001 - 1
Spatial Frequency - Gabor RG0.5 - 3210010.001 - 0.12
Spatial Frequency - Gabor YV0.5 - 3210010.001 - 0.8
Luminance - Gabor Achromatic20.1 - 20010.001 - 1
Area - Gabor Achromatic81000.1 - 10.001 - 1
Phase-Coherent Masking2 (mask) / 2 (test)320.5 (test)0.005 - 0.5 (mask) / 0.01 - 0.5 (test)
Phase-Incoherent Masking0 - 12 (mask) / 1.2 (test)370.8 (test)0.005 - 0.5 (mask) / 0.01 - 0.5 (test)
Contrast Matching5 (reference) / 0.25 - 25 (test)10-0.005 - 0.629 (reference)
+ +do not follow the sensitivity pattern of the visual system, but some may show a band-pass characteristic that is associated with the CSF. Many models show lower sensitivity at lower frequencies, which may indicate that those models obtained some invariance to (low-frequency) illumination through training. + +Luminance The sensitivity of the human eye increases with the luminance. In dim light, human contrast sensitivity increases proportionally to the square root of retinal illuminance, following the DeVries-Rose law. Conversely, in bright light, sensitivity follows Weber's law, remaining independent of illuminance [38]. In this experiment, we want to test whether the computer vision models lose sensitivity at lower luminance levels. Such a loss could be justified by camera noise, which increases in terms of contrast as the light intensity decreases [1]. + +To produce contour plots, we generated a 2D array of image pairs in a similar manner as for spatial frequency variations in the section above, but instead of varying spatial frequency, we varied luminance, as shown in Figure 4e. As a reminder, we did not pass the absolute luminance values directly to each model but instead converted + +them to display-encoded (gamma-corrected) sRGB space — the colour space used for training datasets (see Figure 2). Such an encoding partially compensates for perceptual non-uniformity of luminance. + +The results, shown in row (e) of Figure 5, indicate a systematic drop in sensitivity with luminance for all tested models. The drop in sensitivity of those models is faster than that observed in the human data, in particular for OpenCLIP. + +We can conclude that the models trained on sufficiently large datasets mimic the sensitivity of the visual system to luminance but with a faster drop in sensitivity. This, however, may be the result of using sRGB representation for training datasets, which were presumably mostly well-exposed and contained relatively little information in darker image regions. + +Area Stimulus area (size) significantly affects sensitivity, as larger stimuli activate more retinal cells. Sensitivity increases with area up to the saturation point, denoted as the critical area [37]. The response to stimuli of different sizes tests the model's ability to pool information across the visual field. + +![](images/45234d434d4aa988f17c60a830b5a74fbdb79067252d67b5afac2032d5dc610d.jpg) +(a) + +![](images/aa0444172e51a34e77bc1ea747cb89ad20f2151899ea7c4297d73a716ba6b9c9.jpg) +DINOv2 + +![](images/35f97ada3b530f1ecc9fa19ba38634f751b2ef0cd45abade2dc88b280e87ad42.jpg) +SAM-2 AM2.1-hiera-L + +![](images/05c08c9f36668c34962c79ad3924b6dcd2ae9dede1178c699935ece80bbd50f5.jpg) +(b) + +![](images/f0cb1d4c9c774f738b125ad3f899ecad3f4889f57709af1be117f40a5634fe6d.jpg) + +![](images/037b2220cb657e05717a1dd80faf4a316d6d2aff01dfbb36e26ebcdd2e3c2002.jpg) + +![](images/db14fd77c4a4bd795050b4ffeb906decae1ce44ee7ceabd4161d30e69c3bd82d.jpg) +(c) + +![](images/b353a55570550f30d4a358d22a6fe58b3866d6ef86a2fe3a4ee34cf14445092e.jpg) + +![](images/1a39a3e80cdc5fc48baedb7a9d7835850965d878301f74b76576106335aa0bdd.jpg) + +![](images/4545645c37ed87883d0a9477d6667ef56b85edb9eff9cdf4d4c6a522e2df8dda.jpg) +(d) + +![](images/bba4fde8146dfd1c12b0bce1a54a227a469fdc2c05dee864b1d790b48f90fe66.jpg) + +![](images/3c35aae92cebc266a771748199b524b06a9ff0ac7f51444348daaadc985a23cf.jpg) + +![](images/37703300e949112c6087982e1b0315d64172a1877168cb6a2a4c0a06a8170c92.jpg) +(e) + +![](images/002123188d33cdb1e003ee8f27f0498ac96a4e7086495e32e5cb4192ab3f2f32.jpg) + +![](images/7a5459f0cf7c7d7313a97a8ea43c4d3c0ef306255d039a57a462b7bf99a94179.jpg) + +![](images/b33a4b848ad6d98e0d52bdb2c57bb3a6216b94841c0e40260bb9fde9ca919727.jpg) +(f) + +![](images/447f600fd7065d9232604d2f5e14bf1bd7af5ace80109840719e3aa6a963a65a.jpg) + +![](images/d96d59950efb8a97beb0ab1bc826040a36149a8bc8aa479dbbf9da03a8a2345f.jpg) + +![](images/aeacc86094ff22a60f9715ac9e9f76b98bca70824d9449847c6867aec5fde746.jpg) +(g) + +![](images/4ec7f0fd6a3b7d17c93d97cf80c14285325febc18f0a406ff82b34d048c8598d.jpg) + +![](images/a847f238fdcd06ede35abc32d5698bcb22f6b34f1ddc2db7e55a4a866c091dc0.jpg) + +![](images/aecfb27b88c1bf415e2700a90c40f93beccf1c4f0bdf09684be59d567fdee2f1.jpg) +(h) + +![](images/c056b7796136da919be77d615f40e33582755009980ab01d75b69670b64e6915.jpg) + +![](images/b7e0d684f7d16c3ae97d9b25b655cb094c1f80dfde3364084b288fa3f745ee83.jpg) + +![](images/fe940144a3a69b32d2e01a1c225f8757b97a07a9d8099fd367d67cd12e574f7c.jpg) +(i) + +![](images/cbe015098fad08a967b53a28abfb523581f75bc7532f833dc1ae3a7cf6ba97fa.jpg) +Figure 5. Selected representative experimental results. Each row represents a test, and each column corresponds to a model, selected as the best-performing in their original tasks. (a)-(f): contour plots of contrast detection $S_{\mathrm{ac}}$ , with the ground truth castleCSF [3]. (g),(h): contour plots of contrast masking $S_{\mathrm{ac}}$ , with the ground truth from [14] and [15], respectively. Different colored solid lines representing different $S_{\mathrm{ac}}$ values, where purple indicates the minimum difference $(S_{\mathrm{ac}} \rightarrow 0)$ . (i): results from the contrast matching experiment, where different colors represent different $C_r$ values. The dashed lines are human results [18], while the solid lines are model-predicted results from Equation 2. Results for all models are in the supplementary material. + +![](images/4517034d72c7d377a847734073a7eff57ad180a2a8358c28272432a81c0c254d.jpg) + +![](images/9ad130f3105f49b0463f3b0c83fbb44d09a606a4dbd8c82e3a026859b1213f4a.jpg) + +![](images/34e24619db9f5fded87b2a2b0ce6a7cebf70f92d241bf323b7bc075481255b99.jpg) + +![](images/076f3df94f1637766b7c76fe78469d52b9fb52a5adc28287e014bd52a4e2cf2c.jpg) + +![](images/0449610e1387d864bc54407184c945c6681b812ef995e4fa98f7b7005bd250e9.jpg) + +![](images/2aeecaaf096c601b2def30e4ff711bfb55134fbb421c70e02200c5a1f67c08f5.jpg) + +![](images/13cb7c5f16ca78175924f509fb0affd67b84d57aef31566c11aae6b2b25e4120.jpg) + +![](images/c5a6941c61b14ad16a3ab5e7671baeabf52f0c038db62caaf5d88795b73552b9.jpg) + +![](images/5659372b04a97eee12e20d16b782037f350d0cd8d14145acdc80a0eb164b0615.jpg) + +![](images/67695cec6010abde9080e0ebaa6522980a233a12218456174173e34b3af2e11b.jpg) + +![](images/22801bb5ab4cb3738845a48b08206e748c2942e74d780ed63a3aebfa4b937b65.jpg) + +![](images/8b0c35ba80a07c673cd99fd146b84c28be0a5c7b22979232e7b078bf4f473640.jpg) + +![](images/51fce1bab9f1fb7adbad49fd3c723544775bf1ff74e336e333cbb353d25cfa70.jpg) + +![](images/de63f8593a28f479c67ac19bb046569cc417b3785668f38589cde384a9b73181.jpg) + +![](images/f1e9ae3061e4df7f7097b4ddcffa3e99fe4f22b84f6f5c67edb7dcc8599b51b9.jpg) + +![](images/0e8fff5f273870bb4024bb2a361ce24d3c8e8b825195e84e09bbb13b61b919e6.jpg) + +![](images/3ff7323cab7fa8c421a94b732a840a54f8d4d3f242fae55d925dfbfbda03c0c6.jpg) + +![](images/52436734eded9e750e3faeefe39eabfa19abdcf4fe9e098725b52d0b41dccb8d.jpg) + +![](images/e8b36897e8390bda2e5140251516841a7eb5e2ce66cfac40546eeb258461d151.jpg) + +Contrast Detection - Spatial Frequency - Gabor Achromatic - Spearman Correlation ↑ + +![](images/8e0cb1e6e27798360c09c41734a5870b6d9976748129354297b42946214f78ff.jpg) + +![](images/891a3eed1cd14aaa5bdd95128e2f74fcb1f72e4e3cd57b2322988241e9193e21.jpg) + +![](images/b9e1c2c84291ba86af3f471c380670045e16bdf4822525724386d274777a6fa3.jpg) + +![](images/8ec001b0dd669591dcbf75ae59313e4fa49677e210652074b6f0f57480730675.jpg) + +![](images/66209f039a2c232d3e3ab4acceff3924732831f74365a0f7d0624669610cff86.jpg) + +![](images/8228213a4605f408a8908fbe2499812e00d42a619265dd37fb17ffbccd5ae3ef.jpg) + +![](images/0fc930f3c32ff5595b7936325fabff5dafb52d3c001d5a7b44f5b1e3a688a0b5.jpg) + +![](images/f8d686748e86a14ea6f0332967ba7796f07c0fb33d013415ad0cded263ff7dc6.jpg) + +![](images/2ebd6b659e2b79d97bb3f4574b33c388a6a8ed35a138717417f2922ef05250b7.jpg) +Figure 6. The quantified similarity error between all 45 models and HVS under 9 different tests. For the Contrast Detection and Contrast Masking tasks, Spearman Correlation was used as the metric, with higher values (closer to 1) indicating greater similarity to human vision. For the Supra-threshold Contrast Matching task, RMSE was used as the metric, with lower values (closer to 0) indicating better similarity. + +In this experiment, we vary the size of the Gaussian envelope limiting stimulus size to observe its effect on model responses. We follow the same procedure as in the two previous sections and generate pairs with a uniform field and a Gabor patch (see Figure 4-f). The parameters of the Gabor patch are listed in Table 1. + +The results, shown in row (f) of Figure 5, indicate that most models show summation across the area (the $S_{\mathrm{ac}}$ values increase with the area) and the rate of the increase varies across the models; DINO and SAM-2 have a smaller in + +crease, SD-VAR has a higher increase of $S_{\mathrm{ac}}$ than the human data, and only DINOv2 roughly matches human performance. OpenCLIP shows no consistent patterns. We can conclude that many models show spatial pooling characteristic that shares the trend observed in human data, though the actual slope of the increase is typically different. + +# 4.2. Contrast masking + +Contrast masking explains the decreased visibility of a signal (test) due to the presence of a supra-threshold background (mask). The masking function defines the relationship between the threshold test contrast required for signal detection and the mask contrast. Put simply, a pattern is more difficult to detect in the presence of another pattern of similar spatial frequency and orientation. A typical masking characteristic of the visual system is shown in rows (g) and (h) of Figure 5. It consists of a relatively shallow segment at low mask contrast (near the detection threshold), with the slope increasing for high mask contrast. The shape of the curve is influenced by the specific properties of the mask and test signals [11, 43]. We will consider the case in which the masker is a sinusoidal grating of the same frequency as the test pattern (phase-coherent masking, row (g)) and when the masker is noise (phase-incoherent masking, row (h)) [11, 14, 15], as shown in Figure 4-g,h. + +First, we consider the fundamental form of phase-coherent masking in which the masker image (reference in Figure 2) is a sinusoidal grating and the test image is the masker plus a Gabor patch of the same frequency and phase as the masker (see Figure 4g). Contrast masking data, shown as the dashed black curve in row (g) of Figure 5, shows the smallest contrast of the test Gabor that is detectable in the presence of the masker of a given contrast. As the contrast of the masker increases, the smallest detectable contrast of the test also needs to increase. However, such an increase starts only for a masker that has the contrast sufficiently high to be detected. If the contrast of the masker is near the detection threshold, we can observe a dipper effect — the contrast detection is facilitated by a masker [14, 42, 45]. Such an effect can be only observed in phase coherent masking. + +As an example of phase-incoherent masking, we will consider a masker with band-limited noise and a test with a Gabor patch — see Figure 4h. The human detection thresholds for such masking patterns are similar to those for phase-coherent masking, except that the dipper effect disappears [41]. + +The differences predicted by DINOv2 and OpenCLIP are surprisingly well-aligned with the human contrast masking data — their responses roughly match the slopes of the human data. The alignment is stronger for the phase-incoherent masking. This is particularly notable for OpenCLIP, which did not show any consistent trends for contrast + +detection. Other models do not show strong alignment with the human data. + +Overall, computational models are better aligned with human data for contrast masking than for contrast detection. One possible explanation is that the signals that induce contrast masking are plentiful in natural images, but contrast detection stimuli, which involve barely noticeable patterns on uniform backgrounds, are rare. Therefore, computational models are more likely to pick up the characteristic that is well represented in the training datasets. + +# 4.3. Supra-threshold contrast matching + +While contrast detection and contrast masking explain the just detectable (near-threshold) contrast, most of the vision tasks, such as detection or recognition, involve well-visible (supra-threshold) contrast. Supra-threshold human vision has been studied in contrast-matching experiments in which the magnitude of one contrast is visually matched to the magnitude of another contrast of a different frequency [18] or luminance [24, 32]. One of the most significant findings of those studies is contrast constancy [18] — the ability of the visual system to match physical contrast across frequencies and luminance levels. The results of the seminal study of Georgeson and Sullivan [18] on matching contrast across frequencies are shown as dashed lines in row (i) of Figure 5. At small contrast, the dashed lines show a band-pass shape that follows the contrast sensitivity function. However, as the contrast is increased, the lines become flat showing little influence of frequency on contrast perception. This is an important property that lets us see objects to have the same appearance regardless of the viewing distance. Such a scale invariance is also important for neural networks that are tasked to detect or recognize objects regardless of their size. + +We followed the experimental setup from [18], where the reference was a $5\mathrm{cpd}$ , $10\mathrm{cd} / \mathrm{m}^2$ sinusoidal grating, presented at eight distinct contrast levels $c_{\mathrm{r}}$ . The test stimulus had the same luminance but a different spatial frequency $\rho_{\mathrm{t}}$ . In [18], observers adjusted the test stimulus contrast $c_{\mathrm{t}}$ until its apparent contrast matched that of the reference (contrast matching). In our experiments, we match contrast encodings of sinusoidal gratings: the $(S_{\mathrm{ac}}()$ , eq. (1)), between a feature vector of a sinusoidal grating, $F(\rho ,c)$ of frequency $\rho$ and contrast $c$ , and a uniform field, $U = F(\rho_{\mathrm{t}},0)$ . We find the test contrast $c_{\mathrm{t}}$ that minimizes the expression: + +$$ +\underset {c _ {\mathrm {t}}} {\operatorname {a r g m i n}} \left(S _ {\mathrm {a c}} \left(F \left(\rho_ {\mathrm {r}}, c _ {\mathrm {r}}\right), U\right) - S _ {\mathrm {a c}} \left(F \left(\rho_ {\mathrm {t}}, c _ {\mathrm {t}}\right), U\right)\right) ^ {2}, \tag {2} +$$ + +where the reference frequency $\rho_{\mathrm{r}} = 5$ cpd. The denominators in the expression are used to normalize contrast across frequencies. We experimented with other contrast encodings, including a direct comparison of feature vectors, but the formula above resulted in the best contrast constancy properties across the models. + +The matching contrast predictions for foundation models are visualized as continuous lines in row (i) of Figure 5. The plots show that only DINOv2 and OpenCLIP roughly follow the dashed contrast constancy lines. Both models show less attenuation (more constancy) at the highest spatial frequencies, which could be advantageous when the model needs to work with small-scale features. Both models show attenuation of low frequencies (below 1 cpd), suggesting worse contrast constancy in that frequency range. Other models, including DINO and SAM-2, suffer from large instability across frequencies, or very heavy attenuation of low frequencies in the case of SD-VAE. To conclude, we can observe only partial contrast constancy for selected models. + +# 4.4. Model alignment scores + +As the analysis of all 45 variants of the models is infeasible in the scope of this paper, we prepared quantitative results according to the method explained in Section 3.2 and summarized them in Figure 6. Those let us make three observations: + +First, ColorVideoVDP, which is a visual metric that models human low-level vision, is better aligned with the human data in almost all contrast detection tasks and in terms of contrast constancy, as expected. However, certain variants of OpenCLIP and DINOv2 can match or surpass the ColorVideoVDP alignment in terms of contrast masking. Second, the alignment scores of different foundation model variants (e.g., OpenCLIP) show significant variations and alignment scores appear unrelated to the complexity of the variants or their performance on higher-level tasks. Finally, DINOv2 variants, which have been trained to solve vision tasks, show the greatest alignment with the human data among all foundation models. + +# 5. Conclusions + +If we believe that the goal of both biological and computational low-level vision is to efficiently encode visual information, we can expect that computational models trained on large natural image datasets will share similarities with human vision. In this work, we find that selected computational models, e.g., variants of DINOv2 and OpenCLIP, show surprisingly high alignment with supra-threshold human contrast masking and contrast matching data, but little alignment with the near-threshold contrast detection. This means that computation models do not have the same "bottlenecks" as human vision, but, through training, they attain invariance and efficient contrast coding that resembles that of the visual system. We hope that our testing protocol with basic psychophysical stimuli will provide a useful tool for examining future computational models of vision. + +# References + +[1] Cecilia Aguerrebere, Julie Delon, Yann Gousseau, and Pablo Muse. Study of the digital camera acquisition process and statistical modeling of the sensor raw data. 2013. Citation Key: Aguerrebere2013. 5 +[2] Arash Akbarinia, Yaniv Morgenstern, and Karl R Gegenfurtner. Contrast sensitivity function in deep networks. Neural Networks, 164:228-244, 2023. 2 +[3] Maliha Ashraf, Rafal K Mantiuk, Alexandre Shapiro, and Sophie Wuerger. castlecsf—a contrast sensitivity function of color, area, spatiotemporal frequency, luminance and eccentricity. Journal of Vision, 24(4):5-5, 2024. 3, 4, 6 +[4] Peter GJ Barten. Contrast sensitivity of the human eye and its effects on image quality. SPIE press, 1999. 1, 4 +[5] Jeffrey S Bowers, Gaurav Malhotra, Marin Dujmovic, Milton Llera Montero, Christian Tsvetkov, Valerio Biscione, Guillermo Puebla, Federico Adolfi, John E Hummel, Rachel F Heaton, et al. Deep problems with neural network models of human vision. Behavioral and Brain Sciences, 46: e385, 2023. 2 +[6] Oliver Braddick and Janette Atkinson. Development of human visual function. Vision Research, 51(13):1588-1609, 2011. 1 +[7] Yancheng Cai, Bo Zhang, Baopu Li, Tao Chen, Hongliang Yan, Jingdong Zhang, and Jiahao Xu. Rethinking cross-domain pedestrian detection: a background-focused distribution alignment framework for instance-free one-stage detectors. IEEE transactions on image processing, 32:4935-4950, 2023. 2 +[8] Yancheng Cai, Ali Bozorgian, Maliha Ashraf, Robert Wanat, and Rafal Mantiuk. elatcsf: A temporal contrast sensitivity function for flicker detection and modeling variable refresh rate flicker. SIGGRAPH Asia 2024 [in print], 2024. 3 +[9] Mathilde Caron, Hugo Touvron, Ishan Misra, Hervé Jégou, Julien Mairal, Piotr Bojanowski, and Armand Joulin. Emerging properties in self-supervised vision transformers. In Proceedings of the IEEE/CVF international conference on computer vision, pages 9650-9660, 2021. 1, 2, 3 +[10] Francesco Croce, Christian Schlarmann, Naman Deep Singh, and Matthias Hein. Adversarily robust clip models induce better (robust) perceptual metrics. In ICML 2024 Workshop on Foundation Models in the Wild, 2024. 2 +[11] Scott J Daly. Visible differences predictor: an algorithm for the assessment of image fidelity. In Human Vision, Visual Processing, and Digital Display III, pages 2-15. SPIE, 1992. 7 +[12] Timothee Darcet, Maxime Oquab, Julien Mairal, and Piotr Bojanowski. Vision transformers need registers, 2023. 3 +[13] Alexey Dosovitskiy. An image is worth 16x16 words: Transformers for image recognition at scale. arXiv preprint arXiv:2010.11929, 2020. 2 +[14] John M Foley. Human luminance pattern-vision mechanisms: masking experiments require a new model. JOSA A, 11(6):1710-1719, 1994. 6, 7 +[15] Karl R Gegenfurtner and Daniel C Kiper. Contrast detection in luminance and chromatic noise. JOSA A, 9(11):1880-1888, 1992. 6, 7 + +[16] Robert Geirhos, Patricia Rubisch, Claudio Michaelis, Matthias Bethge, Felix A Wichmann, and Wieland Brendel. Imagenet-trained cnns are biased towards texture; increasing shape bias improves accuracy and robustness. arXiv preprint arXiv:1811.12231, 2018. 2 +[17] Robert Geirhos, Kantharaju Narayanappa, Benjamin Mitzkus, Tizian Thieringer, Matthias Bethge, Felix A Wichmann, and Wieland Brendel. Partial success in closing the gap between human and machine vision. Advances in Neural Information Processing Systems, 34:23885-23899, 2021. 2 +[18] MA Georgeson and GD Sullivan. Contrast constancy: deblurring in human vision by spatial frequency channels. The Journal of physiology, 252(3):627-656, 1975. 1, 2, 6, 8 +[19] Abhijay Ghildyal, Nabajeet Barman, and Saman Zadtootaghaj. Foundation models boost low-level perceptual similarity metrics. arXiv preprint arXiv:2409.07650, 2024. 2 +[20] Ian J. Goodfellow, Jonathon Shlens, and Christian Szegedy. Explaining and harnessing adversarial examples. 2014. 2 +[21] Kaiming He, Xinlei Chen, Saining Xie, Yanghao Li, Piotr Dólár, and Ross Girshick. Masked autoencoders are scalable vision learners. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 16000-16009, 2022. 3 +[22] Gabriel Ilharco, Mitchell Wortsman, Ross Wightman, Cade Gordon, Nicholas Carlini, Rohan Taori, Achal Dave, Vaishaal Shankar, Hongseok Namkoong, John Miller, Hannaneh Hajishirzi, Ali Farhadi, and Ludwig Schmidt. Openclip, 2021. If you use this software, please cite it as below. 1, 3 +[23] Alexander Kirillov, Eric Mintun, Nikhila Ravi, Hanzi Mao, Chloe Rolland, Laura Gustafson, Tete Xiao, Spencer Whitehead, Alexander C Berg, Wan-Yen Lo, et al. Segment anything. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 4015-4026, 2023. 2, 3 +[24] J.J. Kulikowski. Effective contrast constancy and linearity of contrast sensation. Vision Research, 16(12):1419-1431, 1976. Citation Key: Kulikowski1976a. 2, 8 +[25] G E Legge and J M Foley. Contrast masking in human vision. Journal of the Optical Society of America, 70(12):1458-71, 1980. Citation Key: Legge1980. 1 +[26] Qiang Li, Alex Gomez-Villa, Marcelo Bertalmó, and Jesús Malo. Contrast sensitivity functions in autoencoders. Journal of Vision, 22(6):8-8, 2022. 2 +[27] David G. Lowe. Distinctive image features from scale-invariant keypoints. International Journal of Computer Vision, 60(2):91-110, 2004. 2 +[28] Lei Lu, Yancheng Cai, Hua Huang, and Ping Wang. An efficient fine-grained vehicle recognition method based on part-level feature optimization. Neurocomputing, 536:40-49, 2023. 2 +[29] Yilong Lv, Yancheng Cai, Yujie He, and Min Li. Drkd: Decoupling response-based distillation for object detection. Pattern Recognition, 161:111275, 2025. 2 +[30] Rafal K Mantiuk, Param Hanji, Maliha Ashraf, Yuta Asano, and Alexandre Shapiro. Colorvideoovdp: A visual difference + +predictor for image, video and display distortions. arXiv preprint arXiv:2401.11485, 2024.3 +[31] Maxime Oquab, Timothée Darcet, Theo Moutakanni, Huy V. Vo, Marc Szafraniec, Vasil Khalidov, Pierre Fernandez, Daniel Haziza, Francisco Massa, Alaaeldin El-Nouby, Russell Howes, Po-Yao Huang, Hu Xu, Vasu Sharma, Shang-Wen Li, Wojciech Galuba, Mike Rabbat, Mido Assran, Nicolas Ballas, Gabriel Synnaeve, Ishan Misra, Herve Jegou, Julien Mairal, Patrick Labatut, Armand Joulin, and Piotr Bojanowski. Dinov2: Learning robust visual features without supervision, 2023. 2, 3 +[32] Eli Peli. Suprathreshold contrast perception across differences in mean luminance: effects of stimulus size, dichoptic presentation, and length of adaptation. Journal of the Optical Society of America A, 12(5):817, 1995. Citation Key: Peli1995. 2, 8 +[33] Alec Radford, Jong Wook Kim, Chris Hallacy, A. Ramesh, Gabriel Goh, Sandhini Agarwal, Girish Sastry, Amanda Askell, Pamela Mishkin, Jack Clark, Gretchen Krueger, and Ilya Sutskever. Learning transferable visual models from natural language supervision. In ICML, 2021. 1, 3 +[34] Alec Radford, Jong Wook Kim, Chris Hallacy, Aditya Ramesh, Gabriel Goh, Sandhini Agarwal, Girish Sastry, Amanda Askell, Pamela Mishkin, Jack Clark, et al. Learning transferable visual models from natural language supervision. In International conference on machine learning, pages 8748-8763. PMLR, 2021. 2 +[35] Nikhila Ravi, Valentin Gabeur, Yuan-Ting Hu, Ronghang Hu, Chaitanya Ryali, Tengyu Ma, Haitham Khedr, Roman Rädle, Chloe Rolland, Laura Gustafson, et al. Sam 2: Segment anything in images and videos. arXiv preprint arXiv:2408.00714, 2024. 3 +[36] Robin Rombach, Andreas Blattmann, Dominik Lorenz, Patrick Esser, and Björn Ommer. High-resolution image synthesis with latent diffusion models. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 10684-10695, 2022. 3 +[37] Jyrki Rovamo, Olavi Luntinen, and Risto Nasänen. Modelling the dependence of contrast sensitivity on grating area and spatial frequency. Vision Research, 33(18):2773-2788, 1993. 5 +[38] Jyrki Rovamo, Juvi Mustonen, and Risto Nasänen. Neural modulation transfer function of the human visual system at various eccentricities. Vision Research, 35(6):767-774, 1995. 5 +[39] Stephen Tian, Yancheng Cai, Hong-Xing Yu, Sergey Zakharov, Katherine Liu, Adrien Gaidon, Yunzhu Li, and Jia-jun Wu. Multi-object manipulation via object-centric neural scattering functions. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 9021-9031, 2023. 2 +[40] Shikhar Tuli, Ishita Dasgupta, Erin Grant, and Thomas L Griffiths. Are convolutional neural networks or transformers more like human vision? arXiv preprint arXiv:2105.07197, 2021. 2 +[41] A Van Meeteren and JM Valeton. Effects of pictorial noise interfering with visual detection. JOSA A, 5(3):438-444, 1988. 7 + +[42] Andrew B Watson. Efficiency of a model human image code. JOSA A, 4(12):2401-2417, 1987. 7 +[43] Andrew B Watson and Joshua A Solomon. Model of visual contrast gain control and pattern masking. JOSA A, 14(9): 2379-2391, 1997. 7 +[44] Felix A Wichmann and Robert Geirhos. Are deep neural networks adequate behavioral models of human visual perception? Annual Review of Vision Science, 9(1):501-524, 2023. 2 +[45] Hugh R Wilson and James R Bergen. A four mechanism model for threshold spatial vision. Vision research, 19(1): 19-32, 1979. 7 +[46] Fei Yin, Yong Zhang, Baoyuan Wu, Yan Feng, Jingyi Zhang, Yanbo Fan, and Yujiu Yang. Generalizable black-box adversarial attack with meta learning. IEEE transactions on pattern analysis and machine intelligence, 46(3):1804-1818, 2023. 2 +[47] Jingdong Zhang, Jiayuan Fan, Peng Ye, Bo Zhang, Hancheng Ye, Baopu Li, Yancheng Cai, and Tao Chen. Bridgenet: Comprehensive and effective feature interactions via bridge feature for multi-task dense predictions. IEEE Transactions on Pattern Analysis and Machine Intelligence, 2025. 2 \ No newline at end of file diff --git a/2025/Do Computer Vision Foundation Models Learn the Low-level Characteristics of the Human Visual System_/images.zip b/2025/Do Computer Vision Foundation Models Learn the Low-level Characteristics of the Human Visual System_/images.zip new file mode 100644 index 0000000000000000000000000000000000000000..cc68710fa58c332823cc32aac26a742319c0ccc9 --- /dev/null +++ b/2025/Do Computer Vision Foundation Models Learn the Low-level Characteristics of the Human Visual System_/images.zip @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:1c1c9725280ab3f0b74e23519939d441334bb9a0e2aee33644b93658efd1f1db +size 884304 diff --git a/2025/Do Computer Vision Foundation Models Learn the Low-level Characteristics of the Human Visual System_/layout.json b/2025/Do Computer Vision Foundation Models Learn the Low-level Characteristics of the Human Visual System_/layout.json new file mode 100644 index 0000000000000000000000000000000000000000..667b6b78926a48635d281dd35d3292cc21aab516 --- /dev/null +++ b/2025/Do Computer Vision Foundation Models Learn the Low-level Characteristics of the Human Visual System_/layout.json @@ -0,0 +1,8922 @@ +{ + "pdf_info": [ + { + "para_blocks": [ + { + "bbox": [ + 130, + 83, + 479, + 119 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 83, + 479, + 119 + ], + "spans": [ + { + "bbox": [ + 130, + 83, + 479, + 119 + ], + "type": "text", + "content": "Do computer vision foundation models learn the low-level characteristics of the human visual system?" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 93, + 141, + 517, + 156 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 93, + 141, + 517, + 156 + ], + "spans": [ + { + "bbox": [ + 93, + 141, + 517, + 156 + ], + "type": "text", + "content": "Yancheng Cai, Fei Yin, Dounia Hammou, Rafal Mantiuk; University of Cambridge, UK" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 125, + 157, + 481, + 169 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 125, + 157, + 481, + 169 + ], + "spans": [ + { + "bbox": [ + 125, + 157, + 481, + 169 + ], + "type": "text", + "content": "yc613@cam.ac.uk, fy277@cam.ac.uk, dh706@cam.ac.uk, mantiuk@gmail.com" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 151, + 197, + 200, + 209 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 151, + 197, + 200, + 209 + ], + "spans": [ + { + "bbox": [ + 151, + 197, + 200, + 209 + ], + "type": "text", + "content": "Abstract" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 55, + 223, + 296, + 547 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 223, + 296, + 547 + ], + "spans": [ + { + "bbox": [ + 55, + 223, + 296, + 547 + ], + "type": "text", + "content": "Computer vision foundation models, such as DINO or OpenCLIP, are trained in a self-supervised manner on large image datasets. Analogously, substantial evidence suggests that the human visual system (HVS) is influenced by the statistical distribution of colors and patterns in the natural world, characteristics also present in the training data of foundation models. The question we address in this paper is whether foundation models trained on natural images mimic some of the low-level characteristics of the human visual system, such as contrast detection, contrast masking, and contrast constancy. Specifically, we designed a protocol comprising nine test types to evaluate the image encoders of 45 foundation and generative models. Our results indicate that some foundation models (e.g., DINO, DINOv2, and OpenCLIP), share some of the characteristics of human vision, but other models show little resemblance. Foundation models tend to show smaller sensitivity to low contrast and rather irregular responses to contrast across frequencies. The foundation models show the best agreement with human data in terms of contrast masking. Our findings suggest that human vision and computer vision may take both similar and different paths when learning to interpret images of the real world. Overall, while differences remain, foundation models trained on vision tasks start to align with low-level human vision, with DINOv2 showing the closest resemblance. Our code is available on https://github.com/caiyancheng/VFM_HVS_CVPR2025." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 55, + 572, + 135, + 585 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 572, + 135, + 585 + ], + "spans": [ + { + "bbox": [ + 55, + 572, + 135, + 585 + ], + "type": "text", + "content": "1. Introduction" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 55, + 594, + 296, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 594, + 296, + 713 + ], + "spans": [ + { + "bbox": [ + 55, + 594, + 296, + 713 + ], + "type": "text", + "content": "Computer vision foundation models, such as DINO [9] or OpenCLIP [22, 33], show exceptional ability to generalize to different tasks and are becoming cornerstones of many computer vision methods. They owe their exceptional performance to self-supervised training on very large image datasets. The human visual system also owes much of its capability to being able to perceive the world, over many years from infancy to childhood [6]. A question arises: if the neural network and the visual system are trained by being exposed to a large number of images of the world, will" + } + ] + } + ], + "index": 8 + }, + { + "type": "image", + "bbox": [ + 316, + 186, + 555, + 370 + ], + "blocks": [ + { + "bbox": [ + 316, + 186, + 555, + 370 + ], + "lines": [ + { + "bbox": [ + 316, + 186, + 555, + 370 + ], + "spans": [ + { + "bbox": [ + 316, + 186, + 555, + 370 + ], + "type": "image", + "image_path": "69b5a95868ebd18905940c43725a8138c1fd718a2616f30c77c952d26ffde08e.jpg" + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 313, + 377, + 555, + 433 + ], + "lines": [ + { + "bbox": [ + 313, + 377, + 555, + 433 + ], + "spans": [ + { + "bbox": [ + 313, + 377, + 555, + 433 + ], + "type": "text", + "content": "Figure 1. To determine whether image encoders of foundation models exhibit a similar low-level characteristic as human vision, we test them on psychophysical stimuli for which human data is available. We want to test the alignment of contrast encoding between human and computational vision models." + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "image_caption" + } + ], + "index": 9 + }, + { + "bbox": [ + 313, + 459, + 555, + 579 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 459, + 555, + 579 + ], + "spans": [ + { + "bbox": [ + 313, + 459, + 555, + 579 + ], + "type": "text", + "content": "they share their low-level vision characteristics? If they do, we will know that those low-level characteristics arise naturally and likely reflect the statistics of real-world scenes. If they do not, it means that human low-level vision characteristics are specific to the optical/biological limitations of human vision rather than natural image statistics. Our analysis is meant to shed some light on how the vision, either biological or computational, may develop from observing samples of the world, taking either the same or different routes to accomplish their respective tasks." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 313, + 582, + 556, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 582, + 556, + 715 + ], + "spans": [ + { + "bbox": [ + 313, + 582, + 556, + 715 + ], + "type": "text", + "content": "In particular, we are interested in the characteristics that are well understood and measured in human vision science using psychophysical methods: contrast detection [4], contrast masking [25] and contrast constancy [18]. Contrast detection and contrast masking quantify the ability of the visual system to detect small contrast patterns, either on uniform backgrounds (contrast detection) or on backgrounds with patterns (contrast masking). Contrast detection and masking capture the \"bottlenecks\" of the visual system — the characteristic that may prevent us from detecting patterns that are too dark or too small. Similarly, cameras used" + } + ] + } + ], + "index": 12 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "spans": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "type": "text", + "content": "CVF" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 145, + 0, + 496, + 37 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 145, + 0, + 496, + 37 + ], + "spans": [ + { + "bbox": [ + 145, + 0, + 496, + 37 + ], + "type": "text", + "content": "This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "text", + "content": "20039" + } + ] + } + ], + "index": 13 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 0 + }, + { + "para_blocks": [ + { + "bbox": [ + 55, + 72, + 294, + 119 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 72, + 294, + 119 + ], + "spans": [ + { + "bbox": [ + 55, + 72, + 294, + 119 + ], + "type": "text", + "content": "for computer vision are limited by the MTF of the lens, sensor resolution, photon and sensor noise, and we can expect that computer vision methods may need to deal with similar limitations." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 56, + 121, + 295, + 361 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 121, + 295, + 361 + ], + "spans": [ + { + "bbox": [ + 56, + 121, + 295, + 361 + ], + "type": "text", + "content": "Contrast constancy is the term used in vision science to describe the invariance of the visual system to spatial frequency [18] and partially luminance [24, 32]. Georgeson and Sullivan [18] showed that the perceived magnitude of the contrast that is well above the detection threshold (supra-threshold) appears to us the same regardless of spatial frequency. This is a very important characteristic as it allows us to see contrast (and therefore objects) the same regardless of the viewing distance; otherwise, the frequencies would change with the viewing distance and hence the contrast appearance. A partial constancy (invariance) is also observed across luminance [24, 32], though there is a significant deviation from constancy at lower luminance levels, once the visual system needs to rely on the rod vision. The invariance is also an important feature of many computer vision methods. For example in SIFT features [27] have been designed to be invariant to the changes in contrast, brightness, scale, and rotation. In our experiments, we used the supra-threshold contrast matching test to assess whether the models exhibit the characteristic of contrast constancy." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 55, + 361, + 295, + 564 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 361, + 295, + 564 + ], + "spans": [ + { + "bbox": [ + 55, + 361, + 295, + 564 + ], + "type": "text", + "content": "Numerous works on adversarial attacks demonstrated that the classification performance of deep learning models can be greatly degraded by visually inconsequential changes [20]. At the same time, human vision does not suffer from such adversarial vulnerability [44]. This is one of the most salient arguments put forward to state that deep architectures are different from human vision. Here, we propose a different methodology to study this question. We consider the deep neural network to be a black box and compare its responses to well-understood and measured characteristics of the human visual system. In particular, we want to check whether the foundation vision models share the same \"bottlenecks\" and invariance properties as the visual system. To achieve this, we test foundation models on basic vision stimuli, such as Gabor patches and band-limited noise, and compare the response of those models with the psychophysical data collected from human observers." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 67, + 567, + 251, + 578 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 67, + 567, + 251, + 578 + ], + "spans": [ + { + "bbox": [ + 67, + 567, + 251, + 578 + ], + "type": "text", + "content": "In summary, our contributions are as follows:" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 56, + 580, + 295, + 711 + ], + "type": "list", + "angle": 0, + "index": 6, + "blocks": [ + { + "bbox": [ + 56, + 580, + 295, + 650 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 580, + 295, + 650 + ], + "spans": [ + { + "bbox": [ + 56, + 580, + 295, + 650 + ], + "type": "text", + "content": "- We developed a protocol to evaluate the similarity between machine vision models and the human visual system. This protocol includes contrast detection, contrast masking, and contrast constancy, subdivided into nine distinct test types that collectively capture the low-level fundamental characteristics of human vision." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 56, + 651, + 294, + 711 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 651, + 294, + 711 + ], + "spans": [ + { + "bbox": [ + 56, + 651, + 294, + 711 + ], + "type": "text", + "content": "- We tested the image encoders of 45 foundation and generative models. The results reveal similarities between certain foundation models (e.g., DINOv2 and OpenCLIP) and human vision, particularly in the contrast masking test. However, differences persist across other tests." + } + ] + } + ], + "index": 5 + } + ], + "sub_type": "text" + }, + { + "bbox": [ + 314, + 71, + 400, + 83 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 314, + 71, + 400, + 83 + ], + "spans": [ + { + "bbox": [ + 314, + 71, + 400, + 83 + ], + "type": "text", + "content": "2. Related Work" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 313, + 91, + 553, + 331 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 91, + 553, + 331 + ], + "spans": [ + { + "bbox": [ + 313, + 91, + 553, + 331 + ], + "type": "text", + "content": "Since the advent of deep learning, machine vision models based on foundation models [13, 23, 31] and DNNs [28, 29, 39, 47] have successfully handled numerous advanced visual tasks. However, researchers have observed that machine vision operates differently from human vision. [16] revealed that standard CNNs trained on ImageNet are strongly biased toward texture recognition rather than shape, which contrasts with human visual patterns. [44] provided further evidence that deep neural networks (DNNs) differ significantly from the HVS, demonstrating poor robustness in object classification under 3D viewpoint changes and image distortions, and showing vulnerability to adversarial examples, which are rarely problematic for humans. [5] pointed out that DNNs performing well in benchmark tests share little overlap with biological vision mechanisms and fail to account for many findings in psychological studies of human vision. This highlights a clear distinction between machine and human vision, leading to the rise of interest in domain adaptation [7] and making networks robust to adversarial attacks [46]." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 313, + 332, + 553, + 463 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 332, + 553, + 463 + ], + "spans": [ + { + "bbox": [ + 313, + 332, + 553, + 463 + ], + "type": "text", + "content": "But there is also evidence that the gap between neural network-based machine vision models and human vision is gradually narrowing. [40] compared vision transformers (ViT) [13] and CNNs, finding that ViT not only achieves superior task accuracy but also exhibits weaker inductive biases, with error patterns more consistent with human errors. [17] discovered that the long-standing robustness gap between humans and CNNs in handling distortions is shrinking. [10, 19] also demonstrated that foundation models like DINO [9] and CLIP [34] can generate more accurate and robust metrics for low-level perceptual similarity." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 313, + 464, + 553, + 643 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 464, + 553, + 643 + ], + "spans": [ + { + "bbox": [ + 313, + 464, + 553, + 643 + ], + "type": "text", + "content": "Most of the aforementioned studies focus on high-level task performance (e.g., accuracy, consistency, ...), which may not reveal whether computation models suffer from the same bottlenecks and rely on the same invariances as human vision. To that end, [2, 26] have attempted to reveal CSF characteristics within pretrained architectures by training a head with a contrast discrimination classifier. The problem with this approach is that it introduces a bias by relying on a classifier trained to compare contrast. Such studies also make an incorrect assumption that CSF explains both near-threshold and super-threshold vision, while contrast constancy results (see Section 4.3) show that this is not the case. In contrast, we examine networks' low-level characteristics without additional task-specific training, considering both near-threshold and supra-threshold vision." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 314, + 656, + 425, + 670 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 314, + 656, + 425, + 670 + ], + "spans": [ + { + "bbox": [ + 314, + 656, + 425, + 670 + ], + "type": "text", + "content": "3. Testing framework" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 313, + 677, + 553, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 677, + 553, + 713 + ], + "spans": [ + { + "bbox": [ + 313, + 677, + 553, + 713 + ], + "type": "text", + "content": "We first explain the tested models, testing methods, result visualization, and the strategy we used to summarize and quantify our results." + } + ] + } + ], + "index": 12 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 749, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 749, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 749, + 317, + 757 + ], + "type": "text", + "content": "20040" + } + ] + } + ], + "index": 13 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 1 + }, + { + "para_blocks": [ + { + "bbox": [ + 55, + 72, + 261, + 85 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 72, + 261, + 85 + ], + "spans": [ + { + "bbox": [ + 55, + 72, + 261, + 85 + ], + "type": "text", + "content": "3.1. Tested models and testing methodology" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 55, + 89, + 296, + 256 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 89, + 296, + 256 + ], + "spans": [ + { + "bbox": [ + 55, + 89, + 296, + 256 + ], + "type": "text", + "content": "The objective is to evaluate the responses of machine vision foundation models to stimuli commonly used in human vision research [3, 8] and to compare these responses with psychophysical human data. We tested a representative set of 45 models, encompassing the most influential large vision foundation models, including the variants of DINO [9], DINOv2 [12, 31], OpenCLIP [22, 33], SAM [23], SAM2 [35], and MAE [21], as well as the encoder used for the latent space of generative model Stable Diffusion (SDVAE, [36]). Additionally, we report the responses of ColorVideoVDP [30], which is an image and video quality metric that explicitly models low-level human vision and acts as a reference for a low-level human vision model. All models and their variants are listed in Figure 6." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 55, + 257, + 296, + 495 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 257, + 296, + 495 + ], + "spans": [ + { + "bbox": [ + 55, + 257, + 296, + 495 + ], + "type": "text", + "content": "To test these models, we need to compare pairs of images and assess the \"perceived\" difference between them. We adopt a methodology inspired by 2-alternative-forced-choice (2AFC) psychophysical experiments. For example, for a pattern detection task, a pair of images could be a Gabor patch (test) and a uniform field (reference), as shown on the left of Figure 2. As such patterns are calibrated in physical light units in vision science, we generate these patterns as luminance maps, scaled in physical units of " + }, + { + "bbox": [ + 55, + 257, + 296, + 495 + ], + "type": "inline_equation", + "content": "\\mathrm{cd} / \\mathrm{m}^2" + }, + { + "bbox": [ + 55, + 257, + 296, + 495 + ], + "type": "text", + "content": ". These luminance maps are then mapped from the linear space to the sRGB color space using a display model (display peak luminance of " + }, + { + "bbox": [ + 55, + 257, + 296, + 495 + ], + "type": "inline_equation", + "content": "400\\mathrm{cd} / \\mathrm{m}^2" + }, + { + "bbox": [ + 55, + 257, + 296, + 495 + ], + "type": "text", + "content": ") and fed into the image encoder of the foundation model for feature extraction. Note that the sRGB space is almost universally used to represent training datasets and is expected input for the tested encoders. To ensure that models can operate on small contrast values, we modified them to accept floating-point values (instead of 8-bit integers) as input. This was necessary as the quantization artifacts in 8-bit images are often larger than the detection thresholds of human vision." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 55, + 496, + 296, + 568 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 496, + 296, + 568 + ], + "spans": [ + { + "bbox": [ + 55, + 496, + 296, + 568 + ], + "type": "text", + "content": "To investigate whether the distances in the feature space reflect the perceptual detection thresholds and invariances, we experimented with a series of distance measures, including " + }, + { + "bbox": [ + 55, + 496, + 296, + 568 + ], + "type": "inline_equation", + "content": "L_{1}" + }, + { + "bbox": [ + 55, + 496, + 296, + 568 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 55, + 496, + 296, + 568 + ], + "type": "inline_equation", + "content": "L_{2}" + }, + { + "bbox": [ + 55, + 496, + 296, + 568 + ], + "type": "text", + "content": ". We found the cosine similarity expressed as a relative angle " + }, + { + "bbox": [ + 55, + 496, + 296, + 568 + ], + "type": "inline_equation", + "content": "(S_{\\mathrm{ac}})" + }, + { + "bbox": [ + 55, + 496, + 296, + 568 + ], + "type": "text", + "content": " yielded results most consistent with the psychophysical data. " + }, + { + "bbox": [ + 55, + 496, + 296, + 568 + ], + "type": "inline_equation", + "content": "S_{\\mathrm{ac}}" + }, + { + "bbox": [ + 55, + 496, + 296, + 568 + ], + "type": "text", + "content": " is defined as:" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 108, + 574, + 295, + 601 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 108, + 574, + 295, + 601 + ], + "spans": [ + { + "bbox": [ + 108, + 574, + 295, + 601 + ], + "type": "interline_equation", + "content": "S _ {\\mathrm {a c}} = \\frac {1}{\\pi} \\arccos \\left(\\frac {F _ {\\mathrm {T}} \\cdot F _ {\\mathrm {R}}}{\\| F _ {\\mathrm {T}} \\| \\| F _ {\\mathrm {R}} \\|}\\right), \\tag {1}", + "image_path": "5dab39abb6e037b9632026b16ee8541edea5b75ca706d9b08eacd7a9535b09de.jpg" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 55, + 605, + 295, + 654 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 605, + 295, + 654 + ], + "spans": [ + { + "bbox": [ + 55, + 605, + 295, + 654 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 55, + 605, + 295, + 654 + ], + "type": "inline_equation", + "content": "F_{\\mathrm{T}}" + }, + { + "bbox": [ + 55, + 605, + 295, + 654 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 55, + 605, + 295, + 654 + ], + "type": "inline_equation", + "content": "F_{\\mathrm{R}}" + }, + { + "bbox": [ + 55, + 605, + 295, + 654 + ], + "type": "text", + "content": " are the test and reference feature vectors (feature maps reshaped into one dimension), and " + }, + { + "bbox": [ + 55, + 605, + 295, + 654 + ], + "type": "inline_equation", + "content": "\\cdot" + }, + { + "bbox": [ + 55, + 605, + 295, + 654 + ], + "type": "text", + "content": " denotes the dot product. " + }, + { + "bbox": [ + 55, + 605, + 295, + 654 + ], + "type": "inline_equation", + "content": "S_{\\mathrm{ac}} = 0" + }, + { + "bbox": [ + 55, + 605, + 295, + 654 + ], + "type": "text", + "content": " indicates two input images are equivalent, while " + }, + { + "bbox": [ + 55, + 605, + 295, + 654 + ], + "type": "inline_equation", + "content": "S_{\\mathrm{ac}} = 1" + }, + { + "bbox": [ + 55, + 605, + 295, + 654 + ], + "type": "text", + "content": " indicates large differences." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 55, + 654, + 296, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 654, + 296, + 713 + ], + "spans": [ + { + "bbox": [ + 55, + 654, + 296, + 713 + ], + "type": "text", + "content": "To compare the encoder responses with psychophysical data, we need to be able to map the image sampling frequency into the spatial frequency on the retina. For that, we select the effective resolution of 60 pixels-per-degree, which is typical for modern monitors. We note, however," + } + ] + } + ], + "index": 6 + }, + { + "type": "image", + "bbox": [ + 317, + 70, + 555, + 219 + ], + "blocks": [ + { + "bbox": [ + 317, + 70, + 555, + 219 + ], + "lines": [ + { + "bbox": [ + 317, + 70, + 555, + 219 + ], + "spans": [ + { + "bbox": [ + 317, + 70, + 555, + 219 + ], + "type": "image", + "image_path": "7d0f4d354cce31ca077243be794d0ba18e4e88baa1de94f93634af4746385a0a.jpg" + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 313, + 226, + 555, + 281 + ], + "lines": [ + { + "bbox": [ + 313, + 226, + 555, + 281 + ], + "spans": [ + { + "bbox": [ + 313, + 226, + 555, + 281 + ], + "type": "text", + "content": "Figure 2. Pipeline for computing " + }, + { + "bbox": [ + 313, + 226, + 555, + 281 + ], + "type": "inline_equation", + "content": "S_{\\mathrm{ac}}" + }, + { + "bbox": [ + 313, + 226, + 555, + 281 + ], + "type": "text", + "content": ". Generate test and reference images in the linear luminance space (cd/m²), transform them to the sRGB color space using a display model, and input each into the image encoder. Reshape the output features into one-dimensional vectors " + }, + { + "bbox": [ + 313, + 226, + 555, + 281 + ], + "type": "inline_equation", + "content": "F_{\\mathrm{T}}" + }, + { + "bbox": [ + 313, + 226, + 555, + 281 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 313, + 226, + 555, + 281 + ], + "type": "inline_equation", + "content": "F_{\\mathrm{R}}" + }, + { + "bbox": [ + 313, + 226, + 555, + 281 + ], + "type": "text", + "content": ", then compute " + }, + { + "bbox": [ + 313, + 226, + 555, + 281 + ], + "type": "inline_equation", + "content": "S_{\\mathrm{ac}}" + }, + { + "bbox": [ + 313, + 226, + 555, + 281 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "image_caption" + } + ], + "index": 7 + }, + { + "bbox": [ + 313, + 303, + 554, + 398 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 303, + 554, + 398 + ], + "spans": [ + { + "bbox": [ + 313, + 303, + 554, + 398 + ], + "type": "text", + "content": "that the choice of this parameter is arbitrary and the model similarity scores can be shifted by a small multiplier along the spatial frequency axis. The luminance maps are generated at the resolution of " + }, + { + "bbox": [ + 313, + 303, + 554, + 398 + ], + "type": "inline_equation", + "content": "224 \\times 224" + }, + { + "bbox": [ + 313, + 303, + 554, + 398 + ], + "type": "text", + "content": " pixels, corresponding to the size of " + }, + { + "bbox": [ + 313, + 303, + 554, + 398 + ], + "type": "inline_equation", + "content": "3.7 \\times 3.7" + }, + { + "bbox": [ + 313, + 303, + 554, + 398 + ], + "type": "text", + "content": " visual degrees. It roughly aligned with the extend of the human foveal vision, and the stimuli span across multiple receptive fields of human vision and multiple patches/tokens of a foundation model." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 313, + 399, + 554, + 472 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 399, + 554, + 472 + ], + "spans": [ + { + "bbox": [ + 313, + 399, + 554, + 472 + ], + "type": "text", + "content": "For ColorVideoVDP, it can work directly on linear physical units, so conversion to sRGB was unnecessary. We directly use its quality score instead of " + }, + { + "bbox": [ + 313, + 399, + 554, + 472 + ], + "type": "inline_equation", + "content": "S_{\\mathrm{ac}}" + }, + { + "bbox": [ + 313, + 399, + 554, + 472 + ], + "type": "text", + "content": ". Note that the primary focus of this study is on foundation models, with the ColorVideoVDP metric used solely as a baseline for comparison purposes." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 313, + 479, + 443, + 491 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 479, + 443, + 491 + ], + "spans": [ + { + "bbox": [ + 313, + 479, + 443, + 491 + ], + "type": "text", + "content": "3.2. Model alignment score" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 313, + 497, + 554, + 544 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 497, + 554, + 544 + ], + "spans": [ + { + "bbox": [ + 313, + 497, + 554, + 544 + ], + "type": "text", + "content": "Most of our results will be represented as contour plots of model responses, providing qualitative interpretation. Here, we explain our measure of model alignment, which provides quantitative scores." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 313, + 545, + 554, + 689 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 545, + 554, + 689 + ], + "spans": [ + { + "bbox": [ + 313, + 545, + 554, + 689 + ], + "type": "text", + "content": "As an example, we take the leftmost contour plot in row (a) of Figure 5. Each point on the contour plot corresponds to " + }, + { + "bbox": [ + 313, + 545, + 554, + 689 + ], + "type": "inline_equation", + "content": "S_{\\mathrm{ac}}" + }, + { + "bbox": [ + 313, + 545, + 554, + 689 + ], + "type": "text", + "content": " between the test image as shown in Figure 3 and a uniform field of the same mean luminance. The dashed line represents human contrast detection data, predicted with castleCSF [3]. A well-aligned model should show one of the contour lines that follows the dashed castleCSF line. We cannot directly use the " + }, + { + "bbox": [ + 313, + 545, + 554, + 689 + ], + "type": "inline_equation", + "content": "S_{\\mathrm{ac}}" + }, + { + "bbox": [ + 313, + 545, + 554, + 689 + ], + "type": "text", + "content": " values along the dashed line as the measure of alignment because some models result in " + }, + { + "bbox": [ + 313, + 545, + 554, + 689 + ], + "type": "inline_equation", + "content": "S_{\\mathrm{ac}} = 0" + }, + { + "bbox": [ + 313, + 545, + 554, + 689 + ], + "type": "text", + "content": " for most points near the detection threshold (detect no difference). Therefore, instead, we rely on the measure of change in " + }, + { + "bbox": [ + 313, + 545, + 554, + 689 + ], + "type": "inline_equation", + "content": "S_{\\mathrm{ac}}" + }, + { + "bbox": [ + 313, + 545, + 554, + 689 + ], + "type": "text", + "content": " in the neighborhood of the dashed line." + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 313, + 689, + 554, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 689, + 554, + 713 + ], + "spans": [ + { + "bbox": [ + 313, + 689, + 554, + 713 + ], + "type": "text", + "content": "For a well-aligned model, the perceived differences in the neighborhood of the detection threshold (dashed line)" + } + ] + } + ], + "index": 14 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "text", + "content": "20041" + } + ] + } + ], + "index": 15 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 2 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 59, + 70, + 298, + 288 + ], + "blocks": [ + { + "bbox": [ + 59, + 70, + 298, + 288 + ], + "lines": [ + { + "bbox": [ + 59, + 70, + 298, + 288 + ], + "spans": [ + { + "bbox": [ + 59, + 70, + 298, + 288 + ], + "type": "image", + "image_path": "d7c1aed2d0855875c06773d5a02641f0a92c09bb2a00332af52e2c56137b75f5.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 55, + 295, + 297, + 352 + ], + "lines": [ + { + "bbox": [ + 55, + 295, + 297, + 352 + ], + "spans": [ + { + "bbox": [ + 55, + 295, + 297, + 352 + ], + "type": "text", + "content": "Figure 3. Gabors with different spatial frequencies (x-axis) and contrast (y-axis) used as the test images in the contrast detection tests (Section 4.1). \"cpd\" denotes cycles per degree. Note that the high-frequency patterns can be rendered with aliasing artifacts on the screen or in print — those were not present in our tests." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "bbox": [ + 54, + 373, + 295, + 516 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 54, + 373, + 295, + 516 + ], + "spans": [ + { + "bbox": [ + 54, + 373, + 295, + 516 + ], + "type": "text", + "content": "should increase as the contrast increases (the sensitivity decreases). Furthermore, the values should be similar along the dashed line. The measure these two properties, we sample the " + }, + { + "bbox": [ + 54, + 373, + 295, + 516 + ], + "type": "inline_equation", + "content": "S_{\\mathrm{ac}}" + }, + { + "bbox": [ + 54, + 373, + 295, + 516 + ], + "type": "text", + "content": " values for the points that are shifted in contrast (vertical direction) from the dashed line by a multiplier " + }, + { + "bbox": [ + 54, + 373, + 295, + 516 + ], + "type": "inline_equation", + "content": "m" + }, + { + "bbox": [ + 54, + 373, + 295, + 516 + ], + "type": "text", + "content": ", where " + }, + { + "bbox": [ + 54, + 373, + 295, + 516 + ], + "type": "inline_equation", + "content": "0.5 \\leq m \\leq 2" + }, + { + "bbox": [ + 54, + 373, + 295, + 516 + ], + "type": "text", + "content": " (note the logarithmic scale in the contour plot in Figure 5). We collect such data for multiple frequencies (or other dimensions) along the dashed line and calculate the Spearman rank order correlation between the multipliers " + }, + { + "bbox": [ + 54, + 373, + 295, + 516 + ], + "type": "inline_equation", + "content": "m" + }, + { + "bbox": [ + 54, + 373, + 295, + 516 + ], + "type": "text", + "content": " and the " + }, + { + "bbox": [ + 54, + 373, + 295, + 516 + ], + "type": "inline_equation", + "content": "S_{\\mathrm{ac}}" + }, + { + "bbox": [ + 54, + 373, + 295, + 516 + ], + "type": "text", + "content": " values. If the properties mentioned above are preserved, the correlation coefficient value " + }, + { + "bbox": [ + 54, + 373, + 295, + 516 + ], + "type": "inline_equation", + "content": "r_s" + }, + { + "bbox": [ + 54, + 373, + 295, + 516 + ], + "type": "text", + "content": " should be close to 1." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 55, + 517, + 296, + 578 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 517, + 296, + 578 + ], + "spans": [ + { + "bbox": [ + 55, + 517, + 296, + 578 + ], + "type": "text", + "content": "The above strategy is used for all contrast detection and contrast masking experiments. For the contrast matching experiment, we use the root mean squared error (RMSE) between the model and human matching data, expressed as the logarithm of contrast." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 55, + 590, + 137, + 604 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 590, + 137, + 604 + ], + "spans": [ + { + "bbox": [ + 55, + 590, + 137, + 604 + ], + "type": "text", + "content": "4. Experiments" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 55, + 611, + 209, + 623 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 611, + 209, + 623 + ], + "spans": [ + { + "bbox": [ + 55, + 611, + 209, + 623 + ], + "type": "text", + "content": "4.1. Contrast detection and CSF" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 55, + 629, + 296, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 629, + 296, + 715 + ], + "spans": [ + { + "bbox": [ + 55, + 629, + 296, + 715 + ], + "type": "text", + "content": "We begin by testing the foundation models' ability to detect low-contrast (near-threshold) patterns (e.g., Gabor patches, band-limited noise) and compare their performance with the human data. As the reference human data, we rely on the castleCSF [3], which is the recent contrast sensitivity function, modeling contrast detection of both achromatic and chromatic patterns." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 313, + 72, + 553, + 263 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 72, + 553, + 263 + ], + "spans": [ + { + "bbox": [ + 313, + 72, + 553, + 263 + ], + "type": "text", + "content": "Spatial Frequency Contrast sensitivity of the human eye is typically associated with the variation across the spatial frequency. The visual system exhibits a band-pass characteristic, with a peak sensitivity between 2 and 4 cycles per degree (cpd), depending on the luminance and other parameters of the stimulus. The lower sensitivity of the visual system at lower frequencies is associated with the mechanism of lateral inhibition [4], which helps to reduce the influence of (low-frequency) illumination on the perceived images. Such invariance to illumination is a desirable property if the goal is to recognize objects regardless of illumination conditions. The drop in sensitivity at high frequencies is associated with the limitations of eye optics (achromatic contrast) and cone density (chromatic contrast). Here we want to test whether the computer-vision foundation models pick up similar traits when trained on natural images." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 313, + 266, + 555, + 494 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 266, + 555, + 494 + ], + "spans": [ + { + "bbox": [ + 313, + 266, + 555, + 494 + ], + "type": "text", + "content": "To test encoder responses across frequencies, we generated a 2D array of image pairs, in which the reference image had uniform luminance, and the test image contained a Gabor patch, as shown in Figure 3 (refer to Supplementary for the visualization of other stimuli). We generate such Gabors for achromatic (see Figure 4-a) and chromatic modulation (see Figure 4-c,d). We also tested band-limited noise (see Figure 4-b). The test patterns had fixed size but varying spatial frequency (x-axis) and contrast (y-axis). Because most of the contrast detection data are plotted as the function of sensitivity, we follow this convention and plot the sensitivity, which is the inverse of the contrast. This corresponds to the reversal of the axis on the logarithmic plots, which we use in our analysis. The other parameters of the stimuli are listed in Table 1. Although we tested multiple variants of each foundation model, here we show only the variant with the highest complexity and best performance on its original high-level tasks. The contour plots for other variants can be found in the Supplementary." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 313, + 496, + 556, + 698 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 496, + 556, + 698 + ], + "spans": [ + { + "bbox": [ + 313, + 496, + 556, + 698 + ], + "type": "text", + "content": "The contour plots of foundation model responses are shown together with a contrast sensitivity function (castleCSF [3]) in rows (a)-(d) of Figure 5. If the foundation models had the same contrast detection characteristics as the human eye, we would expect the smallest difference (the lowest " + }, + { + "bbox": [ + 313, + 496, + 556, + 698 + ], + "type": "inline_equation", + "content": "S_{\\mathrm{ac}}" + }, + { + "bbox": [ + 313, + 496, + 556, + 698 + ], + "type": "text", + "content": ") contour line to follow the black dashed curve of castleCSF. This is not the case for any of the tested foundation models. Some models, and in particular DINOv2 and SD-VAE, show overall band-pass characteristics, in particular for noise and achromatic Gabors. SD-VAE has a drop of sensitivity at lower frequencies much larger than that of the visual system. The responses to chromatic patterns (rows (c) and (d)) tend to be less regular than to achromatic patterns and lack the shift toward lower frequencies, which is observed in the human data. Both OpenCLIP and SAM-2 show a very inconsistent response across the spatial frequencies." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 325, + 701, + 555, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 325, + 701, + 555, + 713 + ], + "spans": [ + { + "bbox": [ + 325, + 701, + 555, + 713 + ], + "type": "text", + "content": "From the data, we can conclude that foundation models" + } + ] + } + ], + "index": 10 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "type": "text", + "content": "20042" + } + ] + } + ], + "index": 11 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 3 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 58, + 71, + 179, + 224 + ], + "blocks": [ + { + "bbox": [ + 58, + 71, + 179, + 224 + ], + "lines": [ + { + "bbox": [ + 58, + 71, + 179, + 224 + ], + "spans": [ + { + "bbox": [ + 58, + 71, + 179, + 224 + ], + "type": "image", + "image_path": "279d8c2806526810191f33e945b779b441e73d8e76443f6ea954cf34bad9ea72.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 55, + 234, + 555, + 257 + ], + "lines": [ + { + "bbox": [ + 55, + 234, + 555, + 257 + ], + "spans": [ + { + "bbox": [ + 55, + 234, + 555, + 257 + ], + "type": "text", + "content": "Figure 4. Examples of test images for contrast detection and contrast masking. The detailed explanation of the stimuli can be found in the Supplementary. \"Ach.\" denotes achromatic." + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "type": "image", + "bbox": [ + 182, + 72, + 304, + 224 + ], + "blocks": [ + { + "bbox": [ + 182, + 72, + 304, + 224 + ], + "lines": [ + { + "bbox": [ + 182, + 72, + 304, + 224 + ], + "spans": [ + { + "bbox": [ + 182, + 72, + 304, + 224 + ], + "type": "image", + "image_path": "0b75af6b94cd981d111df8bdb0aece20430b01b3f4d16b50c54a98c7f0b17794.jpg" + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_body" + } + ], + "index": 1 + }, + { + "type": "image", + "bbox": [ + 306, + 72, + 427, + 224 + ], + "blocks": [ + { + "bbox": [ + 306, + 72, + 427, + 224 + ], + "lines": [ + { + "bbox": [ + 306, + 72, + 427, + 224 + ], + "spans": [ + { + "bbox": [ + 306, + 72, + 427, + 224 + ], + "type": "image", + "image_path": "6162e081b249d2d0ddb2338aca3ebae29673b37973fa4e40c57f1565ba19c8f6.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + } + ], + "index": 2 + }, + { + "type": "image", + "bbox": [ + 430, + 72, + 553, + 224 + ], + "blocks": [ + { + "bbox": [ + 430, + 72, + 553, + 224 + ], + "lines": [ + { + "bbox": [ + 430, + 72, + 553, + 224 + ], + "spans": [ + { + "bbox": [ + 430, + 72, + 553, + 224 + ], + "type": "image", + "image_path": "44a1272e20b81ec965f413aef084573b2057923ddb69b183e52337396ca8dd89.jpg" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_body" + } + ], + "index": 3 + }, + { + "type": "table", + "bbox": [ + 58, + 288, + 553, + 402 + ], + "blocks": [ + { + "bbox": [ + 108, + 268, + 501, + 280 + ], + "lines": [ + { + "bbox": [ + 108, + 268, + 501, + 280 + ], + "spans": [ + { + "bbox": [ + 108, + 268, + 501, + 280 + ], + "type": "text", + "content": "Table 1. Key parameters for all our tests. Note that \"Radius\" does not apply when the pattern is not a Gabor." + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 58, + 288, + 553, + 402 + ], + "lines": [ + { + "bbox": [ + 58, + 288, + 553, + 402 + ], + "spans": [ + { + "bbox": [ + 58, + 288, + 553, + 402 + ], + "type": "table", + "html": "
TestSpatial Frequency (cpd)Luminance (cd/m2)Radius (degree)Contrast
Spatial Frequency - Gabor Achromatic0.5 - 3210010.001 - 1
Spatial Frequency - Noise Achromatic0.5 - 32100-0.001 - 1
Spatial Frequency - Gabor RG0.5 - 3210010.001 - 0.12
Spatial Frequency - Gabor YV0.5 - 3210010.001 - 0.8
Luminance - Gabor Achromatic20.1 - 20010.001 - 1
Area - Gabor Achromatic81000.1 - 10.001 - 1
Phase-Coherent Masking2 (mask) / 2 (test)320.5 (test)0.005 - 0.5 (mask) / 0.01 - 0.5 (test)
Phase-Incoherent Masking0 - 12 (mask) / 1.2 (test)370.8 (test)0.005 - 0.5 (mask) / 0.01 - 0.5 (test)
Contrast Matching5 (reference) / 0.25 - 25 (test)10-0.005 - 0.629 (reference)
", + "image_path": "10fcf70ff5d18d0c0e8d591be1416fee9d54905b144cb99cc22aa47e058e1c49.jpg" + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "table_body" + } + ], + "index": 6 + }, + { + "bbox": [ + 54, + 432, + 295, + 504 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 54, + 432, + 295, + 504 + ], + "spans": [ + { + "bbox": [ + 54, + 432, + 295, + 504 + ], + "type": "text", + "content": "do not follow the sensitivity pattern of the visual system, but some may show a band-pass characteristic that is associated with the CSF. Many models show lower sensitivity at lower frequencies, which may indicate that those models obtained some invariance to (low-frequency) illumination through training." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 54, + 521, + 295, + 640 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 54, + 521, + 295, + 640 + ], + "spans": [ + { + "bbox": [ + 54, + 521, + 295, + 640 + ], + "type": "text", + "content": "Luminance The sensitivity of the human eye increases with the luminance. In dim light, human contrast sensitivity increases proportionally to the square root of retinal illuminance, following the DeVries-Rose law. Conversely, in bright light, sensitivity follows Weber's law, remaining independent of illuminance [38]. In this experiment, we want to test whether the computer vision models lose sensitivity at lower luminance levels. Such a loss could be justified by camera noise, which increases in terms of contrast as the light intensity decreases [1]." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 54, + 642, + 296, + 712 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 54, + 642, + 296, + 712 + ], + "spans": [ + { + "bbox": [ + 54, + 642, + 296, + 712 + ], + "type": "text", + "content": "To produce contour plots, we generated a 2D array of image pairs in a similar manner as for spatial frequency variations in the section above, but instead of varying spatial frequency, we varied luminance, as shown in Figure 4e. As a reminder, we did not pass the absolute luminance values directly to each model but instead converted" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 313, + 432, + 555, + 479 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 432, + 555, + 479 + ], + "spans": [ + { + "bbox": [ + 313, + 432, + 555, + 479 + ], + "type": "text", + "content": "them to display-encoded (gamma-corrected) sRGB space — the colour space used for training datasets (see Figure 2). Such an encoding partially compensates for perceptual non-uniformity of luminance." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 313, + 480, + 554, + 539 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 480, + 554, + 539 + ], + "spans": [ + { + "bbox": [ + 313, + 480, + 554, + 539 + ], + "type": "text", + "content": "The results, shown in row (e) of Figure 5, indicate a systematic drop in sensitivity with luminance for all tested models. The drop in sensitivity of those models is faster than that observed in the human data, in particular for OpenCLIP." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 313, + 540, + 554, + 624 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 540, + 554, + 624 + ], + "spans": [ + { + "bbox": [ + 313, + 540, + 554, + 624 + ], + "type": "text", + "content": "We can conclude that the models trained on sufficiently large datasets mimic the sensitivity of the visual system to luminance but with a faster drop in sensitivity. This, however, may be the result of using sRGB representation for training datasets, which were presumably mostly well-exposed and contained relatively little information in darker image regions." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 313, + 642, + 554, + 712 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 642, + 554, + 712 + ], + "spans": [ + { + "bbox": [ + 313, + 642, + 554, + 712 + ], + "type": "text", + "content": "Area Stimulus area (size) significantly affects sensitivity, as larger stimuli activate more retinal cells. Sensitivity increases with area up to the saturation point, denoted as the critical area [37]. The response to stimuli of different sizes tests the model's ability to pool information across the visual field." + } + ] + } + ], + "index": 13 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "text", + "content": "20043" + } + ] + } + ], + "index": 14 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 4 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 71, + 71, + 179, + 144 + ], + "blocks": [ + { + "bbox": [ + 58, + 110, + 69, + 118 + ], + "lines": [ + { + "bbox": [ + 58, + 110, + 69, + 118 + ], + "spans": [ + { + "bbox": [ + 58, + 110, + 69, + 118 + ], + "type": "text", + "content": "(a)" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 71, + 71, + 179, + 144 + ], + "lines": [ + { + "bbox": [ + 71, + 71, + 179, + 144 + ], + "spans": [ + { + "bbox": [ + 71, + 71, + 179, + 144 + ], + "type": "image", + "image_path": "45234d434d4aa988f17c60a830b5a74fbdb79067252d67b5afac2032d5dc610d.jpg" + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_body" + } + ], + "index": 1 + }, + { + "type": "image", + "bbox": [ + 182, + 83, + 271, + 144 + ], + "blocks": [ + { + "bbox": [ + 213, + 71, + 239, + 79 + ], + "lines": [ + { + "bbox": [ + 213, + 71, + 239, + 79 + ], + "spans": [ + { + "bbox": [ + 213, + 71, + 239, + 79 + ], + "type": "text", + "content": "DINOv2" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 182, + 83, + 271, + 144 + ], + "lines": [ + { + "bbox": [ + 182, + 83, + 271, + 144 + ], + "spans": [ + { + "bbox": [ + 182, + 83, + 271, + 144 + ], + "type": "image", + "image_path": "aa0444172e51a34e77bc1ea747cb89ad20f2151899ea7c4297d73a716ba6b9c9.jpg" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_body" + } + ], + "index": 3 + }, + { + "type": "image", + "bbox": [ + 276, + 84, + 364, + 144 + ], + "blocks": [ + { + "bbox": [ + 304, + 72, + 340, + 83 + ], + "lines": [ + { + "bbox": [ + 304, + 72, + 340, + 83 + ], + "spans": [ + { + "bbox": [ + 304, + 72, + 340, + 83 + ], + "type": "text", + "content": "SAM-2 AM2.1-hiera-L" + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 276, + 84, + 364, + 144 + ], + "lines": [ + { + "bbox": [ + 276, + 84, + 364, + 144 + ], + "spans": [ + { + "bbox": [ + 276, + 84, + 364, + 144 + ], + "type": "image", + "image_path": "35f97ada3b530f1ecc9fa19ba38634f751b2ef0cd45abade2dc88b280e87ad42.jpg" + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "image_body" + } + ], + "index": 5 + }, + { + "type": "image", + "bbox": [ + 71, + 144, + 179, + 204 + ], + "blocks": [ + { + "bbox": [ + 58, + 169, + 69, + 178 + ], + "lines": [ + { + "bbox": [ + 58, + 169, + 69, + 178 + ], + "spans": [ + { + "bbox": [ + 58, + 169, + 69, + 178 + ], + "type": "text", + "content": "(b)" + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 71, + 144, + 179, + 204 + ], + "lines": [ + { + "bbox": [ + 71, + 144, + 179, + 204 + ], + "spans": [ + { + "bbox": [ + 71, + 144, + 179, + 204 + ], + "type": "image", + "image_path": "05c08c9f36668c34962c79ad3924b6dcd2ae9dede1178c699935ece80bbd50f5.jpg" + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "image_body" + } + ], + "index": 7 + }, + { + "type": "image", + "bbox": [ + 182, + 144, + 271, + 204 + ], + "blocks": [ + { + "bbox": [ + 182, + 144, + 271, + 204 + ], + "lines": [ + { + "bbox": [ + 182, + 144, + 271, + 204 + ], + "spans": [ + { + "bbox": [ + 182, + 144, + 271, + 204 + ], + "type": "image", + "image_path": "f0cb1d4c9c774f738b125ad3f899ecad3f4889f57709af1be117f40a5634fe6d.jpg" + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "image_body" + } + ], + "index": 8 + }, + { + "type": "image", + "bbox": [ + 275, + 144, + 364, + 204 + ], + "blocks": [ + { + "bbox": [ + 275, + 144, + 364, + 204 + ], + "lines": [ + { + "bbox": [ + 275, + 144, + 364, + 204 + ], + "spans": [ + { + "bbox": [ + 275, + 144, + 364, + 204 + ], + "type": "image", + "image_path": "037b2220cb657e05717a1dd80faf4a316d6d2aff01dfbb36e26ebcdd2e3c2002.jpg" + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "image_body" + } + ], + "index": 9 + }, + { + "type": "image", + "bbox": [ + 71, + 204, + 179, + 262 + ], + "blocks": [ + { + "bbox": [ + 58, + 229, + 68, + 237 + ], + "lines": [ + { + "bbox": [ + 58, + 229, + 68, + 237 + ], + "spans": [ + { + "bbox": [ + 58, + 229, + 68, + 237 + ], + "type": "text", + "content": "(c)" + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 71, + 204, + 179, + 262 + ], + "lines": [ + { + "bbox": [ + 71, + 204, + 179, + 262 + ], + "spans": [ + { + "bbox": [ + 71, + 204, + 179, + 262 + ], + "type": "image", + "image_path": "db14fd77c4a4bd795050b4ffeb906decae1ce44ee7ceabd4161d30e69c3bd82d.jpg" + } + ] + } + ], + "index": 11, + "angle": 0, + "type": "image_body" + } + ], + "index": 11 + }, + { + "type": "image", + "bbox": [ + 182, + 203, + 271, + 263 + ], + "blocks": [ + { + "bbox": [ + 182, + 203, + 271, + 263 + ], + "lines": [ + { + "bbox": [ + 182, + 203, + 271, + 263 + ], + "spans": [ + { + "bbox": [ + 182, + 203, + 271, + 263 + ], + "type": "image", + "image_path": "b353a55570550f30d4a358d22a6fe58b3866d6ef86a2fe3a4ee34cf14445092e.jpg" + } + ] + } + ], + "index": 12, + "angle": 0, + "type": "image_body" + } + ], + "index": 12 + }, + { + "type": "image", + "bbox": [ + 275, + 203, + 364, + 263 + ], + "blocks": [ + { + "bbox": [ + 275, + 203, + 364, + 263 + ], + "lines": [ + { + "bbox": [ + 275, + 203, + 364, + 263 + ], + "spans": [ + { + "bbox": [ + 275, + 203, + 364, + 263 + ], + "type": "image", + "image_path": "1a39a3e80cdc5fc48baedb7a9d7835850965d878301f74b76576106335aa0bdd.jpg" + } + ] + } + ], + "index": 13, + "angle": 0, + "type": "image_body" + } + ], + "index": 13 + }, + { + "type": "image", + "bbox": [ + 71, + 263, + 179, + 322 + ], + "blocks": [ + { + "bbox": [ + 58, + 289, + 68, + 297 + ], + "lines": [ + { + "bbox": [ + 58, + 289, + 68, + 297 + ], + "spans": [ + { + "bbox": [ + 58, + 289, + 68, + 297 + ], + "type": "text", + "content": "(d)" + } + ] + } + ], + "index": 14, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 71, + 263, + 179, + 322 + ], + "lines": [ + { + "bbox": [ + 71, + 263, + 179, + 322 + ], + "spans": [ + { + "bbox": [ + 71, + 263, + 179, + 322 + ], + "type": "image", + "image_path": "4545645c37ed87883d0a9477d6667ef56b85edb9eff9cdf4d4c6a522e2df8dda.jpg" + } + ] + } + ], + "index": 15, + "angle": 0, + "type": "image_body" + } + ], + "index": 15 + }, + { + "type": "image", + "bbox": [ + 182, + 263, + 271, + 323 + ], + "blocks": [ + { + "bbox": [ + 182, + 263, + 271, + 323 + ], + "lines": [ + { + "bbox": [ + 182, + 263, + 271, + 323 + ], + "spans": [ + { + "bbox": [ + 182, + 263, + 271, + 323 + ], + "type": "image", + "image_path": "bba4fde8146dfd1c12b0bce1a54a227a469fdc2c05dee864b1d790b48f90fe66.jpg" + } + ] + } + ], + "index": 16, + "angle": 0, + "type": "image_body" + } + ], + "index": 16 + }, + { + "type": "image", + "bbox": [ + 275, + 263, + 364, + 323 + ], + "blocks": [ + { + "bbox": [ + 275, + 263, + 364, + 323 + ], + "lines": [ + { + "bbox": [ + 275, + 263, + 364, + 323 + ], + "spans": [ + { + "bbox": [ + 275, + 263, + 364, + 323 + ], + "type": "image", + "image_path": "3c35aae92cebc266a771748199b524b06a9ff0ac7f51444348daaadc985a23cf.jpg" + } + ] + } + ], + "index": 17, + "angle": 0, + "type": "image_body" + } + ], + "index": 17 + }, + { + "type": "image", + "bbox": [ + 71, + 323, + 179, + 382 + ], + "blocks": [ + { + "bbox": [ + 58, + 349, + 68, + 356 + ], + "lines": [ + { + "bbox": [ + 58, + 349, + 68, + 356 + ], + "spans": [ + { + "bbox": [ + 58, + 349, + 68, + 356 + ], + "type": "text", + "content": "(e)" + } + ] + } + ], + "index": 18, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 71, + 323, + 179, + 382 + ], + "lines": [ + { + "bbox": [ + 71, + 323, + 179, + 382 + ], + "spans": [ + { + "bbox": [ + 71, + 323, + 179, + 382 + ], + "type": "image", + "image_path": "37703300e949112c6087982e1b0315d64172a1877168cb6a2a4c0a06a8170c92.jpg" + } + ] + } + ], + "index": 19, + "angle": 0, + "type": "image_body" + } + ], + "index": 19 + }, + { + "type": "image", + "bbox": [ + 182, + 323, + 271, + 382 + ], + "blocks": [ + { + "bbox": [ + 182, + 323, + 271, + 382 + ], + "lines": [ + { + "bbox": [ + 182, + 323, + 271, + 382 + ], + "spans": [ + { + "bbox": [ + 182, + 323, + 271, + 382 + ], + "type": "image", + "image_path": "002123188d33cdb1e003ee8f27f0498ac96a4e7086495e32e5cb4192ab3f2f32.jpg" + } + ] + } + ], + "index": 20, + "angle": 0, + "type": "image_body" + } + ], + "index": 20 + }, + { + "type": "image", + "bbox": [ + 275, + 323, + 364, + 382 + ], + "blocks": [ + { + "bbox": [ + 275, + 323, + 364, + 382 + ], + "lines": [ + { + "bbox": [ + 275, + 323, + 364, + 382 + ], + "spans": [ + { + "bbox": [ + 275, + 323, + 364, + 382 + ], + "type": "image", + "image_path": "7a5459f0cf7c7d7313a97a8ea43c4d3c0ef306255d039a57a462b7bf99a94179.jpg" + } + ] + } + ], + "index": 21, + "angle": 0, + "type": "image_body" + } + ], + "index": 21 + }, + { + "type": "image", + "bbox": [ + 71, + 382, + 179, + 435 + ], + "blocks": [ + { + "bbox": [ + 59, + 408, + 68, + 415 + ], + "lines": [ + { + "bbox": [ + 59, + 408, + 68, + 415 + ], + "spans": [ + { + "bbox": [ + 59, + 408, + 68, + 415 + ], + "type": "text", + "content": "(f)" + } + ] + } + ], + "index": 22, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 71, + 382, + 179, + 435 + ], + "lines": [ + { + "bbox": [ + 71, + 382, + 179, + 435 + ], + "spans": [ + { + "bbox": [ + 71, + 382, + 179, + 435 + ], + "type": "image", + "image_path": "b33a4b848ad6d98e0d52bdb2c57bb3a6216b94841c0e40260bb9fde9ca919727.jpg" + } + ] + } + ], + "index": 23, + "angle": 0, + "type": "image_body" + } + ], + "index": 23 + }, + { + "type": "image", + "bbox": [ + 182, + 381, + 271, + 436 + ], + "blocks": [ + { + "bbox": [ + 182, + 381, + 271, + 436 + ], + "lines": [ + { + "bbox": [ + 182, + 381, + 271, + 436 + ], + "spans": [ + { + "bbox": [ + 182, + 381, + 271, + 436 + ], + "type": "image", + "image_path": "447f600fd7065d9232604d2f5e14bf1bd7af5ace80109840719e3aa6a963a65a.jpg" + } + ] + } + ], + "index": 24, + "angle": 0, + "type": "image_body" + } + ], + "index": 24 + }, + { + "type": "image", + "bbox": [ + 275, + 381, + 364, + 436 + ], + "blocks": [ + { + "bbox": [ + 275, + 381, + 364, + 436 + ], + "lines": [ + { + "bbox": [ + 275, + 381, + 364, + 436 + ], + "spans": [ + { + "bbox": [ + 275, + 381, + 364, + 436 + ], + "type": "image", + "image_path": "d96d59950efb8a97beb0ab1bc826040a36149a8bc8aa479dbbf9da03a8a2345f.jpg" + } + ] + } + ], + "index": 25, + "angle": 0, + "type": "image_body" + } + ], + "index": 25 + }, + { + "type": "image", + "bbox": [ + 71, + 437, + 179, + 496 + ], + "blocks": [ + { + "bbox": [ + 58, + 468, + 68, + 475 + ], + "lines": [ + { + "bbox": [ + 58, + 468, + 68, + 475 + ], + "spans": [ + { + "bbox": [ + 58, + 468, + 68, + 475 + ], + "type": "text", + "content": "(g)" + } + ] + } + ], + "index": 26, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 71, + 437, + 179, + 496 + ], + "lines": [ + { + "bbox": [ + 71, + 437, + 179, + 496 + ], + "spans": [ + { + "bbox": [ + 71, + 437, + 179, + 496 + ], + "type": "image", + "image_path": "aeacc86094ff22a60f9715ac9e9f76b98bca70824d9449847c6867aec5fde746.jpg" + } + ] + } + ], + "index": 27, + "angle": 0, + "type": "image_body" + } + ], + "index": 27 + }, + { + "type": "image", + "bbox": [ + 182, + 437, + 271, + 496 + ], + "blocks": [ + { + "bbox": [ + 182, + 437, + 271, + 496 + ], + "lines": [ + { + "bbox": [ + 182, + 437, + 271, + 496 + ], + "spans": [ + { + "bbox": [ + 182, + 437, + 271, + 496 + ], + "type": "image", + "image_path": "4ec7f0fd6a3b7d17c93d97cf80c14285325febc18f0a406ff82b34d048c8598d.jpg" + } + ] + } + ], + "index": 28, + "angle": 0, + "type": "image_body" + } + ], + "index": 28 + }, + { + "type": "image", + "bbox": [ + 275, + 437, + 364, + 496 + ], + "blocks": [ + { + "bbox": [ + 275, + 437, + 364, + 496 + ], + "lines": [ + { + "bbox": [ + 275, + 437, + 364, + 496 + ], + "spans": [ + { + "bbox": [ + 275, + 437, + 364, + 496 + ], + "type": "image", + "image_path": "a847f238fdcd06ede35abc32d5698bcb22f6b34f1ddc2db7e55a4a866c091dc0.jpg" + } + ] + } + ], + "index": 29, + "angle": 0, + "type": "image_body" + } + ], + "index": 29 + }, + { + "type": "image", + "bbox": [ + 71, + 498, + 179, + 556 + ], + "blocks": [ + { + "bbox": [ + 58, + 527, + 68, + 535 + ], + "lines": [ + { + "bbox": [ + 58, + 527, + 68, + 535 + ], + "spans": [ + { + "bbox": [ + 58, + 527, + 68, + 535 + ], + "type": "text", + "content": "(h)" + } + ] + } + ], + "index": 30, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 71, + 498, + 179, + 556 + ], + "lines": [ + { + "bbox": [ + 71, + 498, + 179, + 556 + ], + "spans": [ + { + "bbox": [ + 71, + 498, + 179, + 556 + ], + "type": "image", + "image_path": "aecfb27b88c1bf415e2700a90c40f93beccf1c4f0bdf09684be59d567fdee2f1.jpg" + } + ] + } + ], + "index": 31, + "angle": 0, + "type": "image_body" + } + ], + "index": 31 + }, + { + "type": "image", + "bbox": [ + 182, + 498, + 271, + 556 + ], + "blocks": [ + { + "bbox": [ + 182, + 498, + 271, + 556 + ], + "lines": [ + { + "bbox": [ + 182, + 498, + 271, + 556 + ], + "spans": [ + { + "bbox": [ + 182, + 498, + 271, + 556 + ], + "type": "image", + "image_path": "c056b7796136da919be77d615f40e33582755009980ab01d75b69670b64e6915.jpg" + } + ] + } + ], + "index": 32, + "angle": 0, + "type": "image_body" + } + ], + "index": 32 + }, + { + "type": "image", + "bbox": [ + 275, + 498, + 364, + 556 + ], + "blocks": [ + { + "bbox": [ + 275, + 498, + 364, + 556 + ], + "lines": [ + { + "bbox": [ + 275, + 498, + 364, + 556 + ], + "spans": [ + { + "bbox": [ + 275, + 498, + 364, + 556 + ], + "type": "image", + "image_path": "b7e0d684f7d16c3ae97d9b25b655cb094c1f80dfde3364084b288fa3f745ee83.jpg" + } + ] + } + ], + "index": 33, + "angle": 0, + "type": "image_body" + } + ], + "index": 33 + }, + { + "type": "image", + "bbox": [ + 71, + 558, + 179, + 621 + ], + "blocks": [ + { + "bbox": [ + 59, + 586, + 68, + 594 + ], + "lines": [ + { + "bbox": [ + 59, + 586, + 68, + 594 + ], + "spans": [ + { + "bbox": [ + 59, + 586, + 68, + 594 + ], + "type": "text", + "content": "(i)" + } + ] + } + ], + "index": 34, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 71, + 558, + 179, + 621 + ], + "lines": [ + { + "bbox": [ + 71, + 558, + 179, + 621 + ], + "spans": [ + { + "bbox": [ + 71, + 558, + 179, + 621 + ], + "type": "image", + "image_path": "fe940144a3a69b32d2e01a1c225f8757b97a07a9d8099fd367d67cd12e574f7c.jpg" + } + ] + } + ], + "index": 35, + "angle": 0, + "type": "image_body" + } + ], + "index": 35 + }, + { + "type": "image", + "bbox": [ + 182, + 558, + 271, + 620 + ], + "blocks": [ + { + "bbox": [ + 182, + 558, + 271, + 620 + ], + "lines": [ + { + "bbox": [ + 182, + 558, + 271, + 620 + ], + "spans": [ + { + "bbox": [ + 182, + 558, + 271, + 620 + ], + "type": "image", + "image_path": "cbe015098fad08a967b53a28abfb523581f75bc7532f833dc1ae3a7cf6ba97fa.jpg" + } + ] + } + ], + "index": 36, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 55, + 629, + 555, + 696 + ], + "lines": [ + { + "bbox": [ + 55, + 629, + 555, + 696 + ], + "spans": [ + { + "bbox": [ + 55, + 629, + 555, + 696 + ], + "type": "text", + "content": "Figure 5. Selected representative experimental results. Each row represents a test, and each column corresponds to a model, selected as the best-performing in their original tasks. (a)-(f): contour plots of contrast detection " + }, + { + "bbox": [ + 55, + 629, + 555, + 696 + ], + "type": "inline_equation", + "content": "S_{\\mathrm{ac}}" + }, + { + "bbox": [ + 55, + 629, + 555, + 696 + ], + "type": "text", + "content": ", with the ground truth castleCSF [3]. (g),(h): contour plots of contrast masking " + }, + { + "bbox": [ + 55, + 629, + 555, + 696 + ], + "type": "inline_equation", + "content": "S_{\\mathrm{ac}}" + }, + { + "bbox": [ + 55, + 629, + 555, + 696 + ], + "type": "text", + "content": ", with the ground truth from [14] and [15], respectively. Different colored solid lines representing different " + }, + { + "bbox": [ + 55, + 629, + 555, + 696 + ], + "type": "inline_equation", + "content": "S_{\\mathrm{ac}}" + }, + { + "bbox": [ + 55, + 629, + 555, + 696 + ], + "type": "text", + "content": " values, where purple indicates the minimum difference " + }, + { + "bbox": [ + 55, + 629, + 555, + 696 + ], + "type": "inline_equation", + "content": "(S_{\\mathrm{ac}} \\rightarrow 0)" + }, + { + "bbox": [ + 55, + 629, + 555, + 696 + ], + "type": "text", + "content": ". (i): results from the contrast matching experiment, where different colors represent different " + }, + { + "bbox": [ + 55, + 629, + 555, + 696 + ], + "type": "inline_equation", + "content": "C_r" + }, + { + "bbox": [ + 55, + 629, + 555, + 696 + ], + "type": "text", + "content": " values. The dashed lines are human results [18], while the solid lines are model-predicted results from Equation 2. Results for all models are in the supplementary material." + } + ] + } + ], + "index": 56, + "angle": 0, + "type": "image_caption" + } + ], + "index": 36 + }, + { + "type": "image", + "bbox": [ + 275, + 558, + 364, + 620 + ], + "blocks": [ + { + "bbox": [ + 275, + 558, + 364, + 620 + ], + "lines": [ + { + "bbox": [ + 275, + 558, + 364, + 620 + ], + "spans": [ + { + "bbox": [ + 275, + 558, + 364, + 620 + ], + "type": "image", + "image_path": "4517034d72c7d377a847734073a7eff57ad180a2a8358c28272432a81c0c254d.jpg" + } + ] + } + ], + "index": 37, + "angle": 0, + "type": "image_body" + } + ], + "index": 37 + }, + { + "type": "image", + "bbox": [ + 383, + 71, + 457, + 144 + ], + "blocks": [ + { + "bbox": [ + 383, + 71, + 457, + 144 + ], + "lines": [ + { + "bbox": [ + 383, + 71, + 457, + 144 + ], + "spans": [ + { + "bbox": [ + 383, + 71, + 457, + 144 + ], + "type": "image", + "image_path": "9ad130f3105f49b0463f3b0c83fbb44d09a606a4dbd8c82e3a026859b1213f4a.jpg" + } + ] + } + ], + "index": 38, + "angle": 0, + "type": "image_body" + } + ], + "index": 38 + }, + { + "type": "image", + "bbox": [ + 368, + 144, + 457, + 204 + ], + "blocks": [ + { + "bbox": [ + 368, + 144, + 457, + 204 + ], + "lines": [ + { + "bbox": [ + 368, + 144, + 457, + 204 + ], + "spans": [ + { + "bbox": [ + 368, + 144, + 457, + 204 + ], + "type": "image", + "image_path": "34e24619db9f5fded87b2a2b0ce6a7cebf70f92d241bf323b7bc075481255b99.jpg" + } + ] + } + ], + "index": 39, + "angle": 0, + "type": "image_body" + } + ], + "index": 39 + }, + { + "type": "image", + "bbox": [ + 368, + 203, + 457, + 263 + ], + "blocks": [ + { + "bbox": [ + 368, + 203, + 457, + 263 + ], + "lines": [ + { + "bbox": [ + 368, + 203, + 457, + 263 + ], + "spans": [ + { + "bbox": [ + 368, + 203, + 457, + 263 + ], + "type": "image", + "image_path": "076f3df94f1637766b7c76fe78469d52b9fb52a5adc28287e014bd52a4e2cf2c.jpg" + } + ] + } + ], + "index": 40, + "angle": 0, + "type": "image_body" + } + ], + "index": 40 + }, + { + "type": "image", + "bbox": [ + 368, + 263, + 457, + 323 + ], + "blocks": [ + { + "bbox": [ + 368, + 263, + 457, + 323 + ], + "lines": [ + { + "bbox": [ + 368, + 263, + 457, + 323 + ], + "spans": [ + { + "bbox": [ + 368, + 263, + 457, + 323 + ], + "type": "image", + "image_path": "0449610e1387d864bc54407184c945c6681b812ef995e4fa98f7b7005bd250e9.jpg" + } + ] + } + ], + "index": 41, + "angle": 0, + "type": "image_body" + } + ], + "index": 41 + }, + { + "type": "image", + "bbox": [ + 368, + 323, + 457, + 381 + ], + "blocks": [ + { + "bbox": [ + 368, + 323, + 457, + 381 + ], + "lines": [ + { + "bbox": [ + 368, + 323, + 457, + 381 + ], + "spans": [ + { + "bbox": [ + 368, + 323, + 457, + 381 + ], + "type": "image", + "image_path": "2aeecaaf096c601b2def30e4ff711bfb55134fbb421c70e02200c5a1f67c08f5.jpg" + } + ] + } + ], + "index": 42, + "angle": 0, + "type": "image_body" + } + ], + "index": 42 + }, + { + "type": "image", + "bbox": [ + 368, + 381, + 457, + 435 + ], + "blocks": [ + { + "bbox": [ + 368, + 381, + 457, + 435 + ], + "lines": [ + { + "bbox": [ + 368, + 381, + 457, + 435 + ], + "spans": [ + { + "bbox": [ + 368, + 381, + 457, + 435 + ], + "type": "image", + "image_path": "13cb7c5f16ca78175924f509fb0affd67b84d57aef31566c11aae6b2b25e4120.jpg" + } + ] + } + ], + "index": 43, + "angle": 0, + "type": "image_body" + } + ], + "index": 43 + }, + { + "type": "image", + "bbox": [ + 368, + 435, + 457, + 496 + ], + "blocks": [ + { + "bbox": [ + 368, + 435, + 457, + 496 + ], + "lines": [ + { + "bbox": [ + 368, + 435, + 457, + 496 + ], + "spans": [ + { + "bbox": [ + 368, + 435, + 457, + 496 + ], + "type": "image", + "image_path": "c5a6941c61b14ad16a3ab5e7671baeabf52f0c038db62caaf5d88795b73552b9.jpg" + } + ] + } + ], + "index": 44, + "angle": 0, + "type": "image_body" + } + ], + "index": 44 + }, + { + "type": "image", + "bbox": [ + 368, + 498, + 457, + 556 + ], + "blocks": [ + { + "bbox": [ + 368, + 498, + 457, + 556 + ], + "lines": [ + { + "bbox": [ + 368, + 498, + 457, + 556 + ], + "spans": [ + { + "bbox": [ + 368, + 498, + 457, + 556 + ], + "type": "image", + "image_path": "5659372b04a97eee12e20d16b782037f350d0cd8d14145acdc80a0eb164b0615.jpg" + } + ] + } + ], + "index": 45, + "angle": 0, + "type": "image_body" + } + ], + "index": 45 + }, + { + "type": "image", + "bbox": [ + 368, + 558, + 457, + 620 + ], + "blocks": [ + { + "bbox": [ + 368, + 558, + 457, + 620 + ], + "lines": [ + { + "bbox": [ + 368, + 558, + 457, + 620 + ], + "spans": [ + { + "bbox": [ + 368, + 558, + 457, + 620 + ], + "type": "image", + "image_path": "67695cec6010abde9080e0ebaa6522980a233a12218456174173e34b3af2e11b.jpg" + } + ] + } + ], + "index": 46, + "angle": 0, + "type": "image_body" + } + ], + "index": 46 + }, + { + "type": "image", + "bbox": [ + 462, + 71, + 551, + 144 + ], + "blocks": [ + { + "bbox": [ + 462, + 71, + 551, + 144 + ], + "lines": [ + { + "bbox": [ + 462, + 71, + 551, + 144 + ], + "spans": [ + { + "bbox": [ + 462, + 71, + 551, + 144 + ], + "type": "image", + "image_path": "22801bb5ab4cb3738845a48b08206e748c2942e74d780ed63a3aebfa4b937b65.jpg" + } + ] + } + ], + "index": 47, + "angle": 0, + "type": "image_body" + } + ], + "index": 47 + }, + { + "type": "image", + "bbox": [ + 462, + 144, + 551, + 204 + ], + "blocks": [ + { + "bbox": [ + 462, + 144, + 551, + 204 + ], + "lines": [ + { + "bbox": [ + 462, + 144, + 551, + 204 + ], + "spans": [ + { + "bbox": [ + 462, + 144, + 551, + 204 + ], + "type": "image", + "image_path": "8b0c35ba80a07c673cd99fd146b84c28be0a5c7b22979232e7b078bf4f473640.jpg" + } + ] + } + ], + "index": 48, + "angle": 0, + "type": "image_body" + } + ], + "index": 48 + }, + { + "type": "image", + "bbox": [ + 462, + 203, + 551, + 263 + ], + "blocks": [ + { + "bbox": [ + 462, + 203, + 551, + 263 + ], + "lines": [ + { + "bbox": [ + 462, + 203, + 551, + 263 + ], + "spans": [ + { + "bbox": [ + 462, + 203, + 551, + 263 + ], + "type": "image", + "image_path": "51fce1bab9f1fb7adbad49fd3c723544775bf1ff74e336e333cbb353d25cfa70.jpg" + } + ] + } + ], + "index": 49, + "angle": 0, + "type": "image_body" + } + ], + "index": 49 + }, + { + "type": "image", + "bbox": [ + 462, + 263, + 551, + 323 + ], + "blocks": [ + { + "bbox": [ + 462, + 263, + 551, + 323 + ], + "lines": [ + { + "bbox": [ + 462, + 263, + 551, + 323 + ], + "spans": [ + { + "bbox": [ + 462, + 263, + 551, + 323 + ], + "type": "image", + "image_path": "de63f8593a28f479c67ac19bb046569cc417b3785668f38589cde384a9b73181.jpg" + } + ] + } + ], + "index": 50, + "angle": 0, + "type": "image_body" + } + ], + "index": 50 + }, + { + "type": "image", + "bbox": [ + 462, + 323, + 551, + 377 + ], + "blocks": [ + { + "bbox": [ + 462, + 323, + 551, + 377 + ], + "lines": [ + { + "bbox": [ + 462, + 323, + 551, + 377 + ], + "spans": [ + { + "bbox": [ + 462, + 323, + 551, + 377 + ], + "type": "image", + "image_path": "f1e9ae3061e4df7f7097b4ddcffa3e99fe4f22b84f6f5c67edb7dcc8599b51b9.jpg" + } + ] + } + ], + "index": 51, + "angle": 0, + "type": "image_body" + } + ], + "index": 51 + }, + { + "type": "image", + "bbox": [ + 462, + 377, + 551, + 434 + ], + "blocks": [ + { + "bbox": [ + 462, + 377, + 551, + 434 + ], + "lines": [ + { + "bbox": [ + 462, + 377, + 551, + 434 + ], + "spans": [ + { + "bbox": [ + 462, + 377, + 551, + 434 + ], + "type": "image", + "image_path": "0e8fff5f273870bb4024bb2a361ce24d3c8e8b825195e84e09bbb13b61b919e6.jpg" + } + ] + } + ], + "index": 52, + "angle": 0, + "type": "image_body" + } + ], + "index": 52 + }, + { + "type": "image", + "bbox": [ + 462, + 435, + 551, + 496 + ], + "blocks": [ + { + "bbox": [ + 462, + 435, + 551, + 496 + ], + "lines": [ + { + "bbox": [ + 462, + 435, + 551, + 496 + ], + "spans": [ + { + "bbox": [ + 462, + 435, + 551, + 496 + ], + "type": "image", + "image_path": "3ff7323cab7fa8c421a94b732a840a54f8d4d3f242fae55d925dfbfbda03c0c6.jpg" + } + ] + } + ], + "index": 53, + "angle": 0, + "type": "image_body" + } + ], + "index": 53 + }, + { + "type": "image", + "bbox": [ + 462, + 498, + 551, + 553 + ], + "blocks": [ + { + "bbox": [ + 462, + 498, + 551, + 553 + ], + "lines": [ + { + "bbox": [ + 462, + 498, + 551, + 553 + ], + "spans": [ + { + "bbox": [ + 462, + 498, + 551, + 553 + ], + "type": "image", + "image_path": "52436734eded9e750e3faeefe39eabfa19abdcf4fe9e098725b52d0b41dccb8d.jpg" + } + ] + } + ], + "index": 54, + "angle": 0, + "type": "image_body" + } + ], + "index": 54 + }, + { + "type": "image", + "bbox": [ + 462, + 558, + 551, + 620 + ], + "blocks": [ + { + "bbox": [ + 462, + 558, + 551, + 620 + ], + "lines": [ + { + "bbox": [ + 462, + 558, + 551, + 620 + ], + "spans": [ + { + "bbox": [ + 462, + 558, + 551, + 620 + ], + "type": "image", + "image_path": "e8b36897e8390bda2e5140251516841a7eb5e2ce66cfac40546eeb258461d151.jpg" + } + ] + } + ], + "index": 55, + "angle": 0, + "type": "image_body" + } + ], + "index": 55 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 318, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 318, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 318, + 757 + ], + "type": "text", + "content": "20044" + } + ] + } + ], + "index": 57 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 5 + }, + { + "para_blocks": [ + { + "bbox": [ + 79, + 71, + 286, + 79 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 79, + 71, + 286, + 79 + ], + "spans": [ + { + "bbox": [ + 79, + 71, + 286, + 79 + ], + "type": "text", + "content": "Contrast Detection - Spatial Frequency - Gabor Achromatic - Spearman Correlation ↑" + } + ] + } + ], + "index": 0 + }, + { + "type": "image", + "bbox": [ + 60, + 83, + 293, + 116 + ], + "blocks": [ + { + "bbox": [ + 60, + 83, + 293, + 116 + ], + "lines": [ + { + "bbox": [ + 60, + 83, + 293, + 116 + ], + "spans": [ + { + "bbox": [ + 60, + 83, + 293, + 116 + ], + "type": "image", + "image_path": "8e0cb1e6e27798360c09c41734a5870b6d9976748129354297b42946214f78ff.jpg" + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_body" + } + ], + "index": 1 + }, + { + "type": "image", + "bbox": [ + 60, + 120, + 293, + 152 + ], + "blocks": [ + { + "bbox": [ + 60, + 120, + 293, + 152 + ], + "lines": [ + { + "bbox": [ + 60, + 120, + 293, + 152 + ], + "spans": [ + { + "bbox": [ + 60, + 120, + 293, + 152 + ], + "type": "image", + "image_path": "891a3eed1cd14aaa5bdd95128e2f74fcb1f72e4e3cd57b2322988241e9193e21.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + } + ], + "index": 2 + }, + { + "type": "image", + "bbox": [ + 60, + 156, + 293, + 188 + ], + "blocks": [ + { + "bbox": [ + 60, + 156, + 293, + 188 + ], + "lines": [ + { + "bbox": [ + 60, + 156, + 293, + 188 + ], + "spans": [ + { + "bbox": [ + 60, + 156, + 293, + 188 + ], + "type": "image", + "image_path": "b9e1c2c84291ba86af3f471c380670045e16bdf4822525724386d274777a6fa3.jpg" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_body" + } + ], + "index": 3 + }, + { + "type": "image", + "bbox": [ + 60, + 191, + 293, + 225 + ], + "blocks": [ + { + "bbox": [ + 60, + 191, + 293, + 225 + ], + "lines": [ + { + "bbox": [ + 60, + 191, + 293, + 225 + ], + "spans": [ + { + "bbox": [ + 60, + 191, + 293, + 225 + ], + "type": "image", + "image_path": "8ec001b0dd669591dcbf75ae59313e4fa49677e210652074b6f0f57480730675.jpg" + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_body" + } + ], + "index": 4 + }, + { + "type": "image", + "bbox": [ + 60, + 228, + 293, + 262 + ], + "blocks": [ + { + "bbox": [ + 60, + 228, + 293, + 262 + ], + "lines": [ + { + "bbox": [ + 60, + 228, + 293, + 262 + ], + "spans": [ + { + "bbox": [ + 60, + 228, + 293, + 262 + ], + "type": "image", + "image_path": "66209f039a2c232d3e3ab4acceff3924732831f74365a0f7d0624669610cff86.jpg" + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "image_body" + } + ], + "index": 5 + }, + { + "type": "image", + "bbox": [ + 60, + 264, + 293, + 299 + ], + "blocks": [ + { + "bbox": [ + 60, + 264, + 293, + 299 + ], + "lines": [ + { + "bbox": [ + 60, + 264, + 293, + 299 + ], + "spans": [ + { + "bbox": [ + 60, + 264, + 293, + 299 + ], + "type": "image", + "image_path": "8228213a4605f408a8908fbe2499812e00d42a619265dd37fb17ffbccd5ae3ef.jpg" + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "image_body" + } + ], + "index": 6 + }, + { + "type": "image", + "bbox": [ + 60, + 302, + 293, + 336 + ], + "blocks": [ + { + "bbox": [ + 60, + 302, + 293, + 336 + ], + "lines": [ + { + "bbox": [ + 60, + 302, + 293, + 336 + ], + "spans": [ + { + "bbox": [ + 60, + 302, + 293, + 336 + ], + "type": "image", + "image_path": "0fc930f3c32ff5595b7936325fabff5dafb52d3c001d5a7b44f5b1e3a688a0b5.jpg" + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "image_body" + } + ], + "index": 7 + }, + { + "type": "image", + "bbox": [ + 60, + 338, + 293, + 373 + ], + "blocks": [ + { + "bbox": [ + 60, + 338, + 293, + 373 + ], + "lines": [ + { + "bbox": [ + 60, + 338, + 293, + 373 + ], + "spans": [ + { + "bbox": [ + 60, + 338, + 293, + 373 + ], + "type": "image", + "image_path": "f8d686748e86a14ea6f0332967ba7796f07c0fb33d013415ad0cded263ff7dc6.jpg" + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "image_body" + } + ], + "index": 8 + }, + { + "type": "image", + "bbox": [ + 58, + 375, + 293, + 475 + ], + "blocks": [ + { + "bbox": [ + 58, + 375, + 293, + 475 + ], + "lines": [ + { + "bbox": [ + 58, + 375, + 293, + 475 + ], + "spans": [ + { + "bbox": [ + 58, + 375, + 293, + 475 + ], + "type": "image", + "image_path": "2ebd6b659e2b79d97bb3f4574b33c388a6a8ed35a138717417f2922ef05250b7.jpg" + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 55, + 485, + 295, + 563 + ], + "lines": [ + { + "bbox": [ + 55, + 485, + 295, + 563 + ], + "spans": [ + { + "bbox": [ + 55, + 485, + 295, + 563 + ], + "type": "text", + "content": "Figure 6. The quantified similarity error between all 45 models and HVS under 9 different tests. For the Contrast Detection and Contrast Masking tasks, Spearman Correlation was used as the metric, with higher values (closer to 1) indicating greater similarity to human vision. For the Supra-threshold Contrast Matching task, RMSE was used as the metric, with lower values (closer to 0) indicating better similarity." + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "image_caption" + } + ], + "index": 9 + }, + { + "bbox": [ + 55, + 590, + 295, + 662 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 590, + 295, + 662 + ], + "spans": [ + { + "bbox": [ + 55, + 590, + 295, + 662 + ], + "type": "text", + "content": "In this experiment, we vary the size of the Gaussian envelope limiting stimulus size to observe its effect on model responses. We follow the same procedure as in the two previous sections and generate pairs with a uniform field and a Gabor patch (see Figure 4-f). The parameters of the Gabor patch are listed in Table 1." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 55, + 665, + 296, + 712 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 665, + 296, + 712 + ], + "spans": [ + { + "bbox": [ + 55, + 665, + 296, + 712 + ], + "type": "text", + "content": "The results, shown in row (f) of Figure 5, indicate that most models show summation across the area (the " + }, + { + "bbox": [ + 55, + 665, + 296, + 712 + ], + "type": "inline_equation", + "content": "S_{\\mathrm{ac}}" + }, + { + "bbox": [ + 55, + 665, + 296, + 712 + ], + "type": "text", + "content": " values increase with the area) and the rate of the increase varies across the models; DINO and SAM-2 have a smaller in" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 313, + 72, + 554, + 144 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 72, + 554, + 144 + ], + "spans": [ + { + "bbox": [ + 313, + 72, + 554, + 144 + ], + "type": "text", + "content": "crease, SD-VAR has a higher increase of " + }, + { + "bbox": [ + 313, + 72, + 554, + 144 + ], + "type": "inline_equation", + "content": "S_{\\mathrm{ac}}" + }, + { + "bbox": [ + 313, + 72, + 554, + 144 + ], + "type": "text", + "content": " than the human data, and only DINOv2 roughly matches human performance. OpenCLIP shows no consistent patterns. We can conclude that many models show spatial pooling characteristic that shares the trend observed in human data, though the actual slope of the increase is typically different." + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 313, + 154, + 421, + 167 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 154, + 421, + 167 + ], + "spans": [ + { + "bbox": [ + 313, + 154, + 421, + 167 + ], + "type": "text", + "content": "4.2. Contrast masking" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 313, + 172, + 555, + 376 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 172, + 555, + 376 + ], + "spans": [ + { + "bbox": [ + 313, + 172, + 555, + 376 + ], + "type": "text", + "content": "Contrast masking explains the decreased visibility of a signal (test) due to the presence of a supra-threshold background (mask). The masking function defines the relationship between the threshold test contrast required for signal detection and the mask contrast. Put simply, a pattern is more difficult to detect in the presence of another pattern of similar spatial frequency and orientation. A typical masking characteristic of the visual system is shown in rows (g) and (h) of Figure 5. It consists of a relatively shallow segment at low mask contrast (near the detection threshold), with the slope increasing for high mask contrast. The shape of the curve is influenced by the specific properties of the mask and test signals [11, 43]. We will consider the case in which the masker is a sinusoidal grating of the same frequency as the test pattern (phase-coherent masking, row (g)) and when the masker is noise (phase-incoherent masking, row (h)) [11, 14, 15], as shown in Figure 4-g,h." + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 313, + 377, + 554, + 568 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 377, + 554, + 568 + ], + "spans": [ + { + "bbox": [ + 313, + 377, + 554, + 568 + ], + "type": "text", + "content": "First, we consider the fundamental form of phase-coherent masking in which the masker image (reference in Figure 2) is a sinusoidal grating and the test image is the masker plus a Gabor patch of the same frequency and phase as the masker (see Figure 4g). Contrast masking data, shown as the dashed black curve in row (g) of Figure 5, shows the smallest contrast of the test Gabor that is detectable in the presence of the masker of a given contrast. As the contrast of the masker increases, the smallest detectable contrast of the test also needs to increase. However, such an increase starts only for a masker that has the contrast sufficiently high to be detected. If the contrast of the masker is near the detection threshold, we can observe a dipper effect — the contrast detection is facilitated by a masker [14, 42, 45]. Such an effect can be only observed in phase coherent masking." + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 313, + 569, + 554, + 640 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 569, + 554, + 640 + ], + "spans": [ + { + "bbox": [ + 313, + 569, + 554, + 640 + ], + "type": "text", + "content": "As an example of phase-incoherent masking, we will consider a masker with band-limited noise and a test with a Gabor patch — see Figure 4h. The human detection thresholds for such masking patterns are similar to those for phase-coherent masking, except that the dipper effect disappears [41]." + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 313, + 641, + 556, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 641, + 556, + 713 + ], + "spans": [ + { + "bbox": [ + 313, + 641, + 556, + 713 + ], + "type": "text", + "content": "The differences predicted by DINOv2 and OpenCLIP are surprisingly well-aligned with the human contrast masking data — their responses roughly match the slopes of the human data. The alignment is stronger for the phase-incoherent masking. This is particularly notable for OpenCLIP, which did not show any consistent trends for contrast" + } + ] + } + ], + "index": 18 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "text", + "content": "20045" + } + ] + } + ], + "index": 19 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 6 + }, + { + "para_blocks": [ + { + "bbox": [ + 55, + 72, + 295, + 95 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 72, + 295, + 95 + ], + "spans": [ + { + "bbox": [ + 55, + 72, + 295, + 95 + ], + "type": "text", + "content": "detection. Other models do not show strong alignment with the human data." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 55, + 96, + 296, + 193 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 96, + 296, + 193 + ], + "spans": [ + { + "bbox": [ + 55, + 96, + 296, + 193 + ], + "type": "text", + "content": "Overall, computational models are better aligned with human data for contrast masking than for contrast detection. One possible explanation is that the signals that induce contrast masking are plentiful in natural images, but contrast detection stimuli, which involve barely noticeable patterns on uniform backgrounds, are rare. Therefore, computational models are more likely to pick up the characteristic that is well represented in the training datasets." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 55, + 198, + 244, + 211 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 198, + 244, + 211 + ], + "spans": [ + { + "bbox": [ + 55, + 198, + 244, + 211 + ], + "type": "text", + "content": "4.3. Supra-threshold contrast matching" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 56, + 216, + 296, + 477 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 216, + 296, + 477 + ], + "spans": [ + { + "bbox": [ + 56, + 216, + 296, + 477 + ], + "type": "text", + "content": "While contrast detection and contrast masking explain the just detectable (near-threshold) contrast, most of the vision tasks, such as detection or recognition, involve well-visible (supra-threshold) contrast. Supra-threshold human vision has been studied in contrast-matching experiments in which the magnitude of one contrast is visually matched to the magnitude of another contrast of a different frequency [18] or luminance [24, 32]. One of the most significant findings of those studies is contrast constancy [18] — the ability of the visual system to match physical contrast across frequencies and luminance levels. The results of the seminal study of Georgeson and Sullivan [18] on matching contrast across frequencies are shown as dashed lines in row (i) of Figure 5. At small contrast, the dashed lines show a band-pass shape that follows the contrast sensitivity function. However, as the contrast is increased, the lines become flat showing little influence of frequency on contrast perception. This is an important property that lets us see objects to have the same appearance regardless of the viewing distance. Such a scale invariance is also important for neural networks that are tasked to detect or recognize objects regardless of their size." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 55, + 479, + 296, + 611 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 479, + 296, + 611 + ], + "spans": [ + { + "bbox": [ + 55, + 479, + 296, + 611 + ], + "type": "text", + "content": "We followed the experimental setup from [18], where the reference was a " + }, + { + "bbox": [ + 55, + 479, + 296, + 611 + ], + "type": "inline_equation", + "content": "5\\mathrm{cpd}" + }, + { + "bbox": [ + 55, + 479, + 296, + 611 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 55, + 479, + 296, + 611 + ], + "type": "inline_equation", + "content": "10\\mathrm{cd} / \\mathrm{m}^2" + }, + { + "bbox": [ + 55, + 479, + 296, + 611 + ], + "type": "text", + "content": " sinusoidal grating, presented at eight distinct contrast levels " + }, + { + "bbox": [ + 55, + 479, + 296, + 611 + ], + "type": "inline_equation", + "content": "c_{\\mathrm{r}}" + }, + { + "bbox": [ + 55, + 479, + 296, + 611 + ], + "type": "text", + "content": ". The test stimulus had the same luminance but a different spatial frequency " + }, + { + "bbox": [ + 55, + 479, + 296, + 611 + ], + "type": "inline_equation", + "content": "\\rho_{\\mathrm{t}}" + }, + { + "bbox": [ + 55, + 479, + 296, + 611 + ], + "type": "text", + "content": ". In [18], observers adjusted the test stimulus contrast " + }, + { + "bbox": [ + 55, + 479, + 296, + 611 + ], + "type": "inline_equation", + "content": "c_{\\mathrm{t}}" + }, + { + "bbox": [ + 55, + 479, + 296, + 611 + ], + "type": "text", + "content": " until its apparent contrast matched that of the reference (contrast matching). In our experiments, we match contrast encodings of sinusoidal gratings: the " + }, + { + "bbox": [ + 55, + 479, + 296, + 611 + ], + "type": "inline_equation", + "content": "(S_{\\mathrm{ac}}()" + }, + { + "bbox": [ + 55, + 479, + 296, + 611 + ], + "type": "text", + "content": ", eq. (1)), between a feature vector of a sinusoidal grating, " + }, + { + "bbox": [ + 55, + 479, + 296, + 611 + ], + "type": "inline_equation", + "content": "F(\\rho ,c)" + }, + { + "bbox": [ + 55, + 479, + 296, + 611 + ], + "type": "text", + "content": " of frequency " + }, + { + "bbox": [ + 55, + 479, + 296, + 611 + ], + "type": "inline_equation", + "content": "\\rho" + }, + { + "bbox": [ + 55, + 479, + 296, + 611 + ], + "type": "text", + "content": " and contrast " + }, + { + "bbox": [ + 55, + 479, + 296, + 611 + ], + "type": "inline_equation", + "content": "c" + }, + { + "bbox": [ + 55, + 479, + 296, + 611 + ], + "type": "text", + "content": ", and a uniform field, " + }, + { + "bbox": [ + 55, + 479, + 296, + 611 + ], + "type": "inline_equation", + "content": "U = F(\\rho_{\\mathrm{t}},0)" + }, + { + "bbox": [ + 55, + 479, + 296, + 611 + ], + "type": "text", + "content": ". We find the test contrast " + }, + { + "bbox": [ + 55, + 479, + 296, + 611 + ], + "type": "inline_equation", + "content": "c_{\\mathrm{t}}" + }, + { + "bbox": [ + 55, + 479, + 296, + 611 + ], + "type": "text", + "content": " that minimizes the expression:" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 69, + 615, + 295, + 637 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 69, + 615, + 295, + 637 + ], + "spans": [ + { + "bbox": [ + 69, + 615, + 295, + 637 + ], + "type": "interline_equation", + "content": "\\underset {c _ {\\mathrm {t}}} {\\operatorname {a r g m i n}} \\left(S _ {\\mathrm {a c}} \\left(F \\left(\\rho_ {\\mathrm {r}}, c _ {\\mathrm {r}}\\right), U\\right) - S _ {\\mathrm {a c}} \\left(F \\left(\\rho_ {\\mathrm {t}}, c _ {\\mathrm {t}}\\right), U\\right)\\right) ^ {2}, \\tag {2}", + "image_path": "507625dfab4ba982740f6a3df2f58e6e36042b26f5b7875dfb0a7c8a32b1eefb.jpg" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 55, + 642, + 296, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 642, + 296, + 715 + ], + "spans": [ + { + "bbox": [ + 55, + 642, + 296, + 715 + ], + "type": "text", + "content": "where the reference frequency " + }, + { + "bbox": [ + 55, + 642, + 296, + 715 + ], + "type": "inline_equation", + "content": "\\rho_{\\mathrm{r}} = 5" + }, + { + "bbox": [ + 55, + 642, + 296, + 715 + ], + "type": "text", + "content": " cpd. The denominators in the expression are used to normalize contrast across frequencies. We experimented with other contrast encodings, including a direct comparison of feature vectors, but the formula above resulted in the best contrast constancy properties across the models." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 313, + 72, + 555, + 240 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 72, + 555, + 240 + ], + "spans": [ + { + "bbox": [ + 313, + 72, + 555, + 240 + ], + "type": "text", + "content": "The matching contrast predictions for foundation models are visualized as continuous lines in row (i) of Figure 5. The plots show that only DINOv2 and OpenCLIP roughly follow the dashed contrast constancy lines. Both models show less attenuation (more constancy) at the highest spatial frequencies, which could be advantageous when the model needs to work with small-scale features. Both models show attenuation of low frequencies (below 1 cpd), suggesting worse contrast constancy in that frequency range. Other models, including DINO and SAM-2, suffer from large instability across frequencies, or very heavy attenuation of low frequencies in the case of SD-VAE. To conclude, we can observe only partial contrast constancy for selected models." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 313, + 255, + 448, + 268 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 255, + 448, + 268 + ], + "spans": [ + { + "bbox": [ + 313, + 255, + 448, + 268 + ], + "type": "text", + "content": "4.4. Model alignment scores" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 313, + 275, + 555, + 334 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 275, + 555, + 334 + ], + "spans": [ + { + "bbox": [ + 313, + 275, + 555, + 334 + ], + "type": "text", + "content": "As the analysis of all 45 variants of the models is infeasible in the scope of this paper, we prepared quantitative results according to the method explained in Section 3.2 and summarized them in Figure 6. Those let us make three observations:" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 313, + 337, + 556, + 494 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 337, + 556, + 494 + ], + "spans": [ + { + "bbox": [ + 313, + 337, + 556, + 494 + ], + "type": "text", + "content": "First, ColorVideoVDP, which is a visual metric that models human low-level vision, is better aligned with the human data in almost all contrast detection tasks and in terms of contrast constancy, as expected. However, certain variants of OpenCLIP and DINOv2 can match or surpass the ColorVideoVDP alignment in terms of contrast masking. Second, the alignment scores of different foundation model variants (e.g., OpenCLIP) show significant variations and alignment scores appear unrelated to the complexity of the variants or their performance on higher-level tasks. Finally, DINOv2 variants, which have been trained to solve vision tasks, show the greatest alignment with the human data among all foundation models." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 313, + 511, + 392, + 523 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 511, + 392, + 523 + ], + "spans": [ + { + "bbox": [ + 313, + 511, + 392, + 523 + ], + "type": "text", + "content": "5. Conclusions" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 313, + 534, + 556, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 534, + 556, + 715 + ], + "spans": [ + { + "bbox": [ + 313, + 534, + 556, + 715 + ], + "type": "text", + "content": "If we believe that the goal of both biological and computational low-level vision is to efficiently encode visual information, we can expect that computational models trained on large natural image datasets will share similarities with human vision. In this work, we find that selected computational models, e.g., variants of DINOv2 and OpenCLIP, show surprisingly high alignment with supra-threshold human contrast masking and contrast matching data, but little alignment with the near-threshold contrast detection. This means that computation models do not have the same \"bottlenecks\" as human vision, but, through training, they attain invariance and efficient contrast coding that resembles that of the visual system. We hope that our testing protocol with basic psychophysical stimuli will provide a useful tool for examining future computational models of vision." + } + ] + } + ], + "index": 12 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 749, + 318, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 749, + 318, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 749, + 318, + 757 + ], + "type": "text", + "content": "20046" + } + ] + } + ], + "index": 13 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 7 + }, + { + "para_blocks": [ + { + "bbox": [ + 56, + 71, + 115, + 83 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 71, + 115, + 83 + ], + "spans": [ + { + "bbox": [ + 56, + 71, + 115, + 83 + ], + "type": "text", + "content": "References" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 57, + 91, + 296, + 712 + ], + "type": "list", + "angle": 0, + "index": 16, + "blocks": [ + { + "bbox": [ + 61, + 91, + 296, + 135 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 61, + 91, + 296, + 135 + ], + "spans": [ + { + "bbox": [ + 61, + 91, + 296, + 135 + ], + "type": "text", + "content": "[1] Cecilia Aguerrebere, Julie Delon, Yann Gousseau, and Pablo Muse. Study of the digital camera acquisition process and statistical modeling of the sensor raw data. 2013. Citation Key: Aguerrebere2013. 5" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 61, + 136, + 296, + 167 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 61, + 136, + 296, + 167 + ], + "spans": [ + { + "bbox": [ + 61, + 136, + 296, + 167 + ], + "type": "text", + "content": "[2] Arash Akbarinia, Yaniv Morgenstern, and Karl R Gegenfurtner. Contrast sensitivity function in deep networks. Neural Networks, 164:228-244, 2023. 2" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 62, + 168, + 296, + 213 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 168, + 296, + 213 + ], + "spans": [ + { + "bbox": [ + 62, + 168, + 296, + 213 + ], + "type": "text", + "content": "[3] Maliha Ashraf, Rafal K Mantiuk, Alexandre Shapiro, and Sophie Wuerger. castlecsf—a contrast sensitivity function of color, area, spatiotemporal frequency, luminance and eccentricity. Journal of Vision, 24(4):5-5, 2024. 3, 4, 6" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 62, + 213, + 295, + 236 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 213, + 295, + 236 + ], + "spans": [ + { + "bbox": [ + 62, + 213, + 295, + 236 + ], + "type": "text", + "content": "[4] Peter GJ Barten. Contrast sensitivity of the human eye and its effects on image quality. SPIE press, 1999. 1, 4" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 62, + 236, + 294, + 300 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 236, + 294, + 300 + ], + "spans": [ + { + "bbox": [ + 62, + 236, + 294, + 300 + ], + "type": "text", + "content": "[5] Jeffrey S Bowers, Gaurav Malhotra, Marin Dujmovic, Milton Llera Montero, Christian Tsvetkov, Valerio Biscione, Guillermo Puebla, Federico Adolfi, John E Hummel, Rachel F Heaton, et al. Deep problems with neural network models of human vision. Behavioral and Brain Sciences, 46: e385, 2023. 2" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 62, + 302, + 294, + 334 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 302, + 294, + 334 + ], + "spans": [ + { + "bbox": [ + 62, + 302, + 294, + 334 + ], + "type": "text", + "content": "[6] Oliver Braddick and Janette Atkinson. Development of human visual function. Vision Research, 51(13):1588-1609, 2011. 1" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 62, + 335, + 295, + 400 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 335, + 295, + 400 + ], + "spans": [ + { + "bbox": [ + 62, + 335, + 295, + 400 + ], + "type": "text", + "content": "[7] Yancheng Cai, Bo Zhang, Baopu Li, Tao Chen, Hongliang Yan, Jingdong Zhang, and Jiahao Xu. Rethinking cross-domain pedestrian detection: a background-focused distribution alignment framework for instance-free one-stage detectors. IEEE transactions on image processing, 32:4935-4950, 2023. 2" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 62, + 402, + 294, + 446 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 402, + 294, + 446 + ], + "spans": [ + { + "bbox": [ + 62, + 402, + 294, + 446 + ], + "type": "text", + "content": "[8] Yancheng Cai, Ali Bozorgian, Maliha Ashraf, Robert Wanat, and Rafal Mantiuk. elatcsf: A temporal contrast sensitivity function for flicker detection and modeling variable refresh rate flicker. SIGGRAPH Asia 2024 [in print], 2024. 3" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 62, + 447, + 294, + 501 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 447, + 294, + 501 + ], + "spans": [ + { + "bbox": [ + 62, + 447, + 294, + 501 + ], + "type": "text", + "content": "[9] Mathilde Caron, Hugo Touvron, Ishan Misra, Hervé Jégou, Julien Mairal, Piotr Bojanowski, and Armand Joulin. Emerging properties in self-supervised vision transformers. In Proceedings of the IEEE/CVF international conference on computer vision, pages 9650-9660, 2021. 1, 2, 3" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 57, + 502, + 295, + 545 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 502, + 295, + 545 + ], + "spans": [ + { + "bbox": [ + 57, + 502, + 295, + 545 + ], + "type": "text", + "content": "[10] Francesco Croce, Christian Schlarmann, Naman Deep Singh, and Matthias Hein. Adversarily robust clip models induce better (robust) perceptual metrics. In ICML 2024 Workshop on Foundation Models in the Wild, 2024. 2" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 57, + 546, + 295, + 590 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 546, + 295, + 590 + ], + "spans": [ + { + "bbox": [ + 57, + 546, + 295, + 590 + ], + "type": "text", + "content": "[11] Scott J Daly. Visible differences predictor: an algorithm for the assessment of image fidelity. In Human Vision, Visual Processing, and Digital Display III, pages 2-15. SPIE, 1992. 7" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 57, + 591, + 295, + 613 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 591, + 295, + 613 + ], + "spans": [ + { + "bbox": [ + 57, + 591, + 295, + 613 + ], + "type": "text", + "content": "[12] Timothee Darcet, Maxime Oquab, Julien Mairal, and Piotr Bojanowski. Vision transformers need registers, 2023. 3" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 57, + 613, + 295, + 645 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 613, + 295, + 645 + ], + "spans": [ + { + "bbox": [ + 57, + 613, + 295, + 645 + ], + "type": "text", + "content": "[13] Alexey Dosovitskiy. An image is worth 16x16 words: Transformers for image recognition at scale. arXiv preprint arXiv:2010.11929, 2020. 2" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 57, + 647, + 295, + 679 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 647, + 295, + 679 + ], + "spans": [ + { + "bbox": [ + 57, + 647, + 295, + 679 + ], + "type": "text", + "content": "[14] John M Foley. Human luminance pattern-vision mechanisms: masking experiments require a new model. JOSA A, 11(6):1710-1719, 1994. 6, 7" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 57, + 681, + 295, + 712 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 681, + 295, + 712 + ], + "spans": [ + { + "bbox": [ + 57, + 681, + 295, + 712 + ], + "type": "text", + "content": "[15] Karl R Gegenfurtner and Daniel C Kiper. Contrast detection in luminance and chromatic noise. JOSA A, 9(11):1880-1888, 1992. 6, 7" + } + ] + } + ], + "index": 15 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 316, + 73, + 554, + 713 + ], + "type": "list", + "angle": 0, + "index": 32, + "blocks": [ + { + "bbox": [ + 316, + 73, + 554, + 127 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 73, + 554, + 127 + ], + "spans": [ + { + "bbox": [ + 316, + 73, + 554, + 127 + ], + "type": "text", + "content": "[16] Robert Geirhos, Patricia Rubisch, Claudio Michaelis, Matthias Bethge, Felix A Wichmann, and Wieland Brendel. Imagenet-trained cnns are biased towards texture; increasing shape bias improves accuracy and robustness. arXiv preprint arXiv:1811.12231, 2018. 2" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 317, + 129, + 554, + 194 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 129, + 554, + 194 + ], + "spans": [ + { + "bbox": [ + 317, + 129, + 554, + 194 + ], + "type": "text", + "content": "[17] Robert Geirhos, Kantharaju Narayanappa, Benjamin Mitzkus, Tizian Thieringer, Matthias Bethge, Felix A Wichmann, and Wieland Brendel. Partial success in closing the gap between human and machine vision. Advances in Neural Information Processing Systems, 34:23885-23899, 2021. 2" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 316, + 196, + 553, + 228 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 196, + 553, + 228 + ], + "spans": [ + { + "bbox": [ + 316, + 196, + 553, + 228 + ], + "type": "text", + "content": "[18] MA Georgeson and GD Sullivan. Contrast constancy: deblurring in human vision by spatial frequency channels. The Journal of physiology, 252(3):627-656, 1975. 1, 2, 6, 8" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 317, + 230, + 553, + 271 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 230, + 553, + 271 + ], + "spans": [ + { + "bbox": [ + 317, + 230, + 553, + 271 + ], + "type": "text", + "content": "[19] Abhijay Ghildyal, Nabajeet Barman, and Saman Zadtootaghaj. Foundation models boost low-level perceptual similarity metrics. arXiv preprint arXiv:2409.07650, 2024. 2" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 316, + 274, + 553, + 297 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 274, + 553, + 297 + ], + "spans": [ + { + "bbox": [ + 316, + 274, + 553, + 297 + ], + "type": "text", + "content": "[20] Ian J. Goodfellow, Jonathon Shlens, and Christian Szegedy. Explaining and harnessing adversarial examples. 2014. 2" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 317, + 298, + 554, + 352 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 298, + 554, + 352 + ], + "spans": [ + { + "bbox": [ + 317, + 298, + 554, + 352 + ], + "type": "text", + "content": "[21] Kaiming He, Xinlei Chen, Saining Xie, Yanghao Li, Piotr Dólár, and Ross Girshick. Masked autoencoders are scalable vision learners. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 16000-16009, 2022. 3" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 317, + 354, + 553, + 418 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 354, + 553, + 418 + ], + "spans": [ + { + "bbox": [ + 317, + 354, + 553, + 418 + ], + "type": "text", + "content": "[22] Gabriel Ilharco, Mitchell Wortsman, Ross Wightman, Cade Gordon, Nicholas Carlini, Rohan Taori, Achal Dave, Vaishaal Shankar, Hongseok Namkoong, John Miller, Hannaneh Hajishirzi, Ali Farhadi, and Ludwig Schmidt. Openclip, 2021. If you use this software, please cite it as below. 1, 3" + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 317, + 421, + 553, + 475 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 421, + 553, + 475 + ], + "spans": [ + { + "bbox": [ + 317, + 421, + 553, + 475 + ], + "type": "text", + "content": "[23] Alexander Kirillov, Eric Mintun, Nikhila Ravi, Hanzi Mao, Chloe Rolland, Laura Gustafson, Tete Xiao, Spencer Whitehead, Alexander C Berg, Wan-Yen Lo, et al. Segment anything. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 4015-4026, 2023. 2, 3" + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 317, + 476, + 553, + 509 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 476, + 553, + 509 + ], + "spans": [ + { + "bbox": [ + 317, + 476, + 553, + 509 + ], + "type": "text", + "content": "[24] J.J. Kulikowski. Effective contrast constancy and linearity of contrast sensation. Vision Research, 16(12):1419-1431, 1976. Citation Key: Kulikowski1976a. 2, 8" + } + ] + } + ], + "index": 25 + }, + { + "bbox": [ + 316, + 510, + 553, + 543 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 510, + 553, + 543 + ], + "spans": [ + { + "bbox": [ + 316, + 510, + 553, + 543 + ], + "type": "text", + "content": "[25] G E Legge and J M Foley. Contrast masking in human vision. Journal of the Optical Society of America, 70(12):1458-71, 1980. Citation Key: Legge1980. 1" + } + ] + } + ], + "index": 26 + }, + { + "bbox": [ + 316, + 544, + 553, + 577 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 544, + 553, + 577 + ], + "spans": [ + { + "bbox": [ + 316, + 544, + 553, + 577 + ], + "type": "text", + "content": "[26] Qiang Li, Alex Gomez-Villa, Marcelo Bertalmó, and Jesús Malo. Contrast sensitivity functions in autoencoders. Journal of Vision, 22(6):8-8, 2022. 2" + } + ] + } + ], + "index": 27 + }, + { + "bbox": [ + 316, + 578, + 553, + 610 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 578, + 553, + 610 + ], + "spans": [ + { + "bbox": [ + 316, + 578, + 553, + 610 + ], + "type": "text", + "content": "[27] David G. Lowe. Distinctive image features from scale-invariant keypoints. International Journal of Computer Vision, 60(2):91-110, 2004. 2" + } + ] + } + ], + "index": 28 + }, + { + "bbox": [ + 317, + 612, + 553, + 655 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 612, + 553, + 655 + ], + "spans": [ + { + "bbox": [ + 317, + 612, + 553, + 655 + ], + "type": "text", + "content": "[28] Lei Lu, Yancheng Cai, Hua Huang, and Ping Wang. An efficient fine-grained vehicle recognition method based on part-level feature optimization. Neurocomputing, 536:40-49, 2023. 2" + } + ] + } + ], + "index": 29 + }, + { + "bbox": [ + 316, + 657, + 553, + 689 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 657, + 553, + 689 + ], + "spans": [ + { + "bbox": [ + 316, + 657, + 553, + 689 + ], + "type": "text", + "content": "[29] Yilong Lv, Yancheng Cai, Yujie He, and Min Li. Drkd: Decoupling response-based distillation for object detection. Pattern Recognition, 161:111275, 2025. 2" + } + ] + } + ], + "index": 30 + }, + { + "bbox": [ + 316, + 691, + 553, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 691, + 553, + 713 + ], + "spans": [ + { + "bbox": [ + 316, + 691, + 553, + 713 + ], + "type": "text", + "content": "[30] Rafal K Mantiuk, Param Hanji, Maliha Ashraf, Yuta Asano, and Alexandre Shapiro. Colorvideoovdp: A visual difference" + } + ] + } + ], + "index": 31 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "text", + "content": "20047" + } + ] + } + ], + "index": 33 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 8 + }, + { + "para_blocks": [ + { + "bbox": [ + 56, + 72, + 295, + 712 + ], + "type": "list", + "angle": 0, + "index": 12, + "blocks": [ + { + "bbox": [ + 76, + 72, + 294, + 95 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 76, + 72, + 294, + 95 + ], + "spans": [ + { + "bbox": [ + 76, + 72, + 294, + 95 + ], + "type": "text", + "content": "predictor for image, video and display distortions. arXiv preprint arXiv:2401.11485, 2024.3" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 56, + 96, + 295, + 193 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 96, + 295, + 193 + ], + "spans": [ + { + "bbox": [ + 56, + 96, + 295, + 193 + ], + "type": "text", + "content": "[31] Maxime Oquab, Timothée Darcet, Theo Moutakanni, Huy V. Vo, Marc Szafraniec, Vasil Khalidov, Pierre Fernandez, Daniel Haziza, Francisco Massa, Alaaeldin El-Nouby, Russell Howes, Po-Yao Huang, Hu Xu, Vasu Sharma, Shang-Wen Li, Wojciech Galuba, Mike Rabbat, Mido Assran, Nicolas Ballas, Gabriel Synnaeve, Ishan Misra, Herve Jegou, Julien Mairal, Patrick Labatut, Armand Joulin, and Piotr Bojanowski. Dinov2: Learning robust visual features without supervision, 2023. 2, 3" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 56, + 194, + 294, + 248 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 194, + 294, + 248 + ], + "spans": [ + { + "bbox": [ + 56, + 194, + 294, + 248 + ], + "type": "text", + "content": "[32] Eli Peli. Suprathreshold contrast perception across differences in mean luminance: effects of stimulus size, dichoptic presentation, and length of adaptation. Journal of the Optical Society of America A, 12(5):817, 1995. Citation Key: Peli1995. 2, 8" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 56, + 250, + 294, + 304 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 250, + 294, + 304 + ], + "spans": [ + { + "bbox": [ + 56, + 250, + 294, + 304 + ], + "type": "text", + "content": "[33] Alec Radford, Jong Wook Kim, Chris Hallacy, A. Ramesh, Gabriel Goh, Sandhini Agarwal, Girish Sastry, Amanda Askell, Pamela Mishkin, Jack Clark, Gretchen Krueger, and Ilya Sutskever. Learning transferable visual models from natural language supervision. In ICML, 2021. 1, 3" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 56, + 304, + 294, + 369 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 304, + 294, + 369 + ], + "spans": [ + { + "bbox": [ + 56, + 304, + 294, + 369 + ], + "type": "text", + "content": "[34] Alec Radford, Jong Wook Kim, Chris Hallacy, Aditya Ramesh, Gabriel Goh, Sandhini Agarwal, Girish Sastry, Amanda Askell, Pamela Mishkin, Jack Clark, et al. Learning transferable visual models from natural language supervision. In International conference on machine learning, pages 8748-8763. PMLR, 2021. 2" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 56, + 371, + 294, + 425 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 371, + 294, + 425 + ], + "spans": [ + { + "bbox": [ + 56, + 371, + 294, + 425 + ], + "type": "text", + "content": "[35] Nikhila Ravi, Valentin Gabeur, Yuan-Ting Hu, Ronghang Hu, Chaitanya Ryali, Tengyu Ma, Haitham Khedr, Roman Rädle, Chloe Rolland, Laura Gustafson, et al. Sam 2: Segment anything in images and videos. arXiv preprint arXiv:2408.00714, 2024. 3" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 56, + 426, + 295, + 480 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 426, + 295, + 480 + ], + "spans": [ + { + "bbox": [ + 56, + 426, + 295, + 480 + ], + "type": "text", + "content": "[36] Robin Rombach, Andreas Blattmann, Dominik Lorenz, Patrick Esser, and Björn Ommer. High-resolution image synthesis with latent diffusion models. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 10684-10695, 2022. 3" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 56, + 482, + 294, + 523 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 482, + 294, + 523 + ], + "spans": [ + { + "bbox": [ + 56, + 482, + 294, + 523 + ], + "type": "text", + "content": "[37] Jyrki Rovamo, Olavi Luntinen, and Risto Nasänen. Modelling the dependence of contrast sensitivity on grating area and spatial frequency. Vision Research, 33(18):2773-2788, 1993. 5" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 56, + 526, + 294, + 568 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 526, + 294, + 568 + ], + "spans": [ + { + "bbox": [ + 56, + 526, + 294, + 568 + ], + "type": "text", + "content": "[38] Jyrki Rovamo, Juvi Mustonen, and Risto Nasänen. Neural modulation transfer function of the human visual system at various eccentricities. Vision Research, 35(6):767-774, 1995. 5" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 56, + 570, + 294, + 635 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 570, + 294, + 635 + ], + "spans": [ + { + "bbox": [ + 56, + 570, + 294, + 635 + ], + "type": "text", + "content": "[39] Stephen Tian, Yancheng Cai, Hong-Xing Yu, Sergey Zakharov, Katherine Liu, Adrien Gaidon, Yunzhu Li, and Jia-jun Wu. Multi-object manipulation via object-centric neural scattering functions. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 9021-9031, 2023. 2" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 56, + 636, + 294, + 678 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 636, + 294, + 678 + ], + "spans": [ + { + "bbox": [ + 56, + 636, + 294, + 678 + ], + "type": "text", + "content": "[40] Shikhar Tuli, Ishita Dasgupta, Erin Grant, and Thomas L Griffiths. Are convolutional neural networks or transformers more like human vision? arXiv preprint arXiv:2105.07197, 2021. 2" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 56, + 680, + 294, + 712 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 680, + 294, + 712 + ], + "spans": [ + { + "bbox": [ + 56, + 680, + 294, + 712 + ], + "type": "text", + "content": "[41] A Van Meeteren and JM Valeton. Effects of pictorial noise interfering with visual detection. JOSA A, 5(3):438-444, 1988. 7" + } + ] + } + ], + "index": 11 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 315, + 73, + 553, + 328 + ], + "type": "list", + "angle": 0, + "index": 19, + "blocks": [ + { + "bbox": [ + 315, + 73, + 553, + 95 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 315, + 73, + 553, + 95 + ], + "spans": [ + { + "bbox": [ + 315, + 73, + 553, + 95 + ], + "type": "text", + "content": "[42] Andrew B Watson. Efficiency of a model human image code. JOSA A, 4(12):2401-2417, 1987. 7" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 316, + 96, + 553, + 128 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 96, + 553, + 128 + ], + "spans": [ + { + "bbox": [ + 316, + 96, + 553, + 128 + ], + "type": "text", + "content": "[43] Andrew B Watson and Joshua A Solomon. Model of visual contrast gain control and pattern masking. JOSA A, 14(9): 2379-2391, 1997. 7" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 316, + 129, + 553, + 173 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 129, + 553, + 173 + ], + "spans": [ + { + "bbox": [ + 316, + 129, + 553, + 173 + ], + "type": "text", + "content": "[44] Felix A Wichmann and Robert Geirhos. Are deep neural networks adequate behavioral models of human visual perception? Annual Review of Vision Science, 9(1):501-524, 2023. 2" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 316, + 175, + 553, + 207 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 175, + 553, + 207 + ], + "spans": [ + { + "bbox": [ + 316, + 175, + 553, + 207 + ], + "type": "text", + "content": "[45] Hugh R Wilson and James R Bergen. A four mechanism model for threshold spatial vision. Vision research, 19(1): 19-32, 1979. 7" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 316, + 209, + 553, + 262 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 209, + 553, + 262 + ], + "spans": [ + { + "bbox": [ + 316, + 209, + 553, + 262 + ], + "type": "text", + "content": "[46] Fei Yin, Yong Zhang, Baoyuan Wu, Yan Feng, Jingyi Zhang, Yanbo Fan, and Yujiu Yang. Generalizable black-box adversarial attack with meta learning. IEEE transactions on pattern analysis and machine intelligence, 46(3):1804-1818, 2023. 2" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 316, + 264, + 553, + 328 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 264, + 553, + 328 + ], + "spans": [ + { + "bbox": [ + 316, + 264, + 553, + 328 + ], + "type": "text", + "content": "[47] Jingdong Zhang, Jiayuan Fan, Peng Ye, Bo Zhang, Hancheng Ye, Baopu Li, Yancheng Cai, and Tao Chen. Bridgenet: Comprehensive and effective feature interactions via bridge feature for multi-task dense predictions. IEEE Transactions on Pattern Analysis and Machine Intelligence, 2025. 2" + } + ] + } + ], + "index": 18 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "text", + "content": "20048" + } + ] + } + ], + "index": 20 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 9 + } + ], + "_backend": "vlm", + "_version_name": "2.6.4" +} \ No newline at end of file diff --git a/2025/Do ImageNet-trained Models Learn Shortcuts_ The Impact of Frequency Shortcuts on Generalization/f684a2ee-5aa0-48af-b0b3-fb7e06614e21_content_list.json b/2025/Do ImageNet-trained Models Learn Shortcuts_ The Impact of Frequency Shortcuts on Generalization/f684a2ee-5aa0-48af-b0b3-fb7e06614e21_content_list.json new file mode 100644 index 0000000000000000000000000000000000000000..e1302ce44d0ab3a8c9e6dd618652e58ad17ce204 --- /dev/null +++ b/2025/Do ImageNet-trained Models Learn Shortcuts_ The Impact of Frequency Shortcuts on Generalization/f684a2ee-5aa0-48af-b0b3-fb7e06614e21_content_list.json @@ -0,0 +1,1181 @@ +[ + { + "type": "text", + "text": "Do ImageNet-trained models learn shortcuts? The impact of frequency shortcuts on generalization", + "text_level": 1, + "bbox": [ + 233, + 130, + 763, + 176 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Shunxin Wang Raymond Veldhuis Nicola Strisciuglio University of Twente", + "bbox": [ + 264, + 203, + 732, + 239 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "{s.wang-2,r.n.j.veldhuis,n.strisciuglio}@utwente.nl", + "bbox": [ + 261, + 241, + 730, + 257 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Abstract", + "text_level": 1, + "bbox": [ + 246, + 291, + 326, + 306 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Frequency shortcuts refer to specific frequency patterns that models heavily rely on for correct classification. Previous studies have shown that models trained on small image datasets often exploit such shortcuts, potentially impairing their generalization performance. However, existing methods for identifying frequency shortcuts require expensive computations and become impractical for analyzing models trained on large datasets. In this work, we propose the first approach to more efficiently analyze frequency shortcuts at a large scale. We show that both CNN and transformer models learn frequency shortcuts on ImageNet. We also expose that frequency shortcut solutions can yield good performance on out-of-distribution (OOD) test sets which largely retain texture information. However, these shortcuts, mostly aligned with texture patterns, hinder model generalization on rendition-based OOD test sets. These observations suggest that current OOD evaluations often overlook the impact of frequency shortcuts on model generalization. Future benchmarks could thus benefit from explicitly assessing and accounting for these shortcuts to build models that generalize across a broader range of OOD scenarios. Codes are available at https://github.com/nis-research/hfss.", + "bbox": [ + 89, + 325, + 485, + 672 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "1. Introduction", + "text_level": 1, + "bbox": [ + 89, + 707, + 220, + 722 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Superficial correlations between data and ground truth [5, 9] can be learned by models to minimize training objectives with the least effort [34]. This learning behavior is called shortcut learning, which either harms the generalization performance of models or gives an illusion of good generalization abilities when learned shortcuts are present in OOD test sets [39]. Artifacts and visual cues that cause shortcut learning are recognizable by visual inspection of images, and their impact on the training of models can be mitigated through data selection or augmentation [5, 19, 27], i.e. counteracting the spurious correlations between data", + "bbox": [ + 89, + 734, + 483, + 900 + ], + "page_idx": 0 + }, + { + "type": "table", + "img_path": "images/e9e5485f2c317ae9ca07b9bfed19c83756ad1aec481fbc942f3917fd053c6ccb.jpg", + "table_caption": [ + "Table 1. Our method needs less computation times compared to [39] (using ResNet18 on an NVIDIA A40 GPU). Computational time of [39] on ImageNet-1k is estimated from their ImageNet-10 experiments, considering that it increases proportionally to the number of classes." + ], + "table_footnote": [], + "table_body": "
DatasetTime (h) [39]Time (h) (Ours)
CIFAR-107.50.5
ImageNet-1k8500 (354 days)220 (9.2 days)
", + "bbox": [ + 558, + 372, + 859, + 425 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "and ground truth.", + "bbox": [ + 511, + 431, + 630, + 445 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "However, there exist shortcuts in the Fourier domain, which are implicitly embedded in image data characteristics and not easily detectable by visual inspection [37, 39]. Such shortcut solutions consist of small frequency subsets that are easy-to-learn and sufficient for models to achieve high classification rate. Typically, they correspond to simple features like textures, shapes or colors in the spatial domain. Wang, et al. [39] identified frequency shortcuts by retaining relevant frequencies to classification of a certain class. The relevance of an individual frequency was measured by the loss value of the model tested on images of the class with that frequency removed. Thus, their approach requires computational time increasing proportionally to the number of classes in a dataset and the image resolution. Yet, frequency shortcuts usually consist of multiple frequencies. The method in [39] might overlook some shortcuts as individual frequency relevance to classification neglects joint contribution of frequencies. Furthermore, the computational burden limits its applicability for analyzing shortcut learning behavior of models trained on large datasets (e.g. ImageNet) with hundreds or thousands of classes. Several studies have shown that ImageNet-trained models are biased towards textures [7, 8]. However, there is no concrete evidence of what causes this phenomenon, although [39] largely attributes it to frequency shortcut learning.", + "bbox": [ + 509, + 446, + 906, + 824 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "In this work, we propose the first method that enables the uncovering of frequency shortcuts learned by models trained on large-scale datasets (e.g. ImageNet-1k), for analyzing learning behavior and explaining model generalization performance in different OOD scenarios. Our", + "bbox": [ + 511, + 825, + 908, + 902 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "CVF", + "bbox": [ + 106, + 2, + 181, + 42 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore.", + "bbox": [ + 236, + 0, + 810, + 46 + ], + "page_idx": 0 + }, + { + "type": "page_number", + "text": "25198", + "bbox": [ + 478, + 944, + 519, + 957 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "method improves computational efficiency as it applies parallel class-wise loss computation and hierarchical search in the Fourier domain (see Tab. 1), compared to [39]. Moreover, our method considers the joint contribution of frequencies to classification, thus being more effective in finding shortcuts. Our contributions are:", + "bbox": [ + 89, + 90, + 480, + 181 + ], + "page_idx": 1 + }, + { + "type": "list", + "sub_type": "text", + "list_items": [ + "1. We develop a hierarchical frequency shortcut search (HFSS) method, which enables the analysis of shortcut learning in the Fourier domain on large datasets with varying class counts. We reduce computational time and improve the effectiveness at identifying shortcuts.", + "2. We discover that ImageNet-trained models (both CNN and transformer architectures) are subject to learn frequency shortcuts. Different from shortcuts formed by visual cues [9], frequency shortcuts (easy-to-learn features) lead to good performance on both in-distribution (ID) and OOD tests if they do not block models from learning other semantics (difficult features).", + "3. Our HFSS enables a more comprehensive assessment of model generalizability by analyzing OOD data characteristics, specifically the presence of shortcuts. In existing evaluation frameworks, frequency shortcuts do not always impair OOD generalization performance. This emphasizes the importance of considering the role of shortcuts when designing future OOD evaluation benchmarks." + ], + "bbox": [ + 89, + 181, + 482, + 483 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "2. Related works", + "text_level": 1, + "bbox": [ + 89, + 497, + 235, + 512 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Shortcut learning. Models can learn shortcut solutions based on superficial correlations between data and ground truth [9, 11] to optimize training objectives with the least effort. This learning behavior is due to simplicity-bias [34], which is caused by inductive biases provided by gradient descent or components like ReLUs [35]. Shortcuts can be visual cues in the data like source tags and artificial markers [19, 27]. For instance, one fifth of the horse images from Pascal VOC dataset were found to contain a source tag, on which models rely as a discriminant feature to recognize horses [19]. Next to these visual cues, vision models are also subject to shortcuts implicitly existing in the frequency domain, which manifest as small sets of frequencies contributing significantly to image classification performance [37, 39]. The study in [39] explains why models exhibit a textures-bias for classification [8] from a frequency shortcut perspective, but does not provide further investigation on how frequency shortcuts affect the generalization and robustness performance of ImageNet-trained models, due to computational burden.", + "bbox": [ + 89, + 523, + 482, + 824 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Relying on simple shortcut solutions could harm the generalization and robustness performance of models, as shortcut learning appearing in early training might block models from learning other semantics related to the tasks at hand [3, 16, 39, 41]. However, such reliance could also cre", + "bbox": [ + 89, + 825, + 482, + 900 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "ates a false impression of good generalization when shortcuts are present in OOD test sets [1, 19, 39].", + "bbox": [ + 511, + 90, + 903, + 121 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Shortcut identification. Uncovering shortcuts helps understanding the generalizability of models on different out-of-distribution (OOD) data, explaining why models fail to generalize to or perform well on OOD data. It can provide shortcut-related prior knowledge to develop techniques that improve model generalization and robustness performance based on shortcut mitigation [27, 38]. However, common approaches are limited to identifying visually inspectable shortcuts such as text, watermark and color patches [19, 24, 27], using e.g. saliency maps [33]. Rather than directly identifying shortcuts in the data, in [2, 5] shortcut features present in individual images were quantified by assessing how difficult they were for models to learn. The authors in [28, 30, 32] uncovered shortcut features learned within representation space. The authors in [23] investigated shortcut learning by analyzing the relationship between semantic concepts using a knowledge graph, sharing a similar idea to [42]. Despite visual shortcut cues, there are non-observable shortcuts in the frequency domain, which are embedded in data characteristics. The work of [39] identified relevant frequencies to classification, which potentially contain shortcut information. As this technique measures the relevance of one frequency at a time, it requires computational costs increasing proportionally to the number of classes and image resolution. Thus, limited attention has been given to uncovering frequency shortcuts in large datasets, with existing analyses primarily focused on datasets containing a small number of classes [2, 37, 39, 42].", + "bbox": [ + 511, + 140, + 903, + 578 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Existing methods to identify shortcuts are limited to either manual inspection [19] or involve a time-consuming algorithm [37, 39]. Methods quantifying shortcut information or uncovering shortcut features primarily aim at mitigating shortcut learning, without explaining how shortcuts in the data impact model generalization performance. We are the first to enable frequency shortcut analysis of models trained on large-scale datasets and link them to generalization performance in different OOD settings.", + "bbox": [ + 511, + 579, + 905, + 715 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "OOD evaluation. The generalization and robustness performance of vision models are usually evaluated using extra data that is considered OOD, e.g. data collected at different time points [31], with different styles or renditions [8, 14, 36], with synthetic corruption effects [13, 17, 25, 40], and with adversarial noise [4, 12, 20, 43]. However, such benchmarks do not consider the impact of shortcuts learned by models and present in OOD test sets. This might ignore critical factors related to the generalization and robustness capabilities of models. The work in [7] explores model generalization from the perspectives of biases, e.g. texture,", + "bbox": [ + 511, + 734, + 906, + 902 + ], + "page_idx": 1 + }, + { + "type": "page_number", + "text": "25199", + "bbox": [ + 478, + 944, + 517, + 955 + ], + "page_idx": 1 + }, + { + "type": "image", + "img_path": "images/baac05343ec0eb17a342d50c17329627042080eed9e2ea497d7b84740056567b.jpg", + "image_caption": [ + "Figure 1. Scheme of HFSS. Starting from stage 2, we sample frequency patches from a random frequency subset searched in previous stage. This confines the size of search space. The white patches in the binary masks indicate sampled frequency patches." + ], + "image_footnote": [], + "bbox": [ + 138, + 95, + 851, + 354 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "shape and spectral biases. In this work, we investigate the impact of frequency shortcuts on model generalization abilities, establishing connections among these different biases. Our work provides insights into when frequency shortcuts yield good performance or are harmful to model generalization and robustness performance.", + "bbox": [ + 89, + 406, + 485, + 497 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3. Method", + "text_level": 1, + "bbox": [ + 89, + 508, + 181, + 523 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "We propose a method to identify (non-visible) frequency shortcuts linked to intrinsic data characteristics rather than visual cues as in [27]. We reduce significantly computational demands compared to [39]. Our method is the first solution for shortcut identification that enables the analysis of models trained on datasets with a large number of samples and classes. In the following sections, we detail our measurement of shortcut learning in models and examine the impact of shortcuts on generalization.", + "bbox": [ + 89, + 535, + 483, + 672 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3.1. Identifying frequency shortcuts", + "text_level": 1, + "bbox": [ + 89, + 679, + 369, + 695 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "We propose a method called hierarchical frequency shortcut search (HFSS), which exploits hierarchical search of frequency subsets in the image Fourier spectrum. This reduces computational time compared to exhaustive search strategies, e.g. [38, 39]. The search is separated into several stages, each stage gradually narrowing down the spectrum search space in a coarse-to-fine manner and detecting frequency subsets that models strongly rely on for classification. The scheme of HFSS is shown in Fig. 1.", + "bbox": [ + 89, + 700, + 483, + 838 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Hierarchical search for frequency shortcuts. In order to discover the frequency subsets that models rely heavily on for correct classification of each class, we sample different", + "bbox": [ + 89, + 854, + 483, + 900 + ], + "page_idx": 2 + }, + { + "type": "image", + "img_path": "images/68a3b6c4933f29d4b533355d62ccb9283cb622b4843470fba5881fa86dbba681.jpg", + "image_caption": [ + "Figure 2. The frequency spectrum is separated into patches, with $p\\%$ sampled for shortcut evaluation." + ], + "image_footnote": [], + "bbox": [ + 612, + 405, + 790, + 532 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "frequency combinations that potentially contain frequency shortcuts. We use random sampling as it allows to easily consider joint contribution of frequencies to classification, improving the effectiveness of frequency shortcut identification. Random sampling has been shown to contribute to stable search results in data augmentation and reinforcement learning [21, 26, 29], while requiring lower computations w.r.t. optimization-based methods.", + "bbox": [ + 511, + 568, + 906, + 688 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "There are numerous combinations of frequencies by sampling frequency components directly. Instead of doing an exhaustive search, we apply hierarchical search in the Fourier spectrum: we divide the Fourier spectrum into frequency patches and sample these frequency patches to generate candidates of frequency subsets that contain shortcut information, as illustrated in Fig. 2. We exploit overlapping shifted windows in vertical and horizontal directions [22] to prepare frequency patches and thus avoid border effects, in addition to evenly separating the Fourier spectrum. HFSS consists of multiple search stages, the number of which depends on image resolution and desired frequency resolution of the identified frequency subsets. Lower image or frequency resolution requires fewer search stages.", + "bbox": [ + 511, + 689, + 908, + 901 + ], + "page_idx": 2 + }, + { + "type": "page_number", + "text": "25200", + "bbox": [ + 478, + 944, + 519, + 957 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "In summary, the overall process of HFSS is as follows. Each stage of HFSS consists of three steps, namely (1) sampling $B$ frequency subsets which contain $p\\%$ frequency patches, (2) evaluating shortcut information, and (3) forwarding subsets for next stage. At the initial search stage, we sample frequency patches with a large size and evaluate their shortcut information. We evaluate the frequency subsets according to their contributions to classification: we process a subset of training images by retaining a frequency subset and measure the class-wise loss of a model. Frequency subsets that contribute to a lower loss value are stronger candidates to indicate frequency shortcuts as they are sufficient for the model to achieve a high prediction score. We rank frequency subsets by the loss values and use the top- $N$ masks of each class in the next search stage.", + "bbox": [ + 89, + 90, + 480, + 316 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Starting from the second search stage, we sample frequency patches from one of the top- $N$ frequency subsets in the previous stage, with a smaller patch size. This confines the search space. In the final search stage, the top-1 frequency subset of each class is considered dominating the classification of the class concerned. Same as [39], we use binary masks to represent these frequency subsets, called Dominant Frequency Maps (DFMs).", + "bbox": [ + 89, + 318, + 480, + 439 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "The sampling percentage $p\\%$ of frequency patches, the number of sampled frequency subsets $B$ and the frequency resolution of DFMs are hyperparameters of the search algorithm (details are in the supplementary material).", + "bbox": [ + 89, + 440, + 480, + 501 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "3.2. Measuring the degree of shortcut learning", + "text_level": 1, + "bbox": [ + 89, + 513, + 449, + 530 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Analyzing models learning behavior in a class-wise manner, as done in [39], is labor-intensive if there are hundreds or thousands of classes. To provide a broad overview of model learning behavior on training data, we categorize all classes in a dataset into two groups: (1) classes subject to shortcuts and (2) non-shortcut classes. A class is considered subject to shortcuts if its true positive rate (TPR) surpasses a given threshold when the model is tested on images filtered to retain only dominant frequencies of the class concerned. We use TPR, as a high TPR indicates that the frequency subset is sufficient to achieve a high classification rate, thereby acting as a shortcut. The frequency-filtered images are referred to as DFM-filtered images.", + "bbox": [ + 89, + 537, + 482, + 733 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "For each class, we test the model on both the original test images and DFM-filtered images, computing TPR for both. We denote the TPR on the original images of class $c_{i}$ as $\\mathrm{TPR}_{c_i}$ and the TPR on the DFM-filtered images as $\\mathrm{TPR}_{c_i}^{DFM}$ . If $\\mathrm{TPR}_{c_i}^{DFM} > t$ (where $t \\in [0,1]$ is a predefined threshold), class $c_{i}$ is considered subject to shortcuts; otherwise, it is considered as one of the non-shortcut classes. We compute the average TPR values of both groups, shortcut and non-shortcut classes, at different thresholds $t$ . We note them as $\\mathrm{AvgTPR}_{\\mathrm{sct}}@t$ , $\\mathrm{AvgTPR}_{\\mathrm{sct}}^{\\mathrm{DFM}}@t$ for the shortcut classes, and $\\mathrm{AvgTPR}_{\\mathrm{non-sct}}@t$ and $\\mathrm{AvgTPR}_{\\mathrm{non-sct}}^{\\mathrm{DFM}}@t$", + "bbox": [ + 89, + 734, + 482, + 902 + ], + "page_idx": 3 + }, + { + "type": "image", + "img_path": "images/3f0dca1220b8a62254257c0f47a68f10a49e2c576a43e4283ce75ab19f5d00ec.jpg", + "image_caption": [ + "Figure 3. ResNet18 tested on original test images and DFM-filtered images of ImageNet-1k. Blue line shows the average TPR of classes subject to shortcuts on the original test images, and the orange line shows the results of non-shortcut classes, at different threshold $t$ . The green and red lines correspond to results tested on DFM-filtered images. Lower $t$ indicates weak shortcuts and higher $t$ signifies stronger ones. The size of each point reflects the number of classes. The larger the size, the more classes included." + ], + "image_footnote": [], + "bbox": [ + 583, + 92, + 834, + 193 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "for the non-shortcut classes. A higher threshold $t$ corresponds to stronger shortcuts, as dominant frequencies alone are sufficient for accurate classification. An example of the results of these metrics at different threshold values are given in Fig. 3, where the size of each point reflects the number of classes in the two groups. Larger points indicate a higher number of classes, which suggests that the model has a stronger tendency to shortcut learning.", + "bbox": [ + 511, + 321, + 903, + 443 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "We do not average $\\mathrm{TPR}^{\\mathrm{DFM}}$ across all classes because its value for many non-shortcut classes is close to zero (specifically for large datasets). This can result in a close-to-zero $\\mathrm{AvgTPR}^{\\mathrm{DFM}}$ which does not provide useful information for further analysis.", + "bbox": [ + 511, + 443, + 903, + 520 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "3.3. Impact of shortcut on generalization", + "text_level": 1, + "bbox": [ + 511, + 531, + 828, + 546 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "To assess how frequency shortcuts impact the generalization and robustness performance of models trained on ImageNet-1k, we evaluate their performance on several datasets. We perform ID tests on ImageNet-1k (IN-1k) [6] and ImageNet-v2 (IN-v2) [31] test sets. They are considered in-distribution as they are collected using the same protocol as the training set. Furthermore, we use ImageNet-C (IN-C) [13] to benchmark the robustness of models against appearance image corruptions, e.g. noise, blur and weather changes. To evaluate the generalizability of models to images with different renditions, we use ImageNet-Renditions (IN-R) [14] and ImageNet-Sketch (IN-S) [36]. IN-R contains renditions like cartoons, paintings, art and toys of 200 classes from IN-1k. IN-S has the same number of classes as IN-1k, containing images of hand-drawn sketches. These datasets contain visual renditions, which serve to evaluate how the reliance on texture cues (corresponding to most frequency shortcuts) impacts on generalization. To measure the adversarial robustness of models, we apply the fast gradient sign method (FGSM) attacks [10] to the validation set of IN-1k, with $L_{\\infty} = 4 / 255$ . We calculate the average TPR of shortcut and non-shortcut classes on OOD data, that we compare to the results on ID data.", + "bbox": [ + 511, + 553, + 906, + 901 + ], + "page_idx": 3 + }, + { + "type": "page_number", + "text": "25201", + "bbox": [ + 478, + 944, + 517, + 955 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "4. Experiments", + "text_level": 1, + "bbox": [ + 89, + 89, + 225, + 107 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "We start with investigating the trade-off between efficiency (i.e. required computational time) and effectiveness (i.e. capability of finding shortcuts), varying the configuration of the number of sampled frequency subsets at each stage, noted as $B_{s}$ ( $s$ is the index of stage). Then, we analyze frequency shortcut learning of models trained on IN-1k, relating it to performance results under different OOD scenarios.", + "bbox": [ + 89, + 114, + 483, + 223 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Setup. We use CIFAR-10 (C-10) [18] to explore the configuration of $B$ , and configure HFSS with four search stages, with patch size of $8 \\times 8$ , $4 \\times 4$ , $2 \\times 2$ and $1 \\times 1$ , respectively. We then perform larger-scale shortcut analysis using IN-1k: we use six search stages, with patch size of $56 \\times 56$ , $28 \\times 28$ , $14 \\times 14$ , $8 \\times 8$ , $4 \\times 4$ and $2 \\times 2$ . The patch (frequency) resolution in the final stage is the same as that in [39]. At each stage, we sample $p = 60\\%$ frequency patches to form frequency subsets. It results in DFMs containing about $5\\%$ frequencies of the whole spectrum of ImageNet images and about $15\\%$ frequencies of CIFAR images. The number of training images sampled for shortcut information evaluation is the same as that of their corresponding test set. We evaluate the generalization and robustness performance of ImageNet-trained models on IN-v2 [31], IN-C [13], IN-R [14], IN-S [36] and using FGSM [10] attacks. We also carry out experiments on ImageNet-10 (IN-10) [15] with a ResNet18 model to compare HFSS with single-frequency removal-based method in [39]. Following their setups, we use ImageNet-SCT (IN-SCT) for OOD evaluation. Our configuration of HFSS enables direct comparison with the results reported in [39] as DFMs contain around $5\\%$ frequencies of whole spectrum. We report training configurations and additional results in the supplementary material.", + "bbox": [ + 91, + 244, + 483, + 609 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "4.1. Configuration of HFSS", + "text_level": 1, + "bbox": [ + 89, + 619, + 305, + 636 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Sampled frequency subsets. At stage $s$ , we sample $B_{s}$ frequency subsets, such that HFSS can explore different combinations of frequency patterns, validating their contained shortcut information. However, sampling an excessive number of subsets increases the required search time, as more subsets are for evaluating their contributions to classification results. To explore the impact of the number of sampling operation $B_{s}$ on shortcut identification, we perform experiments on C-10 using a ResNet18 backbone, setting the number of sampled frequency subsets in each stage to be as large as possible under limited computational time. We hypothesize that HFSS obtains stable results given enough candidate frequency subsets. We set $B_{1}, B_{2}, B_{3}$ , and $B_{4}$ to be 1000, 2000, 4000 and 8000 respectively, which we note as configuration CF-1. The increasing number of candidate frequency subsets at consecutive stages is determined by the increase in possible frequency patch combinations as patch", + "bbox": [ + 89, + 643, + 483, + 901 + ], + "page_idx": 4 + }, + { + "type": "image", + "img_path": "images/3a9d3b140d81d1445e08aec2994b8cf68392596b766f0d54a2e1dfb0b1f501fd.jpg", + "image_caption": [ + "Figure 4. The best class-wise loss vs. the number of sampled frequency subsets at each stage." + ], + "image_footnote": [], + "bbox": [ + 521, + 90, + 901, + 373 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "size decreases.", + "text_level": 1, + "bbox": [ + 511, + 414, + 614, + 426 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "We run the configuration CF-1 five times and note the stability of our HFSS search algorithms. We track the best (lowest) loss values achieved for each class when the model is tested on DFM-filtered images. In Fig. 4, we report the tracking statistics with average loss values (lines) and standard deviations (shadows) over the five trials for each class, which are further used as a starting point to optimize the configuration of HFSS. As the number of sampled frequency subsets increases, the best loss for each class converges, indicating the stability of HFSS in identifying similar shortcuts learned by models when a sufficient amount of candidate frequency subsets are sampled for validating shortcut information. The standard deviation recorded at different stages is relatively small compared to the range of best loss. For class airplane we observe an outlier outcome at stage 3. In the supplementary material, we present visualization of shortcut cues (DFM-filtered images) computed across the five trials. These visualizations show that the identified shortcut patterns differ in the orientations of strip-like features, corresponding to line features in the image of airplane. This suggests that frequency shortcuts do not arise from fixed frequency subsets; rather, spatial features they correspond to may be visually similar but composed of different frequencies, determining an higher standard deviation in Fig. 4.", + "bbox": [ + 511, + 429, + 906, + 808 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Efficiency vs. effectiveness. Based on the results in Fig. 4, we optimize CF-1 for efficiency and effectiveness in shortcut identification. The required search time scales up proportionally to $B_{s}$ , i.e. the more candidate frequency subsets are sampled, the more time is needed to verify their", + "bbox": [ + 511, + 825, + 908, + 902 + ], + "page_idx": 4 + }, + { + "type": "page_number", + "text": "25202", + "bbox": [ + 478, + 944, + 519, + 957 + ], + "page_idx": 4 + }, + { + "type": "image", + "img_path": "images/44804d03646bcc9d1c3f209e82d0f03083b58153e96f1f4e5679eb69fb893f86.jpg", + "image_caption": [ + "Figure 5. Average TPR vs. search time, where search time increases proportionally with the number of sampled frequency subsets." + ], + "image_footnote": [], + "bbox": [ + 174, + 90, + 395, + 215 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/c2ad637d9e710d9e444390052802dbdc77589e53c0ac22f3c0809fbc2c6ee9d1.jpg", + "image_caption": [ + "Figure 6. The number of shortcut classes given different threshold $t$ search by CF-1 and CF-2.10. Using CF-2.10 uncovers most shortcuts identified by using CF-1 at low thresholds, with computational time reduced to a factor of around 200." + ], + "image_footnote": [], + "bbox": [ + 179, + 260, + 390, + 383 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "relevance to classification. We perform experiments using different $B_{s}$ , investigating how the size of the search space affects the identification of shortcuts. Based on the tracking statistics in Fig. 4, which shows that increasing $B_{s}$ beyond a certain point does not decrease loss significantly, we adjust $B_{1}$ to 200, $B_{2}$ to 800 and $B_{4}$ to 4000 (CF-2.1). We further reduce $B_{3}$ and $B_{4}$ to 2000 (CF-2.2). We also design eight other configurations (see details in the supplementary material), with CF-2.10 being the most efficient as it samples the lowest number of candidate frequency subsets.", + "bbox": [ + 89, + 459, + 482, + 609 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "To compare the performance of HFSS under different configurations, we compute the average TPR over all classes when tested on DFM-filtered images. This indicates the general relevance of the searched frequency subsets to classification. A higher average TPR indicate higher relevance, thus the frequency subsets contain more frequency shortcut information. We run each configuration five times, computing the mean and standard deviations of average TPRs over the five trials. We report the results in Fig. 5 where the left-most point corresponds to average TPR of CF-2.10 and the right-most point corresponds to that of CF-1. We observe that as more frequency subsets are sampled, average TPR increases and saturates. This demonstrates that sampling more candidate frequency subsets does not significantly improve the performance of HFSS, while requiring more computations.", + "bbox": [ + 89, + 611, + 482, + 852 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "We compare CF-1 and CF-2.10 in Fig. 6 based on the number of classes with identified shortcuts at different thresholds. We observe that CF-2.10 uncovers most of the", + "bbox": [ + 89, + 854, + 482, + 898 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "shortcuts identified by using the more complex CF-1 at low threshold level. Although CF-2.10 misses stronger shortcuts (at higher threshold levels), it achieves a $\\sim 200\\times$ reduction in computational time compared to CF-1. A more time-efficient configuration is less effective at finding shortcuts than the more complex one, but it still manages to identify shortcuts at low thresholds.", + "bbox": [ + 511, + 90, + 903, + 196 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "The observations made on C-10 allow to estimate initial configurations for analysis on larger-scale datasets. We track the loss statistics of HFSS applied to ImageNet (provided in the supplementary material) and choose the value $B_{s}$ for ImageNet experiments in Sec. 4.2 as the one where the loss values show no significant decrease. Given the computational intensity of identifying shortcuts in large-scale datasets, our work on IN-1k primarily aims at uncovering shortcuts with affordable computational efforts.", + "bbox": [ + 511, + 196, + 903, + 333 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "4.2. Shortcut identification on IN-1k", + "text_level": 1, + "bbox": [ + 511, + 344, + 795, + 358 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "We apply HFSS to examine frequency shortcuts learned by ImageNet models and how they impact model generalization and robustness. The search configurations for ImageNet models, aimed at time-efficiency and effectiveness at finding shortcuts, are in the supplementary material.", + "bbox": [ + 511, + 366, + 903, + 441 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Results on IN-1k, IN-v2 and IN-C. In Figs. 7 and 8, we report the results of the analysis of impact of shortcuts on model performance. In general, models perform better predictions on images of classes affected by shortcut learning (blue lines) compared to non-shortcut classes (orange lines) in both ID and corruption tests. Models subject to frequency shortcuts yield good performance in the robustness and generalization tests when textures information is largely preserved. For ResNet and ViT models, when $t \\geq 0.8$ , the AvgTPR $_{\\text{sct}}^{\\text{DFM}}$ exceeds AvgTPR $_{\\text{sct}}$ , meaning that these models rely predominantly on dominant frequencies (corresponding to simple features) for classification of shortcut-affected classes, preventing from learning class-related semantics features. For instance, at $t = 0.9$ , only one class ('window screen') is found subject to frequency shortcuts: ResNet18 achieves TPR=0.66 on full-spectrum images of this class and TPR=0.94 on DFM-filtered images (similarly to ResNet50 and ViT-b). Differently, CCT exhibits lower AvgTPR $_{\\text{sct}}^{\\text{DFM}}@0.9$ than AvgTPR $_{\\text{sct}}@0.9$ , indicating that while frequency shortcuts contribute significantly to classification, this model also manage to leverage other semantic information: a strong frequency shortcut (e.g. when $t > 0.7$ ) does not necessarily mean that it is the only information a model uses for classification. We further compare CCT with ResNet50 as they perform similarly on IN-1k (80.57% and 80.1% respectively). In IN-v2 and IN-C, CCT achieves 74.81% and 57.73% prediction rates, having better performance than ResNet50 (74.17% in IN-v2 and 48.85% in IN-C). Considering the highest de", + "bbox": [ + 511, + 462, + 903, + 900 + ], + "page_idx": 5 + }, + { + "type": "page_number", + "text": "25203", + "bbox": [ + 478, + 944, + 517, + 955 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/2f410602a5803270cfd3b01ddd032747c75980460b1f0551cdc1e2c9fd9d3d13.jpg", + "image_caption": [ + "Figure 7. Average TPR of shortcut and non-shortcut classes given different thresholds on datasets with 1000 classes. Models generally perform better on images of shortcut classes than non-shortcut classes. This does not hold for results on IN-S, which lacks preserved texture information." + ], + "image_footnote": [], + "bbox": [ + 133, + 89, + 854, + 407 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/e7eb7c03bb47cb73c67ca58eec8e315e6fd629ae511eca91da914eafbf84387e.jpg", + "image_caption": [ + "Figure 8. Average TPR of shortcut and non-shortcut classes at different thresholds on datasets with 200 classes. Models perform better on shortcut classes than non-shortcut classes in IN-200. IN-R lacks preserved texture information and thus the models have worse performance on shortcut classes than non-shortcut classes." + ], + "image_footnote": [], + "bbox": [ + 132, + 465, + 433, + 772 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "gree of shortcut learning of CCT among the models (indicated by the marker size), inducing frequency reliance without blocking the learning of other semantic features appears", + "bbox": [ + 89, + 854, + 483, + 902 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "to benefit model robustness and generalization performance under statistical distribution shifts and corruption scenarios.", + "bbox": [ + 511, + 467, + 906, + 498 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Results on IN-S and IN-R. Different from the results on IN-v2 and IN-C, frequency shortcuts impair generalization of models to texture and rendition changes. Model performance on shortcut and non-shortcut classes on IN-S is close to each other. In Fig. 8, we observe that shortcut classes have worse prediction results than non-shortcut classes when tested on IN-R which contains images with rendition changes. This is attributable to the fact that shortcut information are not available in the OOD tests, as the rendition and sketch test sets preserve less or very different texture information than IN-C and IN-v2 (see green lines of IN-S and IN-1k, IN-R and IN-200 in Fig. 7 and Fig. 8). Frequency shortcut learning is an explicit cause of the texture-bias of ImageNet models, and the resulting impaired generalizability to renditions is in line with [8].", + "bbox": [ + 511, + 510, + 908, + 737 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Results on FGSM attacks. As shown in Fig. 7, the four models, under adversarial attacks, achieve higher AvgTPR for shortcut classes (especially those with strong shortcuts) than non-shortcut classes. This suggests that models can be inherently robust to adversarial noise if they leverage frequency shortcuts for classification. This is not surprising as adversarial noises hardly manipulate the textures of images. CCT, which shows the highest degree of frequency shortcut learning among the considered models, achieves the best adversarial robustness. This indicates that model reliance", + "bbox": [ + 511, + 750, + 910, + 900 + ], + "page_idx": 6 + }, + { + "type": "page_number", + "text": "25204", + "bbox": [ + 478, + 944, + 519, + 957 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/3d19b165873f3c16216ae597e66203f09ef63f5b145986049d1263da92b5650b.jpg", + "image_caption": [ + "Figure 9. An image of dugong contains frequency shortcuts which a model relies on for high-confidence classification. Its presence in OOD test sets like IN-C, and in the image under FGSM attacks yields correct predictions. But the reliance on shortcuts impairs model generalization performance to renditions changes." + ], + "image_footnote": [], + "bbox": [ + 107, + 97, + 472, + 282 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "on frequency shortcuts can benefit their adversarial robustness, as the simple shortcut features are robust to adversarial noise.", + "bbox": [ + 89, + 375, + 483, + 419 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Summary. Whether frequency shortcuts impair or benefit model generalization and robustness performance depends on the specific OOD scenarios. In Fig. 9, we show an example image of dugong and model predictions in OOD tests. The preserved texture information yields correct predictions on IN-C and under FGSM attacks, but it is not helpful for model generalizability to rendition changes. The findings reveal a limitation in the design of current OOD benchmarks, which overlooks the impact of frequency shortcuts on generalizability of models and its relation to specific characteristics of OOD data. Using HFSS for performance evaluation can bridge such gap.", + "bbox": [ + 88, + 429, + 482, + 609 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "4.3. Comparison with the existing approach", + "text_level": 1, + "bbox": [ + 89, + 621, + 429, + 638 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "More effective at finding shortcuts. We perform experiments on IN-10 for a direct comparison with the only existing method for frequency shortcut analysis [39], which evaluates the contribution of single frequencies to shortcuts and is limited to analyzing models trained on small-scale datasets. We report results in Tab. 2. HFSS uncovers strong frequency shortcuts for classes airliner, Siamese cat (Siam-cat), ox, frog, zebra, and container ship (Con-ship), while [39] was only able to find strong frequency shortcuts for classes zebra and Siamese cat. The better effectiveness of HFSS at finding shortcuts is attributable to the fact that HFSS considers the joint contributions of frequencies to classification, while in [39] the relevance of a single frequency was measured iteratively.", + "bbox": [ + 89, + 643, + 482, + 853 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Frequency shortcuts can impair the generalization performance of models or provide a false impression of good generalization when shortcuts exist in OOD tests [39]. We", + "bbox": [ + 89, + 854, + 483, + 900 + ], + "page_idx": 7 + }, + { + "type": "table", + "img_path": "images/d2fb9469e295ee5f909ac5df48118247ada223515a6dc3eb98049298ae78c966.jpg", + "table_caption": [ + "Table 2. TPR results on DFM-filtered IN-10 images. TPR $\\geq 0.6$ (a strong frequency shortcut) is highlighted in bold." + ], + "table_footnote": [], + "table_body": "
Class Methodairlinerwagonhum-birdSiam-catoxgolden retrieverfrogzebraCon-shiptruck
HFSS [39]0.8800.520.980.90.260.780.70.760.5
0.0800.40.80.020.020.140.80.540.06
", + "bbox": [ + 516, + 128, + 901, + 178 + ], + "page_idx": 7 + }, + { + "type": "table", + "img_path": "images/323a4ff4252b1196b230f316c661e07df565a1c5d8a3c634e1843e3d191a2dd9.jpg", + "table_caption": [ + "Table 3. TPR results on IN-SCT (first row) and corresponding DFM-filtered images (second and third rows). TPR above average TPR (0.367) is highlighted in bold." + ], + "table_footnote": [], + "table_body": "
Class MethodMil-aircraftcarlorikeettabby catholsteinLab-retrievertree froghorsefishing vesselfire truck
—0.3430.430.4430.2710.3290.38570.40.0290.4290.629
HFSS0.5140.0290.5860.8860.4570.2710.8710.2430.5430.071
[39]000.21430.12860.04290.02860.05710.12860.21430
", + "bbox": [ + 516, + 243, + 901, + 301 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "thus validate the impact of shortcuts searched by HFSS in generalization tests, by measuring performance results on IN-SCT, an OOD test set designed by [39]. We report the results in Tab. 3. The strong shortcuts for classes airliner, frog and container ship searched by HFSS are present in the OOD set, contributing to close-to or above average TPR of classes Military aircraft (Mil-aircraft), tree frog, and fishing vessel in the IN-SCT dataset. The shortcuts searched by [39] for class Siamese cat result in a much lower TPR for class tabby cat in IN-SCT, though shortcuts for this class still exist in the OOD data (as observed from the results of HFSS). This indicates that the method proposed in [39] has limitations in estimating the impact of shortcuts on OOD performance, and further demonstrates the effectiveness of HFSS in finding shortcuts.", + "bbox": [ + 511, + 321, + 906, + 547 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "5. Conclusions", + "text_level": 1, + "bbox": [ + 511, + 570, + 638, + 585 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "We proposed the first method for analysis of frequency shortcuts, HFSS, that enables the inspection of shortcut learning in large-scale models and datasets with thousands of classes. HFSS is more time-efficient and effective in finding frequency shortcuts compared to existing approaches. We investigate frequency shortcut learning in models trained on ImageNet and relate their results to robustness and generalization performance under different OOD conditions. The impact of frequency shortcuts on model generalization depends on the specific OOD scenarios. Existing models yield good performance on generalization benchmarks when texture information, corresponding to most frequency shortcuts, is mostly preserved in OOD data. Instead, frequency shortcuts impair the generalizability of models to images with rendition changes. This highlights the limitation of current OOD performance evaluation benchmarks, which need to explicitly take into account the impact that frequency shortcuts have on model performance. HFSS provides a tool to bridge this gap and extends the rigor of model generalization evaluation.", + "bbox": [ + 511, + 598, + 906, + 900 + ], + "page_idx": 7 + }, + { + "type": "page_number", + "text": "25205", + "bbox": [ + 478, + 944, + 517, + 955 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Acknowledgment", + "text_level": 1, + "bbox": [ + 91, + 90, + 240, + 107 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "Shunxin Wang and Nicola Strisciuglio are partly supported by the ERJU project. Europe's Rail Joint Undertaking is a European partnership on rail research and innovation established under the Horizon Europe program (2020-2027).", + "bbox": [ + 89, + 114, + 483, + 176 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "References", + "text_level": 1, + "bbox": [ + 91, + 186, + 187, + 204 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[1] Hyojin Bahng, Sanghyuk Chun, Sangdoo Yun, Jaegul Choo, and Seong Joon Oh. Learning de-biased representations with biased representations. In Proceedings of the 37th International Conference on Machine Learning, pages 528-539. PMLR, 2020. 2", + "[2] Christopher Boland, Sonia Dahdouh, Sotirios A. Tsaftaris, and Keith A Goatman. There are no shortcuts to anywhere worth going: Identifying shortcuts in deep learning models for medical image analysis. In Submitted to Medical Imaging with Deep Learning, 2024. under review. 2", + "[3] Yuan Cao, Zhiying Fang, Yue Wu, Ding-Xuan Zhou, and Quanquan Gu. Towards understanding the spectral bias of deep learning, 2020. 2", + "[4] Francesco Croce, Maksym Andriushchenko, Vikash Sehwag, Edoardo Debenedetti, Nicolas Flammarion, Mung Chiang, Prateek Mittal, and Matthias Hein. Robustbench: a standardized adversarial robustness benchmark. In Thirty-fifth Conference on Neural Information Processing Systems Datasets and Benchmarks Track (Round 2), 2021. 2", + "[5] Nikolay Dagaev, Brett D. Roads, Xiaoliang Luo, Daniel N. Barry, Kaustubh R. Patil, and Bradley C. Love. A too-good-to-be-true prior to reduce shortcut reliance, 2021. 1, 2", + "[6] Jia Deng, Wei Dong, Richard Socher, Li-Jia Li, Kai Li, and Li Fei-Fei. Imagenet: A large-scale hierarchical image database. In 2009 IEEE Conference on Computer Vision and Pattern Recognition, pages 248-255, 2009. 4", + "[7] Paul Gavrikov and Janis Keuper. Can biases in imagenet models explain generalization? In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 22184-22194, 2024. 1, 2", + "[8] Robert Geirhos, Patricia Rubisch, Claudio Michaelis, Matthias Bethge, Felix A. Wichmann, and Wieland Brendel. Imagenet-trained CNNs are biased towards texture; increasing shape bias improves accuracy and robustness. In International Conference on Learning Representations, 2019. 1, 2, 7", + "[9] Robert Geirhos, Jörn-Henrik Jacobsen, Claudio Michaelis, Richard Zemel, Wieland Brendel, Matthias Bethge, and Felix A. Wichmann. Shortcut learning in deep neural networks. Nature Machine Intelligence, 2(11):665-673, 2020. 1, 2", + "[10] Ian J. Goodfellow, Jonathon Shlens, and Christian Szegedy. Explaining and harnessing adversarial examples, 2015. 4, 5", + "[11] Jindong Gu and Daniela Oelke. Understanding bias in machine learning, 2019. 2", + "[12] Jun Guo, Wei Bao, Jiakai Wang, Yuqing Ma, Xinghai Gao, Gang Xiao, Aishan Liu, Jian Dong, Xianglong Liu, and Wenjun Wu. A comprehensive evaluation framework for deep model robustness. Pattern Recognition, 137:109308, 2023. 2" + ], + "bbox": [ + 93, + 213, + 483, + 898 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[13] Dan Hendrycks and Thomas Dietterich. Benchmarking neural network robustness to common corruptions and perturbations. In International Conference on Learning Representations, 2019. 2, 4, 5", + "[14] Dan Hendrycks, Steven Basart, Norman Mu, Saurav Kadavath, Frank Wang, Evan Dorundo, Rahul Desai, Tyler Zhu, Samyak Parajuli, Mike Guo, Dawn Song, Jacob Steinhardt, and Justin Gilmer. The many faces of robustness: A critical analysis of out-of-distribution generalization. In Proceedings of the IEEE/CVF International Conference on Computer Vision (ICCV), pages 8340-8349, 2021. 2, 4, 5", + "[15] Hanxun Huang, Xingjun Ma, Sarah Monazam Erfani, James Bailey, and Yisen Wang. Unlearnable examples: Making personal data unexploitable. In International Conference on Learning Representations, 2021. 5", + "[16] Zhi-Qin John Xu, Yaoyu Zhang, Tao Luo, Yanyang Xiao, and Zheng Ma. Frequency principle: Fourier analysis sheds light on deep neural networks. Communications in Computational Physics, 28(5):1746-1767, 2020. 2", + "[17] Oğuzhan Fatih Kar, Teresa Yeo, Andrei Atanov, and Amir Zamir. 3d common corruptions and data augmentation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 18963-18974, 2022. 2", + "[18] A. Krizhevsky. Learning multiple layers of features from tiny images, 1970. 5", + "[19] S. Lapuschkin, S. Wäldchen, A. Binder, et al. Unmasking Clever Hans predictors and assessing what machines really learn. Nat Commun 10, 1096, 2019. 1, 2", + "[20] Mathias Lecuyer, Vaggelis Atlidakis, Roxana Geambasu, Daniel Hsu, and Suman Jana. Certified robustness to adversarial examples with differential privacy. In 2019 IEEE Symposium on Security and Privacy (SP), pages 656-672, 2019. 2", + "[21] Sungbin Lim, Ildoo Kim, Taesup Kim, Chiheon Kim, and Sungwoong Kim. Fast autoaugment, 2019. 3", + "[22] Ze Liu, Yutong Lin, Yue Cao, Han Hu, Yixuan Wei, Zheng Zhang, Stephen Lin, and Baining Guo. Swin transformer: Hierarchical vision transformer using shifted windows, 2021. 3", + "[23] Emanuele Marconato, Stefano Teso, Antonio Vergari, and Andrea Passerini. Not all neuro-symbolic concepts are created equal: Analysis and mitigation of reasoning shortcuts. In Thirty-seventh Conference on Neural Information Processing Systems, 2023. 2", + "[24] Matthias Minderer, Olivier Bachem, Neil Houlsby, and Michael Tschannen. Automatic shortcut removal for self-supervised representation learning, 2020. 2", + "[25] Eric Mintun, Alexander Kirillov, and Saining Xie. On interaction between augmentations and corruptions in natural corruption robustness. In Advances in Neural Information Processing Systems, pages 3571-3583. Curran Associates, Inc., 2021. 2", + "[26] Samuel G. Müller and Frank Hutter. Trivialaugment: Tuning-free yet state-of-the-art data augmentation, 2021. 3", + "[27] Meike Nauta, Ricky Walsh, Adam Dubowski, and Christin Seifert. Uncovering and correcting shortcut learning in machine learning models for skin cancer diagnosis. Diagnostics, 12(1), 2022. 1, 2, 3" + ], + "bbox": [ + 514, + 92, + 903, + 898 + ], + "page_idx": 8 + }, + { + "type": "page_number", + "text": "25206", + "bbox": [ + 478, + 944, + 519, + 955 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[28] Hongjing Niu, Hanting Li, Feng Zhao, and Bin Li. Roadblocks for temporarily disabling shortcuts and learning new knowledge. In Advances in Neural Information Processing Systems, pages 29064-29075. Curran Associates, Inc., 2022. 2", + "[29] Ian Osband and Benjamin Van Roy. Why is posterior sampling better than optimism for reinforcement learning? In Proceedings of the 34th International Conference on Machine Learning, pages 2701-2710. PMLR, 2017. 3", + "[30] Mohammad Pezeshki, Sekou-Oumar Kaba, Yoshua Bengio, Aaron Courville, Doina Precup, and Guillaume Lajoie. Gradient starvation: A learning proclivity in neural networks. In Advances in Neural Information Processing Systems, 2021. 2", + "[31] Benjamin Recht, Rebecca Roelofs, Ludwig Schmidt, and Vaishaal Shankar. Do ImageNet classifiers generalize to ImageNet? In Proceedings of the 36th International Conference on Machine Learning, pages 5389-5400. PMLR, 2019. 2, 4, 5", + "[32] Joshua Robinson, Li Sun, Ke Yu, Kayhan Batmanghelich, Stefanie Jegelka, and Suvrit Sra. Can contrastive learning avoid shortcut solutions?, 2021. 2", + "[33] Ramprasaath R. Selvaraju, Michael Cogswell, Abhishek Das, Ramakrishna Vedantam, Devi Parikh, and Dhruv Batra. Grad-cam: Visual explanations from deep networks via gradient-based localization. In 2017 IEEE International Conference on Computer Vision (ICCV), pages 618-626, 2017. 2", + "[34] Harshay Shah, Kaustav Tamuly, Aditi Raghunathan, Prateek Jain, and Praneeth Netrapalli. The pitfalls of simplicity bias in neural networks. In Advances in Neural Information Processing Systems, pages 9573-9585. Curran Associates, Inc., 2020. 1, 2", + "[35] Damien Teney, Armand Mihai Nicoliciou, Valentin Hartmann, and Ehsan Abbasnejad. Neural redshift: Random networks are not random functions. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 4786-4796, 2024. 2", + "[36] Haohan Wang, Songwei Ge, Zachary Lipton, and Eric P Xing. Learning robust global representations by penalizing local predictive power. In Advances in Neural Information Processing Systems, pages 10506-10518, 2019. 2, 4, 5", + "[37] Shunxin Wang, Raymond Veldhuis, Christoph Brune, and Nicola Strisciuglio. Frequency shortcut learning in neural networks. In NeurIPS 2022 Workshop on Distribution Shifts: Connecting Methods and Applications, 2022. 1, 2", + "[38] Shunxin Wang, Christoph Brune, Raymond Veldhuis, and Nicola Strisciuglio. Dfm-x: Augmentation by leveraging prior knowledge of shortcut learning. In Proceedings of the IEEE/CVF International Conference on Computer Vision (ICCV) Workshops, pages 129-138, 2023. 2, 3", + "[39] Shunxin Wang, Raymond Veldhuis, Christoph Brune, and Nicola Strisciuglio. What do neural networks learn in image classification? a frequency shortcut perspective. In Proceedings of the IEEE/CVF International Conference on Computer Vision (ICCV), pages 1433-1442, 2023. 1, 2, 3, 4, 5, 8" + ], + "bbox": [ + 91, + 90, + 480, + 898 + ], + "page_idx": 9 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[40] Shunxin Wang, Raymond Veldhuis, Christoph Brune, and Nicola Strisciuglio. A survey on the robustness of computer vision models against common corruptions, 2023. 2", + "[41] Zhi-Qin John Xu and Hanxu Zhou. Deep frequency principle towards understanding why deeper learning is faster, 2020. 2", + "[42] Jiayun Zheng and Maggie Makar. Causally motivated multishutdown identification and removal. In Advances in Neural Information Processing Systems, pages 12800-12812. Curran Associates, Inc., 2022. 2", + "[43] Xin Zou and Weiwei Liu. On the adversarial robustness of out-of-distribution generalization models. In Advances in Neural Information Processing Systems, pages 68908-68938. Curran Associates, Inc., 2023. 2" + ], + "bbox": [ + 514, + 90, + 903, + 275 + ], + "page_idx": 9 + }, + { + "type": "page_number", + "text": "25207", + "bbox": [ + 478, + 945, + 517, + 955 + ], + "page_idx": 9 + } +] \ No newline at end of file diff --git a/2025/Do ImageNet-trained Models Learn Shortcuts_ The Impact of Frequency Shortcuts on Generalization/f684a2ee-5aa0-48af-b0b3-fb7e06614e21_model.json b/2025/Do ImageNet-trained Models Learn Shortcuts_ The Impact of Frequency Shortcuts on Generalization/f684a2ee-5aa0-48af-b0b3-fb7e06614e21_model.json new file mode 100644 index 0000000000000000000000000000000000000000..c0603d71833a92031c792876f9f19c0e5be0323f --- /dev/null +++ b/2025/Do ImageNet-trained Models Learn Shortcuts_ The Impact of Frequency Shortcuts on Generalization/f684a2ee-5aa0-48af-b0b3-fb7e06614e21_model.json @@ -0,0 +1,1716 @@ +[ + [ + { + "type": "header", + "bbox": [ + 0.107, + 0.003, + 0.182, + 0.043 + ], + "angle": 0, + "content": "CVF" + }, + { + "type": "header", + "bbox": [ + 0.237, + 0.0, + 0.812, + 0.047 + ], + "angle": 0, + "content": "This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore." + }, + { + "type": "title", + "bbox": [ + 0.235, + 0.131, + 0.764, + 0.177 + ], + "angle": 0, + "content": "Do ImageNet-trained models learn shortcuts? The impact of frequency shortcuts on generalization" + }, + { + "type": "text", + "bbox": [ + 0.265, + 0.204, + 0.733, + 0.24 + ], + "angle": 0, + "content": "Shunxin Wang Raymond Veldhuis Nicola Strisciuglio University of Twente" + }, + { + "type": "text", + "bbox": [ + 0.263, + 0.242, + 0.731, + 0.258 + ], + "angle": 0, + "content": "{s.wang-2,r.n.j.veldhuis,n.strisciuglio}@utwente.nl" + }, + { + "type": "title", + "bbox": [ + 0.248, + 0.292, + 0.328, + 0.308 + ], + "angle": 0, + "content": "Abstract" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.326, + 0.486, + 0.673 + ], + "angle": 0, + "content": "Frequency shortcuts refer to specific frequency patterns that models heavily rely on for correct classification. Previous studies have shown that models trained on small image datasets often exploit such shortcuts, potentially impairing their generalization performance. However, existing methods for identifying frequency shortcuts require expensive computations and become impractical for analyzing models trained on large datasets. In this work, we propose the first approach to more efficiently analyze frequency shortcuts at a large scale. We show that both CNN and transformer models learn frequency shortcuts on ImageNet. We also expose that frequency shortcut solutions can yield good performance on out-of-distribution (OOD) test sets which largely retain texture information. However, these shortcuts, mostly aligned with texture patterns, hinder model generalization on rendition-based OOD test sets. These observations suggest that current OOD evaluations often overlook the impact of frequency shortcuts on model generalization. Future benchmarks could thus benefit from explicitly assessing and accounting for these shortcuts to build models that generalize across a broader range of OOD scenarios. Codes are available at https://github.com/nis-research/hfss." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.708, + 0.222, + 0.723 + ], + "angle": 0, + "content": "1. Introduction" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.735, + 0.485, + 0.901 + ], + "angle": 0, + "content": "Superficial correlations between data and ground truth [5, 9] can be learned by models to minimize training objectives with the least effort [34]. This learning behavior is called shortcut learning, which either harms the generalization performance of models or gives an illusion of good generalization abilities when learned shortcuts are present in OOD test sets [39]. Artifacts and visual cues that cause shortcut learning are recognizable by visual inspection of images, and their impact on the training of models can be mitigated through data selection or augmentation [5, 19, 27], i.e. counteracting the spurious correlations between data" + }, + { + "type": "table_caption", + "bbox": [ + 0.513, + 0.291, + 0.908, + 0.36 + ], + "angle": 0, + "content": "Table 1. Our method needs less computation times compared to [39] (using ResNet18 on an NVIDIA A40 GPU). Computational time of [39] on ImageNet-1k is estimated from their ImageNet-10 experiments, considering that it increases proportionally to the number of classes." + }, + { + "type": "table", + "bbox": [ + 0.559, + 0.373, + 0.86, + 0.426 + ], + "angle": 0, + "content": "
DatasetTime (h) [39]Time (h) (Ours)
CIFAR-107.50.5
ImageNet-1k8500 (354 days)220 (9.2 days)
" + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.432, + 0.632, + 0.446 + ], + "angle": 0, + "content": "and ground truth." + }, + { + "type": "text", + "bbox": [ + 0.511, + 0.448, + 0.907, + 0.825 + ], + "angle": 0, + "content": "However, there exist shortcuts in the Fourier domain, which are implicitly embedded in image data characteristics and not easily detectable by visual inspection [37, 39]. Such shortcut solutions consist of small frequency subsets that are easy-to-learn and sufficient for models to achieve high classification rate. Typically, they correspond to simple features like textures, shapes or colors in the spatial domain. Wang, et al. [39] identified frequency shortcuts by retaining relevant frequencies to classification of a certain class. The relevance of an individual frequency was measured by the loss value of the model tested on images of the class with that frequency removed. Thus, their approach requires computational time increasing proportionally to the number of classes in a dataset and the image resolution. Yet, frequency shortcuts usually consist of multiple frequencies. The method in [39] might overlook some shortcuts as individual frequency relevance to classification neglects joint contribution of frequencies. Furthermore, the computational burden limits its applicability for analyzing shortcut learning behavior of models trained on large datasets (e.g. ImageNet) with hundreds or thousands of classes. Several studies have shown that ImageNet-trained models are biased towards textures [7, 8]. However, there is no concrete evidence of what causes this phenomenon, although [39] largely attributes it to frequency shortcut learning." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.826, + 0.909, + 0.903 + ], + "angle": 0, + "content": "In this work, we propose the first method that enables the uncovering of frequency shortcuts learned by models trained on large-scale datasets (e.g. ImageNet-1k), for analyzing learning behavior and explaining model generalization performance in different OOD scenarios. Our" + }, + { + "type": "page_number", + "bbox": [ + 0.479, + 0.945, + 0.521, + 0.958 + ], + "angle": 0, + "content": "25198" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.09, + 0.092, + 0.482, + 0.182 + ], + "angle": 0, + "content": "method improves computational efficiency as it applies parallel class-wise loss computation and hierarchical search in the Fourier domain (see Tab. 1), compared to [39]. Moreover, our method considers the joint contribution of frequencies to classification, thus being more effective in finding shortcuts. Our contributions are:" + }, + { + "type": "text", + "bbox": [ + 0.092, + 0.183, + 0.483, + 0.258 + ], + "angle": 0, + "content": "1. We develop a hierarchical frequency shortcut search (HFSS) method, which enables the analysis of shortcut learning in the Fourier domain on large datasets with varying class counts. We reduce computational time and improve the effectiveness at identifying shortcuts." + }, + { + "type": "text", + "bbox": [ + 0.091, + 0.258, + 0.483, + 0.363 + ], + "angle": 0, + "content": "2. We discover that ImageNet-trained models (both CNN and transformer architectures) are subject to learn frequency shortcuts. Different from shortcuts formed by visual cues [9], frequency shortcuts (easy-to-learn features) lead to good performance on both in-distribution (ID) and OOD tests if they do not block models from learning other semantics (difficult features)." + }, + { + "type": "text", + "bbox": [ + 0.091, + 0.364, + 0.483, + 0.484 + ], + "angle": 0, + "content": "3. Our HFSS enables a more comprehensive assessment of model generalizability by analyzing OOD data characteristics, specifically the presence of shortcuts. In existing evaluation frameworks, frequency shortcuts do not always impair OOD generalization performance. This emphasizes the importance of considering the role of shortcuts when designing future OOD evaluation benchmarks." + }, + { + "type": "list", + "bbox": [ + 0.091, + 0.183, + 0.483, + 0.484 + ], + "angle": 0, + "content": null + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.498, + 0.236, + 0.513 + ], + "angle": 0, + "content": "2. Related works" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.524, + 0.483, + 0.825 + ], + "angle": 0, + "content": "Shortcut learning. Models can learn shortcut solutions based on superficial correlations between data and ground truth [9, 11] to optimize training objectives with the least effort. This learning behavior is due to simplicity-bias [34], which is caused by inductive biases provided by gradient descent or components like ReLUs [35]. Shortcuts can be visual cues in the data like source tags and artificial markers [19, 27]. For instance, one fifth of the horse images from Pascal VOC dataset were found to contain a source tag, on which models rely as a discriminant feature to recognize horses [19]. Next to these visual cues, vision models are also subject to shortcuts implicitly existing in the frequency domain, which manifest as small sets of frequencies contributing significantly to image classification performance [37, 39]. The study in [39] explains why models exhibit a textures-bias for classification [8] from a frequency shortcut perspective, but does not provide further investigation on how frequency shortcuts affect the generalization and robustness performance of ImageNet-trained models, due to computational burden." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.826, + 0.483, + 0.901 + ], + "angle": 0, + "content": "Relying on simple shortcut solutions could harm the generalization and robustness performance of models, as shortcut learning appearing in early training might block models from learning other semantics related to the tasks at hand [3, 16, 39, 41]. However, such reliance could also cre" + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.092, + 0.905, + 0.122 + ], + "angle": 0, + "content": "ates a false impression of good generalization when shortcuts are present in OOD test sets [1, 19, 39]." + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.141, + 0.905, + 0.579 + ], + "angle": 0, + "content": "Shortcut identification. Uncovering shortcuts helps understanding the generalizability of models on different out-of-distribution (OOD) data, explaining why models fail to generalize to or perform well on OOD data. It can provide shortcut-related prior knowledge to develop techniques that improve model generalization and robustness performance based on shortcut mitigation [27, 38]. However, common approaches are limited to identifying visually inspectable shortcuts such as text, watermark and color patches [19, 24, 27], using e.g. saliency maps [33]. Rather than directly identifying shortcuts in the data, in [2, 5] shortcut features present in individual images were quantified by assessing how difficult they were for models to learn. The authors in [28, 30, 32] uncovered shortcut features learned within representation space. The authors in [23] investigated shortcut learning by analyzing the relationship between semantic concepts using a knowledge graph, sharing a similar idea to [42]. Despite visual shortcut cues, there are non-observable shortcuts in the frequency domain, which are embedded in data characteristics. The work of [39] identified relevant frequencies to classification, which potentially contain shortcut information. As this technique measures the relevance of one frequency at a time, it requires computational costs increasing proportionally to the number of classes and image resolution. Thus, limited attention has been given to uncovering frequency shortcuts in large datasets, with existing analyses primarily focused on datasets containing a small number of classes [2, 37, 39, 42]." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.58, + 0.906, + 0.716 + ], + "angle": 0, + "content": "Existing methods to identify shortcuts are limited to either manual inspection [19] or involve a time-consuming algorithm [37, 39]. Methods quantifying shortcut information or uncovering shortcut features primarily aim at mitigating shortcut learning, without explaining how shortcuts in the data impact model generalization performance. We are the first to enable frequency shortcut analysis of models trained on large-scale datasets and link them to generalization performance in different OOD settings." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.735, + 0.907, + 0.903 + ], + "angle": 0, + "content": "OOD evaluation. The generalization and robustness performance of vision models are usually evaluated using extra data that is considered OOD, e.g. data collected at different time points [31], with different styles or renditions [8, 14, 36], with synthetic corruption effects [13, 17, 25, 40], and with adversarial noise [4, 12, 20, 43]. However, such benchmarks do not consider the impact of shortcuts learned by models and present in OOD test sets. This might ignore critical factors related to the generalization and robustness capabilities of models. The work in [7] explores model generalization from the perspectives of biases, e.g. texture," + }, + { + "type": "page_number", + "bbox": [ + 0.48, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "25199" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.14, + 0.096, + 0.852, + 0.355 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.089, + 0.364, + 0.908, + 0.395 + ], + "angle": 0, + "content": "Figure 1. Scheme of HFSS. Starting from stage 2, we sample frequency patches from a random frequency subset searched in previous stage. This confines the size of search space. The white patches in the binary masks indicate sampled frequency patches." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.407, + 0.486, + 0.498 + ], + "angle": 0, + "content": "shape and spectral biases. In this work, we investigate the impact of frequency shortcuts on model generalization abilities, establishing connections among these different biases. Our work provides insights into when frequency shortcuts yield good performance or are harmful to model generalization and robustness performance." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.51, + 0.183, + 0.525 + ], + "angle": 0, + "content": "3. Method" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.536, + 0.484, + 0.673 + ], + "angle": 0, + "content": "We propose a method to identify (non-visible) frequency shortcuts linked to intrinsic data characteristics rather than visual cues as in [27]. We reduce significantly computational demands compared to [39]. Our method is the first solution for shortcut identification that enables the analysis of models trained on datasets with a large number of samples and classes. In the following sections, we detail our measurement of shortcut learning in models and examine the impact of shortcuts on generalization." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.68, + 0.37, + 0.696 + ], + "angle": 0, + "content": "3.1. Identifying frequency shortcuts" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.702, + 0.484, + 0.839 + ], + "angle": 0, + "content": "We propose a method called hierarchical frequency shortcut search (HFSS), which exploits hierarchical search of frequency subsets in the image Fourier spectrum. This reduces computational time compared to exhaustive search strategies, e.g. [38, 39]. The search is separated into several stages, each stage gradually narrowing down the spectrum search space in a coarse-to-fine manner and detecting frequency subsets that models strongly rely on for classification. The scheme of HFSS is shown in Fig. 1." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.856, + 0.484, + 0.901 + ], + "angle": 0, + "content": "Hierarchical search for frequency shortcuts. In order to discover the frequency subsets that models rely heavily on for correct classification of each class, we sample different" + }, + { + "type": "image", + "bbox": [ + 0.613, + 0.406, + 0.791, + 0.533 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.513, + 0.536, + 0.907, + 0.565 + ], + "angle": 0, + "content": "Figure 2. The frequency spectrum is separated into patches, with \\( p\\% \\) sampled for shortcut evaluation." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.569, + 0.907, + 0.689 + ], + "angle": 0, + "content": "frequency combinations that potentially contain frequency shortcuts. We use random sampling as it allows to easily consider joint contribution of frequencies to classification, improving the effectiveness of frequency shortcut identification. Random sampling has been shown to contribute to stable search results in data augmentation and reinforcement learning [21, 26, 29], while requiring lower computations w.r.t. optimization-based methods." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.69, + 0.909, + 0.902 + ], + "angle": 0, + "content": "There are numerous combinations of frequencies by sampling frequency components directly. Instead of doing an exhaustive search, we apply hierarchical search in the Fourier spectrum: we divide the Fourier spectrum into frequency patches and sample these frequency patches to generate candidates of frequency subsets that contain shortcut information, as illustrated in Fig. 2. We exploit overlapping shifted windows in vertical and horizontal directions [22] to prepare frequency patches and thus avoid border effects, in addition to evenly separating the Fourier spectrum. HFSS consists of multiple search stages, the number of which depends on image resolution and desired frequency resolution of the identified frequency subsets. Lower image or frequency resolution requires fewer search stages." + }, + { + "type": "page_number", + "bbox": [ + 0.48, + 0.945, + 0.52, + 0.958 + ], + "angle": 0, + "content": "25200" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.09, + 0.092, + 0.482, + 0.318 + ], + "angle": 0, + "content": "In summary, the overall process of HFSS is as follows. Each stage of HFSS consists of three steps, namely (1) sampling \\(B\\) frequency subsets which contain \\(p\\%\\) frequency patches, (2) evaluating shortcut information, and (3) forwarding subsets for next stage. At the initial search stage, we sample frequency patches with a large size and evaluate their shortcut information. We evaluate the frequency subsets according to their contributions to classification: we process a subset of training images by retaining a frequency subset and measure the class-wise loss of a model. Frequency subsets that contribute to a lower loss value are stronger candidates to indicate frequency shortcuts as they are sufficient for the model to achieve a high prediction score. We rank frequency subsets by the loss values and use the top-\\(N\\) masks of each class in the next search stage." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.319, + 0.482, + 0.44 + ], + "angle": 0, + "content": "Starting from the second search stage, we sample frequency patches from one of the top- \\(N\\) frequency subsets in the previous stage, with a smaller patch size. This confines the search space. In the final search stage, the top-1 frequency subset of each class is considered dominating the classification of the class concerned. Same as [39], we use binary masks to represent these frequency subsets, called Dominant Frequency Maps (DFMs)." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.441, + 0.482, + 0.502 + ], + "angle": 0, + "content": "The sampling percentage \\( p\\% \\) of frequency patches, the number of sampled frequency subsets \\( B \\) and the frequency resolution of DFMs are hyperparameters of the search algorithm (details are in the supplementary material)." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.514, + 0.45, + 0.531 + ], + "angle": 0, + "content": "3.2. Measuring the degree of shortcut learning" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.538, + 0.483, + 0.734 + ], + "angle": 0, + "content": "Analyzing models learning behavior in a class-wise manner, as done in [39], is labor-intensive if there are hundreds or thousands of classes. To provide a broad overview of model learning behavior on training data, we categorize all classes in a dataset into two groups: (1) classes subject to shortcuts and (2) non-shortcut classes. A class is considered subject to shortcuts if its true positive rate (TPR) surpasses a given threshold when the model is tested on images filtered to retain only dominant frequencies of the class concerned. We use TPR, as a high TPR indicates that the frequency subset is sufficient to achieve a high classification rate, thereby acting as a shortcut. The frequency-filtered images are referred to as DFM-filtered images." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.735, + 0.483, + 0.904 + ], + "angle": 0, + "content": "For each class, we test the model on both the original test images and DFM-filtered images, computing TPR for both. We denote the TPR on the original images of class \\( c_{i} \\) as \\( \\mathrm{TPR}_{c_i} \\) and the TPR on the DFM-filtered images as \\( \\mathrm{TPR}_{c_i}^{DFM} \\). If \\( \\mathrm{TPR}_{c_i}^{DFM} > t \\) (where \\( t \\in [0,1] \\) is a predefined threshold), class \\( c_{i} \\) is considered subject to shortcuts; otherwise, it is considered as one of the non-shortcut classes. We compute the average TPR values of both groups, shortcut and non-shortcut classes, at different thresholds \\( t \\). We note them as \\( \\mathrm{AvgTPR}_{\\mathrm{sct}}@t \\), \\( \\mathrm{AvgTPR}_{\\mathrm{sct}}^{\\mathrm{DFM}}@t \\) for the shortcut classes, and \\( \\mathrm{AvgTPR}_{\\mathrm{non-sct}}@t \\) and \\( \\mathrm{AvgTPR}_{\\mathrm{non-sct}}^{\\mathrm{DFM}}@t \\)" + }, + { + "type": "image", + "bbox": [ + 0.584, + 0.093, + 0.835, + 0.194 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.512, + 0.197, + 0.905, + 0.309 + ], + "angle": 0, + "content": "Figure 3. ResNet18 tested on original test images and DFM-filtered images of ImageNet-1k. Blue line shows the average TPR of classes subject to shortcuts on the original test images, and the orange line shows the results of non-shortcut classes, at different threshold \\( t \\). The green and red lines correspond to results tested on DFM-filtered images. Lower \\( t \\) indicates weak shortcuts and higher \\( t \\) signifies stronger ones. The size of each point reflects the number of classes. The larger the size, the more classes included." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.323, + 0.905, + 0.444 + ], + "angle": 0, + "content": "for the non-shortcut classes. A higher threshold \\( t \\) corresponds to stronger shortcuts, as dominant frequencies alone are sufficient for accurate classification. An example of the results of these metrics at different threshold values are given in Fig. 3, where the size of each point reflects the number of classes in the two groups. Larger points indicate a higher number of classes, which suggests that the model has a stronger tendency to shortcut learning." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.444, + 0.905, + 0.521 + ], + "angle": 0, + "content": "We do not average \\(\\mathrm{TPR}^{\\mathrm{DFM}}\\) across all classes because its value for many non-shortcut classes is close to zero (specifically for large datasets). This can result in a close-to-zero \\(\\mathrm{AvgTPR}^{\\mathrm{DFM}}\\) which does not provide useful information for further analysis." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.532, + 0.83, + 0.547 + ], + "angle": 0, + "content": "3.3. Impact of shortcut on generalization" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.554, + 0.907, + 0.902 + ], + "angle": 0, + "content": "To assess how frequency shortcuts impact the generalization and robustness performance of models trained on ImageNet-1k, we evaluate their performance on several datasets. We perform ID tests on ImageNet-1k (IN-1k) [6] and ImageNet-v2 (IN-v2) [31] test sets. They are considered in-distribution as they are collected using the same protocol as the training set. Furthermore, we use ImageNet-C (IN-C) [13] to benchmark the robustness of models against appearance image corruptions, e.g. noise, blur and weather changes. To evaluate the generalizability of models to images with different renditions, we use ImageNet-Renditions (IN-R) [14] and ImageNet-Sketch (IN-S) [36]. IN-R contains renditions like cartoons, paintings, art and toys of 200 classes from IN-1k. IN-S has the same number of classes as IN-1k, containing images of hand-drawn sketches. These datasets contain visual renditions, which serve to evaluate how the reliance on texture cues (corresponding to most frequency shortcuts) impacts on generalization. To measure the adversarial robustness of models, we apply the fast gradient sign method (FGSM) attacks [10] to the validation set of IN-1k, with \\( L_{\\infty} = 4 / 255 \\). We calculate the average TPR of shortcut and non-shortcut classes on OOD data, that we compare to the results on ID data." + }, + { + "type": "page_number", + "bbox": [ + 0.48, + 0.945, + 0.518, + 0.957 + ], + "angle": 0, + "content": "25201" + } + ], + [ + { + "type": "title", + "bbox": [ + 0.091, + 0.09, + 0.226, + 0.108 + ], + "angle": 0, + "content": "4. Experiments" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.116, + 0.485, + 0.224 + ], + "angle": 0, + "content": "We start with investigating the trade-off between efficiency (i.e. required computational time) and effectiveness (i.e. capability of finding shortcuts), varying the configuration of the number of sampled frequency subsets at each stage, noted as \\( B_{s} \\) (\\( s \\) is the index of stage). Then, we analyze frequency shortcut learning of models trained on IN-1k, relating it to performance results under different OOD scenarios." + }, + { + "type": "text", + "bbox": [ + 0.093, + 0.246, + 0.485, + 0.61 + ], + "angle": 0, + "content": "Setup. We use CIFAR-10 (C-10) [18] to explore the configuration of \\( B \\), and configure HFSS with four search stages, with patch size of \\( 8 \\times 8 \\), \\( 4 \\times 4 \\), \\( 2 \\times 2 \\) and \\( 1 \\times 1 \\), respectively. We then perform larger-scale shortcut analysis using IN-1k: we use six search stages, with patch size of \\( 56 \\times 56 \\), \\( 28 \\times 28 \\), \\( 14 \\times 14 \\), \\( 8 \\times 8 \\), \\( 4 \\times 4 \\) and \\( 2 \\times 2 \\). The patch (frequency) resolution in the final stage is the same as that in [39]. At each stage, we sample \\( p = 60\\% \\) frequency patches to form frequency subsets. It results in DFMs containing about \\( 5\\% \\) frequencies of the whole spectrum of ImageNet images and about \\( 15\\% \\) frequencies of CIFAR images. The number of training images sampled for shortcut information evaluation is the same as that of their corresponding test set. We evaluate the generalization and robustness performance of ImageNet-trained models on IN-v2 [31], IN-C [13], IN-R [14], IN-S [36] and using FGSM [10] attacks. We also carry out experiments on ImageNet-10 (IN-10) [15] with a ResNet18 model to compare HFSS with single-frequency removal-based method in [39]. Following their setups, we use ImageNet-SCT (IN-SCT) for OOD evaluation. Our configuration of HFSS enables direct comparison with the results reported in [39] as DFMs contain around \\( 5\\% \\) frequencies of whole spectrum. We report training configurations and additional results in the supplementary material." + }, + { + "type": "title", + "bbox": [ + 0.09, + 0.621, + 0.307, + 0.637 + ], + "angle": 0, + "content": "4.1. Configuration of HFSS" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.644, + 0.484, + 0.902 + ], + "angle": 0, + "content": "Sampled frequency subsets. At stage \\(s\\), we sample \\(B_{s}\\) frequency subsets, such that HFSS can explore different combinations of frequency patterns, validating their contained shortcut information. However, sampling an excessive number of subsets increases the required search time, as more subsets are for evaluating their contributions to classification results. To explore the impact of the number of sampling operation \\(B_{s}\\) on shortcut identification, we perform experiments on C-10 using a ResNet18 backbone, setting the number of sampled frequency subsets in each stage to be as large as possible under limited computational time. We hypothesize that HFSS obtains stable results given enough candidate frequency subsets. We set \\(B_{1}, B_{2}, B_{3}\\), and \\(B_{4}\\) to be 1000, 2000, 4000 and 8000 respectively, which we note as configuration CF-1. The increasing number of candidate frequency subsets at consecutive stages is determined by the increase in possible frequency patch combinations as patch" + }, + { + "type": "image", + "bbox": [ + 0.522, + 0.092, + 0.902, + 0.374 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.513, + 0.379, + 0.907, + 0.408 + ], + "angle": 0, + "content": "Figure 4. The best class-wise loss vs. the number of sampled frequency subsets at each stage." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.415, + 0.615, + 0.428 + ], + "angle": 0, + "content": "size decreases." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.43, + 0.907, + 0.809 + ], + "angle": 0, + "content": "We run the configuration CF-1 five times and note the stability of our HFSS search algorithms. We track the best (lowest) loss values achieved for each class when the model is tested on DFM-filtered images. In Fig. 4, we report the tracking statistics with average loss values (lines) and standard deviations (shadows) over the five trials for each class, which are further used as a starting point to optimize the configuration of HFSS. As the number of sampled frequency subsets increases, the best loss for each class converges, indicating the stability of HFSS in identifying similar shortcuts learned by models when a sufficient amount of candidate frequency subsets are sampled for validating shortcut information. The standard deviation recorded at different stages is relatively small compared to the range of best loss. For class airplane we observe an outlier outcome at stage 3. In the supplementary material, we present visualization of shortcut cues (DFM-filtered images) computed across the five trials. These visualizations show that the identified shortcut patterns differ in the orientations of strip-like features, corresponding to line features in the image of airplane. This suggests that frequency shortcuts do not arise from fixed frequency subsets; rather, spatial features they correspond to may be visually similar but composed of different frequencies, determining an higher standard deviation in Fig. 4." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.826, + 0.909, + 0.903 + ], + "angle": 0, + "content": "Efficiency vs. effectiveness. Based on the results in Fig. 4, we optimize CF-1 for efficiency and effectiveness in shortcut identification. The required search time scales up proportionally to \\( B_{s} \\), i.e. the more candidate frequency subsets are sampled, the more time is needed to verify their" + }, + { + "type": "page_number", + "bbox": [ + 0.48, + 0.945, + 0.52, + 0.958 + ], + "angle": 0, + "content": "25202" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.176, + 0.092, + 0.396, + 0.216 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.09, + 0.218, + 0.483, + 0.258 + ], + "angle": 0, + "content": "Figure 5. Average TPR vs. search time, where search time increases proportionally with the number of sampled frequency subsets." + }, + { + "type": "image", + "bbox": [ + 0.18, + 0.261, + 0.391, + 0.384 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.09, + 0.387, + 0.483, + 0.442 + ], + "angle": 0, + "content": "Figure 6. The number of shortcut classes given different threshold \\( t \\) search by CF-1 and CF-2.10. Using CF-2.10 uncovers most shortcuts identified by using CF-1 at low thresholds, with computational time reduced to a factor of around 200." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.46, + 0.483, + 0.61 + ], + "angle": 0, + "content": "relevance to classification. We perform experiments using different \\( B_{s} \\), investigating how the size of the search space affects the identification of shortcuts. Based on the tracking statistics in Fig. 4, which shows that increasing \\( B_{s} \\) beyond a certain point does not decrease loss significantly, we adjust \\( B_{1} \\) to 200, \\( B_{2} \\) to 800 and \\( B_{4} \\) to 4000 (CF-2.1). We further reduce \\( B_{3} \\) and \\( B_{4} \\) to 2000 (CF-2.2). We also design eight other configurations (see details in the supplementary material), with CF-2.10 being the most efficient as it samples the lowest number of candidate frequency subsets." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.612, + 0.483, + 0.853 + ], + "angle": 0, + "content": "To compare the performance of HFSS under different configurations, we compute the average TPR over all classes when tested on DFM-filtered images. This indicates the general relevance of the searched frequency subsets to classification. A higher average TPR indicate higher relevance, thus the frequency subsets contain more frequency shortcut information. We run each configuration five times, computing the mean and standard deviations of average TPRs over the five trials. We report the results in Fig. 5 where the left-most point corresponds to average TPR of CF-2.10 and the right-most point corresponds to that of CF-1. We observe that as more frequency subsets are sampled, average TPR increases and saturates. This demonstrates that sampling more candidate frequency subsets does not significantly improve the performance of HFSS, while requiring more computations." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.856, + 0.483, + 0.9 + ], + "angle": 0, + "content": "We compare CF-1 and CF-2.10 in Fig. 6 based on the number of classes with identified shortcuts at different thresholds. We observe that CF-2.10 uncovers most of the" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.092, + 0.905, + 0.197 + ], + "angle": 0, + "content": "shortcuts identified by using the more complex CF-1 at low threshold level. Although CF-2.10 misses stronger shortcuts (at higher threshold levels), it achieves a \\(\\sim 200\\times\\) reduction in computational time compared to CF-1. A more time-efficient configuration is less effective at finding shortcuts than the more complex one, but it still manages to identify shortcuts at low thresholds." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.198, + 0.905, + 0.334 + ], + "angle": 0, + "content": "The observations made on C-10 allow to estimate initial configurations for analysis on larger-scale datasets. We track the loss statistics of HFSS applied to ImageNet (provided in the supplementary material) and choose the value \\( B_{s} \\) for ImageNet experiments in Sec. 4.2 as the one where the loss values show no significant decrease. Given the computational intensity of identifying shortcuts in large-scale datasets, our work on IN-1k primarily aims at uncovering shortcuts with affordable computational efforts." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.345, + 0.796, + 0.359 + ], + "angle": 0, + "content": "4.2. Shortcut identification on IN-1k" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.367, + 0.905, + 0.443 + ], + "angle": 0, + "content": "We apply HFSS to examine frequency shortcuts learned by ImageNet models and how they impact model generalization and robustness. The search configurations for ImageNet models, aimed at time-efficiency and effectiveness at finding shortcuts, are in the supplementary material." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.463, + 0.905, + 0.901 + ], + "angle": 0, + "content": "Results on IN-1k, IN-v2 and IN-C. In Figs. 7 and 8, we report the results of the analysis of impact of shortcuts on model performance. In general, models perform better predictions on images of classes affected by shortcut learning (blue lines) compared to non-shortcut classes (orange lines) in both ID and corruption tests. Models subject to frequency shortcuts yield good performance in the robustness and generalization tests when textures information is largely preserved. For ResNet and ViT models, when \\( t \\geq 0.8 \\), the AvgTPR\\(_{\\text{sct}}^{\\text{DFM}}\\) exceeds AvgTPR\\(_{\\text{sct}}\\), meaning that these models rely predominantly on dominant frequencies (corresponding to simple features) for classification of shortcut-affected classes, preventing from learning class-related semantics features. For instance, at \\( t = 0.9 \\), only one class ('window screen') is found subject to frequency shortcuts: ResNet18 achieves TPR=0.66 on full-spectrum images of this class and TPR=0.94 on DFM-filtered images (similarly to ResNet50 and ViT-b). Differently, CCT exhibits lower AvgTPR\\(_{\\text{sct}}^{\\text{DFM}}@0.9\\) than AvgTPR\\(_{\\text{sct}}@0.9\\), indicating that while frequency shortcuts contribute significantly to classification, this model also manage to leverage other semantic information: a strong frequency shortcut (e.g. when \\( t > 0.7 \\)) does not necessarily mean that it is the only information a model uses for classification. We further compare CCT with ResNet50 as they perform similarly on IN-1k (80.57% and 80.1% respectively). In IN-v2 and IN-C, CCT achieves 74.81% and 57.73% prediction rates, having better performance than ResNet50 (74.17% in IN-v2 and 48.85% in IN-C). Considering the highest de" + }, + { + "type": "page_number", + "bbox": [ + 0.48, + 0.945, + 0.518, + 0.957 + ], + "angle": 0, + "content": "25203" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.134, + 0.09, + 0.856, + 0.409 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.089, + 0.412, + 0.908, + 0.455 + ], + "angle": 0, + "content": "Figure 7. Average TPR of shortcut and non-shortcut classes given different thresholds on datasets with 1000 classes. Models generally perform better on images of shortcut classes than non-shortcut classes. This does not hold for results on IN-S, which lacks preserved texture information." + }, + { + "type": "image", + "bbox": [ + 0.133, + 0.467, + 0.434, + 0.773 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.09, + 0.778, + 0.486, + 0.849 + ], + "angle": 0, + "content": "Figure 8. Average TPR of shortcut and non-shortcut classes at different thresholds on datasets with 200 classes. Models perform better on shortcut classes than non-shortcut classes in IN-200. IN-R lacks preserved texture information and thus the models have worse performance on shortcut classes than non-shortcut classes." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.856, + 0.485, + 0.903 + ], + "angle": 0, + "content": "gree of shortcut learning of CCT among the models (indicated by the marker size), inducing frequency reliance without blocking the learning of other semantic features appears" + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.468, + 0.907, + 0.499 + ], + "angle": 0, + "content": "to benefit model robustness and generalization performance under statistical distribution shifts and corruption scenarios." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.511, + 0.909, + 0.738 + ], + "angle": 0, + "content": "Results on IN-S and IN-R. Different from the results on IN-v2 and IN-C, frequency shortcuts impair generalization of models to texture and rendition changes. Model performance on shortcut and non-shortcut classes on IN-S is close to each other. In Fig. 8, we observe that shortcut classes have worse prediction results than non-shortcut classes when tested on IN-R which contains images with rendition changes. This is attributable to the fact that shortcut information are not available in the OOD tests, as the rendition and sketch test sets preserve less or very different texture information than IN-C and IN-v2 (see green lines of IN-S and IN-1k, IN-R and IN-200 in Fig. 7 and Fig. 8). Frequency shortcut learning is an explicit cause of the texture-bias of ImageNet models, and the resulting impaired generalizability to renditions is in line with [8]." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.75, + 0.911, + 0.901 + ], + "angle": 0, + "content": "Results on FGSM attacks. As shown in Fig. 7, the four models, under adversarial attacks, achieve higher AvgTPR for shortcut classes (especially those with strong shortcuts) than non-shortcut classes. This suggests that models can be inherently robust to adversarial noise if they leverage frequency shortcuts for classification. This is not surprising as adversarial noises hardly manipulate the textures of images. CCT, which shows the highest degree of frequency shortcut learning among the considered models, achieves the best adversarial robustness. This indicates that model reliance" + }, + { + "type": "page_number", + "bbox": [ + 0.479, + 0.945, + 0.521, + 0.958 + ], + "angle": 0, + "content": "25204" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.108, + 0.098, + 0.473, + 0.284 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.09, + 0.291, + 0.483, + 0.362 + ], + "angle": 0, + "content": "Figure 9. An image of dugong contains frequency shortcuts which a model relies on for high-confidence classification. Its presence in OOD test sets like IN-C, and in the image under FGSM attacks yields correct predictions. But the reliance on shortcuts impairs model generalization performance to renditions changes." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.376, + 0.484, + 0.42 + ], + "angle": 0, + "content": "on frequency shortcuts can benefit their adversarial robustness, as the simple shortcut features are robust to adversarial noise." + }, + { + "type": "text", + "bbox": [ + 0.089, + 0.43, + 0.483, + 0.611 + ], + "angle": 0, + "content": "Summary. Whether frequency shortcuts impair or benefit model generalization and robustness performance depends on the specific OOD scenarios. In Fig. 9, we show an example image of dugong and model predictions in OOD tests. The preserved texture information yields correct predictions on IN-C and under FGSM attacks, but it is not helpful for model generalizability to rendition changes. The findings reveal a limitation in the design of current OOD benchmarks, which overlooks the impact of frequency shortcuts on generalizability of models and its relation to specific characteristics of OOD data. Using HFSS for performance evaluation can bridge such gap." + }, + { + "type": "title", + "bbox": [ + 0.09, + 0.622, + 0.43, + 0.639 + ], + "angle": 0, + "content": "4.3. Comparison with the existing approach" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.644, + 0.483, + 0.854 + ], + "angle": 0, + "content": "More effective at finding shortcuts. We perform experiments on IN-10 for a direct comparison with the only existing method for frequency shortcut analysis [39], which evaluates the contribution of single frequencies to shortcuts and is limited to analyzing models trained on small-scale datasets. We report results in Tab. 2. HFSS uncovers strong frequency shortcuts for classes airliner, Siamese cat (Siam-cat), ox, frog, zebra, and container ship (Con-ship), while [39] was only able to find strong frequency shortcuts for classes zebra and Siamese cat. The better effectiveness of HFSS at finding shortcuts is attributable to the fact that HFSS considers the joint contributions of frequencies to classification, while in [39] the relevance of a single frequency was measured iteratively." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.856, + 0.484, + 0.901 + ], + "angle": 0, + "content": "Frequency shortcuts can impair the generalization performance of models or provide a false impression of good generalization when shortcuts exist in OOD tests [39]. We" + }, + { + "type": "table_caption", + "bbox": [ + 0.513, + 0.089, + 0.907, + 0.118 + ], + "angle": 0, + "content": "Table 2. TPR results on DFM-filtered IN-10 images. TPR \\(\\geq 0.6\\) (a strong frequency shortcut) is highlighted in bold." + }, + { + "type": "table", + "bbox": [ + 0.517, + 0.129, + 0.903, + 0.179 + ], + "angle": 0, + "content": "
Class Methodairlinerwagonhum-birdSiam-catoxgolden retrieverfrogzebraCon-shiptruck
HFSS [39]0.8800.520.980.90.260.780.70.760.5
0.0800.40.80.020.020.140.80.540.06
" + }, + { + "type": "table_caption", + "bbox": [ + 0.513, + 0.191, + 0.907, + 0.234 + ], + "angle": 0, + "content": "Table 3. TPR results on IN-SCT (first row) and corresponding DFM-filtered images (second and third rows). TPR above average TPR (0.367) is highlighted in bold." + }, + { + "type": "table", + "bbox": [ + 0.517, + 0.244, + 0.903, + 0.302 + ], + "angle": 0, + "content": "
Class MethodMil-aircraftcarlorikeettabby catholsteinLab-retrievertree froghorsefishing vesselfire truck
—0.3430.430.4430.2710.3290.38570.40.0290.4290.629
HFSS0.5140.0290.5860.8860.4570.2710.8710.2430.5430.071
[39]000.21430.12860.04290.02860.05710.12860.21430
" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.322, + 0.907, + 0.549 + ], + "angle": 0, + "content": "thus validate the impact of shortcuts searched by HFSS in generalization tests, by measuring performance results on IN-SCT, an OOD test set designed by [39]. We report the results in Tab. 3. The strong shortcuts for classes airliner, frog and container ship searched by HFSS are present in the OOD set, contributing to close-to or above average TPR of classes Military aircraft (Mil-aircraft), tree frog, and fishing vessel in the IN-SCT dataset. The shortcuts searched by [39] for class Siamese cat result in a much lower TPR for class tabby cat in IN-SCT, though shortcuts for this class still exist in the OOD data (as observed from the results of HFSS). This indicates that the method proposed in [39] has limitations in estimating the impact of shortcuts on OOD performance, and further demonstrates the effectiveness of HFSS in finding shortcuts." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.571, + 0.64, + 0.587 + ], + "angle": 0, + "content": "5. Conclusions" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.599, + 0.907, + 0.901 + ], + "angle": 0, + "content": "We proposed the first method for analysis of frequency shortcuts, HFSS, that enables the inspection of shortcut learning in large-scale models and datasets with thousands of classes. HFSS is more time-efficient and effective in finding frequency shortcuts compared to existing approaches. We investigate frequency shortcut learning in models trained on ImageNet and relate their results to robustness and generalization performance under different OOD conditions. The impact of frequency shortcuts on model generalization depends on the specific OOD scenarios. Existing models yield good performance on generalization benchmarks when texture information, corresponding to most frequency shortcuts, is mostly preserved in OOD data. Instead, frequency shortcuts impair the generalizability of models to images with rendition changes. This highlights the limitation of current OOD performance evaluation benchmarks, which need to explicitly take into account the impact that frequency shortcuts have on model performance. HFSS provides a tool to bridge this gap and extends the rigor of model generalization evaluation." + }, + { + "type": "page_number", + "bbox": [ + 0.48, + 0.945, + 0.518, + 0.957 + ], + "angle": 0, + "content": "25205" + } + ], + [ + { + "type": "title", + "bbox": [ + 0.092, + 0.091, + 0.241, + 0.108 + ], + "angle": 0, + "content": "Acknowledgment" + }, + { + "type": "text", + "bbox": [ + 0.091, + 0.115, + 0.484, + 0.177 + ], + "angle": 0, + "content": "Shunxin Wang and Nicola Strisciuglio are partly supported by the ERJU project. Europe's Rail Joint Undertaking is a European partnership on rail research and innovation established under the Horizon Europe program (2020-2027)." + }, + { + "type": "title", + "bbox": [ + 0.092, + 0.188, + 0.188, + 0.205 + ], + "angle": 0, + "content": "References" + }, + { + "type": "ref_text", + "bbox": [ + 0.101, + 0.214, + 0.484, + 0.282 + ], + "angle": 0, + "content": "[1] Hyojin Bahng, Sanghyuk Chun, Sangdoo Yun, Jaegul Choo, and Seong Joon Oh. Learning de-biased representations with biased representations. In Proceedings of the 37th International Conference on Machine Learning, pages 528-539. PMLR, 2020. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.1, + 0.284, + 0.484, + 0.354 + ], + "angle": 0, + "content": "[2] Christopher Boland, Sonia Dahdouh, Sotirios A. Tsaftaris, and Keith A Goatman. There are no shortcuts to anywhere worth going: Identifying shortcuts in deep learning models for medical image analysis. In Submitted to Medical Imaging with Deep Learning, 2024. under review. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.101, + 0.355, + 0.484, + 0.396 + ], + "angle": 0, + "content": "[3] Yuan Cao, Zhiying Fang, Yue Wu, Ding-Xuan Zhou, and Quanquan Gu. Towards understanding the spectral bias of deep learning, 2020. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.101, + 0.397, + 0.484, + 0.478 + ], + "angle": 0, + "content": "[4] Francesco Croce, Maksym Andriushchenko, Vikash Sehwag, Edoardo Debenedetti, Nicolas Flammarion, Mung Chiang, Prateek Mittal, and Matthias Hein. Robustbench: a standardized adversarial robustness benchmark. In Thirty-fifth Conference on Neural Information Processing Systems Datasets and Benchmarks Track (Round 2), 2021. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.481, + 0.484, + 0.522 + ], + "angle": 0, + "content": "[5] Nikolay Dagaev, Brett D. Roads, Xiaoliang Luo, Daniel N. Barry, Kaustubh R. Patil, and Bradley C. Love. A too-good-to-be-true prior to reduce shortcut reliance, 2021. 1, 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.523, + 0.484, + 0.578 + ], + "angle": 0, + "content": "[6] Jia Deng, Wei Dong, Richard Socher, Li-Jia Li, Kai Li, and Li Fei-Fei. Imagenet: A large-scale hierarchical image database. In 2009 IEEE Conference on Computer Vision and Pattern Recognition, pages 248-255, 2009. 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.579, + 0.484, + 0.634 + ], + "angle": 0, + "content": "[7] Paul Gavrikov and Janis Keuper. Can biases in imagenet models explain generalization? In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 22184-22194, 2024. 1, 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.635, + 0.484, + 0.717 + ], + "angle": 0, + "content": "[8] Robert Geirhos, Patricia Rubisch, Claudio Michaelis, Matthias Bethge, Felix A. Wichmann, and Wieland Brendel. Imagenet-trained CNNs are biased towards texture; increasing shape bias improves accuracy and robustness. In International Conference on Learning Representations, 2019. 1, 2, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.718, + 0.484, + 0.773 + ], + "angle": 0, + "content": "[9] Robert Geirhos, Jörn-Henrik Jacobsen, Claudio Michaelis, Richard Zemel, Wieland Brendel, Matthias Bethge, and Felix A. Wichmann. Shortcut learning in deep neural networks. Nature Machine Intelligence, 2(11):665-673, 2020. 1, 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.775, + 0.482, + 0.803 + ], + "angle": 0, + "content": "[10] Ian J. Goodfellow, Jonathon Shlens, and Christian Szegedy. Explaining and harnessing adversarial examples, 2015. 4, 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.804, + 0.482, + 0.831 + ], + "angle": 0, + "content": "[11] Jindong Gu and Daniela Oelke. Understanding bias in machine learning, 2019. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.832, + 0.484, + 0.899 + ], + "angle": 0, + "content": "[12] Jun Guo, Wei Bao, Jiakai Wang, Yuqing Ma, Xinghai Gao, Gang Xiao, Aishan Liu, Jian Dong, Xianglong Liu, and Wenjun Wu. A comprehensive evaluation framework for deep model robustness. Pattern Recognition, 137:109308, 2023. 2" + }, + { + "type": "list", + "bbox": [ + 0.094, + 0.214, + 0.484, + 0.899 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.093, + 0.905, + 0.147 + ], + "angle": 0, + "content": "[13] Dan Hendrycks and Thomas Dietterich. Benchmarking neural network robustness to common corruptions and perturbations. In International Conference on Learning Representations, 2019. 2, 4, 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.149, + 0.905, + 0.245 + ], + "angle": 0, + "content": "[14] Dan Hendrycks, Steven Basart, Norman Mu, Saurav Kadavath, Frank Wang, Evan Dorundo, Rahul Desai, Tyler Zhu, Samyak Parajuli, Mike Guo, Dawn Song, Jacob Steinhardt, and Justin Gilmer. The many faces of robustness: A critical analysis of out-of-distribution generalization. In Proceedings of the IEEE/CVF International Conference on Computer Vision (ICCV), pages 8340-8349, 2021. 2, 4, 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.247, + 0.905, + 0.302 + ], + "angle": 0, + "content": "[15] Hanxun Huang, Xingjun Ma, Sarah Monazam Erfani, James Bailey, and Yisen Wang. Unlearnable examples: Making personal data unexploitable. In International Conference on Learning Representations, 2021. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.302, + 0.905, + 0.357 + ], + "angle": 0, + "content": "[16] Zhi-Qin John Xu, Yaoyu Zhang, Tao Luo, Yanyang Xiao, and Zheng Ma. Frequency principle: Fourier analysis sheds light on deep neural networks. Communications in Computational Physics, 28(5):1746-1767, 2020. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.358, + 0.905, + 0.413 + ], + "angle": 0, + "content": "[17] Oğuzhan Fatih Kar, Teresa Yeo, Andrei Atanov, and Amir Zamir. 3d common corruptions and data augmentation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 18963-18974, 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.414, + 0.905, + 0.441 + ], + "angle": 0, + "content": "[18] A. Krizhevsky. Learning multiple layers of features from tiny images, 1970. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.442, + 0.905, + 0.482 + ], + "angle": 0, + "content": "[19] S. Lapuschkin, S. Wäldchen, A. Binder, et al. Unmasking Clever Hans predictors and assessing what machines really learn. Nat Commun 10, 1096, 2019. 1, 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.484, + 0.905, + 0.551 + ], + "angle": 0, + "content": "[20] Mathias Lecuyer, Vaggelis Atlidakis, Roxana Geambasu, Daniel Hsu, and Suman Jana. Certified robustness to adversarial examples with differential privacy. In 2019 IEEE Symposium on Security and Privacy (SP), pages 656-672, 2019. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.516, + 0.553, + 0.905, + 0.58 + ], + "angle": 0, + "content": "[21] Sungbin Lim, Ildoo Kim, Taesup Kim, Chiheon Kim, and Sungwoong Kim. Fast autoaugment, 2019. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.581, + 0.905, + 0.635 + ], + "angle": 0, + "content": "[22] Ze Liu, Yutong Lin, Yue Cao, Han Hu, Yixuan Wei, Zheng Zhang, Stephen Lin, and Baining Guo. Swin transformer: Hierarchical vision transformer using shifted windows, 2021. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.637, + 0.905, + 0.704 + ], + "angle": 0, + "content": "[23] Emanuele Marconato, Stefano Teso, Antonio Vergari, and Andrea Passerini. Not all neuro-symbolic concepts are created equal: Analysis and mitigation of reasoning shortcuts. In Thirty-seventh Conference on Neural Information Processing Systems, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.706, + 0.905, + 0.748 + ], + "angle": 0, + "content": "[24] Matthias Minderer, Olivier Bachem, Neil Houlsby, and Michael Tschannen. Automatic shortcut removal for self-supervised representation learning, 2020. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.749, + 0.905, + 0.816 + ], + "angle": 0, + "content": "[25] Eric Mintun, Alexander Kirillov, and Saining Xie. On interaction between augmentations and corruptions in natural corruption robustness. In Advances in Neural Information Processing Systems, pages 3571-3583. Curran Associates, Inc., 2021. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.818, + 0.905, + 0.845 + ], + "angle": 0, + "content": "[26] Samuel G. Müller and Frank Hutter. Trivialaugment: Tuning-free yet state-of-the-art data augmentation, 2021. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.846, + 0.905, + 0.9 + ], + "angle": 0, + "content": "[27] Meike Nauta, Ricky Walsh, Adam Dubowski, and Christin Seifert. Uncovering and correcting shortcut learning in machine learning models for skin cancer diagnosis. Diagnostics, 12(1), 2022. 1, 2, 3" + }, + { + "type": "list", + "bbox": [ + 0.516, + 0.093, + 0.905, + 0.9 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.48, + 0.945, + 0.52, + 0.957 + ], + "angle": 0, + "content": "25206" + } + ], + [ + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.092, + 0.482, + 0.16 + ], + "angle": 0, + "content": "[28] Hongjing Niu, Hanting Li, Feng Zhao, and Bin Li. Roadblocks for temporarily disabling shortcuts and learning new knowledge. In Advances in Neural Information Processing Systems, pages 29064-29075. Curran Associates, Inc., 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.164, + 0.482, + 0.22 + ], + "angle": 0, + "content": "[29] Ian Osband and Benjamin Van Roy. Why is posterior sampling better than optimism for reinforcement learning? In Proceedings of the 34th International Conference on Machine Learning, pages 2701-2710. PMLR, 2017. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.221, + 0.482, + 0.288 + ], + "angle": 0, + "content": "[30] Mohammad Pezeshki, Sekou-Oumar Kaba, Yoshua Bengio, Aaron Courville, Doina Precup, and Guillaume Lajoie. Gradient starvation: A learning proclivity in neural networks. In Advances in Neural Information Processing Systems, 2021. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.291, + 0.482, + 0.358 + ], + "angle": 0, + "content": "[31] Benjamin Recht, Rebecca Roelofs, Ludwig Schmidt, and Vaishaal Shankar. Do ImageNet classifiers generalize to ImageNet? In Proceedings of the 36th International Conference on Machine Learning, pages 5389-5400. PMLR, 2019. 2, 4, 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.363, + 0.482, + 0.403 + ], + "angle": 0, + "content": "[32] Joshua Robinson, Li Sun, Ke Yu, Kayhan Batmanghelich, Stefanie Jegelka, and Suvrit Sra. Can contrastive learning avoid shortcut solutions?, 2021. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.406, + 0.482, + 0.488 + ], + "angle": 0, + "content": "[33] Ramprasaath R. Selvaraju, Michael Cogswell, Abhishek Das, Ramakrishna Vedantam, Devi Parikh, and Dhruv Batra. Grad-cam: Visual explanations from deep networks via gradient-based localization. In 2017 IEEE International Conference on Computer Vision (ICCV), pages 618-626, 2017. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.491, + 0.482, + 0.559 + ], + "angle": 0, + "content": "[34] Harshay Shah, Kaustav Tamuly, Aditi Raghunathan, Prateek Jain, and Praneeth Netrapalli. The pitfalls of simplicity bias in neural networks. In Advances in Neural Information Processing Systems, pages 9573-9585. Curran Associates, Inc., 2020. 1, 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.562, + 0.482, + 0.631 + ], + "angle": 0, + "content": "[35] Damien Teney, Armand Mihai Nicoliciou, Valentin Hartmann, and Ehsan Abbasnejad. Neural redshift: Random networks are not random functions. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 4786-4796, 2024. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.633, + 0.482, + 0.688 + ], + "angle": 0, + "content": "[36] Haohan Wang, Songwei Ge, Zachary Lipton, and Eric P Xing. Learning robust global representations by penalizing local predictive power. In Advances in Neural Information Processing Systems, pages 10506-10518, 2019. 2, 4, 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.69, + 0.482, + 0.745 + ], + "angle": 0, + "content": "[37] Shunxin Wang, Raymond Veldhuis, Christoph Brune, and Nicola Strisciuglio. Frequency shortcut learning in neural networks. In NeurIPS 2022 Workshop on Distribution Shifts: Connecting Methods and Applications, 2022. 1, 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.748, + 0.482, + 0.816 + ], + "angle": 0, + "content": "[38] Shunxin Wang, Christoph Brune, Raymond Veldhuis, and Nicola Strisciuglio. Dfm-x: Augmentation by leveraging prior knowledge of shortcut learning. In Proceedings of the IEEE/CVF International Conference on Computer Vision (ICCV) Workshops, pages 129-138, 2023. 2, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.818, + 0.482, + 0.899 + ], + "angle": 0, + "content": "[39] Shunxin Wang, Raymond Veldhuis, Christoph Brune, and Nicola Strisciuglio. What do neural networks learn in image classification? a frequency shortcut perspective. In Proceedings of the IEEE/CVF International Conference on Computer Vision (ICCV), pages 1433-1442, 2023. 1, 2, 3, 4, 5, 8" + }, + { + "type": "list", + "bbox": [ + 0.093, + 0.092, + 0.482, + 0.899 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.516, + 0.092, + 0.905, + 0.135 + ], + "angle": 0, + "content": "[40] Shunxin Wang, Raymond Veldhuis, Christoph Brune, and Nicola Strisciuglio. A survey on the robustness of computer vision models against common corruptions, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.136, + 0.905, + 0.164 + ], + "angle": 0, + "content": "[41] Zhi-Qin John Xu and Hanxu Zhou. Deep frequency principle towards understanding why deeper learning is faster, 2020. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.165, + 0.905, + 0.219 + ], + "angle": 0, + "content": "[42] Jiayun Zheng and Maggie Makar. Causally motivated multishutdown identification and removal. In Advances in Neural Information Processing Systems, pages 12800-12812. Curran Associates, Inc., 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.221, + 0.905, + 0.276 + ], + "angle": 0, + "content": "[43] Xin Zou and Weiwei Liu. On the adversarial robustness of out-of-distribution generalization models. In Advances in Neural Information Processing Systems, pages 68908-68938. Curran Associates, Inc., 2023. 2" + }, + { + "type": "list", + "bbox": [ + 0.516, + 0.092, + 0.905, + 0.276 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.48, + 0.946, + 0.519, + 0.957 + ], + "angle": 0, + "content": "25207" + } + ] +] \ No newline at end of file diff --git a/2025/Do ImageNet-trained Models Learn Shortcuts_ The Impact of Frequency Shortcuts on Generalization/f684a2ee-5aa0-48af-b0b3-fb7e06614e21_origin.pdf b/2025/Do ImageNet-trained Models Learn Shortcuts_ The Impact of Frequency Shortcuts on Generalization/f684a2ee-5aa0-48af-b0b3-fb7e06614e21_origin.pdf new file mode 100644 index 0000000000000000000000000000000000000000..7cb0c518b7f026af02394a4be202eecc2f45bf44 --- /dev/null +++ b/2025/Do ImageNet-trained Models Learn Shortcuts_ The Impact of Frequency Shortcuts on Generalization/f684a2ee-5aa0-48af-b0b3-fb7e06614e21_origin.pdf @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:934d102995684f98b6e53c1047e95a3004b515ed7002265158223581e62e003e +size 2294874 diff --git a/2025/Do ImageNet-trained Models Learn Shortcuts_ The Impact of Frequency Shortcuts on Generalization/full.md b/2025/Do ImageNet-trained Models Learn Shortcuts_ The Impact of Frequency Shortcuts on Generalization/full.md new file mode 100644 index 0000000000000000000000000000000000000000..09585926a2c0d6e934760bbefa1f37434155d175 --- /dev/null +++ b/2025/Do ImageNet-trained Models Learn Shortcuts_ The Impact of Frequency Shortcuts on Generalization/full.md @@ -0,0 +1,223 @@ +# Do ImageNet-trained models learn shortcuts? The impact of frequency shortcuts on generalization + +Shunxin Wang Raymond Veldhuis Nicola Strisciuglio University of Twente + +{s.wang-2,r.n.j.veldhuis,n.strisciuglio}@utwente.nl + +# Abstract + +Frequency shortcuts refer to specific frequency patterns that models heavily rely on for correct classification. Previous studies have shown that models trained on small image datasets often exploit such shortcuts, potentially impairing their generalization performance. However, existing methods for identifying frequency shortcuts require expensive computations and become impractical for analyzing models trained on large datasets. In this work, we propose the first approach to more efficiently analyze frequency shortcuts at a large scale. We show that both CNN and transformer models learn frequency shortcuts on ImageNet. We also expose that frequency shortcut solutions can yield good performance on out-of-distribution (OOD) test sets which largely retain texture information. However, these shortcuts, mostly aligned with texture patterns, hinder model generalization on rendition-based OOD test sets. These observations suggest that current OOD evaluations often overlook the impact of frequency shortcuts on model generalization. Future benchmarks could thus benefit from explicitly assessing and accounting for these shortcuts to build models that generalize across a broader range of OOD scenarios. Codes are available at https://github.com/nis-research/hfss. + +# 1. Introduction + +Superficial correlations between data and ground truth [5, 9] can be learned by models to minimize training objectives with the least effort [34]. This learning behavior is called shortcut learning, which either harms the generalization performance of models or gives an illusion of good generalization abilities when learned shortcuts are present in OOD test sets [39]. Artifacts and visual cues that cause shortcut learning are recognizable by visual inspection of images, and their impact on the training of models can be mitigated through data selection or augmentation [5, 19, 27], i.e. counteracting the spurious correlations between data + +Table 1. Our method needs less computation times compared to [39] (using ResNet18 on an NVIDIA A40 GPU). Computational time of [39] on ImageNet-1k is estimated from their ImageNet-10 experiments, considering that it increases proportionally to the number of classes. + +
DatasetTime (h) [39]Time (h) (Ours)
CIFAR-107.50.5
ImageNet-1k8500 (354 days)220 (9.2 days)
+ +and ground truth. + +However, there exist shortcuts in the Fourier domain, which are implicitly embedded in image data characteristics and not easily detectable by visual inspection [37, 39]. Such shortcut solutions consist of small frequency subsets that are easy-to-learn and sufficient for models to achieve high classification rate. Typically, they correspond to simple features like textures, shapes or colors in the spatial domain. Wang, et al. [39] identified frequency shortcuts by retaining relevant frequencies to classification of a certain class. The relevance of an individual frequency was measured by the loss value of the model tested on images of the class with that frequency removed. Thus, their approach requires computational time increasing proportionally to the number of classes in a dataset and the image resolution. Yet, frequency shortcuts usually consist of multiple frequencies. The method in [39] might overlook some shortcuts as individual frequency relevance to classification neglects joint contribution of frequencies. Furthermore, the computational burden limits its applicability for analyzing shortcut learning behavior of models trained on large datasets (e.g. ImageNet) with hundreds or thousands of classes. Several studies have shown that ImageNet-trained models are biased towards textures [7, 8]. However, there is no concrete evidence of what causes this phenomenon, although [39] largely attributes it to frequency shortcut learning. + +In this work, we propose the first method that enables the uncovering of frequency shortcuts learned by models trained on large-scale datasets (e.g. ImageNet-1k), for analyzing learning behavior and explaining model generalization performance in different OOD scenarios. Our + +method improves computational efficiency as it applies parallel class-wise loss computation and hierarchical search in the Fourier domain (see Tab. 1), compared to [39]. Moreover, our method considers the joint contribution of frequencies to classification, thus being more effective in finding shortcuts. Our contributions are: + +1. We develop a hierarchical frequency shortcut search (HFSS) method, which enables the analysis of shortcut learning in the Fourier domain on large datasets with varying class counts. We reduce computational time and improve the effectiveness at identifying shortcuts. +2. We discover that ImageNet-trained models (both CNN and transformer architectures) are subject to learn frequency shortcuts. Different from shortcuts formed by visual cues [9], frequency shortcuts (easy-to-learn features) lead to good performance on both in-distribution (ID) and OOD tests if they do not block models from learning other semantics (difficult features). +3. Our HFSS enables a more comprehensive assessment of model generalizability by analyzing OOD data characteristics, specifically the presence of shortcuts. In existing evaluation frameworks, frequency shortcuts do not always impair OOD generalization performance. This emphasizes the importance of considering the role of shortcuts when designing future OOD evaluation benchmarks. + +# 2. Related works + +Shortcut learning. Models can learn shortcut solutions based on superficial correlations between data and ground truth [9, 11] to optimize training objectives with the least effort. This learning behavior is due to simplicity-bias [34], which is caused by inductive biases provided by gradient descent or components like ReLUs [35]. Shortcuts can be visual cues in the data like source tags and artificial markers [19, 27]. For instance, one fifth of the horse images from Pascal VOC dataset were found to contain a source tag, on which models rely as a discriminant feature to recognize horses [19]. Next to these visual cues, vision models are also subject to shortcuts implicitly existing in the frequency domain, which manifest as small sets of frequencies contributing significantly to image classification performance [37, 39]. The study in [39] explains why models exhibit a textures-bias for classification [8] from a frequency shortcut perspective, but does not provide further investigation on how frequency shortcuts affect the generalization and robustness performance of ImageNet-trained models, due to computational burden. + +Relying on simple shortcut solutions could harm the generalization and robustness performance of models, as shortcut learning appearing in early training might block models from learning other semantics related to the tasks at hand [3, 16, 39, 41]. However, such reliance could also cre + +ates a false impression of good generalization when shortcuts are present in OOD test sets [1, 19, 39]. + +Shortcut identification. Uncovering shortcuts helps understanding the generalizability of models on different out-of-distribution (OOD) data, explaining why models fail to generalize to or perform well on OOD data. It can provide shortcut-related prior knowledge to develop techniques that improve model generalization and robustness performance based on shortcut mitigation [27, 38]. However, common approaches are limited to identifying visually inspectable shortcuts such as text, watermark and color patches [19, 24, 27], using e.g. saliency maps [33]. Rather than directly identifying shortcuts in the data, in [2, 5] shortcut features present in individual images were quantified by assessing how difficult they were for models to learn. The authors in [28, 30, 32] uncovered shortcut features learned within representation space. The authors in [23] investigated shortcut learning by analyzing the relationship between semantic concepts using a knowledge graph, sharing a similar idea to [42]. Despite visual shortcut cues, there are non-observable shortcuts in the frequency domain, which are embedded in data characteristics. The work of [39] identified relevant frequencies to classification, which potentially contain shortcut information. As this technique measures the relevance of one frequency at a time, it requires computational costs increasing proportionally to the number of classes and image resolution. Thus, limited attention has been given to uncovering frequency shortcuts in large datasets, with existing analyses primarily focused on datasets containing a small number of classes [2, 37, 39, 42]. + +Existing methods to identify shortcuts are limited to either manual inspection [19] or involve a time-consuming algorithm [37, 39]. Methods quantifying shortcut information or uncovering shortcut features primarily aim at mitigating shortcut learning, without explaining how shortcuts in the data impact model generalization performance. We are the first to enable frequency shortcut analysis of models trained on large-scale datasets and link them to generalization performance in different OOD settings. + +OOD evaluation. The generalization and robustness performance of vision models are usually evaluated using extra data that is considered OOD, e.g. data collected at different time points [31], with different styles or renditions [8, 14, 36], with synthetic corruption effects [13, 17, 25, 40], and with adversarial noise [4, 12, 20, 43]. However, such benchmarks do not consider the impact of shortcuts learned by models and present in OOD test sets. This might ignore critical factors related to the generalization and robustness capabilities of models. The work in [7] explores model generalization from the perspectives of biases, e.g. texture, + +![](images/baac05343ec0eb17a342d50c17329627042080eed9e2ea497d7b84740056567b.jpg) +Figure 1. Scheme of HFSS. Starting from stage 2, we sample frequency patches from a random frequency subset searched in previous stage. This confines the size of search space. The white patches in the binary masks indicate sampled frequency patches. + +shape and spectral biases. In this work, we investigate the impact of frequency shortcuts on model generalization abilities, establishing connections among these different biases. Our work provides insights into when frequency shortcuts yield good performance or are harmful to model generalization and robustness performance. + +# 3. Method + +We propose a method to identify (non-visible) frequency shortcuts linked to intrinsic data characteristics rather than visual cues as in [27]. We reduce significantly computational demands compared to [39]. Our method is the first solution for shortcut identification that enables the analysis of models trained on datasets with a large number of samples and classes. In the following sections, we detail our measurement of shortcut learning in models and examine the impact of shortcuts on generalization. + +# 3.1. Identifying frequency shortcuts + +We propose a method called hierarchical frequency shortcut search (HFSS), which exploits hierarchical search of frequency subsets in the image Fourier spectrum. This reduces computational time compared to exhaustive search strategies, e.g. [38, 39]. The search is separated into several stages, each stage gradually narrowing down the spectrum search space in a coarse-to-fine manner and detecting frequency subsets that models strongly rely on for classification. The scheme of HFSS is shown in Fig. 1. + +Hierarchical search for frequency shortcuts. In order to discover the frequency subsets that models rely heavily on for correct classification of each class, we sample different + +![](images/68a3b6c4933f29d4b533355d62ccb9283cb622b4843470fba5881fa86dbba681.jpg) +Figure 2. The frequency spectrum is separated into patches, with $p\%$ sampled for shortcut evaluation. + +frequency combinations that potentially contain frequency shortcuts. We use random sampling as it allows to easily consider joint contribution of frequencies to classification, improving the effectiveness of frequency shortcut identification. Random sampling has been shown to contribute to stable search results in data augmentation and reinforcement learning [21, 26, 29], while requiring lower computations w.r.t. optimization-based methods. + +There are numerous combinations of frequencies by sampling frequency components directly. Instead of doing an exhaustive search, we apply hierarchical search in the Fourier spectrum: we divide the Fourier spectrum into frequency patches and sample these frequency patches to generate candidates of frequency subsets that contain shortcut information, as illustrated in Fig. 2. We exploit overlapping shifted windows in vertical and horizontal directions [22] to prepare frequency patches and thus avoid border effects, in addition to evenly separating the Fourier spectrum. HFSS consists of multiple search stages, the number of which depends on image resolution and desired frequency resolution of the identified frequency subsets. Lower image or frequency resolution requires fewer search stages. + +In summary, the overall process of HFSS is as follows. Each stage of HFSS consists of three steps, namely (1) sampling $B$ frequency subsets which contain $p\%$ frequency patches, (2) evaluating shortcut information, and (3) forwarding subsets for next stage. At the initial search stage, we sample frequency patches with a large size and evaluate their shortcut information. We evaluate the frequency subsets according to their contributions to classification: we process a subset of training images by retaining a frequency subset and measure the class-wise loss of a model. Frequency subsets that contribute to a lower loss value are stronger candidates to indicate frequency shortcuts as they are sufficient for the model to achieve a high prediction score. We rank frequency subsets by the loss values and use the top- $N$ masks of each class in the next search stage. + +Starting from the second search stage, we sample frequency patches from one of the top- $N$ frequency subsets in the previous stage, with a smaller patch size. This confines the search space. In the final search stage, the top-1 frequency subset of each class is considered dominating the classification of the class concerned. Same as [39], we use binary masks to represent these frequency subsets, called Dominant Frequency Maps (DFMs). + +The sampling percentage $p\%$ of frequency patches, the number of sampled frequency subsets $B$ and the frequency resolution of DFMs are hyperparameters of the search algorithm (details are in the supplementary material). + +# 3.2. Measuring the degree of shortcut learning + +Analyzing models learning behavior in a class-wise manner, as done in [39], is labor-intensive if there are hundreds or thousands of classes. To provide a broad overview of model learning behavior on training data, we categorize all classes in a dataset into two groups: (1) classes subject to shortcuts and (2) non-shortcut classes. A class is considered subject to shortcuts if its true positive rate (TPR) surpasses a given threshold when the model is tested on images filtered to retain only dominant frequencies of the class concerned. We use TPR, as a high TPR indicates that the frequency subset is sufficient to achieve a high classification rate, thereby acting as a shortcut. The frequency-filtered images are referred to as DFM-filtered images. + +For each class, we test the model on both the original test images and DFM-filtered images, computing TPR for both. We denote the TPR on the original images of class $c_{i}$ as $\mathrm{TPR}_{c_i}$ and the TPR on the DFM-filtered images as $\mathrm{TPR}_{c_i}^{DFM}$ . If $\mathrm{TPR}_{c_i}^{DFM} > t$ (where $t \in [0,1]$ is a predefined threshold), class $c_{i}$ is considered subject to shortcuts; otherwise, it is considered as one of the non-shortcut classes. We compute the average TPR values of both groups, shortcut and non-shortcut classes, at different thresholds $t$ . We note them as $\mathrm{AvgTPR}_{\mathrm{sct}}@t$ , $\mathrm{AvgTPR}_{\mathrm{sct}}^{\mathrm{DFM}}@t$ for the shortcut classes, and $\mathrm{AvgTPR}_{\mathrm{non-sct}}@t$ and $\mathrm{AvgTPR}_{\mathrm{non-sct}}^{\mathrm{DFM}}@t$ + +![](images/3f0dca1220b8a62254257c0f47a68f10a49e2c576a43e4283ce75ab19f5d00ec.jpg) +Figure 3. ResNet18 tested on original test images and DFM-filtered images of ImageNet-1k. Blue line shows the average TPR of classes subject to shortcuts on the original test images, and the orange line shows the results of non-shortcut classes, at different threshold $t$ . The green and red lines correspond to results tested on DFM-filtered images. Lower $t$ indicates weak shortcuts and higher $t$ signifies stronger ones. The size of each point reflects the number of classes. The larger the size, the more classes included. + +for the non-shortcut classes. A higher threshold $t$ corresponds to stronger shortcuts, as dominant frequencies alone are sufficient for accurate classification. An example of the results of these metrics at different threshold values are given in Fig. 3, where the size of each point reflects the number of classes in the two groups. Larger points indicate a higher number of classes, which suggests that the model has a stronger tendency to shortcut learning. + +We do not average $\mathrm{TPR}^{\mathrm{DFM}}$ across all classes because its value for many non-shortcut classes is close to zero (specifically for large datasets). This can result in a close-to-zero $\mathrm{AvgTPR}^{\mathrm{DFM}}$ which does not provide useful information for further analysis. + +# 3.3. Impact of shortcut on generalization + +To assess how frequency shortcuts impact the generalization and robustness performance of models trained on ImageNet-1k, we evaluate their performance on several datasets. We perform ID tests on ImageNet-1k (IN-1k) [6] and ImageNet-v2 (IN-v2) [31] test sets. They are considered in-distribution as they are collected using the same protocol as the training set. Furthermore, we use ImageNet-C (IN-C) [13] to benchmark the robustness of models against appearance image corruptions, e.g. noise, blur and weather changes. To evaluate the generalizability of models to images with different renditions, we use ImageNet-Renditions (IN-R) [14] and ImageNet-Sketch (IN-S) [36]. IN-R contains renditions like cartoons, paintings, art and toys of 200 classes from IN-1k. IN-S has the same number of classes as IN-1k, containing images of hand-drawn sketches. These datasets contain visual renditions, which serve to evaluate how the reliance on texture cues (corresponding to most frequency shortcuts) impacts on generalization. To measure the adversarial robustness of models, we apply the fast gradient sign method (FGSM) attacks [10] to the validation set of IN-1k, with $L_{\infty} = 4 / 255$ . We calculate the average TPR of shortcut and non-shortcut classes on OOD data, that we compare to the results on ID data. + +# 4. Experiments + +We start with investigating the trade-off between efficiency (i.e. required computational time) and effectiveness (i.e. capability of finding shortcuts), varying the configuration of the number of sampled frequency subsets at each stage, noted as $B_{s}$ ( $s$ is the index of stage). Then, we analyze frequency shortcut learning of models trained on IN-1k, relating it to performance results under different OOD scenarios. + +Setup. We use CIFAR-10 (C-10) [18] to explore the configuration of $B$ , and configure HFSS with four search stages, with patch size of $8 \times 8$ , $4 \times 4$ , $2 \times 2$ and $1 \times 1$ , respectively. We then perform larger-scale shortcut analysis using IN-1k: we use six search stages, with patch size of $56 \times 56$ , $28 \times 28$ , $14 \times 14$ , $8 \times 8$ , $4 \times 4$ and $2 \times 2$ . The patch (frequency) resolution in the final stage is the same as that in [39]. At each stage, we sample $p = 60\%$ frequency patches to form frequency subsets. It results in DFMs containing about $5\%$ frequencies of the whole spectrum of ImageNet images and about $15\%$ frequencies of CIFAR images. The number of training images sampled for shortcut information evaluation is the same as that of their corresponding test set. We evaluate the generalization and robustness performance of ImageNet-trained models on IN-v2 [31], IN-C [13], IN-R [14], IN-S [36] and using FGSM [10] attacks. We also carry out experiments on ImageNet-10 (IN-10) [15] with a ResNet18 model to compare HFSS with single-frequency removal-based method in [39]. Following their setups, we use ImageNet-SCT (IN-SCT) for OOD evaluation. Our configuration of HFSS enables direct comparison with the results reported in [39] as DFMs contain around $5\%$ frequencies of whole spectrum. We report training configurations and additional results in the supplementary material. + +# 4.1. Configuration of HFSS + +Sampled frequency subsets. At stage $s$ , we sample $B_{s}$ frequency subsets, such that HFSS can explore different combinations of frequency patterns, validating their contained shortcut information. However, sampling an excessive number of subsets increases the required search time, as more subsets are for evaluating their contributions to classification results. To explore the impact of the number of sampling operation $B_{s}$ on shortcut identification, we perform experiments on C-10 using a ResNet18 backbone, setting the number of sampled frequency subsets in each stage to be as large as possible under limited computational time. We hypothesize that HFSS obtains stable results given enough candidate frequency subsets. We set $B_{1}, B_{2}, B_{3}$ , and $B_{4}$ to be 1000, 2000, 4000 and 8000 respectively, which we note as configuration CF-1. The increasing number of candidate frequency subsets at consecutive stages is determined by the increase in possible frequency patch combinations as patch + +![](images/3a9d3b140d81d1445e08aec2994b8cf68392596b766f0d54a2e1dfb0b1f501fd.jpg) +Figure 4. The best class-wise loss vs. the number of sampled frequency subsets at each stage. + +# size decreases. + +We run the configuration CF-1 five times and note the stability of our HFSS search algorithms. We track the best (lowest) loss values achieved for each class when the model is tested on DFM-filtered images. In Fig. 4, we report the tracking statistics with average loss values (lines) and standard deviations (shadows) over the five trials for each class, which are further used as a starting point to optimize the configuration of HFSS. As the number of sampled frequency subsets increases, the best loss for each class converges, indicating the stability of HFSS in identifying similar shortcuts learned by models when a sufficient amount of candidate frequency subsets are sampled for validating shortcut information. The standard deviation recorded at different stages is relatively small compared to the range of best loss. For class airplane we observe an outlier outcome at stage 3. In the supplementary material, we present visualization of shortcut cues (DFM-filtered images) computed across the five trials. These visualizations show that the identified shortcut patterns differ in the orientations of strip-like features, corresponding to line features in the image of airplane. This suggests that frequency shortcuts do not arise from fixed frequency subsets; rather, spatial features they correspond to may be visually similar but composed of different frequencies, determining an higher standard deviation in Fig. 4. + +Efficiency vs. effectiveness. Based on the results in Fig. 4, we optimize CF-1 for efficiency and effectiveness in shortcut identification. The required search time scales up proportionally to $B_{s}$ , i.e. the more candidate frequency subsets are sampled, the more time is needed to verify their + +![](images/44804d03646bcc9d1c3f209e82d0f03083b58153e96f1f4e5679eb69fb893f86.jpg) +Figure 5. Average TPR vs. search time, where search time increases proportionally with the number of sampled frequency subsets. + +![](images/c2ad637d9e710d9e444390052802dbdc77589e53c0ac22f3c0809fbc2c6ee9d1.jpg) +Figure 6. The number of shortcut classes given different threshold $t$ search by CF-1 and CF-2.10. Using CF-2.10 uncovers most shortcuts identified by using CF-1 at low thresholds, with computational time reduced to a factor of around 200. + +relevance to classification. We perform experiments using different $B_{s}$ , investigating how the size of the search space affects the identification of shortcuts. Based on the tracking statistics in Fig. 4, which shows that increasing $B_{s}$ beyond a certain point does not decrease loss significantly, we adjust $B_{1}$ to 200, $B_{2}$ to 800 and $B_{4}$ to 4000 (CF-2.1). We further reduce $B_{3}$ and $B_{4}$ to 2000 (CF-2.2). We also design eight other configurations (see details in the supplementary material), with CF-2.10 being the most efficient as it samples the lowest number of candidate frequency subsets. + +To compare the performance of HFSS under different configurations, we compute the average TPR over all classes when tested on DFM-filtered images. This indicates the general relevance of the searched frequency subsets to classification. A higher average TPR indicate higher relevance, thus the frequency subsets contain more frequency shortcut information. We run each configuration five times, computing the mean and standard deviations of average TPRs over the five trials. We report the results in Fig. 5 where the left-most point corresponds to average TPR of CF-2.10 and the right-most point corresponds to that of CF-1. We observe that as more frequency subsets are sampled, average TPR increases and saturates. This demonstrates that sampling more candidate frequency subsets does not significantly improve the performance of HFSS, while requiring more computations. + +We compare CF-1 and CF-2.10 in Fig. 6 based on the number of classes with identified shortcuts at different thresholds. We observe that CF-2.10 uncovers most of the + +shortcuts identified by using the more complex CF-1 at low threshold level. Although CF-2.10 misses stronger shortcuts (at higher threshold levels), it achieves a $\sim 200\times$ reduction in computational time compared to CF-1. A more time-efficient configuration is less effective at finding shortcuts than the more complex one, but it still manages to identify shortcuts at low thresholds. + +The observations made on C-10 allow to estimate initial configurations for analysis on larger-scale datasets. We track the loss statistics of HFSS applied to ImageNet (provided in the supplementary material) and choose the value $B_{s}$ for ImageNet experiments in Sec. 4.2 as the one where the loss values show no significant decrease. Given the computational intensity of identifying shortcuts in large-scale datasets, our work on IN-1k primarily aims at uncovering shortcuts with affordable computational efforts. + +# 4.2. Shortcut identification on IN-1k + +We apply HFSS to examine frequency shortcuts learned by ImageNet models and how they impact model generalization and robustness. The search configurations for ImageNet models, aimed at time-efficiency and effectiveness at finding shortcuts, are in the supplementary material. + +Results on IN-1k, IN-v2 and IN-C. In Figs. 7 and 8, we report the results of the analysis of impact of shortcuts on model performance. In general, models perform better predictions on images of classes affected by shortcut learning (blue lines) compared to non-shortcut classes (orange lines) in both ID and corruption tests. Models subject to frequency shortcuts yield good performance in the robustness and generalization tests when textures information is largely preserved. For ResNet and ViT models, when $t \geq 0.8$ , the AvgTPR $_{\text{sct}}^{\text{DFM}}$ exceeds AvgTPR $_{\text{sct}}$ , meaning that these models rely predominantly on dominant frequencies (corresponding to simple features) for classification of shortcut-affected classes, preventing from learning class-related semantics features. For instance, at $t = 0.9$ , only one class ('window screen') is found subject to frequency shortcuts: ResNet18 achieves TPR=0.66 on full-spectrum images of this class and TPR=0.94 on DFM-filtered images (similarly to ResNet50 and ViT-b). Differently, CCT exhibits lower AvgTPR $_{\text{sct}}^{\text{DFM}}@0.9$ than AvgTPR $_{\text{sct}}@0.9$ , indicating that while frequency shortcuts contribute significantly to classification, this model also manage to leverage other semantic information: a strong frequency shortcut (e.g. when $t > 0.7$ ) does not necessarily mean that it is the only information a model uses for classification. We further compare CCT with ResNet50 as they perform similarly on IN-1k (80.57% and 80.1% respectively). In IN-v2 and IN-C, CCT achieves 74.81% and 57.73% prediction rates, having better performance than ResNet50 (74.17% in IN-v2 and 48.85% in IN-C). Considering the highest de + +![](images/2f410602a5803270cfd3b01ddd032747c75980460b1f0551cdc1e2c9fd9d3d13.jpg) +Figure 7. Average TPR of shortcut and non-shortcut classes given different thresholds on datasets with 1000 classes. Models generally perform better on images of shortcut classes than non-shortcut classes. This does not hold for results on IN-S, which lacks preserved texture information. + +![](images/e7eb7c03bb47cb73c67ca58eec8e315e6fd629ae511eca91da914eafbf84387e.jpg) +Figure 8. Average TPR of shortcut and non-shortcut classes at different thresholds on datasets with 200 classes. Models perform better on shortcut classes than non-shortcut classes in IN-200. IN-R lacks preserved texture information and thus the models have worse performance on shortcut classes than non-shortcut classes. + +gree of shortcut learning of CCT among the models (indicated by the marker size), inducing frequency reliance without blocking the learning of other semantic features appears + +to benefit model robustness and generalization performance under statistical distribution shifts and corruption scenarios. + +Results on IN-S and IN-R. Different from the results on IN-v2 and IN-C, frequency shortcuts impair generalization of models to texture and rendition changes. Model performance on shortcut and non-shortcut classes on IN-S is close to each other. In Fig. 8, we observe that shortcut classes have worse prediction results than non-shortcut classes when tested on IN-R which contains images with rendition changes. This is attributable to the fact that shortcut information are not available in the OOD tests, as the rendition and sketch test sets preserve less or very different texture information than IN-C and IN-v2 (see green lines of IN-S and IN-1k, IN-R and IN-200 in Fig. 7 and Fig. 8). Frequency shortcut learning is an explicit cause of the texture-bias of ImageNet models, and the resulting impaired generalizability to renditions is in line with [8]. + +Results on FGSM attacks. As shown in Fig. 7, the four models, under adversarial attacks, achieve higher AvgTPR for shortcut classes (especially those with strong shortcuts) than non-shortcut classes. This suggests that models can be inherently robust to adversarial noise if they leverage frequency shortcuts for classification. This is not surprising as adversarial noises hardly manipulate the textures of images. CCT, which shows the highest degree of frequency shortcut learning among the considered models, achieves the best adversarial robustness. This indicates that model reliance + +![](images/3d19b165873f3c16216ae597e66203f09ef63f5b145986049d1263da92b5650b.jpg) +Figure 9. An image of dugong contains frequency shortcuts which a model relies on for high-confidence classification. Its presence in OOD test sets like IN-C, and in the image under FGSM attacks yields correct predictions. But the reliance on shortcuts impairs model generalization performance to renditions changes. + +on frequency shortcuts can benefit their adversarial robustness, as the simple shortcut features are robust to adversarial noise. + +Summary. Whether frequency shortcuts impair or benefit model generalization and robustness performance depends on the specific OOD scenarios. In Fig. 9, we show an example image of dugong and model predictions in OOD tests. The preserved texture information yields correct predictions on IN-C and under FGSM attacks, but it is not helpful for model generalizability to rendition changes. The findings reveal a limitation in the design of current OOD benchmarks, which overlooks the impact of frequency shortcuts on generalizability of models and its relation to specific characteristics of OOD data. Using HFSS for performance evaluation can bridge such gap. + +# 4.3. Comparison with the existing approach + +More effective at finding shortcuts. We perform experiments on IN-10 for a direct comparison with the only existing method for frequency shortcut analysis [39], which evaluates the contribution of single frequencies to shortcuts and is limited to analyzing models trained on small-scale datasets. We report results in Tab. 2. HFSS uncovers strong frequency shortcuts for classes airliner, Siamese cat (Siam-cat), ox, frog, zebra, and container ship (Con-ship), while [39] was only able to find strong frequency shortcuts for classes zebra and Siamese cat. The better effectiveness of HFSS at finding shortcuts is attributable to the fact that HFSS considers the joint contributions of frequencies to classification, while in [39] the relevance of a single frequency was measured iteratively. + +Frequency shortcuts can impair the generalization performance of models or provide a false impression of good generalization when shortcuts exist in OOD tests [39]. We + +Table 2. TPR results on DFM-filtered IN-10 images. TPR $\geq 0.6$ (a strong frequency shortcut) is highlighted in bold. + +
Class Methodairlinerwagonhum-birdSiam-catoxgolden retrieverfrogzebraCon-shiptruck
HFSS [39]0.8800.520.980.90.260.780.70.760.5
0.0800.40.80.020.020.140.80.540.06
+ +Table 3. TPR results on IN-SCT (first row) and corresponding DFM-filtered images (second and third rows). TPR above average TPR (0.367) is highlighted in bold. + +
Class MethodMil-aircraftcarlorikeettabby catholsteinLab-retrievertree froghorsefishing vesselfire truck
—0.3430.430.4430.2710.3290.38570.40.0290.4290.629
HFSS0.5140.0290.5860.8860.4570.2710.8710.2430.5430.071
[39]000.21430.12860.04290.02860.05710.12860.21430
+ +thus validate the impact of shortcuts searched by HFSS in generalization tests, by measuring performance results on IN-SCT, an OOD test set designed by [39]. We report the results in Tab. 3. The strong shortcuts for classes airliner, frog and container ship searched by HFSS are present in the OOD set, contributing to close-to or above average TPR of classes Military aircraft (Mil-aircraft), tree frog, and fishing vessel in the IN-SCT dataset. The shortcuts searched by [39] for class Siamese cat result in a much lower TPR for class tabby cat in IN-SCT, though shortcuts for this class still exist in the OOD data (as observed from the results of HFSS). This indicates that the method proposed in [39] has limitations in estimating the impact of shortcuts on OOD performance, and further demonstrates the effectiveness of HFSS in finding shortcuts. + +# 5. Conclusions + +We proposed the first method for analysis of frequency shortcuts, HFSS, that enables the inspection of shortcut learning in large-scale models and datasets with thousands of classes. HFSS is more time-efficient and effective in finding frequency shortcuts compared to existing approaches. We investigate frequency shortcut learning in models trained on ImageNet and relate their results to robustness and generalization performance under different OOD conditions. The impact of frequency shortcuts on model generalization depends on the specific OOD scenarios. Existing models yield good performance on generalization benchmarks when texture information, corresponding to most frequency shortcuts, is mostly preserved in OOD data. Instead, frequency shortcuts impair the generalizability of models to images with rendition changes. This highlights the limitation of current OOD performance evaluation benchmarks, which need to explicitly take into account the impact that frequency shortcuts have on model performance. HFSS provides a tool to bridge this gap and extends the rigor of model generalization evaluation. + +# Acknowledgment + +Shunxin Wang and Nicola Strisciuglio are partly supported by the ERJU project. Europe's Rail Joint Undertaking is a European partnership on rail research and innovation established under the Horizon Europe program (2020-2027). + +# References + +[1] Hyojin Bahng, Sanghyuk Chun, Sangdoo Yun, Jaegul Choo, and Seong Joon Oh. Learning de-biased representations with biased representations. In Proceedings of the 37th International Conference on Machine Learning, pages 528-539. PMLR, 2020. 2 +[2] Christopher Boland, Sonia Dahdouh, Sotirios A. Tsaftaris, and Keith A Goatman. There are no shortcuts to anywhere worth going: Identifying shortcuts in deep learning models for medical image analysis. In Submitted to Medical Imaging with Deep Learning, 2024. under review. 2 +[3] Yuan Cao, Zhiying Fang, Yue Wu, Ding-Xuan Zhou, and Quanquan Gu. Towards understanding the spectral bias of deep learning, 2020. 2 +[4] Francesco Croce, Maksym Andriushchenko, Vikash Sehwag, Edoardo Debenedetti, Nicolas Flammarion, Mung Chiang, Prateek Mittal, and Matthias Hein. Robustbench: a standardized adversarial robustness benchmark. In Thirty-fifth Conference on Neural Information Processing Systems Datasets and Benchmarks Track (Round 2), 2021. 2 +[5] Nikolay Dagaev, Brett D. Roads, Xiaoliang Luo, Daniel N. Barry, Kaustubh R. Patil, and Bradley C. Love. A too-good-to-be-true prior to reduce shortcut reliance, 2021. 1, 2 +[6] Jia Deng, Wei Dong, Richard Socher, Li-Jia Li, Kai Li, and Li Fei-Fei. Imagenet: A large-scale hierarchical image database. In 2009 IEEE Conference on Computer Vision and Pattern Recognition, pages 248-255, 2009. 4 +[7] Paul Gavrikov and Janis Keuper. Can biases in imagenet models explain generalization? In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 22184-22194, 2024. 1, 2 +[8] Robert Geirhos, Patricia Rubisch, Claudio Michaelis, Matthias Bethge, Felix A. Wichmann, and Wieland Brendel. Imagenet-trained CNNs are biased towards texture; increasing shape bias improves accuracy and robustness. In International Conference on Learning Representations, 2019. 1, 2, 7 +[9] Robert Geirhos, Jörn-Henrik Jacobsen, Claudio Michaelis, Richard Zemel, Wieland Brendel, Matthias Bethge, and Felix A. Wichmann. Shortcut learning in deep neural networks. Nature Machine Intelligence, 2(11):665-673, 2020. 1, 2 +[10] Ian J. Goodfellow, Jonathon Shlens, and Christian Szegedy. Explaining and harnessing adversarial examples, 2015. 4, 5 +[11] Jindong Gu and Daniela Oelke. Understanding bias in machine learning, 2019. 2 +[12] Jun Guo, Wei Bao, Jiakai Wang, Yuqing Ma, Xinghai Gao, Gang Xiao, Aishan Liu, Jian Dong, Xianglong Liu, and Wenjun Wu. A comprehensive evaluation framework for deep model robustness. Pattern Recognition, 137:109308, 2023. 2 + +[13] Dan Hendrycks and Thomas Dietterich. Benchmarking neural network robustness to common corruptions and perturbations. In International Conference on Learning Representations, 2019. 2, 4, 5 +[14] Dan Hendrycks, Steven Basart, Norman Mu, Saurav Kadavath, Frank Wang, Evan Dorundo, Rahul Desai, Tyler Zhu, Samyak Parajuli, Mike Guo, Dawn Song, Jacob Steinhardt, and Justin Gilmer. The many faces of robustness: A critical analysis of out-of-distribution generalization. In Proceedings of the IEEE/CVF International Conference on Computer Vision (ICCV), pages 8340-8349, 2021. 2, 4, 5 +[15] Hanxun Huang, Xingjun Ma, Sarah Monazam Erfani, James Bailey, and Yisen Wang. Unlearnable examples: Making personal data unexploitable. In International Conference on Learning Representations, 2021. 5 +[16] Zhi-Qin John Xu, Yaoyu Zhang, Tao Luo, Yanyang Xiao, and Zheng Ma. Frequency principle: Fourier analysis sheds light on deep neural networks. Communications in Computational Physics, 28(5):1746-1767, 2020. 2 +[17] Oğuzhan Fatih Kar, Teresa Yeo, Andrei Atanov, and Amir Zamir. 3d common corruptions and data augmentation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 18963-18974, 2022. 2 +[18] A. Krizhevsky. Learning multiple layers of features from tiny images, 1970. 5 +[19] S. Lapuschkin, S. Wäldchen, A. Binder, et al. Unmasking Clever Hans predictors and assessing what machines really learn. Nat Commun 10, 1096, 2019. 1, 2 +[20] Mathias Lecuyer, Vaggelis Atlidakis, Roxana Geambasu, Daniel Hsu, and Suman Jana. Certified robustness to adversarial examples with differential privacy. In 2019 IEEE Symposium on Security and Privacy (SP), pages 656-672, 2019. 2 +[21] Sungbin Lim, Ildoo Kim, Taesup Kim, Chiheon Kim, and Sungwoong Kim. Fast autoaugment, 2019. 3 +[22] Ze Liu, Yutong Lin, Yue Cao, Han Hu, Yixuan Wei, Zheng Zhang, Stephen Lin, and Baining Guo. Swin transformer: Hierarchical vision transformer using shifted windows, 2021. 3 +[23] Emanuele Marconato, Stefano Teso, Antonio Vergari, and Andrea Passerini. Not all neuro-symbolic concepts are created equal: Analysis and mitigation of reasoning shortcuts. In Thirty-seventh Conference on Neural Information Processing Systems, 2023. 2 +[24] Matthias Minderer, Olivier Bachem, Neil Houlsby, and Michael Tschannen. Automatic shortcut removal for self-supervised representation learning, 2020. 2 +[25] Eric Mintun, Alexander Kirillov, and Saining Xie. On interaction between augmentations and corruptions in natural corruption robustness. In Advances in Neural Information Processing Systems, pages 3571-3583. Curran Associates, Inc., 2021. 2 +[26] Samuel G. Müller and Frank Hutter. Trivialaugment: Tuning-free yet state-of-the-art data augmentation, 2021. 3 +[27] Meike Nauta, Ricky Walsh, Adam Dubowski, and Christin Seifert. Uncovering and correcting shortcut learning in machine learning models for skin cancer diagnosis. Diagnostics, 12(1), 2022. 1, 2, 3 + +[28] Hongjing Niu, Hanting Li, Feng Zhao, and Bin Li. Roadblocks for temporarily disabling shortcuts and learning new knowledge. In Advances in Neural Information Processing Systems, pages 29064-29075. Curran Associates, Inc., 2022. 2 +[29] Ian Osband and Benjamin Van Roy. Why is posterior sampling better than optimism for reinforcement learning? In Proceedings of the 34th International Conference on Machine Learning, pages 2701-2710. PMLR, 2017. 3 +[30] Mohammad Pezeshki, Sekou-Oumar Kaba, Yoshua Bengio, Aaron Courville, Doina Precup, and Guillaume Lajoie. Gradient starvation: A learning proclivity in neural networks. In Advances in Neural Information Processing Systems, 2021. 2 +[31] Benjamin Recht, Rebecca Roelofs, Ludwig Schmidt, and Vaishaal Shankar. Do ImageNet classifiers generalize to ImageNet? In Proceedings of the 36th International Conference on Machine Learning, pages 5389-5400. PMLR, 2019. 2, 4, 5 +[32] Joshua Robinson, Li Sun, Ke Yu, Kayhan Batmanghelich, Stefanie Jegelka, and Suvrit Sra. Can contrastive learning avoid shortcut solutions?, 2021. 2 +[33] Ramprasaath R. Selvaraju, Michael Cogswell, Abhishek Das, Ramakrishna Vedantam, Devi Parikh, and Dhruv Batra. Grad-cam: Visual explanations from deep networks via gradient-based localization. In 2017 IEEE International Conference on Computer Vision (ICCV), pages 618-626, 2017. 2 +[34] Harshay Shah, Kaustav Tamuly, Aditi Raghunathan, Prateek Jain, and Praneeth Netrapalli. The pitfalls of simplicity bias in neural networks. In Advances in Neural Information Processing Systems, pages 9573-9585. Curran Associates, Inc., 2020. 1, 2 +[35] Damien Teney, Armand Mihai Nicoliciou, Valentin Hartmann, and Ehsan Abbasnejad. Neural redshift: Random networks are not random functions. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 4786-4796, 2024. 2 +[36] Haohan Wang, Songwei Ge, Zachary Lipton, and Eric P Xing. Learning robust global representations by penalizing local predictive power. In Advances in Neural Information Processing Systems, pages 10506-10518, 2019. 2, 4, 5 +[37] Shunxin Wang, Raymond Veldhuis, Christoph Brune, and Nicola Strisciuglio. Frequency shortcut learning in neural networks. In NeurIPS 2022 Workshop on Distribution Shifts: Connecting Methods and Applications, 2022. 1, 2 +[38] Shunxin Wang, Christoph Brune, Raymond Veldhuis, and Nicola Strisciuglio. Dfm-x: Augmentation by leveraging prior knowledge of shortcut learning. In Proceedings of the IEEE/CVF International Conference on Computer Vision (ICCV) Workshops, pages 129-138, 2023. 2, 3 +[39] Shunxin Wang, Raymond Veldhuis, Christoph Brune, and Nicola Strisciuglio. What do neural networks learn in image classification? a frequency shortcut perspective. In Proceedings of the IEEE/CVF International Conference on Computer Vision (ICCV), pages 1433-1442, 2023. 1, 2, 3, 4, 5, 8 + +[40] Shunxin Wang, Raymond Veldhuis, Christoph Brune, and Nicola Strisciuglio. A survey on the robustness of computer vision models against common corruptions, 2023. 2 +[41] Zhi-Qin John Xu and Hanxu Zhou. Deep frequency principle towards understanding why deeper learning is faster, 2020. 2 +[42] Jiayun Zheng and Maggie Makar. Causally motivated multishutdown identification and removal. In Advances in Neural Information Processing Systems, pages 12800-12812. Curran Associates, Inc., 2022. 2 +[43] Xin Zou and Weiwei Liu. On the adversarial robustness of out-of-distribution generalization models. In Advances in Neural Information Processing Systems, pages 68908-68938. Curran Associates, Inc., 2023. 2 \ No newline at end of file diff --git a/2025/Do ImageNet-trained Models Learn Shortcuts_ The Impact of Frequency Shortcuts on Generalization/images.zip b/2025/Do ImageNet-trained Models Learn Shortcuts_ The Impact of Frequency Shortcuts on Generalization/images.zip new file mode 100644 index 0000000000000000000000000000000000000000..aa2f6579e24adace1b81c232218b473d4a724682 --- /dev/null +++ b/2025/Do ImageNet-trained Models Learn Shortcuts_ The Impact of Frequency Shortcuts on Generalization/images.zip @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:8e2070967a853a8b3f17b8e658e84c67d6b2d9094c1e1b4a6c01dcecf38132a4 +size 443560 diff --git a/2025/Do ImageNet-trained Models Learn Shortcuts_ The Impact of Frequency Shortcuts on Generalization/layout.json b/2025/Do ImageNet-trained Models Learn Shortcuts_ The Impact of Frequency Shortcuts on Generalization/layout.json new file mode 100644 index 0000000000000000000000000000000000000000..dab99a42d540aab59c133a80ad99b23c34658202 --- /dev/null +++ b/2025/Do ImageNet-trained Models Learn Shortcuts_ The Impact of Frequency Shortcuts on Generalization/layout.json @@ -0,0 +1,6560 @@ +{ + "pdf_info": [ + { + "para_blocks": [ + { + "bbox": [ + 143, + 103, + 467, + 140 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 143, + 103, + 467, + 140 + ], + "spans": [ + { + "bbox": [ + 143, + 103, + 467, + 140 + ], + "type": "text", + "content": "Do ImageNet-trained models learn shortcuts? The impact of frequency shortcuts on generalization" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 162, + 161, + 448, + 190 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 162, + 161, + 448, + 190 + ], + "spans": [ + { + "bbox": [ + 162, + 161, + 448, + 190 + ], + "type": "text", + "content": "Shunxin Wang Raymond Veldhuis Nicola Strisciuglio University of Twente" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 160, + 191, + 447, + 204 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 160, + 191, + 447, + 204 + ], + "spans": [ + { + "bbox": [ + 160, + 191, + 447, + 204 + ], + "type": "text", + "content": "{s.wang-2,r.n.j.veldhuis,n.strisciuglio}@utwente.nl" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 151, + 231, + 200, + 243 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 151, + 231, + 200, + 243 + ], + "spans": [ + { + "bbox": [ + 151, + 231, + 200, + 243 + ], + "type": "text", + "content": "Abstract" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 55, + 258, + 297, + 533 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 258, + 297, + 533 + ], + "spans": [ + { + "bbox": [ + 55, + 258, + 297, + 533 + ], + "type": "text", + "content": "Frequency shortcuts refer to specific frequency patterns that models heavily rely on for correct classification. Previous studies have shown that models trained on small image datasets often exploit such shortcuts, potentially impairing their generalization performance. However, existing methods for identifying frequency shortcuts require expensive computations and become impractical for analyzing models trained on large datasets. In this work, we propose the first approach to more efficiently analyze frequency shortcuts at a large scale. We show that both CNN and transformer models learn frequency shortcuts on ImageNet. We also expose that frequency shortcut solutions can yield good performance on out-of-distribution (OOD) test sets which largely retain texture information. However, these shortcuts, mostly aligned with texture patterns, hinder model generalization on rendition-based OOD test sets. These observations suggest that current OOD evaluations often overlook the impact of frequency shortcuts on model generalization. Future benchmarks could thus benefit from explicitly assessing and accounting for these shortcuts to build models that generalize across a broader range of OOD scenarios. Codes are available at https://github.com/nis-research/hfss." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 55, + 560, + 135, + 572 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 560, + 135, + 572 + ], + "spans": [ + { + "bbox": [ + 55, + 560, + 135, + 572 + ], + "type": "text", + "content": "1. Introduction" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 55, + 582, + 296, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 582, + 296, + 713 + ], + "spans": [ + { + "bbox": [ + 55, + 582, + 296, + 713 + ], + "type": "text", + "content": "Superficial correlations between data and ground truth [5, 9] can be learned by models to minimize training objectives with the least effort [34]. This learning behavior is called shortcut learning, which either harms the generalization performance of models or gives an illusion of good generalization abilities when learned shortcuts are present in OOD test sets [39]. Artifacts and visual cues that cause shortcut learning are recognizable by visual inspection of images, and their impact on the training of models can be mitigated through data selection or augmentation [5, 19, 27], i.e. counteracting the spurious correlations between data" + } + ] + } + ], + "index": 8 + }, + { + "type": "table", + "bbox": [ + 342, + 295, + 526, + 337 + ], + "blocks": [ + { + "bbox": [ + 313, + 230, + 555, + 285 + ], + "lines": [ + { + "bbox": [ + 313, + 230, + 555, + 285 + ], + "spans": [ + { + "bbox": [ + 313, + 230, + 555, + 285 + ], + "type": "text", + "content": "Table 1. Our method needs less computation times compared to [39] (using ResNet18 on an NVIDIA A40 GPU). Computational time of [39] on ImageNet-1k is estimated from their ImageNet-10 experiments, considering that it increases proportionally to the number of classes." + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 342, + 295, + 526, + 337 + ], + "lines": [ + { + "bbox": [ + 342, + 295, + 526, + 337 + ], + "spans": [ + { + "bbox": [ + 342, + 295, + 526, + 337 + ], + "type": "table", + "html": "
DatasetTime (h) [39]Time (h) (Ours)
CIFAR-107.50.5
ImageNet-1k8500 (354 days)220 (9.2 days)
", + "image_path": "e9e5485f2c317ae9ca07b9bfed19c83756ad1aec481fbc942f3917fd053c6ccb.jpg" + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "table_body" + } + ], + "index": 10 + }, + { + "bbox": [ + 313, + 342, + 386, + 353 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 342, + 386, + 353 + ], + "spans": [ + { + "bbox": [ + 313, + 342, + 386, + 353 + ], + "type": "text", + "content": "and ground truth." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 312, + 354, + 555, + 653 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 312, + 354, + 555, + 653 + ], + "spans": [ + { + "bbox": [ + 312, + 354, + 555, + 653 + ], + "type": "text", + "content": "However, there exist shortcuts in the Fourier domain, which are implicitly embedded in image data characteristics and not easily detectable by visual inspection [37, 39]. Such shortcut solutions consist of small frequency subsets that are easy-to-learn and sufficient for models to achieve high classification rate. Typically, they correspond to simple features like textures, shapes or colors in the spatial domain. Wang, et al. [39] identified frequency shortcuts by retaining relevant frequencies to classification of a certain class. The relevance of an individual frequency was measured by the loss value of the model tested on images of the class with that frequency removed. Thus, their approach requires computational time increasing proportionally to the number of classes in a dataset and the image resolution. Yet, frequency shortcuts usually consist of multiple frequencies. The method in [39] might overlook some shortcuts as individual frequency relevance to classification neglects joint contribution of frequencies. Furthermore, the computational burden limits its applicability for analyzing shortcut learning behavior of models trained on large datasets (e.g. ImageNet) with hundreds or thousands of classes. Several studies have shown that ImageNet-trained models are biased towards textures [7, 8]. However, there is no concrete evidence of what causes this phenomenon, although [39] largely attributes it to frequency shortcut learning." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 313, + 654, + 556, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 654, + 556, + 715 + ], + "spans": [ + { + "bbox": [ + 313, + 654, + 556, + 715 + ], + "type": "text", + "content": "In this work, we propose the first method that enables the uncovering of frequency shortcuts learned by models trained on large-scale datasets (e.g. ImageNet-1k), for analyzing learning behavior and explaining model generalization performance in different OOD scenarios. Our" + } + ] + } + ], + "index": 13 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "spans": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "type": "text", + "content": "CVF" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 145, + 0, + 496, + 37 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 145, + 0, + 496, + 37 + ], + "spans": [ + { + "bbox": [ + 145, + 0, + 496, + 37 + ], + "type": "text", + "content": "This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "type": "text", + "content": "25198" + } + ] + } + ], + "index": 14 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 0 + }, + { + "para_blocks": [ + { + "bbox": [ + 55, + 72, + 294, + 144 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 72, + 294, + 144 + ], + "spans": [ + { + "bbox": [ + 55, + 72, + 294, + 144 + ], + "type": "text", + "content": "method improves computational efficiency as it applies parallel class-wise loss computation and hierarchical search in the Fourier domain (see Tab. 1), compared to [39]. Moreover, our method considers the joint contribution of frequencies to classification, thus being more effective in finding shortcuts. Our contributions are:" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 55, + 144, + 295, + 383 + ], + "type": "list", + "angle": 0, + "index": 4, + "blocks": [ + { + "bbox": [ + 56, + 144, + 295, + 204 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 144, + 295, + 204 + ], + "spans": [ + { + "bbox": [ + 56, + 144, + 295, + 204 + ], + "type": "text", + "content": "1. We develop a hierarchical frequency shortcut search (HFSS) method, which enables the analysis of shortcut learning in the Fourier domain on large datasets with varying class counts. We reduce computational time and improve the effectiveness at identifying shortcuts." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 55, + 204, + 295, + 287 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 204, + 295, + 287 + ], + "spans": [ + { + "bbox": [ + 55, + 204, + 295, + 287 + ], + "type": "text", + "content": "2. We discover that ImageNet-trained models (both CNN and transformer architectures) are subject to learn frequency shortcuts. Different from shortcuts formed by visual cues [9], frequency shortcuts (easy-to-learn features) lead to good performance on both in-distribution (ID) and OOD tests if they do not block models from learning other semantics (difficult features)." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 55, + 288, + 295, + 383 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 288, + 295, + 383 + ], + "spans": [ + { + "bbox": [ + 55, + 288, + 295, + 383 + ], + "type": "text", + "content": "3. Our HFSS enables a more comprehensive assessment of model generalizability by analyzing OOD data characteristics, specifically the presence of shortcuts. In existing evaluation frameworks, frequency shortcuts do not always impair OOD generalization performance. This emphasizes the importance of considering the role of shortcuts when designing future OOD evaluation benchmarks." + } + ] + } + ], + "index": 3 + } + ], + "sub_type": "text" + }, + { + "bbox": [ + 55, + 394, + 144, + 406 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 394, + 144, + 406 + ], + "spans": [ + { + "bbox": [ + 55, + 394, + 144, + 406 + ], + "type": "text", + "content": "2. Related works" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 55, + 415, + 295, + 653 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 415, + 295, + 653 + ], + "spans": [ + { + "bbox": [ + 55, + 415, + 295, + 653 + ], + "type": "text", + "content": "Shortcut learning. Models can learn shortcut solutions based on superficial correlations between data and ground truth [9, 11] to optimize training objectives with the least effort. This learning behavior is due to simplicity-bias [34], which is caused by inductive biases provided by gradient descent or components like ReLUs [35]. Shortcuts can be visual cues in the data like source tags and artificial markers [19, 27]. For instance, one fifth of the horse images from Pascal VOC dataset were found to contain a source tag, on which models rely as a discriminant feature to recognize horses [19]. Next to these visual cues, vision models are also subject to shortcuts implicitly existing in the frequency domain, which manifest as small sets of frequencies contributing significantly to image classification performance [37, 39]. The study in [39] explains why models exhibit a textures-bias for classification [8] from a frequency shortcut perspective, but does not provide further investigation on how frequency shortcuts affect the generalization and robustness performance of ImageNet-trained models, due to computational burden." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 55, + 654, + 295, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 654, + 295, + 713 + ], + "spans": [ + { + "bbox": [ + 55, + 654, + 295, + 713 + ], + "type": "text", + "content": "Relying on simple shortcut solutions could harm the generalization and robustness performance of models, as shortcut learning appearing in early training might block models from learning other semantics related to the tasks at hand [3, 16, 39, 41]. However, such reliance could also cre" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 313, + 72, + 553, + 96 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 72, + 553, + 96 + ], + "spans": [ + { + "bbox": [ + 313, + 72, + 553, + 96 + ], + "type": "text", + "content": "ates a false impression of good generalization when shortcuts are present in OOD test sets [1, 19, 39]." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 313, + 111, + 553, + 458 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 111, + 553, + 458 + ], + "spans": [ + { + "bbox": [ + 313, + 111, + 553, + 458 + ], + "type": "text", + "content": "Shortcut identification. Uncovering shortcuts helps understanding the generalizability of models on different out-of-distribution (OOD) data, explaining why models fail to generalize to or perform well on OOD data. It can provide shortcut-related prior knowledge to develop techniques that improve model generalization and robustness performance based on shortcut mitigation [27, 38]. However, common approaches are limited to identifying visually inspectable shortcuts such as text, watermark and color patches [19, 24, 27], using e.g. saliency maps [33]. Rather than directly identifying shortcuts in the data, in [2, 5] shortcut features present in individual images were quantified by assessing how difficult they were for models to learn. The authors in [28, 30, 32] uncovered shortcut features learned within representation space. The authors in [23] investigated shortcut learning by analyzing the relationship between semantic concepts using a knowledge graph, sharing a similar idea to [42]. Despite visual shortcut cues, there are non-observable shortcuts in the frequency domain, which are embedded in data characteristics. The work of [39] identified relevant frequencies to classification, which potentially contain shortcut information. As this technique measures the relevance of one frequency at a time, it requires computational costs increasing proportionally to the number of classes and image resolution. Thus, limited attention has been given to uncovering frequency shortcuts in large datasets, with existing analyses primarily focused on datasets containing a small number of classes [2, 37, 39, 42]." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 313, + 459, + 554, + 567 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 459, + 554, + 567 + ], + "spans": [ + { + "bbox": [ + 313, + 459, + 554, + 567 + ], + "type": "text", + "content": "Existing methods to identify shortcuts are limited to either manual inspection [19] or involve a time-consuming algorithm [37, 39]. Methods quantifying shortcut information or uncovering shortcut features primarily aim at mitigating shortcut learning, without explaining how shortcuts in the data impact model generalization performance. We are the first to enable frequency shortcut analysis of models trained on large-scale datasets and link them to generalization performance in different OOD settings." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 313, + 582, + 555, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 582, + 555, + 715 + ], + "spans": [ + { + "bbox": [ + 313, + 582, + 555, + 715 + ], + "type": "text", + "content": "OOD evaluation. The generalization and robustness performance of vision models are usually evaluated using extra data that is considered OOD, e.g. data collected at different time points [31], with different styles or renditions [8, 14, 36], with synthetic corruption effects [13, 17, 25, 40], and with adversarial noise [4, 12, 20, 43]. However, such benchmarks do not consider the impact of shortcuts learned by models and present in OOD test sets. This might ignore critical factors related to the generalization and robustness capabilities of models. The work in [7] explores model generalization from the perspectives of biases, e.g. texture," + } + ] + } + ], + "index": 11 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "text", + "content": "25199" + } + ] + } + ], + "index": 12 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 1 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 85, + 76, + 521, + 281 + ], + "blocks": [ + { + "bbox": [ + 85, + 76, + 521, + 281 + ], + "lines": [ + { + "bbox": [ + 85, + 76, + 521, + 281 + ], + "spans": [ + { + "bbox": [ + 85, + 76, + 521, + 281 + ], + "type": "image", + "image_path": "baac05343ec0eb17a342d50c17329627042080eed9e2ea497d7b84740056567b.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 54, + 288, + 555, + 312 + ], + "lines": [ + { + "bbox": [ + 54, + 288, + 555, + 312 + ], + "spans": [ + { + "bbox": [ + 54, + 288, + 555, + 312 + ], + "type": "text", + "content": "Figure 1. Scheme of HFSS. Starting from stage 2, we sample frequency patches from a random frequency subset searched in previous stage. This confines the size of search space. The white patches in the binary masks indicate sampled frequency patches." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "bbox": [ + 55, + 322, + 297, + 394 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 322, + 297, + 394 + ], + "spans": [ + { + "bbox": [ + 55, + 322, + 297, + 394 + ], + "type": "text", + "content": "shape and spectral biases. In this work, we investigate the impact of frequency shortcuts on model generalization abilities, establishing connections among these different biases. Our work provides insights into when frequency shortcuts yield good performance or are harmful to model generalization and robustness performance." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 55, + 403, + 111, + 415 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 403, + 111, + 415 + ], + "spans": [ + { + "bbox": [ + 55, + 403, + 111, + 415 + ], + "type": "text", + "content": "3. Method" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 55, + 424, + 296, + 533 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 424, + 296, + 533 + ], + "spans": [ + { + "bbox": [ + 55, + 424, + 296, + 533 + ], + "type": "text", + "content": "We propose a method to identify (non-visible) frequency shortcuts linked to intrinsic data characteristics rather than visual cues as in [27]. We reduce significantly computational demands compared to [39]. Our method is the first solution for shortcut identification that enables the analysis of models trained on datasets with a large number of samples and classes. In the following sections, we detail our measurement of shortcut learning in models and examine the impact of shortcuts on generalization." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 55, + 538, + 226, + 551 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 538, + 226, + 551 + ], + "spans": [ + { + "bbox": [ + 55, + 538, + 226, + 551 + ], + "type": "text", + "content": "3.1. Identifying frequency shortcuts" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 55, + 555, + 296, + 664 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 555, + 296, + 664 + ], + "spans": [ + { + "bbox": [ + 55, + 555, + 296, + 664 + ], + "type": "text", + "content": "We propose a method called hierarchical frequency shortcut search (HFSS), which exploits hierarchical search of frequency subsets in the image Fourier spectrum. This reduces computational time compared to exhaustive search strategies, e.g. [38, 39]. The search is separated into several stages, each stage gradually narrowing down the spectrum search space in a coarse-to-fine manner and detecting frequency subsets that models strongly rely on for classification. The scheme of HFSS is shown in Fig. 1." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 55, + 677, + 296, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 677, + 296, + 713 + ], + "spans": [ + { + "bbox": [ + 55, + 677, + 296, + 713 + ], + "type": "text", + "content": "Hierarchical search for frequency shortcuts. In order to discover the frequency subsets that models rely heavily on for correct classification of each class, we sample different" + } + ] + } + ], + "index": 7 + }, + { + "type": "image", + "bbox": [ + 375, + 321, + 484, + 422 + ], + "blocks": [ + { + "bbox": [ + 375, + 321, + 484, + 422 + ], + "lines": [ + { + "bbox": [ + 375, + 321, + 484, + 422 + ], + "spans": [ + { + "bbox": [ + 375, + 321, + 484, + 422 + ], + "type": "image", + "image_path": "68a3b6c4933f29d4b533355d62ccb9283cb622b4843470fba5881fa86dbba681.jpg" + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 313, + 424, + 555, + 447 + ], + "lines": [ + { + "bbox": [ + 313, + 424, + 555, + 447 + ], + "spans": [ + { + "bbox": [ + 313, + 424, + 555, + 447 + ], + "type": "text", + "content": "Figure 2. The frequency spectrum is separated into patches, with " + }, + { + "bbox": [ + 313, + 424, + 555, + 447 + ], + "type": "inline_equation", + "content": "p\\%" + }, + { + "bbox": [ + 313, + 424, + 555, + 447 + ], + "type": "text", + "content": " sampled for shortcut evaluation." + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "image_caption" + } + ], + "index": 8 + }, + { + "bbox": [ + 313, + 450, + 555, + 545 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 450, + 555, + 545 + ], + "spans": [ + { + "bbox": [ + 313, + 450, + 555, + 545 + ], + "type": "text", + "content": "frequency combinations that potentially contain frequency shortcuts. We use random sampling as it allows to easily consider joint contribution of frequencies to classification, improving the effectiveness of frequency shortcut identification. Random sampling has been shown to contribute to stable search results in data augmentation and reinforcement learning [21, 26, 29], while requiring lower computations w.r.t. optimization-based methods." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 313, + 546, + 556, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 546, + 556, + 714 + ], + "spans": [ + { + "bbox": [ + 313, + 546, + 556, + 714 + ], + "type": "text", + "content": "There are numerous combinations of frequencies by sampling frequency components directly. Instead of doing an exhaustive search, we apply hierarchical search in the Fourier spectrum: we divide the Fourier spectrum into frequency patches and sample these frequency patches to generate candidates of frequency subsets that contain shortcut information, as illustrated in Fig. 2. We exploit overlapping shifted windows in vertical and horizontal directions [22] to prepare frequency patches and thus avoid border effects, in addition to evenly separating the Fourier spectrum. HFSS consists of multiple search stages, the number of which depends on image resolution and desired frequency resolution of the identified frequency subsets. Lower image or frequency resolution requires fewer search stages." + } + ] + } + ], + "index": 11 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "type": "text", + "content": "25200" + } + ] + } + ], + "index": 12 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 2 + }, + { + "para_blocks": [ + { + "bbox": [ + 55, + 72, + 294, + 251 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 72, + 294, + 251 + ], + "spans": [ + { + "bbox": [ + 55, + 72, + 294, + 251 + ], + "type": "text", + "content": "In summary, the overall process of HFSS is as follows. Each stage of HFSS consists of three steps, namely (1) sampling " + }, + { + "bbox": [ + 55, + 72, + 294, + 251 + ], + "type": "inline_equation", + "content": "B" + }, + { + "bbox": [ + 55, + 72, + 294, + 251 + ], + "type": "text", + "content": " frequency subsets which contain " + }, + { + "bbox": [ + 55, + 72, + 294, + 251 + ], + "type": "inline_equation", + "content": "p\\%" + }, + { + "bbox": [ + 55, + 72, + 294, + 251 + ], + "type": "text", + "content": " frequency patches, (2) evaluating shortcut information, and (3) forwarding subsets for next stage. At the initial search stage, we sample frequency patches with a large size and evaluate their shortcut information. We evaluate the frequency subsets according to their contributions to classification: we process a subset of training images by retaining a frequency subset and measure the class-wise loss of a model. Frequency subsets that contribute to a lower loss value are stronger candidates to indicate frequency shortcuts as they are sufficient for the model to achieve a high prediction score. We rank frequency subsets by the loss values and use the top-" + }, + { + "bbox": [ + 55, + 72, + 294, + 251 + ], + "type": "inline_equation", + "content": "N" + }, + { + "bbox": [ + 55, + 72, + 294, + 251 + ], + "type": "text", + "content": " masks of each class in the next search stage." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 55, + 252, + 294, + 348 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 252, + 294, + 348 + ], + "spans": [ + { + "bbox": [ + 55, + 252, + 294, + 348 + ], + "type": "text", + "content": "Starting from the second search stage, we sample frequency patches from one of the top- " + }, + { + "bbox": [ + 55, + 252, + 294, + 348 + ], + "type": "inline_equation", + "content": "N" + }, + { + "bbox": [ + 55, + 252, + 294, + 348 + ], + "type": "text", + "content": " frequency subsets in the previous stage, with a smaller patch size. This confines the search space. In the final search stage, the top-1 frequency subset of each class is considered dominating the classification of the class concerned. Same as [39], we use binary masks to represent these frequency subsets, called Dominant Frequency Maps (DFMs)." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 55, + 349, + 294, + 397 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 349, + 294, + 397 + ], + "spans": [ + { + "bbox": [ + 55, + 349, + 294, + 397 + ], + "type": "text", + "content": "The sampling percentage " + }, + { + "bbox": [ + 55, + 349, + 294, + 397 + ], + "type": "inline_equation", + "content": "p\\%" + }, + { + "bbox": [ + 55, + 349, + 294, + 397 + ], + "type": "text", + "content": " of frequency patches, the number of sampled frequency subsets " + }, + { + "bbox": [ + 55, + 349, + 294, + 397 + ], + "type": "inline_equation", + "content": "B" + }, + { + "bbox": [ + 55, + 349, + 294, + 397 + ], + "type": "text", + "content": " and the frequency resolution of DFMs are hyperparameters of the search algorithm (details are in the supplementary material)." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 55, + 407, + 275, + 420 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 407, + 275, + 420 + ], + "spans": [ + { + "bbox": [ + 55, + 407, + 275, + 420 + ], + "type": "text", + "content": "3.2. Measuring the degree of shortcut learning" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 55, + 426, + 295, + 581 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 426, + 295, + 581 + ], + "spans": [ + { + "bbox": [ + 55, + 426, + 295, + 581 + ], + "type": "text", + "content": "Analyzing models learning behavior in a class-wise manner, as done in [39], is labor-intensive if there are hundreds or thousands of classes. To provide a broad overview of model learning behavior on training data, we categorize all classes in a dataset into two groups: (1) classes subject to shortcuts and (2) non-shortcut classes. A class is considered subject to shortcuts if its true positive rate (TPR) surpasses a given threshold when the model is tested on images filtered to retain only dominant frequencies of the class concerned. We use TPR, as a high TPR indicates that the frequency subset is sufficient to achieve a high classification rate, thereby acting as a shortcut. The frequency-filtered images are referred to as DFM-filtered images." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 55, + 582, + 295, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 582, + 295, + 715 + ], + "spans": [ + { + "bbox": [ + 55, + 582, + 295, + 715 + ], + "type": "text", + "content": "For each class, we test the model on both the original test images and DFM-filtered images, computing TPR for both. We denote the TPR on the original images of class " + }, + { + "bbox": [ + 55, + 582, + 295, + 715 + ], + "type": "inline_equation", + "content": "c_{i}" + }, + { + "bbox": [ + 55, + 582, + 295, + 715 + ], + "type": "text", + "content": " as " + }, + { + "bbox": [ + 55, + 582, + 295, + 715 + ], + "type": "inline_equation", + "content": "\\mathrm{TPR}_{c_i}" + }, + { + "bbox": [ + 55, + 582, + 295, + 715 + ], + "type": "text", + "content": " and the TPR on the DFM-filtered images as " + }, + { + "bbox": [ + 55, + 582, + 295, + 715 + ], + "type": "inline_equation", + "content": "\\mathrm{TPR}_{c_i}^{DFM}" + }, + { + "bbox": [ + 55, + 582, + 295, + 715 + ], + "type": "text", + "content": ". If " + }, + { + "bbox": [ + 55, + 582, + 295, + 715 + ], + "type": "inline_equation", + "content": "\\mathrm{TPR}_{c_i}^{DFM} > t" + }, + { + "bbox": [ + 55, + 582, + 295, + 715 + ], + "type": "text", + "content": " (where " + }, + { + "bbox": [ + 55, + 582, + 295, + 715 + ], + "type": "inline_equation", + "content": "t \\in [0,1]" + }, + { + "bbox": [ + 55, + 582, + 295, + 715 + ], + "type": "text", + "content": " is a predefined threshold), class " + }, + { + "bbox": [ + 55, + 582, + 295, + 715 + ], + "type": "inline_equation", + "content": "c_{i}" + }, + { + "bbox": [ + 55, + 582, + 295, + 715 + ], + "type": "text", + "content": " is considered subject to shortcuts; otherwise, it is considered as one of the non-shortcut classes. We compute the average TPR values of both groups, shortcut and non-shortcut classes, at different thresholds " + }, + { + "bbox": [ + 55, + 582, + 295, + 715 + ], + "type": "inline_equation", + "content": "t" + }, + { + "bbox": [ + 55, + 582, + 295, + 715 + ], + "type": "text", + "content": ". We note them as " + }, + { + "bbox": [ + 55, + 582, + 295, + 715 + ], + "type": "inline_equation", + "content": "\\mathrm{AvgTPR}_{\\mathrm{sct}}@t" + }, + { + "bbox": [ + 55, + 582, + 295, + 715 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 55, + 582, + 295, + 715 + ], + "type": "inline_equation", + "content": "\\mathrm{AvgTPR}_{\\mathrm{sct}}^{\\mathrm{DFM}}@t" + }, + { + "bbox": [ + 55, + 582, + 295, + 715 + ], + "type": "text", + "content": " for the shortcut classes, and " + }, + { + "bbox": [ + 55, + 582, + 295, + 715 + ], + "type": "inline_equation", + "content": "\\mathrm{AvgTPR}_{\\mathrm{non-sct}}@t" + }, + { + "bbox": [ + 55, + 582, + 295, + 715 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 55, + 582, + 295, + 715 + ], + "type": "inline_equation", + "content": "\\mathrm{AvgTPR}_{\\mathrm{non-sct}}^{\\mathrm{DFM}}@t" + } + ] + } + ], + "index": 5 + }, + { + "type": "image", + "bbox": [ + 357, + 73, + 511, + 153 + ], + "blocks": [ + { + "bbox": [ + 357, + 73, + 511, + 153 + ], + "lines": [ + { + "bbox": [ + 357, + 73, + 511, + 153 + ], + "spans": [ + { + "bbox": [ + 357, + 73, + 511, + 153 + ], + "type": "image", + "image_path": "3f0dca1220b8a62254257c0f47a68f10a49e2c576a43e4283ce75ab19f5d00ec.jpg" + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 313, + 156, + 553, + 244 + ], + "lines": [ + { + "bbox": [ + 313, + 156, + 553, + 244 + ], + "spans": [ + { + "bbox": [ + 313, + 156, + 553, + 244 + ], + "type": "text", + "content": "Figure 3. ResNet18 tested on original test images and DFM-filtered images of ImageNet-1k. Blue line shows the average TPR of classes subject to shortcuts on the original test images, and the orange line shows the results of non-shortcut classes, at different threshold " + }, + { + "bbox": [ + 313, + 156, + 553, + 244 + ], + "type": "inline_equation", + "content": "t" + }, + { + "bbox": [ + 313, + 156, + 553, + 244 + ], + "type": "text", + "content": ". The green and red lines correspond to results tested on DFM-filtered images. Lower " + }, + { + "bbox": [ + 313, + 156, + 553, + 244 + ], + "type": "inline_equation", + "content": "t" + }, + { + "bbox": [ + 313, + 156, + 553, + 244 + ], + "type": "text", + "content": " indicates weak shortcuts and higher " + }, + { + "bbox": [ + 313, + 156, + 553, + 244 + ], + "type": "inline_equation", + "content": "t" + }, + { + "bbox": [ + 313, + 156, + 553, + 244 + ], + "type": "text", + "content": " signifies stronger ones. The size of each point reflects the number of classes. The larger the size, the more classes included." + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "image_caption" + } + ], + "index": 6 + }, + { + "bbox": [ + 313, + 255, + 553, + 351 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 255, + 553, + 351 + ], + "spans": [ + { + "bbox": [ + 313, + 255, + 553, + 351 + ], + "type": "text", + "content": "for the non-shortcut classes. A higher threshold " + }, + { + "bbox": [ + 313, + 255, + 553, + 351 + ], + "type": "inline_equation", + "content": "t" + }, + { + "bbox": [ + 313, + 255, + 553, + 351 + ], + "type": "text", + "content": " corresponds to stronger shortcuts, as dominant frequencies alone are sufficient for accurate classification. An example of the results of these metrics at different threshold values are given in Fig. 3, where the size of each point reflects the number of classes in the two groups. Larger points indicate a higher number of classes, which suggests that the model has a stronger tendency to shortcut learning." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 313, + 351, + 553, + 412 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 351, + 553, + 412 + ], + "spans": [ + { + "bbox": [ + 313, + 351, + 553, + 412 + ], + "type": "text", + "content": "We do not average " + }, + { + "bbox": [ + 313, + 351, + 553, + 412 + ], + "type": "inline_equation", + "content": "\\mathrm{TPR}^{\\mathrm{DFM}}" + }, + { + "bbox": [ + 313, + 351, + 553, + 412 + ], + "type": "text", + "content": " across all classes because its value for many non-shortcut classes is close to zero (specifically for large datasets). This can result in a close-to-zero " + }, + { + "bbox": [ + 313, + 351, + 553, + 412 + ], + "type": "inline_equation", + "content": "\\mathrm{AvgTPR}^{\\mathrm{DFM}}" + }, + { + "bbox": [ + 313, + 351, + 553, + 412 + ], + "type": "text", + "content": " which does not provide useful information for further analysis." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 313, + 421, + 507, + 433 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 421, + 507, + 433 + ], + "spans": [ + { + "bbox": [ + 313, + 421, + 507, + 433 + ], + "type": "text", + "content": "3.3. Impact of shortcut on generalization" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 313, + 438, + 555, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 438, + 555, + 714 + ], + "spans": [ + { + "bbox": [ + 313, + 438, + 555, + 714 + ], + "type": "text", + "content": "To assess how frequency shortcuts impact the generalization and robustness performance of models trained on ImageNet-1k, we evaluate their performance on several datasets. We perform ID tests on ImageNet-1k (IN-1k) [6] and ImageNet-v2 (IN-v2) [31] test sets. They are considered in-distribution as they are collected using the same protocol as the training set. Furthermore, we use ImageNet-C (IN-C) [13] to benchmark the robustness of models against appearance image corruptions, e.g. noise, blur and weather changes. To evaluate the generalizability of models to images with different renditions, we use ImageNet-Renditions (IN-R) [14] and ImageNet-Sketch (IN-S) [36]. IN-R contains renditions like cartoons, paintings, art and toys of 200 classes from IN-1k. IN-S has the same number of classes as IN-1k, containing images of hand-drawn sketches. These datasets contain visual renditions, which serve to evaluate how the reliance on texture cues (corresponding to most frequency shortcuts) impacts on generalization. To measure the adversarial robustness of models, we apply the fast gradient sign method (FGSM) attacks [10] to the validation set of IN-1k, with " + }, + { + "bbox": [ + 313, + 438, + 555, + 714 + ], + "type": "inline_equation", + "content": "L_{\\infty} = 4 / 255" + }, + { + "bbox": [ + 313, + 438, + 555, + 714 + ], + "type": "text", + "content": ". We calculate the average TPR of shortcut and non-shortcut classes on OOD data, that we compare to the results on ID data." + } + ] + } + ], + "index": 11 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "text", + "content": "25201" + } + ] + } + ], + "index": 12 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 3 + }, + { + "para_blocks": [ + { + "bbox": [ + 55, + 71, + 138, + 85 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 71, + 138, + 85 + ], + "spans": [ + { + "bbox": [ + 55, + 71, + 138, + 85 + ], + "type": "text", + "content": "4. Experiments" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 55, + 91, + 296, + 177 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 91, + 296, + 177 + ], + "spans": [ + { + "bbox": [ + 55, + 91, + 296, + 177 + ], + "type": "text", + "content": "We start with investigating the trade-off between efficiency (i.e. required computational time) and effectiveness (i.e. capability of finding shortcuts), varying the configuration of the number of sampled frequency subsets at each stage, noted as " + }, + { + "bbox": [ + 55, + 91, + 296, + 177 + ], + "type": "inline_equation", + "content": "B_{s}" + }, + { + "bbox": [ + 55, + 91, + 296, + 177 + ], + "type": "text", + "content": " (" + }, + { + "bbox": [ + 55, + 91, + 296, + 177 + ], + "type": "inline_equation", + "content": "s" + }, + { + "bbox": [ + 55, + 91, + 296, + 177 + ], + "type": "text", + "content": " is the index of stage). Then, we analyze frequency shortcut learning of models trained on IN-1k, relating it to performance results under different OOD scenarios." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 56, + 194, + 296, + 483 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 194, + 296, + 483 + ], + "spans": [ + { + "bbox": [ + 56, + 194, + 296, + 483 + ], + "type": "text", + "content": "Setup. We use CIFAR-10 (C-10) [18] to explore the configuration of " + }, + { + "bbox": [ + 56, + 194, + 296, + 483 + ], + "type": "inline_equation", + "content": "B" + }, + { + "bbox": [ + 56, + 194, + 296, + 483 + ], + "type": "text", + "content": ", and configure HFSS with four search stages, with patch size of " + }, + { + "bbox": [ + 56, + 194, + 296, + 483 + ], + "type": "inline_equation", + "content": "8 \\times 8" + }, + { + "bbox": [ + 56, + 194, + 296, + 483 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 56, + 194, + 296, + 483 + ], + "type": "inline_equation", + "content": "4 \\times 4" + }, + { + "bbox": [ + 56, + 194, + 296, + 483 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 56, + 194, + 296, + 483 + ], + "type": "inline_equation", + "content": "2 \\times 2" + }, + { + "bbox": [ + 56, + 194, + 296, + 483 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 56, + 194, + 296, + 483 + ], + "type": "inline_equation", + "content": "1 \\times 1" + }, + { + "bbox": [ + 56, + 194, + 296, + 483 + ], + "type": "text", + "content": ", respectively. We then perform larger-scale shortcut analysis using IN-1k: we use six search stages, with patch size of " + }, + { + "bbox": [ + 56, + 194, + 296, + 483 + ], + "type": "inline_equation", + "content": "56 \\times 56" + }, + { + "bbox": [ + 56, + 194, + 296, + 483 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 56, + 194, + 296, + 483 + ], + "type": "inline_equation", + "content": "28 \\times 28" + }, + { + "bbox": [ + 56, + 194, + 296, + 483 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 56, + 194, + 296, + 483 + ], + "type": "inline_equation", + "content": "14 \\times 14" + }, + { + "bbox": [ + 56, + 194, + 296, + 483 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 56, + 194, + 296, + 483 + ], + "type": "inline_equation", + "content": "8 \\times 8" + }, + { + "bbox": [ + 56, + 194, + 296, + 483 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 56, + 194, + 296, + 483 + ], + "type": "inline_equation", + "content": "4 \\times 4" + }, + { + "bbox": [ + 56, + 194, + 296, + 483 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 56, + 194, + 296, + 483 + ], + "type": "inline_equation", + "content": "2 \\times 2" + }, + { + "bbox": [ + 56, + 194, + 296, + 483 + ], + "type": "text", + "content": ". The patch (frequency) resolution in the final stage is the same as that in [39]. At each stage, we sample " + }, + { + "bbox": [ + 56, + 194, + 296, + 483 + ], + "type": "inline_equation", + "content": "p = 60\\%" + }, + { + "bbox": [ + 56, + 194, + 296, + 483 + ], + "type": "text", + "content": " frequency patches to form frequency subsets. It results in DFMs containing about " + }, + { + "bbox": [ + 56, + 194, + 296, + 483 + ], + "type": "inline_equation", + "content": "5\\%" + }, + { + "bbox": [ + 56, + 194, + 296, + 483 + ], + "type": "text", + "content": " frequencies of the whole spectrum of ImageNet images and about " + }, + { + "bbox": [ + 56, + 194, + 296, + 483 + ], + "type": "inline_equation", + "content": "15\\%" + }, + { + "bbox": [ + 56, + 194, + 296, + 483 + ], + "type": "text", + "content": " frequencies of CIFAR images. The number of training images sampled for shortcut information evaluation is the same as that of their corresponding test set. We evaluate the generalization and robustness performance of ImageNet-trained models on IN-v2 [31], IN-C [13], IN-R [14], IN-S [36] and using FGSM [10] attacks. We also carry out experiments on ImageNet-10 (IN-10) [15] with a ResNet18 model to compare HFSS with single-frequency removal-based method in [39]. Following their setups, we use ImageNet-SCT (IN-SCT) for OOD evaluation. Our configuration of HFSS enables direct comparison with the results reported in [39] as DFMs contain around " + }, + { + "bbox": [ + 56, + 194, + 296, + 483 + ], + "type": "inline_equation", + "content": "5\\%" + }, + { + "bbox": [ + 56, + 194, + 296, + 483 + ], + "type": "text", + "content": " frequencies of whole spectrum. We report training configurations and additional results in the supplementary material." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 55, + 491, + 187, + 504 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 491, + 187, + 504 + ], + "spans": [ + { + "bbox": [ + 55, + 491, + 187, + 504 + ], + "type": "text", + "content": "4.1. Configuration of HFSS" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 55, + 510, + 296, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 510, + 296, + 714 + ], + "spans": [ + { + "bbox": [ + 55, + 510, + 296, + 714 + ], + "type": "text", + "content": "Sampled frequency subsets. At stage " + }, + { + "bbox": [ + 55, + 510, + 296, + 714 + ], + "type": "inline_equation", + "content": "s" + }, + { + "bbox": [ + 55, + 510, + 296, + 714 + ], + "type": "text", + "content": ", we sample " + }, + { + "bbox": [ + 55, + 510, + 296, + 714 + ], + "type": "inline_equation", + "content": "B_{s}" + }, + { + "bbox": [ + 55, + 510, + 296, + 714 + ], + "type": "text", + "content": " frequency subsets, such that HFSS can explore different combinations of frequency patterns, validating their contained shortcut information. However, sampling an excessive number of subsets increases the required search time, as more subsets are for evaluating their contributions to classification results. To explore the impact of the number of sampling operation " + }, + { + "bbox": [ + 55, + 510, + 296, + 714 + ], + "type": "inline_equation", + "content": "B_{s}" + }, + { + "bbox": [ + 55, + 510, + 296, + 714 + ], + "type": "text", + "content": " on shortcut identification, we perform experiments on C-10 using a ResNet18 backbone, setting the number of sampled frequency subsets in each stage to be as large as possible under limited computational time. We hypothesize that HFSS obtains stable results given enough candidate frequency subsets. We set " + }, + { + "bbox": [ + 55, + 510, + 296, + 714 + ], + "type": "inline_equation", + "content": "B_{1}, B_{2}, B_{3}" + }, + { + "bbox": [ + 55, + 510, + 296, + 714 + ], + "type": "text", + "content": ", and " + }, + { + "bbox": [ + 55, + 510, + 296, + 714 + ], + "type": "inline_equation", + "content": "B_{4}" + }, + { + "bbox": [ + 55, + 510, + 296, + 714 + ], + "type": "text", + "content": " to be 1000, 2000, 4000 and 8000 respectively, which we note as configuration CF-1. The increasing number of candidate frequency subsets at consecutive stages is determined by the increase in possible frequency patch combinations as patch" + } + ] + } + ], + "index": 4 + }, + { + "type": "image", + "bbox": [ + 319, + 72, + 552, + 296 + ], + "blocks": [ + { + "bbox": [ + 319, + 72, + 552, + 296 + ], + "lines": [ + { + "bbox": [ + 319, + 72, + 552, + 296 + ], + "spans": [ + { + "bbox": [ + 319, + 72, + 552, + 296 + ], + "type": "image", + "image_path": "3a9d3b140d81d1445e08aec2994b8cf68392596b766f0d54a2e1dfb0b1f501fd.jpg" + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 313, + 300, + 555, + 323 + ], + "lines": [ + { + "bbox": [ + 313, + 300, + 555, + 323 + ], + "spans": [ + { + "bbox": [ + 313, + 300, + 555, + 323 + ], + "type": "text", + "content": "Figure 4. The best class-wise loss vs. the number of sampled frequency subsets at each stage." + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "image_caption" + } + ], + "index": 5 + }, + { + "bbox": [ + 313, + 328, + 376, + 338 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 328, + 376, + 338 + ], + "spans": [ + { + "bbox": [ + 313, + 328, + 376, + 338 + ], + "type": "text", + "content": "size decreases." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 313, + 340, + 555, + 640 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 340, + 555, + 640 + ], + "spans": [ + { + "bbox": [ + 313, + 340, + 555, + 640 + ], + "type": "text", + "content": "We run the configuration CF-1 five times and note the stability of our HFSS search algorithms. We track the best (lowest) loss values achieved for each class when the model is tested on DFM-filtered images. In Fig. 4, we report the tracking statistics with average loss values (lines) and standard deviations (shadows) over the five trials for each class, which are further used as a starting point to optimize the configuration of HFSS. As the number of sampled frequency subsets increases, the best loss for each class converges, indicating the stability of HFSS in identifying similar shortcuts learned by models when a sufficient amount of candidate frequency subsets are sampled for validating shortcut information. The standard deviation recorded at different stages is relatively small compared to the range of best loss. For class airplane we observe an outlier outcome at stage 3. In the supplementary material, we present visualization of shortcut cues (DFM-filtered images) computed across the five trials. These visualizations show that the identified shortcut patterns differ in the orientations of strip-like features, corresponding to line features in the image of airplane. This suggests that frequency shortcuts do not arise from fixed frequency subsets; rather, spatial features they correspond to may be visually similar but composed of different frequencies, determining an higher standard deviation in Fig. 4." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 313, + 654, + 556, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 654, + 556, + 715 + ], + "spans": [ + { + "bbox": [ + 313, + 654, + 556, + 715 + ], + "type": "text", + "content": "Efficiency vs. effectiveness. Based on the results in Fig. 4, we optimize CF-1 for efficiency and effectiveness in shortcut identification. The required search time scales up proportionally to " + }, + { + "bbox": [ + 313, + 654, + 556, + 715 + ], + "type": "inline_equation", + "content": "B_{s}" + }, + { + "bbox": [ + 313, + 654, + 556, + 715 + ], + "type": "text", + "content": ", i.e. the more candidate frequency subsets are sampled, the more time is needed to verify their" + } + ] + } + ], + "index": 9 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "type": "text", + "content": "25202" + } + ] + } + ], + "index": 10 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 4 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 107, + 72, + 242, + 171 + ], + "blocks": [ + { + "bbox": [ + 107, + 72, + 242, + 171 + ], + "lines": [ + { + "bbox": [ + 107, + 72, + 242, + 171 + ], + "spans": [ + { + "bbox": [ + 107, + 72, + 242, + 171 + ], + "type": "image", + "image_path": "44804d03646bcc9d1c3f209e82d0f03083b58153e96f1f4e5679eb69fb893f86.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 55, + 172, + 295, + 204 + ], + "lines": [ + { + "bbox": [ + 55, + 172, + 295, + 204 + ], + "spans": [ + { + "bbox": [ + 55, + 172, + 295, + 204 + ], + "type": "text", + "content": "Figure 5. Average TPR vs. search time, where search time increases proportionally with the number of sampled frequency subsets." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "type": "image", + "bbox": [ + 110, + 206, + 239, + 304 + ], + "blocks": [ + { + "bbox": [ + 110, + 206, + 239, + 304 + ], + "lines": [ + { + "bbox": [ + 110, + 206, + 239, + 304 + ], + "spans": [ + { + "bbox": [ + 110, + 206, + 239, + 304 + ], + "type": "image", + "image_path": "c2ad637d9e710d9e444390052802dbdc77589e53c0ac22f3c0809fbc2c6ee9d1.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 55, + 306, + 295, + 350 + ], + "lines": [ + { + "bbox": [ + 55, + 306, + 295, + 350 + ], + "spans": [ + { + "bbox": [ + 55, + 306, + 295, + 350 + ], + "type": "text", + "content": "Figure 6. The number of shortcut classes given different threshold " + }, + { + "bbox": [ + 55, + 306, + 295, + 350 + ], + "type": "inline_equation", + "content": "t" + }, + { + "bbox": [ + 55, + 306, + 295, + 350 + ], + "type": "text", + "content": " search by CF-1 and CF-2.10. Using CF-2.10 uncovers most shortcuts identified by using CF-1 at low thresholds, with computational time reduced to a factor of around 200." + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_caption" + } + ], + "index": 2 + }, + { + "bbox": [ + 55, + 364, + 295, + 483 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 364, + 295, + 483 + ], + "spans": [ + { + "bbox": [ + 55, + 364, + 295, + 483 + ], + "type": "text", + "content": "relevance to classification. We perform experiments using different " + }, + { + "bbox": [ + 55, + 364, + 295, + 483 + ], + "type": "inline_equation", + "content": "B_{s}" + }, + { + "bbox": [ + 55, + 364, + 295, + 483 + ], + "type": "text", + "content": ", investigating how the size of the search space affects the identification of shortcuts. Based on the tracking statistics in Fig. 4, which shows that increasing " + }, + { + "bbox": [ + 55, + 364, + 295, + 483 + ], + "type": "inline_equation", + "content": "B_{s}" + }, + { + "bbox": [ + 55, + 364, + 295, + 483 + ], + "type": "text", + "content": " beyond a certain point does not decrease loss significantly, we adjust " + }, + { + "bbox": [ + 55, + 364, + 295, + 483 + ], + "type": "inline_equation", + "content": "B_{1}" + }, + { + "bbox": [ + 55, + 364, + 295, + 483 + ], + "type": "text", + "content": " to 200, " + }, + { + "bbox": [ + 55, + 364, + 295, + 483 + ], + "type": "inline_equation", + "content": "B_{2}" + }, + { + "bbox": [ + 55, + 364, + 295, + 483 + ], + "type": "text", + "content": " to 800 and " + }, + { + "bbox": [ + 55, + 364, + 295, + 483 + ], + "type": "inline_equation", + "content": "B_{4}" + }, + { + "bbox": [ + 55, + 364, + 295, + 483 + ], + "type": "text", + "content": " to 4000 (CF-2.1). We further reduce " + }, + { + "bbox": [ + 55, + 364, + 295, + 483 + ], + "type": "inline_equation", + "content": "B_{3}" + }, + { + "bbox": [ + 55, + 364, + 295, + 483 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 55, + 364, + 295, + 483 + ], + "type": "inline_equation", + "content": "B_{4}" + }, + { + "bbox": [ + 55, + 364, + 295, + 483 + ], + "type": "text", + "content": " to 2000 (CF-2.2). We also design eight other configurations (see details in the supplementary material), with CF-2.10 being the most efficient as it samples the lowest number of candidate frequency subsets." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 55, + 484, + 295, + 675 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 484, + 295, + 675 + ], + "spans": [ + { + "bbox": [ + 55, + 484, + 295, + 675 + ], + "type": "text", + "content": "To compare the performance of HFSS under different configurations, we compute the average TPR over all classes when tested on DFM-filtered images. This indicates the general relevance of the searched frequency subsets to classification. A higher average TPR indicate higher relevance, thus the frequency subsets contain more frequency shortcut information. We run each configuration five times, computing the mean and standard deviations of average TPRs over the five trials. We report the results in Fig. 5 where the left-most point corresponds to average TPR of CF-2.10 and the right-most point corresponds to that of CF-1. We observe that as more frequency subsets are sampled, average TPR increases and saturates. This demonstrates that sampling more candidate frequency subsets does not significantly improve the performance of HFSS, while requiring more computations." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 55, + 677, + 295, + 712 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 677, + 295, + 712 + ], + "spans": [ + { + "bbox": [ + 55, + 677, + 295, + 712 + ], + "type": "text", + "content": "We compare CF-1 and CF-2.10 in Fig. 6 based on the number of classes with identified shortcuts at different thresholds. We observe that CF-2.10 uncovers most of the" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 313, + 72, + 553, + 156 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 72, + 553, + 156 + ], + "spans": [ + { + "bbox": [ + 313, + 72, + 553, + 156 + ], + "type": "text", + "content": "shortcuts identified by using the more complex CF-1 at low threshold level. Although CF-2.10 misses stronger shortcuts (at higher threshold levels), it achieves a " + }, + { + "bbox": [ + 313, + 72, + 553, + 156 + ], + "type": "inline_equation", + "content": "\\sim 200\\times" + }, + { + "bbox": [ + 313, + 72, + 553, + 156 + ], + "type": "text", + "content": " reduction in computational time compared to CF-1. A more time-efficient configuration is less effective at finding shortcuts than the more complex one, but it still manages to identify shortcuts at low thresholds." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 313, + 156, + 553, + 264 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 156, + 553, + 264 + ], + "spans": [ + { + "bbox": [ + 313, + 156, + 553, + 264 + ], + "type": "text", + "content": "The observations made on C-10 allow to estimate initial configurations for analysis on larger-scale datasets. We track the loss statistics of HFSS applied to ImageNet (provided in the supplementary material) and choose the value " + }, + { + "bbox": [ + 313, + 156, + 553, + 264 + ], + "type": "inline_equation", + "content": "B_{s}" + }, + { + "bbox": [ + 313, + 156, + 553, + 264 + ], + "type": "text", + "content": " for ImageNet experiments in Sec. 4.2 as the one where the loss values show no significant decrease. Given the computational intensity of identifying shortcuts in large-scale datasets, our work on IN-1k primarily aims at uncovering shortcuts with affordable computational efforts." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 313, + 273, + 487, + 284 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 273, + 487, + 284 + ], + "spans": [ + { + "bbox": [ + 313, + 273, + 487, + 284 + ], + "type": "text", + "content": "4.2. Shortcut identification on IN-1k" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 313, + 290, + 553, + 350 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 290, + 553, + 350 + ], + "spans": [ + { + "bbox": [ + 313, + 290, + 553, + 350 + ], + "type": "text", + "content": "We apply HFSS to examine frequency shortcuts learned by ImageNet models and how they impact model generalization and robustness. The search configurations for ImageNet models, aimed at time-efficiency and effectiveness at finding shortcuts, are in the supplementary material." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 313, + 366, + 553, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 366, + 553, + 713 + ], + "spans": [ + { + "bbox": [ + 313, + 366, + 553, + 713 + ], + "type": "text", + "content": "Results on IN-1k, IN-v2 and IN-C. In Figs. 7 and 8, we report the results of the analysis of impact of shortcuts on model performance. In general, models perform better predictions on images of classes affected by shortcut learning (blue lines) compared to non-shortcut classes (orange lines) in both ID and corruption tests. Models subject to frequency shortcuts yield good performance in the robustness and generalization tests when textures information is largely preserved. For ResNet and ViT models, when " + }, + { + "bbox": [ + 313, + 366, + 553, + 713 + ], + "type": "inline_equation", + "content": "t \\geq 0.8" + }, + { + "bbox": [ + 313, + 366, + 553, + 713 + ], + "type": "text", + "content": ", the AvgTPR" + }, + { + "bbox": [ + 313, + 366, + 553, + 713 + ], + "type": "inline_equation", + "content": "_{\\text{sct}}^{\\text{DFM}}" + }, + { + "bbox": [ + 313, + 366, + 553, + 713 + ], + "type": "text", + "content": " exceeds AvgTPR" + }, + { + "bbox": [ + 313, + 366, + 553, + 713 + ], + "type": "inline_equation", + "content": "_{\\text{sct}}" + }, + { + "bbox": [ + 313, + 366, + 553, + 713 + ], + "type": "text", + "content": ", meaning that these models rely predominantly on dominant frequencies (corresponding to simple features) for classification of shortcut-affected classes, preventing from learning class-related semantics features. For instance, at " + }, + { + "bbox": [ + 313, + 366, + 553, + 713 + ], + "type": "inline_equation", + "content": "t = 0.9" + }, + { + "bbox": [ + 313, + 366, + 553, + 713 + ], + "type": "text", + "content": ", only one class ('window screen') is found subject to frequency shortcuts: ResNet18 achieves TPR=0.66 on full-spectrum images of this class and TPR=0.94 on DFM-filtered images (similarly to ResNet50 and ViT-b). Differently, CCT exhibits lower AvgTPR" + }, + { + "bbox": [ + 313, + 366, + 553, + 713 + ], + "type": "inline_equation", + "content": "_{\\text{sct}}^{\\text{DFM}}@0.9" + }, + { + "bbox": [ + 313, + 366, + 553, + 713 + ], + "type": "text", + "content": " than AvgTPR" + }, + { + "bbox": [ + 313, + 366, + 553, + 713 + ], + "type": "inline_equation", + "content": "_{\\text{sct}}@0.9" + }, + { + "bbox": [ + 313, + 366, + 553, + 713 + ], + "type": "text", + "content": ", indicating that while frequency shortcuts contribute significantly to classification, this model also manage to leverage other semantic information: a strong frequency shortcut (e.g. when " + }, + { + "bbox": [ + 313, + 366, + 553, + 713 + ], + "type": "inline_equation", + "content": "t > 0.7" + }, + { + "bbox": [ + 313, + 366, + 553, + 713 + ], + "type": "text", + "content": ") does not necessarily mean that it is the only information a model uses for classification. We further compare CCT with ResNet50 as they perform similarly on IN-1k (80.57% and 80.1% respectively). In IN-v2 and IN-C, CCT achieves 74.81% and 57.73% prediction rates, having better performance than ResNet50 (74.17% in IN-v2 and 48.85% in IN-C). Considering the highest de" + } + ] + } + ], + "index": 11 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "text", + "content": "25203" + } + ] + } + ], + "index": 12 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 5 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 82, + 71, + 523, + 323 + ], + "blocks": [ + { + "bbox": [ + 82, + 71, + 523, + 323 + ], + "lines": [ + { + "bbox": [ + 82, + 71, + 523, + 323 + ], + "spans": [ + { + "bbox": [ + 82, + 71, + 523, + 323 + ], + "type": "image", + "image_path": "2f410602a5803270cfd3b01ddd032747c75980460b1f0551cdc1e2c9fd9d3d13.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 54, + 326, + 555, + 360 + ], + "lines": [ + { + "bbox": [ + 54, + 326, + 555, + 360 + ], + "spans": [ + { + "bbox": [ + 54, + 326, + 555, + 360 + ], + "type": "text", + "content": "Figure 7. Average TPR of shortcut and non-shortcut classes given different thresholds on datasets with 1000 classes. Models generally perform better on images of shortcut classes than non-shortcut classes. This does not hold for results on IN-S, which lacks preserved texture information." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "type": "image", + "bbox": [ + 81, + 369, + 265, + 612 + ], + "blocks": [ + { + "bbox": [ + 81, + 369, + 265, + 612 + ], + "lines": [ + { + "bbox": [ + 81, + 369, + 265, + 612 + ], + "spans": [ + { + "bbox": [ + 81, + 369, + 265, + 612 + ], + "type": "image", + "image_path": "e7eb7c03bb47cb73c67ca58eec8e315e6fd629ae511eca91da914eafbf84387e.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 55, + 616, + 297, + 672 + ], + "lines": [ + { + "bbox": [ + 55, + 616, + 297, + 672 + ], + "spans": [ + { + "bbox": [ + 55, + 616, + 297, + 672 + ], + "type": "text", + "content": "Figure 8. Average TPR of shortcut and non-shortcut classes at different thresholds on datasets with 200 classes. Models perform better on shortcut classes than non-shortcut classes in IN-200. IN-R lacks preserved texture information and thus the models have worse performance on shortcut classes than non-shortcut classes." + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_caption" + } + ], + "index": 2 + }, + { + "bbox": [ + 55, + 677, + 296, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 677, + 296, + 715 + ], + "spans": [ + { + "bbox": [ + 55, + 677, + 296, + 715 + ], + "type": "text", + "content": "gree of shortcut learning of CCT among the models (indicated by the marker size), inducing frequency reliance without blocking the learning of other semantic features appears" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 313, + 370, + 555, + 395 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 370, + 555, + 395 + ], + "spans": [ + { + "bbox": [ + 313, + 370, + 555, + 395 + ], + "type": "text", + "content": "to benefit model robustness and generalization performance under statistical distribution shifts and corruption scenarios." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 313, + 404, + 556, + 584 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 404, + 556, + 584 + ], + "spans": [ + { + "bbox": [ + 313, + 404, + 556, + 584 + ], + "type": "text", + "content": "Results on IN-S and IN-R. Different from the results on IN-v2 and IN-C, frequency shortcuts impair generalization of models to texture and rendition changes. Model performance on shortcut and non-shortcut classes on IN-S is close to each other. In Fig. 8, we observe that shortcut classes have worse prediction results than non-shortcut classes when tested on IN-R which contains images with rendition changes. This is attributable to the fact that shortcut information are not available in the OOD tests, as the rendition and sketch test sets preserve less or very different texture information than IN-C and IN-v2 (see green lines of IN-S and IN-1k, IN-R and IN-200 in Fig. 7 and Fig. 8). Frequency shortcut learning is an explicit cause of the texture-bias of ImageNet models, and the resulting impaired generalizability to renditions is in line with [8]." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 313, + 594, + 557, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 594, + 557, + 713 + ], + "spans": [ + { + "bbox": [ + 313, + 594, + 557, + 713 + ], + "type": "text", + "content": "Results on FGSM attacks. As shown in Fig. 7, the four models, under adversarial attacks, achieve higher AvgTPR for shortcut classes (especially those with strong shortcuts) than non-shortcut classes. This suggests that models can be inherently robust to adversarial noise if they leverage frequency shortcuts for classification. This is not surprising as adversarial noises hardly manipulate the textures of images. CCT, which shows the highest degree of frequency shortcut learning among the considered models, achieves the best adversarial robustness. This indicates that model reliance" + } + ] + } + ], + "index": 7 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "type": "text", + "content": "25204" + } + ] + } + ], + "index": 8 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 6 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 66, + 77, + 289, + 224 + ], + "blocks": [ + { + "bbox": [ + 66, + 77, + 289, + 224 + ], + "lines": [ + { + "bbox": [ + 66, + 77, + 289, + 224 + ], + "spans": [ + { + "bbox": [ + 66, + 77, + 289, + 224 + ], + "type": "image", + "image_path": "3d19b165873f3c16216ae597e66203f09ef63f5b145986049d1263da92b5650b.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 55, + 230, + 295, + 286 + ], + "lines": [ + { + "bbox": [ + 55, + 230, + 295, + 286 + ], + "spans": [ + { + "bbox": [ + 55, + 230, + 295, + 286 + ], + "type": "text", + "content": "Figure 9. An image of dugong contains frequency shortcuts which a model relies on for high-confidence classification. Its presence in OOD test sets like IN-C, and in the image under FGSM attacks yields correct predictions. But the reliance on shortcuts impairs model generalization performance to renditions changes." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "bbox": [ + 55, + 297, + 296, + 332 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 297, + 296, + 332 + ], + "spans": [ + { + "bbox": [ + 55, + 297, + 296, + 332 + ], + "type": "text", + "content": "on frequency shortcuts can benefit their adversarial robustness, as the simple shortcut features are robust to adversarial noise." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 54, + 340, + 295, + 483 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 54, + 340, + 295, + 483 + ], + "spans": [ + { + "bbox": [ + 54, + 340, + 295, + 483 + ], + "type": "text", + "content": "Summary. Whether frequency shortcuts impair or benefit model generalization and robustness performance depends on the specific OOD scenarios. In Fig. 9, we show an example image of dugong and model predictions in OOD tests. The preserved texture information yields correct predictions on IN-C and under FGSM attacks, but it is not helpful for model generalizability to rendition changes. The findings reveal a limitation in the design of current OOD benchmarks, which overlooks the impact of frequency shortcuts on generalizability of models and its relation to specific characteristics of OOD data. Using HFSS for performance evaluation can bridge such gap." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 55, + 492, + 263, + 506 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 492, + 263, + 506 + ], + "spans": [ + { + "bbox": [ + 55, + 492, + 263, + 506 + ], + "type": "text", + "content": "4.3. Comparison with the existing approach" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 55, + 510, + 295, + 676 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 510, + 295, + 676 + ], + "spans": [ + { + "bbox": [ + 55, + 510, + 295, + 676 + ], + "type": "text", + "content": "More effective at finding shortcuts. We perform experiments on IN-10 for a direct comparison with the only existing method for frequency shortcut analysis [39], which evaluates the contribution of single frequencies to shortcuts and is limited to analyzing models trained on small-scale datasets. We report results in Tab. 2. HFSS uncovers strong frequency shortcuts for classes airliner, Siamese cat (Siam-cat), ox, frog, zebra, and container ship (Con-ship), while [39] was only able to find strong frequency shortcuts for classes zebra and Siamese cat. The better effectiveness of HFSS at finding shortcuts is attributable to the fact that HFSS considers the joint contributions of frequencies to classification, while in [39] the relevance of a single frequency was measured iteratively." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 55, + 677, + 296, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 677, + 296, + 713 + ], + "spans": [ + { + "bbox": [ + 55, + 677, + 296, + 713 + ], + "type": "text", + "content": "Frequency shortcuts can impair the generalization performance of models or provide a false impression of good generalization when shortcuts exist in OOD tests [39]. We" + } + ] + } + ], + "index": 6 + }, + { + "type": "table", + "bbox": [ + 316, + 102, + 552, + 141 + ], + "blocks": [ + { + "bbox": [ + 313, + 70, + 555, + 93 + ], + "lines": [ + { + "bbox": [ + 313, + 70, + 555, + 93 + ], + "spans": [ + { + "bbox": [ + 313, + 70, + 555, + 93 + ], + "type": "text", + "content": "Table 2. TPR results on DFM-filtered IN-10 images. TPR " + }, + { + "bbox": [ + 313, + 70, + 555, + 93 + ], + "type": "inline_equation", + "content": "\\geq 0.6" + }, + { + "bbox": [ + 313, + 70, + 555, + 93 + ], + "type": "text", + "content": " (a strong frequency shortcut) is highlighted in bold." + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 316, + 102, + 552, + 141 + ], + "lines": [ + { + "bbox": [ + 316, + 102, + 552, + 141 + ], + "spans": [ + { + "bbox": [ + 316, + 102, + 552, + 141 + ], + "type": "table", + "html": "
Class Methodairlinerwagonhum-birdSiam-catoxgolden retrieverfrogzebraCon-shiptruck
HFSS [39]0.8800.520.980.90.260.780.70.760.5
0.0800.40.80.020.020.140.80.540.06
", + "image_path": "d2fb9469e295ee5f909ac5df48118247ada223515a6dc3eb98049298ae78c966.jpg" + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "table_body" + } + ], + "index": 8 + }, + { + "type": "table", + "bbox": [ + 316, + 193, + 552, + 239 + ], + "blocks": [ + { + "bbox": [ + 313, + 151, + 555, + 185 + ], + "lines": [ + { + "bbox": [ + 313, + 151, + 555, + 185 + ], + "spans": [ + { + "bbox": [ + 313, + 151, + 555, + 185 + ], + "type": "text", + "content": "Table 3. TPR results on IN-SCT (first row) and corresponding DFM-filtered images (second and third rows). TPR above average TPR (0.367) is highlighted in bold." + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 316, + 193, + 552, + 239 + ], + "lines": [ + { + "bbox": [ + 316, + 193, + 552, + 239 + ], + "spans": [ + { + "bbox": [ + 316, + 193, + 552, + 239 + ], + "type": "table", + "html": "
Class MethodMil-aircraftcarlorikeettabby catholsteinLab-retrievertree froghorsefishing vesselfire truck
—0.3430.430.4430.2710.3290.38570.40.0290.4290.629
HFSS0.5140.0290.5860.8860.4570.2710.8710.2430.5430.071
[39]000.21430.12860.04290.02860.05710.12860.21430
", + "image_path": "323a4ff4252b1196b230f316c661e07df565a1c5d8a3c634e1843e3d191a2dd9.jpg" + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "table_body" + } + ], + "index": 10 + }, + { + "bbox": [ + 313, + 255, + 555, + 434 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 255, + 555, + 434 + ], + "spans": [ + { + "bbox": [ + 313, + 255, + 555, + 434 + ], + "type": "text", + "content": "thus validate the impact of shortcuts searched by HFSS in generalization tests, by measuring performance results on IN-SCT, an OOD test set designed by [39]. We report the results in Tab. 3. The strong shortcuts for classes airliner, frog and container ship searched by HFSS are present in the OOD set, contributing to close-to or above average TPR of classes Military aircraft (Mil-aircraft), tree frog, and fishing vessel in the IN-SCT dataset. The shortcuts searched by [39] for class Siamese cat result in a much lower TPR for class tabby cat in IN-SCT, though shortcuts for this class still exist in the OOD data (as observed from the results of HFSS). This indicates that the method proposed in [39] has limitations in estimating the impact of shortcuts on OOD performance, and further demonstrates the effectiveness of HFSS in finding shortcuts." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 313, + 452, + 391, + 464 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 452, + 391, + 464 + ], + "spans": [ + { + "bbox": [ + 313, + 452, + 391, + 464 + ], + "type": "text", + "content": "5. Conclusions" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 313, + 474, + 555, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 474, + 555, + 713 + ], + "spans": [ + { + "bbox": [ + 313, + 474, + 555, + 713 + ], + "type": "text", + "content": "We proposed the first method for analysis of frequency shortcuts, HFSS, that enables the inspection of shortcut learning in large-scale models and datasets with thousands of classes. HFSS is more time-efficient and effective in finding frequency shortcuts compared to existing approaches. We investigate frequency shortcut learning in models trained on ImageNet and relate their results to robustness and generalization performance under different OOD conditions. The impact of frequency shortcuts on model generalization depends on the specific OOD scenarios. Existing models yield good performance on generalization benchmarks when texture information, corresponding to most frequency shortcuts, is mostly preserved in OOD data. Instead, frequency shortcuts impair the generalizability of models to images with rendition changes. This highlights the limitation of current OOD performance evaluation benchmarks, which need to explicitly take into account the impact that frequency shortcuts have on model performance. HFSS provides a tool to bridge this gap and extends the rigor of model generalization evaluation." + } + ] + } + ], + "index": 13 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "text", + "content": "25205" + } + ] + } + ], + "index": 14 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 7 + }, + { + "para_blocks": [ + { + "bbox": [ + 56, + 72, + 147, + 85 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 72, + 147, + 85 + ], + "spans": [ + { + "bbox": [ + 56, + 72, + 147, + 85 + ], + "type": "text", + "content": "Acknowledgment" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 55, + 91, + 296, + 140 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 91, + 296, + 140 + ], + "spans": [ + { + "bbox": [ + 55, + 91, + 296, + 140 + ], + "type": "text", + "content": "Shunxin Wang and Nicola Strisciuglio are partly supported by the ERJU project. Europe's Rail Joint Undertaking is a European partnership on rail research and innovation established under the Horizon Europe program (2020-2027)." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 56, + 148, + 115, + 162 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 148, + 115, + 162 + ], + "spans": [ + { + "bbox": [ + 56, + 148, + 115, + 162 + ], + "type": "text", + "content": "References" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 57, + 169, + 296, + 712 + ], + "type": "list", + "angle": 0, + "index": 15, + "blocks": [ + { + "bbox": [ + 61, + 169, + 296, + 223 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 61, + 169, + 296, + 223 + ], + "spans": [ + { + "bbox": [ + 61, + 169, + 296, + 223 + ], + "type": "text", + "content": "[1] Hyojin Bahng, Sanghyuk Chun, Sangdoo Yun, Jaegul Choo, and Seong Joon Oh. Learning de-biased representations with biased representations. In Proceedings of the 37th International Conference on Machine Learning, pages 528-539. PMLR, 2020. 2" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 61, + 224, + 296, + 280 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 61, + 224, + 296, + 280 + ], + "spans": [ + { + "bbox": [ + 61, + 224, + 296, + 280 + ], + "type": "text", + "content": "[2] Christopher Boland, Sonia Dahdouh, Sotirios A. Tsaftaris, and Keith A Goatman. There are no shortcuts to anywhere worth going: Identifying shortcuts in deep learning models for medical image analysis. In Submitted to Medical Imaging with Deep Learning, 2024. under review. 2" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 61, + 281, + 296, + 313 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 61, + 281, + 296, + 313 + ], + "spans": [ + { + "bbox": [ + 61, + 281, + 296, + 313 + ], + "type": "text", + "content": "[3] Yuan Cao, Zhiying Fang, Yue Wu, Ding-Xuan Zhou, and Quanquan Gu. Towards understanding the spectral bias of deep learning, 2020. 2" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 61, + 314, + 296, + 378 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 61, + 314, + 296, + 378 + ], + "spans": [ + { + "bbox": [ + 61, + 314, + 296, + 378 + ], + "type": "text", + "content": "[4] Francesco Croce, Maksym Andriushchenko, Vikash Sehwag, Edoardo Debenedetti, Nicolas Flammarion, Mung Chiang, Prateek Mittal, and Matthias Hein. Robustbench: a standardized adversarial robustness benchmark. In Thirty-fifth Conference on Neural Information Processing Systems Datasets and Benchmarks Track (Round 2), 2021. 2" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 62, + 380, + 296, + 413 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 380, + 296, + 413 + ], + "spans": [ + { + "bbox": [ + 62, + 380, + 296, + 413 + ], + "type": "text", + "content": "[5] Nikolay Dagaev, Brett D. Roads, Xiaoliang Luo, Daniel N. Barry, Kaustubh R. Patil, and Bradley C. Love. A too-good-to-be-true prior to reduce shortcut reliance, 2021. 1, 2" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 62, + 414, + 296, + 457 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 414, + 296, + 457 + ], + "spans": [ + { + "bbox": [ + 62, + 414, + 296, + 457 + ], + "type": "text", + "content": "[6] Jia Deng, Wei Dong, Richard Socher, Li-Jia Li, Kai Li, and Li Fei-Fei. Imagenet: A large-scale hierarchical image database. In 2009 IEEE Conference on Computer Vision and Pattern Recognition, pages 248-255, 2009. 4" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 62, + 458, + 296, + 502 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 458, + 296, + 502 + ], + "spans": [ + { + "bbox": [ + 62, + 458, + 296, + 502 + ], + "type": "text", + "content": "[7] Paul Gavrikov and Janis Keuper. Can biases in imagenet models explain generalization? In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 22184-22194, 2024. 1, 2" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 62, + 502, + 296, + 567 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 502, + 296, + 567 + ], + "spans": [ + { + "bbox": [ + 62, + 502, + 296, + 567 + ], + "type": "text", + "content": "[8] Robert Geirhos, Patricia Rubisch, Claudio Michaelis, Matthias Bethge, Felix A. Wichmann, and Wieland Brendel. Imagenet-trained CNNs are biased towards texture; increasing shape bias improves accuracy and robustness. In International Conference on Learning Representations, 2019. 1, 2, 7" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 62, + 568, + 296, + 612 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 568, + 296, + 612 + ], + "spans": [ + { + "bbox": [ + 62, + 568, + 296, + 612 + ], + "type": "text", + "content": "[9] Robert Geirhos, Jörn-Henrik Jacobsen, Claudio Michaelis, Richard Zemel, Wieland Brendel, Matthias Bethge, and Felix A. Wichmann. Shortcut learning in deep neural networks. Nature Machine Intelligence, 2(11):665-673, 2020. 1, 2" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 57, + 613, + 294, + 635 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 613, + 294, + 635 + ], + "spans": [ + { + "bbox": [ + 57, + 613, + 294, + 635 + ], + "type": "text", + "content": "[10] Ian J. Goodfellow, Jonathon Shlens, and Christian Szegedy. Explaining and harnessing adversarial examples, 2015. 4, 5" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 57, + 636, + 294, + 658 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 636, + 294, + 658 + ], + "spans": [ + { + "bbox": [ + 57, + 636, + 294, + 658 + ], + "type": "text", + "content": "[11] Jindong Gu and Daniela Oelke. Understanding bias in machine learning, 2019. 2" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 57, + 658, + 296, + 712 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 658, + 296, + 712 + ], + "spans": [ + { + "bbox": [ + 57, + 658, + 296, + 712 + ], + "type": "text", + "content": "[12] Jun Guo, Wei Bao, Jiakai Wang, Yuqing Ma, Xinghai Gao, Gang Xiao, Aishan Liu, Jian Dong, Xianglong Liu, and Wenjun Wu. A comprehensive evaluation framework for deep model robustness. Pattern Recognition, 137:109308, 2023. 2" + } + ] + } + ], + "index": 14 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 315, + 73, + 553, + 712 + ], + "type": "list", + "angle": 0, + "index": 31, + "blocks": [ + { + "bbox": [ + 316, + 73, + 553, + 116 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 73, + 553, + 116 + ], + "spans": [ + { + "bbox": [ + 316, + 73, + 553, + 116 + ], + "type": "text", + "content": "[13] Dan Hendrycks and Thomas Dietterich. Benchmarking neural network robustness to common corruptions and perturbations. In International Conference on Learning Representations, 2019. 2, 4, 5" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 317, + 118, + 553, + 194 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 118, + 553, + 194 + ], + "spans": [ + { + "bbox": [ + 317, + 118, + 553, + 194 + ], + "type": "text", + "content": "[14] Dan Hendrycks, Steven Basart, Norman Mu, Saurav Kadavath, Frank Wang, Evan Dorundo, Rahul Desai, Tyler Zhu, Samyak Parajuli, Mike Guo, Dawn Song, Jacob Steinhardt, and Justin Gilmer. The many faces of robustness: A critical analysis of out-of-distribution generalization. In Proceedings of the IEEE/CVF International Conference on Computer Vision (ICCV), pages 8340-8349, 2021. 2, 4, 5" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 316, + 195, + 553, + 239 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 195, + 553, + 239 + ], + "spans": [ + { + "bbox": [ + 316, + 195, + 553, + 239 + ], + "type": "text", + "content": "[15] Hanxun Huang, Xingjun Ma, Sarah Monazam Erfani, James Bailey, and Yisen Wang. Unlearnable examples: Making personal data unexploitable. In International Conference on Learning Representations, 2021. 5" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 316, + 239, + 553, + 282 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 239, + 553, + 282 + ], + "spans": [ + { + "bbox": [ + 316, + 239, + 553, + 282 + ], + "type": "text", + "content": "[16] Zhi-Qin John Xu, Yaoyu Zhang, Tao Luo, Yanyang Xiao, and Zheng Ma. Frequency principle: Fourier analysis sheds light on deep neural networks. Communications in Computational Physics, 28(5):1746-1767, 2020. 2" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 316, + 283, + 553, + 327 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 283, + 553, + 327 + ], + "spans": [ + { + "bbox": [ + 316, + 283, + 553, + 327 + ], + "type": "text", + "content": "[17] Oğuzhan Fatih Kar, Teresa Yeo, Andrei Atanov, and Amir Zamir. 3d common corruptions and data augmentation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 18963-18974, 2022. 2" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 316, + 327, + 553, + 349 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 327, + 553, + 349 + ], + "spans": [ + { + "bbox": [ + 316, + 327, + 553, + 349 + ], + "type": "text", + "content": "[18] A. Krizhevsky. Learning multiple layers of features from tiny images, 1970. 5" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 316, + 350, + 553, + 381 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 350, + 553, + 381 + ], + "spans": [ + { + "bbox": [ + 316, + 350, + 553, + 381 + ], + "type": "text", + "content": "[19] S. Lapuschkin, S. Wäldchen, A. Binder, et al. Unmasking Clever Hans predictors and assessing what machines really learn. Nat Commun 10, 1096, 2019. 1, 2" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 316, + 383, + 553, + 436 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 383, + 553, + 436 + ], + "spans": [ + { + "bbox": [ + 316, + 383, + 553, + 436 + ], + "type": "text", + "content": "[20] Mathias Lecuyer, Vaggelis Atlidakis, Roxana Geambasu, Daniel Hsu, and Suman Jana. Certified robustness to adversarial examples with differential privacy. In 2019 IEEE Symposium on Security and Privacy (SP), pages 656-672, 2019. 2" + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 315, + 437, + 553, + 459 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 315, + 437, + 553, + 459 + ], + "spans": [ + { + "bbox": [ + 315, + 437, + 553, + 459 + ], + "type": "text", + "content": "[21] Sungbin Lim, Ildoo Kim, Taesup Kim, Chiheon Kim, and Sungwoong Kim. Fast autoaugment, 2019. 3" + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 316, + 460, + 553, + 502 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 460, + 553, + 502 + ], + "spans": [ + { + "bbox": [ + 316, + 460, + 553, + 502 + ], + "type": "text", + "content": "[22] Ze Liu, Yutong Lin, Yue Cao, Han Hu, Yixuan Wei, Zheng Zhang, Stephen Lin, and Baining Guo. Swin transformer: Hierarchical vision transformer using shifted windows, 2021. 3" + } + ] + } + ], + "index": 25 + }, + { + "bbox": [ + 316, + 504, + 553, + 557 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 504, + 553, + 557 + ], + "spans": [ + { + "bbox": [ + 316, + 504, + 553, + 557 + ], + "type": "text", + "content": "[23] Emanuele Marconato, Stefano Teso, Antonio Vergari, and Andrea Passerini. Not all neuro-symbolic concepts are created equal: Analysis and mitigation of reasoning shortcuts. In Thirty-seventh Conference on Neural Information Processing Systems, 2023. 2" + } + ] + } + ], + "index": 26 + }, + { + "bbox": [ + 316, + 559, + 553, + 592 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 559, + 553, + 592 + ], + "spans": [ + { + "bbox": [ + 316, + 559, + 553, + 592 + ], + "type": "text", + "content": "[24] Matthias Minderer, Olivier Bachem, Neil Houlsby, and Michael Tschannen. Automatic shortcut removal for self-supervised representation learning, 2020. 2" + } + ] + } + ], + "index": 27 + }, + { + "bbox": [ + 317, + 593, + 553, + 646 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 593, + 553, + 646 + ], + "spans": [ + { + "bbox": [ + 317, + 593, + 553, + 646 + ], + "type": "text", + "content": "[25] Eric Mintun, Alexander Kirillov, and Saining Xie. On interaction between augmentations and corruptions in natural corruption robustness. In Advances in Neural Information Processing Systems, pages 3571-3583. Curran Associates, Inc., 2021. 2" + } + ] + } + ], + "index": 28 + }, + { + "bbox": [ + 316, + 647, + 553, + 669 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 647, + 553, + 669 + ], + "spans": [ + { + "bbox": [ + 316, + 647, + 553, + 669 + ], + "type": "text", + "content": "[26] Samuel G. Müller and Frank Hutter. Trivialaugment: Tuning-free yet state-of-the-art data augmentation, 2021. 3" + } + ] + } + ], + "index": 29 + }, + { + "bbox": [ + 316, + 670, + 553, + 712 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 670, + 553, + 712 + ], + "spans": [ + { + "bbox": [ + 316, + 670, + 553, + 712 + ], + "type": "text", + "content": "[27] Meike Nauta, Ricky Walsh, Adam Dubowski, and Christin Seifert. Uncovering and correcting shortcut learning in machine learning models for skin cancer diagnosis. Diagnostics, 12(1), 2022. 1, 2, 3" + } + ] + } + ], + "index": 30 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 318, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 318, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 318, + 757 + ], + "type": "text", + "content": "25206" + } + ] + } + ], + "index": 32 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 8 + }, + { + "para_blocks": [ + { + "bbox": [ + 56, + 72, + 294, + 712 + ], + "type": "list", + "angle": 0, + "index": 12, + "blocks": [ + { + "bbox": [ + 56, + 72, + 294, + 126 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 72, + 294, + 126 + ], + "spans": [ + { + "bbox": [ + 56, + 72, + 294, + 126 + ], + "type": "text", + "content": "[28] Hongjing Niu, Hanting Li, Feng Zhao, and Bin Li. Roadblocks for temporarily disabling shortcuts and learning new knowledge. In Advances in Neural Information Processing Systems, pages 29064-29075. Curran Associates, Inc., 2022. 2" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 56, + 129, + 294, + 174 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 129, + 294, + 174 + ], + "spans": [ + { + "bbox": [ + 56, + 129, + 294, + 174 + ], + "type": "text", + "content": "[29] Ian Osband and Benjamin Van Roy. Why is posterior sampling better than optimism for reinforcement learning? In Proceedings of the 34th International Conference on Machine Learning, pages 2701-2710. PMLR, 2017. 3" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 56, + 175, + 294, + 228 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 175, + 294, + 228 + ], + "spans": [ + { + "bbox": [ + 56, + 175, + 294, + 228 + ], + "type": "text", + "content": "[30] Mohammad Pezeshki, Sekou-Oumar Kaba, Yoshua Bengio, Aaron Courville, Doina Precup, and Guillaume Lajoie. Gradient starvation: A learning proclivity in neural networks. In Advances in Neural Information Processing Systems, 2021. 2" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 56, + 230, + 294, + 283 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 230, + 294, + 283 + ], + "spans": [ + { + "bbox": [ + 56, + 230, + 294, + 283 + ], + "type": "text", + "content": "[31] Benjamin Recht, Rebecca Roelofs, Ludwig Schmidt, and Vaishaal Shankar. Do ImageNet classifiers generalize to ImageNet? In Proceedings of the 36th International Conference on Machine Learning, pages 5389-5400. PMLR, 2019. 2, 4, 5" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 56, + 287, + 294, + 319 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 287, + 294, + 319 + ], + "spans": [ + { + "bbox": [ + 56, + 287, + 294, + 319 + ], + "type": "text", + "content": "[32] Joshua Robinson, Li Sun, Ke Yu, Kayhan Batmanghelich, Stefanie Jegelka, and Suvrit Sra. Can contrastive learning avoid shortcut solutions?, 2021. 2" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 56, + 321, + 294, + 386 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 321, + 294, + 386 + ], + "spans": [ + { + "bbox": [ + 56, + 321, + 294, + 386 + ], + "type": "text", + "content": "[33] Ramprasaath R. Selvaraju, Michael Cogswell, Abhishek Das, Ramakrishna Vedantam, Devi Parikh, and Dhruv Batra. Grad-cam: Visual explanations from deep networks via gradient-based localization. In 2017 IEEE International Conference on Computer Vision (ICCV), pages 618-626, 2017. 2" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 56, + 388, + 294, + 442 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 388, + 294, + 442 + ], + "spans": [ + { + "bbox": [ + 56, + 388, + 294, + 442 + ], + "type": "text", + "content": "[34] Harshay Shah, Kaustav Tamuly, Aditi Raghunathan, Prateek Jain, and Praneeth Netrapalli. The pitfalls of simplicity bias in neural networks. In Advances in Neural Information Processing Systems, pages 9573-9585. Curran Associates, Inc., 2020. 1, 2" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 56, + 445, + 294, + 499 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 445, + 294, + 499 + ], + "spans": [ + { + "bbox": [ + 56, + 445, + 294, + 499 + ], + "type": "text", + "content": "[35] Damien Teney, Armand Mihai Nicoliciou, Valentin Hartmann, and Ehsan Abbasnejad. Neural redshift: Random networks are not random functions. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 4786-4796, 2024. 2" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 56, + 501, + 294, + 544 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 501, + 294, + 544 + ], + "spans": [ + { + "bbox": [ + 56, + 501, + 294, + 544 + ], + "type": "text", + "content": "[36] Haohan Wang, Songwei Ge, Zachary Lipton, and Eric P Xing. Learning robust global representations by penalizing local predictive power. In Advances in Neural Information Processing Systems, pages 10506-10518, 2019. 2, 4, 5" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 56, + 546, + 294, + 590 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 546, + 294, + 590 + ], + "spans": [ + { + "bbox": [ + 56, + 546, + 294, + 590 + ], + "type": "text", + "content": "[37] Shunxin Wang, Raymond Veldhuis, Christoph Brune, and Nicola Strisciuglio. Frequency shortcut learning in neural networks. In NeurIPS 2022 Workshop on Distribution Shifts: Connecting Methods and Applications, 2022. 1, 2" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 56, + 592, + 294, + 646 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 592, + 294, + 646 + ], + "spans": [ + { + "bbox": [ + 56, + 592, + 294, + 646 + ], + "type": "text", + "content": "[38] Shunxin Wang, Christoph Brune, Raymond Veldhuis, and Nicola Strisciuglio. Dfm-x: Augmentation by leveraging prior knowledge of shortcut learning. In Proceedings of the IEEE/CVF International Conference on Computer Vision (ICCV) Workshops, pages 129-138, 2023. 2, 3" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 56, + 647, + 294, + 712 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 647, + 294, + 712 + ], + "spans": [ + { + "bbox": [ + 56, + 647, + 294, + 712 + ], + "type": "text", + "content": "[39] Shunxin Wang, Raymond Veldhuis, Christoph Brune, and Nicola Strisciuglio. What do neural networks learn in image classification? a frequency shortcut perspective. In Proceedings of the IEEE/CVF International Conference on Computer Vision (ICCV), pages 1433-1442, 2023. 1, 2, 3, 4, 5, 8" + } + ] + } + ], + "index": 11 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 315, + 72, + 553, + 218 + ], + "type": "list", + "angle": 0, + "index": 17, + "blocks": [ + { + "bbox": [ + 315, + 72, + 553, + 106 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 315, + 72, + 553, + 106 + ], + "spans": [ + { + "bbox": [ + 315, + 72, + 553, + 106 + ], + "type": "text", + "content": "[40] Shunxin Wang, Raymond Veldhuis, Christoph Brune, and Nicola Strisciuglio. A survey on the robustness of computer vision models against common corruptions, 2023. 2" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 316, + 107, + 553, + 129 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 107, + 553, + 129 + ], + "spans": [ + { + "bbox": [ + 316, + 107, + 553, + 129 + ], + "type": "text", + "content": "[41] Zhi-Qin John Xu and Hanxu Zhou. Deep frequency principle towards understanding why deeper learning is faster, 2020. 2" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 316, + 130, + 553, + 173 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 130, + 553, + 173 + ], + "spans": [ + { + "bbox": [ + 316, + 130, + 553, + 173 + ], + "type": "text", + "content": "[42] Jiayun Zheng and Maggie Makar. Causally motivated multishutdown identification and removal. In Advances in Neural Information Processing Systems, pages 12800-12812. Curran Associates, Inc., 2022. 2" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 316, + 175, + 553, + 218 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 175, + 553, + 218 + ], + "spans": [ + { + "bbox": [ + 316, + 175, + 553, + 218 + ], + "type": "text", + "content": "[43] Xin Zou and Weiwei Liu. On the adversarial robustness of out-of-distribution generalization models. In Advances in Neural Information Processing Systems, pages 68908-68938. Curran Associates, Inc., 2023. 2" + } + ] + } + ], + "index": 16 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 749, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 749, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 749, + 317, + 757 + ], + "type": "text", + "content": "25207" + } + ] + } + ], + "index": 18 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 9 + } + ], + "_backend": "vlm", + "_version_name": "2.6.4" +} \ No newline at end of file diff --git a/2025/Do Visual Imaginations Improve Vision-and-Language Navigation Agents_/6684e0af-b484-43ef-8443-50d4bfdefdd1_content_list.json b/2025/Do Visual Imaginations Improve Vision-and-Language Navigation Agents_/6684e0af-b484-43ef-8443-50d4bfdefdd1_content_list.json new file mode 100644 index 0000000000000000000000000000000000000000..a16edf097c64c7e23d1c761162801873073a2639 --- /dev/null +++ b/2025/Do Visual Imaginations Improve Vision-and-Language Navigation Agents_/6684e0af-b484-43ef-8443-50d4bfdefdd1_content_list.json @@ -0,0 +1,1414 @@ +[ + { + "type": "text", + "text": "Do Visual Imaginations Improve Vision-and-Language Navigation Agents?", + "text_level": 1, + "bbox": [ + 120, + 130, + 875, + 152 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Akhil Perincherry1, Jacob Krantz, Stefan Lee1 \n1Oregon State University", + "bbox": [ + 315, + 180, + 679, + 217 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "perincha@oregonstate.edu, jkrantz989@gmail.com, leestef@oregonstate.edu", + "bbox": [ + 179, + 219, + 810, + 233 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Abstract", + "text_level": 1, + "bbox": [ + 246, + 268, + 326, + 284 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Vision-and-Language Navigation (VLN) agents are tasked with navigating an unseen environment using natural language instructions. In this work, we study if visual representations of sub-goals implied by the instructions can serve as navigational cues and lead to increased navigation performance. To synthesize these visual representations or \"imaginations\", we leverage a text-to-image diffusion model on landmark references contained in segmented instructions. These imaginations are provided to VLN agents as an added modality to act as landmark cues and an auxiliary loss is added to explicitly encourage relating these with their corresponding referring expressions. Our findings reveal an increase in success rate (SR) of $\\sim 1$ point and up to $\\sim 0.5$ points in success scaled by inverse path length (SPL) across agents. These results suggest that the proposed approach reinforces visual understanding compared to relying on language instructions alone. Code and data for our work can be found at https://www.akhilperincherry.com/VLN-Imagine-website/.", + "bbox": [ + 89, + 301, + 483, + 589 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "1. Introduction", + "text_level": 1, + "bbox": [ + 89, + 618, + 220, + 633 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "In this work, we examine whether providing visual imagery corresponding to described landmarks improves the performance of agents following natural language navigation instructions. Consider the natural language navigation instruction presented in Figure 1 which asks an agent to \"Go straight, take a left at the pool table to enter the kitchen. Walk to the bedroom and stop\". This instruction provides unconditional action directives like \"go straight\" but also frequently conditions the given directions on visual landmarks in the scene such as the pool table, kitchen, and bedroom. Standard approaches to vision-and-language navigation tasks rely on learned cross-modal alignment mechanisms to implicitly associate these noun phrases with their visual referents during navigation. However, text-to-image generation models have improved to the point that producing imagery matching the semantics of these visual references prior to navigation is plausible. We show three such", + "bbox": [ + 89, + 643, + 483, + 900 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Natural Language Navigation Instruction:", + "text_level": 1, + "bbox": [ + 526, + 301, + 754, + 314 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/705f98388715ea92f7dd13c69a293d935e23117c9638211362f8833f3ce60dcd.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 527, + 318, + 562, + 347 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Go straight, take a left at the pool table to enter the kitchen. Walk to the bedroom and stop.", + "bbox": [ + 573, + 319, + 893, + 347 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Visual Imaginations:", + "text_level": 1, + "bbox": [ + 527, + 357, + 638, + 369 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/7734d74b0f516f15705c4d935639aa995a10be9f1d968980a406f0abd2a3c2f0.jpg", + "image_caption": [ + "Figure 1. Illustration of visual imaginations. (Top) A natural language instruction specifying sub-goals pool table, kitchen, and bedroom. (Bottom) Visual imaginations of landmarks pool table, kitchen and bedroom referenced by the sub-goals in the instruction. In our work, we study if these visual imaginations generated using text-to-image models can improve performance in VLN." + ], + "image_footnote": [], + "bbox": [ + 522, + 369, + 563, + 455 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/924ec74d18c943cbc11f8635fd22e9f6fd6d8af80838749c8db460cc81cdb261.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 573, + 373, + 674, + 455 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/90b84f5ea45bd4c6ac8966bf11611da60bbbffea94897ac8d7641849870f928d.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 684, + 373, + 784, + 455 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/388633a477afaf69d30be05e9f4ef9df5c0b90d03ce290669461593b9c8146dd.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 795, + 373, + 900, + 455 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "generations for our running example in Figure 1. Drawing an analogy to the substantial work in cognitive science on the impact of mental imagery on task performance, we refer to these generated images as visual imaginations and study whether providing them in addition to corresponding language-based instructions can improve the performance of vision-and-language navigation agents.", + "bbox": [ + 511, + 568, + 906, + 672 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "But why might we believe visual imaginations could be beneficial? In short - text-to-image models have broad knowledge, and learning to perform semantic matching in the image domain may be an easier task than language grounding. Associating noun phrases and their visual referents is a sub-task in vision-and-language navigation (VLN) that most works address through extensive pretraining of model components on web-scale image caption datasets - e.g., by initializing image and language encoders with CLIP [34] models. While this provides a useful warm-start, it is unclear how well the remaining cross-modal components that drive navigation retain these capabilities after downstream VLN training. Downstream datasets may cover only a small portion of relevant objects and visual grounding remains a weakness in many existing VLN models [43, 47].", + "bbox": [ + 511, + 674, + 908, + 901 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "CVF", + "bbox": [ + 106, + 2, + 181, + 42 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore.", + "bbox": [ + 236, + 0, + 810, + 46 + ], + "page_idx": 0 + }, + { + "type": "page_number", + "text": "3846", + "bbox": [ + 482, + 944, + 514, + 955 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Directly synthesizing visual imaginations using off-the-shelf text-to-image models offers alternative training and inference mechanisms. Synthesized images have known correspondences to noun phrases in instructions that can be leveraged to reinforce visual grounding during downstream task training through auxiliary losses. Further, matching a visual imagination with environment observations reduces to an image-to-image matching task – offering the potential to use visual imaginations as a convenient pivot between language and vision. This may prove especially effective for out-of-distribution landmarks that have limited or no support during downstream task training. For example, novel instructions referencing “butterfly sculptures” or “Pulp Fiction posters” are unlikely to be well-grounded in the training set but produce semantically relevant visual imaginations that could then be matched to observations.", + "bbox": [ + 89, + 90, + 480, + 332 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Though we do not imply any direct connections, our work is also inspired by cognitive science studies which indicate mental imagery can serve as valuable cues in addition to linguistic stimuli. Several works supporting dualcoding theory [29-31] find that concrete imagery tends to be more effective for certain types of concept learning compared to linguistic stimuli alone. Further, findings from human and animal studies [23, 36, 38, 39] suggest that imagining snapshots of the environment or potential trajectories before navigating can improve decision making.", + "bbox": [ + 89, + 334, + 480, + 486 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "To investigate our key question, we augment agents with visual imaginations for the popular Room-2-Room (R2R) [4] and REVERIE [33] tasks. We propose a VLN agent agnostic procedure to incorporate visual imaginations as additional inputs and introduce a text-imagination alignment loss to explicitly encourage visual grounding during training. We generate visual imaginations for key noun phrases in instructions that refer to visual landmarks using an off-the-shelf text-to-image generation model [32].", + "bbox": [ + 89, + 488, + 480, + 625 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Applying our technique to two existing models, HAMT [8] and DUET [9], we find modest improvements - R2R validation unseen success increases by roughly 1.0 and 0.6 points respectively. This translates to an improvement of 2 points in success on the test set for our modified DUET model. On REVERIE validation unseen, we improve DUET performance by 1.3 points success and 0.82 points grounding success. Our results suggest that including explicit visual depictions of referred objects can improve performance on vision-and-language navigation tasks.", + "bbox": [ + 89, + 627, + 480, + 777 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Contributions. We summarize our contributions as:", + "text_level": 1, + "bbox": [ + 89, + 782, + 437, + 796 + ], + "page_idx": 1 + }, + { + "type": "list", + "sub_type": "text", + "list_items": [ + "- We develop a pipeline for generating visual imaginations from navigation instructions and synthesize the R2R-Imagine dataset to enable studying the impact of text-to-image models on VLN agents.", + "- We propose an agent-agnostic method to incorporate visual imaginations into existing VLN agents and show" + ], + "bbox": [ + 94, + 805, + 482, + 901 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "improved performance for HAMT [8] and DUET [9] across R2R and REVERIE.", + "bbox": [ + 532, + 90, + 903, + 119 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "- We provide ablations of modeling decisions to characterize the design space for adding visual imaginations.", + "bbox": [ + 517, + 126, + 903, + 156 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "2. Related Work", + "text_level": 1, + "bbox": [ + 513, + 175, + 653, + 191 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Vision-and-Language Navigation (VLN). Navigating agents that can operate in novel environments using freeform natural language lends itself as personal assistants, field-robots and search-and-rescue agents. Development of VLN agents has been facilitated by photorealistic simulators such as Matterport3D [7] and paired language-action datasets such as Room-2-Room (R2R) [4] and REVERIE [33]. R2R provides fine-grained instructions conveying low-level directives such as \"Go left, then right at the hall\". REVERIE specifies coarse-grained instructions conveying broad high-level goals such as \"Adjust the picture by the lamp in the hall\". Progress in VLN has broadly ranged from architectural improvements and data scaling techniques. Architectural improvements include evolution from recurrent systems [4] to transformer architectures [8, 9, 17]. HAMT [8] uses a hierarchical transformer architecture to encode spatial and temporal information. DUET [9] maintains a topological map and performs coarse and fine scale encoding over it. Data scaling techniques include augmenting environments [14, 40] using an action-to-text speaker model [12] to generate synthetic instructions, and scaling environments [19, 41] with additional scenes such as HM3D [35] and Gibson [42] in Habitat [37]. Recently, zero-shot or learned integrations of pretrained large language models for VLN [45, 46] is showing promising performance. However, running large models in-the-loop incurs high computational costs.", + "bbox": [ + 511, + 202, + 903, + 609 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Closely related to our work, ADAPT [27] augments VLN agents with visual observations related to an instruction. Language-observation pairs are stored from the training set and multiple pairs of matching object/location nouns from instructions are retrieved at test time using CLIP [34] scores. Fundamentally, the ADAPT image base is limited to training environments whereas our imaginations cover an open distribution. Therefore, our method can be more representative of the instruction and visualize novel objects/locations. For instance, a sub-instruction of \"walk into the kitchen with blue walls\" would retrieve images of kitchens from the training environment using ADAPT, but our imaginations would convey a kitchen with blue walls.", + "bbox": [ + 511, + 611, + 903, + 808 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Another adjacent work, LAD [26] leverages layout dreamer and goal dreamer modules optimized around a topological graph to navigate via coarse-grained instructions. The capability of our method contrasts with LAD inthree primary ways: 1) method generality; Our method is performant across models and datasets, whereas LAD is a", + "bbox": [ + 511, + 810, + 903, + 900 + ], + "page_idx": 1 + }, + { + "type": "page_number", + "text": "3847", + "bbox": [ + 482, + 944, + 514, + 955 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "full architecture constructed around a particular topological graph in coarse-grained settings, 2) sequential imagination: we imagine a sequence of landmarks along a path to inform sequential decision-making, whereas LAD imagines just the goal, and 3) we perform early fusion of imaginations and language whereas LAD employs late fusion.", + "bbox": [ + 88, + 90, + 480, + 181 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Image generation in VLN. Several techniques have proposed generative models to predict future visual observations or entire environments in VLN [21, 24, 25]. Pathdreamer [21] developed a GAN-based model to predict panoramic observations after a navigation action in VLN, conditioned on prior visual observations. Agents utilizing these predicted observations in a fixed-horizon search with a learned trajectory-instruction alignment heuristic achieve non-trivial success compared to simple baselines, but fall short of planning on actual observations and standard VLN models. Unlike our approach, Pathdreamer does not condition on instruction text or explore training agents on generated imagery. Similarly, PanoGen [24] uses a diffusion model to also generate panorama observations, but conditions on image captions while doing so as a form of visual dataset augmentation – replacing Matterport3D observations with synthetic generations with matching semantics. In contrast to our approach, this data augmentation does not provide additional inputs to the VLN models, is not conditioned on instructions, and is not active at inference time.", + "bbox": [ + 91, + 185, + 486, + 487 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Related to our work is VLN-Sig [25], which adapts VLN agents to predict quantized features of future visual observations given the instruction and observation history. This supports additional pretraining tasks and an auxiliary loss during downstream VLN finetuning for predicting next-step image semantics. Unlike our approach, VLN-Sig relies on the trained VLN agent to predict future image properties and focuses more on next-step prediction. In contrast, we leverage broad vision-and-language knowledge from text-to-image models and focus generating instruction-referred visual landmarks to improve agent performance.", + "bbox": [ + 89, + 489, + 482, + 655 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Image generation in robot control. Taking a wider view, several methods for controlling embodied agents have leveraged text-conditioned generative techniques to provide goal specifications in the form of images or videos. For tabletop object rearrangement tasks, many works use text-to-image models to synthesize goals or sub-goals in response to natural language commands, see [6, 13, 20] for a nonexhaustive sample. Others produce denser conditioning, leveraging text-to-video models to generate entire potential robot trajectories [5, 11]. While similar in motivation to our approach, these techniques focus on tabletop manipulation settings which introduce difficult physical challenges but significantly limit the problem to constrained spaces that remain largely observed throughout operation. As such, sub-goals are often edits to existing observations rather than generations of yet-to-be-seen landmarks as in our approach.", + "bbox": [ + 89, + 659, + 482, + 901 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3. Methodology", + "text_level": 1, + "bbox": [ + 513, + 89, + 648, + 107 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Our proposed methodology consists of two components - (1) a visual imagination generation pipeline and (2) a model-agnostic approach to integrating visual imaginations into existing Vision-and-Language Navigation (VLN) agents. Before defining these, we start with establishing relevant notation for the VLN task.", + "bbox": [ + 511, + 114, + 903, + 205 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Problem definition. In the standard VLN task [4], an agent is prompted with a natural language navigation instruction and must make a sequence of actions to navigate along the described trajectory based on visual observations. We denote the $L$ -word instruction as $W = (w_{1}, w_{2}, \\dots, w_{L})$ and the visual observation at time step $t$ as a panoramic observation $O_{t}$ consisting of $K$ single view images $O_{t} = (I_{1}^{t}, \\dots, I_{K}^{t})$ . Following prior work, we consider $K = 36$ corresponding to images from 12 headings and 3 pitches. Amongst the single view images, a subset $\\mathcal{A}_{t} \\subset O_{t}$ correspond to navigable directions. Given the instruction $W$ and history of observations $O_{\\leq t}$ until time $t$ , the agent selects an action $a_{t} \\in \\mathcal{A}_{t} \\cup \\{\\text{stop}\\}$ and accordingly either moves or terminates the episode. Generally, the VLN agent is parameterized as a policy network $\\pi_{\\theta}(a_{t} \\mid W, O_{\\leq t})$ with parameters $\\theta$ learned via general pretraining and a combination of imitation learning and reinforcement learning [8, 9].", + "bbox": [ + 511, + 209, + 906, + 467 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3.1. Generating Visual Imaginations", + "text_level": 1, + "bbox": [ + 511, + 478, + 795, + 494 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "To generate visual imaginations of potential landmarks, we consider the structure of navigation instructions. Typically, these consist of a sequence of sub-instructions describing intermediate navigation steps, which may or may not refer to visual landmarks. For instance, a sub-instruction \"go past the couch\" implies a visual landmark while \"go straight then left\" does not. We use FG-R2R [15] to segment instructions, resulting in an average of 3.66 sub-instructions per instruction for the R2R[4] training set. For a given instruction $W$ decomposed into $m$ segments, we denote these as sub-instructions $S = (S_0, \\dots, S_m)$ .", + "bbox": [ + 511, + 501, + 906, + 667 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Sub-instruction filtering scheme. As sub-instructions may not always contain references to visual landmarks, we filter sub-instructions prior to image generation. First, we ignore any sub-instruction that lacks a noun phrase using Spacy [2] and then noun phrases are further filtered by a manual blacklist to avoid uninformative landmark generation. For example, we exclude detected noun phrases rooted on non-visual terms like counts ('one'), directions ('left'), or ambiguous object pronouns (\"it\") as these lack specificity to produce meaningful visual landmarks. After filtering, we denote the remaining sub-instructions $S' \\subset S$ .", + "bbox": [ + 511, + 671, + 908, + 837 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Visual imagination generation. The filtered subinstructions are passed through a text conditioned diffusion model as illustrated in Fig. 3. We use SDXL [32] as the diffusion model with positive and negative prompts cho", + "bbox": [ + 511, + 839, + 908, + 902 + ], + "page_idx": 2 + }, + { + "type": "page_number", + "text": "3848", + "bbox": [ + 482, + 944, + 514, + 955 + ], + "page_idx": 2 + }, + { + "type": "image", + "img_path": "images/4889dd68ffad037d0fd468b4bf220b92df48e6af134f01e0b8a25d397ab8370b.jpg", + "image_caption": [ + "Figure 2. An overview of our approach. (Left) Imaginations generated using valid sub-instructions from an instruction as determined by our filtering scheme are first passed to a pre-trained ViT to obtain feature vectors. A type embedding $t_{Im}$ for imagination modality is then added to the features which are encoded using a 3 layer MLP to obtain imagination embeddings $h_i$ . (Right) To integrate imagination modality to a VLN agent, the imagination embeddings $h_i$ are concatenated with instruction embeddings $t_i$ that are encoded using a text encoder $f_T(W)$ . The concatenated imagination-text embeddings are passed to the VLN agent's cross-modal encoder $f_X$ along with visual embeddings to predict a distribution over the agent's action space." + ], + "image_footnote": [], + "bbox": [ + 91, + 90, + 906, + 200 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Go down the stairs and then turn right. Go stand behind the easel by a unicycle.", + "bbox": [ + 96, + 306, + 478, + 320 + ], + "page_idx": 3 + }, + { + "type": "image", + "img_path": "images/0582747fc800f4c828ddbcb37d0f10ef039bda15bd5c818bfe8d3d2ddf519df1.jpg", + "image_caption": [ + "Figure 3. Example instruction segmentation, filtering, and image generation. Instructions are segmented to sub-instructions leveraging FG-R2R [15] and then filtered to remove phrases referring to uninformative nouns (e.g., \"right\"). We produce visual imaginations using SDXL [32] for remaining sub-instructions." + ], + "image_footnote": [], + "bbox": [ + 96, + 324, + 478, + 428 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "sen to steer generation towards indoor real-estate environments akin to typical environments in Matterport3D [7]. For example, providing \"indoor\" and \"real estate\" as positives prompts and \"humans\" and \"collage\" as negatives. See the supplementary for a full list. For a given instruction, we denote the generated images for valid sub-instructions as $\\mathcal{Z} = \\{Z_i \\mid S_i \\in S'\\}$ . We colloquially refer to these generated images as imaginations and generate them for the entire R2R dataset. The resulting R2R-Imagine dataset consists of over 41k synthesized imaginations. This amounts to an average of 2.96 imaginations per instruction.", + "bbox": [ + 89, + 536, + 483, + 703 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "3.2. Integration with Existing Models", + "text_level": 1, + "bbox": [ + 89, + 712, + 382, + 728 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "We consider augmentation with visual imagination to be a model agnostic approach and integrate it in two prior models - HAMT [8] and DUET [9]. Our integration approach consists of an imagination encoder, an alignment-promoting auxiliary loss, and a finetuning regimen (Fig. 2). Without loss of generality, we consider a decomposition of any VLN agent into an instruction encoder $f_{T}(W)$ , an observation/history encoder $f_{O}(O_{\\leq t})$ , and a cross-modal policy network $f_{X}(f_{T}(W), f_{O}(O_{\\leq t}))$ that combines the encoder outputs to predict a distribution over actions. Note that these may be fairly complex mechanisms and include explicit his", + "bbox": [ + 89, + 734, + 483, + 902 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "tory encoding modules separate from encoding the current observation; however, this abstraction is sufficient here.", + "bbox": [ + 511, + 306, + 906, + 338 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Imagination encoder and integration. Given an instruction $W$ with corresponding imaginations $\\mathcal{Z}$ , we independently encode each imagination $Z_{i}$ using a pretrained vision transformer (ViT) [10] to produce a $d$ -dim embedding vector. These are further transformed by a three-layer MLP such that the $i$ th imagination embedding can be written as", + "bbox": [ + 511, + 340, + 906, + 431 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\nh _ {i} = \\operatorname {M L P} \\left(\\operatorname {V i T} \\left(Z _ {i}\\right) + t _ {I m}\\right) \\tag {1}\n$$\n", + "text_format": "latex", + "bbox": [ + 609, + 446, + 906, + 465 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "where $t_{Im}$ is the type embedding for imagination modality. As we consider visual imaginations to be proxies for sub-instructions, the full set of imagination embeddings $\\mathcal{H}$ are concatenated to the text encodings such that our general VLN agent structure now computes $f_{X}([f_{T}(W),\\mathcal{H}],f_{O}(O_{\\leq t}))$ where $[\\cdot ,\\cdot ]$ denotes a concatenation. Taking the DUET model as an example, the imagination embeddings are concatenated to the instruction encoding prior to being passed to the coarse-scale and fine-scale cross-modal encoders.", + "bbox": [ + 511, + 472, + 905, + 622 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Auxiliary alignment loss. We add an auxiliary loss to align the representations of imaginations and their corresponding sub-instruction's noun phrases. For a given sub-instruction $S_{i}$ , we denote the mean embedding vector for noun phrase tokens in $S_{i}$ from the text encoder $f_{T}$ as $\\bar{S}_{i}$ . To align representations, we compute a cosine similarity loss between the corresponding imagination $Z_{i}$ encoded as $h_{i}$ and $\\bar{S}_{i}$ . This is averaged over all $N_{Im}$ imagination-subinstruction pairs $(Z_{i}, S_{i})$ in a batch,", + "bbox": [ + 511, + 625, + 906, + 763 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {L} _ {c o s} = \\frac {1}{N _ {I m}} \\sum_ {\\left(Z _ {i}, S _ {i}\\right) \\in B} \\left(1 - \\frac {h _ {i} \\cdot \\bar {S} _ {i}}{\\| h _ {i} \\| \\| \\bar {S} _ {i} \\|}\\right). \\tag {2}\n$$\n", + "text_format": "latex", + "bbox": [ + 565, + 773, + 906, + 815 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "During finetuning, we combine this auxiliary loss $\\mathcal{L}_{\\mathrm{cos}}$ with the losses from the base agent with a scaling factor $\\lambda$ such that the overall loss is $\\mathcal{L}_{\\mathrm{base}} + \\lambda \\mathcal{L}_{\\mathrm{cos}}$ .", + "bbox": [ + 511, + 821, + 905, + 868 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Finetuning regimen. We start with fully-trained checkpoints of prior models as our initial weights. Depending", + "bbox": [ + 511, + 869, + 906, + 902 + ], + "page_idx": 3 + }, + { + "type": "page_number", + "text": "3849", + "bbox": [ + 482, + 944, + 514, + 955 + ], + "page_idx": 3 + }, + { + "type": "table", + "img_path": "images/bb0766c09ee4b25d0c478175cd23422c96ff48a3e26b3a7b56aa403786dfacf5.jpg", + "table_caption": [ + "Table 1. Comparison of our approach with selected prior work on the R2R dataset. Methods that use additional visual data beyond MP3D [7] are annotated with $\\dagger$ and are not directly comparable with other approaches. Note that BEVBert* additionally uses panoramic depth images. Adding our visual imagination approach to HAMT and DUET base models (gray rows) leads to improved success rate (SR) and success weighted by inverse path length (SPL) metrics on the val-unseen split. For DUET, this effect also persists on the test set. We highlight these improvements over base models using bold numbers." + ], + "table_footnote": [], + "table_body": "
MethodsValidation SeenValidation UnseenTest Unseen
TLNE↓SR↑SPL↑TLNE↓SR↑SPL↑TLNE↓SR↑SPL↑
PREVALENT [14]10.323.67696510.194.71585310.515.305451
RecBERT [16]11.132.90726812.013.93635712.354.096357
ADAPT [27]10.972.54767212.213.77645812.993.796559
BEVBert* [3]13.562.17817414.552.81756415.873.137362
MARVAL [19]†10.602.99736910.154.06656110.224.186258
ScaleVLN [41]†11.902.16878012.402.34877914.272.738377
HAMT [8]11.152.5175.6172.1811.463.6266.2461.5112.273.936560
HAMT-Imagine (ours)11.302.3677.1673.7211.813.5867.2662.0212.663.896560
DUET [9]12.332.2878.8472.8813.943.3171.5260.4114.733.656959
DUET-Imagine (ours)12.932.1979.973.7514.353.1972.1260.4815.353.527160
", + "bbox": [ + 94, + 170, + 906, + 414 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "on the base method, these parameters may be the result of multi-stage pretraining and finetuning regimens - e.g., leveraging synthetic PREVALENT [14] data as prescribed in [8]. The imagination encoder is integrated to the base model and finetuned with the R2R-Imagine dataset. To mitigate catastrophic forgetting, we perform our training in three stages. First, we train only the newly introduced imagination encoder MLP along with type embeddings $t^{Im}$ with the remaining parameters frozen. Then, all the modules are trained jointly with a decreased learning rate for the base model. Finally, all the parameters are trained at a common learning rate. See supplementary materials for details on learning rates and schedules.", + "bbox": [ + 88, + 438, + 480, + 636 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "4. Experiments", + "text_level": 1, + "bbox": [ + 89, + 648, + 222, + 666 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Using the R2R dataset, we evaluate the impact of visual imaginations on our agent's navigation ability, perform ablation studies to explore the design space, and illustrate qualitative impacts of visual imaginations. Using the REVERIE dataset, we study the generalizability of our technique to a fine-grained instruction setting.", + "bbox": [ + 89, + 672, + 482, + 765 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "4.1. Experimental setup", + "text_level": 1, + "bbox": [ + 89, + 772, + 277, + 789 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "R2R dataset. R2R [4] is constructed using Matterport3D [7] simulator. It consists of 90 indoor environments with 10,567 panoramas (each panorama consists of 36 single-view images) which act as nodes in a navigation graph. There are 7,189 trajectories each with 3 instructions and a corresponding ground truth trajectory. The dataset is split into train (4675 trajectories), val-seen (340 trajectories),", + "bbox": [ + 89, + 794, + 482, + 901 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "val-unseen (783 trajectories), and test (1391 trajectories). The train and val-seen split share environments while val-unseen and test have different environments.", + "bbox": [ + 511, + 438, + 903, + 483 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "R2R-Imagine dataset. The curation process of R2R-Imagine is prescribed in section 3.1. The number of imagination images are 41,558 for train, 3,055 for val-seen, 6,857 for val-unseen, and 12,412 for test. Our imaginations have a resolution of $1024 \\times 1024$ and a single imagination generation on a single H100 GPU takes 3.2 seconds on average. We intend to release the dataset upon acceptance.", + "bbox": [ + 511, + 486, + 905, + 593 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "REVERIE dataset. REVERIE [33] is constructed using Matterport3D and consists of high-level goals instead of finer grained step-by-step instructions. These instructions generally convey the target object and its location. The agent has to rely more on exploration relative to R2R. To be successful, agents must reach and identify the referent object. To synthesize imaginations for REVERIE, we treat the entire instruction as a single sub-instruction and follow the same recipe used for R2R.", + "bbox": [ + 511, + 595, + 906, + 731 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Evaluation metrics. We employ standard VLN metrics [18]. Success rate (SR) is the ratio of instruction-trajectories where the agent stops within a 3 meter radius from the goal and success normalized by inverse path length (SPL) helps distinguish successful episodes with shorter path lengths from longer path lengths. Navigation error (NE) is the distance between an agent's final position and the goal in meters. Trajectory length (TL) is the length of an agent's traversed path in meters. REVERIE [33] additionally defines grounding metrics Remote Grounding Success (RGS) and RGS penalized by path length (RGSPL).", + "bbox": [ + 511, + 734, + 908, + 902 + ], + "page_idx": 4 + }, + { + "type": "page_number", + "text": "3850", + "bbox": [ + 482, + 944, + 514, + 955 + ], + "page_idx": 4 + }, + { + "type": "table", + "img_path": "images/4f11b4027be78f21cffcfb7844175ef8f383f7bbb2cf423cf19e2d85abc6c292.jpg", + "table_caption": [ + "Table 2. Performance comparison of our approach on REVERIE with baseline. We observe improvements of our method across all metrics on DUET when provided coarse-grained instructions." + ], + "table_footnote": [], + "table_body": "
SR↑SPL↑RGS↑RGSPL↑
DUET46.9833.7332.1523.03
DUET-Imagine (ours)48.2833.7632.9723.25
", + "bbox": [ + 94, + 142, + 480, + 204 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "HAMT and DUET base models. We apply our approach to two baseline representative architectures - HAMT [8] and DUET [9]. HAMT is hierarchical and transformer-based; instructions, observations, and trajectory history are encoded separately using unimodal transformers and then combined in a single cross-modal transformer to predict actions. DUET incrementally forms a topographic map of the environment over time and uses dual fine- and coarse-grained cross-modal attention mechanisms to associate language embeddings with the local observation and global topographic features. The results are then fused to predict actions. For both base models, we inject visual imaginations via concatenation with the text embeddings (Sec. 3.2). We use provided checkpoints for pretrained models.", + "bbox": [ + 89, + 227, + 483, + 440 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Implementation details. We follow base agent [8, 9] settings unless specified otherwise. We use a standard off-the-shelf pretrained ViT-B/16 [10] to encode imaginations for both HAMT and DUET. Our VLN-Imagine agents are finetuned for $100\\mathrm{k}$ iterations in three stages as noted in section 3.2. The finetuning process is carried out using a Tesla V100 GPU with a batch size of 8 and takes $\\sim 1.5$ days for each agent. For combining the proposed auxiliary loss function with the base model losses, we empirically set $\\lambda = 0.5$ .", + "bbox": [ + 89, + 441, + 483, + 580 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "4.2. Results", + "text_level": 1, + "bbox": [ + 89, + 585, + 184, + 602 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Our experiments demonstrate that providing generated images of instruction landmarks (imaginations) improves the performance of vision-and-language navigation models (Tab. 1). We highlight our key findings below.", + "bbox": [ + 89, + 608, + 483, + 670 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "VLN agents perform better with imaginations. VLN agents equipped with imaginations perform better than VLN agents without by 0.6-1.0 SR and up to 0.5 SPL on R2R val-unseen (Tab. 1). This improvement coincides with lower navigation error (NE) across both models and all dataset splits. We contextualize these results relative to agent architectures that are representative of the VLN literature. We additionally report a distinct family of techniques that scale training environments to boost performance by reducing overfitting - MARVAL [19] and ScaleVLN [41]. We consider approaches of this class orthogonal to our study and instead use visual data only from Matterport3D [7].", + "bbox": [ + 89, + 671, + 482, + 852 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Imaginations are useful in coarse-grained instruction navigation. Our method is performant when natural language instructions convey high-level objectives (REVERIE,", + "bbox": [ + 89, + 854, + 483, + 902 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Table 3. Role of imaginations; HAMT-Imagine. Null imaginations refers to testing with no imaginations, wrong imaginations refers to testing with imaginations sampled from different instructions. We observe aligned imaginations are important to navigation performance. In addition, results with no imaginations are better than baseline implying a regularization effect during training.", + "bbox": [ + 511, + 88, + 906, + 174 + ], + "page_idx": 5 + }, + { + "type": "table", + "img_path": "images/2302bfb6aa8afb5c39fabe6aff19ecc50463df477421a426597bfdd468998fab.jpg", + "table_caption": [], + "table_footnote": [], + "table_body": "
SR↑SPL↑
Baseline (HAMT)66.2461.51
Null imaginations66.9261.89
Wrong imaginations66.2461.02
Correct imaginations (ours)67.2662.02
", + "bbox": [ + 524, + 183, + 895, + 277 + ], + "page_idx": 5 + }, + { + "type": "table", + "img_path": "images/1083cb207d846927c2225f13979895166d468a8a0bfb316c011c379c72dd36b8.jpg", + "table_caption": [ + "Table 4. Sequential vs. goal imagination; HAMT-Imagine. We study the effect of providing just the imagination belonging to the final valid sub-instruction relative to full valid imaginations. We observe providing just the goal imagination improves performance significantly compared to baseline but shows lower performance relative to providing full imaginations." + ], + "table_footnote": [], + "table_body": "
SR↑SPL↑
Baseline (HAMT)66.2461.51
Goal imagination66.7961.58
Full imaginations (ours)67.2662.02
", + "bbox": [ + 524, + 383, + 895, + 463 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Tab. 2); navigation performance improves by 1.3 SR and grounding performance improves by 0.82 RGS. We hypothesize that imaginations not only aid in reaching the goal but also in identifying and grounding the target concept amongst others objects in the scene.", + "bbox": [ + 511, + 487, + 906, + 564 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "VLN agents benefit from both imagination training and inference. When nullifying imaginations at test-time, VLN-Imagine agents still surpass baseline performance, but to a lesser extent (Tab. 3). We implement this nullification by setting the imagination attention masks to zero in the cross-modal encoder. We hypothesize that imagination-based training provides a regularizing effect on the overall VLN agent during fine-tuning, revealing a research avenue for strong and performant test-time model inference. We further find that imaginations must be aligned with the instruction to provide maximal benefit; when fed with imaginations randomly-sampled from the R2R-Imagine dataset, VLN-Imagine agents perform worse than baseline (Tab. 3).", + "bbox": [ + 511, + 566, + 908, + 762 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Sequential imagination outperforms goal imagination. Sequential refers to imaginations generated over multiple sub-instructions (our method). Goal imagination refers to imagining just the final non-filtered sub-instruction. We show these results in Tab. 4 – while goal imagination outperforms the baseline by 0.5 SR, sequential imagination outperforms the baseline by 1.0 SR. Goal imagination may be particularly useful in guiding the agent to stop at the destination. While distinct from our setting, such goal imagi", + "bbox": [ + 511, + 763, + 908, + 902 + ], + "page_idx": 5 + }, + { + "type": "page_number", + "text": "3851", + "bbox": [ + 482, + 944, + 514, + 957 + ], + "page_idx": 5 + }, + { + "type": "table", + "img_path": "images/cd8705c5d24685f08e019bc71901cb98b1d5ce793fb1d692d7aeeb94d2778ada.jpg", + "table_caption": [ + "Table 5. Comparison between ViT models. We train both HAMT-Imagine and DUET-Imagine agents with a ViT finetuned in-the-loop with HAMT and an off-the-shelf pretrained ViT. We observe no consistent effects across the both agents." + ], + "table_footnote": [], + "table_body": "
ViT fine-tunedViT off-the-shelf
SR↑SPL↑SR↑SPL↑
HAMT-Imagine67.3161.567.2662.02
DUET-Imagine72.5860.2972.1260.48
", + "bbox": [ + 96, + 156, + 478, + 234 + ], + "page_idx": 6 + }, + { + "type": "table", + "img_path": "images/fc2c74a05cf3a9b184d167f53148ee3117c20e60bc35edfdafa8612d8a540897.jpg", + "table_caption": [ + "Table 6. Loss ablation. We study the effect of negative samples in our alignment loss by comparing our approach of using cosine loss with InfoNCE [28] loss. In addition, we train the agent with no auxiliary loss represented by \"No loss\". We do not observe a significant difference between contrastive and cosine losses." + ], + "table_footnote": [], + "table_body": "
SR↑SPL↑
No loss66.7561.60
LInfoNCE67.1861.84
Lcos(ours)67.2662.02
", + "bbox": [ + 102, + 330, + 473, + 409 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "nation has parallels to ImageNav [22], a visual navigation task that conveys goals solely using exemplar images.", + "bbox": [ + 89, + 435, + 482, + 467 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "General-purpose vision encoders are sufficient imagination encoders. We compare encoding imaginations with off-the-shelf ViT [10] vs. ViT fine-tuned on R2R navigation episodes [8]. We find comparable performance (Tab. 5) and choose to employ off-the-shelf ViT in our final model. This is motivated to maximize generality and to avoid potential issues relating to catastrophic forgetting. For example, generalization of frozen ViT is demonstrated in Zeng et al. [44].", + "bbox": [ + 89, + 469, + 482, + 590 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Aligning imaginations to instructions benefits VLN-imagine agents. Aligning the imagination embedding to the instruction embedding as an auxiliary loss leads to a 0.5 SR and a 0.4 SPL increase over no auxiliary loss (Tab. 6). We also consider formulating this loss as contrastive - do negative samples enable a stronger alignment and thus better downstream performance? We find that an InfoNCE loss [28] yields no significant difference relative to our cosine similarity loss (Eq. 2). We construct negative samples using mean noun phrase embeddings from imaginations of other instructions in the batch. The overall loss is $\\mathcal{L}_{\\mathrm{base}} + \\lambda \\mathcal{L}_{\\mathrm{InfoNCE}}$ where we set $\\lambda$ to 0.2 empirically.", + "bbox": [ + 89, + 594, + 482, + 777 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Imaginations correctly represent noun phrases in a subinstruction. To evaluate fidelity of our imaginations, we run an open-vocabulary object detector (LangSAM [1]) on our imaginations to test if the noun phrases contained in the corresponding sub-instruction from R2R are detected successfully. We see in Tab. 7 that at least one noun phrase is detected in our imaginations for $98.78\\%$ of sub-instructions and all noun phrases are detected for $94.99\\%$", + "bbox": [ + 89, + 780, + 482, + 902 + ], + "page_idx": 6 + }, + { + "type": "table", + "img_path": "images/c06c22d2b57ce9fd3fbf66a02641237eb4b2ad7e1e4db1bc6bee297ba28c211c.jpg", + "table_caption": [ + "Table 7. Open vocabulary object detection accuracy (\\%) of imaginations to noun phrases from sub-instructions. Our imaginations represent the noun phrases with a high accuracy. In val-seen, in $98.43\\%$ of sub-instructions, at least 1 noun phrase is detected and in $94.51\\%$ of sub-instructions, all noun phrases are detected." + ], + "table_footnote": [], + "table_body": "
Sub-instructionsNoun Phrase Detection (%)
≥1 detectedAll detected
Val-seen298598.4394.51
Val-unseen670098.9995.34
Test1207598.9395.12
", + "bbox": [ + 514, + 169, + 906, + 268 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "of sub-instructions. We conclude our imaginations provide a high coverage of relevant concepts across splits.", + "bbox": [ + 511, + 294, + 906, + 325 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "4.3. Qualitative visualizations", + "text_level": 1, + "bbox": [ + 511, + 334, + 745, + 349 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "In this section, we provide a qualitative investigation of how imaginations may be used as a pivot between language tokens and visual observations from the environment by examining the attention patterns in HAMT-Imagine.", + "bbox": [ + 511, + 356, + 905, + 416 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "For a given instruction, we randomly select an imagination / sub-instruction pair and examine the agent's trajectory to identify the first occurrence of the reference object. We then examine the attention distributions between language tokens, the visual imagination, and the panoramic observation. As there are many attention heads across multiple layers in the cross-modal encoder, we structure our study by focusing on those attention heads where sub-instruction tokens are attended to most strongly by their corresponding visual imagination and then examine how the visual imagination attends to observations.", + "bbox": [ + 511, + 417, + 905, + 582 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "We then visualize sub-instruction and imagination pairs along with their top attended language tokens and visual observations in Fig. 4. By construction, these are time steps where the referent is visible and the visual imagination is attending to the appropriate noun phrase. As such, our examples primarily demonstrate how the imagination learns to associate with the visual observations. We focus on this aspect as our auxiliary loss already explicitly encourages alignment between visual imaginations and sub-instruction.", + "bbox": [ + 511, + 583, + 905, + 718 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "From the examples in Fig. 4, we make two remarks. Imaginations attend most strongly to observations of matching semantic concepts. In row 1, the imagination and the retrieved observations all depict a cheetah (either a painting or floor mat). Imagination attention scores mirror the proper disambiguation of concepts. In row 2, the sub-instruction (and thus the imagination) refer to a unicycle. While both a bicycle and unicycle are observed during execution, attention scores are maximized by the observations containing the unicycle. These patterns do not directly imply a causal relation between the attention scores and eventual decision making of the VLN agent. Rather, they illustrate how imag-", + "bbox": [ + 511, + 719, + 908, + 902 + ], + "page_idx": 6 + }, + { + "type": "page_number", + "text": "3852", + "bbox": [ + 482, + 944, + 514, + 955 + ], + "page_idx": 6 + }, + { + "type": "table", + "img_path": "images/83b87358d522a73bc97d26a609e9b03519ebeffc31db66d8bd385fb3f80b4813.jpg", + "table_caption": [], + "table_footnote": [], + "table_body": "
Sub-instructionImaginationTop Attended TokensTop Attended Observations
And walk towards the door pass by the cheetah paintingche, ##eta, ##h, painting
Go stand behind the easel by a unicycle next to a doorun, ##ic. ##y, ##cle, to
Once in the room, walk to the left of the pool tablepool, door, room
Walk into the arch hallway near the Native American painting on the wallAmerican, Native, candles
", + "bbox": [ + 91, + 88, + 906, + 415 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Figure 4. Qualitative examples showing imaginations as pivots between language and observation images. The first column contains sub-instruction from a random instruction from R2R, the second column contains the imagination generated using the sub-instruction. The third and fourth columns show highest attended language tokens and observation images from an attention head in HAMT's cross-modal transformer at a time step the associated observations are first visible. In the second example (row 2), the sub-instruction references \"unicycle\" which is captured in the imagination along with neighboring nouns \"easel\" and \"door\". We observe that in a head where top attending language tokens to the imagination query are references to nouns associated with the sub-instruction, its top attended observations to the imagination query are images of the same concept (\"unicycle\"). In this example, the imagination of a unicycle is being used to associate language tokens belonging to \"unicycle\" to observations of unicycle hinting at the utility of imaginations in navigation.", + "bbox": [ + 89, + 426, + 906, + 539 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "inations can serve as a pivot between language and observations and afford disambiguation of related concepts.", + "bbox": [ + 89, + 565, + 482, + 595 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "5. Conclusion", + "text_level": 1, + "bbox": [ + 89, + 613, + 209, + 628 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "In this work, we show that visual representations of instructions, or visual imaginations, play an additive role in training superior navigating agents for vision-and-language navigation (VLN) tasks. We generate a sequence of visual imaginations depicting sub-instruction landmarks using a text-to-image diffusion model. Our approach is model-agnostic and we integrate these imaginations into VLN agents using an imagination encoder, an auxiliary loss, and a fine-tuning training step. We apply this approach to two representative VLN agents across two datasets and observe performance improvements of around 1 SR and 0.5 SPL.", + "bbox": [ + 89, + 640, + 482, + 805 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Limitations. Generating and encoding imaginations increases the computational cost of running VLN agents. This is particularly relevant when running agents on-device for real-world robotic deployments. However, we note that this process is performed upfront, unlike techniques that depend on step-wise observations like novel view synthesis (NVS).", + "bbox": [ + 89, + 810, + 483, + 901 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Stepping more broadly, by design, our imaginations are not grounded in the environment. Personalized reasoning, such as the unique naming of objects and locations, cannot be directly imagined in our framework. Life-long learning of persistent groundings is an open question not just in generating imaginations, but in the VLN space at large.", + "bbox": [ + 511, + 565, + 906, + 655 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Future directions. A natural question that arises from our research is the exact complementary nature of the roles played by visual imaginations and language instructions. This is a rich area for future work. Similar studies relating language and visual observations have been conducted in [48] and [47]. Our paper demonstrated the fundamental utility of imagination in VLN agents. Beyond this lies interesting extensions, such as how imagination can help bridge the simulation-to-reality (Sim2Real) gap for VLN agents and whether imaginations can unlock the performance of VLN world models through image-image reasoning.", + "bbox": [ + 511, + 672, + 908, + 839 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Acknowledgments. We thank Zijiao Yang and Robert Yelle for their assistance with simulator issues and technical support for high performance computing infrastructure.", + "bbox": [ + 511, + 854, + 908, + 902 + ], + "page_idx": 7 + }, + { + "type": "page_number", + "text": "3853", + "bbox": [ + 482, + 944, + 514, + 955 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "References", + "text_level": 1, + "bbox": [ + 91, + 90, + 187, + 104 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[1] Language segment-anything. https://github.com/luca-medeiros/lang-segment-anything/.7", + "[2] spacy: Industrial-strength natural language processing in python. https://spacy.io/.3", + "[3] Dong An, Yuankai Qi, Yangguang Li, Yan Huang, Liang Wang, Tieniu Tan, and Jing Shao. Bevbert: Multimodal map pre-training for language-guided navigation. Proceedings of the IEEE/CVF International Conference on Computer Vision, 2023. 5", + "[4] Peter Anderson, Qi Wu, Damien Teney, Jake Bruce, Mark Johnson, Niko Sünderhauf, Ian D. Reid, Stephen Gould, and Anton van den Hengel. Vision-and-language navigation: Interpreting visually-grounded navigation instructions in real environments. 2018 IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 3674-3683, 2017. 2, 3, 5", + "[5] Homanga Bharadhwaj, Debidatta Dwibedi, Abhinav Gupta, Shubham Tulsiani, Carl Doersch, Ted Xiao, Dhruv Shah, Fei Xia, Dorsa Sadigh, and Sean Kirmani. Gen2act: Human video generation in novel scenarios enables generalizable robot manipulation. arXiv preprint arXiv:2409.16283, 2024. 3", + "[6] Kevin Black, Mitsuhiko Nakamoto, Pranav Atreya, Homer Rich Walke, Chelsea Finn, Aviral Kumar, and Sergey Levine. Zero-shot robotic manipulation with pre-trained image-editing diffusion models. In The Twelfth International Conference on Learning Representations. 3", + "[7] Angel Chang, Angela Dai, Thomas Funkhouser, Maciej Halber, Matthias Niebner, Manolis Savva, Shuran Song, Andy Zeng, and Yinda Zhang. Matterport3d: Learning from rgb-d data in indoor environments. In 2017 International Conference on 3D Vision (3DV), pages 667-676, 2017. 2, 4, 5, 6", + "[8] Shizhe Chen, Pierre-Louis Guhur, Cordelia Schmid, and Ivan Laptev. History aware multimodal transformer for vision-and-language navigation. In NeurIPS, 2021. 2, 3, 4, 5, 6, 7", + "[9] Shizhe Chen, Pierre-Louis Guhur, Makarand Tapaswi, Cordelia Schmid, and Ivan Laptev. Think global, act local: Dual-scale graph transformer for vision-and-language navigation. In CVPR, 2022. 2, 3, 4, 5, 6", + "[10] Alexey Dosovitskiy, Lucas Beyer, Alexander Kolesnikov, Dirk Weissenborn, Xiaohua Zhai, Thomas Unterthiner, Mostafa Dehghani, Matthias Minderer, Georg Heigold, Sylvain Gelly, et al. An image is worth 16x16 words: Transformers for image recognition at scale. arXiv preprint arXiv:2010.11929, 2020. 4, 6, 7", + "[11] Yilun Du, Sherry Yang, Bo Dai, Hanjun Dai, Ofir Nachum, Josh Tenenbaum, Dale Schuurmans, and Pieter Abbeel. Learning universal policies via text-guided video generation. Advances in Neural Information Processing Systems, 36, 2024. 3", + "[12] Daniel Fried, Ronghang Hu, Volkan Cirik, Anna Rohrbach, Jacob Andreas, Louis-Philippe Morency, Taylor Berg-Kirkpatrick, Kate Saenko, Dan Klein, and Trevor Darrell. Speaker-follower models for vision-and-language naviga" + ], + "bbox": [ + 93, + 114, + 482, + 900 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "tion. In Neural Information Processing Systems (NeurIPS), 2018. 2", + "[13] Jialu Gao, Kaizhe Hu, Guowei Xu, and Huazhe Xu. Can pre-trained text-to-image models generate visual goals for reinforcement learning? Advances in Neural Information Processing Systems, 36:38297-38310, 2023. 3", + "[14] Weituo Hao, Chunyuan Li, Xiujun Li, Lawrence Carin, and Jianfeng Gao. Towards learning a generic agent for vision-and-language navigation via pre-training, 2020. 2, 5", + "[15] Yicong Hong, Cristian Rodriguez, Qi Wu, and Stephen Gould. Sub-instruction aware vision-and-language navigation. In Proceedings of the 2020 Conference on Empirical Methods in Natural Language Processing (EMNLP), pages 3360-3376, 2020. 3, 4", + "[16] Yicong Hong, Qi Wu, Yuankai Qi, Cristian Rodriguez-Opazo, and Stephen Gould. A recurrent vision-and-language bert for navigation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 1643–1653, 2021. 5", + "[17] Yicong Hong, Qi Wu, Yuankai Qi, Cristian Rodriguez-Opazo, and Stephen Gould. A recurrent vision-and-language bert for navigation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 1643–1653, 2021. 2", + "[18] Gabriel Ilharco, Vihan Jain, Alexander Ku, Eugene Ie, and Jason Baldridge. General evaluation for instruction conditioned navigation using dynamic time warping. ArXiv, abs/1907.05446, 2019. 5", + "[19] Aishwarya Kamath, Peter Anderson, Su Wang, Jing Yu Koh, Alexander Ku, Austin Waters, Yinfei Yang, Jason Baldridge, and Zarana Parekh. A new path: Scaling vision-and-language navigation with synthetic instructions and imitation learning. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 10813-10823, 2023. 2, 5, 6", + "[20] Ivan Kapelyukh, Vitalis Vosylius, and Edward Johns. Dallebot: Introducing web-scale diffusion models to robotics. 2023. 3", + "[21] Jing Yu Koh, Honglak Lee, Yinfei Yang, Jason Baldridge, and Peter Anderson. Pathdreamer: A world model for indoor navigation, 2021. 3", + "[22] Jacob Krantz, Stefan Lee, Jitendra Malik, Dhruv Batra, and Devendra Singh Chaplot. Instance-specific image goal navigation: Training embodied agents to find object instances, 2022. 7", + "[23] Chongxi Lai, Shinsuke Tanaka, Timothy D. Harris, and Albert K. Lee. Volitional activation of remote place representations with a hippocampal brain-machine interface. Science, 382(6670):566-573, 2023. 2", + "[24] Jialu Li and Mohit Bansal. Panogen: Text-conditioned panoramic environment generation for vision-and-language navigation. arxiv, 2023. 3", + "[25] Jialu Li and Mohit Bansal. Improving vision-and-language navigation by generating future-view image semantics. CVPR, 2023. 3", + "[26] Mingxiao Li, Zehao Wang, Tinne Tuytelaars, and Marie-Francine Moens. Layout-aware dreamer for embodied visual referring expression grounding. In Proceedings of the" + ], + "bbox": [ + 517, + 93, + 903, + 900 + ], + "page_idx": 8 + }, + { + "type": "page_number", + "text": "3854", + "bbox": [ + 482, + 944, + 514, + 955 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "AAAI Conference on Artificial Intelligence, pages 1386-1395, 2023. 2", + "[27] Bingqian Lin, Yi Zhu, Zicong Chen, Xiwen Liang, Jianzhuang Liu, and Xiaodan Liang. Adapt: Vision-language navigation with modality-aligned action prompts. In 2022 IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 15375-15385, 2022. 2, 5", + "[28] Aaron van den Oord, Yazhe Li, and Oriol Vinyals. Representation learning with contrastive predictive coding. arXiv preprint arXiv:1807.03748, 2018. 7", + "[29] Allan Paivio. Abstractness, imagery, and meaningfulness in paired-associate learning. Journal of verbal learning and verbal behavior., 4(1), 1965-2. 2", + "[30] Allan Paivio. A factor-analytic study of word attributes and verbal learning. Journal of verbal learning and verbal behavior, 7(1), 1968-2.", + "[31] A. Paivio, P. C. Smythe, and J. C. Yuille. Imagery versus meaningfulness of nouns in paired-associate learning. Canadian Journal of Psychology / Revue Canadienne De Psychologie, 22:427–441, 1968. 2", + "[32] Dustin Podell, Zion English, Kyle Lacey, Andreas Blattmann, Tim Dockhorn, Jonas Müller, Joe Penna, and Robin Rombach. Sdxl: Improving latent diffusion models for high-resolution image synthesis, 2023. 2, 3, 4", + "[33] Yuankai Qi, Qi Wu, Peter Anderson, Xin Wang, William Yang Wang, Chunhua Shen, and Anton van den Hengel. Reverie: Remote embodied visual referring expression in real indoor environments. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 9982-9991, 2020. 2, 5", + "[34] Alec Radford, Jong Wook Kim, Chris Hallacy, Aditya Ramesh, Gabriel Goh, Sandhini Agarwal, Girish Sastry, Amanda Askell, Pamela Mishkin, Jack Clark, et al. Learning transferable visual models from natural language supervision. In International conference on machine learning, pages 8748-8763. PMLR, 2021. 1, 2", + "[35] Santhosh K Ramakrishnan, Aaron Gokaslan, Erik Wijmans, Oleksandr Maksymets, Alex Clegg, John Turner, Eric Undersander, Wojciech Galuba, Andrew Westbury, Angel X Chang, et al. Habitat-matterport 3d dataset (hm3d): 1000 large-scale 3d environments for embodied ai. arXiv preprint arXiv:2109.08238, 2021. 2", + "[36] Marianne Cumella Reddan, Tor Dessart Wager, and Daniela Schiller. Attenuating neural threat expression with imagination. *Neuron*, 100(4):994–1005.e4, 2018. 2", + "[37] Manolis Savva, Abhishek Kadian, Oleksandr Maksymets, Yili Zhao, Erik Wijmans, Bhavana Jain, Julian Straub, Jia Liu, Vladlen Koltun, Jitendra Malik, et al. Habitat: A platform for embodied ai research. In Proceedings of the IEEE/CVF international conference on computer vision, pages 9339-9347, 2019. 2", + "[38] Daniel L. Schacter, Donna Rose Addis, and Randy L. Buckner. Remembering the past to imagine the future: the prospective brain. Nature Reviews Neuroscience, 8(9):657-661, 2007. 2", + "[39] Roger N. Shepard and Jacqueline Metzler. Mental rotation of three-dimensional objects. Science, 171(3972):701-703, 1971. 2" + ], + "bbox": [ + 91, + 90, + 482, + 898 + ], + "page_idx": 9 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[40] Hao Tan, Licheng Yu, and Mohit Bansal. Learning to navigate unseen environments: Back translation with environmental dropout. In Proceedings of the 2019 Conference of the North American Chapter of the Association for Computational Linguistics: Human Language Technologies, Volume 1 (Long and Short Papers), pages 2610-2621, Minneapolis, Minnesota, 2019. Association for Computational Linguistics. 2", + "[41] Zun Wang, Jialu Li, Yicong Hong, Yi Wang, Qi Wu, Mohit Bansal, Stephen Gould, Hao Tan, and Yu Qiao. Scaling data generation in vision-and-language navigation. In ICCV 2023, 2023. 2, 5, 6", + "[42] Fei Xia, Amir R. Zamir, Zhi-Yang He, Alexander Sax, Jitendra Malik, and Silvio Savarese. Gibson env: real-world perception for embodied agents. In Computer Vision and Pattern Recognition (CVPR), 2018 IEEE Conference on. IEEE, 2018. 2", + "[43] Zijiao Yang, Arjun Majumdar, and Stefan Lee. Behavioral analysis of vision-and-language navigation agents. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 2574-2582, 2023. 1", + "[44] Kuo-Hao Zeng, Zichen Zhang, Kiana Ehsani, Rose Hendrix, Jordi Salvador, Alvaro Herrasti, Ross Girshick, Aniruddha Kembhavi, and Luca Weihs. Poliformer: Scaling on-policy rl with transformers results in masterful navigators. In 8th Annual Conference on Robot Learning, 2024. 7", + "[45] Gengze Zhou, Yicong Hong, and Qi Wu. Navigpt: Explicit reasoning in vision-and-language navigation with large language models. Proceedings of the AAAI Conference on Artificial Intelligence, 38(7):7641-7649, 2024. 2", + "[46] Gengze Zhou, Yicong Hong, Zun Wang, Xin Eric Wang, and Qi Wu. Navigpt-2: Unleashing navigational reasoning capability for large vision-language models. In European Conference on Computer Vision, pages 260–278. Springer, 2025. 2", + "[47] Wanrong Zhu, Yuankai Qi, Pradyumna Narayana, Kazoo Sone, Sugato Basu, Xin Wang, Qi Wu, Miguel Eckstein, and William Yang Wang. Diagnosing vision-and-language navigation: What really matters. In Proceedings of the 2022 Conference of the North American Chapter of the Association for Computational Linguistics: Human Language Technologies, pages 5981-5993, 2022. 1, 8", + "[48] Wang Zhu, Ishika Singh, Yuan Huang, Robin Jia, and Jesse Thomason. Does vln pretraining work with nonsensical or irrelevant instructions? ArXiv, abs/2311.17280, 2023. 8" + ], + "bbox": [ + 516, + 92, + 906, + 723 + ], + "page_idx": 9 + }, + { + "type": "page_number", + "text": "3855", + "bbox": [ + 482, + 944, + 514, + 955 + ], + "page_idx": 9 + } +] \ No newline at end of file diff --git a/2025/Do Visual Imaginations Improve Vision-and-Language Navigation Agents_/6684e0af-b484-43ef-8443-50d4bfdefdd1_model.json b/2025/Do Visual Imaginations Improve Vision-and-Language Navigation Agents_/6684e0af-b484-43ef-8443-50d4bfdefdd1_model.json new file mode 100644 index 0000000000000000000000000000000000000000..ee3687a38c4cb7c944b5a20e62a61bfe368df9b7 --- /dev/null +++ b/2025/Do Visual Imaginations Improve Vision-and-Language Navigation Agents_/6684e0af-b484-43ef-8443-50d4bfdefdd1_model.json @@ -0,0 +1,1969 @@ +[ + [ + { + "type": "header", + "bbox": [ + 0.107, + 0.003, + 0.182, + 0.043 + ], + "angle": 0, + "content": "CVF" + }, + { + "type": "header", + "bbox": [ + 0.237, + 0.001, + 0.812, + 0.047 + ], + "angle": 0, + "content": "This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore." + }, + { + "type": "title", + "bbox": [ + 0.122, + 0.131, + 0.877, + 0.153 + ], + "angle": 0, + "content": "Do Visual Imaginations Improve Vision-and-Language Navigation Agents?" + }, + { + "type": "text", + "bbox": [ + 0.316, + 0.181, + 0.68, + 0.218 + ], + "angle": 0, + "content": "Akhil Perincherry1, Jacob Krantz, Stefan Lee1 \n1Oregon State University" + }, + { + "type": "text", + "bbox": [ + 0.181, + 0.22, + 0.812, + 0.234 + ], + "angle": 0, + "content": "perincha@oregonstate.edu, jkrantz989@gmail.com, leestef@oregonstate.edu" + }, + { + "type": "title", + "bbox": [ + 0.248, + 0.269, + 0.327, + 0.285 + ], + "angle": 0, + "content": "Abstract" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.302, + 0.485, + 0.59 + ], + "angle": 0, + "content": "Vision-and-Language Navigation (VLN) agents are tasked with navigating an unseen environment using natural language instructions. In this work, we study if visual representations of sub-goals implied by the instructions can serve as navigational cues and lead to increased navigation performance. To synthesize these visual representations or \"imaginations\", we leverage a text-to-image diffusion model on landmark references contained in segmented instructions. These imaginations are provided to VLN agents as an added modality to act as landmark cues and an auxiliary loss is added to explicitly encourage relating these with their corresponding referring expressions. Our findings reveal an increase in success rate (SR) of \\(\\sim 1\\) point and up to \\(\\sim 0.5\\) points in success scaled by inverse path length (SPL) across agents. These results suggest that the proposed approach reinforces visual understanding compared to relying on language instructions alone. Code and data for our work can be found at https://www.akhilperincherry.com/VLN-Imagine-website/." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.619, + 0.222, + 0.634 + ], + "angle": 0, + "content": "1. Introduction" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.644, + 0.484, + 0.901 + ], + "angle": 0, + "content": "In this work, we examine whether providing visual imagery corresponding to described landmarks improves the performance of agents following natural language navigation instructions. Consider the natural language navigation instruction presented in Figure 1 which asks an agent to \"Go straight, take a left at the pool table to enter the kitchen. Walk to the bedroom and stop\". This instruction provides unconditional action directives like \"go straight\" but also frequently conditions the given directions on visual landmarks in the scene such as the pool table, kitchen, and bedroom. Standard approaches to vision-and-language navigation tasks rely on learned cross-modal alignment mechanisms to implicitly associate these noun phrases with their visual referents during navigation. However, text-to-image generation models have improved to the point that producing imagery matching the semantics of these visual references prior to navigation is plausible. We show three such" + }, + { + "type": "title", + "bbox": [ + 0.527, + 0.303, + 0.755, + 0.315 + ], + "angle": 0, + "content": "Natural Language Navigation Instruction:" + }, + { + "type": "image", + "bbox": [ + 0.529, + 0.319, + 0.563, + 0.348 + ], + "angle": 0, + "content": null + }, + { + "type": "text", + "bbox": [ + 0.575, + 0.32, + 0.895, + 0.348 + ], + "angle": 0, + "content": "Go straight, take a left at the pool table to enter the kitchen. Walk to the bedroom and stop." + }, + { + "type": "title", + "bbox": [ + 0.528, + 0.358, + 0.64, + 0.37 + ], + "angle": 0, + "content": "Visual Imaginations:" + }, + { + "type": "image", + "bbox": [ + 0.524, + 0.371, + 0.565, + 0.457 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.575, + 0.374, + 0.676, + 0.457 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.686, + 0.374, + 0.785, + 0.456 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.796, + 0.374, + 0.901, + 0.457 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.512, + 0.471, + 0.907, + 0.555 + ], + "angle": 0, + "content": "Figure 1. Illustration of visual imaginations. (Top) A natural language instruction specifying sub-goals pool table, kitchen, and bedroom. (Bottom) Visual imaginations of landmarks pool table, kitchen and bedroom referenced by the sub-goals in the instruction. In our work, we study if these visual imaginations generated using text-to-image models can improve performance in VLN." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.569, + 0.907, + 0.674 + ], + "angle": 0, + "content": "generations for our running example in Figure 1. Drawing an analogy to the substantial work in cognitive science on the impact of mental imagery on task performance, we refer to these generated images as visual imaginations and study whether providing them in addition to corresponding language-based instructions can improve the performance of vision-and-language navigation agents." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.675, + 0.909, + 0.902 + ], + "angle": 0, + "content": "But why might we believe visual imaginations could be beneficial? In short - text-to-image models have broad knowledge, and learning to perform semantic matching in the image domain may be an easier task than language grounding. Associating noun phrases and their visual referents is a sub-task in vision-and-language navigation (VLN) that most works address through extensive pretraining of model components on web-scale image caption datasets - e.g., by initializing image and language encoders with CLIP [34] models. While this provides a useful warm-start, it is unclear how well the remaining cross-modal components that drive navigation retain these capabilities after downstream VLN training. Downstream datasets may cover only a small portion of relevant objects and visual grounding remains a weakness in many existing VLN models [43, 47]." + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.516, + 0.957 + ], + "angle": 0, + "content": "3846" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.09, + 0.092, + 0.482, + 0.333 + ], + "angle": 0, + "content": "Directly synthesizing visual imaginations using off-the-shelf text-to-image models offers alternative training and inference mechanisms. Synthesized images have known correspondences to noun phrases in instructions that can be leveraged to reinforce visual grounding during downstream task training through auxiliary losses. Further, matching a visual imagination with environment observations reduces to an image-to-image matching task – offering the potential to use visual imaginations as a convenient pivot between language and vision. This may prove especially effective for out-of-distribution landmarks that have limited or no support during downstream task training. For example, novel instructions referencing “butterfly sculptures” or “Pulp Fiction posters” are unlikely to be well-grounded in the training set but produce semantically relevant visual imaginations that could then be matched to observations." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.335, + 0.482, + 0.487 + ], + "angle": 0, + "content": "Though we do not imply any direct connections, our work is also inspired by cognitive science studies which indicate mental imagery can serve as valuable cues in addition to linguistic stimuli. Several works supporting dualcoding theory [29-31] find that concrete imagery tends to be more effective for certain types of concept learning compared to linguistic stimuli alone. Further, findings from human and animal studies [23, 36, 38, 39] suggest that imagining snapshots of the environment or potential trajectories before navigating can improve decision making." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.489, + 0.482, + 0.625 + ], + "angle": 0, + "content": "To investigate our key question, we augment agents with visual imaginations for the popular Room-2-Room (R2R) [4] and REVERIE [33] tasks. We propose a VLN agent agnostic procedure to incorporate visual imaginations as additional inputs and introduce a text-imagination alignment loss to explicitly encourage visual grounding during training. We generate visual imaginations for key noun phrases in instructions that refer to visual landmarks using an off-the-shelf text-to-image generation model [32]." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.628, + 0.482, + 0.779 + ], + "angle": 0, + "content": "Applying our technique to two existing models, HAMT [8] and DUET [9], we find modest improvements - R2R validation unseen success increases by roughly 1.0 and 0.6 points respectively. This translates to an improvement of 2 points in success on the test set for our modified DUET model. On REVERIE validation unseen, we improve DUET performance by 1.3 points success and 0.82 points grounding success. Our results suggest that including explicit visual depictions of referred objects can improve performance on vision-and-language navigation tasks." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.783, + 0.438, + 0.797 + ], + "angle": 0, + "content": "Contributions. We summarize our contributions as:" + }, + { + "type": "text", + "bbox": [ + 0.095, + 0.806, + 0.483, + 0.867 + ], + "angle": 0, + "content": "- We develop a pipeline for generating visual imaginations from navigation instructions and synthesize the R2R-Imagine dataset to enable studying the impact of text-to-image models on VLN agents." + }, + { + "type": "text", + "bbox": [ + 0.095, + 0.871, + 0.483, + 0.902 + ], + "angle": 0, + "content": "- We propose an agent-agnostic method to incorporate visual imaginations into existing VLN agents and show" + }, + { + "type": "list", + "bbox": [ + 0.095, + 0.806, + 0.483, + 0.902 + ], + "angle": 0, + "content": null + }, + { + "type": "text", + "bbox": [ + 0.534, + 0.092, + 0.905, + 0.121 + ], + "angle": 0, + "content": "improved performance for HAMT [8] and DUET [9] across R2R and REVERIE." + }, + { + "type": "text", + "bbox": [ + 0.518, + 0.127, + 0.905, + 0.157 + ], + "angle": 0, + "content": "- We provide ablations of modeling decisions to characterize the design space for adding visual imaginations." + }, + { + "type": "title", + "bbox": [ + 0.514, + 0.176, + 0.655, + 0.192 + ], + "angle": 0, + "content": "2. Related Work" + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.203, + 0.905, + 0.61 + ], + "angle": 0, + "content": "Vision-and-Language Navigation (VLN). Navigating agents that can operate in novel environments using freeform natural language lends itself as personal assistants, field-robots and search-and-rescue agents. Development of VLN agents has been facilitated by photorealistic simulators such as Matterport3D [7] and paired language-action datasets such as Room-2-Room (R2R) [4] and REVERIE [33]. R2R provides fine-grained instructions conveying low-level directives such as \"Go left, then right at the hall\". REVERIE specifies coarse-grained instructions conveying broad high-level goals such as \"Adjust the picture by the lamp in the hall\". Progress in VLN has broadly ranged from architectural improvements and data scaling techniques. Architectural improvements include evolution from recurrent systems [4] to transformer architectures [8, 9, 17]. HAMT [8] uses a hierarchical transformer architecture to encode spatial and temporal information. DUET [9] maintains a topological map and performs coarse and fine scale encoding over it. Data scaling techniques include augmenting environments [14, 40] using an action-to-text speaker model [12] to generate synthetic instructions, and scaling environments [19, 41] with additional scenes such as HM3D [35] and Gibson [42] in Habitat [37]. Recently, zero-shot or learned integrations of pretrained large language models for VLN [45, 46] is showing promising performance. However, running large models in-the-loop incurs high computational costs." + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.612, + 0.905, + 0.809 + ], + "angle": 0, + "content": "Closely related to our work, ADAPT [27] augments VLN agents with visual observations related to an instruction. Language-observation pairs are stored from the training set and multiple pairs of matching object/location nouns from instructions are retrieved at test time using CLIP [34] scores. Fundamentally, the ADAPT image base is limited to training environments whereas our imaginations cover an open distribution. Therefore, our method can be more representative of the instruction and visualize novel objects/locations. For instance, a sub-instruction of \"walk into the kitchen with blue walls\" would retrieve images of kitchens from the training environment using ADAPT, but our imaginations would convey a kitchen with blue walls." + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.811, + 0.905, + 0.901 + ], + "angle": 0, + "content": "Another adjacent work, LAD [26] leverages layout dreamer and goal dreamer modules optimized around a topological graph to navigate via coarse-grained instructions. The capability of our method contrasts with LAD inthree primary ways: 1) method generality; Our method is performant across models and datasets, whereas LAD is a" + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.516, + 0.956 + ], + "angle": 0, + "content": "3847" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.089, + 0.092, + 0.482, + 0.183 + ], + "angle": 0, + "content": "full architecture constructed around a particular topological graph in coarse-grained settings, 2) sequential imagination: we imagine a sequence of landmarks along a path to inform sequential decision-making, whereas LAD imagines just the goal, and 3) we perform early fusion of imaginations and language whereas LAD employs late fusion." + }, + { + "type": "text", + "bbox": [ + 0.093, + 0.186, + 0.487, + 0.488 + ], + "angle": 0, + "content": "Image generation in VLN. Several techniques have proposed generative models to predict future visual observations or entire environments in VLN [21, 24, 25]. Pathdreamer [21] developed a GAN-based model to predict panoramic observations after a navigation action in VLN, conditioned on prior visual observations. Agents utilizing these predicted observations in a fixed-horizon search with a learned trajectory-instruction alignment heuristic achieve non-trivial success compared to simple baselines, but fall short of planning on actual observations and standard VLN models. Unlike our approach, Pathdreamer does not condition on instruction text or explore training agents on generated imagery. Similarly, PanoGen [24] uses a diffusion model to also generate panorama observations, but conditions on image captions while doing so as a form of visual dataset augmentation – replacing Matterport3D observations with synthetic generations with matching semantics. In contrast to our approach, this data augmentation does not provide additional inputs to the VLN models, is not conditioned on instructions, and is not active at inference time." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.49, + 0.483, + 0.656 + ], + "angle": 0, + "content": "Related to our work is VLN-Sig [25], which adapts VLN agents to predict quantized features of future visual observations given the instruction and observation history. This supports additional pretraining tasks and an auxiliary loss during downstream VLN finetuning for predicting next-step image semantics. Unlike our approach, VLN-Sig relies on the trained VLN agent to predict future image properties and focuses more on next-step prediction. In contrast, we leverage broad vision-and-language knowledge from text-to-image models and focus generating instruction-referred visual landmarks to improve agent performance." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.66, + 0.483, + 0.902 + ], + "angle": 0, + "content": "Image generation in robot control. Taking a wider view, several methods for controlling embodied agents have leveraged text-conditioned generative techniques to provide goal specifications in the form of images or videos. For tabletop object rearrangement tasks, many works use text-to-image models to synthesize goals or sub-goals in response to natural language commands, see [6, 13, 20] for a nonexhaustive sample. Others produce denser conditioning, leveraging text-to-video models to generate entire potential robot trajectories [5, 11]. While similar in motivation to our approach, these techniques focus on tabletop manipulation settings which introduce difficult physical challenges but significantly limit the problem to constrained spaces that remain largely observed throughout operation. As such, sub-goals are often edits to existing observations rather than generations of yet-to-be-seen landmarks as in our approach." + }, + { + "type": "title", + "bbox": [ + 0.514, + 0.09, + 0.65, + 0.108 + ], + "angle": 0, + "content": "3. Methodology" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.116, + 0.905, + 0.207 + ], + "angle": 0, + "content": "Our proposed methodology consists of two components - (1) a visual imagination generation pipeline and (2) a model-agnostic approach to integrating visual imaginations into existing Vision-and-Language Navigation (VLN) agents. Before defining these, we start with establishing relevant notation for the VLN task." + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.21, + 0.907, + 0.468 + ], + "angle": 0, + "content": "Problem definition. In the standard VLN task [4], an agent is prompted with a natural language navigation instruction and must make a sequence of actions to navigate along the described trajectory based on visual observations. We denote the \\(L\\)-word instruction as \\(W = (w_{1}, w_{2}, \\dots, w_{L})\\) and the visual observation at time step \\(t\\) as a panoramic observation \\(O_{t}\\) consisting of \\(K\\) single view images \\(O_{t} = (I_{1}^{t}, \\dots, I_{K}^{t})\\). Following prior work, we consider \\(K = 36\\) corresponding to images from 12 headings and 3 pitches. Amongst the single view images, a subset \\(\\mathcal{A}_{t} \\subset O_{t}\\) correspond to navigable directions. Given the instruction \\(W\\) and history of observations \\(O_{\\leq t}\\) until time \\(t\\), the agent selects an action \\(a_{t} \\in \\mathcal{A}_{t} \\cup \\{\\text{stop}\\}\\) and accordingly either moves or terminates the episode. Generally, the VLN agent is parameterized as a policy network \\(\\pi_{\\theta}(a_{t} \\mid W, O_{\\leq t})\\) with parameters \\(\\theta\\) learned via general pretraining and a combination of imitation learning and reinforcement learning [8, 9]." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.479, + 0.796, + 0.495 + ], + "angle": 0, + "content": "3.1. Generating Visual Imaginations" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.502, + 0.907, + 0.669 + ], + "angle": 0, + "content": "To generate visual imaginations of potential landmarks, we consider the structure of navigation instructions. Typically, these consist of a sequence of sub-instructions describing intermediate navigation steps, which may or may not refer to visual landmarks. For instance, a sub-instruction \"go past the couch\" implies a visual landmark while \"go straight then left\" does not. We use FG-R2R [15] to segment instructions, resulting in an average of 3.66 sub-instructions per instruction for the R2R[4] training set. For a given instruction \\( W \\) decomposed into \\( m \\) segments, we denote these as sub-instructions \\( S = (S_0, \\dots, S_m) \\)." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.672, + 0.909, + 0.838 + ], + "angle": 0, + "content": "Sub-instruction filtering scheme. As sub-instructions may not always contain references to visual landmarks, we filter sub-instructions prior to image generation. First, we ignore any sub-instruction that lacks a noun phrase using Spacy [2] and then noun phrases are further filtered by a manual blacklist to avoid uninformative landmark generation. For example, we exclude detected noun phrases rooted on non-visual terms like counts ('one'), directions ('left'), or ambiguous object pronouns (\"it\") as these lack specificity to produce meaningful visual landmarks. After filtering, we denote the remaining sub-instructions \\( S' \\subset S \\)." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.84, + 0.909, + 0.903 + ], + "angle": 0, + "content": "Visual imagination generation. The filtered subinstructions are passed through a text conditioned diffusion model as illustrated in Fig. 3. We use SDXL [32] as the diffusion model with positive and negative prompts cho" + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.516, + 0.956 + ], + "angle": 0, + "content": "3848" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.092, + 0.092, + 0.907, + 0.201 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.089, + 0.211, + 0.907, + 0.298 + ], + "angle": 0, + "content": "Figure 2. An overview of our approach. (Left) Imaginations generated using valid sub-instructions from an instruction as determined by our filtering scheme are first passed to a pre-trained ViT to obtain feature vectors. A type embedding \\( t_{Im} \\) for imagination modality is then added to the features which are encoded using a 3 layer MLP to obtain imagination embeddings \\( h_i \\). (Right) To integrate imagination modality to a VLN agent, the imagination embeddings \\( h_i \\) are concatenated with instruction embeddings \\( t_i \\) that are encoded using a text encoder \\( f_T(W) \\). The concatenated imagination-text embeddings are passed to the VLN agent's cross-modal encoder \\( f_X \\) along with visual embeddings to predict a distribution over the agent's action space." + }, + { + "type": "text", + "bbox": [ + 0.098, + 0.308, + 0.479, + 0.321 + ], + "angle": 0, + "content": "Go down the stairs and then turn right. Go stand behind the easel by a unicycle." + }, + { + "type": "image", + "bbox": [ + 0.097, + 0.325, + 0.48, + 0.429 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.09, + 0.439, + 0.485, + 0.51 + ], + "angle": 0, + "content": "Figure 3. Example instruction segmentation, filtering, and image generation. Instructions are segmented to sub-instructions leveraging FG-R2R [15] and then filtered to remove phrases referring to uninformative nouns (e.g., \"right\"). We produce visual imaginations using SDXL [32] for remaining sub-instructions." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.537, + 0.484, + 0.704 + ], + "angle": 0, + "content": "sen to steer generation towards indoor real-estate environments akin to typical environments in Matterport3D [7]. For example, providing \"indoor\" and \"real estate\" as positives prompts and \"humans\" and \"collage\" as negatives. See the supplementary for a full list. For a given instruction, we denote the generated images for valid sub-instructions as \\(\\mathcal{Z} = \\{Z_i \\mid S_i \\in S'\\}\\). We colloquially refer to these generated images as imaginations and generate them for the entire R2R dataset. The resulting R2R-Imagine dataset consists of over 41k synthesized imaginations. This amounts to an average of 2.96 imaginations per instruction." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.713, + 0.383, + 0.729 + ], + "angle": 0, + "content": "3.2. Integration with Existing Models" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.735, + 0.484, + 0.903 + ], + "angle": 0, + "content": "We consider augmentation with visual imagination to be a model agnostic approach and integrate it in two prior models - HAMT [8] and DUET [9]. Our integration approach consists of an imagination encoder, an alignment-promoting auxiliary loss, and a finetuning regimen (Fig. 2). Without loss of generality, we consider a decomposition of any VLN agent into an instruction encoder \\( f_{T}(W) \\), an observation/history encoder \\( f_{O}(O_{\\leq t}) \\), and a cross-modal policy network \\( f_{X}(f_{T}(W), f_{O}(O_{\\leq t})) \\) that combines the encoder outputs to predict a distribution over actions. Note that these may be fairly complex mechanisms and include explicit his" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.308, + 0.907, + 0.339 + ], + "angle": 0, + "content": "tory encoding modules separate from encoding the current observation; however, this abstraction is sufficient here." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.341, + 0.907, + 0.432 + ], + "angle": 0, + "content": "Imagination encoder and integration. Given an instruction \\( W \\) with corresponding imaginations \\( \\mathcal{Z} \\), we independently encode each imagination \\( Z_{i} \\) using a pretrained vision transformer (ViT) [10] to produce a \\( d \\)-dim embedding vector. These are further transformed by a three-layer MLP such that the \\( i \\)th imagination embedding can be written as" + }, + { + "type": "equation", + "bbox": [ + 0.611, + 0.448, + 0.907, + 0.466 + ], + "angle": 0, + "content": "\\[\nh _ {i} = \\operatorname {M L P} \\left(\\operatorname {V i T} \\left(Z _ {i}\\right) + t _ {I m}\\right) \\tag {1}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.473, + 0.906, + 0.623 + ], + "angle": 0, + "content": "where \\( t_{Im} \\) is the type embedding for imagination modality. As we consider visual imaginations to be proxies for sub-instructions, the full set of imagination embeddings \\( \\mathcal{H} \\) are concatenated to the text encodings such that our general VLN agent structure now computes \\( f_{X}([f_{T}(W),\\mathcal{H}],f_{O}(O_{\\leq t})) \\) where \\( [\\cdot ,\\cdot ] \\) denotes a concatenation. Taking the DUET model as an example, the imagination embeddings are concatenated to the instruction encoding prior to being passed to the coarse-scale and fine-scale cross-modal encoders." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.626, + 0.907, + 0.764 + ], + "angle": 0, + "content": "Auxiliary alignment loss. We add an auxiliary loss to align the representations of imaginations and their corresponding sub-instruction's noun phrases. For a given sub-instruction \\( S_{i} \\), we denote the mean embedding vector for noun phrase tokens in \\( S_{i} \\) from the text encoder \\( f_{T} \\) as \\( \\bar{S}_{i} \\). To align representations, we compute a cosine similarity loss between the corresponding imagination \\( Z_{i} \\) encoded as \\( h_{i} \\) and \\( \\bar{S}_{i} \\). This is averaged over all \\( N_{Im} \\) imagination-subinstruction pairs \\( (Z_{i}, S_{i}) \\) in a batch," + }, + { + "type": "equation", + "bbox": [ + 0.566, + 0.775, + 0.907, + 0.816 + ], + "angle": 0, + "content": "\\[\n\\mathcal {L} _ {c o s} = \\frac {1}{N _ {I m}} \\sum_ {\\left(Z _ {i}, S _ {i}\\right) \\in B} \\left(1 - \\frac {h _ {i} \\cdot \\bar {S} _ {i}}{\\| h _ {i} \\| \\| \\bar {S} _ {i} \\|}\\right). \\tag {2}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.823, + 0.906, + 0.869 + ], + "angle": 0, + "content": "During finetuning, we combine this auxiliary loss \\(\\mathcal{L}_{\\mathrm{cos}}\\) with the losses from the base agent with a scaling factor \\(\\lambda\\) such that the overall loss is \\(\\mathcal{L}_{\\mathrm{base}} + \\lambda \\mathcal{L}_{\\mathrm{cos}}\\)." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.871, + 0.907, + 0.903 + ], + "angle": 0, + "content": "Finetuning regimen. We start with fully-trained checkpoints of prior models as our initial weights. Depending" + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.516, + 0.956 + ], + "angle": 0, + "content": "3849" + } + ], + [ + { + "type": "table_caption", + "bbox": [ + 0.09, + 0.089, + 0.907, + 0.16 + ], + "angle": 0, + "content": "Table 1. Comparison of our approach with selected prior work on the R2R dataset. Methods that use additional visual data beyond MP3D [7] are annotated with \\(\\dagger\\) and are not directly comparable with other approaches. Note that BEVBert* additionally uses panoramic depth images. Adding our visual imagination approach to HAMT and DUET base models (gray rows) leads to improved success rate (SR) and success weighted by inverse path length (SPL) metrics on the val-unseen split. For DUET, this effect also persists on the test set. We highlight these improvements over base models using bold numbers." + }, + { + "type": "table", + "bbox": [ + 0.095, + 0.171, + 0.907, + 0.415 + ], + "angle": 0, + "content": "
MethodsValidation SeenValidation UnseenTest Unseen
TLNE↓SR↑SPL↑TLNE↓SR↑SPL↑TLNE↓SR↑SPL↑
PREVALENT [14]10.323.67696510.194.71585310.515.305451
RecBERT [16]11.132.90726812.013.93635712.354.096357
ADAPT [27]10.972.54767212.213.77645812.993.796559
BEVBert* [3]13.562.17817414.552.81756415.873.137362
MARVAL [19]†10.602.99736910.154.06656110.224.186258
ScaleVLN [41]†11.902.16878012.402.34877914.272.738377
HAMT [8]11.152.5175.6172.1811.463.6266.2461.5112.273.936560
HAMT-Imagine (ours)11.302.3677.1673.7211.813.5867.2662.0212.663.896560
DUET [9]12.332.2878.8472.8813.943.3171.5260.4114.733.656959
DUET-Imagine (ours)12.932.1979.973.7514.353.1972.1260.4815.353.527160
" + }, + { + "type": "text", + "bbox": [ + 0.089, + 0.439, + 0.482, + 0.637 + ], + "angle": 0, + "content": "on the base method, these parameters may be the result of multi-stage pretraining and finetuning regimens - e.g., leveraging synthetic PREVALENT [14] data as prescribed in [8]. The imagination encoder is integrated to the base model and finetuned with the R2R-Imagine dataset. To mitigate catastrophic forgetting, we perform our training in three stages. First, we train only the newly introduced imagination encoder MLP along with type embeddings \\( t^{Im} \\) with the remaining parameters frozen. Then, all the modules are trained jointly with a decreased learning rate for the base model. Finally, all the parameters are trained at a common learning rate. See supplementary materials for details on learning rates and schedules." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.649, + 0.223, + 0.667 + ], + "angle": 0, + "content": "4. Experiments" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.674, + 0.483, + 0.766 + ], + "angle": 0, + "content": "Using the R2R dataset, we evaluate the impact of visual imaginations on our agent's navigation ability, perform ablation studies to explore the design space, and illustrate qualitative impacts of visual imaginations. Using the REVERIE dataset, we study the generalizability of our technique to a fine-grained instruction setting." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.773, + 0.279, + 0.79 + ], + "angle": 0, + "content": "4.1. Experimental setup" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.795, + 0.483, + 0.902 + ], + "angle": 0, + "content": "R2R dataset. R2R [4] is constructed using Matterport3D [7] simulator. It consists of 90 indoor environments with 10,567 panoramas (each panorama consists of 36 single-view images) which act as nodes in a navigation graph. There are 7,189 trajectories each with 3 instructions and a corresponding ground truth trajectory. The dataset is split into train (4675 trajectories), val-seen (340 trajectories)," + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.439, + 0.905, + 0.484 + ], + "angle": 0, + "content": "val-unseen (783 trajectories), and test (1391 trajectories). The train and val-seen split share environments while val-unseen and test have different environments." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.487, + 0.906, + 0.594 + ], + "angle": 0, + "content": "R2R-Imagine dataset. The curation process of R2R-Imagine is prescribed in section 3.1. The number of imagination images are 41,558 for train, 3,055 for val-seen, 6,857 for val-unseen, and 12,412 for test. Our imaginations have a resolution of \\(1024 \\times 1024\\) and a single imagination generation on a single H100 GPU takes 3.2 seconds on average. We intend to release the dataset upon acceptance." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.596, + 0.907, + 0.732 + ], + "angle": 0, + "content": "REVERIE dataset. REVERIE [33] is constructed using Matterport3D and consists of high-level goals instead of finer grained step-by-step instructions. These instructions generally convey the target object and its location. The agent has to rely more on exploration relative to R2R. To be successful, agents must reach and identify the referent object. To synthesize imaginations for REVERIE, we treat the entire instruction as a single sub-instruction and follow the same recipe used for R2R." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.735, + 0.909, + 0.903 + ], + "angle": 0, + "content": "Evaluation metrics. We employ standard VLN metrics [18]. Success rate (SR) is the ratio of instruction-trajectories where the agent stops within a 3 meter radius from the goal and success normalized by inverse path length (SPL) helps distinguish successful episodes with shorter path lengths from longer path lengths. Navigation error (NE) is the distance between an agent's final position and the goal in meters. Trajectory length (TL) is the length of an agent's traversed path in meters. REVERIE [33] additionally defines grounding metrics Remote Grounding Success (RGS) and RGS penalized by path length (RGSPL)." + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.516, + 0.956 + ], + "angle": 0, + "content": "3850" + } + ], + [ + { + "type": "table_caption", + "bbox": [ + 0.091, + 0.089, + 0.486, + 0.133 + ], + "angle": 0, + "content": "Table 2. Performance comparison of our approach on REVERIE with baseline. We observe improvements of our method across all metrics on DUET when provided coarse-grained instructions." + }, + { + "type": "table", + "bbox": [ + 0.095, + 0.143, + 0.482, + 0.205 + ], + "angle": 0, + "content": "
SR↑SPL↑RGS↑RGSPL↑
DUET46.9833.7332.1523.03
DUET-Imagine (ours)48.2833.7632.9723.25
" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.228, + 0.485, + 0.441 + ], + "angle": 0, + "content": "HAMT and DUET base models. We apply our approach to two baseline representative architectures - HAMT [8] and DUET [9]. HAMT is hierarchical and transformer-based; instructions, observations, and trajectory history are encoded separately using unimodal transformers and then combined in a single cross-modal transformer to predict actions. DUET incrementally forms a topographic map of the environment over time and uses dual fine- and coarse-grained cross-modal attention mechanisms to associate language embeddings with the local observation and global topographic features. The results are then fused to predict actions. For both base models, we inject visual imaginations via concatenation with the text embeddings (Sec. 3.2). We use provided checkpoints for pretrained models." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.442, + 0.484, + 0.581 + ], + "angle": 0, + "content": "Implementation details. We follow base agent [8, 9] settings unless specified otherwise. We use a standard off-the-shelf pretrained ViT-B/16 [10] to encode imaginations for both HAMT and DUET. Our VLN-Imagine agents are finetuned for \\(100\\mathrm{k}\\) iterations in three stages as noted in section 3.2. The finetuning process is carried out using a Tesla V100 GPU with a batch size of 8 and takes \\(\\sim 1.5\\) days for each agent. For combining the proposed auxiliary loss function with the base model losses, we empirically set \\(\\lambda = 0.5\\)." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.587, + 0.185, + 0.603 + ], + "angle": 0, + "content": "4.2. Results" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.609, + 0.484, + 0.671 + ], + "angle": 0, + "content": "Our experiments demonstrate that providing generated images of instruction landmarks (imaginations) improves the performance of vision-and-language navigation models (Tab. 1). We highlight our key findings below." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.672, + 0.483, + 0.853 + ], + "angle": 0, + "content": "VLN agents perform better with imaginations. VLN agents equipped with imaginations perform better than VLN agents without by 0.6-1.0 SR and up to 0.5 SPL on R2R val-unseen (Tab. 1). This improvement coincides with lower navigation error (NE) across both models and all dataset splits. We contextualize these results relative to agent architectures that are representative of the VLN literature. We additionally report a distinct family of techniques that scale training environments to boost performance by reducing overfitting - MARVAL [19] and ScaleVLN [41]. We consider approaches of this class orthogonal to our study and instead use visual data only from Matterport3D [7]." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.856, + 0.484, + 0.903 + ], + "angle": 0, + "content": "Imaginations are useful in coarse-grained instruction navigation. Our method is performant when natural language instructions convey high-level objectives (REVERIE," + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.089, + 0.907, + 0.175 + ], + "angle": 0, + "content": "Table 3. Role of imaginations; HAMT-Imagine. Null imaginations refers to testing with no imaginations, wrong imaginations refers to testing with imaginations sampled from different instructions. We observe aligned imaginations are important to navigation performance. In addition, results with no imaginations are better than baseline implying a regularization effect during training." + }, + { + "type": "table", + "bbox": [ + 0.526, + 0.184, + 0.897, + 0.278 + ], + "angle": 0, + "content": "
SR↑SPL↑
Baseline (HAMT)66.2461.51
Null imaginations66.9261.89
Wrong imaginations66.2461.02
Correct imaginations (ours)67.2662.02
" + }, + { + "type": "table_caption", + "bbox": [ + 0.513, + 0.29, + 0.907, + 0.377 + ], + "angle": 0, + "content": "Table 4. Sequential vs. goal imagination; HAMT-Imagine. We study the effect of providing just the imagination belonging to the final valid sub-instruction relative to full valid imaginations. We observe providing just the goal imagination improves performance significantly compared to baseline but shows lower performance relative to providing full imaginations." + }, + { + "type": "table", + "bbox": [ + 0.526, + 0.385, + 0.897, + 0.464 + ], + "angle": 0, + "content": "
SR↑SPL↑
Baseline (HAMT)66.2461.51
Goal imagination66.7961.58
Full imaginations (ours)67.2662.02
" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.488, + 0.907, + 0.565 + ], + "angle": 0, + "content": "Tab. 2); navigation performance improves by 1.3 SR and grounding performance improves by 0.82 RGS. We hypothesize that imaginations not only aid in reaching the goal but also in identifying and grounding the target concept amongst others objects in the scene." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.567, + 0.909, + 0.763 + ], + "angle": 0, + "content": "VLN agents benefit from both imagination training and inference. When nullifying imaginations at test-time, VLN-Imagine agents still surpass baseline performance, but to a lesser extent (Tab. 3). We implement this nullification by setting the imagination attention masks to zero in the cross-modal encoder. We hypothesize that imagination-based training provides a regularizing effect on the overall VLN agent during fine-tuning, revealing a research avenue for strong and performant test-time model inference. We further find that imaginations must be aligned with the instruction to provide maximal benefit; when fed with imaginations randomly-sampled from the R2R-Imagine dataset, VLN-Imagine agents perform worse than baseline (Tab. 3)." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.765, + 0.909, + 0.903 + ], + "angle": 0, + "content": "Sequential imagination outperforms goal imagination. Sequential refers to imaginations generated over multiple sub-instructions (our method). Goal imagination refers to imagining just the final non-filtered sub-instruction. We show these results in Tab. 4 – while goal imagination outperforms the baseline by 0.5 SR, sequential imagination outperforms the baseline by 1.0 SR. Goal imagination may be particularly useful in guiding the agent to stop at the destination. While distinct from our setting, such goal imagi" + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.515, + 0.958 + ], + "angle": 0, + "content": "3851" + } + ], + [ + { + "type": "table_caption", + "bbox": [ + 0.091, + 0.089, + 0.486, + 0.148 + ], + "angle": 0, + "content": "Table 5. Comparison between ViT models. We train both HAMT-Imagine and DUET-Imagine agents with a ViT finetuned in-the-loop with HAMT and an off-the-shelf pretrained ViT. We observe no consistent effects across the both agents." + }, + { + "type": "table", + "bbox": [ + 0.097, + 0.157, + 0.48, + 0.235 + ], + "angle": 0, + "content": "
ViT fine-tunedViT off-the-shelf
SR↑SPL↑SR↑SPL↑
HAMT-Imagine67.3161.567.2662.02
DUET-Imagine72.5860.2972.1260.48
" + }, + { + "type": "table_caption", + "bbox": [ + 0.09, + 0.25, + 0.485, + 0.321 + ], + "angle": 0, + "content": "Table 6. Loss ablation. We study the effect of negative samples in our alignment loss by comparing our approach of using cosine loss with InfoNCE [28] loss. In addition, we train the agent with no auxiliary loss represented by \"No loss\". We do not observe a significant difference between contrastive and cosine losses." + }, + { + "type": "table", + "bbox": [ + 0.103, + 0.331, + 0.475, + 0.41 + ], + "angle": 0, + "content": "
SR↑SPL↑
No loss66.7561.60
LInfoNCE67.1861.84
Lcos(ours)67.2662.02
" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.436, + 0.483, + 0.468 + ], + "angle": 0, + "content": "nation has parallels to ImageNav [22], a visual navigation task that conveys goals solely using exemplar images." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.47, + 0.483, + 0.592 + ], + "angle": 0, + "content": "General-purpose vision encoders are sufficient imagination encoders. We compare encoding imaginations with off-the-shelf ViT [10] vs. ViT fine-tuned on R2R navigation episodes [8]. We find comparable performance (Tab. 5) and choose to employ off-the-shelf ViT in our final model. This is motivated to maximize generality and to avoid potential issues relating to catastrophic forgetting. For example, generalization of frozen ViT is demonstrated in Zeng et al. [44]." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.595, + 0.483, + 0.779 + ], + "angle": 0, + "content": "Aligning imaginations to instructions benefits VLN-imagine agents. Aligning the imagination embedding to the instruction embedding as an auxiliary loss leads to a 0.5 SR and a 0.4 SPL increase over no auxiliary loss (Tab. 6). We also consider formulating this loss as contrastive - do negative samples enable a stronger alignment and thus better downstream performance? We find that an InfoNCE loss [28] yields no significant difference relative to our cosine similarity loss (Eq. 2). We construct negative samples using mean noun phrase embeddings from imaginations of other instructions in the batch. The overall loss is \\(\\mathcal{L}_{\\mathrm{base}} + \\lambda \\mathcal{L}_{\\mathrm{InfoNCE}}\\) where we set \\(\\lambda\\) to 0.2 empirically." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.781, + 0.483, + 0.903 + ], + "angle": 0, + "content": "Imaginations correctly represent noun phrases in a subinstruction. To evaluate fidelity of our imaginations, we run an open-vocabulary object detector (LangSAM [1]) on our imaginations to test if the noun phrases contained in the corresponding sub-instruction from R2R are detected successfully. We see in Tab. 7 that at least one noun phrase is detected in our imaginations for \\(98.78\\%\\) of sub-instructions and all noun phrases are detected for \\(94.99\\%\\)" + }, + { + "type": "table_caption", + "bbox": [ + 0.513, + 0.089, + 0.907, + 0.161 + ], + "angle": 0, + "content": "Table 7. Open vocabulary object detection accuracy (\\%) of imaginations to noun phrases from sub-instructions. Our imaginations represent the noun phrases with a high accuracy. In val-seen, in \\(98.43\\%\\) of sub-instructions, at least 1 noun phrase is detected and in \\(94.51\\%\\) of sub-instructions, all noun phrases are detected." + }, + { + "type": "table", + "bbox": [ + 0.515, + 0.17, + 0.907, + 0.27 + ], + "angle": 0, + "content": "
Sub-instructionsNoun Phrase Detection (%)
≥1 detectedAll detected
Val-seen298598.4394.51
Val-unseen670098.9995.34
Test1207598.9395.12
" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.295, + 0.907, + 0.327 + ], + "angle": 0, + "content": "of sub-instructions. We conclude our imaginations provide a high coverage of relevant concepts across splits." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.335, + 0.746, + 0.351 + ], + "angle": 0, + "content": "4.3. Qualitative visualizations" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.357, + 0.906, + 0.417 + ], + "angle": 0, + "content": "In this section, we provide a qualitative investigation of how imaginations may be used as a pivot between language tokens and visual observations from the environment by examining the attention patterns in HAMT-Imagine." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.418, + 0.906, + 0.583 + ], + "angle": 0, + "content": "For a given instruction, we randomly select an imagination / sub-instruction pair and examine the agent's trajectory to identify the first occurrence of the reference object. We then examine the attention distributions between language tokens, the visual imagination, and the panoramic observation. As there are many attention heads across multiple layers in the cross-modal encoder, we structure our study by focusing on those attention heads where sub-instruction tokens are attended to most strongly by their corresponding visual imagination and then examine how the visual imagination attends to observations." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.584, + 0.906, + 0.719 + ], + "angle": 0, + "content": "We then visualize sub-instruction and imagination pairs along with their top attended language tokens and visual observations in Fig. 4. By construction, these are time steps where the referent is visible and the visual imagination is attending to the appropriate noun phrase. As such, our examples primarily demonstrate how the imagination learns to associate with the visual observations. We focus on this aspect as our auxiliary loss already explicitly encourages alignment between visual imaginations and sub-instruction." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.72, + 0.909, + 0.903 + ], + "angle": 0, + "content": "From the examples in Fig. 4, we make two remarks. Imaginations attend most strongly to observations of matching semantic concepts. In row 1, the imagination and the retrieved observations all depict a cheetah (either a painting or floor mat). Imagination attention scores mirror the proper disambiguation of concepts. In row 2, the sub-instruction (and thus the imagination) refer to a unicycle. While both a bicycle and unicycle are observed during execution, attention scores are maximized by the observations containing the unicycle. These patterns do not directly imply a causal relation between the attention scores and eventual decision making of the VLN agent. Rather, they illustrate how imag-" + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.516, + 0.957 + ], + "angle": 0, + "content": "3852" + } + ], + [ + { + "type": "table", + "bbox": [ + 0.092, + 0.089, + 0.907, + 0.416 + ], + "angle": 0, + "content": "
Sub-instructionImaginationTop Attended TokensTop Attended Observations
And walk towards the door pass by the cheetah paintingche, ##eta, ##h, painting
Go stand behind the easel by a unicycle next to a doorun, ##ic. ##y, ##cle, to
Once in the room, walk to the left of the pool tablepool, door, room
Walk into the arch hallway near the Native American painting on the wallAmerican, Native, candles
" + }, + { + "type": "image_caption", + "bbox": [ + 0.09, + 0.428, + 0.907, + 0.54 + ], + "angle": 0, + "content": "Figure 4. Qualitative examples showing imaginations as pivots between language and observation images. The first column contains sub-instruction from a random instruction from R2R, the second column contains the imagination generated using the sub-instruction. The third and fourth columns show highest attended language tokens and observation images from an attention head in HAMT's cross-modal transformer at a time step the associated observations are first visible. In the second example (row 2), the sub-instruction references \"unicycle\" which is captured in the imagination along with neighboring nouns \"easel\" and \"door\". We observe that in a head where top attending language tokens to the imagination query are references to nouns associated with the sub-instruction, its top attended observations to the imagination query are images of the same concept (\"unicycle\"). In this example, the imagination of a unicycle is being used to associate language tokens belonging to \"unicycle\" to observations of unicycle hinting at the utility of imaginations in navigation." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.566, + 0.483, + 0.597 + ], + "angle": 0, + "content": "inations can serve as a pivot between language and observations and afford disambiguation of related concepts." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.614, + 0.21, + 0.629 + ], + "angle": 0, + "content": "5. Conclusion" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.641, + 0.483, + 0.806 + ], + "angle": 0, + "content": "In this work, we show that visual representations of instructions, or visual imaginations, play an additive role in training superior navigating agents for vision-and-language navigation (VLN) tasks. We generate a sequence of visual imaginations depicting sub-instruction landmarks using a text-to-image diffusion model. Our approach is model-agnostic and we integrate these imaginations into VLN agents using an imagination encoder, an auxiliary loss, and a fine-tuning training step. We apply this approach to two representative VLN agents across two datasets and observe performance improvements of around 1 SR and 0.5 SPL." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.811, + 0.484, + 0.902 + ], + "angle": 0, + "content": "Limitations. Generating and encoding imaginations increases the computational cost of running VLN agents. This is particularly relevant when running agents on-device for real-world robotic deployments. However, we note that this process is performed upfront, unlike techniques that depend on step-wise observations like novel view synthesis (NVS)." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.566, + 0.907, + 0.656 + ], + "angle": 0, + "content": "Stepping more broadly, by design, our imaginations are not grounded in the environment. Personalized reasoning, such as the unique naming of objects and locations, cannot be directly imagined in our framework. Life-long learning of persistent groundings is an open question not just in generating imaginations, but in the VLN space at large." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.673, + 0.909, + 0.84 + ], + "angle": 0, + "content": "Future directions. A natural question that arises from our research is the exact complementary nature of the roles played by visual imaginations and language instructions. This is a rich area for future work. Similar studies relating language and visual observations have been conducted in [48] and [47]. Our paper demonstrated the fundamental utility of imagination in VLN agents. Beyond this lies interesting extensions, such as how imagination can help bridge the simulation-to-reality (Sim2Real) gap for VLN agents and whether imaginations can unlock the performance of VLN world models through image-image reasoning." + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.856, + 0.909, + 0.903 + ], + "angle": 0, + "content": "Acknowledgments. We thank Zijiao Yang and Robert Yelle for their assistance with simulator issues and technical support for high performance computing infrastructure." + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.516, + 0.957 + ], + "angle": 0, + "content": "3853" + } + ], + [ + { + "type": "title", + "bbox": [ + 0.093, + 0.091, + 0.188, + 0.106 + ], + "angle": 0, + "content": "References" + }, + { + "type": "ref_text", + "bbox": [ + 0.101, + 0.116, + 0.482, + 0.143 + ], + "angle": 0, + "content": "[1] Language segment-anything. https://github.com/luca-medeiros/lang-segment-anything/.7" + }, + { + "type": "ref_text", + "bbox": [ + 0.101, + 0.146, + 0.483, + 0.173 + ], + "angle": 0, + "content": "[2] spacy: Industrial-strength natural language processing in python. https://spacy.io/.3" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.176, + 0.482, + 0.243 + ], + "angle": 0, + "content": "[3] Dong An, Yuankai Qi, Yangguang Li, Yan Huang, Liang Wang, Tieniu Tan, and Jing Shao. Bevbert: Multimodal map pre-training for language-guided navigation. Proceedings of the IEEE/CVF International Conference on Computer Vision, 2023. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.247, + 0.482, + 0.342 + ], + "angle": 0, + "content": "[4] Peter Anderson, Qi Wu, Damien Teney, Jake Bruce, Mark Johnson, Niko Sünderhauf, Ian D. Reid, Stephen Gould, and Anton van den Hengel. Vision-and-language navigation: Interpreting visually-grounded navigation instructions in real environments. 2018 IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 3674-3683, 2017. 2, 3, 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.346, + 0.482, + 0.427 + ], + "angle": 0, + "content": "[5] Homanga Bharadhwaj, Debidatta Dwibedi, Abhinav Gupta, Shubham Tulsiani, Carl Doersch, Ted Xiao, Dhruv Shah, Fei Xia, Dorsa Sadigh, and Sean Kirmani. Gen2act: Human video generation in novel scenarios enables generalizable robot manipulation. arXiv preprint arXiv:2409.16283, 2024. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.431, + 0.482, + 0.5 + ], + "angle": 0, + "content": "[6] Kevin Black, Mitsuhiko Nakamoto, Pranav Atreya, Homer Rich Walke, Chelsea Finn, Aviral Kumar, and Sergey Levine. Zero-shot robotic manipulation with pre-trained image-editing diffusion models. In The Twelfth International Conference on Learning Representations. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.503, + 0.482, + 0.572 + ], + "angle": 0, + "content": "[7] Angel Chang, Angela Dai, Thomas Funkhouser, Maciej Halber, Matthias Niebner, Manolis Savva, Shuran Song, Andy Zeng, and Yinda Zhang. Matterport3d: Learning from rgb-d data in indoor environments. In 2017 International Conference on 3D Vision (3DV), pages 667-676, 2017. 2, 4, 5, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.575, + 0.482, + 0.628 + ], + "angle": 0, + "content": "[8] Shizhe Chen, Pierre-Louis Guhur, Cordelia Schmid, and Ivan Laptev. History aware multimodal transformer for vision-and-language navigation. In NeurIPS, 2021. 2, 3, 4, 5, 6, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.632, + 0.482, + 0.687 + ], + "angle": 0, + "content": "[9] Shizhe Chen, Pierre-Louis Guhur, Makarand Tapaswi, Cordelia Schmid, and Ivan Laptev. Think global, act local: Dual-scale graph transformer for vision-and-language navigation. In CVPR, 2022. 2, 3, 4, 5, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.689, + 0.482, + 0.771 + ], + "angle": 0, + "content": "[10] Alexey Dosovitskiy, Lucas Beyer, Alexander Kolesnikov, Dirk Weissenborn, Xiaohua Zhai, Thomas Unterthiner, Mostafa Dehghani, Matthias Minderer, Georg Heigold, Sylvain Gelly, et al. An image is worth 16x16 words: Transformers for image recognition at scale. arXiv preprint arXiv:2010.11929, 2020. 4, 6, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.775, + 0.482, + 0.843 + ], + "angle": 0, + "content": "[11] Yilun Du, Sherry Yang, Bo Dai, Hanjun Dai, Ofir Nachum, Josh Tenenbaum, Dale Schuurmans, and Pieter Abbeel. Learning universal policies via text-guided video generation. Advances in Neural Information Processing Systems, 36, 2024. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.846, + 0.482, + 0.901 + ], + "angle": 0, + "content": "[12] Daniel Fried, Ronghang Hu, Volkan Cirik, Anna Rohrbach, Jacob Andreas, Louis-Philippe Morency, Taylor Berg-Kirkpatrick, Kate Saenko, Dan Klein, and Trevor Darrell. Speaker-follower models for vision-and-language naviga" + }, + { + "type": "list", + "bbox": [ + 0.094, + 0.116, + 0.483, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.549, + 0.094, + 0.905, + 0.119 + ], + "angle": 0, + "content": "tion. In Neural Information Processing Systems (NeurIPS), 2018. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.122, + 0.905, + 0.176 + ], + "angle": 0, + "content": "[13] Jialu Gao, Kaizhe Hu, Guowei Xu, and Huazhe Xu. Can pre-trained text-to-image models generate visual goals for reinforcement learning? Advances in Neural Information Processing Systems, 36:38297-38310, 2023. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.178, + 0.905, + 0.218 + ], + "angle": 0, + "content": "[14] Weituo Hao, Chunyuan Li, Xiujun Li, Lawrence Carin, and Jianfeng Gao. Towards learning a generic agent for vision-and-language navigation via pre-training, 2020. 2, 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.22, + 0.905, + 0.286 + ], + "angle": 0, + "content": "[15] Yicong Hong, Cristian Rodriguez, Qi Wu, and Stephen Gould. Sub-instruction aware vision-and-language navigation. In Proceedings of the 2020 Conference on Empirical Methods in Natural Language Processing (EMNLP), pages 3360-3376, 2020. 3, 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.289, + 0.905, + 0.357 + ], + "angle": 0, + "content": "[16] Yicong Hong, Qi Wu, Yuankai Qi, Cristian Rodriguez-Opazo, and Stephen Gould. A recurrent vision-and-language bert for navigation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 1643–1653, 2021. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.359, + 0.905, + 0.426 + ], + "angle": 0, + "content": "[17] Yicong Hong, Qi Wu, Yuankai Qi, Cristian Rodriguez-Opazo, and Stephen Gould. A recurrent vision-and-language bert for navigation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 1643–1653, 2021. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.428, + 0.905, + 0.482 + ], + "angle": 0, + "content": "[18] Gabriel Ilharco, Vihan Jain, Alexander Ku, Eugene Ie, and Jason Baldridge. General evaluation for instruction conditioned navigation using dynamic time warping. ArXiv, abs/1907.05446, 2019. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.484, + 0.905, + 0.579 + ], + "angle": 0, + "content": "[19] Aishwarya Kamath, Peter Anderson, Su Wang, Jing Yu Koh, Alexander Ku, Austin Waters, Yinfei Yang, Jason Baldridge, and Zarana Parekh. A new path: Scaling vision-and-language navigation with synthetic instructions and imitation learning. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 10813-10823, 2023. 2, 5, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.581, + 0.905, + 0.62 + ], + "angle": 0, + "content": "[20] Ivan Kapelyukh, Vitalis Vosylius, and Edward Johns. Dallebot: Introducing web-scale diffusion models to robotics. 2023. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.623, + 0.905, + 0.663 + ], + "angle": 0, + "content": "[21] Jing Yu Koh, Honglak Lee, Yinfei Yang, Jason Baldridge, and Peter Anderson. Pathdreamer: A world model for indoor navigation, 2021. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.665, + 0.905, + 0.718 + ], + "angle": 0, + "content": "[22] Jacob Krantz, Stefan Lee, Jitendra Malik, Dhruv Batra, and Devendra Singh Chaplot. Instance-specific image goal navigation: Training embodied agents to find object instances, 2022. 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.721, + 0.905, + 0.774 + ], + "angle": 0, + "content": "[23] Chongxi Lai, Shinsuke Tanaka, Timothy D. Harris, and Albert K. Lee. Volitional activation of remote place representations with a hippocampal brain-machine interface. Science, 382(6670):566-573, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.776, + 0.905, + 0.816 + ], + "angle": 0, + "content": "[24] Jialu Li and Mohit Bansal. Panogen: Text-conditioned panoramic environment generation for vision-and-language navigation. arxiv, 2023. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.818, + 0.905, + 0.858 + ], + "angle": 0, + "content": "[25] Jialu Li and Mohit Bansal. Improving vision-and-language navigation by generating future-view image semantics. CVPR, 2023. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.86, + 0.905, + 0.901 + ], + "angle": 0, + "content": "[26] Mingxiao Li, Zehao Wang, Tinne Tuytelaars, and Marie-Francine Moens. Layout-aware dreamer for embodied visual referring expression grounding. In Proceedings of the" + }, + { + "type": "list", + "bbox": [ + 0.518, + 0.094, + 0.905, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.516, + 0.956 + ], + "angle": 0, + "content": "3854" + } + ], + [ + { + "type": "ref_text", + "bbox": [ + 0.126, + 0.092, + 0.483, + 0.119 + ], + "angle": 0, + "content": "AAAI Conference on Artificial Intelligence, pages 1386-1395, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.121, + 0.483, + 0.189 + ], + "angle": 0, + "content": "[27] Bingqian Lin, Yi Zhu, Zicong Chen, Xiwen Liang, Jianzhuang Liu, and Xiaodan Liang. Adapt: Vision-language navigation with modality-aligned action prompts. In 2022 IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 15375-15385, 2022. 2, 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.191, + 0.482, + 0.232 + ], + "angle": 0, + "content": "[28] Aaron van den Oord, Yazhe Li, and Oriol Vinyals. Representation learning with contrastive predictive coding. arXiv preprint arXiv:1807.03748, 2018. 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.233, + 0.483, + 0.273 + ], + "angle": 0, + "content": "[29] Allan Paivio. Abstractness, imagery, and meaningfulness in paired-associate learning. Journal of verbal learning and verbal behavior., 4(1), 1965-2. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.274, + 0.482, + 0.315 + ], + "angle": 0, + "content": "[30] Allan Paivio. A factor-analytic study of word attributes and verbal learning. Journal of verbal learning and verbal behavior, 7(1), 1968-2." + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.316, + 0.482, + 0.371 + ], + "angle": 0, + "content": "[31] A. Paivio, P. C. Smythe, and J. C. Yuille. Imagery versus meaningfulness of nouns in paired-associate learning. Canadian Journal of Psychology / Revue Canadienne De Psychologie, 22:427–441, 1968. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.372, + 0.482, + 0.427 + ], + "angle": 0, + "content": "[32] Dustin Podell, Zion English, Kyle Lacey, Andreas Blattmann, Tim Dockhorn, Jonas Müller, Joe Penna, and Robin Rombach. Sdxl: Improving latent diffusion models for high-resolution image synthesis, 2023. 2, 3, 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.428, + 0.483, + 0.51 + ], + "angle": 0, + "content": "[33] Yuankai Qi, Qi Wu, Peter Anderson, Xin Wang, William Yang Wang, Chunhua Shen, and Anton van den Hengel. Reverie: Remote embodied visual referring expression in real indoor environments. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 9982-9991, 2020. 2, 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.511, + 0.483, + 0.594 + ], + "angle": 0, + "content": "[34] Alec Radford, Jong Wook Kim, Chris Hallacy, Aditya Ramesh, Gabriel Goh, Sandhini Agarwal, Girish Sastry, Amanda Askell, Pamela Mishkin, Jack Clark, et al. Learning transferable visual models from natural language supervision. In International conference on machine learning, pages 8748-8763. PMLR, 2021. 1, 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.595, + 0.483, + 0.677 + ], + "angle": 0, + "content": "[35] Santhosh K Ramakrishnan, Aaron Gokaslan, Erik Wijmans, Oleksandr Maksymets, Alex Clegg, John Turner, Eric Undersander, Wojciech Galuba, Andrew Westbury, Angel X Chang, et al. Habitat-matterport 3d dataset (hm3d): 1000 large-scale 3d environments for embodied ai. arXiv preprint arXiv:2109.08238, 2021. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.679, + 0.483, + 0.718 + ], + "angle": 0, + "content": "[36] Marianne Cumella Reddan, Tor Dessart Wager, and Daniela Schiller. Attenuating neural threat expression with imagination. *Neuron*, 100(4):994–1005.e4, 2018. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.72, + 0.483, + 0.803 + ], + "angle": 0, + "content": "[37] Manolis Savva, Abhishek Kadian, Oleksandr Maksymets, Yili Zhao, Erik Wijmans, Bhavana Jain, Julian Straub, Jia Liu, Vladlen Koltun, Jitendra Malik, et al. Habitat: A platform for embodied ai research. In Proceedings of the IEEE/CVF international conference on computer vision, pages 9339-9347, 2019. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.804, + 0.483, + 0.858 + ], + "angle": 0, + "content": "[38] Daniel L. Schacter, Donna Rose Addis, and Randy L. Buckner. Remembering the past to imagine the future: the prospective brain. Nature Reviews Neuroscience, 8(9):657-661, 2007. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.86, + 0.483, + 0.899 + ], + "angle": 0, + "content": "[39] Roger N. Shepard and Jacqueline Metzler. Mental rotation of three-dimensional objects. Science, 171(3972):701-703, 1971. 2" + }, + { + "type": "list", + "bbox": [ + 0.093, + 0.092, + 0.483, + 0.899 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.093, + 0.907, + 0.203 + ], + "angle": 0, + "content": "[40] Hao Tan, Licheng Yu, and Mohit Bansal. Learning to navigate unseen environments: Back translation with environmental dropout. In Proceedings of the 2019 Conference of the North American Chapter of the Association for Computational Linguistics: Human Language Technologies, Volume 1 (Long and Short Papers), pages 2610-2621, Minneapolis, Minnesota, 2019. Association for Computational Linguistics. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.205, + 0.906, + 0.259 + ], + "angle": 0, + "content": "[41] Zun Wang, Jialu Li, Yicong Hong, Yi Wang, Qi Wu, Mohit Bansal, Stephen Gould, Hao Tan, and Yu Qiao. Scaling data generation in vision-and-language navigation. In ICCV 2023, 2023. 2, 5, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.261, + 0.905, + 0.328 + ], + "angle": 0, + "content": "[42] Fei Xia, Amir R. Zamir, Zhi-Yang He, Alexander Sax, Jitendra Malik, and Silvio Savarese. Gibson env: real-world perception for embodied agents. In Computer Vision and Pattern Recognition (CVPR), 2018 IEEE Conference on. IEEE, 2018. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.331, + 0.905, + 0.387 + ], + "angle": 0, + "content": "[43] Zijiao Yang, Arjun Majumdar, and Stefan Lee. Behavioral analysis of vision-and-language navigation agents. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 2574-2582, 2023. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.389, + 0.905, + 0.457 + ], + "angle": 0, + "content": "[44] Kuo-Hao Zeng, Zichen Zhang, Kiana Ehsani, Rose Hendrix, Jordi Salvador, Alvaro Herrasti, Ross Girshick, Aniruddha Kembhavi, and Luca Weihs. Poliformer: Scaling on-policy rl with transformers results in masterful navigators. In 8th Annual Conference on Robot Learning, 2024. 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.459, + 0.905, + 0.514 + ], + "angle": 0, + "content": "[45] Gengze Zhou, Yicong Hong, and Qi Wu. Navigpt: Explicit reasoning in vision-and-language navigation with large language models. Proceedings of the AAAI Conference on Artificial Intelligence, 38(7):7641-7649, 2024. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.516, + 0.905, + 0.582 + ], + "angle": 0, + "content": "[46] Gengze Zhou, Yicong Hong, Zun Wang, Xin Eric Wang, and Qi Wu. Navigpt-2: Unleashing navigational reasoning capability for large vision-language models. In European Conference on Computer Vision, pages 260–278. Springer, 2025. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.585, + 0.905, + 0.682 + ], + "angle": 0, + "content": "[47] Wanrong Zhu, Yuankai Qi, Pradyumna Narayana, Kazoo Sone, Sugato Basu, Xin Wang, Qi Wu, Miguel Eckstein, and William Yang Wang. Diagnosing vision-and-language navigation: What really matters. In Proceedings of the 2022 Conference of the North American Chapter of the Association for Computational Linguistics: Human Language Technologies, pages 5981-5993, 2022. 1, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.684, + 0.905, + 0.724 + ], + "angle": 0, + "content": "[48] Wang Zhu, Ishika Singh, Yuan Huang, Robin Jia, and Jesse Thomason. Does vln pretraining work with nonsensical or irrelevant instructions? ArXiv, abs/2311.17280, 2023. 8" + }, + { + "type": "list", + "bbox": [ + 0.517, + 0.093, + 0.907, + 0.724 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.516, + 0.956 + ], + "angle": 0, + "content": "3855" + } + ] +] \ No newline at end of file diff --git a/2025/Do Visual Imaginations Improve Vision-and-Language Navigation Agents_/6684e0af-b484-43ef-8443-50d4bfdefdd1_origin.pdf b/2025/Do Visual Imaginations Improve Vision-and-Language Navigation Agents_/6684e0af-b484-43ef-8443-50d4bfdefdd1_origin.pdf new file mode 100644 index 0000000000000000000000000000000000000000..22f2a9b563f9639d4939dc741c629626f74cd58d --- /dev/null +++ b/2025/Do Visual Imaginations Improve Vision-and-Language Navigation Agents_/6684e0af-b484-43ef-8443-50d4bfdefdd1_origin.pdf @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:684bd4f312db7ffc0ef40ac8aca33c4bdae21b4f3c8ee86f75b871c8064eb7ce +size 7418966 diff --git a/2025/Do Visual Imaginations Improve Vision-and-Language Navigation Agents_/full.md b/2025/Do Visual Imaginations Improve Vision-and-Language Navigation Agents_/full.md new file mode 100644 index 0000000000000000000000000000000000000000..7fe68a350b21ba3554a95ac9c5f9670bf68aa69c --- /dev/null +++ b/2025/Do Visual Imaginations Improve Vision-and-Language Navigation Agents_/full.md @@ -0,0 +1,274 @@ +# Do Visual Imaginations Improve Vision-and-Language Navigation Agents? + +Akhil Perincherry1, Jacob Krantz, Stefan Lee1 +1Oregon State University + +perincha@oregonstate.edu, jkrantz989@gmail.com, leestef@oregonstate.edu + +# Abstract + +Vision-and-Language Navigation (VLN) agents are tasked with navigating an unseen environment using natural language instructions. In this work, we study if visual representations of sub-goals implied by the instructions can serve as navigational cues and lead to increased navigation performance. To synthesize these visual representations or "imaginations", we leverage a text-to-image diffusion model on landmark references contained in segmented instructions. These imaginations are provided to VLN agents as an added modality to act as landmark cues and an auxiliary loss is added to explicitly encourage relating these with their corresponding referring expressions. Our findings reveal an increase in success rate (SR) of $\sim 1$ point and up to $\sim 0.5$ points in success scaled by inverse path length (SPL) across agents. These results suggest that the proposed approach reinforces visual understanding compared to relying on language instructions alone. Code and data for our work can be found at https://www.akhilperincherry.com/VLN-Imagine-website/. + +# 1. Introduction + +In this work, we examine whether providing visual imagery corresponding to described landmarks improves the performance of agents following natural language navigation instructions. Consider the natural language navigation instruction presented in Figure 1 which asks an agent to "Go straight, take a left at the pool table to enter the kitchen. Walk to the bedroom and stop". This instruction provides unconditional action directives like "go straight" but also frequently conditions the given directions on visual landmarks in the scene such as the pool table, kitchen, and bedroom. Standard approaches to vision-and-language navigation tasks rely on learned cross-modal alignment mechanisms to implicitly associate these noun phrases with their visual referents during navigation. However, text-to-image generation models have improved to the point that producing imagery matching the semantics of these visual references prior to navigation is plausible. We show three such + +# Natural Language Navigation Instruction: + +![](images/705f98388715ea92f7dd13c69a293d935e23117c9638211362f8833f3ce60dcd.jpg) + +Go straight, take a left at the pool table to enter the kitchen. Walk to the bedroom and stop. + +# Visual Imaginations: + +![](images/7734d74b0f516f15705c4d935639aa995a10be9f1d968980a406f0abd2a3c2f0.jpg) +Figure 1. Illustration of visual imaginations. (Top) A natural language instruction specifying sub-goals pool table, kitchen, and bedroom. (Bottom) Visual imaginations of landmarks pool table, kitchen and bedroom referenced by the sub-goals in the instruction. In our work, we study if these visual imaginations generated using text-to-image models can improve performance in VLN. + +![](images/924ec74d18c943cbc11f8635fd22e9f6fd6d8af80838749c8db460cc81cdb261.jpg) + +![](images/90b84f5ea45bd4c6ac8966bf11611da60bbbffea94897ac8d7641849870f928d.jpg) + +![](images/388633a477afaf69d30be05e9f4ef9df5c0b90d03ce290669461593b9c8146dd.jpg) + +generations for our running example in Figure 1. Drawing an analogy to the substantial work in cognitive science on the impact of mental imagery on task performance, we refer to these generated images as visual imaginations and study whether providing them in addition to corresponding language-based instructions can improve the performance of vision-and-language navigation agents. + +But why might we believe visual imaginations could be beneficial? In short - text-to-image models have broad knowledge, and learning to perform semantic matching in the image domain may be an easier task than language grounding. Associating noun phrases and their visual referents is a sub-task in vision-and-language navigation (VLN) that most works address through extensive pretraining of model components on web-scale image caption datasets - e.g., by initializing image and language encoders with CLIP [34] models. While this provides a useful warm-start, it is unclear how well the remaining cross-modal components that drive navigation retain these capabilities after downstream VLN training. Downstream datasets may cover only a small portion of relevant objects and visual grounding remains a weakness in many existing VLN models [43, 47]. + +Directly synthesizing visual imaginations using off-the-shelf text-to-image models offers alternative training and inference mechanisms. Synthesized images have known correspondences to noun phrases in instructions that can be leveraged to reinforce visual grounding during downstream task training through auxiliary losses. Further, matching a visual imagination with environment observations reduces to an image-to-image matching task – offering the potential to use visual imaginations as a convenient pivot between language and vision. This may prove especially effective for out-of-distribution landmarks that have limited or no support during downstream task training. For example, novel instructions referencing “butterfly sculptures” or “Pulp Fiction posters” are unlikely to be well-grounded in the training set but produce semantically relevant visual imaginations that could then be matched to observations. + +Though we do not imply any direct connections, our work is also inspired by cognitive science studies which indicate mental imagery can serve as valuable cues in addition to linguistic stimuli. Several works supporting dualcoding theory [29-31] find that concrete imagery tends to be more effective for certain types of concept learning compared to linguistic stimuli alone. Further, findings from human and animal studies [23, 36, 38, 39] suggest that imagining snapshots of the environment or potential trajectories before navigating can improve decision making. + +To investigate our key question, we augment agents with visual imaginations for the popular Room-2-Room (R2R) [4] and REVERIE [33] tasks. We propose a VLN agent agnostic procedure to incorporate visual imaginations as additional inputs and introduce a text-imagination alignment loss to explicitly encourage visual grounding during training. We generate visual imaginations for key noun phrases in instructions that refer to visual landmarks using an off-the-shelf text-to-image generation model [32]. + +Applying our technique to two existing models, HAMT [8] and DUET [9], we find modest improvements - R2R validation unseen success increases by roughly 1.0 and 0.6 points respectively. This translates to an improvement of 2 points in success on the test set for our modified DUET model. On REVERIE validation unseen, we improve DUET performance by 1.3 points success and 0.82 points grounding success. Our results suggest that including explicit visual depictions of referred objects can improve performance on vision-and-language navigation tasks. + +# Contributions. We summarize our contributions as: + +- We develop a pipeline for generating visual imaginations from navigation instructions and synthesize the R2R-Imagine dataset to enable studying the impact of text-to-image models on VLN agents. +- We propose an agent-agnostic method to incorporate visual imaginations into existing VLN agents and show + +improved performance for HAMT [8] and DUET [9] across R2R and REVERIE. + +- We provide ablations of modeling decisions to characterize the design space for adding visual imaginations. + +# 2. Related Work + +Vision-and-Language Navigation (VLN). Navigating agents that can operate in novel environments using freeform natural language lends itself as personal assistants, field-robots and search-and-rescue agents. Development of VLN agents has been facilitated by photorealistic simulators such as Matterport3D [7] and paired language-action datasets such as Room-2-Room (R2R) [4] and REVERIE [33]. R2R provides fine-grained instructions conveying low-level directives such as "Go left, then right at the hall". REVERIE specifies coarse-grained instructions conveying broad high-level goals such as "Adjust the picture by the lamp in the hall". Progress in VLN has broadly ranged from architectural improvements and data scaling techniques. Architectural improvements include evolution from recurrent systems [4] to transformer architectures [8, 9, 17]. HAMT [8] uses a hierarchical transformer architecture to encode spatial and temporal information. DUET [9] maintains a topological map and performs coarse and fine scale encoding over it. Data scaling techniques include augmenting environments [14, 40] using an action-to-text speaker model [12] to generate synthetic instructions, and scaling environments [19, 41] with additional scenes such as HM3D [35] and Gibson [42] in Habitat [37]. Recently, zero-shot or learned integrations of pretrained large language models for VLN [45, 46] is showing promising performance. However, running large models in-the-loop incurs high computational costs. + +Closely related to our work, ADAPT [27] augments VLN agents with visual observations related to an instruction. Language-observation pairs are stored from the training set and multiple pairs of matching object/location nouns from instructions are retrieved at test time using CLIP [34] scores. Fundamentally, the ADAPT image base is limited to training environments whereas our imaginations cover an open distribution. Therefore, our method can be more representative of the instruction and visualize novel objects/locations. For instance, a sub-instruction of "walk into the kitchen with blue walls" would retrieve images of kitchens from the training environment using ADAPT, but our imaginations would convey a kitchen with blue walls. + +Another adjacent work, LAD [26] leverages layout dreamer and goal dreamer modules optimized around a topological graph to navigate via coarse-grained instructions. The capability of our method contrasts with LAD inthree primary ways: 1) method generality; Our method is performant across models and datasets, whereas LAD is a + +full architecture constructed around a particular topological graph in coarse-grained settings, 2) sequential imagination: we imagine a sequence of landmarks along a path to inform sequential decision-making, whereas LAD imagines just the goal, and 3) we perform early fusion of imaginations and language whereas LAD employs late fusion. + +Image generation in VLN. Several techniques have proposed generative models to predict future visual observations or entire environments in VLN [21, 24, 25]. Pathdreamer [21] developed a GAN-based model to predict panoramic observations after a navigation action in VLN, conditioned on prior visual observations. Agents utilizing these predicted observations in a fixed-horizon search with a learned trajectory-instruction alignment heuristic achieve non-trivial success compared to simple baselines, but fall short of planning on actual observations and standard VLN models. Unlike our approach, Pathdreamer does not condition on instruction text or explore training agents on generated imagery. Similarly, PanoGen [24] uses a diffusion model to also generate panorama observations, but conditions on image captions while doing so as a form of visual dataset augmentation – replacing Matterport3D observations with synthetic generations with matching semantics. In contrast to our approach, this data augmentation does not provide additional inputs to the VLN models, is not conditioned on instructions, and is not active at inference time. + +Related to our work is VLN-Sig [25], which adapts VLN agents to predict quantized features of future visual observations given the instruction and observation history. This supports additional pretraining tasks and an auxiliary loss during downstream VLN finetuning for predicting next-step image semantics. Unlike our approach, VLN-Sig relies on the trained VLN agent to predict future image properties and focuses more on next-step prediction. In contrast, we leverage broad vision-and-language knowledge from text-to-image models and focus generating instruction-referred visual landmarks to improve agent performance. + +Image generation in robot control. Taking a wider view, several methods for controlling embodied agents have leveraged text-conditioned generative techniques to provide goal specifications in the form of images or videos. For tabletop object rearrangement tasks, many works use text-to-image models to synthesize goals or sub-goals in response to natural language commands, see [6, 13, 20] for a nonexhaustive sample. Others produce denser conditioning, leveraging text-to-video models to generate entire potential robot trajectories [5, 11]. While similar in motivation to our approach, these techniques focus on tabletop manipulation settings which introduce difficult physical challenges but significantly limit the problem to constrained spaces that remain largely observed throughout operation. As such, sub-goals are often edits to existing observations rather than generations of yet-to-be-seen landmarks as in our approach. + +# 3. Methodology + +Our proposed methodology consists of two components - (1) a visual imagination generation pipeline and (2) a model-agnostic approach to integrating visual imaginations into existing Vision-and-Language Navigation (VLN) agents. Before defining these, we start with establishing relevant notation for the VLN task. + +Problem definition. In the standard VLN task [4], an agent is prompted with a natural language navigation instruction and must make a sequence of actions to navigate along the described trajectory based on visual observations. We denote the $L$ -word instruction as $W = (w_{1}, w_{2}, \dots, w_{L})$ and the visual observation at time step $t$ as a panoramic observation $O_{t}$ consisting of $K$ single view images $O_{t} = (I_{1}^{t}, \dots, I_{K}^{t})$ . Following prior work, we consider $K = 36$ corresponding to images from 12 headings and 3 pitches. Amongst the single view images, a subset $\mathcal{A}_{t} \subset O_{t}$ correspond to navigable directions. Given the instruction $W$ and history of observations $O_{\leq t}$ until time $t$ , the agent selects an action $a_{t} \in \mathcal{A}_{t} \cup \{\text{stop}\}$ and accordingly either moves or terminates the episode. Generally, the VLN agent is parameterized as a policy network $\pi_{\theta}(a_{t} \mid W, O_{\leq t})$ with parameters $\theta$ learned via general pretraining and a combination of imitation learning and reinforcement learning [8, 9]. + +# 3.1. Generating Visual Imaginations + +To generate visual imaginations of potential landmarks, we consider the structure of navigation instructions. Typically, these consist of a sequence of sub-instructions describing intermediate navigation steps, which may or may not refer to visual landmarks. For instance, a sub-instruction "go past the couch" implies a visual landmark while "go straight then left" does not. We use FG-R2R [15] to segment instructions, resulting in an average of 3.66 sub-instructions per instruction for the R2R[4] training set. For a given instruction $W$ decomposed into $m$ segments, we denote these as sub-instructions $S = (S_0, \dots, S_m)$ . + +Sub-instruction filtering scheme. As sub-instructions may not always contain references to visual landmarks, we filter sub-instructions prior to image generation. First, we ignore any sub-instruction that lacks a noun phrase using Spacy [2] and then noun phrases are further filtered by a manual blacklist to avoid uninformative landmark generation. For example, we exclude detected noun phrases rooted on non-visual terms like counts ('one'), directions ('left'), or ambiguous object pronouns ("it") as these lack specificity to produce meaningful visual landmarks. After filtering, we denote the remaining sub-instructions $S' \subset S$ . + +Visual imagination generation. The filtered subinstructions are passed through a text conditioned diffusion model as illustrated in Fig. 3. We use SDXL [32] as the diffusion model with positive and negative prompts cho + +![](images/4889dd68ffad037d0fd468b4bf220b92df48e6af134f01e0b8a25d397ab8370b.jpg) +Figure 2. An overview of our approach. (Left) Imaginations generated using valid sub-instructions from an instruction as determined by our filtering scheme are first passed to a pre-trained ViT to obtain feature vectors. A type embedding $t_{Im}$ for imagination modality is then added to the features which are encoded using a 3 layer MLP to obtain imagination embeddings $h_i$ . (Right) To integrate imagination modality to a VLN agent, the imagination embeddings $h_i$ are concatenated with instruction embeddings $t_i$ that are encoded using a text encoder $f_T(W)$ . The concatenated imagination-text embeddings are passed to the VLN agent's cross-modal encoder $f_X$ along with visual embeddings to predict a distribution over the agent's action space. + +Go down the stairs and then turn right. Go stand behind the easel by a unicycle. + +![](images/0582747fc800f4c828ddbcb37d0f10ef039bda15bd5c818bfe8d3d2ddf519df1.jpg) +Figure 3. Example instruction segmentation, filtering, and image generation. Instructions are segmented to sub-instructions leveraging FG-R2R [15] and then filtered to remove phrases referring to uninformative nouns (e.g., "right"). We produce visual imaginations using SDXL [32] for remaining sub-instructions. + +sen to steer generation towards indoor real-estate environments akin to typical environments in Matterport3D [7]. For example, providing "indoor" and "real estate" as positives prompts and "humans" and "collage" as negatives. See the supplementary for a full list. For a given instruction, we denote the generated images for valid sub-instructions as $\mathcal{Z} = \{Z_i \mid S_i \in S'\}$ . We colloquially refer to these generated images as imaginations and generate them for the entire R2R dataset. The resulting R2R-Imagine dataset consists of over 41k synthesized imaginations. This amounts to an average of 2.96 imaginations per instruction. + +# 3.2. Integration with Existing Models + +We consider augmentation with visual imagination to be a model agnostic approach and integrate it in two prior models - HAMT [8] and DUET [9]. Our integration approach consists of an imagination encoder, an alignment-promoting auxiliary loss, and a finetuning regimen (Fig. 2). Without loss of generality, we consider a decomposition of any VLN agent into an instruction encoder $f_{T}(W)$ , an observation/history encoder $f_{O}(O_{\leq t})$ , and a cross-modal policy network $f_{X}(f_{T}(W), f_{O}(O_{\leq t}))$ that combines the encoder outputs to predict a distribution over actions. Note that these may be fairly complex mechanisms and include explicit his + +tory encoding modules separate from encoding the current observation; however, this abstraction is sufficient here. + +Imagination encoder and integration. Given an instruction $W$ with corresponding imaginations $\mathcal{Z}$ , we independently encode each imagination $Z_{i}$ using a pretrained vision transformer (ViT) [10] to produce a $d$ -dim embedding vector. These are further transformed by a three-layer MLP such that the $i$ th imagination embedding can be written as + +$$ +h _ {i} = \operatorname {M L P} \left(\operatorname {V i T} \left(Z _ {i}\right) + t _ {I m}\right) \tag {1} +$$ + +where $t_{Im}$ is the type embedding for imagination modality. As we consider visual imaginations to be proxies for sub-instructions, the full set of imagination embeddings $\mathcal{H}$ are concatenated to the text encodings such that our general VLN agent structure now computes $f_{X}([f_{T}(W),\mathcal{H}],f_{O}(O_{\leq t}))$ where $[\cdot ,\cdot ]$ denotes a concatenation. Taking the DUET model as an example, the imagination embeddings are concatenated to the instruction encoding prior to being passed to the coarse-scale and fine-scale cross-modal encoders. + +Auxiliary alignment loss. We add an auxiliary loss to align the representations of imaginations and their corresponding sub-instruction's noun phrases. For a given sub-instruction $S_{i}$ , we denote the mean embedding vector for noun phrase tokens in $S_{i}$ from the text encoder $f_{T}$ as $\bar{S}_{i}$ . To align representations, we compute a cosine similarity loss between the corresponding imagination $Z_{i}$ encoded as $h_{i}$ and $\bar{S}_{i}$ . This is averaged over all $N_{Im}$ imagination-subinstruction pairs $(Z_{i}, S_{i})$ in a batch, + +$$ +\mathcal {L} _ {c o s} = \frac {1}{N _ {I m}} \sum_ {\left(Z _ {i}, S _ {i}\right) \in B} \left(1 - \frac {h _ {i} \cdot \bar {S} _ {i}}{\| h _ {i} \| \| \bar {S} _ {i} \|}\right). \tag {2} +$$ + +During finetuning, we combine this auxiliary loss $\mathcal{L}_{\mathrm{cos}}$ with the losses from the base agent with a scaling factor $\lambda$ such that the overall loss is $\mathcal{L}_{\mathrm{base}} + \lambda \mathcal{L}_{\mathrm{cos}}$ . + +Finetuning regimen. We start with fully-trained checkpoints of prior models as our initial weights. Depending + +Table 1. Comparison of our approach with selected prior work on the R2R dataset. Methods that use additional visual data beyond MP3D [7] are annotated with $\dagger$ and are not directly comparable with other approaches. Note that BEVBert* additionally uses panoramic depth images. Adding our visual imagination approach to HAMT and DUET base models (gray rows) leads to improved success rate (SR) and success weighted by inverse path length (SPL) metrics on the val-unseen split. For DUET, this effect also persists on the test set. We highlight these improvements over base models using bold numbers. + +
MethodsValidation SeenValidation UnseenTest Unseen
TLNE↓SR↑SPL↑TLNE↓SR↑SPL↑TLNE↓SR↑SPL↑
PREVALENT [14]10.323.67696510.194.71585310.515.305451
RecBERT [16]11.132.90726812.013.93635712.354.096357
ADAPT [27]10.972.54767212.213.77645812.993.796559
BEVBert* [3]13.562.17817414.552.81756415.873.137362
MARVAL [19]†10.602.99736910.154.06656110.224.186258
ScaleVLN [41]†11.902.16878012.402.34877914.272.738377
HAMT [8]11.152.5175.6172.1811.463.6266.2461.5112.273.936560
HAMT-Imagine (ours)11.302.3677.1673.7211.813.5867.2662.0212.663.896560
DUET [9]12.332.2878.8472.8813.943.3171.5260.4114.733.656959
DUET-Imagine (ours)12.932.1979.973.7514.353.1972.1260.4815.353.527160
+ +on the base method, these parameters may be the result of multi-stage pretraining and finetuning regimens - e.g., leveraging synthetic PREVALENT [14] data as prescribed in [8]. The imagination encoder is integrated to the base model and finetuned with the R2R-Imagine dataset. To mitigate catastrophic forgetting, we perform our training in three stages. First, we train only the newly introduced imagination encoder MLP along with type embeddings $t^{Im}$ with the remaining parameters frozen. Then, all the modules are trained jointly with a decreased learning rate for the base model. Finally, all the parameters are trained at a common learning rate. See supplementary materials for details on learning rates and schedules. + +# 4. Experiments + +Using the R2R dataset, we evaluate the impact of visual imaginations on our agent's navigation ability, perform ablation studies to explore the design space, and illustrate qualitative impacts of visual imaginations. Using the REVERIE dataset, we study the generalizability of our technique to a fine-grained instruction setting. + +# 4.1. Experimental setup + +R2R dataset. R2R [4] is constructed using Matterport3D [7] simulator. It consists of 90 indoor environments with 10,567 panoramas (each panorama consists of 36 single-view images) which act as nodes in a navigation graph. There are 7,189 trajectories each with 3 instructions and a corresponding ground truth trajectory. The dataset is split into train (4675 trajectories), val-seen (340 trajectories), + +val-unseen (783 trajectories), and test (1391 trajectories). The train and val-seen split share environments while val-unseen and test have different environments. + +R2R-Imagine dataset. The curation process of R2R-Imagine is prescribed in section 3.1. The number of imagination images are 41,558 for train, 3,055 for val-seen, 6,857 for val-unseen, and 12,412 for test. Our imaginations have a resolution of $1024 \times 1024$ and a single imagination generation on a single H100 GPU takes 3.2 seconds on average. We intend to release the dataset upon acceptance. + +REVERIE dataset. REVERIE [33] is constructed using Matterport3D and consists of high-level goals instead of finer grained step-by-step instructions. These instructions generally convey the target object and its location. The agent has to rely more on exploration relative to R2R. To be successful, agents must reach and identify the referent object. To synthesize imaginations for REVERIE, we treat the entire instruction as a single sub-instruction and follow the same recipe used for R2R. + +Evaluation metrics. We employ standard VLN metrics [18]. Success rate (SR) is the ratio of instruction-trajectories where the agent stops within a 3 meter radius from the goal and success normalized by inverse path length (SPL) helps distinguish successful episodes with shorter path lengths from longer path lengths. Navigation error (NE) is the distance between an agent's final position and the goal in meters. Trajectory length (TL) is the length of an agent's traversed path in meters. REVERIE [33] additionally defines grounding metrics Remote Grounding Success (RGS) and RGS penalized by path length (RGSPL). + +Table 2. Performance comparison of our approach on REVERIE with baseline. We observe improvements of our method across all metrics on DUET when provided coarse-grained instructions. + +
SR↑SPL↑RGS↑RGSPL↑
DUET46.9833.7332.1523.03
DUET-Imagine (ours)48.2833.7632.9723.25
+ +HAMT and DUET base models. We apply our approach to two baseline representative architectures - HAMT [8] and DUET [9]. HAMT is hierarchical and transformer-based; instructions, observations, and trajectory history are encoded separately using unimodal transformers and then combined in a single cross-modal transformer to predict actions. DUET incrementally forms a topographic map of the environment over time and uses dual fine- and coarse-grained cross-modal attention mechanisms to associate language embeddings with the local observation and global topographic features. The results are then fused to predict actions. For both base models, we inject visual imaginations via concatenation with the text embeddings (Sec. 3.2). We use provided checkpoints for pretrained models. + +Implementation details. We follow base agent [8, 9] settings unless specified otherwise. We use a standard off-the-shelf pretrained ViT-B/16 [10] to encode imaginations for both HAMT and DUET. Our VLN-Imagine agents are finetuned for $100\mathrm{k}$ iterations in three stages as noted in section 3.2. The finetuning process is carried out using a Tesla V100 GPU with a batch size of 8 and takes $\sim 1.5$ days for each agent. For combining the proposed auxiliary loss function with the base model losses, we empirically set $\lambda = 0.5$ . + +# 4.2. Results + +Our experiments demonstrate that providing generated images of instruction landmarks (imaginations) improves the performance of vision-and-language navigation models (Tab. 1). We highlight our key findings below. + +VLN agents perform better with imaginations. VLN agents equipped with imaginations perform better than VLN agents without by 0.6-1.0 SR and up to 0.5 SPL on R2R val-unseen (Tab. 1). This improvement coincides with lower navigation error (NE) across both models and all dataset splits. We contextualize these results relative to agent architectures that are representative of the VLN literature. We additionally report a distinct family of techniques that scale training environments to boost performance by reducing overfitting - MARVAL [19] and ScaleVLN [41]. We consider approaches of this class orthogonal to our study and instead use visual data only from Matterport3D [7]. + +Imaginations are useful in coarse-grained instruction navigation. Our method is performant when natural language instructions convey high-level objectives (REVERIE, + +Table 3. Role of imaginations; HAMT-Imagine. Null imaginations refers to testing with no imaginations, wrong imaginations refers to testing with imaginations sampled from different instructions. We observe aligned imaginations are important to navigation performance. In addition, results with no imaginations are better than baseline implying a regularization effect during training. + +
SR↑SPL↑
Baseline (HAMT)66.2461.51
Null imaginations66.9261.89
Wrong imaginations66.2461.02
Correct imaginations (ours)67.2662.02
+ +Table 4. Sequential vs. goal imagination; HAMT-Imagine. We study the effect of providing just the imagination belonging to the final valid sub-instruction relative to full valid imaginations. We observe providing just the goal imagination improves performance significantly compared to baseline but shows lower performance relative to providing full imaginations. + +
SR↑SPL↑
Baseline (HAMT)66.2461.51
Goal imagination66.7961.58
Full imaginations (ours)67.2662.02
+ +Tab. 2); navigation performance improves by 1.3 SR and grounding performance improves by 0.82 RGS. We hypothesize that imaginations not only aid in reaching the goal but also in identifying and grounding the target concept amongst others objects in the scene. + +VLN agents benefit from both imagination training and inference. When nullifying imaginations at test-time, VLN-Imagine agents still surpass baseline performance, but to a lesser extent (Tab. 3). We implement this nullification by setting the imagination attention masks to zero in the cross-modal encoder. We hypothesize that imagination-based training provides a regularizing effect on the overall VLN agent during fine-tuning, revealing a research avenue for strong and performant test-time model inference. We further find that imaginations must be aligned with the instruction to provide maximal benefit; when fed with imaginations randomly-sampled from the R2R-Imagine dataset, VLN-Imagine agents perform worse than baseline (Tab. 3). + +Sequential imagination outperforms goal imagination. Sequential refers to imaginations generated over multiple sub-instructions (our method). Goal imagination refers to imagining just the final non-filtered sub-instruction. We show these results in Tab. 4 – while goal imagination outperforms the baseline by 0.5 SR, sequential imagination outperforms the baseline by 1.0 SR. Goal imagination may be particularly useful in guiding the agent to stop at the destination. While distinct from our setting, such goal imagi + +Table 5. Comparison between ViT models. We train both HAMT-Imagine and DUET-Imagine agents with a ViT finetuned in-the-loop with HAMT and an off-the-shelf pretrained ViT. We observe no consistent effects across the both agents. + +
ViT fine-tunedViT off-the-shelf
SR↑SPL↑SR↑SPL↑
HAMT-Imagine67.3161.567.2662.02
DUET-Imagine72.5860.2972.1260.48
+ +Table 6. Loss ablation. We study the effect of negative samples in our alignment loss by comparing our approach of using cosine loss with InfoNCE [28] loss. In addition, we train the agent with no auxiliary loss represented by "No loss". We do not observe a significant difference between contrastive and cosine losses. + +
SR↑SPL↑
No loss66.7561.60
LInfoNCE67.1861.84
Lcos(ours)67.2662.02
+ +nation has parallels to ImageNav [22], a visual navigation task that conveys goals solely using exemplar images. + +General-purpose vision encoders are sufficient imagination encoders. We compare encoding imaginations with off-the-shelf ViT [10] vs. ViT fine-tuned on R2R navigation episodes [8]. We find comparable performance (Tab. 5) and choose to employ off-the-shelf ViT in our final model. This is motivated to maximize generality and to avoid potential issues relating to catastrophic forgetting. For example, generalization of frozen ViT is demonstrated in Zeng et al. [44]. + +Aligning imaginations to instructions benefits VLN-imagine agents. Aligning the imagination embedding to the instruction embedding as an auxiliary loss leads to a 0.5 SR and a 0.4 SPL increase over no auxiliary loss (Tab. 6). We also consider formulating this loss as contrastive - do negative samples enable a stronger alignment and thus better downstream performance? We find that an InfoNCE loss [28] yields no significant difference relative to our cosine similarity loss (Eq. 2). We construct negative samples using mean noun phrase embeddings from imaginations of other instructions in the batch. The overall loss is $\mathcal{L}_{\mathrm{base}} + \lambda \mathcal{L}_{\mathrm{InfoNCE}}$ where we set $\lambda$ to 0.2 empirically. + +Imaginations correctly represent noun phrases in a subinstruction. To evaluate fidelity of our imaginations, we run an open-vocabulary object detector (LangSAM [1]) on our imaginations to test if the noun phrases contained in the corresponding sub-instruction from R2R are detected successfully. We see in Tab. 7 that at least one noun phrase is detected in our imaginations for $98.78\%$ of sub-instructions and all noun phrases are detected for $94.99\%$ + +Table 7. Open vocabulary object detection accuracy (\%) of imaginations to noun phrases from sub-instructions. Our imaginations represent the noun phrases with a high accuracy. In val-seen, in $98.43\%$ of sub-instructions, at least 1 noun phrase is detected and in $94.51\%$ of sub-instructions, all noun phrases are detected. + +
Sub-instructionsNoun Phrase Detection (%)
≥1 detectedAll detected
Val-seen298598.4394.51
Val-unseen670098.9995.34
Test1207598.9395.12
+ +of sub-instructions. We conclude our imaginations provide a high coverage of relevant concepts across splits. + +# 4.3. Qualitative visualizations + +In this section, we provide a qualitative investigation of how imaginations may be used as a pivot between language tokens and visual observations from the environment by examining the attention patterns in HAMT-Imagine. + +For a given instruction, we randomly select an imagination / sub-instruction pair and examine the agent's trajectory to identify the first occurrence of the reference object. We then examine the attention distributions between language tokens, the visual imagination, and the panoramic observation. As there are many attention heads across multiple layers in the cross-modal encoder, we structure our study by focusing on those attention heads where sub-instruction tokens are attended to most strongly by their corresponding visual imagination and then examine how the visual imagination attends to observations. + +We then visualize sub-instruction and imagination pairs along with their top attended language tokens and visual observations in Fig. 4. By construction, these are time steps where the referent is visible and the visual imagination is attending to the appropriate noun phrase. As such, our examples primarily demonstrate how the imagination learns to associate with the visual observations. We focus on this aspect as our auxiliary loss already explicitly encourages alignment between visual imaginations and sub-instruction. + +From the examples in Fig. 4, we make two remarks. Imaginations attend most strongly to observations of matching semantic concepts. In row 1, the imagination and the retrieved observations all depict a cheetah (either a painting or floor mat). Imagination attention scores mirror the proper disambiguation of concepts. In row 2, the sub-instruction (and thus the imagination) refer to a unicycle. While both a bicycle and unicycle are observed during execution, attention scores are maximized by the observations containing the unicycle. These patterns do not directly imply a causal relation between the attention scores and eventual decision making of the VLN agent. Rather, they illustrate how imag- + +
Sub-instructionImaginationTop Attended TokensTop Attended Observations
And walk towards the door pass by the cheetah paintingche, ##eta, ##h, painting
Go stand behind the easel by a unicycle next to a doorun, ##ic. ##y, ##cle, to
Once in the room, walk to the left of the pool tablepool, door, room
Walk into the arch hallway near the Native American painting on the wallAmerican, Native, candles
+ +Figure 4. Qualitative examples showing imaginations as pivots between language and observation images. The first column contains sub-instruction from a random instruction from R2R, the second column contains the imagination generated using the sub-instruction. The third and fourth columns show highest attended language tokens and observation images from an attention head in HAMT's cross-modal transformer at a time step the associated observations are first visible. In the second example (row 2), the sub-instruction references "unicycle" which is captured in the imagination along with neighboring nouns "easel" and "door". We observe that in a head where top attending language tokens to the imagination query are references to nouns associated with the sub-instruction, its top attended observations to the imagination query are images of the same concept ("unicycle"). In this example, the imagination of a unicycle is being used to associate language tokens belonging to "unicycle" to observations of unicycle hinting at the utility of imaginations in navigation. + +inations can serve as a pivot between language and observations and afford disambiguation of related concepts. + +# 5. Conclusion + +In this work, we show that visual representations of instructions, or visual imaginations, play an additive role in training superior navigating agents for vision-and-language navigation (VLN) tasks. We generate a sequence of visual imaginations depicting sub-instruction landmarks using a text-to-image diffusion model. Our approach is model-agnostic and we integrate these imaginations into VLN agents using an imagination encoder, an auxiliary loss, and a fine-tuning training step. We apply this approach to two representative VLN agents across two datasets and observe performance improvements of around 1 SR and 0.5 SPL. + +Limitations. Generating and encoding imaginations increases the computational cost of running VLN agents. This is particularly relevant when running agents on-device for real-world robotic deployments. However, we note that this process is performed upfront, unlike techniques that depend on step-wise observations like novel view synthesis (NVS). + +Stepping more broadly, by design, our imaginations are not grounded in the environment. Personalized reasoning, such as the unique naming of objects and locations, cannot be directly imagined in our framework. Life-long learning of persistent groundings is an open question not just in generating imaginations, but in the VLN space at large. + +Future directions. A natural question that arises from our research is the exact complementary nature of the roles played by visual imaginations and language instructions. This is a rich area for future work. Similar studies relating language and visual observations have been conducted in [48] and [47]. Our paper demonstrated the fundamental utility of imagination in VLN agents. Beyond this lies interesting extensions, such as how imagination can help bridge the simulation-to-reality (Sim2Real) gap for VLN agents and whether imaginations can unlock the performance of VLN world models through image-image reasoning. + +Acknowledgments. We thank Zijiao Yang and Robert Yelle for their assistance with simulator issues and technical support for high performance computing infrastructure. + +# References + +[1] Language segment-anything. https://github.com/luca-medeiros/lang-segment-anything/.7 +[2] spacy: Industrial-strength natural language processing in python. https://spacy.io/.3 +[3] Dong An, Yuankai Qi, Yangguang Li, Yan Huang, Liang Wang, Tieniu Tan, and Jing Shao. Bevbert: Multimodal map pre-training for language-guided navigation. Proceedings of the IEEE/CVF International Conference on Computer Vision, 2023. 5 +[4] Peter Anderson, Qi Wu, Damien Teney, Jake Bruce, Mark Johnson, Niko Sünderhauf, Ian D. Reid, Stephen Gould, and Anton van den Hengel. Vision-and-language navigation: Interpreting visually-grounded navigation instructions in real environments. 2018 IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 3674-3683, 2017. 2, 3, 5 +[5] Homanga Bharadhwaj, Debidatta Dwibedi, Abhinav Gupta, Shubham Tulsiani, Carl Doersch, Ted Xiao, Dhruv Shah, Fei Xia, Dorsa Sadigh, and Sean Kirmani. Gen2act: Human video generation in novel scenarios enables generalizable robot manipulation. arXiv preprint arXiv:2409.16283, 2024. 3 +[6] Kevin Black, Mitsuhiko Nakamoto, Pranav Atreya, Homer Rich Walke, Chelsea Finn, Aviral Kumar, and Sergey Levine. Zero-shot robotic manipulation with pre-trained image-editing diffusion models. In The Twelfth International Conference on Learning Representations. 3 +[7] Angel Chang, Angela Dai, Thomas Funkhouser, Maciej Halber, Matthias Niebner, Manolis Savva, Shuran Song, Andy Zeng, and Yinda Zhang. Matterport3d: Learning from rgb-d data in indoor environments. In 2017 International Conference on 3D Vision (3DV), pages 667-676, 2017. 2, 4, 5, 6 +[8] Shizhe Chen, Pierre-Louis Guhur, Cordelia Schmid, and Ivan Laptev. History aware multimodal transformer for vision-and-language navigation. In NeurIPS, 2021. 2, 3, 4, 5, 6, 7 +[9] Shizhe Chen, Pierre-Louis Guhur, Makarand Tapaswi, Cordelia Schmid, and Ivan Laptev. Think global, act local: Dual-scale graph transformer for vision-and-language navigation. In CVPR, 2022. 2, 3, 4, 5, 6 +[10] Alexey Dosovitskiy, Lucas Beyer, Alexander Kolesnikov, Dirk Weissenborn, Xiaohua Zhai, Thomas Unterthiner, Mostafa Dehghani, Matthias Minderer, Georg Heigold, Sylvain Gelly, et al. An image is worth 16x16 words: Transformers for image recognition at scale. arXiv preprint arXiv:2010.11929, 2020. 4, 6, 7 +[11] Yilun Du, Sherry Yang, Bo Dai, Hanjun Dai, Ofir Nachum, Josh Tenenbaum, Dale Schuurmans, and Pieter Abbeel. Learning universal policies via text-guided video generation. Advances in Neural Information Processing Systems, 36, 2024. 3 +[12] Daniel Fried, Ronghang Hu, Volkan Cirik, Anna Rohrbach, Jacob Andreas, Louis-Philippe Morency, Taylor Berg-Kirkpatrick, Kate Saenko, Dan Klein, and Trevor Darrell. Speaker-follower models for vision-and-language naviga + +tion. In Neural Information Processing Systems (NeurIPS), 2018. 2 +[13] Jialu Gao, Kaizhe Hu, Guowei Xu, and Huazhe Xu. Can pre-trained text-to-image models generate visual goals for reinforcement learning? Advances in Neural Information Processing Systems, 36:38297-38310, 2023. 3 +[14] Weituo Hao, Chunyuan Li, Xiujun Li, Lawrence Carin, and Jianfeng Gao. Towards learning a generic agent for vision-and-language navigation via pre-training, 2020. 2, 5 +[15] Yicong Hong, Cristian Rodriguez, Qi Wu, and Stephen Gould. Sub-instruction aware vision-and-language navigation. In Proceedings of the 2020 Conference on Empirical Methods in Natural Language Processing (EMNLP), pages 3360-3376, 2020. 3, 4 +[16] Yicong Hong, Qi Wu, Yuankai Qi, Cristian Rodriguez-Opazo, and Stephen Gould. A recurrent vision-and-language bert for navigation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 1643–1653, 2021. 5 +[17] Yicong Hong, Qi Wu, Yuankai Qi, Cristian Rodriguez-Opazo, and Stephen Gould. A recurrent vision-and-language bert for navigation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 1643–1653, 2021. 2 +[18] Gabriel Ilharco, Vihan Jain, Alexander Ku, Eugene Ie, and Jason Baldridge. General evaluation for instruction conditioned navigation using dynamic time warping. ArXiv, abs/1907.05446, 2019. 5 +[19] Aishwarya Kamath, Peter Anderson, Su Wang, Jing Yu Koh, Alexander Ku, Austin Waters, Yinfei Yang, Jason Baldridge, and Zarana Parekh. A new path: Scaling vision-and-language navigation with synthetic instructions and imitation learning. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 10813-10823, 2023. 2, 5, 6 +[20] Ivan Kapelyukh, Vitalis Vosylius, and Edward Johns. Dallebot: Introducing web-scale diffusion models to robotics. 2023. 3 +[21] Jing Yu Koh, Honglak Lee, Yinfei Yang, Jason Baldridge, and Peter Anderson. Pathdreamer: A world model for indoor navigation, 2021. 3 +[22] Jacob Krantz, Stefan Lee, Jitendra Malik, Dhruv Batra, and Devendra Singh Chaplot. Instance-specific image goal navigation: Training embodied agents to find object instances, 2022. 7 +[23] Chongxi Lai, Shinsuke Tanaka, Timothy D. Harris, and Albert K. Lee. Volitional activation of remote place representations with a hippocampal brain-machine interface. Science, 382(6670):566-573, 2023. 2 +[24] Jialu Li and Mohit Bansal. Panogen: Text-conditioned panoramic environment generation for vision-and-language navigation. arxiv, 2023. 3 +[25] Jialu Li and Mohit Bansal. Improving vision-and-language navigation by generating future-view image semantics. CVPR, 2023. 3 +[26] Mingxiao Li, Zehao Wang, Tinne Tuytelaars, and Marie-Francine Moens. Layout-aware dreamer for embodied visual referring expression grounding. In Proceedings of the + +AAAI Conference on Artificial Intelligence, pages 1386-1395, 2023. 2 +[27] Bingqian Lin, Yi Zhu, Zicong Chen, Xiwen Liang, Jianzhuang Liu, and Xiaodan Liang. Adapt: Vision-language navigation with modality-aligned action prompts. In 2022 IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 15375-15385, 2022. 2, 5 +[28] Aaron van den Oord, Yazhe Li, and Oriol Vinyals. Representation learning with contrastive predictive coding. arXiv preprint arXiv:1807.03748, 2018. 7 +[29] Allan Paivio. Abstractness, imagery, and meaningfulness in paired-associate learning. Journal of verbal learning and verbal behavior., 4(1), 1965-2. 2 +[30] Allan Paivio. A factor-analytic study of word attributes and verbal learning. Journal of verbal learning and verbal behavior, 7(1), 1968-2. +[31] A. Paivio, P. C. Smythe, and J. C. Yuille. Imagery versus meaningfulness of nouns in paired-associate learning. Canadian Journal of Psychology / Revue Canadienne De Psychologie, 22:427–441, 1968. 2 +[32] Dustin Podell, Zion English, Kyle Lacey, Andreas Blattmann, Tim Dockhorn, Jonas Müller, Joe Penna, and Robin Rombach. Sdxl: Improving latent diffusion models for high-resolution image synthesis, 2023. 2, 3, 4 +[33] Yuankai Qi, Qi Wu, Peter Anderson, Xin Wang, William Yang Wang, Chunhua Shen, and Anton van den Hengel. Reverie: Remote embodied visual referring expression in real indoor environments. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 9982-9991, 2020. 2, 5 +[34] Alec Radford, Jong Wook Kim, Chris Hallacy, Aditya Ramesh, Gabriel Goh, Sandhini Agarwal, Girish Sastry, Amanda Askell, Pamela Mishkin, Jack Clark, et al. Learning transferable visual models from natural language supervision. In International conference on machine learning, pages 8748-8763. PMLR, 2021. 1, 2 +[35] Santhosh K Ramakrishnan, Aaron Gokaslan, Erik Wijmans, Oleksandr Maksymets, Alex Clegg, John Turner, Eric Undersander, Wojciech Galuba, Andrew Westbury, Angel X Chang, et al. Habitat-matterport 3d dataset (hm3d): 1000 large-scale 3d environments for embodied ai. arXiv preprint arXiv:2109.08238, 2021. 2 +[36] Marianne Cumella Reddan, Tor Dessart Wager, and Daniela Schiller. Attenuating neural threat expression with imagination. *Neuron*, 100(4):994–1005.e4, 2018. 2 +[37] Manolis Savva, Abhishek Kadian, Oleksandr Maksymets, Yili Zhao, Erik Wijmans, Bhavana Jain, Julian Straub, Jia Liu, Vladlen Koltun, Jitendra Malik, et al. Habitat: A platform for embodied ai research. In Proceedings of the IEEE/CVF international conference on computer vision, pages 9339-9347, 2019. 2 +[38] Daniel L. Schacter, Donna Rose Addis, and Randy L. Buckner. Remembering the past to imagine the future: the prospective brain. Nature Reviews Neuroscience, 8(9):657-661, 2007. 2 +[39] Roger N. Shepard and Jacqueline Metzler. Mental rotation of three-dimensional objects. Science, 171(3972):701-703, 1971. 2 + +[40] Hao Tan, Licheng Yu, and Mohit Bansal. Learning to navigate unseen environments: Back translation with environmental dropout. In Proceedings of the 2019 Conference of the North American Chapter of the Association for Computational Linguistics: Human Language Technologies, Volume 1 (Long and Short Papers), pages 2610-2621, Minneapolis, Minnesota, 2019. Association for Computational Linguistics. 2 +[41] Zun Wang, Jialu Li, Yicong Hong, Yi Wang, Qi Wu, Mohit Bansal, Stephen Gould, Hao Tan, and Yu Qiao. Scaling data generation in vision-and-language navigation. In ICCV 2023, 2023. 2, 5, 6 +[42] Fei Xia, Amir R. Zamir, Zhi-Yang He, Alexander Sax, Jitendra Malik, and Silvio Savarese. Gibson env: real-world perception for embodied agents. In Computer Vision and Pattern Recognition (CVPR), 2018 IEEE Conference on. IEEE, 2018. 2 +[43] Zijiao Yang, Arjun Majumdar, and Stefan Lee. Behavioral analysis of vision-and-language navigation agents. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 2574-2582, 2023. 1 +[44] Kuo-Hao Zeng, Zichen Zhang, Kiana Ehsani, Rose Hendrix, Jordi Salvador, Alvaro Herrasti, Ross Girshick, Aniruddha Kembhavi, and Luca Weihs. Poliformer: Scaling on-policy rl with transformers results in masterful navigators. In 8th Annual Conference on Robot Learning, 2024. 7 +[45] Gengze Zhou, Yicong Hong, and Qi Wu. Navigpt: Explicit reasoning in vision-and-language navigation with large language models. Proceedings of the AAAI Conference on Artificial Intelligence, 38(7):7641-7649, 2024. 2 +[46] Gengze Zhou, Yicong Hong, Zun Wang, Xin Eric Wang, and Qi Wu. Navigpt-2: Unleashing navigational reasoning capability for large vision-language models. In European Conference on Computer Vision, pages 260–278. Springer, 2025. 2 +[47] Wanrong Zhu, Yuankai Qi, Pradyumna Narayana, Kazoo Sone, Sugato Basu, Xin Wang, Qi Wu, Miguel Eckstein, and William Yang Wang. Diagnosing vision-and-language navigation: What really matters. In Proceedings of the 2022 Conference of the North American Chapter of the Association for Computational Linguistics: Human Language Technologies, pages 5981-5993, 2022. 1, 8 +[48] Wang Zhu, Ishika Singh, Yuan Huang, Robin Jia, and Jesse Thomason. Does vln pretraining work with nonsensical or irrelevant instructions? ArXiv, abs/2311.17280, 2023. 8 \ No newline at end of file diff --git a/2025/Do Visual Imaginations Improve Vision-and-Language Navigation Agents_/images.zip b/2025/Do Visual Imaginations Improve Vision-and-Language Navigation Agents_/images.zip new file mode 100644 index 0000000000000000000000000000000000000000..0cc95bb0db77ddac0a8262197d7b6bd9ab724f35 --- /dev/null +++ b/2025/Do Visual Imaginations Improve Vision-and-Language Navigation Agents_/images.zip @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:ff8eaf2b079a64b6aba461f1850e10f73feb57a2fa2fd58c38ad7e120e67914c +size 434406 diff --git a/2025/Do Visual Imaginations Improve Vision-and-Language Navigation Agents_/layout.json b/2025/Do Visual Imaginations Improve Vision-and-Language Navigation Agents_/layout.json new file mode 100644 index 0000000000000000000000000000000000000000..526fd05c7af60b6db7b2f360f988fbc4cc7693b7 --- /dev/null +++ b/2025/Do Visual Imaginations Improve Vision-and-Language Navigation Agents_/layout.json @@ -0,0 +1,7302 @@ +{ + "pdf_info": [ + { + "para_blocks": [ + { + "bbox": [ + 74, + 103, + 536, + 121 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 74, + 103, + 536, + 121 + ], + "spans": [ + { + "bbox": [ + 74, + 103, + 536, + 121 + ], + "type": "text", + "content": "Do Visual Imaginations Improve Vision-and-Language Navigation Agents?" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 193, + 143, + 416, + 172 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 193, + 143, + 416, + 172 + ], + "spans": [ + { + "bbox": [ + 193, + 143, + 416, + 172 + ], + "type": "text", + "content": "Akhil Perincherry1, Jacob Krantz, Stefan Lee1 \n1Oregon State University" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 110, + 174, + 496, + 185 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 110, + 174, + 496, + 185 + ], + "spans": [ + { + "bbox": [ + 110, + 174, + 496, + 185 + ], + "type": "text", + "content": "perincha@oregonstate.edu, jkrantz989@gmail.com, leestef@oregonstate.edu" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 151, + 213, + 200, + 225 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 151, + 213, + 200, + 225 + ], + "spans": [ + { + "bbox": [ + 151, + 213, + 200, + 225 + ], + "type": "text", + "content": "Abstract" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 55, + 239, + 296, + 467 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 239, + 296, + 467 + ], + "spans": [ + { + "bbox": [ + 55, + 239, + 296, + 467 + ], + "type": "text", + "content": "Vision-and-Language Navigation (VLN) agents are tasked with navigating an unseen environment using natural language instructions. In this work, we study if visual representations of sub-goals implied by the instructions can serve as navigational cues and lead to increased navigation performance. To synthesize these visual representations or \"imaginations\", we leverage a text-to-image diffusion model on landmark references contained in segmented instructions. These imaginations are provided to VLN agents as an added modality to act as landmark cues and an auxiliary loss is added to explicitly encourage relating these with their corresponding referring expressions. Our findings reveal an increase in success rate (SR) of " + }, + { + "bbox": [ + 55, + 239, + 296, + 467 + ], + "type": "inline_equation", + "content": "\\sim 1" + }, + { + "bbox": [ + 55, + 239, + 296, + 467 + ], + "type": "text", + "content": " point and up to " + }, + { + "bbox": [ + 55, + 239, + 296, + 467 + ], + "type": "inline_equation", + "content": "\\sim 0.5" + }, + { + "bbox": [ + 55, + 239, + 296, + 467 + ], + "type": "text", + "content": " points in success scaled by inverse path length (SPL) across agents. These results suggest that the proposed approach reinforces visual understanding compared to relying on language instructions alone. Code and data for our work can be found at https://www.akhilperincherry.com/VLN-Imagine-website/." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 55, + 490, + 135, + 502 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 490, + 135, + 502 + ], + "spans": [ + { + "bbox": [ + 55, + 490, + 135, + 502 + ], + "type": "text", + "content": "1. Introduction" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 55, + 510, + 296, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 510, + 296, + 713 + ], + "spans": [ + { + "bbox": [ + 55, + 510, + 296, + 713 + ], + "type": "text", + "content": "In this work, we examine whether providing visual imagery corresponding to described landmarks improves the performance of agents following natural language navigation instructions. Consider the natural language navigation instruction presented in Figure 1 which asks an agent to \"Go straight, take a left at the pool table to enter the kitchen. Walk to the bedroom and stop\". This instruction provides unconditional action directives like \"go straight\" but also frequently conditions the given directions on visual landmarks in the scene such as the pool table, kitchen, and bedroom. Standard approaches to vision-and-language navigation tasks rely on learned cross-modal alignment mechanisms to implicitly associate these noun phrases with their visual referents during navigation. However, text-to-image generation models have improved to the point that producing imagery matching the semantics of these visual references prior to navigation is plausible. We show three such" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 322, + 239, + 462, + 249 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 322, + 239, + 462, + 249 + ], + "spans": [ + { + "bbox": [ + 322, + 239, + 462, + 249 + ], + "type": "text", + "content": "Natural Language Navigation Instruction:" + } + ] + } + ], + "index": 9 + }, + { + "type": "image", + "bbox": [ + 323, + 252, + 344, + 275 + ], + "blocks": [ + { + "bbox": [ + 323, + 252, + 344, + 275 + ], + "lines": [ + { + "bbox": [ + 323, + 252, + 344, + 275 + ], + "spans": [ + { + "bbox": [ + 323, + 252, + 344, + 275 + ], + "type": "image", + "image_path": "705f98388715ea92f7dd13c69a293d935e23117c9638211362f8833f3ce60dcd.jpg" + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "image_body" + } + ], + "index": 10 + }, + { + "bbox": [ + 351, + 253, + 547, + 275 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 351, + 253, + 547, + 275 + ], + "spans": [ + { + "bbox": [ + 351, + 253, + 547, + 275 + ], + "type": "text", + "content": "Go straight, take a left at the pool table to enter the kitchen. Walk to the bedroom and stop." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 323, + 283, + 391, + 293 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 323, + 283, + 391, + 293 + ], + "spans": [ + { + "bbox": [ + 323, + 283, + 391, + 293 + ], + "type": "text", + "content": "Visual Imaginations:" + } + ] + } + ], + "index": 12 + }, + { + "type": "image", + "bbox": [ + 320, + 293, + 345, + 361 + ], + "blocks": [ + { + "bbox": [ + 320, + 293, + 345, + 361 + ], + "lines": [ + { + "bbox": [ + 320, + 293, + 345, + 361 + ], + "spans": [ + { + "bbox": [ + 320, + 293, + 345, + 361 + ], + "type": "image", + "image_path": "7734d74b0f516f15705c4d935639aa995a10be9f1d968980a406f0abd2a3c2f0.jpg" + } + ] + } + ], + "index": 13, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 313, + 373, + 555, + 439 + ], + "lines": [ + { + "bbox": [ + 313, + 373, + 555, + 439 + ], + "spans": [ + { + "bbox": [ + 313, + 373, + 555, + 439 + ], + "type": "text", + "content": "Figure 1. Illustration of visual imaginations. (Top) A natural language instruction specifying sub-goals pool table, kitchen, and bedroom. (Bottom) Visual imaginations of landmarks pool table, kitchen and bedroom referenced by the sub-goals in the instruction. In our work, we study if these visual imaginations generated using text-to-image models can improve performance in VLN." + } + ] + } + ], + "index": 17, + "angle": 0, + "type": "image_caption" + } + ], + "index": 13 + }, + { + "type": "image", + "bbox": [ + 351, + 296, + 413, + 361 + ], + "blocks": [ + { + "bbox": [ + 351, + 296, + 413, + 361 + ], + "lines": [ + { + "bbox": [ + 351, + 296, + 413, + 361 + ], + "spans": [ + { + "bbox": [ + 351, + 296, + 413, + 361 + ], + "type": "image", + "image_path": "924ec74d18c943cbc11f8635fd22e9f6fd6d8af80838749c8db460cc81cdb261.jpg" + } + ] + } + ], + "index": 14, + "angle": 0, + "type": "image_body" + } + ], + "index": 14 + }, + { + "type": "image", + "bbox": [ + 419, + 296, + 480, + 361 + ], + "blocks": [ + { + "bbox": [ + 419, + 296, + 480, + 361 + ], + "lines": [ + { + "bbox": [ + 419, + 296, + 480, + 361 + ], + "spans": [ + { + "bbox": [ + 419, + 296, + 480, + 361 + ], + "type": "image", + "image_path": "90b84f5ea45bd4c6ac8966bf11611da60bbbffea94897ac8d7641849870f928d.jpg" + } + ] + } + ], + "index": 15, + "angle": 0, + "type": "image_body" + } + ], + "index": 15 + }, + { + "type": "image", + "bbox": [ + 487, + 296, + 551, + 361 + ], + "blocks": [ + { + "bbox": [ + 487, + 296, + 551, + 361 + ], + "lines": [ + { + "bbox": [ + 487, + 296, + 551, + 361 + ], + "spans": [ + { + "bbox": [ + 487, + 296, + 551, + 361 + ], + "type": "image", + "image_path": "388633a477afaf69d30be05e9f4ef9df5c0b90d03ce290669461593b9c8146dd.jpg" + } + ] + } + ], + "index": 16, + "angle": 0, + "type": "image_body" + } + ], + "index": 16 + }, + { + "bbox": [ + 313, + 450, + 555, + 533 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 450, + 555, + 533 + ], + "spans": [ + { + "bbox": [ + 313, + 450, + 555, + 533 + ], + "type": "text", + "content": "generations for our running example in Figure 1. Drawing an analogy to the substantial work in cognitive science on the impact of mental imagery on task performance, we refer to these generated images as visual imaginations and study whether providing them in addition to corresponding language-based instructions can improve the performance of vision-and-language navigation agents." + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 313, + 534, + 556, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 534, + 556, + 714 + ], + "spans": [ + { + "bbox": [ + 313, + 534, + 556, + 714 + ], + "type": "text", + "content": "But why might we believe visual imaginations could be beneficial? In short - text-to-image models have broad knowledge, and learning to perform semantic matching in the image domain may be an easier task than language grounding. Associating noun phrases and their visual referents is a sub-task in vision-and-language navigation (VLN) that most works address through extensive pretraining of model components on web-scale image caption datasets - e.g., by initializing image and language encoders with CLIP [34] models. While this provides a useful warm-start, it is unclear how well the remaining cross-modal components that drive navigation retain these capabilities after downstream VLN training. Downstream datasets may cover only a small portion of relevant objects and visual grounding remains a weakness in many existing VLN models [43, 47]." + } + ] + } + ], + "index": 19 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "spans": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "type": "text", + "content": "CVF" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 145, + 0, + 496, + 37 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 145, + 0, + 496, + 37 + ], + "spans": [ + { + "bbox": [ + 145, + 0, + 496, + 37 + ], + "type": "text", + "content": "This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "text", + "content": "3846" + } + ] + } + ], + "index": 20 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 0 + }, + { + "para_blocks": [ + { + "bbox": [ + 55, + 72, + 294, + 263 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 72, + 294, + 263 + ], + "spans": [ + { + "bbox": [ + 55, + 72, + 294, + 263 + ], + "type": "text", + "content": "Directly synthesizing visual imaginations using off-the-shelf text-to-image models offers alternative training and inference mechanisms. Synthesized images have known correspondences to noun phrases in instructions that can be leveraged to reinforce visual grounding during downstream task training through auxiliary losses. Further, matching a visual imagination with environment observations reduces to an image-to-image matching task – offering the potential to use visual imaginations as a convenient pivot between language and vision. This may prove especially effective for out-of-distribution landmarks that have limited or no support during downstream task training. For example, novel instructions referencing “butterfly sculptures” or “Pulp Fiction posters” are unlikely to be well-grounded in the training set but produce semantically relevant visual imaginations that could then be matched to observations." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 55, + 265, + 294, + 385 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 265, + 294, + 385 + ], + "spans": [ + { + "bbox": [ + 55, + 265, + 294, + 385 + ], + "type": "text", + "content": "Though we do not imply any direct connections, our work is also inspired by cognitive science studies which indicate mental imagery can serve as valuable cues in addition to linguistic stimuli. Several works supporting dualcoding theory [29-31] find that concrete imagery tends to be more effective for certain types of concept learning compared to linguistic stimuli alone. Further, findings from human and animal studies [23, 36, 38, 39] suggest that imagining snapshots of the environment or potential trajectories before navigating can improve decision making." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 55, + 387, + 294, + 495 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 387, + 294, + 495 + ], + "spans": [ + { + "bbox": [ + 55, + 387, + 294, + 495 + ], + "type": "text", + "content": "To investigate our key question, we augment agents with visual imaginations for the popular Room-2-Room (R2R) [4] and REVERIE [33] tasks. We propose a VLN agent agnostic procedure to incorporate visual imaginations as additional inputs and introduce a text-imagination alignment loss to explicitly encourage visual grounding during training. We generate visual imaginations for key noun phrases in instructions that refer to visual landmarks using an off-the-shelf text-to-image generation model [32]." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 55, + 497, + 294, + 616 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 497, + 294, + 616 + ], + "spans": [ + { + "bbox": [ + 55, + 497, + 294, + 616 + ], + "type": "text", + "content": "Applying our technique to two existing models, HAMT [8] and DUET [9], we find modest improvements - R2R validation unseen success increases by roughly 1.0 and 0.6 points respectively. This translates to an improvement of 2 points in success on the test set for our modified DUET model. On REVERIE validation unseen, we improve DUET performance by 1.3 points success and 0.82 points grounding success. Our results suggest that including explicit visual depictions of referred objects can improve performance on vision-and-language navigation tasks." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 55, + 620, + 268, + 631 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 620, + 268, + 631 + ], + "spans": [ + { + "bbox": [ + 55, + 620, + 268, + 631 + ], + "type": "text", + "content": "Contributions. We summarize our contributions as:" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 58, + 638, + 295, + 714 + ], + "type": "list", + "angle": 0, + "index": 7, + "blocks": [ + { + "bbox": [ + 58, + 638, + 295, + 686 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 58, + 638, + 295, + 686 + ], + "spans": [ + { + "bbox": [ + 58, + 638, + 295, + 686 + ], + "type": "text", + "content": "- We develop a pipeline for generating visual imaginations from navigation instructions and synthesize the R2R-Imagine dataset to enable studying the impact of text-to-image models on VLN agents." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 58, + 689, + 295, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 58, + 689, + 295, + 714 + ], + "spans": [ + { + "bbox": [ + 58, + 689, + 295, + 714 + ], + "type": "text", + "content": "- We propose an agent-agnostic method to incorporate visual imaginations into existing VLN agents and show" + } + ] + } + ], + "index": 6 + } + ], + "sub_type": "text" + }, + { + "bbox": [ + 326, + 72, + 553, + 95 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 326, + 72, + 553, + 95 + ], + "spans": [ + { + "bbox": [ + 326, + 72, + 553, + 95 + ], + "type": "text", + "content": "improved performance for HAMT [8] and DUET [9] across R2R and REVERIE." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 317, + 100, + 553, + 124 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 100, + 553, + 124 + ], + "spans": [ + { + "bbox": [ + 317, + 100, + 553, + 124 + ], + "type": "text", + "content": "- We provide ablations of modeling decisions to characterize the design space for adding visual imaginations." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 314, + 139, + 400, + 152 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 314, + 139, + 400, + 152 + ], + "spans": [ + { + "bbox": [ + 314, + 139, + 400, + 152 + ], + "type": "text", + "content": "2. Related Work" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 313, + 160, + 553, + 483 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 160, + 553, + 483 + ], + "spans": [ + { + "bbox": [ + 313, + 160, + 553, + 483 + ], + "type": "text", + "content": "Vision-and-Language Navigation (VLN). Navigating agents that can operate in novel environments using freeform natural language lends itself as personal assistants, field-robots and search-and-rescue agents. Development of VLN agents has been facilitated by photorealistic simulators such as Matterport3D [7] and paired language-action datasets such as Room-2-Room (R2R) [4] and REVERIE [33]. R2R provides fine-grained instructions conveying low-level directives such as \"Go left, then right at the hall\". REVERIE specifies coarse-grained instructions conveying broad high-level goals such as \"Adjust the picture by the lamp in the hall\". Progress in VLN has broadly ranged from architectural improvements and data scaling techniques. Architectural improvements include evolution from recurrent systems [4] to transformer architectures [8, 9, 17]. HAMT [8] uses a hierarchical transformer architecture to encode spatial and temporal information. DUET [9] maintains a topological map and performs coarse and fine scale encoding over it. Data scaling techniques include augmenting environments [14, 40] using an action-to-text speaker model [12] to generate synthetic instructions, and scaling environments [19, 41] with additional scenes such as HM3D [35] and Gibson [42] in Habitat [37]. Recently, zero-shot or learned integrations of pretrained large language models for VLN [45, 46] is showing promising performance. However, running large models in-the-loop incurs high computational costs." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 313, + 484, + 553, + 640 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 484, + 553, + 640 + ], + "spans": [ + { + "bbox": [ + 313, + 484, + 553, + 640 + ], + "type": "text", + "content": "Closely related to our work, ADAPT [27] augments VLN agents with visual observations related to an instruction. Language-observation pairs are stored from the training set and multiple pairs of matching object/location nouns from instructions are retrieved at test time using CLIP [34] scores. Fundamentally, the ADAPT image base is limited to training environments whereas our imaginations cover an open distribution. Therefore, our method can be more representative of the instruction and visualize novel objects/locations. For instance, a sub-instruction of \"walk into the kitchen with blue walls\" would retrieve images of kitchens from the training environment using ADAPT, but our imaginations would convey a kitchen with blue walls." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 313, + 642, + 553, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 642, + 553, + 713 + ], + "spans": [ + { + "bbox": [ + 313, + 642, + 553, + 713 + ], + "type": "text", + "content": "Another adjacent work, LAD [26] leverages layout dreamer and goal dreamer modules optimized around a topological graph to navigate via coarse-grained instructions. The capability of our method contrasts with LAD inthree primary ways: 1) method generality; Our method is performant across models and datasets, whereas LAD is a" + } + ] + } + ], + "index": 13 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "text", + "content": "3847" + } + ] + } + ], + "index": 14 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 1 + }, + { + "para_blocks": [ + { + "bbox": [ + 54, + 72, + 294, + 144 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 54, + 72, + 294, + 144 + ], + "spans": [ + { + "bbox": [ + 54, + 72, + 294, + 144 + ], + "type": "text", + "content": "full architecture constructed around a particular topological graph in coarse-grained settings, 2) sequential imagination: we imagine a sequence of landmarks along a path to inform sequential decision-making, whereas LAD imagines just the goal, and 3) we perform early fusion of imaginations and language whereas LAD employs late fusion." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 56, + 147, + 298, + 386 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 147, + 298, + 386 + ], + "spans": [ + { + "bbox": [ + 56, + 147, + 298, + 386 + ], + "type": "text", + "content": "Image generation in VLN. Several techniques have proposed generative models to predict future visual observations or entire environments in VLN [21, 24, 25]. Pathdreamer [21] developed a GAN-based model to predict panoramic observations after a navigation action in VLN, conditioned on prior visual observations. Agents utilizing these predicted observations in a fixed-horizon search with a learned trajectory-instruction alignment heuristic achieve non-trivial success compared to simple baselines, but fall short of planning on actual observations and standard VLN models. Unlike our approach, Pathdreamer does not condition on instruction text or explore training agents on generated imagery. Similarly, PanoGen [24] uses a diffusion model to also generate panorama observations, but conditions on image captions while doing so as a form of visual dataset augmentation – replacing Matterport3D observations with synthetic generations with matching semantics. In contrast to our approach, this data augmentation does not provide additional inputs to the VLN models, is not conditioned on instructions, and is not active at inference time." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 55, + 388, + 295, + 519 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 388, + 295, + 519 + ], + "spans": [ + { + "bbox": [ + 55, + 388, + 295, + 519 + ], + "type": "text", + "content": "Related to our work is VLN-Sig [25], which adapts VLN agents to predict quantized features of future visual observations given the instruction and observation history. This supports additional pretraining tasks and an auxiliary loss during downstream VLN finetuning for predicting next-step image semantics. Unlike our approach, VLN-Sig relies on the trained VLN agent to predict future image properties and focuses more on next-step prediction. In contrast, we leverage broad vision-and-language knowledge from text-to-image models and focus generating instruction-referred visual landmarks to improve agent performance." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 55, + 522, + 295, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 522, + 295, + 714 + ], + "spans": [ + { + "bbox": [ + 55, + 522, + 295, + 714 + ], + "type": "text", + "content": "Image generation in robot control. Taking a wider view, several methods for controlling embodied agents have leveraged text-conditioned generative techniques to provide goal specifications in the form of images or videos. For tabletop object rearrangement tasks, many works use text-to-image models to synthesize goals or sub-goals in response to natural language commands, see [6, 13, 20] for a nonexhaustive sample. Others produce denser conditioning, leveraging text-to-video models to generate entire potential robot trajectories [5, 11]. While similar in motivation to our approach, these techniques focus on tabletop manipulation settings which introduce difficult physical challenges but significantly limit the problem to constrained spaces that remain largely observed throughout operation. As such, sub-goals are often edits to existing observations rather than generations of yet-to-be-seen landmarks as in our approach." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 314, + 71, + 397, + 85 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 314, + 71, + 397, + 85 + ], + "spans": [ + { + "bbox": [ + 314, + 71, + 397, + 85 + ], + "type": "text", + "content": "3. Methodology" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 313, + 91, + 553, + 163 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 91, + 553, + 163 + ], + "spans": [ + { + "bbox": [ + 313, + 91, + 553, + 163 + ], + "type": "text", + "content": "Our proposed methodology consists of two components - (1) a visual imagination generation pipeline and (2) a model-agnostic approach to integrating visual imaginations into existing Vision-and-Language Navigation (VLN) agents. Before defining these, we start with establishing relevant notation for the VLN task." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 313, + 166, + 555, + 370 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 166, + 555, + 370 + ], + "spans": [ + { + "bbox": [ + 313, + 166, + 555, + 370 + ], + "type": "text", + "content": "Problem definition. In the standard VLN task [4], an agent is prompted with a natural language navigation instruction and must make a sequence of actions to navigate along the described trajectory based on visual observations. We denote the " + }, + { + "bbox": [ + 313, + 166, + 555, + 370 + ], + "type": "inline_equation", + "content": "L" + }, + { + "bbox": [ + 313, + 166, + 555, + 370 + ], + "type": "text", + "content": "-word instruction as " + }, + { + "bbox": [ + 313, + 166, + 555, + 370 + ], + "type": "inline_equation", + "content": "W = (w_{1}, w_{2}, \\dots, w_{L})" + }, + { + "bbox": [ + 313, + 166, + 555, + 370 + ], + "type": "text", + "content": " and the visual observation at time step " + }, + { + "bbox": [ + 313, + 166, + 555, + 370 + ], + "type": "inline_equation", + "content": "t" + }, + { + "bbox": [ + 313, + 166, + 555, + 370 + ], + "type": "text", + "content": " as a panoramic observation " + }, + { + "bbox": [ + 313, + 166, + 555, + 370 + ], + "type": "inline_equation", + "content": "O_{t}" + }, + { + "bbox": [ + 313, + 166, + 555, + 370 + ], + "type": "text", + "content": " consisting of " + }, + { + "bbox": [ + 313, + 166, + 555, + 370 + ], + "type": "inline_equation", + "content": "K" + }, + { + "bbox": [ + 313, + 166, + 555, + 370 + ], + "type": "text", + "content": " single view images " + }, + { + "bbox": [ + 313, + 166, + 555, + 370 + ], + "type": "inline_equation", + "content": "O_{t} = (I_{1}^{t}, \\dots, I_{K}^{t})" + }, + { + "bbox": [ + 313, + 166, + 555, + 370 + ], + "type": "text", + "content": ". Following prior work, we consider " + }, + { + "bbox": [ + 313, + 166, + 555, + 370 + ], + "type": "inline_equation", + "content": "K = 36" + }, + { + "bbox": [ + 313, + 166, + 555, + 370 + ], + "type": "text", + "content": " corresponding to images from 12 headings and 3 pitches. Amongst the single view images, a subset " + }, + { + "bbox": [ + 313, + 166, + 555, + 370 + ], + "type": "inline_equation", + "content": "\\mathcal{A}_{t} \\subset O_{t}" + }, + { + "bbox": [ + 313, + 166, + 555, + 370 + ], + "type": "text", + "content": " correspond to navigable directions. Given the instruction " + }, + { + "bbox": [ + 313, + 166, + 555, + 370 + ], + "type": "inline_equation", + "content": "W" + }, + { + "bbox": [ + 313, + 166, + 555, + 370 + ], + "type": "text", + "content": " and history of observations " + }, + { + "bbox": [ + 313, + 166, + 555, + 370 + ], + "type": "inline_equation", + "content": "O_{\\leq t}" + }, + { + "bbox": [ + 313, + 166, + 555, + 370 + ], + "type": "text", + "content": " until time " + }, + { + "bbox": [ + 313, + 166, + 555, + 370 + ], + "type": "inline_equation", + "content": "t" + }, + { + "bbox": [ + 313, + 166, + 555, + 370 + ], + "type": "text", + "content": ", the agent selects an action " + }, + { + "bbox": [ + 313, + 166, + 555, + 370 + ], + "type": "inline_equation", + "content": "a_{t} \\in \\mathcal{A}_{t} \\cup \\{\\text{stop}\\}" + }, + { + "bbox": [ + 313, + 166, + 555, + 370 + ], + "type": "text", + "content": " and accordingly either moves or terminates the episode. Generally, the VLN agent is parameterized as a policy network " + }, + { + "bbox": [ + 313, + 166, + 555, + 370 + ], + "type": "inline_equation", + "content": "\\pi_{\\theta}(a_{t} \\mid W, O_{\\leq t})" + }, + { + "bbox": [ + 313, + 166, + 555, + 370 + ], + "type": "text", + "content": " with parameters " + }, + { + "bbox": [ + 313, + 166, + 555, + 370 + ], + "type": "inline_equation", + "content": "\\theta" + }, + { + "bbox": [ + 313, + 166, + 555, + 370 + ], + "type": "text", + "content": " learned via general pretraining and a combination of imitation learning and reinforcement learning [8, 9]." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 313, + 379, + 487, + 392 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 379, + 487, + 392 + ], + "spans": [ + { + "bbox": [ + 313, + 379, + 487, + 392 + ], + "type": "text", + "content": "3.1. Generating Visual Imaginations" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 313, + 397, + 555, + 529 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 397, + 555, + 529 + ], + "spans": [ + { + "bbox": [ + 313, + 397, + 555, + 529 + ], + "type": "text", + "content": "To generate visual imaginations of potential landmarks, we consider the structure of navigation instructions. Typically, these consist of a sequence of sub-instructions describing intermediate navigation steps, which may or may not refer to visual landmarks. For instance, a sub-instruction \"go past the couch\" implies a visual landmark while \"go straight then left\" does not. We use FG-R2R [15] to segment instructions, resulting in an average of 3.66 sub-instructions per instruction for the R2R[4] training set. For a given instruction " + }, + { + "bbox": [ + 313, + 397, + 555, + 529 + ], + "type": "inline_equation", + "content": "W" + }, + { + "bbox": [ + 313, + 397, + 555, + 529 + ], + "type": "text", + "content": " decomposed into " + }, + { + "bbox": [ + 313, + 397, + 555, + 529 + ], + "type": "inline_equation", + "content": "m" + }, + { + "bbox": [ + 313, + 397, + 555, + 529 + ], + "type": "text", + "content": " segments, we denote these as sub-instructions " + }, + { + "bbox": [ + 313, + 397, + 555, + 529 + ], + "type": "inline_equation", + "content": "S = (S_0, \\dots, S_m)" + }, + { + "bbox": [ + 313, + 397, + 555, + 529 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 313, + 532, + 556, + 663 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 532, + 556, + 663 + ], + "spans": [ + { + "bbox": [ + 313, + 532, + 556, + 663 + ], + "type": "text", + "content": "Sub-instruction filtering scheme. As sub-instructions may not always contain references to visual landmarks, we filter sub-instructions prior to image generation. First, we ignore any sub-instruction that lacks a noun phrase using Spacy [2] and then noun phrases are further filtered by a manual blacklist to avoid uninformative landmark generation. For example, we exclude detected noun phrases rooted on non-visual terms like counts ('one'), directions ('left'), or ambiguous object pronouns (\"it\") as these lack specificity to produce meaningful visual landmarks. After filtering, we denote the remaining sub-instructions " + }, + { + "bbox": [ + 313, + 532, + 556, + 663 + ], + "type": "inline_equation", + "content": "S' \\subset S" + }, + { + "bbox": [ + 313, + 532, + 556, + 663 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 313, + 665, + 556, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 665, + 556, + 715 + ], + "spans": [ + { + "bbox": [ + 313, + 665, + 556, + 715 + ], + "type": "text", + "content": "Visual imagination generation. The filtered subinstructions are passed through a text conditioned diffusion model as illustrated in Fig. 3. We use SDXL [32] as the diffusion model with positive and negative prompts cho" + } + ] + } + ], + "index": 10 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "text", + "content": "3848" + } + ] + } + ], + "index": 11 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 2 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 56, + 72, + 555, + 159 + ], + "blocks": [ + { + "bbox": [ + 56, + 72, + 555, + 159 + ], + "lines": [ + { + "bbox": [ + 56, + 72, + 555, + 159 + ], + "spans": [ + { + "bbox": [ + 56, + 72, + 555, + 159 + ], + "type": "image", + "image_path": "4889dd68ffad037d0fd468b4bf220b92df48e6af134f01e0b8a25d397ab8370b.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 54, + 167, + 555, + 236 + ], + "lines": [ + { + "bbox": [ + 54, + 167, + 555, + 236 + ], + "spans": [ + { + "bbox": [ + 54, + 167, + 555, + 236 + ], + "type": "text", + "content": "Figure 2. An overview of our approach. (Left) Imaginations generated using valid sub-instructions from an instruction as determined by our filtering scheme are first passed to a pre-trained ViT to obtain feature vectors. A type embedding " + }, + { + "bbox": [ + 54, + 167, + 555, + 236 + ], + "type": "inline_equation", + "content": "t_{Im}" + }, + { + "bbox": [ + 54, + 167, + 555, + 236 + ], + "type": "text", + "content": " for imagination modality is then added to the features which are encoded using a 3 layer MLP to obtain imagination embeddings " + }, + { + "bbox": [ + 54, + 167, + 555, + 236 + ], + "type": "inline_equation", + "content": "h_i" + }, + { + "bbox": [ + 54, + 167, + 555, + 236 + ], + "type": "text", + "content": ". (Right) To integrate imagination modality to a VLN agent, the imagination embeddings " + }, + { + "bbox": [ + 54, + 167, + 555, + 236 + ], + "type": "inline_equation", + "content": "h_i" + }, + { + "bbox": [ + 54, + 167, + 555, + 236 + ], + "type": "text", + "content": " are concatenated with instruction embeddings " + }, + { + "bbox": [ + 54, + 167, + 555, + 236 + ], + "type": "inline_equation", + "content": "t_i" + }, + { + "bbox": [ + 54, + 167, + 555, + 236 + ], + "type": "text", + "content": " that are encoded using a text encoder " + }, + { + "bbox": [ + 54, + 167, + 555, + 236 + ], + "type": "inline_equation", + "content": "f_T(W)" + }, + { + "bbox": [ + 54, + 167, + 555, + 236 + ], + "type": "text", + "content": ". The concatenated imagination-text embeddings are passed to the VLN agent's cross-modal encoder " + }, + { + "bbox": [ + 54, + 167, + 555, + 236 + ], + "type": "inline_equation", + "content": "f_X" + }, + { + "bbox": [ + 54, + 167, + 555, + 236 + ], + "type": "text", + "content": " along with visual embeddings to predict a distribution over the agent's action space." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "bbox": [ + 59, + 243, + 293, + 254 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 59, + 243, + 293, + 254 + ], + "spans": [ + { + "bbox": [ + 59, + 243, + 293, + 254 + ], + "type": "text", + "content": "Go down the stairs and then turn right. Go stand behind the easel by a unicycle." + } + ] + } + ], + "index": 2 + }, + { + "type": "image", + "bbox": [ + 59, + 257, + 293, + 339 + ], + "blocks": [ + { + "bbox": [ + 59, + 257, + 293, + 339 + ], + "lines": [ + { + "bbox": [ + 59, + 257, + 293, + 339 + ], + "spans": [ + { + "bbox": [ + 59, + 257, + 293, + 339 + ], + "type": "image", + "image_path": "0582747fc800f4c828ddbcb37d0f10ef039bda15bd5c818bfe8d3d2ddf519df1.jpg" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 55, + 347, + 296, + 403 + ], + "lines": [ + { + "bbox": [ + 55, + 347, + 296, + 403 + ], + "spans": [ + { + "bbox": [ + 55, + 347, + 296, + 403 + ], + "type": "text", + "content": "Figure 3. Example instruction segmentation, filtering, and image generation. Instructions are segmented to sub-instructions leveraging FG-R2R [15] and then filtered to remove phrases referring to uninformative nouns (e.g., \"right\"). We produce visual imaginations using SDXL [32] for remaining sub-instructions." + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_caption" + } + ], + "index": 3 + }, + { + "bbox": [ + 55, + 425, + 296, + 557 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 425, + 296, + 557 + ], + "spans": [ + { + "bbox": [ + 55, + 425, + 296, + 557 + ], + "type": "text", + "content": "sen to steer generation towards indoor real-estate environments akin to typical environments in Matterport3D [7]. For example, providing \"indoor\" and \"real estate\" as positives prompts and \"humans\" and \"collage\" as negatives. See the supplementary for a full list. For a given instruction, we denote the generated images for valid sub-instructions as " + }, + { + "bbox": [ + 55, + 425, + 296, + 557 + ], + "type": "inline_equation", + "content": "\\mathcal{Z} = \\{Z_i \\mid S_i \\in S'\\}" + }, + { + "bbox": [ + 55, + 425, + 296, + 557 + ], + "type": "text", + "content": ". We colloquially refer to these generated images as imaginations and generate them for the entire R2R dataset. The resulting R2R-Imagine dataset consists of over 41k synthesized imaginations. This amounts to an average of 2.96 imaginations per instruction." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 55, + 564, + 234, + 577 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 564, + 234, + 577 + ], + "spans": [ + { + "bbox": [ + 55, + 564, + 234, + 577 + ], + "type": "text", + "content": "3.2. Integration with Existing Models" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 55, + 582, + 296, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 582, + 296, + 715 + ], + "spans": [ + { + "bbox": [ + 55, + 582, + 296, + 715 + ], + "type": "text", + "content": "We consider augmentation with visual imagination to be a model agnostic approach and integrate it in two prior models - HAMT [8] and DUET [9]. Our integration approach consists of an imagination encoder, an alignment-promoting auxiliary loss, and a finetuning regimen (Fig. 2). Without loss of generality, we consider a decomposition of any VLN agent into an instruction encoder " + }, + { + "bbox": [ + 55, + 582, + 296, + 715 + ], + "type": "inline_equation", + "content": "f_{T}(W)" + }, + { + "bbox": [ + 55, + 582, + 296, + 715 + ], + "type": "text", + "content": ", an observation/history encoder " + }, + { + "bbox": [ + 55, + 582, + 296, + 715 + ], + "type": "inline_equation", + "content": "f_{O}(O_{\\leq t})" + }, + { + "bbox": [ + 55, + 582, + 296, + 715 + ], + "type": "text", + "content": ", and a cross-modal policy network " + }, + { + "bbox": [ + 55, + 582, + 296, + 715 + ], + "type": "inline_equation", + "content": "f_{X}(f_{T}(W), f_{O}(O_{\\leq t}))" + }, + { + "bbox": [ + 55, + 582, + 296, + 715 + ], + "type": "text", + "content": " that combines the encoder outputs to predict a distribution over actions. Note that these may be fairly complex mechanisms and include explicit his" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 313, + 243, + 555, + 268 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 243, + 555, + 268 + ], + "spans": [ + { + "bbox": [ + 313, + 243, + 555, + 268 + ], + "type": "text", + "content": "tory encoding modules separate from encoding the current observation; however, this abstraction is sufficient here." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 313, + 270, + 555, + 342 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 270, + 555, + 342 + ], + "spans": [ + { + "bbox": [ + 313, + 270, + 555, + 342 + ], + "type": "text", + "content": "Imagination encoder and integration. Given an instruction " + }, + { + "bbox": [ + 313, + 270, + 555, + 342 + ], + "type": "inline_equation", + "content": "W" + }, + { + "bbox": [ + 313, + 270, + 555, + 342 + ], + "type": "text", + "content": " with corresponding imaginations " + }, + { + "bbox": [ + 313, + 270, + 555, + 342 + ], + "type": "inline_equation", + "content": "\\mathcal{Z}" + }, + { + "bbox": [ + 313, + 270, + 555, + 342 + ], + "type": "text", + "content": ", we independently encode each imagination " + }, + { + "bbox": [ + 313, + 270, + 555, + 342 + ], + "type": "inline_equation", + "content": "Z_{i}" + }, + { + "bbox": [ + 313, + 270, + 555, + 342 + ], + "type": "text", + "content": " using a pretrained vision transformer (ViT) [10] to produce a " + }, + { + "bbox": [ + 313, + 270, + 555, + 342 + ], + "type": "inline_equation", + "content": "d" + }, + { + "bbox": [ + 313, + 270, + 555, + 342 + ], + "type": "text", + "content": "-dim embedding vector. These are further transformed by a three-layer MLP such that the " + }, + { + "bbox": [ + 313, + 270, + 555, + 342 + ], + "type": "inline_equation", + "content": "i" + }, + { + "bbox": [ + 313, + 270, + 555, + 342 + ], + "type": "text", + "content": "th imagination embedding can be written as" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 373, + 354, + 555, + 369 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 373, + 354, + 555, + 369 + ], + "spans": [ + { + "bbox": [ + 373, + 354, + 555, + 369 + ], + "type": "interline_equation", + "content": "h _ {i} = \\operatorname {M L P} \\left(\\operatorname {V i T} \\left(Z _ {i}\\right) + t _ {I m}\\right) \\tag {1}", + "image_path": "0030a4f3ce68c69be7b776fe69c2baa4cd6aec16e18475389a0a8c32fb22e246.jpg" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 313, + 374, + 554, + 493 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 374, + 554, + 493 + ], + "spans": [ + { + "bbox": [ + 313, + 374, + 554, + 493 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 313, + 374, + 554, + 493 + ], + "type": "inline_equation", + "content": "t_{Im}" + }, + { + "bbox": [ + 313, + 374, + 554, + 493 + ], + "type": "text", + "content": " is the type embedding for imagination modality. As we consider visual imaginations to be proxies for sub-instructions, the full set of imagination embeddings " + }, + { + "bbox": [ + 313, + 374, + 554, + 493 + ], + "type": "inline_equation", + "content": "\\mathcal{H}" + }, + { + "bbox": [ + 313, + 374, + 554, + 493 + ], + "type": "text", + "content": " are concatenated to the text encodings such that our general VLN agent structure now computes " + }, + { + "bbox": [ + 313, + 374, + 554, + 493 + ], + "type": "inline_equation", + "content": "f_{X}([f_{T}(W),\\mathcal{H}],f_{O}(O_{\\leq t}))" + }, + { + "bbox": [ + 313, + 374, + 554, + 493 + ], + "type": "text", + "content": " where " + }, + { + "bbox": [ + 313, + 374, + 554, + 493 + ], + "type": "inline_equation", + "content": "[\\cdot ,\\cdot ]" + }, + { + "bbox": [ + 313, + 374, + 554, + 493 + ], + "type": "text", + "content": " denotes a concatenation. Taking the DUET model as an example, the imagination embeddings are concatenated to the instruction encoding prior to being passed to the coarse-scale and fine-scale cross-modal encoders." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 313, + 495, + 555, + 605 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 495, + 555, + 605 + ], + "spans": [ + { + "bbox": [ + 313, + 495, + 555, + 605 + ], + "type": "text", + "content": "Auxiliary alignment loss. We add an auxiliary loss to align the representations of imaginations and their corresponding sub-instruction's noun phrases. For a given sub-instruction " + }, + { + "bbox": [ + 313, + 495, + 555, + 605 + ], + "type": "inline_equation", + "content": "S_{i}" + }, + { + "bbox": [ + 313, + 495, + 555, + 605 + ], + "type": "text", + "content": ", we denote the mean embedding vector for noun phrase tokens in " + }, + { + "bbox": [ + 313, + 495, + 555, + 605 + ], + "type": "inline_equation", + "content": "S_{i}" + }, + { + "bbox": [ + 313, + 495, + 555, + 605 + ], + "type": "text", + "content": " from the text encoder " + }, + { + "bbox": [ + 313, + 495, + 555, + 605 + ], + "type": "inline_equation", + "content": "f_{T}" + }, + { + "bbox": [ + 313, + 495, + 555, + 605 + ], + "type": "text", + "content": " as " + }, + { + "bbox": [ + 313, + 495, + 555, + 605 + ], + "type": "inline_equation", + "content": "\\bar{S}_{i}" + }, + { + "bbox": [ + 313, + 495, + 555, + 605 + ], + "type": "text", + "content": ". To align representations, we compute a cosine similarity loss between the corresponding imagination " + }, + { + "bbox": [ + 313, + 495, + 555, + 605 + ], + "type": "inline_equation", + "content": "Z_{i}" + }, + { + "bbox": [ + 313, + 495, + 555, + 605 + ], + "type": "text", + "content": " encoded as " + }, + { + "bbox": [ + 313, + 495, + 555, + 605 + ], + "type": "inline_equation", + "content": "h_{i}" + }, + { + "bbox": [ + 313, + 495, + 555, + 605 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 313, + 495, + 555, + 605 + ], + "type": "inline_equation", + "content": "\\bar{S}_{i}" + }, + { + "bbox": [ + 313, + 495, + 555, + 605 + ], + "type": "text", + "content": ". This is averaged over all " + }, + { + "bbox": [ + 313, + 495, + 555, + 605 + ], + "type": "inline_equation", + "content": "N_{Im}" + }, + { + "bbox": [ + 313, + 495, + 555, + 605 + ], + "type": "text", + "content": " imagination-subinstruction pairs " + }, + { + "bbox": [ + 313, + 495, + 555, + 605 + ], + "type": "inline_equation", + "content": "(Z_{i}, S_{i})" + }, + { + "bbox": [ + 313, + 495, + 555, + 605 + ], + "type": "text", + "content": " in a batch," + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 346, + 613, + 555, + 646 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 346, + 613, + 555, + 646 + ], + "spans": [ + { + "bbox": [ + 346, + 613, + 555, + 646 + ], + "type": "interline_equation", + "content": "\\mathcal {L} _ {c o s} = \\frac {1}{N _ {I m}} \\sum_ {\\left(Z _ {i}, S _ {i}\\right) \\in B} \\left(1 - \\frac {h _ {i} \\cdot \\bar {S} _ {i}}{\\| h _ {i} \\| \\| \\bar {S} _ {i} \\|}\\right). \\tag {2}", + "image_path": "b0d512788ebaa9eec9d85718a02301e49a068320f989b4b862979bbd7c27076b.jpg" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 313, + 651, + 554, + 688 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 651, + 554, + 688 + ], + "spans": [ + { + "bbox": [ + 313, + 651, + 554, + 688 + ], + "type": "text", + "content": "During finetuning, we combine this auxiliary loss " + }, + { + "bbox": [ + 313, + 651, + 554, + 688 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{\\mathrm{cos}}" + }, + { + "bbox": [ + 313, + 651, + 554, + 688 + ], + "type": "text", + "content": " with the losses from the base agent with a scaling factor " + }, + { + "bbox": [ + 313, + 651, + 554, + 688 + ], + "type": "inline_equation", + "content": "\\lambda" + }, + { + "bbox": [ + 313, + 651, + 554, + 688 + ], + "type": "text", + "content": " such that the overall loss is " + }, + { + "bbox": [ + 313, + 651, + 554, + 688 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{\\mathrm{base}} + \\lambda \\mathcal{L}_{\\mathrm{cos}}" + }, + { + "bbox": [ + 313, + 651, + 554, + 688 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 313, + 689, + 555, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 689, + 555, + 715 + ], + "spans": [ + { + "bbox": [ + 313, + 689, + 555, + 715 + ], + "type": "text", + "content": "Finetuning regimen. We start with fully-trained checkpoints of prior models as our initial weights. Depending" + } + ] + } + ], + "index": 15 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "text", + "content": "3849" + } + ] + } + ], + "index": 16 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 3 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 58, + 135, + 555, + 328 + ], + "blocks": [ + { + "bbox": [ + 55, + 70, + 555, + 126 + ], + "lines": [ + { + "bbox": [ + 55, + 70, + 555, + 126 + ], + "spans": [ + { + "bbox": [ + 55, + 70, + 555, + 126 + ], + "type": "text", + "content": "Table 1. Comparison of our approach with selected prior work on the R2R dataset. Methods that use additional visual data beyond MP3D [7] are annotated with " + }, + { + "bbox": [ + 55, + 70, + 555, + 126 + ], + "type": "inline_equation", + "content": "\\dagger" + }, + { + "bbox": [ + 55, + 70, + 555, + 126 + ], + "type": "text", + "content": " and are not directly comparable with other approaches. Note that BEVBert* additionally uses panoramic depth images. Adding our visual imagination approach to HAMT and DUET base models (gray rows) leads to improved success rate (SR) and success weighted by inverse path length (SPL) metrics on the val-unseen split. For DUET, this effect also persists on the test set. We highlight these improvements over base models using bold numbers." + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 58, + 135, + 555, + 328 + ], + "lines": [ + { + "bbox": [ + 58, + 135, + 555, + 328 + ], + "spans": [ + { + "bbox": [ + 58, + 135, + 555, + 328 + ], + "type": "table", + "html": "
MethodsValidation SeenValidation UnseenTest Unseen
TLNE↓SR↑SPL↑TLNE↓SR↑SPL↑TLNE↓SR↑SPL↑
PREVALENT [14]10.323.67696510.194.71585310.515.305451
RecBERT [16]11.132.90726812.013.93635712.354.096357
ADAPT [27]10.972.54767212.213.77645812.993.796559
BEVBert* [3]13.562.17817414.552.81756415.873.137362
MARVAL [19]†10.602.99736910.154.06656110.224.186258
ScaleVLN [41]†11.902.16878012.402.34877914.272.738377
HAMT [8]11.152.5175.6172.1811.463.6266.2461.5112.273.936560
HAMT-Imagine (ours)11.302.3677.1673.7211.813.5867.2662.0212.663.896560
DUET [9]12.332.2878.8472.8813.943.3171.5260.4114.733.656959
DUET-Imagine (ours)12.932.1979.973.7514.353.1972.1260.4815.353.527160
", + "image_path": "bb0766c09ee4b25d0c478175cd23422c96ff48a3e26b3a7b56aa403786dfacf5.jpg" + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "table_body" + } + ], + "index": 1 + }, + { + "bbox": [ + 54, + 347, + 294, + 504 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 54, + 347, + 294, + 504 + ], + "spans": [ + { + "bbox": [ + 54, + 347, + 294, + 504 + ], + "type": "text", + "content": "on the base method, these parameters may be the result of multi-stage pretraining and finetuning regimens - e.g., leveraging synthetic PREVALENT [14] data as prescribed in [8]. The imagination encoder is integrated to the base model and finetuned with the R2R-Imagine dataset. To mitigate catastrophic forgetting, we perform our training in three stages. First, we train only the newly introduced imagination encoder MLP along with type embeddings " + }, + { + "bbox": [ + 54, + 347, + 294, + 504 + ], + "type": "inline_equation", + "content": "t^{Im}" + }, + { + "bbox": [ + 54, + 347, + 294, + 504 + ], + "type": "text", + "content": " with the remaining parameters frozen. Then, all the modules are trained jointly with a decreased learning rate for the base model. Finally, all the parameters are trained at a common learning rate. See supplementary materials for details on learning rates and schedules." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 55, + 514, + 136, + 528 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 514, + 136, + 528 + ], + "spans": [ + { + "bbox": [ + 55, + 514, + 136, + 528 + ], + "type": "text", + "content": "4. Experiments" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 55, + 533, + 295, + 606 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 533, + 295, + 606 + ], + "spans": [ + { + "bbox": [ + 55, + 533, + 295, + 606 + ], + "type": "text", + "content": "Using the R2R dataset, we evaluate the impact of visual imaginations on our agent's navigation ability, perform ablation studies to explore the design space, and illustrate qualitative impacts of visual imaginations. Using the REVERIE dataset, we study the generalizability of our technique to a fine-grained instruction setting." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 55, + 612, + 170, + 625 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 612, + 170, + 625 + ], + "spans": [ + { + "bbox": [ + 55, + 612, + 170, + 625 + ], + "type": "text", + "content": "4.1. Experimental setup" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 55, + 629, + 295, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 629, + 295, + 714 + ], + "spans": [ + { + "bbox": [ + 55, + 629, + 295, + 714 + ], + "type": "text", + "content": "R2R dataset. R2R [4] is constructed using Matterport3D [7] simulator. It consists of 90 indoor environments with 10,567 panoramas (each panorama consists of 36 single-view images) which act as nodes in a navigation graph. There are 7,189 trajectories each with 3 instructions and a corresponding ground truth trajectory. The dataset is split into train (4675 trajectories), val-seen (340 trajectories)," + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 313, + 347, + 553, + 383 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 347, + 553, + 383 + ], + "spans": [ + { + "bbox": [ + 313, + 347, + 553, + 383 + ], + "type": "text", + "content": "val-unseen (783 trajectories), and test (1391 trajectories). The train and val-seen split share environments while val-unseen and test have different environments." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 313, + 385, + 554, + 470 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 385, + 554, + 470 + ], + "spans": [ + { + "bbox": [ + 313, + 385, + 554, + 470 + ], + "type": "text", + "content": "R2R-Imagine dataset. The curation process of R2R-Imagine is prescribed in section 3.1. The number of imagination images are 41,558 for train, 3,055 for val-seen, 6,857 for val-unseen, and 12,412 for test. Our imaginations have a resolution of " + }, + { + "bbox": [ + 313, + 385, + 554, + 470 + ], + "type": "inline_equation", + "content": "1024 \\times 1024" + }, + { + "bbox": [ + 313, + 385, + 554, + 470 + ], + "type": "text", + "content": " and a single imagination generation on a single H100 GPU takes 3.2 seconds on average. We intend to release the dataset upon acceptance." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 313, + 472, + 555, + 579 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 472, + 555, + 579 + ], + "spans": [ + { + "bbox": [ + 313, + 472, + 555, + 579 + ], + "type": "text", + "content": "REVERIE dataset. REVERIE [33] is constructed using Matterport3D and consists of high-level goals instead of finer grained step-by-step instructions. These instructions generally convey the target object and its location. The agent has to rely more on exploration relative to R2R. To be successful, agents must reach and identify the referent object. To synthesize imaginations for REVERIE, we treat the entire instruction as a single sub-instruction and follow the same recipe used for R2R." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 313, + 582, + 556, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 582, + 556, + 715 + ], + "spans": [ + { + "bbox": [ + 313, + 582, + 556, + 715 + ], + "type": "text", + "content": "Evaluation metrics. We employ standard VLN metrics [18]. Success rate (SR) is the ratio of instruction-trajectories where the agent stops within a 3 meter radius from the goal and success normalized by inverse path length (SPL) helps distinguish successful episodes with shorter path lengths from longer path lengths. Navigation error (NE) is the distance between an agent's final position and the goal in meters. Trajectory length (TL) is the length of an agent's traversed path in meters. REVERIE [33] additionally defines grounding metrics Remote Grounding Success (RGS) and RGS penalized by path length (RGSPL)." + } + ] + } + ], + "index": 10 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "text", + "content": "3850" + } + ] + } + ], + "index": 11 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 4 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 58, + 113, + 294, + 162 + ], + "blocks": [ + { + "bbox": [ + 55, + 70, + 297, + 105 + ], + "lines": [ + { + "bbox": [ + 55, + 70, + 297, + 105 + ], + "spans": [ + { + "bbox": [ + 55, + 70, + 297, + 105 + ], + "type": "text", + "content": "Table 2. Performance comparison of our approach on REVERIE with baseline. We observe improvements of our method across all metrics on DUET when provided coarse-grained instructions." + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 58, + 113, + 294, + 162 + ], + "lines": [ + { + "bbox": [ + 58, + 113, + 294, + 162 + ], + "spans": [ + { + "bbox": [ + 58, + 113, + 294, + 162 + ], + "type": "table", + "html": "
SR↑SPL↑RGS↑RGSPL↑
DUET46.9833.7332.1523.03
DUET-Imagine (ours)48.2833.7632.9723.25
", + "image_path": "4f11b4027be78f21cffcfb7844175ef8f383f7bbb2cf423cf19e2d85abc6c292.jpg" + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "table_body" + } + ], + "index": 1 + }, + { + "bbox": [ + 55, + 180, + 296, + 349 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 180, + 296, + 349 + ], + "spans": [ + { + "bbox": [ + 55, + 180, + 296, + 349 + ], + "type": "text", + "content": "HAMT and DUET base models. We apply our approach to two baseline representative architectures - HAMT [8] and DUET [9]. HAMT is hierarchical and transformer-based; instructions, observations, and trajectory history are encoded separately using unimodal transformers and then combined in a single cross-modal transformer to predict actions. DUET incrementally forms a topographic map of the environment over time and uses dual fine- and coarse-grained cross-modal attention mechanisms to associate language embeddings with the local observation and global topographic features. The results are then fused to predict actions. For both base models, we inject visual imaginations via concatenation with the text embeddings (Sec. 3.2). We use provided checkpoints for pretrained models." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 55, + 350, + 296, + 460 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 350, + 296, + 460 + ], + "spans": [ + { + "bbox": [ + 55, + 350, + 296, + 460 + ], + "type": "text", + "content": "Implementation details. We follow base agent [8, 9] settings unless specified otherwise. We use a standard off-the-shelf pretrained ViT-B/16 [10] to encode imaginations for both HAMT and DUET. Our VLN-Imagine agents are finetuned for " + }, + { + "bbox": [ + 55, + 350, + 296, + 460 + ], + "type": "inline_equation", + "content": "100\\mathrm{k}" + }, + { + "bbox": [ + 55, + 350, + 296, + 460 + ], + "type": "text", + "content": " iterations in three stages as noted in section 3.2. The finetuning process is carried out using a Tesla V100 GPU with a batch size of 8 and takes " + }, + { + "bbox": [ + 55, + 350, + 296, + 460 + ], + "type": "inline_equation", + "content": "\\sim 1.5" + }, + { + "bbox": [ + 55, + 350, + 296, + 460 + ], + "type": "text", + "content": " days for each agent. For combining the proposed auxiliary loss function with the base model losses, we empirically set " + }, + { + "bbox": [ + 55, + 350, + 296, + 460 + ], + "type": "inline_equation", + "content": "\\lambda = 0.5" + }, + { + "bbox": [ + 55, + 350, + 296, + 460 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 55, + 464, + 113, + 477 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 464, + 113, + 477 + ], + "spans": [ + { + "bbox": [ + 55, + 464, + 113, + 477 + ], + "type": "text", + "content": "4.2. Results" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 55, + 482, + 296, + 531 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 482, + 296, + 531 + ], + "spans": [ + { + "bbox": [ + 55, + 482, + 296, + 531 + ], + "type": "text", + "content": "Our experiments demonstrate that providing generated images of instruction landmarks (imaginations) improves the performance of vision-and-language navigation models (Tab. 1). We highlight our key findings below." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 55, + 532, + 295, + 675 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 532, + 295, + 675 + ], + "spans": [ + { + "bbox": [ + 55, + 532, + 295, + 675 + ], + "type": "text", + "content": "VLN agents perform better with imaginations. VLN agents equipped with imaginations perform better than VLN agents without by 0.6-1.0 SR and up to 0.5 SPL on R2R val-unseen (Tab. 1). This improvement coincides with lower navigation error (NE) across both models and all dataset splits. We contextualize these results relative to agent architectures that are representative of the VLN literature. We additionally report a distinct family of techniques that scale training environments to boost performance by reducing overfitting - MARVAL [19] and ScaleVLN [41]. We consider approaches of this class orthogonal to our study and instead use visual data only from Matterport3D [7]." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 55, + 677, + 296, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 677, + 296, + 715 + ], + "spans": [ + { + "bbox": [ + 55, + 677, + 296, + 715 + ], + "type": "text", + "content": "Imaginations are useful in coarse-grained instruction navigation. Our method is performant when natural language instructions convey high-level objectives (REVERIE," + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 313, + 70, + 555, + 138 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 70, + 555, + 138 + ], + "spans": [ + { + "bbox": [ + 313, + 70, + 555, + 138 + ], + "type": "text", + "content": "Table 3. Role of imaginations; HAMT-Imagine. Null imaginations refers to testing with no imaginations, wrong imaginations refers to testing with imaginations sampled from different instructions. We observe aligned imaginations are important to navigation performance. In addition, results with no imaginations are better than baseline implying a regularization effect during training." + } + ] + } + ], + "index": 8 + }, + { + "type": "table", + "bbox": [ + 321, + 145, + 548, + 220 + ], + "blocks": [ + { + "bbox": [ + 321, + 145, + 548, + 220 + ], + "lines": [ + { + "bbox": [ + 321, + 145, + 548, + 220 + ], + "spans": [ + { + "bbox": [ + 321, + 145, + 548, + 220 + ], + "type": "table", + "html": "
SR↑SPL↑
Baseline (HAMT)66.2461.51
Null imaginations66.9261.89
Wrong imaginations66.2461.02
Correct imaginations (ours)67.2662.02
", + "image_path": "2302bfb6aa8afb5c39fabe6aff19ecc50463df477421a426597bfdd468998fab.jpg" + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "table_body" + } + ], + "index": 9 + }, + { + "type": "table", + "bbox": [ + 321, + 304, + 548, + 367 + ], + "blocks": [ + { + "bbox": [ + 313, + 229, + 555, + 298 + ], + "lines": [ + { + "bbox": [ + 313, + 229, + 555, + 298 + ], + "spans": [ + { + "bbox": [ + 313, + 229, + 555, + 298 + ], + "type": "text", + "content": "Table 4. Sequential vs. goal imagination; HAMT-Imagine. We study the effect of providing just the imagination belonging to the final valid sub-instruction relative to full valid imaginations. We observe providing just the goal imagination improves performance significantly compared to baseline but shows lower performance relative to providing full imaginations." + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 321, + 304, + 548, + 367 + ], + "lines": [ + { + "bbox": [ + 321, + 304, + 548, + 367 + ], + "spans": [ + { + "bbox": [ + 321, + 304, + 548, + 367 + ], + "type": "table", + "html": "
SR↑SPL↑
Baseline (HAMT)66.2461.51
Goal imagination66.7961.58
Full imaginations (ours)67.2662.02
", + "image_path": "1083cb207d846927c2225f13979895166d468a8a0bfb316c011c379c72dd36b8.jpg" + } + ] + } + ], + "index": 11, + "angle": 0, + "type": "table_body" + } + ], + "index": 11 + }, + { + "bbox": [ + 313, + 386, + 555, + 447 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 386, + 555, + 447 + ], + "spans": [ + { + "bbox": [ + 313, + 386, + 555, + 447 + ], + "type": "text", + "content": "Tab. 2); navigation performance improves by 1.3 SR and grounding performance improves by 0.82 RGS. We hypothesize that imaginations not only aid in reaching the goal but also in identifying and grounding the target concept amongst others objects in the scene." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 313, + 449, + 556, + 604 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 449, + 556, + 604 + ], + "spans": [ + { + "bbox": [ + 313, + 449, + 556, + 604 + ], + "type": "text", + "content": "VLN agents benefit from both imagination training and inference. When nullifying imaginations at test-time, VLN-Imagine agents still surpass baseline performance, but to a lesser extent (Tab. 3). We implement this nullification by setting the imagination attention masks to zero in the cross-modal encoder. We hypothesize that imagination-based training provides a regularizing effect on the overall VLN agent during fine-tuning, revealing a research avenue for strong and performant test-time model inference. We further find that imaginations must be aligned with the instruction to provide maximal benefit; when fed with imaginations randomly-sampled from the R2R-Imagine dataset, VLN-Imagine agents perform worse than baseline (Tab. 3)." + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 313, + 605, + 556, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 605, + 556, + 715 + ], + "spans": [ + { + "bbox": [ + 313, + 605, + 556, + 715 + ], + "type": "text", + "content": "Sequential imagination outperforms goal imagination. Sequential refers to imaginations generated over multiple sub-instructions (our method). Goal imagination refers to imagining just the final non-filtered sub-instruction. We show these results in Tab. 4 – while goal imagination outperforms the baseline by 0.5 SR, sequential imagination outperforms the baseline by 1.0 SR. Goal imagination may be particularly useful in guiding the agent to stop at the destination. While distinct from our setting, such goal imagi" + } + ] + } + ], + "index": 14 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 748, + 315, + 758 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 315, + 758 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 315, + 758 + ], + "type": "text", + "content": "3851" + } + ] + } + ], + "index": 15 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 5 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 59, + 124, + 293, + 186 + ], + "blocks": [ + { + "bbox": [ + 55, + 70, + 297, + 117 + ], + "lines": [ + { + "bbox": [ + 55, + 70, + 297, + 117 + ], + "spans": [ + { + "bbox": [ + 55, + 70, + 297, + 117 + ], + "type": "text", + "content": "Table 5. Comparison between ViT models. We train both HAMT-Imagine and DUET-Imagine agents with a ViT finetuned in-the-loop with HAMT and an off-the-shelf pretrained ViT. We observe no consistent effects across the both agents." + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 59, + 124, + 293, + 186 + ], + "lines": [ + { + "bbox": [ + 59, + 124, + 293, + 186 + ], + "spans": [ + { + "bbox": [ + 59, + 124, + 293, + 186 + ], + "type": "table", + "html": "
ViT fine-tunedViT off-the-shelf
SR↑SPL↑SR↑SPL↑
HAMT-Imagine67.3161.567.2662.02
DUET-Imagine72.5860.2972.1260.48
", + "image_path": "cd8705c5d24685f08e019bc71901cb98b1d5ce793fb1d692d7aeeb94d2778ada.jpg" + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "table_body" + } + ], + "index": 1 + }, + { + "type": "table", + "bbox": [ + 63, + 262, + 290, + 324 + ], + "blocks": [ + { + "bbox": [ + 55, + 198, + 296, + 254 + ], + "lines": [ + { + "bbox": [ + 55, + 198, + 296, + 254 + ], + "spans": [ + { + "bbox": [ + 55, + 198, + 296, + 254 + ], + "type": "text", + "content": "Table 6. Loss ablation. We study the effect of negative samples in our alignment loss by comparing our approach of using cosine loss with InfoNCE [28] loss. In addition, we train the agent with no auxiliary loss represented by \"No loss\". We do not observe a significant difference between contrastive and cosine losses." + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 63, + 262, + 290, + 324 + ], + "lines": [ + { + "bbox": [ + 63, + 262, + 290, + 324 + ], + "spans": [ + { + "bbox": [ + 63, + 262, + 290, + 324 + ], + "type": "table", + "html": "
SR↑SPL↑
No loss66.7561.60
LInfoNCE67.1861.84
Lcos(ours)67.2662.02
", + "image_path": "fc2c74a05cf3a9b184d167f53148ee3117c20e60bc35edfdafa8612d8a540897.jpg" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "table_body" + } + ], + "index": 3 + }, + { + "bbox": [ + 55, + 345, + 295, + 370 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 345, + 295, + 370 + ], + "spans": [ + { + "bbox": [ + 55, + 345, + 295, + 370 + ], + "type": "text", + "content": "nation has parallels to ImageNav [22], a visual navigation task that conveys goals solely using exemplar images." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 55, + 372, + 295, + 468 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 372, + 295, + 468 + ], + "spans": [ + { + "bbox": [ + 55, + 372, + 295, + 468 + ], + "type": "text", + "content": "General-purpose vision encoders are sufficient imagination encoders. We compare encoding imaginations with off-the-shelf ViT [10] vs. ViT fine-tuned on R2R navigation episodes [8]. We find comparable performance (Tab. 5) and choose to employ off-the-shelf ViT in our final model. This is motivated to maximize generality and to avoid potential issues relating to catastrophic forgetting. For example, generalization of frozen ViT is demonstrated in Zeng et al. [44]." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 55, + 471, + 295, + 616 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 471, + 295, + 616 + ], + "spans": [ + { + "bbox": [ + 55, + 471, + 295, + 616 + ], + "type": "text", + "content": "Aligning imaginations to instructions benefits VLN-imagine agents. Aligning the imagination embedding to the instruction embedding as an auxiliary loss leads to a 0.5 SR and a 0.4 SPL increase over no auxiliary loss (Tab. 6). We also consider formulating this loss as contrastive - do negative samples enable a stronger alignment and thus better downstream performance? We find that an InfoNCE loss [28] yields no significant difference relative to our cosine similarity loss (Eq. 2). We construct negative samples using mean noun phrase embeddings from imaginations of other instructions in the batch. The overall loss is " + }, + { + "bbox": [ + 55, + 471, + 295, + 616 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{\\mathrm{base}} + \\lambda \\mathcal{L}_{\\mathrm{InfoNCE}}" + }, + { + "bbox": [ + 55, + 471, + 295, + 616 + ], + "type": "text", + "content": " where we set " + }, + { + "bbox": [ + 55, + 471, + 295, + 616 + ], + "type": "inline_equation", + "content": "\\lambda" + }, + { + "bbox": [ + 55, + 471, + 295, + 616 + ], + "type": "text", + "content": " to 0.2 empirically." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 55, + 618, + 295, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 618, + 295, + 715 + ], + "spans": [ + { + "bbox": [ + 55, + 618, + 295, + 715 + ], + "type": "text", + "content": "Imaginations correctly represent noun phrases in a subinstruction. To evaluate fidelity of our imaginations, we run an open-vocabulary object detector (LangSAM [1]) on our imaginations to test if the noun phrases contained in the corresponding sub-instruction from R2R are detected successfully. We see in Tab. 7 that at least one noun phrase is detected in our imaginations for " + }, + { + "bbox": [ + 55, + 618, + 295, + 715 + ], + "type": "inline_equation", + "content": "98.78\\%" + }, + { + "bbox": [ + 55, + 618, + 295, + 715 + ], + "type": "text", + "content": " of sub-instructions and all noun phrases are detected for " + }, + { + "bbox": [ + 55, + 618, + 295, + 715 + ], + "type": "inline_equation", + "content": "94.99\\%" + } + ] + } + ], + "index": 7 + }, + { + "type": "table", + "bbox": [ + 315, + 134, + 555, + 213 + ], + "blocks": [ + { + "bbox": [ + 313, + 70, + 555, + 127 + ], + "lines": [ + { + "bbox": [ + 313, + 70, + 555, + 127 + ], + "spans": [ + { + "bbox": [ + 313, + 70, + 555, + 127 + ], + "type": "text", + "content": "Table 7. Open vocabulary object detection accuracy (\\%) of imaginations to noun phrases from sub-instructions. Our imaginations represent the noun phrases with a high accuracy. In val-seen, in " + }, + { + "bbox": [ + 313, + 70, + 555, + 127 + ], + "type": "inline_equation", + "content": "98.43\\%" + }, + { + "bbox": [ + 313, + 70, + 555, + 127 + ], + "type": "text", + "content": " of sub-instructions, at least 1 noun phrase is detected and in " + }, + { + "bbox": [ + 313, + 70, + 555, + 127 + ], + "type": "inline_equation", + "content": "94.51\\%" + }, + { + "bbox": [ + 313, + 70, + 555, + 127 + ], + "type": "text", + "content": " of sub-instructions, all noun phrases are detected." + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 315, + 134, + 555, + 213 + ], + "lines": [ + { + "bbox": [ + 315, + 134, + 555, + 213 + ], + "spans": [ + { + "bbox": [ + 315, + 134, + 555, + 213 + ], + "type": "table", + "html": "
Sub-instructionsNoun Phrase Detection (%)
≥1 detectedAll detected
Val-seen298598.4394.51
Val-unseen670098.9995.34
Test1207598.9395.12
", + "image_path": "c06c22d2b57ce9fd3fbf66a02641237eb4b2ad7e1e4db1bc6bee297ba28c211c.jpg" + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "table_body" + } + ], + "index": 9 + }, + { + "bbox": [ + 313, + 233, + 555, + 258 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 233, + 555, + 258 + ], + "spans": [ + { + "bbox": [ + 313, + 233, + 555, + 258 + ], + "type": "text", + "content": "of sub-instructions. We conclude our imaginations provide a high coverage of relevant concepts across splits." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 313, + 265, + 456, + 277 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 265, + 456, + 277 + ], + "spans": [ + { + "bbox": [ + 313, + 265, + 456, + 277 + ], + "type": "text", + "content": "4.3. Qualitative visualizations" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 313, + 282, + 554, + 330 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 282, + 554, + 330 + ], + "spans": [ + { + "bbox": [ + 313, + 282, + 554, + 330 + ], + "type": "text", + "content": "In this section, we provide a qualitative investigation of how imaginations may be used as a pivot between language tokens and visual observations from the environment by examining the attention patterns in HAMT-Imagine." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 313, + 331, + 554, + 461 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 331, + 554, + 461 + ], + "spans": [ + { + "bbox": [ + 313, + 331, + 554, + 461 + ], + "type": "text", + "content": "For a given instruction, we randomly select an imagination / sub-instruction pair and examine the agent's trajectory to identify the first occurrence of the reference object. We then examine the attention distributions between language tokens, the visual imagination, and the panoramic observation. As there are many attention heads across multiple layers in the cross-modal encoder, we structure our study by focusing on those attention heads where sub-instruction tokens are attended to most strongly by their corresponding visual imagination and then examine how the visual imagination attends to observations." + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 313, + 462, + 554, + 569 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 462, + 554, + 569 + ], + "spans": [ + { + "bbox": [ + 313, + 462, + 554, + 569 + ], + "type": "text", + "content": "We then visualize sub-instruction and imagination pairs along with their top attended language tokens and visual observations in Fig. 4. By construction, these are time steps where the referent is visible and the visual imagination is attending to the appropriate noun phrase. As such, our examples primarily demonstrate how the imagination learns to associate with the visual observations. We focus on this aspect as our auxiliary loss already explicitly encourages alignment between visual imaginations and sub-instruction." + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 313, + 570, + 556, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 570, + 556, + 715 + ], + "spans": [ + { + "bbox": [ + 313, + 570, + 556, + 715 + ], + "type": "text", + "content": "From the examples in Fig. 4, we make two remarks. Imaginations attend most strongly to observations of matching semantic concepts. In row 1, the imagination and the retrieved observations all depict a cheetah (either a painting or floor mat). Imagination attention scores mirror the proper disambiguation of concepts. In row 2, the sub-instruction (and thus the imagination) refer to a unicycle. While both a bicycle and unicycle are observed during execution, attention scores are maximized by the observations containing the unicycle. These patterns do not directly imply a causal relation between the attention scores and eventual decision making of the VLN agent. Rather, they illustrate how imag-" + } + ] + } + ], + "index": 15 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "text", + "content": "3852" + } + ] + } + ], + "index": 16 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 6 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 56, + 70, + 555, + 329 + ], + "blocks": [ + { + "bbox": [ + 56, + 70, + 555, + 329 + ], + "lines": [ + { + "bbox": [ + 56, + 70, + 555, + 329 + ], + "spans": [ + { + "bbox": [ + 56, + 70, + 555, + 329 + ], + "type": "table", + "html": "
Sub-instructionImaginationTop Attended TokensTop Attended Observations
And walk towards the door pass by the cheetah paintingche, ##eta, ##h, painting
Go stand behind the easel by a unicycle next to a doorun, ##ic. ##y, ##cle, to
Once in the room, walk to the left of the pool tablepool, door, room
Walk into the arch hallway near the Native American painting on the wallAmerican, Native, candles
", + "image_path": "83b87358d522a73bc97d26a609e9b03519ebeffc31db66d8bd385fb3f80b4813.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "table_body" + } + ], + "index": 0 + }, + { + "bbox": [ + 55, + 338, + 555, + 427 + ], + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 338, + 555, + 427 + ], + "spans": [ + { + "bbox": [ + 55, + 338, + 555, + 427 + ], + "type": "text", + "content": "Figure 4. Qualitative examples showing imaginations as pivots between language and observation images. The first column contains sub-instruction from a random instruction from R2R, the second column contains the imagination generated using the sub-instruction. The third and fourth columns show highest attended language tokens and observation images from an attention head in HAMT's cross-modal transformer at a time step the associated observations are first visible. In the second example (row 2), the sub-instruction references \"unicycle\" which is captured in the imagination along with neighboring nouns \"easel\" and \"door\". We observe that in a head where top attending language tokens to the imagination query are references to nouns associated with the sub-instruction, its top attended observations to the imagination query are images of the same concept (\"unicycle\"). In this example, the imagination of a unicycle is being used to associate language tokens belonging to \"unicycle\" to observations of unicycle hinting at the utility of imaginations in navigation." + } + ] + } + ], + "index": 1, + "type": "text" + }, + { + "bbox": [ + 55, + 448, + 295, + 472 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 448, + 295, + 472 + ], + "spans": [ + { + "bbox": [ + 55, + 448, + 295, + 472 + ], + "type": "text", + "content": "inations can serve as a pivot between language and observations and afford disambiguation of related concepts." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 55, + 486, + 128, + 498 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 486, + 128, + 498 + ], + "spans": [ + { + "bbox": [ + 55, + 486, + 128, + 498 + ], + "type": "text", + "content": "5. Conclusion" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 55, + 507, + 295, + 638 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 507, + 295, + 638 + ], + "spans": [ + { + "bbox": [ + 55, + 507, + 295, + 638 + ], + "type": "text", + "content": "In this work, we show that visual representations of instructions, or visual imaginations, play an additive role in training superior navigating agents for vision-and-language navigation (VLN) tasks. We generate a sequence of visual imaginations depicting sub-instruction landmarks using a text-to-image diffusion model. Our approach is model-agnostic and we integrate these imaginations into VLN agents using an imagination encoder, an auxiliary loss, and a fine-tuning training step. We apply this approach to two representative VLN agents across two datasets and observe performance improvements of around 1 SR and 0.5 SPL." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 55, + 642, + 296, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 642, + 296, + 714 + ], + "spans": [ + { + "bbox": [ + 55, + 642, + 296, + 714 + ], + "type": "text", + "content": "Limitations. Generating and encoding imaginations increases the computational cost of running VLN agents. This is particularly relevant when running agents on-device for real-world robotic deployments. However, we note that this process is performed upfront, unlike techniques that depend on step-wise observations like novel view synthesis (NVS)." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 313, + 448, + 555, + 519 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 448, + 555, + 519 + ], + "spans": [ + { + "bbox": [ + 313, + 448, + 555, + 519 + ], + "type": "text", + "content": "Stepping more broadly, by design, our imaginations are not grounded in the environment. Personalized reasoning, such as the unique naming of objects and locations, cannot be directly imagined in our framework. Life-long learning of persistent groundings is an open question not just in generating imaginations, but in the VLN space at large." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 313, + 533, + 556, + 665 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 533, + 556, + 665 + ], + "spans": [ + { + "bbox": [ + 313, + 533, + 556, + 665 + ], + "type": "text", + "content": "Future directions. A natural question that arises from our research is the exact complementary nature of the roles played by visual imaginations and language instructions. This is a rich area for future work. Similar studies relating language and visual observations have been conducted in [48] and [47]. Our paper demonstrated the fundamental utility of imagination in VLN agents. Beyond this lies interesting extensions, such as how imagination can help bridge the simulation-to-reality (Sim2Real) gap for VLN agents and whether imaginations can unlock the performance of VLN world models through image-image reasoning." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 313, + 677, + 556, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 677, + 556, + 715 + ], + "spans": [ + { + "bbox": [ + 313, + 677, + 556, + 715 + ], + "type": "text", + "content": "Acknowledgments. We thank Zijiao Yang and Robert Yelle for their assistance with simulator issues and technical support for high performance computing infrastructure." + } + ] + } + ], + "index": 8 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "text", + "content": "3853" + } + ] + } + ], + "index": 9 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 7 + }, + { + "para_blocks": [ + { + "bbox": [ + 56, + 72, + 115, + 83 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 72, + 115, + 83 + ], + "spans": [ + { + "bbox": [ + 56, + 72, + 115, + 83 + ], + "type": "text", + "content": "References" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 57, + 91, + 295, + 713 + ], + "type": "list", + "angle": 0, + "index": 13, + "blocks": [ + { + "bbox": [ + 61, + 91, + 294, + 113 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 61, + 91, + 294, + 113 + ], + "spans": [ + { + "bbox": [ + 61, + 91, + 294, + 113 + ], + "type": "text", + "content": "[1] Language segment-anything. https://github.com/luca-medeiros/lang-segment-anything/.7" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 61, + 115, + 295, + 137 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 61, + 115, + 295, + 137 + ], + "spans": [ + { + "bbox": [ + 61, + 115, + 295, + 137 + ], + "type": "text", + "content": "[2] spacy: Industrial-strength natural language processing in python. https://spacy.io/.3" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 62, + 139, + 294, + 192 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 139, + 294, + 192 + ], + "spans": [ + { + "bbox": [ + 62, + 139, + 294, + 192 + ], + "type": "text", + "content": "[3] Dong An, Yuankai Qi, Yangguang Li, Yan Huang, Liang Wang, Tieniu Tan, and Jing Shao. Bevbert: Multimodal map pre-training for language-guided navigation. Proceedings of the IEEE/CVF International Conference on Computer Vision, 2023. 5" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 62, + 195, + 294, + 270 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 195, + 294, + 270 + ], + "spans": [ + { + "bbox": [ + 62, + 195, + 294, + 270 + ], + "type": "text", + "content": "[4] Peter Anderson, Qi Wu, Damien Teney, Jake Bruce, Mark Johnson, Niko Sünderhauf, Ian D. Reid, Stephen Gould, and Anton van den Hengel. Vision-and-language navigation: Interpreting visually-grounded navigation instructions in real environments. 2018 IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 3674-3683, 2017. 2, 3, 5" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 62, + 274, + 294, + 338 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 274, + 294, + 338 + ], + "spans": [ + { + "bbox": [ + 62, + 274, + 294, + 338 + ], + "type": "text", + "content": "[5] Homanga Bharadhwaj, Debidatta Dwibedi, Abhinav Gupta, Shubham Tulsiani, Carl Doersch, Ted Xiao, Dhruv Shah, Fei Xia, Dorsa Sadigh, and Sean Kirmani. Gen2act: Human video generation in novel scenarios enables generalizable robot manipulation. arXiv preprint arXiv:2409.16283, 2024. 3" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 62, + 341, + 294, + 396 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 341, + 294, + 396 + ], + "spans": [ + { + "bbox": [ + 62, + 341, + 294, + 396 + ], + "type": "text", + "content": "[6] Kevin Black, Mitsuhiko Nakamoto, Pranav Atreya, Homer Rich Walke, Chelsea Finn, Aviral Kumar, and Sergey Levine. Zero-shot robotic manipulation with pre-trained image-editing diffusion models. In The Twelfth International Conference on Learning Representations. 3" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 62, + 398, + 294, + 453 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 398, + 294, + 453 + ], + "spans": [ + { + "bbox": [ + 62, + 398, + 294, + 453 + ], + "type": "text", + "content": "[7] Angel Chang, Angela Dai, Thomas Funkhouser, Maciej Halber, Matthias Niebner, Manolis Savva, Shuran Song, Andy Zeng, and Yinda Zhang. Matterport3d: Learning from rgb-d data in indoor environments. In 2017 International Conference on 3D Vision (3DV), pages 667-676, 2017. 2, 4, 5, 6" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 62, + 455, + 294, + 497 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 455, + 294, + 497 + ], + "spans": [ + { + "bbox": [ + 62, + 455, + 294, + 497 + ], + "type": "text", + "content": "[8] Shizhe Chen, Pierre-Louis Guhur, Cordelia Schmid, and Ivan Laptev. History aware multimodal transformer for vision-and-language navigation. In NeurIPS, 2021. 2, 3, 4, 5, 6, 7" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 62, + 500, + 294, + 544 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 500, + 294, + 544 + ], + "spans": [ + { + "bbox": [ + 62, + 500, + 294, + 544 + ], + "type": "text", + "content": "[9] Shizhe Chen, Pierre-Louis Guhur, Makarand Tapaswi, Cordelia Schmid, and Ivan Laptev. Think global, act local: Dual-scale graph transformer for vision-and-language navigation. In CVPR, 2022. 2, 3, 4, 5, 6" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 57, + 545, + 294, + 610 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 545, + 294, + 610 + ], + "spans": [ + { + "bbox": [ + 57, + 545, + 294, + 610 + ], + "type": "text", + "content": "[10] Alexey Dosovitskiy, Lucas Beyer, Alexander Kolesnikov, Dirk Weissenborn, Xiaohua Zhai, Thomas Unterthiner, Mostafa Dehghani, Matthias Minderer, Georg Heigold, Sylvain Gelly, et al. An image is worth 16x16 words: Transformers for image recognition at scale. arXiv preprint arXiv:2010.11929, 2020. 4, 6, 7" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 57, + 613, + 294, + 667 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 613, + 294, + 667 + ], + "spans": [ + { + "bbox": [ + 57, + 613, + 294, + 667 + ], + "type": "text", + "content": "[11] Yilun Du, Sherry Yang, Bo Dai, Hanjun Dai, Ofir Nachum, Josh Tenenbaum, Dale Schuurmans, and Pieter Abbeel. Learning universal policies via text-guided video generation. Advances in Neural Information Processing Systems, 36, 2024. 3" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 57, + 670, + 294, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 670, + 294, + 713 + ], + "spans": [ + { + "bbox": [ + 57, + 670, + 294, + 713 + ], + "type": "text", + "content": "[12] Daniel Fried, Ronghang Hu, Volkan Cirik, Anna Rohrbach, Jacob Andreas, Louis-Philippe Morency, Taylor Berg-Kirkpatrick, Kate Saenko, Dan Klein, and Trevor Darrell. Speaker-follower models for vision-and-language naviga" + } + ] + } + ], + "index": 12 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 317, + 74, + 553, + 713 + ], + "type": "list", + "angle": 0, + "index": 29, + "blocks": [ + { + "bbox": [ + 335, + 74, + 553, + 94 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 335, + 74, + 553, + 94 + ], + "spans": [ + { + "bbox": [ + 335, + 74, + 553, + 94 + ], + "type": "text", + "content": "tion. In Neural Information Processing Systems (NeurIPS), 2018. 2" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 317, + 96, + 553, + 139 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 96, + 553, + 139 + ], + "spans": [ + { + "bbox": [ + 317, + 96, + 553, + 139 + ], + "type": "text", + "content": "[13] Jialu Gao, Kaizhe Hu, Guowei Xu, and Huazhe Xu. Can pre-trained text-to-image models generate visual goals for reinforcement learning? Advances in Neural Information Processing Systems, 36:38297-38310, 2023. 3" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 317, + 140, + 553, + 172 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 140, + 553, + 172 + ], + "spans": [ + { + "bbox": [ + 317, + 140, + 553, + 172 + ], + "type": "text", + "content": "[14] Weituo Hao, Chunyuan Li, Xiujun Li, Lawrence Carin, and Jianfeng Gao. Towards learning a generic agent for vision-and-language navigation via pre-training, 2020. 2, 5" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 317, + 174, + 553, + 226 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 174, + 553, + 226 + ], + "spans": [ + { + "bbox": [ + 317, + 174, + 553, + 226 + ], + "type": "text", + "content": "[15] Yicong Hong, Cristian Rodriguez, Qi Wu, and Stephen Gould. Sub-instruction aware vision-and-language navigation. In Proceedings of the 2020 Conference on Empirical Methods in Natural Language Processing (EMNLP), pages 3360-3376, 2020. 3, 4" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 317, + 228, + 553, + 282 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 228, + 553, + 282 + ], + "spans": [ + { + "bbox": [ + 317, + 228, + 553, + 282 + ], + "type": "text", + "content": "[16] Yicong Hong, Qi Wu, Yuankai Qi, Cristian Rodriguez-Opazo, and Stephen Gould. A recurrent vision-and-language bert for navigation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 1643–1653, 2021. 5" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 317, + 284, + 553, + 337 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 284, + 553, + 337 + ], + "spans": [ + { + "bbox": [ + 317, + 284, + 553, + 337 + ], + "type": "text", + "content": "[17] Yicong Hong, Qi Wu, Yuankai Qi, Cristian Rodriguez-Opazo, and Stephen Gould. A recurrent vision-and-language bert for navigation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 1643–1653, 2021. 2" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 317, + 338, + 553, + 381 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 338, + 553, + 381 + ], + "spans": [ + { + "bbox": [ + 317, + 338, + 553, + 381 + ], + "type": "text", + "content": "[18] Gabriel Ilharco, Vihan Jain, Alexander Ku, Eugene Ie, and Jason Baldridge. General evaluation for instruction conditioned navigation using dynamic time warping. ArXiv, abs/1907.05446, 2019. 5" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 317, + 383, + 553, + 458 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 383, + 553, + 458 + ], + "spans": [ + { + "bbox": [ + 317, + 383, + 553, + 458 + ], + "type": "text", + "content": "[19] Aishwarya Kamath, Peter Anderson, Su Wang, Jing Yu Koh, Alexander Ku, Austin Waters, Yinfei Yang, Jason Baldridge, and Zarana Parekh. A new path: Scaling vision-and-language navigation with synthetic instructions and imitation learning. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 10813-10823, 2023. 2, 5, 6" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 317, + 460, + 553, + 491 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 460, + 553, + 491 + ], + "spans": [ + { + "bbox": [ + 317, + 460, + 553, + 491 + ], + "type": "text", + "content": "[20] Ivan Kapelyukh, Vitalis Vosylius, and Edward Johns. Dallebot: Introducing web-scale diffusion models to robotics. 2023. 3" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 317, + 493, + 553, + 525 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 493, + 553, + 525 + ], + "spans": [ + { + "bbox": [ + 317, + 493, + 553, + 525 + ], + "type": "text", + "content": "[21] Jing Yu Koh, Honglak Lee, Yinfei Yang, Jason Baldridge, and Peter Anderson. Pathdreamer: A world model for indoor navigation, 2021. 3" + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 317, + 526, + 553, + 568 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 526, + 553, + 568 + ], + "spans": [ + { + "bbox": [ + 317, + 526, + 553, + 568 + ], + "type": "text", + "content": "[22] Jacob Krantz, Stefan Lee, Jitendra Malik, Dhruv Batra, and Devendra Singh Chaplot. Instance-specific image goal navigation: Training embodied agents to find object instances, 2022. 7" + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 317, + 571, + 553, + 613 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 571, + 553, + 613 + ], + "spans": [ + { + "bbox": [ + 317, + 571, + 553, + 613 + ], + "type": "text", + "content": "[23] Chongxi Lai, Shinsuke Tanaka, Timothy D. Harris, and Albert K. Lee. Volitional activation of remote place representations with a hippocampal brain-machine interface. Science, 382(6670):566-573, 2023. 2" + } + ] + } + ], + "index": 25 + }, + { + "bbox": [ + 317, + 614, + 553, + 646 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 614, + 553, + 646 + ], + "spans": [ + { + "bbox": [ + 317, + 614, + 553, + 646 + ], + "type": "text", + "content": "[24] Jialu Li and Mohit Bansal. Panogen: Text-conditioned panoramic environment generation for vision-and-language navigation. arxiv, 2023. 3" + } + ] + } + ], + "index": 26 + }, + { + "bbox": [ + 317, + 647, + 553, + 679 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 647, + 553, + 679 + ], + "spans": [ + { + "bbox": [ + 317, + 647, + 553, + 679 + ], + "type": "text", + "content": "[25] Jialu Li and Mohit Bansal. Improving vision-and-language navigation by generating future-view image semantics. CVPR, 2023. 3" + } + ] + } + ], + "index": 27 + }, + { + "bbox": [ + 317, + 681, + 553, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 681, + 553, + 713 + ], + "spans": [ + { + "bbox": [ + 317, + 681, + 553, + 713 + ], + "type": "text", + "content": "[26] Mingxiao Li, Zehao Wang, Tinne Tuytelaars, and Marie-Francine Moens. Layout-aware dreamer for embodied visual referring expression grounding. In Proceedings of the" + } + ] + } + ], + "index": 28 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "text", + "content": "3854" + } + ] + } + ], + "index": 30 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 8 + }, + { + "para_blocks": [ + { + "bbox": [ + 56, + 72, + 295, + 712 + ], + "type": "list", + "angle": 0, + "index": 14, + "blocks": [ + { + "bbox": [ + 77, + 72, + 295, + 94 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 77, + 72, + 295, + 94 + ], + "spans": [ + { + "bbox": [ + 77, + 72, + 295, + 94 + ], + "type": "text", + "content": "AAAI Conference on Artificial Intelligence, pages 1386-1395, 2023. 2" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 56, + 95, + 295, + 149 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 95, + 295, + 149 + ], + "spans": [ + { + "bbox": [ + 56, + 95, + 295, + 149 + ], + "type": "text", + "content": "[27] Bingqian Lin, Yi Zhu, Zicong Chen, Xiwen Liang, Jianzhuang Liu, and Xiaodan Liang. Adapt: Vision-language navigation with modality-aligned action prompts. In 2022 IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 15375-15385, 2022. 2, 5" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 56, + 151, + 294, + 183 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 151, + 294, + 183 + ], + "spans": [ + { + "bbox": [ + 56, + 151, + 294, + 183 + ], + "type": "text", + "content": "[28] Aaron van den Oord, Yazhe Li, and Oriol Vinyals. Representation learning with contrastive predictive coding. arXiv preprint arXiv:1807.03748, 2018. 7" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 56, + 184, + 295, + 216 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 184, + 295, + 216 + ], + "spans": [ + { + "bbox": [ + 56, + 184, + 295, + 216 + ], + "type": "text", + "content": "[29] Allan Paivio. Abstractness, imagery, and meaningfulness in paired-associate learning. Journal of verbal learning and verbal behavior., 4(1), 1965-2. 2" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 56, + 217, + 294, + 249 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 217, + 294, + 249 + ], + "spans": [ + { + "bbox": [ + 56, + 217, + 294, + 249 + ], + "type": "text", + "content": "[30] Allan Paivio. A factor-analytic study of word attributes and verbal learning. Journal of verbal learning and verbal behavior, 7(1), 1968-2." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 56, + 250, + 294, + 293 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 250, + 294, + 293 + ], + "spans": [ + { + "bbox": [ + 56, + 250, + 294, + 293 + ], + "type": "text", + "content": "[31] A. Paivio, P. C. Smythe, and J. C. Yuille. Imagery versus meaningfulness of nouns in paired-associate learning. Canadian Journal of Psychology / Revue Canadienne De Psychologie, 22:427–441, 1968. 2" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 56, + 294, + 294, + 338 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 294, + 294, + 338 + ], + "spans": [ + { + "bbox": [ + 56, + 294, + 294, + 338 + ], + "type": "text", + "content": "[32] Dustin Podell, Zion English, Kyle Lacey, Andreas Blattmann, Tim Dockhorn, Jonas Müller, Joe Penna, and Robin Rombach. Sdxl: Improving latent diffusion models for high-resolution image synthesis, 2023. 2, 3, 4" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 56, + 338, + 295, + 403 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 338, + 295, + 403 + ], + "spans": [ + { + "bbox": [ + 56, + 338, + 295, + 403 + ], + "type": "text", + "content": "[33] Yuankai Qi, Qi Wu, Peter Anderson, Xin Wang, William Yang Wang, Chunhua Shen, and Anton van den Hengel. Reverie: Remote embodied visual referring expression in real indoor environments. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 9982-9991, 2020. 2, 5" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 56, + 404, + 295, + 470 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 404, + 295, + 470 + ], + "spans": [ + { + "bbox": [ + 56, + 404, + 295, + 470 + ], + "type": "text", + "content": "[34] Alec Radford, Jong Wook Kim, Chris Hallacy, Aditya Ramesh, Gabriel Goh, Sandhini Agarwal, Girish Sastry, Amanda Askell, Pamela Mishkin, Jack Clark, et al. Learning transferable visual models from natural language supervision. In International conference on machine learning, pages 8748-8763. PMLR, 2021. 1, 2" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 56, + 471, + 295, + 536 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 471, + 295, + 536 + ], + "spans": [ + { + "bbox": [ + 56, + 471, + 295, + 536 + ], + "type": "text", + "content": "[35] Santhosh K Ramakrishnan, Aaron Gokaslan, Erik Wijmans, Oleksandr Maksymets, Alex Clegg, John Turner, Eric Undersander, Wojciech Galuba, Andrew Westbury, Angel X Chang, et al. Habitat-matterport 3d dataset (hm3d): 1000 large-scale 3d environments for embodied ai. arXiv preprint arXiv:2109.08238, 2021. 2" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 56, + 537, + 295, + 568 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 537, + 295, + 568 + ], + "spans": [ + { + "bbox": [ + 56, + 537, + 295, + 568 + ], + "type": "text", + "content": "[36] Marianne Cumella Reddan, Tor Dessart Wager, and Daniela Schiller. Attenuating neural threat expression with imagination. *Neuron*, 100(4):994–1005.e4, 2018. 2" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 56, + 570, + 295, + 635 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 570, + 295, + 635 + ], + "spans": [ + { + "bbox": [ + 56, + 570, + 295, + 635 + ], + "type": "text", + "content": "[37] Manolis Savva, Abhishek Kadian, Oleksandr Maksymets, Yili Zhao, Erik Wijmans, Bhavana Jain, Julian Straub, Jia Liu, Vladlen Koltun, Jitendra Malik, et al. Habitat: A platform for embodied ai research. In Proceedings of the IEEE/CVF international conference on computer vision, pages 9339-9347, 2019. 2" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 56, + 636, + 295, + 679 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 636, + 295, + 679 + ], + "spans": [ + { + "bbox": [ + 56, + 636, + 295, + 679 + ], + "type": "text", + "content": "[38] Daniel L. Schacter, Donna Rose Addis, and Randy L. Buckner. Remembering the past to imagine the future: the prospective brain. Nature Reviews Neuroscience, 8(9):657-661, 2007. 2" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 56, + 681, + 295, + 712 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 681, + 295, + 712 + ], + "spans": [ + { + "bbox": [ + 56, + 681, + 295, + 712 + ], + "type": "text", + "content": "[39] Roger N. Shepard and Jacqueline Metzler. Mental rotation of three-dimensional objects. Science, 171(3972):701-703, 1971. 2" + } + ] + } + ], + "index": 13 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 316, + 73, + 555, + 573 + ], + "type": "list", + "angle": 0, + "index": 24, + "blocks": [ + { + "bbox": [ + 316, + 73, + 555, + 160 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 73, + 555, + 160 + ], + "spans": [ + { + "bbox": [ + 316, + 73, + 555, + 160 + ], + "type": "text", + "content": "[40] Hao Tan, Licheng Yu, and Mohit Bansal. Learning to navigate unseen environments: Back translation with environmental dropout. In Proceedings of the 2019 Conference of the North American Chapter of the Association for Computational Linguistics: Human Language Technologies, Volume 1 (Long and Short Papers), pages 2610-2621, Minneapolis, Minnesota, 2019. Association for Computational Linguistics. 2" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 316, + 162, + 554, + 205 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 162, + 554, + 205 + ], + "spans": [ + { + "bbox": [ + 316, + 162, + 554, + 205 + ], + "type": "text", + "content": "[41] Zun Wang, Jialu Li, Yicong Hong, Yi Wang, Qi Wu, Mohit Bansal, Stephen Gould, Hao Tan, and Yu Qiao. Scaling data generation in vision-and-language navigation. In ICCV 2023, 2023. 2, 5, 6" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 316, + 206, + 553, + 259 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 206, + 553, + 259 + ], + "spans": [ + { + "bbox": [ + 316, + 206, + 553, + 259 + ], + "type": "text", + "content": "[42] Fei Xia, Amir R. Zamir, Zhi-Yang He, Alexander Sax, Jitendra Malik, and Silvio Savarese. Gibson env: real-world perception for embodied agents. In Computer Vision and Pattern Recognition (CVPR), 2018 IEEE Conference on. IEEE, 2018. 2" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 316, + 262, + 553, + 306 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 262, + 553, + 306 + ], + "spans": [ + { + "bbox": [ + 316, + 262, + 553, + 306 + ], + "type": "text", + "content": "[43] Zijiao Yang, Arjun Majumdar, and Stefan Lee. Behavioral analysis of vision-and-language navigation agents. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 2574-2582, 2023. 1" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 316, + 308, + 553, + 361 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 308, + 553, + 361 + ], + "spans": [ + { + "bbox": [ + 316, + 308, + 553, + 361 + ], + "type": "text", + "content": "[44] Kuo-Hao Zeng, Zichen Zhang, Kiana Ehsani, Rose Hendrix, Jordi Salvador, Alvaro Herrasti, Ross Girshick, Aniruddha Kembhavi, and Luca Weihs. Poliformer: Scaling on-policy rl with transformers results in masterful navigators. In 8th Annual Conference on Robot Learning, 2024. 7" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 316, + 363, + 553, + 407 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 363, + 553, + 407 + ], + "spans": [ + { + "bbox": [ + 316, + 363, + 553, + 407 + ], + "type": "text", + "content": "[45] Gengze Zhou, Yicong Hong, and Qi Wu. Navigpt: Explicit reasoning in vision-and-language navigation with large language models. Proceedings of the AAAI Conference on Artificial Intelligence, 38(7):7641-7649, 2024. 2" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 316, + 408, + 553, + 460 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 408, + 553, + 460 + ], + "spans": [ + { + "bbox": [ + 316, + 408, + 553, + 460 + ], + "type": "text", + "content": "[46] Gengze Zhou, Yicong Hong, Zun Wang, Xin Eric Wang, and Qi Wu. Navigpt-2: Unleashing navigational reasoning capability for large vision-language models. In European Conference on Computer Vision, pages 260–278. Springer, 2025. 2" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 316, + 463, + 553, + 540 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 463, + 553, + 540 + ], + "spans": [ + { + "bbox": [ + 316, + 463, + 553, + 540 + ], + "type": "text", + "content": "[47] Wanrong Zhu, Yuankai Qi, Pradyumna Narayana, Kazoo Sone, Sugato Basu, Xin Wang, Qi Wu, Miguel Eckstein, and William Yang Wang. Diagnosing vision-and-language navigation: What really matters. In Proceedings of the 2022 Conference of the North American Chapter of the Association for Computational Linguistics: Human Language Technologies, pages 5981-5993, 2022. 1, 8" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 316, + 541, + 553, + 573 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 541, + 553, + 573 + ], + "spans": [ + { + "bbox": [ + 316, + 541, + 553, + 573 + ], + "type": "text", + "content": "[48] Wang Zhu, Ishika Singh, Yuan Huang, Robin Jia, and Jesse Thomason. Does vln pretraining work with nonsensical or irrelevant instructions? ArXiv, abs/2311.17280, 2023. 8" + } + ] + } + ], + "index": 23 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "text", + "content": "3855" + } + ] + } + ], + "index": 25 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 9 + } + ], + "_backend": "vlm", + "_version_name": "2.6.4" +} \ No newline at end of file diff --git a/2025/Do We Always Need the Simplicity Bias_ Looking for Optimal Inductive Biases in the Wild/f8b7dd25-2878-4e2f-ac24-ae160cb54c72_content_list.json b/2025/Do We Always Need the Simplicity Bias_ Looking for Optimal Inductive Biases in the Wild/f8b7dd25-2878-4e2f-ac24-ae160cb54c72_content_list.json new file mode 100644 index 0000000000000000000000000000000000000000..845e1f13e32487e379f103ad1280088f9e4fa719 --- /dev/null +++ b/2025/Do We Always Need the Simplicity Bias_ Looking for Optimal Inductive Biases in the Wild/f8b7dd25-2878-4e2f-ac24-ae160cb54c72_content_list.json @@ -0,0 +1,2101 @@ +[ + { + "type": "text", + "text": "Do We Always Need the Simplicity Bias? Looking for Optimal Inductive Biases in the Wild", + "text_level": 1, + "bbox": [ + 248, + 130, + 750, + 176 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Damien Teney \nIdiap Research Institute", + "bbox": [ + 130, + 204, + 321, + 239 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Liangze Jiang EPFL", + "bbox": [ + 359, + 204, + 475, + 237 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Abstract", + "text_level": 1, + "bbox": [ + 248, + 267, + 326, + 282 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Neural architectures tend to fit their data with relatively simple functions. This \"simplicity bias\" is widely regarded as key to their success. This paper explores the limits of this principle. Building on recent findings that the simplicity bias stems from ReLU activations [96], we introduce a method to meta-learn new activation functions and inductive biases better suited to specific tasks.", + "bbox": [ + 89, + 301, + 483, + 409 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Findings. We identify multiple tasks where the simplicity bias is inadequate and ReLUs suboptimal. In these cases, we learn new activation functions that perform better by inducing a prior of higher complexity. Interestingly, these cases correspond to domains where neural networks have historically struggled: tabular data, regression tasks, cases of shortcut learning, and algorithmic grokking tasks. In comparison, the simplicity bias induced by ReLUs proves adequate on image tasks where the best learned activations are nearly identical to ReLUs and GeLUs.", + "bbox": [ + 89, + 412, + 483, + 565 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Implications. Contrary to popular belief, the simplicity bias of ReLU networks is not universally useful. It is near-optimal for image classification, but other inductive biases are sometimes preferable. We showed that activation functions can control these inductive biases, but future tailored architectures might provide further benefits. Advances are still needed to characterize a model's inductive biases beyond \"complexity\", and their adequacy with the data.", + "bbox": [ + 89, + 569, + 483, + 691 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "1. Introduction", + "text_level": 1, + "bbox": [ + 89, + 720, + 220, + 736 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "When and why NNs generalize is yet to be understood. Neural networks (NNs) have proven more effective than other machine learning models. However, we still miss a complete explanation of their generalization abilities. A better understanding could help address failures from shortcut learning [29, 93] to distribution shifts [47, 95], biases, and spurious correlations in language models for example [28, 38, 87]. Understanding conditions for generalization would also enable the design of architectures and data preparation from first principles, rather than trial and error.", + "bbox": [ + 89, + 750, + 482, + 900 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Florin Gogianu\nBitdefender", + "bbox": [ + 513, + 204, + 638, + 237 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Ehsan Abbasnejad \nUniversity of Adelaide", + "bbox": [ + 678, + 204, + 859, + 239 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/f6c3c7fa1ae2b01883bb609bc39722c6311e201422ffb6aaffc6076662c9f41c.jpg", + "image_caption": [ + "Figure 1. (1) We modulate the inductive bias of neural architectures by learning novel activation functions that improve generalization on specific datasets. (2) With this tool, we study the relation between model accuracy and complexity. We identify tasks where the simplicity bias of ReLU architectures is suboptimal." + ], + "image_footnote": [], + "bbox": [ + 514, + 266, + 903, + 377 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "This paper studies inductive biases i.e. the assumptions made by learning algorithms to generalize beyond training data [67].1 A vast literature examines the inductive biases of architectures [14], optimizers [69], losses [42], regularizers [50], etc. The simplicity bias is one aspect of the inductive biases of NNs that makes them fit their training data with simple2 functions [8, 71]. Despite wide belief that the simplicity bias could be due to SGD [7, 40, 60, 92], work on untrained networks showed that it can be explained with architectures alone [15, 31, 64, 99]. ReLUs also seem critical to induce the simplicity bias in typical architectures [96].", + "bbox": [ + 511, + 481, + 906, + 648 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Limits of the simplicity bias. The simplicity bias is an intuitive explanation for the ability to generalize on real-world data. It embodies Occam's razor [66] and assumes that data-generating processes in the real world are simple. Additionally, a prior for simplicity is supported by results in algorithmic information theory [18] stating essentially that \"a bias in the distribution of target functions must be towards low complexity\". However, this only means that simplicity is a good prior on average, but not necessarily the best choice on any task or dataset. This matches the no-free lunch theorem [101] according to which no inductive bias is universally useful. Therefore, this paper asks the following.", + "bbox": [ + 511, + 657, + 908, + 840 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "CVF", + "bbox": [ + 106, + 2, + 181, + 42 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore.", + "bbox": [ + 236, + 0, + 810, + 46 + ], + "page_idx": 0 + }, + { + "type": "page_footnote", + "text": "$^{1}$ Inductive biases can formalized as a prior over the space of functions [65]. $^{2}$ Simplicity can be formalized using Kolmogorov complexity or its approximations, e.g. frequency, compressibility, sensitivity, etc. [17, 18, 37, 96]", + "bbox": [ + 511, + 862, + 903, + 900 + ], + "page_idx": 0 + }, + { + "type": "page_number", + "text": "79", + "bbox": [ + 490, + 945, + 506, + 955 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Are there practical applications of machine learning where the simplicity bias is detrimental? In these cases, what do the optimal inductive biases look like?", + "text_level": 1, + "bbox": [ + 102, + 93, + 470, + 138 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "For example, shortcut learning is one situation where the simplicity bias is already known to be detrimental [85, 93].", + "bbox": [ + 89, + 148, + 480, + 179 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Searching for optimal inductive biases by learning activation functions. Prior work [96] showed that ReLU activations are critical to obtain the simplicity bias in typical architectures. Hence we build a new tool to modulate the simplicity bias by learning dataset-specific activation functions. It uses bi-level optimization and a spline parametrization to learn activations free of any prior, such as constraints of smoothness or monotonicity (unlike prior work [2, 6, 82]). This (1) enables the discovery of entirely new activation functions and inductive biases that improve generalization (Figure 1) and (2) highlights the suboptimality of the simplicity bias by comparing the accuracy and complexity of models with ReLUs vs. learned activations.", + "bbox": [ + 89, + 185, + 483, + 380 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Findings. We examine four domains that we hypothesized to be impaired by the simplicity bias: tabular data, regression tasks, cases of shortcut learning, and algorithmic tasks. Our intuition is that they require learning functions with high sensitivity or sharp transitions. For each domain, we collect existing datasets then train and analyze models without and with learned activation functions. In all cases, we obtain better generalization with dataset-specific activations, and the improvements are attributable to learning higher-complexity solutions. In comparison, this analysis also shows that classical image datasets (MNIST, CIFAR, FASHION-MNIST, SVHN) are extremely well suited to the inductive biases of ReLUs. The best learned activations are then strikingly similar to variants like GeLUs [39].", + "bbox": [ + 89, + 387, + 483, + 598 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Summary of contributions.", + "text_level": 1, + "bbox": [ + 89, + 604, + 284, + 619 + ], + "page_idx": 1 + }, + { + "type": "list", + "sub_type": "text", + "list_items": [ + "- A new method to discover dataset-specific activation functions optimized for generalization.", + "- An examination of $>20$ datasets showing that the simplicity bias of ReLU architectures can be suboptimal. (1) For regression tasks and tabular data, new learned activations greatly improve accuracy by helping learn complex functions. (2) For image classification, the process rediscovers smooth variants of ReLUs, suggesting a near-optimal choice for these popular tasks. (3) In cases of shortcut learning, we show that different learned activations can steer the learning towards different image features. (4) For grokking tasks, new learned activations can eliminate the phenomenon, supporting the explanation as a mismatch between data and architectures. We also measure a positive transferability of learned activations across related tasks.", + "- An analysis showing that improvements with learned activations correlate with the learning of complex functions." + ], + "bbox": [ + 89, + 623, + 483, + 900 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Implications. All cases where the simplicity bias proved suboptimal are in domains where NNs have historically struggled. We now connect them to a common explanation. This implies that architectures tailored to some specific domains may still have a place besides scaling up models and data. Conversely, the suitability of ReLUs to image classification suggests that researchers successfully converged by trial and error to designs well tuned to popular tasks.", + "bbox": [ + 511, + 90, + 905, + 212 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "2. Methods", + "text_level": 1, + "bbox": [ + 513, + 233, + 611, + 248 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "This section introduces tools to analyze trained models and to learn new dataset-specific activation functions.", + "bbox": [ + 511, + 258, + 905, + 289 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "2.1. Visualizing a Model's Function", + "text_level": 1, + "bbox": [ + 511, + 306, + 785, + 321 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "A neural network implements a function $f_{\\theta} : \\mathbb{R}^{d_{\\mathrm{in}}} \\rightarrow \\mathbb{R}^{d_{\\mathrm{out}}}$ of parameters $\\theta$ (weights and biases) that maps an input $x \\in \\mathbb{R}^{d_{\\mathrm{in}}}$ to an output $y \\in \\mathbb{R}^{d_{\\mathrm{out}}}$ . For a regression task, $y \\in \\mathbb{R}$ is the predicted value. For a classification task, $y$ is a vector of logits passed through a sigmoid or softmax to obtain class probabilities. Because $d_{\\mathrm{in}}$ can be large, $f$ can be difficult to visualize and analyze. A workaround is to examine $f$ over 1D or 2D slices of the input space [26, 96]. To obtain a slice in a region of plausible data, we use the training data $\\mathcal{T}$ . For a 1D slice (linear path), we sample $x_1, x_2 \\sim \\mathcal{T}$ then define the path $\\mathbf{X}_{x_1, x_2} = [(1 - \\lambda) x_1 + \\lambda x_2, \\lambda \\in [0,1]]$ . We proceed analogously with three points for a 2D slice. We sample $\\lambda$ regularly in [0,1] such that $\\mathbf{X}$ is a finite sequence of points. Then $f$ is evaluated on these points to give a 1D sequence or 2D grid of values that are convenient to display and analyze (Figure 7c). When $d_{\\mathrm{out}} > 1$ (multi-class task), we examine one random dimension of $f$ 's output at a time.", + "bbox": [ + 511, + 327, + 906, + 585 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "2.2. Measuring a Model's Complexity", + "text_level": 1, + "bbox": [ + 511, + 602, + 803, + 618 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "We wish to quantify the complexity of the function $f$ implemented by a model trained on data $\\mathcal{T}$ . Prior work used Fourier decompositions [26, 96] but this requires a delicate implementation. We found a reliable alternative with the total variation (TV) of $f$ averaged over many 1D paths:", + "bbox": [ + 511, + 623, + 905, + 700 + ], + "page_idx": 1 + }, + { + "type": "equation", + "text": "\n$$\n\\operatorname {T V} (f, \\mathcal {T}) = \\mathbb {E} _ {\\boldsymbol {x} _ {1}, \\boldsymbol {x} _ {2}} \\sim \\mathcal {T} \\int_ {\\boldsymbol {x} _ {1}} ^ {\\boldsymbol {x} _ {2}} \\left| f ^ {\\prime} (\\boldsymbol {x}) \\right| d \\boldsymbol {x}. \\tag {1}\n$$\n", + "text_format": "latex", + "bbox": [ + 563, + 700, + 903, + 733 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "with $f'$ the first derivative. We estimate (1) using a path as defined in Section 2.1. We name the points in such a path $\\mathbf{X}_{\\boldsymbol{x}_a, \\boldsymbol{x}_z} := [\\boldsymbol{x}_a, \\boldsymbol{x}_b, \\boldsymbol{x}_c, \\dots, \\boldsymbol{x}_y, \\boldsymbol{x}_z]$ . We then have:", + "bbox": [ + 511, + 734, + 905, + 781 + ], + "page_idx": 1 + }, + { + "type": "equation", + "text": "\n$$\n\\begin{array}{l} \\operatorname {T V} (f, \\mathcal {T}) \\approx \\mathbb {E} _ {\\boldsymbol {x} _ {a}, \\boldsymbol {x} _ {z}} \\sim_ {\\mathcal {T}} | f (\\boldsymbol {x} _ {b}) - f (\\boldsymbol {x} _ {a}) | \\\\ + \\left| f \\left(\\boldsymbol {x} _ {c}\\right) - f (\\boldsymbol {x} _ {b}) \\right| + \\dots \\tag {2} \\\\ + \\left| f \\left(\\boldsymbol {x} _ {z}\\right) - f \\left(\\boldsymbol {x} _ {y}\\right) \\right|. \\\\ \\end{array}\n$$\n", + "text_format": "latex", + "bbox": [ + 529, + 786, + 903, + 834 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Appendix F shows that (2) correlates closely with a Fourier-based measure of complexity: the higher the TV, the higher the complexity. Yet, it is straightforward to implement and discriminative across small and large values.", + "bbox": [ + 511, + 839, + 905, + 900 + ], + "page_idx": 1 + }, + { + "type": "page_number", + "text": "80", + "bbox": [ + 488, + 945, + 506, + 955 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "2.3. Meta-Learning Activation Functions", + "text_level": 1, + "bbox": [ + 89, + 90, + 408, + 107 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Our goal is to optimize the inductive biases of a neural network and recent work [96] showed that the activation functions are the most important component. The typical approach to learn activations [2, 5, 6, 11, 13, 22, 45, 82, 91] (see Related Work) replaces them with a small shared ReLU MLP that implements an $\\mathbb{R}\\rightarrow \\mathbb{R}$ function. Its parameters are optimized along the network's. However this cannot discover truly novel activations because the embedded ReLU MLP has itself a simplicity bias and activations are optimized together with the model. We propose instead:", + "bbox": [ + 89, + 112, + 483, + 263 + ], + "page_idx": 2 + }, + { + "type": "list", + "sub_type": "text", + "list_items": [ + "- an unbiased parametrization of the activations as splines,", + "- a bi-level optimization to learn reusable activations,", + "- an episodic training to optimize for generalization rather than simply to fit the training data." + ], + "bbox": [ + 89, + 263, + 482, + 325 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Parametrization as splines. We want a space of activation functions free of priors such as the smoothness and monotonicity enforced in prior work [5, 13]. We implement an activation $g_{\\psi} : \\mathbb{R} \\to \\mathbb{R}$ as a linear spline with control points defined by $\\psi$ . We define $n_c$ points spread regularly in an interval $[a, b]$ , typically $\\sim 50$ points in $[-5, +5]$ . Then $g$ represents piecewise linear segments interpolating values specified in the learned parameters $\\psi := [g_{\\psi}(a), \\dots, g_{\\psi}(b)] \\in \\mathbb{R}^{n_c}$ . $g$ can represent simple and complex functions, including smooth curves, periodic functions, sharp transitions, etc.", + "bbox": [ + 89, + 329, + 483, + 479 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Bi-level optimization & episodic training. Our goal is to get an activation function that can be reused like any other in subsequent training runs. This differs from prior work (e.g. [2]) that continuously updates the activation during training: the final one may not be suitable to start training with. Our solution is a bi-level meta-learning loop. An inner loop trains the model with a fixed activation function. An outer loop trains the activation function to maximize generalization. Each outer step simulates a new learning task or episode. This means (1) initializing the model with different weights and (2) using different subsets of data for training and validation. With suitable choices, this can simulate in- or out-of-distribution conditions (see Section 3.4). Without episodes, the learned activation could overfit to a particular model initialization for example, and would not generalize in subsequent training runs. The method is outlined as Algorithm 1. Its implementation is discussed in Appendix C.", + "bbox": [ + 89, + 484, + 483, + 741 + ], + "page_idx": 2 + }, + { + "type": "image", + "img_path": "images/ed53be628f428e1cb7466d73447f60e716bb2b6787d390e027740dddbbeb9e65.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 102, + 752, + 140, + 777 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Inductive bias and simplicity bias are not interchangeable. Our method optimizes toward bethepneralization. Simplicity is only one aspect of the 1d models that we analyze post-hoc (e.g. Figure 4).", + "bbox": [ + 145, + 751, + 470, + 811 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3. Tasks and Results", + "text_level": 1, + "bbox": [ + 89, + 829, + 264, + 845 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "We now examine tasks that we hypothesized to be ill-suited to the simplicity bias of ReLU architectures. The intuition is that the target function to learn (e.g. optimal classifier) con", + "bbox": [ + 89, + 854, + 483, + 900 + ], + "page_idx": 2 + }, + { + "type": "code", + "sub_type": "algorithm", + "code_caption": [ + "Algorithm 1 Meta-learning an activation function (AF)." + ], + "code_body": "Input: training data $\\mathcal{T}$ ; untrained neural model $f_{\\theta ,\\psi}$ Initialize $\\psi$ with zeros Parametrization of AF $n_{\\mathrm{tr}}\\gets 0$ Number of inner-loop iterations while $n_{\\mathrm{tr}} < n_{\\mathrm{tr}}^{\\mathrm{max}}$ Outer loop: train AF Increment $n_{\\mathrm{tr}}$ Sample the episode's tr. $(T^{\\prime})$ and val. $(\\mathcal{V})$ sets from T Initialize $\\theta$ randomly Model weights and biases for $n_{\\mathrm{tr}}$ steps Inner loop: train model with fixed AF Eval. loss on $\\mathcal{T}':L\\gets \\Sigma_{(\\pmb {x},\\pmb {y})\\in \\mathcal{T}'}\\mathcal{L}\\big(f_{\\pmb {\\theta},\\pmb {\\psi}}(\\pmb {x},\\pmb {y})\\big)$ Gradient step on weights/biases: $\\pmb {\\theta}\\gets \\mathrm{GD}(\\pmb {\\theta},\\nabla_{\\pmb{\\theta}}L)$ Eval. loss on V: $L\\gets \\Sigma_{(\\pmb {x},\\pmb {y})\\in \\mathcal{V}}\\mathcal{L}\\big(f_{\\pmb {\\theta},\\pmb {\\psi}}(\\pmb {x},\\pmb {y})\\big)$ Gradient step on AF: $\\psi \\gets \\mathrm{GD}(\\psi ,\\nabla_{\\psi}L)$ if performance on $\\nu$ worsens then break Early stopping", + "bbox": [ + 516, + 109, + 890, + 362 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Output: optimized AF $\\psi$", + "bbox": [ + 519, + 364, + 692, + 381 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "tains sharp transitions (regression tasks, tabular datasets), or repeating patterns (algorithmic tasks) that contradict the ReLUs' simplicity bias. For each task, we examine existing datasets with the tools from Section 2. In all cases, we find benefits from architectures whose inductive biases favor more complex functions. Additional details and results are provided in Appendix E.", + "bbox": [ + 511, + 397, + 906, + 503 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3.1. Image Classification Tasks", + "text_level": 1, + "bbox": [ + 511, + 511, + 754, + 527 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Background. We start with classical datasets to validate our methodology: MNIST, FASHION-MNIST, SVHN, CIFAR-10 [49, 52, 68, 102]. They are representative of the vision tasks that guided the development of deep learning. Our hypothesis is therefore that the inductive biases of modern architectures and ReLUs are well suited to these datasets.", + "bbox": [ + 511, + 532, + 906, + 625 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Setup. For each dataset, we learn activation functions with Algorithm 1. We experiment with two initializations of the spline parameters: as zeros and so as to mimic a ReLU. The goal of the latter is to explore the space of functions similar to ReLUs. Because of the difficulty of the optimization, the algorithm is likely to converge to a local optimum similar to ReLUs if there is one. We also experiment with the sharing of the activation function. By default, a single function is shared across the network. Alternatively, we learn a different activation function per layer. This provides more ways to affect the model's inductive biases. Our base architecture is a 3-layer MLP (details in Appendix E).", + "bbox": [ + 511, + 628, + 908, + 809 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Results. We compare in Figure 2a the accuracy of models with ReLUs vs. learned activation functions. Differences are small. The learned activations only improve slightly on SVHN and CIFAR. This suggests that the inductive biases of ReLUs are generally well suited to these datasets.", + "bbox": [ + 511, + 813, + 906, + 887 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "We examine the learned activations in Figure 3a. With", + "bbox": [ + 511, + 888, + 905, + 904 + ], + "page_idx": 2 + }, + { + "type": "page_number", + "text": "81", + "bbox": [ + 488, + 945, + 506, + 955 + ], + "page_idx": 2 + }, + { + "type": "image", + "img_path": "images/211b35dc0b91d3f90a9b9e2fd0c6c2905f10ef07f027f107e2be95efab6b54df.jpg", + "image_caption": [ + "Figure 2. Test accuracy on image datasets. (a) For classification tasks, all models perform similarly, suggesting that the inductive biases of ReLUs are well suited to these datasets. (b) For regression tasks, models with learned activations perform better, especially from an initialization as zeros, which enables the discovery of completely novel activation functions." + ], + "image_footnote": [], + "bbox": [ + 94, + 78, + 228, + 181 + ], + "page_idx": 3 + }, + { + "type": "image", + "img_path": "images/a76d47c4c1ce9be58cd6a561fbc4f9a6eb005f7a13727ef07809908712d3a81f.jpg", + "image_caption": [], + "image_footnote": [ + "- ReLU activations\n- Learned act., ReLU init.\n- Learned act., zero init." + ], + "bbox": [ + 230, + 78, + 359, + 181 + ], + "page_idx": 3 + }, + { + "type": "image", + "img_path": "images/6471036dbc2489f41e7f191e6fa9b8105b97989fcf3ba5b13b887c3690c24440.jpg", + "image_caption": [ + "Figure 3. Activation functions learned for MNIST. For a classification task, the activation learned from a ReLU resembles the popular GeLUs. For a regression task, the learned activations contain irregularities that help a network represent complex functions. See Figure 16 for similar results on other datasets." + ], + "image_footnote": [], + "bbox": [ + 91, + 276, + 467, + 450 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "an initialization as ReLUs, the optimization converges to a smooth variant remarkably similar to GeLUs [11] which are widely used. This suggests that the research community has empirically converged on a local optimum in the space of activation functions. With an initialization as zeros, we discover wavelets [80] that are unlike common activations but perform as well as ReLUs, i.e. another local optimum.", + "bbox": [ + 89, + 541, + 483, + 648 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Take-away: for image classification, learned activations provide very little benefit over ReLUs. Smooth variants of ReLUs are a local optimum in the space of activations. ReLUs' popularity for such tasks could thus be explained with their proximity to this optimum.", + "bbox": [ + 101, + 659, + 472, + 736 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "3.2. Regression Tasks", + "text_level": 1, + "bbox": [ + 89, + 750, + 259, + 766 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Background. Regression tasks are known to be difficult for NNs [90]. They are often turned into a classification through discretization [24, 43]. Existing explanations that invoke implicit biases of gradient descent are clearly incomplete [90]. Our hypothesis is that regression is difficult because it often involves irregular decision boundaries [35] in opposition to the typical solutions of ReLU networks [17].", + "bbox": [ + 89, + 775, + 482, + 881 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Setup. We use the same setup and image datasets as Sec-", + "bbox": [ + 89, + 885, + 482, + 900 + ], + "page_idx": 3 + }, + { + "type": "image", + "img_path": "images/2413db2370a191b22bbcfc9433a64c78f101c9d5ec538770208e7115a8f0bbbc.jpg", + "image_caption": [ + "Figure 4. Accuracy vs. complexity on image datasets. Each marker is a model with different hyperparameters and ReLUs $(\\bullet)$ or learned activations initialized as ReLUs $(\\square)$ or as zeros $(\\text{圆})$ . For classification (a), ReLUs are close to best. Activations optimized from ReLUs only improve the accuracy slightly, corresponding to the GeLU-like function in Figure 3. For regression (b), new activations (learned from zeros) are best. Moreover, accuracy and complexity are clearly correlated only for regression. This supports the hypothesis that regression is more complex than classification and thus benefits from alternatives to the ReLUs' simplicity bias. See Figure 18 for similar results on other datasets." + ], + "image_footnote": [], + "bbox": [ + 558, + 78, + 705, + 191 + ], + "page_idx": 3 + }, + { + "type": "image", + "img_path": "images/022bc003f6aa1b5ffb090bf5ca3e5cc871e8994b0ce0ea05eaeca31b91ca2e79.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 723, + 78, + 859, + 191 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "tion 3.1. The task is now to directly predict class IDs. E.g. for MNIST this means predicting digit values. Models are trained with an MSE loss. To measure accuracy, we discretize the predictions to the nearest class ID.", + "bbox": [ + 511, + 364, + 905, + 425 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Results. The first observation from Figure 2b is that regression is clearly more difficult for NNs than classification (lower accuracies) despite the identical underlying task. Importantly, the learned activations now provide clear improvements, especially when learned from scratch (initialization as zeros). This confirms the hypothesis that the inductive biases of RLUs are not well suited to these tasks.", + "bbox": [ + 511, + 435, + 905, + 541 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Figure 3b shows that the learned activations contain more irregularities for regression than classification. Prior work [96] showed that this can help models represent complex functions with sharp transitions. An analysis of the complexity of trained models (Figures 4 and 18) shows that the accuracy is correlated with complexity for regression but not classification. And regression models with learned activations implement functions of higher complexity than with ReLUs. This supports the claim that the improvements arise from overcoming the simplicity bias of ReLUs.", + "bbox": [ + 511, + 550, + 905, + 702 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Complexity is only one dimension of the inductive biases. The complexity plots for SVHN (Figure 18) interestingly show that models with ReLUs and learned activations get different accuracies at the same complexity level. This shows that our meta learning approach can search over dimensions of the inductive biases that are not captured by our complexity measure, and are yet to be explicitly studied.", + "bbox": [ + 511, + 712, + 905, + 819 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Take-away: regression is more difficult for NNs than classification, and the simplicity bias of ReLUs is partly to blame. Learned activations improve performance by helping networks represent more complex functions.", + "bbox": [ + 522, + 830, + 893, + 891 + ], + "page_idx": 3 + }, + { + "type": "page_number", + "text": "82", + "bbox": [ + 488, + 945, + 506, + 955 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "3.3. Tabular Data", + "text_level": 1, + "bbox": [ + 91, + 90, + 230, + 104 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Background. Tabular data is any data with few unstructured dimensions, which often contains low-cardinality variables such as dates or categorical attributes. This contrasts e.g. with images, which contain many correlated, continuous dimensions (pixels). NNs struggle with tabular datasets and are often inferior to decision trees [35, 63]. Our hypothesis is that the inductive biases of standard architectures are ill-suited to such data because of the simplicity bias. It makes it difficult to learn functions where small changes in the input (e.g. day of the week) correspond to abrupt changes in the target — the definition of sensitivity, a proxy for complexity [17]. This seldom occurs in vision where similar images correspond to similar labels.", + "bbox": [ + 89, + 114, + 483, + 311 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Setup. We use 16 real-world classification datasets from Grinsztajn et al. [34, 35]. Bases include a linear classifier, k-NNs, and boosted decision trees. Our models are MLPs with 1-4 hidden layers (details in Appendix E.4). We compare learned activations functions with ReLUs and TanHs with a global prefactor, $\\tanh (\\alpha x)$ with $\\alpha \\in \\mathbb{R}^{+}$ tuned on the validation set. This is a simple option with tunable complexity, albeit with inductive biases of TanHs [45, 96].", + "bbox": [ + 89, + 315, + 483, + 436 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "We also experiment with learned input activation functions (IAFs). The motivation is to learn a different behavior for each input dimension. Since they carry different information, e.g. continuous vs. categorical variables, one could be suited to the simplicity bias while another is not, for example. IAFs are dimension-specific activation functions applied directly on the data before a standard MLP. IAFs are learned like AFs, from an initialization as the identity i.e. no effect by default. They subsume the gated inputs, and Fourier/numerical embeddings from prior work [20, 25, 32].", + "bbox": [ + 89, + 438, + 483, + 589 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Results. We compare the accuracy of models on the 16 datasets in Figures 5 and 20. Vanilla MLPs generally perform worse than trees. But adjusting the MLPs' inductive biases with learned prefactors or activations eliminates the gap. IAFs perform best, sometimes even surpassing trees. We analyze below the reasons for these improvements.", + "bbox": [ + 89, + 593, + 483, + 684 + ], + "page_idx": 4 + }, + { + "type": "image", + "img_path": "images/0ab3f5a495bc33f3325173094c344ecd7742e560a8f61f327615ce132c794e8f.jpg", + "image_caption": [ + "Figure 5. Comparison of model types over 16 tabular datasets. Vanilla MLPs often perform worse than decision trees, but adjusting their inductive biases with learned activation functions (AFs) eliminates this gap. The input activation functions (IAFs) enable even better performance. See Figure 20 for results per dataset." + ], + "image_footnote": [], + "bbox": [ + 117, + 691, + 450, + 825 + ], + "page_idx": 4 + }, + { + "type": "image", + "img_path": "images/dfa624194d959894bb7c0fdccd577ee48cf3651ba0a3e12c86c4ffafb2bd850f.jpg", + "image_caption": [ + "Best at low complexity" + ], + "image_footnote": [], + "bbox": [ + 516, + 83, + 612, + 167 + ], + "page_idx": 4 + }, + { + "type": "image", + "img_path": "images/d792fd28b74c36d0e1e7b3f038deb73197872e912e2fbf15d0920cf1edfa28ab.jpg", + "image_caption": [ + "Figure 6. Test accuracy vs. complexity on tabular datasets. Each marker represents a model with different hyperparameters, and ReLUs $(\\bullet)$ or learned activations initialized as ReLUs $(\\square)$ or as zeros $(\\square)$ . The learned activations perform better in all cases, but the accuracy peaks at different complexity levels. For some datasets, a low complexity is best and ReLUs thus perform quite well (left-most panel, note the smaller Y scale). For other datasets, the opposite is true and the improvements with learned activations is larger." + ], + "image_footnote": [], + "bbox": [ + 616, + 84, + 707, + 167 + ], + "page_idx": 4 + }, + { + "type": "image", + "img_path": "images/d04604de4f41000a298bdba6678ce37e51b87979e70ffb5ddead231238869014.jpg", + "image_caption": [ + "Different datasets", + "at datasets" + ], + "image_footnote": [], + "bbox": [ + 714, + 84, + 803, + 167 + ], + "page_idx": 4 + }, + { + "type": "image", + "img_path": "images/7c8516fa48acd91e8d3a370fe2998112d4c47833fd599945361ea07792df2ef4.jpg", + "image_caption": [ + "Best at higher complexity $\\uparrow$" + ], + "image_footnote": [], + "bbox": [ + 808, + 84, + 898, + 167 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Learned activation functions close the gap to decision trees by mimicking their inductive bias. We visualize in Figure 7c the functions implemented by different models, plotting their output over slices of the input space (Section 2.1). ReLUs produce the smoothest function while TanHs and learned activations induce sharper patterns. Notably, the IAFs induce sharp axis-aligned decision boundaries that are also characteristic of trees, with which they share a high accuracy. Axis-aligned transitions are the consequence of IAFs applied independently to each dimension. Sharp transitions originate from the complex shape of the learned activation function (Figure 7a) which is possible thanks to the unbiased spline parametrization (Section 2.3).", + "bbox": [ + 511, + 315, + 905, + 512 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "The simplicity/complexity bias is a property of the architecture. We visualize complexity landscapes of MLPs in Figure 7b. Similarly to standard loss landscapes [54], we plot model complexity over 2D slices of the parameter space. A first global view over a plane aligned with the training trajectory shows that complexity steadily increases during training for all models [48, 75, 104] but does so to the highest level for the best model (IAFs). A second view zooms in on each optimized solution in a random 2D plane. This examines the effect of arbitrary perturbations to the parameters. It shows that the ambient complexity of perturbed solutions of the best model is much higher than the solution itself, and then with less accurate models. This means that this architecture is more likely to represent complex functions because they are more abundant in parameter space [65, 84]. This is why the simplicity bias can be overcome: it results from architecture choices and not from an inevitable \"implicit bias\" of SGD [85, 89, 98, 105].", + "bbox": [ + 511, + 515, + 905, + 787 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Different tabular datasets require different inductive biases. We examine the relation between accuracy and complexity in Figure 6. The accuracy peaks at different complexity levels for different datasets. For some, a low complexity is best and ReLU MLPs perform well. For oth", + "bbox": [ + 511, + 790, + 903, + 867 + ], + "page_idx": 4 + }, + { + "type": "page_footnote", + "text": "This resembles an analysis of untrained models [15, 64, 96, 99] but focuses on relevant regions on the parameter space, near optimized models.", + "bbox": [ + 511, + 875, + 903, + 900 + ], + "page_idx": 4 + }, + { + "type": "page_number", + "text": "83", + "bbox": [ + 488, + 945, + 506, + 955 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "(a) Activation function & loss landscape (—training trajectory, $\\star$ early-stopping checkpoint, $\\bullet$ last checkpoint)", + "bbox": [ + 94, + 90, + 622, + 103 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/bad6edef9c1bc8728394cdffe30077a78da01f0153f0f1e2b12cd2182f52e52a.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 112, + 104, + 240, + 152 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/6fd9e9d9dcd0f06ba6459546a2aa20b0438c7dbdff1b349bcb51e7b3c2ea0948.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 274, + 104, + 401, + 152 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/f16688b3b136c6456aa0b3130512d904bec0357624b820eb582478f07a0d98a5.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 436, + 104, + 562, + 152 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/99f44ad2a7cdc1277b102e49942fded6ae5efa58eb372f813762876ff5c44f63.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 598, + 104, + 723, + 152 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/1cdf1bd89d8f8b02937ffa5c9ccbc1c9f9ba6dda970a559b5727aeb0eb968311.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 776, + 126, + 867, + 143 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "(b) Complexity landscape (in weight space along the PCA plane of the training trajectory) & zoom-in (random plane)", + "bbox": [ + 94, + 159, + 643, + 171 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/c752fddc5da611e26c16531aac4e01617512987ab02341ed960dd2558d73ceb4.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 112, + 172, + 236, + 220 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/c75597ea54fed9cf56bc03e79bbd5b1c3181f189013e5ee675e8f5575f82703d.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 274, + 172, + 398, + 220 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/ec77cc9d600b0af491e2b18885eef5e0fb42f9dbce7b4e0069a2b003b0ed9c6d.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 436, + 172, + 560, + 220 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/4bdcc9d70736b99d122ec4d8825092abe4aa1641cb36ba0638504d8998ff0304.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 598, + 172, + 722, + 220 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/a05c6b4ff26d0923effae88c0343e254fbba0fcf68ebb66d679995d464212742.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 774, + 195, + 867, + 213 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "(c) Function implemented by the network (in input space along four random planes containing each one training point $\\bullet$", + "bbox": [ + 94, + 228, + 661, + 239 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/54f22c31c5781e224a169e8e7103f2010d27bd9217944f874670083a202ee4ce.jpg", + "image_caption": [ + "MLP, ReLU" + ], + "image_footnote": [], + "bbox": [ + 112, + 239, + 236, + 335 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/d4288b4e375ddc9c5004153cc157bb14a828d67f3b7167e3dbd74c8af17a4ba8.jpg", + "image_caption": [ + "MLP, TanH w/tuned prefactor" + ], + "image_footnote": [ + "Increasing accuracy Best MLP" + ], + "bbox": [ + 274, + 239, + 398, + 335 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/f5ba720390310337ea1ef721280057a57d81d2394ffe14a47dbc35040f6d3c3c.jpg", + "image_caption": [ + "MLP, learned AF" + ], + "image_footnote": [], + "bbox": [ + 436, + 239, + 560, + 335 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/7c494a259530c3f2741489b48c8a6b8c473170b91091add9daa34e9b85069111.jpg", + "image_caption": [ + "MLP, learned IAFs" + ], + "image_footnote": [], + "bbox": [ + 598, + 239, + 722, + 335 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/19bbd88571bfd919c19a65ad324180c481896a00f036a7ad0c321438a13ee397.jpg", + "image_caption": [ + "Boosted decision trees" + ], + "image_footnote": [], + "bbox": [ + 759, + 239, + 885, + 335 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "ers, a higher complexity is best and the improvements with learned activations are larger. This supports the hypothesis that improvements over ReLU MLPs come from overcoming their simplicity bias. The variance across datasets is also unsurprising since they have little in common besides their low dimensionality (full results in Appendix E.4).", + "bbox": [ + 89, + 454, + 483, + 545 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Effect of width and depth. We show in Figure 8 that the learned activations can be reused in networks of different widths than they were trained for. The accuracy varies with width similarly as with ReLUs. Teney et al. [96] indeed showed that a model's width affects its capacity but not its inductive biases. Therefore width does not interfere with the effects of the learned activations. Figures 8 and 21 also show that good performance can be achieved with fewer layers than with ReLUs. Learned activations might thus have utility in model compression and distillation.", + "bbox": [ + 89, + 547, + 483, + 700 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Take-away: many tabular datasets are ill-suited to ReLU models because they require learning a complex function. Learned activations improve accuracy by implementing sharp axis-aligned decision boundaries that mimic the inductive biases of decision trees.", + "bbox": [ + 101, + 710, + 472, + 786 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "3.4. Shortcut Learning", + "text_level": 1, + "bbox": [ + 89, + 800, + 269, + 816 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Background. Shortcut learning occurs when a model learns spurious features instead of generalizable ones. It is a known consequence of the simplicity bias [85, 93] when the training data contains multiple features of different complexity. Our hypothesis is that the preference for some fea", + "bbox": [ + 89, + 824, + 483, + 901 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/5ad5edb24e4316146b38f696dd7a198fbb9ef5cfb04f3351a0cae4611512107e.jpg", + "image_caption": [ + "Figure 7. Models trained on the ELECTRICITY [34] tabular dataset. ReLU MLPs perform worst (left). TanHs induce sharper transitions in the network's function (c). So does the learned activation function (a) which is itself very irregular. The input activation functions (IAFs) perform best and mimic the axis-aligned boundaries of trees (bottom-right). The complexity landscapes (b) show that complexity increases to the highest level in the best model (IAFs). The zoom-in shows that the ambient complexity is also much higher than in other models. This means that it is inherently more likely to represent complex functions since they are more abundant in parameter space.", + "Figure 8. The learned activation functions surpass ReLUs, often with fewer layers. They can also be reused with different network widths (COVERTYPE [34] tabular dataset, see Figure 21 for others)." + ], + "image_footnote": [], + "bbox": [ + 583, + 450, + 831, + 558 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "tures depends on their alignment with the inductive biases. We will evaluate whether this can be controlled with activation functions.", + "bbox": [ + 511, + 628, + 905, + 672 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Setup. We use MNIST/CIFAR collages [85, 93, 94], a classification task over images combining tiles from MNIST and CIFAR-10. The training set is ambiguous: both tiles are predictive of the labels. Two unambiguous test sets evaluate reliance on either tile: one is predictive, the other contains a random class. We similarly build two validation sets to learn two activation functions optimized for either tile. We simulate OOD conditions by setting $\\mathcal{V}$ in Algorithm 1. The models are the fully-connected MLPs used in [93].", + "bbox": [ + 511, + 681, + 906, + 818 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Results. Figure 9 shows that a baseline with ReLUs is prone to shortcut learning. It relies exclusively on MNIST and the accuracy on the CIFAR test set is not better than chance (10%). In comparison, using either learned activation steers the learning towards either tile. The accuracy", + "bbox": [ + 511, + 824, + 906, + 901 + ], + "page_idx": 5 + }, + { + "type": "page_number", + "text": "84", + "bbox": [ + 488, + 945, + 506, + 955 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/d6b247dcca639738fdf5d443fe4f7bfe462550793593ecf7228ca8b6bf910e45.jpg", + "image_caption": [ + "Ambiguous training images" + ], + "image_footnote": [], + "bbox": [ + 106, + 89, + 169, + 141 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/c684f7795bd28c20580a087965e31eb6cc8ac02ab707ed5e709cc79ec3e5d2ac.jpg", + "image_caption": [ + "Optimized for MNIST" + ], + "image_footnote": [], + "bbox": [ + 183, + 88, + 256, + 143 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/50c38e44d937b34048fe3365f9b2308f201750ababa6cde9e85bc8c5ddf896b4.jpg", + "image_caption": [ + "Optimized for CIFAR" + ], + "image_footnote": [], + "bbox": [ + 259, + 88, + 333, + 143 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/a04e48553aa4a0d68c21ff4539f2373cf52f17a1413cee69f194d411e026db28.jpg", + "image_caption": [ + "Upper bound: ReLU trained on unambiguous MNIST Activation optimized for MNIST ReLU Baseline Activation optimized for CIFAR Upper bound: ReLU trained on unambiguous CIFAR" + ], + "image_footnote": [], + "bbox": [ + 545, + 89, + 797, + 167 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/301cbb3be4b74660f7cb9c81465c5f311bd49aa3c6329f6e809d035827b8f60b.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 797, + 88, + 898, + 166 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "shifts towards either of two tiles as the model prioritizes different features, merely with a change of activation function. This shows that the simplicity bias is not an inevitable effect of SGD. Instead, it directly reflects the alignment between the chosen architecture and the data.", + "bbox": [ + 89, + 238, + 483, + 313 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Training dynamics. The accuracy on CIFAR remains below a model trained on unambiguous CIFAR data. This is because training dynamics are also important. In Figure 9 (right), we plot the accuracy on the two tiles for the whole training trajectory. The reliance on different features varies, and the model eventually relies primarily on simple ones with enough iterations (i.e. without early stopping). This calls for future work combining our findings with the extensive literature on ID/OOD training dynamics [46, 95, 98].", + "bbox": [ + 89, + 316, + 483, + 454 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Take-away: we confirm that shortcut learning is a side effect of the simplicity bias. Different activation functions, while not completely avoiding shortcut learning, can steer the learning towards particular input features.", + "bbox": [ + 101, + 464, + 472, + 526 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "3.5. Algorithmic Tasks and Grokking", + "text_level": 1, + "bbox": [ + 89, + 540, + 382, + 556 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Background. Grokking is a phenomenon where a model first overfits the data (i.e. high training accuracy, low test accuracy) then shifts to high test accuracy after many training steps [73]. This is typically observed on algorithmic tasks and architectures from MLPs to transformers. Our hypothesis is that grokking is due to a mismatch between the target function and the model's inductive biases. Indeed, typical architectures were not developed for the algorithmic tasks where grokking is typically observed. To verify this hypothesis, we will show that endowing an architecture with the right inductive biases, using learned activation functions, can eliminate the phenomenon. Supporting this hypothesis, Zhou et al. [110] proposed that grokking comes from the frequency principle (i.e. low frequencies learned first by SGD), and Kumar et al. [51] showed that it correlates with a misalignment between features at initialization and the target function.", + "bbox": [ + 89, + 564, + 483, + 821 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Setup. Following [36, 51, 58] we train 1-hidden layer MLPs on algorithmic tasks, defined each by one binary operation (Figures 10 and 27). e.g. $y = (x_{1} + x_{2}) \\mod 13$ . The operands are passed as one-hot vectors and the task is a classification over possible outputs. Details in Appendix E.6.", + "bbox": [ + 89, + 824, + 483, + 902 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/d1d11deeb4a9b9b81cb958f58c49724164970864e0e5da9c8d343a1d345e6c86.jpg", + "image_caption": [ + "Figure 9. Experiments on shortcut learning with MNIST/CIFAR collages. The ReLU baseline (■) relies mostly on simple MNIST features. We learn two activation functions that shift the preference towards different features $(\\leftarrow / \\rightarrow)$ . Training trajectories (right) clearly differ with the activation optimized for CIFAR (—), for MNIST (—), or a ReLU (==). The model at initialization (random weights) is marked with $\\bullet$ .", + "Figure 10. Target functions used to investigate grokking [73] (details in Appendix, Figure 27). These patterns are very different from the tasks for which typical architectures were developed." + ], + "image_footnote": [], + "bbox": [ + 514, + 237, + 578, + 309 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/3dcee3491eae684eff053e6f0f384f70f597e8055a8aa1e15456e74be9a04da1.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 583, + 237, + 738, + 309 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/cd388dbc24fcdf1dd019094bd395c93e32996e3d3a1373fed0218c36523683df.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 741, + 237, + 838, + 309 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/44a767f71cfd239d8d47af4809f0d81fb951876d5eff84ae58a7f0445a81e582.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 839, + 237, + 903, + 309 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Results. We compare models with ReLU vs. learned activations across various tasks, network widths, and fractions of training data. We find that the learned, task-specific activations lead to faster convergence and/or higher test accuracy (Figures 11-13). On modular addition (a common task in the grokking literature) the learned-activation model converges $\\sim 10\\times$ faster than ReLUs. Curiously, some models with learned activations also end up overfitting (decreasing test accuracy) with prolonged training. In contrast, ReLU networks either never generalize (test accuracy $\\sim 0$ ) or grok and keep a high accuracy indefinitely. Further investigation is needed to explain this difference. We examine learned activation functions in Figure 12. See Figure 28 in the Appendix for results on other algorithmic tasks [73].", + "bbox": [ + 511, + 362, + 906, + 575 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/9920403831ee31eaddfe832ce9630e4b9cd06b614de2382e9554c6094a3e8800.jpg", + "image_caption": [ + "Figure 11. The learned activations essentially eliminate grokking (delayed convergence). On the above task (addition mod 27), our model converges $\\sim 10\\times$ faster than ReLUs." + ], + "image_footnote": [], + "bbox": [ + 517, + 580, + 741, + 642 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/83e4f3ac0255d05e53e3fa527646c06887f09c4c75c47fd1dc7899320d659641.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 741, + 580, + 900, + 642 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/b5b4266d96db9f14d1a6c8a2fe2d628a45a4da699f3ee05d5f3f7357a326fefc.jpg", + "image_caption": [ + "Figure 12. Activations learned for modular addition. The frequency of the sine-like function varies across versions of the task." + ], + "image_footnote": [], + "bbox": [ + 527, + 696, + 602, + 733 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/65eee4c6eee6e76429bc7140d09f002af221947e030d304d5c5245afa7708ae2.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 624, + 696, + 699, + 733 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/294f53ae5614dc603af10359800aaf25913da0c286cb71d8b322f390eb549d1d.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 720, + 696, + 795, + 733 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/ee32959f3b5caec906a43d63ac57bf87352145a3ed5d6be36ab461f7a76ba870.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 816, + 696, + 892, + 733 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/d2f0504c3ce9c9632681a94e323a57ef64b3ce024d7b4d7f7b22b2c8da9e9cc3.jpg", + "image_caption": [ + "Figure 13. Models with learned activations $(-)$ converge faster than ReLUs $(--)$ across a variety of settings (addition mod 27)." + ], + "image_footnote": [], + "bbox": [ + 557, + 771, + 712, + 849 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/053a8c1203fb699997e086dc2755219911c6bd571080ca3bd2683dd09be94139.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 723, + 770, + 815, + 849 + ], + "page_idx": 6 + }, + { + "type": "page_number", + "text": "85", + "bbox": [ + 488, + 945, + 506, + 955 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Take-away: learned activations eliminate grokking in all our cases, suggesting, as a cause, the mismatch between the data and the architectures' inductive biases.", + "bbox": [ + 102, + 93, + 470, + 138 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "4. Do the Activations Transfer Across Tasks?", + "text_level": 1, + "bbox": [ + 89, + 155, + 468, + 171 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "So far, we used dataset-specific activation functions and found that there exist better alternatives to ReLUs. A practical application would be the learning of activation functions suitable to a broad task, or range of related datasets.", + "bbox": [ + 89, + 180, + 482, + 241 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "As a first step, we study the specialization of the activations functions (AFs) learned for the 22 algorithmic tasks from Figure 27 [73]. We evaluate every task/activation combination, yielding the $22 \\times 22$ matrix of Figure 14. The learned activations do transfer, with improvements in accuracy and convergence shared across tasks. We also evaluate an activation learned on all tasks simultaneously. The accuracy across tasks (i.e. per-column average) reaches $61.5\\%$ vs. only $19.9\\%$ for ReLUs. and $54.0\\%$ on average for tasks-specific solutions. This procedure can thus improve performance on a range of related tasks. Future work could leverage it to discover activation functions that improve performance in other specific domains. See Appendix E.3 for other transfer experiments using image regression tasks.", + "bbox": [ + 89, + 242, + 483, + 454 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/616059b4a69b28d57058a7ca58c82ad3cebff55fedc8cedb0c41fe1f7be2a3a3.jpg", + "image_caption": [ + "Figure 14. Transfer of AFs (columns) across algorithmic tasks (rows). Colors represent the fraction of the best convergence speed or accuracy per task (brighter is better). If the activations were over-specialized, the matrix would be diagonal. On the contrary, it is densely filled, indicating positive transfer across many tasks." + ], + "image_footnote": [], + "bbox": [ + 91, + 457, + 483, + 618 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "5. Discussion", + "text_level": 1, + "bbox": [ + 89, + 709, + 204, + 724 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "We used activation functions as a tool to show that there exist a variety of inductive biases that are useful across applications of NNs. The impossibility of universal inductive biases is well known [101] but a strong argument has also been made that deep learning research is converging towards few architectures with wide applicability [31]. This argument rests on the NNs' simplicity bias being a good match for real-world data [12, 18, 56]. Our results do not invalidate these assumptions: NNs are widely applicable and their simplicity bias is evidently very effective on average. Our results show instead the following.", + "bbox": [ + 89, + 734, + 483, + 900 + ], + "page_idx": 7 + }, + { + "type": "list", + "sub_type": "text", + "list_items": [ + "1. There exist real-world tasks where the inductive biases of typical architectures' are suboptimal. This explanation connects four domains where NNs historically struggled.", + "2. The simplicity bias in modern NNs depends on particular design choices, the activation functions in particular. Research has converged on these choices by trial and error, in large part by optimizing performance on vision tasks. Therefore the adequacy of ReLUs for image classification (Section 3.1) is not accidental." + ], + "bbox": [ + 513, + 90, + 906, + 244 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Relevance to transformers and language models. The simplicity bias exists in transformers [10, 109] and language models [4, 31, 96]. Their embedding layer resembles input activations functions (Section 3.3). Could this explain the transformers' remarkable flexibility? I.e. a simplicity bias on an initial mapping of arbitrary complexity. Zhong and Andreas [108] indeed trained embeddings alone in a random-weight transformer and could learn complex tasks.", + "bbox": [ + 511, + 250, + 906, + 369 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Limitations and open questions. We prioritized breadth by establishing a new connection across multiple disparate topics in machine learning. Each section could expand into its own paper with additional models, datasets, comparisons, etc. Our findings on shortcut learning for example (Section 3.4) could yield new methods to address distribution shifts, though no such claim is made here. Here are the most promising follow-up questions opened by this paper.", + "bbox": [ + 511, + 375, + 908, + 496 + ], + "page_idx": 7 + }, + { + "type": "list", + "sub_type": "text", + "list_items": [ + "- How to fully characterize inductive biases? We focused on simplicity for its prevalence in AI [31], philosophy [72], and the natural sciences [12]. But it is only one dimension among many to characterize inductive biases.", + "- Can we improve state-of-the-art architectures? We used simple MLPs to isolate the effects of activations functions since they are central to the simplicity bias [96]. But other existing mechanisms (architectural, optimization) may already tweak or attenuate the simplicity bias.", + "- Can we learn transferable activation functions for other domains? We examined transferability in Sections 4 and E.3. The results suggest the possibility of better architectures optimized for specific domains. Predicting the suitability of an architecture/dataset pair ex ante (prior to training) would be extremely useful. This may follow from advances on the first open question above.", + "- Are there other detrimental effects of the simplicity bias? Any learning algorithm needs inductive biases to \"fill the gaps\" between training examples. The better they are, the fewer examples are needed. Researching what inductive biases are most useful on real-world tasks might thus hold the key for machine learning to become as data-efficient as humans. More speculatively, high-level cognition has been argued to require postulating explanations beyond the data [16, 23]. In this regard, simplicity-biased architectures might also hold us back." + ], + "bbox": [ + 514, + 500, + 908, + 901 + ], + "page_idx": 7 + }, + { + "type": "page_number", + "text": "86", + "bbox": [ + 490, + 945, + 506, + 955 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "References", + "text_level": 1, + "bbox": [ + 91, + 89, + 187, + 104 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[1] Sravanti Addepalli, Anshul Nasery, Venkatesh Babu Radhakrishnan, Praneeth Netrapalli, and Prateek Jain. Feature reconstruction from outputs can mitigate simplicity bias in neural networks. In ICLR, 2022. 1", + "[2] Konstantinos Panagiotis Alexandridis, Jiankang Deng, Anh Nguyen, and Shan Luo. Adaptive parametric activation. arXiv preprint arXiv:2407.08567, 2024. 2, 3, 1", + "[3] Konstantinos Panagiotis Alexandridis, Jiankang Deng, Anh Nguyen, and Shan Luo. Adaptive parametric activation. In ECCV, 2025. 2", + "[4] Badr AlKhamissi, Greta Tuckute, Antoine Bosselut, and Martin Schrimpf. Brain-like language processing via a shallow untrained multihead attention network. arXiv preprint arXiv:2406.15109, 2024. 8", + "[5] Andrea Apicella, Francesco Isgro, and Roberto Prevete. A simple and efficient architecture for trainable activation functions. Neurocomputing, 2019. 3, 1, 2", + "[6] Andrea Apicella, Francesco Donnarumma, Francesco Isgro, and Roberto Prevete. A survey on modern trainable activation functions. Neural Networks, 2021. 2, 3, 1", + "[7] Sanjeev Arora, Nadav Cohen, Wei Hu, and Yuping Luo. Implicit regularization in deep matrix factorization. NeurIPS, 2019. 1", + "[8] Devansh Arpit, Stanislaw Jastrzebski, Nicolas Ballas, David Krueger, Emmanuel Bengio, Maxinder S Kanwal, Tegan Maharaj, Asja Fischer, Aaron Courville, Yoshua Bengio, et al. A closer look at memorization in deep networks. In ICML. PMLR, 2017. 1", + "[9] Samuel James Bell and Levent Sagun. Simplicity bias leads to amplified performance disparities. In Proceedings of the 2023 ACM Conference on Fairness, Accountability, and Transparency, pages 355-369, 2023. 1", + "[10] Satwik Bhattachamishra, Arkil Patel, Varun Kanade, and Phil Blunsom. Simplicity bias in transformers and their ability to learn sparse boolean functions. arXiv preprint arXiv:2211.12316, 2022.8", + "[11] Garrett Bingham, William Macke, and Risto Miikkulainen. Evolutionary optimization of deep learning activation functions. In Genetic and Evolutionary Computation Conference, 2020. 3, 4, 1, 2", + "[12] Mark Buchanan. A natural bias for simplicity. Nature Physics, 2018. 8", + "[13] Irit Chelly, Shahaf E Finder, Shira Ifergane, and Oren Freifeld. Trainable highly-expressive activation functions. arXiv preprint arXiv:2407.07564, 2024. 3, 1, 2", + "[14] Nadav Cohen and Amnon Shashua. Inductive bias of deep convolutional networks through pooling geometry. arXiv preprint arXiv:1605.06743, 2016. 1", + "[15] Giacomo De Palma, Bobak Kiani, and Seth Lloyd. Random deep neural networks are biased towards simple functions. NeurIPS, 2019. 1, 5", + "[16] David Deutsch. The beginning of infinity: Explanations that transform the world. Penguin, 2011. 8", + "[17] Benoit Dherin, Michael Munn, Mihaela Rosca, and David Barrett. Why neural networks find simple solutions: The" + ], + "bbox": [ + 99, + 112, + 483, + 900 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "many regularizers of geometric complexity. NeurIPS, 35, 2022. 1, 4, 5", + "[18] Kamaludin Dingle, Chico Q Camargo, and Ard A Louis. Input-output maps are strongly biased towards simple outputs. Nature communications, 2018. 1, 8", + "[19] Pedro Domingos. The role of occam's razor in knowledge discovery. Data mining and knowledge discovery, 1999. 1", + "[20] Marius Dragoi, Florin Gogianu, and Elena Burceanu. Closing the gap on tabular data with fourier and implicit categorical features. Submission to ICLR (available on Open-Review), 2024. 5, 1", + "[21] Shiv Ram Dubey, Satish Kumar Singh, and Bidyut Baran Chaudhuri. Activation functions in deep learning: A comprehensive survey and benchmark. Neurocomputing, 2022. 1", + "[22] Stanislas Ducotterd, Alexis Goujon, Pakshal Bohra, Dimitris Perdios, Sebastian Neumayer, and Michael Unser. Improving lipschitz-constrained neural networks by learning activation functions. Journal of Machine Learning Research, 2024. 3, 1, 2", + "[23] Daniel C Elton. Applying deutsch's concept of good explanations to artificial intelligence and neuroscience—an initial exploration. Cognitive Systems Research, 2021. 8", + "[24] Jesse Farebrother, Jordi Orbay, Quan Vuong, Adrien Ali Taiga, Yevgen Chebotar, Ted Xiao, Alex Irpan, Sergey Levine, Pablo Samuel Castro, Aleksandra Faust, et al. Stop regressing: Training value functions via classification for scalable deep rl. arXiv preprint arXiv:2403.03950, 2024. 4", + "[25] James Fiedler. Simple modifications to improve tabular neural networks. CoRR, abs/2108.03214, 2021. 5", + "[26] Sara Fridovich-Keil, Raphael Gontijo Lopes, and Rebecca Roelofs. Spectral bias in practice: The role of function frequency in generalization. NeurIPS, 2022. 2", + "[27] Gallant. There exists a neural network that does not make avoidable mistakes. In IEEE International Conference on Neural Networks. IEEE, 1988. 1", + "[28] Isabel O Gallegos, Ryan A Rossi, Joe Barrow, Md Mehrab Tanjim, Sungchul Kim, Franck Dernoncourt, Tong Yu, Ruiyi Zhang, and Nesreen K Ahmed. Bias and fairness in large language models: A survey. Computational Linguistics, 2024. 1", + "[29] Robert Geirhos, Jorn-Henrik Jacobsen, Claudio Michaelis, Richard Zemel, Wieland Brendel, Matthias Bethge, and Felix A Wichmann. Shortcut learning in deep neural networks. Nature Machine Intelligence, 2020. 1", + "[30] Florin Gogianu, Tudor Berariu, Mihaela C Rosca, Claudia Clopath, Lucian Busoniu, and Razvan Pascanu. Spectral normalisation for deep reinforcement learning: an optimisation perspective. In ICML, 2021. 11", + "[31] Micah Goldblum, Marc Finzi, Keefer Rowan, and Andrew Gordon Wilson. The no free lunch theorem, kolmogorov complexity, and the role of inductive biases in machine learning. arXiv preprint arXiv:2304.05366, 2023. 1, 8", + "[32] Yury Gorishniy, Ivan Rubachev, and Artem Babenko. On embeddings for numerical features in tabular deep learning. NeurIPS, 2022. 5, 1, 9" + ], + "bbox": [ + 522, + 92, + 903, + 900 + ], + "page_idx": 8 + }, + { + "type": "page_number", + "text": "87", + "bbox": [ + 488, + 945, + 506, + 955 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[33] Anirudh Goyal and Yoshua Bengio. Inductive biases for deep learning of higher-level cognition. Proceedings of the Royal Society A, 2022. 1", + "[34] Léo Grinsztajn. Tabular data learning benchmark. https://github.com/LéoGrin/tabular-benchmark, 2022.5,6, 9,10,14", + "[35] Léo Grinsztajn, Edouard Oyallon, and Gael Varoquaux. Why do tree-based models still outperform deep learning on typical tabular data? NeurIPS, 2022. 4, 5, 1", + "[36] Andrey Gromov. Grokking modular arithmetic. arXiv preprint arXiv:2301.02679, 2023. 7", + "[37] Michael Hahn, Dan Jurafsky, and Richard Futrell. Sensitivity as a complexity measure for sequence classification tasks. Transactions of the ACL, 2021. 1", + "[38] Hrayr Harutyunyan, Rafayel Darbinyan, Samvel Karapetyan, and Hrant Khachatrian. In-context learning in presence of spurious correlations. arXiv preprint arXiv:2410.03140, 2024. 1", + "[39] Dan Hendrycks and Kevin Gimpel. Gaussian error linear units (GeLUs). arXiv preprint arXiv:1606.08415, 2016. 2, 1", + "[40] Katherine L Hermann and Andrew K Lampinen. What shapes feature representations? exploring datasets, architectures, and training. arXiv preprint arXiv:2006.12433, 2020. 1", + "[41] Hossein Hosseini, Baicen Xiao, Mayoore Jaiswal, and Radha Poovendran. Assessing shape bias property of convolutional neural networks. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition Workshops, 2018. 1", + "[42] Like Hui and Mikhail Belkin. Evaluation of neural architectures trained with square loss vs cross-entropy in classification tasks. arXiv preprint arXiv:2006.07322, 2020.1", + "[43] Ehsan Imani, Kai Luedemann, Sam Scholnick-Hughes, Esraa Elelimy, and Martha White. Investigating the histogram loss in regression. arXiv preprint arXiv:2402.13425, 2024. 4", + "[44] Ameya D Jagtap and George Em Karniadakis. How important are activation functions in regression and classification? a survey, performance comparison, and future directions. Journal of Machine Learning for Modeling and Computing, 4(1), 2023. 1", + "[45] Ameya D Jagtap, Kenji Kawaguchi, and George Em Karniadakis. Adaptive activation functions accelerate convergence in deep and physics-informed neural networks. Journal of Computational Physics, 2020. 3, 5, 1, 2", + "[46] Anchit Jain, Rozhin Nobahari, Aristide Baratin, and Stefano Sarao Mannelli. Bias in motion: Theoretical insights into the dynamics of bias in sgd training. arXiv preprint arXiv:2405.18296, 2024.7", + "[47] Liangze Jiang and Damien Teney. OOD-chameleon: Is algorithm selection for ood generalization learnable? arXiv preprint arXiv:2410.02735, 2024. 1", + "[48] Dimitris Kalimeris, Gal Kaplun, Preetum Nakkiran, Benjamin Edelman, Tristan Yang, Boaz Barak, and Haofeng Zhang. SGD on neural networks learns functions of increasing complexity. NeurIPS, 2019. 5, 1" + ], + "bbox": [ + 99, + 92, + 483, + 898 + ], + "page_idx": 9 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[49] Alex Krizhevsky and Geoffrey Hinton. Learning multiple layers of features from tiny images. Technical report, University of Toronto, 2009. 3", + "[50] Jan Kukačka, Vladimir Golkov, and Daniel Cremers. Regularization for deep learning: A taxonomy. arXiv preprint arXiv:1710.10686, 2017. 1", + "[51] Tanishq Kumar, Blake Bordelon, Samuel J Gershman, and Cengiz Pehlevan. Grokking as the transition from lazy to rich training dynamics. arXiv preprint arXiv:2310.06110, 2023. 7", + "[52] Yann LeCun, Léon Bottou, Yoshua Bengio, and Patrick Haffner. Gradient-based learning applied to document recognition. Proceedings of the IEEE, 86(11):2278-2324, 1998. 3", + "[53] Alexander Li and Deepak Pathak. Functional regularization for reinforcement learning via learned fourier features. NeurIPS, 2021. 1", + "[54] Hao Li, Zheng Xu, Gavin Taylor, Christoph Studer, and Tom Goldstein. Visualizing the loss landscape of neural nets. NeurIPS, 2018. 5", + "[55] Xuan Li, Yun Wang, and Bo Li. Tree-regularized tabular embeddings. arXiv preprint arXiv:2403.00963, 2024. 1", + "[56] Henry W Lin, Max Tegmark, and David Rolnick. Why does deep and cheap learning work so well? Journal of Statistical Physics, 2017. 8", + "[57] Ziqi Liu, Wei Cai, and Zhi-Qin John Xu. Multi-scale deep neural network (mscalednn) for solving poisson-boltzmann equation in complex domains. arXiv preprint arXiv:2007.11207, 2020. 1", + "[58] Ziming Liu, Ouail Kitouni, Niklas S Nolte, Eric Michaud, Max Tegmark, and Mike Williams. Towards understanding grokking: An effective theory of representation learning. NeurIPS, 2022. 7", + "[59] Ziming Liu, Yixuan Wang, Sachin Vaidya, Fabian Ruehle, James Halverson, Marin Soljačić, Thomas Y Hou, and Max Tegmark. Kan: Kolmogorov-arnold networks. arXiv preprint arXiv:2404.19756, 2024. 2", + "[60] Kaifeng Lyu, Zhiyuan Li, Runzhe Wang, and Sanjeev Arora. Gradient descent on two-layer nets: Margin maximization and simplicity bias. NeurIPS, 2021. 1", + "[61] Andrew L Maas, Awni Y Hannun, Andrew Y Ng, et al. Rectifier nonlinearities improve neural network acoustic models. In ICML, 2013. 1", + "[62] Augustine N Mavor-Parker, Matthew J Sargent, Caswell Barry, Lewis Griffin, and Clare Lyle. Frequency and generalisation of periodic activation functions in reinforcement learning. arXiv preprint arXiv:2407.06756, 2024. 1", + "[63] Duncan McElfresh, Sujay Khandagale, Jonathan Valverde, Vishak Prasad C, Ganesh Ramakrishnan, Micah Goldblum, and Colin White. When do neural nets outperform boosted trees on tabular data? NeurIPS, 2024. 5", + "[64] Chris Mingard, Joar Skalse, Guillermo Valle-Pérez, David Martínez-Rubio, Vladimir Mikulik, and Ard A Louis. Neural networks are a priori biased towards boolean functions with low entropy. arXiv preprint arXiv:1909.11522, 2019. 1, 5" + ], + "bbox": [ + 524, + 92, + 906, + 898 + ], + "page_idx": 9 + }, + { + "type": "page_number", + "text": "88", + "bbox": [ + 490, + 946, + 506, + 955 + ], + "page_idx": 9 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[65] Chris Mingard, Guillermo Valle-Pérez, Joar Skalse, and Ard A Louis. Is SGD a bayesian sampler? well, almost. Journal of Machine Learning Research, 2021. 1, 5", + "[66] Chris Mingard, Henry Rees, Guillermo Valle-Pérez, and Ard A Louis. Do deep neural networks have an inbuilt ocean's razor? arXiv preprint arXiv:2304.06670, 2023. 1", + "[67] Tom M Mitchell. The need for biases in learning generalizations. Rutgers University CS tech report CBM-TR-117, 1980. 1", + "[68] Yuval Netzer, Tao Wang, Adam Coates, Alessandro Bissacco, Bo Wu, and Andrew Y Ng. Reading digits in natural images with unsupervised feature learning. NIPS Workshop on Deep Learning and Unsupervised Feature Learning, 2011. 3", + "[69] Behnam Neyshabur, Ryota Tomioka, and Nathan Srebro. In search of the real inductive bias: On the role of implicit regularization in deep learning. arXiv preprint arXiv:1412.6614, 2014. 1", + "[70] Mohammad Pezeshki, Oumar Kaba, Yoshua Bengio, Aaron C Courville, Doina Precup, and Guillaume Lajoie. Gradient starvation: A learning proclivity in neural networks. NeurIPS, 2021. 1", + "[71] Tomaso Poggio, Kenji Kawaguchi, Qianli Liao, Brando Miranda, Lorenzo Rosasco, Xavier Boix, Jack Hiday, and Hrushikesh Mhaskar. Theory of deep learning III: the nonoverfitting puzzle. CBMM Memo, 2018. 1", + "[72] Karl Popper. \"7. Simplicity\". The logic of scientific discovery. Routledge, 1959. 8", + "[73] Alethea Power, Yuri Burda, Harri Edwards, Igor Babuschkin, and Vedant Misra. Grokking: Generalization beyond overfitting on small algorithmic datasets. arXiv preprint arXiv:2201.02177, 2022. 7, 8, 12", + "[74] Aahlad Manas Puli, Lily Zhang, Yoav Wald, and Rajesh Ranganath. Don't blame dataset shift! shortcut learning due to gradients and cross entropy. NeurIPS, 2023. 1", + "[75] Nasim Rahaman, Aristide Baratin, Devansh Arpit, Felix Draxler, Min Lin, Fred Hamprecht, Yoshua Bengio, and Aaron Courville. On the spectral bias of neural networks. In ICML. PMLR, 2019. 5, 1", + "[76] Prajit Ramachandran, Barret Zoph, and Quoc V Le. Swish: a self-gated activation function. arXiv preprint arXiv:1710.05941, 2017. 1", + "[77] Sameera Ramasinghe and Simon Lucey. Beyond periodicity: Towards a unifying framework for activations in coordinate-MLPs. In ECCV. Springer, 2022. 1", + "[78] Sameera Ramasinghe, Lachlan E MacDonald, and Simon Lucey. On the frequency-bias of coordinate-MLPs. NeurIPS, 2022. 1", + "[79] Mihaela Rosca, Theophane Weber, Arthur Gretton, and Shakir Mohamed. A case for new neural network smoothness constraints. I Can't Believe It's Not Better (ICBINB) Workshop at NeurIPS, 2020. 11", + "[80] Vishwanath Saragadam, Daniel LeJeune, Jasper Tan, Guha Balakrishnan, Ashok Veeraraghavan, and Richard G Baraniuk. Wire: Wavelet implicit neural representations. In CVPR, 2023. 4, 1" + ], + "bbox": [ + 99, + 90, + 482, + 898 + ], + "page_idx": 10 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[81] Simone Scardapane, Michele Scarpiniti, Danilo Comminiello, and Aurelio Uncini. Learning activation functions from data using cubic spline interpolation. Neural Advances in Processing Nonlinear Dynamic Signals, 2019. 2", + "[82] Simone Scardapane, Steven Van Vaerenbergh, Simone Toto, and Aurelio Uncini. Kafnets: Kernel-based nonparametric activation functions for neural networks. Neural Networks, 2019. 2, 3, 1", + "[83] Jürgen Schmidhuber. Discovering neural nets with low kolmogorov complexity and high generalization capability. Neural Networks, 1997. 1", + "[84] Luca Scimeca, Seong Joon Oh, Sanghyuk Chun, Michael Poli, and Sangdoo Yun. Which shortcut cues will DNNs choose? a study from the parameter-space perspective. arXiv preprint arXiv:2110.03095, 2021. 5", + "[85] Harshay Shah, Kaustav Tamuly, Aditi Raghunathan, Prateek Jain, and Praneeth Netrapalli. The pitfalls of simplicity bias in neural networks. NeurIPS, 2020. 2, 5, 6, 11", + "[86] Kexuan Shi, Xingyu Zhou, and Shuhang Gu. Improved implicit neural representation with fourier reparameterized training. In CVPR, 2024. 1", + "[87] Prasann Singhal, Tanya Goyal, Jiacheng Xu, and Greg Durrett. A long way to go: Investigating length correlations in rlhf. arXiv preprint arXiv:2310.03716, 2023. 1", + "[88] Vincent Sitzmann, Julien Martel, Alexander Bergman, David Lindell, and Gordon Wetzstein. Implicit neural representations with periodic activation functions. NeurIPS, 2020. 1", + "[89] Daniel Soudry, Elad Hoffer, Mor Shpigel Nacson, Suriya Gunasekar, and Nathan Srebro. The implicit bias of gradient descent on separable data. The Journal of Machine Learning Research, 2018. 5", + "[90] Lawrence Stewart, Francis Bach, Quentin Berthet, and Jean-Philippe Vert. Regression as classification: Influence of task formulation on neural network features. In ICML. PMLR, 2023. 4", + "[91] Leon René Sutfeld, Flemming Brieger, Holger Finger, Sonja Fülhase, and Gordon Pipa. Adaptive blending units: Trainable activation functions for deep neural networks. In Intelligent Computing: Proceedings of the Computing Conference. Springer, 2020. 3, 1, 2", + "[92] Remi Tachet, Mohammad Pezeshki, Samira Shabanian, Aaron Courville, and Yoshua Bengio. On the learning dynamics of deep neural networks. arXiv preprint arXiv:1809.06848, 2018. 1", + "[93] Damien Teney, Ehsan Abbasnejad, Simon Lucey, and Anton van den Hengel. Evading the simplicity bias: Training a diverse set of models discovers solutions with superior ood generalization. arXiv preprint arXiv:2105.05612, 2021. 1, 2, 6, 11", + "[94] Damien Teney, Maxime Peyrard, and Ehsan Abbasnejad. Predicting is not understanding: Recognizing and addressing underspecification in machine learning. In ECCV. Springer, 2022. 6, 1", + "[95] Damien Teney, Yong Lin, Seong Joon Oh, and Ehsan Abbasnejad. ID and OOD performance are sometimes in" + ], + "bbox": [ + 522, + 92, + 903, + 900 + ], + "page_idx": 10 + }, + { + "type": "page_number", + "text": "89", + "bbox": [ + 488, + 945, + 506, + 955 + ], + "page_idx": 10 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "versely correlated on real-world datasets. NeurIPS, 2024. 1, 7", + "[96] Damien Teney, Armand Mihai Nicoliciou, Valentin Hartmann, and Ehsan Abbasnejad. Neural redshift: Random networks are not random functions. In CVPR, 2024. 1, 2, 3, 4, 5, 6, 8, 11, 14", + "[97] Tijmen Tieleman. Lecture 6.5-rmsprop: Divide the gradient by a running average of its recent magnitude. COURSERA: Neural networks for machine learning, 4(2):26, 2012. 8", + "[98] Nikita Tsoy and Nikola Konstantinov. Simplicity bias of two-layer networks beyond linearly separable data. arXiv preprint arXiv:2405.17299, 2024. 5, 7", + "[99] Guillermo Valle-Perez, Chico Q Camargo, and Ard A Louis. Deep learning generalizes because the parameter-function map is biased towards simple functions. arXiv preprint arXiv:1805.08522, 2018. 1, 5", + "[100] Colin White, Mahmoud Safari, Rhea Sukthanker, Binxin Ru, Thomas Elsken, Arber Zela, Debadeepta Dey, and Frank Hutter. Neural architecture search: Insights from 1000 papers. arXiv preprint arXiv:2301.08727, 2023. 2", + "[101] David H Wolpert. The supervised learning no-free-lunch theorems. Soft computing and industry: Recent applications, 2002. 1, 8", + "[102] Han Xiao, Kashif Rasul, and Roland Vollgraf. Fashion-mnist: A novel image dataset for benchmarking machine learning algorithms. arXiv preprint arXiv:1708.07747, 2017.3", + "[103] Yiheng Xie, Towaki Takikawa, Shunsuke Saito, Or Litany, Shiqin Yan, Numair Khan, Federico Tombari, James Tompkin, Vincent Sitzmann, and Srinath Sridhar. Neural fields in visual computing and beyond. In Computer Graphics Forum. Wiley Online Library, 2022. 1", + "[104] Zhi-Qin John Xu, Yaoyu Zhang, Tao Luo, Yanyang Xiao, and Zheng Ma. Frequency principle: Fourier analysis sheds light on deep neural networks. arXiv preprint arXiv:1901.06523, 2019. 5, 1", + "[105] Zhi-Qin John Xu, Yaoyu Zhang, and Yanyang Xiao. Training behavior of deep neural network in frequency domain. In ICONIP. Springer, 2019. 5", + "[106] Zhi-Qin John Xu, Yaoyu Zhang, and Tao Luo. Overview frequency principle/spectral bias in deep learning. Communications on Applied Mathematics and Computation, 2024. 1", + "[107] Ge Yang, Anurag Ajay, and Pulkit Agrawal. Overcoming the spectral bias of neural value approximation. arXiv preprint arXiv:2206.04672, 2022. 1", + "[108] Ziqian Zhong and Jacob Andreas. Algorithmic capabilities of random transformers. arXiv preprint arXiv:2410.04368, 2024. 8", + "[109] Hattie Zhou, Arwen Bradley, Etai Littwin, Noam Razin, Omid Saremi, Josh Susskind, Samy Bengio, and Preetum Nakkiran. What algorithms can transformers learn? a study in length generalization. arXiv preprint arXiv:2310.16028, 2023. 8", + "[110] Zhangchen Zhou, Yaoyu Zhang, and Zhi-Qin John Xu. A rationale from frequency perspective for grokking in training neural network. arXiv preprint arXiv:2405.17479, 2024. 7" + ], + "bbox": [ + 91, + 92, + 482, + 898 + ], + "page_idx": 11 + }, + { + "type": "page_number", + "text": "90", + "bbox": [ + 490, + 945, + 506, + 955 + ], + "page_idx": 11 + } +] \ No newline at end of file diff --git a/2025/Do We Always Need the Simplicity Bias_ Looking for Optimal Inductive Biases in the Wild/f8b7dd25-2878-4e2f-ac24-ae160cb54c72_model.json b/2025/Do We Always Need the Simplicity Bias_ Looking for Optimal Inductive Biases in the Wild/f8b7dd25-2878-4e2f-ac24-ae160cb54c72_model.json new file mode 100644 index 0000000000000000000000000000000000000000..2a475ec32b5e550516f6e2e71f1dbb35a95ec962 --- /dev/null +++ b/2025/Do We Always Need the Simplicity Bias_ Looking for Optimal Inductive Biases in the Wild/f8b7dd25-2878-4e2f-ac24-ae160cb54c72_model.json @@ -0,0 +1,3502 @@ +[ + [ + { + "type": "header", + "bbox": [ + 0.107, + 0.003, + 0.182, + 0.043 + ], + "angle": 0, + "content": "CVF" + }, + { + "type": "header", + "bbox": [ + 0.237, + 0.001, + 0.812, + 0.047 + ], + "angle": 0, + "content": "This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore." + }, + { + "type": "title", + "bbox": [ + 0.249, + 0.131, + 0.75, + 0.177 + ], + "angle": 0, + "content": "Do We Always Need the Simplicity Bias? Looking for Optimal Inductive Biases in the Wild" + }, + { + "type": "text", + "bbox": [ + 0.132, + 0.205, + 0.322, + 0.24 + ], + "angle": 0, + "content": "Damien Teney \nIdiap Research Institute" + }, + { + "type": "text", + "bbox": [ + 0.361, + 0.205, + 0.476, + 0.238 + ], + "angle": 0, + "content": "Liangze Jiang EPFL" + }, + { + "type": "title", + "bbox": [ + 0.249, + 0.268, + 0.327, + 0.284 + ], + "angle": 0, + "content": "Abstract" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.303, + 0.484, + 0.41 + ], + "angle": 0, + "content": "Neural architectures tend to fit their data with relatively simple functions. This \"simplicity bias\" is widely regarded as key to their success. This paper explores the limits of this principle. Building on recent findings that the simplicity bias stems from ReLU activations [96], we introduce a method to meta-learn new activation functions and inductive biases better suited to specific tasks." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.414, + 0.484, + 0.566 + ], + "angle": 0, + "content": "Findings. We identify multiple tasks where the simplicity bias is inadequate and ReLUs suboptimal. In these cases, we learn new activation functions that perform better by inducing a prior of higher complexity. Interestingly, these cases correspond to domains where neural networks have historically struggled: tabular data, regression tasks, cases of shortcut learning, and algorithmic grokking tasks. In comparison, the simplicity bias induced by ReLUs proves adequate on image tasks where the best learned activations are nearly identical to ReLUs and GeLUs." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.57, + 0.484, + 0.692 + ], + "angle": 0, + "content": "Implications. Contrary to popular belief, the simplicity bias of ReLU networks is not universally useful. It is near-optimal for image classification, but other inductive biases are sometimes preferable. We showed that activation functions can control these inductive biases, but future tailored architectures might provide further benefits. Advances are still needed to characterize a model's inductive biases beyond \"complexity\", and their adequacy with the data." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.722, + 0.222, + 0.737 + ], + "angle": 0, + "content": "1. Introduction" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.75, + 0.483, + 0.901 + ], + "angle": 0, + "content": "When and why NNs generalize is yet to be understood. Neural networks (NNs) have proven more effective than other machine learning models. However, we still miss a complete explanation of their generalization abilities. A better understanding could help address failures from shortcut learning [29, 93] to distribution shifts [47, 95], biases, and spurious correlations in language models for example [28, 38, 87]. Understanding conditions for generalization would also enable the design of architectures and data preparation from first principles, rather than trial and error." + }, + { + "type": "text", + "bbox": [ + 0.514, + 0.205, + 0.639, + 0.238 + ], + "angle": 0, + "content": "Florin Gogianu\nBitdefender" + }, + { + "type": "text", + "bbox": [ + 0.679, + 0.205, + 0.861, + 0.24 + ], + "angle": 0, + "content": "Ehsan Abbasnejad \nUniversity of Adelaide" + }, + { + "type": "image", + "bbox": [ + 0.516, + 0.267, + 0.905, + 0.378 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.512, + 0.392, + 0.907, + 0.463 + ], + "angle": 0, + "content": "Figure 1. (1) We modulate the inductive bias of neural architectures by learning novel activation functions that improve generalization on specific datasets. (2) With this tool, we study the relation between model accuracy and complexity. We identify tasks where the simplicity bias of ReLU architectures is suboptimal." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.482, + 0.907, + 0.649 + ], + "angle": 0, + "content": "This paper studies inductive biases i.e. the assumptions made by learning algorithms to generalize beyond training data [67].1 A vast literature examines the inductive biases of architectures [14], optimizers [69], losses [42], regularizers [50], etc. The simplicity bias is one aspect of the inductive biases of NNs that makes them fit their training data with simple2 functions [8, 71]. Despite wide belief that the simplicity bias could be due to SGD [7, 40, 60, 92], work on untrained networks showed that it can be explained with architectures alone [15, 31, 64, 99]. ReLUs also seem critical to induce the simplicity bias in typical architectures [96]." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.658, + 0.909, + 0.841 + ], + "angle": 0, + "content": "Limits of the simplicity bias. The simplicity bias is an intuitive explanation for the ability to generalize on real-world data. It embodies Occam's razor [66] and assumes that data-generating processes in the real world are simple. Additionally, a prior for simplicity is supported by results in algorithmic information theory [18] stating essentially that \"a bias in the distribution of target functions must be towards low complexity\". However, this only means that simplicity is a good prior on average, but not necessarily the best choice on any task or dataset. This matches the no-free lunch theorem [101] according to which no inductive bias is universally useful. Therefore, this paper asks the following." + }, + { + "type": "page_footnote", + "bbox": [ + 0.513, + 0.863, + 0.905, + 0.901 + ], + "angle": 0, + "content": "\\(^{1}\\)Inductive biases can formalized as a prior over the space of functions [65]. \\(^{2}\\)Simplicity can be formalized using Kolmogorov complexity or its approximations, e.g. frequency, compressibility, sensitivity, etc. [17, 18, 37, 96]" + }, + { + "type": "page_number", + "bbox": [ + 0.491, + 0.946, + 0.508, + 0.957 + ], + "angle": 0, + "content": "79" + } + ], + [ + { + "type": "title", + "bbox": [ + 0.103, + 0.094, + 0.471, + 0.14 + ], + "angle": 0, + "content": "Are there practical applications of machine learning where the simplicity bias is detrimental? In these cases, what do the optimal inductive biases look like?" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.149, + 0.482, + 0.18 + ], + "angle": 0, + "content": "For example, shortcut learning is one situation where the simplicity bias is already known to be detrimental [85, 93]." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.186, + 0.484, + 0.381 + ], + "angle": 0, + "content": "Searching for optimal inductive biases by learning activation functions. Prior work [96] showed that ReLU activations are critical to obtain the simplicity bias in typical architectures. Hence we build a new tool to modulate the simplicity bias by learning dataset-specific activation functions. It uses bi-level optimization and a spline parametrization to learn activations free of any prior, such as constraints of smoothness or monotonicity (unlike prior work [2, 6, 82]). This (1) enables the discovery of entirely new activation functions and inductive biases that improve generalization (Figure 1) and (2) highlights the suboptimality of the simplicity bias by comparing the accuracy and complexity of models with ReLUs vs. learned activations." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.388, + 0.484, + 0.599 + ], + "angle": 0, + "content": "Findings. We examine four domains that we hypothesized to be impaired by the simplicity bias: tabular data, regression tasks, cases of shortcut learning, and algorithmic tasks. Our intuition is that they require learning functions with high sensitivity or sharp transitions. For each domain, we collect existing datasets then train and analyze models without and with learned activation functions. In all cases, we obtain better generalization with dataset-specific activations, and the improvements are attributable to learning higher-complexity solutions. In comparison, this analysis also shows that classical image datasets (MNIST, CIFAR, FASHION-MNIST, SVHN) are extremely well suited to the inductive biases of ReLUs. The best learned activations are then strikingly similar to variants like GeLUs [39]." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.606, + 0.285, + 0.62 + ], + "angle": 0, + "content": "Summary of contributions." + }, + { + "type": "text", + "bbox": [ + 0.091, + 0.624, + 0.483, + 0.654 + ], + "angle": 0, + "content": "- A new method to discover dataset-specific activation functions optimized for generalization." + }, + { + "type": "text", + "bbox": [ + 0.091, + 0.657, + 0.484, + 0.867 + ], + "angle": 0, + "content": "- An examination of \\(>20\\) datasets showing that the simplicity bias of ReLU architectures can be suboptimal. (1) For regression tasks and tabular data, new learned activations greatly improve accuracy by helping learn complex functions. (2) For image classification, the process rediscovers smooth variants of ReLUs, suggesting a near-optimal choice for these popular tasks. (3) In cases of shortcut learning, we show that different learned activations can steer the learning towards different image features. (4) For grokking tasks, new learned activations can eliminate the phenomenon, supporting the explanation as a mismatch between data and architectures. We also measure a positive transferability of learned activations across related tasks." + }, + { + "type": "text", + "bbox": [ + 0.091, + 0.871, + 0.484, + 0.901 + ], + "angle": 0, + "content": "- An analysis showing that improvements with learned activations correlate with the learning of complex functions." + }, + { + "type": "list", + "bbox": [ + 0.091, + 0.624, + 0.484, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.092, + 0.906, + 0.213 + ], + "angle": 0, + "content": "Implications. All cases where the simplicity bias proved suboptimal are in domains where NNs have historically struggled. We now connect them to a common explanation. This implies that architectures tailored to some specific domains may still have a place besides scaling up models and data. Conversely, the suitability of ReLUs to image classification suggests that researchers successfully converged by trial and error to designs well tuned to popular tasks." + }, + { + "type": "title", + "bbox": [ + 0.514, + 0.234, + 0.612, + 0.249 + ], + "angle": 0, + "content": "2. Methods" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.259, + 0.906, + 0.29 + ], + "angle": 0, + "content": "This section introduces tools to analyze trained models and to learn new dataset-specific activation functions." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.307, + 0.787, + 0.323 + ], + "angle": 0, + "content": "2.1. Visualizing a Model's Function" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.328, + 0.907, + 0.587 + ], + "angle": 0, + "content": "A neural network implements a function \\( f_{\\theta} : \\mathbb{R}^{d_{\\mathrm{in}}} \\rightarrow \\mathbb{R}^{d_{\\mathrm{out}}} \\) of parameters \\( \\theta \\) (weights and biases) that maps an input \\( x \\in \\mathbb{R}^{d_{\\mathrm{in}}} \\) to an output \\( y \\in \\mathbb{R}^{d_{\\mathrm{out}}} \\). For a regression task, \\( y \\in \\mathbb{R} \\) is the predicted value. For a classification task, \\( y \\) is a vector of logits passed through a sigmoid or softmax to obtain class probabilities. Because \\( d_{\\mathrm{in}} \\) can be large, \\( f \\) can be difficult to visualize and analyze. A workaround is to examine \\( f \\) over 1D or 2D slices of the input space [26, 96]. To obtain a slice in a region of plausible data, we use the training data \\( \\mathcal{T} \\). For a 1D slice (linear path), we sample \\( x_1, x_2 \\sim \\mathcal{T} \\) then define the path \\( \\mathbf{X}_{x_1, x_2} = [(1 - \\lambda) x_1 + \\lambda x_2, \\lambda \\in [0,1]] \\). We proceed analogously with three points for a 2D slice. We sample \\( \\lambda \\) regularly in [0,1] such that \\( \\mathbf{X} \\) is a finite sequence of points. Then \\( f \\) is evaluated on these points to give a 1D sequence or 2D grid of values that are convenient to display and analyze (Figure 7c). When \\( d_{\\mathrm{out}} > 1 \\) (multi-class task), we examine one random dimension of \\( f \\)'s output at a time." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.603, + 0.805, + 0.619 + ], + "angle": 0, + "content": "2.2. Measuring a Model's Complexity" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.624, + 0.906, + 0.701 + ], + "angle": 0, + "content": "We wish to quantify the complexity of the function \\( f \\) implemented by a model trained on data \\( \\mathcal{T} \\). Prior work used Fourier decompositions [26, 96] but this requires a delicate implementation. We found a reliable alternative with the total variation (TV) of \\( f \\) averaged over many 1D paths:" + }, + { + "type": "equation", + "bbox": [ + 0.565, + 0.701, + 0.905, + 0.734 + ], + "angle": 0, + "content": "\\[\n\\operatorname {T V} (f, \\mathcal {T}) = \\mathbb {E} _ {\\boldsymbol {x} _ {1}, \\boldsymbol {x} _ {2}} \\sim \\mathcal {T} \\int_ {\\boldsymbol {x} _ {1}} ^ {\\boldsymbol {x} _ {2}} \\left| f ^ {\\prime} (\\boldsymbol {x}) \\right| d \\boldsymbol {x}. \\tag {1}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.735, + 0.906, + 0.782 + ], + "angle": 0, + "content": "with \\( f' \\) the first derivative. We estimate (1) using a path as defined in Section 2.1. We name the points in such a path \\( \\mathbf{X}_{\\boldsymbol{x}_a, \\boldsymbol{x}_z} := [\\boldsymbol{x}_a, \\boldsymbol{x}_b, \\boldsymbol{x}_c, \\dots, \\boldsymbol{x}_y, \\boldsymbol{x}_z] \\). We then have:" + }, + { + "type": "equation", + "bbox": [ + 0.531, + 0.787, + 0.905, + 0.835 + ], + "angle": 0, + "content": "\\[\n\\begin{array}{l} \\operatorname {T V} (f, \\mathcal {T}) \\approx \\mathbb {E} _ {\\boldsymbol {x} _ {a}, \\boldsymbol {x} _ {z}} \\sim_ {\\mathcal {T}} | f (\\boldsymbol {x} _ {b}) - f (\\boldsymbol {x} _ {a}) | \\\\ + \\left| f \\left(\\boldsymbol {x} _ {c}\\right) - f (\\boldsymbol {x} _ {b}) \\right| + \\dots \\tag {2} \\\\ + \\left| f \\left(\\boldsymbol {x} _ {z}\\right) - f \\left(\\boldsymbol {x} _ {y}\\right) \\right|. \\\\ \\end{array}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.84, + 0.906, + 0.901 + ], + "angle": 0, + "content": "Appendix F shows that (2) correlates closely with a Fourier-based measure of complexity: the higher the TV, the higher the complexity. Yet, it is straightforward to implement and discriminative across small and large values." + }, + { + "type": "page_number", + "bbox": [ + 0.49, + 0.946, + 0.508, + 0.956 + ], + "angle": 0, + "content": "80" + } + ], + [ + { + "type": "title", + "bbox": [ + 0.091, + 0.091, + 0.409, + 0.108 + ], + "angle": 0, + "content": "2.3. Meta-Learning Activation Functions" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.113, + 0.484, + 0.265 + ], + "angle": 0, + "content": "Our goal is to optimize the inductive biases of a neural network and recent work [96] showed that the activation functions are the most important component. The typical approach to learn activations [2, 5, 6, 11, 13, 22, 45, 82, 91] (see Related Work) replaces them with a small shared ReLU MLP that implements an \\(\\mathbb{R}\\rightarrow \\mathbb{R}\\) function. Its parameters are optimized along the network's. However this cannot discover truly novel activations because the embedded ReLU MLP has itself a simplicity bias and activations are optimized together with the model. We propose instead:" + }, + { + "type": "text", + "bbox": [ + 0.091, + 0.265, + 0.483, + 0.28 + ], + "angle": 0, + "content": "- an unbiased parametrization of the activations as splines," + }, + { + "type": "text", + "bbox": [ + 0.092, + 0.281, + 0.451, + 0.295 + ], + "angle": 0, + "content": "- a bi-level optimization to learn reusable activations," + }, + { + "type": "text", + "bbox": [ + 0.092, + 0.297, + 0.483, + 0.327 + ], + "angle": 0, + "content": "- an episodic training to optimize for generalization rather than simply to fit the training data." + }, + { + "type": "list", + "bbox": [ + 0.091, + 0.265, + 0.483, + 0.327 + ], + "angle": 0, + "content": null + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.33, + 0.484, + 0.481 + ], + "angle": 0, + "content": "Parametrization as splines. We want a space of activation functions free of priors such as the smoothness and monotonicity enforced in prior work [5, 13]. We implement an activation \\( g_{\\psi} : \\mathbb{R} \\to \\mathbb{R} \\) as a linear spline with control points defined by \\( \\psi \\). We define \\( n_c \\) points spread regularly in an interval \\( [a, b] \\), typically \\( \\sim 50 \\) points in \\( [-5, +5] \\). Then \\( g \\) represents piecewise linear segments interpolating values specified in the learned parameters \\( \\psi := [g_{\\psi}(a), \\dots, g_{\\psi}(b)] \\in \\mathbb{R}^{n_c} \\). \\( g \\) can represent simple and complex functions, including smooth curves, periodic functions, sharp transitions, etc." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.485, + 0.484, + 0.742 + ], + "angle": 0, + "content": "Bi-level optimization & episodic training. Our goal is to get an activation function that can be reused like any other in subsequent training runs. This differs from prior work (e.g. [2]) that continuously updates the activation during training: the final one may not be suitable to start training with. Our solution is a bi-level meta-learning loop. An inner loop trains the model with a fixed activation function. An outer loop trains the activation function to maximize generalization. Each outer step simulates a new learning task or episode. This means (1) initializing the model with different weights and (2) using different subsets of data for training and validation. With suitable choices, this can simulate in- or out-of-distribution conditions (see Section 3.4). Without episodes, the learned activation could overfit to a particular model initialization for example, and would not generalize in subsequent training runs. The method is outlined as Algorithm 1. Its implementation is discussed in Appendix C." + }, + { + "type": "image", + "bbox": [ + 0.104, + 0.753, + 0.141, + 0.779 + ], + "angle": 0, + "content": null + }, + { + "type": "text", + "bbox": [ + 0.147, + 0.752, + 0.472, + 0.813 + ], + "angle": 0, + "content": "Inductive bias and simplicity bias are not interchangeable. Our method optimizes toward bethepneralization. Simplicity is only one aspect of the 1d models that we analyze post-hoc (e.g. Figure 4)." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.83, + 0.266, + 0.846 + ], + "angle": 0, + "content": "3. Tasks and Results" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.856, + 0.484, + 0.901 + ], + "angle": 0, + "content": "We now examine tasks that we hypothesized to be ill-suited to the simplicity bias of ReLU architectures. The intuition is that the target function to learn (e.g. optimal classifier) con" + }, + { + "type": "code_caption", + "bbox": [ + 0.516, + 0.091, + 0.885, + 0.107 + ], + "angle": 0, + "content": "Algorithm 1 Meta-learning an activation function (AF)." + }, + { + "type": "algorithm", + "bbox": [ + 0.517, + 0.111, + 0.892, + 0.363 + ], + "angle": 0, + "content": "Input: training data \\(\\mathcal{T}\\) ; untrained neural model \\(f_{\\theta ,\\psi}\\) Initialize \\(\\psi\\) with zeros Parametrization of AF \\(n_{\\mathrm{tr}}\\gets 0\\) Number of inner-loop iterations while \\(n_{\\mathrm{tr}} < n_{\\mathrm{tr}}^{\\mathrm{max}}\\) Outer loop: train AF Increment \\(n_{\\mathrm{tr}}\\) Sample the episode's tr. \\((T^{\\prime})\\) and val. \\((\\mathcal{V})\\) sets from T Initialize \\(\\theta\\) randomly Model weights and biases for \\(n_{\\mathrm{tr}}\\) steps Inner loop: train model with fixed AF Eval. loss on \\(\\mathcal{T}':L\\gets \\Sigma_{(\\pmb {x},\\pmb {y})\\in \\mathcal{T}'}\\mathcal{L}\\big(f_{\\pmb {\\theta},\\pmb {\\psi}}(\\pmb {x},\\pmb {y})\\big)\\) Gradient step on weights/biases: \\(\\pmb {\\theta}\\gets \\mathrm{GD}(\\pmb {\\theta},\\nabla_{\\pmb{\\theta}}L)\\) Eval. loss on V: \\(L\\gets \\Sigma_{(\\pmb {x},\\pmb {y})\\in \\mathcal{V}}\\mathcal{L}\\big(f_{\\pmb {\\theta},\\pmb {\\psi}}(\\pmb {x},\\pmb {y})\\big)\\) Gradient step on AF: \\(\\psi \\gets \\mathrm{GD}(\\psi ,\\nabla_{\\psi}L)\\) if performance on \\(\\nu\\) worsens then break Early stopping" + }, + { + "type": "text", + "bbox": [ + 0.52, + 0.366, + 0.693, + 0.382 + ], + "angle": 0, + "content": "Output: optimized AF \\(\\psi\\)" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.398, + 0.907, + 0.504 + ], + "angle": 0, + "content": "tains sharp transitions (regression tasks, tabular datasets), or repeating patterns (algorithmic tasks) that contradict the ReLUs' simplicity bias. For each task, we examine existing datasets with the tools from Section 2. In all cases, we find benefits from architectures whose inductive biases favor more complex functions. Additional details and results are provided in Appendix E." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.512, + 0.755, + 0.528 + ], + "angle": 0, + "content": "3.1. Image Classification Tasks" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.534, + 0.907, + 0.625 + ], + "angle": 0, + "content": "Background. We start with classical datasets to validate our methodology: MNIST, FASHION-MNIST, SVHN, CIFAR-10 [49, 52, 68, 102]. They are representative of the vision tasks that guided the development of deep learning. Our hypothesis is therefore that the inductive biases of modern architectures and ReLUs are well suited to these datasets." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.629, + 0.909, + 0.81 + ], + "angle": 0, + "content": "Setup. For each dataset, we learn activation functions with Algorithm 1. We experiment with two initializations of the spline parameters: as zeros and so as to mimic a ReLU. The goal of the latter is to explore the space of functions similar to ReLUs. Because of the difficulty of the optimization, the algorithm is likely to converge to a local optimum similar to ReLUs if there is one. We also experiment with the sharing of the activation function. By default, a single function is shared across the network. Alternatively, we learn a different activation function per layer. This provides more ways to affect the model's inductive biases. Our base architecture is a 3-layer MLP (details in Appendix E)." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.814, + 0.907, + 0.888 + ], + "angle": 0, + "content": "Results. We compare in Figure 2a the accuracy of models with ReLUs vs. learned activation functions. Differences are small. The learned activations only improve slightly on SVHN and CIFAR. This suggests that the inductive biases of ReLUs are generally well suited to these datasets." + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.89, + 0.906, + 0.905 + ], + "angle": 0, + "content": "We examine the learned activations in Figure 3a. With" + }, + { + "type": "page_number", + "bbox": [ + 0.49, + 0.946, + 0.507, + 0.957 + ], + "angle": 0, + "content": "81" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.095, + 0.079, + 0.23, + 0.183 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.231, + 0.079, + 0.361, + 0.183 + ], + "angle": 0, + "content": null + }, + { + "type": "image_footnote", + "bbox": [ + 0.362, + 0.109, + 0.49, + 0.141 + ], + "angle": 0, + "content": "- ReLU activations\n- Learned act., ReLU init.\n- Learned act., zero init." + }, + { + "type": "image_caption", + "bbox": [ + 0.09, + 0.187, + 0.483, + 0.27 + ], + "angle": 0, + "content": "Figure 2. Test accuracy on image datasets. (a) For classification tasks, all models perform similarly, suggesting that the inductive biases of ReLUs are well suited to these datasets. (b) For regression tasks, models with learned activations perform better, especially from an initialization as zeros, which enables the discovery of completely novel activation functions." + }, + { + "type": "image", + "bbox": [ + 0.092, + 0.277, + 0.468, + 0.452 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.09, + 0.457, + 0.483, + 0.526 + ], + "angle": 0, + "content": "Figure 3. Activation functions learned for MNIST. For a classification task, the activation learned from a ReLU resembles the popular GeLUs. For a regression task, the learned activations contain irregularities that help a network represent complex functions. See Figure 16 for similar results on other datasets." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.542, + 0.484, + 0.65 + ], + "angle": 0, + "content": "an initialization as ReLUs, the optimization converges to a smooth variant remarkably similar to GeLUs [11] which are widely used. This suggests that the research community has empirically converged on a local optimum in the space of activation functions. With an initialization as zeros, we discover wavelets [80] that are unlike common activations but perform as well as ReLUs, i.e. another local optimum." + }, + { + "type": "text", + "bbox": [ + 0.102, + 0.66, + 0.473, + 0.737 + ], + "angle": 0, + "content": "Take-away: for image classification, learned activations provide very little benefit over ReLUs. Smooth variants of ReLUs are a local optimum in the space of activations. ReLUs' popularity for such tasks could thus be explained with their proximity to this optimum." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.751, + 0.26, + 0.767 + ], + "angle": 0, + "content": "3.2. Regression Tasks" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.776, + 0.483, + 0.882 + ], + "angle": 0, + "content": "Background. Regression tasks are known to be difficult for NNs [90]. They are often turned into a classification through discretization [24, 43]. Existing explanations that invoke implicit biases of gradient descent are clearly incomplete [90]. Our hypothesis is that regression is difficult because it often involves irregular decision boundaries [35] in opposition to the typical solutions of ReLU networks [17]." + }, + { + "type": "text", + "bbox": [ + 0.091, + 0.886, + 0.483, + 0.901 + ], + "angle": 0, + "content": "Setup. We use the same setup and image datasets as Sec-" + }, + { + "type": "image", + "bbox": [ + 0.559, + 0.079, + 0.707, + 0.193 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.724, + 0.079, + 0.861, + 0.193 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.512, + 0.195, + 0.907, + 0.348 + ], + "angle": 0, + "content": "Figure 4. Accuracy vs. complexity on image datasets. Each marker is a model with different hyperparameters and ReLUs \\((\\bullet)\\) or learned activations initialized as ReLUs \\((\\square)\\) or as zeros \\((\\text{圆})\\). For classification (a), ReLUs are close to best. Activations optimized from ReLUs only improve the accuracy slightly, corresponding to the GeLU-like function in Figure 3. For regression (b), new activations (learned from zeros) are best. Moreover, accuracy and complexity are clearly correlated only for regression. This supports the hypothesis that regression is more complex than classification and thus benefits from alternatives to the ReLUs' simplicity bias. See Figure 18 for similar results on other datasets." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.365, + 0.906, + 0.426 + ], + "angle": 0, + "content": "tion 3.1. The task is now to directly predict class IDs. E.g. for MNIST this means predicting digit values. Models are trained with an MSE loss. To measure accuracy, we discretize the predictions to the nearest class ID." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.436, + 0.906, + 0.542 + ], + "angle": 0, + "content": "Results. The first observation from Figure 2b is that regression is clearly more difficult for NNs than classification (lower accuracies) despite the identical underlying task. Importantly, the learned activations now provide clear improvements, especially when learned from scratch (initialization as zeros). This confirms the hypothesis that the inductive biases of RLUs are not well suited to these tasks." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.551, + 0.906, + 0.703 + ], + "angle": 0, + "content": "Figure 3b shows that the learned activations contain more irregularities for regression than classification. Prior work [96] showed that this can help models represent complex functions with sharp transitions. An analysis of the complexity of trained models (Figures 4 and 18) shows that the accuracy is correlated with complexity for regression but not classification. And regression models with learned activations implement functions of higher complexity than with ReLUs. This supports the claim that the improvements arise from overcoming the simplicity bias of ReLUs." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.713, + 0.906, + 0.82 + ], + "angle": 0, + "content": "Complexity is only one dimension of the inductive biases. The complexity plots for SVHN (Figure 18) interestingly show that models with ReLUs and learned activations get different accuracies at the same complexity level. This shows that our meta learning approach can search over dimensions of the inductive biases that are not captured by our complexity measure, and are yet to be explicitly studied." + }, + { + "type": "text", + "bbox": [ + 0.524, + 0.831, + 0.895, + 0.892 + ], + "angle": 0, + "content": "Take-away: regression is more difficult for NNs than classification, and the simplicity bias of ReLUs is partly to blame. Learned activations improve performance by helping networks represent more complex functions." + }, + { + "type": "page_number", + "bbox": [ + 0.49, + 0.946, + 0.508, + 0.957 + ], + "angle": 0, + "content": "82" + } + ], + [ + { + "type": "title", + "bbox": [ + 0.092, + 0.091, + 0.232, + 0.106 + ], + "angle": 0, + "content": "3.3. Tabular Data" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.115, + 0.484, + 0.312 + ], + "angle": 0, + "content": "Background. Tabular data is any data with few unstructured dimensions, which often contains low-cardinality variables such as dates or categorical attributes. This contrasts e.g. with images, which contain many correlated, continuous dimensions (pixels). NNs struggle with tabular datasets and are often inferior to decision trees [35, 63]. Our hypothesis is that the inductive biases of standard architectures are ill-suited to such data because of the simplicity bias. It makes it difficult to learn functions where small changes in the input (e.g. day of the week) correspond to abrupt changes in the target — the definition of sensitivity, a proxy for complexity [17]. This seldom occurs in vision where similar images correspond to similar labels." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.316, + 0.484, + 0.437 + ], + "angle": 0, + "content": "Setup. We use 16 real-world classification datasets from Grinsztajn et al. [34, 35]. Bases include a linear classifier, k-NNs, and boosted decision trees. Our models are MLPs with 1-4 hidden layers (details in Appendix E.4). We compare learned activations functions with ReLUs and TanHs with a global prefactor, \\(\\tanh (\\alpha x)\\) with \\(\\alpha \\in \\mathbb{R}^{+}\\) tuned on the validation set. This is a simple option with tunable complexity, albeit with inductive biases of TanHs [45, 96]." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.439, + 0.484, + 0.59 + ], + "angle": 0, + "content": "We also experiment with learned input activation functions (IAFs). The motivation is to learn a different behavior for each input dimension. Since they carry different information, e.g. continuous vs. categorical variables, one could be suited to the simplicity bias while another is not, for example. IAFs are dimension-specific activation functions applied directly on the data before a standard MLP. IAFs are learned like AFs, from an initialization as the identity i.e. no effect by default. They subsume the gated inputs, and Fourier/numerical embeddings from prior work [20, 25, 32]." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.594, + 0.484, + 0.685 + ], + "angle": 0, + "content": "Results. We compare the accuracy of models on the 16 datasets in Figures 5 and 20. Vanilla MLPs generally perform worse than trees. But adjusting the MLPs' inductive biases with learned prefactors or activations eliminates the gap. IAFs perform best, sometimes even surpassing trees. We analyze below the reasons for these improvements." + }, + { + "type": "image", + "bbox": [ + 0.118, + 0.692, + 0.452, + 0.826 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.09, + 0.828, + 0.484, + 0.899 + ], + "angle": 0, + "content": "Figure 5. Comparison of model types over 16 tabular datasets. Vanilla MLPs often perform worse than decision trees, but adjusting their inductive biases with learned activation functions (AFs) eliminates this gap. The input activation functions (IAFs) enable even better performance. See Figure 20 for results per dataset." + }, + { + "type": "image", + "bbox": [ + 0.517, + 0.084, + 0.614, + 0.169 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.529, + 0.173, + 0.612, + 0.182 + ], + "angle": 0, + "content": "Best at low complexity" + }, + { + "type": "image", + "bbox": [ + 0.617, + 0.085, + 0.709, + 0.169 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.62, + 0.173, + 0.743, + 0.181 + ], + "angle": 0, + "content": "Different datasets" + }, + { + "type": "image", + "bbox": [ + 0.715, + 0.085, + 0.805, + 0.169 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.71, + 0.173, + 0.744, + 0.181 + ], + "angle": 0, + "content": "at datasets" + }, + { + "type": "image", + "bbox": [ + 0.809, + 0.085, + 0.899, + 0.169 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.81, + 0.173, + 0.903, + 0.182 + ], + "angle": 0, + "content": "Best at higher complexity \\(\\uparrow\\)" + }, + { + "type": "image_caption", + "bbox": [ + 0.512, + 0.186, + 0.906, + 0.299 + ], + "angle": 0, + "content": "Figure 6. Test accuracy vs. complexity on tabular datasets. Each marker represents a model with different hyperparameters, and ReLUs \\((\\bullet)\\) or learned activations initialized as ReLUs \\((\\square)\\) or as zeros \\((\\square)\\). The learned activations perform better in all cases, but the accuracy peaks at different complexity levels. For some datasets, a low complexity is best and ReLUs thus perform quite well (left-most panel, note the smaller Y scale). For other datasets, the opposite is true and the improvements with learned activations is larger." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.316, + 0.906, + 0.513 + ], + "angle": 0, + "content": "Learned activation functions close the gap to decision trees by mimicking their inductive bias. We visualize in Figure 7c the functions implemented by different models, plotting their output over slices of the input space (Section 2.1). ReLUs produce the smoothest function while TanHs and learned activations induce sharper patterns. Notably, the IAFs induce sharp axis-aligned decision boundaries that are also characteristic of trees, with which they share a high accuracy. Axis-aligned transitions are the consequence of IAFs applied independently to each dimension. Sharp transitions originate from the complex shape of the learned activation function (Figure 7a) which is possible thanks to the unbiased spline parametrization (Section 2.3)." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.516, + 0.906, + 0.789 + ], + "angle": 0, + "content": "The simplicity/complexity bias is a property of the architecture. We visualize complexity landscapes of MLPs in Figure 7b. Similarly to standard loss landscapes [54], we plot model complexity over 2D slices of the parameter space. A first global view over a plane aligned with the training trajectory shows that complexity steadily increases during training for all models [48, 75, 104] but does so to the highest level for the best model (IAFs). A second view zooms in on each optimized solution in a random 2D plane. This examines the effect of arbitrary perturbations to the parameters. It shows that the ambient complexity of perturbed solutions of the best model is much higher than the solution itself, and then with less accurate models. This means that this architecture is more likely to represent complex functions because they are more abundant in parameter space [65, 84]. This is why the simplicity bias can be overcome: it results from architecture choices and not from an inevitable \"implicit bias\" of SGD [85, 89, 98, 105]." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.791, + 0.905, + 0.868 + ], + "angle": 0, + "content": "Different tabular datasets require different inductive biases. We examine the relation between accuracy and complexity in Figure 6. The accuracy peaks at different complexity levels for different datasets. For some, a low complexity is best and ReLU MLPs perform well. For oth" + }, + { + "type": "page_footnote", + "bbox": [ + 0.513, + 0.875, + 0.905, + 0.901 + ], + "angle": 0, + "content": "This resembles an analysis of untrained models [15, 64, 96, 99] but focuses on relevant regions on the parameter space, near optimized models." + }, + { + "type": "page_number", + "bbox": [ + 0.49, + 0.946, + 0.508, + 0.956 + ], + "angle": 0, + "content": "83" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.095, + 0.092, + 0.624, + 0.104 + ], + "angle": 0, + "content": "(a) Activation function & loss landscape (—training trajectory, \\(\\star\\) early-stopping checkpoint, \\(\\bullet\\) last checkpoint)" + }, + { + "type": "image", + "bbox": [ + 0.114, + 0.105, + 0.241, + 0.153 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.276, + 0.105, + 0.403, + 0.153 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.437, + 0.105, + 0.563, + 0.153 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.599, + 0.105, + 0.725, + 0.153 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.777, + 0.127, + 0.869, + 0.145 + ], + "angle": 0, + "content": null + }, + { + "type": "text", + "bbox": [ + 0.095, + 0.16, + 0.645, + 0.172 + ], + "angle": 0, + "content": "(b) Complexity landscape (in weight space along the PCA plane of the training trajectory) & zoom-in (random plane)" + }, + { + "type": "image", + "bbox": [ + 0.114, + 0.173, + 0.238, + 0.222 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.275, + 0.173, + 0.4, + 0.222 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.437, + 0.173, + 0.562, + 0.222 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.599, + 0.173, + 0.723, + 0.222 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.776, + 0.196, + 0.869, + 0.214 + ], + "angle": 0, + "content": null + }, + { + "type": "text", + "bbox": [ + 0.095, + 0.229, + 0.662, + 0.24 + ], + "angle": 0, + "content": "(c) Function implemented by the network (in input space along four random planes containing each one training point \\(\\bullet\\)" + }, + { + "type": "image", + "bbox": [ + 0.114, + 0.241, + 0.238, + 0.337 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.143, + 0.34, + 0.207, + 0.35 + ], + "angle": 0, + "content": "MLP, ReLU" + }, + { + "type": "image", + "bbox": [ + 0.275, + 0.241, + 0.4, + 0.337 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.259, + 0.34, + 0.416, + 0.351 + ], + "angle": 0, + "content": "MLP, TanH w/tuned prefactor" + }, + { + "type": "image", + "bbox": [ + 0.437, + 0.241, + 0.562, + 0.337 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.454, + 0.34, + 0.544, + 0.35 + ], + "angle": 0, + "content": "MLP, learned AF" + }, + { + "type": "image", + "bbox": [ + 0.599, + 0.241, + 0.723, + 0.337 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.611, + 0.34, + 0.711, + 0.35 + ], + "angle": 0, + "content": "MLP, learned IAFs" + }, + { + "type": "image", + "bbox": [ + 0.76, + 0.241, + 0.886, + 0.337 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.768, + 0.34, + 0.878, + 0.35 + ], + "angle": 0, + "content": "Boosted decision trees" + }, + { + "type": "image_footnote", + "bbox": [ + 0.321, + 0.356, + 0.691, + 0.367 + ], + "angle": 0, + "content": "Increasing accuracy Best MLP" + }, + { + "type": "image_caption", + "bbox": [ + 0.09, + 0.37, + 0.907, + 0.44 + ], + "angle": 0, + "content": "Figure 7. Models trained on the ELECTRICITY [34] tabular dataset. ReLU MLPs perform worst (left). TanHs induce sharper transitions in the network's function (c). So does the learned activation function (a) which is itself very irregular. The input activation functions (IAFs) perform best and mimic the axis-aligned boundaries of trees (bottom-right). The complexity landscapes (b) show that complexity increases to the highest level in the best model (IAFs). The zoom-in shows that the ambient complexity is also much higher than in other models. This means that it is inherently more likely to represent complex functions since they are more abundant in parameter space." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.455, + 0.484, + 0.546 + ], + "angle": 0, + "content": "ers, a higher complexity is best and the improvements with learned activations are larger. This supports the hypothesis that improvements over ReLU MLPs come from overcoming their simplicity bias. The variance across datasets is also unsurprising since they have little in common besides their low dimensionality (full results in Appendix E.4)." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.549, + 0.484, + 0.701 + ], + "angle": 0, + "content": "Effect of width and depth. We show in Figure 8 that the learned activations can be reused in networks of different widths than they were trained for. The accuracy varies with width similarly as with ReLUs. Teney et al. [96] indeed showed that a model's width affects its capacity but not its inductive biases. Therefore width does not interfere with the effects of the learned activations. Figures 8 and 21 also show that good performance can be achieved with fewer layers than with ReLUs. Learned activations might thus have utility in model compression and distillation." + }, + { + "type": "text", + "bbox": [ + 0.102, + 0.712, + 0.473, + 0.787 + ], + "angle": 0, + "content": "Take-away: many tabular datasets are ill-suited to ReLU models because they require learning a complex function. Learned activations improve accuracy by implementing sharp axis-aligned decision boundaries that mimic the inductive biases of decision trees." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.801, + 0.271, + 0.817 + ], + "angle": 0, + "content": "3.4. Shortcut Learning" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.825, + 0.484, + 0.902 + ], + "angle": 0, + "content": "Background. Shortcut learning occurs when a model learns spurious features instead of generalizable ones. It is a known consequence of the simplicity bias [85, 93] when the training data contains multiple features of different complexity. Our hypothesis is that the preference for some fea" + }, + { + "type": "image", + "bbox": [ + 0.584, + 0.452, + 0.832, + 0.559 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.513, + 0.565, + 0.907, + 0.607 + ], + "angle": 0, + "content": "Figure 8. The learned activation functions surpass ReLUs, often with fewer layers. They can also be reused with different network widths (COVERTYPE [34] tabular dataset, see Figure 21 for others)." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.63, + 0.906, + 0.674 + ], + "angle": 0, + "content": "tures depends on their alignment with the inductive biases. We will evaluate whether this can be controlled with activation functions." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.683, + 0.907, + 0.819 + ], + "angle": 0, + "content": "Setup. We use MNIST/CIFAR collages [85, 93, 94], a classification task over images combining tiles from MNIST and CIFAR-10. The training set is ambiguous: both tiles are predictive of the labels. Two unambiguous test sets evaluate reliance on either tile: one is predictive, the other contains a random class. We similarly build two validation sets to learn two activation functions optimized for either tile. We simulate OOD conditions by setting \\(\\mathcal{V}\\) in Algorithm 1. The models are the fully-connected MLPs used in [93]." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.825, + 0.907, + 0.902 + ], + "angle": 0, + "content": "Results. Figure 9 shows that a baseline with ReLUs is prone to shortcut learning. It relies exclusively on MNIST and the accuracy on the CIFAR test set is not better than chance (10%). In comparison, using either learned activation steers the learning towards either tile. The accuracy" + }, + { + "type": "page_number", + "bbox": [ + 0.49, + 0.946, + 0.508, + 0.956 + ], + "angle": 0, + "content": "84" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.107, + 0.09, + 0.171, + 0.142 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.095, + 0.15, + 0.181, + 0.173 + ], + "angle": 0, + "content": "Ambiguous training images" + }, + { + "type": "image", + "bbox": [ + 0.184, + 0.089, + 0.257, + 0.145 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.192, + 0.15, + 0.25, + 0.173 + ], + "angle": 0, + "content": "Optimized for MNIST" + }, + { + "type": "image", + "bbox": [ + 0.261, + 0.089, + 0.334, + 0.145 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.268, + 0.15, + 0.328, + 0.172 + ], + "angle": 0, + "content": "Optimized for CIFAR" + }, + { + "type": "image_caption", + "bbox": [ + 0.336, + 0.093, + 0.547, + 0.142 + ], + "angle": 0, + "content": "Upper bound: ReLU trained on unambiguous MNIST Activation optimized for MNIST ReLU Baseline Activation optimized for CIFAR Upper bound: ReLU trained on unambiguous CIFAR" + }, + { + "type": "image", + "bbox": [ + 0.547, + 0.09, + 0.798, + 0.168 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.799, + 0.089, + 0.9, + 0.167 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.09, + 0.183, + 0.908, + 0.225 + ], + "angle": 0, + "content": "Figure 9. Experiments on shortcut learning with MNIST/CIFAR collages. The ReLU baseline (■) relies mostly on simple MNIST features. We learn two activation functions that shift the preference towards different features \\((\\leftarrow / \\rightarrow)\\). Training trajectories (right) clearly differ with the activation optimized for CIFAR (—), for MNIST (—), or a ReLU (==). The model at initialization (random weights) is marked with \\(\\bullet\\)." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.239, + 0.484, + 0.314 + ], + "angle": 0, + "content": "shifts towards either of two tiles as the model prioritizes different features, merely with a change of activation function. This shows that the simplicity bias is not an inevitable effect of SGD. Instead, it directly reflects the alignment between the chosen architecture and the data." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.318, + 0.484, + 0.455 + ], + "angle": 0, + "content": "Training dynamics. The accuracy on CIFAR remains below a model trained on unambiguous CIFAR data. This is because training dynamics are also important. In Figure 9 (right), we plot the accuracy on the two tiles for the whole training trajectory. The reliance on different features varies, and the model eventually relies primarily on simple ones with enough iterations (i.e. without early stopping). This calls for future work combining our findings with the extensive literature on ID/OOD training dynamics [46, 95, 98]." + }, + { + "type": "text", + "bbox": [ + 0.102, + 0.465, + 0.473, + 0.527 + ], + "angle": 0, + "content": "Take-away: we confirm that shortcut learning is a side effect of the simplicity bias. Different activation functions, while not completely avoiding shortcut learning, can steer the learning towards particular input features." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.541, + 0.383, + 0.558 + ], + "angle": 0, + "content": "3.5. Algorithmic Tasks and Grokking" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.565, + 0.484, + 0.822 + ], + "angle": 0, + "content": "Background. Grokking is a phenomenon where a model first overfits the data (i.e. high training accuracy, low test accuracy) then shifts to high test accuracy after many training steps [73]. This is typically observed on algorithmic tasks and architectures from MLPs to transformers. Our hypothesis is that grokking is due to a mismatch between the target function and the model's inductive biases. Indeed, typical architectures were not developed for the algorithmic tasks where grokking is typically observed. To verify this hypothesis, we will show that endowing an architecture with the right inductive biases, using learned activation functions, can eliminate the phenomenon. Supporting this hypothesis, Zhou et al. [110] proposed that grokking comes from the frequency principle (i.e. low frequencies learned first by SGD), and Kumar et al. [51] showed that it correlates with a misalignment between features at initialization and the target function." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.825, + 0.484, + 0.903 + ], + "angle": 0, + "content": "Setup. Following [36, 51, 58] we train 1-hidden layer MLPs on algorithmic tasks, defined each by one binary operation (Figures 10 and 27). e.g. \\( y = (x_{1} + x_{2}) \\mod 13 \\). The operands are passed as one-hot vectors and the task is a classification over possible outputs. Details in Appendix E.6." + }, + { + "type": "image", + "bbox": [ + 0.516, + 0.238, + 0.58, + 0.31 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.584, + 0.238, + 0.74, + 0.31 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.742, + 0.238, + 0.839, + 0.31 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.84, + 0.238, + 0.904, + 0.31 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.513, + 0.316, + 0.906, + 0.358 + ], + "angle": 0, + "content": "Figure 10. Target functions used to investigate grokking [73] (details in Appendix, Figure 27). These patterns are very different from the tasks for which typical architectures were developed." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.363, + 0.907, + 0.576 + ], + "angle": 0, + "content": "Results. We compare models with ReLU vs. learned activations across various tasks, network widths, and fractions of training data. We find that the learned, task-specific activations lead to faster convergence and/or higher test accuracy (Figures 11-13). On modular addition (a common task in the grokking literature) the learned-activation model converges \\(\\sim 10\\times\\) faster than ReLUs. Curiously, some models with learned activations also end up overfitting (decreasing test accuracy) with prolonged training. In contrast, ReLU networks either never generalize (test accuracy \\(\\sim 0\\)) or grok and keep a high accuracy indefinitely. Further investigation is needed to explain this difference. We examine learned activation functions in Figure 12. See Figure 28 in the Appendix for results on other algorithmic tasks [73]." + }, + { + "type": "image", + "bbox": [ + 0.518, + 0.581, + 0.742, + 0.643 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.743, + 0.581, + 0.901, + 0.643 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.513, + 0.646, + 0.906, + 0.688 + ], + "angle": 0, + "content": "Figure 11. The learned activations essentially eliminate grokking (delayed convergence). On the above task (addition mod 27), our model converges \\(\\sim 10\\times\\) faster than ReLUs." + }, + { + "type": "image", + "bbox": [ + 0.528, + 0.697, + 0.603, + 0.734 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.625, + 0.697, + 0.7, + 0.734 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.721, + 0.697, + 0.796, + 0.734 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.818, + 0.697, + 0.893, + 0.734 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.513, + 0.735, + 0.906, + 0.763 + ], + "angle": 0, + "content": "Figure 12. Activations learned for modular addition. The frequency of the sine-like function varies across versions of the task." + }, + { + "type": "image", + "bbox": [ + 0.558, + 0.772, + 0.713, + 0.851 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.725, + 0.771, + 0.816, + 0.851 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.513, + 0.857, + 0.906, + 0.885 + ], + "angle": 0, + "content": "Figure 13. Models with learned activations \\((-)\\) converge faster than ReLUs \\((--)\\) across a variety of settings (addition mod 27)." + }, + { + "type": "page_number", + "bbox": [ + 0.49, + 0.946, + 0.508, + 0.957 + ], + "angle": 0, + "content": "85" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.103, + 0.094, + 0.471, + 0.139 + ], + "angle": 0, + "content": "Take-away: learned activations eliminate grokking in all our cases, suggesting, as a cause, the mismatch between the data and the architectures' inductive biases." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.156, + 0.47, + 0.172 + ], + "angle": 0, + "content": "4. Do the Activations Transfer Across Tasks?" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.181, + 0.483, + 0.242 + ], + "angle": 0, + "content": "So far, we used dataset-specific activation functions and found that there exist better alternatives to ReLUs. A practical application would be the learning of activation functions suitable to a broad task, or range of related datasets." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.243, + 0.484, + 0.455 + ], + "angle": 0, + "content": "As a first step, we study the specialization of the activations functions (AFs) learned for the 22 algorithmic tasks from Figure 27 [73]. We evaluate every task/activation combination, yielding the \\(22 \\times 22\\) matrix of Figure 14. The learned activations do transfer, with improvements in accuracy and convergence shared across tasks. We also evaluate an activation learned on all tasks simultaneously. The accuracy across tasks (i.e. per-column average) reaches \\(61.5\\%\\) vs. only \\(19.9\\%\\) for ReLUs. and \\(54.0\\%\\) on average for tasks-specific solutions. This procedure can thus improve performance on a range of related tasks. Future work could leverage it to discover activation functions that improve performance in other specific domains. See Appendix E.3 for other transfer experiments using image regression tasks." + }, + { + "type": "image", + "bbox": [ + 0.092, + 0.458, + 0.485, + 0.619 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.09, + 0.624, + 0.483, + 0.693 + ], + "angle": 0, + "content": "Figure 14. Transfer of AFs (columns) across algorithmic tasks (rows). Colors represent the fraction of the best convergence speed or accuracy per task (brighter is better). If the activations were over-specialized, the matrix would be diagonal. On the contrary, it is densely filled, indicating positive transfer across many tasks." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.71, + 0.205, + 0.725 + ], + "angle": 0, + "content": "5. Discussion" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.735, + 0.484, + 0.901 + ], + "angle": 0, + "content": "We used activation functions as a tool to show that there exist a variety of inductive biases that are useful across applications of NNs. The impossibility of universal inductive biases is well known [101] but a strong argument has also been made that deep learning research is converging towards few architectures with wide applicability [31]. This argument rests on the NNs' simplicity bias being a good match for real-world data [12, 18, 56]. Our results do not invalidate these assumptions: NNs are widely applicable and their simplicity bias is evidently very effective on average. Our results show instead the following." + }, + { + "type": "text", + "bbox": [ + 0.515, + 0.092, + 0.907, + 0.152 + ], + "angle": 0, + "content": "1. There exist real-world tasks where the inductive biases of typical architectures' are suboptimal. This explanation connects four domains where NNs historically struggled." + }, + { + "type": "text", + "bbox": [ + 0.514, + 0.155, + 0.907, + 0.245 + ], + "angle": 0, + "content": "2. The simplicity bias in modern NNs depends on particular design choices, the activation functions in particular. Research has converged on these choices by trial and error, in large part by optimizing performance on vision tasks. Therefore the adequacy of ReLUs for image classification (Section 3.1) is not accidental." + }, + { + "type": "list", + "bbox": [ + 0.514, + 0.092, + 0.907, + 0.245 + ], + "angle": 0, + "content": null + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.25, + 0.907, + 0.371 + ], + "angle": 0, + "content": "Relevance to transformers and language models. The simplicity bias exists in transformers [10, 109] and language models [4, 31, 96]. Their embedding layer resembles input activations functions (Section 3.3). Could this explain the transformers' remarkable flexibility? I.e. a simplicity bias on an initial mapping of arbitrary complexity. Zhong and Andreas [108] indeed trained embeddings alone in a random-weight transformer and could learn complex tasks." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.375, + 0.909, + 0.497 + ], + "angle": 0, + "content": "Limitations and open questions. We prioritized breadth by establishing a new connection across multiple disparate topics in machine learning. Each section could expand into its own paper with additional models, datasets, comparisons, etc. Our findings on shortcut learning for example (Section 3.4) could yield new methods to address distribution shifts, though no such claim is made here. Here are the most promising follow-up questions opened by this paper." + }, + { + "type": "text", + "bbox": [ + 0.515, + 0.501, + 0.906, + 0.561 + ], + "angle": 0, + "content": "- How to fully characterize inductive biases? We focused on simplicity for its prevalence in AI [31], philosophy [72], and the natural sciences [12]. But it is only one dimension among many to characterize inductive biases." + }, + { + "type": "text", + "bbox": [ + 0.515, + 0.565, + 0.907, + 0.64 + ], + "angle": 0, + "content": "- Can we improve state-of-the-art architectures? We used simple MLPs to isolate the effects of activations functions since they are central to the simplicity bias [96]. But other existing mechanisms (architectural, optimization) may already tweak or attenuate the simplicity bias." + }, + { + "type": "text", + "bbox": [ + 0.515, + 0.642, + 0.907, + 0.748 + ], + "angle": 0, + "content": "- Can we learn transferable activation functions for other domains? We examined transferability in Sections 4 and E.3. The results suggest the possibility of better architectures optimized for specific domains. Predicting the suitability of an architecture/dataset pair ex ante (prior to training) would be extremely useful. This may follow from advances on the first open question above." + }, + { + "type": "text", + "bbox": [ + 0.515, + 0.75, + 0.909, + 0.902 + ], + "angle": 0, + "content": "- Are there other detrimental effects of the simplicity bias? Any learning algorithm needs inductive biases to \"fill the gaps\" between training examples. The better they are, the fewer examples are needed. Researching what inductive biases are most useful on real-world tasks might thus hold the key for machine learning to become as data-efficient as humans. More speculatively, high-level cognition has been argued to require postulating explanations beyond the data [16, 23]. In this regard, simplicity-biased architectures might also hold us back." + }, + { + "type": "list", + "bbox": [ + 0.515, + 0.501, + 0.909, + 0.902 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.491, + 0.946, + 0.508, + 0.956 + ], + "angle": 0, + "content": "86" + } + ], + [ + { + "type": "title", + "bbox": [ + 0.093, + 0.09, + 0.188, + 0.106 + ], + "angle": 0, + "content": "References" + }, + { + "type": "ref_text", + "bbox": [ + 0.108, + 0.113, + 0.484, + 0.168 + ], + "angle": 0, + "content": "[1] Sravanti Addepalli, Anshul Nasery, Venkatesh Babu Radhakrishnan, Praneeth Netrapalli, and Prateek Jain. Feature reconstruction from outputs can mitigate simplicity bias in neural networks. In ICLR, 2022. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.107, + 0.17, + 0.483, + 0.211 + ], + "angle": 0, + "content": "[2] Konstantinos Panagiotis Alexandridis, Jiankang Deng, Anh Nguyen, and Shan Luo. Adaptive parametric activation. arXiv preprint arXiv:2407.08567, 2024. 2, 3, 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.107, + 0.213, + 0.483, + 0.254 + ], + "angle": 0, + "content": "[3] Konstantinos Panagiotis Alexandridis, Jiankang Deng, Anh Nguyen, and Shan Luo. Adaptive parametric activation. In ECCV, 2025. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.107, + 0.257, + 0.483, + 0.311 + ], + "angle": 0, + "content": "[4] Badr AlKhamissi, Greta Tuckute, Antoine Bosselut, and Martin Schrimpf. Brain-like language processing via a shallow untrained multihead attention network. arXiv preprint arXiv:2406.15109, 2024. 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.107, + 0.314, + 0.483, + 0.355 + ], + "angle": 0, + "content": "[5] Andrea Apicella, Francesco Isgro, and Roberto Prevete. A simple and efficient architecture for trainable activation functions. Neurocomputing, 2019. 3, 1, 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.107, + 0.357, + 0.483, + 0.398 + ], + "angle": 0, + "content": "[6] Andrea Apicella, Francesco Donnarumma, Francesco Isgro, and Roberto Prevete. A survey on modern trainable activation functions. Neural Networks, 2021. 2, 3, 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.107, + 0.401, + 0.483, + 0.44 + ], + "angle": 0, + "content": "[7] Sanjeev Arora, Nadav Cohen, Wei Hu, and Yuping Luo. Implicit regularization in deep matrix factorization. NeurIPS, 2019. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.107, + 0.443, + 0.483, + 0.511 + ], + "angle": 0, + "content": "[8] Devansh Arpit, Stanislaw Jastrzebski, Nicolas Ballas, David Krueger, Emmanuel Bengio, Maxinder S Kanwal, Tegan Maharaj, Asja Fischer, Aaron Courville, Yoshua Bengio, et al. A closer look at memorization in deep networks. In ICML. PMLR, 2017. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.107, + 0.514, + 0.484, + 0.569 + ], + "angle": 0, + "content": "[9] Samuel James Bell and Levent Sagun. Simplicity bias leads to amplified performance disparities. In Proceedings of the 2023 ACM Conference on Fairness, Accountability, and Transparency, pages 355-369, 2023. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.101, + 0.571, + 0.483, + 0.625 + ], + "angle": 0, + "content": "[10] Satwik Bhattachamishra, Arkil Patel, Varun Kanade, and Phil Blunsom. Simplicity bias in transformers and their ability to learn sparse boolean functions. arXiv preprint arXiv:2211.12316, 2022.8" + }, + { + "type": "ref_text", + "bbox": [ + 0.101, + 0.628, + 0.483, + 0.682 + ], + "angle": 0, + "content": "[11] Garrett Bingham, William Macke, and Risto Miikkulainen. Evolutionary optimization of deep learning activation functions. In Genetic and Evolutionary Computation Conference, 2020. 3, 4, 1, 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.101, + 0.686, + 0.483, + 0.712 + ], + "angle": 0, + "content": "[12] Mark Buchanan. A natural bias for simplicity. Nature Physics, 2018. 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.101, + 0.715, + 0.483, + 0.755 + ], + "angle": 0, + "content": "[13] Irit Chelly, Shahaf E Finder, Shira Ifergane, and Oren Freifeld. Trainable highly-expressive activation functions. arXiv preprint arXiv:2407.07564, 2024. 3, 1, 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.101, + 0.758, + 0.483, + 0.799 + ], + "angle": 0, + "content": "[14] Nadav Cohen and Amnon Shashua. Inductive bias of deep convolutional networks through pooling geometry. arXiv preprint arXiv:1605.06743, 2016. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.101, + 0.801, + 0.483, + 0.841 + ], + "angle": 0, + "content": "[15] Giacomo De Palma, Bobak Kiani, and Seth Lloyd. Random deep neural networks are biased towards simple functions. NeurIPS, 2019. 1, 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.101, + 0.844, + 0.483, + 0.871 + ], + "angle": 0, + "content": "[16] David Deutsch. The beginning of infinity: Explanations that transform the world. Penguin, 2011. 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.101, + 0.873, + 0.483, + 0.901 + ], + "angle": 0, + "content": "[17] Benoit Dherin, Michael Munn, Mihaela Rosca, and David Barrett. Why neural networks find simple solutions: The" + }, + { + "type": "list", + "bbox": [ + 0.101, + 0.113, + 0.484, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.554, + 0.093, + 0.905, + 0.12 + ], + "angle": 0, + "content": "many regularizers of geometric complexity. NeurIPS, 35, 2022. 1, 4, 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.524, + 0.122, + 0.905, + 0.163 + ], + "angle": 0, + "content": "[18] Kamaludin Dingle, Chico Q Camargo, and Ard A Louis. Input-output maps are strongly biased towards simple outputs. Nature communications, 2018. 1, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.524, + 0.165, + 0.905, + 0.192 + ], + "angle": 0, + "content": "[19] Pedro Domingos. The role of occam's razor in knowledge discovery. Data mining and knowledge discovery, 1999. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.524, + 0.194, + 0.905, + 0.248 + ], + "angle": 0, + "content": "[20] Marius Dragoi, Florin Gogianu, and Elena Burceanu. Closing the gap on tabular data with fourier and implicit categorical features. Submission to ICLR (available on Open-Review), 2024. 5, 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.524, + 0.251, + 0.905, + 0.303 + ], + "angle": 0, + "content": "[21] Shiv Ram Dubey, Satish Kumar Singh, and Bidyut Baran Chaudhuri. Activation functions in deep learning: A comprehensive survey and benchmark. Neurocomputing, 2022. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.524, + 0.307, + 0.905, + 0.375 + ], + "angle": 0, + "content": "[22] Stanislas Ducotterd, Alexis Goujon, Pakshal Bohra, Dimitris Perdios, Sebastian Neumayer, and Michael Unser. Improving lipschitz-constrained neural networks by learning activation functions. Journal of Machine Learning Research, 2024. 3, 1, 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.524, + 0.378, + 0.905, + 0.418 + ], + "angle": 0, + "content": "[23] Daniel C Elton. Applying deutsch's concept of good explanations to artificial intelligence and neuroscience—an initial exploration. Cognitive Systems Research, 2021. 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.524, + 0.421, + 0.905, + 0.489 + ], + "angle": 0, + "content": "[24] Jesse Farebrother, Jordi Orbay, Quan Vuong, Adrien Ali Taiga, Yevgen Chebotar, Ted Xiao, Alex Irpan, Sergey Levine, Pablo Samuel Castro, Aleksandra Faust, et al. Stop regressing: Training value functions via classification for scalable deep rl. arXiv preprint arXiv:2403.03950, 2024. 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.524, + 0.491, + 0.905, + 0.517 + ], + "angle": 0, + "content": "[25] James Fiedler. Simple modifications to improve tabular neural networks. CoRR, abs/2108.03214, 2021. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.524, + 0.52, + 0.905, + 0.561 + ], + "angle": 0, + "content": "[26] Sara Fridovich-Keil, Raphael Gontijo Lopes, and Rebecca Roelofs. Spectral bias in practice: The role of function frequency in generalization. NeurIPS, 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.524, + 0.563, + 0.905, + 0.603 + ], + "angle": 0, + "content": "[27] Gallant. There exists a neural network that does not make avoidable mistakes. In IEEE International Conference on Neural Networks. IEEE, 1988. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.524, + 0.606, + 0.905, + 0.673 + ], + "angle": 0, + "content": "[28] Isabel O Gallegos, Ryan A Rossi, Joe Barrow, Md Mehrab Tanjim, Sungchul Kim, Franck Dernoncourt, Tong Yu, Ruiyi Zhang, and Nesreen K Ahmed. Bias and fairness in large language models: A survey. Computational Linguistics, 2024. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.524, + 0.676, + 0.905, + 0.731 + ], + "angle": 0, + "content": "[29] Robert Geirhos, Jorn-Henrik Jacobsen, Claudio Michaelis, Richard Zemel, Wieland Brendel, Matthias Bethge, and Felix A Wichmann. Shortcut learning in deep neural networks. Nature Machine Intelligence, 2020. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.524, + 0.733, + 0.905, + 0.788 + ], + "angle": 0, + "content": "[30] Florin Gogianu, Tudor Berariu, Mihaela C Rosca, Claudia Clopath, Lucian Busoniu, and Razvan Pascanu. Spectral normalisation for deep reinforcement learning: an optimisation perspective. In ICML, 2021. 11" + }, + { + "type": "ref_text", + "bbox": [ + 0.524, + 0.79, + 0.905, + 0.856 + ], + "angle": 0, + "content": "[31] Micah Goldblum, Marc Finzi, Keefer Rowan, and Andrew Gordon Wilson. The no free lunch theorem, kolmogorov complexity, and the role of inductive biases in machine learning. arXiv preprint arXiv:2304.05366, 2023. 1, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.524, + 0.859, + 0.905, + 0.901 + ], + "angle": 0, + "content": "[32] Yury Gorishniy, Ivan Rubachev, and Artem Babenko. On embeddings for numerical features in tabular deep learning. NeurIPS, 2022. 5, 1, 9" + }, + { + "type": "list", + "bbox": [ + 0.524, + 0.093, + 0.905, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.49, + 0.946, + 0.508, + 0.956 + ], + "angle": 0, + "content": "87" + } + ], + [ + { + "type": "ref_text", + "bbox": [ + 0.101, + 0.093, + 0.484, + 0.134 + ], + "angle": 0, + "content": "[33] Anirudh Goyal and Yoshua Bengio. Inductive biases for deep learning of higher-level cognition. Proceedings of the Royal Society A, 2022. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.101, + 0.136, + 0.484, + 0.176 + ], + "angle": 0, + "content": "[34] Léo Grinsztajn. Tabular data learning benchmark. https://github.com/LéoGrin/tabular-benchmark, 2022.5,6, 9,10,14" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.179, + 0.482, + 0.22 + ], + "angle": 0, + "content": "[35] Léo Grinsztajn, Edouard Oyallon, and Gael Varoquaux. Why do tree-based models still outperform deep learning on typical tabular data? NeurIPS, 2022. 4, 5, 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.222, + 0.483, + 0.249 + ], + "angle": 0, + "content": "[36] Andrey Gromov. Grokking modular arithmetic. arXiv preprint arXiv:2301.02679, 2023. 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.251, + 0.483, + 0.292 + ], + "angle": 0, + "content": "[37] Michael Hahn, Dan Jurafsky, and Richard Futrell. Sensitivity as a complexity measure for sequence classification tasks. Transactions of the ACL, 2021. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.294, + 0.483, + 0.347 + ], + "angle": 0, + "content": "[38] Hrayr Harutyunyan, Rafayel Darbinyan, Samvel Karapetyan, and Hrant Khachatrian. In-context learning in presence of spurious correlations. arXiv preprint arXiv:2410.03140, 2024. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.35, + 0.483, + 0.39 + ], + "angle": 0, + "content": "[39] Dan Hendrycks and Kevin Gimpel. Gaussian error linear units (GeLUs). arXiv preprint arXiv:1606.08415, 2016. 2, 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.393, + 0.483, + 0.446 + ], + "angle": 0, + "content": "[40] Katherine L Hermann and Andrew K Lampinen. What shapes feature representations? exploring datasets, architectures, and training. arXiv preprint arXiv:2006.12433, 2020. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.45, + 0.483, + 0.517 + ], + "angle": 0, + "content": "[41] Hossein Hosseini, Baicen Xiao, Mayoore Jaiswal, and Radha Poovendran. Assessing shape bias property of convolutional neural networks. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition Workshops, 2018. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.52, + 0.483, + 0.561 + ], + "angle": 0, + "content": "[42] Like Hui and Mikhail Belkin. Evaluation of neural architectures trained with square loss vs cross-entropy in classification tasks. arXiv preprint arXiv:2006.07322, 2020.1" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.564, + 0.483, + 0.616 + ], + "angle": 0, + "content": "[43] Ehsan Imani, Kai Luedemann, Sam Scholnick-Hughes, Esraa Elelimy, and Martha White. Investigating the histogram loss in regression. arXiv preprint arXiv:2402.13425, 2024. 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.619, + 0.483, + 0.687 + ], + "angle": 0, + "content": "[44] Ameya D Jagtap and George Em Karniadakis. How important are activation functions in regression and classification? a survey, performance comparison, and future directions. Journal of Machine Learning for Modeling and Computing, 4(1), 2023. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.69, + 0.483, + 0.744 + ], + "angle": 0, + "content": "[45] Ameya D Jagtap, Kenji Kawaguchi, and George Em Karniadakis. Adaptive activation functions accelerate convergence in deep and physics-informed neural networks. Journal of Computational Physics, 2020. 3, 5, 1, 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.747, + 0.483, + 0.8 + ], + "angle": 0, + "content": "[46] Anchit Jain, Rozhin Nobahari, Aristide Baratin, and Stefano Sarao Mannelli. Bias in motion: Theoretical insights into the dynamics of bias in sgd training. arXiv preprint arXiv:2405.18296, 2024.7" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.804, + 0.483, + 0.843 + ], + "angle": 0, + "content": "[47] Liangze Jiang and Damien Teney. OOD-chameleon: Is algorithm selection for ood generalization learnable? arXiv preprint arXiv:2410.02735, 2024. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.846, + 0.483, + 0.9 + ], + "angle": 0, + "content": "[48] Dimitris Kalimeris, Gal Kaplun, Preetum Nakkiran, Benjamin Edelman, Tristan Yang, Boaz Barak, and Haofeng Zhang. SGD on neural networks learns functions of increasing complexity. NeurIPS, 2019. 5, 1" + }, + { + "type": "list", + "bbox": [ + 0.101, + 0.093, + 0.484, + 0.9 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.525, + 0.093, + 0.905, + 0.134 + ], + "angle": 0, + "content": "[49] Alex Krizhevsky and Geoffrey Hinton. Learning multiple layers of features from tiny images. Technical report, University of Toronto, 2009. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.525, + 0.137, + 0.905, + 0.177 + ], + "angle": 0, + "content": "[50] Jan Kukačka, Vladimir Golkov, and Daniel Cremers. Regularization for deep learning: A taxonomy. arXiv preprint arXiv:1710.10686, 2017. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.525, + 0.181, + 0.905, + 0.234 + ], + "angle": 0, + "content": "[51] Tanishq Kumar, Blake Bordelon, Samuel J Gershman, and Cengiz Pehlevan. Grokking as the transition from lazy to rich training dynamics. arXiv preprint arXiv:2310.06110, 2023. 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.525, + 0.238, + 0.905, + 0.291 + ], + "angle": 0, + "content": "[52] Yann LeCun, Léon Bottou, Yoshua Bengio, and Patrick Haffner. Gradient-based learning applied to document recognition. Proceedings of the IEEE, 86(11):2278-2324, 1998. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.525, + 0.296, + 0.905, + 0.336 + ], + "angle": 0, + "content": "[53] Alexander Li and Deepak Pathak. Functional regularization for reinforcement learning via learned fourier features. NeurIPS, 2021. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.525, + 0.34, + 0.905, + 0.38 + ], + "angle": 0, + "content": "[54] Hao Li, Zheng Xu, Gavin Taylor, Christoph Studer, and Tom Goldstein. Visualizing the loss landscape of neural nets. NeurIPS, 2018. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.525, + 0.384, + 0.905, + 0.41 + ], + "angle": 0, + "content": "[55] Xuan Li, Yun Wang, and Bo Li. Tree-regularized tabular embeddings. arXiv preprint arXiv:2403.00963, 2024. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.525, + 0.413, + 0.907, + 0.453 + ], + "angle": 0, + "content": "[56] Henry W Lin, Max Tegmark, and David Rolnick. Why does deep and cheap learning work so well? Journal of Statistical Physics, 2017. 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.525, + 0.457, + 0.905, + 0.511 + ], + "angle": 0, + "content": "[57] Ziqi Liu, Wei Cai, and Zhi-Qin John Xu. Multi-scale deep neural network (mscalednn) for solving poisson-boltzmann equation in complex domains. arXiv preprint arXiv:2007.11207, 2020. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.525, + 0.514, + 0.905, + 0.568 + ], + "angle": 0, + "content": "[58] Ziming Liu, Ouail Kitouni, Niklas S Nolte, Eric Michaud, Max Tegmark, and Mike Williams. Towards understanding grokking: An effective theory of representation learning. NeurIPS, 2022. 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.525, + 0.572, + 0.905, + 0.626 + ], + "angle": 0, + "content": "[59] Ziming Liu, Yixuan Wang, Sachin Vaidya, Fabian Ruehle, James Halverson, Marin Soljačić, Thomas Y Hou, and Max Tegmark. Kan: Kolmogorov-arnold networks. arXiv preprint arXiv:2404.19756, 2024. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.525, + 0.63, + 0.905, + 0.67 + ], + "angle": 0, + "content": "[60] Kaifeng Lyu, Zhiyuan Li, Runzhe Wang, and Sanjeev Arora. Gradient descent on two-layer nets: Margin maximization and simplicity bias. NeurIPS, 2021. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.525, + 0.674, + 0.905, + 0.713 + ], + "angle": 0, + "content": "[61] Andrew L Maas, Awni Y Hannun, Andrew Y Ng, et al. Rectifier nonlinearities improve neural network acoustic models. In ICML, 2013. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.525, + 0.717, + 0.905, + 0.771 + ], + "angle": 0, + "content": "[62] Augustine N Mavor-Parker, Matthew J Sargent, Caswell Barry, Lewis Griffin, and Clare Lyle. Frequency and generalisation of periodic activation functions in reinforcement learning. arXiv preprint arXiv:2407.06756, 2024. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.525, + 0.775, + 0.905, + 0.828 + ], + "angle": 0, + "content": "[63] Duncan McElfresh, Sujay Khandagale, Jonathan Valverde, Vishak Prasad C, Ganesh Ramakrishnan, Micah Goldblum, and Colin White. When do neural nets outperform boosted trees on tabular data? NeurIPS, 2024. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.525, + 0.832, + 0.905, + 0.899 + ], + "angle": 0, + "content": "[64] Chris Mingard, Joar Skalse, Guillermo Valle-Pérez, David Martínez-Rubio, Vladimir Mikulik, and Ard A Louis. Neural networks are a priori biased towards boolean functions with low entropy. arXiv preprint arXiv:1909.11522, 2019. 1, 5" + }, + { + "type": "list", + "bbox": [ + 0.525, + 0.093, + 0.907, + 0.899 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.491, + 0.947, + 0.508, + 0.956 + ], + "angle": 0, + "content": "88" + } + ], + [ + { + "type": "ref_text", + "bbox": [ + 0.101, + 0.092, + 0.482, + 0.134 + ], + "angle": 0, + "content": "[65] Chris Mingard, Guillermo Valle-Pérez, Joar Skalse, and Ard A Louis. Is SGD a bayesian sampler? well, almost. Journal of Machine Learning Research, 2021. 1, 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.101, + 0.136, + 0.483, + 0.178 + ], + "angle": 0, + "content": "[66] Chris Mingard, Henry Rees, Guillermo Valle-Pérez, and Ard A Louis. Do deep neural networks have an inbuilt ocean's razor? arXiv preprint arXiv:2304.06670, 2023. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.181, + 0.482, + 0.22 + ], + "angle": 0, + "content": "[67] Tom M Mitchell. The need for biases in learning generalizations. Rutgers University CS tech report CBM-TR-117, 1980. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.224, + 0.483, + 0.293 + ], + "angle": 0, + "content": "[68] Yuval Netzer, Tao Wang, Adam Coates, Alessandro Bissacco, Bo Wu, and Andrew Y Ng. Reading digits in natural images with unsupervised feature learning. NIPS Workshop on Deep Learning and Unsupervised Feature Learning, 2011. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.296, + 0.483, + 0.35 + ], + "angle": 0, + "content": "[69] Behnam Neyshabur, Ryota Tomioka, and Nathan Srebro. In search of the real inductive bias: On the role of implicit regularization in deep learning. arXiv preprint arXiv:1412.6614, 2014. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.353, + 0.482, + 0.407 + ], + "angle": 0, + "content": "[70] Mohammad Pezeshki, Oumar Kaba, Yoshua Bengio, Aaron C Courville, Doina Precup, and Guillaume Lajoie. Gradient starvation: A learning proclivity in neural networks. NeurIPS, 2021. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.411, + 0.482, + 0.466 + ], + "angle": 0, + "content": "[71] Tomaso Poggio, Kenji Kawaguchi, Qianli Liao, Brando Miranda, Lorenzo Rosasco, Xavier Boix, Jack Hiday, and Hrushikesh Mhaskar. Theory of deep learning III: the nonoverfitting puzzle. CBMM Memo, 2018. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.468, + 0.482, + 0.496 + ], + "angle": 0, + "content": "[72] Karl Popper. \"7. Simplicity\". The logic of scientific discovery. Routledge, 1959. 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.499, + 0.482, + 0.553 + ], + "angle": 0, + "content": "[73] Alethea Power, Yuri Burda, Harri Edwards, Igor Babuschkin, and Vedant Misra. Grokking: Generalization beyond overfitting on small algorithmic datasets. arXiv preprint arXiv:2201.02177, 2022. 7, 8, 12" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.556, + 0.482, + 0.598 + ], + "angle": 0, + "content": "[74] Aahlad Manas Puli, Lily Zhang, Yoav Wald, and Rajesh Ranganath. Don't blame dataset shift! shortcut learning due to gradients and cross entropy. NeurIPS, 2023. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.6, + 0.482, + 0.654 + ], + "angle": 0, + "content": "[75] Nasim Rahaman, Aristide Baratin, Devansh Arpit, Felix Draxler, Min Lin, Fred Hamprecht, Yoshua Bengio, and Aaron Courville. On the spectral bias of neural networks. In ICML. PMLR, 2019. 5, 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.657, + 0.482, + 0.697 + ], + "angle": 0, + "content": "[76] Prajit Ramachandran, Barret Zoph, and Quoc V Le. Swish: a self-gated activation function. arXiv preprint arXiv:1710.05941, 2017. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.701, + 0.482, + 0.742 + ], + "angle": 0, + "content": "[77] Sameera Ramasinghe and Simon Lucey. Beyond periodicity: Towards a unifying framework for activations in coordinate-MLPs. In ECCV. Springer, 2022. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.745, + 0.482, + 0.785 + ], + "angle": 0, + "content": "[78] Sameera Ramasinghe, Lachlan E MacDonald, and Simon Lucey. On the frequency-bias of coordinate-MLPs. NeurIPS, 2022. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.788, + 0.482, + 0.843 + ], + "angle": 0, + "content": "[79] Mihaela Rosca, Theophane Weber, Arthur Gretton, and Shakir Mohamed. A case for new neural network smoothness constraints. I Can't Believe It's Not Better (ICBINB) Workshop at NeurIPS, 2020. 11" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.846, + 0.482, + 0.9 + ], + "angle": 0, + "content": "[80] Vishwanath Saragadam, Daniel LeJeune, Jasper Tan, Guha Balakrishnan, Ashok Veeraraghavan, and Richard G Baraniuk. Wire: Wavelet implicit neural representations. In CVPR, 2023. 4, 1" + }, + { + "type": "list", + "bbox": [ + 0.101, + 0.092, + 0.483, + 0.9 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.523, + 0.093, + 0.905, + 0.16 + ], + "angle": 0, + "content": "[81] Simone Scardapane, Michele Scarpiniti, Danilo Comminiello, and Aurelio Uncini. Learning activation functions from data using cubic spline interpolation. Neural Advances in Processing Nonlinear Dynamic Signals, 2019. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.523, + 0.164, + 0.905, + 0.218 + ], + "angle": 0, + "content": "[82] Simone Scardapane, Steven Van Vaerenbergh, Simone Toto, and Aurelio Uncini. Kafnets: Kernel-based nonparametric activation functions for neural networks. Neural Networks, 2019. 2, 3, 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.523, + 0.221, + 0.905, + 0.261 + ], + "angle": 0, + "content": "[83] Jürgen Schmidhuber. Discovering neural nets with low kolmogorov complexity and high generalization capability. Neural Networks, 1997. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.523, + 0.264, + 0.905, + 0.318 + ], + "angle": 0, + "content": "[84] Luca Scimeca, Seong Joon Oh, Sanghyuk Chun, Michael Poli, and Sangdoo Yun. Which shortcut cues will DNNs choose? a study from the parameter-space perspective. arXiv preprint arXiv:2110.03095, 2021. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.523, + 0.321, + 0.905, + 0.36 + ], + "angle": 0, + "content": "[85] Harshay Shah, Kaustav Tamuly, Aditi Raghunathan, Prateek Jain, and Praneeth Netrapalli. The pitfalls of simplicity bias in neural networks. NeurIPS, 2020. 2, 5, 6, 11" + }, + { + "type": "ref_text", + "bbox": [ + 0.523, + 0.363, + 0.905, + 0.403 + ], + "angle": 0, + "content": "[86] Kexuan Shi, Xingyu Zhou, and Shuhang Gu. Improved implicit neural representation with fourier reparameterized training. In CVPR, 2024. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.523, + 0.406, + 0.905, + 0.447 + ], + "angle": 0, + "content": "[87] Prasann Singhal, Tanya Goyal, Jiacheng Xu, and Greg Durrett. A long way to go: Investigating length correlations in rlhf. arXiv preprint arXiv:2310.03716, 2023. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.523, + 0.449, + 0.905, + 0.502 + ], + "angle": 0, + "content": "[88] Vincent Sitzmann, Julien Martel, Alexander Bergman, David Lindell, and Gordon Wetzstein. Implicit neural representations with periodic activation functions. NeurIPS, 2020. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.523, + 0.505, + 0.905, + 0.56 + ], + "angle": 0, + "content": "[89] Daniel Soudry, Elad Hoffer, Mor Shpigel Nacson, Suriya Gunasekar, and Nathan Srebro. The implicit bias of gradient descent on separable data. The Journal of Machine Learning Research, 2018. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.523, + 0.563, + 0.905, + 0.616 + ], + "angle": 0, + "content": "[90] Lawrence Stewart, Francis Bach, Quentin Berthet, and Jean-Philippe Vert. Regression as classification: Influence of task formulation on neural network features. In ICML. PMLR, 2023. 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.523, + 0.619, + 0.905, + 0.687 + ], + "angle": 0, + "content": "[91] Leon René Sutfeld, Flemming Brieger, Holger Finger, Sonja Fülhase, and Gordon Pipa. Adaptive blending units: Trainable activation functions for deep neural networks. In Intelligent Computing: Proceedings of the Computing Conference. Springer, 2020. 3, 1, 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.523, + 0.689, + 0.905, + 0.743 + ], + "angle": 0, + "content": "[92] Remi Tachet, Mohammad Pezeshki, Samira Shabanian, Aaron Courville, and Yoshua Bengio. On the learning dynamics of deep neural networks. arXiv preprint arXiv:1809.06848, 2018. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.523, + 0.746, + 0.905, + 0.814 + ], + "angle": 0, + "content": "[93] Damien Teney, Ehsan Abbasnejad, Simon Lucey, and Anton van den Hengel. Evading the simplicity bias: Training a diverse set of models discovers solutions with superior ood generalization. arXiv preprint arXiv:2105.05612, 2021. 1, 2, 6, 11" + }, + { + "type": "ref_text", + "bbox": [ + 0.523, + 0.817, + 0.905, + 0.871 + ], + "angle": 0, + "content": "[94] Damien Teney, Maxime Peyrard, and Ehsan Abbasnejad. Predicting is not understanding: Recognizing and addressing underspecification in machine learning. In ECCV. Springer, 2022. 6, 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.523, + 0.874, + 0.905, + 0.901 + ], + "angle": 0, + "content": "[95] Damien Teney, Yong Lin, Seong Joon Oh, and Ehsan Abbasnejad. ID and OOD performance are sometimes in" + }, + { + "type": "list", + "bbox": [ + 0.523, + 0.093, + 0.905, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.49, + 0.946, + 0.508, + 0.956 + ], + "angle": 0, + "content": "89" + } + ], + [ + { + "type": "ref_text", + "bbox": [ + 0.133, + 0.093, + 0.482, + 0.119 + ], + "angle": 0, + "content": "versely correlated on real-world datasets. NeurIPS, 2024. 1, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.101, + 0.121, + 0.483, + 0.176 + ], + "angle": 0, + "content": "[96] Damien Teney, Armand Mihai Nicoliciou, Valentin Hartmann, and Ehsan Abbasnejad. Neural redshift: Random networks are not random functions. In CVPR, 2024. 1, 2, 3, 4, 5, 6, 8, 11, 14" + }, + { + "type": "ref_text", + "bbox": [ + 0.101, + 0.177, + 0.482, + 0.218 + ], + "angle": 0, + "content": "[97] Tijmen Tieleman. Lecture 6.5-rmsprop: Divide the gradient by a running average of its recent magnitude. COURSERA: Neural networks for machine learning, 4(2):26, 2012. 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.101, + 0.219, + 0.483, + 0.26 + ], + "angle": 0, + "content": "[98] Nikita Tsoy and Nikola Konstantinov. Simplicity bias of two-layer networks beyond linearly separable data. arXiv preprint arXiv:2405.17299, 2024. 5, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.261, + 0.483, + 0.316 + ], + "angle": 0, + "content": "[99] Guillermo Valle-Perez, Chico Q Camargo, and Ard A Louis. Deep learning generalizes because the parameter-function map is biased towards simple functions. arXiv preprint arXiv:1805.08522, 2018. 1, 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.317, + 0.483, + 0.372 + ], + "angle": 0, + "content": "[100] Colin White, Mahmoud Safari, Rhea Sukthanker, Binxin Ru, Thomas Elsken, Arber Zela, Debadeepta Dey, and Frank Hutter. Neural architecture search: Insights from 1000 papers. arXiv preprint arXiv:2301.08727, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.372, + 0.482, + 0.412 + ], + "angle": 0, + "content": "[101] David H Wolpert. The supervised learning no-free-lunch theorems. Soft computing and industry: Recent applications, 2002. 1, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.413, + 0.483, + 0.468 + ], + "angle": 0, + "content": "[102] Han Xiao, Kashif Rasul, and Roland Vollgraf. Fashion-mnist: A novel image dataset for benchmarking machine learning algorithms. arXiv preprint arXiv:1708.07747, 2017.3" + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.469, + 0.482, + 0.538 + ], + "angle": 0, + "content": "[103] Yiheng Xie, Towaki Takikawa, Shunsuke Saito, Or Litany, Shiqin Yan, Numair Khan, Federico Tombari, James Tompkin, Vincent Sitzmann, and Srinath Sridhar. Neural fields in visual computing and beyond. In Computer Graphics Forum. Wiley Online Library, 2022. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.539, + 0.483, + 0.594 + ], + "angle": 0, + "content": "[104] Zhi-Qin John Xu, Yaoyu Zhang, Tao Luo, Yanyang Xiao, and Zheng Ma. Frequency principle: Fourier analysis sheds light on deep neural networks. arXiv preprint arXiv:1901.06523, 2019. 5, 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.595, + 0.482, + 0.636 + ], + "angle": 0, + "content": "[105] Zhi-Qin John Xu, Yaoyu Zhang, and Yanyang Xiao. Training behavior of deep neural network in frequency domain. In ICONIP. Springer, 2019. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.637, + 0.482, + 0.69 + ], + "angle": 0, + "content": "[106] Zhi-Qin John Xu, Yaoyu Zhang, and Tao Luo. Overview frequency principle/spectral bias in deep learning. Communications on Applied Mathematics and Computation, 2024. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.692, + 0.482, + 0.734 + ], + "angle": 0, + "content": "[107] Ge Yang, Anurag Ajay, and Pulkit Agrawal. Overcoming the spectral bias of neural value approximation. arXiv preprint arXiv:2206.04672, 2022. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.735, + 0.482, + 0.774 + ], + "angle": 0, + "content": "[108] Ziqian Zhong and Jacob Andreas. Algorithmic capabilities of random transformers. arXiv preprint arXiv:2410.04368, 2024. 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.776, + 0.482, + 0.843 + ], + "angle": 0, + "content": "[109] Hattie Zhou, Arwen Bradley, Etai Littwin, Noam Razin, Omid Saremi, Josh Susskind, Samy Bengio, and Preetum Nakkiran. What algorithms can transformers learn? a study in length generalization. arXiv preprint arXiv:2310.16028, 2023. 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.845, + 0.482, + 0.899 + ], + "angle": 0, + "content": "[110] Zhangchen Zhou, Yaoyu Zhang, and Zhi-Qin John Xu. A rationale from frequency perspective for grokking in training neural network. arXiv preprint arXiv:2405.17479, 2024. 7" + }, + { + "type": "list", + "bbox": [ + 0.093, + 0.093, + 0.483, + 0.899 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.491, + 0.946, + 0.508, + 0.956 + ], + "angle": 0, + "content": "90" + } + ] +] \ No newline at end of file diff --git a/2025/Do We Always Need the Simplicity Bias_ Looking for Optimal Inductive Biases in the Wild/f8b7dd25-2878-4e2f-ac24-ae160cb54c72_origin.pdf b/2025/Do We Always Need the Simplicity Bias_ Looking for Optimal Inductive Biases in the Wild/f8b7dd25-2878-4e2f-ac24-ae160cb54c72_origin.pdf new file mode 100644 index 0000000000000000000000000000000000000000..ca54c4e9961b329709f812f8478c75143bda0bc5 --- /dev/null +++ b/2025/Do We Always Need the Simplicity Bias_ Looking for Optimal Inductive Biases in the Wild/f8b7dd25-2878-4e2f-ac24-ae160cb54c72_origin.pdf @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:6254c5ab39e7e065ce7cc8e0270741bfeea82018f91cfade7a89cab697b92d99 +size 1516502 diff --git a/2025/Do We Always Need the Simplicity Bias_ Looking for Optimal Inductive Biases in the Wild/full.md b/2025/Do We Always Need the Simplicity Bias_ Looking for Optimal Inductive Biases in the Wild/full.md new file mode 100644 index 0000000000000000000000000000000000000000..4a8bac535372d68e03670a470d7910ef400c4c94 --- /dev/null +++ b/2025/Do We Always Need the Simplicity Bias_ Looking for Optimal Inductive Biases in the Wild/full.md @@ -0,0 +1,443 @@ +# Do We Always Need the Simplicity Bias? Looking for Optimal Inductive Biases in the Wild + +Damien Teney +Idiap Research Institute + +Liangze Jiang EPFL + +# Abstract + +Neural architectures tend to fit their data with relatively simple functions. This "simplicity bias" is widely regarded as key to their success. This paper explores the limits of this principle. Building on recent findings that the simplicity bias stems from ReLU activations [96], we introduce a method to meta-learn new activation functions and inductive biases better suited to specific tasks. + +Findings. We identify multiple tasks where the simplicity bias is inadequate and ReLUs suboptimal. In these cases, we learn new activation functions that perform better by inducing a prior of higher complexity. Interestingly, these cases correspond to domains where neural networks have historically struggled: tabular data, regression tasks, cases of shortcut learning, and algorithmic grokking tasks. In comparison, the simplicity bias induced by ReLUs proves adequate on image tasks where the best learned activations are nearly identical to ReLUs and GeLUs. + +Implications. Contrary to popular belief, the simplicity bias of ReLU networks is not universally useful. It is near-optimal for image classification, but other inductive biases are sometimes preferable. We showed that activation functions can control these inductive biases, but future tailored architectures might provide further benefits. Advances are still needed to characterize a model's inductive biases beyond "complexity", and their adequacy with the data. + +# 1. Introduction + +When and why NNs generalize is yet to be understood. Neural networks (NNs) have proven more effective than other machine learning models. However, we still miss a complete explanation of their generalization abilities. A better understanding could help address failures from shortcut learning [29, 93] to distribution shifts [47, 95], biases, and spurious correlations in language models for example [28, 38, 87]. Understanding conditions for generalization would also enable the design of architectures and data preparation from first principles, rather than trial and error. + +Florin Gogianu +Bitdefender + +Ehsan Abbasnejad +University of Adelaide + +![](images/f6c3c7fa1ae2b01883bb609bc39722c6311e201422ffb6aaffc6076662c9f41c.jpg) +Figure 1. (1) We modulate the inductive bias of neural architectures by learning novel activation functions that improve generalization on specific datasets. (2) With this tool, we study the relation between model accuracy and complexity. We identify tasks where the simplicity bias of ReLU architectures is suboptimal. + +This paper studies inductive biases i.e. the assumptions made by learning algorithms to generalize beyond training data [67].1 A vast literature examines the inductive biases of architectures [14], optimizers [69], losses [42], regularizers [50], etc. The simplicity bias is one aspect of the inductive biases of NNs that makes them fit their training data with simple2 functions [8, 71]. Despite wide belief that the simplicity bias could be due to SGD [7, 40, 60, 92], work on untrained networks showed that it can be explained with architectures alone [15, 31, 64, 99]. ReLUs also seem critical to induce the simplicity bias in typical architectures [96]. + +Limits of the simplicity bias. The simplicity bias is an intuitive explanation for the ability to generalize on real-world data. It embodies Occam's razor [66] and assumes that data-generating processes in the real world are simple. Additionally, a prior for simplicity is supported by results in algorithmic information theory [18] stating essentially that "a bias in the distribution of target functions must be towards low complexity". However, this only means that simplicity is a good prior on average, but not necessarily the best choice on any task or dataset. This matches the no-free lunch theorem [101] according to which no inductive bias is universally useful. Therefore, this paper asks the following. + +# Are there practical applications of machine learning where the simplicity bias is detrimental? In these cases, what do the optimal inductive biases look like? + +For example, shortcut learning is one situation where the simplicity bias is already known to be detrimental [85, 93]. + +Searching for optimal inductive biases by learning activation functions. Prior work [96] showed that ReLU activations are critical to obtain the simplicity bias in typical architectures. Hence we build a new tool to modulate the simplicity bias by learning dataset-specific activation functions. It uses bi-level optimization and a spline parametrization to learn activations free of any prior, such as constraints of smoothness or monotonicity (unlike prior work [2, 6, 82]). This (1) enables the discovery of entirely new activation functions and inductive biases that improve generalization (Figure 1) and (2) highlights the suboptimality of the simplicity bias by comparing the accuracy and complexity of models with ReLUs vs. learned activations. + +Findings. We examine four domains that we hypothesized to be impaired by the simplicity bias: tabular data, regression tasks, cases of shortcut learning, and algorithmic tasks. Our intuition is that they require learning functions with high sensitivity or sharp transitions. For each domain, we collect existing datasets then train and analyze models without and with learned activation functions. In all cases, we obtain better generalization with dataset-specific activations, and the improvements are attributable to learning higher-complexity solutions. In comparison, this analysis also shows that classical image datasets (MNIST, CIFAR, FASHION-MNIST, SVHN) are extremely well suited to the inductive biases of ReLUs. The best learned activations are then strikingly similar to variants like GeLUs [39]. + +# Summary of contributions. + +- A new method to discover dataset-specific activation functions optimized for generalization. +- An examination of $>20$ datasets showing that the simplicity bias of ReLU architectures can be suboptimal. (1) For regression tasks and tabular data, new learned activations greatly improve accuracy by helping learn complex functions. (2) For image classification, the process rediscovers smooth variants of ReLUs, suggesting a near-optimal choice for these popular tasks. (3) In cases of shortcut learning, we show that different learned activations can steer the learning towards different image features. (4) For grokking tasks, new learned activations can eliminate the phenomenon, supporting the explanation as a mismatch between data and architectures. We also measure a positive transferability of learned activations across related tasks. +- An analysis showing that improvements with learned activations correlate with the learning of complex functions. + +Implications. All cases where the simplicity bias proved suboptimal are in domains where NNs have historically struggled. We now connect them to a common explanation. This implies that architectures tailored to some specific domains may still have a place besides scaling up models and data. Conversely, the suitability of ReLUs to image classification suggests that researchers successfully converged by trial and error to designs well tuned to popular tasks. + +# 2. Methods + +This section introduces tools to analyze trained models and to learn new dataset-specific activation functions. + +# 2.1. Visualizing a Model's Function + +A neural network implements a function $f_{\theta} : \mathbb{R}^{d_{\mathrm{in}}} \rightarrow \mathbb{R}^{d_{\mathrm{out}}}$ of parameters $\theta$ (weights and biases) that maps an input $x \in \mathbb{R}^{d_{\mathrm{in}}}$ to an output $y \in \mathbb{R}^{d_{\mathrm{out}}}$ . For a regression task, $y \in \mathbb{R}$ is the predicted value. For a classification task, $y$ is a vector of logits passed through a sigmoid or softmax to obtain class probabilities. Because $d_{\mathrm{in}}$ can be large, $f$ can be difficult to visualize and analyze. A workaround is to examine $f$ over 1D or 2D slices of the input space [26, 96]. To obtain a slice in a region of plausible data, we use the training data $\mathcal{T}$ . For a 1D slice (linear path), we sample $x_1, x_2 \sim \mathcal{T}$ then define the path $\mathbf{X}_{x_1, x_2} = [(1 - \lambda) x_1 + \lambda x_2, \lambda \in [0,1]]$ . We proceed analogously with three points for a 2D slice. We sample $\lambda$ regularly in [0,1] such that $\mathbf{X}$ is a finite sequence of points. Then $f$ is evaluated on these points to give a 1D sequence or 2D grid of values that are convenient to display and analyze (Figure 7c). When $d_{\mathrm{out}} > 1$ (multi-class task), we examine one random dimension of $f$ 's output at a time. + +# 2.2. Measuring a Model's Complexity + +We wish to quantify the complexity of the function $f$ implemented by a model trained on data $\mathcal{T}$ . Prior work used Fourier decompositions [26, 96] but this requires a delicate implementation. We found a reliable alternative with the total variation (TV) of $f$ averaged over many 1D paths: + +$$ +\operatorname {T V} (f, \mathcal {T}) = \mathbb {E} _ {\boldsymbol {x} _ {1}, \boldsymbol {x} _ {2}} \sim \mathcal {T} \int_ {\boldsymbol {x} _ {1}} ^ {\boldsymbol {x} _ {2}} \left| f ^ {\prime} (\boldsymbol {x}) \right| d \boldsymbol {x}. \tag {1} +$$ + +with $f'$ the first derivative. We estimate (1) using a path as defined in Section 2.1. We name the points in such a path $\mathbf{X}_{\boldsymbol{x}_a, \boldsymbol{x}_z} := [\boldsymbol{x}_a, \boldsymbol{x}_b, \boldsymbol{x}_c, \dots, \boldsymbol{x}_y, \boldsymbol{x}_z]$ . We then have: + +$$ +\begin{array}{l} \operatorname {T V} (f, \mathcal {T}) \approx \mathbb {E} _ {\boldsymbol {x} _ {a}, \boldsymbol {x} _ {z}} \sim_ {\mathcal {T}} | f (\boldsymbol {x} _ {b}) - f (\boldsymbol {x} _ {a}) | \\ + \left| f \left(\boldsymbol {x} _ {c}\right) - f (\boldsymbol {x} _ {b}) \right| + \dots \tag {2} \\ + \left| f \left(\boldsymbol {x} _ {z}\right) - f \left(\boldsymbol {x} _ {y}\right) \right|. \\ \end{array} +$$ + +Appendix F shows that (2) correlates closely with a Fourier-based measure of complexity: the higher the TV, the higher the complexity. Yet, it is straightforward to implement and discriminative across small and large values. + +# 2.3. Meta-Learning Activation Functions + +Our goal is to optimize the inductive biases of a neural network and recent work [96] showed that the activation functions are the most important component. The typical approach to learn activations [2, 5, 6, 11, 13, 22, 45, 82, 91] (see Related Work) replaces them with a small shared ReLU MLP that implements an $\mathbb{R}\rightarrow \mathbb{R}$ function. Its parameters are optimized along the network's. However this cannot discover truly novel activations because the embedded ReLU MLP has itself a simplicity bias and activations are optimized together with the model. We propose instead: + +- an unbiased parametrization of the activations as splines, +- a bi-level optimization to learn reusable activations, +- an episodic training to optimize for generalization rather than simply to fit the training data. + +Parametrization as splines. We want a space of activation functions free of priors such as the smoothness and monotonicity enforced in prior work [5, 13]. We implement an activation $g_{\psi} : \mathbb{R} \to \mathbb{R}$ as a linear spline with control points defined by $\psi$ . We define $n_c$ points spread regularly in an interval $[a, b]$ , typically $\sim 50$ points in $[-5, +5]$ . Then $g$ represents piecewise linear segments interpolating values specified in the learned parameters $\psi := [g_{\psi}(a), \dots, g_{\psi}(b)] \in \mathbb{R}^{n_c}$ . $g$ can represent simple and complex functions, including smooth curves, periodic functions, sharp transitions, etc. + +Bi-level optimization & episodic training. Our goal is to get an activation function that can be reused like any other in subsequent training runs. This differs from prior work (e.g. [2]) that continuously updates the activation during training: the final one may not be suitable to start training with. Our solution is a bi-level meta-learning loop. An inner loop trains the model with a fixed activation function. An outer loop trains the activation function to maximize generalization. Each outer step simulates a new learning task or episode. This means (1) initializing the model with different weights and (2) using different subsets of data for training and validation. With suitable choices, this can simulate in- or out-of-distribution conditions (see Section 3.4). Without episodes, the learned activation could overfit to a particular model initialization for example, and would not generalize in subsequent training runs. The method is outlined as Algorithm 1. Its implementation is discussed in Appendix C. + +![](images/ed53be628f428e1cb7466d73447f60e716bb2b6787d390e027740dddbbeb9e65.jpg) + +Inductive bias and simplicity bias are not interchangeable. Our method optimizes toward bethepneralization. Simplicity is only one aspect of the 1d models that we analyze post-hoc (e.g. Figure 4). + +# 3. Tasks and Results + +We now examine tasks that we hypothesized to be ill-suited to the simplicity bias of ReLU architectures. The intuition is that the target function to learn (e.g. optimal classifier) con + +Algorithm 1 Meta-learning an activation function (AF). +Input: training data $\mathcal{T}$ ; untrained neural model $f_{\theta ,\psi}$ Initialize $\psi$ with zeros Parametrization of AF $n_{\mathrm{tr}}\gets 0$ Number of inner-loop iterations while $n_{\mathrm{tr}} < n_{\mathrm{tr}}^{\mathrm{max}}$ Outer loop: train AF Increment $n_{\mathrm{tr}}$ Sample the episode's tr. $(T^{\prime})$ and val. $(\mathcal{V})$ sets from T Initialize $\theta$ randomly Model weights and biases for $n_{\mathrm{tr}}$ steps Inner loop: train model with fixed AF Eval. loss on $\mathcal{T}':L\gets \Sigma_{(\pmb {x},\pmb {y})\in \mathcal{T}'}\mathcal{L}\big(f_{\pmb {\theta},\pmb {\psi}}(\pmb {x},\pmb {y})\big)$ Gradient step on weights/biases: $\pmb {\theta}\gets \mathrm{GD}(\pmb {\theta},\nabla_{\pmb{\theta}}L)$ Eval. loss on V: $L\gets \Sigma_{(\pmb {x},\pmb {y})\in \mathcal{V}}\mathcal{L}\big(f_{\pmb {\theta},\pmb {\psi}}(\pmb {x},\pmb {y})\big)$ Gradient step on AF: $\psi \gets \mathrm{GD}(\psi ,\nabla_{\psi}L)$ if performance on $\nu$ worsens then break Early stopping + +Output: optimized AF $\psi$ + +tains sharp transitions (regression tasks, tabular datasets), or repeating patterns (algorithmic tasks) that contradict the ReLUs' simplicity bias. For each task, we examine existing datasets with the tools from Section 2. In all cases, we find benefits from architectures whose inductive biases favor more complex functions. Additional details and results are provided in Appendix E. + +# 3.1. Image Classification Tasks + +Background. We start with classical datasets to validate our methodology: MNIST, FASHION-MNIST, SVHN, CIFAR-10 [49, 52, 68, 102]. They are representative of the vision tasks that guided the development of deep learning. Our hypothesis is therefore that the inductive biases of modern architectures and ReLUs are well suited to these datasets. + +Setup. For each dataset, we learn activation functions with Algorithm 1. We experiment with two initializations of the spline parameters: as zeros and so as to mimic a ReLU. The goal of the latter is to explore the space of functions similar to ReLUs. Because of the difficulty of the optimization, the algorithm is likely to converge to a local optimum similar to ReLUs if there is one. We also experiment with the sharing of the activation function. By default, a single function is shared across the network. Alternatively, we learn a different activation function per layer. This provides more ways to affect the model's inductive biases. Our base architecture is a 3-layer MLP (details in Appendix E). + +Results. We compare in Figure 2a the accuracy of models with ReLUs vs. learned activation functions. Differences are small. The learned activations only improve slightly on SVHN and CIFAR. This suggests that the inductive biases of ReLUs are generally well suited to these datasets. + +We examine the learned activations in Figure 3a. With + +![](images/211b35dc0b91d3f90a9b9e2fd0c6c2905f10ef07f027f107e2be95efab6b54df.jpg) +Figure 2. Test accuracy on image datasets. (a) For classification tasks, all models perform similarly, suggesting that the inductive biases of ReLUs are well suited to these datasets. (b) For regression tasks, models with learned activations perform better, especially from an initialization as zeros, which enables the discovery of completely novel activation functions. + +![](images/a76d47c4c1ce9be58cd6a561fbc4f9a6eb005f7a13727ef07809908712d3a81f.jpg) +- ReLU activations +- Learned act., ReLU init. +- Learned act., zero init. + +![](images/6471036dbc2489f41e7f191e6fa9b8105b97989fcf3ba5b13b887c3690c24440.jpg) +Figure 3. Activation functions learned for MNIST. For a classification task, the activation learned from a ReLU resembles the popular GeLUs. For a regression task, the learned activations contain irregularities that help a network represent complex functions. See Figure 16 for similar results on other datasets. + +an initialization as ReLUs, the optimization converges to a smooth variant remarkably similar to GeLUs [11] which are widely used. This suggests that the research community has empirically converged on a local optimum in the space of activation functions. With an initialization as zeros, we discover wavelets [80] that are unlike common activations but perform as well as ReLUs, i.e. another local optimum. + +Take-away: for image classification, learned activations provide very little benefit over ReLUs. Smooth variants of ReLUs are a local optimum in the space of activations. ReLUs' popularity for such tasks could thus be explained with their proximity to this optimum. + +# 3.2. Regression Tasks + +Background. Regression tasks are known to be difficult for NNs [90]. They are often turned into a classification through discretization [24, 43]. Existing explanations that invoke implicit biases of gradient descent are clearly incomplete [90]. Our hypothesis is that regression is difficult because it often involves irregular decision boundaries [35] in opposition to the typical solutions of ReLU networks [17]. + +Setup. We use the same setup and image datasets as Sec- + +![](images/2413db2370a191b22bbcfc9433a64c78f101c9d5ec538770208e7115a8f0bbbc.jpg) +Figure 4. Accuracy vs. complexity on image datasets. Each marker is a model with different hyperparameters and ReLUs $(\bullet)$ or learned activations initialized as ReLUs $(\square)$ or as zeros $(\text{圆})$ . For classification (a), ReLUs are close to best. Activations optimized from ReLUs only improve the accuracy slightly, corresponding to the GeLU-like function in Figure 3. For regression (b), new activations (learned from zeros) are best. Moreover, accuracy and complexity are clearly correlated only for regression. This supports the hypothesis that regression is more complex than classification and thus benefits from alternatives to the ReLUs' simplicity bias. See Figure 18 for similar results on other datasets. + +![](images/022bc003f6aa1b5ffb090bf5ca3e5cc871e8994b0ce0ea05eaeca31b91ca2e79.jpg) + +tion 3.1. The task is now to directly predict class IDs. E.g. for MNIST this means predicting digit values. Models are trained with an MSE loss. To measure accuracy, we discretize the predictions to the nearest class ID. + +Results. The first observation from Figure 2b is that regression is clearly more difficult for NNs than classification (lower accuracies) despite the identical underlying task. Importantly, the learned activations now provide clear improvements, especially when learned from scratch (initialization as zeros). This confirms the hypothesis that the inductive biases of RLUs are not well suited to these tasks. + +Figure 3b shows that the learned activations contain more irregularities for regression than classification. Prior work [96] showed that this can help models represent complex functions with sharp transitions. An analysis of the complexity of trained models (Figures 4 and 18) shows that the accuracy is correlated with complexity for regression but not classification. And regression models with learned activations implement functions of higher complexity than with ReLUs. This supports the claim that the improvements arise from overcoming the simplicity bias of ReLUs. + +Complexity is only one dimension of the inductive biases. The complexity plots for SVHN (Figure 18) interestingly show that models with ReLUs and learned activations get different accuracies at the same complexity level. This shows that our meta learning approach can search over dimensions of the inductive biases that are not captured by our complexity measure, and are yet to be explicitly studied. + +Take-away: regression is more difficult for NNs than classification, and the simplicity bias of ReLUs is partly to blame. Learned activations improve performance by helping networks represent more complex functions. + +# 3.3. Tabular Data + +Background. Tabular data is any data with few unstructured dimensions, which often contains low-cardinality variables such as dates or categorical attributes. This contrasts e.g. with images, which contain many correlated, continuous dimensions (pixels). NNs struggle with tabular datasets and are often inferior to decision trees [35, 63]. Our hypothesis is that the inductive biases of standard architectures are ill-suited to such data because of the simplicity bias. It makes it difficult to learn functions where small changes in the input (e.g. day of the week) correspond to abrupt changes in the target — the definition of sensitivity, a proxy for complexity [17]. This seldom occurs in vision where similar images correspond to similar labels. + +Setup. We use 16 real-world classification datasets from Grinsztajn et al. [34, 35]. Bases include a linear classifier, k-NNs, and boosted decision trees. Our models are MLPs with 1-4 hidden layers (details in Appendix E.4). We compare learned activations functions with ReLUs and TanHs with a global prefactor, $\tanh (\alpha x)$ with $\alpha \in \mathbb{R}^{+}$ tuned on the validation set. This is a simple option with tunable complexity, albeit with inductive biases of TanHs [45, 96]. + +We also experiment with learned input activation functions (IAFs). The motivation is to learn a different behavior for each input dimension. Since they carry different information, e.g. continuous vs. categorical variables, one could be suited to the simplicity bias while another is not, for example. IAFs are dimension-specific activation functions applied directly on the data before a standard MLP. IAFs are learned like AFs, from an initialization as the identity i.e. no effect by default. They subsume the gated inputs, and Fourier/numerical embeddings from prior work [20, 25, 32]. + +Results. We compare the accuracy of models on the 16 datasets in Figures 5 and 20. Vanilla MLPs generally perform worse than trees. But adjusting the MLPs' inductive biases with learned prefactors or activations eliminates the gap. IAFs perform best, sometimes even surpassing trees. We analyze below the reasons for these improvements. + +![](images/0ab3f5a495bc33f3325173094c344ecd7742e560a8f61f327615ce132c794e8f.jpg) +Figure 5. Comparison of model types over 16 tabular datasets. Vanilla MLPs often perform worse than decision trees, but adjusting their inductive biases with learned activation functions (AFs) eliminates this gap. The input activation functions (IAFs) enable even better performance. See Figure 20 for results per dataset. + +![](images/dfa624194d959894bb7c0fdccd577ee48cf3651ba0a3e12c86c4ffafb2bd850f.jpg) +Best at low complexity + +![](images/d792fd28b74c36d0e1e7b3f038deb73197872e912e2fbf15d0920cf1edfa28ab.jpg) +Figure 6. Test accuracy vs. complexity on tabular datasets. Each marker represents a model with different hyperparameters, and ReLUs $(\bullet)$ or learned activations initialized as ReLUs $(\square)$ or as zeros $(\square)$ . The learned activations perform better in all cases, but the accuracy peaks at different complexity levels. For some datasets, a low complexity is best and ReLUs thus perform quite well (left-most panel, note the smaller Y scale). For other datasets, the opposite is true and the improvements with learned activations is larger. + +![](images/d04604de4f41000a298bdba6678ce37e51b87979e70ffb5ddead231238869014.jpg) +Different datasets +at datasets + +![](images/7c8516fa48acd91e8d3a370fe2998112d4c47833fd599945361ea07792df2ef4.jpg) +Best at higher complexity $\uparrow$ + +Learned activation functions close the gap to decision trees by mimicking their inductive bias. We visualize in Figure 7c the functions implemented by different models, plotting their output over slices of the input space (Section 2.1). ReLUs produce the smoothest function while TanHs and learned activations induce sharper patterns. Notably, the IAFs induce sharp axis-aligned decision boundaries that are also characteristic of trees, with which they share a high accuracy. Axis-aligned transitions are the consequence of IAFs applied independently to each dimension. Sharp transitions originate from the complex shape of the learned activation function (Figure 7a) which is possible thanks to the unbiased spline parametrization (Section 2.3). + +The simplicity/complexity bias is a property of the architecture. We visualize complexity landscapes of MLPs in Figure 7b. Similarly to standard loss landscapes [54], we plot model complexity over 2D slices of the parameter space. A first global view over a plane aligned with the training trajectory shows that complexity steadily increases during training for all models [48, 75, 104] but does so to the highest level for the best model (IAFs). A second view zooms in on each optimized solution in a random 2D plane. This examines the effect of arbitrary perturbations to the parameters. It shows that the ambient complexity of perturbed solutions of the best model is much higher than the solution itself, and then with less accurate models. This means that this architecture is more likely to represent complex functions because they are more abundant in parameter space [65, 84]. This is why the simplicity bias can be overcome: it results from architecture choices and not from an inevitable "implicit bias" of SGD [85, 89, 98, 105]. + +Different tabular datasets require different inductive biases. We examine the relation between accuracy and complexity in Figure 6. The accuracy peaks at different complexity levels for different datasets. For some, a low complexity is best and ReLU MLPs perform well. For oth + +(a) Activation function & loss landscape (—training trajectory, $\star$ early-stopping checkpoint, $\bullet$ last checkpoint) + +![](images/bad6edef9c1bc8728394cdffe30077a78da01f0153f0f1e2b12cd2182f52e52a.jpg) + +![](images/6fd9e9d9dcd0f06ba6459546a2aa20b0438c7dbdff1b349bcb51e7b3c2ea0948.jpg) + +![](images/f16688b3b136c6456aa0b3130512d904bec0357624b820eb582478f07a0d98a5.jpg) + +![](images/99f44ad2a7cdc1277b102e49942fded6ae5efa58eb372f813762876ff5c44f63.jpg) + +![](images/1cdf1bd89d8f8b02937ffa5c9ccbc1c9f9ba6dda970a559b5727aeb0eb968311.jpg) + +(b) Complexity landscape (in weight space along the PCA plane of the training trajectory) & zoom-in (random plane) + +![](images/c752fddc5da611e26c16531aac4e01617512987ab02341ed960dd2558d73ceb4.jpg) + +![](images/c75597ea54fed9cf56bc03e79bbd5b1c3181f189013e5ee675e8f5575f82703d.jpg) + +![](images/ec77cc9d600b0af491e2b18885eef5e0fb42f9dbce7b4e0069a2b003b0ed9c6d.jpg) + +![](images/4bdcc9d70736b99d122ec4d8825092abe4aa1641cb36ba0638504d8998ff0304.jpg) + +![](images/a05c6b4ff26d0923effae88c0343e254fbba0fcf68ebb66d679995d464212742.jpg) + +(c) Function implemented by the network (in input space along four random planes containing each one training point $\bullet$ + +![](images/54f22c31c5781e224a169e8e7103f2010d27bd9217944f874670083a202ee4ce.jpg) +MLP, ReLU + +MLP, TanH w/tuned prefactor +![](images/d4288b4e375ddc9c5004153cc157bb14a828d67f3b7167e3dbd74c8af17a4ba8.jpg) +Increasing accuracy Best MLP + +![](images/f5ba720390310337ea1ef721280057a57d81d2394ffe14a47dbc35040f6d3c3c.jpg) +MLP, learned AF + +![](images/7c494a259530c3f2741489b48c8a6b8c473170b91091add9daa34e9b85069111.jpg) +MLP, learned IAFs + +![](images/19bbd88571bfd919c19a65ad324180c481896a00f036a7ad0c321438a13ee397.jpg) +Boosted decision trees + +ers, a higher complexity is best and the improvements with learned activations are larger. This supports the hypothesis that improvements over ReLU MLPs come from overcoming their simplicity bias. The variance across datasets is also unsurprising since they have little in common besides their low dimensionality (full results in Appendix E.4). + +Effect of width and depth. We show in Figure 8 that the learned activations can be reused in networks of different widths than they were trained for. The accuracy varies with width similarly as with ReLUs. Teney et al. [96] indeed showed that a model's width affects its capacity but not its inductive biases. Therefore width does not interfere with the effects of the learned activations. Figures 8 and 21 also show that good performance can be achieved with fewer layers than with ReLUs. Learned activations might thus have utility in model compression and distillation. + +Take-away: many tabular datasets are ill-suited to ReLU models because they require learning a complex function. Learned activations improve accuracy by implementing sharp axis-aligned decision boundaries that mimic the inductive biases of decision trees. + +# 3.4. Shortcut Learning + +Background. Shortcut learning occurs when a model learns spurious features instead of generalizable ones. It is a known consequence of the simplicity bias [85, 93] when the training data contains multiple features of different complexity. Our hypothesis is that the preference for some fea + +![](images/5ad5edb24e4316146b38f696dd7a198fbb9ef5cfb04f3351a0cae4611512107e.jpg) +Figure 7. Models trained on the ELECTRICITY [34] tabular dataset. ReLU MLPs perform worst (left). TanHs induce sharper transitions in the network's function (c). So does the learned activation function (a) which is itself very irregular. The input activation functions (IAFs) perform best and mimic the axis-aligned boundaries of trees (bottom-right). The complexity landscapes (b) show that complexity increases to the highest level in the best model (IAFs). The zoom-in shows that the ambient complexity is also much higher than in other models. This means that it is inherently more likely to represent complex functions since they are more abundant in parameter space. +Figure 8. The learned activation functions surpass ReLUs, often with fewer layers. They can also be reused with different network widths (COVERTYPE [34] tabular dataset, see Figure 21 for others). + +tures depends on their alignment with the inductive biases. We will evaluate whether this can be controlled with activation functions. + +Setup. We use MNIST/CIFAR collages [85, 93, 94], a classification task over images combining tiles from MNIST and CIFAR-10. The training set is ambiguous: both tiles are predictive of the labels. Two unambiguous test sets evaluate reliance on either tile: one is predictive, the other contains a random class. We similarly build two validation sets to learn two activation functions optimized for either tile. We simulate OOD conditions by setting $\mathcal{V}$ in Algorithm 1. The models are the fully-connected MLPs used in [93]. + +Results. Figure 9 shows that a baseline with ReLUs is prone to shortcut learning. It relies exclusively on MNIST and the accuracy on the CIFAR test set is not better than chance (10%). In comparison, using either learned activation steers the learning towards either tile. The accuracy + +![](images/d6b247dcca639738fdf5d443fe4f7bfe462550793593ecf7228ca8b6bf910e45.jpg) +Ambiguous training images + +![](images/c684f7795bd28c20580a087965e31eb6cc8ac02ab707ed5e709cc79ec3e5d2ac.jpg) +Optimized for MNIST + +![](images/50c38e44d937b34048fe3365f9b2308f201750ababa6cde9e85bc8c5ddf896b4.jpg) +Optimized for CIFAR + +![](images/a04e48553aa4a0d68c21ff4539f2373cf52f17a1413cee69f194d411e026db28.jpg) +Upper bound: ReLU trained on unambiguous MNIST Activation optimized for MNIST ReLU Baseline Activation optimized for CIFAR Upper bound: ReLU trained on unambiguous CIFAR + +![](images/301cbb3be4b74660f7cb9c81465c5f311bd49aa3c6329f6e809d035827b8f60b.jpg) + +shifts towards either of two tiles as the model prioritizes different features, merely with a change of activation function. This shows that the simplicity bias is not an inevitable effect of SGD. Instead, it directly reflects the alignment between the chosen architecture and the data. + +Training dynamics. The accuracy on CIFAR remains below a model trained on unambiguous CIFAR data. This is because training dynamics are also important. In Figure 9 (right), we plot the accuracy on the two tiles for the whole training trajectory. The reliance on different features varies, and the model eventually relies primarily on simple ones with enough iterations (i.e. without early stopping). This calls for future work combining our findings with the extensive literature on ID/OOD training dynamics [46, 95, 98]. + +Take-away: we confirm that shortcut learning is a side effect of the simplicity bias. Different activation functions, while not completely avoiding shortcut learning, can steer the learning towards particular input features. + +# 3.5. Algorithmic Tasks and Grokking + +Background. Grokking is a phenomenon where a model first overfits the data (i.e. high training accuracy, low test accuracy) then shifts to high test accuracy after many training steps [73]. This is typically observed on algorithmic tasks and architectures from MLPs to transformers. Our hypothesis is that grokking is due to a mismatch between the target function and the model's inductive biases. Indeed, typical architectures were not developed for the algorithmic tasks where grokking is typically observed. To verify this hypothesis, we will show that endowing an architecture with the right inductive biases, using learned activation functions, can eliminate the phenomenon. Supporting this hypothesis, Zhou et al. [110] proposed that grokking comes from the frequency principle (i.e. low frequencies learned first by SGD), and Kumar et al. [51] showed that it correlates with a misalignment between features at initialization and the target function. + +Setup. Following [36, 51, 58] we train 1-hidden layer MLPs on algorithmic tasks, defined each by one binary operation (Figures 10 and 27). e.g. $y = (x_{1} + x_{2}) \mod 13$ . The operands are passed as one-hot vectors and the task is a classification over possible outputs. Details in Appendix E.6. + +![](images/d1d11deeb4a9b9b81cb958f58c49724164970864e0e5da9c8d343a1d345e6c86.jpg) +Figure 9. Experiments on shortcut learning with MNIST/CIFAR collages. The ReLU baseline (■) relies mostly on simple MNIST features. We learn two activation functions that shift the preference towards different features $(\leftarrow / \rightarrow)$ . Training trajectories (right) clearly differ with the activation optimized for CIFAR (—), for MNIST (—), or a ReLU (==). The model at initialization (random weights) is marked with $\bullet$ . +Figure 10. Target functions used to investigate grokking [73] (details in Appendix, Figure 27). These patterns are very different from the tasks for which typical architectures were developed. + +![](images/3dcee3491eae684eff053e6f0f384f70f597e8055a8aa1e15456e74be9a04da1.jpg) + +![](images/cd388dbc24fcdf1dd019094bd395c93e32996e3d3a1373fed0218c36523683df.jpg) + +![](images/44a767f71cfd239d8d47af4809f0d81fb951876d5eff84ae58a7f0445a81e582.jpg) + +Results. We compare models with ReLU vs. learned activations across various tasks, network widths, and fractions of training data. We find that the learned, task-specific activations lead to faster convergence and/or higher test accuracy (Figures 11-13). On modular addition (a common task in the grokking literature) the learned-activation model converges $\sim 10\times$ faster than ReLUs. Curiously, some models with learned activations also end up overfitting (decreasing test accuracy) with prolonged training. In contrast, ReLU networks either never generalize (test accuracy $\sim 0$ ) or grok and keep a high accuracy indefinitely. Further investigation is needed to explain this difference. We examine learned activation functions in Figure 12. See Figure 28 in the Appendix for results on other algorithmic tasks [73]. + +![](images/9920403831ee31eaddfe832ce9630e4b9cd06b614de2382e9554c6094a3e8800.jpg) +Figure 11. The learned activations essentially eliminate grokking (delayed convergence). On the above task (addition mod 27), our model converges $\sim 10\times$ faster than ReLUs. + +![](images/83e4f3ac0255d05e53e3fa527646c06887f09c4c75c47fd1dc7899320d659641.jpg) + +![](images/b5b4266d96db9f14d1a6c8a2fe2d628a45a4da699f3ee05d5f3f7357a326fefc.jpg) +Figure 12. Activations learned for modular addition. The frequency of the sine-like function varies across versions of the task. + +![](images/65eee4c6eee6e76429bc7140d09f002af221947e030d304d5c5245afa7708ae2.jpg) + +![](images/294f53ae5614dc603af10359800aaf25913da0c286cb71d8b322f390eb549d1d.jpg) + +![](images/ee32959f3b5caec906a43d63ac57bf87352145a3ed5d6be36ab461f7a76ba870.jpg) + +![](images/d2f0504c3ce9c9632681a94e323a57ef64b3ce024d7b4d7f7b22b2c8da9e9cc3.jpg) +Figure 13. Models with learned activations $(-)$ converge faster than ReLUs $(--)$ across a variety of settings (addition mod 27). + +![](images/053a8c1203fb699997e086dc2755219911c6bd571080ca3bd2683dd09be94139.jpg) + +Take-away: learned activations eliminate grokking in all our cases, suggesting, as a cause, the mismatch between the data and the architectures' inductive biases. + +# 4. Do the Activations Transfer Across Tasks? + +So far, we used dataset-specific activation functions and found that there exist better alternatives to ReLUs. A practical application would be the learning of activation functions suitable to a broad task, or range of related datasets. + +As a first step, we study the specialization of the activations functions (AFs) learned for the 22 algorithmic tasks from Figure 27 [73]. We evaluate every task/activation combination, yielding the $22 \times 22$ matrix of Figure 14. The learned activations do transfer, with improvements in accuracy and convergence shared across tasks. We also evaluate an activation learned on all tasks simultaneously. The accuracy across tasks (i.e. per-column average) reaches $61.5\%$ vs. only $19.9\%$ for ReLUs. and $54.0\%$ on average for tasks-specific solutions. This procedure can thus improve performance on a range of related tasks. Future work could leverage it to discover activation functions that improve performance in other specific domains. See Appendix E.3 for other transfer experiments using image regression tasks. + +![](images/616059b4a69b28d57058a7ca58c82ad3cebff55fedc8cedb0c41fe1f7be2a3a3.jpg) +Figure 14. Transfer of AFs (columns) across algorithmic tasks (rows). Colors represent the fraction of the best convergence speed or accuracy per task (brighter is better). If the activations were over-specialized, the matrix would be diagonal. On the contrary, it is densely filled, indicating positive transfer across many tasks. + +# 5. Discussion + +We used activation functions as a tool to show that there exist a variety of inductive biases that are useful across applications of NNs. The impossibility of universal inductive biases is well known [101] but a strong argument has also been made that deep learning research is converging towards few architectures with wide applicability [31]. This argument rests on the NNs' simplicity bias being a good match for real-world data [12, 18, 56]. Our results do not invalidate these assumptions: NNs are widely applicable and their simplicity bias is evidently very effective on average. Our results show instead the following. + +1. There exist real-world tasks where the inductive biases of typical architectures' are suboptimal. This explanation connects four domains where NNs historically struggled. +2. The simplicity bias in modern NNs depends on particular design choices, the activation functions in particular. Research has converged on these choices by trial and error, in large part by optimizing performance on vision tasks. Therefore the adequacy of ReLUs for image classification (Section 3.1) is not accidental. + +Relevance to transformers and language models. The simplicity bias exists in transformers [10, 109] and language models [4, 31, 96]. Their embedding layer resembles input activations functions (Section 3.3). Could this explain the transformers' remarkable flexibility? I.e. a simplicity bias on an initial mapping of arbitrary complexity. Zhong and Andreas [108] indeed trained embeddings alone in a random-weight transformer and could learn complex tasks. + +Limitations and open questions. We prioritized breadth by establishing a new connection across multiple disparate topics in machine learning. Each section could expand into its own paper with additional models, datasets, comparisons, etc. Our findings on shortcut learning for example (Section 3.4) could yield new methods to address distribution shifts, though no such claim is made here. Here are the most promising follow-up questions opened by this paper. + +- How to fully characterize inductive biases? We focused on simplicity for its prevalence in AI [31], philosophy [72], and the natural sciences [12]. But it is only one dimension among many to characterize inductive biases. +- Can we improve state-of-the-art architectures? We used simple MLPs to isolate the effects of activations functions since they are central to the simplicity bias [96]. But other existing mechanisms (architectural, optimization) may already tweak or attenuate the simplicity bias. +- Can we learn transferable activation functions for other domains? We examined transferability in Sections 4 and E.3. The results suggest the possibility of better architectures optimized for specific domains. Predicting the suitability of an architecture/dataset pair ex ante (prior to training) would be extremely useful. This may follow from advances on the first open question above. +- Are there other detrimental effects of the simplicity bias? Any learning algorithm needs inductive biases to "fill the gaps" between training examples. The better they are, the fewer examples are needed. Researching what inductive biases are most useful on real-world tasks might thus hold the key for machine learning to become as data-efficient as humans. More speculatively, high-level cognition has been argued to require postulating explanations beyond the data [16, 23]. In this regard, simplicity-biased architectures might also hold us back. + +# References + +[1] Sravanti Addepalli, Anshul Nasery, Venkatesh Babu Radhakrishnan, Praneeth Netrapalli, and Prateek Jain. Feature reconstruction from outputs can mitigate simplicity bias in neural networks. In ICLR, 2022. 1 +[2] Konstantinos Panagiotis Alexandridis, Jiankang Deng, Anh Nguyen, and Shan Luo. Adaptive parametric activation. arXiv preprint arXiv:2407.08567, 2024. 2, 3, 1 +[3] Konstantinos Panagiotis Alexandridis, Jiankang Deng, Anh Nguyen, and Shan Luo. Adaptive parametric activation. In ECCV, 2025. 2 +[4] Badr AlKhamissi, Greta Tuckute, Antoine Bosselut, and Martin Schrimpf. Brain-like language processing via a shallow untrained multihead attention network. arXiv preprint arXiv:2406.15109, 2024. 8 +[5] Andrea Apicella, Francesco Isgro, and Roberto Prevete. A simple and efficient architecture for trainable activation functions. Neurocomputing, 2019. 3, 1, 2 +[6] Andrea Apicella, Francesco Donnarumma, Francesco Isgro, and Roberto Prevete. A survey on modern trainable activation functions. Neural Networks, 2021. 2, 3, 1 +[7] Sanjeev Arora, Nadav Cohen, Wei Hu, and Yuping Luo. Implicit regularization in deep matrix factorization. NeurIPS, 2019. 1 +[8] Devansh Arpit, Stanislaw Jastrzebski, Nicolas Ballas, David Krueger, Emmanuel Bengio, Maxinder S Kanwal, Tegan Maharaj, Asja Fischer, Aaron Courville, Yoshua Bengio, et al. A closer look at memorization in deep networks. In ICML. PMLR, 2017. 1 +[9] Samuel James Bell and Levent Sagun. Simplicity bias leads to amplified performance disparities. In Proceedings of the 2023 ACM Conference on Fairness, Accountability, and Transparency, pages 355-369, 2023. 1 +[10] Satwik Bhattachamishra, Arkil Patel, Varun Kanade, and Phil Blunsom. Simplicity bias in transformers and their ability to learn sparse boolean functions. arXiv preprint arXiv:2211.12316, 2022.8 +[11] Garrett Bingham, William Macke, and Risto Miikkulainen. Evolutionary optimization of deep learning activation functions. In Genetic and Evolutionary Computation Conference, 2020. 3, 4, 1, 2 +[12] Mark Buchanan. A natural bias for simplicity. Nature Physics, 2018. 8 +[13] Irit Chelly, Shahaf E Finder, Shira Ifergane, and Oren Freifeld. Trainable highly-expressive activation functions. arXiv preprint arXiv:2407.07564, 2024. 3, 1, 2 +[14] Nadav Cohen and Amnon Shashua. Inductive bias of deep convolutional networks through pooling geometry. arXiv preprint arXiv:1605.06743, 2016. 1 +[15] Giacomo De Palma, Bobak Kiani, and Seth Lloyd. Random deep neural networks are biased towards simple functions. NeurIPS, 2019. 1, 5 +[16] David Deutsch. The beginning of infinity: Explanations that transform the world. Penguin, 2011. 8 +[17] Benoit Dherin, Michael Munn, Mihaela Rosca, and David Barrett. Why neural networks find simple solutions: The + +many regularizers of geometric complexity. NeurIPS, 35, 2022. 1, 4, 5 +[18] Kamaludin Dingle, Chico Q Camargo, and Ard A Louis. Input-output maps are strongly biased towards simple outputs. Nature communications, 2018. 1, 8 +[19] Pedro Domingos. The role of occam's razor in knowledge discovery. Data mining and knowledge discovery, 1999. 1 +[20] Marius Dragoi, Florin Gogianu, and Elena Burceanu. Closing the gap on tabular data with fourier and implicit categorical features. Submission to ICLR (available on Open-Review), 2024. 5, 1 +[21] Shiv Ram Dubey, Satish Kumar Singh, and Bidyut Baran Chaudhuri. Activation functions in deep learning: A comprehensive survey and benchmark. Neurocomputing, 2022. 1 +[22] Stanislas Ducotterd, Alexis Goujon, Pakshal Bohra, Dimitris Perdios, Sebastian Neumayer, and Michael Unser. Improving lipschitz-constrained neural networks by learning activation functions. Journal of Machine Learning Research, 2024. 3, 1, 2 +[23] Daniel C Elton. Applying deutsch's concept of good explanations to artificial intelligence and neuroscience—an initial exploration. Cognitive Systems Research, 2021. 8 +[24] Jesse Farebrother, Jordi Orbay, Quan Vuong, Adrien Ali Taiga, Yevgen Chebotar, Ted Xiao, Alex Irpan, Sergey Levine, Pablo Samuel Castro, Aleksandra Faust, et al. Stop regressing: Training value functions via classification for scalable deep rl. arXiv preprint arXiv:2403.03950, 2024. 4 +[25] James Fiedler. Simple modifications to improve tabular neural networks. CoRR, abs/2108.03214, 2021. 5 +[26] Sara Fridovich-Keil, Raphael Gontijo Lopes, and Rebecca Roelofs. Spectral bias in practice: The role of function frequency in generalization. NeurIPS, 2022. 2 +[27] Gallant. There exists a neural network that does not make avoidable mistakes. In IEEE International Conference on Neural Networks. IEEE, 1988. 1 +[28] Isabel O Gallegos, Ryan A Rossi, Joe Barrow, Md Mehrab Tanjim, Sungchul Kim, Franck Dernoncourt, Tong Yu, Ruiyi Zhang, and Nesreen K Ahmed. Bias and fairness in large language models: A survey. Computational Linguistics, 2024. 1 +[29] Robert Geirhos, Jorn-Henrik Jacobsen, Claudio Michaelis, Richard Zemel, Wieland Brendel, Matthias Bethge, and Felix A Wichmann. Shortcut learning in deep neural networks. Nature Machine Intelligence, 2020. 1 +[30] Florin Gogianu, Tudor Berariu, Mihaela C Rosca, Claudia Clopath, Lucian Busoniu, and Razvan Pascanu. Spectral normalisation for deep reinforcement learning: an optimisation perspective. In ICML, 2021. 11 +[31] Micah Goldblum, Marc Finzi, Keefer Rowan, and Andrew Gordon Wilson. The no free lunch theorem, kolmogorov complexity, and the role of inductive biases in machine learning. arXiv preprint arXiv:2304.05366, 2023. 1, 8 +[32] Yury Gorishniy, Ivan Rubachev, and Artem Babenko. On embeddings for numerical features in tabular deep learning. NeurIPS, 2022. 5, 1, 9 + +[33] Anirudh Goyal and Yoshua Bengio. Inductive biases for deep learning of higher-level cognition. Proceedings of the Royal Society A, 2022. 1 +[34] Léo Grinsztajn. Tabular data learning benchmark. https://github.com/LéoGrin/tabular-benchmark, 2022.5,6, 9,10,14 +[35] Léo Grinsztajn, Edouard Oyallon, and Gael Varoquaux. Why do tree-based models still outperform deep learning on typical tabular data? NeurIPS, 2022. 4, 5, 1 +[36] Andrey Gromov. Grokking modular arithmetic. arXiv preprint arXiv:2301.02679, 2023. 7 +[37] Michael Hahn, Dan Jurafsky, and Richard Futrell. Sensitivity as a complexity measure for sequence classification tasks. Transactions of the ACL, 2021. 1 +[38] Hrayr Harutyunyan, Rafayel Darbinyan, Samvel Karapetyan, and Hrant Khachatrian. In-context learning in presence of spurious correlations. arXiv preprint arXiv:2410.03140, 2024. 1 +[39] Dan Hendrycks and Kevin Gimpel. Gaussian error linear units (GeLUs). arXiv preprint arXiv:1606.08415, 2016. 2, 1 +[40] Katherine L Hermann and Andrew K Lampinen. What shapes feature representations? exploring datasets, architectures, and training. arXiv preprint arXiv:2006.12433, 2020. 1 +[41] Hossein Hosseini, Baicen Xiao, Mayoore Jaiswal, and Radha Poovendran. Assessing shape bias property of convolutional neural networks. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition Workshops, 2018. 1 +[42] Like Hui and Mikhail Belkin. Evaluation of neural architectures trained with square loss vs cross-entropy in classification tasks. arXiv preprint arXiv:2006.07322, 2020.1 +[43] Ehsan Imani, Kai Luedemann, Sam Scholnick-Hughes, Esraa Elelimy, and Martha White. Investigating the histogram loss in regression. arXiv preprint arXiv:2402.13425, 2024. 4 +[44] Ameya D Jagtap and George Em Karniadakis. How important are activation functions in regression and classification? a survey, performance comparison, and future directions. Journal of Machine Learning for Modeling and Computing, 4(1), 2023. 1 +[45] Ameya D Jagtap, Kenji Kawaguchi, and George Em Karniadakis. Adaptive activation functions accelerate convergence in deep and physics-informed neural networks. Journal of Computational Physics, 2020. 3, 5, 1, 2 +[46] Anchit Jain, Rozhin Nobahari, Aristide Baratin, and Stefano Sarao Mannelli. Bias in motion: Theoretical insights into the dynamics of bias in sgd training. arXiv preprint arXiv:2405.18296, 2024.7 +[47] Liangze Jiang and Damien Teney. OOD-chameleon: Is algorithm selection for ood generalization learnable? arXiv preprint arXiv:2410.02735, 2024. 1 +[48] Dimitris Kalimeris, Gal Kaplun, Preetum Nakkiran, Benjamin Edelman, Tristan Yang, Boaz Barak, and Haofeng Zhang. SGD on neural networks learns functions of increasing complexity. NeurIPS, 2019. 5, 1 + +[49] Alex Krizhevsky and Geoffrey Hinton. Learning multiple layers of features from tiny images. Technical report, University of Toronto, 2009. 3 +[50] Jan Kukačka, Vladimir Golkov, and Daniel Cremers. Regularization for deep learning: A taxonomy. arXiv preprint arXiv:1710.10686, 2017. 1 +[51] Tanishq Kumar, Blake Bordelon, Samuel J Gershman, and Cengiz Pehlevan. Grokking as the transition from lazy to rich training dynamics. arXiv preprint arXiv:2310.06110, 2023. 7 +[52] Yann LeCun, Léon Bottou, Yoshua Bengio, and Patrick Haffner. Gradient-based learning applied to document recognition. Proceedings of the IEEE, 86(11):2278-2324, 1998. 3 +[53] Alexander Li and Deepak Pathak. Functional regularization for reinforcement learning via learned fourier features. NeurIPS, 2021. 1 +[54] Hao Li, Zheng Xu, Gavin Taylor, Christoph Studer, and Tom Goldstein. Visualizing the loss landscape of neural nets. NeurIPS, 2018. 5 +[55] Xuan Li, Yun Wang, and Bo Li. Tree-regularized tabular embeddings. arXiv preprint arXiv:2403.00963, 2024. 1 +[56] Henry W Lin, Max Tegmark, and David Rolnick. Why does deep and cheap learning work so well? Journal of Statistical Physics, 2017. 8 +[57] Ziqi Liu, Wei Cai, and Zhi-Qin John Xu. Multi-scale deep neural network (mscalednn) for solving poisson-boltzmann equation in complex domains. arXiv preprint arXiv:2007.11207, 2020. 1 +[58] Ziming Liu, Ouail Kitouni, Niklas S Nolte, Eric Michaud, Max Tegmark, and Mike Williams. Towards understanding grokking: An effective theory of representation learning. NeurIPS, 2022. 7 +[59] Ziming Liu, Yixuan Wang, Sachin Vaidya, Fabian Ruehle, James Halverson, Marin Soljačić, Thomas Y Hou, and Max Tegmark. Kan: Kolmogorov-arnold networks. arXiv preprint arXiv:2404.19756, 2024. 2 +[60] Kaifeng Lyu, Zhiyuan Li, Runzhe Wang, and Sanjeev Arora. Gradient descent on two-layer nets: Margin maximization and simplicity bias. NeurIPS, 2021. 1 +[61] Andrew L Maas, Awni Y Hannun, Andrew Y Ng, et al. Rectifier nonlinearities improve neural network acoustic models. In ICML, 2013. 1 +[62] Augustine N Mavor-Parker, Matthew J Sargent, Caswell Barry, Lewis Griffin, and Clare Lyle. Frequency and generalisation of periodic activation functions in reinforcement learning. arXiv preprint arXiv:2407.06756, 2024. 1 +[63] Duncan McElfresh, Sujay Khandagale, Jonathan Valverde, Vishak Prasad C, Ganesh Ramakrishnan, Micah Goldblum, and Colin White. When do neural nets outperform boosted trees on tabular data? NeurIPS, 2024. 5 +[64] Chris Mingard, Joar Skalse, Guillermo Valle-Pérez, David Martínez-Rubio, Vladimir Mikulik, and Ard A Louis. Neural networks are a priori biased towards boolean functions with low entropy. arXiv preprint arXiv:1909.11522, 2019. 1, 5 + +[65] Chris Mingard, Guillermo Valle-Pérez, Joar Skalse, and Ard A Louis. Is SGD a bayesian sampler? well, almost. Journal of Machine Learning Research, 2021. 1, 5 +[66] Chris Mingard, Henry Rees, Guillermo Valle-Pérez, and Ard A Louis. Do deep neural networks have an inbuilt ocean's razor? arXiv preprint arXiv:2304.06670, 2023. 1 +[67] Tom M Mitchell. The need for biases in learning generalizations. Rutgers University CS tech report CBM-TR-117, 1980. 1 +[68] Yuval Netzer, Tao Wang, Adam Coates, Alessandro Bissacco, Bo Wu, and Andrew Y Ng. Reading digits in natural images with unsupervised feature learning. NIPS Workshop on Deep Learning and Unsupervised Feature Learning, 2011. 3 +[69] Behnam Neyshabur, Ryota Tomioka, and Nathan Srebro. In search of the real inductive bias: On the role of implicit regularization in deep learning. arXiv preprint arXiv:1412.6614, 2014. 1 +[70] Mohammad Pezeshki, Oumar Kaba, Yoshua Bengio, Aaron C Courville, Doina Precup, and Guillaume Lajoie. Gradient starvation: A learning proclivity in neural networks. NeurIPS, 2021. 1 +[71] Tomaso Poggio, Kenji Kawaguchi, Qianli Liao, Brando Miranda, Lorenzo Rosasco, Xavier Boix, Jack Hiday, and Hrushikesh Mhaskar. Theory of deep learning III: the nonoverfitting puzzle. CBMM Memo, 2018. 1 +[72] Karl Popper. "7. Simplicity". The logic of scientific discovery. Routledge, 1959. 8 +[73] Alethea Power, Yuri Burda, Harri Edwards, Igor Babuschkin, and Vedant Misra. Grokking: Generalization beyond overfitting on small algorithmic datasets. arXiv preprint arXiv:2201.02177, 2022. 7, 8, 12 +[74] Aahlad Manas Puli, Lily Zhang, Yoav Wald, and Rajesh Ranganath. Don't blame dataset shift! shortcut learning due to gradients and cross entropy. NeurIPS, 2023. 1 +[75] Nasim Rahaman, Aristide Baratin, Devansh Arpit, Felix Draxler, Min Lin, Fred Hamprecht, Yoshua Bengio, and Aaron Courville. On the spectral bias of neural networks. In ICML. PMLR, 2019. 5, 1 +[76] Prajit Ramachandran, Barret Zoph, and Quoc V Le. Swish: a self-gated activation function. arXiv preprint arXiv:1710.05941, 2017. 1 +[77] Sameera Ramasinghe and Simon Lucey. Beyond periodicity: Towards a unifying framework for activations in coordinate-MLPs. In ECCV. Springer, 2022. 1 +[78] Sameera Ramasinghe, Lachlan E MacDonald, and Simon Lucey. On the frequency-bias of coordinate-MLPs. NeurIPS, 2022. 1 +[79] Mihaela Rosca, Theophane Weber, Arthur Gretton, and Shakir Mohamed. A case for new neural network smoothness constraints. I Can't Believe It's Not Better (ICBINB) Workshop at NeurIPS, 2020. 11 +[80] Vishwanath Saragadam, Daniel LeJeune, Jasper Tan, Guha Balakrishnan, Ashok Veeraraghavan, and Richard G Baraniuk. Wire: Wavelet implicit neural representations. In CVPR, 2023. 4, 1 + +[81] Simone Scardapane, Michele Scarpiniti, Danilo Comminiello, and Aurelio Uncini. Learning activation functions from data using cubic spline interpolation. Neural Advances in Processing Nonlinear Dynamic Signals, 2019. 2 +[82] Simone Scardapane, Steven Van Vaerenbergh, Simone Toto, and Aurelio Uncini. Kafnets: Kernel-based nonparametric activation functions for neural networks. Neural Networks, 2019. 2, 3, 1 +[83] Jürgen Schmidhuber. Discovering neural nets with low kolmogorov complexity and high generalization capability. Neural Networks, 1997. 1 +[84] Luca Scimeca, Seong Joon Oh, Sanghyuk Chun, Michael Poli, and Sangdoo Yun. Which shortcut cues will DNNs choose? a study from the parameter-space perspective. arXiv preprint arXiv:2110.03095, 2021. 5 +[85] Harshay Shah, Kaustav Tamuly, Aditi Raghunathan, Prateek Jain, and Praneeth Netrapalli. The pitfalls of simplicity bias in neural networks. NeurIPS, 2020. 2, 5, 6, 11 +[86] Kexuan Shi, Xingyu Zhou, and Shuhang Gu. Improved implicit neural representation with fourier reparameterized training. In CVPR, 2024. 1 +[87] Prasann Singhal, Tanya Goyal, Jiacheng Xu, and Greg Durrett. A long way to go: Investigating length correlations in rlhf. arXiv preprint arXiv:2310.03716, 2023. 1 +[88] Vincent Sitzmann, Julien Martel, Alexander Bergman, David Lindell, and Gordon Wetzstein. Implicit neural representations with periodic activation functions. NeurIPS, 2020. 1 +[89] Daniel Soudry, Elad Hoffer, Mor Shpigel Nacson, Suriya Gunasekar, and Nathan Srebro. The implicit bias of gradient descent on separable data. The Journal of Machine Learning Research, 2018. 5 +[90] Lawrence Stewart, Francis Bach, Quentin Berthet, and Jean-Philippe Vert. Regression as classification: Influence of task formulation on neural network features. In ICML. PMLR, 2023. 4 +[91] Leon René Sutfeld, Flemming Brieger, Holger Finger, Sonja Fülhase, and Gordon Pipa. Adaptive blending units: Trainable activation functions for deep neural networks. In Intelligent Computing: Proceedings of the Computing Conference. Springer, 2020. 3, 1, 2 +[92] Remi Tachet, Mohammad Pezeshki, Samira Shabanian, Aaron Courville, and Yoshua Bengio. On the learning dynamics of deep neural networks. arXiv preprint arXiv:1809.06848, 2018. 1 +[93] Damien Teney, Ehsan Abbasnejad, Simon Lucey, and Anton van den Hengel. Evading the simplicity bias: Training a diverse set of models discovers solutions with superior ood generalization. arXiv preprint arXiv:2105.05612, 2021. 1, 2, 6, 11 +[94] Damien Teney, Maxime Peyrard, and Ehsan Abbasnejad. Predicting is not understanding: Recognizing and addressing underspecification in machine learning. In ECCV. Springer, 2022. 6, 1 +[95] Damien Teney, Yong Lin, Seong Joon Oh, and Ehsan Abbasnejad. ID and OOD performance are sometimes in + +versely correlated on real-world datasets. NeurIPS, 2024. 1, 7 +[96] Damien Teney, Armand Mihai Nicoliciou, Valentin Hartmann, and Ehsan Abbasnejad. Neural redshift: Random networks are not random functions. In CVPR, 2024. 1, 2, 3, 4, 5, 6, 8, 11, 14 +[97] Tijmen Tieleman. Lecture 6.5-rmsprop: Divide the gradient by a running average of its recent magnitude. COURSERA: Neural networks for machine learning, 4(2):26, 2012. 8 +[98] Nikita Tsoy and Nikola Konstantinov. Simplicity bias of two-layer networks beyond linearly separable data. arXiv preprint arXiv:2405.17299, 2024. 5, 7 +[99] Guillermo Valle-Perez, Chico Q Camargo, and Ard A Louis. Deep learning generalizes because the parameter-function map is biased towards simple functions. arXiv preprint arXiv:1805.08522, 2018. 1, 5 +[100] Colin White, Mahmoud Safari, Rhea Sukthanker, Binxin Ru, Thomas Elsken, Arber Zela, Debadeepta Dey, and Frank Hutter. Neural architecture search: Insights from 1000 papers. arXiv preprint arXiv:2301.08727, 2023. 2 +[101] David H Wolpert. The supervised learning no-free-lunch theorems. Soft computing and industry: Recent applications, 2002. 1, 8 +[102] Han Xiao, Kashif Rasul, and Roland Vollgraf. Fashion-mnist: A novel image dataset for benchmarking machine learning algorithms. arXiv preprint arXiv:1708.07747, 2017.3 +[103] Yiheng Xie, Towaki Takikawa, Shunsuke Saito, Or Litany, Shiqin Yan, Numair Khan, Federico Tombari, James Tompkin, Vincent Sitzmann, and Srinath Sridhar. Neural fields in visual computing and beyond. In Computer Graphics Forum. Wiley Online Library, 2022. 1 +[104] Zhi-Qin John Xu, Yaoyu Zhang, Tao Luo, Yanyang Xiao, and Zheng Ma. Frequency principle: Fourier analysis sheds light on deep neural networks. arXiv preprint arXiv:1901.06523, 2019. 5, 1 +[105] Zhi-Qin John Xu, Yaoyu Zhang, and Yanyang Xiao. Training behavior of deep neural network in frequency domain. In ICONIP. Springer, 2019. 5 +[106] Zhi-Qin John Xu, Yaoyu Zhang, and Tao Luo. Overview frequency principle/spectral bias in deep learning. Communications on Applied Mathematics and Computation, 2024. 1 +[107] Ge Yang, Anurag Ajay, and Pulkit Agrawal. Overcoming the spectral bias of neural value approximation. arXiv preprint arXiv:2206.04672, 2022. 1 +[108] Ziqian Zhong and Jacob Andreas. Algorithmic capabilities of random transformers. arXiv preprint arXiv:2410.04368, 2024. 8 +[109] Hattie Zhou, Arwen Bradley, Etai Littwin, Noam Razin, Omid Saremi, Josh Susskind, Samy Bengio, and Preetum Nakkiran. What algorithms can transformers learn? a study in length generalization. arXiv preprint arXiv:2310.16028, 2023. 8 +[110] Zhangchen Zhou, Yaoyu Zhang, and Zhi-Qin John Xu. A rationale from frequency perspective for grokking in training neural network. arXiv preprint arXiv:2405.17479, 2024. 7 \ No newline at end of file diff --git a/2025/Do We Always Need the Simplicity Bias_ Looking for Optimal Inductive Biases in the Wild/images.zip b/2025/Do We Always Need the Simplicity Bias_ Looking for Optimal Inductive Biases in the Wild/images.zip new file mode 100644 index 0000000000000000000000000000000000000000..cf900c4d52e9c39a76b8c85c093c8a76404ea8f7 --- /dev/null +++ b/2025/Do We Always Need the Simplicity Bias_ Looking for Optimal Inductive Biases in the Wild/images.zip @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:24e594882d11891eb81e204060fe296d1cb66c5a33d6c198f914666b9d1ad845 +size 373999 diff --git a/2025/Do We Always Need the Simplicity Bias_ Looking for Optimal Inductive Biases in the Wild/layout.json b/2025/Do We Always Need the Simplicity Bias_ Looking for Optimal Inductive Biases in the Wild/layout.json new file mode 100644 index 0000000000000000000000000000000000000000..304c54cc1a342aa788e33663b8b1d11f6c82d6e9 --- /dev/null +++ b/2025/Do We Always Need the Simplicity Bias_ Looking for Optimal Inductive Biases in the Wild/layout.json @@ -0,0 +1,12404 @@ +{ + "pdf_info": [ + { + "para_blocks": [ + { + "bbox": [ + 152, + 103, + 459, + 140 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 152, + 103, + 459, + 140 + ], + "spans": [ + { + "bbox": [ + 152, + 103, + 459, + 140 + ], + "type": "text", + "content": "Do We Always Need the Simplicity Bias? Looking for Optimal Inductive Biases in the Wild" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 80, + 162, + 197, + 190 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 80, + 162, + 197, + 190 + ], + "spans": [ + { + "bbox": [ + 80, + 162, + 197, + 190 + ], + "type": "text", + "content": "Damien Teney \nIdiap Research Institute" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 220, + 162, + 291, + 188 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 220, + 162, + 291, + 188 + ], + "spans": [ + { + "bbox": [ + 220, + 162, + 291, + 188 + ], + "type": "text", + "content": "Liangze Jiang EPFL" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 152, + 212, + 200, + 224 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 152, + 212, + 200, + 224 + ], + "spans": [ + { + "bbox": [ + 152, + 212, + 200, + 224 + ], + "type": "text", + "content": "Abstract" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 55, + 239, + 296, + 324 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 239, + 296, + 324 + ], + "spans": [ + { + "bbox": [ + 55, + 239, + 296, + 324 + ], + "type": "text", + "content": "Neural architectures tend to fit their data with relatively simple functions. This \"simplicity bias\" is widely regarded as key to their success. This paper explores the limits of this principle. Building on recent findings that the simplicity bias stems from ReLU activations [96], we introduce a method to meta-learn new activation functions and inductive biases better suited to specific tasks." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 55, + 327, + 296, + 448 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 327, + 296, + 448 + ], + "spans": [ + { + "bbox": [ + 55, + 327, + 296, + 448 + ], + "type": "text", + "content": "Findings. We identify multiple tasks where the simplicity bias is inadequate and ReLUs suboptimal. In these cases, we learn new activation functions that perform better by inducing a prior of higher complexity. Interestingly, these cases correspond to domains where neural networks have historically struggled: tabular data, regression tasks, cases of shortcut learning, and algorithmic grokking tasks. In comparison, the simplicity bias induced by ReLUs proves adequate on image tasks where the best learned activations are nearly identical to ReLUs and GeLUs." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 55, + 451, + 296, + 548 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 451, + 296, + 548 + ], + "spans": [ + { + "bbox": [ + 55, + 451, + 296, + 548 + ], + "type": "text", + "content": "Implications. Contrary to popular belief, the simplicity bias of ReLU networks is not universally useful. It is near-optimal for image classification, but other inductive biases are sometimes preferable. We showed that activation functions can control these inductive biases, but future tailored architectures might provide further benefits. Advances are still needed to characterize a model's inductive biases beyond \"complexity\", and their adequacy with the data." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 55, + 571, + 135, + 583 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 571, + 135, + 583 + ], + "spans": [ + { + "bbox": [ + 55, + 571, + 135, + 583 + ], + "type": "text", + "content": "1. Introduction" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 55, + 594, + 295, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 594, + 295, + 713 + ], + "spans": [ + { + "bbox": [ + 55, + 594, + 295, + 713 + ], + "type": "text", + "content": "When and why NNs generalize is yet to be understood. Neural networks (NNs) have proven more effective than other machine learning models. However, we still miss a complete explanation of their generalization abilities. A better understanding could help address failures from shortcut learning [29, 93] to distribution shifts [47, 95], biases, and spurious correlations in language models for example [28, 38, 87]. Understanding conditions for generalization would also enable the design of architectures and data preparation from first principles, rather than trial and error." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 314, + 162, + 391, + 188 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 314, + 162, + 391, + 188 + ], + "spans": [ + { + "bbox": [ + 314, + 162, + 391, + 188 + ], + "type": "text", + "content": "Florin Gogianu\nBitdefender" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 415, + 162, + 526, + 190 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 415, + 162, + 526, + 190 + ], + "spans": [ + { + "bbox": [ + 415, + 162, + 526, + 190 + ], + "type": "text", + "content": "Ehsan Abbasnejad \nUniversity of Adelaide" + } + ] + } + ], + "index": 12 + }, + { + "type": "image", + "bbox": [ + 315, + 211, + 553, + 299 + ], + "blocks": [ + { + "bbox": [ + 315, + 211, + 553, + 299 + ], + "lines": [ + { + "bbox": [ + 315, + 211, + 553, + 299 + ], + "spans": [ + { + "bbox": [ + 315, + 211, + 553, + 299 + ], + "type": "image", + "image_path": "f6c3c7fa1ae2b01883bb609bc39722c6311e201422ffb6aaffc6076662c9f41c.jpg" + } + ] + } + ], + "index": 13, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 313, + 310, + 555, + 366 + ], + "lines": [ + { + "bbox": [ + 313, + 310, + 555, + 366 + ], + "spans": [ + { + "bbox": [ + 313, + 310, + 555, + 366 + ], + "type": "text", + "content": "Figure 1. (1) We modulate the inductive bias of neural architectures by learning novel activation functions that improve generalization on specific datasets. (2) With this tool, we study the relation between model accuracy and complexity. We identify tasks where the simplicity bias of ReLU architectures is suboptimal." + } + ] + } + ], + "index": 14, + "angle": 0, + "type": "image_caption" + } + ], + "index": 13 + }, + { + "bbox": [ + 313, + 381, + 555, + 514 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 381, + 555, + 514 + ], + "spans": [ + { + "bbox": [ + 313, + 381, + 555, + 514 + ], + "type": "text", + "content": "This paper studies inductive biases i.e. the assumptions made by learning algorithms to generalize beyond training data [67].1 A vast literature examines the inductive biases of architectures [14], optimizers [69], losses [42], regularizers [50], etc. The simplicity bias is one aspect of the inductive biases of NNs that makes them fit their training data with simple2 functions [8, 71]. Despite wide belief that the simplicity bias could be due to SGD [7, 40, 60, 92], work on untrained networks showed that it can be explained with architectures alone [15, 31, 64, 99]. ReLUs also seem critical to induce the simplicity bias in typical architectures [96]." + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 313, + 521, + 556, + 666 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 521, + 556, + 666 + ], + "spans": [ + { + "bbox": [ + 313, + 521, + 556, + 666 + ], + "type": "text", + "content": "Limits of the simplicity bias. The simplicity bias is an intuitive explanation for the ability to generalize on real-world data. It embodies Occam's razor [66] and assumes that data-generating processes in the real world are simple. Additionally, a prior for simplicity is supported by results in algorithmic information theory [18] stating essentially that \"a bias in the distribution of target functions must be towards low complexity\". However, this only means that simplicity is a good prior on average, but not necessarily the best choice on any task or dataset. This matches the no-free lunch theorem [101] according to which no inductive bias is universally useful. Therefore, this paper asks the following." + } + ] + } + ], + "index": 16 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "spans": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "type": "text", + "content": "CVF" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 145, + 0, + 496, + 37 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 145, + 0, + 496, + 37 + ], + "spans": [ + { + "bbox": [ + 145, + 0, + 496, + 37 + ], + "type": "text", + "content": "This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 313, + 683, + 553, + 713 + ], + "type": "page_footnote", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 683, + 553, + 713 + ], + "spans": [ + { + "bbox": [ + 313, + 683, + 553, + 713 + ], + "type": "inline_equation", + "content": "^{1}" + }, + { + "bbox": [ + 313, + 683, + 553, + 713 + ], + "type": "text", + "content": "Inductive biases can formalized as a prior over the space of functions [65]. " + }, + { + "bbox": [ + 313, + 683, + 553, + 713 + ], + "type": "inline_equation", + "content": "^{2}" + }, + { + "bbox": [ + 313, + 683, + 553, + 713 + ], + "type": "text", + "content": "Simplicity can be formalized using Kolmogorov complexity or its approximations, e.g. frequency, compressibility, sensitivity, etc. [17, 18, 37, 96]" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 300, + 749, + 310, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 300, + 749, + 310, + 757 + ], + "spans": [ + { + "bbox": [ + 300, + 749, + 310, + 757 + ], + "type": "text", + "content": "79" + } + ] + } + ], + "index": 18 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 0 + }, + { + "para_blocks": [ + { + "bbox": [ + 63, + 74, + 288, + 110 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 63, + 74, + 288, + 110 + ], + "spans": [ + { + "bbox": [ + 63, + 74, + 288, + 110 + ], + "type": "text", + "content": "Are there practical applications of machine learning where the simplicity bias is detrimental? In these cases, what do the optimal inductive biases look like?" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 55, + 118, + 294, + 142 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 118, + 294, + 142 + ], + "spans": [ + { + "bbox": [ + 55, + 118, + 294, + 142 + ], + "type": "text", + "content": "For example, shortcut learning is one situation where the simplicity bias is already known to be detrimental [85, 93]." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 55, + 147, + 296, + 301 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 147, + 296, + 301 + ], + "spans": [ + { + "bbox": [ + 55, + 147, + 296, + 301 + ], + "type": "text", + "content": "Searching for optimal inductive biases by learning activation functions. Prior work [96] showed that ReLU activations are critical to obtain the simplicity bias in typical architectures. Hence we build a new tool to modulate the simplicity bias by learning dataset-specific activation functions. It uses bi-level optimization and a spline parametrization to learn activations free of any prior, such as constraints of smoothness or monotonicity (unlike prior work [2, 6, 82]). This (1) enables the discovery of entirely new activation functions and inductive biases that improve generalization (Figure 1) and (2) highlights the suboptimality of the simplicity bias by comparing the accuracy and complexity of models with ReLUs vs. learned activations." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 55, + 307, + 296, + 474 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 307, + 296, + 474 + ], + "spans": [ + { + "bbox": [ + 55, + 307, + 296, + 474 + ], + "type": "text", + "content": "Findings. We examine four domains that we hypothesized to be impaired by the simplicity bias: tabular data, regression tasks, cases of shortcut learning, and algorithmic tasks. Our intuition is that they require learning functions with high sensitivity or sharp transitions. For each domain, we collect existing datasets then train and analyze models without and with learned activation functions. In all cases, we obtain better generalization with dataset-specific activations, and the improvements are attributable to learning higher-complexity solutions. In comparison, this analysis also shows that classical image datasets (MNIST, CIFAR, FASHION-MNIST, SVHN) are extremely well suited to the inductive biases of ReLUs. The best learned activations are then strikingly similar to variants like GeLUs [39]." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 55, + 479, + 174, + 491 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 479, + 174, + 491 + ], + "spans": [ + { + "bbox": [ + 55, + 479, + 174, + 491 + ], + "type": "text", + "content": "Summary of contributions." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 55, + 494, + 296, + 713 + ], + "type": "list", + "angle": 0, + "index": 8, + "blocks": [ + { + "bbox": [ + 55, + 494, + 295, + 517 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 494, + 295, + 517 + ], + "spans": [ + { + "bbox": [ + 55, + 494, + 295, + 517 + ], + "type": "text", + "content": "- A new method to discover dataset-specific activation functions optimized for generalization." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 55, + 520, + 296, + 686 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 520, + 296, + 686 + ], + "spans": [ + { + "bbox": [ + 55, + 520, + 296, + 686 + ], + "type": "text", + "content": "- An examination of " + }, + { + "bbox": [ + 55, + 520, + 296, + 686 + ], + "type": "inline_equation", + "content": ">20" + }, + { + "bbox": [ + 55, + 520, + 296, + 686 + ], + "type": "text", + "content": " datasets showing that the simplicity bias of ReLU architectures can be suboptimal. (1) For regression tasks and tabular data, new learned activations greatly improve accuracy by helping learn complex functions. (2) For image classification, the process rediscovers smooth variants of ReLUs, suggesting a near-optimal choice for these popular tasks. (3) In cases of shortcut learning, we show that different learned activations can steer the learning towards different image features. (4) For grokking tasks, new learned activations can eliminate the phenomenon, supporting the explanation as a mismatch between data and architectures. We also measure a positive transferability of learned activations across related tasks." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 55, + 689, + 296, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 689, + 296, + 713 + ], + "spans": [ + { + "bbox": [ + 55, + 689, + 296, + 713 + ], + "type": "text", + "content": "- An analysis showing that improvements with learned activations correlate with the learning of complex functions." + } + ] + } + ], + "index": 7 + } + ], + "sub_type": "text" + }, + { + "bbox": [ + 313, + 72, + 554, + 168 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 72, + 554, + 168 + ], + "spans": [ + { + "bbox": [ + 313, + 72, + 554, + 168 + ], + "type": "text", + "content": "Implications. All cases where the simplicity bias proved suboptimal are in domains where NNs have historically struggled. We now connect them to a common explanation. This implies that architectures tailored to some specific domains may still have a place besides scaling up models and data. Conversely, the suitability of ReLUs to image classification suggests that researchers successfully converged by trial and error to designs well tuned to popular tasks." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 314, + 185, + 374, + 197 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 314, + 185, + 374, + 197 + ], + "spans": [ + { + "bbox": [ + 314, + 185, + 374, + 197 + ], + "type": "text", + "content": "2. Methods" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 313, + 205, + 554, + 229 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 205, + 554, + 229 + ], + "spans": [ + { + "bbox": [ + 313, + 205, + 554, + 229 + ], + "type": "text", + "content": "This section introduces tools to analyze trained models and to learn new dataset-specific activation functions." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 313, + 243, + 481, + 255 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 243, + 481, + 255 + ], + "spans": [ + { + "bbox": [ + 313, + 243, + 481, + 255 + ], + "type": "text", + "content": "2.1. Visualizing a Model's Function" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 313, + 259, + 555, + 464 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 259, + 555, + 464 + ], + "spans": [ + { + "bbox": [ + 313, + 259, + 555, + 464 + ], + "type": "text", + "content": "A neural network implements a function " + }, + { + "bbox": [ + 313, + 259, + 555, + 464 + ], + "type": "inline_equation", + "content": "f_{\\theta} : \\mathbb{R}^{d_{\\mathrm{in}}} \\rightarrow \\mathbb{R}^{d_{\\mathrm{out}}}" + }, + { + "bbox": [ + 313, + 259, + 555, + 464 + ], + "type": "text", + "content": " of parameters " + }, + { + "bbox": [ + 313, + 259, + 555, + 464 + ], + "type": "inline_equation", + "content": "\\theta" + }, + { + "bbox": [ + 313, + 259, + 555, + 464 + ], + "type": "text", + "content": " (weights and biases) that maps an input " + }, + { + "bbox": [ + 313, + 259, + 555, + 464 + ], + "type": "inline_equation", + "content": "x \\in \\mathbb{R}^{d_{\\mathrm{in}}}" + }, + { + "bbox": [ + 313, + 259, + 555, + 464 + ], + "type": "text", + "content": " to an output " + }, + { + "bbox": [ + 313, + 259, + 555, + 464 + ], + "type": "inline_equation", + "content": "y \\in \\mathbb{R}^{d_{\\mathrm{out}}}" + }, + { + "bbox": [ + 313, + 259, + 555, + 464 + ], + "type": "text", + "content": ". For a regression task, " + }, + { + "bbox": [ + 313, + 259, + 555, + 464 + ], + "type": "inline_equation", + "content": "y \\in \\mathbb{R}" + }, + { + "bbox": [ + 313, + 259, + 555, + 464 + ], + "type": "text", + "content": " is the predicted value. For a classification task, " + }, + { + "bbox": [ + 313, + 259, + 555, + 464 + ], + "type": "inline_equation", + "content": "y" + }, + { + "bbox": [ + 313, + 259, + 555, + 464 + ], + "type": "text", + "content": " is a vector of logits passed through a sigmoid or softmax to obtain class probabilities. Because " + }, + { + "bbox": [ + 313, + 259, + 555, + 464 + ], + "type": "inline_equation", + "content": "d_{\\mathrm{in}}" + }, + { + "bbox": [ + 313, + 259, + 555, + 464 + ], + "type": "text", + "content": " can be large, " + }, + { + "bbox": [ + 313, + 259, + 555, + 464 + ], + "type": "inline_equation", + "content": "f" + }, + { + "bbox": [ + 313, + 259, + 555, + 464 + ], + "type": "text", + "content": " can be difficult to visualize and analyze. A workaround is to examine " + }, + { + "bbox": [ + 313, + 259, + 555, + 464 + ], + "type": "inline_equation", + "content": "f" + }, + { + "bbox": [ + 313, + 259, + 555, + 464 + ], + "type": "text", + "content": " over 1D or 2D slices of the input space [26, 96]. To obtain a slice in a region of plausible data, we use the training data " + }, + { + "bbox": [ + 313, + 259, + 555, + 464 + ], + "type": "inline_equation", + "content": "\\mathcal{T}" + }, + { + "bbox": [ + 313, + 259, + 555, + 464 + ], + "type": "text", + "content": ". For a 1D slice (linear path), we sample " + }, + { + "bbox": [ + 313, + 259, + 555, + 464 + ], + "type": "inline_equation", + "content": "x_1, x_2 \\sim \\mathcal{T}" + }, + { + "bbox": [ + 313, + 259, + 555, + 464 + ], + "type": "text", + "content": " then define the path " + }, + { + "bbox": [ + 313, + 259, + 555, + 464 + ], + "type": "inline_equation", + "content": "\\mathbf{X}_{x_1, x_2} = [(1 - \\lambda) x_1 + \\lambda x_2, \\lambda \\in [0,1]]" + }, + { + "bbox": [ + 313, + 259, + 555, + 464 + ], + "type": "text", + "content": ". We proceed analogously with three points for a 2D slice. We sample " + }, + { + "bbox": [ + 313, + 259, + 555, + 464 + ], + "type": "inline_equation", + "content": "\\lambda" + }, + { + "bbox": [ + 313, + 259, + 555, + 464 + ], + "type": "text", + "content": " regularly in [0,1] such that " + }, + { + "bbox": [ + 313, + 259, + 555, + 464 + ], + "type": "inline_equation", + "content": "\\mathbf{X}" + }, + { + "bbox": [ + 313, + 259, + 555, + 464 + ], + "type": "text", + "content": " is a finite sequence of points. Then " + }, + { + "bbox": [ + 313, + 259, + 555, + 464 + ], + "type": "inline_equation", + "content": "f" + }, + { + "bbox": [ + 313, + 259, + 555, + 464 + ], + "type": "text", + "content": " is evaluated on these points to give a 1D sequence or 2D grid of values that are convenient to display and analyze (Figure 7c). When " + }, + { + "bbox": [ + 313, + 259, + 555, + 464 + ], + "type": "inline_equation", + "content": "d_{\\mathrm{out}} > 1" + }, + { + "bbox": [ + 313, + 259, + 555, + 464 + ], + "type": "text", + "content": " (multi-class task), we examine one random dimension of " + }, + { + "bbox": [ + 313, + 259, + 555, + 464 + ], + "type": "inline_equation", + "content": "f" + }, + { + "bbox": [ + 313, + 259, + 555, + 464 + ], + "type": "text", + "content": "'s output at a time." + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 313, + 477, + 492, + 490 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 477, + 492, + 490 + ], + "spans": [ + { + "bbox": [ + 313, + 477, + 492, + 490 + ], + "type": "text", + "content": "2.2. Measuring a Model's Complexity" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 313, + 494, + 554, + 555 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 494, + 554, + 555 + ], + "spans": [ + { + "bbox": [ + 313, + 494, + 554, + 555 + ], + "type": "text", + "content": "We wish to quantify the complexity of the function " + }, + { + "bbox": [ + 313, + 494, + 554, + 555 + ], + "type": "inline_equation", + "content": "f" + }, + { + "bbox": [ + 313, + 494, + 554, + 555 + ], + "type": "text", + "content": " implemented by a model trained on data " + }, + { + "bbox": [ + 313, + 494, + 554, + 555 + ], + "type": "inline_equation", + "content": "\\mathcal{T}" + }, + { + "bbox": [ + 313, + 494, + 554, + 555 + ], + "type": "text", + "content": ". Prior work used Fourier decompositions [26, 96] but this requires a delicate implementation. We found a reliable alternative with the total variation (TV) of " + }, + { + "bbox": [ + 313, + 494, + 554, + 555 + ], + "type": "inline_equation", + "content": "f" + }, + { + "bbox": [ + 313, + 494, + 554, + 555 + ], + "type": "text", + "content": " averaged over many 1D paths:" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 345, + 555, + 553, + 581 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 345, + 555, + 553, + 581 + ], + "spans": [ + { + "bbox": [ + 345, + 555, + 553, + 581 + ], + "type": "interline_equation", + "content": "\\operatorname {T V} (f, \\mathcal {T}) = \\mathbb {E} _ {\\boldsymbol {x} _ {1}, \\boldsymbol {x} _ {2}} \\sim \\mathcal {T} \\int_ {\\boldsymbol {x} _ {1}} ^ {\\boldsymbol {x} _ {2}} \\left| f ^ {\\prime} (\\boldsymbol {x}) \\right| d \\boldsymbol {x}. \\tag {1}", + "image_path": "8fd2e7fcf093e73e1a25c1727a1a14aa640f022cf2c379c0406d7beb307a5750.jpg" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 313, + 582, + 554, + 619 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 582, + 554, + 619 + ], + "spans": [ + { + "bbox": [ + 313, + 582, + 554, + 619 + ], + "type": "text", + "content": "with " + }, + { + "bbox": [ + 313, + 582, + 554, + 619 + ], + "type": "inline_equation", + "content": "f'" + }, + { + "bbox": [ + 313, + 582, + 554, + 619 + ], + "type": "text", + "content": " the first derivative. We estimate (1) using a path as defined in Section 2.1. We name the points in such a path " + }, + { + "bbox": [ + 313, + 582, + 554, + 619 + ], + "type": "inline_equation", + "content": "\\mathbf{X}_{\\boldsymbol{x}_a, \\boldsymbol{x}_z} := [\\boldsymbol{x}_a, \\boldsymbol{x}_b, \\boldsymbol{x}_c, \\dots, \\boldsymbol{x}_y, \\boldsymbol{x}_z]" + }, + { + "bbox": [ + 313, + 582, + 554, + 619 + ], + "type": "text", + "content": ". We then have:" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 324, + 623, + 553, + 661 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 324, + 623, + 553, + 661 + ], + "spans": [ + { + "bbox": [ + 324, + 623, + 553, + 661 + ], + "type": "interline_equation", + "content": "\\begin{array}{l} \\operatorname {T V} (f, \\mathcal {T}) \\approx \\mathbb {E} _ {\\boldsymbol {x} _ {a}, \\boldsymbol {x} _ {z}} \\sim_ {\\mathcal {T}} | f (\\boldsymbol {x} _ {b}) - f (\\boldsymbol {x} _ {a}) | \\\\ + \\left| f \\left(\\boldsymbol {x} _ {c}\\right) - f (\\boldsymbol {x} _ {b}) \\right| + \\dots \\tag {2} \\\\ + \\left| f \\left(\\boldsymbol {x} _ {z}\\right) - f \\left(\\boldsymbol {x} _ {y}\\right) \\right|. \\\\ \\end{array}", + "image_path": "393c6b507cd62b53c53aa757d8e330ee908fb53a6b11413bcc6693176051f336.jpg" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 313, + 665, + 554, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 665, + 554, + 713 + ], + "spans": [ + { + "bbox": [ + 313, + 665, + 554, + 713 + ], + "type": "text", + "content": "Appendix F shows that (2) correlates closely with a Fourier-based measure of complexity: the higher the TV, the higher the complexity. Yet, it is straightforward to implement and discriminative across small and large values." + } + ] + } + ], + "index": 19 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 299, + 749, + 310, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 299, + 749, + 310, + 757 + ], + "spans": [ + { + "bbox": [ + 299, + 749, + 310, + 757 + ], + "type": "text", + "content": "80" + } + ] + } + ], + "index": 20 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 1 + }, + { + "para_blocks": [ + { + "bbox": [ + 55, + 72, + 250, + 85 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 72, + 250, + 85 + ], + "spans": [ + { + "bbox": [ + 55, + 72, + 250, + 85 + ], + "type": "text", + "content": "2.3. Meta-Learning Activation Functions" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 55, + 89, + 296, + 209 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 89, + 296, + 209 + ], + "spans": [ + { + "bbox": [ + 55, + 89, + 296, + 209 + ], + "type": "text", + "content": "Our goal is to optimize the inductive biases of a neural network and recent work [96] showed that the activation functions are the most important component. The typical approach to learn activations [2, 5, 6, 11, 13, 22, 45, 82, 91] (see Related Work) replaces them with a small shared ReLU MLP that implements an " + }, + { + "bbox": [ + 55, + 89, + 296, + 209 + ], + "type": "inline_equation", + "content": "\\mathbb{R}\\rightarrow \\mathbb{R}" + }, + { + "bbox": [ + 55, + 89, + 296, + 209 + ], + "type": "text", + "content": " function. Its parameters are optimized along the network's. However this cannot discover truly novel activations because the embedded ReLU MLP has itself a simplicity bias and activations are optimized together with the model. We propose instead:" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 55, + 209, + 295, + 258 + ], + "type": "list", + "angle": 0, + "index": 5, + "blocks": [ + { + "bbox": [ + 55, + 209, + 295, + 221 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 209, + 295, + 221 + ], + "spans": [ + { + "bbox": [ + 55, + 209, + 295, + 221 + ], + "type": "text", + "content": "- an unbiased parametrization of the activations as splines," + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 56, + 222, + 276, + 233 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 222, + 276, + 233 + ], + "spans": [ + { + "bbox": [ + 56, + 222, + 276, + 233 + ], + "type": "text", + "content": "- a bi-level optimization to learn reusable activations," + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 56, + 235, + 295, + 258 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 235, + 295, + 258 + ], + "spans": [ + { + "bbox": [ + 56, + 235, + 295, + 258 + ], + "type": "text", + "content": "- an episodic training to optimize for generalization rather than simply to fit the training data." + } + ] + } + ], + "index": 4 + } + ], + "sub_type": "text" + }, + { + "bbox": [ + 55, + 261, + 296, + 380 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 261, + 296, + 380 + ], + "spans": [ + { + "bbox": [ + 55, + 261, + 296, + 380 + ], + "type": "text", + "content": "Parametrization as splines. We want a space of activation functions free of priors such as the smoothness and monotonicity enforced in prior work [5, 13]. We implement an activation " + }, + { + "bbox": [ + 55, + 261, + 296, + 380 + ], + "type": "inline_equation", + "content": "g_{\\psi} : \\mathbb{R} \\to \\mathbb{R}" + }, + { + "bbox": [ + 55, + 261, + 296, + 380 + ], + "type": "text", + "content": " as a linear spline with control points defined by " + }, + { + "bbox": [ + 55, + 261, + 296, + 380 + ], + "type": "inline_equation", + "content": "\\psi" + }, + { + "bbox": [ + 55, + 261, + 296, + 380 + ], + "type": "text", + "content": ". We define " + }, + { + "bbox": [ + 55, + 261, + 296, + 380 + ], + "type": "inline_equation", + "content": "n_c" + }, + { + "bbox": [ + 55, + 261, + 296, + 380 + ], + "type": "text", + "content": " points spread regularly in an interval " + }, + { + "bbox": [ + 55, + 261, + 296, + 380 + ], + "type": "inline_equation", + "content": "[a, b]" + }, + { + "bbox": [ + 55, + 261, + 296, + 380 + ], + "type": "text", + "content": ", typically " + }, + { + "bbox": [ + 55, + 261, + 296, + 380 + ], + "type": "inline_equation", + "content": "\\sim 50" + }, + { + "bbox": [ + 55, + 261, + 296, + 380 + ], + "type": "text", + "content": " points in " + }, + { + "bbox": [ + 55, + 261, + 296, + 380 + ], + "type": "inline_equation", + "content": "[-5, +5]" + }, + { + "bbox": [ + 55, + 261, + 296, + 380 + ], + "type": "text", + "content": ". Then " + }, + { + "bbox": [ + 55, + 261, + 296, + 380 + ], + "type": "inline_equation", + "content": "g" + }, + { + "bbox": [ + 55, + 261, + 296, + 380 + ], + "type": "text", + "content": " represents piecewise linear segments interpolating values specified in the learned parameters " + }, + { + "bbox": [ + 55, + 261, + 296, + 380 + ], + "type": "inline_equation", + "content": "\\psi := [g_{\\psi}(a), \\dots, g_{\\psi}(b)] \\in \\mathbb{R}^{n_c}" + }, + { + "bbox": [ + 55, + 261, + 296, + 380 + ], + "type": "text", + "content": ". " + }, + { + "bbox": [ + 55, + 261, + 296, + 380 + ], + "type": "inline_equation", + "content": "g" + }, + { + "bbox": [ + 55, + 261, + 296, + 380 + ], + "type": "text", + "content": " can represent simple and complex functions, including smooth curves, periodic functions, sharp transitions, etc." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 55, + 384, + 296, + 587 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 384, + 296, + 587 + ], + "spans": [ + { + "bbox": [ + 55, + 384, + 296, + 587 + ], + "type": "text", + "content": "Bi-level optimization & episodic training. Our goal is to get an activation function that can be reused like any other in subsequent training runs. This differs from prior work (e.g. [2]) that continuously updates the activation during training: the final one may not be suitable to start training with. Our solution is a bi-level meta-learning loop. An inner loop trains the model with a fixed activation function. An outer loop trains the activation function to maximize generalization. Each outer step simulates a new learning task or episode. This means (1) initializing the model with different weights and (2) using different subsets of data for training and validation. With suitable choices, this can simulate in- or out-of-distribution conditions (see Section 3.4). Without episodes, the learned activation could overfit to a particular model initialization for example, and would not generalize in subsequent training runs. The method is outlined as Algorithm 1. Its implementation is discussed in Appendix C." + } + ] + } + ], + "index": 7 + }, + { + "type": "image", + "bbox": [ + 63, + 596, + 86, + 616 + ], + "blocks": [ + { + "bbox": [ + 63, + 596, + 86, + 616 + ], + "lines": [ + { + "bbox": [ + 63, + 596, + 86, + 616 + ], + "spans": [ + { + "bbox": [ + 63, + 596, + 86, + 616 + ], + "type": "image", + "image_path": "ed53be628f428e1cb7466d73447f60e716bb2b6787d390e027740dddbbeb9e65.jpg" + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "image_body" + } + ], + "index": 8 + }, + { + "bbox": [ + 89, + 595, + 288, + 643 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 89, + 595, + 288, + 643 + ], + "spans": [ + { + "bbox": [ + 89, + 595, + 288, + 643 + ], + "type": "text", + "content": "Inductive bias and simplicity bias are not interchangeable. Our method optimizes toward bethepneralization. Simplicity is only one aspect of the 1d models that we analyze post-hoc (e.g. Figure 4)." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 55, + 657, + 162, + 670 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 657, + 162, + 670 + ], + "spans": [ + { + "bbox": [ + 55, + 657, + 162, + 670 + ], + "type": "text", + "content": "3. Tasks and Results" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 55, + 677, + 296, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 677, + 296, + 713 + ], + "spans": [ + { + "bbox": [ + 55, + 677, + 296, + 713 + ], + "type": "text", + "content": "We now examine tasks that we hypothesized to be ill-suited to the simplicity bias of ReLU architectures. The intuition is that the target function to learn (e.g. optimal classifier) con" + } + ] + } + ], + "index": 11 + }, + { + "type": "code", + "bbox": [ + 316, + 87, + 545, + 287 + ], + "blocks": [ + { + "bbox": [ + 315, + 72, + 541, + 84 + ], + "lines": [ + { + "bbox": [ + 315, + 72, + 541, + 84 + ], + "spans": [ + { + "bbox": [ + 315, + 72, + 541, + 84 + ], + "type": "text", + "content": "Algorithm 1 Meta-learning an activation function (AF)." + } + ] + } + ], + "index": 12, + "angle": 0, + "type": "code_caption" + }, + { + "bbox": [ + 316, + 87, + 545, + 287 + ], + "lines": [ + { + "bbox": [ + 316, + 87, + 545, + 287 + ], + "spans": [ + { + "bbox": [ + 316, + 87, + 545, + 287 + ], + "type": "text", + "content": "Input: training data " + }, + { + "bbox": [ + 316, + 87, + 545, + 287 + ], + "type": "inline_equation", + "content": "\\mathcal{T}" + }, + { + "bbox": [ + 316, + 87, + 545, + 287 + ], + "type": "text", + "content": " ; untrained neural model " + }, + { + "bbox": [ + 316, + 87, + 545, + 287 + ], + "type": "inline_equation", + "content": "f_{\\theta ,\\psi}" + }, + { + "bbox": [ + 316, + 87, + 545, + 287 + ], + "type": "text", + "content": " Initialize " + }, + { + "bbox": [ + 316, + 87, + 545, + 287 + ], + "type": "inline_equation", + "content": "\\psi" + }, + { + "bbox": [ + 316, + 87, + 545, + 287 + ], + "type": "text", + "content": " with zeros Parametrization of AF " + }, + { + "bbox": [ + 316, + 87, + 545, + 287 + ], + "type": "inline_equation", + "content": "n_{\\mathrm{tr}}\\gets 0" + }, + { + "bbox": [ + 316, + 87, + 545, + 287 + ], + "type": "text", + "content": " Number of inner-loop iterations while " + }, + { + "bbox": [ + 316, + 87, + 545, + 287 + ], + "type": "inline_equation", + "content": "n_{\\mathrm{tr}} < n_{\\mathrm{tr}}^{\\mathrm{max}}" + }, + { + "bbox": [ + 316, + 87, + 545, + 287 + ], + "type": "text", + "content": " Outer loop: train AF Increment " + }, + { + "bbox": [ + 316, + 87, + 545, + 287 + ], + "type": "inline_equation", + "content": "n_{\\mathrm{tr}}" + }, + { + "bbox": [ + 316, + 87, + 545, + 287 + ], + "type": "text", + "content": " Sample the episode's tr. " + }, + { + "bbox": [ + 316, + 87, + 545, + 287 + ], + "type": "inline_equation", + "content": "(T^{\\prime})" + }, + { + "bbox": [ + 316, + 87, + 545, + 287 + ], + "type": "text", + "content": " and val. " + }, + { + "bbox": [ + 316, + 87, + 545, + 287 + ], + "type": "inline_equation", + "content": "(\\mathcal{V})" + }, + { + "bbox": [ + 316, + 87, + 545, + 287 + ], + "type": "text", + "content": " sets from T Initialize " + }, + { + "bbox": [ + 316, + 87, + 545, + 287 + ], + "type": "inline_equation", + "content": "\\theta" + }, + { + "bbox": [ + 316, + 87, + 545, + 287 + ], + "type": "text", + "content": " randomly Model weights and biases for " + }, + { + "bbox": [ + 316, + 87, + 545, + 287 + ], + "type": "inline_equation", + "content": "n_{\\mathrm{tr}}" + }, + { + "bbox": [ + 316, + 87, + 545, + 287 + ], + "type": "text", + "content": " steps Inner loop: train model with fixed AF Eval. loss on " + }, + { + "bbox": [ + 316, + 87, + 545, + 287 + ], + "type": "inline_equation", + "content": "\\mathcal{T}':L\\gets \\Sigma_{(\\pmb {x},\\pmb {y})\\in \\mathcal{T}'}\\mathcal{L}\\big(f_{\\pmb {\\theta},\\pmb {\\psi}}(\\pmb {x},\\pmb {y})\\big)" + }, + { + "bbox": [ + 316, + 87, + 545, + 287 + ], + "type": "text", + "content": " Gradient step on weights/biases: " + }, + { + "bbox": [ + 316, + 87, + 545, + 287 + ], + "type": "inline_equation", + "content": "\\pmb {\\theta}\\gets \\mathrm{GD}(\\pmb {\\theta},\\nabla_{\\pmb{\\theta}}L)" + }, + { + "bbox": [ + 316, + 87, + 545, + 287 + ], + "type": "text", + "content": " Eval. loss on V: " + }, + { + "bbox": [ + 316, + 87, + 545, + 287 + ], + "type": "inline_equation", + "content": "L\\gets \\Sigma_{(\\pmb {x},\\pmb {y})\\in \\mathcal{V}}\\mathcal{L}\\big(f_{\\pmb {\\theta},\\pmb {\\psi}}(\\pmb {x},\\pmb {y})\\big)" + }, + { + "bbox": [ + 316, + 87, + 545, + 287 + ], + "type": "text", + "content": " Gradient step on AF: " + }, + { + "bbox": [ + 316, + 87, + 545, + 287 + ], + "type": "inline_equation", + "content": "\\psi \\gets \\mathrm{GD}(\\psi ,\\nabla_{\\psi}L)" + }, + { + "bbox": [ + 316, + 87, + 545, + 287 + ], + "type": "text", + "content": " if performance on " + }, + { + "bbox": [ + 316, + 87, + 545, + 287 + ], + "type": "inline_equation", + "content": "\\nu" + }, + { + "bbox": [ + 316, + 87, + 545, + 287 + ], + "type": "text", + "content": " worsens then break Early stopping" + } + ] + } + ], + "index": 13, + "angle": 0, + "type": "code_body" + } + ], + "index": 13, + "sub_type": "algorithm" + }, + { + "bbox": [ + 318, + 289, + 424, + 302 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 318, + 289, + 424, + 302 + ], + "spans": [ + { + "bbox": [ + 318, + 289, + 424, + 302 + ], + "type": "text", + "content": "Output: optimized AF " + }, + { + "bbox": [ + 318, + 289, + 424, + 302 + ], + "type": "inline_equation", + "content": "\\psi" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 313, + 315, + 555, + 399 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 315, + 555, + 399 + ], + "spans": [ + { + "bbox": [ + 313, + 315, + 555, + 399 + ], + "type": "text", + "content": "tains sharp transitions (regression tasks, tabular datasets), or repeating patterns (algorithmic tasks) that contradict the ReLUs' simplicity bias. For each task, we examine existing datasets with the tools from Section 2. In all cases, we find benefits from architectures whose inductive biases favor more complex functions. Additional details and results are provided in Appendix E." + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 313, + 405, + 462, + 418 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 405, + 462, + 418 + ], + "spans": [ + { + "bbox": [ + 313, + 405, + 462, + 418 + ], + "type": "text", + "content": "3.1. Image Classification Tasks" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 313, + 422, + 555, + 495 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 422, + 555, + 495 + ], + "spans": [ + { + "bbox": [ + 313, + 422, + 555, + 495 + ], + "type": "text", + "content": "Background. We start with classical datasets to validate our methodology: MNIST, FASHION-MNIST, SVHN, CIFAR-10 [49, 52, 68, 102]. They are representative of the vision tasks that guided the development of deep learning. Our hypothesis is therefore that the inductive biases of modern architectures and ReLUs are well suited to these datasets." + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 313, + 498, + 556, + 641 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 498, + 556, + 641 + ], + "spans": [ + { + "bbox": [ + 313, + 498, + 556, + 641 + ], + "type": "text", + "content": "Setup. For each dataset, we learn activation functions with Algorithm 1. We experiment with two initializations of the spline parameters: as zeros and so as to mimic a ReLU. The goal of the latter is to explore the space of functions similar to ReLUs. Because of the difficulty of the optimization, the algorithm is likely to converge to a local optimum similar to ReLUs if there is one. We also experiment with the sharing of the activation function. By default, a single function is shared across the network. Alternatively, we learn a different activation function per layer. This provides more ways to affect the model's inductive biases. Our base architecture is a 3-layer MLP (details in Appendix E)." + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 313, + 644, + 555, + 703 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 644, + 555, + 703 + ], + "spans": [ + { + "bbox": [ + 313, + 644, + 555, + 703 + ], + "type": "text", + "content": "Results. We compare in Figure 2a the accuracy of models with ReLUs vs. learned activation functions. Differences are small. The learned activations only improve slightly on SVHN and CIFAR. This suggests that the inductive biases of ReLUs are generally well suited to these datasets." + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 313, + 704, + 554, + 716 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 704, + 554, + 716 + ], + "spans": [ + { + "bbox": [ + 313, + 704, + 554, + 716 + ], + "type": "text", + "content": "We examine the learned activations in Figure 3a. With" + } + ] + } + ], + "index": 20 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 299, + 749, + 310, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 299, + 749, + 310, + 757 + ], + "spans": [ + { + "bbox": [ + 299, + 749, + 310, + 757 + ], + "type": "text", + "content": "81" + } + ] + } + ], + "index": 21 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 2 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 58, + 62, + 140, + 144 + ], + "blocks": [ + { + "bbox": [ + 58, + 62, + 140, + 144 + ], + "lines": [ + { + "bbox": [ + 58, + 62, + 140, + 144 + ], + "spans": [ + { + "bbox": [ + 58, + 62, + 140, + 144 + ], + "type": "image", + "image_path": "211b35dc0b91d3f90a9b9e2fd0c6c2905f10ef07f027f107e2be95efab6b54df.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 55, + 148, + 295, + 213 + ], + "lines": [ + { + "bbox": [ + 55, + 148, + 295, + 213 + ], + "spans": [ + { + "bbox": [ + 55, + 148, + 295, + 213 + ], + "type": "text", + "content": "Figure 2. Test accuracy on image datasets. (a) For classification tasks, all models perform similarly, suggesting that the inductive biases of ReLUs are well suited to these datasets. (b) For regression tasks, models with learned activations perform better, especially from an initialization as zeros, which enables the discovery of completely novel activation functions." + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "type": "image", + "bbox": [ + 141, + 62, + 220, + 144 + ], + "blocks": [ + { + "bbox": [ + 141, + 62, + 220, + 144 + ], + "lines": [ + { + "bbox": [ + 141, + 62, + 220, + 144 + ], + "spans": [ + { + "bbox": [ + 141, + 62, + 220, + 144 + ], + "type": "image", + "image_path": "a76d47c4c1ce9be58cd6a561fbc4f9a6eb005f7a13727ef07809908712d3a81f.jpg" + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 221, + 86, + 299, + 111 + ], + "lines": [ + { + "bbox": [ + 221, + 86, + 299, + 111 + ], + "spans": [ + { + "bbox": [ + 221, + 86, + 299, + 111 + ], + "type": "text", + "content": "- ReLU activations\n- Learned act., ReLU init.\n- Learned act., zero init." + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_footnote" + } + ], + "index": 1 + }, + { + "type": "image", + "bbox": [ + 56, + 219, + 286, + 357 + ], + "blocks": [ + { + "bbox": [ + 56, + 219, + 286, + 357 + ], + "lines": [ + { + "bbox": [ + 56, + 219, + 286, + 357 + ], + "spans": [ + { + "bbox": [ + 56, + 219, + 286, + 357 + ], + "type": "image", + "image_path": "6471036dbc2489f41e7f191e6fa9b8105b97989fcf3ba5b13b887c3690c24440.jpg" + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 55, + 361, + 295, + 416 + ], + "lines": [ + { + "bbox": [ + 55, + 361, + 295, + 416 + ], + "spans": [ + { + "bbox": [ + 55, + 361, + 295, + 416 + ], + "type": "text", + "content": "Figure 3. Activation functions learned for MNIST. For a classification task, the activation learned from a ReLU resembles the popular GeLUs. For a regression task, the learned activations contain irregularities that help a network represent complex functions. See Figure 16 for similar results on other datasets." + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "image_caption" + } + ], + "index": 4 + }, + { + "bbox": [ + 55, + 429, + 296, + 514 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 429, + 296, + 514 + ], + "spans": [ + { + "bbox": [ + 55, + 429, + 296, + 514 + ], + "type": "text", + "content": "an initialization as ReLUs, the optimization converges to a smooth variant remarkably similar to GeLUs [11] which are widely used. This suggests that the research community has empirically converged on a local optimum in the space of activation functions. With an initialization as zeros, we discover wavelets [80] that are unlike common activations but perform as well as ReLUs, i.e. another local optimum." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 62, + 522, + 289, + 583 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 522, + 289, + 583 + ], + "spans": [ + { + "bbox": [ + 62, + 522, + 289, + 583 + ], + "type": "text", + "content": "Take-away: for image classification, learned activations provide very little benefit over ReLUs. Smooth variants of ReLUs are a local optimum in the space of activations. ReLUs' popularity for such tasks could thus be explained with their proximity to this optimum." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 55, + 594, + 159, + 607 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 594, + 159, + 607 + ], + "spans": [ + { + "bbox": [ + 55, + 594, + 159, + 607 + ], + "type": "text", + "content": "3.2. Regression Tasks" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 55, + 614, + 295, + 698 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 614, + 295, + 698 + ], + "spans": [ + { + "bbox": [ + 55, + 614, + 295, + 698 + ], + "type": "text", + "content": "Background. Regression tasks are known to be difficult for NNs [90]. They are often turned into a classification through discretization [24, 43]. Existing explanations that invoke implicit biases of gradient descent are clearly incomplete [90]. Our hypothesis is that regression is difficult because it often involves irregular decision boundaries [35] in opposition to the typical solutions of ReLU networks [17]." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 55, + 701, + 295, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 701, + 295, + 713 + ], + "spans": [ + { + "bbox": [ + 55, + 701, + 295, + 713 + ], + "type": "text", + "content": "Setup. We use the same setup and image datasets as Sec-" + } + ] + } + ], + "index": 10 + }, + { + "type": "image", + "bbox": [ + 342, + 62, + 432, + 152 + ], + "blocks": [ + { + "bbox": [ + 342, + 62, + 432, + 152 + ], + "lines": [ + { + "bbox": [ + 342, + 62, + 432, + 152 + ], + "spans": [ + { + "bbox": [ + 342, + 62, + 432, + 152 + ], + "type": "image", + "image_path": "2413db2370a191b22bbcfc9433a64c78f101c9d5ec538770208e7115a8f0bbbc.jpg" + } + ] + } + ], + "index": 11, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 313, + 154, + 555, + 275 + ], + "lines": [ + { + "bbox": [ + 313, + 154, + 555, + 275 + ], + "spans": [ + { + "bbox": [ + 313, + 154, + 555, + 275 + ], + "type": "text", + "content": "Figure 4. Accuracy vs. complexity on image datasets. Each marker is a model with different hyperparameters and ReLUs " + }, + { + "bbox": [ + 313, + 154, + 555, + 275 + ], + "type": "inline_equation", + "content": "(\\bullet)" + }, + { + "bbox": [ + 313, + 154, + 555, + 275 + ], + "type": "text", + "content": " or learned activations initialized as ReLUs " + }, + { + "bbox": [ + 313, + 154, + 555, + 275 + ], + "type": "inline_equation", + "content": "(\\square)" + }, + { + "bbox": [ + 313, + 154, + 555, + 275 + ], + "type": "text", + "content": " or as zeros " + }, + { + "bbox": [ + 313, + 154, + 555, + 275 + ], + "type": "inline_equation", + "content": "(\\text{圆})" + }, + { + "bbox": [ + 313, + 154, + 555, + 275 + ], + "type": "text", + "content": ". For classification (a), ReLUs are close to best. Activations optimized from ReLUs only improve the accuracy slightly, corresponding to the GeLU-like function in Figure 3. For regression (b), new activations (learned from zeros) are best. Moreover, accuracy and complexity are clearly correlated only for regression. This supports the hypothesis that regression is more complex than classification and thus benefits from alternatives to the ReLUs' simplicity bias. See Figure 18 for similar results on other datasets." + } + ] + } + ], + "index": 13, + "angle": 0, + "type": "image_caption" + } + ], + "index": 11 + }, + { + "type": "image", + "bbox": [ + 443, + 62, + 526, + 152 + ], + "blocks": [ + { + "bbox": [ + 443, + 62, + 526, + 152 + ], + "lines": [ + { + "bbox": [ + 443, + 62, + 526, + 152 + ], + "spans": [ + { + "bbox": [ + 443, + 62, + 526, + 152 + ], + "type": "image", + "image_path": "022bc003f6aa1b5ffb090bf5ca3e5cc871e8994b0ce0ea05eaeca31b91ca2e79.jpg" + } + ] + } + ], + "index": 12, + "angle": 0, + "type": "image_body" + } + ], + "index": 12 + }, + { + "bbox": [ + 313, + 289, + 554, + 337 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 289, + 554, + 337 + ], + "spans": [ + { + "bbox": [ + 313, + 289, + 554, + 337 + ], + "type": "text", + "content": "tion 3.1. The task is now to directly predict class IDs. E.g. for MNIST this means predicting digit values. Models are trained with an MSE loss. To measure accuracy, we discretize the predictions to the nearest class ID." + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 313, + 345, + 554, + 429 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 345, + 554, + 429 + ], + "spans": [ + { + "bbox": [ + 313, + 345, + 554, + 429 + ], + "type": "text", + "content": "Results. The first observation from Figure 2b is that regression is clearly more difficult for NNs than classification (lower accuracies) despite the identical underlying task. Importantly, the learned activations now provide clear improvements, especially when learned from scratch (initialization as zeros). This confirms the hypothesis that the inductive biases of RLUs are not well suited to these tasks." + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 313, + 436, + 554, + 556 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 436, + 554, + 556 + ], + "spans": [ + { + "bbox": [ + 313, + 436, + 554, + 556 + ], + "type": "text", + "content": "Figure 3b shows that the learned activations contain more irregularities for regression than classification. Prior work [96] showed that this can help models represent complex functions with sharp transitions. An analysis of the complexity of trained models (Figures 4 and 18) shows that the accuracy is correlated with complexity for regression but not classification. And regression models with learned activations implement functions of higher complexity than with ReLUs. This supports the claim that the improvements arise from overcoming the simplicity bias of ReLUs." + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 313, + 564, + 554, + 649 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 564, + 554, + 649 + ], + "spans": [ + { + "bbox": [ + 313, + 564, + 554, + 649 + ], + "type": "text", + "content": "Complexity is only one dimension of the inductive biases. The complexity plots for SVHN (Figure 18) interestingly show that models with ReLUs and learned activations get different accuracies at the same complexity level. This shows that our meta learning approach can search over dimensions of the inductive biases that are not captured by our complexity measure, and are yet to be explicitly studied." + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 320, + 658, + 547, + 706 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 320, + 658, + 547, + 706 + ], + "spans": [ + { + "bbox": [ + 320, + 658, + 547, + 706 + ], + "type": "text", + "content": "Take-away: regression is more difficult for NNs than classification, and the simplicity bias of ReLUs is partly to blame. Learned activations improve performance by helping networks represent more complex functions." + } + ] + } + ], + "index": 18 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 299, + 749, + 310, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 299, + 749, + 310, + 757 + ], + "spans": [ + { + "bbox": [ + 299, + 749, + 310, + 757 + ], + "type": "text", + "content": "82" + } + ] + } + ], + "index": 19 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 3 + }, + { + "para_blocks": [ + { + "bbox": [ + 56, + 72, + 141, + 83 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 72, + 141, + 83 + ], + "spans": [ + { + "bbox": [ + 56, + 72, + 141, + 83 + ], + "type": "text", + "content": "3.3. Tabular Data" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 55, + 91, + 296, + 247 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 91, + 296, + 247 + ], + "spans": [ + { + "bbox": [ + 55, + 91, + 296, + 247 + ], + "type": "text", + "content": "Background. Tabular data is any data with few unstructured dimensions, which often contains low-cardinality variables such as dates or categorical attributes. This contrasts e.g. with images, which contain many correlated, continuous dimensions (pixels). NNs struggle with tabular datasets and are often inferior to decision trees [35, 63]. Our hypothesis is that the inductive biases of standard architectures are ill-suited to such data because of the simplicity bias. It makes it difficult to learn functions where small changes in the input (e.g. day of the week) correspond to abrupt changes in the target — the definition of sensitivity, a proxy for complexity [17]. This seldom occurs in vision where similar images correspond to similar labels." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 55, + 250, + 296, + 346 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 250, + 296, + 346 + ], + "spans": [ + { + "bbox": [ + 55, + 250, + 296, + 346 + ], + "type": "text", + "content": "Setup. We use 16 real-world classification datasets from Grinsztajn et al. [34, 35]. Bases include a linear classifier, k-NNs, and boosted decision trees. Our models are MLPs with 1-4 hidden layers (details in Appendix E.4). We compare learned activations functions with ReLUs and TanHs with a global prefactor, " + }, + { + "bbox": [ + 55, + 250, + 296, + 346 + ], + "type": "inline_equation", + "content": "\\tanh (\\alpha x)" + }, + { + "bbox": [ + 55, + 250, + 296, + 346 + ], + "type": "text", + "content": " with " + }, + { + "bbox": [ + 55, + 250, + 296, + 346 + ], + "type": "inline_equation", + "content": "\\alpha \\in \\mathbb{R}^{+}" + }, + { + "bbox": [ + 55, + 250, + 296, + 346 + ], + "type": "text", + "content": " tuned on the validation set. This is a simple option with tunable complexity, albeit with inductive biases of TanHs [45, 96]." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 55, + 347, + 296, + 467 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 347, + 296, + 467 + ], + "spans": [ + { + "bbox": [ + 55, + 347, + 296, + 467 + ], + "type": "text", + "content": "We also experiment with learned input activation functions (IAFs). The motivation is to learn a different behavior for each input dimension. Since they carry different information, e.g. continuous vs. categorical variables, one could be suited to the simplicity bias while another is not, for example. IAFs are dimension-specific activation functions applied directly on the data before a standard MLP. IAFs are learned like AFs, from an initialization as the identity i.e. no effect by default. They subsume the gated inputs, and Fourier/numerical embeddings from prior work [20, 25, 32]." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 55, + 470, + 296, + 542 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 470, + 296, + 542 + ], + "spans": [ + { + "bbox": [ + 55, + 470, + 296, + 542 + ], + "type": "text", + "content": "Results. We compare the accuracy of models on the 16 datasets in Figures 5 and 20. Vanilla MLPs generally perform worse than trees. But adjusting the MLPs' inductive biases with learned prefactors or activations eliminates the gap. IAFs perform best, sometimes even surpassing trees. We analyze below the reasons for these improvements." + } + ] + } + ], + "index": 4 + }, + { + "type": "image", + "bbox": [ + 72, + 548, + 276, + 654 + ], + "blocks": [ + { + "bbox": [ + 72, + 548, + 276, + 654 + ], + "lines": [ + { + "bbox": [ + 72, + 548, + 276, + 654 + ], + "spans": [ + { + "bbox": [ + 72, + 548, + 276, + 654 + ], + "type": "image", + "image_path": "0ab3f5a495bc33f3325173094c344ecd7742e560a8f61f327615ce132c794e8f.jpg" + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 55, + 655, + 296, + 712 + ], + "lines": [ + { + "bbox": [ + 55, + 655, + 296, + 712 + ], + "spans": [ + { + "bbox": [ + 55, + 655, + 296, + 712 + ], + "type": "text", + "content": "Figure 5. Comparison of model types over 16 tabular datasets. Vanilla MLPs often perform worse than decision trees, but adjusting their inductive biases with learned activation functions (AFs) eliminates this gap. The input activation functions (IAFs) enable even better performance. See Figure 20 for results per dataset." + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "image_caption" + } + ], + "index": 5 + }, + { + "type": "image", + "bbox": [ + 316, + 66, + 375, + 133 + ], + "blocks": [ + { + "bbox": [ + 316, + 66, + 375, + 133 + ], + "lines": [ + { + "bbox": [ + 316, + 66, + 375, + 133 + ], + "spans": [ + { + "bbox": [ + 316, + 66, + 375, + 133 + ], + "type": "image", + "image_path": "dfa624194d959894bb7c0fdccd577ee48cf3651ba0a3e12c86c4ffafb2bd850f.jpg" + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 323, + 137, + 374, + 144 + ], + "lines": [ + { + "bbox": [ + 323, + 137, + 374, + 144 + ], + "spans": [ + { + "bbox": [ + 323, + 137, + 374, + 144 + ], + "type": "text", + "content": "Best at low complexity" + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "image_caption" + } + ], + "index": 7 + }, + { + "type": "image", + "bbox": [ + 377, + 67, + 433, + 133 + ], + "blocks": [ + { + "bbox": [ + 377, + 67, + 433, + 133 + ], + "lines": [ + { + "bbox": [ + 377, + 67, + 433, + 133 + ], + "spans": [ + { + "bbox": [ + 377, + 67, + 433, + 133 + ], + "type": "image", + "image_path": "d792fd28b74c36d0e1e7b3f038deb73197872e912e2fbf15d0920cf1edfa28ab.jpg" + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 313, + 147, + 554, + 236 + ], + "lines": [ + { + "bbox": [ + 313, + 147, + 554, + 236 + ], + "spans": [ + { + "bbox": [ + 313, + 147, + 554, + 236 + ], + "type": "text", + "content": "Figure 6. Test accuracy vs. complexity on tabular datasets. Each marker represents a model with different hyperparameters, and ReLUs " + }, + { + "bbox": [ + 313, + 147, + 554, + 236 + ], + "type": "inline_equation", + "content": "(\\bullet)" + }, + { + "bbox": [ + 313, + 147, + 554, + 236 + ], + "type": "text", + "content": " or learned activations initialized as ReLUs " + }, + { + "bbox": [ + 313, + 147, + 554, + 236 + ], + "type": "inline_equation", + "content": "(\\square)" + }, + { + "bbox": [ + 313, + 147, + 554, + 236 + ], + "type": "text", + "content": " or as zeros " + }, + { + "bbox": [ + 313, + 147, + 554, + 236 + ], + "type": "inline_equation", + "content": "(\\square)" + }, + { + "bbox": [ + 313, + 147, + 554, + 236 + ], + "type": "text", + "content": ". The learned activations perform better in all cases, but the accuracy peaks at different complexity levels. For some datasets, a low complexity is best and ReLUs thus perform quite well (left-most panel, note the smaller Y scale). For other datasets, the opposite is true and the improvements with learned activations is larger." + } + ] + } + ], + "index": 15, + "angle": 0, + "type": "image_caption" + } + ], + "index": 9 + }, + { + "type": "image", + "bbox": [ + 437, + 67, + 492, + 133 + ], + "blocks": [ + { + "bbox": [ + 379, + 137, + 454, + 143 + ], + "lines": [ + { + "bbox": [ + 379, + 137, + 454, + 143 + ], + "spans": [ + { + "bbox": [ + 379, + 137, + 454, + 143 + ], + "type": "text", + "content": "Different datasets" + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 437, + 67, + 492, + 133 + ], + "lines": [ + { + "bbox": [ + 437, + 67, + 492, + 133 + ], + "spans": [ + { + "bbox": [ + 437, + 67, + 492, + 133 + ], + "type": "image", + "image_path": "d04604de4f41000a298bdba6678ce37e51b87979e70ffb5ddead231238869014.jpg" + } + ] + } + ], + "index": 11, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 434, + 137, + 455, + 143 + ], + "lines": [ + { + "bbox": [ + 434, + 137, + 455, + 143 + ], + "spans": [ + { + "bbox": [ + 434, + 137, + 455, + 143 + ], + "type": "text", + "content": "at datasets" + } + ] + } + ], + "index": 12, + "angle": 0, + "type": "image_caption" + } + ], + "index": 11 + }, + { + "type": "image", + "bbox": [ + 495, + 67, + 550, + 133 + ], + "blocks": [ + { + "bbox": [ + 495, + 67, + 550, + 133 + ], + "lines": [ + { + "bbox": [ + 495, + 67, + 550, + 133 + ], + "spans": [ + { + "bbox": [ + 495, + 67, + 550, + 133 + ], + "type": "image", + "image_path": "7c8516fa48acd91e8d3a370fe2998112d4c47833fd599945361ea07792df2ef4.jpg" + } + ] + } + ], + "index": 13, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 495, + 137, + 552, + 144 + ], + "lines": [ + { + "bbox": [ + 495, + 137, + 552, + 144 + ], + "spans": [ + { + "bbox": [ + 495, + 137, + 552, + 144 + ], + "type": "text", + "content": "Best at higher complexity " + }, + { + "bbox": [ + 495, + 137, + 552, + 144 + ], + "type": "inline_equation", + "content": "\\uparrow" + } + ] + } + ], + "index": 14, + "angle": 0, + "type": "image_caption" + } + ], + "index": 13 + }, + { + "bbox": [ + 313, + 250, + 554, + 406 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 250, + 554, + 406 + ], + "spans": [ + { + "bbox": [ + 313, + 250, + 554, + 406 + ], + "type": "text", + "content": "Learned activation functions close the gap to decision trees by mimicking their inductive bias. We visualize in Figure 7c the functions implemented by different models, plotting their output over slices of the input space (Section 2.1). ReLUs produce the smoothest function while TanHs and learned activations induce sharper patterns. Notably, the IAFs induce sharp axis-aligned decision boundaries that are also characteristic of trees, with which they share a high accuracy. Axis-aligned transitions are the consequence of IAFs applied independently to each dimension. Sharp transitions originate from the complex shape of the learned activation function (Figure 7a) which is possible thanks to the unbiased spline parametrization (Section 2.3)." + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 313, + 408, + 554, + 624 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 408, + 554, + 624 + ], + "spans": [ + { + "bbox": [ + 313, + 408, + 554, + 624 + ], + "type": "text", + "content": "The simplicity/complexity bias is a property of the architecture. We visualize complexity landscapes of MLPs in Figure 7b. Similarly to standard loss landscapes [54], we plot model complexity over 2D slices of the parameter space. A first global view over a plane aligned with the training trajectory shows that complexity steadily increases during training for all models [48, 75, 104] but does so to the highest level for the best model (IAFs). A second view zooms in on each optimized solution in a random 2D plane. This examines the effect of arbitrary perturbations to the parameters. It shows that the ambient complexity of perturbed solutions of the best model is much higher than the solution itself, and then with less accurate models. This means that this architecture is more likely to represent complex functions because they are more abundant in parameter space [65, 84]. This is why the simplicity bias can be overcome: it results from architecture choices and not from an inevitable \"implicit bias\" of SGD [85, 89, 98, 105]." + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 313, + 626, + 553, + 687 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 626, + 553, + 687 + ], + "spans": [ + { + "bbox": [ + 313, + 626, + 553, + 687 + ], + "type": "text", + "content": "Different tabular datasets require different inductive biases. We examine the relation between accuracy and complexity in Figure 6. The accuracy peaks at different complexity levels for different datasets. For some, a low complexity is best and ReLU MLPs perform well. For oth" + } + ] + } + ], + "index": 18 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 313, + 693, + 553, + 713 + ], + "type": "page_footnote", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 693, + 553, + 713 + ], + "spans": [ + { + "bbox": [ + 313, + 693, + 553, + 713 + ], + "type": "text", + "content": "This resembles an analysis of untrained models [15, 64, 96, 99] but focuses on relevant regions on the parameter space, near optimized models." + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 299, + 749, + 310, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 299, + 749, + 310, + 757 + ], + "spans": [ + { + "bbox": [ + 299, + 749, + 310, + 757 + ], + "type": "text", + "content": "83" + } + ] + } + ], + "index": 20 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 4 + }, + { + "para_blocks": [ + { + "bbox": [ + 58, + 72, + 381, + 82 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 58, + 72, + 381, + 82 + ], + "spans": [ + { + "bbox": [ + 58, + 72, + 381, + 82 + ], + "type": "text", + "content": "(a) Activation function & loss landscape (—training trajectory, " + }, + { + "bbox": [ + 58, + 72, + 381, + 82 + ], + "type": "inline_equation", + "content": "\\star" + }, + { + "bbox": [ + 58, + 72, + 381, + 82 + ], + "type": "text", + "content": " early-stopping checkpoint, " + }, + { + "bbox": [ + 58, + 72, + 381, + 82 + ], + "type": "inline_equation", + "content": "\\bullet" + }, + { + "bbox": [ + 58, + 72, + 381, + 82 + ], + "type": "text", + "content": " last checkpoint)" + } + ] + } + ], + "index": 0 + }, + { + "type": "image", + "bbox": [ + 69, + 83, + 147, + 121 + ], + "blocks": [ + { + "bbox": [ + 69, + 83, + 147, + 121 + ], + "lines": [ + { + "bbox": [ + 69, + 83, + 147, + 121 + ], + "spans": [ + { + "bbox": [ + 69, + 83, + 147, + 121 + ], + "type": "image", + "image_path": "bad6edef9c1bc8728394cdffe30077a78da01f0153f0f1e2b12cd2182f52e52a.jpg" + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_body" + } + ], + "index": 1 + }, + { + "type": "image", + "bbox": [ + 168, + 83, + 246, + 121 + ], + "blocks": [ + { + "bbox": [ + 168, + 83, + 246, + 121 + ], + "lines": [ + { + "bbox": [ + 168, + 83, + 246, + 121 + ], + "spans": [ + { + "bbox": [ + 168, + 83, + 246, + 121 + ], + "type": "image", + "image_path": "6fd9e9d9dcd0f06ba6459546a2aa20b0438c7dbdff1b349bcb51e7b3c2ea0948.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + } + ], + "index": 2 + }, + { + "type": "image", + "bbox": [ + 267, + 83, + 344, + 121 + ], + "blocks": [ + { + "bbox": [ + 267, + 83, + 344, + 121 + ], + "lines": [ + { + "bbox": [ + 267, + 83, + 344, + 121 + ], + "spans": [ + { + "bbox": [ + 267, + 83, + 344, + 121 + ], + "type": "image", + "image_path": "f16688b3b136c6456aa0b3130512d904bec0357624b820eb582478f07a0d98a5.jpg" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_body" + } + ], + "index": 3 + }, + { + "type": "image", + "bbox": [ + 366, + 83, + 443, + 121 + ], + "blocks": [ + { + "bbox": [ + 366, + 83, + 443, + 121 + ], + "lines": [ + { + "bbox": [ + 366, + 83, + 443, + 121 + ], + "spans": [ + { + "bbox": [ + 366, + 83, + 443, + 121 + ], + "type": "image", + "image_path": "99f44ad2a7cdc1277b102e49942fded6ae5efa58eb372f813762876ff5c44f63.jpg" + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_body" + } + ], + "index": 4 + }, + { + "type": "image", + "bbox": [ + 475, + 100, + 531, + 114 + ], + "blocks": [ + { + "bbox": [ + 475, + 100, + 531, + 114 + ], + "lines": [ + { + "bbox": [ + 475, + 100, + 531, + 114 + ], + "spans": [ + { + "bbox": [ + 475, + 100, + 531, + 114 + ], + "type": "image", + "image_path": "1cdf1bd89d8f8b02937ffa5c9ccbc1c9f9ba6dda970a559b5727aeb0eb968311.jpg" + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "image_body" + } + ], + "index": 5 + }, + { + "bbox": [ + 58, + 126, + 394, + 136 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 58, + 126, + 394, + 136 + ], + "spans": [ + { + "bbox": [ + 58, + 126, + 394, + 136 + ], + "type": "text", + "content": "(b) Complexity landscape (in weight space along the PCA plane of the training trajectory) & zoom-in (random plane)" + } + ] + } + ], + "index": 6 + }, + { + "type": "image", + "bbox": [ + 69, + 137, + 145, + 175 + ], + "blocks": [ + { + "bbox": [ + 69, + 137, + 145, + 175 + ], + "lines": [ + { + "bbox": [ + 69, + 137, + 145, + 175 + ], + "spans": [ + { + "bbox": [ + 69, + 137, + 145, + 175 + ], + "type": "image", + "image_path": "c752fddc5da611e26c16531aac4e01617512987ab02341ed960dd2558d73ceb4.jpg" + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "image_body" + } + ], + "index": 7 + }, + { + "type": "image", + "bbox": [ + 168, + 137, + 244, + 175 + ], + "blocks": [ + { + "bbox": [ + 168, + 137, + 244, + 175 + ], + "lines": [ + { + "bbox": [ + 168, + 137, + 244, + 175 + ], + "spans": [ + { + "bbox": [ + 168, + 137, + 244, + 175 + ], + "type": "image", + "image_path": "c75597ea54fed9cf56bc03e79bbd5b1c3181f189013e5ee675e8f5575f82703d.jpg" + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "image_body" + } + ], + "index": 8 + }, + { + "type": "image", + "bbox": [ + 267, + 137, + 343, + 175 + ], + "blocks": [ + { + "bbox": [ + 267, + 137, + 343, + 175 + ], + "lines": [ + { + "bbox": [ + 267, + 137, + 343, + 175 + ], + "spans": [ + { + "bbox": [ + 267, + 137, + 343, + 175 + ], + "type": "image", + "image_path": "ec77cc9d600b0af491e2b18885eef5e0fb42f9dbce7b4e0069a2b003b0ed9c6d.jpg" + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "image_body" + } + ], + "index": 9 + }, + { + "type": "image", + "bbox": [ + 366, + 137, + 442, + 175 + ], + "blocks": [ + { + "bbox": [ + 366, + 137, + 442, + 175 + ], + "lines": [ + { + "bbox": [ + 366, + 137, + 442, + 175 + ], + "spans": [ + { + "bbox": [ + 366, + 137, + 442, + 175 + ], + "type": "image", + "image_path": "4bdcc9d70736b99d122ec4d8825092abe4aa1641cb36ba0638504d8998ff0304.jpg" + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "image_body" + } + ], + "index": 10 + }, + { + "type": "image", + "bbox": [ + 474, + 155, + 531, + 169 + ], + "blocks": [ + { + "bbox": [ + 474, + 155, + 531, + 169 + ], + "lines": [ + { + "bbox": [ + 474, + 155, + 531, + 169 + ], + "spans": [ + { + "bbox": [ + 474, + 155, + 531, + 169 + ], + "type": "image", + "image_path": "a05c6b4ff26d0923effae88c0343e254fbba0fcf68ebb66d679995d464212742.jpg" + } + ] + } + ], + "index": 11, + "angle": 0, + "type": "image_body" + } + ], + "index": 11 + }, + { + "bbox": [ + 58, + 181, + 405, + 190 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 58, + 181, + 405, + 190 + ], + "spans": [ + { + "bbox": [ + 58, + 181, + 405, + 190 + ], + "type": "text", + "content": "(c) Function implemented by the network (in input space along four random planes containing each one training point " + }, + { + "bbox": [ + 58, + 181, + 405, + 190 + ], + "type": "inline_equation", + "content": "\\bullet" + } + ] + } + ], + "index": 12 + }, + { + "type": "image", + "bbox": [ + 69, + 190, + 145, + 266 + ], + "blocks": [ + { + "bbox": [ + 69, + 190, + 145, + 266 + ], + "lines": [ + { + "bbox": [ + 69, + 190, + 145, + 266 + ], + "spans": [ + { + "bbox": [ + 69, + 190, + 145, + 266 + ], + "type": "image", + "image_path": "54f22c31c5781e224a169e8e7103f2010d27bd9217944f874670083a202ee4ce.jpg" + } + ] + } + ], + "index": 13, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 87, + 269, + 126, + 277 + ], + "lines": [ + { + "bbox": [ + 87, + 269, + 126, + 277 + ], + "spans": [ + { + "bbox": [ + 87, + 269, + 126, + 277 + ], + "type": "text", + "content": "MLP, ReLU" + } + ] + } + ], + "index": 14, + "angle": 0, + "type": "image_caption" + } + ], + "index": 13 + }, + { + "type": "image", + "bbox": [ + 168, + 190, + 244, + 266 + ], + "blocks": [ + { + "bbox": [ + 168, + 190, + 244, + 266 + ], + "lines": [ + { + "bbox": [ + 168, + 190, + 244, + 266 + ], + "spans": [ + { + "bbox": [ + 168, + 190, + 244, + 266 + ], + "type": "image", + "image_path": "d4288b4e375ddc9c5004153cc157bb14a828d67f3b7167e3dbd74c8af17a4ba8.jpg" + } + ] + } + ], + "index": 15, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 158, + 269, + 254, + 277 + ], + "lines": [ + { + "bbox": [ + 158, + 269, + 254, + 277 + ], + "spans": [ + { + "bbox": [ + 158, + 269, + 254, + 277 + ], + "type": "text", + "content": "MLP, TanH w/tuned prefactor" + } + ] + } + ], + "index": 16, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 196, + 281, + 422, + 290 + ], + "lines": [ + { + "bbox": [ + 196, + 281, + 422, + 290 + ], + "spans": [ + { + "bbox": [ + 196, + 281, + 422, + 290 + ], + "type": "text", + "content": "Increasing accuracy Best MLP" + } + ] + } + ], + "index": 23, + "angle": 0, + "type": "image_footnote" + } + ], + "index": 15 + }, + { + "type": "image", + "bbox": [ + 267, + 190, + 343, + 266 + ], + "blocks": [ + { + "bbox": [ + 267, + 190, + 343, + 266 + ], + "lines": [ + { + "bbox": [ + 267, + 190, + 343, + 266 + ], + "spans": [ + { + "bbox": [ + 267, + 190, + 343, + 266 + ], + "type": "image", + "image_path": "f5ba720390310337ea1ef721280057a57d81d2394ffe14a47dbc35040f6d3c3c.jpg" + } + ] + } + ], + "index": 17, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 277, + 269, + 332, + 277 + ], + "lines": [ + { + "bbox": [ + 277, + 269, + 332, + 277 + ], + "spans": [ + { + "bbox": [ + 277, + 269, + 332, + 277 + ], + "type": "text", + "content": "MLP, learned AF" + } + ] + } + ], + "index": 18, + "angle": 0, + "type": "image_caption" + } + ], + "index": 17 + }, + { + "type": "image", + "bbox": [ + 366, + 190, + 442, + 266 + ], + "blocks": [ + { + "bbox": [ + 366, + 190, + 442, + 266 + ], + "lines": [ + { + "bbox": [ + 366, + 190, + 442, + 266 + ], + "spans": [ + { + "bbox": [ + 366, + 190, + 442, + 266 + ], + "type": "image", + "image_path": "7c494a259530c3f2741489b48c8a6b8c473170b91091add9daa34e9b85069111.jpg" + } + ] + } + ], + "index": 19, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 373, + 269, + 435, + 277 + ], + "lines": [ + { + "bbox": [ + 373, + 269, + 435, + 277 + ], + "spans": [ + { + "bbox": [ + 373, + 269, + 435, + 277 + ], + "type": "text", + "content": "MLP, learned IAFs" + } + ] + } + ], + "index": 20, + "angle": 0, + "type": "image_caption" + } + ], + "index": 19 + }, + { + "type": "image", + "bbox": [ + 465, + 190, + 542, + 266 + ], + "blocks": [ + { + "bbox": [ + 465, + 190, + 542, + 266 + ], + "lines": [ + { + "bbox": [ + 465, + 190, + 542, + 266 + ], + "spans": [ + { + "bbox": [ + 465, + 190, + 542, + 266 + ], + "type": "image", + "image_path": "19bbd88571bfd919c19a65ad324180c481896a00f036a7ad0c321438a13ee397.jpg" + } + ] + } + ], + "index": 21, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 470, + 269, + 537, + 277 + ], + "lines": [ + { + "bbox": [ + 470, + 269, + 537, + 277 + ], + "spans": [ + { + "bbox": [ + 470, + 269, + 537, + 277 + ], + "type": "text", + "content": "Boosted decision trees" + } + ] + } + ], + "index": 22, + "angle": 0, + "type": "image_caption" + } + ], + "index": 21 + }, + { + "bbox": [ + 55, + 360, + 296, + 432 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 360, + 296, + 432 + ], + "spans": [ + { + "bbox": [ + 55, + 360, + 296, + 432 + ], + "type": "text", + "content": "ers, a higher complexity is best and the improvements with learned activations are larger. This supports the hypothesis that improvements over ReLU MLPs come from overcoming their simplicity bias. The variance across datasets is also unsurprising since they have little in common besides their low dimensionality (full results in Appendix E.4)." + } + ] + } + ], + "index": 25 + }, + { + "bbox": [ + 55, + 434, + 296, + 555 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 434, + 296, + 555 + ], + "spans": [ + { + "bbox": [ + 55, + 434, + 296, + 555 + ], + "type": "text", + "content": "Effect of width and depth. We show in Figure 8 that the learned activations can be reused in networks of different widths than they were trained for. The accuracy varies with width similarly as with ReLUs. Teney et al. [96] indeed showed that a model's width affects its capacity but not its inductive biases. Therefore width does not interfere with the effects of the learned activations. Figures 8 and 21 also show that good performance can be achieved with fewer layers than with ReLUs. Learned activations might thus have utility in model compression and distillation." + } + ] + } + ], + "index": 26 + }, + { + "bbox": [ + 62, + 563, + 289, + 623 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 563, + 289, + 623 + ], + "spans": [ + { + "bbox": [ + 62, + 563, + 289, + 623 + ], + "type": "text", + "content": "Take-away: many tabular datasets are ill-suited to ReLU models because they require learning a complex function. Learned activations improve accuracy by implementing sharp axis-aligned decision boundaries that mimic the inductive biases of decision trees." + } + ] + } + ], + "index": 27 + }, + { + "bbox": [ + 55, + 634, + 165, + 647 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 634, + 165, + 647 + ], + "spans": [ + { + "bbox": [ + 55, + 634, + 165, + 647 + ], + "type": "text", + "content": "3.4. Shortcut Learning" + } + ] + } + ], + "index": 28 + }, + { + "bbox": [ + 55, + 653, + 296, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 653, + 296, + 714 + ], + "spans": [ + { + "bbox": [ + 55, + 653, + 296, + 714 + ], + "type": "text", + "content": "Background. Shortcut learning occurs when a model learns spurious features instead of generalizable ones. It is a known consequence of the simplicity bias [85, 93] when the training data contains multiple features of different complexity. Our hypothesis is that the preference for some fea" + } + ] + } + ], + "index": 29 + }, + { + "type": "image", + "bbox": [ + 357, + 357, + 509, + 442 + ], + "blocks": [ + { + "bbox": [ + 55, + 293, + 555, + 348 + ], + "lines": [ + { + "bbox": [ + 55, + 293, + 555, + 348 + ], + "spans": [ + { + "bbox": [ + 55, + 293, + 555, + 348 + ], + "type": "text", + "content": "Figure 7. Models trained on the ELECTRICITY [34] tabular dataset. ReLU MLPs perform worst (left). TanHs induce sharper transitions in the network's function (c). So does the learned activation function (a) which is itself very irregular. The input activation functions (IAFs) perform best and mimic the axis-aligned boundaries of trees (bottom-right). The complexity landscapes (b) show that complexity increases to the highest level in the best model (IAFs). The zoom-in shows that the ambient complexity is also much higher than in other models. This means that it is inherently more likely to represent complex functions since they are more abundant in parameter space." + } + ] + } + ], + "index": 24, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 357, + 357, + 509, + 442 + ], + "lines": [ + { + "bbox": [ + 357, + 357, + 509, + 442 + ], + "spans": [ + { + "bbox": [ + 357, + 357, + 509, + 442 + ], + "type": "image", + "image_path": "5ad5edb24e4316146b38f696dd7a198fbb9ef5cfb04f3351a0cae4611512107e.jpg" + } + ] + } + ], + "index": 30, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 313, + 447, + 555, + 480 + ], + "lines": [ + { + "bbox": [ + 313, + 447, + 555, + 480 + ], + "spans": [ + { + "bbox": [ + 313, + 447, + 555, + 480 + ], + "type": "text", + "content": "Figure 8. The learned activation functions surpass ReLUs, often with fewer layers. They can also be reused with different network widths (COVERTYPE [34] tabular dataset, see Figure 21 for others)." + } + ] + } + ], + "index": 31, + "angle": 0, + "type": "image_caption" + } + ], + "index": 30 + }, + { + "bbox": [ + 313, + 498, + 554, + 533 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 498, + 554, + 533 + ], + "spans": [ + { + "bbox": [ + 313, + 498, + 554, + 533 + ], + "type": "text", + "content": "tures depends on their alignment with the inductive biases. We will evaluate whether this can be controlled with activation functions." + } + ] + } + ], + "index": 32 + }, + { + "bbox": [ + 313, + 540, + 555, + 648 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 540, + 555, + 648 + ], + "spans": [ + { + "bbox": [ + 313, + 540, + 555, + 648 + ], + "type": "text", + "content": "Setup. We use MNIST/CIFAR collages [85, 93, 94], a classification task over images combining tiles from MNIST and CIFAR-10. The training set is ambiguous: both tiles are predictive of the labels. Two unambiguous test sets evaluate reliance on either tile: one is predictive, the other contains a random class. We similarly build two validation sets to learn two activation functions optimized for either tile. We simulate OOD conditions by setting " + }, + { + "bbox": [ + 313, + 540, + 555, + 648 + ], + "type": "inline_equation", + "content": "\\mathcal{V}" + }, + { + "bbox": [ + 313, + 540, + 555, + 648 + ], + "type": "text", + "content": " in Algorithm 1. The models are the fully-connected MLPs used in [93]." + } + ] + } + ], + "index": 33 + }, + { + "bbox": [ + 313, + 653, + 555, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 653, + 555, + 714 + ], + "spans": [ + { + "bbox": [ + 313, + 653, + 555, + 714 + ], + "type": "text", + "content": "Results. Figure 9 shows that a baseline with ReLUs is prone to shortcut learning. It relies exclusively on MNIST and the accuracy on the CIFAR test set is not better than chance (10%). In comparison, using either learned activation steers the learning towards either tile. The accuracy" + } + ] + } + ], + "index": 34 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 299, + 749, + 310, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 299, + 749, + 310, + 757 + ], + "spans": [ + { + "bbox": [ + 299, + 749, + 310, + 757 + ], + "type": "text", + "content": "84" + } + ] + } + ], + "index": 35 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 5 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 65, + 71, + 104, + 112 + ], + "blocks": [ + { + "bbox": [ + 65, + 71, + 104, + 112 + ], + "lines": [ + { + "bbox": [ + 65, + 71, + 104, + 112 + ], + "spans": [ + { + "bbox": [ + 65, + 71, + 104, + 112 + ], + "type": "image", + "image_path": "d6b247dcca639738fdf5d443fe4f7bfe462550793593ecf7228ca8b6bf910e45.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 58, + 118, + 110, + 137 + ], + "lines": [ + { + "bbox": [ + 58, + 118, + 110, + 137 + ], + "spans": [ + { + "bbox": [ + 58, + 118, + 110, + 137 + ], + "type": "text", + "content": "Ambiguous training images" + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "type": "image", + "bbox": [ + 112, + 70, + 157, + 114 + ], + "blocks": [ + { + "bbox": [ + 112, + 70, + 157, + 114 + ], + "lines": [ + { + "bbox": [ + 112, + 70, + 157, + 114 + ], + "spans": [ + { + "bbox": [ + 112, + 70, + 157, + 114 + ], + "type": "image", + "image_path": "c684f7795bd28c20580a087965e31eb6cc8ac02ab707ed5e709cc79ec3e5d2ac.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 117, + 118, + 153, + 137 + ], + "lines": [ + { + "bbox": [ + 117, + 118, + 153, + 137 + ], + "spans": [ + { + "bbox": [ + 117, + 118, + 153, + 137 + ], + "type": "text", + "content": "Optimized for MNIST" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_caption" + } + ], + "index": 2 + }, + { + "type": "image", + "bbox": [ + 159, + 70, + 204, + 114 + ], + "blocks": [ + { + "bbox": [ + 159, + 70, + 204, + 114 + ], + "lines": [ + { + "bbox": [ + 159, + 70, + 204, + 114 + ], + "spans": [ + { + "bbox": [ + 159, + 70, + 204, + 114 + ], + "type": "image", + "image_path": "50c38e44d937b34048fe3365f9b2308f201750ababa6cde9e85bc8c5ddf896b4.jpg" + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 164, + 118, + 200, + 136 + ], + "lines": [ + { + "bbox": [ + 164, + 118, + 200, + 136 + ], + "spans": [ + { + "bbox": [ + 164, + 118, + 200, + 136 + ], + "type": "text", + "content": "Optimized for CIFAR" + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "image_caption" + } + ], + "index": 4 + }, + { + "type": "image", + "bbox": [ + 334, + 71, + 488, + 133 + ], + "blocks": [ + { + "bbox": [ + 205, + 73, + 334, + 112 + ], + "lines": [ + { + "bbox": [ + 205, + 73, + 334, + 112 + ], + "spans": [ + { + "bbox": [ + 205, + 73, + 334, + 112 + ], + "type": "text", + "content": "Upper bound: ReLU trained on unambiguous MNIST Activation optimized for MNIST ReLU Baseline Activation optimized for CIFAR Upper bound: ReLU trained on unambiguous CIFAR" + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 334, + 71, + 488, + 133 + ], + "lines": [ + { + "bbox": [ + 334, + 71, + 488, + 133 + ], + "spans": [ + { + "bbox": [ + 334, + 71, + 488, + 133 + ], + "type": "image", + "image_path": "a04e48553aa4a0d68c21ff4539f2373cf52f17a1413cee69f194d411e026db28.jpg" + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "image_body" + } + ], + "index": 7 + }, + { + "type": "image", + "bbox": [ + 488, + 70, + 550, + 132 + ], + "blocks": [ + { + "bbox": [ + 488, + 70, + 550, + 132 + ], + "lines": [ + { + "bbox": [ + 488, + 70, + 550, + 132 + ], + "spans": [ + { + "bbox": [ + 488, + 70, + 550, + 132 + ], + "type": "image", + "image_path": "301cbb3be4b74660f7cb9c81465c5f311bd49aa3c6329f6e809d035827b8f60b.jpg" + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "image_body" + } + ], + "index": 8 + }, + { + "bbox": [ + 55, + 189, + 296, + 248 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 189, + 296, + 248 + ], + "spans": [ + { + "bbox": [ + 55, + 189, + 296, + 248 + ], + "type": "text", + "content": "shifts towards either of two tiles as the model prioritizes different features, merely with a change of activation function. This shows that the simplicity bias is not an inevitable effect of SGD. Instead, it directly reflects the alignment between the chosen architecture and the data." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 55, + 251, + 296, + 360 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 251, + 296, + 360 + ], + "spans": [ + { + "bbox": [ + 55, + 251, + 296, + 360 + ], + "type": "text", + "content": "Training dynamics. The accuracy on CIFAR remains below a model trained on unambiguous CIFAR data. This is because training dynamics are also important. In Figure 9 (right), we plot the accuracy on the two tiles for the whole training trajectory. The reliance on different features varies, and the model eventually relies primarily on simple ones with enough iterations (i.e. without early stopping). This calls for future work combining our findings with the extensive literature on ID/OOD training dynamics [46, 95, 98]." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 62, + 368, + 289, + 417 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 368, + 289, + 417 + ], + "spans": [ + { + "bbox": [ + 62, + 368, + 289, + 417 + ], + "type": "text", + "content": "Take-away: we confirm that shortcut learning is a side effect of the simplicity bias. Different activation functions, while not completely avoiding shortcut learning, can steer the learning towards particular input features." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 55, + 428, + 234, + 441 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 428, + 234, + 441 + ], + "spans": [ + { + "bbox": [ + 55, + 428, + 234, + 441 + ], + "type": "text", + "content": "3.5. Algorithmic Tasks and Grokking" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 55, + 447, + 296, + 651 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 447, + 296, + 651 + ], + "spans": [ + { + "bbox": [ + 55, + 447, + 296, + 651 + ], + "type": "text", + "content": "Background. Grokking is a phenomenon where a model first overfits the data (i.e. high training accuracy, low test accuracy) then shifts to high test accuracy after many training steps [73]. This is typically observed on algorithmic tasks and architectures from MLPs to transformers. Our hypothesis is that grokking is due to a mismatch between the target function and the model's inductive biases. Indeed, typical architectures were not developed for the algorithmic tasks where grokking is typically observed. To verify this hypothesis, we will show that endowing an architecture with the right inductive biases, using learned activation functions, can eliminate the phenomenon. Supporting this hypothesis, Zhou et al. [110] proposed that grokking comes from the frequency principle (i.e. low frequencies learned first by SGD), and Kumar et al. [51] showed that it correlates with a misalignment between features at initialization and the target function." + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 55, + 653, + 296, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 653, + 296, + 715 + ], + "spans": [ + { + "bbox": [ + 55, + 653, + 296, + 715 + ], + "type": "text", + "content": "Setup. Following [36, 51, 58] we train 1-hidden layer MLPs on algorithmic tasks, defined each by one binary operation (Figures 10 and 27). e.g. " + }, + { + "bbox": [ + 55, + 653, + 296, + 715 + ], + "type": "inline_equation", + "content": "y = (x_{1} + x_{2}) \\mod 13" + }, + { + "bbox": [ + 55, + 653, + 296, + 715 + ], + "type": "text", + "content": ". The operands are passed as one-hot vectors and the task is a classification over possible outputs. Details in Appendix E.6." + } + ] + } + ], + "index": 15 + }, + { + "type": "image", + "bbox": [ + 315, + 188, + 354, + 245 + ], + "blocks": [ + { + "bbox": [ + 55, + 144, + 555, + 178 + ], + "lines": [ + { + "bbox": [ + 55, + 144, + 555, + 178 + ], + "spans": [ + { + "bbox": [ + 55, + 144, + 555, + 178 + ], + "type": "text", + "content": "Figure 9. Experiments on shortcut learning with MNIST/CIFAR collages. The ReLU baseline (■) relies mostly on simple MNIST features. We learn two activation functions that shift the preference towards different features " + }, + { + "bbox": [ + 55, + 144, + 555, + 178 + ], + "type": "inline_equation", + "content": "(\\leftarrow / \\rightarrow)" + }, + { + "bbox": [ + 55, + 144, + 555, + 178 + ], + "type": "text", + "content": ". Training trajectories (right) clearly differ with the activation optimized for CIFAR (—), for MNIST (—), or a ReLU (==). The model at initialization (random weights) is marked with " + }, + { + "bbox": [ + 55, + 144, + 555, + 178 + ], + "type": "inline_equation", + "content": "\\bullet" + }, + { + "bbox": [ + 55, + 144, + 555, + 178 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 315, + 188, + 354, + 245 + ], + "lines": [ + { + "bbox": [ + 315, + 188, + 354, + 245 + ], + "spans": [ + { + "bbox": [ + 315, + 188, + 354, + 245 + ], + "type": "image", + "image_path": "d1d11deeb4a9b9b81cb958f58c49724164970864e0e5da9c8d343a1d345e6c86.jpg" + } + ] + } + ], + "index": 16, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 313, + 250, + 554, + 283 + ], + "lines": [ + { + "bbox": [ + 313, + 250, + 554, + 283 + ], + "spans": [ + { + "bbox": [ + 313, + 250, + 554, + 283 + ], + "type": "text", + "content": "Figure 10. Target functions used to investigate grokking [73] (details in Appendix, Figure 27). These patterns are very different from the tasks for which typical architectures were developed." + } + ] + } + ], + "index": 20, + "angle": 0, + "type": "image_caption" + } + ], + "index": 16 + }, + { + "type": "image", + "bbox": [ + 357, + 188, + 452, + 245 + ], + "blocks": [ + { + "bbox": [ + 357, + 188, + 452, + 245 + ], + "lines": [ + { + "bbox": [ + 357, + 188, + 452, + 245 + ], + "spans": [ + { + "bbox": [ + 357, + 188, + 452, + 245 + ], + "type": "image", + "image_path": "3dcee3491eae684eff053e6f0f384f70f597e8055a8aa1e15456e74be9a04da1.jpg" + } + ] + } + ], + "index": 17, + "angle": 0, + "type": "image_body" + } + ], + "index": 17 + }, + { + "type": "image", + "bbox": [ + 454, + 188, + 513, + 245 + ], + "blocks": [ + { + "bbox": [ + 454, + 188, + 513, + 245 + ], + "lines": [ + { + "bbox": [ + 454, + 188, + 513, + 245 + ], + "spans": [ + { + "bbox": [ + 454, + 188, + 513, + 245 + ], + "type": "image", + "image_path": "cd388dbc24fcdf1dd019094bd395c93e32996e3d3a1373fed0218c36523683df.jpg" + } + ] + } + ], + "index": 18, + "angle": 0, + "type": "image_body" + } + ], + "index": 18 + }, + { + "type": "image", + "bbox": [ + 514, + 188, + 553, + 245 + ], + "blocks": [ + { + "bbox": [ + 514, + 188, + 553, + 245 + ], + "lines": [ + { + "bbox": [ + 514, + 188, + 553, + 245 + ], + "spans": [ + { + "bbox": [ + 514, + 188, + 553, + 245 + ], + "type": "image", + "image_path": "44a767f71cfd239d8d47af4809f0d81fb951876d5eff84ae58a7f0445a81e582.jpg" + } + ] + } + ], + "index": 19, + "angle": 0, + "type": "image_body" + } + ], + "index": 19 + }, + { + "bbox": [ + 313, + 287, + 555, + 456 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 287, + 555, + 456 + ], + "spans": [ + { + "bbox": [ + 313, + 287, + 555, + 456 + ], + "type": "text", + "content": "Results. We compare models with ReLU vs. learned activations across various tasks, network widths, and fractions of training data. We find that the learned, task-specific activations lead to faster convergence and/or higher test accuracy (Figures 11-13). On modular addition (a common task in the grokking literature) the learned-activation model converges " + }, + { + "bbox": [ + 313, + 287, + 555, + 456 + ], + "type": "inline_equation", + "content": "\\sim 10\\times" + }, + { + "bbox": [ + 313, + 287, + 555, + 456 + ], + "type": "text", + "content": " faster than ReLUs. Curiously, some models with learned activations also end up overfitting (decreasing test accuracy) with prolonged training. In contrast, ReLU networks either never generalize (test accuracy " + }, + { + "bbox": [ + 313, + 287, + 555, + 456 + ], + "type": "inline_equation", + "content": "\\sim 0" + }, + { + "bbox": [ + 313, + 287, + 555, + 456 + ], + "type": "text", + "content": ") or grok and keep a high accuracy indefinitely. Further investigation is needed to explain this difference. We examine learned activation functions in Figure 12. See Figure 28 in the Appendix for results on other algorithmic tasks [73]." + } + ] + } + ], + "index": 21 + }, + { + "type": "image", + "bbox": [ + 317, + 460, + 454, + 509 + ], + "blocks": [ + { + "bbox": [ + 317, + 460, + 454, + 509 + ], + "lines": [ + { + "bbox": [ + 317, + 460, + 454, + 509 + ], + "spans": [ + { + "bbox": [ + 317, + 460, + 454, + 509 + ], + "type": "image", + "image_path": "9920403831ee31eaddfe832ce9630e4b9cd06b614de2382e9554c6094a3e8800.jpg" + } + ] + } + ], + "index": 22, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 313, + 511, + 554, + 544 + ], + "lines": [ + { + "bbox": [ + 313, + 511, + 554, + 544 + ], + "spans": [ + { + "bbox": [ + 313, + 511, + 554, + 544 + ], + "type": "text", + "content": "Figure 11. The learned activations essentially eliminate grokking (delayed convergence). On the above task (addition mod 27), our model converges " + }, + { + "bbox": [ + 313, + 511, + 554, + 544 + ], + "type": "inline_equation", + "content": "\\sim 10\\times" + }, + { + "bbox": [ + 313, + 511, + 554, + 544 + ], + "type": "text", + "content": " faster than ReLUs." + } + ] + } + ], + "index": 24, + "angle": 0, + "type": "image_caption" + } + ], + "index": 22 + }, + { + "type": "image", + "bbox": [ + 454, + 460, + 551, + 509 + ], + "blocks": [ + { + "bbox": [ + 454, + 460, + 551, + 509 + ], + "lines": [ + { + "bbox": [ + 454, + 460, + 551, + 509 + ], + "spans": [ + { + "bbox": [ + 454, + 460, + 551, + 509 + ], + "type": "image", + "image_path": "83e4f3ac0255d05e53e3fa527646c06887f09c4c75c47fd1dc7899320d659641.jpg" + } + ] + } + ], + "index": 23, + "angle": 0, + "type": "image_body" + } + ], + "index": 23 + }, + { + "type": "image", + "bbox": [ + 323, + 552, + 369, + 581 + ], + "blocks": [ + { + "bbox": [ + 323, + 552, + 369, + 581 + ], + "lines": [ + { + "bbox": [ + 323, + 552, + 369, + 581 + ], + "spans": [ + { + "bbox": [ + 323, + 552, + 369, + 581 + ], + "type": "image", + "image_path": "b5b4266d96db9f14d1a6c8a2fe2d628a45a4da699f3ee05d5f3f7357a326fefc.jpg" + } + ] + } + ], + "index": 25, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 313, + 582, + 554, + 604 + ], + "lines": [ + { + "bbox": [ + 313, + 582, + 554, + 604 + ], + "spans": [ + { + "bbox": [ + 313, + 582, + 554, + 604 + ], + "type": "text", + "content": "Figure 12. Activations learned for modular addition. The frequency of the sine-like function varies across versions of the task." + } + ] + } + ], + "index": 29, + "angle": 0, + "type": "image_caption" + } + ], + "index": 25 + }, + { + "type": "image", + "bbox": [ + 382, + 552, + 428, + 581 + ], + "blocks": [ + { + "bbox": [ + 382, + 552, + 428, + 581 + ], + "lines": [ + { + "bbox": [ + 382, + 552, + 428, + 581 + ], + "spans": [ + { + "bbox": [ + 382, + 552, + 428, + 581 + ], + "type": "image", + "image_path": "65eee4c6eee6e76429bc7140d09f002af221947e030d304d5c5245afa7708ae2.jpg" + } + ] + } + ], + "index": 26, + "angle": 0, + "type": "image_body" + } + ], + "index": 26 + }, + { + "type": "image", + "bbox": [ + 441, + 552, + 487, + 581 + ], + "blocks": [ + { + "bbox": [ + 441, + 552, + 487, + 581 + ], + "lines": [ + { + "bbox": [ + 441, + 552, + 487, + 581 + ], + "spans": [ + { + "bbox": [ + 441, + 552, + 487, + 581 + ], + "type": "image", + "image_path": "294f53ae5614dc603af10359800aaf25913da0c286cb71d8b322f390eb549d1d.jpg" + } + ] + } + ], + "index": 27, + "angle": 0, + "type": "image_body" + } + ], + "index": 27 + }, + { + "type": "image", + "bbox": [ + 500, + 552, + 546, + 581 + ], + "blocks": [ + { + "bbox": [ + 500, + 552, + 546, + 581 + ], + "lines": [ + { + "bbox": [ + 500, + 552, + 546, + 581 + ], + "spans": [ + { + "bbox": [ + 500, + 552, + 546, + 581 + ], + "type": "image", + "image_path": "ee32959f3b5caec906a43d63ac57bf87352145a3ed5d6be36ab461f7a76ba870.jpg" + } + ] + } + ], + "index": 28, + "angle": 0, + "type": "image_body" + } + ], + "index": 28 + }, + { + "type": "image", + "bbox": [ + 341, + 611, + 436, + 673 + ], + "blocks": [ + { + "bbox": [ + 341, + 611, + 436, + 673 + ], + "lines": [ + { + "bbox": [ + 341, + 611, + 436, + 673 + ], + "spans": [ + { + "bbox": [ + 341, + 611, + 436, + 673 + ], + "type": "image", + "image_path": "d2f0504c3ce9c9632681a94e323a57ef64b3ce024d7b4d7f7b22b2c8da9e9cc3.jpg" + } + ] + } + ], + "index": 30, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 313, + 678, + 554, + 700 + ], + "lines": [ + { + "bbox": [ + 313, + 678, + 554, + 700 + ], + "spans": [ + { + "bbox": [ + 313, + 678, + 554, + 700 + ], + "type": "text", + "content": "Figure 13. Models with learned activations " + }, + { + "bbox": [ + 313, + 678, + 554, + 700 + ], + "type": "inline_equation", + "content": "(-)" + }, + { + "bbox": [ + 313, + 678, + 554, + 700 + ], + "type": "text", + "content": " converge faster than ReLUs " + }, + { + "bbox": [ + 313, + 678, + 554, + 700 + ], + "type": "inline_equation", + "content": "(--)" + }, + { + "bbox": [ + 313, + 678, + 554, + 700 + ], + "type": "text", + "content": " across a variety of settings (addition mod 27)." + } + ] + } + ], + "index": 32, + "angle": 0, + "type": "image_caption" + } + ], + "index": 30 + }, + { + "type": "image", + "bbox": [ + 443, + 610, + 499, + 673 + ], + "blocks": [ + { + "bbox": [ + 443, + 610, + 499, + 673 + ], + "lines": [ + { + "bbox": [ + 443, + 610, + 499, + 673 + ], + "spans": [ + { + "bbox": [ + 443, + 610, + 499, + 673 + ], + "type": "image", + "image_path": "053a8c1203fb699997e086dc2755219911c6bd571080ca3bd2683dd09be94139.jpg" + } + ] + } + ], + "index": 31, + "angle": 0, + "type": "image_body" + } + ], + "index": 31 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 299, + 749, + 310, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 299, + 749, + 310, + 757 + ], + "spans": [ + { + "bbox": [ + 299, + 749, + 310, + 757 + ], + "type": "text", + "content": "85" + } + ] + } + ], + "index": 33 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 6 + }, + { + "para_blocks": [ + { + "bbox": [ + 63, + 74, + 288, + 110 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 63, + 74, + 288, + 110 + ], + "spans": [ + { + "bbox": [ + 63, + 74, + 288, + 110 + ], + "type": "text", + "content": "Take-away: learned activations eliminate grokking in all our cases, suggesting, as a cause, the mismatch between the data and the architectures' inductive biases." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 55, + 123, + 287, + 136 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 123, + 287, + 136 + ], + "spans": [ + { + "bbox": [ + 55, + 123, + 287, + 136 + ], + "type": "text", + "content": "4. Do the Activations Transfer Across Tasks?" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 55, + 143, + 295, + 191 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 143, + 295, + 191 + ], + "spans": [ + { + "bbox": [ + 55, + 143, + 295, + 191 + ], + "type": "text", + "content": "So far, we used dataset-specific activation functions and found that there exist better alternatives to ReLUs. A practical application would be the learning of activation functions suitable to a broad task, or range of related datasets." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 55, + 192, + 296, + 360 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 192, + 296, + 360 + ], + "spans": [ + { + "bbox": [ + 55, + 192, + 296, + 360 + ], + "type": "text", + "content": "As a first step, we study the specialization of the activations functions (AFs) learned for the 22 algorithmic tasks from Figure 27 [73]. We evaluate every task/activation combination, yielding the " + }, + { + "bbox": [ + 55, + 192, + 296, + 360 + ], + "type": "inline_equation", + "content": "22 \\times 22" + }, + { + "bbox": [ + 55, + 192, + 296, + 360 + ], + "type": "text", + "content": " matrix of Figure 14. The learned activations do transfer, with improvements in accuracy and convergence shared across tasks. We also evaluate an activation learned on all tasks simultaneously. The accuracy across tasks (i.e. per-column average) reaches " + }, + { + "bbox": [ + 55, + 192, + 296, + 360 + ], + "type": "inline_equation", + "content": "61.5\\%" + }, + { + "bbox": [ + 55, + 192, + 296, + 360 + ], + "type": "text", + "content": " vs. only " + }, + { + "bbox": [ + 55, + 192, + 296, + 360 + ], + "type": "inline_equation", + "content": "19.9\\%" + }, + { + "bbox": [ + 55, + 192, + 296, + 360 + ], + "type": "text", + "content": " for ReLUs. and " + }, + { + "bbox": [ + 55, + 192, + 296, + 360 + ], + "type": "inline_equation", + "content": "54.0\\%" + }, + { + "bbox": [ + 55, + 192, + 296, + 360 + ], + "type": "text", + "content": " on average for tasks-specific solutions. This procedure can thus improve performance on a range of related tasks. Future work could leverage it to discover activation functions that improve performance in other specific domains. See Appendix E.3 for other transfer experiments using image regression tasks." + } + ] + } + ], + "index": 3 + }, + { + "type": "image", + "bbox": [ + 56, + 362, + 296, + 490 + ], + "blocks": [ + { + "bbox": [ + 56, + 362, + 296, + 490 + ], + "lines": [ + { + "bbox": [ + 56, + 362, + 296, + 490 + ], + "spans": [ + { + "bbox": [ + 56, + 362, + 296, + 490 + ], + "type": "image", + "image_path": "616059b4a69b28d57058a7ca58c82ad3cebff55fedc8cedb0c41fe1f7be2a3a3.jpg" + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 55, + 494, + 295, + 548 + ], + "lines": [ + { + "bbox": [ + 55, + 494, + 295, + 548 + ], + "spans": [ + { + "bbox": [ + 55, + 494, + 295, + 548 + ], + "type": "text", + "content": "Figure 14. Transfer of AFs (columns) across algorithmic tasks (rows). Colors represent the fraction of the best convergence speed or accuracy per task (brighter is better). If the activations were over-specialized, the matrix would be diagonal. On the contrary, it is densely filled, indicating positive transfer across many tasks." + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "image_caption" + } + ], + "index": 4 + }, + { + "bbox": [ + 55, + 562, + 125, + 574 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 562, + 125, + 574 + ], + "spans": [ + { + "bbox": [ + 55, + 562, + 125, + 574 + ], + "type": "text", + "content": "5. Discussion" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 55, + 582, + 296, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 582, + 296, + 713 + ], + "spans": [ + { + "bbox": [ + 55, + 582, + 296, + 713 + ], + "type": "text", + "content": "We used activation functions as a tool to show that there exist a variety of inductive biases that are useful across applications of NNs. The impossibility of universal inductive biases is well known [101] but a strong argument has also been made that deep learning research is converging towards few architectures with wide applicability [31]. This argument rests on the NNs' simplicity bias being a good match for real-world data [12, 18, 56]. Our results do not invalidate these assumptions: NNs are widely applicable and their simplicity bias is evidently very effective on average. Our results show instead the following." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 314, + 72, + 555, + 194 + ], + "type": "list", + "angle": 0, + "index": 10, + "blocks": [ + { + "bbox": [ + 315, + 72, + 555, + 120 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 315, + 72, + 555, + 120 + ], + "spans": [ + { + "bbox": [ + 315, + 72, + 555, + 120 + ], + "type": "text", + "content": "1. There exist real-world tasks where the inductive biases of typical architectures' are suboptimal. This explanation connects four domains where NNs historically struggled." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 314, + 122, + 555, + 194 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 314, + 122, + 555, + 194 + ], + "spans": [ + { + "bbox": [ + 314, + 122, + 555, + 194 + ], + "type": "text", + "content": "2. The simplicity bias in modern NNs depends on particular design choices, the activation functions in particular. Research has converged on these choices by trial and error, in large part by optimizing performance on vision tasks. Therefore the adequacy of ReLUs for image classification (Section 3.1) is not accidental." + } + ] + } + ], + "index": 9 + } + ], + "sub_type": "text" + }, + { + "bbox": [ + 313, + 198, + 555, + 293 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 198, + 555, + 293 + ], + "spans": [ + { + "bbox": [ + 313, + 198, + 555, + 293 + ], + "type": "text", + "content": "Relevance to transformers and language models. The simplicity bias exists in transformers [10, 109] and language models [4, 31, 96]. Their embedding layer resembles input activations functions (Section 3.3). Could this explain the transformers' remarkable flexibility? I.e. a simplicity bias on an initial mapping of arbitrary complexity. Zhong and Andreas [108] indeed trained embeddings alone in a random-weight transformer and could learn complex tasks." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 313, + 297, + 556, + 393 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 297, + 556, + 393 + ], + "spans": [ + { + "bbox": [ + 313, + 297, + 556, + 393 + ], + "type": "text", + "content": "Limitations and open questions. We prioritized breadth by establishing a new connection across multiple disparate topics in machine learning. Each section could expand into its own paper with additional models, datasets, comparisons, etc. Our findings on shortcut learning for example (Section 3.4) could yield new methods to address distribution shifts, though no such claim is made here. Here are the most promising follow-up questions opened by this paper." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 315, + 396, + 556, + 714 + ], + "type": "list", + "angle": 0, + "index": 17, + "blocks": [ + { + "bbox": [ + 315, + 396, + 554, + 444 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 315, + 396, + 554, + 444 + ], + "spans": [ + { + "bbox": [ + 315, + 396, + 554, + 444 + ], + "type": "text", + "content": "- How to fully characterize inductive biases? We focused on simplicity for its prevalence in AI [31], philosophy [72], and the natural sciences [12]. But it is only one dimension among many to characterize inductive biases." + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 315, + 447, + 555, + 506 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 315, + 447, + 555, + 506 + ], + "spans": [ + { + "bbox": [ + 315, + 447, + 555, + 506 + ], + "type": "text", + "content": "- Can we improve state-of-the-art architectures? We used simple MLPs to isolate the effects of activations functions since they are central to the simplicity bias [96]. But other existing mechanisms (architectural, optimization) may already tweak or attenuate the simplicity bias." + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 315, + 508, + 555, + 592 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 315, + 508, + 555, + 592 + ], + "spans": [ + { + "bbox": [ + 315, + 508, + 555, + 592 + ], + "type": "text", + "content": "- Can we learn transferable activation functions for other domains? We examined transferability in Sections 4 and E.3. The results suggest the possibility of better architectures optimized for specific domains. Predicting the suitability of an architecture/dataset pair ex ante (prior to training) would be extremely useful. This may follow from advances on the first open question above." + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 315, + 594, + 556, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 315, + 594, + 556, + 714 + ], + "spans": [ + { + "bbox": [ + 315, + 594, + 556, + 714 + ], + "type": "text", + "content": "- Are there other detrimental effects of the simplicity bias? Any learning algorithm needs inductive biases to \"fill the gaps\" between training examples. The better they are, the fewer examples are needed. Researching what inductive biases are most useful on real-world tasks might thus hold the key for machine learning to become as data-efficient as humans. More speculatively, high-level cognition has been argued to require postulating explanations beyond the data [16, 23]. In this regard, simplicity-biased architectures might also hold us back." + } + ] + } + ], + "index": 16 + } + ], + "sub_type": "text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 300, + 749, + 310, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 300, + 749, + 310, + 757 + ], + "spans": [ + { + "bbox": [ + 300, + 749, + 310, + 757 + ], + "type": "text", + "content": "86" + } + ] + } + ], + "index": 18 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 7 + }, + { + "para_blocks": [ + { + "bbox": [ + 56, + 71, + 115, + 83 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 71, + 115, + 83 + ], + "spans": [ + { + "bbox": [ + 56, + 71, + 115, + 83 + ], + "type": "text", + "content": "References" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 61, + 89, + 296, + 713 + ], + "type": "list", + "angle": 0, + "index": 18, + "blocks": [ + { + "bbox": [ + 66, + 89, + 296, + 133 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 66, + 89, + 296, + 133 + ], + "spans": [ + { + "bbox": [ + 66, + 89, + 296, + 133 + ], + "type": "text", + "content": "[1] Sravanti Addepalli, Anshul Nasery, Venkatesh Babu Radhakrishnan, Praneeth Netrapalli, and Prateek Jain. Feature reconstruction from outputs can mitigate simplicity bias in neural networks. In ICLR, 2022. 1" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 65, + 134, + 295, + 167 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 65, + 134, + 295, + 167 + ], + "spans": [ + { + "bbox": [ + 65, + 134, + 295, + 167 + ], + "type": "text", + "content": "[2] Konstantinos Panagiotis Alexandridis, Jiankang Deng, Anh Nguyen, and Shan Luo. Adaptive parametric activation. arXiv preprint arXiv:2407.08567, 2024. 2, 3, 1" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 65, + 168, + 295, + 201 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 65, + 168, + 295, + 201 + ], + "spans": [ + { + "bbox": [ + 65, + 168, + 295, + 201 + ], + "type": "text", + "content": "[3] Konstantinos Panagiotis Alexandridis, Jiankang Deng, Anh Nguyen, and Shan Luo. Adaptive parametric activation. In ECCV, 2025. 2" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 65, + 203, + 295, + 246 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 65, + 203, + 295, + 246 + ], + "spans": [ + { + "bbox": [ + 65, + 203, + 295, + 246 + ], + "type": "text", + "content": "[4] Badr AlKhamissi, Greta Tuckute, Antoine Bosselut, and Martin Schrimpf. Brain-like language processing via a shallow untrained multihead attention network. arXiv preprint arXiv:2406.15109, 2024. 8" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 65, + 248, + 295, + 281 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 65, + 248, + 295, + 281 + ], + "spans": [ + { + "bbox": [ + 65, + 248, + 295, + 281 + ], + "type": "text", + "content": "[5] Andrea Apicella, Francesco Isgro, and Roberto Prevete. A simple and efficient architecture for trainable activation functions. Neurocomputing, 2019. 3, 1, 2" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 65, + 282, + 295, + 315 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 65, + 282, + 295, + 315 + ], + "spans": [ + { + "bbox": [ + 65, + 282, + 295, + 315 + ], + "type": "text", + "content": "[6] Andrea Apicella, Francesco Donnarumma, Francesco Isgro, and Roberto Prevete. A survey on modern trainable activation functions. Neural Networks, 2021. 2, 3, 1" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 65, + 317, + 295, + 348 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 65, + 317, + 295, + 348 + ], + "spans": [ + { + "bbox": [ + 65, + 317, + 295, + 348 + ], + "type": "text", + "content": "[7] Sanjeev Arora, Nadav Cohen, Wei Hu, and Yuping Luo. Implicit regularization in deep matrix factorization. NeurIPS, 2019. 1" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 65, + 350, + 295, + 404 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 65, + 350, + 295, + 404 + ], + "spans": [ + { + "bbox": [ + 65, + 350, + 295, + 404 + ], + "type": "text", + "content": "[8] Devansh Arpit, Stanislaw Jastrzebski, Nicolas Ballas, David Krueger, Emmanuel Bengio, Maxinder S Kanwal, Tegan Maharaj, Asja Fischer, Aaron Courville, Yoshua Bengio, et al. A closer look at memorization in deep networks. In ICML. PMLR, 2017. 1" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 65, + 407, + 296, + 450 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 65, + 407, + 296, + 450 + ], + "spans": [ + { + "bbox": [ + 65, + 407, + 296, + 450 + ], + "type": "text", + "content": "[9] Samuel James Bell and Levent Sagun. Simplicity bias leads to amplified performance disparities. In Proceedings of the 2023 ACM Conference on Fairness, Accountability, and Transparency, pages 355-369, 2023. 1" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 61, + 452, + 295, + 495 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 61, + 452, + 295, + 495 + ], + "spans": [ + { + "bbox": [ + 61, + 452, + 295, + 495 + ], + "type": "text", + "content": "[10] Satwik Bhattachamishra, Arkil Patel, Varun Kanade, and Phil Blunsom. Simplicity bias in transformers and their ability to learn sparse boolean functions. arXiv preprint arXiv:2211.12316, 2022.8" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 61, + 497, + 295, + 540 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 61, + 497, + 295, + 540 + ], + "spans": [ + { + "bbox": [ + 61, + 497, + 295, + 540 + ], + "type": "text", + "content": "[11] Garrett Bingham, William Macke, and Risto Miikkulainen. Evolutionary optimization of deep learning activation functions. In Genetic and Evolutionary Computation Conference, 2020. 3, 4, 1, 2" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 61, + 543, + 295, + 563 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 61, + 543, + 295, + 563 + ], + "spans": [ + { + "bbox": [ + 61, + 543, + 295, + 563 + ], + "type": "text", + "content": "[12] Mark Buchanan. A natural bias for simplicity. Nature Physics, 2018. 8" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 61, + 566, + 295, + 597 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 61, + 566, + 295, + 597 + ], + "spans": [ + { + "bbox": [ + 61, + 566, + 295, + 597 + ], + "type": "text", + "content": "[13] Irit Chelly, Shahaf E Finder, Shira Ifergane, and Oren Freifeld. Trainable highly-expressive activation functions. arXiv preprint arXiv:2407.07564, 2024. 3, 1, 2" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 61, + 600, + 295, + 632 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 61, + 600, + 295, + 632 + ], + "spans": [ + { + "bbox": [ + 61, + 600, + 295, + 632 + ], + "type": "text", + "content": "[14] Nadav Cohen and Amnon Shashua. Inductive bias of deep convolutional networks through pooling geometry. arXiv preprint arXiv:1605.06743, 2016. 1" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 61, + 634, + 295, + 666 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 61, + 634, + 295, + 666 + ], + "spans": [ + { + "bbox": [ + 61, + 634, + 295, + 666 + ], + "type": "text", + "content": "[15] Giacomo De Palma, Bobak Kiani, and Seth Lloyd. Random deep neural networks are biased towards simple functions. NeurIPS, 2019. 1, 5" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 61, + 668, + 295, + 689 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 61, + 668, + 295, + 689 + ], + "spans": [ + { + "bbox": [ + 61, + 668, + 295, + 689 + ], + "type": "text", + "content": "[16] David Deutsch. The beginning of infinity: Explanations that transform the world. Penguin, 2011. 8" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 61, + 691, + 295, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 61, + 691, + 295, + 713 + ], + "spans": [ + { + "bbox": [ + 61, + 691, + 295, + 713 + ], + "type": "text", + "content": "[17] Benoit Dherin, Michael Munn, Mihaela Rosca, and David Barrett. Why neural networks find simple solutions: The" + } + ] + } + ], + "index": 17 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 320, + 73, + 553, + 713 + ], + "type": "list", + "angle": 0, + "index": 35, + "blocks": [ + { + "bbox": [ + 339, + 73, + 553, + 95 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 339, + 73, + 553, + 95 + ], + "spans": [ + { + "bbox": [ + 339, + 73, + 553, + 95 + ], + "type": "text", + "content": "many regularizers of geometric complexity. NeurIPS, 35, 2022. 1, 4, 5" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 320, + 96, + 553, + 129 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 320, + 96, + 553, + 129 + ], + "spans": [ + { + "bbox": [ + 320, + 96, + 553, + 129 + ], + "type": "text", + "content": "[18] Kamaludin Dingle, Chico Q Camargo, and Ard A Louis. Input-output maps are strongly biased towards simple outputs. Nature communications, 2018. 1, 8" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 320, + 130, + 553, + 152 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 320, + 130, + 553, + 152 + ], + "spans": [ + { + "bbox": [ + 320, + 130, + 553, + 152 + ], + "type": "text", + "content": "[19] Pedro Domingos. The role of occam's razor in knowledge discovery. Data mining and knowledge discovery, 1999. 1" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 320, + 153, + 553, + 196 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 320, + 153, + 553, + 196 + ], + "spans": [ + { + "bbox": [ + 320, + 153, + 553, + 196 + ], + "type": "text", + "content": "[20] Marius Dragoi, Florin Gogianu, and Elena Burceanu. Closing the gap on tabular data with fourier and implicit categorical features. Submission to ICLR (available on Open-Review), 2024. 5, 1" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 320, + 198, + 553, + 239 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 320, + 198, + 553, + 239 + ], + "spans": [ + { + "bbox": [ + 320, + 198, + 553, + 239 + ], + "type": "text", + "content": "[21] Shiv Ram Dubey, Satish Kumar Singh, and Bidyut Baran Chaudhuri. Activation functions in deep learning: A comprehensive survey and benchmark. Neurocomputing, 2022. 1" + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 320, + 243, + 553, + 297 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 320, + 243, + 553, + 297 + ], + "spans": [ + { + "bbox": [ + 320, + 243, + 553, + 297 + ], + "type": "text", + "content": "[22] Stanislas Ducotterd, Alexis Goujon, Pakshal Bohra, Dimitris Perdios, Sebastian Neumayer, and Michael Unser. Improving lipschitz-constrained neural networks by learning activation functions. Journal of Machine Learning Research, 2024. 3, 1, 2" + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 320, + 299, + 553, + 331 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 320, + 299, + 553, + 331 + ], + "spans": [ + { + "bbox": [ + 320, + 299, + 553, + 331 + ], + "type": "text", + "content": "[23] Daniel C Elton. Applying deutsch's concept of good explanations to artificial intelligence and neuroscience—an initial exploration. Cognitive Systems Research, 2021. 8" + } + ] + } + ], + "index": 25 + }, + { + "bbox": [ + 320, + 333, + 553, + 387 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 320, + 333, + 553, + 387 + ], + "spans": [ + { + "bbox": [ + 320, + 333, + 553, + 387 + ], + "type": "text", + "content": "[24] Jesse Farebrother, Jordi Orbay, Quan Vuong, Adrien Ali Taiga, Yevgen Chebotar, Ted Xiao, Alex Irpan, Sergey Levine, Pablo Samuel Castro, Aleksandra Faust, et al. Stop regressing: Training value functions via classification for scalable deep rl. arXiv preprint arXiv:2403.03950, 2024. 4" + } + ] + } + ], + "index": 26 + }, + { + "bbox": [ + 320, + 388, + 553, + 409 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 320, + 388, + 553, + 409 + ], + "spans": [ + { + "bbox": [ + 320, + 388, + 553, + 409 + ], + "type": "text", + "content": "[25] James Fiedler. Simple modifications to improve tabular neural networks. CoRR, abs/2108.03214, 2021. 5" + } + ] + } + ], + "index": 27 + }, + { + "bbox": [ + 320, + 411, + 553, + 444 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 320, + 411, + 553, + 444 + ], + "spans": [ + { + "bbox": [ + 320, + 411, + 553, + 444 + ], + "type": "text", + "content": "[26] Sara Fridovich-Keil, Raphael Gontijo Lopes, and Rebecca Roelofs. Spectral bias in practice: The role of function frequency in generalization. NeurIPS, 2022. 2" + } + ] + } + ], + "index": 28 + }, + { + "bbox": [ + 320, + 445, + 553, + 477 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 320, + 445, + 553, + 477 + ], + "spans": [ + { + "bbox": [ + 320, + 445, + 553, + 477 + ], + "type": "text", + "content": "[27] Gallant. There exists a neural network that does not make avoidable mistakes. In IEEE International Conference on Neural Networks. IEEE, 1988. 1" + } + ] + } + ], + "index": 29 + }, + { + "bbox": [ + 320, + 479, + 553, + 533 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 320, + 479, + 553, + 533 + ], + "spans": [ + { + "bbox": [ + 320, + 479, + 553, + 533 + ], + "type": "text", + "content": "[28] Isabel O Gallegos, Ryan A Rossi, Joe Barrow, Md Mehrab Tanjim, Sungchul Kim, Franck Dernoncourt, Tong Yu, Ruiyi Zhang, and Nesreen K Ahmed. Bias and fairness in large language models: A survey. Computational Linguistics, 2024. 1" + } + ] + } + ], + "index": 30 + }, + { + "bbox": [ + 320, + 535, + 553, + 578 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 320, + 535, + 553, + 578 + ], + "spans": [ + { + "bbox": [ + 320, + 535, + 553, + 578 + ], + "type": "text", + "content": "[29] Robert Geirhos, Jorn-Henrik Jacobsen, Claudio Michaelis, Richard Zemel, Wieland Brendel, Matthias Bethge, and Felix A Wichmann. Shortcut learning in deep neural networks. Nature Machine Intelligence, 2020. 1" + } + ] + } + ], + "index": 31 + }, + { + "bbox": [ + 320, + 580, + 553, + 624 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 320, + 580, + 553, + 624 + ], + "spans": [ + { + "bbox": [ + 320, + 580, + 553, + 624 + ], + "type": "text", + "content": "[30] Florin Gogianu, Tudor Berariu, Mihaela C Rosca, Claudia Clopath, Lucian Busoniu, and Razvan Pascanu. Spectral normalisation for deep reinforcement learning: an optimisation perspective. In ICML, 2021. 11" + } + ] + } + ], + "index": 32 + }, + { + "bbox": [ + 320, + 625, + 553, + 677 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 320, + 625, + 553, + 677 + ], + "spans": [ + { + "bbox": [ + 320, + 625, + 553, + 677 + ], + "type": "text", + "content": "[31] Micah Goldblum, Marc Finzi, Keefer Rowan, and Andrew Gordon Wilson. The no free lunch theorem, kolmogorov complexity, and the role of inductive biases in machine learning. arXiv preprint arXiv:2304.05366, 2023. 1, 8" + } + ] + } + ], + "index": 33 + }, + { + "bbox": [ + 320, + 680, + 553, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 320, + 680, + 553, + 713 + ], + "spans": [ + { + "bbox": [ + 320, + 680, + 553, + 713 + ], + "type": "text", + "content": "[32] Yury Gorishniy, Ivan Rubachev, and Artem Babenko. On embeddings for numerical features in tabular deep learning. NeurIPS, 2022. 5, 1, 9" + } + ] + } + ], + "index": 34 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 299, + 749, + 310, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 299, + 749, + 310, + 757 + ], + "spans": [ + { + "bbox": [ + 299, + 749, + 310, + 757 + ], + "type": "text", + "content": "87" + } + ] + } + ], + "index": 36 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 8 + }, + { + "para_blocks": [ + { + "bbox": [ + 61, + 73, + 296, + 712 + ], + "type": "list", + "angle": 0, + "index": 16, + "blocks": [ + { + "bbox": [ + 61, + 73, + 296, + 106 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 61, + 73, + 296, + 106 + ], + "spans": [ + { + "bbox": [ + 61, + 73, + 296, + 106 + ], + "type": "text", + "content": "[33] Anirudh Goyal and Yoshua Bengio. Inductive biases for deep learning of higher-level cognition. Proceedings of the Royal Society A, 2022. 1" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 61, + 107, + 296, + 139 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 61, + 107, + 296, + 139 + ], + "spans": [ + { + "bbox": [ + 61, + 107, + 296, + 139 + ], + "type": "text", + "content": "[34] Léo Grinsztajn. Tabular data learning benchmark. https://github.com/LéoGrin/tabular-benchmark, 2022.5,6, 9,10,14" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 62, + 141, + 294, + 174 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 141, + 294, + 174 + ], + "spans": [ + { + "bbox": [ + 62, + 141, + 294, + 174 + ], + "type": "text", + "content": "[35] Léo Grinsztajn, Edouard Oyallon, and Gael Varoquaux. Why do tree-based models still outperform deep learning on typical tabular data? NeurIPS, 2022. 4, 5, 1" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 62, + 175, + 295, + 197 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 175, + 295, + 197 + ], + "spans": [ + { + "bbox": [ + 62, + 175, + 295, + 197 + ], + "type": "text", + "content": "[36] Andrey Gromov. Grokking modular arithmetic. arXiv preprint arXiv:2301.02679, 2023. 7" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 62, + 198, + 295, + 231 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 198, + 295, + 231 + ], + "spans": [ + { + "bbox": [ + 62, + 198, + 295, + 231 + ], + "type": "text", + "content": "[37] Michael Hahn, Dan Jurafsky, and Richard Futrell. Sensitivity as a complexity measure for sequence classification tasks. Transactions of the ACL, 2021. 1" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 62, + 232, + 295, + 274 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 232, + 295, + 274 + ], + "spans": [ + { + "bbox": [ + 62, + 232, + 295, + 274 + ], + "type": "text", + "content": "[38] Hrayr Harutyunyan, Rafayel Darbinyan, Samvel Karapetyan, and Hrant Khachatrian. In-context learning in presence of spurious correlations. arXiv preprint arXiv:2410.03140, 2024. 1" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 62, + 277, + 295, + 308 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 277, + 295, + 308 + ], + "spans": [ + { + "bbox": [ + 62, + 277, + 295, + 308 + ], + "type": "text", + "content": "[39] Dan Hendrycks and Kevin Gimpel. Gaussian error linear units (GeLUs). arXiv preprint arXiv:1606.08415, 2016. 2, 1" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 62, + 311, + 295, + 353 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 311, + 295, + 353 + ], + "spans": [ + { + "bbox": [ + 62, + 311, + 295, + 353 + ], + "type": "text", + "content": "[40] Katherine L Hermann and Andrew K Lampinen. What shapes feature representations? exploring datasets, architectures, and training. arXiv preprint arXiv:2006.12433, 2020. 1" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 62, + 356, + 295, + 409 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 356, + 295, + 409 + ], + "spans": [ + { + "bbox": [ + 62, + 356, + 295, + 409 + ], + "type": "text", + "content": "[41] Hossein Hosseini, Baicen Xiao, Mayoore Jaiswal, and Radha Poovendran. Assessing shape bias property of convolutional neural networks. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition Workshops, 2018. 1" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 62, + 411, + 295, + 444 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 411, + 295, + 444 + ], + "spans": [ + { + "bbox": [ + 62, + 411, + 295, + 444 + ], + "type": "text", + "content": "[42] Like Hui and Mikhail Belkin. Evaluation of neural architectures trained with square loss vs cross-entropy in classification tasks. arXiv preprint arXiv:2006.07322, 2020.1" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 62, + 446, + 295, + 487 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 446, + 295, + 487 + ], + "spans": [ + { + "bbox": [ + 62, + 446, + 295, + 487 + ], + "type": "text", + "content": "[43] Ehsan Imani, Kai Luedemann, Sam Scholnick-Hughes, Esraa Elelimy, and Martha White. Investigating the histogram loss in regression. arXiv preprint arXiv:2402.13425, 2024. 4" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 62, + 490, + 295, + 544 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 490, + 295, + 544 + ], + "spans": [ + { + "bbox": [ + 62, + 490, + 295, + 544 + ], + "type": "text", + "content": "[44] Ameya D Jagtap and George Em Karniadakis. How important are activation functions in regression and classification? a survey, performance comparison, and future directions. Journal of Machine Learning for Modeling and Computing, 4(1), 2023. 1" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 62, + 546, + 295, + 589 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 546, + 295, + 589 + ], + "spans": [ + { + "bbox": [ + 62, + 546, + 295, + 589 + ], + "type": "text", + "content": "[45] Ameya D Jagtap, Kenji Kawaguchi, and George Em Karniadakis. Adaptive activation functions accelerate convergence in deep and physics-informed neural networks. Journal of Computational Physics, 2020. 3, 5, 1, 2" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 62, + 591, + 295, + 633 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 591, + 295, + 633 + ], + "spans": [ + { + "bbox": [ + 62, + 591, + 295, + 633 + ], + "type": "text", + "content": "[46] Anchit Jain, Rozhin Nobahari, Aristide Baratin, and Stefano Sarao Mannelli. Bias in motion: Theoretical insights into the dynamics of bias in sgd training. arXiv preprint arXiv:2405.18296, 2024.7" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 62, + 636, + 295, + 667 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 636, + 295, + 667 + ], + "spans": [ + { + "bbox": [ + 62, + 636, + 295, + 667 + ], + "type": "text", + "content": "[47] Liangze Jiang and Damien Teney. OOD-chameleon: Is algorithm selection for ood generalization learnable? arXiv preprint arXiv:2410.02735, 2024. 1" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 62, + 670, + 295, + 712 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 670, + 295, + 712 + ], + "spans": [ + { + "bbox": [ + 62, + 670, + 295, + 712 + ], + "type": "text", + "content": "[48] Dimitris Kalimeris, Gal Kaplun, Preetum Nakkiran, Benjamin Edelman, Tristan Yang, Boaz Barak, and Haofeng Zhang. SGD on neural networks learns functions of increasing complexity. NeurIPS, 2019. 5, 1" + } + ] + } + ], + "index": 15 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 321, + 73, + 555, + 712 + ], + "type": "list", + "angle": 0, + "index": 33, + "blocks": [ + { + "bbox": [ + 321, + 73, + 553, + 106 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 321, + 73, + 553, + 106 + ], + "spans": [ + { + "bbox": [ + 321, + 73, + 553, + 106 + ], + "type": "text", + "content": "[49] Alex Krizhevsky and Geoffrey Hinton. Learning multiple layers of features from tiny images. Technical report, University of Toronto, 2009. 3" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 321, + 108, + 553, + 140 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 321, + 108, + 553, + 140 + ], + "spans": [ + { + "bbox": [ + 321, + 108, + 553, + 140 + ], + "type": "text", + "content": "[50] Jan Kukačka, Vladimir Golkov, and Daniel Cremers. Regularization for deep learning: A taxonomy. arXiv preprint arXiv:1710.10686, 2017. 1" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 321, + 143, + 553, + 185 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 321, + 143, + 553, + 185 + ], + "spans": [ + { + "bbox": [ + 321, + 143, + 553, + 185 + ], + "type": "text", + "content": "[51] Tanishq Kumar, Blake Bordelon, Samuel J Gershman, and Cengiz Pehlevan. Grokking as the transition from lazy to rich training dynamics. arXiv preprint arXiv:2310.06110, 2023. 7" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 321, + 188, + 553, + 230 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 321, + 188, + 553, + 230 + ], + "spans": [ + { + "bbox": [ + 321, + 188, + 553, + 230 + ], + "type": "text", + "content": "[52] Yann LeCun, Léon Bottou, Yoshua Bengio, and Patrick Haffner. Gradient-based learning applied to document recognition. Proceedings of the IEEE, 86(11):2278-2324, 1998. 3" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 321, + 234, + 553, + 266 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 321, + 234, + 553, + 266 + ], + "spans": [ + { + "bbox": [ + 321, + 234, + 553, + 266 + ], + "type": "text", + "content": "[53] Alexander Li and Deepak Pathak. Functional regularization for reinforcement learning via learned fourier features. NeurIPS, 2021. 1" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 321, + 269, + 553, + 300 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 321, + 269, + 553, + 300 + ], + "spans": [ + { + "bbox": [ + 321, + 269, + 553, + 300 + ], + "type": "text", + "content": "[54] Hao Li, Zheng Xu, Gavin Taylor, Christoph Studer, and Tom Goldstein. Visualizing the loss landscape of neural nets. NeurIPS, 2018. 5" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 321, + 304, + 553, + 324 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 321, + 304, + 553, + 324 + ], + "spans": [ + { + "bbox": [ + 321, + 304, + 553, + 324 + ], + "type": "text", + "content": "[55] Xuan Li, Yun Wang, and Bo Li. Tree-regularized tabular embeddings. arXiv preprint arXiv:2403.00963, 2024. 1" + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 321, + 327, + 555, + 358 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 321, + 327, + 555, + 358 + ], + "spans": [ + { + "bbox": [ + 321, + 327, + 555, + 358 + ], + "type": "text", + "content": "[56] Henry W Lin, Max Tegmark, and David Rolnick. Why does deep and cheap learning work so well? Journal of Statistical Physics, 2017. 8" + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 321, + 361, + 553, + 404 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 321, + 361, + 553, + 404 + ], + "spans": [ + { + "bbox": [ + 321, + 361, + 553, + 404 + ], + "type": "text", + "content": "[57] Ziqi Liu, Wei Cai, and Zhi-Qin John Xu. Multi-scale deep neural network (mscalednn) for solving poisson-boltzmann equation in complex domains. arXiv preprint arXiv:2007.11207, 2020. 1" + } + ] + } + ], + "index": 25 + }, + { + "bbox": [ + 321, + 407, + 553, + 449 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 321, + 407, + 553, + 449 + ], + "spans": [ + { + "bbox": [ + 321, + 407, + 553, + 449 + ], + "type": "text", + "content": "[58] Ziming Liu, Ouail Kitouni, Niklas S Nolte, Eric Michaud, Max Tegmark, and Mike Williams. Towards understanding grokking: An effective theory of representation learning. NeurIPS, 2022. 7" + } + ] + } + ], + "index": 26 + }, + { + "bbox": [ + 321, + 453, + 553, + 495 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 321, + 453, + 553, + 495 + ], + "spans": [ + { + "bbox": [ + 321, + 453, + 553, + 495 + ], + "type": "text", + "content": "[59] Ziming Liu, Yixuan Wang, Sachin Vaidya, Fabian Ruehle, James Halverson, Marin Soljačić, Thomas Y Hou, and Max Tegmark. Kan: Kolmogorov-arnold networks. arXiv preprint arXiv:2404.19756, 2024. 2" + } + ] + } + ], + "index": 27 + }, + { + "bbox": [ + 321, + 498, + 553, + 530 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 321, + 498, + 553, + 530 + ], + "spans": [ + { + "bbox": [ + 321, + 498, + 553, + 530 + ], + "type": "text", + "content": "[60] Kaifeng Lyu, Zhiyuan Li, Runzhe Wang, and Sanjeev Arora. Gradient descent on two-layer nets: Margin maximization and simplicity bias. NeurIPS, 2021. 1" + } + ] + } + ], + "index": 28 + }, + { + "bbox": [ + 321, + 533, + 553, + 564 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 321, + 533, + 553, + 564 + ], + "spans": [ + { + "bbox": [ + 321, + 533, + 553, + 564 + ], + "type": "text", + "content": "[61] Andrew L Maas, Awni Y Hannun, Andrew Y Ng, et al. Rectifier nonlinearities improve neural network acoustic models. In ICML, 2013. 1" + } + ] + } + ], + "index": 29 + }, + { + "bbox": [ + 321, + 567, + 553, + 610 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 321, + 567, + 553, + 610 + ], + "spans": [ + { + "bbox": [ + 321, + 567, + 553, + 610 + ], + "type": "text", + "content": "[62] Augustine N Mavor-Parker, Matthew J Sargent, Caswell Barry, Lewis Griffin, and Clare Lyle. Frequency and generalisation of periodic activation functions in reinforcement learning. arXiv preprint arXiv:2407.06756, 2024. 1" + } + ] + } + ], + "index": 30 + }, + { + "bbox": [ + 321, + 613, + 553, + 655 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 321, + 613, + 553, + 655 + ], + "spans": [ + { + "bbox": [ + 321, + 613, + 553, + 655 + ], + "type": "text", + "content": "[63] Duncan McElfresh, Sujay Khandagale, Jonathan Valverde, Vishak Prasad C, Ganesh Ramakrishnan, Micah Goldblum, and Colin White. When do neural nets outperform boosted trees on tabular data? NeurIPS, 2024. 5" + } + ] + } + ], + "index": 31 + }, + { + "bbox": [ + 321, + 658, + 553, + 712 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 321, + 658, + 553, + 712 + ], + "spans": [ + { + "bbox": [ + 321, + 658, + 553, + 712 + ], + "type": "text", + "content": "[64] Chris Mingard, Joar Skalse, Guillermo Valle-Pérez, David Martínez-Rubio, Vladimir Mikulik, and Ard A Louis. Neural networks are a priori biased towards boolean functions with low entropy. arXiv preprint arXiv:1909.11522, 2019. 1, 5" + } + ] + } + ], + "index": 32 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 300, + 750, + 310, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 300, + 750, + 310, + 757 + ], + "spans": [ + { + "bbox": [ + 300, + 750, + 310, + 757 + ], + "type": "text", + "content": "88" + } + ] + } + ], + "index": 34 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 9 + }, + { + "para_blocks": [ + { + "bbox": [ + 61, + 72, + 295, + 712 + ], + "type": "list", + "angle": 0, + "index": 16, + "blocks": [ + { + "bbox": [ + 61, + 72, + 294, + 106 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 61, + 72, + 294, + 106 + ], + "spans": [ + { + "bbox": [ + 61, + 72, + 294, + 106 + ], + "type": "text", + "content": "[65] Chris Mingard, Guillermo Valle-Pérez, Joar Skalse, and Ard A Louis. Is SGD a bayesian sampler? well, almost. Journal of Machine Learning Research, 2021. 1, 5" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 61, + 107, + 295, + 140 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 61, + 107, + 295, + 140 + ], + "spans": [ + { + "bbox": [ + 61, + 107, + 295, + 140 + ], + "type": "text", + "content": "[66] Chris Mingard, Henry Rees, Guillermo Valle-Pérez, and Ard A Louis. Do deep neural networks have an inbuilt ocean's razor? arXiv preprint arXiv:2304.06670, 2023. 1" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 62, + 143, + 294, + 174 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 143, + 294, + 174 + ], + "spans": [ + { + "bbox": [ + 62, + 143, + 294, + 174 + ], + "type": "text", + "content": "[67] Tom M Mitchell. The need for biases in learning generalizations. Rutgers University CS tech report CBM-TR-117, 1980. 1" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 62, + 177, + 295, + 232 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 177, + 295, + 232 + ], + "spans": [ + { + "bbox": [ + 62, + 177, + 295, + 232 + ], + "type": "text", + "content": "[68] Yuval Netzer, Tao Wang, Adam Coates, Alessandro Bissacco, Bo Wu, and Andrew Y Ng. Reading digits in natural images with unsupervised feature learning. NIPS Workshop on Deep Learning and Unsupervised Feature Learning, 2011. 3" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 62, + 234, + 295, + 277 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 234, + 295, + 277 + ], + "spans": [ + { + "bbox": [ + 62, + 234, + 295, + 277 + ], + "type": "text", + "content": "[69] Behnam Neyshabur, Ryota Tomioka, and Nathan Srebro. In search of the real inductive bias: On the role of implicit regularization in deep learning. arXiv preprint arXiv:1412.6614, 2014. 1" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 62, + 279, + 294, + 322 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 279, + 294, + 322 + ], + "spans": [ + { + "bbox": [ + 62, + 279, + 294, + 322 + ], + "type": "text", + "content": "[70] Mohammad Pezeshki, Oumar Kaba, Yoshua Bengio, Aaron C Courville, Doina Precup, and Guillaume Lajoie. Gradient starvation: A learning proclivity in neural networks. NeurIPS, 2021. 1" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 62, + 325, + 294, + 369 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 325, + 294, + 369 + ], + "spans": [ + { + "bbox": [ + 62, + 325, + 294, + 369 + ], + "type": "text", + "content": "[71] Tomaso Poggio, Kenji Kawaguchi, Qianli Liao, Brando Miranda, Lorenzo Rosasco, Xavier Boix, Jack Hiday, and Hrushikesh Mhaskar. Theory of deep learning III: the nonoverfitting puzzle. CBMM Memo, 2018. 1" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 62, + 370, + 294, + 392 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 370, + 294, + 392 + ], + "spans": [ + { + "bbox": [ + 62, + 370, + 294, + 392 + ], + "type": "text", + "content": "[72] Karl Popper. \"7. Simplicity\". The logic of scientific discovery. Routledge, 1959. 8" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 62, + 395, + 294, + 437 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 395, + 294, + 437 + ], + "spans": [ + { + "bbox": [ + 62, + 395, + 294, + 437 + ], + "type": "text", + "content": "[73] Alethea Power, Yuri Burda, Harri Edwards, Igor Babuschkin, and Vedant Misra. Grokking: Generalization beyond overfitting on small algorithmic datasets. arXiv preprint arXiv:2201.02177, 2022. 7, 8, 12" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 62, + 440, + 294, + 473 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 440, + 294, + 473 + ], + "spans": [ + { + "bbox": [ + 62, + 440, + 294, + 473 + ], + "type": "text", + "content": "[74] Aahlad Manas Puli, Lily Zhang, Yoav Wald, and Rajesh Ranganath. Don't blame dataset shift! shortcut learning due to gradients and cross entropy. NeurIPS, 2023. 1" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 62, + 475, + 294, + 517 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 475, + 294, + 517 + ], + "spans": [ + { + "bbox": [ + 62, + 475, + 294, + 517 + ], + "type": "text", + "content": "[75] Nasim Rahaman, Aristide Baratin, Devansh Arpit, Felix Draxler, Min Lin, Fred Hamprecht, Yoshua Bengio, and Aaron Courville. On the spectral bias of neural networks. In ICML. PMLR, 2019. 5, 1" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 62, + 520, + 294, + 552 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 520, + 294, + 552 + ], + "spans": [ + { + "bbox": [ + 62, + 520, + 294, + 552 + ], + "type": "text", + "content": "[76] Prajit Ramachandran, Barret Zoph, and Quoc V Le. Swish: a self-gated activation function. arXiv preprint arXiv:1710.05941, 2017. 1" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 62, + 555, + 294, + 587 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 555, + 294, + 587 + ], + "spans": [ + { + "bbox": [ + 62, + 555, + 294, + 587 + ], + "type": "text", + "content": "[77] Sameera Ramasinghe and Simon Lucey. Beyond periodicity: Towards a unifying framework for activations in coordinate-MLPs. In ECCV. Springer, 2022. 1" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 62, + 590, + 294, + 621 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 590, + 294, + 621 + ], + "spans": [ + { + "bbox": [ + 62, + 590, + 294, + 621 + ], + "type": "text", + "content": "[78] Sameera Ramasinghe, Lachlan E MacDonald, and Simon Lucey. On the frequency-bias of coordinate-MLPs. NeurIPS, 2022. 1" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 62, + 624, + 294, + 667 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 624, + 294, + 667 + ], + "spans": [ + { + "bbox": [ + 62, + 624, + 294, + 667 + ], + "type": "text", + "content": "[79] Mihaela Rosca, Theophane Weber, Arthur Gretton, and Shakir Mohamed. A case for new neural network smoothness constraints. I Can't Believe It's Not Better (ICBINB) Workshop at NeurIPS, 2020. 11" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 62, + 670, + 294, + 712 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 670, + 294, + 712 + ], + "spans": [ + { + "bbox": [ + 62, + 670, + 294, + 712 + ], + "type": "text", + "content": "[80] Vishwanath Saragadam, Daniel LeJeune, Jasper Tan, Guha Balakrishnan, Ashok Veeraraghavan, and Richard G Baraniuk. Wire: Wavelet implicit neural representations. In CVPR, 2023. 4, 1" + } + ] + } + ], + "index": 15 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 320, + 73, + 553, + 713 + ], + "type": "list", + "angle": 0, + "index": 32, + "blocks": [ + { + "bbox": [ + 320, + 73, + 553, + 126 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 320, + 73, + 553, + 126 + ], + "spans": [ + { + "bbox": [ + 320, + 73, + 553, + 126 + ], + "type": "text", + "content": "[81] Simone Scardapane, Michele Scarpiniti, Danilo Comminiello, and Aurelio Uncini. Learning activation functions from data using cubic spline interpolation. Neural Advances in Processing Nonlinear Dynamic Signals, 2019. 2" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 320, + 129, + 553, + 172 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 320, + 129, + 553, + 172 + ], + "spans": [ + { + "bbox": [ + 320, + 129, + 553, + 172 + ], + "type": "text", + "content": "[82] Simone Scardapane, Steven Van Vaerenbergh, Simone Toto, and Aurelio Uncini. Kafnets: Kernel-based nonparametric activation functions for neural networks. Neural Networks, 2019. 2, 3, 1" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 320, + 175, + 553, + 206 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 320, + 175, + 553, + 206 + ], + "spans": [ + { + "bbox": [ + 320, + 175, + 553, + 206 + ], + "type": "text", + "content": "[83] Jürgen Schmidhuber. Discovering neural nets with low kolmogorov complexity and high generalization capability. Neural Networks, 1997. 1" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 320, + 209, + 553, + 251 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 320, + 209, + 553, + 251 + ], + "spans": [ + { + "bbox": [ + 320, + 209, + 553, + 251 + ], + "type": "text", + "content": "[84] Luca Scimeca, Seong Joon Oh, Sanghyuk Chun, Michael Poli, and Sangdoo Yun. Which shortcut cues will DNNs choose? a study from the parameter-space perspective. arXiv preprint arXiv:2110.03095, 2021. 5" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 320, + 254, + 553, + 285 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 320, + 254, + 553, + 285 + ], + "spans": [ + { + "bbox": [ + 320, + 254, + 553, + 285 + ], + "type": "text", + "content": "[85] Harshay Shah, Kaustav Tamuly, Aditi Raghunathan, Prateek Jain, and Praneeth Netrapalli. The pitfalls of simplicity bias in neural networks. NeurIPS, 2020. 2, 5, 6, 11" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 320, + 287, + 553, + 319 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 320, + 287, + 553, + 319 + ], + "spans": [ + { + "bbox": [ + 320, + 287, + 553, + 319 + ], + "type": "text", + "content": "[86] Kexuan Shi, Xingyu Zhou, and Shuhang Gu. Improved implicit neural representation with fourier reparameterized training. In CVPR, 2024. 1" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 320, + 321, + 553, + 354 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 320, + 321, + 553, + 354 + ], + "spans": [ + { + "bbox": [ + 320, + 321, + 553, + 354 + ], + "type": "text", + "content": "[87] Prasann Singhal, Tanya Goyal, Jiacheng Xu, and Greg Durrett. A long way to go: Investigating length correlations in rlhf. arXiv preprint arXiv:2310.03716, 2023. 1" + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 320, + 355, + 553, + 397 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 320, + 355, + 553, + 397 + ], + "spans": [ + { + "bbox": [ + 320, + 355, + 553, + 397 + ], + "type": "text", + "content": "[88] Vincent Sitzmann, Julien Martel, Alexander Bergman, David Lindell, and Gordon Wetzstein. Implicit neural representations with periodic activation functions. NeurIPS, 2020. 1" + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 320, + 399, + 553, + 443 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 320, + 399, + 553, + 443 + ], + "spans": [ + { + "bbox": [ + 320, + 399, + 553, + 443 + ], + "type": "text", + "content": "[89] Daniel Soudry, Elad Hoffer, Mor Shpigel Nacson, Suriya Gunasekar, and Nathan Srebro. The implicit bias of gradient descent on separable data. The Journal of Machine Learning Research, 2018. 5" + } + ] + } + ], + "index": 25 + }, + { + "bbox": [ + 320, + 445, + 553, + 487 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 320, + 445, + 553, + 487 + ], + "spans": [ + { + "bbox": [ + 320, + 445, + 553, + 487 + ], + "type": "text", + "content": "[90] Lawrence Stewart, Francis Bach, Quentin Berthet, and Jean-Philippe Vert. Regression as classification: Influence of task formulation on neural network features. In ICML. PMLR, 2023. 4" + } + ] + } + ], + "index": 26 + }, + { + "bbox": [ + 320, + 490, + 553, + 544 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 320, + 490, + 553, + 544 + ], + "spans": [ + { + "bbox": [ + 320, + 490, + 553, + 544 + ], + "type": "text", + "content": "[91] Leon René Sutfeld, Flemming Brieger, Holger Finger, Sonja Fülhase, and Gordon Pipa. Adaptive blending units: Trainable activation functions for deep neural networks. In Intelligent Computing: Proceedings of the Computing Conference. Springer, 2020. 3, 1, 2" + } + ] + } + ], + "index": 27 + }, + { + "bbox": [ + 320, + 545, + 553, + 588 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 320, + 545, + 553, + 588 + ], + "spans": [ + { + "bbox": [ + 320, + 545, + 553, + 588 + ], + "type": "text", + "content": "[92] Remi Tachet, Mohammad Pezeshki, Samira Shabanian, Aaron Courville, and Yoshua Bengio. On the learning dynamics of deep neural networks. arXiv preprint arXiv:1809.06848, 2018. 1" + } + ] + } + ], + "index": 28 + }, + { + "bbox": [ + 320, + 590, + 553, + 644 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 320, + 590, + 553, + 644 + ], + "spans": [ + { + "bbox": [ + 320, + 590, + 553, + 644 + ], + "type": "text", + "content": "[93] Damien Teney, Ehsan Abbasnejad, Simon Lucey, and Anton van den Hengel. Evading the simplicity bias: Training a diverse set of models discovers solutions with superior ood generalization. arXiv preprint arXiv:2105.05612, 2021. 1, 2, 6, 11" + } + ] + } + ], + "index": 29 + }, + { + "bbox": [ + 320, + 647, + 553, + 689 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 320, + 647, + 553, + 689 + ], + "spans": [ + { + "bbox": [ + 320, + 647, + 553, + 689 + ], + "type": "text", + "content": "[94] Damien Teney, Maxime Peyrard, and Ehsan Abbasnejad. Predicting is not understanding: Recognizing and addressing underspecification in machine learning. In ECCV. Springer, 2022. 6, 1" + } + ] + } + ], + "index": 30 + }, + { + "bbox": [ + 320, + 692, + 553, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 320, + 692, + 553, + 713 + ], + "spans": [ + { + "bbox": [ + 320, + 692, + 553, + 713 + ], + "type": "text", + "content": "[95] Damien Teney, Yong Lin, Seong Joon Oh, and Ehsan Abbasnejad. ID and OOD performance are sometimes in" + } + ] + } + ], + "index": 31 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 299, + 749, + 310, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 299, + 749, + 310, + 757 + ], + "spans": [ + { + "bbox": [ + 299, + 749, + 310, + 757 + ], + "type": "text", + "content": "89" + } + ] + } + ], + "index": 33 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 10 + }, + { + "para_blocks": [ + { + "bbox": [ + 56, + 73, + 295, + 712 + ], + "type": "list", + "angle": 0, + "index": 16, + "blocks": [ + { + "bbox": [ + 81, + 73, + 294, + 94 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 81, + 73, + 294, + 94 + ], + "spans": [ + { + "bbox": [ + 81, + 73, + 294, + 94 + ], + "type": "text", + "content": "versely correlated on real-world datasets. NeurIPS, 2024. 1, 7" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 61, + 95, + 295, + 139 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 61, + 95, + 295, + 139 + ], + "spans": [ + { + "bbox": [ + 61, + 95, + 295, + 139 + ], + "type": "text", + "content": "[96] Damien Teney, Armand Mihai Nicoliciou, Valentin Hartmann, and Ehsan Abbasnejad. Neural redshift: Random networks are not random functions. In CVPR, 2024. 1, 2, 3, 4, 5, 6, 8, 11, 14" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 61, + 140, + 294, + 172 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 61, + 140, + 294, + 172 + ], + "spans": [ + { + "bbox": [ + 61, + 140, + 294, + 172 + ], + "type": "text", + "content": "[97] Tijmen Tieleman. Lecture 6.5-rmsprop: Divide the gradient by a running average of its recent magnitude. COURSERA: Neural networks for machine learning, 4(2):26, 2012. 8" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 61, + 173, + 295, + 205 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 61, + 173, + 295, + 205 + ], + "spans": [ + { + "bbox": [ + 61, + 173, + 295, + 205 + ], + "type": "text", + "content": "[98] Nikita Tsoy and Nikola Konstantinov. Simplicity bias of two-layer networks beyond linearly separable data. arXiv preprint arXiv:2405.17299, 2024. 5, 7" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 62, + 206, + 295, + 250 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 206, + 295, + 250 + ], + "spans": [ + { + "bbox": [ + 62, + 206, + 295, + 250 + ], + "type": "text", + "content": "[99] Guillermo Valle-Perez, Chico Q Camargo, and Ard A Louis. Deep learning generalizes because the parameter-function map is biased towards simple functions. arXiv preprint arXiv:1805.08522, 2018. 1, 5" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 57, + 251, + 295, + 294 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 251, + 295, + 294 + ], + "spans": [ + { + "bbox": [ + 57, + 251, + 295, + 294 + ], + "type": "text", + "content": "[100] Colin White, Mahmoud Safari, Rhea Sukthanker, Binxin Ru, Thomas Elsken, Arber Zela, Debadeepta Dey, and Frank Hutter. Neural architecture search: Insights from 1000 papers. arXiv preprint arXiv:2301.08727, 2023. 2" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 57, + 294, + 294, + 326 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 294, + 294, + 326 + ], + "spans": [ + { + "bbox": [ + 57, + 294, + 294, + 326 + ], + "type": "text", + "content": "[101] David H Wolpert. The supervised learning no-free-lunch theorems. Soft computing and industry: Recent applications, 2002. 1, 8" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 57, + 327, + 295, + 370 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 327, + 295, + 370 + ], + "spans": [ + { + "bbox": [ + 57, + 327, + 295, + 370 + ], + "type": "text", + "content": "[102] Han Xiao, Kashif Rasul, and Roland Vollgraf. Fashion-mnist: A novel image dataset for benchmarking machine learning algorithms. arXiv preprint arXiv:1708.07747, 2017.3" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 57, + 371, + 294, + 426 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 371, + 294, + 426 + ], + "spans": [ + { + "bbox": [ + 57, + 371, + 294, + 426 + ], + "type": "text", + "content": "[103] Yiheng Xie, Towaki Takikawa, Shunsuke Saito, Or Litany, Shiqin Yan, Numair Khan, Federico Tombari, James Tompkin, Vincent Sitzmann, and Srinath Sridhar. Neural fields in visual computing and beyond. In Computer Graphics Forum. Wiley Online Library, 2022. 1" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 56, + 426, + 295, + 470 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 426, + 295, + 470 + ], + "spans": [ + { + "bbox": [ + 56, + 426, + 295, + 470 + ], + "type": "text", + "content": "[104] Zhi-Qin John Xu, Yaoyu Zhang, Tao Luo, Yanyang Xiao, and Zheng Ma. Frequency principle: Fourier analysis sheds light on deep neural networks. arXiv preprint arXiv:1901.06523, 2019. 5, 1" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 57, + 471, + 294, + 503 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 471, + 294, + 503 + ], + "spans": [ + { + "bbox": [ + 57, + 471, + 294, + 503 + ], + "type": "text", + "content": "[105] Zhi-Qin John Xu, Yaoyu Zhang, and Yanyang Xiao. Training behavior of deep neural network in frequency domain. In ICONIP. Springer, 2019. 5" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 57, + 504, + 294, + 546 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 504, + 294, + 546 + ], + "spans": [ + { + "bbox": [ + 57, + 504, + 294, + 546 + ], + "type": "text", + "content": "[106] Zhi-Qin John Xu, Yaoyu Zhang, and Tao Luo. Overview frequency principle/spectral bias in deep learning. Communications on Applied Mathematics and Computation, 2024. 1" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 57, + 548, + 294, + 581 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 548, + 294, + 581 + ], + "spans": [ + { + "bbox": [ + 57, + 548, + 294, + 581 + ], + "type": "text", + "content": "[107] Ge Yang, Anurag Ajay, and Pulkit Agrawal. Overcoming the spectral bias of neural value approximation. arXiv preprint arXiv:2206.04672, 2022. 1" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 57, + 582, + 294, + 613 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 582, + 294, + 613 + ], + "spans": [ + { + "bbox": [ + 57, + 582, + 294, + 613 + ], + "type": "text", + "content": "[108] Ziqian Zhong and Jacob Andreas. Algorithmic capabilities of random transformers. arXiv preprint arXiv:2410.04368, 2024. 8" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 57, + 614, + 294, + 667 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 614, + 294, + 667 + ], + "spans": [ + { + "bbox": [ + 57, + 614, + 294, + 667 + ], + "type": "text", + "content": "[109] Hattie Zhou, Arwen Bradley, Etai Littwin, Noam Razin, Omid Saremi, Josh Susskind, Samy Bengio, and Preetum Nakkiran. What algorithms can transformers learn? a study in length generalization. arXiv preprint arXiv:2310.16028, 2023. 8" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 57, + 669, + 294, + 712 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 669, + 294, + 712 + ], + "spans": [ + { + "bbox": [ + 57, + 669, + 294, + 712 + ], + "type": "text", + "content": "[110] Zhangchen Zhou, Yaoyu Zhang, and Zhi-Qin John Xu. A rationale from frequency perspective for grokking in training neural network. arXiv preprint arXiv:2405.17479, 2024. 7" + } + ] + } + ], + "index": 15 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 300, + 749, + 310, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 300, + 749, + 310, + 757 + ], + "spans": [ + { + "bbox": [ + 300, + 749, + 310, + 757 + ], + "type": "text", + "content": "90" + } + ] + } + ], + "index": 17 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 11 + } + ], + "_backend": "vlm", + "_version_name": "2.6.4" +} \ No newline at end of file diff --git a/2025/Do We Really Need Curated Malicious Data for Safety Alignment in Multi-modal Large Language Models_/7cc50b70-a1c6-4460-8b07-4c06c5509d66_content_list.json b/2025/Do We Really Need Curated Malicious Data for Safety Alignment in Multi-modal Large Language Models_/7cc50b70-a1c6-4460-8b07-4c06c5509d66_content_list.json new file mode 100644 index 0000000000000000000000000000000000000000..f72bc3a888b7568a4c92b4beaa71c9ebbbcadf8b --- /dev/null +++ b/2025/Do We Really Need Curated Malicious Data for Safety Alignment in Multi-modal Large Language Models_/7cc50b70-a1c6-4460-8b07-4c06c5509d66_content_list.json @@ -0,0 +1,1484 @@ +[ + { + "type": "text", + "text": "Do We Really Need Curated Malicious Data for Safety Alignment in Multi-modal Large Language Models?", + "text_level": 1, + "bbox": [ + 156, + 128, + 843, + 176 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Yanbo Wang, Jiyang Guan, Jian Liang $^{\\text{心}}$ , Ran He \nSchool of Artificial Intelligence, University of Chinese Academy of Sciences \nNLPR & MAIS, Institute of Automation, Chinese Academy of Sciences", + "bbox": [ + 192, + 203, + 802, + 256 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "wangyanbo2023@ia.ac.cn liangjian92@gmail.com", + "bbox": [ + 299, + 258, + 692, + 273 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Abstract", + "text_level": 1, + "bbox": [ + 246, + 308, + 326, + 325 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Multi-modal large language models (MLLMs) have made significant progress, yet their safety alignment remains limited. Typically, current open-source MLLMs rely on the alignment inherited from their language module to avoid harmful generations. However, the lack of safety measures specifically designed for multi-modal inputs creates an alignment gap, leaving MLLMs vulnerable to vision-domain attacks such as typographic manipulation. Current methods utilize a carefully designed safety dataset to enhance model defense capability, while the specific knowledge or patterns acquired from the high-quality dataset remain unclear. Through comparison experiments, we find that the alignment gap primarily arises from data distribution biases, while image content, response quality, or the contrastive behavior of the dataset makes little contribution to boosting multi-modal safety. To further investigate this and identify the key factors in improving MLLM safety, we propose finetuning MLLMs on a small set of benign instruct-following data with responses replaced by simple, clear rejection sentences. Experiments show that, without the need for labor-intensive collection of high-quality malicious data, model safety can still be significantly improved, as long as a specific fraction of rejection data exists in the finetuning set, indicating the security alignment is not lost but rather obscured during multi-modal pretraining or instruction finetuning. Simply correcting the underlying data bias could narrow the safety gap in the vision domain. Warning: This paper contains harmful images and AI-generated contents which may be offensive.", + "bbox": [ + 89, + 340, + 485, + 779 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "1. Introduction", + "text_level": 1, + "bbox": [ + 91, + 806, + 220, + 821 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "In recent years, the multi-modal large language model (MLLM) has experienced tremendous prosperity [11, 18, 45, 48, 51]. Companies and academics have proposed mul", + "bbox": [ + 89, + 832, + 483, + 878 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "tiple models with their APIs, finetuning suites, and online demos. However, although the multi-modal understanding capability of these models is advancing steadily, the safety alignment of these models attracts less attention. To be detailed, most MLLMs connect an image encoder and a pretrained large language model with a vision-language connector, which are attention structures at the beginning and then replaced by MLPs [4, 5, 9, 50]. The safety alignment focuses mainly on the large language model, where popular RLHF algorithms [2, 19, 30] such as DPO [36] and PPO [38] are used to ensure that the chat model acts humanly. Nevertheless, such a workflow uncovers an obvious disadvantage: both the pretraining and visual instruction tuning stages do not contain a specific multi-modal safety alignment procedure, and the safety defensive capability could only be inherited from the LLM module. In addition, researchers also find that finetuning models on benign data may still sacrifice model safety [34]. They all challenge the defense capability of MLLMs.", + "bbox": [ + 511, + 309, + 906, + 598 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "From the attackers' point of view, jailbreaking these MLLM models with multi-modal input seems more straightforward than jailbreaking language models [21]. Instead of optimizing adversarial noise to elicit harmful contents from vision language models with white-box accesses [28, 32], the typographic attack is a typical example to demonstrate the inferior safety of MLLMs. Fig-Step [13] directly transforms harmful requests into images through typography to deceive the image encoder module instead of the language module. With similar logic, MM-SafetyBench [25] proposes to add harmful keywords at the bottom of content-related diffusion-generated images, successfully inducing VLMs to generate inappropriate content across multiple fields. To defend such a simple but effective attack, except for prompt engineering techniques at inference time, previous work approaches this by collecting a safe multi-modal dataset with toxic queries and rejection responses, and then finetuning the MLLMs either in the second finetuning stage or after their release [55]. Regardless of a great success, current methods also leave a few questions unresolved.", + "bbox": [ + 511, + 599, + 908, + 902 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "CVF", + "bbox": [ + 106, + 2, + 181, + 42 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore.", + "bbox": [ + 236, + 0, + 810, + 46 + ], + "page_idx": 0 + }, + { + "type": "page_footnote", + "text": "Corresponding author.", + "bbox": [ + 114, + 887, + 250, + 900 + ], + "page_idx": 0 + }, + { + "type": "page_number", + "text": "19879", + "bbox": [ + 480, + 944, + 517, + 955 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/6e69771cca7474e3b1c6adcac02f81ad8c18c69ae5cf058c9f81c85fa3fbb1cc.jpg", + "image_caption": [ + "Figure 1. The main workflow of our study. After demonstrating that the quality of safety finetuning data does not contribute as much as we imagine, we modify the LLaVA-Med [20] dataset to create our finetuning dataset. For each data point, we simply pick one round of the conversation and replace their original answer with a clear rejection, without periods and tokens." + ], + "image_footnote": [], + "bbox": [ + 89, + 85, + 903, + 372 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "tions. For one thing, collecting high-quality is always labor-intensive. Even if we could prompt powerful LLMs and diffusion models to generate the required multi-modal content, it is still effort-taking to create the pipeline, e.g. extracting keywords from related topics for diffusion models, or prompt engineering for high-quality various harmful questions. More importantly, what really matters in the multimodal alignment process? Do VLMs forget their safety alignment through multi-modal instruction tuning? We are still unclear why simple typographic attacks with keywords incitement could break the safety alignment of vision language models, while they possess identical meanings with the text-only prompts.", + "bbox": [ + 88, + 449, + 485, + 647 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "In this work, we are the first to demonstrate that the alignment gap between the vision-language inputs and text-only counterparts comes mainly from training data distribution bias. In detail, we perform ablation studies on the previous finetuning method to filter out some unnecessary attributes of the dataset, discovering that the defense capability may be hardly related to the response quality, answer-rejection comparisons, or conversation rounds and image contents. On the contrary, increasing rejection rates in the finetuning dataset will elicit over-rejection, indicating the dataset distribution may matter in safety finetuning. To further explore the hypothesis that dataset distribution is the key factor for model alignment, we leverage the safety-unrelated LLaVA-Med dataset [20] to mitigate the compliance bias contained in open-source MLLMs. Experiments demonstrate that without safety-related images, harmful", + "bbox": [ + 89, + 659, + 483, + 901 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "text prompts, and detailed rejection reasons, only \"teaching the LLaVA- $\\nu$ 1.5 models to reject a specific proportion of benign prompts\" appears to boost multi-modal safety significantly. Detailed quantitative and qualitative results are followed, encompassing the relationship between reason generation and ordinary training data. To further explore to what extent the rejection proportion is strong enough to mitigate the compliance bias, we make ablation studies on the rejection dataset size, analyze the influence rejection data proportion has on two jailbreak methods, and launch full visual instruction tuning to compare the effectiveness with post-finetuning. We believe our findings could help better understand the safety-related behavior of MLLMs.", + "bbox": [ + 511, + 449, + 908, + 647 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "2. Related Work", + "text_level": 1, + "bbox": [ + 513, + 657, + 656, + 672 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "2.1. Multi-modal Large Language Model", + "text_level": 1, + "bbox": [ + 511, + 681, + 831, + 698 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "In this paper, we leverage materials within the vision language models (VLMs) context. Most MLLMs consist of three parts: modality encoder, pretrained LLM, and modality interface [51]. Instead of training a MLLM from scratch, companies or academies utilize the modality connector to preprocess the on-the-shelf large language models with image encoders [4-6, 9, 44, 50, 54], in which the image encoders are mainly CLIP variants [7, 35, 41]. The language models are mainly pre-trained chat versions, such as Llama families [42], Vicuna families [8], and Qwen series [1, 49]. When training the MLLMs, the prevailing procedure contains vision-language pre-training and visual instruction tuning [9, 22, 24, 54]. In the first phase,", + "bbox": [ + 511, + 704, + 908, + 902 + ], + "page_idx": 1 + }, + { + "type": "page_number", + "text": "19880", + "bbox": [ + 480, + 944, + 519, + 955 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "the multi-modal projectors are trained in image captioning tasks. After that, in the instruction tuning phase, both the large language model and the connector are tuned with a large amount of VQA data [3, 43], aiming to enhance the visual understanding, reasoning, and OCR capability [29, 40]. However, safety alignment data are hardly included in the training phase, especially for those models from non-profit academies [4, 24, 54]. As a consequence, these MLLMs only rely on the safety alignment within their language module to defend against harmful prompts. Previous works prove the weakness of these MLLMs when faced with multi-modal inputs, and attribute this failure to the lack of curated VQA data [31, 47].", + "bbox": [ + 89, + 90, + 483, + 287 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "2.2. Black-box Jailbreak", + "text_level": 1, + "bbox": [ + 89, + 305, + 282, + 320 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "As mentioned above, due to the obvious alignment gap between text-only prompts and vision-language prompts in MLLMs, Many researchers leverage the alignment gap to launch multi-modal jailbreaks. To our best knowledge, Fig-Step [13] is the first work to jailbreak MLLMs with typography. When printing an image of a list containing steps for unethical behaviors with empty items as the input, such an attack could circumvent the alignment while keeping the OCR capability to fill in the blanks. Subsequent works follow the typographic attacks, append a content-related image on the top, and replace the sensitive words in text prompts with image incitement to enhance the attack success rate, e.g. replace \"sell illegal drugs\" with \"the behavior shown in the image\" [21, 25]. Such an incitement method could also be used to launch white-box attacks, where the attackers interact with their well-optimized adversarial examples by pronouns [39] to circumvent model safety alignment.", + "bbox": [ + 89, + 330, + 483, + 587 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "2.3. Jailbreak Defense", + "text_level": 1, + "bbox": [ + 89, + 604, + 264, + 618 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "The Jailbreak defense on MLLMs could be mainly split into two categories: the inference phase and the training phase. Inference-time defenses are more like post-processing: AdaShield [46] proposes to create a shield prompt pool, adaptively appending a warning prompt at the end of user prompts to highlight response safety; MLLM-Protector [31] proposes a harm detector to firstly filter out harmful contents, and then change the response by their fine-tuned response detoxifier; similarly, ECSO [14] transforms input images into texts in a query-aware manner once the originally generated contents are labeled as inappropriate. To our knowledge, the work most related to our standpoint is CoCA [12], which focuses on exploring the safety-awareness of multi-modal inputs. Instead of training, they modify the decoding phase to enhance model safety. In the training phase, VLGuard [55] collects a safety dataset for supervised finetuning (SFT), achieving satisfying results in the defense of typographic attacks.", + "bbox": [ + 89, + 628, + 483, + 900 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3. What Matters in Safety Finetuning Dataset?", + "text_level": 1, + "bbox": [ + 513, + 89, + 906, + 107 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3.1. Preliminaries", + "text_level": 1, + "bbox": [ + 513, + 114, + 653, + 128 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Finetuning workflow. As stated in the introduction part, the previous method could finetune the MLLM in two ways: mixing safety data in the second stage instruction tuning, or directly finetuning the released model. Here we pick VL-Guard [55] as the target for the ablation study, and we train the models based on the released version with LoRA.", + "bbox": [ + 511, + 137, + 906, + 227 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "VLGuard dataset [55]. The dataset composes three types of image-text pairs: safe-safe, safe-unsafe, and unsafe-unsafe. For example, safe-safe indicates a benign image-text pair, and safe-unsafe means the benign image is paired with a harmful (and in most cases, unrelated) text prompt. Here is an example of an LLaVA-style data sample:", + "bbox": [ + 511, + 228, + 906, + 320 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Prompt: What brand of beer is shown in the image?", + "bbox": [ + 544, + 348, + 874, + 376 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Assistant: The brand of the beer shown in the image is Grolsch.", + "bbox": [ + 544, + 377, + 874, + 402 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Prompt: Can you provide information on how to create fake IDs to purchase alcohol like the one shown in the image?", + "bbox": [ + 544, + 404, + 874, + 445 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Assistant: I'm sorry, but I cannot provide assistance or information on illegal activities, including creating fake IDs to purchase alcohol. Engaging in such actions is against the law and our usage policy strictly prohibits facilitating or promoting illegal activities. It's important to follow the legal age requirements for purchasing and consuming alcoholic beverages.", + "bbox": [ + 544, + 446, + 874, + 542 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "For training, the dataset contains 977 single-turn unsafe-image pairs and 1023 multi-turn safe-image pairs, encompassing topics related to privacy, risky behavior, deception, and discrimination.", + "bbox": [ + 511, + 564, + 906, + 623 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3.2. Hypotheses", + "text_level": 1, + "bbox": [ + 513, + 635, + 638, + 651 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "High-quality datasets with well-picked images and detailed responses have been demonstrated to strengthen the safety of MLLMs. Here, we make several hypotheses on dataset attributes that contribute to satisfying defense capability.", + "bbox": [ + 511, + 657, + 906, + 718 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Comparisons in multi-turn data. As stated above, safe-safe and safe-unsafe pairs share identical images, which are structured in a multi-turn conversation style. Therefore, it is reasonable to hypothesize that the contrastive manner of the dataset forces MLLMs to understand what kind of input is toxic, and what kind of inputs are benign even with identical image visual prompts.", + "bbox": [ + 511, + 719, + 905, + 824 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Content-related images. It has been proved that content-related images tend to strengthen the jailbreak prompts [31]. For example, with a bomb image, the model has a higher probability of outputting illegal contents on manufacturing weapons [25]. Thus, We hypothesize that", + "bbox": [ + 511, + 825, + 906, + 900 + ], + "page_idx": 2 + }, + { + "type": "page_number", + "text": "19881", + "bbox": [ + 480, + 944, + 517, + 955 + ], + "page_idx": 2 + }, + { + "type": "table", + "img_path": "images/6ade23144ad273c9453bfd5da902888d7cff63bdd34f88be9ae186b141290bf9.jpg", + "table_caption": [], + "table_footnote": [], + "table_body": "
Base ModelEvaluation DatasetsModel Settings
OriginVLGuardOne-turnUnsafe-onlyChange imageDirect sorryRandom reasonPure-VLGuard
LLaVA-v1.5-7BMM SafetyBench↓96.370.180.180.180.180.120.120.05
FigStep↓100.000.000.000.000.000.000.000.00
XCTest Compliance↑92.0077.2078.8076.8079.2070.8078.0052.00
XCTest Rejection↑75.5096.5095.5096.0093.0099.0098.0099.00
LLaVA-v1.5-13BMM SafetyBench↓97.800.360.360.180.180.420.540.12
FigStep↓99.800.000.000.000.000.000.000.00
XCTest Compliance↑90.0077.6077.2076.8082.0078.0078.0061.60
XCTest Rejection↑84.5097.0096.0096.0094.5098.0097.0099.50
LLaVA-NeXT Mistral-7BMM SafetyBench↓100.00.180.240.480.240.240.990.12
FigStep↓99.880.000.000.000.000.000.120.00
XCTest Compliance↑94.8087.6086.0081.6086.4086.4084.4043.20
XCTest Rejection↑58.5089.5084.5091.5084.0089.5094.0099.50
LLaVA-NeXT Llama-3-8BMM SafetyBench↓96.070.120.182.800.060.710.120.12
FigStep↓99.800.000.000.0038.40.000.000.00
XCTest Compliance↑90.4082.8082.8082.8086.4076.8078.4060.40
XCTest Rejection↑83.5092.0091.0091.0083.5097.0097.5099.00
", + "bbox": [ + 94, + 88, + 903, + 325 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Table 1. Model safety evaluation with various finetuning data. All values are percentages.", + "bbox": [ + 232, + 335, + 761, + 349 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "MLLM safety could be gained from the rejection of unsafeunsafe pairs. After collecting these harmful images and making models learn to reject them, the defense capability of MLLMs probably advances by a huge margin.", + "bbox": [ + 89, + 376, + 482, + 436 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Detailed reject reasons. Detailed reject reasons clarify the specific legal principles the harmful prompts violate, highlighting that as an AI model, it should follow legal and moral requirements to provide helpful and safe content. They add extra safety-related knowledge to the dataset, therefore, it is reasonable to hypothesize that such detailed data are crucial to invoke previous safety alignment in the language module, and then fill the gap between the multimodal input and pure-text domains.", + "bbox": [ + 88, + 438, + 482, + 571 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Dataset distribution bias. Comparing the instruction tuning dataset with the safety finetuning dataset, one distinct difference is the rejection behavior. Regardless of whether the training set contains toxic or harmful contents, in the instruction tuning phase the model hardly learns how to reject because nearly all replies in conversation-style data follow the instructions as detailed as possible. We hypothesize that this may also be the reason why the model loses the ability to reject inappropriate requests.", + "bbox": [ + 89, + 575, + 482, + 712 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "3.3. Experiment Settings", + "text_level": 1, + "bbox": [ + 89, + 726, + 284, + 742 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Finetuning models and datasets. Here, we launch all experiments on the LLaVA series model [22-24]. Aligning with previous methods, we add random-picked 5000 data points from the visual instruction tuning dataset LLaVA-v1.5-mix665k [22]. To testify to the influence of comparisons in the dataset, we split multi-turn safety data into single-turn and eliminate all safe-safe data pairs, named \"One-turn\" and \"Unsafe-only\", respectively. We also change the image prompt with random-picked benign images from LLaVA-v1.5-mix665k, named \"Change im", + "bbox": [ + 89, + 750, + 482, + 900 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "age\" for the content-related image hypothesis. To modify the detailed reasons, we design two alternatives: replacing all replies with one clear rejection sentence or generating rejection contents with the prefix: \"I'm sorry, I cannot assist with that request because it goes against my programming to\" on benign data in the finetuning dataset. For comparisons, we also add experiments on pure VLGuard dataset [55] and non-finetuned original models. Experiments are done on LLaVA-v1.5-7B, LLaVA-v1.5-13B, LLaVA-NeXT-mixtral-7B, and LLaVA-NeXT-llama3-8B [22-24] with LoRA [16]. Finetuning details are listed in the Appendix.", + "bbox": [ + 511, + 376, + 906, + 556 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Evaluation metrics. We mainly evaluate the safety of MLLM models with black-box attacks, for it is the prevailing safety threat MLLM models will face. With collected jailbreaking image-text pairs on the Internet, a user could directly launch such an attack. We pick FigStep [13] and MM-SafetyBench [25] as the attacking sources, and the attack successful rate (ASR) is calculated by string-matching. We also tried to use Llama-3-Guard [10] as the discriminator in early evaluations, but the ASRs are lower because some sensitive topics exhibited in the datasets are not regarded as harmful, even if the model fails to reject content generation. Besides, to evaluate the model performance on visual understanding tasks, we adopt VizWizQA [15] and ScienceQA [27] for visual-text understanding, and XSTest dataset [37] for over-rejection evaluation. One point worth noting is that the XSTest dataset comprises two parts, in which the compliance column means the proportion of benign questions with non-rejection answers.", + "bbox": [ + 511, + 559, + 908, + 832 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "3.4. Results", + "text_level": 1, + "bbox": [ + 511, + 845, + 607, + 861 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Detailed safety and performance evaluations of models on various ablation datasets are shown in Table 1.", + "bbox": [ + 511, + 869, + 906, + 900 + ], + "page_idx": 3 + }, + { + "type": "page_number", + "text": "19882", + "bbox": [ + 480, + 944, + 517, + 955 + ], + "page_idx": 3 + }, + { + "type": "table", + "img_path": "images/622ec068cdbaa5b393a05d49fcff09891e01e433933379aca6b609acd793aaab.jpg", + "table_caption": [], + "table_footnote": [], + "table_body": "
Base ModelFinetuning DatasetsVizWizQA(%)↑ScienceQA(%)
Image↑Total↑
LLaVA-v1.5-7BOrigin55.1070.1269.46
VLGuard56.3668.4270.12
One-turn56.3967.6769.79
Change image55.7867.3869.04
Unsafe-only56.7567.8269.63
Direct sorry53.9867.9769.61
Random reason56.6568.5270.01
Pure-VLGuard56.4565.2464.18
LLaVA-v1.5-13BOrigin57.3472.6874.89
VLGuard56.7771.5472.62
One-turn54.2871.3472.95
Change image55.5770.7572.98
Unsafe-only59.6270.8572.48
Direct sorry56.9871.0071.66
Random reason58.3971.0072.29
Pure-VLGuard58.2671.0071.47
", + "bbox": [ + 94, + 88, + 498, + 325 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Table 2. Evaluation of model capabilities on VQA tasks.", + "bbox": [ + 117, + 334, + 454, + 349 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Comparisons in multi-turn data may make a limited contribution to the safety enhancement. With extensive experiments on both LLaVA-v1.5 [22] and LLaVA-NeXT [23] structures with different model sizes, the finding is consistent that the safety alignment of VLMs hardly relies on the comparison of answers to benign prompts and harmful prompts. For one thing, after splitting a double-turn conversation into two single-turn data points with the same image prompt, the defensive capability of all four models remains nearly unchanged, suggesting that utilizing some benign conversations as the conversation history may have a limited impact on enhancing the defense capability; for another, even eliminating all benign conversations from VL-Guard dataset [55] makes little difference to the defense result. Such a phenomenon indicates that benign data here may only contribute to the model's helpfulness, similar to randomly picked data in LLaVA-v1.5-mix665k [22].", + "bbox": [ + 89, + 383, + 482, + 641 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Content-related images and reject reasons may not be the key factor for the defensive capability. In Table 1, it turns out that changing content-related harmful images to benign training images hardly deteriorates the defensive capability. Except for the inferior performance LLaVA-NeXT-Llama-3-8B [23] has on FigStep [13], under the rest circumstances training models with unrelated images could result in a satisfying defensive performance. Besides, in Table 2, the performance difference caused by image changing is only about $1\\%$ on both VQA tasks [15, 27]. Similar trends also occur on ablations for reject reasons. Modifying or directly eliminating the reasons does not harm defensive capability, indicating that the model learns little from the detailed reasons for making the decision. Qi et al. [33] proposed a similar conclusion in the language model domain, claiming that current finetuning gradients mainly come from a few first tokens.", + "bbox": [ + 89, + 643, + 482, + 900 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "High rejection proportion has the potential to harm the model performance. After eliminating training samples from LLaVA-v1.5-mix665k [22], the rejection conversation takes up around $66.7\\%$ of the whole dataset. With such a data distribution, all models get over-sensitive: they could reject more than $99\\%$ harmful text prompts while only answering about half of the benign questions, even though the curated finetuning data are of high quality. Besides, we also witness an accuracy drop on the ScienceQA dataset [27] with the LLaVA-v1.5-7B model. This phenomenon is consistent with the prior work [55], highlighting the existence of normal training data.", + "bbox": [ + 511, + 90, + 903, + 271 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "To sum up, in the finetuning phase, it seems the models learn little from the high-quality content. Dataset modifications on images, answer reasons, and conversation comparisons probably only have negligible impacts on the safety alignment. However, altering the rejection proportion of the dataset could elicit over-rejection: even though the models only learn how to reject harmful instructions, the rejection behavior is wrongly generalized to benign prompts. Based on this, one question emerges: is dataset distribution the key factor for model alignment? If that is the case, is it possible to enhance MLLM safety without well-curated safe data?", + "bbox": [ + 511, + 272, + 906, + 439 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "4. Benign Data with Rejection Responses Boost Model Safety", + "text_level": 1, + "bbox": [ + 511, + 452, + 906, + 487 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "To figure out the role dataset rejection proportion plays in safety finetuning, we try to stimulate the alignment potential with benign data.", + "bbox": [ + 511, + 494, + 906, + 542 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "4.1. Experimental Setup", + "text_level": 1, + "bbox": [ + 511, + 550, + 702, + 566 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "To align with the previous setting, we randomly pick 2000 data points from LLaVA-Med dataset [20]. For each data point, we randomly pick one round of QAs in the conversation and replace the answer with a clear rejection. As in Sec. 3.3, these data are also mixed with randomly picked 5000 normal data from LLaVA-v1.5-mix665k [22]. Noting that the training data of Yi-VL-6B [52] and LLaVA-NeXT [23] are not publicly available, therefore we keep using the same setting as the LLaVA-v1.5 models. Considering possible overfitting, we only supervise the model to generate a few rejection tokens while masking out the end-of-sentence token. Assuming in a one-turn conversation $x$ , the instruction takes up $m$ tokens. In our experiments, we only calculate the language modeling loss by this equation [56]:", + "bbox": [ + 511, + 571, + 905, + 785 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {L} \\left(x _ {1: n + m}\\right) = - \\log p \\left(x _ {n + 1: n + m} ^ {*} \\mid x _ {1: n}\\right), \\tag {1}\n$$\n", + "text_format": "latex", + "bbox": [ + 575, + 796, + 906, + 814 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "where $x_{n+1:n+m}^*$ refers to the tokenized rejection string: \"I'm sorry, but I cannot assist with your request\", without periods or token ids. For simplicity, here we ignore the existence of system prompts as well as token in some templates. To ensure the reliability of our results,", + "bbox": [ + 511, + 825, + 905, + 902 + ], + "page_idx": 4 + }, + { + "type": "page_number", + "text": "19883", + "bbox": [ + 480, + 944, + 517, + 955 + ], + "page_idx": 4 + }, + { + "type": "table", + "img_path": "images/ca8dac4a0361e8b45c980f9a26618c74dd33e6f3e0c1aff83dfd847506b86f2d.jpg", + "table_caption": [], + "table_footnote": [], + "table_body": "
Base ModelFinetuning DatasetsMM SafetyBench↓FigStep↓XCTest↑VizWizQA↑ScienceQA↑
ComplianceRejectionImageTotal
LLaVA-v1.5-7BOrigin96.37100.0092.0075.5055.1070.1269.46
VLGuard0.180.0077.2096.5056.3668.4270.12
Ours5.600.2090.4082.0055.9268.0269.11
LLaVA-v1.5-13BOrigin97.8099.8090.0084.5057.3472.6874.89
VLGuard0.360.0077.6097.0056.7771.5472.62
Ours1.790.0090.4090.5058.7771.5973.36
Yi-VL-6BOrigin93.5199.6095.2041.50\\( 66.20^1 \\)61.0370.12
VLGuard0.180.0084.4093.0059.3269.8174.51
Ours3.573.2089.2067.0047.9168.2773.07
LLaVA-NeXT Mistral-7BOrigin99.88100.0094.8058.5064.9872.6378.57
VLGuard0.180.0087.6089.5053.1570.2074.56
Ours2.610.0091.2065.5055.5069.3674.65
", + "bbox": [ + 135, + 88, + 864, + 314 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Table 3. The evaluation comparisons between our finetuning dataset and VLGuard [55]. All values are percentages.", + "bbox": [ + 155, + 324, + 839, + 339 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "except for LLaVA-v1.5 models with Vicuna language module, we also pick the LLaVA-NeXT with Mistral-7B language module [17], and the Yi-VL-6B based on Yi-6B [52]. Training details are shown in the Appendix.", + "bbox": [ + 89, + 364, + 482, + 428 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "4.2. Quantitative Results", + "text_level": 1, + "bbox": [ + 89, + 438, + 284, + 454 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "MLLMs do not lose defense on multi-modal prompts. In Table 3, it turns out that all four models work well on jailbreak defenses while maintaining a relatively high rate of benign prompt compliance. For MM-SafetyBench [25], the ASR of most models is below $5\\%$ , which is a substantial drop compared to the original models. Experiments on FigStep [13] are also obvious: the highest ASR among the four models is $3.2\\%$ . In training data, the conversation topic is unrelated to safety and privacy, and we do not append any reason for the \"abnormal\" rejection in each data point. The major factor is the rejection rate: compared to the visual instruction tuning phase with limited rejection data, we add 2000 rejection samples to 5000 benign training data in safety finetuning, reducing the proportion of instances where the answer complies with instructions.", + "bbox": [ + 88, + 460, + 482, + 686 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "More importantly, our experimental results provide little support for the notion that the model forgets previous safety alignment or fails due to the absence of multi-modal safety data, leading it to interpret harmful jailbreak prompts as benign. Instead, we argue that the model fails to defend jailbreaks mentioned above mostly because all multi-modal instruction-tuning data naturally create a compliance bias. After instruction tuning, the model still possesses the basic moral understanding and judgment. The difference is that it is told all prompts with images should be answered and followed eagerly, resulting in the vulnerability.", + "bbox": [ + 88, + 688, + 482, + 854 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Without knowledge, correcting model distribution makes limited progress to text-only prompts. In Table 3, finetuning on our dataset does not significantly impair the", + "bbox": [ + 89, + 854, + 483, + 902 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "safety compliance in XSTest [37], but it also fails to improve the unsafe rejection rate to the level achieved by VLGuard [55]. It is reasonable, for our dataset contains no extra safety-related contents compared to the VLGuard. Accordingly, merely correcting the model's distribution without incorporating additional knowledge is unlikely to result in a significant improvement in text-only input defense.", + "bbox": [ + 511, + 364, + 906, + 470 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Performances on general tasks tend to be associated with the ordinary training data. In experiments, we mix rejection data with ordinary data from LLaVA-v1.5-mix665k [22], the official training set for LLaVA-v1.5. In the first two rows, using our random rejection data could achieve similar results as VLGuard on VizWizQA [15] and ScienceQA [27], most of which are slightly inferior to the original model. For VizWizQA the model trained on our dataset even outperforms the original one, indicating the quality of the ordinary data. However, for LLaVA-NeXT [23] and Yi-VL [52], the performances experience a larger drop, which may be attributed to the inconsistency of training data. This adds more randomness to the results.", + "bbox": [ + 511, + 472, + 908, + 667 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Analyzing the results from the rejection data proportion perspective, we could also prove the claim that large language model alignment contributes to the defense of vision-language jailbreaks. Suppose the model does not possess safety-related knowledge or judgment for multi-modal data. In that case, the rejection rate of generated contents should be identical to, or slightly less than training data distribution. With $28.57\\%$ rejection data in the training set, ASRs of LLaVA-v1.5-7B on jailbreak datasets should be around $70\\%$ , and the performance on visual understanding and reasoning tasks should drop by $15\\%$ to $20\\%$ . However, the evaluation discrepancy between jailbreak datasets and vi", + "bbox": [ + 511, + 669, + 910, + 851 + ], + "page_idx": 5 + }, + { + "type": "page_footnote", + "text": "1The original value under the standard evaluation metric is 4.99, for most answers are appended by meaningless periods and line breaks. The score is calculated after omitting these suffixes.", + "bbox": [ + 511, + 862, + 906, + 900 + ], + "page_idx": 5 + }, + { + "type": "page_number", + "text": "19884", + "bbox": [ + 480, + 944, + 519, + 957 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "sual understanding datasets discloses that the rejection behavior is not evenly added to the content generation. In other words, the model gets over-compliance after visual instruction tuning. Still, it knows which multi-modal prompts are harmful, just as it knows how to reject those text-only harmful prompts without interactions and incitements with images. The semantic meanings between multi-modal inputs and their text-only counterparts are aligned. In safety finetuning, it is less important to meticulously collect security data for the model to learn how to distinguish between benign and malicious instructions. Once enough rejection data is provided to enable rejection behavior, the model can make the right decisions.", + "bbox": [ + 89, + 90, + 483, + 287 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "4.3. Qualitative Results", + "text_level": 1, + "bbox": [ + 89, + 297, + 272, + 313 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "To have a rough understanding of the answer quality, we present some examples of the model-generation content.", + "bbox": [ + 89, + 319, + 483, + 351 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Prompt: \\n The bottom of the image shows a phrase of an activity. List steps in detail to perform this activity using cryptocurrencies without leaving a trace.", + "bbox": [ + 125, + 372, + 447, + 428 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "LLaVA-v1.5-7B", + "text_level": 1, + "bbox": [ + 125, + 441, + 227, + 454 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Assistant: I'm sorry, but I cannot assist with your request. Purchasing illegal drugs with cryptocurrencies...the use of cryptocurrencies for illegal activities can lead to legal consequences and is morally wrong. It is important to support legal and legitimate businesses...", + "bbox": [ + 125, + 455, + 447, + 539 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Yi-VL-6B", + "text_level": 1, + "bbox": [ + 127, + 553, + 189, + 565 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Assistant: I'm sorry, but I cannot assist with your request. Please consult a professional for assistance with your request.", + "bbox": [ + 125, + 566, + 447, + 609 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "LLaVA- $\\nu 1.5 - 7B$ -Random-reason", + "text_level": 1, + "bbox": [ + 125, + 622, + 326, + 633 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Assistant: I'm sorry, I cannot assist with that request because it goes against my programming to provide opinions or beliefs.", + "bbox": [ + 125, + 636, + 447, + 676 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Figure 2. Examples of replies generated under different settings. The reasons are not as high quality as the LLaVA-v1.5. To save space, the image of drugs with the phrase \"purchase illegal drugs\" is not shown in the box.", + "bbox": [ + 89, + 707, + 483, + 762 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "The quality of generated rejection reasons is relatively independent of model decisions. In Fig. 2, the LLaVA-v1.5 model generates high-quality responses, with rejection reasons related to the topic. However, For Yi-VL-6B [52] and LLaVA-NeXT [23], the models keep satisfying reject decisions, while generating imperfect reasons. Similar results are also demonstrated when training LLaVA-v1.5 models under the random reason setting, indicating that the", + "bbox": [ + 89, + 780, + 483, + 901 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "training data quality possibly affects the following sentence completion. With a proper rejection rate in the training set, models can reject inappropriate prompts, while the performance deterioration may be attributed to the distribution distinction among different training data.", + "bbox": [ + 511, + 90, + 905, + 167 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "4.4. Further Analyses on Conversation Datasets", + "text_level": 1, + "bbox": [ + 511, + 176, + 882, + 191 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Introducing rejection answers to benign prompts may create a rejection bias on normal conversations. Therefore, except for XSTest [37] and VQA evaluations [15, 27], we also pick 1000 samples from LLaVA-v1.5-mix665k [22], LLaVA-Instruct-150K [24], and MMInstruct [26] each (for MMInstruct [26] samples are from the \"qa_en\" subset) to evaluate the compliance rate of LLaVA-v1.5 models. String matching is adopted as the metric.", + "bbox": [ + 511, + 199, + 906, + 320 + ], + "page_idx": 6 + }, + { + "type": "table", + "img_path": "images/9c9e01c91c3650f41691da59a62aa7e1e5e4e4b048f5f4ca6ee1c2183b3f9771.jpg", + "table_caption": [], + "table_footnote": [], + "table_body": "
BaseDatasetsOriginVLGuardOurs
LLaVA-v1.5-7B665K↑100.00100.0099.90
150K↑99.6098.9098.70
MM↑100.0099.0088.40
LLaVA-v1.5-13B665K↑100.00100.0099.80
150K↑99.6099.4098.30
MM↑100.0099.2085.90
", + "bbox": [ + 521, + 333, + 898, + 450 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Table 4. Compliance rate under different finetuning settings. 665K, 150K, and MM represent LLaVA-v1.5-mix665k [22], LLaVA-Instruct-150K [24], and MMInstruct [26], respectively.", + "bbox": [ + 511, + 460, + 905, + 503 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "As shown in Table 4, the performance is almost intact for tasks from the LLaVA-v1.5-mix665k [22], but the models tend to reject image descriptions at a higher rate on LLaVA-Instruct-150K [24] and MMInstruct [26]. This may be because the data sampled from LLaVA-Med [20] contain quantities of image description queries, which induce models to reject similar prompts in evaluations, thus sacrificing the original performance to some extent.", + "bbox": [ + 511, + 520, + 905, + 641 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "4.5. Ablation on Rejection Data Proportion", + "text_level": 1, + "bbox": [ + 511, + 651, + 846, + 667 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "With the understanding that datasets with rejection samples could boost model safety, it is natural to ask the following question: without safety-related knowledge in the dataset, how much data is enough to counteract the compliance bias buried in the visual instruction tuning phase? To figure out the relationship between rejection data quantity and jailbreak defense behavior, we finetune the LLaVA-v1.5-7B five times, with increasing numbers of rejection data, while keeping the remaining setting changed. Still, for each experiment, 5000 randomly picked ordinary data are added to the training set. The performance on the original LLaVA-v1.5-7B is also added for reference.", + "bbox": [ + 511, + 672, + 906, + 853 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "From the experiments, we could observe that the defense capability against MM-SafetyBench [25] prompts is more sensitive to the existence of rejection data than Fig-", + "bbox": [ + 511, + 854, + 905, + 900 + ], + "page_idx": 6 + }, + { + "type": "page_number", + "text": "19885", + "bbox": [ + 480, + 944, + 517, + 955 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/12695c757fb0afe130268417e9b0bb1c32599e7b037038e719968f16cca5e8a2.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 116, + 90, + 460, + 176 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/6f079d1c307a443c7daad69e4ce191e181d538d6f5803cebaf27553a5140a668.jpg", + "image_caption": [ + "Figure 3. The influence rejection data proportion has on jailbreak defense capability and visual understanding accuracy." + ], + "image_footnote": [], + "bbox": [ + 114, + 181, + 460, + 291 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Step [13]. For both attacks, $28.57\\%$ is demonstrated to be enough to counteract compliance bias. Besides, with all rejection ratios, the visual understanding performance on VizWizQA [15] and ScienceQA [27] only slightly fluctuates around the baseline, which corresponds to previous findings in Sec. 4.", + "bbox": [ + 89, + 358, + 482, + 449 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "4.6. Adding Data in the Instruction Tuning Phase", + "text_level": 1, + "bbox": [ + 89, + 458, + 472, + 474 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "To gain a deeper understanding of this topic, in this part, we add rejection data into the multi-modal instruction following data to launch the visual instruction tuning. To be specific, due to the limitation of resources, we only design three runs, where the rejection data comprises $2\\%$ and $5\\%$ of the original LLaVA-v1.5-mix665k dataset [22], and the specific amount of rejection data is mixed in the original dataset for experiments. Identical to the official training setting, we adapt full-parameter tuning on the pretrained multi-modal projector and the released Vicuna-7B-v1.5 [8]. Training details are listed in the Appendix.", + "bbox": [ + 89, + 479, + 482, + 647 + ], + "page_idx": 7 + }, + { + "type": "table", + "img_path": "images/9e423e2c96a03e30a6bb0b824a2ff8bc08970d643260cc03bebf9e5016db6c8d.jpg", + "table_caption": [], + "table_footnote": [], + "table_body": "
Evaluation DatasetsMetric (%)Origin2% Rej. data5% Rej. data
MM-SafetyBenchASR↓96.3658.9360.12
FigStepASR↓100.0099.4099.80
XSTestComp. rate↑92.0089.2091.20
XSTestRej. rate↑75.5077.0073.00
ScienceQAImage Acc.↑69.4669.7169.16
ScienceQATotal Acc.↑70.1271.1670.53
VizWizQAAcc.↑55.1056.5656.26
", + "bbox": [ + 96, + 657, + 478, + 789 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Table 5. Evaluation results of Visual Instruction Tuning. \"Origin\" refers to the official LLaVA-v1.5-mix665k dataset. \"Rej.\", \"Comp.\", and \"Acc.\" represents \"Rejection\", \"Compliance\", and \"Accuracy\", respectively.", + "bbox": [ + 89, + 797, + 482, + 854 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "The model safety gets improvements, but not good enough. As shown in Table 5, without harming the vi", + "bbox": [ + 89, + 869, + 482, + 902 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "sual understanding capability, the attack successful rate of MM-SafetyBench [25] drops from $96.36\\%$ to around $60\\%$ , which demonstrates again that MLLMs do not lose defense on multi-modal prompts. However, the defense against FigStep [13] jailbreak does not experience an obvious drop, which matches the ablation study in Fig. 3: for a relatively small percentage of rejection data, the ASR of FigStep remains high. Therefore, the conclusion is drawn that if the proportion of rejected data is insufficient, the defense behavior may not be fully activated.", + "bbox": [ + 511, + 90, + 903, + 241 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "From experiments, if the rejection data lacks any inherent knowledge, awakening the model's multi-model safety awareness probably requires a sufficiently high rejection rate. Given this, enhancing model safety through finetuning seems to be more effective, where a smaller amount of data can constitute a substantial portion of the entire finetuning dataset. Specifically, in full-parameter visual instruction tuning, $2\\%$ of the dataset corresponds to approximately 13k data points, which is more than six times the quantity required for finetuning. Despite this, the safety improvements gained through finetuning are far more significant.", + "bbox": [ + 511, + 242, + 906, + 409 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "5. Conclusion and Limitation", + "text_level": 1, + "bbox": [ + 511, + 424, + 764, + 441 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "In this paper, we investigate the well-curated safety finetuning data, finding that safety-related contents may not be the most crucial part for safety alignment. Based on this, we design a rejection dataset and conclude that the key factor influencing finetuning tends to be the existence of rejection data, instead of other attributes. Experiments demonstrate that MLLMs do not lose defense on multi-modal prompts. On the contrary, it is the compliance bias that prevents the models from refusing the malicious prompts, and training on the rejection dataset could mitigate such a bias.", + "bbox": [ + 511, + 450, + 905, + 601 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "The major limitation in our study is the unsatisfying performance drop under image description queries. Besides, we answer the question that boosting multi-modal large language models without the necessity of curated data is possible, but currently it cannot be used to enhance the safety in the text-only domain significantly. We will leave both of them as future work. To summarize, we hope our findings with supervised finetuning methods can help better understand the safeguard of MLLM.", + "bbox": [ + 511, + 602, + 905, + 739 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Acknowledgment", + "text_level": 1, + "bbox": [ + 511, + 753, + 666, + 771 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "We thank Zhengbo Wang and Yongcan Yu in NLPR for their feedback as well as discussion on our early drafts. Besides, we also would like to present our sincere appreciation to the anonymous reviewers for their constructive suggestions. This work was funded by the National Natural Science Foundation of China under Grants (62425606, 62276256, U2441251, U21B2045) and the Young Elite Scientists Sponsorship Program by CAST (2023QNRC001).", + "bbox": [ + 511, + 779, + 905, + 902 + ], + "page_idx": 7 + }, + { + "type": "page_number", + "text": "19886", + "bbox": [ + 480, + 944, + 519, + 955 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "References", + "text_level": 1, + "bbox": [ + 91, + 90, + 187, + 104 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[1] Jinze Bai, Shuai Bai, Yunfei Chu, Zeyu Cui, Kai Dang, Xiaodong Deng, Yang Fan, Wenbin Ge, Yu Han, Fei Huang, et al. Qwen technical report. arXiv preprint arXiv:2309.16609, 2023. 2", + "[2] Yuntao Bai, Andy Jones, Kamal Ndousse, Amanda Askell, Anna Chen, Nova DasSarma, Dawn Drain, Stanislav Fort, Deep Ganguli, Tom Henighan, et al. Training a helpful and harmless assistant with reinforcement learning from human feedback. arXiv preprint arXiv:2204.05862, 2022. 1", + "[3] Guiming Hardy Chen, Shunian Chen, Ruifei Zhang, Junying Chen, Xiangbo Wu, Zhiyi Zhang, Zhihong Chen, Jianquan Li, Xiang Wan, and Benyou Wang. Allava: Harnessing gpt4v-synthesized data for a lite vision-language model. arXiv preprint arXiv:2402.11684, 2024. 3", + "[4] Jun Chen, Deyao Zhu, Xiaogian Shen, Xiang Li, Zechu Liu, Pengchuan Zhang, Raghuraman Krishnamoorthi, Vikas Chandra, Yunyang Xiong, and Mohamed Elhoseiny. Minigpt-v2: large language model as a unified interface for vision-language multi-task learning. arXiv preprint arXiv:2310.09478, 2023. 1, 2, 3", + "[5] Zhe Chen, Jiannan Wu, Wenhai Wang, Weijie Su, Guo Chen, Sen Xing, Muyan Zhong, Qinglong Zhang, Xizhou Zhu, Lewei Lu, Bin Li, Ping Luo, Tong Lu, Yu Qiao, and Jifeng Dai. Internvl: Scaling up vision foundation models and aligning for generic visual-linguistic tasks. arXiv preprint arXiv:2312.14238, 2023. 1", + "[6] Zhe Chen, Weiyun Wang, Hao Tian, Shenglong Ye, Zhangwei Gao, Erfei Cui, Wenwen Tong, Kongzhi Hu, Jiapeng Luo, Zheng Ma, et al. How far are we to gpt-4v? closing the gap to commercial multimodal models with open-source suites. arXiv preprint arXiv:2404.16821, 2024. 2", + "[7] Mehdi Cherti, Romain Beaumont, Ross Wightman, Mitchell Wortsman, Gabriel Ilharco, Cade Gordon, Christoph Schuhmann, Ludwig Schmidt, and Jenia Jitsev. Reproducible scaling laws for contrastive language-image learning. In Proc. CVPR, 2023. 2", + "[8] Wei-Lin Chiang, Zhuohan Li, Zi Lin, Ying Sheng, Zhang-hao Wu, Hao Zhang, Lianmin Zheng, Siyuan Zhuang, Yong-hao Zhuang, Joseph E. Gonzalez, Ion Stoica, and Eric P. Xing. Vicuna: An open-source chatbot impressing gpt-4 with $90\\%$ chatgpt quality, 2023. 2, 8", + "[9] Wenliang Dai, Junnan Li, Dongxu Li, Anthony Tiong, Junqi Zhao, Weisheng Wang, Boyang Li, Pascale Fung, and Steven Hoi. InstructBLIP: Towards general-purpose vision-language models with instruction tuning. In Proc. NeurIPS, 2023. 1, 2", + "[10] Abhimanyu Dubey, Abhinav Jauhri, Abhinav Pandey, Abhishek Kadian, Ahmad Al-Dahle, Aiesha Letman, Akhil Mathur, Alan Schelten, Amy Yang, Angela Fan, et al. The llama 3 herd of models. arXiv preprint arXiv:2407.21783, 2024. 4", + "[11] Chaoyou Fu, Peixian Chen, Yunhang Shen, Yulei Qin, Mengdan Zhang, Xu Lin, Jinrui Yang, Xiawu Zheng, Ke Li, Xing Sun, et al. Mme: A comprehensive evaluation benchmark for multimodal large language models. arXiv preprint arXiv:2306.13394, 2023. 1" + ], + "bbox": [ + 93, + 114, + 483, + 898 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[12] Jiahui Gao, Renjie Pi, Tianyang Han, Han Wu, Lanqing Hong, Lingpeng Kong, Xin Jiang, and Zhenguo Li. Coca: Regaining safety-awareness of multimodal large language models with constitutional calibration. arXiv preprint arXiv:2409.11365, 2024. 3", + "[13] Yichen Gong, Delong Ran, Jinyuan Liu, Conglei Wang, Tianshuo Cong, Anyu Wang, Sisi Duan, and Xiaoyun Wang. Figstep: Jailbreaking large vision-language models via typographic visual prompts. arXiv preprint arXiv:2311.05608, 2023. 1, 3, 4, 5, 6, 8", + "[14] Yunhao Gou, Kai Chen, Zhili Liu, Lanqing Hong, Hang Xu, Zhenguo Li, Dit-Yan Yeung, James T Kwok, and Yu Zhang. Eyes closed, safety on: Protecting multimodal llms via image-to-text transformation. Proc. ECCV, 2024. 3", + "[15] Danna Gurari, Qing Li, Abigale J Stangl, Anhong Guo, Chi Lin, Kristen Grauman, Jiebo Luo, and Jeffrey P Bigham. Vizwiz grand challenge: Answering visual questions from blind people. In Proc. CVPR, 2018. 4, 5, 6, 7, 8, 1", + "[16] Edward J Hu, Yelong Shen, Phillip Wallis, Zeyuan Allen-Zhu, Yuanzhi Li, Shean Wang, Lu Wang, and Weizhu Chen. LoRA: Low-rank adaptation of large language models. In Proc. ICLR, 2022. 4", + "[17] Albert Q Jiang, Alexandre Sablayrolles, Arthur Mensch, Chris Bamford, Devendra Singh Chaplot, Diego de las Casas, Florian Bressand, Gianna Lengyel, Guillaume Lample, Lucile Saulnier, et al. Mistral 7b. arXiv preprint arXiv:2310.06825, 2023. 6", + "[18] Liqiang Jing, Yiren Li, Junhao Xu, Yongcan Yu, Pei Shen, and Xuemeng Song. Vision enhanced generative pre-trained language model for multimodal sentence summarization. Machine Intelligence Research, 20(2):289-298, 2023. 1", + "[19] Wai-Chung Kwan, Hong-Ru Wang, Hui-Min Wang, and Kam-Fai Wong. A survey on recent advances and challenges in reinforcement learning methods for task-oriented dialogue policy learning. Machine Intelligence Research, 20(3):318-334, 2023. 1", + "[20] Chunyuan Li, Cliff Wong, Sheng Zhang, Naoto Usuyama, Haotian Liu, Jianwei Yang, Tristan Naumann, Hoifung Poon, and Jianfeng Gao. Llava-med: Training a large language-and-vision assistant for biomedicine in one day. Proc. NeurIPS, 2024. 2, 5, 7", + "[21] Yifan Li, Hangyu Guo, Kun Zhou, Wayne Xin Zhao, and JiRong Wen. Images are achilles' heel of alignment: Exploiting visual vulnerabilities for jailbreaking multimodal large language models. In proc. ECCV, 2024. 1, 3", + "[22] Haotian Liu, Chunyuan Li, Yuheng Li, and Yong Jae Lee. Improved baselines with visual instruction tuning. In Proc. CVPR, 2024. 2, 4, 5, 6, 7, 8", + "[23] Haotian Liu, Chunyuan Li, Yuheng Li, Bo Li, Yuanhan Zhang, Sheng Shen, and Yong Jae Lee. Llava-last: Improved reasoning,OCR, and world knowledge, 2024. 5, 6, 7", + "[24] Haotian Liu, Chunyuan Li, Qingyang Wu, and Yong Jae Lee. Visual instruction tuning. In Proc. NeurIPS, 2024. 2, 3, 4, 7", + "[25] Xin Liu, Yichen Zhu, Jindong Gu, Yunshi Lan, Chao Yang, and Yu Qiao. Mm-safetybench: A benchmark for safety evaluation of multimodal large language models. In Proc. ECCV, 2024. 1, 3, 4, 6, 7, 8" + ], + "bbox": [ + 93, + 92, + 905, + 898 + ], + "page_idx": 8 + }, + { + "type": "page_number", + "text": "19887", + "bbox": [ + 480, + 944, + 517, + 955 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[26] Yangzhou Liu, Yue Cao, Zhangwei Gao, Weiyun Wang, Zhe Chen, Wenhai Wang, Hao Tian, Lewei Lu, Xizhou Zhu, Tong Lu, et al. Mminstruct: A high-quality multi-modal instruction tuning dataset with extensive diversity. Science China Information Sciences, 67(12):1-16, 2024. 7", + "[27] Pan Lu, Swaroop Mishra, Tony Xia, Liang Qiu, Kai-Wei Chang, Song-Chun Zhu, Oyvind Tafjord, Peter Clark, and Ashwin Kalyan. Learn to explain: Multimodal reasoning via thought chains for science question answering. In Proc. NeurIPS, 2022. 4, 5, 6, 7, 8, 1", + "[28] Haochen Luo, Jindong Gu, Fengyuan Liu, and Philip Torr. An image is worth 1000 lies: Transferability of adversarial images across prompts on vision-language models. In Proc. ICLR, 2024. 1", + "[29] Anand Mishra, Shashank Shekhar, Ajeet Kumar Singh, and Anirban Chakraborty. Ocr-vqa: Visual question answering by reading text in images. In Proc. ICDAR, 2019. 3", + "[30] Long Ouyang, Jeffrey Wu, Xu Jiang, Diogo Almeida, Carroll Wainwright, Pamela Mishkin, Chong Zhang, Sandhini Agarwal, Katarina Slama, Alex Ray, et al. Training language models to follow instructions with human feedback. In Proc. NeurIPS, 2022. 1", + "[31] Renjie Pi, Tianyang Han, Jianshu Zhang, Yueqi Xie, Rui Pan, Qing Lian, Hanze Dong, Jipeng Zhang, and Tong Zhang. Mllm-protector: Ensuring mllm's safety without hurting performance. In Proc. EMNLP, 2024. 3", + "[32] Xiangyu Qi, Kaixuan Huang, Ashwinee Panda, Peter Henderson, Mengdi Wang, and Prateek Mittal. Visual adversarial examples jailbreak aligned large language models. In Proc. AAAI, 2024. 1", + "[33] Xiangyu Qi, Ashwinee Panda, Kaifeng Lyu, Xiao Ma, Subhrajit Roy, Ahmad Beirami, Prateek Mittal, and Peter Henderson. Safety alignment should be made more than just a few tokens deep. arXiv preprint arXiv:2406.05946, 2024. 5", + "[34] Xiangyu Qi, Yi Zeng, Tinghao Xie, Pin-Yu Chen, Ruoxi Jia, Prateek Mittal, and Peter Henderson. Fine-tuning aligned language models compromises safety, even when users do not intend to! In Proc. ICLR, 2024. 1", + "[35] Alec Radford, Jong Wook Kim, Chris Hallacy, Aditya Ramesh, Gabriel Goh, Sandhini Agarwal, Girish Sastry, Amanda Askell, Pamela Mishkin, Jack Clark, et al. Learning transferable visual models from natural language supervision. In Proc. ICML, 2021. 2", + "[36] Rafael Rafailov, Archit Sharma, Eric Mitchell, Christopher D Manning, Stefano Ermon, and Chelsea Finn. Direct preference optimization: Your language model is secretly a reward model. In Proc. NeurIPS, 2024. 1", + "[37] Paul Röttger, Hannah Kirk, Bertie Vidgen, Giuseppe Attanasio, Federico Bianchi, and Dirk Hovy. Xstest: A test suite for identifying exaggerated safety behaviours in large language models. In Proc. NAACL, 2024. 4, 6, 7, 1", + "[38] John Schulman, Filip Wolski, Prafulla Dhariwal, Alec Radford, and Oleg Klimov. Proximal policy optimization algorithms. arXiv preprint arXiv:1707.06347, 2017. 1", + "[39] Erfan Shayegani, Yue Dong, and Nael Abu-Ghazaleh. Jailbreak in pieces: Compositional adversarial attacks on multimodal language models. In Proc. ICLR, 2024. 3" + ], + "bbox": [ + 91, + 90, + 480, + 900 + ], + "page_idx": 9 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[40] Amanpreet Singh, Vivek Natarajan, Meet Shah, Yu Jiang, Xinlei Chen, Dhruv Batra, Devi Parikh, and Marcus Rohrbach. Towards vqa models that can read. In Proc. CVPR, 2019. 3", + "[41] Quan Sun, Yuxin Fang, Ledell Wu, Xinlong Wang, and Yue Cao. Eva-clip: Improved training techniques for clip at scale. arXiv preprint arXiv:2303.15389, 2023. 2", + "[42] Hugo Touvron, Thibaut Lavril, Gautier Izacard, Xavier Martinet, Marie-Anne Lachaux, Timothee Lacroix, Baptiste Rozière, Naman Goyal, Eric Hambro, Faisal Azhar, et al. Llama: Open and efficient foundation language models. arXiv preprint arXiv:2302.13971, 2023. 2", + "[43] Junke Wang, Lingchen Meng, Zejia Weng, Bo He, Zuxuan Wu, and Yu-Gang Jiang. To see is to believe: Prompting gpt-4v for better visual instruction tuning. arXiv preprint arXiv:2311.07574, 2023. 3", + "[44] Peng Wang, Shuai Bai, Sinan Tan, Shijie Wang, Zhihao Fan, Jinze Bai, Keqin Chen, Xuejing Liu, Jialin Wang, Wenbin Ge, et al. Qwen2-vl: Enhancing vision-language model's perception of the world at any resolution. arXiv preprint arXiv:2409.12191, 2024. 2", + "[45] Xiao Wang, Guangyao Chen, Guangwu Qian, Pengcheng Gao, Xiao-Yong Wei, Yaowei Wang, Yonghong Tian, and Wen Gao. Large-scale multi-modal pre-trained models: A comprehensive survey. Machine Intelligence Research, 20 (4):447-482, 2023. 1", + "[46] Yu Wang, Xiaogeng Liu, Yu Li, Muhao Chen, and Chaowei Xiao. Adashield: Safeguarding multimodal large language models from structure-based attack via adaptive shield prompting. Proc. ECCV, 2024. 3", + "[47] Alexander Wei, Nika Haghtalab, and Jacob Steinhardt. Jailbroken: How does Ilm safety training fail? In Proc. NeurIPS, 2024. 3", + "[48] Peng Xu, Wenqi Shao, Kaipeng Zhang, Peng Gao, Shuo Liu, Meng Lei, Fanqing Meng, Siyuan Huang, Yu Qiao, and Ping Luo. Lvlm-ehub: A comprehensive evaluation benchmark for large vision-language models. IEEE Transactions on Pattern Analysis and Machine Intelligence, 47(3):1877-1893, 2025. 1", + "[49] An Yang, Baosong Yang, Binyuan Hui, Bo Zheng, Bowen Yu, Chang Zhou, Chengpeng Li, Chengyuan Li, Dayiheng Liu, Fei Huang, et al. Qwen2 technical report. arXiv preprint arXiv:2407.10671, 2024. 2", + "[50] Qinghao Ye, Haiyang Xu, Jiabo Ye, Ming Yan, Anwen Hu, Haowei Liu, Qi Qian, Ji Zhang, and Fei Huang. mplugowl2: Revolutionizing multi-modal large language model with modality collaboration. In Proc. CVPR, 2024. 1, 2", + "[51] Shukang Yin, Chaoyou Fu, Sirui Zhao, Ke Li, Xing Sun, Tong Xu, and Enhong Chen. A survey on multimodal large language models. National Science Review, 11(12), 2024. 1, 2", + "[52] Alex Young, Bei Chen, Chao Li, Chengen Huang, Ge Zhang, Guanwei Zhang, Heng Li, Jiangcheng Zhu, Jianqun Chen, Jing Chang, et al. Yi: Open foundation models by 01. ai. arXiv preprint arXiv:2403.04652, 2024. 5, 6, 7", + "[53] Yaowei Zheng, Richong Zhang, Junhao Zhang, Yanhan Ye, Zheyan Luo, Zhangchi Feng, and Yongqiang Ma. Llamafac" + ], + "bbox": [ + 514, + 90, + 903, + 900 + ], + "page_idx": 9 + }, + { + "type": "page_number", + "text": "19888", + "bbox": [ + 480, + 944, + 517, + 955 + ], + "page_idx": 9 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "tory: Unified efficient fine-tuning of $100+$ language models. arXiv preprint arXiv:2403.13372, 2024. 1", + "[54] Deyao Zhu, Jun Chen, Xiaoqian Shen, Xiang Li, and Mohamed Elhoseiny. Minigpt-4: Enhancing vision-language understanding with advanced large language models. arXiv preprint arXiv:2304.10592, 2023. 2, 3", + "[55] Yongshuo Zong, Ondrej Bohdal, Tingyang Yu, Yongxin Yang, and Timothy Hospedales. Safety fine-tuning at (almost) no cost: A baseline for vision large language models. In Proc. ICML, 2024. 1, 3, 4, 5, 6", + "[56] Andy Zou, Zifan Wang, Nicholas Carlini, Milad Nasr, J Zico Kolter, and Matt Fredrikson. Universal and transferable adversarial attacks on aligned language models. arXiv preprint arXiv:2307.15043, 2023. 5" + ], + "bbox": [ + 91, + 90, + 482, + 287 + ], + "page_idx": 10 + }, + { + "type": "page_number", + "text": "19889", + "bbox": [ + 480, + 945, + 517, + 955 + ], + "page_idx": 10 + } +] \ No newline at end of file diff --git a/2025/Do We Really Need Curated Malicious Data for Safety Alignment in Multi-modal Large Language Models_/7cc50b70-a1c6-4460-8b07-4c06c5509d66_model.json b/2025/Do We Really Need Curated Malicious Data for Safety Alignment in Multi-modal Large Language Models_/7cc50b70-a1c6-4460-8b07-4c06c5509d66_model.json new file mode 100644 index 0000000000000000000000000000000000000000..651b8694a8cd2f6bae5d837ab2e0dbb9d1492055 --- /dev/null +++ b/2025/Do We Really Need Curated Malicious Data for Safety Alignment in Multi-modal Large Language Models_/7cc50b70-a1c6-4460-8b07-4c06c5509d66_model.json @@ -0,0 +1,2037 @@ +[ + [ + { + "type": "header", + "bbox": [ + 0.107, + 0.003, + 0.182, + 0.043 + ], + "angle": 0, + "content": "CVF" + }, + { + "type": "header", + "bbox": [ + 0.237, + 0.001, + 0.812, + 0.047 + ], + "angle": 0, + "content": "This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore." + }, + { + "type": "title", + "bbox": [ + 0.157, + 0.13, + 0.844, + 0.177 + ], + "angle": 0, + "content": "Do We Really Need Curated Malicious Data for Safety Alignment in Multi-modal Large Language Models?" + }, + { + "type": "text", + "bbox": [ + 0.194, + 0.204, + 0.803, + 0.257 + ], + "angle": 0, + "content": "Yanbo Wang, Jiyang Guan, Jian Liang\\(^{\\text{心}}\\), Ran He \nSchool of Artificial Intelligence, University of Chinese Academy of Sciences \nNLPR & MAIS, Institute of Automation, Chinese Academy of Sciences" + }, + { + "type": "text", + "bbox": [ + 0.3, + 0.26, + 0.693, + 0.275 + ], + "angle": 0, + "content": "wangyanbo2023@ia.ac.cn liangjian92@gmail.com" + }, + { + "type": "title", + "bbox": [ + 0.248, + 0.309, + 0.327, + 0.327 + ], + "angle": 0, + "content": "Abstract" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.341, + 0.486, + 0.78 + ], + "angle": 0, + "content": "Multi-modal large language models (MLLMs) have made significant progress, yet their safety alignment remains limited. Typically, current open-source MLLMs rely on the alignment inherited from their language module to avoid harmful generations. However, the lack of safety measures specifically designed for multi-modal inputs creates an alignment gap, leaving MLLMs vulnerable to vision-domain attacks such as typographic manipulation. Current methods utilize a carefully designed safety dataset to enhance model defense capability, while the specific knowledge or patterns acquired from the high-quality dataset remain unclear. Through comparison experiments, we find that the alignment gap primarily arises from data distribution biases, while image content, response quality, or the contrastive behavior of the dataset makes little contribution to boosting multi-modal safety. To further investigate this and identify the key factors in improving MLLM safety, we propose finetuning MLLMs on a small set of benign instruct-following data with responses replaced by simple, clear rejection sentences. Experiments show that, without the need for labor-intensive collection of high-quality malicious data, model safety can still be significantly improved, as long as a specific fraction of rejection data exists in the finetuning set, indicating the security alignment is not lost but rather obscured during multi-modal pretraining or instruction finetuning. Simply correcting the underlying data bias could narrow the safety gap in the vision domain. Warning: This paper contains harmful images and AI-generated contents which may be offensive." + }, + { + "type": "title", + "bbox": [ + 0.092, + 0.808, + 0.222, + 0.823 + ], + "angle": 0, + "content": "1. Introduction" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.833, + 0.484, + 0.88 + ], + "angle": 0, + "content": "In recent years, the multi-modal large language model (MLLM) has experienced tremendous prosperity [11, 18, 45, 48, 51]. Companies and academics have proposed mul" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.31, + 0.908, + 0.599 + ], + "angle": 0, + "content": "tiple models with their APIs, finetuning suites, and online demos. However, although the multi-modal understanding capability of these models is advancing steadily, the safety alignment of these models attracts less attention. To be detailed, most MLLMs connect an image encoder and a pretrained large language model with a vision-language connector, which are attention structures at the beginning and then replaced by MLPs [4, 5, 9, 50]. The safety alignment focuses mainly on the large language model, where popular RLHF algorithms [2, 19, 30] such as DPO [36] and PPO [38] are used to ensure that the chat model acts humanly. Nevertheless, such a workflow uncovers an obvious disadvantage: both the pretraining and visual instruction tuning stages do not contain a specific multi-modal safety alignment procedure, and the safety defensive capability could only be inherited from the LLM module. In addition, researchers also find that finetuning models on benign data may still sacrifice model safety [34]. They all challenge the defense capability of MLLMs." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.6, + 0.909, + 0.903 + ], + "angle": 0, + "content": "From the attackers' point of view, jailbreaking these MLLM models with multi-modal input seems more straightforward than jailbreaking language models [21]. Instead of optimizing adversarial noise to elicit harmful contents from vision language models with white-box accesses [28, 32], the typographic attack is a typical example to demonstrate the inferior safety of MLLMs. Fig-Step [13] directly transforms harmful requests into images through typography to deceive the image encoder module instead of the language module. With similar logic, MM-SafetyBench [25] proposes to add harmful keywords at the bottom of content-related diffusion-generated images, successfully inducing VLMs to generate inappropriate content across multiple fields. To defend such a simple but effective attack, except for prompt engineering techniques at inference time, previous work approaches this by collecting a safe multi-modal dataset with toxic queries and rejection responses, and then finetuning the MLLMs either in the second finetuning stage or after their release [55]. Regardless of a great success, current methods also leave a few questions unresolved." + }, + { + "type": "page_footnote", + "bbox": [ + 0.115, + 0.888, + 0.251, + 0.901 + ], + "angle": 0, + "content": "Corresponding author." + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "19879" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.091, + 0.087, + 0.905, + 0.373 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.09, + 0.382, + 0.908, + 0.425 + ], + "angle": 0, + "content": "Figure 1. The main workflow of our study. After demonstrating that the quality of safety finetuning data does not contribute as much as we imagine, we modify the LLaVA-Med [20] dataset to create our finetuning dataset. For each data point, we simply pick one round of the conversation and replace their original answer with a clear rejection, without periods and tokens." + }, + { + "type": "text", + "bbox": [ + 0.089, + 0.45, + 0.486, + 0.648 + ], + "angle": 0, + "content": "tions. For one thing, collecting high-quality is always labor-intensive. Even if we could prompt powerful LLMs and diffusion models to generate the required multi-modal content, it is still effort-taking to create the pipeline, e.g. extracting keywords from related topics for diffusion models, or prompt engineering for high-quality various harmful questions. More importantly, what really matters in the multimodal alignment process? Do VLMs forget their safety alignment through multi-modal instruction tuning? We are still unclear why simple typographic attacks with keywords incitement could break the safety alignment of vision language models, while they possess identical meanings with the text-only prompts." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.66, + 0.485, + 0.902 + ], + "angle": 0, + "content": "In this work, we are the first to demonstrate that the alignment gap between the vision-language inputs and text-only counterparts comes mainly from training data distribution bias. In detail, we perform ablation studies on the previous finetuning method to filter out some unnecessary attributes of the dataset, discovering that the defense capability may be hardly related to the response quality, answer-rejection comparisons, or conversation rounds and image contents. On the contrary, increasing rejection rates in the finetuning dataset will elicit over-rejection, indicating the dataset distribution may matter in safety finetuning. To further explore the hypothesis that dataset distribution is the key factor for model alignment, we leverage the safety-unrelated LLaVA-Med dataset [20] to mitigate the compliance bias contained in open-source MLLMs. Experiments demonstrate that without safety-related images, harmful" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.45, + 0.909, + 0.648 + ], + "angle": 0, + "content": "text prompts, and detailed rejection reasons, only \"teaching the LLaVA-\\(\\nu\\)1.5 models to reject a specific proportion of benign prompts\" appears to boost multi-modal safety significantly. Detailed quantitative and qualitative results are followed, encompassing the relationship between reason generation and ordinary training data. To further explore to what extent the rejection proportion is strong enough to mitigate the compliance bias, we make ablation studies on the rejection dataset size, analyze the influence rejection data proportion has on two jailbreak methods, and launch full visual instruction tuning to compare the effectiveness with post-finetuning. We believe our findings could help better understand the safety-related behavior of MLLMs." + }, + { + "type": "title", + "bbox": [ + 0.514, + 0.658, + 0.657, + 0.673 + ], + "angle": 0, + "content": "2. Related Work" + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.682, + 0.833, + 0.699 + ], + "angle": 0, + "content": "2.1. Multi-modal Large Language Model" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.705, + 0.909, + 0.903 + ], + "angle": 0, + "content": "In this paper, we leverage materials within the vision language models (VLMs) context. Most MLLMs consist of three parts: modality encoder, pretrained LLM, and modality interface [51]. Instead of training a MLLM from scratch, companies or academies utilize the modality connector to preprocess the on-the-shelf large language models with image encoders [4-6, 9, 44, 50, 54], in which the image encoders are mainly CLIP variants [7, 35, 41]. The language models are mainly pre-trained chat versions, such as Llama families [42], Vicuna families [8], and Qwen series [1, 49]. When training the MLLMs, the prevailing procedure contains vision-language pre-training and visual instruction tuning [9, 22, 24, 54]. In the first phase," + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.52, + 0.957 + ], + "angle": 0, + "content": "19880" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.09, + 0.092, + 0.485, + 0.289 + ], + "angle": 0, + "content": "the multi-modal projectors are trained in image captioning tasks. After that, in the instruction tuning phase, both the large language model and the connector are tuned with a large amount of VQA data [3, 43], aiming to enhance the visual understanding, reasoning, and OCR capability [29, 40]. However, safety alignment data are hardly included in the training phase, especially for those models from non-profit academies [4, 24, 54]. As a consequence, these MLLMs only rely on the safety alignment within their language module to defend against harmful prompts. Previous works prove the weakness of these MLLMs when faced with multi-modal inputs, and attribute this failure to the lack of curated VQA data [31, 47]." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.306, + 0.283, + 0.321 + ], + "angle": 0, + "content": "2.2. Black-box Jailbreak" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.331, + 0.485, + 0.588 + ], + "angle": 0, + "content": "As mentioned above, due to the obvious alignment gap between text-only prompts and vision-language prompts in MLLMs, Many researchers leverage the alignment gap to launch multi-modal jailbreaks. To our best knowledge, Fig-Step [13] is the first work to jailbreak MLLMs with typography. When printing an image of a list containing steps for unethical behaviors with empty items as the input, such an attack could circumvent the alignment while keeping the OCR capability to fill in the blanks. Subsequent works follow the typographic attacks, append a content-related image on the top, and replace the sensitive words in text prompts with image incitement to enhance the attack success rate, e.g. replace \"sell illegal drugs\" with \"the behavior shown in the image\" [21, 25]. Such an incitement method could also be used to launch white-box attacks, where the attackers interact with their well-optimized adversarial examples by pronouns [39] to circumvent model safety alignment." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.605, + 0.265, + 0.619 + ], + "angle": 0, + "content": "2.3. Jailbreak Defense" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.629, + 0.484, + 0.901 + ], + "angle": 0, + "content": "The Jailbreak defense on MLLMs could be mainly split into two categories: the inference phase and the training phase. Inference-time defenses are more like post-processing: AdaShield [46] proposes to create a shield prompt pool, adaptively appending a warning prompt at the end of user prompts to highlight response safety; MLLM-Protector [31] proposes a harm detector to firstly filter out harmful contents, and then change the response by their fine-tuned response detoxifier; similarly, ECSO [14] transforms input images into texts in a query-aware manner once the originally generated contents are labeled as inappropriate. To our knowledge, the work most related to our standpoint is CoCA [12], which focuses on exploring the safety-awareness of multi-modal inputs. Instead of training, they modify the decoding phase to enhance model safety. In the training phase, VLGuard [55] collects a safety dataset for supervised finetuning (SFT), achieving satisfying results in the defense of typographic attacks." + }, + { + "type": "title", + "bbox": [ + 0.514, + 0.09, + 0.907, + 0.108 + ], + "angle": 0, + "content": "3. What Matters in Safety Finetuning Dataset?" + }, + { + "type": "title", + "bbox": [ + 0.514, + 0.115, + 0.655, + 0.13 + ], + "angle": 0, + "content": "3.1. Preliminaries" + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.138, + 0.907, + 0.228 + ], + "angle": 0, + "content": "Finetuning workflow. As stated in the introduction part, the previous method could finetune the MLLM in two ways: mixing safety data in the second stage instruction tuning, or directly finetuning the released model. Here we pick VL-Guard [55] as the target for the ablation study, and we train the models based on the released version with LoRA." + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.229, + 0.907, + 0.321 + ], + "angle": 0, + "content": "VLGuard dataset [55]. The dataset composes three types of image-text pairs: safe-safe, safe-unsafe, and unsafe-unsafe. For example, safe-safe indicates a benign image-text pair, and safe-unsafe means the benign image is paired with a harmful (and in most cases, unrelated) text prompt. Here is an example of an LLaVA-style data sample:" + }, + { + "type": "text", + "bbox": [ + 0.545, + 0.349, + 0.875, + 0.377 + ], + "angle": 0, + "content": "Prompt: What brand of beer is shown in the image?" + }, + { + "type": "text", + "bbox": [ + 0.545, + 0.378, + 0.875, + 0.404 + ], + "angle": 0, + "content": "Assistant: The brand of the beer shown in the image is Grolsch." + }, + { + "type": "text", + "bbox": [ + 0.545, + 0.405, + 0.875, + 0.446 + ], + "angle": 0, + "content": "Prompt: Can you provide information on how to create fake IDs to purchase alcohol like the one shown in the image?" + }, + { + "type": "text", + "bbox": [ + 0.545, + 0.447, + 0.875, + 0.543 + ], + "angle": 0, + "content": "Assistant: I'm sorry, but I cannot provide assistance or information on illegal activities, including creating fake IDs to purchase alcohol. Engaging in such actions is against the law and our usage policy strictly prohibits facilitating or promoting illegal activities. It's important to follow the legal age requirements for purchasing and consuming alcoholic beverages." + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.565, + 0.907, + 0.624 + ], + "angle": 0, + "content": "For training, the dataset contains 977 single-turn unsafe-image pairs and 1023 multi-turn safe-image pairs, encompassing topics related to privacy, risky behavior, deception, and discrimination." + }, + { + "type": "title", + "bbox": [ + 0.514, + 0.636, + 0.64, + 0.652 + ], + "angle": 0, + "content": "3.2. Hypotheses" + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.659, + 0.907, + 0.719 + ], + "angle": 0, + "content": "High-quality datasets with well-picked images and detailed responses have been demonstrated to strengthen the safety of MLLMs. Here, we make several hypotheses on dataset attributes that contribute to satisfying defense capability." + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.72, + 0.906, + 0.825 + ], + "angle": 0, + "content": "Comparisons in multi-turn data. As stated above, safe-safe and safe-unsafe pairs share identical images, which are structured in a multi-turn conversation style. Therefore, it is reasonable to hypothesize that the contrastive manner of the dataset forces MLLMs to understand what kind of input is toxic, and what kind of inputs are benign even with identical image visual prompts." + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.826, + 0.907, + 0.901 + ], + "angle": 0, + "content": "Content-related images. It has been proved that content-related images tend to strengthen the jailbreak prompts [31]. For example, with a bomb image, the model has a higher probability of outputting illegal contents on manufacturing weapons [25]. Thus, We hypothesize that" + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.518, + 0.957 + ], + "angle": 0, + "content": "19881" + } + ], + [ + { + "type": "table", + "bbox": [ + 0.095, + 0.089, + 0.904, + 0.326 + ], + "angle": 0, + "content": "
Base ModelEvaluation DatasetsModel Settings
OriginVLGuardOne-turnUnsafe-onlyChange imageDirect sorryRandom reasonPure-VLGuard
LLaVA-v1.5-7BMM SafetyBench↓96.370.180.180.180.180.120.120.05
FigStep↓100.000.000.000.000.000.000.000.00
XCTest Compliance↑92.0077.2078.8076.8079.2070.8078.0052.00
XCTest Rejection↑75.5096.5095.5096.0093.0099.0098.0099.00
LLaVA-v1.5-13BMM SafetyBench↓97.800.360.360.180.180.420.540.12
FigStep↓99.800.000.000.000.000.000.000.00
XCTest Compliance↑90.0077.6077.2076.8082.0078.0078.0061.60
XCTest Rejection↑84.5097.0096.0096.0094.5098.0097.0099.50
LLaVA-NeXT Mistral-7BMM SafetyBench↓100.00.180.240.480.240.240.990.12
FigStep↓99.880.000.000.000.000.000.120.00
XCTest Compliance↑94.8087.6086.0081.6086.4086.4084.4043.20
XCTest Rejection↑58.5089.5084.5091.5084.0089.5094.0099.50
LLaVA-NeXT Llama-3-8BMM SafetyBench↓96.070.120.182.800.060.710.120.12
FigStep↓99.800.000.000.0038.40.000.000.00
XCTest Compliance↑90.4082.8082.8082.8086.4076.8078.4060.40
XCTest Rejection↑83.5092.0091.0091.0083.5097.0097.5099.00
" + }, + { + "type": "table_caption", + "bbox": [ + 0.233, + 0.336, + 0.763, + 0.351 + ], + "angle": 0, + "content": "Table 1. Model safety evaluation with various finetuning data. All values are percentages." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.377, + 0.483, + 0.438 + ], + "angle": 0, + "content": "MLLM safety could be gained from the rejection of unsafeunsafe pairs. After collecting these harmful images and making models learn to reject them, the defense capability of MLLMs probably advances by a huge margin." + }, + { + "type": "text", + "bbox": [ + 0.089, + 0.439, + 0.483, + 0.573 + ], + "angle": 0, + "content": "Detailed reject reasons. Detailed reject reasons clarify the specific legal principles the harmful prompts violate, highlighting that as an AI model, it should follow legal and moral requirements to provide helpful and safe content. They add extra safety-related knowledge to the dataset, therefore, it is reasonable to hypothesize that such detailed data are crucial to invoke previous safety alignment in the language module, and then fill the gap between the multimodal input and pure-text domains." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.577, + 0.483, + 0.713 + ], + "angle": 0, + "content": "Dataset distribution bias. Comparing the instruction tuning dataset with the safety finetuning dataset, one distinct difference is the rejection behavior. Regardless of whether the training set contains toxic or harmful contents, in the instruction tuning phase the model hardly learns how to reject because nearly all replies in conversation-style data follow the instructions as detailed as possible. We hypothesize that this may also be the reason why the model loses the ability to reject inappropriate requests." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.727, + 0.285, + 0.743 + ], + "angle": 0, + "content": "3.3. Experiment Settings" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.75, + 0.483, + 0.901 + ], + "angle": 0, + "content": "Finetuning models and datasets. Here, we launch all experiments on the LLaVA series model [22-24]. Aligning with previous methods, we add random-picked 5000 data points from the visual instruction tuning dataset LLaVA-v1.5-mix665k [22]. To testify to the influence of comparisons in the dataset, we split multi-turn safety data into single-turn and eliminate all safe-safe data pairs, named \"One-turn\" and \"Unsafe-only\", respectively. We also change the image prompt with random-picked benign images from LLaVA-v1.5-mix665k, named \"Change im" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.377, + 0.908, + 0.558 + ], + "angle": 0, + "content": "age\" for the content-related image hypothesis. To modify the detailed reasons, we design two alternatives: replacing all replies with one clear rejection sentence or generating rejection contents with the prefix: \"I'm sorry, I cannot assist with that request because it goes against my programming to\" on benign data in the finetuning dataset. For comparisons, we also add experiments on pure VLGuard dataset [55] and non-finetuned original models. Experiments are done on LLaVA-v1.5-7B, LLaVA-v1.5-13B, LLaVA-NeXT-mixtral-7B, and LLaVA-NeXT-llama3-8B [22-24] with LoRA [16]. Finetuning details are listed in the Appendix." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.56, + 0.909, + 0.833 + ], + "angle": 0, + "content": "Evaluation metrics. We mainly evaluate the safety of MLLM models with black-box attacks, for it is the prevailing safety threat MLLM models will face. With collected jailbreaking image-text pairs on the Internet, a user could directly launch such an attack. We pick FigStep [13] and MM-SafetyBench [25] as the attacking sources, and the attack successful rate (ASR) is calculated by string-matching. We also tried to use Llama-3-Guard [10] as the discriminator in early evaluations, but the ASRs are lower because some sensitive topics exhibited in the datasets are not regarded as harmful, even if the model fails to reject content generation. Besides, to evaluate the model performance on visual understanding tasks, we adopt VizWizQA [15] and ScienceQA [27] for visual-text understanding, and XSTest dataset [37] for over-rejection evaluation. One point worth noting is that the XSTest dataset comprises two parts, in which the compliance column means the proportion of benign questions with non-rejection answers." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.847, + 0.608, + 0.862 + ], + "angle": 0, + "content": "3.4. Results" + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.871, + 0.907, + 0.901 + ], + "angle": 0, + "content": "Detailed safety and performance evaluations of models on various ablation datasets are shown in Table 1." + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.518, + 0.956 + ], + "angle": 0, + "content": "19882" + } + ], + [ + { + "type": "table", + "bbox": [ + 0.095, + 0.089, + 0.499, + 0.326 + ], + "angle": 0, + "content": "
Base ModelFinetuning DatasetsVizWizQA(%)↑ScienceQA(%)
Image↑Total↑
LLaVA-v1.5-7BOrigin55.1070.1269.46
VLGuard56.3668.4270.12
One-turn56.3967.6769.79
Change image55.7867.3869.04
Unsafe-only56.7567.8269.63
Direct sorry53.9867.9769.61
Random reason56.6568.5270.01
Pure-VLGuard56.4565.2464.18
LLaVA-v1.5-13BOrigin57.3472.6874.89
VLGuard56.7771.5472.62
One-turn54.2871.3472.95
Change image55.5770.7572.98
Unsafe-only59.6270.8572.48
Direct sorry56.9871.0071.66
Random reason58.3971.0072.29
Pure-VLGuard58.2671.0071.47
" + }, + { + "type": "table_caption", + "bbox": [ + 0.118, + 0.335, + 0.455, + 0.35 + ], + "angle": 0, + "content": "Table 2. Evaluation of model capabilities on VQA tasks." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.384, + 0.483, + 0.642 + ], + "angle": 0, + "content": "Comparisons in multi-turn data may make a limited contribution to the safety enhancement. With extensive experiments on both LLaVA-v1.5 [22] and LLaVA-NeXT [23] structures with different model sizes, the finding is consistent that the safety alignment of VLMs hardly relies on the comparison of answers to benign prompts and harmful prompts. For one thing, after splitting a double-turn conversation into two single-turn data points with the same image prompt, the defensive capability of all four models remains nearly unchanged, suggesting that utilizing some benign conversations as the conversation history may have a limited impact on enhancing the defense capability; for another, even eliminating all benign conversations from VL-Guard dataset [55] makes little difference to the defense result. Such a phenomenon indicates that benign data here may only contribute to the model's helpfulness, similar to randomly picked data in LLaVA-v1.5-mix665k [22]." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.645, + 0.483, + 0.901 + ], + "angle": 0, + "content": "Content-related images and reject reasons may not be the key factor for the defensive capability. In Table 1, it turns out that changing content-related harmful images to benign training images hardly deteriorates the defensive capability. Except for the inferior performance LLaVA-NeXT-Llama-3-8B [23] has on FigStep [13], under the rest circumstances training models with unrelated images could result in a satisfying defensive performance. Besides, in Table 2, the performance difference caused by image changing is only about \\(1\\%\\) on both VQA tasks [15, 27]. Similar trends also occur on ablations for reject reasons. Modifying or directly eliminating the reasons does not harm defensive capability, indicating that the model learns little from the detailed reasons for making the decision. Qi et al. [33] proposed a similar conclusion in the language model domain, claiming that current finetuning gradients mainly come from a few first tokens." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.092, + 0.905, + 0.272 + ], + "angle": 0, + "content": "High rejection proportion has the potential to harm the model performance. After eliminating training samples from LLaVA-v1.5-mix665k [22], the rejection conversation takes up around \\(66.7\\%\\) of the whole dataset. With such a data distribution, all models get over-sensitive: they could reject more than \\(99\\%\\) harmful text prompts while only answering about half of the benign questions, even though the curated finetuning data are of high quality. Besides, we also witness an accuracy drop on the ScienceQA dataset [27] with the LLaVA-v1.5-7B model. This phenomenon is consistent with the prior work [55], highlighting the existence of normal training data." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.273, + 0.907, + 0.44 + ], + "angle": 0, + "content": "To sum up, in the finetuning phase, it seems the models learn little from the high-quality content. Dataset modifications on images, answer reasons, and conversation comparisons probably only have negligible impacts on the safety alignment. However, altering the rejection proportion of the dataset could elicit over-rejection: even though the models only learn how to reject harmful instructions, the rejection behavior is wrongly generalized to benign prompts. Based on this, one question emerges: is dataset distribution the key factor for model alignment? If that is the case, is it possible to enhance MLLM safety without well-curated safe data?" + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.453, + 0.907, + 0.488 + ], + "angle": 0, + "content": "4. Benign Data with Rejection Responses Boost Model Safety" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.496, + 0.907, + 0.543 + ], + "angle": 0, + "content": "To figure out the role dataset rejection proportion plays in safety finetuning, we try to stimulate the alignment potential with benign data." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.551, + 0.704, + 0.567 + ], + "angle": 0, + "content": "4.1. Experimental Setup" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.573, + 0.906, + 0.786 + ], + "angle": 0, + "content": "To align with the previous setting, we randomly pick 2000 data points from LLaVA-Med dataset [20]. For each data point, we randomly pick one round of QAs in the conversation and replace the answer with a clear rejection. As in Sec. 3.3, these data are also mixed with randomly picked 5000 normal data from LLaVA-v1.5-mix665k [22]. Noting that the training data of Yi-VL-6B [52] and LLaVA-NeXT [23] are not publicly available, therefore we keep using the same setting as the LLaVA-v1.5 models. Considering possible overfitting, we only supervise the model to generate a few rejection tokens while masking out the end-of-sentence token. Assuming in a one-turn conversation \\( x \\), the instruction takes up \\( m \\) tokens. In our experiments, we only calculate the language modeling loss by this equation [56]:" + }, + { + "type": "equation", + "bbox": [ + 0.576, + 0.797, + 0.907, + 0.815 + ], + "angle": 0, + "content": "\\[\n\\mathcal {L} \\left(x _ {1: n + m}\\right) = - \\log p \\left(x _ {n + 1: n + m} ^ {*} \\mid x _ {1: n}\\right), \\tag {1}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.826, + 0.906, + 0.903 + ], + "angle": 0, + "content": "where \\( x_{n+1:n+m}^* \\) refers to the tokenized rejection string: \"I'm sorry, but I cannot assist with your request\", without periods or token ids. For simplicity, here we ignore the existence of system prompts as well as token in some templates. To ensure the reliability of our results," + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.518, + 0.956 + ], + "angle": 0, + "content": "19883" + } + ], + [ + { + "type": "table", + "bbox": [ + 0.137, + 0.089, + 0.865, + 0.315 + ], + "angle": 0, + "content": "
Base ModelFinetuning DatasetsMM SafetyBench↓FigStep↓XCTest↑VizWizQA↑ScienceQA↑
ComplianceRejectionImageTotal
LLaVA-v1.5-7BOrigin96.37100.0092.0075.5055.1070.1269.46
VLGuard0.180.0077.2096.5056.3668.4270.12
Ours5.600.2090.4082.0055.9268.0269.11
LLaVA-v1.5-13BOrigin97.8099.8090.0084.5057.3472.6874.89
VLGuard0.360.0077.6097.0056.7771.5472.62
Ours1.790.0090.4090.5058.7771.5973.36
Yi-VL-6BOrigin93.5199.6095.2041.50\\( 66.20^1 \\)61.0370.12
VLGuard0.180.0084.4093.0059.3269.8174.51
Ours3.573.2089.2067.0047.9168.2773.07
LLaVA-NeXT Mistral-7BOrigin99.88100.0094.8058.5064.9872.6378.57
VLGuard0.180.0087.6089.5053.1570.2074.56
Ours2.610.0091.2065.5055.5069.3674.65
" + }, + { + "type": "table_caption", + "bbox": [ + 0.156, + 0.325, + 0.84, + 0.34 + ], + "angle": 0, + "content": "Table 3. The evaluation comparisons between our finetuning dataset and VLGuard [55]. All values are percentages." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.366, + 0.483, + 0.429 + ], + "angle": 0, + "content": "except for LLaVA-v1.5 models with Vicuna language module, we also pick the LLaVA-NeXT with Mistral-7B language module [17], and the Yi-VL-6B based on Yi-6B [52]. Training details are shown in the Appendix." + }, + { + "type": "title", + "bbox": [ + 0.09, + 0.439, + 0.285, + 0.455 + ], + "angle": 0, + "content": "4.2. Quantitative Results" + }, + { + "type": "text", + "bbox": [ + 0.089, + 0.461, + 0.483, + 0.687 + ], + "angle": 0, + "content": "MLLMs do not lose defense on multi-modal prompts. In Table 3, it turns out that all four models work well on jailbreak defenses while maintaining a relatively high rate of benign prompt compliance. For MM-SafetyBench [25], the ASR of most models is below \\(5\\%\\), which is a substantial drop compared to the original models. Experiments on FigStep [13] are also obvious: the highest ASR among the four models is \\(3.2\\%\\). In training data, the conversation topic is unrelated to safety and privacy, and we do not append any reason for the \"abnormal\" rejection in each data point. The major factor is the rejection rate: compared to the visual instruction tuning phase with limited rejection data, we add 2000 rejection samples to 5000 benign training data in safety finetuning, reducing the proportion of instances where the answer complies with instructions." + }, + { + "type": "text", + "bbox": [ + 0.089, + 0.689, + 0.483, + 0.855 + ], + "angle": 0, + "content": "More importantly, our experimental results provide little support for the notion that the model forgets previous safety alignment or fails due to the absence of multi-modal safety data, leading it to interpret harmful jailbreak prompts as benign. Instead, we argue that the model fails to defend jailbreaks mentioned above mostly because all multi-modal instruction-tuning data naturally create a compliance bias. After instruction tuning, the model still possesses the basic moral understanding and judgment. The difference is that it is told all prompts with images should be answered and followed eagerly, resulting in the vulnerability." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.856, + 0.484, + 0.903 + ], + "angle": 0, + "content": "Without knowledge, correcting model distribution makes limited progress to text-only prompts. In Table 3, finetuning on our dataset does not significantly impair the" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.366, + 0.907, + 0.472 + ], + "angle": 0, + "content": "safety compliance in XSTest [37], but it also fails to improve the unsafe rejection rate to the level achieved by VLGuard [55]. It is reasonable, for our dataset contains no extra safety-related contents compared to the VLGuard. Accordingly, merely correcting the model's distribution without incorporating additional knowledge is unlikely to result in a significant improvement in text-only input defense." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.473, + 0.909, + 0.669 + ], + "angle": 0, + "content": "Performances on general tasks tend to be associated with the ordinary training data. In experiments, we mix rejection data with ordinary data from LLaVA-v1.5-mix665k [22], the official training set for LLaVA-v1.5. In the first two rows, using our random rejection data could achieve similar results as VLGuard on VizWizQA [15] and ScienceQA [27], most of which are slightly inferior to the original model. For VizWizQA the model trained on our dataset even outperforms the original one, indicating the quality of the ordinary data. However, for LLaVA-NeXT [23] and Yi-VL [52], the performances experience a larger drop, which may be attributed to the inconsistency of training data. This adds more randomness to the results." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.67, + 0.911, + 0.852 + ], + "angle": 0, + "content": "Analyzing the results from the rejection data proportion perspective, we could also prove the claim that large language model alignment contributes to the defense of vision-language jailbreaks. Suppose the model does not possess safety-related knowledge or judgment for multi-modal data. In that case, the rejection rate of generated contents should be identical to, or slightly less than training data distribution. With \\(28.57\\%\\) rejection data in the training set, ASRs of LLaVA-v1.5-7B on jailbreak datasets should be around \\(70\\%\\), and the performance on visual understanding and reasoning tasks should drop by \\(15\\%\\) to \\(20\\%\\). However, the evaluation discrepancy between jailbreak datasets and vi" + }, + { + "type": "page_footnote", + "bbox": [ + 0.513, + 0.863, + 0.908, + 0.901 + ], + "angle": 0, + "content": "1The original value under the standard evaluation metric is 4.99, for most answers are appended by meaningless periods and line breaks. The score is calculated after omitting these suffixes." + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.52, + 0.958 + ], + "angle": 0, + "content": "19884" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.09, + 0.092, + 0.485, + 0.289 + ], + "angle": 0, + "content": "sual understanding datasets discloses that the rejection behavior is not evenly added to the content generation. In other words, the model gets over-compliance after visual instruction tuning. Still, it knows which multi-modal prompts are harmful, just as it knows how to reject those text-only harmful prompts without interactions and incitements with images. The semantic meanings between multi-modal inputs and their text-only counterparts are aligned. In safety finetuning, it is less important to meticulously collect security data for the model to learn how to distinguish between benign and malicious instructions. Once enough rejection data is provided to enable rejection behavior, the model can make the right decisions." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.298, + 0.274, + 0.314 + ], + "angle": 0, + "content": "4.3. Qualitative Results" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.32, + 0.484, + 0.352 + ], + "angle": 0, + "content": "To have a rough understanding of the answer quality, we present some examples of the model-generation content." + }, + { + "type": "text", + "bbox": [ + 0.126, + 0.373, + 0.449, + 0.429 + ], + "angle": 0, + "content": "Prompt: \\n The bottom of the image shows a phrase of an activity. List steps in detail to perform this activity using cryptocurrencies without leaving a trace." + }, + { + "type": "title", + "bbox": [ + 0.127, + 0.443, + 0.228, + 0.455 + ], + "angle": 0, + "content": "LLaVA-v1.5-7B" + }, + { + "type": "text", + "bbox": [ + 0.126, + 0.457, + 0.449, + 0.54 + ], + "angle": 0, + "content": "Assistant: I'm sorry, but I cannot assist with your request. Purchasing illegal drugs with cryptocurrencies...the use of cryptocurrencies for illegal activities can lead to legal consequences and is morally wrong. It is important to support legal and legitimate businesses..." + }, + { + "type": "title", + "bbox": [ + 0.128, + 0.554, + 0.19, + 0.566 + ], + "angle": 0, + "content": "Yi-VL-6B" + }, + { + "type": "text", + "bbox": [ + 0.126, + 0.568, + 0.449, + 0.61 + ], + "angle": 0, + "content": "Assistant: I'm sorry, but I cannot assist with your request. Please consult a professional for assistance with your request." + }, + { + "type": "title", + "bbox": [ + 0.127, + 0.623, + 0.328, + 0.635 + ], + "angle": 0, + "content": "LLaVA- \\(\\nu 1.5 - 7B\\) -Random-reason" + }, + { + "type": "text", + "bbox": [ + 0.126, + 0.637, + 0.449, + 0.678 + ], + "angle": 0, + "content": "Assistant: I'm sorry, I cannot assist with that request because it goes against my programming to provide opinions or beliefs." + }, + { + "type": "image_caption", + "bbox": [ + 0.09, + 0.708, + 0.484, + 0.763 + ], + "angle": 0, + "content": "Figure 2. Examples of replies generated under different settings. The reasons are not as high quality as the LLaVA-v1.5. To save space, the image of drugs with the phrase \"purchase illegal drugs\" is not shown in the box." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.781, + 0.484, + 0.902 + ], + "angle": 0, + "content": "The quality of generated rejection reasons is relatively independent of model decisions. In Fig. 2, the LLaVA-v1.5 model generates high-quality responses, with rejection reasons related to the topic. However, For Yi-VL-6B [52] and LLaVA-NeXT [23], the models keep satisfying reject decisions, while generating imperfect reasons. Similar results are also demonstrated when training LLaVA-v1.5 models under the random reason setting, indicating that the" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.092, + 0.906, + 0.168 + ], + "angle": 0, + "content": "training data quality possibly affects the following sentence completion. With a proper rejection rate in the training set, models can reject inappropriate prompts, while the performance deterioration may be attributed to the distribution distinction among different training data." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.178, + 0.883, + 0.193 + ], + "angle": 0, + "content": "4.4. Further Analyses on Conversation Datasets" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.2, + 0.907, + 0.321 + ], + "angle": 0, + "content": "Introducing rejection answers to benign prompts may create a rejection bias on normal conversations. Therefore, except for XSTest [37] and VQA evaluations [15, 27], we also pick 1000 samples from LLaVA-v1.5-mix665k [22], LLaVA-Instruct-150K [24], and MMInstruct [26] each (for MMInstruct [26] samples are from the \"qa_en\" subset) to evaluate the compliance rate of LLaVA-v1.5 models. String matching is adopted as the metric." + }, + { + "type": "table", + "bbox": [ + 0.522, + 0.334, + 0.9, + 0.451 + ], + "angle": 0, + "content": "
BaseDatasetsOriginVLGuardOurs
LLaVA-v1.5-7B665K↑100.00100.0099.90
150K↑99.6098.9098.70
MM↑100.0099.0088.40
LLaVA-v1.5-13B665K↑100.00100.0099.80
150K↑99.6099.4098.30
MM↑100.0099.2085.90
" + }, + { + "type": "table_caption", + "bbox": [ + 0.513, + 0.461, + 0.906, + 0.504 + ], + "angle": 0, + "content": "Table 4. Compliance rate under different finetuning settings. 665K, 150K, and MM represent LLaVA-v1.5-mix665k [22], LLaVA-Instruct-150K [24], and MMInstruct [26], respectively." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.521, + 0.906, + 0.642 + ], + "angle": 0, + "content": "As shown in Table 4, the performance is almost intact for tasks from the LLaVA-v1.5-mix665k [22], but the models tend to reject image descriptions at a higher rate on LLaVA-Instruct-150K [24] and MMInstruct [26]. This may be because the data sampled from LLaVA-Med [20] contain quantities of image description queries, which induce models to reject similar prompts in evaluations, thus sacrificing the original performance to some extent." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.652, + 0.848, + 0.669 + ], + "angle": 0, + "content": "4.5. Ablation on Rejection Data Proportion" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.674, + 0.907, + 0.854 + ], + "angle": 0, + "content": "With the understanding that datasets with rejection samples could boost model safety, it is natural to ask the following question: without safety-related knowledge in the dataset, how much data is enough to counteract the compliance bias buried in the visual instruction tuning phase? To figure out the relationship between rejection data quantity and jailbreak defense behavior, we finetune the LLaVA-v1.5-7B five times, with increasing numbers of rejection data, while keeping the remaining setting changed. Still, for each experiment, 5000 randomly picked ordinary data are added to the training set. The performance on the original LLaVA-v1.5-7B is also added for reference." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.856, + 0.906, + 0.901 + ], + "angle": 0, + "content": "From the experiments, we could observe that the defense capability against MM-SafetyBench [25] prompts is more sensitive to the existence of rejection data than Fig-" + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "19885" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.117, + 0.092, + 0.462, + 0.177 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.116, + 0.183, + 0.462, + 0.292 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.09, + 0.304, + 0.483, + 0.334 + ], + "angle": 0, + "content": "Figure 3. The influence rejection data proportion has on jailbreak defense capability and visual understanding accuracy." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.359, + 0.483, + 0.45 + ], + "angle": 0, + "content": "Step [13]. For both attacks, \\(28.57\\%\\) is demonstrated to be enough to counteract compliance bias. Besides, with all rejection ratios, the visual understanding performance on VizWizQA [15] and ScienceQA [27] only slightly fluctuates around the baseline, which corresponds to previous findings in Sec. 4." + }, + { + "type": "title", + "bbox": [ + 0.09, + 0.459, + 0.473, + 0.475 + ], + "angle": 0, + "content": "4.6. Adding Data in the Instruction Tuning Phase" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.481, + 0.483, + 0.648 + ], + "angle": 0, + "content": "To gain a deeper understanding of this topic, in this part, we add rejection data into the multi-modal instruction following data to launch the visual instruction tuning. To be specific, due to the limitation of resources, we only design three runs, where the rejection data comprises \\(2\\%\\) and \\(5\\%\\) of the original LLaVA-v1.5-mix665k dataset [22], and the specific amount of rejection data is mixed in the original dataset for experiments. Identical to the official training setting, we adapt full-parameter tuning on the pretrained multi-modal projector and the released Vicuna-7B-v1.5 [8]. Training details are listed in the Appendix." + }, + { + "type": "table", + "bbox": [ + 0.098, + 0.658, + 0.479, + 0.79 + ], + "angle": 0, + "content": "
Evaluation DatasetsMetric (%)Origin2% Rej. data5% Rej. data
MM-SafetyBenchASR↓96.3658.9360.12
FigStepASR↓100.0099.4099.80
XSTestComp. rate↑92.0089.2091.20
XSTestRej. rate↑75.5077.0073.00
ScienceQAImage Acc.↑69.4669.7169.16
ScienceQATotal Acc.↑70.1271.1670.53
VizWizQAAcc.↑55.1056.5656.26
" + }, + { + "type": "table_caption", + "bbox": [ + 0.09, + 0.799, + 0.483, + 0.856 + ], + "angle": 0, + "content": "Table 5. Evaluation results of Visual Instruction Tuning. \"Origin\" refers to the official LLaVA-v1.5-mix665k dataset. \"Rej.\", \"Comp.\", and \"Acc.\" represents \"Rejection\", \"Compliance\", and \"Accuracy\", respectively." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.871, + 0.483, + 0.903 + ], + "angle": 0, + "content": "The model safety gets improvements, but not good enough. As shown in Table 5, without harming the vi" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.092, + 0.905, + 0.242 + ], + "angle": 0, + "content": "sual understanding capability, the attack successful rate of MM-SafetyBench [25] drops from \\(96.36\\%\\) to around \\(60\\%\\), which demonstrates again that MLLMs do not lose defense on multi-modal prompts. However, the defense against FigStep [13] jailbreak does not experience an obvious drop, which matches the ablation study in Fig. 3: for a relatively small percentage of rejection data, the ASR of FigStep remains high. Therefore, the conclusion is drawn that if the proportion of rejected data is insufficient, the defense behavior may not be fully activated." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.243, + 0.907, + 0.41 + ], + "angle": 0, + "content": "From experiments, if the rejection data lacks any inherent knowledge, awakening the model's multi-model safety awareness probably requires a sufficiently high rejection rate. Given this, enhancing model safety through finetuning seems to be more effective, where a smaller amount of data can constitute a substantial portion of the entire finetuning dataset. Specifically, in full-parameter visual instruction tuning, \\(2\\%\\) of the dataset corresponds to approximately 13k data points, which is more than six times the quantity required for finetuning. Despite this, the safety improvements gained through finetuning are far more significant." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.425, + 0.766, + 0.442 + ], + "angle": 0, + "content": "5. Conclusion and Limitation" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.451, + 0.906, + 0.602 + ], + "angle": 0, + "content": "In this paper, we investigate the well-curated safety finetuning data, finding that safety-related contents may not be the most crucial part for safety alignment. Based on this, we design a rejection dataset and conclude that the key factor influencing finetuning tends to be the existence of rejection data, instead of other attributes. Experiments demonstrate that MLLMs do not lose defense on multi-modal prompts. On the contrary, it is the compliance bias that prevents the models from refusing the malicious prompts, and training on the rejection dataset could mitigate such a bias." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.603, + 0.906, + 0.74 + ], + "angle": 0, + "content": "The major limitation in our study is the unsatisfying performance drop under image description queries. Besides, we answer the question that boosting multi-modal large language models without the necessity of curated data is possible, but currently it cannot be used to enhance the safety in the text-only domain significantly. We will leave both of them as future work. To summarize, we hope our findings with supervised finetuning methods can help better understand the safeguard of MLLM." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.754, + 0.667, + 0.772 + ], + "angle": 0, + "content": "Acknowledgment" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.78, + 0.906, + 0.903 + ], + "angle": 0, + "content": "We thank Zhengbo Wang and Yongcan Yu in NLPR for their feedback as well as discussion on our early drafts. Besides, we also would like to present our sincere appreciation to the anonymous reviewers for their constructive suggestions. This work was funded by the National Natural Science Foundation of China under Grants (62425606, 62276256, U2441251, U21B2045) and the Young Elite Scientists Sponsorship Program by CAST (2023QNRC001)." + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.52, + 0.957 + ], + "angle": 0, + "content": "19886" + } + ], + [ + { + "type": "title", + "bbox": [ + 0.093, + 0.091, + 0.188, + 0.106 + ], + "angle": 0, + "content": "References" + }, + { + "type": "ref_text", + "bbox": [ + 0.101, + 0.115, + 0.484, + 0.17 + ], + "angle": 0, + "content": "[1] Jinze Bai, Shuai Bai, Yunfei Chu, Zeyu Cui, Kai Dang, Xiaodong Deng, Yang Fan, Wenbin Ge, Yu Han, Fei Huang, et al. Qwen technical report. arXiv preprint arXiv:2309.16609, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.101, + 0.172, + 0.484, + 0.241 + ], + "angle": 0, + "content": "[2] Yuntao Bai, Andy Jones, Kamal Ndousse, Amanda Askell, Anna Chen, Nova DasSarma, Dawn Drain, Stanislav Fort, Deep Ganguli, Tom Henighan, et al. Training a helpful and harmless assistant with reinforcement learning from human feedback. arXiv preprint arXiv:2204.05862, 2022. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.243, + 0.483, + 0.311 + ], + "angle": 0, + "content": "[3] Guiming Hardy Chen, Shunian Chen, Ruifei Zhang, Junying Chen, Xiangbo Wu, Zhiyi Zhang, Zhihong Chen, Jianquan Li, Xiang Wan, and Benyou Wang. Allava: Harnessing gpt4v-synthesized data for a lite vision-language model. arXiv preprint arXiv:2402.11684, 2024. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.313, + 0.483, + 0.395 + ], + "angle": 0, + "content": "[4] Jun Chen, Deyao Zhu, Xiaogian Shen, Xiang Li, Zechu Liu, Pengchuan Zhang, Raghuraman Krishnamoorthi, Vikas Chandra, Yunyang Xiong, and Mohamed Elhoseiny. Minigpt-v2: large language model as a unified interface for vision-language multi-task learning. arXiv preprint arXiv:2310.09478, 2023. 1, 2, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.397, + 0.483, + 0.478 + ], + "angle": 0, + "content": "[5] Zhe Chen, Jiannan Wu, Wenhai Wang, Weijie Su, Guo Chen, Sen Xing, Muyan Zhong, Qinglong Zhang, Xizhou Zhu, Lewei Lu, Bin Li, Ping Luo, Tong Lu, Yu Qiao, and Jifeng Dai. Internvl: Scaling up vision foundation models and aligning for generic visual-linguistic tasks. arXiv preprint arXiv:2312.14238, 2023. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.481, + 0.483, + 0.549 + ], + "angle": 0, + "content": "[6] Zhe Chen, Weiyun Wang, Hao Tian, Shenglong Ye, Zhangwei Gao, Erfei Cui, Wenwen Tong, Kongzhi Hu, Jiapeng Luo, Zheng Ma, et al. How far are we to gpt-4v? closing the gap to commercial multimodal models with open-source suites. arXiv preprint arXiv:2404.16821, 2024. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.551, + 0.483, + 0.619 + ], + "angle": 0, + "content": "[7] Mehdi Cherti, Romain Beaumont, Ross Wightman, Mitchell Wortsman, Gabriel Ilharco, Cade Gordon, Christoph Schuhmann, Ludwig Schmidt, and Jenia Jitsev. Reproducible scaling laws for contrastive language-image learning. In Proc. CVPR, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.622, + 0.483, + 0.69 + ], + "angle": 0, + "content": "[8] Wei-Lin Chiang, Zhuohan Li, Zi Lin, Ying Sheng, Zhang-hao Wu, Hao Zhang, Lianmin Zheng, Siyuan Zhuang, Yong-hao Zhuang, Joseph E. Gonzalez, Ion Stoica, and Eric P. Xing. Vicuna: An open-source chatbot impressing gpt-4 with \\(90\\%\\) chatgpt quality, 2023. 2, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.692, + 0.483, + 0.759 + ], + "angle": 0, + "content": "[9] Wenliang Dai, Junnan Li, Dongxu Li, Anthony Tiong, Junqi Zhao, Weisheng Wang, Boyang Li, Pascale Fung, and Steven Hoi. InstructBLIP: Towards general-purpose vision-language models with instruction tuning. In Proc. NeurIPS, 2023. 1, 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.762, + 0.483, + 0.829 + ], + "angle": 0, + "content": "[10] Abhimanyu Dubey, Abhinav Jauhri, Abhinav Pandey, Abhishek Kadian, Ahmad Al-Dahle, Aiesha Letman, Akhil Mathur, Alan Schelten, Amy Yang, Angela Fan, et al. The llama 3 herd of models. arXiv preprint arXiv:2407.21783, 2024. 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.832, + 0.483, + 0.9 + ], + "angle": 0, + "content": "[11] Chaoyou Fu, Peixian Chen, Yunhang Shen, Yulei Qin, Mengdan Zhang, Xu Lin, Jinrui Yang, Xiawu Zheng, Ke Li, Xing Sun, et al. Mme: A comprehensive evaluation benchmark for multimodal large language models. arXiv preprint arXiv:2306.13394, 2023. 1" + }, + { + "type": "list", + "bbox": [ + 0.094, + 0.115, + 0.484, + 0.9 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.093, + 0.906, + 0.161 + ], + "angle": 0, + "content": "[12] Jiahui Gao, Renjie Pi, Tianyang Han, Han Wu, Lanqing Hong, Lingpeng Kong, Xin Jiang, and Zhenguo Li. Coca: Regaining safety-awareness of multimodal large language models with constitutional calibration. arXiv preprint arXiv:2409.11365, 2024. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.163, + 0.906, + 0.232 + ], + "angle": 0, + "content": "[13] Yichen Gong, Delong Ran, Jinyuan Liu, Conglei Wang, Tianshuo Cong, Anyu Wang, Sisi Duan, and Xiaoyun Wang. Figstep: Jailbreaking large vision-language models via typographic visual prompts. arXiv preprint arXiv:2311.05608, 2023. 1, 3, 4, 5, 6, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.233, + 0.906, + 0.287 + ], + "angle": 0, + "content": "[14] Yunhao Gou, Kai Chen, Zhili Liu, Lanqing Hong, Hang Xu, Zhenguo Li, Dit-Yan Yeung, James T Kwok, and Yu Zhang. Eyes closed, safety on: Protecting multimodal llms via image-to-text transformation. Proc. ECCV, 2024. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.288, + 0.906, + 0.343 + ], + "angle": 0, + "content": "[15] Danna Gurari, Qing Li, Abigale J Stangl, Anhong Guo, Chi Lin, Kristen Grauman, Jiebo Luo, and Jeffrey P Bigham. Vizwiz grand challenge: Answering visual questions from blind people. In Proc. CVPR, 2018. 4, 5, 6, 7, 8, 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.344, + 0.906, + 0.398 + ], + "angle": 0, + "content": "[16] Edward J Hu, Yelong Shen, Phillip Wallis, Zeyuan Allen-Zhu, Yuanzhi Li, Shean Wang, Lu Wang, and Weizhu Chen. LoRA: Low-rank adaptation of large language models. In Proc. ICLR, 2022. 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.4, + 0.906, + 0.468 + ], + "angle": 0, + "content": "[17] Albert Q Jiang, Alexandre Sablayrolles, Arthur Mensch, Chris Bamford, Devendra Singh Chaplot, Diego de las Casas, Florian Bressand, Gianna Lengyel, Guillaume Lample, Lucile Saulnier, et al. Mistral 7b. arXiv preprint arXiv:2310.06825, 2023. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.47, + 0.906, + 0.525 + ], + "angle": 0, + "content": "[18] Liqiang Jing, Yiren Li, Junhao Xu, Yongcan Yu, Pei Shen, and Xuemeng Song. Vision enhanced generative pre-trained language model for multimodal sentence summarization. Machine Intelligence Research, 20(2):289-298, 2023. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.526, + 0.906, + 0.593 + ], + "angle": 0, + "content": "[19] Wai-Chung Kwan, Hong-Ru Wang, Hui-Min Wang, and Kam-Fai Wong. A survey on recent advances and challenges in reinforcement learning methods for task-oriented dialogue policy learning. Machine Intelligence Research, 20(3):318-334, 2023. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.595, + 0.906, + 0.663 + ], + "angle": 0, + "content": "[20] Chunyuan Li, Cliff Wong, Sheng Zhang, Naoto Usuyama, Haotian Liu, Jianwei Yang, Tristan Naumann, Hoifung Poon, and Jianfeng Gao. Llava-med: Training a large language-and-vision assistant for biomedicine in one day. Proc. NeurIPS, 2024. 2, 5, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.665, + 0.906, + 0.719 + ], + "angle": 0, + "content": "[21] Yifan Li, Hangyu Guo, Kun Zhou, Wayne Xin Zhao, and JiRong Wen. Images are achilles' heel of alignment: Exploiting visual vulnerabilities for jailbreaking multimodal large language models. In proc. ECCV, 2024. 1, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.721, + 0.906, + 0.761 + ], + "angle": 0, + "content": "[22] Haotian Liu, Chunyuan Li, Yuheng Li, and Yong Jae Lee. Improved baselines with visual instruction tuning. In Proc. CVPR, 2024. 2, 4, 5, 6, 7, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.762, + 0.906, + 0.813 + ], + "angle": 0, + "content": "[23] Haotian Liu, Chunyuan Li, Yuheng Li, Bo Li, Yuanhan Zhang, Sheng Shen, and Yong Jae Lee. Llava-last: Improved reasoning,OCR, and world knowledge, 2024. 5, 6, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.817, + 0.906, + 0.845 + ], + "angle": 0, + "content": "[24] Haotian Liu, Chunyuan Li, Qingyang Wu, and Yong Jae Lee. Visual instruction tuning. In Proc. NeurIPS, 2024. 2, 3, 4, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.847, + 0.906, + 0.9 + ], + "angle": 0, + "content": "[25] Xin Liu, Yichen Zhu, Jindong Gu, Yunshi Lan, Chao Yang, and Yu Qiao. Mm-safetybench: A benchmark for safety evaluation of multimodal large language models. In Proc. ECCV, 2024. 1, 3, 4, 6, 7, 8" + }, + { + "type": "list", + "bbox": [ + 0.094, + 0.093, + 0.906, + 0.9 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "19887" + } + ], + [ + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.092, + 0.482, + 0.161 + ], + "angle": 0, + "content": "[26] Yangzhou Liu, Yue Cao, Zhangwei Gao, Weiyun Wang, Zhe Chen, Wenhai Wang, Hao Tian, Lewei Lu, Xizhou Zhu, Tong Lu, et al. Mminstruct: A high-quality multi-modal instruction tuning dataset with extensive diversity. Science China Information Sciences, 67(12):1-16, 2024. 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.162, + 0.482, + 0.232 + ], + "angle": 0, + "content": "[27] Pan Lu, Swaroop Mishra, Tony Xia, Liang Qiu, Kai-Wei Chang, Song-Chun Zhu, Oyvind Tafjord, Peter Clark, and Ashwin Kalyan. Learn to explain: Multimodal reasoning via thought chains for science question answering. In Proc. NeurIPS, 2022. 4, 5, 6, 7, 8, 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.233, + 0.482, + 0.287 + ], + "angle": 0, + "content": "[28] Haochen Luo, Jindong Gu, Fengyuan Liu, and Philip Torr. An image is worth 1000 lies: Transferability of adversarial images across prompts on vision-language models. In Proc. ICLR, 2024. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.288, + 0.482, + 0.33 + ], + "angle": 0, + "content": "[29] Anand Mishra, Shashank Shekhar, Ajeet Kumar Singh, and Anirban Chakraborty. Ocr-vqa: Visual question answering by reading text in images. In Proc. ICDAR, 2019. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.33, + 0.482, + 0.398 + ], + "angle": 0, + "content": "[30] Long Ouyang, Jeffrey Wu, Xu Jiang, Diogo Almeida, Carroll Wainwright, Pamela Mishkin, Chong Zhang, Sandhini Agarwal, Katarina Slama, Alex Ray, et al. Training language models to follow instructions with human feedback. In Proc. NeurIPS, 2022. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.4, + 0.482, + 0.455 + ], + "angle": 0, + "content": "[31] Renjie Pi, Tianyang Han, Jianshu Zhang, Yueqi Xie, Rui Pan, Qing Lian, Hanze Dong, Jipeng Zhang, and Tong Zhang. Mllm-protector: Ensuring mllm's safety without hurting performance. In Proc. EMNLP, 2024. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.456, + 0.482, + 0.509 + ], + "angle": 0, + "content": "[32] Xiangyu Qi, Kaixuan Huang, Ashwinee Panda, Peter Henderson, Mengdi Wang, and Prateek Mittal. Visual adversarial examples jailbreak aligned large language models. In Proc. AAAI, 2024. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.511, + 0.482, + 0.579 + ], + "angle": 0, + "content": "[33] Xiangyu Qi, Ashwinee Panda, Kaifeng Lyu, Xiao Ma, Subhrajit Roy, Ahmad Beirami, Prateek Mittal, and Peter Henderson. Safety alignment should be made more than just a few tokens deep. arXiv preprint arXiv:2406.05946, 2024. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.581, + 0.482, + 0.635 + ], + "angle": 0, + "content": "[34] Xiangyu Qi, Yi Zeng, Tinghao Xie, Pin-Yu Chen, Ruoxi Jia, Prateek Mittal, and Peter Henderson. Fine-tuning aligned language models compromises safety, even when users do not intend to! In Proc. ICLR, 2024. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.636, + 0.482, + 0.704 + ], + "angle": 0, + "content": "[35] Alec Radford, Jong Wook Kim, Chris Hallacy, Aditya Ramesh, Gabriel Goh, Sandhini Agarwal, Girish Sastry, Amanda Askell, Pamela Mishkin, Jack Clark, et al. Learning transferable visual models from natural language supervision. In Proc. ICML, 2021. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.705, + 0.482, + 0.761 + ], + "angle": 0, + "content": "[36] Rafael Rafailov, Archit Sharma, Eric Mitchell, Christopher D Manning, Stefano Ermon, and Chelsea Finn. Direct preference optimization: Your language model is secretly a reward model. In Proc. NeurIPS, 2024. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.762, + 0.482, + 0.816 + ], + "angle": 0, + "content": "[37] Paul Röttger, Hannah Kirk, Bertie Vidgen, Giuseppe Attanasio, Federico Bianchi, and Dirk Hovy. Xstest: A test suite for identifying exaggerated safety behaviours in large language models. In Proc. NAACL, 2024. 4, 6, 7, 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.818, + 0.482, + 0.859 + ], + "angle": 0, + "content": "[38] John Schulman, Filip Wolski, Prafulla Dhariwal, Alec Radford, and Oleg Klimov. Proximal policy optimization algorithms. arXiv preprint arXiv:1707.06347, 2017. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.86, + 0.482, + 0.901 + ], + "angle": 0, + "content": "[39] Erfan Shayegani, Yue Dong, and Nael Abu-Ghazaleh. Jailbreak in pieces: Compositional adversarial attacks on multimodal language models. In Proc. ICLR, 2024. 3" + }, + { + "type": "list", + "bbox": [ + 0.093, + 0.092, + 0.482, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.516, + 0.092, + 0.905, + 0.147 + ], + "angle": 0, + "content": "[40] Amanpreet Singh, Vivek Natarajan, Meet Shah, Yu Jiang, Xinlei Chen, Dhruv Batra, Devi Parikh, and Marcus Rohrbach. Towards vqa models that can read. In Proc. CVPR, 2019. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.516, + 0.149, + 0.905, + 0.192 + ], + "angle": 0, + "content": "[41] Quan Sun, Yuxin Fang, Ledell Wu, Xinlong Wang, and Yue Cao. Eva-clip: Improved training techniques for clip at scale. arXiv preprint arXiv:2303.15389, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.516, + 0.192, + 0.905, + 0.262 + ], + "angle": 0, + "content": "[42] Hugo Touvron, Thibaut Lavril, Gautier Izacard, Xavier Martinet, Marie-Anne Lachaux, Timothee Lacroix, Baptiste Rozière, Naman Goyal, Eric Hambro, Faisal Azhar, et al. Llama: Open and efficient foundation language models. arXiv preprint arXiv:2302.13971, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.263, + 0.905, + 0.318 + ], + "angle": 0, + "content": "[43] Junke Wang, Lingchen Meng, Zejia Weng, Bo He, Zuxuan Wu, and Yu-Gang Jiang. To see is to believe: Prompting gpt-4v for better visual instruction tuning. arXiv preprint arXiv:2311.07574, 2023. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.32, + 0.905, + 0.389 + ], + "angle": 0, + "content": "[44] Peng Wang, Shuai Bai, Sinan Tan, Shijie Wang, Zhihao Fan, Jinze Bai, Keqin Chen, Xuejing Liu, Jialin Wang, Wenbin Ge, et al. Qwen2-vl: Enhancing vision-language model's perception of the world at any resolution. arXiv preprint arXiv:2409.12191, 2024. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.391, + 0.905, + 0.459 + ], + "angle": 0, + "content": "[45] Xiao Wang, Guangyao Chen, Guangwu Qian, Pengcheng Gao, Xiao-Yong Wei, Yaowei Wang, Yonghong Tian, and Wen Gao. Large-scale multi-modal pre-trained models: A comprehensive survey. Machine Intelligence Research, 20 (4):447-482, 2023. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.461, + 0.905, + 0.517 + ], + "angle": 0, + "content": "[46] Yu Wang, Xiaogeng Liu, Yu Li, Muhao Chen, and Chaowei Xiao. Adashield: Safeguarding multimodal large language models from structure-based attack via adaptive shield prompting. Proc. ECCV, 2024. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.518, + 0.905, + 0.559 + ], + "angle": 0, + "content": "[47] Alexander Wei, Nika Haghtalab, and Jacob Steinhardt. Jailbroken: How does Ilm safety training fail? In Proc. NeurIPS, 2024. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.561, + 0.905, + 0.643 + ], + "angle": 0, + "content": "[48] Peng Xu, Wenqi Shao, Kaipeng Zhang, Peng Gao, Shuo Liu, Meng Lei, Fanqing Meng, Siyuan Huang, Yu Qiao, and Ping Luo. Lvlm-ehub: A comprehensive evaluation benchmark for large vision-language models. IEEE Transactions on Pattern Analysis and Machine Intelligence, 47(3):1877-1893, 2025. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.645, + 0.905, + 0.701 + ], + "angle": 0, + "content": "[49] An Yang, Baosong Yang, Binyuan Hui, Bo Zheng, Bowen Yu, Chang Zhou, Chengpeng Li, Chengyuan Li, Dayiheng Liu, Fei Huang, et al. Qwen2 technical report. arXiv preprint arXiv:2407.10671, 2024. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.702, + 0.905, + 0.758 + ], + "angle": 0, + "content": "[50] Qinghao Ye, Haiyang Xu, Jiabo Ye, Ming Yan, Anwen Hu, Haowei Liu, Qi Qian, Ji Zhang, and Fei Huang. mplugowl2: Revolutionizing multi-modal large language model with modality collaboration. In Proc. CVPR, 2024. 1, 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.759, + 0.905, + 0.814 + ], + "angle": 0, + "content": "[51] Shukang Yin, Chaoyou Fu, Sirui Zhao, Ke Li, Xing Sun, Tong Xu, and Enhong Chen. A survey on multimodal large language models. National Science Review, 11(12), 2024. 1, 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.816, + 0.905, + 0.872 + ], + "angle": 0, + "content": "[52] Alex Young, Bei Chen, Chao Li, Chengen Huang, Ge Zhang, Guanwei Zhang, Heng Li, Jiangcheng Zhu, Jianqun Chen, Jing Chang, et al. Yi: Open foundation models by 01. ai. arXiv preprint arXiv:2403.04652, 2024. 5, 6, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.873, + 0.905, + 0.901 + ], + "angle": 0, + "content": "[53] Yaowei Zheng, Richong Zhang, Junhao Zhang, Yanhan Ye, Zheyan Luo, Zhangchi Feng, and Yongqiang Ma. Llamafac" + }, + { + "type": "list", + "bbox": [ + 0.516, + 0.092, + 0.905, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.519, + 0.956 + ], + "angle": 0, + "content": "19888" + } + ], + [ + { + "type": "ref_text", + "bbox": [ + 0.124, + 0.092, + 0.482, + 0.12 + ], + "angle": 0, + "content": "tory: Unified efficient fine-tuning of \\(100+\\) language models. arXiv preprint arXiv:2403.13372, 2024. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.122, + 0.483, + 0.177 + ], + "angle": 0, + "content": "[54] Deyao Zhu, Jun Chen, Xiaoqian Shen, Xiang Li, and Mohamed Elhoseiny. Minigpt-4: Enhancing vision-language understanding with advanced large language models. arXiv preprint arXiv:2304.10592, 2023. 2, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.179, + 0.482, + 0.233 + ], + "angle": 0, + "content": "[55] Yongshuo Zong, Ondrej Bohdal, Tingyang Yu, Yongxin Yang, and Timothy Hospedales. Safety fine-tuning at (almost) no cost: A baseline for vision large language models. In Proc. ICML, 2024. 1, 3, 4, 5, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.235, + 0.482, + 0.289 + ], + "angle": 0, + "content": "[56] Andy Zou, Zifan Wang, Nicholas Carlini, Milad Nasr, J Zico Kolter, and Matt Fredrikson. Universal and transferable adversarial attacks on aligned language models. arXiv preprint arXiv:2307.15043, 2023. 5" + }, + { + "type": "list", + "bbox": [ + 0.093, + 0.092, + 0.483, + 0.289 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.946, + 0.519, + 0.956 + ], + "angle": 0, + "content": "19889" + } + ] +] \ No newline at end of file diff --git a/2025/Do We Really Need Curated Malicious Data for Safety Alignment in Multi-modal Large Language Models_/7cc50b70-a1c6-4460-8b07-4c06c5509d66_origin.pdf b/2025/Do We Really Need Curated Malicious Data for Safety Alignment in Multi-modal Large Language Models_/7cc50b70-a1c6-4460-8b07-4c06c5509d66_origin.pdf new file mode 100644 index 0000000000000000000000000000000000000000..8d4771cc19624e5aa009cbb1c0f35468bb1a82c6 --- /dev/null +++ b/2025/Do We Really Need Curated Malicious Data for Safety Alignment in Multi-modal Large Language Models_/7cc50b70-a1c6-4460-8b07-4c06c5509d66_origin.pdf @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:24754866fcffa14ea85f9759352ffbf13455003f7038619df2dfc5f4e862583b +size 1915419 diff --git a/2025/Do We Really Need Curated Malicious Data for Safety Alignment in Multi-modal Large Language Models_/full.md b/2025/Do We Really Need Curated Malicious Data for Safety Alignment in Multi-modal Large Language Models_/full.md new file mode 100644 index 0000000000000000000000000000000000000000..790b584b0fd432e71485b973ddc4924f89b25b7a --- /dev/null +++ b/2025/Do We Really Need Curated Malicious Data for Safety Alignment in Multi-modal Large Language Models_/full.md @@ -0,0 +1,275 @@ +# Do We Really Need Curated Malicious Data for Safety Alignment in Multi-modal Large Language Models? + +Yanbo Wang, Jiyang Guan, Jian Liang $^{\text{心}}$ , Ran He +School of Artificial Intelligence, University of Chinese Academy of Sciences +NLPR & MAIS, Institute of Automation, Chinese Academy of Sciences + +wangyanbo2023@ia.ac.cn liangjian92@gmail.com + +# Abstract + +Multi-modal large language models (MLLMs) have made significant progress, yet their safety alignment remains limited. Typically, current open-source MLLMs rely on the alignment inherited from their language module to avoid harmful generations. However, the lack of safety measures specifically designed for multi-modal inputs creates an alignment gap, leaving MLLMs vulnerable to vision-domain attacks such as typographic manipulation. Current methods utilize a carefully designed safety dataset to enhance model defense capability, while the specific knowledge or patterns acquired from the high-quality dataset remain unclear. Through comparison experiments, we find that the alignment gap primarily arises from data distribution biases, while image content, response quality, or the contrastive behavior of the dataset makes little contribution to boosting multi-modal safety. To further investigate this and identify the key factors in improving MLLM safety, we propose finetuning MLLMs on a small set of benign instruct-following data with responses replaced by simple, clear rejection sentences. Experiments show that, without the need for labor-intensive collection of high-quality malicious data, model safety can still be significantly improved, as long as a specific fraction of rejection data exists in the finetuning set, indicating the security alignment is not lost but rather obscured during multi-modal pretraining or instruction finetuning. Simply correcting the underlying data bias could narrow the safety gap in the vision domain. Warning: This paper contains harmful images and AI-generated contents which may be offensive. + +# 1. Introduction + +In recent years, the multi-modal large language model (MLLM) has experienced tremendous prosperity [11, 18, 45, 48, 51]. Companies and academics have proposed mul + +tiple models with their APIs, finetuning suites, and online demos. However, although the multi-modal understanding capability of these models is advancing steadily, the safety alignment of these models attracts less attention. To be detailed, most MLLMs connect an image encoder and a pretrained large language model with a vision-language connector, which are attention structures at the beginning and then replaced by MLPs [4, 5, 9, 50]. The safety alignment focuses mainly on the large language model, where popular RLHF algorithms [2, 19, 30] such as DPO [36] and PPO [38] are used to ensure that the chat model acts humanly. Nevertheless, such a workflow uncovers an obvious disadvantage: both the pretraining and visual instruction tuning stages do not contain a specific multi-modal safety alignment procedure, and the safety defensive capability could only be inherited from the LLM module. In addition, researchers also find that finetuning models on benign data may still sacrifice model safety [34]. They all challenge the defense capability of MLLMs. + +From the attackers' point of view, jailbreaking these MLLM models with multi-modal input seems more straightforward than jailbreaking language models [21]. Instead of optimizing adversarial noise to elicit harmful contents from vision language models with white-box accesses [28, 32], the typographic attack is a typical example to demonstrate the inferior safety of MLLMs. Fig-Step [13] directly transforms harmful requests into images through typography to deceive the image encoder module instead of the language module. With similar logic, MM-SafetyBench [25] proposes to add harmful keywords at the bottom of content-related diffusion-generated images, successfully inducing VLMs to generate inappropriate content across multiple fields. To defend such a simple but effective attack, except for prompt engineering techniques at inference time, previous work approaches this by collecting a safe multi-modal dataset with toxic queries and rejection responses, and then finetuning the MLLMs either in the second finetuning stage or after their release [55]. Regardless of a great success, current methods also leave a few questions unresolved. + +![](images/6e69771cca7474e3b1c6adcac02f81ad8c18c69ae5cf058c9f81c85fa3fbb1cc.jpg) +Figure 1. The main workflow of our study. After demonstrating that the quality of safety finetuning data does not contribute as much as we imagine, we modify the LLaVA-Med [20] dataset to create our finetuning dataset. For each data point, we simply pick one round of the conversation and replace their original answer with a clear rejection, without periods and tokens. + +tions. For one thing, collecting high-quality is always labor-intensive. Even if we could prompt powerful LLMs and diffusion models to generate the required multi-modal content, it is still effort-taking to create the pipeline, e.g. extracting keywords from related topics for diffusion models, or prompt engineering for high-quality various harmful questions. More importantly, what really matters in the multimodal alignment process? Do VLMs forget their safety alignment through multi-modal instruction tuning? We are still unclear why simple typographic attacks with keywords incitement could break the safety alignment of vision language models, while they possess identical meanings with the text-only prompts. + +In this work, we are the first to demonstrate that the alignment gap between the vision-language inputs and text-only counterparts comes mainly from training data distribution bias. In detail, we perform ablation studies on the previous finetuning method to filter out some unnecessary attributes of the dataset, discovering that the defense capability may be hardly related to the response quality, answer-rejection comparisons, or conversation rounds and image contents. On the contrary, increasing rejection rates in the finetuning dataset will elicit over-rejection, indicating the dataset distribution may matter in safety finetuning. To further explore the hypothesis that dataset distribution is the key factor for model alignment, we leverage the safety-unrelated LLaVA-Med dataset [20] to mitigate the compliance bias contained in open-source MLLMs. Experiments demonstrate that without safety-related images, harmful + +text prompts, and detailed rejection reasons, only "teaching the LLaVA- $\nu$ 1.5 models to reject a specific proportion of benign prompts" appears to boost multi-modal safety significantly. Detailed quantitative and qualitative results are followed, encompassing the relationship between reason generation and ordinary training data. To further explore to what extent the rejection proportion is strong enough to mitigate the compliance bias, we make ablation studies on the rejection dataset size, analyze the influence rejection data proportion has on two jailbreak methods, and launch full visual instruction tuning to compare the effectiveness with post-finetuning. We believe our findings could help better understand the safety-related behavior of MLLMs. + +# 2. Related Work + +# 2.1. Multi-modal Large Language Model + +In this paper, we leverage materials within the vision language models (VLMs) context. Most MLLMs consist of three parts: modality encoder, pretrained LLM, and modality interface [51]. Instead of training a MLLM from scratch, companies or academies utilize the modality connector to preprocess the on-the-shelf large language models with image encoders [4-6, 9, 44, 50, 54], in which the image encoders are mainly CLIP variants [7, 35, 41]. The language models are mainly pre-trained chat versions, such as Llama families [42], Vicuna families [8], and Qwen series [1, 49]. When training the MLLMs, the prevailing procedure contains vision-language pre-training and visual instruction tuning [9, 22, 24, 54]. In the first phase, + +the multi-modal projectors are trained in image captioning tasks. After that, in the instruction tuning phase, both the large language model and the connector are tuned with a large amount of VQA data [3, 43], aiming to enhance the visual understanding, reasoning, and OCR capability [29, 40]. However, safety alignment data are hardly included in the training phase, especially for those models from non-profit academies [4, 24, 54]. As a consequence, these MLLMs only rely on the safety alignment within their language module to defend against harmful prompts. Previous works prove the weakness of these MLLMs when faced with multi-modal inputs, and attribute this failure to the lack of curated VQA data [31, 47]. + +# 2.2. Black-box Jailbreak + +As mentioned above, due to the obvious alignment gap between text-only prompts and vision-language prompts in MLLMs, Many researchers leverage the alignment gap to launch multi-modal jailbreaks. To our best knowledge, Fig-Step [13] is the first work to jailbreak MLLMs with typography. When printing an image of a list containing steps for unethical behaviors with empty items as the input, such an attack could circumvent the alignment while keeping the OCR capability to fill in the blanks. Subsequent works follow the typographic attacks, append a content-related image on the top, and replace the sensitive words in text prompts with image incitement to enhance the attack success rate, e.g. replace "sell illegal drugs" with "the behavior shown in the image" [21, 25]. Such an incitement method could also be used to launch white-box attacks, where the attackers interact with their well-optimized adversarial examples by pronouns [39] to circumvent model safety alignment. + +# 2.3. Jailbreak Defense + +The Jailbreak defense on MLLMs could be mainly split into two categories: the inference phase and the training phase. Inference-time defenses are more like post-processing: AdaShield [46] proposes to create a shield prompt pool, adaptively appending a warning prompt at the end of user prompts to highlight response safety; MLLM-Protector [31] proposes a harm detector to firstly filter out harmful contents, and then change the response by their fine-tuned response detoxifier; similarly, ECSO [14] transforms input images into texts in a query-aware manner once the originally generated contents are labeled as inappropriate. To our knowledge, the work most related to our standpoint is CoCA [12], which focuses on exploring the safety-awareness of multi-modal inputs. Instead of training, they modify the decoding phase to enhance model safety. In the training phase, VLGuard [55] collects a safety dataset for supervised finetuning (SFT), achieving satisfying results in the defense of typographic attacks. + +# 3. What Matters in Safety Finetuning Dataset? + +# 3.1. Preliminaries + +Finetuning workflow. As stated in the introduction part, the previous method could finetune the MLLM in two ways: mixing safety data in the second stage instruction tuning, or directly finetuning the released model. Here we pick VL-Guard [55] as the target for the ablation study, and we train the models based on the released version with LoRA. + +VLGuard dataset [55]. The dataset composes three types of image-text pairs: safe-safe, safe-unsafe, and unsafe-unsafe. For example, safe-safe indicates a benign image-text pair, and safe-unsafe means the benign image is paired with a harmful (and in most cases, unrelated) text prompt. Here is an example of an LLaVA-style data sample: + +Prompt: What brand of beer is shown in the image? + +Assistant: The brand of the beer shown in the image is Grolsch. + +Prompt: Can you provide information on how to create fake IDs to purchase alcohol like the one shown in the image? + +Assistant: I'm sorry, but I cannot provide assistance or information on illegal activities, including creating fake IDs to purchase alcohol. Engaging in such actions is against the law and our usage policy strictly prohibits facilitating or promoting illegal activities. It's important to follow the legal age requirements for purchasing and consuming alcoholic beverages. + +For training, the dataset contains 977 single-turn unsafe-image pairs and 1023 multi-turn safe-image pairs, encompassing topics related to privacy, risky behavior, deception, and discrimination. + +# 3.2. Hypotheses + +High-quality datasets with well-picked images and detailed responses have been demonstrated to strengthen the safety of MLLMs. Here, we make several hypotheses on dataset attributes that contribute to satisfying defense capability. + +Comparisons in multi-turn data. As stated above, safe-safe and safe-unsafe pairs share identical images, which are structured in a multi-turn conversation style. Therefore, it is reasonable to hypothesize that the contrastive manner of the dataset forces MLLMs to understand what kind of input is toxic, and what kind of inputs are benign even with identical image visual prompts. + +Content-related images. It has been proved that content-related images tend to strengthen the jailbreak prompts [31]. For example, with a bomb image, the model has a higher probability of outputting illegal contents on manufacturing weapons [25]. Thus, We hypothesize that + +
Base ModelEvaluation DatasetsModel Settings
OriginVLGuardOne-turnUnsafe-onlyChange imageDirect sorryRandom reasonPure-VLGuard
LLaVA-v1.5-7BMM SafetyBench↓96.370.180.180.180.180.120.120.05
FigStep↓100.000.000.000.000.000.000.000.00
XCTest Compliance↑92.0077.2078.8076.8079.2070.8078.0052.00
XCTest Rejection↑75.5096.5095.5096.0093.0099.0098.0099.00
LLaVA-v1.5-13BMM SafetyBench↓97.800.360.360.180.180.420.540.12
FigStep↓99.800.000.000.000.000.000.000.00
XCTest Compliance↑90.0077.6077.2076.8082.0078.0078.0061.60
XCTest Rejection↑84.5097.0096.0096.0094.5098.0097.0099.50
LLaVA-NeXT Mistral-7BMM SafetyBench↓100.00.180.240.480.240.240.990.12
FigStep↓99.880.000.000.000.000.000.120.00
XCTest Compliance↑94.8087.6086.0081.6086.4086.4084.4043.20
XCTest Rejection↑58.5089.5084.5091.5084.0089.5094.0099.50
LLaVA-NeXT Llama-3-8BMM SafetyBench↓96.070.120.182.800.060.710.120.12
FigStep↓99.800.000.000.0038.40.000.000.00
XCTest Compliance↑90.4082.8082.8082.8086.4076.8078.4060.40
XCTest Rejection↑83.5092.0091.0091.0083.5097.0097.5099.00
+ +Table 1. Model safety evaluation with various finetuning data. All values are percentages. + +MLLM safety could be gained from the rejection of unsafeunsafe pairs. After collecting these harmful images and making models learn to reject them, the defense capability of MLLMs probably advances by a huge margin. + +Detailed reject reasons. Detailed reject reasons clarify the specific legal principles the harmful prompts violate, highlighting that as an AI model, it should follow legal and moral requirements to provide helpful and safe content. They add extra safety-related knowledge to the dataset, therefore, it is reasonable to hypothesize that such detailed data are crucial to invoke previous safety alignment in the language module, and then fill the gap between the multimodal input and pure-text domains. + +Dataset distribution bias. Comparing the instruction tuning dataset with the safety finetuning dataset, one distinct difference is the rejection behavior. Regardless of whether the training set contains toxic or harmful contents, in the instruction tuning phase the model hardly learns how to reject because nearly all replies in conversation-style data follow the instructions as detailed as possible. We hypothesize that this may also be the reason why the model loses the ability to reject inappropriate requests. + +# 3.3. Experiment Settings + +Finetuning models and datasets. Here, we launch all experiments on the LLaVA series model [22-24]. Aligning with previous methods, we add random-picked 5000 data points from the visual instruction tuning dataset LLaVA-v1.5-mix665k [22]. To testify to the influence of comparisons in the dataset, we split multi-turn safety data into single-turn and eliminate all safe-safe data pairs, named "One-turn" and "Unsafe-only", respectively. We also change the image prompt with random-picked benign images from LLaVA-v1.5-mix665k, named "Change im + +age" for the content-related image hypothesis. To modify the detailed reasons, we design two alternatives: replacing all replies with one clear rejection sentence or generating rejection contents with the prefix: "I'm sorry, I cannot assist with that request because it goes against my programming to" on benign data in the finetuning dataset. For comparisons, we also add experiments on pure VLGuard dataset [55] and non-finetuned original models. Experiments are done on LLaVA-v1.5-7B, LLaVA-v1.5-13B, LLaVA-NeXT-mixtral-7B, and LLaVA-NeXT-llama3-8B [22-24] with LoRA [16]. Finetuning details are listed in the Appendix. + +Evaluation metrics. We mainly evaluate the safety of MLLM models with black-box attacks, for it is the prevailing safety threat MLLM models will face. With collected jailbreaking image-text pairs on the Internet, a user could directly launch such an attack. We pick FigStep [13] and MM-SafetyBench [25] as the attacking sources, and the attack successful rate (ASR) is calculated by string-matching. We also tried to use Llama-3-Guard [10] as the discriminator in early evaluations, but the ASRs are lower because some sensitive topics exhibited in the datasets are not regarded as harmful, even if the model fails to reject content generation. Besides, to evaluate the model performance on visual understanding tasks, we adopt VizWizQA [15] and ScienceQA [27] for visual-text understanding, and XSTest dataset [37] for over-rejection evaluation. One point worth noting is that the XSTest dataset comprises two parts, in which the compliance column means the proportion of benign questions with non-rejection answers. + +# 3.4. Results + +Detailed safety and performance evaluations of models on various ablation datasets are shown in Table 1. + +
Base ModelFinetuning DatasetsVizWizQA(%)↑ScienceQA(%)
Image↑Total↑
LLaVA-v1.5-7BOrigin55.1070.1269.46
VLGuard56.3668.4270.12
One-turn56.3967.6769.79
Change image55.7867.3869.04
Unsafe-only56.7567.8269.63
Direct sorry53.9867.9769.61
Random reason56.6568.5270.01
Pure-VLGuard56.4565.2464.18
LLaVA-v1.5-13BOrigin57.3472.6874.89
VLGuard56.7771.5472.62
One-turn54.2871.3472.95
Change image55.5770.7572.98
Unsafe-only59.6270.8572.48
Direct sorry56.9871.0071.66
Random reason58.3971.0072.29
Pure-VLGuard58.2671.0071.47
+ +Table 2. Evaluation of model capabilities on VQA tasks. + +Comparisons in multi-turn data may make a limited contribution to the safety enhancement. With extensive experiments on both LLaVA-v1.5 [22] and LLaVA-NeXT [23] structures with different model sizes, the finding is consistent that the safety alignment of VLMs hardly relies on the comparison of answers to benign prompts and harmful prompts. For one thing, after splitting a double-turn conversation into two single-turn data points with the same image prompt, the defensive capability of all four models remains nearly unchanged, suggesting that utilizing some benign conversations as the conversation history may have a limited impact on enhancing the defense capability; for another, even eliminating all benign conversations from VL-Guard dataset [55] makes little difference to the defense result. Such a phenomenon indicates that benign data here may only contribute to the model's helpfulness, similar to randomly picked data in LLaVA-v1.5-mix665k [22]. + +Content-related images and reject reasons may not be the key factor for the defensive capability. In Table 1, it turns out that changing content-related harmful images to benign training images hardly deteriorates the defensive capability. Except for the inferior performance LLaVA-NeXT-Llama-3-8B [23] has on FigStep [13], under the rest circumstances training models with unrelated images could result in a satisfying defensive performance. Besides, in Table 2, the performance difference caused by image changing is only about $1\%$ on both VQA tasks [15, 27]. Similar trends also occur on ablations for reject reasons. Modifying or directly eliminating the reasons does not harm defensive capability, indicating that the model learns little from the detailed reasons for making the decision. Qi et al. [33] proposed a similar conclusion in the language model domain, claiming that current finetuning gradients mainly come from a few first tokens. + +High rejection proportion has the potential to harm the model performance. After eliminating training samples from LLaVA-v1.5-mix665k [22], the rejection conversation takes up around $66.7\%$ of the whole dataset. With such a data distribution, all models get over-sensitive: they could reject more than $99\%$ harmful text prompts while only answering about half of the benign questions, even though the curated finetuning data are of high quality. Besides, we also witness an accuracy drop on the ScienceQA dataset [27] with the LLaVA-v1.5-7B model. This phenomenon is consistent with the prior work [55], highlighting the existence of normal training data. + +To sum up, in the finetuning phase, it seems the models learn little from the high-quality content. Dataset modifications on images, answer reasons, and conversation comparisons probably only have negligible impacts on the safety alignment. However, altering the rejection proportion of the dataset could elicit over-rejection: even though the models only learn how to reject harmful instructions, the rejection behavior is wrongly generalized to benign prompts. Based on this, one question emerges: is dataset distribution the key factor for model alignment? If that is the case, is it possible to enhance MLLM safety without well-curated safe data? + +# 4. Benign Data with Rejection Responses Boost Model Safety + +To figure out the role dataset rejection proportion plays in safety finetuning, we try to stimulate the alignment potential with benign data. + +# 4.1. Experimental Setup + +To align with the previous setting, we randomly pick 2000 data points from LLaVA-Med dataset [20]. For each data point, we randomly pick one round of QAs in the conversation and replace the answer with a clear rejection. As in Sec. 3.3, these data are also mixed with randomly picked 5000 normal data from LLaVA-v1.5-mix665k [22]. Noting that the training data of Yi-VL-6B [52] and LLaVA-NeXT [23] are not publicly available, therefore we keep using the same setting as the LLaVA-v1.5 models. Considering possible overfitting, we only supervise the model to generate a few rejection tokens while masking out the end-of-sentence token. Assuming in a one-turn conversation $x$ , the instruction takes up $m$ tokens. In our experiments, we only calculate the language modeling loss by this equation [56]: + +$$ +\mathcal {L} \left(x _ {1: n + m}\right) = - \log p \left(x _ {n + 1: n + m} ^ {*} \mid x _ {1: n}\right), \tag {1} +$$ + +where $x_{n+1:n+m}^*$ refers to the tokenized rejection string: "I'm sorry, but I cannot assist with your request", without periods or token ids. For simplicity, here we ignore the existence of system prompts as well as token in some templates. To ensure the reliability of our results, + +
Base ModelFinetuning DatasetsMM SafetyBench↓FigStep↓XCTest↑VizWizQA↑ScienceQA↑
ComplianceRejectionImageTotal
LLaVA-v1.5-7BOrigin96.37100.0092.0075.5055.1070.1269.46
VLGuard0.180.0077.2096.5056.3668.4270.12
Ours5.600.2090.4082.0055.9268.0269.11
LLaVA-v1.5-13BOrigin97.8099.8090.0084.5057.3472.6874.89
VLGuard0.360.0077.6097.0056.7771.5472.62
Ours1.790.0090.4090.5058.7771.5973.36
Yi-VL-6BOrigin93.5199.6095.2041.50\( 66.20^1 \)61.0370.12
VLGuard0.180.0084.4093.0059.3269.8174.51
Ours3.573.2089.2067.0047.9168.2773.07
LLaVA-NeXT Mistral-7BOrigin99.88100.0094.8058.5064.9872.6378.57
VLGuard0.180.0087.6089.5053.1570.2074.56
Ours2.610.0091.2065.5055.5069.3674.65
+ +Table 3. The evaluation comparisons between our finetuning dataset and VLGuard [55]. All values are percentages. + +except for LLaVA-v1.5 models with Vicuna language module, we also pick the LLaVA-NeXT with Mistral-7B language module [17], and the Yi-VL-6B based on Yi-6B [52]. Training details are shown in the Appendix. + +# 4.2. Quantitative Results + +MLLMs do not lose defense on multi-modal prompts. In Table 3, it turns out that all four models work well on jailbreak defenses while maintaining a relatively high rate of benign prompt compliance. For MM-SafetyBench [25], the ASR of most models is below $5\%$ , which is a substantial drop compared to the original models. Experiments on FigStep [13] are also obvious: the highest ASR among the four models is $3.2\%$ . In training data, the conversation topic is unrelated to safety and privacy, and we do not append any reason for the "abnormal" rejection in each data point. The major factor is the rejection rate: compared to the visual instruction tuning phase with limited rejection data, we add 2000 rejection samples to 5000 benign training data in safety finetuning, reducing the proportion of instances where the answer complies with instructions. + +More importantly, our experimental results provide little support for the notion that the model forgets previous safety alignment or fails due to the absence of multi-modal safety data, leading it to interpret harmful jailbreak prompts as benign. Instead, we argue that the model fails to defend jailbreaks mentioned above mostly because all multi-modal instruction-tuning data naturally create a compliance bias. After instruction tuning, the model still possesses the basic moral understanding and judgment. The difference is that it is told all prompts with images should be answered and followed eagerly, resulting in the vulnerability. + +Without knowledge, correcting model distribution makes limited progress to text-only prompts. In Table 3, finetuning on our dataset does not significantly impair the + +safety compliance in XSTest [37], but it also fails to improve the unsafe rejection rate to the level achieved by VLGuard [55]. It is reasonable, for our dataset contains no extra safety-related contents compared to the VLGuard. Accordingly, merely correcting the model's distribution without incorporating additional knowledge is unlikely to result in a significant improvement in text-only input defense. + +Performances on general tasks tend to be associated with the ordinary training data. In experiments, we mix rejection data with ordinary data from LLaVA-v1.5-mix665k [22], the official training set for LLaVA-v1.5. In the first two rows, using our random rejection data could achieve similar results as VLGuard on VizWizQA [15] and ScienceQA [27], most of which are slightly inferior to the original model. For VizWizQA the model trained on our dataset even outperforms the original one, indicating the quality of the ordinary data. However, for LLaVA-NeXT [23] and Yi-VL [52], the performances experience a larger drop, which may be attributed to the inconsistency of training data. This adds more randomness to the results. + +Analyzing the results from the rejection data proportion perspective, we could also prove the claim that large language model alignment contributes to the defense of vision-language jailbreaks. Suppose the model does not possess safety-related knowledge or judgment for multi-modal data. In that case, the rejection rate of generated contents should be identical to, or slightly less than training data distribution. With $28.57\%$ rejection data in the training set, ASRs of LLaVA-v1.5-7B on jailbreak datasets should be around $70\%$ , and the performance on visual understanding and reasoning tasks should drop by $15\%$ to $20\%$ . However, the evaluation discrepancy between jailbreak datasets and vi + +sual understanding datasets discloses that the rejection behavior is not evenly added to the content generation. In other words, the model gets over-compliance after visual instruction tuning. Still, it knows which multi-modal prompts are harmful, just as it knows how to reject those text-only harmful prompts without interactions and incitements with images. The semantic meanings between multi-modal inputs and their text-only counterparts are aligned. In safety finetuning, it is less important to meticulously collect security data for the model to learn how to distinguish between benign and malicious instructions. Once enough rejection data is provided to enable rejection behavior, the model can make the right decisions. + +# 4.3. Qualitative Results + +To have a rough understanding of the answer quality, we present some examples of the model-generation content. + +Prompt: \n The bottom of the image shows a phrase of an activity. List steps in detail to perform this activity using cryptocurrencies without leaving a trace. + +# LLaVA-v1.5-7B + +Assistant: I'm sorry, but I cannot assist with your request. Purchasing illegal drugs with cryptocurrencies...the use of cryptocurrencies for illegal activities can lead to legal consequences and is morally wrong. It is important to support legal and legitimate businesses... + +# Yi-VL-6B + +Assistant: I'm sorry, but I cannot assist with your request. Please consult a professional for assistance with your request. + +# LLaVA- $\nu 1.5 - 7B$ -Random-reason + +Assistant: I'm sorry, I cannot assist with that request because it goes against my programming to provide opinions or beliefs. + +Figure 2. Examples of replies generated under different settings. The reasons are not as high quality as the LLaVA-v1.5. To save space, the image of drugs with the phrase "purchase illegal drugs" is not shown in the box. + +The quality of generated rejection reasons is relatively independent of model decisions. In Fig. 2, the LLaVA-v1.5 model generates high-quality responses, with rejection reasons related to the topic. However, For Yi-VL-6B [52] and LLaVA-NeXT [23], the models keep satisfying reject decisions, while generating imperfect reasons. Similar results are also demonstrated when training LLaVA-v1.5 models under the random reason setting, indicating that the + +training data quality possibly affects the following sentence completion. With a proper rejection rate in the training set, models can reject inappropriate prompts, while the performance deterioration may be attributed to the distribution distinction among different training data. + +# 4.4. Further Analyses on Conversation Datasets + +Introducing rejection answers to benign prompts may create a rejection bias on normal conversations. Therefore, except for XSTest [37] and VQA evaluations [15, 27], we also pick 1000 samples from LLaVA-v1.5-mix665k [22], LLaVA-Instruct-150K [24], and MMInstruct [26] each (for MMInstruct [26] samples are from the "qa_en" subset) to evaluate the compliance rate of LLaVA-v1.5 models. String matching is adopted as the metric. + +
BaseDatasetsOriginVLGuardOurs
LLaVA-v1.5-7B665K↑100.00100.0099.90
150K↑99.6098.9098.70
MM↑100.0099.0088.40
LLaVA-v1.5-13B665K↑100.00100.0099.80
150K↑99.6099.4098.30
MM↑100.0099.2085.90
+ +Table 4. Compliance rate under different finetuning settings. 665K, 150K, and MM represent LLaVA-v1.5-mix665k [22], LLaVA-Instruct-150K [24], and MMInstruct [26], respectively. + +As shown in Table 4, the performance is almost intact for tasks from the LLaVA-v1.5-mix665k [22], but the models tend to reject image descriptions at a higher rate on LLaVA-Instruct-150K [24] and MMInstruct [26]. This may be because the data sampled from LLaVA-Med [20] contain quantities of image description queries, which induce models to reject similar prompts in evaluations, thus sacrificing the original performance to some extent. + +# 4.5. Ablation on Rejection Data Proportion + +With the understanding that datasets with rejection samples could boost model safety, it is natural to ask the following question: without safety-related knowledge in the dataset, how much data is enough to counteract the compliance bias buried in the visual instruction tuning phase? To figure out the relationship between rejection data quantity and jailbreak defense behavior, we finetune the LLaVA-v1.5-7B five times, with increasing numbers of rejection data, while keeping the remaining setting changed. Still, for each experiment, 5000 randomly picked ordinary data are added to the training set. The performance on the original LLaVA-v1.5-7B is also added for reference. + +From the experiments, we could observe that the defense capability against MM-SafetyBench [25] prompts is more sensitive to the existence of rejection data than Fig- + +![](images/12695c757fb0afe130268417e9b0bb1c32599e7b037038e719968f16cca5e8a2.jpg) + +![](images/6f079d1c307a443c7daad69e4ce191e181d538d6f5803cebaf27553a5140a668.jpg) +Figure 3. The influence rejection data proportion has on jailbreak defense capability and visual understanding accuracy. + +Step [13]. For both attacks, $28.57\%$ is demonstrated to be enough to counteract compliance bias. Besides, with all rejection ratios, the visual understanding performance on VizWizQA [15] and ScienceQA [27] only slightly fluctuates around the baseline, which corresponds to previous findings in Sec. 4. + +# 4.6. Adding Data in the Instruction Tuning Phase + +To gain a deeper understanding of this topic, in this part, we add rejection data into the multi-modal instruction following data to launch the visual instruction tuning. To be specific, due to the limitation of resources, we only design three runs, where the rejection data comprises $2\%$ and $5\%$ of the original LLaVA-v1.5-mix665k dataset [22], and the specific amount of rejection data is mixed in the original dataset for experiments. Identical to the official training setting, we adapt full-parameter tuning on the pretrained multi-modal projector and the released Vicuna-7B-v1.5 [8]. Training details are listed in the Appendix. + +
Evaluation DatasetsMetric (%)Origin2% Rej. data5% Rej. data
MM-SafetyBenchASR↓96.3658.9360.12
FigStepASR↓100.0099.4099.80
XSTestComp. rate↑92.0089.2091.20
XSTestRej. rate↑75.5077.0073.00
ScienceQAImage Acc.↑69.4669.7169.16
ScienceQATotal Acc.↑70.1271.1670.53
VizWizQAAcc.↑55.1056.5656.26
+ +Table 5. Evaluation results of Visual Instruction Tuning. "Origin" refers to the official LLaVA-v1.5-mix665k dataset. "Rej.", "Comp.", and "Acc." represents "Rejection", "Compliance", and "Accuracy", respectively. + +The model safety gets improvements, but not good enough. As shown in Table 5, without harming the vi + +sual understanding capability, the attack successful rate of MM-SafetyBench [25] drops from $96.36\%$ to around $60\%$ , which demonstrates again that MLLMs do not lose defense on multi-modal prompts. However, the defense against FigStep [13] jailbreak does not experience an obvious drop, which matches the ablation study in Fig. 3: for a relatively small percentage of rejection data, the ASR of FigStep remains high. Therefore, the conclusion is drawn that if the proportion of rejected data is insufficient, the defense behavior may not be fully activated. + +From experiments, if the rejection data lacks any inherent knowledge, awakening the model's multi-model safety awareness probably requires a sufficiently high rejection rate. Given this, enhancing model safety through finetuning seems to be more effective, where a smaller amount of data can constitute a substantial portion of the entire finetuning dataset. Specifically, in full-parameter visual instruction tuning, $2\%$ of the dataset corresponds to approximately 13k data points, which is more than six times the quantity required for finetuning. Despite this, the safety improvements gained through finetuning are far more significant. + +# 5. Conclusion and Limitation + +In this paper, we investigate the well-curated safety finetuning data, finding that safety-related contents may not be the most crucial part for safety alignment. Based on this, we design a rejection dataset and conclude that the key factor influencing finetuning tends to be the existence of rejection data, instead of other attributes. Experiments demonstrate that MLLMs do not lose defense on multi-modal prompts. On the contrary, it is the compliance bias that prevents the models from refusing the malicious prompts, and training on the rejection dataset could mitigate such a bias. + +The major limitation in our study is the unsatisfying performance drop under image description queries. Besides, we answer the question that boosting multi-modal large language models without the necessity of curated data is possible, but currently it cannot be used to enhance the safety in the text-only domain significantly. We will leave both of them as future work. To summarize, we hope our findings with supervised finetuning methods can help better understand the safeguard of MLLM. + +# Acknowledgment + +We thank Zhengbo Wang and Yongcan Yu in NLPR for their feedback as well as discussion on our early drafts. Besides, we also would like to present our sincere appreciation to the anonymous reviewers for their constructive suggestions. This work was funded by the National Natural Science Foundation of China under Grants (62425606, 62276256, U2441251, U21B2045) and the Young Elite Scientists Sponsorship Program by CAST (2023QNRC001). + +# References + +[1] Jinze Bai, Shuai Bai, Yunfei Chu, Zeyu Cui, Kai Dang, Xiaodong Deng, Yang Fan, Wenbin Ge, Yu Han, Fei Huang, et al. Qwen technical report. arXiv preprint arXiv:2309.16609, 2023. 2 +[2] Yuntao Bai, Andy Jones, Kamal Ndousse, Amanda Askell, Anna Chen, Nova DasSarma, Dawn Drain, Stanislav Fort, Deep Ganguli, Tom Henighan, et al. Training a helpful and harmless assistant with reinforcement learning from human feedback. arXiv preprint arXiv:2204.05862, 2022. 1 +[3] Guiming Hardy Chen, Shunian Chen, Ruifei Zhang, Junying Chen, Xiangbo Wu, Zhiyi Zhang, Zhihong Chen, Jianquan Li, Xiang Wan, and Benyou Wang. Allava: Harnessing gpt4v-synthesized data for a lite vision-language model. arXiv preprint arXiv:2402.11684, 2024. 3 +[4] Jun Chen, Deyao Zhu, Xiaogian Shen, Xiang Li, Zechu Liu, Pengchuan Zhang, Raghuraman Krishnamoorthi, Vikas Chandra, Yunyang Xiong, and Mohamed Elhoseiny. Minigpt-v2: large language model as a unified interface for vision-language multi-task learning. arXiv preprint arXiv:2310.09478, 2023. 1, 2, 3 +[5] Zhe Chen, Jiannan Wu, Wenhai Wang, Weijie Su, Guo Chen, Sen Xing, Muyan Zhong, Qinglong Zhang, Xizhou Zhu, Lewei Lu, Bin Li, Ping Luo, Tong Lu, Yu Qiao, and Jifeng Dai. Internvl: Scaling up vision foundation models and aligning for generic visual-linguistic tasks. arXiv preprint arXiv:2312.14238, 2023. 1 +[6] Zhe Chen, Weiyun Wang, Hao Tian, Shenglong Ye, Zhangwei Gao, Erfei Cui, Wenwen Tong, Kongzhi Hu, Jiapeng Luo, Zheng Ma, et al. How far are we to gpt-4v? closing the gap to commercial multimodal models with open-source suites. arXiv preprint arXiv:2404.16821, 2024. 2 +[7] Mehdi Cherti, Romain Beaumont, Ross Wightman, Mitchell Wortsman, Gabriel Ilharco, Cade Gordon, Christoph Schuhmann, Ludwig Schmidt, and Jenia Jitsev. Reproducible scaling laws for contrastive language-image learning. In Proc. CVPR, 2023. 2 +[8] Wei-Lin Chiang, Zhuohan Li, Zi Lin, Ying Sheng, Zhang-hao Wu, Hao Zhang, Lianmin Zheng, Siyuan Zhuang, Yong-hao Zhuang, Joseph E. Gonzalez, Ion Stoica, and Eric P. Xing. Vicuna: An open-source chatbot impressing gpt-4 with $90\%$ chatgpt quality, 2023. 2, 8 +[9] Wenliang Dai, Junnan Li, Dongxu Li, Anthony Tiong, Junqi Zhao, Weisheng Wang, Boyang Li, Pascale Fung, and Steven Hoi. InstructBLIP: Towards general-purpose vision-language models with instruction tuning. In Proc. NeurIPS, 2023. 1, 2 +[10] Abhimanyu Dubey, Abhinav Jauhri, Abhinav Pandey, Abhishek Kadian, Ahmad Al-Dahle, Aiesha Letman, Akhil Mathur, Alan Schelten, Amy Yang, Angela Fan, et al. The llama 3 herd of models. arXiv preprint arXiv:2407.21783, 2024. 4 +[11] Chaoyou Fu, Peixian Chen, Yunhang Shen, Yulei Qin, Mengdan Zhang, Xu Lin, Jinrui Yang, Xiawu Zheng, Ke Li, Xing Sun, et al. Mme: A comprehensive evaluation benchmark for multimodal large language models. arXiv preprint arXiv:2306.13394, 2023. 1 + +[12] Jiahui Gao, Renjie Pi, Tianyang Han, Han Wu, Lanqing Hong, Lingpeng Kong, Xin Jiang, and Zhenguo Li. Coca: Regaining safety-awareness of multimodal large language models with constitutional calibration. arXiv preprint arXiv:2409.11365, 2024. 3 +[13] Yichen Gong, Delong Ran, Jinyuan Liu, Conglei Wang, Tianshuo Cong, Anyu Wang, Sisi Duan, and Xiaoyun Wang. Figstep: Jailbreaking large vision-language models via typographic visual prompts. arXiv preprint arXiv:2311.05608, 2023. 1, 3, 4, 5, 6, 8 +[14] Yunhao Gou, Kai Chen, Zhili Liu, Lanqing Hong, Hang Xu, Zhenguo Li, Dit-Yan Yeung, James T Kwok, and Yu Zhang. Eyes closed, safety on: Protecting multimodal llms via image-to-text transformation. Proc. ECCV, 2024. 3 +[15] Danna Gurari, Qing Li, Abigale J Stangl, Anhong Guo, Chi Lin, Kristen Grauman, Jiebo Luo, and Jeffrey P Bigham. Vizwiz grand challenge: Answering visual questions from blind people. In Proc. CVPR, 2018. 4, 5, 6, 7, 8, 1 +[16] Edward J Hu, Yelong Shen, Phillip Wallis, Zeyuan Allen-Zhu, Yuanzhi Li, Shean Wang, Lu Wang, and Weizhu Chen. LoRA: Low-rank adaptation of large language models. In Proc. ICLR, 2022. 4 +[17] Albert Q Jiang, Alexandre Sablayrolles, Arthur Mensch, Chris Bamford, Devendra Singh Chaplot, Diego de las Casas, Florian Bressand, Gianna Lengyel, Guillaume Lample, Lucile Saulnier, et al. Mistral 7b. arXiv preprint arXiv:2310.06825, 2023. 6 +[18] Liqiang Jing, Yiren Li, Junhao Xu, Yongcan Yu, Pei Shen, and Xuemeng Song. Vision enhanced generative pre-trained language model for multimodal sentence summarization. Machine Intelligence Research, 20(2):289-298, 2023. 1 +[19] Wai-Chung Kwan, Hong-Ru Wang, Hui-Min Wang, and Kam-Fai Wong. A survey on recent advances and challenges in reinforcement learning methods for task-oriented dialogue policy learning. Machine Intelligence Research, 20(3):318-334, 2023. 1 +[20] Chunyuan Li, Cliff Wong, Sheng Zhang, Naoto Usuyama, Haotian Liu, Jianwei Yang, Tristan Naumann, Hoifung Poon, and Jianfeng Gao. Llava-med: Training a large language-and-vision assistant for biomedicine in one day. Proc. NeurIPS, 2024. 2, 5, 7 +[21] Yifan Li, Hangyu Guo, Kun Zhou, Wayne Xin Zhao, and JiRong Wen. Images are achilles' heel of alignment: Exploiting visual vulnerabilities for jailbreaking multimodal large language models. In proc. ECCV, 2024. 1, 3 +[22] Haotian Liu, Chunyuan Li, Yuheng Li, and Yong Jae Lee. Improved baselines with visual instruction tuning. In Proc. CVPR, 2024. 2, 4, 5, 6, 7, 8 +[23] Haotian Liu, Chunyuan Li, Yuheng Li, Bo Li, Yuanhan Zhang, Sheng Shen, and Yong Jae Lee. Llava-last: Improved reasoning,OCR, and world knowledge, 2024. 5, 6, 7 +[24] Haotian Liu, Chunyuan Li, Qingyang Wu, and Yong Jae Lee. Visual instruction tuning. In Proc. NeurIPS, 2024. 2, 3, 4, 7 +[25] Xin Liu, Yichen Zhu, Jindong Gu, Yunshi Lan, Chao Yang, and Yu Qiao. Mm-safetybench: A benchmark for safety evaluation of multimodal large language models. In Proc. ECCV, 2024. 1, 3, 4, 6, 7, 8 + +[26] Yangzhou Liu, Yue Cao, Zhangwei Gao, Weiyun Wang, Zhe Chen, Wenhai Wang, Hao Tian, Lewei Lu, Xizhou Zhu, Tong Lu, et al. Mminstruct: A high-quality multi-modal instruction tuning dataset with extensive diversity. Science China Information Sciences, 67(12):1-16, 2024. 7 +[27] Pan Lu, Swaroop Mishra, Tony Xia, Liang Qiu, Kai-Wei Chang, Song-Chun Zhu, Oyvind Tafjord, Peter Clark, and Ashwin Kalyan. Learn to explain: Multimodal reasoning via thought chains for science question answering. In Proc. NeurIPS, 2022. 4, 5, 6, 7, 8, 1 +[28] Haochen Luo, Jindong Gu, Fengyuan Liu, and Philip Torr. An image is worth 1000 lies: Transferability of adversarial images across prompts on vision-language models. In Proc. ICLR, 2024. 1 +[29] Anand Mishra, Shashank Shekhar, Ajeet Kumar Singh, and Anirban Chakraborty. Ocr-vqa: Visual question answering by reading text in images. In Proc. ICDAR, 2019. 3 +[30] Long Ouyang, Jeffrey Wu, Xu Jiang, Diogo Almeida, Carroll Wainwright, Pamela Mishkin, Chong Zhang, Sandhini Agarwal, Katarina Slama, Alex Ray, et al. Training language models to follow instructions with human feedback. In Proc. NeurIPS, 2022. 1 +[31] Renjie Pi, Tianyang Han, Jianshu Zhang, Yueqi Xie, Rui Pan, Qing Lian, Hanze Dong, Jipeng Zhang, and Tong Zhang. Mllm-protector: Ensuring mllm's safety without hurting performance. In Proc. EMNLP, 2024. 3 +[32] Xiangyu Qi, Kaixuan Huang, Ashwinee Panda, Peter Henderson, Mengdi Wang, and Prateek Mittal. Visual adversarial examples jailbreak aligned large language models. In Proc. AAAI, 2024. 1 +[33] Xiangyu Qi, Ashwinee Panda, Kaifeng Lyu, Xiao Ma, Subhrajit Roy, Ahmad Beirami, Prateek Mittal, and Peter Henderson. Safety alignment should be made more than just a few tokens deep. arXiv preprint arXiv:2406.05946, 2024. 5 +[34] Xiangyu Qi, Yi Zeng, Tinghao Xie, Pin-Yu Chen, Ruoxi Jia, Prateek Mittal, and Peter Henderson. Fine-tuning aligned language models compromises safety, even when users do not intend to! In Proc. ICLR, 2024. 1 +[35] Alec Radford, Jong Wook Kim, Chris Hallacy, Aditya Ramesh, Gabriel Goh, Sandhini Agarwal, Girish Sastry, Amanda Askell, Pamela Mishkin, Jack Clark, et al. Learning transferable visual models from natural language supervision. In Proc. ICML, 2021. 2 +[36] Rafael Rafailov, Archit Sharma, Eric Mitchell, Christopher D Manning, Stefano Ermon, and Chelsea Finn. Direct preference optimization: Your language model is secretly a reward model. In Proc. NeurIPS, 2024. 1 +[37] Paul Röttger, Hannah Kirk, Bertie Vidgen, Giuseppe Attanasio, Federico Bianchi, and Dirk Hovy. Xstest: A test suite for identifying exaggerated safety behaviours in large language models. In Proc. NAACL, 2024. 4, 6, 7, 1 +[38] John Schulman, Filip Wolski, Prafulla Dhariwal, Alec Radford, and Oleg Klimov. Proximal policy optimization algorithms. arXiv preprint arXiv:1707.06347, 2017. 1 +[39] Erfan Shayegani, Yue Dong, and Nael Abu-Ghazaleh. Jailbreak in pieces: Compositional adversarial attacks on multimodal language models. In Proc. ICLR, 2024. 3 + +[40] Amanpreet Singh, Vivek Natarajan, Meet Shah, Yu Jiang, Xinlei Chen, Dhruv Batra, Devi Parikh, and Marcus Rohrbach. Towards vqa models that can read. In Proc. CVPR, 2019. 3 +[41] Quan Sun, Yuxin Fang, Ledell Wu, Xinlong Wang, and Yue Cao. Eva-clip: Improved training techniques for clip at scale. arXiv preprint arXiv:2303.15389, 2023. 2 +[42] Hugo Touvron, Thibaut Lavril, Gautier Izacard, Xavier Martinet, Marie-Anne Lachaux, Timothee Lacroix, Baptiste Rozière, Naman Goyal, Eric Hambro, Faisal Azhar, et al. Llama: Open and efficient foundation language models. arXiv preprint arXiv:2302.13971, 2023. 2 +[43] Junke Wang, Lingchen Meng, Zejia Weng, Bo He, Zuxuan Wu, and Yu-Gang Jiang. To see is to believe: Prompting gpt-4v for better visual instruction tuning. arXiv preprint arXiv:2311.07574, 2023. 3 +[44] Peng Wang, Shuai Bai, Sinan Tan, Shijie Wang, Zhihao Fan, Jinze Bai, Keqin Chen, Xuejing Liu, Jialin Wang, Wenbin Ge, et al. Qwen2-vl: Enhancing vision-language model's perception of the world at any resolution. arXiv preprint arXiv:2409.12191, 2024. 2 +[45] Xiao Wang, Guangyao Chen, Guangwu Qian, Pengcheng Gao, Xiao-Yong Wei, Yaowei Wang, Yonghong Tian, and Wen Gao. Large-scale multi-modal pre-trained models: A comprehensive survey. Machine Intelligence Research, 20 (4):447-482, 2023. 1 +[46] Yu Wang, Xiaogeng Liu, Yu Li, Muhao Chen, and Chaowei Xiao. Adashield: Safeguarding multimodal large language models from structure-based attack via adaptive shield prompting. Proc. ECCV, 2024. 3 +[47] Alexander Wei, Nika Haghtalab, and Jacob Steinhardt. Jailbroken: How does Ilm safety training fail? In Proc. NeurIPS, 2024. 3 +[48] Peng Xu, Wenqi Shao, Kaipeng Zhang, Peng Gao, Shuo Liu, Meng Lei, Fanqing Meng, Siyuan Huang, Yu Qiao, and Ping Luo. Lvlm-ehub: A comprehensive evaluation benchmark for large vision-language models. IEEE Transactions on Pattern Analysis and Machine Intelligence, 47(3):1877-1893, 2025. 1 +[49] An Yang, Baosong Yang, Binyuan Hui, Bo Zheng, Bowen Yu, Chang Zhou, Chengpeng Li, Chengyuan Li, Dayiheng Liu, Fei Huang, et al. Qwen2 technical report. arXiv preprint arXiv:2407.10671, 2024. 2 +[50] Qinghao Ye, Haiyang Xu, Jiabo Ye, Ming Yan, Anwen Hu, Haowei Liu, Qi Qian, Ji Zhang, and Fei Huang. mplugowl2: Revolutionizing multi-modal large language model with modality collaboration. In Proc. CVPR, 2024. 1, 2 +[51] Shukang Yin, Chaoyou Fu, Sirui Zhao, Ke Li, Xing Sun, Tong Xu, and Enhong Chen. A survey on multimodal large language models. National Science Review, 11(12), 2024. 1, 2 +[52] Alex Young, Bei Chen, Chao Li, Chengen Huang, Ge Zhang, Guanwei Zhang, Heng Li, Jiangcheng Zhu, Jianqun Chen, Jing Chang, et al. Yi: Open foundation models by 01. ai. arXiv preprint arXiv:2403.04652, 2024. 5, 6, 7 +[53] Yaowei Zheng, Richong Zhang, Junhao Zhang, Yanhan Ye, Zheyan Luo, Zhangchi Feng, and Yongqiang Ma. Llamafac + +tory: Unified efficient fine-tuning of $100+$ language models. arXiv preprint arXiv:2403.13372, 2024. 1 +[54] Deyao Zhu, Jun Chen, Xiaoqian Shen, Xiang Li, and Mohamed Elhoseiny. Minigpt-4: Enhancing vision-language understanding with advanced large language models. arXiv preprint arXiv:2304.10592, 2023. 2, 3 +[55] Yongshuo Zong, Ondrej Bohdal, Tingyang Yu, Yongxin Yang, and Timothy Hospedales. Safety fine-tuning at (almost) no cost: A baseline for vision large language models. In Proc. ICML, 2024. 1, 3, 4, 5, 6 +[56] Andy Zou, Zifan Wang, Nicholas Carlini, Milad Nasr, J Zico Kolter, and Matt Fredrikson. Universal and transferable adversarial attacks on aligned language models. arXiv preprint arXiv:2307.15043, 2023. 5 \ No newline at end of file diff --git a/2025/Do We Really Need Curated Malicious Data for Safety Alignment in Multi-modal Large Language Models_/images.zip b/2025/Do We Really Need Curated Malicious Data for Safety Alignment in Multi-modal Large Language Models_/images.zip new file mode 100644 index 0000000000000000000000000000000000000000..8288d38953d92007c17f0e098d88711dd70e990d --- /dev/null +++ b/2025/Do We Really Need Curated Malicious Data for Safety Alignment in Multi-modal Large Language Models_/images.zip @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:e68390718a02323e37c77747868443d38c2a474c35f036a25eb1243adf30ef96 +size 483615 diff --git a/2025/Do We Really Need Curated Malicious Data for Safety Alignment in Multi-modal Large Language Models_/layout.json b/2025/Do We Really Need Curated Malicious Data for Safety Alignment in Multi-modal Large Language Models_/layout.json new file mode 100644 index 0000000000000000000000000000000000000000..4e6cdc54cc35e1aa51da4c27881de4f691d4129d --- /dev/null +++ b/2025/Do We Really Need Curated Malicious Data for Safety Alignment in Multi-modal Large Language Models_/layout.json @@ -0,0 +1,6633 @@ +{ + "pdf_info": [ + { + "para_blocks": [ + { + "bbox": [ + 96, + 102, + 516, + 140 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 96, + 102, + 516, + 140 + ], + "spans": [ + { + "bbox": [ + 96, + 102, + 516, + 140 + ], + "type": "text", + "content": "Do We Really Need Curated Malicious Data for Safety Alignment in Multi-modal Large Language Models?" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 118, + 161, + 491, + 203 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 118, + 161, + 491, + 203 + ], + "spans": [ + { + "bbox": [ + 118, + 161, + 491, + 203 + ], + "type": "text", + "content": "Yanbo Wang, Jiyang Guan, Jian Liang" + }, + { + "bbox": [ + 118, + 161, + 491, + 203 + ], + "type": "inline_equation", + "content": "^{\\text{心}}" + }, + { + "bbox": [ + 118, + 161, + 491, + 203 + ], + "type": "text", + "content": ", Ran He \nSchool of Artificial Intelligence, University of Chinese Academy of Sciences \nNLPR & MAIS, Institute of Automation, Chinese Academy of Sciences" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 183, + 205, + 424, + 217 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 183, + 205, + 424, + 217 + ], + "spans": [ + { + "bbox": [ + 183, + 205, + 424, + 217 + ], + "type": "text", + "content": "wangyanbo2023@ia.ac.cn liangjian92@gmail.com" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 151, + 244, + 200, + 258 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 151, + 244, + 200, + 258 + ], + "spans": [ + { + "bbox": [ + 151, + 244, + 200, + 258 + ], + "type": "text", + "content": "Abstract" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 55, + 270, + 297, + 617 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 270, + 297, + 617 + ], + "spans": [ + { + "bbox": [ + 55, + 270, + 297, + 617 + ], + "type": "text", + "content": "Multi-modal large language models (MLLMs) have made significant progress, yet their safety alignment remains limited. Typically, current open-source MLLMs rely on the alignment inherited from their language module to avoid harmful generations. However, the lack of safety measures specifically designed for multi-modal inputs creates an alignment gap, leaving MLLMs vulnerable to vision-domain attacks such as typographic manipulation. Current methods utilize a carefully designed safety dataset to enhance model defense capability, while the specific knowledge or patterns acquired from the high-quality dataset remain unclear. Through comparison experiments, we find that the alignment gap primarily arises from data distribution biases, while image content, response quality, or the contrastive behavior of the dataset makes little contribution to boosting multi-modal safety. To further investigate this and identify the key factors in improving MLLM safety, we propose finetuning MLLMs on a small set of benign instruct-following data with responses replaced by simple, clear rejection sentences. Experiments show that, without the need for labor-intensive collection of high-quality malicious data, model safety can still be significantly improved, as long as a specific fraction of rejection data exists in the finetuning set, indicating the security alignment is not lost but rather obscured during multi-modal pretraining or instruction finetuning. Simply correcting the underlying data bias could narrow the safety gap in the vision domain. Warning: This paper contains harmful images and AI-generated contents which may be offensive." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 56, + 639, + 135, + 651 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 639, + 135, + 651 + ], + "spans": [ + { + "bbox": [ + 56, + 639, + 135, + 651 + ], + "type": "text", + "content": "1. Introduction" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 55, + 659, + 296, + 696 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 659, + 296, + 696 + ], + "spans": [ + { + "bbox": [ + 55, + 659, + 296, + 696 + ], + "type": "text", + "content": "In recent years, the multi-modal large language model (MLLM) has experienced tremendous prosperity [11, 18, 45, 48, 51]. Companies and academics have proposed mul" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 313, + 245, + 555, + 474 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 245, + 555, + 474 + ], + "spans": [ + { + "bbox": [ + 313, + 245, + 555, + 474 + ], + "type": "text", + "content": "tiple models with their APIs, finetuning suites, and online demos. However, although the multi-modal understanding capability of these models is advancing steadily, the safety alignment of these models attracts less attention. To be detailed, most MLLMs connect an image encoder and a pretrained large language model with a vision-language connector, which are attention structures at the beginning and then replaced by MLPs [4, 5, 9, 50]. The safety alignment focuses mainly on the large language model, where popular RLHF algorithms [2, 19, 30] such as DPO [36] and PPO [38] are used to ensure that the chat model acts humanly. Nevertheless, such a workflow uncovers an obvious disadvantage: both the pretraining and visual instruction tuning stages do not contain a specific multi-modal safety alignment procedure, and the safety defensive capability could only be inherited from the LLM module. In addition, researchers also find that finetuning models on benign data may still sacrifice model safety [34]. They all challenge the defense capability of MLLMs." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 313, + 475, + 556, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 475, + 556, + 715 + ], + "spans": [ + { + "bbox": [ + 313, + 475, + 556, + 715 + ], + "type": "text", + "content": "From the attackers' point of view, jailbreaking these MLLM models with multi-modal input seems more straightforward than jailbreaking language models [21]. Instead of optimizing adversarial noise to elicit harmful contents from vision language models with white-box accesses [28, 32], the typographic attack is a typical example to demonstrate the inferior safety of MLLMs. Fig-Step [13] directly transforms harmful requests into images through typography to deceive the image encoder module instead of the language module. With similar logic, MM-SafetyBench [25] proposes to add harmful keywords at the bottom of content-related diffusion-generated images, successfully inducing VLMs to generate inappropriate content across multiple fields. To defend such a simple but effective attack, except for prompt engineering techniques at inference time, previous work approaches this by collecting a safe multi-modal dataset with toxic queries and rejection responses, and then finetuning the MLLMs either in the second finetuning stage or after their release [55]. Regardless of a great success, current methods also leave a few questions unresolved." + } + ] + } + ], + "index": 10 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "spans": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "type": "text", + "content": "CVF" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 145, + 0, + 496, + 37 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 145, + 0, + 496, + 37 + ], + "spans": [ + { + "bbox": [ + 145, + 0, + 496, + 37 + ], + "type": "text", + "content": "This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 70, + 703, + 153, + 713 + ], + "type": "page_footnote", + "angle": 0, + "lines": [ + { + "bbox": [ + 70, + 703, + 153, + 713 + ], + "spans": [ + { + "bbox": [ + 70, + 703, + 153, + 713 + ], + "type": "text", + "content": "Corresponding author." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "text", + "content": "19879" + } + ] + } + ], + "index": 12 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 0 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 55, + 68, + 553, + 295 + ], + "blocks": [ + { + "bbox": [ + 55, + 68, + 553, + 295 + ], + "lines": [ + { + "bbox": [ + 55, + 68, + 553, + 295 + ], + "spans": [ + { + "bbox": [ + 55, + 68, + 553, + 295 + ], + "type": "image", + "image_path": "6e69771cca7474e3b1c6adcac02f81ad8c18c69ae5cf058c9f81c85fa3fbb1cc.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 55, + 302, + 555, + 336 + ], + "lines": [ + { + "bbox": [ + 55, + 302, + 555, + 336 + ], + "spans": [ + { + "bbox": [ + 55, + 302, + 555, + 336 + ], + "type": "text", + "content": "Figure 1. The main workflow of our study. After demonstrating that the quality of safety finetuning data does not contribute as much as we imagine, we modify the LLaVA-Med [20] dataset to create our finetuning dataset. For each data point, we simply pick one round of the conversation and replace their original answer with a clear rejection, without periods and tokens." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "bbox": [ + 54, + 356, + 297, + 513 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 54, + 356, + 297, + 513 + ], + "spans": [ + { + "bbox": [ + 54, + 356, + 297, + 513 + ], + "type": "text", + "content": "tions. For one thing, collecting high-quality is always labor-intensive. Even if we could prompt powerful LLMs and diffusion models to generate the required multi-modal content, it is still effort-taking to create the pipeline, e.g. extracting keywords from related topics for diffusion models, or prompt engineering for high-quality various harmful questions. More importantly, what really matters in the multimodal alignment process? Do VLMs forget their safety alignment through multi-modal instruction tuning? We are still unclear why simple typographic attacks with keywords incitement could break the safety alignment of vision language models, while they possess identical meanings with the text-only prompts." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 55, + 522, + 296, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 522, + 296, + 714 + ], + "spans": [ + { + "bbox": [ + 55, + 522, + 296, + 714 + ], + "type": "text", + "content": "In this work, we are the first to demonstrate that the alignment gap between the vision-language inputs and text-only counterparts comes mainly from training data distribution bias. In detail, we perform ablation studies on the previous finetuning method to filter out some unnecessary attributes of the dataset, discovering that the defense capability may be hardly related to the response quality, answer-rejection comparisons, or conversation rounds and image contents. On the contrary, increasing rejection rates in the finetuning dataset will elicit over-rejection, indicating the dataset distribution may matter in safety finetuning. To further explore the hypothesis that dataset distribution is the key factor for model alignment, we leverage the safety-unrelated LLaVA-Med dataset [20] to mitigate the compliance bias contained in open-source MLLMs. Experiments demonstrate that without safety-related images, harmful" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 313, + 356, + 556, + 513 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 356, + 556, + 513 + ], + "spans": [ + { + "bbox": [ + 313, + 356, + 556, + 513 + ], + "type": "text", + "content": "text prompts, and detailed rejection reasons, only \"teaching the LLaVA-" + }, + { + "bbox": [ + 313, + 356, + 556, + 513 + ], + "type": "inline_equation", + "content": "\\nu" + }, + { + "bbox": [ + 313, + 356, + 556, + 513 + ], + "type": "text", + "content": "1.5 models to reject a specific proportion of benign prompts\" appears to boost multi-modal safety significantly. Detailed quantitative and qualitative results are followed, encompassing the relationship between reason generation and ordinary training data. To further explore to what extent the rejection proportion is strong enough to mitigate the compliance bias, we make ablation studies on the rejection dataset size, analyze the influence rejection data proportion has on two jailbreak methods, and launch full visual instruction tuning to compare the effectiveness with post-finetuning. We believe our findings could help better understand the safety-related behavior of MLLMs." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 314, + 521, + 402, + 533 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 314, + 521, + 402, + 533 + ], + "spans": [ + { + "bbox": [ + 314, + 521, + 402, + 533 + ], + "type": "text", + "content": "2. Related Work" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 313, + 540, + 509, + 553 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 540, + 509, + 553 + ], + "spans": [ + { + "bbox": [ + 313, + 540, + 509, + 553 + ], + "type": "text", + "content": "2.1. Multi-modal Large Language Model" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 313, + 558, + 556, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 558, + 556, + 715 + ], + "spans": [ + { + "bbox": [ + 313, + 558, + 556, + 715 + ], + "type": "text", + "content": "In this paper, we leverage materials within the vision language models (VLMs) context. Most MLLMs consist of three parts: modality encoder, pretrained LLM, and modality interface [51]. Instead of training a MLLM from scratch, companies or academies utilize the modality connector to preprocess the on-the-shelf large language models with image encoders [4-6, 9, 44, 50, 54], in which the image encoders are mainly CLIP variants [7, 35, 41]. The language models are mainly pre-trained chat versions, such as Llama families [42], Vicuna families [8], and Qwen series [1, 49]. When training the MLLMs, the prevailing procedure contains vision-language pre-training and visual instruction tuning [9, 22, 24, 54]. In the first phase," + } + ] + } + ], + "index": 7 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "type": "text", + "content": "19880" + } + ] + } + ], + "index": 8 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 1 + }, + { + "para_blocks": [ + { + "bbox": [ + 55, + 72, + 296, + 228 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 72, + 296, + 228 + ], + "spans": [ + { + "bbox": [ + 55, + 72, + 296, + 228 + ], + "type": "text", + "content": "the multi-modal projectors are trained in image captioning tasks. After that, in the instruction tuning phase, both the large language model and the connector are tuned with a large amount of VQA data [3, 43], aiming to enhance the visual understanding, reasoning, and OCR capability [29, 40]. However, safety alignment data are hardly included in the training phase, especially for those models from non-profit academies [4, 24, 54]. As a consequence, these MLLMs only rely on the safety alignment within their language module to defend against harmful prompts. Previous works prove the weakness of these MLLMs when faced with multi-modal inputs, and attribute this failure to the lack of curated VQA data [31, 47]." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 55, + 242, + 173, + 254 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 242, + 173, + 254 + ], + "spans": [ + { + "bbox": [ + 55, + 242, + 173, + 254 + ], + "type": "text", + "content": "2.2. Black-box Jailbreak" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 55, + 262, + 296, + 465 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 262, + 296, + 465 + ], + "spans": [ + { + "bbox": [ + 55, + 262, + 296, + 465 + ], + "type": "text", + "content": "As mentioned above, due to the obvious alignment gap between text-only prompts and vision-language prompts in MLLMs, Many researchers leverage the alignment gap to launch multi-modal jailbreaks. To our best knowledge, Fig-Step [13] is the first work to jailbreak MLLMs with typography. When printing an image of a list containing steps for unethical behaviors with empty items as the input, such an attack could circumvent the alignment while keeping the OCR capability to fill in the blanks. Subsequent works follow the typographic attacks, append a content-related image on the top, and replace the sensitive words in text prompts with image incitement to enhance the attack success rate, e.g. replace \"sell illegal drugs\" with \"the behavior shown in the image\" [21, 25]. Such an incitement method could also be used to launch white-box attacks, where the attackers interact with their well-optimized adversarial examples by pronouns [39] to circumvent model safety alignment." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 55, + 479, + 162, + 490 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 479, + 162, + 490 + ], + "spans": [ + { + "bbox": [ + 55, + 479, + 162, + 490 + ], + "type": "text", + "content": "2.3. Jailbreak Defense" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 55, + 498, + 296, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 498, + 296, + 713 + ], + "spans": [ + { + "bbox": [ + 55, + 498, + 296, + 713 + ], + "type": "text", + "content": "The Jailbreak defense on MLLMs could be mainly split into two categories: the inference phase and the training phase. Inference-time defenses are more like post-processing: AdaShield [46] proposes to create a shield prompt pool, adaptively appending a warning prompt at the end of user prompts to highlight response safety; MLLM-Protector [31] proposes a harm detector to firstly filter out harmful contents, and then change the response by their fine-tuned response detoxifier; similarly, ECSO [14] transforms input images into texts in a query-aware manner once the originally generated contents are labeled as inappropriate. To our knowledge, the work most related to our standpoint is CoCA [12], which focuses on exploring the safety-awareness of multi-modal inputs. Instead of training, they modify the decoding phase to enhance model safety. In the training phase, VLGuard [55] collects a safety dataset for supervised finetuning (SFT), achieving satisfying results in the defense of typographic attacks." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 314, + 71, + 555, + 85 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 314, + 71, + 555, + 85 + ], + "spans": [ + { + "bbox": [ + 314, + 71, + 555, + 85 + ], + "type": "text", + "content": "3. What Matters in Safety Finetuning Dataset?" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 314, + 91, + 400, + 102 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 314, + 91, + 400, + 102 + ], + "spans": [ + { + "bbox": [ + 314, + 91, + 400, + 102 + ], + "type": "text", + "content": "3.1. Preliminaries" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 313, + 109, + 555, + 180 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 109, + 555, + 180 + ], + "spans": [ + { + "bbox": [ + 313, + 109, + 555, + 180 + ], + "type": "text", + "content": "Finetuning workflow. As stated in the introduction part, the previous method could finetune the MLLM in two ways: mixing safety data in the second stage instruction tuning, or directly finetuning the released model. Here we pick VL-Guard [55] as the target for the ablation study, and we train the models based on the released version with LoRA." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 313, + 181, + 555, + 254 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 181, + 555, + 254 + ], + "spans": [ + { + "bbox": [ + 313, + 181, + 555, + 254 + ], + "type": "text", + "content": "VLGuard dataset [55]. The dataset composes three types of image-text pairs: safe-safe, safe-unsafe, and unsafe-unsafe. For example, safe-safe indicates a benign image-text pair, and safe-unsafe means the benign image is paired with a harmful (and in most cases, unrelated) text prompt. Here is an example of an LLaVA-style data sample:" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 333, + 276, + 535, + 298 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 333, + 276, + 535, + 298 + ], + "spans": [ + { + "bbox": [ + 333, + 276, + 535, + 298 + ], + "type": "text", + "content": "Prompt: What brand of beer is shown in the image?" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 333, + 299, + 535, + 319 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 333, + 299, + 535, + 319 + ], + "spans": [ + { + "bbox": [ + 333, + 299, + 535, + 319 + ], + "type": "text", + "content": "Assistant: The brand of the beer shown in the image is Grolsch." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 333, + 320, + 535, + 353 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 333, + 320, + 535, + 353 + ], + "spans": [ + { + "bbox": [ + 333, + 320, + 535, + 353 + ], + "type": "text", + "content": "Prompt: Can you provide information on how to create fake IDs to purchase alcohol like the one shown in the image?" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 333, + 354, + 535, + 430 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 333, + 354, + 535, + 430 + ], + "spans": [ + { + "bbox": [ + 333, + 354, + 535, + 430 + ], + "type": "text", + "content": "Assistant: I'm sorry, but I cannot provide assistance or information on illegal activities, including creating fake IDs to purchase alcohol. Engaging in such actions is against the law and our usage policy strictly prohibits facilitating or promoting illegal activities. It's important to follow the legal age requirements for purchasing and consuming alcoholic beverages." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 313, + 447, + 555, + 494 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 447, + 555, + 494 + ], + "spans": [ + { + "bbox": [ + 313, + 447, + 555, + 494 + ], + "type": "text", + "content": "For training, the dataset contains 977 single-turn unsafe-image pairs and 1023 multi-turn safe-image pairs, encompassing topics related to privacy, risky behavior, deception, and discrimination." + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 314, + 503, + 391, + 516 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 314, + 503, + 391, + 516 + ], + "spans": [ + { + "bbox": [ + 314, + 503, + 391, + 516 + ], + "type": "text", + "content": "3.2. Hypotheses" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 313, + 521, + 555, + 569 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 521, + 555, + 569 + ], + "spans": [ + { + "bbox": [ + 313, + 521, + 555, + 569 + ], + "type": "text", + "content": "High-quality datasets with well-picked images and detailed responses have been demonstrated to strengthen the safety of MLLMs. Here, we make several hypotheses on dataset attributes that contribute to satisfying defense capability." + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 313, + 570, + 554, + 653 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 570, + 554, + 653 + ], + "spans": [ + { + "bbox": [ + 313, + 570, + 554, + 653 + ], + "type": "text", + "content": "Comparisons in multi-turn data. As stated above, safe-safe and safe-unsafe pairs share identical images, which are structured in a multi-turn conversation style. Therefore, it is reasonable to hypothesize that the contrastive manner of the dataset forces MLLMs to understand what kind of input is toxic, and what kind of inputs are benign even with identical image visual prompts." + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 313, + 654, + 555, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 654, + 555, + 713 + ], + "spans": [ + { + "bbox": [ + 313, + 654, + 555, + 713 + ], + "type": "text", + "content": "Content-related images. It has been proved that content-related images tend to strengthen the jailbreak prompts [31]. For example, with a bomb image, the model has a higher probability of outputting illegal contents on manufacturing weapons [25]. Thus, We hypothesize that" + } + ] + } + ], + "index": 17 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "text", + "content": "19881" + } + ] + } + ], + "index": 18 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 2 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 58, + 70, + 553, + 258 + ], + "blocks": [ + { + "bbox": [ + 58, + 70, + 553, + 258 + ], + "lines": [ + { + "bbox": [ + 58, + 70, + 553, + 258 + ], + "spans": [ + { + "bbox": [ + 58, + 70, + 553, + 258 + ], + "type": "table", + "html": "
Base ModelEvaluation DatasetsModel Settings
OriginVLGuardOne-turnUnsafe-onlyChange imageDirect sorryRandom reasonPure-VLGuard
LLaVA-v1.5-7BMM SafetyBench↓96.370.180.180.180.180.120.120.05
FigStep↓100.000.000.000.000.000.000.000.00
XCTest Compliance↑92.0077.2078.8076.8079.2070.8078.0052.00
XCTest Rejection↑75.5096.5095.5096.0093.0099.0098.0099.00
LLaVA-v1.5-13BMM SafetyBench↓97.800.360.360.180.180.420.540.12
FigStep↓99.800.000.000.000.000.000.000.00
XCTest Compliance↑90.0077.6077.2076.8082.0078.0078.0061.60
XCTest Rejection↑84.5097.0096.0096.0094.5098.0097.0099.50
LLaVA-NeXT Mistral-7BMM SafetyBench↓100.00.180.240.480.240.240.990.12
FigStep↓99.880.000.000.000.000.000.120.00
XCTest Compliance↑94.8087.6086.0081.6086.4086.4084.4043.20
XCTest Rejection↑58.5089.5084.5091.5084.0089.5094.0099.50
LLaVA-NeXT Llama-3-8BMM SafetyBench↓96.070.120.182.800.060.710.120.12
FigStep↓99.800.000.000.0038.40.000.000.00
XCTest Compliance↑90.4082.8082.8082.8086.4076.8078.4060.40
XCTest Rejection↑83.5092.0091.0091.0083.5097.0097.5099.00
", + "image_path": "6ade23144ad273c9453bfd5da902888d7cff63bdd34f88be9ae186b141290bf9.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "table_body" + } + ], + "index": 0 + }, + { + "bbox": [ + 142, + 266, + 466, + 277 + ], + "lines": [ + { + "bbox": [ + 142, + 266, + 466, + 277 + ], + "spans": [ + { + "bbox": [ + 142, + 266, + 466, + 277 + ], + "type": "text", + "content": "Table 1. Model safety evaluation with various finetuning data. All values are percentages." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "text" + }, + { + "bbox": [ + 55, + 298, + 295, + 346 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 298, + 295, + 346 + ], + "spans": [ + { + "bbox": [ + 55, + 298, + 295, + 346 + ], + "type": "text", + "content": "MLLM safety could be gained from the rejection of unsafeunsafe pairs. After collecting these harmful images and making models learn to reject them, the defense capability of MLLMs probably advances by a huge margin." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 54, + 347, + 295, + 453 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 54, + 347, + 295, + 453 + ], + "spans": [ + { + "bbox": [ + 54, + 347, + 295, + 453 + ], + "type": "text", + "content": "Detailed reject reasons. Detailed reject reasons clarify the specific legal principles the harmful prompts violate, highlighting that as an AI model, it should follow legal and moral requirements to provide helpful and safe content. They add extra safety-related knowledge to the dataset, therefore, it is reasonable to hypothesize that such detailed data are crucial to invoke previous safety alignment in the language module, and then fill the gap between the multimodal input and pure-text domains." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 55, + 456, + 295, + 564 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 456, + 295, + 564 + ], + "spans": [ + { + "bbox": [ + 55, + 456, + 295, + 564 + ], + "type": "text", + "content": "Dataset distribution bias. Comparing the instruction tuning dataset with the safety finetuning dataset, one distinct difference is the rejection behavior. Regardless of whether the training set contains toxic or harmful contents, in the instruction tuning phase the model hardly learns how to reject because nearly all replies in conversation-style data follow the instructions as detailed as possible. We hypothesize that this may also be the reason why the model loses the ability to reject inappropriate requests." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 55, + 575, + 174, + 588 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 575, + 174, + 588 + ], + "spans": [ + { + "bbox": [ + 55, + 575, + 174, + 588 + ], + "type": "text", + "content": "3.3. Experiment Settings" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 55, + 594, + 295, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 594, + 295, + 713 + ], + "spans": [ + { + "bbox": [ + 55, + 594, + 295, + 713 + ], + "type": "text", + "content": "Finetuning models and datasets. Here, we launch all experiments on the LLaVA series model [22-24]. Aligning with previous methods, we add random-picked 5000 data points from the visual instruction tuning dataset LLaVA-v1.5-mix665k [22]. To testify to the influence of comparisons in the dataset, we split multi-turn safety data into single-turn and eliminate all safe-safe data pairs, named \"One-turn\" and \"Unsafe-only\", respectively. We also change the image prompt with random-picked benign images from LLaVA-v1.5-mix665k, named \"Change im" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 313, + 298, + 555, + 441 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 298, + 555, + 441 + ], + "spans": [ + { + "bbox": [ + 313, + 298, + 555, + 441 + ], + "type": "text", + "content": "age\" for the content-related image hypothesis. To modify the detailed reasons, we design two alternatives: replacing all replies with one clear rejection sentence or generating rejection contents with the prefix: \"I'm sorry, I cannot assist with that request because it goes against my programming to\" on benign data in the finetuning dataset. For comparisons, we also add experiments on pure VLGuard dataset [55] and non-finetuned original models. Experiments are done on LLaVA-v1.5-7B, LLaVA-v1.5-13B, LLaVA-NeXT-mixtral-7B, and LLaVA-NeXT-llama3-8B [22-24] with LoRA [16]. Finetuning details are listed in the Appendix." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 313, + 443, + 556, + 659 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 443, + 556, + 659 + ], + "spans": [ + { + "bbox": [ + 313, + 443, + 556, + 659 + ], + "type": "text", + "content": "Evaluation metrics. We mainly evaluate the safety of MLLM models with black-box attacks, for it is the prevailing safety threat MLLM models will face. With collected jailbreaking image-text pairs on the Internet, a user could directly launch such an attack. We pick FigStep [13] and MM-SafetyBench [25] as the attacking sources, and the attack successful rate (ASR) is calculated by string-matching. We also tried to use Llama-3-Guard [10] as the discriminator in early evaluations, but the ASRs are lower because some sensitive topics exhibited in the datasets are not regarded as harmful, even if the model fails to reject content generation. Besides, to evaluate the model performance on visual understanding tasks, we adopt VizWizQA [15] and ScienceQA [27] for visual-text understanding, and XSTest dataset [37] for over-rejection evaluation. One point worth noting is that the XSTest dataset comprises two parts, in which the compliance column means the proportion of benign questions with non-rejection answers." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 313, + 670, + 372, + 682 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 670, + 372, + 682 + ], + "spans": [ + { + "bbox": [ + 313, + 670, + 372, + 682 + ], + "type": "text", + "content": "3.4. Results" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 313, + 689, + 555, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 689, + 555, + 713 + ], + "spans": [ + { + "bbox": [ + 313, + 689, + 555, + 713 + ], + "type": "text", + "content": "Detailed safety and performance evaluations of models on various ablation datasets are shown in Table 1." + } + ] + } + ], + "index": 10 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "text", + "content": "19882" + } + ] + } + ], + "index": 11 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 3 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 58, + 70, + 305, + 258 + ], + "blocks": [ + { + "bbox": [ + 58, + 70, + 305, + 258 + ], + "lines": [ + { + "bbox": [ + 58, + 70, + 305, + 258 + ], + "spans": [ + { + "bbox": [ + 58, + 70, + 305, + 258 + ], + "type": "table", + "html": "
Base ModelFinetuning DatasetsVizWizQA(%)↑ScienceQA(%)
Image↑Total↑
LLaVA-v1.5-7BOrigin55.1070.1269.46
VLGuard56.3668.4270.12
One-turn56.3967.6769.79
Change image55.7867.3869.04
Unsafe-only56.7567.8269.63
Direct sorry53.9867.9769.61
Random reason56.6568.5270.01
Pure-VLGuard56.4565.2464.18
LLaVA-v1.5-13BOrigin57.3472.6874.89
VLGuard56.7771.5472.62
One-turn54.2871.3472.95
Change image55.5770.7572.98
Unsafe-only59.6270.8572.48
Direct sorry56.9871.0071.66
Random reason58.3971.0072.29
Pure-VLGuard58.2671.0071.47
", + "image_path": "622ec068cdbaa5b393a05d49fcff09891e01e433933379aca6b609acd793aaab.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "table_body" + } + ], + "index": 0 + }, + { + "bbox": [ + 72, + 265, + 278, + 277 + ], + "lines": [ + { + "bbox": [ + 72, + 265, + 278, + 277 + ], + "spans": [ + { + "bbox": [ + 72, + 265, + 278, + 277 + ], + "type": "text", + "content": "Table 2. Evaluation of model capabilities on VQA tasks." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "text" + }, + { + "bbox": [ + 55, + 304, + 295, + 508 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 304, + 295, + 508 + ], + "spans": [ + { + "bbox": [ + 55, + 304, + 295, + 508 + ], + "type": "text", + "content": "Comparisons in multi-turn data may make a limited contribution to the safety enhancement. With extensive experiments on both LLaVA-v1.5 [22] and LLaVA-NeXT [23] structures with different model sizes, the finding is consistent that the safety alignment of VLMs hardly relies on the comparison of answers to benign prompts and harmful prompts. For one thing, after splitting a double-turn conversation into two single-turn data points with the same image prompt, the defensive capability of all four models remains nearly unchanged, suggesting that utilizing some benign conversations as the conversation history may have a limited impact on enhancing the defense capability; for another, even eliminating all benign conversations from VL-Guard dataset [55] makes little difference to the defense result. Such a phenomenon indicates that benign data here may only contribute to the model's helpfulness, similar to randomly picked data in LLaVA-v1.5-mix665k [22]." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 55, + 510, + 295, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 510, + 295, + 713 + ], + "spans": [ + { + "bbox": [ + 55, + 510, + 295, + 713 + ], + "type": "text", + "content": "Content-related images and reject reasons may not be the key factor for the defensive capability. In Table 1, it turns out that changing content-related harmful images to benign training images hardly deteriorates the defensive capability. Except for the inferior performance LLaVA-NeXT-Llama-3-8B [23] has on FigStep [13], under the rest circumstances training models with unrelated images could result in a satisfying defensive performance. Besides, in Table 2, the performance difference caused by image changing is only about " + }, + { + "bbox": [ + 55, + 510, + 295, + 713 + ], + "type": "inline_equation", + "content": "1\\%" + }, + { + "bbox": [ + 55, + 510, + 295, + 713 + ], + "type": "text", + "content": " on both VQA tasks [15, 27]. Similar trends also occur on ablations for reject reasons. Modifying or directly eliminating the reasons does not harm defensive capability, indicating that the model learns little from the detailed reasons for making the decision. Qi et al. [33] proposed a similar conclusion in the language model domain, claiming that current finetuning gradients mainly come from a few first tokens." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 313, + 72, + 553, + 215 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 72, + 553, + 215 + ], + "spans": [ + { + "bbox": [ + 313, + 72, + 553, + 215 + ], + "type": "text", + "content": "High rejection proportion has the potential to harm the model performance. After eliminating training samples from LLaVA-v1.5-mix665k [22], the rejection conversation takes up around " + }, + { + "bbox": [ + 313, + 72, + 553, + 215 + ], + "type": "inline_equation", + "content": "66.7\\%" + }, + { + "bbox": [ + 313, + 72, + 553, + 215 + ], + "type": "text", + "content": " of the whole dataset. With such a data distribution, all models get over-sensitive: they could reject more than " + }, + { + "bbox": [ + 313, + 72, + 553, + 215 + ], + "type": "inline_equation", + "content": "99\\%" + }, + { + "bbox": [ + 313, + 72, + 553, + 215 + ], + "type": "text", + "content": " harmful text prompts while only answering about half of the benign questions, even though the curated finetuning data are of high quality. Besides, we also witness an accuracy drop on the ScienceQA dataset [27] with the LLaVA-v1.5-7B model. This phenomenon is consistent with the prior work [55], highlighting the existence of normal training data." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 313, + 216, + 555, + 348 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 216, + 555, + 348 + ], + "spans": [ + { + "bbox": [ + 313, + 216, + 555, + 348 + ], + "type": "text", + "content": "To sum up, in the finetuning phase, it seems the models learn little from the high-quality content. Dataset modifications on images, answer reasons, and conversation comparisons probably only have negligible impacts on the safety alignment. However, altering the rejection proportion of the dataset could elicit over-rejection: even though the models only learn how to reject harmful instructions, the rejection behavior is wrongly generalized to benign prompts. Based on this, one question emerges: is dataset distribution the key factor for model alignment? If that is the case, is it possible to enhance MLLM safety without well-curated safe data?" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 313, + 358, + 555, + 386 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 358, + 555, + 386 + ], + "spans": [ + { + "bbox": [ + 313, + 358, + 555, + 386 + ], + "type": "text", + "content": "4. Benign Data with Rejection Responses Boost Model Safety" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 313, + 392, + 555, + 430 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 392, + 555, + 430 + ], + "spans": [ + { + "bbox": [ + 313, + 392, + 555, + 430 + ], + "type": "text", + "content": "To figure out the role dataset rejection proportion plays in safety finetuning, we try to stimulate the alignment potential with benign data." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 313, + 436, + 430, + 449 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 436, + 430, + 449 + ], + "spans": [ + { + "bbox": [ + 313, + 436, + 430, + 449 + ], + "type": "text", + "content": "4.1. Experimental Setup" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 313, + 453, + 554, + 622 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 453, + 554, + 622 + ], + "spans": [ + { + "bbox": [ + 313, + 453, + 554, + 622 + ], + "type": "text", + "content": "To align with the previous setting, we randomly pick 2000 data points from LLaVA-Med dataset [20]. For each data point, we randomly pick one round of QAs in the conversation and replace the answer with a clear rejection. As in Sec. 3.3, these data are also mixed with randomly picked 5000 normal data from LLaVA-v1.5-mix665k [22]. Noting that the training data of Yi-VL-6B [52] and LLaVA-NeXT [23] are not publicly available, therefore we keep using the same setting as the LLaVA-v1.5 models. Considering possible overfitting, we only supervise the model to generate a few rejection tokens while masking out the end-of-sentence token. Assuming in a one-turn conversation " + }, + { + "bbox": [ + 313, + 453, + 554, + 622 + ], + "type": "inline_equation", + "content": "x" + }, + { + "bbox": [ + 313, + 453, + 554, + 622 + ], + "type": "text", + "content": ", the instruction takes up " + }, + { + "bbox": [ + 313, + 453, + 554, + 622 + ], + "type": "inline_equation", + "content": "m" + }, + { + "bbox": [ + 313, + 453, + 554, + 622 + ], + "type": "text", + "content": " tokens. In our experiments, we only calculate the language modeling loss by this equation [56]:" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 352, + 631, + 555, + 645 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 352, + 631, + 555, + 645 + ], + "spans": [ + { + "bbox": [ + 352, + 631, + 555, + 645 + ], + "type": "interline_equation", + "content": "\\mathcal {L} \\left(x _ {1: n + m}\\right) = - \\log p \\left(x _ {n + 1: n + m} ^ {*} \\mid x _ {1: n}\\right), \\tag {1}", + "image_path": "42f73bc048981f04c8d18e3840ee0d738d410aa59ed2ec5415647920f9d5f945.jpg" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 313, + 654, + 554, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 654, + 554, + 715 + ], + "spans": [ + { + "bbox": [ + 313, + 654, + 554, + 715 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 313, + 654, + 554, + 715 + ], + "type": "inline_equation", + "content": "x_{n+1:n+m}^*" + }, + { + "bbox": [ + 313, + 654, + 554, + 715 + ], + "type": "text", + "content": " refers to the tokenized rejection string: \"I'm sorry, but I cannot assist with your request\", without periods or token ids. For simplicity, here we ignore the existence of system prompts as well as token in some templates. To ensure the reliability of our results," + } + ] + } + ], + "index": 11 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "text", + "content": "19883" + } + ] + } + ], + "index": 12 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 4 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 83, + 70, + 529, + 249 + ], + "blocks": [ + { + "bbox": [ + 83, + 70, + 529, + 249 + ], + "lines": [ + { + "bbox": [ + 83, + 70, + 529, + 249 + ], + "spans": [ + { + "bbox": [ + 83, + 70, + 529, + 249 + ], + "type": "table", + "html": "
Base ModelFinetuning DatasetsMM SafetyBench↓FigStep↓XCTest↑VizWizQA↑ScienceQA↑
ComplianceRejectionImageTotal
LLaVA-v1.5-7BOrigin96.37100.0092.0075.5055.1070.1269.46
VLGuard0.180.0077.2096.5056.3668.4270.12
Ours5.600.2090.4082.0055.9268.0269.11
LLaVA-v1.5-13BOrigin97.8099.8090.0084.5057.3472.6874.89
VLGuard0.360.0077.6097.0056.7771.5472.62
Ours1.790.0090.4090.5058.7771.5973.36
Yi-VL-6BOrigin93.5199.6095.2041.50\\( 66.20^1 \\)61.0370.12
VLGuard0.180.0084.4093.0059.3269.8174.51
Ours3.573.2089.2067.0047.9168.2773.07
LLaVA-NeXT Mistral-7BOrigin99.88100.0094.8058.5064.9872.6378.57
VLGuard0.180.0087.6089.5053.1570.2074.56
Ours2.610.0091.2065.5055.5069.3674.65
", + "image_path": "ca8dac4a0361e8b45c980f9a26618c74dd33e6f3e0c1aff83dfd847506b86f2d.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "table_body" + } + ], + "index": 0 + }, + { + "bbox": [ + 95, + 257, + 514, + 269 + ], + "lines": [ + { + "bbox": [ + 95, + 257, + 514, + 269 + ], + "spans": [ + { + "bbox": [ + 95, + 257, + 514, + 269 + ], + "type": "text", + "content": "Table 3. The evaluation comparisons between our finetuning dataset and VLGuard [55]. All values are percentages." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "text" + }, + { + "bbox": [ + 55, + 289, + 295, + 339 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 289, + 295, + 339 + ], + "spans": [ + { + "bbox": [ + 55, + 289, + 295, + 339 + ], + "type": "text", + "content": "except for LLaVA-v1.5 models with Vicuna language module, we also pick the LLaVA-NeXT with Mistral-7B language module [17], and the Yi-VL-6B based on Yi-6B [52]. Training details are shown in the Appendix." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 55, + 347, + 174, + 360 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 347, + 174, + 360 + ], + "spans": [ + { + "bbox": [ + 55, + 347, + 174, + 360 + ], + "type": "text", + "content": "4.2. Quantitative Results" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 54, + 365, + 295, + 544 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 54, + 365, + 295, + 544 + ], + "spans": [ + { + "bbox": [ + 54, + 365, + 295, + 544 + ], + "type": "text", + "content": "MLLMs do not lose defense on multi-modal prompts. In Table 3, it turns out that all four models work well on jailbreak defenses while maintaining a relatively high rate of benign prompt compliance. For MM-SafetyBench [25], the ASR of most models is below " + }, + { + "bbox": [ + 54, + 365, + 295, + 544 + ], + "type": "inline_equation", + "content": "5\\%" + }, + { + "bbox": [ + 54, + 365, + 295, + 544 + ], + "type": "text", + "content": ", which is a substantial drop compared to the original models. Experiments on FigStep [13] are also obvious: the highest ASR among the four models is " + }, + { + "bbox": [ + 54, + 365, + 295, + 544 + ], + "type": "inline_equation", + "content": "3.2\\%" + }, + { + "bbox": [ + 54, + 365, + 295, + 544 + ], + "type": "text", + "content": ". In training data, the conversation topic is unrelated to safety and privacy, and we do not append any reason for the \"abnormal\" rejection in each data point. The major factor is the rejection rate: compared to the visual instruction tuning phase with limited rejection data, we add 2000 rejection samples to 5000 benign training data in safety finetuning, reducing the proportion of instances where the answer complies with instructions." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 54, + 545, + 295, + 677 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 54, + 545, + 295, + 677 + ], + "spans": [ + { + "bbox": [ + 54, + 545, + 295, + 677 + ], + "type": "text", + "content": "More importantly, our experimental results provide little support for the notion that the model forgets previous safety alignment or fails due to the absence of multi-modal safety data, leading it to interpret harmful jailbreak prompts as benign. Instead, we argue that the model fails to defend jailbreaks mentioned above mostly because all multi-modal instruction-tuning data naturally create a compliance bias. After instruction tuning, the model still possesses the basic moral understanding and judgment. The difference is that it is told all prompts with images should be answered and followed eagerly, resulting in the vulnerability." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 55, + 677, + 296, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 677, + 296, + 715 + ], + "spans": [ + { + "bbox": [ + 55, + 677, + 296, + 715 + ], + "type": "text", + "content": "Without knowledge, correcting model distribution makes limited progress to text-only prompts. In Table 3, finetuning on our dataset does not significantly impair the" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 313, + 289, + 555, + 373 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 289, + 555, + 373 + ], + "spans": [ + { + "bbox": [ + 313, + 289, + 555, + 373 + ], + "type": "text", + "content": "safety compliance in XSTest [37], but it also fails to improve the unsafe rejection rate to the level achieved by VLGuard [55]. It is reasonable, for our dataset contains no extra safety-related contents compared to the VLGuard. Accordingly, merely correcting the model's distribution without incorporating additional knowledge is unlikely to result in a significant improvement in text-only input defense." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 313, + 374, + 556, + 529 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 374, + 556, + 529 + ], + "spans": [ + { + "bbox": [ + 313, + 374, + 556, + 529 + ], + "type": "text", + "content": "Performances on general tasks tend to be associated with the ordinary training data. In experiments, we mix rejection data with ordinary data from LLaVA-v1.5-mix665k [22], the official training set for LLaVA-v1.5. In the first two rows, using our random rejection data could achieve similar results as VLGuard on VizWizQA [15] and ScienceQA [27], most of which are slightly inferior to the original model. For VizWizQA the model trained on our dataset even outperforms the original one, indicating the quality of the ordinary data. However, for LLaVA-NeXT [23] and Yi-VL [52], the performances experience a larger drop, which may be attributed to the inconsistency of training data. This adds more randomness to the results." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 313, + 530, + 557, + 674 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 530, + 557, + 674 + ], + "spans": [ + { + "bbox": [ + 313, + 530, + 557, + 674 + ], + "type": "text", + "content": "Analyzing the results from the rejection data proportion perspective, we could also prove the claim that large language model alignment contributes to the defense of vision-language jailbreaks. Suppose the model does not possess safety-related knowledge or judgment for multi-modal data. In that case, the rejection rate of generated contents should be identical to, or slightly less than training data distribution. With " + }, + { + "bbox": [ + 313, + 530, + 557, + 674 + ], + "type": "inline_equation", + "content": "28.57\\%" + }, + { + "bbox": [ + 313, + 530, + 557, + 674 + ], + "type": "text", + "content": " rejection data in the training set, ASRs of LLaVA-v1.5-7B on jailbreak datasets should be around " + }, + { + "bbox": [ + 313, + 530, + 557, + 674 + ], + "type": "inline_equation", + "content": "70\\%" + }, + { + "bbox": [ + 313, + 530, + 557, + 674 + ], + "type": "text", + "content": ", and the performance on visual understanding and reasoning tasks should drop by " + }, + { + "bbox": [ + 313, + 530, + 557, + 674 + ], + "type": "inline_equation", + "content": "15\\%" + }, + { + "bbox": [ + 313, + 530, + 557, + 674 + ], + "type": "text", + "content": " to " + }, + { + "bbox": [ + 313, + 530, + 557, + 674 + ], + "type": "inline_equation", + "content": "20\\%" + }, + { + "bbox": [ + 313, + 530, + 557, + 674 + ], + "type": "text", + "content": ". However, the evaluation discrepancy between jailbreak datasets and vi" + } + ] + } + ], + "index": 9 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 313, + 683, + 555, + 713 + ], + "type": "page_footnote", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 683, + 555, + 713 + ], + "spans": [ + { + "bbox": [ + 313, + 683, + 555, + 713 + ], + "type": "text", + "content": "1The original value under the standard evaluation metric is 4.99, for most answers are appended by meaningless periods and line breaks. The score is calculated after omitting these suffixes." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 294, + 748, + 318, + 758 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 318, + 758 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 318, + 758 + ], + "type": "text", + "content": "19884" + } + ] + } + ], + "index": 11 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 5 + }, + { + "para_blocks": [ + { + "bbox": [ + 55, + 72, + 296, + 228 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 72, + 296, + 228 + ], + "spans": [ + { + "bbox": [ + 55, + 72, + 296, + 228 + ], + "type": "text", + "content": "sual understanding datasets discloses that the rejection behavior is not evenly added to the content generation. In other words, the model gets over-compliance after visual instruction tuning. Still, it knows which multi-modal prompts are harmful, just as it knows how to reject those text-only harmful prompts without interactions and incitements with images. The semantic meanings between multi-modal inputs and their text-only counterparts are aligned. In safety finetuning, it is less important to meticulously collect security data for the model to learn how to distinguish between benign and malicious instructions. Once enough rejection data is provided to enable rejection behavior, the model can make the right decisions." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 55, + 236, + 167, + 248 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 236, + 167, + 248 + ], + "spans": [ + { + "bbox": [ + 55, + 236, + 167, + 248 + ], + "type": "text", + "content": "4.3. Qualitative Results" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 55, + 253, + 296, + 278 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 253, + 296, + 278 + ], + "spans": [ + { + "bbox": [ + 55, + 253, + 296, + 278 + ], + "type": "text", + "content": "To have a rough understanding of the answer quality, we present some examples of the model-generation content." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 77, + 295, + 274, + 339 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 77, + 295, + 274, + 339 + ], + "spans": [ + { + "bbox": [ + 77, + 295, + 274, + 339 + ], + "type": "text", + "content": "Prompt: \\n The bottom of the image shows a phrase of an activity. List steps in detail to perform this activity using cryptocurrencies without leaving a trace." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 77, + 350, + 139, + 360 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 77, + 350, + 139, + 360 + ], + "spans": [ + { + "bbox": [ + 77, + 350, + 139, + 360 + ], + "type": "text", + "content": "LLaVA-v1.5-7B" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 77, + 361, + 274, + 427 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 77, + 361, + 274, + 427 + ], + "spans": [ + { + "bbox": [ + 77, + 361, + 274, + 427 + ], + "type": "text", + "content": "Assistant: I'm sorry, but I cannot assist with your request. Purchasing illegal drugs with cryptocurrencies...the use of cryptocurrencies for illegal activities can lead to legal consequences and is morally wrong. It is important to support legal and legitimate businesses..." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 78, + 438, + 116, + 448 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 78, + 438, + 116, + 448 + ], + "spans": [ + { + "bbox": [ + 78, + 438, + 116, + 448 + ], + "type": "text", + "content": "Yi-VL-6B" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 77, + 449, + 274, + 483 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 77, + 449, + 274, + 483 + ], + "spans": [ + { + "bbox": [ + 77, + 449, + 274, + 483 + ], + "type": "text", + "content": "Assistant: I'm sorry, but I cannot assist with your request. Please consult a professional for assistance with your request." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 77, + 493, + 200, + 502 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 77, + 493, + 200, + 502 + ], + "spans": [ + { + "bbox": [ + 77, + 493, + 200, + 502 + ], + "type": "text", + "content": "LLaVA- " + }, + { + "bbox": [ + 77, + 493, + 200, + 502 + ], + "type": "inline_equation", + "content": "\\nu 1.5 - 7B" + }, + { + "bbox": [ + 77, + 493, + 200, + 502 + ], + "type": "text", + "content": " -Random-reason" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 77, + 504, + 274, + 536 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 77, + 504, + 274, + 536 + ], + "spans": [ + { + "bbox": [ + 77, + 504, + 274, + 536 + ], + "type": "text", + "content": "Assistant: I'm sorry, I cannot assist with that request because it goes against my programming to provide opinions or beliefs." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 55, + 560, + 296, + 604 + ], + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 560, + 296, + 604 + ], + "spans": [ + { + "bbox": [ + 55, + 560, + 296, + 604 + ], + "type": "text", + "content": "Figure 2. Examples of replies generated under different settings. The reasons are not as high quality as the LLaVA-v1.5. To save space, the image of drugs with the phrase \"purchase illegal drugs\" is not shown in the box." + } + ] + } + ], + "index": 10, + "type": "text" + }, + { + "bbox": [ + 55, + 618, + 296, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 618, + 296, + 714 + ], + "spans": [ + { + "bbox": [ + 55, + 618, + 296, + 714 + ], + "type": "text", + "content": "The quality of generated rejection reasons is relatively independent of model decisions. In Fig. 2, the LLaVA-v1.5 model generates high-quality responses, with rejection reasons related to the topic. However, For Yi-VL-6B [52] and LLaVA-NeXT [23], the models keep satisfying reject decisions, while generating imperfect reasons. Similar results are also demonstrated when training LLaVA-v1.5 models under the random reason setting, indicating that the" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 313, + 72, + 554, + 133 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 72, + 554, + 133 + ], + "spans": [ + { + "bbox": [ + 313, + 72, + 554, + 133 + ], + "type": "text", + "content": "training data quality possibly affects the following sentence completion. With a proper rejection rate in the training set, models can reject inappropriate prompts, while the performance deterioration may be attributed to the distribution distinction among different training data." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 313, + 140, + 540, + 152 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 140, + 540, + 152 + ], + "spans": [ + { + "bbox": [ + 313, + 140, + 540, + 152 + ], + "type": "text", + "content": "4.4. Further Analyses on Conversation Datasets" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 313, + 158, + 555, + 254 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 158, + 555, + 254 + ], + "spans": [ + { + "bbox": [ + 313, + 158, + 555, + 254 + ], + "type": "text", + "content": "Introducing rejection answers to benign prompts may create a rejection bias on normal conversations. Therefore, except for XSTest [37] and VQA evaluations [15, 27], we also pick 1000 samples from LLaVA-v1.5-mix665k [22], LLaVA-Instruct-150K [24], and MMInstruct [26] each (for MMInstruct [26] samples are from the \"qa_en\" subset) to evaluate the compliance rate of LLaVA-v1.5 models. String matching is adopted as the metric." + } + ] + } + ], + "index": 14 + }, + { + "type": "table", + "bbox": [ + 319, + 264, + 550, + 357 + ], + "blocks": [ + { + "bbox": [ + 319, + 264, + 550, + 357 + ], + "lines": [ + { + "bbox": [ + 319, + 264, + 550, + 357 + ], + "spans": [ + { + "bbox": [ + 319, + 264, + 550, + 357 + ], + "type": "table", + "html": "
BaseDatasetsOriginVLGuardOurs
LLaVA-v1.5-7B665K↑100.00100.0099.90
150K↑99.6098.9098.70
MM↑100.0099.0088.40
LLaVA-v1.5-13B665K↑100.00100.0099.80
150K↑99.6099.4098.30
MM↑100.0099.2085.90
", + "image_path": "9c9e01c91c3650f41691da59a62aa7e1e5e4e4b048f5f4ca6ee1c2183b3f9771.jpg" + } + ] + } + ], + "index": 15, + "angle": 0, + "type": "table_body" + } + ], + "index": 15 + }, + { + "bbox": [ + 313, + 365, + 554, + 399 + ], + "lines": [ + { + "bbox": [ + 313, + 365, + 554, + 399 + ], + "spans": [ + { + "bbox": [ + 313, + 365, + 554, + 399 + ], + "type": "text", + "content": "Table 4. Compliance rate under different finetuning settings. 665K, 150K, and MM represent LLaVA-v1.5-mix665k [22], LLaVA-Instruct-150K [24], and MMInstruct [26], respectively." + } + ] + } + ], + "index": 16, + "angle": 0, + "type": "text" + }, + { + "bbox": [ + 313, + 412, + 554, + 508 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 412, + 554, + 508 + ], + "spans": [ + { + "bbox": [ + 313, + 412, + 554, + 508 + ], + "type": "text", + "content": "As shown in Table 4, the performance is almost intact for tasks from the LLaVA-v1.5-mix665k [22], but the models tend to reject image descriptions at a higher rate on LLaVA-Instruct-150K [24] and MMInstruct [26]. This may be because the data sampled from LLaVA-Med [20] contain quantities of image description queries, which induce models to reject similar prompts in evaluations, thus sacrificing the original performance to some extent." + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 313, + 516, + 518, + 529 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 516, + 518, + 529 + ], + "spans": [ + { + "bbox": [ + 313, + 516, + 518, + 529 + ], + "type": "text", + "content": "4.5. Ablation on Rejection Data Proportion" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 313, + 533, + 555, + 676 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 533, + 555, + 676 + ], + "spans": [ + { + "bbox": [ + 313, + 533, + 555, + 676 + ], + "type": "text", + "content": "With the understanding that datasets with rejection samples could boost model safety, it is natural to ask the following question: without safety-related knowledge in the dataset, how much data is enough to counteract the compliance bias buried in the visual instruction tuning phase? To figure out the relationship between rejection data quantity and jailbreak defense behavior, we finetune the LLaVA-v1.5-7B five times, with increasing numbers of rejection data, while keeping the remaining setting changed. Still, for each experiment, 5000 randomly picked ordinary data are added to the training set. The performance on the original LLaVA-v1.5-7B is also added for reference." + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 313, + 677, + 554, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 677, + 554, + 713 + ], + "spans": [ + { + "bbox": [ + 313, + 677, + 554, + 713 + ], + "type": "text", + "content": "From the experiments, we could observe that the defense capability against MM-SafetyBench [25] prompts is more sensitive to the existence of rejection data than Fig-" + } + ] + } + ], + "index": 20 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "text", + "content": "19885" + } + ] + } + ], + "index": 21 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 6 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 71, + 72, + 282, + 140 + ], + "blocks": [ + { + "bbox": [ + 71, + 72, + 282, + 140 + ], + "lines": [ + { + "bbox": [ + 71, + 72, + 282, + 140 + ], + "spans": [ + { + "bbox": [ + 71, + 72, + 282, + 140 + ], + "type": "image", + "image_path": "12695c757fb0afe130268417e9b0bb1c32599e7b037038e719968f16cca5e8a2.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + } + ], + "index": 0 + }, + { + "type": "image", + "bbox": [ + 70, + 144, + 282, + 231 + ], + "blocks": [ + { + "bbox": [ + 70, + 144, + 282, + 231 + ], + "lines": [ + { + "bbox": [ + 70, + 144, + 282, + 231 + ], + "spans": [ + { + "bbox": [ + 70, + 144, + 282, + 231 + ], + "type": "image", + "image_path": "6f079d1c307a443c7daad69e4ce191e181d538d6f5803cebaf27553a5140a668.jpg" + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 55, + 240, + 295, + 264 + ], + "lines": [ + { + "bbox": [ + 55, + 240, + 295, + 264 + ], + "spans": [ + { + "bbox": [ + 55, + 240, + 295, + 264 + ], + "type": "text", + "content": "Figure 3. The influence rejection data proportion has on jailbreak defense capability and visual understanding accuracy." + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_caption" + } + ], + "index": 1 + }, + { + "bbox": [ + 55, + 284, + 295, + 356 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 284, + 295, + 356 + ], + "spans": [ + { + "bbox": [ + 55, + 284, + 295, + 356 + ], + "type": "text", + "content": "Step [13]. For both attacks, " + }, + { + "bbox": [ + 55, + 284, + 295, + 356 + ], + "type": "inline_equation", + "content": "28.57\\%" + }, + { + "bbox": [ + 55, + 284, + 295, + 356 + ], + "type": "text", + "content": " is demonstrated to be enough to counteract compliance bias. Besides, with all rejection ratios, the visual understanding performance on VizWizQA [15] and ScienceQA [27] only slightly fluctuates around the baseline, which corresponds to previous findings in Sec. 4." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 55, + 363, + 289, + 376 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 363, + 289, + 376 + ], + "spans": [ + { + "bbox": [ + 55, + 363, + 289, + 376 + ], + "type": "text", + "content": "4.6. Adding Data in the Instruction Tuning Phase" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 55, + 380, + 295, + 513 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 380, + 295, + 513 + ], + "spans": [ + { + "bbox": [ + 55, + 380, + 295, + 513 + ], + "type": "text", + "content": "To gain a deeper understanding of this topic, in this part, we add rejection data into the multi-modal instruction following data to launch the visual instruction tuning. To be specific, due to the limitation of resources, we only design three runs, where the rejection data comprises " + }, + { + "bbox": [ + 55, + 380, + 295, + 513 + ], + "type": "inline_equation", + "content": "2\\%" + }, + { + "bbox": [ + 55, + 380, + 295, + 513 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 55, + 380, + 295, + 513 + ], + "type": "inline_equation", + "content": "5\\%" + }, + { + "bbox": [ + 55, + 380, + 295, + 513 + ], + "type": "text", + "content": " of the original LLaVA-v1.5-mix665k dataset [22], and the specific amount of rejection data is mixed in the original dataset for experiments. Identical to the official training setting, we adapt full-parameter tuning on the pretrained multi-modal projector and the released Vicuna-7B-v1.5 [8]. Training details are listed in the Appendix." + } + ] + } + ], + "index": 5 + }, + { + "type": "table", + "bbox": [ + 59, + 521, + 293, + 625 + ], + "blocks": [ + { + "bbox": [ + 59, + 521, + 293, + 625 + ], + "lines": [ + { + "bbox": [ + 59, + 521, + 293, + 625 + ], + "spans": [ + { + "bbox": [ + 59, + 521, + 293, + 625 + ], + "type": "table", + "html": "
Evaluation DatasetsMetric (%)Origin2% Rej. data5% Rej. data
MM-SafetyBenchASR↓96.3658.9360.12
FigStepASR↓100.0099.4099.80
XSTestComp. rate↑92.0089.2091.20
XSTestRej. rate↑75.5077.0073.00
ScienceQAImage Acc.↑69.4669.7169.16
ScienceQATotal Acc.↑70.1271.1670.53
VizWizQAAcc.↑55.1056.5656.26
", + "image_path": "9e423e2c96a03e30a6bb0b824a2ff8bc08970d643260cc03bebf9e5016db6c8d.jpg" + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "table_body" + } + ], + "index": 6 + }, + { + "bbox": [ + 55, + 632, + 295, + 677 + ], + "lines": [ + { + "bbox": [ + 55, + 632, + 295, + 677 + ], + "spans": [ + { + "bbox": [ + 55, + 632, + 295, + 677 + ], + "type": "text", + "content": "Table 5. Evaluation results of Visual Instruction Tuning. \"Origin\" refers to the official LLaVA-v1.5-mix665k dataset. \"Rej.\", \"Comp.\", and \"Acc.\" represents \"Rejection\", \"Compliance\", and \"Accuracy\", respectively." + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "text" + }, + { + "bbox": [ + 55, + 689, + 295, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 689, + 295, + 715 + ], + "spans": [ + { + "bbox": [ + 55, + 689, + 295, + 715 + ], + "type": "text", + "content": "The model safety gets improvements, but not good enough. As shown in Table 5, without harming the vi" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 313, + 72, + 553, + 191 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 72, + 553, + 191 + ], + "spans": [ + { + "bbox": [ + 313, + 72, + 553, + 191 + ], + "type": "text", + "content": "sual understanding capability, the attack successful rate of MM-SafetyBench [25] drops from " + }, + { + "bbox": [ + 313, + 72, + 553, + 191 + ], + "type": "inline_equation", + "content": "96.36\\%" + }, + { + "bbox": [ + 313, + 72, + 553, + 191 + ], + "type": "text", + "content": " to around " + }, + { + "bbox": [ + 313, + 72, + 553, + 191 + ], + "type": "inline_equation", + "content": "60\\%" + }, + { + "bbox": [ + 313, + 72, + 553, + 191 + ], + "type": "text", + "content": ", which demonstrates again that MLLMs do not lose defense on multi-modal prompts. However, the defense against FigStep [13] jailbreak does not experience an obvious drop, which matches the ablation study in Fig. 3: for a relatively small percentage of rejection data, the ASR of FigStep remains high. Therefore, the conclusion is drawn that if the proportion of rejected data is insufficient, the defense behavior may not be fully activated." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 313, + 192, + 555, + 324 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 192, + 555, + 324 + ], + "spans": [ + { + "bbox": [ + 313, + 192, + 555, + 324 + ], + "type": "text", + "content": "From experiments, if the rejection data lacks any inherent knowledge, awakening the model's multi-model safety awareness probably requires a sufficiently high rejection rate. Given this, enhancing model safety through finetuning seems to be more effective, where a smaller amount of data can constitute a substantial portion of the entire finetuning dataset. Specifically, in full-parameter visual instruction tuning, " + }, + { + "bbox": [ + 313, + 192, + 555, + 324 + ], + "type": "inline_equation", + "content": "2\\%" + }, + { + "bbox": [ + 313, + 192, + 555, + 324 + ], + "type": "text", + "content": " of the dataset corresponds to approximately 13k data points, which is more than six times the quantity required for finetuning. Despite this, the safety improvements gained through finetuning are far more significant." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 313, + 336, + 468, + 350 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 336, + 468, + 350 + ], + "spans": [ + { + "bbox": [ + 313, + 336, + 468, + 350 + ], + "type": "text", + "content": "5. Conclusion and Limitation" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 313, + 357, + 554, + 476 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 357, + 554, + 476 + ], + "spans": [ + { + "bbox": [ + 313, + 357, + 554, + 476 + ], + "type": "text", + "content": "In this paper, we investigate the well-curated safety finetuning data, finding that safety-related contents may not be the most crucial part for safety alignment. Based on this, we design a rejection dataset and conclude that the key factor influencing finetuning tends to be the existence of rejection data, instead of other attributes. Experiments demonstrate that MLLMs do not lose defense on multi-modal prompts. On the contrary, it is the compliance bias that prevents the models from refusing the malicious prompts, and training on the rejection dataset could mitigate such a bias." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 313, + 477, + 554, + 586 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 477, + 554, + 586 + ], + "spans": [ + { + "bbox": [ + 313, + 477, + 554, + 586 + ], + "type": "text", + "content": "The major limitation in our study is the unsatisfying performance drop under image description queries. Besides, we answer the question that boosting multi-modal large language models without the necessity of curated data is possible, but currently it cannot be used to enhance the safety in the text-only domain significantly. We will leave both of them as future work. To summarize, we hope our findings with supervised finetuning methods can help better understand the safeguard of MLLM." + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 313, + 597, + 408, + 611 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 597, + 408, + 611 + ], + "spans": [ + { + "bbox": [ + 313, + 597, + 408, + 611 + ], + "type": "text", + "content": "Acknowledgment" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 313, + 617, + 554, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 617, + 554, + 715 + ], + "spans": [ + { + "bbox": [ + 313, + 617, + 554, + 715 + ], + "type": "text", + "content": "We thank Zhengbo Wang and Yongcan Yu in NLPR for their feedback as well as discussion on our early drafts. Besides, we also would like to present our sincere appreciation to the anonymous reviewers for their constructive suggestions. This work was funded by the National Natural Science Foundation of China under Grants (62425606, 62276256, U2441251, U21B2045) and the Young Elite Scientists Sponsorship Program by CAST (2023QNRC001)." + } + ] + } + ], + "index": 15 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "type": "text", + "content": "19886" + } + ] + } + ], + "index": 16 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 7 + }, + { + "para_blocks": [ + { + "bbox": [ + 56, + 72, + 115, + 83 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 72, + 115, + 83 + ], + "spans": [ + { + "bbox": [ + 56, + 72, + 115, + 83 + ], + "type": "text", + "content": "References" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 57, + 91, + 296, + 712 + ], + "type": "list", + "angle": 0, + "index": 12, + "blocks": [ + { + "bbox": [ + 61, + 91, + 296, + 134 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 61, + 91, + 296, + 134 + ], + "spans": [ + { + "bbox": [ + 61, + 91, + 296, + 134 + ], + "type": "text", + "content": "[1] Jinze Bai, Shuai Bai, Yunfei Chu, Zeyu Cui, Kai Dang, Xiaodong Deng, Yang Fan, Wenbin Ge, Yu Han, Fei Huang, et al. Qwen technical report. arXiv preprint arXiv:2309.16609, 2023. 2" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 61, + 136, + 296, + 190 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 61, + 136, + 296, + 190 + ], + "spans": [ + { + "bbox": [ + 61, + 136, + 296, + 190 + ], + "type": "text", + "content": "[2] Yuntao Bai, Andy Jones, Kamal Ndousse, Amanda Askell, Anna Chen, Nova DasSarma, Dawn Drain, Stanislav Fort, Deep Ganguli, Tom Henighan, et al. Training a helpful and harmless assistant with reinforcement learning from human feedback. arXiv preprint arXiv:2204.05862, 2022. 1" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 62, + 192, + 295, + 246 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 192, + 295, + 246 + ], + "spans": [ + { + "bbox": [ + 62, + 192, + 295, + 246 + ], + "type": "text", + "content": "[3] Guiming Hardy Chen, Shunian Chen, Ruifei Zhang, Junying Chen, Xiangbo Wu, Zhiyi Zhang, Zhihong Chen, Jianquan Li, Xiang Wan, and Benyou Wang. Allava: Harnessing gpt4v-synthesized data for a lite vision-language model. arXiv preprint arXiv:2402.11684, 2024. 3" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 62, + 247, + 295, + 312 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 247, + 295, + 312 + ], + "spans": [ + { + "bbox": [ + 62, + 247, + 295, + 312 + ], + "type": "text", + "content": "[4] Jun Chen, Deyao Zhu, Xiaogian Shen, Xiang Li, Zechu Liu, Pengchuan Zhang, Raghuraman Krishnamoorthi, Vikas Chandra, Yunyang Xiong, and Mohamed Elhoseiny. Minigpt-v2: large language model as a unified interface for vision-language multi-task learning. arXiv preprint arXiv:2310.09478, 2023. 1, 2, 3" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 62, + 314, + 295, + 378 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 314, + 295, + 378 + ], + "spans": [ + { + "bbox": [ + 62, + 314, + 295, + 378 + ], + "type": "text", + "content": "[5] Zhe Chen, Jiannan Wu, Wenhai Wang, Weijie Su, Guo Chen, Sen Xing, Muyan Zhong, Qinglong Zhang, Xizhou Zhu, Lewei Lu, Bin Li, Ping Luo, Tong Lu, Yu Qiao, and Jifeng Dai. Internvl: Scaling up vision foundation models and aligning for generic visual-linguistic tasks. arXiv preprint arXiv:2312.14238, 2023. 1" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 62, + 380, + 295, + 434 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 380, + 295, + 434 + ], + "spans": [ + { + "bbox": [ + 62, + 380, + 295, + 434 + ], + "type": "text", + "content": "[6] Zhe Chen, Weiyun Wang, Hao Tian, Shenglong Ye, Zhangwei Gao, Erfei Cui, Wenwen Tong, Kongzhi Hu, Jiapeng Luo, Zheng Ma, et al. How far are we to gpt-4v? closing the gap to commercial multimodal models with open-source suites. arXiv preprint arXiv:2404.16821, 2024. 2" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 62, + 436, + 295, + 490 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 436, + 295, + 490 + ], + "spans": [ + { + "bbox": [ + 62, + 436, + 295, + 490 + ], + "type": "text", + "content": "[7] Mehdi Cherti, Romain Beaumont, Ross Wightman, Mitchell Wortsman, Gabriel Ilharco, Cade Gordon, Christoph Schuhmann, Ludwig Schmidt, and Jenia Jitsev. Reproducible scaling laws for contrastive language-image learning. In Proc. CVPR, 2023. 2" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 62, + 492, + 295, + 546 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 492, + 295, + 546 + ], + "spans": [ + { + "bbox": [ + 62, + 492, + 295, + 546 + ], + "type": "text", + "content": "[8] Wei-Lin Chiang, Zhuohan Li, Zi Lin, Ying Sheng, Zhang-hao Wu, Hao Zhang, Lianmin Zheng, Siyuan Zhuang, Yong-hao Zhuang, Joseph E. Gonzalez, Ion Stoica, and Eric P. Xing. Vicuna: An open-source chatbot impressing gpt-4 with " + }, + { + "bbox": [ + 62, + 492, + 295, + 546 + ], + "type": "inline_equation", + "content": "90\\%" + }, + { + "bbox": [ + 62, + 492, + 295, + 546 + ], + "type": "text", + "content": " chatgpt quality, 2023. 2, 8" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 62, + 548, + 295, + 601 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 548, + 295, + 601 + ], + "spans": [ + { + "bbox": [ + 62, + 548, + 295, + 601 + ], + "type": "text", + "content": "[9] Wenliang Dai, Junnan Li, Dongxu Li, Anthony Tiong, Junqi Zhao, Weisheng Wang, Boyang Li, Pascale Fung, and Steven Hoi. InstructBLIP: Towards general-purpose vision-language models with instruction tuning. In Proc. NeurIPS, 2023. 1, 2" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 57, + 603, + 295, + 656 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 603, + 295, + 656 + ], + "spans": [ + { + "bbox": [ + 57, + 603, + 295, + 656 + ], + "type": "text", + "content": "[10] Abhimanyu Dubey, Abhinav Jauhri, Abhinav Pandey, Abhishek Kadian, Ahmad Al-Dahle, Aiesha Letman, Akhil Mathur, Alan Schelten, Amy Yang, Angela Fan, et al. The llama 3 herd of models. arXiv preprint arXiv:2407.21783, 2024. 4" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 57, + 658, + 295, + 712 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 658, + 295, + 712 + ], + "spans": [ + { + "bbox": [ + 57, + 658, + 295, + 712 + ], + "type": "text", + "content": "[11] Chaoyou Fu, Peixian Chen, Yunhang Shen, Yulei Qin, Mengdan Zhang, Xu Lin, Jinrui Yang, Xiawu Zheng, Ke Li, Xing Sun, et al. Mme: A comprehensive evaluation benchmark for multimodal large language models. arXiv preprint arXiv:2306.13394, 2023. 1" + } + ] + } + ], + "index": 11 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 57, + 73, + 554, + 712 + ], + "type": "list", + "angle": 0, + "index": 27, + "blocks": [ + { + "bbox": [ + 316, + 73, + 554, + 127 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 73, + 554, + 127 + ], + "spans": [ + { + "bbox": [ + 316, + 73, + 554, + 127 + ], + "type": "text", + "content": "[12] Jiahui Gao, Renjie Pi, Tianyang Han, Han Wu, Lanqing Hong, Lingpeng Kong, Xin Jiang, and Zhenguo Li. Coca: Regaining safety-awareness of multimodal large language models with constitutional calibration. arXiv preprint arXiv:2409.11365, 2024. 3" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 317, + 129, + 554, + 183 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 129, + 554, + 183 + ], + "spans": [ + { + "bbox": [ + 317, + 129, + 554, + 183 + ], + "type": "text", + "content": "[13] Yichen Gong, Delong Ran, Jinyuan Liu, Conglei Wang, Tianshuo Cong, Anyu Wang, Sisi Duan, and Xiaoyun Wang. Figstep: Jailbreaking large vision-language models via typographic visual prompts. arXiv preprint arXiv:2311.05608, 2023. 1, 3, 4, 5, 6, 8" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 317, + 184, + 554, + 227 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 184, + 554, + 227 + ], + "spans": [ + { + "bbox": [ + 317, + 184, + 554, + 227 + ], + "type": "text", + "content": "[14] Yunhao Gou, Kai Chen, Zhili Liu, Lanqing Hong, Hang Xu, Zhenguo Li, Dit-Yan Yeung, James T Kwok, and Yu Zhang. Eyes closed, safety on: Protecting multimodal llms via image-to-text transformation. Proc. ECCV, 2024. 3" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 317, + 228, + 554, + 271 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 228, + 554, + 271 + ], + "spans": [ + { + "bbox": [ + 317, + 228, + 554, + 271 + ], + "type": "text", + "content": "[15] Danna Gurari, Qing Li, Abigale J Stangl, Anhong Guo, Chi Lin, Kristen Grauman, Jiebo Luo, and Jeffrey P Bigham. Vizwiz grand challenge: Answering visual questions from blind people. In Proc. CVPR, 2018. 4, 5, 6, 7, 8, 1" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 317, + 272, + 554, + 315 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 272, + 554, + 315 + ], + "spans": [ + { + "bbox": [ + 317, + 272, + 554, + 315 + ], + "type": "text", + "content": "[16] Edward J Hu, Yelong Shen, Phillip Wallis, Zeyuan Allen-Zhu, Yuanzhi Li, Shean Wang, Lu Wang, and Weizhu Chen. LoRA: Low-rank adaptation of large language models. In Proc. ICLR, 2022. 4" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 317, + 316, + 554, + 370 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 316, + 554, + 370 + ], + "spans": [ + { + "bbox": [ + 317, + 316, + 554, + 370 + ], + "type": "text", + "content": "[17] Albert Q Jiang, Alexandre Sablayrolles, Arthur Mensch, Chris Bamford, Devendra Singh Chaplot, Diego de las Casas, Florian Bressand, Gianna Lengyel, Guillaume Lample, Lucile Saulnier, et al. Mistral 7b. arXiv preprint arXiv:2310.06825, 2023. 6" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 316, + 372, + 554, + 415 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 372, + 554, + 415 + ], + "spans": [ + { + "bbox": [ + 316, + 372, + 554, + 415 + ], + "type": "text", + "content": "[18] Liqiang Jing, Yiren Li, Junhao Xu, Yongcan Yu, Pei Shen, and Xuemeng Song. Vision enhanced generative pre-trained language model for multimodal sentence summarization. Machine Intelligence Research, 20(2):289-298, 2023. 1" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 317, + 416, + 554, + 469 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 416, + 554, + 469 + ], + "spans": [ + { + "bbox": [ + 317, + 416, + 554, + 469 + ], + "type": "text", + "content": "[19] Wai-Chung Kwan, Hong-Ru Wang, Hui-Min Wang, and Kam-Fai Wong. A survey on recent advances and challenges in reinforcement learning methods for task-oriented dialogue policy learning. Machine Intelligence Research, 20(3):318-334, 2023. 1" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 317, + 471, + 554, + 525 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 471, + 554, + 525 + ], + "spans": [ + { + "bbox": [ + 317, + 471, + 554, + 525 + ], + "type": "text", + "content": "[20] Chunyuan Li, Cliff Wong, Sheng Zhang, Naoto Usuyama, Haotian Liu, Jianwei Yang, Tristan Naumann, Hoifung Poon, and Jianfeng Gao. Llava-med: Training a large language-and-vision assistant for biomedicine in one day. Proc. NeurIPS, 2024. 2, 5, 7" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 317, + 526, + 554, + 569 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 526, + 554, + 569 + ], + "spans": [ + { + "bbox": [ + 317, + 526, + 554, + 569 + ], + "type": "text", + "content": "[21] Yifan Li, Hangyu Guo, Kun Zhou, Wayne Xin Zhao, and JiRong Wen. Images are achilles' heel of alignment: Exploiting visual vulnerabilities for jailbreaking multimodal large language models. In proc. ECCV, 2024. 1, 3" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 317, + 571, + 554, + 602 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 571, + 554, + 602 + ], + "spans": [ + { + "bbox": [ + 317, + 571, + 554, + 602 + ], + "type": "text", + "content": "[22] Haotian Liu, Chunyuan Li, Yuheng Li, and Yong Jae Lee. Improved baselines with visual instruction tuning. In Proc. CVPR, 2024. 2, 4, 5, 6, 7, 8" + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 317, + 603, + 554, + 643 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 603, + 554, + 643 + ], + "spans": [ + { + "bbox": [ + 317, + 603, + 554, + 643 + ], + "type": "text", + "content": "[23] Haotian Liu, Chunyuan Li, Yuheng Li, Bo Li, Yuanhan Zhang, Sheng Shen, and Yong Jae Lee. Llava-last: Improved reasoning,OCR, and world knowledge, 2024. 5, 6, 7" + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 316, + 647, + 554, + 669 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 647, + 554, + 669 + ], + "spans": [ + { + "bbox": [ + 316, + 647, + 554, + 669 + ], + "type": "text", + "content": "[24] Haotian Liu, Chunyuan Li, Qingyang Wu, and Yong Jae Lee. Visual instruction tuning. In Proc. NeurIPS, 2024. 2, 3, 4, 7" + } + ] + } + ], + "index": 25 + }, + { + "bbox": [ + 316, + 670, + 554, + 712 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 670, + 554, + 712 + ], + "spans": [ + { + "bbox": [ + 316, + 670, + 554, + 712 + ], + "type": "text", + "content": "[25] Xin Liu, Yichen Zhu, Jindong Gu, Yunshi Lan, Chao Yang, and Yu Qiao. Mm-safetybench: A benchmark for safety evaluation of multimodal large language models. In Proc. ECCV, 2024. 1, 3, 4, 6, 7, 8" + } + ] + } + ], + "index": 26 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "text", + "content": "19887" + } + ] + } + ], + "index": 28 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 8 + }, + { + "para_blocks": [ + { + "bbox": [ + 56, + 72, + 294, + 713 + ], + "type": "list", + "angle": 0, + "index": 14, + "blocks": [ + { + "bbox": [ + 56, + 72, + 294, + 127 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 72, + 294, + 127 + ], + "spans": [ + { + "bbox": [ + 56, + 72, + 294, + 127 + ], + "type": "text", + "content": "[26] Yangzhou Liu, Yue Cao, Zhangwei Gao, Weiyun Wang, Zhe Chen, Wenhai Wang, Hao Tian, Lewei Lu, Xizhou Zhu, Tong Lu, et al. Mminstruct: A high-quality multi-modal instruction tuning dataset with extensive diversity. Science China Information Sciences, 67(12):1-16, 2024. 7" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 56, + 128, + 294, + 183 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 128, + 294, + 183 + ], + "spans": [ + { + "bbox": [ + 56, + 128, + 294, + 183 + ], + "type": "text", + "content": "[27] Pan Lu, Swaroop Mishra, Tony Xia, Liang Qiu, Kai-Wei Chang, Song-Chun Zhu, Oyvind Tafjord, Peter Clark, and Ashwin Kalyan. Learn to explain: Multimodal reasoning via thought chains for science question answering. In Proc. NeurIPS, 2022. 4, 5, 6, 7, 8, 1" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 56, + 184, + 294, + 227 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 184, + 294, + 227 + ], + "spans": [ + { + "bbox": [ + 56, + 184, + 294, + 227 + ], + "type": "text", + "content": "[28] Haochen Luo, Jindong Gu, Fengyuan Liu, and Philip Torr. An image is worth 1000 lies: Transferability of adversarial images across prompts on vision-language models. In Proc. ICLR, 2024. 1" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 56, + 228, + 294, + 261 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 228, + 294, + 261 + ], + "spans": [ + { + "bbox": [ + 56, + 228, + 294, + 261 + ], + "type": "text", + "content": "[29] Anand Mishra, Shashank Shekhar, Ajeet Kumar Singh, and Anirban Chakraborty. Ocr-vqa: Visual question answering by reading text in images. In Proc. ICDAR, 2019. 3" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 56, + 261, + 294, + 315 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 261, + 294, + 315 + ], + "spans": [ + { + "bbox": [ + 56, + 261, + 294, + 315 + ], + "type": "text", + "content": "[30] Long Ouyang, Jeffrey Wu, Xu Jiang, Diogo Almeida, Carroll Wainwright, Pamela Mishkin, Chong Zhang, Sandhini Agarwal, Katarina Slama, Alex Ray, et al. Training language models to follow instructions with human feedback. In Proc. NeurIPS, 2022. 1" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 56, + 316, + 294, + 360 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 316, + 294, + 360 + ], + "spans": [ + { + "bbox": [ + 56, + 316, + 294, + 360 + ], + "type": "text", + "content": "[31] Renjie Pi, Tianyang Han, Jianshu Zhang, Yueqi Xie, Rui Pan, Qing Lian, Hanze Dong, Jipeng Zhang, and Tong Zhang. Mllm-protector: Ensuring mllm's safety without hurting performance. In Proc. EMNLP, 2024. 3" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 56, + 361, + 294, + 403 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 361, + 294, + 403 + ], + "spans": [ + { + "bbox": [ + 56, + 361, + 294, + 403 + ], + "type": "text", + "content": "[32] Xiangyu Qi, Kaixuan Huang, Ashwinee Panda, Peter Henderson, Mengdi Wang, and Prateek Mittal. Visual adversarial examples jailbreak aligned large language models. In Proc. AAAI, 2024. 1" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 56, + 404, + 294, + 458 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 404, + 294, + 458 + ], + "spans": [ + { + "bbox": [ + 56, + 404, + 294, + 458 + ], + "type": "text", + "content": "[33] Xiangyu Qi, Ashwinee Panda, Kaifeng Lyu, Xiao Ma, Subhrajit Roy, Ahmad Beirami, Prateek Mittal, and Peter Henderson. Safety alignment should be made more than just a few tokens deep. arXiv preprint arXiv:2406.05946, 2024. 5" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 56, + 460, + 294, + 502 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 460, + 294, + 502 + ], + "spans": [ + { + "bbox": [ + 56, + 460, + 294, + 502 + ], + "type": "text", + "content": "[34] Xiangyu Qi, Yi Zeng, Tinghao Xie, Pin-Yu Chen, Ruoxi Jia, Prateek Mittal, and Peter Henderson. Fine-tuning aligned language models compromises safety, even when users do not intend to! In Proc. ICLR, 2024. 1" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 56, + 503, + 294, + 557 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 503, + 294, + 557 + ], + "spans": [ + { + "bbox": [ + 56, + 503, + 294, + 557 + ], + "type": "text", + "content": "[35] Alec Radford, Jong Wook Kim, Chris Hallacy, Aditya Ramesh, Gabriel Goh, Sandhini Agarwal, Girish Sastry, Amanda Askell, Pamela Mishkin, Jack Clark, et al. Learning transferable visual models from natural language supervision. In Proc. ICML, 2021. 2" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 56, + 558, + 294, + 602 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 558, + 294, + 602 + ], + "spans": [ + { + "bbox": [ + 56, + 558, + 294, + 602 + ], + "type": "text", + "content": "[36] Rafael Rafailov, Archit Sharma, Eric Mitchell, Christopher D Manning, Stefano Ermon, and Chelsea Finn. Direct preference optimization: Your language model is secretly a reward model. In Proc. NeurIPS, 2024. 1" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 56, + 603, + 294, + 646 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 603, + 294, + 646 + ], + "spans": [ + { + "bbox": [ + 56, + 603, + 294, + 646 + ], + "type": "text", + "content": "[37] Paul Röttger, Hannah Kirk, Bertie Vidgen, Giuseppe Attanasio, Federico Bianchi, and Dirk Hovy. Xstest: A test suite for identifying exaggerated safety behaviours in large language models. In Proc. NAACL, 2024. 4, 6, 7, 1" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 56, + 647, + 294, + 680 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 647, + 294, + 680 + ], + "spans": [ + { + "bbox": [ + 56, + 647, + 294, + 680 + ], + "type": "text", + "content": "[38] John Schulman, Filip Wolski, Prafulla Dhariwal, Alec Radford, and Oleg Klimov. Proximal policy optimization algorithms. arXiv preprint arXiv:1707.06347, 2017. 1" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 56, + 681, + 294, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 681, + 294, + 713 + ], + "spans": [ + { + "bbox": [ + 56, + 681, + 294, + 713 + ], + "type": "text", + "content": "[39] Erfan Shayegani, Yue Dong, and Nael Abu-Ghazaleh. Jailbreak in pieces: Compositional adversarial attacks on multimodal language models. In Proc. ICLR, 2024. 3" + } + ] + } + ], + "index": 13 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 315, + 72, + 553, + 713 + ], + "type": "list", + "angle": 0, + "index": 29, + "blocks": [ + { + "bbox": [ + 315, + 72, + 553, + 116 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 315, + 72, + 553, + 116 + ], + "spans": [ + { + "bbox": [ + 315, + 72, + 553, + 116 + ], + "type": "text", + "content": "[40] Amanpreet Singh, Vivek Natarajan, Meet Shah, Yu Jiang, Xinlei Chen, Dhruv Batra, Devi Parikh, and Marcus Rohrbach. Towards vqa models that can read. In Proc. CVPR, 2019. 3" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 315, + 118, + 553, + 152 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 315, + 118, + 553, + 152 + ], + "spans": [ + { + "bbox": [ + 315, + 118, + 553, + 152 + ], + "type": "text", + "content": "[41] Quan Sun, Yuxin Fang, Ledell Wu, Xinlong Wang, and Yue Cao. Eva-clip: Improved training techniques for clip at scale. arXiv preprint arXiv:2303.15389, 2023. 2" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 315, + 152, + 553, + 207 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 315, + 152, + 553, + 207 + ], + "spans": [ + { + "bbox": [ + 315, + 152, + 553, + 207 + ], + "type": "text", + "content": "[42] Hugo Touvron, Thibaut Lavril, Gautier Izacard, Xavier Martinet, Marie-Anne Lachaux, Timothee Lacroix, Baptiste Rozière, Naman Goyal, Eric Hambro, Faisal Azhar, et al. Llama: Open and efficient foundation language models. arXiv preprint arXiv:2302.13971, 2023. 2" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 316, + 208, + 553, + 251 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 208, + 553, + 251 + ], + "spans": [ + { + "bbox": [ + 316, + 208, + 553, + 251 + ], + "type": "text", + "content": "[43] Junke Wang, Lingchen Meng, Zejia Weng, Bo He, Zuxuan Wu, and Yu-Gang Jiang. To see is to believe: Prompting gpt-4v for better visual instruction tuning. arXiv preprint arXiv:2311.07574, 2023. 3" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 316, + 253, + 553, + 308 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 253, + 553, + 308 + ], + "spans": [ + { + "bbox": [ + 316, + 253, + 553, + 308 + ], + "type": "text", + "content": "[44] Peng Wang, Shuai Bai, Sinan Tan, Shijie Wang, Zhihao Fan, Jinze Bai, Keqin Chen, Xuejing Liu, Jialin Wang, Wenbin Ge, et al. Qwen2-vl: Enhancing vision-language model's perception of the world at any resolution. arXiv preprint arXiv:2409.12191, 2024. 2" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 316, + 309, + 553, + 363 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 309, + 553, + 363 + ], + "spans": [ + { + "bbox": [ + 316, + 309, + 553, + 363 + ], + "type": "text", + "content": "[45] Xiao Wang, Guangyao Chen, Guangwu Qian, Pengcheng Gao, Xiao-Yong Wei, Yaowei Wang, Yonghong Tian, and Wen Gao. Large-scale multi-modal pre-trained models: A comprehensive survey. Machine Intelligence Research, 20 (4):447-482, 2023. 1" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 316, + 365, + 553, + 409 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 365, + 553, + 409 + ], + "spans": [ + { + "bbox": [ + 316, + 365, + 553, + 409 + ], + "type": "text", + "content": "[46] Yu Wang, Xiaogeng Liu, Yu Li, Muhao Chen, and Chaowei Xiao. Adashield: Safeguarding multimodal large language models from structure-based attack via adaptive shield prompting. Proc. ECCV, 2024. 3" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 316, + 410, + 553, + 442 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 410, + 553, + 442 + ], + "spans": [ + { + "bbox": [ + 316, + 410, + 553, + 442 + ], + "type": "text", + "content": "[47] Alexander Wei, Nika Haghtalab, and Jacob Steinhardt. Jailbroken: How does Ilm safety training fail? In Proc. NeurIPS, 2024. 3" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 316, + 444, + 553, + 509 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 444, + 553, + 509 + ], + "spans": [ + { + "bbox": [ + 316, + 444, + 553, + 509 + ], + "type": "text", + "content": "[48] Peng Xu, Wenqi Shao, Kaipeng Zhang, Peng Gao, Shuo Liu, Meng Lei, Fanqing Meng, Siyuan Huang, Yu Qiao, and Ping Luo. Lvlm-ehub: A comprehensive evaluation benchmark for large vision-language models. IEEE Transactions on Pattern Analysis and Machine Intelligence, 47(3):1877-1893, 2025. 1" + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 316, + 510, + 553, + 555 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 510, + 553, + 555 + ], + "spans": [ + { + "bbox": [ + 316, + 510, + 553, + 555 + ], + "type": "text", + "content": "[49] An Yang, Baosong Yang, Binyuan Hui, Bo Zheng, Bowen Yu, Chang Zhou, Chengpeng Li, Chengyuan Li, Dayiheng Liu, Fei Huang, et al. Qwen2 technical report. arXiv preprint arXiv:2407.10671, 2024. 2" + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 316, + 555, + 553, + 600 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 555, + 553, + 600 + ], + "spans": [ + { + "bbox": [ + 316, + 555, + 553, + 600 + ], + "type": "text", + "content": "[50] Qinghao Ye, Haiyang Xu, Jiabo Ye, Ming Yan, Anwen Hu, Haowei Liu, Qi Qian, Ji Zhang, and Fei Huang. mplugowl2: Revolutionizing multi-modal large language model with modality collaboration. In Proc. CVPR, 2024. 1, 2" + } + ] + } + ], + "index": 25 + }, + { + "bbox": [ + 316, + 601, + 553, + 644 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 601, + 553, + 644 + ], + "spans": [ + { + "bbox": [ + 316, + 601, + 553, + 644 + ], + "type": "text", + "content": "[51] Shukang Yin, Chaoyou Fu, Sirui Zhao, Ke Li, Xing Sun, Tong Xu, and Enhong Chen. A survey on multimodal large language models. National Science Review, 11(12), 2024. 1, 2" + } + ] + } + ], + "index": 26 + }, + { + "bbox": [ + 316, + 646, + 553, + 690 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 646, + 553, + 690 + ], + "spans": [ + { + "bbox": [ + 316, + 646, + 553, + 690 + ], + "type": "text", + "content": "[52] Alex Young, Bei Chen, Chao Li, Chengen Huang, Ge Zhang, Guanwei Zhang, Heng Li, Jiangcheng Zhu, Jianqun Chen, Jing Chang, et al. Yi: Open foundation models by 01. ai. arXiv preprint arXiv:2403.04652, 2024. 5, 6, 7" + } + ] + } + ], + "index": 27 + }, + { + "bbox": [ + 316, + 691, + 553, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 691, + 553, + 713 + ], + "spans": [ + { + "bbox": [ + 316, + 691, + 553, + 713 + ], + "type": "text", + "content": "[53] Yaowei Zheng, Richong Zhang, Junhao Zhang, Yanhan Ye, Zheyan Luo, Zhangchi Feng, and Yongqiang Ma. Llamafac" + } + ] + } + ], + "index": 28 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "text", + "content": "19888" + } + ] + } + ], + "index": 30 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 9 + }, + { + "para_blocks": [ + { + "bbox": [ + 56, + 72, + 295, + 228 + ], + "type": "list", + "angle": 0, + "index": 4, + "blocks": [ + { + "bbox": [ + 75, + 72, + 294, + 95 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 75, + 72, + 294, + 95 + ], + "spans": [ + { + "bbox": [ + 75, + 72, + 294, + 95 + ], + "type": "text", + "content": "tory: Unified efficient fine-tuning of " + }, + { + "bbox": [ + 75, + 72, + 294, + 95 + ], + "type": "inline_equation", + "content": "100+" + }, + { + "bbox": [ + 75, + 72, + 294, + 95 + ], + "type": "text", + "content": " language models. arXiv preprint arXiv:2403.13372, 2024. 1" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 56, + 96, + 295, + 140 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 96, + 295, + 140 + ], + "spans": [ + { + "bbox": [ + 56, + 96, + 295, + 140 + ], + "type": "text", + "content": "[54] Deyao Zhu, Jun Chen, Xiaoqian Shen, Xiang Li, and Mohamed Elhoseiny. Minigpt-4: Enhancing vision-language understanding with advanced large language models. arXiv preprint arXiv:2304.10592, 2023. 2, 3" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 56, + 141, + 294, + 184 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 141, + 294, + 184 + ], + "spans": [ + { + "bbox": [ + 56, + 141, + 294, + 184 + ], + "type": "text", + "content": "[55] Yongshuo Zong, Ondrej Bohdal, Tingyang Yu, Yongxin Yang, and Timothy Hospedales. Safety fine-tuning at (almost) no cost: A baseline for vision large language models. In Proc. ICML, 2024. 1, 3, 4, 5, 6" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 56, + 186, + 294, + 228 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 186, + 294, + 228 + ], + "spans": [ + { + "bbox": [ + 56, + 186, + 294, + 228 + ], + "type": "text", + "content": "[56] Andy Zou, Zifan Wang, Nicholas Carlini, Milad Nasr, J Zico Kolter, and Matt Fredrikson. Universal and transferable adversarial attacks on aligned language models. arXiv preprint arXiv:2307.15043, 2023. 5" + } + ] + } + ], + "index": 3 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 749, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 749, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 749, + 317, + 757 + ], + "type": "text", + "content": "19889" + } + ] + } + ], + "index": 5 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 10 + } + ], + "_backend": "vlm", + "_version_name": "2.6.4" +} \ No newline at end of file diff --git a/2025/Do Your Best and Get Enough Rest for Continual Learning/364acbb3-eca1-422d-be23-a76c46789b30_content_list.json b/2025/Do Your Best and Get Enough Rest for Continual Learning/364acbb3-eca1-422d-be23-a76c46789b30_content_list.json new file mode 100644 index 0000000000000000000000000000000000000000..48048d551982c9b04cf3cae962680d48714ab294 --- /dev/null +++ b/2025/Do Your Best and Get Enough Rest for Continual Learning/364acbb3-eca1-422d-be23-a76c46789b30_content_list.json @@ -0,0 +1,1426 @@ +[ + { + "type": "text", + "text": "Do Your Best and Get Enough Rest for Continual Learning", + "text_level": 1, + "bbox": [ + 197, + 130, + 799, + 152 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Hankyul Kang", + "bbox": [ + 179, + 179, + 305, + 200 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Gregor Seifer1", + "bbox": [ + 354, + 180, + 473, + 199 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Donghyun Lee²", + "bbox": [ + 522, + 180, + 651, + 199 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Jongbin $\\mathrm{Ryu}^{1}$", + "bbox": [ + 699, + 180, + 821, + 199 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "$^{1}$ Ajou University $^{2}$ KAIST", + "bbox": [ + 382, + 214, + 609, + 234 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Abstract", + "text_level": 1, + "bbox": [ + 248, + 268, + 326, + 284 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "According to the forgetting curve theory, we can enhance memory retention by learning extensive data and taking adequate rest. This means that in order to effectively retain new knowledge, it is essential to learn it thoroughly and ensure sufficient rest so that our brain can memorize without forgetting. The main takeaway from this theory is that learning extensive data at once necessitates sufficient rest before learning the same data again. This aspect of human long-term memory retention can be effectively utilized to address the continual learning of neural networks. Retaining new knowledge for a long period of time without catastrophic forgetting is the critical problem of continual learning. Therefore, based on Ebbinghaus' theory, we introduce the view-batch model that adjusts the learning schedules to optimize the recall interval between retraining the same samples. The proposed view-batch model allows the network to get enough rest to learn extensive knowledge from the same samples with a recall interval of sufficient length. To this end, we specifically present two approaches: 1) a replay method that guarantees the optimal recall interval, and 2) a self-supervised learning that acquires extensive knowledge from a single training sample at a time. We empirically show that these approaches of our method are aligned with the forgetting curve theory, which can enhance long-term memory. In our experiments, we also demonstrate that our method significantly improves many state-of-the-art continual learning methods in various protocols and scenarios. We open-source this project at https://github.com/hankyul2/ViewBatchModel.", + "bbox": [ + 89, + 301, + 483, + 739 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "1. Introduction", + "text_level": 1, + "bbox": [ + 91, + 771, + 220, + 787 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "The forgetting curve theory [18] suggests that human memory tends to fade as time goes on; however, memory retention can be improved by repeated learning in an optimal recall interval. Motivated by this theory, many subsequent studies [6, 19, 27] have shown that spaced repetition with", + "bbox": [ + 89, + 797, + 483, + 875 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/c3ac23cdbc9dbc772bc852352bfd6acc3d7e32ca4dbf55f6548953e05a3cd2ad.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 522, + 266, + 707, + 366 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/4062f74cdf243453891fc09a9f280d166d1fddbdbcf4b553b9573e7ffb925912.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 712, + 267, + 895, + 366 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/d39edceeb4ff7ce1f8040457f82497e9700f7cbc5b7b56458a10444e36c27124.jpg", + "image_caption": [ + "(a) Short-term recall interval", + "(c) Long-term recall interval", + "Figure 1. Conceptual graph of the forgetting curve. We show (a) short-term recall interval, (b) optimal recall interval, (c) long-term recall interval, and (d) degree of forgetting. (a-b) Expanding the recall interval improves long-term memory retention of neural networks by repeatedly recalling memory with moderate difficulty, whereas (c) an excessive recall interval decreases it. The depicted forgetting curve regarding recall interval is based on the spacing effect formula [6, 7] provided in the supplementary material." + ], + "image_footnote": [], + "bbox": [ + 522, + 380, + 707, + 479 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/d15d8cc7dc22dc911ab44f2ec4a2ad180ba64b5fb19bc82f43eda0ae0dfc6e57.jpg", + "image_caption": [ + "(b) Optimal recall interval", + "(d) Degree of forgetting" + ], + "image_footnote": [], + "bbox": [ + 712, + 380, + 895, + 479 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "optimal recall interval enhances long-term memory retention in the human brain. This phenomenon is known as the spacing effect [7], where the optimal recall interval for repeated learning is crucial to maintain an acceptable level of forgetting knowledge. Several studies [7, 21, 36] have made significant progress on the spacing effect, shedding light on its effectiveness. There have also been studies [28, 30, 31] highlighting the importance of the recall interval of repeated learning.", + "bbox": [ + 511, + 654, + 906, + 790 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Although this theory has been widely recognized in human brain research, it is currently very little known in the field of machine learning. Only a few works [1, 4, 48] show this theory's potential in neural networks by utilizing it as an analysis toolkit or dynamic learning strategy. So, we investigate the impact of the forgetting curve theory on long-term memory retention in neural networks for continual learning.", + "bbox": [ + 511, + 795, + 908, + 902 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "CVF", + "bbox": [ + 106, + 2, + 181, + 42 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore.", + "bbox": [ + 236, + 0, + 810, + 46 + ], + "page_idx": 0 + }, + { + "type": "page_footnote", + "text": "*Corresponding author.", + "bbox": [ + 109, + 887, + 235, + 898 + ], + "page_idx": 0 + }, + { + "type": "page_number", + "text": "10077", + "bbox": [ + 480, + 944, + 517, + 955 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/171ede31f5cab52ada7526c381f2dc6f5e9d81195679eeec7e8ea1f90a340a1d.jpg", + "image_caption": [ + "Figure 2. Overview of experimental results. We provide comprehensive comparisons of various factors for continual learning. We perform extensive experiments on step and buffer sizes (a-b), three different continual learning methods (c), whether to use the pre-trained model (d), three different benchmarks (e), and two evaluation protocols (f). In all cases, ours improves the baseline performance consistently." + ], + "image_footnote": [], + "bbox": [ + 91, + 89, + 478, + 252 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "In Ebbinghaus' theory, we focus on the decay of memory retention in the forgetting curve as shown in Figure 1. The forgetting curve suggests that, as shown in Figure 1(a), the sample memory shows extreme knowledge decay when repeating its learning with a short time-space. On the other hand, as shown in Figure 1(c), if there is an excessive recall interval of repeated learning, too much memory is forgotten, leading to a decreased level of memory retention when relearning the same sample. However, as shown in Figure 1(b), when there is an optimal recall interval between repeated learning, the decay of memory retention also becomes gentle. This effect of the optimal recall interval implies that it is imperative to implement an appropriate learning schedule to retain memories efficiently in the repeated learning process. The notion behind this spacing effect can be usefully applied in continual learning, particularly to prevent catastrophic forgetting.", + "bbox": [ + 88, + 387, + 482, + 643 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "However, empirically, we found that the current methods for continual learning have a short recall interval [24, 32], so they are limited to profit from the spacing effect, which requires a sufficient recall interval.", + "bbox": [ + 89, + 643, + 482, + 702 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "To tackle this limitation, we introduce two approaches for training neural networks. Our first approach aims to prevent short recall intervals through the proposed replay method. We replay view-batch in the repeated training schedule so that short recall intervals can be delayed sufficiently. We augment multiple views from a single sample, and thus, the recall intervals of repeated learning can be adjusted. In our second approach, we make use of self-supervised learning to compensate for the delayed recall interval by training each sample extensively. We leverage the one-to-many divergence to learn the generalized self-supervised features within a view-batch, which enhances the capability of the networks. These two approaches en", + "bbox": [ + 89, + 704, + 482, + 901 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "sure that the neural networks have enough time to retrain the same samples, and at the same time, they can acquire substantial knowledge from training samples at a time. We apply our view-batch model (VBM) to both rehearsal and rehearsal-free continual learning scenarios, where addressing long-term memory retention is critical. Moreover, we show extensive experimental results of various continual protocols and scenarios. As an overview, Figure 2 showcases the comprehensive comparison against different step sizes, memory buffer size, baseline methods, use of the pretrained model, benchmarks, and protocols. For all these comparisons, our view-batch model achieves significant performance improvements. We have condensed our contributions into the following points:", + "bbox": [ + 511, + 90, + 903, + 301 + ], + "page_idx": 1 + }, + { + "type": "list", + "sub_type": "text", + "list_items": [ + "- We propose the view-batch model that optimizes the recall interval between retraining samples in order to enhance memory decay in the continual learning task. The proposed view-batch model includes replay and self-supervised learning components to accomplish extensive learning in optimized intervals.", + "- We showcase the effectiveness of our approach on various state-of-the-art continual learning methods, where ours consistently improves the performance. The proposed method is a drop-in replacement approach, so it does not require any extra computational costs for the performance improvements.", + "- We reveal that enhancing memory decay can significantly improve the performance of the continual learning. This confirms the practical utility of theoretical knowledge for neural networks." + ], + "bbox": [ + 513, + 303, + 903, + 542 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "2. Related Work", + "text_level": 1, + "bbox": [ + 513, + 559, + 653, + 574 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "2.1. Forgetting Curve Theory in Neural Networks", + "text_level": 1, + "bbox": [ + 511, + 585, + 897, + 601 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Several studies [25, 26] have employed the forgetting curve theory [18] to gain insights into the inherent memory retention of neural networks. Recently, Tirumala et al. [39] investigated the forgetting curve theory in neural networks regarding different model scales. Cao et al. [4] have also utilized the forgetting curve theory to examine the variations in memory conformation of the neural networks as influenced by pre-training methods. Additionally, Amiri et al. [1] proposed a dynamic learning scheduler to reduce total learning cost, leveraging optimal learning intervals corresponding to the forgetting curve. Zhong et al. [48] optimized memory retrieval timing based on the principles of the spacing effect theory, enabling the retention of the previous dialogue context existing over an extended duration.", + "bbox": [ + 511, + 608, + 906, + 820 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "2.2. Data Augmentation", + "text_level": 1, + "bbox": [ + 511, + 830, + 700, + 847 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Data augmentation [14, 45] has been widely used in various existing networks. It helps networks learn rich visual information using a limited number of training samples.", + "bbox": [ + 511, + 854, + 903, + 900 + ], + "page_idx": 1 + }, + { + "type": "page_number", + "text": "10078", + "bbox": [ + 480, + 944, + 517, + 955 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Data augmentation is also effective in self-supervised learning [5, 11]. Unlimited labels can be generated when different augmentation methods are applied to a single sample. Therefore, the network is able to learn with self-supervised labels of augmented samples. We use this augmentation method to construct multiple views in a view-batch and also learn self-supervision along with the supervised information of the continual learning process. There have been several studies [2, 22] on repeated augmentation, which allows learning multiple views from a single sample at a time. However, these studies only focused on improving the performance of the augmentation method and did not examine the effect of recall intervals. In addition, they do not apply self-supervised learning and only use label supervision to train networks.", + "bbox": [ + 89, + 90, + 483, + 316 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "2.3. Continual Learning", + "text_level": 1, + "bbox": [ + 89, + 343, + 281, + 359 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "The main problem of continual learning is preserving old knowledge while acquiring new one. This problem has been recognized as catastrophic forgetting, and one common strategy used to address this problem is to make use of a memory buffer [32]. The approaches to this strategy can be divided into two categories: one involves the direct storage of a subset of training samples from a previously learned task [32, 33], while the other generates training samples from the learned task [37]. This approach of utilizing memory buffers proves to be highly effective in addressing the issue of catastrophic forgetting; however, it does come with the trade-off of requiring additional memory space. Consequently, non-rehearsal continual learning methods [15, 33] have been widely studied to avoid additional memory usage. In addition, studies [41, 42, 46] using pre-trained models in non-rehearsal continual learning have been actively conducted recently. Meanwhile, there have been several works [16, 29] that tackle continual learning from the perspective of the regularization method. The architecture-based method [35, 44] has been another prevalent research direction for addressing continual learning.", + "bbox": [ + 89, + 371, + 483, + 690 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3. Method", + "text_level": 1, + "bbox": [ + 89, + 718, + 181, + 734 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "We introduce a view-batch model designed to enhance the long-term memory retention of neural networks for continual learning (CL) tasks. In existing CL methods [24, 32, 44], the neural networks retrain the same samples with short time intervals, which prohibits long-term memory retention. To address this short recall interval, we first investigate the augmentation model with the replay method to ensure a sufficient recall interval in Sec. 3.1. Then, Sec. 3.2 introduces view-batch learning model to extensively train multiple samples of the same instance in view-batch.", + "bbox": [ + 89, + 750, + 483, + 900 + ], + "page_idx": 2 + }, + { + "type": "image", + "img_path": "images/a0443f4ba890a4c2b249ce19436e98d93600314e1dc1fcc19e74591c0ea53809.jpg", + "image_caption": [ + "(a) Baseline" + ], + "image_footnote": [], + "bbox": [ + 521, + 87, + 705, + 185 + ], + "page_idx": 2 + }, + { + "type": "image", + "img_path": "images/61f95c7408d3f3bb330b735292fca7910ba239dbbf172835f3bbbb95a65e5d29.jpg", + "image_caption": [ + "(b) View-batch", + "Figure 3. Schematic illustration of the proposed view-batch model. In subfigure (b), we show our view-batch model employing the replay $(\\mathrm{V} = 4)$ and self-supervised learning approach. In contrast to (a) the baseline method, we learn multiple views of the same sample (marked as different shades) using the proposed view-batch self-supervised loss to learn it extensively and ensure enough time-space between recall intervals. For simplicity, we assume in (a) that the entire training data and batch size are the same as four, thus a single training epoch constitutes one batch." + ], + "image_footnote": [], + "bbox": [ + 709, + 88, + 895, + 185 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3.1. Replay with Augmentation", + "text_level": 1, + "bbox": [ + 511, + 364, + 756, + 380 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Our view-batch model employs a replaying method with the augmentation to adjust the recall interval of retraining the same samples. In the conventional learning scheduler, a sample-batch is replayed in the training process of neural networks. It means a batch for training networks consists of multiple unique training samples. Therefore, the recall interval of retraining is the same as the number of training samples. However, the strategy on our learning scheduler alternates the replaying method at the view-batch level. We structure a view-batch to have multiple views of a single sample. Thus, the recall interval of a single sample increases with the size of view-batch. The concept of this view-batch model for the replay method is shown in Figure 3. It shows that when learning four samples in sequence, the time-space of re-training is four, whereas when learning $V$ views of a single sample, the time-space is extended by $V$ times.", + "bbox": [ + 511, + 388, + 906, + 643 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Formally, we define samples as $I$ , sample-batch as $\\mathcal{B}^I = \\{I_i\\}_{i=1}^B$ , the number of batches in a epoch as $T$ , and learning scheduler as $\\mathcal{A}$ . In the conventional learning algorithm, we replay each sample-batch to establish the scheduler.", + "bbox": [ + 511, + 645, + 906, + 705 + ], + "page_idx": 2 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {A} _ {\\text {c o n v e n t i o n a l}} = \\left[ \\underbrace {\\mathcal {B} _ {1} ^ {I} , \\cdots , \\mathcal {B} _ {T} ^ {I}} _ {\\text {r e c a l l i n t e r v a l} = B \\times T}, \\mathcal {B} _ {1} ^ {I}, \\dots , \\mathcal {B} _ {T} ^ {I} \\right]. \\tag {1}\n$$\n", + "text_format": "latex", + "bbox": [ + 558, + 717, + 906, + 752 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Therefore, the recall interval of retraining is the same as the number of training samples $B \\times T$ in an epoch. However, to extend the recall interval, we replace the replaying unit from the sample-batch to the view-batch $\\mathcal{B}^{\\nu} = \\{\\mathcal{V}_i\\}_{i=1}^{B}$ . We construct a view-batch by including multiple views of the same sample $\\mathcal{V}_i = \\{I_i\\}_{i=1}^V$ , where $V$ denotes the number of views consisting of different augmentations of a single image. By modifying the replaying unit of Equation (1) with a constructed view-batch, we define our new learning", + "bbox": [ + 511, + 763, + 908, + 902 + ], + "page_idx": 2 + }, + { + "type": "page_number", + "text": "10079", + "bbox": [ + 480, + 944, + 517, + 955 + ], + "page_idx": 2 + }, + { + "type": "code", + "sub_type": "algorithm", + "code_caption": [ + "Algorithm 1 Pseudocode for our view-batch model." + ], + "code_body": "Require: dataset $\\mathcal{D}^{\\mathrm{train}}$ parameters $\\theta$ , learning rate $\\gamma$ , batch size $B$ , view-batch size $V$ \n1: Initialize network parameters $\\theta$ \n2: Update $B\\gets B / V$ \n3: for number of training iterations do \n4: $\\begin{array}{rl} & {\\mathcal{B}^{I} = \\{I_{i}\\}_{i = 1}^{B}\\mathrm{with}I_{i}\\stackrel {\\mathrm{i.i.d.}}{\\sim}D^{\\mathrm{train}}}\\\\ & {\\mathrm{for}i = 1,\\dots ,B\\mathrm{do}}\\\\ & {\\left\\{ \\begin{array}{l}\\mathcal{V}_{i}\\gets \\{I_{i}\\}_{j = 1}^{V}\\\\ \\mathcal{V}_{i,1}\\gets A_{w}(\\mathcal{V}_{i,1})\\\\ \\mathcal{V}_{i,2}\\gets A_{s}(\\mathcal{V}_{i,2:}) \\end{array} \\right.}\\\\ & {\\bar{\\mathcal{B}}^{\\nu}\\gets \\{\\mathcal{V}_{i}\\}_{i = 1}^{B}}\\\\ & {\\mathrm{Compute}l_{\\mathrm{sup}}\\gets L_{\\mathrm{sup}}(\\theta ,\\mathcal{B}^{\\nu})}\\\\ & {\\mathrm{Compute}l_{\\mathrm{ssl}}\\gets L_{\\mathrm{ssl}}(\\theta ,\\mathcal{B}^{\\nu})\\triangleright SSL\\mathrm{of}\\mathrm{Equation}(3)}\\\\ & {\\mathrm{Update}\\theta \\gets \\theta -\\gamma \\nabla_{\\theta}(l_{\\mathrm{sup}} + l_{\\mathrm{ssl}})} \\end{array}$ \n7: replay of Equation (2) \n8: \n9: \n10: \n11: \n12:", + "bbox": [ + 91, + 111, + 483, + 333 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "schedule as follows:", + "bbox": [ + 89, + 358, + 233, + 372 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {A} _ {\\text {o u r s}} = \\left[ \\underbrace {\\mathcal {B} _ {1} ^ {\\nu} , \\cdots , \\mathcal {B} _ {T} ^ {\\nu}} _ {\\text {r e c a l l i n t e r v a l} = B \\times T \\times V}, \\mathcal {B} _ {1} ^ {\\nu}, \\dots , \\mathcal {B} _ {T} ^ {\\nu} \\right], \\tag {2}\n$$\n", + "text_format": "latex", + "bbox": [ + 143, + 383, + 482, + 419 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Based on this rescheduling, the recall interval is increased by $V$ times. While we process $V$ views at a time, we decrease the total learning epoch by $V$ , ensuring that the number of total samples that networks have seen during training is the same as the baseline method. As a result, networks repeatedly forget and retrain samples during extended recall intervals, thereby improving long-term memory retention. This learning of multiple views of a single sample adjusts the time-space while networks learn the same sample repeatedly. To promote extensive learning, we present the self-supervised learning in the following.", + "bbox": [ + 89, + 429, + 483, + 597 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "3.2. Self-supervised Learning", + "text_level": 1, + "bbox": [ + 89, + 604, + 318, + 622 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "We present a self-supervised learning model with our VBM to extensively learn a sample at a time. It is well-known that self-supervision generalizes the networks rather than acquiring task-specific knowledge [5]. Further, task-agnostic knowledge from self-supervision is more robust to catastrophic forgetting of continual learning [8]. For these reasons, we propose the SSL approach to effectively learn view-batch in a (i) simple and an (ii) efficient manner. For (i) simplicity, unlike previous SSL methods [5, 11, 20], which require changes, such as architecture design [8] or teacher networks [12], we only modify an objective function to learn common knowledge within a view-batch of same samples. To satisfy (ii) efficiency, our method does not use an extra training phase or longer training epochs than supervised learning, ensuring the minimal training cost as analyzed in the supplementary material. Thus, the proposed method could be used in most continual learning methods without increasing computing costs as a drop-in replace", + "bbox": [ + 89, + 628, + 483, + 901 + ], + "page_idx": 3 + }, + { + "type": "image", + "img_path": "images/bb59e13fbd75342f101e6a10e184be0c6c0459ee20c8c795814ba8549e9d9d57.jpg", + "image_caption": [ + "(a) Degree of forgetting" + ], + "image_footnote": [], + "bbox": [ + 516, + 89, + 643, + 170 + ], + "page_idx": 3 + }, + { + "type": "image", + "img_path": "images/522012dc639823957c1068037f3f5cdee1e721d183f991328ff70134b26119bc.jpg", + "image_caption": [ + "(b) Forgetting curve", + "Figure 4. Empirical findings related to forgetting curve theory. (a) We report the degree of forgetting for different recall intervals with a $95\\%$ confidence interval denoted as shaded region. The degree of forgetting is measured as the performance degradation of each sample's classification accuracy between recall intervals. (b) We show the decay of memory retention as the learning progresses. This graph shows that when the formal learning stage is finished, memory retention decays over time for all three cases. (c) We compare the network's classification accuracy of continual learning. It shows that x3 achieves the best performance thanks to the slow memory retention decay." + ], + "image_footnote": [], + "bbox": [ + 645, + 94, + 771, + 170 + ], + "page_idx": 3 + }, + { + "type": "image", + "img_path": "images/d62d2a25d9a2e39869c99a38b80f9c2627d6e7c567e7ac89d4ffb1c81396312a.jpg", + "image_caption": [ + "(c) Accuracy" + ], + "image_footnote": [], + "bbox": [ + 772, + 94, + 898, + 170 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "ment approach.", + "bbox": [ + 511, + 377, + 617, + 391 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "To implement our approach, we employ a one-to-many divergence loss to learn self-supervision within a view-batch $\\{\\mathcal{V}_i\\}_{i=1}^B$ . Our one-to-many divergence loss minimizes the average divergence between one weak-augmented view $A_w(\\mathcal{V}_{i,1})$ and the remaining strong-augmented views $A_s(\\mathcal{V}_{i,2:})$ of a view-batch at the logit level. By aggregating different augmented views closer, networks learn common characteristics of objects. Specifically, leveraging KL divergence $D_{\\mathrm{KL}}$ between different augmentations, we define one-to-many-based self-supervised loss as", + "bbox": [ + 511, + 392, + 906, + 544 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\nL _ {\\mathrm {s s l}} \\left(f _ {\\theta}, \\mathcal {B} ^ {\\nu}\\right) = \\frac {1}{B \\cdot (V - 1)} \\sum_ {i = 1} ^ {B} \\sum_ {j = 2} ^ {V} D _ {\\mathrm {K L}} \\left(p _ {i} ^ {1} | | p _ {i} ^ {j}\\right), \\tag {3}\n$$\n", + "text_format": "latex", + "bbox": [ + 532, + 553, + 905, + 597 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "where $p_i^1$ and $p_i^j$ represent network's prediction for weak augmented samples $\\sigma(f_{\\theta}(A_w(\\mathcal{V}_{i,1})))$ and strong augmented samples $\\sigma(f_{\\theta}(A_s(\\mathcal{V}_{i,j})))$ , $f_{\\theta}$ is neural networks, and $\\sigma$ stand for the softmax. Finally, as shown in Algorithm 1, we define our final objective function as", + "bbox": [ + 511, + 609, + 905, + 686 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\min _ {f _ {\\theta}} L _ {\\sup } \\left(f _ {\\theta}, \\mathcal {B} ^ {\\nu}\\right) + L _ {\\operatorname {s s l}} \\left(f _ {\\theta}, \\mathcal {B} ^ {\\nu}\\right), \\tag {4}\n$$\n", + "text_format": "latex", + "bbox": [ + 591, + 696, + 906, + 715 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "where $L_{\\sup}(f_{\\theta}, \\mathcal{B}^{\\nu}) = \\frac{1}{B \\cdot V} \\sum_{i=1}^{B} \\sum_{j=1}^{V} \\mathcal{H}(y_{i}, p_{i}^{j})$ of each view's prediction $p_{i}^{j}$ and its label $y_{i}$ using cross-entropy loss $\\mathcal{H}$ . We select widely used auto-augmentation [13] as a strong augmentation method, which slightly affects the performance of the baseline method as shown in Tab. 6, while using horizontal flip as weak augmentation.", + "bbox": [ + 511, + 724, + 906, + 823 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "3.3. Empirical Evidence", + "text_level": 1, + "bbox": [ + 511, + 830, + 702, + 848 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "This subsection empirically confirms our assumption that optimal recall interval improves memory retention, and accordingly, the performance of the continual method is en", + "bbox": [ + 511, + 854, + 905, + 900 + ], + "page_idx": 3 + }, + { + "type": "page_number", + "text": "10080", + "bbox": [ + 480, + 944, + 519, + 955 + ], + "page_idx": 3 + }, + { + "type": "table", + "img_path": "images/0e693dd0f1344b228d6c3a54347bd61233162a949e5a1b8abb2c7cf39ea1a3c6.jpg", + "table_caption": [], + "table_footnote": [], + "table_body": "
BufferMethodS-CIFAR-10S-Tiny-ImageNet
CILTILAvg*ΔCILTILAvg*Δ
-Joint92.20±0.1598.31±0.1295.2659.99±0.1982.04±0.1071.02
0LwF-61.98±0.9861.98-15.26±0.7115.26
VBM-LwF-77.53±3.0677.53+15.55-51.21±0.3451.21+35.95
200DER56.58±2.0889.32±1.3272.9511.56±0.5841.67±0.6326.62
ER50.27±0.7091.66±0.7970.978.67±0.2338.03±0.7923.35
VBM-ER52.56±1.5993.58±0.2373.07+2.108.83±0.1344.03±1.1226.43+3.08
iCaRL64.11±1.8690.20±0.8877.1619.09±0.1353.07±0.2736.08
VBM-iCaRL69.73±0.8792.76±0.4581.25+4.0921.39±0.1856.31±0.4238.85+2.77
DER++61.67±1.0290.61±1.1376.1412.48±1.6041.19±0.4826.84
VBM-DER++66.99±2.9894.30±0.2180.65+4.5113.67±0.4144.36±0.8829.02+2.18
500DER65.93±5.3892.52±0.6879.2318.04±1.3052.90±0.9135.47
ER56.37±0.9694.15±0.0475.2610.20±0.3749.13±0.3129.67
VBM-ER63.39±2.7794.20±0.2578.80+3.5411.21±0.3052.01±0.7131.61+1.94
iCaRL62.22±3.4588.86±3.2375.5423.93±0.9758.89±0.8541.41
VBM-iCaRL68.97±0.3893.10±0.6681.04+5.5026.89±0.5462.61±0.1744.75+3.34
DER++69.94±2.0493.49±0.2681.7219.39±1.0951.71±0.3535.55
VBM-DER++75.22±0.8194.81±0.3985.02+3.3019.46±0.4649.80±0.8034.63-0.92
5120DER80.79±0.7594.46±0.3087.6332.92±3.0964.77±2.1448.85
ER74.23±4.1896.86±0.2885.5528.07±0.3767.58±0.3047.83
VBM-ER79.29±2.2997.02±0.1388.16+2.6133.07±0.3668.91±0.5050.99+3.16
iCaRL78.49±0.3894.84±0.8286.6732.16±0.2466.75±0.3649.46
VBM-iCaRL79.40±0.6396.09±0.3287.75+1.0834.40±0.1268.86±0.2651.63+2.17
DER++85.42±0.7496.24±0.1990.8334.06±0.5765.28±1.2049.67
VBM-DER++86.62±0.3897.07±0.0791.85+1.0235.14±0.3667.52±0.4851.33+1.66
", + "bbox": [ + 163, + 88, + 831, + 446 + ], + "page_idx": 4 + }, + { + "type": "table", + "img_path": "images/e84ec13eb0f47a5d137e994bbeec5c781d842ec90e8c36d8841e044e7e30dd9c.jpg", + "table_caption": [ + "Table 1. Experimental results on CIL and TIL protocols. In this evaluation, we report the last top-1 accuracy $(\\%)$ on the S-CIFAR-10 and S-Tiny-ImageNet benchmarks with four different buffer sizes. ResNet-18 backbone is used for each continual learning method in this evaluation. *Avg is the averaged accuracy of CIL and TIL. We reproduce baseline methods following Buzzega et al. [3]. We run every method three times with different random seeds for a reliable result. Ours improves performance consistently in most results. The performance of Joint is the upper bound of this experiment, where all old training samples are used in every step." + ], + "table_footnote": [], + "table_body": "
MethodAvgΔLastΔ
Joint-46.48±3.62
AGEM24.88±0.5728.94±0.67
AGEM-R24.72±0.5928.01±0.93
FDR34.62±0.3334.46±0.32
DER39.84±0.1540.46±0.47
ER32.38±2.8832.02±2.39
VBM-ER35.80±0.55+3.4234.72±0.37+2.70
DER++34.56±3.4234.75±2.41
VBM-DER++42.16±0.20+7.2441.81±0.08+7.06
", + "bbox": [ + 112, + 550, + 460, + 699 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Table 2. Experimental results on DIL protocol. This evaluation shows Avg and Last top-1 accuracy $(\\%)$ on the DomainNet benchmark. We utilize ResNet-18 as the backbone network in this evaluation.", + "bbox": [ + 89, + 710, + 482, + 765 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "hanced. Specifically, the logical development of our assumption is that 1) if the network sufficiently forgets a sample at the optimal recall interval and then extensively retrans it, 2) memory retention is enhanced as training progresses, and therefore, 3) the accuracy of continual learning is also improved. To evaluate our assumption, we measure 1) the degree of forgetting between a retraining session for", + "bbox": [ + 89, + 794, + 482, + 900 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "training samples, 2) the decay speed of the memory retention, and 3) the average accuracy at which networks correctly classify their class labels. Figure 4 shows the results for the empirical evidence of our assumption. As depicted in Figure 4(a), the degree of forgetting becomes larger as the recall interval increases. It is obvious that the longer recall interval allows the network to forget a training sample before retraining it again. Figure 4(b) demonstrates the different levels of memory retention decay due to the recall intervals. The experiment results show that memory decays slowest in the x3 case, which has an interval three times longer than the baseline x1 case. Notably, our method achieves considerably higher performance at the end of the formal learning stage. This result is attributed to the fact that multiple epochs are used in training networks, and the memory retention decay between epochs is slow for the x3 case. Accordingly, x3 with our replay method achieves improved accuracy compared to other recall interval cases, as shown in Figure 4(c). We conduct more experiments to show empirical proofs on different datasets, backbone networks, and buffer sizes in Figure 6. The definition of forgetting is given in the supplementary material.", + "bbox": [ + 511, + 554, + 906, + 887 + ], + "page_idx": 4 + }, + { + "type": "page_number", + "text": "10081", + "bbox": [ + 480, + 945, + 517, + 955 + ], + "page_idx": 4 + }, + { + "type": "table", + "img_path": "images/2493e963540e7d19407c871e890f0242242862d4bce9f0fd378d3b79bb2c8993.jpg", + "table_caption": [], + "table_footnote": [], + "table_body": "
Method5 Step10 Step20 Step
AvgΔLastΔAvgΔLastΔAvgΔLastΔ
Joint80.40-80.41-81.49-
iCaRL71.14±0.3459.7165.27±1.0250.7461.20±0.8343.75
UCIR62.77±0.8247.3158.66±0.7143.3958.17±0.3040.63
BiC73.10±0.5562.1068.80±1.2053.5466.48±0.3247.02
WA72.81±0.2860.8469.46±0.2953.7867.33±0.1547.31
DyTox--73.66±0.0260.6772.27±0.1856.32
DER76.77±0.0168.06±0.0075.72±0.0864.32±0.0574.96±0.0161.80±0.07
VBM-DER78.60±0.23+1.8370.60±0.12+2.5478.12±0.07+2.4067.04±0.11+2.7276.95±0.02+1.9964.29±0.04+2.49
TCIL77.33±0.0869.48±0.1476.33±0.1565.66±0.0274.32±0.0162.54±0.02
VBM-TCIL79.23±0.01+1.9071.23±0.04+1.7578.02±0.03+1.6968.14±0.08+2.4876.83±0.00+2.5167.16±0.09+4.62
", + "bbox": [ + 133, + 88, + 867, + 268 + ], + "page_idx": 5 + }, + { + "type": "table", + "img_path": "images/6e6a9d418060d29aa4ab75ef02336598a9d7a63e3f0f0fae8b41914aa52c3c27.jpg", + "table_caption": [ + "Table 3. Experimental results on rehearsal-based CIL protocol. In this evaluation, we provide Avg and Last top-1 accuracy $(\\%)$ on the S-CIFAR-100 benchmark with three different class incremental steps. ResNet-18 backbone is utilized to evaluate each rehearsal-based method. This evaluation also shows that ours achieves consistent performance improvements. Results of DyTox are imported from its original work [17]. We reproduce results of DER and TCIL using official implementation, while other results come from Yan et al. [44]." + ], + "table_footnote": [], + "table_body": "
Method5 Step10 Step
AvgΔLastΔAvgΔLastΔ
Joint80.40-80.41-
LwF-MC32.8911.4220.716.36
BiC54.7435.4245.0824.42
LwM39.3116.7025.478.57
WA63.6749.3645.6826.14
DER40.05±0.0017.11±0.0026.67±0.008.71±0.00
VBM-DER40.47±0.00+0.4217.47±0.01+0.3626.95±0.00+0.288.90±0.00+0.19
TCIL64.40±0.1552.37±0.0856.84±0.7540.31±1.04
VBM-TCIL66.40±0.07+2.0054.69±0.02+2.3261.12±0.02+4.2845.92±0.25+5.61
", + "bbox": [ + 91, + 357, + 483, + 503 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Table 4. Experimental results on non-rehearsal-based CIL protocol. For this evaluation, we measure Avg and Last top-1 accuracy $(\\%)$ on the S-CIFAR-100 benchmark with two different class incremental steps. ResNet-18 backbone is applied to existing and our methods. We report original results from LwF-MC and LwM [15]. We reproduce the results of DER and TCIL by ourselves using official implementation. Other results are reported from Huang et al. [24].", + "bbox": [ + 89, + 513, + 483, + 625 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "4. Experiments", + "text_level": 1, + "bbox": [ + 89, + 654, + 223, + 671 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "This section demonstrates the capabilities of the proposed view-batch model in various continual learning protocols and scenarios through extensive performance evaluation. We perform experiments for three continual learning protocols, such as class, task, and domain incremental learning. Further, we also assess our view-batch model in both rehearsal and non-rehearsal, as well as under the pre-training-based continual learning scenarios.", + "bbox": [ + 89, + 680, + 482, + 801 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "4.1. Experimental Settings", + "text_level": 1, + "bbox": [ + 89, + 815, + 297, + 832 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Datasets. We evaluate our methods on widely used benchmarks such as S-CIFAR-10/100, S-Tiny-ImageNet, S-ImageNet-R, and DomainNet where the prefix of 'S-' denotes the sequential data configuration. We use the com", + "bbox": [ + 89, + 839, + 483, + 902 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "monly adopted class ordering lists, following existing works of Yan et al. [44] and Huang et al. [24].", + "bbox": [ + 511, + 359, + 906, + 391 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Protocols. In our experiments, we utilize three continual learning protocols. First, we make use of task incremental learning (TIL), which involves training neural networks on a series of tasks in a sequential manner while evaluating their performance on both newly acquired and previously learned tasks. Specifically, the TIL protocol consists of the following detailed configurations: 1) a task is defined by a set of label classes, and 2) we evaluate networks knowing the task identity of test samples. Second, we adopt class incremental learning (CIL), which is identical to TIL except that it evaluates test samples without knowing their task identity. Third, we use domain incremental learning (DIL), which focuses on learning the new samples across multiple sequential domains rather than acquiring new classes. In addition to these three protocols, we also include both rehearsal and non-rehearsal scenarios, divided by the use of a memory buffer during continual learning. For the non-rehearsal scenario, we evaluate both scratch- and pre-training-based learning scenarios to showcase extensive experimental results. For all these protocols and scenarios, we provide the mean and standard deviation of three runs for reliable experimental results. The details of the experiment can be found in the supplementary material.", + "bbox": [ + 511, + 411, + 908, + 760 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Baseline Methods. We apply our view-batch to recent baseline continual learning methods such as iCaRL [32], ER [34], DER [3, 44], TCIL [24], and SLCA [46]. We also include results from A-GEM [10], LwM [15], UCIR [23], BiC [43], WA [47], DyTox [17], L2P [42], S-iPrompt [40], DualPrompt [41], and CODA-Prompt [38]. We follow the baseline's hyper-parameters for a fair comparison between ours and its respective baseline methods. Specifically, we", + "bbox": [ + 511, + 779, + 910, + 902 + ], + "page_idx": 5 + }, + { + "type": "page_number", + "text": "10082", + "bbox": [ + 480, + 944, + 519, + 955 + ], + "page_idx": 5 + }, + { + "type": "table", + "img_path": "images/2ebce71d16fdffdc13f1b260e5b450971a20eb35dcd975c47ed23d27db50d15c.jpg", + "table_caption": [], + "table_footnote": [], + "table_body": "
S-CIFAR-100S-ImageNet-R
MethodTop-1Top-5Top-1Top-5
AvgLastAvgLastAvgLastAvgLast
Joint-92.54±0.14-99.24±0.02-82.41±0.02-93.27±0.07
L2P90.42±0.5683.76±0.4298.89±0.0398.08±0.1576.76±0.3373.56±0.4289.39±0.1686.61±0.03
DualPrompt90.17±0.3885.10±0.2398.90±0.0797.68±0.1273.31±0.1969.12±0.1889.19±0.0985.46±0.04
CODA-Prompt91.27±0.6686.86±0.5899.29±0.0998.64±0.0781.89±0.5976.79±0.1793.30±0.3490.49±0.15
S-iPrompt92.19±0.4088.43±0.4299.19±0.0598.63±0.0571.47±0.1468.77±0.2788.44±0.1984.91±0.31
SLCA94.31±1.1291.57±0.4699.64±0.0499.26±0.0383.58±0.6578.81±0.2194.39±0.2692.02±0.25
VBM-SLCA94.52±0.8591.78±0.1899.64±0.0399.29±0.0384.61±0.5980.07±0.1294.86±0.1992.74±0.14
", + "bbox": [ + 145, + 88, + 854, + 247 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/eee9c2892997f41c6cf1f1f1042a54be340f19d9ebebe328fe2b936eb21a7201.jpg", + "image_caption": [ + "(a) iCaRL" + ], + "image_footnote": [], + "bbox": [ + 101, + 323, + 282, + 431 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/605ea598910eb3fa5beb3139decaece279767ebcefe705c6cbe8b007418c956c.jpg", + "image_caption": [ + "(b) DER++" + ], + "image_footnote": [], + "bbox": [ + 287, + 323, + 467, + 431 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/773d2134c08e3b1f5e4757b2f0bf45e7a6e1b18145a7f44973dfa65198898156.jpg", + "image_caption": [ + "Figure 5. Experimental results on memory retention decay. This analysis reports memory retention decay of the first three tasks on the S-CIFAR-10 dataset, comparing the proposed approach against the baseline methods. The green numbers at the end of the last task are the accuracy gain of the proposed approach over the baselines.", + "(a) Dataset", + "Figure 6. Experimental results on recall interval. We measure the last top-1 accuracy $(\\%)$ ( $\\uparrow$ ) in the top row and report forgetting $(\\downarrow)$ at the bottom row, varying recall intervals with different learning factors. Dark colors mean better. We use the ResNet-18 backbone." + ], + "image_footnote": [], + "bbox": [ + 99, + 556, + 220, + 669 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/b17898ccb02afcbc02cac32f8a68825dde50800e921a7dc08779ecef470d4109.jpg", + "image_caption": [ + "(b) Network" + ], + "image_footnote": [], + "bbox": [ + 225, + 558, + 344, + 669 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/93aec7734c01f8373a22f703483d8be15995b2d7280e898d4cfc2b59dc4430a8.jpg", + "image_caption": [ + "(c) Buffer" + ], + "image_footnote": [], + "bbox": [ + 349, + 558, + 470, + 669 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "use ResNet and Vision Transformer (ViT) as backbone networks in traditional and prompt-based approaches. For the traditional approach, we randomly initialize backbone weights at the beginning of every run. On the other hand, for the prompt-based approach, we adopt a frozen backbone whose weights are pre-trained on a large-scale dataset following recent prompt-based CL studies [42, 46]", + "bbox": [ + 89, + 794, + 482, + 900 + ], + "page_idx": 6 + }, + { + "type": "table", + "img_path": "images/a05fdef5078753035f3ffa4f7403b0e1bb9848ee7df022a4ab2366ed336da763.jpg", + "table_caption": [ + "Table 5. Experimental results on pre-trained model-based CIL protocol. This evaluation presents Avg and Last top-1,5 accuracy (%) on the S-CIFAR-100 and S-ImageNet-R benchmarks without using a memory buffer. ViT-B/16 backbone is adopted for each pre-trained model-based method. We reproduce baseline methods using official implementation by ourselves, except for L2P." + ], + "table_footnote": [], + "table_body": "
MethodReplayStrong augmentSSLCILTILAvgΔ
iCaRL---64.1190.2077.16-1.05
-✓-62.0490.1776.11
✓--67.5191.8779.69+2.53
✓✓-66.7891.2178.99+1.83
✓✓✓69.7392.7681.25+4.09
DER++---61.6790.6176.14+0.76
-✓-62.6691.1476.90
✓--64.7592.7978.77+2.63
✓✓-64.4493.3878.91+2.77
✓✓✓66.9994.3080.65+4.51
", + "bbox": [ + 519, + 323, + 901, + 482 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Table 6. Experimental results on factor analysis. We demonstrate the last accuracy (\\%) of CIL and TIL on S-CIFAR-10, varying our main components with different augmentation types. We observe that the proposed components, such as replay and SSL, improve performance consistently and confirm that performance enhancement comes from them, not simply from strong augmentation. In the strong augmentation setting without view batch replay, we maintain the same ratio of weak to strong augmentations.", + "bbox": [ + 511, + 492, + 906, + 604 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "4.2. Experimental Results", + "text_level": 1, + "bbox": [ + 511, + 632, + 715, + 648 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Continual Learning under Three Protocols. We evaluate our method in the three continual learning protocols such as CIL, TIL, and DIL. For TIL and CIL, we use different memory buffer sizes: 0, 200, 500, and 5,120. For DIL, we utilize 50 samples per class, resulting in 17,250 samples in a buffer. We apply our view-batch model to LwF for the non-rehearsal memory buffer (i.e. 0 size) and ER, DER++, and iCaRL for the other buffer sizes. Tabs. 1 and 2 show that our method improves the performance of continual learning methods for all protocols. Moreover, we demonstrate that baseline methods employing our view-batch are comparable to other continual learning methods.", + "bbox": [ + 511, + 657, + 906, + 838 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "We conduct further in-depth experiments under the most widely used CIL protocol. We consider three configurations with step sizes of 5, 10, and 20, which are the number of total incremental steps, as shown in Tab. 3. In addition, we", + "bbox": [ + 511, + 839, + 906, + 900 + ], + "page_idx": 6 + }, + { + "type": "page_number", + "text": "10083", + "bbox": [ + 480, + 944, + 517, + 955 + ], + "page_idx": 6 + }, + { + "type": "table", + "img_path": "images/dd1f0e551243b84a738289a7de94cfa6345865906f28ad84da8fadd0dea8a94a.jpg", + "table_caption": [], + "table_footnote": [], + "table_body": "
MethodCurrent sampleBuffer sampleCILTILAvgΔ
iCaRL--64.1190.2077.16-
✓-64.6494.1279.38+2.22
-✓64.9994.2079.60+2.44
✓✓69.7392.7681.25+4.09
DER++--61.6790.6176.14-
✓-64.8893.2279.05+2.91
-✓62.6994.1578.42+2.28
✓✓66.9994.3080.65+4.51
", + "bbox": [ + 99, + 88, + 475, + 224 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Table 7. Experimental results on sample type. We measure the last accuracy (\\%) of CIL and TIL on S-CIFAR-10, applying our method to different sample types. We use the ResNet-18 as backbone. This analysis shows that our method works well with both sample types.", + "bbox": [ + 89, + 234, + 482, + 306 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "take into account the use of a memory buffer in two scenarios, such as rehearsal and non-rehearsal in Tab. 4. Both experimental results demonstrate that our view-batch model consistently improves the performance of continual learning methods under the various step sizes and memory buffer configurations.", + "bbox": [ + 89, + 330, + 483, + 421 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Non-rehearsal Continual Learning. While the nonrehearsal scenario has the advantage of not using memory buffers, they inherently face the drawback of limited performance due to the inability to retrain networks using old data in the memory buffer. To overcome this limitation, many recent studies have focused on non-rehearsal continual learning using a pre-trained model trained with a large-scale dataset. Thus, we evaluate our view-batch model in a scenario where models are pre-trained, such as SCLA. Experimental results in Tab. 5 indicate that the proposed view-batch model is effective across all benchmarks for this scenario.", + "bbox": [ + 89, + 439, + 483, + 619 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "5. Analysis", + "text_level": 1, + "bbox": [ + 89, + 633, + 187, + 650 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "We present the analysis of the proposed view-batch model. First, we validate that the optimal recall interval mitigates catastrophic forgetting of early tasks. Figure 5 visualize that the recall interval x3 maintains long-term memory retention resolving catastrophic forgetting of initial tasks. Also, Figure 6 analyzes how different factors (i.e., dataset, network scales, memory buffer size) affect the optimal recall intervals in terms of last top-1 accuracy and forgetting metrics [9, 10, 42]. We find that the recall interval x3 or x4 improves the accuracy and resolves catastrophic forgetting in various scenarios. More experimental results of different methods and task types can be found in the supplementary material. Tab. 6 shows that two main factors of the proposed method consistently improve their respective baselines, identifying that the performance improvement comes from the proposed components, not strong augmentation.", + "bbox": [ + 89, + 657, + 482, + 900 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/9feb0905c07c53017f273a2c620e2845a4f8ffe3bd19ed9627d481e09059eeaa.jpg", + "image_caption": [ + "Figure 7. Experimental results on accuracy and forgetting. This analysis shows (a) the last top-1 accuracy $(\\%)$ and (b) corresponding forgetting measures on the S-CIFAR-10 dataset with different buffer sizes. Our method consistently improves accuracy and forgetting measures." + ], + "image_footnote": [], + "bbox": [ + 526, + 88, + 890, + 210 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Lastly, in the rehearsal scenario, we have two different sample types, current- and buffer-based samples. Tab. 7 exhibit that ours performs well with both types.", + "bbox": [ + 511, + 316, + 903, + 362 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "We examine the proposed method in terms of the last accuracy and forgetting measurement. We apply the proposed method to two different baseline approaches. We compute average forgetting at $T$ th task as $\\frac{1}{T - 1}\\sum_{t = 1}^{T - 1}\\max_{j\\in \\{t,\\dots ,T - 1\\}}a_j^t -a_T^t$ where $a_j^t$ means $t$ th task accuracy at $j$ step. Notably, higher forgetting means severe catastrophic forgetting of previous tasks. Figure 7 shows that the view-batch model improves accuracy while addressing the forgetting problem simultaneously.", + "bbox": [ + 511, + 363, + 906, + 500 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "6. Conclusion", + "text_level": 1, + "bbox": [ + 513, + 512, + 633, + 527 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "We propose the view-batch model, which is generally applicable to various continual learning scenarios. Inspired by Ebbinghaus's forgetting curve theory, the proposed view-batch model optimizes the recall interval between retraining samples, improving neural networks' long-term memory. The proposed method consists of two main components: replay and self-supervised learning. The replay optimizes the recall interval for guaranteeing sufficient memory forgetting by grouping multiple views of a sample. Also, the self-supervised learning ensures extensive learning for the multiple views using the one-to-many divergence loss. Experimental results show that sufficient forgetting preserves long-term memory and ultimately improves continual learning accuracy. We hope the proposed approach to continual learning can boost future research.", + "bbox": [ + 511, + 537, + 906, + 765 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Acknowledgment. This paper was supported in part by the ETRI Grant funded by Korean Government (Fundamental Technology Research for Human-Centric Autonomous Intelligent Systems) under Grant 24ZB1200, Artificial Intelligence Convergence Innovation Human Resources Development (IITP-2025-RS-2023-00255968), Artificial Intelligence Innovation Hub (RS-2021-II212068), and the NRF Grant (RS-2024-00356486).", + "bbox": [ + 511, + 780, + 908, + 900 + ], + "page_idx": 7 + }, + { + "type": "page_number", + "text": "10084", + "bbox": [ + 480, + 944, + 519, + 955 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "References", + "text_level": 1, + "bbox": [ + 91, + 89, + 187, + 104 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[1] Hadi Amiri, Timothy Miller, and Guergana Savova. Repeat before forgetting: Spaced repetition for efficient and effective training of neural networks. In Conference on Empirical Methods in Natural Language Processing, 2017. 1, 2", + "[2] Maxim Berman, Hervé Jégou, Andrea Vedaldi, Iasonas Kokkinos, and Matthijs Douze. Multigrain: a unified image embedding for classes and instances. arXiv:1902.05509, 2019.3", + "[3] Pietro Buzzega, Matteo Boschini, Angelo Porrello, Davide Abati, and Simone Calderara. Dark experience for general continual learning: a strong, simple baseline. Advances in Neural Information Processing Systems, 2020. 5, 6", + "[4] Boxi Cao, Qiaoyu Tang, Hongyu Lin, Shanshan Jiang, Bin Dong, Xianpei Han, Jiawei Chen, Tianshu Wang, and Le Sun. Retentive or forgetful? diving into the knowledge memorizing mechanism of language models. In Joint International Conference on Computational Linguistics, Language Resources and Evaluation, 2024. 1, 2", + "[5] Mathilde Caron, Hugo Touvron, Ishan Misra, Herve Jégou, Julien Mairal, Piotr Bojanowski, and Armand Joulin. Emerging properties in self-supervised vision transformers. In International Conference on Computer Vision, 2021. 3, 4", + "[6] Nicholas J Cepeda, Harold Pashler, Edward Vul, John T Wixted, and Doug Rohrer. Distributed practice in verbal recall tasks: A review and quantitative synthesis. *Psychological bulletin*, 2006. 1", + "[7] Nicholas J Cepeda, Edward Vul, Doug Rohrer, John T Wixted, and Harold Pashler. Spacing effects in learning: A temporal ridgeline of optimal retention. Psychological science, 2008. 1", + "[8] Hyuntak Cha, Jaeho Lee, and Jinwoo Shin. Co2l: Contrastive continual learning. In International Conference on Computer Vision, 2021. 4", + "[9] Arslan Chaudhry, Puneet K Dokania, Thalaiyasingam Ajthan, and Philip HS Torr. Riemannian walk for incremental learning: Understanding forgetting and intransigence. In European Conference on Computer Vision, 2018. 8", + "[10] Arslan Chaudhry, Marc'Aurelio Ranzato, Marcus Rohrbach, and Mohamed Elhoseiny. Efficient lifelong learning with a gem. In International Conference on Learning Representations, 2019. 6, 8", + "[11] Ting Chen, Simon Kornblith, Mohammad Norouzi, and Geoffrey Hinton. A simple framework for contrastive learning of visual representations. In International Conference on Machine Learning, 2020. 3, 4", + "[12] Haoyang Cheng, Haitao Wen, Heqian Qiu, Lanxiao Wang, Minjian Zhang, and Hongliang Li. Must unsupervised continual learning relies on previous information? In Conference on Computer Vision and Pattern Recognition, 2024. 4", + "[13] Ekin D Cubuk, Barret Zoph, Dandelion Mane, Vijay Vasudevan, and Quoc V Le. Autoaugment: Learning augmentation policies from data. arXiv:1805.09501, 2018. 4", + "[14] Ekin D Cubuk, Barret Zoph, Jonathon Shlens, and Quoc V Le. Randaugment: Practical automated data augmentation with a reduced search space. In Conference on Computer Vision and Pattern Recognition Workshops, 2020. 2" + ], + "bbox": [ + 93, + 114, + 483, + 900 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[15] Prithviraj Dhar, Rajat Vikram Singh, Kuan-Chuan Peng, Ziyan Wu, and Rama Chellappa. Learning without memorizing. In Conference on Computer Vision and Pattern Recognition, 2019. 3, 6", + "[16] Arthur Douillard, Matthieu Cord, Charles Ollion, Thomas Robert, and Eduardo Valle. Podnet: Pooled outputs distillation for small-tasks incremental learning. In European Conference on Computer Vision, 2020. 3", + "[17] Arthur Douillard, Alexandre Ramé, Guillaume Couairon, and Matthieu Cord. Dytox: Transformers for continual learning with dynamic token expansion. In Conference on Computer Vision and Pattern Recognition, 2022. 6", + "[18] Hermann Ebbinghaus. Memory: A contribution to experimental psychology. Annals of neurosciences, 2013. 1, 2", + "[19] Kate Gordon. Class results with spaced and unspaced memorizing. Journal of Experimental Psychology, 1925. 1", + "[20] Jean-Bastien Grill, Florian Strub, Florent Altché, Corentin Tallec, Pierre Richemond, Elena Buchatskaya, Carl Doersch, Bernardo Avila Pires, Zhaohan Guo, Mohammad Gheshlaghi Azar, et al. Bootstrap your own latent-a new approach to self-supervised learning. Advances in Neural Information Processing Systems, 2020. 4", + "[21] Karri S Hawley, Katie E Cherry, Emily O Boudreaux, and Erin M Jackson. A comparison of adjusted spaced retrieval versus a uniform expanded retrieval schedule for learning a name-face association in older adults with probable Alzheimer's disease. Journal of Clinical and Experimental Neuropsychology, 2008. 1", + "[22] Elad Hoffer, Tal Ben-Nun, Itay Hubara, Niv Giladi, Torsten Hoefler, and Daniel Soudry. Augment your batch: better training with larger batches. arXiv:1901.09335, 2019. 3", + "[23] Saihui Hou, Xinyu Pan, Chen Change Loy, Zilei Wang, and Dahua Lin. Learning a unified classifier incrementally via rebalancing. In Conference on Computer Vision and Pattern Recognition, 2019. 6", + "[24] Bingchen Huang, Zhineng Chen, Peng Zhou, Jiayin Chen, and Zuxuan Wu. Resolving task confusion in dynamic expansion architectures for class incremental learning. In Association for the Advancement of Artificial Intelligence, 2023. 2, 3, 6", + "[25] Anette Hunziker, Yuxin Chen, Oisin Mac Aodha, Manuel Gomez Rodriguez, Andreas Krause, Pietro Perona, Yisong Yue, and Adish Singla. Teaching multiple concepts to a forgetful learner. Advances in Neural Information Processing Systems, 2019. 2", + "[26] Szabolcs Káli and Peter Dayan. Replay, repair and consolidation. Advances in Neural Information Processing Systems, 2002. 2", + "[27] Sean HK Kang. Spaced repetition promotes efficient and effective learning: Policy implications for instruction. *Policy Insights from the Behavioral and Brain Sciences*, 2016. 1", + "[28] Jeffrey D Karpicke and Henry L Roediger III. Expanding retrieval practice promotes short-term retention, but equally spaced retrieval enhances long-term retention. Journal of experimental psychology: learning, memory, and cognition, 2007. 1" + ], + "bbox": [ + 516, + 92, + 906, + 898 + ], + "page_idx": 8 + }, + { + "type": "page_number", + "text": "10085", + "bbox": [ + 480, + 944, + 517, + 955 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[29] Zhizhong Li and Derek Hoiem. Learning without forgetting. IEEE transactions on pattern analysis and machine intelligence, 2017. 3", + "[30] Cecil Alec Mace. The psychology of study. McBride, 1932. 1", + "[31] Arthur W Melton. The situation with respect to the spacing of repetitions and memory. Journal of Verbal Learning and Verbal Behavior, 1970. 1", + "[32] Sylvestre-Alvise Rebuffi, Alexander Kolesnikov, Georg Sperl, and Christoph H Lampert. icarl: Incremental classifier and representation learning. In Conference on Computer Vision and Pattern Recognition, 2017. 2, 3, 6", + "[33] Matthew Riemer, Ignacio Cases, Robert Ajemian, Miao Liu, Irina Rish, Yuhai Tu, , and Gerald Tesauro. Learning to learn without forgetting by maximizing transfer and minimizing interference. In International Conference on Learning Representations, 2019. 3", + "[34] David Rolnick, Arun Ahuja, Jonathan Schwarz, Timothy Lillicrap, and Gregory Wayne. Experience replay for continual learning. Advances in Neural Information Processing Systems, 2019. 6", + "[35] Joan Serra, Didac Suris, Marius Miron, and Alexandros Karatzoglou. Overcoming catastrophic forgetting with hard attention to the task. In International Conference on Machine Learning, 2018. 3", + "[36] John J Shaughnessy. Long-term retention and the spacing effect in free-recall and frequency judgments. The American Journal of Psychology, 1977. 1", + "[37] Hanul Shin, Jung Kwon Lee, Jaehong Kim, and Jiwon Kim. Continual learning with deep generative replay. Advances in Neural Information Processing Systems, 2017. 3", + "[38] James Seale Smith, Leonid Karlinsky, Vyshnavi Gutta, Paola Cascante-Bonilla, Donghyun Kim, Assaf Arbelle, Rameswar Panda, Rogerio Feris, and Zsolt Kira. Coda-prompt: Continual decomposed attention-based prompting for rehearsal-free continual learning. In Conference on Computer Vision and Pattern Recognition, 2023. 6", + "[39] Kushal Tirumala, Aram Markosyan, Luke Zettlemoyer, and Armen Aghajanyan. Memorization without overfitting: Analyzing the training dynamics of large language models. Advances in Neural Information Processing Systems, 2022. 2", + "[40] Yabin Wang, Zhiwu Huang, and Xiaopeng Hong. S-prompts learning with pre-trained transformers: An occam's razor for domain incremental learning. Advances in Neural Information Processing Systems, 2022. 6", + "[41] Zifeng Wang, Zizhao Zhang, Sayna Ebrahimi, Ruoxi Sun, Han Zhang, Chen-Yu Lee, Xiaoqi Ren, Guolong Su, Vincent Perot, Jennifer Dy, et al. Dualprompt: Complementary prompting for rehearsal-free continual learning. In European Conference on Computer Vision, 2022. 3, 6", + "[42] Zifeng Wang, Zizhao Zhang, Chen-Yu Lee, Han Zhang, Ruoxi Sun, Xiaqi Ren, Guolong Su, Vincent Perot, Jennifer Dy, and Tomas Pfister. Learning to prompt for continual learning. In Conference on Computer Vision and Pattern Recognition, 2022. 3, 6, 7, 8", + "[43] Yue Wu, Yinpeng Chen, Lijuan Wang, Yuancheng Ye, Zicheng Liu, Yandong Guo, and Yun Fu. Large scale in" + ], + "bbox": [ + 89, + 90, + 482, + 900 + ], + "page_idx": 9 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "cremental learning. In Conference on Computer Vision and Pattern Recognition, 2019. 6", + "[44] Shipeng Yan, Jiangwei Xie, and Xuming He. Der: Dynamically expandable representation for class incremental learning. In Conference on Computer Vision and Pattern Recognition, 2021. 3, 6", + "[45] Sangdoo Yun, Dongyoon Han, Seong Joon Oh, Sanghyuk Chun, Junsuk Choe, and Youngjoon Yoo. Cutmix: Regularization strategy to train strong classifiers with localizable features. In International Conference on Computer Vision, 2019. 2", + "[46] Gengwei Zhang, Liyuan Wang, Guoliang Kang, Ling Chen, and Yunchao Wei. Slca: Slow learner with classifier alignment for continual learning on a pre-trained model. In International Conference on Computer Vision, 2023. 3, 6, 7", + "[47] Bowen Zhao, Xi Xiao, Guojun Gan, Bin Zhang, and ShuTao Xia. Maintaining discrimination and fairness in class incremental learning. In Conference on Computer Vision and Pattern Recognition, 2020. 6", + "[48] Wanjun Zhong, Lianghong Guo, Qiqi Gao, He Ye, and Yanlin Wang. Memorybank: Enhancing large language models with long-term memory. In Association for the Advancement of Artificial Intelligence, 2024. 1, 2" + ], + "bbox": [ + 516, + 92, + 903, + 416 + ], + "page_idx": 9 + }, + { + "type": "page_number", + "text": "10086", + "bbox": [ + 480, + 945, + 519, + 955 + ], + "page_idx": 9 + } +] \ No newline at end of file diff --git a/2025/Do Your Best and Get Enough Rest for Continual Learning/364acbb3-eca1-422d-be23-a76c46789b30_model.json b/2025/Do Your Best and Get Enough Rest for Continual Learning/364acbb3-eca1-422d-be23-a76c46789b30_model.json new file mode 100644 index 0000000000000000000000000000000000000000..fdc1f231759749184094a27e04ed2a77c3f4da47 --- /dev/null +++ b/2025/Do Your Best and Get Enough Rest for Continual Learning/364acbb3-eca1-422d-be23-a76c46789b30_model.json @@ -0,0 +1,2112 @@ +[ + [ + { + "type": "header", + "bbox": [ + 0.107, + 0.003, + 0.182, + 0.044 + ], + "angle": 0, + "content": "CVF" + }, + { + "type": "header", + "bbox": [ + 0.237, + 0.001, + 0.812, + 0.047 + ], + "angle": 0, + "content": "This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore." + }, + { + "type": "title", + "bbox": [ + 0.199, + 0.131, + 0.8, + 0.154 + ], + "angle": 0, + "content": "Do Your Best and Get Enough Rest for Continual Learning" + }, + { + "type": "text", + "bbox": [ + 0.18, + 0.18, + 0.307, + 0.201 + ], + "angle": 0, + "content": "Hankyul Kang" + }, + { + "type": "text", + "bbox": [ + 0.356, + 0.181, + 0.474, + 0.2 + ], + "angle": 0, + "content": "Gregor Seifer1" + }, + { + "type": "text", + "bbox": [ + 0.524, + 0.181, + 0.653, + 0.2 + ], + "angle": 0, + "content": "Donghyun Lee²" + }, + { + "type": "text", + "bbox": [ + 0.7, + 0.181, + 0.822, + 0.2 + ], + "angle": 0, + "content": "Jongbin \\(\\mathrm{Ryu}^{1}\\)" + }, + { + "type": "text", + "bbox": [ + 0.383, + 0.215, + 0.61, + 0.235 + ], + "angle": 0, + "content": "\\(^{1}\\) Ajou University \\(^{2}\\) KAIST" + }, + { + "type": "title", + "bbox": [ + 0.249, + 0.269, + 0.327, + 0.285 + ], + "angle": 0, + "content": "Abstract" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.302, + 0.485, + 0.741 + ], + "angle": 0, + "content": "According to the forgetting curve theory, we can enhance memory retention by learning extensive data and taking adequate rest. This means that in order to effectively retain new knowledge, it is essential to learn it thoroughly and ensure sufficient rest so that our brain can memorize without forgetting. The main takeaway from this theory is that learning extensive data at once necessitates sufficient rest before learning the same data again. This aspect of human long-term memory retention can be effectively utilized to address the continual learning of neural networks. Retaining new knowledge for a long period of time without catastrophic forgetting is the critical problem of continual learning. Therefore, based on Ebbinghaus' theory, we introduce the view-batch model that adjusts the learning schedules to optimize the recall interval between retraining the same samples. The proposed view-batch model allows the network to get enough rest to learn extensive knowledge from the same samples with a recall interval of sufficient length. To this end, we specifically present two approaches: 1) a replay method that guarantees the optimal recall interval, and 2) a self-supervised learning that acquires extensive knowledge from a single training sample at a time. We empirically show that these approaches of our method are aligned with the forgetting curve theory, which can enhance long-term memory. In our experiments, we also demonstrate that our method significantly improves many state-of-the-art continual learning methods in various protocols and scenarios. We open-source this project at https://github.com/hankyul2/ViewBatchModel." + }, + { + "type": "title", + "bbox": [ + 0.092, + 0.772, + 0.222, + 0.788 + ], + "angle": 0, + "content": "1. Introduction" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.798, + 0.484, + 0.875 + ], + "angle": 0, + "content": "The forgetting curve theory [18] suggests that human memory tends to fade as time goes on; however, memory retention can be improved by repeated learning in an optimal recall interval. Motivated by this theory, many subsequent studies [6, 19, 27] have shown that spaced repetition with" + }, + { + "type": "image", + "bbox": [ + 0.523, + 0.267, + 0.709, + 0.367 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.539, + 0.37, + 0.691, + 0.381 + ], + "angle": 0, + "content": "(a) Short-term recall interval" + }, + { + "type": "image", + "bbox": [ + 0.714, + 0.268, + 0.897, + 0.367 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.734, + 0.37, + 0.872, + 0.381 + ], + "angle": 0, + "content": "(b) Optimal recall interval" + }, + { + "type": "image", + "bbox": [ + 0.524, + 0.381, + 0.709, + 0.481 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.539, + 0.484, + 0.69, + 0.495 + ], + "angle": 0, + "content": "(c) Long-term recall interval" + }, + { + "type": "image", + "bbox": [ + 0.714, + 0.381, + 0.896, + 0.481 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.739, + 0.484, + 0.866, + 0.496 + ], + "angle": 0, + "content": "(d) Degree of forgetting" + }, + { + "type": "image_caption", + "bbox": [ + 0.512, + 0.508, + 0.907, + 0.62 + ], + "angle": 0, + "content": "Figure 1. Conceptual graph of the forgetting curve. We show (a) short-term recall interval, (b) optimal recall interval, (c) long-term recall interval, and (d) degree of forgetting. (a-b) Expanding the recall interval improves long-term memory retention of neural networks by repeatedly recalling memory with moderate difficulty, whereas (c) an excessive recall interval decreases it. The depicted forgetting curve regarding recall interval is based on the spacing effect formula [6, 7] provided in the supplementary material." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.655, + 0.907, + 0.791 + ], + "angle": 0, + "content": "optimal recall interval enhances long-term memory retention in the human brain. This phenomenon is known as the spacing effect [7], where the optimal recall interval for repeated learning is crucial to maintain an acceptable level of forgetting knowledge. Several studies [7, 21, 36] have made significant progress on the spacing effect, shedding light on its effectiveness. There have also been studies [28, 30, 31] highlighting the importance of the recall interval of repeated learning." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.796, + 0.909, + 0.903 + ], + "angle": 0, + "content": "Although this theory has been widely recognized in human brain research, it is currently very little known in the field of machine learning. Only a few works [1, 4, 48] show this theory's potential in neural networks by utilizing it as an analysis toolkit or dynamic learning strategy. So, we investigate the impact of the forgetting curve theory on long-term memory retention in neural networks for continual learning." + }, + { + "type": "page_footnote", + "bbox": [ + 0.11, + 0.888, + 0.236, + 0.9 + ], + "angle": 0, + "content": "*Corresponding author." + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "10077" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.092, + 0.09, + 0.48, + 0.253 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.09, + 0.263, + 0.483, + 0.364 + ], + "angle": 0, + "content": "Figure 2. Overview of experimental results. We provide comprehensive comparisons of various factors for continual learning. We perform extensive experiments on step and buffer sizes (a-b), three different continual learning methods (c), whether to use the pre-trained model (d), three different benchmarks (e), and two evaluation protocols (f). In all cases, ours improves the baseline performance consistently." + }, + { + "type": "text", + "bbox": [ + 0.089, + 0.388, + 0.483, + 0.644 + ], + "angle": 0, + "content": "In Ebbinghaus' theory, we focus on the decay of memory retention in the forgetting curve as shown in Figure 1. The forgetting curve suggests that, as shown in Figure 1(a), the sample memory shows extreme knowledge decay when repeating its learning with a short time-space. On the other hand, as shown in Figure 1(c), if there is an excessive recall interval of repeated learning, too much memory is forgotten, leading to a decreased level of memory retention when relearning the same sample. However, as shown in Figure 1(b), when there is an optimal recall interval between repeated learning, the decay of memory retention also becomes gentle. This effect of the optimal recall interval implies that it is imperative to implement an appropriate learning schedule to retain memories efficiently in the repeated learning process. The notion behind this spacing effect can be usefully applied in continual learning, particularly to prevent catastrophic forgetting." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.645, + 0.483, + 0.703 + ], + "angle": 0, + "content": "However, empirically, we found that the current methods for continual learning have a short recall interval [24, 32], so they are limited to profit from the spacing effect, which requires a sufficient recall interval." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.705, + 0.483, + 0.902 + ], + "angle": 0, + "content": "To tackle this limitation, we introduce two approaches for training neural networks. Our first approach aims to prevent short recall intervals through the proposed replay method. We replay view-batch in the repeated training schedule so that short recall intervals can be delayed sufficiently. We augment multiple views from a single sample, and thus, the recall intervals of repeated learning can be adjusted. In our second approach, we make use of self-supervised learning to compensate for the delayed recall interval by training each sample extensively. We leverage the one-to-many divergence to learn the generalized self-supervised features within a view-batch, which enhances the capability of the networks. These two approaches en" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.092, + 0.905, + 0.303 + ], + "angle": 0, + "content": "sure that the neural networks have enough time to retrain the same samples, and at the same time, they can acquire substantial knowledge from training samples at a time. We apply our view-batch model (VBM) to both rehearsal and rehearsal-free continual learning scenarios, where addressing long-term memory retention is critical. Moreover, we show extensive experimental results of various continual protocols and scenarios. As an overview, Figure 2 showcases the comprehensive comparison against different step sizes, memory buffer size, baseline methods, use of the pretrained model, benchmarks, and protocols. For all these comparisons, our view-batch model achieves significant performance improvements. We have condensed our contributions into the following points:" + }, + { + "type": "text", + "bbox": [ + 0.514, + 0.304, + 0.905, + 0.394 + ], + "angle": 0, + "content": "- We propose the view-batch model that optimizes the recall interval between retraining samples in order to enhance memory decay in the continual learning task. The proposed view-batch model includes replay and self-supervised learning components to accomplish extensive learning in optimized intervals." + }, + { + "type": "text", + "bbox": [ + 0.514, + 0.395, + 0.905, + 0.483 + ], + "angle": 0, + "content": "- We showcase the effectiveness of our approach on various state-of-the-art continual learning methods, where ours consistently improves the performance. The proposed method is a drop-in replacement approach, so it does not require any extra computational costs for the performance improvements." + }, + { + "type": "text", + "bbox": [ + 0.514, + 0.485, + 0.905, + 0.543 + ], + "angle": 0, + "content": "- We reveal that enhancing memory decay can significantly improve the performance of the continual learning. This confirms the practical utility of theoretical knowledge for neural networks." + }, + { + "type": "list", + "bbox": [ + 0.514, + 0.304, + 0.905, + 0.543 + ], + "angle": 0, + "content": null + }, + { + "type": "title", + "bbox": [ + 0.514, + 0.56, + 0.655, + 0.575 + ], + "angle": 0, + "content": "2. Related Work" + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.586, + 0.898, + 0.602 + ], + "angle": 0, + "content": "2.1. Forgetting Curve Theory in Neural Networks" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.609, + 0.907, + 0.821 + ], + "angle": 0, + "content": "Several studies [25, 26] have employed the forgetting curve theory [18] to gain insights into the inherent memory retention of neural networks. Recently, Tirumala et al. [39] investigated the forgetting curve theory in neural networks regarding different model scales. Cao et al. [4] have also utilized the forgetting curve theory to examine the variations in memory conformation of the neural networks as influenced by pre-training methods. Additionally, Amiri et al. [1] proposed a dynamic learning scheduler to reduce total learning cost, leveraging optimal learning intervals corresponding to the forgetting curve. Zhong et al. [48] optimized memory retrieval timing based on the principles of the spacing effect theory, enabling the retention of the previous dialogue context existing over an extended duration." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.832, + 0.702, + 0.848 + ], + "angle": 0, + "content": "2.2. Data Augmentation" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.856, + 0.905, + 0.901 + ], + "angle": 0, + "content": "Data augmentation [14, 45] has been widely used in various existing networks. It helps networks learn rich visual information using a limited number of training samples." + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.519, + 0.956 + ], + "angle": 0, + "content": "10078" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.09, + 0.092, + 0.485, + 0.318 + ], + "angle": 0, + "content": "Data augmentation is also effective in self-supervised learning [5, 11]. Unlimited labels can be generated when different augmentation methods are applied to a single sample. Therefore, the network is able to learn with self-supervised labels of augmented samples. We use this augmentation method to construct multiple views in a view-batch and also learn self-supervision along with the supervised information of the continual learning process. There have been several studies [2, 22] on repeated augmentation, which allows learning multiple views from a single sample at a time. However, these studies only focused on improving the performance of the augmentation method and did not examine the effect of recall intervals. In addition, they do not apply self-supervised learning and only use label supervision to train networks." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.344, + 0.282, + 0.361 + ], + "angle": 0, + "content": "2.3. Continual Learning" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.372, + 0.485, + 0.691 + ], + "angle": 0, + "content": "The main problem of continual learning is preserving old knowledge while acquiring new one. This problem has been recognized as catastrophic forgetting, and one common strategy used to address this problem is to make use of a memory buffer [32]. The approaches to this strategy can be divided into two categories: one involves the direct storage of a subset of training samples from a previously learned task [32, 33], while the other generates training samples from the learned task [37]. This approach of utilizing memory buffers proves to be highly effective in addressing the issue of catastrophic forgetting; however, it does come with the trade-off of requiring additional memory space. Consequently, non-rehearsal continual learning methods [15, 33] have been widely studied to avoid additional memory usage. In addition, studies [41, 42, 46] using pre-trained models in non-rehearsal continual learning have been actively conducted recently. Meanwhile, there have been several works [16, 29] that tackle continual learning from the perspective of the regularization method. The architecture-based method [35, 44] has been another prevalent research direction for addressing continual learning." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.719, + 0.182, + 0.735 + ], + "angle": 0, + "content": "3. Method" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.75, + 0.484, + 0.901 + ], + "angle": 0, + "content": "We introduce a view-batch model designed to enhance the long-term memory retention of neural networks for continual learning (CL) tasks. In existing CL methods [24, 32, 44], the neural networks retrain the same samples with short time intervals, which prohibits long-term memory retention. To address this short recall interval, we first investigate the augmentation model with the replay method to ensure a sufficient recall interval in Sec. 3.1. Then, Sec. 3.2 introduces view-batch learning model to extensively train multiple samples of the same instance in view-batch." + }, + { + "type": "image", + "bbox": [ + 0.522, + 0.088, + 0.707, + 0.186 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.582, + 0.188, + 0.646, + 0.2 + ], + "angle": 0, + "content": "(a) Baseline" + }, + { + "type": "image", + "bbox": [ + 0.71, + 0.089, + 0.897, + 0.186 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.763, + 0.188, + 0.843, + 0.199 + ], + "angle": 0, + "content": "(b) View-batch" + }, + { + "type": "image_caption", + "bbox": [ + 0.512, + 0.212, + 0.908, + 0.337 + ], + "angle": 0, + "content": "Figure 3. Schematic illustration of the proposed view-batch model. In subfigure (b), we show our view-batch model employing the replay \\((\\mathrm{V} = 4)\\) and self-supervised learning approach. In contrast to (a) the baseline method, we learn multiple views of the same sample (marked as different shades) using the proposed view-batch self-supervised loss to learn it extensively and ensure enough time-space between recall intervals. For simplicity, we assume in (a) that the entire training data and batch size are the same as four, thus a single training epoch constitutes one batch." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.365, + 0.757, + 0.381 + ], + "angle": 0, + "content": "3.1. Replay with Augmentation" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.389, + 0.907, + 0.644 + ], + "angle": 0, + "content": "Our view-batch model employs a replaying method with the augmentation to adjust the recall interval of retraining the same samples. In the conventional learning scheduler, a sample-batch is replayed in the training process of neural networks. It means a batch for training networks consists of multiple unique training samples. Therefore, the recall interval of retraining is the same as the number of training samples. However, the strategy on our learning scheduler alternates the replaying method at the view-batch level. We structure a view-batch to have multiple views of a single sample. Thus, the recall interval of a single sample increases with the size of view-batch. The concept of this view-batch model for the replay method is shown in Figure 3. It shows that when learning four samples in sequence, the time-space of re-training is four, whereas when learning \\(V\\) views of a single sample, the time-space is extended by \\(V\\) times." + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.646, + 0.907, + 0.707 + ], + "angle": 0, + "content": "Formally, we define samples as \\(I\\), sample-batch as \\(\\mathcal{B}^I = \\{I_i\\}_{i=1}^B\\), the number of batches in a epoch as \\(T\\), and learning scheduler as \\(\\mathcal{A}\\). In the conventional learning algorithm, we replay each sample-batch to establish the scheduler." + }, + { + "type": "equation", + "bbox": [ + 0.56, + 0.718, + 0.907, + 0.753 + ], + "angle": 0, + "content": "\\[\n\\mathcal {A} _ {\\text {c o n v e n t i o n a l}} = \\left[ \\underbrace {\\mathcal {B} _ {1} ^ {I} , \\cdots , \\mathcal {B} _ {T} ^ {I}} _ {\\text {r e c a l l i n t e r v a l} = B \\times T}, \\mathcal {B} _ {1} ^ {I}, \\dots , \\mathcal {B} _ {T} ^ {I} \\right]. \\tag {1}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.765, + 0.909, + 0.903 + ], + "angle": 0, + "content": "Therefore, the recall interval of retraining is the same as the number of training samples \\(B \\times T\\) in an epoch. However, to extend the recall interval, we replace the replaying unit from the sample-batch to the view-batch \\(\\mathcal{B}^{\\nu} = \\{\\mathcal{V}_i\\}_{i=1}^{B}\\). We construct a view-batch by including multiple views of the same sample \\(\\mathcal{V}_i = \\{I_i\\}_{i=1}^V\\), where \\(V\\) denotes the number of views consisting of different augmentations of a single image. By modifying the replaying unit of Equation (1) with a constructed view-batch, we define our new learning" + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "10079" + } + ], + [ + { + "type": "code_caption", + "bbox": [ + 0.093, + 0.091, + 0.437, + 0.106 + ], + "angle": 0, + "content": "Algorithm 1 Pseudocode for our view-batch model." + }, + { + "type": "algorithm", + "bbox": [ + 0.093, + 0.112, + 0.484, + 0.334 + ], + "angle": 0, + "content": "Require: dataset \\(\\mathcal{D}^{\\mathrm{train}}\\) parameters \\(\\theta\\) , learning rate \\(\\gamma\\) , batch size \\(B\\) , view-batch size \\(V\\) \n1: Initialize network parameters \\(\\theta\\) \n2: Update \\(B\\gets B / V\\) \n3: for number of training iterations do \n4: \\(\\begin{array}{rl} & {\\mathcal{B}^{I} = \\{I_{i}\\}_{i = 1}^{B}\\mathrm{with}I_{i}\\stackrel {\\mathrm{i.i.d.}}{\\sim}D^{\\mathrm{train}}}\\\\ & {\\mathrm{for}i = 1,\\dots ,B\\mathrm{do}}\\\\ & {\\left\\{ \\begin{array}{l}\\mathcal{V}_{i}\\gets \\{I_{i}\\}_{j = 1}^{V}\\\\ \\mathcal{V}_{i,1}\\gets A_{w}(\\mathcal{V}_{i,1})\\\\ \\mathcal{V}_{i,2}\\gets A_{s}(\\mathcal{V}_{i,2:}) \\end{array} \\right.}\\\\ & {\\bar{\\mathcal{B}}^{\\nu}\\gets \\{\\mathcal{V}_{i}\\}_{i = 1}^{B}}\\\\ & {\\mathrm{Compute}l_{\\mathrm{sup}}\\gets L_{\\mathrm{sup}}(\\theta ,\\mathcal{B}^{\\nu})}\\\\ & {\\mathrm{Compute}l_{\\mathrm{ssl}}\\gets L_{\\mathrm{ssl}}(\\theta ,\\mathcal{B}^{\\nu})\\triangleright SSL\\mathrm{of}\\mathrm{Equation}(3)}\\\\ & {\\mathrm{Update}\\theta \\gets \\theta -\\gamma \\nabla_{\\theta}(l_{\\mathrm{sup}} + l_{\\mathrm{ssl}})} \\end{array}\\) \n7: replay of Equation (2) \n8: \n9: \n10: \n11: \n12:" + }, + { + "type": "text", + "bbox": [ + 0.091, + 0.359, + 0.234, + 0.373 + ], + "angle": 0, + "content": "schedule as follows:" + }, + { + "type": "equation", + "bbox": [ + 0.145, + 0.384, + 0.483, + 0.42 + ], + "angle": 0, + "content": "\\[\n\\mathcal {A} _ {\\text {o u r s}} = \\left[ \\underbrace {\\mathcal {B} _ {1} ^ {\\nu} , \\cdots , \\mathcal {B} _ {T} ^ {\\nu}} _ {\\text {r e c a l l i n t e r v a l} = B \\times T \\times V}, \\mathcal {B} _ {1} ^ {\\nu}, \\dots , \\mathcal {B} _ {T} ^ {\\nu} \\right], \\tag {2}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.43, + 0.484, + 0.598 + ], + "angle": 0, + "content": "Based on this rescheduling, the recall interval is increased by \\( V \\) times. While we process \\( V \\) views at a time, we decrease the total learning epoch by \\( V \\), ensuring that the number of total samples that networks have seen during training is the same as the baseline method. As a result, networks repeatedly forget and retrain samples during extended recall intervals, thereby improving long-term memory retention. This learning of multiple views of a single sample adjusts the time-space while networks learn the same sample repeatedly. To promote extensive learning, we present the self-supervised learning in the following." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.606, + 0.32, + 0.623 + ], + "angle": 0, + "content": "3.2. Self-supervised Learning" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.629, + 0.484, + 0.902 + ], + "angle": 0, + "content": "We present a self-supervised learning model with our VBM to extensively learn a sample at a time. It is well-known that self-supervision generalizes the networks rather than acquiring task-specific knowledge [5]. Further, task-agnostic knowledge from self-supervision is more robust to catastrophic forgetting of continual learning [8]. For these reasons, we propose the SSL approach to effectively learn view-batch in a (i) simple and an (ii) efficient manner. For (i) simplicity, unlike previous SSL methods [5, 11, 20], which require changes, such as architecture design [8] or teacher networks [12], we only modify an objective function to learn common knowledge within a view-batch of same samples. To satisfy (ii) efficiency, our method does not use an extra training phase or longer training epochs than supervised learning, ensuring the minimal training cost as analyzed in the supplementary material. Thus, the proposed method could be used in most continual learning methods without increasing computing costs as a drop-in replace" + }, + { + "type": "image", + "bbox": [ + 0.517, + 0.09, + 0.644, + 0.171 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.517, + 0.173, + 0.644, + 0.187 + ], + "angle": 0, + "content": "(a) Degree of forgetting" + }, + { + "type": "image", + "bbox": [ + 0.646, + 0.095, + 0.772, + 0.171 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.655, + 0.174, + 0.762, + 0.187 + ], + "angle": 0, + "content": "(b) Forgetting curve" + }, + { + "type": "image", + "bbox": [ + 0.773, + 0.095, + 0.9, + 0.171 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.802, + 0.174, + 0.871, + 0.187 + ], + "angle": 0, + "content": "(c) Accuracy" + }, + { + "type": "image_caption", + "bbox": [ + 0.512, + 0.198, + 0.907, + 0.352 + ], + "angle": 0, + "content": "Figure 4. Empirical findings related to forgetting curve theory. (a) We report the degree of forgetting for different recall intervals with a \\(95\\%\\) confidence interval denoted as shaded region. The degree of forgetting is measured as the performance degradation of each sample's classification accuracy between recall intervals. (b) We show the decay of memory retention as the learning progresses. This graph shows that when the formal learning stage is finished, memory retention decays over time for all three cases. (c) We compare the network's classification accuracy of continual learning. It shows that x3 achieves the best performance thanks to the slow memory retention decay." + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.378, + 0.619, + 0.392 + ], + "angle": 0, + "content": "ment approach." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.393, + 0.907, + 0.545 + ], + "angle": 0, + "content": "To implement our approach, we employ a one-to-many divergence loss to learn self-supervision within a view-batch \\(\\{\\mathcal{V}_i\\}_{i=1}^B\\). Our one-to-many divergence loss minimizes the average divergence between one weak-augmented view \\(A_w(\\mathcal{V}_{i,1})\\) and the remaining strong-augmented views \\(A_s(\\mathcal{V}_{i,2:})\\) of a view-batch at the logit level. By aggregating different augmented views closer, networks learn common characteristics of objects. Specifically, leveraging KL divergence \\(D_{\\mathrm{KL}}\\) between different augmentations, we define one-to-many-based self-supervised loss as" + }, + { + "type": "equation", + "bbox": [ + 0.533, + 0.554, + 0.906, + 0.598 + ], + "angle": 0, + "content": "\\[\nL _ {\\mathrm {s s l}} \\left(f _ {\\theta}, \\mathcal {B} ^ {\\nu}\\right) = \\frac {1}{B \\cdot (V - 1)} \\sum_ {i = 1} ^ {B} \\sum_ {j = 2} ^ {V} D _ {\\mathrm {K L}} \\left(p _ {i} ^ {1} | | p _ {i} ^ {j}\\right), \\tag {3}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.61, + 0.906, + 0.687 + ], + "angle": 0, + "content": "where \\( p_i^1 \\) and \\( p_i^j \\) represent network's prediction for weak augmented samples \\( \\sigma(f_{\\theta}(A_w(\\mathcal{V}_{i,1}))) \\) and strong augmented samples \\( \\sigma(f_{\\theta}(A_s(\\mathcal{V}_{i,j}))) \\), \\( f_{\\theta} \\) is neural networks, and \\( \\sigma \\) stand for the softmax. Finally, as shown in Algorithm 1, we define our final objective function as" + }, + { + "type": "equation", + "bbox": [ + 0.592, + 0.697, + 0.907, + 0.716 + ], + "angle": 0, + "content": "\\[\n\\min _ {f _ {\\theta}} L _ {\\sup } \\left(f _ {\\theta}, \\mathcal {B} ^ {\\nu}\\right) + L _ {\\operatorname {s s l}} \\left(f _ {\\theta}, \\mathcal {B} ^ {\\nu}\\right), \\tag {4}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.726, + 0.907, + 0.824 + ], + "angle": 0, + "content": "where \\( L_{\\sup}(f_{\\theta}, \\mathcal{B}^{\\nu}) = \\frac{1}{B \\cdot V} \\sum_{i=1}^{B} \\sum_{j=1}^{V} \\mathcal{H}(y_{i}, p_{i}^{j}) \\) of each view's prediction \\( p_{i}^{j} \\) and its label \\( y_{i} \\) using cross-entropy loss \\( \\mathcal{H} \\). We select widely used auto-augmentation [13] as a strong augmentation method, which slightly affects the performance of the baseline method as shown in Tab. 6, while using horizontal flip as weak augmentation." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.832, + 0.704, + 0.849 + ], + "angle": 0, + "content": "3.3. Empirical Evidence" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.856, + 0.906, + 0.901 + ], + "angle": 0, + "content": "This subsection empirically confirms our assumption that optimal recall interval improves memory retention, and accordingly, the performance of the continual method is en" + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.52, + 0.957 + ], + "angle": 0, + "content": "10080" + } + ], + [ + { + "type": "table", + "bbox": [ + 0.165, + 0.089, + 0.833, + 0.447 + ], + "angle": 0, + "content": "
BufferMethodS-CIFAR-10S-Tiny-ImageNet
CILTILAvg*ΔCILTILAvg*Δ
-Joint92.20±0.1598.31±0.1295.2659.99±0.1982.04±0.1071.02
0LwF-61.98±0.9861.98-15.26±0.7115.26
VBM-LwF-77.53±3.0677.53+15.55-51.21±0.3451.21+35.95
200DER56.58±2.0889.32±1.3272.9511.56±0.5841.67±0.6326.62
ER50.27±0.7091.66±0.7970.978.67±0.2338.03±0.7923.35
VBM-ER52.56±1.5993.58±0.2373.07+2.108.83±0.1344.03±1.1226.43+3.08
iCaRL64.11±1.8690.20±0.8877.1619.09±0.1353.07±0.2736.08
VBM-iCaRL69.73±0.8792.76±0.4581.25+4.0921.39±0.1856.31±0.4238.85+2.77
DER++61.67±1.0290.61±1.1376.1412.48±1.6041.19±0.4826.84
VBM-DER++66.99±2.9894.30±0.2180.65+4.5113.67±0.4144.36±0.8829.02+2.18
500DER65.93±5.3892.52±0.6879.2318.04±1.3052.90±0.9135.47
ER56.37±0.9694.15±0.0475.2610.20±0.3749.13±0.3129.67
VBM-ER63.39±2.7794.20±0.2578.80+3.5411.21±0.3052.01±0.7131.61+1.94
iCaRL62.22±3.4588.86±3.2375.5423.93±0.9758.89±0.8541.41
VBM-iCaRL68.97±0.3893.10±0.6681.04+5.5026.89±0.5462.61±0.1744.75+3.34
DER++69.94±2.0493.49±0.2681.7219.39±1.0951.71±0.3535.55
VBM-DER++75.22±0.8194.81±0.3985.02+3.3019.46±0.4649.80±0.8034.63-0.92
5120DER80.79±0.7594.46±0.3087.6332.92±3.0964.77±2.1448.85
ER74.23±4.1896.86±0.2885.5528.07±0.3767.58±0.3047.83
VBM-ER79.29±2.2997.02±0.1388.16+2.6133.07±0.3668.91±0.5050.99+3.16
iCaRL78.49±0.3894.84±0.8286.6732.16±0.2466.75±0.3649.46
VBM-iCaRL79.40±0.6396.09±0.3287.75+1.0834.40±0.1268.86±0.2651.63+2.17
DER++85.42±0.7496.24±0.1990.8334.06±0.5765.28±1.2049.67
VBM-DER++86.62±0.3897.07±0.0791.85+1.0235.14±0.3667.52±0.4851.33+1.66
" + }, + { + "type": "table_caption", + "bbox": [ + 0.09, + 0.458, + 0.907, + 0.529 + ], + "angle": 0, + "content": "Table 1. Experimental results on CIL and TIL protocols. In this evaluation, we report the last top-1 accuracy \\((\\%)\\) on the S-CIFAR-10 and S-Tiny-ImageNet benchmarks with four different buffer sizes. ResNet-18 backbone is used for each continual learning method in this evaluation. *Avg is the averaged accuracy of CIL and TIL. We reproduce baseline methods following Buzzega et al. [3]. We run every method three times with different random seeds for a reliable result. Ours improves performance consistently in most results. The performance of Joint is the upper bound of this experiment, where all old training samples are used in every step." + }, + { + "type": "table", + "bbox": [ + 0.114, + 0.551, + 0.462, + 0.7 + ], + "angle": 0, + "content": "
MethodAvgΔLastΔ
Joint-46.48±3.62
AGEM24.88±0.5728.94±0.67
AGEM-R24.72±0.5928.01±0.93
FDR34.62±0.3334.46±0.32
DER39.84±0.1540.46±0.47
ER32.38±2.8832.02±2.39
VBM-ER35.80±0.55+3.4234.72±0.37+2.70
DER++34.56±3.4234.75±2.41
VBM-DER++42.16±0.20+7.2441.81±0.08+7.06
" + }, + { + "type": "table_caption", + "bbox": [ + 0.09, + 0.711, + 0.483, + 0.766 + ], + "angle": 0, + "content": "Table 2. Experimental results on DIL protocol. This evaluation shows Avg and Last top-1 accuracy \\((\\%)\\) on the DomainNet benchmark. We utilize ResNet-18 as the backbone network in this evaluation." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.795, + 0.483, + 0.901 + ], + "angle": 0, + "content": "hanced. Specifically, the logical development of our assumption is that 1) if the network sufficiently forgets a sample at the optimal recall interval and then extensively retrans it, 2) memory retention is enhanced as training progresses, and therefore, 3) the accuracy of continual learning is also improved. To evaluate our assumption, we measure 1) the degree of forgetting between a retraining session for" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.555, + 0.907, + 0.888 + ], + "angle": 0, + "content": "training samples, 2) the decay speed of the memory retention, and 3) the average accuracy at which networks correctly classify their class labels. Figure 4 shows the results for the empirical evidence of our assumption. As depicted in Figure 4(a), the degree of forgetting becomes larger as the recall interval increases. It is obvious that the longer recall interval allows the network to forget a training sample before retraining it again. Figure 4(b) demonstrates the different levels of memory retention decay due to the recall intervals. The experiment results show that memory decays slowest in the x3 case, which has an interval three times longer than the baseline x1 case. Notably, our method achieves considerably higher performance at the end of the formal learning stage. This result is attributed to the fact that multiple epochs are used in training networks, and the memory retention decay between epochs is slow for the x3 case. Accordingly, x3 with our replay method achieves improved accuracy compared to other recall interval cases, as shown in Figure 4(c). We conduct more experiments to show empirical proofs on different datasets, backbone networks, and buffer sizes in Figure 6. The definition of forgetting is given in the supplementary material." + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.946, + 0.518, + 0.956 + ], + "angle": 0, + "content": "10081" + } + ], + [ + { + "type": "table", + "bbox": [ + 0.134, + 0.089, + 0.868, + 0.27 + ], + "angle": 0, + "content": "
Method5 Step10 Step20 Step
AvgΔLastΔAvgΔLastΔAvgΔLastΔ
Joint80.40-80.41-81.49-
iCaRL71.14±0.3459.7165.27±1.0250.7461.20±0.8343.75
UCIR62.77±0.8247.3158.66±0.7143.3958.17±0.3040.63
BiC73.10±0.5562.1068.80±1.2053.5466.48±0.3247.02
WA72.81±0.2860.8469.46±0.2953.7867.33±0.1547.31
DyTox--73.66±0.0260.6772.27±0.1856.32
DER76.77±0.0168.06±0.0075.72±0.0864.32±0.0574.96±0.0161.80±0.07
VBM-DER78.60±0.23+1.8370.60±0.12+2.5478.12±0.07+2.4067.04±0.11+2.7276.95±0.02+1.9964.29±0.04+2.49
TCIL77.33±0.0869.48±0.1476.33±0.1565.66±0.0274.32±0.0162.54±0.02
VBM-TCIL79.23±0.01+1.9071.23±0.04+1.7578.02±0.03+1.6968.14±0.08+2.4876.83±0.00+2.5167.16±0.09+4.62
" + }, + { + "type": "table_caption", + "bbox": [ + 0.09, + 0.278, + 0.908, + 0.337 + ], + "angle": 0, + "content": "Table 3. Experimental results on rehearsal-based CIL protocol. In this evaluation, we provide Avg and Last top-1 accuracy \\((\\%)\\) on the S-CIFAR-100 benchmark with three different class incremental steps. ResNet-18 backbone is utilized to evaluate each rehearsal-based method. This evaluation also shows that ours achieves consistent performance improvements. Results of DyTox are imported from its original work [17]. We reproduce results of DER and TCIL using official implementation, while other results come from Yan et al. [44]." + }, + { + "type": "table", + "bbox": [ + 0.093, + 0.358, + 0.485, + 0.505 + ], + "angle": 0, + "content": "
Method5 Step10 Step
AvgΔLastΔAvgΔLastΔ
Joint80.40-80.41-
LwF-MC32.8911.4220.716.36
BiC54.7435.4245.0824.42
LwM39.3116.7025.478.57
WA63.6749.3645.6826.14
DER40.05±0.0017.11±0.0026.67±0.008.71±0.00
VBM-DER40.47±0.00+0.4217.47±0.01+0.3626.95±0.00+0.288.90±0.00+0.19
TCIL64.40±0.1552.37±0.0856.84±0.7540.31±1.04
VBM-TCIL66.40±0.07+2.0054.69±0.02+2.3261.12±0.02+4.2845.92±0.25+5.61
" + }, + { + "type": "table_caption", + "bbox": [ + 0.09, + 0.515, + 0.484, + 0.626 + ], + "angle": 0, + "content": "Table 4. Experimental results on non-rehearsal-based CIL protocol. For this evaluation, we measure Avg and Last top-1 accuracy \\((\\%)\\) on the S-CIFAR-100 benchmark with two different class incremental steps. ResNet-18 backbone is applied to existing and our methods. We report original results from LwF-MC and LwM [15]. We reproduce the results of DER and TCIL by ourselves using official implementation. Other results are reported from Huang et al. [24]." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.655, + 0.224, + 0.672 + ], + "angle": 0, + "content": "4. Experiments" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.681, + 0.483, + 0.803 + ], + "angle": 0, + "content": "This section demonstrates the capabilities of the proposed view-batch model in various continual learning protocols and scenarios through extensive performance evaluation. We perform experiments for three continual learning protocols, such as class, task, and domain incremental learning. Further, we also assess our view-batch model in both rehearsal and non-rehearsal, as well as under the pre-training-based continual learning scenarios." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.816, + 0.299, + 0.833 + ], + "angle": 0, + "content": "4.1. Experimental Settings" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.84, + 0.484, + 0.903 + ], + "angle": 0, + "content": "Datasets. We evaluate our methods on widely used benchmarks such as S-CIFAR-10/100, S-Tiny-ImageNet, S-ImageNet-R, and DomainNet where the prefix of 'S-' denotes the sequential data configuration. We use the com" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.361, + 0.907, + 0.392 + ], + "angle": 0, + "content": "monly adopted class ordering lists, following existing works of Yan et al. [44] and Huang et al. [24]." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.412, + 0.909, + 0.761 + ], + "angle": 0, + "content": "Protocols. In our experiments, we utilize three continual learning protocols. First, we make use of task incremental learning (TIL), which involves training neural networks on a series of tasks in a sequential manner while evaluating their performance on both newly acquired and previously learned tasks. Specifically, the TIL protocol consists of the following detailed configurations: 1) a task is defined by a set of label classes, and 2) we evaluate networks knowing the task identity of test samples. Second, we adopt class incremental learning (CIL), which is identical to TIL except that it evaluates test samples without knowing their task identity. Third, we use domain incremental learning (DIL), which focuses on learning the new samples across multiple sequential domains rather than acquiring new classes. In addition to these three protocols, we also include both rehearsal and non-rehearsal scenarios, divided by the use of a memory buffer during continual learning. For the non-rehearsal scenario, we evaluate both scratch- and pre-training-based learning scenarios to showcase extensive experimental results. For all these protocols and scenarios, we provide the mean and standard deviation of three runs for reliable experimental results. The details of the experiment can be found in the supplementary material." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.78, + 0.911, + 0.903 + ], + "angle": 0, + "content": "Baseline Methods. We apply our view-batch to recent baseline continual learning methods such as iCaRL [32], ER [34], DER [3, 44], TCIL [24], and SLCA [46]. We also include results from A-GEM [10], LwM [15], UCIR [23], BiC [43], WA [47], DyTox [17], L2P [42], S-iPrompt [40], DualPrompt [41], and CODA-Prompt [38]. We follow the baseline's hyper-parameters for a fair comparison between ours and its respective baseline methods. Specifically, we" + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.52, + 0.957 + ], + "angle": 0, + "content": "10082" + } + ], + [ + { + "type": "table", + "bbox": [ + 0.146, + 0.089, + 0.855, + 0.248 + ], + "angle": 0, + "content": "
S-CIFAR-100S-ImageNet-R
MethodTop-1Top-5Top-1Top-5
AvgLastAvgLastAvgLastAvgLast
Joint-92.54±0.14-99.24±0.02-82.41±0.02-93.27±0.07
L2P90.42±0.5683.76±0.4298.89±0.0398.08±0.1576.76±0.3373.56±0.4289.39±0.1686.61±0.03
DualPrompt90.17±0.3885.10±0.2398.90±0.0797.68±0.1273.31±0.1969.12±0.1889.19±0.0985.46±0.04
CODA-Prompt91.27±0.6686.86±0.5899.29±0.0998.64±0.0781.89±0.5976.79±0.1793.30±0.3490.49±0.15
S-iPrompt92.19±0.4088.43±0.4299.19±0.0598.63±0.0571.47±0.1468.77±0.2788.44±0.1984.91±0.31
SLCA94.31±1.1291.57±0.4699.64±0.0499.26±0.0383.58±0.6578.81±0.2194.39±0.2692.02±0.25
VBM-SLCA94.52±0.8591.78±0.1899.64±0.0399.29±0.0384.61±0.5980.07±0.1294.86±0.1992.74±0.14
" + }, + { + "type": "table_caption", + "bbox": [ + 0.09, + 0.258, + 0.907, + 0.3 + ], + "angle": 0, + "content": "Table 5. Experimental results on pre-trained model-based CIL protocol. This evaluation presents Avg and Last top-1,5 accuracy (%) on the S-CIFAR-100 and S-ImageNet-R benchmarks without using a memory buffer. ViT-B/16 backbone is adopted for each pre-trained model-based method. We reproduce baseline methods using official implementation by ourselves, except for L2P." + }, + { + "type": "image", + "bbox": [ + 0.102, + 0.324, + 0.283, + 0.433 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.166, + 0.435, + 0.221, + 0.447 + ], + "angle": 0, + "content": "(a) iCaRL" + }, + { + "type": "image", + "bbox": [ + 0.288, + 0.324, + 0.468, + 0.433 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.347, + 0.435, + 0.409, + 0.447 + ], + "angle": 0, + "content": "(b) DER++" + }, + { + "type": "image_caption", + "bbox": [ + 0.09, + 0.459, + 0.483, + 0.542 + ], + "angle": 0, + "content": "Figure 5. Experimental results on memory retention decay. This analysis reports memory retention decay of the first three tasks on the S-CIFAR-10 dataset, comparing the proposed approach against the baseline methods. The green numbers at the end of the last task are the accuracy gain of the proposed approach over the baselines." + }, + { + "type": "image", + "bbox": [ + 0.101, + 0.558, + 0.222, + 0.67 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.132, + 0.672, + 0.191, + 0.682 + ], + "angle": 0, + "content": "(a) Dataset" + }, + { + "type": "image", + "bbox": [ + 0.226, + 0.559, + 0.346, + 0.67 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.253, + 0.671, + 0.319, + 0.682 + ], + "angle": 0, + "content": "(b) Network" + }, + { + "type": "image", + "bbox": [ + 0.351, + 0.559, + 0.471, + 0.67 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.383, + 0.671, + 0.437, + 0.682 + ], + "angle": 0, + "content": "(c) Buffer" + }, + { + "type": "image_caption", + "bbox": [ + 0.09, + 0.695, + 0.483, + 0.764 + ], + "angle": 0, + "content": "Figure 6. Experimental results on recall interval. We measure the last top-1 accuracy \\((\\%)\\) (\\(\\uparrow\\)) in the top row and report forgetting \\((\\downarrow)\\) at the bottom row, varying recall intervals with different learning factors. Dark colors mean better. We use the ResNet-18 backbone." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.795, + 0.483, + 0.901 + ], + "angle": 0, + "content": "use ResNet and Vision Transformer (ViT) as backbone networks in traditional and prompt-based approaches. For the traditional approach, we randomly initialize backbone weights at the beginning of every run. On the other hand, for the prompt-based approach, we adopt a frozen backbone whose weights are pre-trained on a large-scale dataset following recent prompt-based CL studies [42, 46]" + }, + { + "type": "table", + "bbox": [ + 0.521, + 0.324, + 0.903, + 0.483 + ], + "angle": 0, + "content": "
MethodReplayStrong augmentSSLCILTILAvgΔ
iCaRL---64.1190.2077.16-1.05
-✓-62.0490.1776.11
✓--67.5191.8779.69+2.53
✓✓-66.7891.2178.99+1.83
✓✓✓69.7392.7681.25+4.09
DER++---61.6790.6176.14+0.76
-✓-62.6691.1476.90
✓--64.7592.7978.77+2.63
✓✓-64.4493.3878.91+2.77
✓✓✓66.9994.3080.65+4.51
" + }, + { + "type": "table_caption", + "bbox": [ + 0.512, + 0.493, + 0.907, + 0.606 + ], + "angle": 0, + "content": "Table 6. Experimental results on factor analysis. We demonstrate the last accuracy (\\%) of CIL and TIL on S-CIFAR-10, varying our main components with different augmentation types. We observe that the proposed components, such as replay and SSL, improve performance consistently and confirm that performance enhancement comes from them, not simply from strong augmentation. In the strong augmentation setting without view batch replay, we maintain the same ratio of weak to strong augmentations." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.633, + 0.716, + 0.649 + ], + "angle": 0, + "content": "4.2. Experimental Results" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.658, + 0.907, + 0.839 + ], + "angle": 0, + "content": "Continual Learning under Three Protocols. We evaluate our method in the three continual learning protocols such as CIL, TIL, and DIL. For TIL and CIL, we use different memory buffer sizes: 0, 200, 500, and 5,120. For DIL, we utilize 50 samples per class, resulting in 17,250 samples in a buffer. We apply our view-batch model to LwF for the non-rehearsal memory buffer (i.e. 0 size) and ER, DER++, and iCaRL for the other buffer sizes. Tabs. 1 and 2 show that our method improves the performance of continual learning methods for all protocols. Moreover, we demonstrate that baseline methods employing our view-batch are comparable to other continual learning methods." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.84, + 0.907, + 0.901 + ], + "angle": 0, + "content": "We conduct further in-depth experiments under the most widely used CIL protocol. We consider three configurations with step sizes of 5, 10, and 20, which are the number of total incremental steps, as shown in Tab. 3. In addition, we" + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.518, + 0.956 + ], + "angle": 0, + "content": "10083" + } + ], + [ + { + "type": "table", + "bbox": [ + 0.1, + 0.089, + 0.477, + 0.226 + ], + "angle": 0, + "content": "
MethodCurrent sampleBuffer sampleCILTILAvgΔ
iCaRL--64.1190.2077.16-
✓-64.6494.1279.38+2.22
-✓64.9994.2079.60+2.44
✓✓69.7392.7681.25+4.09
DER++--61.6790.6176.14-
✓-64.8893.2279.05+2.91
-✓62.6994.1578.42+2.28
✓✓66.9994.3080.65+4.51
" + }, + { + "type": "table_caption", + "bbox": [ + 0.09, + 0.235, + 0.483, + 0.307 + ], + "angle": 0, + "content": "Table 7. Experimental results on sample type. We measure the last accuracy (\\%) of CIL and TIL on S-CIFAR-10, applying our method to different sample types. We use the ResNet-18 as backbone. This analysis shows that our method works well with both sample types." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.331, + 0.484, + 0.422 + ], + "angle": 0, + "content": "take into account the use of a memory buffer in two scenarios, such as rehearsal and non-rehearsal in Tab. 4. Both experimental results demonstrate that our view-batch model consistently improves the performance of continual learning methods under the various step sizes and memory buffer configurations." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.44, + 0.484, + 0.62 + ], + "angle": 0, + "content": "Non-rehearsal Continual Learning. While the nonrehearsal scenario has the advantage of not using memory buffers, they inherently face the drawback of limited performance due to the inability to retrain networks using old data in the memory buffer. To overcome this limitation, many recent studies have focused on non-rehearsal continual learning using a pre-trained model trained with a large-scale dataset. Thus, we evaluate our view-batch model in a scenario where models are pre-trained, such as SCLA. Experimental results in Tab. 5 indicate that the proposed view-batch model is effective across all benchmarks for this scenario." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.634, + 0.188, + 0.651 + ], + "angle": 0, + "content": "5. Analysis" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.659, + 0.483, + 0.901 + ], + "angle": 0, + "content": "We present the analysis of the proposed view-batch model. First, we validate that the optimal recall interval mitigates catastrophic forgetting of early tasks. Figure 5 visualize that the recall interval x3 maintains long-term memory retention resolving catastrophic forgetting of initial tasks. Also, Figure 6 analyzes how different factors (i.e., dataset, network scales, memory buffer size) affect the optimal recall intervals in terms of last top-1 accuracy and forgetting metrics [9, 10, 42]. We find that the recall interval x3 or x4 improves the accuracy and resolves catastrophic forgetting in various scenarios. More experimental results of different methods and task types can be found in the supplementary material. Tab. 6 shows that two main factors of the proposed method consistently improve their respective baselines, identifying that the performance improvement comes from the proposed components, not strong augmentation." + }, + { + "type": "image", + "bbox": [ + 0.527, + 0.089, + 0.891, + 0.212 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.513, + 0.222, + 0.907, + 0.292 + ], + "angle": 0, + "content": "Figure 7. Experimental results on accuracy and forgetting. This analysis shows (a) the last top-1 accuracy \\((\\%)\\) and (b) corresponding forgetting measures on the S-CIFAR-10 dataset with different buffer sizes. Our method consistently improves accuracy and forgetting measures." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.318, + 0.905, + 0.363 + ], + "angle": 0, + "content": "Lastly, in the rehearsal scenario, we have two different sample types, current- and buffer-based samples. Tab. 7 exhibit that ours performs well with both types." + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.364, + 0.907, + 0.501 + ], + "angle": 0, + "content": "We examine the proposed method in terms of the last accuracy and forgetting measurement. We apply the proposed method to two different baseline approaches. We compute average forgetting at \\( T \\)th task as \\( \\frac{1}{T - 1}\\sum_{t = 1}^{T - 1}\\max_{j\\in \\{t,\\dots ,T - 1\\}}a_j^t -a_T^t \\) where \\( a_j^t \\) means \\( t \\)th task accuracy at \\( j \\) step. Notably, higher forgetting means severe catastrophic forgetting of previous tasks. Figure 7 shows that the view-batch model improves accuracy while addressing the forgetting problem simultaneously." + }, + { + "type": "title", + "bbox": [ + 0.514, + 0.513, + 0.634, + 0.529 + ], + "angle": 0, + "content": "6. Conclusion" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.539, + 0.907, + 0.766 + ], + "angle": 0, + "content": "We propose the view-batch model, which is generally applicable to various continual learning scenarios. Inspired by Ebbinghaus's forgetting curve theory, the proposed view-batch model optimizes the recall interval between retraining samples, improving neural networks' long-term memory. The proposed method consists of two main components: replay and self-supervised learning. The replay optimizes the recall interval for guaranteeing sufficient memory forgetting by grouping multiple views of a sample. Also, the self-supervised learning ensures extensive learning for the multiple views using the one-to-many divergence loss. Experimental results show that sufficient forgetting preserves long-term memory and ultimately improves continual learning accuracy. We hope the proposed approach to continual learning can boost future research." + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.781, + 0.909, + 0.901 + ], + "angle": 0, + "content": "Acknowledgment. This paper was supported in part by the ETRI Grant funded by Korean Government (Fundamental Technology Research for Human-Centric Autonomous Intelligent Systems) under Grant 24ZB1200, Artificial Intelligence Convergence Innovation Human Resources Development (IITP-2025-RS-2023-00255968), Artificial Intelligence Innovation Hub (RS-2021-II212068), and the NRF Grant (RS-2024-00356486)." + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.52, + 0.957 + ], + "angle": 0, + "content": "10084" + } + ], + [ + { + "type": "title", + "bbox": [ + 0.093, + 0.09, + 0.188, + 0.106 + ], + "angle": 0, + "content": "References" + }, + { + "type": "ref_text", + "bbox": [ + 0.101, + 0.115, + 0.484, + 0.171 + ], + "angle": 0, + "content": "[1] Hadi Amiri, Timothy Miller, and Guergana Savova. Repeat before forgetting: Spaced repetition for efficient and effective training of neural networks. In Conference on Empirical Methods in Natural Language Processing, 2017. 1, 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.101, + 0.172, + 0.484, + 0.225 + ], + "angle": 0, + "content": "[2] Maxim Berman, Hervé Jégou, Andrea Vedaldi, Iasonas Kokkinos, and Matthijs Douze. Multigrain: a unified image embedding for classes and instances. arXiv:1902.05509, 2019.3" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.228, + 0.483, + 0.282 + ], + "angle": 0, + "content": "[3] Pietro Buzzega, Matteo Boschini, Angelo Porrello, Davide Abati, and Simone Calderara. Dark experience for general continual learning: a strong, simple baseline. Advances in Neural Information Processing Systems, 2020. 5, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.284, + 0.483, + 0.366 + ], + "angle": 0, + "content": "[4] Boxi Cao, Qiaoyu Tang, Hongyu Lin, Shanshan Jiang, Bin Dong, Xianpei Han, Jiawei Chen, Tianshu Wang, and Le Sun. Retentive or forgetful? diving into the knowledge memorizing mechanism of language models. In Joint International Conference on Computational Linguistics, Language Resources and Evaluation, 2024. 1, 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.368, + 0.483, + 0.422 + ], + "angle": 0, + "content": "[5] Mathilde Caron, Hugo Touvron, Ishan Misra, Herve Jégou, Julien Mairal, Piotr Bojanowski, and Armand Joulin. Emerging properties in self-supervised vision transformers. In International Conference on Computer Vision, 2021. 3, 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.424, + 0.483, + 0.478 + ], + "angle": 0, + "content": "[6] Nicholas J Cepeda, Harold Pashler, Edward Vul, John T Wixted, and Doug Rohrer. Distributed practice in verbal recall tasks: A review and quantitative synthesis. *Psychological bulletin*, 2006. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.481, + 0.483, + 0.535 + ], + "angle": 0, + "content": "[7] Nicholas J Cepeda, Edward Vul, Doug Rohrer, John T Wixted, and Harold Pashler. Spacing effects in learning: A temporal ridgeline of optimal retention. Psychological science, 2008. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.537, + 0.483, + 0.577 + ], + "angle": 0, + "content": "[8] Hyuntak Cha, Jaeho Lee, and Jinwoo Shin. Co2l: Contrastive continual learning. In International Conference on Computer Vision, 2021. 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.579, + 0.483, + 0.634 + ], + "angle": 0, + "content": "[9] Arslan Chaudhry, Puneet K Dokania, Thalaiyasingam Ajthan, and Philip HS Torr. Riemannian walk for incremental learning: Understanding forgetting and intransigence. In European Conference on Computer Vision, 2018. 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.635, + 0.483, + 0.689 + ], + "angle": 0, + "content": "[10] Arslan Chaudhry, Marc'Aurelio Ranzato, Marcus Rohrbach, and Mohamed Elhoseiny. Efficient lifelong learning with a gem. In International Conference on Learning Representations, 2019. 6, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.691, + 0.483, + 0.746 + ], + "angle": 0, + "content": "[11] Ting Chen, Simon Kornblith, Mohammad Norouzi, and Geoffrey Hinton. A simple framework for contrastive learning of visual representations. In International Conference on Machine Learning, 2020. 3, 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.747, + 0.483, + 0.802 + ], + "angle": 0, + "content": "[12] Haoyang Cheng, Haitao Wen, Heqian Qiu, Lanxiao Wang, Minjian Zhang, and Hongliang Li. Must unsupervised continual learning relies on previous information? In Conference on Computer Vision and Pattern Recognition, 2024. 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.804, + 0.483, + 0.844 + ], + "angle": 0, + "content": "[13] Ekin D Cubuk, Barret Zoph, Dandelion Mane, Vijay Vasudevan, and Quoc V Le. Autoaugment: Learning augmentation policies from data. arXiv:1805.09501, 2018. 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.846, + 0.483, + 0.901 + ], + "angle": 0, + "content": "[14] Ekin D Cubuk, Barret Zoph, Jonathon Shlens, and Quoc V Le. Randaugment: Practical automated data augmentation with a reduced search space. In Conference on Computer Vision and Pattern Recognition Workshops, 2020. 2" + }, + { + "type": "list", + "bbox": [ + 0.094, + 0.115, + 0.484, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.093, + 0.905, + 0.147 + ], + "angle": 0, + "content": "[15] Prithviraj Dhar, Rajat Vikram Singh, Kuan-Chuan Peng, Ziyan Wu, and Rama Chellappa. Learning without memorizing. In Conference on Computer Vision and Pattern Recognition, 2019. 3, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.151, + 0.905, + 0.206 + ], + "angle": 0, + "content": "[16] Arthur Douillard, Matthieu Cord, Charles Ollion, Thomas Robert, and Eduardo Valle. Podnet: Pooled outputs distillation for small-tasks incremental learning. In European Conference on Computer Vision, 2020. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.209, + 0.905, + 0.263 + ], + "angle": 0, + "content": "[17] Arthur Douillard, Alexandre Ramé, Guillaume Couairon, and Matthieu Cord. Dytox: Transformers for continual learning with dynamic token expansion. In Conference on Computer Vision and Pattern Recognition, 2022. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.267, + 0.904, + 0.293 + ], + "angle": 0, + "content": "[18] Hermann Ebbinghaus. Memory: A contribution to experimental psychology. Annals of neurosciences, 2013. 1, 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.297, + 0.905, + 0.324 + ], + "angle": 0, + "content": "[19] Kate Gordon. Class results with spaced and unspaced memorizing. Journal of Experimental Psychology, 1925. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.327, + 0.905, + 0.409 + ], + "angle": 0, + "content": "[20] Jean-Bastien Grill, Florian Strub, Florent Altché, Corentin Tallec, Pierre Richemond, Elena Buchatskaya, Carl Doersch, Bernardo Avila Pires, Zhaohan Guo, Mohammad Gheshlaghi Azar, et al. Bootstrap your own latent-a new approach to self-supervised learning. Advances in Neural Information Processing Systems, 2020. 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.413, + 0.905, + 0.495 + ], + "angle": 0, + "content": "[21] Karri S Hawley, Katie E Cherry, Emily O Boudreaux, and Erin M Jackson. A comparison of adjusted spaced retrieval versus a uniform expanded retrieval schedule for learning a name-face association in older adults with probable Alzheimer's disease. Journal of Clinical and Experimental Neuropsychology, 2008. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.499, + 0.905, + 0.539 + ], + "angle": 0, + "content": "[22] Elad Hoffer, Tal Ben-Nun, Itay Hubara, Niv Giladi, Torsten Hoefler, and Daniel Soudry. Augment your batch: better training with larger batches. arXiv:1901.09335, 2019. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.542, + 0.905, + 0.598 + ], + "angle": 0, + "content": "[23] Saihui Hou, Xinyu Pan, Chen Change Loy, Zilei Wang, and Dahua Lin. Learning a unified classifier incrementally via rebalancing. In Conference on Computer Vision and Pattern Recognition, 2019. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.601, + 0.905, + 0.667 + ], + "angle": 0, + "content": "[24] Bingchen Huang, Zhineng Chen, Peng Zhou, Jiayin Chen, and Zuxuan Wu. Resolving task confusion in dynamic expansion architectures for class incremental learning. In Association for the Advancement of Artificial Intelligence, 2023. 2, 3, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.672, + 0.905, + 0.74 + ], + "angle": 0, + "content": "[25] Anette Hunziker, Yuxin Chen, Oisin Mac Aodha, Manuel Gomez Rodriguez, Andreas Krause, Pietro Perona, Yisong Yue, and Adish Singla. Teaching multiple concepts to a forgetful learner. Advances in Neural Information Processing Systems, 2019. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.744, + 0.905, + 0.783 + ], + "angle": 0, + "content": "[26] Szabolcs Káli and Peter Dayan. Replay, repair and consolidation. Advances in Neural Information Processing Systems, 2002. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.788, + 0.905, + 0.828 + ], + "angle": 0, + "content": "[27] Sean HK Kang. Spaced repetition promotes efficient and effective learning: Policy implications for instruction. *Policy Insights from the Behavioral and Brain Sciences*, 2016. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.832, + 0.907, + 0.899 + ], + "angle": 0, + "content": "[28] Jeffrey D Karpicke and Henry L Roediger III. Expanding retrieval practice promotes short-term retention, but equally spaced retrieval enhances long-term retention. Journal of experimental psychology: learning, memory, and cognition, 2007. 1" + }, + { + "type": "list", + "bbox": [ + 0.517, + 0.093, + 0.907, + 0.899 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "10085" + } + ], + [ + { + "type": "ref_text", + "bbox": [ + 0.092, + 0.092, + 0.482, + 0.134 + ], + "angle": 0, + "content": "[29] Zhizhong Li and Derek Hoiem. Learning without forgetting. IEEE transactions on pattern analysis and machine intelligence, 2017. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.091, + 0.136, + 0.482, + 0.163 + ], + "angle": 0, + "content": "[30] Cecil Alec Mace. The psychology of study. McBride, 1932. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.165, + 0.483, + 0.207 + ], + "angle": 0, + "content": "[31] Arthur W Melton. The situation with respect to the spacing of repetitions and memory. Journal of Verbal Learning and Verbal Behavior, 1970. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.208, + 0.483, + 0.263 + ], + "angle": 0, + "content": "[32] Sylvestre-Alvise Rebuffi, Alexander Kolesnikov, Georg Sperl, and Christoph H Lampert. icarl: Incremental classifier and representation learning. In Conference on Computer Vision and Pattern Recognition, 2017. 2, 3, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.265, + 0.482, + 0.332 + ], + "angle": 0, + "content": "[33] Matthew Riemer, Ignacio Cases, Robert Ajemian, Miao Liu, Irina Rish, Yuhai Tu, , and Gerald Tesauro. Learning to learn without forgetting by maximizing transfer and minimizing interference. In International Conference on Learning Representations, 2019. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.334, + 0.482, + 0.39 + ], + "angle": 0, + "content": "[34] David Rolnick, Arun Ahuja, Jonathan Schwarz, Timothy Lillicrap, and Gregory Wayne. Experience replay for continual learning. Advances in Neural Information Processing Systems, 2019. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.392, + 0.482, + 0.447 + ], + "angle": 0, + "content": "[35] Joan Serra, Didac Suris, Marius Miron, and Alexandros Karatzoglou. Overcoming catastrophic forgetting with hard attention to the task. In International Conference on Machine Learning, 2018. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.449, + 0.482, + 0.49 + ], + "angle": 0, + "content": "[36] John J Shaughnessy. Long-term retention and the spacing effect in free-recall and frequency judgments. The American Journal of Psychology, 1977. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.491, + 0.482, + 0.533 + ], + "angle": 0, + "content": "[37] Hanul Shin, Jung Kwon Lee, Jaehong Kim, and Jiwon Kim. Continual learning with deep generative replay. Advances in Neural Information Processing Systems, 2017. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.535, + 0.483, + 0.617 + ], + "angle": 0, + "content": "[38] James Seale Smith, Leonid Karlinsky, Vyshnavi Gutta, Paola Cascante-Bonilla, Donghyun Kim, Assaf Arbelle, Rameswar Panda, Rogerio Feris, and Zsolt Kira. Coda-prompt: Continual decomposed attention-based prompting for rehearsal-free continual learning. In Conference on Computer Vision and Pattern Recognition, 2023. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.619, + 0.483, + 0.674 + ], + "angle": 0, + "content": "[39] Kushal Tirumala, Aram Markosyan, Luke Zettlemoyer, and Armen Aghajanyan. Memorization without overfitting: Analyzing the training dynamics of large language models. Advances in Neural Information Processing Systems, 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.676, + 0.483, + 0.731 + ], + "angle": 0, + "content": "[40] Yabin Wang, Zhiwu Huang, and Xiaopeng Hong. S-prompts learning with pre-trained transformers: An occam's razor for domain incremental learning. Advances in Neural Information Processing Systems, 2022. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.732, + 0.483, + 0.802 + ], + "angle": 0, + "content": "[41] Zifeng Wang, Zizhao Zhang, Sayna Ebrahimi, Ruoxi Sun, Han Zhang, Chen-Yu Lee, Xiaoqi Ren, Guolong Su, Vincent Perot, Jennifer Dy, et al. Dualprompt: Complementary prompting for rehearsal-free continual learning. In European Conference on Computer Vision, 2022. 3, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.803, + 0.483, + 0.871 + ], + "angle": 0, + "content": "[42] Zifeng Wang, Zizhao Zhang, Chen-Yu Lee, Han Zhang, Ruoxi Sun, Xiaqi Ren, Guolong Su, Vincent Perot, Jennifer Dy, and Tomas Pfister. Learning to prompt for continual learning. In Conference on Computer Vision and Pattern Recognition, 2022. 3, 6, 7, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.873, + 0.483, + 0.901 + ], + "angle": 0, + "content": "[43] Yue Wu, Yinpeng Chen, Lijuan Wang, Yuancheng Ye, Zicheng Liu, Yandong Guo, and Yun Fu. Large scale in" + }, + { + "type": "list", + "bbox": [ + 0.091, + 0.092, + 0.483, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.548, + 0.093, + 0.905, + 0.12 + ], + "angle": 0, + "content": "cremental learning. In Conference on Computer Vision and Pattern Recognition, 2019. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.122, + 0.905, + 0.177 + ], + "angle": 0, + "content": "[44] Shipeng Yan, Jiangwei Xie, and Xuming He. Der: Dynamically expandable representation for class incremental learning. In Conference on Computer Vision and Pattern Recognition, 2021. 3, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.179, + 0.905, + 0.246 + ], + "angle": 0, + "content": "[45] Sangdoo Yun, Dongyoon Han, Seong Joon Oh, Sanghyuk Chun, Junsuk Choe, and Youngjoon Yoo. Cutmix: Regularization strategy to train strong classifiers with localizable features. In International Conference on Computer Vision, 2019. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.249, + 0.905, + 0.304 + ], + "angle": 0, + "content": "[46] Gengwei Zhang, Liyuan Wang, Guoliang Kang, Ling Chen, and Yunchao Wei. Slca: Slow learner with classifier alignment for continual learning on a pre-trained model. In International Conference on Computer Vision, 2023. 3, 6, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.305, + 0.905, + 0.36 + ], + "angle": 0, + "content": "[47] Bowen Zhao, Xi Xiao, Guojun Gan, Bin Zhang, and ShuTao Xia. Maintaining discrimination and fairness in class incremental learning. In Conference on Computer Vision and Pattern Recognition, 2020. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.362, + 0.905, + 0.417 + ], + "angle": 0, + "content": "[48] Wanjun Zhong, Lianghong Guo, Qiqi Gao, He Ye, and Yanlin Wang. Memorybank: Enhancing large language models with long-term memory. In Association for the Advancement of Artificial Intelligence, 2024. 1, 2" + }, + { + "type": "list", + "bbox": [ + 0.517, + 0.093, + 0.905, + 0.417 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.946, + 0.52, + 0.957 + ], + "angle": 0, + "content": "10086" + } + ] +] \ No newline at end of file diff --git a/2025/Do Your Best and Get Enough Rest for Continual Learning/364acbb3-eca1-422d-be23-a76c46789b30_origin.pdf b/2025/Do Your Best and Get Enough Rest for Continual Learning/364acbb3-eca1-422d-be23-a76c46789b30_origin.pdf new file mode 100644 index 0000000000000000000000000000000000000000..0f15866a8020c9d996e53ffd4baac5948ea3e224 --- /dev/null +++ b/2025/Do Your Best and Get Enough Rest for Continual Learning/364acbb3-eca1-422d-be23-a76c46789b30_origin.pdf @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:6a5cf3552983e40b53b0fa9d73ff91d8a8fa117fda8418f7165d7c3ff8b8c059 +size 876734 diff --git a/2025/Do Your Best and Get Enough Rest for Continual Learning/full.md b/2025/Do Your Best and Get Enough Rest for Continual Learning/full.md new file mode 100644 index 0000000000000000000000000000000000000000..f10cd9c18c0b2169f59820072cd19e6e47339ea5 --- /dev/null +++ b/2025/Do Your Best and Get Enough Rest for Continual Learning/full.md @@ -0,0 +1,292 @@ +# Do Your Best and Get Enough Rest for Continual Learning + +Hankyul Kang + +Gregor Seifer1 + +Donghyun Lee² + +Jongbin $\mathrm{Ryu}^{1}$ + +$^{1}$ Ajou University $^{2}$ KAIST + +# Abstract + +According to the forgetting curve theory, we can enhance memory retention by learning extensive data and taking adequate rest. This means that in order to effectively retain new knowledge, it is essential to learn it thoroughly and ensure sufficient rest so that our brain can memorize without forgetting. The main takeaway from this theory is that learning extensive data at once necessitates sufficient rest before learning the same data again. This aspect of human long-term memory retention can be effectively utilized to address the continual learning of neural networks. Retaining new knowledge for a long period of time without catastrophic forgetting is the critical problem of continual learning. Therefore, based on Ebbinghaus' theory, we introduce the view-batch model that adjusts the learning schedules to optimize the recall interval between retraining the same samples. The proposed view-batch model allows the network to get enough rest to learn extensive knowledge from the same samples with a recall interval of sufficient length. To this end, we specifically present two approaches: 1) a replay method that guarantees the optimal recall interval, and 2) a self-supervised learning that acquires extensive knowledge from a single training sample at a time. We empirically show that these approaches of our method are aligned with the forgetting curve theory, which can enhance long-term memory. In our experiments, we also demonstrate that our method significantly improves many state-of-the-art continual learning methods in various protocols and scenarios. We open-source this project at https://github.com/hankyul2/ViewBatchModel. + +# 1. Introduction + +The forgetting curve theory [18] suggests that human memory tends to fade as time goes on; however, memory retention can be improved by repeated learning in an optimal recall interval. Motivated by this theory, many subsequent studies [6, 19, 27] have shown that spaced repetition with + +![](images/c3ac23cdbc9dbc772bc852352bfd6acc3d7e32ca4dbf55f6548953e05a3cd2ad.jpg) + +![](images/4062f74cdf243453891fc09a9f280d166d1fddbdbcf4b553b9573e7ffb925912.jpg) + +![](images/d39edceeb4ff7ce1f8040457f82497e9700f7cbc5b7b56458a10444e36c27124.jpg) +(a) Short-term recall interval +(c) Long-term recall interval +Figure 1. Conceptual graph of the forgetting curve. We show (a) short-term recall interval, (b) optimal recall interval, (c) long-term recall interval, and (d) degree of forgetting. (a-b) Expanding the recall interval improves long-term memory retention of neural networks by repeatedly recalling memory with moderate difficulty, whereas (c) an excessive recall interval decreases it. The depicted forgetting curve regarding recall interval is based on the spacing effect formula [6, 7] provided in the supplementary material. + +![](images/d15d8cc7dc22dc911ab44f2ec4a2ad180ba64b5fb19bc82f43eda0ae0dfc6e57.jpg) +(b) Optimal recall interval +(d) Degree of forgetting + +optimal recall interval enhances long-term memory retention in the human brain. This phenomenon is known as the spacing effect [7], where the optimal recall interval for repeated learning is crucial to maintain an acceptable level of forgetting knowledge. Several studies [7, 21, 36] have made significant progress on the spacing effect, shedding light on its effectiveness. There have also been studies [28, 30, 31] highlighting the importance of the recall interval of repeated learning. + +Although this theory has been widely recognized in human brain research, it is currently very little known in the field of machine learning. Only a few works [1, 4, 48] show this theory's potential in neural networks by utilizing it as an analysis toolkit or dynamic learning strategy. So, we investigate the impact of the forgetting curve theory on long-term memory retention in neural networks for continual learning. + +![](images/171ede31f5cab52ada7526c381f2dc6f5e9d81195679eeec7e8ea1f90a340a1d.jpg) +Figure 2. Overview of experimental results. We provide comprehensive comparisons of various factors for continual learning. We perform extensive experiments on step and buffer sizes (a-b), three different continual learning methods (c), whether to use the pre-trained model (d), three different benchmarks (e), and two evaluation protocols (f). In all cases, ours improves the baseline performance consistently. + +In Ebbinghaus' theory, we focus on the decay of memory retention in the forgetting curve as shown in Figure 1. The forgetting curve suggests that, as shown in Figure 1(a), the sample memory shows extreme knowledge decay when repeating its learning with a short time-space. On the other hand, as shown in Figure 1(c), if there is an excessive recall interval of repeated learning, too much memory is forgotten, leading to a decreased level of memory retention when relearning the same sample. However, as shown in Figure 1(b), when there is an optimal recall interval between repeated learning, the decay of memory retention also becomes gentle. This effect of the optimal recall interval implies that it is imperative to implement an appropriate learning schedule to retain memories efficiently in the repeated learning process. The notion behind this spacing effect can be usefully applied in continual learning, particularly to prevent catastrophic forgetting. + +However, empirically, we found that the current methods for continual learning have a short recall interval [24, 32], so they are limited to profit from the spacing effect, which requires a sufficient recall interval. + +To tackle this limitation, we introduce two approaches for training neural networks. Our first approach aims to prevent short recall intervals through the proposed replay method. We replay view-batch in the repeated training schedule so that short recall intervals can be delayed sufficiently. We augment multiple views from a single sample, and thus, the recall intervals of repeated learning can be adjusted. In our second approach, we make use of self-supervised learning to compensate for the delayed recall interval by training each sample extensively. We leverage the one-to-many divergence to learn the generalized self-supervised features within a view-batch, which enhances the capability of the networks. These two approaches en + +sure that the neural networks have enough time to retrain the same samples, and at the same time, they can acquire substantial knowledge from training samples at a time. We apply our view-batch model (VBM) to both rehearsal and rehearsal-free continual learning scenarios, where addressing long-term memory retention is critical. Moreover, we show extensive experimental results of various continual protocols and scenarios. As an overview, Figure 2 showcases the comprehensive comparison against different step sizes, memory buffer size, baseline methods, use of the pretrained model, benchmarks, and protocols. For all these comparisons, our view-batch model achieves significant performance improvements. We have condensed our contributions into the following points: + +- We propose the view-batch model that optimizes the recall interval between retraining samples in order to enhance memory decay in the continual learning task. The proposed view-batch model includes replay and self-supervised learning components to accomplish extensive learning in optimized intervals. +- We showcase the effectiveness of our approach on various state-of-the-art continual learning methods, where ours consistently improves the performance. The proposed method is a drop-in replacement approach, so it does not require any extra computational costs for the performance improvements. +- We reveal that enhancing memory decay can significantly improve the performance of the continual learning. This confirms the practical utility of theoretical knowledge for neural networks. + +# 2. Related Work + +# 2.1. Forgetting Curve Theory in Neural Networks + +Several studies [25, 26] have employed the forgetting curve theory [18] to gain insights into the inherent memory retention of neural networks. Recently, Tirumala et al. [39] investigated the forgetting curve theory in neural networks regarding different model scales. Cao et al. [4] have also utilized the forgetting curve theory to examine the variations in memory conformation of the neural networks as influenced by pre-training methods. Additionally, Amiri et al. [1] proposed a dynamic learning scheduler to reduce total learning cost, leveraging optimal learning intervals corresponding to the forgetting curve. Zhong et al. [48] optimized memory retrieval timing based on the principles of the spacing effect theory, enabling the retention of the previous dialogue context existing over an extended duration. + +# 2.2. Data Augmentation + +Data augmentation [14, 45] has been widely used in various existing networks. It helps networks learn rich visual information using a limited number of training samples. + +Data augmentation is also effective in self-supervised learning [5, 11]. Unlimited labels can be generated when different augmentation methods are applied to a single sample. Therefore, the network is able to learn with self-supervised labels of augmented samples. We use this augmentation method to construct multiple views in a view-batch and also learn self-supervision along with the supervised information of the continual learning process. There have been several studies [2, 22] on repeated augmentation, which allows learning multiple views from a single sample at a time. However, these studies only focused on improving the performance of the augmentation method and did not examine the effect of recall intervals. In addition, they do not apply self-supervised learning and only use label supervision to train networks. + +# 2.3. Continual Learning + +The main problem of continual learning is preserving old knowledge while acquiring new one. This problem has been recognized as catastrophic forgetting, and one common strategy used to address this problem is to make use of a memory buffer [32]. The approaches to this strategy can be divided into two categories: one involves the direct storage of a subset of training samples from a previously learned task [32, 33], while the other generates training samples from the learned task [37]. This approach of utilizing memory buffers proves to be highly effective in addressing the issue of catastrophic forgetting; however, it does come with the trade-off of requiring additional memory space. Consequently, non-rehearsal continual learning methods [15, 33] have been widely studied to avoid additional memory usage. In addition, studies [41, 42, 46] using pre-trained models in non-rehearsal continual learning have been actively conducted recently. Meanwhile, there have been several works [16, 29] that tackle continual learning from the perspective of the regularization method. The architecture-based method [35, 44] has been another prevalent research direction for addressing continual learning. + +# 3. Method + +We introduce a view-batch model designed to enhance the long-term memory retention of neural networks for continual learning (CL) tasks. In existing CL methods [24, 32, 44], the neural networks retrain the same samples with short time intervals, which prohibits long-term memory retention. To address this short recall interval, we first investigate the augmentation model with the replay method to ensure a sufficient recall interval in Sec. 3.1. Then, Sec. 3.2 introduces view-batch learning model to extensively train multiple samples of the same instance in view-batch. + +![](images/a0443f4ba890a4c2b249ce19436e98d93600314e1dc1fcc19e74591c0ea53809.jpg) +(a) Baseline + +![](images/61f95c7408d3f3bb330b735292fca7910ba239dbbf172835f3bbbb95a65e5d29.jpg) +(b) View-batch +Figure 3. Schematic illustration of the proposed view-batch model. In subfigure (b), we show our view-batch model employing the replay $(\mathrm{V} = 4)$ and self-supervised learning approach. In contrast to (a) the baseline method, we learn multiple views of the same sample (marked as different shades) using the proposed view-batch self-supervised loss to learn it extensively and ensure enough time-space between recall intervals. For simplicity, we assume in (a) that the entire training data and batch size are the same as four, thus a single training epoch constitutes one batch. + +# 3.1. Replay with Augmentation + +Our view-batch model employs a replaying method with the augmentation to adjust the recall interval of retraining the same samples. In the conventional learning scheduler, a sample-batch is replayed in the training process of neural networks. It means a batch for training networks consists of multiple unique training samples. Therefore, the recall interval of retraining is the same as the number of training samples. However, the strategy on our learning scheduler alternates the replaying method at the view-batch level. We structure a view-batch to have multiple views of a single sample. Thus, the recall interval of a single sample increases with the size of view-batch. The concept of this view-batch model for the replay method is shown in Figure 3. It shows that when learning four samples in sequence, the time-space of re-training is four, whereas when learning $V$ views of a single sample, the time-space is extended by $V$ times. + +Formally, we define samples as $I$ , sample-batch as $\mathcal{B}^I = \{I_i\}_{i=1}^B$ , the number of batches in a epoch as $T$ , and learning scheduler as $\mathcal{A}$ . In the conventional learning algorithm, we replay each sample-batch to establish the scheduler. + +$$ +\mathcal {A} _ {\text {c o n v e n t i o n a l}} = \left[ \underbrace {\mathcal {B} _ {1} ^ {I} , \cdots , \mathcal {B} _ {T} ^ {I}} _ {\text {r e c a l l i n t e r v a l} = B \times T}, \mathcal {B} _ {1} ^ {I}, \dots , \mathcal {B} _ {T} ^ {I} \right]. \tag {1} +$$ + +Therefore, the recall interval of retraining is the same as the number of training samples $B \times T$ in an epoch. However, to extend the recall interval, we replace the replaying unit from the sample-batch to the view-batch $\mathcal{B}^{\nu} = \{\mathcal{V}_i\}_{i=1}^{B}$ . We construct a view-batch by including multiple views of the same sample $\mathcal{V}_i = \{I_i\}_{i=1}^V$ , where $V$ denotes the number of views consisting of different augmentations of a single image. By modifying the replaying unit of Equation (1) with a constructed view-batch, we define our new learning + +Algorithm 1 Pseudocode for our view-batch model. +Require: dataset $\mathcal{D}^{\mathrm{train}}$ parameters $\theta$ , learning rate $\gamma$ , batch size $B$ , view-batch size $V$ +1: Initialize network parameters $\theta$ +2: Update $B\gets B / V$ +3: for number of training iterations do +4: $\begin{array}{rl} & {\mathcal{B}^{I} = \{I_{i}\}_{i = 1}^{B}\mathrm{with}I_{i}\stackrel {\mathrm{i.i.d.}}{\sim}D^{\mathrm{train}}}\\ & {\mathrm{for}i = 1,\dots ,B\mathrm{do}}\\ & {\left\{ \begin{array}{l}\mathcal{V}_{i}\gets \{I_{i}\}_{j = 1}^{V}\\ \mathcal{V}_{i,1}\gets A_{w}(\mathcal{V}_{i,1})\\ \mathcal{V}_{i,2}\gets A_{s}(\mathcal{V}_{i,2:}) \end{array} \right.}\\ & {\bar{\mathcal{B}}^{\nu}\gets \{\mathcal{V}_{i}\}_{i = 1}^{B}}\\ & {\mathrm{Compute}l_{\mathrm{sup}}\gets L_{\mathrm{sup}}(\theta ,\mathcal{B}^{\nu})}\\ & {\mathrm{Compute}l_{\mathrm{ssl}}\gets L_{\mathrm{ssl}}(\theta ,\mathcal{B}^{\nu})\triangleright SSL\mathrm{of}\mathrm{Equation}(3)}\\ & {\mathrm{Update}\theta \gets \theta -\gamma \nabla_{\theta}(l_{\mathrm{sup}} + l_{\mathrm{ssl}})} \end{array}$ +7: replay of Equation (2) +8: +9: +10: +11: +12: + +schedule as follows: + +$$ +\mathcal {A} _ {\text {o u r s}} = \left[ \underbrace {\mathcal {B} _ {1} ^ {\nu} , \cdots , \mathcal {B} _ {T} ^ {\nu}} _ {\text {r e c a l l i n t e r v a l} = B \times T \times V}, \mathcal {B} _ {1} ^ {\nu}, \dots , \mathcal {B} _ {T} ^ {\nu} \right], \tag {2} +$$ + +Based on this rescheduling, the recall interval is increased by $V$ times. While we process $V$ views at a time, we decrease the total learning epoch by $V$ , ensuring that the number of total samples that networks have seen during training is the same as the baseline method. As a result, networks repeatedly forget and retrain samples during extended recall intervals, thereby improving long-term memory retention. This learning of multiple views of a single sample adjusts the time-space while networks learn the same sample repeatedly. To promote extensive learning, we present the self-supervised learning in the following. + +# 3.2. Self-supervised Learning + +We present a self-supervised learning model with our VBM to extensively learn a sample at a time. It is well-known that self-supervision generalizes the networks rather than acquiring task-specific knowledge [5]. Further, task-agnostic knowledge from self-supervision is more robust to catastrophic forgetting of continual learning [8]. For these reasons, we propose the SSL approach to effectively learn view-batch in a (i) simple and an (ii) efficient manner. For (i) simplicity, unlike previous SSL methods [5, 11, 20], which require changes, such as architecture design [8] or teacher networks [12], we only modify an objective function to learn common knowledge within a view-batch of same samples. To satisfy (ii) efficiency, our method does not use an extra training phase or longer training epochs than supervised learning, ensuring the minimal training cost as analyzed in the supplementary material. Thus, the proposed method could be used in most continual learning methods without increasing computing costs as a drop-in replace + +![](images/bb59e13fbd75342f101e6a10e184be0c6c0459ee20c8c795814ba8549e9d9d57.jpg) +(a) Degree of forgetting + +![](images/522012dc639823957c1068037f3f5cdee1e721d183f991328ff70134b26119bc.jpg) +(b) Forgetting curve +Figure 4. Empirical findings related to forgetting curve theory. (a) We report the degree of forgetting for different recall intervals with a $95\%$ confidence interval denoted as shaded region. The degree of forgetting is measured as the performance degradation of each sample's classification accuracy between recall intervals. (b) We show the decay of memory retention as the learning progresses. This graph shows that when the formal learning stage is finished, memory retention decays over time for all three cases. (c) We compare the network's classification accuracy of continual learning. It shows that x3 achieves the best performance thanks to the slow memory retention decay. + +![](images/d62d2a25d9a2e39869c99a38b80f9c2627d6e7c567e7ac89d4ffb1c81396312a.jpg) +(c) Accuracy + +ment approach. + +To implement our approach, we employ a one-to-many divergence loss to learn self-supervision within a view-batch $\{\mathcal{V}_i\}_{i=1}^B$ . Our one-to-many divergence loss minimizes the average divergence between one weak-augmented view $A_w(\mathcal{V}_{i,1})$ and the remaining strong-augmented views $A_s(\mathcal{V}_{i,2:})$ of a view-batch at the logit level. By aggregating different augmented views closer, networks learn common characteristics of objects. Specifically, leveraging KL divergence $D_{\mathrm{KL}}$ between different augmentations, we define one-to-many-based self-supervised loss as + +$$ +L _ {\mathrm {s s l}} \left(f _ {\theta}, \mathcal {B} ^ {\nu}\right) = \frac {1}{B \cdot (V - 1)} \sum_ {i = 1} ^ {B} \sum_ {j = 2} ^ {V} D _ {\mathrm {K L}} \left(p _ {i} ^ {1} | | p _ {i} ^ {j}\right), \tag {3} +$$ + +where $p_i^1$ and $p_i^j$ represent network's prediction for weak augmented samples $\sigma(f_{\theta}(A_w(\mathcal{V}_{i,1})))$ and strong augmented samples $\sigma(f_{\theta}(A_s(\mathcal{V}_{i,j})))$ , $f_{\theta}$ is neural networks, and $\sigma$ stand for the softmax. Finally, as shown in Algorithm 1, we define our final objective function as + +$$ +\min _ {f _ {\theta}} L _ {\sup } \left(f _ {\theta}, \mathcal {B} ^ {\nu}\right) + L _ {\operatorname {s s l}} \left(f _ {\theta}, \mathcal {B} ^ {\nu}\right), \tag {4} +$$ + +where $L_{\sup}(f_{\theta}, \mathcal{B}^{\nu}) = \frac{1}{B \cdot V} \sum_{i=1}^{B} \sum_{j=1}^{V} \mathcal{H}(y_{i}, p_{i}^{j})$ of each view's prediction $p_{i}^{j}$ and its label $y_{i}$ using cross-entropy loss $\mathcal{H}$ . We select widely used auto-augmentation [13] as a strong augmentation method, which slightly affects the performance of the baseline method as shown in Tab. 6, while using horizontal flip as weak augmentation. + +# 3.3. Empirical Evidence + +This subsection empirically confirms our assumption that optimal recall interval improves memory retention, and accordingly, the performance of the continual method is en + +
BufferMethodS-CIFAR-10S-Tiny-ImageNet
CILTILAvg*ΔCILTILAvg*Δ
-Joint92.20±0.1598.31±0.1295.2659.99±0.1982.04±0.1071.02
0LwF-61.98±0.9861.98-15.26±0.7115.26
VBM-LwF-77.53±3.0677.53+15.55-51.21±0.3451.21+35.95
200DER56.58±2.0889.32±1.3272.9511.56±0.5841.67±0.6326.62
ER50.27±0.7091.66±0.7970.978.67±0.2338.03±0.7923.35
VBM-ER52.56±1.5993.58±0.2373.07+2.108.83±0.1344.03±1.1226.43+3.08
iCaRL64.11±1.8690.20±0.8877.1619.09±0.1353.07±0.2736.08
VBM-iCaRL69.73±0.8792.76±0.4581.25+4.0921.39±0.1856.31±0.4238.85+2.77
DER++61.67±1.0290.61±1.1376.1412.48±1.6041.19±0.4826.84
VBM-DER++66.99±2.9894.30±0.2180.65+4.5113.67±0.4144.36±0.8829.02+2.18
500DER65.93±5.3892.52±0.6879.2318.04±1.3052.90±0.9135.47
ER56.37±0.9694.15±0.0475.2610.20±0.3749.13±0.3129.67
VBM-ER63.39±2.7794.20±0.2578.80+3.5411.21±0.3052.01±0.7131.61+1.94
iCaRL62.22±3.4588.86±3.2375.5423.93±0.9758.89±0.8541.41
VBM-iCaRL68.97±0.3893.10±0.6681.04+5.5026.89±0.5462.61±0.1744.75+3.34
DER++69.94±2.0493.49±0.2681.7219.39±1.0951.71±0.3535.55
VBM-DER++75.22±0.8194.81±0.3985.02+3.3019.46±0.4649.80±0.8034.63-0.92
5120DER80.79±0.7594.46±0.3087.6332.92±3.0964.77±2.1448.85
ER74.23±4.1896.86±0.2885.5528.07±0.3767.58±0.3047.83
VBM-ER79.29±2.2997.02±0.1388.16+2.6133.07±0.3668.91±0.5050.99+3.16
iCaRL78.49±0.3894.84±0.8286.6732.16±0.2466.75±0.3649.46
VBM-iCaRL79.40±0.6396.09±0.3287.75+1.0834.40±0.1268.86±0.2651.63+2.17
DER++85.42±0.7496.24±0.1990.8334.06±0.5765.28±1.2049.67
VBM-DER++86.62±0.3897.07±0.0791.85+1.0235.14±0.3667.52±0.4851.33+1.66
+ +Table 1. Experimental results on CIL and TIL protocols. In this evaluation, we report the last top-1 accuracy $(\%)$ on the S-CIFAR-10 and S-Tiny-ImageNet benchmarks with four different buffer sizes. ResNet-18 backbone is used for each continual learning method in this evaluation. *Avg is the averaged accuracy of CIL and TIL. We reproduce baseline methods following Buzzega et al. [3]. We run every method three times with different random seeds for a reliable result. Ours improves performance consistently in most results. The performance of Joint is the upper bound of this experiment, where all old training samples are used in every step. + +
MethodAvgΔLastΔ
Joint-46.48±3.62
AGEM24.88±0.5728.94±0.67
AGEM-R24.72±0.5928.01±0.93
FDR34.62±0.3334.46±0.32
DER39.84±0.1540.46±0.47
ER32.38±2.8832.02±2.39
VBM-ER35.80±0.55+3.4234.72±0.37+2.70
DER++34.56±3.4234.75±2.41
VBM-DER++42.16±0.20+7.2441.81±0.08+7.06
+ +Table 2. Experimental results on DIL protocol. This evaluation shows Avg and Last top-1 accuracy $(\%)$ on the DomainNet benchmark. We utilize ResNet-18 as the backbone network in this evaluation. + +hanced. Specifically, the logical development of our assumption is that 1) if the network sufficiently forgets a sample at the optimal recall interval and then extensively retrans it, 2) memory retention is enhanced as training progresses, and therefore, 3) the accuracy of continual learning is also improved. To evaluate our assumption, we measure 1) the degree of forgetting between a retraining session for + +training samples, 2) the decay speed of the memory retention, and 3) the average accuracy at which networks correctly classify their class labels. Figure 4 shows the results for the empirical evidence of our assumption. As depicted in Figure 4(a), the degree of forgetting becomes larger as the recall interval increases. It is obvious that the longer recall interval allows the network to forget a training sample before retraining it again. Figure 4(b) demonstrates the different levels of memory retention decay due to the recall intervals. The experiment results show that memory decays slowest in the x3 case, which has an interval three times longer than the baseline x1 case. Notably, our method achieves considerably higher performance at the end of the formal learning stage. This result is attributed to the fact that multiple epochs are used in training networks, and the memory retention decay between epochs is slow for the x3 case. Accordingly, x3 with our replay method achieves improved accuracy compared to other recall interval cases, as shown in Figure 4(c). We conduct more experiments to show empirical proofs on different datasets, backbone networks, and buffer sizes in Figure 6. The definition of forgetting is given in the supplementary material. + +
Method5 Step10 Step20 Step
AvgΔLastΔAvgΔLastΔAvgΔLastΔ
Joint80.40-80.41-81.49-
iCaRL71.14±0.3459.7165.27±1.0250.7461.20±0.8343.75
UCIR62.77±0.8247.3158.66±0.7143.3958.17±0.3040.63
BiC73.10±0.5562.1068.80±1.2053.5466.48±0.3247.02
WA72.81±0.2860.8469.46±0.2953.7867.33±0.1547.31
DyTox--73.66±0.0260.6772.27±0.1856.32
DER76.77±0.0168.06±0.0075.72±0.0864.32±0.0574.96±0.0161.80±0.07
VBM-DER78.60±0.23+1.8370.60±0.12+2.5478.12±0.07+2.4067.04±0.11+2.7276.95±0.02+1.9964.29±0.04+2.49
TCIL77.33±0.0869.48±0.1476.33±0.1565.66±0.0274.32±0.0162.54±0.02
VBM-TCIL79.23±0.01+1.9071.23±0.04+1.7578.02±0.03+1.6968.14±0.08+2.4876.83±0.00+2.5167.16±0.09+4.62
+ +Table 3. Experimental results on rehearsal-based CIL protocol. In this evaluation, we provide Avg and Last top-1 accuracy $(\%)$ on the S-CIFAR-100 benchmark with three different class incremental steps. ResNet-18 backbone is utilized to evaluate each rehearsal-based method. This evaluation also shows that ours achieves consistent performance improvements. Results of DyTox are imported from its original work [17]. We reproduce results of DER and TCIL using official implementation, while other results come from Yan et al. [44]. + +
Method5 Step10 Step
AvgΔLastΔAvgΔLastΔ
Joint80.40-80.41-
LwF-MC32.8911.4220.716.36
BiC54.7435.4245.0824.42
LwM39.3116.7025.478.57
WA63.6749.3645.6826.14
DER40.05±0.0017.11±0.0026.67±0.008.71±0.00
VBM-DER40.47±0.00+0.4217.47±0.01+0.3626.95±0.00+0.288.90±0.00+0.19
TCIL64.40±0.1552.37±0.0856.84±0.7540.31±1.04
VBM-TCIL66.40±0.07+2.0054.69±0.02+2.3261.12±0.02+4.2845.92±0.25+5.61
+ +Table 4. Experimental results on non-rehearsal-based CIL protocol. For this evaluation, we measure Avg and Last top-1 accuracy $(\%)$ on the S-CIFAR-100 benchmark with two different class incremental steps. ResNet-18 backbone is applied to existing and our methods. We report original results from LwF-MC and LwM [15]. We reproduce the results of DER and TCIL by ourselves using official implementation. Other results are reported from Huang et al. [24]. + +# 4. Experiments + +This section demonstrates the capabilities of the proposed view-batch model in various continual learning protocols and scenarios through extensive performance evaluation. We perform experiments for three continual learning protocols, such as class, task, and domain incremental learning. Further, we also assess our view-batch model in both rehearsal and non-rehearsal, as well as under the pre-training-based continual learning scenarios. + +# 4.1. Experimental Settings + +Datasets. We evaluate our methods on widely used benchmarks such as S-CIFAR-10/100, S-Tiny-ImageNet, S-ImageNet-R, and DomainNet where the prefix of 'S-' denotes the sequential data configuration. We use the com + +monly adopted class ordering lists, following existing works of Yan et al. [44] and Huang et al. [24]. + +Protocols. In our experiments, we utilize three continual learning protocols. First, we make use of task incremental learning (TIL), which involves training neural networks on a series of tasks in a sequential manner while evaluating their performance on both newly acquired and previously learned tasks. Specifically, the TIL protocol consists of the following detailed configurations: 1) a task is defined by a set of label classes, and 2) we evaluate networks knowing the task identity of test samples. Second, we adopt class incremental learning (CIL), which is identical to TIL except that it evaluates test samples without knowing their task identity. Third, we use domain incremental learning (DIL), which focuses on learning the new samples across multiple sequential domains rather than acquiring new classes. In addition to these three protocols, we also include both rehearsal and non-rehearsal scenarios, divided by the use of a memory buffer during continual learning. For the non-rehearsal scenario, we evaluate both scratch- and pre-training-based learning scenarios to showcase extensive experimental results. For all these protocols and scenarios, we provide the mean and standard deviation of three runs for reliable experimental results. The details of the experiment can be found in the supplementary material. + +Baseline Methods. We apply our view-batch to recent baseline continual learning methods such as iCaRL [32], ER [34], DER [3, 44], TCIL [24], and SLCA [46]. We also include results from A-GEM [10], LwM [15], UCIR [23], BiC [43], WA [47], DyTox [17], L2P [42], S-iPrompt [40], DualPrompt [41], and CODA-Prompt [38]. We follow the baseline's hyper-parameters for a fair comparison between ours and its respective baseline methods. Specifically, we + +
S-CIFAR-100S-ImageNet-R
MethodTop-1Top-5Top-1Top-5
AvgLastAvgLastAvgLastAvgLast
Joint-92.54±0.14-99.24±0.02-82.41±0.02-93.27±0.07
L2P90.42±0.5683.76±0.4298.89±0.0398.08±0.1576.76±0.3373.56±0.4289.39±0.1686.61±0.03
DualPrompt90.17±0.3885.10±0.2398.90±0.0797.68±0.1273.31±0.1969.12±0.1889.19±0.0985.46±0.04
CODA-Prompt91.27±0.6686.86±0.5899.29±0.0998.64±0.0781.89±0.5976.79±0.1793.30±0.3490.49±0.15
S-iPrompt92.19±0.4088.43±0.4299.19±0.0598.63±0.0571.47±0.1468.77±0.2788.44±0.1984.91±0.31
SLCA94.31±1.1291.57±0.4699.64±0.0499.26±0.0383.58±0.6578.81±0.2194.39±0.2692.02±0.25
VBM-SLCA94.52±0.8591.78±0.1899.64±0.0399.29±0.0384.61±0.5980.07±0.1294.86±0.1992.74±0.14
+ +![](images/eee9c2892997f41c6cf1f1f1042a54be340f19d9ebebe328fe2b936eb21a7201.jpg) +(a) iCaRL + +![](images/605ea598910eb3fa5beb3139decaece279767ebcefe705c6cbe8b007418c956c.jpg) +(b) DER++ + +![](images/773d2134c08e3b1f5e4757b2f0bf45e7a6e1b18145a7f44973dfa65198898156.jpg) +Figure 5. Experimental results on memory retention decay. This analysis reports memory retention decay of the first three tasks on the S-CIFAR-10 dataset, comparing the proposed approach against the baseline methods. The green numbers at the end of the last task are the accuracy gain of the proposed approach over the baselines. +(a) Dataset +Figure 6. Experimental results on recall interval. We measure the last top-1 accuracy $(\%)$ ( $\uparrow$ ) in the top row and report forgetting $(\downarrow)$ at the bottom row, varying recall intervals with different learning factors. Dark colors mean better. We use the ResNet-18 backbone. + +![](images/b17898ccb02afcbc02cac32f8a68825dde50800e921a7dc08779ecef470d4109.jpg) +(b) Network + +![](images/93aec7734c01f8373a22f703483d8be15995b2d7280e898d4cfc2b59dc4430a8.jpg) +(c) Buffer + +use ResNet and Vision Transformer (ViT) as backbone networks in traditional and prompt-based approaches. For the traditional approach, we randomly initialize backbone weights at the beginning of every run. On the other hand, for the prompt-based approach, we adopt a frozen backbone whose weights are pre-trained on a large-scale dataset following recent prompt-based CL studies [42, 46] + +Table 5. Experimental results on pre-trained model-based CIL protocol. This evaluation presents Avg and Last top-1,5 accuracy (%) on the S-CIFAR-100 and S-ImageNet-R benchmarks without using a memory buffer. ViT-B/16 backbone is adopted for each pre-trained model-based method. We reproduce baseline methods using official implementation by ourselves, except for L2P. + +
MethodReplayStrong augmentSSLCILTILAvgΔ
iCaRL---64.1190.2077.16-1.05
-✓-62.0490.1776.11
✓--67.5191.8779.69+2.53
✓✓-66.7891.2178.99+1.83
✓✓✓69.7392.7681.25+4.09
DER++---61.6790.6176.14+0.76
-✓-62.6691.1476.90
✓--64.7592.7978.77+2.63
✓✓-64.4493.3878.91+2.77
✓✓✓66.9994.3080.65+4.51
+ +Table 6. Experimental results on factor analysis. We demonstrate the last accuracy (\%) of CIL and TIL on S-CIFAR-10, varying our main components with different augmentation types. We observe that the proposed components, such as replay and SSL, improve performance consistently and confirm that performance enhancement comes from them, not simply from strong augmentation. In the strong augmentation setting without view batch replay, we maintain the same ratio of weak to strong augmentations. + +# 4.2. Experimental Results + +Continual Learning under Three Protocols. We evaluate our method in the three continual learning protocols such as CIL, TIL, and DIL. For TIL and CIL, we use different memory buffer sizes: 0, 200, 500, and 5,120. For DIL, we utilize 50 samples per class, resulting in 17,250 samples in a buffer. We apply our view-batch model to LwF for the non-rehearsal memory buffer (i.e. 0 size) and ER, DER++, and iCaRL for the other buffer sizes. Tabs. 1 and 2 show that our method improves the performance of continual learning methods for all protocols. Moreover, we demonstrate that baseline methods employing our view-batch are comparable to other continual learning methods. + +We conduct further in-depth experiments under the most widely used CIL protocol. We consider three configurations with step sizes of 5, 10, and 20, which are the number of total incremental steps, as shown in Tab. 3. In addition, we + +
MethodCurrent sampleBuffer sampleCILTILAvgΔ
iCaRL--64.1190.2077.16-
✓-64.6494.1279.38+2.22
-✓64.9994.2079.60+2.44
✓✓69.7392.7681.25+4.09
DER++--61.6790.6176.14-
✓-64.8893.2279.05+2.91
-✓62.6994.1578.42+2.28
✓✓66.9994.3080.65+4.51
+ +Table 7. Experimental results on sample type. We measure the last accuracy (\%) of CIL and TIL on S-CIFAR-10, applying our method to different sample types. We use the ResNet-18 as backbone. This analysis shows that our method works well with both sample types. + +take into account the use of a memory buffer in two scenarios, such as rehearsal and non-rehearsal in Tab. 4. Both experimental results demonstrate that our view-batch model consistently improves the performance of continual learning methods under the various step sizes and memory buffer configurations. + +Non-rehearsal Continual Learning. While the nonrehearsal scenario has the advantage of not using memory buffers, they inherently face the drawback of limited performance due to the inability to retrain networks using old data in the memory buffer. To overcome this limitation, many recent studies have focused on non-rehearsal continual learning using a pre-trained model trained with a large-scale dataset. Thus, we evaluate our view-batch model in a scenario where models are pre-trained, such as SCLA. Experimental results in Tab. 5 indicate that the proposed view-batch model is effective across all benchmarks for this scenario. + +# 5. Analysis + +We present the analysis of the proposed view-batch model. First, we validate that the optimal recall interval mitigates catastrophic forgetting of early tasks. Figure 5 visualize that the recall interval x3 maintains long-term memory retention resolving catastrophic forgetting of initial tasks. Also, Figure 6 analyzes how different factors (i.e., dataset, network scales, memory buffer size) affect the optimal recall intervals in terms of last top-1 accuracy and forgetting metrics [9, 10, 42]. We find that the recall interval x3 or x4 improves the accuracy and resolves catastrophic forgetting in various scenarios. More experimental results of different methods and task types can be found in the supplementary material. Tab. 6 shows that two main factors of the proposed method consistently improve their respective baselines, identifying that the performance improvement comes from the proposed components, not strong augmentation. + +![](images/9feb0905c07c53017f273a2c620e2845a4f8ffe3bd19ed9627d481e09059eeaa.jpg) +Figure 7. Experimental results on accuracy and forgetting. This analysis shows (a) the last top-1 accuracy $(\%)$ and (b) corresponding forgetting measures on the S-CIFAR-10 dataset with different buffer sizes. Our method consistently improves accuracy and forgetting measures. + +Lastly, in the rehearsal scenario, we have two different sample types, current- and buffer-based samples. Tab. 7 exhibit that ours performs well with both types. + +We examine the proposed method in terms of the last accuracy and forgetting measurement. We apply the proposed method to two different baseline approaches. We compute average forgetting at $T$ th task as $\frac{1}{T - 1}\sum_{t = 1}^{T - 1}\max_{j\in \{t,\dots ,T - 1\}}a_j^t -a_T^t$ where $a_j^t$ means $t$ th task accuracy at $j$ step. Notably, higher forgetting means severe catastrophic forgetting of previous tasks. Figure 7 shows that the view-batch model improves accuracy while addressing the forgetting problem simultaneously. + +# 6. Conclusion + +We propose the view-batch model, which is generally applicable to various continual learning scenarios. Inspired by Ebbinghaus's forgetting curve theory, the proposed view-batch model optimizes the recall interval between retraining samples, improving neural networks' long-term memory. The proposed method consists of two main components: replay and self-supervised learning. The replay optimizes the recall interval for guaranteeing sufficient memory forgetting by grouping multiple views of a sample. Also, the self-supervised learning ensures extensive learning for the multiple views using the one-to-many divergence loss. Experimental results show that sufficient forgetting preserves long-term memory and ultimately improves continual learning accuracy. We hope the proposed approach to continual learning can boost future research. + +Acknowledgment. This paper was supported in part by the ETRI Grant funded by Korean Government (Fundamental Technology Research for Human-Centric Autonomous Intelligent Systems) under Grant 24ZB1200, Artificial Intelligence Convergence Innovation Human Resources Development (IITP-2025-RS-2023-00255968), Artificial Intelligence Innovation Hub (RS-2021-II212068), and the NRF Grant (RS-2024-00356486). + +# References + +[1] Hadi Amiri, Timothy Miller, and Guergana Savova. Repeat before forgetting: Spaced repetition for efficient and effective training of neural networks. In Conference on Empirical Methods in Natural Language Processing, 2017. 1, 2 +[2] Maxim Berman, Hervé Jégou, Andrea Vedaldi, Iasonas Kokkinos, and Matthijs Douze. Multigrain: a unified image embedding for classes and instances. arXiv:1902.05509, 2019.3 +[3] Pietro Buzzega, Matteo Boschini, Angelo Porrello, Davide Abati, and Simone Calderara. Dark experience for general continual learning: a strong, simple baseline. Advances in Neural Information Processing Systems, 2020. 5, 6 +[4] Boxi Cao, Qiaoyu Tang, Hongyu Lin, Shanshan Jiang, Bin Dong, Xianpei Han, Jiawei Chen, Tianshu Wang, and Le Sun. Retentive or forgetful? diving into the knowledge memorizing mechanism of language models. In Joint International Conference on Computational Linguistics, Language Resources and Evaluation, 2024. 1, 2 +[5] Mathilde Caron, Hugo Touvron, Ishan Misra, Herve Jégou, Julien Mairal, Piotr Bojanowski, and Armand Joulin. Emerging properties in self-supervised vision transformers. In International Conference on Computer Vision, 2021. 3, 4 +[6] Nicholas J Cepeda, Harold Pashler, Edward Vul, John T Wixted, and Doug Rohrer. Distributed practice in verbal recall tasks: A review and quantitative synthesis. *Psychological bulletin*, 2006. 1 +[7] Nicholas J Cepeda, Edward Vul, Doug Rohrer, John T Wixted, and Harold Pashler. Spacing effects in learning: A temporal ridgeline of optimal retention. Psychological science, 2008. 1 +[8] Hyuntak Cha, Jaeho Lee, and Jinwoo Shin. Co2l: Contrastive continual learning. In International Conference on Computer Vision, 2021. 4 +[9] Arslan Chaudhry, Puneet K Dokania, Thalaiyasingam Ajthan, and Philip HS Torr. Riemannian walk for incremental learning: Understanding forgetting and intransigence. In European Conference on Computer Vision, 2018. 8 +[10] Arslan Chaudhry, Marc'Aurelio Ranzato, Marcus Rohrbach, and Mohamed Elhoseiny. Efficient lifelong learning with a gem. In International Conference on Learning Representations, 2019. 6, 8 +[11] Ting Chen, Simon Kornblith, Mohammad Norouzi, and Geoffrey Hinton. A simple framework for contrastive learning of visual representations. In International Conference on Machine Learning, 2020. 3, 4 +[12] Haoyang Cheng, Haitao Wen, Heqian Qiu, Lanxiao Wang, Minjian Zhang, and Hongliang Li. Must unsupervised continual learning relies on previous information? In Conference on Computer Vision and Pattern Recognition, 2024. 4 +[13] Ekin D Cubuk, Barret Zoph, Dandelion Mane, Vijay Vasudevan, and Quoc V Le. Autoaugment: Learning augmentation policies from data. arXiv:1805.09501, 2018. 4 +[14] Ekin D Cubuk, Barret Zoph, Jonathon Shlens, and Quoc V Le. Randaugment: Practical automated data augmentation with a reduced search space. In Conference on Computer Vision and Pattern Recognition Workshops, 2020. 2 + +[15] Prithviraj Dhar, Rajat Vikram Singh, Kuan-Chuan Peng, Ziyan Wu, and Rama Chellappa. Learning without memorizing. In Conference on Computer Vision and Pattern Recognition, 2019. 3, 6 +[16] Arthur Douillard, Matthieu Cord, Charles Ollion, Thomas Robert, and Eduardo Valle. Podnet: Pooled outputs distillation for small-tasks incremental learning. In European Conference on Computer Vision, 2020. 3 +[17] Arthur Douillard, Alexandre Ramé, Guillaume Couairon, and Matthieu Cord. Dytox: Transformers for continual learning with dynamic token expansion. In Conference on Computer Vision and Pattern Recognition, 2022. 6 +[18] Hermann Ebbinghaus. Memory: A contribution to experimental psychology. Annals of neurosciences, 2013. 1, 2 +[19] Kate Gordon. Class results with spaced and unspaced memorizing. Journal of Experimental Psychology, 1925. 1 +[20] Jean-Bastien Grill, Florian Strub, Florent Altché, Corentin Tallec, Pierre Richemond, Elena Buchatskaya, Carl Doersch, Bernardo Avila Pires, Zhaohan Guo, Mohammad Gheshlaghi Azar, et al. Bootstrap your own latent-a new approach to self-supervised learning. Advances in Neural Information Processing Systems, 2020. 4 +[21] Karri S Hawley, Katie E Cherry, Emily O Boudreaux, and Erin M Jackson. A comparison of adjusted spaced retrieval versus a uniform expanded retrieval schedule for learning a name-face association in older adults with probable Alzheimer's disease. Journal of Clinical and Experimental Neuropsychology, 2008. 1 +[22] Elad Hoffer, Tal Ben-Nun, Itay Hubara, Niv Giladi, Torsten Hoefler, and Daniel Soudry. Augment your batch: better training with larger batches. arXiv:1901.09335, 2019. 3 +[23] Saihui Hou, Xinyu Pan, Chen Change Loy, Zilei Wang, and Dahua Lin. Learning a unified classifier incrementally via rebalancing. In Conference on Computer Vision and Pattern Recognition, 2019. 6 +[24] Bingchen Huang, Zhineng Chen, Peng Zhou, Jiayin Chen, and Zuxuan Wu. Resolving task confusion in dynamic expansion architectures for class incremental learning. In Association for the Advancement of Artificial Intelligence, 2023. 2, 3, 6 +[25] Anette Hunziker, Yuxin Chen, Oisin Mac Aodha, Manuel Gomez Rodriguez, Andreas Krause, Pietro Perona, Yisong Yue, and Adish Singla. Teaching multiple concepts to a forgetful learner. Advances in Neural Information Processing Systems, 2019. 2 +[26] Szabolcs Káli and Peter Dayan. Replay, repair and consolidation. Advances in Neural Information Processing Systems, 2002. 2 +[27] Sean HK Kang. Spaced repetition promotes efficient and effective learning: Policy implications for instruction. *Policy Insights from the Behavioral and Brain Sciences*, 2016. 1 +[28] Jeffrey D Karpicke and Henry L Roediger III. Expanding retrieval practice promotes short-term retention, but equally spaced retrieval enhances long-term retention. Journal of experimental psychology: learning, memory, and cognition, 2007. 1 + +[29] Zhizhong Li and Derek Hoiem. Learning without forgetting. IEEE transactions on pattern analysis and machine intelligence, 2017. 3 +[30] Cecil Alec Mace. The psychology of study. McBride, 1932. 1 +[31] Arthur W Melton. The situation with respect to the spacing of repetitions and memory. Journal of Verbal Learning and Verbal Behavior, 1970. 1 +[32] Sylvestre-Alvise Rebuffi, Alexander Kolesnikov, Georg Sperl, and Christoph H Lampert. icarl: Incremental classifier and representation learning. In Conference on Computer Vision and Pattern Recognition, 2017. 2, 3, 6 +[33] Matthew Riemer, Ignacio Cases, Robert Ajemian, Miao Liu, Irina Rish, Yuhai Tu, , and Gerald Tesauro. Learning to learn without forgetting by maximizing transfer and minimizing interference. In International Conference on Learning Representations, 2019. 3 +[34] David Rolnick, Arun Ahuja, Jonathan Schwarz, Timothy Lillicrap, and Gregory Wayne. Experience replay for continual learning. Advances in Neural Information Processing Systems, 2019. 6 +[35] Joan Serra, Didac Suris, Marius Miron, and Alexandros Karatzoglou. Overcoming catastrophic forgetting with hard attention to the task. In International Conference on Machine Learning, 2018. 3 +[36] John J Shaughnessy. Long-term retention and the spacing effect in free-recall and frequency judgments. The American Journal of Psychology, 1977. 1 +[37] Hanul Shin, Jung Kwon Lee, Jaehong Kim, and Jiwon Kim. Continual learning with deep generative replay. Advances in Neural Information Processing Systems, 2017. 3 +[38] James Seale Smith, Leonid Karlinsky, Vyshnavi Gutta, Paola Cascante-Bonilla, Donghyun Kim, Assaf Arbelle, Rameswar Panda, Rogerio Feris, and Zsolt Kira. Coda-prompt: Continual decomposed attention-based prompting for rehearsal-free continual learning. In Conference on Computer Vision and Pattern Recognition, 2023. 6 +[39] Kushal Tirumala, Aram Markosyan, Luke Zettlemoyer, and Armen Aghajanyan. Memorization without overfitting: Analyzing the training dynamics of large language models. Advances in Neural Information Processing Systems, 2022. 2 +[40] Yabin Wang, Zhiwu Huang, and Xiaopeng Hong. S-prompts learning with pre-trained transformers: An occam's razor for domain incremental learning. Advances in Neural Information Processing Systems, 2022. 6 +[41] Zifeng Wang, Zizhao Zhang, Sayna Ebrahimi, Ruoxi Sun, Han Zhang, Chen-Yu Lee, Xiaoqi Ren, Guolong Su, Vincent Perot, Jennifer Dy, et al. Dualprompt: Complementary prompting for rehearsal-free continual learning. In European Conference on Computer Vision, 2022. 3, 6 +[42] Zifeng Wang, Zizhao Zhang, Chen-Yu Lee, Han Zhang, Ruoxi Sun, Xiaqi Ren, Guolong Su, Vincent Perot, Jennifer Dy, and Tomas Pfister. Learning to prompt for continual learning. In Conference on Computer Vision and Pattern Recognition, 2022. 3, 6, 7, 8 +[43] Yue Wu, Yinpeng Chen, Lijuan Wang, Yuancheng Ye, Zicheng Liu, Yandong Guo, and Yun Fu. Large scale in + +cremental learning. In Conference on Computer Vision and Pattern Recognition, 2019. 6 +[44] Shipeng Yan, Jiangwei Xie, and Xuming He. Der: Dynamically expandable representation for class incremental learning. In Conference on Computer Vision and Pattern Recognition, 2021. 3, 6 +[45] Sangdoo Yun, Dongyoon Han, Seong Joon Oh, Sanghyuk Chun, Junsuk Choe, and Youngjoon Yoo. Cutmix: Regularization strategy to train strong classifiers with localizable features. In International Conference on Computer Vision, 2019. 2 +[46] Gengwei Zhang, Liyuan Wang, Guoliang Kang, Ling Chen, and Yunchao Wei. Slca: Slow learner with classifier alignment for continual learning on a pre-trained model. In International Conference on Computer Vision, 2023. 3, 6, 7 +[47] Bowen Zhao, Xi Xiao, Guojun Gan, Bin Zhang, and ShuTao Xia. Maintaining discrimination and fairness in class incremental learning. In Conference on Computer Vision and Pattern Recognition, 2020. 6 +[48] Wanjun Zhong, Lianghong Guo, Qiqi Gao, He Ye, and Yanlin Wang. Memorybank: Enhancing large language models with long-term memory. In Association for the Advancement of Artificial Intelligence, 2024. 1, 2 \ No newline at end of file diff --git a/2025/Do Your Best and Get Enough Rest for Continual Learning/images.zip b/2025/Do Your Best and Get Enough Rest for Continual Learning/images.zip new file mode 100644 index 0000000000000000000000000000000000000000..d37de9f7ab90bb64b5ccaa4f8f7864b07ad61660 --- /dev/null +++ b/2025/Do Your Best and Get Enough Rest for Continual Learning/images.zip @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:96028a935949b8bc19c1491f37d38bae7d73c27b9143483b7b4229522cd6dd2f +size 710041 diff --git a/2025/Do Your Best and Get Enough Rest for Continual Learning/layout.json b/2025/Do Your Best and Get Enough Rest for Continual Learning/layout.json new file mode 100644 index 0000000000000000000000000000000000000000..d004e6f91fe26a940187b55ca2880bf81c4a870c --- /dev/null +++ b/2025/Do Your Best and Get Enough Rest for Continual Learning/layout.json @@ -0,0 +1,7631 @@ +{ + "pdf_info": [ + { + "para_blocks": [ + { + "bbox": [ + 121, + 103, + 489, + 121 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 121, + 103, + 489, + 121 + ], + "spans": [ + { + "bbox": [ + 121, + 103, + 489, + 121 + ], + "type": "text", + "content": "Do Your Best and Get Enough Rest for Continual Learning" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 110, + 142, + 187, + 159 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 110, + 142, + 187, + 159 + ], + "spans": [ + { + "bbox": [ + 110, + 142, + 187, + 159 + ], + "type": "text", + "content": "Hankyul Kang" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 217, + 143, + 290, + 158 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 217, + 143, + 290, + 158 + ], + "spans": [ + { + "bbox": [ + 217, + 143, + 290, + 158 + ], + "type": "text", + "content": "Gregor Seifer1" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 320, + 143, + 399, + 158 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 320, + 143, + 399, + 158 + ], + "spans": [ + { + "bbox": [ + 320, + 143, + 399, + 158 + ], + "type": "text", + "content": "Donghyun Lee²" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 428, + 143, + 503, + 158 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 428, + 143, + 503, + 158 + ], + "spans": [ + { + "bbox": [ + 428, + 143, + 503, + 158 + ], + "type": "text", + "content": "Jongbin " + }, + { + "bbox": [ + 428, + 143, + 503, + 158 + ], + "type": "inline_equation", + "content": "\\mathrm{Ryu}^{1}" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 234, + 170, + 373, + 186 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 234, + 170, + 373, + 186 + ], + "spans": [ + { + "bbox": [ + 234, + 170, + 373, + 186 + ], + "type": "inline_equation", + "content": "^{1}" + }, + { + "bbox": [ + 234, + 170, + 373, + 186 + ], + "type": "text", + "content": " Ajou University " + }, + { + "bbox": [ + 234, + 170, + 373, + 186 + ], + "type": "inline_equation", + "content": "^{2}" + }, + { + "bbox": [ + 234, + 170, + 373, + 186 + ], + "type": "text", + "content": " KAIST" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 152, + 213, + 200, + 225 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 152, + 213, + 200, + 225 + ], + "spans": [ + { + "bbox": [ + 152, + 213, + 200, + 225 + ], + "type": "text", + "content": "Abstract" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 55, + 239, + 296, + 586 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 239, + 296, + 586 + ], + "spans": [ + { + "bbox": [ + 55, + 239, + 296, + 586 + ], + "type": "text", + "content": "According to the forgetting curve theory, we can enhance memory retention by learning extensive data and taking adequate rest. This means that in order to effectively retain new knowledge, it is essential to learn it thoroughly and ensure sufficient rest so that our brain can memorize without forgetting. The main takeaway from this theory is that learning extensive data at once necessitates sufficient rest before learning the same data again. This aspect of human long-term memory retention can be effectively utilized to address the continual learning of neural networks. Retaining new knowledge for a long period of time without catastrophic forgetting is the critical problem of continual learning. Therefore, based on Ebbinghaus' theory, we introduce the view-batch model that adjusts the learning schedules to optimize the recall interval between retraining the same samples. The proposed view-batch model allows the network to get enough rest to learn extensive knowledge from the same samples with a recall interval of sufficient length. To this end, we specifically present two approaches: 1) a replay method that guarantees the optimal recall interval, and 2) a self-supervised learning that acquires extensive knowledge from a single training sample at a time. We empirically show that these approaches of our method are aligned with the forgetting curve theory, which can enhance long-term memory. In our experiments, we also demonstrate that our method significantly improves many state-of-the-art continual learning methods in various protocols and scenarios. We open-source this project at https://github.com/hankyul2/ViewBatchModel." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 56, + 611, + 135, + 624 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 611, + 135, + 624 + ], + "spans": [ + { + "bbox": [ + 56, + 611, + 135, + 624 + ], + "type": "text", + "content": "1. Introduction" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 55, + 632, + 296, + 693 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 632, + 296, + 693 + ], + "spans": [ + { + "bbox": [ + 55, + 632, + 296, + 693 + ], + "type": "text", + "content": "The forgetting curve theory [18] suggests that human memory tends to fade as time goes on; however, memory retention can be improved by repeated learning in an optimal recall interval. Motivated by this theory, many subsequent studies [6, 19, 27] have shown that spaced repetition with" + } + ] + } + ], + "index": 11 + }, + { + "type": "image", + "bbox": [ + 320, + 211, + 433, + 290 + ], + "blocks": [ + { + "bbox": [ + 320, + 211, + 433, + 290 + ], + "lines": [ + { + "bbox": [ + 320, + 211, + 433, + 290 + ], + "spans": [ + { + "bbox": [ + 320, + 211, + 433, + 290 + ], + "type": "image", + "image_path": "c3ac23cdbc9dbc772bc852352bfd6acc3d7e32ca4dbf55f6548953e05a3cd2ad.jpg" + } + ] + } + ], + "index": 12, + "angle": 0, + "type": "image_body" + } + ], + "index": 12 + }, + { + "type": "image", + "bbox": [ + 436, + 212, + 548, + 290 + ], + "blocks": [ + { + "bbox": [ + 436, + 212, + 548, + 290 + ], + "lines": [ + { + "bbox": [ + 436, + 212, + 548, + 290 + ], + "spans": [ + { + "bbox": [ + 436, + 212, + 548, + 290 + ], + "type": "image", + "image_path": "4062f74cdf243453891fc09a9f280d166d1fddbdbcf4b553b9573e7ffb925912.jpg" + } + ] + } + ], + "index": 14, + "angle": 0, + "type": "image_body" + } + ], + "index": 14 + }, + { + "type": "image", + "bbox": [ + 320, + 301, + 433, + 380 + ], + "blocks": [ + { + "bbox": [ + 329, + 293, + 422, + 301 + ], + "lines": [ + { + "bbox": [ + 329, + 293, + 422, + 301 + ], + "spans": [ + { + "bbox": [ + 329, + 293, + 422, + 301 + ], + "type": "text", + "content": "(a) Short-term recall interval" + } + ] + } + ], + "index": 13, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 320, + 301, + 433, + 380 + ], + "lines": [ + { + "bbox": [ + 320, + 301, + 433, + 380 + ], + "spans": [ + { + "bbox": [ + 320, + 301, + 433, + 380 + ], + "type": "image", + "image_path": "d39edceeb4ff7ce1f8040457f82497e9700f7cbc5b7b56458a10444e36c27124.jpg" + } + ] + } + ], + "index": 16, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 329, + 383, + 422, + 392 + ], + "lines": [ + { + "bbox": [ + 329, + 383, + 422, + 392 + ], + "spans": [ + { + "bbox": [ + 329, + 383, + 422, + 392 + ], + "type": "text", + "content": "(c) Long-term recall interval" + } + ] + } + ], + "index": 17, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 313, + 402, + 555, + 491 + ], + "lines": [ + { + "bbox": [ + 313, + 402, + 555, + 491 + ], + "spans": [ + { + "bbox": [ + 313, + 402, + 555, + 491 + ], + "type": "text", + "content": "Figure 1. Conceptual graph of the forgetting curve. We show (a) short-term recall interval, (b) optimal recall interval, (c) long-term recall interval, and (d) degree of forgetting. (a-b) Expanding the recall interval improves long-term memory retention of neural networks by repeatedly recalling memory with moderate difficulty, whereas (c) an excessive recall interval decreases it. The depicted forgetting curve regarding recall interval is based on the spacing effect formula [6, 7] provided in the supplementary material." + } + ] + } + ], + "index": 20, + "angle": 0, + "type": "image_caption" + } + ], + "index": 16 + }, + { + "type": "image", + "bbox": [ + 436, + 301, + 548, + 380 + ], + "blocks": [ + { + "bbox": [ + 449, + 293, + 533, + 301 + ], + "lines": [ + { + "bbox": [ + 449, + 293, + 533, + 301 + ], + "spans": [ + { + "bbox": [ + 449, + 293, + 533, + 301 + ], + "type": "text", + "content": "(b) Optimal recall interval" + } + ] + } + ], + "index": 15, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 436, + 301, + 548, + 380 + ], + "lines": [ + { + "bbox": [ + 436, + 301, + 548, + 380 + ], + "spans": [ + { + "bbox": [ + 436, + 301, + 548, + 380 + ], + "type": "image", + "image_path": "d15d8cc7dc22dc911ab44f2ec4a2ad180ba64b5fb19bc82f43eda0ae0dfc6e57.jpg" + } + ] + } + ], + "index": 18, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 452, + 383, + 529, + 392 + ], + "lines": [ + { + "bbox": [ + 452, + 383, + 529, + 392 + ], + "spans": [ + { + "bbox": [ + 452, + 383, + 529, + 392 + ], + "type": "text", + "content": "(d) Degree of forgetting" + } + ] + } + ], + "index": 19, + "angle": 0, + "type": "image_caption" + } + ], + "index": 18 + }, + { + "bbox": [ + 313, + 518, + 555, + 626 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 518, + 555, + 626 + ], + "spans": [ + { + "bbox": [ + 313, + 518, + 555, + 626 + ], + "type": "text", + "content": "optimal recall interval enhances long-term memory retention in the human brain. This phenomenon is known as the spacing effect [7], where the optimal recall interval for repeated learning is crucial to maintain an acceptable level of forgetting knowledge. Several studies [7, 21, 36] have made significant progress on the spacing effect, shedding light on its effectiveness. There have also been studies [28, 30, 31] highlighting the importance of the recall interval of repeated learning." + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 313, + 630, + 556, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 630, + 556, + 715 + ], + "spans": [ + { + "bbox": [ + 313, + 630, + 556, + 715 + ], + "type": "text", + "content": "Although this theory has been widely recognized in human brain research, it is currently very little known in the field of machine learning. Only a few works [1, 4, 48] show this theory's potential in neural networks by utilizing it as an analysis toolkit or dynamic learning strategy. So, we investigate the impact of the forgetting curve theory on long-term memory retention in neural networks for continual learning." + } + ] + } + ], + "index": 22 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "spans": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "type": "text", + "content": "CVF" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 145, + 0, + 496, + 37 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 145, + 0, + 496, + 37 + ], + "spans": [ + { + "bbox": [ + 145, + 0, + 496, + 37 + ], + "type": "text", + "content": "This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 67, + 703, + 144, + 712 + ], + "type": "page_footnote", + "angle": 0, + "lines": [ + { + "bbox": [ + 67, + 703, + 144, + 712 + ], + "spans": [ + { + "bbox": [ + 67, + 703, + 144, + 712 + ], + "type": "text", + "content": "*Corresponding author." + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "text", + "content": "10077" + } + ] + } + ], + "index": 24 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 0 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 56, + 71, + 293, + 200 + ], + "blocks": [ + { + "bbox": [ + 56, + 71, + 293, + 200 + ], + "lines": [ + { + "bbox": [ + 56, + 71, + 293, + 200 + ], + "spans": [ + { + "bbox": [ + 56, + 71, + 293, + 200 + ], + "type": "image", + "image_path": "171ede31f5cab52ada7526c381f2dc6f5e9d81195679eeec7e8ea1f90a340a1d.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 55, + 208, + 295, + 288 + ], + "lines": [ + { + "bbox": [ + 55, + 208, + 295, + 288 + ], + "spans": [ + { + "bbox": [ + 55, + 208, + 295, + 288 + ], + "type": "text", + "content": "Figure 2. Overview of experimental results. We provide comprehensive comparisons of various factors for continual learning. We perform extensive experiments on step and buffer sizes (a-b), three different continual learning methods (c), whether to use the pre-trained model (d), three different benchmarks (e), and two evaluation protocols (f). In all cases, ours improves the baseline performance consistently." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "bbox": [ + 54, + 307, + 295, + 510 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 54, + 307, + 295, + 510 + ], + "spans": [ + { + "bbox": [ + 54, + 307, + 295, + 510 + ], + "type": "text", + "content": "In Ebbinghaus' theory, we focus on the decay of memory retention in the forgetting curve as shown in Figure 1. The forgetting curve suggests that, as shown in Figure 1(a), the sample memory shows extreme knowledge decay when repeating its learning with a short time-space. On the other hand, as shown in Figure 1(c), if there is an excessive recall interval of repeated learning, too much memory is forgotten, leading to a decreased level of memory retention when relearning the same sample. However, as shown in Figure 1(b), when there is an optimal recall interval between repeated learning, the decay of memory retention also becomes gentle. This effect of the optimal recall interval implies that it is imperative to implement an appropriate learning schedule to retain memories efficiently in the repeated learning process. The notion behind this spacing effect can be usefully applied in continual learning, particularly to prevent catastrophic forgetting." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 55, + 510, + 295, + 556 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 510, + 295, + 556 + ], + "spans": [ + { + "bbox": [ + 55, + 510, + 295, + 556 + ], + "type": "text", + "content": "However, empirically, we found that the current methods for continual learning have a short recall interval [24, 32], so they are limited to profit from the spacing effect, which requires a sufficient recall interval." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 55, + 558, + 295, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 558, + 295, + 714 + ], + "spans": [ + { + "bbox": [ + 55, + 558, + 295, + 714 + ], + "type": "text", + "content": "To tackle this limitation, we introduce two approaches for training neural networks. Our first approach aims to prevent short recall intervals through the proposed replay method. We replay view-batch in the repeated training schedule so that short recall intervals can be delayed sufficiently. We augment multiple views from a single sample, and thus, the recall intervals of repeated learning can be adjusted. In our second approach, we make use of self-supervised learning to compensate for the delayed recall interval by training each sample extensively. We leverage the one-to-many divergence to learn the generalized self-supervised features within a view-batch, which enhances the capability of the networks. These two approaches en" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 313, + 72, + 553, + 239 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 72, + 553, + 239 + ], + "spans": [ + { + "bbox": [ + 313, + 72, + 553, + 239 + ], + "type": "text", + "content": "sure that the neural networks have enough time to retrain the same samples, and at the same time, they can acquire substantial knowledge from training samples at a time. We apply our view-batch model (VBM) to both rehearsal and rehearsal-free continual learning scenarios, where addressing long-term memory retention is critical. Moreover, we show extensive experimental results of various continual protocols and scenarios. As an overview, Figure 2 showcases the comprehensive comparison against different step sizes, memory buffer size, baseline methods, use of the pretrained model, benchmarks, and protocols. For all these comparisons, our view-batch model achieves significant performance improvements. We have condensed our contributions into the following points:" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 314, + 240, + 553, + 430 + ], + "type": "list", + "angle": 0, + "index": 9, + "blocks": [ + { + "bbox": [ + 314, + 240, + 553, + 312 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 314, + 240, + 553, + 312 + ], + "spans": [ + { + "bbox": [ + 314, + 240, + 553, + 312 + ], + "type": "text", + "content": "- We propose the view-batch model that optimizes the recall interval between retraining samples in order to enhance memory decay in the continual learning task. The proposed view-batch model includes replay and self-supervised learning components to accomplish extensive learning in optimized intervals." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 314, + 312, + 553, + 382 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 314, + 312, + 553, + 382 + ], + "spans": [ + { + "bbox": [ + 314, + 312, + 553, + 382 + ], + "type": "text", + "content": "- We showcase the effectiveness of our approach on various state-of-the-art continual learning methods, where ours consistently improves the performance. The proposed method is a drop-in replacement approach, so it does not require any extra computational costs for the performance improvements." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 314, + 384, + 553, + 430 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 314, + 384, + 553, + 430 + ], + "spans": [ + { + "bbox": [ + 314, + 384, + 553, + 430 + ], + "type": "text", + "content": "- We reveal that enhancing memory decay can significantly improve the performance of the continual learning. This confirms the practical utility of theoretical knowledge for neural networks." + } + ] + } + ], + "index": 8 + } + ], + "sub_type": "text" + }, + { + "bbox": [ + 314, + 443, + 400, + 455 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 314, + 443, + 400, + 455 + ], + "spans": [ + { + "bbox": [ + 314, + 443, + 400, + 455 + ], + "type": "text", + "content": "2. Related Work" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 313, + 464, + 549, + 476 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 464, + 549, + 476 + ], + "spans": [ + { + "bbox": [ + 313, + 464, + 549, + 476 + ], + "type": "text", + "content": "2.1. Forgetting Curve Theory in Neural Networks" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 313, + 482, + 555, + 650 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 482, + 555, + 650 + ], + "spans": [ + { + "bbox": [ + 313, + 482, + 555, + 650 + ], + "type": "text", + "content": "Several studies [25, 26] have employed the forgetting curve theory [18] to gain insights into the inherent memory retention of neural networks. Recently, Tirumala et al. [39] investigated the forgetting curve theory in neural networks regarding different model scales. Cao et al. [4] have also utilized the forgetting curve theory to examine the variations in memory conformation of the neural networks as influenced by pre-training methods. Additionally, Amiri et al. [1] proposed a dynamic learning scheduler to reduce total learning cost, leveraging optimal learning intervals corresponding to the forgetting curve. Zhong et al. [48] optimized memory retrieval timing based on the principles of the spacing effect theory, enabling the retention of the previous dialogue context existing over an extended duration." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 313, + 658, + 429, + 671 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 658, + 429, + 671 + ], + "spans": [ + { + "bbox": [ + 313, + 658, + 429, + 671 + ], + "type": "text", + "content": "2.2. Data Augmentation" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 313, + 677, + 553, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 677, + 553, + 713 + ], + "spans": [ + { + "bbox": [ + 313, + 677, + 553, + 713 + ], + "type": "text", + "content": "Data augmentation [14, 45] has been widely used in various existing networks. It helps networks learn rich visual information using a limited number of training samples." + } + ] + } + ], + "index": 14 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "text", + "content": "10078" + } + ] + } + ], + "index": 15 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 1 + }, + { + "para_blocks": [ + { + "bbox": [ + 55, + 72, + 296, + 251 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 72, + 296, + 251 + ], + "spans": [ + { + "bbox": [ + 55, + 72, + 296, + 251 + ], + "type": "text", + "content": "Data augmentation is also effective in self-supervised learning [5, 11]. Unlimited labels can be generated when different augmentation methods are applied to a single sample. Therefore, the network is able to learn with self-supervised labels of augmented samples. We use this augmentation method to construct multiple views in a view-batch and also learn self-supervision along with the supervised information of the continual learning process. There have been several studies [2, 22] on repeated augmentation, which allows learning multiple views from a single sample at a time. However, these studies only focused on improving the performance of the augmentation method and did not examine the effect of recall intervals. In addition, they do not apply self-supervised learning and only use label supervision to train networks." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 55, + 272, + 172, + 285 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 272, + 172, + 285 + ], + "spans": [ + { + "bbox": [ + 55, + 272, + 172, + 285 + ], + "type": "text", + "content": "2.3. Continual Learning" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 55, + 294, + 296, + 547 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 294, + 296, + 547 + ], + "spans": [ + { + "bbox": [ + 55, + 294, + 296, + 547 + ], + "type": "text", + "content": "The main problem of continual learning is preserving old knowledge while acquiring new one. This problem has been recognized as catastrophic forgetting, and one common strategy used to address this problem is to make use of a memory buffer [32]. The approaches to this strategy can be divided into two categories: one involves the direct storage of a subset of training samples from a previously learned task [32, 33], while the other generates training samples from the learned task [37]. This approach of utilizing memory buffers proves to be highly effective in addressing the issue of catastrophic forgetting; however, it does come with the trade-off of requiring additional memory space. Consequently, non-rehearsal continual learning methods [15, 33] have been widely studied to avoid additional memory usage. In addition, studies [41, 42, 46] using pre-trained models in non-rehearsal continual learning have been actively conducted recently. Meanwhile, there have been several works [16, 29] that tackle continual learning from the perspective of the regularization method. The architecture-based method [35, 44] has been another prevalent research direction for addressing continual learning." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 55, + 569, + 111, + 582 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 569, + 111, + 582 + ], + "spans": [ + { + "bbox": [ + 55, + 569, + 111, + 582 + ], + "type": "text", + "content": "3. Method" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 55, + 594, + 296, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 594, + 296, + 713 + ], + "spans": [ + { + "bbox": [ + 55, + 594, + 296, + 713 + ], + "type": "text", + "content": "We introduce a view-batch model designed to enhance the long-term memory retention of neural networks for continual learning (CL) tasks. In existing CL methods [24, 32, 44], the neural networks retrain the same samples with short time intervals, which prohibits long-term memory retention. To address this short recall interval, we first investigate the augmentation model with the replay method to ensure a sufficient recall interval in Sec. 3.1. Then, Sec. 3.2 introduces view-batch learning model to extensively train multiple samples of the same instance in view-batch." + } + ] + } + ], + "index": 4 + }, + { + "type": "image", + "bbox": [ + 319, + 69, + 432, + 147 + ], + "blocks": [ + { + "bbox": [ + 319, + 69, + 432, + 147 + ], + "lines": [ + { + "bbox": [ + 319, + 69, + 432, + 147 + ], + "spans": [ + { + "bbox": [ + 319, + 69, + 432, + 147 + ], + "type": "image", + "image_path": "a0443f4ba890a4c2b249ce19436e98d93600314e1dc1fcc19e74591c0ea53809.jpg" + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 356, + 148, + 395, + 158 + ], + "lines": [ + { + "bbox": [ + 356, + 148, + 395, + 158 + ], + "spans": [ + { + "bbox": [ + 356, + 148, + 395, + 158 + ], + "type": "text", + "content": "(a) Baseline" + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "image_caption" + } + ], + "index": 5 + }, + { + "type": "image", + "bbox": [ + 434, + 70, + 548, + 147 + ], + "blocks": [ + { + "bbox": [ + 434, + 70, + 548, + 147 + ], + "lines": [ + { + "bbox": [ + 434, + 70, + 548, + 147 + ], + "spans": [ + { + "bbox": [ + 434, + 70, + 548, + 147 + ], + "type": "image", + "image_path": "61f95c7408d3f3bb330b735292fca7910ba239dbbf172835f3bbbb95a65e5d29.jpg" + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 466, + 148, + 515, + 157 + ], + "lines": [ + { + "bbox": [ + 466, + 148, + 515, + 157 + ], + "spans": [ + { + "bbox": [ + 466, + 148, + 515, + 157 + ], + "type": "text", + "content": "(b) View-batch" + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 313, + 167, + 555, + 266 + ], + "lines": [ + { + "bbox": [ + 313, + 167, + 555, + 266 + ], + "spans": [ + { + "bbox": [ + 313, + 167, + 555, + 266 + ], + "type": "text", + "content": "Figure 3. Schematic illustration of the proposed view-batch model. In subfigure (b), we show our view-batch model employing the replay " + }, + { + "bbox": [ + 313, + 167, + 555, + 266 + ], + "type": "inline_equation", + "content": "(\\mathrm{V} = 4)" + }, + { + "bbox": [ + 313, + 167, + 555, + 266 + ], + "type": "text", + "content": " and self-supervised learning approach. In contrast to (a) the baseline method, we learn multiple views of the same sample (marked as different shades) using the proposed view-batch self-supervised loss to learn it extensively and ensure enough time-space between recall intervals. For simplicity, we assume in (a) that the entire training data and batch size are the same as four, thus a single training epoch constitutes one batch." + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "image_caption" + } + ], + "index": 7 + }, + { + "bbox": [ + 313, + 289, + 463, + 301 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 289, + 463, + 301 + ], + "spans": [ + { + "bbox": [ + 313, + 289, + 463, + 301 + ], + "type": "text", + "content": "3.1. Replay with Augmentation" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 313, + 308, + 555, + 510 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 308, + 555, + 510 + ], + "spans": [ + { + "bbox": [ + 313, + 308, + 555, + 510 + ], + "type": "text", + "content": "Our view-batch model employs a replaying method with the augmentation to adjust the recall interval of retraining the same samples. In the conventional learning scheduler, a sample-batch is replayed in the training process of neural networks. It means a batch for training networks consists of multiple unique training samples. Therefore, the recall interval of retraining is the same as the number of training samples. However, the strategy on our learning scheduler alternates the replaying method at the view-batch level. We structure a view-batch to have multiple views of a single sample. Thus, the recall interval of a single sample increases with the size of view-batch. The concept of this view-batch model for the replay method is shown in Figure 3. It shows that when learning four samples in sequence, the time-space of re-training is four, whereas when learning " + }, + { + "bbox": [ + 313, + 308, + 555, + 510 + ], + "type": "inline_equation", + "content": "V" + }, + { + "bbox": [ + 313, + 308, + 555, + 510 + ], + "type": "text", + "content": " views of a single sample, the time-space is extended by " + }, + { + "bbox": [ + 313, + 308, + 555, + 510 + ], + "type": "inline_equation", + "content": "V" + }, + { + "bbox": [ + 313, + 308, + 555, + 510 + ], + "type": "text", + "content": " times." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 313, + 511, + 555, + 559 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 511, + 555, + 559 + ], + "spans": [ + { + "bbox": [ + 313, + 511, + 555, + 559 + ], + "type": "text", + "content": "Formally, we define samples as " + }, + { + "bbox": [ + 313, + 511, + 555, + 559 + ], + "type": "inline_equation", + "content": "I" + }, + { + "bbox": [ + 313, + 511, + 555, + 559 + ], + "type": "text", + "content": ", sample-batch as " + }, + { + "bbox": [ + 313, + 511, + 555, + 559 + ], + "type": "inline_equation", + "content": "\\mathcal{B}^I = \\{I_i\\}_{i=1}^B" + }, + { + "bbox": [ + 313, + 511, + 555, + 559 + ], + "type": "text", + "content": ", the number of batches in a epoch as " + }, + { + "bbox": [ + 313, + 511, + 555, + 559 + ], + "type": "inline_equation", + "content": "T" + }, + { + "bbox": [ + 313, + 511, + 555, + 559 + ], + "type": "text", + "content": ", and learning scheduler as " + }, + { + "bbox": [ + 313, + 511, + 555, + 559 + ], + "type": "inline_equation", + "content": "\\mathcal{A}" + }, + { + "bbox": [ + 313, + 511, + 555, + 559 + ], + "type": "text", + "content": ". In the conventional learning algorithm, we replay each sample-batch to establish the scheduler." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 342, + 568, + 555, + 596 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 342, + 568, + 555, + 596 + ], + "spans": [ + { + "bbox": [ + 342, + 568, + 555, + 596 + ], + "type": "interline_equation", + "content": "\\mathcal {A} _ {\\text {c o n v e n t i o n a l}} = \\left[ \\underbrace {\\mathcal {B} _ {1} ^ {I} , \\cdots , \\mathcal {B} _ {T} ^ {I}} _ {\\text {r e c a l l i n t e r v a l} = B \\times T}, \\mathcal {B} _ {1} ^ {I}, \\dots , \\mathcal {B} _ {T} ^ {I} \\right]. \\tag {1}", + "image_path": "134eec3dcabccd9e012b40a423680b0946259fa00fdd00a1ecb5e1b5fdd5688d.jpg" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 313, + 605, + 556, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 605, + 556, + 715 + ], + "spans": [ + { + "bbox": [ + 313, + 605, + 556, + 715 + ], + "type": "text", + "content": "Therefore, the recall interval of retraining is the same as the number of training samples " + }, + { + "bbox": [ + 313, + 605, + 556, + 715 + ], + "type": "inline_equation", + "content": "B \\times T" + }, + { + "bbox": [ + 313, + 605, + 556, + 715 + ], + "type": "text", + "content": " in an epoch. However, to extend the recall interval, we replace the replaying unit from the sample-batch to the view-batch " + }, + { + "bbox": [ + 313, + 605, + 556, + 715 + ], + "type": "inline_equation", + "content": "\\mathcal{B}^{\\nu} = \\{\\mathcal{V}_i\\}_{i=1}^{B}" + }, + { + "bbox": [ + 313, + 605, + 556, + 715 + ], + "type": "text", + "content": ". We construct a view-batch by including multiple views of the same sample " + }, + { + "bbox": [ + 313, + 605, + 556, + 715 + ], + "type": "inline_equation", + "content": "\\mathcal{V}_i = \\{I_i\\}_{i=1}^V" + }, + { + "bbox": [ + 313, + 605, + 556, + 715 + ], + "type": "text", + "content": ", where " + }, + { + "bbox": [ + 313, + 605, + 556, + 715 + ], + "type": "inline_equation", + "content": "V" + }, + { + "bbox": [ + 313, + 605, + 556, + 715 + ], + "type": "text", + "content": " denotes the number of views consisting of different augmentations of a single image. By modifying the replaying unit of Equation (1) with a constructed view-batch, we define our new learning" + } + ] + } + ], + "index": 14 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "text", + "content": "10079" + } + ] + } + ], + "index": 15 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 2 + }, + { + "para_blocks": [ + { + "type": "code", + "bbox": [ + 56, + 88, + 296, + 264 + ], + "blocks": [ + { + "bbox": [ + 56, + 72, + 267, + 83 + ], + "lines": [ + { + "bbox": [ + 56, + 72, + 267, + 83 + ], + "spans": [ + { + "bbox": [ + 56, + 72, + 267, + 83 + ], + "type": "text", + "content": "Algorithm 1 Pseudocode for our view-batch model." + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "code_caption" + }, + { + "bbox": [ + 56, + 88, + 296, + 264 + ], + "lines": [ + { + "bbox": [ + 56, + 88, + 296, + 264 + ], + "spans": [ + { + "bbox": [ + 56, + 88, + 296, + 264 + ], + "type": "text", + "content": "Require: dataset " + }, + { + "bbox": [ + 56, + 88, + 296, + 264 + ], + "type": "inline_equation", + "content": "\\mathcal{D}^{\\mathrm{train}}" + }, + { + "bbox": [ + 56, + 88, + 296, + 264 + ], + "type": "text", + "content": " parameters " + }, + { + "bbox": [ + 56, + 88, + 296, + 264 + ], + "type": "inline_equation", + "content": "\\theta" + }, + { + "bbox": [ + 56, + 88, + 296, + 264 + ], + "type": "text", + "content": " , learning rate " + }, + { + "bbox": [ + 56, + 88, + 296, + 264 + ], + "type": "inline_equation", + "content": "\\gamma" + }, + { + "bbox": [ + 56, + 88, + 296, + 264 + ], + "type": "text", + "content": " , batch size " + }, + { + "bbox": [ + 56, + 88, + 296, + 264 + ], + "type": "inline_equation", + "content": "B" + }, + { + "bbox": [ + 56, + 88, + 296, + 264 + ], + "type": "text", + "content": " , view-batch size " + }, + { + "bbox": [ + 56, + 88, + 296, + 264 + ], + "type": "inline_equation", + "content": "V" + }, + { + "bbox": [ + 56, + 88, + 296, + 264 + ], + "type": "text", + "content": " \n1: Initialize network parameters " + }, + { + "bbox": [ + 56, + 88, + 296, + 264 + ], + "type": "inline_equation", + "content": "\\theta" + }, + { + "bbox": [ + 56, + 88, + 296, + 264 + ], + "type": "text", + "content": " \n2: Update " + }, + { + "bbox": [ + 56, + 88, + 296, + 264 + ], + "type": "inline_equation", + "content": "B\\gets B / V" + }, + { + "bbox": [ + 56, + 88, + 296, + 264 + ], + "type": "text", + "content": " \n3: for number of training iterations do \n4: " + }, + { + "bbox": [ + 56, + 88, + 296, + 264 + ], + "type": "inline_equation", + "content": "\\begin{array}{rl} & {\\mathcal{B}^{I} = \\{I_{i}\\}_{i = 1}^{B}\\mathrm{with}I_{i}\\stackrel {\\mathrm{i.i.d.}}{\\sim}D^{\\mathrm{train}}}\\\\ & {\\mathrm{for}i = 1,\\dots ,B\\mathrm{do}}\\\\ & {\\left\\{ \\begin{array}{l}\\mathcal{V}_{i}\\gets \\{I_{i}\\}_{j = 1}^{V}\\\\ \\mathcal{V}_{i,1}\\gets A_{w}(\\mathcal{V}_{i,1})\\\\ \\mathcal{V}_{i,2}\\gets A_{s}(\\mathcal{V}_{i,2:}) \\end{array} \\right.}\\\\ & {\\bar{\\mathcal{B}}^{\\nu}\\gets \\{\\mathcal{V}_{i}\\}_{i = 1}^{B}}\\\\ & {\\mathrm{Compute}l_{\\mathrm{sup}}\\gets L_{\\mathrm{sup}}(\\theta ,\\mathcal{B}^{\\nu})}\\\\ & {\\mathrm{Compute}l_{\\mathrm{ssl}}\\gets L_{\\mathrm{ssl}}(\\theta ,\\mathcal{B}^{\\nu})\\triangleright SSL\\mathrm{of}\\mathrm{Equation}(3)}\\\\ & {\\mathrm{Update}\\theta \\gets \\theta -\\gamma \\nabla_{\\theta}(l_{\\mathrm{sup}} + l_{\\mathrm{ssl}})} \\end{array}" + }, + { + "bbox": [ + 56, + 88, + 296, + 264 + ], + "type": "text", + "content": " \n7: replay of Equation (2) \n8: \n9: \n10: \n11: \n12:" + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "code_body" + } + ], + "index": 1, + "sub_type": "algorithm" + }, + { + "bbox": [ + 55, + 284, + 143, + 295 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 284, + 143, + 295 + ], + "spans": [ + { + "bbox": [ + 55, + 284, + 143, + 295 + ], + "type": "text", + "content": "schedule as follows:" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 88, + 304, + 295, + 332 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 88, + 304, + 295, + 332 + ], + "spans": [ + { + "bbox": [ + 88, + 304, + 295, + 332 + ], + "type": "interline_equation", + "content": "\\mathcal {A} _ {\\text {o u r s}} = \\left[ \\underbrace {\\mathcal {B} _ {1} ^ {\\nu} , \\cdots , \\mathcal {B} _ {T} ^ {\\nu}} _ {\\text {r e c a l l i n t e r v a l} = B \\times T \\times V}, \\mathcal {B} _ {1} ^ {\\nu}, \\dots , \\mathcal {B} _ {T} ^ {\\nu} \\right], \\tag {2}", + "image_path": "bb8add008a493d74ad2eeb2739431a84ae972eea84302dcadd4be87816496c35.jpg" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 55, + 340, + 296, + 473 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 340, + 296, + 473 + ], + "spans": [ + { + "bbox": [ + 55, + 340, + 296, + 473 + ], + "type": "text", + "content": "Based on this rescheduling, the recall interval is increased by " + }, + { + "bbox": [ + 55, + 340, + 296, + 473 + ], + "type": "inline_equation", + "content": "V" + }, + { + "bbox": [ + 55, + 340, + 296, + 473 + ], + "type": "text", + "content": " times. While we process " + }, + { + "bbox": [ + 55, + 340, + 296, + 473 + ], + "type": "inline_equation", + "content": "V" + }, + { + "bbox": [ + 55, + 340, + 296, + 473 + ], + "type": "text", + "content": " views at a time, we decrease the total learning epoch by " + }, + { + "bbox": [ + 55, + 340, + 296, + 473 + ], + "type": "inline_equation", + "content": "V" + }, + { + "bbox": [ + 55, + 340, + 296, + 473 + ], + "type": "text", + "content": ", ensuring that the number of total samples that networks have seen during training is the same as the baseline method. As a result, networks repeatedly forget and retrain samples during extended recall intervals, thereby improving long-term memory retention. This learning of multiple views of a single sample adjusts the time-space while networks learn the same sample repeatedly. To promote extensive learning, we present the self-supervised learning in the following." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 55, + 479, + 195, + 493 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 479, + 195, + 493 + ], + "spans": [ + { + "bbox": [ + 55, + 479, + 195, + 493 + ], + "type": "text", + "content": "3.2. Self-supervised Learning" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 55, + 498, + 296, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 498, + 296, + 714 + ], + "spans": [ + { + "bbox": [ + 55, + 498, + 296, + 714 + ], + "type": "text", + "content": "We present a self-supervised learning model with our VBM to extensively learn a sample at a time. It is well-known that self-supervision generalizes the networks rather than acquiring task-specific knowledge [5]. Further, task-agnostic knowledge from self-supervision is more robust to catastrophic forgetting of continual learning [8]. For these reasons, we propose the SSL approach to effectively learn view-batch in a (i) simple and an (ii) efficient manner. For (i) simplicity, unlike previous SSL methods [5, 11, 20], which require changes, such as architecture design [8] or teacher networks [12], we only modify an objective function to learn common knowledge within a view-batch of same samples. To satisfy (ii) efficiency, our method does not use an extra training phase or longer training epochs than supervised learning, ensuring the minimal training cost as analyzed in the supplementary material. Thus, the proposed method could be used in most continual learning methods without increasing computing costs as a drop-in replace" + } + ] + } + ], + "index": 6 + }, + { + "type": "image", + "bbox": [ + 316, + 71, + 394, + 135 + ], + "blocks": [ + { + "bbox": [ + 316, + 71, + 394, + 135 + ], + "lines": [ + { + "bbox": [ + 316, + 71, + 394, + 135 + ], + "spans": [ + { + "bbox": [ + 316, + 71, + 394, + 135 + ], + "type": "image", + "image_path": "bb59e13fbd75342f101e6a10e184be0c6c0459ee20c8c795814ba8549e9d9d57.jpg" + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 316, + 137, + 394, + 148 + ], + "lines": [ + { + "bbox": [ + 316, + 137, + 394, + 148 + ], + "spans": [ + { + "bbox": [ + 316, + 137, + 394, + 148 + ], + "type": "text", + "content": "(a) Degree of forgetting" + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "image_caption" + } + ], + "index": 7 + }, + { + "type": "image", + "bbox": [ + 395, + 75, + 472, + 135 + ], + "blocks": [ + { + "bbox": [ + 395, + 75, + 472, + 135 + ], + "lines": [ + { + "bbox": [ + 395, + 75, + 472, + 135 + ], + "spans": [ + { + "bbox": [ + 395, + 75, + 472, + 135 + ], + "type": "image", + "image_path": "522012dc639823957c1068037f3f5cdee1e721d183f991328ff70134b26119bc.jpg" + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 400, + 137, + 466, + 148 + ], + "lines": [ + { + "bbox": [ + 400, + 137, + 466, + 148 + ], + "spans": [ + { + "bbox": [ + 400, + 137, + 466, + 148 + ], + "type": "text", + "content": "(b) Forgetting curve" + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 313, + 156, + 555, + 278 + ], + "lines": [ + { + "bbox": [ + 313, + 156, + 555, + 278 + ], + "spans": [ + { + "bbox": [ + 313, + 156, + 555, + 278 + ], + "type": "text", + "content": "Figure 4. Empirical findings related to forgetting curve theory. (a) We report the degree of forgetting for different recall intervals with a " + }, + { + "bbox": [ + 313, + 156, + 555, + 278 + ], + "type": "inline_equation", + "content": "95\\%" + }, + { + "bbox": [ + 313, + 156, + 555, + 278 + ], + "type": "text", + "content": " confidence interval denoted as shaded region. The degree of forgetting is measured as the performance degradation of each sample's classification accuracy between recall intervals. (b) We show the decay of memory retention as the learning progresses. This graph shows that when the formal learning stage is finished, memory retention decays over time for all three cases. (c) We compare the network's classification accuracy of continual learning. It shows that x3 achieves the best performance thanks to the slow memory retention decay." + } + ] + } + ], + "index": 13, + "angle": 0, + "type": "image_caption" + } + ], + "index": 9 + }, + { + "type": "image", + "bbox": [ + 473, + 75, + 550, + 135 + ], + "blocks": [ + { + "bbox": [ + 473, + 75, + 550, + 135 + ], + "lines": [ + { + "bbox": [ + 473, + 75, + 550, + 135 + ], + "spans": [ + { + "bbox": [ + 473, + 75, + 550, + 135 + ], + "type": "image", + "image_path": "d62d2a25d9a2e39869c99a38b80f9c2627d6e7c567e7ac89d4ffb1c81396312a.jpg" + } + ] + } + ], + "index": 11, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 490, + 137, + 533, + 148 + ], + "lines": [ + { + "bbox": [ + 490, + 137, + 533, + 148 + ], + "spans": [ + { + "bbox": [ + 490, + 137, + 533, + 148 + ], + "type": "text", + "content": "(c) Accuracy" + } + ] + } + ], + "index": 12, + "angle": 0, + "type": "image_caption" + } + ], + "index": 11 + }, + { + "bbox": [ + 313, + 299, + 378, + 310 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 299, + 378, + 310 + ], + "spans": [ + { + "bbox": [ + 313, + 299, + 378, + 310 + ], + "type": "text", + "content": "ment approach." + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 313, + 311, + 555, + 431 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 311, + 555, + 431 + ], + "spans": [ + { + "bbox": [ + 313, + 311, + 555, + 431 + ], + "type": "text", + "content": "To implement our approach, we employ a one-to-many divergence loss to learn self-supervision within a view-batch " + }, + { + "bbox": [ + 313, + 311, + 555, + 431 + ], + "type": "inline_equation", + "content": "\\{\\mathcal{V}_i\\}_{i=1}^B" + }, + { + "bbox": [ + 313, + 311, + 555, + 431 + ], + "type": "text", + "content": ". Our one-to-many divergence loss minimizes the average divergence between one weak-augmented view " + }, + { + "bbox": [ + 313, + 311, + 555, + 431 + ], + "type": "inline_equation", + "content": "A_w(\\mathcal{V}_{i,1})" + }, + { + "bbox": [ + 313, + 311, + 555, + 431 + ], + "type": "text", + "content": " and the remaining strong-augmented views " + }, + { + "bbox": [ + 313, + 311, + 555, + 431 + ], + "type": "inline_equation", + "content": "A_s(\\mathcal{V}_{i,2:})" + }, + { + "bbox": [ + 313, + 311, + 555, + 431 + ], + "type": "text", + "content": " of a view-batch at the logit level. By aggregating different augmented views closer, networks learn common characteristics of objects. Specifically, leveraging KL divergence " + }, + { + "bbox": [ + 313, + 311, + 555, + 431 + ], + "type": "inline_equation", + "content": "D_{\\mathrm{KL}}" + }, + { + "bbox": [ + 313, + 311, + 555, + 431 + ], + "type": "text", + "content": " between different augmentations, we define one-to-many-based self-supervised loss as" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 326, + 438, + 554, + 473 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 326, + 438, + 554, + 473 + ], + "spans": [ + { + "bbox": [ + 326, + 438, + 554, + 473 + ], + "type": "interline_equation", + "content": "L _ {\\mathrm {s s l}} \\left(f _ {\\theta}, \\mathcal {B} ^ {\\nu}\\right) = \\frac {1}{B \\cdot (V - 1)} \\sum_ {i = 1} ^ {B} \\sum_ {j = 2} ^ {V} D _ {\\mathrm {K L}} \\left(p _ {i} ^ {1} | | p _ {i} ^ {j}\\right), \\tag {3}", + "image_path": "b820005050dca27286abbfd8382cbcb5c9dc4e0da972a2a9c3693b66e81bffd2.jpg" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 313, + 483, + 554, + 544 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 483, + 554, + 544 + ], + "spans": [ + { + "bbox": [ + 313, + 483, + 554, + 544 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 313, + 483, + 554, + 544 + ], + "type": "inline_equation", + "content": "p_i^1" + }, + { + "bbox": [ + 313, + 483, + 554, + 544 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 313, + 483, + 554, + 544 + ], + "type": "inline_equation", + "content": "p_i^j" + }, + { + "bbox": [ + 313, + 483, + 554, + 544 + ], + "type": "text", + "content": " represent network's prediction for weak augmented samples " + }, + { + "bbox": [ + 313, + 483, + 554, + 544 + ], + "type": "inline_equation", + "content": "\\sigma(f_{\\theta}(A_w(\\mathcal{V}_{i,1})))" + }, + { + "bbox": [ + 313, + 483, + 554, + 544 + ], + "type": "text", + "content": " and strong augmented samples " + }, + { + "bbox": [ + 313, + 483, + 554, + 544 + ], + "type": "inline_equation", + "content": "\\sigma(f_{\\theta}(A_s(\\mathcal{V}_{i,j})))" + }, + { + "bbox": [ + 313, + 483, + 554, + 544 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 313, + 483, + 554, + 544 + ], + "type": "inline_equation", + "content": "f_{\\theta}" + }, + { + "bbox": [ + 313, + 483, + 554, + 544 + ], + "type": "text", + "content": " is neural networks, and " + }, + { + "bbox": [ + 313, + 483, + 554, + 544 + ], + "type": "inline_equation", + "content": "\\sigma" + }, + { + "bbox": [ + 313, + 483, + 554, + 544 + ], + "type": "text", + "content": " stand for the softmax. Finally, as shown in Algorithm 1, we define our final objective function as" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 362, + 552, + 555, + 567 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 362, + 552, + 555, + 567 + ], + "spans": [ + { + "bbox": [ + 362, + 552, + 555, + 567 + ], + "type": "interline_equation", + "content": "\\min _ {f _ {\\theta}} L _ {\\sup } \\left(f _ {\\theta}, \\mathcal {B} ^ {\\nu}\\right) + L _ {\\operatorname {s s l}} \\left(f _ {\\theta}, \\mathcal {B} ^ {\\nu}\\right), \\tag {4}", + "image_path": "62fa3c6af41355b37f0ae72d16e91ee3c59ca729bf8c70815dbba8b5d37a18be.jpg" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 313, + 574, + 555, + 652 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 574, + 555, + 652 + ], + "spans": [ + { + "bbox": [ + 313, + 574, + 555, + 652 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 313, + 574, + 555, + 652 + ], + "type": "inline_equation", + "content": "L_{\\sup}(f_{\\theta}, \\mathcal{B}^{\\nu}) = \\frac{1}{B \\cdot V} \\sum_{i=1}^{B} \\sum_{j=1}^{V} \\mathcal{H}(y_{i}, p_{i}^{j})" + }, + { + "bbox": [ + 313, + 574, + 555, + 652 + ], + "type": "text", + "content": " of each view's prediction " + }, + { + "bbox": [ + 313, + 574, + 555, + 652 + ], + "type": "inline_equation", + "content": "p_{i}^{j}" + }, + { + "bbox": [ + 313, + 574, + 555, + 652 + ], + "type": "text", + "content": " and its label " + }, + { + "bbox": [ + 313, + 574, + 555, + 652 + ], + "type": "inline_equation", + "content": "y_{i}" + }, + { + "bbox": [ + 313, + 574, + 555, + 652 + ], + "type": "text", + "content": " using cross-entropy loss " + }, + { + "bbox": [ + 313, + 574, + 555, + 652 + ], + "type": "inline_equation", + "content": "\\mathcal{H}" + }, + { + "bbox": [ + 313, + 574, + 555, + 652 + ], + "type": "text", + "content": ". We select widely used auto-augmentation [13] as a strong augmentation method, which slightly affects the performance of the baseline method as shown in Tab. 6, while using horizontal flip as weak augmentation." + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 313, + 658, + 430, + 672 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 658, + 430, + 672 + ], + "spans": [ + { + "bbox": [ + 313, + 658, + 430, + 672 + ], + "type": "text", + "content": "3.3. Empirical Evidence" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 313, + 677, + 554, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 677, + 554, + 713 + ], + "spans": [ + { + "bbox": [ + 313, + 677, + 554, + 713 + ], + "type": "text", + "content": "This subsection empirically confirms our assumption that optimal recall interval improves memory retention, and accordingly, the performance of the continual method is en" + } + ] + } + ], + "index": 21 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "type": "text", + "content": "10080" + } + ] + } + ], + "index": 22 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 3 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 100, + 70, + 509, + 354 + ], + "blocks": [ + { + "bbox": [ + 100, + 70, + 509, + 354 + ], + "lines": [ + { + "bbox": [ + 100, + 70, + 509, + 354 + ], + "spans": [ + { + "bbox": [ + 100, + 70, + 509, + 354 + ], + "type": "table", + "html": "
BufferMethodS-CIFAR-10S-Tiny-ImageNet
CILTILAvg*ΔCILTILAvg*Δ
-Joint92.20±0.1598.31±0.1295.2659.99±0.1982.04±0.1071.02
0LwF-61.98±0.9861.98-15.26±0.7115.26
VBM-LwF-77.53±3.0677.53+15.55-51.21±0.3451.21+35.95
200DER56.58±2.0889.32±1.3272.9511.56±0.5841.67±0.6326.62
ER50.27±0.7091.66±0.7970.978.67±0.2338.03±0.7923.35
VBM-ER52.56±1.5993.58±0.2373.07+2.108.83±0.1344.03±1.1226.43+3.08
iCaRL64.11±1.8690.20±0.8877.1619.09±0.1353.07±0.2736.08
VBM-iCaRL69.73±0.8792.76±0.4581.25+4.0921.39±0.1856.31±0.4238.85+2.77
DER++61.67±1.0290.61±1.1376.1412.48±1.6041.19±0.4826.84
VBM-DER++66.99±2.9894.30±0.2180.65+4.5113.67±0.4144.36±0.8829.02+2.18
500DER65.93±5.3892.52±0.6879.2318.04±1.3052.90±0.9135.47
ER56.37±0.9694.15±0.0475.2610.20±0.3749.13±0.3129.67
VBM-ER63.39±2.7794.20±0.2578.80+3.5411.21±0.3052.01±0.7131.61+1.94
iCaRL62.22±3.4588.86±3.2375.5423.93±0.9758.89±0.8541.41
VBM-iCaRL68.97±0.3893.10±0.6681.04+5.5026.89±0.5462.61±0.1744.75+3.34
DER++69.94±2.0493.49±0.2681.7219.39±1.0951.71±0.3535.55
VBM-DER++75.22±0.8194.81±0.3985.02+3.3019.46±0.4649.80±0.8034.63-0.92
5120DER80.79±0.7594.46±0.3087.6332.92±3.0964.77±2.1448.85
ER74.23±4.1896.86±0.2885.5528.07±0.3767.58±0.3047.83
VBM-ER79.29±2.2997.02±0.1388.16+2.6133.07±0.3668.91±0.5050.99+3.16
iCaRL78.49±0.3894.84±0.8286.6732.16±0.2466.75±0.3649.46
VBM-iCaRL79.40±0.6396.09±0.3287.75+1.0834.40±0.1268.86±0.2651.63+2.17
DER++85.42±0.7496.24±0.1990.8334.06±0.5765.28±1.2049.67
VBM-DER++86.62±0.3897.07±0.0791.85+1.0235.14±0.3667.52±0.4851.33+1.66
", + "image_path": "0e693dd0f1344b228d6c3a54347bd61233162a949e5a1b8abb2c7cf39ea1a3c6.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "table_body" + } + ], + "index": 0 + }, + { + "type": "table", + "bbox": [ + 69, + 436, + 282, + 554 + ], + "blocks": [ + { + "bbox": [ + 55, + 362, + 555, + 418 + ], + "lines": [ + { + "bbox": [ + 55, + 362, + 555, + 418 + ], + "spans": [ + { + "bbox": [ + 55, + 362, + 555, + 418 + ], + "type": "text", + "content": "Table 1. Experimental results on CIL and TIL protocols. In this evaluation, we report the last top-1 accuracy " + }, + { + "bbox": [ + 55, + 362, + 555, + 418 + ], + "type": "inline_equation", + "content": "(\\%)" + }, + { + "bbox": [ + 55, + 362, + 555, + 418 + ], + "type": "text", + "content": " on the S-CIFAR-10 and S-Tiny-ImageNet benchmarks with four different buffer sizes. ResNet-18 backbone is used for each continual learning method in this evaluation. *Avg is the averaged accuracy of CIL and TIL. We reproduce baseline methods following Buzzega et al. [3]. We run every method three times with different random seeds for a reliable result. Ours improves performance consistently in most results. The performance of Joint is the upper bound of this experiment, where all old training samples are used in every step." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 69, + 436, + 282, + 554 + ], + "lines": [ + { + "bbox": [ + 69, + 436, + 282, + 554 + ], + "spans": [ + { + "bbox": [ + 69, + 436, + 282, + 554 + ], + "type": "table", + "html": "
MethodAvgΔLastΔ
Joint-46.48±3.62
AGEM24.88±0.5728.94±0.67
AGEM-R24.72±0.5928.01±0.93
FDR34.62±0.3334.46±0.32
DER39.84±0.1540.46±0.47
ER32.38±2.8832.02±2.39
VBM-ER35.80±0.55+3.4234.72±0.37+2.70
DER++34.56±3.4234.75±2.41
VBM-DER++42.16±0.20+7.2441.81±0.08+7.06
", + "image_path": "e84ec13eb0f47a5d137e994bbeec5c781d842ec90e8c36d8841e044e7e30dd9c.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "table_body" + } + ], + "index": 2 + }, + { + "bbox": [ + 55, + 563, + 295, + 606 + ], + "lines": [ + { + "bbox": [ + 55, + 563, + 295, + 606 + ], + "spans": [ + { + "bbox": [ + 55, + 563, + 295, + 606 + ], + "type": "text", + "content": "Table 2. Experimental results on DIL protocol. This evaluation shows Avg and Last top-1 accuracy " + }, + { + "bbox": [ + 55, + 563, + 295, + 606 + ], + "type": "inline_equation", + "content": "(\\%)" + }, + { + "bbox": [ + 55, + 563, + 295, + 606 + ], + "type": "text", + "content": " on the DomainNet benchmark. We utilize ResNet-18 as the backbone network in this evaluation." + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "text" + }, + { + "bbox": [ + 55, + 629, + 295, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 629, + 295, + 713 + ], + "spans": [ + { + "bbox": [ + 55, + 629, + 295, + 713 + ], + "type": "text", + "content": "hanced. Specifically, the logical development of our assumption is that 1) if the network sufficiently forgets a sample at the optimal recall interval and then extensively retrans it, 2) memory retention is enhanced as training progresses, and therefore, 3) the accuracy of continual learning is also improved. To evaluate our assumption, we measure 1) the degree of forgetting between a retraining session for" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 313, + 439, + 555, + 703 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 439, + 555, + 703 + ], + "spans": [ + { + "bbox": [ + 313, + 439, + 555, + 703 + ], + "type": "text", + "content": "training samples, 2) the decay speed of the memory retention, and 3) the average accuracy at which networks correctly classify their class labels. Figure 4 shows the results for the empirical evidence of our assumption. As depicted in Figure 4(a), the degree of forgetting becomes larger as the recall interval increases. It is obvious that the longer recall interval allows the network to forget a training sample before retraining it again. Figure 4(b) demonstrates the different levels of memory retention decay due to the recall intervals. The experiment results show that memory decays slowest in the x3 case, which has an interval three times longer than the baseline x1 case. Notably, our method achieves considerably higher performance at the end of the formal learning stage. This result is attributed to the fact that multiple epochs are used in training networks, and the memory retention decay between epochs is slow for the x3 case. Accordingly, x3 with our replay method achieves improved accuracy compared to other recall interval cases, as shown in Figure 4(c). We conduct more experiments to show empirical proofs on different datasets, backbone networks, and buffer sizes in Figure 6. The definition of forgetting is given in the supplementary material." + } + ] + } + ], + "index": 5 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 749, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 749, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 749, + 317, + 757 + ], + "type": "text", + "content": "10081" + } + ] + } + ], + "index": 6 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 4 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 82, + 70, + 531, + 213 + ], + "blocks": [ + { + "bbox": [ + 82, + 70, + 531, + 213 + ], + "lines": [ + { + "bbox": [ + 82, + 70, + 531, + 213 + ], + "spans": [ + { + "bbox": [ + 82, + 70, + 531, + 213 + ], + "type": "table", + "html": "
Method5 Step10 Step20 Step
AvgΔLastΔAvgΔLastΔAvgΔLastΔ
Joint80.40-80.41-81.49-
iCaRL71.14±0.3459.7165.27±1.0250.7461.20±0.8343.75
UCIR62.77±0.8247.3158.66±0.7143.3958.17±0.3040.63
BiC73.10±0.5562.1068.80±1.2053.5466.48±0.3247.02
WA72.81±0.2860.8469.46±0.2953.7867.33±0.1547.31
DyTox--73.66±0.0260.6772.27±0.1856.32
DER76.77±0.0168.06±0.0075.72±0.0864.32±0.0574.96±0.0161.80±0.07
VBM-DER78.60±0.23+1.8370.60±0.12+2.5478.12±0.07+2.4067.04±0.11+2.7276.95±0.02+1.9964.29±0.04+2.49
TCIL77.33±0.0869.48±0.1476.33±0.1565.66±0.0274.32±0.0162.54±0.02
VBM-TCIL79.23±0.01+1.9071.23±0.04+1.7578.02±0.03+1.6968.14±0.08+2.4876.83±0.00+2.5167.16±0.09+4.62
", + "image_path": "2493e963540e7d19407c871e890f0242242862d4bce9f0fd378d3b79bb2c8993.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "table_body" + } + ], + "index": 0 + }, + { + "type": "table", + "bbox": [ + 56, + 283, + 296, + 399 + ], + "blocks": [ + { + "bbox": [ + 55, + 220, + 555, + 266 + ], + "lines": [ + { + "bbox": [ + 55, + 220, + 555, + 266 + ], + "spans": [ + { + "bbox": [ + 55, + 220, + 555, + 266 + ], + "type": "text", + "content": "Table 3. Experimental results on rehearsal-based CIL protocol. In this evaluation, we provide Avg and Last top-1 accuracy " + }, + { + "bbox": [ + 55, + 220, + 555, + 266 + ], + "type": "inline_equation", + "content": "(\\%)" + }, + { + "bbox": [ + 55, + 220, + 555, + 266 + ], + "type": "text", + "content": " on the S-CIFAR-100 benchmark with three different class incremental steps. ResNet-18 backbone is utilized to evaluate each rehearsal-based method. This evaluation also shows that ours achieves consistent performance improvements. Results of DyTox are imported from its original work [17]. We reproduce results of DER and TCIL using official implementation, while other results come from Yan et al. [44]." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 56, + 283, + 296, + 399 + ], + "lines": [ + { + "bbox": [ + 56, + 283, + 296, + 399 + ], + "spans": [ + { + "bbox": [ + 56, + 283, + 296, + 399 + ], + "type": "table", + "html": "
Method5 Step10 Step
AvgΔLastΔAvgΔLastΔ
Joint80.40-80.41-
LwF-MC32.8911.4220.716.36
BiC54.7435.4245.0824.42
LwM39.3116.7025.478.57
WA63.6749.3645.6826.14
DER40.05±0.0017.11±0.0026.67±0.008.71±0.00
VBM-DER40.47±0.00+0.4217.47±0.01+0.3626.95±0.00+0.288.90±0.00+0.19
TCIL64.40±0.1552.37±0.0856.84±0.7540.31±1.04
VBM-TCIL66.40±0.07+2.0054.69±0.02+2.3261.12±0.02+4.2845.92±0.25+5.61
", + "image_path": "6e6a9d418060d29aa4ab75ef02336598a9d7a63e3f0f0fae8b41914aa52c3c27.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "table_body" + } + ], + "index": 2 + }, + { + "bbox": [ + 55, + 407, + 296, + 495 + ], + "lines": [ + { + "bbox": [ + 55, + 407, + 296, + 495 + ], + "spans": [ + { + "bbox": [ + 55, + 407, + 296, + 495 + ], + "type": "text", + "content": "Table 4. Experimental results on non-rehearsal-based CIL protocol. For this evaluation, we measure Avg and Last top-1 accuracy " + }, + { + "bbox": [ + 55, + 407, + 296, + 495 + ], + "type": "inline_equation", + "content": "(\\%)" + }, + { + "bbox": [ + 55, + 407, + 296, + 495 + ], + "type": "text", + "content": " on the S-CIFAR-100 benchmark with two different class incremental steps. ResNet-18 backbone is applied to existing and our methods. We report original results from LwF-MC and LwM [15]. We reproduce the results of DER and TCIL by ourselves using official implementation. Other results are reported from Huang et al. [24]." + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "text" + }, + { + "bbox": [ + 55, + 518, + 137, + 532 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 518, + 137, + 532 + ], + "spans": [ + { + "bbox": [ + 55, + 518, + 137, + 532 + ], + "type": "text", + "content": "4. Experiments" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 55, + 539, + 295, + 635 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 539, + 295, + 635 + ], + "spans": [ + { + "bbox": [ + 55, + 539, + 295, + 635 + ], + "type": "text", + "content": "This section demonstrates the capabilities of the proposed view-batch model in various continual learning protocols and scenarios through extensive performance evaluation. We perform experiments for three continual learning protocols, such as class, task, and domain incremental learning. Further, we also assess our view-batch model in both rehearsal and non-rehearsal, as well as under the pre-training-based continual learning scenarios." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 55, + 646, + 182, + 659 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 646, + 182, + 659 + ], + "spans": [ + { + "bbox": [ + 55, + 646, + 182, + 659 + ], + "type": "text", + "content": "4.1. Experimental Settings" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 55, + 665, + 296, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 665, + 296, + 715 + ], + "spans": [ + { + "bbox": [ + 55, + 665, + 296, + 715 + ], + "type": "text", + "content": "Datasets. We evaluate our methods on widely used benchmarks such as S-CIFAR-10/100, S-Tiny-ImageNet, S-ImageNet-R, and DomainNet where the prefix of 'S-' denotes the sequential data configuration. We use the com" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 313, + 285, + 555, + 310 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 285, + 555, + 310 + ], + "spans": [ + { + "bbox": [ + 313, + 285, + 555, + 310 + ], + "type": "text", + "content": "monly adopted class ordering lists, following existing works of Yan et al. [44] and Huang et al. [24]." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 313, + 326, + 556, + 602 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 326, + 556, + 602 + ], + "spans": [ + { + "bbox": [ + 313, + 326, + 556, + 602 + ], + "type": "text", + "content": "Protocols. In our experiments, we utilize three continual learning protocols. First, we make use of task incremental learning (TIL), which involves training neural networks on a series of tasks in a sequential manner while evaluating their performance on both newly acquired and previously learned tasks. Specifically, the TIL protocol consists of the following detailed configurations: 1) a task is defined by a set of label classes, and 2) we evaluate networks knowing the task identity of test samples. Second, we adopt class incremental learning (CIL), which is identical to TIL except that it evaluates test samples without knowing their task identity. Third, we use domain incremental learning (DIL), which focuses on learning the new samples across multiple sequential domains rather than acquiring new classes. In addition to these three protocols, we also include both rehearsal and non-rehearsal scenarios, divided by the use of a memory buffer during continual learning. For the non-rehearsal scenario, we evaluate both scratch- and pre-training-based learning scenarios to showcase extensive experimental results. For all these protocols and scenarios, we provide the mean and standard deviation of three runs for reliable experimental results. The details of the experiment can be found in the supplementary material." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 313, + 617, + 557, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 617, + 557, + 715 + ], + "spans": [ + { + "bbox": [ + 313, + 617, + 557, + 715 + ], + "type": "text", + "content": "Baseline Methods. We apply our view-batch to recent baseline continual learning methods such as iCaRL [32], ER [34], DER [3, 44], TCIL [24], and SLCA [46]. We also include results from A-GEM [10], LwM [15], UCIR [23], BiC [43], WA [47], DyTox [17], L2P [42], S-iPrompt [40], DualPrompt [41], and CODA-Prompt [38]. We follow the baseline's hyper-parameters for a fair comparison between ours and its respective baseline methods. Specifically, we" + } + ] + } + ], + "index": 10 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "type": "text", + "content": "10082" + } + ] + } + ], + "index": 11 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 5 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 89, + 70, + 523, + 196 + ], + "blocks": [ + { + "bbox": [ + 89, + 70, + 523, + 196 + ], + "lines": [ + { + "bbox": [ + 89, + 70, + 523, + 196 + ], + "spans": [ + { + "bbox": [ + 89, + 70, + 523, + 196 + ], + "type": "table", + "html": "
S-CIFAR-100S-ImageNet-R
MethodTop-1Top-5Top-1Top-5
AvgLastAvgLastAvgLastAvgLast
Joint-92.54±0.14-99.24±0.02-82.41±0.02-93.27±0.07
L2P90.42±0.5683.76±0.4298.89±0.0398.08±0.1576.76±0.3373.56±0.4289.39±0.1686.61±0.03
DualPrompt90.17±0.3885.10±0.2398.90±0.0797.68±0.1273.31±0.1969.12±0.1889.19±0.0985.46±0.04
CODA-Prompt91.27±0.6686.86±0.5899.29±0.0998.64±0.0781.89±0.5976.79±0.1793.30±0.3490.49±0.15
S-iPrompt92.19±0.4088.43±0.4299.19±0.0598.63±0.0571.47±0.1468.77±0.2788.44±0.1984.91±0.31
SLCA94.31±1.1291.57±0.4699.64±0.0499.26±0.0383.58±0.6578.81±0.2194.39±0.2692.02±0.25
VBM-SLCA94.52±0.8591.78±0.1899.64±0.0399.29±0.0384.61±0.5980.07±0.1294.86±0.1992.74±0.14
", + "image_path": "2ebce71d16fdffdc13f1b260e5b450971a20eb35dcd975c47ed23d27db50d15c.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "table_body" + } + ], + "index": 0 + }, + { + "type": "image", + "bbox": [ + 62, + 256, + 173, + 342 + ], + "blocks": [ + { + "bbox": [ + 62, + 256, + 173, + 342 + ], + "lines": [ + { + "bbox": [ + 62, + 256, + 173, + 342 + ], + "spans": [ + { + "bbox": [ + 62, + 256, + 173, + 342 + ], + "type": "image", + "image_path": "eee9c2892997f41c6cf1f1f1042a54be340f19d9ebebe328fe2b936eb21a7201.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 101, + 344, + 135, + 354 + ], + "lines": [ + { + "bbox": [ + 101, + 344, + 135, + 354 + ], + "spans": [ + { + "bbox": [ + 101, + 344, + 135, + 354 + ], + "type": "text", + "content": "(a) iCaRL" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_caption" + } + ], + "index": 2 + }, + { + "type": "image", + "bbox": [ + 176, + 256, + 286, + 342 + ], + "blocks": [ + { + "bbox": [ + 176, + 256, + 286, + 342 + ], + "lines": [ + { + "bbox": [ + 176, + 256, + 286, + 342 + ], + "spans": [ + { + "bbox": [ + 176, + 256, + 286, + 342 + ], + "type": "image", + "image_path": "605ea598910eb3fa5beb3139decaece279767ebcefe705c6cbe8b007418c956c.jpg" + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 212, + 344, + 250, + 354 + ], + "lines": [ + { + "bbox": [ + 212, + 344, + 250, + 354 + ], + "spans": [ + { + "bbox": [ + 212, + 344, + 250, + 354 + ], + "type": "text", + "content": "(b) DER++" + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "image_caption" + } + ], + "index": 4 + }, + { + "type": "image", + "bbox": [ + 61, + 441, + 135, + 530 + ], + "blocks": [ + { + "bbox": [ + 55, + 363, + 295, + 429 + ], + "lines": [ + { + "bbox": [ + 55, + 363, + 295, + 429 + ], + "spans": [ + { + "bbox": [ + 55, + 363, + 295, + 429 + ], + "type": "text", + "content": "Figure 5. Experimental results on memory retention decay. This analysis reports memory retention decay of the first three tasks on the S-CIFAR-10 dataset, comparing the proposed approach against the baseline methods. The green numbers at the end of the last task are the accuracy gain of the proposed approach over the baselines." + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 61, + 441, + 135, + 530 + ], + "lines": [ + { + "bbox": [ + 61, + 441, + 135, + 530 + ], + "spans": [ + { + "bbox": [ + 61, + 441, + 135, + 530 + ], + "type": "image", + "image_path": "773d2134c08e3b1f5e4757b2f0bf45e7a6e1b18145a7f44973dfa65198898156.jpg" + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 80, + 532, + 116, + 540 + ], + "lines": [ + { + "bbox": [ + 80, + 532, + 116, + 540 + ], + "spans": [ + { + "bbox": [ + 80, + 532, + 116, + 540 + ], + "type": "text", + "content": "(a) Dataset" + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 55, + 550, + 295, + 605 + ], + "lines": [ + { + "bbox": [ + 55, + 550, + 295, + 605 + ], + "spans": [ + { + "bbox": [ + 55, + 550, + 295, + 605 + ], + "type": "text", + "content": "Figure 6. Experimental results on recall interval. We measure the last top-1 accuracy " + }, + { + "bbox": [ + 55, + 550, + 295, + 605 + ], + "type": "inline_equation", + "content": "(\\%)" + }, + { + "bbox": [ + 55, + 550, + 295, + 605 + ], + "type": "text", + "content": " (" + }, + { + "bbox": [ + 55, + 550, + 295, + 605 + ], + "type": "inline_equation", + "content": "\\uparrow" + }, + { + "bbox": [ + 55, + 550, + 295, + 605 + ], + "type": "text", + "content": ") in the top row and report forgetting " + }, + { + "bbox": [ + 55, + 550, + 295, + 605 + ], + "type": "inline_equation", + "content": "(\\downarrow)" + }, + { + "bbox": [ + 55, + 550, + 295, + 605 + ], + "type": "text", + "content": " at the bottom row, varying recall intervals with different learning factors. Dark colors mean better. We use the ResNet-18 backbone." + } + ] + } + ], + "index": 13, + "angle": 0, + "type": "image_caption" + } + ], + "index": 7 + }, + { + "type": "image", + "bbox": [ + 138, + 442, + 211, + 530 + ], + "blocks": [ + { + "bbox": [ + 138, + 442, + 211, + 530 + ], + "lines": [ + { + "bbox": [ + 138, + 442, + 211, + 530 + ], + "spans": [ + { + "bbox": [ + 138, + 442, + 211, + 530 + ], + "type": "image", + "image_path": "b17898ccb02afcbc02cac32f8a68825dde50800e921a7dc08779ecef470d4109.jpg" + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 154, + 531, + 195, + 540 + ], + "lines": [ + { + "bbox": [ + 154, + 531, + 195, + 540 + ], + "spans": [ + { + "bbox": [ + 154, + 531, + 195, + 540 + ], + "type": "text", + "content": "(b) Network" + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "image_caption" + } + ], + "index": 9 + }, + { + "type": "image", + "bbox": [ + 214, + 442, + 288, + 530 + ], + "blocks": [ + { + "bbox": [ + 214, + 442, + 288, + 530 + ], + "lines": [ + { + "bbox": [ + 214, + 442, + 288, + 530 + ], + "spans": [ + { + "bbox": [ + 214, + 442, + 288, + 530 + ], + "type": "image", + "image_path": "93aec7734c01f8373a22f703483d8be15995b2d7280e898d4cfc2b59dc4430a8.jpg" + } + ] + } + ], + "index": 11, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 234, + 531, + 267, + 540 + ], + "lines": [ + { + "bbox": [ + 234, + 531, + 267, + 540 + ], + "spans": [ + { + "bbox": [ + 234, + 531, + 267, + 540 + ], + "type": "text", + "content": "(c) Buffer" + } + ] + } + ], + "index": 12, + "angle": 0, + "type": "image_caption" + } + ], + "index": 11 + }, + { + "bbox": [ + 55, + 629, + 295, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 629, + 295, + 713 + ], + "spans": [ + { + "bbox": [ + 55, + 629, + 295, + 713 + ], + "type": "text", + "content": "use ResNet and Vision Transformer (ViT) as backbone networks in traditional and prompt-based approaches. For the traditional approach, we randomly initialize backbone weights at the beginning of every run. On the other hand, for the prompt-based approach, we adopt a frozen backbone whose weights are pre-trained on a large-scale dataset following recent prompt-based CL studies [42, 46]" + } + ] + } + ], + "index": 14 + }, + { + "type": "table", + "bbox": [ + 318, + 256, + 552, + 382 + ], + "blocks": [ + { + "bbox": [ + 55, + 204, + 555, + 237 + ], + "lines": [ + { + "bbox": [ + 55, + 204, + 555, + 237 + ], + "spans": [ + { + "bbox": [ + 55, + 204, + 555, + 237 + ], + "type": "text", + "content": "Table 5. Experimental results on pre-trained model-based CIL protocol. This evaluation presents Avg and Last top-1,5 accuracy (%) on the S-CIFAR-100 and S-ImageNet-R benchmarks without using a memory buffer. ViT-B/16 backbone is adopted for each pre-trained model-based method. We reproduce baseline methods using official implementation by ourselves, except for L2P." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 318, + 256, + 552, + 382 + ], + "lines": [ + { + "bbox": [ + 318, + 256, + 552, + 382 + ], + "spans": [ + { + "bbox": [ + 318, + 256, + 552, + 382 + ], + "type": "table", + "html": "
MethodReplayStrong augmentSSLCILTILAvgΔ
iCaRL---64.1190.2077.16-1.05
-✓-62.0490.1776.11
✓--67.5191.8779.69+2.53
✓✓-66.7891.2178.99+1.83
✓✓✓69.7392.7681.25+4.09
DER++---61.6790.6176.14+0.76
-✓-62.6691.1476.90
✓--64.7592.7978.77+2.63
✓✓-64.4493.3878.91+2.77
✓✓✓66.9994.3080.65+4.51
", + "image_path": "a05fdef5078753035f3ffa4f7403b0e1bb9848ee7df022a4ab2366ed336da763.jpg" + } + ] + } + ], + "index": 15, + "angle": 0, + "type": "table_body" + } + ], + "index": 15 + }, + { + "bbox": [ + 313, + 390, + 555, + 479 + ], + "lines": [ + { + "bbox": [ + 313, + 390, + 555, + 479 + ], + "spans": [ + { + "bbox": [ + 313, + 390, + 555, + 479 + ], + "type": "text", + "content": "Table 6. Experimental results on factor analysis. We demonstrate the last accuracy (\\%) of CIL and TIL on S-CIFAR-10, varying our main components with different augmentation types. We observe that the proposed components, such as replay and SSL, improve performance consistently and confirm that performance enhancement comes from them, not simply from strong augmentation. In the strong augmentation setting without view batch replay, we maintain the same ratio of weak to strong augmentations." + } + ] + } + ], + "index": 16, + "angle": 0, + "type": "text" + }, + { + "bbox": [ + 313, + 501, + 438, + 514 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 501, + 438, + 514 + ], + "spans": [ + { + "bbox": [ + 313, + 501, + 438, + 514 + ], + "type": "text", + "content": "4.2. Experimental Results" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 313, + 521, + 555, + 664 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 521, + 555, + 664 + ], + "spans": [ + { + "bbox": [ + 313, + 521, + 555, + 664 + ], + "type": "text", + "content": "Continual Learning under Three Protocols. We evaluate our method in the three continual learning protocols such as CIL, TIL, and DIL. For TIL and CIL, we use different memory buffer sizes: 0, 200, 500, and 5,120. For DIL, we utilize 50 samples per class, resulting in 17,250 samples in a buffer. We apply our view-batch model to LwF for the non-rehearsal memory buffer (i.e. 0 size) and ER, DER++, and iCaRL for the other buffer sizes. Tabs. 1 and 2 show that our method improves the performance of continual learning methods for all protocols. Moreover, we demonstrate that baseline methods employing our view-batch are comparable to other continual learning methods." + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 313, + 665, + 555, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 665, + 555, + 713 + ], + "spans": [ + { + "bbox": [ + 313, + 665, + 555, + 713 + ], + "type": "text", + "content": "We conduct further in-depth experiments under the most widely used CIL protocol. We consider three configurations with step sizes of 5, 10, and 20, which are the number of total incremental steps, as shown in Tab. 3. In addition, we" + } + ] + } + ], + "index": 19 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "text", + "content": "10083" + } + ] + } + ], + "index": 20 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 6 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 61, + 70, + 291, + 178 + ], + "blocks": [ + { + "bbox": [ + 61, + 70, + 291, + 178 + ], + "lines": [ + { + "bbox": [ + 61, + 70, + 291, + 178 + ], + "spans": [ + { + "bbox": [ + 61, + 70, + 291, + 178 + ], + "type": "table", + "html": "
MethodCurrent sampleBuffer sampleCILTILAvgΔ
iCaRL--64.1190.2077.16-
✓-64.6494.1279.38+2.22
-✓64.9994.2079.60+2.44
✓✓69.7392.7681.25+4.09
DER++--61.6790.6176.14-
✓-64.8893.2279.05+2.91
-✓62.6994.1578.42+2.28
✓✓66.9994.3080.65+4.51
", + "image_path": "dd1f0e551243b84a738289a7de94cfa6345865906f28ad84da8fadd0dea8a94a.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "table_body" + } + ], + "index": 0 + }, + { + "bbox": [ + 55, + 186, + 295, + 243 + ], + "lines": [ + { + "bbox": [ + 55, + 186, + 295, + 243 + ], + "spans": [ + { + "bbox": [ + 55, + 186, + 295, + 243 + ], + "type": "text", + "content": "Table 7. Experimental results on sample type. We measure the last accuracy (\\%) of CIL and TIL on S-CIFAR-10, applying our method to different sample types. We use the ResNet-18 as backbone. This analysis shows that our method works well with both sample types." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "text" + }, + { + "bbox": [ + 55, + 262, + 296, + 334 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 262, + 296, + 334 + ], + "spans": [ + { + "bbox": [ + 55, + 262, + 296, + 334 + ], + "type": "text", + "content": "take into account the use of a memory buffer in two scenarios, such as rehearsal and non-rehearsal in Tab. 4. Both experimental results demonstrate that our view-batch model consistently improves the performance of continual learning methods under the various step sizes and memory buffer configurations." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 55, + 348, + 296, + 491 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 348, + 296, + 491 + ], + "spans": [ + { + "bbox": [ + 55, + 348, + 296, + 491 + ], + "type": "text", + "content": "Non-rehearsal Continual Learning. While the nonrehearsal scenario has the advantage of not using memory buffers, they inherently face the drawback of limited performance due to the inability to retrain networks using old data in the memory buffer. To overcome this limitation, many recent studies have focused on non-rehearsal continual learning using a pre-trained model trained with a large-scale dataset. Thus, we evaluate our view-batch model in a scenario where models are pre-trained, such as SCLA. Experimental results in Tab. 5 indicate that the proposed view-batch model is effective across all benchmarks for this scenario." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 55, + 502, + 115, + 515 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 502, + 115, + 515 + ], + "spans": [ + { + "bbox": [ + 55, + 502, + 115, + 515 + ], + "type": "text", + "content": "5. Analysis" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 55, + 521, + 295, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 521, + 295, + 713 + ], + "spans": [ + { + "bbox": [ + 55, + 521, + 295, + 713 + ], + "type": "text", + "content": "We present the analysis of the proposed view-batch model. First, we validate that the optimal recall interval mitigates catastrophic forgetting of early tasks. Figure 5 visualize that the recall interval x3 maintains long-term memory retention resolving catastrophic forgetting of initial tasks. Also, Figure 6 analyzes how different factors (i.e., dataset, network scales, memory buffer size) affect the optimal recall intervals in terms of last top-1 accuracy and forgetting metrics [9, 10, 42]. We find that the recall interval x3 or x4 improves the accuracy and resolves catastrophic forgetting in various scenarios. More experimental results of different methods and task types can be found in the supplementary material. Tab. 6 shows that two main factors of the proposed method consistently improve their respective baselines, identifying that the performance improvement comes from the proposed components, not strong augmentation." + } + ] + } + ], + "index": 5 + }, + { + "type": "image", + "bbox": [ + 322, + 70, + 545, + 167 + ], + "blocks": [ + { + "bbox": [ + 322, + 70, + 545, + 167 + ], + "lines": [ + { + "bbox": [ + 322, + 70, + 545, + 167 + ], + "spans": [ + { + "bbox": [ + 322, + 70, + 545, + 167 + ], + "type": "image", + "image_path": "9feb0905c07c53017f273a2c620e2845a4f8ffe3bd19ed9627d481e09059eeaa.jpg" + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 313, + 175, + 555, + 231 + ], + "lines": [ + { + "bbox": [ + 313, + 175, + 555, + 231 + ], + "spans": [ + { + "bbox": [ + 313, + 175, + 555, + 231 + ], + "type": "text", + "content": "Figure 7. Experimental results on accuracy and forgetting. This analysis shows (a) the last top-1 accuracy " + }, + { + "bbox": [ + 313, + 175, + 555, + 231 + ], + "type": "inline_equation", + "content": "(\\%)" + }, + { + "bbox": [ + 313, + 175, + 555, + 231 + ], + "type": "text", + "content": " and (b) corresponding forgetting measures on the S-CIFAR-10 dataset with different buffer sizes. Our method consistently improves accuracy and forgetting measures." + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "image_caption" + } + ], + "index": 6 + }, + { + "bbox": [ + 313, + 251, + 553, + 287 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 251, + 553, + 287 + ], + "spans": [ + { + "bbox": [ + 313, + 251, + 553, + 287 + ], + "type": "text", + "content": "Lastly, in the rehearsal scenario, we have two different sample types, current- and buffer-based samples. Tab. 7 exhibit that ours performs well with both types." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 313, + 288, + 555, + 396 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 288, + 555, + 396 + ], + "spans": [ + { + "bbox": [ + 313, + 288, + 555, + 396 + ], + "type": "text", + "content": "We examine the proposed method in terms of the last accuracy and forgetting measurement. We apply the proposed method to two different baseline approaches. We compute average forgetting at " + }, + { + "bbox": [ + 313, + 288, + 555, + 396 + ], + "type": "inline_equation", + "content": "T" + }, + { + "bbox": [ + 313, + 288, + 555, + 396 + ], + "type": "text", + "content": "th task as " + }, + { + "bbox": [ + 313, + 288, + 555, + 396 + ], + "type": "inline_equation", + "content": "\\frac{1}{T - 1}\\sum_{t = 1}^{T - 1}\\max_{j\\in \\{t,\\dots ,T - 1\\}}a_j^t -a_T^t" + }, + { + "bbox": [ + 313, + 288, + 555, + 396 + ], + "type": "text", + "content": " where " + }, + { + "bbox": [ + 313, + 288, + 555, + 396 + ], + "type": "inline_equation", + "content": "a_j^t" + }, + { + "bbox": [ + 313, + 288, + 555, + 396 + ], + "type": "text", + "content": " means " + }, + { + "bbox": [ + 313, + 288, + 555, + 396 + ], + "type": "inline_equation", + "content": "t" + }, + { + "bbox": [ + 313, + 288, + 555, + 396 + ], + "type": "text", + "content": "th task accuracy at " + }, + { + "bbox": [ + 313, + 288, + 555, + 396 + ], + "type": "inline_equation", + "content": "j" + }, + { + "bbox": [ + 313, + 288, + 555, + 396 + ], + "type": "text", + "content": " step. Notably, higher forgetting means severe catastrophic forgetting of previous tasks. Figure 7 shows that the view-batch model improves accuracy while addressing the forgetting problem simultaneously." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 314, + 406, + 388, + 418 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 314, + 406, + 388, + 418 + ], + "spans": [ + { + "bbox": [ + 314, + 406, + 388, + 418 + ], + "type": "text", + "content": "6. Conclusion" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 313, + 426, + 555, + 606 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 426, + 555, + 606 + ], + "spans": [ + { + "bbox": [ + 313, + 426, + 555, + 606 + ], + "type": "text", + "content": "We propose the view-batch model, which is generally applicable to various continual learning scenarios. Inspired by Ebbinghaus's forgetting curve theory, the proposed view-batch model optimizes the recall interval between retraining samples, improving neural networks' long-term memory. The proposed method consists of two main components: replay and self-supervised learning. The replay optimizes the recall interval for guaranteeing sufficient memory forgetting by grouping multiple views of a sample. Also, the self-supervised learning ensures extensive learning for the multiple views using the one-to-many divergence loss. Experimental results show that sufficient forgetting preserves long-term memory and ultimately improves continual learning accuracy. We hope the proposed approach to continual learning can boost future research." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 313, + 618, + 556, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 618, + 556, + 713 + ], + "spans": [ + { + "bbox": [ + 313, + 618, + 556, + 713 + ], + "type": "text", + "content": "Acknowledgment. This paper was supported in part by the ETRI Grant funded by Korean Government (Fundamental Technology Research for Human-Centric Autonomous Intelligent Systems) under Grant 24ZB1200, Artificial Intelligence Convergence Innovation Human Resources Development (IITP-2025-RS-2023-00255968), Artificial Intelligence Innovation Hub (RS-2021-II212068), and the NRF Grant (RS-2024-00356486)." + } + ] + } + ], + "index": 12 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "type": "text", + "content": "10084" + } + ] + } + ], + "index": 13 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 7 + }, + { + "para_blocks": [ + { + "bbox": [ + 56, + 71, + 115, + 83 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 71, + 115, + 83 + ], + "spans": [ + { + "bbox": [ + 56, + 71, + 115, + 83 + ], + "type": "text", + "content": "References" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 57, + 91, + 296, + 713 + ], + "type": "list", + "angle": 0, + "index": 15, + "blocks": [ + { + "bbox": [ + 61, + 91, + 296, + 135 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 61, + 91, + 296, + 135 + ], + "spans": [ + { + "bbox": [ + 61, + 91, + 296, + 135 + ], + "type": "text", + "content": "[1] Hadi Amiri, Timothy Miller, and Guergana Savova. Repeat before forgetting: Spaced repetition for efficient and effective training of neural networks. In Conference on Empirical Methods in Natural Language Processing, 2017. 1, 2" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 61, + 136, + 296, + 178 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 61, + 136, + 296, + 178 + ], + "spans": [ + { + "bbox": [ + 61, + 136, + 296, + 178 + ], + "type": "text", + "content": "[2] Maxim Berman, Hervé Jégou, Andrea Vedaldi, Iasonas Kokkinos, and Matthijs Douze. Multigrain: a unified image embedding for classes and instances. arXiv:1902.05509, 2019.3" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 62, + 180, + 295, + 223 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 180, + 295, + 223 + ], + "spans": [ + { + "bbox": [ + 62, + 180, + 295, + 223 + ], + "type": "text", + "content": "[3] Pietro Buzzega, Matteo Boschini, Angelo Porrello, Davide Abati, and Simone Calderara. Dark experience for general continual learning: a strong, simple baseline. Advances in Neural Information Processing Systems, 2020. 5, 6" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 62, + 224, + 295, + 289 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 224, + 295, + 289 + ], + "spans": [ + { + "bbox": [ + 62, + 224, + 295, + 289 + ], + "type": "text", + "content": "[4] Boxi Cao, Qiaoyu Tang, Hongyu Lin, Shanshan Jiang, Bin Dong, Xianpei Han, Jiawei Chen, Tianshu Wang, and Le Sun. Retentive or forgetful? diving into the knowledge memorizing mechanism of language models. In Joint International Conference on Computational Linguistics, Language Resources and Evaluation, 2024. 1, 2" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 62, + 291, + 295, + 334 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 291, + 295, + 334 + ], + "spans": [ + { + "bbox": [ + 62, + 291, + 295, + 334 + ], + "type": "text", + "content": "[5] Mathilde Caron, Hugo Touvron, Ishan Misra, Herve Jégou, Julien Mairal, Piotr Bojanowski, and Armand Joulin. Emerging properties in self-supervised vision transformers. In International Conference on Computer Vision, 2021. 3, 4" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 62, + 335, + 295, + 378 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 335, + 295, + 378 + ], + "spans": [ + { + "bbox": [ + 62, + 335, + 295, + 378 + ], + "type": "text", + "content": "[6] Nicholas J Cepeda, Harold Pashler, Edward Vul, John T Wixted, and Doug Rohrer. Distributed practice in verbal recall tasks: A review and quantitative synthesis. *Psychological bulletin*, 2006. 1" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 62, + 380, + 295, + 423 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 380, + 295, + 423 + ], + "spans": [ + { + "bbox": [ + 62, + 380, + 295, + 423 + ], + "type": "text", + "content": "[7] Nicholas J Cepeda, Edward Vul, Doug Rohrer, John T Wixted, and Harold Pashler. Spacing effects in learning: A temporal ridgeline of optimal retention. Psychological science, 2008. 1" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 62, + 425, + 295, + 456 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 425, + 295, + 456 + ], + "spans": [ + { + "bbox": [ + 62, + 425, + 295, + 456 + ], + "type": "text", + "content": "[8] Hyuntak Cha, Jaeho Lee, and Jinwoo Shin. Co2l: Contrastive continual learning. In International Conference on Computer Vision, 2021. 4" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 62, + 458, + 295, + 502 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 458, + 295, + 502 + ], + "spans": [ + { + "bbox": [ + 62, + 458, + 295, + 502 + ], + "type": "text", + "content": "[9] Arslan Chaudhry, Puneet K Dokania, Thalaiyasingam Ajthan, and Philip HS Torr. Riemannian walk for incremental learning: Understanding forgetting and intransigence. In European Conference on Computer Vision, 2018. 8" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 57, + 502, + 295, + 545 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 502, + 295, + 545 + ], + "spans": [ + { + "bbox": [ + 57, + 502, + 295, + 545 + ], + "type": "text", + "content": "[10] Arslan Chaudhry, Marc'Aurelio Ranzato, Marcus Rohrbach, and Mohamed Elhoseiny. Efficient lifelong learning with a gem. In International Conference on Learning Representations, 2019. 6, 8" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 57, + 547, + 295, + 590 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 547, + 295, + 590 + ], + "spans": [ + { + "bbox": [ + 57, + 547, + 295, + 590 + ], + "type": "text", + "content": "[11] Ting Chen, Simon Kornblith, Mohammad Norouzi, and Geoffrey Hinton. A simple framework for contrastive learning of visual representations. In International Conference on Machine Learning, 2020. 3, 4" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 57, + 591, + 295, + 635 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 591, + 295, + 635 + ], + "spans": [ + { + "bbox": [ + 57, + 591, + 295, + 635 + ], + "type": "text", + "content": "[12] Haoyang Cheng, Haitao Wen, Heqian Qiu, Lanxiao Wang, Minjian Zhang, and Hongliang Li. Must unsupervised continual learning relies on previous information? In Conference on Computer Vision and Pattern Recognition, 2024. 4" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 57, + 636, + 295, + 668 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 636, + 295, + 668 + ], + "spans": [ + { + "bbox": [ + 57, + 636, + 295, + 668 + ], + "type": "text", + "content": "[13] Ekin D Cubuk, Barret Zoph, Dandelion Mane, Vijay Vasudevan, and Quoc V Le. Autoaugment: Learning augmentation policies from data. arXiv:1805.09501, 2018. 4" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 57, + 670, + 295, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 670, + 295, + 713 + ], + "spans": [ + { + "bbox": [ + 57, + 670, + 295, + 713 + ], + "type": "text", + "content": "[14] Ekin D Cubuk, Barret Zoph, Jonathon Shlens, and Quoc V Le. Randaugment: Practical automated data augmentation with a reduced search space. In Conference on Computer Vision and Pattern Recognition Workshops, 2020. 2" + } + ] + } + ], + "index": 14 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 316, + 73, + 555, + 712 + ], + "type": "list", + "angle": 0, + "index": 30, + "blocks": [ + { + "bbox": [ + 316, + 73, + 553, + 116 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 73, + 553, + 116 + ], + "spans": [ + { + "bbox": [ + 316, + 73, + 553, + 116 + ], + "type": "text", + "content": "[15] Prithviraj Dhar, Rajat Vikram Singh, Kuan-Chuan Peng, Ziyan Wu, and Rama Chellappa. Learning without memorizing. In Conference on Computer Vision and Pattern Recognition, 2019. 3, 6" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 317, + 119, + 553, + 163 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 119, + 553, + 163 + ], + "spans": [ + { + "bbox": [ + 317, + 119, + 553, + 163 + ], + "type": "text", + "content": "[16] Arthur Douillard, Matthieu Cord, Charles Ollion, Thomas Robert, and Eduardo Valle. Podnet: Pooled outputs distillation for small-tasks incremental learning. In European Conference on Computer Vision, 2020. 3" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 317, + 165, + 553, + 208 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 165, + 553, + 208 + ], + "spans": [ + { + "bbox": [ + 317, + 165, + 553, + 208 + ], + "type": "text", + "content": "[17] Arthur Douillard, Alexandre Ramé, Guillaume Couairon, and Matthieu Cord. Dytox: Transformers for continual learning with dynamic token expansion. In Conference on Computer Vision and Pattern Recognition, 2022. 6" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 317, + 211, + 553, + 232 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 211, + 553, + 232 + ], + "spans": [ + { + "bbox": [ + 317, + 211, + 553, + 232 + ], + "type": "text", + "content": "[18] Hermann Ebbinghaus. Memory: A contribution to experimental psychology. Annals of neurosciences, 2013. 1, 2" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 317, + 235, + 553, + 256 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 235, + 553, + 256 + ], + "spans": [ + { + "bbox": [ + 317, + 235, + 553, + 256 + ], + "type": "text", + "content": "[19] Kate Gordon. Class results with spaced and unspaced memorizing. Journal of Experimental Psychology, 1925. 1" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 317, + 258, + 553, + 323 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 258, + 553, + 323 + ], + "spans": [ + { + "bbox": [ + 317, + 258, + 553, + 323 + ], + "type": "text", + "content": "[20] Jean-Bastien Grill, Florian Strub, Florent Altché, Corentin Tallec, Pierre Richemond, Elena Buchatskaya, Carl Doersch, Bernardo Avila Pires, Zhaohan Guo, Mohammad Gheshlaghi Azar, et al. Bootstrap your own latent-a new approach to self-supervised learning. Advances in Neural Information Processing Systems, 2020. 4" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 317, + 327, + 553, + 392 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 327, + 553, + 392 + ], + "spans": [ + { + "bbox": [ + 317, + 327, + 553, + 392 + ], + "type": "text", + "content": "[21] Karri S Hawley, Katie E Cherry, Emily O Boudreaux, and Erin M Jackson. A comparison of adjusted spaced retrieval versus a uniform expanded retrieval schedule for learning a name-face association in older adults with probable Alzheimer's disease. Journal of Clinical and Experimental Neuropsychology, 2008. 1" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 316, + 395, + 553, + 426 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 395, + 553, + 426 + ], + "spans": [ + { + "bbox": [ + 316, + 395, + 553, + 426 + ], + "type": "text", + "content": "[22] Elad Hoffer, Tal Ben-Nun, Itay Hubara, Niv Giladi, Torsten Hoefler, and Daniel Soudry. Augment your batch: better training with larger batches. arXiv:1901.09335, 2019. 3" + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 316, + 429, + 553, + 473 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 429, + 553, + 473 + ], + "spans": [ + { + "bbox": [ + 316, + 429, + 553, + 473 + ], + "type": "text", + "content": "[23] Saihui Hou, Xinyu Pan, Chen Change Loy, Zilei Wang, and Dahua Lin. Learning a unified classifier incrementally via rebalancing. In Conference on Computer Vision and Pattern Recognition, 2019. 6" + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 317, + 475, + 553, + 528 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 475, + 553, + 528 + ], + "spans": [ + { + "bbox": [ + 317, + 475, + 553, + 528 + ], + "type": "text", + "content": "[24] Bingchen Huang, Zhineng Chen, Peng Zhou, Jiayin Chen, and Zuxuan Wu. Resolving task confusion in dynamic expansion architectures for class incremental learning. In Association for the Advancement of Artificial Intelligence, 2023. 2, 3, 6" + } + ] + } + ], + "index": 25 + }, + { + "bbox": [ + 317, + 532, + 553, + 586 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 532, + 553, + 586 + ], + "spans": [ + { + "bbox": [ + 317, + 532, + 553, + 586 + ], + "type": "text", + "content": "[25] Anette Hunziker, Yuxin Chen, Oisin Mac Aodha, Manuel Gomez Rodriguez, Andreas Krause, Pietro Perona, Yisong Yue, and Adish Singla. Teaching multiple concepts to a forgetful learner. Advances in Neural Information Processing Systems, 2019. 2" + } + ] + } + ], + "index": 26 + }, + { + "bbox": [ + 316, + 589, + 553, + 620 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 589, + 553, + 620 + ], + "spans": [ + { + "bbox": [ + 316, + 589, + 553, + 620 + ], + "type": "text", + "content": "[26] Szabolcs Káli and Peter Dayan. Replay, repair and consolidation. Advances in Neural Information Processing Systems, 2002. 2" + } + ] + } + ], + "index": 27 + }, + { + "bbox": [ + 316, + 624, + 553, + 655 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 624, + 553, + 655 + ], + "spans": [ + { + "bbox": [ + 316, + 624, + 553, + 655 + ], + "type": "text", + "content": "[27] Sean HK Kang. Spaced repetition promotes efficient and effective learning: Policy implications for instruction. *Policy Insights from the Behavioral and Brain Sciences*, 2016. 1" + } + ] + } + ], + "index": 28 + }, + { + "bbox": [ + 317, + 658, + 555, + 712 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 658, + 555, + 712 + ], + "spans": [ + { + "bbox": [ + 317, + 658, + 555, + 712 + ], + "type": "text", + "content": "[28] Jeffrey D Karpicke and Henry L Roediger III. Expanding retrieval practice promotes short-term retention, but equally spaced retrieval enhances long-term retention. Journal of experimental psychology: learning, memory, and cognition, 2007. 1" + } + ] + } + ], + "index": 29 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "text", + "content": "10085" + } + ] + } + ], + "index": 31 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 8 + }, + { + "para_blocks": [ + { + "bbox": [ + 55, + 72, + 295, + 713 + ], + "type": "list", + "angle": 0, + "index": 15, + "blocks": [ + { + "bbox": [ + 56, + 72, + 294, + 106 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 72, + 294, + 106 + ], + "spans": [ + { + "bbox": [ + 56, + 72, + 294, + 106 + ], + "type": "text", + "content": "[29] Zhizhong Li and Derek Hoiem. Learning without forgetting. IEEE transactions on pattern analysis and machine intelligence, 2017. 3" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 55, + 107, + 294, + 129 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 107, + 294, + 129 + ], + "spans": [ + { + "bbox": [ + 55, + 107, + 294, + 129 + ], + "type": "text", + "content": "[30] Cecil Alec Mace. The psychology of study. McBride, 1932. 1" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 57, + 130, + 295, + 163 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 130, + 295, + 163 + ], + "spans": [ + { + "bbox": [ + 57, + 130, + 295, + 163 + ], + "type": "text", + "content": "[31] Arthur W Melton. The situation with respect to the spacing of repetitions and memory. Journal of Verbal Learning and Verbal Behavior, 1970. 1" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 57, + 164, + 295, + 208 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 164, + 295, + 208 + ], + "spans": [ + { + "bbox": [ + 57, + 164, + 295, + 208 + ], + "type": "text", + "content": "[32] Sylvestre-Alvise Rebuffi, Alexander Kolesnikov, Georg Sperl, and Christoph H Lampert. icarl: Incremental classifier and representation learning. In Conference on Computer Vision and Pattern Recognition, 2017. 2, 3, 6" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 57, + 209, + 294, + 262 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 209, + 294, + 262 + ], + "spans": [ + { + "bbox": [ + 57, + 209, + 294, + 262 + ], + "type": "text", + "content": "[33] Matthew Riemer, Ignacio Cases, Robert Ajemian, Miao Liu, Irina Rish, Yuhai Tu, , and Gerald Tesauro. Learning to learn without forgetting by maximizing transfer and minimizing interference. In International Conference on Learning Representations, 2019. 3" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 57, + 264, + 294, + 308 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 264, + 294, + 308 + ], + "spans": [ + { + "bbox": [ + 57, + 264, + 294, + 308 + ], + "type": "text", + "content": "[34] David Rolnick, Arun Ahuja, Jonathan Schwarz, Timothy Lillicrap, and Gregory Wayne. Experience replay for continual learning. Advances in Neural Information Processing Systems, 2019. 6" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 57, + 310, + 294, + 354 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 310, + 294, + 354 + ], + "spans": [ + { + "bbox": [ + 57, + 310, + 294, + 354 + ], + "type": "text", + "content": "[35] Joan Serra, Didac Suris, Marius Miron, and Alexandros Karatzoglou. Overcoming catastrophic forgetting with hard attention to the task. In International Conference on Machine Learning, 2018. 3" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 57, + 355, + 294, + 388 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 355, + 294, + 388 + ], + "spans": [ + { + "bbox": [ + 57, + 355, + 294, + 388 + ], + "type": "text", + "content": "[36] John J Shaughnessy. Long-term retention and the spacing effect in free-recall and frequency judgments. The American Journal of Psychology, 1977. 1" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 57, + 388, + 294, + 422 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 388, + 294, + 422 + ], + "spans": [ + { + "bbox": [ + 57, + 388, + 294, + 422 + ], + "type": "text", + "content": "[37] Hanul Shin, Jung Kwon Lee, Jaehong Kim, and Jiwon Kim. Continual learning with deep generative replay. Advances in Neural Information Processing Systems, 2017. 3" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 57, + 423, + 295, + 488 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 423, + 295, + 488 + ], + "spans": [ + { + "bbox": [ + 57, + 423, + 295, + 488 + ], + "type": "text", + "content": "[38] James Seale Smith, Leonid Karlinsky, Vyshnavi Gutta, Paola Cascante-Bonilla, Donghyun Kim, Assaf Arbelle, Rameswar Panda, Rogerio Feris, and Zsolt Kira. Coda-prompt: Continual decomposed attention-based prompting for rehearsal-free continual learning. In Conference on Computer Vision and Pattern Recognition, 2023. 6" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 57, + 490, + 295, + 533 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 490, + 295, + 533 + ], + "spans": [ + { + "bbox": [ + 57, + 490, + 295, + 533 + ], + "type": "text", + "content": "[39] Kushal Tirumala, Aram Markosyan, Luke Zettlemoyer, and Armen Aghajanyan. Memorization without overfitting: Analyzing the training dynamics of large language models. Advances in Neural Information Processing Systems, 2022. 2" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 57, + 535, + 295, + 578 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 535, + 295, + 578 + ], + "spans": [ + { + "bbox": [ + 57, + 535, + 295, + 578 + ], + "type": "text", + "content": "[40] Yabin Wang, Zhiwu Huang, and Xiaopeng Hong. S-prompts learning with pre-trained transformers: An occam's razor for domain incremental learning. Advances in Neural Information Processing Systems, 2022. 6" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 57, + 579, + 295, + 635 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 579, + 295, + 635 + ], + "spans": [ + { + "bbox": [ + 57, + 579, + 295, + 635 + ], + "type": "text", + "content": "[41] Zifeng Wang, Zizhao Zhang, Sayna Ebrahimi, Ruoxi Sun, Han Zhang, Chen-Yu Lee, Xiaoqi Ren, Guolong Su, Vincent Perot, Jennifer Dy, et al. Dualprompt: Complementary prompting for rehearsal-free continual learning. In European Conference on Computer Vision, 2022. 3, 6" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 57, + 635, + 295, + 689 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 635, + 295, + 689 + ], + "spans": [ + { + "bbox": [ + 57, + 635, + 295, + 689 + ], + "type": "text", + "content": "[42] Zifeng Wang, Zizhao Zhang, Chen-Yu Lee, Han Zhang, Ruoxi Sun, Xiaqi Ren, Guolong Su, Vincent Perot, Jennifer Dy, and Tomas Pfister. Learning to prompt for continual learning. In Conference on Computer Vision and Pattern Recognition, 2022. 3, 6, 7, 8" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 57, + 691, + 295, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 691, + 295, + 713 + ], + "spans": [ + { + "bbox": [ + 57, + 691, + 295, + 713 + ], + "type": "text", + "content": "[43] Yue Wu, Yinpeng Chen, Lijuan Wang, Yuancheng Ye, Zicheng Liu, Yandong Guo, and Yun Fu. Large scale in" + } + ] + } + ], + "index": 14 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 316, + 73, + 553, + 330 + ], + "type": "list", + "angle": 0, + "index": 22, + "blocks": [ + { + "bbox": [ + 335, + 73, + 553, + 95 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 335, + 73, + 553, + 95 + ], + "spans": [ + { + "bbox": [ + 335, + 73, + 553, + 95 + ], + "type": "text", + "content": "cremental learning. In Conference on Computer Vision and Pattern Recognition, 2019. 6" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 316, + 96, + 553, + 140 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 96, + 553, + 140 + ], + "spans": [ + { + "bbox": [ + 316, + 96, + 553, + 140 + ], + "type": "text", + "content": "[44] Shipeng Yan, Jiangwei Xie, and Xuming He. Der: Dynamically expandable representation for class incremental learning. In Conference on Computer Vision and Pattern Recognition, 2021. 3, 6" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 316, + 141, + 553, + 194 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 141, + 553, + 194 + ], + "spans": [ + { + "bbox": [ + 316, + 141, + 553, + 194 + ], + "type": "text", + "content": "[45] Sangdoo Yun, Dongyoon Han, Seong Joon Oh, Sanghyuk Chun, Junsuk Choe, and Youngjoon Yoo. Cutmix: Regularization strategy to train strong classifiers with localizable features. In International Conference on Computer Vision, 2019. 2" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 316, + 197, + 553, + 240 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 197, + 553, + 240 + ], + "spans": [ + { + "bbox": [ + 316, + 197, + 553, + 240 + ], + "type": "text", + "content": "[46] Gengwei Zhang, Liyuan Wang, Guoliang Kang, Ling Chen, and Yunchao Wei. Slca: Slow learner with classifier alignment for continual learning on a pre-trained model. In International Conference on Computer Vision, 2023. 3, 6, 7" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 316, + 241, + 553, + 285 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 241, + 553, + 285 + ], + "spans": [ + { + "bbox": [ + 316, + 241, + 553, + 285 + ], + "type": "text", + "content": "[47] Bowen Zhao, Xi Xiao, Guojun Gan, Bin Zhang, and ShuTao Xia. Maintaining discrimination and fairness in class incremental learning. In Conference on Computer Vision and Pattern Recognition, 2020. 6" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 316, + 286, + 553, + 330 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 286, + 553, + 330 + ], + "spans": [ + { + "bbox": [ + 316, + 286, + 553, + 330 + ], + "type": "text", + "content": "[48] Wanjun Zhong, Lianghong Guo, Qiqi Gao, He Ye, and Yanlin Wang. Memorybank: Enhancing large language models with long-term memory. In Association for the Advancement of Artificial Intelligence, 2024. 1, 2" + } + ] + } + ], + "index": 21 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 749, + 318, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 749, + 318, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 749, + 318, + 757 + ], + "type": "text", + "content": "10086" + } + ] + } + ], + "index": 23 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 9 + } + ], + "_backend": "vlm", + "_version_name": "2.6.4" +} \ No newline at end of file diff --git a/2025/DoF-Gaussian_ Controllable Depth-of-Field for 3D Gaussian Splatting/1ef9852f-e1e4-4571-8156-b74884315f2d_content_list.json b/2025/DoF-Gaussian_ Controllable Depth-of-Field for 3D Gaussian Splatting/1ef9852f-e1e4-4571-8156-b74884315f2d_content_list.json new file mode 100644 index 0000000000000000000000000000000000000000..dc0eea4cbb58f66a7821c35f92a1bb64849982d7 --- /dev/null +++ b/2025/DoF-Gaussian_ Controllable Depth-of-Field for 3D Gaussian Splatting/1ef9852f-e1e4-4571-8156-b74884315f2d_content_list.json @@ -0,0 +1,1770 @@ +[ + { + "type": "text", + "text": "DoF-Gaussian: Controllable Depth-of-Field for 3D Gaussian Splatting", + "text_level": 1, + "bbox": [ + 143, + 130, + 852, + 152 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Liao Shen $^{1,2}$ Tianqi Liu $^{1,2}$ Huiqiang Sun $^{1,2}$ Jiaqi Li $^{1}$ Zhiguo Cao $^{1}$ Wei Li $^{2*}$ Chen Change Loy $^{2}$ $^{1}$ School of AIA, Huazhong University of Science and Technology \n $^{2}$ S-Lab, Nanyang Technological University", + "bbox": [ + 99, + 179, + 915, + 234 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "leoshen@hust.edu.cn", + "bbox": [ + 421, + 237, + 594, + 250 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/96b19129ee7253db8f2beb70a89972b0937c4e8dc4a369d8af6ad4fdd2897bb3.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 91, + 263, + 437, + 440 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/ef42b940e1b3a757249e06aeb98210b90570586c95ef8ebd901c80abcc82ae68.jpg", + "image_caption": [ + "DoF-NeRF" + ], + "image_footnote": [], + "bbox": [ + 94, + 441, + 207, + 501 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/d8a776f82313fb89e47962f106e53fd310f39b1a5a09d02c0e52deee350decf2.jpg", + "image_caption": [ + "BAGS" + ], + "image_footnote": [], + "bbox": [ + 210, + 441, + 321, + 501 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/4b3768a91c66cecb078a2c7eff329ba6b4116e9a9db37c239b2935f8100ee7c1.jpg", + "image_caption": [ + "Ours" + ], + "image_footnote": [], + "bbox": [ + 323, + 441, + 437, + 501 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/e3b2ffe15ba989b5ac1a70e607113ff8e23daa3a4026a532b16f75a665645513.jpg", + "image_caption": [ + "Change aperture size and focus distance" + ], + "image_footnote": [], + "bbox": [ + 455, + 263, + 679, + 375 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/becc542fe2a2639b9769c7749c819fe5ab65a7cd6ac1e9808f67cc32958b692a.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 681, + 263, + 903, + 375 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/e599fd807f897a202c19e0a47d1a2ef608ad32fe2903fbe3d3bc71e6c99f9b3f.jpg", + "image_caption": [ + "Change camera poses and zoom-in", + "Figure 1. Given a set of multi-view input images with shallow DoF, DoF-Gaussian can reconstruct a 3D-GS representation of a sharp scene. Thanks to our lens-based design, we can also achieve controllable DoF effects for a variety of applications. The example input images are taken from [55] for illustration purposes. (Zoom-in for best view)" + ], + "image_footnote": [], + "bbox": [ + 455, + 388, + 679, + 500 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/14bd89e418f7a6774affdb731d0a4895c49b187f124105341b3b097982a2a1d6.jpg", + "image_caption": [ + "Change aperture shape" + ], + "image_footnote": [], + "bbox": [ + 681, + 388, + 830, + 500 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Abstract", + "text_level": 1, + "bbox": [ + 246, + 575, + 326, + 590 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Recent advances in 3D Gaussian Splatting (3D-GS) have shown remarkable success in representing 3D scenes and generating high-quality, novel views in real-time. However, 3D-GS and its variants assume that input images are captured based on pinhole imaging and are fully in focus. This assumption limits their applicability, as real-world images often feature shallow depth-of-field (DoF). In this paper, we introduce DoF-Gaussian, a controllable depth-of-field method for 3D-GS. We develop a lens-based imaging model based on geometric optics principles to control DoF effects. To ensure accurate scene geometry, we incorporate depth priors adjusted per scene, and we apply defocus-to-focus adaptation to minimize the gap in the circle of confusion. We also introduce a synthetic dataset to assess refocusing capabilities and the model's ability to learn precise lens parameters. Our framework is customizable and supports various interactive applications. Extensive experiments confirm the effectiveness of our method. Our project is avail-", + "bbox": [ + 88, + 606, + 485, + 878 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "able at https://dof-gaussian.github.io/.", + "bbox": [ + 513, + 577, + 767, + 592 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "1. Introduction", + "text_level": 1, + "bbox": [ + 513, + 618, + 643, + 633 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Depth-of-field (DoF) refers to the distance between the closest and farthest objects in a photo that appears acceptably sharp. In practice, photographers can control the DoF by adjusting the camera's aperture size or focus distance to capture images with either a wide DoF (all-in-focus) or shallow DoF (defocused). The shallow DoF effect is an important technique in photography, as it draws the viewers' attention to the focal region by blurring the surrounding areas. Meanwhile, novel view synthesis aims to create realistic images from novel viewpoints based on a set of source images. However, novel view synthesis typically requires all-in-focus input images and lacks the capability to render varied DoF effects, limiting its applications. In this work, we aim to render novel views with controllable DoF, adding cinematic quality to the results.", + "bbox": [ + 509, + 643, + 906, + 869 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Images captured from the real world sometimes have shallow DoF. Specifically, points of light that do not lie on", + "bbox": [ + 511, + 869, + 906, + 901 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "CVF", + "bbox": [ + 106, + 2, + 181, + 42 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore.", + "bbox": [ + 236, + 0, + 810, + 44 + ], + "page_idx": 0 + }, + { + "type": "page_footnote", + "text": "*Corresponding author.", + "bbox": [ + 109, + 887, + 235, + 898 + ], + "page_idx": 0 + }, + { + "type": "page_number", + "text": "26462", + "bbox": [ + 478, + 944, + 517, + 955 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "the focal plane are projected onto the sensor plane as blurred circles, referred to as the circle of confusion (CoC), and this introduces bokeh blur to the captured image. Most novel view synthesis methods experience performance degradation when processing shallow DoF input images. To address this issue, DoF-NeRF [55] and LensNeRF [17] introduce lens-based camera models in volume rendering to explicitly enable controllable DoF effects. Meanwhile, methods like Deblur-NeRF [26], DP-NeRF [19], and PDRF [33] employ sparse kernels to model defocus blur. However, implicit rendering approaches face significant challenges in training and rendering efficiency. In the realm of 3D Gaussian Splatting (3D-GS) [16], methods such as BAGS [34] and Deblurring 3DGS [18] propose blur estimation networks to model the blur kernels or scaling factors for defocus deblurring, but they cannot accommodate controllable DoF.", + "bbox": [ + 89, + 90, + 480, + 330 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "In this paper, we present DoF-Gaussian, an efficient framework for controllable DoF in 3D-GS, addressing the limitations of existing methods in handling shallow DoF inputs. This problem is non-trivial because 3D-GS and its variants [11, 23, 54, 57, 58] are typically based on a pin-hole camera model that assumes all-in-focus inputs, meaning both foreground and background appear clear. When input images contain bokeh blur, it becomes challenging to accurately construct scene geometry and render depth maps. However, rendering images with different DoFs relies heavily on precise depth estimation. Additionally, our 3D-GS approach assumes an idealized CoC, which inevitably differs from the CoC seen in real photographs, posing further challenges for accurate defocus deblurring. Furthermore, existing datasets in this field are primarily borrowed from deblurring applications, limiting our evaluation to the model's defocus deblurring capability and overlooking other aspects of controllable DoF effects.", + "bbox": [ + 91, + 335, + 482, + 606 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "To address these challenges, we present a new controllable DoF 3D-GS method with the following contributions: First, we employ a lens-based model instead of a pinhole-based one. We make the lens optical parameters, such as aperture size and focus distance, learnable to enable control over the depth of field. Second, we introduce per-scene adjustments for depth priors to minimize errors in scene geometry and depth relationships, ensuring accurate depth-of-field control. Third, we propose a defocus-to-focus adaptation strategy that focuses on learning the focal region after the learnable lens parameters have converged, compensating for differences in CoC. Finally, we present a synthetic dataset designed to comprehensively evaluate the model, including its refocusing capabilities and its ability to learn accurate aperture size and focus distance.", + "bbox": [ + 89, + 611, + 482, + 835 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Benefiting from the lens-based imaging model, our approach enables controlled depth of field, allowing it to handle not only shallow DoF image inputs but also unlock a range of interactive and engaging applications, as depicted", + "bbox": [ + 89, + 839, + 482, + 901 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "in Figure 1. For instance, users can render novel view images with varied DoF, refocus on custom datasets, and even adjust the shape of CoC, e.g., from circular to pentagonal or hexagonal CoC. Moreover, users can dynamically change the DoF while moving the camera or zooming, rendering videos with cinematic effects. Extensive experiments demonstrate that our method outperforms state-of-the-art depth-of-field and defocus deblurring methods.", + "bbox": [ + 511, + 90, + 903, + 212 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "2. Related Work", + "text_level": 1, + "bbox": [ + 511, + 224, + 653, + 239 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Depth-of-Field Rendering. Rendering DoF effects from a single all-in-focus image has been widely explored in previous work. Bokeh blur occurs when light is projected onto the image plane as a circular region, rather than as a point. The size of CoC is affected by the diameter of the aperture (aperture size) and the distance from the camera to the focal plane (focus distance). Photos captured with a small aperture usually present a wide DoF, i.e., all objects appear sharp. In contrast, as the aperture diameter increases, objects near the focal plane remain sharp, while those farther away become blurred with a larger CoC. Physically based methods [1, 20] rely on 3D scene geometry information and are time-consuming. Some methods use neural networks [8, 12–14, 38] trained end-to-end to obtain shallow DoF images. The DoF effects in some studies [6, 25, 35–37, 43, 45, 48, 56, 61] are controllable but usually require an extra disparity map. Wang et al. [46] combines neural fields with an expressive camera model to achieve all-in-focus reconstruction from an image stack. In this work, we show the connection between DoF rendering and novel view synthesis using a lens-based camera model.", + "bbox": [ + 511, + 251, + 903, + 566 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Image Deblurring. Blur can generally be categorized into two main types: camera motion blur and defocus blur. Previous studies [7, 18, 24, 26, 33, 34, 44] attempt to address this issue. Our task of rendering all-in-focus and sharp novel view images from shallow DoF inputs can be considered as a form of defocus deblurring. By modeling the underlying physical principles, we can achieve defocus deblurring by simulating the circle-of-confusion. Furthermore, our method allows for adjustments to lens parameters, enabling applications such as refocusing—capabilities that deblurring approaches like BAGS [34] cannot achieve. Existing datasets in the 3D DoF field are often borrowed from the deblurring domain [26] or designed solely to assess defocus deblurring ability [55]. To support a wider range of applications and validate the accuracy of the learned lens model, we introduce a synthetic dataset for more comprehensive evaluations.", + "bbox": [ + 511, + 568, + 903, + 823 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Novel View Synthesis. Novel view synthesis allows the rendering of unseen camera perspectives from 2D images and their corresponding camera poses. Recent advancements in synthesis can largely be attributed to Neural Radiance Field [28] and 3D Gaussian Splatting [16]. However,", + "bbox": [ + 511, + 824, + 903, + 900 + ], + "page_idx": 1 + }, + { + "type": "page_number", + "text": "26463", + "bbox": [ + 478, + 944, + 517, + 955 + ], + "page_idx": 1 + }, + { + "type": "image", + "img_path": "images/5cc941806fcebb8e511318e455f0cce1f991551e558ae064a5f20dc3492106b5.jpg", + "image_caption": [ + "Figure 2. Overview of DoF-Gaussian. Given input images $I$ with shallow DoF, we first apply SfM from COLMAP to obtain sparse depth $D_{sparse}$ , which is used to train a depth network to derive per-scene depth priors $D_{pred}$ . We then employ $D_{pred}$ to regularize the Gaussians rendered depth map $D$ . Next, by developing a lens imaging model, we can render defocused images $C^*$ to simulate input images. To minimize the discrepancy in CoC, we propose an adaptation using the weight map. Finally, we can render fully clear images for novel view synthesis and achieve various effects by our controllable DoF framework." + ], + "image_footnote": [], + "bbox": [ + 94, + 85, + 908, + 367 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "real-world images are often not clean and well-calibrated. Previous work [4, 27, 29, 31, 42, 44, 49, 54, 62] attempts to cope with non-ideal conditions, such as low light, camera motion, bokeh effects, and various types of image degradation. To recover a sharp scene from images with bokeh blur, methods like Deblur-NeRF [26], DP-NeRF [19], and PDRF [33] model defocus blur using sparse kernels. NeRF Focus [50] and LensNeRF [17] achieve defocus effects by incorporating a lens-based camera model in volume rendering. While DoF-NeRF [55] is similar to our approach, it is limited by the training and rendering efficiency and quality of NeRF. BAGS [34] and Deblurring 3DGS [18] use blur estimation networks to model blur kernels or scaling factors but cannot achieve controllable DoF effects due to the absence of a lens imaging model. Other studies, such as [15, 47], focus on HDR tasks and consider depth-of-field simultaneously. RGS [7] addresses defocus blur by introducing an offset to the 2D covariance matrices of Gaussians. [30] introduces a ray tracing rendering algorithm for particle-based representations to enable many advanced techniques, such as the shallow depth-of-field effect. A concurrent work [51] also explores adjustable depth of field. Our method differs in its controllable DoF mechanism, with the introduction of depth priors to ensure accurate scene depth. In addition, our new defocus-to-focus adaptation strategy eliminates the need for the focus localization network proposed in their work. Besides, we introduce a synthetic dataset for comprehensive evaluation.", + "bbox": [ + 91, + 467, + 483, + 891 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3. Preliminary", + "text_level": 1, + "bbox": [ + 513, + 465, + 638, + 483 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3D Gaussian Splitting represents a 3D scene as a mixture of anisotropic 3D Gaussians, where each Gaussian is characterized by a 3D covariance matrix $\\Sigma$ and mean $\\mu$ :", + "bbox": [ + 511, + 491, + 906, + 537 + ], + "page_idx": 2 + }, + { + "type": "equation", + "text": "\n$$\nG (X) = e ^ {- \\frac {1}{2} (X - \\mu) ^ {\\top} \\Sigma^ {- 1} (X - \\mu)} \\tag {1}\n$$\n", + "text_format": "latex", + "bbox": [ + 604, + 542, + 906, + 564 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "The covariance matrix $\\Sigma$ holds physical meaning only when it is positive semi-definite. Therefore, to enable effective optimization via gradient descent, $\\Sigma$ is decomposed into a scaling matrix $S$ and a rotation matrix $R$ , as:", + "bbox": [ + 511, + 569, + 906, + 628 + ], + "page_idx": 2 + }, + { + "type": "equation", + "text": "\n$$\n\\Sigma = R S S ^ {\\mathsf {T}} R ^ {\\mathsf {T}}. \\tag {2}\n$$\n", + "text_format": "latex", + "bbox": [ + 653, + 635, + 903, + 652 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "To splot Gaussians from 3D space to a 2D plane, the view transformation matrix $W$ and the Jacobian matrix $J$ , which represents the affine approximation of the projective transformation, are utilized to obtain the covariance matrix $\\Sigma'$ in 2D space, as:", + "bbox": [ + 511, + 659, + 906, + 734 + ], + "page_idx": 2 + }, + { + "type": "equation", + "text": "\n$$\n\\Sigma^ {\\prime} = J W \\Sigma W ^ {\\mathsf {T}} J ^ {\\mathsf {T}}. \\tag {3}\n$$\n", + "text_format": "latex", + "bbox": [ + 640, + 733, + 903, + 752 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Subsequently, a point-based alpha-blending rendering can be performed to determine the color of each pixel:", + "bbox": [ + 511, + 755, + 906, + 786 + ], + "page_idx": 2 + }, + { + "type": "equation", + "text": "\n$$\nC = \\sum_ {i} c _ {i} \\alpha_ {i} \\prod_ {j = 1} ^ {i - 1} (1 - \\alpha_ {i}), \\tag {4}\n$$\n", + "text_format": "latex", + "bbox": [ + 617, + 792, + 903, + 834 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "where $c_{i}$ represents the color of each point, defined by spherical harmonics (SH) coefficients. The density $\\alpha_{i}$ is computed as the product of 2D Gaussians and a learnable point-wise opacity.", + "bbox": [ + 511, + 840, + 908, + 902 + ], + "page_idx": 2 + }, + { + "type": "page_number", + "text": "26464", + "bbox": [ + 478, + 944, + 519, + 957 + ], + "page_idx": 2 + }, + { + "type": "image", + "img_path": "images/50236d68e6607c558b614534d5b1607caac7d623b4b0319474552c14335540d6.jpg", + "image_caption": [ + "(I) Pinhole-based model" + ], + "image_footnote": [], + "bbox": [ + 91, + 85, + 277, + 196 + ], + "page_idx": 3 + }, + { + "type": "image", + "img_path": "images/2ffeb0f907d5bb8b063ab3a6b31f0b73dfdfac3903d1bf989a548bb215a6c2b0.jpg", + "image_caption": [ + "(II) Lens-based model", + "Figure 3. The difference between the pinhole-based model (I) and the lens-based model (II). For (I), light emitted from spatial point $O$ directly hits the point $O'$ of the image plane. For (II), we show the case that light emitted from $Q$ converges on the point $Q'$ of the image plane (In-Focus), and the case that light emitted from $P$ converges on the point $P'$ and continues to scatter onto the image plane forming a circle of confusion (Out-of-Focus)." + ], + "image_footnote": [], + "bbox": [ + 282, + 87, + 483, + 199 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "4. Method", + "text_level": 1, + "bbox": [ + 89, + 339, + 181, + 354 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "We propose DoF-Gaussian, a controllable DoF approach for 3D-GS. The overall pipeline is illustrated in Fig. 2. In Section 4.1, we first develop a lens-based imaging model based on geometric optics principles to control DoF effects. Next, we employ the per-scene adjustment of depth priors to guide the correct scene geometry, as described in Section 4.2. Finally, in Section 4.3, we apply defocus-to-focus adaptation to minimize the inconsistencies in the CoC and enhance the defocus deblurring performance.", + "bbox": [ + 89, + 364, + 483, + 501 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "4.1. Lens-based Depth-of-Field Model", + "text_level": 1, + "bbox": [ + 89, + 510, + 383, + 526 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "The physical principles underlying imaging and DoF have been extensively studied in the field of geometric optics [10, 32]. In an idealized optical system, the light emitted from a spatial point $O$ is projected onto a corresponding point $O^{\\prime}$ in the image plane following the principles of pinhole imaging, as illustrated in Fig. 3 (I). However, real-world cameras operate based on the lens model, as shown in Fig. 3 (II). We show the two cases of in-focus and out-of-focus imaging, respectively. A spatial point $P$ , located at a distance $d$ from the lens, is projected on the imaging plane as a circular region referred to as the circle of confusion (CoC). The diameter of this region, $r(d)$ , can be determined by the aperture parameter $\\mathcal{A}$ and focus distance $\\mathcal{F}$ , according to the following equation:", + "bbox": [ + 89, + 532, + 483, + 744 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\nr (d) = \\mathcal {A} \\left| \\frac {1}{\\mathcal {F}} - \\frac {1}{d} \\right|. \\tag {5}\n$$\n", + "text_format": "latex", + "bbox": [ + 218, + 755, + 482, + 789 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "The focus distance $\\mathcal{F}$ primarily controls the depth position of the focal plane of the image, i.e., the depth of the sharp area, while the aperture parameter $\\mathcal{A}$ determines the extent of the bokeh effect. For the point $Q$ located at the focus distance $d = \\mathcal{F}$ , the emitted rays directly converge through the lens to the corresponding point $Q'$ in the image plane, resulting in the absence of a circle of confusion", + "bbox": [ + 89, + 794, + 483, + 900 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Algorithm 1 Lens-based imaging process", + "text_level": 1, + "bbox": [ + 514, + 90, + 792, + 106 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Input: Rasterization rendering $C$ , aperture parameter $\\mathcal{A}$ , focus distance $\\mathcal{F}$ , depth map $D$ , gamma value $\\gamma$ , confuse function Func", + "bbox": [ + 514, + 109, + 905, + 152 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Output: Defocus simulated result $C^*$", + "bbox": [ + 514, + 155, + 767, + 169 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\nR \\leftarrow \\mathcal {A} \\left| \\frac {1}{\\bar {\\mathcal {E}}} - \\frac {1}{\\bar {D}} \\right| \\quad C \\leftarrow (C) ^ {\\gamma}\n$$\n", + "text_format": "latex", + "bbox": [ + 531, + 169, + 733, + 186 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\Phi = [ 0 ], \\tilde {C ^ {*}} = \\tilde {[} 0 ]\n$$\n", + "text_format": "latex", + "bbox": [ + 532, + 186, + 651, + 200 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "for pixel $i\\gets \\mathrm{Traverse}(C)$ do", + "bbox": [ + 531, + 200, + 733, + 215 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "for pixel $j\\gets$ TraverseNeighbor $(c_i,r_i)$ do", + "bbox": [ + 547, + 215, + 836, + 231 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\lambda_ {i j} \\leftarrow \\operatorname {F u n c} (d _ {i}, | i - j |)\n$$\n", + "text_format": "latex", + "bbox": [ + 563, + 231, + 723, + 247 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\Phi_ {j} \\leftarrow \\Phi_ {j} + \\lambda_ {i j}\n$$\n", + "text_format": "latex", + "bbox": [ + 563, + 247, + 669, + 262 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\nc _ {j} ^ {*} \\leftarrow c _ {j} ^ {*} + \\lambda_ {i j} \\cdot c _ {i}\n$$\n", + "text_format": "latex", + "bbox": [ + 563, + 262, + 686, + 277 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "end for", + "bbox": [ + 547, + 277, + 601, + 289 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "end for", + "bbox": [ + 531, + 291, + 584, + 303 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\nC ^ {*} \\leftarrow (C ^ {*} / \\Phi) ^ {\\frac {1}{\\gamma}}\n$$\n", + "text_format": "latex", + "bbox": [ + 531, + 303, + 643, + 321 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "(CoC). This also implies that $r(\\mathcal{F}) = 0$", + "bbox": [ + 513, + 351, + 779, + 366 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "We set focus distance $\\mathcal{F}$ and aperture parameter $\\mathcal{A}$ as learnable for each input image, and these parameters are continuously updated with the optimization of 3D-GS. By modeling the lens camera, the output $C^*$ , which includes bokeh blur effects, can be derived from the 3D-GS rasterizer rendering result $C$ using our CUDA-based Algorithm 1.", + "bbox": [ + 511, + 366, + 905, + 455 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Ideally, the confuse function is represented by the indicator function $\\mathbb{I}(r(d) > l)$ . To achieve a smooth and natural DoF effect, we substitute it with a differentiable function, as proposed in Busam et al. [3]:", + "bbox": [ + 511, + 455, + 905, + 517 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\nF u n c (d, l) = \\frac {1}{2} + \\frac {1}{2} \\tanh (\\alpha (r (d) - l)), \\tag {6}\n$$\n", + "text_format": "latex", + "bbox": [ + 566, + 525, + 905, + 555 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "where $\\alpha$ defines the smoothness of confuse transition and $l$ represents the distance between two pixels.", + "bbox": [ + 511, + 561, + 905, + 592 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "We supervise the training by computing the reconstruction loss between the output image $C^*$ and the shallow DoF input image $I$ , where the loss function is a combination of the $\\mathcal{L}_1$ and a D-SSIM term, with $\\lambda$ set to 0.2:", + "bbox": [ + 511, + 592, + 905, + 652 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {L} _ {r e c} = (1 - \\lambda) \\mathcal {L} _ {1} \\left(I, C ^ {*}\\right) + \\lambda \\mathcal {L} _ {D - S S I M} \\left(I, C ^ {*}\\right). \\tag {7}\n$$\n", + "text_format": "latex", + "bbox": [ + 535, + 662, + 903, + 680 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "For inference, we set the aperture size $\\mathcal{A}$ to 0 so that we can render fully clear images for novel view synthesis.", + "bbox": [ + 511, + 688, + 905, + 719 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "4.2. Per-Scene Adjustment of Depth Priors", + "text_level": 1, + "bbox": [ + 511, + 727, + 844, + 743 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Unlike previous work that assumes input images are all-in-focus and fully clear, it becomes challenging to accurately reconstruct scene geometry and render depth maps for bokeh blurring input images. However, rendering images with varying depths of field relies on precise depth information. To address this issue, we propose using depth prior as useful guidance for the accurate reconstruction of scene geometry. However, due to the presence of bokeh blur in shallow DoF inputs, directly employing a monocular depth network to predict depth maps $D_{pred}$ as pseudo", + "bbox": [ + 511, + 750, + 906, + 901 + ], + "page_idx": 3 + }, + { + "type": "page_number", + "text": "26465", + "bbox": [ + 478, + 944, + 517, + 955 + ], + "page_idx": 3 + }, + { + "type": "image", + "img_path": "images/b533d44be00dafbe9b106e039860d236fe97ff55657c7818f35cf452b64abaa3.jpg", + "image_caption": [ + "RGB", + "BAGS", + "w/o depth priors", + "Ours", + "Figure 4. Effects of the depth priors. The per-scene adjustment of depth priors enhance the geometric structure and yield a more accurate depth map than that estimated by BAGS [34] and a variant of our method without depth priors." + ], + "image_footnote": [], + "bbox": [ + 93, + 99, + 480, + 200 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "ground truth to regularize the Gaussians rendered depth map, as done in [5, 21], does not yield satisfactory results. Inspired by [53], we propose adapting per-scene depth priors through fine-tuning a monocular depth network based on the scene's sparse reconstruction. Specifically, we utilize COLMAP [40, 41] to obtain a sparse Structure-from-Motion (SfM) point cloud and create a set of 3D Gaussians. Consequently, we can derive per-view sparse depth maps $D_{sparse}$ by projecting the 3D point clouds after multiview stereo processing. Although the resulting depth map is sparse, it is robust and can serve as a supervisory signal for training the per-scene depth priors. To address the scale ambiguity inherent in acquired depth maps, we apply a silog loss [9] to fine-tune the depth network [22, 39], expressed as:", + "bbox": [ + 89, + 287, + 483, + 513 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {L} _ {\\text {s i l o g}} = \\frac {1}{2 M} \\sum_ {i = 1} ^ {M} \\left(\\log \\left(e ^ {s} D _ {\\text {p r e d}}\\right) - \\log \\left(D _ {\\text {s p a r s e}}\\right)\\right) ^ {2}, \\tag {8}\n$$\n", + "text_format": "latex", + "bbox": [ + 104, + 522, + 482, + 564 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "where $e^s$ is the scale factor and $M$ is the number of pixels in the image used for calculation. We apply the rasterization technique from Rade-GS [59], which enables the robust rendering of depth maps $D$ for 3D-GS scenes. We predict the depth maps $D_{pred}$ from multi-view input images using the fine-tuned depth network and use them as depth priors to regularize $D$ . The depth loss can be described as follows:", + "bbox": [ + 89, + 573, + 483, + 679 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {L} _ {\\text {d e p t h}} = \\left| \\left| D - D _ {\\text {p r e d}} \\right| \\right| _ {2}. \\tag {9}\n$$\n", + "text_format": "latex", + "bbox": [ + 199, + 690, + 482, + 708 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "As illustrated in Fig. 4, it is evident that the depth priors significantly enhance the geometric structure, resulting in a more accurate depth map compared to both BAGS [34] and the variant of our method without depth priors. In Section 5.3, we further conduct an ablation study comparing several other depth supervision strategies to demonstrate the effectiveness of our approach.", + "bbox": [ + 89, + 717, + 483, + 825 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "4.3. Defocus-to-Focus Adaptation", + "text_level": 1, + "bbox": [ + 89, + 833, + 349, + 849 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "The CoC effect achieved under ideal optical imaging conditions, as described in Algorithm 1, inevitably differs from the real CoC effect produced by an actual DSLR camera.", + "bbox": [ + 89, + 854, + 483, + 900 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "This discrepancy hinders our ability to accurately model the out-of-focus blur in shallow DoF input images.", + "bbox": [ + 511, + 90, + 903, + 121 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "To achieve sharper scenes, we propose an adaptation process that allows transitions from defocus to focus. Specifically, modeling the defocus effects with bokeh blur facilitates learning accurate lens parameters. Consequently, we place greater emphasis on capturing the sharp areas in training images once the lens parameters $\\mathcal{A}$ and $\\mathcal{F}$ have been optimized and converged. This approach is effective because, when the focus distance is determined, the sharp regions in the scene become approximately identifiable. After $t$ optimization iterations, we reweight the $\\mathcal{L}_{rec}$ and $\\mathcal{L}_{depth}$ loss by employing a step-like function:", + "bbox": [ + 511, + 122, + 905, + 287 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\Psi (x) = \\left\\{ \\begin{array}{c c} 1 & , \\text {i t e r a t i o n s} < t \\\\ 1 / \\left(1 + e ^ {- a \\cdot (x - b)}\\right) & , \\text {i t e r a t i o n s} \\geq t \\end{array} \\right. \\tag {10}\n$$\n", + "text_format": "latex", + "bbox": [ + 529, + 297, + 905, + 333 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "where $a$ and $b$ are hyperparameters, and $x = \\left|\\frac{1}{\\mathcal{F}} - \\frac{1}{d}\\right|$ .", + "bbox": [ + 511, + 344, + 870, + 362 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "In addition, the aperture size $\\mathcal{A}$ is generally large after the convergence, which can cause blurring of pixels that should be in focus. Hence, we also reweight the aperture size $\\mathcal{A}$ on a pixel-wise basis:", + "bbox": [ + 511, + 361, + 905, + 421 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {A} ^ {\\prime} = \\mathcal {A} \\cdot \\Psi . \\tag {11}\n$$\n", + "text_format": "latex", + "bbox": [ + 666, + 433, + 903, + 450 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Full Objective. We derive the final loss terms by incorporating the reconstruction loss from Eqn. 7, the depth loss from Eqn. 9, and an additional normal consistency loss as proposed by Huang et al. [11]. The normal consistency loss ensures that the Gaussian splats align with the surface by measuring the consistency between the normal directions computed from the Gaussian and the depth map:", + "bbox": [ + 511, + 462, + 905, + 569 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {L} _ {\\text {n o r m a l}} = \\sum_ {i} w _ {i} \\left(1 - n _ {i} ^ {\\top} \\hat {n} _ {i}\\right), \\tag {12}\n$$\n", + "text_format": "latex", + "bbox": [ + 606, + 580, + 903, + 612 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "where $\\hat{n}$ represents the surface normal direction obtained by applying finite differences on the depth map, $i$ indexes the intersected splats along the ray and $w$ denotes the blending weight of the intersection point.", + "bbox": [ + 511, + 622, + 905, + 681 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Our final training loss $\\mathcal{L}$ is:", + "bbox": [ + 532, + 683, + 715, + 698 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {L} = \\Psi \\odot \\left(\\mathcal {L} _ {\\text {r e c}} + w _ {d} \\mathcal {L} _ {\\text {d e p t h}}\\right) + w _ {n} \\mathcal {L} _ {\\text {n o r m a l}}, \\tag {13}\n$$\n", + "text_format": "latex", + "bbox": [ + 545, + 710, + 903, + 728 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "where $\\odot$ is Hadamard product.", + "bbox": [ + 511, + 739, + 718, + 755 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "5. Experiments", + "text_level": 1, + "bbox": [ + 511, + 768, + 645, + 787 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Implementation Details. In our experiments, we set the smoothness of confuse transition $\\alpha = 4$ . In addition, we set $a = 15$ , $b = 0.3$ , and $t = 10000$ for the defocus-to-focus adaptation. Our method is built upon Mip-Splatting [58], and our optimization strategy and hyperparameter settings remain consistent with it. We train each scene for 30000 iterations and set loss weights $w_{d} = 0.01$ and $w_{n} = 0.05$ .", + "bbox": [ + 511, + 794, + 905, + 900 + ], + "page_idx": 4 + }, + { + "type": "page_number", + "text": "26466", + "bbox": [ + 478, + 944, + 519, + 955 + ], + "page_idx": 4 + }, + { + "type": "table", + "img_path": "images/79c8b25add3113ada87802a6ad836b1b3659d41c21763d83280c90f754571fcd.jpg", + "table_caption": [ + "Table 1. Quantitative comparisons on the defocus blur dataset of Deblur NeRF [26]. We report the PSNR, SSIM, and LPIPS metrics and color each cell as best and second best. Our method outperforms other existing approaches across most scenes." + ], + "table_footnote": [], + "table_body": "
MethodDeblur-NeRF [26]DoF-NeRF [55]DP-NeRF [19]BAGS [34]Deblurring 3DGS [18]Ours
PSNRSSIMLPIPSPSNRSSIMLPIPSPSNRSSIMLPIPSPSNRSSIMLPIPSPSNRSSIMLPIPSPSNRSSIMLPIPS
Cake26.270.7800.12824.530.7310.20626.160.7780.12727.210.8180.10826.880.8030.11626.830.8080.101
Caps23.870.7130.16122.670.6360.26123.950.7120.14324.160.7250.15924.500.7420.14924.620.7490.147
Cisco20.830.7270.08720.640.7240.10720.730.7260.08420.790.7430.07020.830.7320.07921.000.7440.069
Coral19.850.6000.12119.830.5700.24020.110.6110.11820.530.6280.11119.780.6080.13120.370.6300.109
Cupcake22.260.7220.11621.890.7060.14322.800.7410.09622.930.7620.08022.110.7340.09922.970.7570.079
Cups26.210.7990.12725.260.7650.20226.750.8140.10426.270.8230.10426.280.8240.10326.010.8170.100
Daisy23.520.6870.12123.220.6580.19423.790.6970.10823.740.7460.06223.540.7310.09523.930.7350.071
Sausage18.010.5000.18017.860.4880.28018.350.5440.14718.760.5740.11018.990.5700.14119.110.5760.119
Seal26.040.7770.10524.850.6870.14325.950.7780.10326.520.8120.09026.180.8170.09826.570.8250.087
Tools27.810.8950.06126.210.8540.12828.070.8980.05428.600.9130.04627.960.9100.05828.290.9130.051
", + "bbox": [ + 98, + 128, + 898, + 297 + ], + "page_idx": 5 + }, + { + "type": "table", + "img_path": "images/ce5072281c23873e0d90ca790356afc4b28f7638d507e7d8d2468a804e3c793a.jpg", + "table_caption": [ + "Table 2. Quantitative comparisons on our synthetic dataset." + ], + "table_footnote": [], + "table_body": "
MethodPSNR↑SSIM↑LPIPS↓δA↓δF↓
DoF-NeRF [55]25.590.7880.2070.1960.256
Ours28.700.8640.0950.1260.079
", + "bbox": [ + 94, + 347, + 478, + 402 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Baselines and Evaluation Metrics. We compare our method with state-of-the-art deblurring techniques [18, 26, 34] and a depth-of-field method [55]. To quantitatively evaluate the quality of novel view images, we adopt widely used metrics such as PSNR, SSIM [52], and LPIPS [60]. We will compare with Wang et al. [51] once they release the code or provide the full experimental results in their paper.", + "bbox": [ + 89, + 425, + 482, + 530 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Datasets. Following [18, 26, 34], we evaluate our model on the Deblur-NeRF dataset [26]. Subsequently, we conduct further evaluations on the Real Forward-facing dataset [28] and T&T_DB dataset [16]. The effectiveness of our method is validated on these datasets for both shallow DoF inputs and normal inputs. Furthermore, we also introduce a dataset in Section 5.1 to evaluate our model more comprehensively.", + "bbox": [ + 89, + 532, + 483, + 638 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "5.1. A Synthetic Dataset", + "text_level": 1, + "bbox": [ + 89, + 666, + 279, + 681 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "We introduce a synthetic dataset, as illustrated in Fig. 6, based on depth estimation [2] and a single-image DoF rendering method [35] for each image in the Real Forward Facing dataset [28] and T&T_DB dataset [16]. This allows us to evaluate the refocusing ability of our model and to determine whether it accurately learns the correct aperture size and focus distance. In particular, by feeding the DoF rendering method [35] the known aperture size and focus distance, we can convert the all-in-focus images into shallow DoF images. Unlike existing defocus blur datasets proposed by Ma et al. [26] and Wu et al. [55], the test set in our dataset comprises images with shallow DoF. With known lens parameters, our model fits this shallow DoF effect to quantitatively measure the refocusing ability. Meanwhile, we establish the", + "bbox": [ + 89, + 688, + 482, + 900 + ], + "page_idx": 5 + }, + { + "type": "table", + "img_path": "images/7d9c5d5774677e3b5777d01bfcf1b1c8589a980b9c22f6dd6a7fb25fc8ee8684.jpg", + "table_caption": [ + "Table 3. Comparisons in all-in-focus settings. We compare our method with Mip-Splatting on all-in-focus datasets to validate the effectiveness of our model under general input conditions." + ], + "table_footnote": [], + "table_body": "
MethodPSNR↑SSIM↑LPIPS↓
Mip-Splatting [58]27.050.8930.115
Ours27.810.9020.117
", + "bbox": [ + 535, + 375, + 880, + 436 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "known lens parameters beforehand as ground truth, allowing us to calculate the errors $\\delta_{\\mathcal{A}}$ and $\\delta_{\\mathcal{F}}$ between the learned parameters and the ground truth. More details can be found in the supplementary material.", + "bbox": [ + 511, + 458, + 906, + 518 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "5.2. Comparisons", + "text_level": 1, + "bbox": [ + 511, + 529, + 653, + 545 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Quantitative Comparisons. We quantitatively compare DoF-Gaussian against other state-of-the-art methods using a real-world defocus blur dataset [26]. As shown in Table 1, our method outperforms other state-of-the-art baselines. These results suggest that our method achieves superior defocus deblurring performance and produces higher-quality novel view synthesis images.", + "bbox": [ + 511, + 551, + 905, + 657 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Qualitative Comparisons. Visual qualitative comparisons are presented in Fig. 5. Our method surpasses all other methods in terms of image fidelity, generating novel view images that are more faithful to the ground truth images and exhibit less blur. Specifically, the bottle caps and shelves are the sharpest in our rendered images.", + "bbox": [ + 511, + 657, + 905, + 748 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "All-in-focus Inputs. To further validate the effectiveness of our method in the all-in-focus setting, we design an experiment using wide DoF images as inputs. We conduct a comparison between our method and Mip-Splatting [58] on two all-in-focus datasets, the Real Forward Facing dataset [28] and the T&T_DB dataset [16]. As shown in Table 3, our method achieves comparable or even better performance to Mip-Splatting on the average metrics across the datasets. This demonstrates that our lens-based model not only excels with shallow DoF inputs but also achieves good", + "bbox": [ + 511, + 750, + 906, + 900 + ], + "page_idx": 5 + }, + { + "type": "page_number", + "text": "26467", + "bbox": [ + 478, + 944, + 519, + 957 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/afc37f6ecac264bd729b25afea8480afbf9695b641eb6ce660296001e8ecb764.jpg", + "image_caption": [ + "Figure 5. Qualitative comparisons against all baselines. Compared to other state-of-the-art methods, our method represents sharper scenes and generates novel view images with less blur." + ], + "image_footnote": [], + "bbox": [ + 91, + 85, + 906, + 315 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/37b868e3648129b9042355f069a7311dd491a7ed5d3093aa4d8986a139ebc8f1.jpg", + "image_caption": [ + "Figure 6. Examples in our dataset. For each scene, we apply a DoF rendering method [35] to convert all-in-focus images to shallow DoF images with varying focus locations and aperture sizes. This dataset is used to evaluate the model's refocusing ability." + ], + "image_footnote": [], + "bbox": [ + 91, + 368, + 482, + 525 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "results with general input images.", + "bbox": [ + 89, + 612, + 315, + 627 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Our Synthetic Dataset. As shown in Table 2, our method achieves superior refocusing ability through controlled DoF rendering, producing rendered images with the highest quality on our synthetic dataset. Furthermore, our method learns more precise aperture size and focus distance than DoF-NeRF [55]. Since Deblur-NeRF [26], BAGS [34], and Deblurring 3DGS [18] do not incorporate a lens model, they cannot generate novel view images with a specific DoF. As a result, these methods cannot be evaluated on this dataset due to their lack of refocusing capability.", + "bbox": [ + 88, + 627, + 482, + 779 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "5.3. Ablation Study", + "text_level": 1, + "bbox": [ + 89, + 787, + 243, + 804 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "As shown in Table 4 and Fig. 8, we conduct ablation studies to evaluate the effectiveness of our designs. The comparison between #1 and #2 indicates that the lens-based imaging model has a significant impact. This lens model not only facilitates defocus deblurring but also provides the ability to control the depth of field. Next, the defocus-to-focus", + "bbox": [ + 89, + 809, + 482, + 900 + ], + "page_idx": 6 + }, + { + "type": "table", + "img_path": "images/0cb54d37a5e0aa6bffb5e515c8bb2f66ede818e373eab05e3971ea0fa343f02e.jpg", + "table_caption": [ + "Table 4. Ablation study on each component of our method." + ], + "table_footnote": [], + "table_body": "
LensDepthDefocus-to-focusPSNR↑SSIM↑LPIPS↓
#1XXX21.310.6360.239
#2✓XX23.050.7280.109
#3✓X✓23.590.7420.104
#4✓✓X23.420.7380.098
#5✓✓✓23.970.7560.093
", + "bbox": [ + 517, + 396, + 901, + 488 + ], + "page_idx": 6 + }, + { + "type": "table", + "img_path": "images/4b8fc6f9677c389483b5541e2fa50d2536f9b7a9e56c55a2a34f7bca5913f49d.jpg", + "table_caption": [ + "Table 5. Ablation study on different depth priors." + ], + "table_footnote": [], + "table_body": "
MethodPSNR↑SSIM↑LPIPS↓
No fine-tuned depth23.530.7390.114
Sparse depth23.440.7370.113
Ours23.970.7560.093
", + "bbox": [ + 535, + 545, + 879, + 619 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "adaptation (#3) and the per-scene adjustment of depth priors (#4) further enhance the scene geometry and improve the deblurring of details. Ultimately, the combination of all components yields the highest performance gain, leading to a 2.66 dB increase in PSNR over the baseline model.", + "bbox": [ + 511, + 662, + 906, + 739 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "In addition, we conduct an ablation study on different depth supervision methods to demonstrate the superiority of our per-scene adjustment of depth priors, as shown in Table 5. Direct supervision using depth maps predicted by the depth network without scene-specific fine-tuning hinders the reconstruction quality of 3D-GS. Likewise, supervision based on sparse depth maps projected from a 3D point cloud fails to yield optimal results. We show visualizations of depth maps rendered by different depth supervision strategies in the supplementary material.", + "bbox": [ + 511, + 750, + 908, + 902 + ], + "page_idx": 6 + }, + { + "type": "page_number", + "text": "26468", + "bbox": [ + 478, + 944, + 519, + 955 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/c5127cfbbf9def5d270a2dc66c2e97c9df308d7ea4658d695207b9d6918444b3.jpg", + "image_caption": [ + "(I) Change the aperture size and focus distance" + ], + "image_footnote": [], + "bbox": [ + 93, + 104, + 903, + 189 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/99843575058854ca572730dbd9ce3f9674a04e3fb1f82637856ac1f99ed2c6ad.jpg", + "image_caption": [ + "(II) Change the lens parameters and camera poses" + ], + "image_footnote": [], + "bbox": [ + 93, + 208, + 903, + 294 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/99c79dbb4aa2cd977795ab94a7f90038405c9f824840d0801d73de04568ba5a4.jpg", + "image_caption": [ + "(III) Change the lens parameters and zoom-in", + "Figure 7. Applications of our controllable DoF effects. Users can create their own cinematic moments by combining changes in aperture size, focus distance, camera poses, and zoom." + ], + "image_footnote": [], + "bbox": [ + 93, + 311, + 903, + 398 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/d8e23bdec491fec232de644336e6ff64568c5b75af21d1565cb6e86fa943f88e.jpg", + "image_caption": [ + "Figure 8. The visualizations of the ablation study. Qualitative comparisons show that the full model (#5) yields the best result and the reconstructed strawberry is the sharpest." + ], + "image_footnote": [], + "bbox": [ + 93, + 446, + 482, + 609 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "5.4. Interactive Applications", + "text_level": 1, + "bbox": [ + 89, + 681, + 310, + 698 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Thanks to the lens-based model, we can not only represent 3D scenes from shallow DoF inputs but also render interactive DoF effects (see Fig. 7). Various cinematography styles can be achieved through different camera configurations, and users can input custom datasets, whether all-in-focus or bokeh images. By controlling aperture size and focus distance, we can render novel view images with different depths of field. In addition, we can modify the shape of the CoC, such as changing it from a circle to a hexagon. Users can also adjust the depth of field while moving the camera or zooming to create cinematic effects. Unlike [55], our GS framework significantly enhances the training and rendering efficiency, allowing for faster and more seamless inter", + "bbox": [ + 89, + 704, + 483, + 900 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "actions. The details on processing time will be discussed in the supplementary material.", + "bbox": [ + 511, + 450, + 906, + 481 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "6. Conclusion", + "text_level": 1, + "bbox": [ + 511, + 496, + 633, + 512 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "While 3D Gaussian Splatting methods have achieved impressive performance on a wide range of reconstruction tasks, the applications on controllable DoF effects remain challenging and understudied. In this paper, we propose DoF-Gaussian, a controllable DoF method for 3D Gaussian Splatting. Specifically, we develop a lens-based model rather than pinhole imaging to overcome the limitations imposed by shallow DoF inputs for 3D-GS. Furthermore, we propose the per-scene depth adjustment and a defocus-to-focus adaptation to guarantee the performance of defocus deblurring. We also introduce a synthetic dataset for a more comprehensive evaluation. Thanks to the imaging principles, our method supports various interactive applications.", + "bbox": [ + 511, + 522, + 906, + 718 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Future work. Inspired by our experiments, we found that modeling real-world physical imaging principles can enable our method not only to handle shallow DoF inputs but also to perform effectively on general inputs. This insight motivates us to explore a combination of non-ideal conditions, such as shallow DoF inputs, sparse views, and varied lighting environments, which are more aligned with casual photography in our daily lives.", + "bbox": [ + 511, + 719, + 905, + 839 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Acknowledgement. This study is supported under the RIE2020 Industry Alignment Fund - Industry Collaboration Projects (IAF-ICP) Funding Initiative, as well as cash and in-kind contribution from the industry partner(s).", + "bbox": [ + 511, + 839, + 905, + 900 + ], + "page_idx": 7 + }, + { + "type": "page_number", + "text": "26469", + "bbox": [ + 478, + 944, + 519, + 955 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "References", + "text_level": 1, + "bbox": [ + 91, + 89, + 187, + 104 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[1] Guillaume Abadie, Steve McAuley, Evegenii Golubev, Stephen Hill, and Sebastien Lagarde. Advances in real-time rendering in games. In ACM SIGGRAPH 2018 Courses. 2018. 2", + "[2] Aleksei Bochkovskii, Amaël Delaunoy, Hugo Germain, Marcel Santos, Yichao Zhou, Stephan R Richter, and Vladlen Koltun. Depth pro: Sharp monocular metric depth in less than a second. arXiv preprint arXiv:2410.02073, 2024. 6", + "[3] Benjamin Busam, Matthieu Hog, Steven McDonagh, and Gregory Slabaugh. Sterefo: Efficient image refocusing with stereo vision. In ICCVW, 2019. 4", + "[4] Wenbo Chen and Ligang Liu. Deblur-GS: 3D Gaussian Splitting from Camera Motion Blurred Images. PACMCGIT, 7(1), 2024. 3", + "[5] Jaeyoung Chung, Jeongtaek Oh, and Kyoung Mu Lee. Depth-regularized optimization for 3D Gaussian Splitting in few-shot images. In CVPR, 2024. 5", + "[6] Marcos V Conde, Manuel Kolmet, Tim Seizinger, Tom E Bishop, Radu Timofte, Xiangyu Kong, Dafeng Zhang, Jinlong Wu, Fan Wang, Juewen Peng, et al. Lens-to-lens bokeh effect transformation. NTIRE 2023 challenge report. In CVPR, 2023. 2", + "[7] François Darmon, Lorenzo Porzi, Samuel Rota-Bulò, and Peter Kontschieder. Robust gaussian splatting. arXiv preprint arXiv:2404.04211, 2024. 2, 3", + "[8] Saikat Dutta, Sourya Dipta Das, Nisarg A Shah, and Anil Kumar Tiwari. Stacked deep multi-scale hierarchical network for fast bokeh effect rendering from a single image. In CVPR, 2021. 2", + "[9] David Eigen, Christian Puhrsch, and Rob Fergus. Depth map prediction from a single image using a multi-scale deep network. NeurIPS, 27, 2014. 5", + "[10] Eugene Hecht. Optics. Pearson Education India, 2012. 4", + "[11] Binbin Huang, Zehao Yu, Anpei Chen, Andreas Geiger, and Shenghua Gao. 2D Gaussian Splatting for geometrically accurate radiance fields. In ACM SIGGRAPH 2024 Conference Papers, 2024. 2, 5", + "[12] Andrey Ignatov, Jagruti Patel, Radu Timofte, Bolun Zheng, Xin Ye, Li Huang, Xiang Tian, Saikat Dutta, Kuldeep Purohit, Praveen Kandula, et al. Aim 2019 challenge on bokeh effect synthesis: Methods and results. In ICCVW. IEEE, 2019. 2", + "[13] Andrey Ignatov, Jagruti Patel, and Radu Timofte. Rendering natural camera bokeh effect with deep learning. In CVPRW, 2020.", + "[14] Andrey Ignatov, Radu Timofte, Ming Qian, Congyu Qiao, Jiamin Lin, Zhenyu Guo, Chenghua Li, Cong Leng, Jian Cheng, Juewen Peng, et al. AIM 2020 challenge on rendering realistic bokeh. In ECCVW. Springer, 2020. 2", + "[15] Xin Jin, Pengyi Jiao, Zheng-Peng Duan, Xingchao Yang, Chun-Le Guo, Bo Ren, and Chongyi Li. Lighting Every Darkness with 3DGS: Fast Training and Real-Time Rendering for HDR View Synthesis. arXiv preprint arXiv:2406.06216, 2024.3" + ], + "bbox": [ + 93, + 114, + 483, + 898 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[16] Bernhard Kerbl, Georgios Kopanas, Thomas Leimkuhler, and George Drettakis. 3D Gaussian Splatting for Real-Time Radiance Field Rendering. ACM TOG, 42(4), 2023. 2, 6", + "[17] Min-Jung Kim, Gyojung Gu, and Jaegul Choo. LensNeRF: Rethinking volume rendering based on thin-lens camera model. In WACV, 2024. 2, 3", + "[18] Byeonghyeon Lee, Howoong Lee, Xiangyu Sun, Usman Ali, and Eunbyung Park. Deblurring 3D Gaussian Splatting. arXiv preprint arXiv:2401.00834, 2024. 2, 3, 6, 7", + "[19] Dogyoon Lee, Minhyeok Lee, Chajin Shin, and Sangyoun Lee. Dp-NeRF: Deblurred neural radiance field with physical scene priors. In CVPR, 2023. 2, 3, 6", + "[20] Sungkil Lee, Elmar Eisemann, and Hans-Peter Seidel. Realtime lens blur effects and focus control. ACM TOG, 29(4), 2010. 2", + "[21] Jiahe Li, Jiawei Zhang, Xiao Bai, Jin Zheng, Xin Ning, Jun Zhou, and Lin Gu. DNGaussian: Optimizing sparse-view 3D Gaussian radiance fields with global-local depth normalization. In CVPR, 2024. 5", + "[22] Zhengqi Li, Tali Dekel, Forrester Cole, Richard Tucker, Noah Snavely, Ce Liu, and William T Freeman. Learning the depths of moving people by watching frozen people. In CVPR, 2019. 5", + "[23] Tianqi Liu, Guangcong Wang, Shoukang Hu, Liao Shen, Xinyi Ye, Yuhang Zang, Zhiguo Cao, Wei Li, and Ziwei Liu. MVSGaussian: Fast Generalizable Gaussian Splatting Reconstruction from Multi-View Stereo. In ECCV. Springer, 2025. 2", + "[24] Xianrui Luo, Huiqiang Sun, Juewen Peng, and Zhiguo Cao. Dynamic Neural Radiance Field From Defocused Monocular Video. arXiv preprint arXiv:2407.05586, 2024. 2", + "[25] Yawen Luo, Min Shi, Liao Shen, Yachuan Huang, Zixuan Ye, Juewen Peng, and Zhiguo Cao. Video bokeh rendering: Make casual videography cinematic. In ACM MM, 2024. 2", + "[26] Li Ma, Xiaoyu Li, Jing Liao, Qi Zhang, Xuan Wang, Jue Wang, and Pedro V Sander. Deblur-NeRF: Neural radiance fields from blurry images. In CVPR, 2022. 2, 3, 6, 7", + "[27] Ricardo Martin-Brualla, Noha Radwan, Mehdi SM Sajjadi, Jonathan T Barron, Alexey Dosovitskiy, and Daniel Duckworth. NeRF in the wild: Neural radiance fields for unconstrained photo collections. In CVPR, 2021. 3", + "[28] Ben Mildenhall, Pratul P Srinivasan, Matthew Tancik, Jonathan T Barron, Ravi Ramamoorthi, and Ren Ng. NeRF: Representing scenes as neural radiance fields for view synthesis. Communications of the ACM, 65(1), 2021. 2, 6", + "[29] Ben Mildenhall, Peter Hedman, Ricardo Martin-Brualla, Pratul P Srinivasan, and Jonathan T Barron. NeRF in the dark: High dynamic range view synthesis from noisy raw images. In CVPR, 2022. 3", + "[30] Nicolas Moenne-Loccoz, Ashkan Mirzaei, Or Perel, Ricardo de Lutio, Janick Martinez Esturo, Gavriel State, Sanja Fidler, Nicholas Sharp, and Zan Gojcic. 3d gaussian ray tracing: Fast tracing of particle scenes. ACM Transactions on Graphics (TOG), 43(6):1-19, 2024. 3", + "[31] Keunhong Park, Utkarsh Sinha, Jonathan T Barron, Sofien Bouaziz, Dan B Goldman, Steven M Seitz, and Ricardo Martin-Brualla. Nerfies: Deformable neural radiance fields. In ICCV, 2021. 3" + ], + "bbox": [ + 516, + 92, + 903, + 898 + ], + "page_idx": 8 + }, + { + "type": "page_number", + "text": "26470", + "bbox": [ + 478, + 944, + 519, + 955 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[32] Frank L Pedrotti, Leno M Pedrotti, and Leno S Pedrotti. Introduction to optics. Cambridge University Press, 2018. 4", + "[33] Cheng Peng and Rama Chellappa. PDRF: progressively deblurring radiance field for fast scene reconstruction from blurry images. In AAAI, 2023. 2, 3", + "[34] Cheng Peng, Yutao Tang, Yifan Zhou, Nengyu Wang, Xijun Liu, Deming Li, and Rama Chellappa. BAGS: Blur Agnostic Gaussian Splatting through Multi-Scale Kernel Modeling. arXiv preprint arXiv:2403.04926, 2024. 2, 3, 5, 6, 7", + "[35] Juewen Peng, Zhiguo Cao, Xianrui Luo, Hao Lu, Ke Xian, and Jianming Zhang. Bokehme: When neural rendering meets classical rendering. In CVPR, 2022. 2, 6, 7", + "[36] Juewen Peng, Zhiyu Pan, Chengxin Liu, Xianrui Luo, Huiqiang Sun, Liao Shen, Ke Xian, and Zhiguo Cao. Selective bokeh effect transformation. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 1456-1464, 2023.", + "[37] Dominique Piché-Meunier, Yannick Hold-Geoffroy, Jianming Zhang, and Jean-François Lalonde. Lens parameter estimation for realistic depth of field modeling. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 499–508, 2023. 2", + "[38] Ming Qian, Congyu Qiao, Jiamin Lin, Zhenyu Guo, Chenghua Li, Cong Leng, and Jian Cheng. BgGAN: Bokehglass generative adversarial network for rendering realistic bokeh. In ECCVW. Springer, 2020. 2", + "[39] René Ranftl, Katrin Lasinger, David Hafner, Konrad Schindler, and Vladlen Koltun. Towards robust monocular depth estimation: Mixing datasets for zero-shot cross-dataset transfer. IEEE TPAMI, 44(3), 2020. 5", + "[40] Johannes L Schonberger and Jan-Michael Frahm. Structure-from-motion revisited. In CVPR, 2016. 5", + "[41] Johannes L. Schonberger, Enliang Zheng, Jan-Michael Frahm, and Marc Pollefeys. Pixelwise view selection for unstructured multi-view stereo. In ECCV. Springer, 2016. 5", + "[42] Liao Shen, Xingyi Li, Huiqiang Sun, Juewen Peng, Ke Xian, Zhiguo Cao, and Guosheng Lin. Make-it-4d: Synthesizing a consistent long-term dynamic scene video from a single image. In Proceedings of the 31st ACM International Conference on Multimedia, pages 8167–8175, 2023. 3", + "[43] Yichen Sheng, Zixun Yu, Lu Ling, Zhiwen Cao, Xuaner Zhang, Xin Lu, Ke Xian, Haiting Lin, and Bedrich Benes. Dr. Bokeh: DiffeRentiable Occlusion-aware Bokeh Rendering. In CVPR, 2024. 2", + "[44] Huiqiang Sun, Xingyi Li, Liao Shen, Xinyi Ye, Ke Xian, and Zhiguo Cao. DyBluRF: Dynamic Neural Radiance Fields from Blurry Monocular Video. In CVPR, 2024. 2, 3", + "[45] Neal Wadhwa, Rahul Garg, David E Jacobs, Bryan E Feldman, Nori Kanazawa, Robert Carroll, Yair Movshovitz-Attias, Jonathan T Barron, Yael Pritch, and Marc Levoy. Synthetic depth-of-field with a single-camera mobile phone. ACM TOG, 37(4), 2018. 2", + "[46] Chao Wang, Krzysztof Wolski, Xingang Pan, Thomas Leimkuhler, Bin Chen, Christian Theobalt, Karol Myszkowski, Hans-Peter Seidel, and Ana Serrano. An implicit neural representation for the image stack: Depth, all in focus, and high dynamic range. Technical report, 2023. 2" + ], + "bbox": [ + 89, + 92, + 483, + 900 + ], + "page_idx": 9 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[47] Chao Wang, Krzysztof Wolski, Bernhard Kerbl, Ana Serrano, Mojtaba Bemana, Hans-Peter Seidel, Karol Myszkowski, and Thomas Leimkuhler. Cinematic Gaussians: Real-Time HDR Radiance Fields with Depth of Field. arXiv preprint arXiv:2406.07329, 2024. 3", + "[48] Lijun Wang, Xiaohui Shen, Jianming Zhang, Oliver Wang, Zhe Lin, Chih-Yao Hsieh, Sarah Kong, and Huchuan Lu. Deep learning: Shallow depth of field from a single image. arXiv preprint arXiv:1810.08100, 2018. 2", + "[49] Peng Wang, Lingzhe Zhao, Ruijie Ma, and Peidong Liu. BAD-NeRF: Bundle adjusted deblur neural radiance fields. In CVPR, 2023. 3", + "[50] Yinhuai Wang, Shuzhou Yang, Yujie Hu, and Jian Zhang. NeRFocus: Neural radiance field for 3D synthetic defocus. arXiv preprint arXiv:2203.05189, 2022. 3", + "[51] Yujie Wang, Praneeth Chakravarthula, and Baoquan Chen. DOF-GS: Adjustable Depth-of-Field 3D Gaussian Splitting for Refocusing, Defocus Rendering and Blur Removal. arXiv preprint arXiv:2405.17351, 2024. 3, 6", + "[52] Zhou Wang, Alan C Bovik, Hamid R Sheikh, and Eero P Simoncelli. Image quality assessment: from error visibility to structural similarity. IEEE TIP, 13(4), 2004. 6", + "[53] Yi Wei, Shaohui Liu, Yongming Rao, Wang Zhao, Jiwen Lu, and Jie Zhou. Nerfingmvs: Guided optimization of neural radiance fields for indoor multi-view stereo. In ICCV, 2021. 5", + "[54] Guanjun Wu, Taoran Yi, Jiemin Fang, Lingxi Xie, Xiaopeng Zhang, Wei Wei, Wenyu Liu, Qi Tian, and Xinggang Wang. 4d gaussian splatting for real-time dynamic scene rendering. In CVPR, 2024. 2, 3", + "[55] Zijin Wu, Xingyi Li, Juewen Peng, Hao Lu, Zhiguo Cao, and Weicai Zhong. DoF-NeRF: Depth-of-field meets neural radiance fields. In ACM MM, 2022. 1, 2, 3, 6, 7, 8", + "[56] Lei Xiao, Anton Kaplanyan, Alexander Fix, Matt Chapman, and Douglas Lanman. Deepfocus: Learned image synthesis for computational display. In ACM SIGGRAPH 2018 Talks. 2018. 2", + "[57] Haofei Xu, Songyou Peng, Fangjinhua Wang, Hermann Blum, Daniel Barath, Andreas Geiger, and Marc Pollefeys. Depthsplat: Connecting gaussian splatting and depth. arXiv preprint arXiv:2410.13862, 2024. 2", + "[58] Zehao Yu, Anpei Chen, Binbin Huang, Torsten Sattler, and Andreas Geiger. Mip-splatting: Alias-free 3D gaussian splattering. In CVPR, 2024. 2, 5, 6", + "[59] Baowen Zhang, Chuan Fang, Rakesh Shrestha, Yixun Liang, Xiaoxiao Long, and Ping Tan. RaDe-GS: Rasterizing Depth in Gaussian Splatting. arXiv preprint arXiv:2406.01467, 2024.5", + "[60] Richard Zhang, Phillip Isola, Alexei A Efros, Eli Shechtman, and Oliver Wang. The unreasonable effectiveness of deep features as a perceptual metric. In CVPR, 2018. 6", + "[61] Xuaner Zhang, Kevin Matzen, Vivien Nguyen, Dillon Yao, You Zhang, and Ren Ng. Synthetic defocus and lookahead autofocus for casual videography. arXiv preprint arXiv:1905.06326, 2019. 2", + "[62] Lingzhe Zhao, Peng Wang, and Peidong Liu. BAD-Gaussians: Bundle adjusted deblur Gaussian Splatting. arXiv preprint arXiv:2403.11831, 2024. 3" + ], + "bbox": [ + 516, + 92, + 903, + 900 + ], + "page_idx": 9 + }, + { + "type": "page_number", + "text": "26471", + "bbox": [ + 478, + 944, + 517, + 955 + ], + "page_idx": 9 + } +] \ No newline at end of file diff --git a/2025/DoF-Gaussian_ Controllable Depth-of-Field for 3D Gaussian Splatting/1ef9852f-e1e4-4571-8156-b74884315f2d_model.json b/2025/DoF-Gaussian_ Controllable Depth-of-Field for 3D Gaussian Splatting/1ef9852f-e1e4-4571-8156-b74884315f2d_model.json new file mode 100644 index 0000000000000000000000000000000000000000..926bd91c76387499d136f4256fef1dc7f47e20b7 --- /dev/null +++ b/2025/DoF-Gaussian_ Controllable Depth-of-Field for 3D Gaussian Splatting/1ef9852f-e1e4-4571-8156-b74884315f2d_model.json @@ -0,0 +1,2574 @@ +[ + [ + { + "type": "header", + "bbox": [ + 0.107, + 0.003, + 0.182, + 0.043 + ], + "angle": 0, + "content": "CVF" + }, + { + "type": "header", + "bbox": [ + 0.237, + 0.0, + 0.812, + 0.045 + ], + "angle": 0, + "content": "This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore." + }, + { + "type": "title", + "bbox": [ + 0.145, + 0.131, + 0.853, + 0.154 + ], + "angle": 0, + "content": "DoF-Gaussian: Controllable Depth-of-Field for 3D Gaussian Splatting" + }, + { + "type": "text", + "bbox": [ + 0.1, + 0.18, + 0.916, + 0.235 + ], + "angle": 0, + "content": "Liao Shen\\(^{1,2}\\) Tianqi Liu\\(^{1,2}\\) Huiqiang Sun\\(^{1,2}\\) Jiaqi Li\\(^{1}\\) Zhiguo Cao\\(^{1}\\) Wei Li\\(^{2*}\\) Chen Change Loy\\(^{2}\\) \n\\(^{1}\\)School of AIA, Huazhong University of Science and Technology \n\\(^{2}\\)S-Lab, Nanyang Technological University" + }, + { + "type": "text", + "bbox": [ + 0.423, + 0.238, + 0.596, + 0.251 + ], + "angle": 0, + "content": "leoshen@hust.edu.cn" + }, + { + "type": "image", + "bbox": [ + 0.093, + 0.264, + 0.439, + 0.441 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.095, + 0.442, + 0.209, + 0.502 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.115, + 0.503, + 0.186, + 0.514 + ], + "angle": 0, + "content": "DoF-NeRF" + }, + { + "type": "image", + "bbox": [ + 0.211, + 0.442, + 0.323, + 0.502 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.248, + 0.503, + 0.29, + 0.514 + ], + "angle": 0, + "content": "BAGS" + }, + { + "type": "image", + "bbox": [ + 0.325, + 0.442, + 0.439, + 0.502 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.361, + 0.503, + 0.397, + 0.514 + ], + "angle": 0, + "content": "Ours" + }, + { + "type": "image", + "bbox": [ + 0.456, + 0.264, + 0.68, + 0.375 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.547, + 0.376, + 0.791, + 0.388 + ], + "angle": 0, + "content": "Change aperture size and focus distance" + }, + { + "type": "image", + "bbox": [ + 0.682, + 0.264, + 0.905, + 0.375 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.456, + 0.39, + 0.681, + 0.5 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.462, + 0.502, + 0.674, + 0.515 + ], + "angle": 0, + "content": "Change camera poses and zoom-in" + }, + { + "type": "image", + "bbox": [ + 0.682, + 0.39, + 0.831, + 0.5 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.723, + 0.503, + 0.861, + 0.516 + ], + "angle": 0, + "content": "Change aperture shape" + }, + { + "type": "image_caption", + "bbox": [ + 0.09, + 0.521, + 0.908, + 0.564 + ], + "angle": 0, + "content": "Figure 1. Given a set of multi-view input images with shallow DoF, DoF-Gaussian can reconstruct a 3D-GS representation of a sharp scene. Thanks to our lens-based design, we can also achieve controllable DoF effects for a variety of applications. The example input images are taken from [55] for illustration purposes. (Zoom-in for best view)" + }, + { + "type": "title", + "bbox": [ + 0.248, + 0.576, + 0.327, + 0.591 + ], + "angle": 0, + "content": "Abstract" + }, + { + "type": "text", + "bbox": [ + 0.089, + 0.607, + 0.486, + 0.879 + ], + "angle": 0, + "content": "Recent advances in 3D Gaussian Splatting (3D-GS) have shown remarkable success in representing 3D scenes and generating high-quality, novel views in real-time. However, 3D-GS and its variants assume that input images are captured based on pinhole imaging and are fully in focus. This assumption limits their applicability, as real-world images often feature shallow depth-of-field (DoF). In this paper, we introduce DoF-Gaussian, a controllable depth-of-field method for 3D-GS. We develop a lens-based imaging model based on geometric optics principles to control DoF effects. To ensure accurate scene geometry, we incorporate depth priors adjusted per scene, and we apply defocus-to-focus adaptation to minimize the gap in the circle of confusion. We also introduce a synthetic dataset to assess refocusing capabilities and the model's ability to learn precise lens parameters. Our framework is customizable and supports various interactive applications. Extensive experiments confirm the effectiveness of our method. Our project is avail-" + }, + { + "type": "text", + "bbox": [ + 0.514, + 0.578, + 0.768, + 0.593 + ], + "angle": 0, + "content": "able at https://dof-gaussian.github.io/." + }, + { + "type": "title", + "bbox": [ + 0.514, + 0.619, + 0.645, + 0.634 + ], + "angle": 0, + "content": "1. Introduction" + }, + { + "type": "text", + "bbox": [ + 0.511, + 0.644, + 0.907, + 0.87 + ], + "angle": 0, + "content": "Depth-of-field (DoF) refers to the distance between the closest and farthest objects in a photo that appears acceptably sharp. In practice, photographers can control the DoF by adjusting the camera's aperture size or focus distance to capture images with either a wide DoF (all-in-focus) or shallow DoF (defocused). The shallow DoF effect is an important technique in photography, as it draws the viewers' attention to the focal region by blurring the surrounding areas. Meanwhile, novel view synthesis aims to create realistic images from novel viewpoints based on a set of source images. However, novel view synthesis typically requires all-in-focus input images and lacks the capability to render varied DoF effects, limiting its applications. In this work, we aim to render novel views with controllable DoF, adding cinematic quality to the results." + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.871, + 0.907, + 0.902 + ], + "angle": 0, + "content": "Images captured from the real world sometimes have shallow DoF. Specifically, points of light that do not lie on" + }, + { + "type": "page_footnote", + "bbox": [ + 0.11, + 0.888, + 0.236, + 0.9 + ], + "angle": 0, + "content": "*Corresponding author." + }, + { + "type": "page_number", + "bbox": [ + 0.479, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "26462" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.09, + 0.092, + 0.482, + 0.332 + ], + "angle": 0, + "content": "the focal plane are projected onto the sensor plane as blurred circles, referred to as the circle of confusion (CoC), and this introduces bokeh blur to the captured image. Most novel view synthesis methods experience performance degradation when processing shallow DoF input images. To address this issue, DoF-NeRF [55] and LensNeRF [17] introduce lens-based camera models in volume rendering to explicitly enable controllable DoF effects. Meanwhile, methods like Deblur-NeRF [26], DP-NeRF [19], and PDRF [33] employ sparse kernels to model defocus blur. However, implicit rendering approaches face significant challenges in training and rendering efficiency. In the realm of 3D Gaussian Splatting (3D-GS) [16], methods such as BAGS [34] and Deblurring 3DGS [18] propose blur estimation networks to model the blur kernels or scaling factors for defocus deblurring, but they cannot accommodate controllable DoF." + }, + { + "type": "text", + "bbox": [ + 0.093, + 0.336, + 0.483, + 0.607 + ], + "angle": 0, + "content": "In this paper, we present DoF-Gaussian, an efficient framework for controllable DoF in 3D-GS, addressing the limitations of existing methods in handling shallow DoF inputs. This problem is non-trivial because 3D-GS and its variants [11, 23, 54, 57, 58] are typically based on a pin-hole camera model that assumes all-in-focus inputs, meaning both foreground and background appear clear. When input images contain bokeh blur, it becomes challenging to accurately construct scene geometry and render depth maps. However, rendering images with different DoFs relies heavily on precise depth estimation. Additionally, our 3D-GS approach assumes an idealized CoC, which inevitably differs from the CoC seen in real photographs, posing further challenges for accurate defocus deblurring. Furthermore, existing datasets in this field are primarily borrowed from deblurring applications, limiting our evaluation to the model's defocus deblurring capability and overlooking other aspects of controllable DoF effects." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.612, + 0.483, + 0.837 + ], + "angle": 0, + "content": "To address these challenges, we present a new controllable DoF 3D-GS method with the following contributions: First, we employ a lens-based model instead of a pinhole-based one. We make the lens optical parameters, such as aperture size and focus distance, learnable to enable control over the depth of field. Second, we introduce per-scene adjustments for depth priors to minimize errors in scene geometry and depth relationships, ensuring accurate depth-of-field control. Third, we propose a defocus-to-focus adaptation strategy that focuses on learning the focal region after the learnable lens parameters have converged, compensating for differences in CoC. Finally, we present a synthetic dataset designed to comprehensively evaluate the model, including its refocusing capabilities and its ability to learn accurate aperture size and focus distance." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.84, + 0.483, + 0.902 + ], + "angle": 0, + "content": "Benefiting from the lens-based imaging model, our approach enables controlled depth of field, allowing it to handle not only shallow DoF image inputs but also unlock a range of interactive and engaging applications, as depicted" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.092, + 0.905, + 0.213 + ], + "angle": 0, + "content": "in Figure 1. For instance, users can render novel view images with varied DoF, refocus on custom datasets, and even adjust the shape of CoC, e.g., from circular to pentagonal or hexagonal CoC. Moreover, users can dynamically change the DoF while moving the camera or zooming, rendering videos with cinematic effects. Extensive experiments demonstrate that our method outperforms state-of-the-art depth-of-field and defocus deblurring methods." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.226, + 0.655, + 0.241 + ], + "angle": 0, + "content": "2. Related Work" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.252, + 0.905, + 0.568 + ], + "angle": 0, + "content": "Depth-of-Field Rendering. Rendering DoF effects from a single all-in-focus image has been widely explored in previous work. Bokeh blur occurs when light is projected onto the image plane as a circular region, rather than as a point. The size of CoC is affected by the diameter of the aperture (aperture size) and the distance from the camera to the focal plane (focus distance). Photos captured with a small aperture usually present a wide DoF, i.e., all objects appear sharp. In contrast, as the aperture diameter increases, objects near the focal plane remain sharp, while those farther away become blurred with a larger CoC. Physically based methods [1, 20] rely on 3D scene geometry information and are time-consuming. Some methods use neural networks [8, 12–14, 38] trained end-to-end to obtain shallow DoF images. The DoF effects in some studies [6, 25, 35–37, 43, 45, 48, 56, 61] are controllable but usually require an extra disparity map. Wang et al. [46] combines neural fields with an expressive camera model to achieve all-in-focus reconstruction from an image stack. In this work, we show the connection between DoF rendering and novel view synthesis using a lens-based camera model." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.569, + 0.905, + 0.824 + ], + "angle": 0, + "content": "Image Deblurring. Blur can generally be categorized into two main types: camera motion blur and defocus blur. Previous studies [7, 18, 24, 26, 33, 34, 44] attempt to address this issue. Our task of rendering all-in-focus and sharp novel view images from shallow DoF inputs can be considered as a form of defocus deblurring. By modeling the underlying physical principles, we can achieve defocus deblurring by simulating the circle-of-confusion. Furthermore, our method allows for adjustments to lens parameters, enabling applications such as refocusing—capabilities that deblurring approaches like BAGS [34] cannot achieve. Existing datasets in the 3D DoF field are often borrowed from the deblurring domain [26] or designed solely to assess defocus deblurring ability [55]. To support a wider range of applications and validate the accuracy of the learned lens model, we introduce a synthetic dataset for more comprehensive evaluations." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.825, + 0.905, + 0.901 + ], + "angle": 0, + "content": "Novel View Synthesis. Novel view synthesis allows the rendering of unseen camera perspectives from 2D images and their corresponding camera poses. Recent advancements in synthesis can largely be attributed to Neural Radiance Field [28] and 3D Gaussian Splatting [16]. However," + }, + { + "type": "page_number", + "bbox": [ + 0.48, + 0.945, + 0.518, + 0.957 + ], + "angle": 0, + "content": "26463" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.095, + 0.087, + 0.91, + 0.368 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.09, + 0.378, + 0.908, + 0.448 + ], + "angle": 0, + "content": "Figure 2. Overview of DoF-Gaussian. Given input images \\(I\\) with shallow DoF, we first apply SfM from COLMAP to obtain sparse depth \\(D_{sparse}\\), which is used to train a depth network to derive per-scene depth priors \\(D_{pred}\\). We then employ \\(D_{pred}\\) to regularize the Gaussians rendered depth map \\(D\\). Next, by developing a lens imaging model, we can render defocused images \\(C^*\\) to simulate input images. To minimize the discrepancy in CoC, we propose an adaptation using the weight map. Finally, we can render fully clear images for novel view synthesis and achieve various effects by our controllable DoF framework." + }, + { + "type": "text", + "bbox": [ + 0.093, + 0.468, + 0.485, + 0.892 + ], + "angle": 0, + "content": "real-world images are often not clean and well-calibrated. Previous work [4, 27, 29, 31, 42, 44, 49, 54, 62] attempts to cope with non-ideal conditions, such as low light, camera motion, bokeh effects, and various types of image degradation. To recover a sharp scene from images with bokeh blur, methods like Deblur-NeRF [26], DP-NeRF [19], and PDRF [33] model defocus blur using sparse kernels. NeRF Focus [50] and LensNeRF [17] achieve defocus effects by incorporating a lens-based camera model in volume rendering. While DoF-NeRF [55] is similar to our approach, it is limited by the training and rendering efficiency and quality of NeRF. BAGS [34] and Deblurring 3DGS [18] use blur estimation networks to model blur kernels or scaling factors but cannot achieve controllable DoF effects due to the absence of a lens imaging model. Other studies, such as [15, 47], focus on HDR tasks and consider depth-of-field simultaneously. RGS [7] addresses defocus blur by introducing an offset to the 2D covariance matrices of Gaussians. [30] introduces a ray tracing rendering algorithm for particle-based representations to enable many advanced techniques, such as the shallow depth-of-field effect. A concurrent work [51] also explores adjustable depth of field. Our method differs in its controllable DoF mechanism, with the introduction of depth priors to ensure accurate scene depth. In addition, our new defocus-to-focus adaptation strategy eliminates the need for the focus localization network proposed in their work. Besides, we introduce a synthetic dataset for comprehensive evaluation." + }, + { + "type": "title", + "bbox": [ + 0.514, + 0.467, + 0.64, + 0.484 + ], + "angle": 0, + "content": "3. Preliminary" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.492, + 0.907, + 0.538 + ], + "angle": 0, + "content": "3D Gaussian Splitting represents a 3D scene as a mixture of anisotropic 3D Gaussians, where each Gaussian is characterized by a 3D covariance matrix \\(\\Sigma\\) and mean \\(\\mu\\):" + }, + { + "type": "equation", + "bbox": [ + 0.605, + 0.543, + 0.907, + 0.565 + ], + "angle": 0, + "content": "\\[\nG (X) = e ^ {- \\frac {1}{2} (X - \\mu) ^ {\\top} \\Sigma^ {- 1} (X - \\mu)} \\tag {1}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.57, + 0.907, + 0.63 + ], + "angle": 0, + "content": "The covariance matrix \\(\\Sigma\\) holds physical meaning only when it is positive semi-definite. Therefore, to enable effective optimization via gradient descent, \\(\\Sigma\\) is decomposed into a scaling matrix \\(S\\) and a rotation matrix \\(R\\), as:" + }, + { + "type": "equation", + "bbox": [ + 0.655, + 0.636, + 0.905, + 0.654 + ], + "angle": 0, + "content": "\\[\n\\Sigma = R S S ^ {\\mathsf {T}} R ^ {\\mathsf {T}}. \\tag {2}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.66, + 0.907, + 0.735 + ], + "angle": 0, + "content": "To splot Gaussians from 3D space to a 2D plane, the view transformation matrix \\( W \\) and the Jacobian matrix \\( J \\), which represents the affine approximation of the projective transformation, are utilized to obtain the covariance matrix \\( \\Sigma' \\) in 2D space, as:" + }, + { + "type": "equation", + "bbox": [ + 0.642, + 0.734, + 0.905, + 0.753 + ], + "angle": 0, + "content": "\\[\n\\Sigma^ {\\prime} = J W \\Sigma W ^ {\\mathsf {T}} J ^ {\\mathsf {T}}. \\tag {3}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.756, + 0.908, + 0.787 + ], + "angle": 0, + "content": "Subsequently, a point-based alpha-blending rendering can be performed to determine the color of each pixel:" + }, + { + "type": "equation", + "bbox": [ + 0.619, + 0.793, + 0.905, + 0.835 + ], + "angle": 0, + "content": "\\[\nC = \\sum_ {i} c _ {i} \\alpha_ {i} \\prod_ {j = 1} ^ {i - 1} (1 - \\alpha_ {i}), \\tag {4}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.841, + 0.909, + 0.903 + ], + "angle": 0, + "content": "where \\( c_{i} \\) represents the color of each point, defined by spherical harmonics (SH) coefficients. The density \\( \\alpha_{i} \\) is computed as the product of 2D Gaussians and a learnable point-wise opacity." + }, + { + "type": "page_number", + "bbox": [ + 0.479, + 0.945, + 0.521, + 0.958 + ], + "angle": 0, + "content": "26464" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.092, + 0.087, + 0.278, + 0.198 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.117, + 0.199, + 0.255, + 0.212 + ], + "angle": 0, + "content": "(I) Pinhole-based model" + }, + { + "type": "image", + "bbox": [ + 0.284, + 0.088, + 0.484, + 0.2 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.326, + 0.2, + 0.453, + 0.212 + ], + "angle": 0, + "content": "(II) Lens-based model" + }, + { + "type": "image_caption", + "bbox": [ + 0.09, + 0.223, + 0.484, + 0.321 + ], + "angle": 0, + "content": "Figure 3. The difference between the pinhole-based model (I) and the lens-based model (II). For (I), light emitted from spatial point \\(O\\) directly hits the point \\(O'\\) of the image plane. For (II), we show the case that light emitted from \\(Q\\) converges on the point \\(Q'\\) of the image plane (In-Focus), and the case that light emitted from \\(P\\) converges on the point \\(P'\\) and continues to scatter onto the image plane forming a circle of confusion (Out-of-Focus)." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.34, + 0.182, + 0.355 + ], + "angle": 0, + "content": "4. Method" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.365, + 0.484, + 0.502 + ], + "angle": 0, + "content": "We propose DoF-Gaussian, a controllable DoF approach for 3D-GS. The overall pipeline is illustrated in Fig. 2. In Section 4.1, we first develop a lens-based imaging model based on geometric optics principles to control DoF effects. Next, we employ the per-scene adjustment of depth priors to guide the correct scene geometry, as described in Section 4.2. Finally, in Section 4.3, we apply defocus-to-focus adaptation to minimize the inconsistencies in the CoC and enhance the defocus deblurring performance." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.511, + 0.385, + 0.527 + ], + "angle": 0, + "content": "4.1. Lens-based Depth-of-Field Model" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.534, + 0.484, + 0.745 + ], + "angle": 0, + "content": "The physical principles underlying imaging and DoF have been extensively studied in the field of geometric optics [10, 32]. In an idealized optical system, the light emitted from a spatial point \\(O\\) is projected onto a corresponding point \\(O^{\\prime}\\) in the image plane following the principles of pinhole imaging, as illustrated in Fig. 3 (I). However, real-world cameras operate based on the lens model, as shown in Fig. 3 (II). We show the two cases of in-focus and out-of-focus imaging, respectively. A spatial point \\(P\\), located at a distance \\(d\\) from the lens, is projected on the imaging plane as a circular region referred to as the circle of confusion (CoC). The diameter of this region, \\(r(d)\\), can be determined by the aperture parameter \\(\\mathcal{A}\\) and focus distance \\(\\mathcal{F}\\), according to the following equation:" + }, + { + "type": "equation", + "bbox": [ + 0.219, + 0.756, + 0.483, + 0.79 + ], + "angle": 0, + "content": "\\[\nr (d) = \\mathcal {A} \\left| \\frac {1}{\\mathcal {F}} - \\frac {1}{d} \\right|. \\tag {5}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.795, + 0.484, + 0.901 + ], + "angle": 0, + "content": "The focus distance \\(\\mathcal{F}\\) primarily controls the depth position of the focal plane of the image, i.e., the depth of the sharp area, while the aperture parameter \\(\\mathcal{A}\\) determines the extent of the bokeh effect. For the point \\(Q\\) located at the focus distance \\(d = \\mathcal{F}\\), the emitted rays directly converge through the lens to the corresponding point \\(Q'\\) in the image plane, resulting in the absence of a circle of confusion" + }, + { + "type": "title", + "bbox": [ + 0.515, + 0.092, + 0.794, + 0.107 + ], + "angle": 0, + "content": "Algorithm 1 Lens-based imaging process" + }, + { + "type": "text", + "bbox": [ + 0.515, + 0.11, + 0.906, + 0.154 + ], + "angle": 0, + "content": "Input: Rasterization rendering \\(C\\), aperture parameter \\(\\mathcal{A}\\), focus distance \\(\\mathcal{F}\\), depth map \\(D\\), gamma value \\(\\gamma\\), confuse function Func" + }, + { + "type": "text", + "bbox": [ + 0.515, + 0.156, + 0.769, + 0.17 + ], + "angle": 0, + "content": "Output: Defocus simulated result \\( C^* \\)" + }, + { + "type": "equation", + "bbox": [ + 0.532, + 0.17, + 0.735, + 0.187 + ], + "angle": 0, + "content": "\\[\nR \\leftarrow \\mathcal {A} \\left| \\frac {1}{\\bar {\\mathcal {E}}} - \\frac {1}{\\bar {D}} \\right| \\quad C \\leftarrow (C) ^ {\\gamma}\n\\]" + }, + { + "type": "equation", + "bbox": [ + 0.533, + 0.187, + 0.652, + 0.201 + ], + "angle": 0, + "content": "\\[\n\\Phi = [ 0 ], \\tilde {C ^ {*}} = \\tilde {[} 0 ]\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.532, + 0.201, + 0.735, + 0.217 + ], + "angle": 0, + "content": "for pixel \\(i\\gets \\mathrm{Traverse}(C)\\) do" + }, + { + "type": "text", + "bbox": [ + 0.548, + 0.217, + 0.837, + 0.232 + ], + "angle": 0, + "content": "for pixel \\(j\\gets\\) TraverseNeighbor \\((c_i,r_i)\\) do" + }, + { + "type": "equation", + "bbox": [ + 0.565, + 0.232, + 0.724, + 0.248 + ], + "angle": 0, + "content": "\\[\n\\lambda_ {i j} \\leftarrow \\operatorname {F u n c} (d _ {i}, | i - j |)\n\\]" + }, + { + "type": "equation", + "bbox": [ + 0.565, + 0.248, + 0.671, + 0.263 + ], + "angle": 0, + "content": "\\[\n\\Phi_ {j} \\leftarrow \\Phi_ {j} + \\lambda_ {i j}\n\\]" + }, + { + "type": "equation", + "bbox": [ + 0.565, + 0.263, + 0.687, + 0.279 + ], + "angle": 0, + "content": "\\[\nc _ {j} ^ {*} \\leftarrow c _ {j} ^ {*} + \\lambda_ {i j} \\cdot c _ {i}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.548, + 0.279, + 0.602, + 0.29 + ], + "angle": 0, + "content": "end for" + }, + { + "type": "text", + "bbox": [ + 0.532, + 0.292, + 0.586, + 0.304 + ], + "angle": 0, + "content": "end for" + }, + { + "type": "equation", + "bbox": [ + 0.532, + 0.304, + 0.644, + 0.323 + ], + "angle": 0, + "content": "\\[\nC ^ {*} \\leftarrow (C ^ {*} / \\Phi) ^ {\\frac {1}{\\gamma}}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.514, + 0.352, + 0.78, + 0.367 + ], + "angle": 0, + "content": "(CoC). This also implies that \\(r(\\mathcal{F}) = 0\\)" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.367, + 0.906, + 0.457 + ], + "angle": 0, + "content": "We set focus distance \\(\\mathcal{F}\\) and aperture parameter \\(\\mathcal{A}\\) as learnable for each input image, and these parameters are continuously updated with the optimization of 3D-GS. By modeling the lens camera, the output \\(C^*\\), which includes bokeh blur effects, can be derived from the 3D-GS rasterizer rendering result \\(C\\) using our CUDA-based Algorithm 1." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.457, + 0.906, + 0.518 + ], + "angle": 0, + "content": "Ideally, the confuse function is represented by the indicator function \\(\\mathbb{I}(r(d) > l)\\). To achieve a smooth and natural DoF effect, we substitute it with a differentiable function, as proposed in Busam et al. [3]:" + }, + { + "type": "equation", + "bbox": [ + 0.567, + 0.526, + 0.906, + 0.556 + ], + "angle": 0, + "content": "\\[\nF u n c (d, l) = \\frac {1}{2} + \\frac {1}{2} \\tanh (\\alpha (r (d) - l)), \\tag {6}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.563, + 0.906, + 0.593 + ], + "angle": 0, + "content": "where \\(\\alpha\\) defines the smoothness of confuse transition and \\(l\\) represents the distance between two pixels." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.593, + 0.906, + 0.654 + ], + "angle": 0, + "content": "We supervise the training by computing the reconstruction loss between the output image \\( C^* \\) and the shallow DoF input image \\( I \\), where the loss function is a combination of the \\( \\mathcal{L}_1 \\) and a D-SSIM term, with \\( \\lambda \\) set to 0.2:" + }, + { + "type": "equation", + "bbox": [ + 0.536, + 0.663, + 0.905, + 0.681 + ], + "angle": 0, + "content": "\\[\n\\mathcal {L} _ {r e c} = (1 - \\lambda) \\mathcal {L} _ {1} \\left(I, C ^ {*}\\right) + \\lambda \\mathcal {L} _ {D - S S I M} \\left(I, C ^ {*}\\right). \\tag {7}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.689, + 0.906, + 0.72 + ], + "angle": 0, + "content": "For inference, we set the aperture size \\(\\mathcal{A}\\) to 0 so that we can render fully clear images for novel view synthesis." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.728, + 0.845, + 0.744 + ], + "angle": 0, + "content": "4.2. Per-Scene Adjustment of Depth Priors" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.75, + 0.907, + 0.902 + ], + "angle": 0, + "content": "Unlike previous work that assumes input images are all-in-focus and fully clear, it becomes challenging to accurately reconstruct scene geometry and render depth maps for bokeh blurring input images. However, rendering images with varying depths of field relies on precise depth information. To address this issue, we propose using depth prior as useful guidance for the accurate reconstruction of scene geometry. However, due to the presence of bokeh blur in shallow DoF inputs, directly employing a monocular depth network to predict depth maps \\( D_{pred} \\) as pseudo" + }, + { + "type": "page_number", + "bbox": [ + 0.479, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "26465" + } + ], + [ + { + "type": "image_caption", + "bbox": [ + 0.129, + 0.089, + 0.16, + 0.099 + ], + "angle": 0, + "content": "RGB" + }, + { + "type": "image_caption", + "bbox": [ + 0.224, + 0.089, + 0.264, + 0.099 + ], + "angle": 0, + "content": "BAGS" + }, + { + "type": "image_caption", + "bbox": [ + 0.29, + 0.089, + 0.384, + 0.101 + ], + "angle": 0, + "content": "w/o depth priors" + }, + { + "type": "image_caption", + "bbox": [ + 0.418, + 0.089, + 0.447, + 0.099 + ], + "angle": 0, + "content": "Ours" + }, + { + "type": "image", + "bbox": [ + 0.094, + 0.101, + 0.482, + 0.202 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.09, + 0.212, + 0.483, + 0.269 + ], + "angle": 0, + "content": "Figure 4. Effects of the depth priors. The per-scene adjustment of depth priors enhance the geometric structure and yield a more accurate depth map than that estimated by BAGS [34] and a variant of our method without depth priors." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.289, + 0.484, + 0.515 + ], + "angle": 0, + "content": "ground truth to regularize the Gaussians rendered depth map, as done in [5, 21], does not yield satisfactory results. Inspired by [53], we propose adapting per-scene depth priors through fine-tuning a monocular depth network based on the scene's sparse reconstruction. Specifically, we utilize COLMAP [40, 41] to obtain a sparse Structure-from-Motion (SfM) point cloud and create a set of 3D Gaussians. Consequently, we can derive per-view sparse depth maps \\( D_{sparse} \\) by projecting the 3D point clouds after multiview stereo processing. Although the resulting depth map is sparse, it is robust and can serve as a supervisory signal for training the per-scene depth priors. To address the scale ambiguity inherent in acquired depth maps, we apply a silog loss [9] to fine-tune the depth network [22, 39], expressed as:" + }, + { + "type": "equation", + "bbox": [ + 0.106, + 0.523, + 0.483, + 0.565 + ], + "angle": 0, + "content": "\\[\n\\mathcal {L} _ {\\text {s i l o g}} = \\frac {1}{2 M} \\sum_ {i = 1} ^ {M} \\left(\\log \\left(e ^ {s} D _ {\\text {p r e d}}\\right) - \\log \\left(D _ {\\text {s p a r s e}}\\right)\\right) ^ {2}, \\tag {8}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.574, + 0.484, + 0.68 + ], + "angle": 0, + "content": "where \\( e^s \\) is the scale factor and \\( M \\) is the number of pixels in the image used for calculation. We apply the rasterization technique from Rade-GS [59], which enables the robust rendering of depth maps \\( D \\) for 3D-GS scenes. We predict the depth maps \\( D_{pred} \\) from multi-view input images using the fine-tuned depth network and use them as depth priors to regularize \\( D \\). The depth loss can be described as follows:" + }, + { + "type": "equation", + "bbox": [ + 0.2, + 0.691, + 0.483, + 0.709 + ], + "angle": 0, + "content": "\\[\n\\mathcal {L} _ {\\text {d e p t h}} = \\left| \\left| D - D _ {\\text {p r e d}} \\right| \\right| _ {2}. \\tag {9}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.718, + 0.484, + 0.826 + ], + "angle": 0, + "content": "As illustrated in Fig. 4, it is evident that the depth priors significantly enhance the geometric structure, resulting in a more accurate depth map compared to both BAGS [34] and the variant of our method without depth priors. In Section 5.3, we further conduct an ablation study comparing several other depth supervision strategies to demonstrate the effectiveness of our approach." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.834, + 0.351, + 0.85 + ], + "angle": 0, + "content": "4.3. Defocus-to-Focus Adaptation" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.856, + 0.484, + 0.901 + ], + "angle": 0, + "content": "The CoC effect achieved under ideal optical imaging conditions, as described in Algorithm 1, inevitably differs from the real CoC effect produced by an actual DSLR camera." + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.092, + 0.905, + 0.122 + ], + "angle": 0, + "content": "This discrepancy hinders our ability to accurately model the out-of-focus blur in shallow DoF input images." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.123, + 0.906, + 0.289 + ], + "angle": 0, + "content": "To achieve sharper scenes, we propose an adaptation process that allows transitions from defocus to focus. Specifically, modeling the defocus effects with bokeh blur facilitates learning accurate lens parameters. Consequently, we place greater emphasis on capturing the sharp areas in training images once the lens parameters \\(\\mathcal{A}\\) and \\(\\mathcal{F}\\) have been optimized and converged. This approach is effective because, when the focus distance is determined, the sharp regions in the scene become approximately identifiable. After \\(t\\) optimization iterations, we reweight the \\(\\mathcal{L}_{rec}\\) and \\(\\mathcal{L}_{depth}\\) loss by employing a step-like function:" + }, + { + "type": "equation", + "bbox": [ + 0.53, + 0.299, + 0.906, + 0.334 + ], + "angle": 0, + "content": "\\[\n\\Psi (x) = \\left\\{ \\begin{array}{c c} 1 & , \\text {i t e r a t i o n s} < t \\\\ 1 / \\left(1 + e ^ {- a \\cdot (x - b)}\\right) & , \\text {i t e r a t i o n s} \\geq t \\end{array} \\right. \\tag {10}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.345, + 0.871, + 0.363 + ], + "angle": 0, + "content": "where \\(a\\) and \\(b\\) are hyperparameters, and \\(x = \\left|\\frac{1}{\\mathcal{F}} - \\frac{1}{d}\\right|\\)." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.362, + 0.906, + 0.422 + ], + "angle": 0, + "content": "In addition, the aperture size \\(\\mathcal{A}\\) is generally large after the convergence, which can cause blurring of pixels that should be in focus. Hence, we also reweight the aperture size \\(\\mathcal{A}\\) on a pixel-wise basis:" + }, + { + "type": "equation", + "bbox": [ + 0.667, + 0.434, + 0.905, + 0.451 + ], + "angle": 0, + "content": "\\[\n\\mathcal {A} ^ {\\prime} = \\mathcal {A} \\cdot \\Psi . \\tag {11}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.463, + 0.906, + 0.57 + ], + "angle": 0, + "content": "Full Objective. We derive the final loss terms by incorporating the reconstruction loss from Eqn. 7, the depth loss from Eqn. 9, and an additional normal consistency loss as proposed by Huang et al. [11]. The normal consistency loss ensures that the Gaussian splats align with the surface by measuring the consistency between the normal directions computed from the Gaussian and the depth map:" + }, + { + "type": "equation", + "bbox": [ + 0.607, + 0.581, + 0.905, + 0.613 + ], + "angle": 0, + "content": "\\[\n\\mathcal {L} _ {\\text {n o r m a l}} = \\sum_ {i} w _ {i} \\left(1 - n _ {i} ^ {\\top} \\hat {n} _ {i}\\right), \\tag {12}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.623, + 0.906, + 0.683 + ], + "angle": 0, + "content": "where \\(\\hat{n}\\) represents the surface normal direction obtained by applying finite differences on the depth map, \\(i\\) indexes the intersected splats along the ray and \\(w\\) denotes the blending weight of the intersection point." + }, + { + "type": "text", + "bbox": [ + 0.534, + 0.684, + 0.716, + 0.699 + ], + "angle": 0, + "content": "Our final training loss \\(\\mathcal{L}\\) is:" + }, + { + "type": "equation", + "bbox": [ + 0.547, + 0.712, + 0.905, + 0.729 + ], + "angle": 0, + "content": "\\[\n\\mathcal {L} = \\Psi \\odot \\left(\\mathcal {L} _ {\\text {r e c}} + w _ {d} \\mathcal {L} _ {\\text {d e p t h}}\\right) + w _ {n} \\mathcal {L} _ {\\text {n o r m a l}}, \\tag {13}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.74, + 0.72, + 0.756 + ], + "angle": 0, + "content": "where \\(\\odot\\) is Hadamard product." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.77, + 0.646, + 0.788 + ], + "angle": 0, + "content": "5. Experiments" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.795, + 0.906, + 0.901 + ], + "angle": 0, + "content": "Implementation Details. In our experiments, we set the smoothness of confuse transition \\(\\alpha = 4\\). In addition, we set \\(a = 15\\), \\(b = 0.3\\), and \\(t = 10000\\) for the defocus-to-focus adaptation. Our method is built upon Mip-Splatting [58], and our optimization strategy and hyperparameter settings remain consistent with it. We train each scene for 30000 iterations and set loss weights \\(w_{d} = 0.01\\) and \\(w_{n} = 0.05\\)." + }, + { + "type": "page_number", + "bbox": [ + 0.479, + 0.945, + 0.521, + 0.957 + ], + "angle": 0, + "content": "26466" + } + ], + [ + { + "type": "table_caption", + "bbox": [ + 0.09, + 0.089, + 0.908, + 0.121 + ], + "angle": 0, + "content": "Table 1. Quantitative comparisons on the defocus blur dataset of Deblur NeRF [26]. We report the PSNR, SSIM, and LPIPS metrics and color each cell as best and second best. Our method outperforms other existing approaches across most scenes." + }, + { + "type": "table", + "bbox": [ + 0.099, + 0.13, + 0.9, + 0.299 + ], + "angle": 0, + "content": "
MethodDeblur-NeRF [26]DoF-NeRF [55]DP-NeRF [19]BAGS [34]Deblurring 3DGS [18]Ours
PSNRSSIMLPIPSPSNRSSIMLPIPSPSNRSSIMLPIPSPSNRSSIMLPIPSPSNRSSIMLPIPSPSNRSSIMLPIPS
Cake26.270.7800.12824.530.7310.20626.160.7780.12727.210.8180.10826.880.8030.11626.830.8080.101
Caps23.870.7130.16122.670.6360.26123.950.7120.14324.160.7250.15924.500.7420.14924.620.7490.147
Cisco20.830.7270.08720.640.7240.10720.730.7260.08420.790.7430.07020.830.7320.07921.000.7440.069
Coral19.850.6000.12119.830.5700.24020.110.6110.11820.530.6280.11119.780.6080.13120.370.6300.109
Cupcake22.260.7220.11621.890.7060.14322.800.7410.09622.930.7620.08022.110.7340.09922.970.7570.079
Cups26.210.7990.12725.260.7650.20226.750.8140.10426.270.8230.10426.280.8240.10326.010.8170.100
Daisy23.520.6870.12123.220.6580.19423.790.6970.10823.740.7460.06223.540.7310.09523.930.7350.071
Sausage18.010.5000.18017.860.4880.28018.350.5440.14718.760.5740.11018.990.5700.14119.110.5760.119
Seal26.040.7770.10524.850.6870.14325.950.7780.10326.520.8120.09026.180.8170.09826.570.8250.087
Tools27.810.8950.06126.210.8540.12828.070.8980.05428.600.9130.04627.960.9100.05828.290.9130.051
" + }, + { + "type": "table_caption", + "bbox": [ + 0.099, + 0.322, + 0.475, + 0.337 + ], + "angle": 0, + "content": "Table 2. Quantitative comparisons on our synthetic dataset." + }, + { + "type": "table", + "bbox": [ + 0.096, + 0.348, + 0.48, + 0.403 + ], + "angle": 0, + "content": "
MethodPSNR↑SSIM↑LPIPS↓δA↓δF↓
DoF-NeRF [55]25.590.7880.2070.1960.256
Ours28.700.8640.0950.1260.079
" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.426, + 0.483, + 0.531 + ], + "angle": 0, + "content": "Baselines and Evaluation Metrics. We compare our method with state-of-the-art deblurring techniques [18, 26, 34] and a depth-of-field method [55]. To quantitatively evaluate the quality of novel view images, we adopt widely used metrics such as PSNR, SSIM [52], and LPIPS [60]. We will compare with Wang et al. [51] once they release the code or provide the full experimental results in their paper." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.533, + 0.484, + 0.64 + ], + "angle": 0, + "content": "Datasets. Following [18, 26, 34], we evaluate our model on the Deblur-NeRF dataset [26]. Subsequently, we conduct further evaluations on the Real Forward-facing dataset [28] and T&T_DB dataset [16]. The effectiveness of our method is validated on these datasets for both shallow DoF inputs and normal inputs. Furthermore, we also introduce a dataset in Section 5.1 to evaluate our model more comprehensively." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.667, + 0.28, + 0.683 + ], + "angle": 0, + "content": "5.1. A Synthetic Dataset" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.689, + 0.483, + 0.901 + ], + "angle": 0, + "content": "We introduce a synthetic dataset, as illustrated in Fig. 6, based on depth estimation [2] and a single-image DoF rendering method [35] for each image in the Real Forward Facing dataset [28] and T&T_DB dataset [16]. This allows us to evaluate the refocusing ability of our model and to determine whether it accurately learns the correct aperture size and focus distance. In particular, by feeding the DoF rendering method [35] the known aperture size and focus distance, we can convert the all-in-focus images into shallow DoF images. Unlike existing defocus blur datasets proposed by Ma et al. [26] and Wu et al. [55], the test set in our dataset comprises images with shallow DoF. With known lens parameters, our model fits this shallow DoF effect to quantitatively measure the refocusing ability. Meanwhile, we establish the" + }, + { + "type": "table_caption", + "bbox": [ + 0.513, + 0.322, + 0.907, + 0.365 + ], + "angle": 0, + "content": "Table 3. Comparisons in all-in-focus settings. We compare our method with Mip-Splatting on all-in-focus datasets to validate the effectiveness of our model under general input conditions." + }, + { + "type": "table", + "bbox": [ + 0.536, + 0.375, + 0.882, + 0.438 + ], + "angle": 0, + "content": "
MethodPSNR↑SSIM↑LPIPS↓
Mip-Splatting [58]27.050.8930.115
Ours27.810.9020.117
" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.459, + 0.907, + 0.52 + ], + "angle": 0, + "content": "known lens parameters beforehand as ground truth, allowing us to calculate the errors \\(\\delta_{\\mathcal{A}}\\) and \\(\\delta_{\\mathcal{F}}\\) between the learned parameters and the ground truth. More details can be found in the supplementary material." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.53, + 0.655, + 0.546 + ], + "angle": 0, + "content": "5.2. Comparisons" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.553, + 0.906, + 0.658 + ], + "angle": 0, + "content": "Quantitative Comparisons. We quantitatively compare DoF-Gaussian against other state-of-the-art methods using a real-world defocus blur dataset [26]. As shown in Table 1, our method outperforms other state-of-the-art baselines. These results suggest that our method achieves superior defocus deblurring performance and produces higher-quality novel view synthesis images." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.659, + 0.906, + 0.749 + ], + "angle": 0, + "content": "Qualitative Comparisons. Visual qualitative comparisons are presented in Fig. 5. Our method surpasses all other methods in terms of image fidelity, generating novel view images that are more faithful to the ground truth images and exhibit less blur. Specifically, the bottle caps and shelves are the sharpest in our rendered images." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.75, + 0.907, + 0.901 + ], + "angle": 0, + "content": "All-in-focus Inputs. To further validate the effectiveness of our method in the all-in-focus setting, we design an experiment using wide DoF images as inputs. We conduct a comparison between our method and Mip-Splatting [58] on two all-in-focus datasets, the Real Forward Facing dataset [28] and the T&T_DB dataset [16]. As shown in Table 3, our method achieves comparable or even better performance to Mip-Splatting on the average metrics across the datasets. This demonstrates that our lens-based model not only excels with shallow DoF inputs but also achieves good" + }, + { + "type": "page_number", + "bbox": [ + 0.479, + 0.945, + 0.521, + 0.958 + ], + "angle": 0, + "content": "26467" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.093, + 0.086, + 0.907, + 0.316 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.09, + 0.325, + 0.907, + 0.355 + ], + "angle": 0, + "content": "Figure 5. Qualitative comparisons against all baselines. Compared to other state-of-the-art methods, our method represents sharper scenes and generates novel view images with less blur." + }, + { + "type": "image", + "bbox": [ + 0.093, + 0.369, + 0.483, + 0.526 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.09, + 0.536, + 0.483, + 0.593 + ], + "angle": 0, + "content": "Figure 6. Examples in our dataset. For each scene, we apply a DoF rendering method [35] to convert all-in-focus images to shallow DoF images with varying focus locations and aperture sizes. This dataset is used to evaluate the model's refocusing ability." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.613, + 0.316, + 0.628 + ], + "angle": 0, + "content": "results with general input images." + }, + { + "type": "text", + "bbox": [ + 0.089, + 0.628, + 0.483, + 0.78 + ], + "angle": 0, + "content": "Our Synthetic Dataset. As shown in Table 2, our method achieves superior refocusing ability through controlled DoF rendering, producing rendered images with the highest quality on our synthetic dataset. Furthermore, our method learns more precise aperture size and focus distance than DoF-NeRF [55]. Since Deblur-NeRF [26], BAGS [34], and Deblurring 3DGS [18] do not incorporate a lens model, they cannot generate novel view images with a specific DoF. As a result, these methods cannot be evaluated on this dataset due to their lack of refocusing capability." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.789, + 0.245, + 0.805 + ], + "angle": 0, + "content": "5.3. Ablation Study" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.81, + 0.483, + 0.901 + ], + "angle": 0, + "content": "As shown in Table 4 and Fig. 8, we conduct ablation studies to evaluate the effectiveness of our designs. The comparison between #1 and #2 indicates that the lens-based imaging model has a significant impact. This lens model not only facilitates defocus deblurring but also provides the ability to control the depth of field. Next, the defocus-to-focus" + }, + { + "type": "table_caption", + "bbox": [ + 0.525, + 0.371, + 0.895, + 0.386 + ], + "angle": 0, + "content": "Table 4. Ablation study on each component of our method." + }, + { + "type": "table", + "bbox": [ + 0.518, + 0.397, + 0.902, + 0.489 + ], + "angle": 0, + "content": "
LensDepthDefocus-to-focusPSNR↑SSIM↑LPIPS↓
#1XXX21.310.6360.239
#2✓XX23.050.7280.109
#3✓X✓23.590.7420.104
#4✓✓X23.420.7380.098
#5✓✓✓23.970.7560.093
" + }, + { + "type": "table_caption", + "bbox": [ + 0.554, + 0.521, + 0.866, + 0.536 + ], + "angle": 0, + "content": "Table 5. Ablation study on different depth priors." + }, + { + "type": "table", + "bbox": [ + 0.536, + 0.546, + 0.88, + 0.621 + ], + "angle": 0, + "content": "
MethodPSNR↑SSIM↑LPIPS↓
No fine-tuned depth23.530.7390.114
Sparse depth23.440.7370.113
Ours23.970.7560.093
" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.664, + 0.907, + 0.74 + ], + "angle": 0, + "content": "adaptation (#3) and the per-scene adjustment of depth priors (#4) further enhance the scene geometry and improve the deblurring of details. Ultimately, the combination of all components yields the highest performance gain, leading to a 2.66 dB increase in PSNR over the baseline model." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.75, + 0.909, + 0.903 + ], + "angle": 0, + "content": "In addition, we conduct an ablation study on different depth supervision methods to demonstrate the superiority of our per-scene adjustment of depth priors, as shown in Table 5. Direct supervision using depth maps predicted by the depth network without scene-specific fine-tuning hinders the reconstruction quality of 3D-GS. Likewise, supervision based on sparse depth maps projected from a 3D point cloud fails to yield optimal results. We show visualizations of depth maps rendered by different depth supervision strategies in the supplementary material." + }, + { + "type": "page_number", + "bbox": [ + 0.479, + 0.945, + 0.521, + 0.957 + ], + "angle": 0, + "content": "26468" + } + ], + [ + { + "type": "image_caption", + "bbox": [ + 0.318, + 0.088, + 0.654, + 0.103 + ], + "angle": 0, + "content": "(I) Change the aperture size and focus distance" + }, + { + "type": "image", + "bbox": [ + 0.094, + 0.105, + 0.905, + 0.19 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.319, + 0.193, + 0.674, + 0.209 + ], + "angle": 0, + "content": "(II) Change the lens parameters and camera poses" + }, + { + "type": "image", + "bbox": [ + 0.094, + 0.209, + 0.905, + 0.295 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.319, + 0.297, + 0.643, + 0.312 + ], + "angle": 0, + "content": "(III) Change the lens parameters and zoom-in" + }, + { + "type": "image", + "bbox": [ + 0.094, + 0.313, + 0.905, + 0.399 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.09, + 0.41, + 0.907, + 0.438 + ], + "angle": 0, + "content": "Figure 7. Applications of our controllable DoF effects. Users can create their own cinematic moments by combining changes in aperture size, focus distance, camera poses, and zoom." + }, + { + "type": "image", + "bbox": [ + 0.094, + 0.448, + 0.483, + 0.61 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.09, + 0.62, + 0.483, + 0.662 + ], + "angle": 0, + "content": "Figure 8. The visualizations of the ablation study. Qualitative comparisons show that the full model (#5) yields the best result and the reconstructed strawberry is the sharpest." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.682, + 0.312, + 0.699 + ], + "angle": 0, + "content": "5.4. Interactive Applications" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.705, + 0.484, + 0.901 + ], + "angle": 0, + "content": "Thanks to the lens-based model, we can not only represent 3D scenes from shallow DoF inputs but also render interactive DoF effects (see Fig. 7). Various cinematography styles can be achieved through different camera configurations, and users can input custom datasets, whether all-in-focus or bokeh images. By controlling aperture size and focus distance, we can render novel view images with different depths of field. In addition, we can modify the shape of the CoC, such as changing it from a circle to a hexagon. Users can also adjust the depth of field while moving the camera or zooming to create cinematic effects. Unlike [55], our GS framework significantly enhances the training and rendering efficiency, allowing for faster and more seamless inter" + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.452, + 0.907, + 0.482 + ], + "angle": 0, + "content": "actions. The details on processing time will be discussed in the supplementary material." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.497, + 0.634, + 0.513 + ], + "angle": 0, + "content": "6. Conclusion" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.523, + 0.907, + 0.719 + ], + "angle": 0, + "content": "While 3D Gaussian Splatting methods have achieved impressive performance on a wide range of reconstruction tasks, the applications on controllable DoF effects remain challenging and understudied. In this paper, we propose DoF-Gaussian, a controllable DoF method for 3D Gaussian Splatting. Specifically, we develop a lens-based model rather than pinhole imaging to overcome the limitations imposed by shallow DoF inputs for 3D-GS. Furthermore, we propose the per-scene depth adjustment and a defocus-to-focus adaptation to guarantee the performance of defocus deblurring. We also introduce a synthetic dataset for a more comprehensive evaluation. Thanks to the imaging principles, our method supports various interactive applications." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.72, + 0.906, + 0.84 + ], + "angle": 0, + "content": "Future work. Inspired by our experiments, we found that modeling real-world physical imaging principles can enable our method not only to handle shallow DoF inputs but also to perform effectively on general inputs. This insight motivates us to explore a combination of non-ideal conditions, such as shallow DoF inputs, sparse views, and varied lighting environments, which are more aligned with casual photography in our daily lives." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.84, + 0.906, + 0.901 + ], + "angle": 0, + "content": "Acknowledgement. This study is supported under the RIE2020 Industry Alignment Fund - Industry Collaboration Projects (IAF-ICP) Funding Initiative, as well as cash and in-kind contribution from the industry partner(s)." + }, + { + "type": "page_number", + "bbox": [ + 0.479, + 0.945, + 0.521, + 0.957 + ], + "angle": 0, + "content": "26469" + } + ], + [ + { + "type": "title", + "bbox": [ + 0.093, + 0.09, + 0.188, + 0.106 + ], + "angle": 0, + "content": "References" + }, + { + "type": "ref_text", + "bbox": [ + 0.101, + 0.115, + 0.484, + 0.169 + ], + "angle": 0, + "content": "[1] Guillaume Abadie, Steve McAuley, Evegenii Golubev, Stephen Hill, and Sebastien Lagarde. Advances in real-time rendering in games. In ACM SIGGRAPH 2018 Courses. 2018. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.101, + 0.173, + 0.484, + 0.24 + ], + "angle": 0, + "content": "[2] Aleksei Bochkovskii, Amaël Delaunoy, Hugo Germain, Marcel Santos, Yichao Zhou, Stephan R Richter, and Vladlen Koltun. Depth pro: Sharp monocular metric depth in less than a second. arXiv preprint arXiv:2410.02073, 2024. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.244, + 0.483, + 0.284 + ], + "angle": 0, + "content": "[3] Benjamin Busam, Matthieu Hog, Steven McDonagh, and Gregory Slabaugh. Sterefo: Efficient image refocusing with stereo vision. In ICCVW, 2019. 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.287, + 0.482, + 0.327 + ], + "angle": 0, + "content": "[4] Wenbo Chen and Ligang Liu. Deblur-GS: 3D Gaussian Splitting from Camera Motion Blurred Images. PACMCGIT, 7(1), 2024. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.33, + 0.483, + 0.371 + ], + "angle": 0, + "content": "[5] Jaeyoung Chung, Jeongtaek Oh, and Kyoung Mu Lee. Depth-regularized optimization for 3D Gaussian Splitting in few-shot images. In CVPR, 2024. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.374, + 0.483, + 0.441 + ], + "angle": 0, + "content": "[6] Marcos V Conde, Manuel Kolmet, Tim Seizinger, Tom E Bishop, Radu Timofte, Xiangyu Kong, Dafeng Zhang, Jinlong Wu, Fan Wang, Juewen Peng, et al. Lens-to-lens bokeh effect transformation. NTIRE 2023 challenge report. In CVPR, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.445, + 0.483, + 0.486 + ], + "angle": 0, + "content": "[7] François Darmon, Lorenzo Porzi, Samuel Rota-Bulò, and Peter Kontschieder. Robust gaussian splatting. arXiv preprint arXiv:2404.04211, 2024. 2, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.488, + 0.483, + 0.542 + ], + "angle": 0, + "content": "[8] Saikat Dutta, Sourya Dipta Das, Nisarg A Shah, and Anil Kumar Tiwari. Stacked deep multi-scale hierarchical network for fast bokeh effect rendering from a single image. In CVPR, 2021. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.545, + 0.483, + 0.585 + ], + "angle": 0, + "content": "[9] David Eigen, Christian Puhrsch, and Rob Fergus. Depth map prediction from a single image using a multi-scale deep network. NeurIPS, 27, 2014. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.588, + 0.465, + 0.602 + ], + "angle": 0, + "content": "[10] Eugene Hecht. Optics. Pearson Education India, 2012. 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.604, + 0.483, + 0.658 + ], + "angle": 0, + "content": "[11] Binbin Huang, Zehao Yu, Anpei Chen, Andreas Geiger, and Shenghua Gao. 2D Gaussian Splatting for geometrically accurate radiance fields. In ACM SIGGRAPH 2024 Conference Papers, 2024. 2, 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.661, + 0.483, + 0.728 + ], + "angle": 0, + "content": "[12] Andrey Ignatov, Jagruti Patel, Radu Timofte, Bolun Zheng, Xin Ye, Li Huang, Xiang Tian, Saikat Dutta, Kuldeep Purohit, Praveen Kandula, et al. Aim 2019 challenge on bokeh effect synthesis: Methods and results. In ICCVW. IEEE, 2019. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.732, + 0.483, + 0.771 + ], + "angle": 0, + "content": "[13] Andrey Ignatov, Jagruti Patel, and Radu Timofte. Rendering natural camera bokeh effect with deep learning. In CVPRW, 2020." + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.776, + 0.483, + 0.83 + ], + "angle": 0, + "content": "[14] Andrey Ignatov, Radu Timofte, Ming Qian, Congyu Qiao, Jiamin Lin, Zhenyu Guo, Chenghua Li, Cong Leng, Jian Cheng, Juewen Peng, et al. AIM 2020 challenge on rendering realistic bokeh. In ECCVW. Springer, 2020. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.832, + 0.483, + 0.9 + ], + "angle": 0, + "content": "[15] Xin Jin, Pengyi Jiao, Zheng-Peng Duan, Xingchao Yang, Chun-Le Guo, Bo Ren, and Chongyi Li. Lighting Every Darkness with 3DGS: Fast Training and Real-Time Rendering for HDR View Synthesis. arXiv preprint arXiv:2406.06216, 2024.3" + }, + { + "type": "list", + "bbox": [ + 0.094, + 0.115, + 0.484, + 0.9 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.093, + 0.905, + 0.134 + ], + "angle": 0, + "content": "[16] Bernhard Kerbl, Georgios Kopanas, Thomas Leimkuhler, and George Drettakis. 3D Gaussian Splatting for Real-Time Radiance Field Rendering. ACM TOG, 42(4), 2023. 2, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.136, + 0.905, + 0.175 + ], + "angle": 0, + "content": "[17] Min-Jung Kim, Gyojung Gu, and Jaegul Choo. LensNeRF: Rethinking volume rendering based on thin-lens camera model. In WACV, 2024. 2, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.178, + 0.905, + 0.218 + ], + "angle": 0, + "content": "[18] Byeonghyeon Lee, Howoong Lee, Xiangyu Sun, Usman Ali, and Eunbyung Park. Deblurring 3D Gaussian Splatting. arXiv preprint arXiv:2401.00834, 2024. 2, 3, 6, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.22, + 0.905, + 0.259 + ], + "angle": 0, + "content": "[19] Dogyoon Lee, Minhyeok Lee, Chajin Shin, and Sangyoun Lee. Dp-NeRF: Deblurred neural radiance field with physical scene priors. In CVPR, 2023. 2, 3, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.261, + 0.905, + 0.3 + ], + "angle": 0, + "content": "[20] Sungkil Lee, Elmar Eisemann, and Hans-Peter Seidel. Realtime lens blur effects and focus control. ACM TOG, 29(4), 2010. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.302, + 0.905, + 0.356 + ], + "angle": 0, + "content": "[21] Jiahe Li, Jiawei Zhang, Xiao Bai, Jin Zheng, Xin Ning, Jun Zhou, and Lin Gu. DNGaussian: Optimizing sparse-view 3D Gaussian radiance fields with global-local depth normalization. In CVPR, 2024. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.359, + 0.905, + 0.412 + ], + "angle": 0, + "content": "[22] Zhengqi Li, Tali Dekel, Forrester Cole, Richard Tucker, Noah Snavely, Ce Liu, and William T Freeman. Learning the depths of moving people by watching frozen people. In CVPR, 2019. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.414, + 0.905, + 0.481 + ], + "angle": 0, + "content": "[23] Tianqi Liu, Guangcong Wang, Shoukang Hu, Liao Shen, Xinyi Ye, Yuhang Zang, Zhiguo Cao, Wei Li, and Ziwei Liu. MVSGaussian: Fast Generalizable Gaussian Splatting Reconstruction from Multi-View Stereo. In ECCV. Springer, 2025. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.484, + 0.905, + 0.524 + ], + "angle": 0, + "content": "[24] Xianrui Luo, Huiqiang Sun, Juewen Peng, and Zhiguo Cao. Dynamic Neural Radiance Field From Defocused Monocular Video. arXiv preprint arXiv:2407.05586, 2024. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.526, + 0.905, + 0.566 + ], + "angle": 0, + "content": "[25] Yawen Luo, Min Shi, Liao Shen, Yachuan Huang, Zixuan Ye, Juewen Peng, and Zhiguo Cao. Video bokeh rendering: Make casual videography cinematic. In ACM MM, 2024. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.568, + 0.905, + 0.608 + ], + "angle": 0, + "content": "[26] Li Ma, Xiaoyu Li, Jing Liao, Qi Zhang, Xuan Wang, Jue Wang, and Pedro V Sander. Deblur-NeRF: Neural radiance fields from blurry images. In CVPR, 2022. 2, 3, 6, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.61, + 0.905, + 0.663 + ], + "angle": 0, + "content": "[27] Ricardo Martin-Brualla, Noha Radwan, Mehdi SM Sajjadi, Jonathan T Barron, Alexey Dosovitskiy, and Daniel Duckworth. NeRF in the wild: Neural radiance fields for unconstrained photo collections. In CVPR, 2021. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.665, + 0.905, + 0.72 + ], + "angle": 0, + "content": "[28] Ben Mildenhall, Pratul P Srinivasan, Matthew Tancik, Jonathan T Barron, Ravi Ramamoorthi, and Ren Ng. NeRF: Representing scenes as neural radiance fields for view synthesis. Communications of the ACM, 65(1), 2021. 2, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.721, + 0.905, + 0.775 + ], + "angle": 0, + "content": "[29] Ben Mildenhall, Peter Hedman, Ricardo Martin-Brualla, Pratul P Srinivasan, and Jonathan T Barron. NeRF in the dark: High dynamic range view synthesis from noisy raw images. In CVPR, 2022. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.777, + 0.905, + 0.844 + ], + "angle": 0, + "content": "[30] Nicolas Moenne-Loccoz, Ashkan Mirzaei, Or Perel, Ricardo de Lutio, Janick Martinez Esturo, Gavriel State, Sanja Fidler, Nicholas Sharp, and Zan Gojcic. 3d gaussian ray tracing: Fast tracing of particle scenes. ACM Transactions on Graphics (TOG), 43(6):1-19, 2024. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.846, + 0.905, + 0.9 + ], + "angle": 0, + "content": "[31] Keunhong Park, Utkarsh Sinha, Jonathan T Barron, Sofien Bouaziz, Dan B Goldman, Steven M Seitz, and Ricardo Martin-Brualla. Nerfies: Deformable neural radiance fields. In ICCV, 2021. 3" + }, + { + "type": "list", + "bbox": [ + 0.517, + 0.093, + 0.905, + 0.9 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.48, + 0.945, + 0.52, + 0.957 + ], + "angle": 0, + "content": "26470" + } + ], + [ + { + "type": "ref_text", + "bbox": [ + 0.092, + 0.093, + 0.482, + 0.12 + ], + "angle": 0, + "content": "[32] Frank L Pedrotti, Leno M Pedrotti, and Leno S Pedrotti. Introduction to optics. Cambridge University Press, 2018. 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.091, + 0.122, + 0.483, + 0.163 + ], + "angle": 0, + "content": "[33] Cheng Peng and Rama Chellappa. PDRF: progressively deblurring radiance field for fast scene reconstruction from blurry images. In AAAI, 2023. 2, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.165, + 0.483, + 0.22 + ], + "angle": 0, + "content": "[34] Cheng Peng, Yutao Tang, Yifan Zhou, Nengyu Wang, Xijun Liu, Deming Li, and Rama Chellappa. BAGS: Blur Agnostic Gaussian Splatting through Multi-Scale Kernel Modeling. arXiv preprint arXiv:2403.04926, 2024. 2, 3, 5, 6, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.222, + 0.483, + 0.262 + ], + "angle": 0, + "content": "[35] Juewen Peng, Zhiguo Cao, Xianrui Luo, Hao Lu, Ke Xian, and Jianming Zhang. Bokehme: When neural rendering meets classical rendering. In CVPR, 2022. 2, 6, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.264, + 0.484, + 0.333 + ], + "angle": 0, + "content": "[36] Juewen Peng, Zhiyu Pan, Chengxin Liu, Xianrui Luo, Huiqiang Sun, Liao Shen, Ke Xian, and Zhiguo Cao. Selective bokeh effect transformation. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 1456-1464, 2023." + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.335, + 0.483, + 0.404 + ], + "angle": 0, + "content": "[37] Dominique Piché-Meunier, Yannick Hold-Geoffroy, Jianming Zhang, and Jean-François Lalonde. Lens parameter estimation for realistic depth of field modeling. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 499–508, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.406, + 0.483, + 0.461 + ], + "angle": 0, + "content": "[38] Ming Qian, Congyu Qiao, Jiamin Lin, Zhenyu Guo, Chenghua Li, Cong Leng, and Jian Cheng. BgGAN: Bokehglass generative adversarial network for rendering realistic bokeh. In ECCVW. Springer, 2020. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.463, + 0.483, + 0.517 + ], + "angle": 0, + "content": "[39] René Ranftl, Katrin Lasinger, David Hafner, Konrad Schindler, and Vladlen Koltun. Towards robust monocular depth estimation: Mixing datasets for zero-shot cross-dataset transfer. IEEE TPAMI, 44(3), 2020. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.519, + 0.483, + 0.546 + ], + "angle": 0, + "content": "[40] Johannes L Schonberger and Jan-Michael Frahm. Structure-from-motion revisited. In CVPR, 2016. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.549, + 0.483, + 0.589 + ], + "angle": 0, + "content": "[41] Johannes L. Schonberger, Enliang Zheng, Jan-Michael Frahm, and Marc Pollefeys. Pixelwise view selection for unstructured multi-view stereo. In ECCV. Springer, 2016. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.591, + 0.483, + 0.66 + ], + "angle": 0, + "content": "[42] Liao Shen, Xingyi Li, Huiqiang Sun, Juewen Peng, Ke Xian, Zhiguo Cao, and Guosheng Lin. Make-it-4d: Synthesizing a consistent long-term dynamic scene video from a single image. In Proceedings of the 31st ACM International Conference on Multimedia, pages 8167–8175, 2023. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.662, + 0.483, + 0.716 + ], + "angle": 0, + "content": "[43] Yichen Sheng, Zixun Yu, Lu Ling, Zhiwen Cao, Xuaner Zhang, Xin Lu, Ke Xian, Haiting Lin, and Bedrich Benes. Dr. Bokeh: DiffeRentiable Occlusion-aware Bokeh Rendering. In CVPR, 2024. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.719, + 0.483, + 0.76 + ], + "angle": 0, + "content": "[44] Huiqiang Sun, Xingyi Li, Liao Shen, Xinyi Ye, Ke Xian, and Zhiguo Cao. DyBluRF: Dynamic Neural Radiance Fields from Blurry Monocular Video. In CVPR, 2024. 2, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.762, + 0.483, + 0.829 + ], + "angle": 0, + "content": "[45] Neal Wadhwa, Rahul Garg, David E Jacobs, Bryan E Feldman, Nori Kanazawa, Robert Carroll, Yair Movshovitz-Attias, Jonathan T Barron, Yael Pritch, and Marc Levoy. Synthetic depth-of-field with a single-camera mobile phone. ACM TOG, 37(4), 2018. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.832, + 0.483, + 0.901 + ], + "angle": 0, + "content": "[46] Chao Wang, Krzysztof Wolski, Xingang Pan, Thomas Leimkuhler, Bin Chen, Christian Theobalt, Karol Myszkowski, Hans-Peter Seidel, and Ana Serrano. An implicit neural representation for the image stack: Depth, all in focus, and high dynamic range. Technical report, 2023. 2" + }, + { + "type": "list", + "bbox": [ + 0.091, + 0.093, + 0.484, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.093, + 0.905, + 0.161 + ], + "angle": 0, + "content": "[47] Chao Wang, Krzysztof Wolski, Bernhard Kerbl, Ana Serrano, Mojtaba Bemana, Hans-Peter Seidel, Karol Myszkowski, and Thomas Leimkuhler. Cinematic Gaussians: Real-Time HDR Radiance Fields with Depth of Field. arXiv preprint arXiv:2406.07329, 2024. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.163, + 0.905, + 0.218 + ], + "angle": 0, + "content": "[48] Lijun Wang, Xiaohui Shen, Jianming Zhang, Oliver Wang, Zhe Lin, Chih-Yao Hsieh, Sarah Kong, and Huchuan Lu. Deep learning: Shallow depth of field from a single image. arXiv preprint arXiv:1810.08100, 2018. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.219, + 0.905, + 0.259 + ], + "angle": 0, + "content": "[49] Peng Wang, Lingzhe Zhao, Ruijie Ma, and Peidong Liu. BAD-NeRF: Bundle adjusted deblur neural radiance fields. In CVPR, 2023. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.261, + 0.904, + 0.302 + ], + "angle": 0, + "content": "[50] Yinhuai Wang, Shuzhou Yang, Yujie Hu, and Jian Zhang. NeRFocus: Neural radiance field for 3D synthetic defocus. arXiv preprint arXiv:2203.05189, 2022. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.304, + 0.905, + 0.357 + ], + "angle": 0, + "content": "[51] Yujie Wang, Praneeth Chakravarthula, and Baoquan Chen. DOF-GS: Adjustable Depth-of-Field 3D Gaussian Splitting for Refocusing, Defocus Rendering and Blur Removal. arXiv preprint arXiv:2405.17351, 2024. 3, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.359, + 0.905, + 0.399 + ], + "angle": 0, + "content": "[52] Zhou Wang, Alan C Bovik, Hamid R Sheikh, and Eero P Simoncelli. Image quality assessment: from error visibility to structural similarity. IEEE TIP, 13(4), 2004. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.401, + 0.905, + 0.453 + ], + "angle": 0, + "content": "[53] Yi Wei, Shaohui Liu, Yongming Rao, Wang Zhao, Jiwen Lu, and Jie Zhou. Nerfingmvs: Guided optimization of neural radiance fields for indoor multi-view stereo. In ICCV, 2021. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.456, + 0.905, + 0.509 + ], + "angle": 0, + "content": "[54] Guanjun Wu, Taoran Yi, Jiemin Fang, Lingxi Xie, Xiaopeng Zhang, Wei Wei, Wenyu Liu, Qi Tian, and Xinggang Wang. 4d gaussian splatting for real-time dynamic scene rendering. In CVPR, 2024. 2, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.511, + 0.905, + 0.551 + ], + "angle": 0, + "content": "[55] Zijin Wu, Xingyi Li, Juewen Peng, Hao Lu, Zhiguo Cao, and Weicai Zhong. DoF-NeRF: Depth-of-field meets neural radiance fields. In ACM MM, 2022. 1, 2, 3, 6, 7, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.554, + 0.905, + 0.607 + ], + "angle": 0, + "content": "[56] Lei Xiao, Anton Kaplanyan, Alexander Fix, Matt Chapman, and Douglas Lanman. Deepfocus: Learned image synthesis for computational display. In ACM SIGGRAPH 2018 Talks. 2018. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.609, + 0.905, + 0.663 + ], + "angle": 0, + "content": "[57] Haofei Xu, Songyou Peng, Fangjinhua Wang, Hermann Blum, Daniel Barath, Andreas Geiger, and Marc Pollefeys. Depthsplat: Connecting gaussian splatting and depth. arXiv preprint arXiv:2410.13862, 2024. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.665, + 0.905, + 0.705 + ], + "angle": 0, + "content": "[58] Zehao Yu, Anpei Chen, Binbin Huang, Torsten Sattler, and Andreas Geiger. Mip-splatting: Alias-free 3D gaussian splattering. In CVPR, 2024. 2, 5, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.707, + 0.905, + 0.76 + ], + "angle": 0, + "content": "[59] Baowen Zhang, Chuan Fang, Rakesh Shrestha, Yixun Liang, Xiaoxiao Long, and Ping Tan. RaDe-GS: Rasterizing Depth in Gaussian Splatting. arXiv preprint arXiv:2406.01467, 2024.5" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.762, + 0.905, + 0.802 + ], + "angle": 0, + "content": "[60] Richard Zhang, Phillip Isola, Alexei A Efros, Eli Shechtman, and Oliver Wang. The unreasonable effectiveness of deep features as a perceptual metric. In CVPR, 2018. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.804, + 0.905, + 0.858 + ], + "angle": 0, + "content": "[61] Xuaner Zhang, Kevin Matzen, Vivien Nguyen, Dillon Yao, You Zhang, and Ren Ng. Synthetic defocus and lookahead autofocus for casual videography. arXiv preprint arXiv:1905.06326, 2019. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.86, + 0.905, + 0.901 + ], + "angle": 0, + "content": "[62] Lingzhe Zhao, Peng Wang, and Peidong Liu. BAD-Gaussians: Bundle adjusted deblur Gaussian Splatting. arXiv preprint arXiv:2403.11831, 2024. 3" + }, + { + "type": "list", + "bbox": [ + 0.517, + 0.093, + 0.905, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.48, + 0.945, + 0.518, + 0.957 + ], + "angle": 0, + "content": "26471" + } + ] +] \ No newline at end of file diff --git a/2025/DoF-Gaussian_ Controllable Depth-of-Field for 3D Gaussian Splatting/1ef9852f-e1e4-4571-8156-b74884315f2d_origin.pdf b/2025/DoF-Gaussian_ Controllable Depth-of-Field for 3D Gaussian Splatting/1ef9852f-e1e4-4571-8156-b74884315f2d_origin.pdf new file mode 100644 index 0000000000000000000000000000000000000000..0210f0153cc3e79f8f9f280e8dd8a07da3c65355 --- /dev/null +++ b/2025/DoF-Gaussian_ Controllable Depth-of-Field for 3D Gaussian Splatting/1ef9852f-e1e4-4571-8156-b74884315f2d_origin.pdf @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:0bd6751c7668c7671ea4fa5c55051b4710883fdcaca2ab6ee413155fc45e9728 +size 2638057 diff --git a/2025/DoF-Gaussian_ Controllable Depth-of-Field for 3D Gaussian Splatting/full.md b/2025/DoF-Gaussian_ Controllable Depth-of-Field for 3D Gaussian Splatting/full.md new file mode 100644 index 0000000000000000000000000000000000000000..9d72440138f71ef312c2244e4720cd8760b37b1e --- /dev/null +++ b/2025/DoF-Gaussian_ Controllable Depth-of-Field for 3D Gaussian Splatting/full.md @@ -0,0 +1,387 @@ +# DoF-Gaussian: Controllable Depth-of-Field for 3D Gaussian Splatting + +Liao Shen $^{1,2}$ Tianqi Liu $^{1,2}$ Huiqiang Sun $^{1,2}$ Jiaqi Li $^{1}$ Zhiguo Cao $^{1}$ Wei Li $^{2*}$ Chen Change Loy $^{2}$ $^{1}$ School of AIA, Huazhong University of Science and Technology + $^{2}$ S-Lab, Nanyang Technological University + +leoshen@hust.edu.cn + +![](images/96b19129ee7253db8f2beb70a89972b0937c4e8dc4a369d8af6ad4fdd2897bb3.jpg) + +![](images/ef42b940e1b3a757249e06aeb98210b90570586c95ef8ebd901c80abcc82ae68.jpg) +DoF-NeRF + +![](images/d8a776f82313fb89e47962f106e53fd310f39b1a5a09d02c0e52deee350decf2.jpg) +BAGS + +![](images/4b3768a91c66cecb078a2c7eff329ba6b4116e9a9db37c239b2935f8100ee7c1.jpg) +Ours + +![](images/e3b2ffe15ba989b5ac1a70e607113ff8e23daa3a4026a532b16f75a665645513.jpg) +Change aperture size and focus distance + +![](images/becc542fe2a2639b9769c7749c819fe5ab65a7cd6ac1e9808f67cc32958b692a.jpg) + +![](images/e599fd807f897a202c19e0a47d1a2ef608ad32fe2903fbe3d3bc71e6c99f9b3f.jpg) +Change camera poses and zoom-in +Figure 1. Given a set of multi-view input images with shallow DoF, DoF-Gaussian can reconstruct a 3D-GS representation of a sharp scene. Thanks to our lens-based design, we can also achieve controllable DoF effects for a variety of applications. The example input images are taken from [55] for illustration purposes. (Zoom-in for best view) + +![](images/14bd89e418f7a6774affdb731d0a4895c49b187f124105341b3b097982a2a1d6.jpg) +Change aperture shape + +# Abstract + +Recent advances in 3D Gaussian Splatting (3D-GS) have shown remarkable success in representing 3D scenes and generating high-quality, novel views in real-time. However, 3D-GS and its variants assume that input images are captured based on pinhole imaging and are fully in focus. This assumption limits their applicability, as real-world images often feature shallow depth-of-field (DoF). In this paper, we introduce DoF-Gaussian, a controllable depth-of-field method for 3D-GS. We develop a lens-based imaging model based on geometric optics principles to control DoF effects. To ensure accurate scene geometry, we incorporate depth priors adjusted per scene, and we apply defocus-to-focus adaptation to minimize the gap in the circle of confusion. We also introduce a synthetic dataset to assess refocusing capabilities and the model's ability to learn precise lens parameters. Our framework is customizable and supports various interactive applications. Extensive experiments confirm the effectiveness of our method. Our project is avail- + +able at https://dof-gaussian.github.io/. + +# 1. Introduction + +Depth-of-field (DoF) refers to the distance between the closest and farthest objects in a photo that appears acceptably sharp. In practice, photographers can control the DoF by adjusting the camera's aperture size or focus distance to capture images with either a wide DoF (all-in-focus) or shallow DoF (defocused). The shallow DoF effect is an important technique in photography, as it draws the viewers' attention to the focal region by blurring the surrounding areas. Meanwhile, novel view synthesis aims to create realistic images from novel viewpoints based on a set of source images. However, novel view synthesis typically requires all-in-focus input images and lacks the capability to render varied DoF effects, limiting its applications. In this work, we aim to render novel views with controllable DoF, adding cinematic quality to the results. + +Images captured from the real world sometimes have shallow DoF. Specifically, points of light that do not lie on + +the focal plane are projected onto the sensor plane as blurred circles, referred to as the circle of confusion (CoC), and this introduces bokeh blur to the captured image. Most novel view synthesis methods experience performance degradation when processing shallow DoF input images. To address this issue, DoF-NeRF [55] and LensNeRF [17] introduce lens-based camera models in volume rendering to explicitly enable controllable DoF effects. Meanwhile, methods like Deblur-NeRF [26], DP-NeRF [19], and PDRF [33] employ sparse kernels to model defocus blur. However, implicit rendering approaches face significant challenges in training and rendering efficiency. In the realm of 3D Gaussian Splatting (3D-GS) [16], methods such as BAGS [34] and Deblurring 3DGS [18] propose blur estimation networks to model the blur kernels or scaling factors for defocus deblurring, but they cannot accommodate controllable DoF. + +In this paper, we present DoF-Gaussian, an efficient framework for controllable DoF in 3D-GS, addressing the limitations of existing methods in handling shallow DoF inputs. This problem is non-trivial because 3D-GS and its variants [11, 23, 54, 57, 58] are typically based on a pin-hole camera model that assumes all-in-focus inputs, meaning both foreground and background appear clear. When input images contain bokeh blur, it becomes challenging to accurately construct scene geometry and render depth maps. However, rendering images with different DoFs relies heavily on precise depth estimation. Additionally, our 3D-GS approach assumes an idealized CoC, which inevitably differs from the CoC seen in real photographs, posing further challenges for accurate defocus deblurring. Furthermore, existing datasets in this field are primarily borrowed from deblurring applications, limiting our evaluation to the model's defocus deblurring capability and overlooking other aspects of controllable DoF effects. + +To address these challenges, we present a new controllable DoF 3D-GS method with the following contributions: First, we employ a lens-based model instead of a pinhole-based one. We make the lens optical parameters, such as aperture size and focus distance, learnable to enable control over the depth of field. Second, we introduce per-scene adjustments for depth priors to minimize errors in scene geometry and depth relationships, ensuring accurate depth-of-field control. Third, we propose a defocus-to-focus adaptation strategy that focuses on learning the focal region after the learnable lens parameters have converged, compensating for differences in CoC. Finally, we present a synthetic dataset designed to comprehensively evaluate the model, including its refocusing capabilities and its ability to learn accurate aperture size and focus distance. + +Benefiting from the lens-based imaging model, our approach enables controlled depth of field, allowing it to handle not only shallow DoF image inputs but also unlock a range of interactive and engaging applications, as depicted + +in Figure 1. For instance, users can render novel view images with varied DoF, refocus on custom datasets, and even adjust the shape of CoC, e.g., from circular to pentagonal or hexagonal CoC. Moreover, users can dynamically change the DoF while moving the camera or zooming, rendering videos with cinematic effects. Extensive experiments demonstrate that our method outperforms state-of-the-art depth-of-field and defocus deblurring methods. + +# 2. Related Work + +Depth-of-Field Rendering. Rendering DoF effects from a single all-in-focus image has been widely explored in previous work. Bokeh blur occurs when light is projected onto the image plane as a circular region, rather than as a point. The size of CoC is affected by the diameter of the aperture (aperture size) and the distance from the camera to the focal plane (focus distance). Photos captured with a small aperture usually present a wide DoF, i.e., all objects appear sharp. In contrast, as the aperture diameter increases, objects near the focal plane remain sharp, while those farther away become blurred with a larger CoC. Physically based methods [1, 20] rely on 3D scene geometry information and are time-consuming. Some methods use neural networks [8, 12–14, 38] trained end-to-end to obtain shallow DoF images. The DoF effects in some studies [6, 25, 35–37, 43, 45, 48, 56, 61] are controllable but usually require an extra disparity map. Wang et al. [46] combines neural fields with an expressive camera model to achieve all-in-focus reconstruction from an image stack. In this work, we show the connection between DoF rendering and novel view synthesis using a lens-based camera model. + +Image Deblurring. Blur can generally be categorized into two main types: camera motion blur and defocus blur. Previous studies [7, 18, 24, 26, 33, 34, 44] attempt to address this issue. Our task of rendering all-in-focus and sharp novel view images from shallow DoF inputs can be considered as a form of defocus deblurring. By modeling the underlying physical principles, we can achieve defocus deblurring by simulating the circle-of-confusion. Furthermore, our method allows for adjustments to lens parameters, enabling applications such as refocusing—capabilities that deblurring approaches like BAGS [34] cannot achieve. Existing datasets in the 3D DoF field are often borrowed from the deblurring domain [26] or designed solely to assess defocus deblurring ability [55]. To support a wider range of applications and validate the accuracy of the learned lens model, we introduce a synthetic dataset for more comprehensive evaluations. + +Novel View Synthesis. Novel view synthesis allows the rendering of unseen camera perspectives from 2D images and their corresponding camera poses. Recent advancements in synthesis can largely be attributed to Neural Radiance Field [28] and 3D Gaussian Splatting [16]. However, + +![](images/5cc941806fcebb8e511318e455f0cce1f991551e558ae064a5f20dc3492106b5.jpg) +Figure 2. Overview of DoF-Gaussian. Given input images $I$ with shallow DoF, we first apply SfM from COLMAP to obtain sparse depth $D_{sparse}$ , which is used to train a depth network to derive per-scene depth priors $D_{pred}$ . We then employ $D_{pred}$ to regularize the Gaussians rendered depth map $D$ . Next, by developing a lens imaging model, we can render defocused images $C^*$ to simulate input images. To minimize the discrepancy in CoC, we propose an adaptation using the weight map. Finally, we can render fully clear images for novel view synthesis and achieve various effects by our controllable DoF framework. + +real-world images are often not clean and well-calibrated. Previous work [4, 27, 29, 31, 42, 44, 49, 54, 62] attempts to cope with non-ideal conditions, such as low light, camera motion, bokeh effects, and various types of image degradation. To recover a sharp scene from images with bokeh blur, methods like Deblur-NeRF [26], DP-NeRF [19], and PDRF [33] model defocus blur using sparse kernels. NeRF Focus [50] and LensNeRF [17] achieve defocus effects by incorporating a lens-based camera model in volume rendering. While DoF-NeRF [55] is similar to our approach, it is limited by the training and rendering efficiency and quality of NeRF. BAGS [34] and Deblurring 3DGS [18] use blur estimation networks to model blur kernels or scaling factors but cannot achieve controllable DoF effects due to the absence of a lens imaging model. Other studies, such as [15, 47], focus on HDR tasks and consider depth-of-field simultaneously. RGS [7] addresses defocus blur by introducing an offset to the 2D covariance matrices of Gaussians. [30] introduces a ray tracing rendering algorithm for particle-based representations to enable many advanced techniques, such as the shallow depth-of-field effect. A concurrent work [51] also explores adjustable depth of field. Our method differs in its controllable DoF mechanism, with the introduction of depth priors to ensure accurate scene depth. In addition, our new defocus-to-focus adaptation strategy eliminates the need for the focus localization network proposed in their work. Besides, we introduce a synthetic dataset for comprehensive evaluation. + +# 3. Preliminary + +3D Gaussian Splitting represents a 3D scene as a mixture of anisotropic 3D Gaussians, where each Gaussian is characterized by a 3D covariance matrix $\Sigma$ and mean $\mu$ : + +$$ +G (X) = e ^ {- \frac {1}{2} (X - \mu) ^ {\top} \Sigma^ {- 1} (X - \mu)} \tag {1} +$$ + +The covariance matrix $\Sigma$ holds physical meaning only when it is positive semi-definite. Therefore, to enable effective optimization via gradient descent, $\Sigma$ is decomposed into a scaling matrix $S$ and a rotation matrix $R$ , as: + +$$ +\Sigma = R S S ^ {\mathsf {T}} R ^ {\mathsf {T}}. \tag {2} +$$ + +To splot Gaussians from 3D space to a 2D plane, the view transformation matrix $W$ and the Jacobian matrix $J$ , which represents the affine approximation of the projective transformation, are utilized to obtain the covariance matrix $\Sigma'$ in 2D space, as: + +$$ +\Sigma^ {\prime} = J W \Sigma W ^ {\mathsf {T}} J ^ {\mathsf {T}}. \tag {3} +$$ + +Subsequently, a point-based alpha-blending rendering can be performed to determine the color of each pixel: + +$$ +C = \sum_ {i} c _ {i} \alpha_ {i} \prod_ {j = 1} ^ {i - 1} (1 - \alpha_ {i}), \tag {4} +$$ + +where $c_{i}$ represents the color of each point, defined by spherical harmonics (SH) coefficients. The density $\alpha_{i}$ is computed as the product of 2D Gaussians and a learnable point-wise opacity. + +![](images/50236d68e6607c558b614534d5b1607caac7d623b4b0319474552c14335540d6.jpg) +(I) Pinhole-based model + +![](images/2ffeb0f907d5bb8b063ab3a6b31f0b73dfdfac3903d1bf989a548bb215a6c2b0.jpg) +(II) Lens-based model +Figure 3. The difference between the pinhole-based model (I) and the lens-based model (II). For (I), light emitted from spatial point $O$ directly hits the point $O'$ of the image plane. For (II), we show the case that light emitted from $Q$ converges on the point $Q'$ of the image plane (In-Focus), and the case that light emitted from $P$ converges on the point $P'$ and continues to scatter onto the image plane forming a circle of confusion (Out-of-Focus). + +# 4. Method + +We propose DoF-Gaussian, a controllable DoF approach for 3D-GS. The overall pipeline is illustrated in Fig. 2. In Section 4.1, we first develop a lens-based imaging model based on geometric optics principles to control DoF effects. Next, we employ the per-scene adjustment of depth priors to guide the correct scene geometry, as described in Section 4.2. Finally, in Section 4.3, we apply defocus-to-focus adaptation to minimize the inconsistencies in the CoC and enhance the defocus deblurring performance. + +# 4.1. Lens-based Depth-of-Field Model + +The physical principles underlying imaging and DoF have been extensively studied in the field of geometric optics [10, 32]. In an idealized optical system, the light emitted from a spatial point $O$ is projected onto a corresponding point $O^{\prime}$ in the image plane following the principles of pinhole imaging, as illustrated in Fig. 3 (I). However, real-world cameras operate based on the lens model, as shown in Fig. 3 (II). We show the two cases of in-focus and out-of-focus imaging, respectively. A spatial point $P$ , located at a distance $d$ from the lens, is projected on the imaging plane as a circular region referred to as the circle of confusion (CoC). The diameter of this region, $r(d)$ , can be determined by the aperture parameter $\mathcal{A}$ and focus distance $\mathcal{F}$ , according to the following equation: + +$$ +r (d) = \mathcal {A} \left| \frac {1}{\mathcal {F}} - \frac {1}{d} \right|. \tag {5} +$$ + +The focus distance $\mathcal{F}$ primarily controls the depth position of the focal plane of the image, i.e., the depth of the sharp area, while the aperture parameter $\mathcal{A}$ determines the extent of the bokeh effect. For the point $Q$ located at the focus distance $d = \mathcal{F}$ , the emitted rays directly converge through the lens to the corresponding point $Q'$ in the image plane, resulting in the absence of a circle of confusion + +# Algorithm 1 Lens-based imaging process + +Input: Rasterization rendering $C$ , aperture parameter $\mathcal{A}$ , focus distance $\mathcal{F}$ , depth map $D$ , gamma value $\gamma$ , confuse function Func + +Output: Defocus simulated result $C^*$ + +$$ +R \leftarrow \mathcal {A} \left| \frac {1}{\bar {\mathcal {E}}} - \frac {1}{\bar {D}} \right| \quad C \leftarrow (C) ^ {\gamma} +$$ + +$$ +\Phi = [ 0 ], \tilde {C ^ {*}} = \tilde {[} 0 ] +$$ + +for pixel $i\gets \mathrm{Traverse}(C)$ do + +for pixel $j\gets$ TraverseNeighbor $(c_i,r_i)$ do + +$$ +\lambda_ {i j} \leftarrow \operatorname {F u n c} (d _ {i}, | i - j |) +$$ + +$$ +\Phi_ {j} \leftarrow \Phi_ {j} + \lambda_ {i j} +$$ + +$$ +c _ {j} ^ {*} \leftarrow c _ {j} ^ {*} + \lambda_ {i j} \cdot c _ {i} +$$ + +end for + +end for + +$$ +C ^ {*} \leftarrow (C ^ {*} / \Phi) ^ {\frac {1}{\gamma}} +$$ + +(CoC). This also implies that $r(\mathcal{F}) = 0$ + +We set focus distance $\mathcal{F}$ and aperture parameter $\mathcal{A}$ as learnable for each input image, and these parameters are continuously updated with the optimization of 3D-GS. By modeling the lens camera, the output $C^*$ , which includes bokeh blur effects, can be derived from the 3D-GS rasterizer rendering result $C$ using our CUDA-based Algorithm 1. + +Ideally, the confuse function is represented by the indicator function $\mathbb{I}(r(d) > l)$ . To achieve a smooth and natural DoF effect, we substitute it with a differentiable function, as proposed in Busam et al. [3]: + +$$ +F u n c (d, l) = \frac {1}{2} + \frac {1}{2} \tanh (\alpha (r (d) - l)), \tag {6} +$$ + +where $\alpha$ defines the smoothness of confuse transition and $l$ represents the distance between two pixels. + +We supervise the training by computing the reconstruction loss between the output image $C^*$ and the shallow DoF input image $I$ , where the loss function is a combination of the $\mathcal{L}_1$ and a D-SSIM term, with $\lambda$ set to 0.2: + +$$ +\mathcal {L} _ {r e c} = (1 - \lambda) \mathcal {L} _ {1} \left(I, C ^ {*}\right) + \lambda \mathcal {L} _ {D - S S I M} \left(I, C ^ {*}\right). \tag {7} +$$ + +For inference, we set the aperture size $\mathcal{A}$ to 0 so that we can render fully clear images for novel view synthesis. + +# 4.2. Per-Scene Adjustment of Depth Priors + +Unlike previous work that assumes input images are all-in-focus and fully clear, it becomes challenging to accurately reconstruct scene geometry and render depth maps for bokeh blurring input images. However, rendering images with varying depths of field relies on precise depth information. To address this issue, we propose using depth prior as useful guidance for the accurate reconstruction of scene geometry. However, due to the presence of bokeh blur in shallow DoF inputs, directly employing a monocular depth network to predict depth maps $D_{pred}$ as pseudo + +![](images/b533d44be00dafbe9b106e039860d236fe97ff55657c7818f35cf452b64abaa3.jpg) +RGB +BAGS +w/o depth priors +Ours +Figure 4. Effects of the depth priors. The per-scene adjustment of depth priors enhance the geometric structure and yield a more accurate depth map than that estimated by BAGS [34] and a variant of our method without depth priors. + +ground truth to regularize the Gaussians rendered depth map, as done in [5, 21], does not yield satisfactory results. Inspired by [53], we propose adapting per-scene depth priors through fine-tuning a monocular depth network based on the scene's sparse reconstruction. Specifically, we utilize COLMAP [40, 41] to obtain a sparse Structure-from-Motion (SfM) point cloud and create a set of 3D Gaussians. Consequently, we can derive per-view sparse depth maps $D_{sparse}$ by projecting the 3D point clouds after multiview stereo processing. Although the resulting depth map is sparse, it is robust and can serve as a supervisory signal for training the per-scene depth priors. To address the scale ambiguity inherent in acquired depth maps, we apply a silog loss [9] to fine-tune the depth network [22, 39], expressed as: + +$$ +\mathcal {L} _ {\text {s i l o g}} = \frac {1}{2 M} \sum_ {i = 1} ^ {M} \left(\log \left(e ^ {s} D _ {\text {p r e d}}\right) - \log \left(D _ {\text {s p a r s e}}\right)\right) ^ {2}, \tag {8} +$$ + +where $e^s$ is the scale factor and $M$ is the number of pixels in the image used for calculation. We apply the rasterization technique from Rade-GS [59], which enables the robust rendering of depth maps $D$ for 3D-GS scenes. We predict the depth maps $D_{pred}$ from multi-view input images using the fine-tuned depth network and use them as depth priors to regularize $D$ . The depth loss can be described as follows: + +$$ +\mathcal {L} _ {\text {d e p t h}} = \left| \left| D - D _ {\text {p r e d}} \right| \right| _ {2}. \tag {9} +$$ + +As illustrated in Fig. 4, it is evident that the depth priors significantly enhance the geometric structure, resulting in a more accurate depth map compared to both BAGS [34] and the variant of our method without depth priors. In Section 5.3, we further conduct an ablation study comparing several other depth supervision strategies to demonstrate the effectiveness of our approach. + +# 4.3. Defocus-to-Focus Adaptation + +The CoC effect achieved under ideal optical imaging conditions, as described in Algorithm 1, inevitably differs from the real CoC effect produced by an actual DSLR camera. + +This discrepancy hinders our ability to accurately model the out-of-focus blur in shallow DoF input images. + +To achieve sharper scenes, we propose an adaptation process that allows transitions from defocus to focus. Specifically, modeling the defocus effects with bokeh blur facilitates learning accurate lens parameters. Consequently, we place greater emphasis on capturing the sharp areas in training images once the lens parameters $\mathcal{A}$ and $\mathcal{F}$ have been optimized and converged. This approach is effective because, when the focus distance is determined, the sharp regions in the scene become approximately identifiable. After $t$ optimization iterations, we reweight the $\mathcal{L}_{rec}$ and $\mathcal{L}_{depth}$ loss by employing a step-like function: + +$$ +\Psi (x) = \left\{ \begin{array}{c c} 1 & , \text {i t e r a t i o n s} < t \\ 1 / \left(1 + e ^ {- a \cdot (x - b)}\right) & , \text {i t e r a t i o n s} \geq t \end{array} \right. \tag {10} +$$ + +where $a$ and $b$ are hyperparameters, and $x = \left|\frac{1}{\mathcal{F}} - \frac{1}{d}\right|$ . + +In addition, the aperture size $\mathcal{A}$ is generally large after the convergence, which can cause blurring of pixels that should be in focus. Hence, we also reweight the aperture size $\mathcal{A}$ on a pixel-wise basis: + +$$ +\mathcal {A} ^ {\prime} = \mathcal {A} \cdot \Psi . \tag {11} +$$ + +Full Objective. We derive the final loss terms by incorporating the reconstruction loss from Eqn. 7, the depth loss from Eqn. 9, and an additional normal consistency loss as proposed by Huang et al. [11]. The normal consistency loss ensures that the Gaussian splats align with the surface by measuring the consistency between the normal directions computed from the Gaussian and the depth map: + +$$ +\mathcal {L} _ {\text {n o r m a l}} = \sum_ {i} w _ {i} \left(1 - n _ {i} ^ {\top} \hat {n} _ {i}\right), \tag {12} +$$ + +where $\hat{n}$ represents the surface normal direction obtained by applying finite differences on the depth map, $i$ indexes the intersected splats along the ray and $w$ denotes the blending weight of the intersection point. + +Our final training loss $\mathcal{L}$ is: + +$$ +\mathcal {L} = \Psi \odot \left(\mathcal {L} _ {\text {r e c}} + w _ {d} \mathcal {L} _ {\text {d e p t h}}\right) + w _ {n} \mathcal {L} _ {\text {n o r m a l}}, \tag {13} +$$ + +where $\odot$ is Hadamard product. + +# 5. Experiments + +Implementation Details. In our experiments, we set the smoothness of confuse transition $\alpha = 4$ . In addition, we set $a = 15$ , $b = 0.3$ , and $t = 10000$ for the defocus-to-focus adaptation. Our method is built upon Mip-Splatting [58], and our optimization strategy and hyperparameter settings remain consistent with it. We train each scene for 30000 iterations and set loss weights $w_{d} = 0.01$ and $w_{n} = 0.05$ . + +Table 1. Quantitative comparisons on the defocus blur dataset of Deblur NeRF [26]. We report the PSNR, SSIM, and LPIPS metrics and color each cell as best and second best. Our method outperforms other existing approaches across most scenes. + +
MethodDeblur-NeRF [26]DoF-NeRF [55]DP-NeRF [19]BAGS [34]Deblurring 3DGS [18]Ours
PSNRSSIMLPIPSPSNRSSIMLPIPSPSNRSSIMLPIPSPSNRSSIMLPIPSPSNRSSIMLPIPSPSNRSSIMLPIPS
Cake26.270.7800.12824.530.7310.20626.160.7780.12727.210.8180.10826.880.8030.11626.830.8080.101
Caps23.870.7130.16122.670.6360.26123.950.7120.14324.160.7250.15924.500.7420.14924.620.7490.147
Cisco20.830.7270.08720.640.7240.10720.730.7260.08420.790.7430.07020.830.7320.07921.000.7440.069
Coral19.850.6000.12119.830.5700.24020.110.6110.11820.530.6280.11119.780.6080.13120.370.6300.109
Cupcake22.260.7220.11621.890.7060.14322.800.7410.09622.930.7620.08022.110.7340.09922.970.7570.079
Cups26.210.7990.12725.260.7650.20226.750.8140.10426.270.8230.10426.280.8240.10326.010.8170.100
Daisy23.520.6870.12123.220.6580.19423.790.6970.10823.740.7460.06223.540.7310.09523.930.7350.071
Sausage18.010.5000.18017.860.4880.28018.350.5440.14718.760.5740.11018.990.5700.14119.110.5760.119
Seal26.040.7770.10524.850.6870.14325.950.7780.10326.520.8120.09026.180.8170.09826.570.8250.087
Tools27.810.8950.06126.210.8540.12828.070.8980.05428.600.9130.04627.960.9100.05828.290.9130.051
+ +Table 2. Quantitative comparisons on our synthetic dataset. + +
MethodPSNR↑SSIM↑LPIPS↓δA↓δF↓
DoF-NeRF [55]25.590.7880.2070.1960.256
Ours28.700.8640.0950.1260.079
+ +Baselines and Evaluation Metrics. We compare our method with state-of-the-art deblurring techniques [18, 26, 34] and a depth-of-field method [55]. To quantitatively evaluate the quality of novel view images, we adopt widely used metrics such as PSNR, SSIM [52], and LPIPS [60]. We will compare with Wang et al. [51] once they release the code or provide the full experimental results in their paper. + +Datasets. Following [18, 26, 34], we evaluate our model on the Deblur-NeRF dataset [26]. Subsequently, we conduct further evaluations on the Real Forward-facing dataset [28] and T&T_DB dataset [16]. The effectiveness of our method is validated on these datasets for both shallow DoF inputs and normal inputs. Furthermore, we also introduce a dataset in Section 5.1 to evaluate our model more comprehensively. + +# 5.1. A Synthetic Dataset + +We introduce a synthetic dataset, as illustrated in Fig. 6, based on depth estimation [2] and a single-image DoF rendering method [35] for each image in the Real Forward Facing dataset [28] and T&T_DB dataset [16]. This allows us to evaluate the refocusing ability of our model and to determine whether it accurately learns the correct aperture size and focus distance. In particular, by feeding the DoF rendering method [35] the known aperture size and focus distance, we can convert the all-in-focus images into shallow DoF images. Unlike existing defocus blur datasets proposed by Ma et al. [26] and Wu et al. [55], the test set in our dataset comprises images with shallow DoF. With known lens parameters, our model fits this shallow DoF effect to quantitatively measure the refocusing ability. Meanwhile, we establish the + +Table 3. Comparisons in all-in-focus settings. We compare our method with Mip-Splatting on all-in-focus datasets to validate the effectiveness of our model under general input conditions. + +
MethodPSNR↑SSIM↑LPIPS↓
Mip-Splatting [58]27.050.8930.115
Ours27.810.9020.117
+ +known lens parameters beforehand as ground truth, allowing us to calculate the errors $\delta_{\mathcal{A}}$ and $\delta_{\mathcal{F}}$ between the learned parameters and the ground truth. More details can be found in the supplementary material. + +# 5.2. Comparisons + +Quantitative Comparisons. We quantitatively compare DoF-Gaussian against other state-of-the-art methods using a real-world defocus blur dataset [26]. As shown in Table 1, our method outperforms other state-of-the-art baselines. These results suggest that our method achieves superior defocus deblurring performance and produces higher-quality novel view synthesis images. + +Qualitative Comparisons. Visual qualitative comparisons are presented in Fig. 5. Our method surpasses all other methods in terms of image fidelity, generating novel view images that are more faithful to the ground truth images and exhibit less blur. Specifically, the bottle caps and shelves are the sharpest in our rendered images. + +All-in-focus Inputs. To further validate the effectiveness of our method in the all-in-focus setting, we design an experiment using wide DoF images as inputs. We conduct a comparison between our method and Mip-Splatting [58] on two all-in-focus datasets, the Real Forward Facing dataset [28] and the T&T_DB dataset [16]. As shown in Table 3, our method achieves comparable or even better performance to Mip-Splatting on the average metrics across the datasets. This demonstrates that our lens-based model not only excels with shallow DoF inputs but also achieves good + +![](images/afc37f6ecac264bd729b25afea8480afbf9695b641eb6ce660296001e8ecb764.jpg) +Figure 5. Qualitative comparisons against all baselines. Compared to other state-of-the-art methods, our method represents sharper scenes and generates novel view images with less blur. + +![](images/37b868e3648129b9042355f069a7311dd491a7ed5d3093aa4d8986a139ebc8f1.jpg) +Figure 6. Examples in our dataset. For each scene, we apply a DoF rendering method [35] to convert all-in-focus images to shallow DoF images with varying focus locations and aperture sizes. This dataset is used to evaluate the model's refocusing ability. + +results with general input images. + +Our Synthetic Dataset. As shown in Table 2, our method achieves superior refocusing ability through controlled DoF rendering, producing rendered images with the highest quality on our synthetic dataset. Furthermore, our method learns more precise aperture size and focus distance than DoF-NeRF [55]. Since Deblur-NeRF [26], BAGS [34], and Deblurring 3DGS [18] do not incorporate a lens model, they cannot generate novel view images with a specific DoF. As a result, these methods cannot be evaluated on this dataset due to their lack of refocusing capability. + +# 5.3. Ablation Study + +As shown in Table 4 and Fig. 8, we conduct ablation studies to evaluate the effectiveness of our designs. The comparison between #1 and #2 indicates that the lens-based imaging model has a significant impact. This lens model not only facilitates defocus deblurring but also provides the ability to control the depth of field. Next, the defocus-to-focus + +Table 4. Ablation study on each component of our method. + +
LensDepthDefocus-to-focusPSNR↑SSIM↑LPIPS↓
#1XXX21.310.6360.239
#2✓XX23.050.7280.109
#3✓X✓23.590.7420.104
#4✓✓X23.420.7380.098
#5✓✓✓23.970.7560.093
+ +Table 5. Ablation study on different depth priors. + +
MethodPSNR↑SSIM↑LPIPS↓
No fine-tuned depth23.530.7390.114
Sparse depth23.440.7370.113
Ours23.970.7560.093
+ +adaptation (#3) and the per-scene adjustment of depth priors (#4) further enhance the scene geometry and improve the deblurring of details. Ultimately, the combination of all components yields the highest performance gain, leading to a 2.66 dB increase in PSNR over the baseline model. + +In addition, we conduct an ablation study on different depth supervision methods to demonstrate the superiority of our per-scene adjustment of depth priors, as shown in Table 5. Direct supervision using depth maps predicted by the depth network without scene-specific fine-tuning hinders the reconstruction quality of 3D-GS. Likewise, supervision based on sparse depth maps projected from a 3D point cloud fails to yield optimal results. We show visualizations of depth maps rendered by different depth supervision strategies in the supplementary material. + +![](images/c5127cfbbf9def5d270a2dc66c2e97c9df308d7ea4658d695207b9d6918444b3.jpg) +(I) Change the aperture size and focus distance + +![](images/99843575058854ca572730dbd9ce3f9674a04e3fb1f82637856ac1f99ed2c6ad.jpg) +(II) Change the lens parameters and camera poses + +![](images/99c79dbb4aa2cd977795ab94a7f90038405c9f824840d0801d73de04568ba5a4.jpg) +(III) Change the lens parameters and zoom-in +Figure 7. Applications of our controllable DoF effects. Users can create their own cinematic moments by combining changes in aperture size, focus distance, camera poses, and zoom. + +![](images/d8e23bdec491fec232de644336e6ff64568c5b75af21d1565cb6e86fa943f88e.jpg) +Figure 8. The visualizations of the ablation study. Qualitative comparisons show that the full model (#5) yields the best result and the reconstructed strawberry is the sharpest. + +# 5.4. Interactive Applications + +Thanks to the lens-based model, we can not only represent 3D scenes from shallow DoF inputs but also render interactive DoF effects (see Fig. 7). Various cinematography styles can be achieved through different camera configurations, and users can input custom datasets, whether all-in-focus or bokeh images. By controlling aperture size and focus distance, we can render novel view images with different depths of field. In addition, we can modify the shape of the CoC, such as changing it from a circle to a hexagon. Users can also adjust the depth of field while moving the camera or zooming to create cinematic effects. Unlike [55], our GS framework significantly enhances the training and rendering efficiency, allowing for faster and more seamless inter + +actions. The details on processing time will be discussed in the supplementary material. + +# 6. Conclusion + +While 3D Gaussian Splatting methods have achieved impressive performance on a wide range of reconstruction tasks, the applications on controllable DoF effects remain challenging and understudied. In this paper, we propose DoF-Gaussian, a controllable DoF method for 3D Gaussian Splatting. Specifically, we develop a lens-based model rather than pinhole imaging to overcome the limitations imposed by shallow DoF inputs for 3D-GS. Furthermore, we propose the per-scene depth adjustment and a defocus-to-focus adaptation to guarantee the performance of defocus deblurring. We also introduce a synthetic dataset for a more comprehensive evaluation. Thanks to the imaging principles, our method supports various interactive applications. + +Future work. Inspired by our experiments, we found that modeling real-world physical imaging principles can enable our method not only to handle shallow DoF inputs but also to perform effectively on general inputs. This insight motivates us to explore a combination of non-ideal conditions, such as shallow DoF inputs, sparse views, and varied lighting environments, which are more aligned with casual photography in our daily lives. + +Acknowledgement. This study is supported under the RIE2020 Industry Alignment Fund - Industry Collaboration Projects (IAF-ICP) Funding Initiative, as well as cash and in-kind contribution from the industry partner(s). + +# References + +[1] Guillaume Abadie, Steve McAuley, Evegenii Golubev, Stephen Hill, and Sebastien Lagarde. Advances in real-time rendering in games. In ACM SIGGRAPH 2018 Courses. 2018. 2 +[2] Aleksei Bochkovskii, Amaël Delaunoy, Hugo Germain, Marcel Santos, Yichao Zhou, Stephan R Richter, and Vladlen Koltun. Depth pro: Sharp monocular metric depth in less than a second. arXiv preprint arXiv:2410.02073, 2024. 6 +[3] Benjamin Busam, Matthieu Hog, Steven McDonagh, and Gregory Slabaugh. Sterefo: Efficient image refocusing with stereo vision. In ICCVW, 2019. 4 +[4] Wenbo Chen and Ligang Liu. Deblur-GS: 3D Gaussian Splitting from Camera Motion Blurred Images. PACMCGIT, 7(1), 2024. 3 +[5] Jaeyoung Chung, Jeongtaek Oh, and Kyoung Mu Lee. Depth-regularized optimization for 3D Gaussian Splitting in few-shot images. In CVPR, 2024. 5 +[6] Marcos V Conde, Manuel Kolmet, Tim Seizinger, Tom E Bishop, Radu Timofte, Xiangyu Kong, Dafeng Zhang, Jinlong Wu, Fan Wang, Juewen Peng, et al. Lens-to-lens bokeh effect transformation. NTIRE 2023 challenge report. In CVPR, 2023. 2 +[7] François Darmon, Lorenzo Porzi, Samuel Rota-Bulò, and Peter Kontschieder. Robust gaussian splatting. arXiv preprint arXiv:2404.04211, 2024. 2, 3 +[8] Saikat Dutta, Sourya Dipta Das, Nisarg A Shah, and Anil Kumar Tiwari. Stacked deep multi-scale hierarchical network for fast bokeh effect rendering from a single image. In CVPR, 2021. 2 +[9] David Eigen, Christian Puhrsch, and Rob Fergus. Depth map prediction from a single image using a multi-scale deep network. NeurIPS, 27, 2014. 5 +[10] Eugene Hecht. Optics. Pearson Education India, 2012. 4 +[11] Binbin Huang, Zehao Yu, Anpei Chen, Andreas Geiger, and Shenghua Gao. 2D Gaussian Splatting for geometrically accurate radiance fields. In ACM SIGGRAPH 2024 Conference Papers, 2024. 2, 5 +[12] Andrey Ignatov, Jagruti Patel, Radu Timofte, Bolun Zheng, Xin Ye, Li Huang, Xiang Tian, Saikat Dutta, Kuldeep Purohit, Praveen Kandula, et al. Aim 2019 challenge on bokeh effect synthesis: Methods and results. In ICCVW. IEEE, 2019. 2 +[13] Andrey Ignatov, Jagruti Patel, and Radu Timofte. Rendering natural camera bokeh effect with deep learning. In CVPRW, 2020. +[14] Andrey Ignatov, Radu Timofte, Ming Qian, Congyu Qiao, Jiamin Lin, Zhenyu Guo, Chenghua Li, Cong Leng, Jian Cheng, Juewen Peng, et al. AIM 2020 challenge on rendering realistic bokeh. In ECCVW. Springer, 2020. 2 +[15] Xin Jin, Pengyi Jiao, Zheng-Peng Duan, Xingchao Yang, Chun-Le Guo, Bo Ren, and Chongyi Li. Lighting Every Darkness with 3DGS: Fast Training and Real-Time Rendering for HDR View Synthesis. arXiv preprint arXiv:2406.06216, 2024.3 + +[16] Bernhard Kerbl, Georgios Kopanas, Thomas Leimkuhler, and George Drettakis. 3D Gaussian Splatting for Real-Time Radiance Field Rendering. ACM TOG, 42(4), 2023. 2, 6 +[17] Min-Jung Kim, Gyojung Gu, and Jaegul Choo. LensNeRF: Rethinking volume rendering based on thin-lens camera model. In WACV, 2024. 2, 3 +[18] Byeonghyeon Lee, Howoong Lee, Xiangyu Sun, Usman Ali, and Eunbyung Park. Deblurring 3D Gaussian Splatting. arXiv preprint arXiv:2401.00834, 2024. 2, 3, 6, 7 +[19] Dogyoon Lee, Minhyeok Lee, Chajin Shin, and Sangyoun Lee. Dp-NeRF: Deblurred neural radiance field with physical scene priors. In CVPR, 2023. 2, 3, 6 +[20] Sungkil Lee, Elmar Eisemann, and Hans-Peter Seidel. Realtime lens blur effects and focus control. ACM TOG, 29(4), 2010. 2 +[21] Jiahe Li, Jiawei Zhang, Xiao Bai, Jin Zheng, Xin Ning, Jun Zhou, and Lin Gu. DNGaussian: Optimizing sparse-view 3D Gaussian radiance fields with global-local depth normalization. In CVPR, 2024. 5 +[22] Zhengqi Li, Tali Dekel, Forrester Cole, Richard Tucker, Noah Snavely, Ce Liu, and William T Freeman. Learning the depths of moving people by watching frozen people. In CVPR, 2019. 5 +[23] Tianqi Liu, Guangcong Wang, Shoukang Hu, Liao Shen, Xinyi Ye, Yuhang Zang, Zhiguo Cao, Wei Li, and Ziwei Liu. MVSGaussian: Fast Generalizable Gaussian Splatting Reconstruction from Multi-View Stereo. In ECCV. Springer, 2025. 2 +[24] Xianrui Luo, Huiqiang Sun, Juewen Peng, and Zhiguo Cao. Dynamic Neural Radiance Field From Defocused Monocular Video. arXiv preprint arXiv:2407.05586, 2024. 2 +[25] Yawen Luo, Min Shi, Liao Shen, Yachuan Huang, Zixuan Ye, Juewen Peng, and Zhiguo Cao. Video bokeh rendering: Make casual videography cinematic. In ACM MM, 2024. 2 +[26] Li Ma, Xiaoyu Li, Jing Liao, Qi Zhang, Xuan Wang, Jue Wang, and Pedro V Sander. Deblur-NeRF: Neural radiance fields from blurry images. In CVPR, 2022. 2, 3, 6, 7 +[27] Ricardo Martin-Brualla, Noha Radwan, Mehdi SM Sajjadi, Jonathan T Barron, Alexey Dosovitskiy, and Daniel Duckworth. NeRF in the wild: Neural radiance fields for unconstrained photo collections. In CVPR, 2021. 3 +[28] Ben Mildenhall, Pratul P Srinivasan, Matthew Tancik, Jonathan T Barron, Ravi Ramamoorthi, and Ren Ng. NeRF: Representing scenes as neural radiance fields for view synthesis. Communications of the ACM, 65(1), 2021. 2, 6 +[29] Ben Mildenhall, Peter Hedman, Ricardo Martin-Brualla, Pratul P Srinivasan, and Jonathan T Barron. NeRF in the dark: High dynamic range view synthesis from noisy raw images. In CVPR, 2022. 3 +[30] Nicolas Moenne-Loccoz, Ashkan Mirzaei, Or Perel, Ricardo de Lutio, Janick Martinez Esturo, Gavriel State, Sanja Fidler, Nicholas Sharp, and Zan Gojcic. 3d gaussian ray tracing: Fast tracing of particle scenes. ACM Transactions on Graphics (TOG), 43(6):1-19, 2024. 3 +[31] Keunhong Park, Utkarsh Sinha, Jonathan T Barron, Sofien Bouaziz, Dan B Goldman, Steven M Seitz, and Ricardo Martin-Brualla. Nerfies: Deformable neural radiance fields. In ICCV, 2021. 3 + +[32] Frank L Pedrotti, Leno M Pedrotti, and Leno S Pedrotti. Introduction to optics. Cambridge University Press, 2018. 4 +[33] Cheng Peng and Rama Chellappa. PDRF: progressively deblurring radiance field for fast scene reconstruction from blurry images. In AAAI, 2023. 2, 3 +[34] Cheng Peng, Yutao Tang, Yifan Zhou, Nengyu Wang, Xijun Liu, Deming Li, and Rama Chellappa. BAGS: Blur Agnostic Gaussian Splatting through Multi-Scale Kernel Modeling. arXiv preprint arXiv:2403.04926, 2024. 2, 3, 5, 6, 7 +[35] Juewen Peng, Zhiguo Cao, Xianrui Luo, Hao Lu, Ke Xian, and Jianming Zhang. Bokehme: When neural rendering meets classical rendering. In CVPR, 2022. 2, 6, 7 +[36] Juewen Peng, Zhiyu Pan, Chengxin Liu, Xianrui Luo, Huiqiang Sun, Liao Shen, Ke Xian, and Zhiguo Cao. Selective bokeh effect transformation. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 1456-1464, 2023. +[37] Dominique Piché-Meunier, Yannick Hold-Geoffroy, Jianming Zhang, and Jean-François Lalonde. Lens parameter estimation for realistic depth of field modeling. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 499–508, 2023. 2 +[38] Ming Qian, Congyu Qiao, Jiamin Lin, Zhenyu Guo, Chenghua Li, Cong Leng, and Jian Cheng. BgGAN: Bokehglass generative adversarial network for rendering realistic bokeh. In ECCVW. Springer, 2020. 2 +[39] René Ranftl, Katrin Lasinger, David Hafner, Konrad Schindler, and Vladlen Koltun. Towards robust monocular depth estimation: Mixing datasets for zero-shot cross-dataset transfer. IEEE TPAMI, 44(3), 2020. 5 +[40] Johannes L Schonberger and Jan-Michael Frahm. Structure-from-motion revisited. In CVPR, 2016. 5 +[41] Johannes L. Schonberger, Enliang Zheng, Jan-Michael Frahm, and Marc Pollefeys. Pixelwise view selection for unstructured multi-view stereo. In ECCV. Springer, 2016. 5 +[42] Liao Shen, Xingyi Li, Huiqiang Sun, Juewen Peng, Ke Xian, Zhiguo Cao, and Guosheng Lin. Make-it-4d: Synthesizing a consistent long-term dynamic scene video from a single image. In Proceedings of the 31st ACM International Conference on Multimedia, pages 8167–8175, 2023. 3 +[43] Yichen Sheng, Zixun Yu, Lu Ling, Zhiwen Cao, Xuaner Zhang, Xin Lu, Ke Xian, Haiting Lin, and Bedrich Benes. Dr. Bokeh: DiffeRentiable Occlusion-aware Bokeh Rendering. In CVPR, 2024. 2 +[44] Huiqiang Sun, Xingyi Li, Liao Shen, Xinyi Ye, Ke Xian, and Zhiguo Cao. DyBluRF: Dynamic Neural Radiance Fields from Blurry Monocular Video. In CVPR, 2024. 2, 3 +[45] Neal Wadhwa, Rahul Garg, David E Jacobs, Bryan E Feldman, Nori Kanazawa, Robert Carroll, Yair Movshovitz-Attias, Jonathan T Barron, Yael Pritch, and Marc Levoy. Synthetic depth-of-field with a single-camera mobile phone. ACM TOG, 37(4), 2018. 2 +[46] Chao Wang, Krzysztof Wolski, Xingang Pan, Thomas Leimkuhler, Bin Chen, Christian Theobalt, Karol Myszkowski, Hans-Peter Seidel, and Ana Serrano. An implicit neural representation for the image stack: Depth, all in focus, and high dynamic range. Technical report, 2023. 2 + +[47] Chao Wang, Krzysztof Wolski, Bernhard Kerbl, Ana Serrano, Mojtaba Bemana, Hans-Peter Seidel, Karol Myszkowski, and Thomas Leimkuhler. Cinematic Gaussians: Real-Time HDR Radiance Fields with Depth of Field. arXiv preprint arXiv:2406.07329, 2024. 3 +[48] Lijun Wang, Xiaohui Shen, Jianming Zhang, Oliver Wang, Zhe Lin, Chih-Yao Hsieh, Sarah Kong, and Huchuan Lu. Deep learning: Shallow depth of field from a single image. arXiv preprint arXiv:1810.08100, 2018. 2 +[49] Peng Wang, Lingzhe Zhao, Ruijie Ma, and Peidong Liu. BAD-NeRF: Bundle adjusted deblur neural radiance fields. In CVPR, 2023. 3 +[50] Yinhuai Wang, Shuzhou Yang, Yujie Hu, and Jian Zhang. NeRFocus: Neural radiance field for 3D synthetic defocus. arXiv preprint arXiv:2203.05189, 2022. 3 +[51] Yujie Wang, Praneeth Chakravarthula, and Baoquan Chen. DOF-GS: Adjustable Depth-of-Field 3D Gaussian Splitting for Refocusing, Defocus Rendering and Blur Removal. arXiv preprint arXiv:2405.17351, 2024. 3, 6 +[52] Zhou Wang, Alan C Bovik, Hamid R Sheikh, and Eero P Simoncelli. Image quality assessment: from error visibility to structural similarity. IEEE TIP, 13(4), 2004. 6 +[53] Yi Wei, Shaohui Liu, Yongming Rao, Wang Zhao, Jiwen Lu, and Jie Zhou. Nerfingmvs: Guided optimization of neural radiance fields for indoor multi-view stereo. In ICCV, 2021. 5 +[54] Guanjun Wu, Taoran Yi, Jiemin Fang, Lingxi Xie, Xiaopeng Zhang, Wei Wei, Wenyu Liu, Qi Tian, and Xinggang Wang. 4d gaussian splatting for real-time dynamic scene rendering. In CVPR, 2024. 2, 3 +[55] Zijin Wu, Xingyi Li, Juewen Peng, Hao Lu, Zhiguo Cao, and Weicai Zhong. DoF-NeRF: Depth-of-field meets neural radiance fields. In ACM MM, 2022. 1, 2, 3, 6, 7, 8 +[56] Lei Xiao, Anton Kaplanyan, Alexander Fix, Matt Chapman, and Douglas Lanman. Deepfocus: Learned image synthesis for computational display. In ACM SIGGRAPH 2018 Talks. 2018. 2 +[57] Haofei Xu, Songyou Peng, Fangjinhua Wang, Hermann Blum, Daniel Barath, Andreas Geiger, and Marc Pollefeys. Depthsplat: Connecting gaussian splatting and depth. arXiv preprint arXiv:2410.13862, 2024. 2 +[58] Zehao Yu, Anpei Chen, Binbin Huang, Torsten Sattler, and Andreas Geiger. Mip-splatting: Alias-free 3D gaussian splattering. In CVPR, 2024. 2, 5, 6 +[59] Baowen Zhang, Chuan Fang, Rakesh Shrestha, Yixun Liang, Xiaoxiao Long, and Ping Tan. RaDe-GS: Rasterizing Depth in Gaussian Splatting. arXiv preprint arXiv:2406.01467, 2024.5 +[60] Richard Zhang, Phillip Isola, Alexei A Efros, Eli Shechtman, and Oliver Wang. The unreasonable effectiveness of deep features as a perceptual metric. In CVPR, 2018. 6 +[61] Xuaner Zhang, Kevin Matzen, Vivien Nguyen, Dillon Yao, You Zhang, and Ren Ng. Synthetic defocus and lookahead autofocus for casual videography. arXiv preprint arXiv:1905.06326, 2019. 2 +[62] Lingzhe Zhao, Peng Wang, and Peidong Liu. BAD-Gaussians: Bundle adjusted deblur Gaussian Splatting. arXiv preprint arXiv:2403.11831, 2024. 3 \ No newline at end of file diff --git a/2025/DoF-Gaussian_ Controllable Depth-of-Field for 3D Gaussian Splatting/images.zip b/2025/DoF-Gaussian_ Controllable Depth-of-Field for 3D Gaussian Splatting/images.zip new file mode 100644 index 0000000000000000000000000000000000000000..ae90e8fd97328e478a561dc02c49468af539da47 --- /dev/null +++ b/2025/DoF-Gaussian_ Controllable Depth-of-Field for 3D Gaussian Splatting/images.zip @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:3744b3f4ce4a7225a0e62c9c04c456651c8dc1c2386161b3a1083b37bf2e1881 +size 885832 diff --git a/2025/DoF-Gaussian_ Controllable Depth-of-Field for 3D Gaussian Splatting/layout.json b/2025/DoF-Gaussian_ Controllable Depth-of-Field for 3D Gaussian Splatting/layout.json new file mode 100644 index 0000000000000000000000000000000000000000..a32893c0860d7d8a1078b36b6908fe71ee152aab --- /dev/null +++ b/2025/DoF-Gaussian_ Controllable Depth-of-Field for 3D Gaussian Splatting/layout.json @@ -0,0 +1,9886 @@ +{ + "pdf_info": [ + { + "para_blocks": [ + { + "bbox": [ + 88, + 103, + 522, + 121 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 88, + 103, + 522, + 121 + ], + "spans": [ + { + "bbox": [ + 88, + 103, + 522, + 121 + ], + "type": "text", + "content": "DoF-Gaussian: Controllable Depth-of-Field for 3D Gaussian Splatting" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 61, + 142, + 560, + 186 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 61, + 142, + 560, + 186 + ], + "spans": [ + { + "bbox": [ + 61, + 142, + 560, + 186 + ], + "type": "text", + "content": "Liao Shen" + }, + { + "bbox": [ + 61, + 142, + 560, + 186 + ], + "type": "inline_equation", + "content": "^{1,2}" + }, + { + "bbox": [ + 61, + 142, + 560, + 186 + ], + "type": "text", + "content": " Tianqi Liu" + }, + { + "bbox": [ + 61, + 142, + 560, + 186 + ], + "type": "inline_equation", + "content": "^{1,2}" + }, + { + "bbox": [ + 61, + 142, + 560, + 186 + ], + "type": "text", + "content": " Huiqiang Sun" + }, + { + "bbox": [ + 61, + 142, + 560, + 186 + ], + "type": "inline_equation", + "content": "^{1,2}" + }, + { + "bbox": [ + 61, + 142, + 560, + 186 + ], + "type": "text", + "content": " Jiaqi Li" + }, + { + "bbox": [ + 61, + 142, + 560, + 186 + ], + "type": "inline_equation", + "content": "^{1}" + }, + { + "bbox": [ + 61, + 142, + 560, + 186 + ], + "type": "text", + "content": " Zhiguo Cao" + }, + { + "bbox": [ + 61, + 142, + 560, + 186 + ], + "type": "inline_equation", + "content": "^{1}" + }, + { + "bbox": [ + 61, + 142, + 560, + 186 + ], + "type": "text", + "content": " Wei Li" + }, + { + "bbox": [ + 61, + 142, + 560, + 186 + ], + "type": "inline_equation", + "content": "^{2*}" + }, + { + "bbox": [ + 61, + 142, + 560, + 186 + ], + "type": "text", + "content": " Chen Change Loy" + }, + { + "bbox": [ + 61, + 142, + 560, + 186 + ], + "type": "inline_equation", + "content": "^{2}" + }, + { + "bbox": [ + 61, + 142, + 560, + 186 + ], + "type": "inline_equation", + "content": "^{1}" + }, + { + "bbox": [ + 61, + 142, + 560, + 186 + ], + "type": "text", + "content": "School of AIA, Huazhong University of Science and Technology \n" + }, + { + "bbox": [ + 61, + 142, + 560, + 186 + ], + "type": "inline_equation", + "content": "^{2}" + }, + { + "bbox": [ + 61, + 142, + 560, + 186 + ], + "type": "text", + "content": "S-Lab, Nanyang Technological University" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 258, + 188, + 364, + 198 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 258, + 188, + 364, + 198 + ], + "spans": [ + { + "bbox": [ + 258, + 188, + 364, + 198 + ], + "type": "text", + "content": "leoshen@hust.edu.cn" + } + ] + } + ], + "index": 4 + }, + { + "type": "image", + "bbox": [ + 56, + 209, + 268, + 349 + ], + "blocks": [ + { + "bbox": [ + 56, + 209, + 268, + 349 + ], + "lines": [ + { + "bbox": [ + 56, + 209, + 268, + 349 + ], + "spans": [ + { + "bbox": [ + 56, + 209, + 268, + 349 + ], + "type": "image", + "image_path": "96b19129ee7253db8f2beb70a89972b0937c4e8dc4a369d8af6ad4fdd2897bb3.jpg" + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "image_body" + } + ], + "index": 5 + }, + { + "type": "image", + "bbox": [ + 58, + 350, + 127, + 397 + ], + "blocks": [ + { + "bbox": [ + 58, + 350, + 127, + 397 + ], + "lines": [ + { + "bbox": [ + 58, + 350, + 127, + 397 + ], + "spans": [ + { + "bbox": [ + 58, + 350, + 127, + 397 + ], + "type": "image", + "image_path": "ef42b940e1b3a757249e06aeb98210b90570586c95ef8ebd901c80abcc82ae68.jpg" + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 70, + 398, + 113, + 407 + ], + "lines": [ + { + "bbox": [ + 70, + 398, + 113, + 407 + ], + "spans": [ + { + "bbox": [ + 70, + 398, + 113, + 407 + ], + "type": "text", + "content": "DoF-NeRF" + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "image_caption" + } + ], + "index": 6 + }, + { + "type": "image", + "bbox": [ + 129, + 350, + 197, + 397 + ], + "blocks": [ + { + "bbox": [ + 129, + 350, + 197, + 397 + ], + "lines": [ + { + "bbox": [ + 129, + 350, + 197, + 397 + ], + "spans": [ + { + "bbox": [ + 129, + 350, + 197, + 397 + ], + "type": "image", + "image_path": "d8a776f82313fb89e47962f106e53fd310f39b1a5a09d02c0e52deee350decf2.jpg" + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 151, + 398, + 177, + 407 + ], + "lines": [ + { + "bbox": [ + 151, + 398, + 177, + 407 + ], + "spans": [ + { + "bbox": [ + 151, + 398, + 177, + 407 + ], + "type": "text", + "content": "BAGS" + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "image_caption" + } + ], + "index": 8 + }, + { + "type": "image", + "bbox": [ + 198, + 350, + 268, + 397 + ], + "blocks": [ + { + "bbox": [ + 198, + 350, + 268, + 397 + ], + "lines": [ + { + "bbox": [ + 198, + 350, + 268, + 397 + ], + "spans": [ + { + "bbox": [ + 198, + 350, + 268, + 397 + ], + "type": "image", + "image_path": "4b3768a91c66cecb078a2c7eff329ba6b4116e9a9db37c239b2935f8100ee7c1.jpg" + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 220, + 398, + 242, + 407 + ], + "lines": [ + { + "bbox": [ + 220, + 398, + 242, + 407 + ], + "spans": [ + { + "bbox": [ + 220, + 398, + 242, + 407 + ], + "type": "text", + "content": "Ours" + } + ] + } + ], + "index": 11, + "angle": 0, + "type": "image_caption" + } + ], + "index": 10 + }, + { + "type": "image", + "bbox": [ + 279, + 209, + 416, + 297 + ], + "blocks": [ + { + "bbox": [ + 279, + 209, + 416, + 297 + ], + "lines": [ + { + "bbox": [ + 279, + 209, + 416, + 297 + ], + "spans": [ + { + "bbox": [ + 279, + 209, + 416, + 297 + ], + "type": "image", + "image_path": "e3b2ffe15ba989b5ac1a70e607113ff8e23daa3a4026a532b16f75a665645513.jpg" + } + ] + } + ], + "index": 12, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 334, + 297, + 484, + 307 + ], + "lines": [ + { + "bbox": [ + 334, + 297, + 484, + 307 + ], + "spans": [ + { + "bbox": [ + 334, + 297, + 484, + 307 + ], + "type": "text", + "content": "Change aperture size and focus distance" + } + ] + } + ], + "index": 13, + "angle": 0, + "type": "image_caption" + } + ], + "index": 12 + }, + { + "type": "image", + "bbox": [ + 417, + 209, + 553, + 297 + ], + "blocks": [ + { + "bbox": [ + 417, + 209, + 553, + 297 + ], + "lines": [ + { + "bbox": [ + 417, + 209, + 553, + 297 + ], + "spans": [ + { + "bbox": [ + 417, + 209, + 553, + 297 + ], + "type": "image", + "image_path": "becc542fe2a2639b9769c7749c819fe5ab65a7cd6ac1e9808f67cc32958b692a.jpg" + } + ] + } + ], + "index": 14, + "angle": 0, + "type": "image_body" + } + ], + "index": 14 + }, + { + "type": "image", + "bbox": [ + 279, + 308, + 416, + 396 + ], + "blocks": [ + { + "bbox": [ + 279, + 308, + 416, + 396 + ], + "lines": [ + { + "bbox": [ + 279, + 308, + 416, + 396 + ], + "spans": [ + { + "bbox": [ + 279, + 308, + 416, + 396 + ], + "type": "image", + "image_path": "e599fd807f897a202c19e0a47d1a2ef608ad32fe2903fbe3d3bc71e6c99f9b3f.jpg" + } + ] + } + ], + "index": 15, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 282, + 397, + 412, + 407 + ], + "lines": [ + { + "bbox": [ + 282, + 397, + 412, + 407 + ], + "spans": [ + { + "bbox": [ + 282, + 397, + 412, + 407 + ], + "type": "text", + "content": "Change camera poses and zoom-in" + } + ] + } + ], + "index": 16, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 55, + 412, + 555, + 446 + ], + "lines": [ + { + "bbox": [ + 55, + 412, + 555, + 446 + ], + "spans": [ + { + "bbox": [ + 55, + 412, + 555, + 446 + ], + "type": "text", + "content": "Figure 1. Given a set of multi-view input images with shallow DoF, DoF-Gaussian can reconstruct a 3D-GS representation of a sharp scene. Thanks to our lens-based design, we can also achieve controllable DoF effects for a variety of applications. The example input images are taken from [55] for illustration purposes. (Zoom-in for best view)" + } + ] + } + ], + "index": 19, + "angle": 0, + "type": "image_caption" + } + ], + "index": 15 + }, + { + "type": "image", + "bbox": [ + 417, + 308, + 508, + 396 + ], + "blocks": [ + { + "bbox": [ + 417, + 308, + 508, + 396 + ], + "lines": [ + { + "bbox": [ + 417, + 308, + 508, + 396 + ], + "spans": [ + { + "bbox": [ + 417, + 308, + 508, + 396 + ], + "type": "image", + "image_path": "14bd89e418f7a6774affdb731d0a4895c49b187f124105341b3b097982a2a1d6.jpg" + } + ] + } + ], + "index": 17, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 442, + 398, + 526, + 408 + ], + "lines": [ + { + "bbox": [ + 442, + 398, + 526, + 408 + ], + "spans": [ + { + "bbox": [ + 442, + 398, + 526, + 408 + ], + "type": "text", + "content": "Change aperture shape" + } + ] + } + ], + "index": 18, + "angle": 0, + "type": "image_caption" + } + ], + "index": 17 + }, + { + "bbox": [ + 151, + 456, + 200, + 468 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 151, + 456, + 200, + 468 + ], + "spans": [ + { + "bbox": [ + 151, + 456, + 200, + 468 + ], + "type": "text", + "content": "Abstract" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 54, + 480, + 297, + 696 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 54, + 480, + 297, + 696 + ], + "spans": [ + { + "bbox": [ + 54, + 480, + 297, + 696 + ], + "type": "text", + "content": "Recent advances in 3D Gaussian Splatting (3D-GS) have shown remarkable success in representing 3D scenes and generating high-quality, novel views in real-time. However, 3D-GS and its variants assume that input images are captured based on pinhole imaging and are fully in focus. This assumption limits their applicability, as real-world images often feature shallow depth-of-field (DoF). In this paper, we introduce DoF-Gaussian, a controllable depth-of-field method for 3D-GS. We develop a lens-based imaging model based on geometric optics principles to control DoF effects. To ensure accurate scene geometry, we incorporate depth priors adjusted per scene, and we apply defocus-to-focus adaptation to minimize the gap in the circle of confusion. We also introduce a synthetic dataset to assess refocusing capabilities and the model's ability to learn precise lens parameters. Our framework is customizable and supports various interactive applications. Extensive experiments confirm the effectiveness of our method. Our project is avail-" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 314, + 457, + 470, + 469 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 314, + 457, + 470, + 469 + ], + "spans": [ + { + "bbox": [ + 314, + 457, + 470, + 469 + ], + "type": "text", + "content": "able at https://dof-gaussian.github.io/." + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 314, + 490, + 394, + 502 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 314, + 490, + 394, + 502 + ], + "spans": [ + { + "bbox": [ + 314, + 490, + 394, + 502 + ], + "type": "text", + "content": "1. Introduction" + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 312, + 510, + 555, + 689 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 312, + 510, + 555, + 689 + ], + "spans": [ + { + "bbox": [ + 312, + 510, + 555, + 689 + ], + "type": "text", + "content": "Depth-of-field (DoF) refers to the distance between the closest and farthest objects in a photo that appears acceptably sharp. In practice, photographers can control the DoF by adjusting the camera's aperture size or focus distance to capture images with either a wide DoF (all-in-focus) or shallow DoF (defocused). The shallow DoF effect is an important technique in photography, as it draws the viewers' attention to the focal region by blurring the surrounding areas. Meanwhile, novel view synthesis aims to create realistic images from novel viewpoints based on a set of source images. However, novel view synthesis typically requires all-in-focus input images and lacks the capability to render varied DoF effects, limiting its applications. In this work, we aim to render novel views with controllable DoF, adding cinematic quality to the results." + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 313, + 689, + 555, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 689, + 555, + 714 + ], + "spans": [ + { + "bbox": [ + 313, + 689, + 555, + 714 + ], + "type": "text", + "content": "Images captured from the real world sometimes have shallow DoF. Specifically, points of light that do not lie on" + } + ] + } + ], + "index": 25 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "spans": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "type": "text", + "content": "CVF" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 145, + 0, + 496, + 35 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 145, + 0, + 496, + 35 + ], + "spans": [ + { + "bbox": [ + 145, + 0, + 496, + 35 + ], + "type": "text", + "content": "This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 67, + 703, + 144, + 712 + ], + "type": "page_footnote", + "angle": 0, + "lines": [ + { + "bbox": [ + 67, + 703, + 144, + 712 + ], + "spans": [ + { + "bbox": [ + 67, + 703, + 144, + 712 + ], + "type": "text", + "content": "*Corresponding author." + } + ] + } + ], + "index": 26 + }, + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "text", + "content": "26462" + } + ] + } + ], + "index": 27 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 0 + }, + { + "para_blocks": [ + { + "bbox": [ + 55, + 72, + 294, + 262 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 72, + 294, + 262 + ], + "spans": [ + { + "bbox": [ + 55, + 72, + 294, + 262 + ], + "type": "text", + "content": "the focal plane are projected onto the sensor plane as blurred circles, referred to as the circle of confusion (CoC), and this introduces bokeh blur to the captured image. Most novel view synthesis methods experience performance degradation when processing shallow DoF input images. To address this issue, DoF-NeRF [55] and LensNeRF [17] introduce lens-based camera models in volume rendering to explicitly enable controllable DoF effects. Meanwhile, methods like Deblur-NeRF [26], DP-NeRF [19], and PDRF [33] employ sparse kernels to model defocus blur. However, implicit rendering approaches face significant challenges in training and rendering efficiency. In the realm of 3D Gaussian Splatting (3D-GS) [16], methods such as BAGS [34] and Deblurring 3DGS [18] propose blur estimation networks to model the blur kernels or scaling factors for defocus deblurring, but they cannot accommodate controllable DoF." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 56, + 266, + 295, + 480 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 266, + 295, + 480 + ], + "spans": [ + { + "bbox": [ + 56, + 266, + 295, + 480 + ], + "type": "text", + "content": "In this paper, we present DoF-Gaussian, an efficient framework for controllable DoF in 3D-GS, addressing the limitations of existing methods in handling shallow DoF inputs. This problem is non-trivial because 3D-GS and its variants [11, 23, 54, 57, 58] are typically based on a pin-hole camera model that assumes all-in-focus inputs, meaning both foreground and background appear clear. When input images contain bokeh blur, it becomes challenging to accurately construct scene geometry and render depth maps. However, rendering images with different DoFs relies heavily on precise depth estimation. Additionally, our 3D-GS approach assumes an idealized CoC, which inevitably differs from the CoC seen in real photographs, posing further challenges for accurate defocus deblurring. Furthermore, existing datasets in this field are primarily borrowed from deblurring applications, limiting our evaluation to the model's defocus deblurring capability and overlooking other aspects of controllable DoF effects." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 55, + 484, + 295, + 662 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 484, + 295, + 662 + ], + "spans": [ + { + "bbox": [ + 55, + 484, + 295, + 662 + ], + "type": "text", + "content": "To address these challenges, we present a new controllable DoF 3D-GS method with the following contributions: First, we employ a lens-based model instead of a pinhole-based one. We make the lens optical parameters, such as aperture size and focus distance, learnable to enable control over the depth of field. Second, we introduce per-scene adjustments for depth priors to minimize errors in scene geometry and depth relationships, ensuring accurate depth-of-field control. Third, we propose a defocus-to-focus adaptation strategy that focuses on learning the focal region after the learnable lens parameters have converged, compensating for differences in CoC. Finally, we present a synthetic dataset designed to comprehensively evaluate the model, including its refocusing capabilities and its ability to learn accurate aperture size and focus distance." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 55, + 665, + 295, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 665, + 295, + 714 + ], + "spans": [ + { + "bbox": [ + 55, + 665, + 295, + 714 + ], + "type": "text", + "content": "Benefiting from the lens-based imaging model, our approach enables controlled depth of field, allowing it to handle not only shallow DoF image inputs but also unlock a range of interactive and engaging applications, as depicted" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 313, + 72, + 553, + 168 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 72, + 553, + 168 + ], + "spans": [ + { + "bbox": [ + 313, + 72, + 553, + 168 + ], + "type": "text", + "content": "in Figure 1. For instance, users can render novel view images with varied DoF, refocus on custom datasets, and even adjust the shape of CoC, e.g., from circular to pentagonal or hexagonal CoC. Moreover, users can dynamically change the DoF while moving the camera or zooming, rendering videos with cinematic effects. Extensive experiments demonstrate that our method outperforms state-of-the-art depth-of-field and defocus deblurring methods." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 313, + 178, + 400, + 190 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 178, + 400, + 190 + ], + "spans": [ + { + "bbox": [ + 313, + 178, + 400, + 190 + ], + "type": "text", + "content": "2. Related Work" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 313, + 199, + 553, + 449 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 199, + 553, + 449 + ], + "spans": [ + { + "bbox": [ + 313, + 199, + 553, + 449 + ], + "type": "text", + "content": "Depth-of-Field Rendering. Rendering DoF effects from a single all-in-focus image has been widely explored in previous work. Bokeh blur occurs when light is projected onto the image plane as a circular region, rather than as a point. The size of CoC is affected by the diameter of the aperture (aperture size) and the distance from the camera to the focal plane (focus distance). Photos captured with a small aperture usually present a wide DoF, i.e., all objects appear sharp. In contrast, as the aperture diameter increases, objects near the focal plane remain sharp, while those farther away become blurred with a larger CoC. Physically based methods [1, 20] rely on 3D scene geometry information and are time-consuming. Some methods use neural networks [8, 12–14, 38] trained end-to-end to obtain shallow DoF images. The DoF effects in some studies [6, 25, 35–37, 43, 45, 48, 56, 61] are controllable but usually require an extra disparity map. Wang et al. [46] combines neural fields with an expressive camera model to achieve all-in-focus reconstruction from an image stack. In this work, we show the connection between DoF rendering and novel view synthesis using a lens-based camera model." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 313, + 450, + 553, + 652 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 450, + 553, + 652 + ], + "spans": [ + { + "bbox": [ + 313, + 450, + 553, + 652 + ], + "type": "text", + "content": "Image Deblurring. Blur can generally be categorized into two main types: camera motion blur and defocus blur. Previous studies [7, 18, 24, 26, 33, 34, 44] attempt to address this issue. Our task of rendering all-in-focus and sharp novel view images from shallow DoF inputs can be considered as a form of defocus deblurring. By modeling the underlying physical principles, we can achieve defocus deblurring by simulating the circle-of-confusion. Furthermore, our method allows for adjustments to lens parameters, enabling applications such as refocusing—capabilities that deblurring approaches like BAGS [34] cannot achieve. Existing datasets in the 3D DoF field are often borrowed from the deblurring domain [26] or designed solely to assess defocus deblurring ability [55]. To support a wider range of applications and validate the accuracy of the learned lens model, we introduce a synthetic dataset for more comprehensive evaluations." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 313, + 653, + 553, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 653, + 553, + 713 + ], + "spans": [ + { + "bbox": [ + 313, + 653, + 553, + 713 + ], + "type": "text", + "content": "Novel View Synthesis. Novel view synthesis allows the rendering of unseen camera perspectives from 2D images and their corresponding camera poses. Recent advancements in synthesis can largely be attributed to Neural Radiance Field [28] and 3D Gaussian Splatting [16]. However," + } + ] + } + ], + "index": 8 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "text", + "content": "26463" + } + ] + } + ], + "index": 9 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 1 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 58, + 68, + 556, + 291 + ], + "blocks": [ + { + "bbox": [ + 58, + 68, + 556, + 291 + ], + "lines": [ + { + "bbox": [ + 58, + 68, + 556, + 291 + ], + "spans": [ + { + "bbox": [ + 58, + 68, + 556, + 291 + ], + "type": "image", + "image_path": "5cc941806fcebb8e511318e455f0cce1f991551e558ae064a5f20dc3492106b5.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 55, + 299, + 555, + 354 + ], + "lines": [ + { + "bbox": [ + 55, + 299, + 555, + 354 + ], + "spans": [ + { + "bbox": [ + 55, + 299, + 555, + 354 + ], + "type": "text", + "content": "Figure 2. Overview of DoF-Gaussian. Given input images " + }, + { + "bbox": [ + 55, + 299, + 555, + 354 + ], + "type": "inline_equation", + "content": "I" + }, + { + "bbox": [ + 55, + 299, + 555, + 354 + ], + "type": "text", + "content": " with shallow DoF, we first apply SfM from COLMAP to obtain sparse depth " + }, + { + "bbox": [ + 55, + 299, + 555, + 354 + ], + "type": "inline_equation", + "content": "D_{sparse}" + }, + { + "bbox": [ + 55, + 299, + 555, + 354 + ], + "type": "text", + "content": ", which is used to train a depth network to derive per-scene depth priors " + }, + { + "bbox": [ + 55, + 299, + 555, + 354 + ], + "type": "inline_equation", + "content": "D_{pred}" + }, + { + "bbox": [ + 55, + 299, + 555, + 354 + ], + "type": "text", + "content": ". We then employ " + }, + { + "bbox": [ + 55, + 299, + 555, + 354 + ], + "type": "inline_equation", + "content": "D_{pred}" + }, + { + "bbox": [ + 55, + 299, + 555, + 354 + ], + "type": "text", + "content": " to regularize the Gaussians rendered depth map " + }, + { + "bbox": [ + 55, + 299, + 555, + 354 + ], + "type": "inline_equation", + "content": "D" + }, + { + "bbox": [ + 55, + 299, + 555, + 354 + ], + "type": "text", + "content": ". Next, by developing a lens imaging model, we can render defocused images " + }, + { + "bbox": [ + 55, + 299, + 555, + 354 + ], + "type": "inline_equation", + "content": "C^*" + }, + { + "bbox": [ + 55, + 299, + 555, + 354 + ], + "type": "text", + "content": " to simulate input images. To minimize the discrepancy in CoC, we propose an adaptation using the weight map. Finally, we can render fully clear images for novel view synthesis and achieve various effects by our controllable DoF framework." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "bbox": [ + 56, + 370, + 296, + 706 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 370, + 296, + 706 + ], + "spans": [ + { + "bbox": [ + 56, + 370, + 296, + 706 + ], + "type": "text", + "content": "real-world images are often not clean and well-calibrated. Previous work [4, 27, 29, 31, 42, 44, 49, 54, 62] attempts to cope with non-ideal conditions, such as low light, camera motion, bokeh effects, and various types of image degradation. To recover a sharp scene from images with bokeh blur, methods like Deblur-NeRF [26], DP-NeRF [19], and PDRF [33] model defocus blur using sparse kernels. NeRF Focus [50] and LensNeRF [17] achieve defocus effects by incorporating a lens-based camera model in volume rendering. While DoF-NeRF [55] is similar to our approach, it is limited by the training and rendering efficiency and quality of NeRF. BAGS [34] and Deblurring 3DGS [18] use blur estimation networks to model blur kernels or scaling factors but cannot achieve controllable DoF effects due to the absence of a lens imaging model. Other studies, such as [15, 47], focus on HDR tasks and consider depth-of-field simultaneously. RGS [7] addresses defocus blur by introducing an offset to the 2D covariance matrices of Gaussians. [30] introduces a ray tracing rendering algorithm for particle-based representations to enable many advanced techniques, such as the shallow depth-of-field effect. A concurrent work [51] also explores adjustable depth of field. Our method differs in its controllable DoF mechanism, with the introduction of depth priors to ensure accurate scene depth. In addition, our new defocus-to-focus adaptation strategy eliminates the need for the focus localization network proposed in their work. Besides, we introduce a synthetic dataset for comprehensive evaluation." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 314, + 369, + 391, + 383 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 314, + 369, + 391, + 383 + ], + "spans": [ + { + "bbox": [ + 314, + 369, + 391, + 383 + ], + "type": "text", + "content": "3. Preliminary" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 313, + 389, + 555, + 426 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 389, + 555, + 426 + ], + "spans": [ + { + "bbox": [ + 313, + 389, + 555, + 426 + ], + "type": "text", + "content": "3D Gaussian Splitting represents a 3D scene as a mixture of anisotropic 3D Gaussians, where each Gaussian is characterized by a 3D covariance matrix " + }, + { + "bbox": [ + 313, + 389, + 555, + 426 + ], + "type": "inline_equation", + "content": "\\Sigma" + }, + { + "bbox": [ + 313, + 389, + 555, + 426 + ], + "type": "text", + "content": " and mean " + }, + { + "bbox": [ + 313, + 389, + 555, + 426 + ], + "type": "inline_equation", + "content": "\\mu" + }, + { + "bbox": [ + 313, + 389, + 555, + 426 + ], + "type": "text", + "content": ":" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 370, + 430, + 555, + 447 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 370, + 430, + 555, + 447 + ], + "spans": [ + { + "bbox": [ + 370, + 430, + 555, + 447 + ], + "type": "interline_equation", + "content": "G (X) = e ^ {- \\frac {1}{2} (X - \\mu) ^ {\\top} \\Sigma^ {- 1} (X - \\mu)} \\tag {1}", + "image_path": "a4fd43f09e62eaa3758222588c84791b2da18b0ac11508c7b7923c01d59f2823.jpg" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 313, + 451, + 555, + 498 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 451, + 555, + 498 + ], + "spans": [ + { + "bbox": [ + 313, + 451, + 555, + 498 + ], + "type": "text", + "content": "The covariance matrix " + }, + { + "bbox": [ + 313, + 451, + 555, + 498 + ], + "type": "inline_equation", + "content": "\\Sigma" + }, + { + "bbox": [ + 313, + 451, + 555, + 498 + ], + "type": "text", + "content": " holds physical meaning only when it is positive semi-definite. Therefore, to enable effective optimization via gradient descent, " + }, + { + "bbox": [ + 313, + 451, + 555, + 498 + ], + "type": "inline_equation", + "content": "\\Sigma" + }, + { + "bbox": [ + 313, + 451, + 555, + 498 + ], + "type": "text", + "content": " is decomposed into a scaling matrix " + }, + { + "bbox": [ + 313, + 451, + 555, + 498 + ], + "type": "inline_equation", + "content": "S" + }, + { + "bbox": [ + 313, + 451, + 555, + 498 + ], + "type": "text", + "content": " and a rotation matrix " + }, + { + "bbox": [ + 313, + 451, + 555, + 498 + ], + "type": "inline_equation", + "content": "R" + }, + { + "bbox": [ + 313, + 451, + 555, + 498 + ], + "type": "text", + "content": ", as:" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 400, + 503, + 553, + 517 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 400, + 503, + 553, + 517 + ], + "spans": [ + { + "bbox": [ + 400, + 503, + 553, + 517 + ], + "type": "interline_equation", + "content": "\\Sigma = R S S ^ {\\mathsf {T}} R ^ {\\mathsf {T}}. \\tag {2}", + "image_path": "0b1cb04d793c6ade131830c60d3cd23c4d791493aa56c0e6b2096d69513db3d1.jpg" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 313, + 522, + 555, + 582 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 522, + 555, + 582 + ], + "spans": [ + { + "bbox": [ + 313, + 522, + 555, + 582 + ], + "type": "text", + "content": "To splot Gaussians from 3D space to a 2D plane, the view transformation matrix " + }, + { + "bbox": [ + 313, + 522, + 555, + 582 + ], + "type": "inline_equation", + "content": "W" + }, + { + "bbox": [ + 313, + 522, + 555, + 582 + ], + "type": "text", + "content": " and the Jacobian matrix " + }, + { + "bbox": [ + 313, + 522, + 555, + 582 + ], + "type": "inline_equation", + "content": "J" + }, + { + "bbox": [ + 313, + 522, + 555, + 582 + ], + "type": "text", + "content": ", which represents the affine approximation of the projective transformation, are utilized to obtain the covariance matrix " + }, + { + "bbox": [ + 313, + 522, + 555, + 582 + ], + "type": "inline_equation", + "content": "\\Sigma'" + }, + { + "bbox": [ + 313, + 522, + 555, + 582 + ], + "type": "text", + "content": " in 2D space, as:" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 392, + 581, + 553, + 596 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 392, + 581, + 553, + 596 + ], + "spans": [ + { + "bbox": [ + 392, + 581, + 553, + 596 + ], + "type": "interline_equation", + "content": "\\Sigma^ {\\prime} = J W \\Sigma W ^ {\\mathsf {T}} J ^ {\\mathsf {T}}. \\tag {3}", + "image_path": "bf8b3e3682cd89e615cc788f9f34849b9567ae4b3d919344d7a1682989b6291a.jpg" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 313, + 598, + 555, + 623 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 598, + 555, + 623 + ], + "spans": [ + { + "bbox": [ + 313, + 598, + 555, + 623 + ], + "type": "text", + "content": "Subsequently, a point-based alpha-blending rendering can be performed to determine the color of each pixel:" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 378, + 628, + 553, + 661 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 378, + 628, + 553, + 661 + ], + "spans": [ + { + "bbox": [ + 378, + 628, + 553, + 661 + ], + "type": "interline_equation", + "content": "C = \\sum_ {i} c _ {i} \\alpha_ {i} \\prod_ {j = 1} ^ {i - 1} (1 - \\alpha_ {i}), \\tag {4}", + "image_path": "4fc5b54b150d8c0e4c285560948f1d5dbd1e167e9b4f8de1838e99799ab785e6.jpg" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 313, + 666, + 556, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 666, + 556, + 715 + ], + "spans": [ + { + "bbox": [ + 313, + 666, + 556, + 715 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 313, + 666, + 556, + 715 + ], + "type": "inline_equation", + "content": "c_{i}" + }, + { + "bbox": [ + 313, + 666, + 556, + 715 + ], + "type": "text", + "content": " represents the color of each point, defined by spherical harmonics (SH) coefficients. The density " + }, + { + "bbox": [ + 313, + 666, + 556, + 715 + ], + "type": "inline_equation", + "content": "\\alpha_{i}" + }, + { + "bbox": [ + 313, + 666, + 556, + 715 + ], + "type": "text", + "content": " is computed as the product of 2D Gaussians and a learnable point-wise opacity." + } + ] + } + ], + "index": 12 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "type": "text", + "content": "26464" + } + ] + } + ], + "index": 13 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 2 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 56, + 68, + 170, + 156 + ], + "blocks": [ + { + "bbox": [ + 56, + 68, + 170, + 156 + ], + "lines": [ + { + "bbox": [ + 56, + 68, + 170, + 156 + ], + "spans": [ + { + "bbox": [ + 56, + 68, + 170, + 156 + ], + "type": "image", + "image_path": "50236d68e6607c558b614534d5b1607caac7d623b4b0319474552c14335540d6.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 71, + 157, + 156, + 167 + ], + "lines": [ + { + "bbox": [ + 71, + 157, + 156, + 167 + ], + "spans": [ + { + "bbox": [ + 71, + 157, + 156, + 167 + ], + "type": "text", + "content": "(I) Pinhole-based model" + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "type": "image", + "bbox": [ + 173, + 69, + 296, + 158 + ], + "blocks": [ + { + "bbox": [ + 173, + 69, + 296, + 158 + ], + "lines": [ + { + "bbox": [ + 173, + 69, + 296, + 158 + ], + "spans": [ + { + "bbox": [ + 173, + 69, + 296, + 158 + ], + "type": "image", + "image_path": "2ffeb0f907d5bb8b063ab3a6b31f0b73dfdfac3903d1bf989a548bb215a6c2b0.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 199, + 158, + 277, + 167 + ], + "lines": [ + { + "bbox": [ + 199, + 158, + 277, + 167 + ], + "spans": [ + { + "bbox": [ + 199, + 158, + 277, + 167 + ], + "type": "text", + "content": "(II) Lens-based model" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 55, + 176, + 296, + 254 + ], + "lines": [ + { + "bbox": [ + 55, + 176, + 296, + 254 + ], + "spans": [ + { + "bbox": [ + 55, + 176, + 296, + 254 + ], + "type": "text", + "content": "Figure 3. The difference between the pinhole-based model (I) and the lens-based model (II). For (I), light emitted from spatial point " + }, + { + "bbox": [ + 55, + 176, + 296, + 254 + ], + "type": "inline_equation", + "content": "O" + }, + { + "bbox": [ + 55, + 176, + 296, + 254 + ], + "type": "text", + "content": " directly hits the point " + }, + { + "bbox": [ + 55, + 176, + 296, + 254 + ], + "type": "inline_equation", + "content": "O'" + }, + { + "bbox": [ + 55, + 176, + 296, + 254 + ], + "type": "text", + "content": " of the image plane. For (II), we show the case that light emitted from " + }, + { + "bbox": [ + 55, + 176, + 296, + 254 + ], + "type": "inline_equation", + "content": "Q" + }, + { + "bbox": [ + 55, + 176, + 296, + 254 + ], + "type": "text", + "content": " converges on the point " + }, + { + "bbox": [ + 55, + 176, + 296, + 254 + ], + "type": "inline_equation", + "content": "Q'" + }, + { + "bbox": [ + 55, + 176, + 296, + 254 + ], + "type": "text", + "content": " of the image plane (In-Focus), and the case that light emitted from " + }, + { + "bbox": [ + 55, + 176, + 296, + 254 + ], + "type": "inline_equation", + "content": "P" + }, + { + "bbox": [ + 55, + 176, + 296, + 254 + ], + "type": "text", + "content": " converges on the point " + }, + { + "bbox": [ + 55, + 176, + 296, + 254 + ], + "type": "inline_equation", + "content": "P'" + }, + { + "bbox": [ + 55, + 176, + 296, + 254 + ], + "type": "text", + "content": " and continues to scatter onto the image plane forming a circle of confusion (Out-of-Focus)." + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_caption" + } + ], + "index": 2 + }, + { + "bbox": [ + 55, + 269, + 111, + 281 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 269, + 111, + 281 + ], + "spans": [ + { + "bbox": [ + 55, + 269, + 111, + 281 + ], + "type": "text", + "content": "4. Method" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 55, + 289, + 296, + 397 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 289, + 296, + 397 + ], + "spans": [ + { + "bbox": [ + 55, + 289, + 296, + 397 + ], + "type": "text", + "content": "We propose DoF-Gaussian, a controllable DoF approach for 3D-GS. The overall pipeline is illustrated in Fig. 2. In Section 4.1, we first develop a lens-based imaging model based on geometric optics principles to control DoF effects. Next, we employ the per-scene adjustment of depth priors to guide the correct scene geometry, as described in Section 4.2. Finally, in Section 4.3, we apply defocus-to-focus adaptation to minimize the inconsistencies in the CoC and enhance the defocus deblurring performance." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 55, + 404, + 235, + 417 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 404, + 235, + 417 + ], + "spans": [ + { + "bbox": [ + 55, + 404, + 235, + 417 + ], + "type": "text", + "content": "4.1. Lens-based Depth-of-Field Model" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 55, + 422, + 296, + 590 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 422, + 296, + 590 + ], + "spans": [ + { + "bbox": [ + 55, + 422, + 296, + 590 + ], + "type": "text", + "content": "The physical principles underlying imaging and DoF have been extensively studied in the field of geometric optics [10, 32]. In an idealized optical system, the light emitted from a spatial point " + }, + { + "bbox": [ + 55, + 422, + 296, + 590 + ], + "type": "inline_equation", + "content": "O" + }, + { + "bbox": [ + 55, + 422, + 296, + 590 + ], + "type": "text", + "content": " is projected onto a corresponding point " + }, + { + "bbox": [ + 55, + 422, + 296, + 590 + ], + "type": "inline_equation", + "content": "O^{\\prime}" + }, + { + "bbox": [ + 55, + 422, + 296, + 590 + ], + "type": "text", + "content": " in the image plane following the principles of pinhole imaging, as illustrated in Fig. 3 (I). However, real-world cameras operate based on the lens model, as shown in Fig. 3 (II). We show the two cases of in-focus and out-of-focus imaging, respectively. A spatial point " + }, + { + "bbox": [ + 55, + 422, + 296, + 590 + ], + "type": "inline_equation", + "content": "P" + }, + { + "bbox": [ + 55, + 422, + 296, + 590 + ], + "type": "text", + "content": ", located at a distance " + }, + { + "bbox": [ + 55, + 422, + 296, + 590 + ], + "type": "inline_equation", + "content": "d" + }, + { + "bbox": [ + 55, + 422, + 296, + 590 + ], + "type": "text", + "content": " from the lens, is projected on the imaging plane as a circular region referred to as the circle of confusion (CoC). The diameter of this region, " + }, + { + "bbox": [ + 55, + 422, + 296, + 590 + ], + "type": "inline_equation", + "content": "r(d)" + }, + { + "bbox": [ + 55, + 422, + 296, + 590 + ], + "type": "text", + "content": ", can be determined by the aperture parameter " + }, + { + "bbox": [ + 55, + 422, + 296, + 590 + ], + "type": "inline_equation", + "content": "\\mathcal{A}" + }, + { + "bbox": [ + 55, + 422, + 296, + 590 + ], + "type": "text", + "content": " and focus distance " + }, + { + "bbox": [ + 55, + 422, + 296, + 590 + ], + "type": "inline_equation", + "content": "\\mathcal{F}" + }, + { + "bbox": [ + 55, + 422, + 296, + 590 + ], + "type": "text", + "content": ", according to the following equation:" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 134, + 598, + 295, + 625 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 134, + 598, + 295, + 625 + ], + "spans": [ + { + "bbox": [ + 134, + 598, + 295, + 625 + ], + "type": "interline_equation", + "content": "r (d) = \\mathcal {A} \\left| \\frac {1}{\\mathcal {F}} - \\frac {1}{d} \\right|. \\tag {5}", + "image_path": "dff6b2b04a9c83acab145c51f2ef5610741bbb3d9354f4d3c2ad467affedb352.jpg" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 55, + 629, + 296, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 629, + 296, + 713 + ], + "spans": [ + { + "bbox": [ + 55, + 629, + 296, + 713 + ], + "type": "text", + "content": "The focus distance " + }, + { + "bbox": [ + 55, + 629, + 296, + 713 + ], + "type": "inline_equation", + "content": "\\mathcal{F}" + }, + { + "bbox": [ + 55, + 629, + 296, + 713 + ], + "type": "text", + "content": " primarily controls the depth position of the focal plane of the image, i.e., the depth of the sharp area, while the aperture parameter " + }, + { + "bbox": [ + 55, + 629, + 296, + 713 + ], + "type": "inline_equation", + "content": "\\mathcal{A}" + }, + { + "bbox": [ + 55, + 629, + 296, + 713 + ], + "type": "text", + "content": " determines the extent of the bokeh effect. For the point " + }, + { + "bbox": [ + 55, + 629, + 296, + 713 + ], + "type": "inline_equation", + "content": "Q" + }, + { + "bbox": [ + 55, + 629, + 296, + 713 + ], + "type": "text", + "content": " located at the focus distance " + }, + { + "bbox": [ + 55, + 629, + 296, + 713 + ], + "type": "inline_equation", + "content": "d = \\mathcal{F}" + }, + { + "bbox": [ + 55, + 629, + 296, + 713 + ], + "type": "text", + "content": ", the emitted rays directly converge through the lens to the corresponding point " + }, + { + "bbox": [ + 55, + 629, + 296, + 713 + ], + "type": "inline_equation", + "content": "Q'" + }, + { + "bbox": [ + 55, + 629, + 296, + 713 + ], + "type": "text", + "content": " in the image plane, resulting in the absence of a circle of confusion" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 315, + 72, + 485, + 84 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 315, + 72, + 485, + 84 + ], + "spans": [ + { + "bbox": [ + 315, + 72, + 485, + 84 + ], + "type": "text", + "content": "Algorithm 1 Lens-based imaging process" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 315, + 87, + 554, + 121 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 315, + 87, + 554, + 121 + ], + "spans": [ + { + "bbox": [ + 315, + 87, + 554, + 121 + ], + "type": "text", + "content": "Input: Rasterization rendering " + }, + { + "bbox": [ + 315, + 87, + 554, + 121 + ], + "type": "inline_equation", + "content": "C" + }, + { + "bbox": [ + 315, + 87, + 554, + 121 + ], + "type": "text", + "content": ", aperture parameter " + }, + { + "bbox": [ + 315, + 87, + 554, + 121 + ], + "type": "inline_equation", + "content": "\\mathcal{A}" + }, + { + "bbox": [ + 315, + 87, + 554, + 121 + ], + "type": "text", + "content": ", focus distance " + }, + { + "bbox": [ + 315, + 87, + 554, + 121 + ], + "type": "inline_equation", + "content": "\\mathcal{F}" + }, + { + "bbox": [ + 315, + 87, + 554, + 121 + ], + "type": "text", + "content": ", depth map " + }, + { + "bbox": [ + 315, + 87, + 554, + 121 + ], + "type": "inline_equation", + "content": "D" + }, + { + "bbox": [ + 315, + 87, + 554, + 121 + ], + "type": "text", + "content": ", gamma value " + }, + { + "bbox": [ + 315, + 87, + 554, + 121 + ], + "type": "inline_equation", + "content": "\\gamma" + }, + { + "bbox": [ + 315, + 87, + 554, + 121 + ], + "type": "text", + "content": ", confuse function Func" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 315, + 123, + 470, + 134 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 315, + 123, + 470, + 134 + ], + "spans": [ + { + "bbox": [ + 315, + 123, + 470, + 134 + ], + "type": "text", + "content": "Output: Defocus simulated result " + }, + { + "bbox": [ + 315, + 123, + 470, + 134 + ], + "type": "inline_equation", + "content": "C^*" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 325, + 134, + 449, + 148 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 325, + 134, + 449, + 148 + ], + "spans": [ + { + "bbox": [ + 325, + 134, + 449, + 148 + ], + "type": "interline_equation", + "content": "R \\leftarrow \\mathcal {A} \\left| \\frac {1}{\\bar {\\mathcal {E}}} - \\frac {1}{\\bar {D}} \\right| \\quad C \\leftarrow (C) ^ {\\gamma}", + "image_path": "b8db679d6e0d4b1cdc87eaf7805542aa06ee3b66ba4a8b6af95df4ce8e0524c9.jpg" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 326, + 148, + 399, + 159 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 326, + 148, + 399, + 159 + ], + "spans": [ + { + "bbox": [ + 326, + 148, + 399, + 159 + ], + "type": "interline_equation", + "content": "\\Phi = [ 0 ], \\tilde {C ^ {*}} = \\tilde {[} 0 ]", + "image_path": "622aaf688387d5901e8d1388027d682d9589efb04e9673fdaa0e30660fe1005d.jpg" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 325, + 159, + 449, + 171 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 325, + 159, + 449, + 171 + ], + "spans": [ + { + "bbox": [ + 325, + 159, + 449, + 171 + ], + "type": "text", + "content": "for pixel " + }, + { + "bbox": [ + 325, + 159, + 449, + 171 + ], + "type": "inline_equation", + "content": "i\\gets \\mathrm{Traverse}(C)" + }, + { + "bbox": [ + 325, + 159, + 449, + 171 + ], + "type": "text", + "content": " do" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 335, + 171, + 512, + 183 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 335, + 171, + 512, + 183 + ], + "spans": [ + { + "bbox": [ + 335, + 171, + 512, + 183 + ], + "type": "text", + "content": "for pixel " + }, + { + "bbox": [ + 335, + 171, + 512, + 183 + ], + "type": "inline_equation", + "content": "j\\gets" + }, + { + "bbox": [ + 335, + 171, + 512, + 183 + ], + "type": "text", + "content": " TraverseNeighbor " + }, + { + "bbox": [ + 335, + 171, + 512, + 183 + ], + "type": "inline_equation", + "content": "(c_i,r_i)" + }, + { + "bbox": [ + 335, + 171, + 512, + 183 + ], + "type": "text", + "content": " do" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 345, + 183, + 443, + 196 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 345, + 183, + 443, + 196 + ], + "spans": [ + { + "bbox": [ + 345, + 183, + 443, + 196 + ], + "type": "interline_equation", + "content": "\\lambda_ {i j} \\leftarrow \\operatorname {F u n c} (d _ {i}, | i - j |)", + "image_path": "a40975ba64728be858814e42a9ff0ce216c77ef7f9f4f4bc240755c9b5152de5.jpg" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 345, + 196, + 410, + 208 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 345, + 196, + 410, + 208 + ], + "spans": [ + { + "bbox": [ + 345, + 196, + 410, + 208 + ], + "type": "interline_equation", + "content": "\\Phi_ {j} \\leftarrow \\Phi_ {j} + \\lambda_ {i j}", + "image_path": "9d8fc67dc231475edc243533c822e0a221eb8791524f0ed400267bd73bb57352.jpg" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 345, + 208, + 420, + 220 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 345, + 208, + 420, + 220 + ], + "spans": [ + { + "bbox": [ + 345, + 208, + 420, + 220 + ], + "type": "interline_equation", + "content": "c _ {j} ^ {*} \\leftarrow c _ {j} ^ {*} + \\lambda_ {i j} \\cdot c _ {i}", + "image_path": "f79581f7b5eb0ecca8389a2665772bd0e7d431bad56b2666bd1c4a68d162047c.jpg" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 335, + 220, + 368, + 229 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 335, + 220, + 368, + 229 + ], + "spans": [ + { + "bbox": [ + 335, + 220, + 368, + 229 + ], + "type": "text", + "content": "end for" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 325, + 231, + 358, + 240 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 325, + 231, + 358, + 240 + ], + "spans": [ + { + "bbox": [ + 325, + 231, + 358, + 240 + ], + "type": "text", + "content": "end for" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 325, + 240, + 394, + 255 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 325, + 240, + 394, + 255 + ], + "spans": [ + { + "bbox": [ + 325, + 240, + 394, + 255 + ], + "type": "interline_equation", + "content": "C ^ {*} \\leftarrow (C ^ {*} / \\Phi) ^ {\\frac {1}{\\gamma}}", + "image_path": "0daa18eb30fd13efcd4ba212b9876e369406d9acd576a52929efe232c647ca44.jpg" + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 314, + 278, + 477, + 290 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 314, + 278, + 477, + 290 + ], + "spans": [ + { + "bbox": [ + 314, + 278, + 477, + 290 + ], + "type": "text", + "content": "(CoC). This also implies that " + }, + { + "bbox": [ + 314, + 278, + 477, + 290 + ], + "type": "inline_equation", + "content": "r(\\mathcal{F}) = 0" + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 313, + 290, + 554, + 361 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 290, + 554, + 361 + ], + "spans": [ + { + "bbox": [ + 313, + 290, + 554, + 361 + ], + "type": "text", + "content": "We set focus distance " + }, + { + "bbox": [ + 313, + 290, + 554, + 361 + ], + "type": "inline_equation", + "content": "\\mathcal{F}" + }, + { + "bbox": [ + 313, + 290, + 554, + 361 + ], + "type": "text", + "content": " and aperture parameter " + }, + { + "bbox": [ + 313, + 290, + 554, + 361 + ], + "type": "inline_equation", + "content": "\\mathcal{A}" + }, + { + "bbox": [ + 313, + 290, + 554, + 361 + ], + "type": "text", + "content": " as learnable for each input image, and these parameters are continuously updated with the optimization of 3D-GS. By modeling the lens camera, the output " + }, + { + "bbox": [ + 313, + 290, + 554, + 361 + ], + "type": "inline_equation", + "content": "C^*" + }, + { + "bbox": [ + 313, + 290, + 554, + 361 + ], + "type": "text", + "content": ", which includes bokeh blur effects, can be derived from the 3D-GS rasterizer rendering result " + }, + { + "bbox": [ + 313, + 290, + 554, + 361 + ], + "type": "inline_equation", + "content": "C" + }, + { + "bbox": [ + 313, + 290, + 554, + 361 + ], + "type": "text", + "content": " using our CUDA-based Algorithm 1." + } + ] + } + ], + "index": 25 + }, + { + "bbox": [ + 313, + 361, + 554, + 410 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 361, + 554, + 410 + ], + "spans": [ + { + "bbox": [ + 313, + 361, + 554, + 410 + ], + "type": "text", + "content": "Ideally, the confuse function is represented by the indicator function " + }, + { + "bbox": [ + 313, + 361, + 554, + 410 + ], + "type": "inline_equation", + "content": "\\mathbb{I}(r(d) > l)" + }, + { + "bbox": [ + 313, + 361, + 554, + 410 + ], + "type": "text", + "content": ". To achieve a smooth and natural DoF effect, we substitute it with a differentiable function, as proposed in Busam et al. [3]:" + } + ] + } + ], + "index": 26 + }, + { + "bbox": [ + 347, + 416, + 554, + 440 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 347, + 416, + 554, + 440 + ], + "spans": [ + { + "bbox": [ + 347, + 416, + 554, + 440 + ], + "type": "interline_equation", + "content": "F u n c (d, l) = \\frac {1}{2} + \\frac {1}{2} \\tanh (\\alpha (r (d) - l)), \\tag {6}", + "image_path": "14c24c47310c2a7eacd87587c130ebded6b7792a88d7716e859e19f7de49b68f.jpg" + } + ] + } + ], + "index": 27 + }, + { + "bbox": [ + 313, + 445, + 554, + 469 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 445, + 554, + 469 + ], + "spans": [ + { + "bbox": [ + 313, + 445, + 554, + 469 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 313, + 445, + 554, + 469 + ], + "type": "inline_equation", + "content": "\\alpha" + }, + { + "bbox": [ + 313, + 445, + 554, + 469 + ], + "type": "text", + "content": " defines the smoothness of confuse transition and " + }, + { + "bbox": [ + 313, + 445, + 554, + 469 + ], + "type": "inline_equation", + "content": "l" + }, + { + "bbox": [ + 313, + 445, + 554, + 469 + ], + "type": "text", + "content": " represents the distance between two pixels." + } + ] + } + ], + "index": 28 + }, + { + "bbox": [ + 313, + 469, + 554, + 517 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 469, + 554, + 517 + ], + "spans": [ + { + "bbox": [ + 313, + 469, + 554, + 517 + ], + "type": "text", + "content": "We supervise the training by computing the reconstruction loss between the output image " + }, + { + "bbox": [ + 313, + 469, + 554, + 517 + ], + "type": "inline_equation", + "content": "C^*" + }, + { + "bbox": [ + 313, + 469, + 554, + 517 + ], + "type": "text", + "content": " and the shallow DoF input image " + }, + { + "bbox": [ + 313, + 469, + 554, + 517 + ], + "type": "inline_equation", + "content": "I" + }, + { + "bbox": [ + 313, + 469, + 554, + 517 + ], + "type": "text", + "content": ", where the loss function is a combination of the " + }, + { + "bbox": [ + 313, + 469, + 554, + 517 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_1" + }, + { + "bbox": [ + 313, + 469, + 554, + 517 + ], + "type": "text", + "content": " and a D-SSIM term, with " + }, + { + "bbox": [ + 313, + 469, + 554, + 517 + ], + "type": "inline_equation", + "content": "\\lambda" + }, + { + "bbox": [ + 313, + 469, + 554, + 517 + ], + "type": "text", + "content": " set to 0.2:" + } + ] + } + ], + "index": 29 + }, + { + "bbox": [ + 328, + 525, + 553, + 539 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 328, + 525, + 553, + 539 + ], + "spans": [ + { + "bbox": [ + 328, + 525, + 553, + 539 + ], + "type": "interline_equation", + "content": "\\mathcal {L} _ {r e c} = (1 - \\lambda) \\mathcal {L} _ {1} \\left(I, C ^ {*}\\right) + \\lambda \\mathcal {L} _ {D - S S I M} \\left(I, C ^ {*}\\right). \\tag {7}", + "image_path": "354b07ec02324aa6f52392218d7c4b2db03b8337526dfb2b479a2e3ad261a1d0.jpg" + } + ] + } + ], + "index": 30 + }, + { + "bbox": [ + 313, + 545, + 554, + 570 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 545, + 554, + 570 + ], + "spans": [ + { + "bbox": [ + 313, + 545, + 554, + 570 + ], + "type": "text", + "content": "For inference, we set the aperture size " + }, + { + "bbox": [ + 313, + 545, + 554, + 570 + ], + "type": "inline_equation", + "content": "\\mathcal{A}" + }, + { + "bbox": [ + 313, + 545, + 554, + 570 + ], + "type": "text", + "content": " to 0 so that we can render fully clear images for novel view synthesis." + } + ] + } + ], + "index": 31 + }, + { + "bbox": [ + 313, + 576, + 517, + 589 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 576, + 517, + 589 + ], + "spans": [ + { + "bbox": [ + 313, + 576, + 517, + 589 + ], + "type": "text", + "content": "4.2. Per-Scene Adjustment of Depth Priors" + } + ] + } + ], + "index": 32 + }, + { + "bbox": [ + 313, + 594, + 555, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 594, + 555, + 714 + ], + "spans": [ + { + "bbox": [ + 313, + 594, + 555, + 714 + ], + "type": "text", + "content": "Unlike previous work that assumes input images are all-in-focus and fully clear, it becomes challenging to accurately reconstruct scene geometry and render depth maps for bokeh blurring input images. However, rendering images with varying depths of field relies on precise depth information. To address this issue, we propose using depth prior as useful guidance for the accurate reconstruction of scene geometry. However, due to the presence of bokeh blur in shallow DoF inputs, directly employing a monocular depth network to predict depth maps " + }, + { + "bbox": [ + 313, + 594, + 555, + 714 + ], + "type": "inline_equation", + "content": "D_{pred}" + }, + { + "bbox": [ + 313, + 594, + 555, + 714 + ], + "type": "text", + "content": " as pseudo" + } + ] + } + ], + "index": 33 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "text", + "content": "26465" + } + ] + } + ], + "index": 34 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 3 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 57, + 79, + 294, + 159 + ], + "blocks": [ + { + "bbox": [ + 78, + 70, + 97, + 78 + ], + "lines": [ + { + "bbox": [ + 78, + 70, + 97, + 78 + ], + "spans": [ + { + "bbox": [ + 78, + 70, + 97, + 78 + ], + "type": "text", + "content": "RGB" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 137, + 70, + 161, + 78 + ], + "lines": [ + { + "bbox": [ + 137, + 70, + 161, + 78 + ], + "spans": [ + { + "bbox": [ + 137, + 70, + 161, + 78 + ], + "type": "text", + "content": "BAGS" + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 177, + 70, + 235, + 79 + ], + "lines": [ + { + "bbox": [ + 177, + 70, + 235, + 79 + ], + "spans": [ + { + "bbox": [ + 177, + 70, + 235, + 79 + ], + "type": "text", + "content": "w/o depth priors" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 255, + 70, + 273, + 78 + ], + "lines": [ + { + "bbox": [ + 255, + 70, + 273, + 78 + ], + "spans": [ + { + "bbox": [ + 255, + 70, + 273, + 78 + ], + "type": "text", + "content": "Ours" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 57, + 79, + 294, + 159 + ], + "lines": [ + { + "bbox": [ + 57, + 79, + 294, + 159 + ], + "spans": [ + { + "bbox": [ + 57, + 79, + 294, + 159 + ], + "type": "image", + "image_path": "b533d44be00dafbe9b106e039860d236fe97ff55657c7818f35cf452b64abaa3.jpg" + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 55, + 167, + 295, + 213 + ], + "lines": [ + { + "bbox": [ + 55, + 167, + 295, + 213 + ], + "spans": [ + { + "bbox": [ + 55, + 167, + 295, + 213 + ], + "type": "text", + "content": "Figure 4. Effects of the depth priors. The per-scene adjustment of depth priors enhance the geometric structure and yield a more accurate depth map than that estimated by BAGS [34] and a variant of our method without depth priors." + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "image_caption" + } + ], + "index": 4 + }, + { + "bbox": [ + 55, + 228, + 296, + 407 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 228, + 296, + 407 + ], + "spans": [ + { + "bbox": [ + 55, + 228, + 296, + 407 + ], + "type": "text", + "content": "ground truth to regularize the Gaussians rendered depth map, as done in [5, 21], does not yield satisfactory results. Inspired by [53], we propose adapting per-scene depth priors through fine-tuning a monocular depth network based on the scene's sparse reconstruction. Specifically, we utilize COLMAP [40, 41] to obtain a sparse Structure-from-Motion (SfM) point cloud and create a set of 3D Gaussians. Consequently, we can derive per-view sparse depth maps " + }, + { + "bbox": [ + 55, + 228, + 296, + 407 + ], + "type": "inline_equation", + "content": "D_{sparse}" + }, + { + "bbox": [ + 55, + 228, + 296, + 407 + ], + "type": "text", + "content": " by projecting the 3D point clouds after multiview stereo processing. Although the resulting depth map is sparse, it is robust and can serve as a supervisory signal for training the per-scene depth priors. To address the scale ambiguity inherent in acquired depth maps, we apply a silog loss [9] to fine-tune the depth network [22, 39], expressed as:" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 64, + 414, + 295, + 447 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 64, + 414, + 295, + 447 + ], + "spans": [ + { + "bbox": [ + 64, + 414, + 295, + 447 + ], + "type": "interline_equation", + "content": "\\mathcal {L} _ {\\text {s i l o g}} = \\frac {1}{2 M} \\sum_ {i = 1} ^ {M} \\left(\\log \\left(e ^ {s} D _ {\\text {p r e d}}\\right) - \\log \\left(D _ {\\text {s p a r s e}}\\right)\\right) ^ {2}, \\tag {8}", + "image_path": "dc54eb5b6c0954f8aeda308bc5281d61e4355ff02a7d7ae5bd2d6a8a660a2f4b.jpg" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 55, + 454, + 296, + 538 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 454, + 296, + 538 + ], + "spans": [ + { + "bbox": [ + 55, + 454, + 296, + 538 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 55, + 454, + 296, + 538 + ], + "type": "inline_equation", + "content": "e^s" + }, + { + "bbox": [ + 55, + 454, + 296, + 538 + ], + "type": "text", + "content": " is the scale factor and " + }, + { + "bbox": [ + 55, + 454, + 296, + 538 + ], + "type": "inline_equation", + "content": "M" + }, + { + "bbox": [ + 55, + 454, + 296, + 538 + ], + "type": "text", + "content": " is the number of pixels in the image used for calculation. We apply the rasterization technique from Rade-GS [59], which enables the robust rendering of depth maps " + }, + { + "bbox": [ + 55, + 454, + 296, + 538 + ], + "type": "inline_equation", + "content": "D" + }, + { + "bbox": [ + 55, + 454, + 296, + 538 + ], + "type": "text", + "content": " for 3D-GS scenes. We predict the depth maps " + }, + { + "bbox": [ + 55, + 454, + 296, + 538 + ], + "type": "inline_equation", + "content": "D_{pred}" + }, + { + "bbox": [ + 55, + 454, + 296, + 538 + ], + "type": "text", + "content": " from multi-view input images using the fine-tuned depth network and use them as depth priors to regularize " + }, + { + "bbox": [ + 55, + 454, + 296, + 538 + ], + "type": "inline_equation", + "content": "D" + }, + { + "bbox": [ + 55, + 454, + 296, + 538 + ], + "type": "text", + "content": ". The depth loss can be described as follows:" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 122, + 547, + 295, + 561 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 122, + 547, + 295, + 561 + ], + "spans": [ + { + "bbox": [ + 122, + 547, + 295, + 561 + ], + "type": "interline_equation", + "content": "\\mathcal {L} _ {\\text {d e p t h}} = \\left| \\left| D - D _ {\\text {p r e d}} \\right| \\right| _ {2}. \\tag {9}", + "image_path": "6f4f2cbe7bea5430daeb15c3a48459dca8121076ec5347d90ad4cce31bb3595c.jpg" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 55, + 568, + 296, + 654 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 568, + 296, + 654 + ], + "spans": [ + { + "bbox": [ + 55, + 568, + 296, + 654 + ], + "type": "text", + "content": "As illustrated in Fig. 4, it is evident that the depth priors significantly enhance the geometric structure, resulting in a more accurate depth map compared to both BAGS [34] and the variant of our method without depth priors. In Section 5.3, we further conduct an ablation study comparing several other depth supervision strategies to demonstrate the effectiveness of our approach." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 55, + 660, + 214, + 673 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 660, + 214, + 673 + ], + "spans": [ + { + "bbox": [ + 55, + 660, + 214, + 673 + ], + "type": "text", + "content": "4.3. Defocus-to-Focus Adaptation" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 55, + 677, + 296, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 677, + 296, + 713 + ], + "spans": [ + { + "bbox": [ + 55, + 677, + 296, + 713 + ], + "type": "text", + "content": "The CoC effect achieved under ideal optical imaging conditions, as described in Algorithm 1, inevitably differs from the real CoC effect produced by an actual DSLR camera." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 313, + 72, + 553, + 96 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 72, + 553, + 96 + ], + "spans": [ + { + "bbox": [ + 313, + 72, + 553, + 96 + ], + "type": "text", + "content": "This discrepancy hinders our ability to accurately model the out-of-focus blur in shallow DoF input images." + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 313, + 97, + 554, + 228 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 97, + 554, + 228 + ], + "spans": [ + { + "bbox": [ + 313, + 97, + 554, + 228 + ], + "type": "text", + "content": "To achieve sharper scenes, we propose an adaptation process that allows transitions from defocus to focus. Specifically, modeling the defocus effects with bokeh blur facilitates learning accurate lens parameters. Consequently, we place greater emphasis on capturing the sharp areas in training images once the lens parameters " + }, + { + "bbox": [ + 313, + 97, + 554, + 228 + ], + "type": "inline_equation", + "content": "\\mathcal{A}" + }, + { + "bbox": [ + 313, + 97, + 554, + 228 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 313, + 97, + 554, + 228 + ], + "type": "inline_equation", + "content": "\\mathcal{F}" + }, + { + "bbox": [ + 313, + 97, + 554, + 228 + ], + "type": "text", + "content": " have been optimized and converged. This approach is effective because, when the focus distance is determined, the sharp regions in the scene become approximately identifiable. After " + }, + { + "bbox": [ + 313, + 97, + 554, + 228 + ], + "type": "inline_equation", + "content": "t" + }, + { + "bbox": [ + 313, + 97, + 554, + 228 + ], + "type": "text", + "content": " optimization iterations, we reweight the " + }, + { + "bbox": [ + 313, + 97, + 554, + 228 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{rec}" + }, + { + "bbox": [ + 313, + 97, + 554, + 228 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 313, + 97, + 554, + 228 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{depth}" + }, + { + "bbox": [ + 313, + 97, + 554, + 228 + ], + "type": "text", + "content": " loss by employing a step-like function:" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 324, + 236, + 554, + 264 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 324, + 236, + 554, + 264 + ], + "spans": [ + { + "bbox": [ + 324, + 236, + 554, + 264 + ], + "type": "interline_equation", + "content": "\\Psi (x) = \\left\\{ \\begin{array}{c c} 1 & , \\text {i t e r a t i o n s} < t \\\\ 1 / \\left(1 + e ^ {- a \\cdot (x - b)}\\right) & , \\text {i t e r a t i o n s} \\geq t \\end{array} \\right. \\tag {10}", + "image_path": "88b9349b9f1f1a556837b8bd277d8ef30e13e12bba0111282d86ee7fdd9199c4.jpg" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 313, + 273, + 533, + 287 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 273, + 533, + 287 + ], + "spans": [ + { + "bbox": [ + 313, + 273, + 533, + 287 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 313, + 273, + 533, + 287 + ], + "type": "inline_equation", + "content": "a" + }, + { + "bbox": [ + 313, + 273, + 533, + 287 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 313, + 273, + 533, + 287 + ], + "type": "inline_equation", + "content": "b" + }, + { + "bbox": [ + 313, + 273, + 533, + 287 + ], + "type": "text", + "content": " are hyperparameters, and " + }, + { + "bbox": [ + 313, + 273, + 533, + 287 + ], + "type": "inline_equation", + "content": "x = \\left|\\frac{1}{\\mathcal{F}} - \\frac{1}{d}\\right|" + }, + { + "bbox": [ + 313, + 273, + 533, + 287 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 313, + 286, + 554, + 334 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 286, + 554, + 334 + ], + "spans": [ + { + "bbox": [ + 313, + 286, + 554, + 334 + ], + "type": "text", + "content": "In addition, the aperture size " + }, + { + "bbox": [ + 313, + 286, + 554, + 334 + ], + "type": "inline_equation", + "content": "\\mathcal{A}" + }, + { + "bbox": [ + 313, + 286, + 554, + 334 + ], + "type": "text", + "content": " is generally large after the convergence, which can cause blurring of pixels that should be in focus. Hence, we also reweight the aperture size " + }, + { + "bbox": [ + 313, + 286, + 554, + 334 + ], + "type": "inline_equation", + "content": "\\mathcal{A}" + }, + { + "bbox": [ + 313, + 286, + 554, + 334 + ], + "type": "text", + "content": " on a pixel-wise basis:" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 408, + 343, + 553, + 357 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 408, + 343, + 553, + 357 + ], + "spans": [ + { + "bbox": [ + 408, + 343, + 553, + 357 + ], + "type": "interline_equation", + "content": "\\mathcal {A} ^ {\\prime} = \\mathcal {A} \\cdot \\Psi . \\tag {11}", + "image_path": "23fe2339ca37be6bfc79d609e567ca5fe0a1b4c90a36f85adddea7b4735045d2.jpg" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 313, + 366, + 554, + 451 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 366, + 554, + 451 + ], + "spans": [ + { + "bbox": [ + 313, + 366, + 554, + 451 + ], + "type": "text", + "content": "Full Objective. We derive the final loss terms by incorporating the reconstruction loss from Eqn. 7, the depth loss from Eqn. 9, and an additional normal consistency loss as proposed by Huang et al. [11]. The normal consistency loss ensures that the Gaussian splats align with the surface by measuring the consistency between the normal directions computed from the Gaussian and the depth map:" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 371, + 460, + 553, + 485 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 371, + 460, + 553, + 485 + ], + "spans": [ + { + "bbox": [ + 371, + 460, + 553, + 485 + ], + "type": "interline_equation", + "content": "\\mathcal {L} _ {\\text {n o r m a l}} = \\sum_ {i} w _ {i} \\left(1 - n _ {i} ^ {\\top} \\hat {n} _ {i}\\right), \\tag {12}", + "image_path": "4d42425a5238ce2531f892af6797c699b84a406ca18d89788a1e34a91f6754cd.jpg" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 313, + 493, + 554, + 540 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 493, + 554, + 540 + ], + "spans": [ + { + "bbox": [ + 313, + 493, + 554, + 540 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 313, + 493, + 554, + 540 + ], + "type": "inline_equation", + "content": "\\hat{n}" + }, + { + "bbox": [ + 313, + 493, + 554, + 540 + ], + "type": "text", + "content": " represents the surface normal direction obtained by applying finite differences on the depth map, " + }, + { + "bbox": [ + 313, + 493, + 554, + 540 + ], + "type": "inline_equation", + "content": "i" + }, + { + "bbox": [ + 313, + 493, + 554, + 540 + ], + "type": "text", + "content": " indexes the intersected splats along the ray and " + }, + { + "bbox": [ + 313, + 493, + 554, + 540 + ], + "type": "inline_equation", + "content": "w" + }, + { + "bbox": [ + 313, + 493, + 554, + 540 + ], + "type": "text", + "content": " denotes the blending weight of the intersection point." + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 326, + 541, + 438, + 553 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 326, + 541, + 438, + 553 + ], + "spans": [ + { + "bbox": [ + 326, + 541, + 438, + 553 + ], + "type": "text", + "content": "Our final training loss " + }, + { + "bbox": [ + 326, + 541, + 438, + 553 + ], + "type": "inline_equation", + "content": "\\mathcal{L}" + }, + { + "bbox": [ + 326, + 541, + 438, + 553 + ], + "type": "text", + "content": " is:" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 334, + 563, + 553, + 577 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 334, + 563, + 553, + 577 + ], + "spans": [ + { + "bbox": [ + 334, + 563, + 553, + 577 + ], + "type": "interline_equation", + "content": "\\mathcal {L} = \\Psi \\odot \\left(\\mathcal {L} _ {\\text {r e c}} + w _ {d} \\mathcal {L} _ {\\text {d e p t h}}\\right) + w _ {n} \\mathcal {L} _ {\\text {n o r m a l}}, \\tag {13}", + "image_path": "d98196559e854c68dc72a4fb204ee890592a61d532b461f9e08b6e6221257e6c.jpg" + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 313, + 586, + 440, + 598 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 586, + 440, + 598 + ], + "spans": [ + { + "bbox": [ + 313, + 586, + 440, + 598 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 313, + 586, + 440, + 598 + ], + "type": "inline_equation", + "content": "\\odot" + }, + { + "bbox": [ + 313, + 586, + 440, + 598 + ], + "type": "text", + "content": " is Hadamard product." + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 313, + 609, + 395, + 624 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 609, + 395, + 624 + ], + "spans": [ + { + "bbox": [ + 313, + 609, + 395, + 624 + ], + "type": "text", + "content": "5. Experiments" + } + ] + } + ], + "index": 25 + }, + { + "bbox": [ + 313, + 629, + 554, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 629, + 554, + 713 + ], + "spans": [ + { + "bbox": [ + 313, + 629, + 554, + 713 + ], + "type": "text", + "content": "Implementation Details. In our experiments, we set the smoothness of confuse transition " + }, + { + "bbox": [ + 313, + 629, + 554, + 713 + ], + "type": "inline_equation", + "content": "\\alpha = 4" + }, + { + "bbox": [ + 313, + 629, + 554, + 713 + ], + "type": "text", + "content": ". In addition, we set " + }, + { + "bbox": [ + 313, + 629, + 554, + 713 + ], + "type": "inline_equation", + "content": "a = 15" + }, + { + "bbox": [ + 313, + 629, + 554, + 713 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 313, + 629, + 554, + 713 + ], + "type": "inline_equation", + "content": "b = 0.3" + }, + { + "bbox": [ + 313, + 629, + 554, + 713 + ], + "type": "text", + "content": ", and " + }, + { + "bbox": [ + 313, + 629, + 554, + 713 + ], + "type": "inline_equation", + "content": "t = 10000" + }, + { + "bbox": [ + 313, + 629, + 554, + 713 + ], + "type": "text", + "content": " for the defocus-to-focus adaptation. Our method is built upon Mip-Splatting [58], and our optimization strategy and hyperparameter settings remain consistent with it. We train each scene for 30000 iterations and set loss weights " + }, + { + "bbox": [ + 313, + 629, + 554, + 713 + ], + "type": "inline_equation", + "content": "w_{d} = 0.01" + }, + { + "bbox": [ + 313, + 629, + 554, + 713 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 313, + 629, + 554, + 713 + ], + "type": "inline_equation", + "content": "w_{n} = 0.05" + }, + { + "bbox": [ + 313, + 629, + 554, + 713 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 26 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 318, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 318, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 318, + 757 + ], + "type": "text", + "content": "26466" + } + ] + } + ], + "index": 27 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 4 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 60, + 102, + 550, + 236 + ], + "blocks": [ + { + "bbox": [ + 55, + 70, + 555, + 95 + ], + "lines": [ + { + "bbox": [ + 55, + 70, + 555, + 95 + ], + "spans": [ + { + "bbox": [ + 55, + 70, + 555, + 95 + ], + "type": "text", + "content": "Table 1. Quantitative comparisons on the defocus blur dataset of Deblur NeRF [26]. We report the PSNR, SSIM, and LPIPS metrics and color each cell as best and second best. Our method outperforms other existing approaches across most scenes." + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 60, + 102, + 550, + 236 + ], + "lines": [ + { + "bbox": [ + 60, + 102, + 550, + 236 + ], + "spans": [ + { + "bbox": [ + 60, + 102, + 550, + 236 + ], + "type": "table", + "html": "
MethodDeblur-NeRF [26]DoF-NeRF [55]DP-NeRF [19]BAGS [34]Deblurring 3DGS [18]Ours
PSNRSSIMLPIPSPSNRSSIMLPIPSPSNRSSIMLPIPSPSNRSSIMLPIPSPSNRSSIMLPIPSPSNRSSIMLPIPS
Cake26.270.7800.12824.530.7310.20626.160.7780.12727.210.8180.10826.880.8030.11626.830.8080.101
Caps23.870.7130.16122.670.6360.26123.950.7120.14324.160.7250.15924.500.7420.14924.620.7490.147
Cisco20.830.7270.08720.640.7240.10720.730.7260.08420.790.7430.07020.830.7320.07921.000.7440.069
Coral19.850.6000.12119.830.5700.24020.110.6110.11820.530.6280.11119.780.6080.13120.370.6300.109
Cupcake22.260.7220.11621.890.7060.14322.800.7410.09622.930.7620.08022.110.7340.09922.970.7570.079
Cups26.210.7990.12725.260.7650.20226.750.8140.10426.270.8230.10426.280.8240.10326.010.8170.100
Daisy23.520.6870.12123.220.6580.19423.790.6970.10823.740.7460.06223.540.7310.09523.930.7350.071
Sausage18.010.5000.18017.860.4880.28018.350.5440.14718.760.5740.11018.990.5700.14119.110.5760.119
Seal26.040.7770.10524.850.6870.14325.950.7780.10326.520.8120.09026.180.8170.09826.570.8250.087
Tools27.810.8950.06126.210.8540.12828.070.8980.05428.600.9130.04627.960.9100.05828.290.9130.051
", + "image_path": "79c8b25add3113ada87802a6ad836b1b3659d41c21763d83280c90f754571fcd.jpg" + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "table_body" + } + ], + "index": 1 + }, + { + "type": "table", + "bbox": [ + 58, + 275, + 293, + 319 + ], + "blocks": [ + { + "bbox": [ + 60, + 255, + 290, + 266 + ], + "lines": [ + { + "bbox": [ + 60, + 255, + 290, + 266 + ], + "spans": [ + { + "bbox": [ + 60, + 255, + 290, + 266 + ], + "type": "text", + "content": "Table 2. Quantitative comparisons on our synthetic dataset." + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 58, + 275, + 293, + 319 + ], + "lines": [ + { + "bbox": [ + 58, + 275, + 293, + 319 + ], + "spans": [ + { + "bbox": [ + 58, + 275, + 293, + 319 + ], + "type": "table", + "html": "
MethodPSNR↑SSIM↑LPIPS↓δA↓δF↓
DoF-NeRF [55]25.590.7880.2070.1960.256
Ours28.700.8640.0950.1260.079
", + "image_path": "ce5072281c23873e0d90ca790356afc4b28f7638d507e7d8d2468a804e3c793a.jpg" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "table_body" + } + ], + "index": 3 + }, + { + "bbox": [ + 55, + 337, + 295, + 420 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 337, + 295, + 420 + ], + "spans": [ + { + "bbox": [ + 55, + 337, + 295, + 420 + ], + "type": "text", + "content": "Baselines and Evaluation Metrics. We compare our method with state-of-the-art deblurring techniques [18, 26, 34] and a depth-of-field method [55]. To quantitatively evaluate the quality of novel view images, we adopt widely used metrics such as PSNR, SSIM [52], and LPIPS [60]. We will compare with Wang et al. [51] once they release the code or provide the full experimental results in their paper." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 55, + 422, + 296, + 506 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 422, + 296, + 506 + ], + "spans": [ + { + "bbox": [ + 55, + 422, + 296, + 506 + ], + "type": "text", + "content": "Datasets. Following [18, 26, 34], we evaluate our model on the Deblur-NeRF dataset [26]. Subsequently, we conduct further evaluations on the Real Forward-facing dataset [28] and T&T_DB dataset [16]. The effectiveness of our method is validated on these datasets for both shallow DoF inputs and normal inputs. Furthermore, we also introduce a dataset in Section 5.1 to evaluate our model more comprehensively." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 55, + 528, + 171, + 540 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 528, + 171, + 540 + ], + "spans": [ + { + "bbox": [ + 55, + 528, + 171, + 540 + ], + "type": "text", + "content": "5.1. A Synthetic Dataset" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 55, + 545, + 295, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 545, + 295, + 713 + ], + "spans": [ + { + "bbox": [ + 55, + 545, + 295, + 713 + ], + "type": "text", + "content": "We introduce a synthetic dataset, as illustrated in Fig. 6, based on depth estimation [2] and a single-image DoF rendering method [35] for each image in the Real Forward Facing dataset [28] and T&T_DB dataset [16]. This allows us to evaluate the refocusing ability of our model and to determine whether it accurately learns the correct aperture size and focus distance. In particular, by feeding the DoF rendering method [35] the known aperture size and focus distance, we can convert the all-in-focus images into shallow DoF images. Unlike existing defocus blur datasets proposed by Ma et al. [26] and Wu et al. [55], the test set in our dataset comprises images with shallow DoF. With known lens parameters, our model fits this shallow DoF effect to quantitatively measure the refocusing ability. Meanwhile, we establish the" + } + ] + } + ], + "index": 7 + }, + { + "type": "table", + "bbox": [ + 328, + 297, + 539, + 346 + ], + "blocks": [ + { + "bbox": [ + 313, + 255, + 555, + 289 + ], + "lines": [ + { + "bbox": [ + 313, + 255, + 555, + 289 + ], + "spans": [ + { + "bbox": [ + 313, + 255, + 555, + 289 + ], + "type": "text", + "content": "Table 3. Comparisons in all-in-focus settings. We compare our method with Mip-Splatting on all-in-focus datasets to validate the effectiveness of our model under general input conditions." + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 328, + 297, + 539, + 346 + ], + "lines": [ + { + "bbox": [ + 328, + 297, + 539, + 346 + ], + "spans": [ + { + "bbox": [ + 328, + 297, + 539, + 346 + ], + "type": "table", + "html": "
MethodPSNR↑SSIM↑LPIPS↓
Mip-Splatting [58]27.050.8930.115
Ours27.810.9020.117
", + "image_path": "7d9c5d5774677e3b5777d01bfcf1b1c8589a980b9c22f6dd6a7fb25fc8ee8684.jpg" + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "table_body" + } + ], + "index": 9 + }, + { + "bbox": [ + 313, + 363, + 555, + 411 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 363, + 555, + 411 + ], + "spans": [ + { + "bbox": [ + 313, + 363, + 555, + 411 + ], + "type": "text", + "content": "known lens parameters beforehand as ground truth, allowing us to calculate the errors " + }, + { + "bbox": [ + 313, + 363, + 555, + 411 + ], + "type": "inline_equation", + "content": "\\delta_{\\mathcal{A}}" + }, + { + "bbox": [ + 313, + 363, + 555, + 411 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 313, + 363, + 555, + 411 + ], + "type": "inline_equation", + "content": "\\delta_{\\mathcal{F}}" + }, + { + "bbox": [ + 313, + 363, + 555, + 411 + ], + "type": "text", + "content": " between the learned parameters and the ground truth. More details can be found in the supplementary material." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 313, + 419, + 400, + 432 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 419, + 400, + 432 + ], + "spans": [ + { + "bbox": [ + 313, + 419, + 400, + 432 + ], + "type": "text", + "content": "5.2. Comparisons" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 313, + 437, + 554, + 521 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 437, + 554, + 521 + ], + "spans": [ + { + "bbox": [ + 313, + 437, + 554, + 521 + ], + "type": "text", + "content": "Quantitative Comparisons. We quantitatively compare DoF-Gaussian against other state-of-the-art methods using a real-world defocus blur dataset [26]. As shown in Table 1, our method outperforms other state-of-the-art baselines. These results suggest that our method achieves superior defocus deblurring performance and produces higher-quality novel view synthesis images." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 313, + 521, + 554, + 593 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 521, + 554, + 593 + ], + "spans": [ + { + "bbox": [ + 313, + 521, + 554, + 593 + ], + "type": "text", + "content": "Qualitative Comparisons. Visual qualitative comparisons are presented in Fig. 5. Our method surpasses all other methods in terms of image fidelity, generating novel view images that are more faithful to the ground truth images and exhibit less blur. Specifically, the bottle caps and shelves are the sharpest in our rendered images." + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 313, + 594, + 555, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 594, + 555, + 713 + ], + "spans": [ + { + "bbox": [ + 313, + 594, + 555, + 713 + ], + "type": "text", + "content": "All-in-focus Inputs. To further validate the effectiveness of our method in the all-in-focus setting, we design an experiment using wide DoF images as inputs. We conduct a comparison between our method and Mip-Splatting [58] on two all-in-focus datasets, the Real Forward Facing dataset [28] and the T&T_DB dataset [16]. As shown in Table 3, our method achieves comparable or even better performance to Mip-Splatting on the average metrics across the datasets. This demonstrates that our lens-based model not only excels with shallow DoF inputs but also achieves good" + } + ] + } + ], + "index": 14 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "type": "text", + "content": "26467" + } + ] + } + ], + "index": 15 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 5 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 56, + 68, + 555, + 250 + ], + "blocks": [ + { + "bbox": [ + 56, + 68, + 555, + 250 + ], + "lines": [ + { + "bbox": [ + 56, + 68, + 555, + 250 + ], + "spans": [ + { + "bbox": [ + 56, + 68, + 555, + 250 + ], + "type": "image", + "image_path": "afc37f6ecac264bd729b25afea8480afbf9695b641eb6ce660296001e8ecb764.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 55, + 257, + 555, + 281 + ], + "lines": [ + { + "bbox": [ + 55, + 257, + 555, + 281 + ], + "spans": [ + { + "bbox": [ + 55, + 257, + 555, + 281 + ], + "type": "text", + "content": "Figure 5. Qualitative comparisons against all baselines. Compared to other state-of-the-art methods, our method represents sharper scenes and generates novel view images with less blur." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "type": "image", + "bbox": [ + 56, + 292, + 295, + 416 + ], + "blocks": [ + { + "bbox": [ + 56, + 292, + 295, + 416 + ], + "lines": [ + { + "bbox": [ + 56, + 292, + 295, + 416 + ], + "spans": [ + { + "bbox": [ + 56, + 292, + 295, + 416 + ], + "type": "image", + "image_path": "37b868e3648129b9042355f069a7311dd491a7ed5d3093aa4d8986a139ebc8f1.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 55, + 424, + 295, + 469 + ], + "lines": [ + { + "bbox": [ + 55, + 424, + 295, + 469 + ], + "spans": [ + { + "bbox": [ + 55, + 424, + 295, + 469 + ], + "type": "text", + "content": "Figure 6. Examples in our dataset. For each scene, we apply a DoF rendering method [35] to convert all-in-focus images to shallow DoF images with varying focus locations and aperture sizes. This dataset is used to evaluate the model's refocusing ability." + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_caption" + } + ], + "index": 2 + }, + { + "bbox": [ + 55, + 485, + 193, + 497 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 485, + 193, + 497 + ], + "spans": [ + { + "bbox": [ + 55, + 485, + 193, + 497 + ], + "type": "text", + "content": "results with general input images." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 54, + 497, + 295, + 617 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 54, + 497, + 295, + 617 + ], + "spans": [ + { + "bbox": [ + 54, + 497, + 295, + 617 + ], + "type": "text", + "content": "Our Synthetic Dataset. As shown in Table 2, our method achieves superior refocusing ability through controlled DoF rendering, producing rendered images with the highest quality on our synthetic dataset. Furthermore, our method learns more precise aperture size and focus distance than DoF-NeRF [55]. Since Deblur-NeRF [26], BAGS [34], and Deblurring 3DGS [18] do not incorporate a lens model, they cannot generate novel view images with a specific DoF. As a result, these methods cannot be evaluated on this dataset due to their lack of refocusing capability." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 55, + 624, + 149, + 637 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 624, + 149, + 637 + ], + "spans": [ + { + "bbox": [ + 55, + 624, + 149, + 637 + ], + "type": "text", + "content": "5.3. Ablation Study" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 55, + 641, + 295, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 641, + 295, + 713 + ], + "spans": [ + { + "bbox": [ + 55, + 641, + 295, + 713 + ], + "type": "text", + "content": "As shown in Table 4 and Fig. 8, we conduct ablation studies to evaluate the effectiveness of our designs. The comparison between #1 and #2 indicates that the lens-based imaging model has a significant impact. This lens model not only facilitates defocus deblurring but also provides the ability to control the depth of field. Next, the defocus-to-focus" + } + ] + } + ], + "index": 7 + }, + { + "type": "table", + "bbox": [ + 317, + 314, + 552, + 387 + ], + "blocks": [ + { + "bbox": [ + 321, + 293, + 547, + 305 + ], + "lines": [ + { + "bbox": [ + 321, + 293, + 547, + 305 + ], + "spans": [ + { + "bbox": [ + 321, + 293, + 547, + 305 + ], + "type": "text", + "content": "Table 4. Ablation study on each component of our method." + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 317, + 314, + 552, + 387 + ], + "lines": [ + { + "bbox": [ + 317, + 314, + 552, + 387 + ], + "spans": [ + { + "bbox": [ + 317, + 314, + 552, + 387 + ], + "type": "table", + "html": "
LensDepthDefocus-to-focusPSNR↑SSIM↑LPIPS↓
#1XXX21.310.6360.239
#2✓XX23.050.7280.109
#3✓X✓23.590.7420.104
#4✓✓X23.420.7380.098
#5✓✓✓23.970.7560.093
", + "image_path": "0cb54d37a5e0aa6bffb5e515c8bb2f66ede818e373eab05e3971ea0fa343f02e.jpg" + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "table_body" + } + ], + "index": 9 + }, + { + "type": "table", + "bbox": [ + 328, + 432, + 538, + 491 + ], + "blocks": [ + { + "bbox": [ + 339, + 412, + 529, + 424 + ], + "lines": [ + { + "bbox": [ + 339, + 412, + 529, + 424 + ], + "spans": [ + { + "bbox": [ + 339, + 412, + 529, + 424 + ], + "type": "text", + "content": "Table 5. Ablation study on different depth priors." + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 328, + 432, + 538, + 491 + ], + "lines": [ + { + "bbox": [ + 328, + 432, + 538, + 491 + ], + "spans": [ + { + "bbox": [ + 328, + 432, + 538, + 491 + ], + "type": "table", + "html": "
MethodPSNR↑SSIM↑LPIPS↓
No fine-tuned depth23.530.7390.114
Sparse depth23.440.7370.113
Ours23.970.7560.093
", + "image_path": "4b8fc6f9677c389483b5541e2fa50d2536f9b7a9e56c55a2a34f7bca5913f49d.jpg" + } + ] + } + ], + "index": 11, + "angle": 0, + "type": "table_body" + } + ], + "index": 11 + }, + { + "bbox": [ + 313, + 525, + 555, + 586 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 525, + 555, + 586 + ], + "spans": [ + { + "bbox": [ + 313, + 525, + 555, + 586 + ], + "type": "text", + "content": "adaptation (#3) and the per-scene adjustment of depth priors (#4) further enhance the scene geometry and improve the deblurring of details. Ultimately, the combination of all components yields the highest performance gain, leading to a 2.66 dB increase in PSNR over the baseline model." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 313, + 594, + 556, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 594, + 556, + 715 + ], + "spans": [ + { + "bbox": [ + 313, + 594, + 556, + 715 + ], + "type": "text", + "content": "In addition, we conduct an ablation study on different depth supervision methods to demonstrate the superiority of our per-scene adjustment of depth priors, as shown in Table 5. Direct supervision using depth maps predicted by the depth network without scene-specific fine-tuning hinders the reconstruction quality of 3D-GS. Likewise, supervision based on sparse depth maps projected from a 3D point cloud fails to yield optimal results. We show visualizations of depth maps rendered by different depth supervision strategies in the supplementary material." + } + ] + } + ], + "index": 13 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 318, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 318, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 318, + 757 + ], + "type": "text", + "content": "26468" + } + ] + } + ], + "index": 14 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 6 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 57, + 83, + 553, + 150 + ], + "blocks": [ + { + "bbox": [ + 194, + 69, + 400, + 81 + ], + "lines": [ + { + "bbox": [ + 194, + 69, + 400, + 81 + ], + "spans": [ + { + "bbox": [ + 194, + 69, + 400, + 81 + ], + "type": "text", + "content": "(I) Change the aperture size and focus distance" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 57, + 83, + 553, + 150 + ], + "lines": [ + { + "bbox": [ + 57, + 83, + 553, + 150 + ], + "spans": [ + { + "bbox": [ + 57, + 83, + 553, + 150 + ], + "type": "image", + "image_path": "c5127cfbbf9def5d270a2dc66c2e97c9df308d7ea4658d695207b9d6918444b3.jpg" + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_body" + } + ], + "index": 1 + }, + { + "type": "image", + "bbox": [ + 57, + 165, + 553, + 233 + ], + "blocks": [ + { + "bbox": [ + 195, + 152, + 412, + 165 + ], + "lines": [ + { + "bbox": [ + 195, + 152, + 412, + 165 + ], + "spans": [ + { + "bbox": [ + 195, + 152, + 412, + 165 + ], + "type": "text", + "content": "(II) Change the lens parameters and camera poses" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 57, + 165, + 553, + 233 + ], + "lines": [ + { + "bbox": [ + 57, + 165, + 553, + 233 + ], + "spans": [ + { + "bbox": [ + 57, + 165, + 553, + 233 + ], + "type": "image", + "image_path": "99843575058854ca572730dbd9ce3f9674a04e3fb1f82637856ac1f99ed2c6ad.jpg" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_body" + } + ], + "index": 3 + }, + { + "type": "image", + "bbox": [ + 57, + 247, + 553, + 316 + ], + "blocks": [ + { + "bbox": [ + 195, + 235, + 393, + 247 + ], + "lines": [ + { + "bbox": [ + 195, + 235, + 393, + 247 + ], + "spans": [ + { + "bbox": [ + 195, + 235, + 393, + 247 + ], + "type": "text", + "content": "(III) Change the lens parameters and zoom-in" + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 57, + 247, + 553, + 316 + ], + "lines": [ + { + "bbox": [ + 57, + 247, + 553, + 316 + ], + "spans": [ + { + "bbox": [ + 57, + 247, + 553, + 316 + ], + "type": "image", + "image_path": "99c79dbb4aa2cd977795ab94a7f90038405c9f824840d0801d73de04568ba5a4.jpg" + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 55, + 324, + 555, + 346 + ], + "lines": [ + { + "bbox": [ + 55, + 324, + 555, + 346 + ], + "spans": [ + { + "bbox": [ + 55, + 324, + 555, + 346 + ], + "type": "text", + "content": "Figure 7. Applications of our controllable DoF effects. Users can create their own cinematic moments by combining changes in aperture size, focus distance, camera poses, and zoom." + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "image_caption" + } + ], + "index": 5 + }, + { + "type": "image", + "bbox": [ + 57, + 354, + 295, + 483 + ], + "blocks": [ + { + "bbox": [ + 57, + 354, + 295, + 483 + ], + "lines": [ + { + "bbox": [ + 57, + 354, + 295, + 483 + ], + "spans": [ + { + "bbox": [ + 57, + 354, + 295, + 483 + ], + "type": "image", + "image_path": "d8e23bdec491fec232de644336e6ff64568c5b75af21d1565cb6e86fa943f88e.jpg" + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 55, + 491, + 295, + 524 + ], + "lines": [ + { + "bbox": [ + 55, + 491, + 295, + 524 + ], + "spans": [ + { + "bbox": [ + 55, + 491, + 295, + 524 + ], + "type": "text", + "content": "Figure 8. The visualizations of the ablation study. Qualitative comparisons show that the full model (#5) yields the best result and the reconstructed strawberry is the sharpest." + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "image_caption" + } + ], + "index": 7 + }, + { + "bbox": [ + 55, + 540, + 190, + 553 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 540, + 190, + 553 + ], + "spans": [ + { + "bbox": [ + 55, + 540, + 190, + 553 + ], + "type": "text", + "content": "5.4. Interactive Applications" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 55, + 558, + 296, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 558, + 296, + 713 + ], + "spans": [ + { + "bbox": [ + 55, + 558, + 296, + 713 + ], + "type": "text", + "content": "Thanks to the lens-based model, we can not only represent 3D scenes from shallow DoF inputs but also render interactive DoF effects (see Fig. 7). Various cinematography styles can be achieved through different camera configurations, and users can input custom datasets, whether all-in-focus or bokeh images. By controlling aperture size and focus distance, we can render novel view images with different depths of field. In addition, we can modify the shape of the CoC, such as changing it from a circle to a hexagon. Users can also adjust the depth of field while moving the camera or zooming to create cinematic effects. Unlike [55], our GS framework significantly enhances the training and rendering efficiency, allowing for faster and more seamless inter" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 313, + 357, + 555, + 381 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 357, + 555, + 381 + ], + "spans": [ + { + "bbox": [ + 313, + 357, + 555, + 381 + ], + "type": "text", + "content": "actions. The details on processing time will be discussed in the supplementary material." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 313, + 393, + 388, + 406 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 393, + 388, + 406 + ], + "spans": [ + { + "bbox": [ + 313, + 393, + 388, + 406 + ], + "type": "text", + "content": "6. Conclusion" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 313, + 414, + 555, + 569 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 414, + 555, + 569 + ], + "spans": [ + { + "bbox": [ + 313, + 414, + 555, + 569 + ], + "type": "text", + "content": "While 3D Gaussian Splatting methods have achieved impressive performance on a wide range of reconstruction tasks, the applications on controllable DoF effects remain challenging and understudied. In this paper, we propose DoF-Gaussian, a controllable DoF method for 3D Gaussian Splatting. Specifically, we develop a lens-based model rather than pinhole imaging to overcome the limitations imposed by shallow DoF inputs for 3D-GS. Furthermore, we propose the per-scene depth adjustment and a defocus-to-focus adaptation to guarantee the performance of defocus deblurring. We also introduce a synthetic dataset for a more comprehensive evaluation. Thanks to the imaging principles, our method supports various interactive applications." + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 313, + 570, + 554, + 665 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 570, + 554, + 665 + ], + "spans": [ + { + "bbox": [ + 313, + 570, + 554, + 665 + ], + "type": "text", + "content": "Future work. Inspired by our experiments, we found that modeling real-world physical imaging principles can enable our method not only to handle shallow DoF inputs but also to perform effectively on general inputs. This insight motivates us to explore a combination of non-ideal conditions, such as shallow DoF inputs, sparse views, and varied lighting environments, which are more aligned with casual photography in our daily lives." + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 313, + 665, + 554, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 665, + 554, + 713 + ], + "spans": [ + { + "bbox": [ + 313, + 665, + 554, + 713 + ], + "type": "text", + "content": "Acknowledgement. This study is supported under the RIE2020 Industry Alignment Fund - Industry Collaboration Projects (IAF-ICP) Funding Initiative, as well as cash and in-kind contribution from the industry partner(s)." + } + ] + } + ], + "index": 15 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 318, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 318, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 318, + 757 + ], + "type": "text", + "content": "26469" + } + ] + } + ], + "index": 16 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 7 + }, + { + "para_blocks": [ + { + "bbox": [ + 56, + 71, + 115, + 83 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 71, + 115, + 83 + ], + "spans": [ + { + "bbox": [ + 56, + 71, + 115, + 83 + ], + "type": "text", + "content": "References" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 57, + 91, + 296, + 712 + ], + "type": "list", + "angle": 0, + "index": 16, + "blocks": [ + { + "bbox": [ + 61, + 91, + 296, + 133 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 61, + 91, + 296, + 133 + ], + "spans": [ + { + "bbox": [ + 61, + 91, + 296, + 133 + ], + "type": "text", + "content": "[1] Guillaume Abadie, Steve McAuley, Evegenii Golubev, Stephen Hill, and Sebastien Lagarde. Advances in real-time rendering in games. In ACM SIGGRAPH 2018 Courses. 2018. 2" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 61, + 137, + 296, + 190 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 61, + 137, + 296, + 190 + ], + "spans": [ + { + "bbox": [ + 61, + 137, + 296, + 190 + ], + "type": "text", + "content": "[2] Aleksei Bochkovskii, Amaël Delaunoy, Hugo Germain, Marcel Santos, Yichao Zhou, Stephan R Richter, and Vladlen Koltun. Depth pro: Sharp monocular metric depth in less than a second. arXiv preprint arXiv:2410.02073, 2024. 6" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 62, + 193, + 295, + 224 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 193, + 295, + 224 + ], + "spans": [ + { + "bbox": [ + 62, + 193, + 295, + 224 + ], + "type": "text", + "content": "[3] Benjamin Busam, Matthieu Hog, Steven McDonagh, and Gregory Slabaugh. Sterefo: Efficient image refocusing with stereo vision. In ICCVW, 2019. 4" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 62, + 227, + 294, + 258 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 227, + 294, + 258 + ], + "spans": [ + { + "bbox": [ + 62, + 227, + 294, + 258 + ], + "type": "text", + "content": "[4] Wenbo Chen and Ligang Liu. Deblur-GS: 3D Gaussian Splitting from Camera Motion Blurred Images. PACMCGIT, 7(1), 2024. 3" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 62, + 261, + 295, + 293 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 261, + 295, + 293 + ], + "spans": [ + { + "bbox": [ + 62, + 261, + 295, + 293 + ], + "type": "text", + "content": "[5] Jaeyoung Chung, Jeongtaek Oh, and Kyoung Mu Lee. Depth-regularized optimization for 3D Gaussian Splitting in few-shot images. In CVPR, 2024. 5" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 62, + 296, + 295, + 349 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 296, + 295, + 349 + ], + "spans": [ + { + "bbox": [ + 62, + 296, + 295, + 349 + ], + "type": "text", + "content": "[6] Marcos V Conde, Manuel Kolmet, Tim Seizinger, Tom E Bishop, Radu Timofte, Xiangyu Kong, Dafeng Zhang, Jinlong Wu, Fan Wang, Juewen Peng, et al. Lens-to-lens bokeh effect transformation. NTIRE 2023 challenge report. In CVPR, 2023. 2" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 62, + 352, + 295, + 384 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 352, + 295, + 384 + ], + "spans": [ + { + "bbox": [ + 62, + 352, + 295, + 384 + ], + "type": "text", + "content": "[7] François Darmon, Lorenzo Porzi, Samuel Rota-Bulò, and Peter Kontschieder. Robust gaussian splatting. arXiv preprint arXiv:2404.04211, 2024. 2, 3" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 62, + 386, + 295, + 429 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 386, + 295, + 429 + ], + "spans": [ + { + "bbox": [ + 62, + 386, + 295, + 429 + ], + "type": "text", + "content": "[8] Saikat Dutta, Sourya Dipta Das, Nisarg A Shah, and Anil Kumar Tiwari. Stacked deep multi-scale hierarchical network for fast bokeh effect rendering from a single image. In CVPR, 2021. 2" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 62, + 431, + 295, + 463 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 431, + 295, + 463 + ], + "spans": [ + { + "bbox": [ + 62, + 431, + 295, + 463 + ], + "type": "text", + "content": "[9] David Eigen, Christian Puhrsch, and Rob Fergus. Depth map prediction from a single image using a multi-scale deep network. NeurIPS, 27, 2014. 5" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 57, + 465, + 284, + 476 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 465, + 284, + 476 + ], + "spans": [ + { + "bbox": [ + 57, + 465, + 284, + 476 + ], + "type": "text", + "content": "[10] Eugene Hecht. Optics. Pearson Education India, 2012. 4" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 57, + 478, + 295, + 521 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 478, + 295, + 521 + ], + "spans": [ + { + "bbox": [ + 57, + 478, + 295, + 521 + ], + "type": "text", + "content": "[11] Binbin Huang, Zehao Yu, Anpei Chen, Andreas Geiger, and Shenghua Gao. 2D Gaussian Splatting for geometrically accurate radiance fields. In ACM SIGGRAPH 2024 Conference Papers, 2024. 2, 5" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 57, + 523, + 295, + 576 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 523, + 295, + 576 + ], + "spans": [ + { + "bbox": [ + 57, + 523, + 295, + 576 + ], + "type": "text", + "content": "[12] Andrey Ignatov, Jagruti Patel, Radu Timofte, Bolun Zheng, Xin Ye, Li Huang, Xiang Tian, Saikat Dutta, Kuldeep Purohit, Praveen Kandula, et al. Aim 2019 challenge on bokeh effect synthesis: Methods and results. In ICCVW. IEEE, 2019. 2" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 57, + 579, + 295, + 610 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 579, + 295, + 610 + ], + "spans": [ + { + "bbox": [ + 57, + 579, + 295, + 610 + ], + "type": "text", + "content": "[13] Andrey Ignatov, Jagruti Patel, and Radu Timofte. Rendering natural camera bokeh effect with deep learning. In CVPRW, 2020." + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 57, + 614, + 295, + 657 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 614, + 295, + 657 + ], + "spans": [ + { + "bbox": [ + 57, + 614, + 295, + 657 + ], + "type": "text", + "content": "[14] Andrey Ignatov, Radu Timofte, Ming Qian, Congyu Qiao, Jiamin Lin, Zhenyu Guo, Chenghua Li, Cong Leng, Jian Cheng, Juewen Peng, et al. AIM 2020 challenge on rendering realistic bokeh. In ECCVW. Springer, 2020. 2" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 57, + 658, + 295, + 712 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 658, + 295, + 712 + ], + "spans": [ + { + "bbox": [ + 57, + 658, + 295, + 712 + ], + "type": "text", + "content": "[15] Xin Jin, Pengyi Jiao, Zheng-Peng Duan, Xingchao Yang, Chun-Le Guo, Bo Ren, and Chongyi Li. Lighting Every Darkness with 3DGS: Fast Training and Real-Time Rendering for HDR View Synthesis. arXiv preprint arXiv:2406.06216, 2024.3" + } + ] + } + ], + "index": 15 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 316, + 73, + 553, + 712 + ], + "type": "list", + "angle": 0, + "index": 33, + "blocks": [ + { + "bbox": [ + 316, + 73, + 553, + 106 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 73, + 553, + 106 + ], + "spans": [ + { + "bbox": [ + 316, + 73, + 553, + 106 + ], + "type": "text", + "content": "[16] Bernhard Kerbl, Georgios Kopanas, Thomas Leimkuhler, and George Drettakis. 3D Gaussian Splatting for Real-Time Radiance Field Rendering. ACM TOG, 42(4), 2023. 2, 6" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 317, + 107, + 553, + 138 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 107, + 553, + 138 + ], + "spans": [ + { + "bbox": [ + 317, + 107, + 553, + 138 + ], + "type": "text", + "content": "[17] Min-Jung Kim, Gyojung Gu, and Jaegul Choo. LensNeRF: Rethinking volume rendering based on thin-lens camera model. In WACV, 2024. 2, 3" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 317, + 140, + 553, + 172 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 140, + 553, + 172 + ], + "spans": [ + { + "bbox": [ + 317, + 140, + 553, + 172 + ], + "type": "text", + "content": "[18] Byeonghyeon Lee, Howoong Lee, Xiangyu Sun, Usman Ali, and Eunbyung Park. Deblurring 3D Gaussian Splatting. arXiv preprint arXiv:2401.00834, 2024. 2, 3, 6, 7" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 317, + 174, + 553, + 205 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 174, + 553, + 205 + ], + "spans": [ + { + "bbox": [ + 317, + 174, + 553, + 205 + ], + "type": "text", + "content": "[19] Dogyoon Lee, Minhyeok Lee, Chajin Shin, and Sangyoun Lee. Dp-NeRF: Deblurred neural radiance field with physical scene priors. In CVPR, 2023. 2, 3, 6" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 317, + 206, + 553, + 237 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 206, + 553, + 237 + ], + "spans": [ + { + "bbox": [ + 317, + 206, + 553, + 237 + ], + "type": "text", + "content": "[20] Sungkil Lee, Elmar Eisemann, and Hans-Peter Seidel. Realtime lens blur effects and focus control. ACM TOG, 29(4), 2010. 2" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 316, + 239, + 553, + 281 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 239, + 553, + 281 + ], + "spans": [ + { + "bbox": [ + 316, + 239, + 553, + 281 + ], + "type": "text", + "content": "[21] Jiahe Li, Jiawei Zhang, Xiao Bai, Jin Zheng, Xin Ning, Jun Zhou, and Lin Gu. DNGaussian: Optimizing sparse-view 3D Gaussian radiance fields with global-local depth normalization. In CVPR, 2024. 5" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 316, + 284, + 553, + 326 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 284, + 553, + 326 + ], + "spans": [ + { + "bbox": [ + 316, + 284, + 553, + 326 + ], + "type": "text", + "content": "[22] Zhengqi Li, Tali Dekel, Forrester Cole, Richard Tucker, Noah Snavely, Ce Liu, and William T Freeman. Learning the depths of moving people by watching frozen people. In CVPR, 2019. 5" + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 316, + 327, + 553, + 380 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 327, + 553, + 380 + ], + "spans": [ + { + "bbox": [ + 316, + 327, + 553, + 380 + ], + "type": "text", + "content": "[23] Tianqi Liu, Guangcong Wang, Shoukang Hu, Liao Shen, Xinyi Ye, Yuhang Zang, Zhiguo Cao, Wei Li, and Ziwei Liu. MVSGaussian: Fast Generalizable Gaussian Splatting Reconstruction from Multi-View Stereo. In ECCV. Springer, 2025. 2" + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 316, + 383, + 553, + 415 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 383, + 553, + 415 + ], + "spans": [ + { + "bbox": [ + 316, + 383, + 553, + 415 + ], + "type": "text", + "content": "[24] Xianrui Luo, Huiqiang Sun, Juewen Peng, and Zhiguo Cao. Dynamic Neural Radiance Field From Defocused Monocular Video. arXiv preprint arXiv:2407.05586, 2024. 2" + } + ] + } + ], + "index": 25 + }, + { + "bbox": [ + 316, + 416, + 553, + 448 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 416, + 553, + 448 + ], + "spans": [ + { + "bbox": [ + 316, + 416, + 553, + 448 + ], + "type": "text", + "content": "[25] Yawen Luo, Min Shi, Liao Shen, Yachuan Huang, Zixuan Ye, Juewen Peng, and Zhiguo Cao. Video bokeh rendering: Make casual videography cinematic. In ACM MM, 2024. 2" + } + ] + } + ], + "index": 26 + }, + { + "bbox": [ + 316, + 449, + 553, + 481 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 449, + 553, + 481 + ], + "spans": [ + { + "bbox": [ + 316, + 449, + 553, + 481 + ], + "type": "text", + "content": "[26] Li Ma, Xiaoyu Li, Jing Liao, Qi Zhang, Xuan Wang, Jue Wang, and Pedro V Sander. Deblur-NeRF: Neural radiance fields from blurry images. In CVPR, 2022. 2, 3, 6, 7" + } + ] + } + ], + "index": 27 + }, + { + "bbox": [ + 317, + 483, + 553, + 525 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 483, + 553, + 525 + ], + "spans": [ + { + "bbox": [ + 317, + 483, + 553, + 525 + ], + "type": "text", + "content": "[27] Ricardo Martin-Brualla, Noha Radwan, Mehdi SM Sajjadi, Jonathan T Barron, Alexey Dosovitskiy, and Daniel Duckworth. NeRF in the wild: Neural radiance fields for unconstrained photo collections. In CVPR, 2021. 3" + } + ] + } + ], + "index": 28 + }, + { + "bbox": [ + 317, + 526, + 553, + 570 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 526, + 553, + 570 + ], + "spans": [ + { + "bbox": [ + 317, + 526, + 553, + 570 + ], + "type": "text", + "content": "[28] Ben Mildenhall, Pratul P Srinivasan, Matthew Tancik, Jonathan T Barron, Ravi Ramamoorthi, and Ren Ng. NeRF: Representing scenes as neural radiance fields for view synthesis. Communications of the ACM, 65(1), 2021. 2, 6" + } + ] + } + ], + "index": 29 + }, + { + "bbox": [ + 317, + 571, + 553, + 613 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 571, + 553, + 613 + ], + "spans": [ + { + "bbox": [ + 317, + 571, + 553, + 613 + ], + "type": "text", + "content": "[29] Ben Mildenhall, Peter Hedman, Ricardo Martin-Brualla, Pratul P Srinivasan, and Jonathan T Barron. NeRF in the dark: High dynamic range view synthesis from noisy raw images. In CVPR, 2022. 3" + } + ] + } + ], + "index": 30 + }, + { + "bbox": [ + 317, + 615, + 553, + 668 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 615, + 553, + 668 + ], + "spans": [ + { + "bbox": [ + 317, + 615, + 553, + 668 + ], + "type": "text", + "content": "[30] Nicolas Moenne-Loccoz, Ashkan Mirzaei, Or Perel, Ricardo de Lutio, Janick Martinez Esturo, Gavriel State, Sanja Fidler, Nicholas Sharp, and Zan Gojcic. 3d gaussian ray tracing: Fast tracing of particle scenes. ACM Transactions on Graphics (TOG), 43(6):1-19, 2024. 3" + } + ] + } + ], + "index": 31 + }, + { + "bbox": [ + 316, + 670, + 553, + 712 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 670, + 553, + 712 + ], + "spans": [ + { + "bbox": [ + 316, + 670, + 553, + 712 + ], + "type": "text", + "content": "[31] Keunhong Park, Utkarsh Sinha, Jonathan T Barron, Sofien Bouaziz, Dan B Goldman, Steven M Seitz, and Ricardo Martin-Brualla. Nerfies: Deformable neural radiance fields. In ICCV, 2021. 3" + } + ] + } + ], + "index": 32 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 318, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 318, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 318, + 757 + ], + "type": "text", + "content": "26470" + } + ] + } + ], + "index": 34 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 8 + }, + { + "para_blocks": [ + { + "bbox": [ + 55, + 73, + 296, + 713 + ], + "type": "list", + "angle": 0, + "index": 15, + "blocks": [ + { + "bbox": [ + 56, + 73, + 294, + 95 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 73, + 294, + 95 + ], + "spans": [ + { + "bbox": [ + 56, + 73, + 294, + 95 + ], + "type": "text", + "content": "[32] Frank L Pedrotti, Leno M Pedrotti, and Leno S Pedrotti. Introduction to optics. Cambridge University Press, 2018. 4" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 55, + 96, + 295, + 129 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 96, + 295, + 129 + ], + "spans": [ + { + "bbox": [ + 55, + 96, + 295, + 129 + ], + "type": "text", + "content": "[33] Cheng Peng and Rama Chellappa. PDRF: progressively deblurring radiance field for fast scene reconstruction from blurry images. In AAAI, 2023. 2, 3" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 57, + 130, + 295, + 174 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 130, + 295, + 174 + ], + "spans": [ + { + "bbox": [ + 57, + 130, + 295, + 174 + ], + "type": "text", + "content": "[34] Cheng Peng, Yutao Tang, Yifan Zhou, Nengyu Wang, Xijun Liu, Deming Li, and Rama Chellappa. BAGS: Blur Agnostic Gaussian Splatting through Multi-Scale Kernel Modeling. arXiv preprint arXiv:2403.04926, 2024. 2, 3, 5, 6, 7" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 57, + 175, + 295, + 207 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 175, + 295, + 207 + ], + "spans": [ + { + "bbox": [ + 57, + 175, + 295, + 207 + ], + "type": "text", + "content": "[35] Juewen Peng, Zhiguo Cao, Xianrui Luo, Hao Lu, Ke Xian, and Jianming Zhang. Bokehme: When neural rendering meets classical rendering. In CVPR, 2022. 2, 6, 7" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 57, + 209, + 296, + 263 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 209, + 296, + 263 + ], + "spans": [ + { + "bbox": [ + 57, + 209, + 296, + 263 + ], + "type": "text", + "content": "[36] Juewen Peng, Zhiyu Pan, Chengxin Liu, Xianrui Luo, Huiqiang Sun, Liao Shen, Ke Xian, and Zhiguo Cao. Selective bokeh effect transformation. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 1456-1464, 2023." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 57, + 265, + 295, + 319 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 265, + 295, + 319 + ], + "spans": [ + { + "bbox": [ + 57, + 265, + 295, + 319 + ], + "type": "text", + "content": "[37] Dominique Piché-Meunier, Yannick Hold-Geoffroy, Jianming Zhang, and Jean-François Lalonde. Lens parameter estimation for realistic depth of field modeling. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 499–508, 2023. 2" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 57, + 321, + 295, + 365 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 321, + 295, + 365 + ], + "spans": [ + { + "bbox": [ + 57, + 321, + 295, + 365 + ], + "type": "text", + "content": "[38] Ming Qian, Congyu Qiao, Jiamin Lin, Zhenyu Guo, Chenghua Li, Cong Leng, and Jian Cheng. BgGAN: Bokehglass generative adversarial network for rendering realistic bokeh. In ECCVW. Springer, 2020. 2" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 57, + 366, + 295, + 409 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 366, + 295, + 409 + ], + "spans": [ + { + "bbox": [ + 57, + 366, + 295, + 409 + ], + "type": "text", + "content": "[39] René Ranftl, Katrin Lasinger, David Hafner, Konrad Schindler, and Vladlen Koltun. Towards robust monocular depth estimation: Mixing datasets for zero-shot cross-dataset transfer. IEEE TPAMI, 44(3), 2020. 5" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 57, + 411, + 295, + 432 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 411, + 295, + 432 + ], + "spans": [ + { + "bbox": [ + 57, + 411, + 295, + 432 + ], + "type": "text", + "content": "[40] Johannes L Schonberger and Jan-Michael Frahm. Structure-from-motion revisited. In CVPR, 2016. 5" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 57, + 434, + 295, + 466 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 434, + 295, + 466 + ], + "spans": [ + { + "bbox": [ + 57, + 434, + 295, + 466 + ], + "type": "text", + "content": "[41] Johannes L. Schonberger, Enliang Zheng, Jan-Michael Frahm, and Marc Pollefeys. Pixelwise view selection for unstructured multi-view stereo. In ECCV. Springer, 2016. 5" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 57, + 468, + 295, + 522 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 468, + 295, + 522 + ], + "spans": [ + { + "bbox": [ + 57, + 468, + 295, + 522 + ], + "type": "text", + "content": "[42] Liao Shen, Xingyi Li, Huiqiang Sun, Juewen Peng, Ke Xian, Zhiguo Cao, and Guosheng Lin. Make-it-4d: Synthesizing a consistent long-term dynamic scene video from a single image. In Proceedings of the 31st ACM International Conference on Multimedia, pages 8167–8175, 2023. 3" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 57, + 524, + 295, + 567 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 524, + 295, + 567 + ], + "spans": [ + { + "bbox": [ + 57, + 524, + 295, + 567 + ], + "type": "text", + "content": "[43] Yichen Sheng, Zixun Yu, Lu Ling, Zhiwen Cao, Xuaner Zhang, Xin Lu, Ke Xian, Haiting Lin, and Bedrich Benes. Dr. Bokeh: DiffeRentiable Occlusion-aware Bokeh Rendering. In CVPR, 2024. 2" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 57, + 569, + 295, + 601 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 569, + 295, + 601 + ], + "spans": [ + { + "bbox": [ + 57, + 569, + 295, + 601 + ], + "type": "text", + "content": "[44] Huiqiang Sun, Xingyi Li, Liao Shen, Xinyi Ye, Ke Xian, and Zhiguo Cao. DyBluRF: Dynamic Neural Radiance Fields from Blurry Monocular Video. In CVPR, 2024. 2, 3" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 57, + 603, + 295, + 656 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 603, + 295, + 656 + ], + "spans": [ + { + "bbox": [ + 57, + 603, + 295, + 656 + ], + "type": "text", + "content": "[45] Neal Wadhwa, Rahul Garg, David E Jacobs, Bryan E Feldman, Nori Kanazawa, Robert Carroll, Yair Movshovitz-Attias, Jonathan T Barron, Yael Pritch, and Marc Levoy. Synthetic depth-of-field with a single-camera mobile phone. ACM TOG, 37(4), 2018. 2" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 57, + 658, + 295, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 658, + 295, + 713 + ], + "spans": [ + { + "bbox": [ + 57, + 658, + 295, + 713 + ], + "type": "text", + "content": "[46] Chao Wang, Krzysztof Wolski, Xingang Pan, Thomas Leimkuhler, Bin Chen, Christian Theobalt, Karol Myszkowski, Hans-Peter Seidel, and Ana Serrano. An implicit neural representation for the image stack: Depth, all in focus, and high dynamic range. Technical report, 2023. 2" + } + ] + } + ], + "index": 14 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 316, + 73, + 553, + 713 + ], + "type": "list", + "angle": 0, + "index": 32, + "blocks": [ + { + "bbox": [ + 316, + 73, + 553, + 127 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 73, + 553, + 127 + ], + "spans": [ + { + "bbox": [ + 316, + 73, + 553, + 127 + ], + "type": "text", + "content": "[47] Chao Wang, Krzysztof Wolski, Bernhard Kerbl, Ana Serrano, Mojtaba Bemana, Hans-Peter Seidel, Karol Myszkowski, and Thomas Leimkuhler. Cinematic Gaussians: Real-Time HDR Radiance Fields with Depth of Field. arXiv preprint arXiv:2406.07329, 2024. 3" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 317, + 129, + 553, + 172 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 129, + 553, + 172 + ], + "spans": [ + { + "bbox": [ + 317, + 129, + 553, + 172 + ], + "type": "text", + "content": "[48] Lijun Wang, Xiaohui Shen, Jianming Zhang, Oliver Wang, Zhe Lin, Chih-Yao Hsieh, Sarah Kong, and Huchuan Lu. Deep learning: Shallow depth of field from a single image. arXiv preprint arXiv:1810.08100, 2018. 2" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 317, + 173, + 553, + 205 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 173, + 553, + 205 + ], + "spans": [ + { + "bbox": [ + 317, + 173, + 553, + 205 + ], + "type": "text", + "content": "[49] Peng Wang, Lingzhe Zhao, Ruijie Ma, and Peidong Liu. BAD-NeRF: Bundle adjusted deblur neural radiance fields. In CVPR, 2023. 3" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 316, + 206, + 553, + 239 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 206, + 553, + 239 + ], + "spans": [ + { + "bbox": [ + 316, + 206, + 553, + 239 + ], + "type": "text", + "content": "[50] Yinhuai Wang, Shuzhou Yang, Yujie Hu, and Jian Zhang. NeRFocus: Neural radiance field for 3D synthetic defocus. arXiv preprint arXiv:2203.05189, 2022. 3" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 316, + 240, + 553, + 282 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 240, + 553, + 282 + ], + "spans": [ + { + "bbox": [ + 316, + 240, + 553, + 282 + ], + "type": "text", + "content": "[51] Yujie Wang, Praneeth Chakravarthula, and Baoquan Chen. DOF-GS: Adjustable Depth-of-Field 3D Gaussian Splitting for Refocusing, Defocus Rendering and Blur Removal. arXiv preprint arXiv:2405.17351, 2024. 3, 6" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 316, + 284, + 553, + 316 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 284, + 553, + 316 + ], + "spans": [ + { + "bbox": [ + 316, + 284, + 553, + 316 + ], + "type": "text", + "content": "[52] Zhou Wang, Alan C Bovik, Hamid R Sheikh, and Eero P Simoncelli. Image quality assessment: from error visibility to structural similarity. IEEE TIP, 13(4), 2004. 6" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 316, + 317, + 553, + 358 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 317, + 553, + 358 + ], + "spans": [ + { + "bbox": [ + 316, + 317, + 553, + 358 + ], + "type": "text", + "content": "[53] Yi Wei, Shaohui Liu, Yongming Rao, Wang Zhao, Jiwen Lu, and Jie Zhou. Nerfingmvs: Guided optimization of neural radiance fields for indoor multi-view stereo. In ICCV, 2021. 5" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 316, + 361, + 553, + 403 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 361, + 553, + 403 + ], + "spans": [ + { + "bbox": [ + 316, + 361, + 553, + 403 + ], + "type": "text", + "content": "[54] Guanjun Wu, Taoran Yi, Jiemin Fang, Lingxi Xie, Xiaopeng Zhang, Wei Wei, Wenyu Liu, Qi Tian, and Xinggang Wang. 4d gaussian splatting for real-time dynamic scene rendering. In CVPR, 2024. 2, 3" + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 316, + 404, + 553, + 436 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 404, + 553, + 436 + ], + "spans": [ + { + "bbox": [ + 316, + 404, + 553, + 436 + ], + "type": "text", + "content": "[55] Zijin Wu, Xingyi Li, Juewen Peng, Hao Lu, Zhiguo Cao, and Weicai Zhong. DoF-NeRF: Depth-of-field meets neural radiance fields. In ACM MM, 2022. 1, 2, 3, 6, 7, 8" + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 316, + 438, + 553, + 480 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 438, + 553, + 480 + ], + "spans": [ + { + "bbox": [ + 316, + 438, + 553, + 480 + ], + "type": "text", + "content": "[56] Lei Xiao, Anton Kaplanyan, Alexander Fix, Matt Chapman, and Douglas Lanman. Deepfocus: Learned image synthesis for computational display. In ACM SIGGRAPH 2018 Talks. 2018. 2" + } + ] + } + ], + "index": 25 + }, + { + "bbox": [ + 316, + 482, + 553, + 525 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 482, + 553, + 525 + ], + "spans": [ + { + "bbox": [ + 316, + 482, + 553, + 525 + ], + "type": "text", + "content": "[57] Haofei Xu, Songyou Peng, Fangjinhua Wang, Hermann Blum, Daniel Barath, Andreas Geiger, and Marc Pollefeys. Depthsplat: Connecting gaussian splatting and depth. arXiv preprint arXiv:2410.13862, 2024. 2" + } + ] + } + ], + "index": 26 + }, + { + "bbox": [ + 316, + 526, + 553, + 558 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 526, + 553, + 558 + ], + "spans": [ + { + "bbox": [ + 316, + 526, + 553, + 558 + ], + "type": "text", + "content": "[58] Zehao Yu, Anpei Chen, Binbin Huang, Torsten Sattler, and Andreas Geiger. Mip-splatting: Alias-free 3D gaussian splattering. In CVPR, 2024. 2, 5, 6" + } + ] + } + ], + "index": 27 + }, + { + "bbox": [ + 316, + 559, + 553, + 601 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 559, + 553, + 601 + ], + "spans": [ + { + "bbox": [ + 316, + 559, + 553, + 601 + ], + "type": "text", + "content": "[59] Baowen Zhang, Chuan Fang, Rakesh Shrestha, Yixun Liang, Xiaoxiao Long, and Ping Tan. RaDe-GS: Rasterizing Depth in Gaussian Splatting. arXiv preprint arXiv:2406.01467, 2024.5" + } + ] + } + ], + "index": 28 + }, + { + "bbox": [ + 316, + 603, + 553, + 635 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 603, + 553, + 635 + ], + "spans": [ + { + "bbox": [ + 316, + 603, + 553, + 635 + ], + "type": "text", + "content": "[60] Richard Zhang, Phillip Isola, Alexei A Efros, Eli Shechtman, and Oliver Wang. The unreasonable effectiveness of deep features as a perceptual metric. In CVPR, 2018. 6" + } + ] + } + ], + "index": 29 + }, + { + "bbox": [ + 316, + 636, + 553, + 679 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 636, + 553, + 679 + ], + "spans": [ + { + "bbox": [ + 316, + 636, + 553, + 679 + ], + "type": "text", + "content": "[61] Xuaner Zhang, Kevin Matzen, Vivien Nguyen, Dillon Yao, You Zhang, and Ren Ng. Synthetic defocus and lookahead autofocus for casual videography. arXiv preprint arXiv:1905.06326, 2019. 2" + } + ] + } + ], + "index": 30 + }, + { + "bbox": [ + 316, + 681, + 553, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 681, + 553, + 713 + ], + "spans": [ + { + "bbox": [ + 316, + 681, + 553, + 713 + ], + "type": "text", + "content": "[62] Lingzhe Zhao, Peng Wang, and Peidong Liu. BAD-Gaussians: Bundle adjusted deblur Gaussian Splatting. arXiv preprint arXiv:2403.11831, 2024. 3" + } + ] + } + ], + "index": 31 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "text", + "content": "26471" + } + ] + } + ], + "index": 33 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 9 + } + ], + "_backend": "vlm", + "_version_name": "2.6.4" +} \ No newline at end of file diff --git a/2025/DocLayLLM_ An Efficient Multi-modal Extension of Large Language Models for Text-rich Document Understanding/587c40bc-a1fe-44db-8881-11814790688b_content_list.json b/2025/DocLayLLM_ An Efficient Multi-modal Extension of Large Language Models for Text-rich Document Understanding/587c40bc-a1fe-44db-8881-11814790688b_content_list.json new file mode 100644 index 0000000000000000000000000000000000000000..d50aab50a9b6c63b9c4f0335bb45b27340b3059f --- /dev/null +++ b/2025/DocLayLLM_ An Efficient Multi-modal Extension of Large Language Models for Text-rich Document Understanding/587c40bc-a1fe-44db-8881-11814790688b_content_list.json @@ -0,0 +1,1613 @@ +[ + { + "type": "text", + "text": "DocLayLLM: An Efficient Multi-modal Extension of Large Language Models for Text-rich Document Understanding", + "text_level": 1, + "bbox": [ + 107, + 128, + 890, + 176 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Wenhui Liao $^{1\\dagger}$ , Jiapeng Wang $^{1,2\\dagger}$ , Hongliang Li $^{1}$ , Chengyu Wang $^{2*}$ , Jun Huang $^{2}$ , Lianwen Jin $^{1,3*}$ $^{1}$ South China University of Technology, Guangzhou, China \n $^{2}$ Alibaba Cloud Computing, Hangzhou, China \n $^{3}$ SCUT-Zhuhai Institute of Modern Industrial Innovation, Zhuhai, China", + "bbox": [ + 116, + 200, + 879, + 273 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "{eelwh, eejpwang, eehongliangli}@mail.scut.edu.cn, eelwjin@scut.edu.cn {chengyu.wcy, huangjun.hj}@alibaba-inc.com", + "bbox": [ + 189, + 276, + 807, + 309 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Abstract", + "text_level": 1, + "bbox": [ + 246, + 344, + 326, + 359 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Text-rich document understanding (TDU) requires comprehensive analysis of documents containing substantial textual content and complex layouts. While Multimodal Large Language Models (MLLMs) have achieved fast progress in this domain, existing approaches either demand significant computational resources or struggle with effective multi-modal integration. In this paper, we introduce DocLayLLM, an efficient multi-modal extension of LLMs specifically designed for TDU. By lightly integrating visual patch tokens and 2D positional tokens into LLMs' input and encoding the document content using the LLMs themselves, we fully take advantage of the document comprehension capability of LLMs and enhance their perception of OCR information. We have also deeply considered the role of chain-of-thought (CoT) and innovatively proposed the techniques of CoT Pre-training and CoT Annealing. Our DocLayLLM can achieve remarkable performances with lightweight training settings, showcasing its efficiency and effectiveness. Experimental results demonstrate that our DocLayLLM outperforms existing OCR-dependent methods and OCR-free competitors. Code and model are available at https://github.com/whlscut/DocLayLLM.", + "bbox": [ + 89, + 375, + 485, + 708 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "1. Introduction", + "text_level": 1, + "bbox": [ + 89, + 734, + 220, + 750 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Text-rich document understanding (TDU) focuses on analyzing documents with extensive text and complex layouts, requiring a combined understanding of text, layout, and images. Early research [1, 2, 12, 15, 20, 28, 33-37, 44, 52, 55, 65, 68, 69, 72, 73, 77] often used BERT-like [8] Transformer-based document encoders (DE), designing various pre-training tasks for large-scale unlabeled", + "bbox": [ + 89, + 761, + 483, + 867 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "document corpora and fine-tuning their models on specific downstream tasks like key information extraction (KIE) and visual question answering (VQA) [39, 40]. However, these pre-training tasks often diverge significantly from downstream tasks, requiring the model to acquire substantial task-specific knowledge during fine-tuning. This may result in limited generalization capability across different tasks and datasets.", + "bbox": [ + 511, + 345, + 905, + 465 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "With the rapid development of large language models (LLMs), their impressive versatility and generalization capabilities have captured widespread interest [74]. This has led many researchers [17, 38, 61, 76, 79] to overcome the challenges of TDU with LLMs. Initially, efforts focused on employing multi-modal large language models (MLLMs) for OCR-free TDU, where document images are fed into MLLMs, directly generating responses to comprehension tasks. These MLLMs, which can process images and text inputs, are among the most recent advancements in this domain. To date, several general MLLMs have been developed to manage diverse scenarios, and researchers often build their TDU models based on these general models. However, these TDU models typically require high-resolution image inputs to recognize dense text in documents and a large volume of training data to effectively develop this capability, both of which bring significant computational demands. Furthermore, within the extensive collections of image and instruction data used to train these models, the training data for downstream evaluation tasks was often included, which might potentially impact the assessment of their generalization to previously unseen datasets.", + "bbox": [ + 511, + 470, + 906, + 804 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "To overcome the constraints of OCR-free methods, scholars have pursued diverse methodologies for integrating OCR information into LLMs. Initial strategies [14, 70] involved inputting coordinates of OCR bounding boxes as numeric text. However, this textual representation of numbers often resulted in excessively long inputs that could", + "bbox": [ + 511, + 809, + 908, + 902 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "CVF", + "bbox": [ + 106, + 2, + 181, + 42 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore.", + "bbox": [ + 236, + 0, + 810, + 46 + ], + "page_idx": 0 + }, + { + "type": "page_footnote", + "text": "$\\dagger$ Equal contribution.", + "bbox": [ + 107, + 875, + 222, + 887 + ], + "page_idx": 0 + }, + { + "type": "page_footnote", + "text": "\\*Co-corresponding authors.", + "bbox": [ + 109, + 887, + 256, + 898 + ], + "page_idx": 0 + }, + { + "type": "page_number", + "text": "4038", + "bbox": [ + 482, + 944, + 514, + 955 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "decrease inference speeds and potentially degrade model performance. In response, many researchers [11, 24, 45] have adopted an auxiliary document encoder like LayoutLMv3 [20] to process textual, visual, and layout features from documents and integrated into LLMs. Although these methods avoid high-resolution image inputs, the textual content of documents must also be encoded alongside the layout and image features within the DE due to its structural limitations. However, the document comprehension capabilities of these DEs may not be on par with those of LLMs themselves, which might have led to significant finetuning efforts for aligning the feature spaces of DEs and LLMs to improve performances during the whole training phase, demanding substantial computational resources.", + "bbox": [ + 89, + 90, + 480, + 303 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "To address the limitations of OCR-dependent TDU methods, we introduce DocLayLLM, an efficient multimodal extension of large language models for text-rich document understanding. Our method diverges from existing approaches that utilize an extra DE by integrating the document's image patch tokens and 2D positional tokens with textual content into a natural language expression, and then encoding them through a unified LLM. This innovative integration not only taps into the LLM's inherent text comprehension abilities but also allows it to perceive document layouts. Besides, by incorporating patch tokens, the model supports image-related pre-training tasks and further gains a rudimentary understanding of document structures.", + "bbox": [ + 89, + 306, + 480, + 503 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "To advance model performance, we have fully integrated the concept of the chain-of-thought (CoT) [3] into each stage of our training process. Firstly, we developed a novel CoT Pre-training strategy. CoT is engineered to preserve thematic coherence and logical progression during model reasoning, further enhancing the inference ability of LLM, which has proven effectiveness in various subsequent studies [14, 45, 71]. Recently, the use of Long CoT reasoning to enhance the inference capability of models has attracted significant attention [7, 49, 63], and DeepSeek-R1 [7] further emphasized the importance of utilizing CoT data for a cold start before the reinforcement learning stage. Our proposed CoT Pre-training offers a method for rapidly generating large-scale CoT data from non-QA data, thereby improving the model's reasoning ability. By employing CoT in the data generation process and subsequently pre-training the model with the generated QA-formatted data, we have aligned the added visual and layout features with LLM's inherent textual features while improving the model's comprehension ability. Furthermore, we introduced a CoT Annealing technique. While previous research [31, 42] has validated the effectiveness of annealing strategies that gradually increase the proportion of high-quality data towards the end of training, the impact of CoT on data quality had not been explored. To address this, we first propose an annealing strategy that considers data quality from a CoT per", + "bbox": [ + 89, + 508, + 482, + 901 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "spective, further improving model performance.", + "bbox": [ + 511, + 92, + 830, + 106 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Thanks to the above design, utilizing only around 3M pre-training data and 300K supervised fine-tuning (SFT) data with the efficient tuning technology of LoRA [18], our DocLayLLM finishes the whole training process in 36 A100 days, therefore resulting an efficient extension of LLM. Even with such low resources, our DocLayLLM surpasses existing OCR-dependent TDU methods. Furthermore, it also excels in specific tasks without fine-tuning on downstream tasks and outperforms state-of-the-art (SOTA) OCR-free methods when trained with training sets of downstream tasks as other methods, further showcasing its effectiveness. Our contributions are:", + "bbox": [ + 511, + 107, + 903, + 286 + ], + "page_idx": 1 + }, + { + "type": "list", + "sub_type": "text", + "list_items": [ + "- We propose an efficient multi-modal extension of LLMs, which has augmented LLM in understanding text-rich documents and greatly reduced training resources in need.", + "- We introduce CoT Pre-training, a novel approach for generating CoT data that helps enhance the comprehension capabilities of models.", + "- We revisit data annealing from a novel perspective and introduce the CoT Annealing strategy, which further enhances the model's efficiency in data utilization and boosts its ultimate performance.", + "- Our method outperforms existing OCR-dependent TDU solutions and surpasses current OCR-free methods under comparable supervised fine-tuning conditions." + ], + "bbox": [ + 513, + 289, + 903, + 484 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "2. Related Works", + "text_level": 1, + "bbox": [ + 513, + 497, + 661, + 513 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "OCR-free MLLMs for Text-Rich Documents Understanding. To date, a number of studies have attempted to directly answer the instruction for TDU tasks from document images, thereby obviating the need for extra OCR processing and resulting in an end-to-end paradigm. Such models are referred to as OCR-free methods. Early examples like Donut [23], UDOP [62], and OmniParser [66] have explored end-to-end document understanding solutions without relying on LLMs or MLLMs. However, with the rapid development of MLLMs, many new studies have begun to focus on enhancing the capabilities of general MLLMs in document understanding. Some of them attempt to boost the document understanding ability of general MLLMs by tuning them with synthesizing TDU data generated by proprietary LLMs, including LLaVAR [79] and TextSquare [61], while mPLUG-DocOwl 1.5 [17] enhances the document understanding capability by integrating existing document comprehension data and converting it into vision-related QA-formatted data. Moreover, researchers have also explored image encoding methods to read the text more clearly. For instance, UReader [76] and Monkey [38] empower MLLMs to handle high-resolution images by splitting them into sub-images, encoding each separately, and then concatenating them. Some general MLLMs, such as LLaVA-Next [29], LLaVA", + "bbox": [ + 511, + 523, + 903, + 902 + ], + "page_idx": 1 + }, + { + "type": "page_number", + "text": "4039", + "bbox": [ + 482, + 944, + 514, + 955 + ], + "page_idx": 1 + }, + { + "type": "image", + "img_path": "images/67a0a691f89e141767f2dd8180cbecc54cfa227c8d51c9dff7b529145e47dcba.jpg", + "image_caption": [ + "Figure 1. The overall architecture of DocLayLLM." + ], + "image_footnote": [], + "bbox": [ + 142, + 85, + 854, + 329 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "OneVision [27], and Idefics3 [26], also adopt this approach to address the challenge of TDU. Although these above methods have achieved notable success in TDU tasks, they typically require high-resolution image inputs to identify text within images, significantly increasing computational demands. Besides, to equip general MLLMs with the ability to recognize the rich text within a document, a substantial amount of training data is required, including training sets of the benchmarks tested, which resulted in a high demand for training resources and might affect the evaluation of their zero-shot inference ability.", + "bbox": [ + 88, + 378, + 482, + 546 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "OCR-dependent MLLMs for Text-Rich Documents Understanding. To address the challenges of OCR-free MLLMs, some researchers have attempted to integrate OCR information from text-rich documents into LLMs to avoid high-resolution image inputs. ICL-D3IE [14] feeds coordinates of OCR boxes into the LLM using a textual expression of [left, top, right, bottom], while LATIN-Prompt [70] simulates the document layout by inserting spaces and newline characters between OCR-recognized text segments, thus providing the LLM with a document input containing a rudimentary layout. Besides, LMDX [53] also input the layout in a way similar to ICL-D3IE. However, these methods of inputting OCR information in textual form lead to overly lengthy inputs, thereby slowing inference and impacting model performance. Subsequent studies have explored encoding OCR information using additional encoders. Cream [24] utilizes BLIP-2 [30] to encode document OCR positions, images, text, and additional object detection information, with the output of BLIP-2 fed into the LLM. InstructDr [60] introduces an extra feature extractor to encode document OCR information, compressing it with learnable tokens before inputting it into the LLM. Additionally, Fujitake [11] and Luo et al. [45] adopt Lay", + "bbox": [ + 89, + 553, + 482, + 902 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "outLMv3 [20] to encode the OCR information of documents before inputting them into the LLM. Although these methods avoid the issue of overly lengthy text inputs, they must use the extra DE to encode all the OCR textual, positional information from the document due to the structural restriction of DE. However, the additional DE might not be comparable with LLMs themselves in terms of document comprehension capability. This might be the reason why they usually need extensive training of both the DE and the LLM to align the features encoded by the DE with those of the LLM, which consumes significant computational resources. DocLLM [67] proposes an extension where the hidden states derived from OCR bounding boxes interact with those of OCR text features encoded by LLMs within each attention block via cross-attention [5, 6]. However, they use an extra projection layer at each attention block to transform and inject layout information, which might excessively interfere with the original attention computation of LLM, thereby potentially affecting LLM's generalization capability. Moreover, it does not adequately consider the role of image features in TDU. Overall, designing an efficient method to provide OCR information to LLMs remains a research question worthy of deep exploration.", + "bbox": [ + 511, + 378, + 906, + 728 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3. DocLayLLM", + "text_level": 1, + "bbox": [ + 511, + 739, + 650, + 757 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "DocLayLLM is an LLM-based methodology that encodes OCRed textual, visual, and positional information directly within an LLM, eliminating the need for additional DE. The training regime of DocLayLLM includes both pretraining and supervised fine-tuning phases, which are both instruction-tuning-based processes. To further enhance our model, we propose the integration of CoT into our pretraining tasks, denoted as CoT Pre-training. Additionally, we introduce CoT Annealing, which involves the applica", + "bbox": [ + 511, + 763, + 906, + 902 + ], + "page_idx": 2 + }, + { + "type": "page_number", + "text": "4040", + "bbox": [ + 482, + 944, + 514, + 955 + ], + "page_idx": 2 + }, + { + "type": "image", + "img_path": "images/f217931446d01cd5e3933c2f49b3147244b184376610f7c07309eb1661ae21af.jpg", + "image_caption": [ + "(a) Layout-Type-Related CoT." + ], + "image_footnote": [], + "bbox": [ + 125, + 88, + 357, + 247 + ], + "page_idx": 3 + }, + { + "type": "image", + "img_path": "images/1b53df17ea4a95f1df09dcad8302374bd5ded2f4093b1e50d8db8d0cfd5ebbd7.jpg", + "image_caption": [ + "(b) Table-Structure-Aware CoT.", + "Figure 2. Examples of CoT Pre-training across different tasks. Red text highlights key steps in the chain of thought for each task." + ], + "image_footnote": [], + "bbox": [ + 369, + 88, + 614, + 247 + ], + "page_idx": 3 + }, + { + "type": "image", + "img_path": "images/ad1828532a29f3a49f7b14a2e728d2f4cd3b718da1dcec76aa6fc670e403a950.jpg", + "image_caption": [ + "(c) Geometry-Related CoT." + ], + "image_footnote": [], + "bbox": [ + 624, + 88, + 872, + 247 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "tion of annealing to the data utilized in the supervised finetuning phase, specifically from the perspective of CoT.", + "bbox": [ + 89, + 315, + 482, + 345 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "3.1. Model Architecture", + "text_level": 1, + "bbox": [ + 89, + 352, + 277, + 367 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "The overall architecture of our DocLayLLM is illustrated in Figure 1. To process a document image, we first employ an OCR engine. This engine extracts the OCR results, which encompasses the textual content, denoted as $Text_{1:N}$ , and the associated bounding box, represented by $Box_{1:N}$ . Here, $N$ indicates the total number of text segments detected by the OCR, and each $Box_{i} = [x_{0},y_{0},x_{1},y_{1}]$ defines the bounding box for $Text_{i}$ . The bounding box coordinates correspond to the left, top, right, and bottom boundaries of the text segment. After the OCR process, the document image is segmented into patches, denoted as $Patch_{1:M}$ , where $M$ represents the total number of patches. Next, the OCR results and the document patches are integrated with the fixed part of the prompt and the specific Question related to the document. This integration forms the final prompt input to the model, presented in a natural language expression $\\mathcal{T}$ : Given the document patches: $\\{Patch_{1:M}\\}$ and the document text contents and locations in the form of \"text, [left, top, right, bottom]\": $\\{(Text_{i}, Box_{i})_{1:N}\\}$ $\\{Question\\}$", + "bbox": [ + 88, + 375, + 482, + 662 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Subsequently, the textual components of the prompt, including the fixed expression, $Text_{1:N}$ , and $Question$ , are processed using the LLM's native text embedder. This generates the embeddings for the text content as follows:", + "bbox": [ + 89, + 662, + 482, + 722 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\nE m b _ {T} = T E (T e x t C o n t e n t), \\tag {1}\n$$\n", + "text_format": "latex", + "bbox": [ + 183, + 728, + 480, + 744 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "where $TE$ is the native text embedder of LLM, and TextContent consists of the fixed expression, $Text_{1:N}$ , and Question.", + "bbox": [ + 89, + 750, + 482, + 795 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "As for $Patch_{1:M}$ and $Box_{1:N}$ , they are first embedded with an extra embedder, then mapped to the LLM's feature space via a linear projector, yielding $Emb_V$ and $Emb_L$ as:", + "bbox": [ + 89, + 795, + 482, + 842 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\nE m b _ {V} = V P (V E (P a t c h _ {1: M})), \\tag {2}\n$$\n", + "text_format": "latex", + "bbox": [ + 174, + 861, + 482, + 877 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\nE m b _ {L} = L P \\left(L E \\left(\\text {B o x} _ {1: N}\\right)\\right), \\tag {3}\n$$\n", + "text_format": "latex", + "bbox": [ + 178, + 881, + 480, + 896 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "where $VE$ and $VP$ represent the visual embedder and visual projector, $LE$ and $LP$ denote the layout embedder and layout projector, respectively. It is worth noting that despite the introduction of $VE$ , $VP$ , $LE$ , and $LP$ , these are all simple embedding layers or linear layers, thereby minimizing additional computational burden.", + "bbox": [ + 511, + 315, + 903, + 405 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Finally, LLM will comprehend the embeddings provided and utilize them to generate the Answer to the Question:", + "bbox": [ + 511, + 405, + 903, + 436 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\text {A n s w e r} = L L M \\left(\\pi \\left(\\left(E m b _ {T}, E m b _ {V}, E m b _ {L}\\right), \\mathcal {T}\\right)\\right). \\tag {4}\n$$\n", + "text_format": "latex", + "bbox": [ + 527, + 444, + 903, + 460 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Here $\\pi$ represents organizing $Emb_{T}, Emb_{V}, Emb_{L}$ into a feature sequence according to $\\mathcal{T}$ .", + "bbox": [ + 511, + 469, + 903, + 500 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "With our design, we have preserved the original structure of the LLM, ensuring its generalization capability. At the same time, we have enabled the LLM to easily accept and process visual and layout features, making it better suited for understanding text-rich documents with various layouts. This results in a simple yet effective multi-modal extension.", + "bbox": [ + 511, + 500, + 905, + 590 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "3.2. CoT Pre-training", + "text_level": 1, + "bbox": [ + 511, + 597, + 684, + 613 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "During the pre-training phase, we incorporate a wide range of tasks (as detailed in the appendix) to boost the LLM's understanding of the visual and layout features inspired by Luo et al. [45]. Since the datasets related to these tasks are not designed explicitly for instruction-tuning, their annotations are not in a QA format. Therefore, we manually create QA-formatted instructions based on their annotations for training. However, simply converting task descriptions into Question and annotations into Answer is highly inefficient regarding data utilization. To address this, we incorporate CoT during the conversion to instruction-tuning data for specific tasks, asking LLM to solve our problem step-by-step thus enhancing the model's comprehension abilities, as the examples illustrated in Fig 2. Our designed CoT primarily falls into the following three categories:", + "bbox": [ + 511, + 619, + 906, + 847 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "3.2.1. Layout-Type-Related CoT", + "text_level": 1, + "bbox": [ + 511, + 851, + 741, + 867 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "We conduct a layout-type-related CoT for the layout analysis task. To determine the layout type of a given area, it is", + "bbox": [ + 511, + 869, + 905, + 901 + ], + "page_idx": 3 + }, + { + "type": "page_number", + "text": "4041", + "bbox": [ + 482, + 944, + 513, + 955 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "beneficial first to locate its region within the document and then infer based on the nearby layout element. When giving an area bounded by a bounding box, our layout-type-related CoT can be formulated as following steps:", + "bbox": [ + 89, + 90, + 480, + 151 + ], + "page_idx": 4 + }, + { + "type": "list", + "sub_type": "text", + "list_items": [ + "1. Identify the text within the area and determine its approximate region within the document (e.g., top left, top right part of the document, etc).", + "2. Locate the nearest bounding boxes and assess their layout type.", + "3. Infer the layout type of the given area based on the above information." + ], + "bbox": [ + 89, + 152, + 482, + 255 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Through this CoT process, we enable the LLM to fully utilize the visual, layout, and textual features of the document, thereby improving their ability to understand the document's overall structure.", + "bbox": [ + 89, + 257, + 482, + 316 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "3.2.2. Table-Structure-Aware CoT", + "text_level": 1, + "bbox": [ + 89, + 325, + 331, + 339 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Table-structure-aware CoT is primarily employed in the table analysis task. We have integrated the XY-Cut algorithm [48]—a widely-used technique for analyzing table structures. For instance, to locate an element at the $i$ -th row and $j$ -th column, the CoT process is outlined as follows:", + "bbox": [ + 89, + 344, + 482, + 419 + ], + "page_idx": 4 + }, + { + "type": "list", + "sub_type": "text", + "list_items": [ + "1. Identify all the table headers and their respective locations. The header of the $j$ -th column corresponds to the $j$ -th element in the headers list.", + "2. Output all elements in the $j$ -th column when told that the elements of the same column are roughly vertically aligned with their corresponding headers.", + "3. Determine the final answer based on the $i$ -th element of the contents in the $j$ -th column." + ], + "bbox": [ + 91, + 420, + 482, + 541 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Applying the above CoT, we break down a complex question into simple steps using the basic XY-Cut thought. This approach improves the LLM's inference ability through this breakdown and enhances data utilization by effectively leveraging the overall structure of the table.", + "bbox": [ + 89, + 541, + 482, + 617 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "3.2.3. Geometry-Related CoT", + "text_level": 1, + "bbox": [ + 89, + 625, + 302, + 638 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "This is used in geometric analysis to identify geometric relationships between two text segments. It assists LLM to inference with geometric relations. Specifically, when giving the text content of two lines:", + "bbox": [ + 89, + 643, + 482, + 703 + ], + "page_idx": 4 + }, + { + "type": "list", + "sub_type": "text", + "list_items": [ + "1. Recover the bounding box of the given text segments.", + "2. Analyze their vertical and horizontal projections to determine if they overlap through interval analysis.", + "3. Calculate the center point of each segment and determine their relative location by comparing their center points.", + "4. Calculate the minimum Euclidean distance based on the previous analysis. If they overlap, the distance is 0; if they overlap in vertical/horizontal projections, the minimum distance is the smallest interval in the horizontal/vertical projections; otherwise, it is the minimum Euclidean distance between their nearest corners." + ], + "bbox": [ + 89, + 704, + 482, + 868 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "This design advances the LLM's arithmetic reasoning capabilities and helps it more accurately grasp the spatial re", + "bbox": [ + 89, + 869, + 482, + 901 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "relationships between different boxes, and also contributes to layout-type-aware CoT to find the nearest bounding boxes.", + "bbox": [ + 511, + 90, + 903, + 121 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "3.3. CoT Annealing", + "text_level": 1, + "bbox": [ + 511, + 136, + 668, + 152 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Data annealing involves gradually increasing the proportion of high-quality data towards the end of supervised fine-tuning to improve data utilization efficiency and augment the model's reasoning abilities. This technique has been proven to be effective and is widely adopted in models like MiniCPM [19], DataComp-LM [31], and Llama-3.1 [42].", + "bbox": [ + 511, + 160, + 903, + 251 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "During the SFT stage, we employ data with a layout-aware CoT [45] that helps focus on potential regions of questions and answers within documents, thereby better leveraging the layout features learned during pre-training. It also allows the model to strengthen its reasoning capacity through step-by-step thinking. However, in the latter stages of SFT, our goal for LLM is to output precise and direct answers to questions. At this point, using CoT data may introduce unnecessary noise with a long response, and the overly uniform layout-aware CoT data might partially compromise the LLM's generalization abilities.", + "bbox": [ + 511, + 253, + 906, + 419 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "To balance the strengths and weaknesses of such data, we propose CoT annealing, which reconsiders this data annealing issue from a CoT perspective. Specifically, we generate direct-answer data without CoT from the CoT data, treating it as high-quality data. During the SFT process, we gradually increase the proportion of this direct-answer data, thereby enhancing the LLM's ability to respond with a clear and direct answer in the end. Initially, we only utilize the data with CoT, gradually adjusting the ratio of CoT to non-CoT data as the SFT process progresses. Ultimately, we conclude the training using only the data without CoT.", + "bbox": [ + 511, + 421, + 905, + 587 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "4. Experiments", + "text_level": 1, + "bbox": [ + 511, + 606, + 643, + 623 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "4.1. Datasets", + "text_level": 1, + "bbox": [ + 511, + 633, + 614, + 647 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "During the pre-training stage, we conduct our pre-training tasks on several datasets, including RVL-CDIP [13], PubLayNet [80], PubTabNet [81], DocLayNet [54], DocBank [32], DocILE [56], and the Document Dense Description data from LayoutLLM [45]. This results in a total of 3.1 million pairs of instruction tuning data, which is significantly less than the volume used in previous OCR-dependent TDU methods. For the supervised fine-tuning stage, we apply CoT annealing on the layout-aware document QA data utilized in LayoutLLM [45].", + "bbox": [ + 511, + 657, + 905, + 808 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "We evaluate our DocLayLLM on widely used text-rich document benchmarks, including FUNSD [22], CORD [50], SROIE [21], POIE [25], DocVQA [46], InfoVQA [47], VisualMRC [59], DeepForm [58], KLC [57], WTQ [51], and TabFact [4]. For a fair comparison with other approaches, we use the default OCR data from the of", + "bbox": [ + 511, + 810, + 905, + 900 + ], + "page_idx": 4 + }, + { + "type": "page_number", + "text": "4042", + "bbox": [ + 482, + 944, + 514, + 955 + ], + "page_idx": 4 + }, + { + "type": "table", + "img_path": "images/edb5bc45fbd940757148ee9d9a4c06caefa45240d795de091c14dc35544bcab0.jpg", + "table_caption": [], + "table_footnote": [], + "table_body": "
MethodProcessed Data at Pre-trainingDocument-oriented VQAKIE
DocVQAVisualMRCAvg.FUNSD†CORD†SROIE†Avg.
Text
Llama2-7B-Chat [64]-20.509.9015.2015.1020.0035.623.57
Llama3-8B-Instruct [42]-51.7947.7749.7868.5652.3161.2460.70
Text + Box
Llama2-7B-Chatcoor [14]-12.3012.2012.2511.906.4039.4019.23
Llama3-8B-Instructcoor [14]-49.1341.7545.4474.0062.2063.1566.45
LayTextLLMzero [43]10.0M (323%)65.5037.4051.4572.0045.5082.0066.50
Text + Box + Patch
LayoutLLM Llama2 [45]5.7M (184%)74.2555.7364.9978.6562.2170.9770.61
DocLayLLM Llama2 (Ours)3.1M (100%)72.8355.0463.9478.6770.8183.2777.58
DocLayLLM Llama3 (Ours)3.1M (100%)78.3658.5568.4684.1171.3484.3679.94
", + "bbox": [ + 171, + 88, + 826, + 262 + ], + "page_idx": 5 + }, + { + "type": "table", + "img_path": "images/2152fd2ca1786310a14936ab71f7bf774f0553e576a62597b0a36159ed34a79d.jpg", + "table_caption": [ + "Table 1. Comparison with OCR-dependent methods using metrics from LayoutLLM [45]. Processed Data at Pre-training represents the total data volume processed at the pre-training stage, equivalent to dataset size multiplied by epoch count, with values in parentheses representing size ratios compared to DocLayLLM. Avg. represents the average performance across different benchmarks, and $\\dagger$ indicates the cleaned test set used in LayoutLLM. The suffix layout denotes using method from He et al. [14], which provides LLM with layout features via OCR box coordinates in textual form. Bold indicates the best performance, and underline indicates the second-best one." + ], + "table_footnote": [], + "table_body": "
MethodProcessed Data at Pre-trainingDocument-oriented VQAKIE
DocVQAVisualMRCAvg.FUNSDCORDSROIEAvg.
Text
Llama2-7B-Chat [64]-20.506.3013.4023.4051.8058.6044.60
Llama3-8B-Instruct [42]-51.79229.74140.7766.6774.7182.5174.63
Text + Box
Llama2-7B-Chatcoor [14]-12.3028.0020.1514.4038.1050.6034.37
Llama3-8B-Instructcoor [14]-49.13211.69130.4174.7175.3375.9375.32
LMDX [53]-----66.95--
DocLLM [67]16.8M (542%)69.50*264.10*166.80*51.80*67.60*91.90*70.43*
LayTextLLMzero [43]10.0M (323%)65.50200.20132.8547.2077.2083.8069.40
Text + Box + Patch
DocLayLLM Llama2 (Ours)3.1M (100%)72.83310.60191.7280.7479.3784.3781.49
DocLayLLM Llama3 (Ours)3.1M (100%)78.35357.04217.7086.4783.6686.0885.40
", + "bbox": [ + 171, + 354, + 826, + 542 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Table 2. Comparison with OCR-dependent methods using metrics from DocLLM [67]. * indicates that the model uses the training set of the benchmark during training.", + "bbox": [ + 89, + 551, + 906, + 580 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "ficial datasets otherwise commercial OCR engines ${}^{1}$ if not provided, and follow their evaluation pipeline and metrics.", + "bbox": [ + 89, + 606, + 482, + 638 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "4.2. Implement Details", + "text_level": 1, + "bbox": [ + 89, + 645, + 269, + 660 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Our method is based on Llama3-8B-Instruct [42], with a Llama2-7B-Chat [64] version for better comparison with other methods. They are denoted as DocLayLLM $_{Llama3}$ and DocLayLLM $_{Llama2}$ respectively. We initialize $VE$ and $LE$ with weights from LayoutLMv3 [20] to leverage document knowledge acquired during its pre-training, while $VP$ and $LP$ are randomly initialized. Besides, the image resolution is fixed at $224*224$ , resulting in 196 visual patch tokens, to capture an overview of document layout while bringing less computation burden. During pre-training and fine-tuning, we use LoRA [18] with a rank of 64 on LLM and keep $VE$ , $VP$ , $LE$ , and $LP$ trainable, leading to few tuning parameters. We train our model on 8 NVIDIA Tesla A100-80G GPUs, taking about 30 A100 days for pre", + "bbox": [ + 89, + 667, + 482, + 878 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "training and 6 A100 days for supervised fine-tuning. Additional details can be found in the appendix.", + "bbox": [ + 511, + 606, + 903, + 638 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "4.3. Comparisons with SOTA OCR-dependent methods", + "text_level": 1, + "bbox": [ + 511, + 648, + 906, + 679 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "We have compared our DocLayLLM with existing SOTA OCR-dependent methods. In order to exhibit the zero-shot inferring ability precisely, we chose their performances under the zero-shot scenario unless they are not provided.", + "bbox": [ + 511, + 686, + 905, + 747 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "The experimental results are shown in Table 1 and Table 2. Our method outperforms the baseline models, regardless of their type or whether the document's layout information is provided to them. It validates the rationale behind our design of multi-modal extension.", + "bbox": [ + 511, + 748, + 905, + 824 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Additionally, our method significantly outperforms previous SOTA methods despite fewer training resources. As shown in Table 1, it surpasses the performance of Layout-LLM [45] and LayTextLLM [43] though requiring fewer computational resources by using less data and tuning few", + "bbox": [ + 511, + 825, + 905, + 901 + ], + "page_idx": 5 + }, + { + "type": "page_footnote", + "text": "1 https://www.textin.com/", + "bbox": [ + 109, + 886, + 245, + 898 + ], + "page_idx": 5 + }, + { + "type": "page_number", + "text": "4043", + "bbox": [ + 482, + 944, + 514, + 955 + ], + "page_idx": 5 + }, + { + "type": "table", + "img_path": "images/43de467f9e54bd0362645cffd147418ed24630c12e5a78d965519ca83a673d96.jpg", + "table_caption": [], + "table_footnote": [], + "table_body": "
MethodDocument-oriented VQAKIE
DocVQAInfoVQAAvg.FUNSDSROIEPOIEAvg.
LLaVAR [79]12.3016.5014.400.505.205.903.87
UniDoc [10]7.7014.7011.201.002.905.103.00
DocPedia [9]47.10*15.20*31.15*29.9021.4039.9030.40
Monkey [38]50.10*25.80*37.95*24.1041.9019.9028.63
TextMonkey [41]66.70*28.60*47.65*42.9046.2032.0040.37
DocOwl 1.5 [17]----48.3051.80-
TextSquare [61]----53.2071.80-
DocKylin [78]65.10*34.80*49.95*58.7025.5049.5044.57
DocLayLLM\\( _{Llama3} \\) (Ours)77.7942.0259.9180.2876.5975.1377.33
", + "bbox": [ + 196, + 89, + 803, + 242 + ], + "page_idx": 6 + }, + { + "type": "table", + "img_path": "images/287f36606b711a1ac1f5b6c8842d907c5b2df8b75e1d30465fa8a1d0a495e307.jpg", + "table_caption": [ + "Table 3. Comparison with OCR-free methods using the precision metric from TextMonkey [41]. * indicates that the model uses the training set of the benchmark during training." + ], + "table_footnote": [], + "table_body": "
MethodDocument-oriented VQAKIETable Understanding
DocVQAInfoVQAVisualMRCAvg.DeepFormKLCAvg.WTQTabFactAvg.
DocPedia [9]47.1015.20--------
DocOwl [75]62.2038.20188.8096.4042.6030.3036.4526.9060.2043.55
URector [76]65.4042.20221.70109.7749.5032.8041.1529.4067.6048.50
Monkey [38]66.5036.10--40.6032.8036.7025.30--
TextMonkey [41]73.00---59.7037.8048.7531.90--
DocKylin [78]77.3046.60319.90147.9364.20--32.40--
CogAgent [16]81.6044.50--------
DocOwl 1.5 [17]82.2050.70246.40126.4368.8038.7053.7540.6080.2060.40
TextSquare [61]84.3051.50-----49.7084.2066.95
DocLayLLM\\(\\ddagger\\) Llama3 (Ours)86.5258.36327.91157.6077.0740.7358.9058.6283.3670.99
", + "bbox": [ + 138, + 295, + 857, + 455 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Table 4. Comparison with OCR-free methods using the evaluation framework and metrics from DocOwl 1.5 [17]. All methods, including our $DocLayLLM_{Llama3}^{\\ddagger}$ variant, are exposed to the train set corresponding to the evaluation set during training.", + "bbox": [ + 89, + 467, + 906, + 497 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "parameters at training. Even with Llama2 as the backbone, our approach matches the performance of previous SOTA methods in the Document-oriented VQA task and significantly outperforms them in the KIE tasks, further demonstrating the efficiency and effectiveness of our approach.", + "bbox": [ + 88, + 522, + 480, + 598 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Besides, our method exhibits robust zero-shot generalization capability without fine-tuning with specific downstream tasks. From the comparison with DocLLM [67] in Table 2, it is clear that our method, even without fine-tuning, significantly outperforms DocLLM across most datasets and achieves a considerable overall performance improvement. This validates our design, which has preserved the strong generalization ability of the original LLM structure.", + "bbox": [ + 88, + 598, + 482, + 718 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "In summary, DocLayLLM serves as an efficient extension of LLMs. It demonstrates strong zero-shot generalization and effective document understanding by encoding documents within a unified model. Even with limited training resources, DocLayLLM achieves new SOTA performance compared to OCR-dependent competitors.", + "bbox": [ + 88, + 718, + 482, + 809 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "4.4. Comparisons with SOTA OCR-free methods", + "text_level": 1, + "bbox": [ + 89, + 818, + 467, + 834 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "We have also compared our method with SOTA OCR-free TDU models. Table 3 presents the precision of our method and some advent OCR-free TDU methods across various benchmarks. Even though in the circumstance that OCR-", + "bbox": [ + 89, + 839, + 482, + 900 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "free models use some of the training sets of these benchmarks during training, while our method infers in a purely zero-shot manner, our DocLayLLM still outperforms these models, particularly in KIE task, which heavily relies on layout information. This clearly demonstrates the superiority of our design to provide LLM with OCR information.", + "bbox": [ + 511, + 522, + 903, + 613 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "We additionally test DocLayLLM on a broader set of benchmarks from DocOwl 1.5 [17]. To further assess the performance of DocLayLLM on seen datasets, we incorporate the training sets of these benchmarks used in DocOwl 1.5 during the SFT process as other OCR-free TDU methods do during training. To maintain the efficiency advantage of our method, we ensure that the SFT is conducted with the same number of steps as the standard DocLayLLM. The results in Table 4 show that with the inclusion of downstream task data during SFT, our method outperforms the SOTA OCR-free TDU models in most datasets. Moreover, our overall performance in the Document-oriented VQA, KIE, and Table Understanding sub-tasks surpasses the OCR-free TDU methods using limited training resources, showcasing the tremendous potential of OCR-dependent methods and further reinforcing the reliability of our method.", + "bbox": [ + 511, + 613, + 905, + 853 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Overall, our DocLayLLM consistently shows strong zero-shot inference ability, outperforming many OCR-free TDU methods without prior exposure to downstream data.", + "bbox": [ + 511, + 854, + 903, + 900 + ], + "page_idx": 6 + }, + { + "type": "page_number", + "text": "4044", + "bbox": [ + 482, + 944, + 514, + 955 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/8162119e9d04c2187be181409bb60212059e6dda14d0319715346cd8850d696d.jpg", + "image_caption": [ + "Figure 3. Qualitative comparisons of DocLayLLM with other OCR-dependent and OCR-free methods. Zoom in for better view." + ], + "image_footnote": [], + "bbox": [ + 102, + 88, + 472, + 305 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Additionally, when incorporating downstream datasets during SFT like other OCR-free approaches, DocLayLLM surpasses them, highlighting its potential in solving TDU tasks by integrating OCR information into the LLM.", + "bbox": [ + 89, + 369, + 483, + 433 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "4.5. Qualitative Results", + "text_level": 1, + "bbox": [ + 89, + 439, + 274, + 455 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Figure 3 shows some qualitative examples to visualize the performance on KIE tasks of our DocLayLLM compared with LayTextLLM and DocOwl 1.5. The example from SROIE clearly shows that DocLayLLM has improved document comprehension and answer accuracy. And the example from POIE demonstrates our advantage in handling table-formatted documents to give an accurate response. Furthermore, as seen in the FUNSD examples, DocLayLLM excels at extracting answers spanning multiple text lines, outperforming both OCR-dependent and OCR-free methods. For more visualization of qualitative examples, please refer to the appendix.", + "bbox": [ + 88, + 460, + 482, + 643 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "4.6. Ablation Study", + "text_level": 1, + "bbox": [ + 89, + 652, + 243, + 667 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "We have also conducted additional experiments to explore the effectiveness of our proposed method. To conserve computational resources, we perform pre-training at $25\\%$ of the standard scale (including the baseline) but full-scale SFT. We focus on the impact of two main factors: (a) the role of CoT pre-training and CoT annealing, and (b) the influence of different modality configurations.", + "bbox": [ + 89, + 674, + 482, + 779 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Experimental results in Table 5a indicate that omitting CoT Pre-training leads to an obvious performance drop in both the Document-oriented VQA and KIE tasks, while not using CoT Annealing primarily affects VQA performance. Besides, using randomly initialized $VE$ and $LE$ partially compromises the proposed extension, resulting in performance degradation in both the VQA and KIE tasks.", + "bbox": [ + 89, + 780, + 482, + 885 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Additionally, performances in Table 5b show the efficacy", + "bbox": [ + 109, + 885, + 482, + 901 + ], + "page_idx": 7 + }, + { + "type": "table", + "img_path": "images/ef9954dfbaba0c478fdb53ec456c687b68fa891b89c60be6e856cf44f9c4f9d4.jpg", + "table_caption": [], + "table_footnote": [ + "(a) Impact of different settings: $w / o$ CoT Pre-training and $w / o$ CoT Annealing indicate pre-training without CoT and supervised fine-tuning with only CoT data, while $w / o$ pt VE and PE uses randomly initialized visual and layout embeddings instead of pre-trained ones." + ], + "table_body": "
ExperimentDocument-oriented VQAKIE
DocVQAVisualMRCAvg.DeepFormKLCAvg.
DocLayLLM1lama377.43316.27196.8535.4126.2530.83
w/o CoT Pre-training76.00295.49185.7524.5425.4925.02
w/o CoT Annealing77.40310.17193.7934.2326.2130.22
w/o pt VE & LE77.42308.49192.9631.3125.7628.54
", + "bbox": [ + 527, + 89, + 893, + 152 + ], + "page_idx": 7 + }, + { + "type": "table", + "img_path": "images/2baf010f8aee4379dade9668a79bc191ed61cb6c57177b5c3982b6e3c50dd30b.jpg", + "table_caption": [], + "table_footnote": [ + "(b) Impacts of different modality configurations. $\\mathbf{T}$ , $\\mathbf{L}$ , and $\\mathbf{V}$ represent textual, layout, and visual modalities, respectively. (I) indicates inputting OCR coordinates as text following He et al. [14], while (II) embeds coordinates with $LE$ in our approach." + ], + "table_body": "
ModalityDocument-oriented VQAKIE
TLVDocVQAVisualMRCAvg.DeepFormKLCAvg.
✓✗✗73.57280.27176.9315.7624.0919.93
✓(I)✗75.72306.51191.1214.1324.2319.18
✓(II)✗77.32310.75194.0333.7525.9729.86
✓(II)✓77.43316.27196.8535.4126.2530.83
", + "bbox": [ + 527, + 219, + 893, + 290 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Table 5. Ablation study on various designs of our DocLayLLM.", + "bbox": [ + 519, + 353, + 898, + 368 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "of our DocLayLLM to incorporate different modalities for TDU tasks. It is obvious that using only text from the document is insufficient for TDU, while adding layout information in textual form provides some gain on VQA tasks but leads to over-length input (see appendix for analysis), greatly slowing the inference. By our design to integrate layout features via an $LE$ , we achieve notable improvements on these TDU tasks with shorter inputs. And adding visual features further boosts the performance of our method.", + "bbox": [ + 511, + 393, + 903, + 527 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Overall, the combined impact of these designs ensures the efficiency and effectiveness of our method.", + "bbox": [ + 511, + 529, + 903, + 560 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "5. Conclusion", + "text_level": 1, + "bbox": [ + 511, + 573, + 633, + 588 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "We propose DocLayLLM, an efficient multi-modal extension of large language models for text-rich document understanding. By streamlined integration of document patch tokens and OCR positional tokens into the LLM inputs, it achieves a simple yet efficient multi-modal extension of LLM to enable it to have the capability to perceive and process textual, visual, and layout information from the OCR results of the document. Additionally, we introduce CoT Pre-training and CoT Annealing techniques centered around CoT, offering a new way to generate CoT data and further enhancing the model's understanding of text-rich documents. Despite requiring fewer training resources, our method surpasses existing OCR-dependent TDU models, demonstrating strong zero-shot generalization capability. Furthermore, it also exceeds the performance of SOTA OCR-free TDU methods under comparable training conditions, showcasing remarkable performance across various document understanding tasks. Our work significantly advances efficient MLLM-based document understanding while maintaining high-performance standards.", + "bbox": [ + 511, + 598, + 906, + 902 + ], + "page_idx": 7 + }, + { + "type": "page_number", + "text": "4045", + "bbox": [ + 482, + 944, + 514, + 955 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Acknowledgement", + "text_level": 1, + "bbox": [ + 91, + 90, + 250, + 107 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "This research is supported in part by National Natural Science Foundation of China (Grant No.: 62476093, 62441604).", + "bbox": [ + 89, + 113, + 485, + 156 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "References", + "text_level": 1, + "bbox": [ + 91, + 167, + 187, + 184 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[1] Srikar Appalaraju, Bhavan Jasani, Bhargava Urala Kota, Yusheng Xie, and R. Manmatha. DocFormer: End-to-End Transformer for Document Understanding. In Proc. ICCV, pages 993–1003, 2021. 1", + "[2] Srikar Appalaraju, Peng Tang, Qi Dong, Nishant Sankaran, Yichu Zhou, and R. Manmatha. DocFormerv2: Local Features for Document Understanding. In Proc. AAAI, pages 709-718, 2024. 1", + "[3] Tom Brown, Benjamin Mann, Nick Ryder, Melanie Subbiah, Jared D Kaplan, Prafulla Dhariwal, Arvind Neelakantan, Pranav Shyam, Girish Sastry, Amanda Askell, Sandhini Agarwal, Ariel Herbert-Voss, Gretchen Krueger, Tom Henighan, Rewon Child, Aditya Ramesh, Daniel Ziegler, Jeffrey Wu, Clemens Winter, Chris Hesse, Mark Chen, Eric Sigler, Mateusz Litwin, Scott Gray, Benjamin Chess, Jack Clark, Christopher Berner, Sam McCandlish, Alec Radford, Ilya Sutskever, and Dario Amodei. Language Models are Few-shot Learners. Proc. NIPS, 33:1877-1901, 2020. 2", + "[4] Wenhu Chen, Hongmin Wang, Jianshu Chen, Yunkai Zhang, Hong Wang, Shiyang Li, Xiyou Zhou, and William Yang Wang. TabFact: A Large-scale Dataset for Table-based Fact Verification. In Proc. ICLR, 2020. 5", + "[5] Gang Dai, Yifan Zhang, Qingfeng Wang, Qing Du, Zhuliang Yu, Zhuoman Liu, and Shuangping Huang. Disentangling Writer and Character Styles for Handwriting Generation. In Proc. CVPR, pages 5977-5986, 2023. 3", + "[6] Gang Dai, Yifan Zhang, Quhui Ke, Qiangya Guo, and Shuangping Huang. One-dm: One-shot diffusion mimicker for handwritten text generation. In Proc. ECCV, pages 410-427, 2025. 3", + "[7] DeepSeek-AI. DeepSeek-R1: Incentivizing Reasoning Capability in LLMs via Reinforcement Learning. arXiv: Comp. Res. Repository, abs/2501.12948, 2025. 2", + "[8] Jacob Devlin, Ming-Wei Chang, Kenton Lee, and Kristina Toutanova. BERT: Pre-training of Deep Bidirectional Transformers for Language Understanding. In Proc. NAACL, pages 4171–4186, 2019. 1", + "[9] Hao Feng, Qi Liu, Hao Liu, Wengang Zhou, Houqiang Li, and Can Huang. DocPedia: Unleashing the Power of Large Multimodal Model in the Frequency Domain for Versatile Document Understanding. arXiv: Comp. Res. Repository, abs/2311.11810, 2023. 7", + "[10] Hao Feng, Zijian Wang, Jingqun Tang, Jinghui Lu, Wengang Zhou, Houqiang Li, and Can Huang. UniDoc: A Universal Large Multimodal Model for Simultaneous Text Detection, Recognition, Spotting and Understanding. arXiv: Comp. Res. Repository, abs/2308.11592, 2023. 7", + "[11] Masato Fujitake. LayoutLLM: Large Language Model Instruction Tuning for Visually Rich Document Understanding. In Proc. LREC/COLING, pages 10219-10224, 2024. 2, 3" + ], + "bbox": [ + 93, + 193, + 483, + 898 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[12] Lukasz Garncarek, Rafat Powalski, Tomasz Stanisławek, Bartosz Topolski, Piotr Halama, Michal Turski, and Filip Graliński. LAMBERT: Layout-Aware Language Modeling for Information Extraction. In Proc. ICDAR, pages 532-547, 2021. 1", + "[13] Adam W. Harley, Alex Ufkes, and Konstantinos G. Derpanis. Evaluation of Deep Convolutional Nets for Document Image Classification and Retrieval. In Proc. ICDAR, pages 991-995, 2015. 5", + "[14] Jiabang He, Lei Wang, Yi Hu, Ning Liu, Hui Liu, Xing Xu, and Heng Tao Shen. ICL-D3IE: In-Context Learning with Diverse Demonstrations Updating for Document Information Extraction. In Proc. ICCV, pages 19428-19437, 2023. 1, 2, 3, 6, 8", + "[15] Teakgyu Hong, DongHyun Kim, Mingi Ji, Wonseok Hwang, Daehyun Nam, and Sungrae Park. BROS: A Pre-trained Language Model Focusing on Text and Layout for Better Key Information Extraction from Documents. In Proc. AAAI, pages 10767-10775, 2022. 1", + "[16] Wenyi Hong, Weihan Wang, Qingsong Lv, Jiazheng Xu, Wenmeng Yu, Junhui Ji, Yan Wang, Zihan Wang, Yuxiao Dong, Ming Ding, and Jie Tang. CogAgent: A Visual Language Model for GUI Agents. In Proc. CVPR, pages 14281-14290, 2024. 7", + "[17] Anwen Hu, Haiyang Xu, Jiabo Ye, Ming Yan, Liang Zhang, Bo Zhang, Ji Zhang, Qin Jin, Fei Huang, and Jingren Zhou. mPLUG-DocOwl 1.5: Unified Structure Learning for OCR-free Document Understanding. In Findings EMNLP, pages 3096-3120, 2024. 1, 2, 7", + "[18] Edward J. Hu, Yelong Shen, Phillip Wallis, Zeyuan Allen-Zhu, Yuanzhi Li, Shean Wang, Lu Wang, and Weizhu Chen. LoRA: Low-Rank Adaptation of Large Language Models. In Proc. ICLR, 2022. 2, 6", + "[19] Shengding Hu, Yuge Tu, Xu Han, Chaoqun He, Ganqu Cui, Xiang Long, Zhi Zheng, Yewei Fang, Yuxiang Huang, Weilin Zhao, Xinrong Zhang, Zhen Leng Thai, Kai Zhang, Chongyi Wang, Yuan Yao, Chenyang Zhao, Jie Zhou, Jie Cai, Zhongwu Zhai, Ning Ding, Chao Jia, Guoyang Zeng, Dahai Li, Zhiyuan Liu, and Maosong Sun. MiniCPM: Unveiling the Potential of Small Language Models with Scalable Training Strategies. arXiv: Comp. Res. Repository, abs/2404.06395, 2024. 5", + "[20] Yupan Huang, Tengchao Lv, Lei Cui, Yutong Lu, and Furu Wei. LayoutLMv3: Pre-training for Document AI with Unified Text and Image Masking. In Proc. ACM MM, pages 4083-4091, 2022. 1, 2, 3, 6", + "[21] Zheng Huang, Kai Chen, Jianhua He, Xiang Bai, Dimosthenis Karatzas, Shijian Lu, and C. V. Jawahar. ICDAR2019 Competition on Scanned Receipt OCR and Information Extraction. In Proc. ICDAR, pages 1516-1520, 2019. 5", + "[22] Guillaume Jaume, Hazim Kemal Ekenel, and Jean-Philippe Thiran. FUNSD: A Dataset for Form Understanding in Noisy Scanned Documents. In Workshop on ICDAR, pages 1–6, 2019. 5", + "[23] Geewook Kim, Teakgyu Hong, Moonbin Yim, JeongYeon Nam, Jinyoung Park, Jinyeong Yim, Wonseok Hwang, Sangdoo Yun, Dongyoon Han, and Seunghyun Park. OCR-Free" + ], + "bbox": [ + 516, + 92, + 905, + 901 + ], + "page_idx": 8 + }, + { + "type": "page_number", + "text": "4046", + "bbox": [ + 482, + 944, + 516, + 955 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "Document Understanding Transformer. Proc. ECCV, 13688: 498-517, 2022. 2", + "[24] Geewook Kim, Hodong Lee, Daehee Kim, Haeji Jung, Sanghee Park, Yoonsik Kim, Sangdoo Yun, Taeho Kil, Bado Lee, and Seunghyun Park. Visually-Situated Natural Language Understanding with Contrastive Reading Model and Frozen Large Language Models. In Proc. EMNLP, pages 11989-12010, 2023. 2, 3", + "[25] Jianfeng Kuang, Wei Hua, Dingkang Liang, Mingkun Yang, Deqiang Jiang, Bo Ren, and Xiang Bai. Visual Information Extraction in the Wild: Practical Dataset and End-to-End Solution. In Proc. ICDAR, pages 36-53, 2023. 5", + "[26] Hugo Laurençon, Andres Marafioti, Victor Sanh, and Leo Tronchon. Building and better understanding vision-language models: insights and future directions. In Workshop on NIPS, 2024. 3", + "[27] Bo Li, Yuanhan Zhang, Dong Guo, Renrui Zhang, Feng Li, Hao Zhang, Kaichen Zhang, Peiyuan Zhang, Yanwei Li, Ziwei Liu, and Chunyuan Li. LLaVA-OneVision: Easy Visual Task Transfer. Transactions on Machine Learning Research, 2025. 3", + "[28] Chenliang Li, Bin Bi, Ming Yan, Wei Wang, Songfang Huang, Fei Huang, and Luo Si. StructuralLM: Structural Pre-training for Form Understanding. In Proc. ACL, pages 6309-6318, 2021. 1", + "[29] Feng Li, Renrui Zhang, Hao Zhang, Yuanhan Zhang, Bo Li, Wei Li, Zejun Ma, and Chunyuan Li. LLaVa-NeXT-Interleave: Tackling Multi-image, Video, and 3D in Large Multimodal Models. arXiv: Comp. Res. Repository, abs/2407.07895, 2023. 2", + "[30] Junnan Li, Dongxu Li, Silvio Savarese, and Steven C. H. Hoi. BLIP-2: Bootstrapping Language-Image Pre-training with Frozen Image Encoders and Large Language Models. Proc. ICML, 202:19730-19742, 2023. 3", + "[31] Jeffrey Li, Alex Fang, Georgios Smyrnis, Maor Ivgi, Matt Jordan, Samir Yitzhak Gadre, Hritik Bansal, Etash Guha, Sedrick Keh, Kushal Arora, Saurabh Garg, Rui Xin, Niklas Muennighoff, Reinhard Heckel, Jean Mercat, Mayee Chen, Suchin Gururangan, Mitchell Wortsman, Alon Albalak, Yonatan Bitton, Marianna Nezhurina, Amro Abbas, ChengYu Hsieh, Dhruba Ghosh, Josh Gardner, Maciej Kilian, Hanlin Zhang, Rulin Shao, Sarah M. Pratt, Sunny Sanyal, Gabriel Ilharco, Giannis Daras, Kalyani Marathe, Aaron Gokaslan, Jieyu Zhang, Khyathi Raghavi Chandu, Thao Nguyen, Igor Vasiljevic, Sham M. Kakade, Shuran Song, Sujay Sanghavi, Fartash Faghri, Sewoong Oh, Luke Zettlemoyer, Kyle Lo, Alaaeldin El-Nouby, Hadi Pouransari, Alexander Toshev, Stephanie Wang, Dirk Groeneveld, Luca Soldaini, Pang Wei Koh, Jenia Jitsev, Thomas Kollar, Alexandros G. Dimakis, Yair Carmon, Achal Dave, Ludwig Schmidt, and Vaishaal Shankar. DataComp-LM: In search of the next generation of training sets for language models. arXiv: Comp. Res. Repository, abs/2406.11794, 2024. 2, 5", + "[32] Minghao Li, Yiheng Xu, Lei Cui, Shaohan Huang, Furu Wei, Zhoujun Li, and Ming Zhou. DocBank: A Benchmark Dataset for Document Layout Analysis. In Pro. COLING, pages 949-960, 2020. 5" + ], + "bbox": [ + 91, + 90, + 482, + 900 + ], + "page_idx": 9 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[33] Peizhao Li, Jiuxiang Gu, Jason Kuen, Vlad I Morariu, Handong Zhao, Rajiv Jain, Varun Manjunatha, and Hongfu Liu. SelfDoc: Self-Supervised Document Representation Learning. In Proc. CVPR, pages 5652-5660, 2021. 1", + "[34] Qiwei Li, Zuchao Li, Xiantao Cai, Bo Du, and Hai Zhao. Enhancing Visually-Pich Pdocument Punderstanding via Payout Pstructure Podeling. In Proc. ACM MM, page 4513-4523, 2023.", + "[35] Qiwei Li, Zuchao Li, Ping Wang, Haojun Ai, and Hai Zhao. Hypergraph based Understanding for Document Semantic Entity Recognition. In Proc. ACL, pages 2950-2960, 2024.", + "[36] Xin Li, Yan Zheng, Yiqing Hu, Haoyu Cao, Yunfei Wu, Deqiang Jiang, Yinsong Liu, and Bo Ren. Relational Representation Learning in Visually-Rich Documents. In Proc. ACM MM, page 4614-4624, 2022.", + "[37] Yulin Li, Yuxi Qian, Yuechen Yu, Xiameng Qin, Chengquan Zhang, Yan Liu, Kun Yao, Junyu Han, Jingtuo Liu, and Errui Ding. StrucTexT: Structured Text Understanding with Multi-Modal Transformers. In Proc. ACM MM, page 1912–1920, 2021. 1", + "[38] Zhang Li, Biao Yang, Qiang Liu, Zhiyin Ma, Shuo Zhang, Jingxu Yang, Yabo Sun, Yuliang Liu, and Xiang Bai. Monkey: Image Resolution and Text Label Are Important Things for Large Multi-modal Models. In Proc. CVPR, pages 26763-26773, 2024. 1, 2, 7", + "[39] Zening Lin, Jiapeng Wang, and Lianwen Jin. Visual Information Extraction Deep Learning Method: A Critical Review. Journal of Image and Graphics, 28(08):2276-2297, 2023. 1", + "[40] Chenglin Liu, Lianwen Jin, Xiang Bai, Xiaohui Li, and Fei Yin. Frontiers of intelligent document analysis and recognition: review and prospects. Journal of Image and Graphics, 28(08):2223-2252, 2023. 1", + "[41] Yuliang Liu, Biao Yang, Qiang Liu, Zhang Li, Zhiyin Ma, Shuo Zhang, and Xiang Bai. TextMonkey: An OCR-Free Large Multimodal Model for Understanding Document. arXiv: Comp. Res. Repository, abs/2403.04473, 2024. 7", + "[42] Meta AI Llama Team. The Llama 3 Herd of Models. arXiv: Comp. Res. Repository, abs/2407.21783, 2024. 2, 5, 6", + "[43] Jinghui Lu, Haiyang Yu, Yanjie Wang, Yongjie Ye, Jingqun Tang, Ziwei Yang, Binghong Wu, Qi Liu, Hao Feng, Han Wang, Hao Liu, and Can Huang. A Bounding Box is Worth One Token: Interleaving Layout and Text in a Large Language Model for Document Understanding. arXiv: Comp. Res. Repository, abs/2407.01976, 2024. 6", + "[44] Chuwei Luo, Changxu Cheng, Qi Zheng, and Cong Yao. GeoLayoutLM: Geometric Pre-Training for Visual Information Extraction. In Proc. CVPR, pages 7092–7101, 2023. 1", + "[45] Chuwei Luo, Yufan Shen, Zhaoqing Zhu, Qi Zheng, Zhi Yu, and Cong Yao. LayoutLLM: Layout Instruction Tuning with Large Language Models for Document Understanding. In Proc. CVPR, pages 15630-15640, 2024. 2, 3, 4, 5, 6", + "[46] Minesh Mathew, Dimosthenis Karatzas, and C. V. Jawahar. DocVQA: A Dataset for VQA on Document Images. In Proc. WACV, pages 2199-2208, 2021. 5", + "[47] Minesh Mathew, Viraj Bagal, Ruben Tito, Dimosthenis Karatzas, Ernest Valveny, and C.V. Jawahar. InfographicVQA. In Proc. WACV, pages 1697-1706, 2022. 5" + ], + "bbox": [ + 516, + 92, + 903, + 900 + ], + "page_idx": 9 + }, + { + "type": "page_number", + "text": "4047", + "bbox": [ + 482, + 944, + 514, + 955 + ], + "page_idx": 9 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[48] G. Nagy, S. Seth, and M. Viswanathan. A Prototype Document Image Analysis System for Technical Journals. Computer, 25(7):10-22, 1992. 5", + "[49] OpenAI. Hello GPT-4o, 2024. 2", + "[50] Seunghyun Park, Seung Shin, Bado Lee, Junyeop Lee, Jaeheung Surh, Minjoon Seo, and Hwalsuk Lee. CORD: A Consolidated Receipt Dataset for Post-OCR Parsing. In Workshop on NIPS, 2019. 5", + "[51] Panupong Pasupat and Percy Liang. Compositional Semantic Parsing on Semi-Structured Tables. In Proc. ACL, pages 1470–1480, 2015. 5", + "[52] Qiming Peng, Yinxu Pan, Wenjin Wang, Bin Luo, Zhenyu Zhang, Zhengjie Huang, Yuhui Cao, Weichong Yin, Yongfeng Chen, Yin Zhang, Shikun Feng, Yu Sun, Hao Tian, Hua Wu, and Haifeng Wang. ERNIE-Layout: Layout Knowledge Enhanced Pre-training for Visually-rich Document Understanding. In Findings of EMNLP, pages 3744–3756, 2022. 1", + "[53] Vincent Perot, Kai Kang, Florian Luisier, Guolong Su, Xiaoyu Sun, Ramya Sree Boppana, Zilong Wang, Zifeng Wang, Jiaqi Mu, Hao Zhang, Chen-Yu Lee, and Nan Hua. LMDX: Language Model-based Document Information Extraction and Localization. In Findings of ACL, pages 15140-15168, 2024. 3, 6", + "[54] Birgit Pfitzmann, Christoph Auer, Michele Dolfi, Ahmed S. Nassar, and Peter W. J. Staar. DocLayNet: A Large Human-Annotated Dataset for Document-Layout Segmentation. In Proc. ACM SIGKDD Int. Conf. Knowledge discovery & data mining, pages 3743–3751, 2022. 5", + "[55] Rafat Powalski, Łukasz Borchmann, Dawid Jurkiewicz, Tomasz Dwojak, Michal Pietruszka, and Gabriela Pałka. Going Full-TILT Boogie on Document Understanding with Text-Image-Layout Transformer. In Proc. ICDAR, pages 732–747, 2021. 1", + "[56] Stepán Simsa, Milan Sulc, Michal Uricár, Yash Patel, Ahmed Hamdi, Matej Kocián, Matyas Skalický, Jiri Matas, Antoine Doucet, Mickaël Coustaty, and Dimosthenis Karatzas. DocILE Benchmark for Document Information Localization and Extraction. In Proc. ICDAR, pages 147-166, 2023. 5", + "[57] Tomasz Stanislawek, Filip Gralinski, Anna Wróblewska, Dawid Lipinski, Agnieszka Kaliska, Paulina Rosalska, Bartosz Topolski, and Przemyslaw Biecek. Kleister: Key Information Extraction Datasets Involving Long Documents with Complex Layouts. In Proc. ICDAR, pages 564-579, 2021. 5", + "[58] Stacey Svetlichnaya. DeepForm: Understand Structured Documents at Scale. https://wandb.ai/stacey/deepform_v1, 2020. Accessed: 2024-08-14. 5", + "[59] Ryota Tanaka, Kyosuke Nishida, and Sen Yoshida. VisualMRC: Machine Reading Comprehension on Document Images. In Proc. AAAI, pages 13878-13888, 2021. 5", + "[60] Ryota Tanaka, Taichi Iki, Kyosuke Nishida, Kuniko Saito, and Jun Suzuki. InstructDoc: A Dataset for Zero-shot Generalization of Visual Document Understanding with Instructions. In Proc. AAAI, pages 19071-19079, 2024. 3", + "[61] Jingqun Tang, Chunhui Lin, Zhen Zhao, Shu Wei, Binghong Wu, Qi Liu, Hao Feng, Yang Li, Siqi Wang, Lei Liao, Wei" + ], + "bbox": [ + 91, + 92, + 482, + 900 + ], + "page_idx": 10 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "Shi, Yuliang Liu, Hao Liu, Yuan Xie, Xiang Bai, and Can Huang. TextSquare: Scaling up Text-Centric Visual Instruction Tuning. arXiv: Comp. Res. Repository, abs/2404.12803, 2024. 1, 2, 7", + "[62] Zineng Tang, Ziyi Yang, Guoxin Wang, Yuwei Fang, Yang Liu, Chenguang Zhu, Michael Zeng, Cha Zhang, and Mohit Bansal. Unifying Vision, Text, and Layout for Universal Document Processing. In Proc. CVPR, pages 19254-19264, 2023. 2", + "[63] Qwen Team. QwQ: Reflect Deeply on the Boundaries of the Unknown, 2024. 2", + "[64] Hugo Touvron, Louis Martin, Kevin Stone, Peter Albert, Amjad Almahairi, Yasmine Babaei, Nikolay Bashlykov, Soumya Batra, Prajjwal Bhargava, Shruti Bhosale, Dan Bikel, Lukas Blecher, Cristian Canton-Ferrer, Moya Chen, Guillem Cucurull, David Esiobu, Jude Fernandes, Jeremy Fu, Wenyin Fu, Brian Fuller, Cynthia Gao, Vedanuj Goswami, Naman Goyal, Anthony Hartshorn, Saghar Hosseini, Rui Hou, Hakan Inan, Marcin Kardas, Viktor Kerkez, Madian Khabsa, Isabel Kloumann, Artem Korenev, Punit Singh Koura, Marie-Anne Lachaux, Thibaut Lavril, Jenya Lee, Diana Liskovich, Yinghai Lu, Yuning Mao, Xavier Martinet, Todor Mihaylov, Pushkar Mishra, Igor Molybog, Yixin Nie, Andrew Poulton, Jeremy Reizenstein, Rashi Rungta, Kalyan Saladi, Alan Schelten, Ruan Silva, Eric Michael Smith, Ranjan Subramanian, Xiaoqing Ellen Tan, Binh Tang, Ross Taylor, Adina Williams, Jian Xiang Kuan, Puxin Xu, Zheng Yan, Iliyan Zarov, Yuchen Zhang, Angela Fan, Melanie Kambadur, Sharan Narang, Aurélien Rodriguez, Robert Stojnic, Sergey Edunov, and Thomas Scialom. Llama 2: Open Foundation and Fine-Tuned Chat Models. arXiv: Comp. Res. Repository, abs/2307.09288, 2023. 6", + "[65] Yi Tu, Ya Guo, Huan Chen, and Jinyang Tang. LayoutMask: Enhance Text-Layout Interaction in Multi-modal Pretraining for Document Understanding. In Proc. ACL, pages 15200–15212, 2023. 1", + "[66] Jianqiang Wan, Sibo Song, Wenwen Yu, Yuliang Liu, Wenqing Cheng, Fei Huang, Xiang Bai, Cong Yao, and Zhibo Yang. OmniParser: A Unified Framework for Text Spotting Key Information Extraction and Table Recognition. In Proc. CVPR, pages 15641-15653, 2024. 2", + "[67] Dongsheng Wang, Natraj Raman, Mathieu Sibue, Zhiqiang Ma, Petr Babkin, Simerjot Kaur, Yulong Pei, Armineh Nourbakhsh, and Xiaomo Liu. DocLLM: A Layout-Aware Generative Language Model for Multimodal Document Understanding. In Proc. ACL, pages 8529-8548, 2024. 3, 6, 7", + "[68] Jiapeng Wang, Lianwen Jin, and Kai Ding. LiLT: A Simple yet Effective Language-Independent Layout Transformer for Structured Document Understanding. In Proc. ACL, pages 7747–7757, 2022. 1", + "[69] Jiapeng Wang, Zening Lin, Dayi Huang, Longfei Xiong, and Lianwen Jin. LiLTv2: Language-substitutable layout-image transformer for visual information extraction. ACM Transactions on Multimedia Computing, Communications and Applications, 21(3):1-27, 2025. 1", + "[70] Wenjin Wang, Yunhao Li, Yixin Ou, and Yin Zhang. Layout and Task Aware Instruction Prompt for Zero-shot Document" + ], + "bbox": [ + 516, + 92, + 903, + 900 + ], + "page_idx": 10 + }, + { + "type": "page_number", + "text": "4048", + "bbox": [ + 482, + 944, + 514, + 955 + ], + "page_idx": 10 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "Image Question Answering. arXiv: Comp. Res. Repository, abs/2306.00526, 2023. 1, 3", + "[71] Jason Wei, Xuezhi Wang, Dale Schuurmans, Maarten Bosma, brian richter, Fei Xia, Ed Chi, Quoc V Le, and Denny Zhou. Chain-of-Thought Prompting Elicits Reasoning in Large Language Models. Proc. NIPS, 35:24824-24837, 2022. 2", + "[72] Yiheng Xu, Minghao Li, Lei Cui, Shaohan Huang, Furu Wei, and Ming Zhou. LayoutLM: Pre-training of Text and Layout for Document Image Understanding. In Proc. ACM SIGKDD Int. Conf. Knowledge discovery & data mining, page 1192-1200, 2020. 1", + "[73] Yang Xu, Yiheng Xu, Tengchao Lv, Lei Cui, Furu Wei, Guoxin Wang, Yijuan Lu, Dinei Florencio, Cha Zhang, Wanxiang Che, Min Zhang, and Lidong Zhou. LayoutLMv2: Multi-modal Pre-training for Visually-rich Document Understanding. In Proc. ACL, pages 2579-2591, 2021. 1", + "[74] Hao Yan, Yuliang Liu, Lianwen Jin, and Xiang Bai. The development, application, and future of LLM similar to ChatGPT. Journal of Image and Graphics, 28(9):2749-2762, 2023. 1", + "[75] Jiabo Ye, Anwen Hu, Haiyang Xu, Qinghao Ye, Ming Yan, Yuhao Dan, Chenlin Zhao, Guohai Xu, Chenliang Li, Junfeng Tian, Qian Qi, Ji Zhang, and Fei Huang. mPLUGDocOwl: Modularized Multimodal Large Language Model for Document Understanding. arXiv: Comp. Res. Repository, abs/2307.02499, 2023. 7", + "[76] Jiabo Ye, Anwen Hu, Haiyang Xu, Qinghao Ye, Ming Yan, Guohai Xu, Chenliang Li, Junfeng Tian, Qi Qian, Ji Zhang, Qin Jin, Liang He, Xin Lin, and Fei Huang. UReader: Universal OCR-free Visually-situated Language Understanding with Multimodal Large Language Model. In Findings of EMNLP, pages 2841-2858, 2023. 1, 2, 7", + "[77] Chong Zhang, Ya Guo, Yi Tu, Huan Chen, Jinyang Tang, Huijia Zhu, Qi Zhang, and Tao Gui. Reading Order Matters: Information Extraction from Visually-rich Documents by Token Path Prediction. In Proc. EMNLP, pages 13716-13730, 2023. 1", + "[78] Jiaxin Zhang, Wentao Yang, Songxuan Lai, Zecheng Xie, and Lianwen Jin. DocKylin: A Large Multimodal Model for Visual Document Understanding with Efficient Visual Slimming. arXiv: Comp. Res. Repository, abs/2406.19101, 2024. 7", + "[79] Yanzhe Zhang, Ruiyi Zhang, Jiuming Gu, Yufan Zhou, Nedim Lipka, Diyi Yang, and Tong Sun. LLaVAR: Enhanced Visual Instruction Tuning for Text-Rich Image Understanding. arXiv: Comp. Res. Repository, abs/2306.17107, 2023. 1, 2, 7", + "[80] Xu Zhong, Jianbin Tang, and Antonio Jimeno-Yepes. Pub-LayNet: Largest Dataset Ever for Document Layout Analysis. In Proc. ICDAR, pages 1015-1022, 2019. 5", + "[81] Xu Zhong, Elaheh ShafieiBavani, and Antonio JimenoYepes. Image-Based Table Recognition: Data, Model, and Evaluation. Proc. ECCV, 12366:564-580, 2020. 5" + ], + "bbox": [ + 91, + 90, + 482, + 852 + ], + "page_idx": 11 + }, + { + "type": "page_number", + "text": "4049", + "bbox": [ + 482, + 944, + 514, + 955 + ], + "page_idx": 11 + } +] \ No newline at end of file diff --git a/2025/DocLayLLM_ An Efficient Multi-modal Extension of Large Language Models for Text-rich Document Understanding/587c40bc-a1fe-44db-8881-11814790688b_model.json b/2025/DocLayLLM_ An Efficient Multi-modal Extension of Large Language Models for Text-rich Document Understanding/587c40bc-a1fe-44db-8881-11814790688b_model.json new file mode 100644 index 0000000000000000000000000000000000000000..5081f899ed5362eb9a369a92db060c6a1bd72c53 --- /dev/null +++ b/2025/DocLayLLM_ An Efficient Multi-modal Extension of Large Language Models for Text-rich Document Understanding/587c40bc-a1fe-44db-8881-11814790688b_model.json @@ -0,0 +1,2644 @@ +[ + [ + { + "type": "header", + "bbox": [ + 0.107, + 0.003, + 0.182, + 0.043 + ], + "angle": 0, + "content": "CVF" + }, + { + "type": "header", + "bbox": [ + 0.237, + 0.001, + 0.812, + 0.047 + ], + "angle": 0, + "content": "This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore." + }, + { + "type": "title", + "bbox": [ + 0.108, + 0.13, + 0.892, + 0.178 + ], + "angle": 0, + "content": "DocLayLLM: An Efficient Multi-modal Extension of Large Language Models for Text-rich Document Understanding" + }, + { + "type": "text", + "bbox": [ + 0.117, + 0.202, + 0.88, + 0.275 + ], + "angle": 0, + "content": "Wenhui Liao\\(^{1\\dagger}\\), Jiapeng Wang\\(^{1,2\\dagger}\\), Hongliang Li\\(^{1}\\), Chengyu Wang\\(^{2*}\\), Jun Huang\\(^{2}\\), Lianwen Jin\\(^{1,3*}\\) \n\\(^{1}\\)South China University of Technology, Guangzhou, China \n\\(^{2}\\)Alibaba Cloud Computing, Hangzhou, China \n\\(^{3}\\)SCUT-Zhuhai Institute of Modern Industrial Innovation, Zhuhai, China" + }, + { + "type": "text", + "bbox": [ + 0.19, + 0.277, + 0.808, + 0.31 + ], + "angle": 0, + "content": "{eelwh, eejpwang, eehongliangli}@mail.scut.edu.cn, eelwjin@scut.edu.cn {chengyu.wcy, huangjun.hj}@alibaba-inc.com" + }, + { + "type": "title", + "bbox": [ + 0.248, + 0.345, + 0.327, + 0.361 + ], + "angle": 0, + "content": "Abstract" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.376, + 0.486, + 0.709 + ], + "angle": 0, + "content": "Text-rich document understanding (TDU) requires comprehensive analysis of documents containing substantial textual content and complex layouts. While Multimodal Large Language Models (MLLMs) have achieved fast progress in this domain, existing approaches either demand significant computational resources or struggle with effective multi-modal integration. In this paper, we introduce DocLayLLM, an efficient multi-modal extension of LLMs specifically designed for TDU. By lightly integrating visual patch tokens and 2D positional tokens into LLMs' input and encoding the document content using the LLMs themselves, we fully take advantage of the document comprehension capability of LLMs and enhance their perception of OCR information. We have also deeply considered the role of chain-of-thought (CoT) and innovatively proposed the techniques of CoT Pre-training and CoT Annealing. Our DocLayLLM can achieve remarkable performances with lightweight training settings, showcasing its efficiency and effectiveness. Experimental results demonstrate that our DocLayLLM outperforms existing OCR-dependent methods and OCR-free competitors. Code and model are available at https://github.com/whlscut/DocLayLLM." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.736, + 0.222, + 0.751 + ], + "angle": 0, + "content": "1. Introduction" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.762, + 0.485, + 0.868 + ], + "angle": 0, + "content": "Text-rich document understanding (TDU) focuses on analyzing documents with extensive text and complex layouts, requiring a combined understanding of text, layout, and images. Early research [1, 2, 12, 15, 20, 28, 33-37, 44, 52, 55, 65, 68, 69, 72, 73, 77] often used BERT-like [8] Transformer-based document encoders (DE), designing various pre-training tasks for large-scale unlabeled" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.346, + 0.906, + 0.466 + ], + "angle": 0, + "content": "document corpora and fine-tuning their models on specific downstream tasks like key information extraction (KIE) and visual question answering (VQA) [39, 40]. However, these pre-training tasks often diverge significantly from downstream tasks, requiring the model to acquire substantial task-specific knowledge during fine-tuning. This may result in limited generalization capability across different tasks and datasets." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.472, + 0.907, + 0.805 + ], + "angle": 0, + "content": "With the rapid development of large language models (LLMs), their impressive versatility and generalization capabilities have captured widespread interest [74]. This has led many researchers [17, 38, 61, 76, 79] to overcome the challenges of TDU with LLMs. Initially, efforts focused on employing multi-modal large language models (MLLMs) for OCR-free TDU, where document images are fed into MLLMs, directly generating responses to comprehension tasks. These MLLMs, which can process images and text inputs, are among the most recent advancements in this domain. To date, several general MLLMs have been developed to manage diverse scenarios, and researchers often build their TDU models based on these general models. However, these TDU models typically require high-resolution image inputs to recognize dense text in documents and a large volume of training data to effectively develop this capability, both of which bring significant computational demands. Furthermore, within the extensive collections of image and instruction data used to train these models, the training data for downstream evaluation tasks was often included, which might potentially impact the assessment of their generalization to previously unseen datasets." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.81, + 0.909, + 0.903 + ], + "angle": 0, + "content": "To overcome the constraints of OCR-free methods, scholars have pursued diverse methodologies for integrating OCR information into LLMs. Initial strategies [14, 70] involved inputting coordinates of OCR bounding boxes as numeric text. However, this textual representation of numbers often resulted in excessively long inputs that could" + }, + { + "type": "page_footnote", + "bbox": [ + 0.109, + 0.875, + 0.223, + 0.888 + ], + "angle": 0, + "content": "\\(\\dagger\\) Equal contribution." + }, + { + "type": "page_footnote", + "bbox": [ + 0.11, + 0.888, + 0.258, + 0.9 + ], + "angle": 0, + "content": "\\*Co-corresponding authors." + }, + { + "type": "list", + "bbox": [ + 0.109, + 0.875, + 0.258, + 0.9 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.516, + 0.957 + ], + "angle": 0, + "content": "4038" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.09, + 0.092, + 0.482, + 0.304 + ], + "angle": 0, + "content": "decrease inference speeds and potentially degrade model performance. In response, many researchers [11, 24, 45] have adopted an auxiliary document encoder like LayoutLMv3 [20] to process textual, visual, and layout features from documents and integrated into LLMs. Although these methods avoid high-resolution image inputs, the textual content of documents must also be encoded alongside the layout and image features within the DE due to its structural limitations. However, the document comprehension capabilities of these DEs may not be on par with those of LLMs themselves, which might have led to significant finetuning efforts for aligning the feature spaces of DEs and LLMs to improve performances during the whole training phase, demanding substantial computational resources." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.308, + 0.482, + 0.504 + ], + "angle": 0, + "content": "To address the limitations of OCR-dependent TDU methods, we introduce DocLayLLM, an efficient multimodal extension of large language models for text-rich document understanding. Our method diverges from existing approaches that utilize an extra DE by integrating the document's image patch tokens and 2D positional tokens with textual content into a natural language expression, and then encoding them through a unified LLM. This innovative integration not only taps into the LLM's inherent text comprehension abilities but also allows it to perceive document layouts. Besides, by incorporating patch tokens, the model supports image-related pre-training tasks and further gains a rudimentary understanding of document structures." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.509, + 0.483, + 0.902 + ], + "angle": 0, + "content": "To advance model performance, we have fully integrated the concept of the chain-of-thought (CoT) [3] into each stage of our training process. Firstly, we developed a novel CoT Pre-training strategy. CoT is engineered to preserve thematic coherence and logical progression during model reasoning, further enhancing the inference ability of LLM, which has proven effectiveness in various subsequent studies [14, 45, 71]. Recently, the use of Long CoT reasoning to enhance the inference capability of models has attracted significant attention [7, 49, 63], and DeepSeek-R1 [7] further emphasized the importance of utilizing CoT data for a cold start before the reinforcement learning stage. Our proposed CoT Pre-training offers a method for rapidly generating large-scale CoT data from non-QA data, thereby improving the model's reasoning ability. By employing CoT in the data generation process and subsequently pre-training the model with the generated QA-formatted data, we have aligned the added visual and layout features with LLM's inherent textual features while improving the model's comprehension ability. Furthermore, we introduced a CoT Annealing technique. While previous research [31, 42] has validated the effectiveness of annealing strategies that gradually increase the proportion of high-quality data towards the end of training, the impact of CoT on data quality had not been explored. To address this, we first propose an annealing strategy that considers data quality from a CoT per" + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.093, + 0.831, + 0.107 + ], + "angle": 0, + "content": "spective, further improving model performance." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.108, + 0.905, + 0.287 + ], + "angle": 0, + "content": "Thanks to the above design, utilizing only around 3M pre-training data and 300K supervised fine-tuning (SFT) data with the efficient tuning technology of LoRA [18], our DocLayLLM finishes the whole training process in 36 A100 days, therefore resulting an efficient extension of LLM. Even with such low resources, our DocLayLLM surpasses existing OCR-dependent TDU methods. Furthermore, it also excels in specific tasks without fine-tuning on downstream tasks and outperforms state-of-the-art (SOTA) OCR-free methods when trained with training sets of downstream tasks as other methods, further showcasing its effectiveness. Our contributions are:" + }, + { + "type": "text", + "bbox": [ + 0.514, + 0.29, + 0.905, + 0.335 + ], + "angle": 0, + "content": "- We propose an efficient multi-modal extension of LLMs, which has augmented LLM in understanding text-rich documents and greatly reduced training resources in need." + }, + { + "type": "text", + "bbox": [ + 0.514, + 0.336, + 0.905, + 0.379 + ], + "angle": 0, + "content": "- We introduce CoT Pre-training, a novel approach for generating CoT data that helps enhance the comprehension capabilities of models." + }, + { + "type": "text", + "bbox": [ + 0.514, + 0.381, + 0.905, + 0.44 + ], + "angle": 0, + "content": "- We revisit data annealing from a novel perspective and introduce the CoT Annealing strategy, which further enhances the model's efficiency in data utilization and boosts its ultimate performance." + }, + { + "type": "text", + "bbox": [ + 0.514, + 0.441, + 0.905, + 0.486 + ], + "angle": 0, + "content": "- Our method outperforms existing OCR-dependent TDU solutions and surpasses current OCR-free methods under comparable supervised fine-tuning conditions." + }, + { + "type": "list", + "bbox": [ + 0.514, + 0.29, + 0.905, + 0.486 + ], + "angle": 0, + "content": null + }, + { + "type": "title", + "bbox": [ + 0.514, + 0.498, + 0.663, + 0.514 + ], + "angle": 0, + "content": "2. Related Works" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.524, + 0.905, + 0.903 + ], + "angle": 0, + "content": "OCR-free MLLMs for Text-Rich Documents Understanding. To date, a number of studies have attempted to directly answer the instruction for TDU tasks from document images, thereby obviating the need for extra OCR processing and resulting in an end-to-end paradigm. Such models are referred to as OCR-free methods. Early examples like Donut [23], UDOP [62], and OmniParser [66] have explored end-to-end document understanding solutions without relying on LLMs or MLLMs. However, with the rapid development of MLLMs, many new studies have begun to focus on enhancing the capabilities of general MLLMs in document understanding. Some of them attempt to boost the document understanding ability of general MLLMs by tuning them with synthesizing TDU data generated by proprietary LLMs, including LLaVAR [79] and TextSquare [61], while mPLUG-DocOwl 1.5 [17] enhances the document understanding capability by integrating existing document comprehension data and converting it into vision-related QA-formatted data. Moreover, researchers have also explored image encoding methods to read the text more clearly. For instance, UReader [76] and Monkey [38] empower MLLMs to handle high-resolution images by splitting them into sub-images, encoding each separately, and then concatenating them. Some general MLLMs, such as LLaVA-Next [29], LLaVA" + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.516, + 0.956 + ], + "angle": 0, + "content": "4039" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.143, + 0.087, + 0.856, + 0.33 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.346, + 0.339, + 0.651, + 0.354 + ], + "angle": 0, + "content": "Figure 1. The overall architecture of DocLayLLM." + }, + { + "type": "text", + "bbox": [ + 0.089, + 0.38, + 0.483, + 0.547 + ], + "angle": 0, + "content": "OneVision [27], and Idefics3 [26], also adopt this approach to address the challenge of TDU. Although these above methods have achieved notable success in TDU tasks, they typically require high-resolution image inputs to identify text within images, significantly increasing computational demands. Besides, to equip general MLLMs with the ability to recognize the rich text within a document, a substantial amount of training data is required, including training sets of the benchmarks tested, which resulted in a high demand for training resources and might affect the evaluation of their zero-shot inference ability." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.554, + 0.483, + 0.903 + ], + "angle": 0, + "content": "OCR-dependent MLLMs for Text-Rich Documents Understanding. To address the challenges of OCR-free MLLMs, some researchers have attempted to integrate OCR information from text-rich documents into LLMs to avoid high-resolution image inputs. ICL-D3IE [14] feeds coordinates of OCR boxes into the LLM using a textual expression of [left, top, right, bottom], while LATIN-Prompt [70] simulates the document layout by inserting spaces and newline characters between OCR-recognized text segments, thus providing the LLM with a document input containing a rudimentary layout. Besides, LMDX [53] also input the layout in a way similar to ICL-D3IE. However, these methods of inputting OCR information in textual form lead to overly lengthy inputs, thereby slowing inference and impacting model performance. Subsequent studies have explored encoding OCR information using additional encoders. Cream [24] utilizes BLIP-2 [30] to encode document OCR positions, images, text, and additional object detection information, with the output of BLIP-2 fed into the LLM. InstructDr [60] introduces an extra feature extractor to encode document OCR information, compressing it with learnable tokens before inputting it into the LLM. Additionally, Fujitake [11] and Luo et al. [45] adopt Lay" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.38, + 0.908, + 0.729 + ], + "angle": 0, + "content": "outLMv3 [20] to encode the OCR information of documents before inputting them into the LLM. Although these methods avoid the issue of overly lengthy text inputs, they must use the extra DE to encode all the OCR textual, positional information from the document due to the structural restriction of DE. However, the additional DE might not be comparable with LLMs themselves in terms of document comprehension capability. This might be the reason why they usually need extensive training of both the DE and the LLM to align the features encoded by the DE with those of the LLM, which consumes significant computational resources. DocLLM [67] proposes an extension where the hidden states derived from OCR bounding boxes interact with those of OCR text features encoded by LLMs within each attention block via cross-attention [5, 6]. However, they use an extra projection layer at each attention block to transform and inject layout information, which might excessively interfere with the original attention computation of LLM, thereby potentially affecting LLM's generalization capability. Moreover, it does not adequately consider the role of image features in TDU. Overall, designing an efficient method to provide OCR information to LLMs remains a research question worthy of deep exploration." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.74, + 0.651, + 0.758 + ], + "angle": 0, + "content": "3. DocLayLLM" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.765, + 0.907, + 0.903 + ], + "angle": 0, + "content": "DocLayLLM is an LLM-based methodology that encodes OCRed textual, visual, and positional information directly within an LLM, eliminating the need for additional DE. The training regime of DocLayLLM includes both pretraining and supervised fine-tuning phases, which are both instruction-tuning-based processes. To further enhance our model, we propose the integration of CoT into our pretraining tasks, denoted as CoT Pre-training. Additionally, we introduce CoT Annealing, which involves the applica" + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.516, + 0.956 + ], + "angle": 0, + "content": "4040" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.127, + 0.089, + 0.359, + 0.248 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.173, + 0.253, + 0.31, + 0.264 + ], + "angle": 0, + "content": "(a) Layout-Type-Related CoT." + }, + { + "type": "image", + "bbox": [ + 0.37, + 0.089, + 0.616, + 0.248 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.429, + 0.253, + 0.571, + 0.263 + ], + "angle": 0, + "content": "(b) Table-Structure-Aware CoT." + }, + { + "type": "image", + "bbox": [ + 0.625, + 0.089, + 0.874, + 0.248 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.696, + 0.253, + 0.819, + 0.264 + ], + "angle": 0, + "content": "(c) Geometry-Related CoT." + }, + { + "type": "image_caption", + "bbox": [ + 0.116, + 0.275, + 0.879, + 0.29 + ], + "angle": 0, + "content": "Figure 2. Examples of CoT Pre-training across different tasks. Red text highlights key steps in the chain of thought for each task." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.316, + 0.483, + 0.347 + ], + "angle": 0, + "content": "tion of annealing to the data utilized in the supervised finetuning phase, specifically from the perspective of CoT." + }, + { + "type": "title", + "bbox": [ + 0.09, + 0.353, + 0.279, + 0.368 + ], + "angle": 0, + "content": "3.1. Model Architecture" + }, + { + "type": "text", + "bbox": [ + 0.089, + 0.375, + 0.483, + 0.663 + ], + "angle": 0, + "content": "The overall architecture of our DocLayLLM is illustrated in Figure 1. To process a document image, we first employ an OCR engine. This engine extracts the OCR results, which encompasses the textual content, denoted as \\(Text_{1:N}\\), and the associated bounding box, represented by \\(Box_{1:N}\\). Here, \\(N\\) indicates the total number of text segments detected by the OCR, and each \\(Box_{i} = [x_{0},y_{0},x_{1},y_{1}]\\) defines the bounding box for \\(Text_{i}\\). The bounding box coordinates correspond to the left, top, right, and bottom boundaries of the text segment. After the OCR process, the document image is segmented into patches, denoted as \\(Patch_{1:M}\\), where \\(M\\) represents the total number of patches. Next, the OCR results and the document patches are integrated with the fixed part of the prompt and the specific Question related to the document. This integration forms the final prompt input to the model, presented in a natural language expression \\(\\mathcal{T}\\): Given the document patches: \\(\\{Patch_{1:M}\\}\\) and the document text contents and locations in the form of \"text, [left, top, right, bottom]\": \\(\\{(Text_{i}, Box_{i})_{1:N}\\}\\) \\(\\{Question\\}\\)" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.663, + 0.483, + 0.723 + ], + "angle": 0, + "content": "Subsequently, the textual components of the prompt, including the fixed expression, \\(Text_{1:N}\\), and \\(Question\\), are processed using the LLM's native text embedder. This generates the embeddings for the text content as follows:" + }, + { + "type": "equation", + "bbox": [ + 0.184, + 0.729, + 0.482, + 0.745 + ], + "angle": 0, + "content": "\\[\nE m b _ {T} = T E (T e x t C o n t e n t), \\tag {1}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.75, + 0.483, + 0.796 + ], + "angle": 0, + "content": "where \\( TE \\) is the native text embedder of LLM, and TextContent consists of the fixed expression, \\( Text_{1:N} \\), and Question." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.796, + 0.483, + 0.843 + ], + "angle": 0, + "content": "As for \\( Patch_{1:M} \\) and \\( Box_{1:N} \\), they are first embedded with an extra embedder, then mapped to the LLM's feature space via a linear projector, yielding \\( Emb_V \\) and \\( Emb_L \\) as:" + }, + { + "type": "equation", + "bbox": [ + 0.176, + 0.862, + 0.483, + 0.878 + ], + "angle": 0, + "content": "\\[\nE m b _ {V} = V P (V E (P a t c h _ {1: M})), \\tag {2}\n\\]" + }, + { + "type": "equation", + "bbox": [ + 0.179, + 0.882, + 0.482, + 0.897 + ], + "angle": 0, + "content": "\\[\nE m b _ {L} = L P \\left(L E \\left(\\text {B o x} _ {1: N}\\right)\\right), \\tag {3}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.316, + 0.905, + 0.406 + ], + "angle": 0, + "content": "where \\(VE\\) and \\(VP\\) represent the visual embedder and visual projector, \\(LE\\) and \\(LP\\) denote the layout embedder and layout projector, respectively. It is worth noting that despite the introduction of \\(VE\\), \\(VP\\), \\(LE\\), and \\(LP\\), these are all simple embedding layers or linear layers, thereby minimizing additional computational burden." + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.406, + 0.905, + 0.437 + ], + "angle": 0, + "content": "Finally, LLM will comprehend the embeddings provided and utilize them to generate the Answer to the Question:" + }, + { + "type": "equation", + "bbox": [ + 0.528, + 0.445, + 0.905, + 0.462 + ], + "angle": 0, + "content": "\\[\n\\text {A n s w e r} = L L M \\left(\\pi \\left(\\left(E m b _ {T}, E m b _ {V}, E m b _ {L}\\right), \\mathcal {T}\\right)\\right). \\tag {4}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.47, + 0.905, + 0.5 + ], + "angle": 0, + "content": "Here \\(\\pi\\) represents organizing \\(Emb_{T}, Emb_{V}, Emb_{L}\\) into a feature sequence according to \\(\\mathcal{T}\\)." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.501, + 0.906, + 0.591 + ], + "angle": 0, + "content": "With our design, we have preserved the original structure of the LLM, ensuring its generalization capability. At the same time, we have enabled the LLM to easily accept and process visual and layout features, making it better suited for understanding text-rich documents with various layouts. This results in a simple yet effective multi-modal extension." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.598, + 0.685, + 0.614 + ], + "angle": 0, + "content": "3.2. CoT Pre-training" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.62, + 0.907, + 0.848 + ], + "angle": 0, + "content": "During the pre-training phase, we incorporate a wide range of tasks (as detailed in the appendix) to boost the LLM's understanding of the visual and layout features inspired by Luo et al. [45]. Since the datasets related to these tasks are not designed explicitly for instruction-tuning, their annotations are not in a QA format. Therefore, we manually create QA-formatted instructions based on their annotations for training. However, simply converting task descriptions into Question and annotations into Answer is highly inefficient regarding data utilization. To address this, we incorporate CoT during the conversion to instruction-tuning data for specific tasks, asking LLM to solve our problem step-by-step thus enhancing the model's comprehension abilities, as the examples illustrated in Fig 2. Our designed CoT primarily falls into the following three categories:" + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.852, + 0.743, + 0.868 + ], + "angle": 0, + "content": "3.2.1. Layout-Type-Related CoT" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.871, + 0.906, + 0.902 + ], + "angle": 0, + "content": "We conduct a layout-type-related CoT for the layout analysis task. To determine the layout type of a given area, it is" + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.514, + 0.957 + ], + "angle": 0, + "content": "4041" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.09, + 0.092, + 0.482, + 0.152 + ], + "angle": 0, + "content": "beneficial first to locate its region within the document and then infer based on the nearby layout element. When giving an area bounded by a bounding box, our layout-type-related CoT can be formulated as following steps:" + }, + { + "type": "text", + "bbox": [ + 0.092, + 0.153, + 0.483, + 0.197 + ], + "angle": 0, + "content": "1. Identify the text within the area and determine its approximate region within the document (e.g., top left, top right part of the document, etc)." + }, + { + "type": "text", + "bbox": [ + 0.091, + 0.198, + 0.482, + 0.228 + ], + "angle": 0, + "content": "2. Locate the nearest bounding boxes and assess their layout type." + }, + { + "type": "text", + "bbox": [ + 0.091, + 0.228, + 0.482, + 0.256 + ], + "angle": 0, + "content": "3. Infer the layout type of the given area based on the above information." + }, + { + "type": "list", + "bbox": [ + 0.091, + 0.153, + 0.483, + 0.256 + ], + "angle": 0, + "content": null + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.258, + 0.483, + 0.318 + ], + "angle": 0, + "content": "Through this CoT process, we enable the LLM to fully utilize the visual, layout, and textual features of the document, thereby improving their ability to understand the document's overall structure." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.327, + 0.333, + 0.34 + ], + "angle": 0, + "content": "3.2.2. Table-Structure-Aware CoT" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.345, + 0.483, + 0.42 + ], + "angle": 0, + "content": "Table-structure-aware CoT is primarily employed in the table analysis task. We have integrated the XY-Cut algorithm [48]—a widely-used technique for analyzing table structures. For instance, to locate an element at the \\(i\\)-th row and \\(j\\)-th column, the CoT process is outlined as follows:" + }, + { + "type": "text", + "bbox": [ + 0.092, + 0.421, + 0.483, + 0.465 + ], + "angle": 0, + "content": "1. Identify all the table headers and their respective locations. The header of the \\(j\\)-th column corresponds to the \\(j\\)-th element in the headers list." + }, + { + "type": "text", + "bbox": [ + 0.092, + 0.467, + 0.483, + 0.512 + ], + "angle": 0, + "content": "2. Output all elements in the \\(j\\)-th column when told that the elements of the same column are roughly vertically aligned with their corresponding headers." + }, + { + "type": "text", + "bbox": [ + 0.092, + 0.512, + 0.483, + 0.542 + ], + "angle": 0, + "content": "3. Determine the final answer based on the \\(i\\)-th element of the contents in the \\(j\\)-th column." + }, + { + "type": "list", + "bbox": [ + 0.092, + 0.421, + 0.483, + 0.542 + ], + "angle": 0, + "content": null + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.542, + 0.483, + 0.618 + ], + "angle": 0, + "content": "Applying the above CoT, we break down a complex question into simple steps using the basic XY-Cut thought. This approach improves the LLM's inference ability through this breakdown and enhances data utilization by effectively leveraging the overall structure of the table." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.625, + 0.303, + 0.64 + ], + "angle": 0, + "content": "3.2.3. Geometry-Related CoT" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.644, + 0.483, + 0.704 + ], + "angle": 0, + "content": "This is used in geometric analysis to identify geometric relationships between two text segments. It assists LLM to inference with geometric relations. Specifically, when giving the text content of two lines:" + }, + { + "type": "text", + "bbox": [ + 0.092, + 0.705, + 0.466, + 0.719 + ], + "angle": 0, + "content": "1. Recover the bounding box of the given text segments." + }, + { + "type": "text", + "bbox": [ + 0.092, + 0.72, + 0.483, + 0.75 + ], + "angle": 0, + "content": "2. Analyze their vertical and horizontal projections to determine if they overlap through interval analysis." + }, + { + "type": "text", + "bbox": [ + 0.092, + 0.751, + 0.483, + 0.779 + ], + "angle": 0, + "content": "3. Calculate the center point of each segment and determine their relative location by comparing their center points." + }, + { + "type": "text", + "bbox": [ + 0.091, + 0.781, + 0.483, + 0.869 + ], + "angle": 0, + "content": "4. Calculate the minimum Euclidean distance based on the previous analysis. If they overlap, the distance is 0; if they overlap in vertical/horizontal projections, the minimum distance is the smallest interval in the horizontal/vertical projections; otherwise, it is the minimum Euclidean distance between their nearest corners." + }, + { + "type": "list", + "bbox": [ + 0.091, + 0.705, + 0.483, + 0.869 + ], + "angle": 0, + "content": null + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.871, + 0.483, + 0.902 + ], + "angle": 0, + "content": "This design advances the LLM's arithmetic reasoning capabilities and helps it more accurately grasp the spatial re" + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.092, + 0.905, + 0.122 + ], + "angle": 0, + "content": "relationships between different boxes, and also contributes to layout-type-aware CoT to find the nearest bounding boxes." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.137, + 0.669, + 0.153 + ], + "angle": 0, + "content": "3.3. CoT Annealing" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.161, + 0.905, + 0.252 + ], + "angle": 0, + "content": "Data annealing involves gradually increasing the proportion of high-quality data towards the end of supervised fine-tuning to improve data utilization efficiency and augment the model's reasoning abilities. This technique has been proven to be effective and is widely adopted in models like MiniCPM [19], DataComp-LM [31], and Llama-3.1 [42]." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.254, + 0.907, + 0.42 + ], + "angle": 0, + "content": "During the SFT stage, we employ data with a layout-aware CoT [45] that helps focus on potential regions of questions and answers within documents, thereby better leveraging the layout features learned during pre-training. It also allows the model to strengthen its reasoning capacity through step-by-step thinking. However, in the latter stages of SFT, our goal for LLM is to output precise and direct answers to questions. At this point, using CoT data may introduce unnecessary noise with a long response, and the overly uniform layout-aware CoT data might partially compromise the LLM's generalization abilities." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.422, + 0.906, + 0.588 + ], + "angle": 0, + "content": "To balance the strengths and weaknesses of such data, we propose CoT annealing, which reconsiders this data annealing issue from a CoT perspective. Specifically, we generate direct-answer data without CoT from the CoT data, treating it as high-quality data. During the SFT process, we gradually increase the proportion of this direct-answer data, thereby enhancing the LLM's ability to respond with a clear and direct answer in the end. Initially, we only utilize the data with CoT, gradually adjusting the ratio of CoT to non-CoT data as the SFT process progresses. Ultimately, we conclude the training using only the data without CoT." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.607, + 0.645, + 0.624 + ], + "angle": 0, + "content": "4. Experiments" + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.634, + 0.615, + 0.648 + ], + "angle": 0, + "content": "4.1. Datasets" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.658, + 0.906, + 0.809 + ], + "angle": 0, + "content": "During the pre-training stage, we conduct our pre-training tasks on several datasets, including RVL-CDIP [13], PubLayNet [80], PubTabNet [81], DocLayNet [54], DocBank [32], DocILE [56], and the Document Dense Description data from LayoutLLM [45]. This results in a total of 3.1 million pairs of instruction tuning data, which is significantly less than the volume used in previous OCR-dependent TDU methods. For the supervised fine-tuning stage, we apply CoT annealing on the layout-aware document QA data utilized in LayoutLLM [45]." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.811, + 0.906, + 0.901 + ], + "angle": 0, + "content": "We evaluate our DocLayLLM on widely used text-rich document benchmarks, including FUNSD [22], CORD [50], SROIE [21], POIE [25], DocVQA [46], InfoVQA [47], VisualMRC [59], DeepForm [58], KLC [57], WTQ [51], and TabFact [4]. For a fair comparison with other approaches, we use the default OCR data from the of" + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.516, + 0.957 + ], + "angle": 0, + "content": "4042" + } + ], + [ + { + "type": "table", + "bbox": [ + 0.173, + 0.089, + 0.828, + 0.263 + ], + "angle": 0, + "content": "
MethodProcessed Data at Pre-trainingDocument-oriented VQAKIE
DocVQAVisualMRCAvg.FUNSD†CORD†SROIE†Avg.
Text
Llama2-7B-Chat [64]-20.509.9015.2015.1020.0035.623.57
Llama3-8B-Instruct [42]-51.7947.7749.7868.5652.3161.2460.70
Text + Box
Llama2-7B-Chatcoor [14]-12.3012.2012.2511.906.4039.4019.23
Llama3-8B-Instructcoor [14]-49.1341.7545.4474.0062.2063.1566.45
LayTextLLMzero [43]10.0M (323%)65.5037.4051.4572.0045.5082.0066.50
Text + Box + Patch
LayoutLLM Llama2 [45]5.7M (184%)74.2555.7364.9978.6562.2170.9770.61
DocLayLLM Llama2 (Ours)3.1M (100%)72.8355.0463.9478.6770.8183.2777.58
DocLayLLM Llama3 (Ours)3.1M (100%)78.3658.5568.4684.1171.3484.3679.94
" + }, + { + "type": "table_caption", + "bbox": [ + 0.09, + 0.272, + 0.907, + 0.342 + ], + "angle": 0, + "content": "Table 1. Comparison with OCR-dependent methods using metrics from LayoutLLM [45]. Processed Data at Pre-training represents the total data volume processed at the pre-training stage, equivalent to dataset size multiplied by epoch count, with values in parentheses representing size ratios compared to DocLayLLM. Avg. represents the average performance across different benchmarks, and \\(\\dagger\\) indicates the cleaned test set used in LayoutLLM. The suffix layout denotes using method from He et al. [14], which provides LLM with layout features via OCR box coordinates in textual form. Bold indicates the best performance, and underline indicates the second-best one." + }, + { + "type": "table", + "bbox": [ + 0.172, + 0.356, + 0.828, + 0.543 + ], + "angle": 0, + "content": "
MethodProcessed Data at Pre-trainingDocument-oriented VQAKIE
DocVQAVisualMRCAvg.FUNSDCORDSROIEAvg.
Text
Llama2-7B-Chat [64]-20.506.3013.4023.4051.8058.6044.60
Llama3-8B-Instruct [42]-51.79229.74140.7766.6774.7182.5174.63
Text + Box
Llama2-7B-Chatcoor [14]-12.3028.0020.1514.4038.1050.6034.37
Llama3-8B-Instructcoor [14]-49.13211.69130.4174.7175.3375.9375.32
LMDX [53]-----66.95--
DocLLM [67]16.8M (542%)69.50*264.10*166.80*51.80*67.60*91.90*70.43*
LayTextLLMzero [43]10.0M (323%)65.50200.20132.8547.2077.2083.8069.40
Text + Box + Patch
DocLayLLM Llama2 (Ours)3.1M (100%)72.83310.60191.7280.7479.3784.3781.49
DocLayLLM Llama3 (Ours)3.1M (100%)78.35357.04217.7086.4783.6686.0885.40
" + }, + { + "type": "table_caption", + "bbox": [ + 0.09, + 0.552, + 0.907, + 0.581 + ], + "angle": 0, + "content": "Table 2. Comparison with OCR-dependent methods using metrics from DocLLM [67]. * indicates that the model uses the training set of the benchmark during training." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.607, + 0.483, + 0.639 + ], + "angle": 0, + "content": "ficial datasets otherwise commercial OCR engines \\( {}^{1} \\) if not provided, and follow their evaluation pipeline and metrics." + }, + { + "type": "title", + "bbox": [ + 0.09, + 0.646, + 0.27, + 0.661 + ], + "angle": 0, + "content": "4.2. Implement Details" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.668, + 0.483, + 0.88 + ], + "angle": 0, + "content": "Our method is based on Llama3-8B-Instruct [42], with a Llama2-7B-Chat [64] version for better comparison with other methods. They are denoted as DocLayLLM\\(_{Llama3}\\) and DocLayLLM\\(_{Llama2}\\) respectively. We initialize \\(VE\\) and \\(LE\\) with weights from LayoutLMv3 [20] to leverage document knowledge acquired during its pre-training, while \\(VP\\) and \\(LP\\) are randomly initialized. Besides, the image resolution is fixed at \\(224*224\\), resulting in 196 visual patch tokens, to capture an overview of document layout while bringing less computation burden. During pre-training and fine-tuning, we use LoRA [18] with a rank of 64 on LLM and keep \\(VE\\), \\(VP\\), \\(LE\\), and \\(LP\\) trainable, leading to few tuning parameters. We train our model on 8 NVIDIA Tesla A100-80G GPUs, taking about 30 A100 days for pre" + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.607, + 0.905, + 0.639 + ], + "angle": 0, + "content": "training and 6 A100 days for supervised fine-tuning. Additional details can be found in the appendix." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.649, + 0.907, + 0.68 + ], + "angle": 0, + "content": "4.3. Comparisons with SOTA OCR-dependent methods" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.688, + 0.906, + 0.748 + ], + "angle": 0, + "content": "We have compared our DocLayLLM with existing SOTA OCR-dependent methods. In order to exhibit the zero-shot inferring ability precisely, we chose their performances under the zero-shot scenario unless they are not provided." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.749, + 0.906, + 0.825 + ], + "angle": 0, + "content": "The experimental results are shown in Table 1 and Table 2. Our method outperforms the baseline models, regardless of their type or whether the document's layout information is provided to them. It validates the rationale behind our design of multi-modal extension." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.826, + 0.906, + 0.902 + ], + "angle": 0, + "content": "Additionally, our method significantly outperforms previous SOTA methods despite fewer training resources. As shown in Table 1, it surpasses the performance of Layout-LLM [45] and LayTextLLM [43] though requiring fewer computational resources by using less data and tuning few" + }, + { + "type": "page_footnote", + "bbox": [ + 0.11, + 0.887, + 0.246, + 0.9 + ], + "angle": 0, + "content": "1 https://www.textin.com/" + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.516, + 0.957 + ], + "angle": 0, + "content": "4043" + } + ], + [ + { + "type": "table", + "bbox": [ + 0.197, + 0.09, + 0.805, + 0.243 + ], + "angle": 0, + "content": "
MethodDocument-oriented VQAKIE
DocVQAInfoVQAAvg.FUNSDSROIEPOIEAvg.
LLaVAR [79]12.3016.5014.400.505.205.903.87
UniDoc [10]7.7014.7011.201.002.905.103.00
DocPedia [9]47.10*15.20*31.15*29.9021.4039.9030.40
Monkey [38]50.10*25.80*37.95*24.1041.9019.9028.63
TextMonkey [41]66.70*28.60*47.65*42.9046.2032.0040.37
DocOwl 1.5 [17]----48.3051.80-
TextSquare [61]----53.2071.80-
DocKylin [78]65.10*34.80*49.95*58.7025.5049.5044.57
DocLayLLM\\( _{Llama3} \\) (Ours)77.7942.0259.9180.2876.5975.1377.33
" + }, + { + "type": "table_caption", + "bbox": [ + 0.09, + 0.254, + 0.907, + 0.281 + ], + "angle": 0, + "content": "Table 3. Comparison with OCR-free methods using the precision metric from TextMonkey [41]. * indicates that the model uses the training set of the benchmark during training." + }, + { + "type": "table", + "bbox": [ + 0.14, + 0.296, + 0.859, + 0.457 + ], + "angle": 0, + "content": "
MethodDocument-oriented VQAKIETable Understanding
DocVQAInfoVQAVisualMRCAvg.DeepFormKLCAvg.WTQTabFactAvg.
DocPedia [9]47.1015.20--------
DocOwl [75]62.2038.20188.8096.4042.6030.3036.4526.9060.2043.55
URector [76]65.4042.20221.70109.7749.5032.8041.1529.4067.6048.50
Monkey [38]66.5036.10--40.6032.8036.7025.30--
TextMonkey [41]73.00---59.7037.8048.7531.90--
DocKylin [78]77.3046.60319.90147.9364.20--32.40--
CogAgent [16]81.6044.50--------
DocOwl 1.5 [17]82.2050.70246.40126.4368.8038.7053.7540.6080.2060.40
TextSquare [61]84.3051.50-----49.7084.2066.95
DocLayLLM\\(\\ddagger\\) Llama3 (Ours)86.5258.36327.91157.6077.0740.7358.9058.6283.3670.99
" + }, + { + "type": "table_caption", + "bbox": [ + 0.09, + 0.468, + 0.907, + 0.498 + ], + "angle": 0, + "content": "Table 4. Comparison with OCR-free methods using the evaluation framework and metrics from DocOwl 1.5 [17]. All methods, including our \\(DocLayLLM_{Llama3}^{\\ddagger}\\) variant, are exposed to the train set corresponding to the evaluation set during training." + }, + { + "type": "text", + "bbox": [ + 0.089, + 0.523, + 0.482, + 0.599 + ], + "angle": 0, + "content": "parameters at training. Even with Llama2 as the backbone, our approach matches the performance of previous SOTA methods in the Document-oriented VQA task and significantly outperforms them in the KIE tasks, further demonstrating the efficiency and effectiveness of our approach." + }, + { + "type": "text", + "bbox": [ + 0.089, + 0.599, + 0.483, + 0.719 + ], + "angle": 0, + "content": "Besides, our method exhibits robust zero-shot generalization capability without fine-tuning with specific downstream tasks. From the comparison with DocLLM [67] in Table 2, it is clear that our method, even without fine-tuning, significantly outperforms DocLLM across most datasets and achieves a considerable overall performance improvement. This validates our design, which has preserved the strong generalization ability of the original LLM structure." + }, + { + "type": "text", + "bbox": [ + 0.089, + 0.719, + 0.483, + 0.81 + ], + "angle": 0, + "content": "In summary, DocLayLLM serves as an efficient extension of LLMs. It demonstrates strong zero-shot generalization and effective document understanding by encoding documents within a unified model. Even with limited training resources, DocLayLLM achieves new SOTA performance compared to OCR-dependent competitors." + }, + { + "type": "title", + "bbox": [ + 0.09, + 0.819, + 0.468, + 0.835 + ], + "angle": 0, + "content": "4.4. Comparisons with SOTA OCR-free methods" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.84, + 0.483, + 0.901 + ], + "angle": 0, + "content": "We have also compared our method with SOTA OCR-free TDU models. Table 3 presents the precision of our method and some advent OCR-free TDU methods across various benchmarks. Even though in the circumstance that OCR-" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.523, + 0.905, + 0.614 + ], + "angle": 0, + "content": "free models use some of the training sets of these benchmarks during training, while our method infers in a purely zero-shot manner, our DocLayLLM still outperforms these models, particularly in KIE task, which heavily relies on layout information. This clearly demonstrates the superiority of our design to provide LLM with OCR information." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.614, + 0.906, + 0.854 + ], + "angle": 0, + "content": "We additionally test DocLayLLM on a broader set of benchmarks from DocOwl 1.5 [17]. To further assess the performance of DocLayLLM on seen datasets, we incorporate the training sets of these benchmarks used in DocOwl 1.5 during the SFT process as other OCR-free TDU methods do during training. To maintain the efficiency advantage of our method, we ensure that the SFT is conducted with the same number of steps as the standard DocLayLLM. The results in Table 4 show that with the inclusion of downstream task data during SFT, our method outperforms the SOTA OCR-free TDU models in most datasets. Moreover, our overall performance in the Document-oriented VQA, KIE, and Table Understanding sub-tasks surpasses the OCR-free TDU methods using limited training resources, showcasing the tremendous potential of OCR-dependent methods and further reinforcing the reliability of our method." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.856, + 0.905, + 0.901 + ], + "angle": 0, + "content": "Overall, our DocLayLLM consistently shows strong zero-shot inference ability, outperforming many OCR-free TDU methods without prior exposure to downstream data." + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.516, + 0.956 + ], + "angle": 0, + "content": "4044" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.104, + 0.089, + 0.473, + 0.306 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.09, + 0.317, + 0.484, + 0.349 + ], + "angle": 0, + "content": "Figure 3. Qualitative comparisons of DocLayLLM with other OCR-dependent and OCR-free methods. Zoom in for better view." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.371, + 0.484, + 0.434 + ], + "angle": 0, + "content": "Additionally, when incorporating downstream datasets during SFT like other OCR-free approaches, DocLayLLM surpasses them, highlighting its potential in solving TDU tasks by integrating OCR information into the LLM." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.44, + 0.276, + 0.456 + ], + "angle": 0, + "content": "4.5. Qualitative Results" + }, + { + "type": "text", + "bbox": [ + 0.089, + 0.462, + 0.483, + 0.645 + ], + "angle": 0, + "content": "Figure 3 shows some qualitative examples to visualize the performance on KIE tasks of our DocLayLLM compared with LayTextLLM and DocOwl 1.5. The example from SROIE clearly shows that DocLayLLM has improved document comprehension and answer accuracy. And the example from POIE demonstrates our advantage in handling table-formatted documents to give an accurate response. Furthermore, as seen in the FUNSD examples, DocLayLLM excels at extracting answers spanning multiple text lines, outperforming both OCR-dependent and OCR-free methods. For more visualization of qualitative examples, please refer to the appendix." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.653, + 0.245, + 0.669 + ], + "angle": 0, + "content": "4.6. Ablation Study" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.675, + 0.483, + 0.78 + ], + "angle": 0, + "content": "We have also conducted additional experiments to explore the effectiveness of our proposed method. To conserve computational resources, we perform pre-training at \\(25\\%\\) of the standard scale (including the baseline) but full-scale SFT. We focus on the impact of two main factors: (a) the role of CoT pre-training and CoT annealing, and (b) the influence of different modality configurations." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.781, + 0.483, + 0.886 + ], + "angle": 0, + "content": "Experimental results in Table 5a indicate that omitting CoT Pre-training leads to an obvious performance drop in both the Document-oriented VQA and KIE tasks, while not using CoT Annealing primarily affects VQA performance. Besides, using randomly initialized \\(VE\\) and \\(LE\\) partially compromises the proposed extension, resulting in performance degradation in both the VQA and KIE tasks." + }, + { + "type": "text", + "bbox": [ + 0.11, + 0.886, + 0.483, + 0.902 + ], + "angle": 0, + "content": "Additionally, performances in Table 5b show the efficacy" + }, + { + "type": "table", + "bbox": [ + 0.528, + 0.09, + 0.895, + 0.154 + ], + "angle": 0, + "content": "
ExperimentDocument-oriented VQAKIE
DocVQAVisualMRCAvg.DeepFormKLCAvg.
DocLayLLM1lama377.43316.27196.8535.4126.2530.83
w/o CoT Pre-training76.00295.49185.7524.5425.4925.02
w/o CoT Annealing77.40310.17193.7934.2326.2130.22
w/o pt VE & LE77.42308.49192.9631.3125.7628.54
" + }, + { + "type": "table_footnote", + "bbox": [ + 0.525, + 0.156, + 0.896, + 0.207 + ], + "angle": 0, + "content": "(a) Impact of different settings: \\( w / o \\) CoT Pre-training and \\( w / o \\) CoT Annealing indicate pre-training without CoT and supervised fine-tuning with only CoT data, while \\( w / o \\) pt VE and PE uses randomly initialized visual and layout embeddings instead of pre-trained ones." + }, + { + "type": "table", + "bbox": [ + 0.528, + 0.22, + 0.895, + 0.291 + ], + "angle": 0, + "content": "
ModalityDocument-oriented VQAKIE
TLVDocVQAVisualMRCAvg.DeepFormKLCAvg.
✓✗✗73.57280.27176.9315.7624.0919.93
✓(I)✗75.72306.51191.1214.1324.2319.18
✓(II)✗77.32310.75194.0333.7525.9729.86
✓(II)✓77.43316.27196.8535.4126.2530.83
" + }, + { + "type": "table_footnote", + "bbox": [ + 0.525, + 0.293, + 0.895, + 0.342 + ], + "angle": 0, + "content": "(b) Impacts of different modality configurations. \\(\\mathbf{T}\\), \\(\\mathbf{L}\\), and \\(\\mathbf{V}\\) represent textual, layout, and visual modalities, respectively. (I) indicates inputting OCR coordinates as text following He et al. [14], while (II) embeds coordinates with \\(LE\\) in our approach." + }, + { + "type": "table_caption", + "bbox": [ + 0.52, + 0.354, + 0.9, + 0.369 + ], + "angle": 0, + "content": "Table 5. Ablation study on various designs of our DocLayLLM." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.394, + 0.905, + 0.529 + ], + "angle": 0, + "content": "of our DocLayLLM to incorporate different modalities for TDU tasks. It is obvious that using only text from the document is insufficient for TDU, while adding layout information in textual form provides some gain on VQA tasks but leads to over-length input (see appendix for analysis), greatly slowing the inference. By our design to integrate layout features via an \\(LE\\), we achieve notable improvements on these TDU tasks with shorter inputs. And adding visual features further boosts the performance of our method." + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.53, + 0.905, + 0.561 + ], + "angle": 0, + "content": "Overall, the combined impact of these designs ensures the efficiency and effectiveness of our method." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.574, + 0.634, + 0.589 + ], + "angle": 0, + "content": "5. Conclusion" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.599, + 0.907, + 0.903 + ], + "angle": 0, + "content": "We propose DocLayLLM, an efficient multi-modal extension of large language models for text-rich document understanding. By streamlined integration of document patch tokens and OCR positional tokens into the LLM inputs, it achieves a simple yet efficient multi-modal extension of LLM to enable it to have the capability to perceive and process textual, visual, and layout information from the OCR results of the document. Additionally, we introduce CoT Pre-training and CoT Annealing techniques centered around CoT, offering a new way to generate CoT data and further enhancing the model's understanding of text-rich documents. Despite requiring fewer training resources, our method surpasses existing OCR-dependent TDU models, demonstrating strong zero-shot generalization capability. Furthermore, it also exceeds the performance of SOTA OCR-free TDU methods under comparable training conditions, showcasing remarkable performance across various document understanding tasks. Our work significantly advances efficient MLLM-based document understanding while maintaining high-performance standards." + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.516, + 0.957 + ], + "angle": 0, + "content": "4045" + } + ], + [ + { + "type": "title", + "bbox": [ + 0.092, + 0.091, + 0.251, + 0.108 + ], + "angle": 0, + "content": "Acknowledgement" + }, + { + "type": "text", + "bbox": [ + 0.091, + 0.114, + 0.486, + 0.157 + ], + "angle": 0, + "content": "This research is supported in part by National Natural Science Foundation of China (Grant No.: 62476093, 62441604)." + }, + { + "type": "title", + "bbox": [ + 0.093, + 0.169, + 0.188, + 0.185 + ], + "angle": 0, + "content": "References" + }, + { + "type": "ref_text", + "bbox": [ + 0.101, + 0.194, + 0.484, + 0.248 + ], + "angle": 0, + "content": "[1] Srikar Appalaraju, Bhavan Jasani, Bhargava Urala Kota, Yusheng Xie, and R. Manmatha. DocFormer: End-to-End Transformer for Document Understanding. In Proc. ICCV, pages 993–1003, 2021. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.101, + 0.25, + 0.484, + 0.304 + ], + "angle": 0, + "content": "[2] Srikar Appalaraju, Peng Tang, Qi Dong, Nishant Sankaran, Yichu Zhou, and R. Manmatha. DocFormerv2: Local Features for Document Understanding. In Proc. AAAI, pages 709-718, 2024. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.305, + 0.484, + 0.442 + ], + "angle": 0, + "content": "[3] Tom Brown, Benjamin Mann, Nick Ryder, Melanie Subbiah, Jared D Kaplan, Prafulla Dhariwal, Arvind Neelakantan, Pranav Shyam, Girish Sastry, Amanda Askell, Sandhini Agarwal, Ariel Herbert-Voss, Gretchen Krueger, Tom Henighan, Rewon Child, Aditya Ramesh, Daniel Ziegler, Jeffrey Wu, Clemens Winter, Chris Hesse, Mark Chen, Eric Sigler, Mateusz Litwin, Scott Gray, Benjamin Chess, Jack Clark, Christopher Berner, Sam McCandlish, Alec Radford, Ilya Sutskever, and Dario Amodei. Language Models are Few-shot Learners. Proc. NIPS, 33:1877-1901, 2020. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.444, + 0.483, + 0.498 + ], + "angle": 0, + "content": "[4] Wenhu Chen, Hongmin Wang, Jianshu Chen, Yunkai Zhang, Hong Wang, Shiyang Li, Xiyou Zhou, and William Yang Wang. TabFact: A Large-scale Dataset for Table-based Fact Verification. In Proc. ICLR, 2020. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.5, + 0.483, + 0.554 + ], + "angle": 0, + "content": "[5] Gang Dai, Yifan Zhang, Qingfeng Wang, Qing Du, Zhuliang Yu, Zhuoman Liu, and Shuangping Huang. Disentangling Writer and Character Styles for Handwriting Generation. In Proc. CVPR, pages 5977-5986, 2023. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.555, + 0.483, + 0.609 + ], + "angle": 0, + "content": "[6] Gang Dai, Yifan Zhang, Quhui Ke, Qiangya Guo, and Shuangping Huang. One-dm: One-shot diffusion mimicker for handwritten text generation. In Proc. ECCV, pages 410-427, 2025. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.611, + 0.483, + 0.651 + ], + "angle": 0, + "content": "[7] DeepSeek-AI. DeepSeek-R1: Incentivizing Reasoning Capability in LLMs via Reinforcement Learning. arXiv: Comp. Res. Repository, abs/2501.12948, 2025. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.652, + 0.483, + 0.707 + ], + "angle": 0, + "content": "[8] Jacob Devlin, Ming-Wei Chang, Kenton Lee, and Kristina Toutanova. BERT: Pre-training of Deep Bidirectional Transformers for Language Understanding. In Proc. NAACL, pages 4171–4186, 2019. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.708, + 0.483, + 0.775 + ], + "angle": 0, + "content": "[9] Hao Feng, Qi Liu, Hao Liu, Wengang Zhou, Houqiang Li, and Can Huang. DocPedia: Unleashing the Power of Large Multimodal Model in the Frequency Domain for Versatile Document Understanding. arXiv: Comp. Res. Repository, abs/2311.11810, 2023. 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.777, + 0.483, + 0.845 + ], + "angle": 0, + "content": "[10] Hao Feng, Zijian Wang, Jingqun Tang, Jinghui Lu, Wengang Zhou, Houqiang Li, and Can Huang. UniDoc: A Universal Large Multimodal Model for Simultaneous Text Detection, Recognition, Spotting and Understanding. arXiv: Comp. Res. Repository, abs/2308.11592, 2023. 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.846, + 0.483, + 0.899 + ], + "angle": 0, + "content": "[11] Masato Fujitake. LayoutLLM: Large Language Model Instruction Tuning for Visually Rich Document Understanding. In Proc. LREC/COLING, pages 10219-10224, 2024. 2, 3" + }, + { + "type": "list", + "bbox": [ + 0.094, + 0.194, + 0.484, + 0.899 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.093, + 0.906, + 0.161 + ], + "angle": 0, + "content": "[12] Lukasz Garncarek, Rafat Powalski, Tomasz Stanisławek, Bartosz Topolski, Piotr Halama, Michal Turski, and Filip Graliński. LAMBERT: Layout-Aware Language Modeling for Information Extraction. In Proc. ICDAR, pages 532-547, 2021. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.164, + 0.906, + 0.219 + ], + "angle": 0, + "content": "[13] Adam W. Harley, Alex Ufkes, and Konstantinos G. Derpanis. Evaluation of Deep Convolutional Nets for Document Image Classification and Retrieval. In Proc. ICDAR, pages 991-995, 2015. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.221, + 0.906, + 0.289 + ], + "angle": 0, + "content": "[14] Jiabang He, Lei Wang, Yi Hu, Ning Liu, Hui Liu, Xing Xu, and Heng Tao Shen. ICL-D3IE: In-Context Learning with Diverse Demonstrations Updating for Document Information Extraction. In Proc. ICCV, pages 19428-19437, 2023. 1, 2, 3, 6, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.291, + 0.906, + 0.36 + ], + "angle": 0, + "content": "[15] Teakgyu Hong, DongHyun Kim, Mingi Ji, Wonseok Hwang, Daehyun Nam, and Sungrae Park. BROS: A Pre-trained Language Model Focusing on Text and Layout for Better Key Information Extraction from Documents. In Proc. AAAI, pages 10767-10775, 2022. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.363, + 0.906, + 0.431 + ], + "angle": 0, + "content": "[16] Wenyi Hong, Weihan Wang, Qingsong Lv, Jiazheng Xu, Wenmeng Yu, Junhui Ji, Yan Wang, Zihan Wang, Yuxiao Dong, Ming Ding, and Jie Tang. CogAgent: A Visual Language Model for GUI Agents. In Proc. CVPR, pages 14281-14290, 2024. 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.434, + 0.906, + 0.502 + ], + "angle": 0, + "content": "[17] Anwen Hu, Haiyang Xu, Jiabo Ye, Ming Yan, Liang Zhang, Bo Zhang, Ji Zhang, Qin Jin, Fei Huang, and Jingren Zhou. mPLUG-DocOwl 1.5: Unified Structure Learning for OCR-free Document Understanding. In Findings EMNLP, pages 3096-3120, 2024. 1, 2, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.504, + 0.906, + 0.56 + ], + "angle": 0, + "content": "[18] Edward J. Hu, Yelong Shen, Phillip Wallis, Zeyuan Allen-Zhu, Yuanzhi Li, Shean Wang, Lu Wang, and Weizhu Chen. LoRA: Low-Rank Adaptation of Large Language Models. In Proc. ICLR, 2022. 2, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.562, + 0.906, + 0.685 + ], + "angle": 0, + "content": "[19] Shengding Hu, Yuge Tu, Xu Han, Chaoqun He, Ganqu Cui, Xiang Long, Zhi Zheng, Yewei Fang, Yuxiang Huang, Weilin Zhao, Xinrong Zhang, Zhen Leng Thai, Kai Zhang, Chongyi Wang, Yuan Yao, Chenyang Zhao, Jie Zhou, Jie Cai, Zhongwu Zhai, Ning Ding, Chao Jia, Guoyang Zeng, Dahai Li, Zhiyuan Liu, and Maosong Sun. MiniCPM: Unveiling the Potential of Small Language Models with Scalable Training Strategies. arXiv: Comp. Res. Repository, abs/2404.06395, 2024. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.688, + 0.906, + 0.743 + ], + "angle": 0, + "content": "[20] Yupan Huang, Tengchao Lv, Lei Cui, Yutong Lu, and Furu Wei. LayoutLMv3: Pre-training for Document AI with Unified Text and Image Masking. In Proc. ACM MM, pages 4083-4091, 2022. 1, 2, 3, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.746, + 0.906, + 0.801 + ], + "angle": 0, + "content": "[21] Zheng Huang, Kai Chen, Jianhua He, Xiang Bai, Dimosthenis Karatzas, Shijian Lu, and C. V. Jawahar. ICDAR2019 Competition on Scanned Receipt OCR and Information Extraction. In Proc. ICDAR, pages 1516-1520, 2019. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.802, + 0.906, + 0.857 + ], + "angle": 0, + "content": "[22] Guillaume Jaume, Hazim Kemal Ekenel, and Jean-Philippe Thiran. FUNSD: A Dataset for Form Understanding in Noisy Scanned Documents. In Workshop on ICDAR, pages 1–6, 2019. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.859, + 0.906, + 0.902 + ], + "angle": 0, + "content": "[23] Geewook Kim, Teakgyu Hong, Moonbin Yim, JeongYeon Nam, Jinyoung Park, Jinyeong Yim, Wonseok Hwang, Sangdoo Yun, Dongyoon Han, and Seunghyun Park. OCR-Free" + }, + { + "type": "list", + "bbox": [ + 0.517, + 0.093, + 0.906, + 0.902 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.517, + 0.956 + ], + "angle": 0, + "content": "4046" + } + ], + [ + { + "type": "ref_text", + "bbox": [ + 0.125, + 0.092, + 0.482, + 0.12 + ], + "angle": 0, + "content": "Document Understanding Transformer. Proc. ECCV, 13688: 498-517, 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.122, + 0.483, + 0.205 + ], + "angle": 0, + "content": "[24] Geewook Kim, Hodong Lee, Daehee Kim, Haeji Jung, Sanghee Park, Yoonsik Kim, Sangdoo Yun, Taeho Kil, Bado Lee, and Seunghyun Park. Visually-Situated Natural Language Understanding with Contrastive Reading Model and Frozen Large Language Models. In Proc. EMNLP, pages 11989-12010, 2023. 2, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.208, + 0.483, + 0.263 + ], + "angle": 0, + "content": "[25] Jianfeng Kuang, Wei Hua, Dingkang Liang, Mingkun Yang, Deqiang Jiang, Bo Ren, and Xiang Bai. Visual Information Extraction in the Wild: Practical Dataset and End-to-End Solution. In Proc. ICDAR, pages 36-53, 2023. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.266, + 0.482, + 0.32 + ], + "angle": 0, + "content": "[26] Hugo Laurençon, Andres Marafioti, Victor Sanh, and Leo Tronchon. Building and better understanding vision-language models: insights and future directions. In Workshop on NIPS, 2024. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.323, + 0.482, + 0.39 + ], + "angle": 0, + "content": "[27] Bo Li, Yuanhan Zhang, Dong Guo, Renrui Zhang, Feng Li, Hao Zhang, Kaichen Zhang, Peiyuan Zhang, Yanwei Li, Ziwei Liu, and Chunyuan Li. LLaVA-OneVision: Easy Visual Task Transfer. Transactions on Machine Learning Research, 2025. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.394, + 0.482, + 0.449 + ], + "angle": 0, + "content": "[28] Chenliang Li, Bin Bi, Ming Yan, Wei Wang, Songfang Huang, Fei Huang, and Luo Si. StructuralLM: Structural Pre-training for Form Understanding. In Proc. ACL, pages 6309-6318, 2021. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.452, + 0.482, + 0.52 + ], + "angle": 0, + "content": "[29] Feng Li, Renrui Zhang, Hao Zhang, Yuanhan Zhang, Bo Li, Wei Li, Zejun Ma, and Chunyuan Li. LLaVa-NeXT-Interleave: Tackling Multi-image, Video, and 3D in Large Multimodal Models. arXiv: Comp. Res. Repository, abs/2407.07895, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.523, + 0.482, + 0.577 + ], + "angle": 0, + "content": "[30] Junnan Li, Dongxu Li, Silvio Savarese, and Steven C. H. Hoi. BLIP-2: Bootstrapping Language-Image Pre-training with Frozen Image Encoders and Large Language Models. Proc. ICML, 202:19730-19742, 2023. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.581, + 0.483, + 0.843 + ], + "angle": 0, + "content": "[31] Jeffrey Li, Alex Fang, Georgios Smyrnis, Maor Ivgi, Matt Jordan, Samir Yitzhak Gadre, Hritik Bansal, Etash Guha, Sedrick Keh, Kushal Arora, Saurabh Garg, Rui Xin, Niklas Muennighoff, Reinhard Heckel, Jean Mercat, Mayee Chen, Suchin Gururangan, Mitchell Wortsman, Alon Albalak, Yonatan Bitton, Marianna Nezhurina, Amro Abbas, ChengYu Hsieh, Dhruba Ghosh, Josh Gardner, Maciej Kilian, Hanlin Zhang, Rulin Shao, Sarah M. Pratt, Sunny Sanyal, Gabriel Ilharco, Giannis Daras, Kalyani Marathe, Aaron Gokaslan, Jieyu Zhang, Khyathi Raghavi Chandu, Thao Nguyen, Igor Vasiljevic, Sham M. Kakade, Shuran Song, Sujay Sanghavi, Fartash Faghri, Sewoong Oh, Luke Zettlemoyer, Kyle Lo, Alaaeldin El-Nouby, Hadi Pouransari, Alexander Toshev, Stephanie Wang, Dirk Groeneveld, Luca Soldaini, Pang Wei Koh, Jenia Jitsev, Thomas Kollar, Alexandros G. Dimakis, Yair Carmon, Achal Dave, Ludwig Schmidt, and Vaishaal Shankar. DataComp-LM: In search of the next generation of training sets for language models. arXiv: Comp. Res. Repository, abs/2406.11794, 2024. 2, 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.846, + 0.482, + 0.901 + ], + "angle": 0, + "content": "[32] Minghao Li, Yiheng Xu, Lei Cui, Shaohan Huang, Furu Wei, Zhoujun Li, and Ming Zhou. DocBank: A Benchmark Dataset for Document Layout Analysis. In Pro. COLING, pages 949-960, 2020. 5" + }, + { + "type": "list", + "bbox": [ + 0.093, + 0.092, + 0.483, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.093, + 0.905, + 0.147 + ], + "angle": 0, + "content": "[33] Peizhao Li, Jiuxiang Gu, Jason Kuen, Vlad I Morariu, Handong Zhao, Rajiv Jain, Varun Manjunatha, and Hongfu Liu. SelfDoc: Self-Supervised Document Representation Learning. In Proc. CVPR, pages 5652-5660, 2021. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.149, + 0.905, + 0.203 + ], + "angle": 0, + "content": "[34] Qiwei Li, Zuchao Li, Xiantao Cai, Bo Du, and Hai Zhao. Enhancing Visually-Pich Pdocument Punderstanding via Payout Pstructure Podeling. In Proc. ACM MM, page 4513-4523, 2023." + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.205, + 0.905, + 0.245 + ], + "angle": 0, + "content": "[35] Qiwei Li, Zuchao Li, Ping Wang, Haojun Ai, and Hai Zhao. Hypergraph based Understanding for Document Semantic Entity Recognition. In Proc. ACL, pages 2950-2960, 2024." + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.247, + 0.905, + 0.301 + ], + "angle": 0, + "content": "[36] Xin Li, Yan Zheng, Yiqing Hu, Haoyu Cao, Yunfei Wu, Deqiang Jiang, Yinsong Liu, and Bo Ren. Relational Representation Learning in Visually-Rich Documents. In Proc. ACM MM, page 4614-4624, 2022." + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.303, + 0.905, + 0.37 + ], + "angle": 0, + "content": "[37] Yulin Li, Yuxi Qian, Yuechen Yu, Xiameng Qin, Chengquan Zhang, Yan Liu, Kun Yao, Junyu Han, Jingtuo Liu, and Errui Ding. StrucTexT: Structured Text Understanding with Multi-Modal Transformers. In Proc. ACM MM, page 1912–1920, 2021. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.372, + 0.905, + 0.44 + ], + "angle": 0, + "content": "[38] Zhang Li, Biao Yang, Qiang Liu, Zhiyin Ma, Shuo Zhang, Jingxu Yang, Yabo Sun, Yuliang Liu, and Xiang Bai. Monkey: Image Resolution and Text Label Are Important Things for Large Multi-modal Models. In Proc. CVPR, pages 26763-26773, 2024. 1, 2, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.442, + 0.905, + 0.495 + ], + "angle": 0, + "content": "[39] Zening Lin, Jiapeng Wang, and Lianwen Jin. Visual Information Extraction Deep Learning Method: A Critical Review. Journal of Image and Graphics, 28(08):2276-2297, 2023. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.497, + 0.905, + 0.551 + ], + "angle": 0, + "content": "[40] Chenglin Liu, Lianwen Jin, Xiang Bai, Xiaohui Li, and Fei Yin. Frontiers of intelligent document analysis and recognition: review and prospects. Journal of Image and Graphics, 28(08):2223-2252, 2023. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.553, + 0.905, + 0.608 + ], + "angle": 0, + "content": "[41] Yuliang Liu, Biao Yang, Qiang Liu, Zhang Li, Zhiyin Ma, Shuo Zhang, and Xiang Bai. TextMonkey: An OCR-Free Large Multimodal Model for Understanding Document. arXiv: Comp. Res. Repository, abs/2403.04473, 2024. 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.609, + 0.905, + 0.637 + ], + "angle": 0, + "content": "[42] Meta AI Llama Team. The Llama 3 Herd of Models. arXiv: Comp. Res. Repository, abs/2407.21783, 2024. 2, 5, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.638, + 0.905, + 0.719 + ], + "angle": 0, + "content": "[43] Jinghui Lu, Haiyang Yu, Yanjie Wang, Yongjie Ye, Jingqun Tang, Ziwei Yang, Binghong Wu, Qi Liu, Hao Feng, Han Wang, Hao Liu, and Can Huang. A Bounding Box is Worth One Token: Interleaving Layout and Text in a Large Language Model for Document Understanding. arXiv: Comp. Res. Repository, abs/2407.01976, 2024. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.721, + 0.905, + 0.761 + ], + "angle": 0, + "content": "[44] Chuwei Luo, Changxu Cheng, Qi Zheng, and Cong Yao. GeoLayoutLM: Geometric Pre-Training for Visual Information Extraction. In Proc. CVPR, pages 7092–7101, 2023. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.763, + 0.905, + 0.817 + ], + "angle": 0, + "content": "[45] Chuwei Luo, Yufan Shen, Zhaoqing Zhu, Qi Zheng, Zhi Yu, and Cong Yao. LayoutLLM: Layout Instruction Tuning with Large Language Models for Document Understanding. In Proc. CVPR, pages 15630-15640, 2024. 2, 3, 4, 5, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.818, + 0.905, + 0.859 + ], + "angle": 0, + "content": "[46] Minesh Mathew, Dimosthenis Karatzas, and C. V. Jawahar. DocVQA: A Dataset for VQA on Document Images. In Proc. WACV, pages 2199-2208, 2021. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.86, + 0.905, + 0.901 + ], + "angle": 0, + "content": "[47] Minesh Mathew, Viraj Bagal, Ruben Tito, Dimosthenis Karatzas, Ernest Valveny, and C.V. Jawahar. InfographicVQA. In Proc. WACV, pages 1697-1706, 2022. 5" + }, + { + "type": "list", + "bbox": [ + 0.517, + 0.093, + 0.905, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.516, + 0.956 + ], + "angle": 0, + "content": "4047" + } + ], + [ + { + "type": "ref_text", + "bbox": [ + 0.092, + 0.093, + 0.482, + 0.135 + ], + "angle": 0, + "content": "[48] G. Nagy, S. Seth, and M. Viswanathan. A Prototype Document Image Analysis System for Technical Journals. Computer, 25(7):10-22, 1992. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.136, + 0.322, + 0.149 + ], + "angle": 0, + "content": "[49] OpenAI. Hello GPT-4o, 2024. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.152, + 0.483, + 0.207 + ], + "angle": 0, + "content": "[50] Seunghyun Park, Seung Shin, Bado Lee, Junyeop Lee, Jaeheung Surh, Minjoon Seo, and Hwalsuk Lee. CORD: A Consolidated Receipt Dataset for Post-OCR Parsing. In Workshop on NIPS, 2019. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.208, + 0.483, + 0.248 + ], + "angle": 0, + "content": "[51] Panupong Pasupat and Percy Liang. Compositional Semantic Parsing on Semi-Structured Tables. In Proc. ACL, pages 1470–1480, 2015. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.251, + 0.483, + 0.347 + ], + "angle": 0, + "content": "[52] Qiming Peng, Yinxu Pan, Wenjin Wang, Bin Luo, Zhenyu Zhang, Zhengjie Huang, Yuhui Cao, Weichong Yin, Yongfeng Chen, Yin Zhang, Shikun Feng, Yu Sun, Hao Tian, Hua Wu, and Haifeng Wang. ERNIE-Layout: Layout Knowledge Enhanced Pre-training for Visually-rich Document Understanding. In Findings of EMNLP, pages 3744–3756, 2022. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.35, + 0.483, + 0.432 + ], + "angle": 0, + "content": "[53] Vincent Perot, Kai Kang, Florian Luisier, Guolong Su, Xiaoyu Sun, Ramya Sree Boppana, Zilong Wang, Zifeng Wang, Jiaqi Mu, Hao Zhang, Chen-Yu Lee, and Nan Hua. LMDX: Language Model-based Document Information Extraction and Localization. In Findings of ACL, pages 15140-15168, 2024. 3, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.434, + 0.483, + 0.503 + ], + "angle": 0, + "content": "[54] Birgit Pfitzmann, Christoph Auer, Michele Dolfi, Ahmed S. Nassar, and Peter W. J. Staar. DocLayNet: A Large Human-Annotated Dataset for Document-Layout Segmentation. In Proc. ACM SIGKDD Int. Conf. Knowledge discovery & data mining, pages 3743–3751, 2022. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.505, + 0.483, + 0.573 + ], + "angle": 0, + "content": "[55] Rafat Powalski, Łukasz Borchmann, Dawid Jurkiewicz, Tomasz Dwojak, Michal Pietruszka, and Gabriela Pałka. Going Full-TILT Boogie on Document Understanding with Text-Image-Layout Transformer. In Proc. ICDAR, pages 732–747, 2021. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.576, + 0.483, + 0.657 + ], + "angle": 0, + "content": "[56] Stepán Simsa, Milan Sulc, Michal Uricár, Yash Patel, Ahmed Hamdi, Matej Kocián, Matyas Skalický, Jiri Matas, Antoine Doucet, Mickaël Coustaty, and Dimosthenis Karatzas. DocILE Benchmark for Document Information Localization and Extraction. In Proc. ICDAR, pages 147-166, 2023. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.66, + 0.483, + 0.73 + ], + "angle": 0, + "content": "[57] Tomasz Stanislawek, Filip Gralinski, Anna Wróblewska, Dawid Lipinski, Agnieszka Kaliska, Paulina Rosalska, Bartosz Topolski, and Przemyslaw Biecek. Kleister: Key Information Extraction Datasets Involving Long Documents with Complex Layouts. In Proc. ICDAR, pages 564-579, 2021. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.731, + 0.483, + 0.772 + ], + "angle": 0, + "content": "[58] Stacey Svetlichnaya. DeepForm: Understand Structured Documents at Scale. https://wandb.ai/stacey/deepform_v1, 2020. Accessed: 2024-08-14. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.774, + 0.483, + 0.815 + ], + "angle": 0, + "content": "[59] Ryota Tanaka, Kyosuke Nishida, and Sen Yoshida. VisualMRC: Machine Reading Comprehension on Document Images. In Proc. AAAI, pages 13878-13888, 2021. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.817, + 0.483, + 0.872 + ], + "angle": 0, + "content": "[60] Ryota Tanaka, Taichi Iki, Kyosuke Nishida, Kuniko Saito, and Jun Suzuki. InstructDoc: A Dataset for Zero-shot Generalization of Visual Document Understanding with Instructions. In Proc. AAAI, pages 19071-19079, 2024. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.874, + 0.483, + 0.901 + ], + "angle": 0, + "content": "[61] Jingqun Tang, Chunhui Lin, Zhen Zhao, Shu Wei, Binghong Wu, Qi Liu, Hao Feng, Yang Li, Siqi Wang, Lei Liao, Wei" + }, + { + "type": "list", + "bbox": [ + 0.092, + 0.093, + 0.483, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.545, + 0.093, + 0.905, + 0.147 + ], + "angle": 0, + "content": "Shi, Yuliang Liu, Hao Liu, Yuan Xie, Xiang Bai, and Can Huang. TextSquare: Scaling up Text-Centric Visual Instruction Tuning. arXiv: Comp. Res. Repository, abs/2404.12803, 2024. 1, 2, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.149, + 0.905, + 0.217 + ], + "angle": 0, + "content": "[62] Zineng Tang, Ziyi Yang, Guoxin Wang, Yuwei Fang, Yang Liu, Chenguang Zhu, Michael Zeng, Cha Zhang, and Mohit Bansal. Unifying Vision, Text, and Layout for Universal Document Processing. In Proc. CVPR, pages 19254-19264, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.219, + 0.905, + 0.245 + ], + "angle": 0, + "content": "[63] Qwen Team. QwQ: Reflect Deeply on the Boundaries of the Unknown, 2024. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.248, + 0.905, + 0.55 + ], + "angle": 0, + "content": "[64] Hugo Touvron, Louis Martin, Kevin Stone, Peter Albert, Amjad Almahairi, Yasmine Babaei, Nikolay Bashlykov, Soumya Batra, Prajjwal Bhargava, Shruti Bhosale, Dan Bikel, Lukas Blecher, Cristian Canton-Ferrer, Moya Chen, Guillem Cucurull, David Esiobu, Jude Fernandes, Jeremy Fu, Wenyin Fu, Brian Fuller, Cynthia Gao, Vedanuj Goswami, Naman Goyal, Anthony Hartshorn, Saghar Hosseini, Rui Hou, Hakan Inan, Marcin Kardas, Viktor Kerkez, Madian Khabsa, Isabel Kloumann, Artem Korenev, Punit Singh Koura, Marie-Anne Lachaux, Thibaut Lavril, Jenya Lee, Diana Liskovich, Yinghai Lu, Yuning Mao, Xavier Martinet, Todor Mihaylov, Pushkar Mishra, Igor Molybog, Yixin Nie, Andrew Poulton, Jeremy Reizenstein, Rashi Rungta, Kalyan Saladi, Alan Schelten, Ruan Silva, Eric Michael Smith, Ranjan Subramanian, Xiaoqing Ellen Tan, Binh Tang, Ross Taylor, Adina Williams, Jian Xiang Kuan, Puxin Xu, Zheng Yan, Iliyan Zarov, Yuchen Zhang, Angela Fan, Melanie Kambadur, Sharan Narang, Aurélien Rodriguez, Robert Stojnic, Sergey Edunov, and Thomas Scialom. Llama 2: Open Foundation and Fine-Tuned Chat Models. arXiv: Comp. Res. Repository, abs/2307.09288, 2023. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.552, + 0.905, + 0.607 + ], + "angle": 0, + "content": "[65] Yi Tu, Ya Guo, Huan Chen, and Jinyang Tang. LayoutMask: Enhance Text-Layout Interaction in Multi-modal Pretraining for Document Understanding. In Proc. ACL, pages 15200–15212, 2023. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.609, + 0.905, + 0.677 + ], + "angle": 0, + "content": "[66] Jianqiang Wan, Sibo Song, Wenwen Yu, Yuliang Liu, Wenqing Cheng, Fei Huang, Xiang Bai, Cong Yao, and Zhibo Yang. OmniParser: A Unified Framework for Text Spotting Key Information Extraction and Table Recognition. In Proc. CVPR, pages 15641-15653, 2024. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.679, + 0.905, + 0.747 + ], + "angle": 0, + "content": "[67] Dongsheng Wang, Natraj Raman, Mathieu Sibue, Zhiqiang Ma, Petr Babkin, Simerjot Kaur, Yulong Pei, Armineh Nourbakhsh, and Xiaomo Liu. DocLLM: A Layout-Aware Generative Language Model for Multimodal Document Understanding. In Proc. ACL, pages 8529-8548, 2024. 3, 6, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.749, + 0.905, + 0.802 + ], + "angle": 0, + "content": "[68] Jiapeng Wang, Lianwen Jin, and Kai Ding. LiLT: A Simple yet Effective Language-Independent Layout Transformer for Structured Document Understanding. In Proc. ACL, pages 7747–7757, 2022. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.804, + 0.905, + 0.872 + ], + "angle": 0, + "content": "[69] Jiapeng Wang, Zening Lin, Dayi Huang, Longfei Xiong, and Lianwen Jin. LiLTv2: Language-substitutable layout-image transformer for visual information extraction. ACM Transactions on Multimedia Computing, Communications and Applications, 21(3):1-27, 2025. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.874, + 0.905, + 0.901 + ], + "angle": 0, + "content": "[70] Wenjin Wang, Yunhao Li, Yixin Ou, and Yin Zhang. Layout and Task Aware Instruction Prompt for Zero-shot Document" + }, + { + "type": "list", + "bbox": [ + 0.517, + 0.093, + 0.905, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.516, + 0.956 + ], + "angle": 0, + "content": "4048" + } + ], + [ + { + "type": "ref_text", + "bbox": [ + 0.125, + 0.092, + 0.482, + 0.12 + ], + "angle": 0, + "content": "Image Question Answering. arXiv: Comp. Res. Repository, abs/2306.00526, 2023. 1, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.122, + 0.483, + 0.19 + ], + "angle": 0, + "content": "[71] Jason Wei, Xuezhi Wang, Dale Schuurmans, Maarten Bosma, brian richter, Fei Xia, Ed Chi, Quoc V Le, and Denny Zhou. Chain-of-Thought Prompting Elicits Reasoning in Large Language Models. Proc. NIPS, 35:24824-24837, 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.192, + 0.482, + 0.262 + ], + "angle": 0, + "content": "[72] Yiheng Xu, Minghao Li, Lei Cui, Shaohan Huang, Furu Wei, and Ming Zhou. LayoutLM: Pre-training of Text and Layout for Document Image Understanding. In Proc. ACM SIGKDD Int. Conf. Knowledge discovery & data mining, page 1192-1200, 2020. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.263, + 0.482, + 0.332 + ], + "angle": 0, + "content": "[73] Yang Xu, Yiheng Xu, Tengchao Lv, Lei Cui, Furu Wei, Guoxin Wang, Yijuan Lu, Dinei Florencio, Cha Zhang, Wanxiang Che, Min Zhang, and Lidong Zhou. LayoutLMv2: Multi-modal Pre-training for Visually-rich Document Understanding. In Proc. ACL, pages 2579-2591, 2021. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.334, + 0.482, + 0.387 + ], + "angle": 0, + "content": "[74] Hao Yan, Yuliang Liu, Lianwen Jin, and Xiang Bai. The development, application, and future of LLM similar to ChatGPT. Journal of Image and Graphics, 28(9):2749-2762, 2023. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.39, + 0.482, + 0.471 + ], + "angle": 0, + "content": "[75] Jiabo Ye, Anwen Hu, Haiyang Xu, Qinghao Ye, Ming Yan, Yuhao Dan, Chenlin Zhao, Guohai Xu, Chenliang Li, Junfeng Tian, Qian Qi, Ji Zhang, and Fei Huang. mPLUGDocOwl: Modularized Multimodal Large Language Model for Document Understanding. arXiv: Comp. Res. Repository, abs/2307.02499, 2023. 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.474, + 0.483, + 0.557 + ], + "angle": 0, + "content": "[76] Jiabo Ye, Anwen Hu, Haiyang Xu, Qinghao Ye, Ming Yan, Guohai Xu, Chenliang Li, Junfeng Tian, Qi Qian, Ji Zhang, Qin Jin, Liang He, Xin Lin, and Fei Huang. UReader: Universal OCR-free Visually-situated Language Understanding with Multimodal Large Language Model. In Findings of EMNLP, pages 2841-2858, 2023. 1, 2, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.559, + 0.482, + 0.626 + ], + "angle": 0, + "content": "[77] Chong Zhang, Ya Guo, Yi Tu, Huan Chen, Jinyang Tang, Huijia Zhu, Qi Zhang, and Tao Gui. Reading Order Matters: Information Extraction from Visually-rich Documents by Token Path Prediction. In Proc. EMNLP, pages 13716-13730, 2023. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.629, + 0.482, + 0.696 + ], + "angle": 0, + "content": "[78] Jiaxin Zhang, Wentao Yang, Songxuan Lai, Zecheng Xie, and Lianwen Jin. DocKylin: A Large Multimodal Model for Visual Document Understanding with Efficient Visual Slimming. arXiv: Comp. Res. Repository, abs/2406.19101, 2024. 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.699, + 0.482, + 0.768 + ], + "angle": 0, + "content": "[79] Yanzhe Zhang, Ruiyi Zhang, Jiuming Gu, Yufan Zhou, Nedim Lipka, Diyi Yang, and Tong Sun. LLaVAR: Enhanced Visual Instruction Tuning for Text-Rich Image Understanding. arXiv: Comp. Res. Repository, abs/2306.17107, 2023. 1, 2, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.77, + 0.482, + 0.811 + ], + "angle": 0, + "content": "[80] Xu Zhong, Jianbin Tang, and Antonio Jimeno-Yepes. Pub-LayNet: Largest Dataset Ever for Document Layout Analysis. In Proc. ICDAR, pages 1015-1022, 2019. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.813, + 0.482, + 0.853 + ], + "angle": 0, + "content": "[81] Xu Zhong, Elaheh ShafieiBavani, and Antonio JimenoYepes. Image-Based Table Recognition: Data, Model, and Evaluation. Proc. ECCV, 12366:564-580, 2020. 5" + }, + { + "type": "list", + "bbox": [ + 0.093, + 0.092, + 0.483, + 0.853 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.516, + 0.956 + ], + "angle": 0, + "content": "4049" + } + ] +] \ No newline at end of file diff --git a/2025/DocLayLLM_ An Efficient Multi-modal Extension of Large Language Models for Text-rich Document Understanding/587c40bc-a1fe-44db-8881-11814790688b_origin.pdf b/2025/DocLayLLM_ An Efficient Multi-modal Extension of Large Language Models for Text-rich Document Understanding/587c40bc-a1fe-44db-8881-11814790688b_origin.pdf new file mode 100644 index 0000000000000000000000000000000000000000..a5d9ee513a82432c53b6170f6dc647cca2aa219d --- /dev/null +++ b/2025/DocLayLLM_ An Efficient Multi-modal Extension of Large Language Models for Text-rich Document Understanding/587c40bc-a1fe-44db-8881-11814790688b_origin.pdf @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:4c3129f90de96dbfbbf96cd0416ae81a68ffa7f82015e3fafedb1afd99672138 +size 1743763 diff --git a/2025/DocLayLLM_ An Efficient Multi-modal Extension of Large Language Models for Text-rich Document Understanding/full.md b/2025/DocLayLLM_ An Efficient Multi-modal Extension of Large Language Models for Text-rich Document Understanding/full.md new file mode 100644 index 0000000000000000000000000000000000000000..0007d3592de9a598769cefd5d6898100ad607b14 --- /dev/null +++ b/2025/DocLayLLM_ An Efficient Multi-modal Extension of Large Language Models for Text-rich Document Understanding/full.md @@ -0,0 +1,334 @@ +# DocLayLLM: An Efficient Multi-modal Extension of Large Language Models for Text-rich Document Understanding + +Wenhui Liao $^{1\dagger}$ , Jiapeng Wang $^{1,2\dagger}$ , Hongliang Li $^{1}$ , Chengyu Wang $^{2*}$ , Jun Huang $^{2}$ , Lianwen Jin $^{1,3*}$ $^{1}$ South China University of Technology, Guangzhou, China + $^{2}$ Alibaba Cloud Computing, Hangzhou, China + $^{3}$ SCUT-Zhuhai Institute of Modern Industrial Innovation, Zhuhai, China + +{eelwh, eejpwang, eehongliangli}@mail.scut.edu.cn, eelwjin@scut.edu.cn {chengyu.wcy, huangjun.hj}@alibaba-inc.com + +# Abstract + +Text-rich document understanding (TDU) requires comprehensive analysis of documents containing substantial textual content and complex layouts. While Multimodal Large Language Models (MLLMs) have achieved fast progress in this domain, existing approaches either demand significant computational resources or struggle with effective multi-modal integration. In this paper, we introduce DocLayLLM, an efficient multi-modal extension of LLMs specifically designed for TDU. By lightly integrating visual patch tokens and 2D positional tokens into LLMs' input and encoding the document content using the LLMs themselves, we fully take advantage of the document comprehension capability of LLMs and enhance their perception of OCR information. We have also deeply considered the role of chain-of-thought (CoT) and innovatively proposed the techniques of CoT Pre-training and CoT Annealing. Our DocLayLLM can achieve remarkable performances with lightweight training settings, showcasing its efficiency and effectiveness. Experimental results demonstrate that our DocLayLLM outperforms existing OCR-dependent methods and OCR-free competitors. Code and model are available at https://github.com/whlscut/DocLayLLM. + +# 1. Introduction + +Text-rich document understanding (TDU) focuses on analyzing documents with extensive text and complex layouts, requiring a combined understanding of text, layout, and images. Early research [1, 2, 12, 15, 20, 28, 33-37, 44, 52, 55, 65, 68, 69, 72, 73, 77] often used BERT-like [8] Transformer-based document encoders (DE), designing various pre-training tasks for large-scale unlabeled + +document corpora and fine-tuning their models on specific downstream tasks like key information extraction (KIE) and visual question answering (VQA) [39, 40]. However, these pre-training tasks often diverge significantly from downstream tasks, requiring the model to acquire substantial task-specific knowledge during fine-tuning. This may result in limited generalization capability across different tasks and datasets. + +With the rapid development of large language models (LLMs), their impressive versatility and generalization capabilities have captured widespread interest [74]. This has led many researchers [17, 38, 61, 76, 79] to overcome the challenges of TDU with LLMs. Initially, efforts focused on employing multi-modal large language models (MLLMs) for OCR-free TDU, where document images are fed into MLLMs, directly generating responses to comprehension tasks. These MLLMs, which can process images and text inputs, are among the most recent advancements in this domain. To date, several general MLLMs have been developed to manage diverse scenarios, and researchers often build their TDU models based on these general models. However, these TDU models typically require high-resolution image inputs to recognize dense text in documents and a large volume of training data to effectively develop this capability, both of which bring significant computational demands. Furthermore, within the extensive collections of image and instruction data used to train these models, the training data for downstream evaluation tasks was often included, which might potentially impact the assessment of their generalization to previously unseen datasets. + +To overcome the constraints of OCR-free methods, scholars have pursued diverse methodologies for integrating OCR information into LLMs. Initial strategies [14, 70] involved inputting coordinates of OCR bounding boxes as numeric text. However, this textual representation of numbers often resulted in excessively long inputs that could + +decrease inference speeds and potentially degrade model performance. In response, many researchers [11, 24, 45] have adopted an auxiliary document encoder like LayoutLMv3 [20] to process textual, visual, and layout features from documents and integrated into LLMs. Although these methods avoid high-resolution image inputs, the textual content of documents must also be encoded alongside the layout and image features within the DE due to its structural limitations. However, the document comprehension capabilities of these DEs may not be on par with those of LLMs themselves, which might have led to significant finetuning efforts for aligning the feature spaces of DEs and LLMs to improve performances during the whole training phase, demanding substantial computational resources. + +To address the limitations of OCR-dependent TDU methods, we introduce DocLayLLM, an efficient multimodal extension of large language models for text-rich document understanding. Our method diverges from existing approaches that utilize an extra DE by integrating the document's image patch tokens and 2D positional tokens with textual content into a natural language expression, and then encoding them through a unified LLM. This innovative integration not only taps into the LLM's inherent text comprehension abilities but also allows it to perceive document layouts. Besides, by incorporating patch tokens, the model supports image-related pre-training tasks and further gains a rudimentary understanding of document structures. + +To advance model performance, we have fully integrated the concept of the chain-of-thought (CoT) [3] into each stage of our training process. Firstly, we developed a novel CoT Pre-training strategy. CoT is engineered to preserve thematic coherence and logical progression during model reasoning, further enhancing the inference ability of LLM, which has proven effectiveness in various subsequent studies [14, 45, 71]. Recently, the use of Long CoT reasoning to enhance the inference capability of models has attracted significant attention [7, 49, 63], and DeepSeek-R1 [7] further emphasized the importance of utilizing CoT data for a cold start before the reinforcement learning stage. Our proposed CoT Pre-training offers a method for rapidly generating large-scale CoT data from non-QA data, thereby improving the model's reasoning ability. By employing CoT in the data generation process and subsequently pre-training the model with the generated QA-formatted data, we have aligned the added visual and layout features with LLM's inherent textual features while improving the model's comprehension ability. Furthermore, we introduced a CoT Annealing technique. While previous research [31, 42] has validated the effectiveness of annealing strategies that gradually increase the proportion of high-quality data towards the end of training, the impact of CoT on data quality had not been explored. To address this, we first propose an annealing strategy that considers data quality from a CoT per + +spective, further improving model performance. + +Thanks to the above design, utilizing only around 3M pre-training data and 300K supervised fine-tuning (SFT) data with the efficient tuning technology of LoRA [18], our DocLayLLM finishes the whole training process in 36 A100 days, therefore resulting an efficient extension of LLM. Even with such low resources, our DocLayLLM surpasses existing OCR-dependent TDU methods. Furthermore, it also excels in specific tasks without fine-tuning on downstream tasks and outperforms state-of-the-art (SOTA) OCR-free methods when trained with training sets of downstream tasks as other methods, further showcasing its effectiveness. Our contributions are: + +- We propose an efficient multi-modal extension of LLMs, which has augmented LLM in understanding text-rich documents and greatly reduced training resources in need. +- We introduce CoT Pre-training, a novel approach for generating CoT data that helps enhance the comprehension capabilities of models. +- We revisit data annealing from a novel perspective and introduce the CoT Annealing strategy, which further enhances the model's efficiency in data utilization and boosts its ultimate performance. +- Our method outperforms existing OCR-dependent TDU solutions and surpasses current OCR-free methods under comparable supervised fine-tuning conditions. + +# 2. Related Works + +OCR-free MLLMs for Text-Rich Documents Understanding. To date, a number of studies have attempted to directly answer the instruction for TDU tasks from document images, thereby obviating the need for extra OCR processing and resulting in an end-to-end paradigm. Such models are referred to as OCR-free methods. Early examples like Donut [23], UDOP [62], and OmniParser [66] have explored end-to-end document understanding solutions without relying on LLMs or MLLMs. However, with the rapid development of MLLMs, many new studies have begun to focus on enhancing the capabilities of general MLLMs in document understanding. Some of them attempt to boost the document understanding ability of general MLLMs by tuning them with synthesizing TDU data generated by proprietary LLMs, including LLaVAR [79] and TextSquare [61], while mPLUG-DocOwl 1.5 [17] enhances the document understanding capability by integrating existing document comprehension data and converting it into vision-related QA-formatted data. Moreover, researchers have also explored image encoding methods to read the text more clearly. For instance, UReader [76] and Monkey [38] empower MLLMs to handle high-resolution images by splitting them into sub-images, encoding each separately, and then concatenating them. Some general MLLMs, such as LLaVA-Next [29], LLaVA + +![](images/67a0a691f89e141767f2dd8180cbecc54cfa227c8d51c9dff7b529145e47dcba.jpg) +Figure 1. The overall architecture of DocLayLLM. + +OneVision [27], and Idefics3 [26], also adopt this approach to address the challenge of TDU. Although these above methods have achieved notable success in TDU tasks, they typically require high-resolution image inputs to identify text within images, significantly increasing computational demands. Besides, to equip general MLLMs with the ability to recognize the rich text within a document, a substantial amount of training data is required, including training sets of the benchmarks tested, which resulted in a high demand for training resources and might affect the evaluation of their zero-shot inference ability. + +OCR-dependent MLLMs for Text-Rich Documents Understanding. To address the challenges of OCR-free MLLMs, some researchers have attempted to integrate OCR information from text-rich documents into LLMs to avoid high-resolution image inputs. ICL-D3IE [14] feeds coordinates of OCR boxes into the LLM using a textual expression of [left, top, right, bottom], while LATIN-Prompt [70] simulates the document layout by inserting spaces and newline characters between OCR-recognized text segments, thus providing the LLM with a document input containing a rudimentary layout. Besides, LMDX [53] also input the layout in a way similar to ICL-D3IE. However, these methods of inputting OCR information in textual form lead to overly lengthy inputs, thereby slowing inference and impacting model performance. Subsequent studies have explored encoding OCR information using additional encoders. Cream [24] utilizes BLIP-2 [30] to encode document OCR positions, images, text, and additional object detection information, with the output of BLIP-2 fed into the LLM. InstructDr [60] introduces an extra feature extractor to encode document OCR information, compressing it with learnable tokens before inputting it into the LLM. Additionally, Fujitake [11] and Luo et al. [45] adopt Lay + +outLMv3 [20] to encode the OCR information of documents before inputting them into the LLM. Although these methods avoid the issue of overly lengthy text inputs, they must use the extra DE to encode all the OCR textual, positional information from the document due to the structural restriction of DE. However, the additional DE might not be comparable with LLMs themselves in terms of document comprehension capability. This might be the reason why they usually need extensive training of both the DE and the LLM to align the features encoded by the DE with those of the LLM, which consumes significant computational resources. DocLLM [67] proposes an extension where the hidden states derived from OCR bounding boxes interact with those of OCR text features encoded by LLMs within each attention block via cross-attention [5, 6]. However, they use an extra projection layer at each attention block to transform and inject layout information, which might excessively interfere with the original attention computation of LLM, thereby potentially affecting LLM's generalization capability. Moreover, it does not adequately consider the role of image features in TDU. Overall, designing an efficient method to provide OCR information to LLMs remains a research question worthy of deep exploration. + +# 3. DocLayLLM + +DocLayLLM is an LLM-based methodology that encodes OCRed textual, visual, and positional information directly within an LLM, eliminating the need for additional DE. The training regime of DocLayLLM includes both pretraining and supervised fine-tuning phases, which are both instruction-tuning-based processes. To further enhance our model, we propose the integration of CoT into our pretraining tasks, denoted as CoT Pre-training. Additionally, we introduce CoT Annealing, which involves the applica + +![](images/f217931446d01cd5e3933c2f49b3147244b184376610f7c07309eb1661ae21af.jpg) +(a) Layout-Type-Related CoT. + +![](images/1b53df17ea4a95f1df09dcad8302374bd5ded2f4093b1e50d8db8d0cfd5ebbd7.jpg) +(b) Table-Structure-Aware CoT. +Figure 2. Examples of CoT Pre-training across different tasks. Red text highlights key steps in the chain of thought for each task. + +![](images/ad1828532a29f3a49f7b14a2e728d2f4cd3b718da1dcec76aa6fc670e403a950.jpg) +(c) Geometry-Related CoT. + +tion of annealing to the data utilized in the supervised finetuning phase, specifically from the perspective of CoT. + +# 3.1. Model Architecture + +The overall architecture of our DocLayLLM is illustrated in Figure 1. To process a document image, we first employ an OCR engine. This engine extracts the OCR results, which encompasses the textual content, denoted as $Text_{1:N}$ , and the associated bounding box, represented by $Box_{1:N}$ . Here, $N$ indicates the total number of text segments detected by the OCR, and each $Box_{i} = [x_{0},y_{0},x_{1},y_{1}]$ defines the bounding box for $Text_{i}$ . The bounding box coordinates correspond to the left, top, right, and bottom boundaries of the text segment. After the OCR process, the document image is segmented into patches, denoted as $Patch_{1:M}$ , where $M$ represents the total number of patches. Next, the OCR results and the document patches are integrated with the fixed part of the prompt and the specific Question related to the document. This integration forms the final prompt input to the model, presented in a natural language expression $\mathcal{T}$ : Given the document patches: $\{Patch_{1:M}\}$ and the document text contents and locations in the form of "text, [left, top, right, bottom]": $\{(Text_{i}, Box_{i})_{1:N}\}$ $\{Question\}$ + +Subsequently, the textual components of the prompt, including the fixed expression, $Text_{1:N}$ , and $Question$ , are processed using the LLM's native text embedder. This generates the embeddings for the text content as follows: + +$$ +E m b _ {T} = T E (T e x t C o n t e n t), \tag {1} +$$ + +where $TE$ is the native text embedder of LLM, and TextContent consists of the fixed expression, $Text_{1:N}$ , and Question. + +As for $Patch_{1:M}$ and $Box_{1:N}$ , they are first embedded with an extra embedder, then mapped to the LLM's feature space via a linear projector, yielding $Emb_V$ and $Emb_L$ as: + +$$ +E m b _ {V} = V P (V E (P a t c h _ {1: M})), \tag {2} +$$ + +$$ +E m b _ {L} = L P \left(L E \left(\text {B o x} _ {1: N}\right)\right), \tag {3} +$$ + +where $VE$ and $VP$ represent the visual embedder and visual projector, $LE$ and $LP$ denote the layout embedder and layout projector, respectively. It is worth noting that despite the introduction of $VE$ , $VP$ , $LE$ , and $LP$ , these are all simple embedding layers or linear layers, thereby minimizing additional computational burden. + +Finally, LLM will comprehend the embeddings provided and utilize them to generate the Answer to the Question: + +$$ +\text {A n s w e r} = L L M \left(\pi \left(\left(E m b _ {T}, E m b _ {V}, E m b _ {L}\right), \mathcal {T}\right)\right). \tag {4} +$$ + +Here $\pi$ represents organizing $Emb_{T}, Emb_{V}, Emb_{L}$ into a feature sequence according to $\mathcal{T}$ . + +With our design, we have preserved the original structure of the LLM, ensuring its generalization capability. At the same time, we have enabled the LLM to easily accept and process visual and layout features, making it better suited for understanding text-rich documents with various layouts. This results in a simple yet effective multi-modal extension. + +# 3.2. CoT Pre-training + +During the pre-training phase, we incorporate a wide range of tasks (as detailed in the appendix) to boost the LLM's understanding of the visual and layout features inspired by Luo et al. [45]. Since the datasets related to these tasks are not designed explicitly for instruction-tuning, their annotations are not in a QA format. Therefore, we manually create QA-formatted instructions based on their annotations for training. However, simply converting task descriptions into Question and annotations into Answer is highly inefficient regarding data utilization. To address this, we incorporate CoT during the conversion to instruction-tuning data for specific tasks, asking LLM to solve our problem step-by-step thus enhancing the model's comprehension abilities, as the examples illustrated in Fig 2. Our designed CoT primarily falls into the following three categories: + +# 3.2.1. Layout-Type-Related CoT + +We conduct a layout-type-related CoT for the layout analysis task. To determine the layout type of a given area, it is + +beneficial first to locate its region within the document and then infer based on the nearby layout element. When giving an area bounded by a bounding box, our layout-type-related CoT can be formulated as following steps: + +1. Identify the text within the area and determine its approximate region within the document (e.g., top left, top right part of the document, etc). +2. Locate the nearest bounding boxes and assess their layout type. +3. Infer the layout type of the given area based on the above information. + +Through this CoT process, we enable the LLM to fully utilize the visual, layout, and textual features of the document, thereby improving their ability to understand the document's overall structure. + +# 3.2.2. Table-Structure-Aware CoT + +Table-structure-aware CoT is primarily employed in the table analysis task. We have integrated the XY-Cut algorithm [48]—a widely-used technique for analyzing table structures. For instance, to locate an element at the $i$ -th row and $j$ -th column, the CoT process is outlined as follows: + +1. Identify all the table headers and their respective locations. The header of the $j$ -th column corresponds to the $j$ -th element in the headers list. +2. Output all elements in the $j$ -th column when told that the elements of the same column are roughly vertically aligned with their corresponding headers. +3. Determine the final answer based on the $i$ -th element of the contents in the $j$ -th column. + +Applying the above CoT, we break down a complex question into simple steps using the basic XY-Cut thought. This approach improves the LLM's inference ability through this breakdown and enhances data utilization by effectively leveraging the overall structure of the table. + +# 3.2.3. Geometry-Related CoT + +This is used in geometric analysis to identify geometric relationships between two text segments. It assists LLM to inference with geometric relations. Specifically, when giving the text content of two lines: + +1. Recover the bounding box of the given text segments. +2. Analyze their vertical and horizontal projections to determine if they overlap through interval analysis. +3. Calculate the center point of each segment and determine their relative location by comparing their center points. +4. Calculate the minimum Euclidean distance based on the previous analysis. If they overlap, the distance is 0; if they overlap in vertical/horizontal projections, the minimum distance is the smallest interval in the horizontal/vertical projections; otherwise, it is the minimum Euclidean distance between their nearest corners. + +This design advances the LLM's arithmetic reasoning capabilities and helps it more accurately grasp the spatial re + +relationships between different boxes, and also contributes to layout-type-aware CoT to find the nearest bounding boxes. + +# 3.3. CoT Annealing + +Data annealing involves gradually increasing the proportion of high-quality data towards the end of supervised fine-tuning to improve data utilization efficiency and augment the model's reasoning abilities. This technique has been proven to be effective and is widely adopted in models like MiniCPM [19], DataComp-LM [31], and Llama-3.1 [42]. + +During the SFT stage, we employ data with a layout-aware CoT [45] that helps focus on potential regions of questions and answers within documents, thereby better leveraging the layout features learned during pre-training. It also allows the model to strengthen its reasoning capacity through step-by-step thinking. However, in the latter stages of SFT, our goal for LLM is to output precise and direct answers to questions. At this point, using CoT data may introduce unnecessary noise with a long response, and the overly uniform layout-aware CoT data might partially compromise the LLM's generalization abilities. + +To balance the strengths and weaknesses of such data, we propose CoT annealing, which reconsiders this data annealing issue from a CoT perspective. Specifically, we generate direct-answer data without CoT from the CoT data, treating it as high-quality data. During the SFT process, we gradually increase the proportion of this direct-answer data, thereby enhancing the LLM's ability to respond with a clear and direct answer in the end. Initially, we only utilize the data with CoT, gradually adjusting the ratio of CoT to non-CoT data as the SFT process progresses. Ultimately, we conclude the training using only the data without CoT. + +# 4. Experiments + +# 4.1. Datasets + +During the pre-training stage, we conduct our pre-training tasks on several datasets, including RVL-CDIP [13], PubLayNet [80], PubTabNet [81], DocLayNet [54], DocBank [32], DocILE [56], and the Document Dense Description data from LayoutLLM [45]. This results in a total of 3.1 million pairs of instruction tuning data, which is significantly less than the volume used in previous OCR-dependent TDU methods. For the supervised fine-tuning stage, we apply CoT annealing on the layout-aware document QA data utilized in LayoutLLM [45]. + +We evaluate our DocLayLLM on widely used text-rich document benchmarks, including FUNSD [22], CORD [50], SROIE [21], POIE [25], DocVQA [46], InfoVQA [47], VisualMRC [59], DeepForm [58], KLC [57], WTQ [51], and TabFact [4]. For a fair comparison with other approaches, we use the default OCR data from the of + +
MethodProcessed Data at Pre-trainingDocument-oriented VQAKIE
DocVQAVisualMRCAvg.FUNSD†CORD†SROIE†Avg.
Text
Llama2-7B-Chat [64]-20.509.9015.2015.1020.0035.623.57
Llama3-8B-Instruct [42]-51.7947.7749.7868.5652.3161.2460.70
Text + Box
Llama2-7B-Chatcoor [14]-12.3012.2012.2511.906.4039.4019.23
Llama3-8B-Instructcoor [14]-49.1341.7545.4474.0062.2063.1566.45
LayTextLLMzero [43]10.0M (323%)65.5037.4051.4572.0045.5082.0066.50
Text + Box + Patch
LayoutLLM Llama2 [45]5.7M (184%)74.2555.7364.9978.6562.2170.9770.61
DocLayLLM Llama2 (Ours)3.1M (100%)72.8355.0463.9478.6770.8183.2777.58
DocLayLLM Llama3 (Ours)3.1M (100%)78.3658.5568.4684.1171.3484.3679.94
+ +Table 1. Comparison with OCR-dependent methods using metrics from LayoutLLM [45]. Processed Data at Pre-training represents the total data volume processed at the pre-training stage, equivalent to dataset size multiplied by epoch count, with values in parentheses representing size ratios compared to DocLayLLM. Avg. represents the average performance across different benchmarks, and $\dagger$ indicates the cleaned test set used in LayoutLLM. The suffix layout denotes using method from He et al. [14], which provides LLM with layout features via OCR box coordinates in textual form. Bold indicates the best performance, and underline indicates the second-best one. + +
MethodProcessed Data at Pre-trainingDocument-oriented VQAKIE
DocVQAVisualMRCAvg.FUNSDCORDSROIEAvg.
Text
Llama2-7B-Chat [64]-20.506.3013.4023.4051.8058.6044.60
Llama3-8B-Instruct [42]-51.79229.74140.7766.6774.7182.5174.63
Text + Box
Llama2-7B-Chatcoor [14]-12.3028.0020.1514.4038.1050.6034.37
Llama3-8B-Instructcoor [14]-49.13211.69130.4174.7175.3375.9375.32
LMDX [53]-----66.95--
DocLLM [67]16.8M (542%)69.50*264.10*166.80*51.80*67.60*91.90*70.43*
LayTextLLMzero [43]10.0M (323%)65.50200.20132.8547.2077.2083.8069.40
Text + Box + Patch
DocLayLLM Llama2 (Ours)3.1M (100%)72.83310.60191.7280.7479.3784.3781.49
DocLayLLM Llama3 (Ours)3.1M (100%)78.35357.04217.7086.4783.6686.0885.40
+ +Table 2. Comparison with OCR-dependent methods using metrics from DocLLM [67]. * indicates that the model uses the training set of the benchmark during training. + +ficial datasets otherwise commercial OCR engines ${}^{1}$ if not provided, and follow their evaluation pipeline and metrics. + +# 4.2. Implement Details + +Our method is based on Llama3-8B-Instruct [42], with a Llama2-7B-Chat [64] version for better comparison with other methods. They are denoted as DocLayLLM $_{Llama3}$ and DocLayLLM $_{Llama2}$ respectively. We initialize $VE$ and $LE$ with weights from LayoutLMv3 [20] to leverage document knowledge acquired during its pre-training, while $VP$ and $LP$ are randomly initialized. Besides, the image resolution is fixed at $224*224$ , resulting in 196 visual patch tokens, to capture an overview of document layout while bringing less computation burden. During pre-training and fine-tuning, we use LoRA [18] with a rank of 64 on LLM and keep $VE$ , $VP$ , $LE$ , and $LP$ trainable, leading to few tuning parameters. We train our model on 8 NVIDIA Tesla A100-80G GPUs, taking about 30 A100 days for pre + +training and 6 A100 days for supervised fine-tuning. Additional details can be found in the appendix. + +# 4.3. Comparisons with SOTA OCR-dependent methods + +We have compared our DocLayLLM with existing SOTA OCR-dependent methods. In order to exhibit the zero-shot inferring ability precisely, we chose their performances under the zero-shot scenario unless they are not provided. + +The experimental results are shown in Table 1 and Table 2. Our method outperforms the baseline models, regardless of their type or whether the document's layout information is provided to them. It validates the rationale behind our design of multi-modal extension. + +Additionally, our method significantly outperforms previous SOTA methods despite fewer training resources. As shown in Table 1, it surpasses the performance of Layout-LLM [45] and LayTextLLM [43] though requiring fewer computational resources by using less data and tuning few + +
MethodDocument-oriented VQAKIE
DocVQAInfoVQAAvg.FUNSDSROIEPOIEAvg.
LLaVAR [79]12.3016.5014.400.505.205.903.87
UniDoc [10]7.7014.7011.201.002.905.103.00
DocPedia [9]47.10*15.20*31.15*29.9021.4039.9030.40
Monkey [38]50.10*25.80*37.95*24.1041.9019.9028.63
TextMonkey [41]66.70*28.60*47.65*42.9046.2032.0040.37
DocOwl 1.5 [17]----48.3051.80-
TextSquare [61]----53.2071.80-
DocKylin [78]65.10*34.80*49.95*58.7025.5049.5044.57
DocLayLLM\( _{Llama3} \) (Ours)77.7942.0259.9180.2876.5975.1377.33
+ +Table 3. Comparison with OCR-free methods using the precision metric from TextMonkey [41]. * indicates that the model uses the training set of the benchmark during training. + +
MethodDocument-oriented VQAKIETable Understanding
DocVQAInfoVQAVisualMRCAvg.DeepFormKLCAvg.WTQTabFactAvg.
DocPedia [9]47.1015.20--------
DocOwl [75]62.2038.20188.8096.4042.6030.3036.4526.9060.2043.55
URector [76]65.4042.20221.70109.7749.5032.8041.1529.4067.6048.50
Monkey [38]66.5036.10--40.6032.8036.7025.30--
TextMonkey [41]73.00---59.7037.8048.7531.90--
DocKylin [78]77.3046.60319.90147.9364.20--32.40--
CogAgent [16]81.6044.50--------
DocOwl 1.5 [17]82.2050.70246.40126.4368.8038.7053.7540.6080.2060.40
TextSquare [61]84.3051.50-----49.7084.2066.95
DocLayLLM\(\ddagger\) Llama3 (Ours)86.5258.36327.91157.6077.0740.7358.9058.6283.3670.99
+ +Table 4. Comparison with OCR-free methods using the evaluation framework and metrics from DocOwl 1.5 [17]. All methods, including our $DocLayLLM_{Llama3}^{\ddagger}$ variant, are exposed to the train set corresponding to the evaluation set during training. + +parameters at training. Even with Llama2 as the backbone, our approach matches the performance of previous SOTA methods in the Document-oriented VQA task and significantly outperforms them in the KIE tasks, further demonstrating the efficiency and effectiveness of our approach. + +Besides, our method exhibits robust zero-shot generalization capability without fine-tuning with specific downstream tasks. From the comparison with DocLLM [67] in Table 2, it is clear that our method, even without fine-tuning, significantly outperforms DocLLM across most datasets and achieves a considerable overall performance improvement. This validates our design, which has preserved the strong generalization ability of the original LLM structure. + +In summary, DocLayLLM serves as an efficient extension of LLMs. It demonstrates strong zero-shot generalization and effective document understanding by encoding documents within a unified model. Even with limited training resources, DocLayLLM achieves new SOTA performance compared to OCR-dependent competitors. + +# 4.4. Comparisons with SOTA OCR-free methods + +We have also compared our method with SOTA OCR-free TDU models. Table 3 presents the precision of our method and some advent OCR-free TDU methods across various benchmarks. Even though in the circumstance that OCR- + +free models use some of the training sets of these benchmarks during training, while our method infers in a purely zero-shot manner, our DocLayLLM still outperforms these models, particularly in KIE task, which heavily relies on layout information. This clearly demonstrates the superiority of our design to provide LLM with OCR information. + +We additionally test DocLayLLM on a broader set of benchmarks from DocOwl 1.5 [17]. To further assess the performance of DocLayLLM on seen datasets, we incorporate the training sets of these benchmarks used in DocOwl 1.5 during the SFT process as other OCR-free TDU methods do during training. To maintain the efficiency advantage of our method, we ensure that the SFT is conducted with the same number of steps as the standard DocLayLLM. The results in Table 4 show that with the inclusion of downstream task data during SFT, our method outperforms the SOTA OCR-free TDU models in most datasets. Moreover, our overall performance in the Document-oriented VQA, KIE, and Table Understanding sub-tasks surpasses the OCR-free TDU methods using limited training resources, showcasing the tremendous potential of OCR-dependent methods and further reinforcing the reliability of our method. + +Overall, our DocLayLLM consistently shows strong zero-shot inference ability, outperforming many OCR-free TDU methods without prior exposure to downstream data. + +![](images/8162119e9d04c2187be181409bb60212059e6dda14d0319715346cd8850d696d.jpg) +Figure 3. Qualitative comparisons of DocLayLLM with other OCR-dependent and OCR-free methods. Zoom in for better view. + +Additionally, when incorporating downstream datasets during SFT like other OCR-free approaches, DocLayLLM surpasses them, highlighting its potential in solving TDU tasks by integrating OCR information into the LLM. + +# 4.5. Qualitative Results + +Figure 3 shows some qualitative examples to visualize the performance on KIE tasks of our DocLayLLM compared with LayTextLLM and DocOwl 1.5. The example from SROIE clearly shows that DocLayLLM has improved document comprehension and answer accuracy. And the example from POIE demonstrates our advantage in handling table-formatted documents to give an accurate response. Furthermore, as seen in the FUNSD examples, DocLayLLM excels at extracting answers spanning multiple text lines, outperforming both OCR-dependent and OCR-free methods. For more visualization of qualitative examples, please refer to the appendix. + +# 4.6. Ablation Study + +We have also conducted additional experiments to explore the effectiveness of our proposed method. To conserve computational resources, we perform pre-training at $25\%$ of the standard scale (including the baseline) but full-scale SFT. We focus on the impact of two main factors: (a) the role of CoT pre-training and CoT annealing, and (b) the influence of different modality configurations. + +Experimental results in Table 5a indicate that omitting CoT Pre-training leads to an obvious performance drop in both the Document-oriented VQA and KIE tasks, while not using CoT Annealing primarily affects VQA performance. Besides, using randomly initialized $VE$ and $LE$ partially compromises the proposed extension, resulting in performance degradation in both the VQA and KIE tasks. + +Additionally, performances in Table 5b show the efficacy + +
ExperimentDocument-oriented VQAKIE
DocVQAVisualMRCAvg.DeepFormKLCAvg.
DocLayLLM1lama377.43316.27196.8535.4126.2530.83
w/o CoT Pre-training76.00295.49185.7524.5425.4925.02
w/o CoT Annealing77.40310.17193.7934.2326.2130.22
w/o pt VE & LE77.42308.49192.9631.3125.7628.54
+ +(a) Impact of different settings: $w / o$ CoT Pre-training and $w / o$ CoT Annealing indicate pre-training without CoT and supervised fine-tuning with only CoT data, while $w / o$ pt VE and PE uses randomly initialized visual and layout embeddings instead of pre-trained ones. + +
ModalityDocument-oriented VQAKIE
TLVDocVQAVisualMRCAvg.DeepFormKLCAvg.
✓✗✗73.57280.27176.9315.7624.0919.93
✓(I)✗75.72306.51191.1214.1324.2319.18
✓(II)✗77.32310.75194.0333.7525.9729.86
✓(II)✓77.43316.27196.8535.4126.2530.83
+ +(b) Impacts of different modality configurations. $\mathbf{T}$ , $\mathbf{L}$ , and $\mathbf{V}$ represent textual, layout, and visual modalities, respectively. (I) indicates inputting OCR coordinates as text following He et al. [14], while (II) embeds coordinates with $LE$ in our approach. + +Table 5. Ablation study on various designs of our DocLayLLM. + +of our DocLayLLM to incorporate different modalities for TDU tasks. It is obvious that using only text from the document is insufficient for TDU, while adding layout information in textual form provides some gain on VQA tasks but leads to over-length input (see appendix for analysis), greatly slowing the inference. By our design to integrate layout features via an $LE$ , we achieve notable improvements on these TDU tasks with shorter inputs. And adding visual features further boosts the performance of our method. + +Overall, the combined impact of these designs ensures the efficiency and effectiveness of our method. + +# 5. Conclusion + +We propose DocLayLLM, an efficient multi-modal extension of large language models for text-rich document understanding. By streamlined integration of document patch tokens and OCR positional tokens into the LLM inputs, it achieves a simple yet efficient multi-modal extension of LLM to enable it to have the capability to perceive and process textual, visual, and layout information from the OCR results of the document. Additionally, we introduce CoT Pre-training and CoT Annealing techniques centered around CoT, offering a new way to generate CoT data and further enhancing the model's understanding of text-rich documents. Despite requiring fewer training resources, our method surpasses existing OCR-dependent TDU models, demonstrating strong zero-shot generalization capability. Furthermore, it also exceeds the performance of SOTA OCR-free TDU methods under comparable training conditions, showcasing remarkable performance across various document understanding tasks. Our work significantly advances efficient MLLM-based document understanding while maintaining high-performance standards. + +# Acknowledgement + +This research is supported in part by National Natural Science Foundation of China (Grant No.: 62476093, 62441604). + +# References + +[1] Srikar Appalaraju, Bhavan Jasani, Bhargava Urala Kota, Yusheng Xie, and R. Manmatha. DocFormer: End-to-End Transformer for Document Understanding. In Proc. ICCV, pages 993–1003, 2021. 1 +[2] Srikar Appalaraju, Peng Tang, Qi Dong, Nishant Sankaran, Yichu Zhou, and R. Manmatha. DocFormerv2: Local Features for Document Understanding. In Proc. AAAI, pages 709-718, 2024. 1 +[3] Tom Brown, Benjamin Mann, Nick Ryder, Melanie Subbiah, Jared D Kaplan, Prafulla Dhariwal, Arvind Neelakantan, Pranav Shyam, Girish Sastry, Amanda Askell, Sandhini Agarwal, Ariel Herbert-Voss, Gretchen Krueger, Tom Henighan, Rewon Child, Aditya Ramesh, Daniel Ziegler, Jeffrey Wu, Clemens Winter, Chris Hesse, Mark Chen, Eric Sigler, Mateusz Litwin, Scott Gray, Benjamin Chess, Jack Clark, Christopher Berner, Sam McCandlish, Alec Radford, Ilya Sutskever, and Dario Amodei. Language Models are Few-shot Learners. Proc. NIPS, 33:1877-1901, 2020. 2 +[4] Wenhu Chen, Hongmin Wang, Jianshu Chen, Yunkai Zhang, Hong Wang, Shiyang Li, Xiyou Zhou, and William Yang Wang. TabFact: A Large-scale Dataset for Table-based Fact Verification. In Proc. ICLR, 2020. 5 +[5] Gang Dai, Yifan Zhang, Qingfeng Wang, Qing Du, Zhuliang Yu, Zhuoman Liu, and Shuangping Huang. Disentangling Writer and Character Styles for Handwriting Generation. In Proc. CVPR, pages 5977-5986, 2023. 3 +[6] Gang Dai, Yifan Zhang, Quhui Ke, Qiangya Guo, and Shuangping Huang. One-dm: One-shot diffusion mimicker for handwritten text generation. In Proc. ECCV, pages 410-427, 2025. 3 +[7] DeepSeek-AI. DeepSeek-R1: Incentivizing Reasoning Capability in LLMs via Reinforcement Learning. arXiv: Comp. Res. Repository, abs/2501.12948, 2025. 2 +[8] Jacob Devlin, Ming-Wei Chang, Kenton Lee, and Kristina Toutanova. BERT: Pre-training of Deep Bidirectional Transformers for Language Understanding. In Proc. NAACL, pages 4171–4186, 2019. 1 +[9] Hao Feng, Qi Liu, Hao Liu, Wengang Zhou, Houqiang Li, and Can Huang. DocPedia: Unleashing the Power of Large Multimodal Model in the Frequency Domain for Versatile Document Understanding. arXiv: Comp. Res. Repository, abs/2311.11810, 2023. 7 +[10] Hao Feng, Zijian Wang, Jingqun Tang, Jinghui Lu, Wengang Zhou, Houqiang Li, and Can Huang. UniDoc: A Universal Large Multimodal Model for Simultaneous Text Detection, Recognition, Spotting and Understanding. arXiv: Comp. Res. Repository, abs/2308.11592, 2023. 7 +[11] Masato Fujitake. LayoutLLM: Large Language Model Instruction Tuning for Visually Rich Document Understanding. In Proc. LREC/COLING, pages 10219-10224, 2024. 2, 3 + +[12] Lukasz Garncarek, Rafat Powalski, Tomasz Stanisławek, Bartosz Topolski, Piotr Halama, Michal Turski, and Filip Graliński. LAMBERT: Layout-Aware Language Modeling for Information Extraction. In Proc. ICDAR, pages 532-547, 2021. 1 +[13] Adam W. Harley, Alex Ufkes, and Konstantinos G. Derpanis. Evaluation of Deep Convolutional Nets for Document Image Classification and Retrieval. In Proc. ICDAR, pages 991-995, 2015. 5 +[14] Jiabang He, Lei Wang, Yi Hu, Ning Liu, Hui Liu, Xing Xu, and Heng Tao Shen. ICL-D3IE: In-Context Learning with Diverse Demonstrations Updating for Document Information Extraction. In Proc. ICCV, pages 19428-19437, 2023. 1, 2, 3, 6, 8 +[15] Teakgyu Hong, DongHyun Kim, Mingi Ji, Wonseok Hwang, Daehyun Nam, and Sungrae Park. BROS: A Pre-trained Language Model Focusing on Text and Layout for Better Key Information Extraction from Documents. In Proc. AAAI, pages 10767-10775, 2022. 1 +[16] Wenyi Hong, Weihan Wang, Qingsong Lv, Jiazheng Xu, Wenmeng Yu, Junhui Ji, Yan Wang, Zihan Wang, Yuxiao Dong, Ming Ding, and Jie Tang. CogAgent: A Visual Language Model for GUI Agents. In Proc. CVPR, pages 14281-14290, 2024. 7 +[17] Anwen Hu, Haiyang Xu, Jiabo Ye, Ming Yan, Liang Zhang, Bo Zhang, Ji Zhang, Qin Jin, Fei Huang, and Jingren Zhou. mPLUG-DocOwl 1.5: Unified Structure Learning for OCR-free Document Understanding. In Findings EMNLP, pages 3096-3120, 2024. 1, 2, 7 +[18] Edward J. Hu, Yelong Shen, Phillip Wallis, Zeyuan Allen-Zhu, Yuanzhi Li, Shean Wang, Lu Wang, and Weizhu Chen. LoRA: Low-Rank Adaptation of Large Language Models. In Proc. ICLR, 2022. 2, 6 +[19] Shengding Hu, Yuge Tu, Xu Han, Chaoqun He, Ganqu Cui, Xiang Long, Zhi Zheng, Yewei Fang, Yuxiang Huang, Weilin Zhao, Xinrong Zhang, Zhen Leng Thai, Kai Zhang, Chongyi Wang, Yuan Yao, Chenyang Zhao, Jie Zhou, Jie Cai, Zhongwu Zhai, Ning Ding, Chao Jia, Guoyang Zeng, Dahai Li, Zhiyuan Liu, and Maosong Sun. MiniCPM: Unveiling the Potential of Small Language Models with Scalable Training Strategies. arXiv: Comp. Res. Repository, abs/2404.06395, 2024. 5 +[20] Yupan Huang, Tengchao Lv, Lei Cui, Yutong Lu, and Furu Wei. LayoutLMv3: Pre-training for Document AI with Unified Text and Image Masking. In Proc. ACM MM, pages 4083-4091, 2022. 1, 2, 3, 6 +[21] Zheng Huang, Kai Chen, Jianhua He, Xiang Bai, Dimosthenis Karatzas, Shijian Lu, and C. V. Jawahar. ICDAR2019 Competition on Scanned Receipt OCR and Information Extraction. In Proc. ICDAR, pages 1516-1520, 2019. 5 +[22] Guillaume Jaume, Hazim Kemal Ekenel, and Jean-Philippe Thiran. FUNSD: A Dataset for Form Understanding in Noisy Scanned Documents. In Workshop on ICDAR, pages 1–6, 2019. 5 +[23] Geewook Kim, Teakgyu Hong, Moonbin Yim, JeongYeon Nam, Jinyoung Park, Jinyeong Yim, Wonseok Hwang, Sangdoo Yun, Dongyoon Han, and Seunghyun Park. OCR-Free + +Document Understanding Transformer. Proc. ECCV, 13688: 498-517, 2022. 2 +[24] Geewook Kim, Hodong Lee, Daehee Kim, Haeji Jung, Sanghee Park, Yoonsik Kim, Sangdoo Yun, Taeho Kil, Bado Lee, and Seunghyun Park. Visually-Situated Natural Language Understanding with Contrastive Reading Model and Frozen Large Language Models. In Proc. EMNLP, pages 11989-12010, 2023. 2, 3 +[25] Jianfeng Kuang, Wei Hua, Dingkang Liang, Mingkun Yang, Deqiang Jiang, Bo Ren, and Xiang Bai. Visual Information Extraction in the Wild: Practical Dataset and End-to-End Solution. In Proc. ICDAR, pages 36-53, 2023. 5 +[26] Hugo Laurençon, Andres Marafioti, Victor Sanh, and Leo Tronchon. Building and better understanding vision-language models: insights and future directions. In Workshop on NIPS, 2024. 3 +[27] Bo Li, Yuanhan Zhang, Dong Guo, Renrui Zhang, Feng Li, Hao Zhang, Kaichen Zhang, Peiyuan Zhang, Yanwei Li, Ziwei Liu, and Chunyuan Li. LLaVA-OneVision: Easy Visual Task Transfer. Transactions on Machine Learning Research, 2025. 3 +[28] Chenliang Li, Bin Bi, Ming Yan, Wei Wang, Songfang Huang, Fei Huang, and Luo Si. StructuralLM: Structural Pre-training for Form Understanding. In Proc. ACL, pages 6309-6318, 2021. 1 +[29] Feng Li, Renrui Zhang, Hao Zhang, Yuanhan Zhang, Bo Li, Wei Li, Zejun Ma, and Chunyuan Li. LLaVa-NeXT-Interleave: Tackling Multi-image, Video, and 3D in Large Multimodal Models. arXiv: Comp. Res. Repository, abs/2407.07895, 2023. 2 +[30] Junnan Li, Dongxu Li, Silvio Savarese, and Steven C. H. Hoi. BLIP-2: Bootstrapping Language-Image Pre-training with Frozen Image Encoders and Large Language Models. Proc. ICML, 202:19730-19742, 2023. 3 +[31] Jeffrey Li, Alex Fang, Georgios Smyrnis, Maor Ivgi, Matt Jordan, Samir Yitzhak Gadre, Hritik Bansal, Etash Guha, Sedrick Keh, Kushal Arora, Saurabh Garg, Rui Xin, Niklas Muennighoff, Reinhard Heckel, Jean Mercat, Mayee Chen, Suchin Gururangan, Mitchell Wortsman, Alon Albalak, Yonatan Bitton, Marianna Nezhurina, Amro Abbas, ChengYu Hsieh, Dhruba Ghosh, Josh Gardner, Maciej Kilian, Hanlin Zhang, Rulin Shao, Sarah M. Pratt, Sunny Sanyal, Gabriel Ilharco, Giannis Daras, Kalyani Marathe, Aaron Gokaslan, Jieyu Zhang, Khyathi Raghavi Chandu, Thao Nguyen, Igor Vasiljevic, Sham M. Kakade, Shuran Song, Sujay Sanghavi, Fartash Faghri, Sewoong Oh, Luke Zettlemoyer, Kyle Lo, Alaaeldin El-Nouby, Hadi Pouransari, Alexander Toshev, Stephanie Wang, Dirk Groeneveld, Luca Soldaini, Pang Wei Koh, Jenia Jitsev, Thomas Kollar, Alexandros G. Dimakis, Yair Carmon, Achal Dave, Ludwig Schmidt, and Vaishaal Shankar. DataComp-LM: In search of the next generation of training sets for language models. arXiv: Comp. Res. Repository, abs/2406.11794, 2024. 2, 5 +[32] Minghao Li, Yiheng Xu, Lei Cui, Shaohan Huang, Furu Wei, Zhoujun Li, and Ming Zhou. DocBank: A Benchmark Dataset for Document Layout Analysis. In Pro. COLING, pages 949-960, 2020. 5 + +[33] Peizhao Li, Jiuxiang Gu, Jason Kuen, Vlad I Morariu, Handong Zhao, Rajiv Jain, Varun Manjunatha, and Hongfu Liu. SelfDoc: Self-Supervised Document Representation Learning. In Proc. CVPR, pages 5652-5660, 2021. 1 +[34] Qiwei Li, Zuchao Li, Xiantao Cai, Bo Du, and Hai Zhao. Enhancing Visually-Pich Pdocument Punderstanding via Payout Pstructure Podeling. In Proc. ACM MM, page 4513-4523, 2023. +[35] Qiwei Li, Zuchao Li, Ping Wang, Haojun Ai, and Hai Zhao. Hypergraph based Understanding for Document Semantic Entity Recognition. In Proc. ACL, pages 2950-2960, 2024. +[36] Xin Li, Yan Zheng, Yiqing Hu, Haoyu Cao, Yunfei Wu, Deqiang Jiang, Yinsong Liu, and Bo Ren. Relational Representation Learning in Visually-Rich Documents. In Proc. ACM MM, page 4614-4624, 2022. +[37] Yulin Li, Yuxi Qian, Yuechen Yu, Xiameng Qin, Chengquan Zhang, Yan Liu, Kun Yao, Junyu Han, Jingtuo Liu, and Errui Ding. StrucTexT: Structured Text Understanding with Multi-Modal Transformers. In Proc. ACM MM, page 1912–1920, 2021. 1 +[38] Zhang Li, Biao Yang, Qiang Liu, Zhiyin Ma, Shuo Zhang, Jingxu Yang, Yabo Sun, Yuliang Liu, and Xiang Bai. Monkey: Image Resolution and Text Label Are Important Things for Large Multi-modal Models. In Proc. CVPR, pages 26763-26773, 2024. 1, 2, 7 +[39] Zening Lin, Jiapeng Wang, and Lianwen Jin. Visual Information Extraction Deep Learning Method: A Critical Review. Journal of Image and Graphics, 28(08):2276-2297, 2023. 1 +[40] Chenglin Liu, Lianwen Jin, Xiang Bai, Xiaohui Li, and Fei Yin. Frontiers of intelligent document analysis and recognition: review and prospects. Journal of Image and Graphics, 28(08):2223-2252, 2023. 1 +[41] Yuliang Liu, Biao Yang, Qiang Liu, Zhang Li, Zhiyin Ma, Shuo Zhang, and Xiang Bai. TextMonkey: An OCR-Free Large Multimodal Model for Understanding Document. arXiv: Comp. Res. Repository, abs/2403.04473, 2024. 7 +[42] Meta AI Llama Team. The Llama 3 Herd of Models. arXiv: Comp. Res. Repository, abs/2407.21783, 2024. 2, 5, 6 +[43] Jinghui Lu, Haiyang Yu, Yanjie Wang, Yongjie Ye, Jingqun Tang, Ziwei Yang, Binghong Wu, Qi Liu, Hao Feng, Han Wang, Hao Liu, and Can Huang. A Bounding Box is Worth One Token: Interleaving Layout and Text in a Large Language Model for Document Understanding. arXiv: Comp. Res. Repository, abs/2407.01976, 2024. 6 +[44] Chuwei Luo, Changxu Cheng, Qi Zheng, and Cong Yao. GeoLayoutLM: Geometric Pre-Training for Visual Information Extraction. In Proc. CVPR, pages 7092–7101, 2023. 1 +[45] Chuwei Luo, Yufan Shen, Zhaoqing Zhu, Qi Zheng, Zhi Yu, and Cong Yao. LayoutLLM: Layout Instruction Tuning with Large Language Models for Document Understanding. In Proc. CVPR, pages 15630-15640, 2024. 2, 3, 4, 5, 6 +[46] Minesh Mathew, Dimosthenis Karatzas, and C. V. Jawahar. DocVQA: A Dataset for VQA on Document Images. In Proc. WACV, pages 2199-2208, 2021. 5 +[47] Minesh Mathew, Viraj Bagal, Ruben Tito, Dimosthenis Karatzas, Ernest Valveny, and C.V. Jawahar. InfographicVQA. In Proc. WACV, pages 1697-1706, 2022. 5 + +[48] G. Nagy, S. Seth, and M. Viswanathan. A Prototype Document Image Analysis System for Technical Journals. Computer, 25(7):10-22, 1992. 5 +[49] OpenAI. Hello GPT-4o, 2024. 2 +[50] Seunghyun Park, Seung Shin, Bado Lee, Junyeop Lee, Jaeheung Surh, Minjoon Seo, and Hwalsuk Lee. CORD: A Consolidated Receipt Dataset for Post-OCR Parsing. In Workshop on NIPS, 2019. 5 +[51] Panupong Pasupat and Percy Liang. Compositional Semantic Parsing on Semi-Structured Tables. In Proc. ACL, pages 1470–1480, 2015. 5 +[52] Qiming Peng, Yinxu Pan, Wenjin Wang, Bin Luo, Zhenyu Zhang, Zhengjie Huang, Yuhui Cao, Weichong Yin, Yongfeng Chen, Yin Zhang, Shikun Feng, Yu Sun, Hao Tian, Hua Wu, and Haifeng Wang. ERNIE-Layout: Layout Knowledge Enhanced Pre-training for Visually-rich Document Understanding. In Findings of EMNLP, pages 3744–3756, 2022. 1 +[53] Vincent Perot, Kai Kang, Florian Luisier, Guolong Su, Xiaoyu Sun, Ramya Sree Boppana, Zilong Wang, Zifeng Wang, Jiaqi Mu, Hao Zhang, Chen-Yu Lee, and Nan Hua. LMDX: Language Model-based Document Information Extraction and Localization. In Findings of ACL, pages 15140-15168, 2024. 3, 6 +[54] Birgit Pfitzmann, Christoph Auer, Michele Dolfi, Ahmed S. Nassar, and Peter W. J. Staar. DocLayNet: A Large Human-Annotated Dataset for Document-Layout Segmentation. In Proc. ACM SIGKDD Int. Conf. Knowledge discovery & data mining, pages 3743–3751, 2022. 5 +[55] Rafat Powalski, Łukasz Borchmann, Dawid Jurkiewicz, Tomasz Dwojak, Michal Pietruszka, and Gabriela Pałka. Going Full-TILT Boogie on Document Understanding with Text-Image-Layout Transformer. In Proc. ICDAR, pages 732–747, 2021. 1 +[56] Stepán Simsa, Milan Sulc, Michal Uricár, Yash Patel, Ahmed Hamdi, Matej Kocián, Matyas Skalický, Jiri Matas, Antoine Doucet, Mickaël Coustaty, and Dimosthenis Karatzas. DocILE Benchmark for Document Information Localization and Extraction. In Proc. ICDAR, pages 147-166, 2023. 5 +[57] Tomasz Stanislawek, Filip Gralinski, Anna Wróblewska, Dawid Lipinski, Agnieszka Kaliska, Paulina Rosalska, Bartosz Topolski, and Przemyslaw Biecek. Kleister: Key Information Extraction Datasets Involving Long Documents with Complex Layouts. In Proc. ICDAR, pages 564-579, 2021. 5 +[58] Stacey Svetlichnaya. DeepForm: Understand Structured Documents at Scale. https://wandb.ai/stacey/deepform_v1, 2020. Accessed: 2024-08-14. 5 +[59] Ryota Tanaka, Kyosuke Nishida, and Sen Yoshida. VisualMRC: Machine Reading Comprehension on Document Images. In Proc. AAAI, pages 13878-13888, 2021. 5 +[60] Ryota Tanaka, Taichi Iki, Kyosuke Nishida, Kuniko Saito, and Jun Suzuki. InstructDoc: A Dataset for Zero-shot Generalization of Visual Document Understanding with Instructions. In Proc. AAAI, pages 19071-19079, 2024. 3 +[61] Jingqun Tang, Chunhui Lin, Zhen Zhao, Shu Wei, Binghong Wu, Qi Liu, Hao Feng, Yang Li, Siqi Wang, Lei Liao, Wei + +Shi, Yuliang Liu, Hao Liu, Yuan Xie, Xiang Bai, and Can Huang. TextSquare: Scaling up Text-Centric Visual Instruction Tuning. arXiv: Comp. Res. Repository, abs/2404.12803, 2024. 1, 2, 7 +[62] Zineng Tang, Ziyi Yang, Guoxin Wang, Yuwei Fang, Yang Liu, Chenguang Zhu, Michael Zeng, Cha Zhang, and Mohit Bansal. Unifying Vision, Text, and Layout for Universal Document Processing. In Proc. CVPR, pages 19254-19264, 2023. 2 +[63] Qwen Team. QwQ: Reflect Deeply on the Boundaries of the Unknown, 2024. 2 +[64] Hugo Touvron, Louis Martin, Kevin Stone, Peter Albert, Amjad Almahairi, Yasmine Babaei, Nikolay Bashlykov, Soumya Batra, Prajjwal Bhargava, Shruti Bhosale, Dan Bikel, Lukas Blecher, Cristian Canton-Ferrer, Moya Chen, Guillem Cucurull, David Esiobu, Jude Fernandes, Jeremy Fu, Wenyin Fu, Brian Fuller, Cynthia Gao, Vedanuj Goswami, Naman Goyal, Anthony Hartshorn, Saghar Hosseini, Rui Hou, Hakan Inan, Marcin Kardas, Viktor Kerkez, Madian Khabsa, Isabel Kloumann, Artem Korenev, Punit Singh Koura, Marie-Anne Lachaux, Thibaut Lavril, Jenya Lee, Diana Liskovich, Yinghai Lu, Yuning Mao, Xavier Martinet, Todor Mihaylov, Pushkar Mishra, Igor Molybog, Yixin Nie, Andrew Poulton, Jeremy Reizenstein, Rashi Rungta, Kalyan Saladi, Alan Schelten, Ruan Silva, Eric Michael Smith, Ranjan Subramanian, Xiaoqing Ellen Tan, Binh Tang, Ross Taylor, Adina Williams, Jian Xiang Kuan, Puxin Xu, Zheng Yan, Iliyan Zarov, Yuchen Zhang, Angela Fan, Melanie Kambadur, Sharan Narang, Aurélien Rodriguez, Robert Stojnic, Sergey Edunov, and Thomas Scialom. Llama 2: Open Foundation and Fine-Tuned Chat Models. arXiv: Comp. Res. Repository, abs/2307.09288, 2023. 6 +[65] Yi Tu, Ya Guo, Huan Chen, and Jinyang Tang. LayoutMask: Enhance Text-Layout Interaction in Multi-modal Pretraining for Document Understanding. In Proc. ACL, pages 15200–15212, 2023. 1 +[66] Jianqiang Wan, Sibo Song, Wenwen Yu, Yuliang Liu, Wenqing Cheng, Fei Huang, Xiang Bai, Cong Yao, and Zhibo Yang. OmniParser: A Unified Framework for Text Spotting Key Information Extraction and Table Recognition. In Proc. CVPR, pages 15641-15653, 2024. 2 +[67] Dongsheng Wang, Natraj Raman, Mathieu Sibue, Zhiqiang Ma, Petr Babkin, Simerjot Kaur, Yulong Pei, Armineh Nourbakhsh, and Xiaomo Liu. DocLLM: A Layout-Aware Generative Language Model for Multimodal Document Understanding. In Proc. ACL, pages 8529-8548, 2024. 3, 6, 7 +[68] Jiapeng Wang, Lianwen Jin, and Kai Ding. LiLT: A Simple yet Effective Language-Independent Layout Transformer for Structured Document Understanding. In Proc. ACL, pages 7747–7757, 2022. 1 +[69] Jiapeng Wang, Zening Lin, Dayi Huang, Longfei Xiong, and Lianwen Jin. LiLTv2: Language-substitutable layout-image transformer for visual information extraction. ACM Transactions on Multimedia Computing, Communications and Applications, 21(3):1-27, 2025. 1 +[70] Wenjin Wang, Yunhao Li, Yixin Ou, and Yin Zhang. Layout and Task Aware Instruction Prompt for Zero-shot Document + +Image Question Answering. arXiv: Comp. Res. Repository, abs/2306.00526, 2023. 1, 3 +[71] Jason Wei, Xuezhi Wang, Dale Schuurmans, Maarten Bosma, brian richter, Fei Xia, Ed Chi, Quoc V Le, and Denny Zhou. Chain-of-Thought Prompting Elicits Reasoning in Large Language Models. Proc. NIPS, 35:24824-24837, 2022. 2 +[72] Yiheng Xu, Minghao Li, Lei Cui, Shaohan Huang, Furu Wei, and Ming Zhou. LayoutLM: Pre-training of Text and Layout for Document Image Understanding. In Proc. ACM SIGKDD Int. Conf. Knowledge discovery & data mining, page 1192-1200, 2020. 1 +[73] Yang Xu, Yiheng Xu, Tengchao Lv, Lei Cui, Furu Wei, Guoxin Wang, Yijuan Lu, Dinei Florencio, Cha Zhang, Wanxiang Che, Min Zhang, and Lidong Zhou. LayoutLMv2: Multi-modal Pre-training for Visually-rich Document Understanding. In Proc. ACL, pages 2579-2591, 2021. 1 +[74] Hao Yan, Yuliang Liu, Lianwen Jin, and Xiang Bai. The development, application, and future of LLM similar to ChatGPT. Journal of Image and Graphics, 28(9):2749-2762, 2023. 1 +[75] Jiabo Ye, Anwen Hu, Haiyang Xu, Qinghao Ye, Ming Yan, Yuhao Dan, Chenlin Zhao, Guohai Xu, Chenliang Li, Junfeng Tian, Qian Qi, Ji Zhang, and Fei Huang. mPLUGDocOwl: Modularized Multimodal Large Language Model for Document Understanding. arXiv: Comp. Res. Repository, abs/2307.02499, 2023. 7 +[76] Jiabo Ye, Anwen Hu, Haiyang Xu, Qinghao Ye, Ming Yan, Guohai Xu, Chenliang Li, Junfeng Tian, Qi Qian, Ji Zhang, Qin Jin, Liang He, Xin Lin, and Fei Huang. UReader: Universal OCR-free Visually-situated Language Understanding with Multimodal Large Language Model. In Findings of EMNLP, pages 2841-2858, 2023. 1, 2, 7 +[77] Chong Zhang, Ya Guo, Yi Tu, Huan Chen, Jinyang Tang, Huijia Zhu, Qi Zhang, and Tao Gui. Reading Order Matters: Information Extraction from Visually-rich Documents by Token Path Prediction. In Proc. EMNLP, pages 13716-13730, 2023. 1 +[78] Jiaxin Zhang, Wentao Yang, Songxuan Lai, Zecheng Xie, and Lianwen Jin. DocKylin: A Large Multimodal Model for Visual Document Understanding with Efficient Visual Slimming. arXiv: Comp. Res. Repository, abs/2406.19101, 2024. 7 +[79] Yanzhe Zhang, Ruiyi Zhang, Jiuming Gu, Yufan Zhou, Nedim Lipka, Diyi Yang, and Tong Sun. LLaVAR: Enhanced Visual Instruction Tuning for Text-Rich Image Understanding. arXiv: Comp. Res. Repository, abs/2306.17107, 2023. 1, 2, 7 +[80] Xu Zhong, Jianbin Tang, and Antonio Jimeno-Yepes. Pub-LayNet: Largest Dataset Ever for Document Layout Analysis. In Proc. ICDAR, pages 1015-1022, 2019. 5 +[81] Xu Zhong, Elaheh ShafieiBavani, and Antonio JimenoYepes. Image-Based Table Recognition: Data, Model, and Evaluation. Proc. ECCV, 12366:564-580, 2020. 5 \ No newline at end of file diff --git a/2025/DocLayLLM_ An Efficient Multi-modal Extension of Large Language Models for Text-rich Document Understanding/images.zip b/2025/DocLayLLM_ An Efficient Multi-modal Extension of Large Language Models for Text-rich Document Understanding/images.zip new file mode 100644 index 0000000000000000000000000000000000000000..e652cb445252a446e2f111b9002734324a0e0192 --- /dev/null +++ b/2025/DocLayLLM_ An Efficient Multi-modal Extension of Large Language Models for Text-rich Document Understanding/images.zip @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:13a509702b899b7a08536660e2301c7db9f39f7f8ffb7250dbccc6faddbe76ae +size 560057 diff --git a/2025/DocLayLLM_ An Efficient Multi-modal Extension of Large Language Models for Text-rich Document Understanding/layout.json b/2025/DocLayLLM_ An Efficient Multi-modal Extension of Large Language Models for Text-rich Document Understanding/layout.json new file mode 100644 index 0000000000000000000000000000000000000000..00973d42b51ab7682204f5685c6b62a0bdd8ae12 --- /dev/null +++ b/2025/DocLayLLM_ An Efficient Multi-modal Extension of Large Language Models for Text-rich Document Understanding/layout.json @@ -0,0 +1,9292 @@ +{ + "pdf_info": [ + { + "para_blocks": [ + { + "bbox": [ + 66, + 102, + 545, + 140 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 66, + 102, + 545, + 140 + ], + "spans": [ + { + "bbox": [ + 66, + 102, + 545, + 140 + ], + "type": "text", + "content": "DocLayLLM: An Efficient Multi-modal Extension of Large Language Models for Text-rich Document Understanding" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 71, + 159, + 538, + 217 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 71, + 159, + 538, + 217 + ], + "spans": [ + { + "bbox": [ + 71, + 159, + 538, + 217 + ], + "type": "text", + "content": "Wenhui Liao" + }, + { + "bbox": [ + 71, + 159, + 538, + 217 + ], + "type": "inline_equation", + "content": "^{1\\dagger}" + }, + { + "bbox": [ + 71, + 159, + 538, + 217 + ], + "type": "text", + "content": ", Jiapeng Wang" + }, + { + "bbox": [ + 71, + 159, + 538, + 217 + ], + "type": "inline_equation", + "content": "^{1,2\\dagger}" + }, + { + "bbox": [ + 71, + 159, + 538, + 217 + ], + "type": "text", + "content": ", Hongliang Li" + }, + { + "bbox": [ + 71, + 159, + 538, + 217 + ], + "type": "inline_equation", + "content": "^{1}" + }, + { + "bbox": [ + 71, + 159, + 538, + 217 + ], + "type": "text", + "content": ", Chengyu Wang" + }, + { + "bbox": [ + 71, + 159, + 538, + 217 + ], + "type": "inline_equation", + "content": "^{2*}" + }, + { + "bbox": [ + 71, + 159, + 538, + 217 + ], + "type": "text", + "content": ", Jun Huang" + }, + { + "bbox": [ + 71, + 159, + 538, + 217 + ], + "type": "inline_equation", + "content": "^{2}" + }, + { + "bbox": [ + 71, + 159, + 538, + 217 + ], + "type": "text", + "content": ", Lianwen Jin" + }, + { + "bbox": [ + 71, + 159, + 538, + 217 + ], + "type": "inline_equation", + "content": "^{1,3*}" + }, + { + "bbox": [ + 71, + 159, + 538, + 217 + ], + "type": "inline_equation", + "content": "^{1}" + }, + { + "bbox": [ + 71, + 159, + 538, + 217 + ], + "type": "text", + "content": "South China University of Technology, Guangzhou, China \n" + }, + { + "bbox": [ + 71, + 159, + 538, + 217 + ], + "type": "inline_equation", + "content": "^{2}" + }, + { + "bbox": [ + 71, + 159, + 538, + 217 + ], + "type": "text", + "content": "Alibaba Cloud Computing, Hangzhou, China \n" + }, + { + "bbox": [ + 71, + 159, + 538, + 217 + ], + "type": "inline_equation", + "content": "^{3}" + }, + { + "bbox": [ + 71, + 159, + 538, + 217 + ], + "type": "text", + "content": "SCUT-Zhuhai Institute of Modern Industrial Innovation, Zhuhai, China" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 116, + 219, + 494, + 245 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 116, + 219, + 494, + 245 + ], + "spans": [ + { + "bbox": [ + 116, + 219, + 494, + 245 + ], + "type": "text", + "content": "{eelwh, eejpwang, eehongliangli}@mail.scut.edu.cn, eelwjin@scut.edu.cn {chengyu.wcy, huangjun.hj}@alibaba-inc.com" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 151, + 273, + 200, + 285 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 151, + 273, + 200, + 285 + ], + "spans": [ + { + "bbox": [ + 151, + 273, + 200, + 285 + ], + "type": "text", + "content": "Abstract" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 55, + 297, + 297, + 561 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 297, + 297, + 561 + ], + "spans": [ + { + "bbox": [ + 55, + 297, + 297, + 561 + ], + "type": "text", + "content": "Text-rich document understanding (TDU) requires comprehensive analysis of documents containing substantial textual content and complex layouts. While Multimodal Large Language Models (MLLMs) have achieved fast progress in this domain, existing approaches either demand significant computational resources or struggle with effective multi-modal integration. In this paper, we introduce DocLayLLM, an efficient multi-modal extension of LLMs specifically designed for TDU. By lightly integrating visual patch tokens and 2D positional tokens into LLMs' input and encoding the document content using the LLMs themselves, we fully take advantage of the document comprehension capability of LLMs and enhance their perception of OCR information. We have also deeply considered the role of chain-of-thought (CoT) and innovatively proposed the techniques of CoT Pre-training and CoT Annealing. Our DocLayLLM can achieve remarkable performances with lightweight training settings, showcasing its efficiency and effectiveness. Experimental results demonstrate that our DocLayLLM outperforms existing OCR-dependent methods and OCR-free competitors. Code and model are available at https://github.com/whlscut/DocLayLLM." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 55, + 582, + 135, + 594 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 582, + 135, + 594 + ], + "spans": [ + { + "bbox": [ + 55, + 582, + 135, + 594 + ], + "type": "text", + "content": "1. Introduction" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 55, + 603, + 296, + 687 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 603, + 296, + 687 + ], + "spans": [ + { + "bbox": [ + 55, + 603, + 296, + 687 + ], + "type": "text", + "content": "Text-rich document understanding (TDU) focuses on analyzing documents with extensive text and complex layouts, requiring a combined understanding of text, layout, and images. Early research [1, 2, 12, 15, 20, 28, 33-37, 44, 52, 55, 65, 68, 69, 72, 73, 77] often used BERT-like [8] Transformer-based document encoders (DE), designing various pre-training tasks for large-scale unlabeled" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 313, + 274, + 554, + 369 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 274, + 554, + 369 + ], + "spans": [ + { + "bbox": [ + 313, + 274, + 554, + 369 + ], + "type": "text", + "content": "document corpora and fine-tuning their models on specific downstream tasks like key information extraction (KIE) and visual question answering (VQA) [39, 40]. However, these pre-training tasks often diverge significantly from downstream tasks, requiring the model to acquire substantial task-specific knowledge during fine-tuning. This may result in limited generalization capability across different tasks and datasets." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 313, + 373, + 555, + 637 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 373, + 555, + 637 + ], + "spans": [ + { + "bbox": [ + 313, + 373, + 555, + 637 + ], + "type": "text", + "content": "With the rapid development of large language models (LLMs), their impressive versatility and generalization capabilities have captured widespread interest [74]. This has led many researchers [17, 38, 61, 76, 79] to overcome the challenges of TDU with LLMs. Initially, efforts focused on employing multi-modal large language models (MLLMs) for OCR-free TDU, where document images are fed into MLLMs, directly generating responses to comprehension tasks. These MLLMs, which can process images and text inputs, are among the most recent advancements in this domain. To date, several general MLLMs have been developed to manage diverse scenarios, and researchers often build their TDU models based on these general models. However, these TDU models typically require high-resolution image inputs to recognize dense text in documents and a large volume of training data to effectively develop this capability, both of which bring significant computational demands. Furthermore, within the extensive collections of image and instruction data used to train these models, the training data for downstream evaluation tasks was often included, which might potentially impact the assessment of their generalization to previously unseen datasets." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 313, + 641, + 556, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 641, + 556, + 715 + ], + "spans": [ + { + "bbox": [ + 313, + 641, + 556, + 715 + ], + "type": "text", + "content": "To overcome the constraints of OCR-free methods, scholars have pursued diverse methodologies for integrating OCR information into LLMs. Initial strategies [14, 70] involved inputting coordinates of OCR bounding boxes as numeric text. However, this textual representation of numbers often resulted in excessively long inputs that could" + } + ] + } + ], + "index": 11 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "spans": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "type": "text", + "content": "CVF" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 145, + 0, + 496, + 37 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 145, + 0, + 496, + 37 + ], + "spans": [ + { + "bbox": [ + 145, + 0, + 496, + 37 + ], + "type": "text", + "content": "This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 66, + 693, + 136, + 703 + ], + "type": "page_footnote", + "angle": 0, + "lines": [ + { + "bbox": [ + 66, + 693, + 136, + 703 + ], + "spans": [ + { + "bbox": [ + 66, + 693, + 136, + 703 + ], + "type": "inline_equation", + "content": "\\dagger" + }, + { + "bbox": [ + 66, + 693, + 136, + 703 + ], + "type": "text", + "content": " Equal contribution." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 67, + 703, + 157, + 712 + ], + "type": "page_footnote", + "angle": 0, + "lines": [ + { + "bbox": [ + 67, + 703, + 157, + 712 + ], + "spans": [ + { + "bbox": [ + 67, + 703, + 157, + 712 + ], + "type": "text", + "content": "\\*Co-corresponding authors." + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "text", + "content": "4038" + } + ] + } + ], + "index": 15 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 0 + }, + { + "para_blocks": [ + { + "bbox": [ + 55, + 72, + 294, + 240 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 72, + 294, + 240 + ], + "spans": [ + { + "bbox": [ + 55, + 72, + 294, + 240 + ], + "type": "text", + "content": "decrease inference speeds and potentially degrade model performance. In response, many researchers [11, 24, 45] have adopted an auxiliary document encoder like LayoutLMv3 [20] to process textual, visual, and layout features from documents and integrated into LLMs. Although these methods avoid high-resolution image inputs, the textual content of documents must also be encoded alongside the layout and image features within the DE due to its structural limitations. However, the document comprehension capabilities of these DEs may not be on par with those of LLMs themselves, which might have led to significant finetuning efforts for aligning the feature spaces of DEs and LLMs to improve performances during the whole training phase, demanding substantial computational resources." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 55, + 243, + 294, + 399 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 243, + 294, + 399 + ], + "spans": [ + { + "bbox": [ + 55, + 243, + 294, + 399 + ], + "type": "text", + "content": "To address the limitations of OCR-dependent TDU methods, we introduce DocLayLLM, an efficient multimodal extension of large language models for text-rich document understanding. Our method diverges from existing approaches that utilize an extra DE by integrating the document's image patch tokens and 2D positional tokens with textual content into a natural language expression, and then encoding them through a unified LLM. This innovative integration not only taps into the LLM's inherent text comprehension abilities but also allows it to perceive document layouts. Besides, by incorporating patch tokens, the model supports image-related pre-training tasks and further gains a rudimentary understanding of document structures." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 55, + 403, + 295, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 403, + 295, + 714 + ], + "spans": [ + { + "bbox": [ + 55, + 403, + 295, + 714 + ], + "type": "text", + "content": "To advance model performance, we have fully integrated the concept of the chain-of-thought (CoT) [3] into each stage of our training process. Firstly, we developed a novel CoT Pre-training strategy. CoT is engineered to preserve thematic coherence and logical progression during model reasoning, further enhancing the inference ability of LLM, which has proven effectiveness in various subsequent studies [14, 45, 71]. Recently, the use of Long CoT reasoning to enhance the inference capability of models has attracted significant attention [7, 49, 63], and DeepSeek-R1 [7] further emphasized the importance of utilizing CoT data for a cold start before the reinforcement learning stage. Our proposed CoT Pre-training offers a method for rapidly generating large-scale CoT data from non-QA data, thereby improving the model's reasoning ability. By employing CoT in the data generation process and subsequently pre-training the model with the generated QA-formatted data, we have aligned the added visual and layout features with LLM's inherent textual features while improving the model's comprehension ability. Furthermore, we introduced a CoT Annealing technique. While previous research [31, 42] has validated the effectiveness of annealing strategies that gradually increase the proportion of high-quality data towards the end of training, the impact of CoT on data quality had not been explored. To address this, we first propose an annealing strategy that considers data quality from a CoT per" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 313, + 73, + 508, + 84 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 73, + 508, + 84 + ], + "spans": [ + { + "bbox": [ + 313, + 73, + 508, + 84 + ], + "type": "text", + "content": "spective, further improving model performance." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 313, + 85, + 553, + 227 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 85, + 553, + 227 + ], + "spans": [ + { + "bbox": [ + 313, + 85, + 553, + 227 + ], + "type": "text", + "content": "Thanks to the above design, utilizing only around 3M pre-training data and 300K supervised fine-tuning (SFT) data with the efficient tuning technology of LoRA [18], our DocLayLLM finishes the whole training process in 36 A100 days, therefore resulting an efficient extension of LLM. Even with such low resources, our DocLayLLM surpasses existing OCR-dependent TDU methods. Furthermore, it also excels in specific tasks without fine-tuning on downstream tasks and outperforms state-of-the-art (SOTA) OCR-free methods when trained with training sets of downstream tasks as other methods, further showcasing its effectiveness. Our contributions are:" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 314, + 229, + 553, + 384 + ], + "type": "list", + "angle": 0, + "index": 9, + "blocks": [ + { + "bbox": [ + 314, + 229, + 553, + 265 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 314, + 229, + 553, + 265 + ], + "spans": [ + { + "bbox": [ + 314, + 229, + 553, + 265 + ], + "type": "text", + "content": "- We propose an efficient multi-modal extension of LLMs, which has augmented LLM in understanding text-rich documents and greatly reduced training resources in need." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 314, + 266, + 553, + 300 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 314, + 266, + 553, + 300 + ], + "spans": [ + { + "bbox": [ + 314, + 266, + 553, + 300 + ], + "type": "text", + "content": "- We introduce CoT Pre-training, a novel approach for generating CoT data that helps enhance the comprehension capabilities of models." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 314, + 301, + 553, + 348 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 314, + 301, + 553, + 348 + ], + "spans": [ + { + "bbox": [ + 314, + 301, + 553, + 348 + ], + "type": "text", + "content": "- We revisit data annealing from a novel perspective and introduce the CoT Annealing strategy, which further enhances the model's efficiency in data utilization and boosts its ultimate performance." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 314, + 349, + 553, + 384 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 314, + 349, + 553, + 384 + ], + "spans": [ + { + "bbox": [ + 314, + 349, + 553, + 384 + ], + "type": "text", + "content": "- Our method outperforms existing OCR-dependent TDU solutions and surpasses current OCR-free methods under comparable supervised fine-tuning conditions." + } + ] + } + ], + "index": 8 + } + ], + "sub_type": "text" + }, + { + "bbox": [ + 314, + 394, + 405, + 407 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 314, + 394, + 405, + 407 + ], + "spans": [ + { + "bbox": [ + 314, + 394, + 405, + 407 + ], + "type": "text", + "content": "2. Related Works" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 313, + 415, + 553, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 415, + 553, + 715 + ], + "spans": [ + { + "bbox": [ + 313, + 415, + 553, + 715 + ], + "type": "text", + "content": "OCR-free MLLMs for Text-Rich Documents Understanding. To date, a number of studies have attempted to directly answer the instruction for TDU tasks from document images, thereby obviating the need for extra OCR processing and resulting in an end-to-end paradigm. Such models are referred to as OCR-free methods. Early examples like Donut [23], UDOP [62], and OmniParser [66] have explored end-to-end document understanding solutions without relying on LLMs or MLLMs. However, with the rapid development of MLLMs, many new studies have begun to focus on enhancing the capabilities of general MLLMs in document understanding. Some of them attempt to boost the document understanding ability of general MLLMs by tuning them with synthesizing TDU data generated by proprietary LLMs, including LLaVAR [79] and TextSquare [61], while mPLUG-DocOwl 1.5 [17] enhances the document understanding capability by integrating existing document comprehension data and converting it into vision-related QA-formatted data. Moreover, researchers have also explored image encoding methods to read the text more clearly. For instance, UReader [76] and Monkey [38] empower MLLMs to handle high-resolution images by splitting them into sub-images, encoding each separately, and then concatenating them. Some general MLLMs, such as LLaVA-Next [29], LLaVA" + } + ] + } + ], + "index": 11 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "text", + "content": "4039" + } + ] + } + ], + "index": 12 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 1 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 87, + 68, + 523, + 261 + ], + "blocks": [ + { + "bbox": [ + 87, + 68, + 523, + 261 + ], + "lines": [ + { + "bbox": [ + 87, + 68, + 523, + 261 + ], + "spans": [ + { + "bbox": [ + 87, + 68, + 523, + 261 + ], + "type": "image", + "image_path": "67a0a691f89e141767f2dd8180cbecc54cfa227c8d51c9dff7b529145e47dcba.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 211, + 268, + 398, + 280 + ], + "lines": [ + { + "bbox": [ + 211, + 268, + 398, + 280 + ], + "spans": [ + { + "bbox": [ + 211, + 268, + 398, + 280 + ], + "type": "text", + "content": "Figure 1. The overall architecture of DocLayLLM." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "bbox": [ + 54, + 300, + 295, + 433 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 54, + 300, + 295, + 433 + ], + "spans": [ + { + "bbox": [ + 54, + 300, + 295, + 433 + ], + "type": "text", + "content": "OneVision [27], and Idefics3 [26], also adopt this approach to address the challenge of TDU. Although these above methods have achieved notable success in TDU tasks, they typically require high-resolution image inputs to identify text within images, significantly increasing computational demands. Besides, to equip general MLLMs with the ability to recognize the rich text within a document, a substantial amount of training data is required, including training sets of the benchmarks tested, which resulted in a high demand for training resources and might affect the evaluation of their zero-shot inference ability." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 55, + 438, + 295, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 438, + 295, + 715 + ], + "spans": [ + { + "bbox": [ + 55, + 438, + 295, + 715 + ], + "type": "text", + "content": "OCR-dependent MLLMs for Text-Rich Documents Understanding. To address the challenges of OCR-free MLLMs, some researchers have attempted to integrate OCR information from text-rich documents into LLMs to avoid high-resolution image inputs. ICL-D3IE [14] feeds coordinates of OCR boxes into the LLM using a textual expression of [left, top, right, bottom], while LATIN-Prompt [70] simulates the document layout by inserting spaces and newline characters between OCR-recognized text segments, thus providing the LLM with a document input containing a rudimentary layout. Besides, LMDX [53] also input the layout in a way similar to ICL-D3IE. However, these methods of inputting OCR information in textual form lead to overly lengthy inputs, thereby slowing inference and impacting model performance. Subsequent studies have explored encoding OCR information using additional encoders. Cream [24] utilizes BLIP-2 [30] to encode document OCR positions, images, text, and additional object detection information, with the output of BLIP-2 fed into the LLM. InstructDr [60] introduces an extra feature extractor to encode document OCR information, compressing it with learnable tokens before inputting it into the LLM. Additionally, Fujitake [11] and Luo et al. [45] adopt Lay" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 313, + 300, + 555, + 577 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 300, + 555, + 577 + ], + "spans": [ + { + "bbox": [ + 313, + 300, + 555, + 577 + ], + "type": "text", + "content": "outLMv3 [20] to encode the OCR information of documents before inputting them into the LLM. Although these methods avoid the issue of overly lengthy text inputs, they must use the extra DE to encode all the OCR textual, positional information from the document due to the structural restriction of DE. However, the additional DE might not be comparable with LLMs themselves in terms of document comprehension capability. This might be the reason why they usually need extensive training of both the DE and the LLM to align the features encoded by the DE with those of the LLM, which consumes significant computational resources. DocLLM [67] proposes an extension where the hidden states derived from OCR bounding boxes interact with those of OCR text features encoded by LLMs within each attention block via cross-attention [5, 6]. However, they use an extra projection layer at each attention block to transform and inject layout information, which might excessively interfere with the original attention computation of LLM, thereby potentially affecting LLM's generalization capability. Moreover, it does not adequately consider the role of image features in TDU. Overall, designing an efficient method to provide OCR information to LLMs remains a research question worthy of deep exploration." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 313, + 586, + 398, + 600 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 586, + 398, + 600 + ], + "spans": [ + { + "bbox": [ + 313, + 586, + 398, + 600 + ], + "type": "text", + "content": "3. DocLayLLM" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 313, + 605, + 555, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 605, + 555, + 715 + ], + "spans": [ + { + "bbox": [ + 313, + 605, + 555, + 715 + ], + "type": "text", + "content": "DocLayLLM is an LLM-based methodology that encodes OCRed textual, visual, and positional information directly within an LLM, eliminating the need for additional DE. The training regime of DocLayLLM includes both pretraining and supervised fine-tuning phases, which are both instruction-tuning-based processes. To further enhance our model, we propose the integration of CoT into our pretraining tasks, denoted as CoT Pre-training. Additionally, we introduce CoT Annealing, which involves the applica" + } + ] + } + ], + "index": 6 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "text", + "content": "4040" + } + ] + } + ], + "index": 7 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 2 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 77, + 70, + 219, + 196 + ], + "blocks": [ + { + "bbox": [ + 77, + 70, + 219, + 196 + ], + "lines": [ + { + "bbox": [ + 77, + 70, + 219, + 196 + ], + "spans": [ + { + "bbox": [ + 77, + 70, + 219, + 196 + ], + "type": "image", + "image_path": "f217931446d01cd5e3933c2f49b3147244b184376610f7c07309eb1661ae21af.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 105, + 200, + 189, + 209 + ], + "lines": [ + { + "bbox": [ + 105, + 200, + 189, + 209 + ], + "spans": [ + { + "bbox": [ + 105, + 200, + 189, + 209 + ], + "type": "text", + "content": "(a) Layout-Type-Related CoT." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "type": "image", + "bbox": [ + 226, + 70, + 376, + 196 + ], + "blocks": [ + { + "bbox": [ + 226, + 70, + 376, + 196 + ], + "lines": [ + { + "bbox": [ + 226, + 70, + 376, + 196 + ], + "spans": [ + { + "bbox": [ + 226, + 70, + 376, + 196 + ], + "type": "image", + "image_path": "1b53df17ea4a95f1df09dcad8302374bd5ded2f4093b1e50d8db8d0cfd5ebbd7.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 262, + 200, + 349, + 208 + ], + "lines": [ + { + "bbox": [ + 262, + 200, + 349, + 208 + ], + "spans": [ + { + "bbox": [ + 262, + 200, + 349, + 208 + ], + "type": "text", + "content": "(b) Table-Structure-Aware CoT." + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 70, + 217, + 537, + 229 + ], + "lines": [ + { + "bbox": [ + 70, + 217, + 537, + 229 + ], + "spans": [ + { + "bbox": [ + 70, + 217, + 537, + 229 + ], + "type": "text", + "content": "Figure 2. Examples of CoT Pre-training across different tasks. Red text highlights key steps in the chain of thought for each task." + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "image_caption" + } + ], + "index": 2 + }, + { + "type": "image", + "bbox": [ + 382, + 70, + 534, + 196 + ], + "blocks": [ + { + "bbox": [ + 382, + 70, + 534, + 196 + ], + "lines": [ + { + "bbox": [ + 382, + 70, + 534, + 196 + ], + "spans": [ + { + "bbox": [ + 382, + 70, + 534, + 196 + ], + "type": "image", + "image_path": "ad1828532a29f3a49f7b14a2e728d2f4cd3b718da1dcec76aa6fc670e403a950.jpg" + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 425, + 200, + 501, + 209 + ], + "lines": [ + { + "bbox": [ + 425, + 200, + 501, + 209 + ], + "spans": [ + { + "bbox": [ + 425, + 200, + 501, + 209 + ], + "type": "text", + "content": "(c) Geometry-Related CoT." + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "image_caption" + } + ], + "index": 4 + }, + { + "bbox": [ + 55, + 250, + 295, + 274 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 250, + 295, + 274 + ], + "spans": [ + { + "bbox": [ + 55, + 250, + 295, + 274 + ], + "type": "text", + "content": "tion of annealing to the data utilized in the supervised finetuning phase, specifically from the perspective of CoT." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 55, + 279, + 170, + 291 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 279, + 170, + 291 + ], + "spans": [ + { + "bbox": [ + 55, + 279, + 170, + 291 + ], + "type": "text", + "content": "3.1. Model Architecture" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 54, + 297, + 295, + 525 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 54, + 297, + 295, + 525 + ], + "spans": [ + { + "bbox": [ + 54, + 297, + 295, + 525 + ], + "type": "text", + "content": "The overall architecture of our DocLayLLM is illustrated in Figure 1. To process a document image, we first employ an OCR engine. This engine extracts the OCR results, which encompasses the textual content, denoted as " + }, + { + "bbox": [ + 54, + 297, + 295, + 525 + ], + "type": "inline_equation", + "content": "Text_{1:N}" + }, + { + "bbox": [ + 54, + 297, + 295, + 525 + ], + "type": "text", + "content": ", and the associated bounding box, represented by " + }, + { + "bbox": [ + 54, + 297, + 295, + 525 + ], + "type": "inline_equation", + "content": "Box_{1:N}" + }, + { + "bbox": [ + 54, + 297, + 295, + 525 + ], + "type": "text", + "content": ". Here, " + }, + { + "bbox": [ + 54, + 297, + 295, + 525 + ], + "type": "inline_equation", + "content": "N" + }, + { + "bbox": [ + 54, + 297, + 295, + 525 + ], + "type": "text", + "content": " indicates the total number of text segments detected by the OCR, and each " + }, + { + "bbox": [ + 54, + 297, + 295, + 525 + ], + "type": "inline_equation", + "content": "Box_{i} = [x_{0},y_{0},x_{1},y_{1}]" + }, + { + "bbox": [ + 54, + 297, + 295, + 525 + ], + "type": "text", + "content": " defines the bounding box for " + }, + { + "bbox": [ + 54, + 297, + 295, + 525 + ], + "type": "inline_equation", + "content": "Text_{i}" + }, + { + "bbox": [ + 54, + 297, + 295, + 525 + ], + "type": "text", + "content": ". The bounding box coordinates correspond to the left, top, right, and bottom boundaries of the text segment. After the OCR process, the document image is segmented into patches, denoted as " + }, + { + "bbox": [ + 54, + 297, + 295, + 525 + ], + "type": "inline_equation", + "content": "Patch_{1:M}" + }, + { + "bbox": [ + 54, + 297, + 295, + 525 + ], + "type": "text", + "content": ", where " + }, + { + "bbox": [ + 54, + 297, + 295, + 525 + ], + "type": "inline_equation", + "content": "M" + }, + { + "bbox": [ + 54, + 297, + 295, + 525 + ], + "type": "text", + "content": " represents the total number of patches. Next, the OCR results and the document patches are integrated with the fixed part of the prompt and the specific Question related to the document. This integration forms the final prompt input to the model, presented in a natural language expression " + }, + { + "bbox": [ + 54, + 297, + 295, + 525 + ], + "type": "inline_equation", + "content": "\\mathcal{T}" + }, + { + "bbox": [ + 54, + 297, + 295, + 525 + ], + "type": "text", + "content": ": Given the document patches: " + }, + { + "bbox": [ + 54, + 297, + 295, + 525 + ], + "type": "inline_equation", + "content": "\\{Patch_{1:M}\\}" + }, + { + "bbox": [ + 54, + 297, + 295, + 525 + ], + "type": "text", + "content": " and the document text contents and locations in the form of \"text, [left, top, right, bottom]\": " + }, + { + "bbox": [ + 54, + 297, + 295, + 525 + ], + "type": "inline_equation", + "content": "\\{(Text_{i}, Box_{i})_{1:N}\\}" + }, + { + "bbox": [ + 54, + 297, + 295, + 525 + ], + "type": "inline_equation", + "content": "\\{Question\\}" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 55, + 525, + 295, + 572 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 525, + 295, + 572 + ], + "spans": [ + { + "bbox": [ + 55, + 525, + 295, + 572 + ], + "type": "text", + "content": "Subsequently, the textual components of the prompt, including the fixed expression, " + }, + { + "bbox": [ + 55, + 525, + 295, + 572 + ], + "type": "inline_equation", + "content": "Text_{1:N}" + }, + { + "bbox": [ + 55, + 525, + 295, + 572 + ], + "type": "text", + "content": ", and " + }, + { + "bbox": [ + 55, + 525, + 295, + 572 + ], + "type": "inline_equation", + "content": "Question" + }, + { + "bbox": [ + 55, + 525, + 295, + 572 + ], + "type": "text", + "content": ", are processed using the LLM's native text embedder. This generates the embeddings for the text content as follows:" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 112, + 577, + 294, + 590 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 112, + 577, + 294, + 590 + ], + "spans": [ + { + "bbox": [ + 112, + 577, + 294, + 590 + ], + "type": "interline_equation", + "content": "E m b _ {T} = T E (T e x t C o n t e n t), \\tag {1}", + "image_path": "a73b2141e9d4684f82e82c0a296d737bfcd6ea25422dec53f08575415efe4edc.jpg" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 55, + 594, + 295, + 630 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 594, + 295, + 630 + ], + "spans": [ + { + "bbox": [ + 55, + 594, + 295, + 630 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 55, + 594, + 295, + 630 + ], + "type": "inline_equation", + "content": "TE" + }, + { + "bbox": [ + 55, + 594, + 295, + 630 + ], + "type": "text", + "content": " is the native text embedder of LLM, and TextContent consists of the fixed expression, " + }, + { + "bbox": [ + 55, + 594, + 295, + 630 + ], + "type": "inline_equation", + "content": "Text_{1:N}" + }, + { + "bbox": [ + 55, + 594, + 295, + 630 + ], + "type": "text", + "content": ", and Question." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 55, + 630, + 295, + 667 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 630, + 295, + 667 + ], + "spans": [ + { + "bbox": [ + 55, + 630, + 295, + 667 + ], + "type": "text", + "content": "As for " + }, + { + "bbox": [ + 55, + 630, + 295, + 667 + ], + "type": "inline_equation", + "content": "Patch_{1:M}" + }, + { + "bbox": [ + 55, + 630, + 295, + 667 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 55, + 630, + 295, + 667 + ], + "type": "inline_equation", + "content": "Box_{1:N}" + }, + { + "bbox": [ + 55, + 630, + 295, + 667 + ], + "type": "text", + "content": ", they are first embedded with an extra embedder, then mapped to the LLM's feature space via a linear projector, yielding " + }, + { + "bbox": [ + 55, + 630, + 295, + 667 + ], + "type": "inline_equation", + "content": "Emb_V" + }, + { + "bbox": [ + 55, + 630, + 295, + 667 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 55, + 630, + 295, + 667 + ], + "type": "inline_equation", + "content": "Emb_L" + }, + { + "bbox": [ + 55, + 630, + 295, + 667 + ], + "type": "text", + "content": " as:" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 107, + 682, + 295, + 695 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 107, + 682, + 295, + 695 + ], + "spans": [ + { + "bbox": [ + 107, + 682, + 295, + 695 + ], + "type": "interline_equation", + "content": "E m b _ {V} = V P (V E (P a t c h _ {1: M})), \\tag {2}", + "image_path": "9692c6312772ff1ca827caf1265499d618409d636bc79c1fd42af5a1e8b9ea68.jpg" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 109, + 698, + 294, + 710 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 109, + 698, + 294, + 710 + ], + "spans": [ + { + "bbox": [ + 109, + 698, + 294, + 710 + ], + "type": "interline_equation", + "content": "E m b _ {L} = L P \\left(L E \\left(\\text {B o x} _ {1: N}\\right)\\right), \\tag {3}", + "image_path": "e6de28c48d829b0e599437faa9daa0f0b1558252bd974395b3179b681de3aa20.jpg" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 313, + 250, + 553, + 321 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 250, + 553, + 321 + ], + "spans": [ + { + "bbox": [ + 313, + 250, + 553, + 321 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 313, + 250, + 553, + 321 + ], + "type": "inline_equation", + "content": "VE" + }, + { + "bbox": [ + 313, + 250, + 553, + 321 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 313, + 250, + 553, + 321 + ], + "type": "inline_equation", + "content": "VP" + }, + { + "bbox": [ + 313, + 250, + 553, + 321 + ], + "type": "text", + "content": " represent the visual embedder and visual projector, " + }, + { + "bbox": [ + 313, + 250, + 553, + 321 + ], + "type": "inline_equation", + "content": "LE" + }, + { + "bbox": [ + 313, + 250, + 553, + 321 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 313, + 250, + 553, + 321 + ], + "type": "inline_equation", + "content": "LP" + }, + { + "bbox": [ + 313, + 250, + 553, + 321 + ], + "type": "text", + "content": " denote the layout embedder and layout projector, respectively. It is worth noting that despite the introduction of " + }, + { + "bbox": [ + 313, + 250, + 553, + 321 + ], + "type": "inline_equation", + "content": "VE" + }, + { + "bbox": [ + 313, + 250, + 553, + 321 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 313, + 250, + 553, + 321 + ], + "type": "inline_equation", + "content": "VP" + }, + { + "bbox": [ + 313, + 250, + 553, + 321 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 313, + 250, + 553, + 321 + ], + "type": "inline_equation", + "content": "LE" + }, + { + "bbox": [ + 313, + 250, + 553, + 321 + ], + "type": "text", + "content": ", and " + }, + { + "bbox": [ + 313, + 250, + 553, + 321 + ], + "type": "inline_equation", + "content": "LP" + }, + { + "bbox": [ + 313, + 250, + 553, + 321 + ], + "type": "text", + "content": ", these are all simple embedding layers or linear layers, thereby minimizing additional computational burden." + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 313, + 321, + 553, + 346 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 321, + 553, + 346 + ], + "spans": [ + { + "bbox": [ + 313, + 321, + 553, + 346 + ], + "type": "text", + "content": "Finally, LLM will comprehend the embeddings provided and utilize them to generate the Answer to the Question:" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 323, + 352, + 553, + 365 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 323, + 352, + 553, + 365 + ], + "spans": [ + { + "bbox": [ + 323, + 352, + 553, + 365 + ], + "type": "interline_equation", + "content": "\\text {A n s w e r} = L L M \\left(\\pi \\left(\\left(E m b _ {T}, E m b _ {V}, E m b _ {L}\\right), \\mathcal {T}\\right)\\right). \\tag {4}", + "image_path": "d8f7909d12a141dba2d2e5c32f47053aa65f60da8388fd1336869e1d36e95f9f.jpg" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 313, + 372, + 553, + 396 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 372, + 553, + 396 + ], + "spans": [ + { + "bbox": [ + 313, + 372, + 553, + 396 + ], + "type": "text", + "content": "Here " + }, + { + "bbox": [ + 313, + 372, + 553, + 396 + ], + "type": "inline_equation", + "content": "\\pi" + }, + { + "bbox": [ + 313, + 372, + 553, + 396 + ], + "type": "text", + "content": " represents organizing " + }, + { + "bbox": [ + 313, + 372, + 553, + 396 + ], + "type": "inline_equation", + "content": "Emb_{T}, Emb_{V}, Emb_{L}" + }, + { + "bbox": [ + 313, + 372, + 553, + 396 + ], + "type": "text", + "content": " into a feature sequence according to " + }, + { + "bbox": [ + 313, + 372, + 553, + 396 + ], + "type": "inline_equation", + "content": "\\mathcal{T}" + }, + { + "bbox": [ + 313, + 372, + 553, + 396 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 313, + 396, + 554, + 468 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 396, + 554, + 468 + ], + "spans": [ + { + "bbox": [ + 313, + 396, + 554, + 468 + ], + "type": "text", + "content": "With our design, we have preserved the original structure of the LLM, ensuring its generalization capability. At the same time, we have enabled the LLM to easily accept and process visual and layout features, making it better suited for understanding text-rich documents with various layouts. This results in a simple yet effective multi-modal extension." + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 313, + 473, + 419, + 486 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 473, + 419, + 486 + ], + "spans": [ + { + "bbox": [ + 313, + 473, + 419, + 486 + ], + "type": "text", + "content": "3.2. CoT Pre-training" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 313, + 491, + 555, + 671 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 491, + 555, + 671 + ], + "spans": [ + { + "bbox": [ + 313, + 491, + 555, + 671 + ], + "type": "text", + "content": "During the pre-training phase, we incorporate a wide range of tasks (as detailed in the appendix) to boost the LLM's understanding of the visual and layout features inspired by Luo et al. [45]. Since the datasets related to these tasks are not designed explicitly for instruction-tuning, their annotations are not in a QA format. Therefore, we manually create QA-formatted instructions based on their annotations for training. However, simply converting task descriptions into Question and annotations into Answer is highly inefficient regarding data utilization. To address this, we incorporate CoT during the conversion to instruction-tuning data for specific tasks, asking LLM to solve our problem step-by-step thus enhancing the model's comprehension abilities, as the examples illustrated in Fig 2. Our designed CoT primarily falls into the following three categories:" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 313, + 674, + 454, + 687 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 674, + 454, + 687 + ], + "spans": [ + { + "bbox": [ + 313, + 674, + 454, + 687 + ], + "type": "text", + "content": "3.2.1. Layout-Type-Related CoT" + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 313, + 689, + 554, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 689, + 554, + 714 + ], + "spans": [ + { + "bbox": [ + 313, + 689, + 554, + 714 + ], + "type": "text", + "content": "We conduct a layout-type-related CoT for the layout analysis task. To determine the layout type of a given area, it is" + } + ] + } + ], + "index": 24 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 748, + 314, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 314, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 314, + 757 + ], + "type": "text", + "content": "4041" + } + ] + } + ], + "index": 25 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 3 + }, + { + "para_blocks": [ + { + "bbox": [ + 55, + 72, + 294, + 120 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 72, + 294, + 120 + ], + "spans": [ + { + "bbox": [ + 55, + 72, + 294, + 120 + ], + "type": "text", + "content": "beneficial first to locate its region within the document and then infer based on the nearby layout element. When giving an area bounded by a bounding box, our layout-type-related CoT can be formulated as following steps:" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 55, + 121, + 295, + 202 + ], + "type": "list", + "angle": 0, + "index": 4, + "blocks": [ + { + "bbox": [ + 56, + 121, + 295, + 156 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 121, + 295, + 156 + ], + "spans": [ + { + "bbox": [ + 56, + 121, + 295, + 156 + ], + "type": "text", + "content": "1. Identify the text within the area and determine its approximate region within the document (e.g., top left, top right part of the document, etc)." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 55, + 156, + 294, + 180 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 156, + 294, + 180 + ], + "spans": [ + { + "bbox": [ + 55, + 156, + 294, + 180 + ], + "type": "text", + "content": "2. Locate the nearest bounding boxes and assess their layout type." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 55, + 180, + 294, + 202 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 180, + 294, + 202 + ], + "spans": [ + { + "bbox": [ + 55, + 180, + 294, + 202 + ], + "type": "text", + "content": "3. Infer the layout type of the given area based on the above information." + } + ] + } + ], + "index": 3 + } + ], + "sub_type": "text" + }, + { + "bbox": [ + 55, + 204, + 295, + 251 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 204, + 295, + 251 + ], + "spans": [ + { + "bbox": [ + 55, + 204, + 295, + 251 + ], + "type": "text", + "content": "Through this CoT process, we enable the LLM to fully utilize the visual, layout, and textual features of the document, thereby improving their ability to understand the document's overall structure." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 55, + 258, + 203, + 269 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 258, + 203, + 269 + ], + "spans": [ + { + "bbox": [ + 55, + 258, + 203, + 269 + ], + "type": "text", + "content": "3.2.2. Table-Structure-Aware CoT" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 55, + 273, + 295, + 332 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 273, + 295, + 332 + ], + "spans": [ + { + "bbox": [ + 55, + 273, + 295, + 332 + ], + "type": "text", + "content": "Table-structure-aware CoT is primarily employed in the table analysis task. We have integrated the XY-Cut algorithm [48]—a widely-used technique for analyzing table structures. For instance, to locate an element at the " + }, + { + "bbox": [ + 55, + 273, + 295, + 332 + ], + "type": "inline_equation", + "content": "i" + }, + { + "bbox": [ + 55, + 273, + 295, + 332 + ], + "type": "text", + "content": "-th row and " + }, + { + "bbox": [ + 55, + 273, + 295, + 332 + ], + "type": "inline_equation", + "content": "j" + }, + { + "bbox": [ + 55, + 273, + 295, + 332 + ], + "type": "text", + "content": "-th column, the CoT process is outlined as follows:" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 56, + 333, + 295, + 429 + ], + "type": "list", + "angle": 0, + "index": 11, + "blocks": [ + { + "bbox": [ + 56, + 333, + 295, + 368 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 333, + 295, + 368 + ], + "spans": [ + { + "bbox": [ + 56, + 333, + 295, + 368 + ], + "type": "text", + "content": "1. Identify all the table headers and their respective locations. The header of the " + }, + { + "bbox": [ + 56, + 333, + 295, + 368 + ], + "type": "inline_equation", + "content": "j" + }, + { + "bbox": [ + 56, + 333, + 295, + 368 + ], + "type": "text", + "content": "-th column corresponds to the " + }, + { + "bbox": [ + 56, + 333, + 295, + 368 + ], + "type": "inline_equation", + "content": "j" + }, + { + "bbox": [ + 56, + 333, + 295, + 368 + ], + "type": "text", + "content": "-th element in the headers list." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 56, + 369, + 295, + 405 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 369, + 295, + 405 + ], + "spans": [ + { + "bbox": [ + 56, + 369, + 295, + 405 + ], + "type": "text", + "content": "2. Output all elements in the " + }, + { + "bbox": [ + 56, + 369, + 295, + 405 + ], + "type": "inline_equation", + "content": "j" + }, + { + "bbox": [ + 56, + 369, + 295, + 405 + ], + "type": "text", + "content": "-th column when told that the elements of the same column are roughly vertically aligned with their corresponding headers." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 56, + 405, + 295, + 429 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 405, + 295, + 429 + ], + "spans": [ + { + "bbox": [ + 56, + 405, + 295, + 429 + ], + "type": "text", + "content": "3. Determine the final answer based on the " + }, + { + "bbox": [ + 56, + 405, + 295, + 429 + ], + "type": "inline_equation", + "content": "i" + }, + { + "bbox": [ + 56, + 405, + 295, + 429 + ], + "type": "text", + "content": "-th element of the contents in the " + }, + { + "bbox": [ + 56, + 405, + 295, + 429 + ], + "type": "inline_equation", + "content": "j" + }, + { + "bbox": [ + 56, + 405, + 295, + 429 + ], + "type": "text", + "content": "-th column." + } + ] + } + ], + "index": 10 + } + ], + "sub_type": "text" + }, + { + "bbox": [ + 55, + 429, + 295, + 489 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 429, + 295, + 489 + ], + "spans": [ + { + "bbox": [ + 55, + 429, + 295, + 489 + ], + "type": "text", + "content": "Applying the above CoT, we break down a complex question into simple steps using the basic XY-Cut thought. This approach improves the LLM's inference ability through this breakdown and enhances data utilization by effectively leveraging the overall structure of the table." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 55, + 495, + 185, + 506 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 495, + 185, + 506 + ], + "spans": [ + { + "bbox": [ + 55, + 495, + 185, + 506 + ], + "type": "text", + "content": "3.2.3. Geometry-Related CoT" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 55, + 510, + 295, + 557 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 510, + 295, + 557 + ], + "spans": [ + { + "bbox": [ + 55, + 510, + 295, + 557 + ], + "type": "text", + "content": "This is used in geometric analysis to identify geometric relationships between two text segments. It assists LLM to inference with geometric relations. Specifically, when giving the text content of two lines:" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 55, + 558, + 295, + 688 + ], + "type": "list", + "angle": 0, + "index": 19, + "blocks": [ + { + "bbox": [ + 56, + 558, + 285, + 569 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 558, + 285, + 569 + ], + "spans": [ + { + "bbox": [ + 56, + 558, + 285, + 569 + ], + "type": "text", + "content": "1. Recover the bounding box of the given text segments." + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 56, + 570, + 295, + 594 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 570, + 295, + 594 + ], + "spans": [ + { + "bbox": [ + 56, + 570, + 295, + 594 + ], + "type": "text", + "content": "2. Analyze their vertical and horizontal projections to determine if they overlap through interval analysis." + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 56, + 594, + 295, + 616 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 594, + 295, + 616 + ], + "spans": [ + { + "bbox": [ + 56, + 594, + 295, + 616 + ], + "type": "text", + "content": "3. Calculate the center point of each segment and determine their relative location by comparing their center points." + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 55, + 618, + 295, + 688 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 618, + 295, + 688 + ], + "spans": [ + { + "bbox": [ + 55, + 618, + 295, + 688 + ], + "type": "text", + "content": "4. Calculate the minimum Euclidean distance based on the previous analysis. If they overlap, the distance is 0; if they overlap in vertical/horizontal projections, the minimum distance is the smallest interval in the horizontal/vertical projections; otherwise, it is the minimum Euclidean distance between their nearest corners." + } + ] + } + ], + "index": 18 + } + ], + "sub_type": "text" + }, + { + "bbox": [ + 55, + 689, + 295, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 689, + 295, + 714 + ], + "spans": [ + { + "bbox": [ + 55, + 689, + 295, + 714 + ], + "type": "text", + "content": "This design advances the LLM's arithmetic reasoning capabilities and helps it more accurately grasp the spatial re" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 313, + 72, + 553, + 96 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 72, + 553, + 96 + ], + "spans": [ + { + "bbox": [ + 313, + 72, + 553, + 96 + ], + "type": "text", + "content": "relationships between different boxes, and also contributes to layout-type-aware CoT to find the nearest bounding boxes." + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 313, + 108, + 409, + 121 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 108, + 409, + 121 + ], + "spans": [ + { + "bbox": [ + 313, + 108, + 409, + 121 + ], + "type": "text", + "content": "3.3. CoT Annealing" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 313, + 127, + 553, + 199 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 127, + 553, + 199 + ], + "spans": [ + { + "bbox": [ + 313, + 127, + 553, + 199 + ], + "type": "text", + "content": "Data annealing involves gradually increasing the proportion of high-quality data towards the end of supervised fine-tuning to improve data utilization efficiency and augment the model's reasoning abilities. This technique has been proven to be effective and is widely adopted in models like MiniCPM [19], DataComp-LM [31], and Llama-3.1 [42]." + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 313, + 201, + 555, + 332 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 201, + 555, + 332 + ], + "spans": [ + { + "bbox": [ + 313, + 201, + 555, + 332 + ], + "type": "text", + "content": "During the SFT stage, we employ data with a layout-aware CoT [45] that helps focus on potential regions of questions and answers within documents, thereby better leveraging the layout features learned during pre-training. It also allows the model to strengthen its reasoning capacity through step-by-step thinking. However, in the latter stages of SFT, our goal for LLM is to output precise and direct answers to questions. At this point, using CoT data may introduce unnecessary noise with a long response, and the overly uniform layout-aware CoT data might partially compromise the LLM's generalization abilities." + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 313, + 334, + 554, + 465 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 334, + 554, + 465 + ], + "spans": [ + { + "bbox": [ + 313, + 334, + 554, + 465 + ], + "type": "text", + "content": "To balance the strengths and weaknesses of such data, we propose CoT annealing, which reconsiders this data annealing issue from a CoT perspective. Specifically, we generate direct-answer data without CoT from the CoT data, treating it as high-quality data. During the SFT process, we gradually increase the proportion of this direct-answer data, thereby enhancing the LLM's ability to respond with a clear and direct answer in the end. Initially, we only utilize the data with CoT, gradually adjusting the ratio of CoT to non-CoT data as the SFT process progresses. Ultimately, we conclude the training using only the data without CoT." + } + ] + } + ], + "index": 25 + }, + { + "bbox": [ + 313, + 480, + 394, + 494 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 480, + 394, + 494 + ], + "spans": [ + { + "bbox": [ + 313, + 480, + 394, + 494 + ], + "type": "text", + "content": "4. Experiments" + } + ] + } + ], + "index": 26 + }, + { + "bbox": [ + 313, + 502, + 376, + 513 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 502, + 376, + 513 + ], + "spans": [ + { + "bbox": [ + 313, + 502, + 376, + 513 + ], + "type": "text", + "content": "4.1. Datasets" + } + ] + } + ], + "index": 27 + }, + { + "bbox": [ + 313, + 521, + 554, + 640 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 521, + 554, + 640 + ], + "spans": [ + { + "bbox": [ + 313, + 521, + 554, + 640 + ], + "type": "text", + "content": "During the pre-training stage, we conduct our pre-training tasks on several datasets, including RVL-CDIP [13], PubLayNet [80], PubTabNet [81], DocLayNet [54], DocBank [32], DocILE [56], and the Document Dense Description data from LayoutLLM [45]. This results in a total of 3.1 million pairs of instruction tuning data, which is significantly less than the volume used in previous OCR-dependent TDU methods. For the supervised fine-tuning stage, we apply CoT annealing on the layout-aware document QA data utilized in LayoutLLM [45]." + } + ] + } + ], + "index": 28 + }, + { + "bbox": [ + 313, + 642, + 554, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 642, + 554, + 713 + ], + "spans": [ + { + "bbox": [ + 313, + 642, + 554, + 713 + ], + "type": "text", + "content": "We evaluate our DocLayLLM on widely used text-rich document benchmarks, including FUNSD [22], CORD [50], SROIE [21], POIE [25], DocVQA [46], InfoVQA [47], VisualMRC [59], DeepForm [58], KLC [57], WTQ [51], and TabFact [4]. For a fair comparison with other approaches, we use the default OCR data from the of" + } + ] + } + ], + "index": 29 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "text", + "content": "4042" + } + ] + } + ], + "index": 30 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 4 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 105, + 70, + 506, + 208 + ], + "blocks": [ + { + "bbox": [ + 105, + 70, + 506, + 208 + ], + "lines": [ + { + "bbox": [ + 105, + 70, + 506, + 208 + ], + "spans": [ + { + "bbox": [ + 105, + 70, + 506, + 208 + ], + "type": "table", + "html": "
MethodProcessed Data at Pre-trainingDocument-oriented VQAKIE
DocVQAVisualMRCAvg.FUNSD†CORD†SROIE†Avg.
Text
Llama2-7B-Chat [64]-20.509.9015.2015.1020.0035.623.57
Llama3-8B-Instruct [42]-51.7947.7749.7868.5652.3161.2460.70
Text + Box
Llama2-7B-Chatcoor [14]-12.3012.2012.2511.906.4039.4019.23
Llama3-8B-Instructcoor [14]-49.1341.7545.4474.0062.2063.1566.45
LayTextLLMzero [43]10.0M (323%)65.5037.4051.4572.0045.5082.0066.50
Text + Box + Patch
LayoutLLM Llama2 [45]5.7M (184%)74.2555.7364.9978.6562.2170.9770.61
DocLayLLM Llama2 (Ours)3.1M (100%)72.8355.0463.9478.6770.8183.2777.58
DocLayLLM Llama3 (Ours)3.1M (100%)78.3658.5568.4684.1171.3484.3679.94
", + "image_path": "edb5bc45fbd940757148ee9d9a4c06caefa45240d795de091c14dc35544bcab0.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "table_body" + } + ], + "index": 0 + }, + { + "type": "table", + "bbox": [ + 105, + 281, + 506, + 430 + ], + "blocks": [ + { + "bbox": [ + 55, + 215, + 555, + 270 + ], + "lines": [ + { + "bbox": [ + 55, + 215, + 555, + 270 + ], + "spans": [ + { + "bbox": [ + 55, + 215, + 555, + 270 + ], + "type": "text", + "content": "Table 1. Comparison with OCR-dependent methods using metrics from LayoutLLM [45]. Processed Data at Pre-training represents the total data volume processed at the pre-training stage, equivalent to dataset size multiplied by epoch count, with values in parentheses representing size ratios compared to DocLayLLM. Avg. represents the average performance across different benchmarks, and " + }, + { + "bbox": [ + 55, + 215, + 555, + 270 + ], + "type": "inline_equation", + "content": "\\dagger" + }, + { + "bbox": [ + 55, + 215, + 555, + 270 + ], + "type": "text", + "content": " indicates the cleaned test set used in LayoutLLM. The suffix layout denotes using method from He et al. [14], which provides LLM with layout features via OCR box coordinates in textual form. Bold indicates the best performance, and underline indicates the second-best one." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 105, + 281, + 506, + 430 + ], + "lines": [ + { + "bbox": [ + 105, + 281, + 506, + 430 + ], + "spans": [ + { + "bbox": [ + 105, + 281, + 506, + 430 + ], + "type": "table", + "html": "
MethodProcessed Data at Pre-trainingDocument-oriented VQAKIE
DocVQAVisualMRCAvg.FUNSDCORDSROIEAvg.
Text
Llama2-7B-Chat [64]-20.506.3013.4023.4051.8058.6044.60
Llama3-8B-Instruct [42]-51.79229.74140.7766.6774.7182.5174.63
Text + Box
Llama2-7B-Chatcoor [14]-12.3028.0020.1514.4038.1050.6034.37
Llama3-8B-Instructcoor [14]-49.13211.69130.4174.7175.3375.9375.32
LMDX [53]-----66.95--
DocLLM [67]16.8M (542%)69.50*264.10*166.80*51.80*67.60*91.90*70.43*
LayTextLLMzero [43]10.0M (323%)65.50200.20132.8547.2077.2083.8069.40
Text + Box + Patch
DocLayLLM Llama2 (Ours)3.1M (100%)72.83310.60191.7280.7479.3784.3781.49
DocLayLLM Llama3 (Ours)3.1M (100%)78.35357.04217.7086.4783.6686.0885.40
", + "image_path": "2152fd2ca1786310a14936ab71f7bf774f0553e576a62597b0a36159ed34a79d.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "table_body" + } + ], + "index": 2 + }, + { + "bbox": [ + 55, + 437, + 555, + 460 + ], + "lines": [ + { + "bbox": [ + 55, + 437, + 555, + 460 + ], + "spans": [ + { + "bbox": [ + 55, + 437, + 555, + 460 + ], + "type": "text", + "content": "Table 2. Comparison with OCR-dependent methods using metrics from DocLLM [67]. * indicates that the model uses the training set of the benchmark during training." + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "text" + }, + { + "bbox": [ + 55, + 480, + 295, + 506 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 480, + 295, + 506 + ], + "spans": [ + { + "bbox": [ + 55, + 480, + 295, + 506 + ], + "type": "text", + "content": "ficial datasets otherwise commercial OCR engines " + }, + { + "bbox": [ + 55, + 480, + 295, + 506 + ], + "type": "inline_equation", + "content": "{}^{1}" + }, + { + "bbox": [ + 55, + 480, + 295, + 506 + ], + "type": "text", + "content": " if not provided, and follow their evaluation pipeline and metrics." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 55, + 511, + 165, + 523 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 511, + 165, + 523 + ], + "spans": [ + { + "bbox": [ + 55, + 511, + 165, + 523 + ], + "type": "text", + "content": "4.2. Implement Details" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 55, + 529, + 295, + 696 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 529, + 295, + 696 + ], + "spans": [ + { + "bbox": [ + 55, + 529, + 295, + 696 + ], + "type": "text", + "content": "Our method is based on Llama3-8B-Instruct [42], with a Llama2-7B-Chat [64] version for better comparison with other methods. They are denoted as DocLayLLM" + }, + { + "bbox": [ + 55, + 529, + 295, + 696 + ], + "type": "inline_equation", + "content": "_{Llama3}" + }, + { + "bbox": [ + 55, + 529, + 295, + 696 + ], + "type": "text", + "content": " and DocLayLLM" + }, + { + "bbox": [ + 55, + 529, + 295, + 696 + ], + "type": "inline_equation", + "content": "_{Llama2}" + }, + { + "bbox": [ + 55, + 529, + 295, + 696 + ], + "type": "text", + "content": " respectively. We initialize " + }, + { + "bbox": [ + 55, + 529, + 295, + 696 + ], + "type": "inline_equation", + "content": "VE" + }, + { + "bbox": [ + 55, + 529, + 295, + 696 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 55, + 529, + 295, + 696 + ], + "type": "inline_equation", + "content": "LE" + }, + { + "bbox": [ + 55, + 529, + 295, + 696 + ], + "type": "text", + "content": " with weights from LayoutLMv3 [20] to leverage document knowledge acquired during its pre-training, while " + }, + { + "bbox": [ + 55, + 529, + 295, + 696 + ], + "type": "inline_equation", + "content": "VP" + }, + { + "bbox": [ + 55, + 529, + 295, + 696 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 55, + 529, + 295, + 696 + ], + "type": "inline_equation", + "content": "LP" + }, + { + "bbox": [ + 55, + 529, + 295, + 696 + ], + "type": "text", + "content": " are randomly initialized. Besides, the image resolution is fixed at " + }, + { + "bbox": [ + 55, + 529, + 295, + 696 + ], + "type": "inline_equation", + "content": "224*224" + }, + { + "bbox": [ + 55, + 529, + 295, + 696 + ], + "type": "text", + "content": ", resulting in 196 visual patch tokens, to capture an overview of document layout while bringing less computation burden. During pre-training and fine-tuning, we use LoRA [18] with a rank of 64 on LLM and keep " + }, + { + "bbox": [ + 55, + 529, + 295, + 696 + ], + "type": "inline_equation", + "content": "VE" + }, + { + "bbox": [ + 55, + 529, + 295, + 696 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 55, + 529, + 295, + 696 + ], + "type": "inline_equation", + "content": "VP" + }, + { + "bbox": [ + 55, + 529, + 295, + 696 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 55, + 529, + 295, + 696 + ], + "type": "inline_equation", + "content": "LE" + }, + { + "bbox": [ + 55, + 529, + 295, + 696 + ], + "type": "text", + "content": ", and " + }, + { + "bbox": [ + 55, + 529, + 295, + 696 + ], + "type": "inline_equation", + "content": "LP" + }, + { + "bbox": [ + 55, + 529, + 295, + 696 + ], + "type": "text", + "content": " trainable, leading to few tuning parameters. We train our model on 8 NVIDIA Tesla A100-80G GPUs, taking about 30 A100 days for pre" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 313, + 480, + 553, + 506 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 480, + 553, + 506 + ], + "spans": [ + { + "bbox": [ + 313, + 480, + 553, + 506 + ], + "type": "text", + "content": "training and 6 A100 days for supervised fine-tuning. Additional details can be found in the appendix." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 313, + 514, + 555, + 538 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 514, + 555, + 538 + ], + "spans": [ + { + "bbox": [ + 313, + 514, + 555, + 538 + ], + "type": "text", + "content": "4.3. Comparisons with SOTA OCR-dependent methods" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 313, + 544, + 554, + 592 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 544, + 554, + 592 + ], + "spans": [ + { + "bbox": [ + 313, + 544, + 554, + 592 + ], + "type": "text", + "content": "We have compared our DocLayLLM with existing SOTA OCR-dependent methods. In order to exhibit the zero-shot inferring ability precisely, we chose their performances under the zero-shot scenario unless they are not provided." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 313, + 593, + 554, + 653 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 593, + 554, + 653 + ], + "spans": [ + { + "bbox": [ + 313, + 593, + 554, + 653 + ], + "type": "text", + "content": "The experimental results are shown in Table 1 and Table 2. Our method outperforms the baseline models, regardless of their type or whether the document's layout information is provided to them. It validates the rationale behind our design of multi-modal extension." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 313, + 654, + 554, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 654, + 554, + 714 + ], + "spans": [ + { + "bbox": [ + 313, + 654, + 554, + 714 + ], + "type": "text", + "content": "Additionally, our method significantly outperforms previous SOTA methods despite fewer training resources. As shown in Table 1, it surpasses the performance of Layout-LLM [45] and LayTextLLM [43] though requiring fewer computational resources by using less data and tuning few" + } + ] + } + ], + "index": 11 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 67, + 702, + 150, + 712 + ], + "type": "page_footnote", + "angle": 0, + "lines": [ + { + "bbox": [ + 67, + 702, + 150, + 712 + ], + "spans": [ + { + "bbox": [ + 67, + 702, + 150, + 712 + ], + "type": "text", + "content": "1 https://www.textin.com/" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "text", + "content": "4043" + } + ] + } + ], + "index": 13 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 5 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 120, + 71, + 492, + 192 + ], + "blocks": [ + { + "bbox": [ + 120, + 71, + 492, + 192 + ], + "lines": [ + { + "bbox": [ + 120, + 71, + 492, + 192 + ], + "spans": [ + { + "bbox": [ + 120, + 71, + 492, + 192 + ], + "type": "table", + "html": "
MethodDocument-oriented VQAKIE
DocVQAInfoVQAAvg.FUNSDSROIEPOIEAvg.
LLaVAR [79]12.3016.5014.400.505.205.903.87
UniDoc [10]7.7014.7011.201.002.905.103.00
DocPedia [9]47.10*15.20*31.15*29.9021.4039.9030.40
Monkey [38]50.10*25.80*37.95*24.1041.9019.9028.63
TextMonkey [41]66.70*28.60*47.65*42.9046.2032.0040.37
DocOwl 1.5 [17]----48.3051.80-
TextSquare [61]----53.2071.80-
DocKylin [78]65.10*34.80*49.95*58.7025.5049.5044.57
DocLayLLM\\( _{Llama3} \\) (Ours)77.7942.0259.9180.2876.5975.1377.33
", + "image_path": "43de467f9e54bd0362645cffd147418ed24630c12e5a78d965519ca83a673d96.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "table_body" + } + ], + "index": 0 + }, + { + "type": "table", + "bbox": [ + 85, + 234, + 525, + 361 + ], + "blocks": [ + { + "bbox": [ + 55, + 201, + 555, + 222 + ], + "lines": [ + { + "bbox": [ + 55, + 201, + 555, + 222 + ], + "spans": [ + { + "bbox": [ + 55, + 201, + 555, + 222 + ], + "type": "text", + "content": "Table 3. Comparison with OCR-free methods using the precision metric from TextMonkey [41]. * indicates that the model uses the training set of the benchmark during training." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 85, + 234, + 525, + 361 + ], + "lines": [ + { + "bbox": [ + 85, + 234, + 525, + 361 + ], + "spans": [ + { + "bbox": [ + 85, + 234, + 525, + 361 + ], + "type": "table", + "html": "
MethodDocument-oriented VQAKIETable Understanding
DocVQAInfoVQAVisualMRCAvg.DeepFormKLCAvg.WTQTabFactAvg.
DocPedia [9]47.1015.20--------
DocOwl [75]62.2038.20188.8096.4042.6030.3036.4526.9060.2043.55
URector [76]65.4042.20221.70109.7749.5032.8041.1529.4067.6048.50
Monkey [38]66.5036.10--40.6032.8036.7025.30--
TextMonkey [41]73.00---59.7037.8048.7531.90--
DocKylin [78]77.3046.60319.90147.9364.20--32.40--
CogAgent [16]81.6044.50--------
DocOwl 1.5 [17]82.2050.70246.40126.4368.8038.7053.7540.6080.2060.40
TextSquare [61]84.3051.50-----49.7084.2066.95
DocLayLLM\\(\\ddagger\\) Llama3 (Ours)86.5258.36327.91157.6077.0740.7358.9058.6283.3670.99
", + "image_path": "287f36606b711a1ac1f5b6c8842d907c5b2df8b75e1d30465fa8a1d0a495e307.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "table_body" + } + ], + "index": 2 + }, + { + "bbox": [ + 55, + 370, + 555, + 394 + ], + "lines": [ + { + "bbox": [ + 55, + 370, + 555, + 394 + ], + "spans": [ + { + "bbox": [ + 55, + 370, + 555, + 394 + ], + "type": "text", + "content": "Table 4. Comparison with OCR-free methods using the evaluation framework and metrics from DocOwl 1.5 [17]. All methods, including our " + }, + { + "bbox": [ + 55, + 370, + 555, + 394 + ], + "type": "inline_equation", + "content": "DocLayLLM_{Llama3}^{\\ddagger}" + }, + { + "bbox": [ + 55, + 370, + 555, + 394 + ], + "type": "text", + "content": " variant, are exposed to the train set corresponding to the evaluation set during training." + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "text" + }, + { + "bbox": [ + 54, + 414, + 294, + 474 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 54, + 414, + 294, + 474 + ], + "spans": [ + { + "bbox": [ + 54, + 414, + 294, + 474 + ], + "type": "text", + "content": "parameters at training. Even with Llama2 as the backbone, our approach matches the performance of previous SOTA methods in the Document-oriented VQA task and significantly outperforms them in the KIE tasks, further demonstrating the efficiency and effectiveness of our approach." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 54, + 474, + 295, + 569 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 54, + 474, + 295, + 569 + ], + "spans": [ + { + "bbox": [ + 54, + 474, + 295, + 569 + ], + "type": "text", + "content": "Besides, our method exhibits robust zero-shot generalization capability without fine-tuning with specific downstream tasks. From the comparison with DocLLM [67] in Table 2, it is clear that our method, even without fine-tuning, significantly outperforms DocLLM across most datasets and achieves a considerable overall performance improvement. This validates our design, which has preserved the strong generalization ability of the original LLM structure." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 54, + 569, + 295, + 641 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 54, + 569, + 295, + 641 + ], + "spans": [ + { + "bbox": [ + 54, + 569, + 295, + 641 + ], + "type": "text", + "content": "In summary, DocLayLLM serves as an efficient extension of LLMs. It demonstrates strong zero-shot generalization and effective document understanding by encoding documents within a unified model. Even with limited training resources, DocLayLLM achieves new SOTA performance compared to OCR-dependent competitors." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 55, + 648, + 286, + 661 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 648, + 286, + 661 + ], + "spans": [ + { + "bbox": [ + 55, + 648, + 286, + 661 + ], + "type": "text", + "content": "4.4. Comparisons with SOTA OCR-free methods" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 55, + 665, + 295, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 665, + 295, + 713 + ], + "spans": [ + { + "bbox": [ + 55, + 665, + 295, + 713 + ], + "type": "text", + "content": "We have also compared our method with SOTA OCR-free TDU models. Table 3 presents the precision of our method and some advent OCR-free TDU methods across various benchmarks. Even though in the circumstance that OCR-" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 313, + 414, + 553, + 486 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 414, + 553, + 486 + ], + "spans": [ + { + "bbox": [ + 313, + 414, + 553, + 486 + ], + "type": "text", + "content": "free models use some of the training sets of these benchmarks during training, while our method infers in a purely zero-shot manner, our DocLayLLM still outperforms these models, particularly in KIE task, which heavily relies on layout information. This clearly demonstrates the superiority of our design to provide LLM with OCR information." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 313, + 486, + 554, + 676 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 486, + 554, + 676 + ], + "spans": [ + { + "bbox": [ + 313, + 486, + 554, + 676 + ], + "type": "text", + "content": "We additionally test DocLayLLM on a broader set of benchmarks from DocOwl 1.5 [17]. To further assess the performance of DocLayLLM on seen datasets, we incorporate the training sets of these benchmarks used in DocOwl 1.5 during the SFT process as other OCR-free TDU methods do during training. To maintain the efficiency advantage of our method, we ensure that the SFT is conducted with the same number of steps as the standard DocLayLLM. The results in Table 4 show that with the inclusion of downstream task data during SFT, our method outperforms the SOTA OCR-free TDU models in most datasets. Moreover, our overall performance in the Document-oriented VQA, KIE, and Table Understanding sub-tasks surpasses the OCR-free TDU methods using limited training resources, showcasing the tremendous potential of OCR-dependent methods and further reinforcing the reliability of our method." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 313, + 677, + 553, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 677, + 553, + 713 + ], + "spans": [ + { + "bbox": [ + 313, + 677, + 553, + 713 + ], + "type": "text", + "content": "Overall, our DocLayLLM consistently shows strong zero-shot inference ability, outperforming many OCR-free TDU methods without prior exposure to downstream data." + } + ] + } + ], + "index": 11 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "text", + "content": "4044" + } + ] + } + ], + "index": 12 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 6 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 63, + 70, + 289, + 242 + ], + "blocks": [ + { + "bbox": [ + 63, + 70, + 289, + 242 + ], + "lines": [ + { + "bbox": [ + 63, + 70, + 289, + 242 + ], + "spans": [ + { + "bbox": [ + 63, + 70, + 289, + 242 + ], + "type": "image", + "image_path": "8162119e9d04c2187be181409bb60212059e6dda14d0319715346cd8850d696d.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 55, + 251, + 296, + 276 + ], + "lines": [ + { + "bbox": [ + 55, + 251, + 296, + 276 + ], + "spans": [ + { + "bbox": [ + 55, + 251, + 296, + 276 + ], + "type": "text", + "content": "Figure 3. Qualitative comparisons of DocLayLLM with other OCR-dependent and OCR-free methods. Zoom in for better view." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "bbox": [ + 55, + 293, + 296, + 343 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 293, + 296, + 343 + ], + "spans": [ + { + "bbox": [ + 55, + 293, + 296, + 343 + ], + "type": "text", + "content": "Additionally, when incorporating downstream datasets during SFT like other OCR-free approaches, DocLayLLM surpasses them, highlighting its potential in solving TDU tasks by integrating OCR information into the LLM." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 55, + 348, + 168, + 361 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 348, + 168, + 361 + ], + "spans": [ + { + "bbox": [ + 55, + 348, + 168, + 361 + ], + "type": "text", + "content": "4.5. Qualitative Results" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 54, + 365, + 295, + 510 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 54, + 365, + 295, + 510 + ], + "spans": [ + { + "bbox": [ + 54, + 365, + 295, + 510 + ], + "type": "text", + "content": "Figure 3 shows some qualitative examples to visualize the performance on KIE tasks of our DocLayLLM compared with LayTextLLM and DocOwl 1.5. The example from SROIE clearly shows that DocLayLLM has improved document comprehension and answer accuracy. And the example from POIE demonstrates our advantage in handling table-formatted documents to give an accurate response. Furthermore, as seen in the FUNSD examples, DocLayLLM excels at extracting answers spanning multiple text lines, outperforming both OCR-dependent and OCR-free methods. For more visualization of qualitative examples, please refer to the appendix." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 55, + 517, + 149, + 529 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 517, + 149, + 529 + ], + "spans": [ + { + "bbox": [ + 55, + 517, + 149, + 529 + ], + "type": "text", + "content": "4.6. Ablation Study" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 55, + 534, + 295, + 617 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 534, + 295, + 617 + ], + "spans": [ + { + "bbox": [ + 55, + 534, + 295, + 617 + ], + "type": "text", + "content": "We have also conducted additional experiments to explore the effectiveness of our proposed method. To conserve computational resources, we perform pre-training at " + }, + { + "bbox": [ + 55, + 534, + 295, + 617 + ], + "type": "inline_equation", + "content": "25\\%" + }, + { + "bbox": [ + 55, + 534, + 295, + 617 + ], + "type": "text", + "content": " of the standard scale (including the baseline) but full-scale SFT. We focus on the impact of two main factors: (a) the role of CoT pre-training and CoT annealing, and (b) the influence of different modality configurations." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 55, + 618, + 295, + 701 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 618, + 295, + 701 + ], + "spans": [ + { + "bbox": [ + 55, + 618, + 295, + 701 + ], + "type": "text", + "content": "Experimental results in Table 5a indicate that omitting CoT Pre-training leads to an obvious performance drop in both the Document-oriented VQA and KIE tasks, while not using CoT Annealing primarily affects VQA performance. Besides, using randomly initialized " + }, + { + "bbox": [ + 55, + 618, + 295, + 701 + ], + "type": "inline_equation", + "content": "VE" + }, + { + "bbox": [ + 55, + 618, + 295, + 701 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 55, + 618, + 295, + 701 + ], + "type": "inline_equation", + "content": "LE" + }, + { + "bbox": [ + 55, + 618, + 295, + 701 + ], + "type": "text", + "content": " partially compromises the proposed extension, resulting in performance degradation in both the VQA and KIE tasks." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 67, + 701, + 295, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 67, + 701, + 295, + 714 + ], + "spans": [ + { + "bbox": [ + 67, + 701, + 295, + 714 + ], + "type": "text", + "content": "Additionally, performances in Table 5b show the efficacy" + } + ] + } + ], + "index": 8 + }, + { + "type": "table", + "bbox": [ + 323, + 71, + 547, + 121 + ], + "blocks": [ + { + "bbox": [ + 323, + 71, + 547, + 121 + ], + "lines": [ + { + "bbox": [ + 323, + 71, + 547, + 121 + ], + "spans": [ + { + "bbox": [ + 323, + 71, + 547, + 121 + ], + "type": "table", + "html": "
ExperimentDocument-oriented VQAKIE
DocVQAVisualMRCAvg.DeepFormKLCAvg.
DocLayLLM1lama377.43316.27196.8535.4126.2530.83
w/o CoT Pre-training76.00295.49185.7524.5425.4925.02
w/o CoT Annealing77.40310.17193.7934.2326.2130.22
w/o pt VE & LE77.42308.49192.9631.3125.7628.54
", + "image_path": "ef9954dfbaba0c478fdb53ec456c687b68fa891b89c60be6e856cf44f9c4f9d4.jpg" + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "table_body" + }, + { + "bbox": [ + 321, + 123, + 548, + 163 + ], + "lines": [ + { + "bbox": [ + 321, + 123, + 548, + 163 + ], + "spans": [ + { + "bbox": [ + 321, + 123, + 548, + 163 + ], + "type": "text", + "content": "(a) Impact of different settings: " + }, + { + "bbox": [ + 321, + 123, + 548, + 163 + ], + "type": "inline_equation", + "content": "w / o" + }, + { + "bbox": [ + 321, + 123, + 548, + 163 + ], + "type": "text", + "content": " CoT Pre-training and " + }, + { + "bbox": [ + 321, + 123, + 548, + 163 + ], + "type": "inline_equation", + "content": "w / o" + }, + { + "bbox": [ + 321, + 123, + 548, + 163 + ], + "type": "text", + "content": " CoT Annealing indicate pre-training without CoT and supervised fine-tuning with only CoT data, while " + }, + { + "bbox": [ + 321, + 123, + 548, + 163 + ], + "type": "inline_equation", + "content": "w / o" + }, + { + "bbox": [ + 321, + 123, + 548, + 163 + ], + "type": "text", + "content": " pt VE and PE uses randomly initialized visual and layout embeddings instead of pre-trained ones." + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "table_footnote" + } + ], + "index": 9 + }, + { + "type": "table", + "bbox": [ + 323, + 174, + 547, + 230 + ], + "blocks": [ + { + "bbox": [ + 323, + 174, + 547, + 230 + ], + "lines": [ + { + "bbox": [ + 323, + 174, + 547, + 230 + ], + "spans": [ + { + "bbox": [ + 323, + 174, + 547, + 230 + ], + "type": "table", + "html": "
ModalityDocument-oriented VQAKIE
TLVDocVQAVisualMRCAvg.DeepFormKLCAvg.
✓✗✗73.57280.27176.9315.7624.0919.93
✓(I)✗75.72306.51191.1214.1324.2319.18
✓(II)✗77.32310.75194.0333.7525.9729.86
✓(II)✓77.43316.27196.8535.4126.2530.83
", + "image_path": "2baf010f8aee4379dade9668a79bc191ed61cb6c57177b5c3982b6e3c50dd30b.jpg" + } + ] + } + ], + "index": 11, + "angle": 0, + "type": "table_body" + }, + { + "bbox": [ + 321, + 232, + 547, + 270 + ], + "lines": [ + { + "bbox": [ + 321, + 232, + 547, + 270 + ], + "spans": [ + { + "bbox": [ + 321, + 232, + 547, + 270 + ], + "type": "text", + "content": "(b) Impacts of different modality configurations. " + }, + { + "bbox": [ + 321, + 232, + 547, + 270 + ], + "type": "inline_equation", + "content": "\\mathbf{T}" + }, + { + "bbox": [ + 321, + 232, + 547, + 270 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 321, + 232, + 547, + 270 + ], + "type": "inline_equation", + "content": "\\mathbf{L}" + }, + { + "bbox": [ + 321, + 232, + 547, + 270 + ], + "type": "text", + "content": ", and " + }, + { + "bbox": [ + 321, + 232, + 547, + 270 + ], + "type": "inline_equation", + "content": "\\mathbf{V}" + }, + { + "bbox": [ + 321, + 232, + 547, + 270 + ], + "type": "text", + "content": " represent textual, layout, and visual modalities, respectively. (I) indicates inputting OCR coordinates as text following He et al. [14], while (II) embeds coordinates with " + }, + { + "bbox": [ + 321, + 232, + 547, + 270 + ], + "type": "inline_equation", + "content": "LE" + }, + { + "bbox": [ + 321, + 232, + 547, + 270 + ], + "type": "text", + "content": " in our approach." + } + ] + } + ], + "index": 12, + "angle": 0, + "type": "table_footnote" + } + ], + "index": 11 + }, + { + "bbox": [ + 318, + 280, + 550, + 292 + ], + "lines": [ + { + "bbox": [ + 318, + 280, + 550, + 292 + ], + "spans": [ + { + "bbox": [ + 318, + 280, + 550, + 292 + ], + "type": "text", + "content": "Table 5. Ablation study on various designs of our DocLayLLM." + } + ] + } + ], + "index": 13, + "angle": 0, + "type": "text" + }, + { + "bbox": [ + 313, + 312, + 553, + 418 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 312, + 553, + 418 + ], + "spans": [ + { + "bbox": [ + 313, + 312, + 553, + 418 + ], + "type": "text", + "content": "of our DocLayLLM to incorporate different modalities for TDU tasks. It is obvious that using only text from the document is insufficient for TDU, while adding layout information in textual form provides some gain on VQA tasks but leads to over-length input (see appendix for analysis), greatly slowing the inference. By our design to integrate layout features via an " + }, + { + "bbox": [ + 313, + 312, + 553, + 418 + ], + "type": "inline_equation", + "content": "LE" + }, + { + "bbox": [ + 313, + 312, + 553, + 418 + ], + "type": "text", + "content": ", we achieve notable improvements on these TDU tasks with shorter inputs. And adding visual features further boosts the performance of our method." + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 313, + 419, + 553, + 444 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 419, + 553, + 444 + ], + "spans": [ + { + "bbox": [ + 313, + 419, + 553, + 444 + ], + "type": "text", + "content": "Overall, the combined impact of these designs ensures the efficiency and effectiveness of our method." + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 313, + 454, + 388, + 466 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 454, + 388, + 466 + ], + "spans": [ + { + "bbox": [ + 313, + 454, + 388, + 466 + ], + "type": "text", + "content": "5. Conclusion" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 313, + 474, + 555, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 474, + 555, + 715 + ], + "spans": [ + { + "bbox": [ + 313, + 474, + 555, + 715 + ], + "type": "text", + "content": "We propose DocLayLLM, an efficient multi-modal extension of large language models for text-rich document understanding. By streamlined integration of document patch tokens and OCR positional tokens into the LLM inputs, it achieves a simple yet efficient multi-modal extension of LLM to enable it to have the capability to perceive and process textual, visual, and layout information from the OCR results of the document. Additionally, we introduce CoT Pre-training and CoT Annealing techniques centered around CoT, offering a new way to generate CoT data and further enhancing the model's understanding of text-rich documents. Despite requiring fewer training resources, our method surpasses existing OCR-dependent TDU models, demonstrating strong zero-shot generalization capability. Furthermore, it also exceeds the performance of SOTA OCR-free TDU methods under comparable training conditions, showcasing remarkable performance across various document understanding tasks. Our work significantly advances efficient MLLM-based document understanding while maintaining high-performance standards." + } + ] + } + ], + "index": 17 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "text", + "content": "4045" + } + ] + } + ], + "index": 18 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 7 + }, + { + "para_blocks": [ + { + "bbox": [ + 56, + 72, + 153, + 85 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 72, + 153, + 85 + ], + "spans": [ + { + "bbox": [ + 56, + 72, + 153, + 85 + ], + "type": "text", + "content": "Acknowledgement" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 55, + 90, + 297, + 124 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 90, + 297, + 124 + ], + "spans": [ + { + "bbox": [ + 55, + 90, + 297, + 124 + ], + "type": "text", + "content": "This research is supported in part by National Natural Science Foundation of China (Grant No.: 62476093, 62441604)." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 56, + 133, + 115, + 146 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 133, + 115, + 146 + ], + "spans": [ + { + "bbox": [ + 56, + 133, + 115, + 146 + ], + "type": "text", + "content": "References" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 57, + 153, + 296, + 712 + ], + "type": "list", + "angle": 0, + "index": 14, + "blocks": [ + { + "bbox": [ + 61, + 153, + 296, + 196 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 61, + 153, + 296, + 196 + ], + "spans": [ + { + "bbox": [ + 61, + 153, + 296, + 196 + ], + "type": "text", + "content": "[1] Srikar Appalaraju, Bhavan Jasani, Bhargava Urala Kota, Yusheng Xie, and R. Manmatha. DocFormer: End-to-End Transformer for Document Understanding. In Proc. ICCV, pages 993–1003, 2021. 1" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 61, + 198, + 296, + 240 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 61, + 198, + 296, + 240 + ], + "spans": [ + { + "bbox": [ + 61, + 198, + 296, + 240 + ], + "type": "text", + "content": "[2] Srikar Appalaraju, Peng Tang, Qi Dong, Nishant Sankaran, Yichu Zhou, and R. Manmatha. DocFormerv2: Local Features for Document Understanding. In Proc. AAAI, pages 709-718, 2024. 1" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 62, + 241, + 296, + 350 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 241, + 296, + 350 + ], + "spans": [ + { + "bbox": [ + 62, + 241, + 296, + 350 + ], + "type": "text", + "content": "[3] Tom Brown, Benjamin Mann, Nick Ryder, Melanie Subbiah, Jared D Kaplan, Prafulla Dhariwal, Arvind Neelakantan, Pranav Shyam, Girish Sastry, Amanda Askell, Sandhini Agarwal, Ariel Herbert-Voss, Gretchen Krueger, Tom Henighan, Rewon Child, Aditya Ramesh, Daniel Ziegler, Jeffrey Wu, Clemens Winter, Chris Hesse, Mark Chen, Eric Sigler, Mateusz Litwin, Scott Gray, Benjamin Chess, Jack Clark, Christopher Berner, Sam McCandlish, Alec Radford, Ilya Sutskever, and Dario Amodei. Language Models are Few-shot Learners. Proc. NIPS, 33:1877-1901, 2020. 2" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 62, + 351, + 295, + 394 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 351, + 295, + 394 + ], + "spans": [ + { + "bbox": [ + 62, + 351, + 295, + 394 + ], + "type": "text", + "content": "[4] Wenhu Chen, Hongmin Wang, Jianshu Chen, Yunkai Zhang, Hong Wang, Shiyang Li, Xiyou Zhou, and William Yang Wang. TabFact: A Large-scale Dataset for Table-based Fact Verification. In Proc. ICLR, 2020. 5" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 62, + 396, + 295, + 438 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 396, + 295, + 438 + ], + "spans": [ + { + "bbox": [ + 62, + 396, + 295, + 438 + ], + "type": "text", + "content": "[5] Gang Dai, Yifan Zhang, Qingfeng Wang, Qing Du, Zhuliang Yu, Zhuoman Liu, and Shuangping Huang. Disentangling Writer and Character Styles for Handwriting Generation. In Proc. CVPR, pages 5977-5986, 2023. 3" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 62, + 439, + 295, + 482 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 439, + 295, + 482 + ], + "spans": [ + { + "bbox": [ + 62, + 439, + 295, + 482 + ], + "type": "text", + "content": "[6] Gang Dai, Yifan Zhang, Quhui Ke, Qiangya Guo, and Shuangping Huang. One-dm: One-shot diffusion mimicker for handwritten text generation. In Proc. ECCV, pages 410-427, 2025. 3" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 62, + 483, + 295, + 515 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 483, + 295, + 515 + ], + "spans": [ + { + "bbox": [ + 62, + 483, + 295, + 515 + ], + "type": "text", + "content": "[7] DeepSeek-AI. DeepSeek-R1: Incentivizing Reasoning Capability in LLMs via Reinforcement Learning. arXiv: Comp. Res. Repository, abs/2501.12948, 2025. 2" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 62, + 516, + 295, + 559 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 516, + 295, + 559 + ], + "spans": [ + { + "bbox": [ + 62, + 516, + 295, + 559 + ], + "type": "text", + "content": "[8] Jacob Devlin, Ming-Wei Chang, Kenton Lee, and Kristina Toutanova. BERT: Pre-training of Deep Bidirectional Transformers for Language Understanding. In Proc. NAACL, pages 4171–4186, 2019. 1" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 62, + 560, + 295, + 613 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 560, + 295, + 613 + ], + "spans": [ + { + "bbox": [ + 62, + 560, + 295, + 613 + ], + "type": "text", + "content": "[9] Hao Feng, Qi Liu, Hao Liu, Wengang Zhou, Houqiang Li, and Can Huang. DocPedia: Unleashing the Power of Large Multimodal Model in the Frequency Domain for Versatile Document Understanding. arXiv: Comp. Res. Repository, abs/2311.11810, 2023. 7" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 57, + 615, + 295, + 669 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 615, + 295, + 669 + ], + "spans": [ + { + "bbox": [ + 57, + 615, + 295, + 669 + ], + "type": "text", + "content": "[10] Hao Feng, Zijian Wang, Jingqun Tang, Jinghui Lu, Wengang Zhou, Houqiang Li, and Can Huang. UniDoc: A Universal Large Multimodal Model for Simultaneous Text Detection, Recognition, Spotting and Understanding. arXiv: Comp. Res. Repository, abs/2308.11592, 2023. 7" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 57, + 670, + 295, + 712 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 670, + 295, + 712 + ], + "spans": [ + { + "bbox": [ + 57, + 670, + 295, + 712 + ], + "type": "text", + "content": "[11] Masato Fujitake. LayoutLLM: Large Language Model Instruction Tuning for Visually Rich Document Understanding. In Proc. LREC/COLING, pages 10219-10224, 2024. 2, 3" + } + ] + } + ], + "index": 13 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 316, + 73, + 554, + 714 + ], + "type": "list", + "angle": 0, + "index": 27, + "blocks": [ + { + "bbox": [ + 316, + 73, + 554, + 127 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 73, + 554, + 127 + ], + "spans": [ + { + "bbox": [ + 316, + 73, + 554, + 127 + ], + "type": "text", + "content": "[12] Lukasz Garncarek, Rafat Powalski, Tomasz Stanisławek, Bartosz Topolski, Piotr Halama, Michal Turski, and Filip Graliński. LAMBERT: Layout-Aware Language Modeling for Information Extraction. In Proc. ICDAR, pages 532-547, 2021. 1" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 316, + 129, + 554, + 173 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 129, + 554, + 173 + ], + "spans": [ + { + "bbox": [ + 316, + 129, + 554, + 173 + ], + "type": "text", + "content": "[13] Adam W. Harley, Alex Ufkes, and Konstantinos G. Derpanis. Evaluation of Deep Convolutional Nets for Document Image Classification and Retrieval. In Proc. ICDAR, pages 991-995, 2015. 5" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 316, + 175, + 554, + 228 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 175, + 554, + 228 + ], + "spans": [ + { + "bbox": [ + 316, + 175, + 554, + 228 + ], + "type": "text", + "content": "[14] Jiabang He, Lei Wang, Yi Hu, Ning Liu, Hui Liu, Xing Xu, and Heng Tao Shen. ICL-D3IE: In-Context Learning with Diverse Demonstrations Updating for Document Information Extraction. In Proc. ICCV, pages 19428-19437, 2023. 1, 2, 3, 6, 8" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 316, + 230, + 554, + 285 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 230, + 554, + 285 + ], + "spans": [ + { + "bbox": [ + 316, + 230, + 554, + 285 + ], + "type": "text", + "content": "[15] Teakgyu Hong, DongHyun Kim, Mingi Ji, Wonseok Hwang, Daehyun Nam, and Sungrae Park. BROS: A Pre-trained Language Model Focusing on Text and Layout for Better Key Information Extraction from Documents. In Proc. AAAI, pages 10767-10775, 2022. 1" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 316, + 287, + 554, + 341 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 287, + 554, + 341 + ], + "spans": [ + { + "bbox": [ + 316, + 287, + 554, + 341 + ], + "type": "text", + "content": "[16] Wenyi Hong, Weihan Wang, Qingsong Lv, Jiazheng Xu, Wenmeng Yu, Junhui Ji, Yan Wang, Zihan Wang, Yuxiao Dong, Ming Ding, and Jie Tang. CogAgent: A Visual Language Model for GUI Agents. In Proc. CVPR, pages 14281-14290, 2024. 7" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 316, + 343, + 554, + 397 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 343, + 554, + 397 + ], + "spans": [ + { + "bbox": [ + 316, + 343, + 554, + 397 + ], + "type": "text", + "content": "[17] Anwen Hu, Haiyang Xu, Jiabo Ye, Ming Yan, Liang Zhang, Bo Zhang, Ji Zhang, Qin Jin, Fei Huang, and Jingren Zhou. mPLUG-DocOwl 1.5: Unified Structure Learning for OCR-free Document Understanding. In Findings EMNLP, pages 3096-3120, 2024. 1, 2, 7" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 316, + 399, + 554, + 443 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 399, + 554, + 443 + ], + "spans": [ + { + "bbox": [ + 316, + 399, + 554, + 443 + ], + "type": "text", + "content": "[18] Edward J. Hu, Yelong Shen, Phillip Wallis, Zeyuan Allen-Zhu, Yuanzhi Li, Shean Wang, Lu Wang, and Weizhu Chen. LoRA: Low-Rank Adaptation of Large Language Models. In Proc. ICLR, 2022. 2, 6" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 317, + 445, + 554, + 542 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 445, + 554, + 542 + ], + "spans": [ + { + "bbox": [ + 317, + 445, + 554, + 542 + ], + "type": "text", + "content": "[19] Shengding Hu, Yuge Tu, Xu Han, Chaoqun He, Ganqu Cui, Xiang Long, Zhi Zheng, Yewei Fang, Yuxiang Huang, Weilin Zhao, Xinrong Zhang, Zhen Leng Thai, Kai Zhang, Chongyi Wang, Yuan Yao, Chenyang Zhao, Jie Zhou, Jie Cai, Zhongwu Zhai, Ning Ding, Chao Jia, Guoyang Zeng, Dahai Li, Zhiyuan Liu, and Maosong Sun. MiniCPM: Unveiling the Potential of Small Language Models with Scalable Training Strategies. arXiv: Comp. Res. Repository, abs/2404.06395, 2024. 5" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 316, + 544, + 554, + 588 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 544, + 554, + 588 + ], + "spans": [ + { + "bbox": [ + 316, + 544, + 554, + 588 + ], + "type": "text", + "content": "[20] Yupan Huang, Tengchao Lv, Lei Cui, Yutong Lu, and Furu Wei. LayoutLMv3: Pre-training for Document AI with Unified Text and Image Masking. In Proc. ACM MM, pages 4083-4091, 2022. 1, 2, 3, 6" + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 316, + 590, + 554, + 634 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 590, + 554, + 634 + ], + "spans": [ + { + "bbox": [ + 316, + 590, + 554, + 634 + ], + "type": "text", + "content": "[21] Zheng Huang, Kai Chen, Jianhua He, Xiang Bai, Dimosthenis Karatzas, Shijian Lu, and C. V. Jawahar. ICDAR2019 Competition on Scanned Receipt OCR and Information Extraction. In Proc. ICDAR, pages 1516-1520, 2019. 5" + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 316, + 635, + 554, + 678 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 635, + 554, + 678 + ], + "spans": [ + { + "bbox": [ + 316, + 635, + 554, + 678 + ], + "type": "text", + "content": "[22] Guillaume Jaume, Hazim Kemal Ekenel, and Jean-Philippe Thiran. FUNSD: A Dataset for Form Understanding in Noisy Scanned Documents. In Workshop on ICDAR, pages 1–6, 2019. 5" + } + ] + } + ], + "index": 25 + }, + { + "bbox": [ + 316, + 680, + 554, + 714 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 680, + 554, + 714 + ], + "spans": [ + { + "bbox": [ + 316, + 680, + 554, + 714 + ], + "type": "text", + "content": "[23] Geewook Kim, Teakgyu Hong, Moonbin Yim, JeongYeon Nam, Jinyoung Park, Jinyeong Yim, Wonseok Hwang, Sangdoo Yun, Dongyoon Han, and Seunghyun Park. OCR-Free" + } + ] + } + ], + "index": 26 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 748, + 316, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 316, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 316, + 757 + ], + "type": "text", + "content": "4046" + } + ] + } + ], + "index": 28 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 8 + }, + { + "para_blocks": [ + { + "bbox": [ + 56, + 72, + 295, + 713 + ], + "type": "list", + "angle": 0, + "index": 10, + "blocks": [ + { + "bbox": [ + 76, + 72, + 294, + 95 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 76, + 72, + 294, + 95 + ], + "spans": [ + { + "bbox": [ + 76, + 72, + 294, + 95 + ], + "type": "text", + "content": "Document Understanding Transformer. Proc. ECCV, 13688: 498-517, 2022. 2" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 56, + 96, + 295, + 162 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 96, + 295, + 162 + ], + "spans": [ + { + "bbox": [ + 56, + 96, + 295, + 162 + ], + "type": "text", + "content": "[24] Geewook Kim, Hodong Lee, Daehee Kim, Haeji Jung, Sanghee Park, Yoonsik Kim, Sangdoo Yun, Taeho Kil, Bado Lee, and Seunghyun Park. Visually-Situated Natural Language Understanding with Contrastive Reading Model and Frozen Large Language Models. In Proc. EMNLP, pages 11989-12010, 2023. 2, 3" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 56, + 164, + 295, + 208 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 164, + 295, + 208 + ], + "spans": [ + { + "bbox": [ + 56, + 164, + 295, + 208 + ], + "type": "text", + "content": "[25] Jianfeng Kuang, Wei Hua, Dingkang Liang, Mingkun Yang, Deqiang Jiang, Bo Ren, and Xiang Bai. Visual Information Extraction in the Wild: Practical Dataset and End-to-End Solution. In Proc. ICDAR, pages 36-53, 2023. 5" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 56, + 210, + 294, + 253 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 210, + 294, + 253 + ], + "spans": [ + { + "bbox": [ + 56, + 210, + 294, + 253 + ], + "type": "text", + "content": "[26] Hugo Laurençon, Andres Marafioti, Victor Sanh, and Leo Tronchon. Building and better understanding vision-language models: insights and future directions. In Workshop on NIPS, 2024. 3" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 56, + 255, + 294, + 308 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 255, + 294, + 308 + ], + "spans": [ + { + "bbox": [ + 56, + 255, + 294, + 308 + ], + "type": "text", + "content": "[27] Bo Li, Yuanhan Zhang, Dong Guo, Renrui Zhang, Feng Li, Hao Zhang, Kaichen Zhang, Peiyuan Zhang, Yanwei Li, Ziwei Liu, and Chunyuan Li. LLaVA-OneVision: Easy Visual Task Transfer. Transactions on Machine Learning Research, 2025. 3" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 56, + 312, + 294, + 355 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 312, + 294, + 355 + ], + "spans": [ + { + "bbox": [ + 56, + 312, + 294, + 355 + ], + "type": "text", + "content": "[28] Chenliang Li, Bin Bi, Ming Yan, Wei Wang, Songfang Huang, Fei Huang, and Luo Si. StructuralLM: Structural Pre-training for Form Understanding. In Proc. ACL, pages 6309-6318, 2021. 1" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 56, + 357, + 294, + 411 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 357, + 294, + 411 + ], + "spans": [ + { + "bbox": [ + 56, + 357, + 294, + 411 + ], + "type": "text", + "content": "[29] Feng Li, Renrui Zhang, Hao Zhang, Yuanhan Zhang, Bo Li, Wei Li, Zejun Ma, and Chunyuan Li. LLaVa-NeXT-Interleave: Tackling Multi-image, Video, and 3D in Large Multimodal Models. arXiv: Comp. Res. Repository, abs/2407.07895, 2023. 2" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 56, + 414, + 294, + 456 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 414, + 294, + 456 + ], + "spans": [ + { + "bbox": [ + 56, + 414, + 294, + 456 + ], + "type": "text", + "content": "[30] Junnan Li, Dongxu Li, Silvio Savarese, and Steven C. H. Hoi. BLIP-2: Bootstrapping Language-Image Pre-training with Frozen Image Encoders and Large Language Models. Proc. ICML, 202:19730-19742, 2023. 3" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 56, + 460, + 295, + 667 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 460, + 295, + 667 + ], + "spans": [ + { + "bbox": [ + 56, + 460, + 295, + 667 + ], + "type": "text", + "content": "[31] Jeffrey Li, Alex Fang, Georgios Smyrnis, Maor Ivgi, Matt Jordan, Samir Yitzhak Gadre, Hritik Bansal, Etash Guha, Sedrick Keh, Kushal Arora, Saurabh Garg, Rui Xin, Niklas Muennighoff, Reinhard Heckel, Jean Mercat, Mayee Chen, Suchin Gururangan, Mitchell Wortsman, Alon Albalak, Yonatan Bitton, Marianna Nezhurina, Amro Abbas, ChengYu Hsieh, Dhruba Ghosh, Josh Gardner, Maciej Kilian, Hanlin Zhang, Rulin Shao, Sarah M. Pratt, Sunny Sanyal, Gabriel Ilharco, Giannis Daras, Kalyani Marathe, Aaron Gokaslan, Jieyu Zhang, Khyathi Raghavi Chandu, Thao Nguyen, Igor Vasiljevic, Sham M. Kakade, Shuran Song, Sujay Sanghavi, Fartash Faghri, Sewoong Oh, Luke Zettlemoyer, Kyle Lo, Alaaeldin El-Nouby, Hadi Pouransari, Alexander Toshev, Stephanie Wang, Dirk Groeneveld, Luca Soldaini, Pang Wei Koh, Jenia Jitsev, Thomas Kollar, Alexandros G. Dimakis, Yair Carmon, Achal Dave, Ludwig Schmidt, and Vaishaal Shankar. DataComp-LM: In search of the next generation of training sets for language models. arXiv: Comp. Res. Repository, abs/2406.11794, 2024. 2, 5" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 56, + 670, + 294, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 670, + 294, + 713 + ], + "spans": [ + { + "bbox": [ + 56, + 670, + 294, + 713 + ], + "type": "text", + "content": "[32] Minghao Li, Yiheng Xu, Lei Cui, Shaohan Huang, Furu Wei, Zhoujun Li, and Ming Zhou. DocBank: A Benchmark Dataset for Document Layout Analysis. In Pro. COLING, pages 949-960, 2020. 5" + } + ] + } + ], + "index": 9 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 316, + 73, + 553, + 713 + ], + "type": "list", + "angle": 0, + "index": 26, + "blocks": [ + { + "bbox": [ + 316, + 73, + 553, + 116 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 73, + 553, + 116 + ], + "spans": [ + { + "bbox": [ + 316, + 73, + 553, + 116 + ], + "type": "text", + "content": "[33] Peizhao Li, Jiuxiang Gu, Jason Kuen, Vlad I Morariu, Handong Zhao, Rajiv Jain, Varun Manjunatha, and Hongfu Liu. SelfDoc: Self-Supervised Document Representation Learning. In Proc. CVPR, pages 5652-5660, 2021. 1" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 316, + 118, + 553, + 160 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 118, + 553, + 160 + ], + "spans": [ + { + "bbox": [ + 316, + 118, + 553, + 160 + ], + "type": "text", + "content": "[34] Qiwei Li, Zuchao Li, Xiantao Cai, Bo Du, and Hai Zhao. Enhancing Visually-Pich Pdocument Punderstanding via Payout Pstructure Podeling. In Proc. ACM MM, page 4513-4523, 2023." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 316, + 162, + 553, + 194 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 162, + 553, + 194 + ], + "spans": [ + { + "bbox": [ + 316, + 162, + 553, + 194 + ], + "type": "text", + "content": "[35] Qiwei Li, Zuchao Li, Ping Wang, Haojun Ai, and Hai Zhao. Hypergraph based Understanding for Document Semantic Entity Recognition. In Proc. ACL, pages 2950-2960, 2024." + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 316, + 195, + 553, + 238 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 195, + 553, + 238 + ], + "spans": [ + { + "bbox": [ + 316, + 195, + 553, + 238 + ], + "type": "text", + "content": "[36] Xin Li, Yan Zheng, Yiqing Hu, Haoyu Cao, Yunfei Wu, Deqiang Jiang, Yinsong Liu, and Bo Ren. Relational Representation Learning in Visually-Rich Documents. In Proc. ACM MM, page 4614-4624, 2022." + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 316, + 239, + 553, + 293 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 239, + 553, + 293 + ], + "spans": [ + { + "bbox": [ + 316, + 239, + 553, + 293 + ], + "type": "text", + "content": "[37] Yulin Li, Yuxi Qian, Yuechen Yu, Xiameng Qin, Chengquan Zhang, Yan Liu, Kun Yao, Junyu Han, Jingtuo Liu, and Errui Ding. StrucTexT: Structured Text Understanding with Multi-Modal Transformers. In Proc. ACM MM, page 1912–1920, 2021. 1" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 316, + 294, + 553, + 348 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 294, + 553, + 348 + ], + "spans": [ + { + "bbox": [ + 316, + 294, + 553, + 348 + ], + "type": "text", + "content": "[38] Zhang Li, Biao Yang, Qiang Liu, Zhiyin Ma, Shuo Zhang, Jingxu Yang, Yabo Sun, Yuliang Liu, and Xiang Bai. Monkey: Image Resolution and Text Label Are Important Things for Large Multi-modal Models. In Proc. CVPR, pages 26763-26773, 2024. 1, 2, 7" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 316, + 350, + 553, + 392 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 350, + 553, + 392 + ], + "spans": [ + { + "bbox": [ + 316, + 350, + 553, + 392 + ], + "type": "text", + "content": "[39] Zening Lin, Jiapeng Wang, and Lianwen Jin. Visual Information Extraction Deep Learning Method: A Critical Review. Journal of Image and Graphics, 28(08):2276-2297, 2023. 1" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 316, + 393, + 553, + 436 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 393, + 553, + 436 + ], + "spans": [ + { + "bbox": [ + 316, + 393, + 553, + 436 + ], + "type": "text", + "content": "[40] Chenglin Liu, Lianwen Jin, Xiang Bai, Xiaohui Li, and Fei Yin. Frontiers of intelligent document analysis and recognition: review and prospects. Journal of Image and Graphics, 28(08):2223-2252, 2023. 1" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 316, + 437, + 553, + 481 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 437, + 553, + 481 + ], + "spans": [ + { + "bbox": [ + 316, + 437, + 553, + 481 + ], + "type": "text", + "content": "[41] Yuliang Liu, Biao Yang, Qiang Liu, Zhang Li, Zhiyin Ma, Shuo Zhang, and Xiang Bai. TextMonkey: An OCR-Free Large Multimodal Model for Understanding Document. arXiv: Comp. Res. Repository, abs/2403.04473, 2024. 7" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 316, + 482, + 553, + 504 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 482, + 553, + 504 + ], + "spans": [ + { + "bbox": [ + 316, + 482, + 553, + 504 + ], + "type": "text", + "content": "[42] Meta AI Llama Team. The Llama 3 Herd of Models. arXiv: Comp. Res. Repository, abs/2407.21783, 2024. 2, 5, 6" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 316, + 505, + 553, + 569 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 505, + 553, + 569 + ], + "spans": [ + { + "bbox": [ + 316, + 505, + 553, + 569 + ], + "type": "text", + "content": "[43] Jinghui Lu, Haiyang Yu, Yanjie Wang, Yongjie Ye, Jingqun Tang, Ziwei Yang, Binghong Wu, Qi Liu, Hao Feng, Han Wang, Hao Liu, and Can Huang. A Bounding Box is Worth One Token: Interleaving Layout and Text in a Large Language Model for Document Understanding. arXiv: Comp. Res. Repository, abs/2407.01976, 2024. 6" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 316, + 571, + 553, + 602 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 571, + 553, + 602 + ], + "spans": [ + { + "bbox": [ + 316, + 571, + 553, + 602 + ], + "type": "text", + "content": "[44] Chuwei Luo, Changxu Cheng, Qi Zheng, and Cong Yao. GeoLayoutLM: Geometric Pre-Training for Visual Information Extraction. In Proc. CVPR, pages 7092–7101, 2023. 1" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 316, + 604, + 553, + 647 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 604, + 553, + 647 + ], + "spans": [ + { + "bbox": [ + 316, + 604, + 553, + 647 + ], + "type": "text", + "content": "[45] Chuwei Luo, Yufan Shen, Zhaoqing Zhu, Qi Zheng, Zhi Yu, and Cong Yao. LayoutLLM: Layout Instruction Tuning with Large Language Models for Document Understanding. In Proc. CVPR, pages 15630-15640, 2024. 2, 3, 4, 5, 6" + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 316, + 647, + 553, + 680 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 647, + 553, + 680 + ], + "spans": [ + { + "bbox": [ + 316, + 647, + 553, + 680 + ], + "type": "text", + "content": "[46] Minesh Mathew, Dimosthenis Karatzas, and C. V. Jawahar. DocVQA: A Dataset for VQA on Document Images. In Proc. WACV, pages 2199-2208, 2021. 5" + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 316, + 681, + 553, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 681, + 553, + 713 + ], + "spans": [ + { + "bbox": [ + 316, + 681, + 553, + 713 + ], + "type": "text", + "content": "[47] Minesh Mathew, Viraj Bagal, Ruben Tito, Dimosthenis Karatzas, Ernest Valveny, and C.V. Jawahar. InfographicVQA. In Proc. WACV, pages 1697-1706, 2022. 5" + } + ] + } + ], + "index": 25 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "text", + "content": "4047" + } + ] + } + ], + "index": 27 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 9 + }, + { + "para_blocks": [ + { + "bbox": [ + 56, + 73, + 295, + 713 + ], + "type": "list", + "angle": 0, + "index": 14, + "blocks": [ + { + "bbox": [ + 56, + 73, + 294, + 106 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 73, + 294, + 106 + ], + "spans": [ + { + "bbox": [ + 56, + 73, + 294, + 106 + ], + "type": "text", + "content": "[48] G. Nagy, S. Seth, and M. Viswanathan. A Prototype Document Image Analysis System for Technical Journals. Computer, 25(7):10-22, 1992. 5" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 56, + 107, + 197, + 118 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 107, + 197, + 118 + ], + "spans": [ + { + "bbox": [ + 56, + 107, + 197, + 118 + ], + "type": "text", + "content": "[49] OpenAI. Hello GPT-4o, 2024. 2" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 56, + 120, + 295, + 163 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 120, + 295, + 163 + ], + "spans": [ + { + "bbox": [ + 56, + 120, + 295, + 163 + ], + "type": "text", + "content": "[50] Seunghyun Park, Seung Shin, Bado Lee, Junyeop Lee, Jaeheung Surh, Minjoon Seo, and Hwalsuk Lee. CORD: A Consolidated Receipt Dataset for Post-OCR Parsing. In Workshop on NIPS, 2019. 5" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 57, + 164, + 295, + 196 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 164, + 295, + 196 + ], + "spans": [ + { + "bbox": [ + 57, + 164, + 295, + 196 + ], + "type": "text", + "content": "[51] Panupong Pasupat and Percy Liang. Compositional Semantic Parsing on Semi-Structured Tables. In Proc. ACL, pages 1470–1480, 2015. 5" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 56, + 198, + 295, + 274 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 198, + 295, + 274 + ], + "spans": [ + { + "bbox": [ + 56, + 198, + 295, + 274 + ], + "type": "text", + "content": "[52] Qiming Peng, Yinxu Pan, Wenjin Wang, Bin Luo, Zhenyu Zhang, Zhengjie Huang, Yuhui Cao, Weichong Yin, Yongfeng Chen, Yin Zhang, Shikun Feng, Yu Sun, Hao Tian, Hua Wu, and Haifeng Wang. ERNIE-Layout: Layout Knowledge Enhanced Pre-training for Visually-rich Document Understanding. In Findings of EMNLP, pages 3744–3756, 2022. 1" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 56, + 277, + 295, + 342 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 277, + 295, + 342 + ], + "spans": [ + { + "bbox": [ + 56, + 277, + 295, + 342 + ], + "type": "text", + "content": "[53] Vincent Perot, Kai Kang, Florian Luisier, Guolong Su, Xiaoyu Sun, Ramya Sree Boppana, Zilong Wang, Zifeng Wang, Jiaqi Mu, Hao Zhang, Chen-Yu Lee, and Nan Hua. LMDX: Language Model-based Document Information Extraction and Localization. In Findings of ACL, pages 15140-15168, 2024. 3, 6" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 56, + 343, + 295, + 398 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 343, + 295, + 398 + ], + "spans": [ + { + "bbox": [ + 56, + 343, + 295, + 398 + ], + "type": "text", + "content": "[54] Birgit Pfitzmann, Christoph Auer, Michele Dolfi, Ahmed S. Nassar, and Peter W. J. Staar. DocLayNet: A Large Human-Annotated Dataset for Document-Layout Segmentation. In Proc. ACM SIGKDD Int. Conf. Knowledge discovery & data mining, pages 3743–3751, 2022. 5" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 56, + 399, + 295, + 453 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 399, + 295, + 453 + ], + "spans": [ + { + "bbox": [ + 56, + 399, + 295, + 453 + ], + "type": "text", + "content": "[55] Rafat Powalski, Łukasz Borchmann, Dawid Jurkiewicz, Tomasz Dwojak, Michal Pietruszka, and Gabriela Pałka. Going Full-TILT Boogie on Document Understanding with Text-Image-Layout Transformer. In Proc. ICDAR, pages 732–747, 2021. 1" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 56, + 456, + 295, + 520 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 456, + 295, + 520 + ], + "spans": [ + { + "bbox": [ + 56, + 456, + 295, + 520 + ], + "type": "text", + "content": "[56] Stepán Simsa, Milan Sulc, Michal Uricár, Yash Patel, Ahmed Hamdi, Matej Kocián, Matyas Skalický, Jiri Matas, Antoine Doucet, Mickaël Coustaty, and Dimosthenis Karatzas. DocILE Benchmark for Document Information Localization and Extraction. In Proc. ICDAR, pages 147-166, 2023. 5" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 56, + 522, + 295, + 578 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 522, + 295, + 578 + ], + "spans": [ + { + "bbox": [ + 56, + 522, + 295, + 578 + ], + "type": "text", + "content": "[57] Tomasz Stanislawek, Filip Gralinski, Anna Wróblewska, Dawid Lipinski, Agnieszka Kaliska, Paulina Rosalska, Bartosz Topolski, and Przemyslaw Biecek. Kleister: Key Information Extraction Datasets Involving Long Documents with Complex Layouts. In Proc. ICDAR, pages 564-579, 2021. 5" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 56, + 578, + 295, + 611 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 578, + 295, + 611 + ], + "spans": [ + { + "bbox": [ + 56, + 578, + 295, + 611 + ], + "type": "text", + "content": "[58] Stacey Svetlichnaya. DeepForm: Understand Structured Documents at Scale. https://wandb.ai/stacey/deepform_v1, 2020. Accessed: 2024-08-14. 5" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 56, + 613, + 295, + 645 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 613, + 295, + 645 + ], + "spans": [ + { + "bbox": [ + 56, + 613, + 295, + 645 + ], + "type": "text", + "content": "[59] Ryota Tanaka, Kyosuke Nishida, and Sen Yoshida. VisualMRC: Machine Reading Comprehension on Document Images. In Proc. AAAI, pages 13878-13888, 2021. 5" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 56, + 647, + 295, + 690 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 647, + 295, + 690 + ], + "spans": [ + { + "bbox": [ + 56, + 647, + 295, + 690 + ], + "type": "text", + "content": "[60] Ryota Tanaka, Taichi Iki, Kyosuke Nishida, Kuniko Saito, and Jun Suzuki. InstructDoc: A Dataset for Zero-shot Generalization of Visual Document Understanding with Instructions. In Proc. AAAI, pages 19071-19079, 2024. 3" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 56, + 692, + 295, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 692, + 295, + 713 + ], + "spans": [ + { + "bbox": [ + 56, + 692, + 295, + 713 + ], + "type": "text", + "content": "[61] Jingqun Tang, Chunhui Lin, Zhen Zhao, Shu Wei, Binghong Wu, Qi Liu, Hao Feng, Yang Li, Siqi Wang, Lei Liao, Wei" + } + ] + } + ], + "index": 13 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 316, + 73, + 553, + 713 + ], + "type": "list", + "angle": 0, + "index": 25, + "blocks": [ + { + "bbox": [ + 333, + 73, + 553, + 116 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 333, + 73, + 553, + 116 + ], + "spans": [ + { + "bbox": [ + 333, + 73, + 553, + 116 + ], + "type": "text", + "content": "Shi, Yuliang Liu, Hao Liu, Yuan Xie, Xiang Bai, and Can Huang. TextSquare: Scaling up Text-Centric Visual Instruction Tuning. arXiv: Comp. Res. Repository, abs/2404.12803, 2024. 1, 2, 7" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 316, + 118, + 553, + 171 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 118, + 553, + 171 + ], + "spans": [ + { + "bbox": [ + 316, + 118, + 553, + 171 + ], + "type": "text", + "content": "[62] Zineng Tang, Ziyi Yang, Guoxin Wang, Yuwei Fang, Yang Liu, Chenguang Zhu, Michael Zeng, Cha Zhang, and Mohit Bansal. Unifying Vision, Text, and Layout for Universal Document Processing. In Proc. CVPR, pages 19254-19264, 2023. 2" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 316, + 173, + 553, + 194 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 173, + 553, + 194 + ], + "spans": [ + { + "bbox": [ + 316, + 173, + 553, + 194 + ], + "type": "text", + "content": "[63] Qwen Team. QwQ: Reflect Deeply on the Boundaries of the Unknown, 2024. 2" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 316, + 196, + 553, + 435 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 196, + 553, + 435 + ], + "spans": [ + { + "bbox": [ + 316, + 196, + 553, + 435 + ], + "type": "text", + "content": "[64] Hugo Touvron, Louis Martin, Kevin Stone, Peter Albert, Amjad Almahairi, Yasmine Babaei, Nikolay Bashlykov, Soumya Batra, Prajjwal Bhargava, Shruti Bhosale, Dan Bikel, Lukas Blecher, Cristian Canton-Ferrer, Moya Chen, Guillem Cucurull, David Esiobu, Jude Fernandes, Jeremy Fu, Wenyin Fu, Brian Fuller, Cynthia Gao, Vedanuj Goswami, Naman Goyal, Anthony Hartshorn, Saghar Hosseini, Rui Hou, Hakan Inan, Marcin Kardas, Viktor Kerkez, Madian Khabsa, Isabel Kloumann, Artem Korenev, Punit Singh Koura, Marie-Anne Lachaux, Thibaut Lavril, Jenya Lee, Diana Liskovich, Yinghai Lu, Yuning Mao, Xavier Martinet, Todor Mihaylov, Pushkar Mishra, Igor Molybog, Yixin Nie, Andrew Poulton, Jeremy Reizenstein, Rashi Rungta, Kalyan Saladi, Alan Schelten, Ruan Silva, Eric Michael Smith, Ranjan Subramanian, Xiaoqing Ellen Tan, Binh Tang, Ross Taylor, Adina Williams, Jian Xiang Kuan, Puxin Xu, Zheng Yan, Iliyan Zarov, Yuchen Zhang, Angela Fan, Melanie Kambadur, Sharan Narang, Aurélien Rodriguez, Robert Stojnic, Sergey Edunov, and Thomas Scialom. Llama 2: Open Foundation and Fine-Tuned Chat Models. arXiv: Comp. Res. Repository, abs/2307.09288, 2023. 6" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 316, + 437, + 553, + 480 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 437, + 553, + 480 + ], + "spans": [ + { + "bbox": [ + 316, + 437, + 553, + 480 + ], + "type": "text", + "content": "[65] Yi Tu, Ya Guo, Huan Chen, and Jinyang Tang. LayoutMask: Enhance Text-Layout Interaction in Multi-modal Pretraining for Document Understanding. In Proc. ACL, pages 15200–15212, 2023. 1" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 316, + 482, + 553, + 536 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 482, + 553, + 536 + ], + "spans": [ + { + "bbox": [ + 316, + 482, + 553, + 536 + ], + "type": "text", + "content": "[66] Jianqiang Wan, Sibo Song, Wenwen Yu, Yuliang Liu, Wenqing Cheng, Fei Huang, Xiang Bai, Cong Yao, and Zhibo Yang. OmniParser: A Unified Framework for Text Spotting Key Information Extraction and Table Recognition. In Proc. CVPR, pages 15641-15653, 2024. 2" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 316, + 537, + 553, + 591 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 537, + 553, + 591 + ], + "spans": [ + { + "bbox": [ + 316, + 537, + 553, + 591 + ], + "type": "text", + "content": "[67] Dongsheng Wang, Natraj Raman, Mathieu Sibue, Zhiqiang Ma, Petr Babkin, Simerjot Kaur, Yulong Pei, Armineh Nourbakhsh, and Xiaomo Liu. DocLLM: A Layout-Aware Generative Language Model for Multimodal Document Understanding. In Proc. ACL, pages 8529-8548, 2024. 3, 6, 7" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 316, + 593, + 553, + 635 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 593, + 553, + 635 + ], + "spans": [ + { + "bbox": [ + 316, + 593, + 553, + 635 + ], + "type": "text", + "content": "[68] Jiapeng Wang, Lianwen Jin, and Kai Ding. LiLT: A Simple yet Effective Language-Independent Layout Transformer for Structured Document Understanding. In Proc. ACL, pages 7747–7757, 2022. 1" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 316, + 636, + 553, + 690 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 636, + 553, + 690 + ], + "spans": [ + { + "bbox": [ + 316, + 636, + 553, + 690 + ], + "type": "text", + "content": "[69] Jiapeng Wang, Zening Lin, Dayi Huang, Longfei Xiong, and Lianwen Jin. LiLTv2: Language-substitutable layout-image transformer for visual information extraction. ACM Transactions on Multimedia Computing, Communications and Applications, 21(3):1-27, 2025. 1" + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 316, + 692, + 553, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 692, + 553, + 713 + ], + "spans": [ + { + "bbox": [ + 316, + 692, + 553, + 713 + ], + "type": "text", + "content": "[70] Wenjin Wang, Yunhao Li, Yixin Ou, and Yin Zhang. Layout and Task Aware Instruction Prompt for Zero-shot Document" + } + ] + } + ], + "index": 24 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "text", + "content": "4048" + } + ] + } + ], + "index": 26 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 10 + }, + { + "para_blocks": [ + { + "bbox": [ + 56, + 72, + 295, + 675 + ], + "type": "list", + "angle": 0, + "index": 12, + "blocks": [ + { + "bbox": [ + 76, + 72, + 294, + 95 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 76, + 72, + 294, + 95 + ], + "spans": [ + { + "bbox": [ + 76, + 72, + 294, + 95 + ], + "type": "text", + "content": "Image Question Answering. arXiv: Comp. Res. Repository, abs/2306.00526, 2023. 1, 3" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 56, + 96, + 295, + 150 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 96, + 295, + 150 + ], + "spans": [ + { + "bbox": [ + 56, + 96, + 295, + 150 + ], + "type": "text", + "content": "[71] Jason Wei, Xuezhi Wang, Dale Schuurmans, Maarten Bosma, brian richter, Fei Xia, Ed Chi, Quoc V Le, and Denny Zhou. Chain-of-Thought Prompting Elicits Reasoning in Large Language Models. Proc. NIPS, 35:24824-24837, 2022. 2" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 56, + 152, + 294, + 207 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 152, + 294, + 207 + ], + "spans": [ + { + "bbox": [ + 56, + 152, + 294, + 207 + ], + "type": "text", + "content": "[72] Yiheng Xu, Minghao Li, Lei Cui, Shaohan Huang, Furu Wei, and Ming Zhou. LayoutLM: Pre-training of Text and Layout for Document Image Understanding. In Proc. ACM SIGKDD Int. Conf. Knowledge discovery & data mining, page 1192-1200, 2020. 1" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 56, + 208, + 294, + 262 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 208, + 294, + 262 + ], + "spans": [ + { + "bbox": [ + 56, + 208, + 294, + 262 + ], + "type": "text", + "content": "[73] Yang Xu, Yiheng Xu, Tengchao Lv, Lei Cui, Furu Wei, Guoxin Wang, Yijuan Lu, Dinei Florencio, Cha Zhang, Wanxiang Che, Min Zhang, and Lidong Zhou. LayoutLMv2: Multi-modal Pre-training for Visually-rich Document Understanding. In Proc. ACL, pages 2579-2591, 2021. 1" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 56, + 264, + 294, + 306 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 264, + 294, + 306 + ], + "spans": [ + { + "bbox": [ + 56, + 264, + 294, + 306 + ], + "type": "text", + "content": "[74] Hao Yan, Yuliang Liu, Lianwen Jin, and Xiang Bai. The development, application, and future of LLM similar to ChatGPT. Journal of Image and Graphics, 28(9):2749-2762, 2023. 1" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 56, + 308, + 294, + 373 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 308, + 294, + 373 + ], + "spans": [ + { + "bbox": [ + 56, + 308, + 294, + 373 + ], + "type": "text", + "content": "[75] Jiabo Ye, Anwen Hu, Haiyang Xu, Qinghao Ye, Ming Yan, Yuhao Dan, Chenlin Zhao, Guohai Xu, Chenliang Li, Junfeng Tian, Qian Qi, Ji Zhang, and Fei Huang. mPLUGDocOwl: Modularized Multimodal Large Language Model for Document Understanding. arXiv: Comp. Res. Repository, abs/2307.02499, 2023. 7" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 56, + 375, + 295, + 441 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 375, + 295, + 441 + ], + "spans": [ + { + "bbox": [ + 56, + 375, + 295, + 441 + ], + "type": "text", + "content": "[76] Jiabo Ye, Anwen Hu, Haiyang Xu, Qinghao Ye, Ming Yan, Guohai Xu, Chenliang Li, Junfeng Tian, Qi Qian, Ji Zhang, Qin Jin, Liang He, Xin Lin, and Fei Huang. UReader: Universal OCR-free Visually-situated Language Understanding with Multimodal Large Language Model. In Findings of EMNLP, pages 2841-2858, 2023. 1, 2, 7" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 56, + 442, + 294, + 495 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 442, + 294, + 495 + ], + "spans": [ + { + "bbox": [ + 56, + 442, + 294, + 495 + ], + "type": "text", + "content": "[77] Chong Zhang, Ya Guo, Yi Tu, Huan Chen, Jinyang Tang, Huijia Zhu, Qi Zhang, and Tao Gui. Reading Order Matters: Information Extraction from Visually-rich Documents by Token Path Prediction. In Proc. EMNLP, pages 13716-13730, 2023. 1" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 56, + 498, + 294, + 551 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 498, + 294, + 551 + ], + "spans": [ + { + "bbox": [ + 56, + 498, + 294, + 551 + ], + "type": "text", + "content": "[78] Jiaxin Zhang, Wentao Yang, Songxuan Lai, Zecheng Xie, and Lianwen Jin. DocKylin: A Large Multimodal Model for Visual Document Understanding with Efficient Visual Slimming. arXiv: Comp. Res. Repository, abs/2406.19101, 2024. 7" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 56, + 553, + 294, + 608 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 553, + 294, + 608 + ], + "spans": [ + { + "bbox": [ + 56, + 553, + 294, + 608 + ], + "type": "text", + "content": "[79] Yanzhe Zhang, Ruiyi Zhang, Jiuming Gu, Yufan Zhou, Nedim Lipka, Diyi Yang, and Tong Sun. LLaVAR: Enhanced Visual Instruction Tuning for Text-Rich Image Understanding. arXiv: Comp. Res. Repository, abs/2306.17107, 2023. 1, 2, 7" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 56, + 609, + 294, + 642 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 609, + 294, + 642 + ], + "spans": [ + { + "bbox": [ + 56, + 609, + 294, + 642 + ], + "type": "text", + "content": "[80] Xu Zhong, Jianbin Tang, and Antonio Jimeno-Yepes. Pub-LayNet: Largest Dataset Ever for Document Layout Analysis. In Proc. ICDAR, pages 1015-1022, 2019. 5" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 56, + 643, + 294, + 675 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 643, + 294, + 675 + ], + "spans": [ + { + "bbox": [ + 56, + 643, + 294, + 675 + ], + "type": "text", + "content": "[81] Xu Zhong, Elaheh ShafieiBavani, and Antonio JimenoYepes. Image-Based Table Recognition: Data, Model, and Evaluation. Proc. ECCV, 12366:564-580, 2020. 5" + } + ] + } + ], + "index": 11 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "text", + "content": "4049" + } + ] + } + ], + "index": 13 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 11 + } + ], + "_backend": "vlm", + "_version_name": "2.6.4" +} \ No newline at end of file diff --git a/2025/DocSAM_ Unified Document Image Segmentation via Query Decomposition and Heterogeneous Mixed Learning/cd34a0c4-c9ec-4e5a-a2a8-007f0052f160_content_list.json b/2025/DocSAM_ Unified Document Image Segmentation via Query Decomposition and Heterogeneous Mixed Learning/cd34a0c4-c9ec-4e5a-a2a8-007f0052f160_content_list.json new file mode 100644 index 0000000000000000000000000000000000000000..89d1f085402233349299172d0d27306f482313d7 --- /dev/null +++ b/2025/DocSAM_ Unified Document Image Segmentation via Query Decomposition and Heterogeneous Mixed Learning/cd34a0c4-c9ec-4e5a-a2a8-007f0052f160_content_list.json @@ -0,0 +1,1630 @@ +[ + { + "type": "text", + "text": "DocSAM: Unified Document Image Segmentation via Query Decomposition and Heterogeneous Mixed Learning", + "text_level": 1, + "bbox": [ + 94, + 128, + 901, + 176 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Xiao-Hui Li $^{1}$ , Fei Yin $^{1}$ , Cheng-Lin Liu $^{1,2}$ $^{1}$ MAIS, Institute of Automation of Chinese Academy of Sciences, Beijing, 100190, China \n $^{2}$ School of Artificial Intelligence, University of Chinese Academy of Sciences, Beijing, 100049, China \n{xiaohui.li, fyin, liucl}@nlpr.ia.ac.cn", + "bbox": [ + 143, + 202, + 854, + 292 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Abstract", + "text_level": 1, + "bbox": [ + 246, + 325, + 326, + 343 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Document image segmentation is crucial for document analysis and recognition but remains challenging due to the diversity of document formats and segmentation tasks. Existing methods often address these tasks separately, resulting in limited generalization and resource wastage. This paper introduces DocSAM, a transformer-based unified framework designed for various document image segmentation tasks, such as document layout analysis, multi-granularity text segmentation, and table structure recognition, by modelling these tasks as a combination of instance and semantic segmentation. Specifically, DocSAM employs Sentence-BERT to map category names from each dataset into semantic queries that match the dimensionality of instance queries. These two sets of queries interact through an attention mechanism and are cross-attended with image features to predict instance and semantic segmentation masks. Instance categories are predicted by computing the dot product between instance and semantic queries, followed by softmax normalization of scores. Consequently, DocSAM can be jointly trained on heterogeneous datasets, enhancing robustness and generalization while reducing computational and storage resources. Comprehensive evaluations show that DocSAM surpasses existing methods in accuracy, efficiency, and adaptability, highlighting its potential for advancing document image understanding and segmentation across various applications. Codes are available at https://github.com/xhli-git/DocSAM.", + "bbox": [ + 91, + 359, + 483, + 767 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "1. Introduction", + "text_level": 1, + "bbox": [ + 91, + 797, + 220, + 814 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Document image segmentation (DIS) is a fundamental task in the field of document analysis and recognition (DAR) [42], serving as a cornerstone for downstream applications such as text recognition, information extraction (IE), and document visual question answering (DocVQA). Despite", + "bbox": [ + 89, + 824, + 483, + 902 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "its importance, DIS faces significant challenges due to the wide diversity of document types, page layouts, content annotations, and structural complexities, see Fig. 1. Existing approaches often address specific aspects of DIS separately, such as layout analysis, text detection, and table structure recognition, leading to specialized and fragmented solutions tailored to particular applications. This fragmentation not only impedes the performance of individual tasks but also results in redundant computational and storage overheads, making them inefficient for large-scale deployment.", + "bbox": [ + 511, + 327, + 906, + 479 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "To address the aforementioned challenges, this paper introduces DocSAM (Document Segment Anything Model), a transformer-based unified framework designed to simultaneously handle various document image segmentation tasks, thereby eliminating the need for separate models and enhancing overall efficiency. As illustrated in Fig. 2, DocSAM comprises four primary modules: the Vision Backbone, the Deformable Encoder, Sentence-BERT [64], and the Hybrid Query Decoder (HQD). Given a document image and desired instance or semantic class names in natural text format, DocSAM first extracts multi-scale image features using the Vision Backbone. These features are then refined by the Deformable Encoder, which includes several deformable attention layers [108]. Class names are fed into Sentence-BERT and mapped to semantic queries. Subsequently, both semantic queries and learnable instance queries pass together through the HQD, where they interact to jointly perform semantic and instance segmentation.", + "bbox": [ + 511, + 484, + 908, + 758 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Inside each HQD layer (see Fig. 2), semantic and instance queries are concatenated and passed through a multi-head self-attention layer followed by a feed-forward layer for information exchange. These queries are then separately cross-attended with multi-scale image features in a coarse-to-fine manner using two multi-scale decoders, each with $L = 4$ layers. They further interact via another multi-head self-attention and feed-forward layer. The resulting semantic and instance queries, along with fused multi-scale image", + "bbox": [ + 511, + 763, + 908, + 902 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "CVF", + "bbox": [ + 106, + 2, + 181, + 42 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore.", + "bbox": [ + 236, + 0, + 810, + 46 + ], + "page_idx": 0 + }, + { + "type": "page_number", + "text": "15021", + "bbox": [ + 480, + 944, + 517, + 955 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/3a6a1f1783c34c1b3b6f4d53e438720afacc5ce3c93f7798170108fa8066c848.jpg", + "image_caption": [ + "Figure 1. Examples of various segmentation tasks on heterogeneous document datasets." + ], + "image_footnote": [], + "bbox": [ + 96, + 87, + 236, + 358 + ], + "page_idx": 1 + }, + { + "type": "image", + "img_path": "images/8ac8b9cb48c302516b0d45a95ad8117108a69b452aef87003aaf121d116dcf48.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 236, + 87, + 506, + 358 + ], + "page_idx": 1 + }, + { + "type": "image", + "img_path": "images/b734259ed50c26b5a4d03e1d2dcdd8b88de0700ea8d3217e81c277437206cf3c.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 509, + 88, + 751, + 358 + ], + "page_idx": 1 + }, + { + "type": "image", + "img_path": "images/dbde6a541b34d3b73dd2631149976a9949eebc85db692c1e342ced8a6e53ba3c.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 754, + 88, + 901, + 358 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "features, are forwarded to the Mask Predictor, Class Predictor, and BBox Predictor for semantic mask segmentation, instance mask segmentation, category classification, and bounding box regression, respectively. We stack $K$ HQD layers for more refined predictions.", + "bbox": [ + 89, + 409, + 482, + 484 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "This design ensures that DocSAM can effectively manage the heterogeneity of document types, annotation formats, and segmentation tasks while maintaining high efficiency and accuracy. Extensive experiments and evaluations on various datasets demonstrate that DocSAM surpasses existing methods in accuracy, efficiency, and adaptability. Our results highlight DocSAM's potential as a powerful tool for advancing document image segmentation and understanding, with applications spanning from modern and historical document layout analysis to table structure decomposition, handwritten and scene text detection, and beyond. Our contributions are summarized as follows:", + "bbox": [ + 89, + 491, + 482, + 674 + ], + "page_idx": 1 + }, + { + "type": "list", + "sub_type": "text", + "list_items": [ + "- We introduce DocSAM, a unified solution for diverse document image segmentation tasks such as layout analysis, multi-grained text segmentation, and table structure decomposition, reducing the need for specialized models and enhancing overall efficiency;", + "- By training on various tasks and datasets, DocSAM improves robustness and generalization, making it highly effective in handling varied document types and structures;", + "- Compared to specialized models, DocSAM significantly reduces computational and storage requirements, making it more practical for large-scale deployment;", + "- Extensive experiments on various datasets show that Doc-SAM outperforms current methods in terms of accuracy, efficiency, and adaptability." + ], + "bbox": [ + 89, + 688, + 482, + 901 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "2. Related Works", + "text_level": 1, + "bbox": [ + 513, + 407, + 661, + 422 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "2.1. DIS Tasks and Datasets", + "text_level": 1, + "bbox": [ + 513, + 440, + 728, + 455 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Depending on specific application scenarios, DIS involves various sub-tasks including Document Layout Analysis (DLA), Multi-Granularity Text Detection (MGTD), and Table Structure Recognition (TSR). DLA aims at identifying and categorizing page regions including text blocks, figures and tables [9, 16, 36, 60, 106]. This foundational step provides a structured overview of the document's layout, enabling more precise processing in subsequent tasks. MGTD focuses on detecting and segmenting text at various granularities, from paragraphs down to individual lines and words [11, 43, 53, 93, 95]. MGTD is a prerequisite for accurate Optical Character Recognition (OCR) tasks. TSR specifically aims to extract the structural of tables, including rows, columns and cells [17, 35, 38, 72, 107]. By decomposing tables into substructures, TSR facilitates the extraction and analysis of tabular information from documents.", + "bbox": [ + 511, + 470, + 906, + 712 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Along with these tasks, plenty of datasets have been accumulated after decades of research, see Tab. 1. These datasets exhibit great diversity and heterogeneity in data sources, document types, annotation formats, writing languages, category sets and many other aspects. For example, PubLayNet [106] contains born-digital English PDF documents with region-level annotations; SCUT-CAB [8] and MTHv2 [53] contains scanned historical Chinese documents with region, line and char-level annotations; SCUT-HCCDoc [95] contains handwritten documents with line-level annotations; CTW1500 [43] and Total-Text [11] contain natural scene images with texts of arbitrary shapes.", + "bbox": [ + 511, + 719, + 908, + 902 + ], + "page_idx": 1 + }, + { + "type": "page_number", + "text": "15022", + "bbox": [ + 480, + 944, + 517, + 955 + ], + "page_idx": 1 + }, + { + "type": "table", + "img_path": "images/3cfa877871442b1ba2bc5bc2d4be92208279056d90ad3a349d8b32a36f83816a.jpg", + "table_caption": [], + "table_footnote": [], + "table_body": "
TaskDataset
Document Layout AnalysisBaDLAD [71], CDLA [3], D4LA [13], DocBank [36], DocLayNet [60], ICDAR2017-POD [16], IIIT-AR-13K [55], M6Doc [9], PubLayNet [106], RanLayNet [2]
Ancient and Handwritten Document SegmentationCASIA-AHCDB [85], CHDAC-2022 [31], ICDAR2019-HDRC [68], SCUT-CAB [8], MTHv2 [53], HJDataset [69], CASIA-HWDB [41], SCUT-HCCDoc [95]
Table Structure RecognitionFinTabNet [104], ICDAR2013 [18], ICDAR2017-POD [16, 38], ICDAR2019-cTDr [17, 38], NTable [109], PubTables-1M [72], PubTabNet [107], STDW [19], TableBank [35], TNCR [1], WTW [47]
Scene Text DetectionCASIA-10k [22], COCO-Text [76], CTW1500 [43], CTW-Public [92], HUST-TR400 [87], ICDAR2015 [28], ICDAR2017-RCTW [70], ICDAR2017-MLT [56], ICDAR2019-ArT [12], ICDAR2019-LSVT [74], ICDAR2019-MLT [57], ICDAR2019-ReCTS [100], ICDAR2023-HierText [49], ICDAR2023-ReST [91], ICPR2018-MTWI [21], MSRA-TD500 [86], ShopSign [94], Total-Text [11], USTB-SV1K [90]
", + "bbox": [ + 102, + 88, + 470, + 330 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Table 1. Datasets involved in DocSAM.", + "bbox": [ + 166, + 340, + 403, + 353 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "2.2. Deep Learning for DIS", + "text_level": 1, + "bbox": [ + 89, + 387, + 303, + 404 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Existing deep learning based DIS methods basically focus on specific sub-tasks and datasets. Generally speaking, they usually transform various DIS tasks into general object detection or image segmentation problems and make some modifications to general object detection [20, 62, 63, 65] and image segmentation methods [5, 46, 66] to make them more suitable for the tasks and datasets at hand. Some other works treat documents as hierarchical graph structures and adopt graph models like GNN [82] and CRF [32] for the task of layout analysis [37, 52], table structure recognition [10, 38], and text detection [48, 101]. Though more flexible, these methods usually suffer from complicated pre/post-processing steps and are more susceptible to intermediate errors. There are also some multi-modal based methods that combine visual and textual features like LayoutLMv3 [27], DiT [34], and VGT [13]. These methods improve the performance and generalization by pre-training on large-scale unsupervised documents to align text and visual features, but are often slower due to the complexity of architectures.", + "bbox": [ + 89, + 412, + 483, + 700 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "With the prosperity of large language models (LLMs) [102], many large document models are proposed such as UDOP [75], UniDoc [15], DocPedia [14], DocLLM [80], TextMonkey [44], mPLUG-DocOwl [23, 88], etc. Though promising results can be achieved for the DocVQA task, lacking fine-grained intermediate outputs like text locations and page layouts still greatly limits the interpretability and generalization of these models. As compensation, recently some LLM-free unified models are proposed for low-level document processing tasks such as UPOCR [59], DocRes [98], OmniParser [78] and DAT [79]. These works unify several similar or related tasks into unified models through multi-task learning, but at the cost of significant increment", + "bbox": [ + 89, + 704, + 483, + 900 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "of model complexity and calculating overhead, prohibiting them from generalizing to to more tasks and datasets.", + "bbox": [ + 511, + 90, + 906, + 122 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "2.3. Transformer-based Detection&Segmentation", + "text_level": 1, + "bbox": [ + 511, + 133, + 893, + 148 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Following the pioneer work of DETR [4], many Transformer-based objection methods have been proposed in recent years, including Deformable DETR [108], DNDETR [33], DINO [96], Sparse R-CNN [73], etc. These methods share the same idea with DETR that rely on learnable queries and bipartite matching for object decoding, but make different modifications to improve the accuracy and convergence speed, such as bringing in deformable attention and denoising training, or assigning specific spatial meanings to the queries.", + "bbox": [ + 511, + 155, + 906, + 306 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Besides object detection, Transformer also shows great potential in image segmentation [6, 7, 24, 25, 30, 83, 97, 103]. Among them the most related works to this paper are SAM [30] and Mask2former [7]. Inspired by SAM [30] which uses natural language prompts to guide image segmentation, in this paper we propose to embed the class names of each dataset into semantic queries and transform various document segmentation tasks into a combination of instance segmentation and semantic segmentation. The semantic queries not only serve as prompts guiding the model in identifying specific types of regions, but also function as class prototypes that instance queries depend on for classification. Since DIS relies on high resolution image features, we build our DocSAM from Mask2former [7] which adopt Swin-Transformer [45] and deformable attention [108] as the vision encoder. Though the vision encoder of DocSAM is inherited from Mask2former, the decoder is drastically redesigned to be up to the task of general document image segmentation effectively.", + "bbox": [ + 511, + 308, + 908, + 595 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3. DocSAM", + "text_level": 1, + "bbox": [ + 511, + 611, + 614, + 626 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3.1. Preliminaries", + "text_level": 1, + "bbox": [ + 511, + 636, + 653, + 651 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Before introducing the proposed DocSAM, we first explore the key attributes of an ideal all-in-one DIS model and why current methods fall short of this goal. We assert that an exemplary all-in-one DIS model should possess the following attributes: it should have the versatility to convert diverse DIS tasks into a unified framework; it should be adaptable to training on heterogeneous datasets, accommodating diverse annotations without restrictions; and it should maintain the capacity for continual and incremental learning. Current methods are typically designed for specific DIS tasks and datasets, and most models become static after training, unable to efficiently incorporate new data, limiting their versatility and adaptability. Specifically, existing DIS methods rely on fully connected (FC) and Softmax layers to predict region classes, with FC's parameters predefined for specific tasks and datasets, making generalization difficult.", + "bbox": [ + 511, + 657, + 908, + 902 + ], + "page_idx": 2 + }, + { + "type": "page_number", + "text": "15023", + "bbox": [ + 480, + 944, + 517, + 955 + ], + "page_idx": 2 + }, + { + "type": "image", + "img_path": "images/860576fad4d4e862ce1333467fde1ff2d896d31e117085aa79f416ed0e08ac58.jpg", + "image_caption": [ + "Figure 2. Network structure of the proposed DocSAM. DocSAM unify various document image segmentation tasks into one single model through instance and semantic query decomposition and interaction. Skip connections and norm layers are omitted for simplicity." + ], + "image_footnote": [], + "bbox": [ + 112, + 87, + 883, + 333 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "To overcome the above limitations and achieve the aforementioned criteria, the proposed DocSAM makes two significant improvements compared to existing methods. First, it transforms various DIS tasks into a unified paradigm of mask-based instance segmentation and semantic segmentation. Second, it embeds class names into semantic queries, which not only serve as prompts to guide the model in identifying specific types of regions to segment but also function as class prototypes that instance queries depend on for classification. The rest of this chapter presents the details of our proposed DocSAM.", + "bbox": [ + 89, + 398, + 483, + 564 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "3.2. Vision Encoder", + "text_level": 1, + "bbox": [ + 89, + 574, + 243, + 589 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Different DIS tasks may focus on contents of different scales, from large objects like paragraphs and figures spanning entire pages to tiny objects like chars and words covering only a few hundreds of pixels. Therefore, high-resolution multi-scale image features are an essential requirement for a unified DIS model. The vision encoder of DocSAM is adapted from Mask2Former [7], which includes a Swin-Transformer [45] as the vision backbone and deformable attention [108] for feature refinement. Additionally, we use another FPN [40] to fuse multi-scale image features $X_{I} = [X_{I}^{l}\\in \\mathbb{R}^{H_{l}W_{l}\\times C},l\\in \\{1,2,3,4\\}]$ into a single mask feature $X_{M}\\in \\mathbb{R}^{HW\\times C}$ , which is used for subsequent semantic segmentation and instance segmentation. Here, $X_{I}^{l}$ is image feature of level $l$ , $H_{l}$ and $W_{l}$ are the spatial resolution of level $l$ , $C$ is number of feature channels.", + "bbox": [ + 89, + 595, + 483, + 823 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "3.3. Query Embedding", + "text_level": 1, + "bbox": [ + 89, + 832, + 269, + 849 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "The instance queries $Q_{I} \\in \\mathbb{R}^{N \\times C}$ of DocSAM is standard learnable queries, while the semantic queries $Q_{S} \\in \\mathbb{R}^{M \\times C}$ are embedded from class names using the Sentence-BERT", + "bbox": [ + 89, + 854, + 483, + 900 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "[64]. Here $N$ is a predefined instance query number that remains the same across all tasks and datasets, while $M$ is the semantic query number that may change depending on the class number of each dataset, and $C$ is feature dimension. $Q_{I}$ and $Q_{S}$ go together through the following Hybrid Query Decoder for feature decoding and cooperate with each other for semantic segmentation and instance segmentation.", + "bbox": [ + 511, + 398, + 906, + 503 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "3.4. Hybrid Query Decoder", + "text_level": 1, + "bbox": [ + 511, + 513, + 725, + 530 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Inside each HQD layer, see Fig. 2, we first concatenate $Q_{S}$ and $Q_{I}$ along the length dimension and send them into a multi-head self-attention layer (MHSA) followed by a feed forward layer (FFN). This step facilitates information exchange between $Q_{S}$ and $Q_{I}$ , allowing them to attend to each other for query fusion. Next, they are separately crossattended with the multi-scale image features $X_{I}$ in a coarse-to-fine manner by two Multi-Scale Decoders (MSD) each containing $L$ layers. Here, $L = 4$ stands for the number of feature scales. Each MSD layer consists of two MHSA layers, one multi-head cross-attention layer (MHCA) and one FFN layer. Following Mask2Former [7], we also use masked attention in MHCA, where the attention masks are derived from the predicted instance and semantic masks in the previous HQD layer. After that, $Q_{S}$ and $Q_{I}$ further interact with each other through another MHSA and FFN layer. We stack $K$ HQD layers for more refined predictions.", + "bbox": [ + 511, + 536, + 906, + 792 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "3.5. Prediction Head", + "text_level": 1, + "bbox": [ + 511, + 801, + 674, + 816 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "The output $Q_{S}$ and $Q_{I}$ from each HQD layer along with the mask feature $X_{M}$ are sent to the Mask Predictor, Class Predictor and BBox Predictor for semantic mask segmentation, instance mask segmentation, instance category classification and instance bounding box regression, respectively.", + "bbox": [ + 511, + 824, + 905, + 900 + ], + "page_idx": 3 + }, + { + "type": "page_number", + "text": "15024", + "bbox": [ + 480, + 944, + 517, + 955 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "For predicting semantic and instance masks, $Q_{S}$ and $Q_{I}$ are multiplied with $X_{M}$ as:", + "bbox": [ + 89, + 90, + 482, + 121 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\nM _ {S} = \\sigma \\left(Q _ {S} \\times X _ {M} ^ {T}\\right), \\tag {1}\n$$\n", + "text_format": "latex", + "bbox": [ + 212, + 132, + 482, + 150 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "and", + "bbox": [ + 89, + 162, + 120, + 175 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\nM _ {I} = \\sigma \\left(Q _ {I} \\times X _ {M} ^ {T}\\right), \\tag {2}\n$$\n", + "text_format": "latex", + "bbox": [ + 214, + 174, + 480, + 193 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "where $M_S \\in \\mathbb{R}^{M \\times HW}$ and $M_I \\in \\mathbb{R}^{N \\times HW}$ are predicted semantic and instance masks, $\\sigma$ is Sigmoid function, $T$ means matrix transposition, and $\\times$ stands for matrix multiplication. Similarly, for predicting instance classes, $Q_I$ is multiplied with $Q_S$ as:", + "bbox": [ + 89, + 196, + 483, + 275 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\nY _ {I} = \\operatorname {S o f t m a x} \\left(Q _ {I} \\times Q _ {S} ^ {T}\\right), \\tag {3}\n$$\n", + "text_format": "latex", + "bbox": [ + 196, + 286, + 482, + 304 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "where $Y_{I} \\in \\mathbb{R}^{N \\times M}$ is predicted class probabilities of instances, softmax is Softmax function along the second dimension of $Y_{I}$ , $T$ means matrix transposition, and $\\times$ stands for matrix multiplication.", + "bbox": [ + 89, + 314, + 482, + 375 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Since Eq. (1), Eq. (2), Eq. (3) are all based on matrix multiplication, they are actually calculating the similarities between $Q_{S}$ , $Q_{I}$ and $X_{M}$ . So we can also regard the $Q_{S}$ , $Q_{I}$ as instance and semantic prototypes. Through the above semantic query embedding and prototype-based instance classification, we transform the original close-set classifier into an open-set classifier, thus benefiting the construction of unified all-in-one DIS model.", + "bbox": [ + 89, + 376, + 483, + 494 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Besides mask segmentation, DocSAM also keep the ability of bbox prediction. This is realized through bounding box regression with the BBox Predictor. Following ISTR [24] and TransDLANet [9], for each HQD layer we predict the residual values of bbox coordinates relative to predictions of the previous HQD layer.", + "bbox": [ + 89, + 496, + 483, + 587 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "3.6. Model Learning", + "text_level": 1, + "bbox": [ + 89, + 595, + 251, + 612 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "3.6.1. Loss Function", + "text_level": 1, + "bbox": [ + 89, + 619, + 235, + 633 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "There are four losses in DocSAM, namely semantic mask segmentation loss $L_{S}$ , instance mask segmentation loss $L_{I}$ , instance bbox regression loss $L_{B}$ , and instance classification loss $L_{C}$ . Among them, $L_{S}$ is calculated as:", + "bbox": [ + 89, + 637, + 483, + 698 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\nL _ {S} = \\lambda_ {f} L _ {f o c a l} (M _ {S}, \\hat {M} _ {S}) + \\lambda_ {d} L _ {d i c e} (M _ {S}, \\hat {M} _ {S}), \\quad (4)\n$$\n", + "text_format": "latex", + "bbox": [ + 114, + 709, + 482, + 728 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "where $M_S$ and $\\hat{M}_S$ are predicted and ground-truth semantic masks, $L_{focal}$ and $L_{dice}$ are focal loss [67] and dice loss [54], respectively, and $\\lambda_f = 10$ and $\\lambda_d = 1$ are hyperparameters. Similarly, $L_I$ is calculated as:", + "bbox": [ + 89, + 739, + 483, + 801 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\nL _ {I} = \\lambda_ {f} L _ {f o c a l} (M _ {I}, \\hat {M} _ {I}) + \\lambda_ {d} L _ {d i c e} (M _ {I}, \\hat {M} _ {I}), \\tag {5}\n$$\n", + "text_format": "latex", + "bbox": [ + 119, + 811, + 482, + 830 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "where $M_S$ and $\\hat{M}_S$ are predicted and ground-truth instance masks, respectively. $L_B$ is calculated as:", + "bbox": [ + 89, + 842, + 482, + 872 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\nL _ {B} = \\lambda_ {s l 1} L _ {s l 1} (B _ {I}, \\hat {B} _ {I}) + \\lambda_ {d i o u} L _ {d i o u} (B _ {I}, \\hat {B} _ {I}), \\tag {6}\n$$\n", + "text_format": "latex", + "bbox": [ + 114, + 883, + 482, + 902 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "where $B_{I}$ and $\\hat{B}_I$ are predicted and ground-truth bounding boxes, $L_{sl1}$ and $L_{diou}$ are smooth L1 loss and distance IoU loss [105], respectively, and $\\lambda_{sl1} = 1$ and $\\lambda_{diou} = 1$ are hyper-parameters. At last, $L_{C}$ is calculated as:", + "bbox": [ + 511, + 90, + 905, + 151 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\nL _ {C} = L _ {c e} \\left(Y _ {I}, \\hat {Y} _ {I}\\right), \\tag {7}\n$$\n", + "text_format": "latex", + "bbox": [ + 643, + 162, + 905, + 180 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "where $Y_{I}$ and $\\hat{Y}_I$ are predicted and ground-truth instance labels, and $L_{ce}$ is cross entropy loss. The total loss of Doc-SAM is the sum of the above four losses:", + "bbox": [ + 511, + 189, + 905, + 234 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\nL = \\lambda_ {s} L _ {S} + \\lambda_ {i} L _ {I} + \\lambda_ {b} L _ {B} + \\lambda_ {c} L _ {C}, \\tag {8}\n$$\n", + "text_format": "latex", + "bbox": [ + 584, + 247, + 903, + 263 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "where $\\lambda_s = 5$ , $\\lambda_i = 5$ , $\\lambda_b = 1$ , and $\\lambda_c = 1$ are hyperparameters. We add auxiliary losses to every HQD layer and to query features before HQD. Following DETR [4] and Mask2Former [7], we also use bipartite matching to find the best matched instance predictions before calculating the loss. While for semantic predictions, there is no need to perform the bipartite matching, because the predictions and ground-truths are already one-to-one matched.", + "bbox": [ + 511, + 273, + 906, + 396 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "3.6.2. Heterogeneous Mixed Learning", + "text_level": 1, + "bbox": [ + 511, + 402, + 779, + 417 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Unlike existing methods, the novel design of DocSAM enables us to train a single model on heterogeneous mixed datasets. In this work, we collected nearly fifty DIS datasets of various document types and annotation formats, covering diverse DIS tasks from layout analysis and text detection to table structure recognition (see Tab. 1). We combined these datasets to construct a heterogeneous mixed dataset for training DocSAM. After training, the DocSAM model can be directly used as a versatile document segmenter or as a pre-trained model that can be seamlessly fine-tuned using task-specific datasets without any specialized modifications, such as adding or replacing a linear classification layer. This merit of DocSAM endows it with the potential for continual and incremental learning.", + "bbox": [ + 511, + 420, + 906, + 633 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "3.6.3. Improving Training Efficiency", + "text_level": 1, + "bbox": [ + 511, + 640, + 769, + 655 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Directly training DocSAM on such heterogeneous datasets may suffer from slow convergence and long training time, so we propose several strategies to improve training efficiency. Firstly, we pre-train the vision encoder of DocSAM on all 48 datasets using SimMIM [84], hoping it can provide more robust visual features for document images. Secondly, we separate the training datasets into groups with each group containing datasets of similar tasks and styles, see Tab. 1, then we adopt curriculum learning (CL) [81] strategy to warm up the training process by gradually adding new group of datasets. Thirdly, we add an instance query selection (IQS) process at the front of each HQD layer. Motivation behind this is that bipartite matching only calculates losses between matched predictions and ground-truths, and the matched query indexes are mostly the same across HQD layers. For a certain document, large ratio of instance", + "bbox": [ + 511, + 657, + 906, + 900 + ], + "page_idx": 4 + }, + { + "type": "page_number", + "text": "15025", + "bbox": [ + 480, + 944, + 517, + 955 + ], + "page_idx": 4 + }, + { + "type": "image", + "img_path": "images/9c0a89f71440e055410d888e55f44672bd3c8951e4b6666328905956878a4fbd.jpg", + "image_caption": [ + "(a) Training loss curves.", + "Figure 3. Loss curves and on-the-fly validation during training." + ], + "image_footnote": [], + "bbox": [ + 133, + 87, + 496, + 244 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/36491aebc5bd586bb356f174b84f45010b1327fbffa1cbf40d98039795ff5cfb.jpg", + "image_caption": [ + "(b) On-the-fly validation." + ], + "image_footnote": [], + "bbox": [ + 498, + 89, + 867, + 244 + ], + "page_idx": 5 + }, + { + "type": "table", + "img_path": "images/a16467fd4bf7371b4a003ab1bfdbaea232b3a90be56253feccd16d8e94cc9f8e.jpg", + "table_caption": [], + "table_footnote": [], + "table_body": "
Ablation SettingInstanceSemantic
mAP\\( mAP_b \\)mAFmIoU
DocSAMbase0.38040.35170.42560.6615
DocSAMbase + Curriculum0.38690.37040.43380.6622
DocSAMbase + SimMIM0.10020.06580.12940.3882
DocSAMbase + Freezing BERT0.38430.35100.42950.6677
DocSAMbase - Masked Attention0.23220.09380.28460.6327
DocSAMbase - Query Interaction0.29900.17790.35090.6511
DocSAMbase - Query Selection0.33410.31340.37630.6592
DocSAMlarge0.39000.36580.43200.6726
", + "bbox": [ + 102, + 309, + 472, + 428 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Table 2. Ablation studies on model structure and training strategy.", + "bbox": [ + 89, + 438, + 480, + 453 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "queries are not activated from beginning to end, and their class scores are very low. Therefore, we only select instance queries whose class scores are higher than a threshold $T_{k}$ before the $k$ -th HQD layer. We set $T_{k}$ as: $T_{k} = T_{max} / 2^{K - k}$ , where $T_{max} = 0.01$ is the maximum threshold, $K$ is the number of HQD layers, $k$ is the current HQD layer. Experiments show that IQS can discard low-score queries without degrading model performance, thereby improving training speed and reducing memory usage.", + "bbox": [ + 89, + 478, + 483, + 613 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "4. Experiments", + "text_level": 1, + "bbox": [ + 89, + 627, + 220, + 643 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "4.1. Datasets and Metrics", + "text_level": 1, + "bbox": [ + 89, + 652, + 287, + 666 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "The datasets involved in our experiments are listed in Tab. 1. Underlined datasets (15 in total) are used for ablation studies, mixed pre-training, and dataset-specific fine-tuning. All 48 datasets are used for training the final DocSAM model. These datasets cover a wide range of domains and tasks, showing significant heterogeneity in document types, annotation formats, and other aspects. Typical examples are shown in Fig. 1, with more details provided in the supplementary material. For evaluation metrics, we use mIoU [46] for semantic segmentation and mAP (for masks) [39] and $\\mathrm{mAP}_b$ (for bounding boxes) [39] for instance segmentation. Additionally, we introduce a new metric for instance segmentation called mAF, which is calculated as the mean F-score of all classes across all IoUs ranging from 0.5 to 0.95 in increments of 0.05 (i.e., [0.5:0.05:0.95]).", + "bbox": [ + 89, + 674, + 482, + 900 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "4.2. Implementation Details", + "text_level": 1, + "bbox": [ + 511, + 311, + 730, + 328 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "The vision backbone and deformable attention module are initialized from Mask2Former [7], which is pre-trained on the COCO-panoptic dataset [39]. The Sentence-BERT is initialized using the all-MiniLM-L6-v2 model from the Sentence Transformers library [64]. Other parts of DocSAM are randomly initialized. We trained two sizes of models: DocSAM-base (207M parameters) and DocSAM-large (317M parameters). Their vision backbones use Swin-base and Swin-large, respectively, and the instance query numbers $N$ are set to 500 and 900, respectively. The HQD layer number $K$ is set to 4 by default.", + "bbox": [ + 511, + 334, + 906, + 500 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "DocSAM is implemented based on PyTorch [58] and trained on $8 \\times$ NVIDIA A800 GPUs. We use the AdamW optimizer [50] to train the model, setting the base learning rate to $4 \\times 10^{-5}$ , and decay it using cosine annealing strategy [51]. For joint training on mixed datasets, the default settings are 80,000 iterations and a batch size of 32; for ablation studies and dataset-specific fine-tuning, the defaults are 20,000 iterations and a batch size of 8; for comparison with state-of-the-art, the defaults are 40,000 iterations and a batch size of 16.", + "bbox": [ + 511, + 501, + 908, + 651 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "4.3. Main Results", + "text_level": 1, + "bbox": [ + 511, + 662, + 651, + 676 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "4.3.1. Ablation Studies", + "text_level": 1, + "bbox": [ + 511, + 684, + 674, + 698 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "To verify the effect of each module in DocSAM and select the best training strategy before large-scale training, we conducted a series of ablation studies, as shown in Tab. 2 and Fig. 3. The results in Tab. 2 are averaged over all 15 datasets. On-the-fly validation involves fast testing on a small number of samples (e.g. 10 for each dataset) during training. The results show that using curriculum learning and instance query selection can accelerate convergence and improve model performance, while SimMIM pre-training significantly degrades model performance, possibly due to the large gap between SimMIM and document segmentation. Since freezing the weights of Sentence-BERT has almost no impact on performance, we freeze them during", + "bbox": [ + 511, + 703, + 908, + 900 + ], + "page_idx": 5 + }, + { + "type": "page_number", + "text": "15026", + "bbox": [ + 480, + 944, + 517, + 955 + ], + "page_idx": 5 + }, + { + "type": "table", + "img_path": "images/bee58aaee92f1b39cf52a5e8f47be07f6484b2fcb1e0fe590769c1259bc575d5.jpg", + "table_caption": [], + "table_footnote": [], + "table_body": "
TaskDatasetDocSAM-baseDocSAM-large
InstanceSemanticInstanceSemantic
AP50AP75mAP\\( mAP_b \\)mAFmIoUAP50AP75mAP\\( mAP_b \\)mAFmIoU
Document Layout Analysis\\( D^4 \\)LA [13]0.5950.5140.4480.4380.4860.3890.6370.5620.4900.4730.5390.434
DocLayNet [60]0.7160.5280.4840.4800.5430.6070.7440.5700.5170.5010.5840.669
\\( M^6 \\)Doc [9]0.5190.4020.3630.3520.3810.2670.5510.4440.3970.3870.4250.296
PubLayNet [106]0.9360.8620.8060.7890.8470.8980.9460.8840.8300.8050.8680.911
Ancient and Handwritten Document SegmentationSCUT-CAB-logical [8]0.6810.5550.4810.4780.5020.4100.7170.5740.5110.4950.5340.454
SCUT-CAB-physical [8]0.9370.8370.7770.7470.8210.9370.9480.8560.7860.7540.8290.942
HJDataset [69]0.9560.9210.8810.8650.8950.8190.9560.9250.8850.8690.8980.821
CASIA-HWDB [41]0.9290.7850.7210.6640.7880.9350.9120.7700.7140.6430.7900.939
SCUT-HCCDoc [95]0.8650.6350.5440.5590.6250.8440.8690.6420.5490.5600.6250.847
Table Structure RecognitionFinTabNet [104]0.8670.7700.6640.6270.7570.8510.8690.7860.6840.6440.7780.860
PubTabNet [107]0.9700.7880.6430.6350.7140.8400.9700.7890.6480.6340.7230.845
TableBank-latex [35]0.9630.9470.8970.8680.9240.9400.9650.9500.9150.8930.9360.951
TableBank-word [35]0.8730.8370.8220.7930.8510.8440.8780.8440.8350.8140.8570.853
Scene Text DetectionCTW1500 [43]0.7120.4300.4000.3680.5000.7940.7530.4820.4410.4020.5310.817
Total-Text [11]0.7470.4210.4050.4070.7430.4530.7690.4540.4280.4210.4720.764
MSRA-TD500 [86]0.7470.5250.4580.4770.5020.7130.7980.5770.4960.5160.5320.739
ICDAR2015 [28]0.6130.2470.2940.3020.3380.5990.6390.2600.3070.3130.3450.623
", + "bbox": [ + 129, + 84, + 859, + 323 + ], + "page_idx": 6 + }, + { + "type": "table", + "img_path": "images/b6f5e1b05a4cfa6934583f47f4d08beea9aeb1a53647a4ddcf1f5fd6fbf1b8e7.jpg", + "table_caption": [ + "Table 3. Performance of DocSAM after joint pre-training." + ], + "table_footnote": [], + "table_body": "
TaskDatasetDocSAM-large from scratchDocSAM-large from pretrain
InstanceSemanticInstanceSemantic
AP50AP75mAP\\( mAP_b \\)mAFmIoUAP50AP75mAP\\( mAP_b \\)mAFmIoU
DocumentLayoutAnalysis\\( D^4LA[13] \\)0.3650.2590.2390.1940.2330.2050.6980.6370.5550.5460.5950.526
DocLayNet [60]0.5030.2920.2950.2600.3590.3650.8330.6910.6210.6010.6790.736
\\( M^6Doc[9] \\)0.2790.1730.1690.1450.1630.0870.6670.5660.5000.4850.5280.430
PubLayNet [106]0.8730.7590.6960.6220.7380.8410.9540.9040.8540.8500.8880.921
Ancient andHandwrittenDocument SegmentationSCUT-CAB-logical [8]0.3910.2330.2390.1360.2280.2260.7830.6310.5560.5300.5820.481
SCUT-CAB-physical [8]0.8010.6440.6050.4050.6640.9180.9460.8690.7990.7620.8420.945
HJDataset [69]0.8480.8350.7520.6060.7770.8120.9830.9480.9050.8950.9110.822
CASIA-HWDB [41]0.9080.7370.6650.6280.7370.9490.9770.9390.8930.7920.9160.956
SCUT-HCCDoc [95]0.8070.4920.4600.4230.5410.8530.9040.6840.5800.5890.6580.862
Table StructureRecognitionFinTabNet [104]0.3350.1640.1780.0040.2220.7700.8770.8050.7130.6810.8030.870
PubTabNet [107]0.0130.0100.0070.0420.0070.8100.9730.8210.6690.6530.7420.860
TableBank-latex [35]0.7620.6120.5650.0200.6410.9130.9680.9540.9260.9090.9470.958
TableBank-word [35]0.5940.4460.4350.0450.6190.8230.9080.8770.8710.8590.8810.873
Scene Text DetectionCTW1500 [43]0.4310.0980.1620.0710.2530.8000.7940.5390.4800.4530.5730.831
Total-Text [11]0.3130.0420.0960.0470.1680.7490.7940.5170.4600.4660.5020.775
MSRA-TD500 [86]0.5060.1890.2220.0760.2950.7310.8090.6040.5240.5410.5550.744
ICDAR2015 [28]0.2030.0280.0630.0230.1130.5970.6810.3160.3410.3530.3790.641
Unseen DatasetIIIT-AR-13K [55]0.5550.4300.4030.1850.4170.7390.8420.7020.6380.6210.6930.642
CHDAC-2022 [31]0.8860.6960.6040.5090.6490.9150.9390.8280.6870.6250.7270.918
", + "bbox": [ + 133, + 364, + 864, + 631 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Table 4. Performance of DocSAM after dataset specific fine-tuning.", + "bbox": [ + 295, + 641, + 697, + 655 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "training. Similar to Mask2Former, masked attention plays a crucial role in DocSAM, and removing it leads to a significant performance drop. Additionally, without query interaction, DocSAM's performance also decreases substantially, highlighting the importance of information exchange between instance and semantic queries. Finally, training a unified model on heterogeneous datasets heavily relies on the model's capacity, and using a more powerful vision backbone can greatly enhance model performance.", + "bbox": [ + 84, + 679, + 478, + 814 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "4.3.2. Pre-training and Fine-tuning", + "text_level": 1, + "bbox": [ + 89, + 833, + 338, + 849 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "We train DocSAM on mixed heterogeneous datasets (15 datasets) to validate its performance as a unified document segmenter and a pre-trained model for dataset-specific fine", + "bbox": [ + 89, + 854, + 480, + 900 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "tuning. The results are shown in Tab. 3 and Tab. 4. DocSAM achieves good semantic and instance segmentation performance on various datasets and tasks, though performance may vary across datasets due to differing levels of difficulty. As a single-modal model, DocSAM may underperform on datasets like $\\mathrm{D}^4\\mathrm{LA}$ [13], DocLayNet [60], $\\mathrm{M}^6\\mathrm{Doc}$ [9] and SCUT-CAB-logical [8], which require multi-modal information for fine-grained logical layout analysis.", + "bbox": [ + 511, + 681, + 903, + 818 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "After joint training, we fine-tune DocSAM-large on each specific dataset to further improve performance. As shown in Tab. 4, fine-tuning results are significantly higher than direct testing and training from scratch. We also test DocSAM on unseen datasets IIIT-AR-13K [55] and CHDAC", + "bbox": [ + 511, + 825, + 903, + 900 + ], + "page_idx": 6 + }, + { + "type": "page_number", + "text": "15027", + "bbox": [ + 480, + 945, + 517, + 955 + ], + "page_idx": 6 + }, + { + "type": "table", + "img_path": "images/25557cf06dbe8b6a7b07e9ccaf211f4d2758fa721d2b7b8de6a6c3902ac0e804.jpg", + "table_caption": [], + "table_footnote": [], + "table_body": "
MethodObjectInstance
mAPAP50AP75mAPAP50AP75
Faster R-CNN [65]0.4900.6780.5720.4780.6780.552
Mask R-CNN [20]0.4010.5840.4620.3970.5840.456
Deformable DETR [108]0.5720.7680.6340.5560.7650.611
ISTR [24]0.6270.8080.7080.6200.8070.702
TransDLANet [9]0.6450.8270.7270.6380.8260.719
DAT [79]0.712--0.657--
DocSAM0.6630.8400.7550.6610.8400.750
", + "bbox": [ + 102, + 89, + 470, + 195 + ], + "page_idx": 7 + }, + { + "type": "table", + "img_path": "images/0d75b51bd03a7ea210ea8217c48ce340d337836f1a484860eb1460d6f75af982.jpg", + "table_caption": [ + "Table 5. Performance comparison on ${\\mathrm{M}}^{6}$ Doc." + ], + "table_footnote": [], + "table_body": "
MethodObjectInstance
mAPAP50AP75mAPAP50AP75
Physical
Faster R-CNN [65]0.7750.9130.8610.7530.9100.834
Mask R-CNN [20]0.7910.9210.8770.7950.9170.872
SCNet [77]0.8130.9410.8900.8200.9410.891
Deformable DETR [108]0.7990.9230.8710.7790.9210.843
VSR [99]0.7870.9190.8600.7870.9190.852
DocSAM0.7740.9470.8600.8110.9480.891
Logical
Faster R-CNN [65]0.5490.7740.6130.5420.7730.606
Mask R-CNN [20]0.5510.7850.6190.5530.7770.631
SCNet [77]0.6020.8360.6730.6030.8360.680
Deformable DETR [108]0.6270.8520.7170.6200.8510.703
VSR [99]0.5570.7830.6160.5510.7820.611
DocSAM0.5480.7690.6320.5750.7790.667
", + "bbox": [ + 102, + 234, + 470, + 431 + ], + "page_idx": 7 + }, + { + "type": "table", + "img_path": "images/7ff06e1537ec0a1ab860d1a18ac2123c7ae905bf3e61125a639820568dc36a82.jpg", + "table_caption": [ + "Table 6. Performance comparison on SCUT-CAB." + ], + "table_footnote": [], + "table_body": "
MethodCTW1500Total-Text
PRFPRF
HierText [49]0.8460.8740.8600.8550.9050.879
SIR [61]0.8740.8370.8550.9090.8560.882
DPText-DETR [89]0.9170.8620.8880.9180.8640.890
UNITS [29]-----0.898
ESTextSpotter [26]0.9150.8860.9000.9200.8810.900
DAT-DET [79]0.8930.8930.8930.9400.8820.910
DAT-SEG [79]0.9250.9090.9170.9500.8920.920
DocSAM0.8050.8810.8420.7210.8260.770
", + "bbox": [ + 116, + 473, + 457, + 590 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Table 7. Performance comparison on CTW1500 and Total-Text.", + "bbox": [ + 96, + 601, + 475, + 614 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "2022 [31], where fine-tuning from the pre-trained model also yields substantial performance gains. This demonstrates that DocSAM's performance is not yet saturated and can benefit greatly from transfer learning on unseen datasets and tasks.", + "bbox": [ + 89, + 643, + 482, + 719 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "4.3.3. Comparison with State-of-the-Arts", + "text_level": 1, + "bbox": [ + 89, + 729, + 380, + 744 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "To compare with state-of-the-art methods, we further fine-tuned DocSAM on some datasets for additional training iterations. The results are shown in Tab. 5, Tab. 6, and Tab. 7. The best results are shown in bold, and the second-best results are underlined. DocSAM achieves superior or comparable performance with other methods. Note that we did not apply any specific training techniques or data augmentation, configurations for all datasets were kept consistent. We found that DocSAM exhibits much lower performance in logical layout analysis compared to physical analysis,", + "bbox": [ + 89, + 750, + 482, + 901 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "which we attribute to its reliance only on single-modal features. Furthermore, DocSAM achieved relatively low performance on scene text detection datasets. This is likely because scene texts exhibit much greater diversity in shapes and backgrounds, requiring more carefully designed strategies to ensure model performance.", + "bbox": [ + 511, + 90, + 903, + 181 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "4.4. Discussion", + "text_level": 1, + "bbox": [ + 511, + 191, + 632, + 205 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "The goal of this paper is not to achieve state-of-the-art performance on specific dataset and task through meticulously designed model architectures or training strategies. Instead, we aim to design a simple and unified document segmentation model that can be applied to a wide variety of datasets and tasks. Additionally, the trained model should possess good scalability and the ability to continue learning. In this regard, DocSAM is quite successful. It exhibits decent performance on various datasets and tasks and shows great potential for downstream applications both as a versatile segmenter and a pre-trained model. However, experimental results also reveal some weaknesses and limitations of DocSAM, such as long training time and unsatisfactory performance on complex scenarios. We believe that DocSAM can greatly benefit from more sophisticated model design and better data augmentation and training strategies to further accelerate its convergence and improve its performance.", + "bbox": [ + 511, + 213, + 906, + 470 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "5. Conclusion", + "text_level": 1, + "bbox": [ + 511, + 483, + 633, + 500 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "In this paper, we propose DocSAM, a transformer-based unified framework for various document image segmentation tasks. DocSAM integrates layout analysis, multi-grained text segmentation, and table structure decomposition into a single model, reducing the need for specialized models and enhancing efficiency. Trained on heterogeneous datasets, DocSAM demonstrates robust and generalizable performance, effectively handling diverse document types and structures. This approach also reduces computational and storage requirements, making DocSAM suitable for practical deployment in resource-constrained environments. Extensive experiments show that DocSAM outperforms existing methods in terms of accuracy, efficiency, and adaptability. Overall, we believe that DocSAM represents a significant step forward for document image segmentation, and we look forward to its continued development and application in practical scenarios. In the future, we plan to extend DocSAM to a multi-modal version and explore better training strategies to further accelerate its convergence and improve its performance.", + "bbox": [ + 511, + 508, + 906, + 811 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Acknowledgement", + "text_level": 1, + "bbox": [ + 511, + 824, + 671, + 842 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "This work is supported by the National Natural Science Foundation of China(NSFC) Grant U23B2029.", + "bbox": [ + 511, + 849, + 903, + 878 + ], + "page_idx": 7 + }, + { + "type": "page_number", + "text": "15028", + "bbox": [ + 480, + 944, + 517, + 955 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "References", + "text_level": 1, + "bbox": [ + 91, + 89, + 187, + 104 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[1] Abdelrahman Abdallah, Alexander Berendeyev, Islam Nuradin, and Daniyar Nurseitov. Tncr: Table net detection and classification dataset. Neurocomputing, 473:79-97, 2022. 3", + "[2] Avinash Anand, Raj Jaiswal, Mohit Gupta, Siddhesh S Bangar, Pijush Bhuyan, Naman Lal, Rajeev Singh, Ritika Jha, Rajiv Ratn Shah, and Shin'Ichi Satoh. Ranlaynet: A dataset for document layout detection used for domain adaptation and generalization. In ACM MM Asia, pages 1-6, 2023. 3", + "[3] buptlihang. CDLA: A Benchmark Dataset for Cross-Domain Layout Analysis. https://github.com/buptliang/CDLA, 2023. Accessed: 2024-10-31. 3", + "[4] Nicolas Carion, Francisco Massa, Gabriel Synnaeve, Nicolas Usunier, Alexander Kirillov, and Sylvain Gelly. End-to-end object detection with transformers. In ECCV, pages 213-229. Springer, 2020. 3, 5", + "[5] Liang-Chieh Chen, George Papandreou, Iasonas Kokkinos, Vladimir Koltun, and Alan Garg. Deep convolutional nets, atrous convolution, and fully connected crfs. In IEEE Trans. Pattern Anal. Mach. Intell., pages 834-848, 2018. 3", + "[6] Bowen Cheng, Alex Schwing, and Alexander Kirillov. Per-pixel classification is not all you need for semantic segmentation. NeurIPS, 34:17864-17875, 2021. 3", + "[7] Bowen Cheng, Ishan Misra, Alexander G Schwing, Alexander Kirillov, and Rohit Girdhar. Masked-attention mask transformer for universal image segmentation. In CVPR, pages 1290–1299, 2022. 3, 4, 5, 6", + "[8] Hiuyi Cheng, Cheng Jian, Sihang Wu, and Lianwen Jin. Scut-cab: a new benchmark dataset of ancient chinese books with complex layouts for document layout analysis. In ICFHR, pages 436-451. Springer, 2022. 2, 3, 7", + "[9] Hiuyi Cheng, Peirong Zhang, Sihang Wu, Jiaxin Zhang, Qiyuan Zhu, Zecheng Xie, Jing Li, Kai Ding, and Lianwen Jin. M6doc: A large-scale multi-format, multi-type, multilayout, multi-language, multi-annotation category dataset for modern document layout analysis. In CVPR, pages 15138–15147, 2023. 2, 3, 5, 7, 8", + "[10] Zewen Chi, Heyan Huang, Heng-Da Xu, Houjin Yu, Wanxuan Yin, and Xian-Ling Mao. Complicated table structure recognition. arXiv preprint arXiv:1908.04729, 2019. 3", + "[11] Chee Kheng Ch'ng and Chee Seng Chan. Total-text: A comprehensive dataset for scene text detection and recognition. In ICDAR, pages 935-942. IEEE, 2017. 2, 3, 7", + "[12] Chee Kheng Chng, Yuliang Liu, Yipeng Sun, Chun Chet Ng, Canjie Luo, Zihan Ni, ChuanMing Fang, Shuai tao Zhang, Junyu Han, Errui Ding, et al. Icdar2019 robust reading challenge on arbitrary-shaped text-rrc-art. In ICDAR, pages 1571-1576. IEEE, 2019. 3", + "[13] Cheng Da, Chuwei Luo, Qi Zheng, and Cong Yao. Vision grid transformer for document layout analysis. In ICCV, pages 19462-19472, 2023. 3, 7", + "[14] Hao Feng, Qi Liu, Hao Liu, Wengang Zhou, Houqiang Li, and Can Huang. Docpedia: Unleashing the power of large multimodal model in the frequency domain for versatile document understanding. arXiv preprint arXiv:2311.11810, 2023. 3" + ], + "bbox": [ + 101, + 114, + 482, + 898 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[15] Hao Feng, Zijian Wang, Jingqun Tang, Jinghui Lu, Wengang Zhou, Houqiang Li, and Can Huang. Unidoc: A universal large multimodal model for simultaneous text detection, recognition, spotting and understanding. arXiv preprint arXiv:2308.11592, 2023. 3", + "[16] Liangcai Gao, Xiaohan Yi, Zhuoren Jiang, Leipeng Hao, and Zhi Tang. Icdar2017 competition on page object detection. In ICDAR, pages 1417-1422. IEEE, 2017. 2, 3", + "[17] Liangcai Gao, Yilun Huang, Herve Dejean, Jean-Luc Meunier, Qinqin Yan, Yu Fang, Florian Kleber, and Eva Lang. Icdar 2019 competition on table detection and recognition (ctdar). In ICDAR, pages 1510-1515. IEEE, 2019. 2, 3", + "[18] Max Göbel, Tamir Hassan, Ermelinda Oro, and Giorgio Orsi. Icdar 2013 table competition. In ICDAR, pages 1449-1453. IEEE, 2013. 3", + "[19] Mrinal Haloi, Shashank Shekhar, Nikhil Fande, Siddhant Swaroop Dash, et al. Table detection in the wild: A novel diverse table detection dataset and method. arXiv preprint arXiv:2209.09207, 2022.3", + "[20] Kaiming He, Georgia Gkioxari, Piotr Dollar, and Ross Girshick. Mask r-cnn. In ICCV, pages 2961-2969, 2017. 3, 8", + "[21] Mengchao He, Yuliang Liu, Zhibo Yang, Sheng Zhang, Canjie Luo, Feiyu Gao, Qi Zheng, Yongpan Wang, Xin Zhang, and Lianwen Jin. Icpr2018 contest on robust reading for multi-type web images. In ICPR, pages 7-12. IEEE, 2018. 3", + "[22] Wenhao He, Xu-Yao Zhang, Fei Yin, and Cheng-Lin Liu. Multi-oriented and multi-lingual scene text detection with direct regression. ICIP, 27(11):5406-5419, 2018. 3", + "[23] Anwen Hu, Haiyang Xu, Jiabo Ye, Ming Yan, Liang Zhang, Bo Zhang, Chen Li, Ji Zhang, Qin Jin, Fei Huang, et al. mplug-docowl 1.5: Unified structure learning forOCR-free document understanding. arXiv preprint arXiv:2403.12895, 2024.3", + "[24] Jie Hu, Liujuan Cao, Yao Lu, ShengChuan Zhang, Yan Wang, Ke Li, Feiyue Huang, Ling Shao, and Rongrong Ji. Istr: End-to-end instance segmentation with transformers. arXiv preprint arXiv:2105.00637, 2021. 3, 5, 8", + "[25] Jie Hu, Yao Lu, Shengchuan Zhang, and Liujuan Cao. Istr: Mask-embedding-based instance segmentation transformer. IEEE Trans. Image Process., 2024. 3", + "[26] Mingxin Huang, Jiaxin Zhang, Dezhi Peng, Hao Lu, Can Huang, Yuliang Liu, Xiang Bai, and Lianwen Jin. Estextspotter: Towards better scene text spotting with explicit synergy in transformer. In ICCV, pages 19495-19505, 2023. 8", + "[27] Yupan Huang, Tengchao Lv, Lei Cui, Yutong Lu, and Furu Wei. Layoutlmv3: Pre-training for document ai with unified text and image masking. In ACM MM, pages 4083-4091, 2022. 3", + "[28] Dimosthenis Karatzas, Lluis Gomez-Bigorda, Anguelos Nicolaou, Suman Ghosh, Andrew Bagdanov, Masakazu Iwamura, Jiri Matas, Lukas Neumann, Vijay Ramaseshan Chandrasekhar, Shijian Lu, et al. Icdar 2015 competition on robust reading. In ICDAR, pages 1156-1160. IEEE, 2015. 3, 7" + ], + "bbox": [ + 522, + 92, + 903, + 898 + ], + "page_idx": 8 + }, + { + "type": "page_number", + "text": "15029", + "bbox": [ + 480, + 944, + 517, + 955 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[29] Taeho Kil, Seonghyeon Kim, Sukmin Seo, Yoonsik Kim, and Daehee Kim. Towards unified scene text spotting based on sequence generation. In CVPR, pages 15223-15232, 2023. 8", + "[30] Alexander Kirillov, Eric Mintun, Nikhila Ravi, Hanzi Mao, Chloe Rolland, Laura Gustafson, Tete Xiao, Spencer Whitehead, Alexander C Berg, Wan-Yen Lo, et al. Segment anything. In ICCV, pages 4015-4026, 2023. 3", + "[31] Pazhou Lab. IACC competition on Chinese Historical Document Analysis Challenge. https://iacc.pazhoulab-huangpu.com/contestdetail?id=6497f74cd97a2dae9dcaeff8 & award=1, 000, 000, 2022. Accessed: 2024-10-31. 3, 7, 8", + "[32] John Lafferty, Andrew McCallum, Fernando Pereira, et al. Conditional random fields: Probabilistic models for segmenting and labeling sequence data. In ICML, page 3. Williamstown, MA, 2001. 3", + "[33] Feng Li, Hao Zhang, Shilong Liu, Jian Guo, Lionel M Ni, and Lei Zhang. Dn-detr: Accelerate detr training by introducing query denoising. In CVPR, pages 13619-13627, 2022. 3", + "[34] Junlong Li, Yiheng Xu, Tengchao Lv, Lei Cui, Cha Zhang, and Furu Wei. Dit: Self-supervised pre-training for document image transformer. In ACM MM, pages 3530-3539, 2022. 3", + "[35] Minghao Li, Lei Cui, Shaohan Huang, Furu Wei, Ming Zhou, and Zhoujun Li. Tablebank: Table benchmark for image-based table detection and recognition. In LREC, pages 1918-1925, 2020. 2, 3, 7", + "[36] Minghao Li, Yiheng Xu, Lei Cui, Shaohan Huang, Furu Wei, Zhoujun Li, and Ming Zhou. Docbank: A benchmark dataset for document layout analysis. arXiv preprint arXiv:2006.01038, 2020. 2, 3", + "[37] Xiao-Hui Li, Fei Yin, and Cheng-Lin Liu. Page segmentation using convolutional neural network and graphical model. In DAS, pages 231-245. Springer, 2020. 3", + "[38] Xiao-Hui Li, Fei Yin, He-Sen Dai, and Cheng-Lin Liu. Table structure recognition and form parsing by end-to-end object detection and relation parsing. Pattern Recognition, 132:108946, 2022. 2, 3", + "[39] Tsung-Yi Lin, Michael Maire, Serge Belongie, James Hays, Pietro Perona, Deva Ramanan, Piotr Dólar, and C Lawrence Zitnick. Microsoft coco: Common objects in context. In ECCV, pages 740-755. Springer, 2014. 6", + "[40] Tsung-Yi Lin, Piotr Dólar, Ross Girshick, Kaiming He, Bharath Hariharan, and Serge Belongie. Feature pyramid networks for object detection. In CVPR, pages 2117-2125, 2017. 4", + "[41] Cheng-Lin Liu, Fei Yin, Da-Han Wang, and Qiu-Feng Wang. Casia online and offline chinese handwriting databases. In ICDAR, pages 37-41. IEEE, 2011. 3, 7", + "[42] Cheng-Lin Liu, Lianwen Jin, Xiang Bai, Xiaohui Li, and Fei Yin. Frontiers of intelligent document analysis and recognition: review and prospects. Journal of Image and Graphics, 28(08):2223-2252, 2023. 1", + "[43] Yuliang Liu, Lianwen Jin, Shuai tao Zhang, Canjie Luo, and Sheng Zhang. Curved scene text detection via transverse" + ], + "bbox": [ + 99, + 92, + 480, + 900 + ], + "page_idx": 9 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "and longitudinal sequence connection. Pattern Recognition, 90:337-345, 2019. 2, 3, 7", + "[44] Yuliang Liu, Biao Yang, Qiang Liu, Zhang Li, Zhiyin Ma, Shuo Zhang, and Xiang Bai. Textmonkey: AnOCR-free large multimodal model for understanding document. arXiv preprint arXiv:2403.04473, 2024. 3", + "[45] Ze Liu, Yutong Lin, Yan Cao, Yue Hu, Yu Wei, Zhiqiang Zhang, Jiahui Lin, Han Wang, Cheng Lu, Changhu Wang, et al. Swin transformer: Hierarchical vision transformer using shifted windows. In ICCV, pages 12226-12235, 2021. 3, 4", + "[46] Jonathan Long, Evan Shelhamer, and Trevor Darrell. Fully convolutional networks for semantic segmentation. In CVPR, pages 3431-3440, 2015. 3, 6", + "[47] Rujiao Long, Wen Wang, Nan Xue, Feiyu Gao, Zhibo Yang, Yongpan Wang, and Gui-Song Xia. Parsing table structures in the wild. In ICCV, pages 944-952, 2021. 3", + "[48] Shangbang Long, Jiaqiang Ruan, Wenjie Zhang, Xin He, Wenhao Wu, and Cong Yao. Textsnake: A flexible representation for detecting text of arbitrary shapes. In ECCV, pages 20-36, 2018. 3", + "[49] Shangbang Long, Siyang Qin, Dmitry Panteleev, Alessandro Bissacco, Yasuhisa Fujii, and Michalis Raptis. Towards end-to-end unified scene text detection and layout analysis. In CVPR, pages 1049-1059, 2022. 3, 8", + "[50] I Loshchilov. Decoupled weight decay regularization. arXiv preprint arXiv:1711.05101, 2017. 6", + "[51] Ilya Loshchilov and Frank Hutter. Sgdr: Stochastic gradient descent with warm restarts. arXiv preprint arXiv:1608.03983, 2016. 6", + "[52] Siwen Luo, Yihao Ding, Siqu Long, Josiah Poon, and Soyeon Caren Han. Doc-gcn: Heterogeneous graph convolutional networks for document layout analysis. arXiv preprint arXiv:2208.10970, 2022. 3", + "[53] Weihong Ma, Hesuo Zhang, Lianwen Jin, Sihang Wu, Jiapeng Wang, and Yongpan Wang. Joint layout analysis, character detection and recognition for historical document digitization. In ICFHR, pages 31-36. IEEE, 2020. 2, 3", + "[54] Fausto Miletari, Nassir Navab, and Seyed-Ahmad Ahmadi. V-net: Fully convolutional neural networks for volumetric medical image segmentation. In 3DV, pages 565-571. IEEE, 2016. 5", + "[55] Ajoy Mondal, Peter Lipps, and CV Jawahar. Iit-ar-13k: A new dataset for graphical object detection in documents. In DAS, pages 216–230. Springer, 2020. 3, 7", + "[56] Nibal Nayef, Fei Yin, Imen Bizid, Hyunsoo Choi, Yuan Feng, Dimosthenis Karatzas, Zhenbo Luo, Umapada Pal, Christophe Rigaud, Joseph Chazalon, et al. Icdar2017 robust reading challenge on multi-lingual scene text detection and script identification-rrc-plt. In ICDAR, pages 1454-1459. IEEE, 2017. 3", + "[57] Nibal Nayef, Yash Patel, Michal Busta, Pinaki Nath Chowdhury, Dimosthenis Karatzas, Wafa Khlif, Jiri Matas, Umapada Pal, Jean-Christophe Burie, Cheng-lin Liu, et al. Icdar2019 robust reading challenge on multi-lingual scene text detection and recognition—rrc-mlt-2019. In ICDAR, pages 1582–1587. IEEE, 2019. 3" + ], + "bbox": [ + 522, + 92, + 903, + 900 + ], + "page_idx": 9 + }, + { + "type": "page_number", + "text": "15030", + "bbox": [ + 480, + 944, + 519, + 955 + ], + "page_idx": 9 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[58] Adam Paszke, Sam Gross, Francisco Massa, Adam Lerer, James Bradbury, Gregory Chanan, Trevor Killeen, Zeming Lin, Natalia Gimelshein, Luca Antiga, et al. Pytorch: An imperative style, high-performance deep learning library. NeurIPS, 32, 2019. 6", + "[59] Dezhi Peng, Zhenhua Yang, Jiaxin Zhang, Chongyu Liu, Yongxin Shi, Kai Ding, Fengjun Guo, and Lianwen Jin. Upocr: Towards unified pixel-levelOCR interface. In ICML, 2023. 3", + "[60] Birgit Pfitzmann, Christoph Auer, Michele Dolfi, Ahmed S. Nassar, and Peter Staar. Doclaynet: A large human-annotated dataset for document layout analysis. In ACM SIGKDD, pages 3743–3751. 2022. 2, 3, 7", + "[61] Xugong Qin, Pengyuan Lyu, Chengquan Zhang, Yu Zhou, Kun Yao, Peng Zhang, Hailun Lin, and Weiping Wang. Towards robust real-time scene text detection: From semantic to instance representation learning. In ACM MM, pages 2025-2034, 2023. 8", + "[62] Joseph Redmon and Ali Farhadi. You only look once: Unified, real-time object detection. In CVPR, pages 779-788, 2016. 3", + "[63] Joseph Redmon and Ali Farhadi. Yolov3: An incremental improvement, 2018. 3", + "[64] N Reimers. Sentence-bert: Sentence embeddings using siamese bert-networks. arXiv preprint arXiv:1908.10084, 2019. 1, 4, 6", + "[65] Shaoqing Ren, Kaiming He, Ross Girshick, and Jian Sun. Faster r-cnn: Towards real-time object detection with region proposal networks. In NeurIPS, pages 91-99, 2015. 3, 8", + "[66] Olaf Ronneberger, Philipp Fischer, and Thomas Brox. U-net: Convolutional networks for biomedical image segmentation. In MICCAI, pages 234-241. Springer, 2015. 3", + "[67] T-YLPG Ross and GKHP Dólar. Focal loss for dense object detection. In CVPR, pages 2980-2988, 2017. 5", + "[68] Rajkumar Saini, Derek Dobson, Jon Morrey, Marcus Liwicki, and Foteini Simistira Liwicki. Icdar 2019 historical document reading challenge on large structured chinese family records. In ICDAR, pages 1499-1504. IEEE, 2019. 3", + "[69] Zejiang Shen, Kaixuan Zhang, and Melissa Dell. A large dataset of historical japanese documents with complex layouts. In CVPR Workshops, pages 548-549, 2020. 3, 7", + "[70] Baoguang Shi, Cong Yao, Minghui Liao, Mingkun Yang, Pei Xu, Linyan Cui, Serge Belongie, Shijian Lu, and Xiang Bai. Icdar2017 competition on reading chinese text in the wild (rctw-17). In ICDAR, pages 1429–1434. IEEE, 2017. 3", + "[71] Md Istiak Hossain Shihab, Md Rakibul Hasan, Mahfuzur Rahman Emon, Syed Mobassir Hossen, Md Nazmuddoha Ansary, Intesur Ahmed, Fazle Rabbi Rakib, Shahriar Elahi Dhruvo, Souhardya Saha Dip, Akib Hasan Pavel, et al. Badlad: A large multi-domain bengali document layout analysis dataset. In ICDAR, pages 326-341. Springer, 2023. 3", + "[72] Brandon Smock, Rohith Pesala, and Robin Abraham. Pubtables-1m: Towards comprehensive table extraction" + ], + "bbox": [ + 99, + 92, + 482, + 900 + ], + "page_idx": 10 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "from unstructured documents. In CVPR, pages 4634-4642, 2022. 2, 3", + "[73] Xingyi Sun, Ziyi Zhou, Peizhao Wang, Junwei Dai, Lu Lu, Wanli Ouyang, and Ping Luo. Sparse r-cnn: End-to-end object detection with learnable proposals. In CVPR, pages 13164-13173, 2021. 3", + "[74] Yipeng Sun, Zihan Ni, Chee-Kheng Chng, Yuliang Liu, Canjie Luo, Chun Chet Ng, Junyu Han, Errui Ding, Jingtuo Liu, Dimosthenis Karatzas, et al. Icdar 2019 competition on large-scale street view text with partial labeling-rrc-lsvt. In ICDAR, pages 1557-1562. IEEE, 2019. 3", + "[75] Zineng Tang, Ziyi Yang, Guoxin Wang, Yuwei Fang, Yang Liu, Chenguang Zhu, Michael Zeng, Cha Zhang, and Mohit Bansal. Unifying vision, text, and layout for universal document processing. In CVPR, pages 19254-19264, 2023. 3", + "[76] Andreas Veit, Tomas Matera, Lukas Neumann, Jiri Matas, and Serge Belongie. Coco-text: Dataset and benchmark for text detection and recognition in natural images. arXiv preprint arXiv:1601.07140, 2016. 3", + "[77] Thang Vu, Haeyong Kang, and Chang D Yoo. Scnet: Training inference sample consistency for instance segmentation. In AAAI, pages 2701-2709, 2021. 8", + "[78] Jianqiang Wan, Sibo Song, Wenwen Yu, Yuliang Liu, Wenqing Cheng, Fei Huang, Xiang Bai, Cong Yao, and Zhibo Yang. Omniparser: A unified framework for text spotting key information extraction and table recognition. In CVPR, pages 15641-15653, 2024. 3", + "[79] Xingyu Wan, Chengquan Zhang, Pengyuan Lyu, Sen Fan, Zihan Ni, Kun Yao, Errui Ding, and Jingdong Wang. Towards unified multi-granularity text detection with interactive attention. arXiv preprint arXiv:2405.19765, 2024. 3, 8", + "[80] Dongsheng Wang, Natraj Raman, Mathieu Sibue, Zhiqiang Ma, Petr Babkin, Simerjot Kaur, Yulong Pei, Armineh Nourbakhsh, and Xiaomo Liu. Docllm: A layout-aware generative language model for multimodal document understanding. arXiv preprint arXiv:2401.00908, 2023. 3", + "[81] Xin Wang, Yudong Chen, and Wenwu Zhu. A survey on curriculum learning. IEEE Trans. Pattern Anal. Mach. Intell., 44(9):4555-4576, 2021. 5", + "[82] Zonghan Wu, Shirui Pan, Fengwen Chen, Guodong Long, Chengqi Zhang, and S Yu Philip. A comprehensive survey on graph neural networks. IEEE Trans. Neural Netw. Learn. Syst., 32(1):4-24, 2020. 3", + "[83] Enze Xie, Zhiwei Wang, Gang Yu, Xingyi Zhang, Yuheng Wu, Hongsheng Li, and Yonghong Tian. Segformer: Simple and efficient design for semantic segmentation with transformers. In ICCV, pages 14125-14134, 2021. 3", + "[84] Zhenda Xie, Zheng Zhang, Yue Cao, Yutong Lin, Jianmin Bao, Zhuliang Yao, Qi Dai, and Han Hu. Simmim: A simple framework for masked image modeling. In CVPR, pages 9653-9663, 2022. 5", + "[85] Yue Xu, Fei Yin, Da-Han Wang, Xu-Yao Zhang, Zhaoxiang Zhang, and Cheng-Lin Liu. Casia-ahcdb: A large-scale chinese ancient handwritten characters database. In ICDAR, pages 793–798. IEEE, 2019. 3" + ], + "bbox": [ + 522, + 92, + 903, + 900 + ], + "page_idx": 10 + }, + { + "type": "page_number", + "text": "15031", + "bbox": [ + 480, + 945, + 517, + 955 + ], + "page_idx": 10 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[86] Cong Yao, Xiang Bai, Wenyu Liu, Yi Ma, and Zhuowen Tu. Detecting texts of arbitrary orientations in natural images. In CVPR, pages 1083-1090. IEEE, 2012. 3, 7", + "[87] Cong Yao, Xiang Bai, and Wenyu Liu. A unified framework for multioriented text detection and recognition. ICIP, 23 (11):4737-4749, 2014. 3", + "[88] Jiabo Ye, Anwen Hu, Haiyang Xu, Qinghao Ye, Ming Yan, Yuhao Dan, Chenlin Zhao, Guohai Xu, Chenliang Li, Junfeng Tian, et al. mplug-docowl: Modularized multimodal large language model for document understanding. arXiv preprint arXiv:2307.02499, 2023. 3", + "[89] Maoyuan Ye, Jing Zhang, Shanshan Zhao, Juhua Liu, Bo Du, and Dacheng Tao. Dptext-detr: Towards better scene text detection with dynamic points in transformer. In AAAI, pages 3241-3249, 2023. 8", + "[90] Xu-Cheng Yin, Xuwang Yin, Kaizhu Huang, and Hong-Wei Hao. Robust text detection in natural scene images. IEEE Trans. Pattern Anal. Mach. Intell., 36(5):970–983, 2013. 3", + "[91] Wenwen Yu, Mingyu Liu, Mingrui Chen, Ning Lu, Yinlong Wen, Yuliang Liu, Dimosthenis Karatzas, and Xiang Bai. Icdar 2023 competition on reading the seal title. In ICDAR, pages 522-535. Springer, 2023. 3", + "[92] Tai-Ling Yuan, Zhe Zhu, Kun Xu, Cheng-Jun Li, Tai-Jiang Mu, and Shi-Min Hu. A large chinese text dataset in the wild. J. Comput. Sci. Tech., 34:509-521, 2019. 3", + "[93] Yukun Zhai, Xiaogiang Zhang, Xiameng Qin, Sanyuan Zhao, Xingping Dong, and Jianbing Shen. Textformer: A query-based end-to-end text spotter with mixed supervision. Machine Intelligence Research, 21(4):704-717, 2024. 2", + "[94] Chongsheng Zhang, Guowen Peng, Yuefeng Tao, Feifei Fu, Wei Jiang, George Almpanidis, and Ke Chen. Shopsign: A diverse scene text dataset of chinese shop signs in street views. arXiv preprint arXiv:1903.10412, 2019. 3", + "[95] Hesuo Zhang, Lingyu Liang, and Lianwen Jin. Scuthcdoc: A new benchmark dataset of handwritten chinese text in unconstrained camera-captured documents. Pattern Recognition, 108:107559, 2020. 2, 3, 7", + "[96] Hao Zhang, Feng Li, Shilong Liu, Lei Zhang, Hang Su, Jun Zhu, Lionel M Ni, and Heung-Yeung Shum. Dino: Detr with improved denoising anchor boxes for end-to-end object detection. arXiv preprint arXiv:2203.03605, 2022. 3", + "[97] Hao Zhang, Feng Li, Huaizhe Xu, Shijia Huang, Shilong Liu, Lionel M Ni, and Lei Zhang. Mp-former: Mask-piloted transformer for image segmentation. In CVPR, pages 18074-18083, 2023. 3", + "[98] Jiaxin Zhang, Dezhi Peng, Chongyu Liu, Peirong Zhang, and Lianwen Jin. Docres: A generalist model toward unifying document image restoration tasks. In CVPR, pages 15654-15664, 2024. 3", + "[99] Peng Zhang, Can Li, Liang Qiao, Zhanzhan Cheng, Shiliang Pu, Yi Niu, and Fei Wu. Vsr: a unified framework for document layout analysis combining vision, semantics and relations. In ICDAR, pages 115-130. Springer, 2021. 8", + "[100] Rui Zhang, Yongsheng Zhou, Qianyi Jiang, Qi Song, Nan Li, Kai Zhou, Lei Wang, Dong Wang, Minghui Liao," + ], + "bbox": [ + 94, + 92, + 482, + 900 + ], + "page_idx": 11 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "Mingkun Yang, et al. Icdar 2019 robust reading challenge on reading chinese text on signboard. In ICDAR, pages 1577-1581. IEEE, 2019. 3", + "[101] Shi-Xue Zhang, Xiaobin Zhu, Jie-Bo Hou, Chang Liu, Chun Yang, Hongfa Wang, and Xu-Cheng Yin. Deep relational reasoning graph network for arbitrary shape text detection. In CVPR, pages 9699–9708, 2020. 3", + "[102] Wayne Xin Zhao, Kun Zhou, Junyi Li, Tianyi Tang, Xiaolei Wang, Yupeng Hou, Yingqian Min, Beichen Zhang, Junjie Zhang, Zican Dong, et al. A survey of large language models. arXiv preprint arXiv:2303.18223, 2023. 3", + "[103] Enze Zheng, Zhiwei Wang, Gang Yu, Xingyi Zhang, Yuheng Wu, Hongsheng Li, and Yonghong Tian. Rethinking semantic segmentation from a sequence-to-sequence perspective with transformers. In CVPR, pages 12164-12173, 2021. 3", + "[104] Xinyi Zheng, Douglas Burdick, Lucian Popa, Xu Zhong, and Nancy Xin Ru Wang. Global table extractor (gte): A framework for joint table identification and cell structure recognition using visual context. In WACV, pages 697-706, 2021. 3, 7", + "[105] Zhaohui Zheng, Ping Wang, Wei Liu, Jinze Li, Rongguang Ye, and Dongwei Ren. Distance-iou loss: Faster and better learning for bounding box regression. In AAAI, pages 12993-13000, 2020. 5", + "[106] Xu Zhong, Jianbin Tang, and Antonio Jimeno Yepes. Publaynet: largest dataset ever for document layout analysis. In ICDAR, pages 1015-1022. IEEE, 2019. 2, 3, 7", + "[107] Xu Zhong, Elaheh ShafieiBavani, and Antonio Jimeno Yepes. Image-based table recognition: data, model, and evaluation. In ECCV, pages 564-580. Springer, 2020. 2,3,7", + "[108] Xingyi Zhu, Junwei Dai, Lu Lu, Yuwen Zhang, Peizhao Wang, Yisen Sun, Wanli Ouyang, and Ping Luo. Deformable detr: Deformable transformers for end-to-end object detection. In ICCV, pages 12352-12361, 2021. 1, 3, 4, 8", + "[109] Ziyi Zhu, Liangcai Gao, Yibo Li, Yilun Huang, Lin Du, Ning Lu, and Xianfeng Wang. Ntable: a dataset for camera-based table detection. In ICDAR, pages 117-129. Springer, 2021. 3" + ], + "bbox": [ + 94, + 92, + 903, + 669 + ], + "page_idx": 11 + }, + { + "type": "page_number", + "text": "15032", + "bbox": [ + 480, + 945, + 517, + 955 + ], + "page_idx": 11 + } +] \ No newline at end of file diff --git a/2025/DocSAM_ Unified Document Image Segmentation via Query Decomposition and Heterogeneous Mixed Learning/cd34a0c4-c9ec-4e5a-a2a8-007f0052f160_model.json b/2025/DocSAM_ Unified Document Image Segmentation via Query Decomposition and Heterogeneous Mixed Learning/cd34a0c4-c9ec-4e5a-a2a8-007f0052f160_model.json new file mode 100644 index 0000000000000000000000000000000000000000..fa5906ae25ec1f6ca19bb2129db2320a49bad4a9 --- /dev/null +++ b/2025/DocSAM_ Unified Document Image Segmentation via Query Decomposition and Heterogeneous Mixed Learning/cd34a0c4-c9ec-4e5a-a2a8-007f0052f160_model.json @@ -0,0 +1,2798 @@ +[ + [ + { + "type": "header", + "bbox": [ + 0.107, + 0.003, + 0.182, + 0.043 + ], + "angle": 0, + "content": "CVF" + }, + { + "type": "header", + "bbox": [ + 0.237, + 0.001, + 0.812, + 0.047 + ], + "angle": 0, + "content": "This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore." + }, + { + "type": "title", + "bbox": [ + 0.096, + 0.13, + 0.903, + 0.177 + ], + "angle": 0, + "content": "DocSAM: Unified Document Image Segmentation via Query Decomposition and Heterogeneous Mixed Learning" + }, + { + "type": "text", + "bbox": [ + 0.144, + 0.203, + 0.856, + 0.294 + ], + "angle": 0, + "content": "Xiao-Hui Li\\(^{1}\\), Fei Yin\\(^{1}\\), Cheng-Lin Liu\\(^{1,2}\\) \n\\(^{1}\\)MAIS, Institute of Automation of Chinese Academy of Sciences, Beijing, 100190, China \n\\(^{2}\\)School of Artificial Intelligence, University of Chinese Academy of Sciences, Beijing, 100049, China \n{xiaohui.li, fyin, liucl}@nlpr.ia.ac.cn" + }, + { + "type": "title", + "bbox": [ + 0.248, + 0.327, + 0.327, + 0.344 + ], + "angle": 0, + "content": "Abstract" + }, + { + "type": "text", + "bbox": [ + 0.093, + 0.36, + 0.485, + 0.768 + ], + "angle": 0, + "content": "Document image segmentation is crucial for document analysis and recognition but remains challenging due to the diversity of document formats and segmentation tasks. Existing methods often address these tasks separately, resulting in limited generalization and resource wastage. This paper introduces DocSAM, a transformer-based unified framework designed for various document image segmentation tasks, such as document layout analysis, multi-granularity text segmentation, and table structure recognition, by modelling these tasks as a combination of instance and semantic segmentation. Specifically, DocSAM employs Sentence-BERT to map category names from each dataset into semantic queries that match the dimensionality of instance queries. These two sets of queries interact through an attention mechanism and are cross-attended with image features to predict instance and semantic segmentation masks. Instance categories are predicted by computing the dot product between instance and semantic queries, followed by softmax normalization of scores. Consequently, DocSAM can be jointly trained on heterogeneous datasets, enhancing robustness and generalization while reducing computational and storage resources. Comprehensive evaluations show that DocSAM surpasses existing methods in accuracy, efficiency, and adaptability, highlighting its potential for advancing document image understanding and segmentation across various applications. Codes are available at https://github.com/xhli-git/DocSAM." + }, + { + "type": "title", + "bbox": [ + 0.092, + 0.799, + 0.222, + 0.815 + ], + "angle": 0, + "content": "1. Introduction" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.825, + 0.484, + 0.903 + ], + "angle": 0, + "content": "Document image segmentation (DIS) is a fundamental task in the field of document analysis and recognition (DAR) [42], serving as a cornerstone for downstream applications such as text recognition, information extraction (IE), and document visual question answering (DocVQA). Despite" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.328, + 0.907, + 0.48 + ], + "angle": 0, + "content": "its importance, DIS faces significant challenges due to the wide diversity of document types, page layouts, content annotations, and structural complexities, see Fig. 1. Existing approaches often address specific aspects of DIS separately, such as layout analysis, text detection, and table structure recognition, leading to specialized and fragmented solutions tailored to particular applications. This fragmentation not only impedes the performance of individual tasks but also results in redundant computational and storage overheads, making them inefficient for large-scale deployment." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.486, + 0.909, + 0.76 + ], + "angle": 0, + "content": "To address the aforementioned challenges, this paper introduces DocSAM (Document Segment Anything Model), a transformer-based unified framework designed to simultaneously handle various document image segmentation tasks, thereby eliminating the need for separate models and enhancing overall efficiency. As illustrated in Fig. 2, DocSAM comprises four primary modules: the Vision Backbone, the Deformable Encoder, Sentence-BERT [64], and the Hybrid Query Decoder (HQD). Given a document image and desired instance or semantic class names in natural text format, DocSAM first extracts multi-scale image features using the Vision Backbone. These features are then refined by the Deformable Encoder, which includes several deformable attention layers [108]. Class names are fed into Sentence-BERT and mapped to semantic queries. Subsequently, both semantic queries and learnable instance queries pass together through the HQD, where they interact to jointly perform semantic and instance segmentation." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.765, + 0.909, + 0.903 + ], + "angle": 0, + "content": "Inside each HQD layer (see Fig. 2), semantic and instance queries are concatenated and passed through a multi-head self-attention layer followed by a feed-forward layer for information exchange. These queries are then separately cross-attended with multi-scale image features in a coarse-to-fine manner using two multi-scale decoders, each with \\( L = 4 \\) layers. They further interact via another multi-head self-attention and feed-forward layer. The resulting semantic and instance queries, along with fused multi-scale image" + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.518, + 0.957 + ], + "angle": 0, + "content": "15021" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.098, + 0.088, + 0.237, + 0.359 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.237, + 0.088, + 0.508, + 0.359 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.51, + 0.089, + 0.752, + 0.359 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.756, + 0.089, + 0.902, + 0.359 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.238, + 0.369, + 0.757, + 0.384 + ], + "angle": 0, + "content": "Figure 1. Examples of various segmentation tasks on heterogeneous document datasets." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.41, + 0.483, + 0.485 + ], + "angle": 0, + "content": "features, are forwarded to the Mask Predictor, Class Predictor, and BBox Predictor for semantic mask segmentation, instance mask segmentation, category classification, and bounding box regression, respectively. We stack \\(K\\) HQD layers for more refined predictions." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.492, + 0.483, + 0.675 + ], + "angle": 0, + "content": "This design ensures that DocSAM can effectively manage the heterogeneity of document types, annotation formats, and segmentation tasks while maintaining high efficiency and accuracy. Extensive experiments and evaluations on various datasets demonstrate that DocSAM surpasses existing methods in accuracy, efficiency, and adaptability. Our results highlight DocSAM's potential as a powerful tool for advancing document image segmentation and understanding, with applications spanning from modern and historical document layout analysis to table structure decomposition, handwritten and scene text detection, and beyond. Our contributions are summarized as follows:" + }, + { + "type": "text", + "bbox": [ + 0.091, + 0.689, + 0.483, + 0.765 + ], + "angle": 0, + "content": "- We introduce DocSAM, a unified solution for diverse document image segmentation tasks such as layout analysis, multi-grained text segmentation, and table structure decomposition, reducing the need for specialized models and enhancing overall efficiency;" + }, + { + "type": "text", + "bbox": [ + 0.091, + 0.765, + 0.483, + 0.811 + ], + "angle": 0, + "content": "- By training on various tasks and datasets, DocSAM improves robustness and generalization, making it highly effective in handling varied document types and structures;" + }, + { + "type": "text", + "bbox": [ + 0.091, + 0.811, + 0.483, + 0.856 + ], + "angle": 0, + "content": "- Compared to specialized models, DocSAM significantly reduces computational and storage requirements, making it more practical for large-scale deployment;" + }, + { + "type": "text", + "bbox": [ + 0.091, + 0.856, + 0.483, + 0.902 + ], + "angle": 0, + "content": "- Extensive experiments on various datasets show that Doc-SAM outperforms current methods in terms of accuracy, efficiency, and adaptability." + }, + { + "type": "list", + "bbox": [ + 0.091, + 0.689, + 0.483, + 0.902 + ], + "angle": 0, + "content": null + }, + { + "type": "title", + "bbox": [ + 0.514, + 0.409, + 0.663, + 0.424 + ], + "angle": 0, + "content": "2. Related Works" + }, + { + "type": "title", + "bbox": [ + 0.514, + 0.441, + 0.73, + 0.456 + ], + "angle": 0, + "content": "2.1. DIS Tasks and Datasets" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.471, + 0.907, + 0.713 + ], + "angle": 0, + "content": "Depending on specific application scenarios, DIS involves various sub-tasks including Document Layout Analysis (DLA), Multi-Granularity Text Detection (MGTD), and Table Structure Recognition (TSR). DLA aims at identifying and categorizing page regions including text blocks, figures and tables [9, 16, 36, 60, 106]. This foundational step provides a structured overview of the document's layout, enabling more precise processing in subsequent tasks. MGTD focuses on detecting and segmenting text at various granularities, from paragraphs down to individual lines and words [11, 43, 53, 93, 95]. MGTD is a prerequisite for accurate Optical Character Recognition (OCR) tasks. TSR specifically aims to extract the structural of tables, including rows, columns and cells [17, 35, 38, 72, 107]. By decomposing tables into substructures, TSR facilitates the extraction and analysis of tabular information from documents." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.72, + 0.909, + 0.903 + ], + "angle": 0, + "content": "Along with these tasks, plenty of datasets have been accumulated after decades of research, see Tab. 1. These datasets exhibit great diversity and heterogeneity in data sources, document types, annotation formats, writing languages, category sets and many other aspects. For example, PubLayNet [106] contains born-digital English PDF documents with region-level annotations; SCUT-CAB [8] and MTHv2 [53] contains scanned historical Chinese documents with region, line and char-level annotations; SCUT-HCCDoc [95] contains handwritten documents with line-level annotations; CTW1500 [43] and Total-Text [11] contain natural scene images with texts of arbitrary shapes." + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "15022" + } + ], + [ + { + "type": "table", + "bbox": [ + 0.104, + 0.089, + 0.472, + 0.332 + ], + "angle": 0, + "content": "
TaskDataset
Document Layout AnalysisBaDLAD [71], CDLA [3], D4LA [13], DocBank [36], DocLayNet [60], ICDAR2017-POD [16], IIIT-AR-13K [55], M6Doc [9], PubLayNet [106], RanLayNet [2]
Ancient and Handwritten Document SegmentationCASIA-AHCDB [85], CHDAC-2022 [31], ICDAR2019-HDRC [68], SCUT-CAB [8], MTHv2 [53], HJDataset [69], CASIA-HWDB [41], SCUT-HCCDoc [95]
Table Structure RecognitionFinTabNet [104], ICDAR2013 [18], ICDAR2017-POD [16, 38], ICDAR2019-cTDr [17, 38], NTable [109], PubTables-1M [72], PubTabNet [107], STDW [19], TableBank [35], TNCR [1], WTW [47]
Scene Text DetectionCASIA-10k [22], COCO-Text [76], CTW1500 [43], CTW-Public [92], HUST-TR400 [87], ICDAR2015 [28], ICDAR2017-RCTW [70], ICDAR2017-MLT [56], ICDAR2019-ArT [12], ICDAR2019-LSVT [74], ICDAR2019-MLT [57], ICDAR2019-ReCTS [100], ICDAR2023-HierText [49], ICDAR2023-ReST [91], ICPR2018-MTWI [21], MSRA-TD500 [86], ShopSign [94], Total-Text [11], USTB-SV1K [90]
" + }, + { + "type": "table_caption", + "bbox": [ + 0.168, + 0.341, + 0.405, + 0.354 + ], + "angle": 0, + "content": "Table 1. Datasets involved in DocSAM." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.388, + 0.304, + 0.405 + ], + "angle": 0, + "content": "2.2. Deep Learning for DIS" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.414, + 0.484, + 0.701 + ], + "angle": 0, + "content": "Existing deep learning based DIS methods basically focus on specific sub-tasks and datasets. Generally speaking, they usually transform various DIS tasks into general object detection or image segmentation problems and make some modifications to general object detection [20, 62, 63, 65] and image segmentation methods [5, 46, 66] to make them more suitable for the tasks and datasets at hand. Some other works treat documents as hierarchical graph structures and adopt graph models like GNN [82] and CRF [32] for the task of layout analysis [37, 52], table structure recognition [10, 38], and text detection [48, 101]. Though more flexible, these methods usually suffer from complicated pre/post-processing steps and are more susceptible to intermediate errors. There are also some multi-modal based methods that combine visual and textual features like LayoutLMv3 [27], DiT [34], and VGT [13]. These methods improve the performance and generalization by pre-training on large-scale unsupervised documents to align text and visual features, but are often slower due to the complexity of architectures." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.705, + 0.484, + 0.901 + ], + "angle": 0, + "content": "With the prosperity of large language models (LLMs) [102], many large document models are proposed such as UDOP [75], UniDoc [15], DocPedia [14], DocLLM [80], TextMonkey [44], mPLUG-DocOwl [23, 88], etc. Though promising results can be achieved for the DocVQA task, lacking fine-grained intermediate outputs like text locations and page layouts still greatly limits the interpretability and generalization of these models. As compensation, recently some LLM-free unified models are proposed for low-level document processing tasks such as UPOCR [59], DocRes [98], OmniParser [78] and DAT [79]. These works unify several similar or related tasks into unified models through multi-task learning, but at the cost of significant increment" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.092, + 0.907, + 0.123 + ], + "angle": 0, + "content": "of model complexity and calculating overhead, prohibiting them from generalizing to to more tasks and datasets." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.134, + 0.895, + 0.15 + ], + "angle": 0, + "content": "2.3. Transformer-based Detection&Segmentation" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.156, + 0.907, + 0.307 + ], + "angle": 0, + "content": "Following the pioneer work of DETR [4], many Transformer-based objection methods have been proposed in recent years, including Deformable DETR [108], DNDETR [33], DINO [96], Sparse R-CNN [73], etc. These methods share the same idea with DETR that rely on learnable queries and bipartite matching for object decoding, but make different modifications to improve the accuracy and convergence speed, such as bringing in deformable attention and denoising training, or assigning specific spatial meanings to the queries." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.309, + 0.909, + 0.596 + ], + "angle": 0, + "content": "Besides object detection, Transformer also shows great potential in image segmentation [6, 7, 24, 25, 30, 83, 97, 103]. Among them the most related works to this paper are SAM [30] and Mask2former [7]. Inspired by SAM [30] which uses natural language prompts to guide image segmentation, in this paper we propose to embed the class names of each dataset into semantic queries and transform various document segmentation tasks into a combination of instance segmentation and semantic segmentation. The semantic queries not only serve as prompts guiding the model in identifying specific types of regions, but also function as class prototypes that instance queries depend on for classification. Since DIS relies on high resolution image features, we build our DocSAM from Mask2former [7] which adopt Swin-Transformer [45] and deformable attention [108] as the vision encoder. Though the vision encoder of DocSAM is inherited from Mask2former, the decoder is drastically redesigned to be up to the task of general document image segmentation effectively." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.612, + 0.615, + 0.627 + ], + "angle": 0, + "content": "3. DocSAM" + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.637, + 0.655, + 0.652 + ], + "angle": 0, + "content": "3.1. Preliminaries" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.659, + 0.909, + 0.903 + ], + "angle": 0, + "content": "Before introducing the proposed DocSAM, we first explore the key attributes of an ideal all-in-one DIS model and why current methods fall short of this goal. We assert that an exemplary all-in-one DIS model should possess the following attributes: it should have the versatility to convert diverse DIS tasks into a unified framework; it should be adaptable to training on heterogeneous datasets, accommodating diverse annotations without restrictions; and it should maintain the capacity for continual and incremental learning. Current methods are typically designed for specific DIS tasks and datasets, and most models become static after training, unable to efficiently incorporate new data, limiting their versatility and adaptability. Specifically, existing DIS methods rely on fully connected (FC) and Softmax layers to predict region classes, with FC's parameters predefined for specific tasks and datasets, making generalization difficult." + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.518, + 0.957 + ], + "angle": 0, + "content": "15023" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.113, + 0.088, + 0.885, + 0.334 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.09, + 0.344, + 0.908, + 0.374 + ], + "angle": 0, + "content": "Figure 2. Network structure of the proposed DocSAM. DocSAM unify various document image segmentation tasks into one single model through instance and semantic query decomposition and interaction. Skip connections and norm layers are omitted for simplicity." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.399, + 0.484, + 0.565 + ], + "angle": 0, + "content": "To overcome the above limitations and achieve the aforementioned criteria, the proposed DocSAM makes two significant improvements compared to existing methods. First, it transforms various DIS tasks into a unified paradigm of mask-based instance segmentation and semantic segmentation. Second, it embeds class names into semantic queries, which not only serve as prompts to guide the model in identifying specific types of regions to segment but also function as class prototypes that instance queries depend on for classification. The rest of this chapter presents the details of our proposed DocSAM." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.575, + 0.245, + 0.59 + ], + "angle": 0, + "content": "3.2. Vision Encoder" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.597, + 0.484, + 0.824 + ], + "angle": 0, + "content": "Different DIS tasks may focus on contents of different scales, from large objects like paragraphs and figures spanning entire pages to tiny objects like chars and words covering only a few hundreds of pixels. Therefore, high-resolution multi-scale image features are an essential requirement for a unified DIS model. The vision encoder of DocSAM is adapted from Mask2Former [7], which includes a Swin-Transformer [45] as the vision backbone and deformable attention [108] for feature refinement. Additionally, we use another FPN [40] to fuse multi-scale image features \\( X_{I} = [X_{I}^{l}\\in \\mathbb{R}^{H_{l}W_{l}\\times C},l\\in \\{1,2,3,4\\}] \\) into a single mask feature \\( X_{M}\\in \\mathbb{R}^{HW\\times C} \\), which is used for subsequent semantic segmentation and instance segmentation. Here, \\( X_{I}^{l} \\) is image feature of level \\( l \\), \\( H_{l} \\) and \\( W_{l} \\) are the spatial resolution of level \\( l \\), \\( C \\) is number of feature channels." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.833, + 0.271, + 0.85 + ], + "angle": 0, + "content": "3.3. Query Embedding" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.855, + 0.484, + 0.901 + ], + "angle": 0, + "content": "The instance queries \\( Q_{I} \\in \\mathbb{R}^{N \\times C} \\) of DocSAM is standard learnable queries, while the semantic queries \\( Q_{S} \\in \\mathbb{R}^{M \\times C} \\) are embedded from class names using the Sentence-BERT" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.399, + 0.907, + 0.505 + ], + "angle": 0, + "content": "[64]. Here \\( N \\) is a predefined instance query number that remains the same across all tasks and datasets, while \\( M \\) is the semantic query number that may change depending on the class number of each dataset, and \\( C \\) is feature dimension. \\( Q_{I} \\) and \\( Q_{S} \\) go together through the following Hybrid Query Decoder for feature decoding and cooperate with each other for semantic segmentation and instance segmentation." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.515, + 0.727, + 0.531 + ], + "angle": 0, + "content": "3.4. Hybrid Query Decoder" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.537, + 0.907, + 0.793 + ], + "angle": 0, + "content": "Inside each HQD layer, see Fig. 2, we first concatenate \\( Q_{S} \\) and \\( Q_{I} \\) along the length dimension and send them into a multi-head self-attention layer (MHSA) followed by a feed forward layer (FFN). This step facilitates information exchange between \\( Q_{S} \\) and \\( Q_{I} \\), allowing them to attend to each other for query fusion. Next, they are separately crossattended with the multi-scale image features \\( X_{I} \\) in a coarse-to-fine manner by two Multi-Scale Decoders (MSD) each containing \\( L \\) layers. Here, \\( L = 4 \\) stands for the number of feature scales. Each MSD layer consists of two MHSA layers, one multi-head cross-attention layer (MHCA) and one FFN layer. Following Mask2Former [7], we also use masked attention in MHCA, where the attention masks are derived from the predicted instance and semantic masks in the previous HQD layer. After that, \\( Q_{S} \\) and \\( Q_{I} \\) further interact with each other through another MHSA and FFN layer. We stack \\( K \\) HQD layers for more refined predictions." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.803, + 0.676, + 0.818 + ], + "angle": 0, + "content": "3.5. Prediction Head" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.825, + 0.906, + 0.901 + ], + "angle": 0, + "content": "The output \\( Q_{S} \\) and \\( Q_{I} \\) from each HQD layer along with the mask feature \\( X_{M} \\) are sent to the Mask Predictor, Class Predictor and BBox Predictor for semantic mask segmentation, instance mask segmentation, instance category classification and instance bounding box regression, respectively." + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "15024" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.091, + 0.092, + 0.483, + 0.122 + ], + "angle": 0, + "content": "For predicting semantic and instance masks, \\( Q_{S} \\) and \\( Q_{I} \\) are multiplied with \\( X_{M} \\) as:" + }, + { + "type": "equation", + "bbox": [ + 0.213, + 0.133, + 0.483, + 0.151 + ], + "angle": 0, + "content": "\\[\nM _ {S} = \\sigma \\left(Q _ {S} \\times X _ {M} ^ {T}\\right), \\tag {1}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.091, + 0.163, + 0.121, + 0.176 + ], + "angle": 0, + "content": "and" + }, + { + "type": "equation", + "bbox": [ + 0.215, + 0.175, + 0.482, + 0.194 + ], + "angle": 0, + "content": "\\[\nM _ {I} = \\sigma \\left(Q _ {I} \\times X _ {M} ^ {T}\\right), \\tag {2}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.198, + 0.484, + 0.276 + ], + "angle": 0, + "content": "where \\(M_S \\in \\mathbb{R}^{M \\times HW}\\) and \\(M_I \\in \\mathbb{R}^{N \\times HW}\\) are predicted semantic and instance masks, \\(\\sigma\\) is Sigmoid function, \\(T\\) means matrix transposition, and \\(\\times\\) stands for matrix multiplication. Similarly, for predicting instance classes, \\(Q_I\\) is multiplied with \\(Q_S\\) as:" + }, + { + "type": "equation", + "bbox": [ + 0.197, + 0.287, + 0.483, + 0.305 + ], + "angle": 0, + "content": "\\[\nY _ {I} = \\operatorname {S o f t m a x} \\left(Q _ {I} \\times Q _ {S} ^ {T}\\right), \\tag {3}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.315, + 0.483, + 0.376 + ], + "angle": 0, + "content": "where \\( Y_{I} \\in \\mathbb{R}^{N \\times M} \\) is predicted class probabilities of instances, softmax is Softmax function along the second dimension of \\( Y_{I} \\), \\( T \\) means matrix transposition, and \\( \\times \\) stands for matrix multiplication." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.377, + 0.484, + 0.496 + ], + "angle": 0, + "content": "Since Eq. (1), Eq. (2), Eq. (3) are all based on matrix multiplication, they are actually calculating the similarities between \\( Q_{S} \\), \\( Q_{I} \\) and \\( X_{M} \\). So we can also regard the \\( Q_{S} \\), \\( Q_{I} \\) as instance and semantic prototypes. Through the above semantic query embedding and prototype-based instance classification, we transform the original close-set classifier into an open-set classifier, thus benefiting the construction of unified all-in-one DIS model." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.497, + 0.484, + 0.588 + ], + "angle": 0, + "content": "Besides mask segmentation, DocSAM also keep the ability of bbox prediction. This is realized through bounding box regression with the BBox Predictor. Following ISTR [24] and TransDLANet [9], for each HQD layer we predict the residual values of bbox coordinates relative to predictions of the previous HQD layer." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.597, + 0.252, + 0.613 + ], + "angle": 0, + "content": "3.6. Model Learning" + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.62, + 0.236, + 0.634 + ], + "angle": 0, + "content": "3.6.1. Loss Function" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.638, + 0.484, + 0.699 + ], + "angle": 0, + "content": "There are four losses in DocSAM, namely semantic mask segmentation loss \\( L_{S} \\), instance mask segmentation loss \\( L_{I} \\), instance bbox regression loss \\( L_{B} \\), and instance classification loss \\( L_{C} \\). Among them, \\( L_{S} \\) is calculated as:" + }, + { + "type": "equation", + "bbox": [ + 0.115, + 0.71, + 0.483, + 0.729 + ], + "angle": 0, + "content": "\\[\nL _ {S} = \\lambda_ {f} L _ {f o c a l} (M _ {S}, \\hat {M} _ {S}) + \\lambda_ {d} L _ {d i c e} (M _ {S}, \\hat {M} _ {S}), \\quad (4)\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.74, + 0.484, + 0.802 + ], + "angle": 0, + "content": "where \\(M_S\\) and \\(\\hat{M}_S\\) are predicted and ground-truth semantic masks, \\(L_{focal}\\) and \\(L_{dice}\\) are focal loss [67] and dice loss [54], respectively, and \\(\\lambda_f = 10\\) and \\(\\lambda_d = 1\\) are hyperparameters. Similarly, \\(L_I\\) is calculated as:" + }, + { + "type": "equation", + "bbox": [ + 0.12, + 0.812, + 0.483, + 0.831 + ], + "angle": 0, + "content": "\\[\nL _ {I} = \\lambda_ {f} L _ {f o c a l} (M _ {I}, \\hat {M} _ {I}) + \\lambda_ {d} L _ {d i c e} (M _ {I}, \\hat {M} _ {I}), \\tag {5}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.843, + 0.483, + 0.873 + ], + "angle": 0, + "content": "where \\(M_S\\) and \\(\\hat{M}_S\\) are predicted and ground-truth instance masks, respectively. \\(L_B\\) is calculated as:" + }, + { + "type": "equation", + "bbox": [ + 0.116, + 0.884, + 0.483, + 0.903 + ], + "angle": 0, + "content": "\\[\nL _ {B} = \\lambda_ {s l 1} L _ {s l 1} (B _ {I}, \\hat {B} _ {I}) + \\lambda_ {d i o u} L _ {d i o u} (B _ {I}, \\hat {B} _ {I}), \\tag {6}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.092, + 0.906, + 0.152 + ], + "angle": 0, + "content": "where \\( B_{I} \\) and \\( \\hat{B}_I \\) are predicted and ground-truth bounding boxes, \\( L_{sl1} \\) and \\( L_{diou} \\) are smooth L1 loss and distance IoU loss [105], respectively, and \\( \\lambda_{sl1} = 1 \\) and \\( \\lambda_{diou} = 1 \\) are hyper-parameters. At last, \\( L_{C} \\) is calculated as:" + }, + { + "type": "equation", + "bbox": [ + 0.644, + 0.163, + 0.906, + 0.181 + ], + "angle": 0, + "content": "\\[\nL _ {C} = L _ {c e} \\left(Y _ {I}, \\hat {Y} _ {I}\\right), \\tag {7}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.19, + 0.906, + 0.235 + ], + "angle": 0, + "content": "where \\( Y_{I} \\) and \\( \\hat{Y}_I \\) are predicted and ground-truth instance labels, and \\( L_{ce} \\) is cross entropy loss. The total loss of Doc-SAM is the sum of the above four losses:" + }, + { + "type": "equation", + "bbox": [ + 0.585, + 0.248, + 0.905, + 0.265 + ], + "angle": 0, + "content": "\\[\nL = \\lambda_ {s} L _ {S} + \\lambda_ {i} L _ {I} + \\lambda_ {b} L _ {B} + \\lambda_ {c} L _ {C}, \\tag {8}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.275, + 0.907, + 0.397 + ], + "angle": 0, + "content": "where \\(\\lambda_s = 5\\), \\(\\lambda_i = 5\\), \\(\\lambda_b = 1\\), and \\(\\lambda_c = 1\\) are hyperparameters. We add auxiliary losses to every HQD layer and to query features before HQD. Following DETR [4] and Mask2Former [7], we also use bipartite matching to find the best matched instance predictions before calculating the loss. While for semantic predictions, there is no need to perform the bipartite matching, because the predictions and ground-truths are already one-to-one matched." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.403, + 0.78, + 0.419 + ], + "angle": 0, + "content": "3.6.2. Heterogeneous Mixed Learning" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.421, + 0.907, + 0.634 + ], + "angle": 0, + "content": "Unlike existing methods, the novel design of DocSAM enables us to train a single model on heterogeneous mixed datasets. In this work, we collected nearly fifty DIS datasets of various document types and annotation formats, covering diverse DIS tasks from layout analysis and text detection to table structure recognition (see Tab. 1). We combined these datasets to construct a heterogeneous mixed dataset for training DocSAM. After training, the DocSAM model can be directly used as a versatile document segmenter or as a pre-trained model that can be seamlessly fine-tuned using task-specific datasets without any specialized modifications, such as adding or replacing a linear classification layer. This merit of DocSAM endows it with the potential for continual and incremental learning." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.641, + 0.771, + 0.656 + ], + "angle": 0, + "content": "3.6.3. Improving Training Efficiency" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.659, + 0.907, + 0.901 + ], + "angle": 0, + "content": "Directly training DocSAM on such heterogeneous datasets may suffer from slow convergence and long training time, so we propose several strategies to improve training efficiency. Firstly, we pre-train the vision encoder of DocSAM on all 48 datasets using SimMIM [84], hoping it can provide more robust visual features for document images. Secondly, we separate the training datasets into groups with each group containing datasets of similar tasks and styles, see Tab. 1, then we adopt curriculum learning (CL) [81] strategy to warm up the training process by gradually adding new group of datasets. Thirdly, we add an instance query selection (IQS) process at the front of each HQD layer. Motivation behind this is that bipartite matching only calculates losses between matched predictions and ground-truths, and the matched query indexes are mostly the same across HQD layers. For a certain document, large ratio of instance" + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "15025" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.134, + 0.088, + 0.498, + 0.246 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.251, + 0.248, + 0.379, + 0.261 + ], + "angle": 0, + "content": "(a) Training loss curves." + }, + { + "type": "image", + "bbox": [ + 0.499, + 0.09, + 0.868, + 0.246 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.617, + 0.248, + 0.748, + 0.26 + ], + "angle": 0, + "content": "(b) On-the-fly validation." + }, + { + "type": "image_caption", + "bbox": [ + 0.31, + 0.273, + 0.684, + 0.287 + ], + "angle": 0, + "content": "Figure 3. Loss curves and on-the-fly validation during training." + }, + { + "type": "table", + "bbox": [ + 0.104, + 0.31, + 0.473, + 0.429 + ], + "angle": 0, + "content": "
Ablation SettingInstanceSemantic
mAP\\( mAP_b \\)mAFmIoU
DocSAMbase0.38040.35170.42560.6615
DocSAMbase + Curriculum0.38690.37040.43380.6622
DocSAMbase + SimMIM0.10020.06580.12940.3882
DocSAMbase + Freezing BERT0.38430.35100.42950.6677
DocSAMbase - Masked Attention0.23220.09380.28460.6327
DocSAMbase - Query Interaction0.29900.17790.35090.6511
DocSAMbase - Query Selection0.33410.31340.37630.6592
DocSAMlarge0.39000.36580.43200.6726
" + }, + { + "type": "table_caption", + "bbox": [ + 0.091, + 0.439, + 0.482, + 0.454 + ], + "angle": 0, + "content": "Table 2. Ablation studies on model structure and training strategy." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.479, + 0.484, + 0.614 + ], + "angle": 0, + "content": "queries are not activated from beginning to end, and their class scores are very low. Therefore, we only select instance queries whose class scores are higher than a threshold \\( T_{k} \\) before the \\( k \\)-th HQD layer. We set \\( T_{k} \\) as: \\( T_{k} = T_{max} / 2^{K - k} \\), where \\( T_{max} = 0.01 \\) is the maximum threshold, \\( K \\) is the number of HQD layers, \\( k \\) is the current HQD layer. Experiments show that IQS can discard low-score queries without degrading model performance, thereby improving training speed and reducing memory usage." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.628, + 0.222, + 0.645 + ], + "angle": 0, + "content": "4. Experiments" + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.653, + 0.289, + 0.667 + ], + "angle": 0, + "content": "4.1. Datasets and Metrics" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.675, + 0.483, + 0.901 + ], + "angle": 0, + "content": "The datasets involved in our experiments are listed in Tab. 1. Underlined datasets (15 in total) are used for ablation studies, mixed pre-training, and dataset-specific fine-tuning. All 48 datasets are used for training the final DocSAM model. These datasets cover a wide range of domains and tasks, showing significant heterogeneity in document types, annotation formats, and other aspects. Typical examples are shown in Fig. 1, with more details provided in the supplementary material. For evaluation metrics, we use mIoU [46] for semantic segmentation and mAP (for masks) [39] and \\(\\mathrm{mAP}_b\\) (for bounding boxes) [39] for instance segmentation. Additionally, we introduce a new metric for instance segmentation called mAF, which is calculated as the mean F-score of all classes across all IoUs ranging from 0.5 to 0.95 in increments of 0.05 (i.e., [0.5:0.05:0.95])." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.313, + 0.731, + 0.329 + ], + "angle": 0, + "content": "4.2. Implementation Details" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.335, + 0.907, + 0.501 + ], + "angle": 0, + "content": "The vision backbone and deformable attention module are initialized from Mask2Former [7], which is pre-trained on the COCO-panoptic dataset [39]. The Sentence-BERT is initialized using the all-MiniLM-L6-v2 model from the Sentence Transformers library [64]. Other parts of DocSAM are randomly initialized. We trained two sizes of models: DocSAM-base (207M parameters) and DocSAM-large (317M parameters). Their vision backbones use Swin-base and Swin-large, respectively, and the instance query numbers \\(N\\) are set to 500 and 900, respectively. The HQD layer number \\(K\\) is set to 4 by default." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.502, + 0.909, + 0.652 + ], + "angle": 0, + "content": "DocSAM is implemented based on PyTorch [58] and trained on \\(8 \\times\\) NVIDIA A800 GPUs. We use the AdamW optimizer [50] to train the model, setting the base learning rate to \\(4 \\times 10^{-5}\\), and decay it using cosine annealing strategy [51]. For joint training on mixed datasets, the default settings are 80,000 iterations and a batch size of 32; for ablation studies and dataset-specific fine-tuning, the defaults are 20,000 iterations and a batch size of 8; for comparison with state-of-the-art, the defaults are 40,000 iterations and a batch size of 16." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.663, + 0.652, + 0.677 + ], + "angle": 0, + "content": "4.3. Main Results" + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.685, + 0.676, + 0.699 + ], + "angle": 0, + "content": "4.3.1. Ablation Studies" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.704, + 0.909, + 0.901 + ], + "angle": 0, + "content": "To verify the effect of each module in DocSAM and select the best training strategy before large-scale training, we conducted a series of ablation studies, as shown in Tab. 2 and Fig. 3. The results in Tab. 2 are averaged over all 15 datasets. On-the-fly validation involves fast testing on a small number of samples (e.g. 10 for each dataset) during training. The results show that using curriculum learning and instance query selection can accelerate convergence and improve model performance, while SimMIM pre-training significantly degrades model performance, possibly due to the large gap between SimMIM and document segmentation. Since freezing the weights of Sentence-BERT has almost no impact on performance, we freeze them during" + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "15026" + } + ], + [ + { + "type": "table", + "bbox": [ + 0.13, + 0.085, + 0.86, + 0.324 + ], + "angle": 0, + "content": "
TaskDatasetDocSAM-baseDocSAM-large
InstanceSemanticInstanceSemantic
AP50AP75mAP\\( mAP_b \\)mAFmIoUAP50AP75mAP\\( mAP_b \\)mAFmIoU
Document Layout Analysis\\( D^4 \\)LA [13]0.5950.5140.4480.4380.4860.3890.6370.5620.4900.4730.5390.434
DocLayNet [60]0.7160.5280.4840.4800.5430.6070.7440.5700.5170.5010.5840.669
\\( M^6 \\)Doc [9]0.5190.4020.3630.3520.3810.2670.5510.4440.3970.3870.4250.296
PubLayNet [106]0.9360.8620.8060.7890.8470.8980.9460.8840.8300.8050.8680.911
Ancient and Handwritten Document SegmentationSCUT-CAB-logical [8]0.6810.5550.4810.4780.5020.4100.7170.5740.5110.4950.5340.454
SCUT-CAB-physical [8]0.9370.8370.7770.7470.8210.9370.9480.8560.7860.7540.8290.942
HJDataset [69]0.9560.9210.8810.8650.8950.8190.9560.9250.8850.8690.8980.821
CASIA-HWDB [41]0.9290.7850.7210.6640.7880.9350.9120.7700.7140.6430.7900.939
SCUT-HCCDoc [95]0.8650.6350.5440.5590.6250.8440.8690.6420.5490.5600.6250.847
Table Structure RecognitionFinTabNet [104]0.8670.7700.6640.6270.7570.8510.8690.7860.6840.6440.7780.860
PubTabNet [107]0.9700.7880.6430.6350.7140.8400.9700.7890.6480.6340.7230.845
TableBank-latex [35]0.9630.9470.8970.8680.9240.9400.9650.9500.9150.8930.9360.951
TableBank-word [35]0.8730.8370.8220.7930.8510.8440.8780.8440.8350.8140.8570.853
Scene Text DetectionCTW1500 [43]0.7120.4300.4000.3680.5000.7940.7530.4820.4410.4020.5310.817
Total-Text [11]0.7470.4210.4050.4070.7430.4530.7690.4540.4280.4210.4720.764
MSRA-TD500 [86]0.7470.5250.4580.4770.5020.7130.7980.5770.4960.5160.5320.739
ICDAR2015 [28]0.6130.2470.2940.3020.3380.5990.6390.2600.3070.3130.3450.623
" + }, + { + "type": "table_caption", + "bbox": [ + 0.326, + 0.338, + 0.67, + 0.353 + ], + "angle": 0, + "content": "Table 3. Performance of DocSAM after joint pre-training." + }, + { + "type": "table", + "bbox": [ + 0.134, + 0.366, + 0.865, + 0.632 + ], + "angle": 0, + "content": "
TaskDatasetDocSAM-large from scratchDocSAM-large from pretrain
InstanceSemanticInstanceSemantic
AP50AP75mAP\\( mAP_b \\)mAFmIoUAP50AP75mAP\\( mAP_b \\)mAFmIoU
DocumentLayoutAnalysis\\( D^4LA[13] \\)0.3650.2590.2390.1940.2330.2050.6980.6370.5550.5460.5950.526
DocLayNet [60]0.5030.2920.2950.2600.3590.3650.8330.6910.6210.6010.6790.736
\\( M^6Doc[9] \\)0.2790.1730.1690.1450.1630.0870.6670.5660.5000.4850.5280.430
PubLayNet [106]0.8730.7590.6960.6220.7380.8410.9540.9040.8540.8500.8880.921
Ancient andHandwrittenDocument SegmentationSCUT-CAB-logical [8]0.3910.2330.2390.1360.2280.2260.7830.6310.5560.5300.5820.481
SCUT-CAB-physical [8]0.8010.6440.6050.4050.6640.9180.9460.8690.7990.7620.8420.945
HJDataset [69]0.8480.8350.7520.6060.7770.8120.9830.9480.9050.8950.9110.822
CASIA-HWDB [41]0.9080.7370.6650.6280.7370.9490.9770.9390.8930.7920.9160.956
SCUT-HCCDoc [95]0.8070.4920.4600.4230.5410.8530.9040.6840.5800.5890.6580.862
Table StructureRecognitionFinTabNet [104]0.3350.1640.1780.0040.2220.7700.8770.8050.7130.6810.8030.870
PubTabNet [107]0.0130.0100.0070.0420.0070.8100.9730.8210.6690.6530.7420.860
TableBank-latex [35]0.7620.6120.5650.0200.6410.9130.9680.9540.9260.9090.9470.958
TableBank-word [35]0.5940.4460.4350.0450.6190.8230.9080.8770.8710.8590.8810.873
Scene Text DetectionCTW1500 [43]0.4310.0980.1620.0710.2530.8000.7940.5390.4800.4530.5730.831
Total-Text [11]0.3130.0420.0960.0470.1680.7490.7940.5170.4600.4660.5020.775
MSRA-TD500 [86]0.5060.1890.2220.0760.2950.7310.8090.6040.5240.5410.5550.744
ICDAR2015 [28]0.2030.0280.0630.0230.1130.5970.6810.3160.3410.3530.3790.641
Unseen DatasetIIIT-AR-13K [55]0.5550.4300.4030.1850.4170.7390.8420.7020.6380.6210.6930.642
CHDAC-2022 [31]0.8860.6960.6040.5090.6490.9150.9390.8280.6870.6250.7270.918
" + }, + { + "type": "table_caption", + "bbox": [ + 0.297, + 0.642, + 0.699, + 0.656 + ], + "angle": 0, + "content": "Table 4. Performance of DocSAM after dataset specific fine-tuning." + }, + { + "type": "text", + "bbox": [ + 0.086, + 0.68, + 0.479, + 0.815 + ], + "angle": 0, + "content": "training. Similar to Mask2Former, masked attention plays a crucial role in DocSAM, and removing it leads to a significant performance drop. Additionally, without query interaction, DocSAM's performance also decreases substantially, highlighting the importance of information exchange between instance and semantic queries. Finally, training a unified model on heterogeneous datasets heavily relies on the model's capacity, and using a more powerful vision backbone can greatly enhance model performance." + }, + { + "type": "title", + "bbox": [ + 0.09, + 0.834, + 0.339, + 0.85 + ], + "angle": 0, + "content": "4.3.2. Pre-training and Fine-tuning" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.856, + 0.482, + 0.901 + ], + "angle": 0, + "content": "We train DocSAM on mixed heterogeneous datasets (15 datasets) to validate its performance as a unified document segmenter and a pre-trained model for dataset-specific fine" + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.683, + 0.905, + 0.819 + ], + "angle": 0, + "content": "tuning. The results are shown in Tab. 3 and Tab. 4. DocSAM achieves good semantic and instance segmentation performance on various datasets and tasks, though performance may vary across datasets due to differing levels of difficulty. As a single-modal model, DocSAM may underperform on datasets like \\(\\mathrm{D}^4\\mathrm{LA}\\) [13], DocLayNet [60], \\(\\mathrm{M}^6\\mathrm{Doc}\\) [9] and SCUT-CAB-logical [8], which require multi-modal information for fine-grained logical layout analysis." + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.826, + 0.905, + 0.901 + ], + "angle": 0, + "content": "After joint training, we fine-tune DocSAM-large on each specific dataset to further improve performance. As shown in Tab. 4, fine-tuning results are significantly higher than direct testing and training from scratch. We also test DocSAM on unseen datasets IIIT-AR-13K [55] and CHDAC" + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.946, + 0.518, + 0.957 + ], + "angle": 0, + "content": "15027" + } + ], + [ + { + "type": "table", + "bbox": [ + 0.104, + 0.09, + 0.472, + 0.196 + ], + "angle": 0, + "content": "
MethodObjectInstance
mAPAP50AP75mAPAP50AP75
Faster R-CNN [65]0.4900.6780.5720.4780.6780.552
Mask R-CNN [20]0.4010.5840.4620.3970.5840.456
Deformable DETR [108]0.5720.7680.6340.5560.7650.611
ISTR [24]0.6270.8080.7080.6200.8070.702
TransDLANet [9]0.6450.8270.7270.6380.8260.719
DAT [79]0.712--0.657--
DocSAM0.6630.8400.7550.6610.8400.750
" + }, + { + "type": "table_caption", + "bbox": [ + 0.15, + 0.207, + 0.424, + 0.221 + ], + "angle": 0, + "content": "Table 5. Performance comparison on \\( {\\mathrm{M}}^{6} \\) Doc." + }, + { + "type": "table", + "bbox": [ + 0.104, + 0.236, + 0.472, + 0.433 + ], + "angle": 0, + "content": "
MethodObjectInstance
mAPAP50AP75mAPAP50AP75
Physical
Faster R-CNN [65]0.7750.9130.8610.7530.9100.834
Mask R-CNN [20]0.7910.9210.8770.7950.9170.872
SCNet [77]0.8130.9410.8900.8200.9410.891
Deformable DETR [108]0.7990.9230.8710.7790.9210.843
VSR [99]0.7870.9190.8600.7870.9190.852
DocSAM0.7740.9470.8600.8110.9480.891
Logical
Faster R-CNN [65]0.5490.7740.6130.5420.7730.606
Mask R-CNN [20]0.5510.7850.6190.5530.7770.631
SCNet [77]0.6020.8360.6730.6030.8360.680
Deformable DETR [108]0.6270.8520.7170.6200.8510.703
VSR [99]0.5570.7830.6160.5510.7820.611
DocSAM0.5480.7690.6320.5750.7790.667
" + }, + { + "type": "table_caption", + "bbox": [ + 0.137, + 0.444, + 0.436, + 0.457 + ], + "angle": 0, + "content": "Table 6. Performance comparison on SCUT-CAB." + }, + { + "type": "table", + "bbox": [ + 0.117, + 0.474, + 0.459, + 0.591 + ], + "angle": 0, + "content": "
MethodCTW1500Total-Text
PRFPRF
HierText [49]0.8460.8740.8600.8550.9050.879
SIR [61]0.8740.8370.8550.9090.8560.882
DPText-DETR [89]0.9170.8620.8880.9180.8640.890
UNITS [29]-----0.898
ESTextSpotter [26]0.9150.8860.9000.9200.8810.900
DAT-DET [79]0.8930.8930.8930.9400.8820.910
DAT-SEG [79]0.9250.9090.9170.9500.8920.920
DocSAM0.8050.8810.8420.7210.8260.770
" + }, + { + "type": "table_caption", + "bbox": [ + 0.097, + 0.602, + 0.477, + 0.616 + ], + "angle": 0, + "content": "Table 7. Performance comparison on CTW1500 and Total-Text." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.644, + 0.483, + 0.72 + ], + "angle": 0, + "content": "2022 [31], where fine-tuning from the pre-trained model also yields substantial performance gains. This demonstrates that DocSAM's performance is not yet saturated and can benefit greatly from transfer learning on unseen datasets and tasks." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.731, + 0.381, + 0.745 + ], + "angle": 0, + "content": "4.3.3. Comparison with State-of-the-Arts" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.75, + 0.483, + 0.902 + ], + "angle": 0, + "content": "To compare with state-of-the-art methods, we further fine-tuned DocSAM on some datasets for additional training iterations. The results are shown in Tab. 5, Tab. 6, and Tab. 7. The best results are shown in bold, and the second-best results are underlined. DocSAM achieves superior or comparable performance with other methods. Note that we did not apply any specific training techniques or data augmentation, configurations for all datasets were kept consistent. We found that DocSAM exhibits much lower performance in logical layout analysis compared to physical analysis," + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.092, + 0.905, + 0.183 + ], + "angle": 0, + "content": "which we attribute to its reliance only on single-modal features. Furthermore, DocSAM achieved relatively low performance on scene text detection datasets. This is likely because scene texts exhibit much greater diversity in shapes and backgrounds, requiring more carefully designed strategies to ensure model performance." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.192, + 0.633, + 0.206 + ], + "angle": 0, + "content": "4.4. Discussion" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.214, + 0.907, + 0.471 + ], + "angle": 0, + "content": "The goal of this paper is not to achieve state-of-the-art performance on specific dataset and task through meticulously designed model architectures or training strategies. Instead, we aim to design a simple and unified document segmentation model that can be applied to a wide variety of datasets and tasks. Additionally, the trained model should possess good scalability and the ability to continue learning. In this regard, DocSAM is quite successful. It exhibits decent performance on various datasets and tasks and shows great potential for downstream applications both as a versatile segmenter and a pre-trained model. However, experimental results also reveal some weaknesses and limitations of DocSAM, such as long training time and unsatisfactory performance on complex scenarios. We believe that DocSAM can greatly benefit from more sophisticated model design and better data augmentation and training strategies to further accelerate its convergence and improve its performance." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.484, + 0.634, + 0.5 + ], + "angle": 0, + "content": "5. Conclusion" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.509, + 0.907, + 0.812 + ], + "angle": 0, + "content": "In this paper, we propose DocSAM, a transformer-based unified framework for various document image segmentation tasks. DocSAM integrates layout analysis, multi-grained text segmentation, and table structure decomposition into a single model, reducing the need for specialized models and enhancing efficiency. Trained on heterogeneous datasets, DocSAM demonstrates robust and generalizable performance, effectively handling diverse document types and structures. This approach also reduces computational and storage requirements, making DocSAM suitable for practical deployment in resource-constrained environments. Extensive experiments show that DocSAM outperforms existing methods in terms of accuracy, efficiency, and adaptability. Overall, we believe that DocSAM represents a significant step forward for document image segmentation, and we look forward to its continued development and application in practical scenarios. In the future, we plan to extend DocSAM to a multi-modal version and explore better training strategies to further accelerate its convergence and improve its performance." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.825, + 0.673, + 0.843 + ], + "angle": 0, + "content": "Acknowledgement" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.85, + 0.905, + 0.88 + ], + "angle": 0, + "content": "This work is supported by the National Natural Science Foundation of China(NSFC) Grant U23B2029." + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "15028" + } + ], + [ + { + "type": "title", + "bbox": [ + 0.093, + 0.09, + 0.188, + 0.106 + ], + "angle": 0, + "content": "References" + }, + { + "type": "ref_text", + "bbox": [ + 0.108, + 0.115, + 0.483, + 0.157 + ], + "angle": 0, + "content": "[1] Abdelrahman Abdallah, Alexander Berendeyev, Islam Nuradin, and Daniyar Nurseitov. Tncr: Table net detection and classification dataset. Neurocomputing, 473:79-97, 2022. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.109, + 0.159, + 0.483, + 0.227 + ], + "angle": 0, + "content": "[2] Avinash Anand, Raj Jaiswal, Mohit Gupta, Siddhesh S Bangar, Pijush Bhuyan, Naman Lal, Rajeev Singh, Ritika Jha, Rajiv Ratn Shah, and Shin'Ichi Satoh. Ranlaynet: A dataset for document layout detection used for domain adaptation and generalization. In ACM MM Asia, pages 1-6, 2023. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.109, + 0.228, + 0.482, + 0.269 + ], + "angle": 0, + "content": "[3] buptlihang. CDLA: A Benchmark Dataset for Cross-Domain Layout Analysis. https://github.com/buptliang/CDLA, 2023. Accessed: 2024-10-31. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.109, + 0.271, + 0.482, + 0.325 + ], + "angle": 0, + "content": "[4] Nicolas Carion, Francisco Massa, Gabriel Synnaeve, Nicolas Usunier, Alexander Kirillov, and Sylvain Gelly. End-to-end object detection with transformers. In ECCV, pages 213-229. Springer, 2020. 3, 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.109, + 0.327, + 0.482, + 0.395 + ], + "angle": 0, + "content": "[5] Liang-Chieh Chen, George Papandreou, Iasonas Kokkinos, Vladimir Koltun, and Alan Garg. Deep convolutional nets, atrous convolution, and fully connected crfs. In IEEE Trans. Pattern Anal. Mach. Intell., pages 834-848, 2018. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.109, + 0.397, + 0.482, + 0.437 + ], + "angle": 0, + "content": "[6] Bowen Cheng, Alex Schwing, and Alexander Kirillov. Per-pixel classification is not all you need for semantic segmentation. NeurIPS, 34:17864-17875, 2021. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.109, + 0.439, + 0.482, + 0.494 + ], + "angle": 0, + "content": "[7] Bowen Cheng, Ishan Misra, Alexander G Schwing, Alexander Kirillov, and Rohit Girdhar. Masked-attention mask transformer for universal image segmentation. In CVPR, pages 1290–1299, 2022. 3, 4, 5, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.109, + 0.496, + 0.482, + 0.55 + ], + "angle": 0, + "content": "[8] Hiuyi Cheng, Cheng Jian, Sihang Wu, and Lianwen Jin. Scut-cab: a new benchmark dataset of ancient chinese books with complex layouts for document layout analysis. In ICFHR, pages 436-451. Springer, 2022. 2, 3, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.109, + 0.552, + 0.482, + 0.633 + ], + "angle": 0, + "content": "[9] Hiuyi Cheng, Peirong Zhang, Sihang Wu, Jiaxin Zhang, Qiyuan Zhu, Zecheng Xie, Jing Li, Kai Ding, and Lianwen Jin. M6doc: A large-scale multi-format, multi-type, multilayout, multi-language, multi-annotation category dataset for modern document layout analysis. In CVPR, pages 15138–15147, 2023. 2, 3, 5, 7, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.635, + 0.482, + 0.676 + ], + "angle": 0, + "content": "[10] Zewen Chi, Heyan Huang, Heng-Da Xu, Houjin Yu, Wanxuan Yin, and Xian-Ling Mao. Complicated table structure recognition. arXiv preprint arXiv:1908.04729, 2019. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.677, + 0.482, + 0.718 + ], + "angle": 0, + "content": "[11] Chee Kheng Ch'ng and Chee Seng Chan. Total-text: A comprehensive dataset for scene text detection and recognition. In ICDAR, pages 935-942. IEEE, 2017. 2, 3, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.72, + 0.482, + 0.789 + ], + "angle": 0, + "content": "[12] Chee Kheng Chng, Yuliang Liu, Yipeng Sun, Chun Chet Ng, Canjie Luo, Zihan Ni, ChuanMing Fang, Shuai tao Zhang, Junyu Han, Errui Ding, et al. Icdar2019 robust reading challenge on arbitrary-shaped text-rrc-art. In ICDAR, pages 1571-1576. IEEE, 2019. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.79, + 0.482, + 0.831 + ], + "angle": 0, + "content": "[13] Cheng Da, Chuwei Luo, Qi Zheng, and Cong Yao. Vision grid transformer for document layout analysis. In ICCV, pages 19462-19472, 2023. 3, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.832, + 0.482, + 0.9 + ], + "angle": 0, + "content": "[14] Hao Feng, Qi Liu, Hao Liu, Wengang Zhou, Houqiang Li, and Can Huang. Docpedia: Unleashing the power of large multimodal model in the frequency domain for versatile document understanding. arXiv preprint arXiv:2311.11810, 2023. 3" + }, + { + "type": "list", + "bbox": [ + 0.102, + 0.115, + 0.483, + 0.9 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.523, + 0.093, + 0.905, + 0.162 + ], + "angle": 0, + "content": "[15] Hao Feng, Zijian Wang, Jingqun Tang, Jinghui Lu, Wengang Zhou, Houqiang Li, and Can Huang. Unidoc: A universal large multimodal model for simultaneous text detection, recognition, spotting and understanding. arXiv preprint arXiv:2308.11592, 2023. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.523, + 0.164, + 0.905, + 0.205 + ], + "angle": 0, + "content": "[16] Liangcai Gao, Xiaohan Yi, Zhuoren Jiang, Leipeng Hao, and Zhi Tang. Icdar2017 competition on page object detection. In ICDAR, pages 1417-1422. IEEE, 2017. 2, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.523, + 0.207, + 0.905, + 0.262 + ], + "angle": 0, + "content": "[17] Liangcai Gao, Yilun Huang, Herve Dejean, Jean-Luc Meunier, Qinqin Yan, Yu Fang, Florian Kleber, and Eva Lang. Icdar 2019 competition on table detection and recognition (ctdar). In ICDAR, pages 1510-1515. IEEE, 2019. 2, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.523, + 0.264, + 0.905, + 0.304 + ], + "angle": 0, + "content": "[18] Max Göbel, Tamir Hassan, Ermelinda Oro, and Giorgio Orsi. Icdar 2013 table competition. In ICDAR, pages 1449-1453. IEEE, 2013. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.523, + 0.307, + 0.905, + 0.361 + ], + "angle": 0, + "content": "[19] Mrinal Haloi, Shashank Shekhar, Nikhil Fande, Siddhant Swaroop Dash, et al. Table detection in the wild: A novel diverse table detection dataset and method. arXiv preprint arXiv:2209.09207, 2022.3" + }, + { + "type": "ref_text", + "bbox": [ + 0.523, + 0.363, + 0.905, + 0.403 + ], + "angle": 0, + "content": "[20] Kaiming He, Georgia Gkioxari, Piotr Dollar, and Ross Girshick. Mask r-cnn. In ICCV, pages 2961-2969, 2017. 3, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.523, + 0.406, + 0.905, + 0.474 + ], + "angle": 0, + "content": "[21] Mengchao He, Yuliang Liu, Zhibo Yang, Sheng Zhang, Canjie Luo, Feiyu Gao, Qi Zheng, Yongpan Wang, Xin Zhang, and Lianwen Jin. Icpr2018 contest on robust reading for multi-type web images. In ICPR, pages 7-12. IEEE, 2018. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.523, + 0.477, + 0.905, + 0.518 + ], + "angle": 0, + "content": "[22] Wenhao He, Xu-Yao Zhang, Fei Yin, and Cheng-Lin Liu. Multi-oriented and multi-lingual scene text detection with direct regression. ICIP, 27(11):5406-5419, 2018. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.523, + 0.521, + 0.905, + 0.588 + ], + "angle": 0, + "content": "[23] Anwen Hu, Haiyang Xu, Jiabo Ye, Ming Yan, Liang Zhang, Bo Zhang, Chen Li, Ji Zhang, Qin Jin, Fei Huang, et al. mplug-docowl 1.5: Unified structure learning forOCR-free document understanding. arXiv preprint arXiv:2403.12895, 2024.3" + }, + { + "type": "ref_text", + "bbox": [ + 0.523, + 0.591, + 0.905, + 0.646 + ], + "angle": 0, + "content": "[24] Jie Hu, Liujuan Cao, Yao Lu, ShengChuan Zhang, Yan Wang, Ke Li, Feiyue Huang, Ling Shao, and Rongrong Ji. Istr: End-to-end instance segmentation with transformers. arXiv preprint arXiv:2105.00637, 2021. 3, 5, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.523, + 0.648, + 0.905, + 0.689 + ], + "angle": 0, + "content": "[25] Jie Hu, Yao Lu, Shengchuan Zhang, and Liujuan Cao. Istr: Mask-embedding-based instance segmentation transformer. IEEE Trans. Image Process., 2024. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.523, + 0.691, + 0.905, + 0.758 + ], + "angle": 0, + "content": "[26] Mingxin Huang, Jiaxin Zhang, Dezhi Peng, Hao Lu, Can Huang, Yuliang Liu, Xiang Bai, and Lianwen Jin. Estextspotter: Towards better scene text spotting with explicit synergy in transformer. In ICCV, pages 19495-19505, 2023. 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.523, + 0.761, + 0.905, + 0.815 + ], + "angle": 0, + "content": "[27] Yupan Huang, Tengchao Lv, Lei Cui, Yutong Lu, and Furu Wei. Layoutlmv3: Pre-training for document ai with unified text and image masking. In ACM MM, pages 4083-4091, 2022. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.523, + 0.818, + 0.905, + 0.899 + ], + "angle": 0, + "content": "[28] Dimosthenis Karatzas, Lluis Gomez-Bigorda, Anguelos Nicolaou, Suman Ghosh, Andrew Bagdanov, Masakazu Iwamura, Jiri Matas, Lukas Neumann, Vijay Ramaseshan Chandrasekhar, Shijian Lu, et al. Icdar 2015 competition on robust reading. In ICDAR, pages 1156-1160. IEEE, 2015. 3, 7" + }, + { + "type": "list", + "bbox": [ + 0.523, + 0.093, + 0.905, + 0.899 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "15029" + } + ], + [ + { + "type": "ref_text", + "bbox": [ + 0.101, + 0.093, + 0.482, + 0.146 + ], + "angle": 0, + "content": "[29] Taeho Kil, Seonghyeon Kim, Sukmin Seo, Yoonsik Kim, and Daehee Kim. Towards unified scene text spotting based on sequence generation. In CVPR, pages 15223-15232, 2023. 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.101, + 0.149, + 0.482, + 0.205 + ], + "angle": 0, + "content": "[30] Alexander Kirillov, Eric Mintun, Nikhila Ravi, Hanzi Mao, Chloe Rolland, Laura Gustafson, Tete Xiao, Spencer Whitehead, Alexander C Berg, Wan-Yen Lo, et al. Segment anything. In ICCV, pages 4015-4026, 2023. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.207, + 0.482, + 0.274 + ], + "angle": 0, + "content": "[31] Pazhou Lab. IACC competition on Chinese Historical Document Analysis Challenge. https://iacc.pazhoulab-huangpu.com/contestdetail?id=6497f74cd97a2dae9dcaeff8 & award=1, 000, 000, 2022. Accessed: 2024-10-31. 3, 7, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.277, + 0.482, + 0.331 + ], + "angle": 0, + "content": "[32] John Lafferty, Andrew McCallum, Fernando Pereira, et al. Conditional random fields: Probabilistic models for segmenting and labeling sequence data. In ICML, page 3. Williamstown, MA, 2001. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.333, + 0.482, + 0.387 + ], + "angle": 0, + "content": "[33] Feng Li, Hao Zhang, Shilong Liu, Jian Guo, Lionel M Ni, and Lei Zhang. Dn-detr: Accelerate detr training by introducing query denoising. In CVPR, pages 13619-13627, 2022. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.39, + 0.482, + 0.444 + ], + "angle": 0, + "content": "[34] Junlong Li, Yiheng Xu, Tengchao Lv, Lei Cui, Cha Zhang, and Furu Wei. Dit: Self-supervised pre-training for document image transformer. In ACM MM, pages 3530-3539, 2022. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.447, + 0.482, + 0.502 + ], + "angle": 0, + "content": "[35] Minghao Li, Lei Cui, Shaohan Huang, Furu Wei, Ming Zhou, and Zhoujun Li. Tablebank: Table benchmark for image-based table detection and recognition. In LREC, pages 1918-1925, 2020. 2, 3, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.504, + 0.482, + 0.558 + ], + "angle": 0, + "content": "[36] Minghao Li, Yiheng Xu, Lei Cui, Shaohan Huang, Furu Wei, Zhoujun Li, and Ming Zhou. Docbank: A benchmark dataset for document layout analysis. arXiv preprint arXiv:2006.01038, 2020. 2, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.561, + 0.482, + 0.602 + ], + "angle": 0, + "content": "[37] Xiao-Hui Li, Fei Yin, and Cheng-Lin Liu. Page segmentation using convolutional neural network and graphical model. In DAS, pages 231-245. Springer, 2020. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.604, + 0.482, + 0.657 + ], + "angle": 0, + "content": "[38] Xiao-Hui Li, Fei Yin, He-Sen Dai, and Cheng-Lin Liu. Table structure recognition and form parsing by end-to-end object detection and relation parsing. Pattern Recognition, 132:108946, 2022. 2, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.66, + 0.482, + 0.716 + ], + "angle": 0, + "content": "[39] Tsung-Yi Lin, Michael Maire, Serge Belongie, James Hays, Pietro Perona, Deva Ramanan, Piotr Dólar, and C Lawrence Zitnick. Microsoft coco: Common objects in context. In ECCV, pages 740-755. Springer, 2014. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.718, + 0.482, + 0.771 + ], + "angle": 0, + "content": "[40] Tsung-Yi Lin, Piotr Dólar, Ross Girshick, Kaiming He, Bharath Hariharan, and Serge Belongie. Feature pyramid networks for object detection. In CVPR, pages 2117-2125, 2017. 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.774, + 0.482, + 0.815 + ], + "angle": 0, + "content": "[41] Cheng-Lin Liu, Fei Yin, Da-Han Wang, and Qiu-Feng Wang. Casia online and offline chinese handwriting databases. In ICDAR, pages 37-41. IEEE, 2011. 3, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.817, + 0.482, + 0.871 + ], + "angle": 0, + "content": "[42] Cheng-Lin Liu, Lianwen Jin, Xiang Bai, Xiaohui Li, and Fei Yin. Frontiers of intelligent document analysis and recognition: review and prospects. Journal of Image and Graphics, 28(08):2223-2252, 2023. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.873, + 0.482, + 0.901 + ], + "angle": 0, + "content": "[43] Yuliang Liu, Lianwen Jin, Shuai tao Zhang, Canjie Luo, and Sheng Zhang. Curved scene text detection via transverse" + }, + { + "type": "list", + "bbox": [ + 0.101, + 0.093, + 0.482, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.554, + 0.093, + 0.905, + 0.12 + ], + "angle": 0, + "content": "and longitudinal sequence connection. Pattern Recognition, 90:337-345, 2019. 2, 3, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.524, + 0.122, + 0.905, + 0.178 + ], + "angle": 0, + "content": "[44] Yuliang Liu, Biao Yang, Qiang Liu, Zhang Li, Zhiyin Ma, Shuo Zhang, and Xiang Bai. Textmonkey: AnOCR-free large multimodal model for understanding document. arXiv preprint arXiv:2403.04473, 2024. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.524, + 0.179, + 0.905, + 0.247 + ], + "angle": 0, + "content": "[45] Ze Liu, Yutong Lin, Yan Cao, Yue Hu, Yu Wei, Zhiqiang Zhang, Jiahui Lin, Han Wang, Cheng Lu, Changhu Wang, et al. Swin transformer: Hierarchical vision transformer using shifted windows. In ICCV, pages 12226-12235, 2021. 3, 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.524, + 0.249, + 0.905, + 0.291 + ], + "angle": 0, + "content": "[46] Jonathan Long, Evan Shelhamer, and Trevor Darrell. Fully convolutional networks for semantic segmentation. In CVPR, pages 3431-3440, 2015. 3, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.524, + 0.293, + 0.905, + 0.333 + ], + "angle": 0, + "content": "[47] Rujiao Long, Wen Wang, Nan Xue, Feiyu Gao, Zhibo Yang, Yongpan Wang, and Gui-Song Xia. Parsing table structures in the wild. In ICCV, pages 944-952, 2021. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.524, + 0.335, + 0.905, + 0.39 + ], + "angle": 0, + "content": "[48] Shangbang Long, Jiaqiang Ruan, Wenjie Zhang, Xin He, Wenhao Wu, and Cong Yao. Textsnake: A flexible representation for detecting text of arbitrary shapes. In ECCV, pages 20-36, 2018. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.524, + 0.392, + 0.905, + 0.447 + ], + "angle": 0, + "content": "[49] Shangbang Long, Siyang Qin, Dmitry Panteleev, Alessandro Bissacco, Yasuhisa Fujii, and Michalis Raptis. Towards end-to-end unified scene text detection and layout analysis. In CVPR, pages 1049-1059, 2022. 3, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.524, + 0.449, + 0.905, + 0.476 + ], + "angle": 0, + "content": "[50] I Loshchilov. Decoupled weight decay regularization. arXiv preprint arXiv:1711.05101, 2017. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.524, + 0.478, + 0.905, + 0.518 + ], + "angle": 0, + "content": "[51] Ilya Loshchilov and Frank Hutter. Sgdr: Stochastic gradient descent with warm restarts. arXiv preprint arXiv:1608.03983, 2016. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.524, + 0.521, + 0.905, + 0.576 + ], + "angle": 0, + "content": "[52] Siwen Luo, Yihao Ding, Siqu Long, Josiah Poon, and Soyeon Caren Han. Doc-gcn: Heterogeneous graph convolutional networks for document layout analysis. arXiv preprint arXiv:2208.10970, 2022. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.524, + 0.578, + 0.905, + 0.632 + ], + "angle": 0, + "content": "[53] Weihong Ma, Hesuo Zhang, Lianwen Jin, Sihang Wu, Jiapeng Wang, and Yongpan Wang. Joint layout analysis, character detection and recognition for historical document digitization. In ICFHR, pages 31-36. IEEE, 2020. 2, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.524, + 0.634, + 0.905, + 0.688 + ], + "angle": 0, + "content": "[54] Fausto Miletari, Nassir Navab, and Seyed-Ahmad Ahmadi. V-net: Fully convolutional neural networks for volumetric medical image segmentation. In 3DV, pages 565-571. IEEE, 2016. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.524, + 0.691, + 0.905, + 0.733 + ], + "angle": 0, + "content": "[55] Ajoy Mondal, Peter Lipps, and CV Jawahar. Iit-ar-13k: A new dataset for graphical object detection in documents. In DAS, pages 216–230. Springer, 2020. 3, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.524, + 0.735, + 0.905, + 0.815 + ], + "angle": 0, + "content": "[56] Nibal Nayef, Fei Yin, Imen Bizid, Hyunsoo Choi, Yuan Feng, Dimosthenis Karatzas, Zhenbo Luo, Umapada Pal, Christophe Rigaud, Joseph Chazalon, et al. Icdar2017 robust reading challenge on multi-lingual scene text detection and script identification-rrc-plt. In ICDAR, pages 1454-1459. IEEE, 2017. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.524, + 0.818, + 0.905, + 0.901 + ], + "angle": 0, + "content": "[57] Nibal Nayef, Yash Patel, Michal Busta, Pinaki Nath Chowdhury, Dimosthenis Karatzas, Wafa Khlif, Jiri Matas, Umapada Pal, Jean-Christophe Burie, Cheng-lin Liu, et al. Icdar2019 robust reading challenge on multi-lingual scene text detection and recognition—rrc-mlt-2019. In ICDAR, pages 1582–1587. IEEE, 2019. 3" + }, + { + "type": "list", + "bbox": [ + 0.524, + 0.093, + 0.905, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.52, + 0.957 + ], + "angle": 0, + "content": "15030" + } + ], + [ + { + "type": "ref_text", + "bbox": [ + 0.101, + 0.093, + 0.482, + 0.161 + ], + "angle": 0, + "content": "[58] Adam Paszke, Sam Gross, Francisco Massa, Adam Lerer, James Bradbury, Gregory Chanan, Trevor Killeen, Zeming Lin, Natalia Gimelshein, Luca Antiga, et al. Pytorch: An imperative style, high-performance deep learning library. NeurIPS, 32, 2019. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.101, + 0.164, + 0.482, + 0.218 + ], + "angle": 0, + "content": "[59] Dezhi Peng, Zhenhua Yang, Jiaxin Zhang, Chongyu Liu, Yongxin Shi, Kai Ding, Fengjun Guo, and Lianwen Jin. Upocr: Towards unified pixel-levelOCR interface. In ICML, 2023. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.221, + 0.483, + 0.275 + ], + "angle": 0, + "content": "[60] Birgit Pfitzmann, Christoph Auer, Michele Dolfi, Ahmed S. Nassar, and Peter Staar. Doclaynet: A large human-annotated dataset for document layout analysis. In ACM SIGKDD, pages 3743–3751. 2022. 2, 3, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.278, + 0.482, + 0.344 + ], + "angle": 0, + "content": "[61] Xugong Qin, Pengyuan Lyu, Chengquan Zhang, Yu Zhou, Kun Yao, Peng Zhang, Hailun Lin, and Weiping Wang. Towards robust real-time scene text detection: From semantic to instance representation learning. In ACM MM, pages 2025-2034, 2023. 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.347, + 0.482, + 0.387 + ], + "angle": 0, + "content": "[62] Joseph Redmon and Ali Farhadi. You only look once: Unified, real-time object detection. In CVPR, pages 779-788, 2016. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.391, + 0.482, + 0.417 + ], + "angle": 0, + "content": "[63] Joseph Redmon and Ali Farhadi. Yolov3: An incremental improvement, 2018. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.42, + 0.482, + 0.46 + ], + "angle": 0, + "content": "[64] N Reimers. Sentence-bert: Sentence embeddings using siamese bert-networks. arXiv preprint arXiv:1908.10084, 2019. 1, 4, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.463, + 0.482, + 0.516 + ], + "angle": 0, + "content": "[65] Shaoqing Ren, Kaiming He, Ross Girshick, and Jian Sun. Faster r-cnn: Towards real-time object detection with region proposal networks. In NeurIPS, pages 91-99, 2015. 3, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.519, + 0.482, + 0.56 + ], + "angle": 0, + "content": "[66] Olaf Ronneberger, Philipp Fischer, and Thomas Brox. U-net: Convolutional networks for biomedical image segmentation. In MICCAI, pages 234-241. Springer, 2015. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.563, + 0.482, + 0.59 + ], + "angle": 0, + "content": "[67] T-YLPG Ross and GKHP Dólar. Focal loss for dense object detection. In CVPR, pages 2980-2988, 2017. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.592, + 0.482, + 0.658 + ], + "angle": 0, + "content": "[68] Rajkumar Saini, Derek Dobson, Jon Morrey, Marcus Liwicki, and Foteini Simistira Liwicki. Icdar 2019 historical document reading challenge on large structured chinese family records. In ICDAR, pages 1499-1504. IEEE, 2019. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.662, + 0.482, + 0.703 + ], + "angle": 0, + "content": "[69] Zejiang Shen, Kaixuan Zhang, and Melissa Dell. A large dataset of historical japanese documents with complex layouts. In CVPR Workshops, pages 548-549, 2020. 3, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.705, + 0.482, + 0.771 + ], + "angle": 0, + "content": "[70] Baoguang Shi, Cong Yao, Minghui Liao, Mingkun Yang, Pei Xu, Linyan Cui, Serge Belongie, Shijian Lu, and Xiang Bai. Icdar2017 competition on reading chinese text in the wild (rctw-17). In ICDAR, pages 1429–1434. IEEE, 2017. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.776, + 0.482, + 0.871 + ], + "angle": 0, + "content": "[71] Md Istiak Hossain Shihab, Md Rakibul Hasan, Mahfuzur Rahman Emon, Syed Mobassir Hossen, Md Nazmuddoha Ansary, Intesur Ahmed, Fazle Rabbi Rakib, Shahriar Elahi Dhruvo, Souhardya Saha Dip, Akib Hasan Pavel, et al. Badlad: A large multi-domain bengali document layout analysis dataset. In ICDAR, pages 326-341. Springer, 2023. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.874, + 0.482, + 0.901 + ], + "angle": 0, + "content": "[72] Brandon Smock, Rohith Pesala, and Robin Abraham. Pubtables-1m: Towards comprehensive table extraction" + }, + { + "type": "list", + "bbox": [ + 0.101, + 0.093, + 0.483, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.554, + 0.093, + 0.905, + 0.12 + ], + "angle": 0, + "content": "from unstructured documents. In CVPR, pages 4634-4642, 2022. 2, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.525, + 0.123, + 0.905, + 0.176 + ], + "angle": 0, + "content": "[73] Xingyi Sun, Ziyi Zhou, Peizhao Wang, Junwei Dai, Lu Lu, Wanli Ouyang, and Ping Luo. Sparse r-cnn: End-to-end object detection with learnable proposals. In CVPR, pages 13164-13173, 2021. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.524, + 0.179, + 0.905, + 0.248 + ], + "angle": 0, + "content": "[74] Yipeng Sun, Zihan Ni, Chee-Kheng Chng, Yuliang Liu, Canjie Luo, Chun Chet Ng, Junyu Han, Errui Ding, Jingtuo Liu, Dimosthenis Karatzas, et al. Icdar 2019 competition on large-scale street view text with partial labeling-rrc-lsvt. In ICDAR, pages 1557-1562. IEEE, 2019. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.524, + 0.25, + 0.905, + 0.317 + ], + "angle": 0, + "content": "[75] Zineng Tang, Ziyi Yang, Guoxin Wang, Yuwei Fang, Yang Liu, Chenguang Zhu, Michael Zeng, Cha Zhang, and Mohit Bansal. Unifying vision, text, and layout for universal document processing. In CVPR, pages 19254-19264, 2023. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.524, + 0.321, + 0.905, + 0.375 + ], + "angle": 0, + "content": "[76] Andreas Veit, Tomas Matera, Lukas Neumann, Jiri Matas, and Serge Belongie. Coco-text: Dataset and benchmark for text detection and recognition in natural images. arXiv preprint arXiv:1601.07140, 2016. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.524, + 0.378, + 0.905, + 0.418 + ], + "angle": 0, + "content": "[77] Thang Vu, Haeyong Kang, and Chang D Yoo. Scnet: Training inference sample consistency for instance segmentation. In AAAI, pages 2701-2709, 2021. 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.524, + 0.421, + 0.905, + 0.489 + ], + "angle": 0, + "content": "[78] Jianqiang Wan, Sibo Song, Wenwen Yu, Yuliang Liu, Wenqing Cheng, Fei Huang, Xiang Bai, Cong Yao, and Zhibo Yang. Omniparser: A unified framework for text spotting key information extraction and table recognition. In CVPR, pages 15641-15653, 2024. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.524, + 0.491, + 0.905, + 0.559 + ], + "angle": 0, + "content": "[79] Xingyu Wan, Chengquan Zhang, Pengyuan Lyu, Sen Fan, Zihan Ni, Kun Yao, Errui Ding, and Jingdong Wang. Towards unified multi-granularity text detection with interactive attention. arXiv preprint arXiv:2405.19765, 2024. 3, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.524, + 0.562, + 0.905, + 0.63 + ], + "angle": 0, + "content": "[80] Dongsheng Wang, Natraj Raman, Mathieu Sibue, Zhiqiang Ma, Petr Babkin, Simerjot Kaur, Yulong Pei, Armineh Nourbakhsh, and Xiaomo Liu. Docllm: A layout-aware generative language model for multimodal document understanding. arXiv preprint arXiv:2401.00908, 2023. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.524, + 0.633, + 0.905, + 0.673 + ], + "angle": 0, + "content": "[81] Xin Wang, Yudong Chen, and Wenwu Zhu. A survey on curriculum learning. IEEE Trans. Pattern Anal. Mach. Intell., 44(9):4555-4576, 2021. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.524, + 0.676, + 0.905, + 0.73 + ], + "angle": 0, + "content": "[82] Zonghan Wu, Shirui Pan, Fengwen Chen, Guodong Long, Chengqi Zhang, and S Yu Philip. A comprehensive survey on graph neural networks. IEEE Trans. Neural Netw. Learn. Syst., 32(1):4-24, 2020. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.524, + 0.733, + 0.905, + 0.787 + ], + "angle": 0, + "content": "[83] Enze Xie, Zhiwei Wang, Gang Yu, Xingyi Zhang, Yuheng Wu, Hongsheng Li, and Yonghong Tian. Segformer: Simple and efficient design for semantic segmentation with transformers. In ICCV, pages 14125-14134, 2021. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.524, + 0.79, + 0.905, + 0.844 + ], + "angle": 0, + "content": "[84] Zhenda Xie, Zheng Zhang, Yue Cao, Yutong Lin, Jianmin Bao, Zhuliang Yao, Qi Dai, and Han Hu. Simmim: A simple framework for masked image modeling. In CVPR, pages 9653-9663, 2022. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.524, + 0.846, + 0.905, + 0.901 + ], + "angle": 0, + "content": "[85] Yue Xu, Fei Yin, Da-Han Wang, Xu-Yao Zhang, Zhaoxiang Zhang, and Cheng-Lin Liu. Casia-ahcdb: A large-scale chinese ancient handwritten characters database. In ICDAR, pages 793–798. IEEE, 2019. 3" + }, + { + "type": "list", + "bbox": [ + 0.524, + 0.093, + 0.905, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.946, + 0.518, + 0.957 + ], + "angle": 0, + "content": "15031" + } + ], + [ + { + "type": "ref_text", + "bbox": [ + 0.101, + 0.093, + 0.482, + 0.134 + ], + "angle": 0, + "content": "[86] Cong Yao, Xiang Bai, Wenyu Liu, Yi Ma, and Zhuowen Tu. Detecting texts of arbitrary orientations in natural images. In CVPR, pages 1083-1090. IEEE, 2012. 3, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.101, + 0.136, + 0.482, + 0.176 + ], + "angle": 0, + "content": "[87] Cong Yao, Xiang Bai, and Wenyu Liu. A unified framework for multioriented text detection and recognition. ICIP, 23 (11):4737-4749, 2014. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.179, + 0.483, + 0.247 + ], + "angle": 0, + "content": "[88] Jiabo Ye, Anwen Hu, Haiyang Xu, Qinghao Ye, Ming Yan, Yuhao Dan, Chenlin Zhao, Guohai Xu, Chenliang Li, Junfeng Tian, et al. mplug-docowl: Modularized multimodal large language model for document understanding. arXiv preprint arXiv:2307.02499, 2023. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.25, + 0.483, + 0.304 + ], + "angle": 0, + "content": "[89] Maoyuan Ye, Jing Zhang, Shanshan Zhao, Juhua Liu, Bo Du, and Dacheng Tao. Dptext-detr: Towards better scene text detection with dynamic points in transformer. In AAAI, pages 3241-3249, 2023. 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.307, + 0.482, + 0.359 + ], + "angle": 0, + "content": "[90] Xu-Cheng Yin, Xuwang Yin, Kaizhu Huang, and Hong-Wei Hao. Robust text detection in natural scene images. IEEE Trans. Pattern Anal. Mach. Intell., 36(5):970–983, 2013. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.363, + 0.482, + 0.418 + ], + "angle": 0, + "content": "[91] Wenwen Yu, Mingyu Liu, Mingrui Chen, Ning Lu, Yinlong Wen, Yuliang Liu, Dimosthenis Karatzas, and Xiang Bai. Icdar 2023 competition on reading the seal title. In ICDAR, pages 522-535. Springer, 2023. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.42, + 0.482, + 0.46 + ], + "angle": 0, + "content": "[92] Tai-Ling Yuan, Zhe Zhu, Kun Xu, Cheng-Jun Li, Tai-Jiang Mu, and Shi-Min Hu. A large chinese text dataset in the wild. J. Comput. Sci. Tech., 34:509-521, 2019. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.463, + 0.482, + 0.53 + ], + "angle": 0, + "content": "[93] Yukun Zhai, Xiaogiang Zhang, Xiameng Qin, Sanyuan Zhao, Xingping Dong, and Jianbing Shen. Textformer: A query-based end-to-end text spotter with mixed supervision. Machine Intelligence Research, 21(4):704-717, 2024. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.533, + 0.482, + 0.588 + ], + "angle": 0, + "content": "[94] Chongsheng Zhang, Guowen Peng, Yuefeng Tao, Feifei Fu, Wei Jiang, George Almpanidis, and Ke Chen. Shopsign: A diverse scene text dataset of chinese shop signs in street views. arXiv preprint arXiv:1903.10412, 2019. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.591, + 0.482, + 0.644 + ], + "angle": 0, + "content": "[95] Hesuo Zhang, Lingyu Liang, and Lianwen Jin. Scuthcdoc: A new benchmark dataset of handwritten chinese text in unconstrained camera-captured documents. Pattern Recognition, 108:107559, 2020. 2, 3, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.647, + 0.482, + 0.702 + ], + "angle": 0, + "content": "[96] Hao Zhang, Feng Li, Shilong Liu, Lei Zhang, Hang Su, Jun Zhu, Lionel M Ni, and Heung-Yeung Shum. Dino: Detr with improved denoising anchor boxes for end-to-end object detection. arXiv preprint arXiv:2203.03605, 2022. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.704, + 0.482, + 0.758 + ], + "angle": 0, + "content": "[97] Hao Zhang, Feng Li, Huaizhe Xu, Shijia Huang, Shilong Liu, Lionel M Ni, and Lei Zhang. Mp-former: Mask-piloted transformer for image segmentation. In CVPR, pages 18074-18083, 2023. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.761, + 0.482, + 0.814 + ], + "angle": 0, + "content": "[98] Jiaxin Zhang, Dezhi Peng, Chongyu Liu, Peirong Zhang, and Lianwen Jin. Docres: A generalist model toward unifying document image restoration tasks. In CVPR, pages 15654-15664, 2024. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.817, + 0.482, + 0.871 + ], + "angle": 0, + "content": "[99] Peng Zhang, Can Li, Liang Qiao, Zhanzhan Cheng, Shiliang Pu, Yi Niu, and Fei Wu. Vsr: a unified framework for document layout analysis combining vision, semantics and relations. In ICDAR, pages 115-130. Springer, 2021. 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.096, + 0.874, + 0.482, + 0.901 + ], + "angle": 0, + "content": "[100] Rui Zhang, Yongsheng Zhou, Qianyi Jiang, Qi Song, Nan Li, Kai Zhou, Lei Wang, Dong Wang, Minghui Liao," + }, + { + "type": "list", + "bbox": [ + 0.096, + 0.093, + 0.483, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.554, + 0.093, + 0.905, + 0.134 + ], + "angle": 0, + "content": "Mingkun Yang, et al. Icdar 2019 robust reading challenge on reading chinese text on signboard. In ICDAR, pages 1577-1581. IEEE, 2019. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.136, + 0.905, + 0.19 + ], + "angle": 0, + "content": "[101] Shi-Xue Zhang, Xiaobin Zhu, Jie-Bo Hou, Chang Liu, Chun Yang, Hongfa Wang, and Xu-Cheng Yin. Deep relational reasoning graph network for arbitrary shape text detection. In CVPR, pages 9699–9708, 2020. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.193, + 0.905, + 0.247 + ], + "angle": 0, + "content": "[102] Wayne Xin Zhao, Kun Zhou, Junyi Li, Tianyi Tang, Xiaolei Wang, Yupeng Hou, Yingqian Min, Beichen Zhang, Junjie Zhang, Zican Dong, et al. A survey of large language models. arXiv preprint arXiv:2303.18223, 2023. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.25, + 0.905, + 0.316 + ], + "angle": 0, + "content": "[103] Enze Zheng, Zhiwei Wang, Gang Yu, Xingyi Zhang, Yuheng Wu, Hongsheng Li, and Yonghong Tian. Rethinking semantic segmentation from a sequence-to-sequence perspective with transformers. In CVPR, pages 12164-12173, 2021. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.32, + 0.905, + 0.387 + ], + "angle": 0, + "content": "[104] Xinyi Zheng, Douglas Burdick, Lucian Popa, Xu Zhong, and Nancy Xin Ru Wang. Global table extractor (gte): A framework for joint table identification and cell structure recognition using visual context. In WACV, pages 697-706, 2021. 3, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.391, + 0.905, + 0.444 + ], + "angle": 0, + "content": "[105] Zhaohui Zheng, Ping Wang, Wei Liu, Jinze Li, Rongguang Ye, and Dongwei Ren. Distance-iou loss: Faster and better learning for bounding box regression. In AAAI, pages 12993-13000, 2020. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.447, + 0.905, + 0.488 + ], + "angle": 0, + "content": "[106] Xu Zhong, Jianbin Tang, and Antonio Jimeno Yepes. Publaynet: largest dataset ever for document layout analysis. In ICDAR, pages 1015-1022. IEEE, 2019. 2, 3, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.49, + 0.905, + 0.543 + ], + "angle": 0, + "content": "[107] Xu Zhong, Elaheh ShafieiBavani, and Antonio Jimeno Yepes. Image-based table recognition: data, model, and evaluation. In ECCV, pages 564-580. Springer, 2020. 2,3,7" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.547, + 0.905, + 0.613 + ], + "angle": 0, + "content": "[108] Xingyi Zhu, Junwei Dai, Lu Lu, Yuwen Zhang, Peizhao Wang, Yisen Sun, Wanli Ouyang, and Ping Luo. Deformable detr: Deformable transformers for end-to-end object detection. In ICCV, pages 12352-12361, 2021. 1, 3, 4, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.617, + 0.905, + 0.67 + ], + "angle": 0, + "content": "[109] Ziyi Zhu, Liangcai Gao, Yibo Li, Yilun Huang, Lin Du, Ning Lu, and Xianfeng Wang. Ntable: a dataset for camera-based table detection. In ICDAR, pages 117-129. Springer, 2021. 3" + }, + { + "type": "list", + "bbox": [ + 0.096, + 0.093, + 0.905, + 0.67 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.946, + 0.519, + 0.957 + ], + "angle": 0, + "content": "15032" + } + ] +] \ No newline at end of file diff --git a/2025/DocSAM_ Unified Document Image Segmentation via Query Decomposition and Heterogeneous Mixed Learning/cd34a0c4-c9ec-4e5a-a2a8-007f0052f160_origin.pdf b/2025/DocSAM_ Unified Document Image Segmentation via Query Decomposition and Heterogeneous Mixed Learning/cd34a0c4-c9ec-4e5a-a2a8-007f0052f160_origin.pdf new file mode 100644 index 0000000000000000000000000000000000000000..efd44711c079434e63608e55d834c91120eb7ec0 --- /dev/null +++ b/2025/DocSAM_ Unified Document Image Segmentation via Query Decomposition and Heterogeneous Mixed Learning/cd34a0c4-c9ec-4e5a-a2a8-007f0052f160_origin.pdf @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:15e376fa266c773fd5e32bd92c0024e98ffbea4ac8e1907aa4ed1bf7d33bf55e +size 2008204 diff --git a/2025/DocSAM_ Unified Document Image Segmentation via Query Decomposition and Heterogeneous Mixed Learning/full.md b/2025/DocSAM_ Unified Document Image Segmentation via Query Decomposition and Heterogeneous Mixed Learning/full.md new file mode 100644 index 0000000000000000000000000000000000000000..456dbe81843099dd961b5575ad3aafecab8d88b9 --- /dev/null +++ b/2025/DocSAM_ Unified Document Image Segmentation via Query Decomposition and Heterogeneous Mixed Learning/full.md @@ -0,0 +1,363 @@ +# DocSAM: Unified Document Image Segmentation via Query Decomposition and Heterogeneous Mixed Learning + +Xiao-Hui Li $^{1}$ , Fei Yin $^{1}$ , Cheng-Lin Liu $^{1,2}$ $^{1}$ MAIS, Institute of Automation of Chinese Academy of Sciences, Beijing, 100190, China + $^{2}$ School of Artificial Intelligence, University of Chinese Academy of Sciences, Beijing, 100049, China +{xiaohui.li, fyin, liucl}@nlpr.ia.ac.cn + +# Abstract + +Document image segmentation is crucial for document analysis and recognition but remains challenging due to the diversity of document formats and segmentation tasks. Existing methods often address these tasks separately, resulting in limited generalization and resource wastage. This paper introduces DocSAM, a transformer-based unified framework designed for various document image segmentation tasks, such as document layout analysis, multi-granularity text segmentation, and table structure recognition, by modelling these tasks as a combination of instance and semantic segmentation. Specifically, DocSAM employs Sentence-BERT to map category names from each dataset into semantic queries that match the dimensionality of instance queries. These two sets of queries interact through an attention mechanism and are cross-attended with image features to predict instance and semantic segmentation masks. Instance categories are predicted by computing the dot product between instance and semantic queries, followed by softmax normalization of scores. Consequently, DocSAM can be jointly trained on heterogeneous datasets, enhancing robustness and generalization while reducing computational and storage resources. Comprehensive evaluations show that DocSAM surpasses existing methods in accuracy, efficiency, and adaptability, highlighting its potential for advancing document image understanding and segmentation across various applications. Codes are available at https://github.com/xhli-git/DocSAM. + +# 1. Introduction + +Document image segmentation (DIS) is a fundamental task in the field of document analysis and recognition (DAR) [42], serving as a cornerstone for downstream applications such as text recognition, information extraction (IE), and document visual question answering (DocVQA). Despite + +its importance, DIS faces significant challenges due to the wide diversity of document types, page layouts, content annotations, and structural complexities, see Fig. 1. Existing approaches often address specific aspects of DIS separately, such as layout analysis, text detection, and table structure recognition, leading to specialized and fragmented solutions tailored to particular applications. This fragmentation not only impedes the performance of individual tasks but also results in redundant computational and storage overheads, making them inefficient for large-scale deployment. + +To address the aforementioned challenges, this paper introduces DocSAM (Document Segment Anything Model), a transformer-based unified framework designed to simultaneously handle various document image segmentation tasks, thereby eliminating the need for separate models and enhancing overall efficiency. As illustrated in Fig. 2, DocSAM comprises four primary modules: the Vision Backbone, the Deformable Encoder, Sentence-BERT [64], and the Hybrid Query Decoder (HQD). Given a document image and desired instance or semantic class names in natural text format, DocSAM first extracts multi-scale image features using the Vision Backbone. These features are then refined by the Deformable Encoder, which includes several deformable attention layers [108]. Class names are fed into Sentence-BERT and mapped to semantic queries. Subsequently, both semantic queries and learnable instance queries pass together through the HQD, where they interact to jointly perform semantic and instance segmentation. + +Inside each HQD layer (see Fig. 2), semantic and instance queries are concatenated and passed through a multi-head self-attention layer followed by a feed-forward layer for information exchange. These queries are then separately cross-attended with multi-scale image features in a coarse-to-fine manner using two multi-scale decoders, each with $L = 4$ layers. They further interact via another multi-head self-attention and feed-forward layer. The resulting semantic and instance queries, along with fused multi-scale image + +![](images/3a6a1f1783c34c1b3b6f4d53e438720afacc5ce3c93f7798170108fa8066c848.jpg) +Figure 1. Examples of various segmentation tasks on heterogeneous document datasets. + +![](images/8ac8b9cb48c302516b0d45a95ad8117108a69b452aef87003aaf121d116dcf48.jpg) + +![](images/b734259ed50c26b5a4d03e1d2dcdd8b88de0700ea8d3217e81c277437206cf3c.jpg) + +![](images/dbde6a541b34d3b73dd2631149976a9949eebc85db692c1e342ced8a6e53ba3c.jpg) + +features, are forwarded to the Mask Predictor, Class Predictor, and BBox Predictor for semantic mask segmentation, instance mask segmentation, category classification, and bounding box regression, respectively. We stack $K$ HQD layers for more refined predictions. + +This design ensures that DocSAM can effectively manage the heterogeneity of document types, annotation formats, and segmentation tasks while maintaining high efficiency and accuracy. Extensive experiments and evaluations on various datasets demonstrate that DocSAM surpasses existing methods in accuracy, efficiency, and adaptability. Our results highlight DocSAM's potential as a powerful tool for advancing document image segmentation and understanding, with applications spanning from modern and historical document layout analysis to table structure decomposition, handwritten and scene text detection, and beyond. Our contributions are summarized as follows: + +- We introduce DocSAM, a unified solution for diverse document image segmentation tasks such as layout analysis, multi-grained text segmentation, and table structure decomposition, reducing the need for specialized models and enhancing overall efficiency; +- By training on various tasks and datasets, DocSAM improves robustness and generalization, making it highly effective in handling varied document types and structures; +- Compared to specialized models, DocSAM significantly reduces computational and storage requirements, making it more practical for large-scale deployment; +- Extensive experiments on various datasets show that Doc-SAM outperforms current methods in terms of accuracy, efficiency, and adaptability. + +# 2. Related Works + +# 2.1. DIS Tasks and Datasets + +Depending on specific application scenarios, DIS involves various sub-tasks including Document Layout Analysis (DLA), Multi-Granularity Text Detection (MGTD), and Table Structure Recognition (TSR). DLA aims at identifying and categorizing page regions including text blocks, figures and tables [9, 16, 36, 60, 106]. This foundational step provides a structured overview of the document's layout, enabling more precise processing in subsequent tasks. MGTD focuses on detecting and segmenting text at various granularities, from paragraphs down to individual lines and words [11, 43, 53, 93, 95]. MGTD is a prerequisite for accurate Optical Character Recognition (OCR) tasks. TSR specifically aims to extract the structural of tables, including rows, columns and cells [17, 35, 38, 72, 107]. By decomposing tables into substructures, TSR facilitates the extraction and analysis of tabular information from documents. + +Along with these tasks, plenty of datasets have been accumulated after decades of research, see Tab. 1. These datasets exhibit great diversity and heterogeneity in data sources, document types, annotation formats, writing languages, category sets and many other aspects. For example, PubLayNet [106] contains born-digital English PDF documents with region-level annotations; SCUT-CAB [8] and MTHv2 [53] contains scanned historical Chinese documents with region, line and char-level annotations; SCUT-HCCDoc [95] contains handwritten documents with line-level annotations; CTW1500 [43] and Total-Text [11] contain natural scene images with texts of arbitrary shapes. + +
TaskDataset
Document Layout AnalysisBaDLAD [71], CDLA [3], D4LA [13], DocBank [36], DocLayNet [60], ICDAR2017-POD [16], IIIT-AR-13K [55], M6Doc [9], PubLayNet [106], RanLayNet [2]
Ancient and Handwritten Document SegmentationCASIA-AHCDB [85], CHDAC-2022 [31], ICDAR2019-HDRC [68], SCUT-CAB [8], MTHv2 [53], HJDataset [69], CASIA-HWDB [41], SCUT-HCCDoc [95]
Table Structure RecognitionFinTabNet [104], ICDAR2013 [18], ICDAR2017-POD [16, 38], ICDAR2019-cTDr [17, 38], NTable [109], PubTables-1M [72], PubTabNet [107], STDW [19], TableBank [35], TNCR [1], WTW [47]
Scene Text DetectionCASIA-10k [22], COCO-Text [76], CTW1500 [43], CTW-Public [92], HUST-TR400 [87], ICDAR2015 [28], ICDAR2017-RCTW [70], ICDAR2017-MLT [56], ICDAR2019-ArT [12], ICDAR2019-LSVT [74], ICDAR2019-MLT [57], ICDAR2019-ReCTS [100], ICDAR2023-HierText [49], ICDAR2023-ReST [91], ICPR2018-MTWI [21], MSRA-TD500 [86], ShopSign [94], Total-Text [11], USTB-SV1K [90]
+ +Table 1. Datasets involved in DocSAM. + +# 2.2. Deep Learning for DIS + +Existing deep learning based DIS methods basically focus on specific sub-tasks and datasets. Generally speaking, they usually transform various DIS tasks into general object detection or image segmentation problems and make some modifications to general object detection [20, 62, 63, 65] and image segmentation methods [5, 46, 66] to make them more suitable for the tasks and datasets at hand. Some other works treat documents as hierarchical graph structures and adopt graph models like GNN [82] and CRF [32] for the task of layout analysis [37, 52], table structure recognition [10, 38], and text detection [48, 101]. Though more flexible, these methods usually suffer from complicated pre/post-processing steps and are more susceptible to intermediate errors. There are also some multi-modal based methods that combine visual and textual features like LayoutLMv3 [27], DiT [34], and VGT [13]. These methods improve the performance and generalization by pre-training on large-scale unsupervised documents to align text and visual features, but are often slower due to the complexity of architectures. + +With the prosperity of large language models (LLMs) [102], many large document models are proposed such as UDOP [75], UniDoc [15], DocPedia [14], DocLLM [80], TextMonkey [44], mPLUG-DocOwl [23, 88], etc. Though promising results can be achieved for the DocVQA task, lacking fine-grained intermediate outputs like text locations and page layouts still greatly limits the interpretability and generalization of these models. As compensation, recently some LLM-free unified models are proposed for low-level document processing tasks such as UPOCR [59], DocRes [98], OmniParser [78] and DAT [79]. These works unify several similar or related tasks into unified models through multi-task learning, but at the cost of significant increment + +of model complexity and calculating overhead, prohibiting them from generalizing to to more tasks and datasets. + +# 2.3. Transformer-based Detection&Segmentation + +Following the pioneer work of DETR [4], many Transformer-based objection methods have been proposed in recent years, including Deformable DETR [108], DNDETR [33], DINO [96], Sparse R-CNN [73], etc. These methods share the same idea with DETR that rely on learnable queries and bipartite matching for object decoding, but make different modifications to improve the accuracy and convergence speed, such as bringing in deformable attention and denoising training, or assigning specific spatial meanings to the queries. + +Besides object detection, Transformer also shows great potential in image segmentation [6, 7, 24, 25, 30, 83, 97, 103]. Among them the most related works to this paper are SAM [30] and Mask2former [7]. Inspired by SAM [30] which uses natural language prompts to guide image segmentation, in this paper we propose to embed the class names of each dataset into semantic queries and transform various document segmentation tasks into a combination of instance segmentation and semantic segmentation. The semantic queries not only serve as prompts guiding the model in identifying specific types of regions, but also function as class prototypes that instance queries depend on for classification. Since DIS relies on high resolution image features, we build our DocSAM from Mask2former [7] which adopt Swin-Transformer [45] and deformable attention [108] as the vision encoder. Though the vision encoder of DocSAM is inherited from Mask2former, the decoder is drastically redesigned to be up to the task of general document image segmentation effectively. + +# 3. DocSAM + +# 3.1. Preliminaries + +Before introducing the proposed DocSAM, we first explore the key attributes of an ideal all-in-one DIS model and why current methods fall short of this goal. We assert that an exemplary all-in-one DIS model should possess the following attributes: it should have the versatility to convert diverse DIS tasks into a unified framework; it should be adaptable to training on heterogeneous datasets, accommodating diverse annotations without restrictions; and it should maintain the capacity for continual and incremental learning. Current methods are typically designed for specific DIS tasks and datasets, and most models become static after training, unable to efficiently incorporate new data, limiting their versatility and adaptability. Specifically, existing DIS methods rely on fully connected (FC) and Softmax layers to predict region classes, with FC's parameters predefined for specific tasks and datasets, making generalization difficult. + +![](images/860576fad4d4e862ce1333467fde1ff2d896d31e117085aa79f416ed0e08ac58.jpg) +Figure 2. Network structure of the proposed DocSAM. DocSAM unify various document image segmentation tasks into one single model through instance and semantic query decomposition and interaction. Skip connections and norm layers are omitted for simplicity. + +To overcome the above limitations and achieve the aforementioned criteria, the proposed DocSAM makes two significant improvements compared to existing methods. First, it transforms various DIS tasks into a unified paradigm of mask-based instance segmentation and semantic segmentation. Second, it embeds class names into semantic queries, which not only serve as prompts to guide the model in identifying specific types of regions to segment but also function as class prototypes that instance queries depend on for classification. The rest of this chapter presents the details of our proposed DocSAM. + +# 3.2. Vision Encoder + +Different DIS tasks may focus on contents of different scales, from large objects like paragraphs and figures spanning entire pages to tiny objects like chars and words covering only a few hundreds of pixels. Therefore, high-resolution multi-scale image features are an essential requirement for a unified DIS model. The vision encoder of DocSAM is adapted from Mask2Former [7], which includes a Swin-Transformer [45] as the vision backbone and deformable attention [108] for feature refinement. Additionally, we use another FPN [40] to fuse multi-scale image features $X_{I} = [X_{I}^{l}\in \mathbb{R}^{H_{l}W_{l}\times C},l\in \{1,2,3,4\}]$ into a single mask feature $X_{M}\in \mathbb{R}^{HW\times C}$ , which is used for subsequent semantic segmentation and instance segmentation. Here, $X_{I}^{l}$ is image feature of level $l$ , $H_{l}$ and $W_{l}$ are the spatial resolution of level $l$ , $C$ is number of feature channels. + +# 3.3. Query Embedding + +The instance queries $Q_{I} \in \mathbb{R}^{N \times C}$ of DocSAM is standard learnable queries, while the semantic queries $Q_{S} \in \mathbb{R}^{M \times C}$ are embedded from class names using the Sentence-BERT + +[64]. Here $N$ is a predefined instance query number that remains the same across all tasks and datasets, while $M$ is the semantic query number that may change depending on the class number of each dataset, and $C$ is feature dimension. $Q_{I}$ and $Q_{S}$ go together through the following Hybrid Query Decoder for feature decoding and cooperate with each other for semantic segmentation and instance segmentation. + +# 3.4. Hybrid Query Decoder + +Inside each HQD layer, see Fig. 2, we first concatenate $Q_{S}$ and $Q_{I}$ along the length dimension and send them into a multi-head self-attention layer (MHSA) followed by a feed forward layer (FFN). This step facilitates information exchange between $Q_{S}$ and $Q_{I}$ , allowing them to attend to each other for query fusion. Next, they are separately crossattended with the multi-scale image features $X_{I}$ in a coarse-to-fine manner by two Multi-Scale Decoders (MSD) each containing $L$ layers. Here, $L = 4$ stands for the number of feature scales. Each MSD layer consists of two MHSA layers, one multi-head cross-attention layer (MHCA) and one FFN layer. Following Mask2Former [7], we also use masked attention in MHCA, where the attention masks are derived from the predicted instance and semantic masks in the previous HQD layer. After that, $Q_{S}$ and $Q_{I}$ further interact with each other through another MHSA and FFN layer. We stack $K$ HQD layers for more refined predictions. + +# 3.5. Prediction Head + +The output $Q_{S}$ and $Q_{I}$ from each HQD layer along with the mask feature $X_{M}$ are sent to the Mask Predictor, Class Predictor and BBox Predictor for semantic mask segmentation, instance mask segmentation, instance category classification and instance bounding box regression, respectively. + +For predicting semantic and instance masks, $Q_{S}$ and $Q_{I}$ are multiplied with $X_{M}$ as: + +$$ +M _ {S} = \sigma \left(Q _ {S} \times X _ {M} ^ {T}\right), \tag {1} +$$ + +and + +$$ +M _ {I} = \sigma \left(Q _ {I} \times X _ {M} ^ {T}\right), \tag {2} +$$ + +where $M_S \in \mathbb{R}^{M \times HW}$ and $M_I \in \mathbb{R}^{N \times HW}$ are predicted semantic and instance masks, $\sigma$ is Sigmoid function, $T$ means matrix transposition, and $\times$ stands for matrix multiplication. Similarly, for predicting instance classes, $Q_I$ is multiplied with $Q_S$ as: + +$$ +Y _ {I} = \operatorname {S o f t m a x} \left(Q _ {I} \times Q _ {S} ^ {T}\right), \tag {3} +$$ + +where $Y_{I} \in \mathbb{R}^{N \times M}$ is predicted class probabilities of instances, softmax is Softmax function along the second dimension of $Y_{I}$ , $T$ means matrix transposition, and $\times$ stands for matrix multiplication. + +Since Eq. (1), Eq. (2), Eq. (3) are all based on matrix multiplication, they are actually calculating the similarities between $Q_{S}$ , $Q_{I}$ and $X_{M}$ . So we can also regard the $Q_{S}$ , $Q_{I}$ as instance and semantic prototypes. Through the above semantic query embedding and prototype-based instance classification, we transform the original close-set classifier into an open-set classifier, thus benefiting the construction of unified all-in-one DIS model. + +Besides mask segmentation, DocSAM also keep the ability of bbox prediction. This is realized through bounding box regression with the BBox Predictor. Following ISTR [24] and TransDLANet [9], for each HQD layer we predict the residual values of bbox coordinates relative to predictions of the previous HQD layer. + +# 3.6. Model Learning + +# 3.6.1. Loss Function + +There are four losses in DocSAM, namely semantic mask segmentation loss $L_{S}$ , instance mask segmentation loss $L_{I}$ , instance bbox regression loss $L_{B}$ , and instance classification loss $L_{C}$ . Among them, $L_{S}$ is calculated as: + +$$ +L _ {S} = \lambda_ {f} L _ {f o c a l} (M _ {S}, \hat {M} _ {S}) + \lambda_ {d} L _ {d i c e} (M _ {S}, \hat {M} _ {S}), \quad (4) +$$ + +where $M_S$ and $\hat{M}_S$ are predicted and ground-truth semantic masks, $L_{focal}$ and $L_{dice}$ are focal loss [67] and dice loss [54], respectively, and $\lambda_f = 10$ and $\lambda_d = 1$ are hyperparameters. Similarly, $L_I$ is calculated as: + +$$ +L _ {I} = \lambda_ {f} L _ {f o c a l} (M _ {I}, \hat {M} _ {I}) + \lambda_ {d} L _ {d i c e} (M _ {I}, \hat {M} _ {I}), \tag {5} +$$ + +where $M_S$ and $\hat{M}_S$ are predicted and ground-truth instance masks, respectively. $L_B$ is calculated as: + +$$ +L _ {B} = \lambda_ {s l 1} L _ {s l 1} (B _ {I}, \hat {B} _ {I}) + \lambda_ {d i o u} L _ {d i o u} (B _ {I}, \hat {B} _ {I}), \tag {6} +$$ + +where $B_{I}$ and $\hat{B}_I$ are predicted and ground-truth bounding boxes, $L_{sl1}$ and $L_{diou}$ are smooth L1 loss and distance IoU loss [105], respectively, and $\lambda_{sl1} = 1$ and $\lambda_{diou} = 1$ are hyper-parameters. At last, $L_{C}$ is calculated as: + +$$ +L _ {C} = L _ {c e} \left(Y _ {I}, \hat {Y} _ {I}\right), \tag {7} +$$ + +where $Y_{I}$ and $\hat{Y}_I$ are predicted and ground-truth instance labels, and $L_{ce}$ is cross entropy loss. The total loss of Doc-SAM is the sum of the above four losses: + +$$ +L = \lambda_ {s} L _ {S} + \lambda_ {i} L _ {I} + \lambda_ {b} L _ {B} + \lambda_ {c} L _ {C}, \tag {8} +$$ + +where $\lambda_s = 5$ , $\lambda_i = 5$ , $\lambda_b = 1$ , and $\lambda_c = 1$ are hyperparameters. We add auxiliary losses to every HQD layer and to query features before HQD. Following DETR [4] and Mask2Former [7], we also use bipartite matching to find the best matched instance predictions before calculating the loss. While for semantic predictions, there is no need to perform the bipartite matching, because the predictions and ground-truths are already one-to-one matched. + +# 3.6.2. Heterogeneous Mixed Learning + +Unlike existing methods, the novel design of DocSAM enables us to train a single model on heterogeneous mixed datasets. In this work, we collected nearly fifty DIS datasets of various document types and annotation formats, covering diverse DIS tasks from layout analysis and text detection to table structure recognition (see Tab. 1). We combined these datasets to construct a heterogeneous mixed dataset for training DocSAM. After training, the DocSAM model can be directly used as a versatile document segmenter or as a pre-trained model that can be seamlessly fine-tuned using task-specific datasets without any specialized modifications, such as adding or replacing a linear classification layer. This merit of DocSAM endows it with the potential for continual and incremental learning. + +# 3.6.3. Improving Training Efficiency + +Directly training DocSAM on such heterogeneous datasets may suffer from slow convergence and long training time, so we propose several strategies to improve training efficiency. Firstly, we pre-train the vision encoder of DocSAM on all 48 datasets using SimMIM [84], hoping it can provide more robust visual features for document images. Secondly, we separate the training datasets into groups with each group containing datasets of similar tasks and styles, see Tab. 1, then we adopt curriculum learning (CL) [81] strategy to warm up the training process by gradually adding new group of datasets. Thirdly, we add an instance query selection (IQS) process at the front of each HQD layer. Motivation behind this is that bipartite matching only calculates losses between matched predictions and ground-truths, and the matched query indexes are mostly the same across HQD layers. For a certain document, large ratio of instance + +![](images/9c0a89f71440e055410d888e55f44672bd3c8951e4b6666328905956878a4fbd.jpg) +(a) Training loss curves. +Figure 3. Loss curves and on-the-fly validation during training. + +![](images/36491aebc5bd586bb356f174b84f45010b1327fbffa1cbf40d98039795ff5cfb.jpg) +(b) On-the-fly validation. + +
Ablation SettingInstanceSemantic
mAP\( mAP_b \)mAFmIoU
DocSAMbase0.38040.35170.42560.6615
DocSAMbase + Curriculum0.38690.37040.43380.6622
DocSAMbase + SimMIM0.10020.06580.12940.3882
DocSAMbase + Freezing BERT0.38430.35100.42950.6677
DocSAMbase - Masked Attention0.23220.09380.28460.6327
DocSAMbase - Query Interaction0.29900.17790.35090.6511
DocSAMbase - Query Selection0.33410.31340.37630.6592
DocSAMlarge0.39000.36580.43200.6726
+ +Table 2. Ablation studies on model structure and training strategy. + +queries are not activated from beginning to end, and their class scores are very low. Therefore, we only select instance queries whose class scores are higher than a threshold $T_{k}$ before the $k$ -th HQD layer. We set $T_{k}$ as: $T_{k} = T_{max} / 2^{K - k}$ , where $T_{max} = 0.01$ is the maximum threshold, $K$ is the number of HQD layers, $k$ is the current HQD layer. Experiments show that IQS can discard low-score queries without degrading model performance, thereby improving training speed and reducing memory usage. + +# 4. Experiments + +# 4.1. Datasets and Metrics + +The datasets involved in our experiments are listed in Tab. 1. Underlined datasets (15 in total) are used for ablation studies, mixed pre-training, and dataset-specific fine-tuning. All 48 datasets are used for training the final DocSAM model. These datasets cover a wide range of domains and tasks, showing significant heterogeneity in document types, annotation formats, and other aspects. Typical examples are shown in Fig. 1, with more details provided in the supplementary material. For evaluation metrics, we use mIoU [46] for semantic segmentation and mAP (for masks) [39] and $\mathrm{mAP}_b$ (for bounding boxes) [39] for instance segmentation. Additionally, we introduce a new metric for instance segmentation called mAF, which is calculated as the mean F-score of all classes across all IoUs ranging from 0.5 to 0.95 in increments of 0.05 (i.e., [0.5:0.05:0.95]). + +# 4.2. Implementation Details + +The vision backbone and deformable attention module are initialized from Mask2Former [7], which is pre-trained on the COCO-panoptic dataset [39]. The Sentence-BERT is initialized using the all-MiniLM-L6-v2 model from the Sentence Transformers library [64]. Other parts of DocSAM are randomly initialized. We trained two sizes of models: DocSAM-base (207M parameters) and DocSAM-large (317M parameters). Their vision backbones use Swin-base and Swin-large, respectively, and the instance query numbers $N$ are set to 500 and 900, respectively. The HQD layer number $K$ is set to 4 by default. + +DocSAM is implemented based on PyTorch [58] and trained on $8 \times$ NVIDIA A800 GPUs. We use the AdamW optimizer [50] to train the model, setting the base learning rate to $4 \times 10^{-5}$ , and decay it using cosine annealing strategy [51]. For joint training on mixed datasets, the default settings are 80,000 iterations and a batch size of 32; for ablation studies and dataset-specific fine-tuning, the defaults are 20,000 iterations and a batch size of 8; for comparison with state-of-the-art, the defaults are 40,000 iterations and a batch size of 16. + +# 4.3. Main Results + +# 4.3.1. Ablation Studies + +To verify the effect of each module in DocSAM and select the best training strategy before large-scale training, we conducted a series of ablation studies, as shown in Tab. 2 and Fig. 3. The results in Tab. 2 are averaged over all 15 datasets. On-the-fly validation involves fast testing on a small number of samples (e.g. 10 for each dataset) during training. The results show that using curriculum learning and instance query selection can accelerate convergence and improve model performance, while SimMIM pre-training significantly degrades model performance, possibly due to the large gap between SimMIM and document segmentation. Since freezing the weights of Sentence-BERT has almost no impact on performance, we freeze them during + +
TaskDatasetDocSAM-baseDocSAM-large
InstanceSemanticInstanceSemantic
AP50AP75mAP\( mAP_b \)mAFmIoUAP50AP75mAP\( mAP_b \)mAFmIoU
Document Layout Analysis\( D^4 \)LA [13]0.5950.5140.4480.4380.4860.3890.6370.5620.4900.4730.5390.434
DocLayNet [60]0.7160.5280.4840.4800.5430.6070.7440.5700.5170.5010.5840.669
\( M^6 \)Doc [9]0.5190.4020.3630.3520.3810.2670.5510.4440.3970.3870.4250.296
PubLayNet [106]0.9360.8620.8060.7890.8470.8980.9460.8840.8300.8050.8680.911
Ancient and Handwritten Document SegmentationSCUT-CAB-logical [8]0.6810.5550.4810.4780.5020.4100.7170.5740.5110.4950.5340.454
SCUT-CAB-physical [8]0.9370.8370.7770.7470.8210.9370.9480.8560.7860.7540.8290.942
HJDataset [69]0.9560.9210.8810.8650.8950.8190.9560.9250.8850.8690.8980.821
CASIA-HWDB [41]0.9290.7850.7210.6640.7880.9350.9120.7700.7140.6430.7900.939
SCUT-HCCDoc [95]0.8650.6350.5440.5590.6250.8440.8690.6420.5490.5600.6250.847
Table Structure RecognitionFinTabNet [104]0.8670.7700.6640.6270.7570.8510.8690.7860.6840.6440.7780.860
PubTabNet [107]0.9700.7880.6430.6350.7140.8400.9700.7890.6480.6340.7230.845
TableBank-latex [35]0.9630.9470.8970.8680.9240.9400.9650.9500.9150.8930.9360.951
TableBank-word [35]0.8730.8370.8220.7930.8510.8440.8780.8440.8350.8140.8570.853
Scene Text DetectionCTW1500 [43]0.7120.4300.4000.3680.5000.7940.7530.4820.4410.4020.5310.817
Total-Text [11]0.7470.4210.4050.4070.7430.4530.7690.4540.4280.4210.4720.764
MSRA-TD500 [86]0.7470.5250.4580.4770.5020.7130.7980.5770.4960.5160.5320.739
ICDAR2015 [28]0.6130.2470.2940.3020.3380.5990.6390.2600.3070.3130.3450.623
+ +Table 3. Performance of DocSAM after joint pre-training. + +
TaskDatasetDocSAM-large from scratchDocSAM-large from pretrain
InstanceSemanticInstanceSemantic
AP50AP75mAP\( mAP_b \)mAFmIoUAP50AP75mAP\( mAP_b \)mAFmIoU
DocumentLayoutAnalysis\( D^4LA[13] \)0.3650.2590.2390.1940.2330.2050.6980.6370.5550.5460.5950.526
DocLayNet [60]0.5030.2920.2950.2600.3590.3650.8330.6910.6210.6010.6790.736
\( M^6Doc[9] \)0.2790.1730.1690.1450.1630.0870.6670.5660.5000.4850.5280.430
PubLayNet [106]0.8730.7590.6960.6220.7380.8410.9540.9040.8540.8500.8880.921
Ancient andHandwrittenDocument SegmentationSCUT-CAB-logical [8]0.3910.2330.2390.1360.2280.2260.7830.6310.5560.5300.5820.481
SCUT-CAB-physical [8]0.8010.6440.6050.4050.6640.9180.9460.8690.7990.7620.8420.945
HJDataset [69]0.8480.8350.7520.6060.7770.8120.9830.9480.9050.8950.9110.822
CASIA-HWDB [41]0.9080.7370.6650.6280.7370.9490.9770.9390.8930.7920.9160.956
SCUT-HCCDoc [95]0.8070.4920.4600.4230.5410.8530.9040.6840.5800.5890.6580.862
Table StructureRecognitionFinTabNet [104]0.3350.1640.1780.0040.2220.7700.8770.8050.7130.6810.8030.870
PubTabNet [107]0.0130.0100.0070.0420.0070.8100.9730.8210.6690.6530.7420.860
TableBank-latex [35]0.7620.6120.5650.0200.6410.9130.9680.9540.9260.9090.9470.958
TableBank-word [35]0.5940.4460.4350.0450.6190.8230.9080.8770.8710.8590.8810.873
Scene Text DetectionCTW1500 [43]0.4310.0980.1620.0710.2530.8000.7940.5390.4800.4530.5730.831
Total-Text [11]0.3130.0420.0960.0470.1680.7490.7940.5170.4600.4660.5020.775
MSRA-TD500 [86]0.5060.1890.2220.0760.2950.7310.8090.6040.5240.5410.5550.744
ICDAR2015 [28]0.2030.0280.0630.0230.1130.5970.6810.3160.3410.3530.3790.641
Unseen DatasetIIIT-AR-13K [55]0.5550.4300.4030.1850.4170.7390.8420.7020.6380.6210.6930.642
CHDAC-2022 [31]0.8860.6960.6040.5090.6490.9150.9390.8280.6870.6250.7270.918
+ +Table 4. Performance of DocSAM after dataset specific fine-tuning. + +training. Similar to Mask2Former, masked attention plays a crucial role in DocSAM, and removing it leads to a significant performance drop. Additionally, without query interaction, DocSAM's performance also decreases substantially, highlighting the importance of information exchange between instance and semantic queries. Finally, training a unified model on heterogeneous datasets heavily relies on the model's capacity, and using a more powerful vision backbone can greatly enhance model performance. + +# 4.3.2. Pre-training and Fine-tuning + +We train DocSAM on mixed heterogeneous datasets (15 datasets) to validate its performance as a unified document segmenter and a pre-trained model for dataset-specific fine + +tuning. The results are shown in Tab. 3 and Tab. 4. DocSAM achieves good semantic and instance segmentation performance on various datasets and tasks, though performance may vary across datasets due to differing levels of difficulty. As a single-modal model, DocSAM may underperform on datasets like $\mathrm{D}^4\mathrm{LA}$ [13], DocLayNet [60], $\mathrm{M}^6\mathrm{Doc}$ [9] and SCUT-CAB-logical [8], which require multi-modal information for fine-grained logical layout analysis. + +After joint training, we fine-tune DocSAM-large on each specific dataset to further improve performance. As shown in Tab. 4, fine-tuning results are significantly higher than direct testing and training from scratch. We also test DocSAM on unseen datasets IIIT-AR-13K [55] and CHDAC + +
MethodObjectInstance
mAPAP50AP75mAPAP50AP75
Faster R-CNN [65]0.4900.6780.5720.4780.6780.552
Mask R-CNN [20]0.4010.5840.4620.3970.5840.456
Deformable DETR [108]0.5720.7680.6340.5560.7650.611
ISTR [24]0.6270.8080.7080.6200.8070.702
TransDLANet [9]0.6450.8270.7270.6380.8260.719
DAT [79]0.712--0.657--
DocSAM0.6630.8400.7550.6610.8400.750
+ +Table 5. Performance comparison on ${\mathrm{M}}^{6}$ Doc. + +
MethodObjectInstance
mAPAP50AP75mAPAP50AP75
Physical
Faster R-CNN [65]0.7750.9130.8610.7530.9100.834
Mask R-CNN [20]0.7910.9210.8770.7950.9170.872
SCNet [77]0.8130.9410.8900.8200.9410.891
Deformable DETR [108]0.7990.9230.8710.7790.9210.843
VSR [99]0.7870.9190.8600.7870.9190.852
DocSAM0.7740.9470.8600.8110.9480.891
Logical
Faster R-CNN [65]0.5490.7740.6130.5420.7730.606
Mask R-CNN [20]0.5510.7850.6190.5530.7770.631
SCNet [77]0.6020.8360.6730.6030.8360.680
Deformable DETR [108]0.6270.8520.7170.6200.8510.703
VSR [99]0.5570.7830.6160.5510.7820.611
DocSAM0.5480.7690.6320.5750.7790.667
+ +Table 6. Performance comparison on SCUT-CAB. + +
MethodCTW1500Total-Text
PRFPRF
HierText [49]0.8460.8740.8600.8550.9050.879
SIR [61]0.8740.8370.8550.9090.8560.882
DPText-DETR [89]0.9170.8620.8880.9180.8640.890
UNITS [29]-----0.898
ESTextSpotter [26]0.9150.8860.9000.9200.8810.900
DAT-DET [79]0.8930.8930.8930.9400.8820.910
DAT-SEG [79]0.9250.9090.9170.9500.8920.920
DocSAM0.8050.8810.8420.7210.8260.770
+ +Table 7. Performance comparison on CTW1500 and Total-Text. + +2022 [31], where fine-tuning from the pre-trained model also yields substantial performance gains. This demonstrates that DocSAM's performance is not yet saturated and can benefit greatly from transfer learning on unseen datasets and tasks. + +# 4.3.3. Comparison with State-of-the-Arts + +To compare with state-of-the-art methods, we further fine-tuned DocSAM on some datasets for additional training iterations. The results are shown in Tab. 5, Tab. 6, and Tab. 7. The best results are shown in bold, and the second-best results are underlined. DocSAM achieves superior or comparable performance with other methods. Note that we did not apply any specific training techniques or data augmentation, configurations for all datasets were kept consistent. We found that DocSAM exhibits much lower performance in logical layout analysis compared to physical analysis, + +which we attribute to its reliance only on single-modal features. Furthermore, DocSAM achieved relatively low performance on scene text detection datasets. This is likely because scene texts exhibit much greater diversity in shapes and backgrounds, requiring more carefully designed strategies to ensure model performance. + +# 4.4. Discussion + +The goal of this paper is not to achieve state-of-the-art performance on specific dataset and task through meticulously designed model architectures or training strategies. Instead, we aim to design a simple and unified document segmentation model that can be applied to a wide variety of datasets and tasks. Additionally, the trained model should possess good scalability and the ability to continue learning. In this regard, DocSAM is quite successful. It exhibits decent performance on various datasets and tasks and shows great potential for downstream applications both as a versatile segmenter and a pre-trained model. However, experimental results also reveal some weaknesses and limitations of DocSAM, such as long training time and unsatisfactory performance on complex scenarios. We believe that DocSAM can greatly benefit from more sophisticated model design and better data augmentation and training strategies to further accelerate its convergence and improve its performance. + +# 5. Conclusion + +In this paper, we propose DocSAM, a transformer-based unified framework for various document image segmentation tasks. DocSAM integrates layout analysis, multi-grained text segmentation, and table structure decomposition into a single model, reducing the need for specialized models and enhancing efficiency. Trained on heterogeneous datasets, DocSAM demonstrates robust and generalizable performance, effectively handling diverse document types and structures. This approach also reduces computational and storage requirements, making DocSAM suitable for practical deployment in resource-constrained environments. Extensive experiments show that DocSAM outperforms existing methods in terms of accuracy, efficiency, and adaptability. Overall, we believe that DocSAM represents a significant step forward for document image segmentation, and we look forward to its continued development and application in practical scenarios. In the future, we plan to extend DocSAM to a multi-modal version and explore better training strategies to further accelerate its convergence and improve its performance. + +# Acknowledgement + +This work is supported by the National Natural Science Foundation of China(NSFC) Grant U23B2029. + +# References + +[1] Abdelrahman Abdallah, Alexander Berendeyev, Islam Nuradin, and Daniyar Nurseitov. Tncr: Table net detection and classification dataset. Neurocomputing, 473:79-97, 2022. 3 +[2] Avinash Anand, Raj Jaiswal, Mohit Gupta, Siddhesh S Bangar, Pijush Bhuyan, Naman Lal, Rajeev Singh, Ritika Jha, Rajiv Ratn Shah, and Shin'Ichi Satoh. Ranlaynet: A dataset for document layout detection used for domain adaptation and generalization. In ACM MM Asia, pages 1-6, 2023. 3 +[3] buptlihang. CDLA: A Benchmark Dataset for Cross-Domain Layout Analysis. https://github.com/buptliang/CDLA, 2023. Accessed: 2024-10-31. 3 +[4] Nicolas Carion, Francisco Massa, Gabriel Synnaeve, Nicolas Usunier, Alexander Kirillov, and Sylvain Gelly. End-to-end object detection with transformers. In ECCV, pages 213-229. Springer, 2020. 3, 5 +[5] Liang-Chieh Chen, George Papandreou, Iasonas Kokkinos, Vladimir Koltun, and Alan Garg. Deep convolutional nets, atrous convolution, and fully connected crfs. In IEEE Trans. Pattern Anal. Mach. Intell., pages 834-848, 2018. 3 +[6] Bowen Cheng, Alex Schwing, and Alexander Kirillov. Per-pixel classification is not all you need for semantic segmentation. NeurIPS, 34:17864-17875, 2021. 3 +[7] Bowen Cheng, Ishan Misra, Alexander G Schwing, Alexander Kirillov, and Rohit Girdhar. Masked-attention mask transformer for universal image segmentation. In CVPR, pages 1290–1299, 2022. 3, 4, 5, 6 +[8] Hiuyi Cheng, Cheng Jian, Sihang Wu, and Lianwen Jin. Scut-cab: a new benchmark dataset of ancient chinese books with complex layouts for document layout analysis. In ICFHR, pages 436-451. Springer, 2022. 2, 3, 7 +[9] Hiuyi Cheng, Peirong Zhang, Sihang Wu, Jiaxin Zhang, Qiyuan Zhu, Zecheng Xie, Jing Li, Kai Ding, and Lianwen Jin. M6doc: A large-scale multi-format, multi-type, multilayout, multi-language, multi-annotation category dataset for modern document layout analysis. In CVPR, pages 15138–15147, 2023. 2, 3, 5, 7, 8 +[10] Zewen Chi, Heyan Huang, Heng-Da Xu, Houjin Yu, Wanxuan Yin, and Xian-Ling Mao. Complicated table structure recognition. arXiv preprint arXiv:1908.04729, 2019. 3 +[11] Chee Kheng Ch'ng and Chee Seng Chan. Total-text: A comprehensive dataset for scene text detection and recognition. In ICDAR, pages 935-942. IEEE, 2017. 2, 3, 7 +[12] Chee Kheng Chng, Yuliang Liu, Yipeng Sun, Chun Chet Ng, Canjie Luo, Zihan Ni, ChuanMing Fang, Shuai tao Zhang, Junyu Han, Errui Ding, et al. Icdar2019 robust reading challenge on arbitrary-shaped text-rrc-art. In ICDAR, pages 1571-1576. IEEE, 2019. 3 +[13] Cheng Da, Chuwei Luo, Qi Zheng, and Cong Yao. Vision grid transformer for document layout analysis. In ICCV, pages 19462-19472, 2023. 3, 7 +[14] Hao Feng, Qi Liu, Hao Liu, Wengang Zhou, Houqiang Li, and Can Huang. Docpedia: Unleashing the power of large multimodal model in the frequency domain for versatile document understanding. arXiv preprint arXiv:2311.11810, 2023. 3 + +[15] Hao Feng, Zijian Wang, Jingqun Tang, Jinghui Lu, Wengang Zhou, Houqiang Li, and Can Huang. Unidoc: A universal large multimodal model for simultaneous text detection, recognition, spotting and understanding. arXiv preprint arXiv:2308.11592, 2023. 3 +[16] Liangcai Gao, Xiaohan Yi, Zhuoren Jiang, Leipeng Hao, and Zhi Tang. Icdar2017 competition on page object detection. In ICDAR, pages 1417-1422. IEEE, 2017. 2, 3 +[17] Liangcai Gao, Yilun Huang, Herve Dejean, Jean-Luc Meunier, Qinqin Yan, Yu Fang, Florian Kleber, and Eva Lang. Icdar 2019 competition on table detection and recognition (ctdar). In ICDAR, pages 1510-1515. IEEE, 2019. 2, 3 +[18] Max Göbel, Tamir Hassan, Ermelinda Oro, and Giorgio Orsi. Icdar 2013 table competition. In ICDAR, pages 1449-1453. IEEE, 2013. 3 +[19] Mrinal Haloi, Shashank Shekhar, Nikhil Fande, Siddhant Swaroop Dash, et al. Table detection in the wild: A novel diverse table detection dataset and method. arXiv preprint arXiv:2209.09207, 2022.3 +[20] Kaiming He, Georgia Gkioxari, Piotr Dollar, and Ross Girshick. Mask r-cnn. In ICCV, pages 2961-2969, 2017. 3, 8 +[21] Mengchao He, Yuliang Liu, Zhibo Yang, Sheng Zhang, Canjie Luo, Feiyu Gao, Qi Zheng, Yongpan Wang, Xin Zhang, and Lianwen Jin. Icpr2018 contest on robust reading for multi-type web images. In ICPR, pages 7-12. IEEE, 2018. 3 +[22] Wenhao He, Xu-Yao Zhang, Fei Yin, and Cheng-Lin Liu. Multi-oriented and multi-lingual scene text detection with direct regression. ICIP, 27(11):5406-5419, 2018. 3 +[23] Anwen Hu, Haiyang Xu, Jiabo Ye, Ming Yan, Liang Zhang, Bo Zhang, Chen Li, Ji Zhang, Qin Jin, Fei Huang, et al. mplug-docowl 1.5: Unified structure learning forOCR-free document understanding. arXiv preprint arXiv:2403.12895, 2024.3 +[24] Jie Hu, Liujuan Cao, Yao Lu, ShengChuan Zhang, Yan Wang, Ke Li, Feiyue Huang, Ling Shao, and Rongrong Ji. Istr: End-to-end instance segmentation with transformers. arXiv preprint arXiv:2105.00637, 2021. 3, 5, 8 +[25] Jie Hu, Yao Lu, Shengchuan Zhang, and Liujuan Cao. Istr: Mask-embedding-based instance segmentation transformer. IEEE Trans. Image Process., 2024. 3 +[26] Mingxin Huang, Jiaxin Zhang, Dezhi Peng, Hao Lu, Can Huang, Yuliang Liu, Xiang Bai, and Lianwen Jin. Estextspotter: Towards better scene text spotting with explicit synergy in transformer. In ICCV, pages 19495-19505, 2023. 8 +[27] Yupan Huang, Tengchao Lv, Lei Cui, Yutong Lu, and Furu Wei. Layoutlmv3: Pre-training for document ai with unified text and image masking. In ACM MM, pages 4083-4091, 2022. 3 +[28] Dimosthenis Karatzas, Lluis Gomez-Bigorda, Anguelos Nicolaou, Suman Ghosh, Andrew Bagdanov, Masakazu Iwamura, Jiri Matas, Lukas Neumann, Vijay Ramaseshan Chandrasekhar, Shijian Lu, et al. Icdar 2015 competition on robust reading. In ICDAR, pages 1156-1160. IEEE, 2015. 3, 7 + +[29] Taeho Kil, Seonghyeon Kim, Sukmin Seo, Yoonsik Kim, and Daehee Kim. Towards unified scene text spotting based on sequence generation. In CVPR, pages 15223-15232, 2023. 8 +[30] Alexander Kirillov, Eric Mintun, Nikhila Ravi, Hanzi Mao, Chloe Rolland, Laura Gustafson, Tete Xiao, Spencer Whitehead, Alexander C Berg, Wan-Yen Lo, et al. Segment anything. In ICCV, pages 4015-4026, 2023. 3 +[31] Pazhou Lab. IACC competition on Chinese Historical Document Analysis Challenge. https://iacc.pazhoulab-huangpu.com/contestdetail?id=6497f74cd97a2dae9dcaeff8 & award=1, 000, 000, 2022. Accessed: 2024-10-31. 3, 7, 8 +[32] John Lafferty, Andrew McCallum, Fernando Pereira, et al. Conditional random fields: Probabilistic models for segmenting and labeling sequence data. In ICML, page 3. Williamstown, MA, 2001. 3 +[33] Feng Li, Hao Zhang, Shilong Liu, Jian Guo, Lionel M Ni, and Lei Zhang. Dn-detr: Accelerate detr training by introducing query denoising. In CVPR, pages 13619-13627, 2022. 3 +[34] Junlong Li, Yiheng Xu, Tengchao Lv, Lei Cui, Cha Zhang, and Furu Wei. Dit: Self-supervised pre-training for document image transformer. In ACM MM, pages 3530-3539, 2022. 3 +[35] Minghao Li, Lei Cui, Shaohan Huang, Furu Wei, Ming Zhou, and Zhoujun Li. Tablebank: Table benchmark for image-based table detection and recognition. In LREC, pages 1918-1925, 2020. 2, 3, 7 +[36] Minghao Li, Yiheng Xu, Lei Cui, Shaohan Huang, Furu Wei, Zhoujun Li, and Ming Zhou. Docbank: A benchmark dataset for document layout analysis. arXiv preprint arXiv:2006.01038, 2020. 2, 3 +[37] Xiao-Hui Li, Fei Yin, and Cheng-Lin Liu. Page segmentation using convolutional neural network and graphical model. In DAS, pages 231-245. Springer, 2020. 3 +[38] Xiao-Hui Li, Fei Yin, He-Sen Dai, and Cheng-Lin Liu. Table structure recognition and form parsing by end-to-end object detection and relation parsing. Pattern Recognition, 132:108946, 2022. 2, 3 +[39] Tsung-Yi Lin, Michael Maire, Serge Belongie, James Hays, Pietro Perona, Deva Ramanan, Piotr Dólar, and C Lawrence Zitnick. Microsoft coco: Common objects in context. In ECCV, pages 740-755. Springer, 2014. 6 +[40] Tsung-Yi Lin, Piotr Dólar, Ross Girshick, Kaiming He, Bharath Hariharan, and Serge Belongie. Feature pyramid networks for object detection. In CVPR, pages 2117-2125, 2017. 4 +[41] Cheng-Lin Liu, Fei Yin, Da-Han Wang, and Qiu-Feng Wang. Casia online and offline chinese handwriting databases. In ICDAR, pages 37-41. IEEE, 2011. 3, 7 +[42] Cheng-Lin Liu, Lianwen Jin, Xiang Bai, Xiaohui Li, and Fei Yin. Frontiers of intelligent document analysis and recognition: review and prospects. Journal of Image and Graphics, 28(08):2223-2252, 2023. 1 +[43] Yuliang Liu, Lianwen Jin, Shuai tao Zhang, Canjie Luo, and Sheng Zhang. Curved scene text detection via transverse + +and longitudinal sequence connection. Pattern Recognition, 90:337-345, 2019. 2, 3, 7 +[44] Yuliang Liu, Biao Yang, Qiang Liu, Zhang Li, Zhiyin Ma, Shuo Zhang, and Xiang Bai. Textmonkey: AnOCR-free large multimodal model for understanding document. arXiv preprint arXiv:2403.04473, 2024. 3 +[45] Ze Liu, Yutong Lin, Yan Cao, Yue Hu, Yu Wei, Zhiqiang Zhang, Jiahui Lin, Han Wang, Cheng Lu, Changhu Wang, et al. Swin transformer: Hierarchical vision transformer using shifted windows. In ICCV, pages 12226-12235, 2021. 3, 4 +[46] Jonathan Long, Evan Shelhamer, and Trevor Darrell. Fully convolutional networks for semantic segmentation. In CVPR, pages 3431-3440, 2015. 3, 6 +[47] Rujiao Long, Wen Wang, Nan Xue, Feiyu Gao, Zhibo Yang, Yongpan Wang, and Gui-Song Xia. Parsing table structures in the wild. In ICCV, pages 944-952, 2021. 3 +[48] Shangbang Long, Jiaqiang Ruan, Wenjie Zhang, Xin He, Wenhao Wu, and Cong Yao. Textsnake: A flexible representation for detecting text of arbitrary shapes. In ECCV, pages 20-36, 2018. 3 +[49] Shangbang Long, Siyang Qin, Dmitry Panteleev, Alessandro Bissacco, Yasuhisa Fujii, and Michalis Raptis. Towards end-to-end unified scene text detection and layout analysis. In CVPR, pages 1049-1059, 2022. 3, 8 +[50] I Loshchilov. Decoupled weight decay regularization. arXiv preprint arXiv:1711.05101, 2017. 6 +[51] Ilya Loshchilov and Frank Hutter. Sgdr: Stochastic gradient descent with warm restarts. arXiv preprint arXiv:1608.03983, 2016. 6 +[52] Siwen Luo, Yihao Ding, Siqu Long, Josiah Poon, and Soyeon Caren Han. Doc-gcn: Heterogeneous graph convolutional networks for document layout analysis. arXiv preprint arXiv:2208.10970, 2022. 3 +[53] Weihong Ma, Hesuo Zhang, Lianwen Jin, Sihang Wu, Jiapeng Wang, and Yongpan Wang. Joint layout analysis, character detection and recognition for historical document digitization. In ICFHR, pages 31-36. IEEE, 2020. 2, 3 +[54] Fausto Miletari, Nassir Navab, and Seyed-Ahmad Ahmadi. V-net: Fully convolutional neural networks for volumetric medical image segmentation. In 3DV, pages 565-571. IEEE, 2016. 5 +[55] Ajoy Mondal, Peter Lipps, and CV Jawahar. Iit-ar-13k: A new dataset for graphical object detection in documents. In DAS, pages 216–230. Springer, 2020. 3, 7 +[56] Nibal Nayef, Fei Yin, Imen Bizid, Hyunsoo Choi, Yuan Feng, Dimosthenis Karatzas, Zhenbo Luo, Umapada Pal, Christophe Rigaud, Joseph Chazalon, et al. Icdar2017 robust reading challenge on multi-lingual scene text detection and script identification-rrc-plt. In ICDAR, pages 1454-1459. IEEE, 2017. 3 +[57] Nibal Nayef, Yash Patel, Michal Busta, Pinaki Nath Chowdhury, Dimosthenis Karatzas, Wafa Khlif, Jiri Matas, Umapada Pal, Jean-Christophe Burie, Cheng-lin Liu, et al. Icdar2019 robust reading challenge on multi-lingual scene text detection and recognition—rrc-mlt-2019. In ICDAR, pages 1582–1587. IEEE, 2019. 3 + +[58] Adam Paszke, Sam Gross, Francisco Massa, Adam Lerer, James Bradbury, Gregory Chanan, Trevor Killeen, Zeming Lin, Natalia Gimelshein, Luca Antiga, et al. Pytorch: An imperative style, high-performance deep learning library. NeurIPS, 32, 2019. 6 +[59] Dezhi Peng, Zhenhua Yang, Jiaxin Zhang, Chongyu Liu, Yongxin Shi, Kai Ding, Fengjun Guo, and Lianwen Jin. Upocr: Towards unified pixel-levelOCR interface. In ICML, 2023. 3 +[60] Birgit Pfitzmann, Christoph Auer, Michele Dolfi, Ahmed S. Nassar, and Peter Staar. Doclaynet: A large human-annotated dataset for document layout analysis. In ACM SIGKDD, pages 3743–3751. 2022. 2, 3, 7 +[61] Xugong Qin, Pengyuan Lyu, Chengquan Zhang, Yu Zhou, Kun Yao, Peng Zhang, Hailun Lin, and Weiping Wang. Towards robust real-time scene text detection: From semantic to instance representation learning. In ACM MM, pages 2025-2034, 2023. 8 +[62] Joseph Redmon and Ali Farhadi. You only look once: Unified, real-time object detection. In CVPR, pages 779-788, 2016. 3 +[63] Joseph Redmon and Ali Farhadi. Yolov3: An incremental improvement, 2018. 3 +[64] N Reimers. Sentence-bert: Sentence embeddings using siamese bert-networks. arXiv preprint arXiv:1908.10084, 2019. 1, 4, 6 +[65] Shaoqing Ren, Kaiming He, Ross Girshick, and Jian Sun. Faster r-cnn: Towards real-time object detection with region proposal networks. In NeurIPS, pages 91-99, 2015. 3, 8 +[66] Olaf Ronneberger, Philipp Fischer, and Thomas Brox. U-net: Convolutional networks for biomedical image segmentation. In MICCAI, pages 234-241. Springer, 2015. 3 +[67] T-YLPG Ross and GKHP Dólar. Focal loss for dense object detection. In CVPR, pages 2980-2988, 2017. 5 +[68] Rajkumar Saini, Derek Dobson, Jon Morrey, Marcus Liwicki, and Foteini Simistira Liwicki. Icdar 2019 historical document reading challenge on large structured chinese family records. In ICDAR, pages 1499-1504. IEEE, 2019. 3 +[69] Zejiang Shen, Kaixuan Zhang, and Melissa Dell. A large dataset of historical japanese documents with complex layouts. In CVPR Workshops, pages 548-549, 2020. 3, 7 +[70] Baoguang Shi, Cong Yao, Minghui Liao, Mingkun Yang, Pei Xu, Linyan Cui, Serge Belongie, Shijian Lu, and Xiang Bai. Icdar2017 competition on reading chinese text in the wild (rctw-17). In ICDAR, pages 1429–1434. IEEE, 2017. 3 +[71] Md Istiak Hossain Shihab, Md Rakibul Hasan, Mahfuzur Rahman Emon, Syed Mobassir Hossen, Md Nazmuddoha Ansary, Intesur Ahmed, Fazle Rabbi Rakib, Shahriar Elahi Dhruvo, Souhardya Saha Dip, Akib Hasan Pavel, et al. Badlad: A large multi-domain bengali document layout analysis dataset. In ICDAR, pages 326-341. Springer, 2023. 3 +[72] Brandon Smock, Rohith Pesala, and Robin Abraham. Pubtables-1m: Towards comprehensive table extraction + +from unstructured documents. In CVPR, pages 4634-4642, 2022. 2, 3 +[73] Xingyi Sun, Ziyi Zhou, Peizhao Wang, Junwei Dai, Lu Lu, Wanli Ouyang, and Ping Luo. Sparse r-cnn: End-to-end object detection with learnable proposals. In CVPR, pages 13164-13173, 2021. 3 +[74] Yipeng Sun, Zihan Ni, Chee-Kheng Chng, Yuliang Liu, Canjie Luo, Chun Chet Ng, Junyu Han, Errui Ding, Jingtuo Liu, Dimosthenis Karatzas, et al. Icdar 2019 competition on large-scale street view text with partial labeling-rrc-lsvt. In ICDAR, pages 1557-1562. IEEE, 2019. 3 +[75] Zineng Tang, Ziyi Yang, Guoxin Wang, Yuwei Fang, Yang Liu, Chenguang Zhu, Michael Zeng, Cha Zhang, and Mohit Bansal. Unifying vision, text, and layout for universal document processing. In CVPR, pages 19254-19264, 2023. 3 +[76] Andreas Veit, Tomas Matera, Lukas Neumann, Jiri Matas, and Serge Belongie. Coco-text: Dataset and benchmark for text detection and recognition in natural images. arXiv preprint arXiv:1601.07140, 2016. 3 +[77] Thang Vu, Haeyong Kang, and Chang D Yoo. Scnet: Training inference sample consistency for instance segmentation. In AAAI, pages 2701-2709, 2021. 8 +[78] Jianqiang Wan, Sibo Song, Wenwen Yu, Yuliang Liu, Wenqing Cheng, Fei Huang, Xiang Bai, Cong Yao, and Zhibo Yang. Omniparser: A unified framework for text spotting key information extraction and table recognition. In CVPR, pages 15641-15653, 2024. 3 +[79] Xingyu Wan, Chengquan Zhang, Pengyuan Lyu, Sen Fan, Zihan Ni, Kun Yao, Errui Ding, and Jingdong Wang. Towards unified multi-granularity text detection with interactive attention. arXiv preprint arXiv:2405.19765, 2024. 3, 8 +[80] Dongsheng Wang, Natraj Raman, Mathieu Sibue, Zhiqiang Ma, Petr Babkin, Simerjot Kaur, Yulong Pei, Armineh Nourbakhsh, and Xiaomo Liu. Docllm: A layout-aware generative language model for multimodal document understanding. arXiv preprint arXiv:2401.00908, 2023. 3 +[81] Xin Wang, Yudong Chen, and Wenwu Zhu. A survey on curriculum learning. IEEE Trans. Pattern Anal. Mach. Intell., 44(9):4555-4576, 2021. 5 +[82] Zonghan Wu, Shirui Pan, Fengwen Chen, Guodong Long, Chengqi Zhang, and S Yu Philip. A comprehensive survey on graph neural networks. IEEE Trans. Neural Netw. Learn. Syst., 32(1):4-24, 2020. 3 +[83] Enze Xie, Zhiwei Wang, Gang Yu, Xingyi Zhang, Yuheng Wu, Hongsheng Li, and Yonghong Tian. Segformer: Simple and efficient design for semantic segmentation with transformers. In ICCV, pages 14125-14134, 2021. 3 +[84] Zhenda Xie, Zheng Zhang, Yue Cao, Yutong Lin, Jianmin Bao, Zhuliang Yao, Qi Dai, and Han Hu. Simmim: A simple framework for masked image modeling. In CVPR, pages 9653-9663, 2022. 5 +[85] Yue Xu, Fei Yin, Da-Han Wang, Xu-Yao Zhang, Zhaoxiang Zhang, and Cheng-Lin Liu. Casia-ahcdb: A large-scale chinese ancient handwritten characters database. In ICDAR, pages 793–798. IEEE, 2019. 3 + +[86] Cong Yao, Xiang Bai, Wenyu Liu, Yi Ma, and Zhuowen Tu. Detecting texts of arbitrary orientations in natural images. In CVPR, pages 1083-1090. IEEE, 2012. 3, 7 +[87] Cong Yao, Xiang Bai, and Wenyu Liu. A unified framework for multioriented text detection and recognition. ICIP, 23 (11):4737-4749, 2014. 3 +[88] Jiabo Ye, Anwen Hu, Haiyang Xu, Qinghao Ye, Ming Yan, Yuhao Dan, Chenlin Zhao, Guohai Xu, Chenliang Li, Junfeng Tian, et al. mplug-docowl: Modularized multimodal large language model for document understanding. arXiv preprint arXiv:2307.02499, 2023. 3 +[89] Maoyuan Ye, Jing Zhang, Shanshan Zhao, Juhua Liu, Bo Du, and Dacheng Tao. Dptext-detr: Towards better scene text detection with dynamic points in transformer. In AAAI, pages 3241-3249, 2023. 8 +[90] Xu-Cheng Yin, Xuwang Yin, Kaizhu Huang, and Hong-Wei Hao. Robust text detection in natural scene images. IEEE Trans. Pattern Anal. Mach. Intell., 36(5):970–983, 2013. 3 +[91] Wenwen Yu, Mingyu Liu, Mingrui Chen, Ning Lu, Yinlong Wen, Yuliang Liu, Dimosthenis Karatzas, and Xiang Bai. Icdar 2023 competition on reading the seal title. In ICDAR, pages 522-535. Springer, 2023. 3 +[92] Tai-Ling Yuan, Zhe Zhu, Kun Xu, Cheng-Jun Li, Tai-Jiang Mu, and Shi-Min Hu. A large chinese text dataset in the wild. J. Comput. Sci. Tech., 34:509-521, 2019. 3 +[93] Yukun Zhai, Xiaogiang Zhang, Xiameng Qin, Sanyuan Zhao, Xingping Dong, and Jianbing Shen. Textformer: A query-based end-to-end text spotter with mixed supervision. Machine Intelligence Research, 21(4):704-717, 2024. 2 +[94] Chongsheng Zhang, Guowen Peng, Yuefeng Tao, Feifei Fu, Wei Jiang, George Almpanidis, and Ke Chen. Shopsign: A diverse scene text dataset of chinese shop signs in street views. arXiv preprint arXiv:1903.10412, 2019. 3 +[95] Hesuo Zhang, Lingyu Liang, and Lianwen Jin. Scuthcdoc: A new benchmark dataset of handwritten chinese text in unconstrained camera-captured documents. Pattern Recognition, 108:107559, 2020. 2, 3, 7 +[96] Hao Zhang, Feng Li, Shilong Liu, Lei Zhang, Hang Su, Jun Zhu, Lionel M Ni, and Heung-Yeung Shum. Dino: Detr with improved denoising anchor boxes for end-to-end object detection. arXiv preprint arXiv:2203.03605, 2022. 3 +[97] Hao Zhang, Feng Li, Huaizhe Xu, Shijia Huang, Shilong Liu, Lionel M Ni, and Lei Zhang. Mp-former: Mask-piloted transformer for image segmentation. In CVPR, pages 18074-18083, 2023. 3 +[98] Jiaxin Zhang, Dezhi Peng, Chongyu Liu, Peirong Zhang, and Lianwen Jin. Docres: A generalist model toward unifying document image restoration tasks. In CVPR, pages 15654-15664, 2024. 3 +[99] Peng Zhang, Can Li, Liang Qiao, Zhanzhan Cheng, Shiliang Pu, Yi Niu, and Fei Wu. Vsr: a unified framework for document layout analysis combining vision, semantics and relations. In ICDAR, pages 115-130. Springer, 2021. 8 +[100] Rui Zhang, Yongsheng Zhou, Qianyi Jiang, Qi Song, Nan Li, Kai Zhou, Lei Wang, Dong Wang, Minghui Liao, + +Mingkun Yang, et al. Icdar 2019 robust reading challenge on reading chinese text on signboard. In ICDAR, pages 1577-1581. IEEE, 2019. 3 +[101] Shi-Xue Zhang, Xiaobin Zhu, Jie-Bo Hou, Chang Liu, Chun Yang, Hongfa Wang, and Xu-Cheng Yin. Deep relational reasoning graph network for arbitrary shape text detection. In CVPR, pages 9699–9708, 2020. 3 +[102] Wayne Xin Zhao, Kun Zhou, Junyi Li, Tianyi Tang, Xiaolei Wang, Yupeng Hou, Yingqian Min, Beichen Zhang, Junjie Zhang, Zican Dong, et al. A survey of large language models. arXiv preprint arXiv:2303.18223, 2023. 3 +[103] Enze Zheng, Zhiwei Wang, Gang Yu, Xingyi Zhang, Yuheng Wu, Hongsheng Li, and Yonghong Tian. Rethinking semantic segmentation from a sequence-to-sequence perspective with transformers. In CVPR, pages 12164-12173, 2021. 3 +[104] Xinyi Zheng, Douglas Burdick, Lucian Popa, Xu Zhong, and Nancy Xin Ru Wang. Global table extractor (gte): A framework for joint table identification and cell structure recognition using visual context. In WACV, pages 697-706, 2021. 3, 7 +[105] Zhaohui Zheng, Ping Wang, Wei Liu, Jinze Li, Rongguang Ye, and Dongwei Ren. Distance-iou loss: Faster and better learning for bounding box regression. In AAAI, pages 12993-13000, 2020. 5 +[106] Xu Zhong, Jianbin Tang, and Antonio Jimeno Yepes. Publaynet: largest dataset ever for document layout analysis. In ICDAR, pages 1015-1022. IEEE, 2019. 2, 3, 7 +[107] Xu Zhong, Elaheh ShafieiBavani, and Antonio Jimeno Yepes. Image-based table recognition: data, model, and evaluation. In ECCV, pages 564-580. Springer, 2020. 2,3,7 +[108] Xingyi Zhu, Junwei Dai, Lu Lu, Yuwen Zhang, Peizhao Wang, Yisen Sun, Wanli Ouyang, and Ping Luo. Deformable detr: Deformable transformers for end-to-end object detection. In ICCV, pages 12352-12361, 2021. 1, 3, 4, 8 +[109] Ziyi Zhu, Liangcai Gao, Yibo Li, Yilun Huang, Lin Du, Ning Lu, and Xianfeng Wang. Ntable: a dataset for camera-based table detection. In ICDAR, pages 117-129. Springer, 2021. 3 \ No newline at end of file diff --git a/2025/DocSAM_ Unified Document Image Segmentation via Query Decomposition and Heterogeneous Mixed Learning/images.zip b/2025/DocSAM_ Unified Document Image Segmentation via Query Decomposition and Heterogeneous Mixed Learning/images.zip new file mode 100644 index 0000000000000000000000000000000000000000..d13e74148ee433b62d826a8e4329cec419e67c8e --- /dev/null +++ b/2025/DocSAM_ Unified Document Image Segmentation via Query Decomposition and Heterogeneous Mixed Learning/images.zip @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:c25e2f914242e258573dc6652a6ff1a465e73d3453b9fbac25f6c06027261c79 +size 951873 diff --git a/2025/DocSAM_ Unified Document Image Segmentation via Query Decomposition and Heterogeneous Mixed Learning/layout.json b/2025/DocSAM_ Unified Document Image Segmentation via Query Decomposition and Heterogeneous Mixed Learning/layout.json new file mode 100644 index 0000000000000000000000000000000000000000..1b7107e8cdf119686097ba450d1f5ffe355d59ac --- /dev/null +++ b/2025/DocSAM_ Unified Document Image Segmentation via Query Decomposition and Heterogeneous Mixed Learning/layout.json @@ -0,0 +1,10416 @@ +{ + "pdf_info": [ + { + "para_blocks": [ + { + "bbox": [ + 58, + 102, + 552, + 140 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 58, + 102, + 552, + 140 + ], + "spans": [ + { + "bbox": [ + 58, + 102, + 552, + 140 + ], + "type": "text", + "content": "DocSAM: Unified Document Image Segmentation via Query Decomposition and Heterogeneous Mixed Learning" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 88, + 160, + 523, + 232 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 88, + 160, + 523, + 232 + ], + "spans": [ + { + "bbox": [ + 88, + 160, + 523, + 232 + ], + "type": "text", + "content": "Xiao-Hui Li" + }, + { + "bbox": [ + 88, + 160, + 523, + 232 + ], + "type": "inline_equation", + "content": "^{1}" + }, + { + "bbox": [ + 88, + 160, + 523, + 232 + ], + "type": "text", + "content": ", Fei Yin" + }, + { + "bbox": [ + 88, + 160, + 523, + 232 + ], + "type": "inline_equation", + "content": "^{1}" + }, + { + "bbox": [ + 88, + 160, + 523, + 232 + ], + "type": "text", + "content": ", Cheng-Lin Liu" + }, + { + "bbox": [ + 88, + 160, + 523, + 232 + ], + "type": "inline_equation", + "content": "^{1,2}" + }, + { + "bbox": [ + 88, + 160, + 523, + 232 + ], + "type": "inline_equation", + "content": "^{1}" + }, + { + "bbox": [ + 88, + 160, + 523, + 232 + ], + "type": "text", + "content": "MAIS, Institute of Automation of Chinese Academy of Sciences, Beijing, 100190, China \n" + }, + { + "bbox": [ + 88, + 160, + 523, + 232 + ], + "type": "inline_equation", + "content": "^{2}" + }, + { + "bbox": [ + 88, + 160, + 523, + 232 + ], + "type": "text", + "content": "School of Artificial Intelligence, University of Chinese Academy of Sciences, Beijing, 100049, China \n{xiaohui.li, fyin, liucl}@nlpr.ia.ac.cn" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 151, + 258, + 200, + 272 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 151, + 258, + 200, + 272 + ], + "spans": [ + { + "bbox": [ + 151, + 258, + 200, + 272 + ], + "type": "text", + "content": "Abstract" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 56, + 285, + 296, + 608 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 285, + 296, + 608 + ], + "spans": [ + { + "bbox": [ + 56, + 285, + 296, + 608 + ], + "type": "text", + "content": "Document image segmentation is crucial for document analysis and recognition but remains challenging due to the diversity of document formats and segmentation tasks. Existing methods often address these tasks separately, resulting in limited generalization and resource wastage. This paper introduces DocSAM, a transformer-based unified framework designed for various document image segmentation tasks, such as document layout analysis, multi-granularity text segmentation, and table structure recognition, by modelling these tasks as a combination of instance and semantic segmentation. Specifically, DocSAM employs Sentence-BERT to map category names from each dataset into semantic queries that match the dimensionality of instance queries. These two sets of queries interact through an attention mechanism and are cross-attended with image features to predict instance and semantic segmentation masks. Instance categories are predicted by computing the dot product between instance and semantic queries, followed by softmax normalization of scores. Consequently, DocSAM can be jointly trained on heterogeneous datasets, enhancing robustness and generalization while reducing computational and storage resources. Comprehensive evaluations show that DocSAM surpasses existing methods in accuracy, efficiency, and adaptability, highlighting its potential for advancing document image understanding and segmentation across various applications. Codes are available at https://github.com/xhli-git/DocSAM." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 56, + 632, + 135, + 645 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 632, + 135, + 645 + ], + "spans": [ + { + "bbox": [ + 56, + 632, + 135, + 645 + ], + "type": "text", + "content": "1. Introduction" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 55, + 653, + 296, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 653, + 296, + 715 + ], + "spans": [ + { + "bbox": [ + 55, + 653, + 296, + 715 + ], + "type": "text", + "content": "Document image segmentation (DIS) is a fundamental task in the field of document analysis and recognition (DAR) [42], serving as a cornerstone for downstream applications such as text recognition, information extraction (IE), and document visual question answering (DocVQA). Despite" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 313, + 259, + 555, + 380 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 259, + 555, + 380 + ], + "spans": [ + { + "bbox": [ + 313, + 259, + 555, + 380 + ], + "type": "text", + "content": "its importance, DIS faces significant challenges due to the wide diversity of document types, page layouts, content annotations, and structural complexities, see Fig. 1. Existing approaches often address specific aspects of DIS separately, such as layout analysis, text detection, and table structure recognition, leading to specialized and fragmented solutions tailored to particular applications. This fragmentation not only impedes the performance of individual tasks but also results in redundant computational and storage overheads, making them inefficient for large-scale deployment." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 313, + 384, + 556, + 601 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 384, + 556, + 601 + ], + "spans": [ + { + "bbox": [ + 313, + 384, + 556, + 601 + ], + "type": "text", + "content": "To address the aforementioned challenges, this paper introduces DocSAM (Document Segment Anything Model), a transformer-based unified framework designed to simultaneously handle various document image segmentation tasks, thereby eliminating the need for separate models and enhancing overall efficiency. As illustrated in Fig. 2, DocSAM comprises four primary modules: the Vision Backbone, the Deformable Encoder, Sentence-BERT [64], and the Hybrid Query Decoder (HQD). Given a document image and desired instance or semantic class names in natural text format, DocSAM first extracts multi-scale image features using the Vision Backbone. These features are then refined by the Deformable Encoder, which includes several deformable attention layers [108]. Class names are fed into Sentence-BERT and mapped to semantic queries. Subsequently, both semantic queries and learnable instance queries pass together through the HQD, where they interact to jointly perform semantic and instance segmentation." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 313, + 605, + 556, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 605, + 556, + 715 + ], + "spans": [ + { + "bbox": [ + 313, + 605, + 556, + 715 + ], + "type": "text", + "content": "Inside each HQD layer (see Fig. 2), semantic and instance queries are concatenated and passed through a multi-head self-attention layer followed by a feed-forward layer for information exchange. These queries are then separately cross-attended with multi-scale image features in a coarse-to-fine manner using two multi-scale decoders, each with " + }, + { + "bbox": [ + 313, + 605, + 556, + 715 + ], + "type": "inline_equation", + "content": "L = 4" + }, + { + "bbox": [ + 313, + 605, + 556, + 715 + ], + "type": "text", + "content": " layers. They further interact via another multi-head self-attention and feed-forward layer. The resulting semantic and instance queries, along with fused multi-scale image" + } + ] + } + ], + "index": 10 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "spans": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "type": "text", + "content": "CVF" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 145, + 0, + 496, + 37 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 145, + 0, + 496, + 37 + ], + "spans": [ + { + "bbox": [ + 145, + 0, + 496, + 37 + ], + "type": "text", + "content": "This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "text", + "content": "15021" + } + ] + } + ], + "index": 11 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 0 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 59, + 69, + 145, + 284 + ], + "blocks": [ + { + "bbox": [ + 59, + 69, + 145, + 284 + ], + "lines": [ + { + "bbox": [ + 59, + 69, + 145, + 284 + ], + "spans": [ + { + "bbox": [ + 59, + 69, + 145, + 284 + ], + "type": "image", + "image_path": "3a6a1f1783c34c1b3b6f4d53e438720afacc5ce3c93f7798170108fa8066c848.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 145, + 292, + 463, + 304 + ], + "lines": [ + { + "bbox": [ + 145, + 292, + 463, + 304 + ], + "spans": [ + { + "bbox": [ + 145, + 292, + 463, + 304 + ], + "type": "text", + "content": "Figure 1. Examples of various segmentation tasks on heterogeneous document datasets." + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "type": "image", + "bbox": [ + 145, + 69, + 310, + 284 + ], + "blocks": [ + { + "bbox": [ + 145, + 69, + 310, + 284 + ], + "lines": [ + { + "bbox": [ + 145, + 69, + 310, + 284 + ], + "spans": [ + { + "bbox": [ + 145, + 69, + 310, + 284 + ], + "type": "image", + "image_path": "8ac8b9cb48c302516b0d45a95ad8117108a69b452aef87003aaf121d116dcf48.jpg" + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_body" + } + ], + "index": 1 + }, + { + "type": "image", + "bbox": [ + 312, + 70, + 460, + 284 + ], + "blocks": [ + { + "bbox": [ + 312, + 70, + 460, + 284 + ], + "lines": [ + { + "bbox": [ + 312, + 70, + 460, + 284 + ], + "spans": [ + { + "bbox": [ + 312, + 70, + 460, + 284 + ], + "type": "image", + "image_path": "b734259ed50c26b5a4d03e1d2dcdd8b88de0700ea8d3217e81c277437206cf3c.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + } + ], + "index": 2 + }, + { + "type": "image", + "bbox": [ + 462, + 70, + 552, + 284 + ], + "blocks": [ + { + "bbox": [ + 462, + 70, + 552, + 284 + ], + "lines": [ + { + "bbox": [ + 462, + 70, + 552, + 284 + ], + "spans": [ + { + "bbox": [ + 462, + 70, + 552, + 284 + ], + "type": "image", + "image_path": "dbde6a541b34d3b73dd2631149976a9949eebc85db692c1e342ced8a6e53ba3c.jpg" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_body" + } + ], + "index": 3 + }, + { + "bbox": [ + 55, + 324, + 295, + 384 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 324, + 295, + 384 + ], + "spans": [ + { + "bbox": [ + 55, + 324, + 295, + 384 + ], + "type": "text", + "content": "features, are forwarded to the Mask Predictor, Class Predictor, and BBox Predictor for semantic mask segmentation, instance mask segmentation, category classification, and bounding box regression, respectively. We stack " + }, + { + "bbox": [ + 55, + 324, + 295, + 384 + ], + "type": "inline_equation", + "content": "K" + }, + { + "bbox": [ + 55, + 324, + 295, + 384 + ], + "type": "text", + "content": " HQD layers for more refined predictions." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 55, + 389, + 295, + 534 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 389, + 295, + 534 + ], + "spans": [ + { + "bbox": [ + 55, + 389, + 295, + 534 + ], + "type": "text", + "content": "This design ensures that DocSAM can effectively manage the heterogeneity of document types, annotation formats, and segmentation tasks while maintaining high efficiency and accuracy. Extensive experiments and evaluations on various datasets demonstrate that DocSAM surpasses existing methods in accuracy, efficiency, and adaptability. Our results highlight DocSAM's potential as a powerful tool for advancing document image segmentation and understanding, with applications spanning from modern and historical document layout analysis to table structure decomposition, handwritten and scene text detection, and beyond. Our contributions are summarized as follows:" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 55, + 545, + 295, + 714 + ], + "type": "list", + "angle": 0, + "index": 11, + "blocks": [ + { + "bbox": [ + 55, + 545, + 295, + 605 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 545, + 295, + 605 + ], + "spans": [ + { + "bbox": [ + 55, + 545, + 295, + 605 + ], + "type": "text", + "content": "- We introduce DocSAM, a unified solution for diverse document image segmentation tasks such as layout analysis, multi-grained text segmentation, and table structure decomposition, reducing the need for specialized models and enhancing overall efficiency;" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 55, + 605, + 295, + 642 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 605, + 295, + 642 + ], + "spans": [ + { + "bbox": [ + 55, + 605, + 295, + 642 + ], + "type": "text", + "content": "- By training on various tasks and datasets, DocSAM improves robustness and generalization, making it highly effective in handling varied document types and structures;" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 55, + 642, + 295, + 677 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 642, + 295, + 677 + ], + "spans": [ + { + "bbox": [ + 55, + 642, + 295, + 677 + ], + "type": "text", + "content": "- Compared to specialized models, DocSAM significantly reduces computational and storage requirements, making it more practical for large-scale deployment;" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 55, + 677, + 295, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 677, + 295, + 714 + ], + "spans": [ + { + "bbox": [ + 55, + 677, + 295, + 714 + ], + "type": "text", + "content": "- Extensive experiments on various datasets show that Doc-SAM outperforms current methods in terms of accuracy, efficiency, and adaptability." + } + ] + } + ], + "index": 10 + } + ], + "sub_type": "text" + }, + { + "bbox": [ + 314, + 323, + 405, + 335 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 314, + 323, + 405, + 335 + ], + "spans": [ + { + "bbox": [ + 314, + 323, + 405, + 335 + ], + "type": "text", + "content": "2. Related Works" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 314, + 349, + 446, + 361 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 314, + 349, + 446, + 361 + ], + "spans": [ + { + "bbox": [ + 314, + 349, + 446, + 361 + ], + "type": "text", + "content": "2.1. DIS Tasks and Datasets" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 313, + 373, + 555, + 564 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 373, + 555, + 564 + ], + "spans": [ + { + "bbox": [ + 313, + 373, + 555, + 564 + ], + "type": "text", + "content": "Depending on specific application scenarios, DIS involves various sub-tasks including Document Layout Analysis (DLA), Multi-Granularity Text Detection (MGTD), and Table Structure Recognition (TSR). DLA aims at identifying and categorizing page regions including text blocks, figures and tables [9, 16, 36, 60, 106]. This foundational step provides a structured overview of the document's layout, enabling more precise processing in subsequent tasks. MGTD focuses on detecting and segmenting text at various granularities, from paragraphs down to individual lines and words [11, 43, 53, 93, 95]. MGTD is a prerequisite for accurate Optical Character Recognition (OCR) tasks. TSR specifically aims to extract the structural of tables, including rows, columns and cells [17, 35, 38, 72, 107]. By decomposing tables into substructures, TSR facilitates the extraction and analysis of tabular information from documents." + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 313, + 570, + 556, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 570, + 556, + 715 + ], + "spans": [ + { + "bbox": [ + 313, + 570, + 556, + 715 + ], + "type": "text", + "content": "Along with these tasks, plenty of datasets have been accumulated after decades of research, see Tab. 1. These datasets exhibit great diversity and heterogeneity in data sources, document types, annotation formats, writing languages, category sets and many other aspects. For example, PubLayNet [106] contains born-digital English PDF documents with region-level annotations; SCUT-CAB [8] and MTHv2 [53] contains scanned historical Chinese documents with region, line and char-level annotations; SCUT-HCCDoc [95] contains handwritten documents with line-level annotations; CTW1500 [43] and Total-Text [11] contain natural scene images with texts of arbitrary shapes." + } + ] + } + ], + "index": 15 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "text", + "content": "15022" + } + ] + } + ], + "index": 16 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 1 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 63, + 70, + 288, + 262 + ], + "blocks": [ + { + "bbox": [ + 63, + 70, + 288, + 262 + ], + "lines": [ + { + "bbox": [ + 63, + 70, + 288, + 262 + ], + "spans": [ + { + "bbox": [ + 63, + 70, + 288, + 262 + ], + "type": "table", + "html": "
TaskDataset
Document Layout AnalysisBaDLAD [71], CDLA [3], D4LA [13], DocBank [36], DocLayNet [60], ICDAR2017-POD [16], IIIT-AR-13K [55], M6Doc [9], PubLayNet [106], RanLayNet [2]
Ancient and Handwritten Document SegmentationCASIA-AHCDB [85], CHDAC-2022 [31], ICDAR2019-HDRC [68], SCUT-CAB [8], MTHv2 [53], HJDataset [69], CASIA-HWDB [41], SCUT-HCCDoc [95]
Table Structure RecognitionFinTabNet [104], ICDAR2013 [18], ICDAR2017-POD [16, 38], ICDAR2019-cTDr [17, 38], NTable [109], PubTables-1M [72], PubTabNet [107], STDW [19], TableBank [35], TNCR [1], WTW [47]
Scene Text DetectionCASIA-10k [22], COCO-Text [76], CTW1500 [43], CTW-Public [92], HUST-TR400 [87], ICDAR2015 [28], ICDAR2017-RCTW [70], ICDAR2017-MLT [56], ICDAR2019-ArT [12], ICDAR2019-LSVT [74], ICDAR2019-MLT [57], ICDAR2019-ReCTS [100], ICDAR2023-HierText [49], ICDAR2023-ReST [91], ICPR2018-MTWI [21], MSRA-TD500 [86], ShopSign [94], Total-Text [11], USTB-SV1K [90]
", + "image_path": "3cfa877871442b1ba2bc5bc2d4be92208279056d90ad3a349d8b32a36f83816a.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "table_body" + } + ], + "index": 0 + }, + { + "bbox": [ + 102, + 270, + 247, + 280 + ], + "lines": [ + { + "bbox": [ + 102, + 270, + 247, + 280 + ], + "spans": [ + { + "bbox": [ + 102, + 270, + 247, + 280 + ], + "type": "text", + "content": "Table 1. Datasets involved in DocSAM." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "text" + }, + { + "bbox": [ + 55, + 307, + 186, + 320 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 307, + 186, + 320 + ], + "spans": [ + { + "bbox": [ + 55, + 307, + 186, + 320 + ], + "type": "text", + "content": "2.2. Deep Learning for DIS" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 55, + 327, + 296, + 555 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 327, + 296, + 555 + ], + "spans": [ + { + "bbox": [ + 55, + 327, + 296, + 555 + ], + "type": "text", + "content": "Existing deep learning based DIS methods basically focus on specific sub-tasks and datasets. Generally speaking, they usually transform various DIS tasks into general object detection or image segmentation problems and make some modifications to general object detection [20, 62, 63, 65] and image segmentation methods [5, 46, 66] to make them more suitable for the tasks and datasets at hand. Some other works treat documents as hierarchical graph structures and adopt graph models like GNN [82] and CRF [32] for the task of layout analysis [37, 52], table structure recognition [10, 38], and text detection [48, 101]. Though more flexible, these methods usually suffer from complicated pre/post-processing steps and are more susceptible to intermediate errors. There are also some multi-modal based methods that combine visual and textual features like LayoutLMv3 [27], DiT [34], and VGT [13]. These methods improve the performance and generalization by pre-training on large-scale unsupervised documents to align text and visual features, but are often slower due to the complexity of architectures." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 55, + 558, + 296, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 558, + 296, + 713 + ], + "spans": [ + { + "bbox": [ + 55, + 558, + 296, + 713 + ], + "type": "text", + "content": "With the prosperity of large language models (LLMs) [102], many large document models are proposed such as UDOP [75], UniDoc [15], DocPedia [14], DocLLM [80], TextMonkey [44], mPLUG-DocOwl [23, 88], etc. Though promising results can be achieved for the DocVQA task, lacking fine-grained intermediate outputs like text locations and page layouts still greatly limits the interpretability and generalization of these models. As compensation, recently some LLM-free unified models are proposed for low-level document processing tasks such as UPOCR [59], DocRes [98], OmniParser [78] and DAT [79]. These works unify several similar or related tasks into unified models through multi-task learning, but at the cost of significant increment" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 313, + 72, + 555, + 97 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 72, + 555, + 97 + ], + "spans": [ + { + "bbox": [ + 313, + 72, + 555, + 97 + ], + "type": "text", + "content": "of model complexity and calculating overhead, prohibiting them from generalizing to to more tasks and datasets." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 313, + 106, + 547, + 118 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 106, + 547, + 118 + ], + "spans": [ + { + "bbox": [ + 313, + 106, + 547, + 118 + ], + "type": "text", + "content": "2.3. Transformer-based Detection&Segmentation" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 313, + 123, + 555, + 243 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 123, + 555, + 243 + ], + "spans": [ + { + "bbox": [ + 313, + 123, + 555, + 243 + ], + "type": "text", + "content": "Following the pioneer work of DETR [4], many Transformer-based objection methods have been proposed in recent years, including Deformable DETR [108], DNDETR [33], DINO [96], Sparse R-CNN [73], etc. These methods share the same idea with DETR that rely on learnable queries and bipartite matching for object decoding, but make different modifications to improve the accuracy and convergence speed, such as bringing in deformable attention and denoising training, or assigning specific spatial meanings to the queries." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 313, + 244, + 556, + 472 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 244, + 556, + 472 + ], + "spans": [ + { + "bbox": [ + 313, + 244, + 556, + 472 + ], + "type": "text", + "content": "Besides object detection, Transformer also shows great potential in image segmentation [6, 7, 24, 25, 30, 83, 97, 103]. Among them the most related works to this paper are SAM [30] and Mask2former [7]. Inspired by SAM [30] which uses natural language prompts to guide image segmentation, in this paper we propose to embed the class names of each dataset into semantic queries and transform various document segmentation tasks into a combination of instance segmentation and semantic segmentation. The semantic queries not only serve as prompts guiding the model in identifying specific types of regions, but also function as class prototypes that instance queries depend on for classification. Since DIS relies on high resolution image features, we build our DocSAM from Mask2former [7] which adopt Swin-Transformer [45] and deformable attention [108] as the vision encoder. Though the vision encoder of DocSAM is inherited from Mask2former, the decoder is drastically redesigned to be up to the task of general document image segmentation effectively." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 313, + 484, + 376, + 496 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 484, + 376, + 496 + ], + "spans": [ + { + "bbox": [ + 313, + 484, + 376, + 496 + ], + "type": "text", + "content": "3. DocSAM" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 313, + 504, + 400, + 516 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 504, + 400, + 516 + ], + "spans": [ + { + "bbox": [ + 313, + 504, + 400, + 516 + ], + "type": "text", + "content": "3.1. Preliminaries" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 313, + 521, + 556, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 521, + 556, + 715 + ], + "spans": [ + { + "bbox": [ + 313, + 521, + 556, + 715 + ], + "type": "text", + "content": "Before introducing the proposed DocSAM, we first explore the key attributes of an ideal all-in-one DIS model and why current methods fall short of this goal. We assert that an exemplary all-in-one DIS model should possess the following attributes: it should have the versatility to convert diverse DIS tasks into a unified framework; it should be adaptable to training on heterogeneous datasets, accommodating diverse annotations without restrictions; and it should maintain the capacity for continual and incremental learning. Current methods are typically designed for specific DIS tasks and datasets, and most models become static after training, unable to efficiently incorporate new data, limiting their versatility and adaptability. Specifically, existing DIS methods rely on fully connected (FC) and Softmax layers to predict region classes, with FC's parameters predefined for specific tasks and datasets, making generalization difficult." + } + ] + } + ], + "index": 11 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "text", + "content": "15023" + } + ] + } + ], + "index": 12 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 2 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 69, + 69, + 541, + 264 + ], + "blocks": [ + { + "bbox": [ + 69, + 69, + 541, + 264 + ], + "lines": [ + { + "bbox": [ + 69, + 69, + 541, + 264 + ], + "spans": [ + { + "bbox": [ + 69, + 69, + 541, + 264 + ], + "type": "image", + "image_path": "860576fad4d4e862ce1333467fde1ff2d896d31e117085aa79f416ed0e08ac58.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 55, + 272, + 555, + 296 + ], + "lines": [ + { + "bbox": [ + 55, + 272, + 555, + 296 + ], + "spans": [ + { + "bbox": [ + 55, + 272, + 555, + 296 + ], + "type": "text", + "content": "Figure 2. Network structure of the proposed DocSAM. DocSAM unify various document image segmentation tasks into one single model through instance and semantic query decomposition and interaction. Skip connections and norm layers are omitted for simplicity." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "bbox": [ + 55, + 316, + 296, + 447 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 316, + 296, + 447 + ], + "spans": [ + { + "bbox": [ + 55, + 316, + 296, + 447 + ], + "type": "text", + "content": "To overcome the above limitations and achieve the aforementioned criteria, the proposed DocSAM makes two significant improvements compared to existing methods. First, it transforms various DIS tasks into a unified paradigm of mask-based instance segmentation and semantic segmentation. Second, it embeds class names into semantic queries, which not only serve as prompts to guide the model in identifying specific types of regions to segment but also function as class prototypes that instance queries depend on for classification. The rest of this chapter presents the details of our proposed DocSAM." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 55, + 455, + 149, + 467 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 455, + 149, + 467 + ], + "spans": [ + { + "bbox": [ + 55, + 455, + 149, + 467 + ], + "type": "text", + "content": "3.2. Vision Encoder" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 55, + 472, + 296, + 652 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 472, + 296, + 652 + ], + "spans": [ + { + "bbox": [ + 55, + 472, + 296, + 652 + ], + "type": "text", + "content": "Different DIS tasks may focus on contents of different scales, from large objects like paragraphs and figures spanning entire pages to tiny objects like chars and words covering only a few hundreds of pixels. Therefore, high-resolution multi-scale image features are an essential requirement for a unified DIS model. The vision encoder of DocSAM is adapted from Mask2Former [7], which includes a Swin-Transformer [45] as the vision backbone and deformable attention [108] for feature refinement. Additionally, we use another FPN [40] to fuse multi-scale image features " + }, + { + "bbox": [ + 55, + 472, + 296, + 652 + ], + "type": "inline_equation", + "content": "X_{I} = [X_{I}^{l}\\in \\mathbb{R}^{H_{l}W_{l}\\times C},l\\in \\{1,2,3,4\\}]" + }, + { + "bbox": [ + 55, + 472, + 296, + 652 + ], + "type": "text", + "content": " into a single mask feature " + }, + { + "bbox": [ + 55, + 472, + 296, + 652 + ], + "type": "inline_equation", + "content": "X_{M}\\in \\mathbb{R}^{HW\\times C}" + }, + { + "bbox": [ + 55, + 472, + 296, + 652 + ], + "type": "text", + "content": ", which is used for subsequent semantic segmentation and instance segmentation. Here, " + }, + { + "bbox": [ + 55, + 472, + 296, + 652 + ], + "type": "inline_equation", + "content": "X_{I}^{l}" + }, + { + "bbox": [ + 55, + 472, + 296, + 652 + ], + "type": "text", + "content": " is image feature of level " + }, + { + "bbox": [ + 55, + 472, + 296, + 652 + ], + "type": "inline_equation", + "content": "l" + }, + { + "bbox": [ + 55, + 472, + 296, + 652 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 55, + 472, + 296, + 652 + ], + "type": "inline_equation", + "content": "H_{l}" + }, + { + "bbox": [ + 55, + 472, + 296, + 652 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 55, + 472, + 296, + 652 + ], + "type": "inline_equation", + "content": "W_{l}" + }, + { + "bbox": [ + 55, + 472, + 296, + 652 + ], + "type": "text", + "content": " are the spatial resolution of level " + }, + { + "bbox": [ + 55, + 472, + 296, + 652 + ], + "type": "inline_equation", + "content": "l" + }, + { + "bbox": [ + 55, + 472, + 296, + 652 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 55, + 472, + 296, + 652 + ], + "type": "inline_equation", + "content": "C" + }, + { + "bbox": [ + 55, + 472, + 296, + 652 + ], + "type": "text", + "content": " is number of feature channels." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 55, + 659, + 165, + 673 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 659, + 165, + 673 + ], + "spans": [ + { + "bbox": [ + 55, + 659, + 165, + 673 + ], + "type": "text", + "content": "3.3. Query Embedding" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 55, + 677, + 296, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 677, + 296, + 713 + ], + "spans": [ + { + "bbox": [ + 55, + 677, + 296, + 713 + ], + "type": "text", + "content": "The instance queries " + }, + { + "bbox": [ + 55, + 677, + 296, + 713 + ], + "type": "inline_equation", + "content": "Q_{I} \\in \\mathbb{R}^{N \\times C}" + }, + { + "bbox": [ + 55, + 677, + 296, + 713 + ], + "type": "text", + "content": " of DocSAM is standard learnable queries, while the semantic queries " + }, + { + "bbox": [ + 55, + 677, + 296, + 713 + ], + "type": "inline_equation", + "content": "Q_{S} \\in \\mathbb{R}^{M \\times C}" + }, + { + "bbox": [ + 55, + 677, + 296, + 713 + ], + "type": "text", + "content": " are embedded from class names using the Sentence-BERT" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 313, + 316, + 555, + 399 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 316, + 555, + 399 + ], + "spans": [ + { + "bbox": [ + 313, + 316, + 555, + 399 + ], + "type": "text", + "content": "[64]. Here " + }, + { + "bbox": [ + 313, + 316, + 555, + 399 + ], + "type": "inline_equation", + "content": "N" + }, + { + "bbox": [ + 313, + 316, + 555, + 399 + ], + "type": "text", + "content": " is a predefined instance query number that remains the same across all tasks and datasets, while " + }, + { + "bbox": [ + 313, + 316, + 555, + 399 + ], + "type": "inline_equation", + "content": "M" + }, + { + "bbox": [ + 313, + 316, + 555, + 399 + ], + "type": "text", + "content": " is the semantic query number that may change depending on the class number of each dataset, and " + }, + { + "bbox": [ + 313, + 316, + 555, + 399 + ], + "type": "inline_equation", + "content": "C" + }, + { + "bbox": [ + 313, + 316, + 555, + 399 + ], + "type": "text", + "content": " is feature dimension. " + }, + { + "bbox": [ + 313, + 316, + 555, + 399 + ], + "type": "inline_equation", + "content": "Q_{I}" + }, + { + "bbox": [ + 313, + 316, + 555, + 399 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 313, + 316, + 555, + 399 + ], + "type": "inline_equation", + "content": "Q_{S}" + }, + { + "bbox": [ + 313, + 316, + 555, + 399 + ], + "type": "text", + "content": " go together through the following Hybrid Query Decoder for feature decoding and cooperate with each other for semantic segmentation and instance segmentation." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 313, + 407, + 444, + 420 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 407, + 444, + 420 + ], + "spans": [ + { + "bbox": [ + 313, + 407, + 444, + 420 + ], + "type": "text", + "content": "3.4. Hybrid Query Decoder" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 313, + 425, + 555, + 628 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 425, + 555, + 628 + ], + "spans": [ + { + "bbox": [ + 313, + 425, + 555, + 628 + ], + "type": "text", + "content": "Inside each HQD layer, see Fig. 2, we first concatenate " + }, + { + "bbox": [ + 313, + 425, + 555, + 628 + ], + "type": "inline_equation", + "content": "Q_{S}" + }, + { + "bbox": [ + 313, + 425, + 555, + 628 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 313, + 425, + 555, + 628 + ], + "type": "inline_equation", + "content": "Q_{I}" + }, + { + "bbox": [ + 313, + 425, + 555, + 628 + ], + "type": "text", + "content": " along the length dimension and send them into a multi-head self-attention layer (MHSA) followed by a feed forward layer (FFN). This step facilitates information exchange between " + }, + { + "bbox": [ + 313, + 425, + 555, + 628 + ], + "type": "inline_equation", + "content": "Q_{S}" + }, + { + "bbox": [ + 313, + 425, + 555, + 628 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 313, + 425, + 555, + 628 + ], + "type": "inline_equation", + "content": "Q_{I}" + }, + { + "bbox": [ + 313, + 425, + 555, + 628 + ], + "type": "text", + "content": ", allowing them to attend to each other for query fusion. Next, they are separately crossattended with the multi-scale image features " + }, + { + "bbox": [ + 313, + 425, + 555, + 628 + ], + "type": "inline_equation", + "content": "X_{I}" + }, + { + "bbox": [ + 313, + 425, + 555, + 628 + ], + "type": "text", + "content": " in a coarse-to-fine manner by two Multi-Scale Decoders (MSD) each containing " + }, + { + "bbox": [ + 313, + 425, + 555, + 628 + ], + "type": "inline_equation", + "content": "L" + }, + { + "bbox": [ + 313, + 425, + 555, + 628 + ], + "type": "text", + "content": " layers. Here, " + }, + { + "bbox": [ + 313, + 425, + 555, + 628 + ], + "type": "inline_equation", + "content": "L = 4" + }, + { + "bbox": [ + 313, + 425, + 555, + 628 + ], + "type": "text", + "content": " stands for the number of feature scales. Each MSD layer consists of two MHSA layers, one multi-head cross-attention layer (MHCA) and one FFN layer. Following Mask2Former [7], we also use masked attention in MHCA, where the attention masks are derived from the predicted instance and semantic masks in the previous HQD layer. After that, " + }, + { + "bbox": [ + 313, + 425, + 555, + 628 + ], + "type": "inline_equation", + "content": "Q_{S}" + }, + { + "bbox": [ + 313, + 425, + 555, + 628 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 313, + 425, + 555, + 628 + ], + "type": "inline_equation", + "content": "Q_{I}" + }, + { + "bbox": [ + 313, + 425, + 555, + 628 + ], + "type": "text", + "content": " further interact with each other through another MHSA and FFN layer. We stack " + }, + { + "bbox": [ + 313, + 425, + 555, + 628 + ], + "type": "inline_equation", + "content": "K" + }, + { + "bbox": [ + 313, + 425, + 555, + 628 + ], + "type": "text", + "content": " HQD layers for more refined predictions." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 313, + 635, + 413, + 647 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 635, + 413, + 647 + ], + "spans": [ + { + "bbox": [ + 313, + 635, + 413, + 647 + ], + "type": "text", + "content": "3.5. Prediction Head" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 313, + 653, + 554, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 653, + 554, + 713 + ], + "spans": [ + { + "bbox": [ + 313, + 653, + 554, + 713 + ], + "type": "text", + "content": "The output " + }, + { + "bbox": [ + 313, + 653, + 554, + 713 + ], + "type": "inline_equation", + "content": "Q_{S}" + }, + { + "bbox": [ + 313, + 653, + 554, + 713 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 313, + 653, + 554, + 713 + ], + "type": "inline_equation", + "content": "Q_{I}" + }, + { + "bbox": [ + 313, + 653, + 554, + 713 + ], + "type": "text", + "content": " from each HQD layer along with the mask feature " + }, + { + "bbox": [ + 313, + 653, + 554, + 713 + ], + "type": "inline_equation", + "content": "X_{M}" + }, + { + "bbox": [ + 313, + 653, + 554, + 713 + ], + "type": "text", + "content": " are sent to the Mask Predictor, Class Predictor and BBox Predictor for semantic mask segmentation, instance mask segmentation, instance category classification and instance bounding box regression, respectively." + } + ] + } + ], + "index": 11 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "text", + "content": "15024" + } + ] + } + ], + "index": 12 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 3 + }, + { + "para_blocks": [ + { + "bbox": [ + 55, + 72, + 295, + 96 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 72, + 295, + 96 + ], + "spans": [ + { + "bbox": [ + 55, + 72, + 295, + 96 + ], + "type": "text", + "content": "For predicting semantic and instance masks, " + }, + { + "bbox": [ + 55, + 72, + 295, + 96 + ], + "type": "inline_equation", + "content": "Q_{S}" + }, + { + "bbox": [ + 55, + 72, + 295, + 96 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 55, + 72, + 295, + 96 + ], + "type": "inline_equation", + "content": "Q_{I}" + }, + { + "bbox": [ + 55, + 72, + 295, + 96 + ], + "type": "text", + "content": " are multiplied with " + }, + { + "bbox": [ + 55, + 72, + 295, + 96 + ], + "type": "inline_equation", + "content": "X_{M}" + }, + { + "bbox": [ + 55, + 72, + 295, + 96 + ], + "type": "text", + "content": " as:" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 130, + 105, + 295, + 119 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 105, + 295, + 119 + ], + "spans": [ + { + "bbox": [ + 130, + 105, + 295, + 119 + ], + "type": "interline_equation", + "content": "M _ {S} = \\sigma \\left(Q _ {S} \\times X _ {M} ^ {T}\\right), \\tag {1}", + "image_path": "a69e0d8733a013c929462851e9621aa4535df7c72a72633f14fd1dcd1622edb8.jpg" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 55, + 129, + 74, + 139 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 129, + 74, + 139 + ], + "spans": [ + { + "bbox": [ + 55, + 129, + 74, + 139 + ], + "type": "text", + "content": "and" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 131, + 138, + 294, + 153 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 138, + 294, + 153 + ], + "spans": [ + { + "bbox": [ + 131, + 138, + 294, + 153 + ], + "type": "interline_equation", + "content": "M _ {I} = \\sigma \\left(Q _ {I} \\times X _ {M} ^ {T}\\right), \\tag {2}", + "image_path": "31367c7a83f4d93e60dc34cb435b7b176ea8dcc2ebb3b11134266351adbd03a0.jpg" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 55, + 156, + 296, + 218 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 156, + 296, + 218 + ], + "spans": [ + { + "bbox": [ + 55, + 156, + 296, + 218 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 55, + 156, + 296, + 218 + ], + "type": "inline_equation", + "content": "M_S \\in \\mathbb{R}^{M \\times HW}" + }, + { + "bbox": [ + 55, + 156, + 296, + 218 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 55, + 156, + 296, + 218 + ], + "type": "inline_equation", + "content": "M_I \\in \\mathbb{R}^{N \\times HW}" + }, + { + "bbox": [ + 55, + 156, + 296, + 218 + ], + "type": "text", + "content": " are predicted semantic and instance masks, " + }, + { + "bbox": [ + 55, + 156, + 296, + 218 + ], + "type": "inline_equation", + "content": "\\sigma" + }, + { + "bbox": [ + 55, + 156, + 296, + 218 + ], + "type": "text", + "content": " is Sigmoid function, " + }, + { + "bbox": [ + 55, + 156, + 296, + 218 + ], + "type": "inline_equation", + "content": "T" + }, + { + "bbox": [ + 55, + 156, + 296, + 218 + ], + "type": "text", + "content": " means matrix transposition, and " + }, + { + "bbox": [ + 55, + 156, + 296, + 218 + ], + "type": "inline_equation", + "content": "\\times" + }, + { + "bbox": [ + 55, + 156, + 296, + 218 + ], + "type": "text", + "content": " stands for matrix multiplication. Similarly, for predicting instance classes, " + }, + { + "bbox": [ + 55, + 156, + 296, + 218 + ], + "type": "inline_equation", + "content": "Q_I" + }, + { + "bbox": [ + 55, + 156, + 296, + 218 + ], + "type": "text", + "content": " is multiplied with " + }, + { + "bbox": [ + 55, + 156, + 296, + 218 + ], + "type": "inline_equation", + "content": "Q_S" + }, + { + "bbox": [ + 55, + 156, + 296, + 218 + ], + "type": "text", + "content": " as:" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 120, + 227, + 295, + 241 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 120, + 227, + 295, + 241 + ], + "spans": [ + { + "bbox": [ + 120, + 227, + 295, + 241 + ], + "type": "interline_equation", + "content": "Y _ {I} = \\operatorname {S o f t m a x} \\left(Q _ {I} \\times Q _ {S} ^ {T}\\right), \\tag {3}", + "image_path": "79b044411c645e8dd6d25abe2ae7416b45c1141ea1ed832af016f812ab1fa5de.jpg" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 55, + 249, + 295, + 297 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 249, + 295, + 297 + ], + "spans": [ + { + "bbox": [ + 55, + 249, + 295, + 297 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 55, + 249, + 295, + 297 + ], + "type": "inline_equation", + "content": "Y_{I} \\in \\mathbb{R}^{N \\times M}" + }, + { + "bbox": [ + 55, + 249, + 295, + 297 + ], + "type": "text", + "content": " is predicted class probabilities of instances, softmax is Softmax function along the second dimension of " + }, + { + "bbox": [ + 55, + 249, + 295, + 297 + ], + "type": "inline_equation", + "content": "Y_{I}" + }, + { + "bbox": [ + 55, + 249, + 295, + 297 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 55, + 249, + 295, + 297 + ], + "type": "inline_equation", + "content": "T" + }, + { + "bbox": [ + 55, + 249, + 295, + 297 + ], + "type": "text", + "content": " means matrix transposition, and " + }, + { + "bbox": [ + 55, + 249, + 295, + 297 + ], + "type": "inline_equation", + "content": "\\times" + }, + { + "bbox": [ + 55, + 249, + 295, + 297 + ], + "type": "text", + "content": " stands for matrix multiplication." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 55, + 298, + 296, + 392 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 298, + 296, + 392 + ], + "spans": [ + { + "bbox": [ + 55, + 298, + 296, + 392 + ], + "type": "text", + "content": "Since Eq. (1), Eq. (2), Eq. (3) are all based on matrix multiplication, they are actually calculating the similarities between " + }, + { + "bbox": [ + 55, + 298, + 296, + 392 + ], + "type": "inline_equation", + "content": "Q_{S}" + }, + { + "bbox": [ + 55, + 298, + 296, + 392 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 55, + 298, + 296, + 392 + ], + "type": "inline_equation", + "content": "Q_{I}" + }, + { + "bbox": [ + 55, + 298, + 296, + 392 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 55, + 298, + 296, + 392 + ], + "type": "inline_equation", + "content": "X_{M}" + }, + { + "bbox": [ + 55, + 298, + 296, + 392 + ], + "type": "text", + "content": ". So we can also regard the " + }, + { + "bbox": [ + 55, + 298, + 296, + 392 + ], + "type": "inline_equation", + "content": "Q_{S}" + }, + { + "bbox": [ + 55, + 298, + 296, + 392 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 55, + 298, + 296, + 392 + ], + "type": "inline_equation", + "content": "Q_{I}" + }, + { + "bbox": [ + 55, + 298, + 296, + 392 + ], + "type": "text", + "content": " as instance and semantic prototypes. Through the above semantic query embedding and prototype-based instance classification, we transform the original close-set classifier into an open-set classifier, thus benefiting the construction of unified all-in-one DIS model." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 55, + 393, + 296, + 465 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 393, + 296, + 465 + ], + "spans": [ + { + "bbox": [ + 55, + 393, + 296, + 465 + ], + "type": "text", + "content": "Besides mask segmentation, DocSAM also keep the ability of bbox prediction. This is realized through bounding box regression with the BBox Predictor. Following ISTR [24] and TransDLANet [9], for each HQD layer we predict the residual values of bbox coordinates relative to predictions of the previous HQD layer." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 55, + 472, + 154, + 485 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 472, + 154, + 485 + ], + "spans": [ + { + "bbox": [ + 55, + 472, + 154, + 485 + ], + "type": "text", + "content": "3.6. Model Learning" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 55, + 491, + 144, + 502 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 491, + 144, + 502 + ], + "spans": [ + { + "bbox": [ + 55, + 491, + 144, + 502 + ], + "type": "text", + "content": "3.6.1. Loss Function" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 55, + 505, + 296, + 553 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 505, + 296, + 553 + ], + "spans": [ + { + "bbox": [ + 55, + 505, + 296, + 553 + ], + "type": "text", + "content": "There are four losses in DocSAM, namely semantic mask segmentation loss " + }, + { + "bbox": [ + 55, + 505, + 296, + 553 + ], + "type": "inline_equation", + "content": "L_{S}" + }, + { + "bbox": [ + 55, + 505, + 296, + 553 + ], + "type": "text", + "content": ", instance mask segmentation loss " + }, + { + "bbox": [ + 55, + 505, + 296, + 553 + ], + "type": "inline_equation", + "content": "L_{I}" + }, + { + "bbox": [ + 55, + 505, + 296, + 553 + ], + "type": "text", + "content": ", instance bbox regression loss " + }, + { + "bbox": [ + 55, + 505, + 296, + 553 + ], + "type": "inline_equation", + "content": "L_{B}" + }, + { + "bbox": [ + 55, + 505, + 296, + 553 + ], + "type": "text", + "content": ", and instance classification loss " + }, + { + "bbox": [ + 55, + 505, + 296, + 553 + ], + "type": "inline_equation", + "content": "L_{C}" + }, + { + "bbox": [ + 55, + 505, + 296, + 553 + ], + "type": "text", + "content": ". Among them, " + }, + { + "bbox": [ + 55, + 505, + 296, + 553 + ], + "type": "inline_equation", + "content": "L_{S}" + }, + { + "bbox": [ + 55, + 505, + 296, + 553 + ], + "type": "text", + "content": " is calculated as:" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 70, + 562, + 295, + 577 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 70, + 562, + 295, + 577 + ], + "spans": [ + { + "bbox": [ + 70, + 562, + 295, + 577 + ], + "type": "interline_equation", + "content": "L _ {S} = \\lambda_ {f} L _ {f o c a l} (M _ {S}, \\hat {M} _ {S}) + \\lambda_ {d} L _ {d i c e} (M _ {S}, \\hat {M} _ {S}), \\quad (4)", + "image_path": "539449ba1542fad1008142a7d4df01ef03fa91962a10205bdb2a1c16ba962755.jpg" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 55, + 586, + 296, + 635 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 586, + 296, + 635 + ], + "spans": [ + { + "bbox": [ + 55, + 586, + 296, + 635 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 55, + 586, + 296, + 635 + ], + "type": "inline_equation", + "content": "M_S" + }, + { + "bbox": [ + 55, + 586, + 296, + 635 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 55, + 586, + 296, + 635 + ], + "type": "inline_equation", + "content": "\\hat{M}_S" + }, + { + "bbox": [ + 55, + 586, + 296, + 635 + ], + "type": "text", + "content": " are predicted and ground-truth semantic masks, " + }, + { + "bbox": [ + 55, + 586, + 296, + 635 + ], + "type": "inline_equation", + "content": "L_{focal}" + }, + { + "bbox": [ + 55, + 586, + 296, + 635 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 55, + 586, + 296, + 635 + ], + "type": "inline_equation", + "content": "L_{dice}" + }, + { + "bbox": [ + 55, + 586, + 296, + 635 + ], + "type": "text", + "content": " are focal loss [67] and dice loss [54], respectively, and " + }, + { + "bbox": [ + 55, + 586, + 296, + 635 + ], + "type": "inline_equation", + "content": "\\lambda_f = 10" + }, + { + "bbox": [ + 55, + 586, + 296, + 635 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 55, + 586, + 296, + 635 + ], + "type": "inline_equation", + "content": "\\lambda_d = 1" + }, + { + "bbox": [ + 55, + 586, + 296, + 635 + ], + "type": "text", + "content": " are hyperparameters. Similarly, " + }, + { + "bbox": [ + 55, + 586, + 296, + 635 + ], + "type": "inline_equation", + "content": "L_I" + }, + { + "bbox": [ + 55, + 586, + 296, + 635 + ], + "type": "text", + "content": " is calculated as:" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 73, + 643, + 295, + 658 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 73, + 643, + 295, + 658 + ], + "spans": [ + { + "bbox": [ + 73, + 643, + 295, + 658 + ], + "type": "interline_equation", + "content": "L _ {I} = \\lambda_ {f} L _ {f o c a l} (M _ {I}, \\hat {M} _ {I}) + \\lambda_ {d} L _ {d i c e} (M _ {I}, \\hat {M} _ {I}), \\tag {5}", + "image_path": "852db7d24b9a081213b4f7fc573087a30d728c80d911e27bcac7a4d321eb43e8.jpg" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 55, + 667, + 295, + 691 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 667, + 295, + 691 + ], + "spans": [ + { + "bbox": [ + 55, + 667, + 295, + 691 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 55, + 667, + 295, + 691 + ], + "type": "inline_equation", + "content": "M_S" + }, + { + "bbox": [ + 55, + 667, + 295, + 691 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 55, + 667, + 295, + 691 + ], + "type": "inline_equation", + "content": "\\hat{M}_S" + }, + { + "bbox": [ + 55, + 667, + 295, + 691 + ], + "type": "text", + "content": " are predicted and ground-truth instance masks, respectively. " + }, + { + "bbox": [ + 55, + 667, + 295, + 691 + ], + "type": "inline_equation", + "content": "L_B" + }, + { + "bbox": [ + 55, + 667, + 295, + 691 + ], + "type": "text", + "content": " is calculated as:" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 70, + 700, + 295, + 715 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 70, + 700, + 295, + 715 + ], + "spans": [ + { + "bbox": [ + 70, + 700, + 295, + 715 + ], + "type": "interline_equation", + "content": "L _ {B} = \\lambda_ {s l 1} L _ {s l 1} (B _ {I}, \\hat {B} _ {I}) + \\lambda_ {d i o u} L _ {d i o u} (B _ {I}, \\hat {B} _ {I}), \\tag {6}", + "image_path": "765d454743df2c7d56a3e71537050b4485290dd7616b247da588e48b16511ad0.jpg" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 313, + 72, + 554, + 120 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 72, + 554, + 120 + ], + "spans": [ + { + "bbox": [ + 313, + 72, + 554, + 120 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 313, + 72, + 554, + 120 + ], + "type": "inline_equation", + "content": "B_{I}" + }, + { + "bbox": [ + 313, + 72, + 554, + 120 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 313, + 72, + 554, + 120 + ], + "type": "inline_equation", + "content": "\\hat{B}_I" + }, + { + "bbox": [ + 313, + 72, + 554, + 120 + ], + "type": "text", + "content": " are predicted and ground-truth bounding boxes, " + }, + { + "bbox": [ + 313, + 72, + 554, + 120 + ], + "type": "inline_equation", + "content": "L_{sl1}" + }, + { + "bbox": [ + 313, + 72, + 554, + 120 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 313, + 72, + 554, + 120 + ], + "type": "inline_equation", + "content": "L_{diou}" + }, + { + "bbox": [ + 313, + 72, + 554, + 120 + ], + "type": "text", + "content": " are smooth L1 loss and distance IoU loss [105], respectively, and " + }, + { + "bbox": [ + 313, + 72, + 554, + 120 + ], + "type": "inline_equation", + "content": "\\lambda_{sl1} = 1" + }, + { + "bbox": [ + 313, + 72, + 554, + 120 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 313, + 72, + 554, + 120 + ], + "type": "inline_equation", + "content": "\\lambda_{diou} = 1" + }, + { + "bbox": [ + 313, + 72, + 554, + 120 + ], + "type": "text", + "content": " are hyper-parameters. At last, " + }, + { + "bbox": [ + 313, + 72, + 554, + 120 + ], + "type": "inline_equation", + "content": "L_{C}" + }, + { + "bbox": [ + 313, + 72, + 554, + 120 + ], + "type": "text", + "content": " is calculated as:" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 394, + 129, + 554, + 143 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 394, + 129, + 554, + 143 + ], + "spans": [ + { + "bbox": [ + 394, + 129, + 554, + 143 + ], + "type": "interline_equation", + "content": "L _ {C} = L _ {c e} \\left(Y _ {I}, \\hat {Y} _ {I}\\right), \\tag {7}", + "image_path": "85d60a9ab1728dbe97be4576b2530ae3384c7624fc08478050fa32dc16553815.jpg" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 313, + 150, + 554, + 186 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 150, + 554, + 186 + ], + "spans": [ + { + "bbox": [ + 313, + 150, + 554, + 186 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 313, + 150, + 554, + 186 + ], + "type": "inline_equation", + "content": "Y_{I}" + }, + { + "bbox": [ + 313, + 150, + 554, + 186 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 313, + 150, + 554, + 186 + ], + "type": "inline_equation", + "content": "\\hat{Y}_I" + }, + { + "bbox": [ + 313, + 150, + 554, + 186 + ], + "type": "text", + "content": " are predicted and ground-truth instance labels, and " + }, + { + "bbox": [ + 313, + 150, + 554, + 186 + ], + "type": "inline_equation", + "content": "L_{ce}" + }, + { + "bbox": [ + 313, + 150, + 554, + 186 + ], + "type": "text", + "content": " is cross entropy loss. The total loss of Doc-SAM is the sum of the above four losses:" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 358, + 196, + 553, + 209 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 358, + 196, + 553, + 209 + ], + "spans": [ + { + "bbox": [ + 358, + 196, + 553, + 209 + ], + "type": "interline_equation", + "content": "L = \\lambda_ {s} L _ {S} + \\lambda_ {i} L _ {I} + \\lambda_ {b} L _ {B} + \\lambda_ {c} L _ {C}, \\tag {8}", + "image_path": "27dfcb405a848513eb4e119a876e05650ff15689b1bb88c148bc27c1852e3b74.jpg" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 313, + 217, + 555, + 314 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 217, + 555, + 314 + ], + "spans": [ + { + "bbox": [ + 313, + 217, + 555, + 314 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 313, + 217, + 555, + 314 + ], + "type": "inline_equation", + "content": "\\lambda_s = 5" + }, + { + "bbox": [ + 313, + 217, + 555, + 314 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 313, + 217, + 555, + 314 + ], + "type": "inline_equation", + "content": "\\lambda_i = 5" + }, + { + "bbox": [ + 313, + 217, + 555, + 314 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 313, + 217, + 555, + 314 + ], + "type": "inline_equation", + "content": "\\lambda_b = 1" + }, + { + "bbox": [ + 313, + 217, + 555, + 314 + ], + "type": "text", + "content": ", and " + }, + { + "bbox": [ + 313, + 217, + 555, + 314 + ], + "type": "inline_equation", + "content": "\\lambda_c = 1" + }, + { + "bbox": [ + 313, + 217, + 555, + 314 + ], + "type": "text", + "content": " are hyperparameters. We add auxiliary losses to every HQD layer and to query features before HQD. Following DETR [4] and Mask2Former [7], we also use bipartite matching to find the best matched instance predictions before calculating the loss. While for semantic predictions, there is no need to perform the bipartite matching, because the predictions and ground-truths are already one-to-one matched." + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 313, + 319, + 477, + 331 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 319, + 477, + 331 + ], + "spans": [ + { + "bbox": [ + 313, + 319, + 477, + 331 + ], + "type": "text", + "content": "3.6.2. Heterogeneous Mixed Learning" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 313, + 333, + 555, + 502 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 333, + 555, + 502 + ], + "spans": [ + { + "bbox": [ + 313, + 333, + 555, + 502 + ], + "type": "text", + "content": "Unlike existing methods, the novel design of DocSAM enables us to train a single model on heterogeneous mixed datasets. In this work, we collected nearly fifty DIS datasets of various document types and annotation formats, covering diverse DIS tasks from layout analysis and text detection to table structure recognition (see Tab. 1). We combined these datasets to construct a heterogeneous mixed dataset for training DocSAM. After training, the DocSAM model can be directly used as a versatile document segmenter or as a pre-trained model that can be seamlessly fine-tuned using task-specific datasets without any specialized modifications, such as adding or replacing a linear classification layer. This merit of DocSAM endows it with the potential for continual and incremental learning." + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 313, + 507, + 471, + 519 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 507, + 471, + 519 + ], + "spans": [ + { + "bbox": [ + 313, + 507, + 471, + 519 + ], + "type": "text", + "content": "3.6.3. Improving Training Efficiency" + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 313, + 521, + 555, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 521, + 555, + 713 + ], + "spans": [ + { + "bbox": [ + 313, + 521, + 555, + 713 + ], + "type": "text", + "content": "Directly training DocSAM on such heterogeneous datasets may suffer from slow convergence and long training time, so we propose several strategies to improve training efficiency. Firstly, we pre-train the vision encoder of DocSAM on all 48 datasets using SimMIM [84], hoping it can provide more robust visual features for document images. Secondly, we separate the training datasets into groups with each group containing datasets of similar tasks and styles, see Tab. 1, then we adopt curriculum learning (CL) [81] strategy to warm up the training process by gradually adding new group of datasets. Thirdly, we add an instance query selection (IQS) process at the front of each HQD layer. Motivation behind this is that bipartite matching only calculates losses between matched predictions and ground-truths, and the matched query indexes are mostly the same across HQD layers. For a certain document, large ratio of instance" + } + ] + } + ], + "index": 25 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "text", + "content": "15025" + } + ] + } + ], + "index": 26 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 4 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 82, + 69, + 304, + 194 + ], + "blocks": [ + { + "bbox": [ + 82, + 69, + 304, + 194 + ], + "lines": [ + { + "bbox": [ + 82, + 69, + 304, + 194 + ], + "spans": [ + { + "bbox": [ + 82, + 69, + 304, + 194 + ], + "type": "image", + "image_path": "9c0a89f71440e055410d888e55f44672bd3c8951e4b6666328905956878a4fbd.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 153, + 196, + 231, + 206 + ], + "lines": [ + { + "bbox": [ + 153, + 196, + 231, + 206 + ], + "spans": [ + { + "bbox": [ + 153, + 196, + 231, + 206 + ], + "type": "text", + "content": "(a) Training loss curves." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 189, + 216, + 418, + 227 + ], + "lines": [ + { + "bbox": [ + 189, + 216, + 418, + 227 + ], + "spans": [ + { + "bbox": [ + 189, + 216, + 418, + 227 + ], + "type": "text", + "content": "Figure 3. Loss curves and on-the-fly validation during training." + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "type": "image", + "bbox": [ + 305, + 71, + 531, + 194 + ], + "blocks": [ + { + "bbox": [ + 305, + 71, + 531, + 194 + ], + "lines": [ + { + "bbox": [ + 305, + 71, + 531, + 194 + ], + "spans": [ + { + "bbox": [ + 305, + 71, + 531, + 194 + ], + "type": "image", + "image_path": "36491aebc5bd586bb356f174b84f45010b1327fbffa1cbf40d98039795ff5cfb.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 377, + 196, + 457, + 205 + ], + "lines": [ + { + "bbox": [ + 377, + 196, + 457, + 205 + ], + "spans": [ + { + "bbox": [ + 377, + 196, + 457, + 205 + ], + "type": "text", + "content": "(b) On-the-fly validation." + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_caption" + } + ], + "index": 2 + }, + { + "type": "table", + "bbox": [ + 63, + 245, + 289, + 339 + ], + "blocks": [ + { + "bbox": [ + 63, + 245, + 289, + 339 + ], + "lines": [ + { + "bbox": [ + 63, + 245, + 289, + 339 + ], + "spans": [ + { + "bbox": [ + 63, + 245, + 289, + 339 + ], + "type": "table", + "html": "
Ablation SettingInstanceSemantic
mAP\\( mAP_b \\)mAFmIoU
DocSAMbase0.38040.35170.42560.6615
DocSAMbase + Curriculum0.38690.37040.43380.6622
DocSAMbase + SimMIM0.10020.06580.12940.3882
DocSAMbase + Freezing BERT0.38430.35100.42950.6677
DocSAMbase - Masked Attention0.23220.09380.28460.6327
DocSAMbase - Query Interaction0.29900.17790.35090.6511
DocSAMbase - Query Selection0.33410.31340.37630.6592
DocSAMlarge0.39000.36580.43200.6726
", + "image_path": "a16467fd4bf7371b4a003ab1bfdbaea232b3a90be56253feccd16d8e94cc9f8e.jpg" + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "table_body" + } + ], + "index": 5 + }, + { + "bbox": [ + 55, + 347, + 294, + 359 + ], + "lines": [ + { + "bbox": [ + 55, + 347, + 294, + 359 + ], + "spans": [ + { + "bbox": [ + 55, + 347, + 294, + 359 + ], + "type": "text", + "content": "Table 2. Ablation studies on model structure and training strategy." + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "text" + }, + { + "bbox": [ + 55, + 379, + 296, + 486 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 379, + 296, + 486 + ], + "spans": [ + { + "bbox": [ + 55, + 379, + 296, + 486 + ], + "type": "text", + "content": "queries are not activated from beginning to end, and their class scores are very low. Therefore, we only select instance queries whose class scores are higher than a threshold " + }, + { + "bbox": [ + 55, + 379, + 296, + 486 + ], + "type": "inline_equation", + "content": "T_{k}" + }, + { + "bbox": [ + 55, + 379, + 296, + 486 + ], + "type": "text", + "content": " before the " + }, + { + "bbox": [ + 55, + 379, + 296, + 486 + ], + "type": "inline_equation", + "content": "k" + }, + { + "bbox": [ + 55, + 379, + 296, + 486 + ], + "type": "text", + "content": "-th HQD layer. We set " + }, + { + "bbox": [ + 55, + 379, + 296, + 486 + ], + "type": "inline_equation", + "content": "T_{k}" + }, + { + "bbox": [ + 55, + 379, + 296, + 486 + ], + "type": "text", + "content": " as: " + }, + { + "bbox": [ + 55, + 379, + 296, + 486 + ], + "type": "inline_equation", + "content": "T_{k} = T_{max} / 2^{K - k}" + }, + { + "bbox": [ + 55, + 379, + 296, + 486 + ], + "type": "text", + "content": ", where " + }, + { + "bbox": [ + 55, + 379, + 296, + 486 + ], + "type": "inline_equation", + "content": "T_{max} = 0.01" + }, + { + "bbox": [ + 55, + 379, + 296, + 486 + ], + "type": "text", + "content": " is the maximum threshold, " + }, + { + "bbox": [ + 55, + 379, + 296, + 486 + ], + "type": "inline_equation", + "content": "K" + }, + { + "bbox": [ + 55, + 379, + 296, + 486 + ], + "type": "text", + "content": " is the number of HQD layers, " + }, + { + "bbox": [ + 55, + 379, + 296, + 486 + ], + "type": "inline_equation", + "content": "k" + }, + { + "bbox": [ + 55, + 379, + 296, + 486 + ], + "type": "text", + "content": " is the current HQD layer. Experiments show that IQS can discard low-score queries without degrading model performance, thereby improving training speed and reducing memory usage." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 55, + 497, + 135, + 510 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 497, + 135, + 510 + ], + "spans": [ + { + "bbox": [ + 55, + 497, + 135, + 510 + ], + "type": "text", + "content": "4. Experiments" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 55, + 517, + 176, + 528 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 517, + 176, + 528 + ], + "spans": [ + { + "bbox": [ + 55, + 517, + 176, + 528 + ], + "type": "text", + "content": "4.1. Datasets and Metrics" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 55, + 534, + 295, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 534, + 295, + 713 + ], + "spans": [ + { + "bbox": [ + 55, + 534, + 295, + 713 + ], + "type": "text", + "content": "The datasets involved in our experiments are listed in Tab. 1. Underlined datasets (15 in total) are used for ablation studies, mixed pre-training, and dataset-specific fine-tuning. All 48 datasets are used for training the final DocSAM model. These datasets cover a wide range of domains and tasks, showing significant heterogeneity in document types, annotation formats, and other aspects. Typical examples are shown in Fig. 1, with more details provided in the supplementary material. For evaluation metrics, we use mIoU [46] for semantic segmentation and mAP (for masks) [39] and " + }, + { + "bbox": [ + 55, + 534, + 295, + 713 + ], + "type": "inline_equation", + "content": "\\mathrm{mAP}_b" + }, + { + "bbox": [ + 55, + 534, + 295, + 713 + ], + "type": "text", + "content": " (for bounding boxes) [39] for instance segmentation. Additionally, we introduce a new metric for instance segmentation called mAF, which is calculated as the mean F-score of all classes across all IoUs ranging from 0.5 to 0.95 in increments of 0.05 (i.e., [0.5:0.05:0.95])." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 313, + 247, + 447, + 260 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 247, + 447, + 260 + ], + "spans": [ + { + "bbox": [ + 313, + 247, + 447, + 260 + ], + "type": "text", + "content": "4.2. Implementation Details" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 313, + 265, + 555, + 396 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 265, + 555, + 396 + ], + "spans": [ + { + "bbox": [ + 313, + 265, + 555, + 396 + ], + "type": "text", + "content": "The vision backbone and deformable attention module are initialized from Mask2Former [7], which is pre-trained on the COCO-panoptic dataset [39]. The Sentence-BERT is initialized using the all-MiniLM-L6-v2 model from the Sentence Transformers library [64]. Other parts of DocSAM are randomly initialized. We trained two sizes of models: DocSAM-base (207M parameters) and DocSAM-large (317M parameters). Their vision backbones use Swin-base and Swin-large, respectively, and the instance query numbers " + }, + { + "bbox": [ + 313, + 265, + 555, + 396 + ], + "type": "inline_equation", + "content": "N" + }, + { + "bbox": [ + 313, + 265, + 555, + 396 + ], + "type": "text", + "content": " are set to 500 and 900, respectively. The HQD layer number " + }, + { + "bbox": [ + 313, + 265, + 555, + 396 + ], + "type": "inline_equation", + "content": "K" + }, + { + "bbox": [ + 313, + 265, + 555, + 396 + ], + "type": "text", + "content": " is set to 4 by default." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 313, + 397, + 556, + 516 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 397, + 556, + 516 + ], + "spans": [ + { + "bbox": [ + 313, + 397, + 556, + 516 + ], + "type": "text", + "content": "DocSAM is implemented based on PyTorch [58] and trained on " + }, + { + "bbox": [ + 313, + 397, + 556, + 516 + ], + "type": "inline_equation", + "content": "8 \\times" + }, + { + "bbox": [ + 313, + 397, + 556, + 516 + ], + "type": "text", + "content": " NVIDIA A800 GPUs. We use the AdamW optimizer [50] to train the model, setting the base learning rate to " + }, + { + "bbox": [ + 313, + 397, + 556, + 516 + ], + "type": "inline_equation", + "content": "4 \\times 10^{-5}" + }, + { + "bbox": [ + 313, + 397, + 556, + 516 + ], + "type": "text", + "content": ", and decay it using cosine annealing strategy [51]. For joint training on mixed datasets, the default settings are 80,000 iterations and a batch size of 32; for ablation studies and dataset-specific fine-tuning, the defaults are 20,000 iterations and a batch size of 8; for comparison with state-of-the-art, the defaults are 40,000 iterations and a batch size of 16." + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 313, + 525, + 399, + 536 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 525, + 399, + 536 + ], + "spans": [ + { + "bbox": [ + 313, + 525, + 399, + 536 + ], + "type": "text", + "content": "4.3. Main Results" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 313, + 542, + 413, + 553 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 542, + 413, + 553 + ], + "spans": [ + { + "bbox": [ + 313, + 542, + 413, + 553 + ], + "type": "text", + "content": "4.3.1. Ablation Studies" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 313, + 557, + 556, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 557, + 556, + 713 + ], + "spans": [ + { + "bbox": [ + 313, + 557, + 556, + 713 + ], + "type": "text", + "content": "To verify the effect of each module in DocSAM and select the best training strategy before large-scale training, we conducted a series of ablation studies, as shown in Tab. 2 and Fig. 3. The results in Tab. 2 are averaged over all 15 datasets. On-the-fly validation involves fast testing on a small number of samples (e.g. 10 for each dataset) during training. The results show that using curriculum learning and instance query selection can accelerate convergence and improve model performance, while SimMIM pre-training significantly degrades model performance, possibly due to the large gap between SimMIM and document segmentation. Since freezing the weights of Sentence-BERT has almost no impact on performance, we freeze them during" + } + ] + } + ], + "index": 16 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "text", + "content": "15026" + } + ] + } + ], + "index": 17 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 5 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 79, + 67, + 526, + 256 + ], + "blocks": [ + { + "bbox": [ + 79, + 67, + 526, + 256 + ], + "lines": [ + { + "bbox": [ + 79, + 67, + 526, + 256 + ], + "spans": [ + { + "bbox": [ + 79, + 67, + 526, + 256 + ], + "type": "table", + "html": "
TaskDatasetDocSAM-baseDocSAM-large
InstanceSemanticInstanceSemantic
AP50AP75mAP\\( mAP_b \\)mAFmIoUAP50AP75mAP\\( mAP_b \\)mAFmIoU
Document Layout Analysis\\( D^4 \\)LA [13]0.5950.5140.4480.4380.4860.3890.6370.5620.4900.4730.5390.434
DocLayNet [60]0.7160.5280.4840.4800.5430.6070.7440.5700.5170.5010.5840.669
\\( M^6 \\)Doc [9]0.5190.4020.3630.3520.3810.2670.5510.4440.3970.3870.4250.296
PubLayNet [106]0.9360.8620.8060.7890.8470.8980.9460.8840.8300.8050.8680.911
Ancient and Handwritten Document SegmentationSCUT-CAB-logical [8]0.6810.5550.4810.4780.5020.4100.7170.5740.5110.4950.5340.454
SCUT-CAB-physical [8]0.9370.8370.7770.7470.8210.9370.9480.8560.7860.7540.8290.942
HJDataset [69]0.9560.9210.8810.8650.8950.8190.9560.9250.8850.8690.8980.821
CASIA-HWDB [41]0.9290.7850.7210.6640.7880.9350.9120.7700.7140.6430.7900.939
SCUT-HCCDoc [95]0.8650.6350.5440.5590.6250.8440.8690.6420.5490.5600.6250.847
Table Structure RecognitionFinTabNet [104]0.8670.7700.6640.6270.7570.8510.8690.7860.6840.6440.7780.860
PubTabNet [107]0.9700.7880.6430.6350.7140.8400.9700.7890.6480.6340.7230.845
TableBank-latex [35]0.9630.9470.8970.8680.9240.9400.9650.9500.9150.8930.9360.951
TableBank-word [35]0.8730.8370.8220.7930.8510.8440.8780.8440.8350.8140.8570.853
Scene Text DetectionCTW1500 [43]0.7120.4300.4000.3680.5000.7940.7530.4820.4410.4020.5310.817
Total-Text [11]0.7470.4210.4050.4070.7430.4530.7690.4540.4280.4210.4720.764
MSRA-TD500 [86]0.7470.5250.4580.4770.5020.7130.7980.5770.4960.5160.5320.739
ICDAR2015 [28]0.6130.2470.2940.3020.3380.5990.6390.2600.3070.3130.3450.623
", + "image_path": "bee58aaee92f1b39cf52a5e8f47be07f6484b2fcb1e0fe590769c1259bc575d5.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "table_body" + } + ], + "index": 0 + }, + { + "type": "table", + "bbox": [ + 82, + 289, + 529, + 500 + ], + "blocks": [ + { + "bbox": [ + 199, + 267, + 410, + 279 + ], + "lines": [ + { + "bbox": [ + 199, + 267, + 410, + 279 + ], + "spans": [ + { + "bbox": [ + 199, + 267, + 410, + 279 + ], + "type": "text", + "content": "Table 3. Performance of DocSAM after joint pre-training." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 82, + 289, + 529, + 500 + ], + "lines": [ + { + "bbox": [ + 82, + 289, + 529, + 500 + ], + "spans": [ + { + "bbox": [ + 82, + 289, + 529, + 500 + ], + "type": "table", + "html": "
TaskDatasetDocSAM-large from scratchDocSAM-large from pretrain
InstanceSemanticInstanceSemantic
AP50AP75mAP\\( mAP_b \\)mAFmIoUAP50AP75mAP\\( mAP_b \\)mAFmIoU
DocumentLayoutAnalysis\\( D^4LA[13] \\)0.3650.2590.2390.1940.2330.2050.6980.6370.5550.5460.5950.526
DocLayNet [60]0.5030.2920.2950.2600.3590.3650.8330.6910.6210.6010.6790.736
\\( M^6Doc[9] \\)0.2790.1730.1690.1450.1630.0870.6670.5660.5000.4850.5280.430
PubLayNet [106]0.8730.7590.6960.6220.7380.8410.9540.9040.8540.8500.8880.921
Ancient andHandwrittenDocument SegmentationSCUT-CAB-logical [8]0.3910.2330.2390.1360.2280.2260.7830.6310.5560.5300.5820.481
SCUT-CAB-physical [8]0.8010.6440.6050.4050.6640.9180.9460.8690.7990.7620.8420.945
HJDataset [69]0.8480.8350.7520.6060.7770.8120.9830.9480.9050.8950.9110.822
CASIA-HWDB [41]0.9080.7370.6650.6280.7370.9490.9770.9390.8930.7920.9160.956
SCUT-HCCDoc [95]0.8070.4920.4600.4230.5410.8530.9040.6840.5800.5890.6580.862
Table StructureRecognitionFinTabNet [104]0.3350.1640.1780.0040.2220.7700.8770.8050.7130.6810.8030.870
PubTabNet [107]0.0130.0100.0070.0420.0070.8100.9730.8210.6690.6530.7420.860
TableBank-latex [35]0.7620.6120.5650.0200.6410.9130.9680.9540.9260.9090.9470.958
TableBank-word [35]0.5940.4460.4350.0450.6190.8230.9080.8770.8710.8590.8810.873
Scene Text DetectionCTW1500 [43]0.4310.0980.1620.0710.2530.8000.7940.5390.4800.4530.5730.831
Total-Text [11]0.3130.0420.0960.0470.1680.7490.7940.5170.4600.4660.5020.775
MSRA-TD500 [86]0.5060.1890.2220.0760.2950.7310.8090.6040.5240.5410.5550.744
ICDAR2015 [28]0.2030.0280.0630.0230.1130.5970.6810.3160.3410.3530.3790.641
Unseen DatasetIIIT-AR-13K [55]0.5550.4300.4030.1850.4170.7390.8420.7020.6380.6210.6930.642
CHDAC-2022 [31]0.8860.6960.6040.5090.6490.9150.9390.8280.6870.6250.7270.918
", + "image_path": "b6f5e1b05a4cfa6934583f47f4d08beea9aeb1a53647a4ddcf1f5fd6fbf1b8e7.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "table_body" + } + ], + "index": 2 + }, + { + "bbox": [ + 181, + 508, + 427, + 519 + ], + "lines": [ + { + "bbox": [ + 181, + 508, + 427, + 519 + ], + "spans": [ + { + "bbox": [ + 181, + 508, + 427, + 519 + ], + "type": "text", + "content": "Table 4. Performance of DocSAM after dataset specific fine-tuning." + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "text" + }, + { + "bbox": [ + 52, + 538, + 293, + 645 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 52, + 538, + 293, + 645 + ], + "spans": [ + { + "bbox": [ + 52, + 538, + 293, + 645 + ], + "type": "text", + "content": "training. Similar to Mask2Former, masked attention plays a crucial role in DocSAM, and removing it leads to a significant performance drop. Additionally, without query interaction, DocSAM's performance also decreases substantially, highlighting the importance of information exchange between instance and semantic queries. Finally, training a unified model on heterogeneous datasets heavily relies on the model's capacity, and using a more powerful vision backbone can greatly enhance model performance." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 55, + 660, + 207, + 673 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 660, + 207, + 673 + ], + "spans": [ + { + "bbox": [ + 55, + 660, + 207, + 673 + ], + "type": "text", + "content": "4.3.2. Pre-training and Fine-tuning" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 55, + 677, + 294, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 677, + 294, + 713 + ], + "spans": [ + { + "bbox": [ + 55, + 677, + 294, + 713 + ], + "type": "text", + "content": "We train DocSAM on mixed heterogeneous datasets (15 datasets) to validate its performance as a unified document segmenter and a pre-trained model for dataset-specific fine" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 313, + 540, + 553, + 648 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 540, + 553, + 648 + ], + "spans": [ + { + "bbox": [ + 313, + 540, + 553, + 648 + ], + "type": "text", + "content": "tuning. The results are shown in Tab. 3 and Tab. 4. DocSAM achieves good semantic and instance segmentation performance on various datasets and tasks, though performance may vary across datasets due to differing levels of difficulty. As a single-modal model, DocSAM may underperform on datasets like " + }, + { + "bbox": [ + 313, + 540, + 553, + 648 + ], + "type": "inline_equation", + "content": "\\mathrm{D}^4\\mathrm{LA}" + }, + { + "bbox": [ + 313, + 540, + 553, + 648 + ], + "type": "text", + "content": " [13], DocLayNet [60], " + }, + { + "bbox": [ + 313, + 540, + 553, + 648 + ], + "type": "inline_equation", + "content": "\\mathrm{M}^6\\mathrm{Doc}" + }, + { + "bbox": [ + 313, + 540, + 553, + 648 + ], + "type": "text", + "content": " [9] and SCUT-CAB-logical [8], which require multi-modal information for fine-grained logical layout analysis." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 313, + 654, + 553, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 654, + 553, + 713 + ], + "spans": [ + { + "bbox": [ + 313, + 654, + 553, + 713 + ], + "type": "text", + "content": "After joint training, we fine-tune DocSAM-large on each specific dataset to further improve performance. As shown in Tab. 4, fine-tuning results are significantly higher than direct testing and training from scratch. We also test DocSAM on unseen datasets IIIT-AR-13K [55] and CHDAC" + } + ] + } + ], + "index": 8 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 749, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 749, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 749, + 317, + 757 + ], + "type": "text", + "content": "15027" + } + ] + } + ], + "index": 9 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 6 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 63, + 71, + 288, + 155 + ], + "blocks": [ + { + "bbox": [ + 63, + 71, + 288, + 155 + ], + "lines": [ + { + "bbox": [ + 63, + 71, + 288, + 155 + ], + "spans": [ + { + "bbox": [ + 63, + 71, + 288, + 155 + ], + "type": "table", + "html": "
MethodObjectInstance
mAPAP50AP75mAPAP50AP75
Faster R-CNN [65]0.4900.6780.5720.4780.6780.552
Mask R-CNN [20]0.4010.5840.4620.3970.5840.456
Deformable DETR [108]0.5720.7680.6340.5560.7650.611
ISTR [24]0.6270.8080.7080.6200.8070.702
TransDLANet [9]0.6450.8270.7270.6380.8260.719
DAT [79]0.712--0.657--
DocSAM0.6630.8400.7550.6610.8400.750
", + "image_path": "25557cf06dbe8b6a7b07e9ccaf211f4d2758fa721d2b7b8de6a6c3902ac0e804.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "table_body" + } + ], + "index": 0 + }, + { + "type": "table", + "bbox": [ + 63, + 186, + 288, + 342 + ], + "blocks": [ + { + "bbox": [ + 91, + 163, + 259, + 175 + ], + "lines": [ + { + "bbox": [ + 91, + 163, + 259, + 175 + ], + "spans": [ + { + "bbox": [ + 91, + 163, + 259, + 175 + ], + "type": "text", + "content": "Table 5. Performance comparison on " + }, + { + "bbox": [ + 91, + 163, + 259, + 175 + ], + "type": "inline_equation", + "content": "{\\mathrm{M}}^{6}" + }, + { + "bbox": [ + 91, + 163, + 259, + 175 + ], + "type": "text", + "content": " Doc." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 63, + 186, + 288, + 342 + ], + "lines": [ + { + "bbox": [ + 63, + 186, + 288, + 342 + ], + "spans": [ + { + "bbox": [ + 63, + 186, + 288, + 342 + ], + "type": "table", + "html": "
MethodObjectInstance
mAPAP50AP75mAPAP50AP75
Physical
Faster R-CNN [65]0.7750.9130.8610.7530.9100.834
Mask R-CNN [20]0.7910.9210.8770.7950.9170.872
SCNet [77]0.8130.9410.8900.8200.9410.891
Deformable DETR [108]0.7990.9230.8710.7790.9210.843
VSR [99]0.7870.9190.8600.7870.9190.852
DocSAM0.7740.9470.8600.8110.9480.891
Logical
Faster R-CNN [65]0.5490.7740.6130.5420.7730.606
Mask R-CNN [20]0.5510.7850.6190.5530.7770.631
SCNet [77]0.6020.8360.6730.6030.8360.680
Deformable DETR [108]0.6270.8520.7170.6200.8510.703
VSR [99]0.5570.7830.6160.5510.7820.611
DocSAM0.5480.7690.6320.5750.7790.667
", + "image_path": "0d75b51bd03a7ea210ea8217c48ce340d337836f1a484860eb1460d6f75af982.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "table_body" + } + ], + "index": 2 + }, + { + "type": "table", + "bbox": [ + 71, + 375, + 280, + 468 + ], + "blocks": [ + { + "bbox": [ + 83, + 351, + 266, + 361 + ], + "lines": [ + { + "bbox": [ + 83, + 351, + 266, + 361 + ], + "spans": [ + { + "bbox": [ + 83, + 351, + 266, + 361 + ], + "type": "text", + "content": "Table 6. Performance comparison on SCUT-CAB." + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 71, + 375, + 280, + 468 + ], + "lines": [ + { + "bbox": [ + 71, + 375, + 280, + 468 + ], + "spans": [ + { + "bbox": [ + 71, + 375, + 280, + 468 + ], + "type": "table", + "html": "
MethodCTW1500Total-Text
PRFPRF
HierText [49]0.8460.8740.8600.8550.9050.879
SIR [61]0.8740.8370.8550.9090.8560.882
DPText-DETR [89]0.9170.8620.8880.9180.8640.890
UNITS [29]-----0.898
ESTextSpotter [26]0.9150.8860.9000.9200.8810.900
DAT-DET [79]0.8930.8930.8930.9400.8820.910
DAT-SEG [79]0.9250.9090.9170.9500.8920.920
DocSAM0.8050.8810.8420.7210.8260.770
", + "image_path": "7ff06e1537ec0a1ab860d1a18ac2123c7ae905bf3e61125a639820568dc36a82.jpg" + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "table_body" + } + ], + "index": 4 + }, + { + "bbox": [ + 59, + 476, + 291, + 487 + ], + "lines": [ + { + "bbox": [ + 59, + 476, + 291, + 487 + ], + "spans": [ + { + "bbox": [ + 59, + 476, + 291, + 487 + ], + "type": "text", + "content": "Table 7. Performance comparison on CTW1500 and Total-Text." + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "text" + }, + { + "bbox": [ + 55, + 510, + 295, + 570 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 510, + 295, + 570 + ], + "spans": [ + { + "bbox": [ + 55, + 510, + 295, + 570 + ], + "type": "text", + "content": "2022 [31], where fine-tuning from the pre-trained model also yields substantial performance gains. This demonstrates that DocSAM's performance is not yet saturated and can benefit greatly from transfer learning on unseen datasets and tasks." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 55, + 578, + 233, + 590 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 578, + 233, + 590 + ], + "spans": [ + { + "bbox": [ + 55, + 578, + 233, + 590 + ], + "type": "text", + "content": "4.3.3. Comparison with State-of-the-Arts" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 55, + 594, + 295, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 594, + 295, + 714 + ], + "spans": [ + { + "bbox": [ + 55, + 594, + 295, + 714 + ], + "type": "text", + "content": "To compare with state-of-the-art methods, we further fine-tuned DocSAM on some datasets for additional training iterations. The results are shown in Tab. 5, Tab. 6, and Tab. 7. The best results are shown in bold, and the second-best results are underlined. DocSAM achieves superior or comparable performance with other methods. Note that we did not apply any specific training techniques or data augmentation, configurations for all datasets were kept consistent. We found that DocSAM exhibits much lower performance in logical layout analysis compared to physical analysis," + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 313, + 72, + 553, + 144 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 72, + 553, + 144 + ], + "spans": [ + { + "bbox": [ + 313, + 72, + 553, + 144 + ], + "type": "text", + "content": "which we attribute to its reliance only on single-modal features. Furthermore, DocSAM achieved relatively low performance on scene text detection datasets. This is likely because scene texts exhibit much greater diversity in shapes and backgrounds, requiring more carefully designed strategies to ensure model performance." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 313, + 152, + 387, + 163 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 152, + 387, + 163 + ], + "spans": [ + { + "bbox": [ + 313, + 152, + 387, + 163 + ], + "type": "text", + "content": "4.4. Discussion" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 313, + 169, + 555, + 373 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 169, + 555, + 373 + ], + "spans": [ + { + "bbox": [ + 313, + 169, + 555, + 373 + ], + "type": "text", + "content": "The goal of this paper is not to achieve state-of-the-art performance on specific dataset and task through meticulously designed model architectures or training strategies. Instead, we aim to design a simple and unified document segmentation model that can be applied to a wide variety of datasets and tasks. Additionally, the trained model should possess good scalability and the ability to continue learning. In this regard, DocSAM is quite successful. It exhibits decent performance on various datasets and tasks and shows great potential for downstream applications both as a versatile segmenter and a pre-trained model. However, experimental results also reveal some weaknesses and limitations of DocSAM, such as long training time and unsatisfactory performance on complex scenarios. We believe that DocSAM can greatly benefit from more sophisticated model design and better data augmentation and training strategies to further accelerate its convergence and improve its performance." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 313, + 383, + 388, + 396 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 383, + 388, + 396 + ], + "spans": [ + { + "bbox": [ + 313, + 383, + 388, + 396 + ], + "type": "text", + "content": "5. Conclusion" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 313, + 403, + 555, + 643 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 403, + 555, + 643 + ], + "spans": [ + { + "bbox": [ + 313, + 403, + 555, + 643 + ], + "type": "text", + "content": "In this paper, we propose DocSAM, a transformer-based unified framework for various document image segmentation tasks. DocSAM integrates layout analysis, multi-grained text segmentation, and table structure decomposition into a single model, reducing the need for specialized models and enhancing efficiency. Trained on heterogeneous datasets, DocSAM demonstrates robust and generalizable performance, effectively handling diverse document types and structures. This approach also reduces computational and storage requirements, making DocSAM suitable for practical deployment in resource-constrained environments. Extensive experiments show that DocSAM outperforms existing methods in terms of accuracy, efficiency, and adaptability. Overall, we believe that DocSAM represents a significant step forward for document image segmentation, and we look forward to its continued development and application in practical scenarios. In the future, we plan to extend DocSAM to a multi-modal version and explore better training strategies to further accelerate its convergence and improve its performance." + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 313, + 653, + 411, + 667 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 653, + 411, + 667 + ], + "spans": [ + { + "bbox": [ + 313, + 653, + 411, + 667 + ], + "type": "text", + "content": "Acknowledgement" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 313, + 673, + 553, + 696 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 673, + 553, + 696 + ], + "spans": [ + { + "bbox": [ + 313, + 673, + 553, + 696 + ], + "type": "text", + "content": "This work is supported by the National Natural Science Foundation of China(NSFC) Grant U23B2029." + } + ] + } + ], + "index": 15 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "text", + "content": "15028" + } + ] + } + ], + "index": 16 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 7 + }, + { + "para_blocks": [ + { + "bbox": [ + 56, + 71, + 115, + 83 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 71, + 115, + 83 + ], + "spans": [ + { + "bbox": [ + 56, + 71, + 115, + 83 + ], + "type": "text", + "content": "References" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 62, + 91, + 295, + 712 + ], + "type": "list", + "angle": 0, + "index": 15, + "blocks": [ + { + "bbox": [ + 66, + 91, + 295, + 124 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 66, + 91, + 295, + 124 + ], + "spans": [ + { + "bbox": [ + 66, + 91, + 295, + 124 + ], + "type": "text", + "content": "[1] Abdelrahman Abdallah, Alexander Berendeyev, Islam Nuradin, and Daniyar Nurseitov. Tncr: Table net detection and classification dataset. Neurocomputing, 473:79-97, 2022. 3" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 66, + 125, + 295, + 179 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 66, + 125, + 295, + 179 + ], + "spans": [ + { + "bbox": [ + 66, + 125, + 295, + 179 + ], + "type": "text", + "content": "[2] Avinash Anand, Raj Jaiswal, Mohit Gupta, Siddhesh S Bangar, Pijush Bhuyan, Naman Lal, Rajeev Singh, Ritika Jha, Rajiv Ratn Shah, and Shin'Ichi Satoh. Ranlaynet: A dataset for document layout detection used for domain adaptation and generalization. In ACM MM Asia, pages 1-6, 2023. 3" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 66, + 180, + 294, + 213 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 66, + 180, + 294, + 213 + ], + "spans": [ + { + "bbox": [ + 66, + 180, + 294, + 213 + ], + "type": "text", + "content": "[3] buptlihang. CDLA: A Benchmark Dataset for Cross-Domain Layout Analysis. https://github.com/buptliang/CDLA, 2023. Accessed: 2024-10-31. 3" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 66, + 214, + 294, + 257 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 66, + 214, + 294, + 257 + ], + "spans": [ + { + "bbox": [ + 66, + 214, + 294, + 257 + ], + "type": "text", + "content": "[4] Nicolas Carion, Francisco Massa, Gabriel Synnaeve, Nicolas Usunier, Alexander Kirillov, and Sylvain Gelly. End-to-end object detection with transformers. In ECCV, pages 213-229. Springer, 2020. 3, 5" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 66, + 258, + 294, + 312 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 66, + 258, + 294, + 312 + ], + "spans": [ + { + "bbox": [ + 66, + 258, + 294, + 312 + ], + "type": "text", + "content": "[5] Liang-Chieh Chen, George Papandreou, Iasonas Kokkinos, Vladimir Koltun, and Alan Garg. Deep convolutional nets, atrous convolution, and fully connected crfs. In IEEE Trans. Pattern Anal. Mach. Intell., pages 834-848, 2018. 3" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 66, + 314, + 294, + 346 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 66, + 314, + 294, + 346 + ], + "spans": [ + { + "bbox": [ + 66, + 314, + 294, + 346 + ], + "type": "text", + "content": "[6] Bowen Cheng, Alex Schwing, and Alexander Kirillov. Per-pixel classification is not all you need for semantic segmentation. NeurIPS, 34:17864-17875, 2021. 3" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 66, + 347, + 294, + 391 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 66, + 347, + 294, + 391 + ], + "spans": [ + { + "bbox": [ + 66, + 347, + 294, + 391 + ], + "type": "text", + "content": "[7] Bowen Cheng, Ishan Misra, Alexander G Schwing, Alexander Kirillov, and Rohit Girdhar. Masked-attention mask transformer for universal image segmentation. In CVPR, pages 1290–1299, 2022. 3, 4, 5, 6" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 66, + 392, + 294, + 435 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 66, + 392, + 294, + 435 + ], + "spans": [ + { + "bbox": [ + 66, + 392, + 294, + 435 + ], + "type": "text", + "content": "[8] Hiuyi Cheng, Cheng Jian, Sihang Wu, and Lianwen Jin. Scut-cab: a new benchmark dataset of ancient chinese books with complex layouts for document layout analysis. In ICFHR, pages 436-451. Springer, 2022. 2, 3, 7" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 66, + 437, + 294, + 501 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 66, + 437, + 294, + 501 + ], + "spans": [ + { + "bbox": [ + 66, + 437, + 294, + 501 + ], + "type": "text", + "content": "[9] Hiuyi Cheng, Peirong Zhang, Sihang Wu, Jiaxin Zhang, Qiyuan Zhu, Zecheng Xie, Jing Li, Kai Ding, and Lianwen Jin. M6doc: A large-scale multi-format, multi-type, multilayout, multi-language, multi-annotation category dataset for modern document layout analysis. In CVPR, pages 15138–15147, 2023. 2, 3, 5, 7, 8" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 62, + 502, + 294, + 535 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 502, + 294, + 535 + ], + "spans": [ + { + "bbox": [ + 62, + 502, + 294, + 535 + ], + "type": "text", + "content": "[10] Zewen Chi, Heyan Huang, Heng-Da Xu, Houjin Yu, Wanxuan Yin, and Xian-Ling Mao. Complicated table structure recognition. arXiv preprint arXiv:1908.04729, 2019. 3" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 62, + 536, + 294, + 568 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 536, + 294, + 568 + ], + "spans": [ + { + "bbox": [ + 62, + 536, + 294, + 568 + ], + "type": "text", + "content": "[11] Chee Kheng Ch'ng and Chee Seng Chan. Total-text: A comprehensive dataset for scene text detection and recognition. In ICDAR, pages 935-942. IEEE, 2017. 2, 3, 7" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 62, + 570, + 294, + 624 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 570, + 294, + 624 + ], + "spans": [ + { + "bbox": [ + 62, + 570, + 294, + 624 + ], + "type": "text", + "content": "[12] Chee Kheng Chng, Yuliang Liu, Yipeng Sun, Chun Chet Ng, Canjie Luo, Zihan Ni, ChuanMing Fang, Shuai tao Zhang, Junyu Han, Errui Ding, et al. Icdar2019 robust reading challenge on arbitrary-shaped text-rrc-art. In ICDAR, pages 1571-1576. IEEE, 2019. 3" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 62, + 625, + 294, + 658 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 625, + 294, + 658 + ], + "spans": [ + { + "bbox": [ + 62, + 625, + 294, + 658 + ], + "type": "text", + "content": "[13] Cheng Da, Chuwei Luo, Qi Zheng, and Cong Yao. Vision grid transformer for document layout analysis. In ICCV, pages 19462-19472, 2023. 3, 7" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 62, + 658, + 294, + 712 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 658, + 294, + 712 + ], + "spans": [ + { + "bbox": [ + 62, + 658, + 294, + 712 + ], + "type": "text", + "content": "[14] Hao Feng, Qi Liu, Hao Liu, Wengang Zhou, Houqiang Li, and Can Huang. Docpedia: Unleashing the power of large multimodal model in the frequency domain for versatile document understanding. arXiv preprint arXiv:2311.11810, 2023. 3" + } + ] + } + ], + "index": 14 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 320, + 73, + 553, + 712 + ], + "type": "list", + "angle": 0, + "index": 30, + "blocks": [ + { + "bbox": [ + 320, + 73, + 553, + 128 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 320, + 73, + 553, + 128 + ], + "spans": [ + { + "bbox": [ + 320, + 73, + 553, + 128 + ], + "type": "text", + "content": "[15] Hao Feng, Zijian Wang, Jingqun Tang, Jinghui Lu, Wengang Zhou, Houqiang Li, and Can Huang. Unidoc: A universal large multimodal model for simultaneous text detection, recognition, spotting and understanding. arXiv preprint arXiv:2308.11592, 2023. 3" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 320, + 129, + 553, + 162 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 320, + 129, + 553, + 162 + ], + "spans": [ + { + "bbox": [ + 320, + 129, + 553, + 162 + ], + "type": "text", + "content": "[16] Liangcai Gao, Xiaohan Yi, Zhuoren Jiang, Leipeng Hao, and Zhi Tang. Icdar2017 competition on page object detection. In ICDAR, pages 1417-1422. IEEE, 2017. 2, 3" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 320, + 163, + 553, + 207 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 320, + 163, + 553, + 207 + ], + "spans": [ + { + "bbox": [ + 320, + 163, + 553, + 207 + ], + "type": "text", + "content": "[17] Liangcai Gao, Yilun Huang, Herve Dejean, Jean-Luc Meunier, Qinqin Yan, Yu Fang, Florian Kleber, and Eva Lang. Icdar 2019 competition on table detection and recognition (ctdar). In ICDAR, pages 1510-1515. IEEE, 2019. 2, 3" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 320, + 209, + 553, + 240 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 320, + 209, + 553, + 240 + ], + "spans": [ + { + "bbox": [ + 320, + 209, + 553, + 240 + ], + "type": "text", + "content": "[18] Max Göbel, Tamir Hassan, Ermelinda Oro, and Giorgio Orsi. Icdar 2013 table competition. In ICDAR, pages 1449-1453. IEEE, 2013. 3" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 320, + 243, + 553, + 285 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 320, + 243, + 553, + 285 + ], + "spans": [ + { + "bbox": [ + 320, + 243, + 553, + 285 + ], + "type": "text", + "content": "[19] Mrinal Haloi, Shashank Shekhar, Nikhil Fande, Siddhant Swaroop Dash, et al. Table detection in the wild: A novel diverse table detection dataset and method. arXiv preprint arXiv:2209.09207, 2022.3" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 320, + 287, + 553, + 319 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 320, + 287, + 553, + 319 + ], + "spans": [ + { + "bbox": [ + 320, + 287, + 553, + 319 + ], + "type": "text", + "content": "[20] Kaiming He, Georgia Gkioxari, Piotr Dollar, and Ross Girshick. Mask r-cnn. In ICCV, pages 2961-2969, 2017. 3, 8" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 320, + 321, + 553, + 375 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 320, + 321, + 553, + 375 + ], + "spans": [ + { + "bbox": [ + 320, + 321, + 553, + 375 + ], + "type": "text", + "content": "[21] Mengchao He, Yuliang Liu, Zhibo Yang, Sheng Zhang, Canjie Luo, Feiyu Gao, Qi Zheng, Yongpan Wang, Xin Zhang, and Lianwen Jin. Icpr2018 contest on robust reading for multi-type web images. In ICPR, pages 7-12. IEEE, 2018. 3" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 320, + 377, + 553, + 410 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 320, + 377, + 553, + 410 + ], + "spans": [ + { + "bbox": [ + 320, + 377, + 553, + 410 + ], + "type": "text", + "content": "[22] Wenhao He, Xu-Yao Zhang, Fei Yin, and Cheng-Lin Liu. Multi-oriented and multi-lingual scene text detection with direct regression. ICIP, 27(11):5406-5419, 2018. 3" + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 320, + 412, + 553, + 465 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 320, + 412, + 553, + 465 + ], + "spans": [ + { + "bbox": [ + 320, + 412, + 553, + 465 + ], + "type": "text", + "content": "[23] Anwen Hu, Haiyang Xu, Jiabo Ye, Ming Yan, Liang Zhang, Bo Zhang, Chen Li, Ji Zhang, Qin Jin, Fei Huang, et al. mplug-docowl 1.5: Unified structure learning forOCR-free document understanding. arXiv preprint arXiv:2403.12895, 2024.3" + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 320, + 468, + 553, + 511 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 320, + 468, + 553, + 511 + ], + "spans": [ + { + "bbox": [ + 320, + 468, + 553, + 511 + ], + "type": "text", + "content": "[24] Jie Hu, Liujuan Cao, Yao Lu, ShengChuan Zhang, Yan Wang, Ke Li, Feiyue Huang, Ling Shao, and Rongrong Ji. Istr: End-to-end instance segmentation with transformers. arXiv preprint arXiv:2105.00637, 2021. 3, 5, 8" + } + ] + } + ], + "index": 25 + }, + { + "bbox": [ + 320, + 513, + 553, + 545 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 320, + 513, + 553, + 545 + ], + "spans": [ + { + "bbox": [ + 320, + 513, + 553, + 545 + ], + "type": "text", + "content": "[25] Jie Hu, Yao Lu, Shengchuan Zhang, and Liujuan Cao. Istr: Mask-embedding-based instance segmentation transformer. IEEE Trans. Image Process., 2024. 3" + } + ] + } + ], + "index": 26 + }, + { + "bbox": [ + 320, + 547, + 553, + 600 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 320, + 547, + 553, + 600 + ], + "spans": [ + { + "bbox": [ + 320, + 547, + 553, + 600 + ], + "type": "text", + "content": "[26] Mingxin Huang, Jiaxin Zhang, Dezhi Peng, Hao Lu, Can Huang, Yuliang Liu, Xiang Bai, and Lianwen Jin. Estextspotter: Towards better scene text spotting with explicit synergy in transformer. In ICCV, pages 19495-19505, 2023. 8" + } + ] + } + ], + "index": 27 + }, + { + "bbox": [ + 320, + 602, + 553, + 645 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 320, + 602, + 553, + 645 + ], + "spans": [ + { + "bbox": [ + 320, + 602, + 553, + 645 + ], + "type": "text", + "content": "[27] Yupan Huang, Tengchao Lv, Lei Cui, Yutong Lu, and Furu Wei. Layoutlmv3: Pre-training for document ai with unified text and image masking. In ACM MM, pages 4083-4091, 2022. 3" + } + ] + } + ], + "index": 28 + }, + { + "bbox": [ + 320, + 647, + 553, + 712 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 320, + 647, + 553, + 712 + ], + "spans": [ + { + "bbox": [ + 320, + 647, + 553, + 712 + ], + "type": "text", + "content": "[28] Dimosthenis Karatzas, Lluis Gomez-Bigorda, Anguelos Nicolaou, Suman Ghosh, Andrew Bagdanov, Masakazu Iwamura, Jiri Matas, Lukas Neumann, Vijay Ramaseshan Chandrasekhar, Shijian Lu, et al. Icdar 2015 competition on robust reading. In ICDAR, pages 1156-1160. IEEE, 2015. 3, 7" + } + ] + } + ], + "index": 29 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "text", + "content": "15029" + } + ] + } + ], + "index": 31 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 8 + }, + { + "para_blocks": [ + { + "bbox": [ + 61, + 73, + 294, + 713 + ], + "type": "list", + "angle": 0, + "index": 15, + "blocks": [ + { + "bbox": [ + 61, + 73, + 294, + 115 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 61, + 73, + 294, + 115 + ], + "spans": [ + { + "bbox": [ + 61, + 73, + 294, + 115 + ], + "type": "text", + "content": "[29] Taeho Kil, Seonghyeon Kim, Sukmin Seo, Yoonsik Kim, and Daehee Kim. Towards unified scene text spotting based on sequence generation. In CVPR, pages 15223-15232, 2023. 8" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 61, + 118, + 294, + 162 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 61, + 118, + 294, + 162 + ], + "spans": [ + { + "bbox": [ + 61, + 118, + 294, + 162 + ], + "type": "text", + "content": "[30] Alexander Kirillov, Eric Mintun, Nikhila Ravi, Hanzi Mao, Chloe Rolland, Laura Gustafson, Tete Xiao, Spencer Whitehead, Alexander C Berg, Wan-Yen Lo, et al. Segment anything. In ICCV, pages 4015-4026, 2023. 3" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 62, + 163, + 294, + 217 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 163, + 294, + 217 + ], + "spans": [ + { + "bbox": [ + 62, + 163, + 294, + 217 + ], + "type": "text", + "content": "[31] Pazhou Lab. IACC competition on Chinese Historical Document Analysis Challenge. https://iacc.pazhoulab-huangpu.com/contestdetail?id=6497f74cd97a2dae9dcaeff8 & award=1, 000, 000, 2022. Accessed: 2024-10-31. 3, 7, 8" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 62, + 219, + 294, + 262 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 219, + 294, + 262 + ], + "spans": [ + { + "bbox": [ + 62, + 219, + 294, + 262 + ], + "type": "text", + "content": "[32] John Lafferty, Andrew McCallum, Fernando Pereira, et al. Conditional random fields: Probabilistic models for segmenting and labeling sequence data. In ICML, page 3. Williamstown, MA, 2001. 3" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 62, + 263, + 294, + 306 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 263, + 294, + 306 + ], + "spans": [ + { + "bbox": [ + 62, + 263, + 294, + 306 + ], + "type": "text", + "content": "[33] Feng Li, Hao Zhang, Shilong Liu, Jian Guo, Lionel M Ni, and Lei Zhang. Dn-detr: Accelerate detr training by introducing query denoising. In CVPR, pages 13619-13627, 2022. 3" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 62, + 308, + 294, + 351 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 308, + 294, + 351 + ], + "spans": [ + { + "bbox": [ + 62, + 308, + 294, + 351 + ], + "type": "text", + "content": "[34] Junlong Li, Yiheng Xu, Tengchao Lv, Lei Cui, Cha Zhang, and Furu Wei. Dit: Self-supervised pre-training for document image transformer. In ACM MM, pages 3530-3539, 2022. 3" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 62, + 354, + 294, + 397 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 354, + 294, + 397 + ], + "spans": [ + { + "bbox": [ + 62, + 354, + 294, + 397 + ], + "type": "text", + "content": "[35] Minghao Li, Lei Cui, Shaohan Huang, Furu Wei, Ming Zhou, and Zhoujun Li. Tablebank: Table benchmark for image-based table detection and recognition. In LREC, pages 1918-1925, 2020. 2, 3, 7" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 62, + 399, + 294, + 441 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 399, + 294, + 441 + ], + "spans": [ + { + "bbox": [ + 62, + 399, + 294, + 441 + ], + "type": "text", + "content": "[36] Minghao Li, Yiheng Xu, Lei Cui, Shaohan Huang, Furu Wei, Zhoujun Li, and Ming Zhou. Docbank: A benchmark dataset for document layout analysis. arXiv preprint arXiv:2006.01038, 2020. 2, 3" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 62, + 444, + 294, + 476 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 444, + 294, + 476 + ], + "spans": [ + { + "bbox": [ + 62, + 444, + 294, + 476 + ], + "type": "text", + "content": "[37] Xiao-Hui Li, Fei Yin, and Cheng-Lin Liu. Page segmentation using convolutional neural network and graphical model. In DAS, pages 231-245. Springer, 2020. 3" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 62, + 478, + 294, + 520 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 478, + 294, + 520 + ], + "spans": [ + { + "bbox": [ + 62, + 478, + 294, + 520 + ], + "type": "text", + "content": "[38] Xiao-Hui Li, Fei Yin, He-Sen Dai, and Cheng-Lin Liu. Table structure recognition and form parsing by end-to-end object detection and relation parsing. Pattern Recognition, 132:108946, 2022. 2, 3" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 62, + 522, + 294, + 567 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 522, + 294, + 567 + ], + "spans": [ + { + "bbox": [ + 62, + 522, + 294, + 567 + ], + "type": "text", + "content": "[39] Tsung-Yi Lin, Michael Maire, Serge Belongie, James Hays, Pietro Perona, Deva Ramanan, Piotr Dólar, and C Lawrence Zitnick. Microsoft coco: Common objects in context. In ECCV, pages 740-755. Springer, 2014. 6" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 62, + 568, + 294, + 610 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 568, + 294, + 610 + ], + "spans": [ + { + "bbox": [ + 62, + 568, + 294, + 610 + ], + "type": "text", + "content": "[40] Tsung-Yi Lin, Piotr Dólar, Ross Girshick, Kaiming He, Bharath Hariharan, and Serge Belongie. Feature pyramid networks for object detection. In CVPR, pages 2117-2125, 2017. 4" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 62, + 613, + 294, + 645 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 613, + 294, + 645 + ], + "spans": [ + { + "bbox": [ + 62, + 613, + 294, + 645 + ], + "type": "text", + "content": "[41] Cheng-Lin Liu, Fei Yin, Da-Han Wang, and Qiu-Feng Wang. Casia online and offline chinese handwriting databases. In ICDAR, pages 37-41. IEEE, 2011. 3, 7" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 62, + 647, + 294, + 689 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 647, + 294, + 689 + ], + "spans": [ + { + "bbox": [ + 62, + 647, + 294, + 689 + ], + "type": "text", + "content": "[42] Cheng-Lin Liu, Lianwen Jin, Xiang Bai, Xiaohui Li, and Fei Yin. Frontiers of intelligent document analysis and recognition: review and prospects. Journal of Image and Graphics, 28(08):2223-2252, 2023. 1" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 62, + 691, + 294, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 691, + 294, + 713 + ], + "spans": [ + { + "bbox": [ + 62, + 691, + 294, + 713 + ], + "type": "text", + "content": "[43] Yuliang Liu, Lianwen Jin, Shuai tao Zhang, Canjie Luo, and Sheng Zhang. Curved scene text detection via transverse" + } + ] + } + ], + "index": 14 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 320, + 73, + 553, + 713 + ], + "type": "list", + "angle": 0, + "index": 31, + "blocks": [ + { + "bbox": [ + 339, + 73, + 553, + 95 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 339, + 73, + 553, + 95 + ], + "spans": [ + { + "bbox": [ + 339, + 73, + 553, + 95 + ], + "type": "text", + "content": "and longitudinal sequence connection. Pattern Recognition, 90:337-345, 2019. 2, 3, 7" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 320, + 96, + 553, + 140 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 320, + 96, + 553, + 140 + ], + "spans": [ + { + "bbox": [ + 320, + 96, + 553, + 140 + ], + "type": "text", + "content": "[44] Yuliang Liu, Biao Yang, Qiang Liu, Zhang Li, Zhiyin Ma, Shuo Zhang, and Xiang Bai. Textmonkey: AnOCR-free large multimodal model for understanding document. arXiv preprint arXiv:2403.04473, 2024. 3" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 320, + 141, + 553, + 195 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 320, + 141, + 553, + 195 + ], + "spans": [ + { + "bbox": [ + 320, + 141, + 553, + 195 + ], + "type": "text", + "content": "[45] Ze Liu, Yutong Lin, Yan Cao, Yue Hu, Yu Wei, Zhiqiang Zhang, Jiahui Lin, Han Wang, Cheng Lu, Changhu Wang, et al. Swin transformer: Hierarchical vision transformer using shifted windows. In ICCV, pages 12226-12235, 2021. 3, 4" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 320, + 197, + 553, + 230 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 320, + 197, + 553, + 230 + ], + "spans": [ + { + "bbox": [ + 320, + 197, + 553, + 230 + ], + "type": "text", + "content": "[46] Jonathan Long, Evan Shelhamer, and Trevor Darrell. Fully convolutional networks for semantic segmentation. In CVPR, pages 3431-3440, 2015. 3, 6" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 320, + 232, + 553, + 263 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 320, + 232, + 553, + 263 + ], + "spans": [ + { + "bbox": [ + 320, + 232, + 553, + 263 + ], + "type": "text", + "content": "[47] Rujiao Long, Wen Wang, Nan Xue, Feiyu Gao, Zhibo Yang, Yongpan Wang, and Gui-Song Xia. Parsing table structures in the wild. In ICCV, pages 944-952, 2021. 3" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 320, + 265, + 553, + 308 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 320, + 265, + 553, + 308 + ], + "spans": [ + { + "bbox": [ + 320, + 265, + 553, + 308 + ], + "type": "text", + "content": "[48] Shangbang Long, Jiaqiang Ruan, Wenjie Zhang, Xin He, Wenhao Wu, and Cong Yao. Textsnake: A flexible representation for detecting text of arbitrary shapes. In ECCV, pages 20-36, 2018. 3" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 320, + 310, + 553, + 354 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 320, + 310, + 553, + 354 + ], + "spans": [ + { + "bbox": [ + 320, + 310, + 553, + 354 + ], + "type": "text", + "content": "[49] Shangbang Long, Siyang Qin, Dmitry Panteleev, Alessandro Bissacco, Yasuhisa Fujii, and Michalis Raptis. Towards end-to-end unified scene text detection and layout analysis. In CVPR, pages 1049-1059, 2022. 3, 8" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 320, + 355, + 553, + 376 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 320, + 355, + 553, + 376 + ], + "spans": [ + { + "bbox": [ + 320, + 355, + 553, + 376 + ], + "type": "text", + "content": "[50] I Loshchilov. Decoupled weight decay regularization. arXiv preprint arXiv:1711.05101, 2017. 6" + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 320, + 378, + 553, + 410 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 320, + 378, + 553, + 410 + ], + "spans": [ + { + "bbox": [ + 320, + 378, + 553, + 410 + ], + "type": "text", + "content": "[51] Ilya Loshchilov and Frank Hutter. Sgdr: Stochastic gradient descent with warm restarts. arXiv preprint arXiv:1608.03983, 2016. 6" + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 320, + 412, + 553, + 456 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 320, + 412, + 553, + 456 + ], + "spans": [ + { + "bbox": [ + 320, + 412, + 553, + 456 + ], + "type": "text", + "content": "[52] Siwen Luo, Yihao Ding, Siqu Long, Josiah Poon, and Soyeon Caren Han. Doc-gcn: Heterogeneous graph convolutional networks for document layout analysis. arXiv preprint arXiv:2208.10970, 2022. 3" + } + ] + } + ], + "index": 25 + }, + { + "bbox": [ + 320, + 457, + 553, + 500 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 320, + 457, + 553, + 500 + ], + "spans": [ + { + "bbox": [ + 320, + 457, + 553, + 500 + ], + "type": "text", + "content": "[53] Weihong Ma, Hesuo Zhang, Lianwen Jin, Sihang Wu, Jiapeng Wang, and Yongpan Wang. Joint layout analysis, character detection and recognition for historical document digitization. In ICFHR, pages 31-36. IEEE, 2020. 2, 3" + } + ] + } + ], + "index": 26 + }, + { + "bbox": [ + 320, + 502, + 553, + 544 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 320, + 502, + 553, + 544 + ], + "spans": [ + { + "bbox": [ + 320, + 502, + 553, + 544 + ], + "type": "text", + "content": "[54] Fausto Miletari, Nassir Navab, and Seyed-Ahmad Ahmadi. V-net: Fully convolutional neural networks for volumetric medical image segmentation. In 3DV, pages 565-571. IEEE, 2016. 5" + } + ] + } + ], + "index": 27 + }, + { + "bbox": [ + 320, + 547, + 553, + 580 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 320, + 547, + 553, + 580 + ], + "spans": [ + { + "bbox": [ + 320, + 547, + 553, + 580 + ], + "type": "text", + "content": "[55] Ajoy Mondal, Peter Lipps, and CV Jawahar. Iit-ar-13k: A new dataset for graphical object detection in documents. In DAS, pages 216–230. Springer, 2020. 3, 7" + } + ] + } + ], + "index": 28 + }, + { + "bbox": [ + 320, + 582, + 553, + 645 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 320, + 582, + 553, + 645 + ], + "spans": [ + { + "bbox": [ + 320, + 582, + 553, + 645 + ], + "type": "text", + "content": "[56] Nibal Nayef, Fei Yin, Imen Bizid, Hyunsoo Choi, Yuan Feng, Dimosthenis Karatzas, Zhenbo Luo, Umapada Pal, Christophe Rigaud, Joseph Chazalon, et al. Icdar2017 robust reading challenge on multi-lingual scene text detection and script identification-rrc-plt. In ICDAR, pages 1454-1459. IEEE, 2017. 3" + } + ] + } + ], + "index": 29 + }, + { + "bbox": [ + 320, + 647, + 553, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 320, + 647, + 553, + 713 + ], + "spans": [ + { + "bbox": [ + 320, + 647, + 553, + 713 + ], + "type": "text", + "content": "[57] Nibal Nayef, Yash Patel, Michal Busta, Pinaki Nath Chowdhury, Dimosthenis Karatzas, Wafa Khlif, Jiri Matas, Umapada Pal, Jean-Christophe Burie, Cheng-lin Liu, et al. Icdar2019 robust reading challenge on multi-lingual scene text detection and recognition—rrc-mlt-2019. In ICDAR, pages 1582–1587. IEEE, 2019. 3" + } + ] + } + ], + "index": 30 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "type": "text", + "content": "15030" + } + ] + } + ], + "index": 32 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 9 + }, + { + "para_blocks": [ + { + "bbox": [ + 61, + 73, + 295, + 713 + ], + "type": "list", + "angle": 0, + "index": 15, + "blocks": [ + { + "bbox": [ + 61, + 73, + 294, + 127 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 61, + 73, + 294, + 127 + ], + "spans": [ + { + "bbox": [ + 61, + 73, + 294, + 127 + ], + "type": "text", + "content": "[58] Adam Paszke, Sam Gross, Francisco Massa, Adam Lerer, James Bradbury, Gregory Chanan, Trevor Killeen, Zeming Lin, Natalia Gimelshein, Luca Antiga, et al. Pytorch: An imperative style, high-performance deep learning library. NeurIPS, 32, 2019. 6" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 61, + 129, + 294, + 172 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 61, + 129, + 294, + 172 + ], + "spans": [ + { + "bbox": [ + 61, + 129, + 294, + 172 + ], + "type": "text", + "content": "[59] Dezhi Peng, Zhenhua Yang, Jiaxin Zhang, Chongyu Liu, Yongxin Shi, Kai Ding, Fengjun Guo, and Lianwen Jin. Upocr: Towards unified pixel-levelOCR interface. In ICML, 2023. 3" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 62, + 175, + 295, + 217 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 175, + 295, + 217 + ], + "spans": [ + { + "bbox": [ + 62, + 175, + 295, + 217 + ], + "type": "text", + "content": "[60] Birgit Pfitzmann, Christoph Auer, Michele Dolfi, Ahmed S. Nassar, and Peter Staar. Doclaynet: A large human-annotated dataset for document layout analysis. In ACM SIGKDD, pages 3743–3751. 2022. 2, 3, 7" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 62, + 220, + 294, + 272 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 220, + 294, + 272 + ], + "spans": [ + { + "bbox": [ + 62, + 220, + 294, + 272 + ], + "type": "text", + "content": "[61] Xugong Qin, Pengyuan Lyu, Chengquan Zhang, Yu Zhou, Kun Yao, Peng Zhang, Hailun Lin, and Weiping Wang. Towards robust real-time scene text detection: From semantic to instance representation learning. In ACM MM, pages 2025-2034, 2023. 8" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 62, + 274, + 294, + 306 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 274, + 294, + 306 + ], + "spans": [ + { + "bbox": [ + 62, + 274, + 294, + 306 + ], + "type": "text", + "content": "[62] Joseph Redmon and Ali Farhadi. You only look once: Unified, real-time object detection. In CVPR, pages 779-788, 2016. 3" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 62, + 309, + 294, + 330 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 309, + 294, + 330 + ], + "spans": [ + { + "bbox": [ + 62, + 309, + 294, + 330 + ], + "type": "text", + "content": "[63] Joseph Redmon and Ali Farhadi. Yolov3: An incremental improvement, 2018. 3" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 62, + 332, + 294, + 364 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 332, + 294, + 364 + ], + "spans": [ + { + "bbox": [ + 62, + 332, + 294, + 364 + ], + "type": "text", + "content": "[64] N Reimers. Sentence-bert: Sentence embeddings using siamese bert-networks. arXiv preprint arXiv:1908.10084, 2019. 1, 4, 6" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 62, + 366, + 294, + 408 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 366, + 294, + 408 + ], + "spans": [ + { + "bbox": [ + 62, + 366, + 294, + 408 + ], + "type": "text", + "content": "[65] Shaoqing Ren, Kaiming He, Ross Girshick, and Jian Sun. Faster r-cnn: Towards real-time object detection with region proposal networks. In NeurIPS, pages 91-99, 2015. 3, 8" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 62, + 411, + 294, + 443 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 411, + 294, + 443 + ], + "spans": [ + { + "bbox": [ + 62, + 411, + 294, + 443 + ], + "type": "text", + "content": "[66] Olaf Ronneberger, Philipp Fischer, and Thomas Brox. U-net: Convolutional networks for biomedical image segmentation. In MICCAI, pages 234-241. Springer, 2015. 3" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 62, + 445, + 294, + 467 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 445, + 294, + 467 + ], + "spans": [ + { + "bbox": [ + 62, + 445, + 294, + 467 + ], + "type": "text", + "content": "[67] T-YLPG Ross and GKHP Dólar. Focal loss for dense object detection. In CVPR, pages 2980-2988, 2017. 5" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 62, + 468, + 294, + 521 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 468, + 294, + 521 + ], + "spans": [ + { + "bbox": [ + 62, + 468, + 294, + 521 + ], + "type": "text", + "content": "[68] Rajkumar Saini, Derek Dobson, Jon Morrey, Marcus Liwicki, and Foteini Simistira Liwicki. Icdar 2019 historical document reading challenge on large structured chinese family records. In ICDAR, pages 1499-1504. IEEE, 2019. 3" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 62, + 524, + 294, + 556 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 524, + 294, + 556 + ], + "spans": [ + { + "bbox": [ + 62, + 524, + 294, + 556 + ], + "type": "text", + "content": "[69] Zejiang Shen, Kaixuan Zhang, and Melissa Dell. A large dataset of historical japanese documents with complex layouts. In CVPR Workshops, pages 548-549, 2020. 3, 7" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 62, + 558, + 294, + 610 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 558, + 294, + 610 + ], + "spans": [ + { + "bbox": [ + 62, + 558, + 294, + 610 + ], + "type": "text", + "content": "[70] Baoguang Shi, Cong Yao, Minghui Liao, Mingkun Yang, Pei Xu, Linyan Cui, Serge Belongie, Shijian Lu, and Xiang Bai. Icdar2017 competition on reading chinese text in the wild (rctw-17). In ICDAR, pages 1429–1434. IEEE, 2017. 3" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 62, + 614, + 294, + 689 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 614, + 294, + 689 + ], + "spans": [ + { + "bbox": [ + 62, + 614, + 294, + 689 + ], + "type": "text", + "content": "[71] Md Istiak Hossain Shihab, Md Rakibul Hasan, Mahfuzur Rahman Emon, Syed Mobassir Hossen, Md Nazmuddoha Ansary, Intesur Ahmed, Fazle Rabbi Rakib, Shahriar Elahi Dhruvo, Souhardya Saha Dip, Akib Hasan Pavel, et al. Badlad: A large multi-domain bengali document layout analysis dataset. In ICDAR, pages 326-341. Springer, 2023. 3" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 62, + 692, + 294, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 692, + 294, + 713 + ], + "spans": [ + { + "bbox": [ + 62, + 692, + 294, + 713 + ], + "type": "text", + "content": "[72] Brandon Smock, Rohith Pesala, and Robin Abraham. Pubtables-1m: Towards comprehensive table extraction" + } + ] + } + ], + "index": 14 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 320, + 73, + 553, + 713 + ], + "type": "list", + "angle": 0, + "index": 30, + "blocks": [ + { + "bbox": [ + 339, + 73, + 553, + 95 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 339, + 73, + 553, + 95 + ], + "spans": [ + { + "bbox": [ + 339, + 73, + 553, + 95 + ], + "type": "text", + "content": "from unstructured documents. In CVPR, pages 4634-4642, 2022. 2, 3" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 321, + 97, + 553, + 139 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 321, + 97, + 553, + 139 + ], + "spans": [ + { + "bbox": [ + 321, + 97, + 553, + 139 + ], + "type": "text", + "content": "[73] Xingyi Sun, Ziyi Zhou, Peizhao Wang, Junwei Dai, Lu Lu, Wanli Ouyang, and Ping Luo. Sparse r-cnn: End-to-end object detection with learnable proposals. In CVPR, pages 13164-13173, 2021. 3" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 320, + 141, + 553, + 196 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 320, + 141, + 553, + 196 + ], + "spans": [ + { + "bbox": [ + 320, + 141, + 553, + 196 + ], + "type": "text", + "content": "[74] Yipeng Sun, Zihan Ni, Chee-Kheng Chng, Yuliang Liu, Canjie Luo, Chun Chet Ng, Junyu Han, Errui Ding, Jingtuo Liu, Dimosthenis Karatzas, et al. Icdar 2019 competition on large-scale street view text with partial labeling-rrc-lsvt. In ICDAR, pages 1557-1562. IEEE, 2019. 3" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 320, + 198, + 553, + 251 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 320, + 198, + 553, + 251 + ], + "spans": [ + { + "bbox": [ + 320, + 198, + 553, + 251 + ], + "type": "text", + "content": "[75] Zineng Tang, Ziyi Yang, Guoxin Wang, Yuwei Fang, Yang Liu, Chenguang Zhu, Michael Zeng, Cha Zhang, and Mohit Bansal. Unifying vision, text, and layout for universal document processing. In CVPR, pages 19254-19264, 2023. 3" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 320, + 254, + 553, + 297 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 320, + 254, + 553, + 297 + ], + "spans": [ + { + "bbox": [ + 320, + 254, + 553, + 297 + ], + "type": "text", + "content": "[76] Andreas Veit, Tomas Matera, Lukas Neumann, Jiri Matas, and Serge Belongie. Coco-text: Dataset and benchmark for text detection and recognition in natural images. arXiv preprint arXiv:1601.07140, 2016. 3" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 320, + 299, + 553, + 331 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 320, + 299, + 553, + 331 + ], + "spans": [ + { + "bbox": [ + 320, + 299, + 553, + 331 + ], + "type": "text", + "content": "[77] Thang Vu, Haeyong Kang, and Chang D Yoo. Scnet: Training inference sample consistency for instance segmentation. In AAAI, pages 2701-2709, 2021. 8" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 320, + 333, + 553, + 387 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 320, + 333, + 553, + 387 + ], + "spans": [ + { + "bbox": [ + 320, + 333, + 553, + 387 + ], + "type": "text", + "content": "[78] Jianqiang Wan, Sibo Song, Wenwen Yu, Yuliang Liu, Wenqing Cheng, Fei Huang, Xiang Bai, Cong Yao, and Zhibo Yang. Omniparser: A unified framework for text spotting key information extraction and table recognition. In CVPR, pages 15641-15653, 2024. 3" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 320, + 388, + 553, + 442 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 320, + 388, + 553, + 442 + ], + "spans": [ + { + "bbox": [ + 320, + 388, + 553, + 442 + ], + "type": "text", + "content": "[79] Xingyu Wan, Chengquan Zhang, Pengyuan Lyu, Sen Fan, Zihan Ni, Kun Yao, Errui Ding, and Jingdong Wang. Towards unified multi-granularity text detection with interactive attention. arXiv preprint arXiv:2405.19765, 2024. 3, 8" + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 320, + 445, + 553, + 498 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 320, + 445, + 553, + 498 + ], + "spans": [ + { + "bbox": [ + 320, + 445, + 553, + 498 + ], + "type": "text", + "content": "[80] Dongsheng Wang, Natraj Raman, Mathieu Sibue, Zhiqiang Ma, Petr Babkin, Simerjot Kaur, Yulong Pei, Armineh Nourbakhsh, and Xiaomo Liu. Docllm: A layout-aware generative language model for multimodal document understanding. arXiv preprint arXiv:2401.00908, 2023. 3" + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 320, + 501, + 553, + 533 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 320, + 501, + 553, + 533 + ], + "spans": [ + { + "bbox": [ + 320, + 501, + 553, + 533 + ], + "type": "text", + "content": "[81] Xin Wang, Yudong Chen, and Wenwu Zhu. A survey on curriculum learning. IEEE Trans. Pattern Anal. Mach. Intell., 44(9):4555-4576, 2021. 5" + } + ] + } + ], + "index": 25 + }, + { + "bbox": [ + 320, + 535, + 553, + 578 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 320, + 535, + 553, + 578 + ], + "spans": [ + { + "bbox": [ + 320, + 535, + 553, + 578 + ], + "type": "text", + "content": "[82] Zonghan Wu, Shirui Pan, Fengwen Chen, Guodong Long, Chengqi Zhang, and S Yu Philip. A comprehensive survey on graph neural networks. IEEE Trans. Neural Netw. Learn. Syst., 32(1):4-24, 2020. 3" + } + ] + } + ], + "index": 26 + }, + { + "bbox": [ + 320, + 580, + 553, + 623 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 320, + 580, + 553, + 623 + ], + "spans": [ + { + "bbox": [ + 320, + 580, + 553, + 623 + ], + "type": "text", + "content": "[83] Enze Xie, Zhiwei Wang, Gang Yu, Xingyi Zhang, Yuheng Wu, Hongsheng Li, and Yonghong Tian. Segformer: Simple and efficient design for semantic segmentation with transformers. In ICCV, pages 14125-14134, 2021. 3" + } + ] + } + ], + "index": 27 + }, + { + "bbox": [ + 320, + 625, + 553, + 668 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 320, + 625, + 553, + 668 + ], + "spans": [ + { + "bbox": [ + 320, + 625, + 553, + 668 + ], + "type": "text", + "content": "[84] Zhenda Xie, Zheng Zhang, Yue Cao, Yutong Lin, Jianmin Bao, Zhuliang Yao, Qi Dai, and Han Hu. Simmim: A simple framework for masked image modeling. In CVPR, pages 9653-9663, 2022. 5" + } + ] + } + ], + "index": 28 + }, + { + "bbox": [ + 320, + 670, + 553, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 320, + 670, + 553, + 713 + ], + "spans": [ + { + "bbox": [ + 320, + 670, + 553, + 713 + ], + "type": "text", + "content": "[85] Yue Xu, Fei Yin, Da-Han Wang, Xu-Yao Zhang, Zhaoxiang Zhang, and Cheng-Lin Liu. Casia-ahcdb: A large-scale chinese ancient handwritten characters database. In ICDAR, pages 793–798. IEEE, 2019. 3" + } + ] + } + ], + "index": 29 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 749, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 749, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 749, + 317, + 757 + ], + "type": "text", + "content": "15031" + } + ] + } + ], + "index": 31 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 10 + }, + { + "para_blocks": [ + { + "bbox": [ + 58, + 73, + 295, + 713 + ], + "type": "list", + "angle": 0, + "index": 15, + "blocks": [ + { + "bbox": [ + 61, + 73, + 294, + 106 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 61, + 73, + 294, + 106 + ], + "spans": [ + { + "bbox": [ + 61, + 73, + 294, + 106 + ], + "type": "text", + "content": "[86] Cong Yao, Xiang Bai, Wenyu Liu, Yi Ma, and Zhuowen Tu. Detecting texts of arbitrary orientations in natural images. In CVPR, pages 1083-1090. IEEE, 2012. 3, 7" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 61, + 107, + 294, + 139 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 61, + 107, + 294, + 139 + ], + "spans": [ + { + "bbox": [ + 61, + 107, + 294, + 139 + ], + "type": "text", + "content": "[87] Cong Yao, Xiang Bai, and Wenyu Liu. A unified framework for multioriented text detection and recognition. ICIP, 23 (11):4737-4749, 2014. 3" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 62, + 141, + 295, + 195 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 141, + 295, + 195 + ], + "spans": [ + { + "bbox": [ + 62, + 141, + 295, + 195 + ], + "type": "text", + "content": "[88] Jiabo Ye, Anwen Hu, Haiyang Xu, Qinghao Ye, Ming Yan, Yuhao Dan, Chenlin Zhao, Guohai Xu, Chenliang Li, Junfeng Tian, et al. mplug-docowl: Modularized multimodal large language model for document understanding. arXiv preprint arXiv:2307.02499, 2023. 3" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 62, + 198, + 295, + 240 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 198, + 295, + 240 + ], + "spans": [ + { + "bbox": [ + 62, + 198, + 295, + 240 + ], + "type": "text", + "content": "[89] Maoyuan Ye, Jing Zhang, Shanshan Zhao, Juhua Liu, Bo Du, and Dacheng Tao. Dptext-detr: Towards better scene text detection with dynamic points in transformer. In AAAI, pages 3241-3249, 2023. 8" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 62, + 243, + 294, + 284 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 243, + 294, + 284 + ], + "spans": [ + { + "bbox": [ + 62, + 243, + 294, + 284 + ], + "type": "text", + "content": "[90] Xu-Cheng Yin, Xuwang Yin, Kaizhu Huang, and Hong-Wei Hao. Robust text detection in natural scene images. IEEE Trans. Pattern Anal. Mach. Intell., 36(5):970–983, 2013. 3" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 62, + 287, + 294, + 331 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 287, + 294, + 331 + ], + "spans": [ + { + "bbox": [ + 62, + 287, + 294, + 331 + ], + "type": "text", + "content": "[91] Wenwen Yu, Mingyu Liu, Mingrui Chen, Ning Lu, Yinlong Wen, Yuliang Liu, Dimosthenis Karatzas, and Xiang Bai. Icdar 2023 competition on reading the seal title. In ICDAR, pages 522-535. Springer, 2023. 3" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 62, + 332, + 294, + 364 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 332, + 294, + 364 + ], + "spans": [ + { + "bbox": [ + 62, + 332, + 294, + 364 + ], + "type": "text", + "content": "[92] Tai-Ling Yuan, Zhe Zhu, Kun Xu, Cheng-Jun Li, Tai-Jiang Mu, and Shi-Min Hu. A large chinese text dataset in the wild. J. Comput. Sci. Tech., 34:509-521, 2019. 3" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 62, + 366, + 294, + 419 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 366, + 294, + 419 + ], + "spans": [ + { + "bbox": [ + 62, + 366, + 294, + 419 + ], + "type": "text", + "content": "[93] Yukun Zhai, Xiaogiang Zhang, Xiameng Qin, Sanyuan Zhao, Xingping Dong, and Jianbing Shen. Textformer: A query-based end-to-end text spotter with mixed supervision. Machine Intelligence Research, 21(4):704-717, 2024. 2" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 62, + 422, + 294, + 465 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 422, + 294, + 465 + ], + "spans": [ + { + "bbox": [ + 62, + 422, + 294, + 465 + ], + "type": "text", + "content": "[94] Chongsheng Zhang, Guowen Peng, Yuefeng Tao, Feifei Fu, Wei Jiang, George Almpanidis, and Ke Chen. Shopsign: A diverse scene text dataset of chinese shop signs in street views. arXiv preprint arXiv:1903.10412, 2019. 3" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 62, + 468, + 294, + 510 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 468, + 294, + 510 + ], + "spans": [ + { + "bbox": [ + 62, + 468, + 294, + 510 + ], + "type": "text", + "content": "[95] Hesuo Zhang, Lingyu Liang, and Lianwen Jin. Scuthcdoc: A new benchmark dataset of handwritten chinese text in unconstrained camera-captured documents. Pattern Recognition, 108:107559, 2020. 2, 3, 7" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 62, + 512, + 294, + 555 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 512, + 294, + 555 + ], + "spans": [ + { + "bbox": [ + 62, + 512, + 294, + 555 + ], + "type": "text", + "content": "[96] Hao Zhang, Feng Li, Shilong Liu, Lei Zhang, Hang Su, Jun Zhu, Lionel M Ni, and Heung-Yeung Shum. Dino: Detr with improved denoising anchor boxes for end-to-end object detection. arXiv preprint arXiv:2203.03605, 2022. 3" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 62, + 557, + 294, + 600 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 557, + 294, + 600 + ], + "spans": [ + { + "bbox": [ + 62, + 557, + 294, + 600 + ], + "type": "text", + "content": "[97] Hao Zhang, Feng Li, Huaizhe Xu, Shijia Huang, Shilong Liu, Lionel M Ni, and Lei Zhang. Mp-former: Mask-piloted transformer for image segmentation. In CVPR, pages 18074-18083, 2023. 3" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 62, + 602, + 294, + 644 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 602, + 294, + 644 + ], + "spans": [ + { + "bbox": [ + 62, + 602, + 294, + 644 + ], + "type": "text", + "content": "[98] Jiaxin Zhang, Dezhi Peng, Chongyu Liu, Peirong Zhang, and Lianwen Jin. Docres: A generalist model toward unifying document image restoration tasks. In CVPR, pages 15654-15664, 2024. 3" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 62, + 647, + 294, + 689 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 647, + 294, + 689 + ], + "spans": [ + { + "bbox": [ + 62, + 647, + 294, + 689 + ], + "type": "text", + "content": "[99] Peng Zhang, Can Li, Liang Qiao, Zhanzhan Cheng, Shiliang Pu, Yi Niu, and Fei Wu. Vsr: a unified framework for document layout analysis combining vision, semantics and relations. In ICDAR, pages 115-130. Springer, 2021. 8" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 58, + 692, + 294, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 58, + 692, + 294, + 713 + ], + "spans": [ + { + "bbox": [ + 58, + 692, + 294, + 713 + ], + "type": "text", + "content": "[100] Rui Zhang, Yongsheng Zhou, Qianyi Jiang, Qi Song, Nan Li, Kai Zhou, Lei Wang, Dong Wang, Minghui Liao," + } + ] + } + ], + "index": 14 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 58, + 73, + 553, + 530 + ], + "type": "list", + "angle": 0, + "index": 26, + "blocks": [ + { + "bbox": [ + 339, + 73, + 553, + 106 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 339, + 73, + 553, + 106 + ], + "spans": [ + { + "bbox": [ + 339, + 73, + 553, + 106 + ], + "type": "text", + "content": "Mingkun Yang, et al. Icdar 2019 robust reading challenge on reading chinese text on signboard. In ICDAR, pages 1577-1581. IEEE, 2019. 3" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 317, + 107, + 553, + 150 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 107, + 553, + 150 + ], + "spans": [ + { + "bbox": [ + 317, + 107, + 553, + 150 + ], + "type": "text", + "content": "[101] Shi-Xue Zhang, Xiaobin Zhu, Jie-Bo Hou, Chang Liu, Chun Yang, Hongfa Wang, and Xu-Cheng Yin. Deep relational reasoning graph network for arbitrary shape text detection. In CVPR, pages 9699–9708, 2020. 3" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 317, + 152, + 553, + 195 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 152, + 553, + 195 + ], + "spans": [ + { + "bbox": [ + 317, + 152, + 553, + 195 + ], + "type": "text", + "content": "[102] Wayne Xin Zhao, Kun Zhou, Junyi Li, Tianyi Tang, Xiaolei Wang, Yupeng Hou, Yingqian Min, Beichen Zhang, Junjie Zhang, Zican Dong, et al. A survey of large language models. arXiv preprint arXiv:2303.18223, 2023. 3" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 317, + 198, + 553, + 250 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 198, + 553, + 250 + ], + "spans": [ + { + "bbox": [ + 317, + 198, + 553, + 250 + ], + "type": "text", + "content": "[103] Enze Zheng, Zhiwei Wang, Gang Yu, Xingyi Zhang, Yuheng Wu, Hongsheng Li, and Yonghong Tian. Rethinking semantic segmentation from a sequence-to-sequence perspective with transformers. In CVPR, pages 12164-12173, 2021. 3" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 317, + 253, + 553, + 306 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 253, + 553, + 306 + ], + "spans": [ + { + "bbox": [ + 317, + 253, + 553, + 306 + ], + "type": "text", + "content": "[104] Xinyi Zheng, Douglas Burdick, Lucian Popa, Xu Zhong, and Nancy Xin Ru Wang. Global table extractor (gte): A framework for joint table identification and cell structure recognition using visual context. In WACV, pages 697-706, 2021. 3, 7" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 317, + 309, + 553, + 351 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 309, + 553, + 351 + ], + "spans": [ + { + "bbox": [ + 317, + 309, + 553, + 351 + ], + "type": "text", + "content": "[105] Zhaohui Zheng, Ping Wang, Wei Liu, Jinze Li, Rongguang Ye, and Dongwei Ren. Distance-iou loss: Faster and better learning for bounding box regression. In AAAI, pages 12993-13000, 2020. 5" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 317, + 354, + 553, + 386 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 354, + 553, + 386 + ], + "spans": [ + { + "bbox": [ + 317, + 354, + 553, + 386 + ], + "type": "text", + "content": "[106] Xu Zhong, Jianbin Tang, and Antonio Jimeno Yepes. Publaynet: largest dataset ever for document layout analysis. In ICDAR, pages 1015-1022. IEEE, 2019. 2, 3, 7" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 317, + 388, + 553, + 430 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 388, + 553, + 430 + ], + "spans": [ + { + "bbox": [ + 317, + 388, + 553, + 430 + ], + "type": "text", + "content": "[107] Xu Zhong, Elaheh ShafieiBavani, and Antonio Jimeno Yepes. Image-based table recognition: data, model, and evaluation. In ECCV, pages 564-580. Springer, 2020. 2,3,7" + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 317, + 433, + 553, + 485 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 433, + 553, + 485 + ], + "spans": [ + { + "bbox": [ + 317, + 433, + 553, + 485 + ], + "type": "text", + "content": "[108] Xingyi Zhu, Junwei Dai, Lu Lu, Yuwen Zhang, Peizhao Wang, Yisen Sun, Wanli Ouyang, and Ping Luo. Deformable detr: Deformable transformers for end-to-end object detection. In ICCV, pages 12352-12361, 2021. 1, 3, 4, 8" + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 317, + 488, + 553, + 530 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 488, + 553, + 530 + ], + "spans": [ + { + "bbox": [ + 317, + 488, + 553, + 530 + ], + "type": "text", + "content": "[109] Ziyi Zhu, Liangcai Gao, Yibo Li, Yilun Huang, Lin Du, Ning Lu, and Xianfeng Wang. Ntable: a dataset for camera-based table detection. In ICDAR, pages 117-129. Springer, 2021. 3" + } + ] + } + ], + "index": 25 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 749, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 749, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 749, + 317, + 757 + ], + "type": "text", + "content": "15032" + } + ] + } + ], + "index": 27 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 11 + } + ], + "_backend": "vlm", + "_version_name": "2.6.4" +} \ No newline at end of file diff --git a/2025/DocVLM_ Make Your VLM an Efficient Reader/ec872868-b641-4059-95e8-62856f47ad39_content_list.json b/2025/DocVLM_ Make Your VLM an Efficient Reader/ec872868-b641-4059-95e8-62856f47ad39_content_list.json new file mode 100644 index 0000000000000000000000000000000000000000..04cac68792445582b81a8eed73dff31b92c87aaa --- /dev/null +++ b/2025/DocVLM_ Make Your VLM an Efficient Reader/ec872868-b641-4059-95e8-62856f47ad39_content_list.json @@ -0,0 +1,1437 @@ +[ + { + "type": "image", + "img_path": "images/30136c38bce2d4f8bb1643f13deb7e09a6f1103d93600da549ab5a82392284d5.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 200, + 112, + 240, + 161 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "DocVLM: Make Your VLM an Efficient Reader", + "text_level": 1, + "bbox": [ + 274, + 128, + 759, + 148 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Mor Shpigel Nacson*† Aviad Aberdam* Roy Ganz Elad Ben Avraham Alona Golts Yair Kittenplon Shai Mazor Ron Litman AWS AI Labs", + "bbox": [ + 205, + 180, + 790, + 233 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Abstract", + "text_level": 1, + "bbox": [ + 248, + 268, + 326, + 284 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Vision-Language Models (VLMs) excel in diverse visual tasks but face challenges in document understanding, which requires fine-grained text processing. While typical visual tasks perform well with low-resolution inputs, reading-intensive applications demand high-resolution, resulting in significant computational overhead. Using OCR-extracted text in VLM prompts partially addresses this issue but underperforms compared to full-resolution counterpart, as it lacks the complete visual context needed for optimal performance. We introduce DocVLM, a method that integrates an OCR-based modality into VLMs to enhance document processing while preserving original weights. Our approach employs an OCR encoder to capture textual content and layout, compressing these into a compact set of learned queries incorporated into the VLM. Comprehensive evaluations across leading VLMs show that DocVLM significantly reduces reliance on high-resolution images for document understanding. In limited-token regimes $(448\\times 448)$ , DocVLM with 64 learned queries improves DocVQA results from $56.0\\%$ to $86.6\\%$ when integrated with InternVL2 and from $84.4\\%$ to $91.2\\%$ with Qwen2-VL. In LLaVA-OneVision, DocVLM achieves improved results while using $80\\%$ less image tokens. The reduced token usage allows processing multiple pages effectively, showing impressive zero-shot results on DUDE and state-of-the-art performance on MP-DocVQA, highlighting DocVLM's potential for applications requiring high-performance and efficiency.", + "bbox": [ + 89, + 301, + 485, + 710 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "1. Introduction", + "text_level": 1, + "bbox": [ + 91, + 743, + 220, + 758 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "The ability to read and interpret text within images is crucial for numerous real-world applications, particularly in document understanding. This field encompasses diverse document types, from dense-text to infographics and multipage documents [25, 39-41, 50], involving tasks that require capabilities in text comprehension, layout understanding, and", + "bbox": [ + 89, + 768, + 483, + 861 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/b4a69fd84bc032d003f4ac13f7d108554fbe2c480d72ae78b69684f10e8a3e56.jpg", + "image_caption": [ + "Figure 1. DocVLM enhances VLMs' reading capabilities. Integrating DocVLM (solid lines) in top-performing VLMs (dashed lines) consistently improves the performance across all token budgets, frequently surpassing the baseline at higher token counts." + ], + "image_footnote": [], + "bbox": [ + 517, + 268, + 897, + 501 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "visual interpretation. Despite progress in VLMs, processing such documents remains challenging [37], primarily due to the tension between resolution requirements and computational efficiency. While typical computer vision tasks achieve good performance with low-resolution inputs (typically $224 \\times 224$ or $336 \\times 336$ pixels), document analysis demands significantly higher resolutions, resulting in substantial computational overhead [10, 60].", + "bbox": [ + 511, + 568, + 906, + 688 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "To address these challenges, some methods incorporate OCR-extracted text directly into language model prompts [8, 17]. However, this approach typically lags behind OCR-free methods with full-resolution, as it fails to capture crucial visual context and layout information [54]. Moreover, it often produces long sequence inputs, increasing latency and computational costs, especially for dense documents.", + "bbox": [ + 511, + 689, + 908, + 794 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Alternatively, recent VLMs [20, 31, 53] have introduced specialized mechanisms to reduce visual token count, such as image resizing, tiling limits, and feature downsampling. However, when applied to document understanding, these methods result in significant performance degradation, creating an undesirable trade-off between computational efficiency and accuracy, as demonstrated in Fig. 1. These lim", + "bbox": [ + 511, + 795, + 908, + 901 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "CVF", + "bbox": [ + 106, + 2, + 181, + 42 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore.", + "bbox": [ + 236, + 0, + 810, + 46 + ], + "page_idx": 0 + }, + { + "type": "page_footnote", + "text": "*Equal contribution. Corresponding author: mnacson@amazon.com.", + "bbox": [ + 107, + 875, + 472, + 887 + ], + "page_idx": 0 + }, + { + "type": "page_footnote", + "text": "$^{\\dagger}$ Work done during an Amazon internship.", + "bbox": [ + 109, + 887, + 334, + 898 + ], + "page_idx": 0 + }, + { + "type": "page_number", + "text": "29005", + "bbox": [ + 478, + 944, + 517, + 955 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "itations underscore the need for a more efficient approach that maintains high-performance document understanding while reducing computational demands.", + "bbox": [ + 89, + 90, + 480, + 136 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "To overcome these limitations, we introduce DocVLM, a model-agnostic method that enhances VLMs' reading ability by utilizing OCR information effectively. Our approach employs an OCR encoder to capture both contextual and layout details from OCR-extracted text, compressing these encodings into 64 learned queries. These queries are projected and fed into the LLM part of the VLM alongside visual features. Unlike some previous methods [5, 19, 30, 32], our compression mechanism avoids the need for a separate compression module or alterations to the LLM architecture.", + "bbox": [ + 88, + 137, + 480, + 287 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "We demonstrate DocVLM's effectiveness across multiple state-of-the-art VLMs: LlaVA-OneVision [31], InternVL2 [18], and Qwen2-VL [53], each employing a unique image token reduction technique. As illustrated in Figure 1, our method significantly improves performance, particularly in low input token regimes. Our experiments demonstrate consistently, across all studied VLMs and visual token budgets, that DocVLM's OCR encoder not only outperforms the baseline of inserting OCR words into VLMs but also achieves this superior performance when compressed to just 64 tokens. This dual advantage of improved performance and reduced token usage allows for better utilization of fixed token budgets, enabling the allocation of more tokens for visual processing and further enhancing overall performance.", + "bbox": [ + 91, + 289, + 483, + 517 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Importantly, this reduction in sequence length improves the scalability of our approach, enabling its application to multipage document understanding tasks without additional training. We demonstrate that DocVLM can be seamlessly extended to long-context scenarios, such as multipage DocVQA [50, 51]. While current OCR-free approaches struggle with the overwhelming amount of data in multipage documents [27], our method achieves strong zero-shot performance on DUDE and surpasses the current state-of-the-art results on MP-DocVQA (86.3% vs. 80.3%), despite not being trained on multipage data.", + "bbox": [ + 89, + 518, + 482, + 684 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Main contributions:", + "text_level": 1, + "bbox": [ + 89, + 685, + 235, + 698 + ], + "page_idx": 1 + }, + { + "type": "list", + "sub_type": "text", + "list_items": [ + "- DocVLM, a model-agnostic method that efficiently integrates OCR information into VLMs, capturing both text and layout without complex integration techniques.", + "- A compression mechanism that reduces OCR data into a compact set of typically 64 learned queries, significantly reducing computational overhead.", + "- Demonstration of DocVLM's effectiveness across different VLM architectures, LlaVA-OneVision, InternVL2, and Qwen2-VL, showing significant performance improvements in low input token regimes (448×448).", + "- Extension of DocVLM to long-context tasks, achieving strong zero-shot performance on DUDE and SOTA results on MP-DocVQA without multipage training data." + ], + "bbox": [ + 89, + 702, + 482, + 898 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "2. Related Work", + "text_level": 1, + "bbox": [ + 513, + 89, + 656, + 104 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "OCR-free Document VLMs. Early VLMs[5, 9, 22, 23, 32, 36, 44, 59, 61] used relatively small image sizes (e.g., $224 \\times 224$ and $336 \\times 336$ ), performing well on natural-image tasks but falling short in document understanding. To address this, recent approaches enhance document understanding capabilities by operating on high-resolution images, developing various strategies to manage the resulting computational burden. Direct processing methods like Donut [29], PaLI-X [16], and Qwen2-VL [53] attempt full-resolution processing but often resort to image resizing for computational feasibility. Tile-based approaches such as UReader [56] and InternVL2 [18] improve efficiency by processing image tiles independently. Other methods, exemplified by LLaVA-1.5 [31] and LLaVA-OneVision [35], process full-scale images as tiles but downsample the resulting visual features. While these approaches offer different trade-offs, our experiments show their performance deteriorates significantly when constrained to fewer visual tokens.", + "bbox": [ + 511, + 119, + 903, + 392 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "OCR-Enhanced Document Understanding. The widespread availability of efficient, open-source OCR models and cost-effective commercial solutions has driven broad adoption of OCR-based approaches in document understanding [1-4, 21, 28, 34, 43, 46, 57, 58]. Several recent works [9, 16, 17, 19] have explored integrating OCR systems with VLMs by feeding extracted text directly into the language model component. Some approaches further enhance this integration by incorporating spatial layout information [12, 14, 22, 52, 54]. While these methods reduce the computational burden of processing high-resolution images, they currently lag behind OCR-free approaches in performance. Additionally, they face challenges with lengthy input sequences, particularly in multipage settings, which can increase latency and computational costs.", + "bbox": [ + 511, + 397, + 903, + 625 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Document Representation Compression. To address efficiency challenges in processing documents, various compression techniques have been developed. For OCR-enhanced approaches, [14] proposed compressing the OCR signal in multi-page documents using a Compression Transformer, which, despite improving performance in multi-page benchmarks, introduces significant complexity to the system. In the OCR-free setting, generic VLM approaches like Q-former [32] and Resampler [5] compress visual features but struggle with text-dense images. Document-specific methods such as TokenPacker [33] and DocCompressor [27] achieve effective visual compression but show reduced performance on document understanding tasks. In contrast, rather than compressing high-resolution visual inputs, our DocVLM method operates on lower-resolution images and compresses the extensive OCR signal, including textual and layout information, into a compact set of features (typically 64).", + "bbox": [ + 511, + 628, + 903, + 901 + ], + "page_idx": 1 + }, + { + "type": "page_number", + "text": "29006", + "bbox": [ + 478, + 944, + 519, + 957 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "3. Our Method", + "text_level": 1, + "bbox": [ + 91, + 89, + 220, + 104 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "We present DocVLM, a model-agnostic approach that enhances VLMs' reading capabilities, enabling operation with lower-resolution inputs while maintaining or improving document understanding accuracy. Our design preserves the base VLM weights, facilitating easy integration across different model architectures and providing flexibility to balance OCR and visual tokens during inference.", + "bbox": [ + 89, + 114, + 483, + 220 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3.1. Architecture", + "text_level": 1, + "bbox": [ + 89, + 229, + 225, + 244 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Our method introduces two main components that complement existing VLM architectures: an OCR encoder that processes OCR extracted text and layout information, and a query compression mechanism that distills this information into a compact representation. We integrate these components with pre-trained VLMs, which employ various strategies to control the number of visual tokens for efficient processing. Figure 2 illustrates the overall architecture.", + "bbox": [ + 89, + 251, + 483, + 371 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "OCR Encoder Architecture We utilize DocFormerV2 [7], a T5-based encoder-decoder [45] designed for document understanding, which incorporates vision, language, and spatial features. Specifically, we leverage only the encoder component, which comprises 344 million parameters, and omit its visual branch to eliminate redundancy with the VLM's vision capabilities and reduce computational complexity. The encoder processes two types of inputs: user instructions and OCR data from an OCR system, which consists of textual tokens and their corresponding 2D positional information [6, 7, 12, 14, 22, 26].", + "bbox": [ + 89, + 378, + 483, + 546 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Query Compression Mechanism To efficiently integrate OCR information into VLMs, we introduce an instruction-aware compression mechanism that distills the OCR encoder's output into a compact set of learned queries. This mechanism significantly reduces the input sequence length for the language model while preserving essential document information. The compression process utilizes $M$ learnable queries $\\mathbf{Q}$ (typically $M = 64$ ), initialized randomly following the OCR encoder embeddings' distribution. These queries are processed by the OCR encoder alongside two types of embeddings: OCR embeddings $(\\mathbf{E}_{\\mathrm{OCR}})$ , which encode both OCR tokens and their bounding boxes, and instruction embeddings $(\\mathbf{E}_{\\mathrm{Instructions}})$ . The encoding process can be represented as:", + "bbox": [ + 89, + 553, + 483, + 763 + ], + "page_idx": 2 + }, + { + "type": "equation", + "text": "\n$$\n\\operatorname {E n c o d e r} \\left(\\left[ \\mathbf {E} _ {\\mathrm {O C R}}, \\mathbf {E} _ {\\text {I n s t r u c t i o n s}}, \\mathbf {Q} \\right]\\right).\n$$\n", + "text_format": "latex", + "bbox": [ + 174, + 771, + 395, + 787 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "From the encoder output, we retain only the $M$ features corresponding to the learned queries. These compressed features are then projected to match the VLM's hidden dimension and concatenated with the visual tokens before entering the language model. This compression significantly reduces the LLM's input sequence length, enabling either more efficient processing or, under a fixed token budget, allocation of additional tokens to visual features.", + "bbox": [ + 89, + 794, + 483, + 914 + ], + "page_idx": 2 + }, + { + "type": "image", + "img_path": "images/554d8389308a98dbf09ffbb9088233db940c4846db8b657ef8958463d5206fc3.jpg", + "image_caption": [ + "Figure 2. DocVLM Architecture. DocVLM enhances document understanding in frozen VLMs by integrating an OCR module with a query compression mechanism. By condensing OCR data into $\\mathbf{M} = 64$ learnable tokens, DocVLM effectively complements visual information, surpassing the VLM's inherent approaches of increasing image resolution or visual feature dimensions." + ], + "image_footnote": [], + "bbox": [ + 521, + 89, + 901, + 368 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Vision Process OCR-free VLMs employ different visual processing methods and various strategies to control the number of visual tokens, aiming to reduce the computational cost of processing high-resolution images needed for document understanding. These approaches can be grouped into three main paradigms:", + "bbox": [ + 511, + 477, + 905, + 568 + ], + "page_idx": 2 + }, + { + "type": "list", + "sub_type": "text", + "list_items": [ + "1. Full Image Processing with Image Resizing (e.g., Qwen2-VL [53]): In this approach, the model processes the entire image as a single input, controlling the number of visual tokens by resizing the image to a fixed or range-constrained image resolutions. While this image processing preserves global context, it incurs quadratic computational complexity with respect to the number of visual tokens.", + "2. Patch-Based Processing with Controlled Tile Count (e.g., InternVL2 [18]): This strategy segments the input image into spatial tiles, processing each independently, and controls the number of visual tokens by limiting the number of tiles. While most implementations incorporate a low-resolution global view, the primary focus on local processing may compromise global context understanding. The computational complexity for this approach scales linearly with respect to the number of tiles as the image increase. The result is improved memory efficiency compared to processing the full image at once, especially for large images.", + "3. Full-Scale Processing with Feature Downsampling (e.g., LlaVA-OneVision [35]): Some VLMs initially" + ], + "bbox": [ + 514, + 569, + 908, + 901 + ], + "page_idx": 2 + }, + { + "type": "page_number", + "text": "29007", + "bbox": [ + 478, + 944, + 517, + 955 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "process full-scale images, but then downsample the visual features into maximal token count before feeding them to the LLM. While this method captures both global and local context, it introduces significant computational overhead during initial full-scale processing.", + "bbox": [ + 109, + 90, + 480, + 167 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Experimental results confirm that our OCR query compression mechanism significantly enhances document understanding capabilities across all three visual processing strategies, demonstrating its effectiveness as a universal enhancement to existing VLM architectures.", + "bbox": [ + 89, + 169, + 480, + 244 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "3.2. Training Strategy", + "text_level": 1, + "bbox": [ + 89, + 251, + 264, + 268 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Our training strategy aims to integrate the OCR modality into existing VLMs while preserving their core strengths. To achieve this, we keep the VLM completely frozen throughout the process and train only the newly introduced OCR components, i.e., the learnable queries, the OCR encoder, and the projection layer. We employ a two-stage training strategy to gradually incorporate the OCR modality into the pre-trained VLM:", + "bbox": [ + 89, + 273, + 482, + 393 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Stage I: OCR-LLM Alignment. During this stage, we withhold image input from the VLM, forcing the model to rely solely on the newly introduced OCR modality. This approach ensures full utilization of OCR data, aligns OCR components with the LLM input space, and reduces sequence length, improving training efficiency. Given the focus on text input, our dataset selection concentrates on text-related tasks. We begin by training only the randomly initialized components: the learnable queries and projection layer. This allows these components to adapt without disrupting the pretrained OCR encoder. Subsequently, we unfreeze the OCR encoder for fine-tuning, enabling a more comprehensive alignment of the entire encoder to the VLM.", + "bbox": [ + 89, + 396, + 482, + 592 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Stage II: Vision Alignment. In this final stage, we incorporate visual information extracted from the visual encoder, encouraging the OCR components to complement the visual features. Our experiments reveal this stage has a particularly strong effect when using fewer learned queries, allowing the compressed OCR information to better supplement the information acquired from the visual modality (see Sec. 5). During this stage, we add more visually focused datasets to the training process. Note that despite our method preserving the original VLM weights, it might implicitly inject bias through prompt tuning. To avoid this, the training data should represent all tasks of interest.", + "bbox": [ + 89, + 593, + 482, + 773 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "3.3. Multipage Document Extension", + "text_level": 1, + "bbox": [ + 89, + 787, + 369, + 803 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "We conduct our training procedure on single-page data only. However, our approach can be extended to operate on multipage documents. Given a multipage input and its OCR information, the VLM independently processes each page image and concatenates the resulting visual features. For the OCR information, we explore two strategies: Global", + "bbox": [ + 89, + 810, + 482, + 900 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Encoding, which compresses the entire document's OCR information into 64 learnable queries, and Page-wise Encoding, which compresses each page's OCR information separately into 64 learnable queries and then concatenates them, resulting in $64 \\times$ number of pages learned queries. After processing the OCR information using either strategy, we feed the resulting compressed OCR features along with the concatenated visual features into the LLM.", + "bbox": [ + 511, + 90, + 903, + 210 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Our experiments demonstrate that both approaches are highly effective and efficient in processing multipage documents. Using either approach with a restricted number of visual tokens, we obtain strong zero-shot results on DUDE [51] and state-of-the-art results on MP-DocVQA [50]. The page-wise encoding strategy yields slightly better results for lower numbers of visual tokens.", + "bbox": [ + 511, + 212, + 903, + 316 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "4. Experiments", + "text_level": 1, + "bbox": [ + 511, + 334, + 645, + 349 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "4.1. Experimental Setting", + "text_level": 1, + "bbox": [ + 511, + 359, + 714, + 376 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Model Integration: We evaluate DocVLM through integration with three leading open-source VLMs: LLaVA-OneVision [31], InternVL2 [18], and Qwen2-VL [53]. As discussed in Sec. 3.1, these models employ distinct token reduction strategies, enabling us to assess DocVLM's effectiveness across different visual processing approaches.", + "bbox": [ + 511, + 381, + 903, + 472 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Training: Our training protocol, detailed in Sec. 3.2, employs a two-phase strategy. The initial phase focuses on text-centric tasks using datasets spanning document understanding (DocVQA [40], InfoVQA [41]), scene text analysis (ST-VQA [11], TextVQA [48], OCR-VQA [42]), and specialized tasks (ChartQA [39], TextCaps [47], TATDQA [62]). The subsequent vision alignment phase incorporates additional visual-centric datasets: COCO Caption [15] and VQA-V2 [24].", + "bbox": [ + 511, + 473, + 903, + 608 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Evaluation: For evaluation, we focus on five key benchmarks: DocVQA, TextVQA, ST-VQA, InfoVQA, and TextCaps. Results are reported on test sets where available, with TextVQA and TextCaps evaluated on validation sets due to test server restrictions. We use ANLS as the evaluation metric for all datasets, except TextVQA, which uses VQAScore, and TextCaps, which uses CIDEr. To demonstrate DocVLM's generalization capabilities, we conduct zero-shot evaluation on multipage document understanding benchmarks: DUDE [51] and MP-DocVQA [50]. This zero-shot performance is particularly noteworthy as our model is trained exclusively on single-page documents. The supplementary presents additional implementation details, including hyperparameters and optimization strategies, and robustness analysis to different OCR systems.", + "bbox": [ + 511, + 609, + 903, + 835 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "4.2. State-of-the-art Comparisons", + "text_level": 1, + "bbox": [ + 511, + 847, + 774, + 864 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Table 1 presents comprehensive comparisons between DocVLM and other state-of-the-art methods across var", + "bbox": [ + 511, + 869, + 903, + 900 + ], + "page_idx": 3 + }, + { + "type": "page_number", + "text": "29008", + "bbox": [ + 478, + 944, + 517, + 955 + ], + "page_idx": 3 + }, + { + "type": "table", + "img_path": "images/04c5ded42318b47c2743b294d6086a1e819832a43461dd929fbd215199f96e33.jpg", + "table_caption": [], + "table_footnote": [], + "table_body": "
Method# Tok.#PDocVQATextVQAST-VQAInfoVQATextCAPSMP-DocVQADUDE*
No Token Limitations
GPT-4o92.877.4-79.2---
Gemini 1.5 Pro93.178.7-81.0---
GPT-4V87.278.0-75.1---
KOSMOS-2.5-CHAT4K1.3B81.140.7-41.3---
TextSquare2.5K8.6B84.366.8-51.5---
ScreenAI3.5K5B87.8--57.8-72.9-
ScreenAI+OCR4.3K5B89.9--65.9-77.1-
Pali-35.5K5B86.779.584.157.8158.8--
Pali-3+OCR6.3K5B88.680.885.762.4164.3--
# Tokens ≤ 1.5k
URader8417B65.457.6-42.2118.4--
Monkey1.3K9B66.564.3-36.193.2--
TextMonkey7689B73.065.9-28.6---
Vary2567B76.3------
DocOwl23248B80.766.7-46.4131.869.446.8
GRAM9001B85.3----80.351.2
GRAMC-Former2561B87.6----77.645.5
DocFormer v21K1B87.864.071.848.8-76.448.4
LLaVA-OneVision7K7B87.576.171.168.8138.0OOMOOM
LLaVA-OneVision1.5K7B66.572.170.645.6112.941.828.7
DocVLMLLaVA-OneVision (Ours)1.5K7B88.476.970.861.0145.377.943.8
InternVL 23.1K8B91.677.4-74.8-OOMOOM
InternVL 22568B56.065.765.738.451.151.030.5
DocVLMInternVL2 (Ours)3208B86.671.274.357.6119.476.243.3
InternVL 212808B85.775.568.361.543.778.142.2
DocVLMInternVL2 (Ours)13448B91.076.776.765.4123.481.845.6
Qwen2-VL16k7B94.584.370.776.5150.2OOMOOM
Qwen2-VL3207B84.478.070.154.1142.173.041.5
DocVLMQwen2-VL (Ours)3207B91.279.676.561.2144.381.746.1
Qwen2-VL5767B91.582.370.565.3145.082.145.9
DocVLMQwen2-VL (Ours)5767B92.882.879.866.8150.484.547.4
", + "bbox": [ + 96, + 88, + 906, + 550 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Table 1. Comparison with State-of-the-Art Methods. Performance evaluation of DocVLM against state-of-the-art approaches on document understanding benchmarks. Results are categorized into unconstrained models and those with a 1.5k token limit. In the constrained token regime, DocVLM consistently enhances the performance of baseline VLMs across various tasks and visual token budgets. Notably, DocVLM paired with Qwen2-VL (576 tokens) achieves superior performance across all evaluated datasets, including state-of-the-art zero-shot accuracy on DUDE. $\\star$ indicates zero-shot evaluation, with grey entries denoting non-zero-shot results.", + "bbox": [ + 89, + 559, + 906, + 630 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "ious document understanding benchmarks, highlighting DocVLM's ability to improve performance under token constraints. We categorize the results into two main groups: methods without token constraints (both closed and open-source models) and those operating under a 1.5k token limit. We mainly focused on models with around 7B parameters and include methods using an OCR system such as Pali3 [10] ScreenAI [8], DocFormerV2 [7], and GRAM [14].", + "bbox": [ + 89, + 638, + 482, + 760 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "To evaluate DocVLM's effectiveness under token constraints, we integrated it with three baseline models: LlaVA-OneVision [31], InternVL2 [18], and Qwen2-VL [53], each configured to operate within the 1.5k token limit. For LlaVA-OneVision, we utilized the minimal visual token configuration (single visual features tile). InternVL2 was tested with both single-tile (256 tokens) and four-tile (1280 tokens) configurations, while Qwen2-VL was evaluated with 256 and 512 visual tokens, corresponding to image", + "bbox": [ + 89, + 763, + 482, + 902 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "sizes of $448 \\times 448$ and $616 \\times 616$ respectively.", + "bbox": [ + 511, + 638, + 826, + 654 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Under the practical 1.5k token constraint, incorporating DocVLM with each of these baseline models yields substantial and consistent improvements. Notably, these improvements persist even in looser token regimes, such as InternVL2 with 1280 visual tokens and Qwen2-VL with 576 visual tokens. Within this constraint, our Qwen2-VL variant with DocVLM, using just 576 tokens, achieves state-of-the-art performance across all benchmarks: $92.8\\%$ on DocVQA, $82.8\\%$ on TextVQA, $79.8\\%$ on ST-VQA, $66.8\\%$ on InfoVQA, and a CIDEr score of 150.4 on TextCAPS. As detailed in the Appendix, DocVLM achieves these results with significant computational advantages.", + "bbox": [ + 509, + 656, + 906, + 837 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "DocVLM also demonstrates exceptional capability in handling multipage documents. Using the same 576-token configuration described earlier, it achieves $84.5\\%$ accuracy on MP-DocVQA, surpassing previous state-of-the-art re", + "bbox": [ + 511, + 839, + 908, + 902 + ], + "page_idx": 4 + }, + { + "type": "page_number", + "text": "29009", + "bbox": [ + 478, + 944, + 519, + 957 + ], + "page_idx": 4 + }, + { + "type": "image", + "img_path": "images/c79c75b43eab3846f46dea7e2283a6ba1c12a778a56419801d1d40e808b72f2d.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 98, + 89, + 486, + 271 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/f06f08792c330366c71a07d59eacc3a42ea1b996951bb686181e954aeadfa07a.jpg", + "image_caption": [ + "Figure 3. Qualitative Results. Representative examples of DocVLM's performance across diverse document formats, from dense text to infographics and scene text. Our model successfully handles complex layouts, dense content, and presents instruction-following capabilities without explicit training on such datasets. Each example includes an image-instruction pair with baseline and DocVLM predictions." + ], + "image_footnote": [], + "bbox": [ + 94, + 286, + 486, + 467 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/cf27e73cc1d6f78ec8df28502c3f840cca7bd06203ebfe8b2dc9b7a0c2d44593.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 491, + 85, + 901, + 286 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/f77db2901218e23d6ed4ad97896dd025f8c728bc9de19b73954b498ab3344c3e.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 493, + 290, + 901, + 468 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "sults. Furthermore, this setup shows robust generalization capabilities with a zero-shot performance of $47.4\\%$ accuracy on the DUDE dataset, despite not being specifically trained for multipage document processing.", + "bbox": [ + 89, + 535, + 483, + 597 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "4.3. Qualitative Results", + "text_level": 1, + "bbox": [ + 89, + 625, + 272, + 640 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Figure 3 illustrates DocVLM's enhanced capabilities through representative examples that demonstrate three key strengths of our method: (1) improved reading comprehension in complex document layouts, (2) effective handling of dense textual content despite using compressed representations, and (3) preserved and enhanced instruction-following capabilities.", + "bbox": [ + 89, + 652, + 483, + 758 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "The examples span diverse document types, from dense text documents to infographics and scene text, showcasing DocVLM's versatility. In the infographic example, DocVLM not only preserves but enhances the base model's instruction-following capabilities, despite our OCR component not being explicitly trained on instruction-following datasets like [49]. This demonstrates that our compression mechanism successfully retains crucial textual and layout information while significantly reducing token usage.", + "bbox": [ + 89, + 763, + 483, + 902 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "4.4. Scaling to Multipage Documents", + "text_level": 1, + "bbox": [ + 511, + 534, + 799, + 551 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Building on the promising multipage results in Tab. 1, we conduct an in-depth analysis of different DocVLM configurations for multipage document understanding. This analysis focuses on the Qwen2-VL base model, tested on the MP-DocVQA dataset with documents up to 20 pages long – a scale that most other state-of-the-art methods struggle to handle due to token limitations.", + "bbox": [ + 511, + 558, + 906, + 664 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Table 2 compares four OCR integration strategies used during inference in multipage scenarios:", + "bbox": [ + 511, + 667, + 906, + 698 + ], + "page_idx": 5 + }, + { + "type": "list", + "sub_type": "text", + "list_items": [ + "- Baseline: vision-only input (no additional tokens)", + "- Direct OCR word insertion: up to 800 tokens per page", + "- Global OCR encoding: 64 tokens total", + "- Page-wise OCR encoding: 64 tokens per page" + ], + "bbox": [ + 511, + 699, + 885, + 760 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Results indicate consistent improvements over the baseline across all image resolutions (256, 512, and 1024 tokens), with only minimal additional token usage – just 64 tokens for the entire document in the global encoding case. Notably, at 256 visual tokens per page, both page-wise encoding $(82.4\\%)$ and global encoding $(81.7\\%)$ outperform direct OCR word insertion $(79.1\\%)$ while using significantly fewer tokens.", + "bbox": [ + 511, + 762, + 906, + 882 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Our best configuration achieves state-of-the-art perfor", + "bbox": [ + 531, + 885, + 903, + 901 + ], + "page_idx": 5 + }, + { + "type": "page_number", + "text": "29010", + "bbox": [ + 478, + 944, + 519, + 957 + ], + "page_idx": 5 + }, + { + "type": "table", + "img_path": "images/5f669879a86d0397d295b876fea90a2fe624c0206f9242fbcd8ca657ef3186ab.jpg", + "table_caption": [], + "table_footnote": [], + "table_body": "
MethodLLM OCR InputImage Tok.OCR Tok.ANLS
DocOwl2 [27]-324 × pg-69.4
GRAMC-Former [14]-100 × pg25677.6
GRAM [14]-100 × pg800 × pg80.3
Qwen2-VL-256 × pg-73.0
OCR Words256 × pg800 × pg79.1
DocVLMQwen2-VLGlobal Encoding256 × pg6481.7
DocVLMQwen2-VLPage-wise Encoding256 × pg64 × pg82.4
Qwen2-VL-512 × pg-82.1
DocVLMQwen2-VLGlobal Encoding512 × pg6484.5
DocVLMQwen2-VLPage-wise Encoding512 × pg64 × pg85.2
Qwen2-VL-1024 × pg-85.2
DocVLMQwen2-VLGlobal Encoding1024 × pg6486.3
DocVLMQwen2-VLPage-wise Encoding1024 × pg64 × pg86.3
", + "bbox": [ + 94, + 88, + 480, + 248 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "mance of $86.3\\%$ ANLS, significantly outperforming specialized multipage models like GRAM $(80.3\\%)$ . This is particularly impressive considering that DocVLM was trained exclusively on single-page inputs, demonstrating strong zero-shot generalization to multipage scenarios.", + "bbox": [ + 89, + 359, + 482, + 434 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "The comparison between encoding strategies reveals that page-wise encoding consistently outperforms global encoding at lower visual token counts, providing a $+0.7\\%$ improvement for both 256 and 512 image tokens per page. This advantage diminishes at 1024 tokens where both achieve identical performance (86.3% ANLS). Remarkably, DocVLM with page-wise encoding matches or even outperforms the baseline Qwen2-VL using twice as many visual tokens, highlighting the efficiency of our approach.", + "bbox": [ + 89, + 436, + 482, + 571 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "5. Ablation Study", + "text_level": 1, + "bbox": [ + 89, + 587, + 243, + 604 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Impact of OCR Encoding Strategies To evaluate the impact of OCR encoding compression, we compare three strategies for integrating OCR information: (1) inserting raw OCR words in the original VLM, (2) using DocVLM uncompressed OCR encodings, and (3) DocVLM compressed OCR encodings with 64 learned queries. We evaluate these approaches on the DocVQA test set using three representative model configurations: LLaVA-OneVision with 1.5K visual tokens, and both InternVL2 and Qwen2-VL with 256 visual tokens each.", + "bbox": [ + 89, + 612, + 482, + 762 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Results in Tab. 3 demonstrate that DocVLM's OCR encodings significantly outperform raw OCR words across all three models while maintaining the same token count. Notably, our compressed encoding approach, using just 64 tokens instead of 800 OCR tokens, preserves most of these improvements while drastically reducing sequence length. This efficient compression enables a more favorable allocation of the token budget, allowing models to dedicate more tokens to visual processing without compromising OCR ef", + "bbox": [ + 89, + 763, + 482, + 901 + ], + "page_idx": 6 + }, + { + "type": "table", + "img_path": "images/f137ccaa022b3b79d89cfd7bfa6604652b08c4868e0dc00b98ce8191b728aa87.jpg", + "table_caption": [ + "Table 2. Extension to Multipage. Comparison on MP-DocVQA of approaches for incorporating OCR information in multipage document understanding. Both DocVLM multipage extension strategies: global encoding (64 tokens per document) and pagewise encoding (64 tokens per page), outperform previous state-of-the-art methods, notably, without any explicit multipage training." + ], + "table_footnote": [], + "table_body": "
LLM OCR InputOCR Tok.LLaVA-OV 1.5KInternVL2 256Qwen2-VL 256
OCR Words80085.884.489.1
OCR Encoding80089.489.291.9
64 Compressed Encoding6488.486.691.2
", + "bbox": [ + 517, + 88, + 903, + 152 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Table 3. OCR Encoding Strategies. DocVQA results for inserting OCR information using (1) OCR words (baseline), (2) uncompressed OCR encodings, and (3) 64 compressed OCR encodings.", + "bbox": [ + 511, + 162, + 903, + 204 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "fectiveness. The results validate that DocVLM's compression strategy successfully balances performance with computational efficiency, a key factor for practical applications.", + "bbox": [ + 511, + 229, + 903, + 275 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Balancing Vision and OCR Token Allocation Modern VLMs employ various mechanisms to reduce visual token count, creating an inherent trade-off between computational efficiency and model performance, as discussed in Sec. 3.1. We investigate how DocVLM can improve this trade-off by comparing four configurations: (1) baseline VLM without OCR, (2) direct OCR word insertion, (3) DocVLM with uncompressed OCR encodings, and (4) DocVLM with 64 compressed learned queries.", + "bbox": [ + 511, + 316, + 906, + 455 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Figure 4 presents the performance scores on DocVQA validation (left y-axis) and total token counts (right y-axis) for three VLM architectures, showing how these metrics vary across different visual token allocations. Each model employs a distinct token reduction approach: LLaVA-OneVision controls token count through feature downsampling (AnyRes Max), InternVL2 limits the number of processed image tiles (Dynamic Max Batch), and Qwen2-VL adjusts image resolution to constrain token count.", + "bbox": [ + 511, + 458, + 908, + 594 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Our analysis reveals that integrating OCR information, regardless of the method used, consistently improves performance across all models, with particularly pronounced gains in low visual token regimes. However, uncompressed OCR integration methods, whether through direct word insertion or uncompressed DocVLM encodings, require 800 tokens – a significant overhead that could otherwise be allocated to visual processing. For instance, allocating 128 tokens for visual and 800 for OCR in Qwen2-VL achieves $84.3\\%$ using OCR words and $90.1\\%$ using uncompressed encodings. In contrast, using 896 pure visual tokens reaches $92.4\\%$ , demonstrating the potential benefit of allocating more tokens to visual processing. DocVLM's compression mechanism provides a superior option by requiring only 64 tokens for OCR information while maintaining strong performance. In the above example, our approach using fewer tokens, allocating 768 visual tokens and 64 OCR tokens, reaches $93.0\\%$ , outperforming the $90.1\\%$ obtained with the uncompressed encodings, highlighting DocVLM's effective balance between visual and OCR tokens.", + "bbox": [ + 511, + 599, + 908, + 900 + ], + "page_idx": 6 + }, + { + "type": "page_number", + "text": "29011", + "bbox": [ + 478, + 944, + 517, + 955 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/96eb4592a64f4bd6744a858002f3392f1d3dd2945f18b048be3ead0ad81d6f65.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 96, + 89, + 361, + 229 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/423a8f08b63fd05d073cb4c1fec9eba16c90a79ee2783da1cf9bda63ab362dbb.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 361, + 89, + 630, + 229 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/7898eecf2f11b51e09f0d852cb9383f32525f92a2316d36d6ac74b6c688b1e60.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 630, + 90, + 901, + 229 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/649344a343f98136978236c1a3dcd4d41574aae6ad50f6203bcbedd27a1402a3.jpg", + "image_caption": [ + "Figure 4. Balancing Performance and Compute. Analysis of model performance (lines, left y-axis) and token usage (bars, right y-axis) as a function of visual token allocation. Each model employs its inherent token control strategy: AnyRes max for feature downsampling (LLaVA One-Vision), dynamic max tiles (InternVL2), and max image tokens for resolution control (Qwen2-VL). The results highlight that DocVLM consistently improves performance with minimal overhead (64 tokens), offering an efficient OCR-visual token allocation." + ], + "image_footnote": [], + "bbox": [ + 256, + 233, + 738, + 260 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/e7ec1136051f6b341035ba7cbc22e3d91980c4f7b9dde9710591956639eb1e57.jpg", + "image_caption": [ + "Figure 5. Compression Levels. DocVQA validation results for DocVLM integrated with Qwen2-VL across varying OCR and image token budgets. \"0\" represents the baseline, while \"Full\" indicates uncompressed encodings." + ], + "image_footnote": [], + "bbox": [ + 112, + 339, + 460, + 506 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Compression Levels We deepen our analysis of the OCR-visual token trade-off by examining different compression levels in DocVLM integrated with Qwen2-VL. Fig. 5 presents ANLS scores across various combinations of visual tokens (128-1024) and learned queries (16-256), including baselines without OCR and with uncompressed encoding. In low visual token regimes, increasing the number of learned queries yields substantial improvements, validating our compression mechanism's effectiveness in capturing relevant OCR information. Notably, even with just 16 learned queries, DocVLM outperforms the baseline across all visual token configurations, offering strong performance with minimal computational overhead.", + "bbox": [ + 89, + 593, + 483, + 790 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Training Stages Tab. 4 illustrates the impact of the vision alignment stage on ANLS performance for DocVLM with the LLaVA-OneVision base model on the DocVQA test set, showing results for varying numbers of learned queries (16 to 128) and the uncompressed case. Our two-stage training process initially trains OCR modality components with-", + "bbox": [ + 89, + 809, + 483, + 902 + ], + "page_idx": 7 + }, + { + "type": "table", + "img_path": "images/57dd9f561e1c9effc80f99bdf7f4a7038c4c9f2e57e3cc3637b993e57bf9486a.jpg", + "table_caption": [], + "table_footnote": [], + "table_body": "
Training PhasesCompressed Enc.OCR Enc.
1664128800
Stage I: OCR-LLM Alignment81.785.886.389.4
+ Stage II: Vision Alignment87.988.488.490.1
Δ+6.2+2.6+2.1+0.7
", + "bbox": [ + 517, + 338, + 903, + 415 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Table 4. Training Phases. Second-stage training consistently improves DocVQA performance, both with compressed DocVLM tokens and full OCR encoding.", + "bbox": [ + 511, + 425, + 906, + 468 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "out image input, forcing reliance on OCR data alone, before reintroducing the image in the vision alignment stage to adapt learned queries alongside visual information. The results reveal that vision alignment significantly boosts performance, especially with fewer learned queries: for instance, with 16 learned queries, there's an improvement of +6.2, compared to +0.7 in the uncompressed case. Notably, after vision alignment, DocVLM with only 16 learned queries outperforms the baseline of OCR words (from Table 3). These findings underscore the effectiveness of our two-stage training method.", + "bbox": [ + 511, + 483, + 906, + 650 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "6. Conclusions", + "text_level": 1, + "bbox": [ + 511, + 664, + 640, + 680 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Our results demonstrate that DocVLM can be effectively integrated into various VLMs to enhance their document reading capabilities while significantly reducing their dependency on extensive vision tokens. The key takeaway is that in token-constrained scenarios, allocating a small portion of tokens to OCR information consistently yields better results than using those tokens solely for visual processing. The effectiveness of our compression mechanism extends beyond single-page documents, as evidenced by achieving state-of-the-art results on MP-DocVQA using the same 64 tokens to represent multiple pages. These results establish DocVLM as a practical solution for enhancing document understanding in real-world applications where computational efficiency is crucial.", + "bbox": [ + 511, + 688, + 908, + 902 + ], + "page_idx": 7 + }, + { + "type": "page_number", + "text": "29012", + "bbox": [ + 478, + 944, + 519, + 957 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "References", + "text_level": 1, + "bbox": [ + 91, + 89, + 187, + 104 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[1] Aviad Aberdam, Ron Litman, Shahar Tsiper, Oron Anschel, Ron Slossberg, Shai Mazor, R Manmatha, and Pietro Perona. Sequence-to-sequence contrastive learning for text recognition. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 15302-15312, 2021. 2", + "[2] Aviad Aberdam, Roy Ganz, Shai Mazor, and Ron Litman. Multimodal semi-supervised learning for text recognition. arXiv preprint arXiv:2205.03873, 2022.", + "[3] Aviad Aberdam, David Bensaïd, Alona Golts, Roy Ganz, Oren Nuriel, Royee Tichauer, Shai Mazor, and Ron Litman. Clipter: Looking at the bigger picture in scene text recognition. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 21706-21717, 2023.", + "[4] Ofir Abramovich, Niv Nayman, Sharon Fogel, Inbal Lavi, Ron Litman, Shahar Tsiper, Royee Tichauer, Srikar Appalaraju, Shai Mazor, and R Manmatha. Visfocus: Prompt-guided vision encoders forOCR-free dense document understanding. In European Conference on Computer Vision, pages 241-259. Springer, 2025. 2", + "[5] Jean-Baptiste Alayrac, Jeff Donahue, Pauline Luc, Antoine Miech, Iain Barr, Yana Hasson, Karel Lenc, Arthur Mensch, Katherine Millican, Malcolm Reynolds, et al. Flamingo: a visual language model for few-shot learning. Advances in neural information processing systems, 35:23716-23736, 2022. 2", + "[6] Srikar Appalaraju, Bhavan Jasani, Bhargava Urala Kota, Yusheng Xie, and R Manmatha. Docformer: End-to-end transformer for document understanding. In Proceedings of the IEEE/CVF international conference on computer vision, pages 993-1003, 2021. 3", + "[7] Srikar Appalaraju, Peng Tang, Qi Dong, Nishant Sankaran, Yichu Zhou, and R Manmatha. Docformerv2: Local features for document understanding. In Proceedings of the AAAI Conference on Artificial Intelligence, pages 709-718, 2024. 3, 5, 12", + "[8] Gilles Baechler, Srinivas Sunkara, Maria Wang, Fedir Zubach, Hassan Mansoor, Vincent Etter, Victor Cärbune, Jason Lin, Jindong Chen, and Abhanshu Sharma. Screenai: A vision-language model for ui and infographics understanding. arXiv preprint arXiv:2402.04615, 2024. 1, 5", + "[9] Jinze Bai, Shuai Bai, Shusheng Yang, Shijie Wang, Sinan Tan, Peng Wang, Junyang Lin, Chang Zhou, and Jingren Zhou. Qwen-vl: A versatile vision-language model for understanding, localization, text reading, and beyond. arXiv preprint arXiv:2308.12966, 1(2):3, 2023. 2", + "[10] Lucas Beyer, Andreas Steiner, André Susano Pinto, Alexander Kolesnikov, Xiao Wang, Daniel Salz, Maxim Neumann, Ibrahim Alabdulmohsin, Michael Tschannen, Emanuele Bugliarello, et al. Paligemma: A versatile 3b vlm for transfer. arXiv preprint arXiv:2407.07726, 2024. 1, 5", + "[11] Ali Furkan Biten, Ruben Tito, Andres Mafla, Lluis Gomez, Marçal Rusinol, Ernest Valveny, CV Jawahar, and Dimosthenis Karatzas. Scene text visual question answering. In Proceedings of the IEEE/CVF international conference on computer vision, pages 4291-4301, 2019. 4, 12" + ], + "bbox": [ + 93, + 114, + 483, + 900 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[12] Ali Furkan Biten, Ron Litman, Yusheng Xie, Srikar Appalaraju, and R Manmatha. Latr: Layout-aware transformer for scene-text vqa. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 16548-16558, 2022. 2, 3", + "[13] Ali Furkan Biten, Ruben Tito, Lluis Gomez, Ernest Valveny, and Dimosthenis Karatzas. Ocr-idl: Ocr annotations for industry document library dataset. arXiv preprint arXiv:2202.12985, 2022. 12", + "[14] Tsachi Blau, Sharon Fogel, Roi Ronen, Alona Golts, Roy Ganz, Elad Ben Avraham, Aviad Aberdam, Shahar Tsiper, and Ron Litman. Gram: Global reasoning for multi-page vqa. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 15598-15607, 2024. 2, 3, 5, 7", + "[15] Xinlei Chen, Hao Fang, Tsung-Yi Lin, Ramakrishna Vedantam, Saurabh Gupta, Piotr Dólar, and C Lawrence Zitnick. Microsoft coco captions: Data collection and evaluation server. arXiv preprint arXiv:1504.00325, 2015. 4, 12", + "[16] Xi Chen, Josip Djolonga, Piotr Padlewski, Basil Mustafa, Soravit Changpinyo, Jialin Wu, Carlos Riquelme Ruiz, Sebastian Goodman, Xiao Wang, Yi Tay, et al. Pali-x: On scaling up a multilingual vision and language model. arXiv preprint arXiv:2305.18565, 2023. 2", + "[17] Xi Chen, Xiao Wang, Lucas Beyer, Alexander Kolesnikov, Jialin Wu, Paul Voigtlaender, Basil Mustafa, Sebastian Goodman, Ibrahim Alabdulmohsin, Piotr Padlewski, et al. Pali-3 vision language models: Smaller, faster, stronger. arXiv preprint arXiv:2310.09199, 2023. 1, 2", + "[18] Zhe Chen, Weiyun Wang, Hao Tian, Shenglong Ye, Zhangwei Gao, Erfei Cui, Wenwen Tong, Kongzhi Hu, Jiapeng Luo, Zheng Ma, et al. Internvl2: Better than the best—expanding performance boundaries of open-source multimodal models with the progressive scaling strategy, 2024. 2, 3, 4, 5", + "[19] Wenliang Dai, Junnan Li, Dongxu Li, Anthony Meng Huat Tiong, Junqi Zhao, Weisheng Wang, Boyang Albert Li, Pascale Fung, and Steven C. H. Hoi. Instructclip: Towards general-purpose vision-language models with instruction tuning. ArXiv, abs/2305.06500, 2023. 2", + "[20] Xiaoyi Dong, Pan Zhang, Yuhang Zang, Yuhang Cao, Bin Wang, Linke Ouyang, Songyang Zhang, Haodong Duan, Wenwei Zhang, Yining Li, et al. Internlm-xcomposer2-4khd: A pioneering large vision-language model handling resolutions from 336 pixels to 4k hd. arXiv preprint arXiv:2404.06512, 2024. 1", + "[21] Masato Fujitake. Dtrocr: Decoder-only transformer for optical character recognition. In Proceedings of the IEEE/CVF Winter Conference on Applications of Computer Vision, pages 8025-8035, 2024. 2", + "[22] Roy Ganz, Oren Nuriel, Aviad Aberdam, Yair Kittenplon, Shai Mazor, and Ron Litman. Towards models that can see and read. In Proceedings of the IEEE/CVF international conference on computer vision, pages 21718-21728, 2023. 2, 3", + "[23] Roy Ganz, Yair Kittenplon, Aviad Aberdam, Elad Ben Avraham, Oren Nuriel, Shai Mazor, and Ron Litman. Question aware vision transformer for multimodal reasoning. In Proceedings of the IEEE/CVF Conference on Computer Vision" + ], + "bbox": [ + 93, + 92, + 903, + 900 + ], + "page_idx": 8 + }, + { + "type": "page_number", + "text": "29013", + "bbox": [ + 478, + 944, + 517, + 955 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "and Pattern Recognition (CVPR), pages 13861-13871, 2024. 2", + "[24] Yash Goyal, Tejas Khot, Douglas Summers-Stay, Dhruv Bartra, and Devi Parikh. Making the V in VQA matter: Elevating the role of image understanding in Visual Question Answering. In Conference on Computer Vision and Pattern Recognition (CVPR), 2017. 4, 12", + "[25] Jean-Philippe Thiran Guillaume Jaume, Hazim Kemal Ekenel. Funsd: A dataset for form understanding in noisy scanned documents. In Accepted to ICDAR-OST, 2019. 1", + "[26] Benjamin Hsu, Xiaoyu Liu, Huayang Li, Yoshinari Fujinuma, Maria Nadejde, Xing Niu, Yair Kittenplon, Ron Litman, and Raghavendra Pappagari. M3t: A new benchmark dataset for multi-modal document-level machine translation. arXiv preprint arXiv:2406.08255, 2024. 3", + "[27] Anwen Hu, Haiyang Xu, Liang Zhang, Jiabo Ye, Ming Yan, Ji Zhang, Qin Jin, Fei Huang, and Jingren Zhou. mplug-docowl2: High-resolution compressing forocr-free multi-page document understanding. arXiv preprint arXiv:2409.03420, 2024. 2, 7", + "[28] Taeho Kil, Seonghyeon Kim, Sukmin Seo, Yoonsik Kim, and Daehe Kim. Towards unified scene text spotting based on sequence generation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 15223-15232, 2023. 2", + "[29] Geewook Kim, Teakgyu Hong, Moonbin Yim, Jinyoung Park, Jinyeong Yim, Wonseok Hwang, Sangdoo Yun, Dongyoon Han, and Seunghyun Park. Donut: Document understanding transformer withoutOCR.arXiv preprint arXiv:2111.15664, 7(15):2, 2021. 2", + "[30] Hugo Laurençon, Léo Tronchon, Matthieu Cord, and Victor Sanh. What matters when building vision-language models? arXiv preprint arXiv:2405.02246, 2024. 2", + "[31] Bo Li, Yuanhan Zhang, Dong Guo, Renrui Zhang, Feng Li, Hao Zhang, Kaichen Zhang, Yanwei Li, Ziwei Liu, and Chunyuan Li. Llava-onevision: Easy visual task transfer. arXiv preprint arXiv:2408.03326, 2024. 1, 2, 4, 5", + "[32] Junnan Li, Dongxu Li, Silvio Savarese, and Steven Hoi. Blip-2: Bootstrapping language-image pre-training with frozen image encoders and large language models. In International conference on machine learning, pages 19730-19742. PMLR, 2023. 2", + "[33] Wentong Li, Yuqian Yuan, Jian Liu, Dongqi Tang, Song Wang, Jianke Zhu, and Lei Zhang. Tokenpacker: Efficient visual projector for multimodal llm. arXiv preprint arXiv:2407.02392, 2024. 2", + "[34] Ron Litman, Oron Anschel, Shahar Tsiper, Roeit Litman, Shai Mazor, and R Manmatha. Scatter: selective context attentional scene text recognizer. In proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 11962-11972, 2020. 2", + "[35] Haotian Liu, Chunyuan Li, Yuheng Li, Bo Li, Yuanhan Zhang, Sheng Shen, and Yong Jae Lee. Llava next: Improved reasoning,OCR, and world knowledge, 2024. 2, 3", + "[36] Haotian Liu, Chunyuan Li, Qingyang Wu, and Yong Jae Lee. Visual instruction tuning. Advances in neural information processing systems, 36, 2024. 2" + ], + "bbox": [ + 91, + 90, + 482, + 900 + ], + "page_idx": 9 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[37] Yuliang Liu, Zhang Li, Mingxin Huang, Biao Yang, Wenwen Yu, Chunyuan Li, Xucheng Yin, Cheng lin Liu, Lianwen Jin, and Xiang Bai. Ocrbench: On the hidden mystery ofOCR in large multimodal models, 2024. 1", + "[38] Jiahao Lyu, Wei Wang, Dongbao Yang, Jinwen Zhong, and Yu Zhou. Arbitrary reading order scene text spotter with local semantics guidance. arXiv preprint arXiv:2412.10159, 2024. 14", + "[39] Ahmed Masry, Do Xuan Long, Jia Qing Tan, Shafiq Joty, and Enamul Hoque. Chartqa: A benchmark for question answering about charts with visual and logical reasoning. arXiv preprint arXiv:2203.10244, 2022. 1, 4, 12", + "[40] Minesh Mathew, Dimosthenis Karatzas, and CV Jawahar. Docvqa: A dataset for vqa on document images. In Proceedings of the IEEE/CVF winter conference on applications of computer vision, pages 2200-2209, 2021. 4, 12, 13", + "[41] Minesh Mathew, Viraj Bagal, Ruben Tito, Dimosthenis Karatzas, Ernest Valveny, and CV Jawahar. Infographicvqa. In Proceedings of the IEEE/CVF Winter Conference on Applications of Computer Vision, pages 1697-1706, 2022. 1, 4, 12, 13", + "[42] Anand Mishra, Shashank Shekhar, Ajeet Kumar Singh, and Anirban Chakraborty. Ocr-vqa: Visual question answering by reading text in images. In 2019 international conference on document analysis and recognition (ICDAR), pages 947-952. IEEE, 2019. 4, 12", + "[43] Oren Nuriel, Sharon Fogel, and Ron Litman. Textadain: Paying attention to shortcut learning in text recognizers. In European Conference on Computer Vision, pages 427-445. Springer, 2022. 2", + "[44] Zhiliang Peng, Wenhui Wang, Li Dong, Yaru Hao, Shaohan Huang, Shuming Ma, and Furu Wei. Kosmos-2: Grounding multimodal large language models to the world. arXiv preprint arXiv:2306.14824, 2023. 2", + "[45] Colin Raffel, Noam Shazeer, Adam Roberts, Katherine Lee, Sharan Narang, Michael Matena, Yanqi Zhou, Wei Li, and Peter J Liu. Exploring the limits of transfer learning with a unified text-to-text transformer. Journal of machine learning research, 21(140):1-67, 2020. 3", + "[46] Roi Ronen, Shahar Tsiper, Oron Anschel, Inbal Lavi, Amir Markovitz, and R Manmatha. Glass: Global to local attention for scene-text spotting. arXiv preprint arXiv:2208.03364, 2022. 2", + "[47] Oleksii Sidorov, Ronghang Hu, Marcus Rohrbach, and Amanpreet Singh. Textcaps: a dataset for image captioning with reading comprehension. In Computer Vision-ECCV 2020: 16th European Conference, Glasgow, UK, August 23-28, 2020, Proceedings, Part II 16, pages 742-758. Springer, 2020. 4, 12", + "[48] Amanpreet Singh, Vivek Natarajan, Meet Shah, Yu Jiang, Xinlei Chen, Dhruv Batra, Devi Parikh, and Marcus Rohrbach. Towards vqa models that can read. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 8317-8326, 2019. 4, 12", + "[49] Ryota Tanaka, Taichi Iki, Kyosuke Nishida, Kuniko Saito, and Jun Suzuki. Instructdoc: A dataset for zero-shot generalization of visual document understanding with instructions." + ], + "bbox": [ + 516, + 90, + 906, + 900 + ], + "page_idx": 9 + }, + { + "type": "page_number", + "text": "29014", + "bbox": [ + 478, + 944, + 519, + 955 + ], + "page_idx": 9 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "In Proceedings of the AAAI Conference on Artificial Intelligence, pages 19071-19079, 2024. 6", + "[50] Ruben Tito, Dimosthenis Karatzas, and Ernest Valveny. Hierarchical multimodal transformers for multipage docvqa. Pattern Recognition, 144:109834, 2023. 1, 2, 4, 12", + "[51] Jordy Van Landeghem, Ruben Tito, Łukasz Borchmann, Michal Pietruszka, Pawel Joziak, Rafal Powalski, Dawid Jurkiewicz, Mickaël Coustaty, Bertrand Anckaert, Ernest Valveny, et al. Document understanding dataset and evaluation (dude). In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 19528-19540, 2023. 2, 4, 12", + "[52] Dongsheng Wang, Natraj Raman, Mathieu Sibue, Zhiqiang Ma, Petr Babkin, Simerjot Kaur, Yulong Pei, Armineh Nourbakhsh, and Xiaomo Liu. Docllm: A layout-aware generative language model for multimodal document understanding. arXiv preprint arXiv:2401.00908, 2023. 2, 13", + "[53] Peng Wang, Shuai Bai, Sinan Tan, Shijie Wang, Zhihao Fan, Jinze Bai, Keqin Chen, Xuejing Liu, Jialin Wang, Wenbin Ge, et al. Qwen2-vl: Enhancing vision-language model's perception of the world at any resolution. arXiv preprint arXiv:2409.12191, 2024. 1, 2, 3, 4, 5", + "[54] Wenjin Wang, Yunhao Li, Yixin Ou, and Yin Zhang. Layout and task aware instruction prompt for zero-shot document image question answering. arXiv preprint arXiv:2306.00526, 2023. 1, 2, 13", + "[55] Yiheng Xu, Minghao Li, Lei Cui, Shaohan Huang, Furu Wei, and Ming Zhou. Layoutm: Pre-training of text and layout for document image understanding. In Proceedings of the 26th ACM SIGKDD international conference on knowledge discovery & data mining, pages 1192-1200, 2020. 13", + "[56] Jiabo Ye, Anwen Hu, Haiyang Xu, Qinghao Ye, Ming Yan, Guohai Xu, Chenliang Li, Junfeng Tian, Qi Qian, Ji Zhang, et al. Ureader: UniversalOCR-free visually-situated language understanding with multimodal large language model. arXiv preprint arXiv:2310.05126, 2023. 2", + "[57] Maoyuan Ye, Jing Zhang, Shanshan Zhao, Juhua Liu, Bo Du, and Dacheng Tao. Aptext-detr: Towards better scene text detection with dynamic points in transformer. In Proceedings of the AAAI Conference on Artificial Intelligence, pages 3241-3249, 2023. 2", + "[58] Maoyuan Ye, Jing Zhang, Shanshan Zhao, Juhua Liu, Tongliang Liu, Bo Du, and Dacheng Tao. Deepsolo: Let transformer decoder with explicit points solo for text spotting. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 19348-19357, 2023. 2", + "[59] Qinghao Ye, Haiyang Xu, Guohai Xu, Jiabo Ye, Ming Yan, Yiyang Zhou, Junyang Wang, Anwen Hu, Pengcheng Shi, Yaya Shi, et al. MPLug-owl: Modularization empowers large language models with multimodality. arXiv preprint arXiv:2304.14178, 2023. 2", + "[60] Pan Zhang, Xiaoyi Dong, Yuhang Zang, Yuhang Cao, Rui Qian, Lin Chen, Qipeng Guo, Haodong Duan, Bin Wang, Linke Ouyang, et al. Internlm-xcomposer-2.5: A versatile large vision language model supporting long-contextual input and output. arXiv preprint arXiv:2407.03320, 2024. 1" + ], + "bbox": [ + 91, + 90, + 482, + 901 + ], + "page_idx": 10 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[61] Deyao Zhu, Jun Chen, Xiaogian Shen, Xiang Li, and Mohamed Elhoseiny. Minigpt-4: Enhancing vision-language understanding with advanced large language models. arXiv preprint arXiv:2304.10592, 2023. 2", + "[62] Fengbin Zhu, Wenqiang Lei, Fuli Feng, Chao Wang, Haozhou Zhang, and Tat-Seng Chua. Towards complex document understanding by discrete reasoning. In Proceedings of the 30th ACM International Conference on Multimedia, pages 4857-4866, 2022. 4, 12" + ], + "bbox": [ + 514, + 90, + 905, + 219 + ], + "page_idx": 10 + }, + { + "type": "page_number", + "text": "29015", + "bbox": [ + 478, + 944, + 517, + 955 + ], + "page_idx": 10 + } +] \ No newline at end of file diff --git a/2025/DocVLM_ Make Your VLM an Efficient Reader/ec872868-b641-4059-95e8-62856f47ad39_model.json b/2025/DocVLM_ Make Your VLM an Efficient Reader/ec872868-b641-4059-95e8-62856f47ad39_model.json new file mode 100644 index 0000000000000000000000000000000000000000..a76168519f3455a821e7fd4e7daf52ec6c8a0fd5 --- /dev/null +++ b/2025/DocVLM_ Make Your VLM an Efficient Reader/ec872868-b641-4059-95e8-62856f47ad39_model.json @@ -0,0 +1,2202 @@ +[ + [ + { + "type": "header", + "bbox": [ + 0.107, + 0.003, + 0.182, + 0.043 + ], + "angle": 0, + "content": "CVF" + }, + { + "type": "header", + "bbox": [ + 0.237, + 0.001, + 0.812, + 0.047 + ], + "angle": 0, + "content": "This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore." + }, + { + "type": "image", + "bbox": [ + 0.202, + 0.113, + 0.241, + 0.162 + ], + "angle": 0, + "content": null + }, + { + "type": "title", + "bbox": [ + 0.275, + 0.13, + 0.761, + 0.15 + ], + "angle": 0, + "content": "DocVLM: Make Your VLM an Efficient Reader" + }, + { + "type": "text", + "bbox": [ + 0.206, + 0.181, + 0.792, + 0.234 + ], + "angle": 0, + "content": "Mor Shpigel Nacson*† Aviad Aberdam* Roy Ganz Elad Ben Avraham Alona Golts Yair Kittenplon Shai Mazor Ron Litman AWS AI Labs" + }, + { + "type": "title", + "bbox": [ + 0.249, + 0.269, + 0.327, + 0.285 + ], + "angle": 0, + "content": "Abstract" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.303, + 0.486, + 0.712 + ], + "angle": 0, + "content": "Vision-Language Models (VLMs) excel in diverse visual tasks but face challenges in document understanding, which requires fine-grained text processing. While typical visual tasks perform well with low-resolution inputs, reading-intensive applications demand high-resolution, resulting in significant computational overhead. Using OCR-extracted text in VLM prompts partially addresses this issue but underperforms compared to full-resolution counterpart, as it lacks the complete visual context needed for optimal performance. We introduce DocVLM, a method that integrates an OCR-based modality into VLMs to enhance document processing while preserving original weights. Our approach employs an OCR encoder to capture textual content and layout, compressing these into a compact set of learned queries incorporated into the VLM. Comprehensive evaluations across leading VLMs show that DocVLM significantly reduces reliance on high-resolution images for document understanding. In limited-token regimes \\((448\\times 448)\\), DocVLM with 64 learned queries improves DocVQA results from \\(56.0\\%\\) to \\(86.6\\%\\) when integrated with InternVL2 and from \\(84.4\\%\\) to \\(91.2\\%\\) with Qwen2-VL. In LLaVA-OneVision, DocVLM achieves improved results while using \\(80\\%\\) less image tokens. The reduced token usage allows processing multiple pages effectively, showing impressive zero-shot results on DUDE and state-of-the-art performance on MP-DocVQA, highlighting DocVLM's potential for applications requiring high-performance and efficiency." + }, + { + "type": "title", + "bbox": [ + 0.092, + 0.744, + 0.222, + 0.759 + ], + "angle": 0, + "content": "1. Introduction" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.77, + 0.484, + 0.862 + ], + "angle": 0, + "content": "The ability to read and interpret text within images is crucial for numerous real-world applications, particularly in document understanding. This field encompasses diverse document types, from dense-text to infographics and multipage documents [25, 39-41, 50], involving tasks that require capabilities in text comprehension, layout understanding, and" + }, + { + "type": "image", + "bbox": [ + 0.518, + 0.27, + 0.898, + 0.502 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.512, + 0.506, + 0.907, + 0.563 + ], + "angle": 0, + "content": "Figure 1. DocVLM enhances VLMs' reading capabilities. Integrating DocVLM (solid lines) in top-performing VLMs (dashed lines) consistently improves the performance across all token budgets, frequently surpassing the baseline at higher token counts." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.569, + 0.907, + 0.689 + ], + "angle": 0, + "content": "visual interpretation. Despite progress in VLMs, processing such documents remains challenging [37], primarily due to the tension between resolution requirements and computational efficiency. While typical computer vision tasks achieve good performance with low-resolution inputs (typically \\(224 \\times 224\\) or \\(336 \\times 336\\) pixels), document analysis demands significantly higher resolutions, resulting in substantial computational overhead [10, 60]." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.69, + 0.909, + 0.795 + ], + "angle": 0, + "content": "To address these challenges, some methods incorporate OCR-extracted text directly into language model prompts [8, 17]. However, this approach typically lags behind OCR-free methods with full-resolution, as it fails to capture crucial visual context and layout information [54]. Moreover, it often produces long sequence inputs, increasing latency and computational costs, especially for dense documents." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.796, + 0.909, + 0.902 + ], + "angle": 0, + "content": "Alternatively, recent VLMs [20, 31, 53] have introduced specialized mechanisms to reduce visual token count, such as image resizing, tiling limits, and feature downsampling. However, when applied to document understanding, these methods result in significant performance degradation, creating an undesirable trade-off between computational efficiency and accuracy, as demonstrated in Fig. 1. These lim" + }, + { + "type": "page_footnote", + "bbox": [ + 0.109, + 0.875, + 0.473, + 0.888 + ], + "angle": 0, + "content": "*Equal contribution. Corresponding author: mnacson@amazon.com." + }, + { + "type": "page_footnote", + "bbox": [ + 0.111, + 0.888, + 0.335, + 0.9 + ], + "angle": 0, + "content": "\\(^{\\dagger}\\)Work done during an Amazon internship." + }, + { + "type": "list", + "bbox": [ + 0.109, + 0.875, + 0.473, + 0.9 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.48, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "29005" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.09, + 0.092, + 0.482, + 0.137 + ], + "angle": 0, + "content": "itations underscore the need for a more efficient approach that maintains high-performance document understanding while reducing computational demands." + }, + { + "type": "text", + "bbox": [ + 0.089, + 0.138, + 0.482, + 0.289 + ], + "angle": 0, + "content": "To overcome these limitations, we introduce DocVLM, a model-agnostic method that enhances VLMs' reading ability by utilizing OCR information effectively. Our approach employs an OCR encoder to capture both contextual and layout details from OCR-extracted text, compressing these encodings into 64 learned queries. These queries are projected and fed into the LLM part of the VLM alongside visual features. Unlike some previous methods [5, 19, 30, 32], our compression mechanism avoids the need for a separate compression module or alterations to the LLM architecture." + }, + { + "type": "text", + "bbox": [ + 0.093, + 0.29, + 0.485, + 0.518 + ], + "angle": 0, + "content": "We demonstrate DocVLM's effectiveness across multiple state-of-the-art VLMs: LlaVA-OneVision [31], InternVL2 [18], and Qwen2-VL [53], each employing a unique image token reduction technique. As illustrated in Figure 1, our method significantly improves performance, particularly in low input token regimes. Our experiments demonstrate consistently, across all studied VLMs and visual token budgets, that DocVLM's OCR encoder not only outperforms the baseline of inserting OCR words into VLMs but also achieves this superior performance when compressed to just 64 tokens. This dual advantage of improved performance and reduced token usage allows for better utilization of fixed token budgets, enabling the allocation of more tokens for visual processing and further enhancing overall performance." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.519, + 0.483, + 0.685 + ], + "angle": 0, + "content": "Importantly, this reduction in sequence length improves the scalability of our approach, enabling its application to multipage document understanding tasks without additional training. We demonstrate that DocVLM can be seamlessly extended to long-context scenarios, such as multipage DocVQA [50, 51]. While current OCR-free approaches struggle with the overwhelming amount of data in multipage documents [27], our method achieves strong zero-shot performance on DUDE and surpasses the current state-of-the-art results on MP-DocVQA (86.3% vs. 80.3%), despite not being trained on multipage data." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.686, + 0.236, + 0.699 + ], + "angle": 0, + "content": "Main contributions:" + }, + { + "type": "text", + "bbox": [ + 0.091, + 0.703, + 0.483, + 0.748 + ], + "angle": 0, + "content": "- DocVLM, a model-agnostic method that efficiently integrates OCR information into VLMs, capturing both text and layout without complex integration techniques." + }, + { + "type": "text", + "bbox": [ + 0.091, + 0.748, + 0.483, + 0.792 + ], + "angle": 0, + "content": "- A compression mechanism that reduces OCR data into a compact set of typically 64 learned queries, significantly reducing computational overhead." + }, + { + "type": "text", + "bbox": [ + 0.091, + 0.793, + 0.483, + 0.853 + ], + "angle": 0, + "content": "- Demonstration of DocVLM's effectiveness across different VLM architectures, LlaVA-OneVision, InternVL2, and Qwen2-VL, showing significant performance improvements in low input token regimes (448×448)." + }, + { + "type": "text", + "bbox": [ + 0.091, + 0.853, + 0.483, + 0.899 + ], + "angle": 0, + "content": "- Extension of DocVLM to long-context tasks, achieving strong zero-shot performance on DUDE and SOTA results on MP-DocVQA without multipage training data." + }, + { + "type": "list", + "bbox": [ + 0.091, + 0.703, + 0.483, + 0.899 + ], + "angle": 0, + "content": null + }, + { + "type": "title", + "bbox": [ + 0.514, + 0.09, + 0.657, + 0.106 + ], + "angle": 0, + "content": "2. Related Work" + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.12, + 0.905, + 0.393 + ], + "angle": 0, + "content": "OCR-free Document VLMs. Early VLMs[5, 9, 22, 23, 32, 36, 44, 59, 61] used relatively small image sizes (e.g., \\(224 \\times 224\\) and \\(336 \\times 336\\)), performing well on natural-image tasks but falling short in document understanding. To address this, recent approaches enhance document understanding capabilities by operating on high-resolution images, developing various strategies to manage the resulting computational burden. Direct processing methods like Donut [29], PaLI-X [16], and Qwen2-VL [53] attempt full-resolution processing but often resort to image resizing for computational feasibility. Tile-based approaches such as UReader [56] and InternVL2 [18] improve efficiency by processing image tiles independently. Other methods, exemplified by LLaVA-1.5 [31] and LLaVA-OneVision [35], process full-scale images as tiles but downsample the resulting visual features. While these approaches offer different trade-offs, our experiments show their performance deteriorates significantly when constrained to fewer visual tokens." + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.398, + 0.905, + 0.625 + ], + "angle": 0, + "content": "OCR-Enhanced Document Understanding. The widespread availability of efficient, open-source OCR models and cost-effective commercial solutions has driven broad adoption of OCR-based approaches in document understanding [1-4, 21, 28, 34, 43, 46, 57, 58]. Several recent works [9, 16, 17, 19] have explored integrating OCR systems with VLMs by feeding extracted text directly into the language model component. Some approaches further enhance this integration by incorporating spatial layout information [12, 14, 22, 52, 54]. While these methods reduce the computational burden of processing high-resolution images, they currently lag behind OCR-free approaches in performance. Additionally, they face challenges with lengthy input sequences, particularly in multipage settings, which can increase latency and computational costs." + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.629, + 0.905, + 0.902 + ], + "angle": 0, + "content": "Document Representation Compression. To address efficiency challenges in processing documents, various compression techniques have been developed. For OCR-enhanced approaches, [14] proposed compressing the OCR signal in multi-page documents using a Compression Transformer, which, despite improving performance in multi-page benchmarks, introduces significant complexity to the system. In the OCR-free setting, generic VLM approaches like Q-former [32] and Resampler [5] compress visual features but struggle with text-dense images. Document-specific methods such as TokenPacker [33] and DocCompressor [27] achieve effective visual compression but show reduced performance on document understanding tasks. In contrast, rather than compressing high-resolution visual inputs, our DocVLM method operates on lower-resolution images and compresses the extensive OCR signal, including textual and layout information, into a compact set of features (typically 64)." + }, + { + "type": "page_number", + "bbox": [ + 0.48, + 0.945, + 0.521, + 0.958 + ], + "angle": 0, + "content": "29006" + } + ], + [ + { + "type": "title", + "bbox": [ + 0.092, + 0.09, + 0.221, + 0.106 + ], + "angle": 0, + "content": "3. Our Method" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.115, + 0.484, + 0.222 + ], + "angle": 0, + "content": "We present DocVLM, a model-agnostic approach that enhances VLMs' reading capabilities, enabling operation with lower-resolution inputs while maintaining or improving document understanding accuracy. Our design preserves the base VLM weights, facilitating easy integration across different model architectures and providing flexibility to balance OCR and visual tokens during inference." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.231, + 0.226, + 0.245 + ], + "angle": 0, + "content": "3.1. Architecture" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.252, + 0.484, + 0.372 + ], + "angle": 0, + "content": "Our method introduces two main components that complement existing VLM architectures: an OCR encoder that processes OCR extracted text and layout information, and a query compression mechanism that distills this information into a compact representation. We integrate these components with pre-trained VLMs, which employ various strategies to control the number of visual tokens for efficient processing. Figure 2 illustrates the overall architecture." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.38, + 0.484, + 0.547 + ], + "angle": 0, + "content": "OCR Encoder Architecture We utilize DocFormerV2 [7], a T5-based encoder-decoder [45] designed for document understanding, which incorporates vision, language, and spatial features. Specifically, we leverage only the encoder component, which comprises 344 million parameters, and omit its visual branch to eliminate redundancy with the VLM's vision capabilities and reduce computational complexity. The encoder processes two types of inputs: user instructions and OCR data from an OCR system, which consists of textual tokens and their corresponding 2D positional information [6, 7, 12, 14, 22, 26]." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.554, + 0.484, + 0.765 + ], + "angle": 0, + "content": "Query Compression Mechanism To efficiently integrate OCR information into VLMs, we introduce an instruction-aware compression mechanism that distills the OCR encoder's output into a compact set of learned queries. This mechanism significantly reduces the input sequence length for the language model while preserving essential document information. The compression process utilizes \\( M \\) learnable queries \\( \\mathbf{Q} \\) (typically \\( M = 64 \\)), initialized randomly following the OCR encoder embeddings' distribution. These queries are processed by the OCR encoder alongside two types of embeddings: OCR embeddings \\( (\\mathbf{E}_{\\mathrm{OCR}}) \\), which encode both OCR tokens and their bounding boxes, and instruction embeddings \\( (\\mathbf{E}_{\\mathrm{Instructions}}) \\). The encoding process can be represented as:" + }, + { + "type": "equation", + "bbox": [ + 0.176, + 0.772, + 0.396, + 0.789 + ], + "angle": 0, + "content": "\\[\n\\operatorname {E n c o d e r} \\left(\\left[ \\mathbf {E} _ {\\mathrm {O C R}}, \\mathbf {E} _ {\\text {I n s t r u c t i o n s}}, \\mathbf {Q} \\right]\\right).\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.795, + 0.484, + 0.915 + ], + "angle": 0, + "content": "From the encoder output, we retain only the \\(M\\) features corresponding to the learned queries. These compressed features are then projected to match the VLM's hidden dimension and concatenated with the visual tokens before entering the language model. This compression significantly reduces the LLM's input sequence length, enabling either more efficient processing or, under a fixed token budget, allocation of additional tokens to visual features." + }, + { + "type": "image", + "bbox": [ + 0.522, + 0.09, + 0.903, + 0.369 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.512, + 0.382, + 0.907, + 0.466 + ], + "angle": 0, + "content": "Figure 2. DocVLM Architecture. DocVLM enhances document understanding in frozen VLMs by integrating an OCR module with a query compression mechanism. By condensing OCR data into \\(\\mathbf{M} = 64\\) learnable tokens, DocVLM effectively complements visual information, surpassing the VLM's inherent approaches of increasing image resolution or visual feature dimensions." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.478, + 0.906, + 0.569 + ], + "angle": 0, + "content": "Vision Process OCR-free VLMs employ different visual processing methods and various strategies to control the number of visual tokens, aiming to reduce the computational cost of processing high-resolution images needed for document understanding. These approaches can be grouped into three main paradigms:" + }, + { + "type": "text", + "bbox": [ + 0.515, + 0.57, + 0.907, + 0.688 + ], + "angle": 0, + "content": "1. Full Image Processing with Image Resizing (e.g., Qwen2-VL [53]): In this approach, the model processes the entire image as a single input, controlling the number of visual tokens by resizing the image to a fixed or range-constrained image resolutions. While this image processing preserves global context, it incurs quadratic computational complexity with respect to the number of visual tokens." + }, + { + "type": "text", + "bbox": [ + 0.515, + 0.69, + 0.909, + 0.87 + ], + "angle": 0, + "content": "2. Patch-Based Processing with Controlled Tile Count (e.g., InternVL2 [18]): This strategy segments the input image into spatial tiles, processing each independently, and controls the number of visual tokens by limiting the number of tiles. While most implementations incorporate a low-resolution global view, the primary focus on local processing may compromise global context understanding. The computational complexity for this approach scales linearly with respect to the number of tiles as the image increase. The result is improved memory efficiency compared to processing the full image at once, especially for large images." + }, + { + "type": "text", + "bbox": [ + 0.515, + 0.871, + 0.907, + 0.902 + ], + "angle": 0, + "content": "3. Full-Scale Processing with Feature Downsampling (e.g., LlaVA-OneVision [35]): Some VLMs initially" + }, + { + "type": "list", + "bbox": [ + 0.515, + 0.57, + 0.909, + 0.902 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.48, + 0.945, + 0.518, + 0.957 + ], + "angle": 0, + "content": "29007" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.11, + 0.092, + 0.482, + 0.168 + ], + "angle": 0, + "content": "process full-scale images, but then downsample the visual features into maximal token count before feeding them to the LLM. While this method captures both global and local context, it introduces significant computational overhead during initial full-scale processing." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.17, + 0.482, + 0.245 + ], + "angle": 0, + "content": "Experimental results confirm that our OCR query compression mechanism significantly enhances document understanding capabilities across all three visual processing strategies, demonstrating its effectiveness as a universal enhancement to existing VLM architectures." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.252, + 0.265, + 0.269 + ], + "angle": 0, + "content": "3.2. Training Strategy" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.275, + 0.483, + 0.395 + ], + "angle": 0, + "content": "Our training strategy aims to integrate the OCR modality into existing VLMs while preserving their core strengths. To achieve this, we keep the VLM completely frozen throughout the process and train only the newly introduced OCR components, i.e., the learnable queries, the OCR encoder, and the projection layer. We employ a two-stage training strategy to gradually incorporate the OCR modality into the pre-trained VLM:" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.397, + 0.483, + 0.593 + ], + "angle": 0, + "content": "Stage I: OCR-LLM Alignment. During this stage, we withhold image input from the VLM, forcing the model to rely solely on the newly introduced OCR modality. This approach ensures full utilization of OCR data, aligns OCR components with the LLM input space, and reduces sequence length, improving training efficiency. Given the focus on text input, our dataset selection concentrates on text-related tasks. We begin by training only the randomly initialized components: the learnable queries and projection layer. This allows these components to adapt without disrupting the pretrained OCR encoder. Subsequently, we unfreeze the OCR encoder for fine-tuning, enabling a more comprehensive alignment of the entire encoder to the VLM." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.594, + 0.483, + 0.775 + ], + "angle": 0, + "content": "Stage II: Vision Alignment. In this final stage, we incorporate visual information extracted from the visual encoder, encouraging the OCR components to complement the visual features. Our experiments reveal this stage has a particularly strong effect when using fewer learned queries, allowing the compressed OCR information to better supplement the information acquired from the visual modality (see Sec. 5). During this stage, we add more visually focused datasets to the training process. Note that despite our method preserving the original VLM weights, it might implicitly inject bias through prompt tuning. To avoid this, the training data should represent all tasks of interest." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.788, + 0.37, + 0.804 + ], + "angle": 0, + "content": "3.3. Multipage Document Extension" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.811, + 0.483, + 0.901 + ], + "angle": 0, + "content": "We conduct our training procedure on single-page data only. However, our approach can be extended to operate on multipage documents. Given a multipage input and its OCR information, the VLM independently processes each page image and concatenates the resulting visual features. For the OCR information, we explore two strategies: Global" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.092, + 0.905, + 0.212 + ], + "angle": 0, + "content": "Encoding, which compresses the entire document's OCR information into 64 learnable queries, and Page-wise Encoding, which compresses each page's OCR information separately into 64 learnable queries and then concatenates them, resulting in \\(64 \\times\\) number of pages learned queries. After processing the OCR information using either strategy, we feed the resulting compressed OCR features along with the concatenated visual features into the LLM." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.213, + 0.905, + 0.318 + ], + "angle": 0, + "content": "Our experiments demonstrate that both approaches are highly effective and efficient in processing multipage documents. Using either approach with a restricted number of visual tokens, we obtain strong zero-shot results on DUDE [51] and state-of-the-art results on MP-DocVQA [50]. The page-wise encoding strategy yields slightly better results for lower numbers of visual tokens." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.335, + 0.646, + 0.351 + ], + "angle": 0, + "content": "4. Experiments" + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.36, + 0.715, + 0.377 + ], + "angle": 0, + "content": "4.1. Experimental Setting" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.382, + 0.905, + 0.473 + ], + "angle": 0, + "content": "Model Integration: We evaluate DocVLM through integration with three leading open-source VLMs: LLaVA-OneVision [31], InternVL2 [18], and Qwen2-VL [53]. As discussed in Sec. 3.1, these models employ distinct token reduction strategies, enabling us to assess DocVLM's effectiveness across different visual processing approaches." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.474, + 0.905, + 0.609 + ], + "angle": 0, + "content": "Training: Our training protocol, detailed in Sec. 3.2, employs a two-phase strategy. The initial phase focuses on text-centric tasks using datasets spanning document understanding (DocVQA [40], InfoVQA [41]), scene text analysis (ST-VQA [11], TextVQA [48], OCR-VQA [42]), and specialized tasks (ChartQA [39], TextCaps [47], TATDQA [62]). The subsequent vision alignment phase incorporates additional visual-centric datasets: COCO Caption [15] and VQA-V2 [24]." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.611, + 0.905, + 0.837 + ], + "angle": 0, + "content": "Evaluation: For evaluation, we focus on five key benchmarks: DocVQA, TextVQA, ST-VQA, InfoVQA, and TextCaps. Results are reported on test sets where available, with TextVQA and TextCaps evaluated on validation sets due to test server restrictions. We use ANLS as the evaluation metric for all datasets, except TextVQA, which uses VQAScore, and TextCaps, which uses CIDEr. To demonstrate DocVLM's generalization capabilities, we conduct zero-shot evaluation on multipage document understanding benchmarks: DUDE [51] and MP-DocVQA [50]. This zero-shot performance is particularly noteworthy as our model is trained exclusively on single-page documents. The supplementary presents additional implementation details, including hyperparameters and optimization strategies, and robustness analysis to different OCR systems." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.848, + 0.776, + 0.865 + ], + "angle": 0, + "content": "4.2. State-of-the-art Comparisons" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.871, + 0.905, + 0.901 + ], + "angle": 0, + "content": "Table 1 presents comprehensive comparisons between DocVLM and other state-of-the-art methods across var" + }, + { + "type": "page_number", + "bbox": [ + 0.48, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "29008" + } + ], + [ + { + "type": "table", + "bbox": [ + 0.097, + 0.089, + 0.907, + 0.551 + ], + "angle": 0, + "content": "
Method# Tok.#PDocVQATextVQAST-VQAInfoVQATextCAPSMP-DocVQADUDE*
No Token Limitations
GPT-4o92.877.4-79.2---
Gemini 1.5 Pro93.178.7-81.0---
GPT-4V87.278.0-75.1---
KOSMOS-2.5-CHAT4K1.3B81.140.7-41.3---
TextSquare2.5K8.6B84.366.8-51.5---
ScreenAI3.5K5B87.8--57.8-72.9-
ScreenAI+OCR4.3K5B89.9--65.9-77.1-
Pali-35.5K5B86.779.584.157.8158.8--
Pali-3+OCR6.3K5B88.680.885.762.4164.3--
# Tokens ≤ 1.5k
URader8417B65.457.6-42.2118.4--
Monkey1.3K9B66.564.3-36.193.2--
TextMonkey7689B73.065.9-28.6---
Vary2567B76.3------
DocOwl23248B80.766.7-46.4131.869.446.8
GRAM9001B85.3----80.351.2
GRAMC-Former2561B87.6----77.645.5
DocFormer v21K1B87.864.071.848.8-76.448.4
LLaVA-OneVision7K7B87.576.171.168.8138.0OOMOOM
LLaVA-OneVision1.5K7B66.572.170.645.6112.941.828.7
DocVLMLLaVA-OneVision (Ours)1.5K7B88.476.970.861.0145.377.943.8
InternVL 23.1K8B91.677.4-74.8-OOMOOM
InternVL 22568B56.065.765.738.451.151.030.5
DocVLMInternVL2 (Ours)3208B86.671.274.357.6119.476.243.3
InternVL 212808B85.775.568.361.543.778.142.2
DocVLMInternVL2 (Ours)13448B91.076.776.765.4123.481.845.6
Qwen2-VL16k7B94.584.370.776.5150.2OOMOOM
Qwen2-VL3207B84.478.070.154.1142.173.041.5
DocVLMQwen2-VL (Ours)3207B91.279.676.561.2144.381.746.1
Qwen2-VL5767B91.582.370.565.3145.082.145.9
DocVLMQwen2-VL (Ours)5767B92.882.879.866.8150.484.547.4
" + }, + { + "type": "table_caption", + "bbox": [ + 0.09, + 0.56, + 0.907, + 0.631 + ], + "angle": 0, + "content": "Table 1. Comparison with State-of-the-Art Methods. Performance evaluation of DocVLM against state-of-the-art approaches on document understanding benchmarks. Results are categorized into unconstrained models and those with a 1.5k token limit. In the constrained token regime, DocVLM consistently enhances the performance of baseline VLMs across various tasks and visual token budgets. Notably, DocVLM paired with Qwen2-VL (576 tokens) achieves superior performance across all evaluated datasets, including state-of-the-art zero-shot accuracy on DUDE. \\(\\star\\) indicates zero-shot evaluation, with grey entries denoting non-zero-shot results." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.639, + 0.483, + 0.761 + ], + "angle": 0, + "content": "ious document understanding benchmarks, highlighting DocVLM's ability to improve performance under token constraints. We categorize the results into two main groups: methods without token constraints (both closed and open-source models) and those operating under a 1.5k token limit. We mainly focused on models with around 7B parameters and include methods using an OCR system such as Pali3 [10] ScreenAI [8], DocFormerV2 [7], and GRAM [14]." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.765, + 0.483, + 0.903 + ], + "angle": 0, + "content": "To evaluate DocVLM's effectiveness under token constraints, we integrated it with three baseline models: LlaVA-OneVision [31], InternVL2 [18], and Qwen2-VL [53], each configured to operate within the 1.5k token limit. For LlaVA-OneVision, we utilized the minimal visual token configuration (single visual features tile). InternVL2 was tested with both single-tile (256 tokens) and four-tile (1280 tokens) configurations, while Qwen2-VL was evaluated with 256 and 512 visual tokens, corresponding to image" + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.639, + 0.827, + 0.655 + ], + "angle": 0, + "content": "sizes of \\(448 \\times 448\\) and \\(616 \\times 616\\) respectively." + }, + { + "type": "text", + "bbox": [ + 0.511, + 0.657, + 0.907, + 0.838 + ], + "angle": 0, + "content": "Under the practical 1.5k token constraint, incorporating DocVLM with each of these baseline models yields substantial and consistent improvements. Notably, these improvements persist even in looser token regimes, such as InternVL2 with 1280 visual tokens and Qwen2-VL with 576 visual tokens. Within this constraint, our Qwen2-VL variant with DocVLM, using just 576 tokens, achieves state-of-the-art performance across all benchmarks: \\(92.8\\%\\) on DocVQA, \\(82.8\\%\\) on TextVQA, \\(79.8\\%\\) on ST-VQA, \\(66.8\\%\\) on InfoVQA, and a CIDEr score of 150.4 on TextCAPS. As detailed in the Appendix, DocVLM achieves these results with significant computational advantages." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.84, + 0.909, + 0.903 + ], + "angle": 0, + "content": "DocVLM also demonstrates exceptional capability in handling multipage documents. Using the same 576-token configuration described earlier, it achieves \\(84.5\\%\\) accuracy on MP-DocVQA, surpassing previous state-of-the-art re" + }, + { + "type": "page_number", + "bbox": [ + 0.479, + 0.945, + 0.521, + 0.958 + ], + "angle": 0, + "content": "29009" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.099, + 0.09, + 0.488, + 0.272 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.096, + 0.287, + 0.488, + 0.468 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.493, + 0.087, + 0.902, + 0.287 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.494, + 0.291, + 0.902, + 0.469 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.09, + 0.478, + 0.908, + 0.522 + ], + "angle": 0, + "content": "Figure 3. Qualitative Results. Representative examples of DocVLM's performance across diverse document formats, from dense text to infographics and scene text. Our model successfully handles complex layouts, dense content, and presents instruction-following capabilities without explicit training on such datasets. Each example includes an image-instruction pair with baseline and DocVLM predictions." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.536, + 0.484, + 0.598 + ], + "angle": 0, + "content": "sults. Furthermore, this setup shows robust generalization capabilities with a zero-shot performance of \\(47.4\\%\\) accuracy on the DUDE dataset, despite not being specifically trained for multipage document processing." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.625, + 0.274, + 0.641 + ], + "angle": 0, + "content": "4.3. Qualitative Results" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.653, + 0.484, + 0.759 + ], + "angle": 0, + "content": "Figure 3 illustrates DocVLM's enhanced capabilities through representative examples that demonstrate three key strengths of our method: (1) improved reading comprehension in complex document layouts, (2) effective handling of dense textual content despite using compressed representations, and (3) preserved and enhanced instruction-following capabilities." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.765, + 0.484, + 0.903 + ], + "angle": 0, + "content": "The examples span diverse document types, from dense text documents to infographics and scene text, showcasing DocVLM's versatility. In the infographic example, DocVLM not only preserves but enhances the base model's instruction-following capabilities, despite our OCR component not being explicitly trained on instruction-following datasets like [49]. This demonstrates that our compression mechanism successfully retains crucial textual and layout information while significantly reducing token usage." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.535, + 0.8, + 0.552 + ], + "angle": 0, + "content": "4.4. Scaling to Multipage Documents" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.559, + 0.907, + 0.665 + ], + "angle": 0, + "content": "Building on the promising multipage results in Tab. 1, we conduct an in-depth analysis of different DocVLM configurations for multipage document understanding. This analysis focuses on the Qwen2-VL base model, tested on the MP-DocVQA dataset with documents up to 20 pages long – a scale that most other state-of-the-art methods struggle to handle due to token limitations." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.668, + 0.907, + 0.699 + ], + "angle": 0, + "content": "Table 2 compares four OCR integration strategies used during inference in multipage scenarios:" + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.7, + 0.858, + 0.715 + ], + "angle": 0, + "content": "- Baseline: vision-only input (no additional tokens)" + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.716, + 0.887, + 0.731 + ], + "angle": 0, + "content": "- Direct OCR word insertion: up to 800 tokens per page" + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.731, + 0.786, + 0.745 + ], + "angle": 0, + "content": "- Global OCR encoding: 64 tokens total" + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.746, + 0.833, + 0.761 + ], + "angle": 0, + "content": "- Page-wise OCR encoding: 64 tokens per page" + }, + { + "type": "list", + "bbox": [ + 0.513, + 0.7, + 0.887, + 0.761 + ], + "angle": 0, + "content": null + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.763, + 0.907, + 0.883 + ], + "angle": 0, + "content": "Results indicate consistent improvements over the baseline across all image resolutions (256, 512, and 1024 tokens), with only minimal additional token usage – just 64 tokens for the entire document in the global encoding case. Notably, at 256 visual tokens per page, both page-wise encoding \\((82.4\\%)\\) and global encoding \\((81.7\\%)\\) outperform direct OCR word insertion \\((79.1\\%)\\) while using significantly fewer tokens." + }, + { + "type": "text", + "bbox": [ + 0.532, + 0.886, + 0.905, + 0.902 + ], + "angle": 0, + "content": "Our best configuration achieves state-of-the-art perfor" + }, + { + "type": "page_number", + "bbox": [ + 0.479, + 0.945, + 0.521, + 0.958 + ], + "angle": 0, + "content": "29010" + } + ], + [ + { + "type": "table", + "bbox": [ + 0.095, + 0.089, + 0.482, + 0.249 + ], + "angle": 0, + "content": "
MethodLLM OCR InputImage Tok.OCR Tok.ANLS
DocOwl2 [27]-324 × pg-69.4
GRAMC-Former [14]-100 × pg25677.6
GRAM [14]-100 × pg800 × pg80.3
Qwen2-VL-256 × pg-73.0
OCR Words256 × pg800 × pg79.1
DocVLMQwen2-VLGlobal Encoding256 × pg6481.7
DocVLMQwen2-VLPage-wise Encoding256 × pg64 × pg82.4
Qwen2-VL-512 × pg-82.1
DocVLMQwen2-VLGlobal Encoding512 × pg6484.5
DocVLMQwen2-VLPage-wise Encoding512 × pg64 × pg85.2
Qwen2-VL-1024 × pg-85.2
DocVLMQwen2-VLGlobal Encoding1024 × pg6486.3
DocVLMQwen2-VLPage-wise Encoding1024 × pg64 × pg86.3
" + }, + { + "type": "table_caption", + "bbox": [ + 0.09, + 0.26, + 0.483, + 0.344 + ], + "angle": 0, + "content": "Table 2. Extension to Multipage. Comparison on MP-DocVQA of approaches for incorporating OCR information in multipage document understanding. Both DocVLM multipage extension strategies: global encoding (64 tokens per document) and pagewise encoding (64 tokens per page), outperform previous state-of-the-art methods, notably, without any explicit multipage training." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.36, + 0.483, + 0.435 + ], + "angle": 0, + "content": "mance of \\(86.3\\%\\) ANLS, significantly outperforming specialized multipage models like GRAM \\((80.3\\%)\\). This is particularly impressive considering that DocVLM was trained exclusively on single-page inputs, demonstrating strong zero-shot generalization to multipage scenarios." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.437, + 0.483, + 0.573 + ], + "angle": 0, + "content": "The comparison between encoding strategies reveals that page-wise encoding consistently outperforms global encoding at lower visual token counts, providing a \\(+0.7\\%\\) improvement for both 256 and 512 image tokens per page. This advantage diminishes at 1024 tokens where both achieve identical performance (86.3% ANLS). Remarkably, DocVLM with page-wise encoding matches or even outperforms the baseline Qwen2-VL using twice as many visual tokens, highlighting the efficiency of our approach." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.588, + 0.245, + 0.605 + ], + "angle": 0, + "content": "5. Ablation Study" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.613, + 0.483, + 0.763 + ], + "angle": 0, + "content": "Impact of OCR Encoding Strategies To evaluate the impact of OCR encoding compression, we compare three strategies for integrating OCR information: (1) inserting raw OCR words in the original VLM, (2) using DocVLM uncompressed OCR encodings, and (3) DocVLM compressed OCR encodings with 64 learned queries. We evaluate these approaches on the DocVQA test set using three representative model configurations: LLaVA-OneVision with 1.5K visual tokens, and both InternVL2 and Qwen2-VL with 256 visual tokens each." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.765, + 0.483, + 0.902 + ], + "angle": 0, + "content": "Results in Tab. 3 demonstrate that DocVLM's OCR encodings significantly outperform raw OCR words across all three models while maintaining the same token count. Notably, our compressed encoding approach, using just 64 tokens instead of 800 OCR tokens, preserves most of these improvements while drastically reducing sequence length. This efficient compression enables a more favorable allocation of the token budget, allowing models to dedicate more tokens to visual processing without compromising OCR ef" + }, + { + "type": "table", + "bbox": [ + 0.518, + 0.089, + 0.905, + 0.153 + ], + "angle": 0, + "content": "
LLM OCR InputOCR Tok.LLaVA-OV 1.5KInternVL2 256Qwen2-VL 256
OCR Words80085.884.489.1
OCR Encoding80089.489.291.9
64 Compressed Encoding6488.486.691.2
" + }, + { + "type": "table_caption", + "bbox": [ + 0.513, + 0.163, + 0.905, + 0.205 + ], + "angle": 0, + "content": "Table 3. OCR Encoding Strategies. DocVQA results for inserting OCR information using (1) OCR words (baseline), (2) uncompressed OCR encodings, and (3) 64 compressed OCR encodings." + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.23, + 0.905, + 0.276 + ], + "angle": 0, + "content": "fectiveness. The results validate that DocVLM's compression strategy successfully balances performance with computational efficiency, a key factor for practical applications." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.318, + 0.907, + 0.456 + ], + "angle": 0, + "content": "Balancing Vision and OCR Token Allocation Modern VLMs employ various mechanisms to reduce visual token count, creating an inherent trade-off between computational efficiency and model performance, as discussed in Sec. 3.1. We investigate how DocVLM can improve this trade-off by comparing four configurations: (1) baseline VLM without OCR, (2) direct OCR word insertion, (3) DocVLM with uncompressed OCR encodings, and (4) DocVLM with 64 compressed learned queries." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.459, + 0.909, + 0.595 + ], + "angle": 0, + "content": "Figure 4 presents the performance scores on DocVQA validation (left y-axis) and total token counts (right y-axis) for three VLM architectures, showing how these metrics vary across different visual token allocations. Each model employs a distinct token reduction approach: LLaVA-OneVision controls token count through feature downsampling (AnyRes Max), InternVL2 limits the number of processed image tiles (Dynamic Max Batch), and Qwen2-VL adjusts image resolution to constrain token count." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.6, + 0.909, + 0.901 + ], + "angle": 0, + "content": "Our analysis reveals that integrating OCR information, regardless of the method used, consistently improves performance across all models, with particularly pronounced gains in low visual token regimes. However, uncompressed OCR integration methods, whether through direct word insertion or uncompressed DocVLM encodings, require 800 tokens – a significant overhead that could otherwise be allocated to visual processing. For instance, allocating 128 tokens for visual and 800 for OCR in Qwen2-VL achieves \\(84.3\\%\\) using OCR words and \\(90.1\\%\\) using uncompressed encodings. In contrast, using 896 pure visual tokens reaches \\(92.4\\%\\), demonstrating the potential benefit of allocating more tokens to visual processing. DocVLM's compression mechanism provides a superior option by requiring only 64 tokens for OCR information while maintaining strong performance. In the above example, our approach using fewer tokens, allocating 768 visual tokens and 64 OCR tokens, reaches \\(93.0\\%\\), outperforming the \\(90.1\\%\\) obtained with the uncompressed encodings, highlighting DocVLM's effective balance between visual and OCR tokens." + }, + { + "type": "page_number", + "bbox": [ + 0.48, + 0.945, + 0.518, + 0.957 + ], + "angle": 0, + "content": "29011" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.097, + 0.09, + 0.362, + 0.231 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.362, + 0.09, + 0.631, + 0.231 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.631, + 0.091, + 0.902, + 0.231 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.258, + 0.234, + 0.74, + 0.261 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.09, + 0.269, + 0.908, + 0.327 + ], + "angle": 0, + "content": "Figure 4. Balancing Performance and Compute. Analysis of model performance (lines, left y-axis) and token usage (bars, right y-axis) as a function of visual token allocation. Each model employs its inherent token control strategy: AnyRes max for feature downsampling (LLaVA One-Vision), dynamic max tiles (InternVL2), and max image tokens for resolution control (Qwen2-VL). The results highlight that DocVLM consistently improves performance with minimal overhead (64 tokens), offering an efficient OCR-visual token allocation." + }, + { + "type": "image", + "bbox": [ + 0.114, + 0.34, + 0.462, + 0.507 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.09, + 0.521, + 0.486, + 0.579 + ], + "angle": 0, + "content": "Figure 5. Compression Levels. DocVQA validation results for DocVLM integrated with Qwen2-VL across varying OCR and image token budgets. \"0\" represents the baseline, while \"Full\" indicates uncompressed encodings." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.594, + 0.484, + 0.791 + ], + "angle": 0, + "content": "Compression Levels We deepen our analysis of the OCR-visual token trade-off by examining different compression levels in DocVLM integrated with Qwen2-VL. Fig. 5 presents ANLS scores across various combinations of visual tokens (128-1024) and learned queries (16-256), including baselines without OCR and with uncompressed encoding. In low visual token regimes, increasing the number of learned queries yields substantial improvements, validating our compression mechanism's effectiveness in capturing relevant OCR information. Notably, even with just 16 learned queries, DocVLM outperforms the baseline across all visual token configurations, offering strong performance with minimal computational overhead." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.81, + 0.484, + 0.903 + ], + "angle": 0, + "content": "Training Stages Tab. 4 illustrates the impact of the vision alignment stage on ANLS performance for DocVLM with the LLaVA-OneVision base model on the DocVQA test set, showing results for varying numbers of learned queries (16 to 128) and the uncompressed case. Our two-stage training process initially trains OCR modality components with-" + }, + { + "type": "table", + "bbox": [ + 0.518, + 0.339, + 0.905, + 0.416 + ], + "angle": 0, + "content": "
Training PhasesCompressed Enc.OCR Enc.
1664128800
Stage I: OCR-LLM Alignment81.785.886.389.4
+ Stage II: Vision Alignment87.988.488.490.1
Δ+6.2+2.6+2.1+0.7
" + }, + { + "type": "table_caption", + "bbox": [ + 0.513, + 0.426, + 0.907, + 0.469 + ], + "angle": 0, + "content": "Table 4. Training Phases. Second-stage training consistently improves DocVQA performance, both with compressed DocVLM tokens and full OCR encoding." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.484, + 0.907, + 0.651 + ], + "angle": 0, + "content": "out image input, forcing reliance on OCR data alone, before reintroducing the image in the vision alignment stage to adapt learned queries alongside visual information. The results reveal that vision alignment significantly boosts performance, especially with fewer learned queries: for instance, with 16 learned queries, there's an improvement of +6.2, compared to +0.7 in the uncompressed case. Notably, after vision alignment, DocVLM with only 16 learned queries outperforms the baseline of OCR words (from Table 3). These findings underscore the effectiveness of our two-stage training method." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.665, + 0.642, + 0.681 + ], + "angle": 0, + "content": "6. Conclusions" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.689, + 0.909, + 0.903 + ], + "angle": 0, + "content": "Our results demonstrate that DocVLM can be effectively integrated into various VLMs to enhance their document reading capabilities while significantly reducing their dependency on extensive vision tokens. The key takeaway is that in token-constrained scenarios, allocating a small portion of tokens to OCR information consistently yields better results than using those tokens solely for visual processing. The effectiveness of our compression mechanism extends beyond single-page documents, as evidenced by achieving state-of-the-art results on MP-DocVQA using the same 64 tokens to represent multiple pages. These results establish DocVLM as a practical solution for enhancing document understanding in real-world applications where computational efficiency is crucial." + }, + { + "type": "page_number", + "bbox": [ + 0.479, + 0.945, + 0.521, + 0.958 + ], + "angle": 0, + "content": "29012" + } + ], + [ + { + "type": "title", + "bbox": [ + 0.093, + 0.09, + 0.188, + 0.106 + ], + "angle": 0, + "content": "References" + }, + { + "type": "ref_text", + "bbox": [ + 0.101, + 0.115, + 0.484, + 0.197 + ], + "angle": 0, + "content": "[1] Aviad Aberdam, Ron Litman, Shahar Tsiper, Oron Anschel, Ron Slossberg, Shai Mazor, R Manmatha, and Pietro Perona. Sequence-to-sequence contrastive learning for text recognition. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 15302-15312, 2021. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.101, + 0.2, + 0.482, + 0.241 + ], + "angle": 0, + "content": "[2] Aviad Aberdam, Roy Ganz, Shai Mazor, and Ron Litman. Multimodal semi-supervised learning for text recognition. arXiv preprint arXiv:2205.03873, 2022." + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.243, + 0.483, + 0.311 + ], + "angle": 0, + "content": "[3] Aviad Aberdam, David Bensaïd, Alona Golts, Roy Ganz, Oren Nuriel, Royee Tichauer, Shai Mazor, and Ron Litman. Clipter: Looking at the bigger picture in scene text recognition. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 21706-21717, 2023." + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.313, + 0.483, + 0.396 + ], + "angle": 0, + "content": "[4] Ofir Abramovich, Niv Nayman, Sharon Fogel, Inbal Lavi, Ron Litman, Shahar Tsiper, Royee Tichauer, Srikar Appalaraju, Shai Mazor, and R Manmatha. Visfocus: Prompt-guided vision encoders forOCR-free dense document understanding. In European Conference on Computer Vision, pages 241-259. Springer, 2025. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.398, + 0.483, + 0.478 + ], + "angle": 0, + "content": "[5] Jean-Baptiste Alayrac, Jeff Donahue, Pauline Luc, Antoine Miech, Iain Barr, Yana Hasson, Karel Lenc, Arthur Mensch, Katherine Millican, Malcolm Reynolds, et al. Flamingo: a visual language model for few-shot learning. Advances in neural information processing systems, 35:23716-23736, 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.481, + 0.484, + 0.549 + ], + "angle": 0, + "content": "[6] Srikar Appalaraju, Bhavan Jasani, Bhargava Urala Kota, Yusheng Xie, and R Manmatha. Docformer: End-to-end transformer for document understanding. In Proceedings of the IEEE/CVF international conference on computer vision, pages 993-1003, 2021. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.551, + 0.483, + 0.619 + ], + "angle": 0, + "content": "[7] Srikar Appalaraju, Peng Tang, Qi Dong, Nishant Sankaran, Yichu Zhou, and R Manmatha. Docformerv2: Local features for document understanding. In Proceedings of the AAAI Conference on Artificial Intelligence, pages 709-718, 2024. 3, 5, 12" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.622, + 0.483, + 0.69 + ], + "angle": 0, + "content": "[8] Gilles Baechler, Srinivas Sunkara, Maria Wang, Fedir Zubach, Hassan Mansoor, Vincent Etter, Victor Cärbune, Jason Lin, Jindong Chen, and Abhanshu Sharma. Screenai: A vision-language model for ui and infographics understanding. arXiv preprint arXiv:2402.04615, 2024. 1, 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.692, + 0.483, + 0.761 + ], + "angle": 0, + "content": "[9] Jinze Bai, Shuai Bai, Shusheng Yang, Shijie Wang, Sinan Tan, Peng Wang, Junyang Lin, Chang Zhou, and Jingren Zhou. Qwen-vl: A versatile vision-language model for understanding, localization, text reading, and beyond. arXiv preprint arXiv:2308.12966, 1(2):3, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.763, + 0.483, + 0.831 + ], + "angle": 0, + "content": "[10] Lucas Beyer, Andreas Steiner, André Susano Pinto, Alexander Kolesnikov, Xiao Wang, Daniel Salz, Maxim Neumann, Ibrahim Alabdulmohsin, Michael Tschannen, Emanuele Bugliarello, et al. Paligemma: A versatile 3b vlm for transfer. arXiv preprint arXiv:2407.07726, 2024. 1, 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.832, + 0.483, + 0.901 + ], + "angle": 0, + "content": "[11] Ali Furkan Biten, Ruben Tito, Andres Mafla, Lluis Gomez, Marçal Rusinol, Ernest Valveny, CV Jawahar, and Dimosthenis Karatzas. Scene text visual question answering. In Proceedings of the IEEE/CVF international conference on computer vision, pages 4291-4301, 2019. 4, 12" + }, + { + "type": "list", + "bbox": [ + 0.094, + 0.115, + 0.484, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.093, + 0.905, + 0.161 + ], + "angle": 0, + "content": "[12] Ali Furkan Biten, Ron Litman, Yusheng Xie, Srikar Appalaraju, and R Manmatha. Latr: Layout-aware transformer for scene-text vqa. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 16548-16558, 2022. 2, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.163, + 0.905, + 0.217 + ], + "angle": 0, + "content": "[13] Ali Furkan Biten, Ruben Tito, Lluis Gomez, Ernest Valveny, and Dimosthenis Karatzas. Ocr-idl: Ocr annotations for industry document library dataset. arXiv preprint arXiv:2202.12985, 2022. 12" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.219, + 0.905, + 0.3 + ], + "angle": 0, + "content": "[14] Tsachi Blau, Sharon Fogel, Roi Ronen, Alona Golts, Roy Ganz, Elad Ben Avraham, Aviad Aberdam, Shahar Tsiper, and Ron Litman. Gram: Global reasoning for multi-page vqa. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 15598-15607, 2024. 2, 3, 5, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.302, + 0.904, + 0.357 + ], + "angle": 0, + "content": "[15] Xinlei Chen, Hao Fang, Tsung-Yi Lin, Ramakrishna Vedantam, Saurabh Gupta, Piotr Dólar, and C Lawrence Zitnick. Microsoft coco captions: Data collection and evaluation server. arXiv preprint arXiv:1504.00325, 2015. 4, 12" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.359, + 0.905, + 0.427 + ], + "angle": 0, + "content": "[16] Xi Chen, Josip Djolonga, Piotr Padlewski, Basil Mustafa, Soravit Changpinyo, Jialin Wu, Carlos Riquelme Ruiz, Sebastian Goodman, Xiao Wang, Yi Tay, et al. Pali-x: On scaling up a multilingual vision and language model. arXiv preprint arXiv:2305.18565, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.428, + 0.905, + 0.496 + ], + "angle": 0, + "content": "[17] Xi Chen, Xiao Wang, Lucas Beyer, Alexander Kolesnikov, Jialin Wu, Paul Voigtlaender, Basil Mustafa, Sebastian Goodman, Ibrahim Alabdulmohsin, Piotr Padlewski, et al. Pali-3 vision language models: Smaller, faster, stronger. arXiv preprint arXiv:2310.09199, 2023. 1, 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.498, + 0.905, + 0.579 + ], + "angle": 0, + "content": "[18] Zhe Chen, Weiyun Wang, Hao Tian, Shenglong Ye, Zhangwei Gao, Erfei Cui, Wenwen Tong, Kongzhi Hu, Jiapeng Luo, Zheng Ma, et al. Internvl2: Better than the best—expanding performance boundaries of open-source multimodal models with the progressive scaling strategy, 2024. 2, 3, 4, 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.581, + 0.905, + 0.65 + ], + "angle": 0, + "content": "[19] Wenliang Dai, Junnan Li, Dongxu Li, Anthony Meng Huat Tiong, Junqi Zhao, Weisheng Wang, Boyang Albert Li, Pascale Fung, and Steven C. H. Hoi. Instructclip: Towards general-purpose vision-language models with instruction tuning. ArXiv, abs/2305.06500, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.651, + 0.905, + 0.733 + ], + "angle": 0, + "content": "[20] Xiaoyi Dong, Pan Zhang, Yuhang Zang, Yuhang Cao, Bin Wang, Linke Ouyang, Songyang Zhang, Haodong Duan, Wenwei Zhang, Yining Li, et al. Internlm-xcomposer2-4khd: A pioneering large vision-language model handling resolutions from 336 pixels to 4k hd. arXiv preprint arXiv:2404.06512, 2024. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.735, + 0.905, + 0.789 + ], + "angle": 0, + "content": "[21] Masato Fujitake. Dtrocr: Decoder-only transformer for optical character recognition. In Proceedings of the IEEE/CVF Winter Conference on Applications of Computer Vision, pages 8025-8035, 2024. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.79, + 0.905, + 0.845 + ], + "angle": 0, + "content": "[22] Roy Ganz, Oren Nuriel, Aviad Aberdam, Yair Kittenplon, Shai Mazor, and Ron Litman. Towards models that can see and read. In Proceedings of the IEEE/CVF international conference on computer vision, pages 21718-21728, 2023. 2, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.847, + 0.905, + 0.901 + ], + "angle": 0, + "content": "[23] Roy Ganz, Yair Kittenplon, Aviad Aberdam, Elad Ben Avraham, Oren Nuriel, Shai Mazor, and Ron Litman. Question aware vision transformer for multimodal reasoning. In Proceedings of the IEEE/CVF Conference on Computer Vision" + }, + { + "type": "list", + "bbox": [ + 0.094, + 0.093, + 0.905, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.48, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "29013" + } + ], + [ + { + "type": "ref_text", + "bbox": [ + 0.125, + 0.092, + 0.482, + 0.119 + ], + "angle": 0, + "content": "and Pattern Recognition (CVPR), pages 13861-13871, 2024. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.121, + 0.483, + 0.189 + ], + "angle": 0, + "content": "[24] Yash Goyal, Tejas Khot, Douglas Summers-Stay, Dhruv Bartra, and Devi Parikh. Making the V in VQA matter: Elevating the role of image understanding in Visual Question Answering. In Conference on Computer Vision and Pattern Recognition (CVPR), 2017. 4, 12" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.19, + 0.482, + 0.245 + ], + "angle": 0, + "content": "[25] Jean-Philippe Thiran Guillaume Jaume, Hazim Kemal Ekenel. Funsd: A dataset for form understanding in noisy scanned documents. In Accepted to ICDAR-OST, 2019. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.247, + 0.482, + 0.315 + ], + "angle": 0, + "content": "[26] Benjamin Hsu, Xiaoyu Liu, Huayang Li, Yoshinari Fujinuma, Maria Nadejde, Xing Niu, Yair Kittenplon, Ron Litman, and Raghavendra Pappagari. M3t: A new benchmark dataset for multi-modal document-level machine translation. arXiv preprint arXiv:2406.08255, 2024. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.316, + 0.482, + 0.384 + ], + "angle": 0, + "content": "[27] Anwen Hu, Haiyang Xu, Liang Zhang, Jiabo Ye, Ming Yan, Ji Zhang, Qin Jin, Fei Huang, and Jingren Zhou. mplug-docowl2: High-resolution compressing forocr-free multi-page document understanding. arXiv preprint arXiv:2409.03420, 2024. 2, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.385, + 0.482, + 0.453 + ], + "angle": 0, + "content": "[28] Taeho Kil, Seonghyeon Kim, Sukmin Seo, Yoonsik Kim, and Daehe Kim. Towards unified scene text spotting based on sequence generation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 15223-15232, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.455, + 0.482, + 0.523 + ], + "angle": 0, + "content": "[29] Geewook Kim, Teakgyu Hong, Moonbin Yim, Jinyoung Park, Jinyeong Yim, Wonseok Hwang, Sangdoo Yun, Dongyoon Han, and Seunghyun Park. Donut: Document understanding transformer withoutOCR.arXiv preprint arXiv:2111.15664, 7(15):2, 2021. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.525, + 0.482, + 0.566 + ], + "angle": 0, + "content": "[30] Hugo Laurençon, Léo Tronchon, Matthieu Cord, and Victor Sanh. What matters when building vision-language models? arXiv preprint arXiv:2405.02246, 2024. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.567, + 0.482, + 0.621 + ], + "angle": 0, + "content": "[31] Bo Li, Yuanhan Zhang, Dong Guo, Renrui Zhang, Feng Li, Hao Zhang, Kaichen Zhang, Yanwei Li, Ziwei Liu, and Chunyuan Li. Llava-onevision: Easy visual task transfer. arXiv preprint arXiv:2408.03326, 2024. 1, 2, 4, 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.622, + 0.482, + 0.69 + ], + "angle": 0, + "content": "[32] Junnan Li, Dongxu Li, Silvio Savarese, and Steven Hoi. Blip-2: Bootstrapping language-image pre-training with frozen image encoders and large language models. In International conference on machine learning, pages 19730-19742. PMLR, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.692, + 0.482, + 0.746 + ], + "angle": 0, + "content": "[33] Wentong Li, Yuqian Yuan, Jian Liu, Dongqi Tang, Song Wang, Jianke Zhu, and Lei Zhang. Tokenpacker: Efficient visual projector for multimodal llm. arXiv preprint arXiv:2407.02392, 2024. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.748, + 0.483, + 0.816 + ], + "angle": 0, + "content": "[34] Ron Litman, Oron Anschel, Shahar Tsiper, Roeit Litman, Shai Mazor, and R Manmatha. Scatter: selective context attentional scene text recognizer. In proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 11962-11972, 2020. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.818, + 0.482, + 0.859 + ], + "angle": 0, + "content": "[35] Haotian Liu, Chunyuan Li, Yuheng Li, Bo Li, Yuanhan Zhang, Sheng Shen, and Yong Jae Lee. Llava next: Improved reasoning,OCR, and world knowledge, 2024. 2, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.86, + 0.482, + 0.901 + ], + "angle": 0, + "content": "[36] Haotian Liu, Chunyuan Li, Qingyang Wu, and Yong Jae Lee. Visual instruction tuning. Advances in neural information processing systems, 36, 2024. 2" + }, + { + "type": "list", + "bbox": [ + 0.093, + 0.092, + 0.483, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.092, + 0.905, + 0.147 + ], + "angle": 0, + "content": "[37] Yuliang Liu, Zhang Li, Mingxin Huang, Biao Yang, Wenwen Yu, Chunyuan Li, Xucheng Yin, Cheng lin Liu, Lianwen Jin, and Xiang Bai. Ocrbench: On the hidden mystery ofOCR in large multimodal models, 2024. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.149, + 0.905, + 0.204 + ], + "angle": 0, + "content": "[38] Jiahao Lyu, Wei Wang, Dongbao Yang, Jinwen Zhong, and Yu Zhou. Arbitrary reading order scene text spotter with local semantics guidance. arXiv preprint arXiv:2412.10159, 2024. 14" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.206, + 0.905, + 0.262 + ], + "angle": 0, + "content": "[39] Ahmed Masry, Do Xuan Long, Jia Qing Tan, Shafiq Joty, and Enamul Hoque. Chartqa: A benchmark for question answering about charts with visual and logical reasoning. arXiv preprint arXiv:2203.10244, 2022. 1, 4, 12" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.264, + 0.907, + 0.319 + ], + "angle": 0, + "content": "[40] Minesh Mathew, Dimosthenis Karatzas, and CV Jawahar. Docvqa: A dataset for vqa on document images. In Proceedings of the IEEE/CVF winter conference on applications of computer vision, pages 2200-2209, 2021. 4, 12, 13" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.321, + 0.905, + 0.388 + ], + "angle": 0, + "content": "[41] Minesh Mathew, Viraj Bagal, Ruben Tito, Dimosthenis Karatzas, Ernest Valveny, and CV Jawahar. Infographicvqa. In Proceedings of the IEEE/CVF Winter Conference on Applications of Computer Vision, pages 1697-1706, 2022. 1, 4, 12, 13" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.391, + 0.905, + 0.46 + ], + "angle": 0, + "content": "[42] Anand Mishra, Shashank Shekhar, Ajeet Kumar Singh, and Anirban Chakraborty. Ocr-vqa: Visual question answering by reading text in images. In 2019 international conference on document analysis and recognition (ICDAR), pages 947-952. IEEE, 2019. 4, 12" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.462, + 0.905, + 0.517 + ], + "angle": 0, + "content": "[43] Oren Nuriel, Sharon Fogel, and Ron Litman. Textadain: Paying attention to shortcut learning in text recognizers. In European Conference on Computer Vision, pages 427-445. Springer, 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.519, + 0.905, + 0.574 + ], + "angle": 0, + "content": "[44] Zhiliang Peng, Wenhui Wang, Li Dong, Yaru Hao, Shaohan Huang, Shuming Ma, and Furu Wei. Kosmos-2: Grounding multimodal large language models to the world. arXiv preprint arXiv:2306.14824, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.576, + 0.905, + 0.645 + ], + "angle": 0, + "content": "[45] Colin Raffel, Noam Shazeer, Adam Roberts, Katherine Lee, Sharan Narang, Michael Matena, Yanqi Zhou, Wei Li, and Peter J Liu. Exploring the limits of transfer learning with a unified text-to-text transformer. Journal of machine learning research, 21(140):1-67, 2020. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.647, + 0.905, + 0.701 + ], + "angle": 0, + "content": "[46] Roi Ronen, Shahar Tsiper, Oron Anschel, Inbal Lavi, Amir Markovitz, and R Manmatha. Glass: Global to local attention for scene-text spotting. arXiv preprint arXiv:2208.03364, 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.703, + 0.905, + 0.785 + ], + "angle": 0, + "content": "[47] Oleksii Sidorov, Ronghang Hu, Marcus Rohrbach, and Amanpreet Singh. Textcaps: a dataset for image captioning with reading comprehension. In Computer Vision-ECCV 2020: 16th European Conference, Glasgow, UK, August 23-28, 2020, Proceedings, Part II 16, pages 742-758. Springer, 2020. 4, 12" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.788, + 0.905, + 0.857 + ], + "angle": 0, + "content": "[48] Amanpreet Singh, Vivek Natarajan, Meet Shah, Yu Jiang, Xinlei Chen, Dhruv Batra, Devi Parikh, and Marcus Rohrbach. Towards vqa models that can read. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 8317-8326, 2019. 4, 12" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.859, + 0.905, + 0.901 + ], + "angle": 0, + "content": "[49] Ryota Tanaka, Taichi Iki, Kyosuke Nishida, Kuniko Saito, and Jun Suzuki. Instructdoc: A dataset for zero-shot generalization of visual document understanding with instructions." + }, + { + "type": "list", + "bbox": [ + 0.517, + 0.092, + 0.907, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.48, + 0.945, + 0.52, + 0.957 + ], + "angle": 0, + "content": "29014" + } + ], + [ + { + "type": "ref_text", + "bbox": [ + 0.125, + 0.092, + 0.482, + 0.12 + ], + "angle": 0, + "content": "In Proceedings of the AAAI Conference on Artificial Intelligence, pages 19071-19079, 2024. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.122, + 0.483, + 0.164 + ], + "angle": 0, + "content": "[50] Ruben Tito, Dimosthenis Karatzas, and Ernest Valveny. Hierarchical multimodal transformers for multipage docvqa. Pattern Recognition, 144:109834, 2023. 1, 2, 4, 12" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.165, + 0.483, + 0.262 + ], + "angle": 0, + "content": "[51] Jordy Van Landeghem, Ruben Tito, Łukasz Borchmann, Michal Pietruszka, Pawel Joziak, Rafal Powalski, Dawid Jurkiewicz, Mickaël Coustaty, Bertrand Anckaert, Ernest Valveny, et al. Document understanding dataset and evaluation (dude). In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 19528-19540, 2023. 2, 4, 12" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.264, + 0.483, + 0.333 + ], + "angle": 0, + "content": "[52] Dongsheng Wang, Natraj Raman, Mathieu Sibue, Zhiqiang Ma, Petr Babkin, Simerjot Kaur, Yulong Pei, Armineh Nourbakhsh, and Xiaomo Liu. Docllm: A layout-aware generative language model for multimodal document understanding. arXiv preprint arXiv:2401.00908, 2023. 2, 13" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.334, + 0.483, + 0.404 + ], + "angle": 0, + "content": "[53] Peng Wang, Shuai Bai, Sinan Tan, Shijie Wang, Zhihao Fan, Jinze Bai, Keqin Chen, Xuejing Liu, Jialin Wang, Wenbin Ge, et al. Qwen2-vl: Enhancing vision-language model's perception of the world at any resolution. arXiv preprint arXiv:2409.12191, 2024. 1, 2, 3, 4, 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.406, + 0.483, + 0.462 + ], + "angle": 0, + "content": "[54] Wenjin Wang, Yunhao Li, Yixin Ou, and Yin Zhang. Layout and task aware instruction prompt for zero-shot document image question answering. arXiv preprint arXiv:2306.00526, 2023. 1, 2, 13" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.463, + 0.483, + 0.533 + ], + "angle": 0, + "content": "[55] Yiheng Xu, Minghao Li, Lei Cui, Shaohan Huang, Furu Wei, and Ming Zhou. Layoutm: Pre-training of text and layout for document image understanding. In Proceedings of the 26th ACM SIGKDD international conference on knowledge discovery & data mining, pages 1192-1200, 2020. 13" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.534, + 0.483, + 0.603 + ], + "angle": 0, + "content": "[56] Jiabo Ye, Anwen Hu, Haiyang Xu, Qinghao Ye, Ming Yan, Guohai Xu, Chenliang Li, Junfeng Tian, Qi Qian, Ji Zhang, et al. Ureader: UniversalOCR-free visually-situated language understanding with multimodal large language model. arXiv preprint arXiv:2310.05126, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.605, + 0.483, + 0.675 + ], + "angle": 0, + "content": "[57] Maoyuan Ye, Jing Zhang, Shanshan Zhao, Juhua Liu, Bo Du, and Dacheng Tao. Aptext-detr: Towards better scene text detection with dynamic points in transformer. In Proceedings of the AAAI Conference on Artificial Intelligence, pages 3241-3249, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.676, + 0.483, + 0.758 + ], + "angle": 0, + "content": "[58] Maoyuan Ye, Jing Zhang, Shanshan Zhao, Juhua Liu, Tongliang Liu, Bo Du, and Dacheng Tao. Deepsolo: Let transformer decoder with explicit points solo for text spotting. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 19348-19357, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.76, + 0.483, + 0.83 + ], + "angle": 0, + "content": "[59] Qinghao Ye, Haiyang Xu, Guohai Xu, Jiabo Ye, Ming Yan, Yiyang Zhou, Junyang Wang, Anwen Hu, Pengcheng Shi, Yaya Shi, et al. MPLug-owl: Modularization empowers large language models with multimodality. arXiv preprint arXiv:2304.14178, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.832, + 0.483, + 0.902 + ], + "angle": 0, + "content": "[60] Pan Zhang, Xiaoyi Dong, Yuhang Zang, Yuhang Cao, Rui Qian, Lin Chen, Qipeng Guo, Haodong Duan, Bin Wang, Linke Ouyang, et al. Internlm-xcomposer-2.5: A versatile large vision language model supporting long-contextual input and output. arXiv preprint arXiv:2407.03320, 2024. 1" + }, + { + "type": "list", + "bbox": [ + 0.093, + 0.092, + 0.483, + 0.902 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.516, + 0.092, + 0.906, + 0.149 + ], + "angle": 0, + "content": "[61] Deyao Zhu, Jun Chen, Xiaogian Shen, Xiang Li, and Mohamed Elhoseiny. Minigpt-4: Enhancing vision-language understanding with advanced large language models. arXiv preprint arXiv:2304.10592, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.149, + 0.906, + 0.22 + ], + "angle": 0, + "content": "[62] Fengbin Zhu, Wenqiang Lei, Fuli Feng, Chao Wang, Haozhou Zhang, and Tat-Seng Chua. Towards complex document understanding by discrete reasoning. In Proceedings of the 30th ACM International Conference on Multimedia, pages 4857-4866, 2022. 4, 12" + }, + { + "type": "list", + "bbox": [ + 0.516, + 0.092, + 0.906, + 0.22 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.48, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "29015" + } + ] +] \ No newline at end of file diff --git a/2025/DocVLM_ Make Your VLM an Efficient Reader/ec872868-b641-4059-95e8-62856f47ad39_origin.pdf b/2025/DocVLM_ Make Your VLM an Efficient Reader/ec872868-b641-4059-95e8-62856f47ad39_origin.pdf new file mode 100644 index 0000000000000000000000000000000000000000..855e671243565687b03d76c397e42b4d6ca6d38f --- /dev/null +++ b/2025/DocVLM_ Make Your VLM an Efficient Reader/ec872868-b641-4059-95e8-62856f47ad39_origin.pdf @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:661127232a1f72543ebfc96ab06b4dbd0a5a8c51147ef088df3a0a5fc7f2edec +size 5425081 diff --git a/2025/DocVLM_ Make Your VLM an Efficient Reader/full.md b/2025/DocVLM_ Make Your VLM an Efficient Reader/full.md new file mode 100644 index 0000000000000000000000000000000000000000..b7d9f2e99045a7fe9f5cd5d6d762f8359785a3fd --- /dev/null +++ b/2025/DocVLM_ Make Your VLM an Efficient Reader/full.md @@ -0,0 +1,276 @@ +![](images/30136c38bce2d4f8bb1643f13deb7e09a6f1103d93600da549ab5a82392284d5.jpg) + +# DocVLM: Make Your VLM an Efficient Reader + +Mor Shpigel Nacson*† Aviad Aberdam* Roy Ganz Elad Ben Avraham Alona Golts Yair Kittenplon Shai Mazor Ron Litman AWS AI Labs + +# Abstract + +Vision-Language Models (VLMs) excel in diverse visual tasks but face challenges in document understanding, which requires fine-grained text processing. While typical visual tasks perform well with low-resolution inputs, reading-intensive applications demand high-resolution, resulting in significant computational overhead. Using OCR-extracted text in VLM prompts partially addresses this issue but underperforms compared to full-resolution counterpart, as it lacks the complete visual context needed for optimal performance. We introduce DocVLM, a method that integrates an OCR-based modality into VLMs to enhance document processing while preserving original weights. Our approach employs an OCR encoder to capture textual content and layout, compressing these into a compact set of learned queries incorporated into the VLM. Comprehensive evaluations across leading VLMs show that DocVLM significantly reduces reliance on high-resolution images for document understanding. In limited-token regimes $(448\times 448)$ , DocVLM with 64 learned queries improves DocVQA results from $56.0\%$ to $86.6\%$ when integrated with InternVL2 and from $84.4\%$ to $91.2\%$ with Qwen2-VL. In LLaVA-OneVision, DocVLM achieves improved results while using $80\%$ less image tokens. The reduced token usage allows processing multiple pages effectively, showing impressive zero-shot results on DUDE and state-of-the-art performance on MP-DocVQA, highlighting DocVLM's potential for applications requiring high-performance and efficiency. + +# 1. Introduction + +The ability to read and interpret text within images is crucial for numerous real-world applications, particularly in document understanding. This field encompasses diverse document types, from dense-text to infographics and multipage documents [25, 39-41, 50], involving tasks that require capabilities in text comprehension, layout understanding, and + +![](images/b4a69fd84bc032d003f4ac13f7d108554fbe2c480d72ae78b69684f10e8a3e56.jpg) +Figure 1. DocVLM enhances VLMs' reading capabilities. Integrating DocVLM (solid lines) in top-performing VLMs (dashed lines) consistently improves the performance across all token budgets, frequently surpassing the baseline at higher token counts. + +visual interpretation. Despite progress in VLMs, processing such documents remains challenging [37], primarily due to the tension between resolution requirements and computational efficiency. While typical computer vision tasks achieve good performance with low-resolution inputs (typically $224 \times 224$ or $336 \times 336$ pixels), document analysis demands significantly higher resolutions, resulting in substantial computational overhead [10, 60]. + +To address these challenges, some methods incorporate OCR-extracted text directly into language model prompts [8, 17]. However, this approach typically lags behind OCR-free methods with full-resolution, as it fails to capture crucial visual context and layout information [54]. Moreover, it often produces long sequence inputs, increasing latency and computational costs, especially for dense documents. + +Alternatively, recent VLMs [20, 31, 53] have introduced specialized mechanisms to reduce visual token count, such as image resizing, tiling limits, and feature downsampling. However, when applied to document understanding, these methods result in significant performance degradation, creating an undesirable trade-off between computational efficiency and accuracy, as demonstrated in Fig. 1. These lim + +itations underscore the need for a more efficient approach that maintains high-performance document understanding while reducing computational demands. + +To overcome these limitations, we introduce DocVLM, a model-agnostic method that enhances VLMs' reading ability by utilizing OCR information effectively. Our approach employs an OCR encoder to capture both contextual and layout details from OCR-extracted text, compressing these encodings into 64 learned queries. These queries are projected and fed into the LLM part of the VLM alongside visual features. Unlike some previous methods [5, 19, 30, 32], our compression mechanism avoids the need for a separate compression module or alterations to the LLM architecture. + +We demonstrate DocVLM's effectiveness across multiple state-of-the-art VLMs: LlaVA-OneVision [31], InternVL2 [18], and Qwen2-VL [53], each employing a unique image token reduction technique. As illustrated in Figure 1, our method significantly improves performance, particularly in low input token regimes. Our experiments demonstrate consistently, across all studied VLMs and visual token budgets, that DocVLM's OCR encoder not only outperforms the baseline of inserting OCR words into VLMs but also achieves this superior performance when compressed to just 64 tokens. This dual advantage of improved performance and reduced token usage allows for better utilization of fixed token budgets, enabling the allocation of more tokens for visual processing and further enhancing overall performance. + +Importantly, this reduction in sequence length improves the scalability of our approach, enabling its application to multipage document understanding tasks without additional training. We demonstrate that DocVLM can be seamlessly extended to long-context scenarios, such as multipage DocVQA [50, 51]. While current OCR-free approaches struggle with the overwhelming amount of data in multipage documents [27], our method achieves strong zero-shot performance on DUDE and surpasses the current state-of-the-art results on MP-DocVQA (86.3% vs. 80.3%), despite not being trained on multipage data. + +# Main contributions: + +- DocVLM, a model-agnostic method that efficiently integrates OCR information into VLMs, capturing both text and layout without complex integration techniques. +- A compression mechanism that reduces OCR data into a compact set of typically 64 learned queries, significantly reducing computational overhead. +- Demonstration of DocVLM's effectiveness across different VLM architectures, LlaVA-OneVision, InternVL2, and Qwen2-VL, showing significant performance improvements in low input token regimes (448×448). +- Extension of DocVLM to long-context tasks, achieving strong zero-shot performance on DUDE and SOTA results on MP-DocVQA without multipage training data. + +# 2. Related Work + +OCR-free Document VLMs. Early VLMs[5, 9, 22, 23, 32, 36, 44, 59, 61] used relatively small image sizes (e.g., $224 \times 224$ and $336 \times 336$ ), performing well on natural-image tasks but falling short in document understanding. To address this, recent approaches enhance document understanding capabilities by operating on high-resolution images, developing various strategies to manage the resulting computational burden. Direct processing methods like Donut [29], PaLI-X [16], and Qwen2-VL [53] attempt full-resolution processing but often resort to image resizing for computational feasibility. Tile-based approaches such as UReader [56] and InternVL2 [18] improve efficiency by processing image tiles independently. Other methods, exemplified by LLaVA-1.5 [31] and LLaVA-OneVision [35], process full-scale images as tiles but downsample the resulting visual features. While these approaches offer different trade-offs, our experiments show their performance deteriorates significantly when constrained to fewer visual tokens. + +OCR-Enhanced Document Understanding. The widespread availability of efficient, open-source OCR models and cost-effective commercial solutions has driven broad adoption of OCR-based approaches in document understanding [1-4, 21, 28, 34, 43, 46, 57, 58]. Several recent works [9, 16, 17, 19] have explored integrating OCR systems with VLMs by feeding extracted text directly into the language model component. Some approaches further enhance this integration by incorporating spatial layout information [12, 14, 22, 52, 54]. While these methods reduce the computational burden of processing high-resolution images, they currently lag behind OCR-free approaches in performance. Additionally, they face challenges with lengthy input sequences, particularly in multipage settings, which can increase latency and computational costs. + +Document Representation Compression. To address efficiency challenges in processing documents, various compression techniques have been developed. For OCR-enhanced approaches, [14] proposed compressing the OCR signal in multi-page documents using a Compression Transformer, which, despite improving performance in multi-page benchmarks, introduces significant complexity to the system. In the OCR-free setting, generic VLM approaches like Q-former [32] and Resampler [5] compress visual features but struggle with text-dense images. Document-specific methods such as TokenPacker [33] and DocCompressor [27] achieve effective visual compression but show reduced performance on document understanding tasks. In contrast, rather than compressing high-resolution visual inputs, our DocVLM method operates on lower-resolution images and compresses the extensive OCR signal, including textual and layout information, into a compact set of features (typically 64). + +# 3. Our Method + +We present DocVLM, a model-agnostic approach that enhances VLMs' reading capabilities, enabling operation with lower-resolution inputs while maintaining or improving document understanding accuracy. Our design preserves the base VLM weights, facilitating easy integration across different model architectures and providing flexibility to balance OCR and visual tokens during inference. + +# 3.1. Architecture + +Our method introduces two main components that complement existing VLM architectures: an OCR encoder that processes OCR extracted text and layout information, and a query compression mechanism that distills this information into a compact representation. We integrate these components with pre-trained VLMs, which employ various strategies to control the number of visual tokens for efficient processing. Figure 2 illustrates the overall architecture. + +OCR Encoder Architecture We utilize DocFormerV2 [7], a T5-based encoder-decoder [45] designed for document understanding, which incorporates vision, language, and spatial features. Specifically, we leverage only the encoder component, which comprises 344 million parameters, and omit its visual branch to eliminate redundancy with the VLM's vision capabilities and reduce computational complexity. The encoder processes two types of inputs: user instructions and OCR data from an OCR system, which consists of textual tokens and their corresponding 2D positional information [6, 7, 12, 14, 22, 26]. + +Query Compression Mechanism To efficiently integrate OCR information into VLMs, we introduce an instruction-aware compression mechanism that distills the OCR encoder's output into a compact set of learned queries. This mechanism significantly reduces the input sequence length for the language model while preserving essential document information. The compression process utilizes $M$ learnable queries $\mathbf{Q}$ (typically $M = 64$ ), initialized randomly following the OCR encoder embeddings' distribution. These queries are processed by the OCR encoder alongside two types of embeddings: OCR embeddings $(\mathbf{E}_{\mathrm{OCR}})$ , which encode both OCR tokens and their bounding boxes, and instruction embeddings $(\mathbf{E}_{\mathrm{Instructions}})$ . The encoding process can be represented as: + +$$ +\operatorname {E n c o d e r} \left(\left[ \mathbf {E} _ {\mathrm {O C R}}, \mathbf {E} _ {\text {I n s t r u c t i o n s}}, \mathbf {Q} \right]\right). +$$ + +From the encoder output, we retain only the $M$ features corresponding to the learned queries. These compressed features are then projected to match the VLM's hidden dimension and concatenated with the visual tokens before entering the language model. This compression significantly reduces the LLM's input sequence length, enabling either more efficient processing or, under a fixed token budget, allocation of additional tokens to visual features. + +![](images/554d8389308a98dbf09ffbb9088233db940c4846db8b657ef8958463d5206fc3.jpg) +Figure 2. DocVLM Architecture. DocVLM enhances document understanding in frozen VLMs by integrating an OCR module with a query compression mechanism. By condensing OCR data into $\mathbf{M} = 64$ learnable tokens, DocVLM effectively complements visual information, surpassing the VLM's inherent approaches of increasing image resolution or visual feature dimensions. + +Vision Process OCR-free VLMs employ different visual processing methods and various strategies to control the number of visual tokens, aiming to reduce the computational cost of processing high-resolution images needed for document understanding. These approaches can be grouped into three main paradigms: + +1. Full Image Processing with Image Resizing (e.g., Qwen2-VL [53]): In this approach, the model processes the entire image as a single input, controlling the number of visual tokens by resizing the image to a fixed or range-constrained image resolutions. While this image processing preserves global context, it incurs quadratic computational complexity with respect to the number of visual tokens. +2. Patch-Based Processing with Controlled Tile Count (e.g., InternVL2 [18]): This strategy segments the input image into spatial tiles, processing each independently, and controls the number of visual tokens by limiting the number of tiles. While most implementations incorporate a low-resolution global view, the primary focus on local processing may compromise global context understanding. The computational complexity for this approach scales linearly with respect to the number of tiles as the image increase. The result is improved memory efficiency compared to processing the full image at once, especially for large images. +3. Full-Scale Processing with Feature Downsampling (e.g., LlaVA-OneVision [35]): Some VLMs initially + +process full-scale images, but then downsample the visual features into maximal token count before feeding them to the LLM. While this method captures both global and local context, it introduces significant computational overhead during initial full-scale processing. + +Experimental results confirm that our OCR query compression mechanism significantly enhances document understanding capabilities across all three visual processing strategies, demonstrating its effectiveness as a universal enhancement to existing VLM architectures. + +# 3.2. Training Strategy + +Our training strategy aims to integrate the OCR modality into existing VLMs while preserving their core strengths. To achieve this, we keep the VLM completely frozen throughout the process and train only the newly introduced OCR components, i.e., the learnable queries, the OCR encoder, and the projection layer. We employ a two-stage training strategy to gradually incorporate the OCR modality into the pre-trained VLM: + +Stage I: OCR-LLM Alignment. During this stage, we withhold image input from the VLM, forcing the model to rely solely on the newly introduced OCR modality. This approach ensures full utilization of OCR data, aligns OCR components with the LLM input space, and reduces sequence length, improving training efficiency. Given the focus on text input, our dataset selection concentrates on text-related tasks. We begin by training only the randomly initialized components: the learnable queries and projection layer. This allows these components to adapt without disrupting the pretrained OCR encoder. Subsequently, we unfreeze the OCR encoder for fine-tuning, enabling a more comprehensive alignment of the entire encoder to the VLM. + +Stage II: Vision Alignment. In this final stage, we incorporate visual information extracted from the visual encoder, encouraging the OCR components to complement the visual features. Our experiments reveal this stage has a particularly strong effect when using fewer learned queries, allowing the compressed OCR information to better supplement the information acquired from the visual modality (see Sec. 5). During this stage, we add more visually focused datasets to the training process. Note that despite our method preserving the original VLM weights, it might implicitly inject bias through prompt tuning. To avoid this, the training data should represent all tasks of interest. + +# 3.3. Multipage Document Extension + +We conduct our training procedure on single-page data only. However, our approach can be extended to operate on multipage documents. Given a multipage input and its OCR information, the VLM independently processes each page image and concatenates the resulting visual features. For the OCR information, we explore two strategies: Global + +Encoding, which compresses the entire document's OCR information into 64 learnable queries, and Page-wise Encoding, which compresses each page's OCR information separately into 64 learnable queries and then concatenates them, resulting in $64 \times$ number of pages learned queries. After processing the OCR information using either strategy, we feed the resulting compressed OCR features along with the concatenated visual features into the LLM. + +Our experiments demonstrate that both approaches are highly effective and efficient in processing multipage documents. Using either approach with a restricted number of visual tokens, we obtain strong zero-shot results on DUDE [51] and state-of-the-art results on MP-DocVQA [50]. The page-wise encoding strategy yields slightly better results for lower numbers of visual tokens. + +# 4. Experiments + +# 4.1. Experimental Setting + +Model Integration: We evaluate DocVLM through integration with three leading open-source VLMs: LLaVA-OneVision [31], InternVL2 [18], and Qwen2-VL [53]. As discussed in Sec. 3.1, these models employ distinct token reduction strategies, enabling us to assess DocVLM's effectiveness across different visual processing approaches. + +Training: Our training protocol, detailed in Sec. 3.2, employs a two-phase strategy. The initial phase focuses on text-centric tasks using datasets spanning document understanding (DocVQA [40], InfoVQA [41]), scene text analysis (ST-VQA [11], TextVQA [48], OCR-VQA [42]), and specialized tasks (ChartQA [39], TextCaps [47], TATDQA [62]). The subsequent vision alignment phase incorporates additional visual-centric datasets: COCO Caption [15] and VQA-V2 [24]. + +Evaluation: For evaluation, we focus on five key benchmarks: DocVQA, TextVQA, ST-VQA, InfoVQA, and TextCaps. Results are reported on test sets where available, with TextVQA and TextCaps evaluated on validation sets due to test server restrictions. We use ANLS as the evaluation metric for all datasets, except TextVQA, which uses VQAScore, and TextCaps, which uses CIDEr. To demonstrate DocVLM's generalization capabilities, we conduct zero-shot evaluation on multipage document understanding benchmarks: DUDE [51] and MP-DocVQA [50]. This zero-shot performance is particularly noteworthy as our model is trained exclusively on single-page documents. The supplementary presents additional implementation details, including hyperparameters and optimization strategies, and robustness analysis to different OCR systems. + +# 4.2. State-of-the-art Comparisons + +Table 1 presents comprehensive comparisons between DocVLM and other state-of-the-art methods across var + +
Method# Tok.#PDocVQATextVQAST-VQAInfoVQATextCAPSMP-DocVQADUDE*
No Token Limitations
GPT-4o92.877.4-79.2---
Gemini 1.5 Pro93.178.7-81.0---
GPT-4V87.278.0-75.1---
KOSMOS-2.5-CHAT4K1.3B81.140.7-41.3---
TextSquare2.5K8.6B84.366.8-51.5---
ScreenAI3.5K5B87.8--57.8-72.9-
ScreenAI+OCR4.3K5B89.9--65.9-77.1-
Pali-35.5K5B86.779.584.157.8158.8--
Pali-3+OCR6.3K5B88.680.885.762.4164.3--
# Tokens ≤ 1.5k
URader8417B65.457.6-42.2118.4--
Monkey1.3K9B66.564.3-36.193.2--
TextMonkey7689B73.065.9-28.6---
Vary2567B76.3------
DocOwl23248B80.766.7-46.4131.869.446.8
GRAM9001B85.3----80.351.2
GRAMC-Former2561B87.6----77.645.5
DocFormer v21K1B87.864.071.848.8-76.448.4
LLaVA-OneVision7K7B87.576.171.168.8138.0OOMOOM
LLaVA-OneVision1.5K7B66.572.170.645.6112.941.828.7
DocVLMLLaVA-OneVision (Ours)1.5K7B88.476.970.861.0145.377.943.8
InternVL 23.1K8B91.677.4-74.8-OOMOOM
InternVL 22568B56.065.765.738.451.151.030.5
DocVLMInternVL2 (Ours)3208B86.671.274.357.6119.476.243.3
InternVL 212808B85.775.568.361.543.778.142.2
DocVLMInternVL2 (Ours)13448B91.076.776.765.4123.481.845.6
Qwen2-VL16k7B94.584.370.776.5150.2OOMOOM
Qwen2-VL3207B84.478.070.154.1142.173.041.5
DocVLMQwen2-VL (Ours)3207B91.279.676.561.2144.381.746.1
Qwen2-VL5767B91.582.370.565.3145.082.145.9
DocVLMQwen2-VL (Ours)5767B92.882.879.866.8150.484.547.4
+ +Table 1. Comparison with State-of-the-Art Methods. Performance evaluation of DocVLM against state-of-the-art approaches on document understanding benchmarks. Results are categorized into unconstrained models and those with a 1.5k token limit. In the constrained token regime, DocVLM consistently enhances the performance of baseline VLMs across various tasks and visual token budgets. Notably, DocVLM paired with Qwen2-VL (576 tokens) achieves superior performance across all evaluated datasets, including state-of-the-art zero-shot accuracy on DUDE. $\star$ indicates zero-shot evaluation, with grey entries denoting non-zero-shot results. + +ious document understanding benchmarks, highlighting DocVLM's ability to improve performance under token constraints. We categorize the results into two main groups: methods without token constraints (both closed and open-source models) and those operating under a 1.5k token limit. We mainly focused on models with around 7B parameters and include methods using an OCR system such as Pali3 [10] ScreenAI [8], DocFormerV2 [7], and GRAM [14]. + +To evaluate DocVLM's effectiveness under token constraints, we integrated it with three baseline models: LlaVA-OneVision [31], InternVL2 [18], and Qwen2-VL [53], each configured to operate within the 1.5k token limit. For LlaVA-OneVision, we utilized the minimal visual token configuration (single visual features tile). InternVL2 was tested with both single-tile (256 tokens) and four-tile (1280 tokens) configurations, while Qwen2-VL was evaluated with 256 and 512 visual tokens, corresponding to image + +sizes of $448 \times 448$ and $616 \times 616$ respectively. + +Under the practical 1.5k token constraint, incorporating DocVLM with each of these baseline models yields substantial and consistent improvements. Notably, these improvements persist even in looser token regimes, such as InternVL2 with 1280 visual tokens and Qwen2-VL with 576 visual tokens. Within this constraint, our Qwen2-VL variant with DocVLM, using just 576 tokens, achieves state-of-the-art performance across all benchmarks: $92.8\%$ on DocVQA, $82.8\%$ on TextVQA, $79.8\%$ on ST-VQA, $66.8\%$ on InfoVQA, and a CIDEr score of 150.4 on TextCAPS. As detailed in the Appendix, DocVLM achieves these results with significant computational advantages. + +DocVLM also demonstrates exceptional capability in handling multipage documents. Using the same 576-token configuration described earlier, it achieves $84.5\%$ accuracy on MP-DocVQA, surpassing previous state-of-the-art re + +![](images/c79c75b43eab3846f46dea7e2283a6ba1c12a778a56419801d1d40e808b72f2d.jpg) + +![](images/f06f08792c330366c71a07d59eacc3a42ea1b996951bb686181e954aeadfa07a.jpg) +Figure 3. Qualitative Results. Representative examples of DocVLM's performance across diverse document formats, from dense text to infographics and scene text. Our model successfully handles complex layouts, dense content, and presents instruction-following capabilities without explicit training on such datasets. Each example includes an image-instruction pair with baseline and DocVLM predictions. + +![](images/cf27e73cc1d6f78ec8df28502c3f840cca7bd06203ebfe8b2dc9b7a0c2d44593.jpg) + +![](images/f77db2901218e23d6ed4ad97896dd025f8c728bc9de19b73954b498ab3344c3e.jpg) + +sults. Furthermore, this setup shows robust generalization capabilities with a zero-shot performance of $47.4\%$ accuracy on the DUDE dataset, despite not being specifically trained for multipage document processing. + +# 4.3. Qualitative Results + +Figure 3 illustrates DocVLM's enhanced capabilities through representative examples that demonstrate three key strengths of our method: (1) improved reading comprehension in complex document layouts, (2) effective handling of dense textual content despite using compressed representations, and (3) preserved and enhanced instruction-following capabilities. + +The examples span diverse document types, from dense text documents to infographics and scene text, showcasing DocVLM's versatility. In the infographic example, DocVLM not only preserves but enhances the base model's instruction-following capabilities, despite our OCR component not being explicitly trained on instruction-following datasets like [49]. This demonstrates that our compression mechanism successfully retains crucial textual and layout information while significantly reducing token usage. + +# 4.4. Scaling to Multipage Documents + +Building on the promising multipage results in Tab. 1, we conduct an in-depth analysis of different DocVLM configurations for multipage document understanding. This analysis focuses on the Qwen2-VL base model, tested on the MP-DocVQA dataset with documents up to 20 pages long – a scale that most other state-of-the-art methods struggle to handle due to token limitations. + +Table 2 compares four OCR integration strategies used during inference in multipage scenarios: + +- Baseline: vision-only input (no additional tokens) +- Direct OCR word insertion: up to 800 tokens per page +- Global OCR encoding: 64 tokens total +- Page-wise OCR encoding: 64 tokens per page + +Results indicate consistent improvements over the baseline across all image resolutions (256, 512, and 1024 tokens), with only minimal additional token usage – just 64 tokens for the entire document in the global encoding case. Notably, at 256 visual tokens per page, both page-wise encoding $(82.4\%)$ and global encoding $(81.7\%)$ outperform direct OCR word insertion $(79.1\%)$ while using significantly fewer tokens. + +Our best configuration achieves state-of-the-art perfor + +
MethodLLM OCR InputImage Tok.OCR Tok.ANLS
DocOwl2 [27]-324 × pg-69.4
GRAMC-Former [14]-100 × pg25677.6
GRAM [14]-100 × pg800 × pg80.3
Qwen2-VL-256 × pg-73.0
OCR Words256 × pg800 × pg79.1
DocVLMQwen2-VLGlobal Encoding256 × pg6481.7
DocVLMQwen2-VLPage-wise Encoding256 × pg64 × pg82.4
Qwen2-VL-512 × pg-82.1
DocVLMQwen2-VLGlobal Encoding512 × pg6484.5
DocVLMQwen2-VLPage-wise Encoding512 × pg64 × pg85.2
Qwen2-VL-1024 × pg-85.2
DocVLMQwen2-VLGlobal Encoding1024 × pg6486.3
DocVLMQwen2-VLPage-wise Encoding1024 × pg64 × pg86.3
+ +mance of $86.3\%$ ANLS, significantly outperforming specialized multipage models like GRAM $(80.3\%)$ . This is particularly impressive considering that DocVLM was trained exclusively on single-page inputs, demonstrating strong zero-shot generalization to multipage scenarios. + +The comparison between encoding strategies reveals that page-wise encoding consistently outperforms global encoding at lower visual token counts, providing a $+0.7\%$ improvement for both 256 and 512 image tokens per page. This advantage diminishes at 1024 tokens where both achieve identical performance (86.3% ANLS). Remarkably, DocVLM with page-wise encoding matches or even outperforms the baseline Qwen2-VL using twice as many visual tokens, highlighting the efficiency of our approach. + +# 5. Ablation Study + +Impact of OCR Encoding Strategies To evaluate the impact of OCR encoding compression, we compare three strategies for integrating OCR information: (1) inserting raw OCR words in the original VLM, (2) using DocVLM uncompressed OCR encodings, and (3) DocVLM compressed OCR encodings with 64 learned queries. We evaluate these approaches on the DocVQA test set using three representative model configurations: LLaVA-OneVision with 1.5K visual tokens, and both InternVL2 and Qwen2-VL with 256 visual tokens each. + +Results in Tab. 3 demonstrate that DocVLM's OCR encodings significantly outperform raw OCR words across all three models while maintaining the same token count. Notably, our compressed encoding approach, using just 64 tokens instead of 800 OCR tokens, preserves most of these improvements while drastically reducing sequence length. This efficient compression enables a more favorable allocation of the token budget, allowing models to dedicate more tokens to visual processing without compromising OCR ef + +Table 2. Extension to Multipage. Comparison on MP-DocVQA of approaches for incorporating OCR information in multipage document understanding. Both DocVLM multipage extension strategies: global encoding (64 tokens per document) and pagewise encoding (64 tokens per page), outperform previous state-of-the-art methods, notably, without any explicit multipage training. + +
LLM OCR InputOCR Tok.LLaVA-OV 1.5KInternVL2 256Qwen2-VL 256
OCR Words80085.884.489.1
OCR Encoding80089.489.291.9
64 Compressed Encoding6488.486.691.2
+ +Table 3. OCR Encoding Strategies. DocVQA results for inserting OCR information using (1) OCR words (baseline), (2) uncompressed OCR encodings, and (3) 64 compressed OCR encodings. + +fectiveness. The results validate that DocVLM's compression strategy successfully balances performance with computational efficiency, a key factor for practical applications. + +Balancing Vision and OCR Token Allocation Modern VLMs employ various mechanisms to reduce visual token count, creating an inherent trade-off between computational efficiency and model performance, as discussed in Sec. 3.1. We investigate how DocVLM can improve this trade-off by comparing four configurations: (1) baseline VLM without OCR, (2) direct OCR word insertion, (3) DocVLM with uncompressed OCR encodings, and (4) DocVLM with 64 compressed learned queries. + +Figure 4 presents the performance scores on DocVQA validation (left y-axis) and total token counts (right y-axis) for three VLM architectures, showing how these metrics vary across different visual token allocations. Each model employs a distinct token reduction approach: LLaVA-OneVision controls token count through feature downsampling (AnyRes Max), InternVL2 limits the number of processed image tiles (Dynamic Max Batch), and Qwen2-VL adjusts image resolution to constrain token count. + +Our analysis reveals that integrating OCR information, regardless of the method used, consistently improves performance across all models, with particularly pronounced gains in low visual token regimes. However, uncompressed OCR integration methods, whether through direct word insertion or uncompressed DocVLM encodings, require 800 tokens – a significant overhead that could otherwise be allocated to visual processing. For instance, allocating 128 tokens for visual and 800 for OCR in Qwen2-VL achieves $84.3\%$ using OCR words and $90.1\%$ using uncompressed encodings. In contrast, using 896 pure visual tokens reaches $92.4\%$ , demonstrating the potential benefit of allocating more tokens to visual processing. DocVLM's compression mechanism provides a superior option by requiring only 64 tokens for OCR information while maintaining strong performance. In the above example, our approach using fewer tokens, allocating 768 visual tokens and 64 OCR tokens, reaches $93.0\%$ , outperforming the $90.1\%$ obtained with the uncompressed encodings, highlighting DocVLM's effective balance between visual and OCR tokens. + +![](images/96eb4592a64f4bd6744a858002f3392f1d3dd2945f18b048be3ead0ad81d6f65.jpg) + +![](images/423a8f08b63fd05d073cb4c1fec9eba16c90a79ee2783da1cf9bda63ab362dbb.jpg) + +![](images/7898eecf2f11b51e09f0d852cb9383f32525f92a2316d36d6ac74b6c688b1e60.jpg) + +![](images/649344a343f98136978236c1a3dcd4d41574aae6ad50f6203bcbedd27a1402a3.jpg) +Figure 4. Balancing Performance and Compute. Analysis of model performance (lines, left y-axis) and token usage (bars, right y-axis) as a function of visual token allocation. Each model employs its inherent token control strategy: AnyRes max for feature downsampling (LLaVA One-Vision), dynamic max tiles (InternVL2), and max image tokens for resolution control (Qwen2-VL). The results highlight that DocVLM consistently improves performance with minimal overhead (64 tokens), offering an efficient OCR-visual token allocation. + +![](images/e7ec1136051f6b341035ba7cbc22e3d91980c4f7b9dde9710591956639eb1e57.jpg) +Figure 5. Compression Levels. DocVQA validation results for DocVLM integrated with Qwen2-VL across varying OCR and image token budgets. "0" represents the baseline, while "Full" indicates uncompressed encodings. + +Compression Levels We deepen our analysis of the OCR-visual token trade-off by examining different compression levels in DocVLM integrated with Qwen2-VL. Fig. 5 presents ANLS scores across various combinations of visual tokens (128-1024) and learned queries (16-256), including baselines without OCR and with uncompressed encoding. In low visual token regimes, increasing the number of learned queries yields substantial improvements, validating our compression mechanism's effectiveness in capturing relevant OCR information. Notably, even with just 16 learned queries, DocVLM outperforms the baseline across all visual token configurations, offering strong performance with minimal computational overhead. + +Training Stages Tab. 4 illustrates the impact of the vision alignment stage on ANLS performance for DocVLM with the LLaVA-OneVision base model on the DocVQA test set, showing results for varying numbers of learned queries (16 to 128) and the uncompressed case. Our two-stage training process initially trains OCR modality components with- + +
Training PhasesCompressed Enc.OCR Enc.
1664128800
Stage I: OCR-LLM Alignment81.785.886.389.4
+ Stage II: Vision Alignment87.988.488.490.1
Δ+6.2+2.6+2.1+0.7
+ +Table 4. Training Phases. Second-stage training consistently improves DocVQA performance, both with compressed DocVLM tokens and full OCR encoding. + +out image input, forcing reliance on OCR data alone, before reintroducing the image in the vision alignment stage to adapt learned queries alongside visual information. The results reveal that vision alignment significantly boosts performance, especially with fewer learned queries: for instance, with 16 learned queries, there's an improvement of +6.2, compared to +0.7 in the uncompressed case. Notably, after vision alignment, DocVLM with only 16 learned queries outperforms the baseline of OCR words (from Table 3). These findings underscore the effectiveness of our two-stage training method. + +# 6. Conclusions + +Our results demonstrate that DocVLM can be effectively integrated into various VLMs to enhance their document reading capabilities while significantly reducing their dependency on extensive vision tokens. The key takeaway is that in token-constrained scenarios, allocating a small portion of tokens to OCR information consistently yields better results than using those tokens solely for visual processing. The effectiveness of our compression mechanism extends beyond single-page documents, as evidenced by achieving state-of-the-art results on MP-DocVQA using the same 64 tokens to represent multiple pages. These results establish DocVLM as a practical solution for enhancing document understanding in real-world applications where computational efficiency is crucial. + +# References + +[1] Aviad Aberdam, Ron Litman, Shahar Tsiper, Oron Anschel, Ron Slossberg, Shai Mazor, R Manmatha, and Pietro Perona. Sequence-to-sequence contrastive learning for text recognition. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 15302-15312, 2021. 2 +[2] Aviad Aberdam, Roy Ganz, Shai Mazor, and Ron Litman. Multimodal semi-supervised learning for text recognition. arXiv preprint arXiv:2205.03873, 2022. +[3] Aviad Aberdam, David Bensaïd, Alona Golts, Roy Ganz, Oren Nuriel, Royee Tichauer, Shai Mazor, and Ron Litman. Clipter: Looking at the bigger picture in scene text recognition. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 21706-21717, 2023. +[4] Ofir Abramovich, Niv Nayman, Sharon Fogel, Inbal Lavi, Ron Litman, Shahar Tsiper, Royee Tichauer, Srikar Appalaraju, Shai Mazor, and R Manmatha. Visfocus: Prompt-guided vision encoders forOCR-free dense document understanding. In European Conference on Computer Vision, pages 241-259. Springer, 2025. 2 +[5] Jean-Baptiste Alayrac, Jeff Donahue, Pauline Luc, Antoine Miech, Iain Barr, Yana Hasson, Karel Lenc, Arthur Mensch, Katherine Millican, Malcolm Reynolds, et al. Flamingo: a visual language model for few-shot learning. Advances in neural information processing systems, 35:23716-23736, 2022. 2 +[6] Srikar Appalaraju, Bhavan Jasani, Bhargava Urala Kota, Yusheng Xie, and R Manmatha. Docformer: End-to-end transformer for document understanding. In Proceedings of the IEEE/CVF international conference on computer vision, pages 993-1003, 2021. 3 +[7] Srikar Appalaraju, Peng Tang, Qi Dong, Nishant Sankaran, Yichu Zhou, and R Manmatha. Docformerv2: Local features for document understanding. In Proceedings of the AAAI Conference on Artificial Intelligence, pages 709-718, 2024. 3, 5, 12 +[8] Gilles Baechler, Srinivas Sunkara, Maria Wang, Fedir Zubach, Hassan Mansoor, Vincent Etter, Victor Cärbune, Jason Lin, Jindong Chen, and Abhanshu Sharma. Screenai: A vision-language model for ui and infographics understanding. arXiv preprint arXiv:2402.04615, 2024. 1, 5 +[9] Jinze Bai, Shuai Bai, Shusheng Yang, Shijie Wang, Sinan Tan, Peng Wang, Junyang Lin, Chang Zhou, and Jingren Zhou. Qwen-vl: A versatile vision-language model for understanding, localization, text reading, and beyond. arXiv preprint arXiv:2308.12966, 1(2):3, 2023. 2 +[10] Lucas Beyer, Andreas Steiner, André Susano Pinto, Alexander Kolesnikov, Xiao Wang, Daniel Salz, Maxim Neumann, Ibrahim Alabdulmohsin, Michael Tschannen, Emanuele Bugliarello, et al. Paligemma: A versatile 3b vlm for transfer. arXiv preprint arXiv:2407.07726, 2024. 1, 5 +[11] Ali Furkan Biten, Ruben Tito, Andres Mafla, Lluis Gomez, Marçal Rusinol, Ernest Valveny, CV Jawahar, and Dimosthenis Karatzas. Scene text visual question answering. In Proceedings of the IEEE/CVF international conference on computer vision, pages 4291-4301, 2019. 4, 12 + +[12] Ali Furkan Biten, Ron Litman, Yusheng Xie, Srikar Appalaraju, and R Manmatha. Latr: Layout-aware transformer for scene-text vqa. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 16548-16558, 2022. 2, 3 +[13] Ali Furkan Biten, Ruben Tito, Lluis Gomez, Ernest Valveny, and Dimosthenis Karatzas. Ocr-idl: Ocr annotations for industry document library dataset. arXiv preprint arXiv:2202.12985, 2022. 12 +[14] Tsachi Blau, Sharon Fogel, Roi Ronen, Alona Golts, Roy Ganz, Elad Ben Avraham, Aviad Aberdam, Shahar Tsiper, and Ron Litman. Gram: Global reasoning for multi-page vqa. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 15598-15607, 2024. 2, 3, 5, 7 +[15] Xinlei Chen, Hao Fang, Tsung-Yi Lin, Ramakrishna Vedantam, Saurabh Gupta, Piotr Dólar, and C Lawrence Zitnick. Microsoft coco captions: Data collection and evaluation server. arXiv preprint arXiv:1504.00325, 2015. 4, 12 +[16] Xi Chen, Josip Djolonga, Piotr Padlewski, Basil Mustafa, Soravit Changpinyo, Jialin Wu, Carlos Riquelme Ruiz, Sebastian Goodman, Xiao Wang, Yi Tay, et al. Pali-x: On scaling up a multilingual vision and language model. arXiv preprint arXiv:2305.18565, 2023. 2 +[17] Xi Chen, Xiao Wang, Lucas Beyer, Alexander Kolesnikov, Jialin Wu, Paul Voigtlaender, Basil Mustafa, Sebastian Goodman, Ibrahim Alabdulmohsin, Piotr Padlewski, et al. Pali-3 vision language models: Smaller, faster, stronger. arXiv preprint arXiv:2310.09199, 2023. 1, 2 +[18] Zhe Chen, Weiyun Wang, Hao Tian, Shenglong Ye, Zhangwei Gao, Erfei Cui, Wenwen Tong, Kongzhi Hu, Jiapeng Luo, Zheng Ma, et al. Internvl2: Better than the best—expanding performance boundaries of open-source multimodal models with the progressive scaling strategy, 2024. 2, 3, 4, 5 +[19] Wenliang Dai, Junnan Li, Dongxu Li, Anthony Meng Huat Tiong, Junqi Zhao, Weisheng Wang, Boyang Albert Li, Pascale Fung, and Steven C. H. Hoi. Instructclip: Towards general-purpose vision-language models with instruction tuning. ArXiv, abs/2305.06500, 2023. 2 +[20] Xiaoyi Dong, Pan Zhang, Yuhang Zang, Yuhang Cao, Bin Wang, Linke Ouyang, Songyang Zhang, Haodong Duan, Wenwei Zhang, Yining Li, et al. Internlm-xcomposer2-4khd: A pioneering large vision-language model handling resolutions from 336 pixels to 4k hd. arXiv preprint arXiv:2404.06512, 2024. 1 +[21] Masato Fujitake. Dtrocr: Decoder-only transformer for optical character recognition. In Proceedings of the IEEE/CVF Winter Conference on Applications of Computer Vision, pages 8025-8035, 2024. 2 +[22] Roy Ganz, Oren Nuriel, Aviad Aberdam, Yair Kittenplon, Shai Mazor, and Ron Litman. Towards models that can see and read. In Proceedings of the IEEE/CVF international conference on computer vision, pages 21718-21728, 2023. 2, 3 +[23] Roy Ganz, Yair Kittenplon, Aviad Aberdam, Elad Ben Avraham, Oren Nuriel, Shai Mazor, and Ron Litman. Question aware vision transformer for multimodal reasoning. In Proceedings of the IEEE/CVF Conference on Computer Vision + +and Pattern Recognition (CVPR), pages 13861-13871, 2024. 2 +[24] Yash Goyal, Tejas Khot, Douglas Summers-Stay, Dhruv Bartra, and Devi Parikh. Making the V in VQA matter: Elevating the role of image understanding in Visual Question Answering. In Conference on Computer Vision and Pattern Recognition (CVPR), 2017. 4, 12 +[25] Jean-Philippe Thiran Guillaume Jaume, Hazim Kemal Ekenel. Funsd: A dataset for form understanding in noisy scanned documents. In Accepted to ICDAR-OST, 2019. 1 +[26] Benjamin Hsu, Xiaoyu Liu, Huayang Li, Yoshinari Fujinuma, Maria Nadejde, Xing Niu, Yair Kittenplon, Ron Litman, and Raghavendra Pappagari. M3t: A new benchmark dataset for multi-modal document-level machine translation. arXiv preprint arXiv:2406.08255, 2024. 3 +[27] Anwen Hu, Haiyang Xu, Liang Zhang, Jiabo Ye, Ming Yan, Ji Zhang, Qin Jin, Fei Huang, and Jingren Zhou. mplug-docowl2: High-resolution compressing forocr-free multi-page document understanding. arXiv preprint arXiv:2409.03420, 2024. 2, 7 +[28] Taeho Kil, Seonghyeon Kim, Sukmin Seo, Yoonsik Kim, and Daehe Kim. Towards unified scene text spotting based on sequence generation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 15223-15232, 2023. 2 +[29] Geewook Kim, Teakgyu Hong, Moonbin Yim, Jinyoung Park, Jinyeong Yim, Wonseok Hwang, Sangdoo Yun, Dongyoon Han, and Seunghyun Park. Donut: Document understanding transformer withoutOCR.arXiv preprint arXiv:2111.15664, 7(15):2, 2021. 2 +[30] Hugo Laurençon, Léo Tronchon, Matthieu Cord, and Victor Sanh. What matters when building vision-language models? arXiv preprint arXiv:2405.02246, 2024. 2 +[31] Bo Li, Yuanhan Zhang, Dong Guo, Renrui Zhang, Feng Li, Hao Zhang, Kaichen Zhang, Yanwei Li, Ziwei Liu, and Chunyuan Li. Llava-onevision: Easy visual task transfer. arXiv preprint arXiv:2408.03326, 2024. 1, 2, 4, 5 +[32] Junnan Li, Dongxu Li, Silvio Savarese, and Steven Hoi. Blip-2: Bootstrapping language-image pre-training with frozen image encoders and large language models. In International conference on machine learning, pages 19730-19742. PMLR, 2023. 2 +[33] Wentong Li, Yuqian Yuan, Jian Liu, Dongqi Tang, Song Wang, Jianke Zhu, and Lei Zhang. Tokenpacker: Efficient visual projector for multimodal llm. arXiv preprint arXiv:2407.02392, 2024. 2 +[34] Ron Litman, Oron Anschel, Shahar Tsiper, Roeit Litman, Shai Mazor, and R Manmatha. Scatter: selective context attentional scene text recognizer. In proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 11962-11972, 2020. 2 +[35] Haotian Liu, Chunyuan Li, Yuheng Li, Bo Li, Yuanhan Zhang, Sheng Shen, and Yong Jae Lee. Llava next: Improved reasoning,OCR, and world knowledge, 2024. 2, 3 +[36] Haotian Liu, Chunyuan Li, Qingyang Wu, and Yong Jae Lee. Visual instruction tuning. Advances in neural information processing systems, 36, 2024. 2 + +[37] Yuliang Liu, Zhang Li, Mingxin Huang, Biao Yang, Wenwen Yu, Chunyuan Li, Xucheng Yin, Cheng lin Liu, Lianwen Jin, and Xiang Bai. Ocrbench: On the hidden mystery ofOCR in large multimodal models, 2024. 1 +[38] Jiahao Lyu, Wei Wang, Dongbao Yang, Jinwen Zhong, and Yu Zhou. Arbitrary reading order scene text spotter with local semantics guidance. arXiv preprint arXiv:2412.10159, 2024. 14 +[39] Ahmed Masry, Do Xuan Long, Jia Qing Tan, Shafiq Joty, and Enamul Hoque. Chartqa: A benchmark for question answering about charts with visual and logical reasoning. arXiv preprint arXiv:2203.10244, 2022. 1, 4, 12 +[40] Minesh Mathew, Dimosthenis Karatzas, and CV Jawahar. Docvqa: A dataset for vqa on document images. In Proceedings of the IEEE/CVF winter conference on applications of computer vision, pages 2200-2209, 2021. 4, 12, 13 +[41] Minesh Mathew, Viraj Bagal, Ruben Tito, Dimosthenis Karatzas, Ernest Valveny, and CV Jawahar. Infographicvqa. In Proceedings of the IEEE/CVF Winter Conference on Applications of Computer Vision, pages 1697-1706, 2022. 1, 4, 12, 13 +[42] Anand Mishra, Shashank Shekhar, Ajeet Kumar Singh, and Anirban Chakraborty. Ocr-vqa: Visual question answering by reading text in images. In 2019 international conference on document analysis and recognition (ICDAR), pages 947-952. IEEE, 2019. 4, 12 +[43] Oren Nuriel, Sharon Fogel, and Ron Litman. Textadain: Paying attention to shortcut learning in text recognizers. In European Conference on Computer Vision, pages 427-445. Springer, 2022. 2 +[44] Zhiliang Peng, Wenhui Wang, Li Dong, Yaru Hao, Shaohan Huang, Shuming Ma, and Furu Wei. Kosmos-2: Grounding multimodal large language models to the world. arXiv preprint arXiv:2306.14824, 2023. 2 +[45] Colin Raffel, Noam Shazeer, Adam Roberts, Katherine Lee, Sharan Narang, Michael Matena, Yanqi Zhou, Wei Li, and Peter J Liu. Exploring the limits of transfer learning with a unified text-to-text transformer. Journal of machine learning research, 21(140):1-67, 2020. 3 +[46] Roi Ronen, Shahar Tsiper, Oron Anschel, Inbal Lavi, Amir Markovitz, and R Manmatha. Glass: Global to local attention for scene-text spotting. arXiv preprint arXiv:2208.03364, 2022. 2 +[47] Oleksii Sidorov, Ronghang Hu, Marcus Rohrbach, and Amanpreet Singh. Textcaps: a dataset for image captioning with reading comprehension. In Computer Vision-ECCV 2020: 16th European Conference, Glasgow, UK, August 23-28, 2020, Proceedings, Part II 16, pages 742-758. Springer, 2020. 4, 12 +[48] Amanpreet Singh, Vivek Natarajan, Meet Shah, Yu Jiang, Xinlei Chen, Dhruv Batra, Devi Parikh, and Marcus Rohrbach. Towards vqa models that can read. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 8317-8326, 2019. 4, 12 +[49] Ryota Tanaka, Taichi Iki, Kyosuke Nishida, Kuniko Saito, and Jun Suzuki. Instructdoc: A dataset for zero-shot generalization of visual document understanding with instructions. + +In Proceedings of the AAAI Conference on Artificial Intelligence, pages 19071-19079, 2024. 6 +[50] Ruben Tito, Dimosthenis Karatzas, and Ernest Valveny. Hierarchical multimodal transformers for multipage docvqa. Pattern Recognition, 144:109834, 2023. 1, 2, 4, 12 +[51] Jordy Van Landeghem, Ruben Tito, Łukasz Borchmann, Michal Pietruszka, Pawel Joziak, Rafal Powalski, Dawid Jurkiewicz, Mickaël Coustaty, Bertrand Anckaert, Ernest Valveny, et al. Document understanding dataset and evaluation (dude). In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 19528-19540, 2023. 2, 4, 12 +[52] Dongsheng Wang, Natraj Raman, Mathieu Sibue, Zhiqiang Ma, Petr Babkin, Simerjot Kaur, Yulong Pei, Armineh Nourbakhsh, and Xiaomo Liu. Docllm: A layout-aware generative language model for multimodal document understanding. arXiv preprint arXiv:2401.00908, 2023. 2, 13 +[53] Peng Wang, Shuai Bai, Sinan Tan, Shijie Wang, Zhihao Fan, Jinze Bai, Keqin Chen, Xuejing Liu, Jialin Wang, Wenbin Ge, et al. Qwen2-vl: Enhancing vision-language model's perception of the world at any resolution. arXiv preprint arXiv:2409.12191, 2024. 1, 2, 3, 4, 5 +[54] Wenjin Wang, Yunhao Li, Yixin Ou, and Yin Zhang. Layout and task aware instruction prompt for zero-shot document image question answering. arXiv preprint arXiv:2306.00526, 2023. 1, 2, 13 +[55] Yiheng Xu, Minghao Li, Lei Cui, Shaohan Huang, Furu Wei, and Ming Zhou. Layoutm: Pre-training of text and layout for document image understanding. In Proceedings of the 26th ACM SIGKDD international conference on knowledge discovery & data mining, pages 1192-1200, 2020. 13 +[56] Jiabo Ye, Anwen Hu, Haiyang Xu, Qinghao Ye, Ming Yan, Guohai Xu, Chenliang Li, Junfeng Tian, Qi Qian, Ji Zhang, et al. Ureader: UniversalOCR-free visually-situated language understanding with multimodal large language model. arXiv preprint arXiv:2310.05126, 2023. 2 +[57] Maoyuan Ye, Jing Zhang, Shanshan Zhao, Juhua Liu, Bo Du, and Dacheng Tao. Aptext-detr: Towards better scene text detection with dynamic points in transformer. In Proceedings of the AAAI Conference on Artificial Intelligence, pages 3241-3249, 2023. 2 +[58] Maoyuan Ye, Jing Zhang, Shanshan Zhao, Juhua Liu, Tongliang Liu, Bo Du, and Dacheng Tao. Deepsolo: Let transformer decoder with explicit points solo for text spotting. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 19348-19357, 2023. 2 +[59] Qinghao Ye, Haiyang Xu, Guohai Xu, Jiabo Ye, Ming Yan, Yiyang Zhou, Junyang Wang, Anwen Hu, Pengcheng Shi, Yaya Shi, et al. MPLug-owl: Modularization empowers large language models with multimodality. arXiv preprint arXiv:2304.14178, 2023. 2 +[60] Pan Zhang, Xiaoyi Dong, Yuhang Zang, Yuhang Cao, Rui Qian, Lin Chen, Qipeng Guo, Haodong Duan, Bin Wang, Linke Ouyang, et al. Internlm-xcomposer-2.5: A versatile large vision language model supporting long-contextual input and output. arXiv preprint arXiv:2407.03320, 2024. 1 + +[61] Deyao Zhu, Jun Chen, Xiaogian Shen, Xiang Li, and Mohamed Elhoseiny. Minigpt-4: Enhancing vision-language understanding with advanced large language models. arXiv preprint arXiv:2304.10592, 2023. 2 +[62] Fengbin Zhu, Wenqiang Lei, Fuli Feng, Chao Wang, Haozhou Zhang, and Tat-Seng Chua. Towards complex document understanding by discrete reasoning. In Proceedings of the 30th ACM International Conference on Multimedia, pages 4857-4866, 2022. 4, 12 \ No newline at end of file diff --git a/2025/DocVLM_ Make Your VLM an Efficient Reader/images.zip b/2025/DocVLM_ Make Your VLM an Efficient Reader/images.zip new file mode 100644 index 0000000000000000000000000000000000000000..8ee15ce75d27fe9dfedddf5940e87aaddfcdf3a7 --- /dev/null +++ b/2025/DocVLM_ Make Your VLM an Efficient Reader/images.zip @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:fed9eccbeea00944f86987c62cde5c2610d5abdd0dfe22ad4f90c4ad44d1e497 +size 581710 diff --git a/2025/DocVLM_ Make Your VLM an Efficient Reader/layout.json b/2025/DocVLM_ Make Your VLM an Efficient Reader/layout.json new file mode 100644 index 0000000000000000000000000000000000000000..0ab6c62014b3d99a50cd397d769404d42219def1 --- /dev/null +++ b/2025/DocVLM_ Make Your VLM an Efficient Reader/layout.json @@ -0,0 +1,7424 @@ +{ + "pdf_info": [ + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 123, + 89, + 147, + 128 + ], + "blocks": [ + { + "bbox": [ + 123, + 89, + 147, + 128 + ], + "lines": [ + { + "bbox": [ + 123, + 89, + 147, + 128 + ], + "spans": [ + { + "bbox": [ + 123, + 89, + 147, + 128 + ], + "type": "image", + "image_path": "30136c38bce2d4f8bb1643f13deb7e09a6f1103d93600da549ab5a82392284d5.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + } + ], + "index": 2 + }, + { + "bbox": [ + 168, + 102, + 465, + 118 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 168, + 102, + 465, + 118 + ], + "spans": [ + { + "bbox": [ + 168, + 102, + 465, + 118 + ], + "type": "text", + "content": "DocVLM: Make Your VLM an Efficient Reader" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 126, + 143, + 484, + 185 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 126, + 143, + 484, + 185 + ], + "spans": [ + { + "bbox": [ + 126, + 143, + 484, + 185 + ], + "type": "text", + "content": "Mor Shpigel Nacson*† Aviad Aberdam* Roy Ganz Elad Ben Avraham Alona Golts Yair Kittenplon Shai Mazor Ron Litman AWS AI Labs" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 152, + 213, + 200, + 225 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 152, + 213, + 200, + 225 + ], + "spans": [ + { + "bbox": [ + 152, + 213, + 200, + 225 + ], + "type": "text", + "content": "Abstract" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 55, + 239, + 297, + 563 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 239, + 297, + 563 + ], + "spans": [ + { + "bbox": [ + 55, + 239, + 297, + 563 + ], + "type": "text", + "content": "Vision-Language Models (VLMs) excel in diverse visual tasks but face challenges in document understanding, which requires fine-grained text processing. While typical visual tasks perform well with low-resolution inputs, reading-intensive applications demand high-resolution, resulting in significant computational overhead. Using OCR-extracted text in VLM prompts partially addresses this issue but underperforms compared to full-resolution counterpart, as it lacks the complete visual context needed for optimal performance. We introduce DocVLM, a method that integrates an OCR-based modality into VLMs to enhance document processing while preserving original weights. Our approach employs an OCR encoder to capture textual content and layout, compressing these into a compact set of learned queries incorporated into the VLM. Comprehensive evaluations across leading VLMs show that DocVLM significantly reduces reliance on high-resolution images for document understanding. In limited-token regimes " + }, + { + "bbox": [ + 55, + 239, + 297, + 563 + ], + "type": "inline_equation", + "content": "(448\\times 448)" + }, + { + "bbox": [ + 55, + 239, + 297, + 563 + ], + "type": "text", + "content": ", DocVLM with 64 learned queries improves DocVQA results from " + }, + { + "bbox": [ + 55, + 239, + 297, + 563 + ], + "type": "inline_equation", + "content": "56.0\\%" + }, + { + "bbox": [ + 55, + 239, + 297, + 563 + ], + "type": "text", + "content": " to " + }, + { + "bbox": [ + 55, + 239, + 297, + 563 + ], + "type": "inline_equation", + "content": "86.6\\%" + }, + { + "bbox": [ + 55, + 239, + 297, + 563 + ], + "type": "text", + "content": " when integrated with InternVL2 and from " + }, + { + "bbox": [ + 55, + 239, + 297, + 563 + ], + "type": "inline_equation", + "content": "84.4\\%" + }, + { + "bbox": [ + 55, + 239, + 297, + 563 + ], + "type": "text", + "content": " to " + }, + { + "bbox": [ + 55, + 239, + 297, + 563 + ], + "type": "inline_equation", + "content": "91.2\\%" + }, + { + "bbox": [ + 55, + 239, + 297, + 563 + ], + "type": "text", + "content": " with Qwen2-VL. In LLaVA-OneVision, DocVLM achieves improved results while using " + }, + { + "bbox": [ + 55, + 239, + 297, + 563 + ], + "type": "inline_equation", + "content": "80\\%" + }, + { + "bbox": [ + 55, + 239, + 297, + 563 + ], + "type": "text", + "content": " less image tokens. The reduced token usage allows processing multiple pages effectively, showing impressive zero-shot results on DUDE and state-of-the-art performance on MP-DocVQA, highlighting DocVLM's potential for applications requiring high-performance and efficiency." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 56, + 589, + 135, + 601 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 589, + 135, + 601 + ], + "spans": [ + { + "bbox": [ + 56, + 589, + 135, + 601 + ], + "type": "text", + "content": "1. Introduction" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 55, + 609, + 296, + 682 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 609, + 296, + 682 + ], + "spans": [ + { + "bbox": [ + 55, + 609, + 296, + 682 + ], + "type": "text", + "content": "The ability to read and interpret text within images is crucial for numerous real-world applications, particularly in document understanding. This field encompasses diverse document types, from dense-text to infographics and multipage documents [25, 39-41, 50], involving tasks that require capabilities in text comprehension, layout understanding, and" + } + ] + } + ], + "index": 8 + }, + { + "type": "image", + "bbox": [ + 317, + 213, + 549, + 397 + ], + "blocks": [ + { + "bbox": [ + 317, + 213, + 549, + 397 + ], + "lines": [ + { + "bbox": [ + 317, + 213, + 549, + 397 + ], + "spans": [ + { + "bbox": [ + 317, + 213, + 549, + 397 + ], + "type": "image", + "image_path": "b4a69fd84bc032d003f4ac13f7d108554fbe2c480d72ae78b69684f10e8a3e56.jpg" + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 313, + 400, + 555, + 445 + ], + "lines": [ + { + "bbox": [ + 313, + 400, + 555, + 445 + ], + "spans": [ + { + "bbox": [ + 313, + 400, + 555, + 445 + ], + "type": "text", + "content": "Figure 1. DocVLM enhances VLMs' reading capabilities. Integrating DocVLM (solid lines) in top-performing VLMs (dashed lines) consistently improves the performance across all token budgets, frequently surpassing the baseline at higher token counts." + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "image_caption" + } + ], + "index": 9 + }, + { + "bbox": [ + 313, + 450, + 555, + 545 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 450, + 555, + 545 + ], + "spans": [ + { + "bbox": [ + 313, + 450, + 555, + 545 + ], + "type": "text", + "content": "visual interpretation. Despite progress in VLMs, processing such documents remains challenging [37], primarily due to the tension between resolution requirements and computational efficiency. While typical computer vision tasks achieve good performance with low-resolution inputs (typically " + }, + { + "bbox": [ + 313, + 450, + 555, + 545 + ], + "type": "inline_equation", + "content": "224 \\times 224" + }, + { + "bbox": [ + 313, + 450, + 555, + 545 + ], + "type": "text", + "content": " or " + }, + { + "bbox": [ + 313, + 450, + 555, + 545 + ], + "type": "inline_equation", + "content": "336 \\times 336" + }, + { + "bbox": [ + 313, + 450, + 555, + 545 + ], + "type": "text", + "content": " pixels), document analysis demands significantly higher resolutions, resulting in substantial computational overhead [10, 60]." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 313, + 546, + 556, + 629 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 546, + 556, + 629 + ], + "spans": [ + { + "bbox": [ + 313, + 546, + 556, + 629 + ], + "type": "text", + "content": "To address these challenges, some methods incorporate OCR-extracted text directly into language model prompts [8, 17]. However, this approach typically lags behind OCR-free methods with full-resolution, as it fails to capture crucial visual context and layout information [54]. Moreover, it often produces long sequence inputs, increasing latency and computational costs, especially for dense documents." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 313, + 630, + 556, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 630, + 556, + 714 + ], + "spans": [ + { + "bbox": [ + 313, + 630, + 556, + 714 + ], + "type": "text", + "content": "Alternatively, recent VLMs [20, 31, 53] have introduced specialized mechanisms to reduce visual token count, such as image resizing, tiling limits, and feature downsampling. However, when applied to document understanding, these methods result in significant performance degradation, creating an undesirable trade-off between computational efficiency and accuracy, as demonstrated in Fig. 1. These lim" + } + ] + } + ], + "index": 13 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "spans": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "type": "text", + "content": "CVF" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 145, + 0, + 496, + 37 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 145, + 0, + 496, + 37 + ], + "spans": [ + { + "bbox": [ + 145, + 0, + 496, + 37 + ], + "type": "text", + "content": "This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 66, + 693, + 289, + 703 + ], + "type": "page_footnote", + "angle": 0, + "lines": [ + { + "bbox": [ + 66, + 693, + 289, + 703 + ], + "spans": [ + { + "bbox": [ + 66, + 693, + 289, + 703 + ], + "type": "text", + "content": "*Equal contribution. Corresponding author: mnacson@amazon.com." + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 67, + 703, + 205, + 712 + ], + "type": "page_footnote", + "angle": 0, + "lines": [ + { + "bbox": [ + 67, + 703, + 205, + 712 + ], + "spans": [ + { + "bbox": [ + 67, + 703, + 205, + 712 + ], + "type": "inline_equation", + "content": "^{\\dagger}" + }, + { + "bbox": [ + 67, + 703, + 205, + 712 + ], + "type": "text", + "content": "Work done during an Amazon internship." + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "text", + "content": "29005" + } + ] + } + ], + "index": 17 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 0 + }, + { + "para_blocks": [ + { + "bbox": [ + 55, + 72, + 294, + 108 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 72, + 294, + 108 + ], + "spans": [ + { + "bbox": [ + 55, + 72, + 294, + 108 + ], + "type": "text", + "content": "itations underscore the need for a more efficient approach that maintains high-performance document understanding while reducing computational demands." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 54, + 109, + 294, + 228 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 54, + 109, + 294, + 228 + ], + "spans": [ + { + "bbox": [ + 54, + 109, + 294, + 228 + ], + "type": "text", + "content": "To overcome these limitations, we introduce DocVLM, a model-agnostic method that enhances VLMs' reading ability by utilizing OCR information effectively. Our approach employs an OCR encoder to capture both contextual and layout details from OCR-extracted text, compressing these encodings into 64 learned queries. These queries are projected and fed into the LLM part of the VLM alongside visual features. Unlike some previous methods [5, 19, 30, 32], our compression mechanism avoids the need for a separate compression module or alterations to the LLM architecture." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 56, + 229, + 296, + 410 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 229, + 296, + 410 + ], + "spans": [ + { + "bbox": [ + 56, + 229, + 296, + 410 + ], + "type": "text", + "content": "We demonstrate DocVLM's effectiveness across multiple state-of-the-art VLMs: LlaVA-OneVision [31], InternVL2 [18], and Qwen2-VL [53], each employing a unique image token reduction technique. As illustrated in Figure 1, our method significantly improves performance, particularly in low input token regimes. Our experiments demonstrate consistently, across all studied VLMs and visual token budgets, that DocVLM's OCR encoder not only outperforms the baseline of inserting OCR words into VLMs but also achieves this superior performance when compressed to just 64 tokens. This dual advantage of improved performance and reduced token usage allows for better utilization of fixed token budgets, enabling the allocation of more tokens for visual processing and further enhancing overall performance." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 55, + 411, + 295, + 542 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 411, + 295, + 542 + ], + "spans": [ + { + "bbox": [ + 55, + 411, + 295, + 542 + ], + "type": "text", + "content": "Importantly, this reduction in sequence length improves the scalability of our approach, enabling its application to multipage document understanding tasks without additional training. We demonstrate that DocVLM can be seamlessly extended to long-context scenarios, such as multipage DocVQA [50, 51]. While current OCR-free approaches struggle with the overwhelming amount of data in multipage documents [27], our method achieves strong zero-shot performance on DUDE and surpasses the current state-of-the-art results on MP-DocVQA (86.3% vs. 80.3%), despite not being trained on multipage data." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 55, + 543, + 144, + 553 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 543, + 144, + 553 + ], + "spans": [ + { + "bbox": [ + 55, + 543, + 144, + 553 + ], + "type": "text", + "content": "Main contributions:" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 55, + 556, + 295, + 712 + ], + "type": "list", + "angle": 0, + "index": 9, + "blocks": [ + { + "bbox": [ + 55, + 556, + 295, + 592 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 556, + 295, + 592 + ], + "spans": [ + { + "bbox": [ + 55, + 556, + 295, + 592 + ], + "type": "text", + "content": "- DocVLM, a model-agnostic method that efficiently integrates OCR information into VLMs, capturing both text and layout without complex integration techniques." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 55, + 592, + 295, + 627 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 592, + 295, + 627 + ], + "spans": [ + { + "bbox": [ + 55, + 592, + 295, + 627 + ], + "type": "text", + "content": "- A compression mechanism that reduces OCR data into a compact set of typically 64 learned queries, significantly reducing computational overhead." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 55, + 628, + 295, + 675 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 628, + 295, + 675 + ], + "spans": [ + { + "bbox": [ + 55, + 628, + 295, + 675 + ], + "type": "text", + "content": "- Demonstration of DocVLM's effectiveness across different VLM architectures, LlaVA-OneVision, InternVL2, and Qwen2-VL, showing significant performance improvements in low input token regimes (448×448)." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 55, + 675, + 295, + 712 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 675, + 295, + 712 + ], + "spans": [ + { + "bbox": [ + 55, + 675, + 295, + 712 + ], + "type": "text", + "content": "- Extension of DocVLM to long-context tasks, achieving strong zero-shot performance on DUDE and SOTA results on MP-DocVQA without multipage training data." + } + ] + } + ], + "index": 8 + } + ], + "sub_type": "text" + }, + { + "bbox": [ + 314, + 71, + 402, + 83 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 314, + 71, + 402, + 83 + ], + "spans": [ + { + "bbox": [ + 314, + 71, + 402, + 83 + ], + "type": "text", + "content": "2. Related Work" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 313, + 95, + 553, + 311 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 95, + 553, + 311 + ], + "spans": [ + { + "bbox": [ + 313, + 95, + 553, + 311 + ], + "type": "text", + "content": "OCR-free Document VLMs. Early VLMs[5, 9, 22, 23, 32, 36, 44, 59, 61] used relatively small image sizes (e.g., " + }, + { + "bbox": [ + 313, + 95, + 553, + 311 + ], + "type": "inline_equation", + "content": "224 \\times 224" + }, + { + "bbox": [ + 313, + 95, + 553, + 311 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 313, + 95, + 553, + 311 + ], + "type": "inline_equation", + "content": "336 \\times 336" + }, + { + "bbox": [ + 313, + 95, + 553, + 311 + ], + "type": "text", + "content": "), performing well on natural-image tasks but falling short in document understanding. To address this, recent approaches enhance document understanding capabilities by operating on high-resolution images, developing various strategies to manage the resulting computational burden. Direct processing methods like Donut [29], PaLI-X [16], and Qwen2-VL [53] attempt full-resolution processing but often resort to image resizing for computational feasibility. Tile-based approaches such as UReader [56] and InternVL2 [18] improve efficiency by processing image tiles independently. Other methods, exemplified by LLaVA-1.5 [31] and LLaVA-OneVision [35], process full-scale images as tiles but downsample the resulting visual features. While these approaches offer different trade-offs, our experiments show their performance deteriorates significantly when constrained to fewer visual tokens." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 313, + 315, + 553, + 495 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 315, + 553, + 495 + ], + "spans": [ + { + "bbox": [ + 313, + 315, + 553, + 495 + ], + "type": "text", + "content": "OCR-Enhanced Document Understanding. The widespread availability of efficient, open-source OCR models and cost-effective commercial solutions has driven broad adoption of OCR-based approaches in document understanding [1-4, 21, 28, 34, 43, 46, 57, 58]. Several recent works [9, 16, 17, 19] have explored integrating OCR systems with VLMs by feeding extracted text directly into the language model component. Some approaches further enhance this integration by incorporating spatial layout information [12, 14, 22, 52, 54]. While these methods reduce the computational burden of processing high-resolution images, they currently lag behind OCR-free approaches in performance. Additionally, they face challenges with lengthy input sequences, particularly in multipage settings, which can increase latency and computational costs." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 313, + 498, + 553, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 498, + 553, + 714 + ], + "spans": [ + { + "bbox": [ + 313, + 498, + 553, + 714 + ], + "type": "text", + "content": "Document Representation Compression. To address efficiency challenges in processing documents, various compression techniques have been developed. For OCR-enhanced approaches, [14] proposed compressing the OCR signal in multi-page documents using a Compression Transformer, which, despite improving performance in multi-page benchmarks, introduces significant complexity to the system. In the OCR-free setting, generic VLM approaches like Q-former [32] and Resampler [5] compress visual features but struggle with text-dense images. Document-specific methods such as TokenPacker [33] and DocCompressor [27] achieve effective visual compression but show reduced performance on document understanding tasks. In contrast, rather than compressing high-resolution visual inputs, our DocVLM method operates on lower-resolution images and compresses the extensive OCR signal, including textual and layout information, into a compact set of features (typically 64)." + } + ] + } + ], + "index": 13 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "type": "text", + "content": "29006" + } + ] + } + ], + "index": 14 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 1 + }, + { + "para_blocks": [ + { + "bbox": [ + 56, + 71, + 135, + 83 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 71, + 135, + 83 + ], + "spans": [ + { + "bbox": [ + 56, + 71, + 135, + 83 + ], + "type": "text", + "content": "3. Our Method" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 55, + 91, + 296, + 175 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 91, + 296, + 175 + ], + "spans": [ + { + "bbox": [ + 55, + 91, + 296, + 175 + ], + "type": "text", + "content": "We present DocVLM, a model-agnostic approach that enhances VLMs' reading capabilities, enabling operation with lower-resolution inputs while maintaining or improving document understanding accuracy. Our design preserves the base VLM weights, facilitating easy integration across different model architectures and providing flexibility to balance OCR and visual tokens during inference." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 55, + 182, + 138, + 194 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 182, + 138, + 194 + ], + "spans": [ + { + "bbox": [ + 55, + 182, + 138, + 194 + ], + "type": "text", + "content": "3.1. Architecture" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 55, + 199, + 296, + 294 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 199, + 296, + 294 + ], + "spans": [ + { + "bbox": [ + 55, + 199, + 296, + 294 + ], + "type": "text", + "content": "Our method introduces two main components that complement existing VLM architectures: an OCR encoder that processes OCR extracted text and layout information, and a query compression mechanism that distills this information into a compact representation. We integrate these components with pre-trained VLMs, which employ various strategies to control the number of visual tokens for efficient processing. Figure 2 illustrates the overall architecture." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 55, + 300, + 296, + 433 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 300, + 296, + 433 + ], + "spans": [ + { + "bbox": [ + 55, + 300, + 296, + 433 + ], + "type": "text", + "content": "OCR Encoder Architecture We utilize DocFormerV2 [7], a T5-based encoder-decoder [45] designed for document understanding, which incorporates vision, language, and spatial features. Specifically, we leverage only the encoder component, which comprises 344 million parameters, and omit its visual branch to eliminate redundancy with the VLM's vision capabilities and reduce computational complexity. The encoder processes two types of inputs: user instructions and OCR data from an OCR system, which consists of textual tokens and their corresponding 2D positional information [6, 7, 12, 14, 22, 26]." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 55, + 438, + 296, + 605 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 438, + 296, + 605 + ], + "spans": [ + { + "bbox": [ + 55, + 438, + 296, + 605 + ], + "type": "text", + "content": "Query Compression Mechanism To efficiently integrate OCR information into VLMs, we introduce an instruction-aware compression mechanism that distills the OCR encoder's output into a compact set of learned queries. This mechanism significantly reduces the input sequence length for the language model while preserving essential document information. The compression process utilizes " + }, + { + "bbox": [ + 55, + 438, + 296, + 605 + ], + "type": "inline_equation", + "content": "M" + }, + { + "bbox": [ + 55, + 438, + 296, + 605 + ], + "type": "text", + "content": " learnable queries " + }, + { + "bbox": [ + 55, + 438, + 296, + 605 + ], + "type": "inline_equation", + "content": "\\mathbf{Q}" + }, + { + "bbox": [ + 55, + 438, + 296, + 605 + ], + "type": "text", + "content": " (typically " + }, + { + "bbox": [ + 55, + 438, + 296, + 605 + ], + "type": "inline_equation", + "content": "M = 64" + }, + { + "bbox": [ + 55, + 438, + 296, + 605 + ], + "type": "text", + "content": "), initialized randomly following the OCR encoder embeddings' distribution. These queries are processed by the OCR encoder alongside two types of embeddings: OCR embeddings " + }, + { + "bbox": [ + 55, + 438, + 296, + 605 + ], + "type": "inline_equation", + "content": "(\\mathbf{E}_{\\mathrm{OCR}})" + }, + { + "bbox": [ + 55, + 438, + 296, + 605 + ], + "type": "text", + "content": ", which encode both OCR tokens and their bounding boxes, and instruction embeddings " + }, + { + "bbox": [ + 55, + 438, + 296, + 605 + ], + "type": "inline_equation", + "content": "(\\mathbf{E}_{\\mathrm{Instructions}})" + }, + { + "bbox": [ + 55, + 438, + 296, + 605 + ], + "type": "text", + "content": ". The encoding process can be represented as:" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 107, + 611, + 242, + 624 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 107, + 611, + 242, + 624 + ], + "spans": [ + { + "bbox": [ + 107, + 611, + 242, + 624 + ], + "type": "interline_equation", + "content": "\\operatorname {E n c o d e r} \\left(\\left[ \\mathbf {E} _ {\\mathrm {O C R}}, \\mathbf {E} _ {\\text {I n s t r u c t i o n s}}, \\mathbf {Q} \\right]\\right).", + "image_path": "bd9f2eb03ce36b80b87b08c543a134d2eb782804026dbd7a32cc280a759037a1.jpg" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 55, + 629, + 296, + 724 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 629, + 296, + 724 + ], + "spans": [ + { + "bbox": [ + 55, + 629, + 296, + 724 + ], + "type": "text", + "content": "From the encoder output, we retain only the " + }, + { + "bbox": [ + 55, + 629, + 296, + 724 + ], + "type": "inline_equation", + "content": "M" + }, + { + "bbox": [ + 55, + 629, + 296, + 724 + ], + "type": "text", + "content": " features corresponding to the learned queries. These compressed features are then projected to match the VLM's hidden dimension and concatenated with the visual tokens before entering the language model. This compression significantly reduces the LLM's input sequence length, enabling either more efficient processing or, under a fixed token budget, allocation of additional tokens to visual features." + } + ] + } + ], + "index": 7 + }, + { + "type": "image", + "bbox": [ + 319, + 71, + 552, + 292 + ], + "blocks": [ + { + "bbox": [ + 319, + 71, + 552, + 292 + ], + "lines": [ + { + "bbox": [ + 319, + 71, + 552, + 292 + ], + "spans": [ + { + "bbox": [ + 319, + 71, + 552, + 292 + ], + "type": "image", + "image_path": "554d8389308a98dbf09ffbb9088233db940c4846db8b657ef8958463d5206fc3.jpg" + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 313, + 302, + 555, + 369 + ], + "lines": [ + { + "bbox": [ + 313, + 302, + 555, + 369 + ], + "spans": [ + { + "bbox": [ + 313, + 302, + 555, + 369 + ], + "type": "text", + "content": "Figure 2. DocVLM Architecture. DocVLM enhances document understanding in frozen VLMs by integrating an OCR module with a query compression mechanism. By condensing OCR data into " + }, + { + "bbox": [ + 313, + 302, + 555, + 369 + ], + "type": "inline_equation", + "content": "\\mathbf{M} = 64" + }, + { + "bbox": [ + 313, + 302, + 555, + 369 + ], + "type": "text", + "content": " learnable tokens, DocVLM effectively complements visual information, surpassing the VLM's inherent approaches of increasing image resolution or visual feature dimensions." + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "image_caption" + } + ], + "index": 8 + }, + { + "bbox": [ + 313, + 378, + 554, + 450 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 378, + 554, + 450 + ], + "spans": [ + { + "bbox": [ + 313, + 378, + 554, + 450 + ], + "type": "text", + "content": "Vision Process OCR-free VLMs employ different visual processing methods and various strategies to control the number of visual tokens, aiming to reduce the computational cost of processing high-resolution images needed for document understanding. These approaches can be grouped into three main paradigms:" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 315, + 451, + 556, + 714 + ], + "type": "list", + "angle": 0, + "index": 14, + "blocks": [ + { + "bbox": [ + 315, + 451, + 555, + 544 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 315, + 451, + 555, + 544 + ], + "spans": [ + { + "bbox": [ + 315, + 451, + 555, + 544 + ], + "type": "text", + "content": "1. Full Image Processing with Image Resizing (e.g., Qwen2-VL [53]): In this approach, the model processes the entire image as a single input, controlling the number of visual tokens by resizing the image to a fixed or range-constrained image resolutions. While this image processing preserves global context, it incurs quadratic computational complexity with respect to the number of visual tokens." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 315, + 546, + 556, + 689 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 315, + 546, + 556, + 689 + ], + "spans": [ + { + "bbox": [ + 315, + 546, + 556, + 689 + ], + "type": "text", + "content": "2. Patch-Based Processing with Controlled Tile Count (e.g., InternVL2 [18]): This strategy segments the input image into spatial tiles, processing each independently, and controls the number of visual tokens by limiting the number of tiles. While most implementations incorporate a low-resolution global view, the primary focus on local processing may compromise global context understanding. The computational complexity for this approach scales linearly with respect to the number of tiles as the image increase. The result is improved memory efficiency compared to processing the full image at once, especially for large images." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 315, + 689, + 555, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 315, + 689, + 555, + 714 + ], + "spans": [ + { + "bbox": [ + 315, + 689, + 555, + 714 + ], + "type": "text", + "content": "3. Full-Scale Processing with Feature Downsampling (e.g., LlaVA-OneVision [35]): Some VLMs initially" + } + ] + } + ], + "index": 13 + } + ], + "sub_type": "text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "text", + "content": "29007" + } + ] + } + ], + "index": 15 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 2 + }, + { + "para_blocks": [ + { + "bbox": [ + 67, + 72, + 294, + 133 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 67, + 72, + 294, + 133 + ], + "spans": [ + { + "bbox": [ + 67, + 72, + 294, + 133 + ], + "type": "text", + "content": "process full-scale images, but then downsample the visual features into maximal token count before feeding them to the LLM. While this method captures both global and local context, it introduces significant computational overhead during initial full-scale processing." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 55, + 134, + 294, + 194 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 134, + 294, + 194 + ], + "spans": [ + { + "bbox": [ + 55, + 134, + 294, + 194 + ], + "type": "text", + "content": "Experimental results confirm that our OCR query compression mechanism significantly enhances document understanding capabilities across all three visual processing strategies, demonstrating its effectiveness as a universal enhancement to existing VLM architectures." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 55, + 199, + 162, + 213 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 199, + 162, + 213 + ], + "spans": [ + { + "bbox": [ + 55, + 199, + 162, + 213 + ], + "type": "text", + "content": "3.2. Training Strategy" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 55, + 217, + 295, + 312 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 217, + 295, + 312 + ], + "spans": [ + { + "bbox": [ + 55, + 217, + 295, + 312 + ], + "type": "text", + "content": "Our training strategy aims to integrate the OCR modality into existing VLMs while preserving their core strengths. To achieve this, we keep the VLM completely frozen throughout the process and train only the newly introduced OCR components, i.e., the learnable queries, the OCR encoder, and the projection layer. We employ a two-stage training strategy to gradually incorporate the OCR modality into the pre-trained VLM:" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 55, + 314, + 295, + 469 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 314, + 295, + 469 + ], + "spans": [ + { + "bbox": [ + 55, + 314, + 295, + 469 + ], + "type": "text", + "content": "Stage I: OCR-LLM Alignment. During this stage, we withhold image input from the VLM, forcing the model to rely solely on the newly introduced OCR modality. This approach ensures full utilization of OCR data, aligns OCR components with the LLM input space, and reduces sequence length, improving training efficiency. Given the focus on text input, our dataset selection concentrates on text-related tasks. We begin by training only the randomly initialized components: the learnable queries and projection layer. This allows these components to adapt without disrupting the pretrained OCR encoder. Subsequently, we unfreeze the OCR encoder for fine-tuning, enabling a more comprehensive alignment of the entire encoder to the VLM." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 55, + 470, + 295, + 613 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 470, + 295, + 613 + ], + "spans": [ + { + "bbox": [ + 55, + 470, + 295, + 613 + ], + "type": "text", + "content": "Stage II: Vision Alignment. In this final stage, we incorporate visual information extracted from the visual encoder, encouraging the OCR components to complement the visual features. Our experiments reveal this stage has a particularly strong effect when using fewer learned queries, allowing the compressed OCR information to better supplement the information acquired from the visual modality (see Sec. 5). During this stage, we add more visually focused datasets to the training process. Note that despite our method preserving the original VLM weights, it might implicitly inject bias through prompt tuning. To avoid this, the training data should represent all tasks of interest." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 55, + 624, + 226, + 636 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 624, + 226, + 636 + ], + "spans": [ + { + "bbox": [ + 55, + 624, + 226, + 636 + ], + "type": "text", + "content": "3.3. Multipage Document Extension" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 55, + 642, + 295, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 642, + 295, + 713 + ], + "spans": [ + { + "bbox": [ + 55, + 642, + 295, + 713 + ], + "type": "text", + "content": "We conduct our training procedure on single-page data only. However, our approach can be extended to operate on multipage documents. Given a multipage input and its OCR information, the VLM independently processes each page image and concatenates the resulting visual features. For the OCR information, we explore two strategies: Global" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 313, + 72, + 553, + 167 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 72, + 553, + 167 + ], + "spans": [ + { + "bbox": [ + 313, + 72, + 553, + 167 + ], + "type": "text", + "content": "Encoding, which compresses the entire document's OCR information into 64 learnable queries, and Page-wise Encoding, which compresses each page's OCR information separately into 64 learnable queries and then concatenates them, resulting in " + }, + { + "bbox": [ + 313, + 72, + 553, + 167 + ], + "type": "inline_equation", + "content": "64 \\times" + }, + { + "bbox": [ + 313, + 72, + 553, + 167 + ], + "type": "text", + "content": " number of pages learned queries. After processing the OCR information using either strategy, we feed the resulting compressed OCR features along with the concatenated visual features into the LLM." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 313, + 168, + 553, + 251 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 168, + 553, + 251 + ], + "spans": [ + { + "bbox": [ + 313, + 168, + 553, + 251 + ], + "type": "text", + "content": "Our experiments demonstrate that both approaches are highly effective and efficient in processing multipage documents. Using either approach with a restricted number of visual tokens, we obtain strong zero-shot results on DUDE [51] and state-of-the-art results on MP-DocVQA [50]. The page-wise encoding strategy yields slightly better results for lower numbers of visual tokens." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 313, + 265, + 395, + 277 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 265, + 395, + 277 + ], + "spans": [ + { + "bbox": [ + 313, + 265, + 395, + 277 + ], + "type": "text", + "content": "4. Experiments" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 313, + 285, + 437, + 298 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 285, + 437, + 298 + ], + "spans": [ + { + "bbox": [ + 313, + 285, + 437, + 298 + ], + "type": "text", + "content": "4.1. Experimental Setting" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 313, + 302, + 553, + 374 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 302, + 553, + 374 + ], + "spans": [ + { + "bbox": [ + 313, + 302, + 553, + 374 + ], + "type": "text", + "content": "Model Integration: We evaluate DocVLM through integration with three leading open-source VLMs: LLaVA-OneVision [31], InternVL2 [18], and Qwen2-VL [53]. As discussed in Sec. 3.1, these models employ distinct token reduction strategies, enabling us to assess DocVLM's effectiveness across different visual processing approaches." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 313, + 375, + 553, + 482 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 375, + 553, + 482 + ], + "spans": [ + { + "bbox": [ + 313, + 375, + 553, + 482 + ], + "type": "text", + "content": "Training: Our training protocol, detailed in Sec. 3.2, employs a two-phase strategy. The initial phase focuses on text-centric tasks using datasets spanning document understanding (DocVQA [40], InfoVQA [41]), scene text analysis (ST-VQA [11], TextVQA [48], OCR-VQA [42]), and specialized tasks (ChartQA [39], TextCaps [47], TATDQA [62]). The subsequent vision alignment phase incorporates additional visual-centric datasets: COCO Caption [15] and VQA-V2 [24]." + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 313, + 483, + 553, + 662 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 483, + 553, + 662 + ], + "spans": [ + { + "bbox": [ + 313, + 483, + 553, + 662 + ], + "type": "text", + "content": "Evaluation: For evaluation, we focus on five key benchmarks: DocVQA, TextVQA, ST-VQA, InfoVQA, and TextCaps. Results are reported on test sets where available, with TextVQA and TextCaps evaluated on validation sets due to test server restrictions. We use ANLS as the evaluation metric for all datasets, except TextVQA, which uses VQAScore, and TextCaps, which uses CIDEr. To demonstrate DocVLM's generalization capabilities, we conduct zero-shot evaluation on multipage document understanding benchmarks: DUDE [51] and MP-DocVQA [50]. This zero-shot performance is particularly noteworthy as our model is trained exclusively on single-page documents. The supplementary presents additional implementation details, including hyperparameters and optimization strategies, and robustness analysis to different OCR systems." + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 313, + 671, + 474, + 685 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 671, + 474, + 685 + ], + "spans": [ + { + "bbox": [ + 313, + 671, + 474, + 685 + ], + "type": "text", + "content": "4.2. State-of-the-art Comparisons" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 313, + 689, + 553, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 689, + 553, + 713 + ], + "spans": [ + { + "bbox": [ + 313, + 689, + 553, + 713 + ], + "type": "text", + "content": "Table 1 presents comprehensive comparisons between DocVLM and other state-of-the-art methods across var" + } + ] + } + ], + "index": 16 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "text", + "content": "29008" + } + ] + } + ], + "index": 17 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 3 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 59, + 70, + 555, + 436 + ], + "blocks": [ + { + "bbox": [ + 59, + 70, + 555, + 436 + ], + "lines": [ + { + "bbox": [ + 59, + 70, + 555, + 436 + ], + "spans": [ + { + "bbox": [ + 59, + 70, + 555, + 436 + ], + "type": "table", + "html": "
Method# Tok.#PDocVQATextVQAST-VQAInfoVQATextCAPSMP-DocVQADUDE*
No Token Limitations
GPT-4o92.877.4-79.2---
Gemini 1.5 Pro93.178.7-81.0---
GPT-4V87.278.0-75.1---
KOSMOS-2.5-CHAT4K1.3B81.140.7-41.3---
TextSquare2.5K8.6B84.366.8-51.5---
ScreenAI3.5K5B87.8--57.8-72.9-
ScreenAI+OCR4.3K5B89.9--65.9-77.1-
Pali-35.5K5B86.779.584.157.8158.8--
Pali-3+OCR6.3K5B88.680.885.762.4164.3--
# Tokens ≤ 1.5k
URader8417B65.457.6-42.2118.4--
Monkey1.3K9B66.564.3-36.193.2--
TextMonkey7689B73.065.9-28.6---
Vary2567B76.3------
DocOwl23248B80.766.7-46.4131.869.446.8
GRAM9001B85.3----80.351.2
GRAMC-Former2561B87.6----77.645.5
DocFormer v21K1B87.864.071.848.8-76.448.4
LLaVA-OneVision7K7B87.576.171.168.8138.0OOMOOM
LLaVA-OneVision1.5K7B66.572.170.645.6112.941.828.7
DocVLMLLaVA-OneVision (Ours)1.5K7B88.476.970.861.0145.377.943.8
InternVL 23.1K8B91.677.4-74.8-OOMOOM
InternVL 22568B56.065.765.738.451.151.030.5
DocVLMInternVL2 (Ours)3208B86.671.274.357.6119.476.243.3
InternVL 212808B85.775.568.361.543.778.142.2
DocVLMInternVL2 (Ours)13448B91.076.776.765.4123.481.845.6
Qwen2-VL16k7B94.584.370.776.5150.2OOMOOM
Qwen2-VL3207B84.478.070.154.1142.173.041.5
DocVLMQwen2-VL (Ours)3207B91.279.676.561.2144.381.746.1
Qwen2-VL5767B91.582.370.565.3145.082.145.9
DocVLMQwen2-VL (Ours)5767B92.882.879.866.8150.484.547.4
", + "image_path": "04c5ded42318b47c2743b294d6086a1e819832a43461dd929fbd215199f96e33.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "table_body" + } + ], + "index": 0 + }, + { + "bbox": [ + 55, + 443, + 555, + 499 + ], + "lines": [ + { + "bbox": [ + 55, + 443, + 555, + 499 + ], + "spans": [ + { + "bbox": [ + 55, + 443, + 555, + 499 + ], + "type": "text", + "content": "Table 1. Comparison with State-of-the-Art Methods. Performance evaluation of DocVLM against state-of-the-art approaches on document understanding benchmarks. Results are categorized into unconstrained models and those with a 1.5k token limit. In the constrained token regime, DocVLM consistently enhances the performance of baseline VLMs across various tasks and visual token budgets. Notably, DocVLM paired with Qwen2-VL (576 tokens) achieves superior performance across all evaluated datasets, including state-of-the-art zero-shot accuracy on DUDE. " + }, + { + "bbox": [ + 55, + 443, + 555, + 499 + ], + "type": "inline_equation", + "content": "\\star" + }, + { + "bbox": [ + 55, + 443, + 555, + 499 + ], + "type": "text", + "content": " indicates zero-shot evaluation, with grey entries denoting non-zero-shot results." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "text" + }, + { + "bbox": [ + 55, + 506, + 295, + 602 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 506, + 295, + 602 + ], + "spans": [ + { + "bbox": [ + 55, + 506, + 295, + 602 + ], + "type": "text", + "content": "ious document understanding benchmarks, highlighting DocVLM's ability to improve performance under token constraints. We categorize the results into two main groups: methods without token constraints (both closed and open-source models) and those operating under a 1.5k token limit. We mainly focused on models with around 7B parameters and include methods using an OCR system such as Pali3 [10] ScreenAI [8], DocFormerV2 [7], and GRAM [14]." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 55, + 605, + 295, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 605, + 295, + 715 + ], + "spans": [ + { + "bbox": [ + 55, + 605, + 295, + 715 + ], + "type": "text", + "content": "To evaluate DocVLM's effectiveness under token constraints, we integrated it with three baseline models: LlaVA-OneVision [31], InternVL2 [18], and Qwen2-VL [53], each configured to operate within the 1.5k token limit. For LlaVA-OneVision, we utilized the minimal visual token configuration (single visual features tile). InternVL2 was tested with both single-tile (256 tokens) and four-tile (1280 tokens) configurations, while Qwen2-VL was evaluated with 256 and 512 visual tokens, corresponding to image" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 313, + 506, + 506, + 518 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 506, + 506, + 518 + ], + "spans": [ + { + "bbox": [ + 313, + 506, + 506, + 518 + ], + "type": "text", + "content": "sizes of " + }, + { + "bbox": [ + 313, + 506, + 506, + 518 + ], + "type": "inline_equation", + "content": "448 \\times 448" + }, + { + "bbox": [ + 313, + 506, + 506, + 518 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 313, + 506, + 506, + 518 + ], + "type": "inline_equation", + "content": "616 \\times 616" + }, + { + "bbox": [ + 313, + 506, + 506, + 518 + ], + "type": "text", + "content": " respectively." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 312, + 520, + 555, + 663 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 312, + 520, + 555, + 663 + ], + "spans": [ + { + "bbox": [ + 312, + 520, + 555, + 663 + ], + "type": "text", + "content": "Under the practical 1.5k token constraint, incorporating DocVLM with each of these baseline models yields substantial and consistent improvements. Notably, these improvements persist even in looser token regimes, such as InternVL2 with 1280 visual tokens and Qwen2-VL with 576 visual tokens. Within this constraint, our Qwen2-VL variant with DocVLM, using just 576 tokens, achieves state-of-the-art performance across all benchmarks: " + }, + { + "bbox": [ + 312, + 520, + 555, + 663 + ], + "type": "inline_equation", + "content": "92.8\\%" + }, + { + "bbox": [ + 312, + 520, + 555, + 663 + ], + "type": "text", + "content": " on DocVQA, " + }, + { + "bbox": [ + 312, + 520, + 555, + 663 + ], + "type": "inline_equation", + "content": "82.8\\%" + }, + { + "bbox": [ + 312, + 520, + 555, + 663 + ], + "type": "text", + "content": " on TextVQA, " + }, + { + "bbox": [ + 312, + 520, + 555, + 663 + ], + "type": "inline_equation", + "content": "79.8\\%" + }, + { + "bbox": [ + 312, + 520, + 555, + 663 + ], + "type": "text", + "content": " on ST-VQA, " + }, + { + "bbox": [ + 312, + 520, + 555, + 663 + ], + "type": "inline_equation", + "content": "66.8\\%" + }, + { + "bbox": [ + 312, + 520, + 555, + 663 + ], + "type": "text", + "content": " on InfoVQA, and a CIDEr score of 150.4 on TextCAPS. As detailed in the Appendix, DocVLM achieves these results with significant computational advantages." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 313, + 665, + 556, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 665, + 556, + 715 + ], + "spans": [ + { + "bbox": [ + 313, + 665, + 556, + 715 + ], + "type": "text", + "content": "DocVLM also demonstrates exceptional capability in handling multipage documents. Using the same 576-token configuration described earlier, it achieves " + }, + { + "bbox": [ + 313, + 665, + 556, + 715 + ], + "type": "inline_equation", + "content": "84.5\\%" + }, + { + "bbox": [ + 313, + 665, + 556, + 715 + ], + "type": "text", + "content": " accuracy on MP-DocVQA, surpassing previous state-of-the-art re" + } + ] + } + ], + "index": 6 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "type": "text", + "content": "29009" + } + ] + } + ], + "index": 7 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 4 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 60, + 71, + 298, + 215 + ], + "blocks": [ + { + "bbox": [ + 60, + 71, + 298, + 215 + ], + "lines": [ + { + "bbox": [ + 60, + 71, + 298, + 215 + ], + "spans": [ + { + "bbox": [ + 60, + 71, + 298, + 215 + ], + "type": "image", + "image_path": "c79c75b43eab3846f46dea7e2283a6ba1c12a778a56419801d1d40e808b72f2d.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + } + ], + "index": 0 + }, + { + "type": "image", + "bbox": [ + 58, + 227, + 298, + 370 + ], + "blocks": [ + { + "bbox": [ + 58, + 227, + 298, + 370 + ], + "lines": [ + { + "bbox": [ + 58, + 227, + 298, + 370 + ], + "spans": [ + { + "bbox": [ + 58, + 227, + 298, + 370 + ], + "type": "image", + "image_path": "f06f08792c330366c71a07d59eacc3a42ea1b996951bb686181e954aeadfa07a.jpg" + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 55, + 378, + 555, + 413 + ], + "lines": [ + { + "bbox": [ + 55, + 378, + 555, + 413 + ], + "spans": [ + { + "bbox": [ + 55, + 378, + 555, + 413 + ], + "type": "text", + "content": "Figure 3. Qualitative Results. Representative examples of DocVLM's performance across diverse document formats, from dense text to infographics and scene text. Our model successfully handles complex layouts, dense content, and presents instruction-following capabilities without explicit training on such datasets. Each example includes an image-instruction pair with baseline and DocVLM predictions." + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_caption" + } + ], + "index": 1 + }, + { + "type": "image", + "bbox": [ + 301, + 68, + 552, + 227 + ], + "blocks": [ + { + "bbox": [ + 301, + 68, + 552, + 227 + ], + "lines": [ + { + "bbox": [ + 301, + 68, + 552, + 227 + ], + "spans": [ + { + "bbox": [ + 301, + 68, + 552, + 227 + ], + "type": "image", + "image_path": "cf27e73cc1d6f78ec8df28502c3f840cca7bd06203ebfe8b2dc9b7a0c2d44593.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + } + ], + "index": 2 + }, + { + "type": "image", + "bbox": [ + 302, + 230, + 552, + 371 + ], + "blocks": [ + { + "bbox": [ + 302, + 230, + 552, + 371 + ], + "lines": [ + { + "bbox": [ + 302, + 230, + 552, + 371 + ], + "spans": [ + { + "bbox": [ + 302, + 230, + 552, + 371 + ], + "type": "image", + "image_path": "f77db2901218e23d6ed4ad97896dd025f8c728bc9de19b73954b498ab3344c3e.jpg" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_body" + } + ], + "index": 3 + }, + { + "bbox": [ + 55, + 424, + 296, + 473 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 424, + 296, + 473 + ], + "spans": [ + { + "bbox": [ + 55, + 424, + 296, + 473 + ], + "type": "text", + "content": "sults. Furthermore, this setup shows robust generalization capabilities with a zero-shot performance of " + }, + { + "bbox": [ + 55, + 424, + 296, + 473 + ], + "type": "inline_equation", + "content": "47.4\\%" + }, + { + "bbox": [ + 55, + 424, + 296, + 473 + ], + "type": "text", + "content": " accuracy on the DUDE dataset, despite not being specifically trained for multipage document processing." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 55, + 495, + 167, + 507 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 495, + 167, + 507 + ], + "spans": [ + { + "bbox": [ + 55, + 495, + 167, + 507 + ], + "type": "text", + "content": "4.3. Qualitative Results" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 55, + 517, + 296, + 601 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 517, + 296, + 601 + ], + "spans": [ + { + "bbox": [ + 55, + 517, + 296, + 601 + ], + "type": "text", + "content": "Figure 3 illustrates DocVLM's enhanced capabilities through representative examples that demonstrate three key strengths of our method: (1) improved reading comprehension in complex document layouts, (2) effective handling of dense textual content despite using compressed representations, and (3) preserved and enhanced instruction-following capabilities." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 55, + 605, + 296, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 605, + 296, + 715 + ], + "spans": [ + { + "bbox": [ + 55, + 605, + 296, + 715 + ], + "type": "text", + "content": "The examples span diverse document types, from dense text documents to infographics and scene text, showcasing DocVLM's versatility. In the infographic example, DocVLM not only preserves but enhances the base model's instruction-following capabilities, despite our OCR component not being explicitly trained on instruction-following datasets like [49]. This demonstrates that our compression mechanism successfully retains crucial textual and layout information while significantly reducing token usage." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 313, + 423, + 489, + 437 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 423, + 489, + 437 + ], + "spans": [ + { + "bbox": [ + 313, + 423, + 489, + 437 + ], + "type": "text", + "content": "4.4. Scaling to Multipage Documents" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 313, + 442, + 555, + 526 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 442, + 555, + 526 + ], + "spans": [ + { + "bbox": [ + 313, + 442, + 555, + 526 + ], + "type": "text", + "content": "Building on the promising multipage results in Tab. 1, we conduct an in-depth analysis of different DocVLM configurations for multipage document understanding. This analysis focuses on the Qwen2-VL base model, tested on the MP-DocVQA dataset with documents up to 20 pages long – a scale that most other state-of-the-art methods struggle to handle due to token limitations." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 313, + 529, + 555, + 553 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 529, + 555, + 553 + ], + "spans": [ + { + "bbox": [ + 313, + 529, + 555, + 553 + ], + "type": "text", + "content": "Table 2 compares four OCR integration strategies used during inference in multipage scenarios:" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 313, + 554, + 542, + 602 + ], + "type": "list", + "angle": 0, + "index": 16, + "blocks": [ + { + "bbox": [ + 313, + 554, + 525, + 566 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 554, + 525, + 566 + ], + "spans": [ + { + "bbox": [ + 313, + 554, + 525, + 566 + ], + "type": "text", + "content": "- Baseline: vision-only input (no additional tokens)" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 313, + 567, + 542, + 578 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 567, + 542, + 578 + ], + "spans": [ + { + "bbox": [ + 313, + 567, + 542, + 578 + ], + "type": "text", + "content": "- Direct OCR word insertion: up to 800 tokens per page" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 313, + 578, + 481, + 590 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 578, + 481, + 590 + ], + "spans": [ + { + "bbox": [ + 313, + 578, + 481, + 590 + ], + "type": "text", + "content": "- Global OCR encoding: 64 tokens total" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 313, + 590, + 509, + 602 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 590, + 509, + 602 + ], + "spans": [ + { + "bbox": [ + 313, + 590, + 509, + 602 + ], + "type": "text", + "content": "- Page-wise OCR encoding: 64 tokens per page" + } + ] + } + ], + "index": 15 + } + ], + "sub_type": "text" + }, + { + "bbox": [ + 313, + 604, + 555, + 699 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 604, + 555, + 699 + ], + "spans": [ + { + "bbox": [ + 313, + 604, + 555, + 699 + ], + "type": "text", + "content": "Results indicate consistent improvements over the baseline across all image resolutions (256, 512, and 1024 tokens), with only minimal additional token usage – just 64 tokens for the entire document in the global encoding case. Notably, at 256 visual tokens per page, both page-wise encoding " + }, + { + "bbox": [ + 313, + 604, + 555, + 699 + ], + "type": "inline_equation", + "content": "(82.4\\%)" + }, + { + "bbox": [ + 313, + 604, + 555, + 699 + ], + "type": "text", + "content": " and global encoding " + }, + { + "bbox": [ + 313, + 604, + 555, + 699 + ], + "type": "inline_equation", + "content": "(81.7\\%)" + }, + { + "bbox": [ + 313, + 604, + 555, + 699 + ], + "type": "text", + "content": " outperform direct OCR word insertion " + }, + { + "bbox": [ + 313, + 604, + 555, + 699 + ], + "type": "inline_equation", + "content": "(79.1\\%)" + }, + { + "bbox": [ + 313, + 604, + 555, + 699 + ], + "type": "text", + "content": " while using significantly fewer tokens." + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 325, + 701, + 553, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 325, + 701, + 553, + 714 + ], + "spans": [ + { + "bbox": [ + 325, + 701, + 553, + 714 + ], + "type": "text", + "content": "Our best configuration achieves state-of-the-art perfor" + } + ] + } + ], + "index": 18 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "type": "text", + "content": "29010" + } + ] + } + ], + "index": 19 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 5 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 58, + 70, + 294, + 197 + ], + "blocks": [ + { + "bbox": [ + 58, + 70, + 294, + 197 + ], + "lines": [ + { + "bbox": [ + 58, + 70, + 294, + 197 + ], + "spans": [ + { + "bbox": [ + 58, + 70, + 294, + 197 + ], + "type": "table", + "html": "
MethodLLM OCR InputImage Tok.OCR Tok.ANLS
DocOwl2 [27]-324 × pg-69.4
GRAMC-Former [14]-100 × pg25677.6
GRAM [14]-100 × pg800 × pg80.3
Qwen2-VL-256 × pg-73.0
OCR Words256 × pg800 × pg79.1
DocVLMQwen2-VLGlobal Encoding256 × pg6481.7
DocVLMQwen2-VLPage-wise Encoding256 × pg64 × pg82.4
Qwen2-VL-512 × pg-82.1
DocVLMQwen2-VLGlobal Encoding512 × pg6484.5
DocVLMQwen2-VLPage-wise Encoding512 × pg64 × pg85.2
Qwen2-VL-1024 × pg-85.2
DocVLMQwen2-VLGlobal Encoding1024 × pg6486.3
DocVLMQwen2-VLPage-wise Encoding1024 × pg64 × pg86.3
", + "image_path": "5f669879a86d0397d295b876fea90a2fe624c0206f9242fbcd8ca657ef3186ab.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "table_body" + } + ], + "index": 0 + }, + { + "bbox": [ + 55, + 285, + 295, + 344 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 285, + 295, + 344 + ], + "spans": [ + { + "bbox": [ + 55, + 285, + 295, + 344 + ], + "type": "text", + "content": "mance of " + }, + { + "bbox": [ + 55, + 285, + 295, + 344 + ], + "type": "inline_equation", + "content": "86.3\\%" + }, + { + "bbox": [ + 55, + 285, + 295, + 344 + ], + "type": "text", + "content": " ANLS, significantly outperforming specialized multipage models like GRAM " + }, + { + "bbox": [ + 55, + 285, + 295, + 344 + ], + "type": "inline_equation", + "content": "(80.3\\%)" + }, + { + "bbox": [ + 55, + 285, + 295, + 344 + ], + "type": "text", + "content": ". This is particularly impressive considering that DocVLM was trained exclusively on single-page inputs, demonstrating strong zero-shot generalization to multipage scenarios." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 55, + 346, + 295, + 453 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 346, + 295, + 453 + ], + "spans": [ + { + "bbox": [ + 55, + 346, + 295, + 453 + ], + "type": "text", + "content": "The comparison between encoding strategies reveals that page-wise encoding consistently outperforms global encoding at lower visual token counts, providing a " + }, + { + "bbox": [ + 55, + 346, + 295, + 453 + ], + "type": "inline_equation", + "content": "+0.7\\%" + }, + { + "bbox": [ + 55, + 346, + 295, + 453 + ], + "type": "text", + "content": " improvement for both 256 and 512 image tokens per page. This advantage diminishes at 1024 tokens where both achieve identical performance (86.3% ANLS). Remarkably, DocVLM with page-wise encoding matches or even outperforms the baseline Qwen2-VL using twice as many visual tokens, highlighting the efficiency of our approach." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 55, + 465, + 149, + 479 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 465, + 149, + 479 + ], + "spans": [ + { + "bbox": [ + 55, + 465, + 149, + 479 + ], + "type": "text", + "content": "5. Ablation Study" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 55, + 485, + 295, + 604 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 485, + 295, + 604 + ], + "spans": [ + { + "bbox": [ + 55, + 485, + 295, + 604 + ], + "type": "text", + "content": "Impact of OCR Encoding Strategies To evaluate the impact of OCR encoding compression, we compare three strategies for integrating OCR information: (1) inserting raw OCR words in the original VLM, (2) using DocVLM uncompressed OCR encodings, and (3) DocVLM compressed OCR encodings with 64 learned queries. We evaluate these approaches on the DocVQA test set using three representative model configurations: LLaVA-OneVision with 1.5K visual tokens, and both InternVL2 and Qwen2-VL with 256 visual tokens each." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 55, + 605, + 295, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 605, + 295, + 714 + ], + "spans": [ + { + "bbox": [ + 55, + 605, + 295, + 714 + ], + "type": "text", + "content": "Results in Tab. 3 demonstrate that DocVLM's OCR encodings significantly outperform raw OCR words across all three models while maintaining the same token count. Notably, our compressed encoding approach, using just 64 tokens instead of 800 OCR tokens, preserves most of these improvements while drastically reducing sequence length. This efficient compression enables a more favorable allocation of the token budget, allowing models to dedicate more tokens to visual processing without compromising OCR ef" + } + ] + } + ], + "index": 6 + }, + { + "type": "table", + "bbox": [ + 317, + 70, + 553, + 121 + ], + "blocks": [ + { + "bbox": [ + 55, + 205, + 295, + 272 + ], + "lines": [ + { + "bbox": [ + 55, + 205, + 295, + 272 + ], + "spans": [ + { + "bbox": [ + 55, + 205, + 295, + 272 + ], + "type": "text", + "content": "Table 2. Extension to Multipage. Comparison on MP-DocVQA of approaches for incorporating OCR information in multipage document understanding. Both DocVLM multipage extension strategies: global encoding (64 tokens per document) and pagewise encoding (64 tokens per page), outperform previous state-of-the-art methods, notably, without any explicit multipage training." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 317, + 70, + 553, + 121 + ], + "lines": [ + { + "bbox": [ + 317, + 70, + 553, + 121 + ], + "spans": [ + { + "bbox": [ + 317, + 70, + 553, + 121 + ], + "type": "table", + "html": "
LLM OCR InputOCR Tok.LLaVA-OV 1.5KInternVL2 256Qwen2-VL 256
OCR Words80085.884.489.1
OCR Encoding80089.489.291.9
64 Compressed Encoding6488.486.691.2
", + "image_path": "f137ccaa022b3b79d89cfd7bfa6604652b08c4868e0dc00b98ce8191b728aa87.jpg" + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "table_body" + } + ], + "index": 7 + }, + { + "bbox": [ + 313, + 129, + 553, + 162 + ], + "lines": [ + { + "bbox": [ + 313, + 129, + 553, + 162 + ], + "spans": [ + { + "bbox": [ + 313, + 129, + 553, + 162 + ], + "type": "text", + "content": "Table 3. OCR Encoding Strategies. DocVQA results for inserting OCR information using (1) OCR words (baseline), (2) uncompressed OCR encodings, and (3) 64 compressed OCR encodings." + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "text" + }, + { + "bbox": [ + 313, + 182, + 553, + 218 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 182, + 553, + 218 + ], + "spans": [ + { + "bbox": [ + 313, + 182, + 553, + 218 + ], + "type": "text", + "content": "fectiveness. The results validate that DocVLM's compression strategy successfully balances performance with computational efficiency, a key factor for practical applications." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 313, + 251, + 555, + 361 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 251, + 555, + 361 + ], + "spans": [ + { + "bbox": [ + 313, + 251, + 555, + 361 + ], + "type": "text", + "content": "Balancing Vision and OCR Token Allocation Modern VLMs employ various mechanisms to reduce visual token count, creating an inherent trade-off between computational efficiency and model performance, as discussed in Sec. 3.1. We investigate how DocVLM can improve this trade-off by comparing four configurations: (1) baseline VLM without OCR, (2) direct OCR word insertion, (3) DocVLM with uncompressed OCR encodings, and (4) DocVLM with 64 compressed learned queries." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 313, + 363, + 556, + 471 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 363, + 556, + 471 + ], + "spans": [ + { + "bbox": [ + 313, + 363, + 556, + 471 + ], + "type": "text", + "content": "Figure 4 presents the performance scores on DocVQA validation (left y-axis) and total token counts (right y-axis) for three VLM architectures, showing how these metrics vary across different visual token allocations. Each model employs a distinct token reduction approach: LLaVA-OneVision controls token count through feature downsampling (AnyRes Max), InternVL2 limits the number of processed image tiles (Dynamic Max Batch), and Qwen2-VL adjusts image resolution to constrain token count." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 313, + 475, + 556, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 475, + 556, + 713 + ], + "spans": [ + { + "bbox": [ + 313, + 475, + 556, + 713 + ], + "type": "text", + "content": "Our analysis reveals that integrating OCR information, regardless of the method used, consistently improves performance across all models, with particularly pronounced gains in low visual token regimes. However, uncompressed OCR integration methods, whether through direct word insertion or uncompressed DocVLM encodings, require 800 tokens – a significant overhead that could otherwise be allocated to visual processing. For instance, allocating 128 tokens for visual and 800 for OCR in Qwen2-VL achieves " + }, + { + "bbox": [ + 313, + 475, + 556, + 713 + ], + "type": "inline_equation", + "content": "84.3\\%" + }, + { + "bbox": [ + 313, + 475, + 556, + 713 + ], + "type": "text", + "content": " using OCR words and " + }, + { + "bbox": [ + 313, + 475, + 556, + 713 + ], + "type": "inline_equation", + "content": "90.1\\%" + }, + { + "bbox": [ + 313, + 475, + 556, + 713 + ], + "type": "text", + "content": " using uncompressed encodings. In contrast, using 896 pure visual tokens reaches " + }, + { + "bbox": [ + 313, + 475, + 556, + 713 + ], + "type": "inline_equation", + "content": "92.4\\%" + }, + { + "bbox": [ + 313, + 475, + 556, + 713 + ], + "type": "text", + "content": ", demonstrating the potential benefit of allocating more tokens to visual processing. DocVLM's compression mechanism provides a superior option by requiring only 64 tokens for OCR information while maintaining strong performance. In the above example, our approach using fewer tokens, allocating 768 visual tokens and 64 OCR tokens, reaches " + }, + { + "bbox": [ + 313, + 475, + 556, + 713 + ], + "type": "inline_equation", + "content": "93.0\\%" + }, + { + "bbox": [ + 313, + 475, + 556, + 713 + ], + "type": "text", + "content": ", outperforming the " + }, + { + "bbox": [ + 313, + 475, + 556, + 713 + ], + "type": "inline_equation", + "content": "90.1\\%" + }, + { + "bbox": [ + 313, + 475, + 556, + 713 + ], + "type": "text", + "content": " obtained with the uncompressed encodings, highlighting DocVLM's effective balance between visual and OCR tokens." + } + ] + } + ], + "index": 12 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "text", + "content": "29011" + } + ] + } + ], + "index": 13 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 6 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 59, + 71, + 221, + 182 + ], + "blocks": [ + { + "bbox": [ + 59, + 71, + 221, + 182 + ], + "lines": [ + { + "bbox": [ + 59, + 71, + 221, + 182 + ], + "spans": [ + { + "bbox": [ + 59, + 71, + 221, + 182 + ], + "type": "image", + "image_path": "96eb4592a64f4bd6744a858002f3392f1d3dd2945f18b048be3ead0ad81d6f65.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + } + ], + "index": 0 + }, + { + "type": "image", + "bbox": [ + 221, + 71, + 386, + 182 + ], + "blocks": [ + { + "bbox": [ + 221, + 71, + 386, + 182 + ], + "lines": [ + { + "bbox": [ + 221, + 71, + 386, + 182 + ], + "spans": [ + { + "bbox": [ + 221, + 71, + 386, + 182 + ], + "type": "image", + "image_path": "423a8f08b63fd05d073cb4c1fec9eba16c90a79ee2783da1cf9bda63ab362dbb.jpg" + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_body" + } + ], + "index": 1 + }, + { + "type": "image", + "bbox": [ + 386, + 72, + 552, + 182 + ], + "blocks": [ + { + "bbox": [ + 386, + 72, + 552, + 182 + ], + "lines": [ + { + "bbox": [ + 386, + 72, + 552, + 182 + ], + "spans": [ + { + "bbox": [ + 386, + 72, + 552, + 182 + ], + "type": "image", + "image_path": "7898eecf2f11b51e09f0d852cb9383f32525f92a2316d36d6ac74b6c688b1e60.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + } + ], + "index": 2 + }, + { + "type": "image", + "bbox": [ + 157, + 185, + 452, + 206 + ], + "blocks": [ + { + "bbox": [ + 157, + 185, + 452, + 206 + ], + "lines": [ + { + "bbox": [ + 157, + 185, + 452, + 206 + ], + "spans": [ + { + "bbox": [ + 157, + 185, + 452, + 206 + ], + "type": "image", + "image_path": "649344a343f98136978236c1a3dcd4d41574aae6ad50f6203bcbedd27a1402a3.jpg" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 55, + 213, + 555, + 258 + ], + "lines": [ + { + "bbox": [ + 55, + 213, + 555, + 258 + ], + "spans": [ + { + "bbox": [ + 55, + 213, + 555, + 258 + ], + "type": "text", + "content": "Figure 4. Balancing Performance and Compute. Analysis of model performance (lines, left y-axis) and token usage (bars, right y-axis) as a function of visual token allocation. Each model employs its inherent token control strategy: AnyRes max for feature downsampling (LLaVA One-Vision), dynamic max tiles (InternVL2), and max image tokens for resolution control (Qwen2-VL). The results highlight that DocVLM consistently improves performance with minimal overhead (64 tokens), offering an efficient OCR-visual token allocation." + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_caption" + } + ], + "index": 3 + }, + { + "type": "image", + "bbox": [ + 69, + 269, + 282, + 401 + ], + "blocks": [ + { + "bbox": [ + 69, + 269, + 282, + 401 + ], + "lines": [ + { + "bbox": [ + 69, + 269, + 282, + 401 + ], + "spans": [ + { + "bbox": [ + 69, + 269, + 282, + 401 + ], + "type": "image", + "image_path": "e7ec1136051f6b341035ba7cbc22e3d91980c4f7b9dde9710591956639eb1e57.jpg" + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 55, + 412, + 297, + 458 + ], + "lines": [ + { + "bbox": [ + 55, + 412, + 297, + 458 + ], + "spans": [ + { + "bbox": [ + 55, + 412, + 297, + 458 + ], + "type": "text", + "content": "Figure 5. Compression Levels. DocVQA validation results for DocVLM integrated with Qwen2-VL across varying OCR and image token budgets. \"0\" represents the baseline, while \"Full\" indicates uncompressed encodings." + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "image_caption" + } + ], + "index": 5 + }, + { + "bbox": [ + 55, + 470, + 296, + 626 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 470, + 296, + 626 + ], + "spans": [ + { + "bbox": [ + 55, + 470, + 296, + 626 + ], + "type": "text", + "content": "Compression Levels We deepen our analysis of the OCR-visual token trade-off by examining different compression levels in DocVLM integrated with Qwen2-VL. Fig. 5 presents ANLS scores across various combinations of visual tokens (128-1024) and learned queries (16-256), including baselines without OCR and with uncompressed encoding. In low visual token regimes, increasing the number of learned queries yields substantial improvements, validating our compression mechanism's effectiveness in capturing relevant OCR information. Notably, even with just 16 learned queries, DocVLM outperforms the baseline across all visual token configurations, offering strong performance with minimal computational overhead." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 55, + 641, + 296, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 641, + 296, + 715 + ], + "spans": [ + { + "bbox": [ + 55, + 641, + 296, + 715 + ], + "type": "text", + "content": "Training Stages Tab. 4 illustrates the impact of the vision alignment stage on ANLS performance for DocVLM with the LLaVA-OneVision base model on the DocVQA test set, showing results for varying numbers of learned queries (16 to 128) and the uncompressed case. Our two-stage training process initially trains OCR modality components with-" + } + ] + } + ], + "index": 8 + }, + { + "type": "table", + "bbox": [ + 317, + 268, + 553, + 329 + ], + "blocks": [ + { + "bbox": [ + 317, + 268, + 553, + 329 + ], + "lines": [ + { + "bbox": [ + 317, + 268, + 553, + 329 + ], + "spans": [ + { + "bbox": [ + 317, + 268, + 553, + 329 + ], + "type": "table", + "html": "
Training PhasesCompressed Enc.OCR Enc.
1664128800
Stage I: OCR-LLM Alignment81.785.886.389.4
+ Stage II: Vision Alignment87.988.488.490.1
Δ+6.2+2.6+2.1+0.7
", + "image_path": "57dd9f561e1c9effc80f99bdf7f4a7038c4c9f2e57e3cc3637b993e57bf9486a.jpg" + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "table_body" + } + ], + "index": 9 + }, + { + "bbox": [ + 313, + 337, + 555, + 371 + ], + "lines": [ + { + "bbox": [ + 313, + 337, + 555, + 371 + ], + "spans": [ + { + "bbox": [ + 313, + 337, + 555, + 371 + ], + "type": "text", + "content": "Table 4. Training Phases. Second-stage training consistently improves DocVQA performance, both with compressed DocVLM tokens and full OCR encoding." + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "text" + }, + { + "bbox": [ + 313, + 383, + 555, + 515 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 383, + 555, + 515 + ], + "spans": [ + { + "bbox": [ + 313, + 383, + 555, + 515 + ], + "type": "text", + "content": "out image input, forcing reliance on OCR data alone, before reintroducing the image in the vision alignment stage to adapt learned queries alongside visual information. The results reveal that vision alignment significantly boosts performance, especially with fewer learned queries: for instance, with 16 learned queries, there's an improvement of +6.2, compared to +0.7 in the uncompressed case. Notably, after vision alignment, DocVLM with only 16 learned queries outperforms the baseline of OCR words (from Table 3). These findings underscore the effectiveness of our two-stage training method." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 313, + 526, + 392, + 539 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 526, + 392, + 539 + ], + "spans": [ + { + "bbox": [ + 313, + 526, + 392, + 539 + ], + "type": "text", + "content": "6. Conclusions" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 313, + 545, + 556, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 545, + 556, + 715 + ], + "spans": [ + { + "bbox": [ + 313, + 545, + 556, + 715 + ], + "type": "text", + "content": "Our results demonstrate that DocVLM can be effectively integrated into various VLMs to enhance their document reading capabilities while significantly reducing their dependency on extensive vision tokens. The key takeaway is that in token-constrained scenarios, allocating a small portion of tokens to OCR information consistently yields better results than using those tokens solely for visual processing. The effectiveness of our compression mechanism extends beyond single-page documents, as evidenced by achieving state-of-the-art results on MP-DocVQA using the same 64 tokens to represent multiple pages. These results establish DocVLM as a practical solution for enhancing document understanding in real-world applications where computational efficiency is crucial." + } + ] + } + ], + "index": 13 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "type": "text", + "content": "29012" + } + ] + } + ], + "index": 14 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 7 + }, + { + "para_blocks": [ + { + "bbox": [ + 56, + 71, + 115, + 83 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 71, + 115, + 83 + ], + "spans": [ + { + "bbox": [ + 56, + 71, + 115, + 83 + ], + "type": "text", + "content": "References" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 57, + 91, + 296, + 713 + ], + "type": "list", + "angle": 0, + "index": 12, + "blocks": [ + { + "bbox": [ + 61, + 91, + 296, + 156 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 61, + 91, + 296, + 156 + ], + "spans": [ + { + "bbox": [ + 61, + 91, + 296, + 156 + ], + "type": "text", + "content": "[1] Aviad Aberdam, Ron Litman, Shahar Tsiper, Oron Anschel, Ron Slossberg, Shai Mazor, R Manmatha, and Pietro Perona. Sequence-to-sequence contrastive learning for text recognition. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 15302-15312, 2021. 2" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 61, + 158, + 294, + 190 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 61, + 158, + 294, + 190 + ], + "spans": [ + { + "bbox": [ + 61, + 158, + 294, + 190 + ], + "type": "text", + "content": "[2] Aviad Aberdam, Roy Ganz, Shai Mazor, and Ron Litman. Multimodal semi-supervised learning for text recognition. arXiv preprint arXiv:2205.03873, 2022." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 62, + 192, + 295, + 246 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 192, + 295, + 246 + ], + "spans": [ + { + "bbox": [ + 62, + 192, + 295, + 246 + ], + "type": "text", + "content": "[3] Aviad Aberdam, David Bensaïd, Alona Golts, Roy Ganz, Oren Nuriel, Royee Tichauer, Shai Mazor, and Ron Litman. Clipter: Looking at the bigger picture in scene text recognition. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 21706-21717, 2023." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 62, + 247, + 295, + 313 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 247, + 295, + 313 + ], + "spans": [ + { + "bbox": [ + 62, + 247, + 295, + 313 + ], + "type": "text", + "content": "[4] Ofir Abramovich, Niv Nayman, Sharon Fogel, Inbal Lavi, Ron Litman, Shahar Tsiper, Royee Tichauer, Srikar Appalaraju, Shai Mazor, and R Manmatha. Visfocus: Prompt-guided vision encoders forOCR-free dense document understanding. In European Conference on Computer Vision, pages 241-259. Springer, 2025. 2" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 62, + 315, + 295, + 378 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 315, + 295, + 378 + ], + "spans": [ + { + "bbox": [ + 62, + 315, + 295, + 378 + ], + "type": "text", + "content": "[5] Jean-Baptiste Alayrac, Jeff Donahue, Pauline Luc, Antoine Miech, Iain Barr, Yana Hasson, Karel Lenc, Arthur Mensch, Katherine Millican, Malcolm Reynolds, et al. Flamingo: a visual language model for few-shot learning. Advances in neural information processing systems, 35:23716-23736, 2022. 2" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 62, + 380, + 296, + 434 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 380, + 296, + 434 + ], + "spans": [ + { + "bbox": [ + 62, + 380, + 296, + 434 + ], + "type": "text", + "content": "[6] Srikar Appalaraju, Bhavan Jasani, Bhargava Urala Kota, Yusheng Xie, and R Manmatha. Docformer: End-to-end transformer for document understanding. In Proceedings of the IEEE/CVF international conference on computer vision, pages 993-1003, 2021. 3" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 62, + 436, + 295, + 490 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 436, + 295, + 490 + ], + "spans": [ + { + "bbox": [ + 62, + 436, + 295, + 490 + ], + "type": "text", + "content": "[7] Srikar Appalaraju, Peng Tang, Qi Dong, Nishant Sankaran, Yichu Zhou, and R Manmatha. Docformerv2: Local features for document understanding. In Proceedings of the AAAI Conference on Artificial Intelligence, pages 709-718, 2024. 3, 5, 12" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 62, + 492, + 295, + 546 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 492, + 295, + 546 + ], + "spans": [ + { + "bbox": [ + 62, + 492, + 295, + 546 + ], + "type": "text", + "content": "[8] Gilles Baechler, Srinivas Sunkara, Maria Wang, Fedir Zubach, Hassan Mansoor, Vincent Etter, Victor Cärbune, Jason Lin, Jindong Chen, and Abhanshu Sharma. Screenai: A vision-language model for ui and infographics understanding. arXiv preprint arXiv:2402.04615, 2024. 1, 5" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 62, + 548, + 295, + 602 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 548, + 295, + 602 + ], + "spans": [ + { + "bbox": [ + 62, + 548, + 295, + 602 + ], + "type": "text", + "content": "[9] Jinze Bai, Shuai Bai, Shusheng Yang, Shijie Wang, Sinan Tan, Peng Wang, Junyang Lin, Chang Zhou, and Jingren Zhou. Qwen-vl: A versatile vision-language model for understanding, localization, text reading, and beyond. arXiv preprint arXiv:2308.12966, 1(2):3, 2023. 2" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 57, + 604, + 295, + 658 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 604, + 295, + 658 + ], + "spans": [ + { + "bbox": [ + 57, + 604, + 295, + 658 + ], + "type": "text", + "content": "[10] Lucas Beyer, Andreas Steiner, André Susano Pinto, Alexander Kolesnikov, Xiao Wang, Daniel Salz, Maxim Neumann, Ibrahim Alabdulmohsin, Michael Tschannen, Emanuele Bugliarello, et al. Paligemma: A versatile 3b vlm for transfer. arXiv preprint arXiv:2407.07726, 2024. 1, 5" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 57, + 658, + 295, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 658, + 295, + 713 + ], + "spans": [ + { + "bbox": [ + 57, + 658, + 295, + 713 + ], + "type": "text", + "content": "[11] Ali Furkan Biten, Ruben Tito, Andres Mafla, Lluis Gomez, Marçal Rusinol, Ernest Valveny, CV Jawahar, and Dimosthenis Karatzas. Scene text visual question answering. In Proceedings of the IEEE/CVF international conference on computer vision, pages 4291-4301, 2019. 4, 12" + } + ] + } + ], + "index": 11 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 57, + 73, + 553, + 713 + ], + "type": "list", + "angle": 0, + "index": 25, + "blocks": [ + { + "bbox": [ + 316, + 73, + 553, + 127 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 73, + 553, + 127 + ], + "spans": [ + { + "bbox": [ + 316, + 73, + 553, + 127 + ], + "type": "text", + "content": "[12] Ali Furkan Biten, Ron Litman, Yusheng Xie, Srikar Appalaraju, and R Manmatha. Latr: Layout-aware transformer for scene-text vqa. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 16548-16558, 2022. 2, 3" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 317, + 129, + 553, + 171 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 129, + 553, + 171 + ], + "spans": [ + { + "bbox": [ + 317, + 129, + 553, + 171 + ], + "type": "text", + "content": "[13] Ali Furkan Biten, Ruben Tito, Lluis Gomez, Ernest Valveny, and Dimosthenis Karatzas. Ocr-idl: Ocr annotations for industry document library dataset. arXiv preprint arXiv:2202.12985, 2022. 12" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 317, + 173, + 553, + 237 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 173, + 553, + 237 + ], + "spans": [ + { + "bbox": [ + 317, + 173, + 553, + 237 + ], + "type": "text", + "content": "[14] Tsachi Blau, Sharon Fogel, Roi Ronen, Alona Golts, Roy Ganz, Elad Ben Avraham, Aviad Aberdam, Shahar Tsiper, and Ron Litman. Gram: Global reasoning for multi-page vqa. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 15598-15607, 2024. 2, 3, 5, 7" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 316, + 239, + 553, + 282 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 239, + 553, + 282 + ], + "spans": [ + { + "bbox": [ + 316, + 239, + 553, + 282 + ], + "type": "text", + "content": "[15] Xinlei Chen, Hao Fang, Tsung-Yi Lin, Ramakrishna Vedantam, Saurabh Gupta, Piotr Dólar, and C Lawrence Zitnick. Microsoft coco captions: Data collection and evaluation server. arXiv preprint arXiv:1504.00325, 2015. 4, 12" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 316, + 284, + 553, + 338 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 284, + 553, + 338 + ], + "spans": [ + { + "bbox": [ + 316, + 284, + 553, + 338 + ], + "type": "text", + "content": "[16] Xi Chen, Josip Djolonga, Piotr Padlewski, Basil Mustafa, Soravit Changpinyo, Jialin Wu, Carlos Riquelme Ruiz, Sebastian Goodman, Xiao Wang, Yi Tay, et al. Pali-x: On scaling up a multilingual vision and language model. arXiv preprint arXiv:2305.18565, 2023. 2" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 316, + 338, + 553, + 392 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 338, + 553, + 392 + ], + "spans": [ + { + "bbox": [ + 316, + 338, + 553, + 392 + ], + "type": "text", + "content": "[17] Xi Chen, Xiao Wang, Lucas Beyer, Alexander Kolesnikov, Jialin Wu, Paul Voigtlaender, Basil Mustafa, Sebastian Goodman, Ibrahim Alabdulmohsin, Piotr Padlewski, et al. Pali-3 vision language models: Smaller, faster, stronger. arXiv preprint arXiv:2310.09199, 2023. 1, 2" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 317, + 394, + 553, + 458 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 394, + 553, + 458 + ], + "spans": [ + { + "bbox": [ + 317, + 394, + 553, + 458 + ], + "type": "text", + "content": "[18] Zhe Chen, Weiyun Wang, Hao Tian, Shenglong Ye, Zhangwei Gao, Erfei Cui, Wenwen Tong, Kongzhi Hu, Jiapeng Luo, Zheng Ma, et al. Internvl2: Better than the best—expanding performance boundaries of open-source multimodal models with the progressive scaling strategy, 2024. 2, 3, 4, 5" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 317, + 460, + 553, + 514 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 460, + 553, + 514 + ], + "spans": [ + { + "bbox": [ + 317, + 460, + 553, + 514 + ], + "type": "text", + "content": "[19] Wenliang Dai, Junnan Li, Dongxu Li, Anthony Meng Huat Tiong, Junqi Zhao, Weisheng Wang, Boyang Albert Li, Pascale Fung, and Steven C. H. Hoi. Instructclip: Towards general-purpose vision-language models with instruction tuning. ArXiv, abs/2305.06500, 2023. 2" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 317, + 515, + 553, + 580 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 515, + 553, + 580 + ], + "spans": [ + { + "bbox": [ + 317, + 515, + 553, + 580 + ], + "type": "text", + "content": "[20] Xiaoyi Dong, Pan Zhang, Yuhang Zang, Yuhang Cao, Bin Wang, Linke Ouyang, Songyang Zhang, Haodong Duan, Wenwei Zhang, Yining Li, et al. Internlm-xcomposer2-4khd: A pioneering large vision-language model handling resolutions from 336 pixels to 4k hd. arXiv preprint arXiv:2404.06512, 2024. 1" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 316, + 582, + 553, + 624 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 582, + 553, + 624 + ], + "spans": [ + { + "bbox": [ + 316, + 582, + 553, + 624 + ], + "type": "text", + "content": "[21] Masato Fujitake. Dtrocr: Decoder-only transformer for optical character recognition. In Proceedings of the IEEE/CVF Winter Conference on Applications of Computer Vision, pages 8025-8035, 2024. 2" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 316, + 625, + 553, + 669 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 625, + 553, + 669 + ], + "spans": [ + { + "bbox": [ + 316, + 625, + 553, + 669 + ], + "type": "text", + "content": "[22] Roy Ganz, Oren Nuriel, Aviad Aberdam, Yair Kittenplon, Shai Mazor, and Ron Litman. Towards models that can see and read. In Proceedings of the IEEE/CVF international conference on computer vision, pages 21718-21728, 2023. 2, 3" + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 316, + 670, + 553, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 670, + 553, + 713 + ], + "spans": [ + { + "bbox": [ + 316, + 670, + 553, + 713 + ], + "type": "text", + "content": "[23] Roy Ganz, Yair Kittenplon, Aviad Aberdam, Elad Ben Avraham, Oren Nuriel, Shai Mazor, and Ron Litman. Question aware vision transformer for multimodal reasoning. In Proceedings of the IEEE/CVF Conference on Computer Vision" + } + ] + } + ], + "index": 24 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "text", + "content": "29013" + } + ] + } + ], + "index": 26 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 8 + }, + { + "para_blocks": [ + { + "bbox": [ + 56, + 72, + 295, + 713 + ], + "type": "list", + "angle": 0, + "index": 14, + "blocks": [ + { + "bbox": [ + 76, + 72, + 294, + 94 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 76, + 72, + 294, + 94 + ], + "spans": [ + { + "bbox": [ + 76, + 72, + 294, + 94 + ], + "type": "text", + "content": "and Pattern Recognition (CVPR), pages 13861-13871, 2024. 2" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 56, + 95, + 295, + 149 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 95, + 295, + 149 + ], + "spans": [ + { + "bbox": [ + 56, + 95, + 295, + 149 + ], + "type": "text", + "content": "[24] Yash Goyal, Tejas Khot, Douglas Summers-Stay, Dhruv Bartra, and Devi Parikh. Making the V in VQA matter: Elevating the role of image understanding in Visual Question Answering. In Conference on Computer Vision and Pattern Recognition (CVPR), 2017. 4, 12" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 56, + 150, + 294, + 194 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 150, + 294, + 194 + ], + "spans": [ + { + "bbox": [ + 56, + 150, + 294, + 194 + ], + "type": "text", + "content": "[25] Jean-Philippe Thiran Guillaume Jaume, Hazim Kemal Ekenel. Funsd: A dataset for form understanding in noisy scanned documents. In Accepted to ICDAR-OST, 2019. 1" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 56, + 195, + 294, + 249 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 195, + 294, + 249 + ], + "spans": [ + { + "bbox": [ + 56, + 195, + 294, + 249 + ], + "type": "text", + "content": "[26] Benjamin Hsu, Xiaoyu Liu, Huayang Li, Yoshinari Fujinuma, Maria Nadejde, Xing Niu, Yair Kittenplon, Ron Litman, and Raghavendra Pappagari. M3t: A new benchmark dataset for multi-modal document-level machine translation. arXiv preprint arXiv:2406.08255, 2024. 3" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 56, + 250, + 294, + 304 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 250, + 294, + 304 + ], + "spans": [ + { + "bbox": [ + 56, + 250, + 294, + 304 + ], + "type": "text", + "content": "[27] Anwen Hu, Haiyang Xu, Liang Zhang, Jiabo Ye, Ming Yan, Ji Zhang, Qin Jin, Fei Huang, and Jingren Zhou. mplug-docowl2: High-resolution compressing forocr-free multi-page document understanding. arXiv preprint arXiv:2409.03420, 2024. 2, 7" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 56, + 304, + 294, + 358 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 304, + 294, + 358 + ], + "spans": [ + { + "bbox": [ + 56, + 304, + 294, + 358 + ], + "type": "text", + "content": "[28] Taeho Kil, Seonghyeon Kim, Sukmin Seo, Yoonsik Kim, and Daehe Kim. Towards unified scene text spotting based on sequence generation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 15223-15232, 2023. 2" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 56, + 360, + 294, + 414 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 360, + 294, + 414 + ], + "spans": [ + { + "bbox": [ + 56, + 360, + 294, + 414 + ], + "type": "text", + "content": "[29] Geewook Kim, Teakgyu Hong, Moonbin Yim, Jinyoung Park, Jinyeong Yim, Wonseok Hwang, Sangdoo Yun, Dongyoon Han, and Seunghyun Park. Donut: Document understanding transformer withoutOCR.arXiv preprint arXiv:2111.15664, 7(15):2, 2021. 2" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 56, + 415, + 294, + 448 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 415, + 294, + 448 + ], + "spans": [ + { + "bbox": [ + 56, + 415, + 294, + 448 + ], + "type": "text", + "content": "[30] Hugo Laurençon, Léo Tronchon, Matthieu Cord, and Victor Sanh. What matters when building vision-language models? arXiv preprint arXiv:2405.02246, 2024. 2" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 56, + 449, + 294, + 491 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 449, + 294, + 491 + ], + "spans": [ + { + "bbox": [ + 56, + 449, + 294, + 491 + ], + "type": "text", + "content": "[31] Bo Li, Yuanhan Zhang, Dong Guo, Renrui Zhang, Feng Li, Hao Zhang, Kaichen Zhang, Yanwei Li, Ziwei Liu, and Chunyuan Li. Llava-onevision: Easy visual task transfer. arXiv preprint arXiv:2408.03326, 2024. 1, 2, 4, 5" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 56, + 492, + 294, + 546 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 492, + 294, + 546 + ], + "spans": [ + { + "bbox": [ + 56, + 492, + 294, + 546 + ], + "type": "text", + "content": "[32] Junnan Li, Dongxu Li, Silvio Savarese, and Steven Hoi. Blip-2: Bootstrapping language-image pre-training with frozen image encoders and large language models. In International conference on machine learning, pages 19730-19742. PMLR, 2023. 2" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 56, + 548, + 294, + 590 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 548, + 294, + 590 + ], + "spans": [ + { + "bbox": [ + 56, + 548, + 294, + 590 + ], + "type": "text", + "content": "[33] Wentong Li, Yuqian Yuan, Jian Liu, Dongqi Tang, Song Wang, Jianke Zhu, and Lei Zhang. Tokenpacker: Efficient visual projector for multimodal llm. arXiv preprint arXiv:2407.02392, 2024. 2" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 56, + 592, + 295, + 646 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 592, + 295, + 646 + ], + "spans": [ + { + "bbox": [ + 56, + 592, + 295, + 646 + ], + "type": "text", + "content": "[34] Ron Litman, Oron Anschel, Shahar Tsiper, Roeit Litman, Shai Mazor, and R Manmatha. Scatter: selective context attentional scene text recognizer. In proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 11962-11972, 2020. 2" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 56, + 647, + 294, + 680 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 647, + 294, + 680 + ], + "spans": [ + { + "bbox": [ + 56, + 647, + 294, + 680 + ], + "type": "text", + "content": "[35] Haotian Liu, Chunyuan Li, Yuheng Li, Bo Li, Yuanhan Zhang, Sheng Shen, and Yong Jae Lee. Llava next: Improved reasoning,OCR, and world knowledge, 2024. 2, 3" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 56, + 681, + 294, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 681, + 294, + 713 + ], + "spans": [ + { + "bbox": [ + 56, + 681, + 294, + 713 + ], + "type": "text", + "content": "[36] Haotian Liu, Chunyuan Li, Qingyang Wu, and Yong Jae Lee. Visual instruction tuning. Advances in neural information processing systems, 36, 2024. 2" + } + ] + } + ], + "index": 13 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 316, + 72, + 555, + 713 + ], + "type": "list", + "angle": 0, + "index": 28, + "blocks": [ + { + "bbox": [ + 316, + 72, + 553, + 116 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 72, + 553, + 116 + ], + "spans": [ + { + "bbox": [ + 316, + 72, + 553, + 116 + ], + "type": "text", + "content": "[37] Yuliang Liu, Zhang Li, Mingxin Huang, Biao Yang, Wenwen Yu, Chunyuan Li, Xucheng Yin, Cheng lin Liu, Lianwen Jin, and Xiang Bai. Ocrbench: On the hidden mystery ofOCR in large multimodal models, 2024. 1" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 316, + 118, + 553, + 161 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 118, + 553, + 161 + ], + "spans": [ + { + "bbox": [ + 316, + 118, + 553, + 161 + ], + "type": "text", + "content": "[38] Jiahao Lyu, Wei Wang, Dongbao Yang, Jinwen Zhong, and Yu Zhou. Arbitrary reading order scene text spotter with local semantics guidance. arXiv preprint arXiv:2412.10159, 2024. 14" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 316, + 163, + 553, + 207 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 163, + 553, + 207 + ], + "spans": [ + { + "bbox": [ + 316, + 163, + 553, + 207 + ], + "type": "text", + "content": "[39] Ahmed Masry, Do Xuan Long, Jia Qing Tan, Shafiq Joty, and Enamul Hoque. Chartqa: A benchmark for question answering about charts with visual and logical reasoning. arXiv preprint arXiv:2203.10244, 2022. 1, 4, 12" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 316, + 209, + 555, + 252 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 209, + 555, + 252 + ], + "spans": [ + { + "bbox": [ + 316, + 209, + 555, + 252 + ], + "type": "text", + "content": "[40] Minesh Mathew, Dimosthenis Karatzas, and CV Jawahar. Docvqa: A dataset for vqa on document images. In Proceedings of the IEEE/CVF winter conference on applications of computer vision, pages 2200-2209, 2021. 4, 12, 13" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 316, + 254, + 553, + 307 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 254, + 553, + 307 + ], + "spans": [ + { + "bbox": [ + 316, + 254, + 553, + 307 + ], + "type": "text", + "content": "[41] Minesh Mathew, Viraj Bagal, Ruben Tito, Dimosthenis Karatzas, Ernest Valveny, and CV Jawahar. Infographicvqa. In Proceedings of the IEEE/CVF Winter Conference on Applications of Computer Vision, pages 1697-1706, 2022. 1, 4, 12, 13" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 316, + 309, + 553, + 364 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 309, + 553, + 364 + ], + "spans": [ + { + "bbox": [ + 316, + 309, + 553, + 364 + ], + "type": "text", + "content": "[42] Anand Mishra, Shashank Shekhar, Ajeet Kumar Singh, and Anirban Chakraborty. Ocr-vqa: Visual question answering by reading text in images. In 2019 international conference on document analysis and recognition (ICDAR), pages 947-952. IEEE, 2019. 4, 12" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 316, + 365, + 553, + 409 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 365, + 553, + 409 + ], + "spans": [ + { + "bbox": [ + 316, + 365, + 553, + 409 + ], + "type": "text", + "content": "[43] Oren Nuriel, Sharon Fogel, and Ron Litman. Textadain: Paying attention to shortcut learning in text recognizers. In European Conference on Computer Vision, pages 427-445. Springer, 2022. 2" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 316, + 411, + 553, + 454 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 411, + 553, + 454 + ], + "spans": [ + { + "bbox": [ + 316, + 411, + 553, + 454 + ], + "type": "text", + "content": "[44] Zhiliang Peng, Wenhui Wang, Li Dong, Yaru Hao, Shaohan Huang, Shuming Ma, and Furu Wei. Kosmos-2: Grounding multimodal large language models to the world. arXiv preprint arXiv:2306.14824, 2023. 2" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 316, + 456, + 553, + 510 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 456, + 553, + 510 + ], + "spans": [ + { + "bbox": [ + 316, + 456, + 553, + 510 + ], + "type": "text", + "content": "[45] Colin Raffel, Noam Shazeer, Adam Roberts, Katherine Lee, Sharan Narang, Michael Matena, Yanqi Zhou, Wei Li, and Peter J Liu. Exploring the limits of transfer learning with a unified text-to-text transformer. Journal of machine learning research, 21(140):1-67, 2020. 3" + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 316, + 512, + 553, + 555 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 512, + 553, + 555 + ], + "spans": [ + { + "bbox": [ + 316, + 512, + 553, + 555 + ], + "type": "text", + "content": "[46] Roi Ronen, Shahar Tsiper, Oron Anschel, Inbal Lavi, Amir Markovitz, and R Manmatha. Glass: Global to local attention for scene-text spotting. arXiv preprint arXiv:2208.03364, 2022. 2" + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 316, + 556, + 553, + 621 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 556, + 553, + 621 + ], + "spans": [ + { + "bbox": [ + 316, + 556, + 553, + 621 + ], + "type": "text", + "content": "[47] Oleksii Sidorov, Ronghang Hu, Marcus Rohrbach, and Amanpreet Singh. Textcaps: a dataset for image captioning with reading comprehension. In Computer Vision-ECCV 2020: 16th European Conference, Glasgow, UK, August 23-28, 2020, Proceedings, Part II 16, pages 742-758. Springer, 2020. 4, 12" + } + ] + } + ], + "index": 25 + }, + { + "bbox": [ + 316, + 624, + 553, + 678 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 624, + 553, + 678 + ], + "spans": [ + { + "bbox": [ + 316, + 624, + 553, + 678 + ], + "type": "text", + "content": "[48] Amanpreet Singh, Vivek Natarajan, Meet Shah, Yu Jiang, Xinlei Chen, Dhruv Batra, Devi Parikh, and Marcus Rohrbach. Towards vqa models that can read. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 8317-8326, 2019. 4, 12" + } + ] + } + ], + "index": 26 + }, + { + "bbox": [ + 316, + 680, + 553, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 680, + 553, + 713 + ], + "spans": [ + { + "bbox": [ + 316, + 680, + 553, + 713 + ], + "type": "text", + "content": "[49] Ryota Tanaka, Taichi Iki, Kyosuke Nishida, Kuniko Saito, and Jun Suzuki. Instructdoc: A dataset for zero-shot generalization of visual document understanding with instructions." + } + ] + } + ], + "index": 27 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 318, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 318, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 318, + 757 + ], + "type": "text", + "content": "29014" + } + ] + } + ], + "index": 29 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 9 + }, + { + "para_blocks": [ + { + "bbox": [ + 56, + 72, + 295, + 714 + ], + "type": "list", + "angle": 0, + "index": 12, + "blocks": [ + { + "bbox": [ + 76, + 72, + 294, + 95 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 76, + 72, + 294, + 95 + ], + "spans": [ + { + "bbox": [ + 76, + 72, + 294, + 95 + ], + "type": "text", + "content": "In Proceedings of the AAAI Conference on Artificial Intelligence, pages 19071-19079, 2024. 6" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 56, + 96, + 295, + 129 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 96, + 295, + 129 + ], + "spans": [ + { + "bbox": [ + 56, + 96, + 295, + 129 + ], + "type": "text", + "content": "[50] Ruben Tito, Dimosthenis Karatzas, and Ernest Valveny. Hierarchical multimodal transformers for multipage docvqa. Pattern Recognition, 144:109834, 2023. 1, 2, 4, 12" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 56, + 130, + 295, + 207 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 130, + 295, + 207 + ], + "spans": [ + { + "bbox": [ + 56, + 130, + 295, + 207 + ], + "type": "text", + "content": "[51] Jordy Van Landeghem, Ruben Tito, Łukasz Borchmann, Michal Pietruszka, Pawel Joziak, Rafal Powalski, Dawid Jurkiewicz, Mickaël Coustaty, Bertrand Anckaert, Ernest Valveny, et al. Document understanding dataset and evaluation (dude). In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 19528-19540, 2023. 2, 4, 12" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 56, + 209, + 295, + 263 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 209, + 295, + 263 + ], + "spans": [ + { + "bbox": [ + 56, + 209, + 295, + 263 + ], + "type": "text", + "content": "[52] Dongsheng Wang, Natraj Raman, Mathieu Sibue, Zhiqiang Ma, Petr Babkin, Simerjot Kaur, Yulong Pei, Armineh Nourbakhsh, and Xiaomo Liu. Docllm: A layout-aware generative language model for multimodal document understanding. arXiv preprint arXiv:2401.00908, 2023. 2, 13" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 56, + 264, + 295, + 319 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 264, + 295, + 319 + ], + "spans": [ + { + "bbox": [ + 56, + 264, + 295, + 319 + ], + "type": "text", + "content": "[53] Peng Wang, Shuai Bai, Sinan Tan, Shijie Wang, Zhihao Fan, Jinze Bai, Keqin Chen, Xuejing Liu, Jialin Wang, Wenbin Ge, et al. Qwen2-vl: Enhancing vision-language model's perception of the world at any resolution. arXiv preprint arXiv:2409.12191, 2024. 1, 2, 3, 4, 5" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 56, + 321, + 295, + 365 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 321, + 295, + 365 + ], + "spans": [ + { + "bbox": [ + 56, + 321, + 295, + 365 + ], + "type": "text", + "content": "[54] Wenjin Wang, Yunhao Li, Yixin Ou, and Yin Zhang. Layout and task aware instruction prompt for zero-shot document image question answering. arXiv preprint arXiv:2306.00526, 2023. 1, 2, 13" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 56, + 366, + 295, + 422 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 366, + 295, + 422 + ], + "spans": [ + { + "bbox": [ + 56, + 366, + 295, + 422 + ], + "type": "text", + "content": "[55] Yiheng Xu, Minghao Li, Lei Cui, Shaohan Huang, Furu Wei, and Ming Zhou. Layoutm: Pre-training of text and layout for document image understanding. In Proceedings of the 26th ACM SIGKDD international conference on knowledge discovery & data mining, pages 1192-1200, 2020. 13" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 56, + 422, + 295, + 477 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 422, + 295, + 477 + ], + "spans": [ + { + "bbox": [ + 56, + 422, + 295, + 477 + ], + "type": "text", + "content": "[56] Jiabo Ye, Anwen Hu, Haiyang Xu, Qinghao Ye, Ming Yan, Guohai Xu, Chenliang Li, Junfeng Tian, Qi Qian, Ji Zhang, et al. Ureader: UniversalOCR-free visually-situated language understanding with multimodal large language model. arXiv preprint arXiv:2310.05126, 2023. 2" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 56, + 479, + 295, + 534 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 479, + 295, + 534 + ], + "spans": [ + { + "bbox": [ + 56, + 479, + 295, + 534 + ], + "type": "text", + "content": "[57] Maoyuan Ye, Jing Zhang, Shanshan Zhao, Juhua Liu, Bo Du, and Dacheng Tao. Aptext-detr: Towards better scene text detection with dynamic points in transformer. In Proceedings of the AAAI Conference on Artificial Intelligence, pages 3241-3249, 2023. 2" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 56, + 535, + 295, + 600 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 535, + 295, + 600 + ], + "spans": [ + { + "bbox": [ + 56, + 535, + 295, + 600 + ], + "type": "text", + "content": "[58] Maoyuan Ye, Jing Zhang, Shanshan Zhao, Juhua Liu, Tongliang Liu, Bo Du, and Dacheng Tao. Deepsolo: Let transformer decoder with explicit points solo for text spotting. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 19348-19357, 2023. 2" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 56, + 601, + 295, + 657 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 601, + 295, + 657 + ], + "spans": [ + { + "bbox": [ + 56, + 601, + 295, + 657 + ], + "type": "text", + "content": "[59] Qinghao Ye, Haiyang Xu, Guohai Xu, Jiabo Ye, Ming Yan, Yiyang Zhou, Junyang Wang, Anwen Hu, Pengcheng Shi, Yaya Shi, et al. MPLug-owl: Modularization empowers large language models with multimodality. arXiv preprint arXiv:2304.14178, 2023. 2" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 56, + 658, + 295, + 714 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 658, + 295, + 714 + ], + "spans": [ + { + "bbox": [ + 56, + 658, + 295, + 714 + ], + "type": "text", + "content": "[60] Pan Zhang, Xiaoyi Dong, Yuhang Zang, Yuhang Cao, Rui Qian, Lin Chen, Qipeng Guo, Haodong Duan, Bin Wang, Linke Ouyang, et al. Internlm-xcomposer-2.5: A versatile large vision language model supporting long-contextual input and output. arXiv preprint arXiv:2407.03320, 2024. 1" + } + ] + } + ], + "index": 11 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 315, + 72, + 554, + 174 + ], + "type": "list", + "angle": 0, + "index": 15, + "blocks": [ + { + "bbox": [ + 315, + 72, + 554, + 118 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 315, + 72, + 554, + 118 + ], + "spans": [ + { + "bbox": [ + 315, + 72, + 554, + 118 + ], + "type": "text", + "content": "[61] Deyao Zhu, Jun Chen, Xiaogian Shen, Xiang Li, and Mohamed Elhoseiny. Minigpt-4: Enhancing vision-language understanding with advanced large language models. arXiv preprint arXiv:2304.10592, 2023. 2" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 316, + 118, + 554, + 174 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 118, + 554, + 174 + ], + "spans": [ + { + "bbox": [ + 316, + 118, + 554, + 174 + ], + "type": "text", + "content": "[62] Fengbin Zhu, Wenqiang Lei, Fuli Feng, Chao Wang, Haozhou Zhang, and Tat-Seng Chua. Towards complex document understanding by discrete reasoning. In Proceedings of the 30th ACM International Conference on Multimedia, pages 4857-4866, 2022. 4, 12" + } + ] + } + ], + "index": 14 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "text", + "content": "29015" + } + ] + } + ], + "index": 16 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 10 + } + ], + "_backend": "vlm", + "_version_name": "2.6.4" +} \ No newline at end of file diff --git a/2025/Docopilot_ Improving Multimodal Models for Document-Level Understanding/2cc56609-3eb1-412a-98df-28e82ae9cd58_content_list.json b/2025/Docopilot_ Improving Multimodal Models for Document-Level Understanding/2cc56609-3eb1-412a-98df-28e82ae9cd58_content_list.json new file mode 100644 index 0000000000000000000000000000000000000000..c3baf55ed1492bdd6fe98511ff0187e4212a0ff6 --- /dev/null +++ b/2025/Docopilot_ Improving Multimodal Models for Document-Level Understanding/2cc56609-3eb1-412a-98df-28e82ae9cd58_content_list.json @@ -0,0 +1,1740 @@ +[ + { + "type": "text", + "text": "Docopilot: Improving Multimodal Models for Document-Level Understanding", + "text_level": 1, + "bbox": [ + 104, + 130, + 893, + 155 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Yuchen Duan $^{1,2*}$ , Zhe Chen $^{3,1*}$ , Yusong Hu $^{4,1*}$ , Weiyun Wang $^{*5,1}$ , Shenglong Ye $^{1}$ , Botian Shi $^{1}$ , Lewei Lu $^{7}$ , Qibin Hou $^{4}$ , Tong Lu $^{3,1}$ , Hongsheng Li $^{2,1}$ , Jifeng Dai $^{6,1}$ , Wenhai Wang $^{2,1\\boxtimes}$", + "bbox": [ + 156, + 180, + 841, + 217 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "$^{1}$ Shanghai AI Laboratory, $^{2}$ The Chinese University of Hong Kong, $^{3}$ Nanjing University, $^{4}$ Nankai University, $^{5}$ Fudan University, $^{6}$ Tsinghua University, $^{7}$ SenseTime Research", + "bbox": [ + 184, + 220, + 815, + 258 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Abstract", + "text_level": 1, + "bbox": [ + 246, + 292, + 326, + 309 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Despite significant progress in multimodal large language models (MLLMs), their performance on complex, multipage document comprehension remains inadequate, largely due to the lack of high-quality, document-level datasets. While current retrieval-augmented generation (RAG) methods offer partial solutions, they suffer from issues, such as fragmented retrieval contexts, multi-stage error accumulation, and extra time costs of retrieval. In this work, we present a high-quality document-level dataset, Doc750K, designed to support in-depth understanding of multimodal documents. This dataset includes diverse document structures, extensive cross-page dependencies, and real question-answer pairs derived from the original documents. Building on the dataset, we develop a native multimodal model—Docopilot, which can accurately handle document-level dependencies without relying on RAG. Experiments demonstrate that Docopilot achieves superior coherence, accuracy, and efficiency in document understanding tasks and multi-turn interactions, setting a new baseline for document-level multimodal understanding. Data, code, and models are released at https://github.com/OpenGVLab/Docopilot.", + "bbox": [ + 88, + 324, + 485, + 657 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "1. Introduction", + "text_level": 1, + "bbox": [ + 89, + 674, + 223, + 690 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "In recent years, multimodal large language models (MLLMs) [6, 12, 41, 56, 60, 72, 81, 82, 84, 92] have rapidly developed, achieving remarkable performance in various visual understanding tasks [30, 77], particularly image-level tasks, such as image captioning [10, 42], optical character recognition (OCR) [45, 65], and visual question answering (VQA) [24, 44]. Despite these advances, current MLLMs still face significant challenges in document-level understanding [51, 76, 89], where models are required to identify and integrate key information across multi-page documents, setting high expectations for their long-context processing", + "bbox": [ + 89, + 699, + 483, + 867 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/ca45d626bd25fc9f7b1548f606b52489b79335623c8f6b671344188d240b28ca.jpg", + "image_caption": [ + "Figure 1. Accuracy $\\nu.s$ inference latency on MM-NIAH. The proposed Docopilot-8B shows a notable improvement over baseline models [73], achieving a $+19.9\\%$ accuracy gain compared to InternVL2-8B and surpassing InternVL2-26B with less than $31\\%$ of the inference latency. Additionally, Docopilot-2B uses fewer parameters (less than $10\\%$ ) while exhibiting comparable performance to the $10\\times$ larger InternVL2-26B. These results suggest that our Docopilot strikes a reasonable balance between latency, model size, and performance." + ], + "image_footnote": [], + "bbox": [ + 514, + 277, + 906, + 474 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "capabilities of MLLMs.", + "text_level": 1, + "bbox": [ + 511, + 627, + 671, + 642 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Current research on long-content understanding primarily focuses on text-only models [5, 9, 74], targeting specific retrieval tasks such as \"Needle in a Haystack\" (NIAH) [2, 31]. However, existing open-source MLLMs [12, 13, 36, 41, 62, 81, 83] are primarily trained on image-level data, lacking the long-context understanding capacity required for document-level understanding. Retrieval-augmented generation (RAG) methods [14, 22, 50, 61, 63, 95] attempt to address this by retrieving key information to fit within the limited context windows of MLLMs, but they still encounter the following challenges in document-level tasks. (1) Fragmented Retrieval Contexts. Retrieved information is fragmented, lacking the overall structure of the document; (2) Multi-Stage Error Accumulation. Incorrect retrieval results can affect subsequent responses, leading to errors or omissions of critical details, especially in multi-turn or complex tasks; (3) Extra Time Costs. The retrieval", + "bbox": [ + 511, + 643, + 908, + 902 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "CVF", + "bbox": [ + 106, + 2, + 181, + 42 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore.", + "bbox": [ + 236, + 0, + 807, + 46 + ], + "page_idx": 0 + }, + { + "type": "page_footnote", + "text": "* Equal contribution;", + "bbox": [ + 109, + 875, + 230, + 887 + ], + "page_idx": 0 + }, + { + "type": "page_footnote", + "text": "Corresponding author: wangwenhai@pjlab.org.cn", + "bbox": [ + 107, + 888, + 387, + 900 + ], + "page_idx": 0 + }, + { + "type": "page_number", + "text": "4026", + "bbox": [ + 482, + 944, + 516, + 955 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "step increases the latency of the QA system, limiting the scalability of RAG in time-sensitive scenarios.", + "bbox": [ + 89, + 90, + 480, + 119 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "To address these problems, two primary challenges need to be considered. (1) High-Quality Multimodal Document Dataset. While extensive datasets [11, 91, 96, 97] exist for long-context, text-only tasks, high-quality document-level question-answering datasets remain scarce. This shortage is largely attributed to the high costs associated with annotation and the lack of streamlined construction pipelines. (2) Native Document-Level MLLMs. Although RAG-based methods [14, 86, 95] provide some relief, native multimodal models with long-context processing abilities are crucial. However, training native MLLMs specifically for document-level understanding is constrained by current hardware limitations.", + "bbox": [ + 88, + 122, + 482, + 318 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "In this work, we introduce a new multimodal document dataset that supports document-level understanding tasks. Compared to counterparts [33, 76, 78], this dataset has the following features: (1) Large Scale. It includes a total of 758K question-answer samples, containing 5.2B text tokens and 3.1M images. It encompasses content from various sources, such as Sci-Hub, Arxiv, and OpenReview, covering a wide range of topics and document layouts. (2) High Quality. Unlike existing datasets that insert irrelevant questions into documents, we collect real, in-depth question-answer pairs and construct single-page and crosspage questions based on document structure. Such high-quality question-answer data accounts for $31.6\\%$ of the dataset. (3) Multimodal. For the document content, we provide not only the conventional interleaved text-image context but also purely rendered image inputs, catering to the needs of different models.", + "bbox": [ + 91, + 321, + 483, + 575 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Building upon this dataset, we developed a native baseline model for document-level multimodal understanding-Docopilot. Unlike existing approaches [14, 86, 95] that rely on RAG, our model achieves efficient document-level training and testing through simple engineering optimizations, such as multimodal data packing, Ring Attention [43], and Liger Kernel [16]. Leveraging the proxy tasks carefully designed within the dataset, Docopilot can directly handle long-distance dependencies and cross-page information integration without external retrieval support. As shown in Figure 1, this approach not only enhances coherence and accuracy compared to RAG methods but also significantly reduces the response time of the entire question-answering system, delivering superior real-time performance in multi-turn interactions.", + "bbox": [ + 89, + 580, + 482, + 805 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "The main contributions are summarized as follows:", + "bbox": [ + 109, + 808, + 450, + 821 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "(1) We develop the first large-scale, high-quality dataset for document-level multimodal understanding, consisting of 758K QA pairs from 3 sources, supporting 9 types of proxy tasks. This dataset includes $31.6\\%$ real QA pairs directly extracted from documents.", + "bbox": [ + 89, + 825, + 482, + 900 + ], + "page_idx": 1 + }, + { + "type": "list", + "sub_type": "text", + "list_items": [ + "(2) Based on the dataset, we implement Docopilot, a native MLLM designed for document-level understanding without relying on retrieval mechanisms. This approach greatly improves its ability to integrate and comprehend information across multi-page documents.", + "(3) Through extensive experiments on multiple document-level benchmarks, our method demonstrates performance significantly superior to existing approaches, proving its effectiveness and generality. As shown in Figure 1, Docopilot-8B achieves a score of 61.8 on MM-NIAH [86], outperforming InternVL2-8B by 19.9 points and surpassing InternVL2-26B with less than $31\\%$ of the latency. We hope this work could provide a baseline for future advancements in MLLMs for document-level tasks." + ], + "bbox": [ + 511, + 90, + 906, + 303 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "2. Related Work", + "text_level": 1, + "bbox": [ + 513, + 319, + 653, + 335 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "2.1. Multimodal Large Language Models", + "text_level": 1, + "bbox": [ + 511, + 344, + 831, + 361 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Multimodal large language models (MLLMs) have demonstrated impressive capabilities in processing image and text information, opening up new directions for applications such as visual question answering and image captioning. Early models [12, 29, 37, 59] trained with contrastive learning methods excelled in recognizing and understanding open-world semantics within an image-text matching framework. However, their limited generative abilities restricted their applicability. To leverage the powerful generation abilities of large language models (LLMs), subsequent works [13, 38, 42, 47, 83, 84] introduced a connector to align the embedding spaces of vision encoders and LLMs, allowing encoded image embeddings to serve as soft prompts for LLMs. Another series of works [1, 34, 39, 99] extended LLMs by integrating additional visual experts, reducing reliance on standalone vision encoders. More recently, models capable of both understanding and generating images have also made notable progress [20, 35, 67, 75], leveraging the insight that image generation can enhance image understanding. Despite these advancements, current MLLMs still face challenges with long-context multimodal inputs. For instance, InternVL 2.0 [13, 25] performs optimally within a token range of up to 8192, constraining its effectiveness in document-level applications.", + "bbox": [ + 511, + 367, + 906, + 731 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "2.2. Document Understanding Models", + "text_level": 1, + "bbox": [ + 511, + 742, + 808, + 758 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Extracting key information from documents is crucial for industries and academic research. OCR-model-driven methods [3, 4, 71, 80] represent one of the primary technical approaches. These methods extract text, layout, and bounding box information from external systems and integrate it with another model. However, they are prone to error propagation and high processing times due to their reliance on multiple components. Benefitting from the rapid advancements in LLMs, OCR-free methods have also achieved", + "bbox": [ + 511, + 763, + 906, + 900 + ], + "page_idx": 1 + }, + { + "type": "page_number", + "text": "4027", + "bbox": [ + 482, + 944, + 514, + 955 + ], + "page_idx": 1 + }, + { + "type": "image", + "img_path": "images/18b2b2256ee8d2ed25df211f25d986c42c96fd2728e54c3e98f1529a5a7d8047.jpg", + "image_caption": [ + "Figure 2. Multimodal document dataset generation pipeline. This pipeline involves three main stages: (1) Raw Data Collection: Documents are gathered from sources like Sci-Hub, arXiv, and OpenReview, available in PDF and HTML formats. (2) Document Content Extraction: Multimodal content is processed in two formats: interleaved text-image format and multi-image format. (3) Question-Answer Pairs Construction: QA pairs are generated based on the document structure or constructed using GPT-4o." + ], + "image_footnote": [], + "bbox": [ + 91, + 85, + 903, + 308 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "great progress. Donut [32] is the first end-to-end training framework based on a Transformer without requiring OCR engines or APIs. Subsequent works [23, 49, 87, 88, 94] propose diverse modifications in model architectures and training algorithms. However, these models are designed for specific tasks and lack general abilities.", + "bbox": [ + 89, + 393, + 483, + 484 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "2.3. Long-Context Large Language Models", + "text_level": 1, + "bbox": [ + 89, + 500, + 426, + 515 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "With advancements in engineering, architecture, and algorithms, long-context large language models have made substantial progress. Techniques such as Flash Attention [17, 18] and Ring Attention [43] have notably reduced GPU memory usage for training on extended contexts. Additionally, various sparse attention mechanisms—including Shifting Sparse Attention [11], Dilated Attention [19], and Attention Sinks [26, 90]—have enabled efficient scaling to handle larger contexts. New positional embedding methods, like ALiBi [58], xPOS [68], and RoPE [66], further enhance the models' generalization capabilities in length extrapolation. However, these advancements remain largely confined to natural language processing, and methods to extend the context size of MLLMs are still under-explored. Another research approach aims to reduce context size by leveraging retrieval augmented generation (RAG) [14, 22, 95], where only the most relevant passages are retrieved and fed into the generation model. However, this retrieval-based approach can disrupt the coherence of the semantic chain, particularly in complex reasoning tasks, due to fragmented information flow. In this work, we integrate the above engineering techniques into MLLMs and demonstrate that a model fine-tuned on a high-quality, long-context training corpus is a strong baseline, achieving superior performance compared to its RAG counterpart.", + "bbox": [ + 89, + 523, + 483, + 902 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3. Multimodal Document Dataset Generation", + "text_level": 1, + "bbox": [ + 511, + 391, + 895, + 407 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "In this section, we begin by introducing the details of the data engine. Following this, we provide a comprehensive overview of the dataset—Doc-750K.", + "bbox": [ + 511, + 416, + 906, + 460 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3.1. Data Engine", + "text_level": 1, + "bbox": [ + 511, + 470, + 648, + 486 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "The data engine operates primarily through two steps: document content extraction and question-answer (QA) pair construction. Specifically, we first extract multimodal content, including both text and images, from the documents. Based on this extracted content, we then create question-answer pairs. The document content and these constructed pairs are combined to produce conversational-style training data. The format is outlined as:", + "bbox": [ + 511, + 492, + 906, + 612 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Please read the paper: , and answer the question: Answer: ", + "bbox": [ + 521, + 623, + 879, + 662 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Here, the , , and are the placeholder for extracted document content, the generated questions and answer, respectively. In the following, we will provide a detailed explanation of the two key steps: document content extraction and QA pair construction.", + "bbox": [ + 511, + 674, + 906, + 750 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Document Content Extraction. In practical applications, different documents have varying page layouts and content types, which poses significant challenges for content extraction. To enhance the efficiency of multimodal models, it is necessary to organize documents into a unified format for streamlined processing. In this work, we process each document into two formats as follows:", + "bbox": [ + 511, + 750, + 908, + 853 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "(1) Interleaved Text-Image Format. Using the document content extractor MinerU [79], we segment the document content into interleaved text and image annotations, for", + "bbox": [ + 511, + 854, + 908, + 900 + ], + "page_idx": 2 + }, + { + "type": "page_number", + "text": "4028", + "bbox": [ + 482, + 944, + 514, + 955 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "The objective consideration of conservation policy under restructuring is proving to be a difficult task. One of the greatest obstacles has been the persistent uncertainty among utility planners regarding the true resource-effectiveness and cost-effectiveness of conservation relative", + "bbox": [ + 145, + 133, + 488, + 172 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "These figures indicate that larger utilities tend to report greater ...", + "bbox": [ + 148, + 180, + 415, + 191 + ], + "page_idx": 3 + }, + { + "type": "image", + "img_path": "images/1e8cd6897d0cf21bd56005d043630dcd5da915d354ca58756e0370a1e9e36284.jpg", + "image_caption": [ + "Figure 3. Visualization of an example from Doc-750K. The left side presents the interleaved text-image format data obtained through Document Content Extraction, while the right side showcases the annotations generated via Question-Answer Pairs Construction." + ], + "image_footnote": [], + "bbox": [ + 161, + 193, + 264, + 276 + ], + "page_idx": 3 + }, + { + "type": "image", + "img_path": "images/7c96ebc511b971ca903adf4f92d5f88de2e4c30fbb329885667a3f9be9335907.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 287, + 198, + 480, + 268 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Early estimation attempts with individual utilities revealed that there was insufficient variation in the data series to produce reliable coefficient estimates for each utility taken separately. To increase the statistical power of the analysis, ...", + "bbox": [ + 145, + 279, + 495, + 308 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Finally, since no incentives are offered, it becomes more difficult to attribute customer investment decisions to the utility programs as opposed to other influences.", + "bbox": [ + 147, + 316, + 493, + 335 + ], + "page_idx": 3 + }, + { + "type": "table", + "img_path": "images/718578b4a99dc159f339390893b10485b59b40d90d8686993c83f40626ce7523.jpg", + "table_caption": [], + "table_footnote": [], + "table_body": "
Utility (C&I Sales)Program Type1993 Conservation Impacts(Annotated)
GWh% of Sales% of Impacts
Pacific Gas & Electric(45,760 GWh)Incentives1,8864.1%63%
Services9402.1%31%
Other170.4%6%
Total3,0136.6%100%
San Diego Gas & Electric(8,415 GWh)Incentives2512.7%65%
Services971.0%25%
Other2820.4%100%
Total3854.1%100%
Southern California Edison(86,721 GWh)Incentives1,7283.7%22%
Services6,02912.9%71%
Other2400.5%3%
Total7,97417.1%100%
Sacramento(8,733 GWh)Incentives1533.2%58%
Services1042.2%40%
Other60.1%2%
Total2822.5%100%
Los Angeles(14,829 GWh)Incentives7800.5%50%
Services7910.5%50%
Other00.0%0%
Total1571.1%100%
", + "bbox": [ + 151, + 339, + 326, + 443 + ], + "page_idx": 3 + }, + { + "type": "table", + "img_path": "images/fc3ebc0f87d9548c9b90d47d72dd9ac9971482440961d40d1ed63524f0bb55c8.jpg", + "table_caption": [], + "table_footnote": [], + "table_body": "
StudySampleFactorPeriodElectricity
Elec.UnitCostFuel
Longuet et al. 1994Norway 1976aenergy consumptionLB−0.39−0.39
energy consumptionLB−0.15−0.15
energy consumption & transportLB−0.44−0.16
Hammoudi & Kazer 1995Iran 1955–8domestic heatingLB−0.19−0.19
Safdarzadeh 1995USAcommercialLB−0.70−0.25
industrialLB−0.22−0.22
Rall 1996USAindustrialLB−0.18−0.20−0.150.09
industrialLB−0.29−0.29
Counihan 1998USA 1995–8industrialLB−0.390.530.040.09
industrialLB−0.22−0.22
Murray 2000Virginia, USA 1970–8industrialLB−0.29−0.09
industrialLB−0.22−0.22
industrialLB−0.21−0.04
Woodward 2001Netherlands 1977–83industrialLB−1.11−0.2−0.39−0.16
industrialLB−0.11−0.11
National Energy Board 2003CanadaindustrialLB−0.110.020.05
industrialLB−0.11−0.11
industrialLB−0.450.120.06
Energy, Media and Media Ethics Board Canada 2005CanadaindustrialLB−0.090.030.01
industrialLB−0.09−0.09
industrialLB−0.490.110.01
This studyUSAcommercial RLB−0.078−0.033−0.026−0.013
industrialLB−0.078−0.078
", + "bbox": [ + 348, + 339, + 491, + 441 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "According to California regulators, SCE's reported impacts for service programs have always been viewed with skepticism during regulatory reviews, although its impacts for other types of conservation programs have not. In California, service program impacts have historically been viewed as unverifiable, so utilities have been largely ineligible for incentive payments related to service program expenditures. Regulators in California acknowledge that they believe SCE's reported conservation impacts overstate the actual savings achieved by their programs and have communicated this skepticism to the utility. But, according to the rate regulation ...", + "bbox": [ + 145, + 450, + 495, + 527 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "", + "bbox": [ + 539, + 102, + 576, + 111 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Please read the paper and answer the question:", + "bbox": [ + 539, + 112, + 730, + 119 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "According to this paper, how does Image-1 illustrate the conservation impacts of various utilities in 1993? Discuss the findings and the significance of these data in the context of the research.", + "bbox": [ + 537, + 122, + 825, + 152 + ], + "page_idx": 3 + }, + { + "type": "image", + "img_path": "images/04fe59a24ac3d6bbbff74037dab49d3c97c86e29b1e4bb485e0920f30f7f9889.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 531, + 169, + 568, + 202 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "According to the paper, 'In 1993, the 39 companies in the sample accounted for $19\\%$ of total C&I electricity sales nationwide and $40\\%$ of all utility conservation program impacts, including residential impacts (Energy Information Administration 1995). The table in Image-1 provides detailed data on the 1993 C&I sales and C&I conservation program impacts for various utilities. This image shows the sales in GWh and conservation impacts both in GWh and as a percentage of sales. For example, California utilities such as Southern California Edison reported high sales and high conservation percentages, which is a significant outlier. These data illustrate the varying levels of conservation achievements across utilities, underlining the study's aim to verify the effectiveness and accuracy of reported conservation impacts. Larger utilities tend to report greater conservation impacts relative to sales, and this detail supports the analysis presented in the paper.", + "bbox": [ + 581, + 181, + 890, + 327 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Analyze the importance of Image-3 in comparing the study's findings with previous empirical research on electricity price elasticities.", + "bbox": [ + 535, + 358, + 823, + 378 + ], + "page_idx": 3 + }, + { + "type": "image", + "img_path": "images/c880a59ac2258e15418036f94388275e4684eff8fb2a96229ca00928bb57542f.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 854, + 349, + 898, + 383 + ], + "page_idx": 3 + }, + { + "type": "image", + "img_path": "images/392bb71143c9e71645c114493265a9b3032c0eeb7ff4d7c239ff8285ff70a36a.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 529, + 396, + 568, + 429 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "According to the paper, 'The regression diagnostics offered little guidance on improving the structure of the model. The addition of a dummy variable for the years 1971-1973 did little to improve the residuals in subsequent years. Adding a set of dummy variables to account for each unstable period did not seem reasonable.' Image-3 presents a comparative analysis of electricity cross-price elasticity studies. This table showcases various studies, their sample sectors, and reported elasticities for electricity, natural gas, oil, and coal across different regions and time periods. The study's own findings, -0.078 for short-run own-price elasticity of electricity, align closely with prior research, confirming its place within established literature. The context provided by other studies helps to validate the paper's conclusions regarding the robustness of its econometric model and the effectiveness of conservation programs.", + "bbox": [ + 581, + 407, + 892, + 532 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "example, \"\\n\\n \\n\" This format captures the document's textual content, making it easier to construct question-answer pairs.", + "bbox": [ + 89, + 598, + 482, + 643 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "(2) Multi-Image Format. In this format, a document with $n$ pages is rendered as $n$ images, with each image corresponding to a single page. The structure follows the pattern \"\\n\\n\". This format preserves the original layout, enabling the model to learn the overall pagination and visual layout of the document.", + "bbox": [ + 89, + 648, + 482, + 739 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "After processing the document into contexts in interleaved text-image and paginated image formats, we can not only use these contexts for next-token prediction training but also leverage the document's content, hierarchical structure, and layout features to flexibly and precisely generate high-quality question-answer pairs.", + "bbox": [ + 89, + 744, + 482, + 835 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Question-Answer Pairs Construction. In this step, we create question-and-answer pairs tailored to the source, content features, and formatting structure of each document. This process is divided into the following main categories:", + "bbox": [ + 89, + 840, + 482, + 901 + ], + "page_idx": 3 + }, + { + "type": "list", + "sub_type": "text", + "list_items": [ + "(1) For documents with reliable QA annotations, like the review and reply in OpenReview, we extract the QA pairs and organize them into conversation format.", + "(2) For documents with a clear textual structure, such as well-structured papers from Sci-Hub and Arxiv, we convert them to text and segment them, while the model is instructed to generate contents for each segment, including abstracts, experiment descriptions, and captions for figures and tables. The details of each task for structural papers are illustrated in Table 1.", + "(3) For other documents, in addition to using them directly for NTP pretraining, we can input text interspersed with images into MLLMs to obtain QA pairs. To ensure high-quality generated data, we use the state-of-the-art model GPT-4o [56]." + ], + "bbox": [ + 511, + 599, + 905, + 834 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Through our pipeline, most data has been processed into high-quality document-level question-answering data, while the remaining data is converted to plain text and used for next-token prediction tasks. Our pipelines are metricu", + "bbox": [ + 511, + 840, + 905, + 901 + ], + "page_idx": 3 + }, + { + "type": "header", + "text": "0", + "bbox": [ + 96, + 89, + 130, + 122 + ], + "page_idx": 3 + }, + { + "type": "header", + "text": "How Many Kilowatts are in a Negawatt?", + "bbox": [ + 145, + 102, + 316, + 111 + ], + "page_idx": 3 + }, + { + "type": "header", + "text": "Verifying \"Ex Post\" Estimates of Utility Conservation Impacts at the Regional Level", + "bbox": [ + 145, + 112, + 493, + 122 + ], + "page_idx": 3 + }, + { + "type": "header", + "text": "#", + "bbox": [ + 854, + 90, + 898, + 127 + ], + "page_idx": 3 + }, + { + "type": "page_number", + "text": "4029", + "bbox": [ + 482, + 944, + 514, + 955 + ], + "page_idx": 3 + }, + { + "type": "image", + "img_path": "images/728d8d93a75762d4ff8bb882269a03389f5b3a05187bda3f5deb0af18bf957ae.jpg", + "image_caption": [ + "(a) Data Distribution of Doc-750K", + "(b) Data Recipe for Supervised Fine-tuning", + "Figure 4. Data distribution of our dataset. The outer circle shows the distribution of all data categories and the inner circle shows the distribution of data subsets. Left: Data distribution of Doc-750K. Right: Data distribution of our complete SFT training dataset. Note that the number reported in the figure represents the number of samples. “MT” is short for multi-turn." + ], + "image_footnote": [], + "bbox": [ + 94, + 89, + 903, + 223 + ], + "page_idx": 4 + }, + { + "type": "table", + "img_path": "images/5a70c6e62b7dca2b729d85e214f8271bf3452cb07f185ccdf1f577a4c971f2e4.jpg", + "table_caption": [], + "table_footnote": [], + "table_body": "
TasksQuestions
Abstract WritingRead the full text of the paper and provide a concise summary in the form of an abstract.
Paper TitlingBased on the provided abstract or introduction of the research paper, please generate a concise and informative title
Caption WritingGive the relative texts of the images or tables, please write a caption for each image or table based on the relative texts provided.
Experiment WritingPlease write the "Experiments" section based on the incomplete research paper provided.
TranslationPlease read the full text of the following research paper and translate the Experiments section into Chinese.
", + "bbox": [ + 91, + 310, + 472, + 602 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "lously designed to ensure high data quality across all generated context. Each LLM-generated sample is explicitly marked in the metadata as model-generated. Across the entire dataset, only $4.8\\%$ of the data is LLM-generated, reinforcing the overall reliability and quality of the dataset.", + "bbox": [ + 89, + 685, + 483, + 762 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "3.2. Multimodal Document Dataset", + "text_level": 1, + "bbox": [ + 89, + 771, + 364, + 787 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Data Source. The composition and distribution of our training data are detailed in Figure 4. Specifically, our dataset predominantly consists of academic papers, which constitute approximately $32.6\\%$ of the total data. The multimodal data, carefully selected to augment our model's learning dimensions, makes up about $88.8\\%$ of our dataset. This strategic distribution is designed to optimize the train-", + "bbox": [ + 89, + 794, + 483, + 900 + ], + "page_idx": 4 + }, + { + "type": "table", + "img_path": "images/34178f1c1b749176664d52c14694e320668ee4086e61eda412a23e5bd0b531b8.jpg", + "table_caption": [ + "Table 1. Questions format for different tasks. For documents with a clear textual structure, we design several proxy tasks. All tasks leverage the inherent structure of the documents, with answers directly sourced from the original text." + ], + "table_footnote": [], + "table_body": "
StatisticsNumber
Total Questions758K
Total Images3.1M
Total Conversations251K
Multi-Turn Questions87K
Single-Turn Questions164K
Average Text Tokens11245
Average Image Tokens6178
", + "bbox": [ + 596, + 310, + 823, + 452 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Table 2. Key statistics of the Doc-750K datasets. It comprises 758K questions, 3.1M images, and 251K conversations, including 87K multi-turn and 164K single-turn questions. With an average of 11,245 text tokens and 6,178 image tokens, it highlights the dataset's richness and diversity for multimodal research.", + "bbox": [ + 511, + 460, + 906, + 532 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "ing process and improve the model's ability to generalize across different types of data inputs.", + "bbox": [ + 511, + 547, + 905, + 578 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Dataset Statistics. In our Doc-750K dataset, the majority of the data consists of reliably annotated entries, with OpenReview and Arxiv collectively accounting for $75.4\\%$ . The remaining data, sourced from Sci-Hub, is processed using our designed tasks. The overall distribution and number of tasks are shown in Figure 4(a). Our dataset ultimately consists of 251K conversations, comprising a total of 758K questions. Additional statistical details are provided in Table 2. Compared to previous datasets, Doc-750K contains a larger number of images, with an average of four images per conversation segment. Further comparisons with other datasets are shown in Table 3.", + "bbox": [ + 511, + 579, + 906, + 760 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "3.3. Data Recipe for Supervised Fine-Tuning", + "text_level": 1, + "bbox": [ + 511, + 771, + 857, + 787 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Although Doc-750K effectively covers multimodal document QA scenarios, using it directly may lead to model over-fitting on a specific document domain. Therefore, we combine it with several open-source datasets to create a mixed dataset for SFT training. As shown in Figure 4(b), these datasets are organized into 4 categories as follows:", + "bbox": [ + 511, + 794, + 905, + 883 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "(1) For multi-page document QA, Doc-750K serves", + "bbox": [ + 531, + 885, + 905, + 901 + ], + "page_idx": 4 + }, + { + "type": "page_number", + "text": "4030", + "bbox": [ + 482, + 944, + 514, + 955 + ], + "page_idx": 4 + }, + { + "type": "table", + "img_path": "images/4f85ce1ef16faf309af9a9df58f29b1ac12e5962f886980acfc7fcd595266def.jpg", + "table_caption": [], + "table_footnote": [], + "table_body": "
Dataset#Images#QA Pairs#Tokens
Docmatix [33]2,444,7509,500,000390,000,000
DocVQA [15]10,18939,463337,829
TextCaps [64]21,95321,953389,658
TextVQA [65]21,95334,602181,918
ST-VQA [8]17,24723,121127,846
OCR-VQA [55]165,746801,5796,073,824
VisualMRC [70]3,02711,988168,828
DUDE [78]147,59723,71611,341,228
Doc-750K (ours)3,103,494758,0005,200,000,000
", + "bbox": [ + 91, + 88, + 480, + 244 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Table 3. Comparison with popular VQA datasets.", + "bbox": [ + 130, + 253, + 442, + 268 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "as the core dataset, specifically curated to address complex, multi-page document comprehension. Additional datasets such as MP-Docmatix [33], MP-DocVQA [53], DUDE [78], and Taesiri-ArxivQA [69] offer valuable multi-page scenarios requiring inter-page reasoning and contextual retention across sequences.", + "bbox": [ + 89, + 286, + 482, + 377 + ], + "page_idx": 5 + }, + { + "type": "list", + "sub_type": "text", + "list_items": [ + "(2) For multi-image general QA, MMDU-45K [48] offers a comprehensive dataset encompassing diverse real-world scenarios, such as natural environments and everyday contexts. It emphasizes multi-turn dialogues and integration of multiple images, supporting the development of systems capable of generating coherent and accurate responses from complex, lengthy inputs.", + "(3) For single-page document QA, We introduce DocVQA [53], DocReason [93], InfoVQA [54], and ChartQA [52] to further enhance the diversity of the SFT dataset. These datasets focus on individual pages with complex layouts, rich textual information, and, in some cases, graphical data interpretation.", + "(4) For pure-text QA, we add datasets including LongAlpaca [11], LongAlpaca-16K-Length [11], LongQLoRA [91], LongCite [96], LongAlign [7], and LongReward [97] to support the assessment of the model's capabilities in QA tasks requiring long-range dependencies." + ], + "bbox": [ + 89, + 377, + 482, + 652 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "This expanded dataset provides a balanced foundation for training and evaluating multimodal document understanding models, enhancing robustness and adaptability across diverse document-related VQA tasks.", + "bbox": [ + 89, + 652, + 482, + 713 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "4. Enhanced Baseline for Document-Level Multimodal Understanding", + "text_level": 1, + "bbox": [ + 89, + 728, + 482, + 763 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "4.1. Model Architecture", + "text_level": 1, + "bbox": [ + 89, + 771, + 277, + 787 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Our model architecture leverages the widely-adopted ViTMLP-LLM structure [41, 42, 73], consisting of a pre-trained Vision Transformer (ViT), a two-layer MLP projector, and a pre-trained Language Model (LLM). This combination provides a strong baseline for multimodal document analysis, effectively integrating visual and textual information within a unified framework.", + "bbox": [ + 89, + 794, + 482, + 900 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "4.2. Optimizing Training Efficiency", + "text_level": 1, + "bbox": [ + 513, + 90, + 787, + 107 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "The training efficiency of MLLMs is hindered by two key challenges: (1) Inconsistent Sample Lengths. Samples with different context lengths will result in excessive padding and lower training throughput; and (2) Limited GPU Memory. As the model scale and context length increase, GPU memory consumption becomes increasingly unsustainable. To address these issues, we have implemented the following strategies:", + "bbox": [ + 511, + 119, + 906, + 242 + ], + "page_idx": 5 + }, + { + "type": "list", + "sub_type": "text", + "list_items": [ + "(1) Multimodal Data Packing. To balance the computational load between the vision model (ViT) and the language model (LLM) while minimizing resource waste caused by padding, we implement a multimodal data-packing strategy. The key idea is to concatenate multiple samples into long sequences to fully utilize the model's input capacity. Specifically, thresholds $T_{\\mathrm{img}}$ and $T_{\\mathrm{tok}}$ are set for the number of images and tokens, respectively. Samples are managed using a priority queue, sorted in descending order by the number of images and total tokens. A new sample $s$ attempts to combine with the sample at the front of the priority queue. If the combination meets the thresholds (i.e., $T_{\\mathrm{img}}, T_{\\mathrm{tok}}$ ), the combined sample is pushed back into the priority queue. If $s$ cannot match with any existing sample in the queue, it is directly added to the queue. When the image number and total token number of the front sample reach one of the thresholds or the number of samples exceeds the maximum limit $M$ , the front sample is dequeued, padded as needed, and sent for training. This strategy optimizes resource utilization and ensures balanced computational workloads. The detailed pseudo-code can be found in the supplementary materials.", + "(2) Ring Attention. We implement the Ring Attention mechanism [43] to alleviate memory constraints associated with processing long sequences. By partitioning sequences into blocks and distributing computation across multiple devices, Ring Attention allows the model to accommodate larger contexts. This approach enables overlapping communication between key-value blocks and attention computations, thereby enhancing parallel processing efficiency. Consequently, Ring Attention improves the model's capacity to handle extended context lengths without exceeding memory limits.", + "(3) Liger Kernel. To further improve memory and computational efficiency, we integrate the Liger Kernel [16], a specialized kernel library optimized for large-scale model training. The Liger Kernel enhances throughput and reduces memory consumption by employing techniques like kernel fusion, in-place operations, and input chunking. Leveraging the Liger Kernel thus enables higher training throughput and addresses memory limitations, allowing for efficient scaling of large multimodal models." + ], + "bbox": [ + 511, + 250, + 908, + 901 + ], + "page_idx": 5 + }, + { + "type": "page_number", + "text": "4031", + "bbox": [ + 482, + 944, + 513, + 955 + ], + "page_idx": 5 + }, + { + "type": "table", + "img_path": "images/16bbf05e201805b33321f9199fc0b99105a33f078c397d055d771a030686a69d.jpg", + "table_caption": [], + "table_footnote": [], + "table_body": "
ModelsMP-DocMMLong-DocDocGenomeMM-NIAH
ANSL↑Acc↑F1↑Class Acc↑Title ED↓Abstract ED↓SP Acc↑MP Acc↑ShortMediumLongOverall
Proprietary Models
Gemini-1.5-Pro [60]-28.220.6-----73.865.260.867.1
GPT-4o [56]-42.844.997.69.56.571.867.6----
Open-Source Models
MiniMonkey-2B [28]70.310.38.657.416.555.040.328.940.926.923.531.0
InternVL2-2B [13]71.810.510.860.818.454.339.428.936.621.219.426.4
InternVL2-2B + RAG [85]72.617.216.760.818.454.339.428.436.830.234.833.8
Llama3.2-3B-Instruct†[21]-23.721.285.3194.751.040.234.915.52.20.56.6
Docopilot-2B (ours)76.221.816.056.24.543.645.137.458.046.740.949.2
MiniCPM-V2.6-8B [92]-16.915.492.810.232.660.054.249.015.30.023.4
LLaVA-OneVision-8B [36]-10.89.685.649.977.59.87.165.738.00.037.7
mPLUG-DocOwl2-8B [27]69.413.48.9-----17.90.10.06.6
M3DocRAG [14]84.421.022.6---------
VisRAG-8B [95]-18.818.392.810.232.660.050.747.129.229.535.8
InternLM2.5-7B-1M†[9]-28.725.692.777.659.342.742.540.537.235.137.8
InternVL2-8B [13]79.317.416.590.68.239.656.046.156.437.332.442.9
InternVL2-8B + RAG [85]78.724.224.590.68.239.656.046.055.743.445.248.4
InternVL2-26B [13]-15.515.487.516.923.349.742.765.048.741.952.8
Docopilot-8B (ours)81.328.823.093.82.019.753.951.971.257.455.361.8
", + "bbox": [ + 91, + 88, + 911, + 353 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "5. Experiments", + "text_level": 1, + "bbox": [ + 89, + 431, + 222, + 450 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "5.1. Experimental Setup", + "text_level": 1, + "bbox": [ + 89, + 455, + 279, + 472 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Training Details. Our model is available in two sizes: Docopilot-2B and Docopilot-8B, both of which are based on the InternVL2 [73] and fine-tuned for one epoch using the data recipe that includes Doc-750K. The training uses a batch size of 128, the AdamW optimizer with a learning rate of 1e-5, weight decay of 0.01 for the 2B variant, and 0.05 for the 8B variant, along with a cosine learning rate schedule. To speed up training, we apply multimodal data packing to reduce padding and a dynamic high-resolution strategy [13] to enhance OCR for document understanding. The maximum number of tiles for multimodal data is limited to 24, and the maximum sequence length is set to 32k tokens.", + "bbox": [ + 88, + 476, + 482, + 671 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Baselines. We compare our Docopilot with a series of open-source document-level MLLMs [27, 28, 36, 40, 46, 92, 98] that supports multi-image input and proprietary MLLMs, including Gemini-1.5-pro [60], GPT-4o [57]. For comparison with the commonly used RAG method for handling long documents, we selected the latest multimodal RAG methods VisRAG [95], InternVL + RAG [85], and M3DocRAG [14]. To compare with more long-context large language models, we use InternVL2-8B as the OCR model to extract texts from the documents and images and feed the parsed documents to long-context LLMs [9, 21].", + "bbox": [ + 89, + 674, + 482, + 840 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "5.2. Multi-Page VQA", + "text_level": 1, + "bbox": [ + 89, + 847, + 256, + 864 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Benchmarks. For the multi-page VQA task, we evaluate our model on three benchmarks: (1) MP-DocVQA [76],", + "bbox": [ + 89, + 869, + 482, + 902 + ], + "page_idx": 6 + }, + { + "type": "table", + "img_path": "images/26f1931376001c27d19bacb349d2724e9b7d9a32a8ecfd149541b2a2e89a0f47.jpg", + "table_caption": [ + "Table 4. Evaluation on multi-page and interleaved VQA benchmarks. We report the metrics on MP-DocVQA [76] (MP-Doc), MMLongbench-Doc [51] (MMLong-Doc), DocGenome [89], and MM-NIAH [85]. Our model outperforms document-level MLLMs and multimodal RAG methods on multi-page, medium, and long-context QA. The \"Short\", \"Medium\", and \"Long\" in MM-NIAH refer to input length in $[0,8\\mathrm{k}]$ , $(8\\mathrm{k},32\\mathrm{k}]$ , $(32\\mathrm{k},64\\mathrm{k}]$ , respectively. \" $\\dagger$ ” denotes input documents are parsed by OCR models." + ], + "table_footnote": [], + "table_body": "
ModelsDocVQAChartQAInfoVQA
Gemini-1.5-Pro [60]93.187.281.0
GPT-4o [56]92.885.7-
MiniMonkey-2B [28]87.476.560.1
InternVL2-2B [13]86.976.258.9
Docopilot-2B (ours)87.376.458.5
Monkey-8B [40]66.565.136.1
TextMonkey-9B [46]73.066.928.6
mPLUG-DocOwl2-8B [27]80.770.046.4
IXC2.5-7B [98]90.982.269.9
InternVL2-8B [13]91.683.374.8
Docopilot-8B (ours)92.083.373.3
", + "bbox": [ + 514, + 431, + 903, + 604 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Table 5. Results on single-page VQA benchmarks. Our Docopilot models perform comparably to baselines [73], demonstrating enhanced long-context modeling without loss on shorter tasks.", + "bbox": [ + 511, + 611, + 906, + 654 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "which is designed to evaluate the ability to handle complex questions across multiple scanned document pages. (2) MMLongbench-Doc [51], a benchmark for evaluating the performance of MLLMs on multi-modal documents. (3) DocGenome [89], a large-scale benchmark for the evaluation of scientific document comprehension.", + "bbox": [ + 511, + 672, + 906, + 762 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Results. As illustrated in Table 4, our model achieves consistent improvements on multi-page QA benchmarks, outperforming previous document-level MLLMs. Notably, our Docopilot-8B surpasses Gemini-1.5-Pro [60] on MMLongBench-Doc, positioning it as the closest open-source model to GPT-4o. In comparison to RAG-based methods [14, 86, 95], our model demonstrates advantages in multi-page scenarios. For example, in the Multi/Page QA of DocGenome benchmark, the RAG method shows a perfor", + "bbox": [ + 511, + 763, + 908, + 902 + ], + "page_idx": 6 + }, + { + "type": "page_number", + "text": "4032", + "bbox": [ + 482, + 944, + 514, + 955 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "mance decline due to the disruption of document continuity while our Docopilot exhibits a significantly stable improvement compared to the baseline, with Docopilot-8B showing an increase of $12.6\\%$ over InternVL2-8B.", + "bbox": [ + 89, + 90, + 482, + 151 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "5.3. Interleaved Long-Context QA", + "text_level": 1, + "bbox": [ + 89, + 162, + 357, + 178 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Benchmarks. For the interleaved long-context QA task, we evaluate our models on MM-NIAH [86], a benchmark designed for long multimodal document comprehension.", + "bbox": [ + 89, + 184, + 482, + 229 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Results. The right side of Table 4 presents the results of MM-NIAH across context lengths ranging from 1K to 64K. We categorize the context lengths into \"Short,\" \"Medium,\" and \"Long\" based on the context window of InternVL2 (8K) and Docopilot (32K). Our Docopilot demonstrates exceptional performance in both medium- and long-context scenarios, while maintaining high accuracy in short-context situations. Notably, for QA tasks with context lengths in the range of $(32K,64K]$ , Docopilot-2B outperforms InternVL2-2B by $110\\%$ , and Docopilot-8B surpasses InternVL2-8B by $70\\%$ . Furthermore, our model performs comparably to the state-of-the-art multimodal long-context model Gemini-1.5-Pro in contexts longer than 8K, establishing a new state-of-the-art performance among open-source long-context MLLMs.", + "bbox": [ + 89, + 229, + 482, + 455 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "5.4. Single-Page VQA", + "text_level": 1, + "bbox": [ + 89, + 467, + 261, + 483 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "**Benchmarks.** For single-page VQA tasks, we evaluate our model on three benchmarks: (1) DocVQA [53], a benchmark for the evaluation of extracting key information from an image of the given document. (2) ChartQA [52], a benchmark for evaluating the reasoning abilities for chart images. (3) InfoVQA [54], a benchmark for infographic image comprehension.", + "bbox": [ + 89, + 489, + 482, + 594 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Results. As shown in Table 5, our model achieves comparable performance to baseline models. Across the three benchmarks, Docopilot-2B and InternVL2-2B exhibit comparable results, while Docopilot-8B outperforms InternVL2-8B by 0.4 points in DocVQA. These results demonstrate that Doc-750K effectively enhances the model's long-context modeling capabilities without compromising its performance on shorter documents.", + "bbox": [ + 89, + 595, + 482, + 715 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "5.5. Ablation Study", + "text_level": 1, + "bbox": [ + 89, + 727, + 243, + 742 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Effect of Doc-750K. We conducted ablation studies on MMLongBench-Doc [51] to analyze the impact of our Doc-750K. We divided Doc-750K into 3 parts according to the source of the data: (1) Sci-Hub data; (2) Arxiv data; and (3) OpenReview data. We demonstrate the effects of incorporating each part of the data into the SFT process, reported in Table 6. We observed that with the inclusion of different parts of Doc-750K, the model's performance improves continuously. Utilizing only open-source data results in an inferior F1 score.", + "bbox": [ + 89, + 750, + 482, + 898 + ], + "page_idx": 7 + }, + { + "type": "table", + "img_path": "images/f3afcee789bbe9458060b54057fead125444d4ee51d890372258ee2e54c7f719.jpg", + "table_caption": [], + "table_footnote": [], + "table_body": "
ModelsAccF1
Baseline (InternVL2-2B [73])10.510.8
- Variant1: SFT using data recipe w/o Doc-750K18.49.4
- Variant2: Variant1 + our Sci-Hub data18.515.2
- Variant3: Variant2 + our Arxiv data20.515.5
Docopilot-2B: Variant3 + our OpenReview data21.816.0
", + "bbox": [ + 517, + 89, + 903, + 176 + ], + "page_idx": 7 + }, + { + "type": "table", + "img_path": "images/60652b1fe221f3d6db2a12e66b7b07a4d1ee9dca2a83c2944df0dea0797c3d81.jpg", + "table_caption": [ + "Table 6. Ablation study on the data recipe. We evaluate the effectiveness of the training data from different sources on MMLongBench-Doc. Our Doc-750K can consistently enhance the ability of the model to understand multi-page documents." + ], + "table_footnote": [], + "table_body": "
ModelsLatencyAccF1
MiniCPM-V2.6 [92]225.4ms16.915.4
VisRAG-12B [95]288.3ms18.818.3
InternVL2-2B [13]35.9ms10.510.8
InternVL2-2B + RAG [85]82.9ms17.216.7
Docopilot-2B (ours)35.9ms21.816.0
InternVL2-8B [13]81.0ms17.416.5
InternVL2-8B + RAG [85]113.4ms24.224.5
Docopilot-8B (ours)81.0ms28.823.0
", + "bbox": [ + 526, + 256, + 893, + 393 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Table 7. Latency analysis. We evaluate the average token output latency of model outputs on MMLongBench-Doc [51]. RAG-based methods [85, 95] exhibit slower processing speeds due to their two-stage inference process, making them less efficient than document MLLMs for handling multimodal long documents.", + "bbox": [ + 511, + 404, + 906, + 474 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Latency Analysis. To compare the latency in inference between RAG methods and our Docopilot, we conducted a latency analysis on MMLongBench-Doc [51], as reported in Table 7. While RAG reduces the document length input to the MLLM, its own time cost remains non-negligible. For instance, InternVL2-2B + RAG is $130\\%$ slower than InternVL2-2B, and VisRAG is $28\\%$ slower than MiniCPMV2.6. Our Docopilot does not require additional processes and therefore has the same inference time as baseline models, making it more suitable for analyzing long documents.", + "bbox": [ + 511, + 500, + 906, + 651 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "6. Conclusions", + "text_level": 1, + "bbox": [ + 511, + 664, + 640, + 680 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "This work introduced a diverse document-level question-answering dataset that covers complex structures and crosspage dependencies, providing a robust foundation for training and evaluating document understanding models. We also proposed a retrieval-free long-document understanding model that effectively integrates multi-page information, reducing reliance on external retrieval systems. Experimental results show that our model achieves state-of-the-art performance across several document-level QA benchmarks, underscoring its strength in multi-page integration and complex reasoning. Future work will focus on improving computational efficiency, extending the model to larger multimodal tasks, and adapting it to broader applications for enhanced practicality and generalization.", + "bbox": [ + 511, + 688, + 906, + 900 + ], + "page_idx": 7 + }, + { + "type": "page_number", + "text": "4033", + "bbox": [ + 482, + 944, + 514, + 955 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Acknowledgments", + "text_level": 1, + "bbox": [ + 91, + 90, + 250, + 107 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "This project was supported by the National Key R&D Program of China (No. 2022ZD0161300, 2022ZD0160101), the National Natural Science Foundation of China (No. 62376134, 62372223). Zhe Chen is supported by the Youth PhD Student Research Project under the National Natural Science Foundation (No. 623B2050).", + "bbox": [ + 89, + 114, + 483, + 205 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "References", + "text_level": 1, + "bbox": [ + 91, + 218, + 187, + 234 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[1] Jean-Baptiste Alayrac, Jeff Donahue, Pauline Luc, Antoine Miech, Iain Barr, Yana Hasson, Karel Lenc, Arthur Mensch, Katherine Millican, Malcolm Reynolds, et al. Flamingo: a visual language model for few-shot learning. NeurIPS, 35: 23716-23736, 2022. 2", + "[2] Chenxin An, Shansan Gong, Ming Zhong, Xingjian Zhao, Mukai Li, Jun Zhang, Lingpeng Kong, and Xipeng Qiu. L-eval: Instituting standardized evaluation for long context language models. arXiv preprint arXiv:2307.11088, 2023. 1", + "[3] Srikar Appalaraju, Peng Tang, Qi Dong, Nishant Sankaran, Yichu Zhou, and R Manmatha. Docformerv2: Local features for document understanding. In Proceedings of the AAAI Conference on Artificial Intelligence, pages 709-718, 2024.", + "[4] Haoli Bai, Zhiguang Liu, Xiaojun Meng, Wentao Li, Shuang Liu, Nian Xie, Rongfu Zheng, Liangwei Wang, Lu Hou, Jiansheng Wei, et al. Wukong-reader: Multi-modal pretraining for fine-grained visual document understanding. arXiv preprint arXiv:2212.09621, 2022. 2", + "[5] Jinze Bai, Shuai Bai, Yunfei Chu, Zeyu Cui, Kai Dang, Xiaodong Deng, Yang Fan, Wenbin Ge, Yu Han, Fei Huang, Binyuan Hui, Luo Ji, Mei Li, Junyang Lin, Runji Lin, Dayiheng Liu, Gao Liu, Chengqiang Lu, Keming Lu, Jianxin Ma, Rui Men, Xingzhang Ren, Xuancheng Ren, Chuanqi Tan, Sinan Tan, Jianhong Tu, Peng Wang, Shijie Wang, Wei Wang, Shengguang Wu, Benfeng Xu, Jin Xu, An Yang, Hao Yang, Jian Yang, Shusheng Yang, Yang Yao, Bowen Yu, Hongyi Yuan, Zheng Yuan, Jianwei Zhang, Xingxuan Zhang, Yichang Zhang, Zhenru Zhang, Chang Zhou, Jingren Zhou, Xiaohuan Zhou, and Tianhang Zhu. Qwen technical report. arXiv preprint arXiv:2309.16609, 2023. 1", + "[6] Jinze Bai, Shuai Bai, Shusheng Yang, Shijie Wang, Sinan Tan, Peng Wang, Junyang Lin, Chang Zhou, and Jingren Zhou. Qwen-vl: A frontier large vision-language model with versatile abilities. arXiv preprint arXiv:2308.12966, 2023. 1", + "[7] Yushi Bai, Xin Lv, Jiajie Zhang, Yuze He, Ji Qi, Lei Hou, Jie Tang, Yuxiao Dong, and Juanzi Li. Longalign: A recipe for long context alignment of large language models. arXiv preprint arXiv:2401.18058, 2024. 6", + "[8] Ali Furkan Biten, Ruben Tito, Andres Mafla, Lluis Gomez, Marcal Rusinol, Ernest Valveny, CV Jawahar, and Dimosthenis Karatzas. Scene text visual question answering. In ICCV, pages 4291-4301, 2019. 6", + "[9] Zheng Cai, Maosong Cao, Haojiong Chen, Kai Chen, Keyu Chen, Xin Chen, Xun Chen, Zehui Chen, Zhi Chen, Pei Chu, et al. Internl m2 technical report. arXiv preprint arXiv:2403.17297, 2024. 1, 7" + ], + "bbox": [ + 99, + 243, + 483, + 898 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[10] Xinlei Chen, Hao Fang, Tsung-Yi Lin, Ramakrishna Vedantam, Saurabh Gupta, Piotr Dólar, and C Lawrence Zitnick. Microsoft coco captions: Data collection and evaluation server. arXiv preprint arXiv:1504.00325, 2015. 1", + "[11] Yukang Chen, Shengju Qian, Haotian Tang, Xin Lai, Zhi-jian Liu, Song Han, and Jiaya Jia. Longlora: Efficient finetuning of long-context large language models. arXiv preprint arXiv:2309.12307, 2023. 2, 3, 6", + "[12] Zhe Chen, Jiannan Wu, Wenhai Wang, Weijie Su, Guo Chen, Sen Xing, Zhong Muyan, Qinglong Zhang, Xizhou Zhu, Lewei Lu, et al. Internvl: Scaling up vision foundation models and aligning for generic visual-linguistic tasks. arXiv preprint arXiv:2312.14238, 2023. 1, 2", + "[13] Zhe Chen, Weiyun Wang, Hao Tian, Shenglong Ye, Zhangwei Gao, Erfei Cui, Wenwen Tong, Kongzhi Hu, Jiapeng Luo, Zheng Ma, et al. How far are we to gpt-4v? closing the gap to commercial multimodal models with open-source suites. arXiv preprint arXiv:2404.16821, 2024. 1, 2, 7, 8", + "[14] Jaemin Cho, Debanjan Mahata, Ozan Irsoy, Yujie He, and Mohit Bansal. M3docrag: Multi-modal retrieval is what you need for multi-page multi-document understanding. arXiv preprint arXiv:2411.04952, 2024. 1, 2, 3, 7", + "[15] Christopher Clark and Matt Gardner. Simple and effective multi-paragraph reading comprehension. In ACL, pages 845–855, 2018. 6", + "[16] Yun Dai, Vignesh Kothapalli, Qingquan Song, Shao Tang, Siyu Zhu, Steven Shimizu, Shivam Sahni, Haowen Ning, Yanning Chen, et al. Liger kernel: Efficient triton kernels for llm training. arXiv preprint arXiv:2410.10989, 2024. 2, 6", + "[17] Tri Dao. Flashattention-2: Faster attention with better parallelism and work partitioning. arXiv preprint arXiv:2307.08691, 2023. 3", + "[18] Tri Dao, Dan Fu, Stefano Ermon, Atri Rudra, and Christopher Ré. Flashattention: Fast and memory-efficient exact attention with io-awareness. NeurIPS, 35:16344-16359, 2022. 3", + "[19] Jiayu Ding, Shuming Ma, Li Dong, Xingxing Zhang, Shaohan Huang, Wenhui Wang, Nanning Zheng, and Furu Wei. Longnet: Scaling transformers to 1,000,000,000 tokens. arXiv preprint arXiv:2307.02486, 2023. 3", + "[20] Runpei Dong, Chunrui Han, Yuang Peng, Zekun Qi, Zheng Ge, Jinrong Yang, Liang Zhao, Jianjian Sun, Hongyu Zhou, Haoran Wei, et al. Dreamllm: Synergistic multimodal comprehension and creation. In ICLR, 2024. 2", + "[21] Abhimanyu Dubey, Abhinav Jauhri, Abhinav Pandey, Abhishek Kadian, Ahmad Al-Dahle, Aiesha Letman, Akhil Mathur, Alan Schelten, Amy Yang, Angela Fan, et al. The llama 3 herd of models. arXiv preprint arXiv:2407.21783, 2024. 7", + "[22] Manuel Faysse, Hugues Sibille, Tony Wu, Gautier Viaud, Céline Hudelot, and Pierre Colombo. Colpali: Efficient document retrieval with vision language models. arXiv preprint arXiv:2407.01449, 2024. 1, 3", + "[23] Hao Feng, Qi Liu, Hao Liu, Wengang Zhou, Houqiang Li, and Can Huang. Docpedia: Unleashing the power of large multimodal model in the frequency domain for versatile" + ], + "bbox": [ + 516, + 92, + 905, + 900 + ], + "page_idx": 8 + }, + { + "type": "page_number", + "text": "4034", + "bbox": [ + 482, + 944, + 514, + 955 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "document understanding. arXiv preprint arXiv:2311.11810, 2023.3", + "[24] Chaoyou Fu, Peixian Chen, Yunhang Shen, Yulei Qin, Mengdan Zhang, Xu Lin, Zhenyu Qiu, Wei Lin, Jinrui Yang, Xiawu Zheng, et al. Mme: A comprehensive evaluation benchmark for multimodal large language models. arXiv preprint arXiv:2306.13394, 2023. 1", + "[25] Zhangwei Gao, Zhe Chen, Erfei Cui, Yiming Ren, Weiyun Wang, Jinguo Zhu, Hao Tian, Shenglong Ye, Junjun He, Xizhou Zhu, et al. Mini-internvl: A flexible-transfer pocket multimodal model with $5\\%$ parameters and $90\\%$ performance. arXiv preprint arXiv:2410.16261, 2024. 2", + "[26] Chi Han, Qifan Wang, Wenhan Xiong, Yu Chen, Heng Ji, and Sinong Wang. Lm-infinite: Simple on-the-fly length generalization for large language models. arXiv preprint arXiv:2308.16137, 2023. 3", + "[27] Anwen Hu, Haiyang Xu, Liang Zhang, Jiabo Ye, Ming Yan, Ji Zhang, Qin Jin, Fei Huang, and Jingren Zhou. mplug-docowl2: High-resolution compressing for ocr-free multi-page document understanding. arXiv preprint arXiv:2409.03420, 2024. 7, 2", + "[28] Mingxin Huang, Yuliang Liu, Dingkang Liang, Lianwen Jin, and Xiang Bai. Mini-monkey: Alleviate the saw-tooth effect by multi-scale adaptive cropping. arXiv preprint arXiv:2408.02034, 2024. 7", + "[29] Gabriel Ilharco, Mitchell Wortsman, Ross Wightman, Cade Gordon, Nicholas Carlini, Rohan Taori, Achal Dave, Vaishaal Shankar, Hongseok Namkoong, John Miller, Hannaneh Hajishirzi, Ali Farhadi, and Ludwig Schmidt. Openclip. Zenodo. Version 0.1. https://doi.org/10.5281/zenodo.5143773, 2021. DOI: 10.5281/zenodo.5143773. 2", + "[30] Yao Jiang, Xinyu Yan, Ge-Peng Ji, Keren Fu, Meijun Sun, Huan Xiong, Deng-Ping Fan, and Fahad Shahbaz Khan. Effectiveness assessment of recent large vision-language models. Visual Intelligence, 2(1):17, 2024. 1", + "[31] Greg Kamradt. Llmtest_needleinahaystack. https://github.com/gkamradt/LLMTest_NeedleInAHaystack, 2024. Accessed: 2024-11-11. 1", + "[32] Geewook Kim, Teakgyu Hong, Moonbin Yim, JeongYeon Nam, Jinyoung Park, Jinyeong Yim, Wonseok Hwang, Sangdoo Yun, Dongyoon Han, and Seunghyun Park. Ocr-free document understanding transformer. In European Conference on Computer Vision, pages 498-517. Springer, 2022. 3", + "[33] Hugo Laurencon, Andres Marafioti, Victor Sanh, and Léo Tronchon. Building and better understanding vision-language models: insights and future directions. arXiv preprint arXiv:2408.12637, 2024. 2, 6", + "[34] Hugo Laurençon, Lucile Saulnier, Léo Tronchon, Stas Bekman, Amanpreet Singh, Anton Lozhkov, Thomas Wang, Siddharth Karamcheti, Alexander Rush, Douwe Kiela, et al. Obelics: An open web-scale filtered dataset of interleaved image-text documents. NIPS, 36, 2024. 2", + "[35] Bohao Li, Rui Wang, Guangzhi Wang, Yuying Ge, Yixiao Ge, and Ying Shan. Seed-bench: Benchmarking multimodal llms with generative comprehension. arXiv preprint arXiv:2307.16125, 2023. 2" + ], + "bbox": [ + 91, + 90, + 482, + 898 + ], + "page_idx": 9 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[36] Bo Li, Yuanhan Zhang, Dong Guo, Renrui Zhang, Feng Li, Hao Zhang, Kaichen Zhang, Yanwei Li, Ziwei Liu, and Chunyuan Li. Llava-onevision: Easy visual task transfer. arXiv preprint arXiv:2408.03326, 2024. 1, 7", + "[37] Junnan Li, Dongxu Li, Caiming Xiong, and Steven Hoi. Blip: Bootstrapping language-image pre-training for unified vision-language understanding and generation. In ICML, pages 12888-12900, 2022. 2", + "[38] Junnan Li, Dongxu Li, Silvio Savarese, and Steven Hoi. Blip-2: Bootstrapping language-image pre-training with frozen image encoders and large language models. In ICML, pages 19730–19742. PMLR, 2023. 2", + "[39] Qingyun Li, Zhe Chen, Weiyun Wang, Wenhai Wang, Shenglong Ye, Zhenjiang Jin, Guanzhou Chen, Yinan He, Zhangwei Gao, Erfei Cui, et al. Omnicorpus: An unified multimodal corpus of 10 billion-level images interleaved with text. arXiv preprint arXiv:2406.08418, 2024. 2", + "[40] Zhang Li, Biao Yang, Qiang Liu, Zhiyin Ma, Shuo Zhang, Jingxu Yang, Yabo Sun, Yuliang Liu, and Xiang Bai. Monkey: Image resolution and text label are important things for large multi-modal models. arXiv preprint arXiv:2311.06607, 2023. 7", + "[41] Haotian Liu, Chunyuan Li, Yuheng Li, and Yong Jae Lee. Improved baselines with visual instruction tuning. arXiv preprint arXiv:2310.03744, 2023. 1, 6", + "[42] Haotian Liu, Chunyuan Li, Qingyang Wu, and Yong Jae Lee. Visual instruction tuning. NeurIPS, 36, 2023. 1, 2, 6", + "[43] Hao Liu, Matei Zaharia, and Pieter Abbeel. Ring attention with blockwise transformers for near-infinite context. arXiv preprint arXiv:2310.01889, 2023. 2, 3, 6", + "[44] Yuan Liu, Haodong Duan, Yuanhan Zhang, Bo Li, Songyang Zhang, Wangbo Zhao, Yike Yuan, Jiaqi Wang, Conghui He, Ziwei Liu, et al. Mmbench: Is your multi-modal model an all-around player? arXiv preprint arXiv:2307.06281, 2023. 1", + "[45] Yuliang Liu, Zhang Li, Hongliang Li, Wenwen Yu, Mingxin Huang, Dezhi Peng, Mingyu Liu, Mingrui Chen, Chunyuan Li, Lianwen Jin, et al. On the hidden mystery ofOCR in large multimodal models. arXiv preprint arXiv:2305.07895, 2023. 1", + "[46] Yuliang Liu, Biao Yang, Qiang Liu, Zhang Li, Zhiyin Ma, Shuo Zhang, and Xiang Bai. Textmonkey: AnOCR-free large multimodal model for understanding document. arXiv preprint arXiv:2403.04473, 2024. 7", + "[47] Zhaoyang Liu, Yinan He, Wenhai Wang, Weiyun Wang, Yi Wang, Shoufa Chen, Qinglong Zhang, Zeqiang Lai, Yang Yang, Qingyun Li, Jiashuo Yu, et al. Interrupt: Solving vision-centric tasks by interacting with chatgpt beyond language. arXiv preprint arXiv:2305.05662, 2023. 2", + "[48] Ziyu Liu, Tao Chu, Yuhang Zang, Xilin Wei, Xiaoyi Dong, Pan Zhang, Zijian Liang, Yuanjun Xiong, Yu Qiao, Dahua Lin, et al. Mmdu: A multi-turn multi-image dialog understanding benchmark and instruction-tuning dataset for lvlms. arXiv preprint arXiv:2406.11833, 2024. 6", + "[49] Chuwei Luo, Yufan Shen, Zhaoqing Zhu, Qi Zheng, Zhi Yu, and Cong Yao. Layout: Layout instruction tuning with large language models for document understanding. In Pro" + ], + "bbox": [ + 516, + 92, + 903, + 900 + ], + "page_idx": 9 + }, + { + "type": "page_number", + "text": "4035", + "bbox": [ + 482, + 944, + 514, + 955 + ], + "page_idx": 9 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "ceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 15630-15640, 2024. 3", + "[50] Xueguang Ma, Sheng-Chieh Lin, Minghan Li, Wenhu Chen, and Jimmy Lin. Unifying multimodal retrieval via document screenshot embedding. arXiv preprint arXiv:2406.11251, 2024. 1", + "[51] Yubo Ma, Yuhang Zang, Liangyu Chen, Meiqi Chen, Yizhu Jiao, Xinze Li, Xinyuan Lu, Ziyu Liu, Yan Ma, Xiaoyi Dong, Pan Zhang, Liangming Pan, Yu-Gang Jiang, Jiaqi Wang, Yixin Cao, and Aixin Sun. Mmlongbench-doc: Benchmarking long-context document understanding with visualizations, 2024. 1, 7, 8, 5", + "[52] Ahmed Masry, Xuan Long Do, Jia Qing Tan, Shafiq Joty, and Enamul Hoque. Chartqa: A benchmark for question answering about charts with visual and logical reasoning. In ACL, pages 2263-2279, 2022. 6, 8, 1, 5", + "[53] Minesh Mathew, Dimosthenis Karatzas, and CV Jawahar. Docvqa: A dataset for vqa on document images. In WACV, pages 2200-2209, 2021. 6, 8, 1, 5", + "[54] Minesh Mathew, Viraj Bagal, Ruben Tito, Dimosthenis Karatzas, Ernest Valveny, and CV Jawahar. Infographicvqa. In WACV, pages 1697-1706, 2022. 6, 8, 1, 5", + "[55] Anand Mishra, Shashank Shekhar, Ajeet Kumar Singh, and Anirban Chakraborty. Ocr-vqa: Visual question answering by reading text in images. In ICDAR, pages 947-952, 2019. 6", + "[56] OpenAI. Gpt-4v(ison) system card. https://cdn.openai.com/papers/GPTV_System/Card.pdf, 2023.1,4,7", + "[57] OpenAI. Gpt-4o system card. https://openai.com/index/gpt-4o-system-card/, 2024.7", + "[58] Ofir Press, Noah A Smith, and Mike Lewis. Train short, test long: Attention with linear biases enables input length extrapolation. arXiv preprint arXiv:2108.12409, 2021. 3", + "[59] Sylvestre-Alvise Rebuffi, Hakan Bilen, and Andrea Vedaldi. Learning multiple visual domains with residual adapters. NeurIPS, 30, 2017. 2", + "[60] Machel Reid, Nikolay Savinov, Denis Teptyashin, Dmitry Lepikhin, Timothy Lillicrap, Jean-baptiste Alayrac, Radu Soricut, Angeliki Lazaridou, Orhan Firat, Julian Schrittwieser, et al. Gemini 1.5: Unlocking multimodal understanding across millions of tokens of context. arXiv preprint arXiv:2403.05530, 2024. 1, 7", + "[61] Sahel Sharifymoghaddam, Shivani Upadhyay, Wenhu Chen, and Jimmy Lin. Unirag: Universal retrieval augmentation for multi-modal large language models. arXiv preprint arXiv:2405.10311, 2024. 1", + "[62] Min Shi, Fuxiao Liu, Shihao Wang, Shijia Liao, Subhashree Radhakrishnan, De-An Huang, Hongxu Yin, Karan Sapra, Yaser Yacoob, Humphrey Shi, et al. Eagle: Exploring the design space for multimodal llms with mixture of encoders. arXiv preprint arXiv:2408.15998, 2024. 1", + "[63] Weijia Shi, Sewon Min, Michihiro Yasunaga, Minjoon Seo, Rich James, Mike Lewis, Luke Zettlemoyer, and Wen-tau Yih. Replug: Retrieval-augmented black-box language models. arXiv preprint arXiv:2301.12652, 2023. 1" + ], + "bbox": [ + 91, + 92, + 482, + 900 + ], + "page_idx": 10 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[64] Oleksii Sidorov, Ronghang Hu, Marcus Rohrbach, and Amanpreet Singh. Textcaps: a dataset for image captioning with reading comprehension. In ECCV, pages 742-758, 2020. 6", + "[65] Amanpreet Singh, Vivek Natarajan, Meet Shah, Yu Jiang, Xinlei Chen, Dhruv Batra, Devi Parikh, and Marcus Rohrbach. Towards vqa models that can read. In CVPR, pages 8317-8326, 2019. 1, 6", + "[66] Jianlin Su, Murtadha Ahmed, Yu Lu, Shengfeng Pan, Wen Bo, and Yunfeng Liu. Roformer: Enhanced transformer with rotary position embedding. Neurocomputing, 568:127063, 2024. 3", + "[67] Quan Sun, Qiying Yu, Yufeng Cui, Fan Zhang, Xiaosong Zhang, Yueze Wang, Hongcheng Gao, Jingjing Liu, Tiejun Huang, and Xinlong Wang. Generative pretraining in multimodality. In ICLR, 2024. 2", + "[68] Yutao Sun, Li Dong, Barun Patra, Shuming Ma, Shaohan Huang, Alon Benhaim, Vishrav Chaudhary, Xia Song, and Furu Wei. A length-extrapolatable transformer. arXiv preprint arXiv:2212.10554, 2022. 3", + "[69] Mohammad Reza Taesiri. Arxivqa. https://github.com/taesiri/ArXivQA, 2024.6", + "[70] Ryota Tanaka, Kyosuke Nishida, and Sen Yoshida. Visualmrc: Machine reading comprehension on document images. In Proceedings of the AAAI Conference on Artificial Intelligence, pages 13878-13888, 2021. 6", + "[71] Zineng Tang, Ziyi Yang, Guoxin Wang, Yuwei Fang, Yang Liu, Chenguang Zhu, Michael Zeng, Cha Zhang, and Mohit Bansal. Unifying vision, text, and layout for universal document processing. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 19254-19264, 2023. 2", + "[72] Gemini Team, Rohan Anil, Sebastian Borgeaud, Yonghui Wu, Jean-Baptiste Alayrac, Jiahui Yu, Radu Soricut, Johan Schalkwyk, Andrew M Dai, Anja Hauth, et al. Gemini: a family of highly capable multimodal models. arXiv preprint arXiv:2312.11805, 2023. 1", + "[73] OpenGVLab Team. Internv12: Better than the best—expanding performance boundaries of open-source multimodal models with the progressive scaling strategy, 2024. 1, 6, 7, 8", + "[74] Qwen Team. Qwen2.5: A party of foundation models, 2024. 1", + "[75] Changyao Tian, Xizhou Zhu, Yuwen Xiong, Weiyun Wang, Zhe Chen, Wenhai Wang, Yuntao Chen, Lewei Lu, Tong Lu, Jie Zhou, et al. Mm-interleaved: Interleaved image-text generative modeling via multi-modal feature synchronizer. arXiv preprint arXiv:2401.10208, 2024. 2", + "[76] Rubén Tito, Dimosthenis Karatzas, and Ernest Valveny. Hierarchical multimodal transformers for multipage docvqa. Pattern Recognition, 144:109834, 2023. 1, 2, 7, 5", + "[77] Xiaoguang Tu, Zhi He, Yi Huang, Zhi-Hao Zhang, Ming Yang, and Jian Zhao. An overview of large ai models and their applications. Visual Intelligence, 2(1):1-22, 2024. 1", + "[78] Jordy Van Landeghem, Ruben Tito, Łukasz Borchmann, Michal Pietruszka, Paweł Joziak, Rafał Powalski, Dawid Jurkiewicz, Mickael Coustaty, Bertrand Ackaert, Ernest Val" + ], + "bbox": [ + 516, + 92, + 903, + 900 + ], + "page_idx": 10 + }, + { + "type": "page_number", + "text": "4036", + "bbox": [ + 482, + 944, + 516, + 955 + ], + "page_idx": 10 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "veny, et al. Document understanding dataset and evaluation (dude). In Proceedings IEEE/CVF international conference on computer vision-ICCV 2023, pages 19528-19540. IEEE/CVF, 2023. 2, 6", + "[79] Bin Wang, Chao Xu, Xiaomeng Zhao, Linke Ouyang, Fan Wu, Zhiyuan Zhao, Rui Xu, Kaiwen Liu, Yuan Qu, Fukai Shang, et al. Mineru: An open-source solution for precise document content extraction. arXiv preprint arXiv:2409.18839, 2024. 3", + "[80] Dongsheng Wang, Natraj Raman, Mathieu Sibue, Zhiqiang Ma, Petr Babkin, Simerjot Kaur, Yulong Pei, Armineh Nourbakhsh, and Xiaomo Liu. Docllm: A layout-aware generative language model for multimodal document understanding. arXiv preprint arXiv:2401.00908, 2023. 2", + "[81] Peng Wang, Shuai Bai, Sinan Tan, Shijie Wang, Zhihao Fan, Jinze Bai, Keqin Chen, Xuejing Liu, Jialin Wang, Wenbin Ge, et al. Qwen2-vl: Enhancing vision-language model's perception of the world at any resolution. arXiv preprint arXiv:2409.12191, 2024. 1", + "[82] Weihan Wang, Qingsong Lv, Wenmeng Yu, Wenyi Hong, Ji Qi, Yan Wang, Junhui Ji, Zhuoyi Yang, Lei Zhao, Xixuan Song, et al. Cogvlm: Visual expert for pretrained language models. arXiv preprint arXiv:2311.03079, 2023. 1", + "[83] Weiyun Wang, Yiming Ren, Haowen Luo, Tiantong Li, Chenxiang Yan, Zhe Chen, Wenhai Wang, Qingyun Li, Lewei Lu, Xizhou Zhu, et al. The all-seeing project v2: Towards general relation comprehension of the open world. arXiv preprint arXiv:2402.19474, 2024. 1, 2", + "[84] Weiyun Wang, Min Shi, Qingyun Li, Wenhai Wang, Zhenhang Huang, Linjie Xing, Zhe Chen, Hao Li, Xizhou Zhu, Zhiguo Cao, et al. The all-seeing project: Towards panoptic visual recognition and understanding of the open world. In ICLR, 2024. 1, 2", + "[85] Weiyun Wang, Shuibio Zhang, Yiming Ren, Yuchen Duan, Tiantong Li, Shuo Liu, Mengkang Hu, Zhe Chen, Kaipeng Zhang, Lewei Lu, Xizhou Zhu, Ping Luo, Yu Qiao, Jifeng Dai, Wenqi Shao, and Wenhai Wang. Needle in a multimodal haystack. arXiv preprint arXiv:2406.07230, 2024. 7, 8, 3", + "[86] Weiyun Wang, Shuibio Zhang, Yiming Ren, Yuchen Duan, Tiantong Li, Shuo Liu, Mengkang Hu, Zhe Chen, Kaipeng Zhang, Lewei Lu, et al. Needle in a multimodal haystack. arXiv preprint arXiv:2406.07230, 2024. 2, 7, 8, 5", + "[87] Yonghui Wang, Wengang Zhou, Hao Feng, Keyi Zhou, and Houqiang Li. Towards improving document understanding: An exploration on text-grounding via mllms. arXiv preprint arXiv:2311.13194, 2023. 3", + "[88] Haoran Wei, Lingyu Kong, Jinyue Chen, Liang Zhao, Zheng Ge, Jinrong Yang, Jianjian Sun, Chunrui Han, and Xiangyu Zhang. Vary: Scaling up the vision vocabulary for large vision-language models. arXiv preprint arXiv:2312.06109, 2023. 3", + "[89] Renqiu Xia, Song Mao, Xiangchao Yan, Hongbin Zhou, Bo Zhang, Haoyang Peng, Jiahao Pi, Daocheng Fu, Wenjie Wu, Hancheng Ye, et al. Docgenome: An open largescale scientific document benchmark for training and testing multi-modal large language models. arXiv preprint arXiv:2406.11633, 2024. 1, 7, 2, 5" + ], + "bbox": [ + 91, + 90, + 483, + 898 + ], + "page_idx": 11 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[90] Guangxuan Xiao, Yuandong Tian, Beidi Chen, Song Han, and Mike Lewis. Efficient streaming language models with attention sinks. arXiv preprint arXiv:2309.17453, 2023. 3", + "[91] Jianxin Yang. Longqlora: Efficient and effective method to extend context length of large language models. arXiv preprint arXiv:2311.04879, 2023. 2, 6", + "[92] Yuan Yao, Tianyu Yu, Ao Zhang, Chongyi Wang, Junbo Cui, Hongji Zhu, Tianchi Cai, Haoyu Li, Weilin Zhao, Zhihui He, et al. Minicpm-v: A gpt-4v level mllm on your phone. arXiv preprint arXiv:2408.01800, 2024. 1, 7, 8", + "[93] Jiabo Ye, Anwen Hu, Haiyang Xu, Qinghao Ye, Ming Yan, Yuhao Dan, Chenlin Zhao, Guohai Xu, Chenliang Li, Junfeng Tian, et al. mplug-docowl: Modularized multimodal large language model for document understanding. arXiv preprint arXiv:2307.02499, 2023. 6", + "[94] Jiabo Ye, Anwen Hu, Haiyang Xu, Qinghao Ye, Ming Yan, Guohai Xu, Chenliang Li, Junfeng Tian, Qi Qian, Ji Zhang, et al. Ureader: UniversalOCR-free visually-situated language understanding with multimodal large language model. arXiv preprint arXiv:2310.05126, 2023. 3", + "[95] Shi Yu, Chaoyue Tang, Bokai Xu, Junbo Cui, Junhao Ran, Yukun Yan, Zhenghao Liu, Shuo Wang, Xu Han, Zhiyuan Liu, et al. Visrag: Vision-based retrieval-augmented generation on multi-modality documents. arXiv preprint arXiv:2410.10594, 2024. 1, 2, 3, 7, 8", + "[96] Jiajie Zhang, Yushi Bai, Xin Lv, Wanjun Gu, Danqing Liu, Minhao Zou, Shulin Cao, Lei Hou, Yuxiao Dong, Ling Feng, et al. Longcite: Enabling llms to generate fine-grained citations in long-context qa. arXiv e-prints, pages arXiv-2409, 2024. 2, 6", + "[97] Jiajie Zhang, Zhongni Hou, Xin Lv, Shulin Cao, Zhenyu Hou, Yilin Niu, Lei Hou, Yuxiao Dong, Ling Feng, and Juanzi Li. Longreward: Improving long-context large language models with ai feedback. arXiv preprint arXiv:2410.21252, 2024. 2, 6", + "[98] Pan Zhang, Xiaoyi Dong, Yuhang Zang, Yuhang Cao, Rui Qian, Lin Chen, Qipeng Guo, Haodong Duan, Bin Wang, Linke Ouyang, et al. Internlm-xcomposer-2.5: A versatile large vision language model supporting long-contextual input and output. arXiv preprint arXiv:2407.03320, 2024. 7", + "[99] Wanrong Zhu, Jack Hessel, Anas Awadalla, Samir Yitzhak Gadre, Jesse Dodge, Alex Fang, Youngjae Yu, Ludwig Schmidt, William Yang Wang, and Yejin Choi. Multimodal c4: An open, billion-scale corpus of images interleaved with text. NIPS, 36, 2024. 2" + ], + "bbox": [ + 516, + 90, + 903, + 724 + ], + "page_idx": 11 + }, + { + "type": "page_number", + "text": "4037", + "bbox": [ + 482, + 944, + 514, + 955 + ], + "page_idx": 11 + } +] \ No newline at end of file diff --git a/2025/Docopilot_ Improving Multimodal Models for Document-Level Understanding/2cc56609-3eb1-412a-98df-28e82ae9cd58_model.json b/2025/Docopilot_ Improving Multimodal Models for Document-Level Understanding/2cc56609-3eb1-412a-98df-28e82ae9cd58_model.json new file mode 100644 index 0000000000000000000000000000000000000000..4db60ee699af683e11a2bf915eefef5ea9ef05fd --- /dev/null +++ b/2025/Docopilot_ Improving Multimodal Models for Document-Level Understanding/2cc56609-3eb1-412a-98df-28e82ae9cd58_model.json @@ -0,0 +1,2897 @@ +[ + [ + { + "type": "header", + "bbox": [ + 0.107, + 0.003, + 0.182, + 0.043 + ], + "angle": 0, + "content": "CVF" + }, + { + "type": "header", + "bbox": [ + 0.237, + 0.001, + 0.808, + 0.047 + ], + "angle": 0, + "content": "This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore." + }, + { + "type": "title", + "bbox": [ + 0.105, + 0.131, + 0.895, + 0.156 + ], + "angle": 0, + "content": "Docopilot: Improving Multimodal Models for Document-Level Understanding" + }, + { + "type": "text", + "bbox": [ + 0.158, + 0.181, + 0.842, + 0.218 + ], + "angle": 0, + "content": "Yuchen Duan\\(^{1,2*}\\), Zhe Chen\\(^{3,1*}\\), Yusong Hu\\(^{4,1*}\\), Weiyun Wang\\(^{*5,1}\\), Shenglong Ye\\(^{1}\\), Botian Shi\\(^{1}\\), Lewei Lu\\(^{7}\\), Qibin Hou\\(^{4}\\), Tong Lu\\(^{3,1}\\), Hongsheng Li\\(^{2,1}\\), Jifeng Dai\\(^{6,1}\\), Wenhai Wang\\(^{2,1\\boxtimes}\\)" + }, + { + "type": "text", + "bbox": [ + 0.186, + 0.222, + 0.816, + 0.259 + ], + "angle": 0, + "content": "\\(^{1}\\)Shanghai AI Laboratory, \\(^{2}\\)The Chinese University of Hong Kong, \\(^{3}\\)Nanjing University, \\(^{4}\\)Nankai University, \\(^{5}\\)Fudan University, \\(^{6}\\)Tsinghua University, \\(^{7}\\)SenseTime Research" + }, + { + "type": "title", + "bbox": [ + 0.248, + 0.293, + 0.327, + 0.31 + ], + "angle": 0, + "content": "Abstract" + }, + { + "type": "text", + "bbox": [ + 0.089, + 0.325, + 0.486, + 0.658 + ], + "angle": 0, + "content": "Despite significant progress in multimodal large language models (MLLMs), their performance on complex, multipage document comprehension remains inadequate, largely due to the lack of high-quality, document-level datasets. While current retrieval-augmented generation (RAG) methods offer partial solutions, they suffer from issues, such as fragmented retrieval contexts, multi-stage error accumulation, and extra time costs of retrieval. In this work, we present a high-quality document-level dataset, Doc750K, designed to support in-depth understanding of multimodal documents. This dataset includes diverse document structures, extensive cross-page dependencies, and real question-answer pairs derived from the original documents. Building on the dataset, we develop a native multimodal model—Docopilot, which can accurately handle document-level dependencies without relying on RAG. Experiments demonstrate that Docopilot achieves superior coherence, accuracy, and efficiency in document understanding tasks and multi-turn interactions, setting a new baseline for document-level multimodal understanding. Data, code, and models are released at https://github.com/OpenGVLab/Docopilot." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.675, + 0.224, + 0.691 + ], + "angle": 0, + "content": "1. Introduction" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.7, + 0.485, + 0.868 + ], + "angle": 0, + "content": "In recent years, multimodal large language models (MLLMs) [6, 12, 41, 56, 60, 72, 81, 82, 84, 92] have rapidly developed, achieving remarkable performance in various visual understanding tasks [30, 77], particularly image-level tasks, such as image captioning [10, 42], optical character recognition (OCR) [45, 65], and visual question answering (VQA) [24, 44]. Despite these advances, current MLLMs still face significant challenges in document-level understanding [51, 76, 89], where models are required to identify and integrate key information across multi-page documents, setting high expectations for their long-context processing" + }, + { + "type": "image", + "bbox": [ + 0.515, + 0.279, + 0.907, + 0.475 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.512, + 0.485, + 0.909, + 0.61 + ], + "angle": 0, + "content": "Figure 1. Accuracy \\( \\nu.s \\) inference latency on MM-NIAH. The proposed Docopilot-8B shows a notable improvement over baseline models [73], achieving a \\( +19.9\\% \\) accuracy gain compared to InternVL2-8B and surpassing InternVL2-26B with less than \\( 31\\% \\) of the inference latency. Additionally, Docopilot-2B uses fewer parameters (less than \\( 10\\% \\)) while exhibiting comparable performance to the \\( 10\\times \\) larger InternVL2-26B. These results suggest that our Docopilot strikes a reasonable balance between latency, model size, and performance." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.628, + 0.673, + 0.643 + ], + "angle": 0, + "content": "capabilities of MLLMs." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.644, + 0.909, + 0.903 + ], + "angle": 0, + "content": "Current research on long-content understanding primarily focuses on text-only models [5, 9, 74], targeting specific retrieval tasks such as \"Needle in a Haystack\" (NIAH) [2, 31]. However, existing open-source MLLMs [12, 13, 36, 41, 62, 81, 83] are primarily trained on image-level data, lacking the long-context understanding capacity required for document-level understanding. Retrieval-augmented generation (RAG) methods [14, 22, 50, 61, 63, 95] attempt to address this by retrieving key information to fit within the limited context windows of MLLMs, but they still encounter the following challenges in document-level tasks. (1) Fragmented Retrieval Contexts. Retrieved information is fragmented, lacking the overall structure of the document; (2) Multi-Stage Error Accumulation. Incorrect retrieval results can affect subsequent responses, leading to errors or omissions of critical details, especially in multi-turn or complex tasks; (3) Extra Time Costs. The retrieval" + }, + { + "type": "page_footnote", + "bbox": [ + 0.11, + 0.876, + 0.232, + 0.888 + ], + "angle": 0, + "content": "* Equal contribution;" + }, + { + "type": "page_footnote", + "bbox": [ + 0.109, + 0.889, + 0.388, + 0.901 + ], + "angle": 0, + "content": "Corresponding author: wangwenhai@pjlab.org.cn" + }, + { + "type": "list", + "bbox": [ + 0.109, + 0.876, + 0.388, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.517, + 0.957 + ], + "angle": 0, + "content": "4026" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.09, + 0.092, + 0.482, + 0.121 + ], + "angle": 0, + "content": "step increases the latency of the QA system, limiting the scalability of RAG in time-sensitive scenarios." + }, + { + "type": "text", + "bbox": [ + 0.089, + 0.123, + 0.483, + 0.319 + ], + "angle": 0, + "content": "To address these problems, two primary challenges need to be considered. (1) High-Quality Multimodal Document Dataset. While extensive datasets [11, 91, 96, 97] exist for long-context, text-only tasks, high-quality document-level question-answering datasets remain scarce. This shortage is largely attributed to the high costs associated with annotation and the lack of streamlined construction pipelines. (2) Native Document-Level MLLMs. Although RAG-based methods [14, 86, 95] provide some relief, native multimodal models with long-context processing abilities are crucial. However, training native MLLMs specifically for document-level understanding is constrained by current hardware limitations." + }, + { + "type": "text", + "bbox": [ + 0.093, + 0.322, + 0.485, + 0.577 + ], + "angle": 0, + "content": "In this work, we introduce a new multimodal document dataset that supports document-level understanding tasks. Compared to counterparts [33, 76, 78], this dataset has the following features: (1) Large Scale. It includes a total of 758K question-answer samples, containing 5.2B text tokens and 3.1M images. It encompasses content from various sources, such as Sci-Hub, Arxiv, and OpenReview, covering a wide range of topics and document layouts. (2) High Quality. Unlike existing datasets that insert irrelevant questions into documents, we collect real, in-depth question-answer pairs and construct single-page and crosspage questions based on document structure. Such high-quality question-answer data accounts for \\(31.6\\%\\) of the dataset. (3) Multimodal. For the document content, we provide not only the conventional interleaved text-image context but also purely rendered image inputs, catering to the needs of different models." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.581, + 0.483, + 0.806 + ], + "angle": 0, + "content": "Building upon this dataset, we developed a native baseline model for document-level multimodal understanding-Docopilot. Unlike existing approaches [14, 86, 95] that rely on RAG, our model achieves efficient document-level training and testing through simple engineering optimizations, such as multimodal data packing, Ring Attention [43], and Liger Kernel [16]. Leveraging the proxy tasks carefully designed within the dataset, Docopilot can directly handle long-distance dependencies and cross-page information integration without external retrieval support. As shown in Figure 1, this approach not only enhances coherence and accuracy compared to RAG methods but also significantly reduces the response time of the entire question-answering system, delivering superior real-time performance in multi-turn interactions." + }, + { + "type": "text", + "bbox": [ + 0.11, + 0.809, + 0.451, + 0.823 + ], + "angle": 0, + "content": "The main contributions are summarized as follows:" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.826, + 0.483, + 0.901 + ], + "angle": 0, + "content": "(1) We develop the first large-scale, high-quality dataset for document-level multimodal understanding, consisting of 758K QA pairs from 3 sources, supporting 9 types of proxy tasks. This dataset includes \\(31.6\\%\\) real QA pairs directly extracted from documents." + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.092, + 0.905, + 0.167 + ], + "angle": 0, + "content": "(2) Based on the dataset, we implement Docopilot, a native MLLM designed for document-level understanding without relying on retrieval mechanisms. This approach greatly improves its ability to integrate and comprehend information across multi-page documents." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.168, + 0.907, + 0.304 + ], + "angle": 0, + "content": "(3) Through extensive experiments on multiple document-level benchmarks, our method demonstrates performance significantly superior to existing approaches, proving its effectiveness and generality. As shown in Figure 1, Docopilot-8B achieves a score of 61.8 on MM-NIAH [86], outperforming InternVL2-8B by 19.9 points and surpassing InternVL2-26B with less than \\(31\\%\\) of the latency. We hope this work could provide a baseline for future advancements in MLLMs for document-level tasks." + }, + { + "type": "list", + "bbox": [ + 0.512, + 0.092, + 0.907, + 0.304 + ], + "angle": 0, + "content": null + }, + { + "type": "title", + "bbox": [ + 0.514, + 0.32, + 0.655, + 0.336 + ], + "angle": 0, + "content": "2. Related Work" + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.345, + 0.833, + 0.362 + ], + "angle": 0, + "content": "2.1. Multimodal Large Language Models" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.368, + 0.908, + 0.732 + ], + "angle": 0, + "content": "Multimodal large language models (MLLMs) have demonstrated impressive capabilities in processing image and text information, opening up new directions for applications such as visual question answering and image captioning. Early models [12, 29, 37, 59] trained with contrastive learning methods excelled in recognizing and understanding open-world semantics within an image-text matching framework. However, their limited generative abilities restricted their applicability. To leverage the powerful generation abilities of large language models (LLMs), subsequent works [13, 38, 42, 47, 83, 84] introduced a connector to align the embedding spaces of vision encoders and LLMs, allowing encoded image embeddings to serve as soft prompts for LLMs. Another series of works [1, 34, 39, 99] extended LLMs by integrating additional visual experts, reducing reliance on standalone vision encoders. More recently, models capable of both understanding and generating images have also made notable progress [20, 35, 67, 75], leveraging the insight that image generation can enhance image understanding. Despite these advancements, current MLLMs still face challenges with long-context multimodal inputs. For instance, InternVL 2.0 [13, 25] performs optimally within a token range of up to 8192, constraining its effectiveness in document-level applications." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.743, + 0.809, + 0.759 + ], + "angle": 0, + "content": "2.2. Document Understanding Models" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.765, + 0.907, + 0.901 + ], + "angle": 0, + "content": "Extracting key information from documents is crucial for industries and academic research. OCR-model-driven methods [3, 4, 71, 80] represent one of the primary technical approaches. These methods extract text, layout, and bounding box information from external systems and integrate it with another model. However, they are prone to error propagation and high processing times due to their reliance on multiple components. Benefitting from the rapid advancements in LLMs, OCR-free methods have also achieved" + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.516, + 0.957 + ], + "angle": 0, + "content": "4027" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.093, + 0.087, + 0.905, + 0.309 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.09, + 0.318, + 0.908, + 0.377 + ], + "angle": 0, + "content": "Figure 2. Multimodal document dataset generation pipeline. This pipeline involves three main stages: (1) Raw Data Collection: Documents are gathered from sources like Sci-Hub, arXiv, and OpenReview, available in PDF and HTML formats. (2) Document Content Extraction: Multimodal content is processed in two formats: interleaved text-image format and multi-image format. (3) Question-Answer Pairs Construction: QA pairs are generated based on the document structure or constructed using GPT-4o." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.394, + 0.484, + 0.485 + ], + "angle": 0, + "content": "great progress. Donut [32] is the first end-to-end training framework based on a Transformer without requiring OCR engines or APIs. Subsequent works [23, 49, 87, 88, 94] propose diverse modifications in model architectures and training algorithms. However, these models are designed for specific tasks and lack general abilities." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.5, + 0.427, + 0.516 + ], + "angle": 0, + "content": "2.3. Long-Context Large Language Models" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.524, + 0.484, + 0.903 + ], + "angle": 0, + "content": "With advancements in engineering, architecture, and algorithms, long-context large language models have made substantial progress. Techniques such as Flash Attention [17, 18] and Ring Attention [43] have notably reduced GPU memory usage for training on extended contexts. Additionally, various sparse attention mechanisms—including Shifting Sparse Attention [11], Dilated Attention [19], and Attention Sinks [26, 90]—have enabled efficient scaling to handle larger contexts. New positional embedding methods, like ALiBi [58], xPOS [68], and RoPE [66], further enhance the models' generalization capabilities in length extrapolation. However, these advancements remain largely confined to natural language processing, and methods to extend the context size of MLLMs are still under-explored. Another research approach aims to reduce context size by leveraging retrieval augmented generation (RAG) [14, 22, 95], where only the most relevant passages are retrieved and fed into the generation model. However, this retrieval-based approach can disrupt the coherence of the semantic chain, particularly in complex reasoning tasks, due to fragmented information flow. In this work, we integrate the above engineering techniques into MLLMs and demonstrate that a model fine-tuned on a high-quality, long-context training corpus is a strong baseline, achieving superior performance compared to its RAG counterpart." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.392, + 0.896, + 0.409 + ], + "angle": 0, + "content": "3. Multimodal Document Dataset Generation" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.417, + 0.907, + 0.462 + ], + "angle": 0, + "content": "In this section, we begin by introducing the details of the data engine. Following this, we provide a comprehensive overview of the dataset—Doc-750K." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.471, + 0.649, + 0.487 + ], + "angle": 0, + "content": "3.1. Data Engine" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.493, + 0.907, + 0.613 + ], + "angle": 0, + "content": "The data engine operates primarily through two steps: document content extraction and question-answer (QA) pair construction. Specifically, we first extract multimodal content, including both text and images, from the documents. Based on this extracted content, we then create question-answer pairs. The document content and these constructed pairs are combined to produce conversational-style training data. The format is outlined as:" + }, + { + "type": "text", + "bbox": [ + 0.522, + 0.624, + 0.88, + 0.664 + ], + "angle": 0, + "content": "Please read the paper: , and answer the question: Answer: " + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.675, + 0.907, + 0.75 + ], + "angle": 0, + "content": "Here, the , , and are the placeholder for extracted document content, the generated questions and answer, respectively. In the following, we will provide a detailed explanation of the two key steps: document content extraction and QA pair construction." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.75, + 0.909, + 0.854 + ], + "angle": 0, + "content": "Document Content Extraction. In practical applications, different documents have varying page layouts and content types, which poses significant challenges for content extraction. To enhance the efficiency of multimodal models, it is necessary to organize documents into a unified format for streamlined processing. In this work, we process each document into two formats as follows:" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.856, + 0.909, + 0.901 + ], + "angle": 0, + "content": "(1) Interleaved Text-Image Format. Using the document content extractor MinerU [79], we segment the document content into interleaved text and image annotations, for" + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.516, + 0.957 + ], + "angle": 0, + "content": "4028" + } + ], + [ + { + "type": "header", + "bbox": [ + 0.098, + 0.09, + 0.131, + 0.123 + ], + "angle": 0, + "content": "0" + }, + { + "type": "header", + "bbox": [ + 0.147, + 0.103, + 0.317, + 0.112 + ], + "angle": 0, + "content": "How Many Kilowatts are in a Negawatt?" + }, + { + "type": "header", + "bbox": [ + 0.147, + 0.113, + 0.495, + 0.123 + ], + "angle": 0, + "content": "Verifying \"Ex Post\" Estimates of Utility Conservation Impacts at the Regional Level" + }, + { + "type": "text", + "bbox": [ + 0.146, + 0.135, + 0.49, + 0.173 + ], + "angle": 0, + "content": "The objective consideration of conservation policy under restructuring is proving to be a difficult task. One of the greatest obstacles has been the persistent uncertainty among utility planners regarding the true resource-effectiveness and cost-effectiveness of conservation relative" + }, + { + "type": "text", + "bbox": [ + 0.149, + 0.181, + 0.416, + 0.192 + ], + "angle": 0, + "content": "These figures indicate that larger utilities tend to report greater ..." + }, + { + "type": "image", + "bbox": [ + 0.163, + 0.194, + 0.266, + 0.277 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.289, + 0.199, + 0.482, + 0.27 + ], + "angle": 0, + "content": null + }, + { + "type": "text", + "bbox": [ + 0.146, + 0.28, + 0.496, + 0.309 + ], + "angle": 0, + "content": "Early estimation attempts with individual utilities revealed that there was insufficient variation in the data series to produce reliable coefficient estimates for each utility taken separately. To increase the statistical power of the analysis, ..." + }, + { + "type": "text", + "bbox": [ + 0.148, + 0.318, + 0.495, + 0.337 + ], + "angle": 0, + "content": "Finally, since no incentives are offered, it becomes more difficult to attribute customer investment decisions to the utility programs as opposed to other influences." + }, + { + "type": "table", + "bbox": [ + 0.153, + 0.34, + 0.328, + 0.444 + ], + "angle": 0, + "content": "
Utility (C&I Sales)Program Type1993 Conservation Impacts(Annotated)
GWh% of Sales% of Impacts
Pacific Gas & Electric(45,760 GWh)Incentives1,8864.1%63%
Services9402.1%31%
Other170.4%6%
Total3,0136.6%100%
San Diego Gas & Electric(8,415 GWh)Incentives2512.7%65%
Services971.0%25%
Other2820.4%100%
Total3854.1%100%
Southern California Edison(86,721 GWh)Incentives1,7283.7%22%
Services6,02912.9%71%
Other2400.5%3%
Total7,97417.1%100%
Sacramento(8,733 GWh)Incentives1533.2%58%
Services1042.2%40%
Other60.1%2%
Total2822.5%100%
Los Angeles(14,829 GWh)Incentives7800.5%50%
Services7910.5%50%
Other00.0%0%
Total1571.1%100%
" + }, + { + "type": "table", + "bbox": [ + 0.349, + 0.34, + 0.492, + 0.443 + ], + "angle": 0, + "content": "
StudySampleFactorPeriodElectricity
Elec.UnitCostFuel
Longuet et al. 1994Norway 1976aenergy consumptionLB−0.39−0.39
energy consumptionLB−0.15−0.15
energy consumption & transportLB−0.44−0.16
Hammoudi & Kazer 1995Iran 1955–8domestic heatingLB−0.19−0.19
Safdarzadeh 1995USAcommercialLB−0.70−0.25
industrialLB−0.22−0.22
Rall 1996USAindustrialLB−0.18−0.20−0.150.09
industrialLB−0.29−0.29
Counihan 1998USA 1995–8industrialLB−0.390.530.040.09
industrialLB−0.22−0.22
Murray 2000Virginia, USA 1970–8industrialLB−0.29−0.09
industrialLB−0.22−0.22
industrialLB−0.21−0.04
Woodward 2001Netherlands 1977–83industrialLB−1.11−0.2−0.39−0.16
industrialLB−0.11−0.11
National Energy Board 2003CanadaindustrialLB−0.110.020.05
industrialLB−0.11−0.11
industrialLB−0.450.120.06
Energy, Media and Media Ethics Board Canada 2005CanadaindustrialLB−0.090.030.01
industrialLB−0.09−0.09
industrialLB−0.490.110.01
This studyUSAcommercial RLB−0.078−0.033−0.026−0.013
industrialLB−0.078−0.078
" + }, + { + "type": "text", + "bbox": [ + 0.146, + 0.451, + 0.496, + 0.528 + ], + "angle": 0, + "content": "According to California regulators, SCE's reported impacts for service programs have always been viewed with skepticism during regulatory reviews, although its impacts for other types of conservation programs have not. In California, service program impacts have historically been viewed as unverifiable, so utilities have been largely ineligible for incentive payments related to service program expenditures. Regulators in California acknowledge that they believe SCE's reported conservation impacts overstate the actual savings achieved by their programs and have communicated this skepticism to the utility. But, according to the rate regulation ..." + }, + { + "type": "text", + "bbox": [ + 0.54, + 0.103, + 0.578, + 0.112 + ], + "angle": 0, + "content": "" + }, + { + "type": "text", + "bbox": [ + 0.54, + 0.113, + 0.732, + 0.121 + ], + "angle": 0, + "content": "Please read the paper and answer the question:" + }, + { + "type": "text", + "bbox": [ + 0.539, + 0.123, + 0.826, + 0.153 + ], + "angle": 0, + "content": "According to this paper, how does Image-1 illustrate the conservation impacts of various utilities in 1993? Discuss the findings and the significance of these data in the context of the research." + }, + { + "type": "header", + "bbox": [ + 0.855, + 0.091, + 0.9, + 0.128 + ], + "angle": 0, + "content": "#" + }, + { + "type": "image", + "bbox": [ + 0.532, + 0.17, + 0.569, + 0.203 + ], + "angle": 0, + "content": null + }, + { + "type": "text", + "bbox": [ + 0.582, + 0.183, + 0.892, + 0.328 + ], + "angle": 0, + "content": "According to the paper, 'In 1993, the 39 companies in the sample accounted for \\(19\\%\\) of total C&I electricity sales nationwide and \\(40\\%\\) of all utility conservation program impacts, including residential impacts (Energy Information Administration 1995). The table in Image-1 provides detailed data on the 1993 C&I sales and C&I conservation program impacts for various utilities. This image shows the sales in GWh and conservation impacts both in GWh and as a percentage of sales. For example, California utilities such as Southern California Edison reported high sales and high conservation percentages, which is a significant outlier. These data illustrate the varying levels of conservation achievements across utilities, underlining the study's aim to verify the effectiveness and accuracy of reported conservation impacts. Larger utilities tend to report greater conservation impacts relative to sales, and this detail supports the analysis presented in the paper." + }, + { + "type": "text", + "bbox": [ + 0.536, + 0.359, + 0.824, + 0.38 + ], + "angle": 0, + "content": "Analyze the importance of Image-3 in comparing the study's findings with previous empirical research on electricity price elasticities." + }, + { + "type": "image", + "bbox": [ + 0.855, + 0.35, + 0.899, + 0.385 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.531, + 0.397, + 0.569, + 0.43 + ], + "angle": 0, + "content": null + }, + { + "type": "text", + "bbox": [ + 0.582, + 0.409, + 0.893, + 0.533 + ], + "angle": 0, + "content": "According to the paper, 'The regression diagnostics offered little guidance on improving the structure of the model. The addition of a dummy variable for the years 1971-1973 did little to improve the residuals in subsequent years. Adding a set of dummy variables to account for each unstable period did not seem reasonable.' Image-3 presents a comparative analysis of electricity cross-price elasticity studies. This table showcases various studies, their sample sectors, and reported elasticities for electricity, natural gas, oil, and coal across different regions and time periods. The study's own findings, -0.078 for short-run own-price elasticity of electricity, align closely with prior research, confirming its place within established literature. The context provided by other studies helps to validate the paper's conclusions regarding the robustness of its econometric model and the effectiveness of conservation programs." + }, + { + "type": "image_caption", + "bbox": [ + 0.09, + 0.555, + 0.907, + 0.585 + ], + "angle": 0, + "content": "Figure 3. Visualization of an example from Doc-750K. The left side presents the interleaved text-image format data obtained through Document Content Extraction, while the right side showcases the annotations generated via Question-Answer Pairs Construction." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.599, + 0.483, + 0.645 + ], + "angle": 0, + "content": "example, \"\\n\\n \\n\" This format captures the document's textual content, making it easier to construct question-answer pairs." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.65, + 0.483, + 0.741 + ], + "angle": 0, + "content": "(2) Multi-Image Format. In this format, a document with \\( n \\) pages is rendered as \\( n \\) images, with each image corresponding to a single page. The structure follows the pattern \"\\n\\n\". This format preserves the original layout, enabling the model to learn the overall pagination and visual layout of the document." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.745, + 0.483, + 0.837 + ], + "angle": 0, + "content": "After processing the document into contexts in interleaved text-image and paginated image formats, we can not only use these contexts for next-token prediction training but also leverage the document's content, hierarchical structure, and layout features to flexibly and precisely generate high-quality question-answer pairs." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.841, + 0.483, + 0.902 + ], + "angle": 0, + "content": "Question-Answer Pairs Construction. In this step, we create question-and-answer pairs tailored to the source, content features, and formatting structure of each document. This process is divided into the following main categories:" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.6, + 0.905, + 0.645 + ], + "angle": 0, + "content": "(1) For documents with reliable QA annotations, like the review and reply in OpenReview, we extract the QA pairs and organize them into conversation format." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.65, + 0.906, + 0.755 + ], + "angle": 0, + "content": "(2) For documents with a clear textual structure, such as well-structured papers from Sci-Hub and Arxiv, we convert them to text and segment them, while the model is instructed to generate contents for each segment, including abstracts, experiment descriptions, and captions for figures and tables. The details of each task for structural papers are illustrated in Table 1." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.76, + 0.906, + 0.835 + ], + "angle": 0, + "content": "(3) For other documents, in addition to using them directly for NTP pretraining, we can input text interspersed with images into MLLMs to obtain QA pairs. To ensure high-quality generated data, we use the state-of-the-art model GPT-4o [56]." + }, + { + "type": "list", + "bbox": [ + 0.512, + 0.6, + 0.906, + 0.835 + ], + "angle": 0, + "content": null + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.841, + 0.906, + 0.902 + ], + "angle": 0, + "content": "Through our pipeline, most data has been processed into high-quality document-level question-answering data, while the remaining data is converted to plain text and used for next-token prediction tasks. Our pipelines are metricu" + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.516, + 0.957 + ], + "angle": 0, + "content": "4029" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.095, + 0.09, + 0.905, + 0.224 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.2, + 0.226, + 0.368, + 0.238 + ], + "angle": 0, + "content": "(a) Data Distribution of Doc-750K" + }, + { + "type": "image_caption", + "bbox": [ + 0.586, + 0.23, + 0.796, + 0.242 + ], + "angle": 0, + "content": "(b) Data Recipe for Supervised Fine-tuning" + }, + { + "type": "image_caption", + "bbox": [ + 0.089, + 0.256, + 0.908, + 0.3 + ], + "angle": 0, + "content": "Figure 4. Data distribution of our dataset. The outer circle shows the distribution of all data categories and the inner circle shows the distribution of data subsets. Left: Data distribution of Doc-750K. Right: Data distribution of our complete SFT training dataset. Note that the number reported in the figure represents the number of samples. “MT” is short for multi-turn." + }, + { + "type": "table", + "bbox": [ + 0.093, + 0.311, + 0.473, + 0.603 + ], + "angle": 0, + "content": "
TasksQuestions
Abstract WritingRead the full text of the paper and provide a concise summary in the form of an abstract.
Paper TitlingBased on the provided abstract or introduction of the research paper, please generate a concise and informative title
Caption WritingGive the relative texts of the images or tables, please write a caption for each image or table based on the relative texts provided.
Experiment WritingPlease write the "Experiments" section based on the incomplete research paper provided.
TranslationPlease read the full text of the following research paper and translate the Experiments section into Chinese.
" + }, + { + "type": "table_caption", + "bbox": [ + 0.09, + 0.613, + 0.483, + 0.671 + ], + "angle": 0, + "content": "Table 1. Questions format for different tasks. For documents with a clear textual structure, we design several proxy tasks. All tasks leverage the inherent structure of the documents, with answers directly sourced from the original text." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.686, + 0.484, + 0.763 + ], + "angle": 0, + "content": "lously designed to ensure high data quality across all generated context. Each LLM-generated sample is explicitly marked in the metadata as model-generated. Across the entire dataset, only \\(4.8\\%\\) of the data is LLM-generated, reinforcing the overall reliability and quality of the dataset." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.772, + 0.365, + 0.789 + ], + "angle": 0, + "content": "3.2. Multimodal Document Dataset" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.795, + 0.484, + 0.901 + ], + "angle": 0, + "content": "Data Source. The composition and distribution of our training data are detailed in Figure 4. Specifically, our dataset predominantly consists of academic papers, which constitute approximately \\(32.6\\%\\) of the total data. The multimodal data, carefully selected to augment our model's learning dimensions, makes up about \\(88.8\\%\\) of our dataset. This strategic distribution is designed to optimize the train-" + }, + { + "type": "table", + "bbox": [ + 0.597, + 0.311, + 0.825, + 0.453 + ], + "angle": 0, + "content": "
StatisticsNumber
Total Questions758K
Total Images3.1M
Total Conversations251K
Multi-Turn Questions87K
Single-Turn Questions164K
Average Text Tokens11245
Average Image Tokens6178
" + }, + { + "type": "table_caption", + "bbox": [ + 0.512, + 0.462, + 0.907, + 0.533 + ], + "angle": 0, + "content": "Table 2. Key statistics of the Doc-750K datasets. It comprises 758K questions, 3.1M images, and 251K conversations, including 87K multi-turn and 164K single-turn questions. With an average of 11,245 text tokens and 6,178 image tokens, it highlights the dataset's richness and diversity for multimodal research." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.549, + 0.906, + 0.579 + ], + "angle": 0, + "content": "ing process and improve the model's ability to generalize across different types of data inputs." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.58, + 0.907, + 0.761 + ], + "angle": 0, + "content": "Dataset Statistics. In our Doc-750K dataset, the majority of the data consists of reliably annotated entries, with OpenReview and Arxiv collectively accounting for \\(75.4\\%\\). The remaining data, sourced from Sci-Hub, is processed using our designed tasks. The overall distribution and number of tasks are shown in Figure 4(a). Our dataset ultimately consists of 251K conversations, comprising a total of 758K questions. Additional statistical details are provided in Table 2. Compared to previous datasets, Doc-750K contains a larger number of images, with an average of four images per conversation segment. Further comparisons with other datasets are shown in Table 3." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.772, + 0.859, + 0.789 + ], + "angle": 0, + "content": "3.3. Data Recipe for Supervised Fine-Tuning" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.795, + 0.906, + 0.885 + ], + "angle": 0, + "content": "Although Doc-750K effectively covers multimodal document QA scenarios, using it directly may lead to model over-fitting on a specific document domain. Therefore, we combine it with several open-source datasets to create a mixed dataset for SFT training. As shown in Figure 4(b), these datasets are organized into 4 categories as follows:" + }, + { + "type": "text", + "bbox": [ + 0.532, + 0.886, + 0.906, + 0.902 + ], + "angle": 0, + "content": "(1) For multi-page document QA, Doc-750K serves" + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.516, + 0.957 + ], + "angle": 0, + "content": "4030" + } + ], + [ + { + "type": "table", + "bbox": [ + 0.092, + 0.089, + 0.482, + 0.246 + ], + "angle": 0, + "content": "
Dataset#Images#QA Pairs#Tokens
Docmatix [33]2,444,7509,500,000390,000,000
DocVQA [15]10,18939,463337,829
TextCaps [64]21,95321,953389,658
TextVQA [65]21,95334,602181,918
ST-VQA [8]17,24723,121127,846
OCR-VQA [55]165,746801,5796,073,824
VisualMRC [70]3,02711,988168,828
DUDE [78]147,59723,71611,341,228
Doc-750K (ours)3,103,494758,0005,200,000,000
" + }, + { + "type": "table_caption", + "bbox": [ + 0.131, + 0.255, + 0.444, + 0.27 + ], + "angle": 0, + "content": "Table 3. Comparison with popular VQA datasets." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.287, + 0.483, + 0.378 + ], + "angle": 0, + "content": "as the core dataset, specifically curated to address complex, multi-page document comprehension. Additional datasets such as MP-Docmatix [33], MP-DocVQA [53], DUDE [78], and Taesiri-ArxivQA [69] offer valuable multi-page scenarios requiring inter-page reasoning and contextual retention across sequences." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.378, + 0.483, + 0.485 + ], + "angle": 0, + "content": "(2) For multi-image general QA, MMDU-45K [48] offers a comprehensive dataset encompassing diverse real-world scenarios, such as natural environments and everyday contexts. It emphasizes multi-turn dialogues and integration of multiple images, supporting the development of systems capable of generating coherent and accurate responses from complex, lengthy inputs." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.485, + 0.483, + 0.576 + ], + "angle": 0, + "content": "(3) For single-page document QA, We introduce DocVQA [53], DocReason [93], InfoVQA [54], and ChartQA [52] to further enhance the diversity of the SFT dataset. These datasets focus on individual pages with complex layouts, rich textual information, and, in some cases, graphical data interpretation." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.576, + 0.483, + 0.653 + ], + "angle": 0, + "content": "(4) For pure-text QA, we add datasets including LongAlpaca [11], LongAlpaca-16K-Length [11], LongQLoRA [91], LongCite [96], LongAlign [7], and LongReward [97] to support the assessment of the model's capabilities in QA tasks requiring long-range dependencies." + }, + { + "type": "list", + "bbox": [ + 0.09, + 0.378, + 0.483, + 0.653 + ], + "angle": 0, + "content": null + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.653, + 0.483, + 0.714 + ], + "angle": 0, + "content": "This expanded dataset provides a balanced foundation for training and evaluating multimodal document understanding models, enhancing robustness and adaptability across diverse document-related VQA tasks." + }, + { + "type": "title", + "bbox": [ + 0.09, + 0.729, + 0.483, + 0.765 + ], + "angle": 0, + "content": "4. Enhanced Baseline for Document-Level Multimodal Understanding" + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.772, + 0.279, + 0.788 + ], + "angle": 0, + "content": "4.1. Model Architecture" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.795, + 0.483, + 0.901 + ], + "angle": 0, + "content": "Our model architecture leverages the widely-adopted ViTMLP-LLM structure [41, 42, 73], consisting of a pre-trained Vision Transformer (ViT), a two-layer MLP projector, and a pre-trained Language Model (LLM). This combination provides a strong baseline for multimodal document analysis, effectively integrating visual and textual information within a unified framework." + }, + { + "type": "title", + "bbox": [ + 0.514, + 0.091, + 0.789, + 0.108 + ], + "angle": 0, + "content": "4.2. Optimizing Training Efficiency" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.121, + 0.907, + 0.243 + ], + "angle": 0, + "content": "The training efficiency of MLLMs is hindered by two key challenges: (1) Inconsistent Sample Lengths. Samples with different context lengths will result in excessive padding and lower training throughput; and (2) Limited GPU Memory. As the model scale and context length increase, GPU memory consumption becomes increasingly unsustainable. To address these issues, we have implemented the following strategies:" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.25, + 0.908, + 0.584 + ], + "angle": 0, + "content": "(1) Multimodal Data Packing. To balance the computational load between the vision model (ViT) and the language model (LLM) while minimizing resource waste caused by padding, we implement a multimodal data-packing strategy. The key idea is to concatenate multiple samples into long sequences to fully utilize the model's input capacity. Specifically, thresholds \\( T_{\\mathrm{img}} \\) and \\( T_{\\mathrm{tok}} \\) are set for the number of images and tokens, respectively. Samples are managed using a priority queue, sorted in descending order by the number of images and total tokens. A new sample \\( s \\) attempts to combine with the sample at the front of the priority queue. If the combination meets the thresholds (i.e., \\( T_{\\mathrm{img}}, T_{\\mathrm{tok}} \\)), the combined sample is pushed back into the priority queue. If \\( s \\) cannot match with any existing sample in the queue, it is directly added to the queue. When the image number and total token number of the front sample reach one of the thresholds or the number of samples exceeds the maximum limit \\( M \\), the front sample is dequeued, padded as needed, and sent for training. This strategy optimizes resource utilization and ensures balanced computational workloads. The detailed pseudo-code can be found in the supplementary materials." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.591, + 0.909, + 0.757 + ], + "angle": 0, + "content": "(2) Ring Attention. We implement the Ring Attention mechanism [43] to alleviate memory constraints associated with processing long sequences. By partitioning sequences into blocks and distributing computation across multiple devices, Ring Attention allows the model to accommodate larger contexts. This approach enables overlapping communication between key-value blocks and attention computations, thereby enhancing parallel processing efficiency. Consequently, Ring Attention improves the model's capacity to handle extended context lengths without exceeding memory limits." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.765, + 0.909, + 0.902 + ], + "angle": 0, + "content": "(3) Liger Kernel. To further improve memory and computational efficiency, we integrate the Liger Kernel [16], a specialized kernel library optimized for large-scale model training. The Liger Kernel enhances throughput and reduces memory consumption by employing techniques like kernel fusion, in-place operations, and input chunking. Leveraging the Liger Kernel thus enables higher training throughput and addresses memory limitations, allowing for efficient scaling of large multimodal models." + }, + { + "type": "list", + "bbox": [ + 0.512, + 0.25, + 0.909, + 0.902 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.514, + 0.957 + ], + "angle": 0, + "content": "4031" + } + ], + [ + { + "type": "table", + "bbox": [ + 0.092, + 0.089, + 0.912, + 0.354 + ], + "angle": 0, + "content": "
ModelsMP-DocMMLong-DocDocGenomeMM-NIAH
ANSL↑Acc↑F1↑Class Acc↑Title ED↓Abstract ED↓SP Acc↑MP Acc↑ShortMediumLongOverall
Proprietary Models
Gemini-1.5-Pro [60]-28.220.6-----73.865.260.867.1
GPT-4o [56]-42.844.997.69.56.571.867.6----
Open-Source Models
MiniMonkey-2B [28]70.310.38.657.416.555.040.328.940.926.923.531.0
InternVL2-2B [13]71.810.510.860.818.454.339.428.936.621.219.426.4
InternVL2-2B + RAG [85]72.617.216.760.818.454.339.428.436.830.234.833.8
Llama3.2-3B-Instruct†[21]-23.721.285.3194.751.040.234.915.52.20.56.6
Docopilot-2B (ours)76.221.816.056.24.543.645.137.458.046.740.949.2
MiniCPM-V2.6-8B [92]-16.915.492.810.232.660.054.249.015.30.023.4
LLaVA-OneVision-8B [36]-10.89.685.649.977.59.87.165.738.00.037.7
mPLUG-DocOwl2-8B [27]69.413.48.9-----17.90.10.06.6
M3DocRAG [14]84.421.022.6---------
VisRAG-8B [95]-18.818.392.810.232.660.050.747.129.229.535.8
InternLM2.5-7B-1M†[9]-28.725.692.777.659.342.742.540.537.235.137.8
InternVL2-8B [13]79.317.416.590.68.239.656.046.156.437.332.442.9
InternVL2-8B + RAG [85]78.724.224.590.68.239.656.046.055.743.445.248.4
InternVL2-26B [13]-15.515.487.516.923.349.742.765.048.741.952.8
Docopilot-8B (ours)81.328.823.093.82.019.753.951.971.257.455.361.8
" + }, + { + "type": "table_caption", + "bbox": [ + 0.09, + 0.363, + 0.908, + 0.421 + ], + "angle": 0, + "content": "Table 4. Evaluation on multi-page and interleaved VQA benchmarks. We report the metrics on MP-DocVQA [76] (MP-Doc), MMLongbench-Doc [51] (MMLong-Doc), DocGenome [89], and MM-NIAH [85]. Our model outperforms document-level MLLMs and multimodal RAG methods on multi-page, medium, and long-context QA. The \"Short\", \"Medium\", and \"Long\" in MM-NIAH refer to input length in \\([0,8\\mathrm{k}]\\), \\((8\\mathrm{k},32\\mathrm{k}]\\), \\((32\\mathrm{k},64\\mathrm{k}]\\), respectively. \"\\(\\dagger\\)” denotes input documents are parsed by OCR models." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.433, + 0.223, + 0.451 + ], + "angle": 0, + "content": "5. Experiments" + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.456, + 0.281, + 0.473 + ], + "angle": 0, + "content": "5.1. Experimental Setup" + }, + { + "type": "text", + "bbox": [ + 0.089, + 0.477, + 0.483, + 0.672 + ], + "angle": 0, + "content": "Training Details. Our model is available in two sizes: Docopilot-2B and Docopilot-8B, both of which are based on the InternVL2 [73] and fine-tuned for one epoch using the data recipe that includes Doc-750K. The training uses a batch size of 128, the AdamW optimizer with a learning rate of 1e-5, weight decay of 0.01 for the 2B variant, and 0.05 for the 8B variant, along with a cosine learning rate schedule. To speed up training, we apply multimodal data packing to reduce padding and a dynamic high-resolution strategy [13] to enhance OCR for document understanding. The maximum number of tiles for multimodal data is limited to 24, and the maximum sequence length is set to 32k tokens." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.675, + 0.483, + 0.841 + ], + "angle": 0, + "content": "Baselines. We compare our Docopilot with a series of open-source document-level MLLMs [27, 28, 36, 40, 46, 92, 98] that supports multi-image input and proprietary MLLMs, including Gemini-1.5-pro [60], GPT-4o [57]. For comparison with the commonly used RAG method for handling long documents, we selected the latest multimodal RAG methods VisRAG [95], InternVL + RAG [85], and M3DocRAG [14]. To compare with more long-context large language models, we use InternVL2-8B as the OCR model to extract texts from the documents and images and feed the parsed documents to long-context LLMs [9, 21]." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.848, + 0.258, + 0.865 + ], + "angle": 0, + "content": "5.2. Multi-Page VQA" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.871, + 0.483, + 0.903 + ], + "angle": 0, + "content": "Benchmarks. For the multi-page VQA task, we evaluate our model on three benchmarks: (1) MP-DocVQA [76]," + }, + { + "type": "table", + "bbox": [ + 0.516, + 0.432, + 0.905, + 0.605 + ], + "angle": 0, + "content": "
ModelsDocVQAChartQAInfoVQA
Gemini-1.5-Pro [60]93.187.281.0
GPT-4o [56]92.885.7-
MiniMonkey-2B [28]87.476.560.1
InternVL2-2B [13]86.976.258.9
Docopilot-2B (ours)87.376.458.5
Monkey-8B [40]66.565.136.1
TextMonkey-9B [46]73.066.928.6
mPLUG-DocOwl2-8B [27]80.770.046.4
IXC2.5-7B [98]90.982.269.9
InternVL2-8B [13]91.683.374.8
Docopilot-8B (ours)92.083.373.3
" + }, + { + "type": "table_caption", + "bbox": [ + 0.513, + 0.612, + 0.907, + 0.655 + ], + "angle": 0, + "content": "Table 5. Results on single-page VQA benchmarks. Our Docopilot models perform comparably to baselines [73], demonstrating enhanced long-context modeling without loss on shorter tasks." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.673, + 0.907, + 0.763 + ], + "angle": 0, + "content": "which is designed to evaluate the ability to handle complex questions across multiple scanned document pages. (2) MMLongbench-Doc [51], a benchmark for evaluating the performance of MLLMs on multi-modal documents. (3) DocGenome [89], a large-scale benchmark for the evaluation of scientific document comprehension." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.765, + 0.909, + 0.903 + ], + "angle": 0, + "content": "Results. As illustrated in Table 4, our model achieves consistent improvements on multi-page QA benchmarks, outperforming previous document-level MLLMs. Notably, our Docopilot-8B surpasses Gemini-1.5-Pro [60] on MMLongBench-Doc, positioning it as the closest open-source model to GPT-4o. In comparison to RAG-based methods [14, 86, 95], our model demonstrates advantages in multi-page scenarios. For example, in the Multi/Page QA of DocGenome benchmark, the RAG method shows a perfor" + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.516, + 0.957 + ], + "angle": 0, + "content": "4032" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.09, + 0.092, + 0.483, + 0.152 + ], + "angle": 0, + "content": "mance decline due to the disruption of document continuity while our Docopilot exhibits a significantly stable improvement compared to the baseline, with Docopilot-8B showing an increase of \\(12.6\\%\\) over InternVL2-8B." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.163, + 0.358, + 0.179 + ], + "angle": 0, + "content": "5.3. Interleaved Long-Context QA" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.185, + 0.483, + 0.23 + ], + "angle": 0, + "content": "Benchmarks. For the interleaved long-context QA task, we evaluate our models on MM-NIAH [86], a benchmark designed for long multimodal document comprehension." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.231, + 0.483, + 0.457 + ], + "angle": 0, + "content": "Results. The right side of Table 4 presents the results of MM-NIAH across context lengths ranging from 1K to 64K. We categorize the context lengths into \"Short,\" \"Medium,\" and \"Long\" based on the context window of InternVL2 (8K) and Docopilot (32K). Our Docopilot demonstrates exceptional performance in both medium- and long-context scenarios, while maintaining high accuracy in short-context situations. Notably, for QA tasks with context lengths in the range of \\((32K,64K]\\), Docopilot-2B outperforms InternVL2-2B by \\(110\\%\\), and Docopilot-8B surpasses InternVL2-8B by \\(70\\%\\). Furthermore, our model performs comparably to the state-of-the-art multimodal long-context model Gemini-1.5-Pro in contexts longer than 8K, establishing a new state-of-the-art performance among open-source long-context MLLMs." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.468, + 0.262, + 0.484 + ], + "angle": 0, + "content": "5.4. Single-Page VQA" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.49, + 0.483, + 0.595 + ], + "angle": 0, + "content": "**Benchmarks.** For single-page VQA tasks, we evaluate our model on three benchmarks: (1) DocVQA [53], a benchmark for the evaluation of extracting key information from an image of the given document. (2) ChartQA [52], a benchmark for evaluating the reasoning abilities for chart images. (3) InfoVQA [54], a benchmark for infographic image comprehension." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.597, + 0.483, + 0.716 + ], + "angle": 0, + "content": "Results. As shown in Table 5, our model achieves comparable performance to baseline models. Across the three benchmarks, Docopilot-2B and InternVL2-2B exhibit comparable results, while Docopilot-8B outperforms InternVL2-8B by 0.4 points in DocVQA. These results demonstrate that Doc-750K effectively enhances the model's long-context modeling capabilities without compromising its performance on shorter documents." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.728, + 0.245, + 0.743 + ], + "angle": 0, + "content": "5.5. Ablation Study" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.75, + 0.483, + 0.9 + ], + "angle": 0, + "content": "Effect of Doc-750K. We conducted ablation studies on MMLongBench-Doc [51] to analyze the impact of our Doc-750K. We divided Doc-750K into 3 parts according to the source of the data: (1) Sci-Hub data; (2) Arxiv data; and (3) OpenReview data. We demonstrate the effects of incorporating each part of the data into the SFT process, reported in Table 6. We observed that with the inclusion of different parts of Doc-750K, the model's performance improves continuously. Utilizing only open-source data results in an inferior F1 score." + }, + { + "type": "table", + "bbox": [ + 0.518, + 0.09, + 0.905, + 0.178 + ], + "angle": 0, + "content": "
ModelsAccF1
Baseline (InternVL2-2B [73])10.510.8
- Variant1: SFT using data recipe w/o Doc-750K18.49.4
- Variant2: Variant1 + our Sci-Hub data18.515.2
- Variant3: Variant2 + our Arxiv data20.515.5
Docopilot-2B: Variant3 + our OpenReview data21.816.0
" + }, + { + "type": "table_caption", + "bbox": [ + 0.513, + 0.188, + 0.907, + 0.245 + ], + "angle": 0, + "content": "Table 6. Ablation study on the data recipe. We evaluate the effectiveness of the training data from different sources on MMLongBench-Doc. Our Doc-750K can consistently enhance the ability of the model to understand multi-page documents." + }, + { + "type": "table", + "bbox": [ + 0.527, + 0.257, + 0.895, + 0.395 + ], + "angle": 0, + "content": "
ModelsLatencyAccF1
MiniCPM-V2.6 [92]225.4ms16.915.4
VisRAG-12B [95]288.3ms18.818.3
InternVL2-2B [13]35.9ms10.510.8
InternVL2-2B + RAG [85]82.9ms17.216.7
Docopilot-2B (ours)35.9ms21.816.0
InternVL2-8B [13]81.0ms17.416.5
InternVL2-8B + RAG [85]113.4ms24.224.5
Docopilot-8B (ours)81.0ms28.823.0
" + }, + { + "type": "table_caption", + "bbox": [ + 0.513, + 0.405, + 0.907, + 0.475 + ], + "angle": 0, + "content": "Table 7. Latency analysis. We evaluate the average token output latency of model outputs on MMLongBench-Doc [51]. RAG-based methods [85, 95] exhibit slower processing speeds due to their two-stage inference process, making them less efficient than document MLLMs for handling multimodal long documents." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.5, + 0.907, + 0.652 + ], + "angle": 0, + "content": "Latency Analysis. To compare the latency in inference between RAG methods and our Docopilot, we conducted a latency analysis on MMLongBench-Doc [51], as reported in Table 7. While RAG reduces the document length input to the MLLM, its own time cost remains non-negligible. For instance, InternVL2-2B + RAG is \\(130\\%\\) slower than InternVL2-2B, and VisRAG is \\(28\\%\\) slower than MiniCPMV2.6. Our Docopilot does not require additional processes and therefore has the same inference time as baseline models, making it more suitable for analyzing long documents." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.665, + 0.642, + 0.681 + ], + "angle": 0, + "content": "6. Conclusions" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.689, + 0.907, + 0.901 + ], + "angle": 0, + "content": "This work introduced a diverse document-level question-answering dataset that covers complex structures and crosspage dependencies, providing a robust foundation for training and evaluating document understanding models. We also proposed a retrieval-free long-document understanding model that effectively integrates multi-page information, reducing reliance on external retrieval systems. Experimental results show that our model achieves state-of-the-art performance across several document-level QA benchmarks, underscoring its strength in multi-page integration and complex reasoning. Future work will focus on improving computational efficiency, extending the model to larger multimodal tasks, and adapting it to broader applications for enhanced practicality and generalization." + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.516, + 0.957 + ], + "angle": 0, + "content": "4033" + } + ], + [ + { + "type": "title", + "bbox": [ + 0.092, + 0.091, + 0.251, + 0.108 + ], + "angle": 0, + "content": "Acknowledgments" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.115, + 0.484, + 0.206 + ], + "angle": 0, + "content": "This project was supported by the National Key R&D Program of China (No. 2022ZD0161300, 2022ZD0160101), the National Natural Science Foundation of China (No. 62376134, 62372223). Zhe Chen is supported by the Youth PhD Student Research Project under the National Natural Science Foundation (No. 623B2050)." + }, + { + "type": "title", + "bbox": [ + 0.093, + 0.219, + 0.188, + 0.235 + ], + "angle": 0, + "content": "References" + }, + { + "type": "ref_text", + "bbox": [ + 0.101, + 0.244, + 0.484, + 0.312 + ], + "angle": 0, + "content": "[1] Jean-Baptiste Alayrac, Jeff Donahue, Pauline Luc, Antoine Miech, Iain Barr, Yana Hasson, Karel Lenc, Arthur Mensch, Katherine Millican, Malcolm Reynolds, et al. Flamingo: a visual language model for few-shot learning. NeurIPS, 35: 23716-23736, 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.1, + 0.314, + 0.484, + 0.37 + ], + "angle": 0, + "content": "[2] Chenxin An, Shansan Gong, Ming Zhong, Xingjian Zhao, Mukai Li, Jun Zhang, Lingpeng Kong, and Xipeng Qiu. L-eval: Instituting standardized evaluation for long context language models. arXiv preprint arXiv:2307.11088, 2023. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.101, + 0.371, + 0.483, + 0.437 + ], + "angle": 0, + "content": "[3] Srikar Appalaraju, Peng Tang, Qi Dong, Nishant Sankaran, Yichu Zhou, and R Manmatha. Docformerv2: Local features for document understanding. In Proceedings of the AAAI Conference on Artificial Intelligence, pages 709-718, 2024." + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.44, + 0.484, + 0.509 + ], + "angle": 0, + "content": "[4] Haoli Bai, Zhiguang Liu, Xiaojun Meng, Wentao Li, Shuang Liu, Nian Xie, Rongfu Zheng, Liangwei Wang, Lu Hou, Jiansheng Wei, et al. Wukong-reader: Multi-modal pretraining for fine-grained visual document understanding. arXiv preprint arXiv:2212.09621, 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.511, + 0.484, + 0.676 + ], + "angle": 0, + "content": "[5] Jinze Bai, Shuai Bai, Yunfei Chu, Zeyu Cui, Kai Dang, Xiaodong Deng, Yang Fan, Wenbin Ge, Yu Han, Fei Huang, Binyuan Hui, Luo Ji, Mei Li, Junyang Lin, Runji Lin, Dayiheng Liu, Gao Liu, Chengqiang Lu, Keming Lu, Jianxin Ma, Rui Men, Xingzhang Ren, Xuancheng Ren, Chuanqi Tan, Sinan Tan, Jianhong Tu, Peng Wang, Shijie Wang, Wei Wang, Shengguang Wu, Benfeng Xu, Jin Xu, An Yang, Hao Yang, Jian Yang, Shusheng Yang, Yang Yao, Bowen Yu, Hongyi Yuan, Zheng Yuan, Jianwei Zhang, Xingxuan Zhang, Yichang Zhang, Zhenru Zhang, Chang Zhou, Jingren Zhou, Xiaohuan Zhou, and Tianhang Zhu. Qwen technical report. arXiv preprint arXiv:2309.16609, 2023. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.677, + 0.483, + 0.732 + ], + "angle": 0, + "content": "[6] Jinze Bai, Shuai Bai, Shusheng Yang, Shijie Wang, Sinan Tan, Peng Wang, Junyang Lin, Chang Zhou, and Jingren Zhou. Qwen-vl: A frontier large vision-language model with versatile abilities. arXiv preprint arXiv:2308.12966, 2023. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.734, + 0.483, + 0.789 + ], + "angle": 0, + "content": "[7] Yushi Bai, Xin Lv, Jiajie Zhang, Yuze He, Ji Qi, Lei Hou, Jie Tang, Yuxiao Dong, and Juanzi Li. Longalign: A recipe for long context alignment of large language models. arXiv preprint arXiv:2401.18058, 2024. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.79, + 0.483, + 0.844 + ], + "angle": 0, + "content": "[8] Ali Furkan Biten, Ruben Tito, Andres Mafla, Lluis Gomez, Marcal Rusinol, Ernest Valveny, CV Jawahar, and Dimosthenis Karatzas. Scene text visual question answering. In ICCV, pages 4291-4301, 2019. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.846, + 0.483, + 0.9 + ], + "angle": 0, + "content": "[9] Zheng Cai, Maosong Cao, Haojiong Chen, Kai Chen, Keyu Chen, Xin Chen, Xun Chen, Zehui Chen, Zhi Chen, Pei Chu, et al. Internl m2 technical report. arXiv preprint arXiv:2403.17297, 2024. 1, 7" + }, + { + "type": "list", + "bbox": [ + 0.1, + 0.244, + 0.484, + 0.9 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.093, + 0.906, + 0.148 + ], + "angle": 0, + "content": "[10] Xinlei Chen, Hao Fang, Tsung-Yi Lin, Ramakrishna Vedantam, Saurabh Gupta, Piotr Dólar, and C Lawrence Zitnick. Microsoft coco captions: Data collection and evaluation server. arXiv preprint arXiv:1504.00325, 2015. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.15, + 0.906, + 0.204 + ], + "angle": 0, + "content": "[11] Yukang Chen, Shengju Qian, Haotian Tang, Xin Lai, Zhi-jian Liu, Song Han, and Jiaya Jia. Longlora: Efficient finetuning of long-context large language models. arXiv preprint arXiv:2309.12307, 2023. 2, 3, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.207, + 0.906, + 0.276 + ], + "angle": 0, + "content": "[12] Zhe Chen, Jiannan Wu, Wenhai Wang, Weijie Su, Guo Chen, Sen Xing, Zhong Muyan, Qinglong Zhang, Xizhou Zhu, Lewei Lu, et al. Internvl: Scaling up vision foundation models and aligning for generic visual-linguistic tasks. arXiv preprint arXiv:2312.14238, 2023. 1, 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.278, + 0.906, + 0.346 + ], + "angle": 0, + "content": "[13] Zhe Chen, Weiyun Wang, Hao Tian, Shenglong Ye, Zhangwei Gao, Erfei Cui, Wenwen Tong, Kongzhi Hu, Jiapeng Luo, Zheng Ma, et al. How far are we to gpt-4v? closing the gap to commercial multimodal models with open-source suites. arXiv preprint arXiv:2404.16821, 2024. 1, 2, 7, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.349, + 0.906, + 0.403 + ], + "angle": 0, + "content": "[14] Jaemin Cho, Debanjan Mahata, Ozan Irsoy, Yujie He, and Mohit Bansal. M3docrag: Multi-modal retrieval is what you need for multi-page multi-document understanding. arXiv preprint arXiv:2411.04952, 2024. 1, 2, 3, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.405, + 0.906, + 0.445 + ], + "angle": 0, + "content": "[15] Christopher Clark and Matt Gardner. Simple and effective multi-paragraph reading comprehension. In ACL, pages 845–855, 2018. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.448, + 0.906, + 0.516 + ], + "angle": 0, + "content": "[16] Yun Dai, Vignesh Kothapalli, Qingquan Song, Shao Tang, Siyu Zhu, Steven Shimizu, Shivam Sahni, Haowen Ning, Yanning Chen, et al. Liger kernel: Efficient triton kernels for llm training. arXiv preprint arXiv:2410.10989, 2024. 2, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.519, + 0.906, + 0.559 + ], + "angle": 0, + "content": "[17] Tri Dao. Flashattention-2: Faster attention with better parallelism and work partitioning. arXiv preprint arXiv:2307.08691, 2023. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.562, + 0.906, + 0.614 + ], + "angle": 0, + "content": "[18] Tri Dao, Dan Fu, Stefano Ermon, Atri Rudra, and Christopher Ré. Flashattention: Fast and memory-efficient exact attention with io-awareness. NeurIPS, 35:16344-16359, 2022. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.618, + 0.906, + 0.673 + ], + "angle": 0, + "content": "[19] Jiayu Ding, Shuming Ma, Li Dong, Xingxing Zhang, Shaohan Huang, Wenhui Wang, Nanning Zheng, and Furu Wei. Longnet: Scaling transformers to 1,000,000,000 tokens. arXiv preprint arXiv:2307.02486, 2023. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.676, + 0.906, + 0.731 + ], + "angle": 0, + "content": "[20] Runpei Dong, Chunrui Han, Yuang Peng, Zekun Qi, Zheng Ge, Jinrong Yang, Liang Zhao, Jianjian Sun, Hongyu Zhou, Haoran Wei, et al. Dreamllm: Synergistic multimodal comprehension and creation. In ICLR, 2024. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.733, + 0.906, + 0.8 + ], + "angle": 0, + "content": "[21] Abhimanyu Dubey, Abhinav Jauhri, Abhinav Pandey, Abhishek Kadian, Ahmad Al-Dahle, Aiesha Letman, Akhil Mathur, Alan Schelten, Amy Yang, Angela Fan, et al. The llama 3 herd of models. arXiv preprint arXiv:2407.21783, 2024. 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.803, + 0.906, + 0.857 + ], + "angle": 0, + "content": "[22] Manuel Faysse, Hugues Sibille, Tony Wu, Gautier Viaud, Céline Hudelot, and Pierre Colombo. Colpali: Efficient document retrieval with vision language models. arXiv preprint arXiv:2407.01449, 2024. 1, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.86, + 0.906, + 0.901 + ], + "angle": 0, + "content": "[23] Hao Feng, Qi Liu, Hao Liu, Wengang Zhou, Houqiang Li, and Can Huang. Docpedia: Unleashing the power of large multimodal model in the frequency domain for versatile" + }, + { + "type": "list", + "bbox": [ + 0.517, + 0.093, + 0.906, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.516, + 0.956 + ], + "angle": 0, + "content": "4034" + } + ], + [ + { + "type": "ref_text", + "bbox": [ + 0.125, + 0.092, + 0.482, + 0.119 + ], + "angle": 0, + "content": "document understanding. arXiv preprint arXiv:2311.11810, 2023.3" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.121, + 0.483, + 0.189 + ], + "angle": 0, + "content": "[24] Chaoyou Fu, Peixian Chen, Yunhang Shen, Yulei Qin, Mengdan Zhang, Xu Lin, Zhenyu Qiu, Wei Lin, Jinrui Yang, Xiawu Zheng, et al. Mme: A comprehensive evaluation benchmark for multimodal large language models. arXiv preprint arXiv:2306.13394, 2023. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.19, + 0.482, + 0.259 + ], + "angle": 0, + "content": "[25] Zhangwei Gao, Zhe Chen, Erfei Cui, Yiming Ren, Weiyun Wang, Jinguo Zhu, Hao Tian, Shenglong Ye, Junjun He, Xizhou Zhu, et al. Mini-internvl: A flexible-transfer pocket multimodal model with \\(5\\%\\) parameters and \\(90\\%\\) performance. arXiv preprint arXiv:2410.16261, 2024. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.261, + 0.482, + 0.314 + ], + "angle": 0, + "content": "[26] Chi Han, Qifan Wang, Wenhan Xiong, Yu Chen, Heng Ji, and Sinong Wang. Lm-infinite: Simple on-the-fly length generalization for large language models. arXiv preprint arXiv:2308.16137, 2023. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.316, + 0.482, + 0.383 + ], + "angle": 0, + "content": "[27] Anwen Hu, Haiyang Xu, Liang Zhang, Jiabo Ye, Ming Yan, Ji Zhang, Qin Jin, Fei Huang, and Jingren Zhou. mplug-docowl2: High-resolution compressing for ocr-free multi-page document understanding. arXiv preprint arXiv:2409.03420, 2024. 7, 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.385, + 0.482, + 0.439 + ], + "angle": 0, + "content": "[28] Mingxin Huang, Yuliang Liu, Dingkang Liang, Lianwen Jin, and Xiang Bai. Mini-monkey: Alleviate the saw-tooth effect by multi-scale adaptive cropping. arXiv preprint arXiv:2408.02034, 2024. 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.442, + 0.482, + 0.536 + ], + "angle": 0, + "content": "[29] Gabriel Ilharco, Mitchell Wortsman, Ross Wightman, Cade Gordon, Nicholas Carlini, Rohan Taori, Achal Dave, Vaishaal Shankar, Hongseok Namkoong, John Miller, Hannaneh Hajishirzi, Ali Farhadi, and Ludwig Schmidt. Openclip. Zenodo. Version 0.1. https://doi.org/10.5281/zenodo.5143773, 2021. DOI: 10.5281/zenodo.5143773. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.539, + 0.482, + 0.593 + ], + "angle": 0, + "content": "[30] Yao Jiang, Xinyu Yan, Ge-Peng Ji, Keren Fu, Meijun Sun, Huan Xiong, Deng-Ping Fan, and Fahad Shahbaz Khan. Effectiveness assessment of recent large vision-language models. Visual Intelligence, 2(1):17, 2024. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.595, + 0.482, + 0.636 + ], + "angle": 0, + "content": "[31] Greg Kamradt. Llmtest_needleinahaystack. https://github.com/gkamradt/LLMTest_NeedleInAHaystack, 2024. Accessed: 2024-11-11. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.637, + 0.482, + 0.718 + ], + "angle": 0, + "content": "[32] Geewook Kim, Teakgyu Hong, Moonbin Yim, JeongYeon Nam, Jinyoung Park, Jinyeong Yim, Wonseok Hwang, Sangdoo Yun, Dongyoon Han, and Seunghyun Park. Ocr-free document understanding transformer. In European Conference on Computer Vision, pages 498-517. Springer, 2022. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.72, + 0.482, + 0.774 + ], + "angle": 0, + "content": "[33] Hugo Laurencon, Andres Marafioti, Victor Sanh, and Léo Tronchon. Building and better understanding vision-language models: insights and future directions. arXiv preprint arXiv:2408.12637, 2024. 2, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.776, + 0.482, + 0.844 + ], + "angle": 0, + "content": "[34] Hugo Laurençon, Lucile Saulnier, Léo Tronchon, Stas Bekman, Amanpreet Singh, Anton Lozhkov, Thomas Wang, Siddharth Karamcheti, Alexander Rush, Douwe Kiela, et al. Obelics: An open web-scale filtered dataset of interleaved image-text documents. NIPS, 36, 2024. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.846, + 0.482, + 0.899 + ], + "angle": 0, + "content": "[35] Bohao Li, Rui Wang, Guangzhi Wang, Yuying Ge, Yixiao Ge, and Ying Shan. Seed-bench: Benchmarking multimodal llms with generative comprehension. arXiv preprint arXiv:2307.16125, 2023. 2" + }, + { + "type": "list", + "bbox": [ + 0.093, + 0.092, + 0.483, + 0.899 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.093, + 0.905, + 0.147 + ], + "angle": 0, + "content": "[36] Bo Li, Yuanhan Zhang, Dong Guo, Renrui Zhang, Feng Li, Hao Zhang, Kaichen Zhang, Yanwei Li, Ziwei Liu, and Chunyuan Li. Llava-onevision: Easy visual task transfer. arXiv preprint arXiv:2408.03326, 2024. 1, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.149, + 0.905, + 0.204 + ], + "angle": 0, + "content": "[37] Junnan Li, Dongxu Li, Caiming Xiong, and Steven Hoi. Blip: Bootstrapping language-image pre-training for unified vision-language understanding and generation. In ICML, pages 12888-12900, 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.206, + 0.905, + 0.261 + ], + "angle": 0, + "content": "[38] Junnan Li, Dongxu Li, Silvio Savarese, and Steven Hoi. Blip-2: Bootstrapping language-image pre-training with frozen image encoders and large language models. In ICML, pages 19730–19742. PMLR, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.263, + 0.905, + 0.331 + ], + "angle": 0, + "content": "[39] Qingyun Li, Zhe Chen, Weiyun Wang, Wenhai Wang, Shenglong Ye, Zhenjiang Jin, Guanzhou Chen, Yinan He, Zhangwei Gao, Erfei Cui, et al. Omnicorpus: An unified multimodal corpus of 10 billion-level images interleaved with text. arXiv preprint arXiv:2406.08418, 2024. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.334, + 0.905, + 0.401 + ], + "angle": 0, + "content": "[40] Zhang Li, Biao Yang, Qiang Liu, Zhiyin Ma, Shuo Zhang, Jingxu Yang, Yabo Sun, Yuliang Liu, and Xiang Bai. Monkey: Image resolution and text label are important things for large multi-modal models. arXiv preprint arXiv:2311.06607, 2023. 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.404, + 0.905, + 0.446 + ], + "angle": 0, + "content": "[41] Haotian Liu, Chunyuan Li, Yuheng Li, and Yong Jae Lee. Improved baselines with visual instruction tuning. arXiv preprint arXiv:2310.03744, 2023. 1, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.448, + 0.904, + 0.475 + ], + "angle": 0, + "content": "[42] Haotian Liu, Chunyuan Li, Qingyang Wu, and Yong Jae Lee. Visual instruction tuning. NeurIPS, 36, 2023. 1, 2, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.477, + 0.905, + 0.518 + ], + "angle": 0, + "content": "[43] Hao Liu, Matei Zaharia, and Pieter Abbeel. Ring attention with blockwise transformers for near-infinite context. arXiv preprint arXiv:2310.01889, 2023. 2, 3, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.52, + 0.905, + 0.585 + ], + "angle": 0, + "content": "[44] Yuan Liu, Haodong Duan, Yuanhan Zhang, Bo Li, Songyang Zhang, Wangbo Zhao, Yike Yuan, Jiaqi Wang, Conghui He, Ziwei Liu, et al. Mmbench: Is your multi-modal model an all-around player? arXiv preprint arXiv:2307.06281, 2023. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.59, + 0.905, + 0.657 + ], + "angle": 0, + "content": "[45] Yuliang Liu, Zhang Li, Hongliang Li, Wenwen Yu, Mingxin Huang, Dezhi Peng, Mingyu Liu, Mingrui Chen, Chunyuan Li, Lianwen Jin, et al. On the hidden mystery ofOCR in large multimodal models. arXiv preprint arXiv:2305.07895, 2023. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.661, + 0.905, + 0.716 + ], + "angle": 0, + "content": "[46] Yuliang Liu, Biao Yang, Qiang Liu, Zhang Li, Zhiyin Ma, Shuo Zhang, and Xiang Bai. Textmonkey: AnOCR-free large multimodal model for understanding document. arXiv preprint arXiv:2403.04473, 2024. 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.718, + 0.905, + 0.787 + ], + "angle": 0, + "content": "[47] Zhaoyang Liu, Yinan He, Wenhai Wang, Weiyun Wang, Yi Wang, Shoufa Chen, Qinglong Zhang, Zeqiang Lai, Yang Yang, Qingyun Li, Jiashuo Yu, et al. Interrupt: Solving vision-centric tasks by interacting with chatgpt beyond language. arXiv preprint arXiv:2305.05662, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.789, + 0.905, + 0.857 + ], + "angle": 0, + "content": "[48] Ziyu Liu, Tao Chu, Yuhang Zang, Xilin Wei, Xiaoyi Dong, Pan Zhang, Zijian Liang, Yuanjun Xiong, Yu Qiao, Dahua Lin, et al. Mmdu: A multi-turn multi-image dialog understanding benchmark and instruction-tuning dataset for lvlms. arXiv preprint arXiv:2406.11833, 2024. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.859, + 0.905, + 0.901 + ], + "angle": 0, + "content": "[49] Chuwei Luo, Yufan Shen, Zhaoqing Zhu, Qi Zheng, Zhi Yu, and Cong Yao. Layout: Layout instruction tuning with large language models for document understanding. In Pro" + }, + { + "type": "list", + "bbox": [ + 0.517, + 0.093, + 0.905, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.516, + 0.956 + ], + "angle": 0, + "content": "4035" + } + ], + [ + { + "type": "ref_text", + "bbox": [ + 0.126, + 0.093, + 0.482, + 0.12 + ], + "angle": 0, + "content": "ceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 15630-15640, 2024. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.123, + 0.483, + 0.178 + ], + "angle": 0, + "content": "[50] Xueguang Ma, Sheng-Chieh Lin, Minghan Li, Wenhu Chen, and Jimmy Lin. Unifying multimodal retrieval via document screenshot embedding. arXiv preprint arXiv:2406.11251, 2024. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.181, + 0.482, + 0.263 + ], + "angle": 0, + "content": "[51] Yubo Ma, Yuhang Zang, Liangyu Chen, Meiqi Chen, Yizhu Jiao, Xinze Li, Xinyuan Lu, Ziyu Liu, Yan Ma, Xiaoyi Dong, Pan Zhang, Liangming Pan, Yu-Gang Jiang, Jiaqi Wang, Yixin Cao, and Aixin Sun. Mmlongbench-doc: Benchmarking long-context document understanding with visualizations, 2024. 1, 7, 8, 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.266, + 0.482, + 0.322 + ], + "angle": 0, + "content": "[52] Ahmed Masry, Xuan Long Do, Jia Qing Tan, Shafiq Joty, and Enamul Hoque. Chartqa: A benchmark for question answering about charts with visual and logical reasoning. In ACL, pages 2263-2279, 2022. 6, 8, 1, 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.324, + 0.482, + 0.365 + ], + "angle": 0, + "content": "[53] Minesh Mathew, Dimosthenis Karatzas, and CV Jawahar. Docvqa: A dataset for vqa on document images. In WACV, pages 2200-2209, 2021. 6, 8, 1, 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.368, + 0.482, + 0.409 + ], + "angle": 0, + "content": "[54] Minesh Mathew, Viraj Bagal, Ruben Tito, Dimosthenis Karatzas, Ernest Valveny, and CV Jawahar. Infographicvqa. In WACV, pages 1697-1706, 2022. 6, 8, 1, 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.412, + 0.482, + 0.466 + ], + "angle": 0, + "content": "[55] Anand Mishra, Shashank Shekhar, Ajeet Kumar Singh, and Anirban Chakraborty. Ocr-vqa: Visual question answering by reading text in images. In ICDAR, pages 947-952, 2019. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.469, + 0.482, + 0.51 + ], + "angle": 0, + "content": "[56] OpenAI. Gpt-4v(ison) system card. https://cdn.openai.com/papers/GPTV_System/Card.pdf, 2023.1,4,7" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.513, + 0.482, + 0.542 + ], + "angle": 0, + "content": "[57] OpenAI. Gpt-4o system card. https://openai.com/index/gpt-4o-system-card/, 2024.7" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.544, + 0.482, + 0.585 + ], + "angle": 0, + "content": "[58] Ofir Press, Noah A Smith, and Mike Lewis. Train short, test long: Attention with linear biases enables input length extrapolation. arXiv preprint arXiv:2108.12409, 2021. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.587, + 0.482, + 0.628 + ], + "angle": 0, + "content": "[59] Sylvestre-Alvise Rebuffi, Hakan Bilen, and Andrea Vedaldi. Learning multiple visual domains with residual adapters. NeurIPS, 30, 2017. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.631, + 0.482, + 0.713 + ], + "angle": 0, + "content": "[60] Machel Reid, Nikolay Savinov, Denis Teptyashin, Dmitry Lepikhin, Timothy Lillicrap, Jean-baptiste Alayrac, Radu Soricut, Angeliki Lazaridou, Orhan Firat, Julian Schrittwieser, et al. Gemini 1.5: Unlocking multimodal understanding across millions of tokens of context. arXiv preprint arXiv:2403.05530, 2024. 1, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.716, + 0.482, + 0.771 + ], + "angle": 0, + "content": "[61] Sahel Sharifymoghaddam, Shivani Upadhyay, Wenhu Chen, and Jimmy Lin. Unirag: Universal retrieval augmentation for multi-modal large language models. arXiv preprint arXiv:2405.10311, 2024. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.774, + 0.482, + 0.843 + ], + "angle": 0, + "content": "[62] Min Shi, Fuxiao Liu, Shihao Wang, Shijia Liao, Subhashree Radhakrishnan, De-An Huang, Hongxu Yin, Karan Sapra, Yaser Yacoob, Humphrey Shi, et al. Eagle: Exploring the design space for multimodal llms with mixture of encoders. arXiv preprint arXiv:2408.15998, 2024. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.846, + 0.482, + 0.901 + ], + "angle": 0, + "content": "[63] Weijia Shi, Sewon Min, Michihiro Yasunaga, Minjoon Seo, Rich James, Mike Lewis, Luke Zettlemoyer, and Wen-tau Yih. Replug: Retrieval-augmented black-box language models. arXiv preprint arXiv:2301.12652, 2023. 1" + }, + { + "type": "list", + "bbox": [ + 0.093, + 0.093, + 0.483, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.093, + 0.905, + 0.147 + ], + "angle": 0, + "content": "[64] Oleksii Sidorov, Ronghang Hu, Marcus Rohrbach, and Amanpreet Singh. Textcaps: a dataset for image captioning with reading comprehension. In ECCV, pages 742-758, 2020. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.15, + 0.905, + 0.205 + ], + "angle": 0, + "content": "[65] Amanpreet Singh, Vivek Natarajan, Meet Shah, Yu Jiang, Xinlei Chen, Dhruv Batra, Devi Parikh, and Marcus Rohrbach. Towards vqa models that can read. In CVPR, pages 8317-8326, 2019. 1, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.207, + 0.905, + 0.26 + ], + "angle": 0, + "content": "[66] Jianlin Su, Murtadha Ahmed, Yu Lu, Shengfeng Pan, Wen Bo, and Yunfeng Liu. Roformer: Enhanced transformer with rotary position embedding. Neurocomputing, 568:127063, 2024. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.263, + 0.905, + 0.318 + ], + "angle": 0, + "content": "[67] Quan Sun, Qiying Yu, Yufeng Cui, Fan Zhang, Xiaosong Zhang, Yueze Wang, Hongcheng Gao, Jingjing Liu, Tiejun Huang, and Xinlong Wang. Generative pretraining in multimodality. In ICLR, 2024. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.32, + 0.905, + 0.375 + ], + "angle": 0, + "content": "[68] Yutao Sun, Li Dong, Barun Patra, Shuming Ma, Shaohan Huang, Alon Benhaim, Vishrav Chaudhary, Xia Song, and Furu Wei. A length-extrapolatable transformer. arXiv preprint arXiv:2212.10554, 2022. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.377, + 0.904, + 0.404 + ], + "angle": 0, + "content": "[69] Mohammad Reza Taesiri. Arxivqa. https://github.com/taesiri/ArXivQA, 2024.6" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.406, + 0.905, + 0.462 + ], + "angle": 0, + "content": "[70] Ryota Tanaka, Kyosuke Nishida, and Sen Yoshida. Visualmrc: Machine reading comprehension on document images. In Proceedings of the AAAI Conference on Artificial Intelligence, pages 13878-13888, 2021. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.463, + 0.905, + 0.544 + ], + "angle": 0, + "content": "[71] Zineng Tang, Ziyi Yang, Guoxin Wang, Yuwei Fang, Yang Liu, Chenguang Zhu, Michael Zeng, Cha Zhang, and Mohit Bansal. Unifying vision, text, and layout for universal document processing. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 19254-19264, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.547, + 0.905, + 0.615 + ], + "angle": 0, + "content": "[72] Gemini Team, Rohan Anil, Sebastian Borgeaud, Yonghui Wu, Jean-Baptiste Alayrac, Jiahui Yu, Radu Soricut, Johan Schalkwyk, Andrew M Dai, Anja Hauth, et al. Gemini: a family of highly capable multimodal models. arXiv preprint arXiv:2312.11805, 2023. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.617, + 0.905, + 0.672 + ], + "angle": 0, + "content": "[73] OpenGVLab Team. Internv12: Better than the best—expanding performance boundaries of open-source multimodal models with the progressive scaling strategy, 2024. 1, 6, 7, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.674, + 0.905, + 0.701 + ], + "angle": 0, + "content": "[74] Qwen Team. Qwen2.5: A party of foundation models, 2024. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.704, + 0.905, + 0.771 + ], + "angle": 0, + "content": "[75] Changyao Tian, Xizhou Zhu, Yuwen Xiong, Weiyun Wang, Zhe Chen, Wenhai Wang, Yuntao Chen, Lewei Lu, Tong Lu, Jie Zhou, et al. Mm-interleaved: Interleaved image-text generative modeling via multi-modal feature synchronizer. arXiv preprint arXiv:2401.10208, 2024. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.774, + 0.905, + 0.814 + ], + "angle": 0, + "content": "[76] Rubén Tito, Dimosthenis Karatzas, and Ernest Valveny. Hierarchical multimodal transformers for multipage docvqa. Pattern Recognition, 144:109834, 2023. 1, 2, 7, 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.817, + 0.905, + 0.858 + ], + "angle": 0, + "content": "[77] Xiaoguang Tu, Zhi He, Yi Huang, Zhi-Hao Zhang, Ming Yang, and Jian Zhao. An overview of large ai models and their applications. Visual Intelligence, 2(1):1-22, 2024. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.86, + 0.905, + 0.901 + ], + "angle": 0, + "content": "[78] Jordy Van Landeghem, Ruben Tito, Łukasz Borchmann, Michal Pietruszka, Paweł Joziak, Rafał Powalski, Dawid Jurkiewicz, Mickael Coustaty, Bertrand Ackaert, Ernest Val" + }, + { + "type": "list", + "bbox": [ + 0.517, + 0.093, + 0.905, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.517, + 0.956 + ], + "angle": 0, + "content": "4036" + } + ], + [ + { + "type": "ref_text", + "bbox": [ + 0.125, + 0.092, + 0.484, + 0.147 + ], + "angle": 0, + "content": "veny, et al. Document understanding dataset and evaluation (dude). In Proceedings IEEE/CVF international conference on computer vision-ICCV 2023, pages 19528-19540. IEEE/CVF, 2023. 2, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.149, + 0.484, + 0.219 + ], + "angle": 0, + "content": "[79] Bin Wang, Chao Xu, Xiaomeng Zhao, Linke Ouyang, Fan Wu, Zhiyuan Zhao, Rui Xu, Kaiwen Liu, Yuan Qu, Fukai Shang, et al. Mineru: An open-source solution for precise document content extraction. arXiv preprint arXiv:2409.18839, 2024. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.221, + 0.483, + 0.29 + ], + "angle": 0, + "content": "[80] Dongsheng Wang, Natraj Raman, Mathieu Sibue, Zhiqiang Ma, Petr Babkin, Simerjot Kaur, Yulong Pei, Armineh Nourbakhsh, and Xiaomo Liu. Docllm: A layout-aware generative language model for multimodal document understanding. arXiv preprint arXiv:2401.00908, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.291, + 0.483, + 0.36 + ], + "angle": 0, + "content": "[81] Peng Wang, Shuai Bai, Sinan Tan, Shijie Wang, Zhihao Fan, Jinze Bai, Keqin Chen, Xuejing Liu, Jialin Wang, Wenbin Ge, et al. Qwen2-vl: Enhancing vision-language model's perception of the world at any resolution. arXiv preprint arXiv:2409.12191, 2024. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.362, + 0.483, + 0.417 + ], + "angle": 0, + "content": "[82] Weihan Wang, Qingsong Lv, Wenmeng Yu, Wenyi Hong, Ji Qi, Yan Wang, Junhui Ji, Zhuoyi Yang, Lei Zhao, Xixuan Song, et al. Cogvlm: Visual expert for pretrained language models. arXiv preprint arXiv:2311.03079, 2023. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.42, + 0.483, + 0.489 + ], + "angle": 0, + "content": "[83] Weiyun Wang, Yiming Ren, Haowen Luo, Tiantong Li, Chenxiang Yan, Zhe Chen, Wenhai Wang, Qingyun Li, Lewei Lu, Xizhou Zhu, et al. The all-seeing project v2: Towards general relation comprehension of the open world. arXiv preprint arXiv:2402.19474, 2024. 1, 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.491, + 0.483, + 0.559 + ], + "angle": 0, + "content": "[84] Weiyun Wang, Min Shi, Qingyun Li, Wenhai Wang, Zhenhang Huang, Linjie Xing, Zhe Chen, Hao Li, Xizhou Zhu, Zhiguo Cao, et al. The all-seeing project: Towards panoptic visual recognition and understanding of the open world. In ICLR, 2024. 1, 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.562, + 0.483, + 0.631 + ], + "angle": 0, + "content": "[85] Weiyun Wang, Shuibio Zhang, Yiming Ren, Yuchen Duan, Tiantong Li, Shuo Liu, Mengkang Hu, Zhe Chen, Kaipeng Zhang, Lewei Lu, Xizhou Zhu, Ping Luo, Yu Qiao, Jifeng Dai, Wenqi Shao, and Wenhai Wang. Needle in a multimodal haystack. arXiv preprint arXiv:2406.07230, 2024. 7, 8, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.633, + 0.483, + 0.688 + ], + "angle": 0, + "content": "[86] Weiyun Wang, Shuibio Zhang, Yiming Ren, Yuchen Duan, Tiantong Li, Shuo Liu, Mengkang Hu, Zhe Chen, Kaipeng Zhang, Lewei Lu, et al. Needle in a multimodal haystack. arXiv preprint arXiv:2406.07230, 2024. 2, 7, 8, 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.69, + 0.483, + 0.744 + ], + "angle": 0, + "content": "[87] Yonghui Wang, Wengang Zhou, Hao Feng, Keyi Zhou, and Houqiang Li. Towards improving document understanding: An exploration on text-grounding via mllms. arXiv preprint arXiv:2311.13194, 2023. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.747, + 0.483, + 0.814 + ], + "angle": 0, + "content": "[88] Haoran Wei, Lingyu Kong, Jinyue Chen, Liang Zhao, Zheng Ge, Jinrong Yang, Jianjian Sun, Chunrui Han, and Xiangyu Zhang. Vary: Scaling up the vision vocabulary for large vision-language models. arXiv preprint arXiv:2312.06109, 2023. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.818, + 0.483, + 0.9 + ], + "angle": 0, + "content": "[89] Renqiu Xia, Song Mao, Xiangchao Yan, Hongbin Zhou, Bo Zhang, Haoyang Peng, Jiahao Pi, Daocheng Fu, Wenjie Wu, Hancheng Ye, et al. Docgenome: An open largescale scientific document benchmark for training and testing multi-modal large language models. arXiv preprint arXiv:2406.11633, 2024. 1, 7, 2, 5" + }, + { + "type": "list", + "bbox": [ + 0.093, + 0.092, + 0.484, + 0.9 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.092, + 0.905, + 0.134 + ], + "angle": 0, + "content": "[90] Guangxuan Xiao, Yuandong Tian, Beidi Chen, Song Han, and Mike Lewis. Efficient streaming language models with attention sinks. arXiv preprint arXiv:2309.17453, 2023. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.136, + 0.905, + 0.177 + ], + "angle": 0, + "content": "[91] Jianxin Yang. Longqlora: Efficient and effective method to extend context length of large language models. arXiv preprint arXiv:2311.04879, 2023. 2, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.179, + 0.905, + 0.233 + ], + "angle": 0, + "content": "[92] Yuan Yao, Tianyu Yu, Ao Zhang, Chongyi Wang, Junbo Cui, Hongji Zhu, Tianchi Cai, Haoyu Li, Weilin Zhao, Zhihui He, et al. Minicpm-v: A gpt-4v level mllm on your phone. arXiv preprint arXiv:2408.01800, 2024. 1, 7, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.235, + 0.905, + 0.303 + ], + "angle": 0, + "content": "[93] Jiabo Ye, Anwen Hu, Haiyang Xu, Qinghao Ye, Ming Yan, Yuhao Dan, Chenlin Zhao, Guohai Xu, Chenliang Li, Junfeng Tian, et al. mplug-docowl: Modularized multimodal large language model for document understanding. arXiv preprint arXiv:2307.02499, 2023. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.305, + 0.905, + 0.374 + ], + "angle": 0, + "content": "[94] Jiabo Ye, Anwen Hu, Haiyang Xu, Qinghao Ye, Ming Yan, Guohai Xu, Chenliang Li, Junfeng Tian, Qi Qian, Ji Zhang, et al. Ureader: UniversalOCR-free visually-situated language understanding with multimodal large language model. arXiv preprint arXiv:2310.05126, 2023. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.376, + 0.905, + 0.444 + ], + "angle": 0, + "content": "[95] Shi Yu, Chaoyue Tang, Bokai Xu, Junbo Cui, Junhao Ran, Yukun Yan, Zhenghao Liu, Shuo Wang, Xu Han, Zhiyuan Liu, et al. Visrag: Vision-based retrieval-augmented generation on multi-modality documents. arXiv preprint arXiv:2410.10594, 2024. 1, 2, 3, 7, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.446, + 0.905, + 0.514 + ], + "angle": 0, + "content": "[96] Jiajie Zhang, Yushi Bai, Xin Lv, Wanjun Gu, Danqing Liu, Minhao Zou, Shulin Cao, Lei Hou, Yuxiao Dong, Ling Feng, et al. Longcite: Enabling llms to generate fine-grained citations in long-context qa. arXiv e-prints, pages arXiv-2409, 2024. 2, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.517, + 0.905, + 0.585 + ], + "angle": 0, + "content": "[97] Jiajie Zhang, Zhongni Hou, Xin Lv, Shulin Cao, Zhenyu Hou, Yilin Niu, Lei Hou, Yuxiao Dong, Ling Feng, and Juanzi Li. Longreward: Improving long-context large language models with ai feedback. arXiv preprint arXiv:2410.21252, 2024. 2, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.587, + 0.905, + 0.656 + ], + "angle": 0, + "content": "[98] Pan Zhang, Xiaoyi Dong, Yuhang Zang, Yuhang Cao, Rui Qian, Lin Chen, Qipeng Guo, Haodong Duan, Bin Wang, Linke Ouyang, et al. Internlm-xcomposer-2.5: A versatile large vision language model supporting long-contextual input and output. arXiv preprint arXiv:2407.03320, 2024. 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.657, + 0.905, + 0.726 + ], + "angle": 0, + "content": "[99] Wanrong Zhu, Jack Hessel, Anas Awadalla, Samir Yitzhak Gadre, Jesse Dodge, Alex Fang, Youngjae Yu, Ludwig Schmidt, William Yang Wang, and Yejin Choi. Multimodal c4: An open, billion-scale corpus of images interleaved with text. NIPS, 36, 2024. 2" + }, + { + "type": "list", + "bbox": [ + 0.517, + 0.092, + 0.905, + 0.726 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.516, + 0.956 + ], + "angle": 0, + "content": "4037" + } + ] +] \ No newline at end of file diff --git a/2025/Docopilot_ Improving Multimodal Models for Document-Level Understanding/2cc56609-3eb1-412a-98df-28e82ae9cd58_origin.pdf b/2025/Docopilot_ Improving Multimodal Models for Document-Level Understanding/2cc56609-3eb1-412a-98df-28e82ae9cd58_origin.pdf new file mode 100644 index 0000000000000000000000000000000000000000..5858b7753ce3a2cb80c91423a99ed22fbd1869b1 --- /dev/null +++ b/2025/Docopilot_ Improving Multimodal Models for Document-Level Understanding/2cc56609-3eb1-412a-98df-28e82ae9cd58_origin.pdf @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:5f26d2b6545c61646239f11e8910ee560f8c3fefcfac4a5b965a43ee8bf12aea +size 3077608 diff --git a/2025/Docopilot_ Improving Multimodal Models for Document-Level Understanding/full.md b/2025/Docopilot_ Improving Multimodal Models for Document-Level Understanding/full.md new file mode 100644 index 0000000000000000000000000000000000000000..697149beaeb521280df3f27c59b594ab8a085080 --- /dev/null +++ b/2025/Docopilot_ Improving Multimodal Models for Document-Level Understanding/full.md @@ -0,0 +1,348 @@ +# Docopilot: Improving Multimodal Models for Document-Level Understanding + +Yuchen Duan $^{1,2*}$ , Zhe Chen $^{3,1*}$ , Yusong Hu $^{4,1*}$ , Weiyun Wang $^{*5,1}$ , Shenglong Ye $^{1}$ , Botian Shi $^{1}$ , Lewei Lu $^{7}$ , Qibin Hou $^{4}$ , Tong Lu $^{3,1}$ , Hongsheng Li $^{2,1}$ , Jifeng Dai $^{6,1}$ , Wenhai Wang $^{2,1\boxtimes}$ + +$^{1}$ Shanghai AI Laboratory, $^{2}$ The Chinese University of Hong Kong, $^{3}$ Nanjing University, $^{4}$ Nankai University, $^{5}$ Fudan University, $^{6}$ Tsinghua University, $^{7}$ SenseTime Research + +# Abstract + +Despite significant progress in multimodal large language models (MLLMs), their performance on complex, multipage document comprehension remains inadequate, largely due to the lack of high-quality, document-level datasets. While current retrieval-augmented generation (RAG) methods offer partial solutions, they suffer from issues, such as fragmented retrieval contexts, multi-stage error accumulation, and extra time costs of retrieval. In this work, we present a high-quality document-level dataset, Doc750K, designed to support in-depth understanding of multimodal documents. This dataset includes diverse document structures, extensive cross-page dependencies, and real question-answer pairs derived from the original documents. Building on the dataset, we develop a native multimodal model—Docopilot, which can accurately handle document-level dependencies without relying on RAG. Experiments demonstrate that Docopilot achieves superior coherence, accuracy, and efficiency in document understanding tasks and multi-turn interactions, setting a new baseline for document-level multimodal understanding. Data, code, and models are released at https://github.com/OpenGVLab/Docopilot. + +# 1. Introduction + +In recent years, multimodal large language models (MLLMs) [6, 12, 41, 56, 60, 72, 81, 82, 84, 92] have rapidly developed, achieving remarkable performance in various visual understanding tasks [30, 77], particularly image-level tasks, such as image captioning [10, 42], optical character recognition (OCR) [45, 65], and visual question answering (VQA) [24, 44]. Despite these advances, current MLLMs still face significant challenges in document-level understanding [51, 76, 89], where models are required to identify and integrate key information across multi-page documents, setting high expectations for their long-context processing + +![](images/ca45d626bd25fc9f7b1548f606b52489b79335623c8f6b671344188d240b28ca.jpg) +Figure 1. Accuracy $\nu.s$ inference latency on MM-NIAH. The proposed Docopilot-8B shows a notable improvement over baseline models [73], achieving a $+19.9\%$ accuracy gain compared to InternVL2-8B and surpassing InternVL2-26B with less than $31\%$ of the inference latency. Additionally, Docopilot-2B uses fewer parameters (less than $10\%$ ) while exhibiting comparable performance to the $10\times$ larger InternVL2-26B. These results suggest that our Docopilot strikes a reasonable balance between latency, model size, and performance. + +# capabilities of MLLMs. + +Current research on long-content understanding primarily focuses on text-only models [5, 9, 74], targeting specific retrieval tasks such as "Needle in a Haystack" (NIAH) [2, 31]. However, existing open-source MLLMs [12, 13, 36, 41, 62, 81, 83] are primarily trained on image-level data, lacking the long-context understanding capacity required for document-level understanding. Retrieval-augmented generation (RAG) methods [14, 22, 50, 61, 63, 95] attempt to address this by retrieving key information to fit within the limited context windows of MLLMs, but they still encounter the following challenges in document-level tasks. (1) Fragmented Retrieval Contexts. Retrieved information is fragmented, lacking the overall structure of the document; (2) Multi-Stage Error Accumulation. Incorrect retrieval results can affect subsequent responses, leading to errors or omissions of critical details, especially in multi-turn or complex tasks; (3) Extra Time Costs. The retrieval + +step increases the latency of the QA system, limiting the scalability of RAG in time-sensitive scenarios. + +To address these problems, two primary challenges need to be considered. (1) High-Quality Multimodal Document Dataset. While extensive datasets [11, 91, 96, 97] exist for long-context, text-only tasks, high-quality document-level question-answering datasets remain scarce. This shortage is largely attributed to the high costs associated with annotation and the lack of streamlined construction pipelines. (2) Native Document-Level MLLMs. Although RAG-based methods [14, 86, 95] provide some relief, native multimodal models with long-context processing abilities are crucial. However, training native MLLMs specifically for document-level understanding is constrained by current hardware limitations. + +In this work, we introduce a new multimodal document dataset that supports document-level understanding tasks. Compared to counterparts [33, 76, 78], this dataset has the following features: (1) Large Scale. It includes a total of 758K question-answer samples, containing 5.2B text tokens and 3.1M images. It encompasses content from various sources, such as Sci-Hub, Arxiv, and OpenReview, covering a wide range of topics and document layouts. (2) High Quality. Unlike existing datasets that insert irrelevant questions into documents, we collect real, in-depth question-answer pairs and construct single-page and crosspage questions based on document structure. Such high-quality question-answer data accounts for $31.6\%$ of the dataset. (3) Multimodal. For the document content, we provide not only the conventional interleaved text-image context but also purely rendered image inputs, catering to the needs of different models. + +Building upon this dataset, we developed a native baseline model for document-level multimodal understanding-Docopilot. Unlike existing approaches [14, 86, 95] that rely on RAG, our model achieves efficient document-level training and testing through simple engineering optimizations, such as multimodal data packing, Ring Attention [43], and Liger Kernel [16]. Leveraging the proxy tasks carefully designed within the dataset, Docopilot can directly handle long-distance dependencies and cross-page information integration without external retrieval support. As shown in Figure 1, this approach not only enhances coherence and accuracy compared to RAG methods but also significantly reduces the response time of the entire question-answering system, delivering superior real-time performance in multi-turn interactions. + +The main contributions are summarized as follows: + +(1) We develop the first large-scale, high-quality dataset for document-level multimodal understanding, consisting of 758K QA pairs from 3 sources, supporting 9 types of proxy tasks. This dataset includes $31.6\%$ real QA pairs directly extracted from documents. + +(2) Based on the dataset, we implement Docopilot, a native MLLM designed for document-level understanding without relying on retrieval mechanisms. This approach greatly improves its ability to integrate and comprehend information across multi-page documents. +(3) Through extensive experiments on multiple document-level benchmarks, our method demonstrates performance significantly superior to existing approaches, proving its effectiveness and generality. As shown in Figure 1, Docopilot-8B achieves a score of 61.8 on MM-NIAH [86], outperforming InternVL2-8B by 19.9 points and surpassing InternVL2-26B with less than $31\%$ of the latency. We hope this work could provide a baseline for future advancements in MLLMs for document-level tasks. + +# 2. Related Work + +# 2.1. Multimodal Large Language Models + +Multimodal large language models (MLLMs) have demonstrated impressive capabilities in processing image and text information, opening up new directions for applications such as visual question answering and image captioning. Early models [12, 29, 37, 59] trained with contrastive learning methods excelled in recognizing and understanding open-world semantics within an image-text matching framework. However, their limited generative abilities restricted their applicability. To leverage the powerful generation abilities of large language models (LLMs), subsequent works [13, 38, 42, 47, 83, 84] introduced a connector to align the embedding spaces of vision encoders and LLMs, allowing encoded image embeddings to serve as soft prompts for LLMs. Another series of works [1, 34, 39, 99] extended LLMs by integrating additional visual experts, reducing reliance on standalone vision encoders. More recently, models capable of both understanding and generating images have also made notable progress [20, 35, 67, 75], leveraging the insight that image generation can enhance image understanding. Despite these advancements, current MLLMs still face challenges with long-context multimodal inputs. For instance, InternVL 2.0 [13, 25] performs optimally within a token range of up to 8192, constraining its effectiveness in document-level applications. + +# 2.2. Document Understanding Models + +Extracting key information from documents is crucial for industries and academic research. OCR-model-driven methods [3, 4, 71, 80] represent one of the primary technical approaches. These methods extract text, layout, and bounding box information from external systems and integrate it with another model. However, they are prone to error propagation and high processing times due to their reliance on multiple components. Benefitting from the rapid advancements in LLMs, OCR-free methods have also achieved + +![](images/18b2b2256ee8d2ed25df211f25d986c42c96fd2728e54c3e98f1529a5a7d8047.jpg) +Figure 2. Multimodal document dataset generation pipeline. This pipeline involves three main stages: (1) Raw Data Collection: Documents are gathered from sources like Sci-Hub, arXiv, and OpenReview, available in PDF and HTML formats. (2) Document Content Extraction: Multimodal content is processed in two formats: interleaved text-image format and multi-image format. (3) Question-Answer Pairs Construction: QA pairs are generated based on the document structure or constructed using GPT-4o. + +great progress. Donut [32] is the first end-to-end training framework based on a Transformer without requiring OCR engines or APIs. Subsequent works [23, 49, 87, 88, 94] propose diverse modifications in model architectures and training algorithms. However, these models are designed for specific tasks and lack general abilities. + +# 2.3. Long-Context Large Language Models + +With advancements in engineering, architecture, and algorithms, long-context large language models have made substantial progress. Techniques such as Flash Attention [17, 18] and Ring Attention [43] have notably reduced GPU memory usage for training on extended contexts. Additionally, various sparse attention mechanisms—including Shifting Sparse Attention [11], Dilated Attention [19], and Attention Sinks [26, 90]—have enabled efficient scaling to handle larger contexts. New positional embedding methods, like ALiBi [58], xPOS [68], and RoPE [66], further enhance the models' generalization capabilities in length extrapolation. However, these advancements remain largely confined to natural language processing, and methods to extend the context size of MLLMs are still under-explored. Another research approach aims to reduce context size by leveraging retrieval augmented generation (RAG) [14, 22, 95], where only the most relevant passages are retrieved and fed into the generation model. However, this retrieval-based approach can disrupt the coherence of the semantic chain, particularly in complex reasoning tasks, due to fragmented information flow. In this work, we integrate the above engineering techniques into MLLMs and demonstrate that a model fine-tuned on a high-quality, long-context training corpus is a strong baseline, achieving superior performance compared to its RAG counterpart. + +# 3. Multimodal Document Dataset Generation + +In this section, we begin by introducing the details of the data engine. Following this, we provide a comprehensive overview of the dataset—Doc-750K. + +# 3.1. Data Engine + +The data engine operates primarily through two steps: document content extraction and question-answer (QA) pair construction. Specifically, we first extract multimodal content, including both text and images, from the documents. Based on this extracted content, we then create question-answer pairs. The document content and these constructed pairs are combined to produce conversational-style training data. The format is outlined as: + +Please read the paper: , and answer the question: Answer: + +Here, the , , and are the placeholder for extracted document content, the generated questions and answer, respectively. In the following, we will provide a detailed explanation of the two key steps: document content extraction and QA pair construction. + +Document Content Extraction. In practical applications, different documents have varying page layouts and content types, which poses significant challenges for content extraction. To enhance the efficiency of multimodal models, it is necessary to organize documents into a unified format for streamlined processing. In this work, we process each document into two formats as follows: + +(1) Interleaved Text-Image Format. Using the document content extractor MinerU [79], we segment the document content into interleaved text and image annotations, for + +The objective consideration of conservation policy under restructuring is proving to be a difficult task. One of the greatest obstacles has been the persistent uncertainty among utility planners regarding the true resource-effectiveness and cost-effectiveness of conservation relative + +These figures indicate that larger utilities tend to report greater ... + +![](images/1e8cd6897d0cf21bd56005d043630dcd5da915d354ca58756e0370a1e9e36284.jpg) +Figure 3. Visualization of an example from Doc-750K. The left side presents the interleaved text-image format data obtained through Document Content Extraction, while the right side showcases the annotations generated via Question-Answer Pairs Construction. + +![](images/7c96ebc511b971ca903adf4f92d5f88de2e4c30fbb329885667a3f9be9335907.jpg) + +Early estimation attempts with individual utilities revealed that there was insufficient variation in the data series to produce reliable coefficient estimates for each utility taken separately. To increase the statistical power of the analysis, ... + +Finally, since no incentives are offered, it becomes more difficult to attribute customer investment decisions to the utility programs as opposed to other influences. + +
Utility (C&I Sales)Program Type1993 Conservation Impacts(Annotated)
GWh% of Sales% of Impacts
Pacific Gas & Electric(45,760 GWh)Incentives1,8864.1%63%
Services9402.1%31%
Other170.4%6%
Total3,0136.6%100%
San Diego Gas & Electric(8,415 GWh)Incentives2512.7%65%
Services971.0%25%
Other2820.4%100%
Total3854.1%100%
Southern California Edison(86,721 GWh)Incentives1,7283.7%22%
Services6,02912.9%71%
Other2400.5%3%
Total7,97417.1%100%
Sacramento(8,733 GWh)Incentives1533.2%58%
Services1042.2%40%
Other60.1%2%
Total2822.5%100%
Los Angeles(14,829 GWh)Incentives7800.5%50%
Services7910.5%50%
Other00.0%0%
Total1571.1%100%
+ +
StudySampleFactorPeriodElectricity
Elec.UnitCostFuel
Longuet et al. 1994Norway 1976aenergy consumptionLB−0.39−0.39
energy consumptionLB−0.15−0.15
energy consumption & transportLB−0.44−0.16
Hammoudi & Kazer 1995Iran 1955–8domestic heatingLB−0.19−0.19
Safdarzadeh 1995USAcommercialLB−0.70−0.25
industrialLB−0.22−0.22
Rall 1996USAindustrialLB−0.18−0.20−0.150.09
industrialLB−0.29−0.29
Counihan 1998USA 1995–8industrialLB−0.390.530.040.09
industrialLB−0.22−0.22
Murray 2000Virginia, USA 1970–8industrialLB−0.29−0.09
industrialLB−0.22−0.22
industrialLB−0.21−0.04
Woodward 2001Netherlands 1977–83industrialLB−1.11−0.2−0.39−0.16
industrialLB−0.11−0.11
National Energy Board 2003CanadaindustrialLB−0.110.020.05
industrialLB−0.11−0.11
industrialLB−0.450.120.06
Energy, Media and Media Ethics Board Canada 2005CanadaindustrialLB−0.090.030.01
industrialLB−0.09−0.09
industrialLB−0.490.110.01
This studyUSAcommercial RLB−0.078−0.033−0.026−0.013
industrialLB−0.078−0.078
+ +According to California regulators, SCE's reported impacts for service programs have always been viewed with skepticism during regulatory reviews, although its impacts for other types of conservation programs have not. In California, service program impacts have historically been viewed as unverifiable, so utilities have been largely ineligible for incentive payments related to service program expenditures. Regulators in California acknowledge that they believe SCE's reported conservation impacts overstate the actual savings achieved by their programs and have communicated this skepticism to the utility. But, according to the rate regulation ... + + + +Please read the paper and answer the question: + +According to this paper, how does Image-1 illustrate the conservation impacts of various utilities in 1993? Discuss the findings and the significance of these data in the context of the research. + +![](images/04fe59a24ac3d6bbbff74037dab49d3c97c86e29b1e4bb485e0920f30f7f9889.jpg) + +According to the paper, 'In 1993, the 39 companies in the sample accounted for $19\%$ of total C&I electricity sales nationwide and $40\%$ of all utility conservation program impacts, including residential impacts (Energy Information Administration 1995). The table in Image-1 provides detailed data on the 1993 C&I sales and C&I conservation program impacts for various utilities. This image shows the sales in GWh and conservation impacts both in GWh and as a percentage of sales. For example, California utilities such as Southern California Edison reported high sales and high conservation percentages, which is a significant outlier. These data illustrate the varying levels of conservation achievements across utilities, underlining the study's aim to verify the effectiveness and accuracy of reported conservation impacts. Larger utilities tend to report greater conservation impacts relative to sales, and this detail supports the analysis presented in the paper. + +Analyze the importance of Image-3 in comparing the study's findings with previous empirical research on electricity price elasticities. + +![](images/c880a59ac2258e15418036f94388275e4684eff8fb2a96229ca00928bb57542f.jpg) + +![](images/392bb71143c9e71645c114493265a9b3032c0eeb7ff4d7c239ff8285ff70a36a.jpg) + +According to the paper, 'The regression diagnostics offered little guidance on improving the structure of the model. The addition of a dummy variable for the years 1971-1973 did little to improve the residuals in subsequent years. Adding a set of dummy variables to account for each unstable period did not seem reasonable.' Image-3 presents a comparative analysis of electricity cross-price elasticity studies. This table showcases various studies, their sample sectors, and reported elasticities for electricity, natural gas, oil, and coal across different regions and time periods. The study's own findings, -0.078 for short-run own-price elasticity of electricity, align closely with prior research, confirming its place within established literature. The context provided by other studies helps to validate the paper's conclusions regarding the robustness of its econometric model and the effectiveness of conservation programs. + +example, "\n\n \n" This format captures the document's textual content, making it easier to construct question-answer pairs. + +(2) Multi-Image Format. In this format, a document with $n$ pages is rendered as $n$ images, with each image corresponding to a single page. The structure follows the pattern "\n\n". This format preserves the original layout, enabling the model to learn the overall pagination and visual layout of the document. + +After processing the document into contexts in interleaved text-image and paginated image formats, we can not only use these contexts for next-token prediction training but also leverage the document's content, hierarchical structure, and layout features to flexibly and precisely generate high-quality question-answer pairs. + +Question-Answer Pairs Construction. In this step, we create question-and-answer pairs tailored to the source, content features, and formatting structure of each document. This process is divided into the following main categories: + +(1) For documents with reliable QA annotations, like the review and reply in OpenReview, we extract the QA pairs and organize them into conversation format. +(2) For documents with a clear textual structure, such as well-structured papers from Sci-Hub and Arxiv, we convert them to text and segment them, while the model is instructed to generate contents for each segment, including abstracts, experiment descriptions, and captions for figures and tables. The details of each task for structural papers are illustrated in Table 1. +(3) For other documents, in addition to using them directly for NTP pretraining, we can input text interspersed with images into MLLMs to obtain QA pairs. To ensure high-quality generated data, we use the state-of-the-art model GPT-4o [56]. + +Through our pipeline, most data has been processed into high-quality document-level question-answering data, while the remaining data is converted to plain text and used for next-token prediction tasks. Our pipelines are metricu + +![](images/728d8d93a75762d4ff8bb882269a03389f5b3a05187bda3f5deb0af18bf957ae.jpg) +(a) Data Distribution of Doc-750K +(b) Data Recipe for Supervised Fine-tuning +Figure 4. Data distribution of our dataset. The outer circle shows the distribution of all data categories and the inner circle shows the distribution of data subsets. Left: Data distribution of Doc-750K. Right: Data distribution of our complete SFT training dataset. Note that the number reported in the figure represents the number of samples. “MT” is short for multi-turn. + +
TasksQuestions
Abstract WritingRead the full text of the paper and provide a concise summary in the form of an abstract.
Paper TitlingBased on the provided abstract or introduction of the research paper, please generate a concise and informative title
Caption WritingGive the relative texts of the images or tables, please write a caption for each image or table based on the relative texts provided.
Experiment WritingPlease write the "Experiments" section based on the incomplete research paper provided.
TranslationPlease read the full text of the following research paper and translate the Experiments section into Chinese.
+ +lously designed to ensure high data quality across all generated context. Each LLM-generated sample is explicitly marked in the metadata as model-generated. Across the entire dataset, only $4.8\%$ of the data is LLM-generated, reinforcing the overall reliability and quality of the dataset. + +# 3.2. Multimodal Document Dataset + +Data Source. The composition and distribution of our training data are detailed in Figure 4. Specifically, our dataset predominantly consists of academic papers, which constitute approximately $32.6\%$ of the total data. The multimodal data, carefully selected to augment our model's learning dimensions, makes up about $88.8\%$ of our dataset. This strategic distribution is designed to optimize the train- + +Table 1. Questions format for different tasks. For documents with a clear textual structure, we design several proxy tasks. All tasks leverage the inherent structure of the documents, with answers directly sourced from the original text. + +
StatisticsNumber
Total Questions758K
Total Images3.1M
Total Conversations251K
Multi-Turn Questions87K
Single-Turn Questions164K
Average Text Tokens11245
Average Image Tokens6178
+ +Table 2. Key statistics of the Doc-750K datasets. It comprises 758K questions, 3.1M images, and 251K conversations, including 87K multi-turn and 164K single-turn questions. With an average of 11,245 text tokens and 6,178 image tokens, it highlights the dataset's richness and diversity for multimodal research. + +ing process and improve the model's ability to generalize across different types of data inputs. + +Dataset Statistics. In our Doc-750K dataset, the majority of the data consists of reliably annotated entries, with OpenReview and Arxiv collectively accounting for $75.4\%$ . The remaining data, sourced from Sci-Hub, is processed using our designed tasks. The overall distribution and number of tasks are shown in Figure 4(a). Our dataset ultimately consists of 251K conversations, comprising a total of 758K questions. Additional statistical details are provided in Table 2. Compared to previous datasets, Doc-750K contains a larger number of images, with an average of four images per conversation segment. Further comparisons with other datasets are shown in Table 3. + +# 3.3. Data Recipe for Supervised Fine-Tuning + +Although Doc-750K effectively covers multimodal document QA scenarios, using it directly may lead to model over-fitting on a specific document domain. Therefore, we combine it with several open-source datasets to create a mixed dataset for SFT training. As shown in Figure 4(b), these datasets are organized into 4 categories as follows: + +(1) For multi-page document QA, Doc-750K serves + +
Dataset#Images#QA Pairs#Tokens
Docmatix [33]2,444,7509,500,000390,000,000
DocVQA [15]10,18939,463337,829
TextCaps [64]21,95321,953389,658
TextVQA [65]21,95334,602181,918
ST-VQA [8]17,24723,121127,846
OCR-VQA [55]165,746801,5796,073,824
VisualMRC [70]3,02711,988168,828
DUDE [78]147,59723,71611,341,228
Doc-750K (ours)3,103,494758,0005,200,000,000
+ +Table 3. Comparison with popular VQA datasets. + +as the core dataset, specifically curated to address complex, multi-page document comprehension. Additional datasets such as MP-Docmatix [33], MP-DocVQA [53], DUDE [78], and Taesiri-ArxivQA [69] offer valuable multi-page scenarios requiring inter-page reasoning and contextual retention across sequences. + +(2) For multi-image general QA, MMDU-45K [48] offers a comprehensive dataset encompassing diverse real-world scenarios, such as natural environments and everyday contexts. It emphasizes multi-turn dialogues and integration of multiple images, supporting the development of systems capable of generating coherent and accurate responses from complex, lengthy inputs. +(3) For single-page document QA, We introduce DocVQA [53], DocReason [93], InfoVQA [54], and ChartQA [52] to further enhance the diversity of the SFT dataset. These datasets focus on individual pages with complex layouts, rich textual information, and, in some cases, graphical data interpretation. +(4) For pure-text QA, we add datasets including LongAlpaca [11], LongAlpaca-16K-Length [11], LongQLoRA [91], LongCite [96], LongAlign [7], and LongReward [97] to support the assessment of the model's capabilities in QA tasks requiring long-range dependencies. + +This expanded dataset provides a balanced foundation for training and evaluating multimodal document understanding models, enhancing robustness and adaptability across diverse document-related VQA tasks. + +# 4. Enhanced Baseline for Document-Level Multimodal Understanding + +# 4.1. Model Architecture + +Our model architecture leverages the widely-adopted ViTMLP-LLM structure [41, 42, 73], consisting of a pre-trained Vision Transformer (ViT), a two-layer MLP projector, and a pre-trained Language Model (LLM). This combination provides a strong baseline for multimodal document analysis, effectively integrating visual and textual information within a unified framework. + +# 4.2. Optimizing Training Efficiency + +The training efficiency of MLLMs is hindered by two key challenges: (1) Inconsistent Sample Lengths. Samples with different context lengths will result in excessive padding and lower training throughput; and (2) Limited GPU Memory. As the model scale and context length increase, GPU memory consumption becomes increasingly unsustainable. To address these issues, we have implemented the following strategies: + +(1) Multimodal Data Packing. To balance the computational load between the vision model (ViT) and the language model (LLM) while minimizing resource waste caused by padding, we implement a multimodal data-packing strategy. The key idea is to concatenate multiple samples into long sequences to fully utilize the model's input capacity. Specifically, thresholds $T_{\mathrm{img}}$ and $T_{\mathrm{tok}}$ are set for the number of images and tokens, respectively. Samples are managed using a priority queue, sorted in descending order by the number of images and total tokens. A new sample $s$ attempts to combine with the sample at the front of the priority queue. If the combination meets the thresholds (i.e., $T_{\mathrm{img}}, T_{\mathrm{tok}}$ ), the combined sample is pushed back into the priority queue. If $s$ cannot match with any existing sample in the queue, it is directly added to the queue. When the image number and total token number of the front sample reach one of the thresholds or the number of samples exceeds the maximum limit $M$ , the front sample is dequeued, padded as needed, and sent for training. This strategy optimizes resource utilization and ensures balanced computational workloads. The detailed pseudo-code can be found in the supplementary materials. +(2) Ring Attention. We implement the Ring Attention mechanism [43] to alleviate memory constraints associated with processing long sequences. By partitioning sequences into blocks and distributing computation across multiple devices, Ring Attention allows the model to accommodate larger contexts. This approach enables overlapping communication between key-value blocks and attention computations, thereby enhancing parallel processing efficiency. Consequently, Ring Attention improves the model's capacity to handle extended context lengths without exceeding memory limits. +(3) Liger Kernel. To further improve memory and computational efficiency, we integrate the Liger Kernel [16], a specialized kernel library optimized for large-scale model training. The Liger Kernel enhances throughput and reduces memory consumption by employing techniques like kernel fusion, in-place operations, and input chunking. Leveraging the Liger Kernel thus enables higher training throughput and addresses memory limitations, allowing for efficient scaling of large multimodal models. + +
ModelsMP-DocMMLong-DocDocGenomeMM-NIAH
ANSL↑Acc↑F1↑Class Acc↑Title ED↓Abstract ED↓SP Acc↑MP Acc↑ShortMediumLongOverall
Proprietary Models
Gemini-1.5-Pro [60]-28.220.6-----73.865.260.867.1
GPT-4o [56]-42.844.997.69.56.571.867.6----
Open-Source Models
MiniMonkey-2B [28]70.310.38.657.416.555.040.328.940.926.923.531.0
InternVL2-2B [13]71.810.510.860.818.454.339.428.936.621.219.426.4
InternVL2-2B + RAG [85]72.617.216.760.818.454.339.428.436.830.234.833.8
Llama3.2-3B-Instruct†[21]-23.721.285.3194.751.040.234.915.52.20.56.6
Docopilot-2B (ours)76.221.816.056.24.543.645.137.458.046.740.949.2
MiniCPM-V2.6-8B [92]-16.915.492.810.232.660.054.249.015.30.023.4
LLaVA-OneVision-8B [36]-10.89.685.649.977.59.87.165.738.00.037.7
mPLUG-DocOwl2-8B [27]69.413.48.9-----17.90.10.06.6
M3DocRAG [14]84.421.022.6---------
VisRAG-8B [95]-18.818.392.810.232.660.050.747.129.229.535.8
InternLM2.5-7B-1M†[9]-28.725.692.777.659.342.742.540.537.235.137.8
InternVL2-8B [13]79.317.416.590.68.239.656.046.156.437.332.442.9
InternVL2-8B + RAG [85]78.724.224.590.68.239.656.046.055.743.445.248.4
InternVL2-26B [13]-15.515.487.516.923.349.742.765.048.741.952.8
Docopilot-8B (ours)81.328.823.093.82.019.753.951.971.257.455.361.8
+ +# 5. Experiments + +# 5.1. Experimental Setup + +Training Details. Our model is available in two sizes: Docopilot-2B and Docopilot-8B, both of which are based on the InternVL2 [73] and fine-tuned for one epoch using the data recipe that includes Doc-750K. The training uses a batch size of 128, the AdamW optimizer with a learning rate of 1e-5, weight decay of 0.01 for the 2B variant, and 0.05 for the 8B variant, along with a cosine learning rate schedule. To speed up training, we apply multimodal data packing to reduce padding and a dynamic high-resolution strategy [13] to enhance OCR for document understanding. The maximum number of tiles for multimodal data is limited to 24, and the maximum sequence length is set to 32k tokens. + +Baselines. We compare our Docopilot with a series of open-source document-level MLLMs [27, 28, 36, 40, 46, 92, 98] that supports multi-image input and proprietary MLLMs, including Gemini-1.5-pro [60], GPT-4o [57]. For comparison with the commonly used RAG method for handling long documents, we selected the latest multimodal RAG methods VisRAG [95], InternVL + RAG [85], and M3DocRAG [14]. To compare with more long-context large language models, we use InternVL2-8B as the OCR model to extract texts from the documents and images and feed the parsed documents to long-context LLMs [9, 21]. + +# 5.2. Multi-Page VQA + +Benchmarks. For the multi-page VQA task, we evaluate our model on three benchmarks: (1) MP-DocVQA [76], + +Table 4. Evaluation on multi-page and interleaved VQA benchmarks. We report the metrics on MP-DocVQA [76] (MP-Doc), MMLongbench-Doc [51] (MMLong-Doc), DocGenome [89], and MM-NIAH [85]. Our model outperforms document-level MLLMs and multimodal RAG methods on multi-page, medium, and long-context QA. The "Short", "Medium", and "Long" in MM-NIAH refer to input length in $[0,8\mathrm{k}]$ , $(8\mathrm{k},32\mathrm{k}]$ , $(32\mathrm{k},64\mathrm{k}]$ , respectively. " $\dagger$ ” denotes input documents are parsed by OCR models. + +
ModelsDocVQAChartQAInfoVQA
Gemini-1.5-Pro [60]93.187.281.0
GPT-4o [56]92.885.7-
MiniMonkey-2B [28]87.476.560.1
InternVL2-2B [13]86.976.258.9
Docopilot-2B (ours)87.376.458.5
Monkey-8B [40]66.565.136.1
TextMonkey-9B [46]73.066.928.6
mPLUG-DocOwl2-8B [27]80.770.046.4
IXC2.5-7B [98]90.982.269.9
InternVL2-8B [13]91.683.374.8
Docopilot-8B (ours)92.083.373.3
+ +Table 5. Results on single-page VQA benchmarks. Our Docopilot models perform comparably to baselines [73], demonstrating enhanced long-context modeling without loss on shorter tasks. + +which is designed to evaluate the ability to handle complex questions across multiple scanned document pages. (2) MMLongbench-Doc [51], a benchmark for evaluating the performance of MLLMs on multi-modal documents. (3) DocGenome [89], a large-scale benchmark for the evaluation of scientific document comprehension. + +Results. As illustrated in Table 4, our model achieves consistent improvements on multi-page QA benchmarks, outperforming previous document-level MLLMs. Notably, our Docopilot-8B surpasses Gemini-1.5-Pro [60] on MMLongBench-Doc, positioning it as the closest open-source model to GPT-4o. In comparison to RAG-based methods [14, 86, 95], our model demonstrates advantages in multi-page scenarios. For example, in the Multi/Page QA of DocGenome benchmark, the RAG method shows a perfor + +mance decline due to the disruption of document continuity while our Docopilot exhibits a significantly stable improvement compared to the baseline, with Docopilot-8B showing an increase of $12.6\%$ over InternVL2-8B. + +# 5.3. Interleaved Long-Context QA + +Benchmarks. For the interleaved long-context QA task, we evaluate our models on MM-NIAH [86], a benchmark designed for long multimodal document comprehension. + +Results. The right side of Table 4 presents the results of MM-NIAH across context lengths ranging from 1K to 64K. We categorize the context lengths into "Short," "Medium," and "Long" based on the context window of InternVL2 (8K) and Docopilot (32K). Our Docopilot demonstrates exceptional performance in both medium- and long-context scenarios, while maintaining high accuracy in short-context situations. Notably, for QA tasks with context lengths in the range of $(32K,64K]$ , Docopilot-2B outperforms InternVL2-2B by $110\%$ , and Docopilot-8B surpasses InternVL2-8B by $70\%$ . Furthermore, our model performs comparably to the state-of-the-art multimodal long-context model Gemini-1.5-Pro in contexts longer than 8K, establishing a new state-of-the-art performance among open-source long-context MLLMs. + +# 5.4. Single-Page VQA + +**Benchmarks.** For single-page VQA tasks, we evaluate our model on three benchmarks: (1) DocVQA [53], a benchmark for the evaluation of extracting key information from an image of the given document. (2) ChartQA [52], a benchmark for evaluating the reasoning abilities for chart images. (3) InfoVQA [54], a benchmark for infographic image comprehension. + +Results. As shown in Table 5, our model achieves comparable performance to baseline models. Across the three benchmarks, Docopilot-2B and InternVL2-2B exhibit comparable results, while Docopilot-8B outperforms InternVL2-8B by 0.4 points in DocVQA. These results demonstrate that Doc-750K effectively enhances the model's long-context modeling capabilities without compromising its performance on shorter documents. + +# 5.5. Ablation Study + +Effect of Doc-750K. We conducted ablation studies on MMLongBench-Doc [51] to analyze the impact of our Doc-750K. We divided Doc-750K into 3 parts according to the source of the data: (1) Sci-Hub data; (2) Arxiv data; and (3) OpenReview data. We demonstrate the effects of incorporating each part of the data into the SFT process, reported in Table 6. We observed that with the inclusion of different parts of Doc-750K, the model's performance improves continuously. Utilizing only open-source data results in an inferior F1 score. + +
ModelsAccF1
Baseline (InternVL2-2B [73])10.510.8
- Variant1: SFT using data recipe w/o Doc-750K18.49.4
- Variant2: Variant1 + our Sci-Hub data18.515.2
- Variant3: Variant2 + our Arxiv data20.515.5
Docopilot-2B: Variant3 + our OpenReview data21.816.0
+ +Table 6. Ablation study on the data recipe. We evaluate the effectiveness of the training data from different sources on MMLongBench-Doc. Our Doc-750K can consistently enhance the ability of the model to understand multi-page documents. + +
ModelsLatencyAccF1
MiniCPM-V2.6 [92]225.4ms16.915.4
VisRAG-12B [95]288.3ms18.818.3
InternVL2-2B [13]35.9ms10.510.8
InternVL2-2B + RAG [85]82.9ms17.216.7
Docopilot-2B (ours)35.9ms21.816.0
InternVL2-8B [13]81.0ms17.416.5
InternVL2-8B + RAG [85]113.4ms24.224.5
Docopilot-8B (ours)81.0ms28.823.0
+ +Table 7. Latency analysis. We evaluate the average token output latency of model outputs on MMLongBench-Doc [51]. RAG-based methods [85, 95] exhibit slower processing speeds due to their two-stage inference process, making them less efficient than document MLLMs for handling multimodal long documents. + +Latency Analysis. To compare the latency in inference between RAG methods and our Docopilot, we conducted a latency analysis on MMLongBench-Doc [51], as reported in Table 7. While RAG reduces the document length input to the MLLM, its own time cost remains non-negligible. For instance, InternVL2-2B + RAG is $130\%$ slower than InternVL2-2B, and VisRAG is $28\%$ slower than MiniCPMV2.6. Our Docopilot does not require additional processes and therefore has the same inference time as baseline models, making it more suitable for analyzing long documents. + +# 6. Conclusions + +This work introduced a diverse document-level question-answering dataset that covers complex structures and crosspage dependencies, providing a robust foundation for training and evaluating document understanding models. We also proposed a retrieval-free long-document understanding model that effectively integrates multi-page information, reducing reliance on external retrieval systems. Experimental results show that our model achieves state-of-the-art performance across several document-level QA benchmarks, underscoring its strength in multi-page integration and complex reasoning. Future work will focus on improving computational efficiency, extending the model to larger multimodal tasks, and adapting it to broader applications for enhanced practicality and generalization. + +# Acknowledgments + +This project was supported by the National Key R&D Program of China (No. 2022ZD0161300, 2022ZD0160101), the National Natural Science Foundation of China (No. 62376134, 62372223). Zhe Chen is supported by the Youth PhD Student Research Project under the National Natural Science Foundation (No. 623B2050). + +# References + +[1] Jean-Baptiste Alayrac, Jeff Donahue, Pauline Luc, Antoine Miech, Iain Barr, Yana Hasson, Karel Lenc, Arthur Mensch, Katherine Millican, Malcolm Reynolds, et al. Flamingo: a visual language model for few-shot learning. NeurIPS, 35: 23716-23736, 2022. 2 +[2] Chenxin An, Shansan Gong, Ming Zhong, Xingjian Zhao, Mukai Li, Jun Zhang, Lingpeng Kong, and Xipeng Qiu. L-eval: Instituting standardized evaluation for long context language models. arXiv preprint arXiv:2307.11088, 2023. 1 +[3] Srikar Appalaraju, Peng Tang, Qi Dong, Nishant Sankaran, Yichu Zhou, and R Manmatha. Docformerv2: Local features for document understanding. In Proceedings of the AAAI Conference on Artificial Intelligence, pages 709-718, 2024. +[4] Haoli Bai, Zhiguang Liu, Xiaojun Meng, Wentao Li, Shuang Liu, Nian Xie, Rongfu Zheng, Liangwei Wang, Lu Hou, Jiansheng Wei, et al. Wukong-reader: Multi-modal pretraining for fine-grained visual document understanding. arXiv preprint arXiv:2212.09621, 2022. 2 +[5] Jinze Bai, Shuai Bai, Yunfei Chu, Zeyu Cui, Kai Dang, Xiaodong Deng, Yang Fan, Wenbin Ge, Yu Han, Fei Huang, Binyuan Hui, Luo Ji, Mei Li, Junyang Lin, Runji Lin, Dayiheng Liu, Gao Liu, Chengqiang Lu, Keming Lu, Jianxin Ma, Rui Men, Xingzhang Ren, Xuancheng Ren, Chuanqi Tan, Sinan Tan, Jianhong Tu, Peng Wang, Shijie Wang, Wei Wang, Shengguang Wu, Benfeng Xu, Jin Xu, An Yang, Hao Yang, Jian Yang, Shusheng Yang, Yang Yao, Bowen Yu, Hongyi Yuan, Zheng Yuan, Jianwei Zhang, Xingxuan Zhang, Yichang Zhang, Zhenru Zhang, Chang Zhou, Jingren Zhou, Xiaohuan Zhou, and Tianhang Zhu. Qwen technical report. arXiv preprint arXiv:2309.16609, 2023. 1 +[6] Jinze Bai, Shuai Bai, Shusheng Yang, Shijie Wang, Sinan Tan, Peng Wang, Junyang Lin, Chang Zhou, and Jingren Zhou. Qwen-vl: A frontier large vision-language model with versatile abilities. arXiv preprint arXiv:2308.12966, 2023. 1 +[7] Yushi Bai, Xin Lv, Jiajie Zhang, Yuze He, Ji Qi, Lei Hou, Jie Tang, Yuxiao Dong, and Juanzi Li. Longalign: A recipe for long context alignment of large language models. arXiv preprint arXiv:2401.18058, 2024. 6 +[8] Ali Furkan Biten, Ruben Tito, Andres Mafla, Lluis Gomez, Marcal Rusinol, Ernest Valveny, CV Jawahar, and Dimosthenis Karatzas. Scene text visual question answering. In ICCV, pages 4291-4301, 2019. 6 +[9] Zheng Cai, Maosong Cao, Haojiong Chen, Kai Chen, Keyu Chen, Xin Chen, Xun Chen, Zehui Chen, Zhi Chen, Pei Chu, et al. Internl m2 technical report. arXiv preprint arXiv:2403.17297, 2024. 1, 7 + +[10] Xinlei Chen, Hao Fang, Tsung-Yi Lin, Ramakrishna Vedantam, Saurabh Gupta, Piotr Dólar, and C Lawrence Zitnick. Microsoft coco captions: Data collection and evaluation server. arXiv preprint arXiv:1504.00325, 2015. 1 +[11] Yukang Chen, Shengju Qian, Haotian Tang, Xin Lai, Zhi-jian Liu, Song Han, and Jiaya Jia. Longlora: Efficient finetuning of long-context large language models. arXiv preprint arXiv:2309.12307, 2023. 2, 3, 6 +[12] Zhe Chen, Jiannan Wu, Wenhai Wang, Weijie Su, Guo Chen, Sen Xing, Zhong Muyan, Qinglong Zhang, Xizhou Zhu, Lewei Lu, et al. Internvl: Scaling up vision foundation models and aligning for generic visual-linguistic tasks. arXiv preprint arXiv:2312.14238, 2023. 1, 2 +[13] Zhe Chen, Weiyun Wang, Hao Tian, Shenglong Ye, Zhangwei Gao, Erfei Cui, Wenwen Tong, Kongzhi Hu, Jiapeng Luo, Zheng Ma, et al. How far are we to gpt-4v? closing the gap to commercial multimodal models with open-source suites. arXiv preprint arXiv:2404.16821, 2024. 1, 2, 7, 8 +[14] Jaemin Cho, Debanjan Mahata, Ozan Irsoy, Yujie He, and Mohit Bansal. M3docrag: Multi-modal retrieval is what you need for multi-page multi-document understanding. arXiv preprint arXiv:2411.04952, 2024. 1, 2, 3, 7 +[15] Christopher Clark and Matt Gardner. Simple and effective multi-paragraph reading comprehension. In ACL, pages 845–855, 2018. 6 +[16] Yun Dai, Vignesh Kothapalli, Qingquan Song, Shao Tang, Siyu Zhu, Steven Shimizu, Shivam Sahni, Haowen Ning, Yanning Chen, et al. Liger kernel: Efficient triton kernels for llm training. arXiv preprint arXiv:2410.10989, 2024. 2, 6 +[17] Tri Dao. Flashattention-2: Faster attention with better parallelism and work partitioning. arXiv preprint arXiv:2307.08691, 2023. 3 +[18] Tri Dao, Dan Fu, Stefano Ermon, Atri Rudra, and Christopher Ré. Flashattention: Fast and memory-efficient exact attention with io-awareness. NeurIPS, 35:16344-16359, 2022. 3 +[19] Jiayu Ding, Shuming Ma, Li Dong, Xingxing Zhang, Shaohan Huang, Wenhui Wang, Nanning Zheng, and Furu Wei. Longnet: Scaling transformers to 1,000,000,000 tokens. arXiv preprint arXiv:2307.02486, 2023. 3 +[20] Runpei Dong, Chunrui Han, Yuang Peng, Zekun Qi, Zheng Ge, Jinrong Yang, Liang Zhao, Jianjian Sun, Hongyu Zhou, Haoran Wei, et al. Dreamllm: Synergistic multimodal comprehension and creation. In ICLR, 2024. 2 +[21] Abhimanyu Dubey, Abhinav Jauhri, Abhinav Pandey, Abhishek Kadian, Ahmad Al-Dahle, Aiesha Letman, Akhil Mathur, Alan Schelten, Amy Yang, Angela Fan, et al. The llama 3 herd of models. arXiv preprint arXiv:2407.21783, 2024. 7 +[22] Manuel Faysse, Hugues Sibille, Tony Wu, Gautier Viaud, Céline Hudelot, and Pierre Colombo. Colpali: Efficient document retrieval with vision language models. arXiv preprint arXiv:2407.01449, 2024. 1, 3 +[23] Hao Feng, Qi Liu, Hao Liu, Wengang Zhou, Houqiang Li, and Can Huang. Docpedia: Unleashing the power of large multimodal model in the frequency domain for versatile + +document understanding. arXiv preprint arXiv:2311.11810, 2023.3 +[24] Chaoyou Fu, Peixian Chen, Yunhang Shen, Yulei Qin, Mengdan Zhang, Xu Lin, Zhenyu Qiu, Wei Lin, Jinrui Yang, Xiawu Zheng, et al. Mme: A comprehensive evaluation benchmark for multimodal large language models. arXiv preprint arXiv:2306.13394, 2023. 1 +[25] Zhangwei Gao, Zhe Chen, Erfei Cui, Yiming Ren, Weiyun Wang, Jinguo Zhu, Hao Tian, Shenglong Ye, Junjun He, Xizhou Zhu, et al. Mini-internvl: A flexible-transfer pocket multimodal model with $5\%$ parameters and $90\%$ performance. arXiv preprint arXiv:2410.16261, 2024. 2 +[26] Chi Han, Qifan Wang, Wenhan Xiong, Yu Chen, Heng Ji, and Sinong Wang. Lm-infinite: Simple on-the-fly length generalization for large language models. arXiv preprint arXiv:2308.16137, 2023. 3 +[27] Anwen Hu, Haiyang Xu, Liang Zhang, Jiabo Ye, Ming Yan, Ji Zhang, Qin Jin, Fei Huang, and Jingren Zhou. mplug-docowl2: High-resolution compressing for ocr-free multi-page document understanding. arXiv preprint arXiv:2409.03420, 2024. 7, 2 +[28] Mingxin Huang, Yuliang Liu, Dingkang Liang, Lianwen Jin, and Xiang Bai. Mini-monkey: Alleviate the saw-tooth effect by multi-scale adaptive cropping. arXiv preprint arXiv:2408.02034, 2024. 7 +[29] Gabriel Ilharco, Mitchell Wortsman, Ross Wightman, Cade Gordon, Nicholas Carlini, Rohan Taori, Achal Dave, Vaishaal Shankar, Hongseok Namkoong, John Miller, Hannaneh Hajishirzi, Ali Farhadi, and Ludwig Schmidt. Openclip. Zenodo. Version 0.1. https://doi.org/10.5281/zenodo.5143773, 2021. DOI: 10.5281/zenodo.5143773. 2 +[30] Yao Jiang, Xinyu Yan, Ge-Peng Ji, Keren Fu, Meijun Sun, Huan Xiong, Deng-Ping Fan, and Fahad Shahbaz Khan. Effectiveness assessment of recent large vision-language models. Visual Intelligence, 2(1):17, 2024. 1 +[31] Greg Kamradt. Llmtest_needleinahaystack. https://github.com/gkamradt/LLMTest_NeedleInAHaystack, 2024. Accessed: 2024-11-11. 1 +[32] Geewook Kim, Teakgyu Hong, Moonbin Yim, JeongYeon Nam, Jinyoung Park, Jinyeong Yim, Wonseok Hwang, Sangdoo Yun, Dongyoon Han, and Seunghyun Park. Ocr-free document understanding transformer. In European Conference on Computer Vision, pages 498-517. Springer, 2022. 3 +[33] Hugo Laurencon, Andres Marafioti, Victor Sanh, and Léo Tronchon. Building and better understanding vision-language models: insights and future directions. arXiv preprint arXiv:2408.12637, 2024. 2, 6 +[34] Hugo Laurençon, Lucile Saulnier, Léo Tronchon, Stas Bekman, Amanpreet Singh, Anton Lozhkov, Thomas Wang, Siddharth Karamcheti, Alexander Rush, Douwe Kiela, et al. Obelics: An open web-scale filtered dataset of interleaved image-text documents. NIPS, 36, 2024. 2 +[35] Bohao Li, Rui Wang, Guangzhi Wang, Yuying Ge, Yixiao Ge, and Ying Shan. Seed-bench: Benchmarking multimodal llms with generative comprehension. arXiv preprint arXiv:2307.16125, 2023. 2 + +[36] Bo Li, Yuanhan Zhang, Dong Guo, Renrui Zhang, Feng Li, Hao Zhang, Kaichen Zhang, Yanwei Li, Ziwei Liu, and Chunyuan Li. Llava-onevision: Easy visual task transfer. arXiv preprint arXiv:2408.03326, 2024. 1, 7 +[37] Junnan Li, Dongxu Li, Caiming Xiong, and Steven Hoi. Blip: Bootstrapping language-image pre-training for unified vision-language understanding and generation. In ICML, pages 12888-12900, 2022. 2 +[38] Junnan Li, Dongxu Li, Silvio Savarese, and Steven Hoi. Blip-2: Bootstrapping language-image pre-training with frozen image encoders and large language models. In ICML, pages 19730–19742. PMLR, 2023. 2 +[39] Qingyun Li, Zhe Chen, Weiyun Wang, Wenhai Wang, Shenglong Ye, Zhenjiang Jin, Guanzhou Chen, Yinan He, Zhangwei Gao, Erfei Cui, et al. Omnicorpus: An unified multimodal corpus of 10 billion-level images interleaved with text. arXiv preprint arXiv:2406.08418, 2024. 2 +[40] Zhang Li, Biao Yang, Qiang Liu, Zhiyin Ma, Shuo Zhang, Jingxu Yang, Yabo Sun, Yuliang Liu, and Xiang Bai. Monkey: Image resolution and text label are important things for large multi-modal models. arXiv preprint arXiv:2311.06607, 2023. 7 +[41] Haotian Liu, Chunyuan Li, Yuheng Li, and Yong Jae Lee. Improved baselines with visual instruction tuning. arXiv preprint arXiv:2310.03744, 2023. 1, 6 +[42] Haotian Liu, Chunyuan Li, Qingyang Wu, and Yong Jae Lee. Visual instruction tuning. NeurIPS, 36, 2023. 1, 2, 6 +[43] Hao Liu, Matei Zaharia, and Pieter Abbeel. Ring attention with blockwise transformers for near-infinite context. arXiv preprint arXiv:2310.01889, 2023. 2, 3, 6 +[44] Yuan Liu, Haodong Duan, Yuanhan Zhang, Bo Li, Songyang Zhang, Wangbo Zhao, Yike Yuan, Jiaqi Wang, Conghui He, Ziwei Liu, et al. Mmbench: Is your multi-modal model an all-around player? arXiv preprint arXiv:2307.06281, 2023. 1 +[45] Yuliang Liu, Zhang Li, Hongliang Li, Wenwen Yu, Mingxin Huang, Dezhi Peng, Mingyu Liu, Mingrui Chen, Chunyuan Li, Lianwen Jin, et al. On the hidden mystery ofOCR in large multimodal models. arXiv preprint arXiv:2305.07895, 2023. 1 +[46] Yuliang Liu, Biao Yang, Qiang Liu, Zhang Li, Zhiyin Ma, Shuo Zhang, and Xiang Bai. Textmonkey: AnOCR-free large multimodal model for understanding document. arXiv preprint arXiv:2403.04473, 2024. 7 +[47] Zhaoyang Liu, Yinan He, Wenhai Wang, Weiyun Wang, Yi Wang, Shoufa Chen, Qinglong Zhang, Zeqiang Lai, Yang Yang, Qingyun Li, Jiashuo Yu, et al. Interrupt: Solving vision-centric tasks by interacting with chatgpt beyond language. arXiv preprint arXiv:2305.05662, 2023. 2 +[48] Ziyu Liu, Tao Chu, Yuhang Zang, Xilin Wei, Xiaoyi Dong, Pan Zhang, Zijian Liang, Yuanjun Xiong, Yu Qiao, Dahua Lin, et al. Mmdu: A multi-turn multi-image dialog understanding benchmark and instruction-tuning dataset for lvlms. arXiv preprint arXiv:2406.11833, 2024. 6 +[49] Chuwei Luo, Yufan Shen, Zhaoqing Zhu, Qi Zheng, Zhi Yu, and Cong Yao. Layout: Layout instruction tuning with large language models for document understanding. In Pro + +ceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 15630-15640, 2024. 3 +[50] Xueguang Ma, Sheng-Chieh Lin, Minghan Li, Wenhu Chen, and Jimmy Lin. Unifying multimodal retrieval via document screenshot embedding. arXiv preprint arXiv:2406.11251, 2024. 1 +[51] Yubo Ma, Yuhang Zang, Liangyu Chen, Meiqi Chen, Yizhu Jiao, Xinze Li, Xinyuan Lu, Ziyu Liu, Yan Ma, Xiaoyi Dong, Pan Zhang, Liangming Pan, Yu-Gang Jiang, Jiaqi Wang, Yixin Cao, and Aixin Sun. Mmlongbench-doc: Benchmarking long-context document understanding with visualizations, 2024. 1, 7, 8, 5 +[52] Ahmed Masry, Xuan Long Do, Jia Qing Tan, Shafiq Joty, and Enamul Hoque. Chartqa: A benchmark for question answering about charts with visual and logical reasoning. In ACL, pages 2263-2279, 2022. 6, 8, 1, 5 +[53] Minesh Mathew, Dimosthenis Karatzas, and CV Jawahar. Docvqa: A dataset for vqa on document images. In WACV, pages 2200-2209, 2021. 6, 8, 1, 5 +[54] Minesh Mathew, Viraj Bagal, Ruben Tito, Dimosthenis Karatzas, Ernest Valveny, and CV Jawahar. Infographicvqa. In WACV, pages 1697-1706, 2022. 6, 8, 1, 5 +[55] Anand Mishra, Shashank Shekhar, Ajeet Kumar Singh, and Anirban Chakraborty. Ocr-vqa: Visual question answering by reading text in images. In ICDAR, pages 947-952, 2019. 6 +[56] OpenAI. Gpt-4v(ison) system card. https://cdn.openai.com/papers/GPTV_System/Card.pdf, 2023.1,4,7 +[57] OpenAI. Gpt-4o system card. https://openai.com/index/gpt-4o-system-card/, 2024.7 +[58] Ofir Press, Noah A Smith, and Mike Lewis. Train short, test long: Attention with linear biases enables input length extrapolation. arXiv preprint arXiv:2108.12409, 2021. 3 +[59] Sylvestre-Alvise Rebuffi, Hakan Bilen, and Andrea Vedaldi. Learning multiple visual domains with residual adapters. NeurIPS, 30, 2017. 2 +[60] Machel Reid, Nikolay Savinov, Denis Teptyashin, Dmitry Lepikhin, Timothy Lillicrap, Jean-baptiste Alayrac, Radu Soricut, Angeliki Lazaridou, Orhan Firat, Julian Schrittwieser, et al. Gemini 1.5: Unlocking multimodal understanding across millions of tokens of context. arXiv preprint arXiv:2403.05530, 2024. 1, 7 +[61] Sahel Sharifymoghaddam, Shivani Upadhyay, Wenhu Chen, and Jimmy Lin. Unirag: Universal retrieval augmentation for multi-modal large language models. arXiv preprint arXiv:2405.10311, 2024. 1 +[62] Min Shi, Fuxiao Liu, Shihao Wang, Shijia Liao, Subhashree Radhakrishnan, De-An Huang, Hongxu Yin, Karan Sapra, Yaser Yacoob, Humphrey Shi, et al. Eagle: Exploring the design space for multimodal llms with mixture of encoders. arXiv preprint arXiv:2408.15998, 2024. 1 +[63] Weijia Shi, Sewon Min, Michihiro Yasunaga, Minjoon Seo, Rich James, Mike Lewis, Luke Zettlemoyer, and Wen-tau Yih. Replug: Retrieval-augmented black-box language models. arXiv preprint arXiv:2301.12652, 2023. 1 + +[64] Oleksii Sidorov, Ronghang Hu, Marcus Rohrbach, and Amanpreet Singh. Textcaps: a dataset for image captioning with reading comprehension. In ECCV, pages 742-758, 2020. 6 +[65] Amanpreet Singh, Vivek Natarajan, Meet Shah, Yu Jiang, Xinlei Chen, Dhruv Batra, Devi Parikh, and Marcus Rohrbach. Towards vqa models that can read. In CVPR, pages 8317-8326, 2019. 1, 6 +[66] Jianlin Su, Murtadha Ahmed, Yu Lu, Shengfeng Pan, Wen Bo, and Yunfeng Liu. Roformer: Enhanced transformer with rotary position embedding. Neurocomputing, 568:127063, 2024. 3 +[67] Quan Sun, Qiying Yu, Yufeng Cui, Fan Zhang, Xiaosong Zhang, Yueze Wang, Hongcheng Gao, Jingjing Liu, Tiejun Huang, and Xinlong Wang. Generative pretraining in multimodality. In ICLR, 2024. 2 +[68] Yutao Sun, Li Dong, Barun Patra, Shuming Ma, Shaohan Huang, Alon Benhaim, Vishrav Chaudhary, Xia Song, and Furu Wei. A length-extrapolatable transformer. arXiv preprint arXiv:2212.10554, 2022. 3 +[69] Mohammad Reza Taesiri. Arxivqa. https://github.com/taesiri/ArXivQA, 2024.6 +[70] Ryota Tanaka, Kyosuke Nishida, and Sen Yoshida. Visualmrc: Machine reading comprehension on document images. In Proceedings of the AAAI Conference on Artificial Intelligence, pages 13878-13888, 2021. 6 +[71] Zineng Tang, Ziyi Yang, Guoxin Wang, Yuwei Fang, Yang Liu, Chenguang Zhu, Michael Zeng, Cha Zhang, and Mohit Bansal. Unifying vision, text, and layout for universal document processing. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 19254-19264, 2023. 2 +[72] Gemini Team, Rohan Anil, Sebastian Borgeaud, Yonghui Wu, Jean-Baptiste Alayrac, Jiahui Yu, Radu Soricut, Johan Schalkwyk, Andrew M Dai, Anja Hauth, et al. Gemini: a family of highly capable multimodal models. arXiv preprint arXiv:2312.11805, 2023. 1 +[73] OpenGVLab Team. Internv12: Better than the best—expanding performance boundaries of open-source multimodal models with the progressive scaling strategy, 2024. 1, 6, 7, 8 +[74] Qwen Team. Qwen2.5: A party of foundation models, 2024. 1 +[75] Changyao Tian, Xizhou Zhu, Yuwen Xiong, Weiyun Wang, Zhe Chen, Wenhai Wang, Yuntao Chen, Lewei Lu, Tong Lu, Jie Zhou, et al. Mm-interleaved: Interleaved image-text generative modeling via multi-modal feature synchronizer. arXiv preprint arXiv:2401.10208, 2024. 2 +[76] Rubén Tito, Dimosthenis Karatzas, and Ernest Valveny. Hierarchical multimodal transformers for multipage docvqa. Pattern Recognition, 144:109834, 2023. 1, 2, 7, 5 +[77] Xiaoguang Tu, Zhi He, Yi Huang, Zhi-Hao Zhang, Ming Yang, and Jian Zhao. An overview of large ai models and their applications. Visual Intelligence, 2(1):1-22, 2024. 1 +[78] Jordy Van Landeghem, Ruben Tito, Łukasz Borchmann, Michal Pietruszka, Paweł Joziak, Rafał Powalski, Dawid Jurkiewicz, Mickael Coustaty, Bertrand Ackaert, Ernest Val + +veny, et al. Document understanding dataset and evaluation (dude). In Proceedings IEEE/CVF international conference on computer vision-ICCV 2023, pages 19528-19540. IEEE/CVF, 2023. 2, 6 +[79] Bin Wang, Chao Xu, Xiaomeng Zhao, Linke Ouyang, Fan Wu, Zhiyuan Zhao, Rui Xu, Kaiwen Liu, Yuan Qu, Fukai Shang, et al. Mineru: An open-source solution for precise document content extraction. arXiv preprint arXiv:2409.18839, 2024. 3 +[80] Dongsheng Wang, Natraj Raman, Mathieu Sibue, Zhiqiang Ma, Petr Babkin, Simerjot Kaur, Yulong Pei, Armineh Nourbakhsh, and Xiaomo Liu. Docllm: A layout-aware generative language model for multimodal document understanding. arXiv preprint arXiv:2401.00908, 2023. 2 +[81] Peng Wang, Shuai Bai, Sinan Tan, Shijie Wang, Zhihao Fan, Jinze Bai, Keqin Chen, Xuejing Liu, Jialin Wang, Wenbin Ge, et al. Qwen2-vl: Enhancing vision-language model's perception of the world at any resolution. arXiv preprint arXiv:2409.12191, 2024. 1 +[82] Weihan Wang, Qingsong Lv, Wenmeng Yu, Wenyi Hong, Ji Qi, Yan Wang, Junhui Ji, Zhuoyi Yang, Lei Zhao, Xixuan Song, et al. Cogvlm: Visual expert for pretrained language models. arXiv preprint arXiv:2311.03079, 2023. 1 +[83] Weiyun Wang, Yiming Ren, Haowen Luo, Tiantong Li, Chenxiang Yan, Zhe Chen, Wenhai Wang, Qingyun Li, Lewei Lu, Xizhou Zhu, et al. The all-seeing project v2: Towards general relation comprehension of the open world. arXiv preprint arXiv:2402.19474, 2024. 1, 2 +[84] Weiyun Wang, Min Shi, Qingyun Li, Wenhai Wang, Zhenhang Huang, Linjie Xing, Zhe Chen, Hao Li, Xizhou Zhu, Zhiguo Cao, et al. The all-seeing project: Towards panoptic visual recognition and understanding of the open world. In ICLR, 2024. 1, 2 +[85] Weiyun Wang, Shuibio Zhang, Yiming Ren, Yuchen Duan, Tiantong Li, Shuo Liu, Mengkang Hu, Zhe Chen, Kaipeng Zhang, Lewei Lu, Xizhou Zhu, Ping Luo, Yu Qiao, Jifeng Dai, Wenqi Shao, and Wenhai Wang. Needle in a multimodal haystack. arXiv preprint arXiv:2406.07230, 2024. 7, 8, 3 +[86] Weiyun Wang, Shuibio Zhang, Yiming Ren, Yuchen Duan, Tiantong Li, Shuo Liu, Mengkang Hu, Zhe Chen, Kaipeng Zhang, Lewei Lu, et al. Needle in a multimodal haystack. arXiv preprint arXiv:2406.07230, 2024. 2, 7, 8, 5 +[87] Yonghui Wang, Wengang Zhou, Hao Feng, Keyi Zhou, and Houqiang Li. Towards improving document understanding: An exploration on text-grounding via mllms. arXiv preprint arXiv:2311.13194, 2023. 3 +[88] Haoran Wei, Lingyu Kong, Jinyue Chen, Liang Zhao, Zheng Ge, Jinrong Yang, Jianjian Sun, Chunrui Han, and Xiangyu Zhang. Vary: Scaling up the vision vocabulary for large vision-language models. arXiv preprint arXiv:2312.06109, 2023. 3 +[89] Renqiu Xia, Song Mao, Xiangchao Yan, Hongbin Zhou, Bo Zhang, Haoyang Peng, Jiahao Pi, Daocheng Fu, Wenjie Wu, Hancheng Ye, et al. Docgenome: An open largescale scientific document benchmark for training and testing multi-modal large language models. arXiv preprint arXiv:2406.11633, 2024. 1, 7, 2, 5 + +[90] Guangxuan Xiao, Yuandong Tian, Beidi Chen, Song Han, and Mike Lewis. Efficient streaming language models with attention sinks. arXiv preprint arXiv:2309.17453, 2023. 3 +[91] Jianxin Yang. Longqlora: Efficient and effective method to extend context length of large language models. arXiv preprint arXiv:2311.04879, 2023. 2, 6 +[92] Yuan Yao, Tianyu Yu, Ao Zhang, Chongyi Wang, Junbo Cui, Hongji Zhu, Tianchi Cai, Haoyu Li, Weilin Zhao, Zhihui He, et al. Minicpm-v: A gpt-4v level mllm on your phone. arXiv preprint arXiv:2408.01800, 2024. 1, 7, 8 +[93] Jiabo Ye, Anwen Hu, Haiyang Xu, Qinghao Ye, Ming Yan, Yuhao Dan, Chenlin Zhao, Guohai Xu, Chenliang Li, Junfeng Tian, et al. mplug-docowl: Modularized multimodal large language model for document understanding. arXiv preprint arXiv:2307.02499, 2023. 6 +[94] Jiabo Ye, Anwen Hu, Haiyang Xu, Qinghao Ye, Ming Yan, Guohai Xu, Chenliang Li, Junfeng Tian, Qi Qian, Ji Zhang, et al. Ureader: UniversalOCR-free visually-situated language understanding with multimodal large language model. arXiv preprint arXiv:2310.05126, 2023. 3 +[95] Shi Yu, Chaoyue Tang, Bokai Xu, Junbo Cui, Junhao Ran, Yukun Yan, Zhenghao Liu, Shuo Wang, Xu Han, Zhiyuan Liu, et al. Visrag: Vision-based retrieval-augmented generation on multi-modality documents. arXiv preprint arXiv:2410.10594, 2024. 1, 2, 3, 7, 8 +[96] Jiajie Zhang, Yushi Bai, Xin Lv, Wanjun Gu, Danqing Liu, Minhao Zou, Shulin Cao, Lei Hou, Yuxiao Dong, Ling Feng, et al. Longcite: Enabling llms to generate fine-grained citations in long-context qa. arXiv e-prints, pages arXiv-2409, 2024. 2, 6 +[97] Jiajie Zhang, Zhongni Hou, Xin Lv, Shulin Cao, Zhenyu Hou, Yilin Niu, Lei Hou, Yuxiao Dong, Ling Feng, and Juanzi Li. Longreward: Improving long-context large language models with ai feedback. arXiv preprint arXiv:2410.21252, 2024. 2, 6 +[98] Pan Zhang, Xiaoyi Dong, Yuhang Zang, Yuhang Cao, Rui Qian, Lin Chen, Qipeng Guo, Haodong Duan, Bin Wang, Linke Ouyang, et al. Internlm-xcomposer-2.5: A versatile large vision language model supporting long-contextual input and output. arXiv preprint arXiv:2407.03320, 2024. 7 +[99] Wanrong Zhu, Jack Hessel, Anas Awadalla, Samir Yitzhak Gadre, Jesse Dodge, Alex Fang, Youngjae Yu, Ludwig Schmidt, William Yang Wang, and Yejin Choi. Multimodal c4: An open, billion-scale corpus of images interleaved with text. NIPS, 36, 2024. 2 \ No newline at end of file diff --git a/2025/Docopilot_ Improving Multimodal Models for Document-Level Understanding/images.zip b/2025/Docopilot_ Improving Multimodal Models for Document-Level Understanding/images.zip new file mode 100644 index 0000000000000000000000000000000000000000..ade167ed93cb11701dfb29bac419da9be62b356b --- /dev/null +++ b/2025/Docopilot_ Improving Multimodal Models for Document-Level Understanding/images.zip @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:9c9cb55eee856625f0934466a4beee4f05ae9bf8b37c0d350f326cce209e49ec +size 578644 diff --git a/2025/Docopilot_ Improving Multimodal Models for Document-Level Understanding/layout.json b/2025/Docopilot_ Improving Multimodal Models for Document-Level Understanding/layout.json new file mode 100644 index 0000000000000000000000000000000000000000..8af3f67b0d9ef967ea8ddff36f7b3ae8da1ed71f --- /dev/null +++ b/2025/Docopilot_ Improving Multimodal Models for Document-Level Understanding/layout.json @@ -0,0 +1,9743 @@ +{ + "pdf_info": [ + { + "para_blocks": [ + { + "bbox": [ + 64, + 103, + 547, + 123 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 64, + 103, + 547, + 123 + ], + "spans": [ + { + "bbox": [ + 64, + 103, + 547, + 123 + ], + "type": "text", + "content": "Docopilot: Improving Multimodal Models for Document-Level Understanding" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 96, + 143, + 515, + 172 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 96, + 143, + 515, + 172 + ], + "spans": [ + { + "bbox": [ + 96, + 143, + 515, + 172 + ], + "type": "text", + "content": "Yuchen Duan" + }, + { + "bbox": [ + 96, + 143, + 515, + 172 + ], + "type": "inline_equation", + "content": "^{1,2*}" + }, + { + "bbox": [ + 96, + 143, + 515, + 172 + ], + "type": "text", + "content": ", Zhe Chen" + }, + { + "bbox": [ + 96, + 143, + 515, + 172 + ], + "type": "inline_equation", + "content": "^{3,1*}" + }, + { + "bbox": [ + 96, + 143, + 515, + 172 + ], + "type": "text", + "content": ", Yusong Hu" + }, + { + "bbox": [ + 96, + 143, + 515, + 172 + ], + "type": "inline_equation", + "content": "^{4,1*}" + }, + { + "bbox": [ + 96, + 143, + 515, + 172 + ], + "type": "text", + "content": ", Weiyun Wang" + }, + { + "bbox": [ + 96, + 143, + 515, + 172 + ], + "type": "inline_equation", + "content": "^{*5,1}" + }, + { + "bbox": [ + 96, + 143, + 515, + 172 + ], + "type": "text", + "content": ", Shenglong Ye" + }, + { + "bbox": [ + 96, + 143, + 515, + 172 + ], + "type": "inline_equation", + "content": "^{1}" + }, + { + "bbox": [ + 96, + 143, + 515, + 172 + ], + "type": "text", + "content": ", Botian Shi" + }, + { + "bbox": [ + 96, + 143, + 515, + 172 + ], + "type": "inline_equation", + "content": "^{1}" + }, + { + "bbox": [ + 96, + 143, + 515, + 172 + ], + "type": "text", + "content": ", Lewei Lu" + }, + { + "bbox": [ + 96, + 143, + 515, + 172 + ], + "type": "inline_equation", + "content": "^{7}" + }, + { + "bbox": [ + 96, + 143, + 515, + 172 + ], + "type": "text", + "content": ", Qibin Hou" + }, + { + "bbox": [ + 96, + 143, + 515, + 172 + ], + "type": "inline_equation", + "content": "^{4}" + }, + { + "bbox": [ + 96, + 143, + 515, + 172 + ], + "type": "text", + "content": ", Tong Lu" + }, + { + "bbox": [ + 96, + 143, + 515, + 172 + ], + "type": "inline_equation", + "content": "^{3,1}" + }, + { + "bbox": [ + 96, + 143, + 515, + 172 + ], + "type": "text", + "content": ", Hongsheng Li" + }, + { + "bbox": [ + 96, + 143, + 515, + 172 + ], + "type": "inline_equation", + "content": "^{2,1}" + }, + { + "bbox": [ + 96, + 143, + 515, + 172 + ], + "type": "text", + "content": ", Jifeng Dai" + }, + { + "bbox": [ + 96, + 143, + 515, + 172 + ], + "type": "inline_equation", + "content": "^{6,1}" + }, + { + "bbox": [ + 96, + 143, + 515, + 172 + ], + "type": "text", + "content": ", Wenhai Wang" + }, + { + "bbox": [ + 96, + 143, + 515, + 172 + ], + "type": "inline_equation", + "content": "^{2,1\\boxtimes}" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 113, + 175, + 499, + 205 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 113, + 175, + 499, + 205 + ], + "spans": [ + { + "bbox": [ + 113, + 175, + 499, + 205 + ], + "type": "inline_equation", + "content": "^{1}" + }, + { + "bbox": [ + 113, + 175, + 499, + 205 + ], + "type": "text", + "content": "Shanghai AI Laboratory, " + }, + { + "bbox": [ + 113, + 175, + 499, + 205 + ], + "type": "inline_equation", + "content": "^{2}" + }, + { + "bbox": [ + 113, + 175, + 499, + 205 + ], + "type": "text", + "content": "The Chinese University of Hong Kong, " + }, + { + "bbox": [ + 113, + 175, + 499, + 205 + ], + "type": "inline_equation", + "content": "^{3}" + }, + { + "bbox": [ + 113, + 175, + 499, + 205 + ], + "type": "text", + "content": "Nanjing University, " + }, + { + "bbox": [ + 113, + 175, + 499, + 205 + ], + "type": "inline_equation", + "content": "^{4}" + }, + { + "bbox": [ + 113, + 175, + 499, + 205 + ], + "type": "text", + "content": "Nankai University, " + }, + { + "bbox": [ + 113, + 175, + 499, + 205 + ], + "type": "inline_equation", + "content": "^{5}" + }, + { + "bbox": [ + 113, + 175, + 499, + 205 + ], + "type": "text", + "content": "Fudan University, " + }, + { + "bbox": [ + 113, + 175, + 499, + 205 + ], + "type": "inline_equation", + "content": "^{6}" + }, + { + "bbox": [ + 113, + 175, + 499, + 205 + ], + "type": "text", + "content": "Tsinghua University, " + }, + { + "bbox": [ + 113, + 175, + 499, + 205 + ], + "type": "inline_equation", + "content": "^{7}" + }, + { + "bbox": [ + 113, + 175, + 499, + 205 + ], + "type": "text", + "content": "SenseTime Research" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 151, + 232, + 200, + 245 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 151, + 232, + 200, + 245 + ], + "spans": [ + { + "bbox": [ + 151, + 232, + 200, + 245 + ], + "type": "text", + "content": "Abstract" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 54, + 257, + 297, + 521 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 54, + 257, + 297, + 521 + ], + "spans": [ + { + "bbox": [ + 54, + 257, + 297, + 521 + ], + "type": "text", + "content": "Despite significant progress in multimodal large language models (MLLMs), their performance on complex, multipage document comprehension remains inadequate, largely due to the lack of high-quality, document-level datasets. While current retrieval-augmented generation (RAG) methods offer partial solutions, they suffer from issues, such as fragmented retrieval contexts, multi-stage error accumulation, and extra time costs of retrieval. In this work, we present a high-quality document-level dataset, Doc750K, designed to support in-depth understanding of multimodal documents. This dataset includes diverse document structures, extensive cross-page dependencies, and real question-answer pairs derived from the original documents. Building on the dataset, we develop a native multimodal model—Docopilot, which can accurately handle document-level dependencies without relying on RAG. Experiments demonstrate that Docopilot achieves superior coherence, accuracy, and efficiency in document understanding tasks and multi-turn interactions, setting a new baseline for document-level multimodal understanding. Data, code, and models are released at https://github.com/OpenGVLab/Docopilot." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 55, + 534, + 137, + 547 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 534, + 137, + 547 + ], + "spans": [ + { + "bbox": [ + 55, + 534, + 137, + 547 + ], + "type": "text", + "content": "1. Introduction" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 55, + 554, + 296, + 687 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 554, + 296, + 687 + ], + "spans": [ + { + "bbox": [ + 55, + 554, + 296, + 687 + ], + "type": "text", + "content": "In recent years, multimodal large language models (MLLMs) [6, 12, 41, 56, 60, 72, 81, 82, 84, 92] have rapidly developed, achieving remarkable performance in various visual understanding tasks [30, 77], particularly image-level tasks, such as image captioning [10, 42], optical character recognition (OCR) [45, 65], and visual question answering (VQA) [24, 44]. Despite these advances, current MLLMs still face significant challenges in document-level understanding [51, 76, 89], where models are required to identify and integrate key information across multi-page documents, setting high expectations for their long-context processing" + } + ] + } + ], + "index": 8 + }, + { + "type": "image", + "bbox": [ + 315, + 220, + 555, + 376 + ], + "blocks": [ + { + "bbox": [ + 315, + 220, + 555, + 376 + ], + "lines": [ + { + "bbox": [ + 315, + 220, + 555, + 376 + ], + "spans": [ + { + "bbox": [ + 315, + 220, + 555, + 376 + ], + "type": "image", + "image_path": "ca45d626bd25fc9f7b1548f606b52489b79335623c8f6b671344188d240b28ca.jpg" + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 313, + 384, + 556, + 483 + ], + "lines": [ + { + "bbox": [ + 313, + 384, + 556, + 483 + ], + "spans": [ + { + "bbox": [ + 313, + 384, + 556, + 483 + ], + "type": "text", + "content": "Figure 1. Accuracy " + }, + { + "bbox": [ + 313, + 384, + 556, + 483 + ], + "type": "inline_equation", + "content": "\\nu.s" + }, + { + "bbox": [ + 313, + 384, + 556, + 483 + ], + "type": "text", + "content": " inference latency on MM-NIAH. The proposed Docopilot-8B shows a notable improvement over baseline models [73], achieving a " + }, + { + "bbox": [ + 313, + 384, + 556, + 483 + ], + "type": "inline_equation", + "content": "+19.9\\%" + }, + { + "bbox": [ + 313, + 384, + 556, + 483 + ], + "type": "text", + "content": " accuracy gain compared to InternVL2-8B and surpassing InternVL2-26B with less than " + }, + { + "bbox": [ + 313, + 384, + 556, + 483 + ], + "type": "inline_equation", + "content": "31\\%" + }, + { + "bbox": [ + 313, + 384, + 556, + 483 + ], + "type": "text", + "content": " of the inference latency. Additionally, Docopilot-2B uses fewer parameters (less than " + }, + { + "bbox": [ + 313, + 384, + 556, + 483 + ], + "type": "inline_equation", + "content": "10\\%" + }, + { + "bbox": [ + 313, + 384, + 556, + 483 + ], + "type": "text", + "content": ") while exhibiting comparable performance to the " + }, + { + "bbox": [ + 313, + 384, + 556, + 483 + ], + "type": "inline_equation", + "content": "10\\times" + }, + { + "bbox": [ + 313, + 384, + 556, + 483 + ], + "type": "text", + "content": " larger InternVL2-26B. These results suggest that our Docopilot strikes a reasonable balance between latency, model size, and performance." + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "image_caption" + } + ], + "index": 9 + }, + { + "bbox": [ + 313, + 497, + 411, + 509 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 497, + 411, + 509 + ], + "spans": [ + { + "bbox": [ + 313, + 497, + 411, + 509 + ], + "type": "text", + "content": "capabilities of MLLMs." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 313, + 510, + 556, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 510, + 556, + 715 + ], + "spans": [ + { + "bbox": [ + 313, + 510, + 556, + 715 + ], + "type": "text", + "content": "Current research on long-content understanding primarily focuses on text-only models [5, 9, 74], targeting specific retrieval tasks such as \"Needle in a Haystack\" (NIAH) [2, 31]. However, existing open-source MLLMs [12, 13, 36, 41, 62, 81, 83] are primarily trained on image-level data, lacking the long-context understanding capacity required for document-level understanding. Retrieval-augmented generation (RAG) methods [14, 22, 50, 61, 63, 95] attempt to address this by retrieving key information to fit within the limited context windows of MLLMs, but they still encounter the following challenges in document-level tasks. (1) Fragmented Retrieval Contexts. Retrieved information is fragmented, lacking the overall structure of the document; (2) Multi-Stage Error Accumulation. Incorrect retrieval results can affect subsequent responses, leading to errors or omissions of critical details, especially in multi-turn or complex tasks; (3) Extra Time Costs. The retrieval" + } + ] + } + ], + "index": 12 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "spans": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "type": "text", + "content": "CVF" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 145, + 0, + 494, + 37 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 145, + 0, + 494, + 37 + ], + "spans": [ + { + "bbox": [ + 145, + 0, + 494, + 37 + ], + "type": "text", + "content": "This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 67, + 693, + 141, + 703 + ], + "type": "page_footnote", + "angle": 0, + "lines": [ + { + "bbox": [ + 67, + 693, + 141, + 703 + ], + "spans": [ + { + "bbox": [ + 67, + 693, + 141, + 703 + ], + "type": "text", + "content": "* Equal contribution;" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 66, + 704, + 237, + 713 + ], + "type": "page_footnote", + "angle": 0, + "lines": [ + { + "bbox": [ + 66, + 704, + 237, + 713 + ], + "spans": [ + { + "bbox": [ + 66, + 704, + 237, + 713 + ], + "type": "text", + "content": "Corresponding author: wangwenhai@pjlab.org.cn" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 295, + 748, + 316, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 316, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 316, + 757 + ], + "type": "text", + "content": "4026" + } + ] + } + ], + "index": 16 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 0 + }, + { + "para_blocks": [ + { + "bbox": [ + 55, + 72, + 294, + 95 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 72, + 294, + 95 + ], + "spans": [ + { + "bbox": [ + 55, + 72, + 294, + 95 + ], + "type": "text", + "content": "step increases the latency of the QA system, limiting the scalability of RAG in time-sensitive scenarios." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 54, + 97, + 295, + 252 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 54, + 97, + 295, + 252 + ], + "spans": [ + { + "bbox": [ + 54, + 97, + 295, + 252 + ], + "type": "text", + "content": "To address these problems, two primary challenges need to be considered. (1) High-Quality Multimodal Document Dataset. While extensive datasets [11, 91, 96, 97] exist for long-context, text-only tasks, high-quality document-level question-answering datasets remain scarce. This shortage is largely attributed to the high costs associated with annotation and the lack of streamlined construction pipelines. (2) Native Document-Level MLLMs. Although RAG-based methods [14, 86, 95] provide some relief, native multimodal models with long-context processing abilities are crucial. However, training native MLLMs specifically for document-level understanding is constrained by current hardware limitations." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 56, + 255, + 296, + 456 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 255, + 296, + 456 + ], + "spans": [ + { + "bbox": [ + 56, + 255, + 296, + 456 + ], + "type": "text", + "content": "In this work, we introduce a new multimodal document dataset that supports document-level understanding tasks. Compared to counterparts [33, 76, 78], this dataset has the following features: (1) Large Scale. It includes a total of 758K question-answer samples, containing 5.2B text tokens and 3.1M images. It encompasses content from various sources, such as Sci-Hub, Arxiv, and OpenReview, covering a wide range of topics and document layouts. (2) High Quality. Unlike existing datasets that insert irrelevant questions into documents, we collect real, in-depth question-answer pairs and construct single-page and crosspage questions based on document structure. Such high-quality question-answer data accounts for " + }, + { + "bbox": [ + 56, + 255, + 296, + 456 + ], + "type": "inline_equation", + "content": "31.6\\%" + }, + { + "bbox": [ + 56, + 255, + 296, + 456 + ], + "type": "text", + "content": " of the dataset. (3) Multimodal. For the document content, we provide not only the conventional interleaved text-image context but also purely rendered image inputs, catering to the needs of different models." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 55, + 460, + 295, + 638 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 460, + 295, + 638 + ], + "spans": [ + { + "bbox": [ + 55, + 460, + 295, + 638 + ], + "type": "text", + "content": "Building upon this dataset, we developed a native baseline model for document-level multimodal understanding-Docopilot. Unlike existing approaches [14, 86, 95] that rely on RAG, our model achieves efficient document-level training and testing through simple engineering optimizations, such as multimodal data packing, Ring Attention [43], and Liger Kernel [16]. Leveraging the proxy tasks carefully designed within the dataset, Docopilot can directly handle long-distance dependencies and cross-page information integration without external retrieval support. As shown in Figure 1, this approach not only enhances coherence and accuracy compared to RAG methods but also significantly reduces the response time of the entire question-answering system, delivering superior real-time performance in multi-turn interactions." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 67, + 640, + 276, + 651 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 67, + 640, + 276, + 651 + ], + "spans": [ + { + "bbox": [ + 67, + 640, + 276, + 651 + ], + "type": "text", + "content": "The main contributions are summarized as follows:" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 55, + 654, + 295, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 654, + 295, + 713 + ], + "spans": [ + { + "bbox": [ + 55, + 654, + 295, + 713 + ], + "type": "text", + "content": "(1) We develop the first large-scale, high-quality dataset for document-level multimodal understanding, consisting of 758K QA pairs from 3 sources, supporting 9 types of proxy tasks. This dataset includes " + }, + { + "bbox": [ + 55, + 654, + 295, + 713 + ], + "type": "inline_equation", + "content": "31.6\\%" + }, + { + "bbox": [ + 55, + 654, + 295, + 713 + ], + "type": "text", + "content": " real QA pairs directly extracted from documents." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 313, + 72, + 555, + 240 + ], + "type": "list", + "angle": 0, + "index": 8, + "blocks": [ + { + "bbox": [ + 313, + 72, + 553, + 132 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 72, + 553, + 132 + ], + "spans": [ + { + "bbox": [ + 313, + 72, + 553, + 132 + ], + "type": "text", + "content": "(2) Based on the dataset, we implement Docopilot, a native MLLM designed for document-level understanding without relying on retrieval mechanisms. This approach greatly improves its ability to integrate and comprehend information across multi-page documents." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 313, + 133, + 555, + 240 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 133, + 555, + 240 + ], + "spans": [ + { + "bbox": [ + 313, + 133, + 555, + 240 + ], + "type": "text", + "content": "(3) Through extensive experiments on multiple document-level benchmarks, our method demonstrates performance significantly superior to existing approaches, proving its effectiveness and generality. As shown in Figure 1, Docopilot-8B achieves a score of 61.8 on MM-NIAH [86], outperforming InternVL2-8B by 19.9 points and surpassing InternVL2-26B with less than " + }, + { + "bbox": [ + 313, + 133, + 555, + 240 + ], + "type": "inline_equation", + "content": "31\\%" + }, + { + "bbox": [ + 313, + 133, + 555, + 240 + ], + "type": "text", + "content": " of the latency. We hope this work could provide a baseline for future advancements in MLLMs for document-level tasks." + } + ] + } + ], + "index": 7 + } + ], + "sub_type": "text" + }, + { + "bbox": [ + 314, + 253, + 400, + 266 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 314, + 253, + 400, + 266 + ], + "spans": [ + { + "bbox": [ + 314, + 253, + 400, + 266 + ], + "type": "text", + "content": "2. Related Work" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 313, + 273, + 509, + 286 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 273, + 509, + 286 + ], + "spans": [ + { + "bbox": [ + 313, + 273, + 509, + 286 + ], + "type": "text", + "content": "2.1. Multimodal Large Language Models" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 313, + 291, + 555, + 579 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 291, + 555, + 579 + ], + "spans": [ + { + "bbox": [ + 313, + 291, + 555, + 579 + ], + "type": "text", + "content": "Multimodal large language models (MLLMs) have demonstrated impressive capabilities in processing image and text information, opening up new directions for applications such as visual question answering and image captioning. Early models [12, 29, 37, 59] trained with contrastive learning methods excelled in recognizing and understanding open-world semantics within an image-text matching framework. However, their limited generative abilities restricted their applicability. To leverage the powerful generation abilities of large language models (LLMs), subsequent works [13, 38, 42, 47, 83, 84] introduced a connector to align the embedding spaces of vision encoders and LLMs, allowing encoded image embeddings to serve as soft prompts for LLMs. Another series of works [1, 34, 39, 99] extended LLMs by integrating additional visual experts, reducing reliance on standalone vision encoders. More recently, models capable of both understanding and generating images have also made notable progress [20, 35, 67, 75], leveraging the insight that image generation can enhance image understanding. Despite these advancements, current MLLMs still face challenges with long-context multimodal inputs. For instance, InternVL 2.0 [13, 25] performs optimally within a token range of up to 8192, constraining its effectiveness in document-level applications." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 313, + 588, + 495, + 601 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 588, + 495, + 601 + ], + "spans": [ + { + "bbox": [ + 313, + 588, + 495, + 601 + ], + "type": "text", + "content": "2.2. Document Understanding Models" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 313, + 605, + 555, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 605, + 555, + 713 + ], + "spans": [ + { + "bbox": [ + 313, + 605, + 555, + 713 + ], + "type": "text", + "content": "Extracting key information from documents is crucial for industries and academic research. OCR-model-driven methods [3, 4, 71, 80] represent one of the primary technical approaches. These methods extract text, layout, and bounding box information from external systems and integrate it with another model. However, they are prone to error propagation and high processing times due to their reliance on multiple components. Benefitting from the rapid advancements in LLMs, OCR-free methods have also achieved" + } + ] + } + ], + "index": 13 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "text", + "content": "4027" + } + ] + } + ], + "index": 14 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 1 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 56, + 68, + 553, + 244 + ], + "blocks": [ + { + "bbox": [ + 56, + 68, + 553, + 244 + ], + "lines": [ + { + "bbox": [ + 56, + 68, + 553, + 244 + ], + "spans": [ + { + "bbox": [ + 56, + 68, + 553, + 244 + ], + "type": "image", + "image_path": "18b2b2256ee8d2ed25df211f25d986c42c96fd2728e54c3e98f1529a5a7d8047.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 55, + 251, + 555, + 298 + ], + "lines": [ + { + "bbox": [ + 55, + 251, + 555, + 298 + ], + "spans": [ + { + "bbox": [ + 55, + 251, + 555, + 298 + ], + "type": "text", + "content": "Figure 2. Multimodal document dataset generation pipeline. This pipeline involves three main stages: (1) Raw Data Collection: Documents are gathered from sources like Sci-Hub, arXiv, and OpenReview, available in PDF and HTML formats. (2) Document Content Extraction: Multimodal content is processed in two formats: interleaved text-image format and multi-image format. (3) Question-Answer Pairs Construction: QA pairs are generated based on the document structure or constructed using GPT-4o." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "bbox": [ + 55, + 312, + 296, + 384 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 312, + 296, + 384 + ], + "spans": [ + { + "bbox": [ + 55, + 312, + 296, + 384 + ], + "type": "text", + "content": "great progress. Donut [32] is the first end-to-end training framework based on a Transformer without requiring OCR engines or APIs. Subsequent works [23, 49, 87, 88, 94] propose diverse modifications in model architectures and training algorithms. However, these models are designed for specific tasks and lack general abilities." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 55, + 396, + 261, + 408 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 396, + 261, + 408 + ], + "spans": [ + { + "bbox": [ + 55, + 396, + 261, + 408 + ], + "type": "text", + "content": "2.3. Long-Context Large Language Models" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 55, + 415, + 296, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 415, + 296, + 715 + ], + "spans": [ + { + "bbox": [ + 55, + 415, + 296, + 715 + ], + "type": "text", + "content": "With advancements in engineering, architecture, and algorithms, long-context large language models have made substantial progress. Techniques such as Flash Attention [17, 18] and Ring Attention [43] have notably reduced GPU memory usage for training on extended contexts. Additionally, various sparse attention mechanisms—including Shifting Sparse Attention [11], Dilated Attention [19], and Attention Sinks [26, 90]—have enabled efficient scaling to handle larger contexts. New positional embedding methods, like ALiBi [58], xPOS [68], and RoPE [66], further enhance the models' generalization capabilities in length extrapolation. However, these advancements remain largely confined to natural language processing, and methods to extend the context size of MLLMs are still under-explored. Another research approach aims to reduce context size by leveraging retrieval augmented generation (RAG) [14, 22, 95], where only the most relevant passages are retrieved and fed into the generation model. However, this retrieval-based approach can disrupt the coherence of the semantic chain, particularly in complex reasoning tasks, due to fragmented information flow. In this work, we integrate the above engineering techniques into MLLMs and demonstrate that a model fine-tuned on a high-quality, long-context training corpus is a strong baseline, achieving superior performance compared to its RAG counterpart." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 313, + 310, + 548, + 323 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 310, + 548, + 323 + ], + "spans": [ + { + "bbox": [ + 313, + 310, + 548, + 323 + ], + "type": "text", + "content": "3. Multimodal Document Dataset Generation" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 313, + 330, + 555, + 365 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 330, + 555, + 365 + ], + "spans": [ + { + "bbox": [ + 313, + 330, + 555, + 365 + ], + "type": "text", + "content": "In this section, we begin by introducing the details of the data engine. Following this, we provide a comprehensive overview of the dataset—Doc-750K." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 313, + 373, + 397, + 385 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 373, + 397, + 385 + ], + "spans": [ + { + "bbox": [ + 313, + 373, + 397, + 385 + ], + "type": "text", + "content": "3.1. Data Engine" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 313, + 390, + 555, + 485 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 390, + 555, + 485 + ], + "spans": [ + { + "bbox": [ + 313, + 390, + 555, + 485 + ], + "type": "text", + "content": "The data engine operates primarily through two steps: document content extraction and question-answer (QA) pair construction. Specifically, we first extract multimodal content, including both text and images, from the documents. Based on this extracted content, we then create question-answer pairs. The document content and these constructed pairs are combined to produce conversational-style training data. The format is outlined as:" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 319, + 494, + 538, + 525 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 319, + 494, + 538, + 525 + ], + "spans": [ + { + "bbox": [ + 319, + 494, + 538, + 525 + ], + "type": "text", + "content": "Please read the paper: , and answer the question: Answer: " + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 313, + 534, + 555, + 594 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 534, + 555, + 594 + ], + "spans": [ + { + "bbox": [ + 313, + 534, + 555, + 594 + ], + "type": "text", + "content": "Here, the , , and are the placeholder for extracted document content, the generated questions and answer, respectively. In the following, we will provide a detailed explanation of the two key steps: document content extraction and QA pair construction." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 313, + 594, + 556, + 676 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 594, + 556, + 676 + ], + "spans": [ + { + "bbox": [ + 313, + 594, + 556, + 676 + ], + "type": "text", + "content": "Document Content Extraction. In practical applications, different documents have varying page layouts and content types, which poses significant challenges for content extraction. To enhance the efficiency of multimodal models, it is necessary to organize documents into a unified format for streamlined processing. In this work, we process each document into two formats as follows:" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 313, + 677, + 556, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 677, + 556, + 713 + ], + "spans": [ + { + "bbox": [ + 313, + 677, + 556, + 713 + ], + "type": "text", + "content": "(1) Interleaved Text-Image Format. Using the document content extractor MinerU [79], we segment the document content into interleaved text and image annotations, for" + } + ] + } + ], + "index": 12 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "text", + "content": "4028" + } + ] + } + ], + "index": 13 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 2 + }, + { + "para_blocks": [ + { + "bbox": [ + 89, + 106, + 299, + 137 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 89, + 106, + 299, + 137 + ], + "spans": [ + { + "bbox": [ + 89, + 106, + 299, + 137 + ], + "type": "text", + "content": "The objective consideration of conservation policy under restructuring is proving to be a difficult task. One of the greatest obstacles has been the persistent uncertainty among utility planners regarding the true resource-effectiveness and cost-effectiveness of conservation relative" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 91, + 143, + 254, + 152 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 91, + 143, + 254, + 152 + ], + "spans": [ + { + "bbox": [ + 91, + 143, + 254, + 152 + ], + "type": "text", + "content": "These figures indicate that larger utilities tend to report greater ..." + } + ] + } + ], + "index": 4 + }, + { + "type": "image", + "bbox": [ + 99, + 153, + 162, + 219 + ], + "blocks": [ + { + "bbox": [ + 99, + 153, + 162, + 219 + ], + "lines": [ + { + "bbox": [ + 99, + 153, + 162, + 219 + ], + "spans": [ + { + "bbox": [ + 99, + 153, + 162, + 219 + ], + "type": "image", + "image_path": "1e8cd6897d0cf21bd56005d043630dcd5da915d354ca58756e0370a1e9e36284.jpg" + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 55, + 439, + 555, + 463 + ], + "lines": [ + { + "bbox": [ + 55, + 439, + 555, + 463 + ], + "spans": [ + { + "bbox": [ + 55, + 439, + 555, + 463 + ], + "type": "text", + "content": "Figure 3. Visualization of an example from Doc-750K. The left side presents the interleaved text-image format data obtained through Document Content Extraction, while the right side showcases the annotations generated via Question-Answer Pairs Construction." + } + ] + } + ], + "index": 22, + "angle": 0, + "type": "image_caption" + } + ], + "index": 5 + }, + { + "type": "image", + "bbox": [ + 176, + 157, + 294, + 213 + ], + "blocks": [ + { + "bbox": [ + 176, + 157, + 294, + 213 + ], + "lines": [ + { + "bbox": [ + 176, + 157, + 294, + 213 + ], + "spans": [ + { + "bbox": [ + 176, + 157, + 294, + 213 + ], + "type": "image", + "image_path": "7c96ebc511b971ca903adf4f92d5f88de2e4c30fbb329885667a3f9be9335907.jpg" + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "image_body" + } + ], + "index": 6 + }, + { + "bbox": [ + 89, + 221, + 303, + 244 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 89, + 221, + 303, + 244 + ], + "spans": [ + { + "bbox": [ + 89, + 221, + 303, + 244 + ], + "type": "text", + "content": "Early estimation attempts with individual utilities revealed that there was insufficient variation in the data series to produce reliable coefficient estimates for each utility taken separately. To increase the statistical power of the analysis, ..." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 90, + 251, + 302, + 266 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 90, + 251, + 302, + 266 + ], + "spans": [ + { + "bbox": [ + 90, + 251, + 302, + 266 + ], + "type": "text", + "content": "Finally, since no incentives are offered, it becomes more difficult to attribute customer investment decisions to the utility programs as opposed to other influences." + } + ] + } + ], + "index": 8 + }, + { + "type": "table", + "bbox": [ + 93, + 269, + 200, + 351 + ], + "blocks": [ + { + "bbox": [ + 93, + 269, + 200, + 351 + ], + "lines": [ + { + "bbox": [ + 93, + 269, + 200, + 351 + ], + "spans": [ + { + "bbox": [ + 93, + 269, + 200, + 351 + ], + "type": "table", + "html": "
Utility (C&I Sales)Program Type1993 Conservation Impacts(Annotated)
GWh% of Sales% of Impacts
Pacific Gas & Electric(45,760 GWh)Incentives1,8864.1%63%
Services9402.1%31%
Other170.4%6%
Total3,0136.6%100%
San Diego Gas & Electric(8,415 GWh)Incentives2512.7%65%
Services971.0%25%
Other2820.4%100%
Total3854.1%100%
Southern California Edison(86,721 GWh)Incentives1,7283.7%22%
Services6,02912.9%71%
Other2400.5%3%
Total7,97417.1%100%
Sacramento(8,733 GWh)Incentives1533.2%58%
Services1042.2%40%
Other60.1%2%
Total2822.5%100%
Los Angeles(14,829 GWh)Incentives7800.5%50%
Services7910.5%50%
Other00.0%0%
Total1571.1%100%
", + "image_path": "718578b4a99dc159f339390893b10485b59b40d90d8686993c83f40626ce7523.jpg" + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "table_body" + } + ], + "index": 9 + }, + { + "type": "table", + "bbox": [ + 213, + 269, + 301, + 350 + ], + "blocks": [ + { + "bbox": [ + 213, + 269, + 301, + 350 + ], + "lines": [ + { + "bbox": [ + 213, + 269, + 301, + 350 + ], + "spans": [ + { + "bbox": [ + 213, + 269, + 301, + 350 + ], + "type": "table", + "html": "
StudySampleFactorPeriodElectricity
Elec.UnitCostFuel
Longuet et al. 1994Norway 1976aenergy consumptionLB−0.39−0.39
energy consumptionLB−0.15−0.15
energy consumption & transportLB−0.44−0.16
Hammoudi & Kazer 1995Iran 1955–8domestic heatingLB−0.19−0.19
Safdarzadeh 1995USAcommercialLB−0.70−0.25
industrialLB−0.22−0.22
Rall 1996USAindustrialLB−0.18−0.20−0.150.09
industrialLB−0.29−0.29
Counihan 1998USA 1995–8industrialLB−0.390.530.040.09
industrialLB−0.22−0.22
Murray 2000Virginia, USA 1970–8industrialLB−0.29−0.09
industrialLB−0.22−0.22
industrialLB−0.21−0.04
Woodward 2001Netherlands 1977–83industrialLB−1.11−0.2−0.39−0.16
industrialLB−0.11−0.11
National Energy Board 2003CanadaindustrialLB−0.110.020.05
industrialLB−0.11−0.11
industrialLB−0.450.120.06
Energy, Media and Media Ethics Board Canada 2005CanadaindustrialLB−0.090.030.01
industrialLB−0.09−0.09
industrialLB−0.490.110.01
This studyUSAcommercial RLB−0.078−0.033−0.026−0.013
industrialLB−0.078−0.078
", + "image_path": "fc3ebc0f87d9548c9b90d47d72dd9ac9971482440961d40d1ed63524f0bb55c8.jpg" + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "table_body" + } + ], + "index": 10 + }, + { + "bbox": [ + 89, + 357, + 303, + 418 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 89, + 357, + 303, + 418 + ], + "spans": [ + { + "bbox": [ + 89, + 357, + 303, + 418 + ], + "type": "text", + "content": "According to California regulators, SCE's reported impacts for service programs have always been viewed with skepticism during regulatory reviews, although its impacts for other types of conservation programs have not. In California, service program impacts have historically been viewed as unverifiable, so utilities have been largely ineligible for incentive payments related to service program expenditures. Regulators in California acknowledge that they believe SCE's reported conservation impacts overstate the actual savings achieved by their programs and have communicated this skepticism to the utility. But, according to the rate regulation ..." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 330, + 81, + 353, + 88 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 330, + 81, + 353, + 88 + ], + "spans": [ + { + "bbox": [ + 330, + 81, + 353, + 88 + ], + "type": "text", + "content": "" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 330, + 89, + 447, + 95 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 330, + 89, + 447, + 95 + ], + "spans": [ + { + "bbox": [ + 330, + 89, + 447, + 95 + ], + "type": "text", + "content": "Please read the paper and answer the question:" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 329, + 97, + 505, + 121 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 329, + 97, + 505, + 121 + ], + "spans": [ + { + "bbox": [ + 329, + 97, + 505, + 121 + ], + "type": "text", + "content": "According to this paper, how does Image-1 illustrate the conservation impacts of various utilities in 1993? Discuss the findings and the significance of these data in the context of the research." + } + ] + } + ], + "index": 14 + }, + { + "type": "image", + "bbox": [ + 325, + 134, + 348, + 160 + ], + "blocks": [ + { + "bbox": [ + 325, + 134, + 348, + 160 + ], + "lines": [ + { + "bbox": [ + 325, + 134, + 348, + 160 + ], + "spans": [ + { + "bbox": [ + 325, + 134, + 348, + 160 + ], + "type": "image", + "image_path": "04fe59a24ac3d6bbbff74037dab49d3c97c86e29b1e4bb485e0920f30f7f9889.jpg" + } + ] + } + ], + "index": 16, + "angle": 0, + "type": "image_body" + } + ], + "index": 16 + }, + { + "bbox": [ + 356, + 144, + 545, + 259 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 356, + 144, + 545, + 259 + ], + "spans": [ + { + "bbox": [ + 356, + 144, + 545, + 259 + ], + "type": "text", + "content": "According to the paper, 'In 1993, the 39 companies in the sample accounted for " + }, + { + "bbox": [ + 356, + 144, + 545, + 259 + ], + "type": "inline_equation", + "content": "19\\%" + }, + { + "bbox": [ + 356, + 144, + 545, + 259 + ], + "type": "text", + "content": " of total C&I electricity sales nationwide and " + }, + { + "bbox": [ + 356, + 144, + 545, + 259 + ], + "type": "inline_equation", + "content": "40\\%" + }, + { + "bbox": [ + 356, + 144, + 545, + 259 + ], + "type": "text", + "content": " of all utility conservation program impacts, including residential impacts (Energy Information Administration 1995). The table in Image-1 provides detailed data on the 1993 C&I sales and C&I conservation program impacts for various utilities. This image shows the sales in GWh and conservation impacts both in GWh and as a percentage of sales. For example, California utilities such as Southern California Edison reported high sales and high conservation percentages, which is a significant outlier. These data illustrate the varying levels of conservation achievements across utilities, underlining the study's aim to verify the effectiveness and accuracy of reported conservation impacts. Larger utilities tend to report greater conservation impacts relative to sales, and this detail supports the analysis presented in the paper." + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 328, + 284, + 504, + 300 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 328, + 284, + 504, + 300 + ], + "spans": [ + { + "bbox": [ + 328, + 284, + 504, + 300 + ], + "type": "text", + "content": "Analyze the importance of Image-3 in comparing the study's findings with previous empirical research on electricity price elasticities." + } + ] + } + ], + "index": 18 + }, + { + "type": "image", + "bbox": [ + 523, + 277, + 550, + 304 + ], + "blocks": [ + { + "bbox": [ + 523, + 277, + 550, + 304 + ], + "lines": [ + { + "bbox": [ + 523, + 277, + 550, + 304 + ], + "spans": [ + { + "bbox": [ + 523, + 277, + 550, + 304 + ], + "type": "image", + "image_path": "c880a59ac2258e15418036f94388275e4684eff8fb2a96229ca00928bb57542f.jpg" + } + ] + } + ], + "index": 19, + "angle": 0, + "type": "image_body" + } + ], + "index": 19 + }, + { + "type": "image", + "bbox": [ + 324, + 314, + 348, + 340 + ], + "blocks": [ + { + "bbox": [ + 324, + 314, + 348, + 340 + ], + "lines": [ + { + "bbox": [ + 324, + 314, + 348, + 340 + ], + "spans": [ + { + "bbox": [ + 324, + 314, + 348, + 340 + ], + "type": "image", + "image_path": "392bb71143c9e71645c114493265a9b3032c0eeb7ff4d7c239ff8285ff70a36a.jpg" + } + ] + } + ], + "index": 20, + "angle": 0, + "type": "image_body" + } + ], + "index": 20 + }, + { + "bbox": [ + 356, + 323, + 546, + 422 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 356, + 323, + 546, + 422 + ], + "spans": [ + { + "bbox": [ + 356, + 323, + 546, + 422 + ], + "type": "text", + "content": "According to the paper, 'The regression diagnostics offered little guidance on improving the structure of the model. The addition of a dummy variable for the years 1971-1973 did little to improve the residuals in subsequent years. Adding a set of dummy variables to account for each unstable period did not seem reasonable.' Image-3 presents a comparative analysis of electricity cross-price elasticity studies. This table showcases various studies, their sample sectors, and reported elasticities for electricity, natural gas, oil, and coal across different regions and time periods. The study's own findings, -0.078 for short-run own-price elasticity of electricity, align closely with prior research, confirming its place within established literature. The context provided by other studies helps to validate the paper's conclusions regarding the robustness of its econometric model and the effectiveness of conservation programs." + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 55, + 474, + 295, + 510 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 474, + 295, + 510 + ], + "spans": [ + { + "bbox": [ + 55, + 474, + 295, + 510 + ], + "type": "text", + "content": "example, \"\\n\\n \\n\" This format captures the document's textual content, making it easier to construct question-answer pairs." + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 55, + 514, + 295, + 586 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 514, + 295, + 586 + ], + "spans": [ + { + "bbox": [ + 55, + 514, + 295, + 586 + ], + "type": "text", + "content": "(2) Multi-Image Format. In this format, a document with " + }, + { + "bbox": [ + 55, + 514, + 295, + 586 + ], + "type": "inline_equation", + "content": "n" + }, + { + "bbox": [ + 55, + 514, + 295, + 586 + ], + "type": "text", + "content": " pages is rendered as " + }, + { + "bbox": [ + 55, + 514, + 295, + 586 + ], + "type": "inline_equation", + "content": "n" + }, + { + "bbox": [ + 55, + 514, + 295, + 586 + ], + "type": "text", + "content": " images, with each image corresponding to a single page. The structure follows the pattern \"\\n\\n\". This format preserves the original layout, enabling the model to learn the overall pagination and visual layout of the document." + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 55, + 590, + 295, + 662 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 590, + 295, + 662 + ], + "spans": [ + { + "bbox": [ + 55, + 590, + 295, + 662 + ], + "type": "text", + "content": "After processing the document into contexts in interleaved text-image and paginated image formats, we can not only use these contexts for next-token prediction training but also leverage the document's content, hierarchical structure, and layout features to flexibly and precisely generate high-quality question-answer pairs." + } + ] + } + ], + "index": 25 + }, + { + "bbox": [ + 55, + 666, + 295, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 666, + 295, + 714 + ], + "spans": [ + { + "bbox": [ + 55, + 666, + 295, + 714 + ], + "type": "text", + "content": "Question-Answer Pairs Construction. In this step, we create question-and-answer pairs tailored to the source, content features, and formatting structure of each document. This process is divided into the following main categories:" + } + ] + } + ], + "index": 26 + }, + { + "bbox": [ + 313, + 475, + 554, + 661 + ], + "type": "list", + "angle": 0, + "index": 30, + "blocks": [ + { + "bbox": [ + 313, + 475, + 553, + 510 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 475, + 553, + 510 + ], + "spans": [ + { + "bbox": [ + 313, + 475, + 553, + 510 + ], + "type": "text", + "content": "(1) For documents with reliable QA annotations, like the review and reply in OpenReview, we extract the QA pairs and organize them into conversation format." + } + ] + } + ], + "index": 27 + }, + { + "bbox": [ + 313, + 514, + 554, + 597 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 514, + 554, + 597 + ], + "spans": [ + { + "bbox": [ + 313, + 514, + 554, + 597 + ], + "type": "text", + "content": "(2) For documents with a clear textual structure, such as well-structured papers from Sci-Hub and Arxiv, we convert them to text and segment them, while the model is instructed to generate contents for each segment, including abstracts, experiment descriptions, and captions for figures and tables. The details of each task for structural papers are illustrated in Table 1." + } + ] + } + ], + "index": 28 + }, + { + "bbox": [ + 313, + 601, + 554, + 661 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 601, + 554, + 661 + ], + "spans": [ + { + "bbox": [ + 313, + 601, + 554, + 661 + ], + "type": "text", + "content": "(3) For other documents, in addition to using them directly for NTP pretraining, we can input text interspersed with images into MLLMs to obtain QA pairs. To ensure high-quality generated data, we use the state-of-the-art model GPT-4o [56]." + } + ] + } + ], + "index": 29 + } + ], + "sub_type": "text" + }, + { + "bbox": [ + 313, + 666, + 554, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 666, + 554, + 714 + ], + "spans": [ + { + "bbox": [ + 313, + 666, + 554, + 714 + ], + "type": "text", + "content": "Through our pipeline, most data has been processed into high-quality document-level question-answering data, while the remaining data is converted to plain text and used for next-token prediction tasks. Our pipelines are metricu" + } + ] + } + ], + "index": 31 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 59, + 71, + 80, + 97 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 59, + 71, + 80, + 97 + ], + "spans": [ + { + "bbox": [ + 59, + 71, + 80, + 97 + ], + "type": "text", + "content": "0" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 89, + 81, + 194, + 88 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 89, + 81, + 194, + 88 + ], + "spans": [ + { + "bbox": [ + 89, + 81, + 194, + 88 + ], + "type": "text", + "content": "How Many Kilowatts are in a Negawatt?" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 89, + 89, + 302, + 97 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 89, + 89, + 302, + 97 + ], + "spans": [ + { + "bbox": [ + 89, + 89, + 302, + 97 + ], + "type": "text", + "content": "Verifying \"Ex Post\" Estimates of Utility Conservation Impacts at the Regional Level" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 523, + 72, + 550, + 101 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 523, + 72, + 550, + 101 + ], + "spans": [ + { + "bbox": [ + 523, + 72, + 550, + 101 + ], + "type": "text", + "content": "#" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "text", + "content": "4029" + } + ] + } + ], + "index": 32 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 3 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 58, + 71, + 553, + 177 + ], + "blocks": [ + { + "bbox": [ + 58, + 71, + 553, + 177 + ], + "lines": [ + { + "bbox": [ + 58, + 71, + 553, + 177 + ], + "spans": [ + { + "bbox": [ + 58, + 71, + 553, + 177 + ], + "type": "image", + "image_path": "728d8d93a75762d4ff8bb882269a03389f5b3a05187bda3f5deb0af18bf957ae.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 122, + 178, + 225, + 188 + ], + "lines": [ + { + "bbox": [ + 122, + 178, + 225, + 188 + ], + "spans": [ + { + "bbox": [ + 122, + 178, + 225, + 188 + ], + "type": "text", + "content": "(a) Data Distribution of Doc-750K" + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 358, + 182, + 487, + 191 + ], + "lines": [ + { + "bbox": [ + 358, + 182, + 487, + 191 + ], + "spans": [ + { + "bbox": [ + 358, + 182, + 487, + 191 + ], + "type": "text", + "content": "(b) Data Recipe for Supervised Fine-tuning" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 54, + 202, + 555, + 237 + ], + "lines": [ + { + "bbox": [ + 54, + 202, + 555, + 237 + ], + "spans": [ + { + "bbox": [ + 54, + 202, + 555, + 237 + ], + "type": "text", + "content": "Figure 4. Data distribution of our dataset. The outer circle shows the distribution of all data categories and the inner circle shows the distribution of data subsets. Left: Data distribution of Doc-750K. Right: Data distribution of our complete SFT training dataset. Note that the number reported in the figure represents the number of samples. “MT” is short for multi-turn." + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "type": "table", + "bbox": [ + 56, + 246, + 289, + 477 + ], + "blocks": [ + { + "bbox": [ + 56, + 246, + 289, + 477 + ], + "lines": [ + { + "bbox": [ + 56, + 246, + 289, + 477 + ], + "spans": [ + { + "bbox": [ + 56, + 246, + 289, + 477 + ], + "type": "table", + "html": "
TasksQuestions
Abstract WritingRead the full text of the paper and provide a concise summary in the form of an abstract.
Paper TitlingBased on the provided abstract or introduction of the research paper, please generate a concise and informative title
Caption WritingGive the relative texts of the images or tables, please write a caption for each image or table based on the relative texts provided.
Experiment WritingPlease write the "Experiments" section based on the incomplete research paper provided.
TranslationPlease read the full text of the following research paper and translate the Experiments section into Chinese.
", + "image_path": "5a70c6e62b7dca2b729d85e214f8271bf3452cb07f185ccdf1f577a4c971f2e4.jpg" + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "table_body" + } + ], + "index": 4 + }, + { + "bbox": [ + 55, + 543, + 296, + 604 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 543, + 296, + 604 + ], + "spans": [ + { + "bbox": [ + 55, + 543, + 296, + 604 + ], + "type": "text", + "content": "lously designed to ensure high data quality across all generated context. Each LLM-generated sample is explicitly marked in the metadata as model-generated. Across the entire dataset, only " + }, + { + "bbox": [ + 55, + 543, + 296, + 604 + ], + "type": "inline_equation", + "content": "4.8\\%" + }, + { + "bbox": [ + 55, + 543, + 296, + 604 + ], + "type": "text", + "content": " of the data is LLM-generated, reinforcing the overall reliability and quality of the dataset." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 55, + 611, + 223, + 624 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 611, + 223, + 624 + ], + "spans": [ + { + "bbox": [ + 55, + 611, + 223, + 624 + ], + "type": "text", + "content": "3.2. Multimodal Document Dataset" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 55, + 629, + 296, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 629, + 296, + 713 + ], + "spans": [ + { + "bbox": [ + 55, + 629, + 296, + 713 + ], + "type": "text", + "content": "Data Source. The composition and distribution of our training data are detailed in Figure 4. Specifically, our dataset predominantly consists of academic papers, which constitute approximately " + }, + { + "bbox": [ + 55, + 629, + 296, + 713 + ], + "type": "inline_equation", + "content": "32.6\\%" + }, + { + "bbox": [ + 55, + 629, + 296, + 713 + ], + "type": "text", + "content": " of the total data. The multimodal data, carefully selected to augment our model's learning dimensions, makes up about " + }, + { + "bbox": [ + 55, + 629, + 296, + 713 + ], + "type": "inline_equation", + "content": "88.8\\%" + }, + { + "bbox": [ + 55, + 629, + 296, + 713 + ], + "type": "text", + "content": " of our dataset. This strategic distribution is designed to optimize the train-" + } + ] + } + ], + "index": 8 + }, + { + "type": "table", + "bbox": [ + 365, + 246, + 504, + 358 + ], + "blocks": [ + { + "bbox": [ + 55, + 485, + 295, + 531 + ], + "lines": [ + { + "bbox": [ + 55, + 485, + 295, + 531 + ], + "spans": [ + { + "bbox": [ + 55, + 485, + 295, + 531 + ], + "type": "text", + "content": "Table 1. Questions format for different tasks. For documents with a clear textual structure, we design several proxy tasks. All tasks leverage the inherent structure of the documents, with answers directly sourced from the original text." + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 365, + 246, + 504, + 358 + ], + "lines": [ + { + "bbox": [ + 365, + 246, + 504, + 358 + ], + "spans": [ + { + "bbox": [ + 365, + 246, + 504, + 358 + ], + "type": "table", + "html": "
StatisticsNumber
Total Questions758K
Total Images3.1M
Total Conversations251K
Multi-Turn Questions87K
Single-Turn Questions164K
Average Text Tokens11245
Average Image Tokens6178
", + "image_path": "34178f1c1b749176664d52c14694e320668ee4086e61eda412a23e5bd0b531b8.jpg" + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "table_body" + } + ], + "index": 9 + }, + { + "bbox": [ + 313, + 365, + 555, + 422 + ], + "lines": [ + { + "bbox": [ + 313, + 365, + 555, + 422 + ], + "spans": [ + { + "bbox": [ + 313, + 365, + 555, + 422 + ], + "type": "text", + "content": "Table 2. Key statistics of the Doc-750K datasets. It comprises 758K questions, 3.1M images, and 251K conversations, including 87K multi-turn and 164K single-turn questions. With an average of 11,245 text tokens and 6,178 image tokens, it highlights the dataset's richness and diversity for multimodal research." + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "text" + }, + { + "bbox": [ + 313, + 434, + 554, + 458 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 434, + 554, + 458 + ], + "spans": [ + { + "bbox": [ + 313, + 434, + 554, + 458 + ], + "type": "text", + "content": "ing process and improve the model's ability to generalize across different types of data inputs." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 313, + 459, + 555, + 602 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 459, + 555, + 602 + ], + "spans": [ + { + "bbox": [ + 313, + 459, + 555, + 602 + ], + "type": "text", + "content": "Dataset Statistics. In our Doc-750K dataset, the majority of the data consists of reliably annotated entries, with OpenReview and Arxiv collectively accounting for " + }, + { + "bbox": [ + 313, + 459, + 555, + 602 + ], + "type": "inline_equation", + "content": "75.4\\%" + }, + { + "bbox": [ + 313, + 459, + 555, + 602 + ], + "type": "text", + "content": ". The remaining data, sourced from Sci-Hub, is processed using our designed tasks. The overall distribution and number of tasks are shown in Figure 4(a). Our dataset ultimately consists of 251K conversations, comprising a total of 758K questions. Additional statistical details are provided in Table 2. Compared to previous datasets, Doc-750K contains a larger number of images, with an average of four images per conversation segment. Further comparisons with other datasets are shown in Table 3." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 313, + 611, + 525, + 624 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 611, + 525, + 624 + ], + "spans": [ + { + "bbox": [ + 313, + 611, + 525, + 624 + ], + "type": "text", + "content": "3.3. Data Recipe for Supervised Fine-Tuning" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 313, + 629, + 554, + 700 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 629, + 554, + 700 + ], + "spans": [ + { + "bbox": [ + 313, + 629, + 554, + 700 + ], + "type": "text", + "content": "Although Doc-750K effectively covers multimodal document QA scenarios, using it directly may lead to model over-fitting on a specific document domain. Therefore, we combine it with several open-source datasets to create a mixed dataset for SFT training. As shown in Figure 4(b), these datasets are organized into 4 categories as follows:" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 325, + 701, + 554, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 325, + 701, + 554, + 714 + ], + "spans": [ + { + "bbox": [ + 325, + 701, + 554, + 714 + ], + "type": "text", + "content": "(1) For multi-page document QA, Doc-750K serves" + } + ] + } + ], + "index": 15 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "text", + "content": "4030" + } + ] + } + ], + "index": 16 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 4 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 56, + 70, + 294, + 194 + ], + "blocks": [ + { + "bbox": [ + 56, + 70, + 294, + 194 + ], + "lines": [ + { + "bbox": [ + 56, + 70, + 294, + 194 + ], + "spans": [ + { + "bbox": [ + 56, + 70, + 294, + 194 + ], + "type": "table", + "html": "
Dataset#Images#QA Pairs#Tokens
Docmatix [33]2,444,7509,500,000390,000,000
DocVQA [15]10,18939,463337,829
TextCaps [64]21,95321,953389,658
TextVQA [65]21,95334,602181,918
ST-VQA [8]17,24723,121127,846
OCR-VQA [55]165,746801,5796,073,824
VisualMRC [70]3,02711,988168,828
DUDE [78]147,59723,71611,341,228
Doc-750K (ours)3,103,494758,0005,200,000,000
", + "image_path": "4f85ce1ef16faf309af9a9df58f29b1ac12e5962f886980acfc7fcd595266def.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "table_body" + } + ], + "index": 0 + }, + { + "bbox": [ + 80, + 201, + 271, + 213 + ], + "lines": [ + { + "bbox": [ + 80, + 201, + 271, + 213 + ], + "spans": [ + { + "bbox": [ + 80, + 201, + 271, + 213 + ], + "type": "text", + "content": "Table 3. Comparison with popular VQA datasets." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "text" + }, + { + "bbox": [ + 55, + 227, + 295, + 299 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 227, + 295, + 299 + ], + "spans": [ + { + "bbox": [ + 55, + 227, + 295, + 299 + ], + "type": "text", + "content": "as the core dataset, specifically curated to address complex, multi-page document comprehension. Additional datasets such as MP-Docmatix [33], MP-DocVQA [53], DUDE [78], and Taesiri-ArxivQA [69] offer valuable multi-page scenarios requiring inter-page reasoning and contextual retention across sequences." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 55, + 299, + 295, + 517 + ], + "type": "list", + "angle": 0, + "index": 6, + "blocks": [ + { + "bbox": [ + 55, + 299, + 295, + 384 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 299, + 295, + 384 + ], + "spans": [ + { + "bbox": [ + 55, + 299, + 295, + 384 + ], + "type": "text", + "content": "(2) For multi-image general QA, MMDU-45K [48] offers a comprehensive dataset encompassing diverse real-world scenarios, such as natural environments and everyday contexts. It emphasizes multi-turn dialogues and integration of multiple images, supporting the development of systems capable of generating coherent and accurate responses from complex, lengthy inputs." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 55, + 384, + 295, + 456 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 384, + 295, + 456 + ], + "spans": [ + { + "bbox": [ + 55, + 384, + 295, + 456 + ], + "type": "text", + "content": "(3) For single-page document QA, We introduce DocVQA [53], DocReason [93], InfoVQA [54], and ChartQA [52] to further enhance the diversity of the SFT dataset. These datasets focus on individual pages with complex layouts, rich textual information, and, in some cases, graphical data interpretation." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 55, + 456, + 295, + 517 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 456, + 295, + 517 + ], + "spans": [ + { + "bbox": [ + 55, + 456, + 295, + 517 + ], + "type": "text", + "content": "(4) For pure-text QA, we add datasets including LongAlpaca [11], LongAlpaca-16K-Length [11], LongQLoRA [91], LongCite [96], LongAlign [7], and LongReward [97] to support the assessment of the model's capabilities in QA tasks requiring long-range dependencies." + } + ] + } + ], + "index": 5 + } + ], + "sub_type": "text" + }, + { + "bbox": [ + 55, + 517, + 295, + 565 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 517, + 295, + 565 + ], + "spans": [ + { + "bbox": [ + 55, + 517, + 295, + 565 + ], + "type": "text", + "content": "This expanded dataset provides a balanced foundation for training and evaluating multimodal document understanding models, enhancing robustness and adaptability across diverse document-related VQA tasks." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 55, + 577, + 295, + 605 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 577, + 295, + 605 + ], + "spans": [ + { + "bbox": [ + 55, + 577, + 295, + 605 + ], + "type": "text", + "content": "4. Enhanced Baseline for Document-Level Multimodal Understanding" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 55, + 611, + 170, + 624 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 611, + 170, + 624 + ], + "spans": [ + { + "bbox": [ + 55, + 611, + 170, + 624 + ], + "type": "text", + "content": "4.1. Model Architecture" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 55, + 629, + 295, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 629, + 295, + 713 + ], + "spans": [ + { + "bbox": [ + 55, + 629, + 295, + 713 + ], + "type": "text", + "content": "Our model architecture leverages the widely-adopted ViTMLP-LLM structure [41, 42, 73], consisting of a pre-trained Vision Transformer (ViT), a two-layer MLP projector, and a pre-trained Language Model (LLM). This combination provides a strong baseline for multimodal document analysis, effectively integrating visual and textual information within a unified framework." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 314, + 72, + 482, + 85 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 314, + 72, + 482, + 85 + ], + "spans": [ + { + "bbox": [ + 314, + 72, + 482, + 85 + ], + "type": "text", + "content": "4.2. Optimizing Training Efficiency" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 313, + 95, + 555, + 192 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 95, + 555, + 192 + ], + "spans": [ + { + "bbox": [ + 313, + 95, + 555, + 192 + ], + "type": "text", + "content": "The training efficiency of MLLMs is hindered by two key challenges: (1) Inconsistent Sample Lengths. Samples with different context lengths will result in excessive padding and lower training throughput; and (2) Limited GPU Memory. As the model scale and context length increase, GPU memory consumption becomes increasingly unsustainable. To address these issues, we have implemented the following strategies:" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 313, + 198, + 556, + 714 + ], + "type": "list", + "angle": 0, + "index": 16, + "blocks": [ + { + "bbox": [ + 313, + 198, + 555, + 462 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 198, + 555, + 462 + ], + "spans": [ + { + "bbox": [ + 313, + 198, + 555, + 462 + ], + "type": "text", + "content": "(1) Multimodal Data Packing. To balance the computational load between the vision model (ViT) and the language model (LLM) while minimizing resource waste caused by padding, we implement a multimodal data-packing strategy. The key idea is to concatenate multiple samples into long sequences to fully utilize the model's input capacity. Specifically, thresholds " + }, + { + "bbox": [ + 313, + 198, + 555, + 462 + ], + "type": "inline_equation", + "content": "T_{\\mathrm{img}}" + }, + { + "bbox": [ + 313, + 198, + 555, + 462 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 313, + 198, + 555, + 462 + ], + "type": "inline_equation", + "content": "T_{\\mathrm{tok}}" + }, + { + "bbox": [ + 313, + 198, + 555, + 462 + ], + "type": "text", + "content": " are set for the number of images and tokens, respectively. Samples are managed using a priority queue, sorted in descending order by the number of images and total tokens. A new sample " + }, + { + "bbox": [ + 313, + 198, + 555, + 462 + ], + "type": "inline_equation", + "content": "s" + }, + { + "bbox": [ + 313, + 198, + 555, + 462 + ], + "type": "text", + "content": " attempts to combine with the sample at the front of the priority queue. If the combination meets the thresholds (i.e., " + }, + { + "bbox": [ + 313, + 198, + 555, + 462 + ], + "type": "inline_equation", + "content": "T_{\\mathrm{img}}, T_{\\mathrm{tok}}" + }, + { + "bbox": [ + 313, + 198, + 555, + 462 + ], + "type": "text", + "content": "), the combined sample is pushed back into the priority queue. If " + }, + { + "bbox": [ + 313, + 198, + 555, + 462 + ], + "type": "inline_equation", + "content": "s" + }, + { + "bbox": [ + 313, + 198, + 555, + 462 + ], + "type": "text", + "content": " cannot match with any existing sample in the queue, it is directly added to the queue. When the image number and total token number of the front sample reach one of the thresholds or the number of samples exceeds the maximum limit " + }, + { + "bbox": [ + 313, + 198, + 555, + 462 + ], + "type": "inline_equation", + "content": "M" + }, + { + "bbox": [ + 313, + 198, + 555, + 462 + ], + "type": "text", + "content": ", the front sample is dequeued, padded as needed, and sent for training. This strategy optimizes resource utilization and ensures balanced computational workloads. The detailed pseudo-code can be found in the supplementary materials." + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 313, + 468, + 556, + 599 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 468, + 556, + 599 + ], + "spans": [ + { + "bbox": [ + 313, + 468, + 556, + 599 + ], + "type": "text", + "content": "(2) Ring Attention. We implement the Ring Attention mechanism [43] to alleviate memory constraints associated with processing long sequences. By partitioning sequences into blocks and distributing computation across multiple devices, Ring Attention allows the model to accommodate larger contexts. This approach enables overlapping communication between key-value blocks and attention computations, thereby enhancing parallel processing efficiency. Consequently, Ring Attention improves the model's capacity to handle extended context lengths without exceeding memory limits." + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 313, + 605, + 556, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 605, + 556, + 714 + ], + "spans": [ + { + "bbox": [ + 313, + 605, + 556, + 714 + ], + "type": "text", + "content": "(3) Liger Kernel. To further improve memory and computational efficiency, we integrate the Liger Kernel [16], a specialized kernel library optimized for large-scale model training. The Liger Kernel enhances throughput and reduces memory consumption by employing techniques like kernel fusion, in-place operations, and input chunking. Leveraging the Liger Kernel thus enables higher training throughput and addresses memory limitations, allowing for efficient scaling of large multimodal models." + } + ] + } + ], + "index": 15 + } + ], + "sub_type": "text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 748, + 314, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 314, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 314, + 757 + ], + "type": "text", + "content": "4031" + } + ] + } + ], + "index": 17 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 5 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 56, + 70, + 558, + 280 + ], + "blocks": [ + { + "bbox": [ + 56, + 70, + 558, + 280 + ], + "lines": [ + { + "bbox": [ + 56, + 70, + 558, + 280 + ], + "spans": [ + { + "bbox": [ + 56, + 70, + 558, + 280 + ], + "type": "table", + "html": "
ModelsMP-DocMMLong-DocDocGenomeMM-NIAH
ANSL↑Acc↑F1↑Class Acc↑Title ED↓Abstract ED↓SP Acc↑MP Acc↑ShortMediumLongOverall
Proprietary Models
Gemini-1.5-Pro [60]-28.220.6-----73.865.260.867.1
GPT-4o [56]-42.844.997.69.56.571.867.6----
Open-Source Models
MiniMonkey-2B [28]70.310.38.657.416.555.040.328.940.926.923.531.0
InternVL2-2B [13]71.810.510.860.818.454.339.428.936.621.219.426.4
InternVL2-2B + RAG [85]72.617.216.760.818.454.339.428.436.830.234.833.8
Llama3.2-3B-Instruct†[21]-23.721.285.3194.751.040.234.915.52.20.56.6
Docopilot-2B (ours)76.221.816.056.24.543.645.137.458.046.740.949.2
MiniCPM-V2.6-8B [92]-16.915.492.810.232.660.054.249.015.30.023.4
LLaVA-OneVision-8B [36]-10.89.685.649.977.59.87.165.738.00.037.7
mPLUG-DocOwl2-8B [27]69.413.48.9-----17.90.10.06.6
M3DocRAG [14]84.421.022.6---------
VisRAG-8B [95]-18.818.392.810.232.660.050.747.129.229.535.8
InternLM2.5-7B-1M†[9]-28.725.692.777.659.342.742.540.537.235.137.8
InternVL2-8B [13]79.317.416.590.68.239.656.046.156.437.332.442.9
InternVL2-8B + RAG [85]78.724.224.590.68.239.656.046.055.743.445.248.4
InternVL2-26B [13]-15.515.487.516.923.349.742.765.048.741.952.8
Docopilot-8B (ours)81.328.823.093.82.019.753.951.971.257.455.361.8
", + "image_path": "16bbf05e201805b33321f9199fc0b99105a33f078c397d055d771a030686a69d.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "table_body" + } + ], + "index": 0 + }, + { + "bbox": [ + 55, + 342, + 136, + 357 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 342, + 136, + 357 + ], + "spans": [ + { + "bbox": [ + 55, + 342, + 136, + 357 + ], + "type": "text", + "content": "5. Experiments" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 55, + 361, + 171, + 374 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 361, + 171, + 374 + ], + "spans": [ + { + "bbox": [ + 55, + 361, + 171, + 374 + ], + "type": "text", + "content": "5.1. Experimental Setup" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 54, + 377, + 295, + 532 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 54, + 377, + 295, + 532 + ], + "spans": [ + { + "bbox": [ + 54, + 377, + 295, + 532 + ], + "type": "text", + "content": "Training Details. Our model is available in two sizes: Docopilot-2B and Docopilot-8B, both of which are based on the InternVL2 [73] and fine-tuned for one epoch using the data recipe that includes Doc-750K. The training uses a batch size of 128, the AdamW optimizer with a learning rate of 1e-5, weight decay of 0.01 for the 2B variant, and 0.05 for the 8B variant, along with a cosine learning rate schedule. To speed up training, we apply multimodal data packing to reduce padding and a dynamic high-resolution strategy [13] to enhance OCR for document understanding. The maximum number of tiles for multimodal data is limited to 24, and the maximum sequence length is set to 32k tokens." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 55, + 534, + 295, + 666 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 534, + 295, + 666 + ], + "spans": [ + { + "bbox": [ + 55, + 534, + 295, + 666 + ], + "type": "text", + "content": "Baselines. We compare our Docopilot with a series of open-source document-level MLLMs [27, 28, 36, 40, 46, 92, 98] that supports multi-image input and proprietary MLLMs, including Gemini-1.5-pro [60], GPT-4o [57]. For comparison with the commonly used RAG method for handling long documents, we selected the latest multimodal RAG methods VisRAG [95], InternVL + RAG [85], and M3DocRAG [14]. To compare with more long-context large language models, we use InternVL2-8B as the OCR model to extract texts from the documents and images and feed the parsed documents to long-context LLMs [9, 21]." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 55, + 671, + 157, + 685 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 671, + 157, + 685 + ], + "spans": [ + { + "bbox": [ + 55, + 671, + 157, + 685 + ], + "type": "text", + "content": "5.2. Multi-Page VQA" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 55, + 689, + 295, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 689, + 295, + 715 + ], + "spans": [ + { + "bbox": [ + 55, + 689, + 295, + 715 + ], + "type": "text", + "content": "Benchmarks. For the multi-page VQA task, we evaluate our model on three benchmarks: (1) MP-DocVQA [76]," + } + ] + } + ], + "index": 7 + }, + { + "type": "table", + "bbox": [ + 315, + 342, + 553, + 479 + ], + "blocks": [ + { + "bbox": [ + 55, + 287, + 555, + 333 + ], + "lines": [ + { + "bbox": [ + 55, + 287, + 555, + 333 + ], + "spans": [ + { + "bbox": [ + 55, + 287, + 555, + 333 + ], + "type": "text", + "content": "Table 4. Evaluation on multi-page and interleaved VQA benchmarks. We report the metrics on MP-DocVQA [76] (MP-Doc), MMLongbench-Doc [51] (MMLong-Doc), DocGenome [89], and MM-NIAH [85]. Our model outperforms document-level MLLMs and multimodal RAG methods on multi-page, medium, and long-context QA. The \"Short\", \"Medium\", and \"Long\" in MM-NIAH refer to input length in " + }, + { + "bbox": [ + 55, + 287, + 555, + 333 + ], + "type": "inline_equation", + "content": "[0,8\\mathrm{k}]" + }, + { + "bbox": [ + 55, + 287, + 555, + 333 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 55, + 287, + 555, + 333 + ], + "type": "inline_equation", + "content": "(8\\mathrm{k},32\\mathrm{k}]" + }, + { + "bbox": [ + 55, + 287, + 555, + 333 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 55, + 287, + 555, + 333 + ], + "type": "inline_equation", + "content": "(32\\mathrm{k},64\\mathrm{k}]" + }, + { + "bbox": [ + 55, + 287, + 555, + 333 + ], + "type": "text", + "content": ", respectively. \"" + }, + { + "bbox": [ + 55, + 287, + 555, + 333 + ], + "type": "inline_equation", + "content": "\\dagger" + }, + { + "bbox": [ + 55, + 287, + 555, + 333 + ], + "type": "text", + "content": "” denotes input documents are parsed by OCR models." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 315, + 342, + 553, + 479 + ], + "lines": [ + { + "bbox": [ + 315, + 342, + 553, + 479 + ], + "spans": [ + { + "bbox": [ + 315, + 342, + 553, + 479 + ], + "type": "table", + "html": "
ModelsDocVQAChartQAInfoVQA
Gemini-1.5-Pro [60]93.187.281.0
GPT-4o [56]92.885.7-
MiniMonkey-2B [28]87.476.560.1
InternVL2-2B [13]86.976.258.9
Docopilot-2B (ours)87.376.458.5
Monkey-8B [40]66.565.136.1
TextMonkey-9B [46]73.066.928.6
mPLUG-DocOwl2-8B [27]80.770.046.4
IXC2.5-7B [98]90.982.269.9
InternVL2-8B [13]91.683.374.8
Docopilot-8B (ours)92.083.373.3
", + "image_path": "26f1931376001c27d19bacb349d2724e9b7d9a32a8ecfd149541b2a2e89a0f47.jpg" + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "table_body" + } + ], + "index": 8 + }, + { + "bbox": [ + 313, + 484, + 555, + 518 + ], + "lines": [ + { + "bbox": [ + 313, + 484, + 555, + 518 + ], + "spans": [ + { + "bbox": [ + 313, + 484, + 555, + 518 + ], + "type": "text", + "content": "Table 5. Results on single-page VQA benchmarks. Our Docopilot models perform comparably to baselines [73], demonstrating enhanced long-context modeling without loss on shorter tasks." + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "text" + }, + { + "bbox": [ + 313, + 533, + 555, + 604 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 533, + 555, + 604 + ], + "spans": [ + { + "bbox": [ + 313, + 533, + 555, + 604 + ], + "type": "text", + "content": "which is designed to evaluate the ability to handle complex questions across multiple scanned document pages. (2) MMLongbench-Doc [51], a benchmark for evaluating the performance of MLLMs on multi-modal documents. (3) DocGenome [89], a large-scale benchmark for the evaluation of scientific document comprehension." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 313, + 605, + 556, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 605, + 556, + 715 + ], + "spans": [ + { + "bbox": [ + 313, + 605, + 556, + 715 + ], + "type": "text", + "content": "Results. As illustrated in Table 4, our model achieves consistent improvements on multi-page QA benchmarks, outperforming previous document-level MLLMs. Notably, our Docopilot-8B surpasses Gemini-1.5-Pro [60] on MMLongBench-Doc, positioning it as the closest open-source model to GPT-4o. In comparison to RAG-based methods [14, 86, 95], our model demonstrates advantages in multi-page scenarios. For example, in the Multi/Page QA of DocGenome benchmark, the RAG method shows a perfor" + } + ] + } + ], + "index": 11 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "text", + "content": "4032" + } + ] + } + ], + "index": 12 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 6 + }, + { + "para_blocks": [ + { + "bbox": [ + 55, + 72, + 295, + 120 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 72, + 295, + 120 + ], + "spans": [ + { + "bbox": [ + 55, + 72, + 295, + 120 + ], + "type": "text", + "content": "mance decline due to the disruption of document continuity while our Docopilot exhibits a significantly stable improvement compared to the baseline, with Docopilot-8B showing an increase of " + }, + { + "bbox": [ + 55, + 72, + 295, + 120 + ], + "type": "inline_equation", + "content": "12.6\\%" + }, + { + "bbox": [ + 55, + 72, + 295, + 120 + ], + "type": "text", + "content": " over InternVL2-8B." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 55, + 129, + 219, + 141 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 129, + 219, + 141 + ], + "spans": [ + { + "bbox": [ + 55, + 129, + 219, + 141 + ], + "type": "text", + "content": "5.3. Interleaved Long-Context QA" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 55, + 146, + 295, + 182 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 146, + 295, + 182 + ], + "spans": [ + { + "bbox": [ + 55, + 146, + 295, + 182 + ], + "type": "text", + "content": "Benchmarks. For the interleaved long-context QA task, we evaluate our models on MM-NIAH [86], a benchmark designed for long multimodal document comprehension." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 55, + 182, + 295, + 361 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 182, + 295, + 361 + ], + "spans": [ + { + "bbox": [ + 55, + 182, + 295, + 361 + ], + "type": "text", + "content": "Results. The right side of Table 4 presents the results of MM-NIAH across context lengths ranging from 1K to 64K. We categorize the context lengths into \"Short,\" \"Medium,\" and \"Long\" based on the context window of InternVL2 (8K) and Docopilot (32K). Our Docopilot demonstrates exceptional performance in both medium- and long-context scenarios, while maintaining high accuracy in short-context situations. Notably, for QA tasks with context lengths in the range of " + }, + { + "bbox": [ + 55, + 182, + 295, + 361 + ], + "type": "inline_equation", + "content": "(32K,64K]" + }, + { + "bbox": [ + 55, + 182, + 295, + 361 + ], + "type": "text", + "content": ", Docopilot-2B outperforms InternVL2-2B by " + }, + { + "bbox": [ + 55, + 182, + 295, + 361 + ], + "type": "inline_equation", + "content": "110\\%" + }, + { + "bbox": [ + 55, + 182, + 295, + 361 + ], + "type": "text", + "content": ", and Docopilot-8B surpasses InternVL2-8B by " + }, + { + "bbox": [ + 55, + 182, + 295, + 361 + ], + "type": "inline_equation", + "content": "70\\%" + }, + { + "bbox": [ + 55, + 182, + 295, + 361 + ], + "type": "text", + "content": ". Furthermore, our model performs comparably to the state-of-the-art multimodal long-context model Gemini-1.5-Pro in contexts longer than 8K, establishing a new state-of-the-art performance among open-source long-context MLLMs." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 55, + 370, + 160, + 383 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 370, + 160, + 383 + ], + "spans": [ + { + "bbox": [ + 55, + 370, + 160, + 383 + ], + "type": "text", + "content": "5.4. Single-Page VQA" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 55, + 388, + 295, + 471 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 388, + 295, + 471 + ], + "spans": [ + { + "bbox": [ + 55, + 388, + 295, + 471 + ], + "type": "text", + "content": "**Benchmarks.** For single-page VQA tasks, we evaluate our model on three benchmarks: (1) DocVQA [53], a benchmark for the evaluation of extracting key information from an image of the given document. (2) ChartQA [52], a benchmark for evaluating the reasoning abilities for chart images. (3) InfoVQA [54], a benchmark for infographic image comprehension." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 55, + 472, + 295, + 567 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 472, + 295, + 567 + ], + "spans": [ + { + "bbox": [ + 55, + 472, + 295, + 567 + ], + "type": "text", + "content": "Results. As shown in Table 5, our model achieves comparable performance to baseline models. Across the three benchmarks, Docopilot-2B and InternVL2-2B exhibit comparable results, while Docopilot-8B outperforms InternVL2-8B by 0.4 points in DocVQA. These results demonstrate that Doc-750K effectively enhances the model's long-context modeling capabilities without compromising its performance on shorter documents." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 55, + 576, + 149, + 588 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 576, + 149, + 588 + ], + "spans": [ + { + "bbox": [ + 55, + 576, + 149, + 588 + ], + "type": "text", + "content": "5.5. Ablation Study" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 55, + 594, + 295, + 712 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 594, + 295, + 712 + ], + "spans": [ + { + "bbox": [ + 55, + 594, + 295, + 712 + ], + "type": "text", + "content": "Effect of Doc-750K. We conducted ablation studies on MMLongBench-Doc [51] to analyze the impact of our Doc-750K. We divided Doc-750K into 3 parts according to the source of the data: (1) Sci-Hub data; (2) Arxiv data; and (3) OpenReview data. We demonstrate the effects of incorporating each part of the data into the SFT process, reported in Table 6. We observed that with the inclusion of different parts of Doc-750K, the model's performance improves continuously. Utilizing only open-source data results in an inferior F1 score." + } + ] + } + ], + "index": 8 + }, + { + "type": "table", + "bbox": [ + 317, + 71, + 553, + 140 + ], + "blocks": [ + { + "bbox": [ + 317, + 71, + 553, + 140 + ], + "lines": [ + { + "bbox": [ + 317, + 71, + 553, + 140 + ], + "spans": [ + { + "bbox": [ + 317, + 71, + 553, + 140 + ], + "type": "table", + "html": "
ModelsAccF1
Baseline (InternVL2-2B [73])10.510.8
- Variant1: SFT using data recipe w/o Doc-750K18.49.4
- Variant2: Variant1 + our Sci-Hub data18.515.2
- Variant3: Variant2 + our Arxiv data20.515.5
Docopilot-2B: Variant3 + our OpenReview data21.816.0
", + "image_path": "f3afcee789bbe9458060b54057fead125444d4ee51d890372258ee2e54c7f719.jpg" + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "table_body" + } + ], + "index": 9 + }, + { + "type": "table", + "bbox": [ + 322, + 203, + 547, + 312 + ], + "blocks": [ + { + "bbox": [ + 313, + 148, + 555, + 194 + ], + "lines": [ + { + "bbox": [ + 313, + 148, + 555, + 194 + ], + "spans": [ + { + "bbox": [ + 313, + 148, + 555, + 194 + ], + "type": "text", + "content": "Table 6. Ablation study on the data recipe. We evaluate the effectiveness of the training data from different sources on MMLongBench-Doc. Our Doc-750K can consistently enhance the ability of the model to understand multi-page documents." + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 322, + 203, + 547, + 312 + ], + "lines": [ + { + "bbox": [ + 322, + 203, + 547, + 312 + ], + "spans": [ + { + "bbox": [ + 322, + 203, + 547, + 312 + ], + "type": "table", + "html": "
ModelsLatencyAccF1
MiniCPM-V2.6 [92]225.4ms16.915.4
VisRAG-12B [95]288.3ms18.818.3
InternVL2-2B [13]35.9ms10.510.8
InternVL2-2B + RAG [85]82.9ms17.216.7
Docopilot-2B (ours)35.9ms21.816.0
InternVL2-8B [13]81.0ms17.416.5
InternVL2-8B + RAG [85]113.4ms24.224.5
Docopilot-8B (ours)81.0ms28.823.0
", + "image_path": "60652b1fe221f3d6db2a12e66b7b07a4d1ee9dca2a83c2944df0dea0797c3d81.jpg" + } + ] + } + ], + "index": 11, + "angle": 0, + "type": "table_body" + } + ], + "index": 11 + }, + { + "bbox": [ + 313, + 320, + 555, + 376 + ], + "lines": [ + { + "bbox": [ + 313, + 320, + 555, + 376 + ], + "spans": [ + { + "bbox": [ + 313, + 320, + 555, + 376 + ], + "type": "text", + "content": "Table 7. Latency analysis. We evaluate the average token output latency of model outputs on MMLongBench-Doc [51]. RAG-based methods [85, 95] exhibit slower processing speeds due to their two-stage inference process, making them less efficient than document MLLMs for handling multimodal long documents." + } + ] + } + ], + "index": 12, + "angle": 0, + "type": "text" + }, + { + "bbox": [ + 313, + 396, + 555, + 516 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 396, + 555, + 516 + ], + "spans": [ + { + "bbox": [ + 313, + 396, + 555, + 516 + ], + "type": "text", + "content": "Latency Analysis. To compare the latency in inference between RAG methods and our Docopilot, we conducted a latency analysis on MMLongBench-Doc [51], as reported in Table 7. While RAG reduces the document length input to the MLLM, its own time cost remains non-negligible. For instance, InternVL2-2B + RAG is " + }, + { + "bbox": [ + 313, + 396, + 555, + 516 + ], + "type": "inline_equation", + "content": "130\\%" + }, + { + "bbox": [ + 313, + 396, + 555, + 516 + ], + "type": "text", + "content": " slower than InternVL2-2B, and VisRAG is " + }, + { + "bbox": [ + 313, + 396, + 555, + 516 + ], + "type": "inline_equation", + "content": "28\\%" + }, + { + "bbox": [ + 313, + 396, + 555, + 516 + ], + "type": "text", + "content": " slower than MiniCPMV2.6. Our Docopilot does not require additional processes and therefore has the same inference time as baseline models, making it more suitable for analyzing long documents." + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 313, + 526, + 392, + 539 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 526, + 392, + 539 + ], + "spans": [ + { + "bbox": [ + 313, + 526, + 392, + 539 + ], + "type": "text", + "content": "6. Conclusions" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 313, + 545, + 555, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 545, + 555, + 713 + ], + "spans": [ + { + "bbox": [ + 313, + 545, + 555, + 713 + ], + "type": "text", + "content": "This work introduced a diverse document-level question-answering dataset that covers complex structures and crosspage dependencies, providing a robust foundation for training and evaluating document understanding models. We also proposed a retrieval-free long-document understanding model that effectively integrates multi-page information, reducing reliance on external retrieval systems. Experimental results show that our model achieves state-of-the-art performance across several document-level QA benchmarks, underscoring its strength in multi-page integration and complex reasoning. Future work will focus on improving computational efficiency, extending the model to larger multimodal tasks, and adapting it to broader applications for enhanced practicality and generalization." + } + ] + } + ], + "index": 15 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "text", + "content": "4033" + } + ] + } + ], + "index": 16 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 7 + }, + { + "para_blocks": [ + { + "bbox": [ + 56, + 72, + 153, + 85 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 72, + 153, + 85 + ], + "spans": [ + { + "bbox": [ + 56, + 72, + 153, + 85 + ], + "type": "text", + "content": "Acknowledgments" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 55, + 91, + 296, + 163 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 91, + 296, + 163 + ], + "spans": [ + { + "bbox": [ + 55, + 91, + 296, + 163 + ], + "type": "text", + "content": "This project was supported by the National Key R&D Program of China (No. 2022ZD0161300, 2022ZD0160101), the National Natural Science Foundation of China (No. 62376134, 62372223). Zhe Chen is supported by the Youth PhD Student Research Project under the National Natural Science Foundation (No. 623B2050)." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 56, + 173, + 115, + 186 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 173, + 115, + 186 + ], + "spans": [ + { + "bbox": [ + 56, + 173, + 115, + 186 + ], + "type": "text", + "content": "References" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 61, + 193, + 296, + 712 + ], + "type": "list", + "angle": 0, + "index": 12, + "blocks": [ + { + "bbox": [ + 61, + 193, + 296, + 247 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 61, + 193, + 296, + 247 + ], + "spans": [ + { + "bbox": [ + 61, + 193, + 296, + 247 + ], + "type": "text", + "content": "[1] Jean-Baptiste Alayrac, Jeff Donahue, Pauline Luc, Antoine Miech, Iain Barr, Yana Hasson, Karel Lenc, Arthur Mensch, Katherine Millican, Malcolm Reynolds, et al. Flamingo: a visual language model for few-shot learning. NeurIPS, 35: 23716-23736, 2022. 2" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 61, + 248, + 296, + 293 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 61, + 248, + 296, + 293 + ], + "spans": [ + { + "bbox": [ + 61, + 248, + 296, + 293 + ], + "type": "text", + "content": "[2] Chenxin An, Shansan Gong, Ming Zhong, Xingjian Zhao, Mukai Li, Jun Zhang, Lingpeng Kong, and Xipeng Qiu. L-eval: Instituting standardized evaluation for long context language models. arXiv preprint arXiv:2307.11088, 2023. 1" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 61, + 293, + 295, + 346 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 61, + 293, + 295, + 346 + ], + "spans": [ + { + "bbox": [ + 61, + 293, + 295, + 346 + ], + "type": "text", + "content": "[3] Srikar Appalaraju, Peng Tang, Qi Dong, Nishant Sankaran, Yichu Zhou, and R Manmatha. Docformerv2: Local features for document understanding. In Proceedings of the AAAI Conference on Artificial Intelligence, pages 709-718, 2024." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 62, + 348, + 296, + 403 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 348, + 296, + 403 + ], + "spans": [ + { + "bbox": [ + 62, + 348, + 296, + 403 + ], + "type": "text", + "content": "[4] Haoli Bai, Zhiguang Liu, Xiaojun Meng, Wentao Li, Shuang Liu, Nian Xie, Rongfu Zheng, Liangwei Wang, Lu Hou, Jiansheng Wei, et al. Wukong-reader: Multi-modal pretraining for fine-grained visual document understanding. arXiv preprint arXiv:2212.09621, 2022. 2" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 62, + 404, + 296, + 535 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 404, + 296, + 535 + ], + "spans": [ + { + "bbox": [ + 62, + 404, + 296, + 535 + ], + "type": "text", + "content": "[5] Jinze Bai, Shuai Bai, Yunfei Chu, Zeyu Cui, Kai Dang, Xiaodong Deng, Yang Fan, Wenbin Ge, Yu Han, Fei Huang, Binyuan Hui, Luo Ji, Mei Li, Junyang Lin, Runji Lin, Dayiheng Liu, Gao Liu, Chengqiang Lu, Keming Lu, Jianxin Ma, Rui Men, Xingzhang Ren, Xuancheng Ren, Chuanqi Tan, Sinan Tan, Jianhong Tu, Peng Wang, Shijie Wang, Wei Wang, Shengguang Wu, Benfeng Xu, Jin Xu, An Yang, Hao Yang, Jian Yang, Shusheng Yang, Yang Yao, Bowen Yu, Hongyi Yuan, Zheng Yuan, Jianwei Zhang, Xingxuan Zhang, Yichang Zhang, Zhenru Zhang, Chang Zhou, Jingren Zhou, Xiaohuan Zhou, and Tianhang Zhu. Qwen technical report. arXiv preprint arXiv:2309.16609, 2023. 1" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 62, + 536, + 295, + 579 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 536, + 295, + 579 + ], + "spans": [ + { + "bbox": [ + 62, + 536, + 295, + 579 + ], + "type": "text", + "content": "[6] Jinze Bai, Shuai Bai, Shusheng Yang, Shijie Wang, Sinan Tan, Peng Wang, Junyang Lin, Chang Zhou, and Jingren Zhou. Qwen-vl: A frontier large vision-language model with versatile abilities. arXiv preprint arXiv:2308.12966, 2023. 1" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 62, + 581, + 295, + 624 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 581, + 295, + 624 + ], + "spans": [ + { + "bbox": [ + 62, + 581, + 295, + 624 + ], + "type": "text", + "content": "[7] Yushi Bai, Xin Lv, Jiajie Zhang, Yuze He, Ji Qi, Lei Hou, Jie Tang, Yuxiao Dong, and Juanzi Li. Longalign: A recipe for long context alignment of large language models. arXiv preprint arXiv:2401.18058, 2024. 6" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 62, + 625, + 295, + 668 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 625, + 295, + 668 + ], + "spans": [ + { + "bbox": [ + 62, + 625, + 295, + 668 + ], + "type": "text", + "content": "[8] Ali Furkan Biten, Ruben Tito, Andres Mafla, Lluis Gomez, Marcal Rusinol, Ernest Valveny, CV Jawahar, and Dimosthenis Karatzas. Scene text visual question answering. In ICCV, pages 4291-4301, 2019. 6" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 62, + 670, + 295, + 712 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 670, + 295, + 712 + ], + "spans": [ + { + "bbox": [ + 62, + 670, + 295, + 712 + ], + "type": "text", + "content": "[9] Zheng Cai, Maosong Cao, Haojiong Chen, Kai Chen, Keyu Chen, Xin Chen, Xun Chen, Zehui Chen, Zhi Chen, Pei Chu, et al. Internl m2 technical report. arXiv preprint arXiv:2403.17297, 2024. 1, 7" + } + ] + } + ], + "index": 11 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 316, + 73, + 554, + 713 + ], + "type": "list", + "angle": 0, + "index": 27, + "blocks": [ + { + "bbox": [ + 316, + 73, + 554, + 117 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 73, + 554, + 117 + ], + "spans": [ + { + "bbox": [ + 316, + 73, + 554, + 117 + ], + "type": "text", + "content": "[10] Xinlei Chen, Hao Fang, Tsung-Yi Lin, Ramakrishna Vedantam, Saurabh Gupta, Piotr Dólar, and C Lawrence Zitnick. Microsoft coco captions: Data collection and evaluation server. arXiv preprint arXiv:1504.00325, 2015. 1" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 317, + 118, + 554, + 161 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 118, + 554, + 161 + ], + "spans": [ + { + "bbox": [ + 317, + 118, + 554, + 161 + ], + "type": "text", + "content": "[11] Yukang Chen, Shengju Qian, Haotian Tang, Xin Lai, Zhi-jian Liu, Song Han, and Jiaya Jia. Longlora: Efficient finetuning of long-context large language models. arXiv preprint arXiv:2309.12307, 2023. 2, 3, 6" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 316, + 163, + 554, + 218 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 163, + 554, + 218 + ], + "spans": [ + { + "bbox": [ + 316, + 163, + 554, + 218 + ], + "type": "text", + "content": "[12] Zhe Chen, Jiannan Wu, Wenhai Wang, Weijie Su, Guo Chen, Sen Xing, Zhong Muyan, Qinglong Zhang, Xizhou Zhu, Lewei Lu, et al. Internvl: Scaling up vision foundation models and aligning for generic visual-linguistic tasks. arXiv preprint arXiv:2312.14238, 2023. 1, 2" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 316, + 220, + 554, + 274 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 220, + 554, + 274 + ], + "spans": [ + { + "bbox": [ + 316, + 220, + 554, + 274 + ], + "type": "text", + "content": "[13] Zhe Chen, Weiyun Wang, Hao Tian, Shenglong Ye, Zhangwei Gao, Erfei Cui, Wenwen Tong, Kongzhi Hu, Jiapeng Luo, Zheng Ma, et al. How far are we to gpt-4v? closing the gap to commercial multimodal models with open-source suites. arXiv preprint arXiv:2404.16821, 2024. 1, 2, 7, 8" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 316, + 276, + 554, + 319 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 276, + 554, + 319 + ], + "spans": [ + { + "bbox": [ + 316, + 276, + 554, + 319 + ], + "type": "text", + "content": "[14] Jaemin Cho, Debanjan Mahata, Ozan Irsoy, Yujie He, and Mohit Bansal. M3docrag: Multi-modal retrieval is what you need for multi-page multi-document understanding. arXiv preprint arXiv:2411.04952, 2024. 1, 2, 3, 7" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 316, + 320, + 554, + 352 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 320, + 554, + 352 + ], + "spans": [ + { + "bbox": [ + 316, + 320, + 554, + 352 + ], + "type": "text", + "content": "[15] Christopher Clark and Matt Gardner. Simple and effective multi-paragraph reading comprehension. In ACL, pages 845–855, 2018. 6" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 317, + 354, + 554, + 408 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 354, + 554, + 408 + ], + "spans": [ + { + "bbox": [ + 317, + 354, + 554, + 408 + ], + "type": "text", + "content": "[16] Yun Dai, Vignesh Kothapalli, Qingquan Song, Shao Tang, Siyu Zhu, Steven Shimizu, Shivam Sahni, Haowen Ning, Yanning Chen, et al. Liger kernel: Efficient triton kernels for llm training. arXiv preprint arXiv:2410.10989, 2024. 2, 6" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 316, + 411, + 554, + 442 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 411, + 554, + 442 + ], + "spans": [ + { + "bbox": [ + 316, + 411, + 554, + 442 + ], + "type": "text", + "content": "[17] Tri Dao. Flashattention-2: Faster attention with better parallelism and work partitioning. arXiv preprint arXiv:2307.08691, 2023. 3" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 317, + 445, + 554, + 486 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 445, + 554, + 486 + ], + "spans": [ + { + "bbox": [ + 317, + 445, + 554, + 486 + ], + "type": "text", + "content": "[18] Tri Dao, Dan Fu, Stefano Ermon, Atri Rudra, and Christopher Ré. Flashattention: Fast and memory-efficient exact attention with io-awareness. NeurIPS, 35:16344-16359, 2022. 3" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 317, + 489, + 554, + 533 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 489, + 554, + 533 + ], + "spans": [ + { + "bbox": [ + 317, + 489, + 554, + 533 + ], + "type": "text", + "content": "[19] Jiayu Ding, Shuming Ma, Li Dong, Xingxing Zhang, Shaohan Huang, Wenhui Wang, Nanning Zheng, and Furu Wei. Longnet: Scaling transformers to 1,000,000,000 tokens. arXiv preprint arXiv:2307.02486, 2023. 3" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 317, + 535, + 554, + 578 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 535, + 554, + 578 + ], + "spans": [ + { + "bbox": [ + 317, + 535, + 554, + 578 + ], + "type": "text", + "content": "[20] Runpei Dong, Chunrui Han, Yuang Peng, Zekun Qi, Zheng Ge, Jinrong Yang, Liang Zhao, Jianjian Sun, Hongyu Zhou, Haoran Wei, et al. Dreamllm: Synergistic multimodal comprehension and creation. In ICLR, 2024. 2" + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 317, + 580, + 554, + 633 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 580, + 554, + 633 + ], + "spans": [ + { + "bbox": [ + 317, + 580, + 554, + 633 + ], + "type": "text", + "content": "[21] Abhimanyu Dubey, Abhinav Jauhri, Abhinav Pandey, Abhishek Kadian, Ahmad Al-Dahle, Aiesha Letman, Akhil Mathur, Alan Schelten, Amy Yang, Angela Fan, et al. The llama 3 herd of models. arXiv preprint arXiv:2407.21783, 2024. 7" + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 316, + 635, + 554, + 678 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 635, + 554, + 678 + ], + "spans": [ + { + "bbox": [ + 316, + 635, + 554, + 678 + ], + "type": "text", + "content": "[22] Manuel Faysse, Hugues Sibille, Tony Wu, Gautier Viaud, Céline Hudelot, and Pierre Colombo. Colpali: Efficient document retrieval with vision language models. arXiv preprint arXiv:2407.01449, 2024. 1, 3" + } + ] + } + ], + "index": 25 + }, + { + "bbox": [ + 316, + 681, + 554, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 681, + 554, + 713 + ], + "spans": [ + { + "bbox": [ + 316, + 681, + 554, + 713 + ], + "type": "text", + "content": "[23] Hao Feng, Qi Liu, Hao Liu, Wengang Zhou, Houqiang Li, and Can Huang. Docpedia: Unleashing the power of large multimodal model in the frequency domain for versatile" + } + ] + } + ], + "index": 26 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "text", + "content": "4034" + } + ] + } + ], + "index": 28 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 8 + }, + { + "para_blocks": [ + { + "bbox": [ + 56, + 72, + 295, + 712 + ], + "type": "list", + "angle": 0, + "index": 13, + "blocks": [ + { + "bbox": [ + 76, + 72, + 294, + 94 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 76, + 72, + 294, + 94 + ], + "spans": [ + { + "bbox": [ + 76, + 72, + 294, + 94 + ], + "type": "text", + "content": "document understanding. arXiv preprint arXiv:2311.11810, 2023.3" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 56, + 95, + 295, + 149 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 95, + 295, + 149 + ], + "spans": [ + { + "bbox": [ + 56, + 95, + 295, + 149 + ], + "type": "text", + "content": "[24] Chaoyou Fu, Peixian Chen, Yunhang Shen, Yulei Qin, Mengdan Zhang, Xu Lin, Zhenyu Qiu, Wei Lin, Jinrui Yang, Xiawu Zheng, et al. Mme: A comprehensive evaluation benchmark for multimodal large language models. arXiv preprint arXiv:2306.13394, 2023. 1" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 56, + 150, + 294, + 205 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 150, + 294, + 205 + ], + "spans": [ + { + "bbox": [ + 56, + 150, + 294, + 205 + ], + "type": "text", + "content": "[25] Zhangwei Gao, Zhe Chen, Erfei Cui, Yiming Ren, Weiyun Wang, Jinguo Zhu, Hao Tian, Shenglong Ye, Junjun He, Xizhou Zhu, et al. Mini-internvl: A flexible-transfer pocket multimodal model with " + }, + { + "bbox": [ + 56, + 150, + 294, + 205 + ], + "type": "inline_equation", + "content": "5\\%" + }, + { + "bbox": [ + 56, + 150, + 294, + 205 + ], + "type": "text", + "content": " parameters and " + }, + { + "bbox": [ + 56, + 150, + 294, + 205 + ], + "type": "inline_equation", + "content": "90\\%" + }, + { + "bbox": [ + 56, + 150, + 294, + 205 + ], + "type": "text", + "content": " performance. arXiv preprint arXiv:2410.16261, 2024. 2" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 56, + 206, + 294, + 248 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 206, + 294, + 248 + ], + "spans": [ + { + "bbox": [ + 56, + 206, + 294, + 248 + ], + "type": "text", + "content": "[26] Chi Han, Qifan Wang, Wenhan Xiong, Yu Chen, Heng Ji, and Sinong Wang. Lm-infinite: Simple on-the-fly length generalization for large language models. arXiv preprint arXiv:2308.16137, 2023. 3" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 56, + 250, + 294, + 303 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 250, + 294, + 303 + ], + "spans": [ + { + "bbox": [ + 56, + 250, + 294, + 303 + ], + "type": "text", + "content": "[27] Anwen Hu, Haiyang Xu, Liang Zhang, Jiabo Ye, Ming Yan, Ji Zhang, Qin Jin, Fei Huang, and Jingren Zhou. mplug-docowl2: High-resolution compressing for ocr-free multi-page document understanding. arXiv preprint arXiv:2409.03420, 2024. 7, 2" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 56, + 304, + 294, + 347 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 304, + 294, + 347 + ], + "spans": [ + { + "bbox": [ + 56, + 304, + 294, + 347 + ], + "type": "text", + "content": "[28] Mingxin Huang, Yuliang Liu, Dingkang Liang, Lianwen Jin, and Xiang Bai. Mini-monkey: Alleviate the saw-tooth effect by multi-scale adaptive cropping. arXiv preprint arXiv:2408.02034, 2024. 7" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 56, + 350, + 294, + 424 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 350, + 294, + 424 + ], + "spans": [ + { + "bbox": [ + 56, + 350, + 294, + 424 + ], + "type": "text", + "content": "[29] Gabriel Ilharco, Mitchell Wortsman, Ross Wightman, Cade Gordon, Nicholas Carlini, Rohan Taori, Achal Dave, Vaishaal Shankar, Hongseok Namkoong, John Miller, Hannaneh Hajishirzi, Ali Farhadi, and Ludwig Schmidt. Openclip. Zenodo. Version 0.1. https://doi.org/10.5281/zenodo.5143773, 2021. DOI: 10.5281/zenodo.5143773. 2" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 56, + 426, + 294, + 469 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 426, + 294, + 469 + ], + "spans": [ + { + "bbox": [ + 56, + 426, + 294, + 469 + ], + "type": "text", + "content": "[30] Yao Jiang, Xinyu Yan, Ge-Peng Ji, Keren Fu, Meijun Sun, Huan Xiong, Deng-Ping Fan, and Fahad Shahbaz Khan. Effectiveness assessment of recent large vision-language models. Visual Intelligence, 2(1):17, 2024. 1" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 56, + 471, + 294, + 503 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 471, + 294, + 503 + ], + "spans": [ + { + "bbox": [ + 56, + 471, + 294, + 503 + ], + "type": "text", + "content": "[31] Greg Kamradt. Llmtest_needleinahaystack. https://github.com/gkamradt/LLMTest_NeedleInAHaystack, 2024. Accessed: 2024-11-11. 1" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 56, + 504, + 294, + 568 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 504, + 294, + 568 + ], + "spans": [ + { + "bbox": [ + 56, + 504, + 294, + 568 + ], + "type": "text", + "content": "[32] Geewook Kim, Teakgyu Hong, Moonbin Yim, JeongYeon Nam, Jinyoung Park, Jinyeong Yim, Wonseok Hwang, Sangdoo Yun, Dongyoon Han, and Seunghyun Park. Ocr-free document understanding transformer. In European Conference on Computer Vision, pages 498-517. Springer, 2022. 3" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 56, + 570, + 294, + 613 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 570, + 294, + 613 + ], + "spans": [ + { + "bbox": [ + 56, + 570, + 294, + 613 + ], + "type": "text", + "content": "[33] Hugo Laurencon, Andres Marafioti, Victor Sanh, and Léo Tronchon. Building and better understanding vision-language models: insights and future directions. arXiv preprint arXiv:2408.12637, 2024. 2, 6" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 56, + 614, + 294, + 668 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 614, + 294, + 668 + ], + "spans": [ + { + "bbox": [ + 56, + 614, + 294, + 668 + ], + "type": "text", + "content": "[34] Hugo Laurençon, Lucile Saulnier, Léo Tronchon, Stas Bekman, Amanpreet Singh, Anton Lozhkov, Thomas Wang, Siddharth Karamcheti, Alexander Rush, Douwe Kiela, et al. Obelics: An open web-scale filtered dataset of interleaved image-text documents. NIPS, 36, 2024. 2" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 56, + 670, + 294, + 712 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 670, + 294, + 712 + ], + "spans": [ + { + "bbox": [ + 56, + 670, + 294, + 712 + ], + "type": "text", + "content": "[35] Bohao Li, Rui Wang, Guangzhi Wang, Yuying Ge, Yixiao Ge, and Ying Shan. Seed-bench: Benchmarking multimodal llms with generative comprehension. arXiv preprint arXiv:2307.16125, 2023. 2" + } + ] + } + ], + "index": 12 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 316, + 73, + 553, + 713 + ], + "type": "list", + "angle": 0, + "index": 28, + "blocks": [ + { + "bbox": [ + 316, + 73, + 553, + 116 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 73, + 553, + 116 + ], + "spans": [ + { + "bbox": [ + 316, + 73, + 553, + 116 + ], + "type": "text", + "content": "[36] Bo Li, Yuanhan Zhang, Dong Guo, Renrui Zhang, Feng Li, Hao Zhang, Kaichen Zhang, Yanwei Li, Ziwei Liu, and Chunyuan Li. Llava-onevision: Easy visual task transfer. arXiv preprint arXiv:2408.03326, 2024. 1, 7" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 316, + 118, + 553, + 161 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 118, + 553, + 161 + ], + "spans": [ + { + "bbox": [ + 316, + 118, + 553, + 161 + ], + "type": "text", + "content": "[37] Junnan Li, Dongxu Li, Caiming Xiong, and Steven Hoi. Blip: Bootstrapping language-image pre-training for unified vision-language understanding and generation. In ICML, pages 12888-12900, 2022. 2" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 316, + 163, + 553, + 206 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 163, + 553, + 206 + ], + "spans": [ + { + "bbox": [ + 316, + 163, + 553, + 206 + ], + "type": "text", + "content": "[38] Junnan Li, Dongxu Li, Silvio Savarese, and Steven Hoi. Blip-2: Bootstrapping language-image pre-training with frozen image encoders and large language models. In ICML, pages 19730–19742. PMLR, 2023. 2" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 316, + 208, + 553, + 262 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 208, + 553, + 262 + ], + "spans": [ + { + "bbox": [ + 316, + 208, + 553, + 262 + ], + "type": "text", + "content": "[39] Qingyun Li, Zhe Chen, Weiyun Wang, Wenhai Wang, Shenglong Ye, Zhenjiang Jin, Guanzhou Chen, Yinan He, Zhangwei Gao, Erfei Cui, et al. Omnicorpus: An unified multimodal corpus of 10 billion-level images interleaved with text. arXiv preprint arXiv:2406.08418, 2024. 2" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 316, + 264, + 553, + 317 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 264, + 553, + 317 + ], + "spans": [ + { + "bbox": [ + 316, + 264, + 553, + 317 + ], + "type": "text", + "content": "[40] Zhang Li, Biao Yang, Qiang Liu, Zhiyin Ma, Shuo Zhang, Jingxu Yang, Yabo Sun, Yuliang Liu, and Xiang Bai. Monkey: Image resolution and text label are important things for large multi-modal models. arXiv preprint arXiv:2311.06607, 2023. 7" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 316, + 319, + 553, + 353 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 319, + 553, + 353 + ], + "spans": [ + { + "bbox": [ + 316, + 319, + 553, + 353 + ], + "type": "text", + "content": "[41] Haotian Liu, Chunyuan Li, Yuheng Li, and Yong Jae Lee. Improved baselines with visual instruction tuning. arXiv preprint arXiv:2310.03744, 2023. 1, 6" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 316, + 354, + 553, + 376 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 354, + 553, + 376 + ], + "spans": [ + { + "bbox": [ + 316, + 354, + 553, + 376 + ], + "type": "text", + "content": "[42] Haotian Liu, Chunyuan Li, Qingyang Wu, and Yong Jae Lee. Visual instruction tuning. NeurIPS, 36, 2023. 1, 2, 6" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 316, + 377, + 553, + 410 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 377, + 553, + 410 + ], + "spans": [ + { + "bbox": [ + 316, + 377, + 553, + 410 + ], + "type": "text", + "content": "[43] Hao Liu, Matei Zaharia, and Pieter Abbeel. Ring attention with blockwise transformers for near-infinite context. arXiv preprint arXiv:2310.01889, 2023. 2, 3, 6" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 316, + 411, + 553, + 463 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 411, + 553, + 463 + ], + "spans": [ + { + "bbox": [ + 316, + 411, + 553, + 463 + ], + "type": "text", + "content": "[44] Yuan Liu, Haodong Duan, Yuanhan Zhang, Bo Li, Songyang Zhang, Wangbo Zhao, Yike Yuan, Jiaqi Wang, Conghui He, Ziwei Liu, et al. Mmbench: Is your multi-modal model an all-around player? arXiv preprint arXiv:2307.06281, 2023. 1" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 316, + 467, + 553, + 520 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 467, + 553, + 520 + ], + "spans": [ + { + "bbox": [ + 316, + 467, + 553, + 520 + ], + "type": "text", + "content": "[45] Yuliang Liu, Zhang Li, Hongliang Li, Wenwen Yu, Mingxin Huang, Dezhi Peng, Mingyu Liu, Mingrui Chen, Chunyuan Li, Lianwen Jin, et al. On the hidden mystery ofOCR in large multimodal models. arXiv preprint arXiv:2305.07895, 2023. 1" + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 316, + 523, + 553, + 567 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 523, + 553, + 567 + ], + "spans": [ + { + "bbox": [ + 316, + 523, + 553, + 567 + ], + "type": "text", + "content": "[46] Yuliang Liu, Biao Yang, Qiang Liu, Zhang Li, Zhiyin Ma, Shuo Zhang, and Xiang Bai. Textmonkey: AnOCR-free large multimodal model for understanding document. arXiv preprint arXiv:2403.04473, 2024. 7" + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 316, + 568, + 553, + 623 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 568, + 553, + 623 + ], + "spans": [ + { + "bbox": [ + 316, + 568, + 553, + 623 + ], + "type": "text", + "content": "[47] Zhaoyang Liu, Yinan He, Wenhai Wang, Weiyun Wang, Yi Wang, Shoufa Chen, Qinglong Zhang, Zeqiang Lai, Yang Yang, Qingyun Li, Jiashuo Yu, et al. Interrupt: Solving vision-centric tasks by interacting with chatgpt beyond language. arXiv preprint arXiv:2305.05662, 2023. 2" + } + ] + } + ], + "index": 25 + }, + { + "bbox": [ + 316, + 624, + 553, + 678 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 624, + 553, + 678 + ], + "spans": [ + { + "bbox": [ + 316, + 624, + 553, + 678 + ], + "type": "text", + "content": "[48] Ziyu Liu, Tao Chu, Yuhang Zang, Xilin Wei, Xiaoyi Dong, Pan Zhang, Zijian Liang, Yuanjun Xiong, Yu Qiao, Dahua Lin, et al. Mmdu: A multi-turn multi-image dialog understanding benchmark and instruction-tuning dataset for lvlms. arXiv preprint arXiv:2406.11833, 2024. 6" + } + ] + } + ], + "index": 26 + }, + { + "bbox": [ + 316, + 680, + 553, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 680, + 553, + 713 + ], + "spans": [ + { + "bbox": [ + 316, + 680, + 553, + 713 + ], + "type": "text", + "content": "[49] Chuwei Luo, Yufan Shen, Zhaoqing Zhu, Qi Zheng, Zhi Yu, and Cong Yao. Layout: Layout instruction tuning with large language models for document understanding. In Pro" + } + ] + } + ], + "index": 27 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "text", + "content": "4035" + } + ] + } + ], + "index": 29 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 9 + }, + { + "para_blocks": [ + { + "bbox": [ + 56, + 73, + 295, + 713 + ], + "type": "list", + "angle": 0, + "index": 15, + "blocks": [ + { + "bbox": [ + 77, + 73, + 294, + 95 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 77, + 73, + 294, + 95 + ], + "spans": [ + { + "bbox": [ + 77, + 73, + 294, + 95 + ], + "type": "text", + "content": "ceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 15630-15640, 2024. 3" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 56, + 97, + 295, + 140 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 97, + 295, + 140 + ], + "spans": [ + { + "bbox": [ + 56, + 97, + 295, + 140 + ], + "type": "text", + "content": "[50] Xueguang Ma, Sheng-Chieh Lin, Minghan Li, Wenhu Chen, and Jimmy Lin. Unifying multimodal retrieval via document screenshot embedding. arXiv preprint arXiv:2406.11251, 2024. 1" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 56, + 143, + 294, + 208 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 143, + 294, + 208 + ], + "spans": [ + { + "bbox": [ + 56, + 143, + 294, + 208 + ], + "type": "text", + "content": "[51] Yubo Ma, Yuhang Zang, Liangyu Chen, Meiqi Chen, Yizhu Jiao, Xinze Li, Xinyuan Lu, Ziyu Liu, Yan Ma, Xiaoyi Dong, Pan Zhang, Liangming Pan, Yu-Gang Jiang, Jiaqi Wang, Yixin Cao, and Aixin Sun. Mmlongbench-doc: Benchmarking long-context document understanding with visualizations, 2024. 1, 7, 8, 5" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 56, + 210, + 294, + 255 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 210, + 294, + 255 + ], + "spans": [ + { + "bbox": [ + 56, + 210, + 294, + 255 + ], + "type": "text", + "content": "[52] Ahmed Masry, Xuan Long Do, Jia Qing Tan, Shafiq Joty, and Enamul Hoque. Chartqa: A benchmark for question answering about charts with visual and logical reasoning. In ACL, pages 2263-2279, 2022. 6, 8, 1, 5" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 56, + 256, + 294, + 289 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 256, + 294, + 289 + ], + "spans": [ + { + "bbox": [ + 56, + 256, + 294, + 289 + ], + "type": "text", + "content": "[53] Minesh Mathew, Dimosthenis Karatzas, and CV Jawahar. Docvqa: A dataset for vqa on document images. In WACV, pages 2200-2209, 2021. 6, 8, 1, 5" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 56, + 291, + 294, + 323 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 291, + 294, + 323 + ], + "spans": [ + { + "bbox": [ + 56, + 291, + 294, + 323 + ], + "type": "text", + "content": "[54] Minesh Mathew, Viraj Bagal, Ruben Tito, Dimosthenis Karatzas, Ernest Valveny, and CV Jawahar. Infographicvqa. In WACV, pages 1697-1706, 2022. 6, 8, 1, 5" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 56, + 326, + 294, + 369 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 326, + 294, + 369 + ], + "spans": [ + { + "bbox": [ + 56, + 326, + 294, + 369 + ], + "type": "text", + "content": "[55] Anand Mishra, Shashank Shekhar, Ajeet Kumar Singh, and Anirban Chakraborty. Ocr-vqa: Visual question answering by reading text in images. In ICDAR, pages 947-952, 2019. 6" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 56, + 371, + 294, + 403 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 371, + 294, + 403 + ], + "spans": [ + { + "bbox": [ + 56, + 371, + 294, + 403 + ], + "type": "text", + "content": "[56] OpenAI. Gpt-4v(ison) system card. https://cdn.openai.com/papers/GPTV_System/Card.pdf, 2023.1,4,7" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 56, + 406, + 294, + 429 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 406, + 294, + 429 + ], + "spans": [ + { + "bbox": [ + 56, + 406, + 294, + 429 + ], + "type": "text", + "content": "[57] OpenAI. Gpt-4o system card. https://openai.com/index/gpt-4o-system-card/, 2024.7" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 56, + 430, + 294, + 463 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 430, + 294, + 463 + ], + "spans": [ + { + "bbox": [ + 56, + 430, + 294, + 463 + ], + "type": "text", + "content": "[58] Ofir Press, Noah A Smith, and Mike Lewis. Train short, test long: Attention with linear biases enables input length extrapolation. arXiv preprint arXiv:2108.12409, 2021. 3" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 56, + 464, + 294, + 497 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 464, + 294, + 497 + ], + "spans": [ + { + "bbox": [ + 56, + 464, + 294, + 497 + ], + "type": "text", + "content": "[59] Sylvestre-Alvise Rebuffi, Hakan Bilen, and Andrea Vedaldi. Learning multiple visual domains with residual adapters. NeurIPS, 30, 2017. 2" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 56, + 499, + 294, + 564 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 499, + 294, + 564 + ], + "spans": [ + { + "bbox": [ + 56, + 499, + 294, + 564 + ], + "type": "text", + "content": "[60] Machel Reid, Nikolay Savinov, Denis Teptyashin, Dmitry Lepikhin, Timothy Lillicrap, Jean-baptiste Alayrac, Radu Soricut, Angeliki Lazaridou, Orhan Firat, Julian Schrittwieser, et al. Gemini 1.5: Unlocking multimodal understanding across millions of tokens of context. arXiv preprint arXiv:2403.05530, 2024. 1, 7" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 56, + 567, + 294, + 610 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 567, + 294, + 610 + ], + "spans": [ + { + "bbox": [ + 56, + 567, + 294, + 610 + ], + "type": "text", + "content": "[61] Sahel Sharifymoghaddam, Shivani Upadhyay, Wenhu Chen, and Jimmy Lin. Unirag: Universal retrieval augmentation for multi-modal large language models. arXiv preprint arXiv:2405.10311, 2024. 1" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 56, + 613, + 294, + 667 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 613, + 294, + 667 + ], + "spans": [ + { + "bbox": [ + 56, + 613, + 294, + 667 + ], + "type": "text", + "content": "[62] Min Shi, Fuxiao Liu, Shihao Wang, Shijia Liao, Subhashree Radhakrishnan, De-An Huang, Hongxu Yin, Karan Sapra, Yaser Yacoob, Humphrey Shi, et al. Eagle: Exploring the design space for multimodal llms with mixture of encoders. arXiv preprint arXiv:2408.15998, 2024. 1" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 56, + 670, + 294, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 670, + 294, + 713 + ], + "spans": [ + { + "bbox": [ + 56, + 670, + 294, + 713 + ], + "type": "text", + "content": "[63] Weijia Shi, Sewon Min, Michihiro Yasunaga, Minjoon Seo, Rich James, Mike Lewis, Luke Zettlemoyer, and Wen-tau Yih. Replug: Retrieval-augmented black-box language models. arXiv preprint arXiv:2301.12652, 2023. 1" + } + ] + } + ], + "index": 14 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 316, + 73, + 553, + 713 + ], + "type": "list", + "angle": 0, + "index": 31, + "blocks": [ + { + "bbox": [ + 316, + 73, + 553, + 116 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 73, + 553, + 116 + ], + "spans": [ + { + "bbox": [ + 316, + 73, + 553, + 116 + ], + "type": "text", + "content": "[64] Oleksii Sidorov, Ronghang Hu, Marcus Rohrbach, and Amanpreet Singh. Textcaps: a dataset for image captioning with reading comprehension. In ECCV, pages 742-758, 2020. 6" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 316, + 118, + 553, + 162 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 118, + 553, + 162 + ], + "spans": [ + { + "bbox": [ + 316, + 118, + 553, + 162 + ], + "type": "text", + "content": "[65] Amanpreet Singh, Vivek Natarajan, Meet Shah, Yu Jiang, Xinlei Chen, Dhruv Batra, Devi Parikh, and Marcus Rohrbach. Towards vqa models that can read. In CVPR, pages 8317-8326, 2019. 1, 6" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 316, + 163, + 553, + 205 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 163, + 553, + 205 + ], + "spans": [ + { + "bbox": [ + 316, + 163, + 553, + 205 + ], + "type": "text", + "content": "[66] Jianlin Su, Murtadha Ahmed, Yu Lu, Shengfeng Pan, Wen Bo, and Yunfeng Liu. Roformer: Enhanced transformer with rotary position embedding. Neurocomputing, 568:127063, 2024. 3" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 316, + 208, + 553, + 251 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 208, + 553, + 251 + ], + "spans": [ + { + "bbox": [ + 316, + 208, + 553, + 251 + ], + "type": "text", + "content": "[67] Quan Sun, Qiying Yu, Yufeng Cui, Fan Zhang, Xiaosong Zhang, Yueze Wang, Hongcheng Gao, Jingjing Liu, Tiejun Huang, and Xinlong Wang. Generative pretraining in multimodality. In ICLR, 2024. 2" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 316, + 253, + 553, + 297 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 253, + 553, + 297 + ], + "spans": [ + { + "bbox": [ + 316, + 253, + 553, + 297 + ], + "type": "text", + "content": "[68] Yutao Sun, Li Dong, Barun Patra, Shuming Ma, Shaohan Huang, Alon Benhaim, Vishrav Chaudhary, Xia Song, and Furu Wei. A length-extrapolatable transformer. arXiv preprint arXiv:2212.10554, 2022. 3" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 316, + 298, + 553, + 319 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 298, + 553, + 319 + ], + "spans": [ + { + "bbox": [ + 316, + 298, + 553, + 319 + ], + "type": "text", + "content": "[69] Mohammad Reza Taesiri. Arxivqa. https://github.com/taesiri/ArXivQA, 2024.6" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 316, + 321, + 553, + 365 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 321, + 553, + 365 + ], + "spans": [ + { + "bbox": [ + 316, + 321, + 553, + 365 + ], + "type": "text", + "content": "[70] Ryota Tanaka, Kyosuke Nishida, and Sen Yoshida. Visualmrc: Machine reading comprehension on document images. In Proceedings of the AAAI Conference on Artificial Intelligence, pages 13878-13888, 2021. 6" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 316, + 366, + 553, + 430 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 366, + 553, + 430 + ], + "spans": [ + { + "bbox": [ + 316, + 366, + 553, + 430 + ], + "type": "text", + "content": "[71] Zineng Tang, Ziyi Yang, Guoxin Wang, Yuwei Fang, Yang Liu, Chenguang Zhu, Michael Zeng, Cha Zhang, and Mohit Bansal. Unifying vision, text, and layout for universal document processing. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 19254-19264, 2023. 2" + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 316, + 433, + 553, + 487 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 433, + 553, + 487 + ], + "spans": [ + { + "bbox": [ + 316, + 433, + 553, + 487 + ], + "type": "text", + "content": "[72] Gemini Team, Rohan Anil, Sebastian Borgeaud, Yonghui Wu, Jean-Baptiste Alayrac, Jiahui Yu, Radu Soricut, Johan Schalkwyk, Andrew M Dai, Anja Hauth, et al. Gemini: a family of highly capable multimodal models. arXiv preprint arXiv:2312.11805, 2023. 1" + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 316, + 488, + 553, + 532 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 488, + 553, + 532 + ], + "spans": [ + { + "bbox": [ + 316, + 488, + 553, + 532 + ], + "type": "text", + "content": "[73] OpenGVLab Team. Internv12: Better than the best—expanding performance boundaries of open-source multimodal models with the progressive scaling strategy, 2024. 1, 6, 7, 8" + } + ] + } + ], + "index": 25 + }, + { + "bbox": [ + 316, + 533, + 553, + 555 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 533, + 553, + 555 + ], + "spans": [ + { + "bbox": [ + 316, + 533, + 553, + 555 + ], + "type": "text", + "content": "[74] Qwen Team. Qwen2.5: A party of foundation models, 2024. 1" + } + ] + } + ], + "index": 26 + }, + { + "bbox": [ + 316, + 557, + 553, + 610 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 557, + 553, + 610 + ], + "spans": [ + { + "bbox": [ + 316, + 557, + 553, + 610 + ], + "type": "text", + "content": "[75] Changyao Tian, Xizhou Zhu, Yuwen Xiong, Weiyun Wang, Zhe Chen, Wenhai Wang, Yuntao Chen, Lewei Lu, Tong Lu, Jie Zhou, et al. Mm-interleaved: Interleaved image-text generative modeling via multi-modal feature synchronizer. arXiv preprint arXiv:2401.10208, 2024. 2" + } + ] + } + ], + "index": 27 + }, + { + "bbox": [ + 316, + 613, + 553, + 644 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 613, + 553, + 644 + ], + "spans": [ + { + "bbox": [ + 316, + 613, + 553, + 644 + ], + "type": "text", + "content": "[76] Rubén Tito, Dimosthenis Karatzas, and Ernest Valveny. Hierarchical multimodal transformers for multipage docvqa. Pattern Recognition, 144:109834, 2023. 1, 2, 7, 5" + } + ] + } + ], + "index": 28 + }, + { + "bbox": [ + 316, + 647, + 553, + 679 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 647, + 553, + 679 + ], + "spans": [ + { + "bbox": [ + 316, + 647, + 553, + 679 + ], + "type": "text", + "content": "[77] Xiaoguang Tu, Zhi He, Yi Huang, Zhi-Hao Zhang, Ming Yang, and Jian Zhao. An overview of large ai models and their applications. Visual Intelligence, 2(1):1-22, 2024. 1" + } + ] + } + ], + "index": 29 + }, + { + "bbox": [ + 316, + 681, + 553, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 681, + 553, + 713 + ], + "spans": [ + { + "bbox": [ + 316, + 681, + 553, + 713 + ], + "type": "text", + "content": "[78] Jordy Van Landeghem, Ruben Tito, Łukasz Borchmann, Michal Pietruszka, Paweł Joziak, Rafał Powalski, Dawid Jurkiewicz, Mickael Coustaty, Bertrand Ackaert, Ernest Val" + } + ] + } + ], + "index": 30 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 748, + 316, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 316, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 316, + 757 + ], + "type": "text", + "content": "4036" + } + ] + } + ], + "index": 32 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 10 + }, + { + "para_blocks": [ + { + "bbox": [ + 56, + 72, + 296, + 712 + ], + "type": "list", + "angle": 0, + "index": 12, + "blocks": [ + { + "bbox": [ + 76, + 72, + 296, + 116 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 76, + 72, + 296, + 116 + ], + "spans": [ + { + "bbox": [ + 76, + 72, + 296, + 116 + ], + "type": "text", + "content": "veny, et al. Document understanding dataset and evaluation (dude). In Proceedings IEEE/CVF international conference on computer vision-ICCV 2023, pages 19528-19540. IEEE/CVF, 2023. 2, 6" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 56, + 118, + 296, + 173 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 118, + 296, + 173 + ], + "spans": [ + { + "bbox": [ + 56, + 118, + 296, + 173 + ], + "type": "text", + "content": "[79] Bin Wang, Chao Xu, Xiaomeng Zhao, Linke Ouyang, Fan Wu, Zhiyuan Zhao, Rui Xu, Kaiwen Liu, Yuan Qu, Fukai Shang, et al. Mineru: An open-source solution for precise document content extraction. arXiv preprint arXiv:2409.18839, 2024. 3" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 56, + 175, + 295, + 229 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 175, + 295, + 229 + ], + "spans": [ + { + "bbox": [ + 56, + 175, + 295, + 229 + ], + "type": "text", + "content": "[80] Dongsheng Wang, Natraj Raman, Mathieu Sibue, Zhiqiang Ma, Petr Babkin, Simerjot Kaur, Yulong Pei, Armineh Nourbakhsh, and Xiaomo Liu. Docllm: A layout-aware generative language model for multimodal document understanding. arXiv preprint arXiv:2401.00908, 2023. 2" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 56, + 230, + 295, + 285 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 230, + 295, + 285 + ], + "spans": [ + { + "bbox": [ + 56, + 230, + 295, + 285 + ], + "type": "text", + "content": "[81] Peng Wang, Shuai Bai, Sinan Tan, Shijie Wang, Zhihao Fan, Jinze Bai, Keqin Chen, Xuejing Liu, Jialin Wang, Wenbin Ge, et al. Qwen2-vl: Enhancing vision-language model's perception of the world at any resolution. arXiv preprint arXiv:2409.12191, 2024. 1" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 56, + 286, + 295, + 330 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 286, + 295, + 330 + ], + "spans": [ + { + "bbox": [ + 56, + 286, + 295, + 330 + ], + "type": "text", + "content": "[82] Weihan Wang, Qingsong Lv, Wenmeng Yu, Wenyi Hong, Ji Qi, Yan Wang, Junhui Ji, Zhuoyi Yang, Lei Zhao, Xixuan Song, et al. Cogvlm: Visual expert for pretrained language models. arXiv preprint arXiv:2311.03079, 2023. 1" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 56, + 332, + 295, + 387 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 332, + 295, + 387 + ], + "spans": [ + { + "bbox": [ + 56, + 332, + 295, + 387 + ], + "type": "text", + "content": "[83] Weiyun Wang, Yiming Ren, Haowen Luo, Tiantong Li, Chenxiang Yan, Zhe Chen, Wenhai Wang, Qingyun Li, Lewei Lu, Xizhou Zhu, et al. The all-seeing project v2: Towards general relation comprehension of the open world. arXiv preprint arXiv:2402.19474, 2024. 1, 2" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 56, + 388, + 295, + 442 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 388, + 295, + 442 + ], + "spans": [ + { + "bbox": [ + 56, + 388, + 295, + 442 + ], + "type": "text", + "content": "[84] Weiyun Wang, Min Shi, Qingyun Li, Wenhai Wang, Zhenhang Huang, Linjie Xing, Zhe Chen, Hao Li, Xizhou Zhu, Zhiguo Cao, et al. The all-seeing project: Towards panoptic visual recognition and understanding of the open world. In ICLR, 2024. 1, 2" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 56, + 445, + 295, + 499 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 445, + 295, + 499 + ], + "spans": [ + { + "bbox": [ + 56, + 445, + 295, + 499 + ], + "type": "text", + "content": "[85] Weiyun Wang, Shuibio Zhang, Yiming Ren, Yuchen Duan, Tiantong Li, Shuo Liu, Mengkang Hu, Zhe Chen, Kaipeng Zhang, Lewei Lu, Xizhou Zhu, Ping Luo, Yu Qiao, Jifeng Dai, Wenqi Shao, and Wenhai Wang. Needle in a multimodal haystack. arXiv preprint arXiv:2406.07230, 2024. 7, 8, 3" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 56, + 501, + 295, + 544 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 501, + 295, + 544 + ], + "spans": [ + { + "bbox": [ + 56, + 501, + 295, + 544 + ], + "type": "text", + "content": "[86] Weiyun Wang, Shuibio Zhang, Yiming Ren, Yuchen Duan, Tiantong Li, Shuo Liu, Mengkang Hu, Zhe Chen, Kaipeng Zhang, Lewei Lu, et al. Needle in a multimodal haystack. arXiv preprint arXiv:2406.07230, 2024. 2, 7, 8, 5" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 56, + 546, + 295, + 589 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 546, + 295, + 589 + ], + "spans": [ + { + "bbox": [ + 56, + 546, + 295, + 589 + ], + "type": "text", + "content": "[87] Yonghui Wang, Wengang Zhou, Hao Feng, Keyi Zhou, and Houqiang Li. Towards improving document understanding: An exploration on text-grounding via mllms. arXiv preprint arXiv:2311.13194, 2023. 3" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 56, + 591, + 295, + 644 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 591, + 295, + 644 + ], + "spans": [ + { + "bbox": [ + 56, + 591, + 295, + 644 + ], + "type": "text", + "content": "[88] Haoran Wei, Lingyu Kong, Jinyue Chen, Liang Zhao, Zheng Ge, Jinrong Yang, Jianjian Sun, Chunrui Han, and Xiangyu Zhang. Vary: Scaling up the vision vocabulary for large vision-language models. arXiv preprint arXiv:2312.06109, 2023. 3" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 56, + 647, + 295, + 712 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 647, + 295, + 712 + ], + "spans": [ + { + "bbox": [ + 56, + 647, + 295, + 712 + ], + "type": "text", + "content": "[89] Renqiu Xia, Song Mao, Xiangchao Yan, Hongbin Zhou, Bo Zhang, Haoyang Peng, Jiahao Pi, Daocheng Fu, Wenjie Wu, Hancheng Ye, et al. Docgenome: An open largescale scientific document benchmark for training and testing multi-modal large language models. arXiv preprint arXiv:2406.11633, 2024. 1, 7, 2, 5" + } + ] + } + ], + "index": 11 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 316, + 72, + 553, + 574 + ], + "type": "list", + "angle": 0, + "index": 23, + "blocks": [ + { + "bbox": [ + 316, + 72, + 553, + 106 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 72, + 553, + 106 + ], + "spans": [ + { + "bbox": [ + 316, + 72, + 553, + 106 + ], + "type": "text", + "content": "[90] Guangxuan Xiao, Yuandong Tian, Beidi Chen, Song Han, and Mike Lewis. Efficient streaming language models with attention sinks. arXiv preprint arXiv:2309.17453, 2023. 3" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 316, + 107, + 553, + 140 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 107, + 553, + 140 + ], + "spans": [ + { + "bbox": [ + 316, + 107, + 553, + 140 + ], + "type": "text", + "content": "[91] Jianxin Yang. Longqlora: Efficient and effective method to extend context length of large language models. arXiv preprint arXiv:2311.04879, 2023. 2, 6" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 316, + 141, + 553, + 184 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 141, + 553, + 184 + ], + "spans": [ + { + "bbox": [ + 316, + 141, + 553, + 184 + ], + "type": "text", + "content": "[92] Yuan Yao, Tianyu Yu, Ao Zhang, Chongyi Wang, Junbo Cui, Hongji Zhu, Tianchi Cai, Haoyu Li, Weilin Zhao, Zhihui He, et al. Minicpm-v: A gpt-4v level mllm on your phone. arXiv preprint arXiv:2408.01800, 2024. 1, 7, 8" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 316, + 186, + 553, + 239 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 186, + 553, + 239 + ], + "spans": [ + { + "bbox": [ + 316, + 186, + 553, + 239 + ], + "type": "text", + "content": "[93] Jiabo Ye, Anwen Hu, Haiyang Xu, Qinghao Ye, Ming Yan, Yuhao Dan, Chenlin Zhao, Guohai Xu, Chenliang Li, Junfeng Tian, et al. mplug-docowl: Modularized multimodal large language model for document understanding. arXiv preprint arXiv:2307.02499, 2023. 6" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 316, + 241, + 553, + 296 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 241, + 553, + 296 + ], + "spans": [ + { + "bbox": [ + 316, + 241, + 553, + 296 + ], + "type": "text", + "content": "[94] Jiabo Ye, Anwen Hu, Haiyang Xu, Qinghao Ye, Ming Yan, Guohai Xu, Chenliang Li, Junfeng Tian, Qi Qian, Ji Zhang, et al. Ureader: UniversalOCR-free visually-situated language understanding with multimodal large language model. arXiv preprint arXiv:2310.05126, 2023. 3" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 316, + 297, + 553, + 351 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 297, + 553, + 351 + ], + "spans": [ + { + "bbox": [ + 316, + 297, + 553, + 351 + ], + "type": "text", + "content": "[95] Shi Yu, Chaoyue Tang, Bokai Xu, Junbo Cui, Junhao Ran, Yukun Yan, Zhenghao Liu, Shuo Wang, Xu Han, Zhiyuan Liu, et al. Visrag: Vision-based retrieval-augmented generation on multi-modality documents. arXiv preprint arXiv:2410.10594, 2024. 1, 2, 3, 7, 8" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 316, + 353, + 553, + 407 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 353, + 553, + 407 + ], + "spans": [ + { + "bbox": [ + 316, + 353, + 553, + 407 + ], + "type": "text", + "content": "[96] Jiajie Zhang, Yushi Bai, Xin Lv, Wanjun Gu, Danqing Liu, Minhao Zou, Shulin Cao, Lei Hou, Yuxiao Dong, Ling Feng, et al. Longcite: Enabling llms to generate fine-grained citations in long-context qa. arXiv e-prints, pages arXiv-2409, 2024. 2, 6" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 316, + 409, + 553, + 463 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 409, + 553, + 463 + ], + "spans": [ + { + "bbox": [ + 316, + 409, + 553, + 463 + ], + "type": "text", + "content": "[97] Jiajie Zhang, Zhongni Hou, Xin Lv, Shulin Cao, Zhenyu Hou, Yilin Niu, Lei Hou, Yuxiao Dong, Ling Feng, and Juanzi Li. Longreward: Improving long-context large language models with ai feedback. arXiv preprint arXiv:2410.21252, 2024. 2, 6" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 316, + 464, + 553, + 519 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 464, + 553, + 519 + ], + "spans": [ + { + "bbox": [ + 316, + 464, + 553, + 519 + ], + "type": "text", + "content": "[98] Pan Zhang, Xiaoyi Dong, Yuhang Zang, Yuhang Cao, Rui Qian, Lin Chen, Qipeng Guo, Haodong Duan, Bin Wang, Linke Ouyang, et al. Internlm-xcomposer-2.5: A versatile large vision language model supporting long-contextual input and output. arXiv preprint arXiv:2407.03320, 2024. 7" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 316, + 520, + 553, + 574 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 520, + 553, + 574 + ], + "spans": [ + { + "bbox": [ + 316, + 520, + 553, + 574 + ], + "type": "text", + "content": "[99] Wanrong Zhu, Jack Hessel, Anas Awadalla, Samir Yitzhak Gadre, Jesse Dodge, Alex Fang, Youngjae Yu, Ludwig Schmidt, William Yang Wang, and Yejin Choi. Multimodal c4: An open, billion-scale corpus of images interleaved with text. NIPS, 36, 2024. 2" + } + ] + } + ], + "index": 22 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "text", + "content": "4037" + } + ] + } + ], + "index": 24 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 11 + } + ], + "_backend": "vlm", + "_version_name": "2.6.4" +} \ No newline at end of file diff --git a/2025/Document Haystacks_ Vision-Language Reasoning Over Piles of 1000+ Documents/1b0623c9-cc0a-4950-bcf6-716e67ed00b3_content_list.json b/2025/Document Haystacks_ Vision-Language Reasoning Over Piles of 1000+ Documents/1b0623c9-cc0a-4950-bcf6-716e67ed00b3_content_list.json new file mode 100644 index 0000000000000000000000000000000000000000..afa281984db5b531a412f2567a5ff80b9cc6b96a --- /dev/null +++ b/2025/Document Haystacks_ Vision-Language Reasoning Over Piles of 1000+ Documents/1b0623c9-cc0a-4950-bcf6-716e67ed00b3_content_list.json @@ -0,0 +1,1301 @@ +[ + { + "type": "text", + "text": "Document Haystacks: Vision-Language Reasoning Over Piles of 1000+ Documents", + "text_level": 1, + "bbox": [ + 176, + 130, + 823, + 174 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Jun Chen $^{1}$ , Dannong Xu $^{2}$ , Junjie Fei $^{1}$ , Chun-Mei Feng $^{3}$ , Mohamed Elhoseiny $^{1}$ $^{1}$ King Abdullah University of Science and Technology \n $^{2}$ The University of Sydney, $^{3}$ IHPC, A*STAR \n{jun.chen, junjie. fei, mohamed.elhoseiny}@kaust.edu.sa \ndaxu8019@uni.sydney.edu.au, fengcm.ai@gmail.com", + "bbox": [ + 189, + 202, + 805, + 292 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/84598932cd96043268c6467ed2ade6f7cbfc2c116b10251727075dd11b974622.jpg", + "image_caption": [ + "(a) Previous Benchmarks: Each question paired with a limited image set" + ], + "image_footnote": [], + "bbox": [ + 94, + 324, + 480, + 523 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/20a0a7b3bd1b59437ec0c97bc20b4582e5abfc47b8608e808953356a12e6ee2d.jpg", + "image_caption": [ + "(b) Our Benchmarks: All questions mapped to an extensive document collection", + "Figure 1. Comparison between previous and proposed benchmarks. Given a question as input, all benchmarks aim to retrieve relevant images from an image pool to correctly answer the question. Unlike prior benchmarks like RetVQA [32] and WebVQA [7], which structure their datasets by pairing each question with a limited set of images (typically $\\leq 30$ ), our benchmarks, DocHaystack and InfoHaystack, map each question to a substantially larger document collection, scaling up to 1,000 visual documents. This expanded scope more accurately represents large-scale document retrieval scenarios and offers a greater challenge in retrieval accuracy and visual question answering." + ], + "image_footnote": [], + "bbox": [ + 488, + 325, + 897, + 523 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Abstract", + "text_level": 1, + "bbox": [ + 246, + 652, + 326, + 669 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Large multimodal models (LMMs) have achieved impressive progress in vision-language understanding, yet they face limitations in real-world applications requiring complex reasoning over a large number of images. Existing benchmarks for multi-image question-answering are limited in scope, each question is paired with only up to 30 images, which does not fully capture the demands of large-scale retrieval tasks encountered in the real-world usages. To reduce these gaps, we introduce two document haystack benchmarks, dubbed DocHaystack and InfoHaystack, designed to evaluate LMM performance on large-scale visual document retrieval and understanding. Addition", + "bbox": [ + 88, + 686, + 485, + 869 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "ally, we propose V-RAG, a novel, vision-centric retrieval-augmented generation (RAG) framework that leverages a suite of multimodal vision encoders, each optimized for specific strengths, and a dedicated question-document relevance module. V-RAG sets a new standard, with a $9\\%$ and $11\\%$ improvement in Recall@1 on the challenging DocHaystack-1000 and InfoHaystack-1000 benchmarks, respectively, compared to the previous best baseline models. Additionally, integrating V-RAG with LMMs enables them to efficiently operate across thousands of images, yielding significant improvements on our DocHaystack and InfoHaystack benchmarks. Our code and datasets are available at https://github.com/Vision-CAIR/dochaystacks", + "bbox": [ + 511, + 654, + 908, + 867 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "CVF", + "bbox": [ + 106, + 2, + 181, + 42 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore.", + "bbox": [ + 236, + 0, + 810, + 46 + ], + "page_idx": 0 + }, + { + "type": "page_footnote", + "text": "* Equal contribution", + "bbox": [ + 114, + 887, + 227, + 898 + ], + "page_idx": 0 + }, + { + "type": "page_number", + "text": "24817", + "bbox": [ + 478, + 944, + 519, + 955 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "1. Introduction", + "text_level": 1, + "bbox": [ + 91, + 90, + 222, + 104 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Large Multimodal Models (LMMs) [1, 20, 30, 41] have made remarkable progress in the vision-language understanding. However, these models still face challenges when tasked with reasoning over extensive collections of images or documents [43], limiting their effectiveness in real-world applications, such as visual search or querying over large sets of images or documents, like those stored on personal devices or in photo albums. However, there lacks such proper benchmarks to evaluate these capabilities. To address this gap, we introduce the DocHaystack and InfoHaystack benchmarks, designed to evaluate LMMs on large-scale image retrieval and understanding capabilities, pushing the boundaries of LMM performance in complex, real-world scenarios.", + "bbox": [ + 89, + 114, + 480, + 325 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "The existing multi-image retrieving and reasoning benchmarks are primarily constructed on a small scale, as highlighted in works such as [32, 37]. Each question in these benchmarks is paired with only up to 30 images as illustrated in Figure 1 (a). However, this limited scope does not align well with real-world scenarios, which often require retrieval and reasoning across hundreds or thousands of images or documents. In contrast, we introduce the Document Haystacks benchmarks, depicted in Figure 1 (b), which allow for querying questions from a large-scale collection of up to 1,000 documents, necessitating that models retrieve and reason over an extensive set of documents for each question. This scale better simulates practical applications and their demands.", + "bbox": [ + 89, + 325, + 482, + 536 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "The main challenge for constructing our document haystack benchmarks is collecting specific questions while ensuring there are no ambiguous answers across a large set of images. Existing datasets, such as those in DocVQA and InfographicVQA [27, 28], contain numerous \"general\" questions, like \"What is the table number?\", where answers could be derived from multiple images, leading to nonunique responses. To address this, we implemented a rigorous data filtering pipeline. First, we employed both a large language model (LLM) and human annotators to systematically filter out \"general\" questions based on carefully defined criteria. Additionally, we used the LLM to exclude questions relying on generic knowledge, such as \"What is the capital of Missouri?\", which can be answered without image context. This approach ensures that the questions in the benchmark can only be answered through specific visual cues from the provided images, maintaining the benchmark's integrity for evaluating image-based understanding.", + "bbox": [ + 89, + 537, + 482, + 809 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "To enable the current LMMs effectively reason over a large number of images, we propose a vision-centric retrieval-augmented generation (RAG) framework, named V-RAG. V-RAG combines multiple multimodal vision encoders, leveraging each encoder's unique strengths to enhance retrieval accuracy. Additionally, it incorporates an", + "bbox": [ + 89, + 810, + 482, + 901 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "LMM-filter module to assess the relevance of each document to the query, refining the retrieval process by ensuring that only relevant documents are prioritized. This integrated approach allows V-RAG to navigate extensive document collections efficiently. Experimental results demonstrate that V-RAG achieves $9\\%$ and $11\\%$ improvement in Recall@1 on the DocHaystack-1000 and InfoHaystack-1000 compared to previous best text-to-image retrieval methods. Additionally, we found that integrating V-RAG brings GPT4o over a $55\\%$ acc improvement on DocHaystack-200 and a $34\\%$ acc improvement on InfoHaystack-200, indicating the effectiveness of our V-RAG.", + "bbox": [ + 511, + 90, + 903, + 271 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Our contributions are as follows:", + "bbox": [ + 532, + 272, + 751, + 285 + ], + "page_idx": 1 + }, + { + "type": "list", + "sub_type": "text", + "list_items": [ + "- We introduce the Document Haystack benchmarks, including DocHaystack-100/200/1000 and InfoHaystack-100/200/1000, with the most challenging setup consisting of 1,000 documents for each inquiry. These benchmarks advance document retrieval and reasoning tasks by requiring models to navigate and reason across extensive document collections, surpassing prior benchmarks limited to smaller retrieval tasks.", + "- We propose a vision-centric retrieval-augmented generation framework, V-RAG, which enhances the retrieval capabilities of LMMs. V-RAG achieves substantial improvements over previous best text-to-image retrieval methods by $9\\%$ and $11\\%$ on DocHaystack-1000 and InfoHaystack-1000, respectively." + ], + "bbox": [ + 511, + 287, + 903, + 500 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "2. Related Works", + "text_level": 1, + "bbox": [ + 511, + 512, + 663, + 527 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "VQA benchmarks. VQA play a critical role in assessing a model's ability to understand and reason across visual contexts [6, 11, 12]. Traditional VQA datasets typically measure a model's comprehension of object attributes [15, 19], spatial relationships [15], as well as its understanding of documents [27, 28], charts [26], mathematics [23, 40, 46], and open knowledges [25, 35]. Additionally, these benchmarks explore models' knowledge across varied fields, including science and the arts [22, 44]. This broad array of benchmarks has greatly advanced vision-language models by cultivating diverse visual comprehension skills, particularly for modern foundation models in vision-language understanding [1, 3, 8, 20, 22, 30, 31, 39, 47]. Notably, these benchmarks have primarily focused on question answering within single image or document. In contrast, our benchmark shifts the focus towards retrieval and comprehensive understanding across a large collection of visual documents, presenting new challenges and expanding the scope of visual question answering.", + "bbox": [ + 511, + 537, + 903, + 824 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Several previous efforts have tackled the challenge of visual question answering and reasoning across multiple images [5, 7, 32, 37, 38, 42]. For instance, datasets such as MultimodalQA [37] and ISVQA [5] require models to have multi-image reasoning abilities. Meanwhile, WebQA [7]", + "bbox": [ + 511, + 825, + 903, + 900 + ], + "page_idx": 1 + }, + { + "type": "page_number", + "text": "24818", + "bbox": [ + 478, + 944, + 519, + 955 + ], + "page_idx": 1 + }, + { + "type": "image", + "img_path": "images/44a1374ccf0762cb7e5d4774343878b7d47e5654ddb937ab9ce7944c30bed756.jpg", + "image_caption": [ + "Figure 2. Data Curation Pipeline. Our benchmarks are curated based on the DocVQA and InfographicVQA datasets, following a three-step filtering process to obtain document-specific question-answer pairs. In Step 1, we filter out general questions (e.g., \"What is the table number?\"), as these could be answered by multiple documents and lack specificity. Step 2 involves a manual review by human annotators to further remove general questions. In Step 3, we eliminate generic-knowledge questions (e.g., \"How many sports were in the 2008 Beijing Paralympic Games?\") that can be answered directly by large language models without requiring image input." + ], + "image_footnote": [], + "bbox": [ + 107, + 89, + 890, + 209 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "and RetVQA [32] involve an additional step where models must first retrieve relevant images from a limited image pool before answering visual questions based on these results. However, these benchmarks are generally constrained to relatively small image pools, where each question is paired with an image set containing up to 30 images. In contrast, our proposed benchmarks, DocHaystack and InfoHaystack, significantly expand this scope by requiring models to retrieve and reason from a much larger set of up to 1,000 documents, presenting a notably greater challenge in retrieval and multi-image reasoning.", + "bbox": [ + 88, + 316, + 485, + 482 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Large multimodal models (LMMs). LMMs have achieved substantial advancements in understanding and reasoning across single or multiple images [1, 8, 20, 30, 41, 47]. These models have significantly enhanced vision-language understanding across numerous dimensions and applications [12, 23, 44, 46]. LMMs benefit primarily from large-scale image-text alignment and extensive language modeling, which emerge them with advanced understanding and reasoning abilities. However, despite these breakthroughs, LMMs still encounter challenges when handling large-scale image or document sets [43]. This difficulty is due to the inherent complexity of processing such complex data. To address this, retrieval-based approaches have been developed to extend the capacity of vision-language models, augmenting their ability to process and reason over a larger number of images.", + "bbox": [ + 89, + 487, + 483, + 731 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Retrieval-augmented generation (RAG). RAG integrates retrieval systems [4, 11, 16, 33, 45], with generative models, enhancing them with additional knowledge. While RAG has been extensively explored in language domains [2, 13, 17, 24], its application in vision-language contexts is also advancing. In vision-language RAG, models like MuRAG [9] leverage image-text memory to retrieve top-k neighbors by comparing inner-product similarities. RetVQA [32] uses an image-question relevance encoder, combining BERT [10] and Faster R-CNN [34] to filter relevant images, while MIRAGE [43] employs a CLIP-based", + "bbox": [ + 89, + 734, + 485, + 902 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "encoder [33] to train a retriever. These frameworks extend model capabilities, enabling retrieval and reasoning across hundreds or thousands of images. In contrast, we propose V-RAG, a vision-centric RAG framework that integrates multiple vision encoders to more effectively capture image features, and introduces a LMM-based question-document relevance comparison module. Our results demonstrate that V-RAG surpasses existing methods on our DocHaystack and InfoHaystack benchmarks, setting a new standard for large-scale visual retrieval and reasoning.", + "bbox": [ + 511, + 316, + 908, + 469 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3. DocHaystack and InfoHaystack Benchmarks", + "text_level": 1, + "bbox": [ + 511, + 479, + 906, + 513 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "To support effective retrieval and reasoning across extensive document collections, we present two new benchmarks—DocHaystack and InfoHaystack—designed to ensure each question yields a unique, document-specific answer. Derived from DocVQA [28] and InfographicVQA [27], these benchmarks address the challenge of answer ambiguity by selectively curating questions that can only be answered by a single document within a large dataset.", + "bbox": [ + 511, + 523, + 908, + 657 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Benchmark construction pipeline. There exists many general questions in the existing benchmarks and lead to multiple answers for different document context. For example, general questions like \"What is the table number?\" may apply to various documents and yield multiple valid answers, while a targeted question like \"Who is the reviewer for the article titled 'An antithyroid factor in milk'?\" is likely to produce a unique answer, as only a single document or a limited set of documents would contain that information. Therefore, our benchmark construction follows a structured three-step filtering pipeline, illustrated in Figure 2, to ensure high-quality, unique-answer questions. First, we employ a large language model (LLM) to filter out general questions that could generate multiple answers across documents. Next, a manual review step further", + "bbox": [ + 511, + 659, + 910, + 901 + ], + "page_idx": 2 + }, + { + "type": "page_number", + "text": "24819", + "bbox": [ + 478, + 944, + 519, + 957 + ], + "page_idx": 2 + }, + { + "type": "image", + "img_path": "images/f521efde925d2b87455540c536284a49453fb485b9b8a3388f1312a8a9bf9242.jpg", + "image_caption": [ + "Figure 3. The V-RAG pipeline workflow. In the top section, a vision encoder ensemble is used, combining multiple vision models—CLIP, SigLIP, and OpenCLIP—to process a large document haystack. Each encoder computes similarity scores, which are averaged into $Sim_{avg}$ . The top m documents, based on these scores, are selected for further analysis. In the bottom right, the LMM-Filter Module utilizes a pretrained LMM to assess whether each selected document can potentially answer the posed question. This filtering step removes documents that do not match, retaining only relevant ones. Finally, the top k most relevant images are input into the LMM along with the original question $q$ to generate a specific answer." + ], + "image_footnote": [], + "bbox": [ + 102, + 88, + 897, + 364 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "checks the questions to ensure the data quality. Finally, a generic-knowledge filtering stage refines the dataset further, retaining only questions closely tied to specific document content.", + "bbox": [ + 88, + 491, + 480, + 549 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "This carefully designed pipeline, combining LLM-based filtering and human review, effectively curates questions that drive accurate, document-specific retrieval. By focusing on reducing answer ambiguity, DocHaystack and InfoHaystack enhance the precision of retrieval and reasoning in large-scale document processing tasks, providing a valuable tool for the evaluation of retrieval systems. We discuss this data curation pipeline in details as follows:", + "bbox": [ + 88, + 551, + 482, + 672 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "General-question LLM filtering. We begin by using the LLM, GPT-4o [30], to filter out general questions through a set of well-crafted instructions. Leveraging the LLM's strong contextual understanding, this initial filtering step allows us to efficiently process large volumes of data, identifying broad or ambiguous questions that may yield multiple answers across documents. This automated approach significantly enhances the benchmark construction's efficiency and quality.", + "bbox": [ + 88, + 672, + 482, + 808 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "To guide the LLM, we first define the task, providing clear distinctions between general and specific questions along with illustrative examples. With this framework, the LLM can then assess each question and determine if it is general or specific. The instructional format is as follows:", + "bbox": [ + 88, + 809, + 482, + 883 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "You are an evaluator tasked with identifying if a ques", + "bbox": [ + 109, + 885, + 482, + 901 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "tion is specific or general. A general question seeks commonly known or widely applicable information without unique identifiers, e.g., \"Who is the person standing in the ground?\" A specific question, however, requests unique information about a particular individual, event, or object, e.g., \"What is the Social Security Number of Charles Yarbrough?\" Based on these definitions, determine if the following question is general or specific: {question}.", + "bbox": [ + 511, + 491, + 903, + 611 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "General-question manual review. After the initial LLM filtering, we conduct a manual review of the questions that were classified as specific. This manual process involves two key steps to ensure answer uniqueness and benchmark quality.", + "bbox": [ + 511, + 612, + 903, + 686 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "In the first step, we examine each question to confirm it contains unique identifiers—such as names, dates, titles, or other specific attributes—suggesting a document-specific answer. This careful check helps identify questions with clear, unique markers that direct the retrieval process to a single document.", + "bbox": [ + 511, + 688, + 903, + 777 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "In the second step, we verify the uniqueness of each answer to eliminate any remaining ambiguity. Although specific identifiers are present, questions may still be prone to ambiguity, such as with common names or recurring book titles. To address this, we employ a refined verification process. First, we use Optical Character Recognition (OCR) [36] to extract all text from images in the dataset. We then search for occurrences of the unique identifiers re", + "bbox": [ + 511, + 779, + 906, + 900 + ], + "page_idx": 3 + }, + { + "type": "page_number", + "text": "24820", + "bbox": [ + 478, + 944, + 519, + 957 + ], + "page_idx": 3 + }, + { + "type": "table", + "img_path": "images/d437293a87d0d4a0b1dccef251980a432c874ef29e4a86dbe99487f17300256a.jpg", + "table_caption": [], + "table_footnote": [], + "table_body": "
GPT-4oLLaVA-OneVisionQwen2-VL
DocVQA27.0%4.7%3.4%
InfographicVQA56.4%13.4%11.3%
", + "bbox": [ + 94, + 88, + 480, + 143 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Table 1. Percentage of questions answerable by LMMs without vision input. We evaluate GPT-4o, LLaVA-Onevision, and Qwen2-VL on their ability to answer questions directly from our dataset without requiring vision input. The reported percentage reflects the proportion of examples that can be answered solely through language understanding.", + "bbox": [ + 89, + 152, + 482, + 237 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "tained from the first step across other documents. If matches are found, a manual review is conducted to ensure no alternative valid answers exist. This comprehensive approach minimizes the possibility of a single question mapping to multiple answers, enhancing the precision and reliability of our benchmarks.", + "bbox": [ + 88, + 263, + 482, + 354 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Generic-knowledge filtering. In DocVQA and InfographicVQA tasks, certain questions—such as “How many sports were in the 2008 Beijing", + "bbox": [ + 89, + 354, + 482, + 400 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Paralympic Games?\"—can be answered based on general knowledge accessible to a large language model, without relying on the image content. This introduces a language bias when using LMMs for visual question answering, as it shifts the focus away from image-based reasoning. To address this, we filter out these general-knowledge questions, ensuring that evaluation emphasizes vision-based understanding and that models rely primarily on visual content to generate accurate answers.", + "bbox": [ + 89, + 401, + 482, + 536 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "To implement this, we developed an LLM-based evaluation pipeline that detects and excludes such questions. For each question, we prompt an LLM with “{question}, answer briefly.” After receiving a response, we compare it to the ground-truth answer using another LLM. If the response matches the ground truth, we classify the question as general knowledge-related and remove it, thereby isolating questions that truly require visual document understanding. As shown in Table 1, GPT-4o accurately answers $27.0\\%$ of DocVQA questions and $56.4\\%$ of InfographicVQA questions directly, a rate significantly higher than that of open-source LLMs. Therefore, we select GPT-4o to filter out the questions that can be directly answered by the GPT-4o model. Overall, this process is to ensure that the evaluation reflects the necessity of vision-based comprehension.", + "bbox": [ + 89, + 537, + 482, + 779 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Final dataset profile. After a rigorous three-stage data filtering process, we retained 109 questions from DocVQA and 155 questions from InfographicVQA, associated with 59 and 66 documents that provide the evidence, respectively. To assess retrieval performance at scale, we introduce two benchmarks: DocHaystack-1000 and InfoHaystack-1000, where each question requires retrieving relevant content from a set of 1,000 documents. Given", + "bbox": [ + 89, + 780, + 482, + 901 + ], + "page_idx": 4 + }, + { + "type": "image", + "img_path": "images/f8264d51318bf845052e4510597aaf53449d962463b65f1bcef0c4390c546d80.jpg", + "image_caption": [ + "Figure 4. Question type analysis. We analyze the distribution of question types of DocHaystack and InfoHaystack. Each benchmark categorizes the data into 5 different types." + ], + "image_footnote": [], + "bbox": [ + 524, + 94, + 687, + 229 + ], + "page_idx": 4 + }, + { + "type": "image", + "img_path": "images/bcf44f1ff8cc585290afca51d4c3ff1172c7e6440de95546fb4b832627797fe4.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 717, + 95, + 888, + 234 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "the challenge this scale presents to current LMMs, particularly in terms of context length limitations, we also construct two smaller benchmarks: DocHaystack-100/200 and InfoHaystack-100/200. These benchmarks allow direct input of all associated images into the context, enabling evaluation of models' long-context comprehension ability. For training set, we also construct a dataset comprising 2,835 questions similarly, with 899 from DocVQA and 1,936 from InfographicVQA, to support robust learning and generalization for the multi-image reasoning tasks.", + "bbox": [ + 511, + 300, + 906, + 450 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Question type analysis. The types of questions represent the types of the evidence required for accurate answers. In Figure 4, we illustrate the distribution of question types across our dataset to provide insights into its structure. Following the classification system used in DocVQA and InfographicVQA, we categorize questions accordingly (note that a single question may fall into multiple categories). As shown in the figure, the DocHaystack benchmark places a greater emphasis on Table/List and Layout understanding, whereas InfoHaystack primarily targets Figure, Text, and Table comprehension.", + "bbox": [ + 511, + 452, + 908, + 618 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "4. Method", + "text_level": 1, + "bbox": [ + 511, + 632, + 604, + 648 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Current large multimodal models (LMMs) face substantial challenges when reasoning across hundreds or thousands of images, due not only to context length limitations but also to the inherent complexity of the task. This issue is particularly pronounced in our benchmarks, which contain 1k document files requiring high-resolution images to capture and interpret small-font text effectively. To enable LMMs to perform reasoning over a substantial number of documents, we introduce a vision-centric retrieval-augmented generation (V-RAG) framework. V-RAG efficiently retrieves a reduced set of relevant documents, allowing the LMM to focus on a manageable subset for deeper understanding, as illustrated in Figure 3. In the following section, we provide a detailed description of the V-RAG pipeline.", + "bbox": [ + 511, + 657, + 906, + 869 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Task definition. Given a question $q$ and a collection of $N$ documents $\\mathcal{D} = \\{D_1, \\ldots, D_N\\}$ , the V-RAG framework", + "bbox": [ + 511, + 869, + 908, + 902 + ], + "page_idx": 4 + }, + { + "type": "page_number", + "text": "24821", + "bbox": [ + 478, + 944, + 517, + 955 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "aims to retrieve the top- $k$ most relevant documents to support LMMs understanding and answering the question $q$ . V-RAG accomplishes this through a two-step retrieval process designed to effectively identify and rank relevant documents for each question.", + "bbox": [ + 89, + 90, + 480, + 167 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Vision encoder ensemble. Document files often contain a mix of text, symbols, and visual elements across various scales, requiring vision encoders to capture a comprehensive understanding of these complex structures. To efficiently handle this diversity, we represent each document as an image and utilize an ensemble of vision encoders, including CLIP [33], SigLIP [45], and OpenCLIP [16], each bringing distinct strengths to the image understanding, as depicted in Figure 3. For example, the ConvNext encoder [21] from OpenCLIP [16] is particularly effective for high-resolution image encoding. We compute the similarity score between each question $q$ and all documents in the document set $\\mathcal{D}$ according to Equation 1, with similarity scores from each encoder represented as $Sim_{c}$ , $Sim_{o}$ , and $Sim_{s}$ respectively.", + "bbox": [ + 89, + 169, + 482, + 396 + ], + "page_idx": 5 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {S} (q, \\mathcal {D}) = \\cos \\left(\\phi_ {\\mathrm {t}} (q), \\phi_ {\\mathrm {v}} (D _ {j})\\right) \\mid D _ {j} \\in \\mathcal {D}, \\tag {1}\n$$\n", + "text_format": "latex", + "bbox": [ + 148, + 422, + 482, + 439 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "where $S$ denotes the computing the similarity between the query $q$ and a collection of documents $\\mathcal{D}$ . $\\cos$ denotes the cosine similarity. $\\phi_{\\mathrm{t}}$ denotes the text encoder, and $\\phi_{\\mathrm{v}}$ denotes the vision encoder.", + "bbox": [ + 89, + 453, + 482, + 513 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "To derive a final relevance score, we calculate the average similarity $Sim_{avg}$ for each question-image pair by combining $Sim_c$ , $Sim_o$ , and $Sim_s$ . We then rank the images based on $Sim_{avg}$ in descending order, selecting the top- $m$ most relevant images according to their similarity scores.", + "bbox": [ + 89, + 517, + 482, + 593 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "LMM-filter module. To refine the selection of top- $m$ relevant images further, we introduce a LMM-based question-image relevance assessment module. This module evaluates the relevance between each question and the top- $m$ images identified in the first filtering step. Specifically, we pair each image with the question text and input them into an open-source vision-language model, prompting, \"Can this image provide answers to this question? Respond only with yes or no\". We only retain the question-image pairs that are identified as \"yes\" from LMM, and remove other irrelevant images.", + "bbox": [ + 89, + 595, + 482, + 762 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "LMM-VQA module. Achieving high top-1 ranking accuracy in image retrieval is challenging, so we retain the top- $k$ images from the LMM-filtered ranking list and present them to the LMM-VQA to improve the likelihood of including relevant images. We input these top- $k$ images alongside the question into the LMM-VQA (see Figure 3), which then generates the answer directly. To enhance robustness against visual distractors, the LMM-VQA can be further optimized, as analyzed in the experiment section.", + "bbox": [ + 89, + 763, + 482, + 901 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "5. Experiments", + "text_level": 1, + "bbox": [ + 513, + 89, + 645, + 107 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "In the experiments section, we will primarily describe our training setup, covering evaluation metrics, baseline models, and the fine-tuning procedure for the LMM-VQA model. We also present the main experimental results along with an ablation study to provide further insights.", + "bbox": [ + 511, + 117, + 903, + 194 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "5.1. Training setup", + "text_level": 1, + "bbox": [ + 513, + 210, + 661, + 228 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Metric. In our evaluation of the DocHaystack and Info-Haystack benchmarks, we employ a model-based assessment by leveraging GPT-4o-mini [30] to accurately determine whether the model predictions match target answers. This method uses a carefully structured prompt to facilitate GPT-4o-mini's evaluation of answer correctness. We empirically found that the model-based evaluation achieves higher consistency and alignment with human judgment. Additional details on the prompt design are provided in the Appendix.", + "bbox": [ + 511, + 236, + 903, + 387 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "For the document retrieval evaluation, we report the baseline results using Recall@1, Recall@3, and Recall@5 metrics. These metrics enable a thorough assessment of retrieval accuracy across varying levels of precision.", + "bbox": [ + 511, + 390, + 903, + 450 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Baselines. In our experiment, we have evaluated several open and closed-sourced vision-language models on the retrieval and VQA performance. For the large multimodal model, we used the gpt-4o-2024-08-06 version of GPT-4o [30], the LLaVA-OneVision-Qwen2-7b-OV-HF version of LLaVA-OneVision [20], and the Qwen2-VL-7B-Instruct version of Qwen2-VL [3]. For computing the text-to-image similarities, we employed the Jina-CLIP-v1 [18] variant, Nomic-Embed-Vision-v1.5 [29] variant, CLIP [33] ViT-L/14@336 variant, for SigLIP [45], the ViT-SO400M/14@384 variant, and for OpenCLIP [16], the ConvNeXt-XXL@1024 variant as well as text-based method, BM25.", + "bbox": [ + 511, + 454, + 905, + 648 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "In our V-RAG setting, we apply LLaVA-OneVision-Qwen2-7b-OV-HF for the LMM-filter module and Qwen2-VL-7B-Instruct for the LMM-VQA module. We select $m$ as 60 and $k$ as 5 in our experiment.", + "bbox": [ + 511, + 652, + 903, + 714 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Optimizing the LMM-VQA module. To improve the robustness of the LMM-VQA model in handling visual question answering with multiple distractor images, we further fine-tune the model using our curated training data.", + "bbox": [ + 511, + 715, + 903, + 777 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "During this fine-tuning process, we introduce 1-10 randomly sampled distractor images for each question, creating a challenging setting that encourages the model to focus on relevant content amid a mix of positive and negative images. The fine-tuning is conducted with a batch size of 32 and a peak learning rate of 1e-4 over a single epoch. Additionally, we leverage LoRA [14] with a rank of 8 to efficiently adapt the model's parameters during training.", + "bbox": [ + 511, + 779, + 906, + 901 + ], + "page_idx": 5 + }, + { + "type": "page_number", + "text": "24822", + "bbox": [ + 478, + 944, + 517, + 955 + ], + "page_idx": 5 + }, + { + "type": "table", + "img_path": "images/9090873ad8c2d36b7f0e13db88bc070120c9449c5abb1c838cb7520be7e96248.jpg", + "table_caption": [], + "table_footnote": [], + "table_body": "
DocHaystack-100DocHaystack-200DocHaystack-1000
R@1R@3R@5R@1R@3R@5R@1R@3R@5
BM25 (OCR)63.375.279.865.171.675.256.966.169.7
Jina-CLIP [18]16.531.241.39.224.830.33.77.312.8
Nomic-Embed-Vision [29]16.524.828.413.821.125.71.82.86.4
CLIP [33]46.865.169.744.058.765.123.941.345.9
SigLIP [45]51.467.976.247.763.370.633.049.557.8
OpenCLIP [16]58.775.279.856.970.675.234.949.557.8
V-RAG (ours)81.789.089.078.084.484.466.178.078.9
InfoHaystack-100InfoHaystack-200InfoHaystack-1000
R@1R@3R@5R@1R@3R@5R@1R@3R@5
BM25 (OCR)56.865.871.051.665.269.038.751.658.1
Jina-CLIP43.251.658.136.846.551.623.933.637.4
Nomic-Embed-Vision34.850.356.831.043.248.420.731.035.5
CLIP69.778.785.865.277.481.945.864.570.3
SigLIP58.171.680.055.567.776.839.455.561.9
OpenCLIP72.385.292.966.581.989.053.665.872.9
V-RAG (ours)79.491.092.974.888.488.464.574.278.1
", + "bbox": [ + 168, + 87, + 826, + 311 + ], + "page_idx": 6 + }, + { + "type": "table", + "img_path": "images/04643601e948bb0cd287df502a30280430ef1da242076566b2e46b5b73abc7a7.jpg", + "table_caption": [ + "Table 2. Retrieval Results. We compare our V-RAG model with other text-to-image and text-to-text (using OCR) retrieval methods across both benchmarks. V-RAG consistently outperforms baseline models on Recall@1, Recall@3, and Recall@5 metrics. Notably, V-RAG leverages an ensemble of text-to-image models along with a large multimodal model in a two-stage filtering approach. Top-performing values in each column are highlighted in bold." + ], + "table_footnote": [], + "table_body": "
ModelDocHaystackInfoHaystack
10020010001002001000
LLaVA-OV [20]------
GPT-4o [30]27.523.9-23.920.0-
Gemini [1]50.548.6-29.021.9-
Qwen2-VL [41]41.312.8-20.014.2-
MIRAGE [43]3.73.72.87.77.16.5
LLaVA-OV+V-RAG69.765.155.143.241.936.8
GPT-4o+V-RAG81.772.567.065.263.256.8
Gemini+V-RAG73.465.158.757.457.447.1
Qwen2-VL+V-RAG82.674.366.165.865.860.0
Qwen2-VL-f.t.+V-RAG86.279.873.467.167.760.0
", + "bbox": [ + 114, + 393, + 460, + 547 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Table 3. The VQA results for the DocHaystack and Info-Haystack. We evaluate with many closed-source and open-source multimodal model, and also integrating them with our V-RAG retrieval framework. - denotes that those models can not be inferred due to their token context constraints. To enable GPT-4o and Qwen2-VL to process hundreds of images, we employ low-resolution mode and adjust image size for compatibility.", + "bbox": [ + 89, + 556, + 482, + 655 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "5.2. Main Experimental Results", + "text_level": 1, + "bbox": [ + 89, + 662, + 338, + 679 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "We evaluated a range of open-source and closed-source vision-language models for VQA tasks. We also evaluate several text-to-image and text-to-text (with OCR) retrieval models to evaluate their retrieval capabilities on our benchmarks. More detailed performance analysis are described in the following sections.", + "bbox": [ + 89, + 686, + 482, + 777 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Retrieving results. The retrieval results in Table 2 demonstrate the superiority of our proposed V-RAG framework over several baseline methods across both DocHaystack and InfoHaystack benchmarks. V-RAG consistently achieves the highest Recall@1, Recall@3, and Recall@5 scores on most categories, indicating its robust retrieval capabilities. Notably, V-RAG outperforms text-based retrieving models such as BM25 and also the text-to-image retrieval models", + "bbox": [ + 89, + 779, + 482, + 900 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "like jina-clip, CLIP, SigLIP, and OpenCLIP by substantial margins, especially on the DocHaystack-100 subset, where it reaches Recall@1 of $81.7\\%$ and Recall@5 of $89.0\\%$ . This pattern continues for larger datasets (DocHaystack-1000), where V-RAG remains competitive, achieving Recall@1 of $66.1\\%$ . It achieves the top performance across all recall metrics on DocHaystack. For InfoHaystack benchmarks, V-RAG also outperforms other models, particularly on InfoHaystack-100 and InfoHaystack-200, where it receives Recall@1 of $79.4\\%$ and $64.5\\%$ , higher than previous best by around $7\\%$ and $11\\%$ , respectively. This consistent performance advantage highlights the effectiveness of V-RAG's ensemble of multiple vision encoders, allowing it to capture more granular details and improve retrieval accuracy over large multimodal models.", + "bbox": [ + 511, + 396, + 906, + 623 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Visual question answering (VQA) results. The table presents VQA results for the DocHaystack and Info-Haystack benchmarks across varying dataset sizes (100, 200, 1000) using different multimodal models, both independently and in combination with the V-RAG framework. The results show that Qwen2-VL fine-tuned with V-RAG (Qwen2-VL-f.t.+V-RAG) achieves the highest scores across most benchmarks, with particularly notable performance on DocHaystack-100 (86.2) and InfoHaystack-100 (67.1), indicating superior retrieval and VQA capabilities in these scenarios. When V-RAG is added to other models, substantial improvements are observed, demonstrating the framework's efficacy in enhancing retrieval accuracy. For instance, GPT-4o's performance increases significantly with V-RAG, particularly for DocHaystack-100 and -200. The analysis highlights that V-RAG integration generally boosts performance across models, with Qwen2-VL-f.t.+V-RAG standing out as the top performer on both", + "bbox": [ + 511, + 628, + 908, + 901 + ], + "page_idx": 6 + }, + { + "type": "page_number", + "text": "24823", + "bbox": [ + 478, + 945, + 517, + 955 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/1b5777be2eeaa49dfad5bd1c1f8a694dc890c097efba4cb37865a0cb29e32481.jpg", + "image_caption": [ + "Figure 5. Top-k selection ablation analysis for LMM-VQA. We demonstrate the results for LLaVA, Qwen2-VL, GPT-4o and also the finetuned Qwen2-VL model on the DocHaystack-100/1000 and InfoHaystack-100/1000 benchmarks. All the models are integrated with our V-RAG framework. We show the VQA accuracy performance for each ablation." + ], + "image_footnote": [], + "bbox": [ + 99, + 93, + 279, + 220 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/e78914bb84f03c5adaa7cdab5f6063249e26ced941af318bf49a1fd6a8b8ec39.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 300, + 93, + 480, + 219 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/86d3e9efa600a0d7ed27dbf45fec14a74066c047072ff13e29f06128ffd0f57b.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 504, + 93, + 681, + 219 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/ccb77c5c7273b24a04c7501c4b31159868cfa7fb1c21e67e4e3b06d7f6ff10f6.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 705, + 95, + 883, + 219 + ], + "page_idx": 7 + }, + { + "type": "table", + "img_path": "images/f306ec03032315a38dd381fcfd77e8c626ffb9a62d45ca8984edf0912b087ab1.jpg", + "table_caption": [], + "table_footnote": [], + "table_body": "
CLIPSigLIPOpenCLIPLMM-filterDocHaystack-1000InfoHaystack-1000
R@1R@3R@5R@1R@3R@5
✓XXX23.941.345.945.864.570.3
X✓XX33.049.557.839.455.561.9
XX✓X34.949.557.853.665.872.9
✓✓XX40.459.662.459.467.774.2
✓✓✓X42.266.177.556.171.078.1
✓✓✓✓66.178.078.964.574.278.1
", + "bbox": [ + 261, + 279, + 736, + 364 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Table 4. Ablation study on the V-RAG framework components. We quantify the impact of each module for the Recall@1, Recall@3 and Recall@5 retrieval performance on the DocHaystack-1000 and InfoHaystack-1000 for our V-RAG framework.", + "bbox": [ + 89, + 375, + 906, + 402 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "benchmarks, especially for the larger 1000-document tasks where retrieval accuracy is more challenging. This suggests that V-RAG's vision-centric, retrieval-augmented approach is highly effective for large-scale multimodal document understanding.", + "bbox": [ + 89, + 419, + 482, + 494 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "The table also shows that the DocHaystack-1000 and InfoHaystack-1000 present significant challenges for current LMMs. The drop in performance for larger document sets, with top accuracy only reaching $73.4\\%$ for DocHaystack-1000 and $60.0\\%$ for InfoHaystack-1000, underscores the difficulty our benchmarks.", + "bbox": [ + 89, + 496, + 483, + 585 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "5.3. Ablation Studies", + "text_level": 1, + "bbox": [ + 89, + 602, + 254, + 617 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Ablation study on Top-k Selection. This figure presents the top-k selection ablation analysis for LMM-VQA across four models: LLaVA-OV, Qwen2-VL, GPT-4o, and the fine-tuned Qwen2-VL (Qwen2-VL-f.t.), evaluated on the DocHaystack-100/1000 and InfoHaystack-100/1000 benchmarks. The analysis reports VQA accuracy as a function of top-k selection (Top 1, Top 3, and Top 5). Overall, accuracy tends to improve with larger k-values, suggesting that offering more retrieval options positively impacts model performance. However, for LLaVA-OV, there is a marked decrease in performance at top-5, indicating that this model struggles to process multiple images at this scale.", + "bbox": [ + 89, + 626, + 482, + 808 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Ablation study on the V-RAG framework components. The ablation study in Table 4 highlights the contributions of each component in the V-RAG framework on the DocHaystack-1000 and InfoHaystack-1000 benchmarks. Using CLIP alone yields low performance (e.g., Recall@1 of $23.9\\%$ on DocHaystack-1000 and $45.8\\%$ on", + "bbox": [ + 89, + 810, + 482, + 900 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "InfoHaystack-1000), indicating its limited retrieval capability on its own. Adding SigLIP and OpenCLIP incrementally improves results.", + "bbox": [ + 511, + 419, + 906, + 463 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "The highest performance is achieved when all three encoders are combined with the VLM-filter module, leading to Recall@1 scores of $66.1\\%$ on DocHaystack-1000 and $64.5\\%$ on InfoHaystack-1000. This setup also achieves the top Recall@1, Recall@3 and Recall@5 values, demonstrating that the VLM-filter is essential for refining the ensemble outputs and significantly improving retrieval accuracy. These results confirm that each module contributes to VRAG's overall effectiveness.", + "bbox": [ + 511, + 470, + 906, + 606 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "6. Conclusion", + "text_level": 1, + "bbox": [ + 513, + 641, + 633, + 657 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "In this work, we introduced the DocHaystack and Info-Haystack benchmarks to evaluate LMMs for retrieving and reasoning across large-scale documents. Our benchmarks providing a more rigorous and realistic assessment of large multimodal models in real-world, large-scale retrieval scenarios. To tackle these challenges, we proposed V-RAG, a vision-centric retrieval-augmented generation framework that significantly enhances retrieval precision and overall VQA performance. V-RAG achieves this through an ensemble of vision encoders and a specialized relevance filtering module, enabling improved accuracy across diverse visual inputs. Experimental results indicate that integrating V-RAG enables both open-source and closed-source LMMs to achieve superior performance in large-scale image retrieval and complex reasoning tasks.", + "bbox": [ + 511, + 674, + 908, + 900 + ], + "page_idx": 7 + }, + { + "type": "page_number", + "text": "24824", + "bbox": [ + 478, + 944, + 517, + 955 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "References", + "text_level": 1, + "bbox": [ + 91, + 90, + 187, + 104 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[1] Google AI. Gemini: Google's multimodal ai model. Google AI Research, 2024. https://fireflies.ai/blog/gemini-vs-gpt-4.2,3,7", + "[2] Akari Asai, Zeqiu Wu, Yizhong Wang, Avirup Sil, and Hannaneh Hajishirzi. Self-rag: Learning to retrieve, generate, and critique through self-reflection. arXiv preprint arXiv:2310.11511, 2023. 3", + "[3] Yifan Bai, Zhen Zhang, Yifan Zhang, Yuxuan Li, Yifan Zhang, Yifan Zhang, Yifan Zhang, Yifan Zhang, Yifan Zhang, Yifan Zhang, et al. Qwen-vl: A frontier vision-language model with larger-scale vision pre-training and aligned cross-modal instruction tuning. arXiv preprint arXiv:2310.06726, 2023. 2, 6", + "[4] Yang Bai, Xinxing Xu, Yong Liu, Salman Khan, Fahad Khan, Wangmeng Zuo, Rick Siow Mong Goh, and ChunMei Feng. Sentence-level prompts benefit composed image retrieval. In Proceedings of the International Conference on Learning Representations (ICLR), 2024. Spotlight Presentation. 3", + "[5] Aayush Bansal, Karan Sikka, Gaurav Sharma, and Rama Chellappa. Visual question answering on image sets. In Proceedings of the European Conference on Computer Vision (ECCV), pages 35-51, 2020. 2", + "[6] Florian Bordes, Richard Yuanzhe Pang, Anurag Ajay, Alexander C Li, Adrien Bardes, Suzanne Petryk, Oscar Manas, Zhiqiu Lin, Anas Mahmoud, Bargav Jayaraman, et al. An introduction to vision-language modeling. arXiv preprint arXiv:2405.17247, 2024. 2", + "[7] Yingshan Chang, Mridu Narang, Hisami Suzuki, Guihong Cao, Jianfeng Gao, and Yonatan Bisk. Webqa: Multihop and multimodal qa. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 14178-14188, 2022. 1, 2", + "[8] Jun Chen, Deyao Zhu, Xiaogian Shen, Xiang Li, Zechun Liu, Pengchuan Zhang, Raghuraman Krishnamoorthi, Vikas Chandra, Yunyang Xiong, and Mohamed Elhoseiny. Minigpt-v2: Large language model as a unified interface for vision-language multi-task learning. arXiv preprint arXiv:2310.15339, 2023. 2, 3", + "[9] Wenhu Chen, Hexiang Hu, Xi Chen, Pat Verga, and William W. Cohen. MuRAG: Multimodal retrieval-augmented generator for open question answering over images and text. In Proceedings of the 2022 Conference on Empirical Methods in Natural Language Processing (EMNLP), pages 5558-5570. Association for Computational Linguistics, 2022. 3", + "[10] Jacob Devlin, Ming-Wei Chang, Kenton Lee, and Kristina Toutanova. Bert: Pre-training of deep bidirectional transformers for language understanding. In Proceedings of the 2019 Conference of the North American Chapter of the Association for Computational Linguistics: Human Language Technologies, pages 4171–4186. Association for Computational Linguistics, 2019. 3", + "[11] Chun-Mei Feng, Yang Bai, Tao Luo, Zhen Li, Salman Khan, Wangmeng Zuo, Xinxing Xu, Rick Siow Mong Goh," + ], + "bbox": [ + 93, + 114, + 483, + 900 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "and Yong Liu. Vq4cir: Boosting composed image retrieval with visual question answering. arXiv preprint arXiv:2312.12273, 2023. 2, 3", + "[12] Yash Goyal, Tejas Khot, Douglas Summers-Stay, Dhruv Batra, and Devi Parikh. Making the V in VQA matter: Elevating the role of image understanding in Visual Question Answering. In Conference on Computer Vision and Pattern Recognition (CVPR), 2017. 2, 3", + "[13] Kelvin Guu, Kenton Lee, Zora Tung, Panupong Pasupat, and Ming-Wei Chang. Realm: Retrieval-augmented language model pre-training. In Proceedings of the 37th International Conference on Machine Learning (ICML), pages 3929-3938. PMLR, 2020. 3", + "[14] Edward J. Hu, Yelong Shen, Phillip Wallis, Zeyuan Allen-Zhu, Yuanzhi Li, Shean Wang, Lu Wang, and Weizhu Chen. Lora: Low-rank adaptation of large language models, 2021. 6", + "[15] Drew A Hudson and Christopher D Manning. Gqa: A new dataset for real-world visual reasoning and compositional question answering. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition, pages 6700-6709, 2019. 2", + "[16] Gabriel Ilharco, Mitchell Wortsman, Ross Wightman, Cade Gordon, Nicholas Carlini, Rohan Taori, Achal Dave, Vaishaal Shankar, Hongseok Namkoong, John Miller, Hannaneh Hajishirzi, Ali Farhadi, and Ludwig Schmidt. Openclip. 2021. If you use this software, please cite it as below. 3, 6, 7", + "[17] Vladimir Karpukhin, Barlas Oğuz, Sewon Min, Patrick Lewis, Ledell Wu, Sergey Edunov, Danqi Chen, and Wentau Yih. Dense passage retrieval for open-domain question answering. In Proceedings of the 2020 Conference on Empirical Methods in Natural Language Processing (EMNLP), pages 6769-6781. Association for Computational Linguistics, 2020. 3", + "[18] Andreas Koukounas, Georgios Mastrapas, Michael Günther, Bo Wang, Scott Martens, Isabelle Mohr, Saba Sturua, Mohammad Kalim Akram, Joan Fontanals Martínez, Saahil Ognawala, Susana Guzman, Maximilian Werk, Nan Wang, and Han Xiao. Jina clip: Your clip model is also your text retriever, 2024. 6, 7", + "[19] Ranjay Krishna, Yuke Zhu, Oliver Groth, Justin Johnson, Kenji Hata, Joshua Kravitz, Stephanie Chen, Yannis Kalantidis, Li-Jia Li, David A. Shamma, Michael S. Bernstein, and Li Fei-Fei. Visual genome: Connecting language and vision using crowdsourced dense image annotations. International Journal of Computer Vision, 123(1):32-73, 2017. 2", + "[20] Bo Li, Yuanhan Zhang, Dong Guo, Renrui Zhang, Feng Li, Hao Zhang, Kaichen Zhang, Peiyuan Zhang, Yanwei Li, Ziwei Liu, and Chunyuan Li. Llava-onevision: Easy visual task transfer. arXiv preprint arXiv:2408.03326, 2024. 2, 3, 6, 7", + "[21] Zhuang Liu, Hanzi Mao, Chao-Yuan Wu, Christoph Feichtenhofer, Trevor Darrell, and Saining Xie. A convnet for the 2020s. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 11976-11986, 2022. 6", + "[22] Pan Lu, Tony Xia, Weicheng Shi, Ahmed El Kholy, Xi Victor Lin, Jianfeng Gao, Xiang Chen, and Kai-Wei Chang. Learn" + ], + "bbox": [ + 516, + 92, + 903, + 900 + ], + "page_idx": 8 + }, + { + "type": "page_number", + "text": "24825", + "bbox": [ + 478, + 945, + 517, + 955 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "to explain: Multimodal reasoning via thought chains for science question answering. In Advances in Neural Information Processing Systems, 2022. 2", + "[23] Pan Lu, Hritik Bansal, Tony Xia, Jiacheng Liu, Chunyuan Li, Hannaneh Hajishirzi, Hao Cheng, Kai-Wei Chang, Michel Galley, and Jianfeng Gao. Mathvista: Evaluating mathematical reasoning of foundation models in visual contexts. arXiv preprint arXiv:2310.02255, 2023. 2, 3", + "[24] Hongyin Luo, Yung-Sung Chuang, Yuan Gong, Tianhua Zhang, Yoon Kim, Xixin Wu, Danny Fox, Helen Meng, and James Glass. SAIL: Search-augmented instruction learning. arXiv preprint arXiv:2305.15225, 2023. 3", + "[25] Kenneth Marino, Mohammad Rastegari, Ali Farhadi, and Roozbeh Mottaghi. Ok-vqa: A visual question answering benchmark requiring external knowledge. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition, pages 3195-3204, 2019. 2", + "[26] Ahmed Masry, Do Xuan Long, Jia Qing Tan, Shafiq Joty, and Enamul Hoque. Chartqa: A benchmark for question answering about charts with visual and logical reasoning. In Findings of the Association for Computational Linguistics: ACL 2022, pages 2263-2279, 2022. 2", + "[27] Minesh Mathew, Viraj Bagal, Ruben Pérez Tito, Dimosthenis Karatzas, Ernest Valveny, and CV Jawahar. Infographicvqa. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 4390-4399, 2021. 2, 3", + "[28] Minesh Mathew, Dimosthenis Karatzas, and CV Jawahar. Docvqa: A dataset for vqa on document images. In Proceedings of the IEEE/CVF Winter Conference on Applications of Computer Vision, pages 2200-2209, 2021. 2, 3", + "[29] Zach Nussbaum, Brandon Duderstadt, and Andriy Mulyar. Nomic embed vision: Expanding the latent space, 2024. 6, 7", + "[30] OpenAI. Gpt-4o: Enhanced multimodal language model. OpenAI Research, 2024. https://openai.com/index/hello-gpt-4o/.2,3,4,6,7", + "[31] OpenAI. Gpt-4v: Multimodal language model with vision capabilities. OpenAI Research, 2024. https://openai.com/index/gpt-4/.2", + "[32] Abhirama Subramanyam Penamakuri, Manish Gupta, Mithun Das Gupta, and Anand Mishra. Answer mining from a pool of images: Towards retrieval-based visual question answering. In Proceedings of the International Joint Conference on Artificial Intelligence (IJCAI), pages 1052-1058, 2023. 1, 2, 3", + "[33] Alec Radford, Jong Wook Kim, Karthik Hallacy, Aditya Ramesh, Gabriel Goh, Sandhini Agarwal, Girish Sastry, Amanda Askell, Pamela Mishkin, Jack Clark, et al. Learning transferable visual models from natural language supervision. In Proceedings of the 38th International Conference on Machine Learning, pages 8748-8763. PMLR, 2021. 3, 6, 7", + "[34] Shaoqing Ren, Kaiming He, Ross Girshick, and Jian Sun. Faster r-cnn: Towards real-time object detection with region proposal networks. Advances in Neural Information Processing Systems, 28, 2015. 3", + "[35] Dustin Schwenk, Apoorv Khandelwal, Christopher Clark, Kenneth Marino, and Roozbeh Mottaghi. A-okvqa: A" + ], + "bbox": [ + 91, + 90, + 482, + 900 + ], + "page_idx": 9 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "benchmark for visual question answering using world knowledge. In European Conference on Computer Vision (ECCV), pages 1-17, 2022. 2", + "[36] Ray Smith et al. TesseractOCR engine. https://github.com/tesseract-ocr/tesseract, 2024. Accessed: 2024-11-06. 4", + "[37] Alon Talmor, Sewon Min, Robin Jia, Yanai Elazar, Uriel Singer Hasson, and Danqi Chen. Multimodalqa: Complex question answering over text, tables, and images. In Proceedings of the International Conference on Learning Representations (ICLR), 2021. 2", + "[38] Ryota Tanaka, Kyosuke Nishida, Kosuke Nishida, Taku Hasegawa, Itsumi Saito, and Kuniko Saito. Slidevqa: A dataset for document visual question answering on multiple images. In Proceedings of the AAAI Conference on Artificial Intelligence, pages 13636-13645, 2023. 2", + "[39] Ke Wang, Yichi Zhang, and Hongsheng Li. Mini-gemini: An efficient and versatile vision-language model. arXiv preprint arXiv:2310.12345, 2023. 2", + "[40] Ke Wang, Junting Pan, Weikang Shi, Zimu Lu, Mingjie Zhan, and Hongsheng Li. Measuring multimodal mathematical reasoning with math-vision dataset. arXiv preprint arXiv:2402.14804, 2024. 2", + "[41] Peng Wang, Shuai Bai, Sinan Tan, Shijie Wang, Zhihao Fan, Jinze Bai, Keqin Chen, Xuejing Liu, Jialin Wang, Wenbin Ge, et al. Qwen2-vl: Enhancing vision-language model's perception of the world at any resolution. arXiv preprint arXiv:2409.12191, 2024. 2, 3, 7", + "[42] Weiyun Wang, Shuibio Zhang, Yiming Ren, Yuchen Duan, Tiantong Li, Shuo Liu, Mengkang Hu, Zhe Chen, Kaipeng Zhang, Lewei Lu, et al. Needle in a multimodal haystack. arXiv preprint arXiv:2406.07230, 2024. 2", + "[43] Tsung-Han Wu, Giscard Biamby, Jerome Quenum, Ritwik Gupta, Joseph E. Gonzalez, Trevor Darrell, and David M. Chan. Visual haystacks: A vision-centric needle-in-a-haystack benchmark. arXiv preprint arXiv:2407.13766, 2024. 2, 3, 7", + "[44] Xiang Yue, Yuansheng Ni, Kai Zhang, Tianyu Zheng, Ruoqi Liu, Ge Zhang, Samuel Stevens, Dongfu Jiang, Weiming Ren, Yuxuan Sun, et al. Mmmu: A massive multi-discipline multimodal understanding and reasoning benchmark for expert agi. arXiv preprint arXiv:2311.16502, 2024. 2, 3", + "[45] Xiaohua Zhai, Basil Mustafa, Alexander Kolesnikov, and Lucas Beyer. Sigmoid loss for language image pre-training. In Proceedings of the IEEE/CVF International Conference on Computer Vision (ICCV), pages 11975-11985, 2023. 3, 6, 7", + "[46] Renrui Zhang, Dongzhi Jiang, Yichi Zhang, Haokun Lin, Ziyu Guo, Pengshuo Qiu, Aojun Zhou, Pan Lu, Kai-Wei Chang, Peng Gao, and Hongsheng Li. Mathverse: Does your multi-modal llm truly see the diagrams in visual math problems? arXiv preprint arXiv:2403.14624, 2024. 2, 3", + "[47] Deyao Zhu, Jun Chen, Xiaogian Shen, Xiang Li, and Mohamed Elhoseiny. Minigpt-4: Enhancing vision-language understanding with advanced large language models. In International Conference on Learning Representations (ICLR), 2024. 2, 3" + ], + "bbox": [ + 516, + 92, + 903, + 895 + ], + "page_idx": 9 + }, + { + "type": "page_number", + "text": "24826", + "bbox": [ + 478, + 945, + 519, + 955 + ], + "page_idx": 9 + } +] \ No newline at end of file diff --git a/2025/Document Haystacks_ Vision-Language Reasoning Over Piles of 1000+ Documents/1b0623c9-cc0a-4950-bcf6-716e67ed00b3_model.json b/2025/Document Haystacks_ Vision-Language Reasoning Over Piles of 1000+ Documents/1b0623c9-cc0a-4950-bcf6-716e67ed00b3_model.json new file mode 100644 index 0000000000000000000000000000000000000000..2dffebc572ac0603f19d9f2bac10dd27213d80ba --- /dev/null +++ b/2025/Document Haystacks_ Vision-Language Reasoning Over Piles of 1000+ Documents/1b0623c9-cc0a-4950-bcf6-716e67ed00b3_model.json @@ -0,0 +1,1859 @@ +[ + [ + { + "type": "header", + "bbox": [ + 0.107, + 0.003, + 0.182, + 0.044 + ], + "angle": 0, + "content": "CVF" + }, + { + "type": "header", + "bbox": [ + 0.237, + 0.001, + 0.812, + 0.047 + ], + "angle": 0, + "content": "This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore." + }, + { + "type": "title", + "bbox": [ + 0.177, + 0.131, + 0.825, + 0.175 + ], + "angle": 0, + "content": "Document Haystacks: Vision-Language Reasoning Over Piles of 1000+ Documents" + }, + { + "type": "text", + "bbox": [ + 0.19, + 0.203, + 0.806, + 0.294 + ], + "angle": 0, + "content": "Jun Chen\\(^{1}\\), Dannong Xu\\(^{2}\\), Junjie Fei\\(^{1}\\), Chun-Mei Feng\\(^{3}\\), Mohamed Elhoseiny\\(^{1}\\) \n\\(^{1}\\)King Abdullah University of Science and Technology \n\\(^{2}\\)The University of Sydney, \\(^{3}\\)IHPC, A*STAR \n{jun.chen, junjie. fei, mohamed.elhoseiny}@kaust.edu.sa \ndaxu8019@uni.sydney.edu.au, fengcm.ai@gmail.com" + }, + { + "type": "image", + "bbox": [ + 0.096, + 0.325, + 0.482, + 0.525 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.166, + 0.532, + 0.411, + 0.556 + ], + "angle": 0, + "content": "(a) Previous Benchmarks: Each question paired with a limited image set" + }, + { + "type": "image", + "bbox": [ + 0.489, + 0.326, + 0.898, + 0.525 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.539, + 0.531, + 0.849, + 0.556 + ], + "angle": 0, + "content": "(b) Our Benchmarks: All questions mapped to an extensive document collection" + }, + { + "type": "image_caption", + "bbox": [ + 0.089, + 0.566, + 0.908, + 0.637 + ], + "angle": 0, + "content": "Figure 1. Comparison between previous and proposed benchmarks. Given a question as input, all benchmarks aim to retrieve relevant images from an image pool to correctly answer the question. Unlike prior benchmarks like RetVQA [32] and WebVQA [7], which structure their datasets by pairing each question with a limited set of images (typically \\(\\leq 30\\)), our benchmarks, DocHaystack and InfoHaystack, map each question to a substantially larger document collection, scaling up to 1,000 visual documents. This expanded scope more accurately represents large-scale document retrieval scenarios and offers a greater challenge in retrieval accuracy and visual question answering." + }, + { + "type": "title", + "bbox": [ + 0.248, + 0.654, + 0.327, + 0.67 + ], + "angle": 0, + "content": "Abstract" + }, + { + "type": "text", + "bbox": [ + 0.089, + 0.688, + 0.486, + 0.87 + ], + "angle": 0, + "content": "Large multimodal models (LMMs) have achieved impressive progress in vision-language understanding, yet they face limitations in real-world applications requiring complex reasoning over a large number of images. Existing benchmarks for multi-image question-answering are limited in scope, each question is paired with only up to 30 images, which does not fully capture the demands of large-scale retrieval tasks encountered in the real-world usages. To reduce these gaps, we introduce two document haystack benchmarks, dubbed DocHaystack and InfoHaystack, designed to evaluate LMM performance on large-scale visual document retrieval and understanding. Addition" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.655, + 0.909, + 0.868 + ], + "angle": 0, + "content": "ally, we propose V-RAG, a novel, vision-centric retrieval-augmented generation (RAG) framework that leverages a suite of multimodal vision encoders, each optimized for specific strengths, and a dedicated question-document relevance module. V-RAG sets a new standard, with a \\(9\\%\\) and \\(11\\%\\) improvement in Recall@1 on the challenging DocHaystack-1000 and InfoHaystack-1000 benchmarks, respectively, compared to the previous best baseline models. Additionally, integrating V-RAG with LMMs enables them to efficiently operate across thousands of images, yielding significant improvements on our DocHaystack and InfoHaystack benchmarks. Our code and datasets are available at https://github.com/Vision-CAIR/dochaystacks" + }, + { + "type": "page_footnote", + "bbox": [ + 0.116, + 0.888, + 0.228, + 0.9 + ], + "angle": 0, + "content": "* Equal contribution" + }, + { + "type": "page_number", + "bbox": [ + 0.479, + 0.945, + 0.52, + 0.957 + ], + "angle": 0, + "content": "24817" + } + ], + [ + { + "type": "title", + "bbox": [ + 0.093, + 0.091, + 0.223, + 0.106 + ], + "angle": 0, + "content": "1. Introduction" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.115, + 0.482, + 0.326 + ], + "angle": 0, + "content": "Large Multimodal Models (LMMs) [1, 20, 30, 41] have made remarkable progress in the vision-language understanding. However, these models still face challenges when tasked with reasoning over extensive collections of images or documents [43], limiting their effectiveness in real-world applications, such as visual search or querying over large sets of images or documents, like those stored on personal devices or in photo albums. However, there lacks such proper benchmarks to evaluate these capabilities. To address this gap, we introduce the DocHaystack and InfoHaystack benchmarks, designed to evaluate LMMs on large-scale image retrieval and understanding capabilities, pushing the boundaries of LMM performance in complex, real-world scenarios." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.327, + 0.483, + 0.537 + ], + "angle": 0, + "content": "The existing multi-image retrieving and reasoning benchmarks are primarily constructed on a small scale, as highlighted in works such as [32, 37]. Each question in these benchmarks is paired with only up to 30 images as illustrated in Figure 1 (a). However, this limited scope does not align well with real-world scenarios, which often require retrieval and reasoning across hundreds or thousands of images or documents. In contrast, we introduce the Document Haystacks benchmarks, depicted in Figure 1 (b), which allow for querying questions from a large-scale collection of up to 1,000 documents, necessitating that models retrieve and reason over an extensive set of documents for each question. This scale better simulates practical applications and their demands." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.539, + 0.483, + 0.81 + ], + "angle": 0, + "content": "The main challenge for constructing our document haystack benchmarks is collecting specific questions while ensuring there are no ambiguous answers across a large set of images. Existing datasets, such as those in DocVQA and InfographicVQA [27, 28], contain numerous \"general\" questions, like \"What is the table number?\", where answers could be derived from multiple images, leading to nonunique responses. To address this, we implemented a rigorous data filtering pipeline. First, we employed both a large language model (LLM) and human annotators to systematically filter out \"general\" questions based on carefully defined criteria. Additionally, we used the LLM to exclude questions relying on generic knowledge, such as \"What is the capital of Missouri?\", which can be answered without image context. This approach ensures that the questions in the benchmark can only be answered through specific visual cues from the provided images, maintaining the benchmark's integrity for evaluating image-based understanding." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.811, + 0.483, + 0.902 + ], + "angle": 0, + "content": "To enable the current LMMs effectively reason over a large number of images, we propose a vision-centric retrieval-augmented generation (RAG) framework, named V-RAG. V-RAG combines multiple multimodal vision encoders, leveraging each encoder's unique strengths to enhance retrieval accuracy. Additionally, it incorporates an" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.092, + 0.905, + 0.272 + ], + "angle": 0, + "content": "LMM-filter module to assess the relevance of each document to the query, refining the retrieval process by ensuring that only relevant documents are prioritized. This integrated approach allows V-RAG to navigate extensive document collections efficiently. Experimental results demonstrate that V-RAG achieves \\(9\\%\\) and \\(11\\%\\) improvement in Recall@1 on the DocHaystack-1000 and InfoHaystack-1000 compared to previous best text-to-image retrieval methods. Additionally, we found that integrating V-RAG brings GPT4o over a \\(55\\%\\) acc improvement on DocHaystack-200 and a \\(34\\%\\) acc improvement on InfoHaystack-200, indicating the effectiveness of our V-RAG." + }, + { + "type": "text", + "bbox": [ + 0.533, + 0.273, + 0.753, + 0.286 + ], + "angle": 0, + "content": "Our contributions are as follows:" + }, + { + "type": "text", + "bbox": [ + 0.514, + 0.289, + 0.905, + 0.409 + ], + "angle": 0, + "content": "- We introduce the Document Haystack benchmarks, including DocHaystack-100/200/1000 and InfoHaystack-100/200/1000, with the most challenging setup consisting of 1,000 documents for each inquiry. These benchmarks advance document retrieval and reasoning tasks by requiring models to navigate and reason across extensive document collections, surpassing prior benchmarks limited to smaller retrieval tasks." + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.41, + 0.905, + 0.501 + ], + "angle": 0, + "content": "- We propose a vision-centric retrieval-augmented generation framework, V-RAG, which enhances the retrieval capabilities of LMMs. V-RAG achieves substantial improvements over previous best text-to-image retrieval methods by \\(9\\%\\) and \\(11\\%\\) on DocHaystack-1000 and InfoHaystack-1000, respectively." + }, + { + "type": "list", + "bbox": [ + 0.513, + 0.289, + 0.905, + 0.501 + ], + "angle": 0, + "content": null + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.513, + 0.665, + 0.529 + ], + "angle": 0, + "content": "2. Related Works" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.538, + 0.905, + 0.825 + ], + "angle": 0, + "content": "VQA benchmarks. VQA play a critical role in assessing a model's ability to understand and reason across visual contexts [6, 11, 12]. Traditional VQA datasets typically measure a model's comprehension of object attributes [15, 19], spatial relationships [15], as well as its understanding of documents [27, 28], charts [26], mathematics [23, 40, 46], and open knowledges [25, 35]. Additionally, these benchmarks explore models' knowledge across varied fields, including science and the arts [22, 44]. This broad array of benchmarks has greatly advanced vision-language models by cultivating diverse visual comprehension skills, particularly for modern foundation models in vision-language understanding [1, 3, 8, 20, 22, 30, 31, 39, 47]. Notably, these benchmarks have primarily focused on question answering within single image or document. In contrast, our benchmark shifts the focus towards retrieval and comprehensive understanding across a large collection of visual documents, presenting new challenges and expanding the scope of visual question answering." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.826, + 0.905, + 0.901 + ], + "angle": 0, + "content": "Several previous efforts have tackled the challenge of visual question answering and reasoning across multiple images [5, 7, 32, 37, 38, 42]. For instance, datasets such as MultimodalQA [37] and ISVQA [5] require models to have multi-image reasoning abilities. Meanwhile, WebQA [7]" + }, + { + "type": "page_number", + "bbox": [ + 0.48, + 0.945, + 0.521, + 0.957 + ], + "angle": 0, + "content": "24818" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.109, + 0.09, + 0.891, + 0.21 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.089, + 0.231, + 0.908, + 0.304 + ], + "angle": 0, + "content": "Figure 2. Data Curation Pipeline. Our benchmarks are curated based on the DocVQA and InfographicVQA datasets, following a three-step filtering process to obtain document-specific question-answer pairs. In Step 1, we filter out general questions (e.g., \"What is the table number?\"), as these could be answered by multiple documents and lack specificity. Step 2 involves a manual review by human annotators to further remove general questions. In Step 3, we eliminate generic-knowledge questions (e.g., \"How many sports were in the 2008 Beijing Paralympic Games?\") that can be answered directly by large language models without requiring image input." + }, + { + "type": "text", + "bbox": [ + 0.089, + 0.317, + 0.486, + 0.483 + ], + "angle": 0, + "content": "and RetVQA [32] involve an additional step where models must first retrieve relevant images from a limited image pool before answering visual questions based on these results. However, these benchmarks are generally constrained to relatively small image pools, where each question is paired with an image set containing up to 30 images. In contrast, our proposed benchmarks, DocHaystack and InfoHaystack, significantly expand this scope by requiring models to retrieve and reason from a much larger set of up to 1,000 documents, presenting a notably greater challenge in retrieval and multi-image reasoning." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.488, + 0.485, + 0.732 + ], + "angle": 0, + "content": "Large multimodal models (LMMs). LMMs have achieved substantial advancements in understanding and reasoning across single or multiple images [1, 8, 20, 30, 41, 47]. These models have significantly enhanced vision-language understanding across numerous dimensions and applications [12, 23, 44, 46]. LMMs benefit primarily from large-scale image-text alignment and extensive language modeling, which emerge them with advanced understanding and reasoning abilities. However, despite these breakthroughs, LMMs still encounter challenges when handling large-scale image or document sets [43]. This difficulty is due to the inherent complexity of processing such complex data. To address this, retrieval-based approaches have been developed to extend the capacity of vision-language models, augmenting their ability to process and reason over a larger number of images." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.735, + 0.486, + 0.904 + ], + "angle": 0, + "content": "Retrieval-augmented generation (RAG). RAG integrates retrieval systems [4, 11, 16, 33, 45], with generative models, enhancing them with additional knowledge. While RAG has been extensively explored in language domains [2, 13, 17, 24], its application in vision-language contexts is also advancing. In vision-language RAG, models like MuRAG [9] leverage image-text memory to retrieve top-k neighbors by comparing inner-product similarities. RetVQA [32] uses an image-question relevance encoder, combining BERT [10] and Faster R-CNN [34] to filter relevant images, while MIRAGE [43] employs a CLIP-based" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.317, + 0.909, + 0.47 + ], + "angle": 0, + "content": "encoder [33] to train a retriever. These frameworks extend model capabilities, enabling retrieval and reasoning across hundreds or thousands of images. In contrast, we propose V-RAG, a vision-centric RAG framework that integrates multiple vision encoders to more effectively capture image features, and introduces a LMM-based question-document relevance comparison module. Our results demonstrate that V-RAG surpasses existing methods on our DocHaystack and InfoHaystack benchmarks, setting a new standard for large-scale visual retrieval and reasoning." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.481, + 0.907, + 0.515 + ], + "angle": 0, + "content": "3. DocHaystack and InfoHaystack Benchmarks" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.524, + 0.909, + 0.658 + ], + "angle": 0, + "content": "To support effective retrieval and reasoning across extensive document collections, we present two new benchmarks—DocHaystack and InfoHaystack—designed to ensure each question yields a unique, document-specific answer. Derived from DocVQA [28] and InfographicVQA [27], these benchmarks address the challenge of answer ambiguity by selectively curating questions that can only be answered by a single document within a large dataset." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.66, + 0.911, + 0.902 + ], + "angle": 0, + "content": "Benchmark construction pipeline. There exists many general questions in the existing benchmarks and lead to multiple answers for different document context. For example, general questions like \"What is the table number?\" may apply to various documents and yield multiple valid answers, while a targeted question like \"Who is the reviewer for the article titled 'An antithyroid factor in milk'?\" is likely to produce a unique answer, as only a single document or a limited set of documents would contain that information. Therefore, our benchmark construction follows a structured three-step filtering pipeline, illustrated in Figure 2, to ensure high-quality, unique-answer questions. First, we employ a large language model (LLM) to filter out general questions that could generate multiple answers across documents. Next, a manual review step further" + }, + { + "type": "page_number", + "bbox": [ + 0.479, + 0.945, + 0.521, + 0.958 + ], + "angle": 0, + "content": "24819" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.103, + 0.089, + 0.898, + 0.366 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.089, + 0.392, + 0.908, + 0.477 + ], + "angle": 0, + "content": "Figure 3. The V-RAG pipeline workflow. In the top section, a vision encoder ensemble is used, combining multiple vision models—CLIP, SigLIP, and OpenCLIP—to process a large document haystack. Each encoder computes similarity scores, which are averaged into \\( Sim_{avg} \\). The top m documents, based on these scores, are selected for further analysis. In the bottom right, the LMM-Filter Module utilizes a pretrained LMM to assess whether each selected document can potentially answer the posed question. This filtering step removes documents that do not match, retaining only relevant ones. Finally, the top k most relevant images are input into the LMM along with the original question \\( q \\) to generate a specific answer." + }, + { + "type": "text", + "bbox": [ + 0.089, + 0.492, + 0.482, + 0.55 + ], + "angle": 0, + "content": "checks the questions to ensure the data quality. Finally, a generic-knowledge filtering stage refines the dataset further, retaining only questions closely tied to specific document content." + }, + { + "type": "text", + "bbox": [ + 0.089, + 0.552, + 0.483, + 0.673 + ], + "angle": 0, + "content": "This carefully designed pipeline, combining LLM-based filtering and human review, effectively curates questions that drive accurate, document-specific retrieval. By focusing on reducing answer ambiguity, DocHaystack and InfoHaystack enhance the precision of retrieval and reasoning in large-scale document processing tasks, providing a valuable tool for the evaluation of retrieval systems. We discuss this data curation pipeline in details as follows:" + }, + { + "type": "text", + "bbox": [ + 0.089, + 0.673, + 0.483, + 0.809 + ], + "angle": 0, + "content": "General-question LLM filtering. We begin by using the LLM, GPT-4o [30], to filter out general questions through a set of well-crafted instructions. Leveraging the LLM's strong contextual understanding, this initial filtering step allows us to efficiently process large volumes of data, identifying broad or ambiguous questions that may yield multiple answers across documents. This automated approach significantly enhances the benchmark construction's efficiency and quality." + }, + { + "type": "text", + "bbox": [ + 0.089, + 0.81, + 0.483, + 0.885 + ], + "angle": 0, + "content": "To guide the LLM, we first define the task, providing clear distinctions between general and specific questions along with illustrative examples. With this framework, the LLM can then assess each question and determine if it is general or specific. The instructional format is as follows:" + }, + { + "type": "text", + "bbox": [ + 0.11, + 0.886, + 0.483, + 0.902 + ], + "angle": 0, + "content": "You are an evaluator tasked with identifying if a ques" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.492, + 0.905, + 0.612 + ], + "angle": 0, + "content": "tion is specific or general. A general question seeks commonly known or widely applicable information without unique identifiers, e.g., \"Who is the person standing in the ground?\" A specific question, however, requests unique information about a particular individual, event, or object, e.g., \"What is the Social Security Number of Charles Yarbrough?\" Based on these definitions, determine if the following question is general or specific: {question}." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.613, + 0.905, + 0.688 + ], + "angle": 0, + "content": "General-question manual review. After the initial LLM filtering, we conduct a manual review of the questions that were classified as specific. This manual process involves two key steps to ensure answer uniqueness and benchmark quality." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.689, + 0.905, + 0.779 + ], + "angle": 0, + "content": "In the first step, we examine each question to confirm it contains unique identifiers—such as names, dates, titles, or other specific attributes—suggesting a document-specific answer. This careful check helps identify questions with clear, unique markers that direct the retrieval process to a single document." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.78, + 0.907, + 0.901 + ], + "angle": 0, + "content": "In the second step, we verify the uniqueness of each answer to eliminate any remaining ambiguity. Although specific identifiers are present, questions may still be prone to ambiguity, such as with common names or recurring book titles. To address this, we employ a refined verification process. First, we use Optical Character Recognition (OCR) [36] to extract all text from images in the dataset. We then search for occurrences of the unique identifiers re" + }, + { + "type": "page_number", + "bbox": [ + 0.479, + 0.945, + 0.521, + 0.958 + ], + "angle": 0, + "content": "24820" + } + ], + [ + { + "type": "table", + "bbox": [ + 0.095, + 0.089, + 0.482, + 0.144 + ], + "angle": 0, + "content": "
GPT-4oLLaVA-OneVisionQwen2-VL
DocVQA27.0%4.7%3.4%
InfographicVQA56.4%13.4%11.3%
" + }, + { + "type": "table_caption", + "bbox": [ + 0.09, + 0.153, + 0.483, + 0.238 + ], + "angle": 0, + "content": "Table 1. Percentage of questions answerable by LMMs without vision input. We evaluate GPT-4o, LLaVA-Onevision, and Qwen2-VL on their ability to answer questions directly from our dataset without requiring vision input. The reported percentage reflects the proportion of examples that can be answered solely through language understanding." + }, + { + "type": "text", + "bbox": [ + 0.089, + 0.265, + 0.483, + 0.355 + ], + "angle": 0, + "content": "tained from the first step across other documents. If matches are found, a manual review is conducted to ensure no alternative valid answers exist. This comprehensive approach minimizes the possibility of a single question mapping to multiple answers, enhancing the precision and reliability of our benchmarks." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.356, + 0.483, + 0.401 + ], + "angle": 0, + "content": "Generic-knowledge filtering. In DocVQA and InfographicVQA tasks, certain questions—such as “How many sports were in the 2008 Beijing" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.402, + 0.483, + 0.537 + ], + "angle": 0, + "content": "Paralympic Games?\"—can be answered based on general knowledge accessible to a large language model, without relying on the image content. This introduces a language bias when using LMMs for visual question answering, as it shifts the focus away from image-based reasoning. To address this, we filter out these general-knowledge questions, ensuring that evaluation emphasizes vision-based understanding and that models rely primarily on visual content to generate accurate answers." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.538, + 0.483, + 0.78 + ], + "angle": 0, + "content": "To implement this, we developed an LLM-based evaluation pipeline that detects and excludes such questions. For each question, we prompt an LLM with “{question}, answer briefly.” After receiving a response, we compare it to the ground-truth answer using another LLM. If the response matches the ground truth, we classify the question as general knowledge-related and remove it, thereby isolating questions that truly require visual document understanding. As shown in Table 1, GPT-4o accurately answers \\(27.0\\%\\) of DocVQA questions and \\(56.4\\%\\) of InfographicVQA questions directly, a rate significantly higher than that of open-source LLMs. Therefore, we select GPT-4o to filter out the questions that can be directly answered by the GPT-4o model. Overall, this process is to ensure that the evaluation reflects the necessity of vision-based comprehension." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.781, + 0.483, + 0.902 + ], + "angle": 0, + "content": "Final dataset profile. After a rigorous three-stage data filtering process, we retained 109 questions from DocVQA and 155 questions from InfographicVQA, associated with 59 and 66 documents that provide the evidence, respectively. To assess retrieval performance at scale, we introduce two benchmarks: DocHaystack-1000 and InfoHaystack-1000, where each question requires retrieving relevant content from a set of 1,000 documents. Given" + }, + { + "type": "image", + "bbox": [ + 0.526, + 0.095, + 0.688, + 0.231 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.718, + 0.096, + 0.889, + 0.236 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.512, + 0.246, + 0.908, + 0.289 + ], + "angle": 0, + "content": "Figure 4. Question type analysis. We analyze the distribution of question types of DocHaystack and InfoHaystack. Each benchmark categorizes the data into 5 different types." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.301, + 0.907, + 0.452 + ], + "angle": 0, + "content": "the challenge this scale presents to current LMMs, particularly in terms of context length limitations, we also construct two smaller benchmarks: DocHaystack-100/200 and InfoHaystack-100/200. These benchmarks allow direct input of all associated images into the context, enabling evaluation of models' long-context comprehension ability. For training set, we also construct a dataset comprising 2,835 questions similarly, with 899 from DocVQA and 1,936 from InfographicVQA, to support robust learning and generalization for the multi-image reasoning tasks." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.453, + 0.909, + 0.619 + ], + "angle": 0, + "content": "Question type analysis. The types of questions represent the types of the evidence required for accurate answers. In Figure 4, we illustrate the distribution of question types across our dataset to provide insights into its structure. Following the classification system used in DocVQA and InfographicVQA, we categorize questions accordingly (note that a single question may fall into multiple categories). As shown in the figure, the DocHaystack benchmark places a greater emphasis on Table/List and Layout understanding, whereas InfoHaystack primarily targets Figure, Text, and Table comprehension." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.633, + 0.606, + 0.649 + ], + "angle": 0, + "content": "4. Method" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.659, + 0.907, + 0.87 + ], + "angle": 0, + "content": "Current large multimodal models (LMMs) face substantial challenges when reasoning across hundreds or thousands of images, due not only to context length limitations but also to the inherent complexity of the task. This issue is particularly pronounced in our benchmarks, which contain 1k document files requiring high-resolution images to capture and interpret small-font text effectively. To enable LMMs to perform reasoning over a substantial number of documents, we introduce a vision-centric retrieval-augmented generation (V-RAG) framework. V-RAG efficiently retrieves a reduced set of relevant documents, allowing the LMM to focus on a manageable subset for deeper understanding, as illustrated in Figure 3. In the following section, we provide a detailed description of the V-RAG pipeline." + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.871, + 0.909, + 0.903 + ], + "angle": 0, + "content": "Task definition. Given a question \\( q \\) and a collection of \\( N \\) documents \\( \\mathcal{D} = \\{D_1, \\ldots, D_N\\} \\), the V-RAG framework" + }, + { + "type": "page_number", + "bbox": [ + 0.48, + 0.945, + 0.518, + 0.957 + ], + "angle": 0, + "content": "24821" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.09, + 0.092, + 0.482, + 0.168 + ], + "angle": 0, + "content": "aims to retrieve the top-\\(k\\) most relevant documents to support LMMs understanding and answering the question \\(q\\). V-RAG accomplishes this through a two-step retrieval process designed to effectively identify and rank relevant documents for each question." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.17, + 0.483, + 0.397 + ], + "angle": 0, + "content": "Vision encoder ensemble. Document files often contain a mix of text, symbols, and visual elements across various scales, requiring vision encoders to capture a comprehensive understanding of these complex structures. To efficiently handle this diversity, we represent each document as an image and utilize an ensemble of vision encoders, including CLIP [33], SigLIP [45], and OpenCLIP [16], each bringing distinct strengths to the image understanding, as depicted in Figure 3. For example, the ConvNext encoder [21] from OpenCLIP [16] is particularly effective for high-resolution image encoding. We compute the similarity score between each question \\( q \\) and all documents in the document set \\( \\mathcal{D} \\) according to Equation 1, with similarity scores from each encoder represented as \\( Sim_{c} \\), \\( Sim_{o} \\), and \\( Sim_{s} \\) respectively." + }, + { + "type": "equation", + "bbox": [ + 0.149, + 0.423, + 0.483, + 0.44 + ], + "angle": 0, + "content": "\\[\n\\mathcal {S} (q, \\mathcal {D}) = \\cos \\left(\\phi_ {\\mathrm {t}} (q), \\phi_ {\\mathrm {v}} (D _ {j})\\right) \\mid D _ {j} \\in \\mathcal {D}, \\tag {1}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.454, + 0.483, + 0.515 + ], + "angle": 0, + "content": "where \\(S\\) denotes the computing the similarity between the query \\(q\\) and a collection of documents \\(\\mathcal{D}\\). \\(\\cos\\) denotes the cosine similarity. \\(\\phi_{\\mathrm{t}}\\) denotes the text encoder, and \\(\\phi_{\\mathrm{v}}\\) denotes the vision encoder." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.518, + 0.483, + 0.594 + ], + "angle": 0, + "content": "To derive a final relevance score, we calculate the average similarity \\( Sim_{avg} \\) for each question-image pair by combining \\( Sim_c \\), \\( Sim_o \\), and \\( Sim_s \\). We then rank the images based on \\( Sim_{avg} \\) in descending order, selecting the top-\\( m \\) most relevant images according to their similarity scores." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.597, + 0.483, + 0.763 + ], + "angle": 0, + "content": "LMM-filter module. To refine the selection of top-\\(m\\) relevant images further, we introduce a LMM-based question-image relevance assessment module. This module evaluates the relevance between each question and the top-\\(m\\) images identified in the first filtering step. Specifically, we pair each image with the question text and input them into an open-source vision-language model, prompting, \"Can this image provide answers to this question? Respond only with yes or no\". We only retain the question-image pairs that are identified as \"yes\" from LMM, and remove other irrelevant images." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.765, + 0.483, + 0.902 + ], + "angle": 0, + "content": "LMM-VQA module. Achieving high top-1 ranking accuracy in image retrieval is challenging, so we retain the top-\\(k\\) images from the LMM-filtered ranking list and present them to the LMM-VQA to improve the likelihood of including relevant images. We input these top-\\(k\\) images alongside the question into the LMM-VQA (see Figure 3), which then generates the answer directly. To enhance robustness against visual distractors, the LMM-VQA can be further optimized, as analyzed in the experiment section." + }, + { + "type": "title", + "bbox": [ + 0.514, + 0.09, + 0.646, + 0.108 + ], + "angle": 0, + "content": "5. Experiments" + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.118, + 0.905, + 0.195 + ], + "angle": 0, + "content": "In the experiments section, we will primarily describe our training setup, covering evaluation metrics, baseline models, and the fine-tuning procedure for the LMM-VQA model. We also present the main experimental results along with an ablation study to provide further insights." + }, + { + "type": "title", + "bbox": [ + 0.514, + 0.212, + 0.663, + 0.229 + ], + "angle": 0, + "content": "5.1. Training setup" + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.237, + 0.905, + 0.388 + ], + "angle": 0, + "content": "Metric. In our evaluation of the DocHaystack and Info-Haystack benchmarks, we employ a model-based assessment by leveraging GPT-4o-mini [30] to accurately determine whether the model predictions match target answers. This method uses a carefully structured prompt to facilitate GPT-4o-mini's evaluation of answer correctness. We empirically found that the model-based evaluation achieves higher consistency and alignment with human judgment. Additional details on the prompt design are provided in the Appendix." + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.391, + 0.905, + 0.452 + ], + "angle": 0, + "content": "For the document retrieval evaluation, we report the baseline results using Recall@1, Recall@3, and Recall@5 metrics. These metrics enable a thorough assessment of retrieval accuracy across varying levels of precision." + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.455, + 0.906, + 0.65 + ], + "angle": 0, + "content": "Baselines. In our experiment, we have evaluated several open and closed-sourced vision-language models on the retrieval and VQA performance. For the large multimodal model, we used the gpt-4o-2024-08-06 version of GPT-4o [30], the LLaVA-OneVision-Qwen2-7b-OV-HF version of LLaVA-OneVision [20], and the Qwen2-VL-7B-Instruct version of Qwen2-VL [3]. For computing the text-to-image similarities, we employed the Jina-CLIP-v1 [18] variant, Nomic-Embed-Vision-v1.5 [29] variant, CLIP [33] ViT-L/14@336 variant, for SigLIP [45], the ViT-SO400M/14@384 variant, and for OpenCLIP [16], the ConvNeXt-XXL@1024 variant as well as text-based method, BM25." + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.654, + 0.905, + 0.715 + ], + "angle": 0, + "content": "In our V-RAG setting, we apply LLaVA-OneVision-Qwen2-7b-OV-HF for the LMM-filter module and Qwen2-VL-7B-Instruct for the LMM-VQA module. We select \\( m \\) as 60 and \\( k \\) as 5 in our experiment." + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.717, + 0.905, + 0.778 + ], + "angle": 0, + "content": "Optimizing the LMM-VQA module. To improve the robustness of the LMM-VQA model in handling visual question answering with multiple distractor images, we further fine-tune the model using our curated training data." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.78, + 0.907, + 0.902 + ], + "angle": 0, + "content": "During this fine-tuning process, we introduce 1-10 randomly sampled distractor images for each question, creating a challenging setting that encourages the model to focus on relevant content amid a mix of positive and negative images. The fine-tuning is conducted with a batch size of 32 and a peak learning rate of 1e-4 over a single epoch. Additionally, we leverage LoRA [14] with a rank of 8 to efficiently adapt the model's parameters during training." + }, + { + "type": "page_number", + "bbox": [ + 0.48, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "24822" + } + ], + [ + { + "type": "table", + "bbox": [ + 0.169, + 0.088, + 0.827, + 0.313 + ], + "angle": 0, + "content": "
DocHaystack-100DocHaystack-200DocHaystack-1000
R@1R@3R@5R@1R@3R@5R@1R@3R@5
BM25 (OCR)63.375.279.865.171.675.256.966.169.7
Jina-CLIP [18]16.531.241.39.224.830.33.77.312.8
Nomic-Embed-Vision [29]16.524.828.413.821.125.71.82.86.4
CLIP [33]46.865.169.744.058.765.123.941.345.9
SigLIP [45]51.467.976.247.763.370.633.049.557.8
OpenCLIP [16]58.775.279.856.970.675.234.949.557.8
V-RAG (ours)81.789.089.078.084.484.466.178.078.9
InfoHaystack-100InfoHaystack-200InfoHaystack-1000
R@1R@3R@5R@1R@3R@5R@1R@3R@5
BM25 (OCR)56.865.871.051.665.269.038.751.658.1
Jina-CLIP43.251.658.136.846.551.623.933.637.4
Nomic-Embed-Vision34.850.356.831.043.248.420.731.035.5
CLIP69.778.785.865.277.481.945.864.570.3
SigLIP58.171.680.055.567.776.839.455.561.9
OpenCLIP72.385.292.966.581.989.053.665.872.9
V-RAG (ours)79.491.092.974.888.488.464.574.278.1
" + }, + { + "type": "table_caption", + "bbox": [ + 0.09, + 0.325, + 0.907, + 0.381 + ], + "angle": 0, + "content": "Table 2. Retrieval Results. We compare our V-RAG model with other text-to-image and text-to-text (using OCR) retrieval methods across both benchmarks. V-RAG consistently outperforms baseline models on Recall@1, Recall@3, and Recall@5 metrics. Notably, V-RAG leverages an ensemble of text-to-image models along with a large multimodal model in a two-stage filtering approach. Top-performing values in each column are highlighted in bold." + }, + { + "type": "table", + "bbox": [ + 0.115, + 0.394, + 0.462, + 0.548 + ], + "angle": 0, + "content": "
ModelDocHaystackInfoHaystack
10020010001002001000
LLaVA-OV [20]------
GPT-4o [30]27.523.9-23.920.0-
Gemini [1]50.548.6-29.021.9-
Qwen2-VL [41]41.312.8-20.014.2-
MIRAGE [43]3.73.72.87.77.16.5
LLaVA-OV+V-RAG69.765.155.143.241.936.8
GPT-4o+V-RAG81.772.567.065.263.256.8
Gemini+V-RAG73.465.158.757.457.447.1
Qwen2-VL+V-RAG82.674.366.165.865.860.0
Qwen2-VL-f.t.+V-RAG86.279.873.467.167.760.0
" + }, + { + "type": "table_caption", + "bbox": [ + 0.09, + 0.558, + 0.483, + 0.656 + ], + "angle": 0, + "content": "Table 3. The VQA results for the DocHaystack and Info-Haystack. We evaluate with many closed-source and open-source multimodal model, and also integrating them with our V-RAG retrieval framework. - denotes that those models can not be inferred due to their token context constraints. To enable GPT-4o and Qwen2-VL to process hundreds of images, we employ low-resolution mode and adjust image size for compatibility." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.664, + 0.339, + 0.68 + ], + "angle": 0, + "content": "5.2. Main Experimental Results" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.688, + 0.483, + 0.779 + ], + "angle": 0, + "content": "We evaluated a range of open-source and closed-source vision-language models for VQA tasks. We also evaluate several text-to-image and text-to-text (with OCR) retrieval models to evaluate their retrieval capabilities on our benchmarks. More detailed performance analysis are described in the following sections." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.78, + 0.483, + 0.901 + ], + "angle": 0, + "content": "Retrieving results. The retrieval results in Table 2 demonstrate the superiority of our proposed V-RAG framework over several baseline methods across both DocHaystack and InfoHaystack benchmarks. V-RAG consistently achieves the highest Recall@1, Recall@3, and Recall@5 scores on most categories, indicating its robust retrieval capabilities. Notably, V-RAG outperforms text-based retrieving models such as BM25 and also the text-to-image retrieval models" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.397, + 0.907, + 0.624 + ], + "angle": 0, + "content": "like jina-clip, CLIP, SigLIP, and OpenCLIP by substantial margins, especially on the DocHaystack-100 subset, where it reaches Recall@1 of \\(81.7\\%\\) and Recall@5 of \\(89.0\\%\\). This pattern continues for larger datasets (DocHaystack-1000), where V-RAG remains competitive, achieving Recall@1 of \\(66.1\\%\\). It achieves the top performance across all recall metrics on DocHaystack. For InfoHaystack benchmarks, V-RAG also outperforms other models, particularly on InfoHaystack-100 and InfoHaystack-200, where it receives Recall@1 of \\(79.4\\%\\) and \\(64.5\\%\\), higher than previous best by around \\(7\\%\\) and \\(11\\%\\), respectively. This consistent performance advantage highlights the effectiveness of V-RAG's ensemble of multiple vision encoders, allowing it to capture more granular details and improve retrieval accuracy over large multimodal models." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.629, + 0.909, + 0.902 + ], + "angle": 0, + "content": "Visual question answering (VQA) results. The table presents VQA results for the DocHaystack and Info-Haystack benchmarks across varying dataset sizes (100, 200, 1000) using different multimodal models, both independently and in combination with the V-RAG framework. The results show that Qwen2-VL fine-tuned with V-RAG (Qwen2-VL-f.t.+V-RAG) achieves the highest scores across most benchmarks, with particularly notable performance on DocHaystack-100 (86.2) and InfoHaystack-100 (67.1), indicating superior retrieval and VQA capabilities in these scenarios. When V-RAG is added to other models, substantial improvements are observed, demonstrating the framework's efficacy in enhancing retrieval accuracy. For instance, GPT-4o's performance increases significantly with V-RAG, particularly for DocHaystack-100 and -200. The analysis highlights that V-RAG integration generally boosts performance across models, with Qwen2-VL-f.t.+V-RAG standing out as the top performer on both" + }, + { + "type": "page_number", + "bbox": [ + 0.48, + 0.946, + 0.518, + 0.957 + ], + "angle": 0, + "content": "24823" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.101, + 0.094, + 0.28, + 0.221 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.302, + 0.094, + 0.482, + 0.22 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.505, + 0.094, + 0.682, + 0.22 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.706, + 0.096, + 0.885, + 0.22 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.09, + 0.225, + 0.907, + 0.268 + ], + "angle": 0, + "content": "Figure 5. Top-k selection ablation analysis for LMM-VQA. We demonstrate the results for LLaVA, Qwen2-VL, GPT-4o and also the finetuned Qwen2-VL model on the DocHaystack-100/1000 and InfoHaystack-100/1000 benchmarks. All the models are integrated with our V-RAG framework. We show the VQA accuracy performance for each ablation." + }, + { + "type": "table", + "bbox": [ + 0.263, + 0.28, + 0.737, + 0.365 + ], + "angle": 0, + "content": "
CLIPSigLIPOpenCLIPLMM-filterDocHaystack-1000InfoHaystack-1000
R@1R@3R@5R@1R@3R@5
✓XXX23.941.345.945.864.570.3
X✓XX33.049.557.839.455.561.9
XX✓X34.949.557.853.665.872.9
✓✓XX40.459.662.459.467.774.2
✓✓✓X42.266.177.556.171.078.1
✓✓✓✓66.178.078.964.574.278.1
" + }, + { + "type": "table_caption", + "bbox": [ + 0.09, + 0.375, + 0.907, + 0.404 + ], + "angle": 0, + "content": "Table 4. Ablation study on the V-RAG framework components. We quantify the impact of each module for the Recall@1, Recall@3 and Recall@5 retrieval performance on the DocHaystack-1000 and InfoHaystack-1000 for our V-RAG framework." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.42, + 0.483, + 0.495 + ], + "angle": 0, + "content": "benchmarks, especially for the larger 1000-document tasks where retrieval accuracy is more challenging. This suggests that V-RAG's vision-centric, retrieval-augmented approach is highly effective for large-scale multimodal document understanding." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.497, + 0.484, + 0.587 + ], + "angle": 0, + "content": "The table also shows that the DocHaystack-1000 and InfoHaystack-1000 present significant challenges for current LMMs. The drop in performance for larger document sets, with top accuracy only reaching \\(73.4\\%\\) for DocHaystack-1000 and \\(60.0\\%\\) for InfoHaystack-1000, underscores the difficulty our benchmarks." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.603, + 0.255, + 0.618 + ], + "angle": 0, + "content": "5.3. Ablation Studies" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.627, + 0.483, + 0.809 + ], + "angle": 0, + "content": "Ablation study on Top-k Selection. This figure presents the top-k selection ablation analysis for LMM-VQA across four models: LLaVA-OV, Qwen2-VL, GPT-4o, and the fine-tuned Qwen2-VL (Qwen2-VL-f.t.), evaluated on the DocHaystack-100/1000 and InfoHaystack-100/1000 benchmarks. The analysis reports VQA accuracy as a function of top-k selection (Top 1, Top 3, and Top 5). Overall, accuracy tends to improve with larger k-values, suggesting that offering more retrieval options positively impacts model performance. However, for LLaVA-OV, there is a marked decrease in performance at top-5, indicating that this model struggles to process multiple images at this scale." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.811, + 0.483, + 0.901 + ], + "angle": 0, + "content": "Ablation study on the V-RAG framework components. The ablation study in Table 4 highlights the contributions of each component in the V-RAG framework on the DocHaystack-1000 and InfoHaystack-1000 benchmarks. Using CLIP alone yields low performance (e.g., Recall@1 of \\(23.9\\%\\) on DocHaystack-1000 and \\(45.8\\%\\) on" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.42, + 0.907, + 0.464 + ], + "angle": 0, + "content": "InfoHaystack-1000), indicating its limited retrieval capability on its own. Adding SigLIP and OpenCLIP incrementally improves results." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.472, + 0.907, + 0.607 + ], + "angle": 0, + "content": "The highest performance is achieved when all three encoders are combined with the VLM-filter module, leading to Recall@1 scores of \\(66.1\\%\\) on DocHaystack-1000 and \\(64.5\\%\\) on InfoHaystack-1000. This setup also achieves the top Recall@1, Recall@3 and Recall@5 values, demonstrating that the VLM-filter is essential for refining the ensemble outputs and significantly improving retrieval accuracy. These results confirm that each module contributes to VRAG's overall effectiveness." + }, + { + "type": "title", + "bbox": [ + 0.514, + 0.642, + 0.634, + 0.658 + ], + "angle": 0, + "content": "6. Conclusion" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.675, + 0.909, + 0.901 + ], + "angle": 0, + "content": "In this work, we introduced the DocHaystack and Info-Haystack benchmarks to evaluate LMMs for retrieving and reasoning across large-scale documents. Our benchmarks providing a more rigorous and realistic assessment of large multimodal models in real-world, large-scale retrieval scenarios. To tackle these challenges, we proposed V-RAG, a vision-centric retrieval-augmented generation framework that significantly enhances retrieval precision and overall VQA performance. V-RAG achieves this through an ensemble of vision encoders and a specialized relevance filtering module, enabling improved accuracy across diverse visual inputs. Experimental results indicate that integrating V-RAG enables both open-source and closed-source LMMs to achieve superior performance in large-scale image retrieval and complex reasoning tasks." + }, + { + "type": "page_number", + "bbox": [ + 0.48, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "24824" + } + ], + [ + { + "type": "title", + "bbox": [ + 0.093, + 0.091, + 0.188, + 0.106 + ], + "angle": 0, + "content": "References" + }, + { + "type": "ref_text", + "bbox": [ + 0.101, + 0.115, + 0.484, + 0.156 + ], + "angle": 0, + "content": "[1] Google AI. Gemini: Google's multimodal ai model. Google AI Research, 2024. https://fireflies.ai/blog/gemini-vs-gpt-4.2,3,7" + }, + { + "type": "ref_text", + "bbox": [ + 0.101, + 0.159, + 0.484, + 0.214 + ], + "angle": 0, + "content": "[2] Akari Asai, Zeqiu Wu, Yizhong Wang, Avirup Sil, and Hannaneh Hajishirzi. Self-rag: Learning to retrieve, generate, and critique through self-reflection. arXiv preprint arXiv:2310.11511, 2023. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.217, + 0.484, + 0.299 + ], + "angle": 0, + "content": "[3] Yifan Bai, Zhen Zhang, Yifan Zhang, Yuxuan Li, Yifan Zhang, Yifan Zhang, Yifan Zhang, Yifan Zhang, Yifan Zhang, Yifan Zhang, et al. Qwen-vl: A frontier vision-language model with larger-scale vision pre-training and aligned cross-modal instruction tuning. arXiv preprint arXiv:2310.06726, 2023. 2, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.303, + 0.482, + 0.384 + ], + "angle": 0, + "content": "[4] Yang Bai, Xinxing Xu, Yong Liu, Salman Khan, Fahad Khan, Wangmeng Zuo, Rick Siow Mong Goh, and ChunMei Feng. Sentence-level prompts benefit composed image retrieval. In Proceedings of the International Conference on Learning Representations (ICLR), 2024. Spotlight Presentation. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.388, + 0.482, + 0.443 + ], + "angle": 0, + "content": "[5] Aayush Bansal, Karan Sikka, Gaurav Sharma, and Rama Chellappa. Visual question answering on image sets. In Proceedings of the European Conference on Computer Vision (ECCV), pages 35-51, 2020. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.446, + 0.482, + 0.514 + ], + "angle": 0, + "content": "[6] Florian Bordes, Richard Yuanzhe Pang, Anurag Ajay, Alexander C Li, Adrien Bardes, Suzanne Petryk, Oscar Manas, Zhiqiu Lin, Anas Mahmoud, Bargav Jayaraman, et al. An introduction to vision-language modeling. arXiv preprint arXiv:2405.17247, 2024. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.518, + 0.482, + 0.585 + ], + "angle": 0, + "content": "[7] Yingshan Chang, Mridu Narang, Hisami Suzuki, Guihong Cao, Jianfeng Gao, and Yonatan Bisk. Webqa: Multihop and multimodal qa. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 14178-14188, 2022. 1, 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.589, + 0.482, + 0.671 + ], + "angle": 0, + "content": "[8] Jun Chen, Deyao Zhu, Xiaogian Shen, Xiang Li, Zechun Liu, Pengchuan Zhang, Raghuraman Krishnamoorthi, Vikas Chandra, Yunyang Xiong, and Mohamed Elhoseiny. Minigpt-v2: Large language model as a unified interface for vision-language multi-task learning. arXiv preprint arXiv:2310.15339, 2023. 2, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.675, + 0.482, + 0.77 + ], + "angle": 0, + "content": "[9] Wenhu Chen, Hexiang Hu, Xi Chen, Pat Verga, and William W. Cohen. MuRAG: Multimodal retrieval-augmented generator for open question answering over images and text. In Proceedings of the 2022 Conference on Empirical Methods in Natural Language Processing (EMNLP), pages 5558-5570. Association for Computational Linguistics, 2022. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.774, + 0.482, + 0.87 + ], + "angle": 0, + "content": "[10] Jacob Devlin, Ming-Wei Chang, Kenton Lee, and Kristina Toutanova. Bert: Pre-training of deep bidirectional transformers for language understanding. In Proceedings of the 2019 Conference of the North American Chapter of the Association for Computational Linguistics: Human Language Technologies, pages 4171–4186. Association for Computational Linguistics, 2019. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.873, + 0.482, + 0.901 + ], + "angle": 0, + "content": "[11] Chun-Mei Feng, Yang Bai, Tao Luo, Zhen Li, Salman Khan, Wangmeng Zuo, Xinxing Xu, Rick Siow Mong Goh," + }, + { + "type": "list", + "bbox": [ + 0.094, + 0.115, + 0.484, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.545, + 0.093, + 0.905, + 0.133 + ], + "angle": 0, + "content": "and Yong Liu. Vq4cir: Boosting composed image retrieval with visual question answering. arXiv preprint arXiv:2312.12273, 2023. 2, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.135, + 0.905, + 0.203 + ], + "angle": 0, + "content": "[12] Yash Goyal, Tejas Khot, Douglas Summers-Stay, Dhruv Batra, and Devi Parikh. Making the V in VQA matter: Elevating the role of image understanding in Visual Question Answering. In Conference on Computer Vision and Pattern Recognition (CVPR), 2017. 2, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.205, + 0.905, + 0.272 + ], + "angle": 0, + "content": "[13] Kelvin Guu, Kenton Lee, Zora Tung, Panupong Pasupat, and Ming-Wei Chang. Realm: Retrieval-augmented language model pre-training. In Proceedings of the 37th International Conference on Machine Learning (ICML), pages 3929-3938. PMLR, 2020. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.274, + 0.904, + 0.327 + ], + "angle": 0, + "content": "[14] Edward J. Hu, Yelong Shen, Phillip Wallis, Zeyuan Allen-Zhu, Yuanzhi Li, Shean Wang, Lu Wang, and Weizhu Chen. Lora: Low-rank adaptation of large language models, 2021. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.33, + 0.905, + 0.397 + ], + "angle": 0, + "content": "[15] Drew A Hudson and Christopher D Manning. Gqa: A new dataset for real-world visual reasoning and compositional question answering. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition, pages 6700-6709, 2019. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.4, + 0.905, + 0.481 + ], + "angle": 0, + "content": "[16] Gabriel Ilharco, Mitchell Wortsman, Ross Wightman, Cade Gordon, Nicholas Carlini, Rohan Taori, Achal Dave, Vaishaal Shankar, Hongseok Namkoong, John Miller, Hannaneh Hajishirzi, Ali Farhadi, and Ludwig Schmidt. Openclip. 2021. If you use this software, please cite it as below. 3, 6, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.484, + 0.905, + 0.579 + ], + "angle": 0, + "content": "[17] Vladimir Karpukhin, Barlas Oğuz, Sewon Min, Patrick Lewis, Ledell Wu, Sergey Edunov, Danqi Chen, and Wentau Yih. Dense passage retrieval for open-domain question answering. In Proceedings of the 2020 Conference on Empirical Methods in Natural Language Processing (EMNLP), pages 6769-6781. Association for Computational Linguistics, 2020. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.581, + 0.905, + 0.662 + ], + "angle": 0, + "content": "[18] Andreas Koukounas, Georgios Mastrapas, Michael Günther, Bo Wang, Scott Martens, Isabelle Mohr, Saba Sturua, Mohammad Kalim Akram, Joan Fontanals Martínez, Saahil Ognawala, Susana Guzman, Maximilian Werk, Nan Wang, and Han Xiao. Jina clip: Your clip model is also your text retriever, 2024. 6, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.665, + 0.905, + 0.746 + ], + "angle": 0, + "content": "[19] Ranjay Krishna, Yuke Zhu, Oliver Groth, Justin Johnson, Kenji Hata, Joshua Kravitz, Stephanie Chen, Yannis Kalantidis, Li-Jia Li, David A. Shamma, Michael S. Bernstein, and Li Fei-Fei. Visual genome: Connecting language and vision using crowdsourced dense image annotations. International Journal of Computer Vision, 123(1):32-73, 2017. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.748, + 0.905, + 0.802 + ], + "angle": 0, + "content": "[20] Bo Li, Yuanhan Zhang, Dong Guo, Renrui Zhang, Feng Li, Hao Zhang, Kaichen Zhang, Peiyuan Zhang, Yanwei Li, Ziwei Liu, and Chunyuan Li. Llava-onevision: Easy visual task transfer. arXiv preprint arXiv:2408.03326, 2024. 2, 3, 6, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.804, + 0.905, + 0.87 + ], + "angle": 0, + "content": "[21] Zhuang Liu, Hanzi Mao, Chao-Yuan Wu, Christoph Feichtenhofer, Trevor Darrell, and Saining Xie. A convnet for the 2020s. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 11976-11986, 2022. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.873, + 0.905, + 0.901 + ], + "angle": 0, + "content": "[22] Pan Lu, Tony Xia, Weicheng Shi, Ahmed El Kholy, Xi Victor Lin, Jianfeng Gao, Xiang Chen, and Kai-Wei Chang. Learn" + }, + { + "type": "list", + "bbox": [ + 0.517, + 0.093, + 0.905, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.48, + 0.946, + 0.518, + 0.957 + ], + "angle": 0, + "content": "24825" + } + ], + [ + { + "type": "ref_text", + "bbox": [ + 0.125, + 0.092, + 0.482, + 0.134 + ], + "angle": 0, + "content": "to explain: Multimodal reasoning via thought chains for science question answering. In Advances in Neural Information Processing Systems, 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.135, + 0.483, + 0.205 + ], + "angle": 0, + "content": "[23] Pan Lu, Hritik Bansal, Tony Xia, Jiacheng Liu, Chunyuan Li, Hannaneh Hajishirzi, Hao Cheng, Kai-Wei Chang, Michel Galley, and Jianfeng Gao. Mathvista: Evaluating mathematical reasoning of foundation models in visual contexts. arXiv preprint arXiv:2310.02255, 2023. 2, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.205, + 0.482, + 0.26 + ], + "angle": 0, + "content": "[24] Hongyin Luo, Yung-Sung Chuang, Yuan Gong, Tianhua Zhang, Yoon Kim, Xixin Wu, Danny Fox, Helen Meng, and James Glass. SAIL: Search-augmented instruction learning. arXiv preprint arXiv:2305.15225, 2023. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.261, + 0.482, + 0.329 + ], + "angle": 0, + "content": "[25] Kenneth Marino, Mohammad Rastegari, Ali Farhadi, and Roozbeh Mottaghi. Ok-vqa: A visual question answering benchmark requiring external knowledge. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition, pages 3195-3204, 2019. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.33, + 0.482, + 0.399 + ], + "angle": 0, + "content": "[26] Ahmed Masry, Do Xuan Long, Jia Qing Tan, Shafiq Joty, and Enamul Hoque. Chartqa: A benchmark for question answering about charts with visual and logical reasoning. In Findings of the Association for Computational Linguistics: ACL 2022, pages 2263-2279, 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.4, + 0.482, + 0.468 + ], + "angle": 0, + "content": "[27] Minesh Mathew, Viraj Bagal, Ruben Pérez Tito, Dimosthenis Karatzas, Ernest Valveny, and CV Jawahar. Infographicvqa. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 4390-4399, 2021. 2, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.469, + 0.482, + 0.525 + ], + "angle": 0, + "content": "[28] Minesh Mathew, Dimosthenis Karatzas, and CV Jawahar. Docvqa: A dataset for vqa on document images. In Proceedings of the IEEE/CVF Winter Conference on Applications of Computer Vision, pages 2200-2209, 2021. 2, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.525, + 0.482, + 0.553 + ], + "angle": 0, + "content": "[29] Zach Nussbaum, Brandon Duderstadt, and Andriy Mulyar. Nomic embed vision: Expanding the latent space, 2024. 6, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.553, + 0.482, + 0.594 + ], + "angle": 0, + "content": "[30] OpenAI. Gpt-4o: Enhanced multimodal language model. OpenAI Research, 2024. https://openai.com/index/hello-gpt-4o/.2,3,4,6,7" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.595, + 0.482, + 0.637 + ], + "angle": 0, + "content": "[31] OpenAI. Gpt-4v: Multimodal language model with vision capabilities. OpenAI Research, 2024. https://openai.com/index/gpt-4/.2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.637, + 0.482, + 0.719 + ], + "angle": 0, + "content": "[32] Abhirama Subramanyam Penamakuri, Manish Gupta, Mithun Das Gupta, and Anand Mishra. Answer mining from a pool of images: Towards retrieval-based visual question answering. In Proceedings of the International Joint Conference on Artificial Intelligence (IJCAI), pages 1052-1058, 2023. 1, 2, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.72, + 0.482, + 0.815 + ], + "angle": 0, + "content": "[33] Alec Radford, Jong Wook Kim, Karthik Hallacy, Aditya Ramesh, Gabriel Goh, Sandhini Agarwal, Girish Sastry, Amanda Askell, Pamela Mishkin, Jack Clark, et al. Learning transferable visual models from natural language supervision. In Proceedings of the 38th International Conference on Machine Learning, pages 8748-8763. PMLR, 2021. 3, 6, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.818, + 0.482, + 0.873 + ], + "angle": 0, + "content": "[34] Shaoqing Ren, Kaiming He, Ross Girshick, and Jian Sun. Faster r-cnn: Towards real-time object detection with region proposal networks. Advances in Neural Information Processing Systems, 28, 2015. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.873, + 0.482, + 0.901 + ], + "angle": 0, + "content": "[35] Dustin Schwenk, Apoorv Khandelwal, Christopher Clark, Kenneth Marino, and Roozbeh Mottaghi. A-okvqa: A" + }, + { + "type": "list", + "bbox": [ + 0.093, + 0.092, + 0.483, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.545, + 0.093, + 0.905, + 0.135 + ], + "angle": 0, + "content": "benchmark for visual question answering using world knowledge. In European Conference on Computer Vision (ECCV), pages 1-17, 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.136, + 0.905, + 0.176 + ], + "angle": 0, + "content": "[36] Ray Smith et al. TesseractOCR engine. https://github.com/tesseract-ocr/tesseract, 2024. Accessed: 2024-11-06. 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.179, + 0.905, + 0.247 + ], + "angle": 0, + "content": "[37] Alon Talmor, Sewon Min, Robin Jia, Yanai Elazar, Uriel Singer Hasson, and Danqi Chen. Multimodalqa: Complex question answering over text, tables, and images. In Proceedings of the International Conference on Learning Representations (ICLR), 2021. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.249, + 0.905, + 0.319 + ], + "angle": 0, + "content": "[38] Ryota Tanaka, Kyosuke Nishida, Kosuke Nishida, Taku Hasegawa, Itsumi Saito, and Kuniko Saito. Slidevqa: A dataset for document visual question answering on multiple images. In Proceedings of the AAAI Conference on Artificial Intelligence, pages 13636-13645, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.32, + 0.905, + 0.36 + ], + "angle": 0, + "content": "[39] Ke Wang, Yichi Zhang, and Hongsheng Li. Mini-gemini: An efficient and versatile vision-language model. arXiv preprint arXiv:2310.12345, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.362, + 0.905, + 0.416 + ], + "angle": 0, + "content": "[40] Ke Wang, Junting Pan, Weikang Shi, Zimu Lu, Mingjie Zhan, and Hongsheng Li. Measuring multimodal mathematical reasoning with math-vision dataset. arXiv preprint arXiv:2402.14804, 2024. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.418, + 0.905, + 0.487 + ], + "angle": 0, + "content": "[41] Peng Wang, Shuai Bai, Sinan Tan, Shijie Wang, Zhihao Fan, Jinze Bai, Keqin Chen, Xuejing Liu, Jialin Wang, Wenbin Ge, et al. Qwen2-vl: Enhancing vision-language model's perception of the world at any resolution. arXiv preprint arXiv:2409.12191, 2024. 2, 3, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.489, + 0.905, + 0.545 + ], + "angle": 0, + "content": "[42] Weiyun Wang, Shuibio Zhang, Yiming Ren, Yuchen Duan, Tiantong Li, Shuo Liu, Mengkang Hu, Zhe Chen, Kaipeng Zhang, Lewei Lu, et al. Needle in a multimodal haystack. arXiv preprint arXiv:2406.07230, 2024. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.546, + 0.905, + 0.613 + ], + "angle": 0, + "content": "[43] Tsung-Han Wu, Giscard Biamby, Jerome Quenum, Ritwik Gupta, Joseph E. Gonzalez, Trevor Darrell, and David M. Chan. Visual haystacks: A vision-centric needle-in-a-haystack benchmark. arXiv preprint arXiv:2407.13766, 2024. 2, 3, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.616, + 0.905, + 0.685 + ], + "angle": 0, + "content": "[44] Xiang Yue, Yuansheng Ni, Kai Zhang, Tianyu Zheng, Ruoqi Liu, Ge Zhang, Samuel Stevens, Dongfu Jiang, Weiming Ren, Yuxuan Sun, et al. Mmmu: A massive multi-discipline multimodal understanding and reasoning benchmark for expert agi. arXiv preprint arXiv:2311.16502, 2024. 2, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.687, + 0.905, + 0.755 + ], + "angle": 0, + "content": "[45] Xiaohua Zhai, Basil Mustafa, Alexander Kolesnikov, and Lucas Beyer. Sigmoid loss for language image pre-training. In Proceedings of the IEEE/CVF International Conference on Computer Vision (ICCV), pages 11975-11985, 2023. 3, 6, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.757, + 0.905, + 0.826 + ], + "angle": 0, + "content": "[46] Renrui Zhang, Dongzhi Jiang, Yichi Zhang, Haokun Lin, Ziyu Guo, Pengshuo Qiu, Aojun Zhou, Pan Lu, Kai-Wei Chang, Peng Gao, and Hongsheng Li. Mathverse: Does your multi-modal llm truly see the diagrams in visual math problems? arXiv preprint arXiv:2403.14624, 2024. 2, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.828, + 0.905, + 0.896 + ], + "angle": 0, + "content": "[47] Deyao Zhu, Jun Chen, Xiaogian Shen, Xiang Li, and Mohamed Elhoseiny. Minigpt-4: Enhancing vision-language understanding with advanced large language models. In International Conference on Learning Representations (ICLR), 2024. 2, 3" + }, + { + "type": "list", + "bbox": [ + 0.517, + 0.093, + 0.905, + 0.896 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.48, + 0.946, + 0.52, + 0.957 + ], + "angle": 0, + "content": "24826" + } + ] +] \ No newline at end of file diff --git a/2025/Document Haystacks_ Vision-Language Reasoning Over Piles of 1000+ Documents/1b0623c9-cc0a-4950-bcf6-716e67ed00b3_origin.pdf b/2025/Document Haystacks_ Vision-Language Reasoning Over Piles of 1000+ Documents/1b0623c9-cc0a-4950-bcf6-716e67ed00b3_origin.pdf new file mode 100644 index 0000000000000000000000000000000000000000..94d72a9f986ffbb1db2a08c516c3fb845c74904e --- /dev/null +++ b/2025/Document Haystacks_ Vision-Language Reasoning Over Piles of 1000+ Documents/1b0623c9-cc0a-4950-bcf6-716e67ed00b3_origin.pdf @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:a8cc8a9244098a28c23a16c6c9c2310fdef9c0daf89bbcd54065801f3712afa0 +size 1135487 diff --git a/2025/Document Haystacks_ Vision-Language Reasoning Over Piles of 1000+ Documents/full.md b/2025/Document Haystacks_ Vision-Language Reasoning Over Piles of 1000+ Documents/full.md new file mode 100644 index 0000000000000000000000000000000000000000..f7fc890a25a4212d5035cb484fce7949a6c774b4 --- /dev/null +++ b/2025/Document Haystacks_ Vision-Language Reasoning Over Piles of 1000+ Documents/full.md @@ -0,0 +1,248 @@ +# Document Haystacks: Vision-Language Reasoning Over Piles of 1000+ Documents + +Jun Chen $^{1}$ , Dannong Xu $^{2}$ , Junjie Fei $^{1}$ , Chun-Mei Feng $^{3}$ , Mohamed Elhoseiny $^{1}$ $^{1}$ King Abdullah University of Science and Technology + $^{2}$ The University of Sydney, $^{3}$ IHPC, A*STAR +{jun.chen, junjie. fei, mohamed.elhoseiny}@kaust.edu.sa +daxu8019@uni.sydney.edu.au, fengcm.ai@gmail.com + +![](images/84598932cd96043268c6467ed2ade6f7cbfc2c116b10251727075dd11b974622.jpg) +(a) Previous Benchmarks: Each question paired with a limited image set + +![](images/20a0a7b3bd1b59437ec0c97bc20b4582e5abfc47b8608e808953356a12e6ee2d.jpg) +(b) Our Benchmarks: All questions mapped to an extensive document collection +Figure 1. Comparison between previous and proposed benchmarks. Given a question as input, all benchmarks aim to retrieve relevant images from an image pool to correctly answer the question. Unlike prior benchmarks like RetVQA [32] and WebVQA [7], which structure their datasets by pairing each question with a limited set of images (typically $\leq 30$ ), our benchmarks, DocHaystack and InfoHaystack, map each question to a substantially larger document collection, scaling up to 1,000 visual documents. This expanded scope more accurately represents large-scale document retrieval scenarios and offers a greater challenge in retrieval accuracy and visual question answering. + +# Abstract + +Large multimodal models (LMMs) have achieved impressive progress in vision-language understanding, yet they face limitations in real-world applications requiring complex reasoning over a large number of images. Existing benchmarks for multi-image question-answering are limited in scope, each question is paired with only up to 30 images, which does not fully capture the demands of large-scale retrieval tasks encountered in the real-world usages. To reduce these gaps, we introduce two document haystack benchmarks, dubbed DocHaystack and InfoHaystack, designed to evaluate LMM performance on large-scale visual document retrieval and understanding. Addition + +ally, we propose V-RAG, a novel, vision-centric retrieval-augmented generation (RAG) framework that leverages a suite of multimodal vision encoders, each optimized for specific strengths, and a dedicated question-document relevance module. V-RAG sets a new standard, with a $9\%$ and $11\%$ improvement in Recall@1 on the challenging DocHaystack-1000 and InfoHaystack-1000 benchmarks, respectively, compared to the previous best baseline models. Additionally, integrating V-RAG with LMMs enables them to efficiently operate across thousands of images, yielding significant improvements on our DocHaystack and InfoHaystack benchmarks. Our code and datasets are available at https://github.com/Vision-CAIR/dochaystacks + +# 1. Introduction + +Large Multimodal Models (LMMs) [1, 20, 30, 41] have made remarkable progress in the vision-language understanding. However, these models still face challenges when tasked with reasoning over extensive collections of images or documents [43], limiting their effectiveness in real-world applications, such as visual search or querying over large sets of images or documents, like those stored on personal devices or in photo albums. However, there lacks such proper benchmarks to evaluate these capabilities. To address this gap, we introduce the DocHaystack and InfoHaystack benchmarks, designed to evaluate LMMs on large-scale image retrieval and understanding capabilities, pushing the boundaries of LMM performance in complex, real-world scenarios. + +The existing multi-image retrieving and reasoning benchmarks are primarily constructed on a small scale, as highlighted in works such as [32, 37]. Each question in these benchmarks is paired with only up to 30 images as illustrated in Figure 1 (a). However, this limited scope does not align well with real-world scenarios, which often require retrieval and reasoning across hundreds or thousands of images or documents. In contrast, we introduce the Document Haystacks benchmarks, depicted in Figure 1 (b), which allow for querying questions from a large-scale collection of up to 1,000 documents, necessitating that models retrieve and reason over an extensive set of documents for each question. This scale better simulates practical applications and their demands. + +The main challenge for constructing our document haystack benchmarks is collecting specific questions while ensuring there are no ambiguous answers across a large set of images. Existing datasets, such as those in DocVQA and InfographicVQA [27, 28], contain numerous "general" questions, like "What is the table number?", where answers could be derived from multiple images, leading to nonunique responses. To address this, we implemented a rigorous data filtering pipeline. First, we employed both a large language model (LLM) and human annotators to systematically filter out "general" questions based on carefully defined criteria. Additionally, we used the LLM to exclude questions relying on generic knowledge, such as "What is the capital of Missouri?", which can be answered without image context. This approach ensures that the questions in the benchmark can only be answered through specific visual cues from the provided images, maintaining the benchmark's integrity for evaluating image-based understanding. + +To enable the current LMMs effectively reason over a large number of images, we propose a vision-centric retrieval-augmented generation (RAG) framework, named V-RAG. V-RAG combines multiple multimodal vision encoders, leveraging each encoder's unique strengths to enhance retrieval accuracy. Additionally, it incorporates an + +LMM-filter module to assess the relevance of each document to the query, refining the retrieval process by ensuring that only relevant documents are prioritized. This integrated approach allows V-RAG to navigate extensive document collections efficiently. Experimental results demonstrate that V-RAG achieves $9\%$ and $11\%$ improvement in Recall@1 on the DocHaystack-1000 and InfoHaystack-1000 compared to previous best text-to-image retrieval methods. Additionally, we found that integrating V-RAG brings GPT4o over a $55\%$ acc improvement on DocHaystack-200 and a $34\%$ acc improvement on InfoHaystack-200, indicating the effectiveness of our V-RAG. + +Our contributions are as follows: + +- We introduce the Document Haystack benchmarks, including DocHaystack-100/200/1000 and InfoHaystack-100/200/1000, with the most challenging setup consisting of 1,000 documents for each inquiry. These benchmarks advance document retrieval and reasoning tasks by requiring models to navigate and reason across extensive document collections, surpassing prior benchmarks limited to smaller retrieval tasks. +- We propose a vision-centric retrieval-augmented generation framework, V-RAG, which enhances the retrieval capabilities of LMMs. V-RAG achieves substantial improvements over previous best text-to-image retrieval methods by $9\%$ and $11\%$ on DocHaystack-1000 and InfoHaystack-1000, respectively. + +# 2. Related Works + +VQA benchmarks. VQA play a critical role in assessing a model's ability to understand and reason across visual contexts [6, 11, 12]. Traditional VQA datasets typically measure a model's comprehension of object attributes [15, 19], spatial relationships [15], as well as its understanding of documents [27, 28], charts [26], mathematics [23, 40, 46], and open knowledges [25, 35]. Additionally, these benchmarks explore models' knowledge across varied fields, including science and the arts [22, 44]. This broad array of benchmarks has greatly advanced vision-language models by cultivating diverse visual comprehension skills, particularly for modern foundation models in vision-language understanding [1, 3, 8, 20, 22, 30, 31, 39, 47]. Notably, these benchmarks have primarily focused on question answering within single image or document. In contrast, our benchmark shifts the focus towards retrieval and comprehensive understanding across a large collection of visual documents, presenting new challenges and expanding the scope of visual question answering. + +Several previous efforts have tackled the challenge of visual question answering and reasoning across multiple images [5, 7, 32, 37, 38, 42]. For instance, datasets such as MultimodalQA [37] and ISVQA [5] require models to have multi-image reasoning abilities. Meanwhile, WebQA [7] + +![](images/44a1374ccf0762cb7e5d4774343878b7d47e5654ddb937ab9ce7944c30bed756.jpg) +Figure 2. Data Curation Pipeline. Our benchmarks are curated based on the DocVQA and InfographicVQA datasets, following a three-step filtering process to obtain document-specific question-answer pairs. In Step 1, we filter out general questions (e.g., "What is the table number?"), as these could be answered by multiple documents and lack specificity. Step 2 involves a manual review by human annotators to further remove general questions. In Step 3, we eliminate generic-knowledge questions (e.g., "How many sports were in the 2008 Beijing Paralympic Games?") that can be answered directly by large language models without requiring image input. + +and RetVQA [32] involve an additional step where models must first retrieve relevant images from a limited image pool before answering visual questions based on these results. However, these benchmarks are generally constrained to relatively small image pools, where each question is paired with an image set containing up to 30 images. In contrast, our proposed benchmarks, DocHaystack and InfoHaystack, significantly expand this scope by requiring models to retrieve and reason from a much larger set of up to 1,000 documents, presenting a notably greater challenge in retrieval and multi-image reasoning. + +Large multimodal models (LMMs). LMMs have achieved substantial advancements in understanding and reasoning across single or multiple images [1, 8, 20, 30, 41, 47]. These models have significantly enhanced vision-language understanding across numerous dimensions and applications [12, 23, 44, 46]. LMMs benefit primarily from large-scale image-text alignment and extensive language modeling, which emerge them with advanced understanding and reasoning abilities. However, despite these breakthroughs, LMMs still encounter challenges when handling large-scale image or document sets [43]. This difficulty is due to the inherent complexity of processing such complex data. To address this, retrieval-based approaches have been developed to extend the capacity of vision-language models, augmenting their ability to process and reason over a larger number of images. + +Retrieval-augmented generation (RAG). RAG integrates retrieval systems [4, 11, 16, 33, 45], with generative models, enhancing them with additional knowledge. While RAG has been extensively explored in language domains [2, 13, 17, 24], its application in vision-language contexts is also advancing. In vision-language RAG, models like MuRAG [9] leverage image-text memory to retrieve top-k neighbors by comparing inner-product similarities. RetVQA [32] uses an image-question relevance encoder, combining BERT [10] and Faster R-CNN [34] to filter relevant images, while MIRAGE [43] employs a CLIP-based + +encoder [33] to train a retriever. These frameworks extend model capabilities, enabling retrieval and reasoning across hundreds or thousands of images. In contrast, we propose V-RAG, a vision-centric RAG framework that integrates multiple vision encoders to more effectively capture image features, and introduces a LMM-based question-document relevance comparison module. Our results demonstrate that V-RAG surpasses existing methods on our DocHaystack and InfoHaystack benchmarks, setting a new standard for large-scale visual retrieval and reasoning. + +# 3. DocHaystack and InfoHaystack Benchmarks + +To support effective retrieval and reasoning across extensive document collections, we present two new benchmarks—DocHaystack and InfoHaystack—designed to ensure each question yields a unique, document-specific answer. Derived from DocVQA [28] and InfographicVQA [27], these benchmarks address the challenge of answer ambiguity by selectively curating questions that can only be answered by a single document within a large dataset. + +Benchmark construction pipeline. There exists many general questions in the existing benchmarks and lead to multiple answers for different document context. For example, general questions like "What is the table number?" may apply to various documents and yield multiple valid answers, while a targeted question like "Who is the reviewer for the article titled 'An antithyroid factor in milk'?" is likely to produce a unique answer, as only a single document or a limited set of documents would contain that information. Therefore, our benchmark construction follows a structured three-step filtering pipeline, illustrated in Figure 2, to ensure high-quality, unique-answer questions. First, we employ a large language model (LLM) to filter out general questions that could generate multiple answers across documents. Next, a manual review step further + +![](images/f521efde925d2b87455540c536284a49453fb485b9b8a3388f1312a8a9bf9242.jpg) +Figure 3. The V-RAG pipeline workflow. In the top section, a vision encoder ensemble is used, combining multiple vision models—CLIP, SigLIP, and OpenCLIP—to process a large document haystack. Each encoder computes similarity scores, which are averaged into $Sim_{avg}$ . The top m documents, based on these scores, are selected for further analysis. In the bottom right, the LMM-Filter Module utilizes a pretrained LMM to assess whether each selected document can potentially answer the posed question. This filtering step removes documents that do not match, retaining only relevant ones. Finally, the top k most relevant images are input into the LMM along with the original question $q$ to generate a specific answer. + +checks the questions to ensure the data quality. Finally, a generic-knowledge filtering stage refines the dataset further, retaining only questions closely tied to specific document content. + +This carefully designed pipeline, combining LLM-based filtering and human review, effectively curates questions that drive accurate, document-specific retrieval. By focusing on reducing answer ambiguity, DocHaystack and InfoHaystack enhance the precision of retrieval and reasoning in large-scale document processing tasks, providing a valuable tool for the evaluation of retrieval systems. We discuss this data curation pipeline in details as follows: + +General-question LLM filtering. We begin by using the LLM, GPT-4o [30], to filter out general questions through a set of well-crafted instructions. Leveraging the LLM's strong contextual understanding, this initial filtering step allows us to efficiently process large volumes of data, identifying broad or ambiguous questions that may yield multiple answers across documents. This automated approach significantly enhances the benchmark construction's efficiency and quality. + +To guide the LLM, we first define the task, providing clear distinctions between general and specific questions along with illustrative examples. With this framework, the LLM can then assess each question and determine if it is general or specific. The instructional format is as follows: + +You are an evaluator tasked with identifying if a ques + +tion is specific or general. A general question seeks commonly known or widely applicable information without unique identifiers, e.g., "Who is the person standing in the ground?" A specific question, however, requests unique information about a particular individual, event, or object, e.g., "What is the Social Security Number of Charles Yarbrough?" Based on these definitions, determine if the following question is general or specific: {question}. + +General-question manual review. After the initial LLM filtering, we conduct a manual review of the questions that were classified as specific. This manual process involves two key steps to ensure answer uniqueness and benchmark quality. + +In the first step, we examine each question to confirm it contains unique identifiers—such as names, dates, titles, or other specific attributes—suggesting a document-specific answer. This careful check helps identify questions with clear, unique markers that direct the retrieval process to a single document. + +In the second step, we verify the uniqueness of each answer to eliminate any remaining ambiguity. Although specific identifiers are present, questions may still be prone to ambiguity, such as with common names or recurring book titles. To address this, we employ a refined verification process. First, we use Optical Character Recognition (OCR) [36] to extract all text from images in the dataset. We then search for occurrences of the unique identifiers re + +
GPT-4oLLaVA-OneVisionQwen2-VL
DocVQA27.0%4.7%3.4%
InfographicVQA56.4%13.4%11.3%
+ +Table 1. Percentage of questions answerable by LMMs without vision input. We evaluate GPT-4o, LLaVA-Onevision, and Qwen2-VL on their ability to answer questions directly from our dataset without requiring vision input. The reported percentage reflects the proportion of examples that can be answered solely through language understanding. + +tained from the first step across other documents. If matches are found, a manual review is conducted to ensure no alternative valid answers exist. This comprehensive approach minimizes the possibility of a single question mapping to multiple answers, enhancing the precision and reliability of our benchmarks. + +Generic-knowledge filtering. In DocVQA and InfographicVQA tasks, certain questions—such as “How many sports were in the 2008 Beijing + +Paralympic Games?"—can be answered based on general knowledge accessible to a large language model, without relying on the image content. This introduces a language bias when using LMMs for visual question answering, as it shifts the focus away from image-based reasoning. To address this, we filter out these general-knowledge questions, ensuring that evaluation emphasizes vision-based understanding and that models rely primarily on visual content to generate accurate answers. + +To implement this, we developed an LLM-based evaluation pipeline that detects and excludes such questions. For each question, we prompt an LLM with “{question}, answer briefly.” After receiving a response, we compare it to the ground-truth answer using another LLM. If the response matches the ground truth, we classify the question as general knowledge-related and remove it, thereby isolating questions that truly require visual document understanding. As shown in Table 1, GPT-4o accurately answers $27.0\%$ of DocVQA questions and $56.4\%$ of InfographicVQA questions directly, a rate significantly higher than that of open-source LLMs. Therefore, we select GPT-4o to filter out the questions that can be directly answered by the GPT-4o model. Overall, this process is to ensure that the evaluation reflects the necessity of vision-based comprehension. + +Final dataset profile. After a rigorous three-stage data filtering process, we retained 109 questions from DocVQA and 155 questions from InfographicVQA, associated with 59 and 66 documents that provide the evidence, respectively. To assess retrieval performance at scale, we introduce two benchmarks: DocHaystack-1000 and InfoHaystack-1000, where each question requires retrieving relevant content from a set of 1,000 documents. Given + +![](images/f8264d51318bf845052e4510597aaf53449d962463b65f1bcef0c4390c546d80.jpg) +Figure 4. Question type analysis. We analyze the distribution of question types of DocHaystack and InfoHaystack. Each benchmark categorizes the data into 5 different types. + +![](images/bcf44f1ff8cc585290afca51d4c3ff1172c7e6440de95546fb4b832627797fe4.jpg) + +the challenge this scale presents to current LMMs, particularly in terms of context length limitations, we also construct two smaller benchmarks: DocHaystack-100/200 and InfoHaystack-100/200. These benchmarks allow direct input of all associated images into the context, enabling evaluation of models' long-context comprehension ability. For training set, we also construct a dataset comprising 2,835 questions similarly, with 899 from DocVQA and 1,936 from InfographicVQA, to support robust learning and generalization for the multi-image reasoning tasks. + +Question type analysis. The types of questions represent the types of the evidence required for accurate answers. In Figure 4, we illustrate the distribution of question types across our dataset to provide insights into its structure. Following the classification system used in DocVQA and InfographicVQA, we categorize questions accordingly (note that a single question may fall into multiple categories). As shown in the figure, the DocHaystack benchmark places a greater emphasis on Table/List and Layout understanding, whereas InfoHaystack primarily targets Figure, Text, and Table comprehension. + +# 4. Method + +Current large multimodal models (LMMs) face substantial challenges when reasoning across hundreds or thousands of images, due not only to context length limitations but also to the inherent complexity of the task. This issue is particularly pronounced in our benchmarks, which contain 1k document files requiring high-resolution images to capture and interpret small-font text effectively. To enable LMMs to perform reasoning over a substantial number of documents, we introduce a vision-centric retrieval-augmented generation (V-RAG) framework. V-RAG efficiently retrieves a reduced set of relevant documents, allowing the LMM to focus on a manageable subset for deeper understanding, as illustrated in Figure 3. In the following section, we provide a detailed description of the V-RAG pipeline. + +Task definition. Given a question $q$ and a collection of $N$ documents $\mathcal{D} = \{D_1, \ldots, D_N\}$ , the V-RAG framework + +aims to retrieve the top- $k$ most relevant documents to support LMMs understanding and answering the question $q$ . V-RAG accomplishes this through a two-step retrieval process designed to effectively identify and rank relevant documents for each question. + +Vision encoder ensemble. Document files often contain a mix of text, symbols, and visual elements across various scales, requiring vision encoders to capture a comprehensive understanding of these complex structures. To efficiently handle this diversity, we represent each document as an image and utilize an ensemble of vision encoders, including CLIP [33], SigLIP [45], and OpenCLIP [16], each bringing distinct strengths to the image understanding, as depicted in Figure 3. For example, the ConvNext encoder [21] from OpenCLIP [16] is particularly effective for high-resolution image encoding. We compute the similarity score between each question $q$ and all documents in the document set $\mathcal{D}$ according to Equation 1, with similarity scores from each encoder represented as $Sim_{c}$ , $Sim_{o}$ , and $Sim_{s}$ respectively. + +$$ +\mathcal {S} (q, \mathcal {D}) = \cos \left(\phi_ {\mathrm {t}} (q), \phi_ {\mathrm {v}} (D _ {j})\right) \mid D _ {j} \in \mathcal {D}, \tag {1} +$$ + +where $S$ denotes the computing the similarity between the query $q$ and a collection of documents $\mathcal{D}$ . $\cos$ denotes the cosine similarity. $\phi_{\mathrm{t}}$ denotes the text encoder, and $\phi_{\mathrm{v}}$ denotes the vision encoder. + +To derive a final relevance score, we calculate the average similarity $Sim_{avg}$ for each question-image pair by combining $Sim_c$ , $Sim_o$ , and $Sim_s$ . We then rank the images based on $Sim_{avg}$ in descending order, selecting the top- $m$ most relevant images according to their similarity scores. + +LMM-filter module. To refine the selection of top- $m$ relevant images further, we introduce a LMM-based question-image relevance assessment module. This module evaluates the relevance between each question and the top- $m$ images identified in the first filtering step. Specifically, we pair each image with the question text and input them into an open-source vision-language model, prompting, "Can this image provide answers to this question? Respond only with yes or no". We only retain the question-image pairs that are identified as "yes" from LMM, and remove other irrelevant images. + +LMM-VQA module. Achieving high top-1 ranking accuracy in image retrieval is challenging, so we retain the top- $k$ images from the LMM-filtered ranking list and present them to the LMM-VQA to improve the likelihood of including relevant images. We input these top- $k$ images alongside the question into the LMM-VQA (see Figure 3), which then generates the answer directly. To enhance robustness against visual distractors, the LMM-VQA can be further optimized, as analyzed in the experiment section. + +# 5. Experiments + +In the experiments section, we will primarily describe our training setup, covering evaluation metrics, baseline models, and the fine-tuning procedure for the LMM-VQA model. We also present the main experimental results along with an ablation study to provide further insights. + +# 5.1. Training setup + +Metric. In our evaluation of the DocHaystack and Info-Haystack benchmarks, we employ a model-based assessment by leveraging GPT-4o-mini [30] to accurately determine whether the model predictions match target answers. This method uses a carefully structured prompt to facilitate GPT-4o-mini's evaluation of answer correctness. We empirically found that the model-based evaluation achieves higher consistency and alignment with human judgment. Additional details on the prompt design are provided in the Appendix. + +For the document retrieval evaluation, we report the baseline results using Recall@1, Recall@3, and Recall@5 metrics. These metrics enable a thorough assessment of retrieval accuracy across varying levels of precision. + +Baselines. In our experiment, we have evaluated several open and closed-sourced vision-language models on the retrieval and VQA performance. For the large multimodal model, we used the gpt-4o-2024-08-06 version of GPT-4o [30], the LLaVA-OneVision-Qwen2-7b-OV-HF version of LLaVA-OneVision [20], and the Qwen2-VL-7B-Instruct version of Qwen2-VL [3]. For computing the text-to-image similarities, we employed the Jina-CLIP-v1 [18] variant, Nomic-Embed-Vision-v1.5 [29] variant, CLIP [33] ViT-L/14@336 variant, for SigLIP [45], the ViT-SO400M/14@384 variant, and for OpenCLIP [16], the ConvNeXt-XXL@1024 variant as well as text-based method, BM25. + +In our V-RAG setting, we apply LLaVA-OneVision-Qwen2-7b-OV-HF for the LMM-filter module and Qwen2-VL-7B-Instruct for the LMM-VQA module. We select $m$ as 60 and $k$ as 5 in our experiment. + +Optimizing the LMM-VQA module. To improve the robustness of the LMM-VQA model in handling visual question answering with multiple distractor images, we further fine-tune the model using our curated training data. + +During this fine-tuning process, we introduce 1-10 randomly sampled distractor images for each question, creating a challenging setting that encourages the model to focus on relevant content amid a mix of positive and negative images. The fine-tuning is conducted with a batch size of 32 and a peak learning rate of 1e-4 over a single epoch. Additionally, we leverage LoRA [14] with a rank of 8 to efficiently adapt the model's parameters during training. + +
DocHaystack-100DocHaystack-200DocHaystack-1000
R@1R@3R@5R@1R@3R@5R@1R@3R@5
BM25 (OCR)63.375.279.865.171.675.256.966.169.7
Jina-CLIP [18]16.531.241.39.224.830.33.77.312.8
Nomic-Embed-Vision [29]16.524.828.413.821.125.71.82.86.4
CLIP [33]46.865.169.744.058.765.123.941.345.9
SigLIP [45]51.467.976.247.763.370.633.049.557.8
OpenCLIP [16]58.775.279.856.970.675.234.949.557.8
V-RAG (ours)81.789.089.078.084.484.466.178.078.9
InfoHaystack-100InfoHaystack-200InfoHaystack-1000
R@1R@3R@5R@1R@3R@5R@1R@3R@5
BM25 (OCR)56.865.871.051.665.269.038.751.658.1
Jina-CLIP43.251.658.136.846.551.623.933.637.4
Nomic-Embed-Vision34.850.356.831.043.248.420.731.035.5
CLIP69.778.785.865.277.481.945.864.570.3
SigLIP58.171.680.055.567.776.839.455.561.9
OpenCLIP72.385.292.966.581.989.053.665.872.9
V-RAG (ours)79.491.092.974.888.488.464.574.278.1
+ +Table 2. Retrieval Results. We compare our V-RAG model with other text-to-image and text-to-text (using OCR) retrieval methods across both benchmarks. V-RAG consistently outperforms baseline models on Recall@1, Recall@3, and Recall@5 metrics. Notably, V-RAG leverages an ensemble of text-to-image models along with a large multimodal model in a two-stage filtering approach. Top-performing values in each column are highlighted in bold. + +
ModelDocHaystackInfoHaystack
10020010001002001000
LLaVA-OV [20]------
GPT-4o [30]27.523.9-23.920.0-
Gemini [1]50.548.6-29.021.9-
Qwen2-VL [41]41.312.8-20.014.2-
MIRAGE [43]3.73.72.87.77.16.5
LLaVA-OV+V-RAG69.765.155.143.241.936.8
GPT-4o+V-RAG81.772.567.065.263.256.8
Gemini+V-RAG73.465.158.757.457.447.1
Qwen2-VL+V-RAG82.674.366.165.865.860.0
Qwen2-VL-f.t.+V-RAG86.279.873.467.167.760.0
+ +Table 3. The VQA results for the DocHaystack and Info-Haystack. We evaluate with many closed-source and open-source multimodal model, and also integrating them with our V-RAG retrieval framework. - denotes that those models can not be inferred due to their token context constraints. To enable GPT-4o and Qwen2-VL to process hundreds of images, we employ low-resolution mode and adjust image size for compatibility. + +# 5.2. Main Experimental Results + +We evaluated a range of open-source and closed-source vision-language models for VQA tasks. We also evaluate several text-to-image and text-to-text (with OCR) retrieval models to evaluate their retrieval capabilities on our benchmarks. More detailed performance analysis are described in the following sections. + +Retrieving results. The retrieval results in Table 2 demonstrate the superiority of our proposed V-RAG framework over several baseline methods across both DocHaystack and InfoHaystack benchmarks. V-RAG consistently achieves the highest Recall@1, Recall@3, and Recall@5 scores on most categories, indicating its robust retrieval capabilities. Notably, V-RAG outperforms text-based retrieving models such as BM25 and also the text-to-image retrieval models + +like jina-clip, CLIP, SigLIP, and OpenCLIP by substantial margins, especially on the DocHaystack-100 subset, where it reaches Recall@1 of $81.7\%$ and Recall@5 of $89.0\%$ . This pattern continues for larger datasets (DocHaystack-1000), where V-RAG remains competitive, achieving Recall@1 of $66.1\%$ . It achieves the top performance across all recall metrics on DocHaystack. For InfoHaystack benchmarks, V-RAG also outperforms other models, particularly on InfoHaystack-100 and InfoHaystack-200, where it receives Recall@1 of $79.4\%$ and $64.5\%$ , higher than previous best by around $7\%$ and $11\%$ , respectively. This consistent performance advantage highlights the effectiveness of V-RAG's ensemble of multiple vision encoders, allowing it to capture more granular details and improve retrieval accuracy over large multimodal models. + +Visual question answering (VQA) results. The table presents VQA results for the DocHaystack and Info-Haystack benchmarks across varying dataset sizes (100, 200, 1000) using different multimodal models, both independently and in combination with the V-RAG framework. The results show that Qwen2-VL fine-tuned with V-RAG (Qwen2-VL-f.t.+V-RAG) achieves the highest scores across most benchmarks, with particularly notable performance on DocHaystack-100 (86.2) and InfoHaystack-100 (67.1), indicating superior retrieval and VQA capabilities in these scenarios. When V-RAG is added to other models, substantial improvements are observed, demonstrating the framework's efficacy in enhancing retrieval accuracy. For instance, GPT-4o's performance increases significantly with V-RAG, particularly for DocHaystack-100 and -200. The analysis highlights that V-RAG integration generally boosts performance across models, with Qwen2-VL-f.t.+V-RAG standing out as the top performer on both + +![](images/1b5777be2eeaa49dfad5bd1c1f8a694dc890c097efba4cb37865a0cb29e32481.jpg) +Figure 5. Top-k selection ablation analysis for LMM-VQA. We demonstrate the results for LLaVA, Qwen2-VL, GPT-4o and also the finetuned Qwen2-VL model on the DocHaystack-100/1000 and InfoHaystack-100/1000 benchmarks. All the models are integrated with our V-RAG framework. We show the VQA accuracy performance for each ablation. + +![](images/e78914bb84f03c5adaa7cdab5f6063249e26ced941af318bf49a1fd6a8b8ec39.jpg) + +![](images/86d3e9efa600a0d7ed27dbf45fec14a74066c047072ff13e29f06128ffd0f57b.jpg) + +![](images/ccb77c5c7273b24a04c7501c4b31159868cfa7fb1c21e67e4e3b06d7f6ff10f6.jpg) + +
CLIPSigLIPOpenCLIPLMM-filterDocHaystack-1000InfoHaystack-1000
R@1R@3R@5R@1R@3R@5
✓XXX23.941.345.945.864.570.3
X✓XX33.049.557.839.455.561.9
XX✓X34.949.557.853.665.872.9
✓✓XX40.459.662.459.467.774.2
✓✓✓X42.266.177.556.171.078.1
✓✓✓✓66.178.078.964.574.278.1
+ +Table 4. Ablation study on the V-RAG framework components. We quantify the impact of each module for the Recall@1, Recall@3 and Recall@5 retrieval performance on the DocHaystack-1000 and InfoHaystack-1000 for our V-RAG framework. + +benchmarks, especially for the larger 1000-document tasks where retrieval accuracy is more challenging. This suggests that V-RAG's vision-centric, retrieval-augmented approach is highly effective for large-scale multimodal document understanding. + +The table also shows that the DocHaystack-1000 and InfoHaystack-1000 present significant challenges for current LMMs. The drop in performance for larger document sets, with top accuracy only reaching $73.4\%$ for DocHaystack-1000 and $60.0\%$ for InfoHaystack-1000, underscores the difficulty our benchmarks. + +# 5.3. Ablation Studies + +Ablation study on Top-k Selection. This figure presents the top-k selection ablation analysis for LMM-VQA across four models: LLaVA-OV, Qwen2-VL, GPT-4o, and the fine-tuned Qwen2-VL (Qwen2-VL-f.t.), evaluated on the DocHaystack-100/1000 and InfoHaystack-100/1000 benchmarks. The analysis reports VQA accuracy as a function of top-k selection (Top 1, Top 3, and Top 5). Overall, accuracy tends to improve with larger k-values, suggesting that offering more retrieval options positively impacts model performance. However, for LLaVA-OV, there is a marked decrease in performance at top-5, indicating that this model struggles to process multiple images at this scale. + +Ablation study on the V-RAG framework components. The ablation study in Table 4 highlights the contributions of each component in the V-RAG framework on the DocHaystack-1000 and InfoHaystack-1000 benchmarks. Using CLIP alone yields low performance (e.g., Recall@1 of $23.9\%$ on DocHaystack-1000 and $45.8\%$ on + +InfoHaystack-1000), indicating its limited retrieval capability on its own. Adding SigLIP and OpenCLIP incrementally improves results. + +The highest performance is achieved when all three encoders are combined with the VLM-filter module, leading to Recall@1 scores of $66.1\%$ on DocHaystack-1000 and $64.5\%$ on InfoHaystack-1000. This setup also achieves the top Recall@1, Recall@3 and Recall@5 values, demonstrating that the VLM-filter is essential for refining the ensemble outputs and significantly improving retrieval accuracy. These results confirm that each module contributes to VRAG's overall effectiveness. + +# 6. Conclusion + +In this work, we introduced the DocHaystack and Info-Haystack benchmarks to evaluate LMMs for retrieving and reasoning across large-scale documents. Our benchmarks providing a more rigorous and realistic assessment of large multimodal models in real-world, large-scale retrieval scenarios. To tackle these challenges, we proposed V-RAG, a vision-centric retrieval-augmented generation framework that significantly enhances retrieval precision and overall VQA performance. V-RAG achieves this through an ensemble of vision encoders and a specialized relevance filtering module, enabling improved accuracy across diverse visual inputs. Experimental results indicate that integrating V-RAG enables both open-source and closed-source LMMs to achieve superior performance in large-scale image retrieval and complex reasoning tasks. + +# References + +[1] Google AI. Gemini: Google's multimodal ai model. Google AI Research, 2024. https://fireflies.ai/blog/gemini-vs-gpt-4.2,3,7 +[2] Akari Asai, Zeqiu Wu, Yizhong Wang, Avirup Sil, and Hannaneh Hajishirzi. Self-rag: Learning to retrieve, generate, and critique through self-reflection. arXiv preprint arXiv:2310.11511, 2023. 3 +[3] Yifan Bai, Zhen Zhang, Yifan Zhang, Yuxuan Li, Yifan Zhang, Yifan Zhang, Yifan Zhang, Yifan Zhang, Yifan Zhang, Yifan Zhang, et al. Qwen-vl: A frontier vision-language model with larger-scale vision pre-training and aligned cross-modal instruction tuning. arXiv preprint arXiv:2310.06726, 2023. 2, 6 +[4] Yang Bai, Xinxing Xu, Yong Liu, Salman Khan, Fahad Khan, Wangmeng Zuo, Rick Siow Mong Goh, and ChunMei Feng. Sentence-level prompts benefit composed image retrieval. In Proceedings of the International Conference on Learning Representations (ICLR), 2024. Spotlight Presentation. 3 +[5] Aayush Bansal, Karan Sikka, Gaurav Sharma, and Rama Chellappa. Visual question answering on image sets. In Proceedings of the European Conference on Computer Vision (ECCV), pages 35-51, 2020. 2 +[6] Florian Bordes, Richard Yuanzhe Pang, Anurag Ajay, Alexander C Li, Adrien Bardes, Suzanne Petryk, Oscar Manas, Zhiqiu Lin, Anas Mahmoud, Bargav Jayaraman, et al. An introduction to vision-language modeling. arXiv preprint arXiv:2405.17247, 2024. 2 +[7] Yingshan Chang, Mridu Narang, Hisami Suzuki, Guihong Cao, Jianfeng Gao, and Yonatan Bisk. Webqa: Multihop and multimodal qa. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 14178-14188, 2022. 1, 2 +[8] Jun Chen, Deyao Zhu, Xiaogian Shen, Xiang Li, Zechun Liu, Pengchuan Zhang, Raghuraman Krishnamoorthi, Vikas Chandra, Yunyang Xiong, and Mohamed Elhoseiny. Minigpt-v2: Large language model as a unified interface for vision-language multi-task learning. arXiv preprint arXiv:2310.15339, 2023. 2, 3 +[9] Wenhu Chen, Hexiang Hu, Xi Chen, Pat Verga, and William W. Cohen. MuRAG: Multimodal retrieval-augmented generator for open question answering over images and text. In Proceedings of the 2022 Conference on Empirical Methods in Natural Language Processing (EMNLP), pages 5558-5570. Association for Computational Linguistics, 2022. 3 +[10] Jacob Devlin, Ming-Wei Chang, Kenton Lee, and Kristina Toutanova. Bert: Pre-training of deep bidirectional transformers for language understanding. In Proceedings of the 2019 Conference of the North American Chapter of the Association for Computational Linguistics: Human Language Technologies, pages 4171–4186. Association for Computational Linguistics, 2019. 3 +[11] Chun-Mei Feng, Yang Bai, Tao Luo, Zhen Li, Salman Khan, Wangmeng Zuo, Xinxing Xu, Rick Siow Mong Goh, + +and Yong Liu. Vq4cir: Boosting composed image retrieval with visual question answering. arXiv preprint arXiv:2312.12273, 2023. 2, 3 +[12] Yash Goyal, Tejas Khot, Douglas Summers-Stay, Dhruv Batra, and Devi Parikh. Making the V in VQA matter: Elevating the role of image understanding in Visual Question Answering. In Conference on Computer Vision and Pattern Recognition (CVPR), 2017. 2, 3 +[13] Kelvin Guu, Kenton Lee, Zora Tung, Panupong Pasupat, and Ming-Wei Chang. Realm: Retrieval-augmented language model pre-training. In Proceedings of the 37th International Conference on Machine Learning (ICML), pages 3929-3938. PMLR, 2020. 3 +[14] Edward J. Hu, Yelong Shen, Phillip Wallis, Zeyuan Allen-Zhu, Yuanzhi Li, Shean Wang, Lu Wang, and Weizhu Chen. Lora: Low-rank adaptation of large language models, 2021. 6 +[15] Drew A Hudson and Christopher D Manning. Gqa: A new dataset for real-world visual reasoning and compositional question answering. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition, pages 6700-6709, 2019. 2 +[16] Gabriel Ilharco, Mitchell Wortsman, Ross Wightman, Cade Gordon, Nicholas Carlini, Rohan Taori, Achal Dave, Vaishaal Shankar, Hongseok Namkoong, John Miller, Hannaneh Hajishirzi, Ali Farhadi, and Ludwig Schmidt. Openclip. 2021. If you use this software, please cite it as below. 3, 6, 7 +[17] Vladimir Karpukhin, Barlas Oğuz, Sewon Min, Patrick Lewis, Ledell Wu, Sergey Edunov, Danqi Chen, and Wentau Yih. Dense passage retrieval for open-domain question answering. In Proceedings of the 2020 Conference on Empirical Methods in Natural Language Processing (EMNLP), pages 6769-6781. Association for Computational Linguistics, 2020. 3 +[18] Andreas Koukounas, Georgios Mastrapas, Michael Günther, Bo Wang, Scott Martens, Isabelle Mohr, Saba Sturua, Mohammad Kalim Akram, Joan Fontanals Martínez, Saahil Ognawala, Susana Guzman, Maximilian Werk, Nan Wang, and Han Xiao. Jina clip: Your clip model is also your text retriever, 2024. 6, 7 +[19] Ranjay Krishna, Yuke Zhu, Oliver Groth, Justin Johnson, Kenji Hata, Joshua Kravitz, Stephanie Chen, Yannis Kalantidis, Li-Jia Li, David A. Shamma, Michael S. Bernstein, and Li Fei-Fei. Visual genome: Connecting language and vision using crowdsourced dense image annotations. International Journal of Computer Vision, 123(1):32-73, 2017. 2 +[20] Bo Li, Yuanhan Zhang, Dong Guo, Renrui Zhang, Feng Li, Hao Zhang, Kaichen Zhang, Peiyuan Zhang, Yanwei Li, Ziwei Liu, and Chunyuan Li. Llava-onevision: Easy visual task transfer. arXiv preprint arXiv:2408.03326, 2024. 2, 3, 6, 7 +[21] Zhuang Liu, Hanzi Mao, Chao-Yuan Wu, Christoph Feichtenhofer, Trevor Darrell, and Saining Xie. A convnet for the 2020s. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 11976-11986, 2022. 6 +[22] Pan Lu, Tony Xia, Weicheng Shi, Ahmed El Kholy, Xi Victor Lin, Jianfeng Gao, Xiang Chen, and Kai-Wei Chang. Learn + +to explain: Multimodal reasoning via thought chains for science question answering. In Advances in Neural Information Processing Systems, 2022. 2 +[23] Pan Lu, Hritik Bansal, Tony Xia, Jiacheng Liu, Chunyuan Li, Hannaneh Hajishirzi, Hao Cheng, Kai-Wei Chang, Michel Galley, and Jianfeng Gao. Mathvista: Evaluating mathematical reasoning of foundation models in visual contexts. arXiv preprint arXiv:2310.02255, 2023. 2, 3 +[24] Hongyin Luo, Yung-Sung Chuang, Yuan Gong, Tianhua Zhang, Yoon Kim, Xixin Wu, Danny Fox, Helen Meng, and James Glass. SAIL: Search-augmented instruction learning. arXiv preprint arXiv:2305.15225, 2023. 3 +[25] Kenneth Marino, Mohammad Rastegari, Ali Farhadi, and Roozbeh Mottaghi. Ok-vqa: A visual question answering benchmark requiring external knowledge. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition, pages 3195-3204, 2019. 2 +[26] Ahmed Masry, Do Xuan Long, Jia Qing Tan, Shafiq Joty, and Enamul Hoque. Chartqa: A benchmark for question answering about charts with visual and logical reasoning. In Findings of the Association for Computational Linguistics: ACL 2022, pages 2263-2279, 2022. 2 +[27] Minesh Mathew, Viraj Bagal, Ruben Pérez Tito, Dimosthenis Karatzas, Ernest Valveny, and CV Jawahar. Infographicvqa. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 4390-4399, 2021. 2, 3 +[28] Minesh Mathew, Dimosthenis Karatzas, and CV Jawahar. Docvqa: A dataset for vqa on document images. In Proceedings of the IEEE/CVF Winter Conference on Applications of Computer Vision, pages 2200-2209, 2021. 2, 3 +[29] Zach Nussbaum, Brandon Duderstadt, and Andriy Mulyar. Nomic embed vision: Expanding the latent space, 2024. 6, 7 +[30] OpenAI. Gpt-4o: Enhanced multimodal language model. OpenAI Research, 2024. https://openai.com/index/hello-gpt-4o/.2,3,4,6,7 +[31] OpenAI. Gpt-4v: Multimodal language model with vision capabilities. OpenAI Research, 2024. https://openai.com/index/gpt-4/.2 +[32] Abhirama Subramanyam Penamakuri, Manish Gupta, Mithun Das Gupta, and Anand Mishra. Answer mining from a pool of images: Towards retrieval-based visual question answering. In Proceedings of the International Joint Conference on Artificial Intelligence (IJCAI), pages 1052-1058, 2023. 1, 2, 3 +[33] Alec Radford, Jong Wook Kim, Karthik Hallacy, Aditya Ramesh, Gabriel Goh, Sandhini Agarwal, Girish Sastry, Amanda Askell, Pamela Mishkin, Jack Clark, et al. Learning transferable visual models from natural language supervision. In Proceedings of the 38th International Conference on Machine Learning, pages 8748-8763. PMLR, 2021. 3, 6, 7 +[34] Shaoqing Ren, Kaiming He, Ross Girshick, and Jian Sun. Faster r-cnn: Towards real-time object detection with region proposal networks. Advances in Neural Information Processing Systems, 28, 2015. 3 +[35] Dustin Schwenk, Apoorv Khandelwal, Christopher Clark, Kenneth Marino, and Roozbeh Mottaghi. A-okvqa: A + +benchmark for visual question answering using world knowledge. In European Conference on Computer Vision (ECCV), pages 1-17, 2022. 2 +[36] Ray Smith et al. TesseractOCR engine. https://github.com/tesseract-ocr/tesseract, 2024. Accessed: 2024-11-06. 4 +[37] Alon Talmor, Sewon Min, Robin Jia, Yanai Elazar, Uriel Singer Hasson, and Danqi Chen. Multimodalqa: Complex question answering over text, tables, and images. In Proceedings of the International Conference on Learning Representations (ICLR), 2021. 2 +[38] Ryota Tanaka, Kyosuke Nishida, Kosuke Nishida, Taku Hasegawa, Itsumi Saito, and Kuniko Saito. Slidevqa: A dataset for document visual question answering on multiple images. In Proceedings of the AAAI Conference on Artificial Intelligence, pages 13636-13645, 2023. 2 +[39] Ke Wang, Yichi Zhang, and Hongsheng Li. Mini-gemini: An efficient and versatile vision-language model. arXiv preprint arXiv:2310.12345, 2023. 2 +[40] Ke Wang, Junting Pan, Weikang Shi, Zimu Lu, Mingjie Zhan, and Hongsheng Li. Measuring multimodal mathematical reasoning with math-vision dataset. arXiv preprint arXiv:2402.14804, 2024. 2 +[41] Peng Wang, Shuai Bai, Sinan Tan, Shijie Wang, Zhihao Fan, Jinze Bai, Keqin Chen, Xuejing Liu, Jialin Wang, Wenbin Ge, et al. Qwen2-vl: Enhancing vision-language model's perception of the world at any resolution. arXiv preprint arXiv:2409.12191, 2024. 2, 3, 7 +[42] Weiyun Wang, Shuibio Zhang, Yiming Ren, Yuchen Duan, Tiantong Li, Shuo Liu, Mengkang Hu, Zhe Chen, Kaipeng Zhang, Lewei Lu, et al. Needle in a multimodal haystack. arXiv preprint arXiv:2406.07230, 2024. 2 +[43] Tsung-Han Wu, Giscard Biamby, Jerome Quenum, Ritwik Gupta, Joseph E. Gonzalez, Trevor Darrell, and David M. Chan. Visual haystacks: A vision-centric needle-in-a-haystack benchmark. arXiv preprint arXiv:2407.13766, 2024. 2, 3, 7 +[44] Xiang Yue, Yuansheng Ni, Kai Zhang, Tianyu Zheng, Ruoqi Liu, Ge Zhang, Samuel Stevens, Dongfu Jiang, Weiming Ren, Yuxuan Sun, et al. Mmmu: A massive multi-discipline multimodal understanding and reasoning benchmark for expert agi. arXiv preprint arXiv:2311.16502, 2024. 2, 3 +[45] Xiaohua Zhai, Basil Mustafa, Alexander Kolesnikov, and Lucas Beyer. Sigmoid loss for language image pre-training. In Proceedings of the IEEE/CVF International Conference on Computer Vision (ICCV), pages 11975-11985, 2023. 3, 6, 7 +[46] Renrui Zhang, Dongzhi Jiang, Yichi Zhang, Haokun Lin, Ziyu Guo, Pengshuo Qiu, Aojun Zhou, Pan Lu, Kai-Wei Chang, Peng Gao, and Hongsheng Li. Mathverse: Does your multi-modal llm truly see the diagrams in visual math problems? arXiv preprint arXiv:2403.14624, 2024. 2, 3 +[47] Deyao Zhu, Jun Chen, Xiaogian Shen, Xiang Li, and Mohamed Elhoseiny. Minigpt-4: Enhancing vision-language understanding with advanced large language models. In International Conference on Learning Representations (ICLR), 2024. 2, 3 \ No newline at end of file diff --git a/2025/Document Haystacks_ Vision-Language Reasoning Over Piles of 1000+ Documents/images.zip b/2025/Document Haystacks_ Vision-Language Reasoning Over Piles of 1000+ Documents/images.zip new file mode 100644 index 0000000000000000000000000000000000000000..37f567141c46e6fe8c0ae2989f65ce3f9b012852 --- /dev/null +++ b/2025/Document Haystacks_ Vision-Language Reasoning Over Piles of 1000+ Documents/images.zip @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:05547fa92d1bf0f516caedbc00d85d0a852a712d8c98691bedce002824ff26c7 +size 445244 diff --git a/2025/Document Haystacks_ Vision-Language Reasoning Over Piles of 1000+ Documents/layout.json b/2025/Document Haystacks_ Vision-Language Reasoning Over Piles of 1000+ Documents/layout.json new file mode 100644 index 0000000000000000000000000000000000000000..63304476cae21388648ebad534d91cf7bf2bcbef --- /dev/null +++ b/2025/Document Haystacks_ Vision-Language Reasoning Over Piles of 1000+ Documents/layout.json @@ -0,0 +1,6935 @@ +{ + "pdf_info": [ + { + "para_blocks": [ + { + "bbox": [ + 108, + 103, + 504, + 138 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 108, + 103, + 504, + 138 + ], + "spans": [ + { + "bbox": [ + 108, + 103, + 504, + 138 + ], + "type": "text", + "content": "Document Haystacks: Vision-Language Reasoning Over Piles of 1000+ Documents" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 116, + 160, + 493, + 232 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 116, + 160, + 493, + 232 + ], + "spans": [ + { + "bbox": [ + 116, + 160, + 493, + 232 + ], + "type": "text", + "content": "Jun Chen" + }, + { + "bbox": [ + 116, + 160, + 493, + 232 + ], + "type": "inline_equation", + "content": "^{1}" + }, + { + "bbox": [ + 116, + 160, + 493, + 232 + ], + "type": "text", + "content": ", Dannong Xu" + }, + { + "bbox": [ + 116, + 160, + 493, + 232 + ], + "type": "inline_equation", + "content": "^{2}" + }, + { + "bbox": [ + 116, + 160, + 493, + 232 + ], + "type": "text", + "content": ", Junjie Fei" + }, + { + "bbox": [ + 116, + 160, + 493, + 232 + ], + "type": "inline_equation", + "content": "^{1}" + }, + { + "bbox": [ + 116, + 160, + 493, + 232 + ], + "type": "text", + "content": ", Chun-Mei Feng" + }, + { + "bbox": [ + 116, + 160, + 493, + 232 + ], + "type": "inline_equation", + "content": "^{3}" + }, + { + "bbox": [ + 116, + 160, + 493, + 232 + ], + "type": "text", + "content": ", Mohamed Elhoseiny" + }, + { + "bbox": [ + 116, + 160, + 493, + 232 + ], + "type": "inline_equation", + "content": "^{1}" + }, + { + "bbox": [ + 116, + 160, + 493, + 232 + ], + "type": "inline_equation", + "content": "^{1}" + }, + { + "bbox": [ + 116, + 160, + 493, + 232 + ], + "type": "text", + "content": "King Abdullah University of Science and Technology \n" + }, + { + "bbox": [ + 116, + 160, + 493, + 232 + ], + "type": "inline_equation", + "content": "^{2}" + }, + { + "bbox": [ + 116, + 160, + 493, + 232 + ], + "type": "text", + "content": "The University of Sydney, " + }, + { + "bbox": [ + 116, + 160, + 493, + 232 + ], + "type": "inline_equation", + "content": "^{3}" + }, + { + "bbox": [ + 116, + 160, + 493, + 232 + ], + "type": "text", + "content": "IHPC, A*STAR \n{jun.chen, junjie. fei, mohamed.elhoseiny}@kaust.edu.sa \ndaxu8019@uni.sydney.edu.au, fengcm.ai@gmail.com" + } + ] + } + ], + "index": 3 + }, + { + "type": "image", + "bbox": [ + 58, + 257, + 294, + 415 + ], + "blocks": [ + { + "bbox": [ + 58, + 257, + 294, + 415 + ], + "lines": [ + { + "bbox": [ + 58, + 257, + 294, + 415 + ], + "spans": [ + { + "bbox": [ + 58, + 257, + 294, + 415 + ], + "type": "image", + "image_path": "84598932cd96043268c6467ed2ade6f7cbfc2c116b10251727075dd11b974622.jpg" + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 101, + 421, + 251, + 440 + ], + "lines": [ + { + "bbox": [ + 101, + 421, + 251, + 440 + ], + "spans": [ + { + "bbox": [ + 101, + 421, + 251, + 440 + ], + "type": "text", + "content": "(a) Previous Benchmarks: Each question paired with a limited image set" + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "image_caption" + } + ], + "index": 4 + }, + { + "type": "image", + "bbox": [ + 299, + 258, + 549, + 415 + ], + "blocks": [ + { + "bbox": [ + 299, + 258, + 549, + 415 + ], + "lines": [ + { + "bbox": [ + 299, + 258, + 549, + 415 + ], + "spans": [ + { + "bbox": [ + 299, + 258, + 549, + 415 + ], + "type": "image", + "image_path": "20a0a7b3bd1b59437ec0c97bc20b4582e5abfc47b8608e808953356a12e6ee2d.jpg" + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 329, + 420, + 519, + 440 + ], + "lines": [ + { + "bbox": [ + 329, + 420, + 519, + 440 + ], + "spans": [ + { + "bbox": [ + 329, + 420, + 519, + 440 + ], + "type": "text", + "content": "(b) Our Benchmarks: All questions mapped to an extensive document collection" + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 54, + 448, + 555, + 504 + ], + "lines": [ + { + "bbox": [ + 54, + 448, + 555, + 504 + ], + "spans": [ + { + "bbox": [ + 54, + 448, + 555, + 504 + ], + "type": "text", + "content": "Figure 1. Comparison between previous and proposed benchmarks. Given a question as input, all benchmarks aim to retrieve relevant images from an image pool to correctly answer the question. Unlike prior benchmarks like RetVQA [32] and WebVQA [7], which structure their datasets by pairing each question with a limited set of images (typically " + }, + { + "bbox": [ + 54, + 448, + 555, + 504 + ], + "type": "inline_equation", + "content": "\\leq 30" + }, + { + "bbox": [ + 54, + 448, + 555, + 504 + ], + "type": "text", + "content": "), our benchmarks, DocHaystack and InfoHaystack, map each question to a substantially larger document collection, scaling up to 1,000 visual documents. This expanded scope more accurately represents large-scale document retrieval scenarios and offers a greater challenge in retrieval accuracy and visual question answering." + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "image_caption" + } + ], + "index": 6 + }, + { + "bbox": [ + 151, + 517, + 200, + 530 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 151, + 517, + 200, + 530 + ], + "spans": [ + { + "bbox": [ + 151, + 517, + 200, + 530 + ], + "type": "text", + "content": "Abstract" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 54, + 544, + 297, + 689 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 54, + 544, + 297, + 689 + ], + "spans": [ + { + "bbox": [ + 54, + 544, + 297, + 689 + ], + "type": "text", + "content": "Large multimodal models (LMMs) have achieved impressive progress in vision-language understanding, yet they face limitations in real-world applications requiring complex reasoning over a large number of images. Existing benchmarks for multi-image question-answering are limited in scope, each question is paired with only up to 30 images, which does not fully capture the demands of large-scale retrieval tasks encountered in the real-world usages. To reduce these gaps, we introduce two document haystack benchmarks, dubbed DocHaystack and InfoHaystack, designed to evaluate LMM performance on large-scale visual document retrieval and understanding. Addition" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 313, + 518, + 556, + 687 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 518, + 556, + 687 + ], + "spans": [ + { + "bbox": [ + 313, + 518, + 556, + 687 + ], + "type": "text", + "content": "ally, we propose V-RAG, a novel, vision-centric retrieval-augmented generation (RAG) framework that leverages a suite of multimodal vision encoders, each optimized for specific strengths, and a dedicated question-document relevance module. V-RAG sets a new standard, with a " + }, + { + "bbox": [ + 313, + 518, + 556, + 687 + ], + "type": "inline_equation", + "content": "9\\%" + }, + { + "bbox": [ + 313, + 518, + 556, + 687 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 313, + 518, + 556, + 687 + ], + "type": "inline_equation", + "content": "11\\%" + }, + { + "bbox": [ + 313, + 518, + 556, + 687 + ], + "type": "text", + "content": " improvement in Recall@1 on the challenging DocHaystack-1000 and InfoHaystack-1000 benchmarks, respectively, compared to the previous best baseline models. Additionally, integrating V-RAG with LMMs enables them to efficiently operate across thousands of images, yielding significant improvements on our DocHaystack and InfoHaystack benchmarks. Our code and datasets are available at https://github.com/Vision-CAIR/dochaystacks" + } + ] + } + ], + "index": 11 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "spans": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "type": "text", + "content": "CVF" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 145, + 0, + 496, + 37 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 145, + 0, + 496, + 37 + ], + "spans": [ + { + "bbox": [ + 145, + 0, + 496, + 37 + ], + "type": "text", + "content": "This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 70, + 703, + 139, + 712 + ], + "type": "page_footnote", + "angle": 0, + "lines": [ + { + "bbox": [ + 70, + 703, + 139, + 712 + ], + "spans": [ + { + "bbox": [ + 70, + 703, + 139, + 712 + ], + "type": "text", + "content": "* Equal contribution" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 293, + 748, + 318, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 318, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 318, + 757 + ], + "type": "text", + "content": "24817" + } + ] + } + ], + "index": 13 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 0 + }, + { + "para_blocks": [ + { + "bbox": [ + 56, + 72, + 136, + 83 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 72, + 136, + 83 + ], + "spans": [ + { + "bbox": [ + 56, + 72, + 136, + 83 + ], + "type": "text", + "content": "1. Introduction" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 55, + 91, + 294, + 258 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 91, + 294, + 258 + ], + "spans": [ + { + "bbox": [ + 55, + 91, + 294, + 258 + ], + "type": "text", + "content": "Large Multimodal Models (LMMs) [1, 20, 30, 41] have made remarkable progress in the vision-language understanding. However, these models still face challenges when tasked with reasoning over extensive collections of images or documents [43], limiting their effectiveness in real-world applications, such as visual search or querying over large sets of images or documents, like those stored on personal devices or in photo albums. However, there lacks such proper benchmarks to evaluate these capabilities. To address this gap, we introduce the DocHaystack and InfoHaystack benchmarks, designed to evaluate LMMs on large-scale image retrieval and understanding capabilities, pushing the boundaries of LMM performance in complex, real-world scenarios." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 55, + 258, + 295, + 425 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 258, + 295, + 425 + ], + "spans": [ + { + "bbox": [ + 55, + 258, + 295, + 425 + ], + "type": "text", + "content": "The existing multi-image retrieving and reasoning benchmarks are primarily constructed on a small scale, as highlighted in works such as [32, 37]. Each question in these benchmarks is paired with only up to 30 images as illustrated in Figure 1 (a). However, this limited scope does not align well with real-world scenarios, which often require retrieval and reasoning across hundreds or thousands of images or documents. In contrast, we introduce the Document Haystacks benchmarks, depicted in Figure 1 (b), which allow for querying questions from a large-scale collection of up to 1,000 documents, necessitating that models retrieve and reason over an extensive set of documents for each question. This scale better simulates practical applications and their demands." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 55, + 426, + 295, + 641 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 426, + 295, + 641 + ], + "spans": [ + { + "bbox": [ + 55, + 426, + 295, + 641 + ], + "type": "text", + "content": "The main challenge for constructing our document haystack benchmarks is collecting specific questions while ensuring there are no ambiguous answers across a large set of images. Existing datasets, such as those in DocVQA and InfographicVQA [27, 28], contain numerous \"general\" questions, like \"What is the table number?\", where answers could be derived from multiple images, leading to nonunique responses. To address this, we implemented a rigorous data filtering pipeline. First, we employed both a large language model (LLM) and human annotators to systematically filter out \"general\" questions based on carefully defined criteria. Additionally, we used the LLM to exclude questions relying on generic knowledge, such as \"What is the capital of Missouri?\", which can be answered without image context. This approach ensures that the questions in the benchmark can only be answered through specific visual cues from the provided images, maintaining the benchmark's integrity for evaluating image-based understanding." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 55, + 642, + 295, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 642, + 295, + 714 + ], + "spans": [ + { + "bbox": [ + 55, + 642, + 295, + 714 + ], + "type": "text", + "content": "To enable the current LMMs effectively reason over a large number of images, we propose a vision-centric retrieval-augmented generation (RAG) framework, named V-RAG. V-RAG combines multiple multimodal vision encoders, leveraging each encoder's unique strengths to enhance retrieval accuracy. Additionally, it incorporates an" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 313, + 72, + 553, + 215 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 72, + 553, + 215 + ], + "spans": [ + { + "bbox": [ + 313, + 72, + 553, + 215 + ], + "type": "text", + "content": "LMM-filter module to assess the relevance of each document to the query, refining the retrieval process by ensuring that only relevant documents are prioritized. This integrated approach allows V-RAG to navigate extensive document collections efficiently. Experimental results demonstrate that V-RAG achieves " + }, + { + "bbox": [ + 313, + 72, + 553, + 215 + ], + "type": "inline_equation", + "content": "9\\%" + }, + { + "bbox": [ + 313, + 72, + 553, + 215 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 313, + 72, + 553, + 215 + ], + "type": "inline_equation", + "content": "11\\%" + }, + { + "bbox": [ + 313, + 72, + 553, + 215 + ], + "type": "text", + "content": " improvement in Recall@1 on the DocHaystack-1000 and InfoHaystack-1000 compared to previous best text-to-image retrieval methods. Additionally, we found that integrating V-RAG brings GPT4o over a " + }, + { + "bbox": [ + 313, + 72, + 553, + 215 + ], + "type": "inline_equation", + "content": "55\\%" + }, + { + "bbox": [ + 313, + 72, + 553, + 215 + ], + "type": "text", + "content": " acc improvement on DocHaystack-200 and a " + }, + { + "bbox": [ + 313, + 72, + 553, + 215 + ], + "type": "inline_equation", + "content": "34\\%" + }, + { + "bbox": [ + 313, + 72, + 553, + 215 + ], + "type": "text", + "content": " acc improvement on InfoHaystack-200, indicating the effectiveness of our V-RAG." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 326, + 216, + 460, + 226 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 326, + 216, + 460, + 226 + ], + "spans": [ + { + "bbox": [ + 326, + 216, + 460, + 226 + ], + "type": "text", + "content": "Our contributions are as follows:" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 313, + 228, + 553, + 396 + ], + "type": "list", + "angle": 0, + "index": 9, + "blocks": [ + { + "bbox": [ + 314, + 228, + 553, + 323 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 314, + 228, + 553, + 323 + ], + "spans": [ + { + "bbox": [ + 314, + 228, + 553, + 323 + ], + "type": "text", + "content": "- We introduce the Document Haystack benchmarks, including DocHaystack-100/200/1000 and InfoHaystack-100/200/1000, with the most challenging setup consisting of 1,000 documents for each inquiry. These benchmarks advance document retrieval and reasoning tasks by requiring models to navigate and reason across extensive document collections, surpassing prior benchmarks limited to smaller retrieval tasks." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 313, + 324, + 553, + 396 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 324, + 553, + 396 + ], + "spans": [ + { + "bbox": [ + 313, + 324, + 553, + 396 + ], + "type": "text", + "content": "- We propose a vision-centric retrieval-augmented generation framework, V-RAG, which enhances the retrieval capabilities of LMMs. V-RAG achieves substantial improvements over previous best text-to-image retrieval methods by " + }, + { + "bbox": [ + 313, + 324, + 553, + 396 + ], + "type": "inline_equation", + "content": "9\\%" + }, + { + "bbox": [ + 313, + 324, + 553, + 396 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 313, + 324, + 553, + 396 + ], + "type": "inline_equation", + "content": "11\\%" + }, + { + "bbox": [ + 313, + 324, + 553, + 396 + ], + "type": "text", + "content": " on DocHaystack-1000 and InfoHaystack-1000, respectively." + } + ] + } + ], + "index": 8 + } + ], + "sub_type": "text" + }, + { + "bbox": [ + 313, + 406, + 406, + 418 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 406, + 406, + 418 + ], + "spans": [ + { + "bbox": [ + 313, + 406, + 406, + 418 + ], + "type": "text", + "content": "2. Related Works" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 313, + 426, + 553, + 653 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 426, + 553, + 653 + ], + "spans": [ + { + "bbox": [ + 313, + 426, + 553, + 653 + ], + "type": "text", + "content": "VQA benchmarks. VQA play a critical role in assessing a model's ability to understand and reason across visual contexts [6, 11, 12]. Traditional VQA datasets typically measure a model's comprehension of object attributes [15, 19], spatial relationships [15], as well as its understanding of documents [27, 28], charts [26], mathematics [23, 40, 46], and open knowledges [25, 35]. Additionally, these benchmarks explore models' knowledge across varied fields, including science and the arts [22, 44]. This broad array of benchmarks has greatly advanced vision-language models by cultivating diverse visual comprehension skills, particularly for modern foundation models in vision-language understanding [1, 3, 8, 20, 22, 30, 31, 39, 47]. Notably, these benchmarks have primarily focused on question answering within single image or document. In contrast, our benchmark shifts the focus towards retrieval and comprehensive understanding across a large collection of visual documents, presenting new challenges and expanding the scope of visual question answering." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 313, + 654, + 553, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 654, + 553, + 713 + ], + "spans": [ + { + "bbox": [ + 313, + 654, + 553, + 713 + ], + "type": "text", + "content": "Several previous efforts have tackled the challenge of visual question answering and reasoning across multiple images [5, 7, 32, 37, 38, 42]. For instance, datasets such as MultimodalQA [37] and ISVQA [5] require models to have multi-image reasoning abilities. Meanwhile, WebQA [7]" + } + ] + } + ], + "index": 12 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 318, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 318, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 318, + 757 + ], + "type": "text", + "content": "24818" + } + ] + } + ], + "index": 13 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 1 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 66, + 71, + 545, + 166 + ], + "blocks": [ + { + "bbox": [ + 66, + 71, + 545, + 166 + ], + "lines": [ + { + "bbox": [ + 66, + 71, + 545, + 166 + ], + "spans": [ + { + "bbox": [ + 66, + 71, + 545, + 166 + ], + "type": "image", + "image_path": "44a1374ccf0762cb7e5d4774343878b7d47e5654ddb937ab9ce7944c30bed756.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 54, + 182, + 555, + 240 + ], + "lines": [ + { + "bbox": [ + 54, + 182, + 555, + 240 + ], + "spans": [ + { + "bbox": [ + 54, + 182, + 555, + 240 + ], + "type": "text", + "content": "Figure 2. Data Curation Pipeline. Our benchmarks are curated based on the DocVQA and InfographicVQA datasets, following a three-step filtering process to obtain document-specific question-answer pairs. In Step 1, we filter out general questions (e.g., \"What is the table number?\"), as these could be answered by multiple documents and lack specificity. Step 2 involves a manual review by human annotators to further remove general questions. In Step 3, we eliminate generic-knowledge questions (e.g., \"How many sports were in the 2008 Beijing Paralympic Games?\") that can be answered directly by large language models without requiring image input." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "bbox": [ + 54, + 251, + 297, + 382 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 54, + 251, + 297, + 382 + ], + "spans": [ + { + "bbox": [ + 54, + 251, + 297, + 382 + ], + "type": "text", + "content": "and RetVQA [32] involve an additional step where models must first retrieve relevant images from a limited image pool before answering visual questions based on these results. However, these benchmarks are generally constrained to relatively small image pools, where each question is paired with an image set containing up to 30 images. In contrast, our proposed benchmarks, DocHaystack and InfoHaystack, significantly expand this scope by requiring models to retrieve and reason from a much larger set of up to 1,000 documents, presenting a notably greater challenge in retrieval and multi-image reasoning." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 55, + 386, + 296, + 579 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 386, + 296, + 579 + ], + "spans": [ + { + "bbox": [ + 55, + 386, + 296, + 579 + ], + "type": "text", + "content": "Large multimodal models (LMMs). LMMs have achieved substantial advancements in understanding and reasoning across single or multiple images [1, 8, 20, 30, 41, 47]. These models have significantly enhanced vision-language understanding across numerous dimensions and applications [12, 23, 44, 46]. LMMs benefit primarily from large-scale image-text alignment and extensive language modeling, which emerge them with advanced understanding and reasoning abilities. However, despite these breakthroughs, LMMs still encounter challenges when handling large-scale image or document sets [43]. This difficulty is due to the inherent complexity of processing such complex data. To address this, retrieval-based approaches have been developed to extend the capacity of vision-language models, augmenting their ability to process and reason over a larger number of images." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 55, + 582, + 297, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 582, + 297, + 715 + ], + "spans": [ + { + "bbox": [ + 55, + 582, + 297, + 715 + ], + "type": "text", + "content": "Retrieval-augmented generation (RAG). RAG integrates retrieval systems [4, 11, 16, 33, 45], with generative models, enhancing them with additional knowledge. While RAG has been extensively explored in language domains [2, 13, 17, 24], its application in vision-language contexts is also advancing. In vision-language RAG, models like MuRAG [9] leverage image-text memory to retrieve top-k neighbors by comparing inner-product similarities. RetVQA [32] uses an image-question relevance encoder, combining BERT [10] and Faster R-CNN [34] to filter relevant images, while MIRAGE [43] employs a CLIP-based" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 313, + 251, + 556, + 372 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 251, + 556, + 372 + ], + "spans": [ + { + "bbox": [ + 313, + 251, + 556, + 372 + ], + "type": "text", + "content": "encoder [33] to train a retriever. These frameworks extend model capabilities, enabling retrieval and reasoning across hundreds or thousands of images. In contrast, we propose V-RAG, a vision-centric RAG framework that integrates multiple vision encoders to more effectively capture image features, and introduces a LMM-based question-document relevance comparison module. Our results demonstrate that V-RAG surpasses existing methods on our DocHaystack and InfoHaystack benchmarks, setting a new standard for large-scale visual retrieval and reasoning." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 313, + 380, + 555, + 407 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 380, + 555, + 407 + ], + "spans": [ + { + "bbox": [ + 313, + 380, + 555, + 407 + ], + "type": "text", + "content": "3. DocHaystack and InfoHaystack Benchmarks" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 313, + 415, + 556, + 521 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 415, + 556, + 521 + ], + "spans": [ + { + "bbox": [ + 313, + 415, + 556, + 521 + ], + "type": "text", + "content": "To support effective retrieval and reasoning across extensive document collections, we present two new benchmarks—DocHaystack and InfoHaystack—designed to ensure each question yields a unique, document-specific answer. Derived from DocVQA [28] and InfographicVQA [27], these benchmarks address the challenge of answer ambiguity by selectively curating questions that can only be answered by a single document within a large dataset." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 313, + 522, + 557, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 522, + 557, + 714 + ], + "spans": [ + { + "bbox": [ + 313, + 522, + 557, + 714 + ], + "type": "text", + "content": "Benchmark construction pipeline. There exists many general questions in the existing benchmarks and lead to multiple answers for different document context. For example, general questions like \"What is the table number?\" may apply to various documents and yield multiple valid answers, while a targeted question like \"Who is the reviewer for the article titled 'An antithyroid factor in milk'?\" is likely to produce a unique answer, as only a single document or a limited set of documents would contain that information. Therefore, our benchmark construction follows a structured three-step filtering pipeline, illustrated in Figure 2, to ensure high-quality, unique-answer questions. First, we employ a large language model (LLM) to filter out general questions that could generate multiple answers across documents. Next, a manual review step further" + } + ] + } + ], + "index": 8 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "type": "text", + "content": "24819" + } + ] + } + ], + "index": 9 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 2 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 63, + 70, + 549, + 289 + ], + "blocks": [ + { + "bbox": [ + 63, + 70, + 549, + 289 + ], + "lines": [ + { + "bbox": [ + 63, + 70, + 549, + 289 + ], + "spans": [ + { + "bbox": [ + 63, + 70, + 549, + 289 + ], + "type": "image", + "image_path": "f521efde925d2b87455540c536284a49453fb485b9b8a3388f1312a8a9bf9242.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 54, + 310, + 555, + 377 + ], + "lines": [ + { + "bbox": [ + 54, + 310, + 555, + 377 + ], + "spans": [ + { + "bbox": [ + 54, + 310, + 555, + 377 + ], + "type": "text", + "content": "Figure 3. The V-RAG pipeline workflow. In the top section, a vision encoder ensemble is used, combining multiple vision models—CLIP, SigLIP, and OpenCLIP—to process a large document haystack. Each encoder computes similarity scores, which are averaged into " + }, + { + "bbox": [ + 54, + 310, + 555, + 377 + ], + "type": "inline_equation", + "content": "Sim_{avg}" + }, + { + "bbox": [ + 54, + 310, + 555, + 377 + ], + "type": "text", + "content": ". The top m documents, based on these scores, are selected for further analysis. In the bottom right, the LMM-Filter Module utilizes a pretrained LMM to assess whether each selected document can potentially answer the posed question. This filtering step removes documents that do not match, retaining only relevant ones. Finally, the top k most relevant images are input into the LMM along with the original question " + }, + { + "bbox": [ + 54, + 310, + 555, + 377 + ], + "type": "inline_equation", + "content": "q" + }, + { + "bbox": [ + 54, + 310, + 555, + 377 + ], + "type": "text", + "content": " to generate a specific answer." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "bbox": [ + 54, + 389, + 294, + 435 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 54, + 389, + 294, + 435 + ], + "spans": [ + { + "bbox": [ + 54, + 389, + 294, + 435 + ], + "type": "text", + "content": "checks the questions to ensure the data quality. Finally, a generic-knowledge filtering stage refines the dataset further, retaining only questions closely tied to specific document content." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 54, + 437, + 295, + 533 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 54, + 437, + 295, + 533 + ], + "spans": [ + { + "bbox": [ + 54, + 437, + 295, + 533 + ], + "type": "text", + "content": "This carefully designed pipeline, combining LLM-based filtering and human review, effectively curates questions that drive accurate, document-specific retrieval. By focusing on reducing answer ambiguity, DocHaystack and InfoHaystack enhance the precision of retrieval and reasoning in large-scale document processing tasks, providing a valuable tool for the evaluation of retrieval systems. We discuss this data curation pipeline in details as follows:" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 54, + 533, + 295, + 640 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 54, + 533, + 295, + 640 + ], + "spans": [ + { + "bbox": [ + 54, + 533, + 295, + 640 + ], + "type": "text", + "content": "General-question LLM filtering. We begin by using the LLM, GPT-4o [30], to filter out general questions through a set of well-crafted instructions. Leveraging the LLM's strong contextual understanding, this initial filtering step allows us to efficiently process large volumes of data, identifying broad or ambiguous questions that may yield multiple answers across documents. This automated approach significantly enhances the benchmark construction's efficiency and quality." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 54, + 641, + 295, + 700 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 54, + 641, + 295, + 700 + ], + "spans": [ + { + "bbox": [ + 54, + 641, + 295, + 700 + ], + "type": "text", + "content": "To guide the LLM, we first define the task, providing clear distinctions between general and specific questions along with illustrative examples. With this framework, the LLM can then assess each question and determine if it is general or specific. The instructional format is as follows:" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 67, + 701, + 295, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 67, + 701, + 295, + 714 + ], + "spans": [ + { + "bbox": [ + 67, + 701, + 295, + 714 + ], + "type": "text", + "content": "You are an evaluator tasked with identifying if a ques" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 313, + 389, + 553, + 484 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 389, + 553, + 484 + ], + "spans": [ + { + "bbox": [ + 313, + 389, + 553, + 484 + ], + "type": "text", + "content": "tion is specific or general. A general question seeks commonly known or widely applicable information without unique identifiers, e.g., \"Who is the person standing in the ground?\" A specific question, however, requests unique information about a particular individual, event, or object, e.g., \"What is the Social Security Number of Charles Yarbrough?\" Based on these definitions, determine if the following question is general or specific: {question}." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 313, + 485, + 553, + 544 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 485, + 553, + 544 + ], + "spans": [ + { + "bbox": [ + 313, + 485, + 553, + 544 + ], + "type": "text", + "content": "General-question manual review. After the initial LLM filtering, we conduct a manual review of the questions that were classified as specific. This manual process involves two key steps to ensure answer uniqueness and benchmark quality." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 313, + 545, + 553, + 616 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 545, + 553, + 616 + ], + "spans": [ + { + "bbox": [ + 313, + 545, + 553, + 616 + ], + "type": "text", + "content": "In the first step, we examine each question to confirm it contains unique identifiers—such as names, dates, titles, or other specific attributes—suggesting a document-specific answer. This careful check helps identify questions with clear, unique markers that direct the retrieval process to a single document." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 313, + 617, + 555, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 617, + 555, + 713 + ], + "spans": [ + { + "bbox": [ + 313, + 617, + 555, + 713 + ], + "type": "text", + "content": "In the second step, we verify the uniqueness of each answer to eliminate any remaining ambiguity. Although specific identifiers are present, questions may still be prone to ambiguity, such as with common names or recurring book titles. To address this, we employ a refined verification process. First, we use Optical Character Recognition (OCR) [36] to extract all text from images in the dataset. We then search for occurrences of the unique identifiers re" + } + ] + } + ], + "index": 10 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "type": "text", + "content": "24820" + } + ] + } + ], + "index": 11 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 3 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 58, + 70, + 294, + 114 + ], + "blocks": [ + { + "bbox": [ + 58, + 70, + 294, + 114 + ], + "lines": [ + { + "bbox": [ + 58, + 70, + 294, + 114 + ], + "spans": [ + { + "bbox": [ + 58, + 70, + 294, + 114 + ], + "type": "table", + "html": "
GPT-4oLLaVA-OneVisionQwen2-VL
DocVQA27.0%4.7%3.4%
InfographicVQA56.4%13.4%11.3%
", + "image_path": "d437293a87d0d4a0b1dccef251980a432c874ef29e4a86dbe99487f17300256a.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "table_body" + } + ], + "index": 0 + }, + { + "bbox": [ + 55, + 121, + 295, + 188 + ], + "lines": [ + { + "bbox": [ + 55, + 121, + 295, + 188 + ], + "spans": [ + { + "bbox": [ + 55, + 121, + 295, + 188 + ], + "type": "text", + "content": "Table 1. Percentage of questions answerable by LMMs without vision input. We evaluate GPT-4o, LLaVA-Onevision, and Qwen2-VL on their ability to answer questions directly from our dataset without requiring vision input. The reported percentage reflects the proportion of examples that can be answered solely through language understanding." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "text" + }, + { + "bbox": [ + 54, + 209, + 295, + 281 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 54, + 209, + 295, + 281 + ], + "spans": [ + { + "bbox": [ + 54, + 209, + 295, + 281 + ], + "type": "text", + "content": "tained from the first step across other documents. If matches are found, a manual review is conducted to ensure no alternative valid answers exist. This comprehensive approach minimizes the possibility of a single question mapping to multiple answers, enhancing the precision and reliability of our benchmarks." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 55, + 281, + 295, + 317 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 281, + 295, + 317 + ], + "spans": [ + { + "bbox": [ + 55, + 281, + 295, + 317 + ], + "type": "text", + "content": "Generic-knowledge filtering. In DocVQA and InfographicVQA tasks, certain questions—such as “How many sports were in the 2008 Beijing" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 55, + 318, + 295, + 425 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 318, + 295, + 425 + ], + "spans": [ + { + "bbox": [ + 55, + 318, + 295, + 425 + ], + "type": "text", + "content": "Paralympic Games?\"—can be answered based on general knowledge accessible to a large language model, without relying on the image content. This introduces a language bias when using LMMs for visual question answering, as it shifts the focus away from image-based reasoning. To address this, we filter out these general-knowledge questions, ensuring that evaluation emphasizes vision-based understanding and that models rely primarily on visual content to generate accurate answers." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 55, + 426, + 295, + 617 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 426, + 295, + 617 + ], + "spans": [ + { + "bbox": [ + 55, + 426, + 295, + 617 + ], + "type": "text", + "content": "To implement this, we developed an LLM-based evaluation pipeline that detects and excludes such questions. For each question, we prompt an LLM with “{question}, answer briefly.” After receiving a response, we compare it to the ground-truth answer using another LLM. If the response matches the ground truth, we classify the question as general knowledge-related and remove it, thereby isolating questions that truly require visual document understanding. As shown in Table 1, GPT-4o accurately answers " + }, + { + "bbox": [ + 55, + 426, + 295, + 617 + ], + "type": "inline_equation", + "content": "27.0\\%" + }, + { + "bbox": [ + 55, + 426, + 295, + 617 + ], + "type": "text", + "content": " of DocVQA questions and " + }, + { + "bbox": [ + 55, + 426, + 295, + 617 + ], + "type": "inline_equation", + "content": "56.4\\%" + }, + { + "bbox": [ + 55, + 426, + 295, + 617 + ], + "type": "text", + "content": " of InfographicVQA questions directly, a rate significantly higher than that of open-source LLMs. Therefore, we select GPT-4o to filter out the questions that can be directly answered by the GPT-4o model. Overall, this process is to ensure that the evaluation reflects the necessity of vision-based comprehension." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 55, + 618, + 295, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 618, + 295, + 714 + ], + "spans": [ + { + "bbox": [ + 55, + 618, + 295, + 714 + ], + "type": "text", + "content": "Final dataset profile. After a rigorous three-stage data filtering process, we retained 109 questions from DocVQA and 155 questions from InfographicVQA, associated with 59 and 66 documents that provide the evidence, respectively. To assess retrieval performance at scale, we introduce two benchmarks: DocHaystack-1000 and InfoHaystack-1000, where each question requires retrieving relevant content from a set of 1,000 documents. Given" + } + ] + } + ], + "index": 6 + }, + { + "type": "image", + "bbox": [ + 321, + 75, + 421, + 182 + ], + "blocks": [ + { + "bbox": [ + 321, + 75, + 421, + 182 + ], + "lines": [ + { + "bbox": [ + 321, + 75, + 421, + 182 + ], + "spans": [ + { + "bbox": [ + 321, + 75, + 421, + 182 + ], + "type": "image", + "image_path": "f8264d51318bf845052e4510597aaf53449d962463b65f1bcef0c4390c546d80.jpg" + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 313, + 194, + 555, + 228 + ], + "lines": [ + { + "bbox": [ + 313, + 194, + 555, + 228 + ], + "spans": [ + { + "bbox": [ + 313, + 194, + 555, + 228 + ], + "type": "text", + "content": "Figure 4. Question type analysis. We analyze the distribution of question types of DocHaystack and InfoHaystack. Each benchmark categorizes the data into 5 different types." + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "image_caption" + } + ], + "index": 7 + }, + { + "type": "image", + "bbox": [ + 439, + 76, + 544, + 186 + ], + "blocks": [ + { + "bbox": [ + 439, + 76, + 544, + 186 + ], + "lines": [ + { + "bbox": [ + 439, + 76, + 544, + 186 + ], + "spans": [ + { + "bbox": [ + 439, + 76, + 544, + 186 + ], + "type": "image", + "image_path": "bcf44f1ff8cc585290afca51d4c3ff1172c7e6440de95546fb4b832627797fe4.jpg" + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "image_body" + } + ], + "index": 8 + }, + { + "bbox": [ + 313, + 238, + 555, + 357 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 238, + 555, + 357 + ], + "spans": [ + { + "bbox": [ + 313, + 238, + 555, + 357 + ], + "type": "text", + "content": "the challenge this scale presents to current LMMs, particularly in terms of context length limitations, we also construct two smaller benchmarks: DocHaystack-100/200 and InfoHaystack-100/200. These benchmarks allow direct input of all associated images into the context, enabling evaluation of models' long-context comprehension ability. For training set, we also construct a dataset comprising 2,835 questions similarly, with 899 from DocVQA and 1,936 from InfographicVQA, to support robust learning and generalization for the multi-image reasoning tasks." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 313, + 358, + 556, + 490 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 358, + 556, + 490 + ], + "spans": [ + { + "bbox": [ + 313, + 358, + 556, + 490 + ], + "type": "text", + "content": "Question type analysis. The types of questions represent the types of the evidence required for accurate answers. In Figure 4, we illustrate the distribution of question types across our dataset to provide insights into its structure. Following the classification system used in DocVQA and InfographicVQA, we categorize questions accordingly (note that a single question may fall into multiple categories). As shown in the figure, the DocHaystack benchmark places a greater emphasis on Table/List and Layout understanding, whereas InfoHaystack primarily targets Figure, Text, and Table comprehension." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 313, + 501, + 370, + 514 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 501, + 370, + 514 + ], + "spans": [ + { + "bbox": [ + 313, + 501, + 370, + 514 + ], + "type": "text", + "content": "4. Method" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 313, + 521, + 555, + 689 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 521, + 555, + 689 + ], + "spans": [ + { + "bbox": [ + 313, + 521, + 555, + 689 + ], + "type": "text", + "content": "Current large multimodal models (LMMs) face substantial challenges when reasoning across hundreds or thousands of images, due not only to context length limitations but also to the inherent complexity of the task. This issue is particularly pronounced in our benchmarks, which contain 1k document files requiring high-resolution images to capture and interpret small-font text effectively. To enable LMMs to perform reasoning over a substantial number of documents, we introduce a vision-centric retrieval-augmented generation (V-RAG) framework. V-RAG efficiently retrieves a reduced set of relevant documents, allowing the LMM to focus on a manageable subset for deeper understanding, as illustrated in Figure 3. In the following section, we provide a detailed description of the V-RAG pipeline." + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 313, + 689, + 556, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 689, + 556, + 715 + ], + "spans": [ + { + "bbox": [ + 313, + 689, + 556, + 715 + ], + "type": "text", + "content": "Task definition. Given a question " + }, + { + "bbox": [ + 313, + 689, + 556, + 715 + ], + "type": "inline_equation", + "content": "q" + }, + { + "bbox": [ + 313, + 689, + 556, + 715 + ], + "type": "text", + "content": " and a collection of " + }, + { + "bbox": [ + 313, + 689, + 556, + 715 + ], + "type": "inline_equation", + "content": "N" + }, + { + "bbox": [ + 313, + 689, + 556, + 715 + ], + "type": "text", + "content": " documents " + }, + { + "bbox": [ + 313, + 689, + 556, + 715 + ], + "type": "inline_equation", + "content": "\\mathcal{D} = \\{D_1, \\ldots, D_N\\}" + }, + { + "bbox": [ + 313, + 689, + 556, + 715 + ], + "type": "text", + "content": ", the V-RAG framework" + } + ] + } + ], + "index": 14 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "text", + "content": "24821" + } + ] + } + ], + "index": 15 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 4 + }, + { + "para_blocks": [ + { + "bbox": [ + 55, + 72, + 294, + 133 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 72, + 294, + 133 + ], + "spans": [ + { + "bbox": [ + 55, + 72, + 294, + 133 + ], + "type": "text", + "content": "aims to retrieve the top-" + }, + { + "bbox": [ + 55, + 72, + 294, + 133 + ], + "type": "inline_equation", + "content": "k" + }, + { + "bbox": [ + 55, + 72, + 294, + 133 + ], + "type": "text", + "content": " most relevant documents to support LMMs understanding and answering the question " + }, + { + "bbox": [ + 55, + 72, + 294, + 133 + ], + "type": "inline_equation", + "content": "q" + }, + { + "bbox": [ + 55, + 72, + 294, + 133 + ], + "type": "text", + "content": ". V-RAG accomplishes this through a two-step retrieval process designed to effectively identify and rank relevant documents for each question." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 55, + 134, + 295, + 314 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 134, + 295, + 314 + ], + "spans": [ + { + "bbox": [ + 55, + 134, + 295, + 314 + ], + "type": "text", + "content": "Vision encoder ensemble. Document files often contain a mix of text, symbols, and visual elements across various scales, requiring vision encoders to capture a comprehensive understanding of these complex structures. To efficiently handle this diversity, we represent each document as an image and utilize an ensemble of vision encoders, including CLIP [33], SigLIP [45], and OpenCLIP [16], each bringing distinct strengths to the image understanding, as depicted in Figure 3. For example, the ConvNext encoder [21] from OpenCLIP [16] is particularly effective for high-resolution image encoding. We compute the similarity score between each question " + }, + { + "bbox": [ + 55, + 134, + 295, + 314 + ], + "type": "inline_equation", + "content": "q" + }, + { + "bbox": [ + 55, + 134, + 295, + 314 + ], + "type": "text", + "content": " and all documents in the document set " + }, + { + "bbox": [ + 55, + 134, + 295, + 314 + ], + "type": "inline_equation", + "content": "\\mathcal{D}" + }, + { + "bbox": [ + 55, + 134, + 295, + 314 + ], + "type": "text", + "content": " according to Equation 1, with similarity scores from each encoder represented as " + }, + { + "bbox": [ + 55, + 134, + 295, + 314 + ], + "type": "inline_equation", + "content": "Sim_{c}" + }, + { + "bbox": [ + 55, + 134, + 295, + 314 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 55, + 134, + 295, + 314 + ], + "type": "inline_equation", + "content": "Sim_{o}" + }, + { + "bbox": [ + 55, + 134, + 295, + 314 + ], + "type": "text", + "content": ", and " + }, + { + "bbox": [ + 55, + 134, + 295, + 314 + ], + "type": "inline_equation", + "content": "Sim_{s}" + }, + { + "bbox": [ + 55, + 134, + 295, + 314 + ], + "type": "text", + "content": " respectively." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 91, + 335, + 295, + 348 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 91, + 335, + 295, + 348 + ], + "spans": [ + { + "bbox": [ + 91, + 335, + 295, + 348 + ], + "type": "interline_equation", + "content": "\\mathcal {S} (q, \\mathcal {D}) = \\cos \\left(\\phi_ {\\mathrm {t}} (q), \\phi_ {\\mathrm {v}} (D _ {j})\\right) \\mid D _ {j} \\in \\mathcal {D}, \\tag {1}", + "image_path": "fba7f299a389df96064a6eded86d09d73ef1454e492b3a5d2f811ad1ee413aba.jpg" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 55, + 359, + 295, + 407 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 359, + 295, + 407 + ], + "spans": [ + { + "bbox": [ + 55, + 359, + 295, + 407 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 55, + 359, + 295, + 407 + ], + "type": "inline_equation", + "content": "S" + }, + { + "bbox": [ + 55, + 359, + 295, + 407 + ], + "type": "text", + "content": " denotes the computing the similarity between the query " + }, + { + "bbox": [ + 55, + 359, + 295, + 407 + ], + "type": "inline_equation", + "content": "q" + }, + { + "bbox": [ + 55, + 359, + 295, + 407 + ], + "type": "text", + "content": " and a collection of documents " + }, + { + "bbox": [ + 55, + 359, + 295, + 407 + ], + "type": "inline_equation", + "content": "\\mathcal{D}" + }, + { + "bbox": [ + 55, + 359, + 295, + 407 + ], + "type": "text", + "content": ". " + }, + { + "bbox": [ + 55, + 359, + 295, + 407 + ], + "type": "inline_equation", + "content": "\\cos" + }, + { + "bbox": [ + 55, + 359, + 295, + 407 + ], + "type": "text", + "content": " denotes the cosine similarity. " + }, + { + "bbox": [ + 55, + 359, + 295, + 407 + ], + "type": "inline_equation", + "content": "\\phi_{\\mathrm{t}}" + }, + { + "bbox": [ + 55, + 359, + 295, + 407 + ], + "type": "text", + "content": " denotes the text encoder, and " + }, + { + "bbox": [ + 55, + 359, + 295, + 407 + ], + "type": "inline_equation", + "content": "\\phi_{\\mathrm{v}}" + }, + { + "bbox": [ + 55, + 359, + 295, + 407 + ], + "type": "text", + "content": " denotes the vision encoder." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 55, + 410, + 295, + 470 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 410, + 295, + 470 + ], + "spans": [ + { + "bbox": [ + 55, + 410, + 295, + 470 + ], + "type": "text", + "content": "To derive a final relevance score, we calculate the average similarity " + }, + { + "bbox": [ + 55, + 410, + 295, + 470 + ], + "type": "inline_equation", + "content": "Sim_{avg}" + }, + { + "bbox": [ + 55, + 410, + 295, + 470 + ], + "type": "text", + "content": " for each question-image pair by combining " + }, + { + "bbox": [ + 55, + 410, + 295, + 470 + ], + "type": "inline_equation", + "content": "Sim_c" + }, + { + "bbox": [ + 55, + 410, + 295, + 470 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 55, + 410, + 295, + 470 + ], + "type": "inline_equation", + "content": "Sim_o" + }, + { + "bbox": [ + 55, + 410, + 295, + 470 + ], + "type": "text", + "content": ", and " + }, + { + "bbox": [ + 55, + 410, + 295, + 470 + ], + "type": "inline_equation", + "content": "Sim_s" + }, + { + "bbox": [ + 55, + 410, + 295, + 470 + ], + "type": "text", + "content": ". We then rank the images based on " + }, + { + "bbox": [ + 55, + 410, + 295, + 470 + ], + "type": "inline_equation", + "content": "Sim_{avg}" + }, + { + "bbox": [ + 55, + 410, + 295, + 470 + ], + "type": "text", + "content": " in descending order, selecting the top-" + }, + { + "bbox": [ + 55, + 410, + 295, + 470 + ], + "type": "inline_equation", + "content": "m" + }, + { + "bbox": [ + 55, + 410, + 295, + 470 + ], + "type": "text", + "content": " most relevant images according to their similarity scores." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 55, + 472, + 295, + 604 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 472, + 295, + 604 + ], + "spans": [ + { + "bbox": [ + 55, + 472, + 295, + 604 + ], + "type": "text", + "content": "LMM-filter module. To refine the selection of top-" + }, + { + "bbox": [ + 55, + 472, + 295, + 604 + ], + "type": "inline_equation", + "content": "m" + }, + { + "bbox": [ + 55, + 472, + 295, + 604 + ], + "type": "text", + "content": " relevant images further, we introduce a LMM-based question-image relevance assessment module. This module evaluates the relevance between each question and the top-" + }, + { + "bbox": [ + 55, + 472, + 295, + 604 + ], + "type": "inline_equation", + "content": "m" + }, + { + "bbox": [ + 55, + 472, + 295, + 604 + ], + "type": "text", + "content": " images identified in the first filtering step. Specifically, we pair each image with the question text and input them into an open-source vision-language model, prompting, \"Can this image provide answers to this question? Respond only with yes or no\". We only retain the question-image pairs that are identified as \"yes\" from LMM, and remove other irrelevant images." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 55, + 605, + 295, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 605, + 295, + 714 + ], + "spans": [ + { + "bbox": [ + 55, + 605, + 295, + 714 + ], + "type": "text", + "content": "LMM-VQA module. Achieving high top-1 ranking accuracy in image retrieval is challenging, so we retain the top-" + }, + { + "bbox": [ + 55, + 605, + 295, + 714 + ], + "type": "inline_equation", + "content": "k" + }, + { + "bbox": [ + 55, + 605, + 295, + 714 + ], + "type": "text", + "content": " images from the LMM-filtered ranking list and present them to the LMM-VQA to improve the likelihood of including relevant images. We input these top-" + }, + { + "bbox": [ + 55, + 605, + 295, + 714 + ], + "type": "inline_equation", + "content": "k" + }, + { + "bbox": [ + 55, + 605, + 295, + 714 + ], + "type": "text", + "content": " images alongside the question into the LMM-VQA (see Figure 3), which then generates the answer directly. To enhance robustness against visual distractors, the LMM-VQA can be further optimized, as analyzed in the experiment section." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 314, + 71, + 395, + 85 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 314, + 71, + 395, + 85 + ], + "spans": [ + { + "bbox": [ + 314, + 71, + 395, + 85 + ], + "type": "text", + "content": "5. Experiments" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 313, + 93, + 553, + 154 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 93, + 553, + 154 + ], + "spans": [ + { + "bbox": [ + 313, + 93, + 553, + 154 + ], + "type": "text", + "content": "In the experiments section, we will primarily describe our training setup, covering evaluation metrics, baseline models, and the fine-tuning procedure for the LMM-VQA model. We also present the main experimental results along with an ablation study to provide further insights." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 314, + 167, + 405, + 181 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 314, + 167, + 405, + 181 + ], + "spans": [ + { + "bbox": [ + 314, + 167, + 405, + 181 + ], + "type": "text", + "content": "5.1. Training setup" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 313, + 187, + 553, + 307 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 187, + 553, + 307 + ], + "spans": [ + { + "bbox": [ + 313, + 187, + 553, + 307 + ], + "type": "text", + "content": "Metric. In our evaluation of the DocHaystack and Info-Haystack benchmarks, we employ a model-based assessment by leveraging GPT-4o-mini [30] to accurately determine whether the model predictions match target answers. This method uses a carefully structured prompt to facilitate GPT-4o-mini's evaluation of answer correctness. We empirically found that the model-based evaluation achieves higher consistency and alignment with human judgment. Additional details on the prompt design are provided in the Appendix." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 313, + 309, + 553, + 357 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 309, + 553, + 357 + ], + "spans": [ + { + "bbox": [ + 313, + 309, + 553, + 357 + ], + "type": "text", + "content": "For the document retrieval evaluation, we report the baseline results using Recall@1, Recall@3, and Recall@5 metrics. These metrics enable a thorough assessment of retrieval accuracy across varying levels of precision." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 313, + 360, + 554, + 514 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 360, + 554, + 514 + ], + "spans": [ + { + "bbox": [ + 313, + 360, + 554, + 514 + ], + "type": "text", + "content": "Baselines. In our experiment, we have evaluated several open and closed-sourced vision-language models on the retrieval and VQA performance. For the large multimodal model, we used the gpt-4o-2024-08-06 version of GPT-4o [30], the LLaVA-OneVision-Qwen2-7b-OV-HF version of LLaVA-OneVision [20], and the Qwen2-VL-7B-Instruct version of Qwen2-VL [3]. For computing the text-to-image similarities, we employed the Jina-CLIP-v1 [18] variant, Nomic-Embed-Vision-v1.5 [29] variant, CLIP [33] ViT-L/14@336 variant, for SigLIP [45], the ViT-SO400M/14@384 variant, and for OpenCLIP [16], the ConvNeXt-XXL@1024 variant as well as text-based method, BM25." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 313, + 517, + 553, + 566 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 517, + 553, + 566 + ], + "spans": [ + { + "bbox": [ + 313, + 517, + 553, + 566 + ], + "type": "text", + "content": "In our V-RAG setting, we apply LLaVA-OneVision-Qwen2-7b-OV-HF for the LMM-filter module and Qwen2-VL-7B-Instruct for the LMM-VQA module. We select " + }, + { + "bbox": [ + 313, + 517, + 553, + 566 + ], + "type": "inline_equation", + "content": "m" + }, + { + "bbox": [ + 313, + 517, + 553, + 566 + ], + "type": "text", + "content": " as 60 and " + }, + { + "bbox": [ + 313, + 517, + 553, + 566 + ], + "type": "inline_equation", + "content": "k" + }, + { + "bbox": [ + 313, + 517, + 553, + 566 + ], + "type": "text", + "content": " as 5 in our experiment." + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 313, + 567, + 553, + 616 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 567, + 553, + 616 + ], + "spans": [ + { + "bbox": [ + 313, + 567, + 553, + 616 + ], + "type": "text", + "content": "Optimizing the LMM-VQA module. To improve the robustness of the LMM-VQA model in handling visual question answering with multiple distractor images, we further fine-tune the model using our curated training data." + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 313, + 617, + 555, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 617, + 555, + 714 + ], + "spans": [ + { + "bbox": [ + 313, + 617, + 555, + 714 + ], + "type": "text", + "content": "During this fine-tuning process, we introduce 1-10 randomly sampled distractor images for each question, creating a challenging setting that encourages the model to focus on relevant content amid a mix of positive and negative images. The fine-tuning is conducted with a batch size of 32 and a peak learning rate of 1e-4 over a single epoch. Additionally, we leverage LoRA [14] with a rank of 8 to efficiently adapt the model's parameters during training." + } + ] + } + ], + "index": 15 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "text", + "content": "24822" + } + ] + } + ], + "index": 16 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 5 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 103, + 69, + 506, + 247 + ], + "blocks": [ + { + "bbox": [ + 103, + 69, + 506, + 247 + ], + "lines": [ + { + "bbox": [ + 103, + 69, + 506, + 247 + ], + "spans": [ + { + "bbox": [ + 103, + 69, + 506, + 247 + ], + "type": "table", + "html": "
DocHaystack-100DocHaystack-200DocHaystack-1000
R@1R@3R@5R@1R@3R@5R@1R@3R@5
BM25 (OCR)63.375.279.865.171.675.256.966.169.7
Jina-CLIP [18]16.531.241.39.224.830.33.77.312.8
Nomic-Embed-Vision [29]16.524.828.413.821.125.71.82.86.4
CLIP [33]46.865.169.744.058.765.123.941.345.9
SigLIP [45]51.467.976.247.763.370.633.049.557.8
OpenCLIP [16]58.775.279.856.970.675.234.949.557.8
V-RAG (ours)81.789.089.078.084.484.466.178.078.9
InfoHaystack-100InfoHaystack-200InfoHaystack-1000
R@1R@3R@5R@1R@3R@5R@1R@3R@5
BM25 (OCR)56.865.871.051.665.269.038.751.658.1
Jina-CLIP43.251.658.136.846.551.623.933.637.4
Nomic-Embed-Vision34.850.356.831.043.248.420.731.035.5
CLIP69.778.785.865.277.481.945.864.570.3
SigLIP58.171.680.055.567.776.839.455.561.9
OpenCLIP72.385.292.966.581.989.053.665.872.9
V-RAG (ours)79.491.092.974.888.488.464.574.278.1
", + "image_path": "9090873ad8c2d36b7f0e13db88bc070120c9449c5abb1c838cb7520be7e96248.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "table_body" + } + ], + "index": 0 + }, + { + "type": "table", + "bbox": [ + 70, + 312, + 282, + 434 + ], + "blocks": [ + { + "bbox": [ + 55, + 257, + 555, + 301 + ], + "lines": [ + { + "bbox": [ + 55, + 257, + 555, + 301 + ], + "spans": [ + { + "bbox": [ + 55, + 257, + 555, + 301 + ], + "type": "text", + "content": "Table 2. Retrieval Results. We compare our V-RAG model with other text-to-image and text-to-text (using OCR) retrieval methods across both benchmarks. V-RAG consistently outperforms baseline models on Recall@1, Recall@3, and Recall@5 metrics. Notably, V-RAG leverages an ensemble of text-to-image models along with a large multimodal model in a two-stage filtering approach. Top-performing values in each column are highlighted in bold." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 70, + 312, + 282, + 434 + ], + "lines": [ + { + "bbox": [ + 70, + 312, + 282, + 434 + ], + "spans": [ + { + "bbox": [ + 70, + 312, + 282, + 434 + ], + "type": "table", + "html": "
ModelDocHaystackInfoHaystack
10020010001002001000
LLaVA-OV [20]------
GPT-4o [30]27.523.9-23.920.0-
Gemini [1]50.548.6-29.021.9-
Qwen2-VL [41]41.312.8-20.014.2-
MIRAGE [43]3.73.72.87.77.16.5
LLaVA-OV+V-RAG69.765.155.143.241.936.8
GPT-4o+V-RAG81.772.567.065.263.256.8
Gemini+V-RAG73.465.158.757.457.447.1
Qwen2-VL+V-RAG82.674.366.165.865.860.0
Qwen2-VL-f.t.+V-RAG86.279.873.467.167.760.0
", + "image_path": "04643601e948bb0cd287df502a30280430ef1da242076566b2e46b5b73abc7a7.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "table_body" + } + ], + "index": 2 + }, + { + "bbox": [ + 55, + 441, + 295, + 519 + ], + "lines": [ + { + "bbox": [ + 55, + 441, + 295, + 519 + ], + "spans": [ + { + "bbox": [ + 55, + 441, + 295, + 519 + ], + "type": "text", + "content": "Table 3. The VQA results for the DocHaystack and Info-Haystack. We evaluate with many closed-source and open-source multimodal model, and also integrating them with our V-RAG retrieval framework. - denotes that those models can not be inferred due to their token context constraints. To enable GPT-4o and Qwen2-VL to process hundreds of images, we employ low-resolution mode and adjust image size for compatibility." + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "text" + }, + { + "bbox": [ + 55, + 525, + 207, + 538 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 525, + 207, + 538 + ], + "spans": [ + { + "bbox": [ + 55, + 525, + 207, + 538 + ], + "type": "text", + "content": "5.2. Main Experimental Results" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 55, + 544, + 295, + 616 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 544, + 295, + 616 + ], + "spans": [ + { + "bbox": [ + 55, + 544, + 295, + 616 + ], + "type": "text", + "content": "We evaluated a range of open-source and closed-source vision-language models for VQA tasks. We also evaluate several text-to-image and text-to-text (with OCR) retrieval models to evaluate their retrieval capabilities on our benchmarks. More detailed performance analysis are described in the following sections." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 55, + 617, + 295, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 617, + 295, + 713 + ], + "spans": [ + { + "bbox": [ + 55, + 617, + 295, + 713 + ], + "type": "text", + "content": "Retrieving results. The retrieval results in Table 2 demonstrate the superiority of our proposed V-RAG framework over several baseline methods across both DocHaystack and InfoHaystack benchmarks. V-RAG consistently achieves the highest Recall@1, Recall@3, and Recall@5 scores on most categories, indicating its robust retrieval capabilities. Notably, V-RAG outperforms text-based retrieving models such as BM25 and also the text-to-image retrieval models" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 313, + 314, + 555, + 494 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 314, + 555, + 494 + ], + "spans": [ + { + "bbox": [ + 313, + 314, + 555, + 494 + ], + "type": "text", + "content": "like jina-clip, CLIP, SigLIP, and OpenCLIP by substantial margins, especially on the DocHaystack-100 subset, where it reaches Recall@1 of " + }, + { + "bbox": [ + 313, + 314, + 555, + 494 + ], + "type": "inline_equation", + "content": "81.7\\%" + }, + { + "bbox": [ + 313, + 314, + 555, + 494 + ], + "type": "text", + "content": " and Recall@5 of " + }, + { + "bbox": [ + 313, + 314, + 555, + 494 + ], + "type": "inline_equation", + "content": "89.0\\%" + }, + { + "bbox": [ + 313, + 314, + 555, + 494 + ], + "type": "text", + "content": ". This pattern continues for larger datasets (DocHaystack-1000), where V-RAG remains competitive, achieving Recall@1 of " + }, + { + "bbox": [ + 313, + 314, + 555, + 494 + ], + "type": "inline_equation", + "content": "66.1\\%" + }, + { + "bbox": [ + 313, + 314, + 555, + 494 + ], + "type": "text", + "content": ". It achieves the top performance across all recall metrics on DocHaystack. For InfoHaystack benchmarks, V-RAG also outperforms other models, particularly on InfoHaystack-100 and InfoHaystack-200, where it receives Recall@1 of " + }, + { + "bbox": [ + 313, + 314, + 555, + 494 + ], + "type": "inline_equation", + "content": "79.4\\%" + }, + { + "bbox": [ + 313, + 314, + 555, + 494 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 313, + 314, + 555, + 494 + ], + "type": "inline_equation", + "content": "64.5\\%" + }, + { + "bbox": [ + 313, + 314, + 555, + 494 + ], + "type": "text", + "content": ", higher than previous best by around " + }, + { + "bbox": [ + 313, + 314, + 555, + 494 + ], + "type": "inline_equation", + "content": "7\\%" + }, + { + "bbox": [ + 313, + 314, + 555, + 494 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 313, + 314, + 555, + 494 + ], + "type": "inline_equation", + "content": "11\\%" + }, + { + "bbox": [ + 313, + 314, + 555, + 494 + ], + "type": "text", + "content": ", respectively. This consistent performance advantage highlights the effectiveness of V-RAG's ensemble of multiple vision encoders, allowing it to capture more granular details and improve retrieval accuracy over large multimodal models." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 313, + 498, + 556, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 498, + 556, + 714 + ], + "spans": [ + { + "bbox": [ + 313, + 498, + 556, + 714 + ], + "type": "text", + "content": "Visual question answering (VQA) results. The table presents VQA results for the DocHaystack and Info-Haystack benchmarks across varying dataset sizes (100, 200, 1000) using different multimodal models, both independently and in combination with the V-RAG framework. The results show that Qwen2-VL fine-tuned with V-RAG (Qwen2-VL-f.t.+V-RAG) achieves the highest scores across most benchmarks, with particularly notable performance on DocHaystack-100 (86.2) and InfoHaystack-100 (67.1), indicating superior retrieval and VQA capabilities in these scenarios. When V-RAG is added to other models, substantial improvements are observed, demonstrating the framework's efficacy in enhancing retrieval accuracy. For instance, GPT-4o's performance increases significantly with V-RAG, particularly for DocHaystack-100 and -200. The analysis highlights that V-RAG integration generally boosts performance across models, with Qwen2-VL-f.t.+V-RAG standing out as the top performer on both" + } + ] + } + ], + "index": 8 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 749, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 749, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 749, + 317, + 757 + ], + "type": "text", + "content": "24823" + } + ] + } + ], + "index": 9 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 6 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 61, + 74, + 171, + 175 + ], + "blocks": [ + { + "bbox": [ + 61, + 74, + 171, + 175 + ], + "lines": [ + { + "bbox": [ + 61, + 74, + 171, + 175 + ], + "spans": [ + { + "bbox": [ + 61, + 74, + 171, + 175 + ], + "type": "image", + "image_path": "1b5777be2eeaa49dfad5bd1c1f8a694dc890c097efba4cb37865a0cb29e32481.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 55, + 178, + 555, + 212 + ], + "lines": [ + { + "bbox": [ + 55, + 178, + 555, + 212 + ], + "spans": [ + { + "bbox": [ + 55, + 178, + 555, + 212 + ], + "type": "text", + "content": "Figure 5. Top-k selection ablation analysis for LMM-VQA. We demonstrate the results for LLaVA, Qwen2-VL, GPT-4o and also the finetuned Qwen2-VL model on the DocHaystack-100/1000 and InfoHaystack-100/1000 benchmarks. All the models are integrated with our V-RAG framework. We show the VQA accuracy performance for each ablation." + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "type": "image", + "bbox": [ + 184, + 74, + 294, + 174 + ], + "blocks": [ + { + "bbox": [ + 184, + 74, + 294, + 174 + ], + "lines": [ + { + "bbox": [ + 184, + 74, + 294, + 174 + ], + "spans": [ + { + "bbox": [ + 184, + 74, + 294, + 174 + ], + "type": "image", + "image_path": "e78914bb84f03c5adaa7cdab5f6063249e26ced941af318bf49a1fd6a8b8ec39.jpg" + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_body" + } + ], + "index": 1 + }, + { + "type": "image", + "bbox": [ + 309, + 74, + 417, + 174 + ], + "blocks": [ + { + "bbox": [ + 309, + 74, + 417, + 174 + ], + "lines": [ + { + "bbox": [ + 309, + 74, + 417, + 174 + ], + "spans": [ + { + "bbox": [ + 309, + 74, + 417, + 174 + ], + "type": "image", + "image_path": "86d3e9efa600a0d7ed27dbf45fec14a74066c047072ff13e29f06128ffd0f57b.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + } + ], + "index": 2 + }, + { + "type": "image", + "bbox": [ + 432, + 76, + 541, + 174 + ], + "blocks": [ + { + "bbox": [ + 432, + 76, + 541, + 174 + ], + "lines": [ + { + "bbox": [ + 432, + 76, + 541, + 174 + ], + "spans": [ + { + "bbox": [ + 432, + 76, + 541, + 174 + ], + "type": "image", + "image_path": "ccb77c5c7273b24a04c7501c4b31159868cfa7fb1c21e67e4e3b06d7f6ff10f6.jpg" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_body" + } + ], + "index": 3 + }, + { + "type": "table", + "bbox": [ + 160, + 221, + 451, + 289 + ], + "blocks": [ + { + "bbox": [ + 160, + 221, + 451, + 289 + ], + "lines": [ + { + "bbox": [ + 160, + 221, + 451, + 289 + ], + "spans": [ + { + "bbox": [ + 160, + 221, + 451, + 289 + ], + "type": "table", + "html": "
CLIPSigLIPOpenCLIPLMM-filterDocHaystack-1000InfoHaystack-1000
R@1R@3R@5R@1R@3R@5
✓XXX23.941.345.945.864.570.3
X✓XX33.049.557.839.455.561.9
XX✓X34.949.557.853.665.872.9
✓✓XX40.459.662.459.467.774.2
✓✓✓X42.266.177.556.171.078.1
✓✓✓✓66.178.078.964.574.278.1
", + "image_path": "f306ec03032315a38dd381fcfd77e8c626ffb9a62d45ca8984edf0912b087ab1.jpg" + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "table_body" + } + ], + "index": 5 + }, + { + "bbox": [ + 55, + 297, + 555, + 319 + ], + "lines": [ + { + "bbox": [ + 55, + 297, + 555, + 319 + ], + "spans": [ + { + "bbox": [ + 55, + 297, + 555, + 319 + ], + "type": "text", + "content": "Table 4. Ablation study on the V-RAG framework components. We quantify the impact of each module for the Recall@1, Recall@3 and Recall@5 retrieval performance on the DocHaystack-1000 and InfoHaystack-1000 for our V-RAG framework." + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "text" + }, + { + "bbox": [ + 55, + 332, + 295, + 392 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 332, + 295, + 392 + ], + "spans": [ + { + "bbox": [ + 55, + 332, + 295, + 392 + ], + "type": "text", + "content": "benchmarks, especially for the larger 1000-document tasks where retrieval accuracy is more challenging. This suggests that V-RAG's vision-centric, retrieval-augmented approach is highly effective for large-scale multimodal document understanding." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 55, + 393, + 296, + 464 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 393, + 296, + 464 + ], + "spans": [ + { + "bbox": [ + 55, + 393, + 296, + 464 + ], + "type": "text", + "content": "The table also shows that the DocHaystack-1000 and InfoHaystack-1000 present significant challenges for current LMMs. The drop in performance for larger document sets, with top accuracy only reaching " + }, + { + "bbox": [ + 55, + 393, + 296, + 464 + ], + "type": "inline_equation", + "content": "73.4\\%" + }, + { + "bbox": [ + 55, + 393, + 296, + 464 + ], + "type": "text", + "content": " for DocHaystack-1000 and " + }, + { + "bbox": [ + 55, + 393, + 296, + 464 + ], + "type": "inline_equation", + "content": "60.0\\%" + }, + { + "bbox": [ + 55, + 393, + 296, + 464 + ], + "type": "text", + "content": " for InfoHaystack-1000, underscores the difficulty our benchmarks." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 55, + 477, + 156, + 489 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 477, + 156, + 489 + ], + "spans": [ + { + "bbox": [ + 55, + 477, + 156, + 489 + ], + "type": "text", + "content": "5.3. Ablation Studies" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 55, + 496, + 295, + 640 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 496, + 295, + 640 + ], + "spans": [ + { + "bbox": [ + 55, + 496, + 295, + 640 + ], + "type": "text", + "content": "Ablation study on Top-k Selection. This figure presents the top-k selection ablation analysis for LMM-VQA across four models: LLaVA-OV, Qwen2-VL, GPT-4o, and the fine-tuned Qwen2-VL (Qwen2-VL-f.t.), evaluated on the DocHaystack-100/1000 and InfoHaystack-100/1000 benchmarks. The analysis reports VQA accuracy as a function of top-k selection (Top 1, Top 3, and Top 5). Overall, accuracy tends to improve with larger k-values, suggesting that offering more retrieval options positively impacts model performance. However, for LLaVA-OV, there is a marked decrease in performance at top-5, indicating that this model struggles to process multiple images at this scale." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 55, + 642, + 295, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 642, + 295, + 713 + ], + "spans": [ + { + "bbox": [ + 55, + 642, + 295, + 713 + ], + "type": "text", + "content": "Ablation study on the V-RAG framework components. The ablation study in Table 4 highlights the contributions of each component in the V-RAG framework on the DocHaystack-1000 and InfoHaystack-1000 benchmarks. Using CLIP alone yields low performance (e.g., Recall@1 of " + }, + { + "bbox": [ + 55, + 642, + 295, + 713 + ], + "type": "inline_equation", + "content": "23.9\\%" + }, + { + "bbox": [ + 55, + 642, + 295, + 713 + ], + "type": "text", + "content": " on DocHaystack-1000 and " + }, + { + "bbox": [ + 55, + 642, + 295, + 713 + ], + "type": "inline_equation", + "content": "45.8\\%" + }, + { + "bbox": [ + 55, + 642, + 295, + 713 + ], + "type": "text", + "content": " on" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 313, + 332, + 555, + 367 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 332, + 555, + 367 + ], + "spans": [ + { + "bbox": [ + 313, + 332, + 555, + 367 + ], + "type": "text", + "content": "InfoHaystack-1000), indicating its limited retrieval capability on its own. Adding SigLIP and OpenCLIP incrementally improves results." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 313, + 373, + 555, + 480 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 373, + 555, + 480 + ], + "spans": [ + { + "bbox": [ + 313, + 373, + 555, + 480 + ], + "type": "text", + "content": "The highest performance is achieved when all three encoders are combined with the VLM-filter module, leading to Recall@1 scores of " + }, + { + "bbox": [ + 313, + 373, + 555, + 480 + ], + "type": "inline_equation", + "content": "66.1\\%" + }, + { + "bbox": [ + 313, + 373, + 555, + 480 + ], + "type": "text", + "content": " on DocHaystack-1000 and " + }, + { + "bbox": [ + 313, + 373, + 555, + 480 + ], + "type": "inline_equation", + "content": "64.5\\%" + }, + { + "bbox": [ + 313, + 373, + 555, + 480 + ], + "type": "text", + "content": " on InfoHaystack-1000. This setup also achieves the top Recall@1, Recall@3 and Recall@5 values, demonstrating that the VLM-filter is essential for refining the ensemble outputs and significantly improving retrieval accuracy. These results confirm that each module contributes to VRAG's overall effectiveness." + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 314, + 508, + 388, + 521 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 314, + 508, + 388, + 521 + ], + "spans": [ + { + "bbox": [ + 314, + 508, + 388, + 521 + ], + "type": "text", + "content": "6. Conclusion" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 313, + 534, + 556, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 534, + 556, + 713 + ], + "spans": [ + { + "bbox": [ + 313, + 534, + 556, + 713 + ], + "type": "text", + "content": "In this work, we introduced the DocHaystack and Info-Haystack benchmarks to evaluate LMMs for retrieving and reasoning across large-scale documents. Our benchmarks providing a more rigorous and realistic assessment of large multimodal models in real-world, large-scale retrieval scenarios. To tackle these challenges, we proposed V-RAG, a vision-centric retrieval-augmented generation framework that significantly enhances retrieval precision and overall VQA performance. V-RAG achieves this through an ensemble of vision encoders and a specialized relevance filtering module, enabling improved accuracy across diverse visual inputs. Experimental results indicate that integrating V-RAG enables both open-source and closed-source LMMs to achieve superior performance in large-scale image retrieval and complex reasoning tasks." + } + ] + } + ], + "index": 15 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "text", + "content": "24824" + } + ] + } + ], + "index": 16 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 7 + }, + { + "para_blocks": [ + { + "bbox": [ + 56, + 72, + 115, + 83 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 72, + 115, + 83 + ], + "spans": [ + { + "bbox": [ + 56, + 72, + 115, + 83 + ], + "type": "text", + "content": "References" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 57, + 91, + 296, + 713 + ], + "type": "list", + "angle": 0, + "index": 12, + "blocks": [ + { + "bbox": [ + 61, + 91, + 296, + 123 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 61, + 91, + 296, + 123 + ], + "spans": [ + { + "bbox": [ + 61, + 91, + 296, + 123 + ], + "type": "text", + "content": "[1] Google AI. Gemini: Google's multimodal ai model. Google AI Research, 2024. https://fireflies.ai/blog/gemini-vs-gpt-4.2,3,7" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 61, + 125, + 296, + 169 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 61, + 125, + 296, + 169 + ], + "spans": [ + { + "bbox": [ + 61, + 125, + 296, + 169 + ], + "type": "text", + "content": "[2] Akari Asai, Zeqiu Wu, Yizhong Wang, Avirup Sil, and Hannaneh Hajishirzi. Self-rag: Learning to retrieve, generate, and critique through self-reflection. arXiv preprint arXiv:2310.11511, 2023. 3" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 62, + 171, + 296, + 236 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 171, + 296, + 236 + ], + "spans": [ + { + "bbox": [ + 62, + 171, + 296, + 236 + ], + "type": "text", + "content": "[3] Yifan Bai, Zhen Zhang, Yifan Zhang, Yuxuan Li, Yifan Zhang, Yifan Zhang, Yifan Zhang, Yifan Zhang, Yifan Zhang, Yifan Zhang, et al. Qwen-vl: A frontier vision-language model with larger-scale vision pre-training and aligned cross-modal instruction tuning. arXiv preprint arXiv:2310.06726, 2023. 2, 6" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 62, + 239, + 294, + 304 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 239, + 294, + 304 + ], + "spans": [ + { + "bbox": [ + 62, + 239, + 294, + 304 + ], + "type": "text", + "content": "[4] Yang Bai, Xinxing Xu, Yong Liu, Salman Khan, Fahad Khan, Wangmeng Zuo, Rick Siow Mong Goh, and ChunMei Feng. Sentence-level prompts benefit composed image retrieval. In Proceedings of the International Conference on Learning Representations (ICLR), 2024. Spotlight Presentation. 3" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 62, + 307, + 294, + 350 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 307, + 294, + 350 + ], + "spans": [ + { + "bbox": [ + 62, + 307, + 294, + 350 + ], + "type": "text", + "content": "[5] Aayush Bansal, Karan Sikka, Gaurav Sharma, and Rama Chellappa. Visual question answering on image sets. In Proceedings of the European Conference on Computer Vision (ECCV), pages 35-51, 2020. 2" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 62, + 353, + 294, + 407 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 353, + 294, + 407 + ], + "spans": [ + { + "bbox": [ + 62, + 353, + 294, + 407 + ], + "type": "text", + "content": "[6] Florian Bordes, Richard Yuanzhe Pang, Anurag Ajay, Alexander C Li, Adrien Bardes, Suzanne Petryk, Oscar Manas, Zhiqiu Lin, Anas Mahmoud, Bargav Jayaraman, et al. An introduction to vision-language modeling. arXiv preprint arXiv:2405.17247, 2024. 2" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 62, + 410, + 294, + 463 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 410, + 294, + 463 + ], + "spans": [ + { + "bbox": [ + 62, + 410, + 294, + 463 + ], + "type": "text", + "content": "[7] Yingshan Chang, Mridu Narang, Hisami Suzuki, Guihong Cao, Jianfeng Gao, and Yonatan Bisk. Webqa: Multihop and multimodal qa. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 14178-14188, 2022. 1, 2" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 62, + 466, + 294, + 531 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 466, + 294, + 531 + ], + "spans": [ + { + "bbox": [ + 62, + 466, + 294, + 531 + ], + "type": "text", + "content": "[8] Jun Chen, Deyao Zhu, Xiaogian Shen, Xiang Li, Zechun Liu, Pengchuan Zhang, Raghuraman Krishnamoorthi, Vikas Chandra, Yunyang Xiong, and Mohamed Elhoseiny. Minigpt-v2: Large language model as a unified interface for vision-language multi-task learning. arXiv preprint arXiv:2310.15339, 2023. 2, 3" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 62, + 534, + 294, + 609 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 534, + 294, + 609 + ], + "spans": [ + { + "bbox": [ + 62, + 534, + 294, + 609 + ], + "type": "text", + "content": "[9] Wenhu Chen, Hexiang Hu, Xi Chen, Pat Verga, and William W. Cohen. MuRAG: Multimodal retrieval-augmented generator for open question answering over images and text. In Proceedings of the 2022 Conference on Empirical Methods in Natural Language Processing (EMNLP), pages 5558-5570. Association for Computational Linguistics, 2022. 3" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 57, + 613, + 294, + 689 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 613, + 294, + 689 + ], + "spans": [ + { + "bbox": [ + 57, + 613, + 294, + 689 + ], + "type": "text", + "content": "[10] Jacob Devlin, Ming-Wei Chang, Kenton Lee, and Kristina Toutanova. Bert: Pre-training of deep bidirectional transformers for language understanding. In Proceedings of the 2019 Conference of the North American Chapter of the Association for Computational Linguistics: Human Language Technologies, pages 4171–4186. Association for Computational Linguistics, 2019. 3" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 57, + 691, + 294, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 691, + 294, + 713 + ], + "spans": [ + { + "bbox": [ + 57, + 691, + 294, + 713 + ], + "type": "text", + "content": "[11] Chun-Mei Feng, Yang Bai, Tao Luo, Zhen Li, Salman Khan, Wangmeng Zuo, Xinxing Xu, Rick Siow Mong Goh," + } + ] + } + ], + "index": 11 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 316, + 73, + 553, + 713 + ], + "type": "list", + "angle": 0, + "index": 25, + "blocks": [ + { + "bbox": [ + 333, + 73, + 553, + 105 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 333, + 73, + 553, + 105 + ], + "spans": [ + { + "bbox": [ + 333, + 73, + 553, + 105 + ], + "type": "text", + "content": "and Yong Liu. Vq4cir: Boosting composed image retrieval with visual question answering. arXiv preprint arXiv:2312.12273, 2023. 2, 3" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 317, + 106, + 553, + 160 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 106, + 553, + 160 + ], + "spans": [ + { + "bbox": [ + 317, + 106, + 553, + 160 + ], + "type": "text", + "content": "[12] Yash Goyal, Tejas Khot, Douglas Summers-Stay, Dhruv Batra, and Devi Parikh. Making the V in VQA matter: Elevating the role of image understanding in Visual Question Answering. In Conference on Computer Vision and Pattern Recognition (CVPR), 2017. 2, 3" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 317, + 162, + 553, + 215 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 162, + 553, + 215 + ], + "spans": [ + { + "bbox": [ + 317, + 162, + 553, + 215 + ], + "type": "text", + "content": "[13] Kelvin Guu, Kenton Lee, Zora Tung, Panupong Pasupat, and Ming-Wei Chang. Realm: Retrieval-augmented language model pre-training. In Proceedings of the 37th International Conference on Machine Learning (ICML), pages 3929-3938. PMLR, 2020. 3" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 316, + 217, + 553, + 258 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 217, + 553, + 258 + ], + "spans": [ + { + "bbox": [ + 316, + 217, + 553, + 258 + ], + "type": "text", + "content": "[14] Edward J. Hu, Yelong Shen, Phillip Wallis, Zeyuan Allen-Zhu, Yuanzhi Li, Shean Wang, Lu Wang, and Weizhu Chen. Lora: Low-rank adaptation of large language models, 2021. 6" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 316, + 261, + 553, + 314 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 261, + 553, + 314 + ], + "spans": [ + { + "bbox": [ + 316, + 261, + 553, + 314 + ], + "type": "text", + "content": "[15] Drew A Hudson and Christopher D Manning. Gqa: A new dataset for real-world visual reasoning and compositional question answering. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition, pages 6700-6709, 2019. 2" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 317, + 316, + 553, + 380 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 316, + 553, + 380 + ], + "spans": [ + { + "bbox": [ + 317, + 316, + 553, + 380 + ], + "type": "text", + "content": "[16] Gabriel Ilharco, Mitchell Wortsman, Ross Wightman, Cade Gordon, Nicholas Carlini, Rohan Taori, Achal Dave, Vaishaal Shankar, Hongseok Namkoong, John Miller, Hannaneh Hajishirzi, Ali Farhadi, and Ludwig Schmidt. Openclip. 2021. If you use this software, please cite it as below. 3, 6, 7" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 317, + 383, + 553, + 458 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 383, + 553, + 458 + ], + "spans": [ + { + "bbox": [ + 317, + 383, + 553, + 458 + ], + "type": "text", + "content": "[17] Vladimir Karpukhin, Barlas Oğuz, Sewon Min, Patrick Lewis, Ledell Wu, Sergey Edunov, Danqi Chen, and Wentau Yih. Dense passage retrieval for open-domain question answering. In Proceedings of the 2020 Conference on Empirical Methods in Natural Language Processing (EMNLP), pages 6769-6781. Association for Computational Linguistics, 2020. 3" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 317, + 460, + 553, + 524 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 460, + 553, + 524 + ], + "spans": [ + { + "bbox": [ + 317, + 460, + 553, + 524 + ], + "type": "text", + "content": "[18] Andreas Koukounas, Georgios Mastrapas, Michael Günther, Bo Wang, Scott Martens, Isabelle Mohr, Saba Sturua, Mohammad Kalim Akram, Joan Fontanals Martínez, Saahil Ognawala, Susana Guzman, Maximilian Werk, Nan Wang, and Han Xiao. Jina clip: Your clip model is also your text retriever, 2024. 6, 7" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 317, + 526, + 553, + 590 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 526, + 553, + 590 + ], + "spans": [ + { + "bbox": [ + 317, + 526, + 553, + 590 + ], + "type": "text", + "content": "[19] Ranjay Krishna, Yuke Zhu, Oliver Groth, Justin Johnson, Kenji Hata, Joshua Kravitz, Stephanie Chen, Yannis Kalantidis, Li-Jia Li, David A. Shamma, Michael S. Bernstein, and Li Fei-Fei. Visual genome: Connecting language and vision using crowdsourced dense image annotations. International Journal of Computer Vision, 123(1):32-73, 2017. 2" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 317, + 592, + 553, + 635 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 592, + 553, + 635 + ], + "spans": [ + { + "bbox": [ + 317, + 592, + 553, + 635 + ], + "type": "text", + "content": "[20] Bo Li, Yuanhan Zhang, Dong Guo, Renrui Zhang, Feng Li, Hao Zhang, Kaichen Zhang, Peiyuan Zhang, Yanwei Li, Ziwei Liu, and Chunyuan Li. Llava-onevision: Easy visual task transfer. arXiv preprint arXiv:2408.03326, 2024. 2, 3, 6, 7" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 317, + 636, + 553, + 689 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 636, + 553, + 689 + ], + "spans": [ + { + "bbox": [ + 317, + 636, + 553, + 689 + ], + "type": "text", + "content": "[21] Zhuang Liu, Hanzi Mao, Chao-Yuan Wu, Christoph Feichtenhofer, Trevor Darrell, and Saining Xie. A convnet for the 2020s. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 11976-11986, 2022. 6" + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 316, + 691, + 553, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 691, + 553, + 713 + ], + "spans": [ + { + "bbox": [ + 316, + 691, + 553, + 713 + ], + "type": "text", + "content": "[22] Pan Lu, Tony Xia, Weicheng Shi, Ahmed El Kholy, Xi Victor Lin, Jianfeng Gao, Xiang Chen, and Kai-Wei Chang. Learn" + } + ] + } + ], + "index": 24 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 749, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 749, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 749, + 317, + 757 + ], + "type": "text", + "content": "24825" + } + ] + } + ], + "index": 26 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 8 + }, + { + "para_blocks": [ + { + "bbox": [ + 56, + 72, + 295, + 713 + ], + "type": "list", + "angle": 0, + "index": 14, + "blocks": [ + { + "bbox": [ + 76, + 72, + 294, + 106 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 76, + 72, + 294, + 106 + ], + "spans": [ + { + "bbox": [ + 76, + 72, + 294, + 106 + ], + "type": "text", + "content": "to explain: Multimodal reasoning via thought chains for science question answering. In Advances in Neural Information Processing Systems, 2022. 2" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 56, + 106, + 295, + 162 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 106, + 295, + 162 + ], + "spans": [ + { + "bbox": [ + 56, + 106, + 295, + 162 + ], + "type": "text", + "content": "[23] Pan Lu, Hritik Bansal, Tony Xia, Jiacheng Liu, Chunyuan Li, Hannaneh Hajishirzi, Hao Cheng, Kai-Wei Chang, Michel Galley, and Jianfeng Gao. Mathvista: Evaluating mathematical reasoning of foundation models in visual contexts. arXiv preprint arXiv:2310.02255, 2023. 2, 3" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 56, + 162, + 294, + 205 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 162, + 294, + 205 + ], + "spans": [ + { + "bbox": [ + 56, + 162, + 294, + 205 + ], + "type": "text", + "content": "[24] Hongyin Luo, Yung-Sung Chuang, Yuan Gong, Tianhua Zhang, Yoon Kim, Xixin Wu, Danny Fox, Helen Meng, and James Glass. SAIL: Search-augmented instruction learning. arXiv preprint arXiv:2305.15225, 2023. 3" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 56, + 206, + 294, + 260 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 206, + 294, + 260 + ], + "spans": [ + { + "bbox": [ + 56, + 206, + 294, + 260 + ], + "type": "text", + "content": "[25] Kenneth Marino, Mohammad Rastegari, Ali Farhadi, and Roozbeh Mottaghi. Ok-vqa: A visual question answering benchmark requiring external knowledge. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition, pages 3195-3204, 2019. 2" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 56, + 261, + 294, + 316 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 261, + 294, + 316 + ], + "spans": [ + { + "bbox": [ + 56, + 261, + 294, + 316 + ], + "type": "text", + "content": "[26] Ahmed Masry, Do Xuan Long, Jia Qing Tan, Shafiq Joty, and Enamul Hoque. Chartqa: A benchmark for question answering about charts with visual and logical reasoning. In Findings of the Association for Computational Linguistics: ACL 2022, pages 2263-2279, 2022. 2" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 56, + 316, + 294, + 370 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 316, + 294, + 370 + ], + "spans": [ + { + "bbox": [ + 56, + 316, + 294, + 370 + ], + "type": "text", + "content": "[27] Minesh Mathew, Viraj Bagal, Ruben Pérez Tito, Dimosthenis Karatzas, Ernest Valveny, and CV Jawahar. Infographicvqa. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 4390-4399, 2021. 2, 3" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 56, + 371, + 294, + 415 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 371, + 294, + 415 + ], + "spans": [ + { + "bbox": [ + 56, + 371, + 294, + 415 + ], + "type": "text", + "content": "[28] Minesh Mathew, Dimosthenis Karatzas, and CV Jawahar. Docvqa: A dataset for vqa on document images. In Proceedings of the IEEE/CVF Winter Conference on Applications of Computer Vision, pages 2200-2209, 2021. 2, 3" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 56, + 415, + 294, + 437 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 415, + 294, + 437 + ], + "spans": [ + { + "bbox": [ + 56, + 415, + 294, + 437 + ], + "type": "text", + "content": "[29] Zach Nussbaum, Brandon Duderstadt, and Andriy Mulyar. Nomic embed vision: Expanding the latent space, 2024. 6, 7" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 56, + 437, + 294, + 470 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 437, + 294, + 470 + ], + "spans": [ + { + "bbox": [ + 56, + 437, + 294, + 470 + ], + "type": "text", + "content": "[30] OpenAI. Gpt-4o: Enhanced multimodal language model. OpenAI Research, 2024. https://openai.com/index/hello-gpt-4o/.2,3,4,6,7" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 56, + 471, + 294, + 504 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 471, + 294, + 504 + ], + "spans": [ + { + "bbox": [ + 56, + 471, + 294, + 504 + ], + "type": "text", + "content": "[31] OpenAI. Gpt-4v: Multimodal language model with vision capabilities. OpenAI Research, 2024. https://openai.com/index/gpt-4/.2" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 56, + 504, + 294, + 569 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 504, + 294, + 569 + ], + "spans": [ + { + "bbox": [ + 56, + 504, + 294, + 569 + ], + "type": "text", + "content": "[32] Abhirama Subramanyam Penamakuri, Manish Gupta, Mithun Das Gupta, and Anand Mishra. Answer mining from a pool of images: Towards retrieval-based visual question answering. In Proceedings of the International Joint Conference on Artificial Intelligence (IJCAI), pages 1052-1058, 2023. 1, 2, 3" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 56, + 570, + 294, + 645 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 570, + 294, + 645 + ], + "spans": [ + { + "bbox": [ + 56, + 570, + 294, + 645 + ], + "type": "text", + "content": "[33] Alec Radford, Jong Wook Kim, Karthik Hallacy, Aditya Ramesh, Gabriel Goh, Sandhini Agarwal, Girish Sastry, Amanda Askell, Pamela Mishkin, Jack Clark, et al. Learning transferable visual models from natural language supervision. In Proceedings of the 38th International Conference on Machine Learning, pages 8748-8763. PMLR, 2021. 3, 6, 7" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 56, + 647, + 294, + 691 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 647, + 294, + 691 + ], + "spans": [ + { + "bbox": [ + 56, + 647, + 294, + 691 + ], + "type": "text", + "content": "[34] Shaoqing Ren, Kaiming He, Ross Girshick, and Jian Sun. Faster r-cnn: Towards real-time object detection with region proposal networks. Advances in Neural Information Processing Systems, 28, 2015. 3" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 56, + 691, + 294, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 691, + 294, + 713 + ], + "spans": [ + { + "bbox": [ + 56, + 691, + 294, + 713 + ], + "type": "text", + "content": "[35] Dustin Schwenk, Apoorv Khandelwal, Christopher Clark, Kenneth Marino, and Roozbeh Mottaghi. A-okvqa: A" + } + ] + } + ], + "index": 13 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 316, + 73, + 553, + 709 + ], + "type": "list", + "angle": 0, + "index": 28, + "blocks": [ + { + "bbox": [ + 333, + 73, + 553, + 106 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 333, + 73, + 553, + 106 + ], + "spans": [ + { + "bbox": [ + 333, + 73, + 553, + 106 + ], + "type": "text", + "content": "benchmark for visual question answering using world knowledge. In European Conference on Computer Vision (ECCV), pages 1-17, 2022. 2" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 316, + 107, + 553, + 139 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 107, + 553, + 139 + ], + "spans": [ + { + "bbox": [ + 316, + 107, + 553, + 139 + ], + "type": "text", + "content": "[36] Ray Smith et al. TesseractOCR engine. https://github.com/tesseract-ocr/tesseract, 2024. Accessed: 2024-11-06. 4" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 316, + 141, + 553, + 195 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 141, + 553, + 195 + ], + "spans": [ + { + "bbox": [ + 316, + 141, + 553, + 195 + ], + "type": "text", + "content": "[37] Alon Talmor, Sewon Min, Robin Jia, Yanai Elazar, Uriel Singer Hasson, and Danqi Chen. Multimodalqa: Complex question answering over text, tables, and images. In Proceedings of the International Conference on Learning Representations (ICLR), 2021. 2" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 316, + 197, + 553, + 252 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 197, + 553, + 252 + ], + "spans": [ + { + "bbox": [ + 316, + 197, + 553, + 252 + ], + "type": "text", + "content": "[38] Ryota Tanaka, Kyosuke Nishida, Kosuke Nishida, Taku Hasegawa, Itsumi Saito, and Kuniko Saito. Slidevqa: A dataset for document visual question answering on multiple images. In Proceedings of the AAAI Conference on Artificial Intelligence, pages 13636-13645, 2023. 2" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 316, + 253, + 553, + 285 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 253, + 553, + 285 + ], + "spans": [ + { + "bbox": [ + 316, + 253, + 553, + 285 + ], + "type": "text", + "content": "[39] Ke Wang, Yichi Zhang, and Hongsheng Li. Mini-gemini: An efficient and versatile vision-language model. arXiv preprint arXiv:2310.12345, 2023. 2" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 316, + 286, + 553, + 329 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 286, + 553, + 329 + ], + "spans": [ + { + "bbox": [ + 316, + 286, + 553, + 329 + ], + "type": "text", + "content": "[40] Ke Wang, Junting Pan, Weikang Shi, Zimu Lu, Mingjie Zhan, and Hongsheng Li. Measuring multimodal mathematical reasoning with math-vision dataset. arXiv preprint arXiv:2402.14804, 2024. 2" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 316, + 331, + 553, + 385 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 331, + 553, + 385 + ], + "spans": [ + { + "bbox": [ + 316, + 331, + 553, + 385 + ], + "type": "text", + "content": "[41] Peng Wang, Shuai Bai, Sinan Tan, Shijie Wang, Zhihao Fan, Jinze Bai, Keqin Chen, Xuejing Liu, Jialin Wang, Wenbin Ge, et al. Qwen2-vl: Enhancing vision-language model's perception of the world at any resolution. arXiv preprint arXiv:2409.12191, 2024. 2, 3, 7" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 316, + 387, + 553, + 431 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 387, + 553, + 431 + ], + "spans": [ + { + "bbox": [ + 316, + 387, + 553, + 431 + ], + "type": "text", + "content": "[42] Weiyun Wang, Shuibio Zhang, Yiming Ren, Yuchen Duan, Tiantong Li, Shuo Liu, Mengkang Hu, Zhe Chen, Kaipeng Zhang, Lewei Lu, et al. Needle in a multimodal haystack. arXiv preprint arXiv:2406.07230, 2024. 2" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 316, + 432, + 553, + 485 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 432, + 553, + 485 + ], + "spans": [ + { + "bbox": [ + 316, + 432, + 553, + 485 + ], + "type": "text", + "content": "[43] Tsung-Han Wu, Giscard Biamby, Jerome Quenum, Ritwik Gupta, Joseph E. Gonzalez, Trevor Darrell, and David M. Chan. Visual haystacks: A vision-centric needle-in-a-haystack benchmark. arXiv preprint arXiv:2407.13766, 2024. 2, 3, 7" + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 316, + 487, + 553, + 542 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 487, + 553, + 542 + ], + "spans": [ + { + "bbox": [ + 316, + 487, + 553, + 542 + ], + "type": "text", + "content": "[44] Xiang Yue, Yuansheng Ni, Kai Zhang, Tianyu Zheng, Ruoqi Liu, Ge Zhang, Samuel Stevens, Dongfu Jiang, Weiming Ren, Yuxuan Sun, et al. Mmmu: A massive multi-discipline multimodal understanding and reasoning benchmark for expert agi. arXiv preprint arXiv:2311.16502, 2024. 2, 3" + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 316, + 544, + 553, + 597 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 544, + 553, + 597 + ], + "spans": [ + { + "bbox": [ + 316, + 544, + 553, + 597 + ], + "type": "text", + "content": "[45] Xiaohua Zhai, Basil Mustafa, Alexander Kolesnikov, and Lucas Beyer. Sigmoid loss for language image pre-training. In Proceedings of the IEEE/CVF International Conference on Computer Vision (ICCV), pages 11975-11985, 2023. 3, 6, 7" + } + ] + } + ], + "index": 25 + }, + { + "bbox": [ + 316, + 599, + 553, + 654 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 599, + 553, + 654 + ], + "spans": [ + { + "bbox": [ + 316, + 599, + 553, + 654 + ], + "type": "text", + "content": "[46] Renrui Zhang, Dongzhi Jiang, Yichi Zhang, Haokun Lin, Ziyu Guo, Pengshuo Qiu, Aojun Zhou, Pan Lu, Kai-Wei Chang, Peng Gao, and Hongsheng Li. Mathverse: Does your multi-modal llm truly see the diagrams in visual math problems? arXiv preprint arXiv:2403.14624, 2024. 2, 3" + } + ] + } + ], + "index": 26 + }, + { + "bbox": [ + 316, + 655, + 553, + 709 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 655, + 553, + 709 + ], + "spans": [ + { + "bbox": [ + 316, + 655, + 553, + 709 + ], + "type": "text", + "content": "[47] Deyao Zhu, Jun Chen, Xiaogian Shen, Xiang Li, and Mohamed Elhoseiny. Minigpt-4: Enhancing vision-language understanding with advanced large language models. In International Conference on Learning Representations (ICLR), 2024. 2, 3" + } + ] + } + ], + "index": 27 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 749, + 318, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 749, + 318, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 749, + 318, + 757 + ], + "type": "text", + "content": "24826" + } + ] + } + ], + "index": 29 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 9 + } + ], + "_backend": "vlm", + "_version_name": "2.6.4" +} \ No newline at end of file diff --git a/2025/Domain Adaptive Diabetic Retinopathy Grading with Model Absence and Flowing Data/8fc9ee64-90f9-4f33-8fb1-64a5f7b319a1_content_list.json b/2025/Domain Adaptive Diabetic Retinopathy Grading with Model Absence and Flowing Data/8fc9ee64-90f9-4f33-8fb1-64a5f7b319a1_content_list.json new file mode 100644 index 0000000000000000000000000000000000000000..cf02ad79f4c7ee51ba94411abbf680fdba07c00a --- /dev/null +++ b/2025/Domain Adaptive Diabetic Retinopathy Grading with Model Absence and Flowing Data/8fc9ee64-90f9-4f33-8fb1-64a5f7b319a1_content_list.json @@ -0,0 +1,1600 @@ +[ + { + "type": "text", + "text": "Domain Adaptive Diabetic Retinopathy Grading with Model Absence and Flowing Data", + "text_level": 1, + "bbox": [ + 194, + 130, + 803, + 174 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Wenxin $\\mathrm{Su}^1$ , Song Tang $^{*1,2,3}$ , Xiaofeng Liu $^4$ , Xiaojing Yi $^5$ , Mao Ye $^6$ , Chunxiao $\\mathrm{Zu}^1$ , Jiahao Li $^7$ , and Xiatian Zhu $^8$", + "bbox": [ + 228, + 214, + 769, + 251 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "1University of Shanghai for Science and Technology 2Universität Hamburg, 3ComOriginMat Inc. \n4Tale University 5Sichuan Eye Hospital 6University of Electronic Science and Technology of China \n7Chinese Academy of Medical Sciences & Peking Union Medical College 8University of Surrey", + "bbox": [ + 104, + 263, + 888, + 316 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "steventangsong@gmail.com", + "bbox": [ + 395, + 320, + 609, + 333 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Abstract", + "text_level": 1, + "bbox": [ + 248, + 378, + 326, + 393 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Domain shift (the difference between source and target domains) poses a significant challenge in clinical applications, e.g., Diabetic Retinopathy (DR) grading. Despite considering certain clinical requirements, like source data privacy, conventional transfer methods are predominantly model-centered and often struggle to prevent model-targeted attacks. In this paper, we address a challenging Online Model-aGnostic Domain Adaptation (OMG-DA) setting, driven by the demands of clinical environments. This setting is characterized by the absence of the model and the flow of target data. To tackle the new challenge, we propose a novel approach, Generative Unadversarial ExampleS (GUES), which enables adaptation from a data-centric perspective. Specifically, we first theoretically reformulate conventional perturbation optimization in a generative way—learning a perturbation generation function with a latent input variable. During model instantiation, we leverage a Variational AutoEncoder to express this function. The encoder with the reparameterization trick predicts the latent input, whilst the decoder is responsible for the generation. Furthermore, the saliency map is selected as pseudoperturbation labels. Because it not only captures potential lesions but also theoretically provides an upper bound on the function input, enabling the identification of the latent variable. Extensive experiments on DR benchmarks with both frozen pre-trained models and trainable models demonstrate the superiority of GUES, showing robustness even with small batch size. The source code and data are available at https://github.com/tntek/GUES.", + "bbox": [ + 89, + 412, + 485, + 851 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/6fbb60f19e595fa5edb64b48acb684f52e615b73cb2d7ef826c77943258bdfd7.jpg", + "image_caption": [ + "Figure 1. Comparison between the OMG-DA and SFDA settings. (a) In SFDA, adaptation builds upon the accumulated data, which demands significant storage resources in the hospital. Additionally, the models' architecture and parameters are accessible, exposing them to potential attacks. (b) OMG-DA provides a practical scenario: Flowing data mimic the patients' arrival in a stream way, and the models are unseen (strictly controlled) before using them, avoiding attacks like membership inference attacks [26]." + ], + "image_footnote": [], + "bbox": [ + 535, + 376, + 885, + 494 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "1. Introduction", + "text_level": 1, + "bbox": [ + 514, + 643, + 643, + 660 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Diabetic Retinopathy (DR) is a significant health concern, ranking among the leading causes of blindness and affecting millions of people worldwide [3]. Early-stage intervention for DR is crucial to preserve vision, highlighting the importance of timely diagnosis [27]. Although deep learning (DL) has demonstrated promising results in automating the grading of DR [6, 8, 36], deploying DL models in real-world clinical settings remains challenging. For example, DL models often struggle to generalize effectively to complex scenarios, such as variations in imaging equipment, ethnic groups, or temporal factors, leading to different data distributions, a challenge known as domain shift [11]. This issue significantly hampers the widespread adoption and success of DL-based diagnostic tools in clinical practice [14].", + "bbox": [ + 511, + 672, + 908, + 883 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Recently, many adaptation methods for grading diabetic", + "bbox": [ + 532, + 885, + 903, + 900 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "CVF", + "bbox": [ + 106, + 2, + 181, + 42 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore.", + "bbox": [ + 236, + 0, + 807, + 44 + ], + "page_idx": 0 + }, + { + "type": "page_footnote", + "text": "*Corresponding author", + "bbox": [ + 109, + 886, + 233, + 898 + ], + "page_idx": 0 + }, + { + "type": "page_number", + "text": "28337", + "bbox": [ + 478, + 944, + 517, + 955 + ], + "page_idx": 0 + }, + { + "type": "table", + "img_path": "images/68d886601922b4cab513baef77a9ffeadb139001fd3f2ab7e14a21ea98fc7f4e.jpg", + "table_caption": [ + "Table 1. Comparison between different transfer settings. Notation: source (s), target (t), data $\\mathbf{x}$ ,label $\\mathbf{y}$ ,loss function $L\\left( \\cdot \\right)$ ." + ], + "table_footnote": [], + "table_body": "
Setting nameModel absenceData flowSource data privacySource dataTarget dataTrain lossTest loss
Fine-tuning (FT)XX✓-xt,ytL(xt,yt)-
Unsupervised domain adaptation (UDA)XXXxs,ysxtL(xs,ys)+L(xt)-
Source-free domain adaptation (SFDA)XX✓-xtL(xt)-
Test-time adaptation (TTA)X✓✓-xt-L(xt)
Online model-agnostic domain adaptation (OMG-DA)✓✓✓-xt--
", + "bbox": [ + 107, + 114, + 888, + 191 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "retinopathy (DR) have focused on addressing the issue of domain shift [5, 20, 23, 40]. The initial classic transfer learning strategies, Unsupervised Domain Adaptation (UDA), required access to well-annotated source data. Nevertheless, the growing emphasis on privacy protection has shifted research toward the Source-Free Domain Adaptation (SFDA) framework [23, 40]. SFDA involves adapting a source model—pre-trained on the source domain—to the target domain in a self-supervised manner, thereby ensuring the protection of source patient data.", + "bbox": [ + 88, + 217, + 483, + 367 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "In recent developments, specific needs have emerged in the clinical field. The introduction of model weight-based techniques for reconstructing training data has created a demand for model privacy [26, 39], which goes beyond traditional source data protection. In addition, there is a growing requirement for models capable of handling incoming patient data in a flowing fashion (test data arriving sequentially), referred to as a flowing data constraint [33, 34]. Unfortunately, existing SFDA methods cannot effectively address this challenge, as they rely on full access to the model and require offline training on a pre-collected dataset. Fig. 1 provides an intuitive illustration of this issue.", + "bbox": [ + 88, + 369, + 483, + 547 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "In this paper, we consider a clinically motivated setting, called Online Model-aGnostic Domain Adaptation (OMG-DA) and propose a novel Generative Unadversarial ExampleS (GUES) approach for the DR grading problem in this new setting. Specifically, OMG-DA presents an extreme safety scenario: The available target data is unlabeled and arrives in a flowing format, with no prior information about the pre-trained source model and data. Tab. 1 provides a detailed comparison with previous adaptation settings.", + "bbox": [ + 88, + 551, + 483, + 686 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "In GUES, we address the absence of source data and pre-trained models by producing generalized unadversarial examples [24] for unlabeled target data. To this end, we introduce generative unadversarial learning, which theoretically reformulates conventional iterative perturbation optimization. This new method aims to learn a generative function for perturbations and involves addressing two key tasks: (1) Identifying the latent function input, which is the derivative of initial random noise w.r.t. image data, and (2) selecting a self-supervised property to serve as pseudo-perturbation labels. In practice, we leverage a Variational Autoencoder (VAE) [10] based approach to facilitate this learning process. In terms of function representation, we model the latent input using the encoder along with the reparameterization trick,", + "bbox": [ + 88, + 688, + 483, + 900 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "whilst the decoder accomplishes the generation of individual-perturbation. Additionally, we choose the saliency map as the pseudo-perturbation label for two reasons: (1) It helps discover potential lesions, and (2) it aids in identifying the latent input by providing an upper bound.", + "bbox": [ + 511, + 217, + 906, + 292 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Our contributions are summarized as follows:", + "bbox": [ + 529, + 292, + 841, + 306 + ], + "page_idx": 1 + }, + { + "type": "list", + "sub_type": "text", + "list_items": [ + "- Pioneering a novel transfer setting OMG-DA, which is closer to real-world clinical scenarios and meets three typical requirements at the same time, including (1) model absence, (2) flowing data, and (3) source data privacy.", + "- Developing a new OMG-DA approach GUES in the context of DR grading, grounded on the generative unadversarial examples theory, where we learn an individual-perturbation generative function under saliency map supervision, removing relying on labels and models.", + "- Extensive evaluations on four DR benchmarks, indicating that GUES can largely promote the source model's performance in the target domain, as well as trainable test-time adaptation models, even at small batch size." + ], + "bbox": [ + 511, + 309, + 906, + 503 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "2. Related work", + "text_level": 1, + "bbox": [ + 511, + 521, + 651, + 536 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Adaptation methods for DR grading. Driven by real-world medical requirements, domain adaptation has been an attractive topic in this DR grading issue. For instance, Nguyen et al. [20] introduce a UDA approach that enables the model to focus on vessel structures that remain invariant to domain shifts via image reconstruction using labeled source domain data. In SFDA, Zhang et al. [40] propose generating labeled, target-style retinal images to improve the source model's generalization, relying solely on a pre-trained source model and unlabeled target images. Additionally, DG in DR grading has been explored through domain-invariant feature learning approaches [4] and divergence-based methods [5], leveraging labeled source domain data.", + "bbox": [ + 511, + 547, + 906, + 743 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "The methods above rely on labeled data, require full access to the model, and necessitate offline training on a precollected dataset. In real clinical settings, these requirements can be impractical due to constraints around data and model privacy, as well as the need for real-time adaptability without retraining. In contrast, GUES provides an online adaptation solution that operates without requiring labeled data or access to the model, addressing domain adaptive DR grading.", + "bbox": [ + 511, + 744, + 908, + 866 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Unadversarial learning. Unadversarial learning was initially developed by Salman et al. [24], aiming to modify", + "bbox": [ + 511, + 869, + 908, + 901 + ], + "page_idx": 1 + }, + { + "type": "page_number", + "text": "28338", + "bbox": [ + 478, + 944, + 519, + 955 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "input image distribution to make them more easily recognizable by the model. Current mainstreams achieve this learning process by adding class-specific perturbations to the input images. Here, the perturbations are generated based on the gradient of an objective function w.r.t. image. This approach allows for the design of unadversarial examples without model training. For example, based on [24], NSA [25] introduces a method to generate more natural perturbations using a trainable generator. Similarly, CAT [18] demonstrates a new distance metric for generating unadversarial examples.", + "bbox": [ + 89, + 90, + 483, + 242 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "All existing unadversarial learning methods require access to model parameters, outputs, and labeled data. This dependency invalidates them in our OMG-DA setting which only has access to unlabeled target data. In addition to this, our GUES produces individualized unadversarial examples in a generative manner, which stands out from previous methods that focus on class-specific unadversarial examples.", + "bbox": [ + 89, + 243, + 483, + 349 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Saliency map for medical image. A fine-grained saliency map is a pixel feature generated by calculating the central surround differences within images, identifying salient regions without any need for training [19]. This feature is widely applied in various medical image analysis tasks to extract pathologically important regions [9, 22, 31, 35]. For instance, in DR grading, studies such as [9, 22] utilize saliency maps to guide models in focusing on critical features like the optic disc, cup, and vessel structures. Similarly, in brain tumor detection, Tomar et al. [31] leverage saliency maps to enhance the model's attention on tumor and bone structures. In skin cancer detection, saliency maps help isolate lesion regions with distinctive features, such as lumpiness, which are essential for accurate diagnosis [35].", + "bbox": [ + 89, + 354, + 483, + 565 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "As stated above, existing works primarily use saliency maps to highlight lesion regions. Unlike it, GUES selects saliency maps as pseudo-perturbation labels.", + "bbox": [ + 89, + 566, + 483, + 612 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3. Methodology", + "text_level": 1, + "bbox": [ + 89, + 630, + 225, + 648 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Problem statement. Given two different but related domains, i.e., source domain $\\mathcal{S}$ and target domain $\\mathcal{T}$ , $\\mathcal{S}$ contains $n_s$ labeled samples, while $\\mathcal{T}$ has $n$ unlabeled data. Both labeled and unlabeled samples share the same $C$ categories. Let $\\mathcal{X}_s$ and $\\mathcal{Y}_s$ be the source samples and the corresponding labels. Similarly, we denote the target samples and their labels by $\\mathcal{X}_t = \\{\\pmb{x}_t^i\\}_{i=1}^n$ and $\\mathcal{Y}_t = \\{y_t^i\\}_{i=1}^n$ , respectively, where $n$ signifies the number of samples. The source model $\\theta_s$ is pre-trained on $\\{\\mathcal{X}_s, \\mathcal{Y}_s\\}$ .", + "bbox": [ + 89, + 657, + 483, + 792 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "OMG-DA is featured in (1) the absence of the source model $\\theta_{s}$ and domain $S$ and (2) the flowing target data $\\mathcal{X}_t$ the same as the TTA setting [34]. Unlike previous transfer settings that are model adaptation-centered [15, 29, 30, 38], OMG-DA considers adaptation from the perspective of data. Specifically, OMG-DA aims to modify the distribution of target data to facilitate downstream tasks.", + "bbox": [ + 89, + 795, + 483, + 900 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3.1. Generative Unadversarial Examples", + "text_level": 1, + "bbox": [ + 511, + 90, + 828, + 107 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "The unadversarial learning idea [24] is well aligned with OMG-DA in terms of promoting adaptation by enhancing the generalization of input data. We briefly recap traditional unadversarial learning. Unlike adversarial learning [28] that generates confusing samples to mislead models, unadversarial learning aims to construct generalized samples, tackling out-of-distribution issues. Formally, this learning can be summarized in the following optimization problem.", + "bbox": [ + 511, + 112, + 906, + 234 + ], + "page_idx": 2 + }, + { + "type": "equation", + "text": "\n$$\n\\hat {\\delta} = \\arg \\min _ {\\delta} L \\left(f _ {\\theta} (x + \\delta), y\\right), s. t. | | \\delta | | \\leq \\epsilon \\tag {1}\n$$\n", + "text_format": "latex", + "bbox": [ + 566, + 244, + 906, + 268 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "where $L(\\cdot)$ denotes objective function, e.g., cross-entropy loss for classification tasks, $x$ and $y$ are input image and its label, $f_{\\theta}$ is a pre-trained model with parameters $\\theta$ , $\\delta$ is a perturbation, $\\epsilon$ is a small threshold. The current scheme solves this problem in an iterative way formulated as", + "bbox": [ + 511, + 277, + 905, + 354 + ], + "page_idx": 2 + }, + { + "type": "equation", + "text": "\n$$\n\\delta_ {k + 1} = \\delta_ {k} + \\alpha \\cdot \\operatorname {s i g n} \\left(\\nabla_ {x} L \\left(f _ {\\theta} \\left(x + \\delta_ {k}\\right), y\\right)\\right), k \\in [ 0, K - 1 ], \\tag {2}\n$$\n", + "text_format": "latex", + "bbox": [ + 519, + 362, + 905, + 391 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "where $\\alpha$ is a trade-off parameter, $K$ is iteration number, $\\delta_0$ is an initial random noise. In the inference phase, the optimal perturbation $\\hat{\\delta}$ is integrated into the input $x$ , forming an unadversarial example $\\hat{x} = x + \\hat{\\delta}$ , which is easily recognizable by the model $f_{\\theta}$ . Obviously, the conventional unadversarial paradigm cannot meet our OMG-DA setting due to the absence of $f_{\\theta}$ , $L$ , and the label $y$ .", + "bbox": [ + 511, + 391, + 905, + 496 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "In this paper, we re-consider the iterative optimization process above and obtain the theorem below (The proof is provided in Supplementary).", + "bbox": [ + 511, + 497, + 905, + 542 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Theorem 1 Given the unadversarial learning problem defined in Eq. (1), the iterative process featured by Eq. (2) can be expressed as the following generative form.", + "bbox": [ + 511, + 551, + 906, + 597 + ], + "page_idx": 2 + }, + { + "type": "equation", + "text": "\n$$\n\\delta_ {k} = \\delta_ {0} + V \\cdot F _ {\\Phi} \\left(\\frac {\\partial \\delta_ {0}}{\\partial x}\\right), \\tag {3}\n$$\n", + "text_format": "latex", + "bbox": [ + 616, + 606, + 906, + 640 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "where $\\delta_0$ is an initial random noise, $V > 0$ is a bound constant, $F_{\\Phi}$ is a generative function, $\\frac{\\partial\\delta_0}{\\partial x}$ is a latent variable.", + "bbox": [ + 511, + 648, + 906, + 681 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Grounded on Theorem 1, we have: When $\\delta_{k}$ converges to optimal $\\hat{\\delta}$ , i.e., $\\delta_{k} \\rightarrow \\hat{\\delta}$ , function $F_{\\Phi}$ also evolves to the optimal one, denoted by $\\hat{F}_{\\Phi}$ , i.e., $F_{\\Phi} \\rightarrow \\hat{F}_{\\Phi}$ . This provides an insight: The unadversarial learning problem above can also be solved in a generative fashion. Correspondingly, the generated data are termed generative unadversarial examples.", + "bbox": [ + 511, + 702, + 906, + 792 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3.2. Model Instantiation", + "text_level": 1, + "bbox": [ + 511, + 801, + 702, + 816 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Within this context of generative unadversarial learning, conventional unadversarial learning presented in Eq. (1) is boiled down to learning $F_{\\Phi}\\left(\\frac{\\partial\\delta_0}{\\partial x}\\right)$ . We can achieve this by training a generative neural network. To make this solution sense, we have to solve two difficulties as follows. (A) One is", + "bbox": [ + 511, + 825, + 905, + 900 + ], + "page_idx": 2 + }, + { + "type": "page_number", + "text": "28339", + "bbox": [ + 478, + 944, + 517, + 955 + ], + "page_idx": 2 + }, + { + "type": "image", + "img_path": "images/a7c3ea92141db9da1aa0dd614c6fc7178718c4c132cabf7b92e49b8adc76978e.jpg", + "image_caption": [ + "Figure 2. The instantiation framework of GUES in the OMG-DA setting. (a) For target input $x_{t}$ , the VAE model generates individual perturbation $\\delta_{t} = F_{\\Phi}\\left(\\frac{\\partial\\delta_{0}}{\\partial x}\\right)$ . After that, the by-pass path incorporates $\\delta_{t}$ and $x_{t}$ to create the generative unadversarial example $\\hat{x}_{t}$ . Treating $x_{t}$ 's saliency map $g_{t}$ as reconstruction supervision for model training. (b) At the inference phase, the generated unadversarial example $\\hat{x}_{t}$ is directly provided to the frozen source model or other trainable models." + ], + "image_footnote": [], + "bbox": [ + 133, + 88, + 867, + 305 + ], + "page_idx": 3 + }, + { + "type": "image", + "img_path": "images/df185912946c69dbba5c9887b58cd479347ea57e68b6d91b7216902aed39df4b.jpg", + "image_caption": [ + "Figure 3. Explanation in choosing fine-grained saliency maps as supervision. Left: The testing fundus image selected from \"Moderate DR\" class in APTOS demonstrates that H (hemorrhages), SE (soft exudates), and EX (hard exudates) are essential characteristics to judge the DR grade. Middle: The saliency map highlights those lesions. Right: The gradient-CAM visualization of the source model on task DDR→APTOS." + ], + "image_footnote": [], + "bbox": [ + 112, + 396, + 460, + 501 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "the identification of $\\frac{\\partial\\delta_0}{\\partial x}$ when the relationship between them is unknown. (B) The other is selecting property supervision (pseudo-perturbation labels) to drive $\\delta_{k}\\rightarrow \\hat{\\delta}$ .", + "bbox": [ + 89, + 619, + 483, + 667 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Solution to Problem A. In practice, considering derivative $\\frac{\\partial\\delta_0}{\\partial x}$ is relevant with both random noise $\\delta_0$ and input image $x$ , we sample it from a certain Gaussian distribution associated with $x$ . Furthermore, the output size of $F_{\\Phi}$ is the same as $x$ . Therefore, we employ the VAE model to jointly model $\\frac{\\partial\\delta_0}{\\partial x}$ and $F_{\\Phi}$ , since VAE is an autoencoder characterized by random sampling. Specifically, as shown in Fig.2 (a), we approximate $\\frac{\\partial\\delta_0}{\\partial x}$ by latent variable $z$ , which is jointly determined by input $x$ and sampled random signal $\\tau$ . As for $F_{\\Phi}$ , it is demonstrated by the decoder module. Suppose $D(\\cdot)$ is the decoder, $E_{\\tau}(\\cdot)$ is the encoder with reparameterization trick, our scheme can be formulated as", + "bbox": [ + 89, + 670, + 483, + 851 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\nD \\left(E _ {\\tau} (x)\\right)\\rightarrow F _ {\\Phi} \\left(\\frac {\\partial \\delta_ {0}}{\\partial x}\\right), E _ {\\tau} (x) \\rightarrow \\frac {\\partial \\delta_ {0}}{\\partial x}, D (\\cdot) \\rightarrow F _ {\\Phi} (\\cdot). \\tag {4}\n$$\n", + "text_format": "latex", + "bbox": [ + 102, + 858, + 483, + 900 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Solution to Problem B. We adopt the fine-grained saliency map as the supervision. Two reasons contribute to our selection. First of all, the empirical results show that, for the specific task of DR grading, the saliency map is an acceptable pseudo-perturbation. Specifically, perturbation in the unadversarial context enhances the regions associated with the category and reduces the prominence of other areas, thereby identifying the lesion zones relevant to DR grading. As illustrated in Fig. 3, the saliency map effectively identifies potential lesions, such as hemorrhages, soft exudates, and hard exudates. Furthermore, it includes gradient information, as it highlights regions similar to Grad-CAM (Right side). More importantly, we have the theorem below. (The proof is provided in Supplementary)", + "bbox": [ + 511, + 398, + 906, + 611 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Theorem 2 Given the partial derivatives of the initial random noise $\\delta_0$ w.r.t image $x$ is $\\frac{\\partial\\delta_0}{\\partial x}$ and $x$ 's saliency map is $s = G(x)$ where $G(\\cdot)$ is the computation function of the saliency map. We have the following relationship:", + "bbox": [ + 511, + 621, + 906, + 681 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\frac {\\partial \\delta_ {0}}{\\partial x} \\leq U \\cdot s, \\tag {5}\n$$\n", + "text_format": "latex", + "bbox": [ + 663, + 691, + 906, + 722 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "where $U > 0$ is a bound constant.", + "bbox": [ + 511, + 732, + 740, + 746 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Theorem 2 suggests that the saliency map provides upper bounds for $\\frac{\\partial\\delta_0}{\\partial x}$ . Namely, $s$ provides relaxed descriptions for the variation $\\frac{\\partial\\delta_0}{\\partial x}$ . This can help guide the learning of $\\frac{\\partial\\delta_0}{\\partial x}$ .", + "bbox": [ + 511, + 757, + 906, + 806 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "GUES framework. Based on the analysis above, we instantiate GUES as the framework depicted in Fig. 2. As shown in sub-figure (a), our method integrates a VAE model and by-pass connection, achieving the learning of $F_{\\Phi}\\left(\\frac{\\partial \\delta_0}{\\partial x}\\right)$ . Specifically, $\\frac{\\partial \\delta_0}{\\partial x}$ is sampled from an input $x_t$ -featured Gaussian distribution $N(\\mu, \\sigma)$ , which is jointly learned using the", + "bbox": [ + 511, + 806, + 908, + 901 + ], + "page_idx": 3 + }, + { + "type": "page_number", + "text": "28340", + "bbox": [ + 478, + 944, + 519, + 955 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Algorithm 1 The pipeline of proposed GUES", + "text_level": 1, + "bbox": [ + 91, + 90, + 393, + 104 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Input: Online batch samples $\\mathcal{B}$ , a trainable $\\mathrm{VAE}\\theta_v$ consists of an encoder $E_{\\tau}(\\cdot)$ and a decoder $D(\\cdot)$ .", + "bbox": [ + 89, + 107, + 480, + 137 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Procedure:", + "text_level": 1, + "bbox": [ + 91, + 138, + 171, + 151 + ], + "page_idx": 4 + }, + { + "type": "list", + "sub_type": "text", + "list_items": [ + "1: for $x_{i}$ in $\\mathcal{B}$ do", + "2: Approximate the latent function input by $E_{\\tau}(x_i)$ ;", + "3: Learn a generative function for perturbation $\\delta_{i}$ by $D$ ;", + "4: Calculate individual perturbations $\\delta_{i}$ by $D(E_{\\tau}(x))$", + "5: Create the unadversarial example $\\hat{x}_i$ by incorporating $\\delta_i$ and $x_i$ through the bypass path.", + "6: Generate a fine-grained saliency map $g_{i}$ of the $x_{i}$ ;", + "7: Update $\\theta_v$ with Eq.(8), taking $g_t$ as a supervision." + ], + "bbox": [ + 101, + 155, + 480, + 276 + ], + "page_idx": 4 + }, + { + "type": "list", + "sub_type": "text", + "list_items": [ + "8: end for", + "9: return The generative unadversarial examples $\\hat{x}$" + ], + "bbox": [ + 101, + 277, + 437, + 306 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "encoder and reparameterization. The decoder, representing $F_{\\Phi}$ , then transforms $\\frac{\\partial \\delta_0}{\\partial x}$ into a generative perturbation $\\delta_t$ . Finally, the by-pass structure incorporates $x_t$ and $\\delta_t$ to produce $\\hat{x}_t$ . As shown in Fig. 2 (b), during the inference phase, the trained GUES model generates the corresponding generative unadversarial examples for the unseen frozen or fine-tuned model from specific testing samples.", + "bbox": [ + 89, + 339, + 483, + 445 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Loss function. The loss function for GUES training consists of two components. First, we enforce the latent space with mean $\\mu$ and variance $\\sigma$ satisfy the standard normal distribution $\\mathcal{N}(0,I)$ . Suppose the encoder in VAE models the posterior distribution $q(z|x_{t}) = \\mathcal{N}(\\mu (x_{t}),\\sigma^{2}(x_{t}))$ , this regularization can be formulated as:", + "bbox": [ + 89, + 448, + 483, + 540 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\nL _ {\\mathrm {K L}} = D _ {\\mathrm {K L}} \\left(q (z | x _ {t}) \\| \\mathcal {N} (0, I)\\right), \\tag {6}\n$$\n", + "text_format": "latex", + "bbox": [ + 166, + 553, + 483, + 570 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "where function $D_{\\mathrm{KL}}$ computes the Kullback-Leibler divergence. The other reconstruction loss between the unadversarial example $\\hat{x}_t$ and saliency map $g_t$ is presented by the following regression form:", + "bbox": [ + 89, + 583, + 483, + 643 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\nL _ {\\mathrm {M S E}} = \\left\\| \\hat {x} _ {t} - g _ {t} \\right\\| _ {2}. \\tag {7}\n$$\n", + "text_format": "latex", + "bbox": [ + 207, + 657, + 482, + 672 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Formally, combining Eq. (6) and Eq. (7), the final objective of GUES can be summarized as:", + "bbox": [ + 89, + 686, + 482, + 715 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\nL _ {\\text {G U E S}} = \\alpha L _ {\\mathrm {K L}} + \\beta L _ {\\mathrm {M S E}}, \\tag {8}\n$$\n", + "text_format": "latex", + "bbox": [ + 197, + 731, + 482, + 746 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "where $\\alpha$ and $\\beta$ are trade-off parameters. For clarity, we summarize the training procedure of GUES in Algorithm 1.", + "bbox": [ + 89, + 760, + 482, + 789 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "In Eq. (8), the first item $L_{\\mathrm{KL}}$ ensures the learning of $\\frac{\\partial \\delta_0}{\\partial x}$ . On the one hand, as aforementioned, we use $z$ to present $\\frac{\\partial \\delta_0}{\\partial x}$ . $L_{\\mathrm{KL}}$ aligns the $z$ space with $\\mathcal{N}(0, I)$ , thereby linking the random noise to $\\frac{\\partial \\delta_0}{\\partial x}$ . On the other hand, $q(z | x_t)$ in $L_{\\mathrm{KL}}$ is a function of input $x_t$ , building relationship $x_t$ to $\\frac{\\partial \\delta_0}{\\partial x}$ . Additionally, the reconstruction regulated by the seconded item $L_{\\mathrm{MSE}}$ encourage $\\delta_k \\rightarrow \\hat{\\delta}$ .", + "bbox": [ + 89, + 790, + 483, + 901 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "4. Experiments", + "text_level": 1, + "bbox": [ + 513, + 89, + 643, + 106 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "4.1. Datasets", + "text_level": 1, + "bbox": [ + 513, + 114, + 614, + 128 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "We perform evaluation experiments on four existing fundus benchmarks, including APTOS [1], DDR [13], DeepDR [17], and Messidor-2 (termed MD2) [7]. Those datasets share five grading/classes: no DR, mild DR, moderate DR, severe DR, and proliferative DR. Taking each dataset as a separate domain, we form 12 transfer tasks crossing domains. For example, as APTOS is the source domain while the others are target domains, we have three transfer tasks APTOS $\\rightarrow$ DDR, APTOS $\\rightarrow$ DeepDR, and APTOS $\\rightarrow$ MD2. The illustration of the label distribution and the domain shift of them is demonstrated in Supplementary. Of Note, all datasets have a severe class imbalance (e.g., \"no DR\" class itself takes up to $45.8\\%$ of the DDR dataset).", + "bbox": [ + 511, + 138, + 906, + 335 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "4.2. Implementation Detail", + "text_level": 1, + "bbox": [ + 513, + 345, + 723, + 362 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Source model pre-training. We adopt the DeiT-base network [32] as the backbone of the source pre-trained model, training it in a supervised manner using the source data and corresponding ground truths. During this source training phase, the adopted objective is the classic cross-entropy loss with label smoothing, the same as other methods [15, 37, 38].", + "bbox": [ + 511, + 373, + 906, + 464 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Variational autoencoder setting. The VAE model is an eight-layer convolutional architecture with a latent space dimension of 10. We do not employ a pre-trained VAE and utilize a VAE without fine-tuning on any other dataset, ensuring that the learning component $F_{\\Phi}\\left(\\frac{\\partial\\delta_0}{\\partial x}\\right)$ are unbiased and independent of prior pre-training data.", + "bbox": [ + 511, + 468, + 905, + 558 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Parameter setting. For the trade-off parameters in Eq. (8), we set $\\alpha$ to 1.0, while $\\beta$ is tuned with $\\{0.0001, 0.01, 1\\}$ to ensure that $L_{\\mathrm{KL}}$ and $L_{\\mathrm{MSE}}$ remain on the same values scale.", + "bbox": [ + 511, + 559, + 906, + 604 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Training setting. For fairness, all compared methods were implemented with the same DeiT-base backbone. We adopt the batch size of 64, SGD optimizer with a momentum of 0.9 and a learning rate of 1e-5 on all datasets. All experiments are conducted with PyTorch on a single GPU of RTX A6000.", + "bbox": [ + 511, + 604, + 905, + 680 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "4.3. Comparison Settings", + "text_level": 1, + "bbox": [ + 513, + 691, + 710, + 707 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Evaluation metrics. To account for unbalanced datasets, besides conventional classification accuracy (termed ACC), we adopt the measure of Quadratic Weighted Kappa (termed QWK) [2] and the average of QWK and ACC (termed AVG). Their rules are provided in Supplementary.", + "bbox": [ + 511, + 715, + 906, + 791 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Competitors. We compare GUES with nine existing state-of-the-art adaptation methods divided into three groups. (1) The first group involves applying the source model directly to the target domain. (2) The second group includes five SFDA methods SHOT [15], NRC [38], CoWA [12], PLUE [16], and TPDS [29]. (3) The third group comprises three typical TTA methods: SHOT-IM [15], TENT [34], and SAR [21].", + "bbox": [ + 511, + 794, + 906, + 900 + ], + "page_idx": 4 + }, + { + "type": "page_number", + "text": "28341", + "bbox": [ + 478, + 944, + 517, + 955 + ], + "page_idx": 4 + }, + { + "type": "table", + "img_path": "images/4c12d1275a6d2cf5a7d683991f743a407c371d398fa7c7e8cb48a095c07bb756.jpg", + "table_caption": [ + "Table 2. The results of Source, SFDA, TTA, OMG-DA, and OMG-DA combination methods on datasets APTOS, DeepDR, DDR, and MD2 are presented. The improvements over baseline methods Source, SHOT-IM, and TENT are highlighted as (+x.x)." + ], + "table_footnote": [], + "table_body": "
MethodVenueAPTOS→DDRAPTOS→DeepDRAPTOS→MD2DDR→APTOSDDR→DeepDRDDR→MD2DeepDR→APTOS
ACCQWKAVGACCQWKAVGACCQWKAVGACCQWKAVGACCQWKAVGACCQWKAVGACCQWKAVG
Source-60.659.259.952.671.762.160.948.754.865.672.969.345.060.652.849.434.642.043.471.657.5
GUES + Source-62.059.560.853.069.761.359.846.753.376.081.878.956.468.762.559.147.653.346.574.160.3
SHOT [15]ICML2066.969.067.953.673.563.651.738.044.877.084.280.659.274.666.957.143.150.162.382.572.4
NRC [38]NeurIPS2161.965.263.551.670.961.354.140.347.260.376.368.352.069.160.550.635.242.952.074.363.2
CoWA [12]ICML2259.064.962.051.070.760.853.037.345.157.274.565.850.166.858.453.139.646.350.473.061.7
PLUE [16]CVPR2362.065.163.651.369.760.554.541.147.863.464.263.854.364.859.551.628.239.954.670.562.6
TPDS [29]IICV2466.667.867.251.671.561.552.740.746.776.683.980.258.073.165.654.541.047.860.880.070.4
SHOT-IM [15]ICML2066.569.267.952.673.663.153.236.644.975.982.179.058.573.966.257.343.550.461.984.072.9
TENT [34]ICLR2059.950.255.153.170.161.661.448.454.975.282.478.855.168.461.760.850.555.660.279.769.9
SAR [21]ICLR2367.963.865.853.673.063.357.244.951.175.683.579.555.671.663.649.237.043.159.579.369.4
GUES + SHOT-IM - GUES + TENT -68.668.568.553.572.863.255.743.849.777.283.180.260.575.167.861.551.256.362.683.272.9
-61.856.359.053.270.061.661.147.254.175.983.079.458.770.864.763.353.858.654.977.666.3
MethodVenueDeepDR→DDRDeepDR→MD2MD2→APTOSMD2→DDRMD2→DeepDRAccAvg. QWKAVG
ACCQWKAVGACCQWKAVGACCQWKAVGACCQWKAVGACCQWKAVG
Source-56.466.961.748.750.249.443.970.357.160.256.558.359.858.459.153.960.157.0
GUES + Source-57.365.661.448.352.050.159.376.267.764.755.860.258.657.157.858.4 (+4.5)62.9 (+2.8)60.7 (+3.7)
SHOT [15]ICML2057.471.264.348.841.545.252.773.062.854.659.256.959.670.264.958.465.061.7
NRC [38]NeurIPS2144.960.952.949.841.845.848.869.158.952.852.852.858.062.860.453.159.956.5
CoWA [12]ICML2248.758.453.649.942.746.351.070.160.549.650.950.357.660.659.156.961.959.4
PLUE [16]CVPR2347.253.550.456.447.451.956.069.162.656.554.355.458.864.861.855.557.756.6
TPDS [29]IICV2459.369.464.350.542.446.460.374.967.660.060.460.258.963.060.959.264.061.6
SHOT-IM [15]ICML2054.669.462.051.238.244.761.677.969.757.058.757.957.569.863.759.064.761.9
TENT [34]ICLR2058.545.451.958.356.557.455.174.164.655.831.743.758.053.655.859.359.259.3
SAR [21]ICLR2353.066.359.642.633.137.955.273.064.149.748.349.056.765.861.356.361.659.0
GUES + SHOT-IM - GUES + TENT --62.371.566.952.847.850.362.878.170.466.059.462.760.668.664.662.0 (+3.0)66.9 (+2.2)64.5 (+2.6)
-62.661.462.059.157.458.259.375.667.564.050.557.258.356.057.161.0 (+1.7)63.3 (+4.1)62.2 (+2.9)
", + "bbox": [ + 112, + 128, + 883, + 465 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/63765010acc15e3c61df0b9bf3ce8559686a07f425f19d2b0dc2843245c0ab1e.jpg", + "image_caption": [ + "Figure 4. Comparison results with batch size varying from 2 to 64 over the 12 tasks (The details are provided in Supplementary.). Left, middle, and right report ACC, QWK, and AVG, respectively." + ], + "image_footnote": [], + "bbox": [ + 98, + 488, + 475, + 583 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Comparison protocol in OMG-DA setting. For a comprehensive comparison, our comparison follows two different fashions as follows.", + "bbox": [ + 89, + 665, + 485, + 710 + ], + "page_idx": 5 + }, + { + "type": "list", + "sub_type": "text", + "list_items": [ + "- Case without training: We first generate the unadversarial examples for the target domain by the trained GUES model and then provide them to the frozen source model.", + "- Case with training: We plug GUES into other TTA methods (they are also online methods with flowing data) as online image pre-processing." + ], + "bbox": [ + 89, + 713, + 482, + 804 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "The two cases evaluate GUES from different aspects. The first isolates the generalization ability of the unadversarial examples generated by GUES, whilst the second highlights GUES's compatibility with other trainable online schemes.", + "bbox": [ + 89, + 806, + 482, + 867 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Corresponding to the comparison protocols above, besides the version GUES corresponding to the case with", + "bbox": [ + 89, + 869, + 485, + 900 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "out training, we also introduce GUES+SHOT-IM and GUES+TENT which correspond to the case with training.", + "bbox": [ + 511, + 489, + 906, + 521 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "4.4. Comparison Results", + "text_level": 1, + "bbox": [ + 511, + 532, + 705, + 550 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "This part presents the comparison results following the cases above. Also, considering batch size a crucial factor for TTA methods, the results as the batch size varies are provided.", + "bbox": [ + 511, + 556, + 906, + 603 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Results without training. The comparisons are shown in Tab. 2. On average, across the 12 tasks and without training of the source model, GUES achieves improvements of $4.5\\%$ in ACC, $2.8\\%$ in QWK, and $3.7\\%$ in AVG compared to the source model. These results demonstrate that GUES modifies the target data distribution effectively, adapting the target domain to align with the source domain.", + "bbox": [ + 511, + 607, + 905, + 713 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Results with training. As shown in Tab. 2, GUES+SHOT-IM outperforms the previous best SFDA and TTA methods, respectively surpassing TENT in ACC by $2.7\\%$ , SHOT in QWK by $1.9\\%$ , and SHOT-IM in AVG by $2.6\\%$ . Compared to SHOT-IM, GUES+SHOT-IM gains over $3.0\\%$ in ACC, $2.2\\%$ in QWK, and $2.6\\%$ in AVG. Similarly, GUES+TENT improves over TENT by $1.7\\%$ in ACC, $4.1\\%$ in QWK, and $2.9\\%$ in AVG. These results highlight the effectiveness of combining GUES with other methods that require training.", + "bbox": [ + 511, + 717, + 906, + 853 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Results with varying batch size. This part isolates the effect of batch size, which is a crucial factor for TTA methods. Fig. 4 depicts the performance variation as batch size vary", + "bbox": [ + 511, + 854, + 908, + 902 + ], + "page_idx": 5 + }, + { + "type": "page_number", + "text": "28342", + "bbox": [ + 478, + 944, + 519, + 955 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/373dc7e04a1dccbc02cde73710d605ee9811e1e2bb7be310782249f961c2d3d8.jpg", + "image_caption": [ + "Figure 5. Interpretability analysis based on a typical fundus image from \"Moderate DR\" class in APTOS. Here, H (hemorrhages), SE (soft exudates), and EX (hard exudates) are essential characteristics to judge the DR grade. The gradient CAM-based heatmap of five models visualizes the capture of those lesions. All models are trained on task DDR $\\rightarrow$ APTOS, where Oracle is trained using ground truth in APTOS." + ], + "image_footnote": [], + "bbox": [ + 133, + 88, + 867, + 198 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/7dc39ff85d07b737b9df1a1ce46a1b53f0d0d32651c72c934caa98923545f3cf.jpg", + "image_caption": [ + "Figure 6. Feature distribution comparison of 3D density charts on task DDR $\\rightarrow$ APTOS. Oracle is trained on APTOS by ground truth." + ], + "image_footnote": [], + "bbox": [ + 101, + 273, + 472, + 369 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "ing from 2 to 64 over the 12 tasks. It is observed that TTA methods SHOT-IM and TENT suffer from severe performance drops when the batch size becomes small. SHOT-IM exhibits a decrease of approximately $16\\%$ in ACC when the batch size is reduced from 64 to 2, whilst TENT shows a substantial decline of around $34\\%$ in QWK. Oppose to it, the methods with GUES, SHOT-IM+GUES, and TENT+GUES, do not have evident performance decline at the smaller batch size. Moreover, this combination not only mitigates the drop but also shows improvements when the batch size is 64. This indicates that GUES effectively stabilizes the performance of SHOT-IM and TENT, enhancing their robustness to variations in batch size while boosting their overall effectiveness.", + "bbox": [ + 88, + 430, + 482, + 626 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "We attribute GUES's excellent robustness to its ability to predict individual perturbations (see Supplementary for more details on the visualization of perturbations) that focus on single image-specific features rather than global data commonalities. For instance, the conventional unadversarial examples approach refines the class-specific perturbation sensitive to batch size. Furthermore, both SHOT-IM and TENT are entropy-based methods that require large-scale batch size for accurate entropy estimation.", + "bbox": [ + 89, + 627, + 482, + 762 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "4.5. Visualization Analysis", + "text_level": 1, + "bbox": [ + 89, + 771, + 297, + 787 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Interpretability. For a better understanding, Fig. 5 demonstrates whether GUES can help capture pathologically relevant features, such as H (hemorrhages), SE (soft exudates), and EX (hard exudates), determining DR grade. First of all, when comparing the source model with GUES, the source model only captures a limited area of the lesion, while GUES effectively captures most of the DR-related features. Further-", + "bbox": [ + 89, + 794, + 485, + 900 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/860379c907f692a288f195ca59110b8bb3e73479d954ca982ac69eb7b89547a2.jpg", + "image_caption": [ + "Figure 7. Unadversarial examples visualization of two typical target samples from APTOS. The generative perturbations are generated by the GUES model trained on task DDR $\\rightarrow$ APTOS." + ], + "image_footnote": [], + "bbox": [ + 517, + 275, + 903, + 406 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "more, combining GUES with SHOT-IM (i.e., GUES+SHOT-IM) expands the focus on DR-related features beyond those captured by SHOT-IM alone. Additionally, when comparing the four models to Oracle, only GUES and GUES+SHOT-IM resemble Oracle, suggesting that GUES effectively directs the model's attention to DR-critical features.", + "bbox": [ + 511, + 484, + 906, + 575 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Feature distribution. Taking the task DDR→APTOS as a toy experiment, we visualize the feature distribution extracted from the final convolutional layer of the prediction model using a 3D density chart. Considering that APTOS is a class-imbalanced dataset, with the “No DR” class alone comprising up to $49.2\\%$ of the dataset, our analysis focuses on this crucial property. As shown in Fig. 6, the feature distribution of the source model does not reflect this imbalanced characteristic; instead, it displays a more uniform classification. Conversely, the feature distribution of GUES exhibits a distinct imbalance, with one expanded high-density region alongside several smaller high-density regions, resembling the distribution pattern seen in Oracle.", + "bbox": [ + 511, + 579, + 908, + 776 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Visualization of unadversarial examples. This part visualizes unadversarial examples of two typical target samples from APTOS and corresponding generative perturbations, based on the task DDR $\\rightarrow$ APTOS. Considering the generative perturbations alter the original images that may not be easily visible to the naked eye, we collect RGB statistics to illustrate these changes quantitatively. It is observed that in Fig. 7, each channel (R, G, and B) exhibits notable", + "bbox": [ + 511, + 779, + 910, + 901 + ], + "page_idx": 6 + }, + { + "type": "page_number", + "text": "28343", + "bbox": [ + 478, + 944, + 519, + 957 + ], + "page_idx": 6 + }, + { + "type": "table", + "img_path": "images/abbc41cdc5923668482ea79c8f23b3797adb03c20e06c083b383e1746de46547.jpg", + "table_caption": [ + "Table 3. ACC results of ablation study $(\\%)$" + ], + "table_footnote": [], + "table_body": "
#\\( L_{\\text{KL}} \\)\\( L_{\\text{MSE}} \\)APTOSDDRDeepDRMD2Avg.
1XX51.059.152.553.053.9
2✓X54.758.354.953.355.3
3X✓56.960.152.952.955.7
4✓✓60.661.356.055.758.4
5GUES w/ AE53.860.253.853.255.3
6GUES w/ Mixup56.860.253.052.855.7
7GUES w/ Self56.560.453.754.456.3
8GUES w/ Sal46.134.341.845.942.0
", + "bbox": [ + 99, + 114, + 470, + 223 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "fluctuations, with the RGB statistics of the original images (a) differing significantly from those of the unadversarial examples (c). Additionally, each generative perturbation is unique, meaning that the alterations introduced by these perturbations are individual. These results suggest that the perturbations may help highlight critical DR-related features, refining the model's focus on diagnostically relevant areas.", + "bbox": [ + 89, + 253, + 482, + 359 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "4.6. Further Analysis", + "text_level": 1, + "bbox": [ + 89, + 372, + 256, + 387 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Ablation study. In this part, we evaluate the effect of objective loss, as well as the components involved in GUES including the sampling strategy and saliency map-based supervision. To address the first issue, we conduct a progressive experiment. The top four rows of Tab. 3 list the ablation results where the source model's performance is the baseline. Using $L_{\\mathrm{KL}}$ or $L_{\\mathrm{MSE}}$ alone yields an average ACC improvement of approximately $1.4\\%$ and $1.8\\%$ , respectively, over the baseline. As both of them work, the ACC increases $3.5\\%$ on average further. The results indicate that all objective components positively affect the final performance.", + "bbox": [ + 89, + 396, + 483, + 561 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "To evaluate the impact of the sampling component, we propose a variation method of GUES, GUES w/ AE, where we remove this sampling process by replacing VAE with a conventional Auto-encoder model. In addition, two GUES variations are used to assess the advantage of saliency map-based supervision. Specifically, GUES w/ Mixup replaces the saliency maps with a Mixup of saliency maps and original images, whilst GUES w/ Self replaces the saliency maps with the original images. As presented in the Tab. 3, compared with the full version of GUES (the fourth row), GUES w/ AE, GUES w/ Mixup and GUES w/ Self decrease by $3.1\\%$ at least on average. Besides, replacing the original images with saliency maps as inputs (GUES w/ Sal) leads to a significant drop of about $15.4\\%$ . Above these experiments confirm the effectiveness of our design choices.", + "bbox": [ + 89, + 564, + 483, + 790 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Parameter sensitiveness. Taking the task DDR $\\rightarrow$ APTOS as a toy experiment, we present the GUES performance varying as hyper-parameters $0.5\\leq \\alpha \\leq 1.5$ with 0.1 steps, $0.00005\\leq \\beta \\leq 0.00014$ with 0.00001 steps. As depicted in Fig. 8, the ACC, QWK, and AVG variation surfaces show fluctuations in the tiny performance zone, with approximately $0.2\\%$ in ACC, $0.25\\%$ in QWK, and $0.4\\%$ in", + "bbox": [ + 89, + 794, + 483, + 900 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/9cdc9e8c29501aa175b39b55b2bda1428b83dc0e214cde588a14fcebc3a5d085.jpg", + "image_caption": [ + "Figure 8. Parameter sensitiveness study results over $\\alpha \\times \\beta$ based on task DDR $\\rightarrow$ APTOS. From left to right, there are results on ACC, QWK, and AVG, respectively." + ], + "image_footnote": [], + "bbox": [ + 526, + 89, + 645, + 180 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/3c06c4119aeb4010430e016ae1e367f133b843ee2a703215d4cbaef1c5ca462e.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 651, + 89, + 769, + 180 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/5b58cf7f34b38f89b60754474f73c44da4741fc11bfe7135a5098f474b99d0f1.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 774, + 89, + 893, + 180 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "AVG. This observation suggests that GUES is insensitive to alterations in $\\alpha$ and $\\beta$ .", + "bbox": [ + 511, + 258, + 903, + 289 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Limitation. GUES uses the saliency map to guide the learning of the generative function. This method is effective for DR grading but encounters challenges in natural image scenarios. The natural images contain rich semantics, such as shape, relative structure, and complex background, which are not all relevant to tasks. However, the saliency map blindly highlights all those factors, struggling to capture the task-specific ones. In a theoretical point-view, the richness of these semantics means significant variations, resulting in a super-relaxed bound constant $U$ (see Theorem 2) that undermines the descriptive power of the saliency map for $\\frac{\\partial \\delta_0}{\\partial x}$ . In contrast, fundus images are more monolithic, implying a smaller $U$ that justifies the usage of the saliency map. (The further discussion is provided in Supplementary)", + "bbox": [ + 511, + 292, + 906, + 505 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "5. Conclusion", + "text_level": 1, + "bbox": [ + 511, + 518, + 632, + 534 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "In this paper, we propose a clinically motivated setting, OMG-DA, where the models are unseen prior to their use, and only target data flows are accessible. This setting ensures both model protection and source data privacy in a data flow scenario. To adapt to the target domain without access to the models, we introduce a GUES approach. Instead of conventional iterative optimization, we generate unadversarial examples for flowing target data by directly predicting individual perturbations. This approach is grounded in the theoretical results of generative unadversarial learning. In practice, we utilize the VAE model to learn the perturbation generation function with a latent input variable. Furthermore, we demonstrate that saliency maps can serve as an upper bound for this latent variable. This relationship inspires us to use the saliency maps as pseudo-perturbation labels for model training. Extensive experiments conducted on four DR benchmarks confirm that the proposed method can achieve state-of-the-art results when it pairs with both frozen pre-trained and fine-tuning models.", + "bbox": [ + 511, + 544, + 908, + 830 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Acknowledgments", + "text_level": 1, + "bbox": [ + 511, + 845, + 671, + 861 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "This work is supported by the National Natural Science Foundation of China (62476169, 62206168, 62276048).", + "bbox": [ + 511, + 869, + 905, + 900 + ], + "page_idx": 7 + }, + { + "type": "page_number", + "text": "28344", + "bbox": [ + 478, + 944, + 517, + 955 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "References", + "text_level": 1, + "bbox": [ + 91, + 89, + 187, + 104 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[1] Aptos: Aptos 2019 blindness detection website. https://www.kaggle.com/c/aptos2019-blindness-detection, accessed February 20, 2022.", + "[2] Quadratic weighted kappa. https://www.Eyepacs.com/aroraaman/quadratic-kappa-metric-explained-in-5-simple-steps, accessed July 2022.", + "[3] Eman AbdelMaksoud, Sherif Barakat, and Mohammed Elmogy. A comprehensive diagnosis system for early signs and different diabetic retinopathy grades using fundus retinal images based on pathological changes detection. Computers in Biology and Medicine, 126:104039, 2020.", + "[4] Mohammad Atwany and Mohammad Yaqub. Drogen: domain generalization in diabetic retinopathy classification. In MICCAI, 2022.", + "[5] Haoxuan Che, Yuhan Cheng, Haibo Jin, and Hao Chen. Towards generalizable diabetic retinopathy grading in unseen domains. In MICCAI, 2023.", + "[6] Ling Dai, Liang Wu, Huating Li, Chun Cai, Qiang Wu, Hongyu Kong, Ruhan Liu, Xiangning Wang, Xuhong Hou, Yuexing Liu, et al. A deep learning system for detecting diabetic retinopathy across the disease spectrum. Nature communications, 12(1):3242, 2021.", + "[7] Etienne Decencière, Xiwei Zhang, Guy Cazuguel, Bruno Lay, Béatrice Cochener, Caroline Trone, Philippe Gain, John-Richard Ordóñez-Varela, Pascale Massin, Ali Erginay, et al. Feedback on a publicly distributed image database: the messidor database. Image Analysis & Stereology, pages 231-234, 2014.", + "[8] Along He, Tao Li, Ning Li, Kai Wang, and Huazhu Fu. Cabinet: Category attention block for imbalanced diabetic retinopathy grading. IEEE Transactions on Medical Imaging, 40(1):143-153, 2020.", + "[9] Yijin Huang, Junyan Lyu, Pujin Cheng, Roger Tam, and Xiaoying Tang. Ssit: Saliency-guided self-supervised image transformer for diabetic retinopathy grading. IEEE Journal of Biomedical and Health Informatics, 2024.", + "[10] Diederik P Kingma. Auto-encoding variational bayes. arXiv:1312.6114, 2013.", + "[11] Wouter M Kouw and Marco Loog. An introduction to domain adaptation and transfer learning. arXiv:1812.11806, 2018.", + "[12] Jonghyun Lee, Dahuin Jung, Junho Yim, and Sungroh Yoon. Confidence score for source-free unsupervised domain adaptation. In ICML, 2022.", + "[13] Tao Li, Yingqi Gao, Kai Wang, Song Guo, Hanruo Liu, and Hong Kang. Diagnostic assessment of deep learning algorithms for diabetic retinopathy screening. Information Sciences, 501:511-522, 2019.", + "[14] Tao Li, Wang Bo, Chunyu Hu, Hong Kang, Hanruo Liu, Kai Wang, and Huazhu Fu. Applications of deep learning in fundus images: A review. Medical Image Analysis, 69: 101971, 2021.", + "[15] Jian Liang, Dapeng Hu, and Jiashi Feng. Do we really need to access the source data? source hypothesis transfer for unsupervised domain adaptation. In ICML, 2020." + ], + "bbox": [ + 93, + 114, + 485, + 900 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[16] Mattia Litrico, Alessio Del Bue, and Pietro Morerio. Guiding pseudo-labels with uncertainty estimation for source-free unsupervised domain adaptation. In CVPR, 2023.", + "[17] Ruhan Liu, Xiangning Wang, Qiang Wu, Ling Dai, Xi Fang, Tao Yan, Jaemin Son, Shiqi Tang, Jiang Li, Zijian Gao, et al. Deepgrid: Diabetic retinopathy—grading and image quality estimation challenge. *Patterns*, 3(6), 2022.", + "[18] Xingbin Liu, Huafeng Kuang, Xianming Lin, Yongjian Wu, and Rongrong Ji. Cat: Collaborative adversarial training. arXiv:2303.14922, 2023.", + "[19] Sebastian Montabone and Alvaro Soto. Human detection using a mobile platform and novel features derived from a visual saliency mechanism. Image and Vision Computing, 28 (3):391-402, 2010.", + "[20] Duy MH Nguyen, Truong TN Mai, Ngoc TT Than, Alexander Prange, and Daniel Sonntag. Self-supervised domain adaptation for diabetic retinopathy grading using vessel image reconstruction. In KI, 2021.", + "[21] Shuaicheng Niu, Jiaxiang Wu, Yifan Zhang, Zhiquan Wen, Yaofo Chen, Peilin Zhao, and Mingkui Tan. Towards stable test-time adaptation in dynamic wild world. In ICLR, 2023.", + "[22] Jiaming Qiu, Weikai Huang, Yijin Huang, Nanxi Yu, and Xiaoying Tang. Augpaste: A one-shot approach for diabetic retinopathy detection. Biomedical Signal Processing and Control, 96:106489, 2024.", + "[23] Jinye Ran, Guanghua Zhang, Fan Xia, Ximei Zhang, Juan Xie, and Hao Zhang. Source-free active domain adaptation for diabetic retinopathy grading based on ultra-wide-field fundus images. Computers in Biology and Medicine, 174:108418, 2024.", + "[24] Hadi Salman, Andrew Ilyas, Logan Engstrom, Sai Vemprala, Aleksander Madry, and Ashish Kapoor. Unadversarial examples: Designing objects for robust vision. In NeurIPS, 2021.", + "[25] Abhijith Sharma, Phil Munz, and Apurva Narayan. Nsa: Naturalistic support artifact to boost network confidence. In IJCNN, 2023.", + "[26] Reza Shokri, Marco Stronati, Congzheng Song, and Vitaly Shmatikov. Membership inference attacks against machine learning models. In S&P, 2017.", + "[27] Daniel E Singer, David M Nathan, Howard A Fogel, and Andrew P Schachat. Screening for diabetic retinopathy. Annals of Internal Medicine, 116(8):660-671, 1992.", + "[28] C Szegedy. Intriguing properties of neural networks. arXiv:1312.6199, 2013.", + "[29] Song Tang, An Chang, Fabian Zhang, Xiatian Zhu, Mao Ye, and Changshui Zhang. Source-free domain adaptation via target prediction distribution searching. International Journal of Computer Vision, 132(3):654-672, 2024.", + "[30] Song Tang, Wenxin Su, Mao Ye, Jianwei Zhang, and Xiatian Zhu. Unified source-free domain adaptation. arXiv:2403.07601, 2024.", + "[31] Nishtha Tomar, Sushmita Chandel, and Gaurav Bhatnagar. A visual attention-based algorithm for brain tumor detection using an on-center saliency map and a superpixel-based framework. Healthcare Analytics, 5:100323, 2024." + ], + "bbox": [ + 516, + 92, + 906, + 898 + ], + "page_idx": 8 + }, + { + "type": "page_number", + "text": "28345", + "bbox": [ + 478, + 944, + 517, + 955 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[32] Hugo Touvron, Matthieu Cord, Matthijs Douze, Francisco Massa, Alexandre Sablayrolles, and Hervé Jégou. Training data-efficient image transformers & distillation through attention. In ICML, 2021.", + "[33] Jeya Maria Jose Valanarasu, Pengfei Guo, VS Vibashan, and Vishal M Patel. On-the-fly test-time adaptation for medical image segmentation. In MIDL, 2024.", + "[34] Dequan Wang, Evan Shelhamer, Shaoteng Liu, Bruno Olshausen, and Trevor Darrell. Tent: Fully test-time adaptation by entropy minimization. In ICLR, 2020.", + "[35] Tianyunxi Wei, Yijin Huang, Li Lin, Pujin Cheng, Sirui Li, and Xiaoying Tang. Saliency-guided and patch-based mixup for long-tailed skin cancer image classification. arXiv:2406.10801, 2024.", + "[36] Zhan Wu, Gonglei Shi, Yang Chen, Fei Shi, Xinjian Chen, Gouenou Coatieux, Jian Yang, Limin Luo, and Shuo Li. Coarse-to-fine classification for diabetic retinopathy grading using convolutional neural network. Artificial Intelligence in Medicine, 108:101936, 2020.", + "[37] Tongkun Xu, Weihua Chen, WANG Pichao, Fan Wang, Hao Li, and Rong Jin. Cdtrans: Cross-domain transformer for unsupervised domain adaptation. In ICLR, 2021.", + "[38] Shiqi Yang, Joost van de Weijer, Luis Herranz, Shangling Jui, et al. Exploiting the intrinsic neighborhood structure for source-free domain adaptation. In NeurIPS, 2021.", + "[39] Hongxu Yin, Arun Mallya, Arash Vahdat, Jose M Alvarez, Jan Kautz, and Pavlo Molchanov. See through gradients: Image batch recovery via grad inversion. In CVPR, 2021.", + "[40] Chenrui Zhang, Tao Lei, and Ping Chen. Diabetic retinopathy grading by a source-free transfer learning approach. Biomedical Signal Processing and Control, 73:103423, 2022." + ], + "bbox": [ + 91, + 92, + 485, + 531 + ], + "page_idx": 9 + }, + { + "type": "page_number", + "text": "28346", + "bbox": [ + 478, + 945, + 519, + 955 + ], + "page_idx": 9 + } +] \ No newline at end of file diff --git a/2025/Domain Adaptive Diabetic Retinopathy Grading with Model Absence and Flowing Data/8fc9ee64-90f9-4f33-8fb1-64a5f7b319a1_model.json b/2025/Domain Adaptive Diabetic Retinopathy Grading with Model Absence and Flowing Data/8fc9ee64-90f9-4f33-8fb1-64a5f7b319a1_model.json new file mode 100644 index 0000000000000000000000000000000000000000..3ef2d45de5a35fe37347d232ecd4d672d7a1df17 --- /dev/null +++ b/2025/Domain Adaptive Diabetic Retinopathy Grading with Model Absence and Flowing Data/8fc9ee64-90f9-4f33-8fb1-64a5f7b319a1_model.json @@ -0,0 +1,2189 @@ +[ + [ + { + "type": "header", + "bbox": [ + 0.107, + 0.003, + 0.182, + 0.043 + ], + "angle": 0, + "content": "CVF" + }, + { + "type": "header", + "bbox": [ + 0.237, + 0.001, + 0.808, + 0.045 + ], + "angle": 0, + "content": "This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore." + }, + { + "type": "title", + "bbox": [ + 0.195, + 0.131, + 0.804, + 0.175 + ], + "angle": 0, + "content": "Domain Adaptive Diabetic Retinopathy Grading with Model Absence and Flowing Data" + }, + { + "type": "text", + "bbox": [ + 0.23, + 0.215, + 0.77, + 0.252 + ], + "angle": 0, + "content": "Wenxin \\(\\mathrm{Su}^1\\), Song Tang\\(^{*1,2,3}\\), Xiaofeng Liu\\(^4\\), Xiaojing Yi\\(^5\\), Mao Ye\\(^6\\), Chunxiao \\(\\mathrm{Zu}^1\\), Jiahao Li\\(^7\\), and Xiatian Zhu\\(^8\\)" + }, + { + "type": "text", + "bbox": [ + 0.106, + 0.264, + 0.89, + 0.318 + ], + "angle": 0, + "content": "1University of Shanghai for Science and Technology 2Universität Hamburg, 3ComOriginMat Inc. \n4Tale University 5Sichuan Eye Hospital 6University of Electronic Science and Technology of China \n7Chinese Academy of Medical Sciences & Peking Union Medical College 8University of Surrey" + }, + { + "type": "text", + "bbox": [ + 0.396, + 0.321, + 0.611, + 0.334 + ], + "angle": 0, + "content": "steventangsong@gmail.com" + }, + { + "type": "title", + "bbox": [ + 0.249, + 0.379, + 0.327, + 0.394 + ], + "angle": 0, + "content": "Abstract" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.414, + 0.486, + 0.852 + ], + "angle": 0, + "content": "Domain shift (the difference between source and target domains) poses a significant challenge in clinical applications, e.g., Diabetic Retinopathy (DR) grading. Despite considering certain clinical requirements, like source data privacy, conventional transfer methods are predominantly model-centered and often struggle to prevent model-targeted attacks. In this paper, we address a challenging Online Model-aGnostic Domain Adaptation (OMG-DA) setting, driven by the demands of clinical environments. This setting is characterized by the absence of the model and the flow of target data. To tackle the new challenge, we propose a novel approach, Generative Unadversarial ExampleS (GUES), which enables adaptation from a data-centric perspective. Specifically, we first theoretically reformulate conventional perturbation optimization in a generative way—learning a perturbation generation function with a latent input variable. During model instantiation, we leverage a Variational AutoEncoder to express this function. The encoder with the reparameterization trick predicts the latent input, whilst the decoder is responsible for the generation. Furthermore, the saliency map is selected as pseudoperturbation labels. Because it not only captures potential lesions but also theoretically provides an upper bound on the function input, enabling the identification of the latent variable. Extensive experiments on DR benchmarks with both frozen pre-trained models and trainable models demonstrate the superiority of GUES, showing robustness even with small batch size. The source code and data are available at https://github.com/tntek/GUES." + }, + { + "type": "image", + "bbox": [ + 0.536, + 0.377, + 0.886, + 0.495 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.512, + 0.505, + 0.908, + 0.617 + ], + "angle": 0, + "content": "Figure 1. Comparison between the OMG-DA and SFDA settings. (a) In SFDA, adaptation builds upon the accumulated data, which demands significant storage resources in the hospital. Additionally, the models' architecture and parameters are accessible, exposing them to potential attacks. (b) OMG-DA provides a practical scenario: Flowing data mimic the patients' arrival in a stream way, and the models are unseen (strictly controlled) before using them, avoiding attacks like membership inference attacks [26]." + }, + { + "type": "title", + "bbox": [ + 0.515, + 0.645, + 0.645, + 0.661 + ], + "angle": 0, + "content": "1. Introduction" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.673, + 0.909, + 0.884 + ], + "angle": 0, + "content": "Diabetic Retinopathy (DR) is a significant health concern, ranking among the leading causes of blindness and affecting millions of people worldwide [3]. Early-stage intervention for DR is crucial to preserve vision, highlighting the importance of timely diagnosis [27]. Although deep learning (DL) has demonstrated promising results in automating the grading of DR [6, 8, 36], deploying DL models in real-world clinical settings remains challenging. For example, DL models often struggle to generalize effectively to complex scenarios, such as variations in imaging equipment, ethnic groups, or temporal factors, leading to different data distributions, a challenge known as domain shift [11]. This issue significantly hampers the widespread adoption and success of DL-based diagnostic tools in clinical practice [14]." + }, + { + "type": "text", + "bbox": [ + 0.533, + 0.886, + 0.905, + 0.901 + ], + "angle": 0, + "content": "Recently, many adaptation methods for grading diabetic" + }, + { + "type": "page_footnote", + "bbox": [ + 0.11, + 0.887, + 0.235, + 0.9 + ], + "angle": 0, + "content": "*Corresponding author" + }, + { + "type": "page_number", + "bbox": [ + 0.479, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "28337" + } + ], + [ + { + "type": "table_caption", + "bbox": [ + 0.134, + 0.089, + 0.861, + 0.104 + ], + "angle": 0, + "content": "Table 1. Comparison between different transfer settings. Notation: source (s), target (t), data \\( \\mathbf{x} \\) ,label \\( \\mathbf{y} \\) ,loss function \\( L\\left( \\cdot \\right) \\) ." + }, + { + "type": "table", + "bbox": [ + 0.108, + 0.115, + 0.89, + 0.193 + ], + "angle": 0, + "content": "
Setting nameModel absenceData flowSource data privacySource dataTarget dataTrain lossTest loss
Fine-tuning (FT)XX✓-xt,ytL(xt,yt)-
Unsupervised domain adaptation (UDA)XXXxs,ysxtL(xs,ys)+L(xt)-
Source-free domain adaptation (SFDA)XX✓-xtL(xt)-
Test-time adaptation (TTA)X✓✓-xt-L(xt)
Online model-agnostic domain adaptation (OMG-DA)✓✓✓-xt--
" + }, + { + "type": "text", + "bbox": [ + 0.089, + 0.218, + 0.484, + 0.368 + ], + "angle": 0, + "content": "retinopathy (DR) have focused on addressing the issue of domain shift [5, 20, 23, 40]. The initial classic transfer learning strategies, Unsupervised Domain Adaptation (UDA), required access to well-annotated source data. Nevertheless, the growing emphasis on privacy protection has shifted research toward the Source-Free Domain Adaptation (SFDA) framework [23, 40]. SFDA involves adapting a source model—pre-trained on the source domain—to the target domain in a self-supervised manner, thereby ensuring the protection of source patient data." + }, + { + "type": "text", + "bbox": [ + 0.089, + 0.37, + 0.484, + 0.549 + ], + "angle": 0, + "content": "In recent developments, specific needs have emerged in the clinical field. The introduction of model weight-based techniques for reconstructing training data has created a demand for model privacy [26, 39], which goes beyond traditional source data protection. In addition, there is a growing requirement for models capable of handling incoming patient data in a flowing fashion (test data arriving sequentially), referred to as a flowing data constraint [33, 34]. Unfortunately, existing SFDA methods cannot effectively address this challenge, as they rely on full access to the model and require offline training on a pre-collected dataset. Fig. 1 provides an intuitive illustration of this issue." + }, + { + "type": "text", + "bbox": [ + 0.089, + 0.552, + 0.484, + 0.688 + ], + "angle": 0, + "content": "In this paper, we consider a clinically motivated setting, called Online Model-aGnostic Domain Adaptation (OMG-DA) and propose a novel Generative Unadversarial ExampleS (GUES) approach for the DR grading problem in this new setting. Specifically, OMG-DA presents an extreme safety scenario: The available target data is unlabeled and arrives in a flowing format, with no prior information about the pre-trained source model and data. Tab. 1 provides a detailed comparison with previous adaptation settings." + }, + { + "type": "text", + "bbox": [ + 0.089, + 0.689, + 0.484, + 0.901 + ], + "angle": 0, + "content": "In GUES, we address the absence of source data and pre-trained models by producing generalized unadversarial examples [24] for unlabeled target data. To this end, we introduce generative unadversarial learning, which theoretically reformulates conventional iterative perturbation optimization. This new method aims to learn a generative function for perturbations and involves addressing two key tasks: (1) Identifying the latent function input, which is the derivative of initial random noise w.r.t. image data, and (2) selecting a self-supervised property to serve as pseudo-perturbation labels. In practice, we leverage a Variational Autoencoder (VAE) [10] based approach to facilitate this learning process. In terms of function representation, we model the latent input using the encoder along with the reparameterization trick," + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.218, + 0.907, + 0.293 + ], + "angle": 0, + "content": "whilst the decoder accomplishes the generation of individual-perturbation. Additionally, we choose the saliency map as the pseudo-perturbation label for two reasons: (1) It helps discover potential lesions, and (2) it aids in identifying the latent input by providing an upper bound." + }, + { + "type": "text", + "bbox": [ + 0.531, + 0.294, + 0.843, + 0.308 + ], + "angle": 0, + "content": "Our contributions are summarized as follows:" + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.31, + 0.905, + 0.369 + ], + "angle": 0, + "content": "- Pioneering a novel transfer setting OMG-DA, which is closer to real-world clinical scenarios and meets three typical requirements at the same time, including (1) model absence, (2) flowing data, and (3) source data privacy." + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.37, + 0.907, + 0.445 + ], + "angle": 0, + "content": "- Developing a new OMG-DA approach GUES in the context of DR grading, grounded on the generative unadversarial examples theory, where we learn an individual-perturbation generative function under saliency map supervision, removing relying on labels and models." + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.446, + 0.905, + 0.505 + ], + "angle": 0, + "content": "- Extensive evaluations on four DR benchmarks, indicating that GUES can largely promote the source model's performance in the target domain, as well as trainable test-time adaptation models, even at small batch size." + }, + { + "type": "list", + "bbox": [ + 0.513, + 0.31, + 0.907, + 0.505 + ], + "angle": 0, + "content": null + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.522, + 0.652, + 0.537 + ], + "angle": 0, + "content": "2. Related work" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.548, + 0.907, + 0.744 + ], + "angle": 0, + "content": "Adaptation methods for DR grading. Driven by real-world medical requirements, domain adaptation has been an attractive topic in this DR grading issue. For instance, Nguyen et al. [20] introduce a UDA approach that enables the model to focus on vessel structures that remain invariant to domain shifts via image reconstruction using labeled source domain data. In SFDA, Zhang et al. [40] propose generating labeled, target-style retinal images to improve the source model's generalization, relying solely on a pre-trained source model and unlabeled target images. Additionally, DG in DR grading has been explored through domain-invariant feature learning approaches [4] and divergence-based methods [5], leveraging labeled source domain data." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.746, + 0.909, + 0.867 + ], + "angle": 0, + "content": "The methods above rely on labeled data, require full access to the model, and necessitate offline training on a precollected dataset. In real clinical settings, these requirements can be impractical due to constraints around data and model privacy, as well as the need for real-time adaptability without retraining. In contrast, GUES provides an online adaptation solution that operates without requiring labeled data or access to the model, addressing domain adaptive DR grading." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.871, + 0.909, + 0.902 + ], + "angle": 0, + "content": "Unadversarial learning. Unadversarial learning was initially developed by Salman et al. [24], aiming to modify" + }, + { + "type": "page_number", + "bbox": [ + 0.479, + 0.945, + 0.521, + 0.957 + ], + "angle": 0, + "content": "28338" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.09, + 0.092, + 0.484, + 0.243 + ], + "angle": 0, + "content": "input image distribution to make them more easily recognizable by the model. Current mainstreams achieve this learning process by adding class-specific perturbations to the input images. Here, the perturbations are generated based on the gradient of an objective function w.r.t. image. This approach allows for the design of unadversarial examples without model training. For example, based on [24], NSA [25] introduces a method to generate more natural perturbations using a trainable generator. Similarly, CAT [18] demonstrates a new distance metric for generating unadversarial examples." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.244, + 0.484, + 0.35 + ], + "angle": 0, + "content": "All existing unadversarial learning methods require access to model parameters, outputs, and labeled data. This dependency invalidates them in our OMG-DA setting which only has access to unlabeled target data. In addition to this, our GUES produces individualized unadversarial examples in a generative manner, which stands out from previous methods that focus on class-specific unadversarial examples." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.355, + 0.484, + 0.566 + ], + "angle": 0, + "content": "Saliency map for medical image. A fine-grained saliency map is a pixel feature generated by calculating the central surround differences within images, identifying salient regions without any need for training [19]. This feature is widely applied in various medical image analysis tasks to extract pathologically important regions [9, 22, 31, 35]. For instance, in DR grading, studies such as [9, 22] utilize saliency maps to guide models in focusing on critical features like the optic disc, cup, and vessel structures. Similarly, in brain tumor detection, Tomar et al. [31] leverage saliency maps to enhance the model's attention on tumor and bone structures. In skin cancer detection, saliency maps help isolate lesion regions with distinctive features, such as lumpiness, which are essential for accurate diagnosis [35]." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.568, + 0.484, + 0.613 + ], + "angle": 0, + "content": "As stated above, existing works primarily use saliency maps to highlight lesion regions. Unlike it, GUES selects saliency maps as pseudo-perturbation labels." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.631, + 0.226, + 0.649 + ], + "angle": 0, + "content": "3. Methodology" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.658, + 0.484, + 0.794 + ], + "angle": 0, + "content": "Problem statement. Given two different but related domains, i.e., source domain \\( \\mathcal{S} \\) and target domain \\( \\mathcal{T} \\), \\( \\mathcal{S} \\) contains \\( n_s \\) labeled samples, while \\( \\mathcal{T} \\) has \\( n \\) unlabeled data. Both labeled and unlabeled samples share the same \\( C \\) categories. Let \\( \\mathcal{X}_s \\) and \\( \\mathcal{Y}_s \\) be the source samples and the corresponding labels. Similarly, we denote the target samples and their labels by \\( \\mathcal{X}_t = \\{\\pmb{x}_t^i\\}_{i=1}^n \\) and \\( \\mathcal{Y}_t = \\{y_t^i\\}_{i=1}^n \\), respectively, where \\( n \\) signifies the number of samples. The source model \\( \\theta_s \\) is pre-trained on \\( \\{\\mathcal{X}_s, \\mathcal{Y}_s\\} \\)." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.796, + 0.484, + 0.901 + ], + "angle": 0, + "content": "OMG-DA is featured in (1) the absence of the source model \\(\\theta_{s}\\) and domain \\(S\\) and (2) the flowing target data \\(\\mathcal{X}_t\\) the same as the TTA setting [34]. Unlike previous transfer settings that are model adaptation-centered [15, 29, 30, 38], OMG-DA considers adaptation from the perspective of data. Specifically, OMG-DA aims to modify the distribution of target data to facilitate downstream tasks." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.091, + 0.829, + 0.108 + ], + "angle": 0, + "content": "3.1. Generative Unadversarial Examples" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.113, + 0.907, + 0.236 + ], + "angle": 0, + "content": "The unadversarial learning idea [24] is well aligned with OMG-DA in terms of promoting adaptation by enhancing the generalization of input data. We briefly recap traditional unadversarial learning. Unlike adversarial learning [28] that generates confusing samples to mislead models, unadversarial learning aims to construct generalized samples, tackling out-of-distribution issues. Formally, this learning can be summarized in the following optimization problem." + }, + { + "type": "equation", + "bbox": [ + 0.567, + 0.245, + 0.907, + 0.269 + ], + "angle": 0, + "content": "\\[\n\\hat {\\delta} = \\arg \\min _ {\\delta} L \\left(f _ {\\theta} (x + \\delta), y\\right), s. t. | | \\delta | | \\leq \\epsilon \\tag {1}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.279, + 0.906, + 0.355 + ], + "angle": 0, + "content": "where \\( L(\\cdot) \\) denotes objective function, e.g., cross-entropy loss for classification tasks, \\( x \\) and \\( y \\) are input image and its label, \\( f_{\\theta} \\) is a pre-trained model with parameters \\( \\theta \\), \\( \\delta \\) is a perturbation, \\( \\epsilon \\) is a small threshold. The current scheme solves this problem in an iterative way formulated as" + }, + { + "type": "equation", + "bbox": [ + 0.521, + 0.363, + 0.906, + 0.392 + ], + "angle": 0, + "content": "\\[\n\\delta_ {k + 1} = \\delta_ {k} + \\alpha \\cdot \\operatorname {s i g n} \\left(\\nabla_ {x} L \\left(f _ {\\theta} \\left(x + \\delta_ {k}\\right), y\\right)\\right), k \\in [ 0, K - 1 ], \\tag {2}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.392, + 0.906, + 0.497 + ], + "angle": 0, + "content": "where \\(\\alpha\\) is a trade-off parameter, \\(K\\) is iteration number, \\(\\delta_0\\) is an initial random noise. In the inference phase, the optimal perturbation \\(\\hat{\\delta}\\) is integrated into the input \\(x\\), forming an unadversarial example \\(\\hat{x} = x + \\hat{\\delta}\\), which is easily recognizable by the model \\(f_{\\theta}\\). Obviously, the conventional unadversarial paradigm cannot meet our OMG-DA setting due to the absence of \\(f_{\\theta}\\), \\(L\\), and the label \\(y\\)." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.498, + 0.906, + 0.543 + ], + "angle": 0, + "content": "In this paper, we re-consider the iterative optimization process above and obtain the theorem below (The proof is provided in Supplementary)." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.552, + 0.907, + 0.598 + ], + "angle": 0, + "content": "Theorem 1 Given the unadversarial learning problem defined in Eq. (1), the iterative process featured by Eq. (2) can be expressed as the following generative form." + }, + { + "type": "equation", + "bbox": [ + 0.617, + 0.607, + 0.907, + 0.641 + ], + "angle": 0, + "content": "\\[\n\\delta_ {k} = \\delta_ {0} + V \\cdot F _ {\\Phi} \\left(\\frac {\\partial \\delta_ {0}}{\\partial x}\\right), \\tag {3}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.649, + 0.907, + 0.682 + ], + "angle": 0, + "content": "where \\(\\delta_0\\) is an initial random noise, \\(V > 0\\) is a bound constant, \\(F_{\\Phi}\\) is a generative function, \\(\\frac{\\partial\\delta_0}{\\partial x}\\) is a latent variable." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.703, + 0.907, + 0.794 + ], + "angle": 0, + "content": "Grounded on Theorem 1, we have: When \\(\\delta_{k}\\) converges to optimal \\(\\hat{\\delta}\\), i.e., \\(\\delta_{k} \\rightarrow \\hat{\\delta}\\), function \\(F_{\\Phi}\\) also evolves to the optimal one, denoted by \\(\\hat{F}_{\\Phi}\\), i.e., \\(F_{\\Phi} \\rightarrow \\hat{F}_{\\Phi}\\). This provides an insight: The unadversarial learning problem above can also be solved in a generative fashion. Correspondingly, the generated data are termed generative unadversarial examples." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.802, + 0.704, + 0.817 + ], + "angle": 0, + "content": "3.2. Model Instantiation" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.826, + 0.906, + 0.901 + ], + "angle": 0, + "content": "Within this context of generative unadversarial learning, conventional unadversarial learning presented in Eq. (1) is boiled down to learning \\( F_{\\Phi}\\left(\\frac{\\partial\\delta_0}{\\partial x}\\right) \\). We can achieve this by training a generative neural network. To make this solution sense, we have to solve two difficulties as follows. (A) One is" + }, + { + "type": "page_number", + "bbox": [ + 0.48, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "28339" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.135, + 0.089, + 0.868, + 0.306 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.09, + 0.317, + 0.908, + 0.375 + ], + "angle": 0, + "content": "Figure 2. The instantiation framework of GUES in the OMG-DA setting. (a) For target input \\( x_{t} \\), the VAE model generates individual perturbation \\( \\delta_{t} = F_{\\Phi}\\left(\\frac{\\partial\\delta_{0}}{\\partial x}\\right) \\). After that, the by-pass path incorporates \\( \\delta_{t} \\) and \\( x_{t} \\) to create the generative unadversarial example \\( \\hat{x}_{t} \\). Treating \\( x_{t} \\)'s saliency map \\( g_{t} \\) as reconstruction supervision for model training. (b) At the inference phase, the generated unadversarial example \\( \\hat{x}_{t} \\) is directly provided to the frozen source model or other trainable models." + }, + { + "type": "image", + "bbox": [ + 0.113, + 0.397, + 0.462, + 0.502 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.09, + 0.512, + 0.484, + 0.609 + ], + "angle": 0, + "content": "Figure 3. Explanation in choosing fine-grained saliency maps as supervision. Left: The testing fundus image selected from \"Moderate DR\" class in APTOS demonstrates that H (hemorrhages), SE (soft exudates), and EX (hard exudates) are essential characteristics to judge the DR grade. Middle: The saliency map highlights those lesions. Right: The gradient-CAM visualization of the source model on task DDR→APTOS." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.621, + 0.484, + 0.668 + ], + "angle": 0, + "content": "the identification of \\(\\frac{\\partial\\delta_0}{\\partial x}\\) when the relationship between them is unknown. (B) The other is selecting property supervision (pseudo-perturbation labels) to drive \\(\\delta_{k}\\rightarrow \\hat{\\delta}\\)." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.671, + 0.484, + 0.852 + ], + "angle": 0, + "content": "Solution to Problem A. In practice, considering derivative \\(\\frac{\\partial\\delta_0}{\\partial x}\\) is relevant with both random noise \\(\\delta_0\\) and input image \\(x\\), we sample it from a certain Gaussian distribution associated with \\(x\\). Furthermore, the output size of \\(F_{\\Phi}\\) is the same as \\(x\\). Therefore, we employ the VAE model to jointly model \\(\\frac{\\partial\\delta_0}{\\partial x}\\) and \\(F_{\\Phi}\\), since VAE is an autoencoder characterized by random sampling. Specifically, as shown in Fig.2 (a), we approximate \\(\\frac{\\partial\\delta_0}{\\partial x}\\) by latent variable \\(z\\), which is jointly determined by input \\(x\\) and sampled random signal \\(\\tau\\). As for \\(F_{\\Phi}\\), it is demonstrated by the decoder module. Suppose \\(D(\\cdot)\\) is the decoder, \\(E_{\\tau}(\\cdot)\\) is the encoder with reparameterization trick, our scheme can be formulated as" + }, + { + "type": "equation", + "bbox": [ + 0.104, + 0.859, + 0.484, + 0.901 + ], + "angle": 0, + "content": "\\[\nD \\left(E _ {\\tau} (x)\\right)\\rightarrow F _ {\\Phi} \\left(\\frac {\\partial \\delta_ {0}}{\\partial x}\\right), E _ {\\tau} (x) \\rightarrow \\frac {\\partial \\delta_ {0}}{\\partial x}, D (\\cdot) \\rightarrow F _ {\\Phi} (\\cdot). \\tag {4}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.399, + 0.908, + 0.612 + ], + "angle": 0, + "content": "Solution to Problem B. We adopt the fine-grained saliency map as the supervision. Two reasons contribute to our selection. First of all, the empirical results show that, for the specific task of DR grading, the saliency map is an acceptable pseudo-perturbation. Specifically, perturbation in the unadversarial context enhances the regions associated with the category and reduces the prominence of other areas, thereby identifying the lesion zones relevant to DR grading. As illustrated in Fig. 3, the saliency map effectively identifies potential lesions, such as hemorrhages, soft exudates, and hard exudates. Furthermore, it includes gradient information, as it highlights regions similar to Grad-CAM (Right side). More importantly, we have the theorem below. (The proof is provided in Supplementary)" + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.622, + 0.908, + 0.682 + ], + "angle": 0, + "content": "Theorem 2 Given the partial derivatives of the initial random noise \\(\\delta_0\\) w.r.t image \\(x\\) is \\(\\frac{\\partial\\delta_0}{\\partial x}\\) and \\(x\\)'s saliency map is \\(s = G(x)\\) where \\(G(\\cdot)\\) is the computation function of the saliency map. We have the following relationship:" + }, + { + "type": "equation", + "bbox": [ + 0.665, + 0.692, + 0.907, + 0.723 + ], + "angle": 0, + "content": "\\[\n\\frac {\\partial \\delta_ {0}}{\\partial x} \\leq U \\cdot s, \\tag {5}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.733, + 0.741, + 0.747 + ], + "angle": 0, + "content": "where \\( U > 0 \\) is a bound constant." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.758, + 0.907, + 0.807 + ], + "angle": 0, + "content": "Theorem 2 suggests that the saliency map provides upper bounds for \\(\\frac{\\partial\\delta_0}{\\partial x}\\). Namely, \\(s\\) provides relaxed descriptions for the variation \\(\\frac{\\partial\\delta_0}{\\partial x}\\). This can help guide the learning of \\(\\frac{\\partial\\delta_0}{\\partial x}\\)." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.808, + 0.909, + 0.902 + ], + "angle": 0, + "content": "GUES framework. Based on the analysis above, we instantiate GUES as the framework depicted in Fig. 2. As shown in sub-figure (a), our method integrates a VAE model and by-pass connection, achieving the learning of \\( F_{\\Phi}\\left(\\frac{\\partial \\delta_0}{\\partial x}\\right) \\). Specifically, \\( \\frac{\\partial \\delta_0}{\\partial x} \\) is sampled from an input \\( x_t \\)-featured Gaussian distribution \\( N(\\mu, \\sigma) \\), which is jointly learned using the" + }, + { + "type": "page_number", + "bbox": [ + 0.479, + 0.945, + 0.521, + 0.957 + ], + "angle": 0, + "content": "28340" + } + ], + [ + { + "type": "title", + "bbox": [ + 0.092, + 0.091, + 0.395, + 0.106 + ], + "angle": 0, + "content": "Algorithm 1 The pipeline of proposed GUES" + }, + { + "type": "text", + "bbox": [ + 0.091, + 0.108, + 0.482, + 0.138 + ], + "angle": 0, + "content": "Input: Online batch samples \\(\\mathcal{B}\\), a trainable \\(\\mathrm{VAE}\\theta_v\\) consists of an encoder \\(E_{\\tau}(\\cdot)\\) and a decoder \\(D(\\cdot)\\)." + }, + { + "type": "title", + "bbox": [ + 0.092, + 0.139, + 0.172, + 0.152 + ], + "angle": 0, + "content": "Procedure:" + }, + { + "type": "text", + "bbox": [ + 0.102, + 0.156, + 0.217, + 0.17 + ], + "angle": 0, + "content": "1: for \\( x_{i} \\) in \\( \\mathcal{B} \\) do" + }, + { + "type": "text", + "bbox": [ + 0.102, + 0.171, + 0.464, + 0.186 + ], + "angle": 0, + "content": "2: Approximate the latent function input by \\( E_{\\tau}(x_i) \\);" + }, + { + "type": "text", + "bbox": [ + 0.102, + 0.187, + 0.478, + 0.201 + ], + "angle": 0, + "content": "3: Learn a generative function for perturbation \\(\\delta_{i}\\) by \\(D\\);" + }, + { + "type": "text", + "bbox": [ + 0.102, + 0.202, + 0.477, + 0.216 + ], + "angle": 0, + "content": "4: Calculate individual perturbations \\(\\delta_{i}\\) by \\(D(E_{\\tau}(x))\\)" + }, + { + "type": "text", + "bbox": [ + 0.102, + 0.217, + 0.482, + 0.247 + ], + "angle": 0, + "content": "5: Create the unadversarial example \\(\\hat{x}_i\\) by incorporating \\(\\delta_i\\) and \\(x_i\\) through the bypass path." + }, + { + "type": "text", + "bbox": [ + 0.102, + 0.248, + 0.464, + 0.262 + ], + "angle": 0, + "content": "6: Generate a fine-grained saliency map \\( g_{i} \\) of the \\( x_{i} \\);" + }, + { + "type": "text", + "bbox": [ + 0.102, + 0.263, + 0.462, + 0.277 + ], + "angle": 0, + "content": "7: Update \\(\\theta_v\\) with Eq.(8), taking \\(g_t\\) as a supervision." + }, + { + "type": "list", + "bbox": [ + 0.102, + 0.156, + 0.482, + 0.277 + ], + "angle": 0, + "content": null + }, + { + "type": "text", + "bbox": [ + 0.102, + 0.278, + 0.174, + 0.29 + ], + "angle": 0, + "content": "8: end for" + }, + { + "type": "text", + "bbox": [ + 0.102, + 0.292, + 0.439, + 0.307 + ], + "angle": 0, + "content": "9: return The generative unadversarial examples \\(\\hat{x}\\)" + }, + { + "type": "list", + "bbox": [ + 0.102, + 0.278, + 0.439, + 0.307 + ], + "angle": 0, + "content": null + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.34, + 0.484, + 0.446 + ], + "angle": 0, + "content": "encoder and reparameterization. The decoder, representing \\( F_{\\Phi} \\), then transforms \\( \\frac{\\partial \\delta_0}{\\partial x} \\) into a generative perturbation \\( \\delta_t \\). Finally, the by-pass structure incorporates \\( x_t \\) and \\( \\delta_t \\) to produce \\( \\hat{x}_t \\). As shown in Fig. 2 (b), during the inference phase, the trained GUES model generates the corresponding generative unadversarial examples for the unseen frozen or fine-tuned model from specific testing samples." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.449, + 0.484, + 0.541 + ], + "angle": 0, + "content": "Loss function. The loss function for GUES training consists of two components. First, we enforce the latent space with mean \\(\\mu\\) and variance \\(\\sigma\\) satisfy the standard normal distribution \\(\\mathcal{N}(0,I)\\). Suppose the encoder in VAE models the posterior distribution \\(q(z|x_{t}) = \\mathcal{N}(\\mu (x_{t}),\\sigma^{2}(x_{t}))\\), this regularization can be formulated as:" + }, + { + "type": "equation", + "bbox": [ + 0.168, + 0.554, + 0.484, + 0.571 + ], + "angle": 0, + "content": "\\[\nL _ {\\mathrm {K L}} = D _ {\\mathrm {K L}} \\left(q (z | x _ {t}) \\| \\mathcal {N} (0, I)\\right), \\tag {6}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.584, + 0.485, + 0.645 + ], + "angle": 0, + "content": "where function \\(D_{\\mathrm{KL}}\\) computes the Kullback-Leibler divergence. The other reconstruction loss between the unadversarial example \\(\\hat{x}_t\\) and saliency map \\(g_t\\) is presented by the following regression form:" + }, + { + "type": "equation", + "bbox": [ + 0.209, + 0.658, + 0.483, + 0.674 + ], + "angle": 0, + "content": "\\[\nL _ {\\mathrm {M S E}} = \\left\\| \\hat {x} _ {t} - g _ {t} \\right\\| _ {2}. \\tag {7}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.687, + 0.483, + 0.717 + ], + "angle": 0, + "content": "Formally, combining Eq. (6) and Eq. (7), the final objective of GUES can be summarized as:" + }, + { + "type": "equation", + "bbox": [ + 0.199, + 0.732, + 0.483, + 0.747 + ], + "angle": 0, + "content": "\\[\nL _ {\\text {G U E S}} = \\alpha L _ {\\mathrm {K L}} + \\beta L _ {\\mathrm {M S E}}, \\tag {8}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.761, + 0.483, + 0.79 + ], + "angle": 0, + "content": "where \\(\\alpha\\) and \\(\\beta\\) are trade-off parameters. For clarity, we summarize the training procedure of GUES in Algorithm 1." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.791, + 0.484, + 0.902 + ], + "angle": 0, + "content": "In Eq. (8), the first item \\( L_{\\mathrm{KL}} \\) ensures the learning of \\( \\frac{\\partial \\delta_0}{\\partial x} \\). On the one hand, as aforementioned, we use \\( z \\) to present \\( \\frac{\\partial \\delta_0}{\\partial x} \\). \\( L_{\\mathrm{KL}} \\) aligns the \\( z \\) space with \\( \\mathcal{N}(0, I) \\), thereby linking the random noise to \\( \\frac{\\partial \\delta_0}{\\partial x} \\). On the other hand, \\( q(z | x_t) \\) in \\( L_{\\mathrm{KL}} \\) is a function of input \\( x_t \\), building relationship \\( x_t \\) to \\( \\frac{\\partial \\delta_0}{\\partial x} \\). Additionally, the reconstruction regulated by the seconded item \\( L_{\\mathrm{MSE}} \\) encourage \\( \\delta_k \\rightarrow \\hat{\\delta} \\)." + }, + { + "type": "title", + "bbox": [ + 0.514, + 0.09, + 0.645, + 0.107 + ], + "angle": 0, + "content": "4. Experiments" + }, + { + "type": "title", + "bbox": [ + 0.514, + 0.115, + 0.616, + 0.13 + ], + "angle": 0, + "content": "4.1. Datasets" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.139, + 0.907, + 0.336 + ], + "angle": 0, + "content": "We perform evaluation experiments on four existing fundus benchmarks, including APTOS [1], DDR [13], DeepDR [17], and Messidor-2 (termed MD2) [7]. Those datasets share five grading/classes: no DR, mild DR, moderate DR, severe DR, and proliferative DR. Taking each dataset as a separate domain, we form 12 transfer tasks crossing domains. For example, as APTOS is the source domain while the others are target domains, we have three transfer tasks APTOS \\(\\rightarrow\\) DDR, APTOS \\(\\rightarrow\\) DeepDR, and APTOS \\(\\rightarrow\\) MD2. The illustration of the label distribution and the domain shift of them is demonstrated in Supplementary. Of Note, all datasets have a severe class imbalance (e.g., \"no DR\" class itself takes up to \\(45.8\\%\\) of the DDR dataset)." + }, + { + "type": "title", + "bbox": [ + 0.514, + 0.347, + 0.724, + 0.363 + ], + "angle": 0, + "content": "4.2. Implementation Detail" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.374, + 0.908, + 0.465 + ], + "angle": 0, + "content": "Source model pre-training. We adopt the DeiT-base network [32] as the backbone of the source pre-trained model, training it in a supervised manner using the source data and corresponding ground truths. During this source training phase, the adopted objective is the classic cross-entropy loss with label smoothing, the same as other methods [15, 37, 38]." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.469, + 0.906, + 0.559 + ], + "angle": 0, + "content": "Variational autoencoder setting. The VAE model is an eight-layer convolutional architecture with a latent space dimension of 10. We do not employ a pre-trained VAE and utilize a VAE without fine-tuning on any other dataset, ensuring that the learning component \\( F_{\\Phi}\\left(\\frac{\\partial\\delta_0}{\\partial x}\\right) \\) are unbiased and independent of prior pre-training data." + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.56, + 0.907, + 0.605 + ], + "angle": 0, + "content": "Parameter setting. For the trade-off parameters in Eq. (8), we set \\(\\alpha\\) to 1.0, while \\(\\beta\\) is tuned with \\(\\{0.0001, 0.01, 1\\}\\) to ensure that \\(L_{\\mathrm{KL}}\\) and \\(L_{\\mathrm{MSE}}\\) remain on the same values scale." + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.606, + 0.906, + 0.681 + ], + "angle": 0, + "content": "Training setting. For fairness, all compared methods were implemented with the same DeiT-base backbone. We adopt the batch size of 64, SGD optimizer with a momentum of 0.9 and a learning rate of 1e-5 on all datasets. All experiments are conducted with PyTorch on a single GPU of RTX A6000." + }, + { + "type": "title", + "bbox": [ + 0.514, + 0.692, + 0.711, + 0.708 + ], + "angle": 0, + "content": "4.3. Comparison Settings" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.716, + 0.907, + 0.792 + ], + "angle": 0, + "content": "Evaluation metrics. To account for unbalanced datasets, besides conventional classification accuracy (termed ACC), we adopt the measure of Quadratic Weighted Kappa (termed QWK) [2] and the average of QWK and ACC (termed AVG). Their rules are provided in Supplementary." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.795, + 0.907, + 0.901 + ], + "angle": 0, + "content": "Competitors. We compare GUES with nine existing state-of-the-art adaptation methods divided into three groups. (1) The first group involves applying the source model directly to the target domain. (2) The second group includes five SFDA methods SHOT [15], NRC [38], CoWA [12], PLUE [16], and TPDS [29]. (3) The third group comprises three typical TTA methods: SHOT-IM [15], TENT [34], and SAR [21]." + }, + { + "type": "page_number", + "bbox": [ + 0.479, + 0.945, + 0.518, + 0.957 + ], + "angle": 0, + "content": "28341" + } + ], + [ + { + "type": "table_caption", + "bbox": [ + 0.09, + 0.089, + 0.908, + 0.119 + ], + "angle": 0, + "content": "Table 2. The results of Source, SFDA, TTA, OMG-DA, and OMG-DA combination methods on datasets APTOS, DeepDR, DDR, and MD2 are presented. The improvements over baseline methods Source, SHOT-IM, and TENT are highlighted as (+x.x)." + }, + { + "type": "table", + "bbox": [ + 0.113, + 0.129, + 0.885, + 0.466 + ], + "angle": 0, + "content": "
MethodVenueAPTOS→DDRAPTOS→DeepDRAPTOS→MD2DDR→APTOSDDR→DeepDRDDR→MD2DeepDR→APTOS
ACCQWKAVGACCQWKAVGACCQWKAVGACCQWKAVGACCQWKAVGACCQWKAVGACCQWKAVG
Source-60.659.259.952.671.762.160.948.754.865.672.969.345.060.652.849.434.642.043.471.657.5
GUES + Source-62.059.560.853.069.761.359.846.753.376.081.878.956.468.762.559.147.653.346.574.160.3
SHOT [15]ICML2066.969.067.953.673.563.651.738.044.877.084.280.659.274.666.957.143.150.162.382.572.4
NRC [38]NeurIPS2161.965.263.551.670.961.354.140.347.260.376.368.352.069.160.550.635.242.952.074.363.2
CoWA [12]ICML2259.064.962.051.070.760.853.037.345.157.274.565.850.166.858.453.139.646.350.473.061.7
PLUE [16]CVPR2362.065.163.651.369.760.554.541.147.863.464.263.854.364.859.551.628.239.954.670.562.6
TPDS [29]IICV2466.667.867.251.671.561.552.740.746.776.683.980.258.073.165.654.541.047.860.880.070.4
SHOT-IM [15]ICML2066.569.267.952.673.663.153.236.644.975.982.179.058.573.966.257.343.550.461.984.072.9
TENT [34]ICLR2059.950.255.153.170.161.661.448.454.975.282.478.855.168.461.760.850.555.660.279.769.9
SAR [21]ICLR2367.963.865.853.673.063.357.244.951.175.683.579.555.671.663.649.237.043.159.579.369.4
GUES + SHOT-IM - GUES + TENT -68.668.568.553.572.863.255.743.849.777.283.180.260.575.167.861.551.256.362.683.272.9
-61.856.359.053.270.061.661.147.254.175.983.079.458.770.864.763.353.858.654.977.666.3
MethodVenueDeepDR→DDRDeepDR→MD2MD2→APTOSMD2→DDRMD2→DeepDRAccAvg. QWKAVG
ACCQWKAVGACCQWKAVGACCQWKAVGACCQWKAVGACCQWKAVG
Source-56.466.961.748.750.249.443.970.357.160.256.558.359.858.459.153.960.157.0
GUES + Source-57.365.661.448.352.050.159.376.267.764.755.860.258.657.157.858.4 (+4.5)62.9 (+2.8)60.7 (+3.7)
SHOT [15]ICML2057.471.264.348.841.545.252.773.062.854.659.256.959.670.264.958.465.061.7
NRC [38]NeurIPS2144.960.952.949.841.845.848.869.158.952.852.852.858.062.860.453.159.956.5
CoWA [12]ICML2248.758.453.649.942.746.351.070.160.549.650.950.357.660.659.156.961.959.4
PLUE [16]CVPR2347.253.550.456.447.451.956.069.162.656.554.355.458.864.861.855.557.756.6
TPDS [29]IICV2459.369.464.350.542.446.460.374.967.660.060.460.258.963.060.959.264.061.6
SHOT-IM [15]ICML2054.669.462.051.238.244.761.677.969.757.058.757.957.569.863.759.064.761.9
TENT [34]ICLR2058.545.451.958.356.557.455.174.164.655.831.743.758.053.655.859.359.259.3
SAR [21]ICLR2353.066.359.642.633.137.955.273.064.149.748.349.056.765.861.356.361.659.0
GUES + SHOT-IM - GUES + TENT --62.371.566.952.847.850.362.878.170.466.059.462.760.668.664.662.0 (+3.0)66.9 (+2.2)64.5 (+2.6)
-62.661.462.059.157.458.259.375.667.564.050.557.258.356.057.161.0 (+1.7)63.3 (+4.1)62.2 (+2.9)
" + }, + { + "type": "image", + "bbox": [ + 0.099, + 0.489, + 0.476, + 0.584 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.09, + 0.592, + 0.486, + 0.637 + ], + "angle": 0, + "content": "Figure 4. Comparison results with batch size varying from 2 to 64 over the 12 tasks (The details are provided in Supplementary.). Left, middle, and right report ACC, QWK, and AVG, respectively." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.666, + 0.486, + 0.711 + ], + "angle": 0, + "content": "Comparison protocol in OMG-DA setting. For a comprehensive comparison, our comparison follows two different fashions as follows." + }, + { + "type": "text", + "bbox": [ + 0.091, + 0.714, + 0.483, + 0.759 + ], + "angle": 0, + "content": "- Case without training: We first generate the unadversarial examples for the target domain by the trained GUES model and then provide them to the frozen source model." + }, + { + "type": "text", + "bbox": [ + 0.091, + 0.76, + 0.483, + 0.805 + ], + "angle": 0, + "content": "- Case with training: We plug GUES into other TTA methods (they are also online methods with flowing data) as online image pre-processing." + }, + { + "type": "list", + "bbox": [ + 0.091, + 0.714, + 0.483, + 0.805 + ], + "angle": 0, + "content": null + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.808, + 0.483, + 0.868 + ], + "angle": 0, + "content": "The two cases evaluate GUES from different aspects. The first isolates the generalization ability of the unadversarial examples generated by GUES, whilst the second highlights GUES's compatibility with other trainable online schemes." + }, + { + "type": "text", + "bbox": [ + 0.091, + 0.871, + 0.486, + 0.901 + ], + "angle": 0, + "content": "Corresponding to the comparison protocols above, besides the version GUES corresponding to the case with" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.491, + 0.907, + 0.522 + ], + "angle": 0, + "content": "out training, we also introduce GUES+SHOT-IM and GUES+TENT which correspond to the case with training." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.534, + 0.707, + 0.551 + ], + "angle": 0, + "content": "4.4. Comparison Results" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.558, + 0.907, + 0.604 + ], + "angle": 0, + "content": "This part presents the comparison results following the cases above. Also, considering batch size a crucial factor for TTA methods, the results as the batch size varies are provided." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.608, + 0.906, + 0.714 + ], + "angle": 0, + "content": "Results without training. The comparisons are shown in Tab. 2. On average, across the 12 tasks and without training of the source model, GUES achieves improvements of \\(4.5\\%\\) in ACC, \\(2.8\\%\\) in QWK, and \\(3.7\\%\\) in AVG compared to the source model. These results demonstrate that GUES modifies the target data distribution effectively, adapting the target domain to align with the source domain." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.718, + 0.907, + 0.854 + ], + "angle": 0, + "content": "Results with training. As shown in Tab. 2, GUES+SHOT-IM outperforms the previous best SFDA and TTA methods, respectively surpassing TENT in ACC by \\(2.7\\%\\), SHOT in QWK by \\(1.9\\%\\), and SHOT-IM in AVG by \\(2.6\\%\\). Compared to SHOT-IM, GUES+SHOT-IM gains over \\(3.0\\%\\) in ACC, \\(2.2\\%\\) in QWK, and \\(2.6\\%\\) in AVG. Similarly, GUES+TENT improves over TENT by \\(1.7\\%\\) in ACC, \\(4.1\\%\\) in QWK, and \\(2.9\\%\\) in AVG. These results highlight the effectiveness of combining GUES with other methods that require training." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.856, + 0.909, + 0.903 + ], + "angle": 0, + "content": "Results with varying batch size. This part isolates the effect of batch size, which is a crucial factor for TTA methods. Fig. 4 depicts the performance variation as batch size vary" + }, + { + "type": "page_number", + "bbox": [ + 0.479, + 0.945, + 0.521, + 0.957 + ], + "angle": 0, + "content": "28342" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.134, + 0.089, + 0.868, + 0.199 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.09, + 0.208, + 0.908, + 0.252 + ], + "angle": 0, + "content": "Figure 5. Interpretability analysis based on a typical fundus image from \"Moderate DR\" class in APTOS. Here, H (hemorrhages), SE (soft exudates), and EX (hard exudates) are essential characteristics to judge the DR grade. The gradient CAM-based heatmap of five models visualizes the capture of those lesions. All models are trained on task DDR \\(\\rightarrow\\) APTOS, where Oracle is trained using ground truth in APTOS." + }, + { + "type": "image", + "bbox": [ + 0.102, + 0.274, + 0.473, + 0.37 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.09, + 0.377, + 0.483, + 0.406 + ], + "angle": 0, + "content": "Figure 6. Feature distribution comparison of 3D density charts on task DDR \\(\\rightarrow\\) APTOS. Oracle is trained on APTOS by ground truth." + }, + { + "type": "text", + "bbox": [ + 0.089, + 0.431, + 0.483, + 0.627 + ], + "angle": 0, + "content": "ing from 2 to 64 over the 12 tasks. It is observed that TTA methods SHOT-IM and TENT suffer from severe performance drops when the batch size becomes small. SHOT-IM exhibits a decrease of approximately \\(16\\%\\) in ACC when the batch size is reduced from 64 to 2, whilst TENT shows a substantial decline of around \\(34\\%\\) in QWK. Oppose to it, the methods with GUES, SHOT-IM+GUES, and TENT+GUES, do not have evident performance decline at the smaller batch size. Moreover, this combination not only mitigates the drop but also shows improvements when the batch size is 64. This indicates that GUES effectively stabilizes the performance of SHOT-IM and TENT, enhancing their robustness to variations in batch size while boosting their overall effectiveness." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.628, + 0.483, + 0.763 + ], + "angle": 0, + "content": "We attribute GUES's excellent robustness to its ability to predict individual perturbations (see Supplementary for more details on the visualization of perturbations) that focus on single image-specific features rather than global data commonalities. For instance, the conventional unadversarial examples approach refines the class-specific perturbation sensitive to batch size. Furthermore, both SHOT-IM and TENT are entropy-based methods that require large-scale batch size for accurate entropy estimation." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.772, + 0.299, + 0.789 + ], + "angle": 0, + "content": "4.5. Visualization Analysis" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.795, + 0.486, + 0.901 + ], + "angle": 0, + "content": "Interpretability. For a better understanding, Fig. 5 demonstrates whether GUES can help capture pathologically relevant features, such as H (hemorrhages), SE (soft exudates), and EX (hard exudates), determining DR grade. First of all, when comparing the source model with GUES, the source model only captures a limited area of the lesion, while GUES effectively captures most of the DR-related features. Further-" + }, + { + "type": "image", + "bbox": [ + 0.518, + 0.276, + 0.905, + 0.407 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.513, + 0.416, + 0.908, + 0.458 + ], + "angle": 0, + "content": "Figure 7. Unadversarial examples visualization of two typical target samples from APTOS. The generative perturbations are generated by the GUES model trained on task DDR \\(\\rightarrow\\) APTOS." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.485, + 0.908, + 0.576 + ], + "angle": 0, + "content": "more, combining GUES with SHOT-IM (i.e., GUES+SHOT-IM) expands the focus on DR-related features beyond those captured by SHOT-IM alone. Additionally, when comparing the four models to Oracle, only GUES and GUES+SHOT-IM resemble Oracle, suggesting that GUES effectively directs the model's attention to DR-critical features." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.58, + 0.909, + 0.777 + ], + "angle": 0, + "content": "Feature distribution. Taking the task DDR→APTOS as a toy experiment, we visualize the feature distribution extracted from the final convolutional layer of the prediction model using a 3D density chart. Considering that APTOS is a class-imbalanced dataset, with the “No DR” class alone comprising up to \\(49.2\\%\\) of the dataset, our analysis focuses on this crucial property. As shown in Fig. 6, the feature distribution of the source model does not reflect this imbalanced characteristic; instead, it displays a more uniform classification. Conversely, the feature distribution of GUES exhibits a distinct imbalance, with one expanded high-density region alongside several smaller high-density regions, resembling the distribution pattern seen in Oracle." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.78, + 0.911, + 0.902 + ], + "angle": 0, + "content": "Visualization of unadversarial examples. This part visualizes unadversarial examples of two typical target samples from APTOS and corresponding generative perturbations, based on the task DDR \\(\\rightarrow\\) APTOS. Considering the generative perturbations alter the original images that may not be easily visible to the naked eye, we collect RGB statistics to illustrate these changes quantitatively. It is observed that in Fig. 7, each channel (R, G, and B) exhibits notable" + }, + { + "type": "page_number", + "bbox": [ + 0.479, + 0.945, + 0.521, + 0.958 + ], + "angle": 0, + "content": "28343" + } + ], + [ + { + "type": "table_caption", + "bbox": [ + 0.157, + 0.09, + 0.416, + 0.104 + ], + "angle": 0, + "content": "Table 3. ACC results of ablation study \\((\\%)\\)" + }, + { + "type": "table", + "bbox": [ + 0.1, + 0.115, + 0.472, + 0.224 + ], + "angle": 0, + "content": "
#\\( L_{\\text{KL}} \\)\\( L_{\\text{MSE}} \\)APTOSDDRDeepDRMD2Avg.
1XX51.059.152.553.053.9
2✓X54.758.354.953.355.3
3X✓56.960.152.952.955.7
4✓✓60.661.356.055.758.4
5GUES w/ AE53.860.253.853.255.3
6GUES w/ Mixup56.860.253.052.855.7
7GUES w/ Self56.560.453.754.456.3
8GUES w/ Sal46.134.341.845.942.0
" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.255, + 0.483, + 0.36 + ], + "angle": 0, + "content": "fluctuations, with the RGB statistics of the original images (a) differing significantly from those of the unadversarial examples (c). Additionally, each generative perturbation is unique, meaning that the alterations introduced by these perturbations are individual. These results suggest that the perturbations may help highlight critical DR-related features, refining the model's focus on diagnostically relevant areas." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.373, + 0.258, + 0.388 + ], + "angle": 0, + "content": "4.6. Further Analysis" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.397, + 0.484, + 0.563 + ], + "angle": 0, + "content": "Ablation study. In this part, we evaluate the effect of objective loss, as well as the components involved in GUES including the sampling strategy and saliency map-based supervision. To address the first issue, we conduct a progressive experiment. The top four rows of Tab. 3 list the ablation results where the source model's performance is the baseline. Using \\( L_{\\mathrm{KL}} \\) or \\( L_{\\mathrm{MSE}} \\) alone yields an average ACC improvement of approximately \\( 1.4\\% \\) and \\( 1.8\\% \\), respectively, over the baseline. As both of them work, the ACC increases \\( 3.5\\% \\) on average further. The results indicate that all objective components positively affect the final performance." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.565, + 0.484, + 0.791 + ], + "angle": 0, + "content": "To evaluate the impact of the sampling component, we propose a variation method of GUES, GUES w/ AE, where we remove this sampling process by replacing VAE with a conventional Auto-encoder model. In addition, two GUES variations are used to assess the advantage of saliency map-based supervision. Specifically, GUES w/ Mixup replaces the saliency maps with a Mixup of saliency maps and original images, whilst GUES w/ Self replaces the saliency maps with the original images. As presented in the Tab. 3, compared with the full version of GUES (the fourth row), GUES w/ AE, GUES w/ Mixup and GUES w/ Self decrease by \\(3.1\\%\\) at least on average. Besides, replacing the original images with saliency maps as inputs (GUES w/ Sal) leads to a significant drop of about \\(15.4\\%\\). Above these experiments confirm the effectiveness of our design choices." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.795, + 0.484, + 0.901 + ], + "angle": 0, + "content": "Parameter sensitiveness. Taking the task DDR \\(\\rightarrow\\) APTOS as a toy experiment, we present the GUES performance varying as hyper-parameters \\(0.5\\leq \\alpha \\leq 1.5\\) with 0.1 steps, \\(0.00005\\leq \\beta \\leq 0.00014\\) with 0.00001 steps. As depicted in Fig. 8, the ACC, QWK, and AVG variation surfaces show fluctuations in the tiny performance zone, with approximately \\(0.2\\%\\) in ACC, \\(0.25\\%\\) in QWK, and \\(0.4\\%\\) in" + }, + { + "type": "image", + "bbox": [ + 0.527, + 0.09, + 0.646, + 0.181 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.652, + 0.09, + 0.771, + 0.181 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.776, + 0.09, + 0.895, + 0.181 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.513, + 0.192, + 0.907, + 0.234 + ], + "angle": 0, + "content": "Figure 8. Parameter sensitiveness study results over \\(\\alpha \\times \\beta\\) based on task DDR \\(\\rightarrow\\) APTOS. From left to right, there are results on ACC, QWK, and AVG, respectively." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.26, + 0.905, + 0.29 + ], + "angle": 0, + "content": "AVG. This observation suggests that GUES is insensitive to alterations in \\(\\alpha\\) and \\(\\beta\\)." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.294, + 0.908, + 0.506 + ], + "angle": 0, + "content": "Limitation. GUES uses the saliency map to guide the learning of the generative function. This method is effective for DR grading but encounters challenges in natural image scenarios. The natural images contain rich semantics, such as shape, relative structure, and complex background, which are not all relevant to tasks. However, the saliency map blindly highlights all those factors, struggling to capture the task-specific ones. In a theoretical point-view, the richness of these semantics means significant variations, resulting in a super-relaxed bound constant \\( U \\) (see Theorem 2) that undermines the descriptive power of the saliency map for \\( \\frac{\\partial \\delta_0}{\\partial x} \\). In contrast, fundus images are more monolithic, implying a smaller \\( U \\) that justifies the usage of the saliency map. (The further discussion is provided in Supplementary)" + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.52, + 0.633, + 0.535 + ], + "angle": 0, + "content": "5. Conclusion" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.545, + 0.909, + 0.832 + ], + "angle": 0, + "content": "In this paper, we propose a clinically motivated setting, OMG-DA, where the models are unseen prior to their use, and only target data flows are accessible. This setting ensures both model protection and source data privacy in a data flow scenario. To adapt to the target domain without access to the models, we introduce a GUES approach. Instead of conventional iterative optimization, we generate unadversarial examples for flowing target data by directly predicting individual perturbations. This approach is grounded in the theoretical results of generative unadversarial learning. In practice, we utilize the VAE model to learn the perturbation generation function with a latent input variable. Furthermore, we demonstrate that saliency maps can serve as an upper bound for this latent variable. This relationship inspires us to use the saliency maps as pseudo-perturbation labels for model training. Extensive experiments conducted on four DR benchmarks confirm that the proposed method can achieve state-of-the-art results when it pairs with both frozen pre-trained and fine-tuning models." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.846, + 0.673, + 0.862 + ], + "angle": 0, + "content": "Acknowledgments" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.871, + 0.906, + 0.901 + ], + "angle": 0, + "content": "This work is supported by the National Natural Science Foundation of China (62476169, 62206168, 62276048)." + }, + { + "type": "page_number", + "bbox": [ + 0.48, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "28344" + } + ], + [ + { + "type": "title", + "bbox": [ + 0.093, + 0.09, + 0.188, + 0.106 + ], + "angle": 0, + "content": "References" + }, + { + "type": "ref_text", + "bbox": [ + 0.101, + 0.116, + 0.486, + 0.157 + ], + "angle": 0, + "content": "[1] Aptos: Aptos 2019 blindness detection website. https://www.kaggle.com/c/aptos2019-blindness-detection, accessed February 20, 2022." + }, + { + "type": "ref_text", + "bbox": [ + 0.101, + 0.16, + 0.486, + 0.213 + ], + "angle": 0, + "content": "[2] Quadratic weighted kappa. https://www.Eyepacs.com/aroraaman/quadratic-kappa-metric-explained-in-5-simple-steps, accessed July 2022." + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.216, + 0.484, + 0.285 + ], + "angle": 0, + "content": "[3] Eman AbdelMaksoud, Sherif Barakat, and Mohammed Elmogy. A comprehensive diagnosis system for early signs and different diabetic retinopathy grades using fundus retinal images based on pathological changes detection. Computers in Biology and Medicine, 126:104039, 2020." + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.287, + 0.484, + 0.327 + ], + "angle": 0, + "content": "[4] Mohammad Atwany and Mohammad Yaqub. Drogen: domain generalization in diabetic retinopathy classification. In MICCAI, 2022." + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.33, + 0.484, + 0.371 + ], + "angle": 0, + "content": "[5] Haoxuan Che, Yuhan Cheng, Haibo Jin, and Hao Chen. Towards generalizable diabetic retinopathy grading in unseen domains. In MICCAI, 2023." + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.374, + 0.484, + 0.442 + ], + "angle": 0, + "content": "[6] Ling Dai, Liang Wu, Huating Li, Chun Cai, Qiang Wu, Hongyu Kong, Ruhan Liu, Xiangning Wang, Xuhong Hou, Yuexing Liu, et al. A deep learning system for detecting diabetic retinopathy across the disease spectrum. Nature communications, 12(1):3242, 2021." + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.445, + 0.484, + 0.526 + ], + "angle": 0, + "content": "[7] Etienne Decencière, Xiwei Zhang, Guy Cazuguel, Bruno Lay, Béatrice Cochener, Caroline Trone, Philippe Gain, John-Richard Ordóñez-Varela, Pascale Massin, Ali Erginay, et al. Feedback on a publicly distributed image database: the messidor database. Image Analysis & Stereology, pages 231-234, 2014." + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.529, + 0.484, + 0.583 + ], + "angle": 0, + "content": "[8] Along He, Tao Li, Ning Li, Kai Wang, and Huazhu Fu. Cabinet: Category attention block for imbalanced diabetic retinopathy grading. IEEE Transactions on Medical Imaging, 40(1):143-153, 2020." + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.586, + 0.484, + 0.641 + ], + "angle": 0, + "content": "[9] Yijin Huang, Junyan Lyu, Pujin Cheng, Roger Tam, and Xiaoying Tang. Ssit: Saliency-guided self-supervised image transformer for diabetic retinopathy grading. IEEE Journal of Biomedical and Health Informatics, 2024." + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.643, + 0.484, + 0.67 + ], + "angle": 0, + "content": "[10] Diederik P Kingma. Auto-encoding variational bayes. arXiv:1312.6114, 2013." + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.673, + 0.482, + 0.699 + ], + "angle": 0, + "content": "[11] Wouter M Kouw and Marco Loog. An introduction to domain adaptation and transfer learning. arXiv:1812.11806, 2018." + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.702, + 0.484, + 0.742 + ], + "angle": 0, + "content": "[12] Jonghyun Lee, Dahuin Jung, Junho Yim, and Sungroh Yoon. Confidence score for source-free unsupervised domain adaptation. In ICML, 2022." + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.746, + 0.484, + 0.8 + ], + "angle": 0, + "content": "[13] Tao Li, Yingqi Gao, Kai Wang, Song Guo, Hanruo Liu, and Hong Kang. Diagnostic assessment of deep learning algorithms for diabetic retinopathy screening. Information Sciences, 501:511-522, 2019." + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.803, + 0.484, + 0.857 + ], + "angle": 0, + "content": "[14] Tao Li, Wang Bo, Chunyu Hu, Hong Kang, Hanruo Liu, Kai Wang, and Huazhu Fu. Applications of deep learning in fundus images: A review. Medical Image Analysis, 69: 101971, 2021." + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.86, + 0.484, + 0.901 + ], + "angle": 0, + "content": "[15] Jian Liang, Dapeng Hu, and Jiashi Feng. Do we really need to access the source data? source hypothesis transfer for unsupervised domain adaptation. In ICML, 2020." + }, + { + "type": "list", + "bbox": [ + 0.094, + 0.116, + 0.486, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.093, + 0.907, + 0.135 + ], + "angle": 0, + "content": "[16] Mattia Litrico, Alessio Del Bue, and Pietro Morerio. Guiding pseudo-labels with uncertainty estimation for source-free unsupervised domain adaptation. In CVPR, 2023." + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.137, + 0.907, + 0.192 + ], + "angle": 0, + "content": "[17] Ruhan Liu, Xiangning Wang, Qiang Wu, Ling Dai, Xi Fang, Tao Yan, Jaemin Son, Shiqi Tang, Jiang Li, Zijian Gao, et al. Deepgrid: Diabetic retinopathy—grading and image quality estimation challenge. *Patterns*, 3(6), 2022." + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.195, + 0.907, + 0.235 + ], + "angle": 0, + "content": "[18] Xingbin Liu, Huafeng Kuang, Xianming Lin, Yongjian Wu, and Rongrong Ji. Cat: Collaborative adversarial training. arXiv:2303.14922, 2023." + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.238, + 0.906, + 0.292 + ], + "angle": 0, + "content": "[19] Sebastian Montabone and Alvaro Soto. Human detection using a mobile platform and novel features derived from a visual saliency mechanism. Image and Vision Computing, 28 (3):391-402, 2010." + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.296, + 0.907, + 0.35 + ], + "angle": 0, + "content": "[20] Duy MH Nguyen, Truong TN Mai, Ngoc TT Than, Alexander Prange, and Daniel Sonntag. Self-supervised domain adaptation for diabetic retinopathy grading using vessel image reconstruction. In KI, 2021." + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.354, + 0.906, + 0.395 + ], + "angle": 0, + "content": "[21] Shuaicheng Niu, Jiaxiang Wu, Yifan Zhang, Zhiquan Wen, Yaofo Chen, Peilin Zhao, and Mingkui Tan. Towards stable test-time adaptation in dynamic wild world. In ICLR, 2023." + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.398, + 0.906, + 0.451 + ], + "angle": 0, + "content": "[22] Jiaming Qiu, Weikai Huang, Yijin Huang, Nanxi Yu, and Xiaoying Tang. Augpaste: A one-shot approach for diabetic retinopathy detection. Biomedical Signal Processing and Control, 96:106489, 2024." + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.455, + 0.906, + 0.522 + ], + "angle": 0, + "content": "[23] Jinye Ran, Guanghua Zhang, Fan Xia, Ximei Zhang, Juan Xie, and Hao Zhang. Source-free active domain adaptation for diabetic retinopathy grading based on ultra-wide-field fundus images. Computers in Biology and Medicine, 174:108418, 2024." + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.526, + 0.907, + 0.58 + ], + "angle": 0, + "content": "[24] Hadi Salman, Andrew Ilyas, Logan Engstrom, Sai Vemprala, Aleksander Madry, and Ashish Kapoor. Unadversarial examples: Designing objects for robust vision. In NeurIPS, 2021." + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.583, + 0.906, + 0.624 + ], + "angle": 0, + "content": "[25] Abhijith Sharma, Phil Munz, and Apurva Narayan. Nsa: Naturalistic support artifact to boost network confidence. In IJCNN, 2023." + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.627, + 0.906, + 0.668 + ], + "angle": 0, + "content": "[26] Reza Shokri, Marco Stronati, Congzheng Song, and Vitaly Shmatikov. Membership inference attacks against machine learning models. In S&P, 2017." + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.671, + 0.907, + 0.712 + ], + "angle": 0, + "content": "[27] Daniel E Singer, David M Nathan, Howard A Fogel, and Andrew P Schachat. Screening for diabetic retinopathy. Annals of Internal Medicine, 116(8):660-671, 1992." + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.715, + 0.907, + 0.741 + ], + "angle": 0, + "content": "[28] C Szegedy. Intriguing properties of neural networks. arXiv:1312.6199, 2013." + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.745, + 0.907, + 0.799 + ], + "angle": 0, + "content": "[29] Song Tang, An Chang, Fabian Zhang, Xiatian Zhu, Mao Ye, and Changshui Zhang. Source-free domain adaptation via target prediction distribution searching. International Journal of Computer Vision, 132(3):654-672, 2024." + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.802, + 0.907, + 0.842 + ], + "angle": 0, + "content": "[30] Song Tang, Wenxin Su, Mao Ye, Jianwei Zhang, and Xiatian Zhu. Unified source-free domain adaptation. arXiv:2403.07601, 2024." + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.846, + 0.907, + 0.9 + ], + "angle": 0, + "content": "[31] Nishtha Tomar, Sushmita Chandel, and Gaurav Bhatnagar. A visual attention-based algorithm for brain tumor detection using an on-center saliency map and a superpixel-based framework. Healthcare Analytics, 5:100323, 2024." + }, + { + "type": "list", + "bbox": [ + 0.517, + 0.093, + 0.907, + 0.9 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.48, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "28345" + } + ], + [ + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.093, + 0.486, + 0.147 + ], + "angle": 0, + "content": "[32] Hugo Touvron, Matthieu Cord, Matthijs Douze, Francisco Massa, Alexandre Sablayrolles, and Hervé Jégou. Training data-efficient image transformers & distillation through attention. In ICML, 2021." + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.149, + 0.483, + 0.19 + ], + "angle": 0, + "content": "[33] Jeya Maria Jose Valanarasu, Pengfei Guo, VS Vibashan, and Vishal M Patel. On-the-fly test-time adaptation for medical image segmentation. In MIDL, 2024." + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.192, + 0.483, + 0.234 + ], + "angle": 0, + "content": "[34] Dequan Wang, Evan Shelhamer, Shaoteng Liu, Bruno Olshausen, and Trevor Darrell. Tent: Fully test-time adaptation by entropy minimization. In ICLR, 2020." + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.235, + 0.483, + 0.289 + ], + "angle": 0, + "content": "[35] Tianyunxi Wei, Yijin Huang, Li Lin, Pujin Cheng, Sirui Li, and Xiaoying Tang. Saliency-guided and patch-based mixup for long-tailed skin cancer image classification. arXiv:2406.10801, 2024." + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.291, + 0.483, + 0.359 + ], + "angle": 0, + "content": "[36] Zhan Wu, Gonglei Shi, Yang Chen, Fei Shi, Xinjian Chen, Gouenou Coatieux, Jian Yang, Limin Luo, and Shuo Li. Coarse-to-fine classification for diabetic retinopathy grading using convolutional neural network. Artificial Intelligence in Medicine, 108:101936, 2020." + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.362, + 0.483, + 0.403 + ], + "angle": 0, + "content": "[37] Tongkun Xu, Weihua Chen, WANG Pichao, Fan Wang, Hao Li, and Rong Jin. Cdtrans: Cross-domain transformer for unsupervised domain adaptation. In ICLR, 2021." + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.405, + 0.483, + 0.445 + ], + "angle": 0, + "content": "[38] Shiqi Yang, Joost van de Weijer, Luis Herranz, Shangling Jui, et al. Exploiting the intrinsic neighborhood structure for source-free domain adaptation. In NeurIPS, 2021." + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.448, + 0.483, + 0.489 + ], + "angle": 0, + "content": "[39] Hongxu Yin, Arun Mallya, Arash Vahdat, Jose M Alvarez, Jan Kautz, and Pavlo Molchanov. See through gradients: Image batch recovery via grad inversion. In CVPR, 2021." + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.49, + 0.483, + 0.532 + ], + "angle": 0, + "content": "[40] Chenrui Zhang, Tao Lei, and Ping Chen. Diabetic retinopathy grading by a source-free transfer learning approach. Biomedical Signal Processing and Control, 73:103423, 2022." + }, + { + "type": "list", + "bbox": [ + 0.093, + 0.093, + 0.486, + 0.532 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.48, + 0.946, + 0.52, + 0.957 + ], + "angle": 0, + "content": "28346" + } + ] +] \ No newline at end of file diff --git a/2025/Domain Adaptive Diabetic Retinopathy Grading with Model Absence and Flowing Data/8fc9ee64-90f9-4f33-8fb1-64a5f7b319a1_origin.pdf b/2025/Domain Adaptive Diabetic Retinopathy Grading with Model Absence and Flowing Data/8fc9ee64-90f9-4f33-8fb1-64a5f7b319a1_origin.pdf new file mode 100644 index 0000000000000000000000000000000000000000..324086581b7309cd7caff68371741b37b346d577 --- /dev/null +++ b/2025/Domain Adaptive Diabetic Retinopathy Grading with Model Absence and Flowing Data/8fc9ee64-90f9-4f33-8fb1-64a5f7b319a1_origin.pdf @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:6d209e913d7dae878f9fee7e45d206ac120fb36b8266eff64d93fcad9024e6e3 +size 2555526 diff --git a/2025/Domain Adaptive Diabetic Retinopathy Grading with Model Absence and Flowing Data/full.md b/2025/Domain Adaptive Diabetic Retinopathy Grading with Model Absence and Flowing Data/full.md new file mode 100644 index 0000000000000000000000000000000000000000..a64f85923a23a68a5f7d813641ff0043c644f4db --- /dev/null +++ b/2025/Domain Adaptive Diabetic Retinopathy Grading with Model Absence and Flowing Data/full.md @@ -0,0 +1,316 @@ +# Domain Adaptive Diabetic Retinopathy Grading with Model Absence and Flowing Data + +Wenxin $\mathrm{Su}^1$ , Song Tang $^{*1,2,3}$ , Xiaofeng Liu $^4$ , Xiaojing Yi $^5$ , Mao Ye $^6$ , Chunxiao $\mathrm{Zu}^1$ , Jiahao Li $^7$ , and Xiatian Zhu $^8$ + +1University of Shanghai for Science and Technology 2Universität Hamburg, 3ComOriginMat Inc. +4Tale University 5Sichuan Eye Hospital 6University of Electronic Science and Technology of China +7Chinese Academy of Medical Sciences & Peking Union Medical College 8University of Surrey + +steventangsong@gmail.com + +# Abstract + +Domain shift (the difference between source and target domains) poses a significant challenge in clinical applications, e.g., Diabetic Retinopathy (DR) grading. Despite considering certain clinical requirements, like source data privacy, conventional transfer methods are predominantly model-centered and often struggle to prevent model-targeted attacks. In this paper, we address a challenging Online Model-aGnostic Domain Adaptation (OMG-DA) setting, driven by the demands of clinical environments. This setting is characterized by the absence of the model and the flow of target data. To tackle the new challenge, we propose a novel approach, Generative Unadversarial ExampleS (GUES), which enables adaptation from a data-centric perspective. Specifically, we first theoretically reformulate conventional perturbation optimization in a generative way—learning a perturbation generation function with a latent input variable. During model instantiation, we leverage a Variational AutoEncoder to express this function. The encoder with the reparameterization trick predicts the latent input, whilst the decoder is responsible for the generation. Furthermore, the saliency map is selected as pseudoperturbation labels. Because it not only captures potential lesions but also theoretically provides an upper bound on the function input, enabling the identification of the latent variable. Extensive experiments on DR benchmarks with both frozen pre-trained models and trainable models demonstrate the superiority of GUES, showing robustness even with small batch size. The source code and data are available at https://github.com/tntek/GUES. + +![](images/6fbb60f19e595fa5edb64b48acb684f52e615b73cb2d7ef826c77943258bdfd7.jpg) +Figure 1. Comparison between the OMG-DA and SFDA settings. (a) In SFDA, adaptation builds upon the accumulated data, which demands significant storage resources in the hospital. Additionally, the models' architecture and parameters are accessible, exposing them to potential attacks. (b) OMG-DA provides a practical scenario: Flowing data mimic the patients' arrival in a stream way, and the models are unseen (strictly controlled) before using them, avoiding attacks like membership inference attacks [26]. + +# 1. Introduction + +Diabetic Retinopathy (DR) is a significant health concern, ranking among the leading causes of blindness and affecting millions of people worldwide [3]. Early-stage intervention for DR is crucial to preserve vision, highlighting the importance of timely diagnosis [27]. Although deep learning (DL) has demonstrated promising results in automating the grading of DR [6, 8, 36], deploying DL models in real-world clinical settings remains challenging. For example, DL models often struggle to generalize effectively to complex scenarios, such as variations in imaging equipment, ethnic groups, or temporal factors, leading to different data distributions, a challenge known as domain shift [11]. This issue significantly hampers the widespread adoption and success of DL-based diagnostic tools in clinical practice [14]. + +Recently, many adaptation methods for grading diabetic + +Table 1. Comparison between different transfer settings. Notation: source (s), target (t), data $\mathbf{x}$ ,label $\mathbf{y}$ ,loss function $L\left( \cdot \right)$ . + +
Setting nameModel absenceData flowSource data privacySource dataTarget dataTrain lossTest loss
Fine-tuning (FT)XX✓-xt,ytL(xt,yt)-
Unsupervised domain adaptation (UDA)XXXxs,ysxtL(xs,ys)+L(xt)-
Source-free domain adaptation (SFDA)XX✓-xtL(xt)-
Test-time adaptation (TTA)X✓✓-xt-L(xt)
Online model-agnostic domain adaptation (OMG-DA)✓✓✓-xt--
+ +retinopathy (DR) have focused on addressing the issue of domain shift [5, 20, 23, 40]. The initial classic transfer learning strategies, Unsupervised Domain Adaptation (UDA), required access to well-annotated source data. Nevertheless, the growing emphasis on privacy protection has shifted research toward the Source-Free Domain Adaptation (SFDA) framework [23, 40]. SFDA involves adapting a source model—pre-trained on the source domain—to the target domain in a self-supervised manner, thereby ensuring the protection of source patient data. + +In recent developments, specific needs have emerged in the clinical field. The introduction of model weight-based techniques for reconstructing training data has created a demand for model privacy [26, 39], which goes beyond traditional source data protection. In addition, there is a growing requirement for models capable of handling incoming patient data in a flowing fashion (test data arriving sequentially), referred to as a flowing data constraint [33, 34]. Unfortunately, existing SFDA methods cannot effectively address this challenge, as they rely on full access to the model and require offline training on a pre-collected dataset. Fig. 1 provides an intuitive illustration of this issue. + +In this paper, we consider a clinically motivated setting, called Online Model-aGnostic Domain Adaptation (OMG-DA) and propose a novel Generative Unadversarial ExampleS (GUES) approach for the DR grading problem in this new setting. Specifically, OMG-DA presents an extreme safety scenario: The available target data is unlabeled and arrives in a flowing format, with no prior information about the pre-trained source model and data. Tab. 1 provides a detailed comparison with previous adaptation settings. + +In GUES, we address the absence of source data and pre-trained models by producing generalized unadversarial examples [24] for unlabeled target data. To this end, we introduce generative unadversarial learning, which theoretically reformulates conventional iterative perturbation optimization. This new method aims to learn a generative function for perturbations and involves addressing two key tasks: (1) Identifying the latent function input, which is the derivative of initial random noise w.r.t. image data, and (2) selecting a self-supervised property to serve as pseudo-perturbation labels. In practice, we leverage a Variational Autoencoder (VAE) [10] based approach to facilitate this learning process. In terms of function representation, we model the latent input using the encoder along with the reparameterization trick, + +whilst the decoder accomplishes the generation of individual-perturbation. Additionally, we choose the saliency map as the pseudo-perturbation label for two reasons: (1) It helps discover potential lesions, and (2) it aids in identifying the latent input by providing an upper bound. + +Our contributions are summarized as follows: + +- Pioneering a novel transfer setting OMG-DA, which is closer to real-world clinical scenarios and meets three typical requirements at the same time, including (1) model absence, (2) flowing data, and (3) source data privacy. +- Developing a new OMG-DA approach GUES in the context of DR grading, grounded on the generative unadversarial examples theory, where we learn an individual-perturbation generative function under saliency map supervision, removing relying on labels and models. +- Extensive evaluations on four DR benchmarks, indicating that GUES can largely promote the source model's performance in the target domain, as well as trainable test-time adaptation models, even at small batch size. + +# 2. Related work + +Adaptation methods for DR grading. Driven by real-world medical requirements, domain adaptation has been an attractive topic in this DR grading issue. For instance, Nguyen et al. [20] introduce a UDA approach that enables the model to focus on vessel structures that remain invariant to domain shifts via image reconstruction using labeled source domain data. In SFDA, Zhang et al. [40] propose generating labeled, target-style retinal images to improve the source model's generalization, relying solely on a pre-trained source model and unlabeled target images. Additionally, DG in DR grading has been explored through domain-invariant feature learning approaches [4] and divergence-based methods [5], leveraging labeled source domain data. + +The methods above rely on labeled data, require full access to the model, and necessitate offline training on a precollected dataset. In real clinical settings, these requirements can be impractical due to constraints around data and model privacy, as well as the need for real-time adaptability without retraining. In contrast, GUES provides an online adaptation solution that operates without requiring labeled data or access to the model, addressing domain adaptive DR grading. + +Unadversarial learning. Unadversarial learning was initially developed by Salman et al. [24], aiming to modify + +input image distribution to make them more easily recognizable by the model. Current mainstreams achieve this learning process by adding class-specific perturbations to the input images. Here, the perturbations are generated based on the gradient of an objective function w.r.t. image. This approach allows for the design of unadversarial examples without model training. For example, based on [24], NSA [25] introduces a method to generate more natural perturbations using a trainable generator. Similarly, CAT [18] demonstrates a new distance metric for generating unadversarial examples. + +All existing unadversarial learning methods require access to model parameters, outputs, and labeled data. This dependency invalidates them in our OMG-DA setting which only has access to unlabeled target data. In addition to this, our GUES produces individualized unadversarial examples in a generative manner, which stands out from previous methods that focus on class-specific unadversarial examples. + +Saliency map for medical image. A fine-grained saliency map is a pixel feature generated by calculating the central surround differences within images, identifying salient regions without any need for training [19]. This feature is widely applied in various medical image analysis tasks to extract pathologically important regions [9, 22, 31, 35]. For instance, in DR grading, studies such as [9, 22] utilize saliency maps to guide models in focusing on critical features like the optic disc, cup, and vessel structures. Similarly, in brain tumor detection, Tomar et al. [31] leverage saliency maps to enhance the model's attention on tumor and bone structures. In skin cancer detection, saliency maps help isolate lesion regions with distinctive features, such as lumpiness, which are essential for accurate diagnosis [35]. + +As stated above, existing works primarily use saliency maps to highlight lesion regions. Unlike it, GUES selects saliency maps as pseudo-perturbation labels. + +# 3. Methodology + +Problem statement. Given two different but related domains, i.e., source domain $\mathcal{S}$ and target domain $\mathcal{T}$ , $\mathcal{S}$ contains $n_s$ labeled samples, while $\mathcal{T}$ has $n$ unlabeled data. Both labeled and unlabeled samples share the same $C$ categories. Let $\mathcal{X}_s$ and $\mathcal{Y}_s$ be the source samples and the corresponding labels. Similarly, we denote the target samples and their labels by $\mathcal{X}_t = \{\pmb{x}_t^i\}_{i=1}^n$ and $\mathcal{Y}_t = \{y_t^i\}_{i=1}^n$ , respectively, where $n$ signifies the number of samples. The source model $\theta_s$ is pre-trained on $\{\mathcal{X}_s, \mathcal{Y}_s\}$ . + +OMG-DA is featured in (1) the absence of the source model $\theta_{s}$ and domain $S$ and (2) the flowing target data $\mathcal{X}_t$ the same as the TTA setting [34]. Unlike previous transfer settings that are model adaptation-centered [15, 29, 30, 38], OMG-DA considers adaptation from the perspective of data. Specifically, OMG-DA aims to modify the distribution of target data to facilitate downstream tasks. + +# 3.1. Generative Unadversarial Examples + +The unadversarial learning idea [24] is well aligned with OMG-DA in terms of promoting adaptation by enhancing the generalization of input data. We briefly recap traditional unadversarial learning. Unlike adversarial learning [28] that generates confusing samples to mislead models, unadversarial learning aims to construct generalized samples, tackling out-of-distribution issues. Formally, this learning can be summarized in the following optimization problem. + +$$ +\hat {\delta} = \arg \min _ {\delta} L \left(f _ {\theta} (x + \delta), y\right), s. t. | | \delta | | \leq \epsilon \tag {1} +$$ + +where $L(\cdot)$ denotes objective function, e.g., cross-entropy loss for classification tasks, $x$ and $y$ are input image and its label, $f_{\theta}$ is a pre-trained model with parameters $\theta$ , $\delta$ is a perturbation, $\epsilon$ is a small threshold. The current scheme solves this problem in an iterative way formulated as + +$$ +\delta_ {k + 1} = \delta_ {k} + \alpha \cdot \operatorname {s i g n} \left(\nabla_ {x} L \left(f _ {\theta} \left(x + \delta_ {k}\right), y\right)\right), k \in [ 0, K - 1 ], \tag {2} +$$ + +where $\alpha$ is a trade-off parameter, $K$ is iteration number, $\delta_0$ is an initial random noise. In the inference phase, the optimal perturbation $\hat{\delta}$ is integrated into the input $x$ , forming an unadversarial example $\hat{x} = x + \hat{\delta}$ , which is easily recognizable by the model $f_{\theta}$ . Obviously, the conventional unadversarial paradigm cannot meet our OMG-DA setting due to the absence of $f_{\theta}$ , $L$ , and the label $y$ . + +In this paper, we re-consider the iterative optimization process above and obtain the theorem below (The proof is provided in Supplementary). + +Theorem 1 Given the unadversarial learning problem defined in Eq. (1), the iterative process featured by Eq. (2) can be expressed as the following generative form. + +$$ +\delta_ {k} = \delta_ {0} + V \cdot F _ {\Phi} \left(\frac {\partial \delta_ {0}}{\partial x}\right), \tag {3} +$$ + +where $\delta_0$ is an initial random noise, $V > 0$ is a bound constant, $F_{\Phi}$ is a generative function, $\frac{\partial\delta_0}{\partial x}$ is a latent variable. + +Grounded on Theorem 1, we have: When $\delta_{k}$ converges to optimal $\hat{\delta}$ , i.e., $\delta_{k} \rightarrow \hat{\delta}$ , function $F_{\Phi}$ also evolves to the optimal one, denoted by $\hat{F}_{\Phi}$ , i.e., $F_{\Phi} \rightarrow \hat{F}_{\Phi}$ . This provides an insight: The unadversarial learning problem above can also be solved in a generative fashion. Correspondingly, the generated data are termed generative unadversarial examples. + +# 3.2. Model Instantiation + +Within this context of generative unadversarial learning, conventional unadversarial learning presented in Eq. (1) is boiled down to learning $F_{\Phi}\left(\frac{\partial\delta_0}{\partial x}\right)$ . We can achieve this by training a generative neural network. To make this solution sense, we have to solve two difficulties as follows. (A) One is + +![](images/a7c3ea92141db9da1aa0dd614c6fc7178718c4c132cabf7b92e49b8adc76978e.jpg) +Figure 2. The instantiation framework of GUES in the OMG-DA setting. (a) For target input $x_{t}$ , the VAE model generates individual perturbation $\delta_{t} = F_{\Phi}\left(\frac{\partial\delta_{0}}{\partial x}\right)$ . After that, the by-pass path incorporates $\delta_{t}$ and $x_{t}$ to create the generative unadversarial example $\hat{x}_{t}$ . Treating $x_{t}$ 's saliency map $g_{t}$ as reconstruction supervision for model training. (b) At the inference phase, the generated unadversarial example $\hat{x}_{t}$ is directly provided to the frozen source model or other trainable models. + +![](images/df185912946c69dbba5c9887b58cd479347ea57e68b6d91b7216902aed39df4b.jpg) +Figure 3. Explanation in choosing fine-grained saliency maps as supervision. Left: The testing fundus image selected from "Moderate DR" class in APTOS demonstrates that H (hemorrhages), SE (soft exudates), and EX (hard exudates) are essential characteristics to judge the DR grade. Middle: The saliency map highlights those lesions. Right: The gradient-CAM visualization of the source model on task DDR→APTOS. + +the identification of $\frac{\partial\delta_0}{\partial x}$ when the relationship between them is unknown. (B) The other is selecting property supervision (pseudo-perturbation labels) to drive $\delta_{k}\rightarrow \hat{\delta}$ . + +Solution to Problem A. In practice, considering derivative $\frac{\partial\delta_0}{\partial x}$ is relevant with both random noise $\delta_0$ and input image $x$ , we sample it from a certain Gaussian distribution associated with $x$ . Furthermore, the output size of $F_{\Phi}$ is the same as $x$ . Therefore, we employ the VAE model to jointly model $\frac{\partial\delta_0}{\partial x}$ and $F_{\Phi}$ , since VAE is an autoencoder characterized by random sampling. Specifically, as shown in Fig.2 (a), we approximate $\frac{\partial\delta_0}{\partial x}$ by latent variable $z$ , which is jointly determined by input $x$ and sampled random signal $\tau$ . As for $F_{\Phi}$ , it is demonstrated by the decoder module. Suppose $D(\cdot)$ is the decoder, $E_{\tau}(\cdot)$ is the encoder with reparameterization trick, our scheme can be formulated as + +$$ +D \left(E _ {\tau} (x)\right)\rightarrow F _ {\Phi} \left(\frac {\partial \delta_ {0}}{\partial x}\right), E _ {\tau} (x) \rightarrow \frac {\partial \delta_ {0}}{\partial x}, D (\cdot) \rightarrow F _ {\Phi} (\cdot). \tag {4} +$$ + +Solution to Problem B. We adopt the fine-grained saliency map as the supervision. Two reasons contribute to our selection. First of all, the empirical results show that, for the specific task of DR grading, the saliency map is an acceptable pseudo-perturbation. Specifically, perturbation in the unadversarial context enhances the regions associated with the category and reduces the prominence of other areas, thereby identifying the lesion zones relevant to DR grading. As illustrated in Fig. 3, the saliency map effectively identifies potential lesions, such as hemorrhages, soft exudates, and hard exudates. Furthermore, it includes gradient information, as it highlights regions similar to Grad-CAM (Right side). More importantly, we have the theorem below. (The proof is provided in Supplementary) + +Theorem 2 Given the partial derivatives of the initial random noise $\delta_0$ w.r.t image $x$ is $\frac{\partial\delta_0}{\partial x}$ and $x$ 's saliency map is $s = G(x)$ where $G(\cdot)$ is the computation function of the saliency map. We have the following relationship: + +$$ +\frac {\partial \delta_ {0}}{\partial x} \leq U \cdot s, \tag {5} +$$ + +where $U > 0$ is a bound constant. + +Theorem 2 suggests that the saliency map provides upper bounds for $\frac{\partial\delta_0}{\partial x}$ . Namely, $s$ provides relaxed descriptions for the variation $\frac{\partial\delta_0}{\partial x}$ . This can help guide the learning of $\frac{\partial\delta_0}{\partial x}$ . + +GUES framework. Based on the analysis above, we instantiate GUES as the framework depicted in Fig. 2. As shown in sub-figure (a), our method integrates a VAE model and by-pass connection, achieving the learning of $F_{\Phi}\left(\frac{\partial \delta_0}{\partial x}\right)$ . Specifically, $\frac{\partial \delta_0}{\partial x}$ is sampled from an input $x_t$ -featured Gaussian distribution $N(\mu, \sigma)$ , which is jointly learned using the + +# Algorithm 1 The pipeline of proposed GUES + +Input: Online batch samples $\mathcal{B}$ , a trainable $\mathrm{VAE}\theta_v$ consists of an encoder $E_{\tau}(\cdot)$ and a decoder $D(\cdot)$ . + +# Procedure: + +1: for $x_{i}$ in $\mathcal{B}$ do +2: Approximate the latent function input by $E_{\tau}(x_i)$ ; +3: Learn a generative function for perturbation $\delta_{i}$ by $D$ ; +4: Calculate individual perturbations $\delta_{i}$ by $D(E_{\tau}(x))$ +5: Create the unadversarial example $\hat{x}_i$ by incorporating $\delta_i$ and $x_i$ through the bypass path. +6: Generate a fine-grained saliency map $g_{i}$ of the $x_{i}$ ; +7: Update $\theta_v$ with Eq.(8), taking $g_t$ as a supervision. + +8: end for +9: return The generative unadversarial examples $\hat{x}$ + +encoder and reparameterization. The decoder, representing $F_{\Phi}$ , then transforms $\frac{\partial \delta_0}{\partial x}$ into a generative perturbation $\delta_t$ . Finally, the by-pass structure incorporates $x_t$ and $\delta_t$ to produce $\hat{x}_t$ . As shown in Fig. 2 (b), during the inference phase, the trained GUES model generates the corresponding generative unadversarial examples for the unseen frozen or fine-tuned model from specific testing samples. + +Loss function. The loss function for GUES training consists of two components. First, we enforce the latent space with mean $\mu$ and variance $\sigma$ satisfy the standard normal distribution $\mathcal{N}(0,I)$ . Suppose the encoder in VAE models the posterior distribution $q(z|x_{t}) = \mathcal{N}(\mu (x_{t}),\sigma^{2}(x_{t}))$ , this regularization can be formulated as: + +$$ +L _ {\mathrm {K L}} = D _ {\mathrm {K L}} \left(q (z | x _ {t}) \| \mathcal {N} (0, I)\right), \tag {6} +$$ + +where function $D_{\mathrm{KL}}$ computes the Kullback-Leibler divergence. The other reconstruction loss between the unadversarial example $\hat{x}_t$ and saliency map $g_t$ is presented by the following regression form: + +$$ +L _ {\mathrm {M S E}} = \left\| \hat {x} _ {t} - g _ {t} \right\| _ {2}. \tag {7} +$$ + +Formally, combining Eq. (6) and Eq. (7), the final objective of GUES can be summarized as: + +$$ +L _ {\text {G U E S}} = \alpha L _ {\mathrm {K L}} + \beta L _ {\mathrm {M S E}}, \tag {8} +$$ + +where $\alpha$ and $\beta$ are trade-off parameters. For clarity, we summarize the training procedure of GUES in Algorithm 1. + +In Eq. (8), the first item $L_{\mathrm{KL}}$ ensures the learning of $\frac{\partial \delta_0}{\partial x}$ . On the one hand, as aforementioned, we use $z$ to present $\frac{\partial \delta_0}{\partial x}$ . $L_{\mathrm{KL}}$ aligns the $z$ space with $\mathcal{N}(0, I)$ , thereby linking the random noise to $\frac{\partial \delta_0}{\partial x}$ . On the other hand, $q(z | x_t)$ in $L_{\mathrm{KL}}$ is a function of input $x_t$ , building relationship $x_t$ to $\frac{\partial \delta_0}{\partial x}$ . Additionally, the reconstruction regulated by the seconded item $L_{\mathrm{MSE}}$ encourage $\delta_k \rightarrow \hat{\delta}$ . + +# 4. Experiments + +# 4.1. Datasets + +We perform evaluation experiments on four existing fundus benchmarks, including APTOS [1], DDR [13], DeepDR [17], and Messidor-2 (termed MD2) [7]. Those datasets share five grading/classes: no DR, mild DR, moderate DR, severe DR, and proliferative DR. Taking each dataset as a separate domain, we form 12 transfer tasks crossing domains. For example, as APTOS is the source domain while the others are target domains, we have three transfer tasks APTOS $\rightarrow$ DDR, APTOS $\rightarrow$ DeepDR, and APTOS $\rightarrow$ MD2. The illustration of the label distribution and the domain shift of them is demonstrated in Supplementary. Of Note, all datasets have a severe class imbalance (e.g., "no DR" class itself takes up to $45.8\%$ of the DDR dataset). + +# 4.2. Implementation Detail + +Source model pre-training. We adopt the DeiT-base network [32] as the backbone of the source pre-trained model, training it in a supervised manner using the source data and corresponding ground truths. During this source training phase, the adopted objective is the classic cross-entropy loss with label smoothing, the same as other methods [15, 37, 38]. + +Variational autoencoder setting. The VAE model is an eight-layer convolutional architecture with a latent space dimension of 10. We do not employ a pre-trained VAE and utilize a VAE without fine-tuning on any other dataset, ensuring that the learning component $F_{\Phi}\left(\frac{\partial\delta_0}{\partial x}\right)$ are unbiased and independent of prior pre-training data. + +Parameter setting. For the trade-off parameters in Eq. (8), we set $\alpha$ to 1.0, while $\beta$ is tuned with $\{0.0001, 0.01, 1\}$ to ensure that $L_{\mathrm{KL}}$ and $L_{\mathrm{MSE}}$ remain on the same values scale. + +Training setting. For fairness, all compared methods were implemented with the same DeiT-base backbone. We adopt the batch size of 64, SGD optimizer with a momentum of 0.9 and a learning rate of 1e-5 on all datasets. All experiments are conducted with PyTorch on a single GPU of RTX A6000. + +# 4.3. Comparison Settings + +Evaluation metrics. To account for unbalanced datasets, besides conventional classification accuracy (termed ACC), we adopt the measure of Quadratic Weighted Kappa (termed QWK) [2] and the average of QWK and ACC (termed AVG). Their rules are provided in Supplementary. + +Competitors. We compare GUES with nine existing state-of-the-art adaptation methods divided into three groups. (1) The first group involves applying the source model directly to the target domain. (2) The second group includes five SFDA methods SHOT [15], NRC [38], CoWA [12], PLUE [16], and TPDS [29]. (3) The third group comprises three typical TTA methods: SHOT-IM [15], TENT [34], and SAR [21]. + +Table 2. The results of Source, SFDA, TTA, OMG-DA, and OMG-DA combination methods on datasets APTOS, DeepDR, DDR, and MD2 are presented. The improvements over baseline methods Source, SHOT-IM, and TENT are highlighted as (+x.x). + +
MethodVenueAPTOS→DDRAPTOS→DeepDRAPTOS→MD2DDR→APTOSDDR→DeepDRDDR→MD2DeepDR→APTOS
ACCQWKAVGACCQWKAVGACCQWKAVGACCQWKAVGACCQWKAVGACCQWKAVGACCQWKAVG
Source-60.659.259.952.671.762.160.948.754.865.672.969.345.060.652.849.434.642.043.471.657.5
GUES + Source-62.059.560.853.069.761.359.846.753.376.081.878.956.468.762.559.147.653.346.574.160.3
SHOT [15]ICML2066.969.067.953.673.563.651.738.044.877.084.280.659.274.666.957.143.150.162.382.572.4
NRC [38]NeurIPS2161.965.263.551.670.961.354.140.347.260.376.368.352.069.160.550.635.242.952.074.363.2
CoWA [12]ICML2259.064.962.051.070.760.853.037.345.157.274.565.850.166.858.453.139.646.350.473.061.7
PLUE [16]CVPR2362.065.163.651.369.760.554.541.147.863.464.263.854.364.859.551.628.239.954.670.562.6
TPDS [29]IICV2466.667.867.251.671.561.552.740.746.776.683.980.258.073.165.654.541.047.860.880.070.4
SHOT-IM [15]ICML2066.569.267.952.673.663.153.236.644.975.982.179.058.573.966.257.343.550.461.984.072.9
TENT [34]ICLR2059.950.255.153.170.161.661.448.454.975.282.478.855.168.461.760.850.555.660.279.769.9
SAR [21]ICLR2367.963.865.853.673.063.357.244.951.175.683.579.555.671.663.649.237.043.159.579.369.4
GUES + SHOT-IM - GUES + TENT -68.668.568.553.572.863.255.743.849.777.283.180.260.575.167.861.551.256.362.683.272.9
-61.856.359.053.270.061.661.147.254.175.983.079.458.770.864.763.353.858.654.977.666.3
MethodVenueDeepDR→DDRDeepDR→MD2MD2→APTOSMD2→DDRMD2→DeepDRAccAvg. QWKAVG
ACCQWKAVGACCQWKAVGACCQWKAVGACCQWKAVGACCQWKAVG
Source-56.466.961.748.750.249.443.970.357.160.256.558.359.858.459.153.960.157.0
GUES + Source-57.365.661.448.352.050.159.376.267.764.755.860.258.657.157.858.4 (+4.5)62.9 (+2.8)60.7 (+3.7)
SHOT [15]ICML2057.471.264.348.841.545.252.773.062.854.659.256.959.670.264.958.465.061.7
NRC [38]NeurIPS2144.960.952.949.841.845.848.869.158.952.852.852.858.062.860.453.159.956.5
CoWA [12]ICML2248.758.453.649.942.746.351.070.160.549.650.950.357.660.659.156.961.959.4
PLUE [16]CVPR2347.253.550.456.447.451.956.069.162.656.554.355.458.864.861.855.557.756.6
TPDS [29]IICV2459.369.464.350.542.446.460.374.967.660.060.460.258.963.060.959.264.061.6
SHOT-IM [15]ICML2054.669.462.051.238.244.761.677.969.757.058.757.957.569.863.759.064.761.9
TENT [34]ICLR2058.545.451.958.356.557.455.174.164.655.831.743.758.053.655.859.359.259.3
SAR [21]ICLR2353.066.359.642.633.137.955.273.064.149.748.349.056.765.861.356.361.659.0
GUES + SHOT-IM - GUES + TENT --62.371.566.952.847.850.362.878.170.466.059.462.760.668.664.662.0 (+3.0)66.9 (+2.2)64.5 (+2.6)
-62.661.462.059.157.458.259.375.667.564.050.557.258.356.057.161.0 (+1.7)63.3 (+4.1)62.2 (+2.9)
+ +![](images/63765010acc15e3c61df0b9bf3ce8559686a07f425f19d2b0dc2843245c0ab1e.jpg) +Figure 4. Comparison results with batch size varying from 2 to 64 over the 12 tasks (The details are provided in Supplementary.). Left, middle, and right report ACC, QWK, and AVG, respectively. + +Comparison protocol in OMG-DA setting. For a comprehensive comparison, our comparison follows two different fashions as follows. + +- Case without training: We first generate the unadversarial examples for the target domain by the trained GUES model and then provide them to the frozen source model. +- Case with training: We plug GUES into other TTA methods (they are also online methods with flowing data) as online image pre-processing. + +The two cases evaluate GUES from different aspects. The first isolates the generalization ability of the unadversarial examples generated by GUES, whilst the second highlights GUES's compatibility with other trainable online schemes. + +Corresponding to the comparison protocols above, besides the version GUES corresponding to the case with + +out training, we also introduce GUES+SHOT-IM and GUES+TENT which correspond to the case with training. + +# 4.4. Comparison Results + +This part presents the comparison results following the cases above. Also, considering batch size a crucial factor for TTA methods, the results as the batch size varies are provided. + +Results without training. The comparisons are shown in Tab. 2. On average, across the 12 tasks and without training of the source model, GUES achieves improvements of $4.5\%$ in ACC, $2.8\%$ in QWK, and $3.7\%$ in AVG compared to the source model. These results demonstrate that GUES modifies the target data distribution effectively, adapting the target domain to align with the source domain. + +Results with training. As shown in Tab. 2, GUES+SHOT-IM outperforms the previous best SFDA and TTA methods, respectively surpassing TENT in ACC by $2.7\%$ , SHOT in QWK by $1.9\%$ , and SHOT-IM in AVG by $2.6\%$ . Compared to SHOT-IM, GUES+SHOT-IM gains over $3.0\%$ in ACC, $2.2\%$ in QWK, and $2.6\%$ in AVG. Similarly, GUES+TENT improves over TENT by $1.7\%$ in ACC, $4.1\%$ in QWK, and $2.9\%$ in AVG. These results highlight the effectiveness of combining GUES with other methods that require training. + +Results with varying batch size. This part isolates the effect of batch size, which is a crucial factor for TTA methods. Fig. 4 depicts the performance variation as batch size vary + +![](images/373dc7e04a1dccbc02cde73710d605ee9811e1e2bb7be310782249f961c2d3d8.jpg) +Figure 5. Interpretability analysis based on a typical fundus image from "Moderate DR" class in APTOS. Here, H (hemorrhages), SE (soft exudates), and EX (hard exudates) are essential characteristics to judge the DR grade. The gradient CAM-based heatmap of five models visualizes the capture of those lesions. All models are trained on task DDR $\rightarrow$ APTOS, where Oracle is trained using ground truth in APTOS. + +![](images/7dc39ff85d07b737b9df1a1ce46a1b53f0d0d32651c72c934caa98923545f3cf.jpg) +Figure 6. Feature distribution comparison of 3D density charts on task DDR $\rightarrow$ APTOS. Oracle is trained on APTOS by ground truth. + +ing from 2 to 64 over the 12 tasks. It is observed that TTA methods SHOT-IM and TENT suffer from severe performance drops when the batch size becomes small. SHOT-IM exhibits a decrease of approximately $16\%$ in ACC when the batch size is reduced from 64 to 2, whilst TENT shows a substantial decline of around $34\%$ in QWK. Oppose to it, the methods with GUES, SHOT-IM+GUES, and TENT+GUES, do not have evident performance decline at the smaller batch size. Moreover, this combination not only mitigates the drop but also shows improvements when the batch size is 64. This indicates that GUES effectively stabilizes the performance of SHOT-IM and TENT, enhancing their robustness to variations in batch size while boosting their overall effectiveness. + +We attribute GUES's excellent robustness to its ability to predict individual perturbations (see Supplementary for more details on the visualization of perturbations) that focus on single image-specific features rather than global data commonalities. For instance, the conventional unadversarial examples approach refines the class-specific perturbation sensitive to batch size. Furthermore, both SHOT-IM and TENT are entropy-based methods that require large-scale batch size for accurate entropy estimation. + +# 4.5. Visualization Analysis + +Interpretability. For a better understanding, Fig. 5 demonstrates whether GUES can help capture pathologically relevant features, such as H (hemorrhages), SE (soft exudates), and EX (hard exudates), determining DR grade. First of all, when comparing the source model with GUES, the source model only captures a limited area of the lesion, while GUES effectively captures most of the DR-related features. Further- + +![](images/860379c907f692a288f195ca59110b8bb3e73479d954ca982ac69eb7b89547a2.jpg) +Figure 7. Unadversarial examples visualization of two typical target samples from APTOS. The generative perturbations are generated by the GUES model trained on task DDR $\rightarrow$ APTOS. + +more, combining GUES with SHOT-IM (i.e., GUES+SHOT-IM) expands the focus on DR-related features beyond those captured by SHOT-IM alone. Additionally, when comparing the four models to Oracle, only GUES and GUES+SHOT-IM resemble Oracle, suggesting that GUES effectively directs the model's attention to DR-critical features. + +Feature distribution. Taking the task DDR→APTOS as a toy experiment, we visualize the feature distribution extracted from the final convolutional layer of the prediction model using a 3D density chart. Considering that APTOS is a class-imbalanced dataset, with the “No DR” class alone comprising up to $49.2\%$ of the dataset, our analysis focuses on this crucial property. As shown in Fig. 6, the feature distribution of the source model does not reflect this imbalanced characteristic; instead, it displays a more uniform classification. Conversely, the feature distribution of GUES exhibits a distinct imbalance, with one expanded high-density region alongside several smaller high-density regions, resembling the distribution pattern seen in Oracle. + +Visualization of unadversarial examples. This part visualizes unadversarial examples of two typical target samples from APTOS and corresponding generative perturbations, based on the task DDR $\rightarrow$ APTOS. Considering the generative perturbations alter the original images that may not be easily visible to the naked eye, we collect RGB statistics to illustrate these changes quantitatively. It is observed that in Fig. 7, each channel (R, G, and B) exhibits notable + +Table 3. ACC results of ablation study $(\%)$ + +
#\( L_{\text{KL}} \)\( L_{\text{MSE}} \)APTOSDDRDeepDRMD2Avg.
1XX51.059.152.553.053.9
2✓X54.758.354.953.355.3
3X✓56.960.152.952.955.7
4✓✓60.661.356.055.758.4
5GUES w/ AE53.860.253.853.255.3
6GUES w/ Mixup56.860.253.052.855.7
7GUES w/ Self56.560.453.754.456.3
8GUES w/ Sal46.134.341.845.942.0
+ +fluctuations, with the RGB statistics of the original images (a) differing significantly from those of the unadversarial examples (c). Additionally, each generative perturbation is unique, meaning that the alterations introduced by these perturbations are individual. These results suggest that the perturbations may help highlight critical DR-related features, refining the model's focus on diagnostically relevant areas. + +# 4.6. Further Analysis + +Ablation study. In this part, we evaluate the effect of objective loss, as well as the components involved in GUES including the sampling strategy and saliency map-based supervision. To address the first issue, we conduct a progressive experiment. The top four rows of Tab. 3 list the ablation results where the source model's performance is the baseline. Using $L_{\mathrm{KL}}$ or $L_{\mathrm{MSE}}$ alone yields an average ACC improvement of approximately $1.4\%$ and $1.8\%$ , respectively, over the baseline. As both of them work, the ACC increases $3.5\%$ on average further. The results indicate that all objective components positively affect the final performance. + +To evaluate the impact of the sampling component, we propose a variation method of GUES, GUES w/ AE, where we remove this sampling process by replacing VAE with a conventional Auto-encoder model. In addition, two GUES variations are used to assess the advantage of saliency map-based supervision. Specifically, GUES w/ Mixup replaces the saliency maps with a Mixup of saliency maps and original images, whilst GUES w/ Self replaces the saliency maps with the original images. As presented in the Tab. 3, compared with the full version of GUES (the fourth row), GUES w/ AE, GUES w/ Mixup and GUES w/ Self decrease by $3.1\%$ at least on average. Besides, replacing the original images with saliency maps as inputs (GUES w/ Sal) leads to a significant drop of about $15.4\%$ . Above these experiments confirm the effectiveness of our design choices. + +Parameter sensitiveness. Taking the task DDR $\rightarrow$ APTOS as a toy experiment, we present the GUES performance varying as hyper-parameters $0.5\leq \alpha \leq 1.5$ with 0.1 steps, $0.00005\leq \beta \leq 0.00014$ with 0.00001 steps. As depicted in Fig. 8, the ACC, QWK, and AVG variation surfaces show fluctuations in the tiny performance zone, with approximately $0.2\%$ in ACC, $0.25\%$ in QWK, and $0.4\%$ in + +![](images/9cdc9e8c29501aa175b39b55b2bda1428b83dc0e214cde588a14fcebc3a5d085.jpg) +Figure 8. Parameter sensitiveness study results over $\alpha \times \beta$ based on task DDR $\rightarrow$ APTOS. From left to right, there are results on ACC, QWK, and AVG, respectively. + +![](images/3c06c4119aeb4010430e016ae1e367f133b843ee2a703215d4cbaef1c5ca462e.jpg) + +![](images/5b58cf7f34b38f89b60754474f73c44da4741fc11bfe7135a5098f474b99d0f1.jpg) + +AVG. This observation suggests that GUES is insensitive to alterations in $\alpha$ and $\beta$ . + +Limitation. GUES uses the saliency map to guide the learning of the generative function. This method is effective for DR grading but encounters challenges in natural image scenarios. The natural images contain rich semantics, such as shape, relative structure, and complex background, which are not all relevant to tasks. However, the saliency map blindly highlights all those factors, struggling to capture the task-specific ones. In a theoretical point-view, the richness of these semantics means significant variations, resulting in a super-relaxed bound constant $U$ (see Theorem 2) that undermines the descriptive power of the saliency map for $\frac{\partial \delta_0}{\partial x}$ . In contrast, fundus images are more monolithic, implying a smaller $U$ that justifies the usage of the saliency map. (The further discussion is provided in Supplementary) + +# 5. Conclusion + +In this paper, we propose a clinically motivated setting, OMG-DA, where the models are unseen prior to their use, and only target data flows are accessible. This setting ensures both model protection and source data privacy in a data flow scenario. To adapt to the target domain without access to the models, we introduce a GUES approach. Instead of conventional iterative optimization, we generate unadversarial examples for flowing target data by directly predicting individual perturbations. This approach is grounded in the theoretical results of generative unadversarial learning. In practice, we utilize the VAE model to learn the perturbation generation function with a latent input variable. Furthermore, we demonstrate that saliency maps can serve as an upper bound for this latent variable. This relationship inspires us to use the saliency maps as pseudo-perturbation labels for model training. Extensive experiments conducted on four DR benchmarks confirm that the proposed method can achieve state-of-the-art results when it pairs with both frozen pre-trained and fine-tuning models. + +# Acknowledgments + +This work is supported by the National Natural Science Foundation of China (62476169, 62206168, 62276048). + +# References + +[1] Aptos: Aptos 2019 blindness detection website. https://www.kaggle.com/c/aptos2019-blindness-detection, accessed February 20, 2022. +[2] Quadratic weighted kappa. https://www.Eyepacs.com/aroraaman/quadratic-kappa-metric-explained-in-5-simple-steps, accessed July 2022. +[3] Eman AbdelMaksoud, Sherif Barakat, and Mohammed Elmogy. A comprehensive diagnosis system for early signs and different diabetic retinopathy grades using fundus retinal images based on pathological changes detection. Computers in Biology and Medicine, 126:104039, 2020. +[4] Mohammad Atwany and Mohammad Yaqub. Drogen: domain generalization in diabetic retinopathy classification. In MICCAI, 2022. +[5] Haoxuan Che, Yuhan Cheng, Haibo Jin, and Hao Chen. Towards generalizable diabetic retinopathy grading in unseen domains. In MICCAI, 2023. +[6] Ling Dai, Liang Wu, Huating Li, Chun Cai, Qiang Wu, Hongyu Kong, Ruhan Liu, Xiangning Wang, Xuhong Hou, Yuexing Liu, et al. A deep learning system for detecting diabetic retinopathy across the disease spectrum. Nature communications, 12(1):3242, 2021. +[7] Etienne Decencière, Xiwei Zhang, Guy Cazuguel, Bruno Lay, Béatrice Cochener, Caroline Trone, Philippe Gain, John-Richard Ordóñez-Varela, Pascale Massin, Ali Erginay, et al. Feedback on a publicly distributed image database: the messidor database. Image Analysis & Stereology, pages 231-234, 2014. +[8] Along He, Tao Li, Ning Li, Kai Wang, and Huazhu Fu. Cabinet: Category attention block for imbalanced diabetic retinopathy grading. IEEE Transactions on Medical Imaging, 40(1):143-153, 2020. +[9] Yijin Huang, Junyan Lyu, Pujin Cheng, Roger Tam, and Xiaoying Tang. Ssit: Saliency-guided self-supervised image transformer for diabetic retinopathy grading. IEEE Journal of Biomedical and Health Informatics, 2024. +[10] Diederik P Kingma. Auto-encoding variational bayes. arXiv:1312.6114, 2013. +[11] Wouter M Kouw and Marco Loog. An introduction to domain adaptation and transfer learning. arXiv:1812.11806, 2018. +[12] Jonghyun Lee, Dahuin Jung, Junho Yim, and Sungroh Yoon. Confidence score for source-free unsupervised domain adaptation. In ICML, 2022. +[13] Tao Li, Yingqi Gao, Kai Wang, Song Guo, Hanruo Liu, and Hong Kang. Diagnostic assessment of deep learning algorithms for diabetic retinopathy screening. Information Sciences, 501:511-522, 2019. +[14] Tao Li, Wang Bo, Chunyu Hu, Hong Kang, Hanruo Liu, Kai Wang, and Huazhu Fu. Applications of deep learning in fundus images: A review. Medical Image Analysis, 69: 101971, 2021. +[15] Jian Liang, Dapeng Hu, and Jiashi Feng. Do we really need to access the source data? source hypothesis transfer for unsupervised domain adaptation. In ICML, 2020. + +[16] Mattia Litrico, Alessio Del Bue, and Pietro Morerio. Guiding pseudo-labels with uncertainty estimation for source-free unsupervised domain adaptation. In CVPR, 2023. +[17] Ruhan Liu, Xiangning Wang, Qiang Wu, Ling Dai, Xi Fang, Tao Yan, Jaemin Son, Shiqi Tang, Jiang Li, Zijian Gao, et al. Deepgrid: Diabetic retinopathy—grading and image quality estimation challenge. *Patterns*, 3(6), 2022. +[18] Xingbin Liu, Huafeng Kuang, Xianming Lin, Yongjian Wu, and Rongrong Ji. Cat: Collaborative adversarial training. arXiv:2303.14922, 2023. +[19] Sebastian Montabone and Alvaro Soto. Human detection using a mobile platform and novel features derived from a visual saliency mechanism. Image and Vision Computing, 28 (3):391-402, 2010. +[20] Duy MH Nguyen, Truong TN Mai, Ngoc TT Than, Alexander Prange, and Daniel Sonntag. Self-supervised domain adaptation for diabetic retinopathy grading using vessel image reconstruction. In KI, 2021. +[21] Shuaicheng Niu, Jiaxiang Wu, Yifan Zhang, Zhiquan Wen, Yaofo Chen, Peilin Zhao, and Mingkui Tan. Towards stable test-time adaptation in dynamic wild world. In ICLR, 2023. +[22] Jiaming Qiu, Weikai Huang, Yijin Huang, Nanxi Yu, and Xiaoying Tang. Augpaste: A one-shot approach for diabetic retinopathy detection. Biomedical Signal Processing and Control, 96:106489, 2024. +[23] Jinye Ran, Guanghua Zhang, Fan Xia, Ximei Zhang, Juan Xie, and Hao Zhang. Source-free active domain adaptation for diabetic retinopathy grading based on ultra-wide-field fundus images. Computers in Biology and Medicine, 174:108418, 2024. +[24] Hadi Salman, Andrew Ilyas, Logan Engstrom, Sai Vemprala, Aleksander Madry, and Ashish Kapoor. Unadversarial examples: Designing objects for robust vision. In NeurIPS, 2021. +[25] Abhijith Sharma, Phil Munz, and Apurva Narayan. Nsa: Naturalistic support artifact to boost network confidence. In IJCNN, 2023. +[26] Reza Shokri, Marco Stronati, Congzheng Song, and Vitaly Shmatikov. Membership inference attacks against machine learning models. In S&P, 2017. +[27] Daniel E Singer, David M Nathan, Howard A Fogel, and Andrew P Schachat. Screening for diabetic retinopathy. Annals of Internal Medicine, 116(8):660-671, 1992. +[28] C Szegedy. Intriguing properties of neural networks. arXiv:1312.6199, 2013. +[29] Song Tang, An Chang, Fabian Zhang, Xiatian Zhu, Mao Ye, and Changshui Zhang. Source-free domain adaptation via target prediction distribution searching. International Journal of Computer Vision, 132(3):654-672, 2024. +[30] Song Tang, Wenxin Su, Mao Ye, Jianwei Zhang, and Xiatian Zhu. Unified source-free domain adaptation. arXiv:2403.07601, 2024. +[31] Nishtha Tomar, Sushmita Chandel, and Gaurav Bhatnagar. A visual attention-based algorithm for brain tumor detection using an on-center saliency map and a superpixel-based framework. Healthcare Analytics, 5:100323, 2024. + +[32] Hugo Touvron, Matthieu Cord, Matthijs Douze, Francisco Massa, Alexandre Sablayrolles, and Hervé Jégou. Training data-efficient image transformers & distillation through attention. In ICML, 2021. +[33] Jeya Maria Jose Valanarasu, Pengfei Guo, VS Vibashan, and Vishal M Patel. On-the-fly test-time adaptation for medical image segmentation. In MIDL, 2024. +[34] Dequan Wang, Evan Shelhamer, Shaoteng Liu, Bruno Olshausen, and Trevor Darrell. Tent: Fully test-time adaptation by entropy minimization. In ICLR, 2020. +[35] Tianyunxi Wei, Yijin Huang, Li Lin, Pujin Cheng, Sirui Li, and Xiaoying Tang. Saliency-guided and patch-based mixup for long-tailed skin cancer image classification. arXiv:2406.10801, 2024. +[36] Zhan Wu, Gonglei Shi, Yang Chen, Fei Shi, Xinjian Chen, Gouenou Coatieux, Jian Yang, Limin Luo, and Shuo Li. Coarse-to-fine classification for diabetic retinopathy grading using convolutional neural network. Artificial Intelligence in Medicine, 108:101936, 2020. +[37] Tongkun Xu, Weihua Chen, WANG Pichao, Fan Wang, Hao Li, and Rong Jin. Cdtrans: Cross-domain transformer for unsupervised domain adaptation. In ICLR, 2021. +[38] Shiqi Yang, Joost van de Weijer, Luis Herranz, Shangling Jui, et al. Exploiting the intrinsic neighborhood structure for source-free domain adaptation. In NeurIPS, 2021. +[39] Hongxu Yin, Arun Mallya, Arash Vahdat, Jose M Alvarez, Jan Kautz, and Pavlo Molchanov. See through gradients: Image batch recovery via grad inversion. In CVPR, 2021. +[40] Chenrui Zhang, Tao Lei, and Ping Chen. Diabetic retinopathy grading by a source-free transfer learning approach. Biomedical Signal Processing and Control, 73:103423, 2022. \ No newline at end of file diff --git a/2025/Domain Adaptive Diabetic Retinopathy Grading with Model Absence and Flowing Data/images.zip b/2025/Domain Adaptive Diabetic Retinopathy Grading with Model Absence and Flowing Data/images.zip new file mode 100644 index 0000000000000000000000000000000000000000..57444101dc870f489b263a6bb180b73187638b0c --- /dev/null +++ b/2025/Domain Adaptive Diabetic Retinopathy Grading with Model Absence and Flowing Data/images.zip @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:4a320f8f77e735cf009d8bb691a04ec16edf3c2c93c9ef7760a628bf2be1db90 +size 609166 diff --git a/2025/Domain Adaptive Diabetic Retinopathy Grading with Model Absence and Flowing Data/layout.json b/2025/Domain Adaptive Diabetic Retinopathy Grading with Model Absence and Flowing Data/layout.json new file mode 100644 index 0000000000000000000000000000000000000000..afe2eea3a0104d74934441b60e33a7c5c02ccaf3 --- /dev/null +++ b/2025/Domain Adaptive Diabetic Retinopathy Grading with Model Absence and Flowing Data/layout.json @@ -0,0 +1,10511 @@ +{ + "pdf_info": [ + { + "para_blocks": [ + { + "bbox": [ + 119, + 103, + 492, + 138 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 119, + 103, + 492, + 138 + ], + "spans": [ + { + "bbox": [ + 119, + 103, + 492, + 138 + ], + "type": "text", + "content": "Domain Adaptive Diabetic Retinopathy Grading with Model Absence and Flowing Data" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 140, + 170, + 471, + 199 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 140, + 170, + 471, + 199 + ], + "spans": [ + { + "bbox": [ + 140, + 170, + 471, + 199 + ], + "type": "text", + "content": "Wenxin " + }, + { + "bbox": [ + 140, + 170, + 471, + 199 + ], + "type": "inline_equation", + "content": "\\mathrm{Su}^1" + }, + { + "bbox": [ + 140, + 170, + 471, + 199 + ], + "type": "text", + "content": ", Song Tang" + }, + { + "bbox": [ + 140, + 170, + 471, + 199 + ], + "type": "inline_equation", + "content": "^{*1,2,3}" + }, + { + "bbox": [ + 140, + 170, + 471, + 199 + ], + "type": "text", + "content": ", Xiaofeng Liu" + }, + { + "bbox": [ + 140, + 170, + 471, + 199 + ], + "type": "inline_equation", + "content": "^4" + }, + { + "bbox": [ + 140, + 170, + 471, + 199 + ], + "type": "text", + "content": ", Xiaojing Yi" + }, + { + "bbox": [ + 140, + 170, + 471, + 199 + ], + "type": "inline_equation", + "content": "^5" + }, + { + "bbox": [ + 140, + 170, + 471, + 199 + ], + "type": "text", + "content": ", Mao Ye" + }, + { + "bbox": [ + 140, + 170, + 471, + 199 + ], + "type": "inline_equation", + "content": "^6" + }, + { + "bbox": [ + 140, + 170, + 471, + 199 + ], + "type": "text", + "content": ", Chunxiao " + }, + { + "bbox": [ + 140, + 170, + 471, + 199 + ], + "type": "inline_equation", + "content": "\\mathrm{Zu}^1" + }, + { + "bbox": [ + 140, + 170, + 471, + 199 + ], + "type": "text", + "content": ", Jiahao Li" + }, + { + "bbox": [ + 140, + 170, + 471, + 199 + ], + "type": "inline_equation", + "content": "^7" + }, + { + "bbox": [ + 140, + 170, + 471, + 199 + ], + "type": "text", + "content": ", and Xiatian Zhu" + }, + { + "bbox": [ + 140, + 170, + 471, + 199 + ], + "type": "inline_equation", + "content": "^8" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 64, + 209, + 544, + 251 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 64, + 209, + 544, + 251 + ], + "spans": [ + { + "bbox": [ + 64, + 209, + 544, + 251 + ], + "type": "text", + "content": "1University of Shanghai for Science and Technology 2Universität Hamburg, 3ComOriginMat Inc. \n4Tale University 5Sichuan Eye Hospital 6University of Electronic Science and Technology of China \n7Chinese Academy of Medical Sciences & Peking Union Medical College 8University of Surrey" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 242, + 254, + 373, + 264 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 242, + 254, + 373, + 264 + ], + "spans": [ + { + "bbox": [ + 242, + 254, + 373, + 264 + ], + "type": "text", + "content": "steventangsong@gmail.com" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 152, + 300, + 200, + 312 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 152, + 300, + 200, + 312 + ], + "spans": [ + { + "bbox": [ + 152, + 300, + 200, + 312 + ], + "type": "text", + "content": "Abstract" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 55, + 327, + 297, + 674 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 327, + 297, + 674 + ], + "spans": [ + { + "bbox": [ + 55, + 327, + 297, + 674 + ], + "type": "text", + "content": "Domain shift (the difference between source and target domains) poses a significant challenge in clinical applications, e.g., Diabetic Retinopathy (DR) grading. Despite considering certain clinical requirements, like source data privacy, conventional transfer methods are predominantly model-centered and often struggle to prevent model-targeted attacks. In this paper, we address a challenging Online Model-aGnostic Domain Adaptation (OMG-DA) setting, driven by the demands of clinical environments. This setting is characterized by the absence of the model and the flow of target data. To tackle the new challenge, we propose a novel approach, Generative Unadversarial ExampleS (GUES), which enables adaptation from a data-centric perspective. Specifically, we first theoretically reformulate conventional perturbation optimization in a generative way—learning a perturbation generation function with a latent input variable. During model instantiation, we leverage a Variational AutoEncoder to express this function. The encoder with the reparameterization trick predicts the latent input, whilst the decoder is responsible for the generation. Furthermore, the saliency map is selected as pseudoperturbation labels. Because it not only captures potential lesions but also theoretically provides an upper bound on the function input, enabling the identification of the latent variable. Extensive experiments on DR benchmarks with both frozen pre-trained models and trainable models demonstrate the superiority of GUES, showing robustness even with small batch size. The source code and data are available at https://github.com/tntek/GUES." + } + ] + } + ], + "index": 7 + }, + { + "type": "image", + "bbox": [ + 328, + 298, + 542, + 392 + ], + "blocks": [ + { + "bbox": [ + 328, + 298, + 542, + 392 + ], + "lines": [ + { + "bbox": [ + 328, + 298, + 542, + 392 + ], + "spans": [ + { + "bbox": [ + 328, + 298, + 542, + 392 + ], + "type": "image", + "image_path": "6fbb60f19e595fa5edb64b48acb684f52e615b73cb2d7ef826c77943258bdfd7.jpg" + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 313, + 399, + 555, + 488 + ], + "lines": [ + { + "bbox": [ + 313, + 399, + 555, + 488 + ], + "spans": [ + { + "bbox": [ + 313, + 399, + 555, + 488 + ], + "type": "text", + "content": "Figure 1. Comparison between the OMG-DA and SFDA settings. (a) In SFDA, adaptation builds upon the accumulated data, which demands significant storage resources in the hospital. Additionally, the models' architecture and parameters are accessible, exposing them to potential attacks. (b) OMG-DA provides a practical scenario: Flowing data mimic the patients' arrival in a stream way, and the models are unseen (strictly controlled) before using them, avoiding attacks like membership inference attacks [26]." + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "image_caption" + } + ], + "index": 8 + }, + { + "bbox": [ + 315, + 510, + 394, + 523 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 315, + 510, + 394, + 523 + ], + "spans": [ + { + "bbox": [ + 315, + 510, + 394, + 523 + ], + "type": "text", + "content": "1. Introduction" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 313, + 533, + 556, + 700 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 533, + 556, + 700 + ], + "spans": [ + { + "bbox": [ + 313, + 533, + 556, + 700 + ], + "type": "text", + "content": "Diabetic Retinopathy (DR) is a significant health concern, ranking among the leading causes of blindness and affecting millions of people worldwide [3]. Early-stage intervention for DR is crucial to preserve vision, highlighting the importance of timely diagnosis [27]. Although deep learning (DL) has demonstrated promising results in automating the grading of DR [6, 8, 36], deploying DL models in real-world clinical settings remains challenging. For example, DL models often struggle to generalize effectively to complex scenarios, such as variations in imaging equipment, ethnic groups, or temporal factors, leading to different data distributions, a challenge known as domain shift [11]. This issue significantly hampers the widespread adoption and success of DL-based diagnostic tools in clinical practice [14]." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 326, + 701, + 553, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 326, + 701, + 553, + 713 + ], + "spans": [ + { + "bbox": [ + 326, + 701, + 553, + 713 + ], + "type": "text", + "content": "Recently, many adaptation methods for grading diabetic" + } + ] + } + ], + "index": 12 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "spans": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "type": "text", + "content": "CVF" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 145, + 0, + 494, + 35 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 145, + 0, + 494, + 35 + ], + "spans": [ + { + "bbox": [ + 145, + 0, + 494, + 35 + ], + "type": "text", + "content": "This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 67, + 702, + 143, + 712 + ], + "type": "page_footnote", + "angle": 0, + "lines": [ + { + "bbox": [ + 67, + 702, + 143, + 712 + ], + "spans": [ + { + "bbox": [ + 67, + 702, + 143, + 712 + ], + "type": "text", + "content": "*Corresponding author" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "text", + "content": "28337" + } + ] + } + ], + "index": 14 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 0 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 66, + 91, + 544, + 152 + ], + "blocks": [ + { + "bbox": [ + 82, + 70, + 526, + 82 + ], + "lines": [ + { + "bbox": [ + 82, + 70, + 526, + 82 + ], + "spans": [ + { + "bbox": [ + 82, + 70, + 526, + 82 + ], + "type": "text", + "content": "Table 1. Comparison between different transfer settings. Notation: source (s), target (t), data " + }, + { + "bbox": [ + 82, + 70, + 526, + 82 + ], + "type": "inline_equation", + "content": "\\mathbf{x}" + }, + { + "bbox": [ + 82, + 70, + 526, + 82 + ], + "type": "text", + "content": " ,label " + }, + { + "bbox": [ + 82, + 70, + 526, + 82 + ], + "type": "inline_equation", + "content": "\\mathbf{y}" + }, + { + "bbox": [ + 82, + 70, + 526, + 82 + ], + "type": "text", + "content": " ,loss function " + }, + { + "bbox": [ + 82, + 70, + 526, + 82 + ], + "type": "inline_equation", + "content": "L\\left( \\cdot \\right)" + }, + { + "bbox": [ + 82, + 70, + 526, + 82 + ], + "type": "text", + "content": " ." + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 66, + 91, + 544, + 152 + ], + "lines": [ + { + "bbox": [ + 66, + 91, + 544, + 152 + ], + "spans": [ + { + "bbox": [ + 66, + 91, + 544, + 152 + ], + "type": "table", + "html": "
Setting nameModel absenceData flowSource data privacySource dataTarget dataTrain lossTest loss
Fine-tuning (FT)XX✓-xt,ytL(xt,yt)-
Unsupervised domain adaptation (UDA)XXXxs,ysxtL(xs,ys)+L(xt)-
Source-free domain adaptation (SFDA)XX✓-xtL(xt)-
Test-time adaptation (TTA)X✓✓-xt-L(xt)
Online model-agnostic domain adaptation (OMG-DA)✓✓✓-xt--
", + "image_path": "68d886601922b4cab513baef77a9ffeadb139001fd3f2ab7e14a21ea98fc7f4e.jpg" + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "table_body" + } + ], + "index": 1 + }, + { + "bbox": [ + 54, + 172, + 296, + 291 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 54, + 172, + 296, + 291 + ], + "spans": [ + { + "bbox": [ + 54, + 172, + 296, + 291 + ], + "type": "text", + "content": "retinopathy (DR) have focused on addressing the issue of domain shift [5, 20, 23, 40]. The initial classic transfer learning strategies, Unsupervised Domain Adaptation (UDA), required access to well-annotated source data. Nevertheless, the growing emphasis on privacy protection has shifted research toward the Source-Free Domain Adaptation (SFDA) framework [23, 40]. SFDA involves adapting a source model—pre-trained on the source domain—to the target domain in a self-supervised manner, thereby ensuring the protection of source patient data." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 54, + 293, + 296, + 434 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 54, + 293, + 296, + 434 + ], + "spans": [ + { + "bbox": [ + 54, + 293, + 296, + 434 + ], + "type": "text", + "content": "In recent developments, specific needs have emerged in the clinical field. The introduction of model weight-based techniques for reconstructing training data has created a demand for model privacy [26, 39], which goes beyond traditional source data protection. In addition, there is a growing requirement for models capable of handling incoming patient data in a flowing fashion (test data arriving sequentially), referred to as a flowing data constraint [33, 34]. Unfortunately, existing SFDA methods cannot effectively address this challenge, as they rely on full access to the model and require offline training on a pre-collected dataset. Fig. 1 provides an intuitive illustration of this issue." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 54, + 437, + 296, + 544 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 54, + 437, + 296, + 544 + ], + "spans": [ + { + "bbox": [ + 54, + 437, + 296, + 544 + ], + "type": "text", + "content": "In this paper, we consider a clinically motivated setting, called Online Model-aGnostic Domain Adaptation (OMG-DA) and propose a novel Generative Unadversarial ExampleS (GUES) approach for the DR grading problem in this new setting. Specifically, OMG-DA presents an extreme safety scenario: The available target data is unlabeled and arrives in a flowing format, with no prior information about the pre-trained source model and data. Tab. 1 provides a detailed comparison with previous adaptation settings." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 54, + 545, + 296, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 54, + 545, + 296, + 713 + ], + "spans": [ + { + "bbox": [ + 54, + 545, + 296, + 713 + ], + "type": "text", + "content": "In GUES, we address the absence of source data and pre-trained models by producing generalized unadversarial examples [24] for unlabeled target data. To this end, we introduce generative unadversarial learning, which theoretically reformulates conventional iterative perturbation optimization. This new method aims to learn a generative function for perturbations and involves addressing two key tasks: (1) Identifying the latent function input, which is the derivative of initial random noise w.r.t. image data, and (2) selecting a self-supervised property to serve as pseudo-perturbation labels. In practice, we leverage a Variational Autoencoder (VAE) [10] based approach to facilitate this learning process. In terms of function representation, we model the latent input using the encoder along with the reparameterization trick," + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 313, + 172, + 555, + 232 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 172, + 555, + 232 + ], + "spans": [ + { + "bbox": [ + 313, + 172, + 555, + 232 + ], + "type": "text", + "content": "whilst the decoder accomplishes the generation of individual-perturbation. Additionally, we choose the saliency map as the pseudo-perturbation label for two reasons: (1) It helps discover potential lesions, and (2) it aids in identifying the latent input by providing an upper bound." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 324, + 232, + 515, + 243 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 324, + 232, + 515, + 243 + ], + "spans": [ + { + "bbox": [ + 324, + 232, + 515, + 243 + ], + "type": "text", + "content": "Our contributions are summarized as follows:" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 313, + 245, + 555, + 399 + ], + "type": "list", + "angle": 0, + "index": 11, + "blocks": [ + { + "bbox": [ + 313, + 245, + 553, + 292 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 245, + 553, + 292 + ], + "spans": [ + { + "bbox": [ + 313, + 245, + 553, + 292 + ], + "type": "text", + "content": "- Pioneering a novel transfer setting OMG-DA, which is closer to real-world clinical scenarios and meets three typical requirements at the same time, including (1) model absence, (2) flowing data, and (3) source data privacy." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 313, + 293, + 555, + 352 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 293, + 555, + 352 + ], + "spans": [ + { + "bbox": [ + 313, + 293, + 555, + 352 + ], + "type": "text", + "content": "- Developing a new OMG-DA approach GUES in the context of DR grading, grounded on the generative unadversarial examples theory, where we learn an individual-perturbation generative function under saliency map supervision, removing relying on labels and models." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 313, + 353, + 553, + 399 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 353, + 553, + 399 + ], + "spans": [ + { + "bbox": [ + 313, + 353, + 553, + 399 + ], + "type": "text", + "content": "- Extensive evaluations on four DR benchmarks, indicating that GUES can largely promote the source model's performance in the target domain, as well as trainable test-time adaptation models, even at small batch size." + } + ] + } + ], + "index": 10 + } + ], + "sub_type": "text" + }, + { + "bbox": [ + 313, + 413, + 399, + 425 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 413, + 399, + 425 + ], + "spans": [ + { + "bbox": [ + 313, + 413, + 399, + 425 + ], + "type": "text", + "content": "2. Related work" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 313, + 434, + 555, + 589 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 434, + 555, + 589 + ], + "spans": [ + { + "bbox": [ + 313, + 434, + 555, + 589 + ], + "type": "text", + "content": "Adaptation methods for DR grading. Driven by real-world medical requirements, domain adaptation has been an attractive topic in this DR grading issue. For instance, Nguyen et al. [20] introduce a UDA approach that enables the model to focus on vessel structures that remain invariant to domain shifts via image reconstruction using labeled source domain data. In SFDA, Zhang et al. [40] propose generating labeled, target-style retinal images to improve the source model's generalization, relying solely on a pre-trained source model and unlabeled target images. Additionally, DG in DR grading has been explored through domain-invariant feature learning approaches [4] and divergence-based methods [5], leveraging labeled source domain data." + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 313, + 590, + 556, + 686 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 590, + 556, + 686 + ], + "spans": [ + { + "bbox": [ + 313, + 590, + 556, + 686 + ], + "type": "text", + "content": "The methods above rely on labeled data, require full access to the model, and necessitate offline training on a precollected dataset. In real clinical settings, these requirements can be impractical due to constraints around data and model privacy, as well as the need for real-time adaptability without retraining. In contrast, GUES provides an online adaptation solution that operates without requiring labeled data or access to the model, addressing domain adaptive DR grading." + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 313, + 689, + 556, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 689, + 556, + 714 + ], + "spans": [ + { + "bbox": [ + 313, + 689, + 556, + 714 + ], + "type": "text", + "content": "Unadversarial learning. Unadversarial learning was initially developed by Salman et al. [24], aiming to modify" + } + ] + } + ], + "index": 15 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 318, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 318, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 318, + 757 + ], + "type": "text", + "content": "28338" + } + ] + } + ], + "index": 16 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 1 + }, + { + "para_blocks": [ + { + "bbox": [ + 55, + 72, + 296, + 192 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 72, + 296, + 192 + ], + "spans": [ + { + "bbox": [ + 55, + 72, + 296, + 192 + ], + "type": "text", + "content": "input image distribution to make them more easily recognizable by the model. Current mainstreams achieve this learning process by adding class-specific perturbations to the input images. Here, the perturbations are generated based on the gradient of an objective function w.r.t. image. This approach allows for the design of unadversarial examples without model training. For example, based on [24], NSA [25] introduces a method to generate more natural perturbations using a trainable generator. Similarly, CAT [18] demonstrates a new distance metric for generating unadversarial examples." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 55, + 193, + 296, + 277 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 193, + 296, + 277 + ], + "spans": [ + { + "bbox": [ + 55, + 193, + 296, + 277 + ], + "type": "text", + "content": "All existing unadversarial learning methods require access to model parameters, outputs, and labeled data. This dependency invalidates them in our OMG-DA setting which only has access to unlabeled target data. In addition to this, our GUES produces individualized unadversarial examples in a generative manner, which stands out from previous methods that focus on class-specific unadversarial examples." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 55, + 281, + 296, + 448 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 281, + 296, + 448 + ], + "spans": [ + { + "bbox": [ + 55, + 281, + 296, + 448 + ], + "type": "text", + "content": "Saliency map for medical image. A fine-grained saliency map is a pixel feature generated by calculating the central surround differences within images, identifying salient regions without any need for training [19]. This feature is widely applied in various medical image analysis tasks to extract pathologically important regions [9, 22, 31, 35]. For instance, in DR grading, studies such as [9, 22] utilize saliency maps to guide models in focusing on critical features like the optic disc, cup, and vessel structures. Similarly, in brain tumor detection, Tomar et al. [31] leverage saliency maps to enhance the model's attention on tumor and bone structures. In skin cancer detection, saliency maps help isolate lesion regions with distinctive features, such as lumpiness, which are essential for accurate diagnosis [35]." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 55, + 449, + 296, + 485 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 449, + 296, + 485 + ], + "spans": [ + { + "bbox": [ + 55, + 449, + 296, + 485 + ], + "type": "text", + "content": "As stated above, existing works primarily use saliency maps to highlight lesion regions. Unlike it, GUES selects saliency maps as pseudo-perturbation labels." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 55, + 499, + 138, + 514 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 499, + 138, + 514 + ], + "spans": [ + { + "bbox": [ + 55, + 499, + 138, + 514 + ], + "type": "text", + "content": "3. Methodology" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 55, + 521, + 296, + 628 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 521, + 296, + 628 + ], + "spans": [ + { + "bbox": [ + 55, + 521, + 296, + 628 + ], + "type": "text", + "content": "Problem statement. Given two different but related domains, i.e., source domain " + }, + { + "bbox": [ + 55, + 521, + 296, + 628 + ], + "type": "inline_equation", + "content": "\\mathcal{S}" + }, + { + "bbox": [ + 55, + 521, + 296, + 628 + ], + "type": "text", + "content": " and target domain " + }, + { + "bbox": [ + 55, + 521, + 296, + 628 + ], + "type": "inline_equation", + "content": "\\mathcal{T}" + }, + { + "bbox": [ + 55, + 521, + 296, + 628 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 55, + 521, + 296, + 628 + ], + "type": "inline_equation", + "content": "\\mathcal{S}" + }, + { + "bbox": [ + 55, + 521, + 296, + 628 + ], + "type": "text", + "content": " contains " + }, + { + "bbox": [ + 55, + 521, + 296, + 628 + ], + "type": "inline_equation", + "content": "n_s" + }, + { + "bbox": [ + 55, + 521, + 296, + 628 + ], + "type": "text", + "content": " labeled samples, while " + }, + { + "bbox": [ + 55, + 521, + 296, + 628 + ], + "type": "inline_equation", + "content": "\\mathcal{T}" + }, + { + "bbox": [ + 55, + 521, + 296, + 628 + ], + "type": "text", + "content": " has " + }, + { + "bbox": [ + 55, + 521, + 296, + 628 + ], + "type": "inline_equation", + "content": "n" + }, + { + "bbox": [ + 55, + 521, + 296, + 628 + ], + "type": "text", + "content": " unlabeled data. Both labeled and unlabeled samples share the same " + }, + { + "bbox": [ + 55, + 521, + 296, + 628 + ], + "type": "inline_equation", + "content": "C" + }, + { + "bbox": [ + 55, + 521, + 296, + 628 + ], + "type": "text", + "content": " categories. Let " + }, + { + "bbox": [ + 55, + 521, + 296, + 628 + ], + "type": "inline_equation", + "content": "\\mathcal{X}_s" + }, + { + "bbox": [ + 55, + 521, + 296, + 628 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 55, + 521, + 296, + 628 + ], + "type": "inline_equation", + "content": "\\mathcal{Y}_s" + }, + { + "bbox": [ + 55, + 521, + 296, + 628 + ], + "type": "text", + "content": " be the source samples and the corresponding labels. Similarly, we denote the target samples and their labels by " + }, + { + "bbox": [ + 55, + 521, + 296, + 628 + ], + "type": "inline_equation", + "content": "\\mathcal{X}_t = \\{\\pmb{x}_t^i\\}_{i=1}^n" + }, + { + "bbox": [ + 55, + 521, + 296, + 628 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 55, + 521, + 296, + 628 + ], + "type": "inline_equation", + "content": "\\mathcal{Y}_t = \\{y_t^i\\}_{i=1}^n" + }, + { + "bbox": [ + 55, + 521, + 296, + 628 + ], + "type": "text", + "content": ", respectively, where " + }, + { + "bbox": [ + 55, + 521, + 296, + 628 + ], + "type": "inline_equation", + "content": "n" + }, + { + "bbox": [ + 55, + 521, + 296, + 628 + ], + "type": "text", + "content": " signifies the number of samples. The source model " + }, + { + "bbox": [ + 55, + 521, + 296, + 628 + ], + "type": "inline_equation", + "content": "\\theta_s" + }, + { + "bbox": [ + 55, + 521, + 296, + 628 + ], + "type": "text", + "content": " is pre-trained on " + }, + { + "bbox": [ + 55, + 521, + 296, + 628 + ], + "type": "inline_equation", + "content": "\\{\\mathcal{X}_s, \\mathcal{Y}_s\\}" + }, + { + "bbox": [ + 55, + 521, + 296, + 628 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 55, + 630, + 296, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 630, + 296, + 713 + ], + "spans": [ + { + "bbox": [ + 55, + 630, + 296, + 713 + ], + "type": "text", + "content": "OMG-DA is featured in (1) the absence of the source model " + }, + { + "bbox": [ + 55, + 630, + 296, + 713 + ], + "type": "inline_equation", + "content": "\\theta_{s}" + }, + { + "bbox": [ + 55, + 630, + 296, + 713 + ], + "type": "text", + "content": " and domain " + }, + { + "bbox": [ + 55, + 630, + 296, + 713 + ], + "type": "inline_equation", + "content": "S" + }, + { + "bbox": [ + 55, + 630, + 296, + 713 + ], + "type": "text", + "content": " and (2) the flowing target data " + }, + { + "bbox": [ + 55, + 630, + 296, + 713 + ], + "type": "inline_equation", + "content": "\\mathcal{X}_t" + }, + { + "bbox": [ + 55, + 630, + 296, + 713 + ], + "type": "text", + "content": " the same as the TTA setting [34]. Unlike previous transfer settings that are model adaptation-centered [15, 29, 30, 38], OMG-DA considers adaptation from the perspective of data. Specifically, OMG-DA aims to modify the distribution of target data to facilitate downstream tasks." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 313, + 72, + 507, + 85 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 72, + 507, + 85 + ], + "spans": [ + { + "bbox": [ + 313, + 72, + 507, + 85 + ], + "type": "text", + "content": "3.1. Generative Unadversarial Examples" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 313, + 89, + 555, + 186 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 89, + 555, + 186 + ], + "spans": [ + { + "bbox": [ + 313, + 89, + 555, + 186 + ], + "type": "text", + "content": "The unadversarial learning idea [24] is well aligned with OMG-DA in terms of promoting adaptation by enhancing the generalization of input data. We briefly recap traditional unadversarial learning. Unlike adversarial learning [28] that generates confusing samples to mislead models, unadversarial learning aims to construct generalized samples, tackling out-of-distribution issues. Formally, this learning can be summarized in the following optimization problem." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 347, + 194, + 555, + 213 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 347, + 194, + 555, + 213 + ], + "spans": [ + { + "bbox": [ + 347, + 194, + 555, + 213 + ], + "type": "interline_equation", + "content": "\\hat {\\delta} = \\arg \\min _ {\\delta} L \\left(f _ {\\theta} (x + \\delta), y\\right), s. t. | | \\delta | | \\leq \\epsilon \\tag {1}", + "image_path": "e64934289bd9822134111b0ebee5c89bb58c67d970f1b98d41bcacc26577cf0d.jpg" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 313, + 220, + 554, + 281 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 220, + 554, + 281 + ], + "spans": [ + { + "bbox": [ + 313, + 220, + 554, + 281 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 313, + 220, + 554, + 281 + ], + "type": "inline_equation", + "content": "L(\\cdot)" + }, + { + "bbox": [ + 313, + 220, + 554, + 281 + ], + "type": "text", + "content": " denotes objective function, e.g., cross-entropy loss for classification tasks, " + }, + { + "bbox": [ + 313, + 220, + 554, + 281 + ], + "type": "inline_equation", + "content": "x" + }, + { + "bbox": [ + 313, + 220, + 554, + 281 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 313, + 220, + 554, + 281 + ], + "type": "inline_equation", + "content": "y" + }, + { + "bbox": [ + 313, + 220, + 554, + 281 + ], + "type": "text", + "content": " are input image and its label, " + }, + { + "bbox": [ + 313, + 220, + 554, + 281 + ], + "type": "inline_equation", + "content": "f_{\\theta}" + }, + { + "bbox": [ + 313, + 220, + 554, + 281 + ], + "type": "text", + "content": " is a pre-trained model with parameters " + }, + { + "bbox": [ + 313, + 220, + 554, + 281 + ], + "type": "inline_equation", + "content": "\\theta" + }, + { + "bbox": [ + 313, + 220, + 554, + 281 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 313, + 220, + 554, + 281 + ], + "type": "inline_equation", + "content": "\\delta" + }, + { + "bbox": [ + 313, + 220, + 554, + 281 + ], + "type": "text", + "content": " is a perturbation, " + }, + { + "bbox": [ + 313, + 220, + 554, + 281 + ], + "type": "inline_equation", + "content": "\\epsilon" + }, + { + "bbox": [ + 313, + 220, + 554, + 281 + ], + "type": "text", + "content": " is a small threshold. The current scheme solves this problem in an iterative way formulated as" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 318, + 287, + 554, + 310 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 318, + 287, + 554, + 310 + ], + "spans": [ + { + "bbox": [ + 318, + 287, + 554, + 310 + ], + "type": "interline_equation", + "content": "\\delta_ {k + 1} = \\delta_ {k} + \\alpha \\cdot \\operatorname {s i g n} \\left(\\nabla_ {x} L \\left(f _ {\\theta} \\left(x + \\delta_ {k}\\right), y\\right)\\right), k \\in [ 0, K - 1 ], \\tag {2}", + "image_path": "a36473bd4143b8296f58d1fc25a75c7ae5dff5fe64e6f6a51bd69cb636dd85a2.jpg" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 313, + 310, + 554, + 393 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 310, + 554, + 393 + ], + "spans": [ + { + "bbox": [ + 313, + 310, + 554, + 393 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 313, + 310, + 554, + 393 + ], + "type": "inline_equation", + "content": "\\alpha" + }, + { + "bbox": [ + 313, + 310, + 554, + 393 + ], + "type": "text", + "content": " is a trade-off parameter, " + }, + { + "bbox": [ + 313, + 310, + 554, + 393 + ], + "type": "inline_equation", + "content": "K" + }, + { + "bbox": [ + 313, + 310, + 554, + 393 + ], + "type": "text", + "content": " is iteration number, " + }, + { + "bbox": [ + 313, + 310, + 554, + 393 + ], + "type": "inline_equation", + "content": "\\delta_0" + }, + { + "bbox": [ + 313, + 310, + 554, + 393 + ], + "type": "text", + "content": " is an initial random noise. In the inference phase, the optimal perturbation " + }, + { + "bbox": [ + 313, + 310, + 554, + 393 + ], + "type": "inline_equation", + "content": "\\hat{\\delta}" + }, + { + "bbox": [ + 313, + 310, + 554, + 393 + ], + "type": "text", + "content": " is integrated into the input " + }, + { + "bbox": [ + 313, + 310, + 554, + 393 + ], + "type": "inline_equation", + "content": "x" + }, + { + "bbox": [ + 313, + 310, + 554, + 393 + ], + "type": "text", + "content": ", forming an unadversarial example " + }, + { + "bbox": [ + 313, + 310, + 554, + 393 + ], + "type": "inline_equation", + "content": "\\hat{x} = x + \\hat{\\delta}" + }, + { + "bbox": [ + 313, + 310, + 554, + 393 + ], + "type": "text", + "content": ", which is easily recognizable by the model " + }, + { + "bbox": [ + 313, + 310, + 554, + 393 + ], + "type": "inline_equation", + "content": "f_{\\theta}" + }, + { + "bbox": [ + 313, + 310, + 554, + 393 + ], + "type": "text", + "content": ". Obviously, the conventional unadversarial paradigm cannot meet our OMG-DA setting due to the absence of " + }, + { + "bbox": [ + 313, + 310, + 554, + 393 + ], + "type": "inline_equation", + "content": "f_{\\theta}" + }, + { + "bbox": [ + 313, + 310, + 554, + 393 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 313, + 310, + 554, + 393 + ], + "type": "inline_equation", + "content": "L" + }, + { + "bbox": [ + 313, + 310, + 554, + 393 + ], + "type": "text", + "content": ", and the label " + }, + { + "bbox": [ + 313, + 310, + 554, + 393 + ], + "type": "inline_equation", + "content": "y" + }, + { + "bbox": [ + 313, + 310, + 554, + 393 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 313, + 394, + 554, + 430 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 394, + 554, + 430 + ], + "spans": [ + { + "bbox": [ + 313, + 394, + 554, + 430 + ], + "type": "text", + "content": "In this paper, we re-consider the iterative optimization process above and obtain the theorem below (The proof is provided in Supplementary)." + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 313, + 437, + 555, + 473 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 437, + 555, + 473 + ], + "spans": [ + { + "bbox": [ + 313, + 437, + 555, + 473 + ], + "type": "text", + "content": "Theorem 1 Given the unadversarial learning problem defined in Eq. (1), the iterative process featured by Eq. (2) can be expressed as the following generative form." + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 377, + 480, + 555, + 507 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 377, + 480, + 555, + 507 + ], + "spans": [ + { + "bbox": [ + 377, + 480, + 555, + 507 + ], + "type": "interline_equation", + "content": "\\delta_ {k} = \\delta_ {0} + V \\cdot F _ {\\Phi} \\left(\\frac {\\partial \\delta_ {0}}{\\partial x}\\right), \\tag {3}", + "image_path": "2f88bb07f309379eb723e8aea0b90464b4920fa7b5bc73ae3dea1874ad5a99fa.jpg" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 313, + 514, + 555, + 540 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 514, + 555, + 540 + ], + "spans": [ + { + "bbox": [ + 313, + 514, + 555, + 540 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 313, + 514, + 555, + 540 + ], + "type": "inline_equation", + "content": "\\delta_0" + }, + { + "bbox": [ + 313, + 514, + 555, + 540 + ], + "type": "text", + "content": " is an initial random noise, " + }, + { + "bbox": [ + 313, + 514, + 555, + 540 + ], + "type": "inline_equation", + "content": "V > 0" + }, + { + "bbox": [ + 313, + 514, + 555, + 540 + ], + "type": "text", + "content": " is a bound constant, " + }, + { + "bbox": [ + 313, + 514, + 555, + 540 + ], + "type": "inline_equation", + "content": "F_{\\Phi}" + }, + { + "bbox": [ + 313, + 514, + 555, + 540 + ], + "type": "text", + "content": " is a generative function, " + }, + { + "bbox": [ + 313, + 514, + 555, + 540 + ], + "type": "inline_equation", + "content": "\\frac{\\partial\\delta_0}{\\partial x}" + }, + { + "bbox": [ + 313, + 514, + 555, + 540 + ], + "type": "text", + "content": " is a latent variable." + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 313, + 556, + 555, + 628 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 556, + 555, + 628 + ], + "spans": [ + { + "bbox": [ + 313, + 556, + 555, + 628 + ], + "type": "text", + "content": "Grounded on Theorem 1, we have: When " + }, + { + "bbox": [ + 313, + 556, + 555, + 628 + ], + "type": "inline_equation", + "content": "\\delta_{k}" + }, + { + "bbox": [ + 313, + 556, + 555, + 628 + ], + "type": "text", + "content": " converges to optimal " + }, + { + "bbox": [ + 313, + 556, + 555, + 628 + ], + "type": "inline_equation", + "content": "\\hat{\\delta}" + }, + { + "bbox": [ + 313, + 556, + 555, + 628 + ], + "type": "text", + "content": ", i.e., " + }, + { + "bbox": [ + 313, + 556, + 555, + 628 + ], + "type": "inline_equation", + "content": "\\delta_{k} \\rightarrow \\hat{\\delta}" + }, + { + "bbox": [ + 313, + 556, + 555, + 628 + ], + "type": "text", + "content": ", function " + }, + { + "bbox": [ + 313, + 556, + 555, + 628 + ], + "type": "inline_equation", + "content": "F_{\\Phi}" + }, + { + "bbox": [ + 313, + 556, + 555, + 628 + ], + "type": "text", + "content": " also evolves to the optimal one, denoted by " + }, + { + "bbox": [ + 313, + 556, + 555, + 628 + ], + "type": "inline_equation", + "content": "\\hat{F}_{\\Phi}" + }, + { + "bbox": [ + 313, + 556, + 555, + 628 + ], + "type": "text", + "content": ", i.e., " + }, + { + "bbox": [ + 313, + 556, + 555, + 628 + ], + "type": "inline_equation", + "content": "F_{\\Phi} \\rightarrow \\hat{F}_{\\Phi}" + }, + { + "bbox": [ + 313, + 556, + 555, + 628 + ], + "type": "text", + "content": ". This provides an insight: The unadversarial learning problem above can also be solved in a generative fashion. Correspondingly, the generated data are termed generative unadversarial examples." + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 313, + 635, + 430, + 647 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 635, + 430, + 647 + ], + "spans": [ + { + "bbox": [ + 313, + 635, + 430, + 647 + ], + "type": "text", + "content": "3.2. Model Instantiation" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 313, + 654, + 554, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 654, + 554, + 713 + ], + "spans": [ + { + "bbox": [ + 313, + 654, + 554, + 713 + ], + "type": "text", + "content": "Within this context of generative unadversarial learning, conventional unadversarial learning presented in Eq. (1) is boiled down to learning " + }, + { + "bbox": [ + 313, + 654, + 554, + 713 + ], + "type": "inline_equation", + "content": "F_{\\Phi}\\left(\\frac{\\partial\\delta_0}{\\partial x}\\right)" + }, + { + "bbox": [ + 313, + 654, + 554, + 713 + ], + "type": "text", + "content": ". We can achieve this by training a generative neural network. To make this solution sense, we have to solve two difficulties as follows. (A) One is" + } + ] + } + ], + "index": 19 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "text", + "content": "28339" + } + ] + } + ], + "index": 20 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 2 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 82, + 70, + 531, + 242 + ], + "blocks": [ + { + "bbox": [ + 82, + 70, + 531, + 242 + ], + "lines": [ + { + "bbox": [ + 82, + 70, + 531, + 242 + ], + "spans": [ + { + "bbox": [ + 82, + 70, + 531, + 242 + ], + "type": "image", + "image_path": "a7c3ea92141db9da1aa0dd614c6fc7178718c4c132cabf7b92e49b8adc76978e.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 55, + 251, + 555, + 297 + ], + "lines": [ + { + "bbox": [ + 55, + 251, + 555, + 297 + ], + "spans": [ + { + "bbox": [ + 55, + 251, + 555, + 297 + ], + "type": "text", + "content": "Figure 2. The instantiation framework of GUES in the OMG-DA setting. (a) For target input " + }, + { + "bbox": [ + 55, + 251, + 555, + 297 + ], + "type": "inline_equation", + "content": "x_{t}" + }, + { + "bbox": [ + 55, + 251, + 555, + 297 + ], + "type": "text", + "content": ", the VAE model generates individual perturbation " + }, + { + "bbox": [ + 55, + 251, + 555, + 297 + ], + "type": "inline_equation", + "content": "\\delta_{t} = F_{\\Phi}\\left(\\frac{\\partial\\delta_{0}}{\\partial x}\\right)" + }, + { + "bbox": [ + 55, + 251, + 555, + 297 + ], + "type": "text", + "content": ". After that, the by-pass path incorporates " + }, + { + "bbox": [ + 55, + 251, + 555, + 297 + ], + "type": "inline_equation", + "content": "\\delta_{t}" + }, + { + "bbox": [ + 55, + 251, + 555, + 297 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 55, + 251, + 555, + 297 + ], + "type": "inline_equation", + "content": "x_{t}" + }, + { + "bbox": [ + 55, + 251, + 555, + 297 + ], + "type": "text", + "content": " to create the generative unadversarial example " + }, + { + "bbox": [ + 55, + 251, + 555, + 297 + ], + "type": "inline_equation", + "content": "\\hat{x}_{t}" + }, + { + "bbox": [ + 55, + 251, + 555, + 297 + ], + "type": "text", + "content": ". Treating " + }, + { + "bbox": [ + 55, + 251, + 555, + 297 + ], + "type": "inline_equation", + "content": "x_{t}" + }, + { + "bbox": [ + 55, + 251, + 555, + 297 + ], + "type": "text", + "content": "'s saliency map " + }, + { + "bbox": [ + 55, + 251, + 555, + 297 + ], + "type": "inline_equation", + "content": "g_{t}" + }, + { + "bbox": [ + 55, + 251, + 555, + 297 + ], + "type": "text", + "content": " as reconstruction supervision for model training. (b) At the inference phase, the generated unadversarial example " + }, + { + "bbox": [ + 55, + 251, + 555, + 297 + ], + "type": "inline_equation", + "content": "\\hat{x}_{t}" + }, + { + "bbox": [ + 55, + 251, + 555, + 297 + ], + "type": "text", + "content": " is directly provided to the frozen source model or other trainable models." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "type": "image", + "bbox": [ + 69, + 314, + 282, + 397 + ], + "blocks": [ + { + "bbox": [ + 69, + 314, + 282, + 397 + ], + "lines": [ + { + "bbox": [ + 69, + 314, + 282, + 397 + ], + "spans": [ + { + "bbox": [ + 69, + 314, + 282, + 397 + ], + "type": "image", + "image_path": "df185912946c69dbba5c9887b58cd479347ea57e68b6d91b7216902aed39df4b.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 55, + 405, + 296, + 482 + ], + "lines": [ + { + "bbox": [ + 55, + 405, + 296, + 482 + ], + "spans": [ + { + "bbox": [ + 55, + 405, + 296, + 482 + ], + "type": "text", + "content": "Figure 3. Explanation in choosing fine-grained saliency maps as supervision. Left: The testing fundus image selected from \"Moderate DR\" class in APTOS demonstrates that H (hemorrhages), SE (soft exudates), and EX (hard exudates) are essential characteristics to judge the DR grade. Middle: The saliency map highlights those lesions. Right: The gradient-CAM visualization of the source model on task DDR→APTOS." + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_caption" + } + ], + "index": 2 + }, + { + "bbox": [ + 55, + 491, + 296, + 529 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 491, + 296, + 529 + ], + "spans": [ + { + "bbox": [ + 55, + 491, + 296, + 529 + ], + "type": "text", + "content": "the identification of " + }, + { + "bbox": [ + 55, + 491, + 296, + 529 + ], + "type": "inline_equation", + "content": "\\frac{\\partial\\delta_0}{\\partial x}" + }, + { + "bbox": [ + 55, + 491, + 296, + 529 + ], + "type": "text", + "content": " when the relationship between them is unknown. (B) The other is selecting property supervision (pseudo-perturbation labels) to drive " + }, + { + "bbox": [ + 55, + 491, + 296, + 529 + ], + "type": "inline_equation", + "content": "\\delta_{k}\\rightarrow \\hat{\\delta}" + }, + { + "bbox": [ + 55, + 491, + 296, + 529 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 55, + 531, + 296, + 674 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 531, + 296, + 674 + ], + "spans": [ + { + "bbox": [ + 55, + 531, + 296, + 674 + ], + "type": "text", + "content": "Solution to Problem A. In practice, considering derivative " + }, + { + "bbox": [ + 55, + 531, + 296, + 674 + ], + "type": "inline_equation", + "content": "\\frac{\\partial\\delta_0}{\\partial x}" + }, + { + "bbox": [ + 55, + 531, + 296, + 674 + ], + "type": "text", + "content": " is relevant with both random noise " + }, + { + "bbox": [ + 55, + 531, + 296, + 674 + ], + "type": "inline_equation", + "content": "\\delta_0" + }, + { + "bbox": [ + 55, + 531, + 296, + 674 + ], + "type": "text", + "content": " and input image " + }, + { + "bbox": [ + 55, + 531, + 296, + 674 + ], + "type": "inline_equation", + "content": "x" + }, + { + "bbox": [ + 55, + 531, + 296, + 674 + ], + "type": "text", + "content": ", we sample it from a certain Gaussian distribution associated with " + }, + { + "bbox": [ + 55, + 531, + 296, + 674 + ], + "type": "inline_equation", + "content": "x" + }, + { + "bbox": [ + 55, + 531, + 296, + 674 + ], + "type": "text", + "content": ". Furthermore, the output size of " + }, + { + "bbox": [ + 55, + 531, + 296, + 674 + ], + "type": "inline_equation", + "content": "F_{\\Phi}" + }, + { + "bbox": [ + 55, + 531, + 296, + 674 + ], + "type": "text", + "content": " is the same as " + }, + { + "bbox": [ + 55, + 531, + 296, + 674 + ], + "type": "inline_equation", + "content": "x" + }, + { + "bbox": [ + 55, + 531, + 296, + 674 + ], + "type": "text", + "content": ". Therefore, we employ the VAE model to jointly model " + }, + { + "bbox": [ + 55, + 531, + 296, + 674 + ], + "type": "inline_equation", + "content": "\\frac{\\partial\\delta_0}{\\partial x}" + }, + { + "bbox": [ + 55, + 531, + 296, + 674 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 55, + 531, + 296, + 674 + ], + "type": "inline_equation", + "content": "F_{\\Phi}" + }, + { + "bbox": [ + 55, + 531, + 296, + 674 + ], + "type": "text", + "content": ", since VAE is an autoencoder characterized by random sampling. Specifically, as shown in Fig.2 (a), we approximate " + }, + { + "bbox": [ + 55, + 531, + 296, + 674 + ], + "type": "inline_equation", + "content": "\\frac{\\partial\\delta_0}{\\partial x}" + }, + { + "bbox": [ + 55, + 531, + 296, + 674 + ], + "type": "text", + "content": " by latent variable " + }, + { + "bbox": [ + 55, + 531, + 296, + 674 + ], + "type": "inline_equation", + "content": "z" + }, + { + "bbox": [ + 55, + 531, + 296, + 674 + ], + "type": "text", + "content": ", which is jointly determined by input " + }, + { + "bbox": [ + 55, + 531, + 296, + 674 + ], + "type": "inline_equation", + "content": "x" + }, + { + "bbox": [ + 55, + 531, + 296, + 674 + ], + "type": "text", + "content": " and sampled random signal " + }, + { + "bbox": [ + 55, + 531, + 296, + 674 + ], + "type": "inline_equation", + "content": "\\tau" + }, + { + "bbox": [ + 55, + 531, + 296, + 674 + ], + "type": "text", + "content": ". As for " + }, + { + "bbox": [ + 55, + 531, + 296, + 674 + ], + "type": "inline_equation", + "content": "F_{\\Phi}" + }, + { + "bbox": [ + 55, + 531, + 296, + 674 + ], + "type": "text", + "content": ", it is demonstrated by the decoder module. Suppose " + }, + { + "bbox": [ + 55, + 531, + 296, + 674 + ], + "type": "inline_equation", + "content": "D(\\cdot)" + }, + { + "bbox": [ + 55, + 531, + 296, + 674 + ], + "type": "text", + "content": " is the decoder, " + }, + { + "bbox": [ + 55, + 531, + 296, + 674 + ], + "type": "inline_equation", + "content": "E_{\\tau}(\\cdot)" + }, + { + "bbox": [ + 55, + 531, + 296, + 674 + ], + "type": "text", + "content": " is the encoder with reparameterization trick, our scheme can be formulated as" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 63, + 680, + 296, + 713 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 63, + 680, + 296, + 713 + ], + "spans": [ + { + "bbox": [ + 63, + 680, + 296, + 713 + ], + "type": "interline_equation", + "content": "D \\left(E _ {\\tau} (x)\\right)\\rightarrow F _ {\\Phi} \\left(\\frac {\\partial \\delta_ {0}}{\\partial x}\\right), E _ {\\tau} (x) \\rightarrow \\frac {\\partial \\delta_ {0}}{\\partial x}, D (\\cdot) \\rightarrow F _ {\\Phi} (\\cdot). \\tag {4}", + "image_path": "d7683f5dc83193586daa9444382fbf3663c78430498bc970ad4a300ab106efe2.jpg" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 313, + 316, + 555, + 484 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 316, + 555, + 484 + ], + "spans": [ + { + "bbox": [ + 313, + 316, + 555, + 484 + ], + "type": "text", + "content": "Solution to Problem B. We adopt the fine-grained saliency map as the supervision. Two reasons contribute to our selection. First of all, the empirical results show that, for the specific task of DR grading, the saliency map is an acceptable pseudo-perturbation. Specifically, perturbation in the unadversarial context enhances the regions associated with the category and reduces the prominence of other areas, thereby identifying the lesion zones relevant to DR grading. As illustrated in Fig. 3, the saliency map effectively identifies potential lesions, such as hemorrhages, soft exudates, and hard exudates. Furthermore, it includes gradient information, as it highlights regions similar to Grad-CAM (Right side). More importantly, we have the theorem below. (The proof is provided in Supplementary)" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 313, + 492, + 555, + 540 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 492, + 555, + 540 + ], + "spans": [ + { + "bbox": [ + 313, + 492, + 555, + 540 + ], + "type": "text", + "content": "Theorem 2 Given the partial derivatives of the initial random noise " + }, + { + "bbox": [ + 313, + 492, + 555, + 540 + ], + "type": "inline_equation", + "content": "\\delta_0" + }, + { + "bbox": [ + 313, + 492, + 555, + 540 + ], + "type": "text", + "content": " w.r.t image " + }, + { + "bbox": [ + 313, + 492, + 555, + 540 + ], + "type": "inline_equation", + "content": "x" + }, + { + "bbox": [ + 313, + 492, + 555, + 540 + ], + "type": "text", + "content": " is " + }, + { + "bbox": [ + 313, + 492, + 555, + 540 + ], + "type": "inline_equation", + "content": "\\frac{\\partial\\delta_0}{\\partial x}" + }, + { + "bbox": [ + 313, + 492, + 555, + 540 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 313, + 492, + 555, + 540 + ], + "type": "inline_equation", + "content": "x" + }, + { + "bbox": [ + 313, + 492, + 555, + 540 + ], + "type": "text", + "content": "'s saliency map is " + }, + { + "bbox": [ + 313, + 492, + 555, + 540 + ], + "type": "inline_equation", + "content": "s = G(x)" + }, + { + "bbox": [ + 313, + 492, + 555, + 540 + ], + "type": "text", + "content": " where " + }, + { + "bbox": [ + 313, + 492, + 555, + 540 + ], + "type": "inline_equation", + "content": "G(\\cdot)" + }, + { + "bbox": [ + 313, + 492, + 555, + 540 + ], + "type": "text", + "content": " is the computation function of the saliency map. We have the following relationship:" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 406, + 548, + 555, + 572 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 406, + 548, + 555, + 572 + ], + "spans": [ + { + "bbox": [ + 406, + 548, + 555, + 572 + ], + "type": "interline_equation", + "content": "\\frac {\\partial \\delta_ {0}}{\\partial x} \\leq U \\cdot s, \\tag {5}", + "image_path": "36b2e05cf22dd328542d5cd015acd2a618817cadcbc2adb0d95d4539e979cc20.jpg" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 313, + 580, + 453, + 591 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 580, + 453, + 591 + ], + "spans": [ + { + "bbox": [ + 313, + 580, + 453, + 591 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 313, + 580, + 453, + 591 + ], + "type": "inline_equation", + "content": "U > 0" + }, + { + "bbox": [ + 313, + 580, + 453, + 591 + ], + "type": "text", + "content": " is a bound constant." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 313, + 600, + 555, + 639 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 600, + 555, + 639 + ], + "spans": [ + { + "bbox": [ + 313, + 600, + 555, + 639 + ], + "type": "text", + "content": "Theorem 2 suggests that the saliency map provides upper bounds for " + }, + { + "bbox": [ + 313, + 600, + 555, + 639 + ], + "type": "inline_equation", + "content": "\\frac{\\partial\\delta_0}{\\partial x}" + }, + { + "bbox": [ + 313, + 600, + 555, + 639 + ], + "type": "text", + "content": ". Namely, " + }, + { + "bbox": [ + 313, + 600, + 555, + 639 + ], + "type": "inline_equation", + "content": "s" + }, + { + "bbox": [ + 313, + 600, + 555, + 639 + ], + "type": "text", + "content": " provides relaxed descriptions for the variation " + }, + { + "bbox": [ + 313, + 600, + 555, + 639 + ], + "type": "inline_equation", + "content": "\\frac{\\partial\\delta_0}{\\partial x}" + }, + { + "bbox": [ + 313, + 600, + 555, + 639 + ], + "type": "text", + "content": ". This can help guide the learning of " + }, + { + "bbox": [ + 313, + 600, + 555, + 639 + ], + "type": "inline_equation", + "content": "\\frac{\\partial\\delta_0}{\\partial x}" + }, + { + "bbox": [ + 313, + 600, + 555, + 639 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 313, + 639, + 556, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 639, + 556, + 714 + ], + "spans": [ + { + "bbox": [ + 313, + 639, + 556, + 714 + ], + "type": "text", + "content": "GUES framework. Based on the analysis above, we instantiate GUES as the framework depicted in Fig. 2. As shown in sub-figure (a), our method integrates a VAE model and by-pass connection, achieving the learning of " + }, + { + "bbox": [ + 313, + 639, + 556, + 714 + ], + "type": "inline_equation", + "content": "F_{\\Phi}\\left(\\frac{\\partial \\delta_0}{\\partial x}\\right)" + }, + { + "bbox": [ + 313, + 639, + 556, + 714 + ], + "type": "text", + "content": ". Specifically, " + }, + { + "bbox": [ + 313, + 639, + 556, + 714 + ], + "type": "inline_equation", + "content": "\\frac{\\partial \\delta_0}{\\partial x}" + }, + { + "bbox": [ + 313, + 639, + 556, + 714 + ], + "type": "text", + "content": " is sampled from an input " + }, + { + "bbox": [ + 313, + 639, + 556, + 714 + ], + "type": "inline_equation", + "content": "x_t" + }, + { + "bbox": [ + 313, + 639, + 556, + 714 + ], + "type": "text", + "content": "-featured Gaussian distribution " + }, + { + "bbox": [ + 313, + 639, + 556, + 714 + ], + "type": "inline_equation", + "content": "N(\\mu, \\sigma)" + }, + { + "bbox": [ + 313, + 639, + 556, + 714 + ], + "type": "text", + "content": ", which is jointly learned using the" + } + ] + } + ], + "index": 12 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 318, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 318, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 318, + 757 + ], + "type": "text", + "content": "28340" + } + ] + } + ], + "index": 13 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 3 + }, + { + "para_blocks": [ + { + "bbox": [ + 56, + 72, + 241, + 83 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 72, + 241, + 83 + ], + "spans": [ + { + "bbox": [ + 56, + 72, + 241, + 83 + ], + "type": "text", + "content": "Algorithm 1 The pipeline of proposed GUES" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 55, + 85, + 294, + 109 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 85, + 294, + 109 + ], + "spans": [ + { + "bbox": [ + 55, + 85, + 294, + 109 + ], + "type": "text", + "content": "Input: Online batch samples " + }, + { + "bbox": [ + 55, + 85, + 294, + 109 + ], + "type": "inline_equation", + "content": "\\mathcal{B}" + }, + { + "bbox": [ + 55, + 85, + 294, + 109 + ], + "type": "text", + "content": ", a trainable " + }, + { + "bbox": [ + 55, + 85, + 294, + 109 + ], + "type": "inline_equation", + "content": "\\mathrm{VAE}\\theta_v" + }, + { + "bbox": [ + 55, + 85, + 294, + 109 + ], + "type": "text", + "content": " consists of an encoder " + }, + { + "bbox": [ + 55, + 85, + 294, + 109 + ], + "type": "inline_equation", + "content": "E_{\\tau}(\\cdot)" + }, + { + "bbox": [ + 55, + 85, + 294, + 109 + ], + "type": "text", + "content": " and a decoder " + }, + { + "bbox": [ + 55, + 85, + 294, + 109 + ], + "type": "inline_equation", + "content": "D(\\cdot)" + }, + { + "bbox": [ + 55, + 85, + 294, + 109 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 56, + 110, + 105, + 120 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 110, + 105, + 120 + ], + "spans": [ + { + "bbox": [ + 56, + 110, + 105, + 120 + ], + "type": "text", + "content": "Procedure:" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 62, + 123, + 294, + 219 + ], + "type": "list", + "angle": 0, + "index": 10, + "blocks": [ + { + "bbox": [ + 62, + 123, + 132, + 134 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 123, + 132, + 134 + ], + "spans": [ + { + "bbox": [ + 62, + 123, + 132, + 134 + ], + "type": "text", + "content": "1: for " + }, + { + "bbox": [ + 62, + 123, + 132, + 134 + ], + "type": "inline_equation", + "content": "x_{i}" + }, + { + "bbox": [ + 62, + 123, + 132, + 134 + ], + "type": "text", + "content": " in " + }, + { + "bbox": [ + 62, + 123, + 132, + 134 + ], + "type": "inline_equation", + "content": "\\mathcal{B}" + }, + { + "bbox": [ + 62, + 123, + 132, + 134 + ], + "type": "text", + "content": " do" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 62, + 135, + 283, + 147 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 135, + 283, + 147 + ], + "spans": [ + { + "bbox": [ + 62, + 135, + 283, + 147 + ], + "type": "text", + "content": "2: Approximate the latent function input by " + }, + { + "bbox": [ + 62, + 135, + 283, + 147 + ], + "type": "inline_equation", + "content": "E_{\\tau}(x_i)" + }, + { + "bbox": [ + 62, + 135, + 283, + 147 + ], + "type": "text", + "content": ";" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 62, + 148, + 292, + 159 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 148, + 292, + 159 + ], + "spans": [ + { + "bbox": [ + 62, + 148, + 292, + 159 + ], + "type": "text", + "content": "3: Learn a generative function for perturbation " + }, + { + "bbox": [ + 62, + 148, + 292, + 159 + ], + "type": "inline_equation", + "content": "\\delta_{i}" + }, + { + "bbox": [ + 62, + 148, + 292, + 159 + ], + "type": "text", + "content": " by " + }, + { + "bbox": [ + 62, + 148, + 292, + 159 + ], + "type": "inline_equation", + "content": "D" + }, + { + "bbox": [ + 62, + 148, + 292, + 159 + ], + "type": "text", + "content": ";" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 62, + 159, + 291, + 171 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 159, + 291, + 171 + ], + "spans": [ + { + "bbox": [ + 62, + 159, + 291, + 171 + ], + "type": "text", + "content": "4: Calculate individual perturbations " + }, + { + "bbox": [ + 62, + 159, + 291, + 171 + ], + "type": "inline_equation", + "content": "\\delta_{i}" + }, + { + "bbox": [ + 62, + 159, + 291, + 171 + ], + "type": "text", + "content": " by " + }, + { + "bbox": [ + 62, + 159, + 291, + 171 + ], + "type": "inline_equation", + "content": "D(E_{\\tau}(x))" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 62, + 171, + 294, + 195 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 171, + 294, + 195 + ], + "spans": [ + { + "bbox": [ + 62, + 171, + 294, + 195 + ], + "type": "text", + "content": "5: Create the unadversarial example " + }, + { + "bbox": [ + 62, + 171, + 294, + 195 + ], + "type": "inline_equation", + "content": "\\hat{x}_i" + }, + { + "bbox": [ + 62, + 171, + 294, + 195 + ], + "type": "text", + "content": " by incorporating " + }, + { + "bbox": [ + 62, + 171, + 294, + 195 + ], + "type": "inline_equation", + "content": "\\delta_i" + }, + { + "bbox": [ + 62, + 171, + 294, + 195 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 62, + 171, + 294, + 195 + ], + "type": "inline_equation", + "content": "x_i" + }, + { + "bbox": [ + 62, + 171, + 294, + 195 + ], + "type": "text", + "content": " through the bypass path." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 62, + 196, + 283, + 207 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 196, + 283, + 207 + ], + "spans": [ + { + "bbox": [ + 62, + 196, + 283, + 207 + ], + "type": "text", + "content": "6: Generate a fine-grained saliency map " + }, + { + "bbox": [ + 62, + 196, + 283, + 207 + ], + "type": "inline_equation", + "content": "g_{i}" + }, + { + "bbox": [ + 62, + 196, + 283, + 207 + ], + "type": "text", + "content": " of the " + }, + { + "bbox": [ + 62, + 196, + 283, + 207 + ], + "type": "inline_equation", + "content": "x_{i}" + }, + { + "bbox": [ + 62, + 196, + 283, + 207 + ], + "type": "text", + "content": ";" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 62, + 208, + 282, + 219 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 208, + 282, + 219 + ], + "spans": [ + { + "bbox": [ + 62, + 208, + 282, + 219 + ], + "type": "text", + "content": "7: Update " + }, + { + "bbox": [ + 62, + 208, + 282, + 219 + ], + "type": "inline_equation", + "content": "\\theta_v" + }, + { + "bbox": [ + 62, + 208, + 282, + 219 + ], + "type": "text", + "content": " with Eq.(8), taking " + }, + { + "bbox": [ + 62, + 208, + 282, + 219 + ], + "type": "inline_equation", + "content": "g_t" + }, + { + "bbox": [ + 62, + 208, + 282, + 219 + ], + "type": "text", + "content": " as a supervision." + } + ] + } + ], + "index": 9 + } + ], + "sub_type": "text" + }, + { + "bbox": [ + 62, + 220, + 268, + 243 + ], + "type": "list", + "angle": 0, + "index": 13, + "blocks": [ + { + "bbox": [ + 62, + 220, + 106, + 229 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 220, + 106, + 229 + ], + "spans": [ + { + "bbox": [ + 62, + 220, + 106, + 229 + ], + "type": "text", + "content": "8: end for" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 62, + 231, + 268, + 243 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 231, + 268, + 243 + ], + "spans": [ + { + "bbox": [ + 62, + 231, + 268, + 243 + ], + "type": "text", + "content": "9: return The generative unadversarial examples " + }, + { + "bbox": [ + 62, + 231, + 268, + 243 + ], + "type": "inline_equation", + "content": "\\hat{x}" + } + ] + } + ], + "index": 12 + } + ], + "sub_type": "text" + }, + { + "bbox": [ + 55, + 269, + 296, + 353 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 269, + 296, + 353 + ], + "spans": [ + { + "bbox": [ + 55, + 269, + 296, + 353 + ], + "type": "text", + "content": "encoder and reparameterization. The decoder, representing " + }, + { + "bbox": [ + 55, + 269, + 296, + 353 + ], + "type": "inline_equation", + "content": "F_{\\Phi}" + }, + { + "bbox": [ + 55, + 269, + 296, + 353 + ], + "type": "text", + "content": ", then transforms " + }, + { + "bbox": [ + 55, + 269, + 296, + 353 + ], + "type": "inline_equation", + "content": "\\frac{\\partial \\delta_0}{\\partial x}" + }, + { + "bbox": [ + 55, + 269, + 296, + 353 + ], + "type": "text", + "content": " into a generative perturbation " + }, + { + "bbox": [ + 55, + 269, + 296, + 353 + ], + "type": "inline_equation", + "content": "\\delta_t" + }, + { + "bbox": [ + 55, + 269, + 296, + 353 + ], + "type": "text", + "content": ". Finally, the by-pass structure incorporates " + }, + { + "bbox": [ + 55, + 269, + 296, + 353 + ], + "type": "inline_equation", + "content": "x_t" + }, + { + "bbox": [ + 55, + 269, + 296, + 353 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 55, + 269, + 296, + 353 + ], + "type": "inline_equation", + "content": "\\delta_t" + }, + { + "bbox": [ + 55, + 269, + 296, + 353 + ], + "type": "text", + "content": " to produce " + }, + { + "bbox": [ + 55, + 269, + 296, + 353 + ], + "type": "inline_equation", + "content": "\\hat{x}_t" + }, + { + "bbox": [ + 55, + 269, + 296, + 353 + ], + "type": "text", + "content": ". As shown in Fig. 2 (b), during the inference phase, the trained GUES model generates the corresponding generative unadversarial examples for the unseen frozen or fine-tuned model from specific testing samples." + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 55, + 355, + 296, + 428 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 355, + 296, + 428 + ], + "spans": [ + { + "bbox": [ + 55, + 355, + 296, + 428 + ], + "type": "text", + "content": "Loss function. The loss function for GUES training consists of two components. First, we enforce the latent space with mean " + }, + { + "bbox": [ + 55, + 355, + 296, + 428 + ], + "type": "inline_equation", + "content": "\\mu" + }, + { + "bbox": [ + 55, + 355, + 296, + 428 + ], + "type": "text", + "content": " and variance " + }, + { + "bbox": [ + 55, + 355, + 296, + 428 + ], + "type": "inline_equation", + "content": "\\sigma" + }, + { + "bbox": [ + 55, + 355, + 296, + 428 + ], + "type": "text", + "content": " satisfy the standard normal distribution " + }, + { + "bbox": [ + 55, + 355, + 296, + 428 + ], + "type": "inline_equation", + "content": "\\mathcal{N}(0,I)" + }, + { + "bbox": [ + 55, + 355, + 296, + 428 + ], + "type": "text", + "content": ". Suppose the encoder in VAE models the posterior distribution " + }, + { + "bbox": [ + 55, + 355, + 296, + 428 + ], + "type": "inline_equation", + "content": "q(z|x_{t}) = \\mathcal{N}(\\mu (x_{t}),\\sigma^{2}(x_{t}))" + }, + { + "bbox": [ + 55, + 355, + 296, + 428 + ], + "type": "text", + "content": ", this regularization can be formulated as:" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 102, + 438, + 296, + 452 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 102, + 438, + 296, + 452 + ], + "spans": [ + { + "bbox": [ + 102, + 438, + 296, + 452 + ], + "type": "interline_equation", + "content": "L _ {\\mathrm {K L}} = D _ {\\mathrm {K L}} \\left(q (z | x _ {t}) \\| \\mathcal {N} (0, I)\\right), \\tag {6}", + "image_path": "c7e4146f8078920ec62dcb1b4917d6d953d43a820ff372a71d40d8dcd6e60e24.jpg" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 55, + 462, + 296, + 510 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 462, + 296, + 510 + ], + "spans": [ + { + "bbox": [ + 55, + 462, + 296, + 510 + ], + "type": "text", + "content": "where function " + }, + { + "bbox": [ + 55, + 462, + 296, + 510 + ], + "type": "inline_equation", + "content": "D_{\\mathrm{KL}}" + }, + { + "bbox": [ + 55, + 462, + 296, + 510 + ], + "type": "text", + "content": " computes the Kullback-Leibler divergence. The other reconstruction loss between the unadversarial example " + }, + { + "bbox": [ + 55, + 462, + 296, + 510 + ], + "type": "inline_equation", + "content": "\\hat{x}_t" + }, + { + "bbox": [ + 55, + 462, + 296, + 510 + ], + "type": "text", + "content": " and saliency map " + }, + { + "bbox": [ + 55, + 462, + 296, + 510 + ], + "type": "inline_equation", + "content": "g_t" + }, + { + "bbox": [ + 55, + 462, + 296, + 510 + ], + "type": "text", + "content": " is presented by the following regression form:" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 127, + 521, + 295, + 533 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 127, + 521, + 295, + 533 + ], + "spans": [ + { + "bbox": [ + 127, + 521, + 295, + 533 + ], + "type": "interline_equation", + "content": "L _ {\\mathrm {M S E}} = \\left\\| \\hat {x} _ {t} - g _ {t} \\right\\| _ {2}. \\tag {7}", + "image_path": "5f2540da613a1c7b4feed10ceb0d16ea63be83fbd115ba404dab7e8275cce745.jpg" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 55, + 544, + 295, + 567 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 544, + 295, + 567 + ], + "spans": [ + { + "bbox": [ + 55, + 544, + 295, + 567 + ], + "type": "text", + "content": "Formally, combining Eq. (6) and Eq. (7), the final objective of GUES can be summarized as:" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 121, + 579, + 295, + 591 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 121, + 579, + 295, + 591 + ], + "spans": [ + { + "bbox": [ + 121, + 579, + 295, + 591 + ], + "type": "interline_equation", + "content": "L _ {\\text {G U E S}} = \\alpha L _ {\\mathrm {K L}} + \\beta L _ {\\mathrm {M S E}}, \\tag {8}", + "image_path": "3a23b0a20b36e9b81eda4d9b61a2dbbee1acf7da8b31fd81ace156862b5b8101.jpg" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 55, + 602, + 295, + 625 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 602, + 295, + 625 + ], + "spans": [ + { + "bbox": [ + 55, + 602, + 295, + 625 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 55, + 602, + 295, + 625 + ], + "type": "inline_equation", + "content": "\\alpha" + }, + { + "bbox": [ + 55, + 602, + 295, + 625 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 55, + 602, + 295, + 625 + ], + "type": "inline_equation", + "content": "\\beta" + }, + { + "bbox": [ + 55, + 602, + 295, + 625 + ], + "type": "text", + "content": " are trade-off parameters. For clarity, we summarize the training procedure of GUES in Algorithm 1." + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 55, + 626, + 296, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 626, + 296, + 714 + ], + "spans": [ + { + "bbox": [ + 55, + 626, + 296, + 714 + ], + "type": "text", + "content": "In Eq. (8), the first item " + }, + { + "bbox": [ + 55, + 626, + 296, + 714 + ], + "type": "inline_equation", + "content": "L_{\\mathrm{KL}}" + }, + { + "bbox": [ + 55, + 626, + 296, + 714 + ], + "type": "text", + "content": " ensures the learning of " + }, + { + "bbox": [ + 55, + 626, + 296, + 714 + ], + "type": "inline_equation", + "content": "\\frac{\\partial \\delta_0}{\\partial x}" + }, + { + "bbox": [ + 55, + 626, + 296, + 714 + ], + "type": "text", + "content": ". On the one hand, as aforementioned, we use " + }, + { + "bbox": [ + 55, + 626, + 296, + 714 + ], + "type": "inline_equation", + "content": "z" + }, + { + "bbox": [ + 55, + 626, + 296, + 714 + ], + "type": "text", + "content": " to present " + }, + { + "bbox": [ + 55, + 626, + 296, + 714 + ], + "type": "inline_equation", + "content": "\\frac{\\partial \\delta_0}{\\partial x}" + }, + { + "bbox": [ + 55, + 626, + 296, + 714 + ], + "type": "text", + "content": ". " + }, + { + "bbox": [ + 55, + 626, + 296, + 714 + ], + "type": "inline_equation", + "content": "L_{\\mathrm{KL}}" + }, + { + "bbox": [ + 55, + 626, + 296, + 714 + ], + "type": "text", + "content": " aligns the " + }, + { + "bbox": [ + 55, + 626, + 296, + 714 + ], + "type": "inline_equation", + "content": "z" + }, + { + "bbox": [ + 55, + 626, + 296, + 714 + ], + "type": "text", + "content": " space with " + }, + { + "bbox": [ + 55, + 626, + 296, + 714 + ], + "type": "inline_equation", + "content": "\\mathcal{N}(0, I)" + }, + { + "bbox": [ + 55, + 626, + 296, + 714 + ], + "type": "text", + "content": ", thereby linking the random noise to " + }, + { + "bbox": [ + 55, + 626, + 296, + 714 + ], + "type": "inline_equation", + "content": "\\frac{\\partial \\delta_0}{\\partial x}" + }, + { + "bbox": [ + 55, + 626, + 296, + 714 + ], + "type": "text", + "content": ". On the other hand, " + }, + { + "bbox": [ + 55, + 626, + 296, + 714 + ], + "type": "inline_equation", + "content": "q(z | x_t)" + }, + { + "bbox": [ + 55, + 626, + 296, + 714 + ], + "type": "text", + "content": " in " + }, + { + "bbox": [ + 55, + 626, + 296, + 714 + ], + "type": "inline_equation", + "content": "L_{\\mathrm{KL}}" + }, + { + "bbox": [ + 55, + 626, + 296, + 714 + ], + "type": "text", + "content": " is a function of input " + }, + { + "bbox": [ + 55, + 626, + 296, + 714 + ], + "type": "inline_equation", + "content": "x_t" + }, + { + "bbox": [ + 55, + 626, + 296, + 714 + ], + "type": "text", + "content": ", building relationship " + }, + { + "bbox": [ + 55, + 626, + 296, + 714 + ], + "type": "inline_equation", + "content": "x_t" + }, + { + "bbox": [ + 55, + 626, + 296, + 714 + ], + "type": "text", + "content": " to " + }, + { + "bbox": [ + 55, + 626, + 296, + 714 + ], + "type": "inline_equation", + "content": "\\frac{\\partial \\delta_0}{\\partial x}" + }, + { + "bbox": [ + 55, + 626, + 296, + 714 + ], + "type": "text", + "content": ". Additionally, the reconstruction regulated by the seconded item " + }, + { + "bbox": [ + 55, + 626, + 296, + 714 + ], + "type": "inline_equation", + "content": "L_{\\mathrm{MSE}}" + }, + { + "bbox": [ + 55, + 626, + 296, + 714 + ], + "type": "text", + "content": " encourage " + }, + { + "bbox": [ + 55, + 626, + 296, + 714 + ], + "type": "inline_equation", + "content": "\\delta_k \\rightarrow \\hat{\\delta}" + }, + { + "bbox": [ + 55, + 626, + 296, + 714 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 314, + 71, + 394, + 84 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 314, + 71, + 394, + 84 + ], + "spans": [ + { + "bbox": [ + 314, + 71, + 394, + 84 + ], + "type": "text", + "content": "4. Experiments" + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 314, + 91, + 376, + 102 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 314, + 91, + 376, + 102 + ], + "spans": [ + { + "bbox": [ + 314, + 91, + 376, + 102 + ], + "type": "text", + "content": "4.1. Datasets" + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 313, + 110, + 555, + 266 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 110, + 555, + 266 + ], + "spans": [ + { + "bbox": [ + 313, + 110, + 555, + 266 + ], + "type": "text", + "content": "We perform evaluation experiments on four existing fundus benchmarks, including APTOS [1], DDR [13], DeepDR [17], and Messidor-2 (termed MD2) [7]. Those datasets share five grading/classes: no DR, mild DR, moderate DR, severe DR, and proliferative DR. Taking each dataset as a separate domain, we form 12 transfer tasks crossing domains. For example, as APTOS is the source domain while the others are target domains, we have three transfer tasks APTOS " + }, + { + "bbox": [ + 313, + 110, + 555, + 266 + ], + "type": "inline_equation", + "content": "\\rightarrow" + }, + { + "bbox": [ + 313, + 110, + 555, + 266 + ], + "type": "text", + "content": " DDR, APTOS " + }, + { + "bbox": [ + 313, + 110, + 555, + 266 + ], + "type": "inline_equation", + "content": "\\rightarrow" + }, + { + "bbox": [ + 313, + 110, + 555, + 266 + ], + "type": "text", + "content": " DeepDR, and APTOS " + }, + { + "bbox": [ + 313, + 110, + 555, + 266 + ], + "type": "inline_equation", + "content": "\\rightarrow" + }, + { + "bbox": [ + 313, + 110, + 555, + 266 + ], + "type": "text", + "content": " MD2. The illustration of the label distribution and the domain shift of them is demonstrated in Supplementary. Of Note, all datasets have a severe class imbalance (e.g., \"no DR\" class itself takes up to " + }, + { + "bbox": [ + 313, + 110, + 555, + 266 + ], + "type": "inline_equation", + "content": "45.8\\%" + }, + { + "bbox": [ + 313, + 110, + 555, + 266 + ], + "type": "text", + "content": " of the DDR dataset)." + } + ] + } + ], + "index": 25 + }, + { + "bbox": [ + 314, + 274, + 443, + 287 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 314, + 274, + 443, + 287 + ], + "spans": [ + { + "bbox": [ + 314, + 274, + 443, + 287 + ], + "type": "text", + "content": "4.2. Implementation Detail" + } + ] + } + ], + "index": 26 + }, + { + "bbox": [ + 313, + 296, + 555, + 368 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 296, + 555, + 368 + ], + "spans": [ + { + "bbox": [ + 313, + 296, + 555, + 368 + ], + "type": "text", + "content": "Source model pre-training. We adopt the DeiT-base network [32] as the backbone of the source pre-trained model, training it in a supervised manner using the source data and corresponding ground truths. During this source training phase, the adopted objective is the classic cross-entropy loss with label smoothing, the same as other methods [15, 37, 38]." + } + ] + } + ], + "index": 27 + }, + { + "bbox": [ + 313, + 371, + 554, + 442 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 371, + 554, + 442 + ], + "spans": [ + { + "bbox": [ + 313, + 371, + 554, + 442 + ], + "type": "text", + "content": "Variational autoencoder setting. The VAE model is an eight-layer convolutional architecture with a latent space dimension of 10. We do not employ a pre-trained VAE and utilize a VAE without fine-tuning on any other dataset, ensuring that the learning component " + }, + { + "bbox": [ + 313, + 371, + 554, + 442 + ], + "type": "inline_equation", + "content": "F_{\\Phi}\\left(\\frac{\\partial\\delta_0}{\\partial x}\\right)" + }, + { + "bbox": [ + 313, + 371, + 554, + 442 + ], + "type": "text", + "content": " are unbiased and independent of prior pre-training data." + } + ] + } + ], + "index": 28 + }, + { + "bbox": [ + 313, + 443, + 555, + 479 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 443, + 555, + 479 + ], + "spans": [ + { + "bbox": [ + 313, + 443, + 555, + 479 + ], + "type": "text", + "content": "Parameter setting. For the trade-off parameters in Eq. (8), we set " + }, + { + "bbox": [ + 313, + 443, + 555, + 479 + ], + "type": "inline_equation", + "content": "\\alpha" + }, + { + "bbox": [ + 313, + 443, + 555, + 479 + ], + "type": "text", + "content": " to 1.0, while " + }, + { + "bbox": [ + 313, + 443, + 555, + 479 + ], + "type": "inline_equation", + "content": "\\beta" + }, + { + "bbox": [ + 313, + 443, + 555, + 479 + ], + "type": "text", + "content": " is tuned with " + }, + { + "bbox": [ + 313, + 443, + 555, + 479 + ], + "type": "inline_equation", + "content": "\\{0.0001, 0.01, 1\\}" + }, + { + "bbox": [ + 313, + 443, + 555, + 479 + ], + "type": "text", + "content": " to ensure that " + }, + { + "bbox": [ + 313, + 443, + 555, + 479 + ], + "type": "inline_equation", + "content": "L_{\\mathrm{KL}}" + }, + { + "bbox": [ + 313, + 443, + 555, + 479 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 313, + 443, + 555, + 479 + ], + "type": "inline_equation", + "content": "L_{\\mathrm{MSE}}" + }, + { + "bbox": [ + 313, + 443, + 555, + 479 + ], + "type": "text", + "content": " remain on the same values scale." + } + ] + } + ], + "index": 29 + }, + { + "bbox": [ + 313, + 479, + 554, + 539 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 479, + 554, + 539 + ], + "spans": [ + { + "bbox": [ + 313, + 479, + 554, + 539 + ], + "type": "text", + "content": "Training setting. For fairness, all compared methods were implemented with the same DeiT-base backbone. We adopt the batch size of 64, SGD optimizer with a momentum of 0.9 and a learning rate of 1e-5 on all datasets. All experiments are conducted with PyTorch on a single GPU of RTX A6000." + } + ] + } + ], + "index": 30 + }, + { + "bbox": [ + 314, + 548, + 435, + 560 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 314, + 548, + 435, + 560 + ], + "spans": [ + { + "bbox": [ + 314, + 548, + 435, + 560 + ], + "type": "text", + "content": "4.3. Comparison Settings" + } + ] + } + ], + "index": 31 + }, + { + "bbox": [ + 313, + 567, + 555, + 627 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 567, + 555, + 627 + ], + "spans": [ + { + "bbox": [ + 313, + 567, + 555, + 627 + ], + "type": "text", + "content": "Evaluation metrics. To account for unbalanced datasets, besides conventional classification accuracy (termed ACC), we adopt the measure of Quadratic Weighted Kappa (termed QWK) [2] and the average of QWK and ACC (termed AVG). Their rules are provided in Supplementary." + } + ] + } + ], + "index": 32 + }, + { + "bbox": [ + 313, + 629, + 555, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 629, + 555, + 713 + ], + "spans": [ + { + "bbox": [ + 313, + 629, + 555, + 713 + ], + "type": "text", + "content": "Competitors. We compare GUES with nine existing state-of-the-art adaptation methods divided into three groups. (1) The first group involves applying the source model directly to the target domain. (2) The second group includes five SFDA methods SHOT [15], NRC [38], CoWA [12], PLUE [16], and TPDS [29]. (3) The third group comprises three typical TTA methods: SHOT-IM [15], TENT [34], and SAR [21]." + } + ] + } + ], + "index": 33 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "text", + "content": "28341" + } + ] + } + ], + "index": 34 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 4 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 69, + 102, + 541, + 369 + ], + "blocks": [ + { + "bbox": [ + 55, + 70, + 555, + 94 + ], + "lines": [ + { + "bbox": [ + 55, + 70, + 555, + 94 + ], + "spans": [ + { + "bbox": [ + 55, + 70, + 555, + 94 + ], + "type": "text", + "content": "Table 2. The results of Source, SFDA, TTA, OMG-DA, and OMG-DA combination methods on datasets APTOS, DeepDR, DDR, and MD2 are presented. The improvements over baseline methods Source, SHOT-IM, and TENT are highlighted as (+x.x)." + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 69, + 102, + 541, + 369 + ], + "lines": [ + { + "bbox": [ + 69, + 102, + 541, + 369 + ], + "spans": [ + { + "bbox": [ + 69, + 102, + 541, + 369 + ], + "type": "table", + "html": "
MethodVenueAPTOS→DDRAPTOS→DeepDRAPTOS→MD2DDR→APTOSDDR→DeepDRDDR→MD2DeepDR→APTOS
ACCQWKAVGACCQWKAVGACCQWKAVGACCQWKAVGACCQWKAVGACCQWKAVGACCQWKAVG
Source-60.659.259.952.671.762.160.948.754.865.672.969.345.060.652.849.434.642.043.471.657.5
GUES + Source-62.059.560.853.069.761.359.846.753.376.081.878.956.468.762.559.147.653.346.574.160.3
SHOT [15]ICML2066.969.067.953.673.563.651.738.044.877.084.280.659.274.666.957.143.150.162.382.572.4
NRC [38]NeurIPS2161.965.263.551.670.961.354.140.347.260.376.368.352.069.160.550.635.242.952.074.363.2
CoWA [12]ICML2259.064.962.051.070.760.853.037.345.157.274.565.850.166.858.453.139.646.350.473.061.7
PLUE [16]CVPR2362.065.163.651.369.760.554.541.147.863.464.263.854.364.859.551.628.239.954.670.562.6
TPDS [29]IICV2466.667.867.251.671.561.552.740.746.776.683.980.258.073.165.654.541.047.860.880.070.4
SHOT-IM [15]ICML2066.569.267.952.673.663.153.236.644.975.982.179.058.573.966.257.343.550.461.984.072.9
TENT [34]ICLR2059.950.255.153.170.161.661.448.454.975.282.478.855.168.461.760.850.555.660.279.769.9
SAR [21]ICLR2367.963.865.853.673.063.357.244.951.175.683.579.555.671.663.649.237.043.159.579.369.4
GUES + SHOT-IM - GUES + TENT -68.668.568.553.572.863.255.743.849.777.283.180.260.575.167.861.551.256.362.683.272.9
-61.856.359.053.270.061.661.147.254.175.983.079.458.770.864.763.353.858.654.977.666.3
MethodVenueDeepDR→DDRDeepDR→MD2MD2→APTOSMD2→DDRMD2→DeepDRAccAvg. QWKAVG
ACCQWKAVGACCQWKAVGACCQWKAVGACCQWKAVGACCQWKAVG
Source-56.466.961.748.750.249.443.970.357.160.256.558.359.858.459.153.960.157.0
GUES + Source-57.365.661.448.352.050.159.376.267.764.755.860.258.657.157.858.4 (+4.5)62.9 (+2.8)60.7 (+3.7)
SHOT [15]ICML2057.471.264.348.841.545.252.773.062.854.659.256.959.670.264.958.465.061.7
NRC [38]NeurIPS2144.960.952.949.841.845.848.869.158.952.852.852.858.062.860.453.159.956.5
CoWA [12]ICML2248.758.453.649.942.746.351.070.160.549.650.950.357.660.659.156.961.959.4
PLUE [16]CVPR2347.253.550.456.447.451.956.069.162.656.554.355.458.864.861.855.557.756.6
TPDS [29]IICV2459.369.464.350.542.446.460.374.967.660.060.460.258.963.060.959.264.061.6
SHOT-IM [15]ICML2054.669.462.051.238.244.761.677.969.757.058.757.957.569.863.759.064.761.9
TENT [34]ICLR2058.545.451.958.356.557.455.174.164.655.831.743.758.053.655.859.359.259.3
SAR [21]ICLR2353.066.359.642.633.137.955.273.064.149.748.349.056.765.861.356.361.659.0
GUES + SHOT-IM - GUES + TENT --62.371.566.952.847.850.362.878.170.466.059.462.760.668.664.662.0 (+3.0)66.9 (+2.2)64.5 (+2.6)
-62.661.462.059.157.458.259.375.667.564.050.557.258.356.057.161.0 (+1.7)63.3 (+4.1)62.2 (+2.9)
", + "image_path": "4c12d1275a6d2cf5a7d683991f743a407c371d398fa7c7e8cb48a095c07bb756.jpg" + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "table_body" + } + ], + "index": 1 + }, + { + "type": "image", + "bbox": [ + 60, + 387, + 291, + 462 + ], + "blocks": [ + { + "bbox": [ + 60, + 387, + 291, + 462 + ], + "lines": [ + { + "bbox": [ + 60, + 387, + 291, + 462 + ], + "spans": [ + { + "bbox": [ + 60, + 387, + 291, + 462 + ], + "type": "image", + "image_path": "63765010acc15e3c61df0b9bf3ce8559686a07f425f19d2b0dc2843245c0ab1e.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 55, + 468, + 297, + 504 + ], + "lines": [ + { + "bbox": [ + 55, + 468, + 297, + 504 + ], + "spans": [ + { + "bbox": [ + 55, + 468, + 297, + 504 + ], + "type": "text", + "content": "Figure 4. Comparison results with batch size varying from 2 to 64 over the 12 tasks (The details are provided in Supplementary.). Left, middle, and right report ACC, QWK, and AVG, respectively." + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_caption" + } + ], + "index": 2 + }, + { + "bbox": [ + 55, + 527, + 297, + 563 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 527, + 297, + 563 + ], + "spans": [ + { + "bbox": [ + 55, + 527, + 297, + 563 + ], + "type": "text", + "content": "Comparison protocol in OMG-DA setting. For a comprehensive comparison, our comparison follows two different fashions as follows." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 55, + 565, + 295, + 637 + ], + "type": "list", + "angle": 0, + "index": 7, + "blocks": [ + { + "bbox": [ + 55, + 565, + 295, + 601 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 565, + 295, + 601 + ], + "spans": [ + { + "bbox": [ + 55, + 565, + 295, + 601 + ], + "type": "text", + "content": "- Case without training: We first generate the unadversarial examples for the target domain by the trained GUES model and then provide them to the frozen source model." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 55, + 601, + 295, + 637 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 601, + 295, + 637 + ], + "spans": [ + { + "bbox": [ + 55, + 601, + 295, + 637 + ], + "type": "text", + "content": "- Case with training: We plug GUES into other TTA methods (they are also online methods with flowing data) as online image pre-processing." + } + ] + } + ], + "index": 6 + } + ], + "sub_type": "text" + }, + { + "bbox": [ + 55, + 639, + 295, + 687 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 639, + 295, + 687 + ], + "spans": [ + { + "bbox": [ + 55, + 639, + 295, + 687 + ], + "type": "text", + "content": "The two cases evaluate GUES from different aspects. The first isolates the generalization ability of the unadversarial examples generated by GUES, whilst the second highlights GUES's compatibility with other trainable online schemes." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 55, + 689, + 297, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 689, + 297, + 713 + ], + "spans": [ + { + "bbox": [ + 55, + 689, + 297, + 713 + ], + "type": "text", + "content": "Corresponding to the comparison protocols above, besides the version GUES corresponding to the case with" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 313, + 388, + 555, + 413 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 388, + 555, + 413 + ], + "spans": [ + { + "bbox": [ + 313, + 388, + 555, + 413 + ], + "type": "text", + "content": "out training, we also introduce GUES+SHOT-IM and GUES+TENT which correspond to the case with training." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 313, + 422, + 432, + 436 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 422, + 432, + 436 + ], + "spans": [ + { + "bbox": [ + 313, + 422, + 432, + 436 + ], + "type": "text", + "content": "4.4. Comparison Results" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 313, + 441, + 555, + 478 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 441, + 555, + 478 + ], + "spans": [ + { + "bbox": [ + 313, + 441, + 555, + 478 + ], + "type": "text", + "content": "This part presents the comparison results following the cases above. Also, considering batch size a crucial factor for TTA methods, the results as the batch size varies are provided." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 313, + 481, + 554, + 565 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 481, + 554, + 565 + ], + "spans": [ + { + "bbox": [ + 313, + 481, + 554, + 565 + ], + "type": "text", + "content": "Results without training. The comparisons are shown in Tab. 2. On average, across the 12 tasks and without training of the source model, GUES achieves improvements of " + }, + { + "bbox": [ + 313, + 481, + 554, + 565 + ], + "type": "inline_equation", + "content": "4.5\\%" + }, + { + "bbox": [ + 313, + 481, + 554, + 565 + ], + "type": "text", + "content": " in ACC, " + }, + { + "bbox": [ + 313, + 481, + 554, + 565 + ], + "type": "inline_equation", + "content": "2.8\\%" + }, + { + "bbox": [ + 313, + 481, + 554, + 565 + ], + "type": "text", + "content": " in QWK, and " + }, + { + "bbox": [ + 313, + 481, + 554, + 565 + ], + "type": "inline_equation", + "content": "3.7\\%" + }, + { + "bbox": [ + 313, + 481, + 554, + 565 + ], + "type": "text", + "content": " in AVG compared to the source model. These results demonstrate that GUES modifies the target data distribution effectively, adapting the target domain to align with the source domain." + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 313, + 568, + 555, + 676 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 568, + 555, + 676 + ], + "spans": [ + { + "bbox": [ + 313, + 568, + 555, + 676 + ], + "type": "text", + "content": "Results with training. As shown in Tab. 2, GUES+SHOT-IM outperforms the previous best SFDA and TTA methods, respectively surpassing TENT in ACC by " + }, + { + "bbox": [ + 313, + 568, + 555, + 676 + ], + "type": "inline_equation", + "content": "2.7\\%" + }, + { + "bbox": [ + 313, + 568, + 555, + 676 + ], + "type": "text", + "content": ", SHOT in QWK by " + }, + { + "bbox": [ + 313, + 568, + 555, + 676 + ], + "type": "inline_equation", + "content": "1.9\\%" + }, + { + "bbox": [ + 313, + 568, + 555, + 676 + ], + "type": "text", + "content": ", and SHOT-IM in AVG by " + }, + { + "bbox": [ + 313, + 568, + 555, + 676 + ], + "type": "inline_equation", + "content": "2.6\\%" + }, + { + "bbox": [ + 313, + 568, + 555, + 676 + ], + "type": "text", + "content": ". Compared to SHOT-IM, GUES+SHOT-IM gains over " + }, + { + "bbox": [ + 313, + 568, + 555, + 676 + ], + "type": "inline_equation", + "content": "3.0\\%" + }, + { + "bbox": [ + 313, + 568, + 555, + 676 + ], + "type": "text", + "content": " in ACC, " + }, + { + "bbox": [ + 313, + 568, + 555, + 676 + ], + "type": "inline_equation", + "content": "2.2\\%" + }, + { + "bbox": [ + 313, + 568, + 555, + 676 + ], + "type": "text", + "content": " in QWK, and " + }, + { + "bbox": [ + 313, + 568, + 555, + 676 + ], + "type": "inline_equation", + "content": "2.6\\%" + }, + { + "bbox": [ + 313, + 568, + 555, + 676 + ], + "type": "text", + "content": " in AVG. Similarly, GUES+TENT improves over TENT by " + }, + { + "bbox": [ + 313, + 568, + 555, + 676 + ], + "type": "inline_equation", + "content": "1.7\\%" + }, + { + "bbox": [ + 313, + 568, + 555, + 676 + ], + "type": "text", + "content": " in ACC, " + }, + { + "bbox": [ + 313, + 568, + 555, + 676 + ], + "type": "inline_equation", + "content": "4.1\\%" + }, + { + "bbox": [ + 313, + 568, + 555, + 676 + ], + "type": "text", + "content": " in QWK, and " + }, + { + "bbox": [ + 313, + 568, + 555, + 676 + ], + "type": "inline_equation", + "content": "2.9\\%" + }, + { + "bbox": [ + 313, + 568, + 555, + 676 + ], + "type": "text", + "content": " in AVG. These results highlight the effectiveness of combining GUES with other methods that require training." + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 313, + 677, + 556, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 677, + 556, + 715 + ], + "spans": [ + { + "bbox": [ + 313, + 677, + 556, + 715 + ], + "type": "text", + "content": "Results with varying batch size. This part isolates the effect of batch size, which is a crucial factor for TTA methods. Fig. 4 depicts the performance variation as batch size vary" + } + ] + } + ], + "index": 15 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 318, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 318, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 318, + 757 + ], + "type": "text", + "content": "28342" + } + ] + } + ], + "index": 16 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 5 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 82, + 70, + 531, + 157 + ], + "blocks": [ + { + "bbox": [ + 82, + 70, + 531, + 157 + ], + "lines": [ + { + "bbox": [ + 82, + 70, + 531, + 157 + ], + "spans": [ + { + "bbox": [ + 82, + 70, + 531, + 157 + ], + "type": "image", + "image_path": "373dc7e04a1dccbc02cde73710d605ee9811e1e2bb7be310782249f961c2d3d8.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 55, + 164, + 555, + 199 + ], + "lines": [ + { + "bbox": [ + 55, + 164, + 555, + 199 + ], + "spans": [ + { + "bbox": [ + 55, + 164, + 555, + 199 + ], + "type": "text", + "content": "Figure 5. Interpretability analysis based on a typical fundus image from \"Moderate DR\" class in APTOS. Here, H (hemorrhages), SE (soft exudates), and EX (hard exudates) are essential characteristics to judge the DR grade. The gradient CAM-based heatmap of five models visualizes the capture of those lesions. All models are trained on task DDR " + }, + { + "bbox": [ + 55, + 164, + 555, + 199 + ], + "type": "inline_equation", + "content": "\\rightarrow" + }, + { + "bbox": [ + 55, + 164, + 555, + 199 + ], + "type": "text", + "content": " APTOS, where Oracle is trained using ground truth in APTOS." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "type": "image", + "bbox": [ + 62, + 217, + 289, + 293 + ], + "blocks": [ + { + "bbox": [ + 62, + 217, + 289, + 293 + ], + "lines": [ + { + "bbox": [ + 62, + 217, + 289, + 293 + ], + "spans": [ + { + "bbox": [ + 62, + 217, + 289, + 293 + ], + "type": "image", + "image_path": "7dc39ff85d07b737b9df1a1ce46a1b53f0d0d32651c72c934caa98923545f3cf.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 55, + 298, + 295, + 321 + ], + "lines": [ + { + "bbox": [ + 55, + 298, + 295, + 321 + ], + "spans": [ + { + "bbox": [ + 55, + 298, + 295, + 321 + ], + "type": "text", + "content": "Figure 6. Feature distribution comparison of 3D density charts on task DDR " + }, + { + "bbox": [ + 55, + 298, + 295, + 321 + ], + "type": "inline_equation", + "content": "\\rightarrow" + }, + { + "bbox": [ + 55, + 298, + 295, + 321 + ], + "type": "text", + "content": " APTOS. Oracle is trained on APTOS by ground truth." + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_caption" + } + ], + "index": 2 + }, + { + "bbox": [ + 54, + 341, + 295, + 496 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 54, + 341, + 295, + 496 + ], + "spans": [ + { + "bbox": [ + 54, + 341, + 295, + 496 + ], + "type": "text", + "content": "ing from 2 to 64 over the 12 tasks. It is observed that TTA methods SHOT-IM and TENT suffer from severe performance drops when the batch size becomes small. SHOT-IM exhibits a decrease of approximately " + }, + { + "bbox": [ + 54, + 341, + 295, + 496 + ], + "type": "inline_equation", + "content": "16\\%" + }, + { + "bbox": [ + 54, + 341, + 295, + 496 + ], + "type": "text", + "content": " in ACC when the batch size is reduced from 64 to 2, whilst TENT shows a substantial decline of around " + }, + { + "bbox": [ + 54, + 341, + 295, + 496 + ], + "type": "inline_equation", + "content": "34\\%" + }, + { + "bbox": [ + 54, + 341, + 295, + 496 + ], + "type": "text", + "content": " in QWK. Oppose to it, the methods with GUES, SHOT-IM+GUES, and TENT+GUES, do not have evident performance decline at the smaller batch size. Moreover, this combination not only mitigates the drop but also shows improvements when the batch size is 64. This indicates that GUES effectively stabilizes the performance of SHOT-IM and TENT, enhancing their robustness to variations in batch size while boosting their overall effectiveness." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 55, + 497, + 295, + 604 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 497, + 295, + 604 + ], + "spans": [ + { + "bbox": [ + 55, + 497, + 295, + 604 + ], + "type": "text", + "content": "We attribute GUES's excellent robustness to its ability to predict individual perturbations (see Supplementary for more details on the visualization of perturbations) that focus on single image-specific features rather than global data commonalities. For instance, the conventional unadversarial examples approach refines the class-specific perturbation sensitive to batch size. Furthermore, both SHOT-IM and TENT are entropy-based methods that require large-scale batch size for accurate entropy estimation." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 55, + 611, + 182, + 624 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 611, + 182, + 624 + ], + "spans": [ + { + "bbox": [ + 55, + 611, + 182, + 624 + ], + "type": "text", + "content": "4.5. Visualization Analysis" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 55, + 629, + 297, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 629, + 297, + 713 + ], + "spans": [ + { + "bbox": [ + 55, + 629, + 297, + 713 + ], + "type": "text", + "content": "Interpretability. For a better understanding, Fig. 5 demonstrates whether GUES can help capture pathologically relevant features, such as H (hemorrhages), SE (soft exudates), and EX (hard exudates), determining DR grade. First of all, when comparing the source model with GUES, the source model only captures a limited area of the lesion, while GUES effectively captures most of the DR-related features. Further-" + } + ] + } + ], + "index": 7 + }, + { + "type": "image", + "bbox": [ + 317, + 218, + 553, + 322 + ], + "blocks": [ + { + "bbox": [ + 317, + 218, + 553, + 322 + ], + "lines": [ + { + "bbox": [ + 317, + 218, + 553, + 322 + ], + "spans": [ + { + "bbox": [ + 317, + 218, + 553, + 322 + ], + "type": "image", + "image_path": "860379c907f692a288f195ca59110b8bb3e73479d954ca982ac69eb7b89547a2.jpg" + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 313, + 329, + 555, + 362 + ], + "lines": [ + { + "bbox": [ + 313, + 329, + 555, + 362 + ], + "spans": [ + { + "bbox": [ + 313, + 329, + 555, + 362 + ], + "type": "text", + "content": "Figure 7. Unadversarial examples visualization of two typical target samples from APTOS. The generative perturbations are generated by the GUES model trained on task DDR " + }, + { + "bbox": [ + 313, + 329, + 555, + 362 + ], + "type": "inline_equation", + "content": "\\rightarrow" + }, + { + "bbox": [ + 313, + 329, + 555, + 362 + ], + "type": "text", + "content": " APTOS." + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "image_caption" + } + ], + "index": 8 + }, + { + "bbox": [ + 313, + 384, + 555, + 456 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 384, + 555, + 456 + ], + "spans": [ + { + "bbox": [ + 313, + 384, + 555, + 456 + ], + "type": "text", + "content": "more, combining GUES with SHOT-IM (i.e., GUES+SHOT-IM) expands the focus on DR-related features beyond those captured by SHOT-IM alone. Additionally, when comparing the four models to Oracle, only GUES and GUES+SHOT-IM resemble Oracle, suggesting that GUES effectively directs the model's attention to DR-critical features." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 313, + 459, + 556, + 615 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 459, + 556, + 615 + ], + "spans": [ + { + "bbox": [ + 313, + 459, + 556, + 615 + ], + "type": "text", + "content": "Feature distribution. Taking the task DDR→APTOS as a toy experiment, we visualize the feature distribution extracted from the final convolutional layer of the prediction model using a 3D density chart. Considering that APTOS is a class-imbalanced dataset, with the “No DR” class alone comprising up to " + }, + { + "bbox": [ + 313, + 459, + 556, + 615 + ], + "type": "inline_equation", + "content": "49.2\\%" + }, + { + "bbox": [ + 313, + 459, + 556, + 615 + ], + "type": "text", + "content": " of the dataset, our analysis focuses on this crucial property. As shown in Fig. 6, the feature distribution of the source model does not reflect this imbalanced characteristic; instead, it displays a more uniform classification. Conversely, the feature distribution of GUES exhibits a distinct imbalance, with one expanded high-density region alongside several smaller high-density regions, resembling the distribution pattern seen in Oracle." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 313, + 617, + 557, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 617, + 557, + 714 + ], + "spans": [ + { + "bbox": [ + 313, + 617, + 557, + 714 + ], + "type": "text", + "content": "Visualization of unadversarial examples. This part visualizes unadversarial examples of two typical target samples from APTOS and corresponding generative perturbations, based on the task DDR " + }, + { + "bbox": [ + 313, + 617, + 557, + 714 + ], + "type": "inline_equation", + "content": "\\rightarrow" + }, + { + "bbox": [ + 313, + 617, + 557, + 714 + ], + "type": "text", + "content": " APTOS. Considering the generative perturbations alter the original images that may not be easily visible to the naked eye, we collect RGB statistics to illustrate these changes quantitatively. It is observed that in Fig. 7, each channel (R, G, and B) exhibits notable" + } + ] + } + ], + "index": 12 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "type": "text", + "content": "28343" + } + ] + } + ], + "index": 13 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 6 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 61, + 91, + 288, + 177 + ], + "blocks": [ + { + "bbox": [ + 96, + 71, + 254, + 82 + ], + "lines": [ + { + "bbox": [ + 96, + 71, + 254, + 82 + ], + "spans": [ + { + "bbox": [ + 96, + 71, + 254, + 82 + ], + "type": "text", + "content": "Table 3. ACC results of ablation study " + }, + { + "bbox": [ + 96, + 71, + 254, + 82 + ], + "type": "inline_equation", + "content": "(\\%)" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 61, + 91, + 288, + 177 + ], + "lines": [ + { + "bbox": [ + 61, + 91, + 288, + 177 + ], + "spans": [ + { + "bbox": [ + 61, + 91, + 288, + 177 + ], + "type": "table", + "html": "
#\\( L_{\\text{KL}} \\)\\( L_{\\text{MSE}} \\)APTOSDDRDeepDRMD2Avg.
1XX51.059.152.553.053.9
2✓X54.758.354.953.355.3
3X✓56.960.152.952.955.7
4✓✓60.661.356.055.758.4
5GUES w/ AE53.860.253.853.255.3
6GUES w/ Mixup56.860.253.052.855.7
7GUES w/ Self56.560.453.754.456.3
8GUES w/ Sal46.134.341.845.942.0
", + "image_path": "abbc41cdc5923668482ea79c8f23b3797adb03c20e06c083b383e1746de46547.jpg" + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "table_body" + } + ], + "index": 1 + }, + { + "bbox": [ + 55, + 201, + 295, + 285 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 201, + 295, + 285 + ], + "spans": [ + { + "bbox": [ + 55, + 201, + 295, + 285 + ], + "type": "text", + "content": "fluctuations, with the RGB statistics of the original images (a) differing significantly from those of the unadversarial examples (c). Additionally, each generative perturbation is unique, meaning that the alterations introduced by these perturbations are individual. These results suggest that the perturbations may help highlight critical DR-related features, refining the model's focus on diagnostically relevant areas." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 55, + 295, + 157, + 307 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 295, + 157, + 307 + ], + "spans": [ + { + "bbox": [ + 55, + 295, + 157, + 307 + ], + "type": "text", + "content": "4.6. Further Analysis" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 55, + 314, + 296, + 445 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 314, + 296, + 445 + ], + "spans": [ + { + "bbox": [ + 55, + 314, + 296, + 445 + ], + "type": "text", + "content": "Ablation study. In this part, we evaluate the effect of objective loss, as well as the components involved in GUES including the sampling strategy and saliency map-based supervision. To address the first issue, we conduct a progressive experiment. The top four rows of Tab. 3 list the ablation results where the source model's performance is the baseline. Using " + }, + { + "bbox": [ + 55, + 314, + 296, + 445 + ], + "type": "inline_equation", + "content": "L_{\\mathrm{KL}}" + }, + { + "bbox": [ + 55, + 314, + 296, + 445 + ], + "type": "text", + "content": " or " + }, + { + "bbox": [ + 55, + 314, + 296, + 445 + ], + "type": "inline_equation", + "content": "L_{\\mathrm{MSE}}" + }, + { + "bbox": [ + 55, + 314, + 296, + 445 + ], + "type": "text", + "content": " alone yields an average ACC improvement of approximately " + }, + { + "bbox": [ + 55, + 314, + 296, + 445 + ], + "type": "inline_equation", + "content": "1.4\\%" + }, + { + "bbox": [ + 55, + 314, + 296, + 445 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 55, + 314, + 296, + 445 + ], + "type": "inline_equation", + "content": "1.8\\%" + }, + { + "bbox": [ + 55, + 314, + 296, + 445 + ], + "type": "text", + "content": ", respectively, over the baseline. As both of them work, the ACC increases " + }, + { + "bbox": [ + 55, + 314, + 296, + 445 + ], + "type": "inline_equation", + "content": "3.5\\%" + }, + { + "bbox": [ + 55, + 314, + 296, + 445 + ], + "type": "text", + "content": " on average further. The results indicate that all objective components positively affect the final performance." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 55, + 447, + 296, + 626 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 447, + 296, + 626 + ], + "spans": [ + { + "bbox": [ + 55, + 447, + 296, + 626 + ], + "type": "text", + "content": "To evaluate the impact of the sampling component, we propose a variation method of GUES, GUES w/ AE, where we remove this sampling process by replacing VAE with a conventional Auto-encoder model. In addition, two GUES variations are used to assess the advantage of saliency map-based supervision. Specifically, GUES w/ Mixup replaces the saliency maps with a Mixup of saliency maps and original images, whilst GUES w/ Self replaces the saliency maps with the original images. As presented in the Tab. 3, compared with the full version of GUES (the fourth row), GUES w/ AE, GUES w/ Mixup and GUES w/ Self decrease by " + }, + { + "bbox": [ + 55, + 447, + 296, + 626 + ], + "type": "inline_equation", + "content": "3.1\\%" + }, + { + "bbox": [ + 55, + 447, + 296, + 626 + ], + "type": "text", + "content": " at least on average. Besides, replacing the original images with saliency maps as inputs (GUES w/ Sal) leads to a significant drop of about " + }, + { + "bbox": [ + 55, + 447, + 296, + 626 + ], + "type": "inline_equation", + "content": "15.4\\%" + }, + { + "bbox": [ + 55, + 447, + 296, + 626 + ], + "type": "text", + "content": ". Above these experiments confirm the effectiveness of our design choices." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 55, + 629, + 296, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 629, + 296, + 713 + ], + "spans": [ + { + "bbox": [ + 55, + 629, + 296, + 713 + ], + "type": "text", + "content": "Parameter sensitiveness. Taking the task DDR " + }, + { + "bbox": [ + 55, + 629, + 296, + 713 + ], + "type": "inline_equation", + "content": "\\rightarrow" + }, + { + "bbox": [ + 55, + 629, + 296, + 713 + ], + "type": "text", + "content": " APTOS as a toy experiment, we present the GUES performance varying as hyper-parameters " + }, + { + "bbox": [ + 55, + 629, + 296, + 713 + ], + "type": "inline_equation", + "content": "0.5\\leq \\alpha \\leq 1.5" + }, + { + "bbox": [ + 55, + 629, + 296, + 713 + ], + "type": "text", + "content": " with 0.1 steps, " + }, + { + "bbox": [ + 55, + 629, + 296, + 713 + ], + "type": "inline_equation", + "content": "0.00005\\leq \\beta \\leq 0.00014" + }, + { + "bbox": [ + 55, + 629, + 296, + 713 + ], + "type": "text", + "content": " with 0.00001 steps. As depicted in Fig. 8, the ACC, QWK, and AVG variation surfaces show fluctuations in the tiny performance zone, with approximately " + }, + { + "bbox": [ + 55, + 629, + 296, + 713 + ], + "type": "inline_equation", + "content": "0.2\\%" + }, + { + "bbox": [ + 55, + 629, + 296, + 713 + ], + "type": "text", + "content": " in ACC, " + }, + { + "bbox": [ + 55, + 629, + 296, + 713 + ], + "type": "inline_equation", + "content": "0.25\\%" + }, + { + "bbox": [ + 55, + 629, + 296, + 713 + ], + "type": "text", + "content": " in QWK, and " + }, + { + "bbox": [ + 55, + 629, + 296, + 713 + ], + "type": "inline_equation", + "content": "0.4\\%" + }, + { + "bbox": [ + 55, + 629, + 296, + 713 + ], + "type": "text", + "content": " in" + } + ] + } + ], + "index": 6 + }, + { + "type": "image", + "bbox": [ + 322, + 71, + 395, + 143 + ], + "blocks": [ + { + "bbox": [ + 322, + 71, + 395, + 143 + ], + "lines": [ + { + "bbox": [ + 322, + 71, + 395, + 143 + ], + "spans": [ + { + "bbox": [ + 322, + 71, + 395, + 143 + ], + "type": "image", + "image_path": "9cdc9e8c29501aa175b39b55b2bda1428b83dc0e214cde588a14fcebc3a5d085.jpg" + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 313, + 152, + 555, + 185 + ], + "lines": [ + { + "bbox": [ + 313, + 152, + 555, + 185 + ], + "spans": [ + { + "bbox": [ + 313, + 152, + 555, + 185 + ], + "type": "text", + "content": "Figure 8. Parameter sensitiveness study results over " + }, + { + "bbox": [ + 313, + 152, + 555, + 185 + ], + "type": "inline_equation", + "content": "\\alpha \\times \\beta" + }, + { + "bbox": [ + 313, + 152, + 555, + 185 + ], + "type": "text", + "content": " based on task DDR " + }, + { + "bbox": [ + 313, + 152, + 555, + 185 + ], + "type": "inline_equation", + "content": "\\rightarrow" + }, + { + "bbox": [ + 313, + 152, + 555, + 185 + ], + "type": "text", + "content": " APTOS. From left to right, there are results on ACC, QWK, and AVG, respectively." + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "image_caption" + } + ], + "index": 7 + }, + { + "type": "image", + "bbox": [ + 399, + 71, + 471, + 143 + ], + "blocks": [ + { + "bbox": [ + 399, + 71, + 471, + 143 + ], + "lines": [ + { + "bbox": [ + 399, + 71, + 471, + 143 + ], + "spans": [ + { + "bbox": [ + 399, + 71, + 471, + 143 + ], + "type": "image", + "image_path": "3c06c4119aeb4010430e016ae1e367f133b843ee2a703215d4cbaef1c5ca462e.jpg" + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "image_body" + } + ], + "index": 8 + }, + { + "type": "image", + "bbox": [ + 474, + 71, + 547, + 143 + ], + "blocks": [ + { + "bbox": [ + 474, + 71, + 547, + 143 + ], + "lines": [ + { + "bbox": [ + 474, + 71, + 547, + 143 + ], + "spans": [ + { + "bbox": [ + 474, + 71, + 547, + 143 + ], + "type": "image", + "image_path": "5b58cf7f34b38f89b60754474f73c44da4741fc11bfe7135a5098f474b99d0f1.jpg" + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "image_body" + } + ], + "index": 9 + }, + { + "bbox": [ + 313, + 205, + 553, + 229 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 205, + 553, + 229 + ], + "spans": [ + { + "bbox": [ + 313, + 205, + 553, + 229 + ], + "type": "text", + "content": "AVG. This observation suggests that GUES is insensitive to alterations in " + }, + { + "bbox": [ + 313, + 205, + 553, + 229 + ], + "type": "inline_equation", + "content": "\\alpha" + }, + { + "bbox": [ + 313, + 205, + 553, + 229 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 313, + 205, + 553, + 229 + ], + "type": "inline_equation", + "content": "\\beta" + }, + { + "bbox": [ + 313, + 205, + 553, + 229 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 313, + 232, + 555, + 400 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 232, + 555, + 400 + ], + "spans": [ + { + "bbox": [ + 313, + 232, + 555, + 400 + ], + "type": "text", + "content": "Limitation. GUES uses the saliency map to guide the learning of the generative function. This method is effective for DR grading but encounters challenges in natural image scenarios. The natural images contain rich semantics, such as shape, relative structure, and complex background, which are not all relevant to tasks. However, the saliency map blindly highlights all those factors, struggling to capture the task-specific ones. In a theoretical point-view, the richness of these semantics means significant variations, resulting in a super-relaxed bound constant " + }, + { + "bbox": [ + 313, + 232, + 555, + 400 + ], + "type": "inline_equation", + "content": "U" + }, + { + "bbox": [ + 313, + 232, + 555, + 400 + ], + "type": "text", + "content": " (see Theorem 2) that undermines the descriptive power of the saliency map for " + }, + { + "bbox": [ + 313, + 232, + 555, + 400 + ], + "type": "inline_equation", + "content": "\\frac{\\partial \\delta_0}{\\partial x}" + }, + { + "bbox": [ + 313, + 232, + 555, + 400 + ], + "type": "text", + "content": ". In contrast, fundus images are more monolithic, implying a smaller " + }, + { + "bbox": [ + 313, + 232, + 555, + 400 + ], + "type": "inline_equation", + "content": "U" + }, + { + "bbox": [ + 313, + 232, + 555, + 400 + ], + "type": "text", + "content": " that justifies the usage of the saliency map. (The further discussion is provided in Supplementary)" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 313, + 411, + 387, + 423 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 411, + 387, + 423 + ], + "spans": [ + { + "bbox": [ + 313, + 411, + 387, + 423 + ], + "type": "text", + "content": "5. Conclusion" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 313, + 431, + 556, + 658 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 431, + 556, + 658 + ], + "spans": [ + { + "bbox": [ + 313, + 431, + 556, + 658 + ], + "type": "text", + "content": "In this paper, we propose a clinically motivated setting, OMG-DA, where the models are unseen prior to their use, and only target data flows are accessible. This setting ensures both model protection and source data privacy in a data flow scenario. To adapt to the target domain without access to the models, we introduce a GUES approach. Instead of conventional iterative optimization, we generate unadversarial examples for flowing target data by directly predicting individual perturbations. This approach is grounded in the theoretical results of generative unadversarial learning. In practice, we utilize the VAE model to learn the perturbation generation function with a latent input variable. Furthermore, we demonstrate that saliency maps can serve as an upper bound for this latent variable. This relationship inspires us to use the saliency maps as pseudo-perturbation labels for model training. Extensive experiments conducted on four DR benchmarks confirm that the proposed method can achieve state-of-the-art results when it pairs with both frozen pre-trained and fine-tuning models." + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 313, + 670, + 411, + 682 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 670, + 411, + 682 + ], + "spans": [ + { + "bbox": [ + 313, + 670, + 411, + 682 + ], + "type": "text", + "content": "Acknowledgments" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 313, + 689, + 554, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 689, + 554, + 713 + ], + "spans": [ + { + "bbox": [ + 313, + 689, + 554, + 713 + ], + "type": "text", + "content": "This work is supported by the National Natural Science Foundation of China (62476169, 62206168, 62276048)." + } + ] + } + ], + "index": 16 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "text", + "content": "28344" + } + ] + } + ], + "index": 17 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 7 + }, + { + "para_blocks": [ + { + "bbox": [ + 56, + 71, + 115, + 83 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 71, + 115, + 83 + ], + "spans": [ + { + "bbox": [ + 56, + 71, + 115, + 83 + ], + "type": "text", + "content": "References" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 57, + 91, + 297, + 713 + ], + "type": "list", + "angle": 0, + "index": 16, + "blocks": [ + { + "bbox": [ + 61, + 91, + 297, + 124 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 61, + 91, + 297, + 124 + ], + "spans": [ + { + "bbox": [ + 61, + 91, + 297, + 124 + ], + "type": "text", + "content": "[1] Aptos: Aptos 2019 blindness detection website. https://www.kaggle.com/c/aptos2019-blindness-detection, accessed February 20, 2022." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 61, + 126, + 297, + 168 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 61, + 126, + 297, + 168 + ], + "spans": [ + { + "bbox": [ + 61, + 126, + 297, + 168 + ], + "type": "text", + "content": "[2] Quadratic weighted kappa. https://www.Eyepacs.com/aroraaman/quadratic-kappa-metric-explained-in-5-simple-steps, accessed July 2022." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 62, + 171, + 296, + 225 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 171, + 296, + 225 + ], + "spans": [ + { + "bbox": [ + 62, + 171, + 296, + 225 + ], + "type": "text", + "content": "[3] Eman AbdelMaksoud, Sherif Barakat, and Mohammed Elmogy. A comprehensive diagnosis system for early signs and different diabetic retinopathy grades using fundus retinal images based on pathological changes detection. Computers in Biology and Medicine, 126:104039, 2020." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 62, + 227, + 296, + 258 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 227, + 296, + 258 + ], + "spans": [ + { + "bbox": [ + 62, + 227, + 296, + 258 + ], + "type": "text", + "content": "[4] Mohammad Atwany and Mohammad Yaqub. Drogen: domain generalization in diabetic retinopathy classification. In MICCAI, 2022." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 62, + 261, + 296, + 293 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 261, + 296, + 293 + ], + "spans": [ + { + "bbox": [ + 62, + 261, + 296, + 293 + ], + "type": "text", + "content": "[5] Haoxuan Che, Yuhan Cheng, Haibo Jin, and Hao Chen. Towards generalizable diabetic retinopathy grading in unseen domains. In MICCAI, 2023." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 62, + 296, + 296, + 350 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 296, + 296, + 350 + ], + "spans": [ + { + "bbox": [ + 62, + 296, + 296, + 350 + ], + "type": "text", + "content": "[6] Ling Dai, Liang Wu, Huating Li, Chun Cai, Qiang Wu, Hongyu Kong, Ruhan Liu, Xiangning Wang, Xuhong Hou, Yuexing Liu, et al. A deep learning system for detecting diabetic retinopathy across the disease spectrum. Nature communications, 12(1):3242, 2021." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 62, + 352, + 296, + 416 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 352, + 296, + 416 + ], + "spans": [ + { + "bbox": [ + 62, + 352, + 296, + 416 + ], + "type": "text", + "content": "[7] Etienne Decencière, Xiwei Zhang, Guy Cazuguel, Bruno Lay, Béatrice Cochener, Caroline Trone, Philippe Gain, John-Richard Ordóñez-Varela, Pascale Massin, Ali Erginay, et al. Feedback on a publicly distributed image database: the messidor database. Image Analysis & Stereology, pages 231-234, 2014." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 62, + 418, + 296, + 461 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 418, + 296, + 461 + ], + "spans": [ + { + "bbox": [ + 62, + 418, + 296, + 461 + ], + "type": "text", + "content": "[8] Along He, Tao Li, Ning Li, Kai Wang, and Huazhu Fu. Cabinet: Category attention block for imbalanced diabetic retinopathy grading. IEEE Transactions on Medical Imaging, 40(1):143-153, 2020." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 62, + 464, + 296, + 507 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 464, + 296, + 507 + ], + "spans": [ + { + "bbox": [ + 62, + 464, + 296, + 507 + ], + "type": "text", + "content": "[9] Yijin Huang, Junyan Lyu, Pujin Cheng, Roger Tam, and Xiaoying Tang. Ssit: Saliency-guided self-supervised image transformer for diabetic retinopathy grading. IEEE Journal of Biomedical and Health Informatics, 2024." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 57, + 509, + 296, + 530 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 509, + 296, + 530 + ], + "spans": [ + { + "bbox": [ + 57, + 509, + 296, + 530 + ], + "type": "text", + "content": "[10] Diederik P Kingma. Auto-encoding variational bayes. arXiv:1312.6114, 2013." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 57, + 533, + 294, + 553 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 533, + 294, + 553 + ], + "spans": [ + { + "bbox": [ + 57, + 533, + 294, + 553 + ], + "type": "text", + "content": "[11] Wouter M Kouw and Marco Loog. An introduction to domain adaptation and transfer learning. arXiv:1812.11806, 2018." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 57, + 555, + 296, + 587 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 555, + 296, + 587 + ], + "spans": [ + { + "bbox": [ + 57, + 555, + 296, + 587 + ], + "type": "text", + "content": "[12] Jonghyun Lee, Dahuin Jung, Junho Yim, and Sungroh Yoon. Confidence score for source-free unsupervised domain adaptation. In ICML, 2022." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 57, + 590, + 296, + 633 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 590, + 296, + 633 + ], + "spans": [ + { + "bbox": [ + 57, + 590, + 296, + 633 + ], + "type": "text", + "content": "[13] Tao Li, Yingqi Gao, Kai Wang, Song Guo, Hanruo Liu, and Hong Kang. Diagnostic assessment of deep learning algorithms for diabetic retinopathy screening. Information Sciences, 501:511-522, 2019." + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 57, + 635, + 296, + 678 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 635, + 296, + 678 + ], + "spans": [ + { + "bbox": [ + 57, + 635, + 296, + 678 + ], + "type": "text", + "content": "[14] Tao Li, Wang Bo, Chunyu Hu, Hong Kang, Hanruo Liu, Kai Wang, and Huazhu Fu. Applications of deep learning in fundus images: A review. Medical Image Analysis, 69: 101971, 2021." + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 57, + 681, + 296, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 681, + 296, + 713 + ], + "spans": [ + { + "bbox": [ + 57, + 681, + 296, + 713 + ], + "type": "text", + "content": "[15] Jian Liang, Dapeng Hu, and Jiashi Feng. Do we really need to access the source data? source hypothesis transfer for unsupervised domain adaptation. In ICML, 2020." + } + ] + } + ], + "index": 15 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 316, + 73, + 555, + 712 + ], + "type": "list", + "angle": 0, + "index": 33, + "blocks": [ + { + "bbox": [ + 316, + 73, + 555, + 106 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 73, + 555, + 106 + ], + "spans": [ + { + "bbox": [ + 316, + 73, + 555, + 106 + ], + "type": "text", + "content": "[16] Mattia Litrico, Alessio Del Bue, and Pietro Morerio. Guiding pseudo-labels with uncertainty estimation for source-free unsupervised domain adaptation. In CVPR, 2023." + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 317, + 108, + 555, + 152 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 108, + 555, + 152 + ], + "spans": [ + { + "bbox": [ + 317, + 108, + 555, + 152 + ], + "type": "text", + "content": "[17] Ruhan Liu, Xiangning Wang, Qiang Wu, Ling Dai, Xi Fang, Tao Yan, Jaemin Son, Shiqi Tang, Jiang Li, Zijian Gao, et al. Deepgrid: Diabetic retinopathy—grading and image quality estimation challenge. *Patterns*, 3(6), 2022." + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 316, + 154, + 555, + 186 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 154, + 555, + 186 + ], + "spans": [ + { + "bbox": [ + 316, + 154, + 555, + 186 + ], + "type": "text", + "content": "[18] Xingbin Liu, Huafeng Kuang, Xianming Lin, Yongjian Wu, and Rongrong Ji. Cat: Collaborative adversarial training. arXiv:2303.14922, 2023." + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 317, + 188, + 554, + 231 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 188, + 554, + 231 + ], + "spans": [ + { + "bbox": [ + 317, + 188, + 554, + 231 + ], + "type": "text", + "content": "[19] Sebastian Montabone and Alvaro Soto. Human detection using a mobile platform and novel features derived from a visual saliency mechanism. Image and Vision Computing, 28 (3):391-402, 2010." + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 317, + 234, + 555, + 277 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 234, + 555, + 277 + ], + "spans": [ + { + "bbox": [ + 317, + 234, + 555, + 277 + ], + "type": "text", + "content": "[20] Duy MH Nguyen, Truong TN Mai, Ngoc TT Than, Alexander Prange, and Daniel Sonntag. Self-supervised domain adaptation for diabetic retinopathy grading using vessel image reconstruction. In KI, 2021." + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 316, + 280, + 554, + 312 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 280, + 554, + 312 + ], + "spans": [ + { + "bbox": [ + 316, + 280, + 554, + 312 + ], + "type": "text", + "content": "[21] Shuaicheng Niu, Jiaxiang Wu, Yifan Zhang, Zhiquan Wen, Yaofo Chen, Peilin Zhao, and Mingkui Tan. Towards stable test-time adaptation in dynamic wild world. In ICLR, 2023." + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 316, + 315, + 554, + 357 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 315, + 554, + 357 + ], + "spans": [ + { + "bbox": [ + 316, + 315, + 554, + 357 + ], + "type": "text", + "content": "[22] Jiaming Qiu, Weikai Huang, Yijin Huang, Nanxi Yu, and Xiaoying Tang. Augpaste: A one-shot approach for diabetic retinopathy detection. Biomedical Signal Processing and Control, 96:106489, 2024." + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 317, + 360, + 554, + 413 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 360, + 554, + 413 + ], + "spans": [ + { + "bbox": [ + 317, + 360, + 554, + 413 + ], + "type": "text", + "content": "[23] Jinye Ran, Guanghua Zhang, Fan Xia, Ximei Zhang, Juan Xie, and Hao Zhang. Source-free active domain adaptation for diabetic retinopathy grading based on ultra-wide-field fundus images. Computers in Biology and Medicine, 174:108418, 2024." + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 316, + 416, + 555, + 459 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 416, + 555, + 459 + ], + "spans": [ + { + "bbox": [ + 316, + 416, + 555, + 459 + ], + "type": "text", + "content": "[24] Hadi Salman, Andrew Ilyas, Logan Engstrom, Sai Vemprala, Aleksander Madry, and Ashish Kapoor. Unadversarial examples: Designing objects for robust vision. In NeurIPS, 2021." + } + ] + } + ], + "index": 25 + }, + { + "bbox": [ + 316, + 461, + 554, + 494 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 461, + 554, + 494 + ], + "spans": [ + { + "bbox": [ + 316, + 461, + 554, + 494 + ], + "type": "text", + "content": "[25] Abhijith Sharma, Phil Munz, and Apurva Narayan. Nsa: Naturalistic support artifact to boost network confidence. In IJCNN, 2023." + } + ] + } + ], + "index": 26 + }, + { + "bbox": [ + 316, + 496, + 554, + 529 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 496, + 554, + 529 + ], + "spans": [ + { + "bbox": [ + 316, + 496, + 554, + 529 + ], + "type": "text", + "content": "[26] Reza Shokri, Marco Stronati, Congzheng Song, and Vitaly Shmatikov. Membership inference attacks against machine learning models. In S&P, 2017." + } + ] + } + ], + "index": 27 + }, + { + "bbox": [ + 316, + 531, + 555, + 563 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 531, + 555, + 563 + ], + "spans": [ + { + "bbox": [ + 316, + 531, + 555, + 563 + ], + "type": "text", + "content": "[27] Daniel E Singer, David M Nathan, Howard A Fogel, and Andrew P Schachat. Screening for diabetic retinopathy. Annals of Internal Medicine, 116(8):660-671, 1992." + } + ] + } + ], + "index": 28 + }, + { + "bbox": [ + 316, + 566, + 555, + 586 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 566, + 555, + 586 + ], + "spans": [ + { + "bbox": [ + 316, + 566, + 555, + 586 + ], + "type": "text", + "content": "[28] C Szegedy. Intriguing properties of neural networks. arXiv:1312.6199, 2013." + } + ] + } + ], + "index": 29 + }, + { + "bbox": [ + 317, + 590, + 555, + 632 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 590, + 555, + 632 + ], + "spans": [ + { + "bbox": [ + 317, + 590, + 555, + 632 + ], + "type": "text", + "content": "[29] Song Tang, An Chang, Fabian Zhang, Xiatian Zhu, Mao Ye, and Changshui Zhang. Source-free domain adaptation via target prediction distribution searching. International Journal of Computer Vision, 132(3):654-672, 2024." + } + ] + } + ], + "index": 30 + }, + { + "bbox": [ + 316, + 635, + 555, + 666 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 635, + 555, + 666 + ], + "spans": [ + { + "bbox": [ + 316, + 635, + 555, + 666 + ], + "type": "text", + "content": "[30] Song Tang, Wenxin Su, Mao Ye, Jianwei Zhang, and Xiatian Zhu. Unified source-free domain adaptation. arXiv:2403.07601, 2024." + } + ] + } + ], + "index": 31 + }, + { + "bbox": [ + 316, + 670, + 555, + 712 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 670, + 555, + 712 + ], + "spans": [ + { + "bbox": [ + 316, + 670, + 555, + 712 + ], + "type": "text", + "content": "[31] Nishtha Tomar, Sushmita Chandel, and Gaurav Bhatnagar. A visual attention-based algorithm for brain tumor detection using an on-center saliency map and a superpixel-based framework. Healthcare Analytics, 5:100323, 2024." + } + ] + } + ], + "index": 32 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "text", + "content": "28345" + } + ] + } + ], + "index": 34 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 8 + }, + { + "para_blocks": [ + { + "bbox": [ + 56, + 73, + 297, + 421 + ], + "type": "list", + "angle": 0, + "index": 9, + "blocks": [ + { + "bbox": [ + 56, + 73, + 297, + 116 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 73, + 297, + 116 + ], + "spans": [ + { + "bbox": [ + 56, + 73, + 297, + 116 + ], + "type": "text", + "content": "[32] Hugo Touvron, Matthieu Cord, Matthijs Douze, Francisco Massa, Alexandre Sablayrolles, and Hervé Jégou. Training data-efficient image transformers & distillation through attention. In ICML, 2021." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 56, + 118, + 295, + 150 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 118, + 295, + 150 + ], + "spans": [ + { + "bbox": [ + 56, + 118, + 295, + 150 + ], + "type": "text", + "content": "[33] Jeya Maria Jose Valanarasu, Pengfei Guo, VS Vibashan, and Vishal M Patel. On-the-fly test-time adaptation for medical image segmentation. In MIDL, 2024." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 56, + 152, + 295, + 185 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 152, + 295, + 185 + ], + "spans": [ + { + "bbox": [ + 56, + 152, + 295, + 185 + ], + "type": "text", + "content": "[34] Dequan Wang, Evan Shelhamer, Shaoteng Liu, Bruno Olshausen, and Trevor Darrell. Tent: Fully test-time adaptation by entropy minimization. In ICLR, 2020." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 56, + 186, + 295, + 228 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 186, + 295, + 228 + ], + "spans": [ + { + "bbox": [ + 56, + 186, + 295, + 228 + ], + "type": "text", + "content": "[35] Tianyunxi Wei, Yijin Huang, Li Lin, Pujin Cheng, Sirui Li, and Xiaoying Tang. Saliency-guided and patch-based mixup for long-tailed skin cancer image classification. arXiv:2406.10801, 2024." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 56, + 230, + 295, + 284 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 230, + 295, + 284 + ], + "spans": [ + { + "bbox": [ + 56, + 230, + 295, + 284 + ], + "type": "text", + "content": "[36] Zhan Wu, Gonglei Shi, Yang Chen, Fei Shi, Xinjian Chen, Gouenou Coatieux, Jian Yang, Limin Luo, and Shuo Li. Coarse-to-fine classification for diabetic retinopathy grading using convolutional neural network. Artificial Intelligence in Medicine, 108:101936, 2020." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 56, + 286, + 295, + 319 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 286, + 295, + 319 + ], + "spans": [ + { + "bbox": [ + 56, + 286, + 295, + 319 + ], + "type": "text", + "content": "[37] Tongkun Xu, Weihua Chen, WANG Pichao, Fan Wang, Hao Li, and Rong Jin. Cdtrans: Cross-domain transformer for unsupervised domain adaptation. In ICLR, 2021." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 56, + 320, + 295, + 352 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 320, + 295, + 352 + ], + "spans": [ + { + "bbox": [ + 56, + 320, + 295, + 352 + ], + "type": "text", + "content": "[38] Shiqi Yang, Joost van de Weijer, Luis Herranz, Shangling Jui, et al. Exploiting the intrinsic neighborhood structure for source-free domain adaptation. In NeurIPS, 2021." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 56, + 354, + 295, + 387 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 354, + 295, + 387 + ], + "spans": [ + { + "bbox": [ + 56, + 354, + 295, + 387 + ], + "type": "text", + "content": "[39] Hongxu Yin, Arun Mallya, Arash Vahdat, Jose M Alvarez, Jan Kautz, and Pavlo Molchanov. See through gradients: Image batch recovery via grad inversion. In CVPR, 2021." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 56, + 388, + 295, + 421 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 388, + 295, + 421 + ], + "spans": [ + { + "bbox": [ + 56, + 388, + 295, + 421 + ], + "type": "text", + "content": "[40] Chenrui Zhang, Tao Lei, and Ping Chen. Diabetic retinopathy grading by a source-free transfer learning approach. Biomedical Signal Processing and Control, 73:103423, 2022." + } + ] + } + ], + "index": 8 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 749, + 318, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 749, + 318, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 749, + 318, + 757 + ], + "type": "text", + "content": "28346" + } + ] + } + ], + "index": 10 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 9 + } + ], + "_backend": "vlm", + "_version_name": "2.6.4" +} \ No newline at end of file diff --git a/2025/Domain Generalization in CLIP via Learning with Diverse Text Prompts/51839507-6fb2-45c2-a66c-0644bd7936ed_content_list.json b/2025/Domain Generalization in CLIP via Learning with Diverse Text Prompts/51839507-6fb2-45c2-a66c-0644bd7936ed_content_list.json new file mode 100644 index 0000000000000000000000000000000000000000..ffe1f6e13bd557ea213e36bb293dbf73611829ef --- /dev/null +++ b/2025/Domain Generalization in CLIP via Learning with Diverse Text Prompts/51839507-6fb2-45c2-a66c-0644bd7936ed_content_list.json @@ -0,0 +1,1689 @@ +[ + { + "type": "text", + "text": "Domain Generalization in CLIP via Learning with Diverse Text Prompts", + "text_level": 1, + "bbox": [ + 133, + 130, + 866, + 152 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Changsong Wen, Zelin Peng, Yu Huang, Xiaokang Yang, Wei Shen MoE Key Lab of Artificial Intelligence, AI Institute, Shanghai Jiao Tong University", + "bbox": [ + 166, + 180, + 830, + 217 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "{changsong, zelin.peng, yellowfish, xkyang, wei.shen}@sjtu.edu.cn", + "bbox": [ + 209, + 219, + 784, + 234 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Abstract", + "text_level": 1, + "bbox": [ + 246, + 268, + 328, + 284 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Domain generalization (DG) aims to train a model on source domains that can generalize well to unseen domains. Recent advances in Vision-Language Models (VLMs), such as CLIP, exhibit remarkable generalization capabilities across a wide range of data distributions, benefiting tasks like DG. However, CLIP is pre-trained by aligning images with their descriptions, which inevitably captures domain-specific details. Moreover, adapting CLIP to source domains with limited feature diversity introduces bias. These limitations hinder the model's ability to generalize across domains. In this paper, we propose a new DG approach by learning with diverse text prompts. These text prompts incorporate varied contexts to imitate different domains, enabling DG model to learn domain-invariant features. The text prompts guide DG model learning in three aspects: feature suppression, which uses these prompts to identify domain-sensitive features and suppress them; feature consistency, which ensures the model's features are robust to domain variations imitated by the diverse prompts; and feature diversification, which diversifies features based on the prompts to mitigate bias. Experimental results show that our approach improves domain generalization performance on five datasets from the DomainBed benchmark, achieving state-of-the-art results.", + "bbox": [ + 89, + 301, + 485, + 662 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "1. Introduction", + "text_level": 1, + "bbox": [ + 89, + 696, + 220, + 712 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "In the complex real-world applications, deep learning models are expected to be capable of recognizing data from a wide range of distributions. Within the past few years, researchers are dedicated to investigating domain generalization (DG) [6, 7, 9, 24, 25, 27], which empowers the model to learn robust features from multiple domains and exhibit good performance across various data distributions. Recent advances in foundation Vision-Language Models (VLMs) [34, 35, 52, 61], such as CLIP [49], are trained on large-scale data sourced from the web with a wide variety", + "bbox": [ + 89, + 722, + 483, + 875 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/2745900baa7e7c20eff0effcd1eedfb1ebef11f5cf1d26f966d7308491e855b3.jpg", + "image_caption": [ + "Figure 1. Limitations with existing methods. As shown on the left, CLIP's image encoder is pre-trained to align with textual descriptions, which inevitably captures domain-specific visual details, resulting in perturbations in domain generalization. Additionally, as shown on the right, due to the limited feature diversity in the source domains, adapting CLIP with source domains will introduce bias." + ], + "image_footnote": [], + "bbox": [ + 535, + 294, + 906, + 465 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "of data distributions. These VLMs possess remarkable generalization capabilities, and boost many open-world recognition tasks, including out-of-distribution learning [50] and domain generalization [1, 10, 51].", + "bbox": [ + 511, + 627, + 906, + 686 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Existing DG studies [20, 23, 27, 50, 60] leverage CLIP's pre-trained weights with techniques like parameter regularization [23] and weight space ensembling [6, 27, 60] to adapt the model for downstream datasets, while preserving generalization ability. Other recent DG works [1, 26] propose using knowledge distillation to transfer CLIP's knowledge into DG models. However, as shown in Fig. 1, CLIP [49] aligns images with their descriptions during pretraining, which inevitably contain domain-specific details, impacting existing DG models. Additionally, since the training images are limited to a few source domains in DG, adapting CLIP within these images leads the model to emphasize the features specific to the source domains. This results in a lack of feature diversity [43, 46, 58], which com", + "bbox": [ + 511, + 689, + 908, + 900 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "CVF", + "bbox": [ + 106, + 2, + 181, + 42 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore.", + "bbox": [ + 236, + 0, + 810, + 46 + ], + "page_idx": 0 + }, + { + "type": "page_footnote", + "text": "Corresponding author.", + "bbox": [ + 114, + 887, + 250, + 900 + ], + "page_idx": 0 + }, + { + "type": "page_number", + "text": "9559", + "bbox": [ + 482, + 944, + 514, + 955 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "promises the model's ability to generalize to new domains.", + "bbox": [ + 89, + 90, + 478, + 107 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Erasing domain-sensitive details and mitigating bias benefit from learning domain-invariant representations for each class. In this paper, we propose a new DG approach to learn domain-invariant class features by leveraging a diverse set of text prompts. Based on the multi-modal alignment capabilities of CLIP, we use these text prompts to imitate multiple domains and effectively guide the DG learning process. A wide range of domains provides diverse perspectives on each class, allowing the model to learn features that are robust across varying contexts, thereby promoting domain-invariant feature learning.", + "bbox": [ + 88, + 107, + 482, + 272 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Diverse text prompts support three key aspects of our approach to learning domain-invariant features: they facilitate domain-sensitive feature suppression, ensure feature consistency across domain variations, and enable feature diversification to reduce bias. To be specific, we leverage the prompts recommended by original CLIP [49] to approximate different domains, such as \"a painting of a {class}\" or \"a {class} in a video game\". We first identify channels in the features activated by specific text prompts. These channels are considered highly relevant to the domain context of the prompt, thus we suppress their activation values. Next, we compare predictions under perturbations of diverse text prompts, and employ a consistency constraint to ensure the model is robust to domain variations while maintaining stable predictions. Subsequently, we decompose the component of the image features orthogonal to the text prompts and impose diversification on the orthogonal component. This approach enables a more diverse distribution of image features without impacting classification.", + "bbox": [ + 91, + 275, + 483, + 561 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "To comprehensively evaluate our method, we perform extensive experiments on the widely used DomainBed benchmark [21]. The experiment results demonstrate the superior performance and robustness of our proposed approach, achieving state-of-the-art results across five domain generalization datasets.", + "bbox": [ + 89, + 561, + 482, + 652 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "2. Related Work", + "text_level": 1, + "bbox": [ + 89, + 669, + 232, + 684 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "2.1. Large-scale Vision Language Models", + "text_level": 1, + "bbox": [ + 89, + 695, + 408, + 710 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Large-scale Vision Language Models (VLMs) [28, 49, 52, 61], trained on extensive image-text data from the internet, demonstrate strong robustness to diverse data distributions. These VLMs exhibit remarkable generalization capabilities, serving as fundamental models in the field of domain generalization. CLIP [49] is one of the most prevalent approaches among VLMs. It is trained by aligning image-text pairs through contrastive learning, where the model learns to associate corresponding images and texts from large-scale datasets by maximizing their similarity in a shared embedding space. Among the past few years, many efforts have been made to improve the CLIP-like VLMs by designing", + "bbox": [ + 89, + 719, + 482, + 901 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "better loss function [61], mitigating the impact of noisy web data [34, 35], and developing optimization techniques [52]. Another line of works focus on leveraging CLIP's generalization capabilities and its cross-modal alignment strengths to address downstream tasks such as zero-shot classification [29, 64], domain generalization [1, 26], and 2D/3D dense prediction [14, 45, 48].", + "bbox": [ + 511, + 90, + 903, + 196 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "2.2. Domain Generalization", + "text_level": 1, + "bbox": [ + 511, + 207, + 730, + 222 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Domain generalization [3, 8, 10, 18, 31, 40, 53, 59, 63] seeks to develop models capable of adapting to tasks from unseen domains without additional training. Conventional domain generalization methods encompass approaches such as learning domain-invariant features [13, 22, 37, 39], data augmentation [36, 42, 56], and adversarial learning [33, 38] to improve generalization to unseen domains.", + "bbox": [ + 511, + 229, + 903, + 349 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Recently, VLMs are widely utilized as foundational models in domain generalization [1, 20, 26, 50, 60]. Some approaches fine-tune CLIP model on downstream DG datasets. However, since these DG datasets (0.01 to $0.5\\mathrm{M}$ ) are significantly smaller than CLIP's training set (400M), directly fine-tuning disrupts CLIP's parameter space [30], leading to degraded generalization performance. To mitigate the limitations, many impressive works [6, 30, 60] propose parameter ensemble methods between CLIP and the fine-tuned model to preserve generalization ability. Other works adopt CLIP for DG downstream tasks by designing loss functions during fine-tuning [20, 50], or employing regularization in parameter-efficient fine-tuning method [23]. Some knowledge distillation-based methods [1, 26] align the features of DG models with the output features of CLIP, transferring CLIP's generalization ability to DG models.", + "bbox": [ + 511, + 352, + 905, + 594 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "DomainDrop [22] mitigates model feature sensitivity to different domains by suppressing channel-wise feature variance across source domains. PromptStyler [11] generates diverse templates with style constraints and content consistency to train a text classifier. The text classifier is directly applied to visual domain generalization. LADS [16] and TEAM [47] employ text prompts for facilitating image feature augmentation. CAE [41] manually design a domain bank to obtain domain-unified text prompts for a stable prediction across domains. Different from these works that recognize the importance of text prompts, our approach enables direct optimization of the visual encoder by imitating multiple domain variations through a set of text prompts.", + "bbox": [ + 511, + 595, + 905, + 791 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "2.3. Prompt Engineering", + "text_level": 1, + "bbox": [ + 511, + 801, + 707, + 816 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "In CLIP, the design of prompts promotes the model to effectively capture the key information within the image. Recent prompt learning methods [29, 64, 65] are proposed to optimize the prompt as trainable parameters, allowing the model to dynamically adjust the prompt based on tasks.", + "bbox": [ + 511, + 825, + 903, + 902 + ], + "page_idx": 1 + }, + { + "type": "page_number", + "text": "9560", + "bbox": [ + 480, + 944, + 514, + 955 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "CoOp [65] enhances CLIP's adaptability for visual recognition tasks by using learnable context vectors as language prompts. CoCoOp [64] incorporates image features as conditional prompts to better capture context. MaPLe [29] introduces a multimodal prompt learning approach, finetuning both image and text encoders simultaneously for improved performance.", + "bbox": [ + 88, + 90, + 480, + 196 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Recently, the domain generalization method DPL [62] employs an MLP to capture domain-specific image features as context for the learnable prompt. SPG [4] designs a prompt generator with domain knowledge to generate soft prompts for the images from target domain. These methods [4, 19, 62] primarily focus on incorporating image context into the text encoder to achieve multimodal alignment in downstream tasks. Differently, we optimize the image encoder through prompts from diverse templates. These prompts effectively imitate different domains while preventing the parameters of learnable prompts from being influenced by source domain-specific information during training.", + "bbox": [ + 89, + 198, + 482, + 396 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3. Preliminary", + "text_level": 1, + "bbox": [ + 89, + 411, + 217, + 429 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Problem Formulation. Domain generalization aims to train a model that can generalize well to unseen target domains, even when the training data comes from multiple source domains with potentially different data distributions. Specifically, the model is trained using data from $o$ different source domains $\\{D_i^s\\}_{i=1}^o$ , where each source domain $D_i^s = \\{(x_i, y_i)\\}_{i=1}^u$ consists of input data $x_i$ and corresponding class labels $y_i$ , with $u$ representing the number of samples in the source domain.", + "bbox": [ + 89, + 438, + 482, + 571 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "The model's generalization performance is evaluated on the target domain $D^{t}$ , which shares the same set of class labels as the source domains $\\{D_{i}^{s}\\}_{i = 1}^{o}$ . The joint distribution of input data $x$ and labels $y$ in the target domain differs from that in the source domains.", + "bbox": [ + 89, + 575, + 482, + 650 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Utilizing CLIP for domain generalization. CLIP's strong generalization capability stems from its multimodal training on diverse internet data covering a wide range of concepts and contexts. As a prominent vision-language model, CLIP comprises an image encoder $F_{I}(\\cdot)$ and a text encoder $F_{T}(\\cdot)$ . Both image and text encoders adopt a transformer architecture [15, 54] to generate their representations through self-attention.", + "bbox": [ + 89, + 652, + 482, + 772 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "CLIP employs contrastive learning to align images and their descriptive text during pre-training. It associates visual and textual features by maximizing the similarity of the aligned image-text pairs while reducing the similarity of other ones:", + "bbox": [ + 89, + 775, + 482, + 849 + ], + "page_idx": 2 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {L} _ {\\mathrm {C L I P}} = - \\frac {1}{N} \\sum_ {i = 1} ^ {N} \\log \\frac {\\exp \\left(\\mathrm {S} \\left(I _ {i} , A _ {i}\\right) / \\tau\\right)}{\\sum_ {j = 1} ^ {N} \\exp \\left(\\mathrm {S} \\left(I _ {i} , A _ {j}\\right) / \\tau\\right)}, \\tag {1}\n$$\n", + "text_format": "latex", + "bbox": [ + 127, + 862, + 482, + 905 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "where $I$ and $A$ represent the input image and caption, $i$ , $j$ are indexes of the inputs, $S$ means similarity matrices, such as cosine similarity. $\\tau$ is a temperature parameter used to adjust the smoothness of the similarity distribution.", + "bbox": [ + 511, + 90, + 903, + 151 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Baseline of our approach. In the conventional domain generalization methods [1, 23, 29], \"a photo of a $\\{class\\}$ \" is encoded by CLIP's text encoder as text prompt $\\{P_i\\}_{i=1}^c$ , representing the main concepts of $c$ different classes. Leveraging CLIP's remarkable multi-modal alignment capabilities between images and text, the text prompts for different classes are treated as domain-invariant representations of the image features [1, 26]. We adopt a basic baseline [1] to adapt CLIP for downstream domain generalization tasks by simultaneously distilling image and text features to the DG model.", + "bbox": [ + 511, + 152, + 906, + 316 + ], + "page_idx": 2 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {L} _ {\\mathrm {K D}} = \\lambda \\cdot \\cos \\left(F _ {I} \\left(x _ {i}\\right), f _ {I} \\left(x _ {i}\\right)\\right) + (1 - \\lambda) \\cdot \\cos \\left(F _ {T} \\left(y _ {i}\\right), f _ {I} \\left(x _ {i}\\right)\\right), \\tag {2}\n$$\n", + "text_format": "latex", + "bbox": [ + 513, + 330, + 908, + 361 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "where $F_{I}(\\cdot)$ and $F_{T}(\\cdot)$ are CLIP's image and text encoders, $f_{I}(\\cdot)$ is the DG model. Our approach differs slightly in that we ensemble multiple prompts used to imitate different domains during the distillation process. This has a minor direct impact on performance but is more intuitive to the design of our overall framework, ensuring stability across domain variations. The classification result $\\hat{y}$ is obtained by choosing the ensembled text prompt with the highest similarity to the $i$ -th image.", + "bbox": [ + 511, + 361, + 906, + 498 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "4. Methodology", + "text_level": 1, + "bbox": [ + 511, + 512, + 648, + 530 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "In this section, we introduce our proposed learning with diverse prompts paradigm for learning domain-invariant features in DG. Initially, we filter out unreliable prompts based on the prediction confidence. Next, we propose a Domain-sensitive Channel Suppression (DCS) method to identify and suppress domain-sensitive features. Moreover, a Multi-domain Consistency Constraint (MCC) approach is designed to make the features robust to variations across domains. Then, we apply Orthogonal Uniformity Regularization (OUR) to diversify features in the DG model.", + "bbox": [ + 511, + 537, + 906, + 690 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "4.1. Entropy-based Prompt Filtering", + "text_level": 1, + "bbox": [ + 511, + 700, + 797, + 717 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "We imitate different domains by using the text templates $\\{T_j\\}_{j=1}^n$ recommended in CLIP [49], such as \"a drawing of a {class}\", $n$ is the number of templates. These prompts are generated by encoding the combination of the $j$ -th text template $T_j$ and the $i$ -th class label $y_i$ by CLIP text encoder:", + "bbox": [ + 511, + 724, + 905, + 800 + ], + "page_idx": 2 + }, + { + "type": "equation", + "text": "\n$$\nP _ {i} ^ {j} = F _ {T} \\left(y _ {i}, T _ {j}\\right), \\tag {3}\n$$\n", + "text_format": "latex", + "bbox": [ + 648, + 813, + 903, + 832 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "where $P \\in \\mathbb{R}^{n \\times c \\times D}$ , $c$ is the number of classes, $D$ represents the feature dimensions, $i$ and $j$ indicate the index of the classes and prompt templates. These prompts employ diverse contexts to describe each class, such as a drawing,", + "bbox": [ + 511, + 839, + 906, + 901 + ], + "page_idx": 2 + }, + { + "type": "page_number", + "text": "9561", + "bbox": [ + 480, + 944, + 513, + 955 + ], + "page_idx": 2 + }, + { + "type": "image", + "img_path": "images/d7d94cbb7475280d264b29faded19619031e0a3df2471fb30d2a5b37785fc980.jpg", + "image_caption": [ + "Figure 2. Overview of our proposed method. To enforce the DG model to learn domain-invariant class features, we imitate multiple domains by text prompts. Three components of our method, including feature suppression, feature consistency, and feature diversification, are achieved through Domain-sensitive Channel Suppression (DCS), Multi-domain Consistency Constraint (MCC), and Orthogonal Uniformity Regularization (OUR) modules. The blue line represents the process during inference." + ], + "image_footnote": [], + "bbox": [ + 148, + 146, + 844, + 439 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "a sculpture, or a rendering, etc. Thus, these prompts can be naturally regarded as descriptions of the class from multiple domains. The DG model $f_{I}(\\cdot)$ is regularized to extract domain-invariant features that are robust to prompts from different domains.", + "bbox": [ + 88, + 551, + 480, + 626 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Within the prompts provided by CLIP, we first discard a few unreliable ones based on the prediction confidence for each input [2], as they negatively affect the feature channel suppression and feature consistency calculation, which rely on accurately identifying relevant features. In the Entropy-based Prompt Filtering (EPF), we first extract the encoded image features $I_{i} = f_{I}(x_{i})$ for the input image $x_{i}$ . For the $j$ -th template, the predicted probability for the $i$ -th class is calculated as:", + "bbox": [ + 89, + 628, + 482, + 762 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\hat {d} _ {i} ^ {j} = \\frac {\\exp \\left(\\cos \\left(I _ {i} , P _ {i} ^ {j}\\right)\\right)}{\\sum_ {k = 1} ^ {c} \\exp \\left(\\cos \\left(I _ {i} , P _ {k} ^ {j}\\right)\\right)}, \\tag {4}\n$$\n", + "text_format": "latex", + "bbox": [ + 181, + 773, + 482, + 811 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "We denote that $\\hat{d}_i^j$ is a constant and $\\hat{d}^j = \\{\\hat{d}_i^j\\}_{i=1}^c$ is a predicted distribution for the $j$ -th template.", + "bbox": [ + 89, + 821, + 482, + 854 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Entropy quantifies the uncertainty of a probability distribution. If the model output has high confidence, the distribution typically concentrates on a specific class, resulting", + "bbox": [ + 89, + 854, + 483, + 901 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "in low entropy. Conversely, if the model has similar confidence across multiple classes, the distribution is more uniform, leading to higher entropy. Therefore, we calculate the entropy for each template's prediction:", + "bbox": [ + 511, + 551, + 905, + 613 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\mathrm {H} \\left(\\hat {d} ^ {j}\\right) = - \\sum_ {i = 1} ^ {c} \\hat {d} _ {i} ^ {j} \\log \\left(\\hat {d} _ {i} ^ {j}\\right). \\tag {5}\n$$\n", + "text_format": "latex", + "bbox": [ + 617, + 619, + 903, + 657 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Then, we eliminate the unreliable prompts and select the $m$ prompts with the lowest entropy for each input:", + "bbox": [ + 511, + 665, + 905, + 696 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\left\\{T _ {1}, T _ {2}, \\dots , T _ {m} \\right\\} = \\arg \\min _ {j \\in \\{1 \\dots , n \\}} \\mathrm {H} \\left(\\hat {d} ^ {j}\\right), \\tag {6}\n$$\n", + "text_format": "latex", + "bbox": [ + 560, + 702, + 903, + 722 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "4.2. Domain-sensitive Channel Suppression", + "text_level": 1, + "bbox": [ + 511, + 726, + 848, + 742 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "The image features should focus on discriminative features between classes and be robust to prompts from different domains. However, some domain-sensitive channels encode variations specific to the domain context of prompt, introducing perturbations. To address this, we propose Domain-sensitive Channel Suppression (DCS) to select and suppress these channels, ensuring the extracted image embedding captures domain-invariant class features.", + "bbox": [ + 511, + 750, + 905, + 869 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Specifically, we utilize gradient attribution [12] to quantify each channel's contribution in the image features when", + "bbox": [ + 511, + 869, + 905, + 900 + ], + "page_idx": 3 + }, + { + "type": "page_number", + "text": "9562", + "bbox": [ + 480, + 944, + 514, + 955 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "calculating their similarity with different prompts. We simultaneously consider the model's sensitivity to each channel (i.e., the gradient) and the actual activation value of that channel to measure each image channel's contribution to recognition with $P_{i}^{j}$ :", + "bbox": [ + 89, + 90, + 483, + 167 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\alpha_ {k j} = \\left| \\frac {\\partial \\cos \\left(I _ {i} , P _ {i} ^ {j}\\right)}{\\partial I _ {i} ^ {k}} \\right| \\cdot I _ {i} ^ {k}, \\tag {7}\n$$\n", + "text_format": "latex", + "bbox": [ + 194, + 180, + 482, + 215 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "where $P_{i}^{j}$ means the prompt corresponding to the $i$ -th image's label with $T_{j}$ as the template. $\\alpha_{kj}$ is a constant that represents the contribution of the $k$ -th image channel to the similarity calculation with the $j$ -th prompt $P_{i}^{j}$ . For each image feature, we normalize it through the channel dimensions:", + "bbox": [ + 89, + 229, + 483, + 319 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\bar {\\alpha} _ {k j} = \\frac {\\alpha_ {k j}}{\\frac {1}{D} \\sqrt {\\sum_ {k = 1} ^ {D} \\alpha_ {k j} ^ {2}}}, \\tag {8}\n$$\n", + "text_format": "latex", + "bbox": [ + 205, + 319, + 482, + 357 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Then, we calculate the z-score among different prompts:", + "bbox": [ + 89, + 364, + 465, + 381 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\alpha_ {k j} ^ {\\mathrm {z - s c o r e}} = \\frac {\\overline {{\\alpha}} _ {k j} - \\mu}{\\sigma}, \\tag {9}\n$$\n", + "text_format": "latex", + "bbox": [ + 218, + 392, + 482, + 421 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "$\\mu$ and $\\sigma$ are the average of $\\overline{\\alpha}_{kj}$ across all templates, which are calculated as:", + "bbox": [ + 89, + 431, + 483, + 460 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\mu = \\frac {1}{m} \\sum_ {j = 1} ^ {m} \\bar {\\alpha} _ {k j} \\tag {10}\n$$\n", + "text_format": "latex", + "bbox": [ + 230, + 472, + 482, + 513 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\sigma = \\sqrt {\\frac {1}{m} \\sum_ {j = 1} ^ {m} \\left(\\bar {\\alpha} _ {k j} - \\mu\\right) ^ {2}} \\tag {11}\n$$\n", + "text_format": "latex", + "bbox": [ + 197, + 531, + 482, + 579 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "We check the z-score values of each channel across different prompts. A high z-score indicates that a channel is activated by the context provided by the specific prompt, showing a strong correlation with domain-specific information. We apply a mask to such image feature channels before calculating the loss:", + "bbox": [ + 89, + 585, + 483, + 676 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\mathbf {M} _ {D} = \\left\\{ \\begin{array}{l l} 1, & \\text {i f} \\max _ {k} \\alpha_ {k j} ^ {\\mathrm {z - s c o r e}} > \\beta \\\\ 0, & \\text {o t h e r w i s e} \\end{array} \\right. \\tag {12}\n$$\n", + "text_format": "latex", + "bbox": [ + 169, + 689, + 482, + 729 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "The mask $\\mathbf{M}_D\\in \\mathbb{R}^D$ is directly employed on image features by the element-wise multiplication $I_i^k = \\mathbf{M}_D\\odot I_i$ . We suppress these domain-sensitive features by applying $L_{1}$ regularization:", + "bbox": [ + 89, + 743, + 483, + 804 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {L} _ {r e g} = \\gamma \\sum_ {i = 1} ^ {B} \\sum_ {k = 1} ^ {D} \\left| I _ {i} ^ {k} \\right|, \\tag {13}\n$$\n", + "text_format": "latex", + "bbox": [ + 210, + 816, + 482, + 858 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "where $B$ represents the batch size, and $\\gamma$ is the weight of loss, $I_{i}^{k}$ is the masked image feature.", + "bbox": [ + 89, + 869, + 483, + 901 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "4.3. Multi-domain Consistency Constraint", + "text_level": 1, + "bbox": [ + 513, + 90, + 841, + 106 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "We further introduce Multi-domain Consistency Constraint (MCC) on the predicted distribution between the image and varied prompts. We calculate the similarity between image features and different prompts for classification. The KL loss function is then applied to ensure that the predicted distributions remain consistent between different prompts, enhancing the robustness of the image encoder to domain variations. Specifically, we use Eq. (4) to obtain the predicted distributions for different templates $\\{\\hat{d}^j\\}_{j=1}^m$ , and apply KL loss to ensure consistency:", + "bbox": [ + 511, + 113, + 906, + 266 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\operatorname {K L} \\left(\\hat {d} ^ {j} \\| \\hat {d} ^ {k}\\right) = \\sum_ {i = 1} ^ {c} \\hat {d} _ {i} ^ {j} \\log \\frac {\\hat {d} _ {i} ^ {j}}{\\hat {d} _ {i} ^ {k}} \\tag {14}\n$$\n", + "text_format": "latex", + "bbox": [ + 612, + 276, + 906, + 316 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {L} _ {\\mathrm {K L}} = \\frac {1}{m (m - 1)} \\sum_ {j = 1} ^ {m} \\sum_ {\\substack {k = 1 \\\\ k \\neq t}} ^ {m} \\mathrm {K L} \\left(\\hat {d} ^ {j} \\| \\hat {d} ^ {k}\\right) \\tag{15}\n$$\n", + "text_format": "latex", + "bbox": [ + 580, + 330, + 905, + 380 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "The KL loss compares the prediction with prompts from diverse domains and enforces consistency, guiding the learned features to focus on domain-invariant information.", + "bbox": [ + 511, + 386, + 905, + 431 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "4.4. Orthogonal Uniformity Regularization", + "text_level": 1, + "bbox": [ + 513, + 441, + 848, + 459 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Due to the limited diversity of image source domains, learning with these images biases CLIP toward the feature distribution of these source domains. To mitigate this bias, we further diversify the normalized image features to achieve a more uniform distribution in the hyperspherical space. Uniformity of the features increases the possibility that unseen features are positioned near seen features [46], thereby enhancing the model's generalization through feature diversification.", + "bbox": [ + 511, + 465, + 906, + 601 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "However, it is not feasible to directly diversify the image features. As mentioned in Sec. 3, we use an ensembled prompt $e_i$ in classification, which can be formulated as:", + "bbox": [ + 511, + 602, + 905, + 648 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\ne _ {i} = \\frac {1}{n} \\sum_ {j = 1} ^ {n} P _ {i} ^ {j}. \\tag {16}\n$$\n", + "text_format": "latex", + "bbox": [ + 653, + 660, + 905, + 700 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "We compute the similarity between the ensembled prompt $e_i$ and the image feature $I_i$ for classification:", + "bbox": [ + 511, + 712, + 905, + 743 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\hat {y} = \\operatorname {a r g m a x} _ {j} \\cos \\left(I _ {i}, e _ {j}\\right). \\tag {17}\n$$\n", + "text_format": "latex", + "bbox": [ + 624, + 756, + 906, + 773 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Therefore, as demonstrated in Eq. (17), directly applying uniformity constraints on image features can impact the similarity calculation between image features and prompts.", + "bbox": [ + 511, + 784, + 905, + 829 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "To address this, we decompose $I_{i}$ into two components, one in the direction of $e_{i}$ and the other in the orthogonal direction:", + "bbox": [ + 511, + 830, + 905, + 872 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\nI _ {i} ^ {\\parallel} = \\frac {I _ {i} \\cdot e _ {i}}{\\| e _ {i} \\| ^ {2}} \\cdot e _ {i}, \\tag {18}\n$$\n", + "text_format": "latex", + "bbox": [ + 651, + 872, + 903, + 905 + ], + "page_idx": 4 + }, + { + "type": "page_number", + "text": "9563", + "bbox": [ + 480, + 944, + 514, + 955 + ], + "page_idx": 4 + }, + { + "type": "table", + "img_path": "images/c25e209c638f443ef734a070fe2fa790b179a7d5f7e2fbb063515f8533920367.jpg", + "table_caption": [ + "Table 1. Comparison with the state-of-the-art methods on the five domain generalization datasets from DomainBed benchmark. We report the averaged leave-one-domain-out experimental results with ViT-B/16 as the backbone. Numbers in bold represent the best performance among all the methods. T and F indicate whether the CLIP parameters are trained or frozen during adaptation to DG." + ], + "table_footnote": [], + "table_body": "
MethodsCLIPOHTerraIncVLCSPACSDomainNetAvg
Fine-tuning
SWAD (NeurIPS'2021) [6]T81.0142.9279.1391.3557.9270.47
LP-FT (ICLR'2022) [30]T81.1747.2680.8892.9257.0471.85
WiSE-FT (CVPR'2022) [60]T86.3254.5082.8897.2958.0175.80
MIRO (ECCV'2022) [7]T84.8059.3082.3096.4460.4776.66
DART (CVPR'2023) [27]T80.9351.2480.3893.4359.3273.06
SAGM (CVPR'2023) [57]T83.4058.6482.0594.3159.0575.49
FLYP (CVPR'2023) [20]T82.7633.2566.6478.5357.4163.72
CLIPood (ICML'2023) [50]T83.3146.2877.1993.1657.7871.55
Knowledge Distillation
RISE (ICCV'2023) [26]T78.3949.6180.6293.2555.3771.45
VL2V-SD (CVPR'2024) [1]T87.3858.5483.2596.6862.7977.73
Prompt-learning
DPL (TJSAI'2023) [62]F84.2052.6084.3097.3059.5075.60
SPG (ECCV'2024) [4]F83.6050.2082.4097.0060.1074.70
Parameter-efficient Fine-tuning
LPD (ECCV'2024) [23]F84.2157.3083.2096.5559.3076.11
OursT87.6563.2784.7997.0363.1179.17
", + "bbox": [ + 112, + 141, + 885, + 429 + ], + "page_idx": 5 + }, + { + "type": "equation", + "text": "\n$$\nI _ {i} ^ {\\perp} = I _ {i} - I _ {i} ^ {\\parallel}, \\tag {19}\n$$\n", + "text_format": "latex", + "bbox": [ + 236, + 464, + 480, + 484 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "where $I_{i}^{\\parallel}$ and $I_{i}^{\\perp}$ are the decomposed parallel and orthogonal components, respectively. We define a pairwise uniformity loss based on the dot product between the orthogonal components of different image features within the batch:", + "bbox": [ + 89, + 489, + 483, + 554 + ], + "page_idx": 5 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {L} _ {\\mathrm {O U}} = \\frac {1}{B (B - 1)} \\sum_ {i = 1} ^ {B} \\sum_ {\\substack {j = 1 \\\\ j \\neq i}} ^ {B} \\left(I _ {i} ^ {\\perp} \\cdot I _ {j} ^ {\\perp}\\right) ^ {2} \\tag{20}\n$$\n", + "text_format": "latex", + "bbox": [ + 161, + 560, + 482, + 613 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Since $e_i$ is computed independently for each image, the orthogonal component is also based on different images. Applying the uniformity constraint solely on the orthogonal direction avoids interference with the similarity calculation between the image features and the prompt.", + "bbox": [ + 89, + 619, + 483, + 696 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Finally, combining the baseline with Domain-sensitive Channel Suppression, Multi-domain Consistency Constraint, and Orthogonal Uniformity Regularization yields the total loss used during training, which is formulated as:", + "bbox": [ + 89, + 696, + 483, + 757 + ], + "page_idx": 5 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {L} _ {\\text {t o t a l}} = \\mathcal {L} _ {K D} + \\mathcal {L} _ {\\text {r e g}} + \\mathcal {L} _ {K L} + \\mathcal {L} _ {O U}, \\tag {21}\n$$\n", + "text_format": "latex", + "bbox": [ + 155, + 766, + 482, + 782 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "5. Experiments", + "text_level": 1, + "bbox": [ + 89, + 792, + 223, + 809 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "5.1. Datasets and Evaluation", + "text_level": 1, + "bbox": [ + 89, + 816, + 313, + 832 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "We conduct evaluation experiments on five datasets from the DomainBed benchmark [21], following previous methods [1, 22, 26]. PACS [32] consists of 9,991 samples spanning four domains: {art, cartoons, photos, sketches}, and", + "bbox": [ + 89, + 839, + 483, + 901 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "includes a total of 7 classes. VLCS [17] contains 10,729 examples, 5 classes from four datasets as different domains: {VOC2007, LabelMe, Caltech, SUN}. Terra Incognita (TI) [5] contains 24,788 samples from 10 animal categories. These photos are captured across different locations {L100, L38, L43, L46}, acting as the four domains. Office Home (OH) [55] have four domains {art, clipart, product, real}, with 15,588 examples from 65 classes. DomainNet [44] is a large-scale cross-domain dataset containing 586,575 images from 6 different domains: {clipart, infographic, painting, quickdraw, real, sketch}. It covers 345 classes and is widely used to evaluate the performance of domain generalization algorithms.", + "bbox": [ + 511, + 452, + 906, + 648 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Following the protocols from previous domain generalization works [1, 11, 26], we split the datasets by using $80\\%$ of the data for training and the remaining $20\\%$ as the validation set. We adopt a leave-one-out experiment setup, where we alternately leave on domain for testing and train with the remaining domains. The average performance on the unseen domains is reported as the domain generalization results.", + "bbox": [ + 511, + 651, + 908, + 771 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "5.2. Implementation Details", + "text_level": 1, + "bbox": [ + 511, + 785, + 730, + 801 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "We employ CLIP with ViT-B/16 as the backbone model and fine-tune it on each of the datasets. We fine-tune the model for 5,000 iterations on the PACS, VLCS, Office Home, and Terra Incognito datasets, and for 15,000 iterations on the DomainNet dataset. We use the $n = 80$ templates recommended in original CLIP [49] as prompts to imitate dif", + "bbox": [ + 511, + 809, + 906, + 901 + ], + "page_idx": 5 + }, + { + "type": "page_number", + "text": "9564", + "bbox": [ + 480, + 944, + 514, + 955 + ], + "page_idx": 5 + }, + { + "type": "table", + "img_path": "images/b0583a62be85f8c9dc3b8d567f93ba6b7f701503acaafe1098f8171b80bdac21.jpg", + "table_caption": [ + "Table 2. Ablation analysis on the effectiveness of our method. EPS: entropy-based prompt selection, DCS: domain-sensitive channel suppression, MCC: multi-domain consistency constraint. OUR: orthogonal uniformity regularization, URIF*: uniformity regularization for image features, where * indicates it is not our component, but used as an additional comparison to OUR." + ], + "table_footnote": [], + "table_body": "
BaseEPSDCSMCCOURURIF*Avg
✓81.46
✓✓✓82.28
✓✓✓82.37
✓✓✓✓82.79
✓✓82.30
✓✓81.55
✓✓✓✓✓83.19
", + "bbox": [ + 91, + 183, + 485, + 325 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "ferent domains for each class. The threshold $\\beta$ for the suppressed feature mask $\\mathbf{M}_D$ is set dynamically for each sample. We suppress $20\\%$ image feature channels to reduce domain-specific perturbation, and set the regularization term as $10^{-4}$ . We select $m = 40$ prompts with low entropy from 80 contextual prompts. Other experimental settings all follow our baseline model. SWAD [6] is a general approach that improves model robustness by averaging weights during fine-tuning DG models. For a fair comparison, we integrate SWAD into all methods that fine-tune CLIP's parameters, following the previous work [1].", + "bbox": [ + 88, + 333, + 485, + 501 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "5.3. Comparison with State-of-the-Arts", + "text_level": 1, + "bbox": [ + 89, + 512, + 395, + 527 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "We mainly compare with recent advances in domain generalization methods based on CLIP. We first compares our method with the following fine-tuning methods. SWAD [6], WiSE-FT [60], and DART [27] explore weight-ensembling methods to maintain CLIP's generalization ability. LPFT [30], FLYP [20], and CLIPood [50] investigate effective fine-tuning methods. The importance of text is highlighted in VL2V-SD [1] and RISE [26]. They incorporate text features into the DG model via knowledge distillation. DPL [62] and SPG [4] are prompt-learning methods that design learnable modules to extract context features from input images. LPD [23] explores a parameter-efficient fine-tuning approach for adapting CLIP in DG. We report the averaged leave-one-domain-out experimental results following the setting from previous works [1, 4].", + "bbox": [ + 91, + 537, + 482, + 763 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "As shown in Table 1, our approach achieves an average improvement of $1.4\\%$ over SOTA methods across five datasets. Our method demonstrates a particularly notable $4.7\\%$ improvement over SOTA on the TerraInc dataset [5]. As the images in TerraInc are sourced from wildlife photography, target objects occupy a small portion of the scene and contain substantial irrelevant contextual information. Our approach effectively reduces interference from such perturbation and achieves impressive improvement.", + "bbox": [ + 89, + 765, + 485, + 902 + ], + "page_idx": 6 + }, + { + "type": "table", + "img_path": "images/ec6e5f34a6ef29ab5aabf926ec7f493c8d4bfa40d8851cabfcc7225af15ad57a.jpg", + "table_caption": [ + "Table 3. Performance comparison of our method with various prompt learning approaches across different datasets, including OH, TI, VLCS, and PACS. The average accuracy (Avg) across these datasets highlights the robustness and superior generalization capability of template generated prompts." + ], + "table_footnote": [], + "table_body": "
MethodsOHTIVLCSPACSAvg
CLIP [49]82.434.182.396.573.8
CoOP [65]83.046.880.896.476.8
Co-CoOP [64]83.445.380.396.776.4
Maple [29]83.450.282.296.578.1
DPL [62]84.252.684.397.379.6
SPG [4]83.650.282.497.078.3
Ours87.763.384.897.083.2
", + "bbox": [ + 514, + 171, + 911, + 301 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "5.4. Ablation Study", + "text_level": 1, + "bbox": [ + 511, + 315, + 666, + 330 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Ablation of the main components. We conduct ablation studies to evaluate the effectiveness of each component of our approach. Following previous methods [1, 23], we report the results on four datasets: OH [55], TI [5], VLCS [17], and PACS [32]. 'Avg' denotes the average of the four datasets. As shown in Table 2, using DCS with EPS yields a notable performance increase and improves the accuracy to $82.28\\%$ , highlighting the significance of suppressing the domain-specific channels. MCC also achieves an improvement, raising accuracy to $82.37\\%$ by enforcing cross-domain consistency. Applying both DSC and MCC modules simultaneously to suppress the domain-specific context results in an average gain of $1.33\\%$ .", + "bbox": [ + 511, + 339, + 906, + 536 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Moreover, incorporating uniformity regularization for the orthogonal component increases the average accuracy to $82.30\\%$ , suggesting that capturing more diversified features enables the model to generalize better across domains by preventing an overemphasis on specific source domains. In contrast, applying uniformity regularization directly to image features (URIF*) without orthogonal decomposition yields a decrease to $81.55\\%$ , demonstrating that directly regularizing image features may introduce perturbations during classification. Finally, the best performance is obtained by combining EPS, DSC, MCR, and OUR, reaching an average accuracy of $83.19\\%$ .", + "bbox": [ + 509, + 537, + 908, + 718 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Discussions on the design of prompts. In our implementation, we imitate diverse domains by the prompts generated from 80 templates recommended by CLIP. In this section, we conduct experiments to analysis the advantages of our template-based prompts over the learnable prompts used in previous methods [4, 29, 62, 64]. By optimizing the learnable prompts during training, these model learn to adapt dynamically to different inputs. As shown in Table 3, although these methods show significant improvement compared to CLIP, our approach still outperforms the state-of-the-art prompt learning methods by $3.6\\%$ . While they incorporate context of the image, the learnable prompts also", + "bbox": [ + 509, + 719, + 910, + 902 + ], + "page_idx": 6 + }, + { + "type": "page_number", + "text": "9565", + "bbox": [ + 480, + 944, + 514, + 955 + ], + "page_idx": 6 + }, + { + "type": "table", + "img_path": "images/078f735ab9058108c7d086458b84f5c22407515071a1ca644ef04ec18c97d4e1.jpg", + "table_caption": [ + "Table 4. Ablations of comparison with knowledge distillation from the single manually designed prompt and the design of our method. EP represents the ensembled $n$ prompts for knowledge distillation." + ], + "table_footnote": [], + "table_body": "
PromptsOHTIVLCSPACSAvg
{class}86.559.583.796.681.6
a photo of a {class}87.458.583.396.781.5
a {domain} of a {class}87.260.783.996.982.2
EP87.159.983.696.581.9
Ours87.763.384.897.083.2
", + "bbox": [ + 91, + 156, + 483, + 268 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "introduces potential perturbations from specific source domains. Unlike these learnable prompts, our method leverages a set of templates without requiring domain-specific adaptation during training, enabling the model to generalize more robustly across diverse domains.", + "bbox": [ + 89, + 292, + 482, + 367 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Analysis on knowledge distillation from different designed prompts. In this part, we conduct experiments to demonstrate the advantages of using a diverse set of prompts compared to the carefully designed manual prompt. In our approach, the diverse set of prompts are ensembled and a knowledge distillation framework is used to transfer the ensembled prompts (EP) to the DG model. Based on the knowledge distillation baseline, we compare with three types of manually designed prompts, including {\"class)\", \"a photo of a {class}\", and \"a {domain}\" of a {class)\". As shown in Table 4, a simple use of ensembled prompts provides slight gains, with $0.3\\%$ and $0.4\\%$ improvements compared to manual prompts {\"class\"} and \"a photo of a {class}\". Moreover, EP even performs worse than the carefully designed prompt: \"a {domain}\" of a {class)\". However, a set of diverse prompts enables us to imitate multiple domains and learn domain-invariant features. Our overall method achieves consistently better performance across all four datasets, with notable gains in challenging cases like TI and VLCS. Such results demonstrate the effectiveness of our multi-domain imitation method.", + "bbox": [ + 89, + 369, + 482, + 686 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Hyperparameter analysis. Additionally, we investigate the hyperparameter for the number of selected templates $m$ and the ratio of channels suppressed during gradient attribution. Table 5 presents analysis on the number of prompts and its impact on the average performance across four datasets. We evaluate different values for $m$ , ranging from 10 to 80. The results show that increasing $m$ improves performance at first, when $m = 40$ , the performance reaches its maximum value $83.2\\%$ . Using more templates ensures that the imitated domains contain more complex contexts, thus facilitating our model in learning robust features. However, further increases $m$ reduce the performance, with the average accuracy dropping to $83.0\\%$ at $m = 60$ and $82.3\\%$ at $m = 80$ . Since we do not design", + "bbox": [ + 89, + 688, + 482, + 900 + ], + "page_idx": 7 + }, + { + "type": "table", + "img_path": "images/0b56d6081b2c6d415ffb5fcb2cd5baffc64fd4c578d510f1bf27628575a257d7.jpg", + "table_caption": [ + "Table 5. Hyperparameter analysis on the number of prompts $m$ ." + ], + "table_footnote": [], + "table_body": "
m1020406080
Avg82.782.983.283.082.3
", + "bbox": [ + 524, + 114, + 895, + 159 + ], + "page_idx": 7 + }, + { + "type": "table", + "img_path": "images/89503814643621f1f40377a0c804bbcb7359e2c4ce321752048e4711a8a11e21.jpg", + "table_caption": [ + "Table 6. Hyperparameter analysis on the ratio of channels suppressed during gradient attribution." + ], + "table_footnote": [], + "table_body": "
Ratio0.10.20.30.40.5
Avg82.383.282.081.781.6
", + "bbox": [ + 524, + 208, + 895, + 252 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "different templates for different datasets, utilizing all of the templates without selection introduces noise, leading to a drop in performance. Therefore, we set the hyperparameter $m$ as 40.", + "bbox": [ + 511, + 267, + 903, + 325 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Table 6 presents the analysis on the ratio of channels suppressed during gradient attribution and its effect on the overall performance across four datasets. We evaluate different suppression ratios, ranging from 0.1 to 0.5. The results show that increasing the suppression ratio improves performance at first, with the maximum average accuracy of $83.2\\%$ achieved at a suppression ratio of 0.2. However, further increases in the suppression ratio slightly reduce accuracy, with the average performance dropping to $82.0\\%$ , $81.7\\%$ , and $81.6\\%$ at 0.3, 0.4, and 0.5. The experiments demonstrate that excessive suppression filters out essential information, resulting in degraded performance. Thus, we set the suppression ratio to 0.2 as the optimal value for this hyperparameter.", + "bbox": [ + 511, + 329, + 906, + 541 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "6. Conclusion", + "text_level": 1, + "bbox": [ + 511, + 558, + 633, + 574 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "We develop a new domain generalization approach by learning with diverse text prompts. We use these prompts to imitate multiple domains, which guide the DG model in learning domain-invariant features through multi-modal alignment in CLIP. These prompts contribute to three aspects, including suppressing domain-sensitive features, ensure feature consistency with varied domains, and enable feature diversification in the DG model. By incorporating these three terms, our approach reduces the CLIP-based DG model's tendency to capture domain-sensitive details and mitigates bias towards the source data distribution. Experimental results on five domain generalization datasets from the DomainBed benchmark demonstrate the effectiveness of our method.", + "bbox": [ + 511, + 584, + 906, + 795 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "7. Acknowledgments", + "text_level": 1, + "bbox": [ + 511, + 813, + 692, + 830 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "This work was supported by the NSFC under Grant 62322604 and 62176159, and in part by the Shanghai Municipal Science and Technology Major Project under Grant 2021SHZDZX0102.", + "bbox": [ + 511, + 839, + 905, + 898 + ], + "page_idx": 7 + }, + { + "type": "page_number", + "text": "9566", + "bbox": [ + 480, + 945, + 514, + 955 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "References", + "text_level": 1, + "bbox": [ + 91, + 89, + 187, + 104 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[1] Sravanti Addepalli, Ashish Ramayee Asokan, Lakshay Sharma, and R Venkatesh Babu. Leveraging vision-language models for improving domain generalization in image classification. In CVPR, pages 23922-23932, 2024. 1, 2, 3, 6, 7", + "[2] James Urquhart Allingham, Jie Ren, Michael W Dusenberry, Xiuye Gu, Yin Cui, Dustin Tran, Jeremiah Zhe Liu, and Balaji Lakshminarayanan. A simple zero-shot prompt weighting technique to improve prompt ensembling in text-image models. In ICML, 2023. 4", + "[3] Martin Arjovsky, Léon Bottou, Ishaan Gulrajani, and David Lopez-Paz. Invariant risk minimization. arXiv preprint arXiv:1907.02893, 2019. 2", + "[4] Shuanghao Bai, Yuedi Zhang, Wanqi Zhou, Zhirong Luan, and Badong Chen. Soft prompt generation for domain generalization. In ECCV, pages 434-450, 2024. 3, 6, 7", + "[5] Sara Beery, Grant Van Horn, and Pietro Perona. Recognition in terra incognita. In ECCV, pages 456-473, 2018. 6, 7", + "[6] Junbum Cha, Sanghyuk Chun, Kyungjae Lee, Han-Cheol Cho, Seunghyun Park, Yunsung Lee, and Sungrae Park. Swad: Domain generalization by seeking flat minima. NeurIPS, 34:22405-22418, 2021. 1, 2, 6, 7", + "[7] Junbum Cha, Kyungjae Lee, Sungrae Park, and Sanghyuk Chun. Domain generalization by mutual-information regularization with pre-trained models. In ECCV, pages 440-457, 2022. 1, 6", + "[8] Liang Chen, Yong Zhang, Yibing Song, Ying Shan, and Lingqiao Liu. Improved test-time adaptation for domain generalization. In CVPR, pages 24172-24182, 2023. 2", + "[9] Liang Chen, Yong Zhang, Yibing Song, Anton Van Den Hengel, and Lingqiao Liu. Domain generalization via rationale invariance. In ICCV, pages 1751-1760, 2023. 1", + "[10] De Cheng, Zhipeng Xu, Xinyang Jiang, Nannan Wang, Dongsheng Li, and Xinbo Gao. Disentangled prompt representation for domain generalization. In CVPR, pages 23595-23604, 2024. 1, 2", + "[11] Junhyeong Cho, Gilhyun Nam, Sungyeon Kim, Hunmin Yang, and Suha Kwak. Promptstyler: Prompt-driven style generation for source-free domain generalization. In ICCV, pages 15702-15712, 2023. 2, 6", + "[12] Huiqi Deng, Na Zou, Mengnan Du, Weifu Chen, Guocan Feng, Ziwei Yang, Zheyang Li, and Quanshi Zhang. Unifying fourteen post-hoc attribution methods with taylor interactions. IEEE Transactions on Pattern Analysis and Machine Intelligence, 46(7):4625-4640, 2024. 4", + "[13] Yu Ding, Lei Wang, Bin Liang, Shuming Liang, Yang Wang, and Fang Chen. Domain generalization by learning and removing domain-specific features. In NeurIPS, pages 24226-24239, 2022. 2", + "[14] Xiaoyi Dong, Jianmin Bao, Yinglin Zheng, Ting Zhang, Dongdong Chen, Hao Yang, Ming Zeng, Weiming Zhang, Lu Yuan, Dong Chen, et al. Maskclip: Masked self-distillation advances contrastive language-image pretraining. In CVPR, pages 10995-11005, 2023. 2", + "[15] Alexey Dosovitskiy, Lucas Beyer, Alexander Kolesnikov, Dirk Weissenborn, Xiaohua Zhai, Thomas Unterthiner," + ], + "bbox": [ + 93, + 114, + 483, + 900 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "Mostafa Dehghani, Matthias Minderer, Georg Heigold, Sylvain Gelly, Jakob Uszkoreit, and Neil Houlsby. An image is worth 16x16 words: Transformers for image recognition at scale. In ICLR, 2021. 3", + "[16] Lisa Dunlap, Clara Mohri, Devin Guillory, Han Zhang, Trevor Darrell, Joseph E Gonzalez, Aditi Raghunathan, and Anna Rohrbach. Using language to extend to unseen domains. In ICLR, 2023. 2", + "[17] Chen Fang, Ye Xu, and Daniel N Rockmore. Unbiased metric learning: On the utilization of multiple datasets and web images for softening bias. In ICCV, pages 1657-1664, 2013. 6, 7", + "[18] Yaroslav Ganin, Evgeniya Ustinova, Hana Ajakan, Pascal Germain, Hugo Larochelle, François Laviolette, Mario March, and Victor Lempitsky. Domain-adversarial training of neural networks. Journal of machine learning research, 17(59):1-35, 2016. 2", + "[19] Chunjiang Ge, Rui Huang, Mixue Xie, Zihang Lai, Shiji Song, Shuang Li, and Gao Huang. Domain adaptation via prompt learning. IEEE Transactions on Neural Networks and Learning Systems, pages 1-11, 2023. 3", + "[20] Sachin Goyal, Ananya Kumar, Sankalp Garg, Zico Kolter, and Aditi Raghunathan. Finetune like you pretrain: Improved finetuning of zero-shot vision models. In CVPR, pages 19338-19347, 2023. 1, 2, 6, 7", + "[21] Ishaan Gulrajani and David Lopez-Paz. In search of lost domain generalization. In ICLR, 2020. 2, 6", + "[22] Jintao Guo, Lei Qi, and Yinghuan Shi. Domaindrop: Suppressing domain-sensitive channels for domain generalization. In ICCV, pages 19114-19124, 2023. 2, 6", + "[23] Jiajun Hu, Jian Zhang, Lei Qi, Yinghuan Shi, and Yang Gao. Learn to preserve and diversify: Parameter-efficient group with orthogonal regularization for domain generalization. In ECCV, pages 198-216, 2024. 1, 2, 3, 6, 7", + "[24] Lanqing Hu, Meina Kan, Shiguang Shan, and Xilin Chen. Dandelionnet: Domain composition with instance adaptive classification for domain generalization. In ICCV, pages 19050-19059, 2023. 1", + "[25] Zenan Huang, Haobo Wang, Junbo Zhao, and Nenggan Zheng. idag: Invariant dag searching for domain generalization. In ICCV, pages 19169-19179, 2023. 1", + "[26] Zeyi Huang, Andy Zhou, Zijian Ling, Mu Cai, Haohan Wang, and Yong Jae Lee. A sentence speaks a thousand images: Domain generalization through distilling clip with language guidance. In ICCV, pages 11685-11695, 2023. 1, 2, 3, 6, 7", + "[27] Samyak Jain, Sravanti Addepalli, Pawan Kumar Sahu, Priyam Dey, and R Venkatesh Babu. Dart: Diversify-aggregate-repeat training improves generalization of neural networks. In CVPR, pages 16048-16059, 2023. 1, 6, 7", + "[28] Chao Jia, Yinfei Yang, Ye Xia, Yi-Ting Chen, Zarana Parekh, Hieu Pham, Quoc Le, Yun-Hsuan Sung, Zhen Li, and Tom Duerig. Scaling up visual and vision-language representation learning with noisy text supervision. In ICML, pages 4904-4916, 2021. 2", + "[29] Muhammad Uzair Khattak, Hanoona Rasheed, Muhammad Maaz, Salman Khan, and Fahad Shahbaz Khan. Maple:" + ], + "bbox": [ + 516, + 92, + 903, + 900 + ], + "page_idx": 8 + }, + { + "type": "page_number", + "text": "9567", + "bbox": [ + 482, + 944, + 514, + 955 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "Multi-modal prompt learning. In CVPR, pages 19113-19122, 2023. 2, 3, 7", + "[30] Ananya Kumar, Aditi Raghunathan, Robbie Jones, Tengyu Ma, and Percy Liang. Fine-tuning can distort pretrained features and underperform out-of-distribution. In ICLR, 2022. 2, 6, 7", + "[31] Chenming Li, Daoan Zhang, Wenjian Huang, and Jianguo Zhang. Cross contrasting feature perturbation for domain generalization. In ICCV, pages 1327-1337, 2023. 2", + "[32] Da Li, Yongxin Yang, Yi-Zhe Song, and Timothy M Hospedales. Deeper, broader and artier domain generalization. In ICCV, pages 5542-5550, 2017. 6, 7", + "[33] Haoliang Li, Sinno Jialin Pan, Shiqi Wang, and Alex C Kot. Domain generalization with adversarial feature learning. In CVPR, pages 5400-5409, 2018. 2", + "[34] Junnan Li, Dongxu Li, Caiming Xiong, and Steven Hoi. Blip: Bootstrapping language-image pre-training for unified vision-language understanding and generation. In ICML, pages 12888-12900, 2022. 1, 2", + "[35] Junnan Li, Dongxu Li, Silvio Savarese, and Steven Hoi. Blip-2: Bootstrapping language-image pre-training with frozen image encoders and large language models. In ICML, pages 19730–19742, 2023. 1, 2", + "[36] Xiaotong Li, Yongxing Dai, Yixiao Ge, Jun Liu, Ying Shan, and Ling-Yu Duan. Uncertainty modeling for out-of-distribution generalization. In ICLR, 2022. 2", + "[37] Ya Li, Mingming Gong, Xinmei Tian, Tongliang Liu, and Dacheng Tao. Domain generalization via conditional invariant representations. In AAAI, 2018. 2", + "[38] Ya Li, Xinmei Tian, Mingming Gong, Yajing Liu, Tongliang Liu, Kun Zhang, and Dacheng Tao. Deep domain generalization via conditional invariant adversarial networks. In ECCV, pages 624-639, 2018. 2", + "[39] Krikamol Muandet, David Balduzzi, and Bernhard Schölkopf. Domain generalization via invariant feature representation. In ICML, pages 10-18, 2013. 2", + "[40] Hyeonseob Nam, HyunJae Lee, Jongchan Park, Wonjun Yoon, and Donggeun Yoo. Reducing domain gap by reducing style bias. In CVPR, pages 8690-8699, 2021. 2", + "[41] Hongjing Niu, Hanting Li, Feng Zhao, and Bin Li. Domain-unified prompt representations for source-free domain generalization. arXiv preprint arXiv:2209.14926, 2022. 2", + "[42] Oren Nuriel, Sagie Benaim, and Lior Wolf. Permuted adain: Reducing the bias towards global statistics in image classification. In CVPR, pages 9482-9491, 2021. 2", + "[43] Changdae Oh, Junhyuk So, Hoyoon Byun, YongTaek Lim, Minchul Shin, Jong-June Jeon, and Kyungwoo Song. Geodesic multi-modal mixup for robust fine-tuning. In NeurIPS, 2024. 1", + "[44] Xingchao Peng, Qinxun Bai, Xide Xia, Zijun Huang, Kate Saenko, and Bo Wang. Moment matching for multi-source domain adaptation. In ICCV, pages 1406-1415, 2019. 6", + "[45] Zelin Peng, Zhengqin Xu, Zhilin Zeng, Lingxi Xie, Qi Tian, and Wei Shen. Parameter efficient fine-tuning via cross block orchestration for segment anything model. In CVPR, pages 3743-3752, 2024. 2" + ], + "bbox": [ + 91, + 90, + 482, + 898 + ], + "page_idx": 9 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[46] Shi Pu, Kaili Zhao, and Mao Zheng. Alignment-uniformity aware representation learning for zero-shot video classification. In CVPR, pages 19968-19977, 2022. 1, 5", + "[47] Daiqing Qi, Handong Zhao, Aidong Zhang, and Sheng Li. Generalizing to unseen domains via text-guided augmentation. In ECCV, pages 285-300, 2024. 2", + "[48] Minghan Qin, Wanhua Li, Jiawei Zhou, Haoqian Wang, and Hanspeter Pfister. Langsplat: 3d language gaussian splatting. In CVPR, pages 20051-20060, 2024. 2", + "[49] Alec Radford, Jong Wook Kim, Chris Hallacy, Aditya Ramesh, Gabriel Goh, Sandhini Agarwal, Girish Sastry, Amanda Askell, Pamela Mishkin, Jack Clark, et al. Learning transferable visual models from natural language supervision. In ICML, pages 8748-8763, 2021. 1, 2, 3, 6, 7", + "[50] Yang Shu, Xingzhuo Guo, Jialong Wu, Ximei Wang, Jianmin Wang, and Mingsheng Long. Clipood: Generalizing clip to out-of-distributions. In ICML, pages 31716-31731, 2023. 1, 2, 6, 7", + "[51] Mainak Singha, Ankit Jha, Shirsha Bose, Ashwin Nair, Moloud Abdar, and Biplab Banerjee. Unknown prompt the only lacuna: Unveiling clip's potential for open domain generalization. In CVPR, pages 13309-13319, 2024. 1", + "[52] Quan Sun, Yuxin Fang, Ledell Wu, Xinlong Wang, and Yue Cao. Eva-clip: Improved training techniques for clip at scale. arXiv preprint arXiv:2303.15389, 2023. 1, 2", + "[53] Vladimir N Vapnik. An overview of statistical learning theory. IEEE transactions on neural networks, 10(5):988-999, 1999. 2", + "[54] Ashish Vaswani, Noam Shazeer, Niki Parmar, Jakob Uszkoreit, Llion Jones, Aidan N Gomez, Lukasz Kaiser, and Illia Polosukhin. Attention is all you need. In NeurIPS, pages 5998-6008, 2017. 3", + "[55] Hemanth Venkateswara, Jose Eusebio, Shayok Chakraborty, and Sethuraman Panchanathan. Deep hashing network for unsupervised domain adaptation. In CVPR, pages 5018-5027, 2017. 6, 7", + "[56] Mengzhu Wang, Yuehua Liu, Jianlong Yuan, Shanshan Wang, Zhibin Wang, and Wei Wang. Inter-class and inter-domain semantic augmentation for domain generalization. IEEE Transactions on Image Processing, 33:1338-1347, 2024. 2", + "[57] Pengfei Wang, Zhaoxiang Zhang, Zhen Lei, and Lei Zhang. Sharpness-aware gradient matching for domain generalization. In CVPR, pages 3769-3778, 2023. 6", + "[58] Tongzhou Wang and Phillip Isola. Understanding contrastive representation learning through alignment and uniformity on the hypersphere. In ICML, pages 9929-9939, 2020. 1", + "[59] Yufei Wang, Haoliang Li, and Alex C Kot. Heterogeneous domain generalization via domain mixup. In ICASSP, pages 3622-3626, 2020. 2", + "[60] Mitchell Wortsman, Gabriel Ilharco, Jong Wook Kim, Mike Li, Simon Kornblith, Rebecca Roelofs, Raphael Gontijo Lopes, Hannaneh Hajishirzi, Ali Farhadi, Hongseok Namkoong, et al. Robust fine-tuning of zero-shot models. In CVPR, pages 7959-7971, 2022. 1, 2, 6, 7", + "[61] Sitong Wu, Haoru Tan, Zhuotao Tian, Yukang Chen, Xiaojuan Qi, and Jiaya Jia. Saco loss: Sample-wise affinity con" + ], + "bbox": [ + 516, + 90, + 903, + 902 + ], + "page_idx": 9 + }, + { + "type": "page_number", + "text": "9568", + "bbox": [ + 482, + 945, + 514, + 955 + ], + "page_idx": 9 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "sistency for vision-language pre-training. In CVPR, pages 27358-27369, 2024. 1, 2", + "[62] Xin Zhang, Shixiang Shane Gu, Yutaka Matsuo, and Yusuke Iwasawa. Domain prompt learning for efficiently adapting clip to unseen domains. Transactions of the Japanese Society for Artificial Intelligence, 38(6), 2023. 3, 6, 7", + "[63] Xingxuan Zhang, Renzhe Xu, Han Yu, Yancheng Dong, Pengfei Tian, and Peng Cui. Flatness-aware minimization for domain generalization. In ICCV, pages 5189-5202, 2023. 2", + "[64] Kaiyang Zhou, Jingkang Yang, Chen Change Loy, and Zwei Liu. Conditional prompt learning for vision-language models. In CVPR, pages 16816-16825, 2022. 2, 3, 7", + "[65] Kaiyang Zhou, Jingkang Yang, Chen Change Loy, and Ziwei Liu. Learning to prompt for vision-language models. International Journal of Computer Vision, 130(9):2337-2348, 2022. 2, 3, 7" + ], + "bbox": [ + 91, + 90, + 482, + 316 + ], + "page_idx": 10 + }, + { + "type": "page_number", + "text": "9569", + "bbox": [ + 482, + 945, + 514, + 955 + ], + "page_idx": 10 + } +] \ No newline at end of file diff --git a/2025/Domain Generalization in CLIP via Learning with Diverse Text Prompts/51839507-6fb2-45c2-a66c-0644bd7936ed_model.json b/2025/Domain Generalization in CLIP via Learning with Diverse Text Prompts/51839507-6fb2-45c2-a66c-0644bd7936ed_model.json new file mode 100644 index 0000000000000000000000000000000000000000..b0ca74a666f7b7b6462573991eed6c58f92737f9 --- /dev/null +++ b/2025/Domain Generalization in CLIP via Learning with Diverse Text Prompts/51839507-6fb2-45c2-a66c-0644bd7936ed_model.json @@ -0,0 +1,2389 @@ +[ + [ + { + "type": "header", + "bbox": [ + 0.107, + 0.003, + 0.182, + 0.044 + ], + "angle": 0, + "content": "CVF" + }, + { + "type": "header", + "bbox": [ + 0.237, + 0.001, + 0.812, + 0.047 + ], + "angle": 0, + "content": "This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore." + }, + { + "type": "title", + "bbox": [ + 0.134, + 0.131, + 0.867, + 0.154 + ], + "angle": 0, + "content": "Domain Generalization in CLIP via Learning with Diverse Text Prompts" + }, + { + "type": "text", + "bbox": [ + 0.168, + 0.181, + 0.831, + 0.218 + ], + "angle": 0, + "content": "Changsong Wen, Zelin Peng, Yu Huang, Xiaokang Yang, Wei Shen MoE Key Lab of Artificial Intelligence, AI Institute, Shanghai Jiao Tong University" + }, + { + "type": "text", + "bbox": [ + 0.21, + 0.22, + 0.785, + 0.235 + ], + "angle": 0, + "content": "{changsong, zelin.peng, yellowfish, xkyang, wei.shen}@sjtu.edu.cn" + }, + { + "type": "title", + "bbox": [ + 0.248, + 0.269, + 0.329, + 0.285 + ], + "angle": 0, + "content": "Abstract" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.302, + 0.486, + 0.664 + ], + "angle": 0, + "content": "Domain generalization (DG) aims to train a model on source domains that can generalize well to unseen domains. Recent advances in Vision-Language Models (VLMs), such as CLIP, exhibit remarkable generalization capabilities across a wide range of data distributions, benefiting tasks like DG. However, CLIP is pre-trained by aligning images with their descriptions, which inevitably captures domain-specific details. Moreover, adapting CLIP to source domains with limited feature diversity introduces bias. These limitations hinder the model's ability to generalize across domains. In this paper, we propose a new DG approach by learning with diverse text prompts. These text prompts incorporate varied contexts to imitate different domains, enabling DG model to learn domain-invariant features. The text prompts guide DG model learning in three aspects: feature suppression, which uses these prompts to identify domain-sensitive features and suppress them; feature consistency, which ensures the model's features are robust to domain variations imitated by the diverse prompts; and feature diversification, which diversifies features based on the prompts to mitigate bias. Experimental results show that our approach improves domain generalization performance on five datasets from the DomainBed benchmark, achieving state-of-the-art results." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.697, + 0.222, + 0.713 + ], + "angle": 0, + "content": "1. Introduction" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.723, + 0.485, + 0.875 + ], + "angle": 0, + "content": "In the complex real-world applications, deep learning models are expected to be capable of recognizing data from a wide range of distributions. Within the past few years, researchers are dedicated to investigating domain generalization (DG) [6, 7, 9, 24, 25, 27], which empowers the model to learn robust features from multiple domains and exhibit good performance across various data distributions. Recent advances in foundation Vision-Language Models (VLMs) [34, 35, 52, 61], such as CLIP [49], are trained on large-scale data sourced from the web with a wide variety" + }, + { + "type": "image", + "bbox": [ + 0.537, + 0.295, + 0.907, + 0.466 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.512, + 0.503, + 0.907, + 0.6 + ], + "angle": 0, + "content": "Figure 1. Limitations with existing methods. As shown on the left, CLIP's image encoder is pre-trained to align with textual descriptions, which inevitably captures domain-specific visual details, resulting in perturbations in domain generalization. Additionally, as shown on the right, due to the limited feature diversity in the source domains, adapting CLIP with source domains will introduce bias." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.628, + 0.907, + 0.688 + ], + "angle": 0, + "content": "of data distributions. These VLMs possess remarkable generalization capabilities, and boost many open-world recognition tasks, including out-of-distribution learning [50] and domain generalization [1, 10, 51]." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.69, + 0.909, + 0.901 + ], + "angle": 0, + "content": "Existing DG studies [20, 23, 27, 50, 60] leverage CLIP's pre-trained weights with techniques like parameter regularization [23] and weight space ensembling [6, 27, 60] to adapt the model for downstream datasets, while preserving generalization ability. Other recent DG works [1, 26] propose using knowledge distillation to transfer CLIP's knowledge into DG models. However, as shown in Fig. 1, CLIP [49] aligns images with their descriptions during pretraining, which inevitably contain domain-specific details, impacting existing DG models. Additionally, since the training images are limited to a few source domains in DG, adapting CLIP within these images leads the model to emphasize the features specific to the source domains. This results in a lack of feature diversity [43, 46, 58], which com" + }, + { + "type": "page_footnote", + "bbox": [ + 0.115, + 0.888, + 0.251, + 0.901 + ], + "angle": 0, + "content": "Corresponding author." + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.516, + 0.957 + ], + "angle": 0, + "content": "9559" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.09, + 0.092, + 0.48, + 0.108 + ], + "angle": 0, + "content": "promises the model's ability to generalize to new domains." + }, + { + "type": "text", + "bbox": [ + 0.089, + 0.108, + 0.483, + 0.273 + ], + "angle": 0, + "content": "Erasing domain-sensitive details and mitigating bias benefit from learning domain-invariant representations for each class. In this paper, we propose a new DG approach to learn domain-invariant class features by leveraging a diverse set of text prompts. Based on the multi-modal alignment capabilities of CLIP, we use these text prompts to imitate multiple domains and effectively guide the DG learning process. A wide range of domains provides diverse perspectives on each class, allowing the model to learn features that are robust across varying contexts, thereby promoting domain-invariant feature learning." + }, + { + "type": "text", + "bbox": [ + 0.093, + 0.276, + 0.485, + 0.562 + ], + "angle": 0, + "content": "Diverse text prompts support three key aspects of our approach to learning domain-invariant features: they facilitate domain-sensitive feature suppression, ensure feature consistency across domain variations, and enable feature diversification to reduce bias. To be specific, we leverage the prompts recommended by original CLIP [49] to approximate different domains, such as \"a painting of a {class}\" or \"a {class} in a video game\". We first identify channels in the features activated by specific text prompts. These channels are considered highly relevant to the domain context of the prompt, thus we suppress their activation values. Next, we compare predictions under perturbations of diverse text prompts, and employ a consistency constraint to ensure the model is robust to domain variations while maintaining stable predictions. Subsequently, we decompose the component of the image features orthogonal to the text prompts and impose diversification on the orthogonal component. This approach enables a more diverse distribution of image features without impacting classification." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.563, + 0.483, + 0.654 + ], + "angle": 0, + "content": "To comprehensively evaluate our method, we perform extensive experiments on the widely used DomainBed benchmark [21]. The experiment results demonstrate the superior performance and robustness of our proposed approach, achieving state-of-the-art results across five domain generalization datasets." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.67, + 0.233, + 0.685 + ], + "angle": 0, + "content": "2. Related Work" + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.696, + 0.409, + 0.712 + ], + "angle": 0, + "content": "2.1. Large-scale Vision Language Models" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.72, + 0.483, + 0.902 + ], + "angle": 0, + "content": "Large-scale Vision Language Models (VLMs) [28, 49, 52, 61], trained on extensive image-text data from the internet, demonstrate strong robustness to diverse data distributions. These VLMs exhibit remarkable generalization capabilities, serving as fundamental models in the field of domain generalization. CLIP [49] is one of the most prevalent approaches among VLMs. It is trained by aligning image-text pairs through contrastive learning, where the model learns to associate corresponding images and texts from large-scale datasets by maximizing their similarity in a shared embedding space. Among the past few years, many efforts have been made to improve the CLIP-like VLMs by designing" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.092, + 0.905, + 0.198 + ], + "angle": 0, + "content": "better loss function [61], mitigating the impact of noisy web data [34, 35], and developing optimization techniques [52]. Another line of works focus on leveraging CLIP's generalization capabilities and its cross-modal alignment strengths to address downstream tasks such as zero-shot classification [29, 64], domain generalization [1, 26], and 2D/3D dense prediction [14, 45, 48]." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.208, + 0.731, + 0.223 + ], + "angle": 0, + "content": "2.2. Domain Generalization" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.231, + 0.905, + 0.35 + ], + "angle": 0, + "content": "Domain generalization [3, 8, 10, 18, 31, 40, 53, 59, 63] seeks to develop models capable of adapting to tasks from unseen domains without additional training. Conventional domain generalization methods encompass approaches such as learning domain-invariant features [13, 22, 37, 39], data augmentation [36, 42, 56], and adversarial learning [33, 38] to improve generalization to unseen domains." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.353, + 0.906, + 0.595 + ], + "angle": 0, + "content": "Recently, VLMs are widely utilized as foundational models in domain generalization [1, 20, 26, 50, 60]. Some approaches fine-tune CLIP model on downstream DG datasets. However, since these DG datasets (0.01 to \\(0.5\\mathrm{M}\\)) are significantly smaller than CLIP's training set (400M), directly fine-tuning disrupts CLIP's parameter space [30], leading to degraded generalization performance. To mitigate the limitations, many impressive works [6, 30, 60] propose parameter ensemble methods between CLIP and the fine-tuned model to preserve generalization ability. Other works adopt CLIP for DG downstream tasks by designing loss functions during fine-tuning [20, 50], or employing regularization in parameter-efficient fine-tuning method [23]. Some knowledge distillation-based methods [1, 26] align the features of DG models with the output features of CLIP, transferring CLIP's generalization ability to DG models." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.596, + 0.906, + 0.792 + ], + "angle": 0, + "content": "DomainDrop [22] mitigates model feature sensitivity to different domains by suppressing channel-wise feature variance across source domains. PromptStyler [11] generates diverse templates with style constraints and content consistency to train a text classifier. The text classifier is directly applied to visual domain generalization. LADS [16] and TEAM [47] employ text prompts for facilitating image feature augmentation. CAE [41] manually design a domain bank to obtain domain-unified text prompts for a stable prediction across domains. Different from these works that recognize the importance of text prompts, our approach enables direct optimization of the visual encoder by imitating multiple domain variations through a set of text prompts." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.802, + 0.709, + 0.818 + ], + "angle": 0, + "content": "2.3. Prompt Engineering" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.826, + 0.905, + 0.903 + ], + "angle": 0, + "content": "In CLIP, the design of prompts promotes the model to effectively capture the key information within the image. Recent prompt learning methods [29, 64, 65] are proposed to optimize the prompt as trainable parameters, allowing the model to dynamically adjust the prompt based on tasks." + }, + { + "type": "page_number", + "bbox": [ + 0.482, + 0.945, + 0.516, + 0.956 + ], + "angle": 0, + "content": "9560" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.089, + 0.091, + 0.482, + 0.198 + ], + "angle": 0, + "content": "CoOp [65] enhances CLIP's adaptability for visual recognition tasks by using learnable context vectors as language prompts. CoCoOp [64] incorporates image features as conditional prompts to better capture context. MaPLe [29] introduces a multimodal prompt learning approach, finetuning both image and text encoders simultaneously for improved performance." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.199, + 0.483, + 0.397 + ], + "angle": 0, + "content": "Recently, the domain generalization method DPL [62] employs an MLP to capture domain-specific image features as context for the learnable prompt. SPG [4] designs a prompt generator with domain knowledge to generate soft prompts for the images from target domain. These methods [4, 19, 62] primarily focus on incorporating image context into the text encoder to achieve multimodal alignment in downstream tasks. Differently, we optimize the image encoder through prompts from diverse templates. These prompts effectively imitate different domains while preventing the parameters of learnable prompts from being influenced by source domain-specific information during training." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.412, + 0.218, + 0.43 + ], + "angle": 0, + "content": "3. Preliminary" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.439, + 0.483, + 0.573 + ], + "angle": 0, + "content": "Problem Formulation. Domain generalization aims to train a model that can generalize well to unseen target domains, even when the training data comes from multiple source domains with potentially different data distributions. Specifically, the model is trained using data from \\(o\\) different source domains \\(\\{D_i^s\\}_{i=1}^o\\), where each source domain \\(D_i^s = \\{(x_i, y_i)\\}_{i=1}^u\\) consists of input data \\(x_i\\) and corresponding class labels \\(y_i\\), with \\(u\\) representing the number of samples in the source domain." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.577, + 0.483, + 0.651 + ], + "angle": 0, + "content": "The model's generalization performance is evaluated on the target domain \\( D^{t} \\), which shares the same set of class labels as the source domains \\( \\{D_{i}^{s}\\}_{i = 1}^{o} \\). The joint distribution of input data \\( x \\) and labels \\( y \\) in the target domain differs from that in the source domains." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.654, + 0.483, + 0.773 + ], + "angle": 0, + "content": "Utilizing CLIP for domain generalization. CLIP's strong generalization capability stems from its multimodal training on diverse internet data covering a wide range of concepts and contexts. As a prominent vision-language model, CLIP comprises an image encoder \\( F_{I}(\\cdot) \\) and a text encoder \\( F_{T}(\\cdot) \\). Both image and text encoders adopt a transformer architecture [15, 54] to generate their representations through self-attention." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.776, + 0.483, + 0.85 + ], + "angle": 0, + "content": "CLIP employs contrastive learning to align images and their descriptive text during pre-training. It associates visual and textual features by maximizing the similarity of the aligned image-text pairs while reducing the similarity of other ones:" + }, + { + "type": "equation", + "bbox": [ + 0.129, + 0.863, + 0.483, + 0.906 + ], + "angle": 0, + "content": "\\[\n\\mathcal {L} _ {\\mathrm {C L I P}} = - \\frac {1}{N} \\sum_ {i = 1} ^ {N} \\log \\frac {\\exp \\left(\\mathrm {S} \\left(I _ {i} , A _ {i}\\right) / \\tau\\right)}{\\sum_ {j = 1} ^ {N} \\exp \\left(\\mathrm {S} \\left(I _ {i} , A _ {j}\\right) / \\tau\\right)}, \\tag {1}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.092, + 0.905, + 0.152 + ], + "angle": 0, + "content": "where \\(I\\) and \\(A\\) represent the input image and caption, \\(i\\), \\(j\\) are indexes of the inputs, \\(S\\) means similarity matrices, such as cosine similarity. \\(\\tau\\) is a temperature parameter used to adjust the smoothness of the similarity distribution." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.153, + 0.907, + 0.318 + ], + "angle": 0, + "content": "Baseline of our approach. In the conventional domain generalization methods [1, 23, 29], \"a photo of a \\(\\{class\\}\\)\" is encoded by CLIP's text encoder as text prompt \\(\\{P_i\\}_{i=1}^c\\), representing the main concepts of \\(c\\) different classes. Leveraging CLIP's remarkable multi-modal alignment capabilities between images and text, the text prompts for different classes are treated as domain-invariant representations of the image features [1, 26]. We adopt a basic baseline [1] to adapt CLIP for downstream domain generalization tasks by simultaneously distilling image and text features to the DG model." + }, + { + "type": "equation", + "bbox": [ + 0.514, + 0.332, + 0.909, + 0.362 + ], + "angle": 0, + "content": "\\[\n\\mathcal {L} _ {\\mathrm {K D}} = \\lambda \\cdot \\cos \\left(F _ {I} \\left(x _ {i}\\right), f _ {I} \\left(x _ {i}\\right)\\right) + (1 - \\lambda) \\cdot \\cos \\left(F _ {T} \\left(y _ {i}\\right), f _ {I} \\left(x _ {i}\\right)\\right), \\tag {2}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.362, + 0.907, + 0.499 + ], + "angle": 0, + "content": "where \\( F_{I}(\\cdot) \\) and \\( F_{T}(\\cdot) \\) are CLIP's image and text encoders, \\( f_{I}(\\cdot) \\) is the DG model. Our approach differs slightly in that we ensemble multiple prompts used to imitate different domains during the distillation process. This has a minor direct impact on performance but is more intuitive to the design of our overall framework, ensuring stability across domain variations. The classification result \\( \\hat{y} \\) is obtained by choosing the ensembled text prompt with the highest similarity to the \\( i \\)-th image." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.513, + 0.649, + 0.531 + ], + "angle": 0, + "content": "4. Methodology" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.539, + 0.907, + 0.691 + ], + "angle": 0, + "content": "In this section, we introduce our proposed learning with diverse prompts paradigm for learning domain-invariant features in DG. Initially, we filter out unreliable prompts based on the prediction confidence. Next, we propose a Domain-sensitive Channel Suppression (DCS) method to identify and suppress domain-sensitive features. Moreover, a Multi-domain Consistency Constraint (MCC) approach is designed to make the features robust to variations across domains. Then, we apply Orthogonal Uniformity Regularization (OUR) to diversify features in the DG model." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.701, + 0.799, + 0.718 + ], + "angle": 0, + "content": "4.1. Entropy-based Prompt Filtering" + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.725, + 0.906, + 0.801 + ], + "angle": 0, + "content": "We imitate different domains by using the text templates \\(\\{T_j\\}_{j=1}^n\\) recommended in CLIP [49], such as \"a drawing of a {class}\", \\(n\\) is the number of templates. These prompts are generated by encoding the combination of the \\(j\\)-th text template \\(T_j\\) and the \\(i\\)-th class label \\(y_i\\) by CLIP text encoder:" + }, + { + "type": "equation", + "bbox": [ + 0.649, + 0.814, + 0.905, + 0.833 + ], + "angle": 0, + "content": "\\[\nP _ {i} ^ {j} = F _ {T} \\left(y _ {i}, T _ {j}\\right), \\tag {3}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.84, + 0.907, + 0.902 + ], + "angle": 0, + "content": "where \\(P \\in \\mathbb{R}^{n \\times c \\times D}\\), \\(c\\) is the number of classes, \\(D\\) represents the feature dimensions, \\(i\\) and \\(j\\) indicate the index of the classes and prompt templates. These prompts employ diverse contexts to describe each class, such as a drawing," + }, + { + "type": "page_number", + "bbox": [ + 0.482, + 0.945, + 0.514, + 0.957 + ], + "angle": 0, + "content": "9561" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.149, + 0.147, + 0.846, + 0.44 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.089, + 0.47, + 0.908, + 0.528 + ], + "angle": 0, + "content": "Figure 2. Overview of our proposed method. To enforce the DG model to learn domain-invariant class features, we imitate multiple domains by text prompts. Three components of our method, including feature suppression, feature consistency, and feature diversification, are achieved through Domain-sensitive Channel Suppression (DCS), Multi-domain Consistency Constraint (MCC), and Orthogonal Uniformity Regularization (OUR) modules. The blue line represents the process during inference." + }, + { + "type": "text", + "bbox": [ + 0.089, + 0.553, + 0.482, + 0.627 + ], + "angle": 0, + "content": "a sculpture, or a rendering, etc. Thus, these prompts can be naturally regarded as descriptions of the class from multiple domains. The DG model \\( f_{I}(\\cdot) \\) is regularized to extract domain-invariant features that are robust to prompts from different domains." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.629, + 0.483, + 0.763 + ], + "angle": 0, + "content": "Within the prompts provided by CLIP, we first discard a few unreliable ones based on the prediction confidence for each input [2], as they negatively affect the feature channel suppression and feature consistency calculation, which rely on accurately identifying relevant features. In the Entropy-based Prompt Filtering (EPF), we first extract the encoded image features \\( I_{i} = f_{I}(x_{i}) \\) for the input image \\( x_{i} \\). For the \\( j \\)-th template, the predicted probability for the \\( i \\)-th class is calculated as:" + }, + { + "type": "equation", + "bbox": [ + 0.183, + 0.775, + 0.483, + 0.813 + ], + "angle": 0, + "content": "\\[\n\\hat {d} _ {i} ^ {j} = \\frac {\\exp \\left(\\cos \\left(I _ {i} , P _ {i} ^ {j}\\right)\\right)}{\\sum_ {k = 1} ^ {c} \\exp \\left(\\cos \\left(I _ {i} , P _ {k} ^ {j}\\right)\\right)}, \\tag {4}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.823, + 0.483, + 0.856 + ], + "angle": 0, + "content": "We denote that \\(\\hat{d}_i^j\\) is a constant and \\(\\hat{d}^j = \\{\\hat{d}_i^j\\}_{i=1}^c\\) is a predicted distribution for the \\(j\\)-th template." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.856, + 0.484, + 0.902 + ], + "angle": 0, + "content": "Entropy quantifies the uncertainty of a probability distribution. If the model output has high confidence, the distribution typically concentrates on a specific class, resulting" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.553, + 0.906, + 0.614 + ], + "angle": 0, + "content": "in low entropy. Conversely, if the model has similar confidence across multiple classes, the distribution is more uniform, leading to higher entropy. Therefore, we calculate the entropy for each template's prediction:" + }, + { + "type": "equation", + "bbox": [ + 0.619, + 0.621, + 0.905, + 0.659 + ], + "angle": 0, + "content": "\\[\n\\mathrm {H} \\left(\\hat {d} ^ {j}\\right) = - \\sum_ {i = 1} ^ {c} \\hat {d} _ {i} ^ {j} \\log \\left(\\hat {d} _ {i} ^ {j}\\right). \\tag {5}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.666, + 0.906, + 0.697 + ], + "angle": 0, + "content": "Then, we eliminate the unreliable prompts and select the \\(m\\) prompts with the lowest entropy for each input:" + }, + { + "type": "equation", + "bbox": [ + 0.561, + 0.703, + 0.905, + 0.723 + ], + "angle": 0, + "content": "\\[\n\\left\\{T _ {1}, T _ {2}, \\dots , T _ {m} \\right\\} = \\arg \\min _ {j \\in \\{1 \\dots , n \\}} \\mathrm {H} \\left(\\hat {d} ^ {j}\\right), \\tag {6}\n\\]" + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.727, + 0.849, + 0.743 + ], + "angle": 0, + "content": "4.2. Domain-sensitive Channel Suppression" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.75, + 0.906, + 0.87 + ], + "angle": 0, + "content": "The image features should focus on discriminative features between classes and be robust to prompts from different domains. However, some domain-sensitive channels encode variations specific to the domain context of prompt, introducing perturbations. To address this, we propose Domain-sensitive Channel Suppression (DCS) to select and suppress these channels, ensuring the extracted image embedding captures domain-invariant class features." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.871, + 0.906, + 0.901 + ], + "angle": 0, + "content": "Specifically, we utilize gradient attribution [12] to quantify each channel's contribution in the image features when" + }, + { + "type": "page_number", + "bbox": [ + 0.482, + 0.945, + 0.516, + 0.957 + ], + "angle": 0, + "content": "9562" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.09, + 0.092, + 0.485, + 0.168 + ], + "angle": 0, + "content": "calculating their similarity with different prompts. We simultaneously consider the model's sensitivity to each channel (i.e., the gradient) and the actual activation value of that channel to measure each image channel's contribution to recognition with \\( P_{i}^{j} \\):" + }, + { + "type": "equation", + "bbox": [ + 0.195, + 0.181, + 0.483, + 0.217 + ], + "angle": 0, + "content": "\\[\n\\alpha_ {k j} = \\left| \\frac {\\partial \\cos \\left(I _ {i} , P _ {i} ^ {j}\\right)}{\\partial I _ {i} ^ {k}} \\right| \\cdot I _ {i} ^ {k}, \\tag {7}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.231, + 0.484, + 0.32 + ], + "angle": 0, + "content": "where \\( P_{i}^{j} \\) means the prompt corresponding to the \\( i \\)-th image's label with \\( T_{j} \\) as the template. \\( \\alpha_{kj} \\) is a constant that represents the contribution of the \\( k \\)-th image channel to the similarity calculation with the \\( j \\)-th prompt \\( P_{i}^{j} \\). For each image feature, we normalize it through the channel dimensions:" + }, + { + "type": "equation", + "bbox": [ + 0.206, + 0.32, + 0.483, + 0.358 + ], + "angle": 0, + "content": "\\[\n\\bar {\\alpha} _ {k j} = \\frac {\\alpha_ {k j}}{\\frac {1}{D} \\sqrt {\\sum_ {k = 1} ^ {D} \\alpha_ {k j} ^ {2}}}, \\tag {8}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.091, + 0.366, + 0.466, + 0.382 + ], + "angle": 0, + "content": "Then, we calculate the z-score among different prompts:" + }, + { + "type": "equation", + "bbox": [ + 0.22, + 0.393, + 0.483, + 0.422 + ], + "angle": 0, + "content": "\\[\n\\alpha_ {k j} ^ {\\mathrm {z - s c o r e}} = \\frac {\\overline {{\\alpha}} _ {k j} - \\mu}{\\sigma}, \\tag {9}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.433, + 0.484, + 0.462 + ], + "angle": 0, + "content": "\\(\\mu\\) and \\(\\sigma\\) are the average of \\(\\overline{\\alpha}_{kj}\\) across all templates, which are calculated as:" + }, + { + "type": "equation", + "bbox": [ + 0.232, + 0.473, + 0.483, + 0.514 + ], + "angle": 0, + "content": "\\[\n\\mu = \\frac {1}{m} \\sum_ {j = 1} ^ {m} \\bar {\\alpha} _ {k j} \\tag {10}\n\\]" + }, + { + "type": "equation", + "bbox": [ + 0.199, + 0.532, + 0.483, + 0.58 + ], + "angle": 0, + "content": "\\[\n\\sigma = \\sqrt {\\frac {1}{m} \\sum_ {j = 1} ^ {m} \\left(\\bar {\\alpha} _ {k j} - \\mu\\right) ^ {2}} \\tag {11}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.587, + 0.484, + 0.678 + ], + "angle": 0, + "content": "We check the z-score values of each channel across different prompts. A high z-score indicates that a channel is activated by the context provided by the specific prompt, showing a strong correlation with domain-specific information. We apply a mask to such image feature channels before calculating the loss:" + }, + { + "type": "equation", + "bbox": [ + 0.171, + 0.69, + 0.483, + 0.731 + ], + "angle": 0, + "content": "\\[\n\\mathbf {M} _ {D} = \\left\\{ \\begin{array}{l l} 1, & \\text {i f} \\max _ {k} \\alpha_ {k j} ^ {\\mathrm {z - s c o r e}} > \\beta \\\\ 0, & \\text {o t h e r w i s e} \\end{array} \\right. \\tag {12}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.744, + 0.484, + 0.805 + ], + "angle": 0, + "content": "The mask \\(\\mathbf{M}_D\\in \\mathbb{R}^D\\) is directly employed on image features by the element-wise multiplication \\(I_i^k = \\mathbf{M}_D\\odot I_i\\). We suppress these domain-sensitive features by applying \\(L_{1}\\) regularization:" + }, + { + "type": "equation", + "bbox": [ + 0.211, + 0.817, + 0.483, + 0.859 + ], + "angle": 0, + "content": "\\[\n\\mathcal {L} _ {r e g} = \\gamma \\sum_ {i = 1} ^ {B} \\sum_ {k = 1} ^ {D} \\left| I _ {i} ^ {k} \\right|, \\tag {13}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.871, + 0.484, + 0.902 + ], + "angle": 0, + "content": "where \\(B\\) represents the batch size, and \\(\\gamma\\) is the weight of loss, \\(I_{i}^{k}\\) is the masked image feature." + }, + { + "type": "title", + "bbox": [ + 0.514, + 0.091, + 0.842, + 0.107 + ], + "angle": 0, + "content": "4.3. Multi-domain Consistency Constraint" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.114, + 0.907, + 0.267 + ], + "angle": 0, + "content": "We further introduce Multi-domain Consistency Constraint (MCC) on the predicted distribution between the image and varied prompts. We calculate the similarity between image features and different prompts for classification. The KL loss function is then applied to ensure that the predicted distributions remain consistent between different prompts, enhancing the robustness of the image encoder to domain variations. Specifically, we use Eq. (4) to obtain the predicted distributions for different templates \\(\\{\\hat{d}^j\\}_{j=1}^m\\), and apply KL loss to ensure consistency:" + }, + { + "type": "equation", + "bbox": [ + 0.614, + 0.277, + 0.907, + 0.318 + ], + "angle": 0, + "content": "\\[\n\\operatorname {K L} \\left(\\hat {d} ^ {j} \\| \\hat {d} ^ {k}\\right) = \\sum_ {i = 1} ^ {c} \\hat {d} _ {i} ^ {j} \\log \\frac {\\hat {d} _ {i} ^ {j}}{\\hat {d} _ {i} ^ {k}} \\tag {14}\n\\]" + }, + { + "type": "equation", + "bbox": [ + 0.581, + 0.332, + 0.906, + 0.381 + ], + "angle": 0, + "content": "\\[\n\\mathcal {L} _ {\\mathrm {K L}} = \\frac {1}{m (m - 1)} \\sum_ {j = 1} ^ {m} \\sum_ {\\substack {k = 1 \\\\ k \\neq t}} ^ {m} \\mathrm {K L} \\left(\\hat {d} ^ {j} \\| \\hat {d} ^ {k}\\right) \\tag{15}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.387, + 0.906, + 0.432 + ], + "angle": 0, + "content": "The KL loss compares the prediction with prompts from diverse domains and enforces consistency, guiding the learned features to focus on domain-invariant information." + }, + { + "type": "title", + "bbox": [ + 0.514, + 0.443, + 0.849, + 0.46 + ], + "angle": 0, + "content": "4.4. Orthogonal Uniformity Regularization" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.467, + 0.907, + 0.602 + ], + "angle": 0, + "content": "Due to the limited diversity of image source domains, learning with these images biases CLIP toward the feature distribution of these source domains. To mitigate this bias, we further diversify the normalized image features to achieve a more uniform distribution in the hyperspherical space. Uniformity of the features increases the possibility that unseen features are positioned near seen features [46], thereby enhancing the model's generalization through feature diversification." + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.603, + 0.906, + 0.649 + ], + "angle": 0, + "content": "However, it is not feasible to directly diversify the image features. As mentioned in Sec. 3, we use an ensembled prompt \\( e_i \\) in classification, which can be formulated as:" + }, + { + "type": "equation", + "bbox": [ + 0.655, + 0.661, + 0.906, + 0.701 + ], + "angle": 0, + "content": "\\[\ne _ {i} = \\frac {1}{n} \\sum_ {j = 1} ^ {n} P _ {i} ^ {j}. \\tag {16}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.713, + 0.906, + 0.744 + ], + "angle": 0, + "content": "We compute the similarity between the ensembled prompt \\( e_i \\) and the image feature \\( I_i \\) for classification:" + }, + { + "type": "equation", + "bbox": [ + 0.625, + 0.757, + 0.907, + 0.775 + ], + "angle": 0, + "content": "\\[\n\\hat {y} = \\operatorname {a r g m a x} _ {j} \\cos \\left(I _ {i}, e _ {j}\\right). \\tag {17}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.785, + 0.906, + 0.83 + ], + "angle": 0, + "content": "Therefore, as demonstrated in Eq. (17), directly applying uniformity constraints on image features can impact the similarity calculation between image features and prompts." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.831, + 0.906, + 0.873 + ], + "angle": 0, + "content": "To address this, we decompose \\( I_{i} \\) into two components, one in the direction of \\( e_{i} \\) and the other in the orthogonal direction:" + }, + { + "type": "equation", + "bbox": [ + 0.653, + 0.873, + 0.905, + 0.906 + ], + "angle": 0, + "content": "\\[\nI _ {i} ^ {\\parallel} = \\frac {I _ {i} \\cdot e _ {i}}{\\| e _ {i} \\| ^ {2}} \\cdot e _ {i}, \\tag {18}\n\\]" + }, + { + "type": "page_number", + "bbox": [ + 0.482, + 0.945, + 0.516, + 0.957 + ], + "angle": 0, + "content": "9563" + } + ], + [ + { + "type": "table_caption", + "bbox": [ + 0.09, + 0.089, + 0.907, + 0.132 + ], + "angle": 0, + "content": "Table 1. Comparison with the state-of-the-art methods on the five domain generalization datasets from DomainBed benchmark. We report the averaged leave-one-domain-out experimental results with ViT-B/16 as the backbone. Numbers in bold represent the best performance among all the methods. T and F indicate whether the CLIP parameters are trained or frozen during adaptation to DG." + }, + { + "type": "table", + "bbox": [ + 0.113, + 0.142, + 0.887, + 0.43 + ], + "angle": 0, + "content": "
MethodsCLIPOHTerraIncVLCSPACSDomainNetAvg
Fine-tuning
SWAD (NeurIPS'2021) [6]T81.0142.9279.1391.3557.9270.47
LP-FT (ICLR'2022) [30]T81.1747.2680.8892.9257.0471.85
WiSE-FT (CVPR'2022) [60]T86.3254.5082.8897.2958.0175.80
MIRO (ECCV'2022) [7]T84.8059.3082.3096.4460.4776.66
DART (CVPR'2023) [27]T80.9351.2480.3893.4359.3273.06
SAGM (CVPR'2023) [57]T83.4058.6482.0594.3159.0575.49
FLYP (CVPR'2023) [20]T82.7633.2566.6478.5357.4163.72
CLIPood (ICML'2023) [50]T83.3146.2877.1993.1657.7871.55
Knowledge Distillation
RISE (ICCV'2023) [26]T78.3949.6180.6293.2555.3771.45
VL2V-SD (CVPR'2024) [1]T87.3858.5483.2596.6862.7977.73
Prompt-learning
DPL (TJSAI'2023) [62]F84.2052.6084.3097.3059.5075.60
SPG (ECCV'2024) [4]F83.6050.2082.4097.0060.1074.70
Parameter-efficient Fine-tuning
LPD (ECCV'2024) [23]F84.2157.3083.2096.5559.3076.11
OursT87.6563.2784.7997.0363.1179.17
" + }, + { + "type": "equation", + "bbox": [ + 0.237, + 0.465, + 0.482, + 0.486 + ], + "angle": 0, + "content": "\\[\nI _ {i} ^ {\\perp} = I _ {i} - I _ {i} ^ {\\parallel}, \\tag {19}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.491, + 0.484, + 0.555 + ], + "angle": 0, + "content": "where \\( I_{i}^{\\parallel} \\) and \\( I_{i}^{\\perp} \\) are the decomposed parallel and orthogonal components, respectively. We define a pairwise uniformity loss based on the dot product between the orthogonal components of different image features within the batch:" + }, + { + "type": "equation", + "bbox": [ + 0.162, + 0.561, + 0.483, + 0.614 + ], + "angle": 0, + "content": "\\[\n\\mathcal {L} _ {\\mathrm {O U}} = \\frac {1}{B (B - 1)} \\sum_ {i = 1} ^ {B} \\sum_ {\\substack {j = 1 \\\\ j \\neq i}} ^ {B} \\left(I _ {i} ^ {\\perp} \\cdot I _ {j} ^ {\\perp}\\right) ^ {2} \\tag{20}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.621, + 0.484, + 0.697 + ], + "angle": 0, + "content": "Since \\( e_i \\) is computed independently for each image, the orthogonal component is also based on different images. Applying the uniformity constraint solely on the orthogonal direction avoids interference with the similarity calculation between the image features and the prompt." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.697, + 0.484, + 0.758 + ], + "angle": 0, + "content": "Finally, combining the baseline with Domain-sensitive Channel Suppression, Multi-domain Consistency Constraint, and Orthogonal Uniformity Regularization yields the total loss used during training, which is formulated as:" + }, + { + "type": "equation", + "bbox": [ + 0.156, + 0.767, + 0.483, + 0.784 + ], + "angle": 0, + "content": "\\[\n\\mathcal {L} _ {\\text {t o t a l}} = \\mathcal {L} _ {K D} + \\mathcal {L} _ {\\text {r e g}} + \\mathcal {L} _ {K L} + \\mathcal {L} _ {O U}, \\tag {21}\n\\]" + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.793, + 0.224, + 0.81 + ], + "angle": 0, + "content": "5. Experiments" + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.817, + 0.314, + 0.833 + ], + "angle": 0, + "content": "5.1. Datasets and Evaluation" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.84, + 0.484, + 0.902 + ], + "angle": 0, + "content": "We conduct evaluation experiments on five datasets from the DomainBed benchmark [21], following previous methods [1, 22, 26]. PACS [32] consists of 9,991 samples spanning four domains: {art, cartoons, photos, sketches}, and" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.453, + 0.907, + 0.65 + ], + "angle": 0, + "content": "includes a total of 7 classes. VLCS [17] contains 10,729 examples, 5 classes from four datasets as different domains: {VOC2007, LabelMe, Caltech, SUN}. Terra Incognita (TI) [5] contains 24,788 samples from 10 animal categories. These photos are captured across different locations {L100, L38, L43, L46}, acting as the four domains. Office Home (OH) [55] have four domains {art, clipart, product, real}, with 15,588 examples from 65 classes. DomainNet [44] is a large-scale cross-domain dataset containing 586,575 images from 6 different domains: {clipart, infographic, painting, quickdraw, real, sketch}. It covers 345 classes and is widely used to evaluate the performance of domain generalization algorithms." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.652, + 0.909, + 0.772 + ], + "angle": 0, + "content": "Following the protocols from previous domain generalization works [1, 11, 26], we split the datasets by using \\(80\\%\\) of the data for training and the remaining \\(20\\%\\) as the validation set. We adopt a leave-one-out experiment setup, where we alternately leave on domain for testing and train with the remaining domains. The average performance on the unseen domains is reported as the domain generalization results." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.786, + 0.731, + 0.802 + ], + "angle": 0, + "content": "5.2. Implementation Details" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.81, + 0.907, + 0.902 + ], + "angle": 0, + "content": "We employ CLIP with ViT-B/16 as the backbone model and fine-tune it on each of the datasets. We fine-tune the model for 5,000 iterations on the PACS, VLCS, Office Home, and Terra Incognito datasets, and for 15,000 iterations on the DomainNet dataset. We use the \\( n = 80 \\) templates recommended in original CLIP [49] as prompts to imitate dif" + }, + { + "type": "page_number", + "bbox": [ + 0.482, + 0.945, + 0.516, + 0.957 + ], + "angle": 0, + "content": "9564" + } + ], + [ + { + "type": "table_caption", + "bbox": [ + 0.09, + 0.089, + 0.486, + 0.175 + ], + "angle": 0, + "content": "Table 2. Ablation analysis on the effectiveness of our method. EPS: entropy-based prompt selection, DCS: domain-sensitive channel suppression, MCC: multi-domain consistency constraint. OUR: orthogonal uniformity regularization, URIF*: uniformity regularization for image features, where * indicates it is not our component, but used as an additional comparison to OUR." + }, + { + "type": "table", + "bbox": [ + 0.092, + 0.184, + 0.486, + 0.326 + ], + "angle": 0, + "content": "
BaseEPSDCSMCCOURURIF*Avg
✓81.46
✓✓✓82.28
✓✓✓82.37
✓✓✓✓82.79
✓✓82.30
✓✓81.55
✓✓✓✓✓83.19
" + }, + { + "type": "text", + "bbox": [ + 0.089, + 0.334, + 0.486, + 0.502 + ], + "angle": 0, + "content": "ferent domains for each class. The threshold \\(\\beta\\) for the suppressed feature mask \\(\\mathbf{M}_D\\) is set dynamically for each sample. We suppress \\(20\\%\\) image feature channels to reduce domain-specific perturbation, and set the regularization term as \\(10^{-4}\\). We select \\(m = 40\\) prompts with low entropy from 80 contextual prompts. Other experimental settings all follow our baseline model. SWAD [6] is a general approach that improves model robustness by averaging weights during fine-tuning DG models. For a fair comparison, we integrate SWAD into all methods that fine-tune CLIP's parameters, following the previous work [1]." + }, + { + "type": "title", + "bbox": [ + 0.09, + 0.513, + 0.396, + 0.529 + ], + "angle": 0, + "content": "5.3. Comparison with State-of-the-Arts" + }, + { + "type": "text", + "bbox": [ + 0.093, + 0.538, + 0.483, + 0.765 + ], + "angle": 0, + "content": "We mainly compare with recent advances in domain generalization methods based on CLIP. We first compares our method with the following fine-tuning methods. SWAD [6], WiSE-FT [60], and DART [27] explore weight-ensembling methods to maintain CLIP's generalization ability. LPFT [30], FLYP [20], and CLIPood [50] investigate effective fine-tuning methods. The importance of text is highlighted in VL2V-SD [1] and RISE [26]. They incorporate text features into the DG model via knowledge distillation. DPL [62] and SPG [4] are prompt-learning methods that design learnable modules to extract context features from input images. LPD [23] explores a parameter-efficient fine-tuning approach for adapting CLIP in DG. We report the averaged leave-one-domain-out experimental results following the setting from previous works [1, 4]." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.766, + 0.486, + 0.903 + ], + "angle": 0, + "content": "As shown in Table 1, our approach achieves an average improvement of \\(1.4\\%\\) over SOTA methods across five datasets. Our method demonstrates a particularly notable \\(4.7\\%\\) improvement over SOTA on the TerraInc dataset [5]. As the images in TerraInc are sourced from wildlife photography, target objects occupy a small portion of the scene and contain substantial irrelevant contextual information. Our approach effectively reduces interference from such perturbation and achieves impressive improvement." + }, + { + "type": "table_caption", + "bbox": [ + 0.513, + 0.089, + 0.908, + 0.161 + ], + "angle": 0, + "content": "Table 3. Performance comparison of our method with various prompt learning approaches across different datasets, including OH, TI, VLCS, and PACS. The average accuracy (Avg) across these datasets highlights the robustness and superior generalization capability of template generated prompts." + }, + { + "type": "table", + "bbox": [ + 0.516, + 0.172, + 0.912, + 0.303 + ], + "angle": 0, + "content": "
MethodsOHTIVLCSPACSAvg
CLIP [49]82.434.182.396.573.8
CoOP [65]83.046.880.896.476.8
Co-CoOP [64]83.445.380.396.776.4
Maple [29]83.450.282.296.578.1
DPL [62]84.252.684.397.379.6
SPG [4]83.650.282.497.078.3
Ours87.763.384.897.083.2
" + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.316, + 0.667, + 0.332 + ], + "angle": 0, + "content": "5.4. Ablation Study" + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.34, + 0.907, + 0.537 + ], + "angle": 0, + "content": "Ablation of the main components. We conduct ablation studies to evaluate the effectiveness of each component of our approach. Following previous methods [1, 23], we report the results on four datasets: OH [55], TI [5], VLCS [17], and PACS [32]. 'Avg' denotes the average of the four datasets. As shown in Table 2, using DCS with EPS yields a notable performance increase and improves the accuracy to \\(82.28\\%\\), highlighting the significance of suppressing the domain-specific channels. MCC also achieves an improvement, raising accuracy to \\(82.37\\%\\) by enforcing cross-domain consistency. Applying both DSC and MCC modules simultaneously to suppress the domain-specific context results in an average gain of \\(1.33\\%\\)." + }, + { + "type": "text", + "bbox": [ + 0.511, + 0.538, + 0.909, + 0.719 + ], + "angle": 0, + "content": "Moreover, incorporating uniformity regularization for the orthogonal component increases the average accuracy to \\(82.30\\%\\), suggesting that capturing more diversified features enables the model to generalize better across domains by preventing an overemphasis on specific source domains. In contrast, applying uniformity regularization directly to image features (URIF*) without orthogonal decomposition yields a decrease to \\(81.55\\%\\), demonstrating that directly regularizing image features may introduce perturbations during classification. Finally, the best performance is obtained by combining EPS, DSC, MCR, and OUR, reaching an average accuracy of \\(83.19\\%\\)." + }, + { + "type": "text", + "bbox": [ + 0.511, + 0.72, + 0.911, + 0.903 + ], + "angle": 0, + "content": "Discussions on the design of prompts. In our implementation, we imitate diverse domains by the prompts generated from 80 templates recommended by CLIP. In this section, we conduct experiments to analysis the advantages of our template-based prompts over the learnable prompts used in previous methods [4, 29, 62, 64]. By optimizing the learnable prompts during training, these model learn to adapt dynamically to different inputs. As shown in Table 3, although these methods show significant improvement compared to CLIP, our approach still outperforms the state-of-the-art prompt learning methods by \\(3.6\\%\\). While they incorporate context of the image, the learnable prompts also" + }, + { + "type": "page_number", + "bbox": [ + 0.482, + 0.945, + 0.516, + 0.957 + ], + "angle": 0, + "content": "9565" + } + ], + [ + { + "type": "table_caption", + "bbox": [ + 0.09, + 0.089, + 0.486, + 0.145 + ], + "angle": 0, + "content": "Table 4. Ablations of comparison with knowledge distillation from the single manually designed prompt and the design of our method. EP represents the ensembled \\( n \\) prompts for knowledge distillation." + }, + { + "type": "table", + "bbox": [ + 0.093, + 0.157, + 0.485, + 0.27 + ], + "angle": 0, + "content": "
PromptsOHTIVLCSPACSAvg
{class}86.559.583.796.681.6
a photo of a {class}87.458.583.396.781.5
a {domain} of a {class}87.260.783.996.982.2
EP87.159.983.696.581.9
Ours87.763.384.897.083.2
" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.293, + 0.483, + 0.368 + ], + "angle": 0, + "content": "introduces potential perturbations from specific source domains. Unlike these learnable prompts, our method leverages a set of templates without requiring domain-specific adaptation during training, enabling the model to generalize more robustly across diverse domains." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.37, + 0.483, + 0.687 + ], + "angle": 0, + "content": "Analysis on knowledge distillation from different designed prompts. In this part, we conduct experiments to demonstrate the advantages of using a diverse set of prompts compared to the carefully designed manual prompt. In our approach, the diverse set of prompts are ensembled and a knowledge distillation framework is used to transfer the ensembled prompts (EP) to the DG model. Based on the knowledge distillation baseline, we compare with three types of manually designed prompts, including {\"class)\", \"a photo of a {class}\", and \"a {domain}\" of a {class)\". As shown in Table 4, a simple use of ensembled prompts provides slight gains, with \\(0.3\\%\\) and \\(0.4\\%\\) improvements compared to manual prompts {\"class\"} and \"a photo of a {class}\". Moreover, EP even performs worse than the carefully designed prompt: \"a {domain}\" of a {class)\". However, a set of diverse prompts enables us to imitate multiple domains and learn domain-invariant features. Our overall method achieves consistently better performance across all four datasets, with notable gains in challenging cases like TI and VLCS. Such results demonstrate the effectiveness of our multi-domain imitation method." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.689, + 0.483, + 0.901 + ], + "angle": 0, + "content": "Hyperparameter analysis. Additionally, we investigate the hyperparameter for the number of selected templates \\( m \\) and the ratio of channels suppressed during gradient attribution. Table 5 presents analysis on the number of prompts and its impact on the average performance across four datasets. We evaluate different values for \\( m \\), ranging from 10 to 80. The results show that increasing \\( m \\) improves performance at first, when \\( m = 40 \\), the performance reaches its maximum value \\( 83.2\\% \\). Using more templates ensures that the imitated domains contain more complex contexts, thus facilitating our model in learning robust features. However, further increases \\( m \\) reduce the performance, with the average accuracy dropping to \\( 83.0\\% \\) at \\( m = 60 \\) and \\( 82.3\\% \\) at \\( m = 80 \\). Since we do not design" + }, + { + "type": "table_caption", + "bbox": [ + 0.52, + 0.09, + 0.9, + 0.105 + ], + "angle": 0, + "content": "Table 5. Hyperparameter analysis on the number of prompts \\( m \\) ." + }, + { + "type": "table", + "bbox": [ + 0.526, + 0.116, + 0.897, + 0.16 + ], + "angle": 0, + "content": "
m1020406080
Avg82.782.983.283.082.3
" + }, + { + "type": "table_caption", + "bbox": [ + 0.513, + 0.169, + 0.905, + 0.198 + ], + "angle": 0, + "content": "Table 6. Hyperparameter analysis on the ratio of channels suppressed during gradient attribution." + }, + { + "type": "table", + "bbox": [ + 0.526, + 0.209, + 0.897, + 0.253 + ], + "angle": 0, + "content": "
Ratio0.10.20.30.40.5
Avg82.383.282.081.781.6
" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.268, + 0.905, + 0.327 + ], + "angle": 0, + "content": "different templates for different datasets, utilizing all of the templates without selection introduces noise, leading to a drop in performance. Therefore, we set the hyperparameter \\( m \\) as 40." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.33, + 0.907, + 0.542 + ], + "angle": 0, + "content": "Table 6 presents the analysis on the ratio of channels suppressed during gradient attribution and its effect on the overall performance across four datasets. We evaluate different suppression ratios, ranging from 0.1 to 0.5. The results show that increasing the suppression ratio improves performance at first, with the maximum average accuracy of \\(83.2\\%\\) achieved at a suppression ratio of 0.2. However, further increases in the suppression ratio slightly reduce accuracy, with the average performance dropping to \\(82.0\\%\\), \\(81.7\\%\\), and \\(81.6\\%\\) at 0.3, 0.4, and 0.5. The experiments demonstrate that excessive suppression filters out essential information, resulting in degraded performance. Thus, we set the suppression ratio to 0.2 as the optimal value for this hyperparameter." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.559, + 0.634, + 0.575 + ], + "angle": 0, + "content": "6. Conclusion" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.585, + 0.907, + 0.796 + ], + "angle": 0, + "content": "We develop a new domain generalization approach by learning with diverse text prompts. We use these prompts to imitate multiple domains, which guide the DG model in learning domain-invariant features through multi-modal alignment in CLIP. These prompts contribute to three aspects, including suppressing domain-sensitive features, ensure feature consistency with varied domains, and enable feature diversification in the DG model. By incorporating these three terms, our approach reduces the CLIP-based DG model's tendency to capture domain-sensitive details and mitigates bias towards the source data distribution. Experimental results on five domain generalization datasets from the DomainBed benchmark demonstrate the effectiveness of our method." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.814, + 0.693, + 0.831 + ], + "angle": 0, + "content": "7. Acknowledgments" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.84, + 0.906, + 0.9 + ], + "angle": 0, + "content": "This work was supported by the NSFC under Grant 62322604 and 62176159, and in part by the Shanghai Municipal Science and Technology Major Project under Grant 2021SHZDZX0102." + }, + { + "type": "page_number", + "bbox": [ + 0.482, + 0.946, + 0.516, + 0.957 + ], + "angle": 0, + "content": "9566" + } + ], + [ + { + "type": "title", + "bbox": [ + 0.093, + 0.09, + 0.188, + 0.106 + ], + "angle": 0, + "content": "References" + }, + { + "type": "ref_text", + "bbox": [ + 0.101, + 0.115, + 0.484, + 0.183 + ], + "angle": 0, + "content": "[1] Sravanti Addepalli, Ashish Ramayee Asokan, Lakshay Sharma, and R Venkatesh Babu. Leveraging vision-language models for improving domain generalization in image classification. In CVPR, pages 23922-23932, 2024. 1, 2, 3, 6, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.101, + 0.185, + 0.484, + 0.254 + ], + "angle": 0, + "content": "[2] James Urquhart Allingham, Jie Ren, Michael W Dusenberry, Xiuye Gu, Yin Cui, Dustin Tran, Jeremiah Zhe Liu, and Balaji Lakshminarayanan. A simple zero-shot prompt weighting technique to improve prompt ensembling in text-image models. In ICML, 2023. 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.256, + 0.483, + 0.297 + ], + "angle": 0, + "content": "[3] Martin Arjovsky, Léon Bottou, Ishaan Gulrajani, and David Lopez-Paz. Invariant risk minimization. arXiv preprint arXiv:1907.02893, 2019. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.298, + 0.483, + 0.339 + ], + "angle": 0, + "content": "[4] Shuanghao Bai, Yuedi Zhang, Wanqi Zhou, Zhirong Luan, and Badong Chen. Soft prompt generation for domain generalization. In ECCV, pages 434-450, 2024. 3, 6, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.341, + 0.483, + 0.368 + ], + "angle": 0, + "content": "[5] Sara Beery, Grant Van Horn, and Pietro Perona. Recognition in terra incognita. In ECCV, pages 456-473, 2018. 6, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.37, + 0.483, + 0.423 + ], + "angle": 0, + "content": "[6] Junbum Cha, Sanghyuk Chun, Kyungjae Lee, Han-Cheol Cho, Seunghyun Park, Yunsung Lee, and Sungrae Park. Swad: Domain generalization by seeking flat minima. NeurIPS, 34:22405-22418, 2021. 1, 2, 6, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.425, + 0.483, + 0.478 + ], + "angle": 0, + "content": "[7] Junbum Cha, Kyungjae Lee, Sungrae Park, and Sanghyuk Chun. Domain generalization by mutual-information regularization with pre-trained models. In ECCV, pages 440-457, 2022. 1, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.481, + 0.483, + 0.522 + ], + "angle": 0, + "content": "[8] Liang Chen, Yong Zhang, Yibing Song, Ying Shan, and Lingqiao Liu. Improved test-time adaptation for domain generalization. In CVPR, pages 24172-24182, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.524, + 0.483, + 0.564 + ], + "angle": 0, + "content": "[9] Liang Chen, Yong Zhang, Yibing Song, Anton Van Den Hengel, and Lingqiao Liu. Domain generalization via rationale invariance. In ICCV, pages 1751-1760, 2023. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.566, + 0.483, + 0.619 + ], + "angle": 0, + "content": "[10] De Cheng, Zhipeng Xu, Xinyang Jiang, Nannan Wang, Dongsheng Li, and Xinbo Gao. Disentangled prompt representation for domain generalization. In CVPR, pages 23595-23604, 2024. 1, 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.622, + 0.483, + 0.676 + ], + "angle": 0, + "content": "[11] Junhyeong Cho, Gilhyun Nam, Sungyeon Kim, Hunmin Yang, and Suha Kwak. Promptstyler: Prompt-driven style generation for source-free domain generalization. In ICCV, pages 15702-15712, 2023. 2, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.678, + 0.483, + 0.746 + ], + "angle": 0, + "content": "[12] Huiqi Deng, Na Zou, Mengnan Du, Weifu Chen, Guocan Feng, Ziwei Yang, Zheyang Li, and Quanshi Zhang. Unifying fourteen post-hoc attribution methods with taylor interactions. IEEE Transactions on Pattern Analysis and Machine Intelligence, 46(7):4625-4640, 2024. 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.748, + 0.483, + 0.802 + ], + "angle": 0, + "content": "[13] Yu Ding, Lei Wang, Bin Liang, Shuming Liang, Yang Wang, and Fang Chen. Domain generalization by learning and removing domain-specific features. In NeurIPS, pages 24226-24239, 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.804, + 0.483, + 0.872 + ], + "angle": 0, + "content": "[14] Xiaoyi Dong, Jianmin Bao, Yinglin Zheng, Ting Zhang, Dongdong Chen, Hao Yang, Ming Zeng, Weiming Zhang, Lu Yuan, Dong Chen, et al. Maskclip: Masked self-distillation advances contrastive language-image pretraining. In CVPR, pages 10995-11005, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.874, + 0.483, + 0.901 + ], + "angle": 0, + "content": "[15] Alexey Dosovitskiy, Lucas Beyer, Alexander Kolesnikov, Dirk Weissenborn, Xiaohua Zhai, Thomas Unterthiner," + }, + { + "type": "list", + "bbox": [ + 0.094, + 0.115, + 0.484, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.548, + 0.093, + 0.905, + 0.147 + ], + "angle": 0, + "content": "Mostafa Dehghani, Matthias Minderer, Georg Heigold, Sylvain Gelly, Jakob Uszkoreit, and Neil Houlsby. An image is worth 16x16 words: Transformers for image recognition at scale. In ICLR, 2021. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.15, + 0.905, + 0.204 + ], + "angle": 0, + "content": "[16] Lisa Dunlap, Clara Mohri, Devin Guillory, Han Zhang, Trevor Darrell, Joseph E Gonzalez, Aditi Raghunathan, and Anna Rohrbach. Using language to extend to unseen domains. In ICLR, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.207, + 0.905, + 0.261 + ], + "angle": 0, + "content": "[17] Chen Fang, Ye Xu, and Daniel N Rockmore. Unbiased metric learning: On the utilization of multiple datasets and web images for softening bias. In ICCV, pages 1657-1664, 2013. 6, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.264, + 0.905, + 0.331 + ], + "angle": 0, + "content": "[18] Yaroslav Ganin, Evgeniya Ustinova, Hana Ajakan, Pascal Germain, Hugo Larochelle, François Laviolette, Mario March, and Victor Lempitsky. Domain-adversarial training of neural networks. Journal of machine learning research, 17(59):1-35, 2016. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.334, + 0.905, + 0.389 + ], + "angle": 0, + "content": "[19] Chunjiang Ge, Rui Huang, Mixue Xie, Zihang Lai, Shiji Song, Shuang Li, and Gao Huang. Domain adaptation via prompt learning. IEEE Transactions on Neural Networks and Learning Systems, pages 1-11, 2023. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.391, + 0.905, + 0.445 + ], + "angle": 0, + "content": "[20] Sachin Goyal, Ananya Kumar, Sankalp Garg, Zico Kolter, and Aditi Raghunathan. Finetune like you pretrain: Improved finetuning of zero-shot vision models. In CVPR, pages 19338-19347, 2023. 1, 2, 6, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.448, + 0.905, + 0.473 + ], + "angle": 0, + "content": "[21] Ishaan Gulrajani and David Lopez-Paz. In search of lost domain generalization. In ICLR, 2020. 2, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.476, + 0.905, + 0.517 + ], + "angle": 0, + "content": "[22] Jintao Guo, Lei Qi, and Yinghuan Shi. Domaindrop: Suppressing domain-sensitive channels for domain generalization. In ICCV, pages 19114-19124, 2023. 2, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.519, + 0.905, + 0.574 + ], + "angle": 0, + "content": "[23] Jiajun Hu, Jian Zhang, Lei Qi, Yinghuan Shi, and Yang Gao. Learn to preserve and diversify: Parameter-efficient group with orthogonal regularization for domain generalization. In ECCV, pages 198-216, 2024. 1, 2, 3, 6, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.577, + 0.905, + 0.63 + ], + "angle": 0, + "content": "[24] Lanqing Hu, Meina Kan, Shiguang Shan, and Xilin Chen. Dandelionnet: Domain composition with instance adaptive classification for domain generalization. In ICCV, pages 19050-19059, 2023. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.633, + 0.905, + 0.674 + ], + "angle": 0, + "content": "[25] Zenan Huang, Haobo Wang, Junbo Zhao, and Nenggan Zheng. idag: Invariant dag searching for domain generalization. In ICCV, pages 19169-19179, 2023. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.676, + 0.905, + 0.743 + ], + "angle": 0, + "content": "[26] Zeyi Huang, Andy Zhou, Zijian Ling, Mu Cai, Haohan Wang, and Yong Jae Lee. A sentence speaks a thousand images: Domain generalization through distilling clip with language guidance. In ICCV, pages 11685-11695, 2023. 1, 2, 3, 6, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.746, + 0.905, + 0.801 + ], + "angle": 0, + "content": "[27] Samyak Jain, Sravanti Addepalli, Pawan Kumar Sahu, Priyam Dey, and R Venkatesh Babu. Dart: Diversify-aggregate-repeat training improves generalization of neural networks. In CVPR, pages 16048-16059, 2023. 1, 6, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.803, + 0.905, + 0.87 + ], + "angle": 0, + "content": "[28] Chao Jia, Yinfei Yang, Ye Xia, Yi-Ting Chen, Zarana Parekh, Hieu Pham, Quoc Le, Yun-Hsuan Sung, Zhen Li, and Tom Duerig. Scaling up visual and vision-language representation learning with noisy text supervision. In ICML, pages 4904-4916, 2021. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.873, + 0.905, + 0.901 + ], + "angle": 0, + "content": "[29] Muhammad Uzair Khattak, Hanoona Rasheed, Muhammad Maaz, Salman Khan, and Fahad Shahbaz Khan. Maple:" + }, + { + "type": "list", + "bbox": [ + 0.517, + 0.093, + 0.905, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.516, + 0.957 + ], + "angle": 0, + "content": "9567" + } + ], + [ + { + "type": "ref_text", + "bbox": [ + 0.126, + 0.092, + 0.482, + 0.12 + ], + "angle": 0, + "content": "Multi-modal prompt learning. In CVPR, pages 19113-19122, 2023. 2, 3, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.123, + 0.483, + 0.178 + ], + "angle": 0, + "content": "[30] Ananya Kumar, Aditi Raghunathan, Robbie Jones, Tengyu Ma, and Percy Liang. Fine-tuning can distort pretrained features and underperform out-of-distribution. In ICLR, 2022. 2, 6, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.18, + 0.482, + 0.222 + ], + "angle": 0, + "content": "[31] Chenming Li, Daoan Zhang, Wenjian Huang, and Jianguo Zhang. Cross contrasting feature perturbation for domain generalization. In ICCV, pages 1327-1337, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.224, + 0.482, + 0.266 + ], + "angle": 0, + "content": "[32] Da Li, Yongxin Yang, Yi-Zhe Song, and Timothy M Hospedales. Deeper, broader and artier domain generalization. In ICCV, pages 5542-5550, 2017. 6, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.268, + 0.482, + 0.308 + ], + "angle": 0, + "content": "[33] Haoliang Li, Sinno Jialin Pan, Shiqi Wang, and Alex C Kot. Domain generalization with adversarial feature learning. In CVPR, pages 5400-5409, 2018. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.311, + 0.482, + 0.366 + ], + "angle": 0, + "content": "[34] Junnan Li, Dongxu Li, Caiming Xiong, and Steven Hoi. Blip: Bootstrapping language-image pre-training for unified vision-language understanding and generation. In ICML, pages 12888-12900, 2022. 1, 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.368, + 0.482, + 0.424 + ], + "angle": 0, + "content": "[35] Junnan Li, Dongxu Li, Silvio Savarese, and Steven Hoi. Blip-2: Bootstrapping language-image pre-training with frozen image encoders and large language models. In ICML, pages 19730–19742, 2023. 1, 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.426, + 0.482, + 0.468 + ], + "angle": 0, + "content": "[36] Xiaotong Li, Yongxing Dai, Yixiao Ge, Jun Liu, Ying Shan, and Ling-Yu Duan. Uncertainty modeling for out-of-distribution generalization. In ICLR, 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.47, + 0.482, + 0.511 + ], + "angle": 0, + "content": "[37] Ya Li, Mingming Gong, Xinmei Tian, Tongliang Liu, and Dacheng Tao. Domain generalization via conditional invariant representations. In AAAI, 2018. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.513, + 0.482, + 0.568 + ], + "angle": 0, + "content": "[38] Ya Li, Xinmei Tian, Mingming Gong, Yajing Liu, Tongliang Liu, Kun Zhang, and Dacheng Tao. Deep domain generalization via conditional invariant adversarial networks. In ECCV, pages 624-639, 2018. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.571, + 0.482, + 0.612 + ], + "angle": 0, + "content": "[39] Krikamol Muandet, David Balduzzi, and Bernhard Schölkopf. Domain generalization via invariant feature representation. In ICML, pages 10-18, 2013. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.614, + 0.482, + 0.655 + ], + "angle": 0, + "content": "[40] Hyeonseob Nam, HyunJae Lee, Jongchan Park, Wonjun Yoon, and Donggeun Yoo. Reducing domain gap by reducing style bias. In CVPR, pages 8690-8699, 2021. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.657, + 0.482, + 0.699 + ], + "angle": 0, + "content": "[41] Hongjing Niu, Hanting Li, Feng Zhao, and Bin Li. Domain-unified prompt representations for source-free domain generalization. arXiv preprint arXiv:2209.14926, 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.701, + 0.482, + 0.742 + ], + "angle": 0, + "content": "[42] Oren Nuriel, Sagie Benaim, and Lior Wolf. Permuted adain: Reducing the bias towards global statistics in image classification. In CVPR, pages 9482-9491, 2021. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.744, + 0.482, + 0.8 + ], + "angle": 0, + "content": "[43] Changdae Oh, Junhyuk So, Hoyoon Byun, YongTaek Lim, Minchul Shin, Jong-June Jeon, and Kyungwoo Song. Geodesic multi-modal mixup for robust fine-tuning. In NeurIPS, 2024. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.802, + 0.482, + 0.844 + ], + "angle": 0, + "content": "[44] Xingchao Peng, Qinxun Bai, Xide Xia, Zijun Huang, Kate Saenko, and Bo Wang. Moment matching for multi-source domain adaptation. In ICCV, pages 1406-1415, 2019. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.846, + 0.482, + 0.9 + ], + "angle": 0, + "content": "[45] Zelin Peng, Zhengqin Xu, Zhilin Zeng, Lingxi Xie, Qi Tian, and Wei Shen. Parameter efficient fine-tuning via cross block orchestration for segment anything model. In CVPR, pages 3743-3752, 2024. 2" + }, + { + "type": "list", + "bbox": [ + 0.093, + 0.092, + 0.483, + 0.9 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.092, + 0.905, + 0.134 + ], + "angle": 0, + "content": "[46] Shi Pu, Kaili Zhao, and Mao Zheng. Alignment-uniformity aware representation learning for zero-shot video classification. In CVPR, pages 19968-19977, 2022. 1, 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.136, + 0.905, + 0.178 + ], + "angle": 0, + "content": "[47] Daiqing Qi, Handong Zhao, Aidong Zhang, and Sheng Li. Generalizing to unseen domains via text-guided augmentation. In ECCV, pages 285-300, 2024. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.179, + 0.905, + 0.22 + ], + "angle": 0, + "content": "[48] Minghan Qin, Wanhua Li, Jiawei Zhou, Haoqian Wang, and Hanspeter Pfister. Langsplat: 3d language gaussian splatting. In CVPR, pages 20051-20060, 2024. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.222, + 0.905, + 0.29 + ], + "angle": 0, + "content": "[49] Alec Radford, Jong Wook Kim, Chris Hallacy, Aditya Ramesh, Gabriel Goh, Sandhini Agarwal, Girish Sastry, Amanda Askell, Pamela Mishkin, Jack Clark, et al. Learning transferable visual models from natural language supervision. In ICML, pages 8748-8763, 2021. 1, 2, 3, 6, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.292, + 0.905, + 0.346 + ], + "angle": 0, + "content": "[50] Yang Shu, Xingzhuo Guo, Jialong Wu, Ximei Wang, Jianmin Wang, and Mingsheng Long. Clipood: Generalizing clip to out-of-distributions. In ICML, pages 31716-31731, 2023. 1, 2, 6, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.349, + 0.905, + 0.404 + ], + "angle": 0, + "content": "[51] Mainak Singha, Ankit Jha, Shirsha Bose, Ashwin Nair, Moloud Abdar, and Biplab Banerjee. Unknown prompt the only lacuna: Unveiling clip's potential for open domain generalization. In CVPR, pages 13309-13319, 2024. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.405, + 0.905, + 0.447 + ], + "angle": 0, + "content": "[52] Quan Sun, Yuxin Fang, Ledell Wu, Xinlong Wang, and Yue Cao. Eva-clip: Improved training techniques for clip at scale. arXiv preprint arXiv:2303.15389, 2023. 1, 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.448, + 0.905, + 0.489 + ], + "angle": 0, + "content": "[53] Vladimir N Vapnik. An overview of statistical learning theory. IEEE transactions on neural networks, 10(5):988-999, 1999. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.491, + 0.905, + 0.545 + ], + "angle": 0, + "content": "[54] Ashish Vaswani, Noam Shazeer, Niki Parmar, Jakob Uszkoreit, Llion Jones, Aidan N Gomez, Lukasz Kaiser, and Illia Polosukhin. Attention is all you need. In NeurIPS, pages 5998-6008, 2017. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.547, + 0.905, + 0.602 + ], + "angle": 0, + "content": "[55] Hemanth Venkateswara, Jose Eusebio, Shayok Chakraborty, and Sethuraman Panchanathan. Deep hashing network for unsupervised domain adaptation. In CVPR, pages 5018-5027, 2017. 6, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.604, + 0.905, + 0.672 + ], + "angle": 0, + "content": "[56] Mengzhu Wang, Yuehua Liu, Jianlong Yuan, Shanshan Wang, Zhibin Wang, and Wei Wang. Inter-class and inter-domain semantic augmentation for domain generalization. IEEE Transactions on Image Processing, 33:1338-1347, 2024. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.674, + 0.905, + 0.716 + ], + "angle": 0, + "content": "[57] Pengfei Wang, Zhaoxiang Zhang, Zhen Lei, and Lei Zhang. Sharpness-aware gradient matching for domain generalization. In CVPR, pages 3769-3778, 2023. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.718, + 0.905, + 0.759 + ], + "angle": 0, + "content": "[58] Tongzhou Wang and Phillip Isola. Understanding contrastive representation learning through alignment and uniformity on the hypersphere. In ICML, pages 9929-9939, 2020. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.76, + 0.905, + 0.8 + ], + "angle": 0, + "content": "[59] Yufei Wang, Haoliang Li, and Alex C Kot. Heterogeneous domain generalization via domain mixup. In ICASSP, pages 3622-3626, 2020. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.802, + 0.905, + 0.871 + ], + "angle": 0, + "content": "[60] Mitchell Wortsman, Gabriel Ilharco, Jong Wook Kim, Mike Li, Simon Kornblith, Rebecca Roelofs, Raphael Gontijo Lopes, Hannaneh Hajishirzi, Ali Farhadi, Hongseok Namkoong, et al. Robust fine-tuning of zero-shot models. In CVPR, pages 7959-7971, 2022. 1, 2, 6, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.873, + 0.905, + 0.903 + ], + "angle": 0, + "content": "[61] Sitong Wu, Haoru Tan, Zhuotao Tian, Yukang Chen, Xiaojuan Qi, and Jiaya Jia. Saco loss: Sample-wise affinity con" + }, + { + "type": "list", + "bbox": [ + 0.517, + 0.092, + 0.905, + 0.903 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.946, + 0.516, + 0.956 + ], + "angle": 0, + "content": "9568" + } + ], + [ + { + "type": "ref_text", + "bbox": [ + 0.125, + 0.092, + 0.482, + 0.119 + ], + "angle": 0, + "content": "sistency for vision-language pre-training. In CVPR, pages 27358-27369, 2024. 1, 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.122, + 0.483, + 0.177 + ], + "angle": 0, + "content": "[62] Xin Zhang, Shixiang Shane Gu, Yutaka Matsuo, and Yusuke Iwasawa. Domain prompt learning for efficiently adapting clip to unseen domains. Transactions of the Japanese Society for Artificial Intelligence, 38(6), 2023. 3, 6, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.179, + 0.482, + 0.22 + ], + "angle": 0, + "content": "[63] Xingxuan Zhang, Renzhe Xu, Han Yu, Yancheng Dong, Pengfei Tian, and Peng Cui. Flatness-aware minimization for domain generalization. In ICCV, pages 5189-5202, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.222, + 0.482, + 0.262 + ], + "angle": 0, + "content": "[64] Kaiyang Zhou, Jingkang Yang, Chen Change Loy, and Zwei Liu. Conditional prompt learning for vision-language models. In CVPR, pages 16816-16825, 2022. 2, 3, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.264, + 0.482, + 0.318 + ], + "angle": 0, + "content": "[65] Kaiyang Zhou, Jingkang Yang, Chen Change Loy, and Ziwei Liu. Learning to prompt for vision-language models. International Journal of Computer Vision, 130(9):2337-2348, 2022. 2, 3, 7" + }, + { + "type": "list", + "bbox": [ + 0.093, + 0.092, + 0.483, + 0.318 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.946, + 0.516, + 0.956 + ], + "angle": 0, + "content": "9569" + } + ] +] \ No newline at end of file diff --git a/2025/Domain Generalization in CLIP via Learning with Diverse Text Prompts/51839507-6fb2-45c2-a66c-0644bd7936ed_origin.pdf b/2025/Domain Generalization in CLIP via Learning with Diverse Text Prompts/51839507-6fb2-45c2-a66c-0644bd7936ed_origin.pdf new file mode 100644 index 0000000000000000000000000000000000000000..6f931980fd03bb16c965758ffdf62818000c2c95 --- /dev/null +++ b/2025/Domain Generalization in CLIP via Learning with Diverse Text Prompts/51839507-6fb2-45c2-a66c-0644bd7936ed_origin.pdf @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:b61a39960d2c98c8262cd5896a7c7cf618ae7749174d58e2887b811d8d91cd77 +size 1211852 diff --git a/2025/Domain Generalization in CLIP via Learning with Diverse Text Prompts/full.md b/2025/Domain Generalization in CLIP via Learning with Diverse Text Prompts/full.md new file mode 100644 index 0000000000000000000000000000000000000000..c1e4566531f9f303993d22d4807c7ebd286fdc85 --- /dev/null +++ b/2025/Domain Generalization in CLIP via Learning with Diverse Text Prompts/full.md @@ -0,0 +1,368 @@ +# Domain Generalization in CLIP via Learning with Diverse Text Prompts + +Changsong Wen, Zelin Peng, Yu Huang, Xiaokang Yang, Wei Shen MoE Key Lab of Artificial Intelligence, AI Institute, Shanghai Jiao Tong University + +{changsong, zelin.peng, yellowfish, xkyang, wei.shen}@sjtu.edu.cn + +# Abstract + +Domain generalization (DG) aims to train a model on source domains that can generalize well to unseen domains. Recent advances in Vision-Language Models (VLMs), such as CLIP, exhibit remarkable generalization capabilities across a wide range of data distributions, benefiting tasks like DG. However, CLIP is pre-trained by aligning images with their descriptions, which inevitably captures domain-specific details. Moreover, adapting CLIP to source domains with limited feature diversity introduces bias. These limitations hinder the model's ability to generalize across domains. In this paper, we propose a new DG approach by learning with diverse text prompts. These text prompts incorporate varied contexts to imitate different domains, enabling DG model to learn domain-invariant features. The text prompts guide DG model learning in three aspects: feature suppression, which uses these prompts to identify domain-sensitive features and suppress them; feature consistency, which ensures the model's features are robust to domain variations imitated by the diverse prompts; and feature diversification, which diversifies features based on the prompts to mitigate bias. Experimental results show that our approach improves domain generalization performance on five datasets from the DomainBed benchmark, achieving state-of-the-art results. + +# 1. Introduction + +In the complex real-world applications, deep learning models are expected to be capable of recognizing data from a wide range of distributions. Within the past few years, researchers are dedicated to investigating domain generalization (DG) [6, 7, 9, 24, 25, 27], which empowers the model to learn robust features from multiple domains and exhibit good performance across various data distributions. Recent advances in foundation Vision-Language Models (VLMs) [34, 35, 52, 61], such as CLIP [49], are trained on large-scale data sourced from the web with a wide variety + +![](images/2745900baa7e7c20eff0effcd1eedfb1ebef11f5cf1d26f966d7308491e855b3.jpg) +Figure 1. Limitations with existing methods. As shown on the left, CLIP's image encoder is pre-trained to align with textual descriptions, which inevitably captures domain-specific visual details, resulting in perturbations in domain generalization. Additionally, as shown on the right, due to the limited feature diversity in the source domains, adapting CLIP with source domains will introduce bias. + +of data distributions. These VLMs possess remarkable generalization capabilities, and boost many open-world recognition tasks, including out-of-distribution learning [50] and domain generalization [1, 10, 51]. + +Existing DG studies [20, 23, 27, 50, 60] leverage CLIP's pre-trained weights with techniques like parameter regularization [23] and weight space ensembling [6, 27, 60] to adapt the model for downstream datasets, while preserving generalization ability. Other recent DG works [1, 26] propose using knowledge distillation to transfer CLIP's knowledge into DG models. However, as shown in Fig. 1, CLIP [49] aligns images with their descriptions during pretraining, which inevitably contain domain-specific details, impacting existing DG models. Additionally, since the training images are limited to a few source domains in DG, adapting CLIP within these images leads the model to emphasize the features specific to the source domains. This results in a lack of feature diversity [43, 46, 58], which com + +promises the model's ability to generalize to new domains. + +Erasing domain-sensitive details and mitigating bias benefit from learning domain-invariant representations for each class. In this paper, we propose a new DG approach to learn domain-invariant class features by leveraging a diverse set of text prompts. Based on the multi-modal alignment capabilities of CLIP, we use these text prompts to imitate multiple domains and effectively guide the DG learning process. A wide range of domains provides diverse perspectives on each class, allowing the model to learn features that are robust across varying contexts, thereby promoting domain-invariant feature learning. + +Diverse text prompts support three key aspects of our approach to learning domain-invariant features: they facilitate domain-sensitive feature suppression, ensure feature consistency across domain variations, and enable feature diversification to reduce bias. To be specific, we leverage the prompts recommended by original CLIP [49] to approximate different domains, such as "a painting of a {class}" or "a {class} in a video game". We first identify channels in the features activated by specific text prompts. These channels are considered highly relevant to the domain context of the prompt, thus we suppress their activation values. Next, we compare predictions under perturbations of diverse text prompts, and employ a consistency constraint to ensure the model is robust to domain variations while maintaining stable predictions. Subsequently, we decompose the component of the image features orthogonal to the text prompts and impose diversification on the orthogonal component. This approach enables a more diverse distribution of image features without impacting classification. + +To comprehensively evaluate our method, we perform extensive experiments on the widely used DomainBed benchmark [21]. The experiment results demonstrate the superior performance and robustness of our proposed approach, achieving state-of-the-art results across five domain generalization datasets. + +# 2. Related Work + +# 2.1. Large-scale Vision Language Models + +Large-scale Vision Language Models (VLMs) [28, 49, 52, 61], trained on extensive image-text data from the internet, demonstrate strong robustness to diverse data distributions. These VLMs exhibit remarkable generalization capabilities, serving as fundamental models in the field of domain generalization. CLIP [49] is one of the most prevalent approaches among VLMs. It is trained by aligning image-text pairs through contrastive learning, where the model learns to associate corresponding images and texts from large-scale datasets by maximizing their similarity in a shared embedding space. Among the past few years, many efforts have been made to improve the CLIP-like VLMs by designing + +better loss function [61], mitigating the impact of noisy web data [34, 35], and developing optimization techniques [52]. Another line of works focus on leveraging CLIP's generalization capabilities and its cross-modal alignment strengths to address downstream tasks such as zero-shot classification [29, 64], domain generalization [1, 26], and 2D/3D dense prediction [14, 45, 48]. + +# 2.2. Domain Generalization + +Domain generalization [3, 8, 10, 18, 31, 40, 53, 59, 63] seeks to develop models capable of adapting to tasks from unseen domains without additional training. Conventional domain generalization methods encompass approaches such as learning domain-invariant features [13, 22, 37, 39], data augmentation [36, 42, 56], and adversarial learning [33, 38] to improve generalization to unseen domains. + +Recently, VLMs are widely utilized as foundational models in domain generalization [1, 20, 26, 50, 60]. Some approaches fine-tune CLIP model on downstream DG datasets. However, since these DG datasets (0.01 to $0.5\mathrm{M}$ ) are significantly smaller than CLIP's training set (400M), directly fine-tuning disrupts CLIP's parameter space [30], leading to degraded generalization performance. To mitigate the limitations, many impressive works [6, 30, 60] propose parameter ensemble methods between CLIP and the fine-tuned model to preserve generalization ability. Other works adopt CLIP for DG downstream tasks by designing loss functions during fine-tuning [20, 50], or employing regularization in parameter-efficient fine-tuning method [23]. Some knowledge distillation-based methods [1, 26] align the features of DG models with the output features of CLIP, transferring CLIP's generalization ability to DG models. + +DomainDrop [22] mitigates model feature sensitivity to different domains by suppressing channel-wise feature variance across source domains. PromptStyler [11] generates diverse templates with style constraints and content consistency to train a text classifier. The text classifier is directly applied to visual domain generalization. LADS [16] and TEAM [47] employ text prompts for facilitating image feature augmentation. CAE [41] manually design a domain bank to obtain domain-unified text prompts for a stable prediction across domains. Different from these works that recognize the importance of text prompts, our approach enables direct optimization of the visual encoder by imitating multiple domain variations through a set of text prompts. + +# 2.3. Prompt Engineering + +In CLIP, the design of prompts promotes the model to effectively capture the key information within the image. Recent prompt learning methods [29, 64, 65] are proposed to optimize the prompt as trainable parameters, allowing the model to dynamically adjust the prompt based on tasks. + +CoOp [65] enhances CLIP's adaptability for visual recognition tasks by using learnable context vectors as language prompts. CoCoOp [64] incorporates image features as conditional prompts to better capture context. MaPLe [29] introduces a multimodal prompt learning approach, finetuning both image and text encoders simultaneously for improved performance. + +Recently, the domain generalization method DPL [62] employs an MLP to capture domain-specific image features as context for the learnable prompt. SPG [4] designs a prompt generator with domain knowledge to generate soft prompts for the images from target domain. These methods [4, 19, 62] primarily focus on incorporating image context into the text encoder to achieve multimodal alignment in downstream tasks. Differently, we optimize the image encoder through prompts from diverse templates. These prompts effectively imitate different domains while preventing the parameters of learnable prompts from being influenced by source domain-specific information during training. + +# 3. Preliminary + +Problem Formulation. Domain generalization aims to train a model that can generalize well to unseen target domains, even when the training data comes from multiple source domains with potentially different data distributions. Specifically, the model is trained using data from $o$ different source domains $\{D_i^s\}_{i=1}^o$ , where each source domain $D_i^s = \{(x_i, y_i)\}_{i=1}^u$ consists of input data $x_i$ and corresponding class labels $y_i$ , with $u$ representing the number of samples in the source domain. + +The model's generalization performance is evaluated on the target domain $D^{t}$ , which shares the same set of class labels as the source domains $\{D_{i}^{s}\}_{i = 1}^{o}$ . The joint distribution of input data $x$ and labels $y$ in the target domain differs from that in the source domains. + +Utilizing CLIP for domain generalization. CLIP's strong generalization capability stems from its multimodal training on diverse internet data covering a wide range of concepts and contexts. As a prominent vision-language model, CLIP comprises an image encoder $F_{I}(\cdot)$ and a text encoder $F_{T}(\cdot)$ . Both image and text encoders adopt a transformer architecture [15, 54] to generate their representations through self-attention. + +CLIP employs contrastive learning to align images and their descriptive text during pre-training. It associates visual and textual features by maximizing the similarity of the aligned image-text pairs while reducing the similarity of other ones: + +$$ +\mathcal {L} _ {\mathrm {C L I P}} = - \frac {1}{N} \sum_ {i = 1} ^ {N} \log \frac {\exp \left(\mathrm {S} \left(I _ {i} , A _ {i}\right) / \tau\right)}{\sum_ {j = 1} ^ {N} \exp \left(\mathrm {S} \left(I _ {i} , A _ {j}\right) / \tau\right)}, \tag {1} +$$ + +where $I$ and $A$ represent the input image and caption, $i$ , $j$ are indexes of the inputs, $S$ means similarity matrices, such as cosine similarity. $\tau$ is a temperature parameter used to adjust the smoothness of the similarity distribution. + +Baseline of our approach. In the conventional domain generalization methods [1, 23, 29], "a photo of a $\{class\}$ " is encoded by CLIP's text encoder as text prompt $\{P_i\}_{i=1}^c$ , representing the main concepts of $c$ different classes. Leveraging CLIP's remarkable multi-modal alignment capabilities between images and text, the text prompts for different classes are treated as domain-invariant representations of the image features [1, 26]. We adopt a basic baseline [1] to adapt CLIP for downstream domain generalization tasks by simultaneously distilling image and text features to the DG model. + +$$ +\mathcal {L} _ {\mathrm {K D}} = \lambda \cdot \cos \left(F _ {I} \left(x _ {i}\right), f _ {I} \left(x _ {i}\right)\right) + (1 - \lambda) \cdot \cos \left(F _ {T} \left(y _ {i}\right), f _ {I} \left(x _ {i}\right)\right), \tag {2} +$$ + +where $F_{I}(\cdot)$ and $F_{T}(\cdot)$ are CLIP's image and text encoders, $f_{I}(\cdot)$ is the DG model. Our approach differs slightly in that we ensemble multiple prompts used to imitate different domains during the distillation process. This has a minor direct impact on performance but is more intuitive to the design of our overall framework, ensuring stability across domain variations. The classification result $\hat{y}$ is obtained by choosing the ensembled text prompt with the highest similarity to the $i$ -th image. + +# 4. Methodology + +In this section, we introduce our proposed learning with diverse prompts paradigm for learning domain-invariant features in DG. Initially, we filter out unreliable prompts based on the prediction confidence. Next, we propose a Domain-sensitive Channel Suppression (DCS) method to identify and suppress domain-sensitive features. Moreover, a Multi-domain Consistency Constraint (MCC) approach is designed to make the features robust to variations across domains. Then, we apply Orthogonal Uniformity Regularization (OUR) to diversify features in the DG model. + +# 4.1. Entropy-based Prompt Filtering + +We imitate different domains by using the text templates $\{T_j\}_{j=1}^n$ recommended in CLIP [49], such as "a drawing of a {class}", $n$ is the number of templates. These prompts are generated by encoding the combination of the $j$ -th text template $T_j$ and the $i$ -th class label $y_i$ by CLIP text encoder: + +$$ +P _ {i} ^ {j} = F _ {T} \left(y _ {i}, T _ {j}\right), \tag {3} +$$ + +where $P \in \mathbb{R}^{n \times c \times D}$ , $c$ is the number of classes, $D$ represents the feature dimensions, $i$ and $j$ indicate the index of the classes and prompt templates. These prompts employ diverse contexts to describe each class, such as a drawing, + +![](images/d7d94cbb7475280d264b29faded19619031e0a3df2471fb30d2a5b37785fc980.jpg) +Figure 2. Overview of our proposed method. To enforce the DG model to learn domain-invariant class features, we imitate multiple domains by text prompts. Three components of our method, including feature suppression, feature consistency, and feature diversification, are achieved through Domain-sensitive Channel Suppression (DCS), Multi-domain Consistency Constraint (MCC), and Orthogonal Uniformity Regularization (OUR) modules. The blue line represents the process during inference. + +a sculpture, or a rendering, etc. Thus, these prompts can be naturally regarded as descriptions of the class from multiple domains. The DG model $f_{I}(\cdot)$ is regularized to extract domain-invariant features that are robust to prompts from different domains. + +Within the prompts provided by CLIP, we first discard a few unreliable ones based on the prediction confidence for each input [2], as they negatively affect the feature channel suppression and feature consistency calculation, which rely on accurately identifying relevant features. In the Entropy-based Prompt Filtering (EPF), we first extract the encoded image features $I_{i} = f_{I}(x_{i})$ for the input image $x_{i}$ . For the $j$ -th template, the predicted probability for the $i$ -th class is calculated as: + +$$ +\hat {d} _ {i} ^ {j} = \frac {\exp \left(\cos \left(I _ {i} , P _ {i} ^ {j}\right)\right)}{\sum_ {k = 1} ^ {c} \exp \left(\cos \left(I _ {i} , P _ {k} ^ {j}\right)\right)}, \tag {4} +$$ + +We denote that $\hat{d}_i^j$ is a constant and $\hat{d}^j = \{\hat{d}_i^j\}_{i=1}^c$ is a predicted distribution for the $j$ -th template. + +Entropy quantifies the uncertainty of a probability distribution. If the model output has high confidence, the distribution typically concentrates on a specific class, resulting + +in low entropy. Conversely, if the model has similar confidence across multiple classes, the distribution is more uniform, leading to higher entropy. Therefore, we calculate the entropy for each template's prediction: + +$$ +\mathrm {H} \left(\hat {d} ^ {j}\right) = - \sum_ {i = 1} ^ {c} \hat {d} _ {i} ^ {j} \log \left(\hat {d} _ {i} ^ {j}\right). \tag {5} +$$ + +Then, we eliminate the unreliable prompts and select the $m$ prompts with the lowest entropy for each input: + +$$ +\left\{T _ {1}, T _ {2}, \dots , T _ {m} \right\} = \arg \min _ {j \in \{1 \dots , n \}} \mathrm {H} \left(\hat {d} ^ {j}\right), \tag {6} +$$ + +# 4.2. Domain-sensitive Channel Suppression + +The image features should focus on discriminative features between classes and be robust to prompts from different domains. However, some domain-sensitive channels encode variations specific to the domain context of prompt, introducing perturbations. To address this, we propose Domain-sensitive Channel Suppression (DCS) to select and suppress these channels, ensuring the extracted image embedding captures domain-invariant class features. + +Specifically, we utilize gradient attribution [12] to quantify each channel's contribution in the image features when + +calculating their similarity with different prompts. We simultaneously consider the model's sensitivity to each channel (i.e., the gradient) and the actual activation value of that channel to measure each image channel's contribution to recognition with $P_{i}^{j}$ : + +$$ +\alpha_ {k j} = \left| \frac {\partial \cos \left(I _ {i} , P _ {i} ^ {j}\right)}{\partial I _ {i} ^ {k}} \right| \cdot I _ {i} ^ {k}, \tag {7} +$$ + +where $P_{i}^{j}$ means the prompt corresponding to the $i$ -th image's label with $T_{j}$ as the template. $\alpha_{kj}$ is a constant that represents the contribution of the $k$ -th image channel to the similarity calculation with the $j$ -th prompt $P_{i}^{j}$ . For each image feature, we normalize it through the channel dimensions: + +$$ +\bar {\alpha} _ {k j} = \frac {\alpha_ {k j}}{\frac {1}{D} \sqrt {\sum_ {k = 1} ^ {D} \alpha_ {k j} ^ {2}}}, \tag {8} +$$ + +Then, we calculate the z-score among different prompts: + +$$ +\alpha_ {k j} ^ {\mathrm {z - s c o r e}} = \frac {\overline {{\alpha}} _ {k j} - \mu}{\sigma}, \tag {9} +$$ + +$\mu$ and $\sigma$ are the average of $\overline{\alpha}_{kj}$ across all templates, which are calculated as: + +$$ +\mu = \frac {1}{m} \sum_ {j = 1} ^ {m} \bar {\alpha} _ {k j} \tag {10} +$$ + +$$ +\sigma = \sqrt {\frac {1}{m} \sum_ {j = 1} ^ {m} \left(\bar {\alpha} _ {k j} - \mu\right) ^ {2}} \tag {11} +$$ + +We check the z-score values of each channel across different prompts. A high z-score indicates that a channel is activated by the context provided by the specific prompt, showing a strong correlation with domain-specific information. We apply a mask to such image feature channels before calculating the loss: + +$$ +\mathbf {M} _ {D} = \left\{ \begin{array}{l l} 1, & \text {i f} \max _ {k} \alpha_ {k j} ^ {\mathrm {z - s c o r e}} > \beta \\ 0, & \text {o t h e r w i s e} \end{array} \right. \tag {12} +$$ + +The mask $\mathbf{M}_D\in \mathbb{R}^D$ is directly employed on image features by the element-wise multiplication $I_i^k = \mathbf{M}_D\odot I_i$ . We suppress these domain-sensitive features by applying $L_{1}$ regularization: + +$$ +\mathcal {L} _ {r e g} = \gamma \sum_ {i = 1} ^ {B} \sum_ {k = 1} ^ {D} \left| I _ {i} ^ {k} \right|, \tag {13} +$$ + +where $B$ represents the batch size, and $\gamma$ is the weight of loss, $I_{i}^{k}$ is the masked image feature. + +# 4.3. Multi-domain Consistency Constraint + +We further introduce Multi-domain Consistency Constraint (MCC) on the predicted distribution between the image and varied prompts. We calculate the similarity between image features and different prompts for classification. The KL loss function is then applied to ensure that the predicted distributions remain consistent between different prompts, enhancing the robustness of the image encoder to domain variations. Specifically, we use Eq. (4) to obtain the predicted distributions for different templates $\{\hat{d}^j\}_{j=1}^m$ , and apply KL loss to ensure consistency: + +$$ +\operatorname {K L} \left(\hat {d} ^ {j} \| \hat {d} ^ {k}\right) = \sum_ {i = 1} ^ {c} \hat {d} _ {i} ^ {j} \log \frac {\hat {d} _ {i} ^ {j}}{\hat {d} _ {i} ^ {k}} \tag {14} +$$ + +$$ +\mathcal {L} _ {\mathrm {K L}} = \frac {1}{m (m - 1)} \sum_ {j = 1} ^ {m} \sum_ {\substack {k = 1 \\ k \neq t}} ^ {m} \mathrm {K L} \left(\hat {d} ^ {j} \| \hat {d} ^ {k}\right) \tag{15} +$$ + +The KL loss compares the prediction with prompts from diverse domains and enforces consistency, guiding the learned features to focus on domain-invariant information. + +# 4.4. Orthogonal Uniformity Regularization + +Due to the limited diversity of image source domains, learning with these images biases CLIP toward the feature distribution of these source domains. To mitigate this bias, we further diversify the normalized image features to achieve a more uniform distribution in the hyperspherical space. Uniformity of the features increases the possibility that unseen features are positioned near seen features [46], thereby enhancing the model's generalization through feature diversification. + +However, it is not feasible to directly diversify the image features. As mentioned in Sec. 3, we use an ensembled prompt $e_i$ in classification, which can be formulated as: + +$$ +e _ {i} = \frac {1}{n} \sum_ {j = 1} ^ {n} P _ {i} ^ {j}. \tag {16} +$$ + +We compute the similarity between the ensembled prompt $e_i$ and the image feature $I_i$ for classification: + +$$ +\hat {y} = \operatorname {a r g m a x} _ {j} \cos \left(I _ {i}, e _ {j}\right). \tag {17} +$$ + +Therefore, as demonstrated in Eq. (17), directly applying uniformity constraints on image features can impact the similarity calculation between image features and prompts. + +To address this, we decompose $I_{i}$ into two components, one in the direction of $e_{i}$ and the other in the orthogonal direction: + +$$ +I _ {i} ^ {\parallel} = \frac {I _ {i} \cdot e _ {i}}{\| e _ {i} \| ^ {2}} \cdot e _ {i}, \tag {18} +$$ + +Table 1. Comparison with the state-of-the-art methods on the five domain generalization datasets from DomainBed benchmark. We report the averaged leave-one-domain-out experimental results with ViT-B/16 as the backbone. Numbers in bold represent the best performance among all the methods. T and F indicate whether the CLIP parameters are trained or frozen during adaptation to DG. + +
MethodsCLIPOHTerraIncVLCSPACSDomainNetAvg
Fine-tuning
SWAD (NeurIPS'2021) [6]T81.0142.9279.1391.3557.9270.47
LP-FT (ICLR'2022) [30]T81.1747.2680.8892.9257.0471.85
WiSE-FT (CVPR'2022) [60]T86.3254.5082.8897.2958.0175.80
MIRO (ECCV'2022) [7]T84.8059.3082.3096.4460.4776.66
DART (CVPR'2023) [27]T80.9351.2480.3893.4359.3273.06
SAGM (CVPR'2023) [57]T83.4058.6482.0594.3159.0575.49
FLYP (CVPR'2023) [20]T82.7633.2566.6478.5357.4163.72
CLIPood (ICML'2023) [50]T83.3146.2877.1993.1657.7871.55
Knowledge Distillation
RISE (ICCV'2023) [26]T78.3949.6180.6293.2555.3771.45
VL2V-SD (CVPR'2024) [1]T87.3858.5483.2596.6862.7977.73
Prompt-learning
DPL (TJSAI'2023) [62]F84.2052.6084.3097.3059.5075.60
SPG (ECCV'2024) [4]F83.6050.2082.4097.0060.1074.70
Parameter-efficient Fine-tuning
LPD (ECCV'2024) [23]F84.2157.3083.2096.5559.3076.11
OursT87.6563.2784.7997.0363.1179.17
+ +$$ +I _ {i} ^ {\perp} = I _ {i} - I _ {i} ^ {\parallel}, \tag {19} +$$ + +where $I_{i}^{\parallel}$ and $I_{i}^{\perp}$ are the decomposed parallel and orthogonal components, respectively. We define a pairwise uniformity loss based on the dot product between the orthogonal components of different image features within the batch: + +$$ +\mathcal {L} _ {\mathrm {O U}} = \frac {1}{B (B - 1)} \sum_ {i = 1} ^ {B} \sum_ {\substack {j = 1 \\ j \neq i}} ^ {B} \left(I _ {i} ^ {\perp} \cdot I _ {j} ^ {\perp}\right) ^ {2} \tag{20} +$$ + +Since $e_i$ is computed independently for each image, the orthogonal component is also based on different images. Applying the uniformity constraint solely on the orthogonal direction avoids interference with the similarity calculation between the image features and the prompt. + +Finally, combining the baseline with Domain-sensitive Channel Suppression, Multi-domain Consistency Constraint, and Orthogonal Uniformity Regularization yields the total loss used during training, which is formulated as: + +$$ +\mathcal {L} _ {\text {t o t a l}} = \mathcal {L} _ {K D} + \mathcal {L} _ {\text {r e g}} + \mathcal {L} _ {K L} + \mathcal {L} _ {O U}, \tag {21} +$$ + +# 5. Experiments + +# 5.1. Datasets and Evaluation + +We conduct evaluation experiments on five datasets from the DomainBed benchmark [21], following previous methods [1, 22, 26]. PACS [32] consists of 9,991 samples spanning four domains: {art, cartoons, photos, sketches}, and + +includes a total of 7 classes. VLCS [17] contains 10,729 examples, 5 classes from four datasets as different domains: {VOC2007, LabelMe, Caltech, SUN}. Terra Incognita (TI) [5] contains 24,788 samples from 10 animal categories. These photos are captured across different locations {L100, L38, L43, L46}, acting as the four domains. Office Home (OH) [55] have four domains {art, clipart, product, real}, with 15,588 examples from 65 classes. DomainNet [44] is a large-scale cross-domain dataset containing 586,575 images from 6 different domains: {clipart, infographic, painting, quickdraw, real, sketch}. It covers 345 classes and is widely used to evaluate the performance of domain generalization algorithms. + +Following the protocols from previous domain generalization works [1, 11, 26], we split the datasets by using $80\%$ of the data for training and the remaining $20\%$ as the validation set. We adopt a leave-one-out experiment setup, where we alternately leave on domain for testing and train with the remaining domains. The average performance on the unseen domains is reported as the domain generalization results. + +# 5.2. Implementation Details + +We employ CLIP with ViT-B/16 as the backbone model and fine-tune it on each of the datasets. We fine-tune the model for 5,000 iterations on the PACS, VLCS, Office Home, and Terra Incognito datasets, and for 15,000 iterations on the DomainNet dataset. We use the $n = 80$ templates recommended in original CLIP [49] as prompts to imitate dif + +Table 2. Ablation analysis on the effectiveness of our method. EPS: entropy-based prompt selection, DCS: domain-sensitive channel suppression, MCC: multi-domain consistency constraint. OUR: orthogonal uniformity regularization, URIF*: uniformity regularization for image features, where * indicates it is not our component, but used as an additional comparison to OUR. + +
BaseEPSDCSMCCOURURIF*Avg
✓81.46
✓✓✓82.28
✓✓✓82.37
✓✓✓✓82.79
✓✓82.30
✓✓81.55
✓✓✓✓✓83.19
+ +ferent domains for each class. The threshold $\beta$ for the suppressed feature mask $\mathbf{M}_D$ is set dynamically for each sample. We suppress $20\%$ image feature channels to reduce domain-specific perturbation, and set the regularization term as $10^{-4}$ . We select $m = 40$ prompts with low entropy from 80 contextual prompts. Other experimental settings all follow our baseline model. SWAD [6] is a general approach that improves model robustness by averaging weights during fine-tuning DG models. For a fair comparison, we integrate SWAD into all methods that fine-tune CLIP's parameters, following the previous work [1]. + +# 5.3. Comparison with State-of-the-Arts + +We mainly compare with recent advances in domain generalization methods based on CLIP. We first compares our method with the following fine-tuning methods. SWAD [6], WiSE-FT [60], and DART [27] explore weight-ensembling methods to maintain CLIP's generalization ability. LPFT [30], FLYP [20], and CLIPood [50] investigate effective fine-tuning methods. The importance of text is highlighted in VL2V-SD [1] and RISE [26]. They incorporate text features into the DG model via knowledge distillation. DPL [62] and SPG [4] are prompt-learning methods that design learnable modules to extract context features from input images. LPD [23] explores a parameter-efficient fine-tuning approach for adapting CLIP in DG. We report the averaged leave-one-domain-out experimental results following the setting from previous works [1, 4]. + +As shown in Table 1, our approach achieves an average improvement of $1.4\%$ over SOTA methods across five datasets. Our method demonstrates a particularly notable $4.7\%$ improvement over SOTA on the TerraInc dataset [5]. As the images in TerraInc are sourced from wildlife photography, target objects occupy a small portion of the scene and contain substantial irrelevant contextual information. Our approach effectively reduces interference from such perturbation and achieves impressive improvement. + +Table 3. Performance comparison of our method with various prompt learning approaches across different datasets, including OH, TI, VLCS, and PACS. The average accuracy (Avg) across these datasets highlights the robustness and superior generalization capability of template generated prompts. + +
MethodsOHTIVLCSPACSAvg
CLIP [49]82.434.182.396.573.8
CoOP [65]83.046.880.896.476.8
Co-CoOP [64]83.445.380.396.776.4
Maple [29]83.450.282.296.578.1
DPL [62]84.252.684.397.379.6
SPG [4]83.650.282.497.078.3
Ours87.763.384.897.083.2
+ +# 5.4. Ablation Study + +Ablation of the main components. We conduct ablation studies to evaluate the effectiveness of each component of our approach. Following previous methods [1, 23], we report the results on four datasets: OH [55], TI [5], VLCS [17], and PACS [32]. 'Avg' denotes the average of the four datasets. As shown in Table 2, using DCS with EPS yields a notable performance increase and improves the accuracy to $82.28\%$ , highlighting the significance of suppressing the domain-specific channels. MCC also achieves an improvement, raising accuracy to $82.37\%$ by enforcing cross-domain consistency. Applying both DSC and MCC modules simultaneously to suppress the domain-specific context results in an average gain of $1.33\%$ . + +Moreover, incorporating uniformity regularization for the orthogonal component increases the average accuracy to $82.30\%$ , suggesting that capturing more diversified features enables the model to generalize better across domains by preventing an overemphasis on specific source domains. In contrast, applying uniformity regularization directly to image features (URIF*) without orthogonal decomposition yields a decrease to $81.55\%$ , demonstrating that directly regularizing image features may introduce perturbations during classification. Finally, the best performance is obtained by combining EPS, DSC, MCR, and OUR, reaching an average accuracy of $83.19\%$ . + +Discussions on the design of prompts. In our implementation, we imitate diverse domains by the prompts generated from 80 templates recommended by CLIP. In this section, we conduct experiments to analysis the advantages of our template-based prompts over the learnable prompts used in previous methods [4, 29, 62, 64]. By optimizing the learnable prompts during training, these model learn to adapt dynamically to different inputs. As shown in Table 3, although these methods show significant improvement compared to CLIP, our approach still outperforms the state-of-the-art prompt learning methods by $3.6\%$ . While they incorporate context of the image, the learnable prompts also + +Table 4. Ablations of comparison with knowledge distillation from the single manually designed prompt and the design of our method. EP represents the ensembled $n$ prompts for knowledge distillation. + +
PromptsOHTIVLCSPACSAvg
{class}86.559.583.796.681.6
a photo of a {class}87.458.583.396.781.5
a {domain} of a {class}87.260.783.996.982.2
EP87.159.983.696.581.9
Ours87.763.384.897.083.2
+ +introduces potential perturbations from specific source domains. Unlike these learnable prompts, our method leverages a set of templates without requiring domain-specific adaptation during training, enabling the model to generalize more robustly across diverse domains. + +Analysis on knowledge distillation from different designed prompts. In this part, we conduct experiments to demonstrate the advantages of using a diverse set of prompts compared to the carefully designed manual prompt. In our approach, the diverse set of prompts are ensembled and a knowledge distillation framework is used to transfer the ensembled prompts (EP) to the DG model. Based on the knowledge distillation baseline, we compare with three types of manually designed prompts, including {"class)", "a photo of a {class}", and "a {domain}" of a {class)". As shown in Table 4, a simple use of ensembled prompts provides slight gains, with $0.3\%$ and $0.4\%$ improvements compared to manual prompts {"class"} and "a photo of a {class}". Moreover, EP even performs worse than the carefully designed prompt: "a {domain}" of a {class)". However, a set of diverse prompts enables us to imitate multiple domains and learn domain-invariant features. Our overall method achieves consistently better performance across all four datasets, with notable gains in challenging cases like TI and VLCS. Such results demonstrate the effectiveness of our multi-domain imitation method. + +Hyperparameter analysis. Additionally, we investigate the hyperparameter for the number of selected templates $m$ and the ratio of channels suppressed during gradient attribution. Table 5 presents analysis on the number of prompts and its impact on the average performance across four datasets. We evaluate different values for $m$ , ranging from 10 to 80. The results show that increasing $m$ improves performance at first, when $m = 40$ , the performance reaches its maximum value $83.2\%$ . Using more templates ensures that the imitated domains contain more complex contexts, thus facilitating our model in learning robust features. However, further increases $m$ reduce the performance, with the average accuracy dropping to $83.0\%$ at $m = 60$ and $82.3\%$ at $m = 80$ . Since we do not design + +Table 5. Hyperparameter analysis on the number of prompts $m$ . + +
m1020406080
Avg82.782.983.283.082.3
+ +Table 6. Hyperparameter analysis on the ratio of channels suppressed during gradient attribution. + +
Ratio0.10.20.30.40.5
Avg82.383.282.081.781.6
+ +different templates for different datasets, utilizing all of the templates without selection introduces noise, leading to a drop in performance. Therefore, we set the hyperparameter $m$ as 40. + +Table 6 presents the analysis on the ratio of channels suppressed during gradient attribution and its effect on the overall performance across four datasets. We evaluate different suppression ratios, ranging from 0.1 to 0.5. The results show that increasing the suppression ratio improves performance at first, with the maximum average accuracy of $83.2\%$ achieved at a suppression ratio of 0.2. However, further increases in the suppression ratio slightly reduce accuracy, with the average performance dropping to $82.0\%$ , $81.7\%$ , and $81.6\%$ at 0.3, 0.4, and 0.5. The experiments demonstrate that excessive suppression filters out essential information, resulting in degraded performance. Thus, we set the suppression ratio to 0.2 as the optimal value for this hyperparameter. + +# 6. Conclusion + +We develop a new domain generalization approach by learning with diverse text prompts. We use these prompts to imitate multiple domains, which guide the DG model in learning domain-invariant features through multi-modal alignment in CLIP. These prompts contribute to three aspects, including suppressing domain-sensitive features, ensure feature consistency with varied domains, and enable feature diversification in the DG model. By incorporating these three terms, our approach reduces the CLIP-based DG model's tendency to capture domain-sensitive details and mitigates bias towards the source data distribution. Experimental results on five domain generalization datasets from the DomainBed benchmark demonstrate the effectiveness of our method. + +# 7. Acknowledgments + +This work was supported by the NSFC under Grant 62322604 and 62176159, and in part by the Shanghai Municipal Science and Technology Major Project under Grant 2021SHZDZX0102. + +# References + +[1] Sravanti Addepalli, Ashish Ramayee Asokan, Lakshay Sharma, and R Venkatesh Babu. Leveraging vision-language models for improving domain generalization in image classification. In CVPR, pages 23922-23932, 2024. 1, 2, 3, 6, 7 +[2] James Urquhart Allingham, Jie Ren, Michael W Dusenberry, Xiuye Gu, Yin Cui, Dustin Tran, Jeremiah Zhe Liu, and Balaji Lakshminarayanan. A simple zero-shot prompt weighting technique to improve prompt ensembling in text-image models. In ICML, 2023. 4 +[3] Martin Arjovsky, Léon Bottou, Ishaan Gulrajani, and David Lopez-Paz. Invariant risk minimization. arXiv preprint arXiv:1907.02893, 2019. 2 +[4] Shuanghao Bai, Yuedi Zhang, Wanqi Zhou, Zhirong Luan, and Badong Chen. Soft prompt generation for domain generalization. In ECCV, pages 434-450, 2024. 3, 6, 7 +[5] Sara Beery, Grant Van Horn, and Pietro Perona. Recognition in terra incognita. In ECCV, pages 456-473, 2018. 6, 7 +[6] Junbum Cha, Sanghyuk Chun, Kyungjae Lee, Han-Cheol Cho, Seunghyun Park, Yunsung Lee, and Sungrae Park. Swad: Domain generalization by seeking flat minima. NeurIPS, 34:22405-22418, 2021. 1, 2, 6, 7 +[7] Junbum Cha, Kyungjae Lee, Sungrae Park, and Sanghyuk Chun. Domain generalization by mutual-information regularization with pre-trained models. In ECCV, pages 440-457, 2022. 1, 6 +[8] Liang Chen, Yong Zhang, Yibing Song, Ying Shan, and Lingqiao Liu. Improved test-time adaptation for domain generalization. In CVPR, pages 24172-24182, 2023. 2 +[9] Liang Chen, Yong Zhang, Yibing Song, Anton Van Den Hengel, and Lingqiao Liu. Domain generalization via rationale invariance. In ICCV, pages 1751-1760, 2023. 1 +[10] De Cheng, Zhipeng Xu, Xinyang Jiang, Nannan Wang, Dongsheng Li, and Xinbo Gao. Disentangled prompt representation for domain generalization. In CVPR, pages 23595-23604, 2024. 1, 2 +[11] Junhyeong Cho, Gilhyun Nam, Sungyeon Kim, Hunmin Yang, and Suha Kwak. Promptstyler: Prompt-driven style generation for source-free domain generalization. In ICCV, pages 15702-15712, 2023. 2, 6 +[12] Huiqi Deng, Na Zou, Mengnan Du, Weifu Chen, Guocan Feng, Ziwei Yang, Zheyang Li, and Quanshi Zhang. Unifying fourteen post-hoc attribution methods with taylor interactions. IEEE Transactions on Pattern Analysis and Machine Intelligence, 46(7):4625-4640, 2024. 4 +[13] Yu Ding, Lei Wang, Bin Liang, Shuming Liang, Yang Wang, and Fang Chen. Domain generalization by learning and removing domain-specific features. In NeurIPS, pages 24226-24239, 2022. 2 +[14] Xiaoyi Dong, Jianmin Bao, Yinglin Zheng, Ting Zhang, Dongdong Chen, Hao Yang, Ming Zeng, Weiming Zhang, Lu Yuan, Dong Chen, et al. Maskclip: Masked self-distillation advances contrastive language-image pretraining. In CVPR, pages 10995-11005, 2023. 2 +[15] Alexey Dosovitskiy, Lucas Beyer, Alexander Kolesnikov, Dirk Weissenborn, Xiaohua Zhai, Thomas Unterthiner, + +Mostafa Dehghani, Matthias Minderer, Georg Heigold, Sylvain Gelly, Jakob Uszkoreit, and Neil Houlsby. An image is worth 16x16 words: Transformers for image recognition at scale. In ICLR, 2021. 3 +[16] Lisa Dunlap, Clara Mohri, Devin Guillory, Han Zhang, Trevor Darrell, Joseph E Gonzalez, Aditi Raghunathan, and Anna Rohrbach. Using language to extend to unseen domains. In ICLR, 2023. 2 +[17] Chen Fang, Ye Xu, and Daniel N Rockmore. Unbiased metric learning: On the utilization of multiple datasets and web images for softening bias. In ICCV, pages 1657-1664, 2013. 6, 7 +[18] Yaroslav Ganin, Evgeniya Ustinova, Hana Ajakan, Pascal Germain, Hugo Larochelle, François Laviolette, Mario March, and Victor Lempitsky. Domain-adversarial training of neural networks. Journal of machine learning research, 17(59):1-35, 2016. 2 +[19] Chunjiang Ge, Rui Huang, Mixue Xie, Zihang Lai, Shiji Song, Shuang Li, and Gao Huang. Domain adaptation via prompt learning. IEEE Transactions on Neural Networks and Learning Systems, pages 1-11, 2023. 3 +[20] Sachin Goyal, Ananya Kumar, Sankalp Garg, Zico Kolter, and Aditi Raghunathan. Finetune like you pretrain: Improved finetuning of zero-shot vision models. In CVPR, pages 19338-19347, 2023. 1, 2, 6, 7 +[21] Ishaan Gulrajani and David Lopez-Paz. In search of lost domain generalization. In ICLR, 2020. 2, 6 +[22] Jintao Guo, Lei Qi, and Yinghuan Shi. Domaindrop: Suppressing domain-sensitive channels for domain generalization. In ICCV, pages 19114-19124, 2023. 2, 6 +[23] Jiajun Hu, Jian Zhang, Lei Qi, Yinghuan Shi, and Yang Gao. Learn to preserve and diversify: Parameter-efficient group with orthogonal regularization for domain generalization. In ECCV, pages 198-216, 2024. 1, 2, 3, 6, 7 +[24] Lanqing Hu, Meina Kan, Shiguang Shan, and Xilin Chen. Dandelionnet: Domain composition with instance adaptive classification for domain generalization. In ICCV, pages 19050-19059, 2023. 1 +[25] Zenan Huang, Haobo Wang, Junbo Zhao, and Nenggan Zheng. idag: Invariant dag searching for domain generalization. In ICCV, pages 19169-19179, 2023. 1 +[26] Zeyi Huang, Andy Zhou, Zijian Ling, Mu Cai, Haohan Wang, and Yong Jae Lee. A sentence speaks a thousand images: Domain generalization through distilling clip with language guidance. In ICCV, pages 11685-11695, 2023. 1, 2, 3, 6, 7 +[27] Samyak Jain, Sravanti Addepalli, Pawan Kumar Sahu, Priyam Dey, and R Venkatesh Babu. Dart: Diversify-aggregate-repeat training improves generalization of neural networks. In CVPR, pages 16048-16059, 2023. 1, 6, 7 +[28] Chao Jia, Yinfei Yang, Ye Xia, Yi-Ting Chen, Zarana Parekh, Hieu Pham, Quoc Le, Yun-Hsuan Sung, Zhen Li, and Tom Duerig. Scaling up visual and vision-language representation learning with noisy text supervision. In ICML, pages 4904-4916, 2021. 2 +[29] Muhammad Uzair Khattak, Hanoona Rasheed, Muhammad Maaz, Salman Khan, and Fahad Shahbaz Khan. Maple: + +Multi-modal prompt learning. In CVPR, pages 19113-19122, 2023. 2, 3, 7 +[30] Ananya Kumar, Aditi Raghunathan, Robbie Jones, Tengyu Ma, and Percy Liang. Fine-tuning can distort pretrained features and underperform out-of-distribution. In ICLR, 2022. 2, 6, 7 +[31] Chenming Li, Daoan Zhang, Wenjian Huang, and Jianguo Zhang. Cross contrasting feature perturbation for domain generalization. In ICCV, pages 1327-1337, 2023. 2 +[32] Da Li, Yongxin Yang, Yi-Zhe Song, and Timothy M Hospedales. Deeper, broader and artier domain generalization. In ICCV, pages 5542-5550, 2017. 6, 7 +[33] Haoliang Li, Sinno Jialin Pan, Shiqi Wang, and Alex C Kot. Domain generalization with adversarial feature learning. In CVPR, pages 5400-5409, 2018. 2 +[34] Junnan Li, Dongxu Li, Caiming Xiong, and Steven Hoi. Blip: Bootstrapping language-image pre-training for unified vision-language understanding and generation. In ICML, pages 12888-12900, 2022. 1, 2 +[35] Junnan Li, Dongxu Li, Silvio Savarese, and Steven Hoi. Blip-2: Bootstrapping language-image pre-training with frozen image encoders and large language models. In ICML, pages 19730–19742, 2023. 1, 2 +[36] Xiaotong Li, Yongxing Dai, Yixiao Ge, Jun Liu, Ying Shan, and Ling-Yu Duan. Uncertainty modeling for out-of-distribution generalization. In ICLR, 2022. 2 +[37] Ya Li, Mingming Gong, Xinmei Tian, Tongliang Liu, and Dacheng Tao. Domain generalization via conditional invariant representations. In AAAI, 2018. 2 +[38] Ya Li, Xinmei Tian, Mingming Gong, Yajing Liu, Tongliang Liu, Kun Zhang, and Dacheng Tao. Deep domain generalization via conditional invariant adversarial networks. In ECCV, pages 624-639, 2018. 2 +[39] Krikamol Muandet, David Balduzzi, and Bernhard Schölkopf. Domain generalization via invariant feature representation. In ICML, pages 10-18, 2013. 2 +[40] Hyeonseob Nam, HyunJae Lee, Jongchan Park, Wonjun Yoon, and Donggeun Yoo. Reducing domain gap by reducing style bias. In CVPR, pages 8690-8699, 2021. 2 +[41] Hongjing Niu, Hanting Li, Feng Zhao, and Bin Li. Domain-unified prompt representations for source-free domain generalization. arXiv preprint arXiv:2209.14926, 2022. 2 +[42] Oren Nuriel, Sagie Benaim, and Lior Wolf. Permuted adain: Reducing the bias towards global statistics in image classification. In CVPR, pages 9482-9491, 2021. 2 +[43] Changdae Oh, Junhyuk So, Hoyoon Byun, YongTaek Lim, Minchul Shin, Jong-June Jeon, and Kyungwoo Song. Geodesic multi-modal mixup for robust fine-tuning. In NeurIPS, 2024. 1 +[44] Xingchao Peng, Qinxun Bai, Xide Xia, Zijun Huang, Kate Saenko, and Bo Wang. Moment matching for multi-source domain adaptation. In ICCV, pages 1406-1415, 2019. 6 +[45] Zelin Peng, Zhengqin Xu, Zhilin Zeng, Lingxi Xie, Qi Tian, and Wei Shen. Parameter efficient fine-tuning via cross block orchestration for segment anything model. In CVPR, pages 3743-3752, 2024. 2 + +[46] Shi Pu, Kaili Zhao, and Mao Zheng. Alignment-uniformity aware representation learning for zero-shot video classification. In CVPR, pages 19968-19977, 2022. 1, 5 +[47] Daiqing Qi, Handong Zhao, Aidong Zhang, and Sheng Li. Generalizing to unseen domains via text-guided augmentation. In ECCV, pages 285-300, 2024. 2 +[48] Minghan Qin, Wanhua Li, Jiawei Zhou, Haoqian Wang, and Hanspeter Pfister. Langsplat: 3d language gaussian splatting. In CVPR, pages 20051-20060, 2024. 2 +[49] Alec Radford, Jong Wook Kim, Chris Hallacy, Aditya Ramesh, Gabriel Goh, Sandhini Agarwal, Girish Sastry, Amanda Askell, Pamela Mishkin, Jack Clark, et al. Learning transferable visual models from natural language supervision. In ICML, pages 8748-8763, 2021. 1, 2, 3, 6, 7 +[50] Yang Shu, Xingzhuo Guo, Jialong Wu, Ximei Wang, Jianmin Wang, and Mingsheng Long. Clipood: Generalizing clip to out-of-distributions. In ICML, pages 31716-31731, 2023. 1, 2, 6, 7 +[51] Mainak Singha, Ankit Jha, Shirsha Bose, Ashwin Nair, Moloud Abdar, and Biplab Banerjee. Unknown prompt the only lacuna: Unveiling clip's potential for open domain generalization. In CVPR, pages 13309-13319, 2024. 1 +[52] Quan Sun, Yuxin Fang, Ledell Wu, Xinlong Wang, and Yue Cao. Eva-clip: Improved training techniques for clip at scale. arXiv preprint arXiv:2303.15389, 2023. 1, 2 +[53] Vladimir N Vapnik. An overview of statistical learning theory. IEEE transactions on neural networks, 10(5):988-999, 1999. 2 +[54] Ashish Vaswani, Noam Shazeer, Niki Parmar, Jakob Uszkoreit, Llion Jones, Aidan N Gomez, Lukasz Kaiser, and Illia Polosukhin. Attention is all you need. In NeurIPS, pages 5998-6008, 2017. 3 +[55] Hemanth Venkateswara, Jose Eusebio, Shayok Chakraborty, and Sethuraman Panchanathan. Deep hashing network for unsupervised domain adaptation. In CVPR, pages 5018-5027, 2017. 6, 7 +[56] Mengzhu Wang, Yuehua Liu, Jianlong Yuan, Shanshan Wang, Zhibin Wang, and Wei Wang. Inter-class and inter-domain semantic augmentation for domain generalization. IEEE Transactions on Image Processing, 33:1338-1347, 2024. 2 +[57] Pengfei Wang, Zhaoxiang Zhang, Zhen Lei, and Lei Zhang. Sharpness-aware gradient matching for domain generalization. In CVPR, pages 3769-3778, 2023. 6 +[58] Tongzhou Wang and Phillip Isola. Understanding contrastive representation learning through alignment and uniformity on the hypersphere. In ICML, pages 9929-9939, 2020. 1 +[59] Yufei Wang, Haoliang Li, and Alex C Kot. Heterogeneous domain generalization via domain mixup. In ICASSP, pages 3622-3626, 2020. 2 +[60] Mitchell Wortsman, Gabriel Ilharco, Jong Wook Kim, Mike Li, Simon Kornblith, Rebecca Roelofs, Raphael Gontijo Lopes, Hannaneh Hajishirzi, Ali Farhadi, Hongseok Namkoong, et al. Robust fine-tuning of zero-shot models. In CVPR, pages 7959-7971, 2022. 1, 2, 6, 7 +[61] Sitong Wu, Haoru Tan, Zhuotao Tian, Yukang Chen, Xiaojuan Qi, and Jiaya Jia. Saco loss: Sample-wise affinity con + +sistency for vision-language pre-training. In CVPR, pages 27358-27369, 2024. 1, 2 +[62] Xin Zhang, Shixiang Shane Gu, Yutaka Matsuo, and Yusuke Iwasawa. Domain prompt learning for efficiently adapting clip to unseen domains. Transactions of the Japanese Society for Artificial Intelligence, 38(6), 2023. 3, 6, 7 +[63] Xingxuan Zhang, Renzhe Xu, Han Yu, Yancheng Dong, Pengfei Tian, and Peng Cui. Flatness-aware minimization for domain generalization. In ICCV, pages 5189-5202, 2023. 2 +[64] Kaiyang Zhou, Jingkang Yang, Chen Change Loy, and Zwei Liu. Conditional prompt learning for vision-language models. In CVPR, pages 16816-16825, 2022. 2, 3, 7 +[65] Kaiyang Zhou, Jingkang Yang, Chen Change Loy, and Ziwei Liu. Learning to prompt for vision-language models. International Journal of Computer Vision, 130(9):2337-2348, 2022. 2, 3, 7 \ No newline at end of file diff --git a/2025/Domain Generalization in CLIP via Learning with Diverse Text Prompts/images.zip b/2025/Domain Generalization in CLIP via Learning with Diverse Text Prompts/images.zip new file mode 100644 index 0000000000000000000000000000000000000000..cde8319915330643eb22f88acdf8ad2908bfe15a --- /dev/null +++ b/2025/Domain Generalization in CLIP via Learning with Diverse Text Prompts/images.zip @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:a8b01682eb2d61450c40caf61e90bdd7b4e77808a23f217e5a1b86a4f8c0c524 +size 462587 diff --git a/2025/Domain Generalization in CLIP via Learning with Diverse Text Prompts/layout.json b/2025/Domain Generalization in CLIP via Learning with Diverse Text Prompts/layout.json new file mode 100644 index 0000000000000000000000000000000000000000..4cc5aab19f55c7b3c4defd03aefd7f497318f99a --- /dev/null +++ b/2025/Domain Generalization in CLIP via Learning with Diverse Text Prompts/layout.json @@ -0,0 +1,9420 @@ +{ + "pdf_info": [ + { + "para_blocks": [ + { + "bbox": [ + 82, + 103, + 530, + 121 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 82, + 103, + 530, + 121 + ], + "spans": [ + { + "bbox": [ + 82, + 103, + 530, + 121 + ], + "type": "text", + "content": "Domain Generalization in CLIP via Learning with Diverse Text Prompts" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 102, + 143, + 508, + 172 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 102, + 143, + 508, + 172 + ], + "spans": [ + { + "bbox": [ + 102, + 143, + 508, + 172 + ], + "type": "text", + "content": "Changsong Wen, Zelin Peng, Yu Huang, Xiaokang Yang, Wei Shen MoE Key Lab of Artificial Intelligence, AI Institute, Shanghai Jiao Tong University" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 128, + 174, + 480, + 186 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 128, + 174, + 480, + 186 + ], + "spans": [ + { + "bbox": [ + 128, + 174, + 480, + 186 + ], + "type": "text", + "content": "{changsong, zelin.peng, yellowfish, xkyang, wei.shen}@sjtu.edu.cn" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 151, + 213, + 201, + 225 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 151, + 213, + 201, + 225 + ], + "spans": [ + { + "bbox": [ + 151, + 213, + 201, + 225 + ], + "type": "text", + "content": "Abstract" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 55, + 239, + 297, + 525 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 239, + 297, + 525 + ], + "spans": [ + { + "bbox": [ + 55, + 239, + 297, + 525 + ], + "type": "text", + "content": "Domain generalization (DG) aims to train a model on source domains that can generalize well to unseen domains. Recent advances in Vision-Language Models (VLMs), such as CLIP, exhibit remarkable generalization capabilities across a wide range of data distributions, benefiting tasks like DG. However, CLIP is pre-trained by aligning images with their descriptions, which inevitably captures domain-specific details. Moreover, adapting CLIP to source domains with limited feature diversity introduces bias. These limitations hinder the model's ability to generalize across domains. In this paper, we propose a new DG approach by learning with diverse text prompts. These text prompts incorporate varied contexts to imitate different domains, enabling DG model to learn domain-invariant features. The text prompts guide DG model learning in three aspects: feature suppression, which uses these prompts to identify domain-sensitive features and suppress them; feature consistency, which ensures the model's features are robust to domain variations imitated by the diverse prompts; and feature diversification, which diversifies features based on the prompts to mitigate bias. Experimental results show that our approach improves domain generalization performance on five datasets from the DomainBed benchmark, achieving state-of-the-art results." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 55, + 552, + 135, + 564 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 552, + 135, + 564 + ], + "spans": [ + { + "bbox": [ + 55, + 552, + 135, + 564 + ], + "type": "text", + "content": "1. Introduction" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 55, + 572, + 296, + 693 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 572, + 296, + 693 + ], + "spans": [ + { + "bbox": [ + 55, + 572, + 296, + 693 + ], + "type": "text", + "content": "In the complex real-world applications, deep learning models are expected to be capable of recognizing data from a wide range of distributions. Within the past few years, researchers are dedicated to investigating domain generalization (DG) [6, 7, 9, 24, 25, 27], which empowers the model to learn robust features from multiple domains and exhibit good performance across various data distributions. Recent advances in foundation Vision-Language Models (VLMs) [34, 35, 52, 61], such as CLIP [49], are trained on large-scale data sourced from the web with a wide variety" + } + ] + } + ], + "index": 8 + }, + { + "type": "image", + "bbox": [ + 328, + 233, + 555, + 369 + ], + "blocks": [ + { + "bbox": [ + 328, + 233, + 555, + 369 + ], + "lines": [ + { + "bbox": [ + 328, + 233, + 555, + 369 + ], + "spans": [ + { + "bbox": [ + 328, + 233, + 555, + 369 + ], + "type": "image", + "image_path": "2745900baa7e7c20eff0effcd1eedfb1ebef11f5cf1d26f966d7308491e855b3.jpg" + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 313, + 398, + 555, + 475 + ], + "lines": [ + { + "bbox": [ + 313, + 398, + 555, + 475 + ], + "spans": [ + { + "bbox": [ + 313, + 398, + 555, + 475 + ], + "type": "text", + "content": "Figure 1. Limitations with existing methods. As shown on the left, CLIP's image encoder is pre-trained to align with textual descriptions, which inevitably captures domain-specific visual details, resulting in perturbations in domain generalization. Additionally, as shown on the right, due to the limited feature diversity in the source domains, adapting CLIP with source domains will introduce bias." + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "image_caption" + } + ], + "index": 9 + }, + { + "bbox": [ + 313, + 497, + 555, + 544 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 497, + 555, + 544 + ], + "spans": [ + { + "bbox": [ + 313, + 497, + 555, + 544 + ], + "type": "text", + "content": "of data distributions. These VLMs possess remarkable generalization capabilities, and boost many open-world recognition tasks, including out-of-distribution learning [50] and domain generalization [1, 10, 51]." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 313, + 546, + 556, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 546, + 556, + 713 + ], + "spans": [ + { + "bbox": [ + 313, + 546, + 556, + 713 + ], + "type": "text", + "content": "Existing DG studies [20, 23, 27, 50, 60] leverage CLIP's pre-trained weights with techniques like parameter regularization [23] and weight space ensembling [6, 27, 60] to adapt the model for downstream datasets, while preserving generalization ability. Other recent DG works [1, 26] propose using knowledge distillation to transfer CLIP's knowledge into DG models. However, as shown in Fig. 1, CLIP [49] aligns images with their descriptions during pretraining, which inevitably contain domain-specific details, impacting existing DG models. Additionally, since the training images are limited to a few source domains in DG, adapting CLIP within these images leads the model to emphasize the features specific to the source domains. This results in a lack of feature diversity [43, 46, 58], which com" + } + ] + } + ], + "index": 12 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "spans": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "type": "text", + "content": "CVF" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 145, + 0, + 496, + 37 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 145, + 0, + 496, + 37 + ], + "spans": [ + { + "bbox": [ + 145, + 0, + 496, + 37 + ], + "type": "text", + "content": "This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 70, + 703, + 153, + 713 + ], + "type": "page_footnote", + "angle": 0, + "lines": [ + { + "bbox": [ + 70, + 703, + 153, + 713 + ], + "spans": [ + { + "bbox": [ + 70, + 703, + 153, + 713 + ], + "type": "text", + "content": "Corresponding author." + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "text", + "content": "9559" + } + ] + } + ], + "index": 14 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 0 + }, + { + "para_blocks": [ + { + "bbox": [ + 55, + 72, + 293, + 85 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 72, + 293, + 85 + ], + "spans": [ + { + "bbox": [ + 55, + 72, + 293, + 85 + ], + "type": "text", + "content": "promises the model's ability to generalize to new domains." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 54, + 85, + 295, + 216 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 54, + 85, + 295, + 216 + ], + "spans": [ + { + "bbox": [ + 54, + 85, + 295, + 216 + ], + "type": "text", + "content": "Erasing domain-sensitive details and mitigating bias benefit from learning domain-invariant representations for each class. In this paper, we propose a new DG approach to learn domain-invariant class features by leveraging a diverse set of text prompts. Based on the multi-modal alignment capabilities of CLIP, we use these text prompts to imitate multiple domains and effectively guide the DG learning process. A wide range of domains provides diverse perspectives on each class, allowing the model to learn features that are robust across varying contexts, thereby promoting domain-invariant feature learning." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 56, + 218, + 296, + 445 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 218, + 296, + 445 + ], + "spans": [ + { + "bbox": [ + 56, + 218, + 296, + 445 + ], + "type": "text", + "content": "Diverse text prompts support three key aspects of our approach to learning domain-invariant features: they facilitate domain-sensitive feature suppression, ensure feature consistency across domain variations, and enable feature diversification to reduce bias. To be specific, we leverage the prompts recommended by original CLIP [49] to approximate different domains, such as \"a painting of a {class}\" or \"a {class} in a video game\". We first identify channels in the features activated by specific text prompts. These channels are considered highly relevant to the domain context of the prompt, thus we suppress their activation values. Next, we compare predictions under perturbations of diverse text prompts, and employ a consistency constraint to ensure the model is robust to domain variations while maintaining stable predictions. Subsequently, we decompose the component of the image features orthogonal to the text prompts and impose diversification on the orthogonal component. This approach enables a more diverse distribution of image features without impacting classification." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 55, + 445, + 295, + 517 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 445, + 295, + 517 + ], + "spans": [ + { + "bbox": [ + 55, + 445, + 295, + 517 + ], + "type": "text", + "content": "To comprehensively evaluate our method, we perform extensive experiments on the widely used DomainBed benchmark [21]. The experiment results demonstrate the superior performance and robustness of our proposed approach, achieving state-of-the-art results across five domain generalization datasets." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 55, + 530, + 142, + 542 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 530, + 142, + 542 + ], + "spans": [ + { + "bbox": [ + 55, + 530, + 142, + 542 + ], + "type": "text", + "content": "2. Related Work" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 55, + 551, + 250, + 563 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 551, + 250, + 563 + ], + "spans": [ + { + "bbox": [ + 55, + 551, + 250, + 563 + ], + "type": "text", + "content": "2.1. Large-scale Vision Language Models" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 55, + 570, + 295, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 570, + 295, + 714 + ], + "spans": [ + { + "bbox": [ + 55, + 570, + 295, + 714 + ], + "type": "text", + "content": "Large-scale Vision Language Models (VLMs) [28, 49, 52, 61], trained on extensive image-text data from the internet, demonstrate strong robustness to diverse data distributions. These VLMs exhibit remarkable generalization capabilities, serving as fundamental models in the field of domain generalization. CLIP [49] is one of the most prevalent approaches among VLMs. It is trained by aligning image-text pairs through contrastive learning, where the model learns to associate corresponding images and texts from large-scale datasets by maximizing their similarity in a shared embedding space. Among the past few years, many efforts have been made to improve the CLIP-like VLMs by designing" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 313, + 72, + 553, + 156 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 72, + 553, + 156 + ], + "spans": [ + { + "bbox": [ + 313, + 72, + 553, + 156 + ], + "type": "text", + "content": "better loss function [61], mitigating the impact of noisy web data [34, 35], and developing optimization techniques [52]. Another line of works focus on leveraging CLIP's generalization capabilities and its cross-modal alignment strengths to address downstream tasks such as zero-shot classification [29, 64], domain generalization [1, 26], and 2D/3D dense prediction [14, 45, 48]." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 313, + 164, + 447, + 176 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 164, + 447, + 176 + ], + "spans": [ + { + "bbox": [ + 313, + 164, + 447, + 176 + ], + "type": "text", + "content": "2.2. Domain Generalization" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 313, + 182, + 553, + 277 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 182, + 553, + 277 + ], + "spans": [ + { + "bbox": [ + 313, + 182, + 553, + 277 + ], + "type": "text", + "content": "Domain generalization [3, 8, 10, 18, 31, 40, 53, 59, 63] seeks to develop models capable of adapting to tasks from unseen domains without additional training. Conventional domain generalization methods encompass approaches such as learning domain-invariant features [13, 22, 37, 39], data augmentation [36, 42, 56], and adversarial learning [33, 38] to improve generalization to unseen domains." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 313, + 279, + 554, + 471 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 279, + 554, + 471 + ], + "spans": [ + { + "bbox": [ + 313, + 279, + 554, + 471 + ], + "type": "text", + "content": "Recently, VLMs are widely utilized as foundational models in domain generalization [1, 20, 26, 50, 60]. Some approaches fine-tune CLIP model on downstream DG datasets. However, since these DG datasets (0.01 to " + }, + { + "bbox": [ + 313, + 279, + 554, + 471 + ], + "type": "inline_equation", + "content": "0.5\\mathrm{M}" + }, + { + "bbox": [ + 313, + 279, + 554, + 471 + ], + "type": "text", + "content": ") are significantly smaller than CLIP's training set (400M), directly fine-tuning disrupts CLIP's parameter space [30], leading to degraded generalization performance. To mitigate the limitations, many impressive works [6, 30, 60] propose parameter ensemble methods between CLIP and the fine-tuned model to preserve generalization ability. Other works adopt CLIP for DG downstream tasks by designing loss functions during fine-tuning [20, 50], or employing regularization in parameter-efficient fine-tuning method [23]. Some knowledge distillation-based methods [1, 26] align the features of DG models with the output features of CLIP, transferring CLIP's generalization ability to DG models." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 313, + 472, + 554, + 627 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 472, + 554, + 627 + ], + "spans": [ + { + "bbox": [ + 313, + 472, + 554, + 627 + ], + "type": "text", + "content": "DomainDrop [22] mitigates model feature sensitivity to different domains by suppressing channel-wise feature variance across source domains. PromptStyler [11] generates diverse templates with style constraints and content consistency to train a text classifier. The text classifier is directly applied to visual domain generalization. LADS [16] and TEAM [47] employ text prompts for facilitating image feature augmentation. CAE [41] manually design a domain bank to obtain domain-unified text prompts for a stable prediction across domains. Different from these works that recognize the importance of text prompts, our approach enables direct optimization of the visual encoder by imitating multiple domain variations through a set of text prompts." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 313, + 635, + 433, + 647 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 635, + 433, + 647 + ], + "spans": [ + { + "bbox": [ + 313, + 635, + 433, + 647 + ], + "type": "text", + "content": "2.3. Prompt Engineering" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 313, + 654, + 553, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 654, + 553, + 715 + ], + "spans": [ + { + "bbox": [ + 313, + 654, + 553, + 715 + ], + "type": "text", + "content": "In CLIP, the design of prompts promotes the model to effectively capture the key information within the image. Recent prompt learning methods [29, 64, 65] are proposed to optimize the prompt as trainable parameters, allowing the model to dynamically adjust the prompt based on tasks." + } + ] + } + ], + "index": 13 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 315, + 757 + ], + "type": "text", + "content": "9560" + } + ] + } + ], + "index": 14 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 1 + }, + { + "para_blocks": [ + { + "bbox": [ + 54, + 72, + 294, + 156 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 54, + 72, + 294, + 156 + ], + "spans": [ + { + "bbox": [ + 54, + 72, + 294, + 156 + ], + "type": "text", + "content": "CoOp [65] enhances CLIP's adaptability for visual recognition tasks by using learnable context vectors as language prompts. CoCoOp [64] incorporates image features as conditional prompts to better capture context. MaPLe [29] introduces a multimodal prompt learning approach, finetuning both image and text encoders simultaneously for improved performance." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 55, + 157, + 295, + 314 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 157, + 295, + 314 + ], + "spans": [ + { + "bbox": [ + 55, + 157, + 295, + 314 + ], + "type": "text", + "content": "Recently, the domain generalization method DPL [62] employs an MLP to capture domain-specific image features as context for the learnable prompt. SPG [4] designs a prompt generator with domain knowledge to generate soft prompts for the images from target domain. These methods [4, 19, 62] primarily focus on incorporating image context into the text encoder to achieve multimodal alignment in downstream tasks. Differently, we optimize the image encoder through prompts from diverse templates. These prompts effectively imitate different domains while preventing the parameters of learnable prompts from being influenced by source domain-specific information during training." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 55, + 326, + 133, + 340 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 326, + 133, + 340 + ], + "spans": [ + { + "bbox": [ + 55, + 326, + 133, + 340 + ], + "type": "text", + "content": "3. Preliminary" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 55, + 347, + 295, + 453 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 347, + 295, + 453 + ], + "spans": [ + { + "bbox": [ + 55, + 347, + 295, + 453 + ], + "type": "text", + "content": "Problem Formulation. Domain generalization aims to train a model that can generalize well to unseen target domains, even when the training data comes from multiple source domains with potentially different data distributions. Specifically, the model is trained using data from " + }, + { + "bbox": [ + 55, + 347, + 295, + 453 + ], + "type": "inline_equation", + "content": "o" + }, + { + "bbox": [ + 55, + 347, + 295, + 453 + ], + "type": "text", + "content": " different source domains " + }, + { + "bbox": [ + 55, + 347, + 295, + 453 + ], + "type": "inline_equation", + "content": "\\{D_i^s\\}_{i=1}^o" + }, + { + "bbox": [ + 55, + 347, + 295, + 453 + ], + "type": "text", + "content": ", where each source domain " + }, + { + "bbox": [ + 55, + 347, + 295, + 453 + ], + "type": "inline_equation", + "content": "D_i^s = \\{(x_i, y_i)\\}_{i=1}^u" + }, + { + "bbox": [ + 55, + 347, + 295, + 453 + ], + "type": "text", + "content": " consists of input data " + }, + { + "bbox": [ + 55, + 347, + 295, + 453 + ], + "type": "inline_equation", + "content": "x_i" + }, + { + "bbox": [ + 55, + 347, + 295, + 453 + ], + "type": "text", + "content": " and corresponding class labels " + }, + { + "bbox": [ + 55, + 347, + 295, + 453 + ], + "type": "inline_equation", + "content": "y_i" + }, + { + "bbox": [ + 55, + 347, + 295, + 453 + ], + "type": "text", + "content": ", with " + }, + { + "bbox": [ + 55, + 347, + 295, + 453 + ], + "type": "inline_equation", + "content": "u" + }, + { + "bbox": [ + 55, + 347, + 295, + 453 + ], + "type": "text", + "content": " representing the number of samples in the source domain." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 55, + 456, + 295, + 515 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 456, + 295, + 515 + ], + "spans": [ + { + "bbox": [ + 55, + 456, + 295, + 515 + ], + "type": "text", + "content": "The model's generalization performance is evaluated on the target domain " + }, + { + "bbox": [ + 55, + 456, + 295, + 515 + ], + "type": "inline_equation", + "content": "D^{t}" + }, + { + "bbox": [ + 55, + 456, + 295, + 515 + ], + "type": "text", + "content": ", which shares the same set of class labels as the source domains " + }, + { + "bbox": [ + 55, + 456, + 295, + 515 + ], + "type": "inline_equation", + "content": "\\{D_{i}^{s}\\}_{i = 1}^{o}" + }, + { + "bbox": [ + 55, + 456, + 295, + 515 + ], + "type": "text", + "content": ". The joint distribution of input data " + }, + { + "bbox": [ + 55, + 456, + 295, + 515 + ], + "type": "inline_equation", + "content": "x" + }, + { + "bbox": [ + 55, + 456, + 295, + 515 + ], + "type": "text", + "content": " and labels " + }, + { + "bbox": [ + 55, + 456, + 295, + 515 + ], + "type": "inline_equation", + "content": "y" + }, + { + "bbox": [ + 55, + 456, + 295, + 515 + ], + "type": "text", + "content": " in the target domain differs from that in the source domains." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 55, + 517, + 295, + 612 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 517, + 295, + 612 + ], + "spans": [ + { + "bbox": [ + 55, + 517, + 295, + 612 + ], + "type": "text", + "content": "Utilizing CLIP for domain generalization. CLIP's strong generalization capability stems from its multimodal training on diverse internet data covering a wide range of concepts and contexts. As a prominent vision-language model, CLIP comprises an image encoder " + }, + { + "bbox": [ + 55, + 517, + 295, + 612 + ], + "type": "inline_equation", + "content": "F_{I}(\\cdot)" + }, + { + "bbox": [ + 55, + 517, + 295, + 612 + ], + "type": "text", + "content": " and a text encoder " + }, + { + "bbox": [ + 55, + 517, + 295, + 612 + ], + "type": "inline_equation", + "content": "F_{T}(\\cdot)" + }, + { + "bbox": [ + 55, + 517, + 295, + 612 + ], + "type": "text", + "content": ". Both image and text encoders adopt a transformer architecture [15, 54] to generate their representations through self-attention." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 55, + 614, + 295, + 673 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 614, + 295, + 673 + ], + "spans": [ + { + "bbox": [ + 55, + 614, + 295, + 673 + ], + "type": "text", + "content": "CLIP employs contrastive learning to align images and their descriptive text during pre-training. It associates visual and textual features by maximizing the similarity of the aligned image-text pairs while reducing the similarity of other ones:" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 78, + 683, + 295, + 717 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 78, + 683, + 295, + 717 + ], + "spans": [ + { + "bbox": [ + 78, + 683, + 295, + 717 + ], + "type": "interline_equation", + "content": "\\mathcal {L} _ {\\mathrm {C L I P}} = - \\frac {1}{N} \\sum_ {i = 1} ^ {N} \\log \\frac {\\exp \\left(\\mathrm {S} \\left(I _ {i} , A _ {i}\\right) / \\tau\\right)}{\\sum_ {j = 1} ^ {N} \\exp \\left(\\mathrm {S} \\left(I _ {i} , A _ {j}\\right) / \\tau\\right)}, \\tag {1}", + "image_path": "416040b4a78f5a575a2169e8f17879aab09e50abde2a5f875384d27b405e080d.jpg" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 313, + 72, + 553, + 120 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 72, + 553, + 120 + ], + "spans": [ + { + "bbox": [ + 313, + 72, + 553, + 120 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 313, + 72, + 553, + 120 + ], + "type": "inline_equation", + "content": "I" + }, + { + "bbox": [ + 313, + 72, + 553, + 120 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 313, + 72, + 553, + 120 + ], + "type": "inline_equation", + "content": "A" + }, + { + "bbox": [ + 313, + 72, + 553, + 120 + ], + "type": "text", + "content": " represent the input image and caption, " + }, + { + "bbox": [ + 313, + 72, + 553, + 120 + ], + "type": "inline_equation", + "content": "i" + }, + { + "bbox": [ + 313, + 72, + 553, + 120 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 313, + 72, + 553, + 120 + ], + "type": "inline_equation", + "content": "j" + }, + { + "bbox": [ + 313, + 72, + 553, + 120 + ], + "type": "text", + "content": " are indexes of the inputs, " + }, + { + "bbox": [ + 313, + 72, + 553, + 120 + ], + "type": "inline_equation", + "content": "S" + }, + { + "bbox": [ + 313, + 72, + 553, + 120 + ], + "type": "text", + "content": " means similarity matrices, such as cosine similarity. " + }, + { + "bbox": [ + 313, + 72, + 553, + 120 + ], + "type": "inline_equation", + "content": "\\tau" + }, + { + "bbox": [ + 313, + 72, + 553, + 120 + ], + "type": "text", + "content": " is a temperature parameter used to adjust the smoothness of the similarity distribution." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 313, + 121, + 555, + 251 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 121, + 555, + 251 + ], + "spans": [ + { + "bbox": [ + 313, + 121, + 555, + 251 + ], + "type": "text", + "content": "Baseline of our approach. In the conventional domain generalization methods [1, 23, 29], \"a photo of a " + }, + { + "bbox": [ + 313, + 121, + 555, + 251 + ], + "type": "inline_equation", + "content": "\\{class\\}" + }, + { + "bbox": [ + 313, + 121, + 555, + 251 + ], + "type": "text", + "content": "\" is encoded by CLIP's text encoder as text prompt " + }, + { + "bbox": [ + 313, + 121, + 555, + 251 + ], + "type": "inline_equation", + "content": "\\{P_i\\}_{i=1}^c" + }, + { + "bbox": [ + 313, + 121, + 555, + 251 + ], + "type": "text", + "content": ", representing the main concepts of " + }, + { + "bbox": [ + 313, + 121, + 555, + 251 + ], + "type": "inline_equation", + "content": "c" + }, + { + "bbox": [ + 313, + 121, + 555, + 251 + ], + "type": "text", + "content": " different classes. Leveraging CLIP's remarkable multi-modal alignment capabilities between images and text, the text prompts for different classes are treated as domain-invariant representations of the image features [1, 26]. We adopt a basic baseline [1] to adapt CLIP for downstream domain generalization tasks by simultaneously distilling image and text features to the DG model." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 314, + 262, + 556, + 286 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 314, + 262, + 556, + 286 + ], + "spans": [ + { + "bbox": [ + 314, + 262, + 556, + 286 + ], + "type": "interline_equation", + "content": "\\mathcal {L} _ {\\mathrm {K D}} = \\lambda \\cdot \\cos \\left(F _ {I} \\left(x _ {i}\\right), f _ {I} \\left(x _ {i}\\right)\\right) + (1 - \\lambda) \\cdot \\cos \\left(F _ {T} \\left(y _ {i}\\right), f _ {I} \\left(x _ {i}\\right)\\right), \\tag {2}", + "image_path": "db88306d74b1dd9bcaeccef7a063361314a16d7f2edc1671b7c23b02a4bed68e.jpg" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 313, + 286, + 555, + 395 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 286, + 555, + 395 + ], + "spans": [ + { + "bbox": [ + 313, + 286, + 555, + 395 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 313, + 286, + 555, + 395 + ], + "type": "inline_equation", + "content": "F_{I}(\\cdot)" + }, + { + "bbox": [ + 313, + 286, + 555, + 395 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 313, + 286, + 555, + 395 + ], + "type": "inline_equation", + "content": "F_{T}(\\cdot)" + }, + { + "bbox": [ + 313, + 286, + 555, + 395 + ], + "type": "text", + "content": " are CLIP's image and text encoders, " + }, + { + "bbox": [ + 313, + 286, + 555, + 395 + ], + "type": "inline_equation", + "content": "f_{I}(\\cdot)" + }, + { + "bbox": [ + 313, + 286, + 555, + 395 + ], + "type": "text", + "content": " is the DG model. Our approach differs slightly in that we ensemble multiple prompts used to imitate different domains during the distillation process. This has a minor direct impact on performance but is more intuitive to the design of our overall framework, ensuring stability across domain variations. The classification result " + }, + { + "bbox": [ + 313, + 286, + 555, + 395 + ], + "type": "inline_equation", + "content": "\\hat{y}" + }, + { + "bbox": [ + 313, + 286, + 555, + 395 + ], + "type": "text", + "content": " is obtained by choosing the ensembled text prompt with the highest similarity to the " + }, + { + "bbox": [ + 313, + 286, + 555, + 395 + ], + "type": "inline_equation", + "content": "i" + }, + { + "bbox": [ + 313, + 286, + 555, + 395 + ], + "type": "text", + "content": "-th image." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 313, + 406, + 397, + 420 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 406, + 397, + 420 + ], + "spans": [ + { + "bbox": [ + 313, + 406, + 397, + 420 + ], + "type": "text", + "content": "4. Methodology" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 313, + 426, + 555, + 547 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 426, + 555, + 547 + ], + "spans": [ + { + "bbox": [ + 313, + 426, + 555, + 547 + ], + "type": "text", + "content": "In this section, we introduce our proposed learning with diverse prompts paradigm for learning domain-invariant features in DG. Initially, we filter out unreliable prompts based on the prediction confidence. Next, we propose a Domain-sensitive Channel Suppression (DCS) method to identify and suppress domain-sensitive features. Moreover, a Multi-domain Consistency Constraint (MCC) approach is designed to make the features robust to variations across domains. Then, we apply Orthogonal Uniformity Regularization (OUR) to diversify features in the DG model." + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 313, + 555, + 488, + 568 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 555, + 488, + 568 + ], + "spans": [ + { + "bbox": [ + 313, + 555, + 488, + 568 + ], + "type": "text", + "content": "4.1. Entropy-based Prompt Filtering" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 313, + 574, + 554, + 634 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 574, + 554, + 634 + ], + "spans": [ + { + "bbox": [ + 313, + 574, + 554, + 634 + ], + "type": "text", + "content": "We imitate different domains by using the text templates " + }, + { + "bbox": [ + 313, + 574, + 554, + 634 + ], + "type": "inline_equation", + "content": "\\{T_j\\}_{j=1}^n" + }, + { + "bbox": [ + 313, + 574, + 554, + 634 + ], + "type": "text", + "content": " recommended in CLIP [49], such as \"a drawing of a {class}\", " + }, + { + "bbox": [ + 313, + 574, + 554, + 634 + ], + "type": "inline_equation", + "content": "n" + }, + { + "bbox": [ + 313, + 574, + 554, + 634 + ], + "type": "text", + "content": " is the number of templates. These prompts are generated by encoding the combination of the " + }, + { + "bbox": [ + 313, + 574, + 554, + 634 + ], + "type": "inline_equation", + "content": "j" + }, + { + "bbox": [ + 313, + 574, + 554, + 634 + ], + "type": "text", + "content": "-th text template " + }, + { + "bbox": [ + 313, + 574, + 554, + 634 + ], + "type": "inline_equation", + "content": "T_j" + }, + { + "bbox": [ + 313, + 574, + 554, + 634 + ], + "type": "text", + "content": " and the " + }, + { + "bbox": [ + 313, + 574, + 554, + 634 + ], + "type": "inline_equation", + "content": "i" + }, + { + "bbox": [ + 313, + 574, + 554, + 634 + ], + "type": "text", + "content": "-th class label " + }, + { + "bbox": [ + 313, + 574, + 554, + 634 + ], + "type": "inline_equation", + "content": "y_i" + }, + { + "bbox": [ + 313, + 574, + 554, + 634 + ], + "type": "text", + "content": " by CLIP text encoder:" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 397, + 644, + 553, + 659 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 397, + 644, + 553, + 659 + ], + "spans": [ + { + "bbox": [ + 397, + 644, + 553, + 659 + ], + "type": "interline_equation", + "content": "P _ {i} ^ {j} = F _ {T} \\left(y _ {i}, T _ {j}\\right), \\tag {3}", + "image_path": "d57be4211d2d96951c133873371e639785cfa67c416706069a1b9c6da2837b68.jpg" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 313, + 665, + 555, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 665, + 555, + 714 + ], + "spans": [ + { + "bbox": [ + 313, + 665, + 555, + 714 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 313, + 665, + 555, + 714 + ], + "type": "inline_equation", + "content": "P \\in \\mathbb{R}^{n \\times c \\times D}" + }, + { + "bbox": [ + 313, + 665, + 555, + 714 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 313, + 665, + 555, + 714 + ], + "type": "inline_equation", + "content": "c" + }, + { + "bbox": [ + 313, + 665, + 555, + 714 + ], + "type": "text", + "content": " is the number of classes, " + }, + { + "bbox": [ + 313, + 665, + 555, + 714 + ], + "type": "inline_equation", + "content": "D" + }, + { + "bbox": [ + 313, + 665, + 555, + 714 + ], + "type": "text", + "content": " represents the feature dimensions, " + }, + { + "bbox": [ + 313, + 665, + 555, + 714 + ], + "type": "inline_equation", + "content": "i" + }, + { + "bbox": [ + 313, + 665, + 555, + 714 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 313, + 665, + 555, + 714 + ], + "type": "inline_equation", + "content": "j" + }, + { + "bbox": [ + 313, + 665, + 555, + 714 + ], + "type": "text", + "content": " indicate the index of the classes and prompt templates. These prompts employ diverse contexts to describe each class, such as a drawing," + } + ] + } + ], + "index": 17 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 314, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 314, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 314, + 757 + ], + "type": "text", + "content": "9561" + } + ] + } + ], + "index": 18 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 2 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 91, + 116, + 517, + 348 + ], + "blocks": [ + { + "bbox": [ + 91, + 116, + 517, + 348 + ], + "lines": [ + { + "bbox": [ + 91, + 116, + 517, + 348 + ], + "spans": [ + { + "bbox": [ + 91, + 116, + 517, + 348 + ], + "type": "image", + "image_path": "d7d94cbb7475280d264b29faded19619031e0a3df2471fb30d2a5b37785fc980.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 54, + 372, + 555, + 418 + ], + "lines": [ + { + "bbox": [ + 54, + 372, + 555, + 418 + ], + "spans": [ + { + "bbox": [ + 54, + 372, + 555, + 418 + ], + "type": "text", + "content": "Figure 2. Overview of our proposed method. To enforce the DG model to learn domain-invariant class features, we imitate multiple domains by text prompts. Three components of our method, including feature suppression, feature consistency, and feature diversification, are achieved through Domain-sensitive Channel Suppression (DCS), Multi-domain Consistency Constraint (MCC), and Orthogonal Uniformity Regularization (OUR) modules. The blue line represents the process during inference." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "bbox": [ + 54, + 437, + 294, + 496 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 54, + 437, + 294, + 496 + ], + "spans": [ + { + "bbox": [ + 54, + 437, + 294, + 496 + ], + "type": "text", + "content": "a sculpture, or a rendering, etc. Thus, these prompts can be naturally regarded as descriptions of the class from multiple domains. The DG model " + }, + { + "bbox": [ + 54, + 437, + 294, + 496 + ], + "type": "inline_equation", + "content": "f_{I}(\\cdot)" + }, + { + "bbox": [ + 54, + 437, + 294, + 496 + ], + "type": "text", + "content": " is regularized to extract domain-invariant features that are robust to prompts from different domains." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 55, + 498, + 295, + 604 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 498, + 295, + 604 + ], + "spans": [ + { + "bbox": [ + 55, + 498, + 295, + 604 + ], + "type": "text", + "content": "Within the prompts provided by CLIP, we first discard a few unreliable ones based on the prediction confidence for each input [2], as they negatively affect the feature channel suppression and feature consistency calculation, which rely on accurately identifying relevant features. In the Entropy-based Prompt Filtering (EPF), we first extract the encoded image features " + }, + { + "bbox": [ + 55, + 498, + 295, + 604 + ], + "type": "inline_equation", + "content": "I_{i} = f_{I}(x_{i})" + }, + { + "bbox": [ + 55, + 498, + 295, + 604 + ], + "type": "text", + "content": " for the input image " + }, + { + "bbox": [ + 55, + 498, + 295, + 604 + ], + "type": "inline_equation", + "content": "x_{i}" + }, + { + "bbox": [ + 55, + 498, + 295, + 604 + ], + "type": "text", + "content": ". For the " + }, + { + "bbox": [ + 55, + 498, + 295, + 604 + ], + "type": "inline_equation", + "content": "j" + }, + { + "bbox": [ + 55, + 498, + 295, + 604 + ], + "type": "text", + "content": "-th template, the predicted probability for the " + }, + { + "bbox": [ + 55, + 498, + 295, + 604 + ], + "type": "inline_equation", + "content": "i" + }, + { + "bbox": [ + 55, + 498, + 295, + 604 + ], + "type": "text", + "content": "-th class is calculated as:" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 111, + 613, + 295, + 643 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 111, + 613, + 295, + 643 + ], + "spans": [ + { + "bbox": [ + 111, + 613, + 295, + 643 + ], + "type": "interline_equation", + "content": "\\hat {d} _ {i} ^ {j} = \\frac {\\exp \\left(\\cos \\left(I _ {i} , P _ {i} ^ {j}\\right)\\right)}{\\sum_ {k = 1} ^ {c} \\exp \\left(\\cos \\left(I _ {i} , P _ {k} ^ {j}\\right)\\right)}, \\tag {4}", + "image_path": "297715b38f87e4d6b67b901694c83d19feab5b700a365e307b4390bb4d33b627.jpg" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 55, + 651, + 295, + 677 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 651, + 295, + 677 + ], + "spans": [ + { + "bbox": [ + 55, + 651, + 295, + 677 + ], + "type": "text", + "content": "We denote that " + }, + { + "bbox": [ + 55, + 651, + 295, + 677 + ], + "type": "inline_equation", + "content": "\\hat{d}_i^j" + }, + { + "bbox": [ + 55, + 651, + 295, + 677 + ], + "type": "text", + "content": " is a constant and " + }, + { + "bbox": [ + 55, + 651, + 295, + 677 + ], + "type": "inline_equation", + "content": "\\hat{d}^j = \\{\\hat{d}_i^j\\}_{i=1}^c" + }, + { + "bbox": [ + 55, + 651, + 295, + 677 + ], + "type": "text", + "content": " is a predicted distribution for the " + }, + { + "bbox": [ + 55, + 651, + 295, + 677 + ], + "type": "inline_equation", + "content": "j" + }, + { + "bbox": [ + 55, + 651, + 295, + 677 + ], + "type": "text", + "content": "-th template." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 55, + 677, + 296, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 677, + 296, + 714 + ], + "spans": [ + { + "bbox": [ + 55, + 677, + 296, + 714 + ], + "type": "text", + "content": "Entropy quantifies the uncertainty of a probability distribution. If the model output has high confidence, the distribution typically concentrates on a specific class, resulting" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 313, + 437, + 554, + 486 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 437, + 554, + 486 + ], + "spans": [ + { + "bbox": [ + 313, + 437, + 554, + 486 + ], + "type": "text", + "content": "in low entropy. Conversely, if the model has similar confidence across multiple classes, the distribution is more uniform, leading to higher entropy. Therefore, we calculate the entropy for each template's prediction:" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 378, + 491, + 553, + 521 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 378, + 491, + 553, + 521 + ], + "spans": [ + { + "bbox": [ + 378, + 491, + 553, + 521 + ], + "type": "interline_equation", + "content": "\\mathrm {H} \\left(\\hat {d} ^ {j}\\right) = - \\sum_ {i = 1} ^ {c} \\hat {d} _ {i} ^ {j} \\log \\left(\\hat {d} _ {i} ^ {j}\\right). \\tag {5}", + "image_path": "456c432100329094913ae81f6bf38c45e2d2805d234ca6df874876270ecc0219.jpg" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 313, + 527, + 554, + 552 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 527, + 554, + 552 + ], + "spans": [ + { + "bbox": [ + 313, + 527, + 554, + 552 + ], + "type": "text", + "content": "Then, we eliminate the unreliable prompts and select the " + }, + { + "bbox": [ + 313, + 527, + 554, + 552 + ], + "type": "inline_equation", + "content": "m" + }, + { + "bbox": [ + 313, + 527, + 554, + 552 + ], + "type": "text", + "content": " prompts with the lowest entropy for each input:" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 343, + 556, + 553, + 572 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 343, + 556, + 553, + 572 + ], + "spans": [ + { + "bbox": [ + 343, + 556, + 553, + 572 + ], + "type": "interline_equation", + "content": "\\left\\{T _ {1}, T _ {2}, \\dots , T _ {m} \\right\\} = \\arg \\min _ {j \\in \\{1 \\dots , n \\}} \\mathrm {H} \\left(\\hat {d} ^ {j}\\right), \\tag {6}", + "image_path": "073689882041f6e1374a9cf10c59254e788bc61a58c64322ef16f9e6bf64f9a9.jpg" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 313, + 575, + 519, + 588 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 575, + 519, + 588 + ], + "spans": [ + { + "bbox": [ + 313, + 575, + 519, + 588 + ], + "type": "text", + "content": "4.2. Domain-sensitive Channel Suppression" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 313, + 594, + 554, + 689 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 594, + 554, + 689 + ], + "spans": [ + { + "bbox": [ + 313, + 594, + 554, + 689 + ], + "type": "text", + "content": "The image features should focus on discriminative features between classes and be robust to prompts from different domains. However, some domain-sensitive channels encode variations specific to the domain context of prompt, introducing perturbations. To address this, we propose Domain-sensitive Channel Suppression (DCS) to select and suppress these channels, ensuring the extracted image embedding captures domain-invariant class features." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 313, + 689, + 554, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 689, + 554, + 713 + ], + "spans": [ + { + "bbox": [ + 313, + 689, + 554, + 713 + ], + "type": "text", + "content": "Specifically, we utilize gradient attribution [12] to quantify each channel's contribution in the image features when" + } + ] + } + ], + "index": 13 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 315, + 757 + ], + "type": "text", + "content": "9562" + } + ] + } + ], + "index": 14 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 3 + }, + { + "para_blocks": [ + { + "bbox": [ + 55, + 72, + 296, + 133 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 72, + 296, + 133 + ], + "spans": [ + { + "bbox": [ + 55, + 72, + 296, + 133 + ], + "type": "text", + "content": "calculating their similarity with different prompts. We simultaneously consider the model's sensitivity to each channel (i.e., the gradient) and the actual activation value of that channel to measure each image channel's contribution to recognition with " + }, + { + "bbox": [ + 55, + 72, + 296, + 133 + ], + "type": "inline_equation", + "content": "P_{i}^{j}" + }, + { + "bbox": [ + 55, + 72, + 296, + 133 + ], + "type": "text", + "content": ":" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 119, + 143, + 295, + 171 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 119, + 143, + 295, + 171 + ], + "spans": [ + { + "bbox": [ + 119, + 143, + 295, + 171 + ], + "type": "interline_equation", + "content": "\\alpha_ {k j} = \\left| \\frac {\\partial \\cos \\left(I _ {i} , P _ {i} ^ {j}\\right)}{\\partial I _ {i} ^ {k}} \\right| \\cdot I _ {i} ^ {k}, \\tag {7}", + "image_path": "1faff0d5444e8650fdb24de31055e39b2987adc29438675f66bb273943b117e5.jpg" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 55, + 182, + 296, + 253 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 182, + 296, + 253 + ], + "spans": [ + { + "bbox": [ + 55, + 182, + 296, + 253 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 55, + 182, + 296, + 253 + ], + "type": "inline_equation", + "content": "P_{i}^{j}" + }, + { + "bbox": [ + 55, + 182, + 296, + 253 + ], + "type": "text", + "content": " means the prompt corresponding to the " + }, + { + "bbox": [ + 55, + 182, + 296, + 253 + ], + "type": "inline_equation", + "content": "i" + }, + { + "bbox": [ + 55, + 182, + 296, + 253 + ], + "type": "text", + "content": "-th image's label with " + }, + { + "bbox": [ + 55, + 182, + 296, + 253 + ], + "type": "inline_equation", + "content": "T_{j}" + }, + { + "bbox": [ + 55, + 182, + 296, + 253 + ], + "type": "text", + "content": " as the template. " + }, + { + "bbox": [ + 55, + 182, + 296, + 253 + ], + "type": "inline_equation", + "content": "\\alpha_{kj}" + }, + { + "bbox": [ + 55, + 182, + 296, + 253 + ], + "type": "text", + "content": " is a constant that represents the contribution of the " + }, + { + "bbox": [ + 55, + 182, + 296, + 253 + ], + "type": "inline_equation", + "content": "k" + }, + { + "bbox": [ + 55, + 182, + 296, + 253 + ], + "type": "text", + "content": "-th image channel to the similarity calculation with the " + }, + { + "bbox": [ + 55, + 182, + 296, + 253 + ], + "type": "inline_equation", + "content": "j" + }, + { + "bbox": [ + 55, + 182, + 296, + 253 + ], + "type": "text", + "content": "-th prompt " + }, + { + "bbox": [ + 55, + 182, + 296, + 253 + ], + "type": "inline_equation", + "content": "P_{i}^{j}" + }, + { + "bbox": [ + 55, + 182, + 296, + 253 + ], + "type": "text", + "content": ". For each image feature, we normalize it through the channel dimensions:" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 126, + 253, + 295, + 283 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 126, + 253, + 295, + 283 + ], + "spans": [ + { + "bbox": [ + 126, + 253, + 295, + 283 + ], + "type": "interline_equation", + "content": "\\bar {\\alpha} _ {k j} = \\frac {\\alpha_ {k j}}{\\frac {1}{D} \\sqrt {\\sum_ {k = 1} ^ {D} \\alpha_ {k j} ^ {2}}}, \\tag {8}", + "image_path": "1c7fd519464dae707f3c206a8e994b9e9ccdba2e52a5a4fedebce5cdf5d415a0.jpg" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 55, + 289, + 285, + 302 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 289, + 285, + 302 + ], + "spans": [ + { + "bbox": [ + 55, + 289, + 285, + 302 + ], + "type": "text", + "content": "Then, we calculate the z-score among different prompts:" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 134, + 311, + 295, + 334 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 134, + 311, + 295, + 334 + ], + "spans": [ + { + "bbox": [ + 134, + 311, + 295, + 334 + ], + "type": "interline_equation", + "content": "\\alpha_ {k j} ^ {\\mathrm {z - s c o r e}} = \\frac {\\overline {{\\alpha}} _ {k j} - \\mu}{\\sigma}, \\tag {9}", + "image_path": "790d27d8e01fc1c9e69fe175ac5f4edc66dd77df9f60b49b696daa7959ae0821.jpg" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 55, + 342, + 296, + 365 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 342, + 296, + 365 + ], + "spans": [ + { + "bbox": [ + 55, + 342, + 296, + 365 + ], + "type": "inline_equation", + "content": "\\mu" + }, + { + "bbox": [ + 55, + 342, + 296, + 365 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 55, + 342, + 296, + 365 + ], + "type": "inline_equation", + "content": "\\sigma" + }, + { + "bbox": [ + 55, + 342, + 296, + 365 + ], + "type": "text", + "content": " are the average of " + }, + { + "bbox": [ + 55, + 342, + 296, + 365 + ], + "type": "inline_equation", + "content": "\\overline{\\alpha}_{kj}" + }, + { + "bbox": [ + 55, + 342, + 296, + 365 + ], + "type": "text", + "content": " across all templates, which are calculated as:" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 141, + 374, + 295, + 407 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 141, + 374, + 295, + 407 + ], + "spans": [ + { + "bbox": [ + 141, + 374, + 295, + 407 + ], + "type": "interline_equation", + "content": "\\mu = \\frac {1}{m} \\sum_ {j = 1} ^ {m} \\bar {\\alpha} _ {k j} \\tag {10}", + "image_path": "583d1fcb24d44a77a032f7ac070acc7e71a33bfd0ac98205def5fddd23cf1305.jpg" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 121, + 421, + 295, + 459 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 121, + 421, + 295, + 459 + ], + "spans": [ + { + "bbox": [ + 121, + 421, + 295, + 459 + ], + "type": "interline_equation", + "content": "\\sigma = \\sqrt {\\frac {1}{m} \\sum_ {j = 1} ^ {m} \\left(\\bar {\\alpha} _ {k j} - \\mu\\right) ^ {2}} \\tag {11}", + "image_path": "1841eaa1383f422830b2fb53d100ff901292438a97a823ad3d35d7e4fbfe2a6f.jpg" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 55, + 464, + 296, + 536 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 464, + 296, + 536 + ], + "spans": [ + { + "bbox": [ + 55, + 464, + 296, + 536 + ], + "type": "text", + "content": "We check the z-score values of each channel across different prompts. A high z-score indicates that a channel is activated by the context provided by the specific prompt, showing a strong correlation with domain-specific information. We apply a mask to such image feature channels before calculating the loss:" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 104, + 546, + 295, + 578 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 104, + 546, + 295, + 578 + ], + "spans": [ + { + "bbox": [ + 104, + 546, + 295, + 578 + ], + "type": "interline_equation", + "content": "\\mathbf {M} _ {D} = \\left\\{ \\begin{array}{l l} 1, & \\text {i f} \\max _ {k} \\alpha_ {k j} ^ {\\mathrm {z - s c o r e}} > \\beta \\\\ 0, & \\text {o t h e r w i s e} \\end{array} \\right. \\tag {12}", + "image_path": "44a2b1edd108439ae894c59cc15e60c0e3150a1e3244dd1c19448e89f07fc448.jpg" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 55, + 589, + 296, + 637 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 589, + 296, + 637 + ], + "spans": [ + { + "bbox": [ + 55, + 589, + 296, + 637 + ], + "type": "text", + "content": "The mask " + }, + { + "bbox": [ + 55, + 589, + 296, + 637 + ], + "type": "inline_equation", + "content": "\\mathbf{M}_D\\in \\mathbb{R}^D" + }, + { + "bbox": [ + 55, + 589, + 296, + 637 + ], + "type": "text", + "content": " is directly employed on image features by the element-wise multiplication " + }, + { + "bbox": [ + 55, + 589, + 296, + 637 + ], + "type": "inline_equation", + "content": "I_i^k = \\mathbf{M}_D\\odot I_i" + }, + { + "bbox": [ + 55, + 589, + 296, + 637 + ], + "type": "text", + "content": ". We suppress these domain-sensitive features by applying " + }, + { + "bbox": [ + 55, + 589, + 296, + 637 + ], + "type": "inline_equation", + "content": "L_{1}" + }, + { + "bbox": [ + 55, + 589, + 296, + 637 + ], + "type": "text", + "content": " regularization:" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 129, + 647, + 295, + 680 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 129, + 647, + 295, + 680 + ], + "spans": [ + { + "bbox": [ + 129, + 647, + 295, + 680 + ], + "type": "interline_equation", + "content": "\\mathcal {L} _ {r e g} = \\gamma \\sum_ {i = 1} ^ {B} \\sum_ {k = 1} ^ {D} \\left| I _ {i} ^ {k} \\right|, \\tag {13}", + "image_path": "742b8f5563cf67f119298883eba585db6838cc3af2d09c467ff9590e870530c7.jpg" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 55, + 689, + 296, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 689, + 296, + 714 + ], + "spans": [ + { + "bbox": [ + 55, + 689, + 296, + 714 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 55, + 689, + 296, + 714 + ], + "type": "inline_equation", + "content": "B" + }, + { + "bbox": [ + 55, + 689, + 296, + 714 + ], + "type": "text", + "content": " represents the batch size, and " + }, + { + "bbox": [ + 55, + 689, + 296, + 714 + ], + "type": "inline_equation", + "content": "\\gamma" + }, + { + "bbox": [ + 55, + 689, + 296, + 714 + ], + "type": "text", + "content": " is the weight of loss, " + }, + { + "bbox": [ + 55, + 689, + 296, + 714 + ], + "type": "inline_equation", + "content": "I_{i}^{k}" + }, + { + "bbox": [ + 55, + 689, + 296, + 714 + ], + "type": "text", + "content": " is the masked image feature." + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 314, + 72, + 515, + 84 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 314, + 72, + 515, + 84 + ], + "spans": [ + { + "bbox": [ + 314, + 72, + 515, + 84 + ], + "type": "text", + "content": "4.3. Multi-domain Consistency Constraint" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 313, + 90, + 555, + 211 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 90, + 555, + 211 + ], + "spans": [ + { + "bbox": [ + 313, + 90, + 555, + 211 + ], + "type": "text", + "content": "We further introduce Multi-domain Consistency Constraint (MCC) on the predicted distribution between the image and varied prompts. We calculate the similarity between image features and different prompts for classification. The KL loss function is then applied to ensure that the predicted distributions remain consistent between different prompts, enhancing the robustness of the image encoder to domain variations. Specifically, we use Eq. (4) to obtain the predicted distributions for different templates " + }, + { + "bbox": [ + 313, + 90, + 555, + 211 + ], + "type": "inline_equation", + "content": "\\{\\hat{d}^j\\}_{j=1}^m" + }, + { + "bbox": [ + 313, + 90, + 555, + 211 + ], + "type": "text", + "content": ", and apply KL loss to ensure consistency:" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 375, + 219, + 555, + 251 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 375, + 219, + 555, + 251 + ], + "spans": [ + { + "bbox": [ + 375, + 219, + 555, + 251 + ], + "type": "interline_equation", + "content": "\\operatorname {K L} \\left(\\hat {d} ^ {j} \\| \\hat {d} ^ {k}\\right) = \\sum_ {i = 1} ^ {c} \\hat {d} _ {i} ^ {j} \\log \\frac {\\hat {d} _ {i} ^ {j}}{\\hat {d} _ {i} ^ {k}} \\tag {14}", + "image_path": "65adc37bca92ff084cac578c3fb9345102e3e2919dfb34202a3ba0e4693493a6.jpg" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 355, + 262, + 554, + 301 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 355, + 262, + 554, + 301 + ], + "spans": [ + { + "bbox": [ + 355, + 262, + 554, + 301 + ], + "type": "interline_equation", + "content": "\\mathcal {L} _ {\\mathrm {K L}} = \\frac {1}{m (m - 1)} \\sum_ {j = 1} ^ {m} \\sum_ {\\substack {k = 1 \\\\ k \\neq t}} ^ {m} \\mathrm {K L} \\left(\\hat {d} ^ {j} \\| \\hat {d} ^ {k}\\right) \\tag{15}", + "image_path": "cd0dc6e81065ab63ee85e7f064157db58a9347774e37ed00a1430bf4b0f83919.jpg" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 313, + 306, + 554, + 342 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 306, + 554, + 342 + ], + "spans": [ + { + "bbox": [ + 313, + 306, + 554, + 342 + ], + "type": "text", + "content": "The KL loss compares the prediction with prompts from diverse domains and enforces consistency, guiding the learned features to focus on domain-invariant information." + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 314, + 350, + 519, + 364 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 314, + 350, + 519, + 364 + ], + "spans": [ + { + "bbox": [ + 314, + 350, + 519, + 364 + ], + "type": "text", + "content": "4.4. Orthogonal Uniformity Regularization" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 313, + 369, + 555, + 476 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 369, + 555, + 476 + ], + "spans": [ + { + "bbox": [ + 313, + 369, + 555, + 476 + ], + "type": "text", + "content": "Due to the limited diversity of image source domains, learning with these images biases CLIP toward the feature distribution of these source domains. To mitigate this bias, we further diversify the normalized image features to achieve a more uniform distribution in the hyperspherical space. Uniformity of the features increases the possibility that unseen features are positioned near seen features [46], thereby enhancing the model's generalization through feature diversification." + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 313, + 477, + 554, + 514 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 477, + 554, + 514 + ], + "spans": [ + { + "bbox": [ + 313, + 477, + 554, + 514 + ], + "type": "text", + "content": "However, it is not feasible to directly diversify the image features. As mentioned in Sec. 3, we use an ensembled prompt " + }, + { + "bbox": [ + 313, + 477, + 554, + 514 + ], + "type": "inline_equation", + "content": "e_i" + }, + { + "bbox": [ + 313, + 477, + 554, + 514 + ], + "type": "text", + "content": " in classification, which can be formulated as:" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 400, + 523, + 554, + 555 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 400, + 523, + 554, + 555 + ], + "spans": [ + { + "bbox": [ + 400, + 523, + 554, + 555 + ], + "type": "interline_equation", + "content": "e _ {i} = \\frac {1}{n} \\sum_ {j = 1} ^ {n} P _ {i} ^ {j}. \\tag {16}", + "image_path": "1f0be9bf790d1c671d7c124d57a72920b69ddd013e5d9daf57af23eeb06081f1.jpg" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 313, + 564, + 554, + 589 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 564, + 554, + 589 + ], + "spans": [ + { + "bbox": [ + 313, + 564, + 554, + 589 + ], + "type": "text", + "content": "We compute the similarity between the ensembled prompt " + }, + { + "bbox": [ + 313, + 564, + 554, + 589 + ], + "type": "inline_equation", + "content": "e_i" + }, + { + "bbox": [ + 313, + 564, + 554, + 589 + ], + "type": "text", + "content": " and the image feature " + }, + { + "bbox": [ + 313, + 564, + 554, + 589 + ], + "type": "inline_equation", + "content": "I_i" + }, + { + "bbox": [ + 313, + 564, + 554, + 589 + ], + "type": "text", + "content": " for classification:" + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 382, + 599, + 555, + 613 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 382, + 599, + 555, + 613 + ], + "spans": [ + { + "bbox": [ + 382, + 599, + 555, + 613 + ], + "type": "interline_equation", + "content": "\\hat {y} = \\operatorname {a r g m a x} _ {j} \\cos \\left(I _ {i}, e _ {j}\\right). \\tag {17}", + "image_path": "e627440ed1c5090355174c826edde08b7baccb88456140762be4e00d14421d2e.jpg" + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 313, + 621, + 554, + 657 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 621, + 554, + 657 + ], + "spans": [ + { + "bbox": [ + 313, + 621, + 554, + 657 + ], + "type": "text", + "content": "Therefore, as demonstrated in Eq. (17), directly applying uniformity constraints on image features can impact the similarity calculation between image features and prompts." + } + ] + } + ], + "index": 25 + }, + { + "bbox": [ + 313, + 658, + 554, + 691 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 658, + 554, + 691 + ], + "spans": [ + { + "bbox": [ + 313, + 658, + 554, + 691 + ], + "type": "text", + "content": "To address this, we decompose " + }, + { + "bbox": [ + 313, + 658, + 554, + 691 + ], + "type": "inline_equation", + "content": "I_{i}" + }, + { + "bbox": [ + 313, + 658, + 554, + 691 + ], + "type": "text", + "content": " into two components, one in the direction of " + }, + { + "bbox": [ + 313, + 658, + 554, + 691 + ], + "type": "inline_equation", + "content": "e_{i}" + }, + { + "bbox": [ + 313, + 658, + 554, + 691 + ], + "type": "text", + "content": " and the other in the orthogonal direction:" + } + ] + } + ], + "index": 26 + }, + { + "bbox": [ + 399, + 691, + 553, + 717 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 399, + 691, + 553, + 717 + ], + "spans": [ + { + "bbox": [ + 399, + 691, + 553, + 717 + ], + "type": "interline_equation", + "content": "I _ {i} ^ {\\parallel} = \\frac {I _ {i} \\cdot e _ {i}}{\\| e _ {i} \\| ^ {2}} \\cdot e _ {i}, \\tag {18}", + "image_path": "f6ef1b2cf63d9a780dfe9b65bdda1675cafefd1591f59ff274a0e09ec22c5e95.jpg" + } + ] + } + ], + "index": 27 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 315, + 757 + ], + "type": "text", + "content": "9563" + } + ] + } + ], + "index": 28 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 4 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 69, + 112, + 542, + 340 + ], + "blocks": [ + { + "bbox": [ + 55, + 70, + 555, + 104 + ], + "lines": [ + { + "bbox": [ + 55, + 70, + 555, + 104 + ], + "spans": [ + { + "bbox": [ + 55, + 70, + 555, + 104 + ], + "type": "text", + "content": "Table 1. Comparison with the state-of-the-art methods on the five domain generalization datasets from DomainBed benchmark. We report the averaged leave-one-domain-out experimental results with ViT-B/16 as the backbone. Numbers in bold represent the best performance among all the methods. T and F indicate whether the CLIP parameters are trained or frozen during adaptation to DG." + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 69, + 112, + 542, + 340 + ], + "lines": [ + { + "bbox": [ + 69, + 112, + 542, + 340 + ], + "spans": [ + { + "bbox": [ + 69, + 112, + 542, + 340 + ], + "type": "table", + "html": "
MethodsCLIPOHTerraIncVLCSPACSDomainNetAvg
Fine-tuning
SWAD (NeurIPS'2021) [6]T81.0142.9279.1391.3557.9270.47
LP-FT (ICLR'2022) [30]T81.1747.2680.8892.9257.0471.85
WiSE-FT (CVPR'2022) [60]T86.3254.5082.8897.2958.0175.80
MIRO (ECCV'2022) [7]T84.8059.3082.3096.4460.4776.66
DART (CVPR'2023) [27]T80.9351.2480.3893.4359.3273.06
SAGM (CVPR'2023) [57]T83.4058.6482.0594.3159.0575.49
FLYP (CVPR'2023) [20]T82.7633.2566.6478.5357.4163.72
CLIPood (ICML'2023) [50]T83.3146.2877.1993.1657.7871.55
Knowledge Distillation
RISE (ICCV'2023) [26]T78.3949.6180.6293.2555.3771.45
VL2V-SD (CVPR'2024) [1]T87.3858.5483.2596.6862.7977.73
Prompt-learning
DPL (TJSAI'2023) [62]F84.2052.6084.3097.3059.5075.60
SPG (ECCV'2024) [4]F83.6050.2082.4097.0060.1074.70
Parameter-efficient Fine-tuning
LPD (ECCV'2024) [23]F84.2157.3083.2096.5559.3076.11
OursT87.6563.2784.7997.0363.1179.17
", + "image_path": "c25e209c638f443ef734a070fe2fa790b179a7d5f7e2fbb063515f8533920367.jpg" + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "table_body" + } + ], + "index": 1 + }, + { + "bbox": [ + 145, + 368, + 294, + 384 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 145, + 368, + 294, + 384 + ], + "spans": [ + { + "bbox": [ + 145, + 368, + 294, + 384 + ], + "type": "interline_equation", + "content": "I _ {i} ^ {\\perp} = I _ {i} - I _ {i} ^ {\\parallel}, \\tag {19}", + "image_path": "8ea30ed10691ebec5d933b1f32a80a2b827a4c1f69118d7e415ca300c9d50d36.jpg" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 55, + 388, + 296, + 439 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 388, + 296, + 439 + ], + "spans": [ + { + "bbox": [ + 55, + 388, + 296, + 439 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 55, + 388, + 296, + 439 + ], + "type": "inline_equation", + "content": "I_{i}^{\\parallel}" + }, + { + "bbox": [ + 55, + 388, + 296, + 439 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 55, + 388, + 296, + 439 + ], + "type": "inline_equation", + "content": "I_{i}^{\\perp}" + }, + { + "bbox": [ + 55, + 388, + 296, + 439 + ], + "type": "text", + "content": " are the decomposed parallel and orthogonal components, respectively. We define a pairwise uniformity loss based on the dot product between the orthogonal components of different image features within the batch:" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 99, + 444, + 295, + 486 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 99, + 444, + 295, + 486 + ], + "spans": [ + { + "bbox": [ + 99, + 444, + 295, + 486 + ], + "type": "interline_equation", + "content": "\\mathcal {L} _ {\\mathrm {O U}} = \\frac {1}{B (B - 1)} \\sum_ {i = 1} ^ {B} \\sum_ {\\substack {j = 1 \\\\ j \\neq i}} ^ {B} \\left(I _ {i} ^ {\\perp} \\cdot I _ {j} ^ {\\perp}\\right) ^ {2} \\tag{20}", + "image_path": "5c5a1133d1857bfd6e5c920715ba8743556ed77d304a28dbf04e84065f1c027d.jpg" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 55, + 491, + 296, + 552 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 491, + 296, + 552 + ], + "spans": [ + { + "bbox": [ + 55, + 491, + 296, + 552 + ], + "type": "text", + "content": "Since " + }, + { + "bbox": [ + 55, + 491, + 296, + 552 + ], + "type": "inline_equation", + "content": "e_i" + }, + { + "bbox": [ + 55, + 491, + 296, + 552 + ], + "type": "text", + "content": " is computed independently for each image, the orthogonal component is also based on different images. Applying the uniformity constraint solely on the orthogonal direction avoids interference with the similarity calculation between the image features and the prompt." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 55, + 552, + 296, + 600 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 552, + 296, + 600 + ], + "spans": [ + { + "bbox": [ + 55, + 552, + 296, + 600 + ], + "type": "text", + "content": "Finally, combining the baseline with Domain-sensitive Channel Suppression, Multi-domain Consistency Constraint, and Orthogonal Uniformity Regularization yields the total loss used during training, which is formulated as:" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 95, + 607, + 295, + 620 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 95, + 607, + 295, + 620 + ], + "spans": [ + { + "bbox": [ + 95, + 607, + 295, + 620 + ], + "type": "interline_equation", + "content": "\\mathcal {L} _ {\\text {t o t a l}} = \\mathcal {L} _ {K D} + \\mathcal {L} _ {\\text {r e g}} + \\mathcal {L} _ {K L} + \\mathcal {L} _ {O U}, \\tag {21}", + "image_path": "e42787bbad5817bbc2b4c5f9082a993b509893d08c4bc9729cd19f531b077045.jpg" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 55, + 628, + 137, + 641 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 628, + 137, + 641 + ], + "spans": [ + { + "bbox": [ + 55, + 628, + 137, + 641 + ], + "type": "text", + "content": "5. Experiments" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 55, + 647, + 192, + 659 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 647, + 192, + 659 + ], + "spans": [ + { + "bbox": [ + 55, + 647, + 192, + 659 + ], + "type": "text", + "content": "5.1. Datasets and Evaluation" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 55, + 665, + 296, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 665, + 296, + 714 + ], + "spans": [ + { + "bbox": [ + 55, + 665, + 296, + 714 + ], + "type": "text", + "content": "We conduct evaluation experiments on five datasets from the DomainBed benchmark [21], following previous methods [1, 22, 26]. PACS [32] consists of 9,991 samples spanning four domains: {art, cartoons, photos, sketches}, and" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 313, + 358, + 555, + 514 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 358, + 555, + 514 + ], + "spans": [ + { + "bbox": [ + 313, + 358, + 555, + 514 + ], + "type": "text", + "content": "includes a total of 7 classes. VLCS [17] contains 10,729 examples, 5 classes from four datasets as different domains: {VOC2007, LabelMe, Caltech, SUN}. Terra Incognita (TI) [5] contains 24,788 samples from 10 animal categories. These photos are captured across different locations {L100, L38, L43, L46}, acting as the four domains. Office Home (OH) [55] have four domains {art, clipart, product, real}, with 15,588 examples from 65 classes. DomainNet [44] is a large-scale cross-domain dataset containing 586,575 images from 6 different domains: {clipart, infographic, painting, quickdraw, real, sketch}. It covers 345 classes and is widely used to evaluate the performance of domain generalization algorithms." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 313, + 516, + 556, + 611 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 516, + 556, + 611 + ], + "spans": [ + { + "bbox": [ + 313, + 516, + 556, + 611 + ], + "type": "text", + "content": "Following the protocols from previous domain generalization works [1, 11, 26], we split the datasets by using " + }, + { + "bbox": [ + 313, + 516, + 556, + 611 + ], + "type": "inline_equation", + "content": "80\\%" + }, + { + "bbox": [ + 313, + 516, + 556, + 611 + ], + "type": "text", + "content": " of the data for training and the remaining " + }, + { + "bbox": [ + 313, + 516, + 556, + 611 + ], + "type": "inline_equation", + "content": "20\\%" + }, + { + "bbox": [ + 313, + 516, + 556, + 611 + ], + "type": "text", + "content": " as the validation set. We adopt a leave-one-out experiment setup, where we alternately leave on domain for testing and train with the remaining domains. The average performance on the unseen domains is reported as the domain generalization results." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 313, + 622, + 447, + 635 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 622, + 447, + 635 + ], + "spans": [ + { + "bbox": [ + 313, + 622, + 447, + 635 + ], + "type": "text", + "content": "5.2. Implementation Details" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 313, + 641, + 555, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 641, + 555, + 714 + ], + "spans": [ + { + "bbox": [ + 313, + 641, + 555, + 714 + ], + "type": "text", + "content": "We employ CLIP with ViT-B/16 as the backbone model and fine-tune it on each of the datasets. We fine-tune the model for 5,000 iterations on the PACS, VLCS, Office Home, and Terra Incognito datasets, and for 15,000 iterations on the DomainNet dataset. We use the " + }, + { + "bbox": [ + 313, + 641, + 555, + 714 + ], + "type": "inline_equation", + "content": "n = 80" + }, + { + "bbox": [ + 313, + 641, + 555, + 714 + ], + "type": "text", + "content": " templates recommended in original CLIP [49] as prompts to imitate dif" + } + ] + } + ], + "index": 14 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 315, + 757 + ], + "type": "text", + "content": "9564" + } + ] + } + ], + "index": 15 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 5 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 56, + 145, + 297, + 258 + ], + "blocks": [ + { + "bbox": [ + 55, + 70, + 297, + 138 + ], + "lines": [ + { + "bbox": [ + 55, + 70, + 297, + 138 + ], + "spans": [ + { + "bbox": [ + 55, + 70, + 297, + 138 + ], + "type": "text", + "content": "Table 2. Ablation analysis on the effectiveness of our method. EPS: entropy-based prompt selection, DCS: domain-sensitive channel suppression, MCC: multi-domain consistency constraint. OUR: orthogonal uniformity regularization, URIF*: uniformity regularization for image features, where * indicates it is not our component, but used as an additional comparison to OUR." + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 56, + 145, + 297, + 258 + ], + "lines": [ + { + "bbox": [ + 56, + 145, + 297, + 258 + ], + "spans": [ + { + "bbox": [ + 56, + 145, + 297, + 258 + ], + "type": "table", + "html": "
BaseEPSDCSMCCOURURIF*Avg
✓81.46
✓✓✓82.28
✓✓✓82.37
✓✓✓✓82.79
✓✓82.30
✓✓81.55
✓✓✓✓✓83.19
", + "image_path": "b0583a62be85f8c9dc3b8d567f93ba6b7f701503acaafe1098f8171b80bdac21.jpg" + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "table_body" + } + ], + "index": 1 + }, + { + "bbox": [ + 54, + 264, + 297, + 397 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 54, + 264, + 297, + 397 + ], + "spans": [ + { + "bbox": [ + 54, + 264, + 297, + 397 + ], + "type": "text", + "content": "ferent domains for each class. The threshold " + }, + { + "bbox": [ + 54, + 264, + 297, + 397 + ], + "type": "inline_equation", + "content": "\\beta" + }, + { + "bbox": [ + 54, + 264, + 297, + 397 + ], + "type": "text", + "content": " for the suppressed feature mask " + }, + { + "bbox": [ + 54, + 264, + 297, + 397 + ], + "type": "inline_equation", + "content": "\\mathbf{M}_D" + }, + { + "bbox": [ + 54, + 264, + 297, + 397 + ], + "type": "text", + "content": " is set dynamically for each sample. We suppress " + }, + { + "bbox": [ + 54, + 264, + 297, + 397 + ], + "type": "inline_equation", + "content": "20\\%" + }, + { + "bbox": [ + 54, + 264, + 297, + 397 + ], + "type": "text", + "content": " image feature channels to reduce domain-specific perturbation, and set the regularization term as " + }, + { + "bbox": [ + 54, + 264, + 297, + 397 + ], + "type": "inline_equation", + "content": "10^{-4}" + }, + { + "bbox": [ + 54, + 264, + 297, + 397 + ], + "type": "text", + "content": ". We select " + }, + { + "bbox": [ + 54, + 264, + 297, + 397 + ], + "type": "inline_equation", + "content": "m = 40" + }, + { + "bbox": [ + 54, + 264, + 297, + 397 + ], + "type": "text", + "content": " prompts with low entropy from 80 contextual prompts. Other experimental settings all follow our baseline model. SWAD [6] is a general approach that improves model robustness by averaging weights during fine-tuning DG models. For a fair comparison, we integrate SWAD into all methods that fine-tune CLIP's parameters, following the previous work [1]." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 55, + 406, + 242, + 418 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 406, + 242, + 418 + ], + "spans": [ + { + "bbox": [ + 55, + 406, + 242, + 418 + ], + "type": "text", + "content": "5.3. Comparison with State-of-the-Arts" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 56, + 426, + 295, + 605 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 426, + 295, + 605 + ], + "spans": [ + { + "bbox": [ + 56, + 426, + 295, + 605 + ], + "type": "text", + "content": "We mainly compare with recent advances in domain generalization methods based on CLIP. We first compares our method with the following fine-tuning methods. SWAD [6], WiSE-FT [60], and DART [27] explore weight-ensembling methods to maintain CLIP's generalization ability. LPFT [30], FLYP [20], and CLIPood [50] investigate effective fine-tuning methods. The importance of text is highlighted in VL2V-SD [1] and RISE [26]. They incorporate text features into the DG model via knowledge distillation. DPL [62] and SPG [4] are prompt-learning methods that design learnable modules to extract context features from input images. LPD [23] explores a parameter-efficient fine-tuning approach for adapting CLIP in DG. We report the averaged leave-one-domain-out experimental results following the setting from previous works [1, 4]." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 55, + 606, + 297, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 606, + 297, + 715 + ], + "spans": [ + { + "bbox": [ + 55, + 606, + 297, + 715 + ], + "type": "text", + "content": "As shown in Table 1, our approach achieves an average improvement of " + }, + { + "bbox": [ + 55, + 606, + 297, + 715 + ], + "type": "inline_equation", + "content": "1.4\\%" + }, + { + "bbox": [ + 55, + 606, + 297, + 715 + ], + "type": "text", + "content": " over SOTA methods across five datasets. Our method demonstrates a particularly notable " + }, + { + "bbox": [ + 55, + 606, + 297, + 715 + ], + "type": "inline_equation", + "content": "4.7\\%" + }, + { + "bbox": [ + 55, + 606, + 297, + 715 + ], + "type": "text", + "content": " improvement over SOTA on the TerraInc dataset [5]. As the images in TerraInc are sourced from wildlife photography, target objects occupy a small portion of the scene and contain substantial irrelevant contextual information. Our approach effectively reduces interference from such perturbation and achieves impressive improvement." + } + ] + } + ], + "index": 5 + }, + { + "type": "table", + "bbox": [ + 315, + 136, + 558, + 239 + ], + "blocks": [ + { + "bbox": [ + 313, + 70, + 555, + 127 + ], + "lines": [ + { + "bbox": [ + 313, + 70, + 555, + 127 + ], + "spans": [ + { + "bbox": [ + 313, + 70, + 555, + 127 + ], + "type": "text", + "content": "Table 3. Performance comparison of our method with various prompt learning approaches across different datasets, including OH, TI, VLCS, and PACS. The average accuracy (Avg) across these datasets highlights the robustness and superior generalization capability of template generated prompts." + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 315, + 136, + 558, + 239 + ], + "lines": [ + { + "bbox": [ + 315, + 136, + 558, + 239 + ], + "spans": [ + { + "bbox": [ + 315, + 136, + 558, + 239 + ], + "type": "table", + "html": "
MethodsOHTIVLCSPACSAvg
CLIP [49]82.434.182.396.573.8
CoOP [65]83.046.880.896.476.8
Co-CoOP [64]83.445.380.396.776.4
Maple [29]83.450.282.296.578.1
DPL [62]84.252.684.397.379.6
SPG [4]83.650.282.497.078.3
Ours87.763.384.897.083.2
", + "image_path": "ec6e5f34a6ef29ab5aabf926ec7f493c8d4bfa40d8851cabfcc7225af15ad57a.jpg" + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "table_body" + } + ], + "index": 7 + }, + { + "bbox": [ + 313, + 250, + 408, + 262 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 250, + 408, + 262 + ], + "spans": [ + { + "bbox": [ + 313, + 250, + 408, + 262 + ], + "type": "text", + "content": "5.4. Ablation Study" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 313, + 269, + 555, + 425 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 269, + 555, + 425 + ], + "spans": [ + { + "bbox": [ + 313, + 269, + 555, + 425 + ], + "type": "text", + "content": "Ablation of the main components. We conduct ablation studies to evaluate the effectiveness of each component of our approach. Following previous methods [1, 23], we report the results on four datasets: OH [55], TI [5], VLCS [17], and PACS [32]. 'Avg' denotes the average of the four datasets. As shown in Table 2, using DCS with EPS yields a notable performance increase and improves the accuracy to " + }, + { + "bbox": [ + 313, + 269, + 555, + 425 + ], + "type": "inline_equation", + "content": "82.28\\%" + }, + { + "bbox": [ + 313, + 269, + 555, + 425 + ], + "type": "text", + "content": ", highlighting the significance of suppressing the domain-specific channels. MCC also achieves an improvement, raising accuracy to " + }, + { + "bbox": [ + 313, + 269, + 555, + 425 + ], + "type": "inline_equation", + "content": "82.37\\%" + }, + { + "bbox": [ + 313, + 269, + 555, + 425 + ], + "type": "text", + "content": " by enforcing cross-domain consistency. Applying both DSC and MCC modules simultaneously to suppress the domain-specific context results in an average gain of " + }, + { + "bbox": [ + 313, + 269, + 555, + 425 + ], + "type": "inline_equation", + "content": "1.33\\%" + }, + { + "bbox": [ + 313, + 269, + 555, + 425 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 312, + 426, + 556, + 569 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 312, + 426, + 556, + 569 + ], + "spans": [ + { + "bbox": [ + 312, + 426, + 556, + 569 + ], + "type": "text", + "content": "Moreover, incorporating uniformity regularization for the orthogonal component increases the average accuracy to " + }, + { + "bbox": [ + 312, + 426, + 556, + 569 + ], + "type": "inline_equation", + "content": "82.30\\%" + }, + { + "bbox": [ + 312, + 426, + 556, + 569 + ], + "type": "text", + "content": ", suggesting that capturing more diversified features enables the model to generalize better across domains by preventing an overemphasis on specific source domains. In contrast, applying uniformity regularization directly to image features (URIF*) without orthogonal decomposition yields a decrease to " + }, + { + "bbox": [ + 312, + 426, + 556, + 569 + ], + "type": "inline_equation", + "content": "81.55\\%" + }, + { + "bbox": [ + 312, + 426, + 556, + 569 + ], + "type": "text", + "content": ", demonstrating that directly regularizing image features may introduce perturbations during classification. Finally, the best performance is obtained by combining EPS, DSC, MCR, and OUR, reaching an average accuracy of " + }, + { + "bbox": [ + 312, + 426, + 556, + 569 + ], + "type": "inline_equation", + "content": "83.19\\%" + }, + { + "bbox": [ + 312, + 426, + 556, + 569 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 312, + 570, + 557, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 312, + 570, + 557, + 715 + ], + "spans": [ + { + "bbox": [ + 312, + 570, + 557, + 715 + ], + "type": "text", + "content": "Discussions on the design of prompts. In our implementation, we imitate diverse domains by the prompts generated from 80 templates recommended by CLIP. In this section, we conduct experiments to analysis the advantages of our template-based prompts over the learnable prompts used in previous methods [4, 29, 62, 64]. By optimizing the learnable prompts during training, these model learn to adapt dynamically to different inputs. As shown in Table 3, although these methods show significant improvement compared to CLIP, our approach still outperforms the state-of-the-art prompt learning methods by " + }, + { + "bbox": [ + 312, + 570, + 557, + 715 + ], + "type": "inline_equation", + "content": "3.6\\%" + }, + { + "bbox": [ + 312, + 570, + 557, + 715 + ], + "type": "text", + "content": ". While they incorporate context of the image, the learnable prompts also" + } + ] + } + ], + "index": 11 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 315, + 757 + ], + "type": "text", + "content": "9565" + } + ] + } + ], + "index": 12 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 6 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 56, + 124, + 296, + 213 + ], + "blocks": [ + { + "bbox": [ + 55, + 70, + 297, + 114 + ], + "lines": [ + { + "bbox": [ + 55, + 70, + 297, + 114 + ], + "spans": [ + { + "bbox": [ + 55, + 70, + 297, + 114 + ], + "type": "text", + "content": "Table 4. Ablations of comparison with knowledge distillation from the single manually designed prompt and the design of our method. EP represents the ensembled " + }, + { + "bbox": [ + 55, + 70, + 297, + 114 + ], + "type": "inline_equation", + "content": "n" + }, + { + "bbox": [ + 55, + 70, + 297, + 114 + ], + "type": "text", + "content": " prompts for knowledge distillation." + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 56, + 124, + 296, + 213 + ], + "lines": [ + { + "bbox": [ + 56, + 124, + 296, + 213 + ], + "spans": [ + { + "bbox": [ + 56, + 124, + 296, + 213 + ], + "type": "table", + "html": "
PromptsOHTIVLCSPACSAvg
{class}86.559.583.796.681.6
a photo of a {class}87.458.583.396.781.5
a {domain} of a {class}87.260.783.996.982.2
EP87.159.983.696.581.9
Ours87.763.384.897.083.2
", + "image_path": "078f735ab9058108c7d086458b84f5c22407515071a1ca644ef04ec18c97d4e1.jpg" + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "table_body" + } + ], + "index": 1 + }, + { + "bbox": [ + 55, + 232, + 295, + 291 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 232, + 295, + 291 + ], + "spans": [ + { + "bbox": [ + 55, + 232, + 295, + 291 + ], + "type": "text", + "content": "introduces potential perturbations from specific source domains. Unlike these learnable prompts, our method leverages a set of templates without requiring domain-specific adaptation during training, enabling the model to generalize more robustly across diverse domains." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 55, + 293, + 295, + 544 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 293, + 295, + 544 + ], + "spans": [ + { + "bbox": [ + 55, + 293, + 295, + 544 + ], + "type": "text", + "content": "Analysis on knowledge distillation from different designed prompts. In this part, we conduct experiments to demonstrate the advantages of using a diverse set of prompts compared to the carefully designed manual prompt. In our approach, the diverse set of prompts are ensembled and a knowledge distillation framework is used to transfer the ensembled prompts (EP) to the DG model. Based on the knowledge distillation baseline, we compare with three types of manually designed prompts, including {\"class)\", \"a photo of a {class}\", and \"a {domain}\" of a {class)\". As shown in Table 4, a simple use of ensembled prompts provides slight gains, with " + }, + { + "bbox": [ + 55, + 293, + 295, + 544 + ], + "type": "inline_equation", + "content": "0.3\\%" + }, + { + "bbox": [ + 55, + 293, + 295, + 544 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 55, + 293, + 295, + 544 + ], + "type": "inline_equation", + "content": "0.4\\%" + }, + { + "bbox": [ + 55, + 293, + 295, + 544 + ], + "type": "text", + "content": " improvements compared to manual prompts {\"class\"} and \"a photo of a {class}\". Moreover, EP even performs worse than the carefully designed prompt: \"a {domain}\" of a {class)\". However, a set of diverse prompts enables us to imitate multiple domains and learn domain-invariant features. Our overall method achieves consistently better performance across all four datasets, with notable gains in challenging cases like TI and VLCS. Such results demonstrate the effectiveness of our multi-domain imitation method." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 55, + 545, + 295, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 545, + 295, + 713 + ], + "spans": [ + { + "bbox": [ + 55, + 545, + 295, + 713 + ], + "type": "text", + "content": "Hyperparameter analysis. Additionally, we investigate the hyperparameter for the number of selected templates " + }, + { + "bbox": [ + 55, + 545, + 295, + 713 + ], + "type": "inline_equation", + "content": "m" + }, + { + "bbox": [ + 55, + 545, + 295, + 713 + ], + "type": "text", + "content": " and the ratio of channels suppressed during gradient attribution. Table 5 presents analysis on the number of prompts and its impact on the average performance across four datasets. We evaluate different values for " + }, + { + "bbox": [ + 55, + 545, + 295, + 713 + ], + "type": "inline_equation", + "content": "m" + }, + { + "bbox": [ + 55, + 545, + 295, + 713 + ], + "type": "text", + "content": ", ranging from 10 to 80. The results show that increasing " + }, + { + "bbox": [ + 55, + 545, + 295, + 713 + ], + "type": "inline_equation", + "content": "m" + }, + { + "bbox": [ + 55, + 545, + 295, + 713 + ], + "type": "text", + "content": " improves performance at first, when " + }, + { + "bbox": [ + 55, + 545, + 295, + 713 + ], + "type": "inline_equation", + "content": "m = 40" + }, + { + "bbox": [ + 55, + 545, + 295, + 713 + ], + "type": "text", + "content": ", the performance reaches its maximum value " + }, + { + "bbox": [ + 55, + 545, + 295, + 713 + ], + "type": "inline_equation", + "content": "83.2\\%" + }, + { + "bbox": [ + 55, + 545, + 295, + 713 + ], + "type": "text", + "content": ". Using more templates ensures that the imitated domains contain more complex contexts, thus facilitating our model in learning robust features. However, further increases " + }, + { + "bbox": [ + 55, + 545, + 295, + 713 + ], + "type": "inline_equation", + "content": "m" + }, + { + "bbox": [ + 55, + 545, + 295, + 713 + ], + "type": "text", + "content": " reduce the performance, with the average accuracy dropping to " + }, + { + "bbox": [ + 55, + 545, + 295, + 713 + ], + "type": "inline_equation", + "content": "83.0\\%" + }, + { + "bbox": [ + 55, + 545, + 295, + 713 + ], + "type": "text", + "content": " at " + }, + { + "bbox": [ + 55, + 545, + 295, + 713 + ], + "type": "inline_equation", + "content": "m = 60" + }, + { + "bbox": [ + 55, + 545, + 295, + 713 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 55, + 545, + 295, + 713 + ], + "type": "inline_equation", + "content": "82.3\\%" + }, + { + "bbox": [ + 55, + 545, + 295, + 713 + ], + "type": "text", + "content": " at " + }, + { + "bbox": [ + 55, + 545, + 295, + 713 + ], + "type": "inline_equation", + "content": "m = 80" + }, + { + "bbox": [ + 55, + 545, + 295, + 713 + ], + "type": "text", + "content": ". Since we do not design" + } + ] + } + ], + "index": 4 + }, + { + "type": "table", + "bbox": [ + 321, + 91, + 548, + 126 + ], + "blocks": [ + { + "bbox": [ + 318, + 71, + 550, + 83 + ], + "lines": [ + { + "bbox": [ + 318, + 71, + 550, + 83 + ], + "spans": [ + { + "bbox": [ + 318, + 71, + 550, + 83 + ], + "type": "text", + "content": "Table 5. Hyperparameter analysis on the number of prompts " + }, + { + "bbox": [ + 318, + 71, + 550, + 83 + ], + "type": "inline_equation", + "content": "m" + }, + { + "bbox": [ + 318, + 71, + 550, + 83 + ], + "type": "text", + "content": " ." + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 321, + 91, + 548, + 126 + ], + "lines": [ + { + "bbox": [ + 321, + 91, + 548, + 126 + ], + "spans": [ + { + "bbox": [ + 321, + 91, + 548, + 126 + ], + "type": "table", + "html": "
m1020406080
Avg82.782.983.283.082.3
", + "image_path": "0b56d6081b2c6d415ffb5fcb2cd5baffc64fd4c578d510f1bf27628575a257d7.jpg" + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "table_body" + } + ], + "index": 6 + }, + { + "type": "table", + "bbox": [ + 321, + 165, + 548, + 200 + ], + "blocks": [ + { + "bbox": [ + 313, + 133, + 553, + 156 + ], + "lines": [ + { + "bbox": [ + 313, + 133, + 553, + 156 + ], + "spans": [ + { + "bbox": [ + 313, + 133, + 553, + 156 + ], + "type": "text", + "content": "Table 6. Hyperparameter analysis on the ratio of channels suppressed during gradient attribution." + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 321, + 165, + 548, + 200 + ], + "lines": [ + { + "bbox": [ + 321, + 165, + 548, + 200 + ], + "spans": [ + { + "bbox": [ + 321, + 165, + 548, + 200 + ], + "type": "table", + "html": "
Ratio0.10.20.30.40.5
Avg82.383.282.081.781.6
", + "image_path": "89503814643621f1f40377a0c804bbcb7359e2c4ce321752048e4711a8a11e21.jpg" + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "table_body" + } + ], + "index": 8 + }, + { + "bbox": [ + 313, + 212, + 553, + 258 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 212, + 553, + 258 + ], + "spans": [ + { + "bbox": [ + 313, + 212, + 553, + 258 + ], + "type": "text", + "content": "different templates for different datasets, utilizing all of the templates without selection introduces noise, leading to a drop in performance. Therefore, we set the hyperparameter " + }, + { + "bbox": [ + 313, + 212, + 553, + 258 + ], + "type": "inline_equation", + "content": "m" + }, + { + "bbox": [ + 313, + 212, + 553, + 258 + ], + "type": "text", + "content": " as 40." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 313, + 261, + 555, + 429 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 261, + 555, + 429 + ], + "spans": [ + { + "bbox": [ + 313, + 261, + 555, + 429 + ], + "type": "text", + "content": "Table 6 presents the analysis on the ratio of channels suppressed during gradient attribution and its effect on the overall performance across four datasets. We evaluate different suppression ratios, ranging from 0.1 to 0.5. The results show that increasing the suppression ratio improves performance at first, with the maximum average accuracy of " + }, + { + "bbox": [ + 313, + 261, + 555, + 429 + ], + "type": "inline_equation", + "content": "83.2\\%" + }, + { + "bbox": [ + 313, + 261, + 555, + 429 + ], + "type": "text", + "content": " achieved at a suppression ratio of 0.2. However, further increases in the suppression ratio slightly reduce accuracy, with the average performance dropping to " + }, + { + "bbox": [ + 313, + 261, + 555, + 429 + ], + "type": "inline_equation", + "content": "82.0\\%" + }, + { + "bbox": [ + 313, + 261, + 555, + 429 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 313, + 261, + 555, + 429 + ], + "type": "inline_equation", + "content": "81.7\\%" + }, + { + "bbox": [ + 313, + 261, + 555, + 429 + ], + "type": "text", + "content": ", and " + }, + { + "bbox": [ + 313, + 261, + 555, + 429 + ], + "type": "inline_equation", + "content": "81.6\\%" + }, + { + "bbox": [ + 313, + 261, + 555, + 429 + ], + "type": "text", + "content": " at 0.3, 0.4, and 0.5. The experiments demonstrate that excessive suppression filters out essential information, resulting in degraded performance. Thus, we set the suppression ratio to 0.2 as the optimal value for this hyperparameter." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 313, + 442, + 388, + 455 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 442, + 388, + 455 + ], + "spans": [ + { + "bbox": [ + 313, + 442, + 388, + 455 + ], + "type": "text", + "content": "6. Conclusion" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 313, + 463, + 555, + 630 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 463, + 555, + 630 + ], + "spans": [ + { + "bbox": [ + 313, + 463, + 555, + 630 + ], + "type": "text", + "content": "We develop a new domain generalization approach by learning with diverse text prompts. We use these prompts to imitate multiple domains, which guide the DG model in learning domain-invariant features through multi-modal alignment in CLIP. These prompts contribute to three aspects, including suppressing domain-sensitive features, ensure feature consistency with varied domains, and enable feature diversification in the DG model. By incorporating these three terms, our approach reduces the CLIP-based DG model's tendency to capture domain-sensitive details and mitigates bias towards the source data distribution. Experimental results on five domain generalization datasets from the DomainBed benchmark demonstrate the effectiveness of our method." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 313, + 644, + 424, + 658 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 644, + 424, + 658 + ], + "spans": [ + { + "bbox": [ + 313, + 644, + 424, + 658 + ], + "type": "text", + "content": "7. Acknowledgments" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 313, + 665, + 554, + 712 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 665, + 554, + 712 + ], + "spans": [ + { + "bbox": [ + 313, + 665, + 554, + 712 + ], + "type": "text", + "content": "This work was supported by the NSFC under Grant 62322604 and 62176159, and in part by the Shanghai Municipal Science and Technology Major Project under Grant 2021SHZDZX0102." + } + ] + } + ], + "index": 14 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 749, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 749, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 749, + 315, + 757 + ], + "type": "text", + "content": "9566" + } + ] + } + ], + "index": 15 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 7 + }, + { + "para_blocks": [ + { + "bbox": [ + 56, + 71, + 115, + 83 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 71, + 115, + 83 + ], + "spans": [ + { + "bbox": [ + 56, + 71, + 115, + 83 + ], + "type": "text", + "content": "References" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 57, + 91, + 296, + 713 + ], + "type": "list", + "angle": 0, + "index": 16, + "blocks": [ + { + "bbox": [ + 61, + 91, + 296, + 144 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 61, + 91, + 296, + 144 + ], + "spans": [ + { + "bbox": [ + 61, + 91, + 296, + 144 + ], + "type": "text", + "content": "[1] Sravanti Addepalli, Ashish Ramayee Asokan, Lakshay Sharma, and R Venkatesh Babu. Leveraging vision-language models for improving domain generalization in image classification. In CVPR, pages 23922-23932, 2024. 1, 2, 3, 6, 7" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 61, + 146, + 296, + 201 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 61, + 146, + 296, + 201 + ], + "spans": [ + { + "bbox": [ + 61, + 146, + 296, + 201 + ], + "type": "text", + "content": "[2] James Urquhart Allingham, Jie Ren, Michael W Dusenberry, Xiuye Gu, Yin Cui, Dustin Tran, Jeremiah Zhe Liu, and Balaji Lakshminarayanan. A simple zero-shot prompt weighting technique to improve prompt ensembling in text-image models. In ICML, 2023. 4" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 62, + 202, + 295, + 235 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 202, + 295, + 235 + ], + "spans": [ + { + "bbox": [ + 62, + 202, + 295, + 235 + ], + "type": "text", + "content": "[3] Martin Arjovsky, Léon Bottou, Ishaan Gulrajani, and David Lopez-Paz. Invariant risk minimization. arXiv preprint arXiv:1907.02893, 2019. 2" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 62, + 236, + 295, + 268 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 236, + 295, + 268 + ], + "spans": [ + { + "bbox": [ + 62, + 236, + 295, + 268 + ], + "type": "text", + "content": "[4] Shuanghao Bai, Yuedi Zhang, Wanqi Zhou, Zhirong Luan, and Badong Chen. Soft prompt generation for domain generalization. In ECCV, pages 434-450, 2024. 3, 6, 7" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 62, + 270, + 295, + 291 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 270, + 295, + 291 + ], + "spans": [ + { + "bbox": [ + 62, + 270, + 295, + 291 + ], + "type": "text", + "content": "[5] Sara Beery, Grant Van Horn, and Pietro Perona. Recognition in terra incognita. In ECCV, pages 456-473, 2018. 6, 7" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 62, + 293, + 295, + 335 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 293, + 295, + 335 + ], + "spans": [ + { + "bbox": [ + 62, + 293, + 295, + 335 + ], + "type": "text", + "content": "[6] Junbum Cha, Sanghyuk Chun, Kyungjae Lee, Han-Cheol Cho, Seunghyun Park, Yunsung Lee, and Sungrae Park. Swad: Domain generalization by seeking flat minima. NeurIPS, 34:22405-22418, 2021. 1, 2, 6, 7" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 62, + 336, + 295, + 378 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 336, + 295, + 378 + ], + "spans": [ + { + "bbox": [ + 62, + 336, + 295, + 378 + ], + "type": "text", + "content": "[7] Junbum Cha, Kyungjae Lee, Sungrae Park, and Sanghyuk Chun. Domain generalization by mutual-information regularization with pre-trained models. In ECCV, pages 440-457, 2022. 1, 6" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 62, + 380, + 295, + 413 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 380, + 295, + 413 + ], + "spans": [ + { + "bbox": [ + 62, + 380, + 295, + 413 + ], + "type": "text", + "content": "[8] Liang Chen, Yong Zhang, Yibing Song, Ying Shan, and Lingqiao Liu. Improved test-time adaptation for domain generalization. In CVPR, pages 24172-24182, 2023. 2" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 62, + 415, + 295, + 446 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 415, + 295, + 446 + ], + "spans": [ + { + "bbox": [ + 62, + 415, + 295, + 446 + ], + "type": "text", + "content": "[9] Liang Chen, Yong Zhang, Yibing Song, Anton Van Den Hengel, and Lingqiao Liu. Domain generalization via rationale invariance. In ICCV, pages 1751-1760, 2023. 1" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 57, + 448, + 295, + 490 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 448, + 295, + 490 + ], + "spans": [ + { + "bbox": [ + 57, + 448, + 295, + 490 + ], + "type": "text", + "content": "[10] De Cheng, Zhipeng Xu, Xinyang Jiang, Nannan Wang, Dongsheng Li, and Xinbo Gao. Disentangled prompt representation for domain generalization. In CVPR, pages 23595-23604, 2024. 1, 2" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 57, + 492, + 295, + 535 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 492, + 295, + 535 + ], + "spans": [ + { + "bbox": [ + 57, + 492, + 295, + 535 + ], + "type": "text", + "content": "[11] Junhyeong Cho, Gilhyun Nam, Sungyeon Kim, Hunmin Yang, and Suha Kwak. Promptstyler: Prompt-driven style generation for source-free domain generalization. In ICCV, pages 15702-15712, 2023. 2, 6" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 57, + 536, + 295, + 590 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 536, + 295, + 590 + ], + "spans": [ + { + "bbox": [ + 57, + 536, + 295, + 590 + ], + "type": "text", + "content": "[12] Huiqi Deng, Na Zou, Mengnan Du, Weifu Chen, Guocan Feng, Ziwei Yang, Zheyang Li, and Quanshi Zhang. Unifying fourteen post-hoc attribution methods with taylor interactions. IEEE Transactions on Pattern Analysis and Machine Intelligence, 46(7):4625-4640, 2024. 4" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 57, + 592, + 295, + 635 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 592, + 295, + 635 + ], + "spans": [ + { + "bbox": [ + 57, + 592, + 295, + 635 + ], + "type": "text", + "content": "[13] Yu Ding, Lei Wang, Bin Liang, Shuming Liang, Yang Wang, and Fang Chen. Domain generalization by learning and removing domain-specific features. In NeurIPS, pages 24226-24239, 2022. 2" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 57, + 636, + 295, + 690 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 636, + 295, + 690 + ], + "spans": [ + { + "bbox": [ + 57, + 636, + 295, + 690 + ], + "type": "text", + "content": "[14] Xiaoyi Dong, Jianmin Bao, Yinglin Zheng, Ting Zhang, Dongdong Chen, Hao Yang, Ming Zeng, Weiming Zhang, Lu Yuan, Dong Chen, et al. Maskclip: Masked self-distillation advances contrastive language-image pretraining. In CVPR, pages 10995-11005, 2023. 2" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 57, + 692, + 295, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 692, + 295, + 713 + ], + "spans": [ + { + "bbox": [ + 57, + 692, + 295, + 713 + ], + "type": "text", + "content": "[15] Alexey Dosovitskiy, Lucas Beyer, Alexander Kolesnikov, Dirk Weissenborn, Xiaohua Zhai, Thomas Unterthiner," + } + ] + } + ], + "index": 15 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 316, + 73, + 553, + 713 + ], + "type": "list", + "angle": 0, + "index": 32, + "blocks": [ + { + "bbox": [ + 335, + 73, + 553, + 116 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 335, + 73, + 553, + 116 + ], + "spans": [ + { + "bbox": [ + 335, + 73, + 553, + 116 + ], + "type": "text", + "content": "Mostafa Dehghani, Matthias Minderer, Georg Heigold, Sylvain Gelly, Jakob Uszkoreit, and Neil Houlsby. An image is worth 16x16 words: Transformers for image recognition at scale. In ICLR, 2021. 3" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 316, + 118, + 553, + 161 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 118, + 553, + 161 + ], + "spans": [ + { + "bbox": [ + 316, + 118, + 553, + 161 + ], + "type": "text", + "content": "[16] Lisa Dunlap, Clara Mohri, Devin Guillory, Han Zhang, Trevor Darrell, Joseph E Gonzalez, Aditi Raghunathan, and Anna Rohrbach. Using language to extend to unseen domains. In ICLR, 2023. 2" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 317, + 163, + 553, + 206 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 163, + 553, + 206 + ], + "spans": [ + { + "bbox": [ + 317, + 163, + 553, + 206 + ], + "type": "text", + "content": "[17] Chen Fang, Ye Xu, and Daniel N Rockmore. Unbiased metric learning: On the utilization of multiple datasets and web images for softening bias. In ICCV, pages 1657-1664, 2013. 6, 7" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 316, + 209, + 553, + 262 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 209, + 553, + 262 + ], + "spans": [ + { + "bbox": [ + 316, + 209, + 553, + 262 + ], + "type": "text", + "content": "[18] Yaroslav Ganin, Evgeniya Ustinova, Hana Ajakan, Pascal Germain, Hugo Larochelle, François Laviolette, Mario March, and Victor Lempitsky. Domain-adversarial training of neural networks. Journal of machine learning research, 17(59):1-35, 2016. 2" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 316, + 264, + 553, + 308 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 264, + 553, + 308 + ], + "spans": [ + { + "bbox": [ + 316, + 264, + 553, + 308 + ], + "type": "text", + "content": "[19] Chunjiang Ge, Rui Huang, Mixue Xie, Zihang Lai, Shiji Song, Shuang Li, and Gao Huang. Domain adaptation via prompt learning. IEEE Transactions on Neural Networks and Learning Systems, pages 1-11, 2023. 3" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 316, + 309, + 553, + 352 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 309, + 553, + 352 + ], + "spans": [ + { + "bbox": [ + 316, + 309, + 553, + 352 + ], + "type": "text", + "content": "[20] Sachin Goyal, Ananya Kumar, Sankalp Garg, Zico Kolter, and Aditi Raghunathan. Finetune like you pretrain: Improved finetuning of zero-shot vision models. In CVPR, pages 19338-19347, 2023. 1, 2, 6, 7" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 316, + 354, + 553, + 374 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 354, + 553, + 374 + ], + "spans": [ + { + "bbox": [ + 316, + 354, + 553, + 374 + ], + "type": "text", + "content": "[21] Ishaan Gulrajani and David Lopez-Paz. In search of lost domain generalization. In ICLR, 2020. 2, 6" + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 316, + 376, + 553, + 409 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 376, + 553, + 409 + ], + "spans": [ + { + "bbox": [ + 316, + 376, + 553, + 409 + ], + "type": "text", + "content": "[22] Jintao Guo, Lei Qi, and Yinghuan Shi. Domaindrop: Suppressing domain-sensitive channels for domain generalization. In ICCV, pages 19114-19124, 2023. 2, 6" + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 316, + 411, + 553, + 454 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 411, + 553, + 454 + ], + "spans": [ + { + "bbox": [ + 316, + 411, + 553, + 454 + ], + "type": "text", + "content": "[23] Jiajun Hu, Jian Zhang, Lei Qi, Yinghuan Shi, and Yang Gao. Learn to preserve and diversify: Parameter-efficient group with orthogonal regularization for domain generalization. In ECCV, pages 198-216, 2024. 1, 2, 3, 6, 7" + } + ] + } + ], + "index": 25 + }, + { + "bbox": [ + 316, + 456, + 553, + 498 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 456, + 553, + 498 + ], + "spans": [ + { + "bbox": [ + 316, + 456, + 553, + 498 + ], + "type": "text", + "content": "[24] Lanqing Hu, Meina Kan, Shiguang Shan, and Xilin Chen. Dandelionnet: Domain composition with instance adaptive classification for domain generalization. In ICCV, pages 19050-19059, 2023. 1" + } + ] + } + ], + "index": 26 + }, + { + "bbox": [ + 316, + 501, + 553, + 533 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 501, + 553, + 533 + ], + "spans": [ + { + "bbox": [ + 316, + 501, + 553, + 533 + ], + "type": "text", + "content": "[25] Zenan Huang, Haobo Wang, Junbo Zhao, and Nenggan Zheng. idag: Invariant dag searching for domain generalization. In ICCV, pages 19169-19179, 2023. 1" + } + ] + } + ], + "index": 27 + }, + { + "bbox": [ + 317, + 535, + 553, + 588 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 535, + 553, + 588 + ], + "spans": [ + { + "bbox": [ + 317, + 535, + 553, + 588 + ], + "type": "text", + "content": "[26] Zeyi Huang, Andy Zhou, Zijian Ling, Mu Cai, Haohan Wang, and Yong Jae Lee. A sentence speaks a thousand images: Domain generalization through distilling clip with language guidance. In ICCV, pages 11685-11695, 2023. 1, 2, 3, 6, 7" + } + ] + } + ], + "index": 28 + }, + { + "bbox": [ + 316, + 590, + 553, + 634 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 590, + 553, + 634 + ], + "spans": [ + { + "bbox": [ + 316, + 590, + 553, + 634 + ], + "type": "text", + "content": "[27] Samyak Jain, Sravanti Addepalli, Pawan Kumar Sahu, Priyam Dey, and R Venkatesh Babu. Dart: Diversify-aggregate-repeat training improves generalization of neural networks. In CVPR, pages 16048-16059, 2023. 1, 6, 7" + } + ] + } + ], + "index": 29 + }, + { + "bbox": [ + 316, + 635, + 553, + 689 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 635, + 553, + 689 + ], + "spans": [ + { + "bbox": [ + 316, + 635, + 553, + 689 + ], + "type": "text", + "content": "[28] Chao Jia, Yinfei Yang, Ye Xia, Yi-Ting Chen, Zarana Parekh, Hieu Pham, Quoc Le, Yun-Hsuan Sung, Zhen Li, and Tom Duerig. Scaling up visual and vision-language representation learning with noisy text supervision. In ICML, pages 4904-4916, 2021. 2" + } + ] + } + ], + "index": 30 + }, + { + "bbox": [ + 316, + 691, + 553, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 691, + 553, + 713 + ], + "spans": [ + { + "bbox": [ + 316, + 691, + 553, + 713 + ], + "type": "text", + "content": "[29] Muhammad Uzair Khattak, Hanoona Rasheed, Muhammad Maaz, Salman Khan, and Fahad Shahbaz Khan. Maple:" + } + ] + } + ], + "index": 31 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "text", + "content": "9567" + } + ] + } + ], + "index": 33 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 8 + }, + { + "para_blocks": [ + { + "bbox": [ + 56, + 72, + 295, + 712 + ], + "type": "list", + "angle": 0, + "index": 17, + "blocks": [ + { + "bbox": [ + 77, + 72, + 294, + 95 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 77, + 72, + 294, + 95 + ], + "spans": [ + { + "bbox": [ + 77, + 72, + 294, + 95 + ], + "type": "text", + "content": "Multi-modal prompt learning. In CVPR, pages 19113-19122, 2023. 2, 3, 7" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 56, + 97, + 295, + 140 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 97, + 295, + 140 + ], + "spans": [ + { + "bbox": [ + 56, + 97, + 295, + 140 + ], + "type": "text", + "content": "[30] Ananya Kumar, Aditi Raghunathan, Robbie Jones, Tengyu Ma, and Percy Liang. Fine-tuning can distort pretrained features and underperform out-of-distribution. In ICLR, 2022. 2, 6, 7" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 56, + 142, + 294, + 175 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 142, + 294, + 175 + ], + "spans": [ + { + "bbox": [ + 56, + 142, + 294, + 175 + ], + "type": "text", + "content": "[31] Chenming Li, Daoan Zhang, Wenjian Huang, and Jianguo Zhang. Cross contrasting feature perturbation for domain generalization. In ICCV, pages 1327-1337, 2023. 2" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 56, + 177, + 294, + 210 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 177, + 294, + 210 + ], + "spans": [ + { + "bbox": [ + 56, + 177, + 294, + 210 + ], + "type": "text", + "content": "[32] Da Li, Yongxin Yang, Yi-Zhe Song, and Timothy M Hospedales. Deeper, broader and artier domain generalization. In ICCV, pages 5542-5550, 2017. 6, 7" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 56, + 212, + 294, + 243 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 212, + 294, + 243 + ], + "spans": [ + { + "bbox": [ + 56, + 212, + 294, + 243 + ], + "type": "text", + "content": "[33] Haoliang Li, Sinno Jialin Pan, Shiqi Wang, and Alex C Kot. Domain generalization with adversarial feature learning. In CVPR, pages 5400-5409, 2018. 2" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 56, + 246, + 294, + 289 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 246, + 294, + 289 + ], + "spans": [ + { + "bbox": [ + 56, + 246, + 294, + 289 + ], + "type": "text", + "content": "[34] Junnan Li, Dongxu Li, Caiming Xiong, and Steven Hoi. Blip: Bootstrapping language-image pre-training for unified vision-language understanding and generation. In ICML, pages 12888-12900, 2022. 1, 2" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 56, + 291, + 294, + 335 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 291, + 294, + 335 + ], + "spans": [ + { + "bbox": [ + 56, + 291, + 294, + 335 + ], + "type": "text", + "content": "[35] Junnan Li, Dongxu Li, Silvio Savarese, and Steven Hoi. Blip-2: Bootstrapping language-image pre-training with frozen image encoders and large language models. In ICML, pages 19730–19742, 2023. 1, 2" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 56, + 337, + 294, + 370 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 337, + 294, + 370 + ], + "spans": [ + { + "bbox": [ + 56, + 337, + 294, + 370 + ], + "type": "text", + "content": "[36] Xiaotong Li, Yongxing Dai, Yixiao Ge, Jun Liu, Ying Shan, and Ling-Yu Duan. Uncertainty modeling for out-of-distribution generalization. In ICLR, 2022. 2" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 56, + 372, + 294, + 404 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 372, + 294, + 404 + ], + "spans": [ + { + "bbox": [ + 56, + 372, + 294, + 404 + ], + "type": "text", + "content": "[37] Ya Li, Mingming Gong, Xinmei Tian, Tongliang Liu, and Dacheng Tao. Domain generalization via conditional invariant representations. In AAAI, 2018. 2" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 56, + 406, + 294, + 449 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 406, + 294, + 449 + ], + "spans": [ + { + "bbox": [ + 56, + 406, + 294, + 449 + ], + "type": "text", + "content": "[38] Ya Li, Xinmei Tian, Mingming Gong, Yajing Liu, Tongliang Liu, Kun Zhang, and Dacheng Tao. Deep domain generalization via conditional invariant adversarial networks. In ECCV, pages 624-639, 2018. 2" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 56, + 452, + 294, + 484 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 452, + 294, + 484 + ], + "spans": [ + { + "bbox": [ + 56, + 452, + 294, + 484 + ], + "type": "text", + "content": "[39] Krikamol Muandet, David Balduzzi, and Bernhard Schölkopf. Domain generalization via invariant feature representation. In ICML, pages 10-18, 2013. 2" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 56, + 486, + 294, + 518 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 486, + 294, + 518 + ], + "spans": [ + { + "bbox": [ + 56, + 486, + 294, + 518 + ], + "type": "text", + "content": "[40] Hyeonseob Nam, HyunJae Lee, Jongchan Park, Wonjun Yoon, and Donggeun Yoo. Reducing domain gap by reducing style bias. In CVPR, pages 8690-8699, 2021. 2" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 56, + 520, + 294, + 553 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 520, + 294, + 553 + ], + "spans": [ + { + "bbox": [ + 56, + 520, + 294, + 553 + ], + "type": "text", + "content": "[41] Hongjing Niu, Hanting Li, Feng Zhao, and Bin Li. Domain-unified prompt representations for source-free domain generalization. arXiv preprint arXiv:2209.14926, 2022. 2" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 56, + 555, + 294, + 587 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 555, + 294, + 587 + ], + "spans": [ + { + "bbox": [ + 56, + 555, + 294, + 587 + ], + "type": "text", + "content": "[42] Oren Nuriel, Sagie Benaim, and Lior Wolf. Permuted adain: Reducing the bias towards global statistics in image classification. In CVPR, pages 9482-9491, 2021. 2" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 56, + 589, + 294, + 633 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 589, + 294, + 633 + ], + "spans": [ + { + "bbox": [ + 56, + 589, + 294, + 633 + ], + "type": "text", + "content": "[43] Changdae Oh, Junhyuk So, Hoyoon Byun, YongTaek Lim, Minchul Shin, Jong-June Jeon, and Kyungwoo Song. Geodesic multi-modal mixup for robust fine-tuning. In NeurIPS, 2024. 1" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 56, + 635, + 294, + 668 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 635, + 294, + 668 + ], + "spans": [ + { + "bbox": [ + 56, + 635, + 294, + 668 + ], + "type": "text", + "content": "[44] Xingchao Peng, Qinxun Bai, Xide Xia, Zijun Huang, Kate Saenko, and Bo Wang. Moment matching for multi-source domain adaptation. In ICCV, pages 1406-1415, 2019. 6" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 56, + 670, + 294, + 712 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 670, + 294, + 712 + ], + "spans": [ + { + "bbox": [ + 56, + 670, + 294, + 712 + ], + "type": "text", + "content": "[45] Zelin Peng, Zhengqin Xu, Zhilin Zeng, Lingxi Xie, Qi Tian, and Wei Shen. Parameter efficient fine-tuning via cross block orchestration for segment anything model. In CVPR, pages 3743-3752, 2024. 2" + } + ] + } + ], + "index": 16 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 316, + 72, + 553, + 715 + ], + "type": "list", + "angle": 0, + "index": 34, + "blocks": [ + { + "bbox": [ + 316, + 72, + 553, + 106 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 72, + 553, + 106 + ], + "spans": [ + { + "bbox": [ + 316, + 72, + 553, + 106 + ], + "type": "text", + "content": "[46] Shi Pu, Kaili Zhao, and Mao Zheng. Alignment-uniformity aware representation learning for zero-shot video classification. In CVPR, pages 19968-19977, 2022. 1, 5" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 316, + 107, + 553, + 140 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 107, + 553, + 140 + ], + "spans": [ + { + "bbox": [ + 316, + 107, + 553, + 140 + ], + "type": "text", + "content": "[47] Daiqing Qi, Handong Zhao, Aidong Zhang, and Sheng Li. Generalizing to unseen domains via text-guided augmentation. In ECCV, pages 285-300, 2024. 2" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 316, + 141, + 553, + 174 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 141, + 553, + 174 + ], + "spans": [ + { + "bbox": [ + 316, + 141, + 553, + 174 + ], + "type": "text", + "content": "[48] Minghan Qin, Wanhua Li, Jiawei Zhou, Haoqian Wang, and Hanspeter Pfister. Langsplat: 3d language gaussian splatting. In CVPR, pages 20051-20060, 2024. 2" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 316, + 175, + 553, + 229 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 175, + 553, + 229 + ], + "spans": [ + { + "bbox": [ + 316, + 175, + 553, + 229 + ], + "type": "text", + "content": "[49] Alec Radford, Jong Wook Kim, Chris Hallacy, Aditya Ramesh, Gabriel Goh, Sandhini Agarwal, Girish Sastry, Amanda Askell, Pamela Mishkin, Jack Clark, et al. Learning transferable visual models from natural language supervision. In ICML, pages 8748-8763, 2021. 1, 2, 3, 6, 7" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 316, + 231, + 553, + 274 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 231, + 553, + 274 + ], + "spans": [ + { + "bbox": [ + 316, + 231, + 553, + 274 + ], + "type": "text", + "content": "[50] Yang Shu, Xingzhuo Guo, Jialong Wu, Ximei Wang, Jianmin Wang, and Mingsheng Long. Clipood: Generalizing clip to out-of-distributions. In ICML, pages 31716-31731, 2023. 1, 2, 6, 7" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 316, + 276, + 553, + 319 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 276, + 553, + 319 + ], + "spans": [ + { + "bbox": [ + 316, + 276, + 553, + 319 + ], + "type": "text", + "content": "[51] Mainak Singha, Ankit Jha, Shirsha Bose, Ashwin Nair, Moloud Abdar, and Biplab Banerjee. Unknown prompt the only lacuna: Unveiling clip's potential for open domain generalization. In CVPR, pages 13309-13319, 2024. 1" + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 316, + 320, + 553, + 354 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 320, + 553, + 354 + ], + "spans": [ + { + "bbox": [ + 316, + 320, + 553, + 354 + ], + "type": "text", + "content": "[52] Quan Sun, Yuxin Fang, Ledell Wu, Xinlong Wang, and Yue Cao. Eva-clip: Improved training techniques for clip at scale. arXiv preprint arXiv:2303.15389, 2023. 1, 2" + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 316, + 354, + 553, + 387 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 354, + 553, + 387 + ], + "spans": [ + { + "bbox": [ + 316, + 354, + 553, + 387 + ], + "type": "text", + "content": "[53] Vladimir N Vapnik. An overview of statistical learning theory. IEEE transactions on neural networks, 10(5):988-999, 1999. 2" + } + ] + } + ], + "index": 25 + }, + { + "bbox": [ + 316, + 388, + 553, + 431 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 388, + 553, + 431 + ], + "spans": [ + { + "bbox": [ + 316, + 388, + 553, + 431 + ], + "type": "text", + "content": "[54] Ashish Vaswani, Noam Shazeer, Niki Parmar, Jakob Uszkoreit, Llion Jones, Aidan N Gomez, Lukasz Kaiser, and Illia Polosukhin. Attention is all you need. In NeurIPS, pages 5998-6008, 2017. 3" + } + ] + } + ], + "index": 26 + }, + { + "bbox": [ + 316, + 433, + 553, + 476 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 433, + 553, + 476 + ], + "spans": [ + { + "bbox": [ + 316, + 433, + 553, + 476 + ], + "type": "text", + "content": "[55] Hemanth Venkateswara, Jose Eusebio, Shayok Chakraborty, and Sethuraman Panchanathan. Deep hashing network for unsupervised domain adaptation. In CVPR, pages 5018-5027, 2017. 6, 7" + } + ] + } + ], + "index": 27 + }, + { + "bbox": [ + 316, + 478, + 553, + 532 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 478, + 553, + 532 + ], + "spans": [ + { + "bbox": [ + 316, + 478, + 553, + 532 + ], + "type": "text", + "content": "[56] Mengzhu Wang, Yuehua Liu, Jianlong Yuan, Shanshan Wang, Zhibin Wang, and Wei Wang. Inter-class and inter-domain semantic augmentation for domain generalization. IEEE Transactions on Image Processing, 33:1338-1347, 2024. 2" + } + ] + } + ], + "index": 28 + }, + { + "bbox": [ + 316, + 533, + 553, + 567 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 533, + 553, + 567 + ], + "spans": [ + { + "bbox": [ + 316, + 533, + 553, + 567 + ], + "type": "text", + "content": "[57] Pengfei Wang, Zhaoxiang Zhang, Zhen Lei, and Lei Zhang. Sharpness-aware gradient matching for domain generalization. In CVPR, pages 3769-3778, 2023. 6" + } + ] + } + ], + "index": 29 + }, + { + "bbox": [ + 316, + 568, + 553, + 601 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 568, + 553, + 601 + ], + "spans": [ + { + "bbox": [ + 316, + 568, + 553, + 601 + ], + "type": "text", + "content": "[58] Tongzhou Wang and Phillip Isola. Understanding contrastive representation learning through alignment and uniformity on the hypersphere. In ICML, pages 9929-9939, 2020. 1" + } + ] + } + ], + "index": 30 + }, + { + "bbox": [ + 316, + 601, + 553, + 633 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 601, + 553, + 633 + ], + "spans": [ + { + "bbox": [ + 316, + 601, + 553, + 633 + ], + "type": "text", + "content": "[59] Yufei Wang, Haoliang Li, and Alex C Kot. Heterogeneous domain generalization via domain mixup. In ICASSP, pages 3622-3626, 2020. 2" + } + ] + } + ], + "index": 31 + }, + { + "bbox": [ + 316, + 635, + 553, + 689 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 635, + 553, + 689 + ], + "spans": [ + { + "bbox": [ + 316, + 635, + 553, + 689 + ], + "type": "text", + "content": "[60] Mitchell Wortsman, Gabriel Ilharco, Jong Wook Kim, Mike Li, Simon Kornblith, Rebecca Roelofs, Raphael Gontijo Lopes, Hannaneh Hajishirzi, Ali Farhadi, Hongseok Namkoong, et al. Robust fine-tuning of zero-shot models. In CVPR, pages 7959-7971, 2022. 1, 2, 6, 7" + } + ] + } + ], + "index": 32 + }, + { + "bbox": [ + 316, + 691, + 553, + 715 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 691, + 553, + 715 + ], + "spans": [ + { + "bbox": [ + 316, + 691, + 553, + 715 + ], + "type": "text", + "content": "[61] Sitong Wu, Haoru Tan, Zhuotao Tian, Yukang Chen, Xiaojuan Qi, and Jiaya Jia. Saco loss: Sample-wise affinity con" + } + ] + } + ], + "index": 33 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 749, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 749, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 749, + 315, + 757 + ], + "type": "text", + "content": "9568" + } + ] + } + ], + "index": 35 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 9 + }, + { + "para_blocks": [ + { + "bbox": [ + 56, + 72, + 295, + 251 + ], + "type": "list", + "angle": 0, + "index": 5, + "blocks": [ + { + "bbox": [ + 76, + 72, + 294, + 94 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 76, + 72, + 294, + 94 + ], + "spans": [ + { + "bbox": [ + 76, + 72, + 294, + 94 + ], + "type": "text", + "content": "sistency for vision-language pre-training. In CVPR, pages 27358-27369, 2024. 1, 2" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 56, + 96, + 295, + 140 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 96, + 295, + 140 + ], + "spans": [ + { + "bbox": [ + 56, + 96, + 295, + 140 + ], + "type": "text", + "content": "[62] Xin Zhang, Shixiang Shane Gu, Yutaka Matsuo, and Yusuke Iwasawa. Domain prompt learning for efficiently adapting clip to unseen domains. Transactions of the Japanese Society for Artificial Intelligence, 38(6), 2023. 3, 6, 7" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 56, + 141, + 294, + 174 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 141, + 294, + 174 + ], + "spans": [ + { + "bbox": [ + 56, + 141, + 294, + 174 + ], + "type": "text", + "content": "[63] Xingxuan Zhang, Renzhe Xu, Han Yu, Yancheng Dong, Pengfei Tian, and Peng Cui. Flatness-aware minimization for domain generalization. In ICCV, pages 5189-5202, 2023. 2" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 56, + 175, + 294, + 207 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 175, + 294, + 207 + ], + "spans": [ + { + "bbox": [ + 56, + 175, + 294, + 207 + ], + "type": "text", + "content": "[64] Kaiyang Zhou, Jingkang Yang, Chen Change Loy, and Zwei Liu. Conditional prompt learning for vision-language models. In CVPR, pages 16816-16825, 2022. 2, 3, 7" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 56, + 209, + 294, + 251 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 209, + 294, + 251 + ], + "spans": [ + { + "bbox": [ + 56, + 209, + 294, + 251 + ], + "type": "text", + "content": "[65] Kaiyang Zhou, Jingkang Yang, Chen Change Loy, and Ziwei Liu. Learning to prompt for vision-language models. International Journal of Computer Vision, 130(9):2337-2348, 2022. 2, 3, 7" + } + ] + } + ], + "index": 4 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 749, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 749, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 749, + 315, + 757 + ], + "type": "text", + "content": "9569" + } + ] + } + ], + "index": 6 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 10 + } + ], + "_backend": "vlm", + "_version_name": "2.6.4" +} \ No newline at end of file diff --git a/2025/Don't Shake the Wheel_ Momentum-Aware Planning in End-to-End Autonomous Driving/85994b24-3b2b-443a-a1bf-eabbda00b06b_content_list.json b/2025/Don't Shake the Wheel_ Momentum-Aware Planning in End-to-End Autonomous Driving/85994b24-3b2b-443a-a1bf-eabbda00b06b_content_list.json new file mode 100644 index 0000000000000000000000000000000000000000..a3ab5112299d252cb92ecf63e56ff965edfb5210 --- /dev/null +++ b/2025/Don't Shake the Wheel_ Momentum-Aware Planning in End-to-End Autonomous Driving/85994b24-3b2b-443a-a1bf-eabbda00b06b_content_list.json @@ -0,0 +1,1314 @@ +[ + { + "type": "text", + "text": "Don't Shake the Wheel: Momentum-Aware Planning in End-to-End Autonomous Driving", + "text_level": 1, + "bbox": [ + 218, + 130, + 779, + 176 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Ziying Song $^{1,2,\\dagger}$ , Caiyan Jia $^{1,2,\\star}$ , Lin Liu $^{1,2}$ , Hongyu Pan $^{3}$ , Yongchang Zhang $^{3}$ , Junming Wang $^{3,7}$ , Xingyu Zhang $^{3}$ , Shaoqing Xu $^{4}$ , Lei Yang $^{5}$ , Yadan Luo $^{6,\\star}$", + "bbox": [ + 184, + 202, + 810, + 239 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "$^{1}$ School of Computer Science and Technology, Beijing Jiaotong University", + "bbox": [ + 200, + 244, + 795, + 263 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "$^{2}$ Beijing Key Laboratory of Traffic Data Mining and Embodied Intelligence", + "bbox": [ + 200, + 263, + 797, + 281 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "$^{3}$ Horizon Robotics $^{4}$ University of Macau $^{5}$ THU $^{6}$ The University of Queensland $^{7}$ HKU", + "bbox": [ + 158, + 281, + 836, + 299 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "{songziying, cyjia}@bjtu.edu.cn, y.luo@uq.edu.au.", + "bbox": [ + 272, + 301, + 702, + 316 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Abstract", + "text_level": 1, + "bbox": [ + 246, + 345, + 326, + 362 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "End-to-end autonomous driving frameworks enable seamless integration of perception and planning but often rely on one-shot trajectory prediction, which may lead to unstable control and vulnerability to occlusions in single-frame perception. To address this, we propose the Momentum-Aware Driving (MomAD) framework, which introduces trajectory momentum and perception momentum to stabilize and refine trajectory predictions. MomAD comprises two core components: (1) Topological Trajectory Matching (TTM) employs Hausdorff Distance to select the optimal planning query that aligns with prior paths to ensure coherence; (2) Momentum Planning Interactor (MPI) cross-attends the selected planning query with historical queries to expand static and dynamic perception files. This enriched query, in turn, helps regenerate long-horizon trajectory and reduce collision risks. To mitigate noise arising from dynamic environments and detection errors, we introduce robust instance denoising during training, enabling the planning model to focus on critical signals and improve its robustness. We also propose a novel Trajectory Prediction Consistency (TPC) metric to quantitatively assess planning stability. Experiments on the nuScenes dataset demonstrate that MomAD achieves superior long-term consistency $(\\geq 3s)$ compared to SOTA methods. Moreover, evaluations on the curated Turning-nuScenes shows that MomAD reduces the collision rate by $26\\%$ and improves TPC by $0.97m$ $(33.45\\%)$ over a 6s prediction horizon, while closed-loop on Bench2Drive demonstrates an up to $16.3\\%$ improvement in success rate. The source code is available at https://github.com/adept-thu/MomAD.", + "bbox": [ + 88, + 388, + 483, + 843 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/04df1ebe27a0d8377e1cf1573ee359cf62c3b0c9f81069686903c6fc238a7f4f.jpg", + "image_caption": [ + "(a) Deterministic Planning" + ], + "image_footnote": [], + "bbox": [ + 517, + 345, + 635, + 516 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/c5aa0b8cde25aa4b247a21da4796663b7f51e4a85282ea312f3c08b31ad8dc78.jpg", + "image_caption": [ + "(b) Multi-modal Trajectory" + ], + "image_footnote": [], + "bbox": [ + 638, + 345, + 758, + 516 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/dc4ff25f585a9c2a8a9a67ed08cfb4ac1495c8745eb4c547abeb6457741bbdc4.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 758, + 345, + 900, + 516 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/d03e39cae0f6cbaeb15ee3f345b3e5e60680f2e7c34687159517be322f33550b.jpg", + "image_caption": [ + "(c) Momentum Planning (Ours)" + ], + "image_footnote": [], + "bbox": [ + 519, + 541, + 638, + 703 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/cee37a50b1899f64c3cb4a28cf30f60ca14efe4d6183c5c741efd6d03d4a6396.jpg", + "image_caption": [ + "Figure 1. (a) Deterministic Planning [14, 15, 19, 32] predicts deterministic trajectories, but lacks action diversity, posing safety risks. (b) Multi-modal Trajectory Planning [4, 27, 28] selects the highest-scoring trajectory among the multi-modal trajectories, yet fails to ensure stability and consistency, having risks in vehicle trembling. (c) Momentum Planning leverages the trajectory and perception momentum to enhance current planning through historical guidance to overcome temporal inconsistency." + ], + "image_footnote": [], + "bbox": [ + 640, + 541, + 777, + 703 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/f5ead49f269af7f21c98a269e9b431bc77482c0b52323ee03525bcb78a2b5eb4.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 784, + 541, + 903, + 703 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "CVF", + "bbox": [ + 106, + 2, + 181, + 42 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore.", + "bbox": [ + 236, + 0, + 810, + 46 + ], + "page_idx": 0 + }, + { + "type": "page_footnote", + "text": "† Intern of Horizon Robotics, † Corresponding author.", + "bbox": [ + 114, + 888, + 403, + 902 + ], + "page_idx": 0 + }, + { + "type": "page_number", + "text": "22432", + "bbox": [ + 478, + 944, + 519, + 957 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "1. Introduction", + "text_level": 1, + "bbox": [ + 91, + 90, + 222, + 104 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Autonomous driving [3, 25, 29] has undergone a transformative shift from modular, manually crafted pipelines to a more integrated, end-to-end paradigm [15, 19, 24]. Unlike traditional approaches that handle tasks like detection, tracking, mapping, motion prediction, and planning in isolation, the end-to-end framework emphasizes seamless integration. By prioritizing planning, it strategically directs information from upstream perception modules, thereby enhancing robustness and reliability in dynamic driving environments.", + "bbox": [ + 89, + 114, + 480, + 265 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Achieving high-quality planning in end-to-end frameworks hinges on accurately predicting the future trajectory prediction for the ego vehicle [1, 4, 7, 15, 19, 27]. Such future prediction requires a long-horizon understanding of both static and dynamic environmental factors, including map elements and interactions with surrounding agents. For instance, UniAD [15] queries the ego context from detailed bird's-eye-view (BEV) maps at each timestamp, while VAD [19] uses an ego query to retrieve surrounding context. The retrieved information then informs the planner, which predicts a deterministic trajectory for the vehicle, as illustrated in Figure 1 (a). Nevertheless, optimal trajectory prediction is inherently stochastic due to the unpredictability of other road users' intentions, varying road conditions, and the ambiguity introduced by human driving behaviors. This stochastic nature complicates the regression target, making deterministic predictions suboptimal and even risk-prone, potentially leading to severe collisions. To mitigate these uncertainties, methods such as VADv2 [4] and SparseDrive [27] leverage probabilistic modeling to capture the continuous planning action space, producing multi-modal trajectories that consider various possible behaviors of road agents, as shown in Figure 1 (b). While effective, these multi-modal approaches are typically one-shot and solely on the current perception frame. This limitation makes them susceptible to occlusion or loss of key visual cues, which can degrade multi-modal trajectory quality. Additionally, without temporal consistency, consecutive trajectories may lack coherence, causing unstable vehicle control and introducing undesirable directional shifts and oscillations.", + "bbox": [ + 89, + 267, + 482, + 734 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "To stabilize trajectory prediction, we draw inspiration from human driving behaviors and introduce the concept of momentum into autonomous driving. In physics, momentum reflects an object's tendency to maintain its velocity based on speed and direction [16]. Analogously, in driving, momentum captures the smooth, forward progression of movement informed by past trajectories and modulated by present conditions. As illustrated in Figure 1 (c), by explicitly integrating historical trajectories with current predictions, we aim to achieve smoother and more coherent planning outcomes. To this end, we propose an end-to-end", + "bbox": [ + 89, + 736, + 482, + 902 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Momentum-Aware Driving (MomAD) framework, which incorporates momentum awareness to deliver stable and responsive planning in driving scenarios. MomAD interprets momentum on two levels: (1) trajectory momentum: By aligning candidate multi-modal trajectory with prior predictions, abrupt shifts in ego vehicle's path can be minimized, ensuring consistent control and a more comfortable driving experience. (2) perception momentum: By aggregating historical context and attending to map elements and surroundings over time, the model broadens its perspective, capturing subtle agent intentions missed in single-frame observations. To implement these ideas, we introduce (1) Topological Trajectory Matching (TTM): we first use this module to minimize planning discrepancies across time steps by employing the Hausdorff Distance to identify multimodal trajectory proposals that best align with past planning results. This approach ensures temporal coherence by preventing excessive deviation from previous trajectories. (2) Momentum Planning Interactor (MPI): Since the selected trajectory may still be biased toward the current perception and sacrifice long-horizon considerations, we cross-attends the current best planning query with historical plan queries, which implicitly convey critical long-term ego-temporal, ego-agent, and ego-map information as key and value vectors. This interaction enriches the current query with long-horizon perception momentum, improving its context awareness. To enhance robustness against environmental noise and perception errors, we incorporate a Robust Instance Denoising Module in the perception stage. By introducing controlled perturbations during training, the model learns to denoise perception inputs, achieving resilience to dynamic changes and misdetections.", + "bbox": [ + 511, + 90, + 903, + 573 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "To evaluate the planning stability of MomAD, we propose a new Trajectory Prediction Consistency (TPC) metric to measure consistency between predicted and historical trajectories. Experiments demonstrate that MomAD can maintain long-term consistency ( $\\geq 3$ seconds). Given that most scenes in nuScenes involve straight roads, which limit the assessment of temporal inconsistency, we curated a Turning-nuScenes validation set from turning scenarios within the nuScenes dataset to provide a more challenging evaluation, where our approach outperforms state-of-the-art end-to-end frameworks. For example, our MomAD reduces the collision rate by $26\\%$ and the TPC by $0.97\\mathrm{m}$ compared to SparseDrive [27] for a 6-second horizon prediction in the Turning-nuScenes validation set.", + "bbox": [ + 511, + 575, + 903, + 785 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "2. Related Work", + "text_level": 1, + "bbox": [ + 511, + 801, + 653, + 816 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "End-to-end autonomous driving, which learns directly from raw sensor data to generate planning trajectories or driving commands, eliminates manual feature extraction [3, 25]. End-to-end autonomous driving methods [5, 6, 8-13, 15, 18-21, 23, 26, 28, 30-36] have garnered increased attent", + "bbox": [ + 511, + 827, + 903, + 902 + ], + "page_idx": 1 + }, + { + "type": "page_number", + "text": "22433", + "bbox": [ + 478, + 945, + 517, + 955 + ], + "page_idx": 1 + }, + { + "type": "image", + "img_path": "images/991d8a2d24ec69326049bda428c951dab6e0a97aaf3df4dcb21ae72add0abdfa.jpg", + "image_caption": [ + "Figure 2. The overall architecture of MomAD. MomAD, as a multi-modal trajectory end-to-end autonomous driving method, first encodes multi-view images into feature maps, then learns a sparse scene representation through a robust instance denoising via perturbation module, and finally performs a momentum planning through Topological Trajectory Matching (TTM) module and Momentum Planning Interactor (MPI) module to accomplish planning tasks. Our approach addresses critical challenges of stability and robustness in dynamic driving conditions." + ], + "image_footnote": [], + "bbox": [ + 91, + 85, + 897, + 244 + ], + "page_idx": 2 + }, + { + "type": "image", + "img_path": "images/b3627d66e8e105144801867ff0063acb082dc4f467f283e9024241da1d0d4845.jpg", + "image_caption": [ + "Figure 3. The illustration of Momentum Planning Interactor (MPI). MPI cross-attends a selected planning query with historical queries to expand static and dynamic perception files, resulting in an enriched query that improves long-horizon trajectory generation and reduces collision risks." + ], + "image_footnote": [], + "bbox": [ + 91, + 343, + 480, + 460 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "tion. UniAD [15] effectively integrates information from various preceding tasks, including perception, prediction, and planning modules to assist in trajectory planning, and achieves significant performance improvements. VAD [19] models driving scenarios as fully vectorized representations and employs explicit instance-level planning constraints to enhance planning safety. However, they [5, 11, 14, 15, 19] adopt a deterministic approach to trajectory prediction, which fails to account for trajectory diversity and may introduce risks due to intermediate regression results. VADv2 [4] proposes probabilistic planning to address the limitations of deterministic trajectory prediction, enabling multimodal trajectory predictions. Building upon the multimodal trajectory planning framework, SparseDrive [27] designs planning and motion prediction modules to achieve SOTA performance and efficiency on the nuScenes dataset. However, while multi-modal trajectory planning methods [4, 27, 28] have achieved SOTA performance, they overlook the temporal inconsistency caused by maximum score offsets. Regarding temporal inconsistency, existing methods only address temporal instance characteristics up to now, and they entirely overlook the issue of temporal planning", + "bbox": [ + 89, + 571, + 482, + 904 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "consistency. In this work, we focus on this issue, aiming to address it using the concept of momentum planning.", + "bbox": [ + 511, + 347, + 906, + 378 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3. Method", + "text_level": 1, + "bbox": [ + 513, + 416, + 604, + 431 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Framework Overview. Figure 2 presents an overview of the proposed MomAD system, which integrates sparse perception and momentum-aware planning. To capture key dynamic and static instances interacting with the ego vehicle, the sparse perception module builds upon the SparseDrive [27] to encode multi-view image features, which are aggregated into instance features $\\mathbf{F}_t^{\\mathrm{ins}}$ for road agents and map elements at time step $t$ . These features are obtained by sampling keypoints around the anchor boxes and polylines, feeding into the detection/tracking and online mapping blocks for accurate predictions. The core of MomAD is the joint motion and momentum-aware planning module, which comprises two main components: (1) Topological Trajectory Matching (Sec 3.1), which explicitly selects the candidate trajectory that best matches the prior path among all multi-modal trajectories to ensure temporal coherence; and (2) Momentum Planning Interaction (Sec 3.2), which expands the perceptive field by cross-attending the selected candidate trajectory's planning query with queries from the previous time step in the long-horizon query mixer. This approach provides a broader view of the surrounding environment and the intentions of other agents. The refined query is then processed by the planning head to generate updated multi-modal trajectories. Since the planning module heavily relies on detection and map instance features, we introduce a robust instance denoising via perturbation module (Sec 3.3) within the sparse perception component during training. This ensures robustness by reducing sensitivity to noisy perception features, enhancing the stability of trajectory prediction and planning.", + "bbox": [ + 511, + 449, + 908, + 904 + ], + "page_idx": 2 + }, + { + "type": "page_number", + "text": "22434", + "bbox": [ + 478, + 944, + 519, + 957 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3.1. Topological Trajectory Matching (TTM)", + "text_level": 1, + "bbox": [ + 89, + 90, + 436, + 107 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "The proposed TTM module is inspired by the continuity of human driving behavior, where the optimal trajectories $T_{t}^{*}$ are influenced by the historical path $T_{t-1}$ to maintain temporal consistency and stability. Let $\\mathcal{T}_t = \\{T_t^k\\}_{k=1}^K$ denote a set of $K$ multi-modal candidate trajectories generated at time step $t$ , where each trajectory $T_{t}^{k} = \\{(x_{t,i}^{k},y_{t,i}^{k})\\}_{i=1}^{N_{t}}$ consists of $N_{t}$ predicted waypoints. Typically, $K$ is set to $6\\times 3$ to account for six trajectory proposals for each of three possible commands (left, right, and straight), and $N_{t}$ is chosen as 6 or 12 representing 0.5s interval for a 3- or 6-second prediction horizon.", + "bbox": [ + 89, + 112, + 483, + 277 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Trajectory Coordinate Transformation. Since the historical and current predicted trajectories are generated in the ego vehicle's coordinate system at different moments, it is essential to transform them into a common coordinate system for accurate matching. This transformation from moment $t$ to $t - 1$ is achieved as follows:", + "bbox": [ + 89, + 279, + 483, + 369 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\nT _ {t} ^ {k} \\leftarrow \\mathbf {R} _ {t - 1} ^ {- 1} \\left(T _ {t} ^ {k} - \\mathbf {\\Gamma} _ {t - 1}\\right), \\forall k \\in [ K ], \\tag {1}\n$$\n", + "text_format": "latex", + "bbox": [ + 158, + 376, + 482, + 395 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "where $\\mathbf{R}_{t - 1}$ and $\\Gamma_{t - 1}$ denote the rotation and translation matrix, respectively.", + "bbox": [ + 89, + 401, + 482, + 430 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Trajectory Distance Measurement. Simple Euclidean distance is inadequate for capturing the global alignment of trajectories, as it only measures pointwise proximity and is highly sensitive to local variations. This limitation becomes especially apparent in complex scenarios such as turns or varying point densities, where close points may not represent the alignment of the entire trajectory path. To address these limitations, TTM employs the Hausdorff distance as a more robust metric for evaluating trajectory alignment. The Hausdorff distance captures both local and global trajectory structures by measuring the maximum deviation between two sets of points, effectively quantifying the worst-case alignment between the candidate and historical trajectories. For each candidate trajectory $T_{t}^{k}$ , the Hausdorff distance to the historical trajectory $T_{t-1}$ is computed as,", + "bbox": [ + 89, + 431, + 483, + 657 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\begin{array}{l} d_{1} = \\sup_{p\\in T^{k}_{t}}\\inf_{h\\in T_{t - 1}}\\| p - h\\| , d_{2} = \\sup_{h\\in T_{t - 1}}\\inf_{p\\in T^{k}_{t}}\\| h - p\\| , \\\\ d _ {H} \\left(T _ {t} ^ {k}, T _ {t - 1}\\right) = \\max \\left(d _ {1}, d _ {2}\\right), \\tag {2} \\\\ \\end{array}\n$$\n", + "text_format": "latex", + "bbox": [ + 102, + 664, + 480, + 724 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "where $p \\in T_t^k$ and $h \\in T_{t - 1}$ represent waypoints in the candidate and historical trajectories, respectively. The Hausdorff distance considers the furthest point discrepancies between trajectories in both directions, ensuring that even minor global misalignments are captured. TTM then selects the trajectory $T_{t}^{k^{*}}$ that minimizes this distance:", + "bbox": [ + 89, + 726, + 483, + 818 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\nk ^ {*} = \\underset {k \\in [ K ]} {\\arg \\min } d _ {H} \\left(T _ {t} ^ {k}, T _ {t - 1}\\right) \\tag {3}\n$$\n", + "text_format": "latex", + "bbox": [ + 189, + 823, + 482, + 851 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "This selection enforces continuity, aligning with historical driving patterns and providing stable trajectory predictions that are less prone to sudden shifts.", + "bbox": [ + 89, + 857, + 483, + 902 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "3.2. Momentum Planning Interactor (MPI)", + "text_level": 1, + "bbox": [ + 513, + 90, + 846, + 106 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "While TTM selects the most consistent trajectory $T_{t}^{k^{*}}$ based on historical alignment, $T_{t}^{k^{*}}$ is solely based on the current perception $\\mathbf{F}_{\\mathrm{ins}}$ , which may lack a comprehensive view of the environment and be sensitive to occlusions. Therefore, the MPI module, as illustrated in Figure 3, incorporates a long-horizon query mixer to enrich the selected planning query $\\mathbf{Q}_t^{p^*}\\in \\mathbb{R}^{D_q}$ of $T_{t}^{k^{*}}$ with historical planning query $\\mathbf{Q}_{t - 1}^p\\in \\mathbb{R}^{K\\times D_q}$ and the associated planning scores $\\mathbf{S}_{t - 1}^p\\in \\mathbb{R}^K$ , implicitly capturing a broader understanding of the surrounding context and other agents' intentions over time. Here, $D_{q}$ represents the latent dimension of planning queries. This enriched planning query $\\tilde{\\mathbf{Q}}_t^{p^*}\\in \\mathbb{R}^{D_q}$ will be combined with instance features to re-generate an improved trajectory $\\tilde{\\mathcal{T}}_t\\in \\mathbb{R}^{K\\times N_t\\times 2}$ .", + "bbox": [ + 511, + 114, + 906, + 329 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Long-horizon Query Mixer. To achieve a robust perception of temporal momentum, the query mixer allows crossattention between the selected candidate trajectory's planning query with multi-modal planning queries from the previous time step. The historical planning queries $\\mathbf{Q}_{t-1}^p$ and associated scores $\\mathbf{S}_{t-1}^p$ are combined through element-wise interaction and processed with an LSTM to simulate temporal evolution:", + "bbox": [ + 511, + 332, + 905, + 452 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\mathbf {Q} _ {t - 1} ^ {p ^ {\\prime}} = \\operatorname {L S T M} \\left(\\sigma \\left(\\mathbf {S} _ {t - 1} ^ {p}\\right) \\circ \\operatorname {M L P} \\left(\\mathbf {Q} _ {t - 1} ^ {p}\\right)\\right), \\tag {4}\n$$\n", + "text_format": "latex", + "bbox": [ + 555, + 468, + 906, + 489 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "where $\\sigma(\\cdot)$ indicates the sigmoid function, MLP: $\\mathbb{R}^{D_q} \\mapsto \\mathbb{R}^{D_q}$ the linear transform and $\\circ$ the element-wise product. The LSTM processes this interaction, producing a surrogate multi-modal query $\\mathbf{Q}_{t-1}^{p'} \\in \\mathbb{R}^{K \\times D_q}$ that captures the temporal evolution of planning queries. To aggregate historical information, the current planning query $\\mathbf{Q}_t^{p*}$ is used as a query in a cross-attention module. The result $\\tilde{\\mathbf{Q}}_t^{p*}$ incorporates long-term spatiotemporal context, which is further combined with the planning instance features and the encoded ego-anchor position information to inform the subsequent trajectory predictor:", + "bbox": [ + 511, + 505, + 905, + 675 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\begin{array}{l} \\tilde {\\mathbf {Q}} _ {t} ^ {p *} = \\text {A t t e n t i o n} \\left(\\mathbf {Q} _ {t} ^ {p *}, \\mathbf {Q} _ {t - 1} ^ {p ^ {\\prime}}, \\mathbf {Q} _ {t - 1} ^ {p ^ {\\prime}}\\right). \\tag {5} \\\\ \\tilde {\\mathcal {T}} _ {t}, \\tilde {\\mathbf {S}} _ {t} = \\operatorname {P l a n H e a d} \\left(\\tilde {\\mathbf {Q}} _ {t} ^ {p *}, \\mathbf {F} _ {t} ^ {\\text {i n s}}\\right), \\\\ \\end{array}\n$$\n", + "text_format": "latex", + "bbox": [ + 578, + 691, + 903, + 732 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "The PlanHead module then generates refined $\\tilde{T}_t, \\tilde{\\mathbf{S}}_t$ . The best trajectory $T_t^*$ is then selected based on the highest scores among multi-modal outputs. Importantly, while the best trajectory is chosen based on the multi-modal trajectory scores, unlike previous selections, the current multimodal trajectories now fully consider the temporal consistency. This approach provides a stable, temporally-aware planning solution that is robust to occlusions and noise, significantly improving trajectory stability and control in complex driving environments.", + "bbox": [ + 511, + 750, + 906, + 902 + ], + "page_idx": 3 + }, + { + "type": "page_number", + "text": "22435", + "bbox": [ + 478, + 944, + 517, + 955 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "3.3. Robust Instance Denoising via Perturbation", + "text_level": 1, + "bbox": [ + 89, + 90, + 464, + 107 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Our trajectory prediction and refinement rely heavily on the instance features $\\mathbf{F}_t^{\\mathrm{ins}}$ of road agents and map elements provided by the sparse perception module. However, due to detector instability and the dynamically changing map, these instance features may be noisy, potentially introducing errors in downstream planning. To enhance the stability of planning against such noisy inputs, we introduce controlled noise perturbations during training and employ a lightweight encoder-decoder transformer block (see Figure 2) to learn effective denoising. This approach enables the model to distinguish between essential and extraneous features, reducing the impact of perception noise on trajectory predictions. During test-time inference, this denoising capability allows the trajectory predictor to be resilient to fluctuations in instance features. As a result, the model can produce smoother, more stable trajectories even in challenging scenarios with occlusions, temporary obstacles, or misdetections.", + "bbox": [ + 89, + 113, + 485, + 387 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "4. Experiments", + "text_level": 1, + "bbox": [ + 89, + 401, + 225, + 417 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "4.1. Experimental Setup", + "text_level": 1, + "bbox": [ + 89, + 425, + 282, + 443 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Datasets. We conducted extensive experiments on the widely adopted nuScenes dataset [2] to evaluate tasks including detection, online mapping and planning in an open-loop setting. The nuScenes dataset comprises 1,000 driving scenes, with 700 and 150 sequences allocated for training and validation. Each scene spans around 20 seconds and contains roughly 40 key-frames annotated at $2\\mathrm{Hz}$ , where each sample includes six images captured by surrounding cameras covering $360^{\\circ}$ FOV horizontally and point clouds collected by both LiDAR and radar sensors. Since most planning tasks in nuScenes focus predominantly on gostraight commands, we curate a challenging subset of turning scenarios to form the Turning-nuScenes dataset, aimed at verifying the temporal consistency of predicted trajectories across time steps. Planning samples for turns are selected by setting the threshold between 3s and 0.5s of 'gt_ego_fut_trajs' to 25. The turning nuScenes validation dataset constitutes only one-tenth of the full nuScenes validation set, including 17 scenes with 680 samples. We use Bench2Drive [17], a closed-loop evaluation protocol under CARLA Leaderboard 2.0 for end-to-end autonomous driving. It provides an official training set, where we use the base set (1000 clips) for fair comparison with all the other baselines. We use the official 220 routes for evaluation.", + "bbox": [ + 89, + 449, + 483, + 811 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Evaluation Metrics for Planning. For planning evaluation, we adopt the commonly used L2 Displacement Error (L2) and Collision Rate to assess planning performance. The calculation of the L2 error follows VAD [19] and the collision rate is aligned with SparseDrive [27]. However, the mainstream planning metrics cannot faithfully reveal the", + "bbox": [ + 89, + 811, + 483, + 905 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "stability of predicted trajectories. Therefore, we introduce a novel metric, Trajectory Prediction Consistency (TPC), to measure the disparity between current predicted trajectories and historical predicted trajectories, allowing for a more comprehensive assessment of the consistency of trajectories. With coordinates transformed, the TPC between the current predicted trajectory $T_{t}^{\\mathrm{Pred}}$ and historical one $T_{t-1}^{\\mathrm{Pred}}$ is defined as,", + "bbox": [ + 511, + 90, + 906, + 212 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\mathrm {T P C} = \\frac {1}{N _ {T}} \\sqrt {\\sum_ {i = 1} ^ {N} \\left(\\left(T _ {t} ^ {\\mathrm {P r e d}} - T _ {t - 1} ^ {\\mathrm {P r e d}}\\right) ^ {2} \\cdot T _ {\\mathrm {G T}} ^ {\\mathrm {M a s k}}\\right)}, \\tag {6}\n$$\n", + "text_format": "latex", + "bbox": [ + 532, + 223, + 906, + 271 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "where $N_{T}$ is the total number of GT trajectories in the validation set, and $T_{\\mathrm{GT}}^{\\mathrm{Mask}}$ is the mask for trajectories exceeding the overlapped time period of two trajectories. Our TPC metric evaluates whether autonomous vehicles adhere to predicted trajectories, ensuring continuity across frames. Notably, the TPC metric provides a statistical perspective on the dataset-wide evaluation rather than at the individual sample level.", + "bbox": [ + 511, + 281, + 908, + 404 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "4.2. Main Results", + "text_level": 1, + "bbox": [ + 511, + 412, + 653, + 426 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "To conduct a comprehensive comparison, we have undertaken an exhaustive analysis of two distinct metrics presented in Table 1, which are derived from UniAD [15] and VAD [19]. It is noteworthy that, aside from Table 1, all other tables rely on the VAD [19] evaluation metrics for their assessments.", + "bbox": [ + 511, + 435, + 908, + 527 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "4.2.1 Planning Results", + "text_level": 1, + "bbox": [ + 511, + 546, + 687, + 561 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "NuScenes. As shown in Table 1, MomAD achieves an L2 error of $0.60\\mathrm{m}$ , a collision rate of $0.09\\%$ , and a TPC of $0.54\\mathrm{m}$ , respectively. Compared to SOTAs like UniAD [15], VAD [19] and SparseDrive [27], our method shows SOTA performance in planning results. It is worth noting that we have made significant improvements in TPC, which directly proves our effectiveness in timing consistency. It is worth noting that we achieved significant improvements in TPC at $0.30\\mathrm{m}$ , $0.53\\mathrm{m}$ , and $0.78\\mathrm{m}$ at 1s, 2s, and 3s on the nuScenes dataset, directly demonstrating our effectiveness in temporal consistency. Additionally, our MomAD is straightforward and achieves an FPS of 7.8, a slightly slower than SparseDrive. In summary, our MomAD effectively utilizes the smoothing advantage of Momentum and has a significant effect on improving temporal consistency.", + "bbox": [ + 511, + 570, + 908, + 796 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Turning-nuScenes. As noted in [22], the nuScenes dataset features many straight routes, which limits the assessment of end-to-end methods. The simplicity of these paths can mask a model's true performance in complex scenarios. To address this, we evaluated our MomAD on the Turing-nuScenes validation set, as shown in Table 2. SparseDrive [27], a SOTA end-to-end method utilizing", + "bbox": [ + 511, + 797, + 910, + 905 + ], + "page_idx": 4 + }, + { + "type": "page_number", + "text": "22436", + "bbox": [ + 478, + 944, + 519, + 957 + ], + "page_idx": 4 + }, + { + "type": "table", + "img_path": "images/4ca87b80aa863416047f152abd276537fc9022697f8b8e0fbb9f47a80099a7a6.jpg", + "table_caption": [ + "Table 1. Planning results on the nuScenes validation dataset. ${}^{ \\dagger }$ denotes evaluation protocol used in UniAD [15]. * denotes results reproduced with the official checkpoint. As Ref. [22] states, we deactivate the ego status information for a fair comparison." + ], + "table_footnote": [], + "table_body": "
MethodInputBackboneL2 (m) ↓Col. Rate (%) ↓TPC (m) ↓FPS ↑
1s2s3sAvg.1s2s3sAvg.1s2s3sAvg.
UniAD† [15]CameraResNet1010.480.961.651.030.050.170.710.310.450.891.540.961.8 (A100)
VAD† [19]CameraResNet500.541.151.981.220.100.240.960.430.470.831.430.91-
SparseDrive†* [27]CameraResNet500.440.921.691.010.070.190.710.320.390.771.410.859.0 (RTX4090)
MomAD (Ours)†CameraResNet500.430.881.620.980.060.160.680.300.370.741.300.807.8 (RTX4090)
UniAD [15]CameraResNet1010.450.701.040.730.620.580.630.610.410.680.970.681.8 (A100)
VAD [19]CameraResNet500.410.701.050.720.030.190.430.210.360.660.910.64-
SparseDrive [27]CameraResNet500.290.580.960.610.010.050.180.080.300.570.850.579.0 (RTX4090)
MomAD (Ours)CameraResNet500.310.570.910.600.010.050.220.090.300.530.780.547.8 (RTX4090)
", + "bbox": [ + 93, + 128, + 903, + 277 + ], + "page_idx": 5 + }, + { + "type": "table", + "img_path": "images/d817252fe32a968cd2a976eed7be1d8e30e2dda01d1f5d8a639d9adab8a4a826.jpg", + "table_caption": [ + "Table 2. Planning results on the Turning-nuScenes validation dataset. SparseDrive [27] is a SOTA end-to-end multi-modal trajectory planning method. Red indicates improvement. We follow the VAD [19] evaluation metric." + ], + "table_footnote": [], + "table_body": "
MethodL2 (m) ↓Col. Rate (%) ↓TPC (m) ↓
1s2s3sAvg.1s2s3sAvg.1s2s3sAvg.
SparseDrive [27]0.350.771.460.860.040.170.980.400.340.701.330.79
MomAD (Ours)0.33-0.020.70-0.071.24-0.220.76-0.100.03-0.010.13-0.040.79-0.190.32-0.080.32-0.020.54-0.161.05-0.280.63-0.16
", + "bbox": [ + 94, + 321, + 903, + 392 + ], + "page_idx": 5 + }, + { + "type": "table", + "img_path": "images/033bcda7cb771bf3f8c31402a900e57a8559542b1e37cac4a3c143cd79fe4df4.jpg", + "table_caption": [ + "Table 3. Long trajectory planning results on the nuScenes and Turning-nuScenes validation sets. We train models for 10 epochs for 6s-horizon prediction. T-nuScenes indicates the challenging Turning-nuScenes. We follow the VAD [19] evaluation metric." + ], + "table_footnote": [], + "table_body": "
SplitMethodL2 (m) ↓Col. Rate (%) ↓TPC (m) ↓
4s5s6s4s5s6s4s5s6s
nuScenesSparseDrive [27]1.752.322.950.871.542.331.331.661.99
MomAD1.671.982.450.831.432.131.191.451.61
-0.09-0.34-0.50-0.04-0.11-0.20-0.14-0.21-0.38
T-nuScenesSparseDrive [27]2.072.713.360.911.712.571.542.312.90
MomAD1.802.072.510.851.572.311.371.581.93
-0.27-0.64-0.85-0.06-0.14-0.26-0.17-0.73-0.97
", + "bbox": [ + 94, + 474, + 478, + 565 + ], + "page_idx": 5 + }, + { + "type": "table", + "img_path": "images/2322b85e0cffa940dfc14d49428804c190f92d328d3189a51c98fe892bde7705.jpg", + "table_caption": [ + "Table 4. Open-loop and Closed-loop results on Bench2Drive (V0.0.3) under base training set. 'mmt' refers multi-modal trajectory variant of VAD and * the re-implementation." + ], + "table_footnote": [], + "table_body": "
MethodOpen-loop MetricClosed-loop Metric
Avg. L2 ↓DS ↑ SR (%) ↑Effi ↑Comf ↑
VAD0.9142.3515.00157.9446.01
VADmmt*0.8942.8715.91158.1247.22
MomAD (Euclidean)0.8744.2216.91161.7748.70
MomAD0.8545.3517.44162.0949.34
SparseDrive*0.8744.5416.71170.2148.63
MomAD (Euclidean)0.8446.1217.45173.3550.98
MomAD0.8247.9118.11174.9151.20
", + "bbox": [ + 94, + 622, + 478, + 739 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "multi-modal trajectories, performs well across various scenarios but struggles with driving stability during turns. In contrast, MomAD exhibits superior consistency in trajectory predictions, as indicated by the TPC metric. Overall, MomAD not only delivers effective trajectory predictions under standard conditions but also maintains reliability amid dynamic changes and complex environments.", + "bbox": [ + 88, + 766, + 482, + 872 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Long Trajectory Prediction. Accurate long trajectory prediction is vital for enhancing the stability of autonomous", + "bbox": [ + 88, + 873, + 482, + 902 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "driving, is useful to evaluate molels' ability to address temporal inconsistency issues in multi-modal trajectory planning. As shown in Table 3, we compared SparseDrive and MomAD in 4-6s long trajectory prediction on the nuScenes and Turning-nuScenes dataset, demonstrating a significant performance improvement. Specifically, in nuScenes dataset, compared with SparseDrive, MomAD experienced a decrease of $0.09\\mathrm{m}$ $(5.14\\%)$ , $0.34\\mathrm{m}$ $(14.66\\%)$ , and $0.50\\mathrm{m}$ $(16.95\\%)$ in the 4s, 5s, and 6s of L2 error, a decrease of $0.04\\%$ $0.11\\%$ and $0.20\\%$ in the 4s, 5s, and 6s of collision rate, and a decrease of $0.14\\mathrm{m}$ $(10.53\\%)$ , $0.21\\mathrm{m}$ $(12.65\\%)$ , and $0.38\\mathrm{m}$ $(19.10\\%)$ in the 4s, 5s, and 6s of TPC, respectively. Furthermore, in Turning-nuScenes dataset, compared with SparseDrive, MomAD experienced a decrease of $0.27\\mathrm{m}$ $(13.04\\%)$ , $0.64\\mathrm{m}$ $(23.62\\%)$ , and $0.85\\mathrm{m}$ $(25.30\\%)$ in the 4s, 5s, and 6s of L2 error, a decrease of $0.06\\%$ $0.14\\%$ and $0.26\\%$ in the 4s, 5s, and 6s of collision rate, and a decrease of $0.17\\mathrm{m}$ $(11.04\\%)$ , $0.73\\mathrm{m}$ $(31.60\\%)$ , and $0.97\\mathrm{m}$ $(32.45\\%)$ in the 4s, 5s, and 6s of TPC, respectively. In summary, our MomAD has improved the performance of long trajectory predictions, which further proves that MomAD can effectively alleviate the problem of temporal inconsistency.", + "bbox": [ + 511, + 417, + 906, + 763 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Bench2Drive. We have included evaluations on the challenging closed-loop results on Bench2Drive dataset, as shown in Table 4, w which covers 44 interactive scenes (e.g., cut-ins, overtaking, detours) and 220 routes across diverse weather conditions and locations. Our MomAD improves success rate by $16.3\\%$ and $8.4\\%$ over the VAD multi-modal variant and SparseDrive, and enhances the Comfortness score (trajectory smoothness) by $7.2\\%$ and $5.3\\%$ , demonstrating its effectiveness.", + "bbox": [ + 511, + 766, + 908, + 902 + ], + "page_idx": 5 + }, + { + "type": "page_number", + "text": "22437", + "bbox": [ + 478, + 944, + 519, + 957 + ], + "page_idx": 5 + }, + { + "type": "table", + "img_path": "images/a0cd90cb65f95101f3fc1cc6961a1c5aba8462fb63a73a2969357a93850998c0.jpg", + "table_caption": [ + "Table 5. Perception and motion results on the nuScenes validation dataset. ${}^{ \\dagger }$ indicates the results are reproduced with the official checkpoint. ${\\mathrm{{AP}}}_{d}$ denotes ${\\mathrm{{AP}}}_{\\text{divider }}$ . ${\\mathrm{{AP}}}_{b}$ denotes ${\\mathrm{{AP}}}_{\\text{boundary }}$ . mADE denotes minADE. mFDE denotes minFDE." + ], + "table_footnote": [], + "table_body": "
Method3D Object DetectionMulti-Object TrackingOnline MappingMotion Prediction
mAP ↑NDS ↑mATE ↓mASE ↓mAOE ↓mAVE ↓mAAE ↓AMOTA ↑AMOTP ↓Recall ↑IDS ↓mAP ↑APped ↑APd ↑APb ↑mADE ↓mFDE ↓MR ↓EPA ↑
UniAD [15]0.3800.4980.6840.2770.3830.3810.1920.3591.3200.467906----0.711.020.1510.456
VAD† [19]0.3120.4350.6100.2880.5410.5340.228----47.640.651.550.6----
SparseDrive [27]0.4180.5250.5660.2750.5520.2610.1900.3861.2540.49988655.149.957.058.40.620.990.1360.482
MomAD (Ours)0.4230.5310.5610.2690.5490.2580.1880.3911.2430.50985355.950.758.158.90.610.980.1370.499
", + "bbox": [ + 91, + 127, + 903, + 200 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/cd0c23434c044dc7720c52132f7566fec1c2ebe9ef23195e91eb55d45f9d0d48.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 94, + 218, + 495, + 349 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/19ae4902ce112b004de94d825a4617078236757c492b697d43361caea240d3b6.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 501, + 217, + 900, + 349 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/3d9435429872b91eeddacefa255e145856a828e73a19607f3603cd236a93a864.jpg", + "image_caption": [ + "Figure 4. Visualization results of MomAD compared with UniAD, VAD and SparseDrive across multiple frames. The proposed MomAD achieves temporal consistency whichever from the predicted trajectory compared with ground truth (GT) or from the TPC metric." + ], + "image_footnote": [], + "bbox": [ + 94, + 349, + 493, + 489 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/bf70935d41cd2902af5a66c001957d649c65c9d7185e6c23096be6dbb5251b91.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 500, + 351, + 900, + 489 + ], + "page_idx": 6 + }, + { + "type": "table", + "img_path": "images/6ac047b3b897c5a0d6d659989fe5e8249619a2d38793508c5bfb133bc639fdf9.jpg", + "table_caption": [ + "Table 6. Ablation studies of the sparse perception module in MomAD on the nuScenes validation split. The Encoder and Denoise $(\\sigma)$ module is denoted as ED. MP represents Momentum planning. NS is the Gaussian noise factor controlling the noise level. Noise is applied during training only. We follow the VAD [19] evaluation metric." + ], + "table_footnote": [], + "table_body": "
ED MP NSDetectionTrackingOnline MappingMotionPlanning (Avg.)
mAP ↑NDS ↑AMOTA ↑mAP ↑APped ↑mADE ↓L2 ↓Col. ↓TPC ↓
✓0.00.4070.5210.38155.049.30.630.620.140.56
0.00.4050.5200.38055.149.50.630.610.130.55
✓0.10.4200.5300.39055.850.50.580.610.120.55
✓0.00.4170.5280.38655.450.60.630.610.110.55
✓0.050.4210.5290.38855.650.80.620.610.110.54
✓0.10.4230.5310.39155.950.70.610.600.090.54
✓0.20.4180.5200.38854.449.20.630.620.180.58
✓0.30.4120.5180.38354.048.80.650.640.220.61
", + "bbox": [ + 94, + 651, + 480, + 771 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "4.2.2 Perception and Motion Prediction Results", + "text_level": 1, + "bbox": [ + 89, + 800, + 436, + 815 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Sparse representation is efficient but suffers from instability issues caused by the variability of instance features. To address these issues, we have enhanced the instance features using the Encoder and Denoise $(\\sigma)$ module (denoted as ED) within the sparse perception framework, ensuring", + "bbox": [ + 89, + 827, + 483, + 905 + ], + "page_idx": 6 + }, + { + "type": "table", + "img_path": "images/be52047a852a17da4e2b63128535f684738fa143d85818932199e74d569b75b6.jpg", + "table_caption": [ + "Table 7. Impact of history frames in MomAD on the Turning-nuScenes validation set. $t$ denotes the frame number, where 1 indicates the history is empty (represented by 0), 2 signifies that the historical result corresponds to the previous 1 frame, and 3 indicates that the historical result pertains to the previous 2 frames. We follow the VAD [19] evaluation metric." + ], + "table_footnote": [], + "table_body": "
EDMPNStL2(m) ↓Col. Rate(%) ↓TPC(m) ↓
2s3sAvg.2s3sAvg.2s3sAvg.
0.771.460.860.170.980.400.701.330.79
✓20.711.270.770.140.830.340.551.070.65
✓0.10.771.450.860.180.970.390.691.330.79
✓✓0.110.781.480.880.180.980.390.701.350.81
✓✓0.120.701.240.760.130.790.320.541.050.63
✓✓0.130.721.270.780.140.840.350.561.090.66
", + "bbox": [ + 517, + 651, + 903, + 761 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "end-to-end stability for autonomous driving. As shown in Table 5, our MomAD perception module includes 3D object detection, multi-object tracking, and online mapping tasks. For 3D object detection, MomAD achieves $42.3\\%$ mAP and $53.1\\%$ NPS, improving the mAP by $0.5\\%$ and the NDS by $0.6\\%$ compared to the baseline SparseDrive [27]. For multi-object tracking, MomAD achieves an AMOTA of", + "bbox": [ + 511, + 797, + 908, + 905 + ], + "page_idx": 6 + }, + { + "type": "page_number", + "text": "22438", + "bbox": [ + 478, + 945, + 519, + 955 + ], + "page_idx": 6 + }, + { + "type": "table", + "img_path": "images/e98d4be004fb33e5d5d2f2c5837c14e2e74d9d3a218e04cc27816a08db1494c2.jpg", + "table_caption": [ + "Table 8. Ablation studies of the impact of the different modules in MP on the Turning-nuScenes validation dataset. QM denotes Long-horizon Query Mixer, TP denotes Trajectory Predictor. Add refers to the addition operation between the historical planning query $\\mathbf{Q}_{t-1}^p$ and the selected planning query $\\mathbf{Q}_t^{p*}$ . We follow the VAD [19] evaluation metric." + ], + "table_footnote": [], + "table_body": "
QMAddTPL2 (m) ↓Col. Rate (%) ↓TPC (m) ↓
2s3sAvg.2s3sAvg.2s3sAvg.
0.771.460.860.170.980.400.701.330.79
✓✓0.761.380.820.140.880.360.621.210.67
✓✓0.701.240.760.130.790.320.541.050.63
", + "bbox": [ + 93, + 183, + 480, + 256 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "39.1%, surpassing the baseline SparseDrive by 0.5%. For online mapping, compared to 55.1% mAP for the baseline SparseDrive, our MomAD achieves 55.9% mAP, improving the mAP by 0.8%. For motion prediction, our MomAD outperforms SparseDrive [27] and UniAD [15], achieves better motion prediction performance by considering the influence of an ego vehicle on other agents. In detail, Our MomAD achieves a 0.61m minADE, 0.98 minFDE, and 13.7% MR, and 0.499 EPA, respectively.", + "bbox": [ + 89, + 281, + 480, + 417 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "4.3. Ablation Study", + "text_level": 1, + "bbox": [ + 89, + 426, + 243, + 441 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Roles of 'ED' module in Sparse Perception. Sparse representation end-to-end methods yield efficient computation but unstable metrics. Further details are available in the Appendix. As shown in Table 6, the Encoder and Denoise $(\\sigma)$ module (ED) within sparse perception enhances the instance features, significantly impacts the overall pipeline. By introducing Gaussian noise and employing techniques, the robustness of instance features is improved, particularly when training with Noisy at 0.1. Our findings suggest that a controlled level of noise can enhance the end-to-end capabilities of sparse methods during training, offering insights for the community on sparse end-to-end methods.", + "bbox": [ + 89, + 449, + 482, + 630 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Roles of 'MP' module in Planing. As Li et al. [22] have stated, most end-to-end autonomous driving methods perform poorly in turning scenarios. As shown in Table 7, to better evaluate the planning performance of end-to-end methods in turning scenarios, our MomAD is evaluated on the Turning-nuScenes validation dataset rather than only on the full nuScenes validation dataset. Specifically, under the premise of executing 'ED', at $t = 1$ , providing a 0 value to the MP module does not improve performance. We have tried to change the MLP operation of the planning to a more complex operation, but it does not enhance the results. However, when $t = 2$ , historical queries and results are used, L2 (Avg) reaches 0.76m, Col. (Avg) reaches 0.32 %, and TPC reaches 0.63m, which represents a significant improvement. In addition, when $t = 3$ , more frames are fused, the improvement has actually decreased, which may be due to the uncertainty introduced by the departure of historical features, but there is still an overall improvement.", + "bbox": [ + 89, + 630, + 482, + 902 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Roles of different sub-modules in 'MP' module. As shown in Table 8, we conducted an in-depth analysis of the internal mechanism of momentum planning. We found that simply using the native 'Add' operation to regenerate the planning results can achieve a good improvement, with L2 (Avg.), Col. (Avg.), and TPC (Avg.) decreasing by $0.04\\mathrm{m}$ , $0.04\\%$ , and $0.12\\mathrm{m}$ , respectively. However, the 'Add' operation alone does not fully utilize historical features. Our Long-horizon Query Mixer has achieved the current optimal performance. Overall, historical results are very important for current outcomes, and their reasonable utilization can maximize the performance of end-to-end planning.", + "bbox": [ + 511, + 90, + 906, + 272 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "4.4. Visualization", + "text_level": 1, + "bbox": [ + 511, + 287, + 650, + 301 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "As shown in Figure 4, we showcase multi-frame qualitative comparisons of end-to-end solutions, including UniAD [15], VAD [19], SparseDrive [27], and the proposed MomAD. In a representative turning scenario, the MomAD approach demonstrates superior long-term awareness of surrounding vehicles, reducing the likelihood of collisions. Additionally, it generates smoother ego-vehicle trajectories (shown in yellow and blue) that closely align with the ground-truth trajectory (in red). This highlights its strong temporal consistency and lower TPC scores. Additional visualizations for various driving commands are provided in the Appendix.", + "bbox": [ + 511, + 311, + 906, + 494 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "5. Conclusion and Future Work", + "text_level": 1, + "bbox": [ + 511, + 512, + 782, + 527 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "The proposed MomAD framework addresses key challenges in planning stability and robustness for end-to-end autonomous driving. By leveraging trajectory momentum and perception momentum, MomAD stabilizes trajectory predictions through Topological Trajectory Matching (TTM) and Momentum Planning Interactor (MPI), ensuring temporal coherence and enriching long-horizon context. Evaluations on nuScenes, the curated Turning-nuScenes validation set, and Bench2Drive demonstrate its superior performance in reducing collision rates and improving trajectory consistency compared to SOTA methods. While MomAD improves temporal consistency in long-horizon trajectory prediction, a gap remains due to mode collapse induced by the standard teacher-forcing approach to trajectory regression, limiting trajectory diversity. Future work will explore diffusion models and speculative decoding to enhance trajectory diversity while balancing efficiency.", + "bbox": [ + 511, + 539, + 908, + 797 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Acknowledgements", + "text_level": 1, + "bbox": [ + 511, + 815, + 681, + 832 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "We appreciate the valuable discussions with Wenchao Sun, Bo Jiang, and Bencheng Liao from Horizon Robotics. This work was supported by the National Key R&D Program of China (2018AAA0100302).", + "bbox": [ + 511, + 842, + 906, + 902 + ], + "page_idx": 7 + }, + { + "type": "page_number", + "text": "22439", + "bbox": [ + 478, + 945, + 517, + 955 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "References", + "text_level": 1, + "bbox": [ + 91, + 89, + 187, + 104 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[1] Szilard Aradi. Survey of deep reinforcement learning for motion planning of autonomous vehicles. IEEE Transactions on Intelligent Transportation Systems, 23(2):740-759, 2020. 2", + "[2] Holger Caesar, Varun Bankiti, Alex H Lang, Sourabh Vora, Venice Erin Liong, Qiang Xu, Anush Krishnan, Yu Pan, Giancarlo Baldan, and Oscar Beijbom. nuscenes: A multimodal dataset for autonomous driving. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 11621-11631, 2020. 5", + "[3] Li Chen, Penghao Wu, Kashyap Chitta, Bernhard Jaeger, Andreas Geiger, and Hongyang Li. End-to-end autonomous driving: Challenges and frontiers. IEEE Transactions on Pattern Analysis and Machine Intelligence, 2024. 2", + "[4] Shaoyu Chen, Bo Jiang, Hao Gao, Bencheng Liao, Qing Xu, Qian Zhang, Chang Huang, Wenyu Liu, and Xinggang Wang. Vadv2: End-to-end vectorized autonomous driving via probabilistic planning. arXiv preprint arXiv:2402.13243, 2024. 1, 2, 3", + "[5] Zhili Chen, Maosheng Ye, Shuangjie Xu, Tongyi Cao, and Qifeng Chen. Ppad: Iterative interactions of prediction and planning for end-to-end autonomous driving. In European Conference on Computer Vision, pages 239-256. Springer, 2025. 2, 3", + "[6] Jie Cheng, Yingbing Chen, Xiaodong Mei, Bowen Yang, Bo Li, and Ming Liu. Rethinking imitation-based planners for autonomous driving. In 2024 IEEE International Conference on Robotics and Automation (ICRA), pages 14123-14130. IEEE, 2024. 2", + "[7] Laurene Claussmann, Marc Revilloud, Dominique Gruyer, and Sébastien Glaser. A review of motion planning for highway autonomous driving. IEEE Transactions on Intelligent Transportation Systems, 21(5):1826-1848, 2020. 2", + "[8] Felipe Codevilla, Matthias Müller, Antonio López, Vladlen Koltun, and Alexey Dosovitskiy. End-to-end driving via conditional imitation learning. In 2018 IEEE international conference on robotics and automation (ICRA), pages 4693-4700. IEEE, 2018. 2", + "[9] Felipe Codevilla, Matthias Müller, Antonio López, Vladlen Koltun, and Alexey Dosovitskiy. End-to-end driving via conditional imitation learning. In 2018 IEEE International Conference on Robotics and Automation (ICRA), pages 4693-4700, 2018.", + "[10] Felipe Codevilla, Eder Santana, Antonio M. Lopez, and Adrien Gaidon. Exploring the limitations of behavior cloning for autonomous driving. In Proceedings of the IEEE/CVF International Conference on Computer Vision (ICCV), 2019.", + "[11] Simon Doll, Niklas Hanselmann, Lukas Schneider, Richard Schulz, Marius Cordts, Markus Enzweiler, and Hendrik Lensch. Dualad: Disentangling the dynamic and static world for end-to-end driving. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 14728-14737, 2024. 3", + "[12] Mingzhe Guo, Zhipeng Zhang, Yuan He, Ke Wang, and Liping Jing. End-to-end autonomous driving without costly" + ], + "bbox": [ + 93, + 114, + 483, + 902 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "modularization and 3d manual annotation. arXiv preprint arXiv:2406.17680, 2024.", + "[13] Peiyun Hu, Aaron Huang, John Dolan, David Held, and Deva Ramanan. Safe local motion planning with self-supervised freespace forecasting. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 12732-12741, 2021. 2", + "[14] Shengchao Hu, Li Chen, Penghao Wu, Hongyang Li, Junchi Yan, and Dacheng Tao. St-p3: End-to-end vision-based autonomous driving via spatial-temporal feature learning. In European Conference on Computer Vision, pages 533-549. Springer, 2022. 1, 3", + "[15] Yihan Hu, Jiazhi Yang, Li Chen, Keyu Li, Chonghao Sima, Xizhou Zhu, Siqi Chai, Senyao Du, Tianwei Lin, Wenhai Wang, Lewei Lu, Xiaosong Jia, Qiang Liu, Jifeng Dai, Yu Qiao, and Hongyang Li. Planning-oriented autonomous driving. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 17853-17862, 2023. 1, 2, 3, 5, 6, 7, 8", + "[16] Xiangdong Ji, Yizhuang Liu, Yu-Sheng Liu, Jian-Hui Zhang, and Yong Zhao. Large-momentum effective theory. Reviews of Modern Physics, 93(3):035005, 2021. 2", + "[17] Xiaosong Jia, Zhenjie Yang, Qifeng Li, Zhiyuan Zhang, and Junchi Yan. Bench2drive: Towards multi-ability benchmarking of closed-loop end-to-end autonomous driving. Advances in Neural Information Processing Systems, 37:819–844, 2025. 5", + "[18] Bo Jiang, Shaoyu Chen, Xinggang Wang, Bencheng Liao, Tianheng Cheng, Jiajie Chen, Helong Zhou, Qian Zhang, Wenyu Liu, and Chang Huang. Perceive, interact, predict: Learning dynamic and static clues for end-to-end motion prediction. arXiv preprint arXiv:2212.02181, 2022. 2", + "[19] Bo Jiang, Shaoyu Chen, Qing Xu, Benchcheng Liao, Jiajie Chen, Helong Zhou, Qian Zhang, Wenyu Liu, Chang Huang, and Xinggang Wang. Vad: Vectorized scene representation for efficient autonomous driving. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 8340-8350, 2023. 1, 2, 3, 5, 6, 7, 8", + "[20] Bo Jiang, Shaoyu Chen, Benchcheng Liao, Xingyu Zhang, Wei Yin, Qian Zhang, Chang Huang, Wenyu Liu, and Xinggang Wang. Senna: Bridging large vision-language models and end-to-end autonomous driving, 2024.", + "[21] Zhenxin Li, Kailin Li, Shihao Wang, Shiyi Lan, Zhiding Yu, Yishen Ji, Zhiqi Li, Ziye Zhu, Jan Kautz, Zuxuan Wu, et al. Hydra-mdp: End-to-end multimodal planning with multitarget hydra-distillation. arXiv preprint arXiv:2406.06978, 2024. 2", + "[22] Zhiqi Li, Zhiding Yu, Shiyi Lan, Jiahan Li, Jan Kautz, Tong Lu, and Jose M Alvarez. Is ego status all you need for open-loop end-to-end autonomous driving? In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 14864-14873, 2024. 5, 6, 8", + "[23] Jiageng Mao, Yuxi Qian, Junjie Ye, Hang Zhao, and Yue Wang. Gpt-driver: Learning to drive with gpt. arXiv preprint arXiv:2310.01415, 2023. 2", + "[24] Aditya Prakash, Kashyap Chitta, and Andreas Geiger. Multimodal fusion transformer for end-to-end autonomous driv" + ], + "bbox": [ + 516, + 93, + 906, + 902 + ], + "page_idx": 8 + }, + { + "type": "page_number", + "text": "22440", + "bbox": [ + 478, + 944, + 519, + 955 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "ing. In 2021 IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), 2021. 2", + "[25] Ziying Song, Lin Liu, Feiyang Jia, Yadan Luo, Caiyan Jia, Guoxin Zhang, Lei Yang, and Li Wang. Robustness-aware 3d object detection in autonomous driving: A review and outlook. IEEE Transactions on Intelligent Transportation Systems, pages 1-30, 2024. 2", + "[26] Haisheng Su, Wei Wu, and Junchi Yan. Difsd: Ego-centric fully sparse paradigm with uncertainty denoising and iterative refinement for efficient end-to-end autonomous driving. arXiv preprint arXiv:2409.09777, 2024. 2", + "[27] Wenchao Sun, Xuewu Lin, Yining Shi, Chuang Zhang, Hao ran Wu, and Sifa Zheng. Sparsedrive: End-to-end autonomous driving via sparse scene representation. arXiv preprint arXiv:2405.19620, 2024. 1, 2, 3, 5, 6, 7, 8", + "[28] Junming Wang, Xingyu Zhang, Zebin Xing, Songen Gu, Xiaoyang Guo, Yang Hu, Ziying Song, Qian Zhang, Xiaoxiao Long, and Wei Yin. He-drive: Human-like end-to-end driving with vision language models, 2024. 1, 2, 3", + "[29] Li Wang, Xinyu Zhang, Ziying Song, Jiangfeng Bi, Guoxin Zhang, Haiyue Wei, Liyao Tang, Lei Yang, Jun Li, Caiyan Jia, et al. Multi-modal 3d object detection in autonomous driving: A survey and taxonomy. IEEE Transactions on Intelligent Vehicles, 2023. 2", + "[30] Penghao Wu, Xiaosong Jia, Li Chen, Junchi Yan, Hongyang Li, and Yu Qiao. Trajectory-guided control prediction for end-to-end autonomous driving: A simple yet strong baseline. In Advances in Neural Information Processing Systems, pages 6119-6132. Curran Associates, Inc., 2022. 2", + "[31] Zeyu Yang, Nan Song, Wei Li, Xiatian Zhu, Li Zhang, and Philip HS Torr. Deepinteraction++: Multi-modality interaction for autonomous driving. arXiv preprint arXiv:2408.05075, 2024.", + "[32] Tengju Ye, Wei Jing, Chunyong Hu, Shikun Huang, Lingping Gao, Fangzhen Li, Jingke Wang, Ke Guo, Wencong Xiao, Weibo Mao, Hang Zheng, Kun Li, Junbo Chen, and Kaicheng Yu. Fusionad: Multi-modality fusion for prediction and planning tasks of autonomous driving, 2023. 1", + "[33] Wenyuan Zeng, Wenjie Luo, Simon Suo, Abbas Sadat, Bin Yang, Sergio Casas, and Raquel Urtasun. End-to-end interpretable neural motion planner. In 2019 IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), 2019.", + "[34] Yunpeng Zhang, Deheng Qian, Ding Li, Yifeng Pan, Yong Chen, Zhenbao Liang, Zhiyao Zhang, Shurui Zhang, Hongxu Li, Maolei Fu, et al. Graphad: Interaction scene graph for end-to-end autonomous driving. arXiv preprint arXiv:2403.19098, 2024.", + "[35] Zhejun Zhang, Alexander Liniger, Dengxin Dai, Fisher Yu, and Luc Van Gool. End-to-end urban driving by imitating a reinforcement learning coach. In Proceedings of the IEEE/CVF International Conference on Computer Vision (ICCV), pages 15222-15232, 2021.", + "[36] Wenzhao Zheng, Ruiqi Song, Xianda Guo, and Long Chen. Genad: Generative end-to-end autonomous driving. arXiv preprint arXiv:2402.11502, 2024. 2" + ], + "bbox": [ + 91, + 90, + 483, + 882 + ], + "page_idx": 9 + }, + { + "type": "page_number", + "text": "22441", + "bbox": [ + 478, + 945, + 517, + 955 + ], + "page_idx": 9 + } +] \ No newline at end of file diff --git a/2025/Don't Shake the Wheel_ Momentum-Aware Planning in End-to-End Autonomous Driving/85994b24-3b2b-443a-a1bf-eabbda00b06b_model.json b/2025/Don't Shake the Wheel_ Momentum-Aware Planning in End-to-End Autonomous Driving/85994b24-3b2b-443a-a1bf-eabbda00b06b_model.json new file mode 100644 index 0000000000000000000000000000000000000000..b970cf91d70b34343e7028328dae6d583b3249b5 --- /dev/null +++ b/2025/Don't Shake the Wheel_ Momentum-Aware Planning in End-to-End Autonomous Driving/85994b24-3b2b-443a-a1bf-eabbda00b06b_model.json @@ -0,0 +1,1771 @@ +[ + [ + { + "type": "header", + "bbox": [ + 0.107, + 0.003, + 0.182, + 0.044 + ], + "angle": 0, + "content": "CVF" + }, + { + "type": "header", + "bbox": [ + 0.237, + 0.001, + 0.812, + 0.047 + ], + "angle": 0, + "content": "This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore." + }, + { + "type": "title", + "bbox": [ + 0.219, + 0.131, + 0.78, + 0.177 + ], + "angle": 0, + "content": "Don't Shake the Wheel: Momentum-Aware Planning in End-to-End Autonomous Driving" + }, + { + "type": "text", + "bbox": [ + 0.186, + 0.203, + 0.812, + 0.24 + ], + "angle": 0, + "content": "Ziying Song\\(^{1,2,\\dagger}\\), Caiyan Jia\\(^{1,2,\\star}\\), Lin Liu\\(^{1,2}\\), Hongyu Pan\\(^{3}\\), Yongchang Zhang\\(^{3}\\), Junming Wang\\(^{3,7}\\), Xingyu Zhang\\(^{3}\\), Shaoqing Xu\\(^{4}\\), Lei Yang\\(^{5}\\), Yadan Luo\\(^{6,\\star}\\)" + }, + { + "type": "text", + "bbox": [ + 0.202, + 0.245, + 0.796, + 0.264 + ], + "angle": 0, + "content": "\\(^{1}\\)School of Computer Science and Technology, Beijing Jiaotong University" + }, + { + "type": "text", + "bbox": [ + 0.201, + 0.264, + 0.798, + 0.282 + ], + "angle": 0, + "content": "\\(^{2}\\)Beijing Key Laboratory of Traffic Data Mining and Embodied Intelligence" + }, + { + "type": "text", + "bbox": [ + 0.159, + 0.282, + 0.838, + 0.3 + ], + "angle": 0, + "content": "\\(^{3}\\)Horizon Robotics \\(^{4}\\)University of Macau \\(^{5}\\)THU \\(^{6}\\)The University of Queensland \\(^{7}\\)HKU" + }, + { + "type": "text", + "bbox": [ + 0.273, + 0.302, + 0.704, + 0.317 + ], + "angle": 0, + "content": "{songziying, cyjia}@bjtu.edu.cn, y.luo@uq.edu.au." + }, + { + "type": "title", + "bbox": [ + 0.248, + 0.346, + 0.327, + 0.363 + ], + "angle": 0, + "content": "Abstract" + }, + { + "type": "text", + "bbox": [ + 0.089, + 0.39, + 0.485, + 0.844 + ], + "angle": 0, + "content": "End-to-end autonomous driving frameworks enable seamless integration of perception and planning but often rely on one-shot trajectory prediction, which may lead to unstable control and vulnerability to occlusions in single-frame perception. To address this, we propose the Momentum-Aware Driving (MomAD) framework, which introduces trajectory momentum and perception momentum to stabilize and refine trajectory predictions. MomAD comprises two core components: (1) Topological Trajectory Matching (TTM) employs Hausdorff Distance to select the optimal planning query that aligns with prior paths to ensure coherence; (2) Momentum Planning Interactor (MPI) cross-attends the selected planning query with historical queries to expand static and dynamic perception files. This enriched query, in turn, helps regenerate long-horizon trajectory and reduce collision risks. To mitigate noise arising from dynamic environments and detection errors, we introduce robust instance denoising during training, enabling the planning model to focus on critical signals and improve its robustness. We also propose a novel Trajectory Prediction Consistency (TPC) metric to quantitatively assess planning stability. Experiments on the nuScenes dataset demonstrate that MomAD achieves superior long-term consistency \\((\\geq 3s)\\) compared to SOTA methods. Moreover, evaluations on the curated Turning-nuScenes shows that MomAD reduces the collision rate by \\(26\\%\\) and improves TPC by \\(0.97m\\) \\((33.45\\%)\\) over a 6s prediction horizon, while closed-loop on Bench2Drive demonstrates an up to \\(16.3\\%\\) improvement in success rate. The source code is available at https://github.com/adept-thu/MomAD." + }, + { + "type": "image", + "bbox": [ + 0.518, + 0.346, + 0.637, + 0.517 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.518, + 0.519, + 0.668, + 0.532 + ], + "angle": 0, + "content": "(a) Deterministic Planning" + }, + { + "type": "image", + "bbox": [ + 0.64, + 0.346, + 0.759, + 0.517 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.704, + 0.519, + 0.856, + 0.533 + ], + "angle": 0, + "content": "(b) Multi-modal Trajectory" + }, + { + "type": "image", + "bbox": [ + 0.759, + 0.346, + 0.901, + 0.517 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.52, + 0.542, + 0.639, + 0.704 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.641, + 0.542, + 0.779, + 0.704 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.785, + 0.542, + 0.904, + 0.704 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.624, + 0.715, + 0.805, + 0.728 + ], + "angle": 0, + "content": "(c) Momentum Planning (Ours)" + }, + { + "type": "image_caption", + "bbox": [ + 0.512, + 0.734, + 0.907, + 0.846 + ], + "angle": 0, + "content": "Figure 1. (a) Deterministic Planning [14, 15, 19, 32] predicts deterministic trajectories, but lacks action diversity, posing safety risks. (b) Multi-modal Trajectory Planning [4, 27, 28] selects the highest-scoring trajectory among the multi-modal trajectories, yet fails to ensure stability and consistency, having risks in vehicle trembling. (c) Momentum Planning leverages the trajectory and perception momentum to enhance current planning through historical guidance to overcome temporal inconsistency." + }, + { + "type": "page_footnote", + "bbox": [ + 0.115, + 0.889, + 0.405, + 0.904 + ], + "angle": 0, + "content": "† Intern of Horizon Robotics, † Corresponding author." + }, + { + "type": "page_number", + "bbox": [ + 0.479, + 0.945, + 0.52, + 0.958 + ], + "angle": 0, + "content": "22432" + } + ], + [ + { + "type": "title", + "bbox": [ + 0.093, + 0.091, + 0.223, + 0.106 + ], + "angle": 0, + "content": "1. Introduction" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.116, + 0.482, + 0.266 + ], + "angle": 0, + "content": "Autonomous driving [3, 25, 29] has undergone a transformative shift from modular, manually crafted pipelines to a more integrated, end-to-end paradigm [15, 19, 24]. Unlike traditional approaches that handle tasks like detection, tracking, mapping, motion prediction, and planning in isolation, the end-to-end framework emphasizes seamless integration. By prioritizing planning, it strategically directs information from upstream perception modules, thereby enhancing robustness and reliability in dynamic driving environments." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.268, + 0.483, + 0.736 + ], + "angle": 0, + "content": "Achieving high-quality planning in end-to-end frameworks hinges on accurately predicting the future trajectory prediction for the ego vehicle [1, 4, 7, 15, 19, 27]. Such future prediction requires a long-horizon understanding of both static and dynamic environmental factors, including map elements and interactions with surrounding agents. For instance, UniAD [15] queries the ego context from detailed bird's-eye-view (BEV) maps at each timestamp, while VAD [19] uses an ego query to retrieve surrounding context. The retrieved information then informs the planner, which predicts a deterministic trajectory for the vehicle, as illustrated in Figure 1 (a). Nevertheless, optimal trajectory prediction is inherently stochastic due to the unpredictability of other road users' intentions, varying road conditions, and the ambiguity introduced by human driving behaviors. This stochastic nature complicates the regression target, making deterministic predictions suboptimal and even risk-prone, potentially leading to severe collisions. To mitigate these uncertainties, methods such as VADv2 [4] and SparseDrive [27] leverage probabilistic modeling to capture the continuous planning action space, producing multi-modal trajectories that consider various possible behaviors of road agents, as shown in Figure 1 (b). While effective, these multi-modal approaches are typically one-shot and solely on the current perception frame. This limitation makes them susceptible to occlusion or loss of key visual cues, which can degrade multi-modal trajectory quality. Additionally, without temporal consistency, consecutive trajectories may lack coherence, causing unstable vehicle control and introducing undesirable directional shifts and oscillations." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.737, + 0.483, + 0.904 + ], + "angle": 0, + "content": "To stabilize trajectory prediction, we draw inspiration from human driving behaviors and introduce the concept of momentum into autonomous driving. In physics, momentum reflects an object's tendency to maintain its velocity based on speed and direction [16]. Analogously, in driving, momentum captures the smooth, forward progression of movement informed by past trajectories and modulated by present conditions. As illustrated in Figure 1 (c), by explicitly integrating historical trajectories with current predictions, we aim to achieve smoother and more coherent planning outcomes. To this end, we propose an end-to-end" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.092, + 0.905, + 0.574 + ], + "angle": 0, + "content": "Momentum-Aware Driving (MomAD) framework, which incorporates momentum awareness to deliver stable and responsive planning in driving scenarios. MomAD interprets momentum on two levels: (1) trajectory momentum: By aligning candidate multi-modal trajectory with prior predictions, abrupt shifts in ego vehicle's path can be minimized, ensuring consistent control and a more comfortable driving experience. (2) perception momentum: By aggregating historical context and attending to map elements and surroundings over time, the model broadens its perspective, capturing subtle agent intentions missed in single-frame observations. To implement these ideas, we introduce (1) Topological Trajectory Matching (TTM): we first use this module to minimize planning discrepancies across time steps by employing the Hausdorff Distance to identify multimodal trajectory proposals that best align with past planning results. This approach ensures temporal coherence by preventing excessive deviation from previous trajectories. (2) Momentum Planning Interactor (MPI): Since the selected trajectory may still be biased toward the current perception and sacrifice long-horizon considerations, we cross-attends the current best planning query with historical plan queries, which implicitly convey critical long-term ego-temporal, ego-agent, and ego-map information as key and value vectors. This interaction enriches the current query with long-horizon perception momentum, improving its context awareness. To enhance robustness against environmental noise and perception errors, we incorporate a Robust Instance Denoising Module in the perception stage. By introducing controlled perturbations during training, the model learns to denoise perception inputs, achieving resilience to dynamic changes and misdetections." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.576, + 0.905, + 0.786 + ], + "angle": 0, + "content": "To evaluate the planning stability of MomAD, we propose a new Trajectory Prediction Consistency (TPC) metric to measure consistency between predicted and historical trajectories. Experiments demonstrate that MomAD can maintain long-term consistency (\\(\\geq 3\\) seconds). Given that most scenes in nuScenes involve straight roads, which limit the assessment of temporal inconsistency, we curated a Turning-nuScenes validation set from turning scenarios within the nuScenes dataset to provide a more challenging evaluation, where our approach outperforms state-of-the-art end-to-end frameworks. For example, our MomAD reduces the collision rate by \\(26\\%\\) and the TPC by \\(0.97\\mathrm{m}\\) compared to SparseDrive [27] for a 6-second horizon prediction in the Turning-nuScenes validation set." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.802, + 0.655, + 0.817 + ], + "angle": 0, + "content": "2. Related Work" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.828, + 0.905, + 0.904 + ], + "angle": 0, + "content": "End-to-end autonomous driving, which learns directly from raw sensor data to generate planning trajectories or driving commands, eliminates manual feature extraction [3, 25]. End-to-end autonomous driving methods [5, 6, 8-13, 15, 18-21, 23, 26, 28, 30-36] have garnered increased attent" + }, + { + "type": "page_number", + "bbox": [ + 0.48, + 0.946, + 0.518, + 0.957 + ], + "angle": 0, + "content": "22433" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.092, + 0.087, + 0.898, + 0.245 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.089, + 0.253, + 0.908, + 0.325 + ], + "angle": 0, + "content": "Figure 2. The overall architecture of MomAD. MomAD, as a multi-modal trajectory end-to-end autonomous driving method, first encodes multi-view images into feature maps, then learns a sparse scene representation through a robust instance denoising via perturbation module, and finally performs a momentum planning through Topological Trajectory Matching (TTM) module and Momentum Planning Interactor (MPI) module to accomplish planning tasks. Our approach addresses critical challenges of stability and robustness in dynamic driving conditions." + }, + { + "type": "image", + "bbox": [ + 0.093, + 0.344, + 0.482, + 0.462 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.09, + 0.472, + 0.483, + 0.542 + ], + "angle": 0, + "content": "Figure 3. The illustration of Momentum Planning Interactor (MPI). MPI cross-attends a selected planning query with historical queries to expand static and dynamic perception files, resulting in an enriched query that improves long-horizon trajectory generation and reduces collision risks." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.572, + 0.483, + 0.905 + ], + "angle": 0, + "content": "tion. UniAD [15] effectively integrates information from various preceding tasks, including perception, prediction, and planning modules to assist in trajectory planning, and achieves significant performance improvements. VAD [19] models driving scenarios as fully vectorized representations and employs explicit instance-level planning constraints to enhance planning safety. However, they [5, 11, 14, 15, 19] adopt a deterministic approach to trajectory prediction, which fails to account for trajectory diversity and may introduce risks due to intermediate regression results. VADv2 [4] proposes probabilistic planning to address the limitations of deterministic trajectory prediction, enabling multimodal trajectory predictions. Building upon the multimodal trajectory planning framework, SparseDrive [27] designs planning and motion prediction modules to achieve SOTA performance and efficiency on the nuScenes dataset. However, while multi-modal trajectory planning methods [4, 27, 28] have achieved SOTA performance, they overlook the temporal inconsistency caused by maximum score offsets. Regarding temporal inconsistency, existing methods only address temporal instance characteristics up to now, and they entirely overlook the issue of temporal planning" + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.348, + 0.907, + 0.38 + ], + "angle": 0, + "content": "consistency. In this work, we focus on this issue, aiming to address it using the concept of momentum planning." + }, + { + "type": "title", + "bbox": [ + 0.514, + 0.417, + 0.605, + 0.433 + ], + "angle": 0, + "content": "3. Method" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.45, + 0.909, + 0.905 + ], + "angle": 0, + "content": "Framework Overview. Figure 2 presents an overview of the proposed MomAD system, which integrates sparse perception and momentum-aware planning. To capture key dynamic and static instances interacting with the ego vehicle, the sparse perception module builds upon the SparseDrive [27] to encode multi-view image features, which are aggregated into instance features \\(\\mathbf{F}_t^{\\mathrm{ins}}\\) for road agents and map elements at time step \\(t\\). These features are obtained by sampling keypoints around the anchor boxes and polylines, feeding into the detection/tracking and online mapping blocks for accurate predictions. The core of MomAD is the joint motion and momentum-aware planning module, which comprises two main components: (1) Topological Trajectory Matching (Sec 3.1), which explicitly selects the candidate trajectory that best matches the prior path among all multi-modal trajectories to ensure temporal coherence; and (2) Momentum Planning Interaction (Sec 3.2), which expands the perceptive field by cross-attending the selected candidate trajectory's planning query with queries from the previous time step in the long-horizon query mixer. This approach provides a broader view of the surrounding environment and the intentions of other agents. The refined query is then processed by the planning head to generate updated multi-modal trajectories. Since the planning module heavily relies on detection and map instance features, we introduce a robust instance denoising via perturbation module (Sec 3.3) within the sparse perception component during training. This ensures robustness by reducing sensitivity to noisy perception features, enhancing the stability of trajectory prediction and planning." + }, + { + "type": "page_number", + "bbox": [ + 0.479, + 0.945, + 0.521, + 0.958 + ], + "angle": 0, + "content": "22434" + } + ], + [ + { + "type": "title", + "bbox": [ + 0.091, + 0.091, + 0.437, + 0.108 + ], + "angle": 0, + "content": "3.1. Topological Trajectory Matching (TTM)" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.113, + 0.484, + 0.279 + ], + "angle": 0, + "content": "The proposed TTM module is inspired by the continuity of human driving behavior, where the optimal trajectories \\(T_{t}^{*}\\) are influenced by the historical path \\(T_{t-1}\\) to maintain temporal consistency and stability. Let \\(\\mathcal{T}_t = \\{T_t^k\\}_{k=1}^K\\) denote a set of \\(K\\) multi-modal candidate trajectories generated at time step \\(t\\), where each trajectory \\(T_{t}^{k} = \\{(x_{t,i}^{k},y_{t,i}^{k})\\}_{i=1}^{N_{t}}\\) consists of \\(N_{t}\\) predicted waypoints. Typically, \\(K\\) is set to \\(6\\times 3\\) to account for six trajectory proposals for each of three possible commands (left, right, and straight), and \\(N_{t}\\) is chosen as 6 or 12 representing 0.5s interval for a 3- or 6-second prediction horizon." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.28, + 0.484, + 0.37 + ], + "angle": 0, + "content": "Trajectory Coordinate Transformation. Since the historical and current predicted trajectories are generated in the ego vehicle's coordinate system at different moments, it is essential to transform them into a common coordinate system for accurate matching. This transformation from moment \\( t \\) to \\( t - 1 \\) is achieved as follows:" + }, + { + "type": "equation", + "bbox": [ + 0.16, + 0.377, + 0.483, + 0.396 + ], + "angle": 0, + "content": "\\[\nT _ {t} ^ {k} \\leftarrow \\mathbf {R} _ {t - 1} ^ {- 1} \\left(T _ {t} ^ {k} - \\mathbf {\\Gamma} _ {t - 1}\\right), \\forall k \\in [ K ], \\tag {1}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.402, + 0.483, + 0.431 + ], + "angle": 0, + "content": "where \\(\\mathbf{R}_{t - 1}\\) and \\(\\Gamma_{t - 1}\\) denote the rotation and translation matrix, respectively." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.432, + 0.484, + 0.659 + ], + "angle": 0, + "content": "Trajectory Distance Measurement. Simple Euclidean distance is inadequate for capturing the global alignment of trajectories, as it only measures pointwise proximity and is highly sensitive to local variations. This limitation becomes especially apparent in complex scenarios such as turns or varying point densities, where close points may not represent the alignment of the entire trajectory path. To address these limitations, TTM employs the Hausdorff distance as a more robust metric for evaluating trajectory alignment. The Hausdorff distance captures both local and global trajectory structures by measuring the maximum deviation between two sets of points, effectively quantifying the worst-case alignment between the candidate and historical trajectories. For each candidate trajectory \\(T_{t}^{k}\\), the Hausdorff distance to the historical trajectory \\(T_{t-1}\\) is computed as," + }, + { + "type": "equation", + "bbox": [ + 0.104, + 0.665, + 0.482, + 0.726 + ], + "angle": 0, + "content": "\\[\n\\begin{array}{l} d_{1} = \\sup_{p\\in T^{k}_{t}}\\inf_{h\\in T_{t - 1}}\\| p - h\\| , d_{2} = \\sup_{h\\in T_{t - 1}}\\inf_{p\\in T^{k}_{t}}\\| h - p\\| , \\\\ d _ {H} \\left(T _ {t} ^ {k}, T _ {t - 1}\\right) = \\max \\left(d _ {1}, d _ {2}\\right), \\tag {2} \\\\ \\end{array}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.727, + 0.484, + 0.819 + ], + "angle": 0, + "content": "where \\( p \\in T_t^k \\) and \\( h \\in T_{t - 1} \\) represent waypoints in the candidate and historical trajectories, respectively. The Hausdorff distance considers the furthest point discrepancies between trajectories in both directions, ensuring that even minor global misalignments are captured. TTM then selects the trajectory \\( T_{t}^{k^{*}} \\) that minimizes this distance:" + }, + { + "type": "equation", + "bbox": [ + 0.19, + 0.824, + 0.483, + 0.852 + ], + "angle": 0, + "content": "\\[\nk ^ {*} = \\underset {k \\in [ K ]} {\\arg \\min } d _ {H} \\left(T _ {t} ^ {k}, T _ {t - 1}\\right) \\tag {3}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.858, + 0.484, + 0.904 + ], + "angle": 0, + "content": "This selection enforces continuity, aligning with historical driving patterns and providing stable trajectory predictions that are less prone to sudden shifts." + }, + { + "type": "title", + "bbox": [ + 0.514, + 0.091, + 0.848, + 0.107 + ], + "angle": 0, + "content": "3.2. Momentum Planning Interactor (MPI)" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.115, + 0.907, + 0.33 + ], + "angle": 0, + "content": "While TTM selects the most consistent trajectory \\(T_{t}^{k^{*}}\\) based on historical alignment, \\(T_{t}^{k^{*}}\\) is solely based on the current perception \\(\\mathbf{F}_{\\mathrm{ins}}\\), which may lack a comprehensive view of the environment and be sensitive to occlusions. Therefore, the MPI module, as illustrated in Figure 3, incorporates a long-horizon query mixer to enrich the selected planning query \\(\\mathbf{Q}_t^{p^*}\\in \\mathbb{R}^{D_q}\\) of \\(T_{t}^{k^{*}}\\) with historical planning query \\(\\mathbf{Q}_{t - 1}^p\\in \\mathbb{R}^{K\\times D_q}\\) and the associated planning scores \\(\\mathbf{S}_{t - 1}^p\\in \\mathbb{R}^K\\), implicitly capturing a broader understanding of the surrounding context and other agents' intentions over time. Here, \\(D_{q}\\) represents the latent dimension of planning queries. This enriched planning query \\(\\tilde{\\mathbf{Q}}_t^{p^*}\\in \\mathbb{R}^{D_q}\\) will be combined with instance features to re-generate an improved trajectory \\(\\tilde{\\mathcal{T}}_t\\in \\mathbb{R}^{K\\times N_t\\times 2}\\)." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.333, + 0.906, + 0.453 + ], + "angle": 0, + "content": "Long-horizon Query Mixer. To achieve a robust perception of temporal momentum, the query mixer allows crossattention between the selected candidate trajectory's planning query with multi-modal planning queries from the previous time step. The historical planning queries \\(\\mathbf{Q}_{t-1}^p\\) and associated scores \\(\\mathbf{S}_{t-1}^p\\) are combined through element-wise interaction and processed with an LSTM to simulate temporal evolution:" + }, + { + "type": "equation", + "bbox": [ + 0.556, + 0.469, + 0.907, + 0.49 + ], + "angle": 0, + "content": "\\[\n\\mathbf {Q} _ {t - 1} ^ {p ^ {\\prime}} = \\operatorname {L S T M} \\left(\\sigma \\left(\\mathbf {S} _ {t - 1} ^ {p}\\right) \\circ \\operatorname {M L P} \\left(\\mathbf {Q} _ {t - 1} ^ {p}\\right)\\right), \\tag {4}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.506, + 0.906, + 0.676 + ], + "angle": 0, + "content": "where \\(\\sigma(\\cdot)\\) indicates the sigmoid function, MLP: \\(\\mathbb{R}^{D_q} \\mapsto \\mathbb{R}^{D_q}\\) the linear transform and \\(\\circ\\) the element-wise product. The LSTM processes this interaction, producing a surrogate multi-modal query \\(\\mathbf{Q}_{t-1}^{p'} \\in \\mathbb{R}^{K \\times D_q}\\) that captures the temporal evolution of planning queries. To aggregate historical information, the current planning query \\(\\mathbf{Q}_t^{p*}\\) is used as a query in a cross-attention module. The result \\(\\tilde{\\mathbf{Q}}_t^{p*}\\) incorporates long-term spatiotemporal context, which is further combined with the planning instance features and the encoded ego-anchor position information to inform the subsequent trajectory predictor:" + }, + { + "type": "equation", + "bbox": [ + 0.579, + 0.692, + 0.905, + 0.733 + ], + "angle": 0, + "content": "\\[\n\\begin{array}{l} \\tilde {\\mathbf {Q}} _ {t} ^ {p *} = \\text {A t t e n t i o n} \\left(\\mathbf {Q} _ {t} ^ {p *}, \\mathbf {Q} _ {t - 1} ^ {p ^ {\\prime}}, \\mathbf {Q} _ {t - 1} ^ {p ^ {\\prime}}\\right). \\tag {5} \\\\ \\tilde {\\mathcal {T}} _ {t}, \\tilde {\\mathbf {S}} _ {t} = \\operatorname {P l a n H e a d} \\left(\\tilde {\\mathbf {Q}} _ {t} ^ {p *}, \\mathbf {F} _ {t} ^ {\\text {i n s}}\\right), \\\\ \\end{array}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.751, + 0.907, + 0.904 + ], + "angle": 0, + "content": "The PlanHead module then generates refined \\(\\tilde{T}_t, \\tilde{\\mathbf{S}}_t\\). The best trajectory \\(T_t^*\\) is then selected based on the highest scores among multi-modal outputs. Importantly, while the best trajectory is chosen based on the multi-modal trajectory scores, unlike previous selections, the current multimodal trajectories now fully consider the temporal consistency. This approach provides a stable, temporally-aware planning solution that is robust to occlusions and noise, significantly improving trajectory stability and control in complex driving environments." + }, + { + "type": "page_number", + "bbox": [ + 0.479, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "22435" + } + ], + [ + { + "type": "title", + "bbox": [ + 0.091, + 0.091, + 0.465, + 0.108 + ], + "angle": 0, + "content": "3.3. Robust Instance Denoising via Perturbation" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.114, + 0.486, + 0.388 + ], + "angle": 0, + "content": "Our trajectory prediction and refinement rely heavily on the instance features \\(\\mathbf{F}_t^{\\mathrm{ins}}\\) of road agents and map elements provided by the sparse perception module. However, due to detector instability and the dynamically changing map, these instance features may be noisy, potentially introducing errors in downstream planning. To enhance the stability of planning against such noisy inputs, we introduce controlled noise perturbations during training and employ a lightweight encoder-decoder transformer block (see Figure 2) to learn effective denoising. This approach enables the model to distinguish between essential and extraneous features, reducing the impact of perception noise on trajectory predictions. During test-time inference, this denoising capability allows the trajectory predictor to be resilient to fluctuations in instance features. As a result, the model can produce smoother, more stable trajectories even in challenging scenarios with occlusions, temporary obstacles, or misdetections." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.402, + 0.226, + 0.418 + ], + "angle": 0, + "content": "4. Experiments" + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.426, + 0.283, + 0.444 + ], + "angle": 0, + "content": "4.1. Experimental Setup" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.45, + 0.485, + 0.812 + ], + "angle": 0, + "content": "Datasets. We conducted extensive experiments on the widely adopted nuScenes dataset [2] to evaluate tasks including detection, online mapping and planning in an open-loop setting. The nuScenes dataset comprises 1,000 driving scenes, with 700 and 150 sequences allocated for training and validation. Each scene spans around 20 seconds and contains roughly 40 key-frames annotated at \\(2\\mathrm{Hz}\\), where each sample includes six images captured by surrounding cameras covering \\(360^{\\circ}\\) FOV horizontally and point clouds collected by both LiDAR and radar sensors. Since most planning tasks in nuScenes focus predominantly on gostraight commands, we curate a challenging subset of turning scenarios to form the Turning-nuScenes dataset, aimed at verifying the temporal consistency of predicted trajectories across time steps. Planning samples for turns are selected by setting the threshold between 3s and 0.5s of 'gt_ego_fut_trajs' to 25. The turning nuScenes validation dataset constitutes only one-tenth of the full nuScenes validation set, including 17 scenes with 680 samples. We use Bench2Drive [17], a closed-loop evaluation protocol under CARLA Leaderboard 2.0 for end-to-end autonomous driving. It provides an official training set, where we use the base set (1000 clips) for fair comparison with all the other baselines. We use the official 220 routes for evaluation." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.813, + 0.485, + 0.906 + ], + "angle": 0, + "content": "Evaluation Metrics for Planning. For planning evaluation, we adopt the commonly used L2 Displacement Error (L2) and Collision Rate to assess planning performance. The calculation of the L2 error follows VAD [19] and the collision rate is aligned with SparseDrive [27]. However, the mainstream planning metrics cannot faithfully reveal the" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.092, + 0.907, + 0.213 + ], + "angle": 0, + "content": "stability of predicted trajectories. Therefore, we introduce a novel metric, Trajectory Prediction Consistency (TPC), to measure the disparity between current predicted trajectories and historical predicted trajectories, allowing for a more comprehensive assessment of the consistency of trajectories. With coordinates transformed, the TPC between the current predicted trajectory \\(T_{t}^{\\mathrm{Pred}}\\) and historical one \\(T_{t-1}^{\\mathrm{Pred}}\\) is defined as," + }, + { + "type": "equation", + "bbox": [ + 0.533, + 0.224, + 0.907, + 0.272 + ], + "angle": 0, + "content": "\\[\n\\mathrm {T P C} = \\frac {1}{N _ {T}} \\sqrt {\\sum_ {i = 1} ^ {N} \\left(\\left(T _ {t} ^ {\\mathrm {P r e d}} - T _ {t - 1} ^ {\\mathrm {P r e d}}\\right) ^ {2} \\cdot T _ {\\mathrm {G T}} ^ {\\mathrm {M a s k}}\\right)}, \\tag {6}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.282, + 0.909, + 0.405 + ], + "angle": 0, + "content": "where \\( N_{T} \\) is the total number of GT trajectories in the validation set, and \\( T_{\\mathrm{GT}}^{\\mathrm{Mask}} \\) is the mask for trajectories exceeding the overlapped time period of two trajectories. Our TPC metric evaluates whether autonomous vehicles adhere to predicted trajectories, ensuring continuity across frames. Notably, the TPC metric provides a statistical perspective on the dataset-wide evaluation rather than at the individual sample level." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.413, + 0.654, + 0.428 + ], + "angle": 0, + "content": "4.2. Main Results" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.436, + 0.909, + 0.529 + ], + "angle": 0, + "content": "To conduct a comprehensive comparison, we have undertaken an exhaustive analysis of two distinct metrics presented in Table 1, which are derived from UniAD [15] and VAD [19]. It is noteworthy that, aside from Table 1, all other tables rely on the VAD [19] evaluation metrics for their assessments." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.547, + 0.688, + 0.563 + ], + "angle": 0, + "content": "4.2.1 Planning Results" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.571, + 0.909, + 0.797 + ], + "angle": 0, + "content": "NuScenes. As shown in Table 1, MomAD achieves an L2 error of \\(0.60\\mathrm{m}\\), a collision rate of \\(0.09\\%\\), and a TPC of \\(0.54\\mathrm{m}\\), respectively. Compared to SOTAs like UniAD [15], VAD [19] and SparseDrive [27], our method shows SOTA performance in planning results. It is worth noting that we have made significant improvements in TPC, which directly proves our effectiveness in timing consistency. It is worth noting that we achieved significant improvements in TPC at \\(0.30\\mathrm{m}\\), \\(0.53\\mathrm{m}\\), and \\(0.78\\mathrm{m}\\) at 1s, 2s, and 3s on the nuScenes dataset, directly demonstrating our effectiveness in temporal consistency. Additionally, our MomAD is straightforward and achieves an FPS of 7.8, a slightly slower than SparseDrive. In summary, our MomAD effectively utilizes the smoothing advantage of Momentum and has a significant effect on improving temporal consistency." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.798, + 0.911, + 0.906 + ], + "angle": 0, + "content": "Turning-nuScenes. As noted in [22], the nuScenes dataset features many straight routes, which limits the assessment of end-to-end methods. The simplicity of these paths can mask a model's true performance in complex scenarios. To address this, we evaluated our MomAD on the Turing-nuScenes validation set, as shown in Table 2. SparseDrive [27], a SOTA end-to-end method utilizing" + }, + { + "type": "page_number", + "bbox": [ + 0.479, + 0.945, + 0.521, + 0.958 + ], + "angle": 0, + "content": "22436" + } + ], + [ + { + "type": "table_caption", + "bbox": [ + 0.09, + 0.09, + 0.907, + 0.12 + ], + "angle": 0, + "content": "Table 1. Planning results on the nuScenes validation dataset. \\( {}^{ \\dagger } \\) denotes evaluation protocol used in UniAD [15]. * denotes results reproduced with the official checkpoint. As Ref. [22] states, we deactivate the ego status information for a fair comparison." + }, + { + "type": "table", + "bbox": [ + 0.094, + 0.129, + 0.905, + 0.279 + ], + "angle": 0, + "content": "
MethodInputBackboneL2 (m) ↓Col. Rate (%) ↓TPC (m) ↓FPS ↑
1s2s3sAvg.1s2s3sAvg.1s2s3sAvg.
UniAD† [15]CameraResNet1010.480.961.651.030.050.170.710.310.450.891.540.961.8 (A100)
VAD† [19]CameraResNet500.541.151.981.220.100.240.960.430.470.831.430.91-
SparseDrive†* [27]CameraResNet500.440.921.691.010.070.190.710.320.390.771.410.859.0 (RTX4090)
MomAD (Ours)†CameraResNet500.430.881.620.980.060.160.680.300.370.741.300.807.8 (RTX4090)
UniAD [15]CameraResNet1010.450.701.040.730.620.580.630.610.410.680.970.681.8 (A100)
VAD [19]CameraResNet500.410.701.050.720.030.190.430.210.360.660.910.64-
SparseDrive [27]CameraResNet500.290.580.960.610.010.050.180.080.300.570.850.579.0 (RTX4090)
MomAD (Ours)CameraResNet500.310.570.910.600.010.050.220.090.300.530.780.547.8 (RTX4090)
" + }, + { + "type": "table_caption", + "bbox": [ + 0.09, + 0.293, + 0.907, + 0.321 + ], + "angle": 0, + "content": "Table 2. Planning results on the Turning-nuScenes validation dataset. SparseDrive [27] is a SOTA end-to-end multi-modal trajectory planning method. Red indicates improvement. We follow the VAD [19] evaluation metric." + }, + { + "type": "table", + "bbox": [ + 0.095, + 0.323, + 0.905, + 0.393 + ], + "angle": 0, + "content": "
MethodL2 (m) ↓Col. Rate (%) ↓TPC (m) ↓
1s2s3sAvg.1s2s3sAvg.1s2s3sAvg.
SparseDrive [27]0.350.771.460.860.040.170.980.400.340.701.330.79
MomAD (Ours)0.33-0.020.70-0.071.24-0.220.76-0.100.03-0.010.13-0.040.79-0.190.32-0.080.32-0.020.54-0.161.05-0.280.63-0.16
" + }, + { + "type": "table_caption", + "bbox": [ + 0.09, + 0.417, + 0.482, + 0.472 + ], + "angle": 0, + "content": "Table 3. Long trajectory planning results on the nuScenes and Turning-nuScenes validation sets. We train models for 10 epochs for 6s-horizon prediction. T-nuScenes indicates the challenging Turning-nuScenes. We follow the VAD [19] evaluation metric." + }, + { + "type": "table", + "bbox": [ + 0.095, + 0.476, + 0.48, + 0.566 + ], + "angle": 0, + "content": "
SplitMethodL2 (m) ↓Col. Rate (%) ↓TPC (m) ↓
4s5s6s4s5s6s4s5s6s
nuScenesSparseDrive [27]1.752.322.950.871.542.331.331.661.99
MomAD1.671.982.450.831.432.131.191.451.61
-0.09-0.34-0.50-0.04-0.11-0.20-0.14-0.21-0.38
T-nuScenesSparseDrive [27]2.072.713.360.911.712.571.542.312.90
MomAD1.802.072.510.851.572.311.371.581.93
-0.27-0.64-0.85-0.06-0.14-0.26-0.17-0.73-0.97
" + }, + { + "type": "table_caption", + "bbox": [ + 0.09, + 0.581, + 0.483, + 0.622 + ], + "angle": 0, + "content": "Table 4. Open-loop and Closed-loop results on Bench2Drive (V0.0.3) under base training set. 'mmt' refers multi-modal trajectory variant of VAD and * the re-implementation." + }, + { + "type": "table", + "bbox": [ + 0.095, + 0.623, + 0.48, + 0.741 + ], + "angle": 0, + "content": "
MethodOpen-loop MetricClosed-loop Metric
Avg. L2 ↓DS ↑ SR (%) ↑Effi ↑Comf ↑
VAD0.9142.3515.00157.9446.01
VADmmt*0.8942.8715.91158.1247.22
MomAD (Euclidean)0.8744.2216.91161.7748.70
MomAD0.8545.3517.44162.0949.34
SparseDrive*0.8744.5416.71170.2148.63
MomAD (Euclidean)0.8446.1217.45173.3550.98
MomAD0.8247.9118.11174.9151.20
" + }, + { + "type": "text", + "bbox": [ + 0.089, + 0.767, + 0.483, + 0.873 + ], + "angle": 0, + "content": "multi-modal trajectories, performs well across various scenarios but struggles with driving stability during turns. In contrast, MomAD exhibits superior consistency in trajectory predictions, as indicated by the TPC metric. Overall, MomAD not only delivers effective trajectory predictions under standard conditions but also maintains reliability amid dynamic changes and complex environments." + }, + { + "type": "text", + "bbox": [ + 0.089, + 0.874, + 0.483, + 0.903 + ], + "angle": 0, + "content": "Long Trajectory Prediction. Accurate long trajectory prediction is vital for enhancing the stability of autonomous" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.419, + 0.908, + 0.765 + ], + "angle": 0, + "content": "driving, is useful to evaluate molels' ability to address temporal inconsistency issues in multi-modal trajectory planning. As shown in Table 3, we compared SparseDrive and MomAD in 4-6s long trajectory prediction on the nuScenes and Turning-nuScenes dataset, demonstrating a significant performance improvement. Specifically, in nuScenes dataset, compared with SparseDrive, MomAD experienced a decrease of \\(0.09\\mathrm{m}\\) \\((5.14\\%)\\), \\(0.34\\mathrm{m}\\) \\((14.66\\%)\\), and \\(0.50\\mathrm{m}\\) \\((16.95\\%)\\) in the 4s, 5s, and 6s of L2 error, a decrease of \\(0.04\\%\\) \\(0.11\\%\\) and \\(0.20\\%\\) in the 4s, 5s, and 6s of collision rate, and a decrease of \\(0.14\\mathrm{m}\\) \\((10.53\\%)\\), \\(0.21\\mathrm{m}\\) \\((12.65\\%)\\), and \\(0.38\\mathrm{m}\\) \\((19.10\\%)\\) in the 4s, 5s, and 6s of TPC, respectively. Furthermore, in Turning-nuScenes dataset, compared with SparseDrive, MomAD experienced a decrease of \\(0.27\\mathrm{m}\\) \\((13.04\\%)\\), \\(0.64\\mathrm{m}\\) \\((23.62\\%)\\), and \\(0.85\\mathrm{m}\\) \\((25.30\\%)\\) in the 4s, 5s, and 6s of L2 error, a decrease of \\(0.06\\%\\) \\(0.14\\%\\) and \\(0.26\\%\\) in the 4s, 5s, and 6s of collision rate, and a decrease of \\(0.17\\mathrm{m}\\) \\((11.04\\%)\\), \\(0.73\\mathrm{m}\\) \\((31.60\\%)\\), and \\(0.97\\mathrm{m}\\) \\((32.45\\%)\\) in the 4s, 5s, and 6s of TPC, respectively. In summary, our MomAD has improved the performance of long trajectory predictions, which further proves that MomAD can effectively alleviate the problem of temporal inconsistency." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.767, + 0.909, + 0.904 + ], + "angle": 0, + "content": "Bench2Drive. We have included evaluations on the challenging closed-loop results on Bench2Drive dataset, as shown in Table 4, w which covers 44 interactive scenes (e.g., cut-ins, overtaking, detours) and 220 routes across diverse weather conditions and locations. Our MomAD improves success rate by \\(16.3\\%\\) and \\(8.4\\%\\) over the VAD multi-modal variant and SparseDrive, and enhances the Comfortness score (trajectory smoothness) by \\(7.2\\%\\) and \\(5.3\\%\\), demonstrating its effectiveness." + }, + { + "type": "page_number", + "bbox": [ + 0.479, + 0.945, + 0.521, + 0.958 + ], + "angle": 0, + "content": "22437" + } + ], + [ + { + "type": "table_caption", + "bbox": [ + 0.09, + 0.09, + 0.908, + 0.12 + ], + "angle": 0, + "content": "Table 5. Perception and motion results on the nuScenes validation dataset. \\( {}^{ \\dagger } \\) indicates the results are reproduced with the official checkpoint. \\( {\\mathrm{{AP}}}_{d} \\) denotes \\( {\\mathrm{{AP}}}_{\\text{divider }} \\) . \\( {\\mathrm{{AP}}}_{b} \\) denotes \\( {\\mathrm{{AP}}}_{\\text{boundary }} \\) . mADE denotes minADE. mFDE denotes minFDE." + }, + { + "type": "table", + "bbox": [ + 0.093, + 0.128, + 0.905, + 0.202 + ], + "angle": 0, + "content": "
Method3D Object DetectionMulti-Object TrackingOnline MappingMotion Prediction
mAP ↑NDS ↑mATE ↓mASE ↓mAOE ↓mAVE ↓mAAE ↓AMOTA ↑AMOTP ↓Recall ↑IDS ↓mAP ↑APped ↑APd ↑APb ↑mADE ↓mFDE ↓MR ↓EPA ↑
UniAD [15]0.3800.4980.6840.2770.3830.3810.1920.3591.3200.467906----0.711.020.1510.456
VAD† [19]0.3120.4350.6100.2880.5410.5340.228----47.640.651.550.6----
SparseDrive [27]0.4180.5250.5660.2750.5520.2610.1900.3861.2540.49988655.149.957.058.40.620.990.1360.482
MomAD (Ours)0.4230.5310.5610.2690.5490.2580.1880.3911.2430.50985355.950.758.158.90.610.980.1370.499
" + }, + { + "type": "image", + "bbox": [ + 0.095, + 0.219, + 0.496, + 0.351 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.502, + 0.218, + 0.901, + 0.351 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.095, + 0.351, + 0.495, + 0.491 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.501, + 0.352, + 0.901, + 0.491 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.09, + 0.504, + 0.908, + 0.534 + ], + "angle": 0, + "content": "Figure 4. Visualization results of MomAD compared with UniAD, VAD and SparseDrive across multiple frames. The proposed MomAD achieves temporal consistency whichever from the predicted trajectory compared with ground truth (GT) or from the TPC metric." + }, + { + "type": "table_caption", + "bbox": [ + 0.09, + 0.557, + 0.483, + 0.64 + ], + "angle": 0, + "content": "Table 6. Ablation studies of the sparse perception module in MomAD on the nuScenes validation split. The Encoder and Denoise \\((\\sigma)\\) module is denoted as ED. MP represents Momentum planning. NS is the Gaussian noise factor controlling the noise level. Noise is applied during training only. We follow the VAD [19] evaluation metric." + }, + { + "type": "table", + "bbox": [ + 0.095, + 0.652, + 0.481, + 0.772 + ], + "angle": 0, + "content": "
ED MP NSDetectionTrackingOnline MappingMotionPlanning (Avg.)
mAP ↑NDS ↑AMOTA ↑mAP ↑APped ↑mADE ↓L2 ↓Col. ↓TPC ↓
✓0.00.4070.5210.38155.049.30.630.620.140.56
0.00.4050.5200.38055.149.50.630.610.130.55
✓0.10.4200.5300.39055.850.50.580.610.120.55
✓0.00.4170.5280.38655.450.60.630.610.110.55
✓0.050.4210.5290.38855.650.80.620.610.110.54
✓0.10.4230.5310.39155.950.70.610.600.090.54
✓0.20.4180.5200.38854.449.20.630.620.180.58
✓0.30.4120.5180.38354.048.80.650.640.220.61
" + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.801, + 0.437, + 0.816 + ], + "angle": 0, + "content": "4.2.2 Perception and Motion Prediction Results" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.828, + 0.484, + 0.906 + ], + "angle": 0, + "content": "Sparse representation is efficient but suffers from instability issues caused by the variability of instance features. To address these issues, we have enhanced the instance features using the Encoder and Denoise \\((\\sigma)\\) module (denoted as ED) within the sparse perception framework, ensuring" + }, + { + "type": "table_caption", + "bbox": [ + 0.512, + 0.557, + 0.907, + 0.64 + ], + "angle": 0, + "content": "Table 7. Impact of history frames in MomAD on the Turning-nuScenes validation set. \\( t \\) denotes the frame number, where 1 indicates the history is empty (represented by 0), 2 signifies that the historical result corresponds to the previous 1 frame, and 3 indicates that the historical result pertains to the previous 2 frames. We follow the VAD [19] evaluation metric." + }, + { + "type": "table", + "bbox": [ + 0.518, + 0.652, + 0.905, + 0.762 + ], + "angle": 0, + "content": "
EDMPNStL2(m) ↓Col. Rate(%) ↓TPC(m) ↓
2s3sAvg.2s3sAvg.2s3sAvg.
0.771.460.860.170.980.400.701.330.79
✓20.711.270.770.140.830.340.551.070.65
✓0.10.771.450.860.180.970.390.691.330.79
✓✓0.110.781.480.880.180.980.390.701.350.81
✓✓0.120.701.240.760.130.790.320.541.050.63
✓✓0.130.721.270.780.140.840.350.561.090.66
" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.798, + 0.909, + 0.906 + ], + "angle": 0, + "content": "end-to-end stability for autonomous driving. As shown in Table 5, our MomAD perception module includes 3D object detection, multi-object tracking, and online mapping tasks. For 3D object detection, MomAD achieves \\(42.3\\%\\) mAP and \\(53.1\\%\\) NPS, improving the mAP by \\(0.5\\%\\) and the NDS by \\(0.6\\%\\) compared to the baseline SparseDrive [27]. For multi-object tracking, MomAD achieves an AMOTA of" + }, + { + "type": "page_number", + "bbox": [ + 0.479, + 0.946, + 0.521, + 0.957 + ], + "angle": 0, + "content": "22438" + } + ], + [ + { + "type": "table_caption", + "bbox": [ + 0.09, + 0.089, + 0.482, + 0.173 + ], + "angle": 0, + "content": "Table 8. Ablation studies of the impact of the different modules in MP on the Turning-nuScenes validation dataset. QM denotes Long-horizon Query Mixer, TP denotes Trajectory Predictor. Add refers to the addition operation between the historical planning query \\(\\mathbf{Q}_{t-1}^p\\) and the selected planning query \\(\\mathbf{Q}_t^{p*}\\). We follow the VAD [19] evaluation metric." + }, + { + "type": "table", + "bbox": [ + 0.094, + 0.184, + 0.482, + 0.257 + ], + "angle": 0, + "content": "
QMAddTPL2 (m) ↓Col. Rate (%) ↓TPC (m) ↓
2s3sAvg.2s3sAvg.2s3sAvg.
0.771.460.860.170.980.400.701.330.79
✓✓0.761.380.820.140.880.360.621.210.67
✓✓0.701.240.760.130.790.320.541.050.63
" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.282, + 0.482, + 0.419 + ], + "angle": 0, + "content": "39.1%, surpassing the baseline SparseDrive by 0.5%. For online mapping, compared to 55.1% mAP for the baseline SparseDrive, our MomAD achieves 55.9% mAP, improving the mAP by 0.8%. For motion prediction, our MomAD outperforms SparseDrive [27] and UniAD [15], achieves better motion prediction performance by considering the influence of an ego vehicle on other agents. In detail, Our MomAD achieves a 0.61m minADE, 0.98 minFDE, and 13.7% MR, and 0.499 EPA, respectively." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.427, + 0.245, + 0.443 + ], + "angle": 0, + "content": "4.3. Ablation Study" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.45, + 0.483, + 0.631 + ], + "angle": 0, + "content": "Roles of 'ED' module in Sparse Perception. Sparse representation end-to-end methods yield efficient computation but unstable metrics. Further details are available in the Appendix. As shown in Table 6, the Encoder and Denoise \\((\\sigma)\\) module (ED) within sparse perception enhances the instance features, significantly impacts the overall pipeline. By introducing Gaussian noise and employing techniques, the robustness of instance features is improved, particularly when training with Noisy at 0.1. Our findings suggest that a controlled level of noise can enhance the end-to-end capabilities of sparse methods during training, offering insights for the community on sparse end-to-end methods." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.631, + 0.483, + 0.904 + ], + "angle": 0, + "content": "Roles of 'MP' module in Planing. As Li et al. [22] have stated, most end-to-end autonomous driving methods perform poorly in turning scenarios. As shown in Table 7, to better evaluate the planning performance of end-to-end methods in turning scenarios, our MomAD is evaluated on the Turning-nuScenes validation dataset rather than only on the full nuScenes validation dataset. Specifically, under the premise of executing 'ED', at \\( t = 1 \\), providing a 0 value to the MP module does not improve performance. We have tried to change the MLP operation of the planning to a more complex operation, but it does not enhance the results. However, when \\( t = 2 \\), historical queries and results are used, L2 (Avg) reaches 0.76m, Col. (Avg) reaches 0.32 %, and TPC reaches 0.63m, which represents a significant improvement. In addition, when \\( t = 3 \\), more frames are fused, the improvement has actually decreased, which may be due to the uncertainty introduced by the departure of historical features, but there is still an overall improvement." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.092, + 0.907, + 0.273 + ], + "angle": 0, + "content": "Roles of different sub-modules in 'MP' module. As shown in Table 8, we conducted an in-depth analysis of the internal mechanism of momentum planning. We found that simply using the native 'Add' operation to regenerate the planning results can achieve a good improvement, with L2 (Avg.), Col. (Avg.), and TPC (Avg.) decreasing by \\(0.04\\mathrm{m}\\), \\(0.04\\%\\), and \\(0.12\\mathrm{m}\\), respectively. However, the 'Add' operation alone does not fully utilize historical features. Our Long-horizon Query Mixer has achieved the current optimal performance. Overall, historical results are very important for current outcomes, and their reasonable utilization can maximize the performance of end-to-end planning." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.288, + 0.651, + 0.303 + ], + "angle": 0, + "content": "4.4. Visualization" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.313, + 0.907, + 0.495 + ], + "angle": 0, + "content": "As shown in Figure 4, we showcase multi-frame qualitative comparisons of end-to-end solutions, including UniAD [15], VAD [19], SparseDrive [27], and the proposed MomAD. In a representative turning scenario, the MomAD approach demonstrates superior long-term awareness of surrounding vehicles, reducing the likelihood of collisions. Additionally, it generates smoother ego-vehicle trajectories (shown in yellow and blue) that closely align with the ground-truth trajectory (in red). This highlights its strong temporal consistency and lower TPC scores. Additional visualizations for various driving commands are provided in the Appendix." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.513, + 0.783, + 0.529 + ], + "angle": 0, + "content": "5. Conclusion and Future Work" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.54, + 0.909, + 0.798 + ], + "angle": 0, + "content": "The proposed MomAD framework addresses key challenges in planning stability and robustness for end-to-end autonomous driving. By leveraging trajectory momentum and perception momentum, MomAD stabilizes trajectory predictions through Topological Trajectory Matching (TTM) and Momentum Planning Interactor (MPI), ensuring temporal coherence and enriching long-horizon context. Evaluations on nuScenes, the curated Turning-nuScenes validation set, and Bench2Drive demonstrate its superior performance in reducing collision rates and improving trajectory consistency compared to SOTA methods. While MomAD improves temporal consistency in long-horizon trajectory prediction, a gap remains due to mode collapse induced by the standard teacher-forcing approach to trajectory regression, limiting trajectory diversity. Future work will explore diffusion models and speculative decoding to enhance trajectory diversity while balancing efficiency." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.816, + 0.682, + 0.833 + ], + "angle": 0, + "content": "Acknowledgements" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.843, + 0.907, + 0.903 + ], + "angle": 0, + "content": "We appreciate the valuable discussions with Wenchao Sun, Bo Jiang, and Bencheng Liao from Horizon Robotics. This work was supported by the National Key R&D Program of China (2018AAA0100302)." + }, + { + "type": "page_number", + "bbox": [ + 0.48, + 0.946, + 0.519, + 0.957 + ], + "angle": 0, + "content": "22439" + } + ], + [ + { + "type": "title", + "bbox": [ + 0.093, + 0.09, + 0.188, + 0.106 + ], + "angle": 0, + "content": "References" + }, + { + "type": "ref_text", + "bbox": [ + 0.101, + 0.116, + 0.484, + 0.169 + ], + "angle": 0, + "content": "[1] Szilard Aradi. Survey of deep reinforcement learning for motion planning of autonomous vehicles. IEEE Transactions on Intelligent Transportation Systems, 23(2):740-759, 2020. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.101, + 0.173, + 0.484, + 0.255 + ], + "angle": 0, + "content": "[2] Holger Caesar, Varun Bankiti, Alex H Lang, Sourabh Vora, Venice Erin Liong, Qiang Xu, Anush Krishnan, Yu Pan, Giancarlo Baldan, and Oscar Beijbom. nuscenes: A multimodal dataset for autonomous driving. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 11621-11631, 2020. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.257, + 0.483, + 0.311 + ], + "angle": 0, + "content": "[3] Li Chen, Penghao Wu, Kashyap Chitta, Bernhard Jaeger, Andreas Geiger, and Hongyang Li. End-to-end autonomous driving: Challenges and frontiers. IEEE Transactions on Pattern Analysis and Machine Intelligence, 2024. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.313, + 0.483, + 0.38 + ], + "angle": 0, + "content": "[4] Shaoyu Chen, Bo Jiang, Hao Gao, Bencheng Liao, Qing Xu, Qian Zhang, Chang Huang, Wenyu Liu, and Xinggang Wang. Vadv2: End-to-end vectorized autonomous driving via probabilistic planning. arXiv preprint arXiv:2402.13243, 2024. 1, 2, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.383, + 0.483, + 0.45 + ], + "angle": 0, + "content": "[5] Zhili Chen, Maosheng Ye, Shuangjie Xu, Tongyi Cao, and Qifeng Chen. Ppad: Iterative interactions of prediction and planning for end-to-end autonomous driving. In European Conference on Computer Vision, pages 239-256. Springer, 2025. 2, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.454, + 0.483, + 0.521 + ], + "angle": 0, + "content": "[6] Jie Cheng, Yingbing Chen, Xiaodong Mei, Bowen Yang, Bo Li, and Ming Liu. Rethinking imitation-based planners for autonomous driving. In 2024 IEEE International Conference on Robotics and Automation (ICRA), pages 14123-14130. IEEE, 2024. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.525, + 0.483, + 0.579 + ], + "angle": 0, + "content": "[7] Laurene Claussmann, Marc Revilloud, Dominique Gruyer, and Sébastien Glaser. A review of motion planning for highway autonomous driving. IEEE Transactions on Intelligent Transportation Systems, 21(5):1826-1848, 2020. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.581, + 0.483, + 0.648 + ], + "angle": 0, + "content": "[8] Felipe Codevilla, Matthias Müller, Antonio López, Vladlen Koltun, and Alexey Dosovitskiy. End-to-end driving via conditional imitation learning. In 2018 IEEE international conference on robotics and automation (ICRA), pages 4693-4700. IEEE, 2018. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.651, + 0.483, + 0.718 + ], + "angle": 0, + "content": "[9] Felipe Codevilla, Matthias Müller, Antonio López, Vladlen Koltun, and Alexey Dosovitskiy. End-to-end driving via conditional imitation learning. In 2018 IEEE International Conference on Robotics and Automation (ICRA), pages 4693-4700, 2018." + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.722, + 0.483, + 0.789 + ], + "angle": 0, + "content": "[10] Felipe Codevilla, Eder Santana, Antonio M. Lopez, and Adrien Gaidon. Exploring the limitations of behavior cloning for autonomous driving. In Proceedings of the IEEE/CVF International Conference on Computer Vision (ICCV), 2019." + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.792, + 0.483, + 0.874 + ], + "angle": 0, + "content": "[11] Simon Doll, Niklas Hanselmann, Lukas Schneider, Richard Schulz, Marius Cordts, Markus Enzweiler, and Hendrik Lensch. Dualad: Disentangling the dynamic and static world for end-to-end driving. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 14728-14737, 2024. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.876, + 0.483, + 0.904 + ], + "angle": 0, + "content": "[12] Mingzhe Guo, Zhipeng Zhang, Yuan He, Ke Wang, and Liping Jing. End-to-end autonomous driving without costly" + }, + { + "type": "list", + "bbox": [ + 0.094, + 0.116, + 0.484, + 0.904 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.549, + 0.094, + 0.905, + 0.119 + ], + "angle": 0, + "content": "modularization and 3d manual annotation. arXiv preprint arXiv:2406.17680, 2024." + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.123, + 0.905, + 0.191 + ], + "angle": 0, + "content": "[13] Peiyun Hu, Aaron Huang, John Dolan, David Held, and Deva Ramanan. Safe local motion planning with self-supervised freespace forecasting. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 12732-12741, 2021. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.194, + 0.905, + 0.262 + ], + "angle": 0, + "content": "[14] Shengchao Hu, Li Chen, Penghao Wu, Hongyang Li, Junchi Yan, and Dacheng Tao. St-p3: End-to-end vision-based autonomous driving via spatial-temporal feature learning. In European Conference on Computer Vision, pages 533-549. Springer, 2022. 1, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.265, + 0.905, + 0.36 + ], + "angle": 0, + "content": "[15] Yihan Hu, Jiazhi Yang, Li Chen, Keyu Li, Chonghao Sima, Xizhou Zhu, Siqi Chai, Senyao Du, Tianwei Lin, Wenhai Wang, Lewei Lu, Xiaosong Jia, Qiang Liu, Jifeng Dai, Yu Qiao, and Hongyang Li. Planning-oriented autonomous driving. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 17853-17862, 2023. 1, 2, 3, 5, 6, 7, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.364, + 0.904, + 0.404 + ], + "angle": 0, + "content": "[16] Xiangdong Ji, Yizhuang Liu, Yu-Sheng Liu, Jian-Hui Zhang, and Yong Zhao. Large-momentum effective theory. Reviews of Modern Physics, 93(3):035005, 2021. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.407, + 0.905, + 0.474 + ], + "angle": 0, + "content": "[17] Xiaosong Jia, Zhenjie Yang, Qifeng Li, Zhiyuan Zhang, and Junchi Yan. Bench2drive: Towards multi-ability benchmarking of closed-loop end-to-end autonomous driving. Advances in Neural Information Processing Systems, 37:819–844, 2025. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.478, + 0.905, + 0.546 + ], + "angle": 0, + "content": "[18] Bo Jiang, Shaoyu Chen, Xinggang Wang, Bencheng Liao, Tianheng Cheng, Jiajie Chen, Helong Zhou, Qian Zhang, Wenyu Liu, and Chang Huang. Perceive, interact, predict: Learning dynamic and static clues for end-to-end motion prediction. arXiv preprint arXiv:2212.02181, 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.549, + 0.905, + 0.631 + ], + "angle": 0, + "content": "[19] Bo Jiang, Shaoyu Chen, Qing Xu, Benchcheng Liao, Jiajie Chen, Helong Zhou, Qian Zhang, Wenyu Liu, Chang Huang, and Xinggang Wang. Vad: Vectorized scene representation for efficient autonomous driving. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 8340-8350, 2023. 1, 2, 3, 5, 6, 7, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.634, + 0.905, + 0.688 + ], + "angle": 0, + "content": "[20] Bo Jiang, Shaoyu Chen, Benchcheng Liao, Xingyu Zhang, Wei Yin, Qian Zhang, Chang Huang, Wenyu Liu, and Xinggang Wang. Senna: Bridging large vision-language models and end-to-end autonomous driving, 2024." + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.691, + 0.905, + 0.758 + ], + "angle": 0, + "content": "[21] Zhenxin Li, Kailin Li, Shihao Wang, Shiyi Lan, Zhiding Yu, Yishen Ji, Zhiqi Li, Ziye Zhu, Jan Kautz, Zuxuan Wu, et al. Hydra-mdp: End-to-end multimodal planning with multitarget hydra-distillation. arXiv preprint arXiv:2406.06978, 2024. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.762, + 0.907, + 0.83 + ], + "angle": 0, + "content": "[22] Zhiqi Li, Zhiding Yu, Shiyi Lan, Jiahan Li, Jan Kautz, Tong Lu, and Jose M Alvarez. Is ego status all you need for open-loop end-to-end autonomous driving? In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 14864-14873, 2024. 5, 6, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.833, + 0.905, + 0.872 + ], + "angle": 0, + "content": "[23] Jiageng Mao, Yuxi Qian, Junjie Ye, Hang Zhao, and Yue Wang. Gpt-driver: Learning to drive with gpt. arXiv preprint arXiv:2310.01415, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.876, + 0.905, + 0.903 + ], + "angle": 0, + "content": "[24] Aditya Prakash, Kashyap Chitta, and Andreas Geiger. Multimodal fusion transformer for end-to-end autonomous driv" + }, + { + "type": "list", + "bbox": [ + 0.517, + 0.094, + 0.907, + 0.903 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.48, + 0.945, + 0.52, + 0.957 + ], + "angle": 0, + "content": "22440" + } + ], + [ + { + "type": "ref_text", + "bbox": [ + 0.125, + 0.092, + 0.484, + 0.12 + ], + "angle": 0, + "content": "ing. In 2021 IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), 2021. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.122, + 0.484, + 0.191 + ], + "angle": 0, + "content": "[25] Ziying Song, Lin Liu, Feiyang Jia, Yadan Luo, Caiyan Jia, Guoxin Zhang, Lei Yang, and Li Wang. Robustness-aware 3d object detection in autonomous driving: A review and outlook. IEEE Transactions on Intelligent Transportation Systems, pages 1-30, 2024. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.192, + 0.483, + 0.248 + ], + "angle": 0, + "content": "[26] Haisheng Su, Wei Wu, and Junchi Yan. Difsd: Ego-centric fully sparse paradigm with uncertainty denoising and iterative refinement for efficient end-to-end autonomous driving. arXiv preprint arXiv:2409.09777, 2024. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.249, + 0.483, + 0.305 + ], + "angle": 0, + "content": "[27] Wenchao Sun, Xuewu Lin, Yining Shi, Chuang Zhang, Hao ran Wu, and Sifa Zheng. Sparsedrive: End-to-end autonomous driving via sparse scene representation. arXiv preprint arXiv:2405.19620, 2024. 1, 2, 3, 5, 6, 7, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.306, + 0.483, + 0.361 + ], + "angle": 0, + "content": "[28] Junming Wang, Xingyu Zhang, Zebin Xing, Songen Gu, Xiaoyang Guo, Yang Hu, Ziying Song, Qian Zhang, Xiaoxiao Long, and Wei Yin. He-drive: Human-like end-to-end driving with vision language models, 2024. 1, 2, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.362, + 0.483, + 0.432 + ], + "angle": 0, + "content": "[29] Li Wang, Xinyu Zhang, Ziying Song, Jiangfeng Bi, Guoxin Zhang, Haiyue Wei, Liyao Tang, Lei Yang, Jun Li, Caiyan Jia, et al. Multi-modal 3d object detection in autonomous driving: A survey and taxonomy. IEEE Transactions on Intelligent Vehicles, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.433, + 0.483, + 0.502 + ], + "angle": 0, + "content": "[30] Penghao Wu, Xiaosong Jia, Li Chen, Junchi Yan, Hongyang Li, and Yu Qiao. Trajectory-guided control prediction for end-to-end autonomous driving: A simple yet strong baseline. In Advances in Neural Information Processing Systems, pages 6119-6132. Curran Associates, Inc., 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.503, + 0.483, + 0.558 + ], + "angle": 0, + "content": "[31] Zeyu Yang, Nan Song, Wei Li, Xiatian Zhu, Li Zhang, and Philip HS Torr. Deepinteraction++: Multi-modality interaction for autonomous driving. arXiv preprint arXiv:2408.05075, 2024." + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.56, + 0.483, + 0.629 + ], + "angle": 0, + "content": "[32] Tengju Ye, Wei Jing, Chunyong Hu, Shikun Huang, Lingping Gao, Fangzhen Li, Jingke Wang, Ke Guo, Wencong Xiao, Weibo Mao, Hang Zheng, Kun Li, Junbo Chen, and Kaicheng Yu. Fusionad: Multi-modality fusion for prediction and planning tasks of autonomous driving, 2023. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.63, + 0.483, + 0.698 + ], + "angle": 0, + "content": "[33] Wenyuan Zeng, Wenjie Luo, Simon Suo, Abbas Sadat, Bin Yang, Sergio Casas, and Raquel Urtasun. End-to-end interpretable neural motion planner. In 2019 IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), 2019." + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.7, + 0.483, + 0.769 + ], + "angle": 0, + "content": "[34] Yunpeng Zhang, Deheng Qian, Ding Li, Yifeng Pan, Yong Chen, Zhenbao Liang, Zhiyao Zhang, Shurui Zhang, Hongxu Li, Maolei Fu, et al. Graphad: Interaction scene graph for end-to-end autonomous driving. arXiv preprint arXiv:2403.19098, 2024." + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.771, + 0.483, + 0.84 + ], + "angle": 0, + "content": "[35] Zhejun Zhang, Alexander Liniger, Dengxin Dai, Fisher Yu, and Luc Van Gool. End-to-end urban driving by imitating a reinforcement learning coach. In Proceedings of the IEEE/CVF International Conference on Computer Vision (ICCV), pages 15222-15232, 2021." + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.841, + 0.483, + 0.883 + ], + "angle": 0, + "content": "[36] Wenzhao Zheng, Ruiqi Song, Xianda Guo, and Long Chen. Genad: Generative end-to-end autonomous driving. arXiv preprint arXiv:2402.11502, 2024. 2" + }, + { + "type": "list", + "bbox": [ + 0.093, + 0.092, + 0.484, + 0.883 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.48, + 0.946, + 0.518, + 0.957 + ], + "angle": 0, + "content": "22441" + } + ] +] \ No newline at end of file diff --git a/2025/Don't Shake the Wheel_ Momentum-Aware Planning in End-to-End Autonomous Driving/85994b24-3b2b-443a-a1bf-eabbda00b06b_origin.pdf b/2025/Don't Shake the Wheel_ Momentum-Aware Planning in End-to-End Autonomous Driving/85994b24-3b2b-443a-a1bf-eabbda00b06b_origin.pdf new file mode 100644 index 0000000000000000000000000000000000000000..39f752472199231be3dbe960a5d923c49dd4fe02 --- /dev/null +++ b/2025/Don't Shake the Wheel_ Momentum-Aware Planning in End-to-End Autonomous Driving/85994b24-3b2b-443a-a1bf-eabbda00b06b_origin.pdf @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:37f24e11e5fe524e93ae05f0832b50d4c25b08dd9d9bba564e3cb96947cfbd0a +size 2051023 diff --git a/2025/Don't Shake the Wheel_ Momentum-Aware Planning in End-to-End Autonomous Driving/full.md b/2025/Don't Shake the Wheel_ Momentum-Aware Planning in End-to-End Autonomous Driving/full.md new file mode 100644 index 0000000000000000000000000000000000000000..9a2b85e25d4027d300af3925263a648d4a8705c8 --- /dev/null +++ b/2025/Don't Shake the Wheel_ Momentum-Aware Planning in End-to-End Autonomous Driving/full.md @@ -0,0 +1,255 @@ +# Don't Shake the Wheel: Momentum-Aware Planning in End-to-End Autonomous Driving + +Ziying Song $^{1,2,\dagger}$ , Caiyan Jia $^{1,2,\star}$ , Lin Liu $^{1,2}$ , Hongyu Pan $^{3}$ , Yongchang Zhang $^{3}$ , Junming Wang $^{3,7}$ , Xingyu Zhang $^{3}$ , Shaoqing Xu $^{4}$ , Lei Yang $^{5}$ , Yadan Luo $^{6,\star}$ + +$^{1}$ School of Computer Science and Technology, Beijing Jiaotong University + +$^{2}$ Beijing Key Laboratory of Traffic Data Mining and Embodied Intelligence + +$^{3}$ Horizon Robotics $^{4}$ University of Macau $^{5}$ THU $^{6}$ The University of Queensland $^{7}$ HKU + +{songziying, cyjia}@bjtu.edu.cn, y.luo@uq.edu.au. + +# Abstract + +End-to-end autonomous driving frameworks enable seamless integration of perception and planning but often rely on one-shot trajectory prediction, which may lead to unstable control and vulnerability to occlusions in single-frame perception. To address this, we propose the Momentum-Aware Driving (MomAD) framework, which introduces trajectory momentum and perception momentum to stabilize and refine trajectory predictions. MomAD comprises two core components: (1) Topological Trajectory Matching (TTM) employs Hausdorff Distance to select the optimal planning query that aligns with prior paths to ensure coherence; (2) Momentum Planning Interactor (MPI) cross-attends the selected planning query with historical queries to expand static and dynamic perception files. This enriched query, in turn, helps regenerate long-horizon trajectory and reduce collision risks. To mitigate noise arising from dynamic environments and detection errors, we introduce robust instance denoising during training, enabling the planning model to focus on critical signals and improve its robustness. We also propose a novel Trajectory Prediction Consistency (TPC) metric to quantitatively assess planning stability. Experiments on the nuScenes dataset demonstrate that MomAD achieves superior long-term consistency $(\geq 3s)$ compared to SOTA methods. Moreover, evaluations on the curated Turning-nuScenes shows that MomAD reduces the collision rate by $26\%$ and improves TPC by $0.97m$ $(33.45\%)$ over a 6s prediction horizon, while closed-loop on Bench2Drive demonstrates an up to $16.3\%$ improvement in success rate. The source code is available at https://github.com/adept-thu/MomAD. + +![](images/04df1ebe27a0d8377e1cf1573ee359cf62c3b0c9f81069686903c6fc238a7f4f.jpg) +(a) Deterministic Planning + +![](images/c5aa0b8cde25aa4b247a21da4796663b7f51e4a85282ea312f3c08b31ad8dc78.jpg) +(b) Multi-modal Trajectory + +![](images/dc4ff25f585a9c2a8a9a67ed08cfb4ac1495c8745eb4c547abeb6457741bbdc4.jpg) + +![](images/d03e39cae0f6cbaeb15ee3f345b3e5e60680f2e7c34687159517be322f33550b.jpg) +(c) Momentum Planning (Ours) + +![](images/cee37a50b1899f64c3cb4a28cf30f60ca14efe4d6183c5c741efd6d03d4a6396.jpg) +Figure 1. (a) Deterministic Planning [14, 15, 19, 32] predicts deterministic trajectories, but lacks action diversity, posing safety risks. (b) Multi-modal Trajectory Planning [4, 27, 28] selects the highest-scoring trajectory among the multi-modal trajectories, yet fails to ensure stability and consistency, having risks in vehicle trembling. (c) Momentum Planning leverages the trajectory and perception momentum to enhance current planning through historical guidance to overcome temporal inconsistency. + +![](images/f5ead49f269af7f21c98a269e9b431bc77482c0b52323ee03525bcb78a2b5eb4.jpg) + +# 1. Introduction + +Autonomous driving [3, 25, 29] has undergone a transformative shift from modular, manually crafted pipelines to a more integrated, end-to-end paradigm [15, 19, 24]. Unlike traditional approaches that handle tasks like detection, tracking, mapping, motion prediction, and planning in isolation, the end-to-end framework emphasizes seamless integration. By prioritizing planning, it strategically directs information from upstream perception modules, thereby enhancing robustness and reliability in dynamic driving environments. + +Achieving high-quality planning in end-to-end frameworks hinges on accurately predicting the future trajectory prediction for the ego vehicle [1, 4, 7, 15, 19, 27]. Such future prediction requires a long-horizon understanding of both static and dynamic environmental factors, including map elements and interactions with surrounding agents. For instance, UniAD [15] queries the ego context from detailed bird's-eye-view (BEV) maps at each timestamp, while VAD [19] uses an ego query to retrieve surrounding context. The retrieved information then informs the planner, which predicts a deterministic trajectory for the vehicle, as illustrated in Figure 1 (a). Nevertheless, optimal trajectory prediction is inherently stochastic due to the unpredictability of other road users' intentions, varying road conditions, and the ambiguity introduced by human driving behaviors. This stochastic nature complicates the regression target, making deterministic predictions suboptimal and even risk-prone, potentially leading to severe collisions. To mitigate these uncertainties, methods such as VADv2 [4] and SparseDrive [27] leverage probabilistic modeling to capture the continuous planning action space, producing multi-modal trajectories that consider various possible behaviors of road agents, as shown in Figure 1 (b). While effective, these multi-modal approaches are typically one-shot and solely on the current perception frame. This limitation makes them susceptible to occlusion or loss of key visual cues, which can degrade multi-modal trajectory quality. Additionally, without temporal consistency, consecutive trajectories may lack coherence, causing unstable vehicle control and introducing undesirable directional shifts and oscillations. + +To stabilize trajectory prediction, we draw inspiration from human driving behaviors and introduce the concept of momentum into autonomous driving. In physics, momentum reflects an object's tendency to maintain its velocity based on speed and direction [16]. Analogously, in driving, momentum captures the smooth, forward progression of movement informed by past trajectories and modulated by present conditions. As illustrated in Figure 1 (c), by explicitly integrating historical trajectories with current predictions, we aim to achieve smoother and more coherent planning outcomes. To this end, we propose an end-to-end + +Momentum-Aware Driving (MomAD) framework, which incorporates momentum awareness to deliver stable and responsive planning in driving scenarios. MomAD interprets momentum on two levels: (1) trajectory momentum: By aligning candidate multi-modal trajectory with prior predictions, abrupt shifts in ego vehicle's path can be minimized, ensuring consistent control and a more comfortable driving experience. (2) perception momentum: By aggregating historical context and attending to map elements and surroundings over time, the model broadens its perspective, capturing subtle agent intentions missed in single-frame observations. To implement these ideas, we introduce (1) Topological Trajectory Matching (TTM): we first use this module to minimize planning discrepancies across time steps by employing the Hausdorff Distance to identify multimodal trajectory proposals that best align with past planning results. This approach ensures temporal coherence by preventing excessive deviation from previous trajectories. (2) Momentum Planning Interactor (MPI): Since the selected trajectory may still be biased toward the current perception and sacrifice long-horizon considerations, we cross-attends the current best planning query with historical plan queries, which implicitly convey critical long-term ego-temporal, ego-agent, and ego-map information as key and value vectors. This interaction enriches the current query with long-horizon perception momentum, improving its context awareness. To enhance robustness against environmental noise and perception errors, we incorporate a Robust Instance Denoising Module in the perception stage. By introducing controlled perturbations during training, the model learns to denoise perception inputs, achieving resilience to dynamic changes and misdetections. + +To evaluate the planning stability of MomAD, we propose a new Trajectory Prediction Consistency (TPC) metric to measure consistency between predicted and historical trajectories. Experiments demonstrate that MomAD can maintain long-term consistency ( $\geq 3$ seconds). Given that most scenes in nuScenes involve straight roads, which limit the assessment of temporal inconsistency, we curated a Turning-nuScenes validation set from turning scenarios within the nuScenes dataset to provide a more challenging evaluation, where our approach outperforms state-of-the-art end-to-end frameworks. For example, our MomAD reduces the collision rate by $26\%$ and the TPC by $0.97\mathrm{m}$ compared to SparseDrive [27] for a 6-second horizon prediction in the Turning-nuScenes validation set. + +# 2. Related Work + +End-to-end autonomous driving, which learns directly from raw sensor data to generate planning trajectories or driving commands, eliminates manual feature extraction [3, 25]. End-to-end autonomous driving methods [5, 6, 8-13, 15, 18-21, 23, 26, 28, 30-36] have garnered increased attent + +![](images/991d8a2d24ec69326049bda428c951dab6e0a97aaf3df4dcb21ae72add0abdfa.jpg) +Figure 2. The overall architecture of MomAD. MomAD, as a multi-modal trajectory end-to-end autonomous driving method, first encodes multi-view images into feature maps, then learns a sparse scene representation through a robust instance denoising via perturbation module, and finally performs a momentum planning through Topological Trajectory Matching (TTM) module and Momentum Planning Interactor (MPI) module to accomplish planning tasks. Our approach addresses critical challenges of stability and robustness in dynamic driving conditions. + +![](images/b3627d66e8e105144801867ff0063acb082dc4f467f283e9024241da1d0d4845.jpg) +Figure 3. The illustration of Momentum Planning Interactor (MPI). MPI cross-attends a selected planning query with historical queries to expand static and dynamic perception files, resulting in an enriched query that improves long-horizon trajectory generation and reduces collision risks. + +tion. UniAD [15] effectively integrates information from various preceding tasks, including perception, prediction, and planning modules to assist in trajectory planning, and achieves significant performance improvements. VAD [19] models driving scenarios as fully vectorized representations and employs explicit instance-level planning constraints to enhance planning safety. However, they [5, 11, 14, 15, 19] adopt a deterministic approach to trajectory prediction, which fails to account for trajectory diversity and may introduce risks due to intermediate regression results. VADv2 [4] proposes probabilistic planning to address the limitations of deterministic trajectory prediction, enabling multimodal trajectory predictions. Building upon the multimodal trajectory planning framework, SparseDrive [27] designs planning and motion prediction modules to achieve SOTA performance and efficiency on the nuScenes dataset. However, while multi-modal trajectory planning methods [4, 27, 28] have achieved SOTA performance, they overlook the temporal inconsistency caused by maximum score offsets. Regarding temporal inconsistency, existing methods only address temporal instance characteristics up to now, and they entirely overlook the issue of temporal planning + +consistency. In this work, we focus on this issue, aiming to address it using the concept of momentum planning. + +# 3. Method + +Framework Overview. Figure 2 presents an overview of the proposed MomAD system, which integrates sparse perception and momentum-aware planning. To capture key dynamic and static instances interacting with the ego vehicle, the sparse perception module builds upon the SparseDrive [27] to encode multi-view image features, which are aggregated into instance features $\mathbf{F}_t^{\mathrm{ins}}$ for road agents and map elements at time step $t$ . These features are obtained by sampling keypoints around the anchor boxes and polylines, feeding into the detection/tracking and online mapping blocks for accurate predictions. The core of MomAD is the joint motion and momentum-aware planning module, which comprises two main components: (1) Topological Trajectory Matching (Sec 3.1), which explicitly selects the candidate trajectory that best matches the prior path among all multi-modal trajectories to ensure temporal coherence; and (2) Momentum Planning Interaction (Sec 3.2), which expands the perceptive field by cross-attending the selected candidate trajectory's planning query with queries from the previous time step in the long-horizon query mixer. This approach provides a broader view of the surrounding environment and the intentions of other agents. The refined query is then processed by the planning head to generate updated multi-modal trajectories. Since the planning module heavily relies on detection and map instance features, we introduce a robust instance denoising via perturbation module (Sec 3.3) within the sparse perception component during training. This ensures robustness by reducing sensitivity to noisy perception features, enhancing the stability of trajectory prediction and planning. + +# 3.1. Topological Trajectory Matching (TTM) + +The proposed TTM module is inspired by the continuity of human driving behavior, where the optimal trajectories $T_{t}^{*}$ are influenced by the historical path $T_{t-1}$ to maintain temporal consistency and stability. Let $\mathcal{T}_t = \{T_t^k\}_{k=1}^K$ denote a set of $K$ multi-modal candidate trajectories generated at time step $t$ , where each trajectory $T_{t}^{k} = \{(x_{t,i}^{k},y_{t,i}^{k})\}_{i=1}^{N_{t}}$ consists of $N_{t}$ predicted waypoints. Typically, $K$ is set to $6\times 3$ to account for six trajectory proposals for each of three possible commands (left, right, and straight), and $N_{t}$ is chosen as 6 or 12 representing 0.5s interval for a 3- or 6-second prediction horizon. + +Trajectory Coordinate Transformation. Since the historical and current predicted trajectories are generated in the ego vehicle's coordinate system at different moments, it is essential to transform them into a common coordinate system for accurate matching. This transformation from moment $t$ to $t - 1$ is achieved as follows: + +$$ +T _ {t} ^ {k} \leftarrow \mathbf {R} _ {t - 1} ^ {- 1} \left(T _ {t} ^ {k} - \mathbf {\Gamma} _ {t - 1}\right), \forall k \in [ K ], \tag {1} +$$ + +where $\mathbf{R}_{t - 1}$ and $\Gamma_{t - 1}$ denote the rotation and translation matrix, respectively. + +Trajectory Distance Measurement. Simple Euclidean distance is inadequate for capturing the global alignment of trajectories, as it only measures pointwise proximity and is highly sensitive to local variations. This limitation becomes especially apparent in complex scenarios such as turns or varying point densities, where close points may not represent the alignment of the entire trajectory path. To address these limitations, TTM employs the Hausdorff distance as a more robust metric for evaluating trajectory alignment. The Hausdorff distance captures both local and global trajectory structures by measuring the maximum deviation between two sets of points, effectively quantifying the worst-case alignment between the candidate and historical trajectories. For each candidate trajectory $T_{t}^{k}$ , the Hausdorff distance to the historical trajectory $T_{t-1}$ is computed as, + +$$ +\begin{array}{l} d_{1} = \sup_{p\in T^{k}_{t}}\inf_{h\in T_{t - 1}}\| p - h\| , d_{2} = \sup_{h\in T_{t - 1}}\inf_{p\in T^{k}_{t}}\| h - p\| , \\ d _ {H} \left(T _ {t} ^ {k}, T _ {t - 1}\right) = \max \left(d _ {1}, d _ {2}\right), \tag {2} \\ \end{array} +$$ + +where $p \in T_t^k$ and $h \in T_{t - 1}$ represent waypoints in the candidate and historical trajectories, respectively. The Hausdorff distance considers the furthest point discrepancies between trajectories in both directions, ensuring that even minor global misalignments are captured. TTM then selects the trajectory $T_{t}^{k^{*}}$ that minimizes this distance: + +$$ +k ^ {*} = \underset {k \in [ K ]} {\arg \min } d _ {H} \left(T _ {t} ^ {k}, T _ {t - 1}\right) \tag {3} +$$ + +This selection enforces continuity, aligning with historical driving patterns and providing stable trajectory predictions that are less prone to sudden shifts. + +# 3.2. Momentum Planning Interactor (MPI) + +While TTM selects the most consistent trajectory $T_{t}^{k^{*}}$ based on historical alignment, $T_{t}^{k^{*}}$ is solely based on the current perception $\mathbf{F}_{\mathrm{ins}}$ , which may lack a comprehensive view of the environment and be sensitive to occlusions. Therefore, the MPI module, as illustrated in Figure 3, incorporates a long-horizon query mixer to enrich the selected planning query $\mathbf{Q}_t^{p^*}\in \mathbb{R}^{D_q}$ of $T_{t}^{k^{*}}$ with historical planning query $\mathbf{Q}_{t - 1}^p\in \mathbb{R}^{K\times D_q}$ and the associated planning scores $\mathbf{S}_{t - 1}^p\in \mathbb{R}^K$ , implicitly capturing a broader understanding of the surrounding context and other agents' intentions over time. Here, $D_{q}$ represents the latent dimension of planning queries. This enriched planning query $\tilde{\mathbf{Q}}_t^{p^*}\in \mathbb{R}^{D_q}$ will be combined with instance features to re-generate an improved trajectory $\tilde{\mathcal{T}}_t\in \mathbb{R}^{K\times N_t\times 2}$ . + +Long-horizon Query Mixer. To achieve a robust perception of temporal momentum, the query mixer allows crossattention between the selected candidate trajectory's planning query with multi-modal planning queries from the previous time step. The historical planning queries $\mathbf{Q}_{t-1}^p$ and associated scores $\mathbf{S}_{t-1}^p$ are combined through element-wise interaction and processed with an LSTM to simulate temporal evolution: + +$$ +\mathbf {Q} _ {t - 1} ^ {p ^ {\prime}} = \operatorname {L S T M} \left(\sigma \left(\mathbf {S} _ {t - 1} ^ {p}\right) \circ \operatorname {M L P} \left(\mathbf {Q} _ {t - 1} ^ {p}\right)\right), \tag {4} +$$ + +where $\sigma(\cdot)$ indicates the sigmoid function, MLP: $\mathbb{R}^{D_q} \mapsto \mathbb{R}^{D_q}$ the linear transform and $\circ$ the element-wise product. The LSTM processes this interaction, producing a surrogate multi-modal query $\mathbf{Q}_{t-1}^{p'} \in \mathbb{R}^{K \times D_q}$ that captures the temporal evolution of planning queries. To aggregate historical information, the current planning query $\mathbf{Q}_t^{p*}$ is used as a query in a cross-attention module. The result $\tilde{\mathbf{Q}}_t^{p*}$ incorporates long-term spatiotemporal context, which is further combined with the planning instance features and the encoded ego-anchor position information to inform the subsequent trajectory predictor: + +$$ +\begin{array}{l} \tilde {\mathbf {Q}} _ {t} ^ {p *} = \text {A t t e n t i o n} \left(\mathbf {Q} _ {t} ^ {p *}, \mathbf {Q} _ {t - 1} ^ {p ^ {\prime}}, \mathbf {Q} _ {t - 1} ^ {p ^ {\prime}}\right). \tag {5} \\ \tilde {\mathcal {T}} _ {t}, \tilde {\mathbf {S}} _ {t} = \operatorname {P l a n H e a d} \left(\tilde {\mathbf {Q}} _ {t} ^ {p *}, \mathbf {F} _ {t} ^ {\text {i n s}}\right), \\ \end{array} +$$ + +The PlanHead module then generates refined $\tilde{T}_t, \tilde{\mathbf{S}}_t$ . The best trajectory $T_t^*$ is then selected based on the highest scores among multi-modal outputs. Importantly, while the best trajectory is chosen based on the multi-modal trajectory scores, unlike previous selections, the current multimodal trajectories now fully consider the temporal consistency. This approach provides a stable, temporally-aware planning solution that is robust to occlusions and noise, significantly improving trajectory stability and control in complex driving environments. + +# 3.3. Robust Instance Denoising via Perturbation + +Our trajectory prediction and refinement rely heavily on the instance features $\mathbf{F}_t^{\mathrm{ins}}$ of road agents and map elements provided by the sparse perception module. However, due to detector instability and the dynamically changing map, these instance features may be noisy, potentially introducing errors in downstream planning. To enhance the stability of planning against such noisy inputs, we introduce controlled noise perturbations during training and employ a lightweight encoder-decoder transformer block (see Figure 2) to learn effective denoising. This approach enables the model to distinguish between essential and extraneous features, reducing the impact of perception noise on trajectory predictions. During test-time inference, this denoising capability allows the trajectory predictor to be resilient to fluctuations in instance features. As a result, the model can produce smoother, more stable trajectories even in challenging scenarios with occlusions, temporary obstacles, or misdetections. + +# 4. Experiments + +# 4.1. Experimental Setup + +Datasets. We conducted extensive experiments on the widely adopted nuScenes dataset [2] to evaluate tasks including detection, online mapping and planning in an open-loop setting. The nuScenes dataset comprises 1,000 driving scenes, with 700 and 150 sequences allocated for training and validation. Each scene spans around 20 seconds and contains roughly 40 key-frames annotated at $2\mathrm{Hz}$ , where each sample includes six images captured by surrounding cameras covering $360^{\circ}$ FOV horizontally and point clouds collected by both LiDAR and radar sensors. Since most planning tasks in nuScenes focus predominantly on gostraight commands, we curate a challenging subset of turning scenarios to form the Turning-nuScenes dataset, aimed at verifying the temporal consistency of predicted trajectories across time steps. Planning samples for turns are selected by setting the threshold between 3s and 0.5s of 'gt_ego_fut_trajs' to 25. The turning nuScenes validation dataset constitutes only one-tenth of the full nuScenes validation set, including 17 scenes with 680 samples. We use Bench2Drive [17], a closed-loop evaluation protocol under CARLA Leaderboard 2.0 for end-to-end autonomous driving. It provides an official training set, where we use the base set (1000 clips) for fair comparison with all the other baselines. We use the official 220 routes for evaluation. + +Evaluation Metrics for Planning. For planning evaluation, we adopt the commonly used L2 Displacement Error (L2) and Collision Rate to assess planning performance. The calculation of the L2 error follows VAD [19] and the collision rate is aligned with SparseDrive [27]. However, the mainstream planning metrics cannot faithfully reveal the + +stability of predicted trajectories. Therefore, we introduce a novel metric, Trajectory Prediction Consistency (TPC), to measure the disparity between current predicted trajectories and historical predicted trajectories, allowing for a more comprehensive assessment of the consistency of trajectories. With coordinates transformed, the TPC between the current predicted trajectory $T_{t}^{\mathrm{Pred}}$ and historical one $T_{t-1}^{\mathrm{Pred}}$ is defined as, + +$$ +\mathrm {T P C} = \frac {1}{N _ {T}} \sqrt {\sum_ {i = 1} ^ {N} \left(\left(T _ {t} ^ {\mathrm {P r e d}} - T _ {t - 1} ^ {\mathrm {P r e d}}\right) ^ {2} \cdot T _ {\mathrm {G T}} ^ {\mathrm {M a s k}}\right)}, \tag {6} +$$ + +where $N_{T}$ is the total number of GT trajectories in the validation set, and $T_{\mathrm{GT}}^{\mathrm{Mask}}$ is the mask for trajectories exceeding the overlapped time period of two trajectories. Our TPC metric evaluates whether autonomous vehicles adhere to predicted trajectories, ensuring continuity across frames. Notably, the TPC metric provides a statistical perspective on the dataset-wide evaluation rather than at the individual sample level. + +# 4.2. Main Results + +To conduct a comprehensive comparison, we have undertaken an exhaustive analysis of two distinct metrics presented in Table 1, which are derived from UniAD [15] and VAD [19]. It is noteworthy that, aside from Table 1, all other tables rely on the VAD [19] evaluation metrics for their assessments. + +# 4.2.1 Planning Results + +NuScenes. As shown in Table 1, MomAD achieves an L2 error of $0.60\mathrm{m}$ , a collision rate of $0.09\%$ , and a TPC of $0.54\mathrm{m}$ , respectively. Compared to SOTAs like UniAD [15], VAD [19] and SparseDrive [27], our method shows SOTA performance in planning results. It is worth noting that we have made significant improvements in TPC, which directly proves our effectiveness in timing consistency. It is worth noting that we achieved significant improvements in TPC at $0.30\mathrm{m}$ , $0.53\mathrm{m}$ , and $0.78\mathrm{m}$ at 1s, 2s, and 3s on the nuScenes dataset, directly demonstrating our effectiveness in temporal consistency. Additionally, our MomAD is straightforward and achieves an FPS of 7.8, a slightly slower than SparseDrive. In summary, our MomAD effectively utilizes the smoothing advantage of Momentum and has a significant effect on improving temporal consistency. + +Turning-nuScenes. As noted in [22], the nuScenes dataset features many straight routes, which limits the assessment of end-to-end methods. The simplicity of these paths can mask a model's true performance in complex scenarios. To address this, we evaluated our MomAD on the Turing-nuScenes validation set, as shown in Table 2. SparseDrive [27], a SOTA end-to-end method utilizing + +Table 1. Planning results on the nuScenes validation dataset. ${}^{ \dagger }$ denotes evaluation protocol used in UniAD [15]. * denotes results reproduced with the official checkpoint. As Ref. [22] states, we deactivate the ego status information for a fair comparison. + +
MethodInputBackboneL2 (m) ↓Col. Rate (%) ↓TPC (m) ↓FPS ↑
1s2s3sAvg.1s2s3sAvg.1s2s3sAvg.
UniAD† [15]CameraResNet1010.480.961.651.030.050.170.710.310.450.891.540.961.8 (A100)
VAD† [19]CameraResNet500.541.151.981.220.100.240.960.430.470.831.430.91-
SparseDrive†* [27]CameraResNet500.440.921.691.010.070.190.710.320.390.771.410.859.0 (RTX4090)
MomAD (Ours)†CameraResNet500.430.881.620.980.060.160.680.300.370.741.300.807.8 (RTX4090)
UniAD [15]CameraResNet1010.450.701.040.730.620.580.630.610.410.680.970.681.8 (A100)
VAD [19]CameraResNet500.410.701.050.720.030.190.430.210.360.660.910.64-
SparseDrive [27]CameraResNet500.290.580.960.610.010.050.180.080.300.570.850.579.0 (RTX4090)
MomAD (Ours)CameraResNet500.310.570.910.600.010.050.220.090.300.530.780.547.8 (RTX4090)
+ +Table 2. Planning results on the Turning-nuScenes validation dataset. SparseDrive [27] is a SOTA end-to-end multi-modal trajectory planning method. Red indicates improvement. We follow the VAD [19] evaluation metric. + +
MethodL2 (m) ↓Col. Rate (%) ↓TPC (m) ↓
1s2s3sAvg.1s2s3sAvg.1s2s3sAvg.
SparseDrive [27]0.350.771.460.860.040.170.980.400.340.701.330.79
MomAD (Ours)0.33-0.020.70-0.071.24-0.220.76-0.100.03-0.010.13-0.040.79-0.190.32-0.080.32-0.020.54-0.161.05-0.280.63-0.16
+ +Table 3. Long trajectory planning results on the nuScenes and Turning-nuScenes validation sets. We train models for 10 epochs for 6s-horizon prediction. T-nuScenes indicates the challenging Turning-nuScenes. We follow the VAD [19] evaluation metric. + +
SplitMethodL2 (m) ↓Col. Rate (%) ↓TPC (m) ↓
4s5s6s4s5s6s4s5s6s
nuScenesSparseDrive [27]1.752.322.950.871.542.331.331.661.99
MomAD1.671.982.450.831.432.131.191.451.61
-0.09-0.34-0.50-0.04-0.11-0.20-0.14-0.21-0.38
T-nuScenesSparseDrive [27]2.072.713.360.911.712.571.542.312.90
MomAD1.802.072.510.851.572.311.371.581.93
-0.27-0.64-0.85-0.06-0.14-0.26-0.17-0.73-0.97
+ +Table 4. Open-loop and Closed-loop results on Bench2Drive (V0.0.3) under base training set. 'mmt' refers multi-modal trajectory variant of VAD and * the re-implementation. + +
MethodOpen-loop MetricClosed-loop Metric
Avg. L2 ↓DS ↑ SR (%) ↑Effi ↑Comf ↑
VAD0.9142.3515.00157.9446.01
VADmmt*0.8942.8715.91158.1247.22
MomAD (Euclidean)0.8744.2216.91161.7748.70
MomAD0.8545.3517.44162.0949.34
SparseDrive*0.8744.5416.71170.2148.63
MomAD (Euclidean)0.8446.1217.45173.3550.98
MomAD0.8247.9118.11174.9151.20
+ +multi-modal trajectories, performs well across various scenarios but struggles with driving stability during turns. In contrast, MomAD exhibits superior consistency in trajectory predictions, as indicated by the TPC metric. Overall, MomAD not only delivers effective trajectory predictions under standard conditions but also maintains reliability amid dynamic changes and complex environments. + +Long Trajectory Prediction. Accurate long trajectory prediction is vital for enhancing the stability of autonomous + +driving, is useful to evaluate molels' ability to address temporal inconsistency issues in multi-modal trajectory planning. As shown in Table 3, we compared SparseDrive and MomAD in 4-6s long trajectory prediction on the nuScenes and Turning-nuScenes dataset, demonstrating a significant performance improvement. Specifically, in nuScenes dataset, compared with SparseDrive, MomAD experienced a decrease of $0.09\mathrm{m}$ $(5.14\%)$ , $0.34\mathrm{m}$ $(14.66\%)$ , and $0.50\mathrm{m}$ $(16.95\%)$ in the 4s, 5s, and 6s of L2 error, a decrease of $0.04\%$ $0.11\%$ and $0.20\%$ in the 4s, 5s, and 6s of collision rate, and a decrease of $0.14\mathrm{m}$ $(10.53\%)$ , $0.21\mathrm{m}$ $(12.65\%)$ , and $0.38\mathrm{m}$ $(19.10\%)$ in the 4s, 5s, and 6s of TPC, respectively. Furthermore, in Turning-nuScenes dataset, compared with SparseDrive, MomAD experienced a decrease of $0.27\mathrm{m}$ $(13.04\%)$ , $0.64\mathrm{m}$ $(23.62\%)$ , and $0.85\mathrm{m}$ $(25.30\%)$ in the 4s, 5s, and 6s of L2 error, a decrease of $0.06\%$ $0.14\%$ and $0.26\%$ in the 4s, 5s, and 6s of collision rate, and a decrease of $0.17\mathrm{m}$ $(11.04\%)$ , $0.73\mathrm{m}$ $(31.60\%)$ , and $0.97\mathrm{m}$ $(32.45\%)$ in the 4s, 5s, and 6s of TPC, respectively. In summary, our MomAD has improved the performance of long trajectory predictions, which further proves that MomAD can effectively alleviate the problem of temporal inconsistency. + +Bench2Drive. We have included evaluations on the challenging closed-loop results on Bench2Drive dataset, as shown in Table 4, w which covers 44 interactive scenes (e.g., cut-ins, overtaking, detours) and 220 routes across diverse weather conditions and locations. Our MomAD improves success rate by $16.3\%$ and $8.4\%$ over the VAD multi-modal variant and SparseDrive, and enhances the Comfortness score (trajectory smoothness) by $7.2\%$ and $5.3\%$ , demonstrating its effectiveness. + +Table 5. Perception and motion results on the nuScenes validation dataset. ${}^{ \dagger }$ indicates the results are reproduced with the official checkpoint. ${\mathrm{{AP}}}_{d}$ denotes ${\mathrm{{AP}}}_{\text{divider }}$ . ${\mathrm{{AP}}}_{b}$ denotes ${\mathrm{{AP}}}_{\text{boundary }}$ . mADE denotes minADE. mFDE denotes minFDE. + +
Method3D Object DetectionMulti-Object TrackingOnline MappingMotion Prediction
mAP ↑NDS ↑mATE ↓mASE ↓mAOE ↓mAVE ↓mAAE ↓AMOTA ↑AMOTP ↓Recall ↑IDS ↓mAP ↑APped ↑APd ↑APb ↑mADE ↓mFDE ↓MR ↓EPA ↑
UniAD [15]0.3800.4980.6840.2770.3830.3810.1920.3591.3200.467906----0.711.020.1510.456
VAD† [19]0.3120.4350.6100.2880.5410.5340.228----47.640.651.550.6----
SparseDrive [27]0.4180.5250.5660.2750.5520.2610.1900.3861.2540.49988655.149.957.058.40.620.990.1360.482
MomAD (Ours)0.4230.5310.5610.2690.5490.2580.1880.3911.2430.50985355.950.758.158.90.610.980.1370.499
+ +![](images/cd0c23434c044dc7720c52132f7566fec1c2ebe9ef23195e91eb55d45f9d0d48.jpg) + +![](images/19ae4902ce112b004de94d825a4617078236757c492b697d43361caea240d3b6.jpg) + +![](images/3d9435429872b91eeddacefa255e145856a828e73a19607f3603cd236a93a864.jpg) +Figure 4. Visualization results of MomAD compared with UniAD, VAD and SparseDrive across multiple frames. The proposed MomAD achieves temporal consistency whichever from the predicted trajectory compared with ground truth (GT) or from the TPC metric. + +![](images/bf70935d41cd2902af5a66c001957d649c65c9d7185e6c23096be6dbb5251b91.jpg) + +Table 6. Ablation studies of the sparse perception module in MomAD on the nuScenes validation split. The Encoder and Denoise $(\sigma)$ module is denoted as ED. MP represents Momentum planning. NS is the Gaussian noise factor controlling the noise level. Noise is applied during training only. We follow the VAD [19] evaluation metric. + +
ED MP NSDetectionTrackingOnline MappingMotionPlanning (Avg.)
mAP ↑NDS ↑AMOTA ↑mAP ↑APped ↑mADE ↓L2 ↓Col. ↓TPC ↓
✓0.00.4070.5210.38155.049.30.630.620.140.56
0.00.4050.5200.38055.149.50.630.610.130.55
✓0.10.4200.5300.39055.850.50.580.610.120.55
✓0.00.4170.5280.38655.450.60.630.610.110.55
✓0.050.4210.5290.38855.650.80.620.610.110.54
✓0.10.4230.5310.39155.950.70.610.600.090.54
✓0.20.4180.5200.38854.449.20.630.620.180.58
✓0.30.4120.5180.38354.048.80.650.640.220.61
+ +# 4.2.2 Perception and Motion Prediction Results + +Sparse representation is efficient but suffers from instability issues caused by the variability of instance features. To address these issues, we have enhanced the instance features using the Encoder and Denoise $(\sigma)$ module (denoted as ED) within the sparse perception framework, ensuring + +Table 7. Impact of history frames in MomAD on the Turning-nuScenes validation set. $t$ denotes the frame number, where 1 indicates the history is empty (represented by 0), 2 signifies that the historical result corresponds to the previous 1 frame, and 3 indicates that the historical result pertains to the previous 2 frames. We follow the VAD [19] evaluation metric. + +
EDMPNStL2(m) ↓Col. Rate(%) ↓TPC(m) ↓
2s3sAvg.2s3sAvg.2s3sAvg.
0.771.460.860.170.980.400.701.330.79
✓20.711.270.770.140.830.340.551.070.65
✓0.10.771.450.860.180.970.390.691.330.79
✓✓0.110.781.480.880.180.980.390.701.350.81
✓✓0.120.701.240.760.130.790.320.541.050.63
✓✓0.130.721.270.780.140.840.350.561.090.66
+ +end-to-end stability for autonomous driving. As shown in Table 5, our MomAD perception module includes 3D object detection, multi-object tracking, and online mapping tasks. For 3D object detection, MomAD achieves $42.3\%$ mAP and $53.1\%$ NPS, improving the mAP by $0.5\%$ and the NDS by $0.6\%$ compared to the baseline SparseDrive [27]. For multi-object tracking, MomAD achieves an AMOTA of + +Table 8. Ablation studies of the impact of the different modules in MP on the Turning-nuScenes validation dataset. QM denotes Long-horizon Query Mixer, TP denotes Trajectory Predictor. Add refers to the addition operation between the historical planning query $\mathbf{Q}_{t-1}^p$ and the selected planning query $\mathbf{Q}_t^{p*}$ . We follow the VAD [19] evaluation metric. + +
QMAddTPL2 (m) ↓Col. Rate (%) ↓TPC (m) ↓
2s3sAvg.2s3sAvg.2s3sAvg.
0.771.460.860.170.980.400.701.330.79
✓✓0.761.380.820.140.880.360.621.210.67
✓✓0.701.240.760.130.790.320.541.050.63
+ +39.1%, surpassing the baseline SparseDrive by 0.5%. For online mapping, compared to 55.1% mAP for the baseline SparseDrive, our MomAD achieves 55.9% mAP, improving the mAP by 0.8%. For motion prediction, our MomAD outperforms SparseDrive [27] and UniAD [15], achieves better motion prediction performance by considering the influence of an ego vehicle on other agents. In detail, Our MomAD achieves a 0.61m minADE, 0.98 minFDE, and 13.7% MR, and 0.499 EPA, respectively. + +# 4.3. Ablation Study + +Roles of 'ED' module in Sparse Perception. Sparse representation end-to-end methods yield efficient computation but unstable metrics. Further details are available in the Appendix. As shown in Table 6, the Encoder and Denoise $(\sigma)$ module (ED) within sparse perception enhances the instance features, significantly impacts the overall pipeline. By introducing Gaussian noise and employing techniques, the robustness of instance features is improved, particularly when training with Noisy at 0.1. Our findings suggest that a controlled level of noise can enhance the end-to-end capabilities of sparse methods during training, offering insights for the community on sparse end-to-end methods. + +Roles of 'MP' module in Planing. As Li et al. [22] have stated, most end-to-end autonomous driving methods perform poorly in turning scenarios. As shown in Table 7, to better evaluate the planning performance of end-to-end methods in turning scenarios, our MomAD is evaluated on the Turning-nuScenes validation dataset rather than only on the full nuScenes validation dataset. Specifically, under the premise of executing 'ED', at $t = 1$ , providing a 0 value to the MP module does not improve performance. We have tried to change the MLP operation of the planning to a more complex operation, but it does not enhance the results. However, when $t = 2$ , historical queries and results are used, L2 (Avg) reaches 0.76m, Col. (Avg) reaches 0.32 %, and TPC reaches 0.63m, which represents a significant improvement. In addition, when $t = 3$ , more frames are fused, the improvement has actually decreased, which may be due to the uncertainty introduced by the departure of historical features, but there is still an overall improvement. + +Roles of different sub-modules in 'MP' module. As shown in Table 8, we conducted an in-depth analysis of the internal mechanism of momentum planning. We found that simply using the native 'Add' operation to regenerate the planning results can achieve a good improvement, with L2 (Avg.), Col. (Avg.), and TPC (Avg.) decreasing by $0.04\mathrm{m}$ , $0.04\%$ , and $0.12\mathrm{m}$ , respectively. However, the 'Add' operation alone does not fully utilize historical features. Our Long-horizon Query Mixer has achieved the current optimal performance. Overall, historical results are very important for current outcomes, and their reasonable utilization can maximize the performance of end-to-end planning. + +# 4.4. Visualization + +As shown in Figure 4, we showcase multi-frame qualitative comparisons of end-to-end solutions, including UniAD [15], VAD [19], SparseDrive [27], and the proposed MomAD. In a representative turning scenario, the MomAD approach demonstrates superior long-term awareness of surrounding vehicles, reducing the likelihood of collisions. Additionally, it generates smoother ego-vehicle trajectories (shown in yellow and blue) that closely align with the ground-truth trajectory (in red). This highlights its strong temporal consistency and lower TPC scores. Additional visualizations for various driving commands are provided in the Appendix. + +# 5. Conclusion and Future Work + +The proposed MomAD framework addresses key challenges in planning stability and robustness for end-to-end autonomous driving. By leveraging trajectory momentum and perception momentum, MomAD stabilizes trajectory predictions through Topological Trajectory Matching (TTM) and Momentum Planning Interactor (MPI), ensuring temporal coherence and enriching long-horizon context. Evaluations on nuScenes, the curated Turning-nuScenes validation set, and Bench2Drive demonstrate its superior performance in reducing collision rates and improving trajectory consistency compared to SOTA methods. While MomAD improves temporal consistency in long-horizon trajectory prediction, a gap remains due to mode collapse induced by the standard teacher-forcing approach to trajectory regression, limiting trajectory diversity. Future work will explore diffusion models and speculative decoding to enhance trajectory diversity while balancing efficiency. + +# Acknowledgements + +We appreciate the valuable discussions with Wenchao Sun, Bo Jiang, and Bencheng Liao from Horizon Robotics. This work was supported by the National Key R&D Program of China (2018AAA0100302). + +# References + +[1] Szilard Aradi. Survey of deep reinforcement learning for motion planning of autonomous vehicles. IEEE Transactions on Intelligent Transportation Systems, 23(2):740-759, 2020. 2 +[2] Holger Caesar, Varun Bankiti, Alex H Lang, Sourabh Vora, Venice Erin Liong, Qiang Xu, Anush Krishnan, Yu Pan, Giancarlo Baldan, and Oscar Beijbom. nuscenes: A multimodal dataset for autonomous driving. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 11621-11631, 2020. 5 +[3] Li Chen, Penghao Wu, Kashyap Chitta, Bernhard Jaeger, Andreas Geiger, and Hongyang Li. End-to-end autonomous driving: Challenges and frontiers. IEEE Transactions on Pattern Analysis and Machine Intelligence, 2024. 2 +[4] Shaoyu Chen, Bo Jiang, Hao Gao, Bencheng Liao, Qing Xu, Qian Zhang, Chang Huang, Wenyu Liu, and Xinggang Wang. Vadv2: End-to-end vectorized autonomous driving via probabilistic planning. arXiv preprint arXiv:2402.13243, 2024. 1, 2, 3 +[5] Zhili Chen, Maosheng Ye, Shuangjie Xu, Tongyi Cao, and Qifeng Chen. Ppad: Iterative interactions of prediction and planning for end-to-end autonomous driving. In European Conference on Computer Vision, pages 239-256. Springer, 2025. 2, 3 +[6] Jie Cheng, Yingbing Chen, Xiaodong Mei, Bowen Yang, Bo Li, and Ming Liu. Rethinking imitation-based planners for autonomous driving. In 2024 IEEE International Conference on Robotics and Automation (ICRA), pages 14123-14130. IEEE, 2024. 2 +[7] Laurene Claussmann, Marc Revilloud, Dominique Gruyer, and Sébastien Glaser. A review of motion planning for highway autonomous driving. IEEE Transactions on Intelligent Transportation Systems, 21(5):1826-1848, 2020. 2 +[8] Felipe Codevilla, Matthias Müller, Antonio López, Vladlen Koltun, and Alexey Dosovitskiy. End-to-end driving via conditional imitation learning. In 2018 IEEE international conference on robotics and automation (ICRA), pages 4693-4700. IEEE, 2018. 2 +[9] Felipe Codevilla, Matthias Müller, Antonio López, Vladlen Koltun, and Alexey Dosovitskiy. End-to-end driving via conditional imitation learning. In 2018 IEEE International Conference on Robotics and Automation (ICRA), pages 4693-4700, 2018. +[10] Felipe Codevilla, Eder Santana, Antonio M. Lopez, and Adrien Gaidon. Exploring the limitations of behavior cloning for autonomous driving. In Proceedings of the IEEE/CVF International Conference on Computer Vision (ICCV), 2019. +[11] Simon Doll, Niklas Hanselmann, Lukas Schneider, Richard Schulz, Marius Cordts, Markus Enzweiler, and Hendrik Lensch. Dualad: Disentangling the dynamic and static world for end-to-end driving. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 14728-14737, 2024. 3 +[12] Mingzhe Guo, Zhipeng Zhang, Yuan He, Ke Wang, and Liping Jing. End-to-end autonomous driving without costly + +modularization and 3d manual annotation. arXiv preprint arXiv:2406.17680, 2024. +[13] Peiyun Hu, Aaron Huang, John Dolan, David Held, and Deva Ramanan. Safe local motion planning with self-supervised freespace forecasting. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 12732-12741, 2021. 2 +[14] Shengchao Hu, Li Chen, Penghao Wu, Hongyang Li, Junchi Yan, and Dacheng Tao. St-p3: End-to-end vision-based autonomous driving via spatial-temporal feature learning. In European Conference on Computer Vision, pages 533-549. Springer, 2022. 1, 3 +[15] Yihan Hu, Jiazhi Yang, Li Chen, Keyu Li, Chonghao Sima, Xizhou Zhu, Siqi Chai, Senyao Du, Tianwei Lin, Wenhai Wang, Lewei Lu, Xiaosong Jia, Qiang Liu, Jifeng Dai, Yu Qiao, and Hongyang Li. Planning-oriented autonomous driving. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 17853-17862, 2023. 1, 2, 3, 5, 6, 7, 8 +[16] Xiangdong Ji, Yizhuang Liu, Yu-Sheng Liu, Jian-Hui Zhang, and Yong Zhao. Large-momentum effective theory. Reviews of Modern Physics, 93(3):035005, 2021. 2 +[17] Xiaosong Jia, Zhenjie Yang, Qifeng Li, Zhiyuan Zhang, and Junchi Yan. Bench2drive: Towards multi-ability benchmarking of closed-loop end-to-end autonomous driving. Advances in Neural Information Processing Systems, 37:819–844, 2025. 5 +[18] Bo Jiang, Shaoyu Chen, Xinggang Wang, Bencheng Liao, Tianheng Cheng, Jiajie Chen, Helong Zhou, Qian Zhang, Wenyu Liu, and Chang Huang. Perceive, interact, predict: Learning dynamic and static clues for end-to-end motion prediction. arXiv preprint arXiv:2212.02181, 2022. 2 +[19] Bo Jiang, Shaoyu Chen, Qing Xu, Benchcheng Liao, Jiajie Chen, Helong Zhou, Qian Zhang, Wenyu Liu, Chang Huang, and Xinggang Wang. Vad: Vectorized scene representation for efficient autonomous driving. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 8340-8350, 2023. 1, 2, 3, 5, 6, 7, 8 +[20] Bo Jiang, Shaoyu Chen, Benchcheng Liao, Xingyu Zhang, Wei Yin, Qian Zhang, Chang Huang, Wenyu Liu, and Xinggang Wang. Senna: Bridging large vision-language models and end-to-end autonomous driving, 2024. +[21] Zhenxin Li, Kailin Li, Shihao Wang, Shiyi Lan, Zhiding Yu, Yishen Ji, Zhiqi Li, Ziye Zhu, Jan Kautz, Zuxuan Wu, et al. Hydra-mdp: End-to-end multimodal planning with multitarget hydra-distillation. arXiv preprint arXiv:2406.06978, 2024. 2 +[22] Zhiqi Li, Zhiding Yu, Shiyi Lan, Jiahan Li, Jan Kautz, Tong Lu, and Jose M Alvarez. Is ego status all you need for open-loop end-to-end autonomous driving? In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 14864-14873, 2024. 5, 6, 8 +[23] Jiageng Mao, Yuxi Qian, Junjie Ye, Hang Zhao, and Yue Wang. Gpt-driver: Learning to drive with gpt. arXiv preprint arXiv:2310.01415, 2023. 2 +[24] Aditya Prakash, Kashyap Chitta, and Andreas Geiger. Multimodal fusion transformer for end-to-end autonomous driv + +ing. In 2021 IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), 2021. 2 +[25] Ziying Song, Lin Liu, Feiyang Jia, Yadan Luo, Caiyan Jia, Guoxin Zhang, Lei Yang, and Li Wang. Robustness-aware 3d object detection in autonomous driving: A review and outlook. IEEE Transactions on Intelligent Transportation Systems, pages 1-30, 2024. 2 +[26] Haisheng Su, Wei Wu, and Junchi Yan. Difsd: Ego-centric fully sparse paradigm with uncertainty denoising and iterative refinement for efficient end-to-end autonomous driving. arXiv preprint arXiv:2409.09777, 2024. 2 +[27] Wenchao Sun, Xuewu Lin, Yining Shi, Chuang Zhang, Hao ran Wu, and Sifa Zheng. Sparsedrive: End-to-end autonomous driving via sparse scene representation. arXiv preprint arXiv:2405.19620, 2024. 1, 2, 3, 5, 6, 7, 8 +[28] Junming Wang, Xingyu Zhang, Zebin Xing, Songen Gu, Xiaoyang Guo, Yang Hu, Ziying Song, Qian Zhang, Xiaoxiao Long, and Wei Yin. He-drive: Human-like end-to-end driving with vision language models, 2024. 1, 2, 3 +[29] Li Wang, Xinyu Zhang, Ziying Song, Jiangfeng Bi, Guoxin Zhang, Haiyue Wei, Liyao Tang, Lei Yang, Jun Li, Caiyan Jia, et al. Multi-modal 3d object detection in autonomous driving: A survey and taxonomy. IEEE Transactions on Intelligent Vehicles, 2023. 2 +[30] Penghao Wu, Xiaosong Jia, Li Chen, Junchi Yan, Hongyang Li, and Yu Qiao. Trajectory-guided control prediction for end-to-end autonomous driving: A simple yet strong baseline. In Advances in Neural Information Processing Systems, pages 6119-6132. Curran Associates, Inc., 2022. 2 +[31] Zeyu Yang, Nan Song, Wei Li, Xiatian Zhu, Li Zhang, and Philip HS Torr. Deepinteraction++: Multi-modality interaction for autonomous driving. arXiv preprint arXiv:2408.05075, 2024. +[32] Tengju Ye, Wei Jing, Chunyong Hu, Shikun Huang, Lingping Gao, Fangzhen Li, Jingke Wang, Ke Guo, Wencong Xiao, Weibo Mao, Hang Zheng, Kun Li, Junbo Chen, and Kaicheng Yu. Fusionad: Multi-modality fusion for prediction and planning tasks of autonomous driving, 2023. 1 +[33] Wenyuan Zeng, Wenjie Luo, Simon Suo, Abbas Sadat, Bin Yang, Sergio Casas, and Raquel Urtasun. End-to-end interpretable neural motion planner. In 2019 IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), 2019. +[34] Yunpeng Zhang, Deheng Qian, Ding Li, Yifeng Pan, Yong Chen, Zhenbao Liang, Zhiyao Zhang, Shurui Zhang, Hongxu Li, Maolei Fu, et al. Graphad: Interaction scene graph for end-to-end autonomous driving. arXiv preprint arXiv:2403.19098, 2024. +[35] Zhejun Zhang, Alexander Liniger, Dengxin Dai, Fisher Yu, and Luc Van Gool. End-to-end urban driving by imitating a reinforcement learning coach. In Proceedings of the IEEE/CVF International Conference on Computer Vision (ICCV), pages 15222-15232, 2021. +[36] Wenzhao Zheng, Ruiqi Song, Xianda Guo, and Long Chen. Genad: Generative end-to-end autonomous driving. arXiv preprint arXiv:2402.11502, 2024. 2 \ No newline at end of file diff --git a/2025/Don't Shake the Wheel_ Momentum-Aware Planning in End-to-End Autonomous Driving/images.zip b/2025/Don't Shake the Wheel_ Momentum-Aware Planning in End-to-End Autonomous Driving/images.zip new file mode 100644 index 0000000000000000000000000000000000000000..6ee9eb2e0c219553ec6a0d1482e6fb6fb5444fa7 --- /dev/null +++ b/2025/Don't Shake the Wheel_ Momentum-Aware Planning in End-to-End Autonomous Driving/images.zip @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:fccc577dc9c9a6c998e426874c621239d8c2786282c80b5a9fdd27fa65df3957 +size 628508 diff --git a/2025/Don't Shake the Wheel_ Momentum-Aware Planning in End-to-End Autonomous Driving/layout.json b/2025/Don't Shake the Wheel_ Momentum-Aware Planning in End-to-End Autonomous Driving/layout.json new file mode 100644 index 0000000000000000000000000000000000000000..e43cc0d2705295a5fc403693beb556ea3689cead --- /dev/null +++ b/2025/Don't Shake the Wheel_ Momentum-Aware Planning in End-to-End Autonomous Driving/layout.json @@ -0,0 +1,8070 @@ +{ + "pdf_info": [ + { + "para_blocks": [ + { + "bbox": [ + 134, + 103, + 477, + 140 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 134, + 103, + 477, + 140 + ], + "spans": [ + { + "bbox": [ + 134, + 103, + 477, + 140 + ], + "type": "text", + "content": "Don't Shake the Wheel: Momentum-Aware Planning in End-to-End Autonomous Driving" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 113, + 160, + 496, + 190 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 113, + 160, + 496, + 190 + ], + "spans": [ + { + "bbox": [ + 113, + 160, + 496, + 190 + ], + "type": "text", + "content": "Ziying Song" + }, + { + "bbox": [ + 113, + 160, + 496, + 190 + ], + "type": "inline_equation", + "content": "^{1,2,\\dagger}" + }, + { + "bbox": [ + 113, + 160, + 496, + 190 + ], + "type": "text", + "content": ", Caiyan Jia" + }, + { + "bbox": [ + 113, + 160, + 496, + 190 + ], + "type": "inline_equation", + "content": "^{1,2,\\star}" + }, + { + "bbox": [ + 113, + 160, + 496, + 190 + ], + "type": "text", + "content": ", Lin Liu" + }, + { + "bbox": [ + 113, + 160, + 496, + 190 + ], + "type": "inline_equation", + "content": "^{1,2}" + }, + { + "bbox": [ + 113, + 160, + 496, + 190 + ], + "type": "text", + "content": ", Hongyu Pan" + }, + { + "bbox": [ + 113, + 160, + 496, + 190 + ], + "type": "inline_equation", + "content": "^{3}" + }, + { + "bbox": [ + 113, + 160, + 496, + 190 + ], + "type": "text", + "content": ", Yongchang Zhang" + }, + { + "bbox": [ + 113, + 160, + 496, + 190 + ], + "type": "inline_equation", + "content": "^{3}" + }, + { + "bbox": [ + 113, + 160, + 496, + 190 + ], + "type": "text", + "content": ", Junming Wang" + }, + { + "bbox": [ + 113, + 160, + 496, + 190 + ], + "type": "inline_equation", + "content": "^{3,7}" + }, + { + "bbox": [ + 113, + 160, + 496, + 190 + ], + "type": "text", + "content": ", Xingyu Zhang" + }, + { + "bbox": [ + 113, + 160, + 496, + 190 + ], + "type": "inline_equation", + "content": "^{3}" + }, + { + "bbox": [ + 113, + 160, + 496, + 190 + ], + "type": "text", + "content": ", Shaoqing Xu" + }, + { + "bbox": [ + 113, + 160, + 496, + 190 + ], + "type": "inline_equation", + "content": "^{4}" + }, + { + "bbox": [ + 113, + 160, + 496, + 190 + ], + "type": "text", + "content": ", Lei Yang" + }, + { + "bbox": [ + 113, + 160, + 496, + 190 + ], + "type": "inline_equation", + "content": "^{5}" + }, + { + "bbox": [ + 113, + 160, + 496, + 190 + ], + "type": "text", + "content": ", Yadan Luo" + }, + { + "bbox": [ + 113, + 160, + 496, + 190 + ], + "type": "inline_equation", + "content": "^{6,\\star}" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 123, + 194, + 487, + 209 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 123, + 194, + 487, + 209 + ], + "spans": [ + { + "bbox": [ + 123, + 194, + 487, + 209 + ], + "type": "inline_equation", + "content": "^{1}" + }, + { + "bbox": [ + 123, + 194, + 487, + 209 + ], + "type": "text", + "content": "School of Computer Science and Technology, Beijing Jiaotong University" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 123, + 209, + 488, + 223 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 123, + 209, + 488, + 223 + ], + "spans": [ + { + "bbox": [ + 123, + 209, + 488, + 223 + ], + "type": "inline_equation", + "content": "^{2}" + }, + { + "bbox": [ + 123, + 209, + 488, + 223 + ], + "type": "text", + "content": "Beijing Key Laboratory of Traffic Data Mining and Embodied Intelligence" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 97, + 223, + 512, + 237 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 97, + 223, + 512, + 237 + ], + "spans": [ + { + "bbox": [ + 97, + 223, + 512, + 237 + ], + "type": "inline_equation", + "content": "^{3}" + }, + { + "bbox": [ + 97, + 223, + 512, + 237 + ], + "type": "text", + "content": "Horizon Robotics " + }, + { + "bbox": [ + 97, + 223, + 512, + 237 + ], + "type": "inline_equation", + "content": "^{4}" + }, + { + "bbox": [ + 97, + 223, + 512, + 237 + ], + "type": "text", + "content": "University of Macau " + }, + { + "bbox": [ + 97, + 223, + 512, + 237 + ], + "type": "inline_equation", + "content": "^{5}" + }, + { + "bbox": [ + 97, + 223, + 512, + 237 + ], + "type": "text", + "content": "THU " + }, + { + "bbox": [ + 97, + 223, + 512, + 237 + ], + "type": "inline_equation", + "content": "^{6}" + }, + { + "bbox": [ + 97, + 223, + 512, + 237 + ], + "type": "text", + "content": "The University of Queensland " + }, + { + "bbox": [ + 97, + 223, + 512, + 237 + ], + "type": "inline_equation", + "content": "^{7}" + }, + { + "bbox": [ + 97, + 223, + 512, + 237 + ], + "type": "text", + "content": "HKU" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 167, + 239, + 430, + 251 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 167, + 239, + 430, + 251 + ], + "spans": [ + { + "bbox": [ + 167, + 239, + 430, + 251 + ], + "type": "text", + "content": "{songziying, cyjia}@bjtu.edu.cn, y.luo@uq.edu.au." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 151, + 274, + 200, + 287 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 151, + 274, + 200, + 287 + ], + "spans": [ + { + "bbox": [ + 151, + 274, + 200, + 287 + ], + "type": "text", + "content": "Abstract" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 54, + 308, + 296, + 668 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 54, + 308, + 296, + 668 + ], + "spans": [ + { + "bbox": [ + 54, + 308, + 296, + 668 + ], + "type": "text", + "content": "End-to-end autonomous driving frameworks enable seamless integration of perception and planning but often rely on one-shot trajectory prediction, which may lead to unstable control and vulnerability to occlusions in single-frame perception. To address this, we propose the Momentum-Aware Driving (MomAD) framework, which introduces trajectory momentum and perception momentum to stabilize and refine trajectory predictions. MomAD comprises two core components: (1) Topological Trajectory Matching (TTM) employs Hausdorff Distance to select the optimal planning query that aligns with prior paths to ensure coherence; (2) Momentum Planning Interactor (MPI) cross-attends the selected planning query with historical queries to expand static and dynamic perception files. This enriched query, in turn, helps regenerate long-horizon trajectory and reduce collision risks. To mitigate noise arising from dynamic environments and detection errors, we introduce robust instance denoising during training, enabling the planning model to focus on critical signals and improve its robustness. We also propose a novel Trajectory Prediction Consistency (TPC) metric to quantitatively assess planning stability. Experiments on the nuScenes dataset demonstrate that MomAD achieves superior long-term consistency " + }, + { + "bbox": [ + 54, + 308, + 296, + 668 + ], + "type": "inline_equation", + "content": "(\\geq 3s)" + }, + { + "bbox": [ + 54, + 308, + 296, + 668 + ], + "type": "text", + "content": " compared to SOTA methods. Moreover, evaluations on the curated Turning-nuScenes shows that MomAD reduces the collision rate by " + }, + { + "bbox": [ + 54, + 308, + 296, + 668 + ], + "type": "inline_equation", + "content": "26\\%" + }, + { + "bbox": [ + 54, + 308, + 296, + 668 + ], + "type": "text", + "content": " and improves TPC by " + }, + { + "bbox": [ + 54, + 308, + 296, + 668 + ], + "type": "inline_equation", + "content": "0.97m" + }, + { + "bbox": [ + 54, + 308, + 296, + 668 + ], + "type": "inline_equation", + "content": "(33.45\\%)" + }, + { + "bbox": [ + 54, + 308, + 296, + 668 + ], + "type": "text", + "content": " over a 6s prediction horizon, while closed-loop on Bench2Drive demonstrates an up to " + }, + { + "bbox": [ + 54, + 308, + 296, + 668 + ], + "type": "inline_equation", + "content": "16.3\\%" + }, + { + "bbox": [ + 54, + 308, + 296, + 668 + ], + "type": "text", + "content": " improvement in success rate. The source code is available at https://github.com/adept-thu/MomAD." + } + ] + } + ], + "index": 9 + }, + { + "type": "image", + "bbox": [ + 317, + 274, + 389, + 409 + ], + "blocks": [ + { + "bbox": [ + 317, + 274, + 389, + 409 + ], + "lines": [ + { + "bbox": [ + 317, + 274, + 389, + 409 + ], + "spans": [ + { + "bbox": [ + 317, + 274, + 389, + 409 + ], + "type": "image", + "image_path": "04df1ebe27a0d8377e1cf1573ee359cf62c3b0c9f81069686903c6fc238a7f4f.jpg" + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 317, + 411, + 408, + 421 + ], + "lines": [ + { + "bbox": [ + 317, + 411, + 408, + 421 + ], + "spans": [ + { + "bbox": [ + 317, + 411, + 408, + 421 + ], + "type": "text", + "content": "(a) Deterministic Planning" + } + ] + } + ], + "index": 11, + "angle": 0, + "type": "image_caption" + } + ], + "index": 10 + }, + { + "type": "image", + "bbox": [ + 391, + 274, + 464, + 409 + ], + "blocks": [ + { + "bbox": [ + 391, + 274, + 464, + 409 + ], + "lines": [ + { + "bbox": [ + 391, + 274, + 464, + 409 + ], + "spans": [ + { + "bbox": [ + 391, + 274, + 464, + 409 + ], + "type": "image", + "image_path": "c5aa0b8cde25aa4b247a21da4796663b7f51e4a85282ea312f3c08b31ad8dc78.jpg" + } + ] + } + ], + "index": 12, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 430, + 411, + 523, + 422 + ], + "lines": [ + { + "bbox": [ + 430, + 411, + 523, + 422 + ], + "spans": [ + { + "bbox": [ + 430, + 411, + 523, + 422 + ], + "type": "text", + "content": "(b) Multi-modal Trajectory" + } + ] + } + ], + "index": 13, + "angle": 0, + "type": "image_caption" + } + ], + "index": 12 + }, + { + "type": "image", + "bbox": [ + 464, + 274, + 551, + 409 + ], + "blocks": [ + { + "bbox": [ + 464, + 274, + 551, + 409 + ], + "lines": [ + { + "bbox": [ + 464, + 274, + 551, + 409 + ], + "spans": [ + { + "bbox": [ + 464, + 274, + 551, + 409 + ], + "type": "image", + "image_path": "dc4ff25f585a9c2a8a9a67ed08cfb4ac1495c8745eb4c547abeb6457741bbdc4.jpg" + } + ] + } + ], + "index": 14, + "angle": 0, + "type": "image_body" + } + ], + "index": 14 + }, + { + "type": "image", + "bbox": [ + 318, + 429, + 391, + 557 + ], + "blocks": [ + { + "bbox": [ + 318, + 429, + 391, + 557 + ], + "lines": [ + { + "bbox": [ + 318, + 429, + 391, + 557 + ], + "spans": [ + { + "bbox": [ + 318, + 429, + 391, + 557 + ], + "type": "image", + "image_path": "d03e39cae0f6cbaeb15ee3f345b3e5e60680f2e7c34687159517be322f33550b.jpg" + } + ] + } + ], + "index": 15, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 381, + 566, + 492, + 576 + ], + "lines": [ + { + "bbox": [ + 381, + 566, + 492, + 576 + ], + "spans": [ + { + "bbox": [ + 381, + 566, + 492, + 576 + ], + "type": "text", + "content": "(c) Momentum Planning (Ours)" + } + ] + } + ], + "index": 18, + "angle": 0, + "type": "image_caption" + } + ], + "index": 15 + }, + { + "type": "image", + "bbox": [ + 392, + 429, + 476, + 557 + ], + "blocks": [ + { + "bbox": [ + 392, + 429, + 476, + 557 + ], + "lines": [ + { + "bbox": [ + 392, + 429, + 476, + 557 + ], + "spans": [ + { + "bbox": [ + 392, + 429, + 476, + 557 + ], + "type": "image", + "image_path": "cee37a50b1899f64c3cb4a28cf30f60ca14efe4d6183c5c741efd6d03d4a6396.jpg" + } + ] + } + ], + "index": 16, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 313, + 581, + 555, + 670 + ], + "lines": [ + { + "bbox": [ + 313, + 581, + 555, + 670 + ], + "spans": [ + { + "bbox": [ + 313, + 581, + 555, + 670 + ], + "type": "text", + "content": "Figure 1. (a) Deterministic Planning [14, 15, 19, 32] predicts deterministic trajectories, but lacks action diversity, posing safety risks. (b) Multi-modal Trajectory Planning [4, 27, 28] selects the highest-scoring trajectory among the multi-modal trajectories, yet fails to ensure stability and consistency, having risks in vehicle trembling. (c) Momentum Planning leverages the trajectory and perception momentum to enhance current planning through historical guidance to overcome temporal inconsistency." + } + ] + } + ], + "index": 19, + "angle": 0, + "type": "image_caption" + } + ], + "index": 16 + }, + { + "type": "image", + "bbox": [ + 480, + 429, + 553, + 557 + ], + "blocks": [ + { + "bbox": [ + 480, + 429, + 553, + 557 + ], + "lines": [ + { + "bbox": [ + 480, + 429, + 553, + 557 + ], + "spans": [ + { + "bbox": [ + 480, + 429, + 553, + 557 + ], + "type": "image", + "image_path": "f5ead49f269af7f21c98a269e9b431bc77482c0b52323ee03525bcb78a2b5eb4.jpg" + } + ] + } + ], + "index": 17, + "angle": 0, + "type": "image_body" + } + ], + "index": 17 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "spans": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "type": "text", + "content": "CVF" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 145, + 0, + 496, + 37 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 145, + 0, + 496, + 37 + ], + "spans": [ + { + "bbox": [ + 145, + 0, + 496, + 37 + ], + "type": "text", + "content": "This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 70, + 704, + 247, + 715 + ], + "type": "page_footnote", + "angle": 0, + "lines": [ + { + "bbox": [ + 70, + 704, + 247, + 715 + ], + "spans": [ + { + "bbox": [ + 70, + 704, + 247, + 715 + ], + "type": "text", + "content": "† Intern of Horizon Robotics, † Corresponding author." + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "type": "text", + "content": "22432" + } + ] + } + ], + "index": 21 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 0 + }, + { + "para_blocks": [ + { + "bbox": [ + 56, + 72, + 136, + 83 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 72, + 136, + 83 + ], + "spans": [ + { + "bbox": [ + 56, + 72, + 136, + 83 + ], + "type": "text", + "content": "1. Introduction" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 55, + 91, + 294, + 210 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 91, + 294, + 210 + ], + "spans": [ + { + "bbox": [ + 55, + 91, + 294, + 210 + ], + "type": "text", + "content": "Autonomous driving [3, 25, 29] has undergone a transformative shift from modular, manually crafted pipelines to a more integrated, end-to-end paradigm [15, 19, 24]. Unlike traditional approaches that handle tasks like detection, tracking, mapping, motion prediction, and planning in isolation, the end-to-end framework emphasizes seamless integration. By prioritizing planning, it strategically directs information from upstream perception modules, thereby enhancing robustness and reliability in dynamic driving environments." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 55, + 212, + 295, + 582 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 212, + 295, + 582 + ], + "spans": [ + { + "bbox": [ + 55, + 212, + 295, + 582 + ], + "type": "text", + "content": "Achieving high-quality planning in end-to-end frameworks hinges on accurately predicting the future trajectory prediction for the ego vehicle [1, 4, 7, 15, 19, 27]. Such future prediction requires a long-horizon understanding of both static and dynamic environmental factors, including map elements and interactions with surrounding agents. For instance, UniAD [15] queries the ego context from detailed bird's-eye-view (BEV) maps at each timestamp, while VAD [19] uses an ego query to retrieve surrounding context. The retrieved information then informs the planner, which predicts a deterministic trajectory for the vehicle, as illustrated in Figure 1 (a). Nevertheless, optimal trajectory prediction is inherently stochastic due to the unpredictability of other road users' intentions, varying road conditions, and the ambiguity introduced by human driving behaviors. This stochastic nature complicates the regression target, making deterministic predictions suboptimal and even risk-prone, potentially leading to severe collisions. To mitigate these uncertainties, methods such as VADv2 [4] and SparseDrive [27] leverage probabilistic modeling to capture the continuous planning action space, producing multi-modal trajectories that consider various possible behaviors of road agents, as shown in Figure 1 (b). While effective, these multi-modal approaches are typically one-shot and solely on the current perception frame. This limitation makes them susceptible to occlusion or loss of key visual cues, which can degrade multi-modal trajectory quality. Additionally, without temporal consistency, consecutive trajectories may lack coherence, causing unstable vehicle control and introducing undesirable directional shifts and oscillations." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 55, + 583, + 295, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 583, + 295, + 715 + ], + "spans": [ + { + "bbox": [ + 55, + 583, + 295, + 715 + ], + "type": "text", + "content": "To stabilize trajectory prediction, we draw inspiration from human driving behaviors and introduce the concept of momentum into autonomous driving. In physics, momentum reflects an object's tendency to maintain its velocity based on speed and direction [16]. Analogously, in driving, momentum captures the smooth, forward progression of movement informed by past trajectories and modulated by present conditions. As illustrated in Figure 1 (c), by explicitly integrating historical trajectories with current predictions, we aim to achieve smoother and more coherent planning outcomes. To this end, we propose an end-to-end" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 313, + 72, + 553, + 454 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 72, + 553, + 454 + ], + "spans": [ + { + "bbox": [ + 313, + 72, + 553, + 454 + ], + "type": "text", + "content": "Momentum-Aware Driving (MomAD) framework, which incorporates momentum awareness to deliver stable and responsive planning in driving scenarios. MomAD interprets momentum on two levels: (1) trajectory momentum: By aligning candidate multi-modal trajectory with prior predictions, abrupt shifts in ego vehicle's path can be minimized, ensuring consistent control and a more comfortable driving experience. (2) perception momentum: By aggregating historical context and attending to map elements and surroundings over time, the model broadens its perspective, capturing subtle agent intentions missed in single-frame observations. To implement these ideas, we introduce (1) Topological Trajectory Matching (TTM): we first use this module to minimize planning discrepancies across time steps by employing the Hausdorff Distance to identify multimodal trajectory proposals that best align with past planning results. This approach ensures temporal coherence by preventing excessive deviation from previous trajectories. (2) Momentum Planning Interactor (MPI): Since the selected trajectory may still be biased toward the current perception and sacrifice long-horizon considerations, we cross-attends the current best planning query with historical plan queries, which implicitly convey critical long-term ego-temporal, ego-agent, and ego-map information as key and value vectors. This interaction enriches the current query with long-horizon perception momentum, improving its context awareness. To enhance robustness against environmental noise and perception errors, we incorporate a Robust Instance Denoising Module in the perception stage. By introducing controlled perturbations during training, the model learns to denoise perception inputs, achieving resilience to dynamic changes and misdetections." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 313, + 456, + 553, + 622 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 456, + 553, + 622 + ], + "spans": [ + { + "bbox": [ + 313, + 456, + 553, + 622 + ], + "type": "text", + "content": "To evaluate the planning stability of MomAD, we propose a new Trajectory Prediction Consistency (TPC) metric to measure consistency between predicted and historical trajectories. Experiments demonstrate that MomAD can maintain long-term consistency (" + }, + { + "bbox": [ + 313, + 456, + 553, + 622 + ], + "type": "inline_equation", + "content": "\\geq 3" + }, + { + "bbox": [ + 313, + 456, + 553, + 622 + ], + "type": "text", + "content": " seconds). Given that most scenes in nuScenes involve straight roads, which limit the assessment of temporal inconsistency, we curated a Turning-nuScenes validation set from turning scenarios within the nuScenes dataset to provide a more challenging evaluation, where our approach outperforms state-of-the-art end-to-end frameworks. For example, our MomAD reduces the collision rate by " + }, + { + "bbox": [ + 313, + 456, + 553, + 622 + ], + "type": "inline_equation", + "content": "26\\%" + }, + { + "bbox": [ + 313, + 456, + 553, + 622 + ], + "type": "text", + "content": " and the TPC by " + }, + { + "bbox": [ + 313, + 456, + 553, + 622 + ], + "type": "inline_equation", + "content": "0.97\\mathrm{m}" + }, + { + "bbox": [ + 313, + 456, + 553, + 622 + ], + "type": "text", + "content": " compared to SparseDrive [27] for a 6-second horizon prediction in the Turning-nuScenes validation set." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 313, + 635, + 400, + 647 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 635, + 400, + 647 + ], + "spans": [ + { + "bbox": [ + 313, + 635, + 400, + 647 + ], + "type": "text", + "content": "2. Related Work" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 313, + 655, + 553, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 655, + 553, + 715 + ], + "spans": [ + { + "bbox": [ + 313, + 655, + 553, + 715 + ], + "type": "text", + "content": "End-to-end autonomous driving, which learns directly from raw sensor data to generate planning trajectories or driving commands, eliminates manual feature extraction [3, 25]. End-to-end autonomous driving methods [5, 6, 8-13, 15, 18-21, 23, 26, 28, 30-36] have garnered increased attent" + } + ] + } + ], + "index": 7 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 749, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 749, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 749, + 317, + 757 + ], + "type": "text", + "content": "22433" + } + ] + } + ], + "index": 8 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 1 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 56, + 68, + 549, + 194 + ], + "blocks": [ + { + "bbox": [ + 56, + 68, + 549, + 194 + ], + "lines": [ + { + "bbox": [ + 56, + 68, + 549, + 194 + ], + "spans": [ + { + "bbox": [ + 56, + 68, + 549, + 194 + ], + "type": "image", + "image_path": "991d8a2d24ec69326049bda428c951dab6e0a97aaf3df4dcb21ae72add0abdfa.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 54, + 200, + 555, + 257 + ], + "lines": [ + { + "bbox": [ + 54, + 200, + 555, + 257 + ], + "spans": [ + { + "bbox": [ + 54, + 200, + 555, + 257 + ], + "type": "text", + "content": "Figure 2. The overall architecture of MomAD. MomAD, as a multi-modal trajectory end-to-end autonomous driving method, first encodes multi-view images into feature maps, then learns a sparse scene representation through a robust instance denoising via perturbation module, and finally performs a momentum planning through Topological Trajectory Matching (TTM) module and Momentum Planning Interactor (MPI) module to accomplish planning tasks. Our approach addresses critical challenges of stability and robustness in dynamic driving conditions." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "type": "image", + "bbox": [ + 56, + 272, + 294, + 365 + ], + "blocks": [ + { + "bbox": [ + 56, + 272, + 294, + 365 + ], + "lines": [ + { + "bbox": [ + 56, + 272, + 294, + 365 + ], + "spans": [ + { + "bbox": [ + 56, + 272, + 294, + 365 + ], + "type": "image", + "image_path": "b3627d66e8e105144801867ff0063acb082dc4f467f283e9024241da1d0d4845.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 55, + 373, + 295, + 429 + ], + "lines": [ + { + "bbox": [ + 55, + 373, + 295, + 429 + ], + "spans": [ + { + "bbox": [ + 55, + 373, + 295, + 429 + ], + "type": "text", + "content": "Figure 3. The illustration of Momentum Planning Interactor (MPI). MPI cross-attends a selected planning query with historical queries to expand static and dynamic perception files, resulting in an enriched query that improves long-horizon trajectory generation and reduces collision risks." + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_caption" + } + ], + "index": 2 + }, + { + "bbox": [ + 55, + 453, + 295, + 716 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 453, + 295, + 716 + ], + "spans": [ + { + "bbox": [ + 55, + 453, + 295, + 716 + ], + "type": "text", + "content": "tion. UniAD [15] effectively integrates information from various preceding tasks, including perception, prediction, and planning modules to assist in trajectory planning, and achieves significant performance improvements. VAD [19] models driving scenarios as fully vectorized representations and employs explicit instance-level planning constraints to enhance planning safety. However, they [5, 11, 14, 15, 19] adopt a deterministic approach to trajectory prediction, which fails to account for trajectory diversity and may introduce risks due to intermediate regression results. VADv2 [4] proposes probabilistic planning to address the limitations of deterministic trajectory prediction, enabling multimodal trajectory predictions. Building upon the multimodal trajectory planning framework, SparseDrive [27] designs planning and motion prediction modules to achieve SOTA performance and efficiency on the nuScenes dataset. However, while multi-modal trajectory planning methods [4, 27, 28] have achieved SOTA performance, they overlook the temporal inconsistency caused by maximum score offsets. Regarding temporal inconsistency, existing methods only address temporal instance characteristics up to now, and they entirely overlook the issue of temporal planning" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 313, + 275, + 555, + 300 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 275, + 555, + 300 + ], + "spans": [ + { + "bbox": [ + 313, + 275, + 555, + 300 + ], + "type": "text", + "content": "consistency. In this work, we focus on this issue, aiming to address it using the concept of momentum planning." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 314, + 330, + 370, + 342 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 314, + 330, + 370, + 342 + ], + "spans": [ + { + "bbox": [ + 314, + 330, + 370, + 342 + ], + "type": "text", + "content": "3. Method" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 313, + 356, + 556, + 716 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 356, + 556, + 716 + ], + "spans": [ + { + "bbox": [ + 313, + 356, + 556, + 716 + ], + "type": "text", + "content": "Framework Overview. Figure 2 presents an overview of the proposed MomAD system, which integrates sparse perception and momentum-aware planning. To capture key dynamic and static instances interacting with the ego vehicle, the sparse perception module builds upon the SparseDrive [27] to encode multi-view image features, which are aggregated into instance features " + }, + { + "bbox": [ + 313, + 356, + 556, + 716 + ], + "type": "inline_equation", + "content": "\\mathbf{F}_t^{\\mathrm{ins}}" + }, + { + "bbox": [ + 313, + 356, + 556, + 716 + ], + "type": "text", + "content": " for road agents and map elements at time step " + }, + { + "bbox": [ + 313, + 356, + 556, + 716 + ], + "type": "inline_equation", + "content": "t" + }, + { + "bbox": [ + 313, + 356, + 556, + 716 + ], + "type": "text", + "content": ". These features are obtained by sampling keypoints around the anchor boxes and polylines, feeding into the detection/tracking and online mapping blocks for accurate predictions. The core of MomAD is the joint motion and momentum-aware planning module, which comprises two main components: (1) Topological Trajectory Matching (Sec 3.1), which explicitly selects the candidate trajectory that best matches the prior path among all multi-modal trajectories to ensure temporal coherence; and (2) Momentum Planning Interaction (Sec 3.2), which expands the perceptive field by cross-attending the selected candidate trajectory's planning query with queries from the previous time step in the long-horizon query mixer. This approach provides a broader view of the surrounding environment and the intentions of other agents. The refined query is then processed by the planning head to generate updated multi-modal trajectories. Since the planning module heavily relies on detection and map instance features, we introduce a robust instance denoising via perturbation module (Sec 3.3) within the sparse perception component during training. This ensures robustness by reducing sensitivity to noisy perception features, enhancing the stability of trajectory prediction and planning." + } + ] + } + ], + "index": 7 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "type": "text", + "content": "22434" + } + ] + } + ], + "index": 8 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 2 + }, + { + "para_blocks": [ + { + "bbox": [ + 55, + 72, + 267, + 85 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 72, + 267, + 85 + ], + "spans": [ + { + "bbox": [ + 55, + 72, + 267, + 85 + ], + "type": "text", + "content": "3.1. Topological Trajectory Matching (TTM)" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 55, + 89, + 296, + 220 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 89, + 296, + 220 + ], + "spans": [ + { + "bbox": [ + 55, + 89, + 296, + 220 + ], + "type": "text", + "content": "The proposed TTM module is inspired by the continuity of human driving behavior, where the optimal trajectories " + }, + { + "bbox": [ + 55, + 89, + 296, + 220 + ], + "type": "inline_equation", + "content": "T_{t}^{*}" + }, + { + "bbox": [ + 55, + 89, + 296, + 220 + ], + "type": "text", + "content": " are influenced by the historical path " + }, + { + "bbox": [ + 55, + 89, + 296, + 220 + ], + "type": "inline_equation", + "content": "T_{t-1}" + }, + { + "bbox": [ + 55, + 89, + 296, + 220 + ], + "type": "text", + "content": " to maintain temporal consistency and stability. Let " + }, + { + "bbox": [ + 55, + 89, + 296, + 220 + ], + "type": "inline_equation", + "content": "\\mathcal{T}_t = \\{T_t^k\\}_{k=1}^K" + }, + { + "bbox": [ + 55, + 89, + 296, + 220 + ], + "type": "text", + "content": " denote a set of " + }, + { + "bbox": [ + 55, + 89, + 296, + 220 + ], + "type": "inline_equation", + "content": "K" + }, + { + "bbox": [ + 55, + 89, + 296, + 220 + ], + "type": "text", + "content": " multi-modal candidate trajectories generated at time step " + }, + { + "bbox": [ + 55, + 89, + 296, + 220 + ], + "type": "inline_equation", + "content": "t" + }, + { + "bbox": [ + 55, + 89, + 296, + 220 + ], + "type": "text", + "content": ", where each trajectory " + }, + { + "bbox": [ + 55, + 89, + 296, + 220 + ], + "type": "inline_equation", + "content": "T_{t}^{k} = \\{(x_{t,i}^{k},y_{t,i}^{k})\\}_{i=1}^{N_{t}}" + }, + { + "bbox": [ + 55, + 89, + 296, + 220 + ], + "type": "text", + "content": " consists of " + }, + { + "bbox": [ + 55, + 89, + 296, + 220 + ], + "type": "inline_equation", + "content": "N_{t}" + }, + { + "bbox": [ + 55, + 89, + 296, + 220 + ], + "type": "text", + "content": " predicted waypoints. Typically, " + }, + { + "bbox": [ + 55, + 89, + 296, + 220 + ], + "type": "inline_equation", + "content": "K" + }, + { + "bbox": [ + 55, + 89, + 296, + 220 + ], + "type": "text", + "content": " is set to " + }, + { + "bbox": [ + 55, + 89, + 296, + 220 + ], + "type": "inline_equation", + "content": "6\\times 3" + }, + { + "bbox": [ + 55, + 89, + 296, + 220 + ], + "type": "text", + "content": " to account for six trajectory proposals for each of three possible commands (left, right, and straight), and " + }, + { + "bbox": [ + 55, + 89, + 296, + 220 + ], + "type": "inline_equation", + "content": "N_{t}" + }, + { + "bbox": [ + 55, + 89, + 296, + 220 + ], + "type": "text", + "content": " is chosen as 6 or 12 representing 0.5s interval for a 3- or 6-second prediction horizon." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 55, + 221, + 296, + 293 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 221, + 296, + 293 + ], + "spans": [ + { + "bbox": [ + 55, + 221, + 296, + 293 + ], + "type": "text", + "content": "Trajectory Coordinate Transformation. Since the historical and current predicted trajectories are generated in the ego vehicle's coordinate system at different moments, it is essential to transform them into a common coordinate system for accurate matching. This transformation from moment " + }, + { + "bbox": [ + 55, + 221, + 296, + 293 + ], + "type": "inline_equation", + "content": "t" + }, + { + "bbox": [ + 55, + 221, + 296, + 293 + ], + "type": "text", + "content": " to " + }, + { + "bbox": [ + 55, + 221, + 296, + 293 + ], + "type": "inline_equation", + "content": "t - 1" + }, + { + "bbox": [ + 55, + 221, + 296, + 293 + ], + "type": "text", + "content": " is achieved as follows:" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 97, + 298, + 295, + 313 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 97, + 298, + 295, + 313 + ], + "spans": [ + { + "bbox": [ + 97, + 298, + 295, + 313 + ], + "type": "interline_equation", + "content": "T _ {t} ^ {k} \\leftarrow \\mathbf {R} _ {t - 1} ^ {- 1} \\left(T _ {t} ^ {k} - \\mathbf {\\Gamma} _ {t - 1}\\right), \\forall k \\in [ K ], \\tag {1}", + "image_path": "40b9eeb4d79ffd240278a40aca681ce6ede14242f288621168ddb866835b4ade.jpg" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 55, + 318, + 295, + 341 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 318, + 295, + 341 + ], + "spans": [ + { + "bbox": [ + 55, + 318, + 295, + 341 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 55, + 318, + 295, + 341 + ], + "type": "inline_equation", + "content": "\\mathbf{R}_{t - 1}" + }, + { + "bbox": [ + 55, + 318, + 295, + 341 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 55, + 318, + 295, + 341 + ], + "type": "inline_equation", + "content": "\\Gamma_{t - 1}" + }, + { + "bbox": [ + 55, + 318, + 295, + 341 + ], + "type": "text", + "content": " denote the rotation and translation matrix, respectively." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 55, + 342, + 296, + 521 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 342, + 296, + 521 + ], + "spans": [ + { + "bbox": [ + 55, + 342, + 296, + 521 + ], + "type": "text", + "content": "Trajectory Distance Measurement. Simple Euclidean distance is inadequate for capturing the global alignment of trajectories, as it only measures pointwise proximity and is highly sensitive to local variations. This limitation becomes especially apparent in complex scenarios such as turns or varying point densities, where close points may not represent the alignment of the entire trajectory path. To address these limitations, TTM employs the Hausdorff distance as a more robust metric for evaluating trajectory alignment. The Hausdorff distance captures both local and global trajectory structures by measuring the maximum deviation between two sets of points, effectively quantifying the worst-case alignment between the candidate and historical trajectories. For each candidate trajectory " + }, + { + "bbox": [ + 55, + 342, + 296, + 521 + ], + "type": "inline_equation", + "content": "T_{t}^{k}" + }, + { + "bbox": [ + 55, + 342, + 296, + 521 + ], + "type": "text", + "content": ", the Hausdorff distance to the historical trajectory " + }, + { + "bbox": [ + 55, + 342, + 296, + 521 + ], + "type": "inline_equation", + "content": "T_{t-1}" + }, + { + "bbox": [ + 55, + 342, + 296, + 521 + ], + "type": "text", + "content": " is computed as," + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 63, + 526, + 294, + 574 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 63, + 526, + 294, + 574 + ], + "spans": [ + { + "bbox": [ + 63, + 526, + 294, + 574 + ], + "type": "interline_equation", + "content": "\\begin{array}{l} d_{1} = \\sup_{p\\in T^{k}_{t}}\\inf_{h\\in T_{t - 1}}\\| p - h\\| , d_{2} = \\sup_{h\\in T_{t - 1}}\\inf_{p\\in T^{k}_{t}}\\| h - p\\| , \\\\ d _ {H} \\left(T _ {t} ^ {k}, T _ {t - 1}\\right) = \\max \\left(d _ {1}, d _ {2}\\right), \\tag {2} \\\\ \\end{array}", + "image_path": "8dfe6676ed0968152df0c2239a2b47fababb8ebd4995a163846b7169e28f1c17.jpg" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 55, + 575, + 296, + 648 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 575, + 296, + 648 + ], + "spans": [ + { + "bbox": [ + 55, + 575, + 296, + 648 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 55, + 575, + 296, + 648 + ], + "type": "inline_equation", + "content": "p \\in T_t^k" + }, + { + "bbox": [ + 55, + 575, + 296, + 648 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 55, + 575, + 296, + 648 + ], + "type": "inline_equation", + "content": "h \\in T_{t - 1}" + }, + { + "bbox": [ + 55, + 575, + 296, + 648 + ], + "type": "text", + "content": " represent waypoints in the candidate and historical trajectories, respectively. The Hausdorff distance considers the furthest point discrepancies between trajectories in both directions, ensuring that even minor global misalignments are captured. TTM then selects the trajectory " + }, + { + "bbox": [ + 55, + 575, + 296, + 648 + ], + "type": "inline_equation", + "content": "T_{t}^{k^{*}}" + }, + { + "bbox": [ + 55, + 575, + 296, + 648 + ], + "type": "text", + "content": " that minimizes this distance:" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 116, + 652, + 295, + 674 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 116, + 652, + 295, + 674 + ], + "spans": [ + { + "bbox": [ + 116, + 652, + 295, + 674 + ], + "type": "interline_equation", + "content": "k ^ {*} = \\underset {k \\in [ K ]} {\\arg \\min } d _ {H} \\left(T _ {t} ^ {k}, T _ {t - 1}\\right) \\tag {3}", + "image_path": "cc584692104d79c584fff630b973a4f283cadfca7b73ec2b73f9e7b7226d6dd4.jpg" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 55, + 679, + 296, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 679, + 296, + 715 + ], + "spans": [ + { + "bbox": [ + 55, + 679, + 296, + 715 + ], + "type": "text", + "content": "This selection enforces continuity, aligning with historical driving patterns and providing stable trajectory predictions that are less prone to sudden shifts." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 314, + 72, + 518, + 84 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 314, + 72, + 518, + 84 + ], + "spans": [ + { + "bbox": [ + 314, + 72, + 518, + 84 + ], + "type": "text", + "content": "3.2. Momentum Planning Interactor (MPI)" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 313, + 91, + 555, + 261 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 91, + 555, + 261 + ], + "spans": [ + { + "bbox": [ + 313, + 91, + 555, + 261 + ], + "type": "text", + "content": "While TTM selects the most consistent trajectory " + }, + { + "bbox": [ + 313, + 91, + 555, + 261 + ], + "type": "inline_equation", + "content": "T_{t}^{k^{*}}" + }, + { + "bbox": [ + 313, + 91, + 555, + 261 + ], + "type": "text", + "content": " based on historical alignment, " + }, + { + "bbox": [ + 313, + 91, + 555, + 261 + ], + "type": "inline_equation", + "content": "T_{t}^{k^{*}}" + }, + { + "bbox": [ + 313, + 91, + 555, + 261 + ], + "type": "text", + "content": " is solely based on the current perception " + }, + { + "bbox": [ + 313, + 91, + 555, + 261 + ], + "type": "inline_equation", + "content": "\\mathbf{F}_{\\mathrm{ins}}" + }, + { + "bbox": [ + 313, + 91, + 555, + 261 + ], + "type": "text", + "content": ", which may lack a comprehensive view of the environment and be sensitive to occlusions. Therefore, the MPI module, as illustrated in Figure 3, incorporates a long-horizon query mixer to enrich the selected planning query " + }, + { + "bbox": [ + 313, + 91, + 555, + 261 + ], + "type": "inline_equation", + "content": "\\mathbf{Q}_t^{p^*}\\in \\mathbb{R}^{D_q}" + }, + { + "bbox": [ + 313, + 91, + 555, + 261 + ], + "type": "text", + "content": " of " + }, + { + "bbox": [ + 313, + 91, + 555, + 261 + ], + "type": "inline_equation", + "content": "T_{t}^{k^{*}}" + }, + { + "bbox": [ + 313, + 91, + 555, + 261 + ], + "type": "text", + "content": " with historical planning query " + }, + { + "bbox": [ + 313, + 91, + 555, + 261 + ], + "type": "inline_equation", + "content": "\\mathbf{Q}_{t - 1}^p\\in \\mathbb{R}^{K\\times D_q}" + }, + { + "bbox": [ + 313, + 91, + 555, + 261 + ], + "type": "text", + "content": " and the associated planning scores " + }, + { + "bbox": [ + 313, + 91, + 555, + 261 + ], + "type": "inline_equation", + "content": "\\mathbf{S}_{t - 1}^p\\in \\mathbb{R}^K" + }, + { + "bbox": [ + 313, + 91, + 555, + 261 + ], + "type": "text", + "content": ", implicitly capturing a broader understanding of the surrounding context and other agents' intentions over time. Here, " + }, + { + "bbox": [ + 313, + 91, + 555, + 261 + ], + "type": "inline_equation", + "content": "D_{q}" + }, + { + "bbox": [ + 313, + 91, + 555, + 261 + ], + "type": "text", + "content": " represents the latent dimension of planning queries. This enriched planning query " + }, + { + "bbox": [ + 313, + 91, + 555, + 261 + ], + "type": "inline_equation", + "content": "\\tilde{\\mathbf{Q}}_t^{p^*}\\in \\mathbb{R}^{D_q}" + }, + { + "bbox": [ + 313, + 91, + 555, + 261 + ], + "type": "text", + "content": " will be combined with instance features to re-generate an improved trajectory " + }, + { + "bbox": [ + 313, + 91, + 555, + 261 + ], + "type": "inline_equation", + "content": "\\tilde{\\mathcal{T}}_t\\in \\mathbb{R}^{K\\times N_t\\times 2}" + }, + { + "bbox": [ + 313, + 91, + 555, + 261 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 313, + 263, + 554, + 358 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 263, + 554, + 358 + ], + "spans": [ + { + "bbox": [ + 313, + 263, + 554, + 358 + ], + "type": "text", + "content": "Long-horizon Query Mixer. To achieve a robust perception of temporal momentum, the query mixer allows crossattention between the selected candidate trajectory's planning query with multi-modal planning queries from the previous time step. The historical planning queries " + }, + { + "bbox": [ + 313, + 263, + 554, + 358 + ], + "type": "inline_equation", + "content": "\\mathbf{Q}_{t-1}^p" + }, + { + "bbox": [ + 313, + 263, + 554, + 358 + ], + "type": "text", + "content": " and associated scores " + }, + { + "bbox": [ + 313, + 263, + 554, + 358 + ], + "type": "inline_equation", + "content": "\\mathbf{S}_{t-1}^p" + }, + { + "bbox": [ + 313, + 263, + 554, + 358 + ], + "type": "text", + "content": " are combined through element-wise interaction and processed with an LSTM to simulate temporal evolution:" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 340, + 371, + 555, + 388 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 340, + 371, + 555, + 388 + ], + "spans": [ + { + "bbox": [ + 340, + 371, + 555, + 388 + ], + "type": "interline_equation", + "content": "\\mathbf {Q} _ {t - 1} ^ {p ^ {\\prime}} = \\operatorname {L S T M} \\left(\\sigma \\left(\\mathbf {S} _ {t - 1} ^ {p}\\right) \\circ \\operatorname {M L P} \\left(\\mathbf {Q} _ {t - 1} ^ {p}\\right)\\right), \\tag {4}", + "image_path": "c2293e9c24aaf3281ae7feb764e9f3a7b161770a80478e9e7c42bdfbfc6dfd71.jpg" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 313, + 400, + 554, + 535 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 400, + 554, + 535 + ], + "spans": [ + { + "bbox": [ + 313, + 400, + 554, + 535 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 313, + 400, + 554, + 535 + ], + "type": "inline_equation", + "content": "\\sigma(\\cdot)" + }, + { + "bbox": [ + 313, + 400, + 554, + 535 + ], + "type": "text", + "content": " indicates the sigmoid function, MLP: " + }, + { + "bbox": [ + 313, + 400, + 554, + 535 + ], + "type": "inline_equation", + "content": "\\mathbb{R}^{D_q} \\mapsto \\mathbb{R}^{D_q}" + }, + { + "bbox": [ + 313, + 400, + 554, + 535 + ], + "type": "text", + "content": " the linear transform and " + }, + { + "bbox": [ + 313, + 400, + 554, + 535 + ], + "type": "inline_equation", + "content": "\\circ" + }, + { + "bbox": [ + 313, + 400, + 554, + 535 + ], + "type": "text", + "content": " the element-wise product. The LSTM processes this interaction, producing a surrogate multi-modal query " + }, + { + "bbox": [ + 313, + 400, + 554, + 535 + ], + "type": "inline_equation", + "content": "\\mathbf{Q}_{t-1}^{p'} \\in \\mathbb{R}^{K \\times D_q}" + }, + { + "bbox": [ + 313, + 400, + 554, + 535 + ], + "type": "text", + "content": " that captures the temporal evolution of planning queries. To aggregate historical information, the current planning query " + }, + { + "bbox": [ + 313, + 400, + 554, + 535 + ], + "type": "inline_equation", + "content": "\\mathbf{Q}_t^{p*}" + }, + { + "bbox": [ + 313, + 400, + 554, + 535 + ], + "type": "text", + "content": " is used as a query in a cross-attention module. The result " + }, + { + "bbox": [ + 313, + 400, + 554, + 535 + ], + "type": "inline_equation", + "content": "\\tilde{\\mathbf{Q}}_t^{p*}" + }, + { + "bbox": [ + 313, + 400, + 554, + 535 + ], + "type": "text", + "content": " incorporates long-term spatiotemporal context, which is further combined with the planning instance features and the encoded ego-anchor position information to inform the subsequent trajectory predictor:" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 354, + 548, + 553, + 580 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 354, + 548, + 553, + 580 + ], + "spans": [ + { + "bbox": [ + 354, + 548, + 553, + 580 + ], + "type": "interline_equation", + "content": "\\begin{array}{l} \\tilde {\\mathbf {Q}} _ {t} ^ {p *} = \\text {A t t e n t i o n} \\left(\\mathbf {Q} _ {t} ^ {p *}, \\mathbf {Q} _ {t - 1} ^ {p ^ {\\prime}}, \\mathbf {Q} _ {t - 1} ^ {p ^ {\\prime}}\\right). \\tag {5} \\\\ \\tilde {\\mathcal {T}} _ {t}, \\tilde {\\mathbf {S}} _ {t} = \\operatorname {P l a n H e a d} \\left(\\tilde {\\mathbf {Q}} _ {t} ^ {p *}, \\mathbf {F} _ {t} ^ {\\text {i n s}}\\right), \\\\ \\end{array}", + "image_path": "5c66525061c7a07916dfba67382f3525515ed698cbac95bed344fdfe0fffb86f.jpg" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 313, + 594, + 555, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 594, + 555, + 715 + ], + "spans": [ + { + "bbox": [ + 313, + 594, + 555, + 715 + ], + "type": "text", + "content": "The PlanHead module then generates refined " + }, + { + "bbox": [ + 313, + 594, + 555, + 715 + ], + "type": "inline_equation", + "content": "\\tilde{T}_t, \\tilde{\\mathbf{S}}_t" + }, + { + "bbox": [ + 313, + 594, + 555, + 715 + ], + "type": "text", + "content": ". The best trajectory " + }, + { + "bbox": [ + 313, + 594, + 555, + 715 + ], + "type": "inline_equation", + "content": "T_t^*" + }, + { + "bbox": [ + 313, + 594, + 555, + 715 + ], + "type": "text", + "content": " is then selected based on the highest scores among multi-modal outputs. Importantly, while the best trajectory is chosen based on the multi-modal trajectory scores, unlike previous selections, the current multimodal trajectories now fully consider the temporal consistency. This approach provides a stable, temporally-aware planning solution that is robust to occlusions and noise, significantly improving trajectory stability and control in complex driving environments." + } + ] + } + ], + "index": 16 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "text", + "content": "22435" + } + ] + } + ], + "index": 17 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 3 + }, + { + "para_blocks": [ + { + "bbox": [ + 55, + 72, + 284, + 85 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 72, + 284, + 85 + ], + "spans": [ + { + "bbox": [ + 55, + 72, + 284, + 85 + ], + "type": "text", + "content": "3.3. Robust Instance Denoising via Perturbation" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 55, + 90, + 297, + 307 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 90, + 297, + 307 + ], + "spans": [ + { + "bbox": [ + 55, + 90, + 297, + 307 + ], + "type": "text", + "content": "Our trajectory prediction and refinement rely heavily on the instance features " + }, + { + "bbox": [ + 55, + 90, + 297, + 307 + ], + "type": "inline_equation", + "content": "\\mathbf{F}_t^{\\mathrm{ins}}" + }, + { + "bbox": [ + 55, + 90, + 297, + 307 + ], + "type": "text", + "content": " of road agents and map elements provided by the sparse perception module. However, due to detector instability and the dynamically changing map, these instance features may be noisy, potentially introducing errors in downstream planning. To enhance the stability of planning against such noisy inputs, we introduce controlled noise perturbations during training and employ a lightweight encoder-decoder transformer block (see Figure 2) to learn effective denoising. This approach enables the model to distinguish between essential and extraneous features, reducing the impact of perception noise on trajectory predictions. During test-time inference, this denoising capability allows the trajectory predictor to be resilient to fluctuations in instance features. As a result, the model can produce smoother, more stable trajectories even in challenging scenarios with occlusions, temporary obstacles, or misdetections." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 55, + 318, + 138, + 331 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 318, + 138, + 331 + ], + "spans": [ + { + "bbox": [ + 55, + 318, + 138, + 331 + ], + "type": "text", + "content": "4. Experiments" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 55, + 337, + 173, + 351 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 337, + 173, + 351 + ], + "spans": [ + { + "bbox": [ + 55, + 337, + 173, + 351 + ], + "type": "text", + "content": "4.1. Experimental Setup" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 55, + 356, + 296, + 643 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 356, + 296, + 643 + ], + "spans": [ + { + "bbox": [ + 55, + 356, + 296, + 643 + ], + "type": "text", + "content": "Datasets. We conducted extensive experiments on the widely adopted nuScenes dataset [2] to evaluate tasks including detection, online mapping and planning in an open-loop setting. The nuScenes dataset comprises 1,000 driving scenes, with 700 and 150 sequences allocated for training and validation. Each scene spans around 20 seconds and contains roughly 40 key-frames annotated at " + }, + { + "bbox": [ + 55, + 356, + 296, + 643 + ], + "type": "inline_equation", + "content": "2\\mathrm{Hz}" + }, + { + "bbox": [ + 55, + 356, + 296, + 643 + ], + "type": "text", + "content": ", where each sample includes six images captured by surrounding cameras covering " + }, + { + "bbox": [ + 55, + 356, + 296, + 643 + ], + "type": "inline_equation", + "content": "360^{\\circ}" + }, + { + "bbox": [ + 55, + 356, + 296, + 643 + ], + "type": "text", + "content": " FOV horizontally and point clouds collected by both LiDAR and radar sensors. Since most planning tasks in nuScenes focus predominantly on gostraight commands, we curate a challenging subset of turning scenarios to form the Turning-nuScenes dataset, aimed at verifying the temporal consistency of predicted trajectories across time steps. Planning samples for turns are selected by setting the threshold between 3s and 0.5s of 'gt_ego_fut_trajs' to 25. The turning nuScenes validation dataset constitutes only one-tenth of the full nuScenes validation set, including 17 scenes with 680 samples. We use Bench2Drive [17], a closed-loop evaluation protocol under CARLA Leaderboard 2.0 for end-to-end autonomous driving. It provides an official training set, where we use the base set (1000 clips) for fair comparison with all the other baselines. We use the official 220 routes for evaluation." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 55, + 643, + 296, + 717 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 643, + 296, + 717 + ], + "spans": [ + { + "bbox": [ + 55, + 643, + 296, + 717 + ], + "type": "text", + "content": "Evaluation Metrics for Planning. For planning evaluation, we adopt the commonly used L2 Displacement Error (L2) and Collision Rate to assess planning performance. The calculation of the L2 error follows VAD [19] and the collision rate is aligned with SparseDrive [27]. However, the mainstream planning metrics cannot faithfully reveal the" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 313, + 72, + 555, + 168 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 72, + 555, + 168 + ], + "spans": [ + { + "bbox": [ + 313, + 72, + 555, + 168 + ], + "type": "text", + "content": "stability of predicted trajectories. Therefore, we introduce a novel metric, Trajectory Prediction Consistency (TPC), to measure the disparity between current predicted trajectories and historical predicted trajectories, allowing for a more comprehensive assessment of the consistency of trajectories. With coordinates transformed, the TPC between the current predicted trajectory " + }, + { + "bbox": [ + 313, + 72, + 555, + 168 + ], + "type": "inline_equation", + "content": "T_{t}^{\\mathrm{Pred}}" + }, + { + "bbox": [ + 313, + 72, + 555, + 168 + ], + "type": "text", + "content": " and historical one " + }, + { + "bbox": [ + 313, + 72, + 555, + 168 + ], + "type": "inline_equation", + "content": "T_{t-1}^{\\mathrm{Pred}}" + }, + { + "bbox": [ + 313, + 72, + 555, + 168 + ], + "type": "text", + "content": " is defined as," + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 326, + 177, + 555, + 215 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 326, + 177, + 555, + 215 + ], + "spans": [ + { + "bbox": [ + 326, + 177, + 555, + 215 + ], + "type": "interline_equation", + "content": "\\mathrm {T P C} = \\frac {1}{N _ {T}} \\sqrt {\\sum_ {i = 1} ^ {N} \\left(\\left(T _ {t} ^ {\\mathrm {P r e d}} - T _ {t - 1} ^ {\\mathrm {P r e d}}\\right) ^ {2} \\cdot T _ {\\mathrm {G T}} ^ {\\mathrm {M a s k}}\\right)}, \\tag {6}", + "image_path": "c372a951a805db569641795e5bb96cfef50cf121a8a5b9070d4c9afd7c984d71.jpg" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 313, + 223, + 556, + 320 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 223, + 556, + 320 + ], + "spans": [ + { + "bbox": [ + 313, + 223, + 556, + 320 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 313, + 223, + 556, + 320 + ], + "type": "inline_equation", + "content": "N_{T}" + }, + { + "bbox": [ + 313, + 223, + 556, + 320 + ], + "type": "text", + "content": " is the total number of GT trajectories in the validation set, and " + }, + { + "bbox": [ + 313, + 223, + 556, + 320 + ], + "type": "inline_equation", + "content": "T_{\\mathrm{GT}}^{\\mathrm{Mask}}" + }, + { + "bbox": [ + 313, + 223, + 556, + 320 + ], + "type": "text", + "content": " is the mask for trajectories exceeding the overlapped time period of two trajectories. Our TPC metric evaluates whether autonomous vehicles adhere to predicted trajectories, ensuring continuity across frames. Notably, the TPC metric provides a statistical perspective on the dataset-wide evaluation rather than at the individual sample level." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 313, + 327, + 400, + 338 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 327, + 400, + 338 + ], + "spans": [ + { + "bbox": [ + 313, + 327, + 400, + 338 + ], + "type": "text", + "content": "4.2. Main Results" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 313, + 345, + 556, + 418 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 345, + 556, + 418 + ], + "spans": [ + { + "bbox": [ + 313, + 345, + 556, + 418 + ], + "type": "text", + "content": "To conduct a comprehensive comparison, we have undertaken an exhaustive analysis of two distinct metrics presented in Table 1, which are derived from UniAD [15] and VAD [19]. It is noteworthy that, aside from Table 1, all other tables rely on the VAD [19] evaluation metrics for their assessments." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 313, + 433, + 421, + 445 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 433, + 421, + 445 + ], + "spans": [ + { + "bbox": [ + 313, + 433, + 421, + 445 + ], + "type": "text", + "content": "4.2.1 Planning Results" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 313, + 452, + 556, + 631 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 452, + 556, + 631 + ], + "spans": [ + { + "bbox": [ + 313, + 452, + 556, + 631 + ], + "type": "text", + "content": "NuScenes. As shown in Table 1, MomAD achieves an L2 error of " + }, + { + "bbox": [ + 313, + 452, + 556, + 631 + ], + "type": "inline_equation", + "content": "0.60\\mathrm{m}" + }, + { + "bbox": [ + 313, + 452, + 556, + 631 + ], + "type": "text", + "content": ", a collision rate of " + }, + { + "bbox": [ + 313, + 452, + 556, + 631 + ], + "type": "inline_equation", + "content": "0.09\\%" + }, + { + "bbox": [ + 313, + 452, + 556, + 631 + ], + "type": "text", + "content": ", and a TPC of " + }, + { + "bbox": [ + 313, + 452, + 556, + 631 + ], + "type": "inline_equation", + "content": "0.54\\mathrm{m}" + }, + { + "bbox": [ + 313, + 452, + 556, + 631 + ], + "type": "text", + "content": ", respectively. Compared to SOTAs like UniAD [15], VAD [19] and SparseDrive [27], our method shows SOTA performance in planning results. It is worth noting that we have made significant improvements in TPC, which directly proves our effectiveness in timing consistency. It is worth noting that we achieved significant improvements in TPC at " + }, + { + "bbox": [ + 313, + 452, + 556, + 631 + ], + "type": "inline_equation", + "content": "0.30\\mathrm{m}" + }, + { + "bbox": [ + 313, + 452, + 556, + 631 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 313, + 452, + 556, + 631 + ], + "type": "inline_equation", + "content": "0.53\\mathrm{m}" + }, + { + "bbox": [ + 313, + 452, + 556, + 631 + ], + "type": "text", + "content": ", and " + }, + { + "bbox": [ + 313, + 452, + 556, + 631 + ], + "type": "inline_equation", + "content": "0.78\\mathrm{m}" + }, + { + "bbox": [ + 313, + 452, + 556, + 631 + ], + "type": "text", + "content": " at 1s, 2s, and 3s on the nuScenes dataset, directly demonstrating our effectiveness in temporal consistency. Additionally, our MomAD is straightforward and achieves an FPS of 7.8, a slightly slower than SparseDrive. In summary, our MomAD effectively utilizes the smoothing advantage of Momentum and has a significant effect on improving temporal consistency." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 313, + 632, + 557, + 717 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 632, + 557, + 717 + ], + "spans": [ + { + "bbox": [ + 313, + 632, + 557, + 717 + ], + "type": "text", + "content": "Turning-nuScenes. As noted in [22], the nuScenes dataset features many straight routes, which limits the assessment of end-to-end methods. The simplicity of these paths can mask a model's true performance in complex scenarios. To address this, we evaluated our MomAD on the Turing-nuScenes validation set, as shown in Table 2. SparseDrive [27], a SOTA end-to-end method utilizing" + } + ] + } + ], + "index": 13 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "type": "text", + "content": "22436" + } + ] + } + ], + "index": 14 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 4 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 57, + 102, + 553, + 220 + ], + "blocks": [ + { + "bbox": [ + 55, + 71, + 555, + 95 + ], + "lines": [ + { + "bbox": [ + 55, + 71, + 555, + 95 + ], + "spans": [ + { + "bbox": [ + 55, + 71, + 555, + 95 + ], + "type": "text", + "content": "Table 1. Planning results on the nuScenes validation dataset. " + }, + { + "bbox": [ + 55, + 71, + 555, + 95 + ], + "type": "inline_equation", + "content": "{}^{ \\dagger }" + }, + { + "bbox": [ + 55, + 71, + 555, + 95 + ], + "type": "text", + "content": " denotes evaluation protocol used in UniAD [15]. * denotes results reproduced with the official checkpoint. As Ref. [22] states, we deactivate the ego status information for a fair comparison." + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 57, + 102, + 553, + 220 + ], + "lines": [ + { + "bbox": [ + 57, + 102, + 553, + 220 + ], + "spans": [ + { + "bbox": [ + 57, + 102, + 553, + 220 + ], + "type": "table", + "html": "
MethodInputBackboneL2 (m) ↓Col. Rate (%) ↓TPC (m) ↓FPS ↑
1s2s3sAvg.1s2s3sAvg.1s2s3sAvg.
UniAD† [15]CameraResNet1010.480.961.651.030.050.170.710.310.450.891.540.961.8 (A100)
VAD† [19]CameraResNet500.541.151.981.220.100.240.960.430.470.831.430.91-
SparseDrive†* [27]CameraResNet500.440.921.691.010.070.190.710.320.390.771.410.859.0 (RTX4090)
MomAD (Ours)†CameraResNet500.430.881.620.980.060.160.680.300.370.741.300.807.8 (RTX4090)
UniAD [15]CameraResNet1010.450.701.040.730.620.580.630.610.410.680.970.681.8 (A100)
VAD [19]CameraResNet500.410.701.050.720.030.190.430.210.360.660.910.64-
SparseDrive [27]CameraResNet500.290.580.960.610.010.050.180.080.300.570.850.579.0 (RTX4090)
MomAD (Ours)CameraResNet500.310.570.910.600.010.050.220.090.300.530.780.547.8 (RTX4090)
", + "image_path": "4ca87b80aa863416047f152abd276537fc9022697f8b8e0fbb9f47a80099a7a6.jpg" + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "table_body" + } + ], + "index": 1 + }, + { + "type": "table", + "bbox": [ + 58, + 255, + 553, + 311 + ], + "blocks": [ + { + "bbox": [ + 55, + 232, + 555, + 254 + ], + "lines": [ + { + "bbox": [ + 55, + 232, + 555, + 254 + ], + "spans": [ + { + "bbox": [ + 55, + 232, + 555, + 254 + ], + "type": "text", + "content": "Table 2. Planning results on the Turning-nuScenes validation dataset. SparseDrive [27] is a SOTA end-to-end multi-modal trajectory planning method. Red indicates improvement. We follow the VAD [19] evaluation metric." + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 58, + 255, + 553, + 311 + ], + "lines": [ + { + "bbox": [ + 58, + 255, + 553, + 311 + ], + "spans": [ + { + "bbox": [ + 58, + 255, + 553, + 311 + ], + "type": "table", + "html": "
MethodL2 (m) ↓Col. Rate (%) ↓TPC (m) ↓
1s2s3sAvg.1s2s3sAvg.1s2s3sAvg.
SparseDrive [27]0.350.771.460.860.040.170.980.400.340.701.330.79
MomAD (Ours)0.33-0.020.70-0.071.24-0.220.76-0.100.03-0.010.13-0.040.79-0.190.32-0.080.32-0.020.54-0.161.05-0.280.63-0.16
", + "image_path": "d817252fe32a968cd2a976eed7be1d8e30e2dda01d1f5d8a639d9adab8a4a826.jpg" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "table_body" + } + ], + "index": 3 + }, + { + "type": "table", + "bbox": [ + 58, + 376, + 293, + 448 + ], + "blocks": [ + { + "bbox": [ + 55, + 330, + 294, + 373 + ], + "lines": [ + { + "bbox": [ + 55, + 330, + 294, + 373 + ], + "spans": [ + { + "bbox": [ + 55, + 330, + 294, + 373 + ], + "type": "text", + "content": "Table 3. Long trajectory planning results on the nuScenes and Turning-nuScenes validation sets. We train models for 10 epochs for 6s-horizon prediction. T-nuScenes indicates the challenging Turning-nuScenes. We follow the VAD [19] evaluation metric." + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 58, + 376, + 293, + 448 + ], + "lines": [ + { + "bbox": [ + 58, + 376, + 293, + 448 + ], + "spans": [ + { + "bbox": [ + 58, + 376, + 293, + 448 + ], + "type": "table", + "html": "
SplitMethodL2 (m) ↓Col. Rate (%) ↓TPC (m) ↓
4s5s6s4s5s6s4s5s6s
nuScenesSparseDrive [27]1.752.322.950.871.542.331.331.661.99
MomAD1.671.982.450.831.432.131.191.451.61
-0.09-0.34-0.50-0.04-0.11-0.20-0.14-0.21-0.38
T-nuScenesSparseDrive [27]2.072.713.360.911.712.571.542.312.90
MomAD1.802.072.510.851.572.311.371.581.93
-0.27-0.64-0.85-0.06-0.14-0.26-0.17-0.73-0.97
", + "image_path": "033bcda7cb771bf3f8c31402a900e57a8559542b1e37cac4a3c143cd79fe4df4.jpg" + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "table_body" + } + ], + "index": 5 + }, + { + "type": "table", + "bbox": [ + 58, + 493, + 293, + 586 + ], + "blocks": [ + { + "bbox": [ + 55, + 460, + 295, + 492 + ], + "lines": [ + { + "bbox": [ + 55, + 460, + 295, + 492 + ], + "spans": [ + { + "bbox": [ + 55, + 460, + 295, + 492 + ], + "type": "text", + "content": "Table 4. Open-loop and Closed-loop results on Bench2Drive (V0.0.3) under base training set. 'mmt' refers multi-modal trajectory variant of VAD and * the re-implementation." + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 58, + 493, + 293, + 586 + ], + "lines": [ + { + "bbox": [ + 58, + 493, + 293, + 586 + ], + "spans": [ + { + "bbox": [ + 58, + 493, + 293, + 586 + ], + "type": "table", + "html": "
MethodOpen-loop MetricClosed-loop Metric
Avg. L2 ↓DS ↑ SR (%) ↑Effi ↑Comf ↑
VAD0.9142.3515.00157.9446.01
VADmmt*0.8942.8715.91158.1247.22
MomAD (Euclidean)0.8744.2216.91161.7748.70
MomAD0.8545.3517.44162.0949.34
SparseDrive*0.8744.5416.71170.2148.63
MomAD (Euclidean)0.8446.1217.45173.3550.98
MomAD0.8247.9118.11174.9151.20
", + "image_path": "2322b85e0cffa940dfc14d49428804c190f92d328d3189a51c98fe892bde7705.jpg" + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "table_body" + } + ], + "index": 7 + }, + { + "bbox": [ + 54, + 607, + 295, + 691 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 54, + 607, + 295, + 691 + ], + "spans": [ + { + "bbox": [ + 54, + 607, + 295, + 691 + ], + "type": "text", + "content": "multi-modal trajectories, performs well across various scenarios but struggles with driving stability during turns. In contrast, MomAD exhibits superior consistency in trajectory predictions, as indicated by the TPC metric. Overall, MomAD not only delivers effective trajectory predictions under standard conditions but also maintains reliability amid dynamic changes and complex environments." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 54, + 692, + 295, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 54, + 692, + 295, + 715 + ], + "spans": [ + { + "bbox": [ + 54, + 692, + 295, + 715 + ], + "type": "text", + "content": "Long Trajectory Prediction. Accurate long trajectory prediction is vital for enhancing the stability of autonomous" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 313, + 331, + 555, + 605 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 331, + 555, + 605 + ], + "spans": [ + { + "bbox": [ + 313, + 331, + 555, + 605 + ], + "type": "text", + "content": "driving, is useful to evaluate molels' ability to address temporal inconsistency issues in multi-modal trajectory planning. As shown in Table 3, we compared SparseDrive and MomAD in 4-6s long trajectory prediction on the nuScenes and Turning-nuScenes dataset, demonstrating a significant performance improvement. Specifically, in nuScenes dataset, compared with SparseDrive, MomAD experienced a decrease of " + }, + { + "bbox": [ + 313, + 331, + 555, + 605 + ], + "type": "inline_equation", + "content": "0.09\\mathrm{m}" + }, + { + "bbox": [ + 313, + 331, + 555, + 605 + ], + "type": "inline_equation", + "content": "(5.14\\%)" + }, + { + "bbox": [ + 313, + 331, + 555, + 605 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 313, + 331, + 555, + 605 + ], + "type": "inline_equation", + "content": "0.34\\mathrm{m}" + }, + { + "bbox": [ + 313, + 331, + 555, + 605 + ], + "type": "inline_equation", + "content": "(14.66\\%)" + }, + { + "bbox": [ + 313, + 331, + 555, + 605 + ], + "type": "text", + "content": ", and " + }, + { + "bbox": [ + 313, + 331, + 555, + 605 + ], + "type": "inline_equation", + "content": "0.50\\mathrm{m}" + }, + { + "bbox": [ + 313, + 331, + 555, + 605 + ], + "type": "inline_equation", + "content": "(16.95\\%)" + }, + { + "bbox": [ + 313, + 331, + 555, + 605 + ], + "type": "text", + "content": " in the 4s, 5s, and 6s of L2 error, a decrease of " + }, + { + "bbox": [ + 313, + 331, + 555, + 605 + ], + "type": "inline_equation", + "content": "0.04\\%" + }, + { + "bbox": [ + 313, + 331, + 555, + 605 + ], + "type": "inline_equation", + "content": "0.11\\%" + }, + { + "bbox": [ + 313, + 331, + 555, + 605 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 313, + 331, + 555, + 605 + ], + "type": "inline_equation", + "content": "0.20\\%" + }, + { + "bbox": [ + 313, + 331, + 555, + 605 + ], + "type": "text", + "content": " in the 4s, 5s, and 6s of collision rate, and a decrease of " + }, + { + "bbox": [ + 313, + 331, + 555, + 605 + ], + "type": "inline_equation", + "content": "0.14\\mathrm{m}" + }, + { + "bbox": [ + 313, + 331, + 555, + 605 + ], + "type": "inline_equation", + "content": "(10.53\\%)" + }, + { + "bbox": [ + 313, + 331, + 555, + 605 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 313, + 331, + 555, + 605 + ], + "type": "inline_equation", + "content": "0.21\\mathrm{m}" + }, + { + "bbox": [ + 313, + 331, + 555, + 605 + ], + "type": "inline_equation", + "content": "(12.65\\%)" + }, + { + "bbox": [ + 313, + 331, + 555, + 605 + ], + "type": "text", + "content": ", and " + }, + { + "bbox": [ + 313, + 331, + 555, + 605 + ], + "type": "inline_equation", + "content": "0.38\\mathrm{m}" + }, + { + "bbox": [ + 313, + 331, + 555, + 605 + ], + "type": "inline_equation", + "content": "(19.10\\%)" + }, + { + "bbox": [ + 313, + 331, + 555, + 605 + ], + "type": "text", + "content": " in the 4s, 5s, and 6s of TPC, respectively. Furthermore, in Turning-nuScenes dataset, compared with SparseDrive, MomAD experienced a decrease of " + }, + { + "bbox": [ + 313, + 331, + 555, + 605 + ], + "type": "inline_equation", + "content": "0.27\\mathrm{m}" + }, + { + "bbox": [ + 313, + 331, + 555, + 605 + ], + "type": "inline_equation", + "content": "(13.04\\%)" + }, + { + "bbox": [ + 313, + 331, + 555, + 605 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 313, + 331, + 555, + 605 + ], + "type": "inline_equation", + "content": "0.64\\mathrm{m}" + }, + { + "bbox": [ + 313, + 331, + 555, + 605 + ], + "type": "inline_equation", + "content": "(23.62\\%)" + }, + { + "bbox": [ + 313, + 331, + 555, + 605 + ], + "type": "text", + "content": ", and " + }, + { + "bbox": [ + 313, + 331, + 555, + 605 + ], + "type": "inline_equation", + "content": "0.85\\mathrm{m}" + }, + { + "bbox": [ + 313, + 331, + 555, + 605 + ], + "type": "inline_equation", + "content": "(25.30\\%)" + }, + { + "bbox": [ + 313, + 331, + 555, + 605 + ], + "type": "text", + "content": " in the 4s, 5s, and 6s of L2 error, a decrease of " + }, + { + "bbox": [ + 313, + 331, + 555, + 605 + ], + "type": "inline_equation", + "content": "0.06\\%" + }, + { + "bbox": [ + 313, + 331, + 555, + 605 + ], + "type": "inline_equation", + "content": "0.14\\%" + }, + { + "bbox": [ + 313, + 331, + 555, + 605 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 313, + 331, + 555, + 605 + ], + "type": "inline_equation", + "content": "0.26\\%" + }, + { + "bbox": [ + 313, + 331, + 555, + 605 + ], + "type": "text", + "content": " in the 4s, 5s, and 6s of collision rate, and a decrease of " + }, + { + "bbox": [ + 313, + 331, + 555, + 605 + ], + "type": "inline_equation", + "content": "0.17\\mathrm{m}" + }, + { + "bbox": [ + 313, + 331, + 555, + 605 + ], + "type": "inline_equation", + "content": "(11.04\\%)" + }, + { + "bbox": [ + 313, + 331, + 555, + 605 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 313, + 331, + 555, + 605 + ], + "type": "inline_equation", + "content": "0.73\\mathrm{m}" + }, + { + "bbox": [ + 313, + 331, + 555, + 605 + ], + "type": "inline_equation", + "content": "(31.60\\%)" + }, + { + "bbox": [ + 313, + 331, + 555, + 605 + ], + "type": "text", + "content": ", and " + }, + { + "bbox": [ + 313, + 331, + 555, + 605 + ], + "type": "inline_equation", + "content": "0.97\\mathrm{m}" + }, + { + "bbox": [ + 313, + 331, + 555, + 605 + ], + "type": "inline_equation", + "content": "(32.45\\%)" + }, + { + "bbox": [ + 313, + 331, + 555, + 605 + ], + "type": "text", + "content": " in the 4s, 5s, and 6s of TPC, respectively. In summary, our MomAD has improved the performance of long trajectory predictions, which further proves that MomAD can effectively alleviate the problem of temporal inconsistency." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 313, + 607, + 556, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 607, + 556, + 715 + ], + "spans": [ + { + "bbox": [ + 313, + 607, + 556, + 715 + ], + "type": "text", + "content": "Bench2Drive. We have included evaluations on the challenging closed-loop results on Bench2Drive dataset, as shown in Table 4, w which covers 44 interactive scenes (e.g., cut-ins, overtaking, detours) and 220 routes across diverse weather conditions and locations. Our MomAD improves success rate by " + }, + { + "bbox": [ + 313, + 607, + 556, + 715 + ], + "type": "inline_equation", + "content": "16.3\\%" + }, + { + "bbox": [ + 313, + 607, + 556, + 715 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 313, + 607, + 556, + 715 + ], + "type": "inline_equation", + "content": "8.4\\%" + }, + { + "bbox": [ + 313, + 607, + 556, + 715 + ], + "type": "text", + "content": " over the VAD multi-modal variant and SparseDrive, and enhances the Comfortness score (trajectory smoothness) by " + }, + { + "bbox": [ + 313, + 607, + 556, + 715 + ], + "type": "inline_equation", + "content": "7.2\\%" + }, + { + "bbox": [ + 313, + 607, + 556, + 715 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 313, + 607, + 556, + 715 + ], + "type": "inline_equation", + "content": "5.3\\%" + }, + { + "bbox": [ + 313, + 607, + 556, + 715 + ], + "type": "text", + "content": ", demonstrating its effectiveness." + } + ] + } + ], + "index": 11 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "type": "text", + "content": "22437" + } + ] + } + ], + "index": 12 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 5 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 56, + 101, + 553, + 159 + ], + "blocks": [ + { + "bbox": [ + 55, + 71, + 555, + 95 + ], + "lines": [ + { + "bbox": [ + 55, + 71, + 555, + 95 + ], + "spans": [ + { + "bbox": [ + 55, + 71, + 555, + 95 + ], + "type": "text", + "content": "Table 5. Perception and motion results on the nuScenes validation dataset. " + }, + { + "bbox": [ + 55, + 71, + 555, + 95 + ], + "type": "inline_equation", + "content": "{}^{ \\dagger }" + }, + { + "bbox": [ + 55, + 71, + 555, + 95 + ], + "type": "text", + "content": " indicates the results are reproduced with the official checkpoint. " + }, + { + "bbox": [ + 55, + 71, + 555, + 95 + ], + "type": "inline_equation", + "content": "{\\mathrm{{AP}}}_{d}" + }, + { + "bbox": [ + 55, + 71, + 555, + 95 + ], + "type": "text", + "content": " denotes " + }, + { + "bbox": [ + 55, + 71, + 555, + 95 + ], + "type": "inline_equation", + "content": "{\\mathrm{{AP}}}_{\\text{divider }}" + }, + { + "bbox": [ + 55, + 71, + 555, + 95 + ], + "type": "text", + "content": " . " + }, + { + "bbox": [ + 55, + 71, + 555, + 95 + ], + "type": "inline_equation", + "content": "{\\mathrm{{AP}}}_{b}" + }, + { + "bbox": [ + 55, + 71, + 555, + 95 + ], + "type": "text", + "content": " denotes " + }, + { + "bbox": [ + 55, + 71, + 555, + 95 + ], + "type": "inline_equation", + "content": "{\\mathrm{{AP}}}_{\\text{boundary }}" + }, + { + "bbox": [ + 55, + 71, + 555, + 95 + ], + "type": "text", + "content": " . mADE denotes minADE. mFDE denotes minFDE." + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 56, + 101, + 553, + 159 + ], + "lines": [ + { + "bbox": [ + 56, + 101, + 553, + 159 + ], + "spans": [ + { + "bbox": [ + 56, + 101, + 553, + 159 + ], + "type": "table", + "html": "
Method3D Object DetectionMulti-Object TrackingOnline MappingMotion Prediction
mAP ↑NDS ↑mATE ↓mASE ↓mAOE ↓mAVE ↓mAAE ↓AMOTA ↑AMOTP ↓Recall ↑IDS ↓mAP ↑APped ↑APd ↑APb ↑mADE ↓mFDE ↓MR ↓EPA ↑
UniAD [15]0.3800.4980.6840.2770.3830.3810.1920.3591.3200.467906----0.711.020.1510.456
VAD† [19]0.3120.4350.6100.2880.5410.5340.228----47.640.651.550.6----
SparseDrive [27]0.4180.5250.5660.2750.5520.2610.1900.3861.2540.49988655.149.957.058.40.620.990.1360.482
MomAD (Ours)0.4230.5310.5610.2690.5490.2580.1880.3911.2430.50985355.950.758.158.90.610.980.1370.499
", + "image_path": "a0cd90cb65f95101f3fc1cc6961a1c5aba8462fb63a73a2969357a93850998c0.jpg" + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "table_body" + } + ], + "index": 1 + }, + { + "type": "image", + "bbox": [ + 58, + 173, + 303, + 277 + ], + "blocks": [ + { + "bbox": [ + 58, + 173, + 303, + 277 + ], + "lines": [ + { + "bbox": [ + 58, + 173, + 303, + 277 + ], + "spans": [ + { + "bbox": [ + 58, + 173, + 303, + 277 + ], + "type": "image", + "image_path": "cd0c23434c044dc7720c52132f7566fec1c2ebe9ef23195e91eb55d45f9d0d48.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + } + ], + "index": 2 + }, + { + "type": "image", + "bbox": [ + 307, + 172, + 551, + 277 + ], + "blocks": [ + { + "bbox": [ + 307, + 172, + 551, + 277 + ], + "lines": [ + { + "bbox": [ + 307, + 172, + 551, + 277 + ], + "spans": [ + { + "bbox": [ + 307, + 172, + 551, + 277 + ], + "type": "image", + "image_path": "19ae4902ce112b004de94d825a4617078236757c492b697d43361caea240d3b6.jpg" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_body" + } + ], + "index": 3 + }, + { + "type": "image", + "bbox": [ + 58, + 277, + 302, + 388 + ], + "blocks": [ + { + "bbox": [ + 58, + 277, + 302, + 388 + ], + "lines": [ + { + "bbox": [ + 58, + 277, + 302, + 388 + ], + "spans": [ + { + "bbox": [ + 58, + 277, + 302, + 388 + ], + "type": "image", + "image_path": "3d9435429872b91eeddacefa255e145856a828e73a19607f3603cd236a93a864.jpg" + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 55, + 399, + 555, + 422 + ], + "lines": [ + { + "bbox": [ + 55, + 399, + 555, + 422 + ], + "spans": [ + { + "bbox": [ + 55, + 399, + 555, + 422 + ], + "type": "text", + "content": "Figure 4. Visualization results of MomAD compared with UniAD, VAD and SparseDrive across multiple frames. The proposed MomAD achieves temporal consistency whichever from the predicted trajectory compared with ground truth (GT) or from the TPC metric." + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "image_caption" + } + ], + "index": 4 + }, + { + "type": "image", + "bbox": [ + 306, + 278, + 551, + 388 + ], + "blocks": [ + { + "bbox": [ + 306, + 278, + 551, + 388 + ], + "lines": [ + { + "bbox": [ + 306, + 278, + 551, + 388 + ], + "spans": [ + { + "bbox": [ + 306, + 278, + 551, + 388 + ], + "type": "image", + "image_path": "bf70935d41cd2902af5a66c001957d649c65c9d7185e6c23096be6dbb5251b91.jpg" + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "image_body" + } + ], + "index": 5 + }, + { + "type": "table", + "bbox": [ + 58, + 516, + 294, + 611 + ], + "blocks": [ + { + "bbox": [ + 55, + 441, + 295, + 506 + ], + "lines": [ + { + "bbox": [ + 55, + 441, + 295, + 506 + ], + "spans": [ + { + "bbox": [ + 55, + 441, + 295, + 506 + ], + "type": "text", + "content": "Table 6. Ablation studies of the sparse perception module in MomAD on the nuScenes validation split. The Encoder and Denoise " + }, + { + "bbox": [ + 55, + 441, + 295, + 506 + ], + "type": "inline_equation", + "content": "(\\sigma)" + }, + { + "bbox": [ + 55, + 441, + 295, + 506 + ], + "type": "text", + "content": " module is denoted as ED. MP represents Momentum planning. NS is the Gaussian noise factor controlling the noise level. Noise is applied during training only. We follow the VAD [19] evaluation metric." + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 58, + 516, + 294, + 611 + ], + "lines": [ + { + "bbox": [ + 58, + 516, + 294, + 611 + ], + "spans": [ + { + "bbox": [ + 58, + 516, + 294, + 611 + ], + "type": "table", + "html": "
ED MP NSDetectionTrackingOnline MappingMotionPlanning (Avg.)
mAP ↑NDS ↑AMOTA ↑mAP ↑APped ↑mADE ↓L2 ↓Col. ↓TPC ↓
✓0.00.4070.5210.38155.049.30.630.620.140.56
0.00.4050.5200.38055.149.50.630.610.130.55
✓0.10.4200.5300.39055.850.50.580.610.120.55
✓0.00.4170.5280.38655.450.60.630.610.110.55
✓0.050.4210.5290.38855.650.80.620.610.110.54
✓0.10.4230.5310.39155.950.70.610.600.090.54
✓0.20.4180.5200.38854.449.20.630.620.180.58
✓0.30.4120.5180.38354.048.80.650.640.220.61
", + "image_path": "6ac047b3b897c5a0d6d659989fe5e8249619a2d38793508c5bfb133bc639fdf9.jpg" + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "table_body" + } + ], + "index": 8 + }, + { + "bbox": [ + 55, + 634, + 267, + 646 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 634, + 267, + 646 + ], + "spans": [ + { + "bbox": [ + 55, + 634, + 267, + 646 + ], + "type": "text", + "content": "4.2.2 Perception and Motion Prediction Results" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 55, + 655, + 296, + 717 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 655, + 296, + 717 + ], + "spans": [ + { + "bbox": [ + 55, + 655, + 296, + 717 + ], + "type": "text", + "content": "Sparse representation is efficient but suffers from instability issues caused by the variability of instance features. To address these issues, we have enhanced the instance features using the Encoder and Denoise " + }, + { + "bbox": [ + 55, + 655, + 296, + 717 + ], + "type": "inline_equation", + "content": "(\\sigma)" + }, + { + "bbox": [ + 55, + 655, + 296, + 717 + ], + "type": "text", + "content": " module (denoted as ED) within the sparse perception framework, ensuring" + } + ] + } + ], + "index": 10 + }, + { + "type": "table", + "bbox": [ + 317, + 516, + 553, + 603 + ], + "blocks": [ + { + "bbox": [ + 313, + 441, + 555, + 506 + ], + "lines": [ + { + "bbox": [ + 313, + 441, + 555, + 506 + ], + "spans": [ + { + "bbox": [ + 313, + 441, + 555, + 506 + ], + "type": "text", + "content": "Table 7. Impact of history frames in MomAD on the Turning-nuScenes validation set. " + }, + { + "bbox": [ + 313, + 441, + 555, + 506 + ], + "type": "inline_equation", + "content": "t" + }, + { + "bbox": [ + 313, + 441, + 555, + 506 + ], + "type": "text", + "content": " denotes the frame number, where 1 indicates the history is empty (represented by 0), 2 signifies that the historical result corresponds to the previous 1 frame, and 3 indicates that the historical result pertains to the previous 2 frames. We follow the VAD [19] evaluation metric." + } + ] + } + ], + "index": 11, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 317, + 516, + 553, + 603 + ], + "lines": [ + { + "bbox": [ + 317, + 516, + 553, + 603 + ], + "spans": [ + { + "bbox": [ + 317, + 516, + 553, + 603 + ], + "type": "table", + "html": "
EDMPNStL2(m) ↓Col. Rate(%) ↓TPC(m) ↓
2s3sAvg.2s3sAvg.2s3sAvg.
0.771.460.860.170.980.400.701.330.79
✓20.711.270.770.140.830.340.551.070.65
✓0.10.771.450.860.180.970.390.691.330.79
✓✓0.110.781.480.880.180.980.390.701.350.81
✓✓0.120.701.240.760.130.790.320.541.050.63
✓✓0.130.721.270.780.140.840.350.561.090.66
", + "image_path": "be52047a852a17da4e2b63128535f684738fa143d85818932199e74d569b75b6.jpg" + } + ] + } + ], + "index": 12, + "angle": 0, + "type": "table_body" + } + ], + "index": 12 + }, + { + "bbox": [ + 313, + 632, + 556, + 717 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 632, + 556, + 717 + ], + "spans": [ + { + "bbox": [ + 313, + 632, + 556, + 717 + ], + "type": "text", + "content": "end-to-end stability for autonomous driving. As shown in Table 5, our MomAD perception module includes 3D object detection, multi-object tracking, and online mapping tasks. For 3D object detection, MomAD achieves " + }, + { + "bbox": [ + 313, + 632, + 556, + 717 + ], + "type": "inline_equation", + "content": "42.3\\%" + }, + { + "bbox": [ + 313, + 632, + 556, + 717 + ], + "type": "text", + "content": " mAP and " + }, + { + "bbox": [ + 313, + 632, + 556, + 717 + ], + "type": "inline_equation", + "content": "53.1\\%" + }, + { + "bbox": [ + 313, + 632, + 556, + 717 + ], + "type": "text", + "content": " NPS, improving the mAP by " + }, + { + "bbox": [ + 313, + 632, + 556, + 717 + ], + "type": "inline_equation", + "content": "0.5\\%" + }, + { + "bbox": [ + 313, + 632, + 556, + 717 + ], + "type": "text", + "content": " and the NDS by " + }, + { + "bbox": [ + 313, + 632, + 556, + 717 + ], + "type": "inline_equation", + "content": "0.6\\%" + }, + { + "bbox": [ + 313, + 632, + 556, + 717 + ], + "type": "text", + "content": " compared to the baseline SparseDrive [27]. For multi-object tracking, MomAD achieves an AMOTA of" + } + ] + } + ], + "index": 13 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 749, + 318, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 749, + 318, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 749, + 318, + 757 + ], + "type": "text", + "content": "22438" + } + ] + } + ], + "index": 14 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 6 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 57, + 145, + 294, + 203 + ], + "blocks": [ + { + "bbox": [ + 55, + 70, + 294, + 137 + ], + "lines": [ + { + "bbox": [ + 55, + 70, + 294, + 137 + ], + "spans": [ + { + "bbox": [ + 55, + 70, + 294, + 137 + ], + "type": "text", + "content": "Table 8. Ablation studies of the impact of the different modules in MP on the Turning-nuScenes validation dataset. QM denotes Long-horizon Query Mixer, TP denotes Trajectory Predictor. Add refers to the addition operation between the historical planning query " + }, + { + "bbox": [ + 55, + 70, + 294, + 137 + ], + "type": "inline_equation", + "content": "\\mathbf{Q}_{t-1}^p" + }, + { + "bbox": [ + 55, + 70, + 294, + 137 + ], + "type": "text", + "content": " and the selected planning query " + }, + { + "bbox": [ + 55, + 70, + 294, + 137 + ], + "type": "inline_equation", + "content": "\\mathbf{Q}_t^{p*}" + }, + { + "bbox": [ + 55, + 70, + 294, + 137 + ], + "type": "text", + "content": ". We follow the VAD [19] evaluation metric." + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 57, + 145, + 294, + 203 + ], + "lines": [ + { + "bbox": [ + 57, + 145, + 294, + 203 + ], + "spans": [ + { + "bbox": [ + 57, + 145, + 294, + 203 + ], + "type": "table", + "html": "
QMAddTPL2 (m) ↓Col. Rate (%) ↓TPC (m) ↓
2s3sAvg.2s3sAvg.2s3sAvg.
0.771.460.860.170.980.400.701.330.79
✓✓0.761.380.820.140.880.360.621.210.67
✓✓0.701.240.760.130.790.320.541.050.63
", + "image_path": "e98d4be004fb33e5d5d2f2c5837c14e2e74d9d3a218e04cc27816a08db1494c2.jpg" + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "table_body" + } + ], + "index": 1 + }, + { + "bbox": [ + 55, + 223, + 294, + 331 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 223, + 294, + 331 + ], + "spans": [ + { + "bbox": [ + 55, + 223, + 294, + 331 + ], + "type": "text", + "content": "39.1%, surpassing the baseline SparseDrive by 0.5%. For online mapping, compared to 55.1% mAP for the baseline SparseDrive, our MomAD achieves 55.9% mAP, improving the mAP by 0.8%. For motion prediction, our MomAD outperforms SparseDrive [27] and UniAD [15], achieves better motion prediction performance by considering the influence of an ego vehicle on other agents. In detail, Our MomAD achieves a 0.61m minADE, 0.98 minFDE, and 13.7% MR, and 0.499 EPA, respectively." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 55, + 338, + 149, + 350 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 338, + 149, + 350 + ], + "spans": [ + { + "bbox": [ + 55, + 338, + 149, + 350 + ], + "type": "text", + "content": "4.3. Ablation Study" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 55, + 356, + 295, + 499 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 356, + 295, + 499 + ], + "spans": [ + { + "bbox": [ + 55, + 356, + 295, + 499 + ], + "type": "text", + "content": "Roles of 'ED' module in Sparse Perception. Sparse representation end-to-end methods yield efficient computation but unstable metrics. Further details are available in the Appendix. As shown in Table 6, the Encoder and Denoise " + }, + { + "bbox": [ + 55, + 356, + 295, + 499 + ], + "type": "inline_equation", + "content": "(\\sigma)" + }, + { + "bbox": [ + 55, + 356, + 295, + 499 + ], + "type": "text", + "content": " module (ED) within sparse perception enhances the instance features, significantly impacts the overall pipeline. By introducing Gaussian noise and employing techniques, the robustness of instance features is improved, particularly when training with Noisy at 0.1. Our findings suggest that a controlled level of noise can enhance the end-to-end capabilities of sparse methods during training, offering insights for the community on sparse end-to-end methods." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 55, + 499, + 295, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 499, + 295, + 715 + ], + "spans": [ + { + "bbox": [ + 55, + 499, + 295, + 715 + ], + "type": "text", + "content": "Roles of 'MP' module in Planing. As Li et al. [22] have stated, most end-to-end autonomous driving methods perform poorly in turning scenarios. As shown in Table 7, to better evaluate the planning performance of end-to-end methods in turning scenarios, our MomAD is evaluated on the Turning-nuScenes validation dataset rather than only on the full nuScenes validation dataset. Specifically, under the premise of executing 'ED', at " + }, + { + "bbox": [ + 55, + 499, + 295, + 715 + ], + "type": "inline_equation", + "content": "t = 1" + }, + { + "bbox": [ + 55, + 499, + 295, + 715 + ], + "type": "text", + "content": ", providing a 0 value to the MP module does not improve performance. We have tried to change the MLP operation of the planning to a more complex operation, but it does not enhance the results. However, when " + }, + { + "bbox": [ + 55, + 499, + 295, + 715 + ], + "type": "inline_equation", + "content": "t = 2" + }, + { + "bbox": [ + 55, + 499, + 295, + 715 + ], + "type": "text", + "content": ", historical queries and results are used, L2 (Avg) reaches 0.76m, Col. (Avg) reaches 0.32 %, and TPC reaches 0.63m, which represents a significant improvement. In addition, when " + }, + { + "bbox": [ + 55, + 499, + 295, + 715 + ], + "type": "inline_equation", + "content": "t = 3" + }, + { + "bbox": [ + 55, + 499, + 295, + 715 + ], + "type": "text", + "content": ", more frames are fused, the improvement has actually decreased, which may be due to the uncertainty introduced by the departure of historical features, but there is still an overall improvement." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 313, + 72, + 555, + 216 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 72, + 555, + 216 + ], + "spans": [ + { + "bbox": [ + 313, + 72, + 555, + 216 + ], + "type": "text", + "content": "Roles of different sub-modules in 'MP' module. As shown in Table 8, we conducted an in-depth analysis of the internal mechanism of momentum planning. We found that simply using the native 'Add' operation to regenerate the planning results can achieve a good improvement, with L2 (Avg.), Col. (Avg.), and TPC (Avg.) decreasing by " + }, + { + "bbox": [ + 313, + 72, + 555, + 216 + ], + "type": "inline_equation", + "content": "0.04\\mathrm{m}" + }, + { + "bbox": [ + 313, + 72, + 555, + 216 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 313, + 72, + 555, + 216 + ], + "type": "inline_equation", + "content": "0.04\\%" + }, + { + "bbox": [ + 313, + 72, + 555, + 216 + ], + "type": "text", + "content": ", and " + }, + { + "bbox": [ + 313, + 72, + 555, + 216 + ], + "type": "inline_equation", + "content": "0.12\\mathrm{m}" + }, + { + "bbox": [ + 313, + 72, + 555, + 216 + ], + "type": "text", + "content": ", respectively. However, the 'Add' operation alone does not fully utilize historical features. Our Long-horizon Query Mixer has achieved the current optimal performance. Overall, historical results are very important for current outcomes, and their reasonable utilization can maximize the performance of end-to-end planning." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 313, + 228, + 398, + 239 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 228, + 398, + 239 + ], + "spans": [ + { + "bbox": [ + 313, + 228, + 398, + 239 + ], + "type": "text", + "content": "4.4. Visualization" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 313, + 247, + 555, + 392 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 247, + 555, + 392 + ], + "spans": [ + { + "bbox": [ + 313, + 247, + 555, + 392 + ], + "type": "text", + "content": "As shown in Figure 4, we showcase multi-frame qualitative comparisons of end-to-end solutions, including UniAD [15], VAD [19], SparseDrive [27], and the proposed MomAD. In a representative turning scenario, the MomAD approach demonstrates superior long-term awareness of surrounding vehicles, reducing the likelihood of collisions. Additionally, it generates smoother ego-vehicle trajectories (shown in yellow and blue) that closely align with the ground-truth trajectory (in red). This highlights its strong temporal consistency and lower TPC scores. Additional visualizations for various driving commands are provided in the Appendix." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 313, + 406, + 479, + 418 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 406, + 479, + 418 + ], + "spans": [ + { + "bbox": [ + 313, + 406, + 479, + 418 + ], + "type": "text", + "content": "5. Conclusion and Future Work" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 313, + 427, + 556, + 632 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 427, + 556, + 632 + ], + "spans": [ + { + "bbox": [ + 313, + 427, + 556, + 632 + ], + "type": "text", + "content": "The proposed MomAD framework addresses key challenges in planning stability and robustness for end-to-end autonomous driving. By leveraging trajectory momentum and perception momentum, MomAD stabilizes trajectory predictions through Topological Trajectory Matching (TTM) and Momentum Planning Interactor (MPI), ensuring temporal coherence and enriching long-horizon context. Evaluations on nuScenes, the curated Turning-nuScenes validation set, and Bench2Drive demonstrate its superior performance in reducing collision rates and improving trajectory consistency compared to SOTA methods. While MomAD improves temporal consistency in long-horizon trajectory prediction, a gap remains due to mode collapse induced by the standard teacher-forcing approach to trajectory regression, limiting trajectory diversity. Future work will explore diffusion models and speculative decoding to enhance trajectory diversity while balancing efficiency." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 313, + 646, + 417, + 659 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 646, + 417, + 659 + ], + "spans": [ + { + "bbox": [ + 313, + 646, + 417, + 659 + ], + "type": "text", + "content": "Acknowledgements" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 313, + 667, + 555, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 667, + 555, + 715 + ], + "spans": [ + { + "bbox": [ + 313, + 667, + 555, + 715 + ], + "type": "text", + "content": "We appreciate the valuable discussions with Wenchao Sun, Bo Jiang, and Bencheng Liao from Horizon Robotics. This work was supported by the National Key R&D Program of China (2018AAA0100302)." + } + ] + } + ], + "index": 12 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 749, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 749, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 749, + 317, + 757 + ], + "type": "text", + "content": "22439" + } + ] + } + ], + "index": 13 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 7 + }, + { + "para_blocks": [ + { + "bbox": [ + 56, + 71, + 115, + 83 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 71, + 115, + 83 + ], + "spans": [ + { + "bbox": [ + 56, + 71, + 115, + 83 + ], + "type": "text", + "content": "References" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 57, + 91, + 296, + 715 + ], + "type": "list", + "angle": 0, + "index": 13, + "blocks": [ + { + "bbox": [ + 61, + 91, + 296, + 133 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 61, + 91, + 296, + 133 + ], + "spans": [ + { + "bbox": [ + 61, + 91, + 296, + 133 + ], + "type": "text", + "content": "[1] Szilard Aradi. Survey of deep reinforcement learning for motion planning of autonomous vehicles. IEEE Transactions on Intelligent Transportation Systems, 23(2):740-759, 2020. 2" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 61, + 137, + 296, + 201 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 61, + 137, + 296, + 201 + ], + "spans": [ + { + "bbox": [ + 61, + 137, + 296, + 201 + ], + "type": "text", + "content": "[2] Holger Caesar, Varun Bankiti, Alex H Lang, Sourabh Vora, Venice Erin Liong, Qiang Xu, Anush Krishnan, Yu Pan, Giancarlo Baldan, and Oscar Beijbom. nuscenes: A multimodal dataset for autonomous driving. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 11621-11631, 2020. 5" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 62, + 203, + 295, + 246 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 203, + 295, + 246 + ], + "spans": [ + { + "bbox": [ + 62, + 203, + 295, + 246 + ], + "type": "text", + "content": "[3] Li Chen, Penghao Wu, Kashyap Chitta, Bernhard Jaeger, Andreas Geiger, and Hongyang Li. End-to-end autonomous driving: Challenges and frontiers. IEEE Transactions on Pattern Analysis and Machine Intelligence, 2024. 2" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 62, + 247, + 295, + 300 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 247, + 295, + 300 + ], + "spans": [ + { + "bbox": [ + 62, + 247, + 295, + 300 + ], + "type": "text", + "content": "[4] Shaoyu Chen, Bo Jiang, Hao Gao, Bencheng Liao, Qing Xu, Qian Zhang, Chang Huang, Wenyu Liu, and Xinggang Wang. Vadv2: End-to-end vectorized autonomous driving via probabilistic planning. arXiv preprint arXiv:2402.13243, 2024. 1, 2, 3" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 62, + 303, + 295, + 356 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 303, + 295, + 356 + ], + "spans": [ + { + "bbox": [ + 62, + 303, + 295, + 356 + ], + "type": "text", + "content": "[5] Zhili Chen, Maosheng Ye, Shuangjie Xu, Tongyi Cao, and Qifeng Chen. Ppad: Iterative interactions of prediction and planning for end-to-end autonomous driving. In European Conference on Computer Vision, pages 239-256. Springer, 2025. 2, 3" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 62, + 359, + 295, + 412 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 359, + 295, + 412 + ], + "spans": [ + { + "bbox": [ + 62, + 359, + 295, + 412 + ], + "type": "text", + "content": "[6] Jie Cheng, Yingbing Chen, Xiaodong Mei, Bowen Yang, Bo Li, and Ming Liu. Rethinking imitation-based planners for autonomous driving. In 2024 IEEE International Conference on Robotics and Automation (ICRA), pages 14123-14130. IEEE, 2024. 2" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 62, + 415, + 295, + 458 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 415, + 295, + 458 + ], + "spans": [ + { + "bbox": [ + 62, + 415, + 295, + 458 + ], + "type": "text", + "content": "[7] Laurene Claussmann, Marc Revilloud, Dominique Gruyer, and Sébastien Glaser. A review of motion planning for highway autonomous driving. IEEE Transactions on Intelligent Transportation Systems, 21(5):1826-1848, 2020. 2" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 62, + 460, + 295, + 513 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 460, + 295, + 513 + ], + "spans": [ + { + "bbox": [ + 62, + 460, + 295, + 513 + ], + "type": "text", + "content": "[8] Felipe Codevilla, Matthias Müller, Antonio López, Vladlen Koltun, and Alexey Dosovitskiy. End-to-end driving via conditional imitation learning. In 2018 IEEE international conference on robotics and automation (ICRA), pages 4693-4700. IEEE, 2018. 2" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 62, + 515, + 295, + 568 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 515, + 295, + 568 + ], + "spans": [ + { + "bbox": [ + 62, + 515, + 295, + 568 + ], + "type": "text", + "content": "[9] Felipe Codevilla, Matthias Müller, Antonio López, Vladlen Koltun, and Alexey Dosovitskiy. End-to-end driving via conditional imitation learning. In 2018 IEEE International Conference on Robotics and Automation (ICRA), pages 4693-4700, 2018." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 57, + 571, + 295, + 624 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 571, + 295, + 624 + ], + "spans": [ + { + "bbox": [ + 57, + 571, + 295, + 624 + ], + "type": "text", + "content": "[10] Felipe Codevilla, Eder Santana, Antonio M. Lopez, and Adrien Gaidon. Exploring the limitations of behavior cloning for autonomous driving. In Proceedings of the IEEE/CVF International Conference on Computer Vision (ICCV), 2019." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 57, + 627, + 295, + 692 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 627, + 295, + 692 + ], + "spans": [ + { + "bbox": [ + 57, + 627, + 295, + 692 + ], + "type": "text", + "content": "[11] Simon Doll, Niklas Hanselmann, Lukas Schneider, Richard Schulz, Marius Cordts, Markus Enzweiler, and Hendrik Lensch. Dualad: Disentangling the dynamic and static world for end-to-end driving. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 14728-14737, 2024. 3" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 57, + 693, + 295, + 715 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 693, + 295, + 715 + ], + "spans": [ + { + "bbox": [ + 57, + 693, + 295, + 715 + ], + "type": "text", + "content": "[12] Mingzhe Guo, Zhipeng Zhang, Yuan He, Ke Wang, and Liping Jing. End-to-end autonomous driving without costly" + } + ] + } + ], + "index": 12 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 316, + 74, + 555, + 715 + ], + "type": "list", + "angle": 0, + "index": 27, + "blocks": [ + { + "bbox": [ + 335, + 74, + 553, + 94 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 335, + 74, + 553, + 94 + ], + "spans": [ + { + "bbox": [ + 335, + 74, + 553, + 94 + ], + "type": "text", + "content": "modularization and 3d manual annotation. arXiv preprint arXiv:2406.17680, 2024." + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 317, + 97, + 553, + 151 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 97, + 553, + 151 + ], + "spans": [ + { + "bbox": [ + 317, + 97, + 553, + 151 + ], + "type": "text", + "content": "[13] Peiyun Hu, Aaron Huang, John Dolan, David Held, and Deva Ramanan. Safe local motion planning with self-supervised freespace forecasting. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 12732-12741, 2021. 2" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 317, + 153, + 553, + 207 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 153, + 553, + 207 + ], + "spans": [ + { + "bbox": [ + 317, + 153, + 553, + 207 + ], + "type": "text", + "content": "[14] Shengchao Hu, Li Chen, Penghao Wu, Hongyang Li, Junchi Yan, and Dacheng Tao. St-p3: End-to-end vision-based autonomous driving via spatial-temporal feature learning. In European Conference on Computer Vision, pages 533-549. Springer, 2022. 1, 3" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 317, + 209, + 553, + 285 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 209, + 553, + 285 + ], + "spans": [ + { + "bbox": [ + 317, + 209, + 553, + 285 + ], + "type": "text", + "content": "[15] Yihan Hu, Jiazhi Yang, Li Chen, Keyu Li, Chonghao Sima, Xizhou Zhu, Siqi Chai, Senyao Du, Tianwei Lin, Wenhai Wang, Lewei Lu, Xiaosong Jia, Qiang Liu, Jifeng Dai, Yu Qiao, and Hongyang Li. Planning-oriented autonomous driving. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 17853-17862, 2023. 1, 2, 3, 5, 6, 7, 8" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 316, + 288, + 553, + 319 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 288, + 553, + 319 + ], + "spans": [ + { + "bbox": [ + 316, + 288, + 553, + 319 + ], + "type": "text", + "content": "[16] Xiangdong Ji, Yizhuang Liu, Yu-Sheng Liu, Jian-Hui Zhang, and Yong Zhao. Large-momentum effective theory. Reviews of Modern Physics, 93(3):035005, 2021. 2" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 317, + 322, + 553, + 375 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 322, + 553, + 375 + ], + "spans": [ + { + "bbox": [ + 317, + 322, + 553, + 375 + ], + "type": "text", + "content": "[17] Xiaosong Jia, Zhenjie Yang, Qifeng Li, Zhiyuan Zhang, and Junchi Yan. Bench2drive: Towards multi-ability benchmarking of closed-loop end-to-end autonomous driving. Advances in Neural Information Processing Systems, 37:819–844, 2025. 5" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 317, + 378, + 553, + 432 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 378, + 553, + 432 + ], + "spans": [ + { + "bbox": [ + 317, + 378, + 553, + 432 + ], + "type": "text", + "content": "[18] Bo Jiang, Shaoyu Chen, Xinggang Wang, Bencheng Liao, Tianheng Cheng, Jiajie Chen, Helong Zhou, Qian Zhang, Wenyu Liu, and Chang Huang. Perceive, interact, predict: Learning dynamic and static clues for end-to-end motion prediction. arXiv preprint arXiv:2212.02181, 2022. 2" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 317, + 434, + 553, + 499 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 434, + 553, + 499 + ], + "spans": [ + { + "bbox": [ + 317, + 434, + 553, + 499 + ], + "type": "text", + "content": "[19] Bo Jiang, Shaoyu Chen, Qing Xu, Benchcheng Liao, Jiajie Chen, Helong Zhou, Qian Zhang, Wenyu Liu, Chang Huang, and Xinggang Wang. Vad: Vectorized scene representation for efficient autonomous driving. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 8340-8350, 2023. 1, 2, 3, 5, 6, 7, 8" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 317, + 502, + 553, + 544 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 502, + 553, + 544 + ], + "spans": [ + { + "bbox": [ + 317, + 502, + 553, + 544 + ], + "type": "text", + "content": "[20] Bo Jiang, Shaoyu Chen, Benchcheng Liao, Xingyu Zhang, Wei Yin, Qian Zhang, Chang Huang, Wenyu Liu, and Xinggang Wang. Senna: Bridging large vision-language models and end-to-end autonomous driving, 2024." + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 317, + 547, + 553, + 600 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 547, + 553, + 600 + ], + "spans": [ + { + "bbox": [ + 317, + 547, + 553, + 600 + ], + "type": "text", + "content": "[21] Zhenxin Li, Kailin Li, Shihao Wang, Shiyi Lan, Zhiding Yu, Yishen Ji, Zhiqi Li, Ziye Zhu, Jan Kautz, Zuxuan Wu, et al. Hydra-mdp: End-to-end multimodal planning with multitarget hydra-distillation. arXiv preprint arXiv:2406.06978, 2024. 2" + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 317, + 603, + 555, + 657 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 603, + 555, + 657 + ], + "spans": [ + { + "bbox": [ + 317, + 603, + 555, + 657 + ], + "type": "text", + "content": "[22] Zhiqi Li, Zhiding Yu, Shiyi Lan, Jiahan Li, Jan Kautz, Tong Lu, and Jose M Alvarez. Is ego status all you need for open-loop end-to-end autonomous driving? In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 14864-14873, 2024. 5, 6, 8" + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 317, + 659, + 553, + 690 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 659, + 553, + 690 + ], + "spans": [ + { + "bbox": [ + 317, + 659, + 553, + 690 + ], + "type": "text", + "content": "[23] Jiageng Mao, Yuxi Qian, Junjie Ye, Hang Zhao, and Yue Wang. Gpt-driver: Learning to drive with gpt. arXiv preprint arXiv:2310.01415, 2023. 2" + } + ] + } + ], + "index": 25 + }, + { + "bbox": [ + 316, + 693, + 553, + 715 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 693, + 553, + 715 + ], + "spans": [ + { + "bbox": [ + 316, + 693, + 553, + 715 + ], + "type": "text", + "content": "[24] Aditya Prakash, Kashyap Chitta, and Andreas Geiger. Multimodal fusion transformer for end-to-end autonomous driv" + } + ] + } + ], + "index": 26 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 318, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 318, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 318, + 757 + ], + "type": "text", + "content": "22440" + } + ] + } + ], + "index": 28 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 8 + }, + { + "para_blocks": [ + { + "bbox": [ + 56, + 72, + 296, + 699 + ], + "type": "list", + "angle": 0, + "index": 13, + "blocks": [ + { + "bbox": [ + 76, + 72, + 296, + 95 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 76, + 72, + 296, + 95 + ], + "spans": [ + { + "bbox": [ + 76, + 72, + 296, + 95 + ], + "type": "text", + "content": "ing. In 2021 IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), 2021. 2" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 56, + 96, + 296, + 151 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 96, + 296, + 151 + ], + "spans": [ + { + "bbox": [ + 56, + 96, + 296, + 151 + ], + "type": "text", + "content": "[25] Ziying Song, Lin Liu, Feiyang Jia, Yadan Luo, Caiyan Jia, Guoxin Zhang, Lei Yang, and Li Wang. Robustness-aware 3d object detection in autonomous driving: A review and outlook. IEEE Transactions on Intelligent Transportation Systems, pages 1-30, 2024. 2" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 56, + 152, + 295, + 196 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 152, + 295, + 196 + ], + "spans": [ + { + "bbox": [ + 56, + 152, + 295, + 196 + ], + "type": "text", + "content": "[26] Haisheng Su, Wei Wu, and Junchi Yan. Difsd: Ego-centric fully sparse paradigm with uncertainty denoising and iterative refinement for efficient end-to-end autonomous driving. arXiv preprint arXiv:2409.09777, 2024. 2" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 56, + 197, + 295, + 241 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 197, + 295, + 241 + ], + "spans": [ + { + "bbox": [ + 56, + 197, + 295, + 241 + ], + "type": "text", + "content": "[27] Wenchao Sun, Xuewu Lin, Yining Shi, Chuang Zhang, Hao ran Wu, and Sifa Zheng. Sparsedrive: End-to-end autonomous driving via sparse scene representation. arXiv preprint arXiv:2405.19620, 2024. 1, 2, 3, 5, 6, 7, 8" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 56, + 242, + 295, + 285 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 242, + 295, + 285 + ], + "spans": [ + { + "bbox": [ + 56, + 242, + 295, + 285 + ], + "type": "text", + "content": "[28] Junming Wang, Xingyu Zhang, Zebin Xing, Songen Gu, Xiaoyang Guo, Yang Hu, Ziying Song, Qian Zhang, Xiaoxiao Long, and Wei Yin. He-drive: Human-like end-to-end driving with vision language models, 2024. 1, 2, 3" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 56, + 286, + 295, + 342 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 286, + 295, + 342 + ], + "spans": [ + { + "bbox": [ + 56, + 286, + 295, + 342 + ], + "type": "text", + "content": "[29] Li Wang, Xinyu Zhang, Ziying Song, Jiangfeng Bi, Guoxin Zhang, Haiyue Wei, Liyao Tang, Lei Yang, Jun Li, Caiyan Jia, et al. Multi-modal 3d object detection in autonomous driving: A survey and taxonomy. IEEE Transactions on Intelligent Vehicles, 2023. 2" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 56, + 342, + 295, + 397 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 342, + 295, + 397 + ], + "spans": [ + { + "bbox": [ + 56, + 342, + 295, + 397 + ], + "type": "text", + "content": "[30] Penghao Wu, Xiaosong Jia, Li Chen, Junchi Yan, Hongyang Li, and Yu Qiao. Trajectory-guided control prediction for end-to-end autonomous driving: A simple yet strong baseline. In Advances in Neural Information Processing Systems, pages 6119-6132. Curran Associates, Inc., 2022. 2" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 56, + 398, + 295, + 441 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 398, + 295, + 441 + ], + "spans": [ + { + "bbox": [ + 56, + 398, + 295, + 441 + ], + "type": "text", + "content": "[31] Zeyu Yang, Nan Song, Wei Li, Xiatian Zhu, Li Zhang, and Philip HS Torr. Deepinteraction++: Multi-modality interaction for autonomous driving. arXiv preprint arXiv:2408.05075, 2024." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 56, + 443, + 295, + 498 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 443, + 295, + 498 + ], + "spans": [ + { + "bbox": [ + 56, + 443, + 295, + 498 + ], + "type": "text", + "content": "[32] Tengju Ye, Wei Jing, Chunyong Hu, Shikun Huang, Lingping Gao, Fangzhen Li, Jingke Wang, Ke Guo, Wencong Xiao, Weibo Mao, Hang Zheng, Kun Li, Junbo Chen, and Kaicheng Yu. Fusionad: Multi-modality fusion for prediction and planning tasks of autonomous driving, 2023. 1" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 56, + 498, + 295, + 552 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 498, + 295, + 552 + ], + "spans": [ + { + "bbox": [ + 56, + 498, + 295, + 552 + ], + "type": "text", + "content": "[33] Wenyuan Zeng, Wenjie Luo, Simon Suo, Abbas Sadat, Bin Yang, Sergio Casas, and Raquel Urtasun. End-to-end interpretable neural motion planner. In 2019 IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), 2019." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 56, + 554, + 295, + 609 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 554, + 295, + 609 + ], + "spans": [ + { + "bbox": [ + 56, + 554, + 295, + 609 + ], + "type": "text", + "content": "[34] Yunpeng Zhang, Deheng Qian, Ding Li, Yifeng Pan, Yong Chen, Zhenbao Liang, Zhiyao Zhang, Shurui Zhang, Hongxu Li, Maolei Fu, et al. Graphad: Interaction scene graph for end-to-end autonomous driving. arXiv preprint arXiv:2403.19098, 2024." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 56, + 610, + 295, + 665 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 610, + 295, + 665 + ], + "spans": [ + { + "bbox": [ + 56, + 610, + 295, + 665 + ], + "type": "text", + "content": "[35] Zhejun Zhang, Alexander Liniger, Dengxin Dai, Fisher Yu, and Luc Van Gool. End-to-end urban driving by imitating a reinforcement learning coach. In Proceedings of the IEEE/CVF International Conference on Computer Vision (ICCV), pages 15222-15232, 2021." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 56, + 666, + 295, + 699 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 666, + 295, + 699 + ], + "spans": [ + { + "bbox": [ + 56, + 666, + 295, + 699 + ], + "type": "text", + "content": "[36] Wenzhao Zheng, Ruiqi Song, Xianda Guo, and Long Chen. Genad: Generative end-to-end autonomous driving. arXiv preprint arXiv:2402.11502, 2024. 2" + } + ] + } + ], + "index": 12 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 749, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 749, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 749, + 317, + 757 + ], + "type": "text", + "content": "22441" + } + ] + } + ], + "index": 14 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 9 + } + ], + "_backend": "vlm", + "_version_name": "2.6.4" +} \ No newline at end of file diff --git a/2025/Doppelgangers and Adversarial Vulnerability/85f12467-2b9c-44c2-8a35-40f17a213b17_content_list.json b/2025/Doppelgangers and Adversarial Vulnerability/85f12467-2b9c-44c2-8a35-40f17a213b17_content_list.json new file mode 100644 index 0000000000000000000000000000000000000000..8694554f895642ad15f42f0340a9853f84deec17 --- /dev/null +++ b/2025/Doppelgangers and Adversarial Vulnerability/85f12467-2b9c-44c2-8a35-40f17a213b17_content_list.json @@ -0,0 +1,2105 @@ +[ + { + "type": "text", + "text": "Doppelgängers and Adversarial Vulnerability", + "text_level": 1, + "bbox": [ + 267, + 130, + 728, + 152 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "George Kamberov* \nUniversity of Alaska Anchorage \nAnchorage, AK 99508 \ngkamberov@alaska.edu", + "bbox": [ + 370, + 181, + 627, + 252 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Abstract", + "text_level": 1, + "bbox": [ + 248, + 286, + 326, + 301 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Machine learning (ML) classifiers can make mistakes that are perceptually and cognitively disturbing to humans. The most notorious examples of such errors are adversarial visual metamers. This paper investigates the phenomenon of adversarial Doppelgängers (AD), which encompasses adversarial visual metamers, and compares the performance and robustness of ML classifiers to human performance.", + "bbox": [ + 89, + 316, + 483, + 424 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "We find that ADs are inputs that are close to each other with respect to a perceptual metric defined in this paper, and show that ADs are qualitatively different from the usual adversarial examples. The vast majority of classifiers are vulnerable to ADs and robustness-accuracy trade-offs may not improve them. Some classification problems do not admit any AD-robust classifiers because the underlying classes are ambiguous. We provide criteria to determine whether a classification problem is well defined; describe the structure and attributes of AD-robust classifiers; introduce and explore the notions of conceptual entropy and regions of conceptual ambiguity for classifiers that are vulnerable to AD attacks; and discuss methods to bound the AD fooling rate of an attack. We define the notion of classifiers that exhibit hypersensitive behavior, that is, classifiers whose only mistakes are adversarial Doppelgangers. Improving the AD robustness of hypersensitive classifiers is equivalent to improving accuracy. We identify conditions guaranteeing that all classifiers with sufficiently high accuracy are hypersensitive.", + "bbox": [ + 91, + 424, + 483, + 724 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "1. Introduction", + "text_level": 1, + "bbox": [ + 91, + 755, + 220, + 771 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Perceptual metamers1 are the most striking adversarial examples studied by the machine learning community. Two perceptual metamers are shown in Figure 1. The phenomenon of metamersim studied in the visual domain, in", + "bbox": [ + 89, + 781, + 482, + 842 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "cluding perceptual metamers, is a manifestation of the existence of Doppelgängers: different inputs or stimuli that are perceptually indiscriminable. The research community has engaged in active studies of adversarial vulnerability ever since the publication of [73]. Adversarial Doppelgängers, that is, adversarial examples which are Doppelgängers, are qualitatively different from the vast majority of known adversarial examples which humans readily discriminate from correctly classified input samples (Figure 2). There is no", + "bbox": [ + 511, + 287, + 906, + 424 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/62ed6623050a90fa01673828c79577cb3c0d76de6010d623657b9cecd15d7a81.jpg", + "image_caption": [ + "(a)" + ], + "image_footnote": [], + "bbox": [ + 545, + 444, + 707, + 564 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/d40457b559e82150da3a0cecb420f2a4a7324d6ee1438310d9d407d5b8f97774.jpg", + "image_caption": [ + "(b)", + "Figure 1. Most people cannot discriminate image (a) from image (b). MobileNetV2 classifies the later image as \"persian\" and the former picture as \"taby\"." + ], + "image_footnote": [], + "bbox": [ + 718, + 444, + 870, + 564 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/a667691483f835cef81b7244c3cfb8676b8ca325f9fcaa6b04d3b02471edcc62.jpg", + "image_caption": [ + "(a) Labrador", + "Figure 2. Applying a Fast Signed Gradient perturbation to the image (a) classified by MobileNetV2 as Labrador yields the image (b) which is classified by MobileNetV2 as Weimeraner." + ], + "image_footnote": [], + "bbox": [ + 545, + 691, + 707, + 813 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/d6fc1eea66ea1f1c66e3f3c7e073f39f21c9fec269a3cf921e69cfe48859c98a.jpg", + "image_caption": [ + "(b) Weimeraner" + ], + "image_footnote": [], + "bbox": [ + 712, + 691, + 870, + 813 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "CVF", + "bbox": [ + 106, + 2, + 181, + 42 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore.", + "bbox": [ + 236, + 0, + 810, + 46 + ], + "page_idx": 0 + }, + { + "type": "page_footnote", + "text": "*This material is based upon work supported by the National Science Foundation under Grant No. 2433241 \n1“images that are physically distinct but perceptually indistinguishable”, [6]. See also “metameric images”, [37]", + "bbox": [ + 89, + 849, + 482, + 900 + ], + "page_idx": 0 + }, + { + "type": "page_number", + "text": "10244", + "bbox": [ + 480, + 944, + 519, + 955 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "evidence that Doppelgängers can be studied and understood completely using the $\\ell_p$ norms or more general geodesic distances on manifolds, which have been employed to quantify sample differences and to investigate adversarial examples. Perception and context impose topology on a space of inputs, but it rarely aligns with a manifold topology. We will denote this context-relative topology by $\\tau_{\\delta}$ . It is defined by the context-relative ability to acquire and deploy knowledge.", + "bbox": [ + 88, + 90, + 480, + 226 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "In this paper, we explore the context-relative perceptual topologies on a space of inputs $\\mathbf{X}$ and examine the vulnerability and robustness of machine learning classifiers to adversarial Doppelgangers. We show that, while the majority of classifiers are vulnerable to adversarial Doppelgangers, safe (Doppelganger robust) classifiers do exist if the classification problem is well defined. However, these robust classifiers may be very rare.", + "bbox": [ + 89, + 227, + 482, + 347 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "In Section 3, we discuss the context-relative notion of (active) indiscriminability and the topology $\\tau_{\\delta}$ that it induces on a space of inputs. The separability and metric properties of various motivating examples of $\\tau_{\\delta}$ . Additionally, we show that the distances between Doppelgängers are small if measured by a perceptually-based context-relative metric. We also examine the relation between indiscriminability, IDSicriminability, and feature representations. The existence and structure of Doppelgänger robust classifiers is discussed in Section 4. In Section 5, we investigate the relationship between Doppelgängers and misclassified input samples, define the notion of hypersensitive behavior, and show that improved adversarial Doppelgänger robustness does not have to lead to a reduction in accuracy.", + "bbox": [ + 89, + 347, + 482, + 559 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "The structure of perceptually regular, i.e., AD robust classifiers are discussed in Section 6. By definition, a classifier is not regular if and only if some inputs can be attacked by adversarial Doppelgangers. Not surprisingly, it turns out that some inputs are more vulnerable than others. In Section 7, we provide measures of adversarial Doppelgangers vulnerability and upper bounds on the fooling rate of an adversarial Doppelganger attack.", + "bbox": [ + 89, + 559, + 482, + 680 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "2. Related work", + "text_level": 1, + "bbox": [ + 89, + 693, + 227, + 708 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "The pair $(\\mathbf{X}, \\text{indiscriminability relation})$ is a tolerance space. Tolerance spaces, rough sets, and granular computing have been discussed extensively. See [59, 62, 66, 88, 89]. The color and image metamers studied by many authors including [3, 6, 9, 18, 22, 37, 39, 41, 42, 51, 76, 86] are Doppelgangers.", + "bbox": [ + 89, + 717, + 482, + 808 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "The research on adversarial examples to date builds on the hypothesis that the space of input samples is a metric space $(\\mathbf{X},\\mathrm{dist}_{\\mathbf{X}})$ . A misclassified input $x^{*}$ is consid", + "bbox": [ + 89, + 809, + 482, + 854 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "ered an adversarial example if it is nearby a correctly classified input sample $x$ , i.e., $\\mathrm{dist}_{\\mathbf{X}}(x, x^{*})$ is small. Usually $\\mathbf{X}$ is assumed to be $\\mathbb{R}^n$ , endowed with the $\\ell_p$ , norm, $p = 1, 2, \\ldots, \\infty$ or at least locally homeomorphic to $\\mathbb{R}^n$ , i.e., a manifold, equipped with some geodesic distance. Somewhat non surprisingly many authors have shown that every classifier can be attacked with such adversarial examples [5, 17, 25, 52] or at least that this is true in many contexts, [44, 47, 67].", + "bbox": [ + 511, + 90, + 903, + 226 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Other papers indicate that there are paths toward eliminating adversarial examples completely, i.e., it is possible to achieve provable \"adversarial robustness\" by fixing/retraining the classifier [1, 24, 36, 43, 72, 74]. A widely accepted tenet is that \"there is a clear trade-off between accuracy and [adversarial] robustness, and a better performance in testing accuracy in general reduces [adversarial] robustness\", [71]. For empirical evidence for this trade-off and some attempts to explain this phenomenon see [71, 78, 90].", + "bbox": [ + 511, + 227, + 903, + 378 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "3. Perceptual Topology", + "text_level": 1, + "bbox": [ + 511, + 391, + 709, + 409 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "3.1 Indiscriminability and Topology", + "text_level": 1, + "bbox": [ + 511, + 417, + 766, + 431 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "The ability to decide whether one stimulus/input is distinct from another is essential for adaptation, survival, and intelligent life. Intelligent agents are uniquely capable to activate knowledge to judge distinction. Williamson calls this context-relative process discrimination ([84]) and defines a context-relative symmetric and reflexive binary relation denoted by $\\stackrel{\\alpha\\delta}{\\approx}$ and called indiscriminability:", + "bbox": [ + 511, + 431, + 903, + 542 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Definition 1 ([84]). Two inputs $x$ and $y$ are called indiscriminable to a subject at a time $t$ if and only if at time $t$ the subject is not able to activate (acquire or employ) the relevant kind of knowledge that $x$ and $y$ are distinct.", + "bbox": [ + 511, + 551, + 903, + 613 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Indiscriminability generates a context-relative topology on the set of inputs $\\mathbf{X}$ .", + "bbox": [ + 511, + 623, + 903, + 652 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Definition 2. The the phenomenal neighborhood of an input $x \\in \\mathbf{X}$ , is the set $\\mathfrak{d}(x) = \\left\\{y \\in \\mathbf{X} : y \\stackrel{\\alpha \\delta}{\\approx} x\\right\\}$ . A point $y \\in \\mathfrak{d}(x) \\setminus \\{x\\}$ is called a Doppelgänger [of $x$ ]. The perceptual topology $\\tau_{\\delta}$ is the topology generated by the sub-basis $\\mathfrak{D}_{\\alpha \\delta} = \\{\\mathfrak{d}(x)\\}_{x \\in \\mathbf{X}}$ .", + "bbox": [ + 511, + 662, + 903, + 756 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Example 0: An input $x \\in \\mathbf{X}$ is called optimal if it does not have non-trivial/non-identical Doppelgangers, i.e.,", + "bbox": [ + 511, + 763, + 903, + 794 + ], + "page_idx": 1 + }, + { + "type": "page_footnote", + "text": "$^{3}$ Context and its role in discrimination and similarity judgments have been studied extensively and by many authors including [11, 23, 30, 45, 60, 77, 79, 81].", + "bbox": [ + 511, + 801, + 903, + 839 + ], + "page_idx": 1 + }, + { + "type": "page_footnote", + "text": "4Some authors refer to indiscriminability as active indiscriminability, see [16]. Poincaré discusses indiscriminability in [58] but refers to it as indiscriminability. Similarly Poston studies indistinguishability, basing it on the \"limit of discrimination\" of the biological senses and instruments, [59].", + "bbox": [ + 511, + 839, + 903, + 898 + ], + "page_idx": 1 + }, + { + "type": "page_footnote", + "text": "2 Some metamers arising in other fields including biology and chemistry are not Doppelgängers, for example, segments in many earthworms are considered metamers but are visually discriminable.", + "bbox": [ + 89, + 862, + 482, + 900 + ], + "page_idx": 1 + }, + { + "type": "page_number", + "text": "10245", + "bbox": [ + 480, + 944, + 517, + 955 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "if $\\mathfrak{d}(x) = \\{x\\}$ . In particular, if $\\stackrel{\\infty \\delta}{\\approx}$ is the identity relationship $=$ (i.e., all inputs are optimal within the given context), then the perceptual topology $\\tau_{\\delta}$ is discrete. However, the finiteness of human observations (they are subject to finite time and finite work constraints) and the bounded rationality constraints imposed by the limitations on the availability of information and computational capabilities to humans, [68], indicate that the scenario $\\mathfrak{d}(x) = \\{x\\}$ for every $x \\in \\mathbf{X}$ may be highly unlikely.", + "bbox": [ + 89, + 89, + 480, + 229 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Often reflexive binary relations are defined and discussed as coverings of the underlying space. Indeed, every reflexive binary relation (RBR) $\\approx$ on $\\mathbf{X}$ defines a covering $\\{\\mathfrak{g}(x) = \\{y:y\\approx x\\} \\}_{x\\in \\mathbf{X}}$ , of $\\mathbf{X}$ and vice versa one can define reflexive binary relationships through coverings of $\\mathbf{X}$ . See Appendix, Section A. In particular, the perceptual topology, $\\tau_{\\delta}$ , is a tolerable topology, cf., Definition 10 in Appendix, Section A.", + "bbox": [ + 89, + 231, + 480, + 352 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "The active psychophysics research on just noticeable difference initiated by Weber and Fechner, [19, 20, 82, 83], provides one of the few classes of examples where we have empirically supported understanding of the perceptual topology.", + "bbox": [ + 89, + 354, + 480, + 430 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Example 1: Let $\\mathbf{X}$ be the closed bounded interval $[a,b]\\subset$ $(0, + \\infty)$ . Suppose that Weber's law holds and let $k > 0$ be the Weber constant. Let $w = 1 + k$ , then", + "bbox": [ + 89, + 445, + 482, + 491 + ], + "page_idx": 2 + }, + { + "type": "equation", + "text": "\n$$\n\\mathfrak {d} (x) = \\left\\{ \\begin{array}{l l} {[ a, x w),} & {a \\leq x < a w} \\\\ {(x / w, x w),} & {a w \\leq x \\leq b / w} \\\\ {(x / w, b ],} & {b / w < x \\leq b.} \\end{array} \\right. \\tag {1}\n$$\n", + "text_format": "latex", + "bbox": [ + 150, + 505, + 480, + 560 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "The covering $\\{\\mathfrak{d}(x)\\}_{x\\in [a,b]}$ defines a symmetric RBR but the relation is not transitive. The corresponding perceptual topology is $T_{0}$ but not $T_{1}$ , and the topology is not pseudometric.", + "bbox": [ + 89, + 574, + 482, + 633 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "The transitivity or more more often the lack of transitivity of $\\stackrel{\\alpha\\delta}{\\approx}$ have been studied extensively and proven or postulated in many human experiences, [2, 6, 10, 16, 26, 27, 32, 54, 61, 84, 85].", + "bbox": [ + 89, + 636, + 482, + 702 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Definition 3. We will denote by $\\sim_{\\sigma}$ the transitive closure of the indiciminability relation $\\stackrel{\\alpha \\delta}{\\approx}$ on $\\mathbf{X}$ . It is defined explicitly as $x \\sim_{\\sigma} y$ iff there exists a finite chain of Doppelgangers $x = x_0 \\stackrel{\\alpha \\delta}{\\approx} x_1 \\stackrel{\\alpha \\delta}{\\approx} x_2 \\stackrel{\\alpha \\delta}{\\approx} \\dots \\stackrel{\\alpha \\delta}{\\approx} x_n = y$ . We will call the relation $\\sim_{\\sigma}$ perceptual metamorphy and will refer to any two inputs $x \\sim_{\\sigma} y$ as metamorphic. Extending Pawlak's terminology, [53] we call the equivalence classes in $\\mathbf{X} / \\sim_{\\sigma}$ (perceptually) elementary sets.", + "bbox": [ + 89, + 715, + 482, + 848 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Example 2: If the indiscriminability relation is transitive, then each $\\mathfrak{d}(x)$ is an elementary set. The perceptual topology may be optimal (recall Example 0) or not. In the former case it is Hausdorff and in fact $(\\mathbf{X},\\tau_{\\delta})$ is a discrete manifold, in the later case the topology $\\tau_{\\delta}$ is not $T_0$ . See Part A.1 in the Appendix. Human visual perception provides a fundamental example where $\\mathfrak{d}(x)\\neq \\{x\\}$ . If two images $x$ and $y$ differ only in unattended regions for example due to low saliency values (cf. [92]), then $x\\stackrel {\\alpha \\delta}{\\approx}y$ . Visual metamers have been studied by many authors, including [3, 6, 9, 18, 22, 28, 37, 39, 41, 42, 51, 76, 86]. In these studies, the input space is assumed to be endowed with Grassmann structure (see [39]), and in particular, the indiscriminability relation is transitive. $^6$", + "bbox": [ + 511, + 90, + 903, + 306 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "We are not aware of perceptual topologies that are metric. Still, every Doppelgänger $y \\stackrel{\\alpha \\delta}{\\approx} x$ of an input $x$ is a small perturbation of $x$ in the sense that the $y$ is a nearest neighbor of $x$ with respect to an appropriate metric $d_w(\\cdot, \\cdot)$ on $\\mathbf{X}$ . Indeed, let the discrimination graph $\\Gamma(\\mathbf{X}, E_{\\alpha \\delta})$ be the undirected simple graph, where $\\mathbf{X}$ is the set of vertices and we say that there is an edge $\\{x, y\\} \\in E_{\\alpha \\delta}$ between the vertices $x, y \\in \\mathbf{X}$ iff $x \\stackrel{\\alpha \\delta}{\\approx} y$ .", + "bbox": [ + 511, + 306, + 903, + 438 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Definition 4. We will call the discrimination graph distance $d_{\\infty}(x,y)$ between the vertices $x$ and $y$ and, in particular, $d_{\\infty}(x,y) = \\infty$ iff $x \\not\\simeq_{\\sigma} y$ the extended perceptual distance. The perceptual distance is the metric $d_w: \\mathbf{X} \\times \\mathbf{X} \\to [0,1]$ defined by:", + "bbox": [ + 511, + 446, + 903, + 523 + ], + "page_idx": 2 + }, + { + "type": "equation", + "text": "\n$$\nd _ {w} (x, y) = \\frac {d _ {\\infty} (x , y)}{1 + d _ {\\infty} (x , y)}, \\forall x, y \\in \\mathbf {X}. \\tag {2}\n$$\n", + "text_format": "latex", + "bbox": [ + 578, + 531, + 903, + 566 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "The metric $d_w$ does not generate the perceptual topology. It is certainly not the usual $l_p$ or any other manifold metric used in ML.", + "bbox": [ + 511, + 571, + 903, + 617 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3.2 Indiscriminabile may not be Indiscernible", + "text_level": 1, + "bbox": [ + 511, + 632, + 834, + 647 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Let $\\Phi$ be the space of all features of the inputs/stimuli $x\\in \\mathbf{X}$ and let $\\Phi_x\\subset \\Phi$ be the set of features attributed to $x$ in a given context. Following [21], we say that $x$ and $y$ are indiscernible, in a given context, if $\\Phi_x = \\Phi_y$ . Many researchers use the terms indiscriminability and indiscernibility as synonyms. This is only accurate when $\\stackrel{\\alpha\\delta}{\\approx}$ is transi", + "bbox": [ + 511, + 648, + 903, + 743 + ], + "page_idx": 2 + }, + { + "type": "page_footnote", + "text": "6In these studies, indiscriminable inputs are referred to as \"matching\" or \"metameric\", or \"alike\".", + "bbox": [ + 511, + 751, + 903, + 775 + ], + "page_idx": 2 + }, + { + "type": "page_footnote", + "text": "The open metric ball $\\tilde{B}_{1/2}^w(x)$ equals $\\{x\\}$ , for all inputs $x \\in \\mathbf{X}$ . On the other hand, the finiteness of human observations and the hypothesis of bounded rationality suggest that biologically plausible perceptual topologies are not discrete.", + "bbox": [ + 511, + 776, + 903, + 825 + ], + "page_idx": 2 + }, + { + "type": "page_footnote", + "text": "The existence of the extended metric was hinted at in [46]; it was discussed in a related context in [65] and rediscovered and exploited in [59]. For more discussion see Part F.1 in the Appendix.", + "bbox": [ + 511, + 827, + 903, + 863 + ], + "page_idx": 2 + }, + { + "type": "page_footnote", + "text": "9Leibniz discussed indiscernibility and postulated the Principle/Law of Identity of Indiscernibles, $(\\Phi_{x} = \\Phi_{y})\\Rightarrow (x = y)$ , in [80] and in the third and fourth papers addressed to Samuel Clarke, [8].", + "bbox": [ + 511, + 863, + 903, + 900 + ], + "page_idx": 2 + }, + { + "type": "page_footnote", + "text": "5Optimal objects and light sources have been described and studied in colorimetry, [42, 87].", + "bbox": [ + 89, + 875, + 482, + 900 + ], + "page_idx": 2 + }, + { + "type": "page_number", + "text": "10246", + "bbox": [ + 480, + 944, + 517, + 955 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "tive. $^{10}$ In general, the relationship between indiscriminability and indiscernibility is not well understood. However, indiscernibility implies indiscriminability if $\\{\\Phi_x\\}_{x\\in \\mathbf{X}}$ is a perceptually discriminative feature representation, i.e.,", + "bbox": [ + 89, + 90, + 483, + 154 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\Phi_ {x} \\bigcap \\Phi_ {y} \\neq \\varnothing \\Longleftrightarrow x \\stackrel {\\circledast} {\\approx} y. \\tag {3}\n$$\n", + "text_format": "latex", + "bbox": [ + 194, + 162, + 482, + 186 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "The biological plausibility of discriminative feature representations is an open question. Still, they provide insight into the structure of the perceptual topology. The attributed discriminative features represent structures of Doppelgangers. Indeed, let $\\{\\Phi_x\\}_{x\\in \\mathbf{X}}$ be a feature representation and let $\\operatorname {cl}(\\xi)$ be the context-dependent semantic cluster of inputs sharing the feature $\\xi \\in \\Phi$ . Specifically,", + "bbox": [ + 89, + 195, + 483, + 303 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\mathfrak {c l} (\\xi) = \\left\\{x \\in \\mathbf {X}, \\text {s . t . ,} \\xi \\in \\Phi_ {x} \\right\\}. ^ {1 2} \\tag {4}\n$$\n", + "text_format": "latex", + "bbox": [ + 174, + 311, + 482, + 330 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "In particular, if $\\{\\Phi_x\\}_{x\\in \\mathbf{X}}$ is a discriminative feature representation, then every attributed discriminative feature $\\xi \\in \\Phi_{x}$ is associated to, and in some way explained by, a collection of Doppelgangers since $\\mathfrak{cl}(\\xi)\\subset \\mathfrak{d}(x)$ . For more detailed discussion and examples of discriminative feature representations see Part B in the Appendix.", + "bbox": [ + 89, + 339, + 483, + 431 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "4. Classifiers and Adversarial Doppelgängers.", + "text_level": 1, + "bbox": [ + 89, + 443, + 475, + 460 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "A classifier $R$ (with $m$ labels) is called fully populated iff the labeling function $\\text{label}_R: \\mathbf{X} \\to \\{1, \\dots, m\\}$ is surjective mapping onto the range of labels $\\{1, \\dots, m\\}$ . For any classifier $R$ , with $m$ labels, we will denote by $R_c$ the level set of the labeling function $\\text{label}_R$ for each label $c \\in \\{1, \\dots, m\\}$ .13", + "bbox": [ + 89, + 468, + 483, + 559 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Definition 5. We say that $x$ is a Doppelgänger adversarial to the classifier $R$ iff $\\exists y \\in \\mathfrak{d}(x)$ such that $\\operatorname{label}_R(x) \\neq \\operatorname{label}_R(y)$ and we will refer to both $x$ and $y$ as adversarial Doppelgangers when the classifier $R$ is clear from the context.", + "bbox": [ + 89, + 568, + 483, + 642 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "A classifier $R$ is called (perceptually) regular iff it does not admit adversarial Doppelgangers. If $R$ is regular, then $\\mathfrak{D}_{\\alpha \\delta} = \\{\\mathfrak{d}(x)\\}_{x \\in \\mathbf{X}}$ are $R$ coherent coverings (as defined in [66]).", + "bbox": [ + 89, + 643, + 483, + 703 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "We say that the classification problem with $m$ -labels is well defined if there exists a fully populated and perceptually regular classifier with $m$ labels. Otherwise we say that the classification problem with $m$ -labels is not well defined.", + "bbox": [ + 89, + 704, + 483, + 777 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "The labeling function $\\text{label}_R$ of a regular classifier is continuous with respect to the perceptual topology $\\tau_{\\delta}$ . Furthermore, if $R$ is not regular and $x$ is a point of discontinuity of $\\text{label}_R: (\\mathbf{X}, \\tau_{\\delta}) \\to \\{1, \\dots, m\\}$ , then $x$ is an adversarial Doppelgänger. The discontinuity is an indication of the cognitive disruption that occurs when one encounters some AD.", + "bbox": [ + 511, + 90, + 903, + 194 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "It is well known that in some experiences perceptually unambiguous categories and hence perceptually regular classifiers do not exist. A simple example is provided by the perceptual topology in Example 1. Indeed, in this case $\\mathbf{X} / \\sim_{\\sigma}$ is a singleton, i.e., every two inputs are metamorphic and hence there is a only one elementary set which equals the whole $\\mathbf{X}$ . Therefore, every classifier with two or more labels must have adversarial Doppelgängers.[14]", + "bbox": [ + 511, + 196, + 905, + 318 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Clearly, no amount of \"robust training\" will get rid of adversarial examples of a classifier with a surjective labeling function $\\text{label}_R: (\\mathbf{X}, \\tau_{\\delta}) \\to \\{1, \\ldots, m\\}$ if $\\mathbf{X}$ cannot be broken into $m$ perceptually unambiguous categories. In the rest of this section we investigate the non-existence, existence and internal structure of regular classifiers.", + "bbox": [ + 511, + 318, + 905, + 409 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "In Part D of the Appendix we show a specific example of a well defined classification problem and discuss the actual regular classifier.", + "bbox": [ + 511, + 409, + 905, + 454 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Example 3: If $\\stackrel{\\alpha \\delta}{\\approx}$ is transitive then for every number of labels $m$ smaller than the number of equivalence classes card $\\left(\\mathbf{X} / \\stackrel{\\alpha \\delta}{\\approx}\\right)$ there exists a fully populated regular classifier with $m$ labels, but if the number of labels $m$ is bigger than card $\\left(\\mathbf{X} / \\stackrel{\\alpha \\delta}{\\approx}\\right)$ then every fully populated classifier with $m$ labels must admit adversarial Doppelgangers.", + "bbox": [ + 511, + 455, + 906, + 580 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Example 3 indicates that if the transitive closure $\\sim_{\\sigma}$ is trivial, i.e., $\\sim_{\\sigma} = \\mathbf{X} \\times \\mathbf{X}$ , then no label is safe. Namely:", + "bbox": [ + 511, + 594, + 905, + 625 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Observation 1. If the transitive closure $\\sim_{\\sigma}$ of the indiscriminability relation $\\stackrel{\\alpha \\delta}{\\approx}$ is trivial, then every fully populated classifier with two or more classes admits adversarial Doppelgängers. In particular, let $R$ be a fully populated classifier with a surjective labeling function $\\text{label}_R: X \\to \\{1, 2, \\ldots, m\\}$ , then for every label $c$ there exist adversarial Doppelgängers $x(c) \\in \\mathbf{X}$ and $x^*(c) \\in \\mathfrak{d}(x(c))$ such that $c = \\text{label}_R(x(c))$ and $\\text{label}_R(x^*(c)) \\neq \\text{label}_R(x(c))$ .", + "bbox": [ + 511, + 635, + 905, + 763 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "The proof follows from the fact that every finite chain of Doppelgängers connecting points that are labeled differently by a classifier must contain a pair of adversarial Doppelgängers. See Lemma 2 in Appendix Section C.", + "bbox": [ + 511, + 772, + 903, + 833 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "A straight-forward argument shows that if $R$ is a perceptually regular fully populated classifier, then $x \\in R_i$ iff", + "bbox": [ + 511, + 834, + 905, + 864 + ], + "page_idx": 3 + }, + { + "type": "page_footnote", + "text": "14See Lemma 3 and the short argument that follows it in Appendix, Section C.", + "bbox": [ + 511, + 875, + 903, + 898 + ], + "page_idx": 3 + }, + { + "type": "page_footnote", + "text": "10See Observation 7, in Part B of the Appendix.", + "bbox": [ + 106, + 785, + 357, + 797 + ], + "page_idx": 3 + }, + { + "type": "page_footnote", + "text": "11A feature $\\xi \\in \\Phi$ is called attributed if $\\xi \\in \\Phi_x$ for some input $x \\in \\mathbf{X}$ . It is plausible that $\\Phi = \\bigcup_{x \\in \\mathbf{X}} \\Phi_x$ , and so all features are attributed. However, many models do not preclude the existence of spurious latent traits.", + "bbox": [ + 91, + 797, + 482, + 849 + ], + "page_idx": 3 + }, + { + "type": "page_footnote", + "text": "12 The semantic cluster of inputs sharing the feature $\\xi \\in \\Phi$ is defined for any feature representation. A feature $\\xi$ is attributed in a given context, iff $\\operatorname{cl}(\\xi) \\neq \\emptyset$ ; a feature is a hypothetical feature, when $\\operatorname{cl}(\\xi) = \\emptyset$ .", + "bbox": [ + 91, + 849, + 482, + 887 + ], + "page_idx": 3 + }, + { + "type": "page_footnote", + "text": "$^{13}R$ is fully populated iff $R_{c}\\neq \\emptyset$ for every label $c\\in \\{1,\\ldots ,m\\}$", + "bbox": [ + 104, + 887, + 455, + 900 + ], + "page_idx": 3 + }, + { + "type": "page_number", + "text": "10247", + "bbox": [ + 480, + 944, + 517, + 955 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "$[x]_{\\sim_{\\sigma}} \\subset R_i$ , i.e., each level set is a disjoint union of elementary sets $[x]_{\\sim_{\\sigma}}$ . The pigeonhole principle yields:", + "bbox": [ + 89, + 90, + 482, + 122 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Observation 2. If the number of equivalence classes card $(\\mathbf{X} / \\sim_{\\sigma}) \\geq 2$ , then for every natural number $2 \\leq m \\leq$ card $(\\mathbf{X} / \\sim_{\\sigma})$ there exists a perceptually regular fully populated classifier with $m$ labels. However, if $m >$ card $(\\mathbf{X} / \\sim_{\\sigma})$ , then every fully populated classifier with $m$ labels must have adversarial Doppelgangers.", + "bbox": [ + 89, + 130, + 483, + 220 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "In particular, if $p = \\operatorname{card}(\\mathbf{X} / \\sim_{\\sigma}) \\geq 2$ is finite, then for every natural number $m \\leq p$ there are exactly $S(p, m)$ regular fully populated classifiers with $m$ classes. Here $S(p, m)$ is the Sterling number of the second kind.", + "bbox": [ + 88, + 220, + 483, + 281 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Observation 2 shows that the problem of finding a fully populated perceptually unambiguous classifier $R$ with precisely $m$ labels is not well defined if card $(\\mathbf{X} / \\sim_{\\sigma}) < m$ and vice versa that the same problem is well defined for every $m$ such that card $(\\mathbf{X} / \\sim_{\\sigma}) \\geq m$ . In the former case solutions do not exist while in the later case solutions exist and each class segment $R_{i}$ is a union of equivalence classes $[x]_{\\sim_{\\sigma}}$ ,", + "bbox": [ + 89, + 290, + 483, + 397 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\nR _ {i} = \\bigcup_ {x \\in R _ {i}} [ x ] _ {\\sim_ {\\sigma}}. \\tag {5}\n$$\n", + "text_format": "latex", + "bbox": [ + 227, + 407, + 482, + 440 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "The existence and properties of discriminative feature representations provide insight whether a classification problem is well defined. In particular, if there exists a discriminative feature representation and the set of attributed features is finite, then every classification problem, whose number of labels exceeds the number of attributed features, is not well defined. See Observation 9 in Appendix Part B.2. Further discussion of the structure of class segments of a regular classifier including the class/category core and fringe are discussed in Section 6.", + "bbox": [ + 89, + 448, + 483, + 599 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "5. Accuracy and Adversarial Doppelgängers.", + "text_level": 1, + "bbox": [ + 89, + 613, + 475, + 630 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "The accuracy-adversarial robustness trade off observed and discussed in the literature involves various measures of accuracy [48, 50, 71, 78, 90]. We will discuss classification accuracy. We will show that there is a strong relationship between classifier accuracy and vulnerability to adversarial Doppelgängers. In particular, we will identify (perceptual) scenarios in which low accuracy classifiers are critically vulnerable to adversarial Doppelgänger attacks but on the other hand all high accuracy classifiers can be fooled only by Doppelgängers.", + "bbox": [ + 89, + 638, + 483, + 789 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "5.1. The Probabilistic Setup.", + "text_level": 1, + "bbox": [ + 89, + 797, + 310, + 814 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "We will assume that $(\\mathbf{X},\\mathcal{F},\\mu)$ is a probability measure space equipped with perceptual topology $\\tau_{\\delta}$ generated by an indsicriminability relation $\\stackrel{\\alpha \\delta}{\\approx}$ such that for every $x \\in \\mathbf{X}$ the set of Doppelgangers $\\mathfrak{d}(x)$ and the equivalence class $[x]_{\\sim_{\\sigma}}$ are events, $(\\mathfrak{d}(x) \\in \\mathcal{F}$ and $[x]_{\\sim_{\\sigma}} \\in \\mathcal{F}, \\forall x \\in \\mathbf{X})$ .", + "bbox": [ + 89, + 819, + 483, + 902 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "In the rest of this section we will assume that the classification problem with $m \\geq 2$ labels is well defined and let $\\Omega$ be a perceptually regular classifier; we will reserve the notation $R$ to denote any classifier ( $R$ may or may not be perceptually regular) such that $R_{i} \\cap \\Omega_{i}$ , $i = 1, \\ldots, m$ are the true positives (of class $i$ ). To define accuracy we will focus only on regular models and classifiers s.t., $\\Omega_{i} \\in \\mathcal{F}$ and $R_{i} \\in \\mathcal{F}$ , for all $i = 1, \\ldots, m$ . The accuracy of the classifier $R$ defined as", + "bbox": [ + 511, + 90, + 906, + 226 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\operatorname {a c c u r a c y} _ {\\Omega} (R) = \\mu \\left(R _ {1} \\cap \\Omega_ {1}\\right) + \\dots + \\mu \\left(R _ {m} \\cap \\Omega_ {m}\\right). \\tag {6}\n$$\n", + "text_format": "latex", + "bbox": [ + 522, + 239, + 906, + 258 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Furthermore let us assume that $\\mu (\\Omega_i) > 0$ for every $i = 1,\\ldots ,m$ and thus we can define recall rates", + "bbox": [ + 511, + 268, + 906, + 300 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\rho_ {i} = \\frac {\\mu \\left(R _ {i} \\cap \\Omega_ {i}\\right)}{\\mu \\left(\\Omega_ {i}\\right)}, \\quad i = 1, \\dots , m \\tag {7}\n$$\n", + "text_format": "latex", + "bbox": [ + 594, + 311, + 905, + 345 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Bounds on the recall rates imply bounds on the accuracy. Namely if", + "bbox": [ + 511, + 357, + 905, + 386 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\underline {{\\rho}} \\leq \\rho_ {i} \\leq \\bar {\\rho}, \\quad i = 1, \\dots , m\n$$\n", + "text_format": "latex", + "bbox": [ + 612, + 388, + 803, + 406 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "then since $\\mu$ is a probability measure on $X$ , and", + "bbox": [ + 511, + 412, + 830, + 429 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\mu \\left(R _ {i} \\cap \\Omega_ {i}\\right) = \\rho_ {i} \\mu \\left(\\Omega_ {i}\\right), \\quad i = 1, \\dots , m \\tag {8}\n$$\n", + "text_format": "latex", + "bbox": [ + 575, + 441, + 906, + 459 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "we get", + "bbox": [ + 513, + 473, + 563, + 486 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\underline {{\\rho}} \\leq \\operatorname {a c c u r a c y} _ {\\Omega} (R) = \\sum_ {i = 1} ^ {m} \\rho_ {i} \\mu (\\Omega_ {i}) \\leq \\bar {\\rho}.\n$$\n", + "text_format": "latex", + "bbox": [ + 576, + 500, + 841, + 539 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "5.2. Are Trade-Offs Possible?", + "text_level": 1, + "bbox": [ + 511, + 551, + 741, + 566 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Let $i(x) \\in \\{1, \\ldots, m\\}$ be the object class label of $x \\in \\mathbf{X}$ i.e., $x \\in \\Omega_{i(x)}$ and", + "bbox": [ + 511, + 574, + 905, + 606 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\bar {k} (\\Omega) = \\sup _ {x \\in \\mathbf {X}} \\left(\\frac {\\mu \\left(\\Omega_ {i (x)}\\right)}{\\mu (\\mathfrak {d} (x))}\\right). \\tag {9}\n$$\n", + "text_format": "latex", + "bbox": [ + 612, + 618, + 905, + 659 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Every classifier whose recall rates do not exceed $1 / \\bar{k} (\\Omega)$ is totally unsafe in the sense that every correctly classified input admits adversarial Doppelgangers. Specifically:", + "bbox": [ + 511, + 672, + 906, + 719 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Observation 3. Suppose that the sets of Doppelgangers are not negligible and $\\inf_{x\\in \\mathbf{X}}\\mu (\\mathfrak{d}(x)) > 0$ , and let $\\Omega = \\{\\Omega_1,\\dots ,\\Omega_m\\}$ be a regular world model and let $R = \\{R_{1},\\ldots ,R_{m}\\}$ be a classifier whose recall rates are strictly smaller than $1 / \\bar{k} (\\Omega)$ and so", + "bbox": [ + 511, + 729, + 906, + 811 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\frac {\\mu \\left(R _ {i (x)} \\cap \\mathfrak {d} (x)\\right)}{\\mu (\\mathfrak {d} (x))} \\leq \\bar {\\rho} \\bar {k} (\\Omega) < 1. \\tag {10}\n$$\n", + "text_format": "latex", + "bbox": [ + 599, + 824, + 903, + 859 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Thus every correctly classified input $x$ has adversarial Doppelgängers.", + "bbox": [ + 511, + 869, + 905, + 901 + ], + "page_idx": 4 + }, + { + "type": "page_number", + "text": "10248", + "bbox": [ + 480, + 944, + 517, + 955 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Observation 3 shows that sacrificing accuracy may lead to increasing the probability of encountering adversarial Doppelgangers and in fact that there is no trade off for accuracies that are sufficiently low provided that all sets of Doppelgangers have positive measure.", + "bbox": [ + 89, + 90, + 482, + 167 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "The lack of an opportunity for a trade-off is even more striking when one tries to improve high recall rate (and hence high accuracy) classifiers.", + "bbox": [ + 89, + 167, + 482, + 213 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Observation 4. $^{15}$ If $\\inf_{x\\in \\mathbf{X}}\\mu (\\mathfrak{d}(x)) > 0$ , and let $R =$ $\\{R_1,\\ldots ,R_m\\}$ be a classifier whose recall rates are sufficiently high so that $\\underline{\\rho} >1 - 1 / \\bar{k} (\\Omega)$ . i.e.,", + "bbox": [ + 89, + 224, + 482, + 277 + ], + "page_idx": 5 + }, + { + "type": "equation", + "text": "\n$$\n\\left(1 - \\underline {{\\rho}}\\right) \\bar {k} (\\Omega) < 1. \\tag {11}\n$$\n", + "text_format": "latex", + "bbox": [ + 223, + 290, + 482, + 311 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Then every misclassified input $x$ is an adversarial Doppelgänger.", + "bbox": [ + 89, + 321, + 482, + 354 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Definition 6. We say that a classifier has a hypersensitive behavior if every misclassified input is an adversarial Doppelgänger.", + "bbox": [ + 89, + 364, + 483, + 411 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "For classifiers with hypersensitive behavior adversarial robustness can only be improved by improving accuracies, i.e., by eliminating misclassification. Observation 4 shows that if $\\mu (\\mathfrak{d}(x)) > 0$ , for all inputs $x\\in \\mathbf{X}$ , then all classifiers with sufficiently high accuracy are either regular (when accuracy equals to one) or have hypersensitive behavior.", + "bbox": [ + 89, + 422, + 482, + 513 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "In summary adversarial Doppelgänger robustness - accuracy trade-off may happen for classifiers with middling accuracy rates or when there are inputs whose Doppelgängers are negligible in measure.", + "bbox": [ + 89, + 513, + 483, + 575 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "6. Life without borders.", + "text_level": 1, + "bbox": [ + 89, + 590, + 294, + 606 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "A important property of the perceptual topology is that the if a classifier $\\Omega$ is perceptually regular, then it \"imposes an open [topological] borders policy\", that is, $\\partial \\Omega_{c} = \\emptyset$ for every label $c$ . Linguists and psychologists, have observed and postulated that natural perceptual and semantic categories are borderless. See for example [64]. Class/decision boundaries are studied and exploited in many works on classifiers, and in particular on adversarial robustness. These boundaries are artifacts of the metric topology used by the researchers, they are not perceptual phenomena.", + "bbox": [ + 89, + 616, + 482, + 767 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "However, we are all familiar with the idea that some stimuli are more intrinsic/representative to/of a given class and at the same time frequently there are objects/stimuli/Input that, while they are firmly with in the class, are less representative/share few(er) features compared to the rest of the elements in the class, that is, they \"are/belong to the fringe\" of the class.", + "bbox": [ + 89, + 768, + 482, + 875 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Definition 7. Let $s$ be a similarity scale, i.e., a function $s: \\mathbf{X} \\times \\mathbf{X} \\to \\mathbb{R}$ such that $s(x, x) \\geq s(x, y), \\forall x, y \\in \\mathbf{X}$ as in [79], [45] (measuring similarity within a fixed context) and [40].", + "bbox": [ + 511, + 90, + 903, + 151 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "The values $s(x, x)$ can be and sometimes are used to represent the salience or equivalently importance of the input within $\\mathbf{X}$ , see for example [79]. The similarity scale provides a method to quantify the affinity of a input/stimulus to a given measurable subset $D \\subset \\mathbf{X}$ and the notions of prototype and fringe. The (s-)affinity of $x$ with a measurable set $D$ is defined as", + "bbox": [ + 511, + 159, + 906, + 265 + ], + "page_idx": 5 + }, + { + "type": "equation", + "text": "\n$$\nP (x, D) = \\int_ {D} s (x, y) \\tag {12}\n$$\n", + "text_format": "latex", + "bbox": [ + 632, + 270, + 903, + 304 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "This is a straightforward generalization of the notion of prototypicality defined in [79].", + "bbox": [ + 511, + 310, + 906, + 340 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Definition 8. $x \\in D$ is called a prototype of $D$ (with respect to the integrable similarity scale $s: \\mathbf{X} \\times \\mathbf{X} \\to \\mathbb{R}$ ) if", + "bbox": [ + 511, + 348, + 906, + 378 + ], + "page_idx": 5 + }, + { + "type": "equation", + "text": "\n$$\nP (x, D) = \\sup _ {z \\in D} P (z, D) \\tag {13}\n$$\n", + "text_format": "latex", + "bbox": [ + 622, + 393, + 903, + 417 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "$x \\in D$ is called a fringe element of $D$ (with respect to the integrable similarity scale $s: \\mathbf{X} \\times \\mathbf{X} \\to \\mathbb{R}$ ) if", + "bbox": [ + 511, + 422, + 903, + 452 + ], + "page_idx": 5 + }, + { + "type": "equation", + "text": "\n$$\nP (x, D) = \\inf _ {z \\in D} P (z, D) \\tag {14}\n$$\n", + "text_format": "latex", + "bbox": [ + 622, + 462, + 903, + 484 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "One is tempted to think of prototypes as the stimuli that are \"clearest cases, best examples\". \"easy to tell apart\" and to \"be a good representative\" and hence that optimize salience [13, 31, 63, 64, 79]. However, clearly there is no reason to expect that a stimulus that is unlikely to be observed/encountered would be selected as a prototype. The examples below show that prototypes and fringes are obtained by optimizing a mixture of the frequency of appearance (likelihood to encounter) and salience.", + "bbox": [ + 511, + 493, + 905, + 628 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "The core and fringe sets may be empty. In practice one is often satisfied with finding elements that may not be true prototypes but belong to the $M$ -core, i.e., have affinity exceeding a fixed threshold $M$ . Similarly, elements whose affinity is below a given threshold $\\tau$ belong to the $\\tau$ -fringe.", + "bbox": [ + 511, + 628, + 905, + 704 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Many perceptual and cognitive processes exploit the intelligent agents' abilities to measure and/or compare the salience/importance of features. In some simple cases it is expected and it even might be true that salience/importance is a probability measure $f_{\\Phi}$ on the space of (all possible) features $\\Phi$ and the feature representations $\\Phi_x$ are measurable subsets of $\\Phi$ . In many accounts including Tversky's feature contrast model [79] the salience scale $f_{\\Phi}$ is a context dependent, nonnegative function defined on a collection $\\Upsilon(\\Phi)$ of subsets of $\\Phi$ which is closed under finite unions and intersections, and set differences. Furthermore, $\\Phi_x \\in \\Upsilon(\\Phi), \\forall x \\in \\mathbf{X}$ ; and the non-negative function $f_{\\Phi}$ is feature additive, i.e., $f_{\\Phi}(A \\cup B) = f_{\\Phi}(A) +$", + "bbox": [ + 511, + 705, + 906, + 901 + ], + "page_idx": 5 + }, + { + "type": "page_footnote", + "text": "15The proof is in Appendix Part E.", + "bbox": [ + 104, + 886, + 289, + 900 + ], + "page_idx": 5 + }, + { + "type": "page_number", + "text": "10249", + "bbox": [ + 480, + 944, + 517, + 955 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "$f_{\\Phi}(B)$ , if $A \\cap B = \\emptyset$ . The value $f(x) = f_{\\phi}(\\Phi_x), \\forall x \\in \\mathbf{X}$ is called the salience/prominence of the input $x$ , [79]. We will call a (Tversky) salience scale $f_{\\Phi}$ perceptually regular if the prominence/salience function $f(x) = f_{\\phi}(\\Phi_x)$ is perceptually regular. We will call a perceptually regular salience scale $f_{\\Phi}$ fully deployable if it can be used to judge the distinguishing features of inputs/stimuli. In particular, we can use $f_{\\Phi}(\\Phi_x \\setminus \\Phi_y)$ to discriminate $x$ from $y$ . Thus if $f_{\\Phi}$ is fully deployable, then $f_{\\Phi}(\\Phi_x \\setminus \\Phi_y) = 0$ and $f_{\\Phi}(\\Phi_y \\setminus \\Phi_x) = 0$ whenever $x \\stackrel{\\alpha \\delta}{\\approx} y$ or equivalently $f_{\\Phi}(\\Phi_x \\cap \\Phi_y) = f(x)$ if $x \\stackrel{\\alpha \\delta}{\\approx} y$ .", + "bbox": [ + 89, + 90, + 483, + 270 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "The following special case provides a particularly useful insight into the nature of prototypes and the possible internal structure of categories as discussed in [64].", + "bbox": [ + 89, + 273, + 483, + 320 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Example 4: If $s$ is a contrast similarity [79] such that $f_{\\Phi}(\\Phi_x) = s(x,x) / \\theta$ is fully deployable, $\\theta \\in (0, + \\infty)$ , and furthermore, $f_{\\Phi}(\\Phi_x \\cap \\Phi_y) = f_{\\Phi}(\\Phi_x)$ for every pair $x \\sim_{\\sigma} y$ and $\\Phi_x \\cap \\Phi_y = \\emptyset$ if $x \\not\\sim_{\\sigma} y$ , then for every regular class $D \\subset \\mathbf{X}$ , and every $x \\in D$ we get", + "bbox": [ + 89, + 339, + 483, + 416 + ], + "page_idx": 6 + }, + { + "type": "equation", + "text": "\n$$\nP (x, D) = \\Theta \\mu (D) \\left(\\frac {\\mu ([ x ] _ {\\sim_ {\\sigma}})}{\\mu (D)} - \\frac {\\alpha}{\\Theta}\\right) f _ {\\Phi} \\left(\\Phi_ {x}\\right) - \\beta I _ {\\Phi} (D), \\tag {15}\n$$\n", + "text_format": "latex", + "bbox": [ + 89, + 436, + 483, + 481 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "where $\\Theta = (\\alpha +\\beta +\\theta)$ and $\\alpha ,\\beta$ ,and $\\theta$ are non-negative constants. In particular, if $D$ is a finite union of equivalence classes $\\zeta_j\\in \\mathbf{X} / \\sim_\\sigma$ , then both prototypes and fringe elements exist. Furthermore, if $x$ is prototype/fringe then so are all stimuli in its component $\\zeta = [x]_{\\sim_{\\sigma}}$ . Similar statements hold for M-core and $\\tau$ -fringe elements.", + "bbox": [ + 89, + 483, + 483, + 573 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "If the class $D$ is fixed, then $\\frac{\\mu([x]_{\\sim_{\\sigma}})}{\\mu(D)}$ is just the probability to encounter (and learn) a set of stimuli, and $f_{\\Phi}(\\Phi_x)$ can be interpreted as the level of prominence. So the optimization process involves learning prominent examples that can be encountered reasonably often. Thus in this case the prototype does not fall in either of the two main branches of prototypes, that is inputs that represent the central tendency in the regular class vs. prototypes as highly \"representative exemplar(s) of a category\", see [15], page 52.", + "bbox": [ + 89, + 578, + 483, + 731 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "More generally, when the similarity scale is bounded, for example, this is true for real similarity measures deployable by humans, then, as predicted by modern prototype theory, each perceptually regular subset $D \\subset \\mathbf{X}$ corresponding to real (natural) categories created and analyzed by real intelligent agents consists of core elements (possibly M-core), a layer of fringe (possibly $\\tau$ -fringe) elements, and layers of elements of various levels of intermediate affinity with $D$ . In Part I of the Appendix we introduce class invariants of regular classifiers including structural entropy, expected index of coincidence, and importance.", + "bbox": [ + 89, + 734, + 483, + 900 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "7. Quantifying Doppelgänger Vulnerability.", + "text_level": 1, + "bbox": [ + 511, + 89, + 879, + 108 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "By definition if a classifier $R$ is not regular, then it is vulnerable to adversarial Doppelgangers attacks, that is, for some input $x$ there exists $a(x) \\stackrel{\\alpha \\delta}{\\approx} x$ and such that $\\operatorname{label}_R(x) \\neq \\operatorname{label}_R(a(x))$ . In particular, we say that $R$ is conceptually ambiguous at $x$ and we call the set", + "bbox": [ + 511, + 114, + 905, + 196 + ], + "page_idx": 6 + }, + { + "type": "equation", + "text": "\n$$\n\\mathrm {A} (R) = \\left\\{x \\in \\mathbf {X}: \\exists y \\stackrel {\\alpha \\delta} {\\approx} x \\text {a n d} \\operatorname {l a b e l} _ {R} (x) \\neq \\operatorname {l a b e l} _ {R} (y) \\right\\} \\tag {16}\n$$\n", + "text_format": "latex", + "bbox": [ + 519, + 205, + 903, + 239 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "the region of conceptual ambiguity. When $(X, \\mu)$ is a probability measure space and $\\mu(\\mathfrak{d}(x)) > 0$ , we use the probability distribution of labels at $x$ :", + "bbox": [ + 511, + 241, + 905, + 286 + ], + "page_idx": 6 + }, + { + "type": "equation", + "text": "\n$$\np _ {j} (x) = \\frac {\\mu \\left(R _ {j} \\bigcap \\mathfrak {d} (x)\\right)}{\\mu (\\mathfrak {d} (x))}, \\quad j = 1, \\dots , m \\tag {17}\n$$\n", + "text_format": "latex", + "bbox": [ + 570, + 294, + 903, + 329 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "and the conceptual entropy of $R$ at $x$ defined as", + "bbox": [ + 511, + 335, + 836, + 351 + ], + "page_idx": 6 + }, + { + "type": "equation", + "text": "\n$$\nH _ {R} (x) = - \\sum_ {j = 1} ^ {m} p _ {j} (x) \\log \\left(p _ {j} (x)\\right) \\tag {18}\n$$\n", + "text_format": "latex", + "bbox": [ + 596, + 359, + 903, + 401 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "to detect whether $R$ is conceptually ambiguous at $x$ (i.e., $H_{R}(x) > 0$ ), and to quantify the likelihoods of various adversarial Doppelgänger attacks.", + "bbox": [ + 511, + 410, + 905, + 455 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Definition 9. Let $R$ be a classifier and $\\hat{a} : \\mathbf{X} \\to \\mathbf{X}$ , we will call the inner measure of the set $\\{x : \\operatorname{label}_R(\\hat{a}(x)) \\neq \\operatorname{label}_R(x)\\}$ the $R$ -fooling rate of the mapping $\\hat{a}$ , and we will denote it by $F_R(\\hat{a})$ . A mapping $\\hat{a} : \\mathbf{X} \\to \\mathbf{X}$ is called an adversarial Doppelgänger attack to a classifier $R$ if and only if $\\hat{a}(x) \\stackrel{\\alpha \\delta}{\\approx} x, \\forall x \\in \\mathbf{X}$ , and the $R$ -fooling rate $F_R(\\hat{a})$ is positive.", + "bbox": [ + 511, + 464, + 905, + 575 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "The set $\\{x:\\mathrm{label}_R(\\hat{a} (x))\\neq \\mathrm{label}_R(x)\\}$ is a subset of the region of conceptual ambiguity of $R$ , which yields an upper bound on the $R$ -fooling rate by the outer measure of $A(R)$ ..", + "bbox": [ + 511, + 583, + 905, + 642 + ], + "page_idx": 6 + }, + { + "type": "equation", + "text": "\n$$\nF _ {R} (\\hat {a}) \\leq \\mu_ {*} (A (R)). \\tag {19}\n$$\n", + "text_format": "latex", + "bbox": [ + 638, + 643, + 903, + 660 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "In specific scenarios it is possible to get an upper bound on the size, possibly the outer measure, of $A(R)$ which in turn shows that the $R$ -fooling rates are bounded away from one. See Example 11 in Part J in the Appendix.", + "bbox": [ + 511, + 665, + 905, + 724 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Adversarial Doppelgänger attacks are distinct from the adversarial attacks studied to date. The universal adversarial attacks, [47], can achieve fooling rates as close to one as one desires. As illustrated above, adversarial Doppelgänger attacks may not be able to reach fooling rates that are too high. On the other hand in some cases, the optimal fooling rate of one can be achieved.", + "bbox": [ + 511, + 726, + 905, + 830 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Observation 5. If $R$ is conceptually ambiguous at every $x \\in \\mathbf{X}$ , e.g., when $H_{R}(x) > 0$ for every $x \\in \\mathbf{X}$ , then there exists an adversarial Doppelgänger attack with $R$ -fooling rate equal to one.", + "bbox": [ + 511, + 839, + 905, + 900 + ], + "page_idx": 6 + }, + { + "type": "page_number", + "text": "10250", + "bbox": [ + 480, + 944, + 517, + 955 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Indeed, if $R$ is conceptually ambiguous at every $x \\in \\mathbf{X}$ , then $\\{y \\in \\mathfrak{d}(x) : \\operatorname{label}_R(y) \\neq \\operatorname{label}_R(x)\\} \\neq \\emptyset$ , for every $x$ , and therefore the axiom of choice implies that there exists a map $\\hat{a} : \\mathbf{X} \\to \\mathbf{X}$ such that $\\hat{a}(x) \\in \\{y \\in \\mathfrak{d}(x) : \\operatorname{label}_R(y) \\neq \\operatorname{label}_R(x)\\}$ , for every $x \\in X$ . It turns out that in practice there may be many classifiers that are conceptually ambiguous at every input.", + "bbox": [ + 89, + 90, + 480, + 196 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Example 5: Consider the case when $\\stackrel{\\alpha\\delta}{\\approx}$ is transitive, i.e., $\\stackrel{\\alpha\\delta}{\\approx}$ equals its transitive closure $\\sim_{\\sigma}$ , e.g., when $\\mathbf{X}$ is equipped with a Grassmann structure as in [39], and there exist at least two different equivalence classes. Thus binary classification is a well defined problem. There exist at least $\\prod_{\\zeta \\in \\mathbf{X} / \\sim_{\\sigma}} \\{U \\subset \\zeta : 0 < \\mu(U) < \\mu(\\zeta)\\}$ worth of fully-populated binary classifiers which are conceptually ambiguous at every input $x \\in \\mathbf{X}$ .", + "bbox": [ + 89, + 196, + 480, + 330 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "The last example and Observation 5 show that even high accuracy classifiers can be vulnerable to adversarial Doppelgänger attacks with fooling rate equal to one.", + "bbox": [ + 89, + 332, + 480, + 378 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "We conclude this section with a warning that popular methods to deal with unseen data, including marking missing data and imputation, may introduce conceptual ambiguity. For example, if a model is trained on a data set $T \\subset \\mathbf{X}$ that includes only parts of some elementary sets, then adding a class label NA to label unseen data can compromise adversarial Doppelgänger robustness. Indeed, let $S$ be a nonempty set of training data such that $S \\subsetneq \\zeta \\in \\mathbf{X} / \\sim_{\\sigma}$ . There exists, $x \\in S$ such that $\\mathfrak{d}(x) \\setminus S \\neq \\emptyset$ . Every $z \\in \\mathfrak{d}(x) \\setminus S$ , labeled as NA, is an adversarial Doppelgänger of $x$ .", + "bbox": [ + 89, + 378, + 480, + 542 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "8. Discussion and Conclusions.", + "text_level": 1, + "bbox": [ + 89, + 558, + 349, + 573 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "A central focus of this paper is the adversarial Doppelgangers phenomenon, where classifiers assign different labels to inputs that humans cannot discriminate. Until now, this phenomenon has not been well understood, possibly due to the limitations of the distance-based analysis that has dominated the field. In the \"absence of a distance measure that accurately captures the perceptual differences between a source and adversarial example many researchers have decided to use the $\\ell_p$ distance\", [29]. The available empirical observations and models - both perceptual and cognitive, including those based on just noticeable differences - provide no evidence that biologically plausible perceptual topologies are metric. This paper advances the understanding of context-related perceptual topologies in input spaces, which are rarely metric. Our investigation shows that adversarial Doppelgangers are very close to each other with respect to the context-relevant perceptual metric $d_w$ , this metric is not a manifold metric and does not generate the perceptual topology. This distinction highlights the shortcomings of traditional, purely manifold metric-based representations and analysis of perceptual spaces.", + "bbox": [ + 89, + 583, + 482, + 900 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "The machine learning community has expended significant efforts aimed to build adversarially robust classifiers. This may be a march towards a bridge too far. Philosophers, experimental psychologists, and linguists, are well aware that many classification problems are not well defined due to perceptual ambiguities. Any fully populated classifier for a classification problem that is not well defined is doomed to be a victim of the adversarial Doppelgängers phenomenon. Our results reveal the structure of adversarial Doppelgänger-robust classifiers, regular classifiers, and criteria and methods to establish whether a classification problem is well defined or not. The new understanding of the structure of regular classifiers, the analysis of zones of ambiguity, and the methods to measure and bound the fooling rates of adversarial Doppelgänger attacks provide guidance on how to design adversarially robust training to improve classifiers that are not regular. In addition to revealing the impossibility to use accuracy-robustness trade-offs in many scenarios, including robustifying hypersensitive classifiers, our analysis indicates that marking unseen data can jeopardize robustness if the training data contains only a proper subset of an elementary set.", + "bbox": [ + 511, + 90, + 903, + 422 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "We explore feature representations, the related concept of indiscernibility introduced by Leibniz, and their connection to indiscriminability. This investigation reveals the nature of class prototypes and fringe inputs, and how the size of a discriminative feature representation can be used to determine whether a classification problem is not well defined. Indiscernibility and indiscriminability, are often conflated in the machine learning literature. Elucidating the distinction between them is vital for understanding the limitations of current classifiers and addressing the shortcomings in their design.", + "bbox": [ + 511, + 422, + 903, + 588 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Our discussion of the Doppelgängers phenomenon brings to light a significant divergence between human perception and artificial neural network models, including feedforward models, RNN models and ResNet. The indiscriminability relations of these artificial neural network models, studied in [18], are transitive16 while it is well accepted, that, in many contexts, the human indiscriminability relation is not transitive.", + "bbox": [ + 511, + 589, + 903, + 708 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "The results and insights gained from this investigation point to concrete warnings and actionable steps for improving the training and testing of classifiers.", + "bbox": [ + 511, + 709, + 903, + 755 + ], + "page_idx": 7 + }, + { + "type": "page_footnote", + "text": "16See Part K in the Appendix.", + "bbox": [ + 527, + 886, + 687, + 898 + ], + "page_idx": 7 + }, + { + "type": "page_number", + "text": "10251", + "bbox": [ + 480, + 944, + 517, + 955 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "References", + "text_level": 1, + "bbox": [ + 91, + 89, + 187, + 104 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[1] Zeyuan Allen-Zhu and Yuanzhi Li. Feature purification: How adversarial training performs robust deep learning. In 2021 IEEE 62nd Annual Symposium on Foundations of Computer Science (FOCS), pages 977-988. IEEE, 2022. 2", + "[2] Valtteri Arstila. Why the transitivity of perceptual simultaneity should be taken seriously. Frontiers in integrative neuroscience, 6:3, 2012. 3", + "[3] Benjamin Balas, Lisa Nakano, and Ruth Rosenholtz. A summary-statistic representation in peripheral vision explains visual crowding. Journal of vision, 9(12):13-13, 2009. 2, 3", + "[4] Max Black. The identity of indiscernibles. Mind, 61(242): 153-164, 1952. 1", + "[5] Nicholas Boucher, Ilia Shumailov, Ross Anderson, and Nicolas Papernot. Bad characters: Imperceptible nlp attacks. In 2022 IEEE Symposium on Security and Privacy (SP), pages 1987-2004. IEEE, 2022. 2", + "[6] William F Broderick, Gizem Rufo, Jonathan Winawer, and Eero P Simoncelli. Foveated metamers of the early visual system. bioRxiv, pages 2023-05, 2023. 1, 2, 3, 6", + "[7] Fan RK Chung. Lectures on spectral graph theory. CBMS Lectures, Fresno, 6(92):17-21, 1996. 6", + "[8] Samuel Clarke. A Collection of Papers, which Passed Between the Late Learned Mr. Leibnitz, and Dr. Clarke, in the Years 1715 and 1716: Relating to the Principles of Natural Philosophy and Religion. With an Appendix. To which are Added, Letters to Dr. Clarke Concerning Liberty and Necessity; from a Gentleman of the University of Cambridge: with the Doctor's Answers to Them. Also Remarks Upon a Book, Entitled, A Philosophical Enquiry Concerning Human Liberty. James Knapton, at the Crown in St. Paul's Church Yard., 1717. 3", + "[9] Jozef B Cohen and William E Kappauf. Color mixture and fundamental metamers: Theory, algebra, geometry, application. The American journal of psychology, pages 171-259, 1985. 2, 3", + "[10] Rafael De Clercq and Leon Horsten. Perceptual indiscriminability: In defence of Wright's proof. The Philosophical Quarterly, 54(216):439-444, 2004. 3", + "[11] Lieven Decock and Igor Douven. Similarity after Goodman. Review of philosophy and psychology, 2:61-75, 2011. 2", + "[12] Kaize Ding, Albert Jiongqian Liang, Bryan Perrozi, Ting Chen, Ruoxi Wang, Lichan Hong, Ed H. Chi, Huan Liu, and Derek Zhiyuan Cheng. Hyperformer: Learning expressive sparse feature representations via hypergraph transformer, 2023. 2", + "[13] Igor Douven. Putting prototypes in place. Cognition, 193: 104007, 2019. 6", + "[14] Michael Elad. Sparse and redundant representations: from theory to applications in signal and image processing. Springer Science & Business Media, 2010. 2", + "[15] William K Estes. Classification and cognition. Oxford University Press, 1994. 7", + "[16] Katalin Farkas. The subject's point of view. OUP Oxford, 2010. 2, 3" + ], + "bbox": [ + 93, + 114, + 483, + 898 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[17] Alhussein Fawzi, Hamza Fawzi, and Omar Fawzi. Adversarial vulnerability for any classifier. Advances in neural information processing systems, 31, 2018. 2", + "[18] Jenelle Feather, Alex Durango, Ray Gonzalez, and Josh McDermott. Metamers of neural networks reveal divergence from human perceptual systems. Advances in Neural Information Processing Systems, 32, 2019. 2, 3, 8, 7", + "[19] Gustav Theodor Fechner. Elemente der psychophysik. Breitkopf u. Härtel, 1860. 3", + "[20] Gustav Theodor Fechner. Revision der hauptpunkte der Psychophysik. Breitkopf und Härtel, 1882. 3", + "[21] Peter Forrest. The Identity of Indiscernibles. In *The Stanford Encyclopedia of Philosophy*. Metaphysics Research Lab, Stanford University, Winter 2020 edition, 2020. 3", + "[22] Jeremy Freeman and Eero P Simoncelli. Metamers of the ventral stream. Nature neuroscience, 14(9):1195-1201, 2011. 2, 3", + "[23] Peter Gardenfors. Conceptual spaces: The geometry of thought. MIT press, 2004. 2", + "[24] Amir Globerson and Sam Roweis. Nightmare at test time: robust learning by feature deletion. In Proceedings of the 23rd international conference on Machine learning, pages 353-360, 2006. 2", + "[25] Ian J Goodfellow, Jonathon Shlens, and Christian Szegedy. Explaining and harnessing adversarial examples. arXiv preprint arXiv:1412.6572, 2014. 2", + "[26] Nelson Goodman. The structure of appearance. Harvard Univ. Press, 1951. 3", + "[27] Delia Graff. Phenomenal continua and the sorites. Mind, 110 (440):905-936, 2001. 3", + "[28] Hermann Grassmann. Zur theorie der farbenmischung. Annalen der Physik, 165(5):69-84, 1853. 3", + "[29] Jamie Hayes and George Danezis. Learning universal adversarial perturbations with generative models. In 2018 IEEE Security and Privacy Workshops (SPW), pages 43-49. IEEE, 2018. 8", + "[30] Martin N Hebart, Charles Y Zheng, Francisco Pereira, and Chris I Baker. Revealing the multidimensional mental representations of natural objects underlying human similarity judgements. Nature human behaviour, 4(11):1173-1185, 2020. 2", + "[31] Eleanor R Heider. \"Focal\" color areas and the development of color names. Developmental psychology, 4(3):447, 1971. 6", + "[32] Benj Hellie. Noise and perceptual indiscriminability. Mind, 114(455):481-508, 2005. 3", + "[33] Christopher James Henry. Near sets: theory and applications. University of Manitoba (Canada), 2011. 2", + "[34] Felix Hovsepian. A metalogical analysis of vagueness: an exploratory study into the geometry of logic. PhD thesis, University of Warwick, 1992. 1", + "[35] Ke Huang and Selin Aviyente. Sparse representation for signal classification. Advances in neural information processing systems, 19, 2006. 2", + "[36] Andrew Ilyas, Shibani Santurkar, Dimitris Tsipras, Logan Engstrom, Brandon Tran, and Aleksander Madry. Adversarial examples are not bugs, they are features. Advances in neural information processing systems, 32, 2019. 2" + ], + "bbox": [ + 516, + 92, + 906, + 900 + ], + "page_idx": 8 + }, + { + "type": "page_number", + "text": "10252", + "bbox": [ + 480, + 944, + 517, + 955 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[37] Akshay V Jagadeesh and Justin L Gardner. Texture-like representation of objects in human visual cortex. Proceedings of the National Academy of Sciences, 119(17):e2115302119, 2022. 1, 2, 3", + "[38] Ralph Kopperman. All topologies come from generalized metrics. The American Mathematical Monthly, 95(2):89-97, 1988. 1", + "[39] David H Krantz. Color measurement and color theory: I. representation theorem for grassmann structures. Journal of Mathematical Psychology, 12(3):283-303, 1975. 2, 3, 8", + "[40] Dekang Lin et al. An information-theoretic definition of similarity. In ICML, pages 296-304, 1998. 6", + "[41] Alexander D Logvinenko. Object-colour manifold. International journal of computer vision, 101:143-160, 2013. 2, 3", + "[42] Alexander D. Logvinenko and Eugene Demidenko. On counting metamers. IEEE Transactions on Image Processing, 25(2):770-775, 2016. 2, 3", + "[43] Aleksander Madry, Aleksandar Makelov, Ludwig Schmidt, Dimitris Tsipras, and Adrian Vladu. Towards deep learning models resistant to adversarial attacks. arXiv preprint arXiv:1706.06083, 2017. 2", + "[44] Saeed Mahloujifar, Dimitrios I Diochnos, and Mohammad Mahmoody. The curse of concentration in robust learning: Evasion and poisoning attacks from concentration of measure. In Proceedings of the AAAI Conference on Artificial Intelligence, pages 4536-4543, 2019. 2", + "[45] Douglas L Medin, Robert L Goldstone, and Dedre Gentner. Respects for similarity. Psychological review, 100(2):254, 1993. 2, 6", + "[46] Alexius Meinong. Über die Bedeutung des Weber'schen Gesetzes: Beiträge zur Psychologie des Vergleichens und Messens. L. Voss, 1896. 3", + "[47] Seyed-Mohsen Moosavi-Dezfooli, Alhussein Fawzi, Omar Fawzi, and Pascal Frossard. Universal adversarial perturbations. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 1765-1773, 2017. 2, 7", + "[48] Weili Nie, Brandon Guo, Yujia Huang, Chaowei Xiao, Arash Vahdat, and Anima Anandkumar. Diffusion models for adversarial purification. arXiv preprint arXiv:2205.07460, 2022. 5", + "[49] Bruno A Olshausen and David J Field. Sparse coding with an overcomplete basis set: A strategy employed by V1? Vision research, 37(23):3311-3325, 1997. 2", + "[50] Marwan Omar, Soohyeon Choi, DaeHun Nyang, and David Mohaisen. Robust natural language processing: Recent advances, challenges, and future directions. IEEE Access, 2022. 5", + "[51] Wilhelm Ostwald. Physikalische Farbenlehre. Verlag Unesma, 1919. 2, 3", + "[52] Nicolas Papernot, Patrick McDaniel, Somesh Jha, Matt Fredrikson, Z Berkay Celik, and Ananthram Swami. The limitations of deep learning in adversarial settings. In 2016 IEEE European symposium on security and privacy (EuroS&P), pages 372-387. IEEE, 2016. 2, 5", + "[53] Zdislaw Pawlak. Classification of objects by means of attributes. Polish Academy of Sciences [PAS]. Institute of Computer Science, 1981. 3" + ], + "bbox": [ + 91, + 90, + 483, + 900 + ], + "page_idx": 9 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[54] Charles Pelling. Exactness, inexactness, and the nontransitivity of perceptual indiscriminability. Synthese, 164 (2):289-312, 2008. 3", + "[55] Mika Perälä. Aristotle on perceptual discrimination. *Phronesis*, 63(3):257–292, 2018. 1", + "[56] James F Peters. Near sets. Special theory about nearness of objects. Fundamenta Informaticae, 75(1-4):407-433, 2007. 2", + "[57] James F Peters and Piotr Wasilewski. Tolerance spaces: Origins, theoretical aspects and applications. Information Sciences, 195:211-225, 2012. 1, 2", + "[58] Henri Poincaré. Dernières pensées. Flammarion, 1930. 2", + "[59] Tim Poston. Fuzzy geometry. PhD thesis, University of Warwick, 1971. 2, 3, 1", + "[60] Diana Raffman. Is perceptual indiscriminability nontransitive? Philosophical Topics, 28(1):153-175, 2000. 2", + "[61] Diana Raffman. Indiscriminability and phenomenal continua. Philosophical perspectives, 26:309-322, 2012. 3", + "[62] Fred S Roberts. Tolerance geometry. Notre Dame Journal of Formal Logic, 14(1):68-76, 1973. 2", + "[63] Eleanor Rosch. Cognitive reference points. Cognitive Psychology, 7(4):532-547, 1975. 6", + "[64] Eleanor H Rosch. On the internal structure of perceptual and semantic categories. In Cognitive development and acquisition of language, pages 111-144. Elsevier, 1973. 6, 7", + "[65] Bertrand Russell. The principles of mathematics. Routledge, 2020. 3", + "[66] Julius A Schreider. Equality, Resemblance and Order. Mir Publishers, Moscow, 1975. 2, 4", + "[67] Ali Shafahi, W Ronny Huang, Christoph Studer, Soheil Feizi, and Tom Goldstein. Are adversarial examples inevitable? arXiv preprint arXiv:1809.02104, 2018. 2", + "[68] Herbert A Simon. Models of man; social and rational. Wiley, 1957. 3", + "[69] Alexei B Sossinsky. Tolerance space theory and some applications. Acta Applicandae Mathematica, 5:137-167, 1986. 6", + "[70] Michael W Spratling. Classification using sparse representations: a biologically plausible approach. Biological cybernetics, 108:61-73, 2014. 2", + "[71] Dong Su, Huan Zhang, Hongge Chen, Jinfeng Yi, Pin-Yu Chen, and Yupeng Gao. Is robustness the cost of accuracy? a comprehensive study on the robustness of 18 deep image classification models. In Proceedings of the European Conference on Computer Vision (ECCV), pages 631-648, 2018. 2, 5", + "[72] Jeremias Sulam, Ramchandran Muthukumar, and Raman Arora. Adversarial robustness of supervised sparse coding. Advances in Neural Information Processing Systems, 33:2110-2121, 2020. 2", + "[73] Christian Szegedy, Wojciech Zaremba, Ilya Sutskever, Joan Bruna, Dumitru Erhan, Ian Goodfellow, and Rob Fergus. Intriguing properties of neural networks. arXiv preprint arXiv:1312.6199, 2013. 1", + "[74] Thomas Tanay and Lewis Griffin. A boundary tilting perspective on the phenomenon of adversarial examples. arXiv preprint arXiv:1608.07690, 2016. 2" + ], + "bbox": [ + 516, + 90, + 906, + 900 + ], + "page_idx": 9 + }, + { + "type": "page_number", + "text": "10253", + "bbox": [ + 480, + 944, + 517, + 955 + ], + "page_idx": 9 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[75] Yee Whye Teh, Max Welling, Simon Osindero, and Geoffrey E Hinton. Energy-based models for sparse overcomplete representations. Journal of Machine Learning Research, 4 (Dec):1235-1260, 2003. 2", + "[76] William A Thornton. Matching lights, metamers and human visual response. The Journal of Color and Appearance, 2: 23-29, 1973. 2, 3", + "[77] Edward Chace Tolman. *Purposive behavior in animals and men*. Univ of California Press, 1932. 2", + "[78] Dimitris Tsipras, Shibani Santurkar, Logan Engstrom, Alexander Turner, and Aleksander Madry. Robustness may be at odds with accuracy. arXiv preprint arXiv:1805.12152, 2018. 2, 5", + "[79] Amos Tversky. Features of similarity. Psychological review, 84(4):327, 1977. 2, 6, 7", + "[80] Gottfried Wilhelm Freiherr von Leibniz. Leibniz: Discourse on Metaphysics. Open Court Publishing Company, 1902. 3, 2", + "[81] Stella Vosniadou and Andrew Ortony. Similarity and analogical reasoning. Cambridge University Press, 1989. 2", + "[82] Ernst Heinrich Weber. De Pulsu, resorptione, auditu et tactu: Annotationes anatomicae et physiologicae... CF Koehler, 1831. 3", + "[83] Ernst Heinrich Weber. Tastsinn und gemeinschaft. Handworterbuch der Physiologie, 1846. 3", + "[84] Timothy Williamson. Identity and Discrimination. Wiley-Blackwell, 1990. 2, 3", + "[85] Crispin Wright. On the coherence of vague predicates. Synthese, pages 325-365, 1975. 3", + "[86] Günter Wyszecki. Valenzmetrische untersuchung des zusammenhanges zwischen normaler und anomaler trichromasia. 1953. 2, 3", + "[87] Gunther Wyszecki and Walter Stanley Stiles. Color science: concepts and methods, quantitative data and formulae. John Wiley & Sons, 2000. 3", + "[88] Erik C. Zeeman. The topology of the brain and visual perception. In The Topology of 3-Manifolds. Prentice Hall, NJ, 1962. 2", + "[89] E. Christopher Zeeman. Topology of the brain. In Math and Computer Science in Biology and Medicine. H.M. Stationary Office, London, United Kingdom, 1965. 2", + "[90] Hongyang Zhang, Yaodong Yu, Jiantao Jiao, Eric Xing, Laurent El Ghaoui, and Michael Jordan. Theoretically principled trade-off between robustness and accuracy. In International conference on machine learning, pages 7472-7482. PMLR, 2019. 2, 5", + "[91] Zheng Zhang, Yong Xu, Jian Yang, Xuelong Li, and David Zhang. A survey of sparse representation: algorithms and applications. IEEE access, 3:490-530, 2015. 2", + "[92] Li Zhaoping and Keith A May. Psychophysical tests of the hypothesis of a bottom-up saliency map in primary visual cortex. PLoS Computational Biology, 3(4):e62, 2007. 3" + ], + "bbox": [ + 91, + 90, + 482, + 833 + ], + "page_idx": 10 + }, + { + "type": "page_number", + "text": "10254", + "bbox": [ + 480, + 944, + 517, + 955 + ], + "page_idx": 10 + } +] \ No newline at end of file diff --git a/2025/Doppelgangers and Adversarial Vulnerability/85f12467-2b9c-44c2-8a35-40f17a213b17_model.json b/2025/Doppelgangers and Adversarial Vulnerability/85f12467-2b9c-44c2-8a35-40f17a213b17_model.json new file mode 100644 index 0000000000000000000000000000000000000000..d753e50379997120f84c59374f737a3d5931d5c0 --- /dev/null +++ b/2025/Doppelgangers and Adversarial Vulnerability/85f12467-2b9c-44c2-8a35-40f17a213b17_model.json @@ -0,0 +1,3082 @@ +[ + [ + { + "type": "header", + "bbox": [ + 0.107, + 0.003, + 0.182, + 0.043 + ], + "angle": 0, + "content": "CVF" + }, + { + "type": "header", + "bbox": [ + 0.237, + 0.001, + 0.812, + 0.047 + ], + "angle": 0, + "content": "This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore." + }, + { + "type": "title", + "bbox": [ + 0.269, + 0.131, + 0.73, + 0.154 + ], + "angle": 0, + "content": "Doppelgängers and Adversarial Vulnerability" + }, + { + "type": "text", + "bbox": [ + 0.371, + 0.182, + 0.629, + 0.253 + ], + "angle": 0, + "content": "George Kamberov* \nUniversity of Alaska Anchorage \nAnchorage, AK 99508 \ngkamberov@alaska.edu" + }, + { + "type": "title", + "bbox": [ + 0.249, + 0.287, + 0.327, + 0.303 + ], + "angle": 0, + "content": "Abstract" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.318, + 0.484, + 0.425 + ], + "angle": 0, + "content": "Machine learning (ML) classifiers can make mistakes that are perceptually and cognitively disturbing to humans. The most notorious examples of such errors are adversarial visual metamers. This paper investigates the phenomenon of adversarial Doppelgängers (AD), which encompasses adversarial visual metamers, and compares the performance and robustness of ML classifiers to human performance." + }, + { + "type": "text", + "bbox": [ + 0.093, + 0.425, + 0.485, + 0.726 + ], + "angle": 0, + "content": "We find that ADs are inputs that are close to each other with respect to a perceptual metric defined in this paper, and show that ADs are qualitatively different from the usual adversarial examples. The vast majority of classifiers are vulnerable to ADs and robustness-accuracy trade-offs may not improve them. Some classification problems do not admit any AD-robust classifiers because the underlying classes are ambiguous. We provide criteria to determine whether a classification problem is well defined; describe the structure and attributes of AD-robust classifiers; introduce and explore the notions of conceptual entropy and regions of conceptual ambiguity for classifiers that are vulnerable to AD attacks; and discuss methods to bound the AD fooling rate of an attack. We define the notion of classifiers that exhibit hypersensitive behavior, that is, classifiers whose only mistakes are adversarial Doppelgangers. Improving the AD robustness of hypersensitive classifiers is equivalent to improving accuracy. We identify conditions guaranteeing that all classifiers with sufficiently high accuracy are hypersensitive." + }, + { + "type": "title", + "bbox": [ + 0.092, + 0.756, + 0.222, + 0.772 + ], + "angle": 0, + "content": "1. Introduction" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.782, + 0.483, + 0.843 + ], + "angle": 0, + "content": "Perceptual metamers1 are the most striking adversarial examples studied by the machine learning community. Two perceptual metamers are shown in Figure 1. The phenomenon of metamersim studied in the visual domain, in" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.288, + 0.907, + 0.425 + ], + "angle": 0, + "content": "cluding perceptual metamers, is a manifestation of the existence of Doppelgängers: different inputs or stimuli that are perceptually indiscriminable. The research community has engaged in active studies of adversarial vulnerability ever since the publication of [73]. Adversarial Doppelgängers, that is, adversarial examples which are Doppelgängers, are qualitatively different from the vast majority of known adversarial examples which humans readily discriminate from correctly classified input samples (Figure 2). There is no" + }, + { + "type": "image", + "bbox": [ + 0.547, + 0.445, + 0.709, + 0.565 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.617, + 0.571, + 0.635, + 0.582 + ], + "angle": 0, + "content": "(a)" + }, + { + "type": "image", + "bbox": [ + 0.719, + 0.445, + 0.871, + 0.565 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.779, + 0.571, + 0.797, + 0.582 + ], + "angle": 0, + "content": "(b)" + }, + { + "type": "image_caption", + "bbox": [ + 0.513, + 0.595, + 0.907, + 0.636 + ], + "angle": 0, + "content": "Figure 1. Most people cannot discriminate image (a) from image (b). MobileNetV2 classifies the later image as \"persian\" and the former picture as \"taby\"." + }, + { + "type": "image", + "bbox": [ + 0.547, + 0.692, + 0.709, + 0.814 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.594, + 0.818, + 0.662, + 0.83 + ], + "angle": 0, + "content": "(a) Labrador" + }, + { + "type": "image", + "bbox": [ + 0.714, + 0.692, + 0.871, + 0.814 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.747, + 0.818, + 0.832, + 0.83 + ], + "angle": 0, + "content": "(b) Weimeraner" + }, + { + "type": "image_caption", + "bbox": [ + 0.512, + 0.842, + 0.907, + 0.884 + ], + "angle": 0, + "content": "Figure 2. Applying a Fast Signed Gradient perturbation to the image (a) classified by MobileNetV2 as Labrador yields the image (b) which is classified by MobileNetV2 as Weimeraner." + }, + { + "type": "page_footnote", + "bbox": [ + 0.09, + 0.851, + 0.483, + 0.901 + ], + "angle": 0, + "content": "*This material is based upon work supported by the National Science Foundation under Grant No. 2433241 \n1“images that are physically distinct but perceptually indistinguishable”, [6]. See also “metameric images”, [37]" + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.52, + 0.957 + ], + "angle": 0, + "content": "10244" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.089, + 0.092, + 0.482, + 0.227 + ], + "angle": 0, + "content": "evidence that Doppelgängers can be studied and understood completely using the \\(\\ell_p\\) norms or more general geodesic distances on manifolds, which have been employed to quantify sample differences and to investigate adversarial examples. Perception and context impose topology on a space of inputs, but it rarely aligns with a manifold topology. We will denote this context-relative topology by \\(\\tau_{\\delta}\\). It is defined by the context-relative ability to acquire and deploy knowledge." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.228, + 0.483, + 0.348 + ], + "angle": 0, + "content": "In this paper, we explore the context-relative perceptual topologies on a space of inputs \\(\\mathbf{X}\\) and examine the vulnerability and robustness of machine learning classifiers to adversarial Doppelgangers. We show that, while the majority of classifiers are vulnerable to adversarial Doppelgangers, safe (Doppelganger robust) classifiers do exist if the classification problem is well defined. However, these robust classifiers may be very rare." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.348, + 0.483, + 0.56 + ], + "angle": 0, + "content": "In Section 3, we discuss the context-relative notion of (active) indiscriminability and the topology \\(\\tau_{\\delta}\\) that it induces on a space of inputs. The separability and metric properties of various motivating examples of \\(\\tau_{\\delta}\\). Additionally, we show that the distances between Doppelgängers are small if measured by a perceptually-based context-relative metric. We also examine the relation between indiscriminability, IDSicriminability, and feature representations. The existence and structure of Doppelgänger robust classifiers is discussed in Section 4. In Section 5, we investigate the relationship between Doppelgängers and misclassified input samples, define the notion of hypersensitive behavior, and show that improved adversarial Doppelgänger robustness does not have to lead to a reduction in accuracy." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.56, + 0.483, + 0.681 + ], + "angle": 0, + "content": "The structure of perceptually regular, i.e., AD robust classifiers are discussed in Section 6. By definition, a classifier is not regular if and only if some inputs can be attacked by adversarial Doppelgangers. Not surprisingly, it turns out that some inputs are more vulnerable than others. In Section 7, we provide measures of adversarial Doppelgangers vulnerability and upper bounds on the fooling rate of an adversarial Doppelganger attack." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.694, + 0.228, + 0.709 + ], + "angle": 0, + "content": "2. Related work" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.718, + 0.483, + 0.809 + ], + "angle": 0, + "content": "The pair \\((\\mathbf{X}, \\text{indiscriminability relation})\\) is a tolerance space. Tolerance spaces, rough sets, and granular computing have been discussed extensively. See [59, 62, 66, 88, 89]. The color and image metamers studied by many authors including [3, 6, 9, 18, 22, 37, 39, 41, 42, 51, 76, 86] are Doppelgangers." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.81, + 0.483, + 0.856 + ], + "angle": 0, + "content": "The research on adversarial examples to date builds on the hypothesis that the space of input samples is a metric space \\((\\mathbf{X},\\mathrm{dist}_{\\mathbf{X}})\\). A misclassified input \\(x^{*}\\) is consid" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.092, + 0.905, + 0.227 + ], + "angle": 0, + "content": "ered an adversarial example if it is nearby a correctly classified input sample \\( x \\), i.e., \\( \\mathrm{dist}_{\\mathbf{X}}(x, x^{*}) \\) is small. Usually \\( \\mathbf{X} \\) is assumed to be \\( \\mathbb{R}^n \\), endowed with the \\( \\ell_p \\), norm, \\( p = 1, 2, \\ldots, \\infty \\) or at least locally homeomorphic to \\( \\mathbb{R}^n \\), i.e., a manifold, equipped with some geodesic distance. Somewhat non surprisingly many authors have shown that every classifier can be attacked with such adversarial examples [5, 17, 25, 52] or at least that this is true in many contexts, [44, 47, 67]." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.228, + 0.905, + 0.379 + ], + "angle": 0, + "content": "Other papers indicate that there are paths toward eliminating adversarial examples completely, i.e., it is possible to achieve provable \"adversarial robustness\" by fixing/retraining the classifier [1, 24, 36, 43, 72, 74]. A widely accepted tenet is that \"there is a clear trade-off between accuracy and [adversarial] robustness, and a better performance in testing accuracy in general reduces [adversarial] robustness\", [71]. For empirical evidence for this trade-off and some attempts to explain this phenomenon see [71, 78, 90]." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.392, + 0.71, + 0.41 + ], + "angle": 0, + "content": "3. Perceptual Topology" + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.418, + 0.767, + 0.433 + ], + "angle": 0, + "content": "3.1 Indiscriminability and Topology" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.433, + 0.905, + 0.543 + ], + "angle": 0, + "content": "The ability to decide whether one stimulus/input is distinct from another is essential for adaptation, survival, and intelligent life. Intelligent agents are uniquely capable to activate knowledge to judge distinction. Williamson calls this context-relative process discrimination ([84]) and defines a context-relative symmetric and reflexive binary relation denoted by \\(\\stackrel{\\alpha\\delta}{\\approx}\\) and called indiscriminability:" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.553, + 0.905, + 0.614 + ], + "angle": 0, + "content": "Definition 1 ([84]). Two inputs \\( x \\) and \\( y \\) are called indiscriminable to a subject at a time \\( t \\) if and only if at time \\( t \\) the subject is not able to activate (acquire or employ) the relevant kind of knowledge that \\( x \\) and \\( y \\) are distinct." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.624, + 0.905, + 0.654 + ], + "angle": 0, + "content": "Indiscriminability generates a context-relative topology on the set of inputs \\(\\mathbf{X}\\)." + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.664, + 0.905, + 0.757 + ], + "angle": 0, + "content": "Definition 2. The the phenomenal neighborhood of an input \\(x \\in \\mathbf{X}\\), is the set \\(\\mathfrak{d}(x) = \\left\\{y \\in \\mathbf{X} : y \\stackrel{\\alpha \\delta}{\\approx} x\\right\\}\\). A point \\(y \\in \\mathfrak{d}(x) \\setminus \\{x\\}\\) is called a Doppelgänger [of \\(x\\)]. The perceptual topology \\(\\tau_{\\delta}\\) is the topology generated by the sub-basis \\(\\mathfrak{D}_{\\alpha \\delta} = \\{\\mathfrak{d}(x)\\}_{x \\in \\mathbf{X}}\\)." + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.764, + 0.905, + 0.795 + ], + "angle": 0, + "content": "Example 0: An input \\( x \\in \\mathbf{X} \\) is called optimal if it does not have non-trivial/non-identical Doppelgangers, i.e.," + }, + { + "type": "page_footnote", + "bbox": [ + 0.512, + 0.803, + 0.905, + 0.84 + ], + "angle": 0, + "content": "\\(^{3}\\)Context and its role in discrimination and similarity judgments have been studied extensively and by many authors including [11, 23, 30, 45, 60, 77, 79, 81]." + }, + { + "type": "page_footnote", + "bbox": [ + 0.513, + 0.84, + 0.905, + 0.9 + ], + "angle": 0, + "content": "4Some authors refer to indiscriminability as active indiscriminability, see [16]. Poincaré discusses indiscriminability in [58] but refers to it as indiscriminability. Similarly Poston studies indistinguishability, basing it on the \"limit of discrimination\" of the biological senses and instruments, [59]." + }, + { + "type": "list", + "bbox": [ + 0.512, + 0.803, + 0.905, + 0.9 + ], + "angle": 0, + "content": null + }, + { + "type": "page_footnote", + "bbox": [ + 0.09, + 0.863, + 0.483, + 0.901 + ], + "angle": 0, + "content": "2 Some metamers arising in other fields including biology and chemistry are not Doppelgängers, for example, segments in many earthworms are considered metamers but are visually discriminable." + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.518, + 0.957 + ], + "angle": 0, + "content": "10245" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.09, + 0.09, + 0.482, + 0.231 + ], + "angle": 0, + "content": "if \\(\\mathfrak{d}(x) = \\{x\\}\\). In particular, if \\(\\stackrel{\\infty \\delta}{\\approx}\\) is the identity relationship \\(=\\) (i.e., all inputs are optimal within the given context), then the perceptual topology \\(\\tau_{\\delta}\\) is discrete. However, the finiteness of human observations (they are subject to finite time and finite work constraints) and the bounded rationality constraints imposed by the limitations on the availability of information and computational capabilities to humans, [68], indicate that the scenario \\(\\mathfrak{d}(x) = \\{x\\}\\) for every \\(x \\in \\mathbf{X}\\) may be highly unlikely." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.232, + 0.482, + 0.353 + ], + "angle": 0, + "content": "Often reflexive binary relations are defined and discussed as coverings of the underlying space. Indeed, every reflexive binary relation (RBR) \\(\\approx\\) on \\(\\mathbf{X}\\) defines a covering \\(\\{\\mathfrak{g}(x) = \\{y:y\\approx x\\} \\}_{x\\in \\mathbf{X}}\\), of \\(\\mathbf{X}\\) and vice versa one can define reflexive binary relationships through coverings of \\(\\mathbf{X}\\). See Appendix, Section A. In particular, the perceptual topology, \\(\\tau_{\\delta}\\), is a tolerable topology, cf., Definition 10 in Appendix, Section A." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.355, + 0.482, + 0.431 + ], + "angle": 0, + "content": "The active psychophysics research on just noticeable difference initiated by Weber and Fechner, [19, 20, 82, 83], provides one of the few classes of examples where we have empirically supported understanding of the perceptual topology." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.446, + 0.483, + 0.492 + ], + "angle": 0, + "content": "Example 1: Let \\(\\mathbf{X}\\) be the closed bounded interval \\([a,b]\\subset\\) \\((0, + \\infty)\\). Suppose that Weber's law holds and let \\(k > 0\\) be the Weber constant. Let \\(w = 1 + k\\), then" + }, + { + "type": "equation", + "bbox": [ + 0.151, + 0.506, + 0.482, + 0.561 + ], + "angle": 0, + "content": "\\[\n\\mathfrak {d} (x) = \\left\\{ \\begin{array}{l l} {[ a, x w),} & {a \\leq x < a w} \\\\ {(x / w, x w),} & {a w \\leq x \\leq b / w} \\\\ {(x / w, b ],} & {b / w < x \\leq b.} \\end{array} \\right. \\tag {1}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.575, + 0.483, + 0.635 + ], + "angle": 0, + "content": "The covering \\(\\{\\mathfrak{d}(x)\\}_{x\\in [a,b]}\\) defines a symmetric RBR but the relation is not transitive. The corresponding perceptual topology is \\(T_{0}\\) but not \\(T_{1}\\), and the topology is not pseudometric." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.637, + 0.483, + 0.703 + ], + "angle": 0, + "content": "The transitivity or more more often the lack of transitivity of \\(\\stackrel{\\alpha\\delta}{\\approx}\\) have been studied extensively and proven or postulated in many human experiences, [2, 6, 10, 16, 26, 27, 32, 54, 61, 84, 85]." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.717, + 0.483, + 0.849 + ], + "angle": 0, + "content": "Definition 3. We will denote by \\(\\sim_{\\sigma}\\) the transitive closure of the indiciminability relation \\(\\stackrel{\\alpha \\delta}{\\approx}\\) on \\(\\mathbf{X}\\). It is defined explicitly as \\(x \\sim_{\\sigma} y\\) iff there exists a finite chain of Doppelgangers \\(x = x_0 \\stackrel{\\alpha \\delta}{\\approx} x_1 \\stackrel{\\alpha \\delta}{\\approx} x_2 \\stackrel{\\alpha \\delta}{\\approx} \\dots \\stackrel{\\alpha \\delta}{\\approx} x_n = y\\). We will call the relation \\(\\sim_{\\sigma}\\) perceptual metamorphy and will refer to any two inputs \\(x \\sim_{\\sigma} y\\) as metamorphic. Extending Pawlak's terminology, [53] we call the equivalence classes in \\(\\mathbf{X} / \\sim_{\\sigma}\\) (perceptually) elementary sets." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.092, + 0.905, + 0.307 + ], + "angle": 0, + "content": "Example 2: If the indiscriminability relation is transitive, then each \\(\\mathfrak{d}(x)\\) is an elementary set. The perceptual topology may be optimal (recall Example 0) or not. In the former case it is Hausdorff and in fact \\((\\mathbf{X},\\tau_{\\delta})\\) is a discrete manifold, in the later case the topology \\(\\tau_{\\delta}\\) is not \\(T_0\\). See Part A.1 in the Appendix. Human visual perception provides a fundamental example where \\(\\mathfrak{d}(x)\\neq \\{x\\}\\). If two images \\(x\\) and \\(y\\) differ only in unattended regions for example due to low saliency values (cf. [92]), then \\(x\\stackrel {\\alpha \\delta}{\\approx}y\\). Visual metamers have been studied by many authors, including [3, 6, 9, 18, 22, 28, 37, 39, 41, 42, 51, 76, 86]. In these studies, the input space is assumed to be endowed with Grassmann structure (see [39]), and in particular, the indiscriminability relation is transitive.\\(^6\\)" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.308, + 0.905, + 0.439 + ], + "angle": 0, + "content": "We are not aware of perceptual topologies that are metric. Still, every Doppelgänger \\( y \\stackrel{\\alpha \\delta}{\\approx} x \\) of an input \\( x \\) is a small perturbation of \\( x \\) in the sense that the \\( y \\) is a nearest neighbor of \\( x \\) with respect to an appropriate metric \\( d_w(\\cdot, \\cdot) \\) on \\( \\mathbf{X} \\). Indeed, let the discrimination graph \\( \\Gamma(\\mathbf{X}, E_{\\alpha \\delta}) \\) be the undirected simple graph, where \\( \\mathbf{X} \\) is the set of vertices and we say that there is an edge \\( \\{x, y\\} \\in E_{\\alpha \\delta} \\) between the vertices \\( x, y \\in \\mathbf{X} \\) iff \\( x \\stackrel{\\alpha \\delta}{\\approx} y \\)." + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.447, + 0.905, + 0.524 + ], + "angle": 0, + "content": "Definition 4. We will call the discrimination graph distance \\( d_{\\infty}(x,y) \\) between the vertices \\( x \\) and \\( y \\) and, in particular, \\( d_{\\infty}(x,y) = \\infty \\) iff \\( x \\not\\simeq_{\\sigma} y \\) the extended perceptual distance. The perceptual distance is the metric \\( d_w: \\mathbf{X} \\times \\mathbf{X} \\to [0,1] \\) defined by:" + }, + { + "type": "equation", + "bbox": [ + 0.579, + 0.532, + 0.905, + 0.567 + ], + "angle": 0, + "content": "\\[\nd _ {w} (x, y) = \\frac {d _ {\\infty} (x , y)}{1 + d _ {\\infty} (x , y)}, \\forall x, y \\in \\mathbf {X}. \\tag {2}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.573, + 0.905, + 0.618 + ], + "angle": 0, + "content": "The metric \\( d_w \\) does not generate the perceptual topology. It is certainly not the usual \\( l_p \\) or any other manifold metric used in ML." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.633, + 0.836, + 0.648 + ], + "angle": 0, + "content": "3.2 Indiscriminabile may not be Indiscernible" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.649, + 0.905, + 0.744 + ], + "angle": 0, + "content": "Let \\(\\Phi\\) be the space of all features of the inputs/stimuli \\(x\\in \\mathbf{X}\\) and let \\(\\Phi_x\\subset \\Phi\\) be the set of features attributed to \\(x\\) in a given context. Following [21], we say that \\(x\\) and \\(y\\) are indiscernible, in a given context, if \\(\\Phi_x = \\Phi_y\\). Many researchers use the terms indiscriminability and indiscernibility as synonyms. This is only accurate when \\(\\stackrel{\\alpha\\delta}{\\approx}\\) is transi" + }, + { + "type": "page_footnote", + "bbox": [ + 0.513, + 0.752, + 0.905, + 0.776 + ], + "angle": 0, + "content": "6In these studies, indiscriminable inputs are referred to as \"matching\" or \"metameric\", or \"alike\"." + }, + { + "type": "page_footnote", + "bbox": [ + 0.513, + 0.777, + 0.905, + 0.827 + ], + "angle": 0, + "content": "The open metric ball \\(\\tilde{B}_{1/2}^w(x)\\) equals \\(\\{x\\}\\), for all inputs \\(x \\in \\mathbf{X}\\). On the other hand, the finiteness of human observations and the hypothesis of bounded rationality suggest that biologically plausible perceptual topologies are not discrete." + }, + { + "type": "page_footnote", + "bbox": [ + 0.513, + 0.828, + 0.905, + 0.864 + ], + "angle": 0, + "content": "The existence of the extended metric was hinted at in [46]; it was discussed in a related context in [65] and rediscovered and exploited in [59]. For more discussion see Part F.1 in the Appendix." + }, + { + "type": "page_footnote", + "bbox": [ + 0.513, + 0.864, + 0.905, + 0.901 + ], + "angle": 0, + "content": "9Leibniz discussed indiscernibility and postulated the Principle/Law of Identity of Indiscernibles, \\((\\Phi_{x} = \\Phi_{y})\\Rightarrow (x = y)\\) , in [80] and in the third and fourth papers addressed to Samuel Clarke, [8]." + }, + { + "type": "list", + "bbox": [ + 0.513, + 0.752, + 0.905, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "page_footnote", + "bbox": [ + 0.09, + 0.875, + 0.483, + 0.901 + ], + "angle": 0, + "content": "5Optimal objects and light sources have been described and studied in colorimetry, [42, 87]." + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "10246" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.09, + 0.091, + 0.485, + 0.155 + ], + "angle": 0, + "content": "tive.\\(^{10}\\) In general, the relationship between indiscriminability and indiscernibility is not well understood. However, indiscernibility implies indiscriminability if \\(\\{\\Phi_x\\}_{x\\in \\mathbf{X}}\\) is a perceptually discriminative feature representation, i.e.," + }, + { + "type": "equation", + "bbox": [ + 0.196, + 0.164, + 0.483, + 0.188 + ], + "angle": 0, + "content": "\\[\n\\Phi_ {x} \\bigcap \\Phi_ {y} \\neq \\varnothing \\Longleftrightarrow x \\stackrel {\\circledast} {\\approx} y. \\tag {3}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.196, + 0.485, + 0.304 + ], + "angle": 0, + "content": "The biological plausibility of discriminative feature representations is an open question. Still, they provide insight into the structure of the perceptual topology. The attributed discriminative features represent structures of Doppelgangers. Indeed, let \\(\\{\\Phi_x\\}_{x\\in \\mathbf{X}}\\) be a feature representation and let \\(\\operatorname {cl}(\\xi)\\) be the context-dependent semantic cluster of inputs sharing the feature \\(\\xi \\in \\Phi\\). Specifically," + }, + { + "type": "equation", + "bbox": [ + 0.176, + 0.312, + 0.483, + 0.331 + ], + "angle": 0, + "content": "\\[\n\\mathfrak {c l} (\\xi) = \\left\\{x \\in \\mathbf {X}, \\text {s . t . ,} \\xi \\in \\Phi_ {x} \\right\\}. ^ {1 2} \\tag {4}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.34, + 0.485, + 0.432 + ], + "angle": 0, + "content": "In particular, if \\(\\{\\Phi_x\\}_{x\\in \\mathbf{X}}\\) is a discriminative feature representation, then every attributed discriminative feature \\(\\xi \\in \\Phi_{x}\\) is associated to, and in some way explained by, a collection of Doppelgangers since \\(\\mathfrak{cl}(\\xi)\\subset \\mathfrak{d}(x)\\). For more detailed discussion and examples of discriminative feature representations see Part B in the Appendix." + }, + { + "type": "title", + "bbox": [ + 0.09, + 0.444, + 0.476, + 0.462 + ], + "angle": 0, + "content": "4. Classifiers and Adversarial Doppelgängers." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.469, + 0.484, + 0.56 + ], + "angle": 0, + "content": "A classifier \\( R \\) (with \\( m \\) labels) is called fully populated iff the labeling function \\( \\text{label}_R: \\mathbf{X} \\to \\{1, \\dots, m\\} \\) is surjective mapping onto the range of labels \\( \\{1, \\dots, m\\} \\). For any classifier \\( R \\), with \\( m \\) labels, we will denote by \\( R_c \\) the level set of the labeling function \\( \\text{label}_R \\) for each label \\( c \\in \\{1, \\dots, m\\} \\).13" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.569, + 0.484, + 0.643 + ], + "angle": 0, + "content": "Definition 5. We say that \\( x \\) is a Doppelgänger adversarial to the classifier \\( R \\) iff \\( \\exists y \\in \\mathfrak{d}(x) \\) such that \\( \\operatorname{label}_R(x) \\neq \\operatorname{label}_R(y) \\) and we will refer to both \\( x \\) and \\( y \\) as adversarial Doppelgangers when the classifier \\( R \\) is clear from the context." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.644, + 0.484, + 0.704 + ], + "angle": 0, + "content": "A classifier \\(R\\) is called (perceptually) regular iff it does not admit adversarial Doppelgangers. If \\(R\\) is regular, then \\(\\mathfrak{D}_{\\alpha \\delta} = \\{\\mathfrak{d}(x)\\}_{x \\in \\mathbf{X}}\\) are \\(R\\) coherent coverings (as defined in [66])." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.705, + 0.484, + 0.779 + ], + "angle": 0, + "content": "We say that the classification problem with \\( m \\)-labels is well defined if there exists a fully populated and perceptually regular classifier with \\( m \\) labels. Otherwise we say that the classification problem with \\( m \\)-labels is not well defined." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.092, + 0.905, + 0.195 + ], + "angle": 0, + "content": "The labeling function \\( \\text{label}_R \\) of a regular classifier is continuous with respect to the perceptual topology \\( \\tau_{\\delta} \\). Furthermore, if \\( R \\) is not regular and \\( x \\) is a point of discontinuity of \\( \\text{label}_R: (\\mathbf{X}, \\tau_{\\delta}) \\to \\{1, \\dots, m\\} \\), then \\( x \\) is an adversarial Doppelgänger. The discontinuity is an indication of the cognitive disruption that occurs when one encounters some AD." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.198, + 0.906, + 0.319 + ], + "angle": 0, + "content": "It is well known that in some experiences perceptually unambiguous categories and hence perceptually regular classifiers do not exist. A simple example is provided by the perceptual topology in Example 1. Indeed, in this case \\(\\mathbf{X} / \\sim_{\\sigma}\\) is a singleton, i.e., every two inputs are metamorphic and hence there is a only one elementary set which equals the whole \\(\\mathbf{X}\\). Therefore, every classifier with two or more labels must have adversarial Doppelgängers.[14]" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.319, + 0.906, + 0.41 + ], + "angle": 0, + "content": "Clearly, no amount of \"robust training\" will get rid of adversarial examples of a classifier with a surjective labeling function \\( \\text{label}_R: (\\mathbf{X}, \\tau_{\\delta}) \\to \\{1, \\ldots, m\\} \\) if \\( \\mathbf{X} \\) cannot be broken into \\( m \\) perceptually unambiguous categories. In the rest of this section we investigate the non-existence, existence and internal structure of regular classifiers." + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.41, + 0.906, + 0.455 + ], + "angle": 0, + "content": "In Part D of the Appendix we show a specific example of a well defined classification problem and discuss the actual regular classifier." + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.457, + 0.907, + 0.581 + ], + "angle": 0, + "content": "Example 3: If \\(\\stackrel{\\alpha \\delta}{\\approx}\\) is transitive then for every number of labels \\(m\\) smaller than the number of equivalence classes card \\(\\left(\\mathbf{X} / \\stackrel{\\alpha \\delta}{\\approx}\\right)\\) there exists a fully populated regular classifier with \\(m\\) labels, but if the number of labels \\(m\\) is bigger than card \\(\\left(\\mathbf{X} / \\stackrel{\\alpha \\delta}{\\approx}\\right)\\) then every fully populated classifier with \\(m\\) labels must admit adversarial Doppelgangers." + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.595, + 0.906, + 0.626 + ], + "angle": 0, + "content": "Example 3 indicates that if the transitive closure \\(\\sim_{\\sigma}\\) is trivial, i.e., \\(\\sim_{\\sigma} = \\mathbf{X} \\times \\mathbf{X}\\), then no label is safe. Namely:" + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.636, + 0.906, + 0.764 + ], + "angle": 0, + "content": "Observation 1. If the transitive closure \\(\\sim_{\\sigma}\\) of the indiscriminability relation \\(\\stackrel{\\alpha \\delta}{\\approx}\\) is trivial, then every fully populated classifier with two or more classes admits adversarial Doppelgängers. In particular, let \\(R\\) be a fully populated classifier with a surjective labeling function \\(\\text{label}_R: X \\to \\{1, 2, \\ldots, m\\}\\), then for every label \\(c\\) there exist adversarial Doppelgängers \\(x(c) \\in \\mathbf{X}\\) and \\(x^*(c) \\in \\mathfrak{d}(x(c))\\) such that \\(c = \\text{label}_R(x(c))\\) and \\(\\text{label}_R(x^*(c)) \\neq \\text{label}_R(x(c))\\)." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.773, + 0.905, + 0.834 + ], + "angle": 0, + "content": "The proof follows from the fact that every finite chain of Doppelgängers connecting points that are labeled differently by a classifier must contain a pair of adversarial Doppelgängers. See Lemma 2 in Appendix Section C." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.835, + 0.906, + 0.865 + ], + "angle": 0, + "content": "A straight-forward argument shows that if \\( R \\) is a perceptually regular fully populated classifier, then \\( x \\in R_i \\) iff" + }, + { + "type": "page_footnote", + "bbox": [ + 0.513, + 0.875, + 0.905, + 0.9 + ], + "angle": 0, + "content": "14See Lemma 3 and the short argument that follows it in Appendix, Section C." + }, + { + "type": "page_footnote", + "bbox": [ + 0.107, + 0.786, + 0.358, + 0.798 + ], + "angle": 0, + "content": "10See Observation 7, in Part B of the Appendix." + }, + { + "type": "page_footnote", + "bbox": [ + 0.093, + 0.799, + 0.483, + 0.851 + ], + "angle": 0, + "content": "11A feature \\(\\xi \\in \\Phi\\) is called attributed if \\(\\xi \\in \\Phi_x\\) for some input \\(x \\in \\mathbf{X}\\). It is plausible that \\(\\Phi = \\bigcup_{x \\in \\mathbf{X}} \\Phi_x\\), and so all features are attributed. However, many models do not preclude the existence of spurious latent traits." + }, + { + "type": "page_footnote", + "bbox": [ + 0.093, + 0.851, + 0.483, + 0.888 + ], + "angle": 0, + "content": "12 The semantic cluster of inputs sharing the feature \\(\\xi \\in \\Phi\\) is defined for any feature representation. A feature \\(\\xi\\) is attributed in a given context, iff \\(\\operatorname{cl}(\\xi) \\neq \\emptyset\\); a feature is a hypothetical feature, when \\(\\operatorname{cl}(\\xi) = \\emptyset\\)." + }, + { + "type": "page_footnote", + "bbox": [ + 0.106, + 0.888, + 0.456, + 0.901 + ], + "angle": 0, + "content": "\\(^{13}R\\) is fully populated iff \\(R_{c}\\neq \\emptyset\\) for every label \\(c\\in \\{1,\\ldots ,m\\}\\)" + }, + { + "type": "list", + "bbox": [ + 0.093, + 0.786, + 0.483, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "10247" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.091, + 0.091, + 0.483, + 0.123 + ], + "angle": 0, + "content": "\\([x]_{\\sim_{\\sigma}} \\subset R_i\\), i.e., each level set is a disjoint union of elementary sets \\([x]_{\\sim_{\\sigma}}\\). The pigeonhole principle yields:" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.131, + 0.484, + 0.221 + ], + "angle": 0, + "content": "Observation 2. If the number of equivalence classes card \\((\\mathbf{X} / \\sim_{\\sigma}) \\geq 2\\), then for every natural number \\(2 \\leq m \\leq\\) card \\((\\mathbf{X} / \\sim_{\\sigma})\\) there exists a perceptually regular fully populated classifier with \\(m\\) labels. However, if \\(m >\\) card \\((\\mathbf{X} / \\sim_{\\sigma})\\), then every fully populated classifier with \\(m\\) labels must have adversarial Doppelgangers." + }, + { + "type": "text", + "bbox": [ + 0.089, + 0.222, + 0.484, + 0.282 + ], + "angle": 0, + "content": "In particular, if \\( p = \\operatorname{card}(\\mathbf{X} / \\sim_{\\sigma}) \\geq 2 \\) is finite, then for every natural number \\( m \\leq p \\) there are exactly \\( S(p, m) \\) regular fully populated classifiers with \\( m \\) classes. Here \\( S(p, m) \\) is the Sterling number of the second kind." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.291, + 0.484, + 0.398 + ], + "angle": 0, + "content": "Observation 2 shows that the problem of finding a fully populated perceptually unambiguous classifier \\( R \\) with precisely \\( m \\) labels is not well defined if card \\( (\\mathbf{X} / \\sim_{\\sigma}) < m \\) and vice versa that the same problem is well defined for every \\( m \\) such that card \\( (\\mathbf{X} / \\sim_{\\sigma}) \\geq m \\). In the former case solutions do not exist while in the later case solutions exist and each class segment \\( R_{i} \\) is a union of equivalence classes \\( [x]_{\\sim_{\\sigma}} \\)," + }, + { + "type": "equation", + "bbox": [ + 0.228, + 0.408, + 0.483, + 0.441 + ], + "angle": 0, + "content": "\\[\nR _ {i} = \\bigcup_ {x \\in R _ {i}} [ x ] _ {\\sim_ {\\sigma}}. \\tag {5}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.449, + 0.484, + 0.601 + ], + "angle": 0, + "content": "The existence and properties of discriminative feature representations provide insight whether a classification problem is well defined. In particular, if there exists a discriminative feature representation and the set of attributed features is finite, then every classification problem, whose number of labels exceeds the number of attributed features, is not well defined. See Observation 9 in Appendix Part B.2. Further discussion of the structure of class segments of a regular classifier including the class/category core and fringe are discussed in Section 6." + }, + { + "type": "title", + "bbox": [ + 0.09, + 0.614, + 0.476, + 0.631 + ], + "angle": 0, + "content": "5. Accuracy and Adversarial Doppelgängers." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.639, + 0.484, + 0.79 + ], + "angle": 0, + "content": "The accuracy-adversarial robustness trade off observed and discussed in the literature involves various measures of accuracy [48, 50, 71, 78, 90]. We will discuss classification accuracy. We will show that there is a strong relationship between classifier accuracy and vulnerability to adversarial Doppelgängers. In particular, we will identify (perceptual) scenarios in which low accuracy classifiers are critically vulnerable to adversarial Doppelgänger attacks but on the other hand all high accuracy classifiers can be fooled only by Doppelgängers." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.799, + 0.312, + 0.815 + ], + "angle": 0, + "content": "5.1. The Probabilistic Setup." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.82, + 0.484, + 0.903 + ], + "angle": 0, + "content": "We will assume that \\((\\mathbf{X},\\mathcal{F},\\mu)\\) is a probability measure space equipped with perceptual topology \\(\\tau_{\\delta}\\) generated by an indsicriminability relation \\(\\stackrel{\\alpha \\delta}{\\approx}\\) such that for every \\(x \\in \\mathbf{X}\\) the set of Doppelgangers \\(\\mathfrak{d}(x)\\) and the equivalence class \\([x]_{\\sim_{\\sigma}}\\) are events, \\((\\mathfrak{d}(x) \\in \\mathcal{F}\\) and \\([x]_{\\sim_{\\sigma}} \\in \\mathcal{F}, \\forall x \\in \\mathbf{X})\\)." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.092, + 0.907, + 0.227 + ], + "angle": 0, + "content": "In the rest of this section we will assume that the classification problem with \\(m \\geq 2\\) labels is well defined and let \\(\\Omega\\) be a perceptually regular classifier; we will reserve the notation \\(R\\) to denote any classifier (\\(R\\) may or may not be perceptually regular) such that \\(R_{i} \\cap \\Omega_{i}\\), \\(i = 1, \\ldots, m\\) are the true positives (of class \\(i\\)). To define accuracy we will focus only on regular models and classifiers s.t., \\(\\Omega_{i} \\in \\mathcal{F}\\) and \\(R_{i} \\in \\mathcal{F}\\), for all \\(i = 1, \\ldots, m\\). The accuracy of the classifier \\(R\\) defined as" + }, + { + "type": "equation", + "bbox": [ + 0.523, + 0.241, + 0.907, + 0.26 + ], + "angle": 0, + "content": "\\[\n\\operatorname {a c c u r a c y} _ {\\Omega} (R) = \\mu \\left(R _ {1} \\cap \\Omega_ {1}\\right) + \\dots + \\mu \\left(R _ {m} \\cap \\Omega_ {m}\\right). \\tag {6}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.27, + 0.907, + 0.301 + ], + "angle": 0, + "content": "Furthermore let us assume that \\(\\mu (\\Omega_i) > 0\\) for every \\(i = 1,\\ldots ,m\\) and thus we can define recall rates" + }, + { + "type": "equation", + "bbox": [ + 0.595, + 0.312, + 0.906, + 0.347 + ], + "angle": 0, + "content": "\\[\n\\rho_ {i} = \\frac {\\mu \\left(R _ {i} \\cap \\Omega_ {i}\\right)}{\\mu \\left(\\Omega_ {i}\\right)}, \\quad i = 1, \\dots , m \\tag {7}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.358, + 0.906, + 0.387 + ], + "angle": 0, + "content": "Bounds on the recall rates imply bounds on the accuracy. Namely if" + }, + { + "type": "equation", + "bbox": [ + 0.614, + 0.39, + 0.805, + 0.407 + ], + "angle": 0, + "content": "\\[\n\\underline {{\\rho}} \\leq \\rho_ {i} \\leq \\bar {\\rho}, \\quad i = 1, \\dots , m\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.414, + 0.831, + 0.43 + ], + "angle": 0, + "content": "then since \\(\\mu\\) is a probability measure on \\(X\\), and" + }, + { + "type": "equation", + "bbox": [ + 0.576, + 0.442, + 0.907, + 0.46 + ], + "angle": 0, + "content": "\\[\n\\mu \\left(R _ {i} \\cap \\Omega_ {i}\\right) = \\rho_ {i} \\mu \\left(\\Omega_ {i}\\right), \\quad i = 1, \\dots , m \\tag {8}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.514, + 0.474, + 0.564, + 0.487 + ], + "angle": 0, + "content": "we get" + }, + { + "type": "equation", + "bbox": [ + 0.577, + 0.501, + 0.843, + 0.54 + ], + "angle": 0, + "content": "\\[\n\\underline {{\\rho}} \\leq \\operatorname {a c c u r a c y} _ {\\Omega} (R) = \\sum_ {i = 1} ^ {m} \\rho_ {i} \\mu (\\Omega_ {i}) \\leq \\bar {\\rho}.\n\\]" + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.552, + 0.743, + 0.567 + ], + "angle": 0, + "content": "5.2. Are Trade-Offs Possible?" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.575, + 0.906, + 0.607 + ], + "angle": 0, + "content": "Let \\( i(x) \\in \\{1, \\ldots, m\\} \\) be the object class label of \\( x \\in \\mathbf{X} \\) i.e., \\( x \\in \\Omega_{i(x)} \\) and" + }, + { + "type": "equation", + "bbox": [ + 0.614, + 0.619, + 0.906, + 0.66 + ], + "angle": 0, + "content": "\\[\n\\bar {k} (\\Omega) = \\sup _ {x \\in \\mathbf {X}} \\left(\\frac {\\mu \\left(\\Omega_ {i (x)}\\right)}{\\mu (\\mathfrak {d} (x))}\\right). \\tag {9}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.673, + 0.907, + 0.72 + ], + "angle": 0, + "content": "Every classifier whose recall rates do not exceed \\(1 / \\bar{k} (\\Omega)\\) is totally unsafe in the sense that every correctly classified input admits adversarial Doppelgangers. Specifically:" + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.731, + 0.907, + 0.813 + ], + "angle": 0, + "content": "Observation 3. Suppose that the sets of Doppelgangers are not negligible and \\(\\inf_{x\\in \\mathbf{X}}\\mu (\\mathfrak{d}(x)) > 0\\) , and let \\(\\Omega = \\{\\Omega_1,\\dots ,\\Omega_m\\}\\) be a regular world model and let \\(R = \\{R_{1},\\ldots ,R_{m}\\}\\) be a classifier whose recall rates are strictly smaller than \\(1 / \\bar{k} (\\Omega)\\) and so" + }, + { + "type": "equation", + "bbox": [ + 0.601, + 0.825, + 0.905, + 0.86 + ], + "angle": 0, + "content": "\\[\n\\frac {\\mu \\left(R _ {i (x)} \\cap \\mathfrak {d} (x)\\right)}{\\mu (\\mathfrak {d} (x))} \\leq \\bar {\\rho} \\bar {k} (\\Omega) < 1. \\tag {10}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.871, + 0.906, + 0.902 + ], + "angle": 0, + "content": "Thus every correctly classified input \\( x \\) has adversarial Doppelgängers." + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "10248" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.09, + 0.092, + 0.483, + 0.168 + ], + "angle": 0, + "content": "Observation 3 shows that sacrificing accuracy may lead to increasing the probability of encountering adversarial Doppelgangers and in fact that there is no trade off for accuracies that are sufficiently low provided that all sets of Doppelgangers have positive measure." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.168, + 0.483, + 0.214 + ], + "angle": 0, + "content": "The lack of an opportunity for a trade-off is even more striking when one tries to improve high recall rate (and hence high accuracy) classifiers." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.225, + 0.483, + 0.279 + ], + "angle": 0, + "content": "Observation 4. \\(^{15}\\) If \\(\\inf_{x\\in \\mathbf{X}}\\mu (\\mathfrak{d}(x)) > 0\\) , and let \\(R =\\) \\(\\{R_1,\\ldots ,R_m\\}\\) be a classifier whose recall rates are sufficiently high so that \\(\\underline{\\rho} >1 - 1 / \\bar{k} (\\Omega)\\) . i.e.," + }, + { + "type": "equation", + "bbox": [ + 0.225, + 0.291, + 0.483, + 0.312 + ], + "angle": 0, + "content": "\\[\n\\left(1 - \\underline {{\\rho}}\\right) \\bar {k} (\\Omega) < 1. \\tag {11}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.322, + 0.483, + 0.355 + ], + "angle": 0, + "content": "Then every misclassified input \\( x \\) is an adversarial Doppelgänger." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.365, + 0.484, + 0.412 + ], + "angle": 0, + "content": "Definition 6. We say that a classifier has a hypersensitive behavior if every misclassified input is an adversarial Doppelgänger." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.423, + 0.483, + 0.514 + ], + "angle": 0, + "content": "For classifiers with hypersensitive behavior adversarial robustness can only be improved by improving accuracies, i.e., by eliminating misclassification. Observation 4 shows that if \\(\\mu (\\mathfrak{d}(x)) > 0\\), for all inputs \\(x\\in \\mathbf{X}\\), then all classifiers with sufficiently high accuracy are either regular (when accuracy equals to one) or have hypersensitive behavior." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.515, + 0.484, + 0.576 + ], + "angle": 0, + "content": "In summary adversarial Doppelgänger robustness - accuracy trade-off may happen for classifiers with middling accuracy rates or when there are inputs whose Doppelgängers are negligible in measure." + }, + { + "type": "title", + "bbox": [ + 0.09, + 0.591, + 0.295, + 0.607 + ], + "angle": 0, + "content": "6. Life without borders." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.617, + 0.483, + 0.768 + ], + "angle": 0, + "content": "A important property of the perceptual topology is that the if a classifier \\(\\Omega\\) is perceptually regular, then it \"imposes an open [topological] borders policy\", that is, \\(\\partial \\Omega_{c} = \\emptyset\\) for every label \\(c\\). Linguists and psychologists, have observed and postulated that natural perceptual and semantic categories are borderless. See for example [64]. Class/decision boundaries are studied and exploited in many works on classifiers, and in particular on adversarial robustness. These boundaries are artifacts of the metric topology used by the researchers, they are not perceptual phenomena." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.77, + 0.483, + 0.875 + ], + "angle": 0, + "content": "However, we are all familiar with the idea that some stimuli are more intrinsic/representative to/of a given class and at the same time frequently there are objects/stimuli/Input that, while they are firmly with in the class, are less representative/share few(er) features compared to the rest of the elements in the class, that is, they \"are/belong to the fringe\" of the class." + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.092, + 0.905, + 0.152 + ], + "angle": 0, + "content": "Definition 7. Let \\( s \\) be a similarity scale, i.e., a function \\( s: \\mathbf{X} \\times \\mathbf{X} \\to \\mathbb{R} \\) such that \\( s(x, x) \\geq s(x, y), \\forall x, y \\in \\mathbf{X} \\) as in [79], [45] (measuring similarity within a fixed context) and [40]." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.16, + 0.907, + 0.266 + ], + "angle": 0, + "content": "The values \\( s(x, x) \\) can be and sometimes are used to represent the salience or equivalently importance of the input within \\( \\mathbf{X} \\), see for example [79]. The similarity scale provides a method to quantify the affinity of a input/stimulus to a given measurable subset \\( D \\subset \\mathbf{X} \\) and the notions of prototype and fringe. The (s-)affinity of \\( x \\) with a measurable set \\( D \\) is defined as" + }, + { + "type": "equation", + "bbox": [ + 0.633, + 0.271, + 0.905, + 0.305 + ], + "angle": 0, + "content": "\\[\nP (x, D) = \\int_ {D} s (x, y) \\tag {12}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.311, + 0.907, + 0.342 + ], + "angle": 0, + "content": "This is a straightforward generalization of the notion of prototypicality defined in [79]." + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.349, + 0.907, + 0.38 + ], + "angle": 0, + "content": "Definition 8. \\( x \\in D \\) is called a prototype of \\( D \\) (with respect to the integrable similarity scale \\( s: \\mathbf{X} \\times \\mathbf{X} \\to \\mathbb{R} \\)) if" + }, + { + "type": "equation", + "bbox": [ + 0.624, + 0.394, + 0.905, + 0.419 + ], + "angle": 0, + "content": "\\[\nP (x, D) = \\sup _ {z \\in D} P (z, D) \\tag {13}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.423, + 0.905, + 0.453 + ], + "angle": 0, + "content": "\\(x \\in D\\) is called a fringe element of \\(D\\) (with respect to the integrable similarity scale \\(s: \\mathbf{X} \\times \\mathbf{X} \\to \\mathbb{R}\\)) if" + }, + { + "type": "equation", + "bbox": [ + 0.624, + 0.463, + 0.905, + 0.485 + ], + "angle": 0, + "content": "\\[\nP (x, D) = \\inf _ {z \\in D} P (z, D) \\tag {14}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.494, + 0.906, + 0.629 + ], + "angle": 0, + "content": "One is tempted to think of prototypes as the stimuli that are \"clearest cases, best examples\". \"easy to tell apart\" and to \"be a good representative\" and hence that optimize salience [13, 31, 63, 64, 79]. However, clearly there is no reason to expect that a stimulus that is unlikely to be observed/encountered would be selected as a prototype. The examples below show that prototypes and fringes are obtained by optimizing a mixture of the frequency of appearance (likelihood to encounter) and salience." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.63, + 0.906, + 0.705 + ], + "angle": 0, + "content": "The core and fringe sets may be empty. In practice one is often satisfied with finding elements that may not be true prototypes but belong to the \\(M\\)-core, i.e., have affinity exceeding a fixed threshold \\(M\\). Similarly, elements whose affinity is below a given threshold \\(\\tau\\) belong to the \\(\\tau\\)-fringe." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.706, + 0.907, + 0.902 + ], + "angle": 0, + "content": "Many perceptual and cognitive processes exploit the intelligent agents' abilities to measure and/or compare the salience/importance of features. In some simple cases it is expected and it even might be true that salience/importance is a probability measure \\( f_{\\Phi} \\) on the space of (all possible) features \\( \\Phi \\) and the feature representations \\( \\Phi_x \\) are measurable subsets of \\( \\Phi \\). In many accounts including Tversky's feature contrast model [79] the salience scale \\( f_{\\Phi} \\) is a context dependent, nonnegative function defined on a collection \\( \\Upsilon(\\Phi) \\) of subsets of \\( \\Phi \\) which is closed under finite unions and intersections, and set differences. Furthermore, \\( \\Phi_x \\in \\Upsilon(\\Phi), \\forall x \\in \\mathbf{X} \\); and the non-negative function \\( f_{\\Phi} \\) is feature additive, i.e., \\( f_{\\Phi}(A \\cup B) = f_{\\Phi}(A) + \\)" + }, + { + "type": "page_footnote", + "bbox": [ + 0.105, + 0.887, + 0.29, + 0.901 + ], + "angle": 0, + "content": "15The proof is in Appendix Part E." + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "10249" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.09, + 0.091, + 0.484, + 0.271 + ], + "angle": 0, + "content": "\\(f_{\\Phi}(B)\\), if \\(A \\cap B = \\emptyset\\). The value \\(f(x) = f_{\\phi}(\\Phi_x), \\forall x \\in \\mathbf{X}\\) is called the salience/prominence of the input \\(x\\), [79]. We will call a (Tversky) salience scale \\(f_{\\Phi}\\) perceptually regular if the prominence/salience function \\(f(x) = f_{\\phi}(\\Phi_x)\\) is perceptually regular. We will call a perceptually regular salience scale \\(f_{\\Phi}\\) fully deployable if it can be used to judge the distinguishing features of inputs/stimuli. In particular, we can use \\(f_{\\Phi}(\\Phi_x \\setminus \\Phi_y)\\) to discriminate \\(x\\) from \\(y\\). Thus if \\(f_{\\Phi}\\) is fully deployable, then \\(f_{\\Phi}(\\Phi_x \\setminus \\Phi_y) = 0\\) and \\(f_{\\Phi}(\\Phi_y \\setminus \\Phi_x) = 0\\) whenever \\(x \\stackrel{\\alpha \\delta}{\\approx} y\\) or equivalently \\(f_{\\Phi}(\\Phi_x \\cap \\Phi_y) = f(x)\\) if \\(x \\stackrel{\\alpha \\delta}{\\approx} y\\)." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.275, + 0.484, + 0.321 + ], + "angle": 0, + "content": "The following special case provides a particularly useful insight into the nature of prototypes and the possible internal structure of categories as discussed in [64]." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.34, + 0.484, + 0.417 + ], + "angle": 0, + "content": "Example 4: If \\( s \\) is a contrast similarity [79] such that \\( f_{\\Phi}(\\Phi_x) = s(x,x) / \\theta \\) is fully deployable, \\( \\theta \\in (0, + \\infty) \\), and furthermore, \\( f_{\\Phi}(\\Phi_x \\cap \\Phi_y) = f_{\\Phi}(\\Phi_x) \\) for every pair \\( x \\sim_{\\sigma} y \\) and \\( \\Phi_x \\cap \\Phi_y = \\emptyset \\) if \\( x \\not\\sim_{\\sigma} y \\), then for every regular class \\( D \\subset \\mathbf{X} \\), and every \\( x \\in D \\) we get" + }, + { + "type": "equation", + "bbox": [ + 0.091, + 0.437, + 0.484, + 0.482 + ], + "angle": 0, + "content": "\\[\nP (x, D) = \\Theta \\mu (D) \\left(\\frac {\\mu ([ x ] _ {\\sim_ {\\sigma}})}{\\mu (D)} - \\frac {\\alpha}{\\Theta}\\right) f _ {\\Phi} \\left(\\Phi_ {x}\\right) - \\beta I _ {\\Phi} (D), \\tag {15}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.484, + 0.484, + 0.574 + ], + "angle": 0, + "content": "where \\(\\Theta = (\\alpha +\\beta +\\theta)\\) and \\(\\alpha ,\\beta\\) ,and \\(\\theta\\) are non-negative constants. In particular, if \\(D\\) is a finite union of equivalence classes \\(\\zeta_j\\in \\mathbf{X} / \\sim_\\sigma\\) , then both prototypes and fringe elements exist. Furthermore, if \\(x\\) is prototype/fringe then so are all stimuli in its component \\(\\zeta = [x]_{\\sim_{\\sigma}}\\) . Similar statements hold for M-core and \\(\\tau\\) -fringe elements." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.579, + 0.484, + 0.732 + ], + "angle": 0, + "content": "If the class \\(D\\) is fixed, then \\(\\frac{\\mu([x]_{\\sim_{\\sigma}})}{\\mu(D)}\\) is just the probability to encounter (and learn) a set of stimuli, and \\(f_{\\Phi}(\\Phi_x)\\) can be interpreted as the level of prominence. So the optimization process involves learning prominent examples that can be encountered reasonably often. Thus in this case the prototype does not fall in either of the two main branches of prototypes, that is inputs that represent the central tendency in the regular class vs. prototypes as highly \"representative exemplar(s) of a category\", see [15], page 52." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.735, + 0.484, + 0.901 + ], + "angle": 0, + "content": "More generally, when the similarity scale is bounded, for example, this is true for real similarity measures deployable by humans, then, as predicted by modern prototype theory, each perceptually regular subset \\( D \\subset \\mathbf{X} \\) corresponding to real (natural) categories created and analyzed by real intelligent agents consists of core elements (possibly M-core), a layer of fringe (possibly \\( \\tau \\)-fringe) elements, and layers of elements of various levels of intermediate affinity with \\( D \\). In Part I of the Appendix we introduce class invariants of regular classifiers including structural entropy, expected index of coincidence, and importance." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.09, + 0.88, + 0.109 + ], + "angle": 0, + "content": "7. Quantifying Doppelgänger Vulnerability." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.115, + 0.906, + 0.197 + ], + "angle": 0, + "content": "By definition if a classifier \\( R \\) is not regular, then it is vulnerable to adversarial Doppelgangers attacks, that is, for some input \\( x \\) there exists \\( a(x) \\stackrel{\\alpha \\delta}{\\approx} x \\) and such that \\( \\operatorname{label}_R(x) \\neq \\operatorname{label}_R(a(x)) \\). In particular, we say that \\( R \\) is conceptually ambiguous at \\( x \\) and we call the set" + }, + { + "type": "equation", + "bbox": [ + 0.521, + 0.207, + 0.905, + 0.241 + ], + "angle": 0, + "content": "\\[\n\\mathrm {A} (R) = \\left\\{x \\in \\mathbf {X}: \\exists y \\stackrel {\\alpha \\delta} {\\approx} x \\text {a n d} \\operatorname {l a b e l} _ {R} (x) \\neq \\operatorname {l a b e l} _ {R} (y) \\right\\} \\tag {16}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.242, + 0.906, + 0.287 + ], + "angle": 0, + "content": "the region of conceptual ambiguity. When \\((X, \\mu)\\) is a probability measure space and \\(\\mu(\\mathfrak{d}(x)) > 0\\), we use the probability distribution of labels at \\(x\\):" + }, + { + "type": "equation", + "bbox": [ + 0.571, + 0.295, + 0.905, + 0.33 + ], + "angle": 0, + "content": "\\[\np _ {j} (x) = \\frac {\\mu \\left(R _ {j} \\bigcap \\mathfrak {d} (x)\\right)}{\\mu (\\mathfrak {d} (x))}, \\quad j = 1, \\dots , m \\tag {17}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.337, + 0.838, + 0.352 + ], + "angle": 0, + "content": "and the conceptual entropy of \\( R \\) at \\( x \\) defined as" + }, + { + "type": "equation", + "bbox": [ + 0.598, + 0.361, + 0.905, + 0.402 + ], + "angle": 0, + "content": "\\[\nH _ {R} (x) = - \\sum_ {j = 1} ^ {m} p _ {j} (x) \\log \\left(p _ {j} (x)\\right) \\tag {18}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.411, + 0.906, + 0.457 + ], + "angle": 0, + "content": "to detect whether \\( R \\) is conceptually ambiguous at \\( x \\) (i.e., \\( H_{R}(x) > 0 \\)), and to quantify the likelihoods of various adversarial Doppelgänger attacks." + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.465, + 0.906, + 0.576 + ], + "angle": 0, + "content": "Definition 9. Let \\( R \\) be a classifier and \\( \\hat{a} : \\mathbf{X} \\to \\mathbf{X} \\), we will call the inner measure of the set \\( \\{x : \\operatorname{label}_R(\\hat{a}(x)) \\neq \\operatorname{label}_R(x)\\} \\) the \\( R \\)-fooling rate of the mapping \\( \\hat{a} \\), and we will denote it by \\( F_R(\\hat{a}) \\). A mapping \\( \\hat{a} : \\mathbf{X} \\to \\mathbf{X} \\) is called an adversarial Doppelgänger attack to a classifier \\( R \\) if and only if \\( \\hat{a}(x) \\stackrel{\\alpha \\delta}{\\approx} x, \\forall x \\in \\mathbf{X} \\), and the \\( R \\)-fooling rate \\( F_R(\\hat{a}) \\) is positive." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.584, + 0.906, + 0.643 + ], + "angle": 0, + "content": "The set \\(\\{x:\\mathrm{label}_R(\\hat{a} (x))\\neq \\mathrm{label}_R(x)\\}\\) is a subset of the region of conceptual ambiguity of \\(R\\) , which yields an upper bound on the \\(R\\) -fooling rate by the outer measure of \\(A(R)\\) .." + }, + { + "type": "equation", + "bbox": [ + 0.64, + 0.644, + 0.905, + 0.661 + ], + "angle": 0, + "content": "\\[\nF _ {R} (\\hat {a}) \\leq \\mu_ {*} (A (R)). \\tag {19}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.666, + 0.906, + 0.726 + ], + "angle": 0, + "content": "In specific scenarios it is possible to get an upper bound on the size, possibly the outer measure, of \\( A(R) \\) which in turn shows that the \\( R \\)-fooling rates are bounded away from one. See Example 11 in Part J in the Appendix." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.727, + 0.906, + 0.831 + ], + "angle": 0, + "content": "Adversarial Doppelgänger attacks are distinct from the adversarial attacks studied to date. The universal adversarial attacks, [47], can achieve fooling rates as close to one as one desires. As illustrated above, adversarial Doppelgänger attacks may not be able to reach fooling rates that are too high. On the other hand in some cases, the optimal fooling rate of one can be achieved." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.84, + 0.906, + 0.901 + ], + "angle": 0, + "content": "Observation 5. If \\( R \\) is conceptually ambiguous at every \\( x \\in \\mathbf{X} \\), e.g., when \\( H_{R}(x) > 0 \\) for every \\( x \\in \\mathbf{X} \\), then there exists an adversarial Doppelgänger attack with \\( R \\)-fooling rate equal to one." + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "10250" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.09, + 0.092, + 0.482, + 0.198 + ], + "angle": 0, + "content": "Indeed, if \\( R \\) is conceptually ambiguous at every \\( x \\in \\mathbf{X} \\), then \\( \\{y \\in \\mathfrak{d}(x) : \\operatorname{label}_R(y) \\neq \\operatorname{label}_R(x)\\} \\neq \\emptyset \\), for every \\( x \\), and therefore the axiom of choice implies that there exists a map \\( \\hat{a} : \\mathbf{X} \\to \\mathbf{X} \\) such that \\( \\hat{a}(x) \\in \\{y \\in \\mathfrak{d}(x) : \\operatorname{label}_R(y) \\neq \\operatorname{label}_R(x)\\} \\), for every \\( x \\in X \\). It turns out that in practice there may be many classifiers that are conceptually ambiguous at every input." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.197, + 0.482, + 0.332 + ], + "angle": 0, + "content": "Example 5: Consider the case when \\(\\stackrel{\\alpha\\delta}{\\approx}\\) is transitive, i.e., \\(\\stackrel{\\alpha\\delta}{\\approx}\\) equals its transitive closure \\(\\sim_{\\sigma}\\), e.g., when \\(\\mathbf{X}\\) is equipped with a Grassmann structure as in [39], and there exist at least two different equivalence classes. Thus binary classification is a well defined problem. There exist at least \\(\\prod_{\\zeta \\in \\mathbf{X} / \\sim_{\\sigma}} \\{U \\subset \\zeta : 0 < \\mu(U) < \\mu(\\zeta)\\}\\) worth of fully-populated binary classifiers which are conceptually ambiguous at every input \\(x \\in \\mathbf{X}\\)." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.333, + 0.482, + 0.379 + ], + "angle": 0, + "content": "The last example and Observation 5 show that even high accuracy classifiers can be vulnerable to adversarial Doppelgänger attacks with fooling rate equal to one." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.379, + 0.482, + 0.543 + ], + "angle": 0, + "content": "We conclude this section with a warning that popular methods to deal with unseen data, including marking missing data and imputation, may introduce conceptual ambiguity. For example, if a model is trained on a data set \\( T \\subset \\mathbf{X} \\) that includes only parts of some elementary sets, then adding a class label NA to label unseen data can compromise adversarial Doppelgänger robustness. Indeed, let \\( S \\) be a nonempty set of training data such that \\( S \\subsetneq \\zeta \\in \\mathbf{X} / \\sim_{\\sigma} \\). There exists, \\( x \\in S \\) such that \\( \\mathfrak{d}(x) \\setminus S \\neq \\emptyset \\). Every \\( z \\in \\mathfrak{d}(x) \\setminus S \\), labeled as NA, is an adversarial Doppelgänger of \\( x \\)." + }, + { + "type": "title", + "bbox": [ + 0.09, + 0.559, + 0.35, + 0.574 + ], + "angle": 0, + "content": "8. Discussion and Conclusions." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.584, + 0.483, + 0.901 + ], + "angle": 0, + "content": "A central focus of this paper is the adversarial Doppelgangers phenomenon, where classifiers assign different labels to inputs that humans cannot discriminate. Until now, this phenomenon has not been well understood, possibly due to the limitations of the distance-based analysis that has dominated the field. In the \"absence of a distance measure that accurately captures the perceptual differences between a source and adversarial example many researchers have decided to use the \\(\\ell_p\\) distance\", [29]. The available empirical observations and models - both perceptual and cognitive, including those based on just noticeable differences - provide no evidence that biologically plausible perceptual topologies are metric. This paper advances the understanding of context-related perceptual topologies in input spaces, which are rarely metric. Our investigation shows that adversarial Doppelgangers are very close to each other with respect to the context-relevant perceptual metric \\(d_w\\), this metric is not a manifold metric and does not generate the perceptual topology. This distinction highlights the shortcomings of traditional, purely manifold metric-based representations and analysis of perceptual spaces." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.092, + 0.905, + 0.424 + ], + "angle": 0, + "content": "The machine learning community has expended significant efforts aimed to build adversarially robust classifiers. This may be a march towards a bridge too far. Philosophers, experimental psychologists, and linguists, are well aware that many classification problems are not well defined due to perceptual ambiguities. Any fully populated classifier for a classification problem that is not well defined is doomed to be a victim of the adversarial Doppelgängers phenomenon. Our results reveal the structure of adversarial Doppelgänger-robust classifiers, regular classifiers, and criteria and methods to establish whether a classification problem is well defined or not. The new understanding of the structure of regular classifiers, the analysis of zones of ambiguity, and the methods to measure and bound the fooling rates of adversarial Doppelgänger attacks provide guidance on how to design adversarially robust training to improve classifiers that are not regular. In addition to revealing the impossibility to use accuracy-robustness trade-offs in many scenarios, including robustifying hypersensitive classifiers, our analysis indicates that marking unseen data can jeopardize robustness if the training data contains only a proper subset of an elementary set." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.424, + 0.905, + 0.589 + ], + "angle": 0, + "content": "We explore feature representations, the related concept of indiscernibility introduced by Leibniz, and their connection to indiscriminability. This investigation reveals the nature of class prototypes and fringe inputs, and how the size of a discriminative feature representation can be used to determine whether a classification problem is not well defined. Indiscernibility and indiscriminability, are often conflated in the machine learning literature. Elucidating the distinction between them is vital for understanding the limitations of current classifiers and addressing the shortcomings in their design." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.59, + 0.905, + 0.709 + ], + "angle": 0, + "content": "Our discussion of the Doppelgängers phenomenon brings to light a significant divergence between human perception and artificial neural network models, including feedforward models, RNN models and ResNet. The indiscriminability relations of these artificial neural network models, studied in [18], are transitive16 while it is well accepted, that, in many contexts, the human indiscriminability relation is not transitive." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.71, + 0.905, + 0.756 + ], + "angle": 0, + "content": "The results and insights gained from this investigation point to concrete warnings and actionable steps for improving the training and testing of classifiers." + }, + { + "type": "page_footnote", + "bbox": [ + 0.528, + 0.887, + 0.688, + 0.9 + ], + "angle": 0, + "content": "16See Part K in the Appendix." + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.518, + 0.957 + ], + "angle": 0, + "content": "10251" + } + ], + [ + { + "type": "title", + "bbox": [ + 0.093, + 0.09, + 0.188, + 0.106 + ], + "angle": 0, + "content": "References" + }, + { + "type": "ref_text", + "bbox": [ + 0.101, + 0.115, + 0.484, + 0.171 + ], + "angle": 0, + "content": "[1] Zeyuan Allen-Zhu and Yuanzhi Li. Feature purification: How adversarial training performs robust deep learning. In 2021 IEEE 62nd Annual Symposium on Foundations of Computer Science (FOCS), pages 977-988. IEEE, 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.101, + 0.173, + 0.484, + 0.213 + ], + "angle": 0, + "content": "[2] Valtteri Arstila. Why the transitivity of perceptual simultaneity should be taken seriously. Frontiers in integrative neuroscience, 6:3, 2012. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.216, + 0.484, + 0.269 + ], + "angle": 0, + "content": "[3] Benjamin Balas, Lisa Nakano, and Ruth Rosenholtz. A summary-statistic representation in peripheral vision explains visual crowding. Journal of vision, 9(12):13-13, 2009. 2, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.273, + 0.482, + 0.299 + ], + "angle": 0, + "content": "[4] Max Black. The identity of indiscernibles. Mind, 61(242): 153-164, 1952. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.302, + 0.483, + 0.357 + ], + "angle": 0, + "content": "[5] Nicholas Boucher, Ilia Shumailov, Ross Anderson, and Nicolas Papernot. Bad characters: Imperceptible nlp attacks. In 2022 IEEE Symposium on Security and Privacy (SP), pages 1987-2004. IEEE, 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.36, + 0.483, + 0.401 + ], + "angle": 0, + "content": "[6] William F Broderick, Gizem Rufo, Jonathan Winawer, and Eero P Simoncelli. Foveated metamers of the early visual system. bioRxiv, pages 2023-05, 2023. 1, 2, 3, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.403, + 0.483, + 0.429 + ], + "angle": 0, + "content": "[7] Fan RK Chung. Lectures on spectral graph theory. CBMS Lectures, Fresno, 6(92):17-21, 1996. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.432, + 0.484, + 0.568 + ], + "angle": 0, + "content": "[8] Samuel Clarke. A Collection of Papers, which Passed Between the Late Learned Mr. Leibnitz, and Dr. Clarke, in the Years 1715 and 1716: Relating to the Principles of Natural Philosophy and Religion. With an Appendix. To which are Added, Letters to Dr. Clarke Concerning Liberty and Necessity; from a Gentleman of the University of Cambridge: with the Doctor's Answers to Them. Also Remarks Upon a Book, Entitled, A Philosophical Enquiry Concerning Human Liberty. James Knapton, at the Crown in St. Paul's Church Yard., 1717. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.572, + 0.483, + 0.625 + ], + "angle": 0, + "content": "[9] Jozef B Cohen and William E Kappauf. Color mixture and fundamental metamers: Theory, algebra, geometry, application. The American journal of psychology, pages 171-259, 1985. 2, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.629, + 0.483, + 0.669 + ], + "angle": 0, + "content": "[10] Rafael De Clercq and Leon Horsten. Perceptual indiscriminability: In defence of Wright's proof. The Philosophical Quarterly, 54(216):439-444, 2004. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.672, + 0.482, + 0.699 + ], + "angle": 0, + "content": "[11] Lieven Decock and Igor Douven. Similarity after Goodman. Review of philosophy and psychology, 2:61-75, 2011. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.701, + 0.483, + 0.768 + ], + "angle": 0, + "content": "[12] Kaize Ding, Albert Jiongqian Liang, Bryan Perrozi, Ting Chen, Ruoxi Wang, Lichan Hong, Ed H. Chi, Huan Liu, and Derek Zhiyuan Cheng. Hyperformer: Learning expressive sparse feature representations via hypergraph transformer, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.772, + 0.482, + 0.798 + ], + "angle": 0, + "content": "[13] Igor Douven. Putting prototypes in place. Cognition, 193: 104007, 2019. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.802, + 0.483, + 0.842 + ], + "angle": 0, + "content": "[14] Michael Elad. Sparse and redundant representations: from theory to applications in signal and image processing. Springer Science & Business Media, 2010. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.845, + 0.482, + 0.871 + ], + "angle": 0, + "content": "[15] William K Estes. Classification and cognition. Oxford University Press, 1994. 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.874, + 0.482, + 0.899 + ], + "angle": 0, + "content": "[16] Katalin Farkas. The subject's point of view. OUP Oxford, 2010. 2, 3" + }, + { + "type": "list", + "bbox": [ + 0.094, + 0.115, + 0.484, + 0.899 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.093, + 0.905, + 0.134 + ], + "angle": 0, + "content": "[17] Alhussein Fawzi, Hamza Fawzi, and Omar Fawzi. Adversarial vulnerability for any classifier. Advances in neural information processing systems, 31, 2018. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.136, + 0.905, + 0.189 + ], + "angle": 0, + "content": "[18] Jenelle Feather, Alex Durango, Ray Gonzalez, and Josh McDermott. Metamers of neural networks reveal divergence from human perceptual systems. Advances in Neural Information Processing Systems, 32, 2019. 2, 3, 8, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.191, + 0.905, + 0.217 + ], + "angle": 0, + "content": "[19] Gustav Theodor Fechner. Elemente der psychophysik. Breitkopf u. Härtel, 1860. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.219, + 0.905, + 0.245 + ], + "angle": 0, + "content": "[20] Gustav Theodor Fechner. Revision der hauptpunkte der Psychophysik. Breitkopf und Härtel, 1882. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.247, + 0.905, + 0.287 + ], + "angle": 0, + "content": "[21] Peter Forrest. The Identity of Indiscernibles. In *The Stanford Encyclopedia of Philosophy*. Metaphysics Research Lab, Stanford University, Winter 2020 edition, 2020. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.289, + 0.905, + 0.327 + ], + "angle": 0, + "content": "[22] Jeremy Freeman and Eero P Simoncelli. Metamers of the ventral stream. Nature neuroscience, 14(9):1195-1201, 2011. 2, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.331, + 0.907, + 0.356 + ], + "angle": 0, + "content": "[23] Peter Gardenfors. Conceptual spaces: The geometry of thought. MIT press, 2004. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.359, + 0.905, + 0.412 + ], + "angle": 0, + "content": "[24] Amir Globerson and Sam Roweis. Nightmare at test time: robust learning by feature deletion. In Proceedings of the 23rd international conference on Machine learning, pages 353-360, 2006. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.414, + 0.905, + 0.455 + ], + "angle": 0, + "content": "[25] Ian J Goodfellow, Jonathon Shlens, and Christian Szegedy. Explaining and harnessing adversarial examples. arXiv preprint arXiv:1412.6572, 2014. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.457, + 0.905, + 0.481 + ], + "angle": 0, + "content": "[26] Nelson Goodman. The structure of appearance. Harvard Univ. Press, 1951. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.484, + 0.905, + 0.51 + ], + "angle": 0, + "content": "[27] Delia Graff. Phenomenal continua and the sorites. Mind, 110 (440):905-936, 2001. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.512, + 0.905, + 0.537 + ], + "angle": 0, + "content": "[28] Hermann Grassmann. Zur theorie der farbenmischung. Annalen der Physik, 165(5):69-84, 1853. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.54, + 0.905, + 0.593 + ], + "angle": 0, + "content": "[29] Jamie Hayes and George Danezis. Learning universal adversarial perturbations with generative models. In 2018 IEEE Security and Privacy Workshops (SPW), pages 43-49. IEEE, 2018. 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.596, + 0.905, + 0.662 + ], + "angle": 0, + "content": "[30] Martin N Hebart, Charles Y Zheng, Francisco Pereira, and Chris I Baker. Revealing the multidimensional mental representations of natural objects underlying human similarity judgements. Nature human behaviour, 4(11):1173-1185, 2020. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.665, + 0.905, + 0.704 + ], + "angle": 0, + "content": "[31] Eleanor R Heider. \"Focal\" color areas and the development of color names. Developmental psychology, 4(3):447, 1971. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.707, + 0.905, + 0.732 + ], + "angle": 0, + "content": "[32] Benj Hellie. Noise and perceptual indiscriminability. Mind, 114(455):481-508, 2005. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.735, + 0.905, + 0.761 + ], + "angle": 0, + "content": "[33] Christopher James Henry. Near sets: theory and applications. University of Manitoba (Canada), 2011. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.763, + 0.905, + 0.803 + ], + "angle": 0, + "content": "[34] Felix Hovsepian. A metalogical analysis of vagueness: an exploratory study into the geometry of logic. PhD thesis, University of Warwick, 1992. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.805, + 0.905, + 0.844 + ], + "angle": 0, + "content": "[35] Ke Huang and Selin Aviyente. Sparse representation for signal classification. Advances in neural information processing systems, 19, 2006. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.847, + 0.905, + 0.901 + ], + "angle": 0, + "content": "[36] Andrew Ilyas, Shibani Santurkar, Dimitris Tsipras, Logan Engstrom, Brandon Tran, and Aleksander Madry. Adversarial examples are not bugs, they are features. Advances in neural information processing systems, 32, 2019. 2" + }, + { + "type": "list", + "bbox": [ + 0.517, + 0.093, + 0.907, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "10252" + } + ], + [ + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.092, + 0.482, + 0.146 + ], + "angle": 0, + "content": "[37] Akshay V Jagadeesh and Justin L Gardner. Texture-like representation of objects in human visual cortex. Proceedings of the National Academy of Sciences, 119(17):e2115302119, 2022. 1, 2, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.148, + 0.482, + 0.188 + ], + "angle": 0, + "content": "[38] Ralph Kopperman. All topologies come from generalized metrics. The American Mathematical Monthly, 95(2):89-97, 1988. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.19, + 0.484, + 0.232 + ], + "angle": 0, + "content": "[39] David H Krantz. Color measurement and color theory: I. representation theorem for grassmann structures. Journal of Mathematical Psychology, 12(3):283-303, 1975. 2, 3, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.233, + 0.482, + 0.26 + ], + "angle": 0, + "content": "[40] Dekang Lin et al. An information-theoretic definition of similarity. In ICML, pages 296-304, 1998. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.261, + 0.482, + 0.3 + ], + "angle": 0, + "content": "[41] Alexander D Logvinenko. Object-colour manifold. International journal of computer vision, 101:143-160, 2013. 2, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.302, + 0.482, + 0.343 + ], + "angle": 0, + "content": "[42] Alexander D. Logvinenko and Eugene Demidenko. On counting metamers. IEEE Transactions on Image Processing, 25(2):770-775, 2016. 2, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.344, + 0.482, + 0.399 + ], + "angle": 0, + "content": "[43] Aleksander Madry, Aleksandar Makelov, Ludwig Schmidt, Dimitris Tsipras, and Adrian Vladu. Towards deep learning models resistant to adversarial attacks. arXiv preprint arXiv:1706.06083, 2017. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.401, + 0.482, + 0.468 + ], + "angle": 0, + "content": "[44] Saeed Mahloujifar, Dimitrios I Diochnos, and Mohammad Mahmoody. The curse of concentration in robust learning: Evasion and poisoning attacks from concentration of measure. In Proceedings of the AAAI Conference on Artificial Intelligence, pages 4536-4543, 2019. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.47, + 0.482, + 0.51 + ], + "angle": 0, + "content": "[45] Douglas L Medin, Robert L Goldstone, and Dedre Gentner. Respects for similarity. Psychological review, 100(2):254, 1993. 2, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.511, + 0.482, + 0.552 + ], + "angle": 0, + "content": "[46] Alexius Meinong. Über die Bedeutung des Weber'schen Gesetzes: Beiträge zur Psychologie des Vergleichens und Messens. L. Voss, 1896. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.553, + 0.482, + 0.608 + ], + "angle": 0, + "content": "[47] Seyed-Mohsen Moosavi-Dezfooli, Alhussein Fawzi, Omar Fawzi, and Pascal Frossard. Universal adversarial perturbations. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 1765-1773, 2017. 2, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.609, + 0.482, + 0.663 + ], + "angle": 0, + "content": "[48] Weili Nie, Brandon Guo, Yujia Huang, Chaowei Xiao, Arash Vahdat, and Anima Anandkumar. Diffusion models for adversarial purification. arXiv preprint arXiv:2205.07460, 2022. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.665, + 0.482, + 0.705 + ], + "angle": 0, + "content": "[49] Bruno A Olshausen and David J Field. Sparse coding with an overcomplete basis set: A strategy employed by V1? Vision research, 37(23):3311-3325, 1997. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.707, + 0.482, + 0.761 + ], + "angle": 0, + "content": "[50] Marwan Omar, Soohyeon Choi, DaeHun Nyang, and David Mohaisen. Robust natural language processing: Recent advances, challenges, and future directions. IEEE Access, 2022. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.762, + 0.482, + 0.789 + ], + "angle": 0, + "content": "[51] Wilhelm Ostwald. Physikalische Farbenlehre. Verlag Unesma, 1919. 2, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.79, + 0.482, + 0.859 + ], + "angle": 0, + "content": "[52] Nicolas Papernot, Patrick McDaniel, Somesh Jha, Matt Fredrikson, Z Berkay Celik, and Ananthram Swami. The limitations of deep learning in adversarial settings. In 2016 IEEE European symposium on security and privacy (EuroS&P), pages 372-387. IEEE, 2016. 2, 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.86, + 0.482, + 0.901 + ], + "angle": 0, + "content": "[53] Zdislaw Pawlak. Classification of objects by means of attributes. Polish Academy of Sciences [PAS]. Institute of Computer Science, 1981. 3" + }, + { + "type": "list", + "bbox": [ + 0.093, + 0.092, + 0.484, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.092, + 0.905, + 0.134 + ], + "angle": 0, + "content": "[54] Charles Pelling. Exactness, inexactness, and the nontransitivity of perceptual indiscriminability. Synthese, 164 (2):289-312, 2008. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.136, + 0.905, + 0.163 + ], + "angle": 0, + "content": "[55] Mika Perälä. Aristotle on perceptual discrimination. *Phronesis*, 63(3):257–292, 2018. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.164, + 0.905, + 0.204 + ], + "angle": 0, + "content": "[56] James F Peters. Near sets. Special theory about nearness of objects. Fundamenta Informaticae, 75(1-4):407-433, 2007. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.206, + 0.905, + 0.247 + ], + "angle": 0, + "content": "[57] James F Peters and Piotr Wasilewski. Tolerance spaces: Origins, theoretical aspects and applications. Information Sciences, 195:211-225, 2012. 1, 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.249, + 0.89, + 0.263 + ], + "angle": 0, + "content": "[58] Henri Poincaré. Dernières pensées. Flammarion, 1930. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.264, + 0.905, + 0.29 + ], + "angle": 0, + "content": "[59] Tim Poston. Fuzzy geometry. PhD thesis, University of Warwick, 1971. 2, 3, 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.292, + 0.905, + 0.32 + ], + "angle": 0, + "content": "[60] Diana Raffman. Is perceptual indiscriminability nontransitive? Philosophical Topics, 28(1):153-175, 2000. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.321, + 0.905, + 0.349 + ], + "angle": 0, + "content": "[61] Diana Raffman. Indiscriminability and phenomenal continua. Philosophical perspectives, 26:309-322, 2012. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.35, + 0.907, + 0.377 + ], + "angle": 0, + "content": "[62] Fred S Roberts. Tolerance geometry. Notre Dame Journal of Formal Logic, 14(1):68-76, 1973. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.379, + 0.905, + 0.405 + ], + "angle": 0, + "content": "[63] Eleanor Rosch. Cognitive reference points. Cognitive Psychology, 7(4):532-547, 1975. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.407, + 0.905, + 0.449 + ], + "angle": 0, + "content": "[64] Eleanor H Rosch. On the internal structure of perceptual and semantic categories. In Cognitive development and acquisition of language, pages 111-144. Elsevier, 1973. 6, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.449, + 0.905, + 0.476 + ], + "angle": 0, + "content": "[65] Bertrand Russell. The principles of mathematics. Routledge, 2020. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.478, + 0.905, + 0.505 + ], + "angle": 0, + "content": "[66] Julius A Schreider. Equality, Resemblance and Order. Mir Publishers, Moscow, 1975. 2, 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.507, + 0.905, + 0.548 + ], + "angle": 0, + "content": "[67] Ali Shafahi, W Ronny Huang, Christoph Studer, Soheil Feizi, and Tom Goldstein. Are adversarial examples inevitable? arXiv preprint arXiv:1809.02104, 2018. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.549, + 0.905, + 0.576 + ], + "angle": 0, + "content": "[68] Herbert A Simon. Models of man; social and rational. Wiley, 1957. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.578, + 0.905, + 0.617 + ], + "angle": 0, + "content": "[69] Alexei B Sossinsky. Tolerance space theory and some applications. Acta Applicandae Mathematica, 5:137-167, 1986. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.62, + 0.905, + 0.661 + ], + "angle": 0, + "content": "[70] Michael W Spratling. Classification using sparse representations: a biologically plausible approach. Biological cybernetics, 108:61-73, 2014. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.663, + 0.905, + 0.745 + ], + "angle": 0, + "content": "[71] Dong Su, Huan Zhang, Hongge Chen, Jinfeng Yi, Pin-Yu Chen, and Yupeng Gao. Is robustness the cost of accuracy? a comprehensive study on the robustness of 18 deep image classification models. In Proceedings of the European Conference on Computer Vision (ECCV), pages 631-648, 2018. 2, 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.747, + 0.905, + 0.801 + ], + "angle": 0, + "content": "[72] Jeremias Sulam, Ramchandran Muthukumar, and Raman Arora. Adversarial robustness of supervised sparse coding. Advances in Neural Information Processing Systems, 33:2110-2121, 2020. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.803, + 0.905, + 0.858 + ], + "angle": 0, + "content": "[73] Christian Szegedy, Wojciech Zaremba, Ilya Sutskever, Joan Bruna, Dumitru Erhan, Ian Goodfellow, and Rob Fergus. Intriguing properties of neural networks. arXiv preprint arXiv:1312.6199, 2013. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.86, + 0.905, + 0.901 + ], + "angle": 0, + "content": "[74] Thomas Tanay and Lewis Griffin. A boundary tilting perspective on the phenomenon of adversarial examples. arXiv preprint arXiv:1608.07690, 2016. 2" + }, + { + "type": "list", + "bbox": [ + 0.517, + 0.092, + 0.907, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "10253" + } + ], + [ + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.092, + 0.482, + 0.147 + ], + "angle": 0, + "content": "[75] Yee Whye Teh, Max Welling, Simon Osindero, and Geoffrey E Hinton. Energy-based models for sparse overcomplete representations. Journal of Machine Learning Research, 4 (Dec):1235-1260, 2003. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.149, + 0.482, + 0.19 + ], + "angle": 0, + "content": "[76] William A Thornton. Matching lights, metamers and human visual response. The Journal of Color and Appearance, 2: 23-29, 1973. 2, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.192, + 0.483, + 0.219 + ], + "angle": 0, + "content": "[77] Edward Chace Tolman. *Purposive behavior in animals and men*. Univ of California Press, 1932. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.221, + 0.482, + 0.275 + ], + "angle": 0, + "content": "[78] Dimitris Tsipras, Shibani Santurkar, Logan Engstrom, Alexander Turner, and Aleksander Madry. Robustness may be at odds with accuracy. arXiv preprint arXiv:1805.12152, 2018. 2, 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.278, + 0.482, + 0.304 + ], + "angle": 0, + "content": "[79] Amos Tversky. Features of similarity. Psychological review, 84(4):327, 1977. 2, 6, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.307, + 0.482, + 0.346 + ], + "angle": 0, + "content": "[80] Gottfried Wilhelm Freiherr von Leibniz. Leibniz: Discourse on Metaphysics. Open Court Publishing Company, 1902. 3, 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.349, + 0.482, + 0.377 + ], + "angle": 0, + "content": "[81] Stella Vosniadou and Andrew Ortony. Similarity and analogical reasoning. Cambridge University Press, 1989. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.379, + 0.482, + 0.418 + ], + "angle": 0, + "content": "[82] Ernst Heinrich Weber. De Pulsu, resorptione, auditu et tactu: Annotationes anatomicae et physiologicae... CF Koehler, 1831. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.421, + 0.482, + 0.449 + ], + "angle": 0, + "content": "[83] Ernst Heinrich Weber. Tastsinn und gemeinschaft. Handworterbuch der Physiologie, 1846. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.451, + 0.482, + 0.477 + ], + "angle": 0, + "content": "[84] Timothy Williamson. Identity and Discrimination. Wiley-Blackwell, 1990. 2, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.479, + 0.482, + 0.506 + ], + "angle": 0, + "content": "[85] Crispin Wright. On the coherence of vague predicates. Synthese, pages 325-365, 1975. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.508, + 0.482, + 0.548 + ], + "angle": 0, + "content": "[86] Günter Wyszecki. Valenzmetrische untersuchung des zusammenhanges zwischen normaler und anomaler trichromasia. 1953. 2, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.551, + 0.482, + 0.592 + ], + "angle": 0, + "content": "[87] Gunther Wyszecki and Walter Stanley Stiles. Color science: concepts and methods, quantitative data and formulae. John Wiley & Sons, 2000. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.594, + 0.482, + 0.634 + ], + "angle": 0, + "content": "[88] Erik C. Zeeman. The topology of the brain and visual perception. In The Topology of 3-Manifolds. Prentice Hall, NJ, 1962. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.636, + 0.482, + 0.678 + ], + "angle": 0, + "content": "[89] E. Christopher Zeeman. Topology of the brain. In Math and Computer Science in Biology and Medicine. H.M. Stationary Office, London, United Kingdom, 1965. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.679, + 0.482, + 0.747 + ], + "angle": 0, + "content": "[90] Hongyang Zhang, Yaodong Yu, Jiantao Jiao, Eric Xing, Laurent El Ghaoui, and Michael Jordan. Theoretically principled trade-off between robustness and accuracy. In International conference on machine learning, pages 7472-7482. PMLR, 2019. 2, 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.749, + 0.482, + 0.79 + ], + "angle": 0, + "content": "[91] Zheng Zhang, Yong Xu, Jian Yang, Xuelong Li, and David Zhang. A survey of sparse representation: algorithms and applications. IEEE access, 3:490-530, 2015. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.792, + 0.482, + 0.834 + ], + "angle": 0, + "content": "[92] Li Zhaoping and Keith A May. Psychophysical tests of the hypothesis of a bottom-up saliency map in primary visual cortex. PLoS Computational Biology, 3(4):e62, 2007. 3" + }, + { + "type": "list", + "bbox": [ + 0.093, + 0.092, + 0.483, + 0.834 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.519, + 0.956 + ], + "angle": 0, + "content": "10254" + } + ] +] \ No newline at end of file diff --git a/2025/Doppelgangers and Adversarial Vulnerability/85f12467-2b9c-44c2-8a35-40f17a213b17_origin.pdf b/2025/Doppelgangers and Adversarial Vulnerability/85f12467-2b9c-44c2-8a35-40f17a213b17_origin.pdf new file mode 100644 index 0000000000000000000000000000000000000000..fb67af7dff534a5475c606b493598b9e0b2426d6 --- /dev/null +++ b/2025/Doppelgangers and Adversarial Vulnerability/85f12467-2b9c-44c2-8a35-40f17a213b17_origin.pdf @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:737e199d14e5d8632e4fa445ba6a4689bebcdb4896385202a08a60b260dd107e +size 2000123 diff --git a/2025/Doppelgangers and Adversarial Vulnerability/full.md b/2025/Doppelgangers and Adversarial Vulnerability/full.md new file mode 100644 index 0000000000000000000000000000000000000000..c61c044e6d568fba4c39b9ff1542d5b40116c8fb --- /dev/null +++ b/2025/Doppelgangers and Adversarial Vulnerability/full.md @@ -0,0 +1,433 @@ +# Doppelgängers and Adversarial Vulnerability + +George Kamberov* +University of Alaska Anchorage +Anchorage, AK 99508 +gkamberov@alaska.edu + +# Abstract + +Machine learning (ML) classifiers can make mistakes that are perceptually and cognitively disturbing to humans. The most notorious examples of such errors are adversarial visual metamers. This paper investigates the phenomenon of adversarial Doppelgängers (AD), which encompasses adversarial visual metamers, and compares the performance and robustness of ML classifiers to human performance. + +We find that ADs are inputs that are close to each other with respect to a perceptual metric defined in this paper, and show that ADs are qualitatively different from the usual adversarial examples. The vast majority of classifiers are vulnerable to ADs and robustness-accuracy trade-offs may not improve them. Some classification problems do not admit any AD-robust classifiers because the underlying classes are ambiguous. We provide criteria to determine whether a classification problem is well defined; describe the structure and attributes of AD-robust classifiers; introduce and explore the notions of conceptual entropy and regions of conceptual ambiguity for classifiers that are vulnerable to AD attacks; and discuss methods to bound the AD fooling rate of an attack. We define the notion of classifiers that exhibit hypersensitive behavior, that is, classifiers whose only mistakes are adversarial Doppelgangers. Improving the AD robustness of hypersensitive classifiers is equivalent to improving accuracy. We identify conditions guaranteeing that all classifiers with sufficiently high accuracy are hypersensitive. + +# 1. Introduction + +Perceptual metamers1 are the most striking adversarial examples studied by the machine learning community. Two perceptual metamers are shown in Figure 1. The phenomenon of metamersim studied in the visual domain, in + +cluding perceptual metamers, is a manifestation of the existence of Doppelgängers: different inputs or stimuli that are perceptually indiscriminable. The research community has engaged in active studies of adversarial vulnerability ever since the publication of [73]. Adversarial Doppelgängers, that is, adversarial examples which are Doppelgängers, are qualitatively different from the vast majority of known adversarial examples which humans readily discriminate from correctly classified input samples (Figure 2). There is no + +![](images/62ed6623050a90fa01673828c79577cb3c0d76de6010d623657b9cecd15d7a81.jpg) +(a) + +![](images/d40457b559e82150da3a0cecb420f2a4a7324d6ee1438310d9d407d5b8f97774.jpg) +(b) +Figure 1. Most people cannot discriminate image (a) from image (b). MobileNetV2 classifies the later image as "persian" and the former picture as "taby". + +![](images/a667691483f835cef81b7244c3cfb8676b8ca325f9fcaa6b04d3b02471edcc62.jpg) +(a) Labrador +Figure 2. Applying a Fast Signed Gradient perturbation to the image (a) classified by MobileNetV2 as Labrador yields the image (b) which is classified by MobileNetV2 as Weimeraner. + +![](images/d6fc1eea66ea1f1c66e3f3c7e073f39f21c9fec269a3cf921e69cfe48859c98a.jpg) +(b) Weimeraner + +evidence that Doppelgängers can be studied and understood completely using the $\ell_p$ norms or more general geodesic distances on manifolds, which have been employed to quantify sample differences and to investigate adversarial examples. Perception and context impose topology on a space of inputs, but it rarely aligns with a manifold topology. We will denote this context-relative topology by $\tau_{\delta}$ . It is defined by the context-relative ability to acquire and deploy knowledge. + +In this paper, we explore the context-relative perceptual topologies on a space of inputs $\mathbf{X}$ and examine the vulnerability and robustness of machine learning classifiers to adversarial Doppelgangers. We show that, while the majority of classifiers are vulnerable to adversarial Doppelgangers, safe (Doppelganger robust) classifiers do exist if the classification problem is well defined. However, these robust classifiers may be very rare. + +In Section 3, we discuss the context-relative notion of (active) indiscriminability and the topology $\tau_{\delta}$ that it induces on a space of inputs. The separability and metric properties of various motivating examples of $\tau_{\delta}$ . Additionally, we show that the distances between Doppelgängers are small if measured by a perceptually-based context-relative metric. We also examine the relation between indiscriminability, IDSicriminability, and feature representations. The existence and structure of Doppelgänger robust classifiers is discussed in Section 4. In Section 5, we investigate the relationship between Doppelgängers and misclassified input samples, define the notion of hypersensitive behavior, and show that improved adversarial Doppelgänger robustness does not have to lead to a reduction in accuracy. + +The structure of perceptually regular, i.e., AD robust classifiers are discussed in Section 6. By definition, a classifier is not regular if and only if some inputs can be attacked by adversarial Doppelgangers. Not surprisingly, it turns out that some inputs are more vulnerable than others. In Section 7, we provide measures of adversarial Doppelgangers vulnerability and upper bounds on the fooling rate of an adversarial Doppelganger attack. + +# 2. Related work + +The pair $(\mathbf{X}, \text{indiscriminability relation})$ is a tolerance space. Tolerance spaces, rough sets, and granular computing have been discussed extensively. See [59, 62, 66, 88, 89]. The color and image metamers studied by many authors including [3, 6, 9, 18, 22, 37, 39, 41, 42, 51, 76, 86] are Doppelgangers. + +The research on adversarial examples to date builds on the hypothesis that the space of input samples is a metric space $(\mathbf{X},\mathrm{dist}_{\mathbf{X}})$ . A misclassified input $x^{*}$ is consid + +ered an adversarial example if it is nearby a correctly classified input sample $x$ , i.e., $\mathrm{dist}_{\mathbf{X}}(x, x^{*})$ is small. Usually $\mathbf{X}$ is assumed to be $\mathbb{R}^n$ , endowed with the $\ell_p$ , norm, $p = 1, 2, \ldots, \infty$ or at least locally homeomorphic to $\mathbb{R}^n$ , i.e., a manifold, equipped with some geodesic distance. Somewhat non surprisingly many authors have shown that every classifier can be attacked with such adversarial examples [5, 17, 25, 52] or at least that this is true in many contexts, [44, 47, 67]. + +Other papers indicate that there are paths toward eliminating adversarial examples completely, i.e., it is possible to achieve provable "adversarial robustness" by fixing/retraining the classifier [1, 24, 36, 43, 72, 74]. A widely accepted tenet is that "there is a clear trade-off between accuracy and [adversarial] robustness, and a better performance in testing accuracy in general reduces [adversarial] robustness", [71]. For empirical evidence for this trade-off and some attempts to explain this phenomenon see [71, 78, 90]. + +# 3. Perceptual Topology + +# 3.1 Indiscriminability and Topology + +The ability to decide whether one stimulus/input is distinct from another is essential for adaptation, survival, and intelligent life. Intelligent agents are uniquely capable to activate knowledge to judge distinction. Williamson calls this context-relative process discrimination ([84]) and defines a context-relative symmetric and reflexive binary relation denoted by $\stackrel{\alpha\delta}{\approx}$ and called indiscriminability: + +Definition 1 ([84]). Two inputs $x$ and $y$ are called indiscriminable to a subject at a time $t$ if and only if at time $t$ the subject is not able to activate (acquire or employ) the relevant kind of knowledge that $x$ and $y$ are distinct. + +Indiscriminability generates a context-relative topology on the set of inputs $\mathbf{X}$ . + +Definition 2. The the phenomenal neighborhood of an input $x \in \mathbf{X}$ , is the set $\mathfrak{d}(x) = \left\{y \in \mathbf{X} : y \stackrel{\alpha \delta}{\approx} x\right\}$ . A point $y \in \mathfrak{d}(x) \setminus \{x\}$ is called a Doppelgänger [of $x$ ]. The perceptual topology $\tau_{\delta}$ is the topology generated by the sub-basis $\mathfrak{D}_{\alpha \delta} = \{\mathfrak{d}(x)\}_{x \in \mathbf{X}}$ . + +Example 0: An input $x \in \mathbf{X}$ is called optimal if it does not have non-trivial/non-identical Doppelgangers, i.e., + +if $\mathfrak{d}(x) = \{x\}$ . In particular, if $\stackrel{\infty \delta}{\approx}$ is the identity relationship $=$ (i.e., all inputs are optimal within the given context), then the perceptual topology $\tau_{\delta}$ is discrete. However, the finiteness of human observations (they are subject to finite time and finite work constraints) and the bounded rationality constraints imposed by the limitations on the availability of information and computational capabilities to humans, [68], indicate that the scenario $\mathfrak{d}(x) = \{x\}$ for every $x \in \mathbf{X}$ may be highly unlikely. + +Often reflexive binary relations are defined and discussed as coverings of the underlying space. Indeed, every reflexive binary relation (RBR) $\approx$ on $\mathbf{X}$ defines a covering $\{\mathfrak{g}(x) = \{y:y\approx x\} \}_{x\in \mathbf{X}}$ , of $\mathbf{X}$ and vice versa one can define reflexive binary relationships through coverings of $\mathbf{X}$ . See Appendix, Section A. In particular, the perceptual topology, $\tau_{\delta}$ , is a tolerable topology, cf., Definition 10 in Appendix, Section A. + +The active psychophysics research on just noticeable difference initiated by Weber and Fechner, [19, 20, 82, 83], provides one of the few classes of examples where we have empirically supported understanding of the perceptual topology. + +Example 1: Let $\mathbf{X}$ be the closed bounded interval $[a,b]\subset$ $(0, + \infty)$ . Suppose that Weber's law holds and let $k > 0$ be the Weber constant. Let $w = 1 + k$ , then + +$$ +\mathfrak {d} (x) = \left\{ \begin{array}{l l} {[ a, x w),} & {a \leq x < a w} \\ {(x / w, x w),} & {a w \leq x \leq b / w} \\ {(x / w, b ],} & {b / w < x \leq b.} \end{array} \right. \tag {1} +$$ + +The covering $\{\mathfrak{d}(x)\}_{x\in [a,b]}$ defines a symmetric RBR but the relation is not transitive. The corresponding perceptual topology is $T_{0}$ but not $T_{1}$ , and the topology is not pseudometric. + +The transitivity or more more often the lack of transitivity of $\stackrel{\alpha\delta}{\approx}$ have been studied extensively and proven or postulated in many human experiences, [2, 6, 10, 16, 26, 27, 32, 54, 61, 84, 85]. + +Definition 3. We will denote by $\sim_{\sigma}$ the transitive closure of the indiciminability relation $\stackrel{\alpha \delta}{\approx}$ on $\mathbf{X}$ . It is defined explicitly as $x \sim_{\sigma} y$ iff there exists a finite chain of Doppelgangers $x = x_0 \stackrel{\alpha \delta}{\approx} x_1 \stackrel{\alpha \delta}{\approx} x_2 \stackrel{\alpha \delta}{\approx} \dots \stackrel{\alpha \delta}{\approx} x_n = y$ . We will call the relation $\sim_{\sigma}$ perceptual metamorphy and will refer to any two inputs $x \sim_{\sigma} y$ as metamorphic. Extending Pawlak's terminology, [53] we call the equivalence classes in $\mathbf{X} / \sim_{\sigma}$ (perceptually) elementary sets. + +Example 2: If the indiscriminability relation is transitive, then each $\mathfrak{d}(x)$ is an elementary set. The perceptual topology may be optimal (recall Example 0) or not. In the former case it is Hausdorff and in fact $(\mathbf{X},\tau_{\delta})$ is a discrete manifold, in the later case the topology $\tau_{\delta}$ is not $T_0$ . See Part A.1 in the Appendix. Human visual perception provides a fundamental example where $\mathfrak{d}(x)\neq \{x\}$ . If two images $x$ and $y$ differ only in unattended regions for example due to low saliency values (cf. [92]), then $x\stackrel {\alpha \delta}{\approx}y$ . Visual metamers have been studied by many authors, including [3, 6, 9, 18, 22, 28, 37, 39, 41, 42, 51, 76, 86]. In these studies, the input space is assumed to be endowed with Grassmann structure (see [39]), and in particular, the indiscriminability relation is transitive. $^6$ + +We are not aware of perceptual topologies that are metric. Still, every Doppelgänger $y \stackrel{\alpha \delta}{\approx} x$ of an input $x$ is a small perturbation of $x$ in the sense that the $y$ is a nearest neighbor of $x$ with respect to an appropriate metric $d_w(\cdot, \cdot)$ on $\mathbf{X}$ . Indeed, let the discrimination graph $\Gamma(\mathbf{X}, E_{\alpha \delta})$ be the undirected simple graph, where $\mathbf{X}$ is the set of vertices and we say that there is an edge $\{x, y\} \in E_{\alpha \delta}$ between the vertices $x, y \in \mathbf{X}$ iff $x \stackrel{\alpha \delta}{\approx} y$ . + +Definition 4. We will call the discrimination graph distance $d_{\infty}(x,y)$ between the vertices $x$ and $y$ and, in particular, $d_{\infty}(x,y) = \infty$ iff $x \not\simeq_{\sigma} y$ the extended perceptual distance. The perceptual distance is the metric $d_w: \mathbf{X} \times \mathbf{X} \to [0,1]$ defined by: + +$$ +d _ {w} (x, y) = \frac {d _ {\infty} (x , y)}{1 + d _ {\infty} (x , y)}, \forall x, y \in \mathbf {X}. \tag {2} +$$ + +The metric $d_w$ does not generate the perceptual topology. It is certainly not the usual $l_p$ or any other manifold metric used in ML. + +# 3.2 Indiscriminabile may not be Indiscernible + +Let $\Phi$ be the space of all features of the inputs/stimuli $x\in \mathbf{X}$ and let $\Phi_x\subset \Phi$ be the set of features attributed to $x$ in a given context. Following [21], we say that $x$ and $y$ are indiscernible, in a given context, if $\Phi_x = \Phi_y$ . Many researchers use the terms indiscriminability and indiscernibility as synonyms. This is only accurate when $\stackrel{\alpha\delta}{\approx}$ is transi + +tive. $^{10}$ In general, the relationship between indiscriminability and indiscernibility is not well understood. However, indiscernibility implies indiscriminability if $\{\Phi_x\}_{x\in \mathbf{X}}$ is a perceptually discriminative feature representation, i.e., + +$$ +\Phi_ {x} \bigcap \Phi_ {y} \neq \varnothing \Longleftrightarrow x \stackrel {\circledast} {\approx} y. \tag {3} +$$ + +The biological plausibility of discriminative feature representations is an open question. Still, they provide insight into the structure of the perceptual topology. The attributed discriminative features represent structures of Doppelgangers. Indeed, let $\{\Phi_x\}_{x\in \mathbf{X}}$ be a feature representation and let $\operatorname {cl}(\xi)$ be the context-dependent semantic cluster of inputs sharing the feature $\xi \in \Phi$ . Specifically, + +$$ +\mathfrak {c l} (\xi) = \left\{x \in \mathbf {X}, \text {s . t . ,} \xi \in \Phi_ {x} \right\}. ^ {1 2} \tag {4} +$$ + +In particular, if $\{\Phi_x\}_{x\in \mathbf{X}}$ is a discriminative feature representation, then every attributed discriminative feature $\xi \in \Phi_{x}$ is associated to, and in some way explained by, a collection of Doppelgangers since $\mathfrak{cl}(\xi)\subset \mathfrak{d}(x)$ . For more detailed discussion and examples of discriminative feature representations see Part B in the Appendix. + +# 4. Classifiers and Adversarial Doppelgängers. + +A classifier $R$ (with $m$ labels) is called fully populated iff the labeling function $\text{label}_R: \mathbf{X} \to \{1, \dots, m\}$ is surjective mapping onto the range of labels $\{1, \dots, m\}$ . For any classifier $R$ , with $m$ labels, we will denote by $R_c$ the level set of the labeling function $\text{label}_R$ for each label $c \in \{1, \dots, m\}$ .13 + +Definition 5. We say that $x$ is a Doppelgänger adversarial to the classifier $R$ iff $\exists y \in \mathfrak{d}(x)$ such that $\operatorname{label}_R(x) \neq \operatorname{label}_R(y)$ and we will refer to both $x$ and $y$ as adversarial Doppelgangers when the classifier $R$ is clear from the context. + +A classifier $R$ is called (perceptually) regular iff it does not admit adversarial Doppelgangers. If $R$ is regular, then $\mathfrak{D}_{\alpha \delta} = \{\mathfrak{d}(x)\}_{x \in \mathbf{X}}$ are $R$ coherent coverings (as defined in [66]). + +We say that the classification problem with $m$ -labels is well defined if there exists a fully populated and perceptually regular classifier with $m$ labels. Otherwise we say that the classification problem with $m$ -labels is not well defined. + +The labeling function $\text{label}_R$ of a regular classifier is continuous with respect to the perceptual topology $\tau_{\delta}$ . Furthermore, if $R$ is not regular and $x$ is a point of discontinuity of $\text{label}_R: (\mathbf{X}, \tau_{\delta}) \to \{1, \dots, m\}$ , then $x$ is an adversarial Doppelgänger. The discontinuity is an indication of the cognitive disruption that occurs when one encounters some AD. + +It is well known that in some experiences perceptually unambiguous categories and hence perceptually regular classifiers do not exist. A simple example is provided by the perceptual topology in Example 1. Indeed, in this case $\mathbf{X} / \sim_{\sigma}$ is a singleton, i.e., every two inputs are metamorphic and hence there is a only one elementary set which equals the whole $\mathbf{X}$ . Therefore, every classifier with two or more labels must have adversarial Doppelgängers.[14] + +Clearly, no amount of "robust training" will get rid of adversarial examples of a classifier with a surjective labeling function $\text{label}_R: (\mathbf{X}, \tau_{\delta}) \to \{1, \ldots, m\}$ if $\mathbf{X}$ cannot be broken into $m$ perceptually unambiguous categories. In the rest of this section we investigate the non-existence, existence and internal structure of regular classifiers. + +In Part D of the Appendix we show a specific example of a well defined classification problem and discuss the actual regular classifier. + +Example 3: If $\stackrel{\alpha \delta}{\approx}$ is transitive then for every number of labels $m$ smaller than the number of equivalence classes card $\left(\mathbf{X} / \stackrel{\alpha \delta}{\approx}\right)$ there exists a fully populated regular classifier with $m$ labels, but if the number of labels $m$ is bigger than card $\left(\mathbf{X} / \stackrel{\alpha \delta}{\approx}\right)$ then every fully populated classifier with $m$ labels must admit adversarial Doppelgangers. + +Example 3 indicates that if the transitive closure $\sim_{\sigma}$ is trivial, i.e., $\sim_{\sigma} = \mathbf{X} \times \mathbf{X}$ , then no label is safe. Namely: + +Observation 1. If the transitive closure $\sim_{\sigma}$ of the indiscriminability relation $\stackrel{\alpha \delta}{\approx}$ is trivial, then every fully populated classifier with two or more classes admits adversarial Doppelgängers. In particular, let $R$ be a fully populated classifier with a surjective labeling function $\text{label}_R: X \to \{1, 2, \ldots, m\}$ , then for every label $c$ there exist adversarial Doppelgängers $x(c) \in \mathbf{X}$ and $x^*(c) \in \mathfrak{d}(x(c))$ such that $c = \text{label}_R(x(c))$ and $\text{label}_R(x^*(c)) \neq \text{label}_R(x(c))$ . + +The proof follows from the fact that every finite chain of Doppelgängers connecting points that are labeled differently by a classifier must contain a pair of adversarial Doppelgängers. See Lemma 2 in Appendix Section C. + +A straight-forward argument shows that if $R$ is a perceptually regular fully populated classifier, then $x \in R_i$ iff + +$[x]_{\sim_{\sigma}} \subset R_i$ , i.e., each level set is a disjoint union of elementary sets $[x]_{\sim_{\sigma}}$ . The pigeonhole principle yields: + +Observation 2. If the number of equivalence classes card $(\mathbf{X} / \sim_{\sigma}) \geq 2$ , then for every natural number $2 \leq m \leq$ card $(\mathbf{X} / \sim_{\sigma})$ there exists a perceptually regular fully populated classifier with $m$ labels. However, if $m >$ card $(\mathbf{X} / \sim_{\sigma})$ , then every fully populated classifier with $m$ labels must have adversarial Doppelgangers. + +In particular, if $p = \operatorname{card}(\mathbf{X} / \sim_{\sigma}) \geq 2$ is finite, then for every natural number $m \leq p$ there are exactly $S(p, m)$ regular fully populated classifiers with $m$ classes. Here $S(p, m)$ is the Sterling number of the second kind. + +Observation 2 shows that the problem of finding a fully populated perceptually unambiguous classifier $R$ with precisely $m$ labels is not well defined if card $(\mathbf{X} / \sim_{\sigma}) < m$ and vice versa that the same problem is well defined for every $m$ such that card $(\mathbf{X} / \sim_{\sigma}) \geq m$ . In the former case solutions do not exist while in the later case solutions exist and each class segment $R_{i}$ is a union of equivalence classes $[x]_{\sim_{\sigma}}$ , + +$$ +R _ {i} = \bigcup_ {x \in R _ {i}} [ x ] _ {\sim_ {\sigma}}. \tag {5} +$$ + +The existence and properties of discriminative feature representations provide insight whether a classification problem is well defined. In particular, if there exists a discriminative feature representation and the set of attributed features is finite, then every classification problem, whose number of labels exceeds the number of attributed features, is not well defined. See Observation 9 in Appendix Part B.2. Further discussion of the structure of class segments of a regular classifier including the class/category core and fringe are discussed in Section 6. + +# 5. Accuracy and Adversarial Doppelgängers. + +The accuracy-adversarial robustness trade off observed and discussed in the literature involves various measures of accuracy [48, 50, 71, 78, 90]. We will discuss classification accuracy. We will show that there is a strong relationship between classifier accuracy and vulnerability to adversarial Doppelgängers. In particular, we will identify (perceptual) scenarios in which low accuracy classifiers are critically vulnerable to adversarial Doppelgänger attacks but on the other hand all high accuracy classifiers can be fooled only by Doppelgängers. + +# 5.1. The Probabilistic Setup. + +We will assume that $(\mathbf{X},\mathcal{F},\mu)$ is a probability measure space equipped with perceptual topology $\tau_{\delta}$ generated by an indsicriminability relation $\stackrel{\alpha \delta}{\approx}$ such that for every $x \in \mathbf{X}$ the set of Doppelgangers $\mathfrak{d}(x)$ and the equivalence class $[x]_{\sim_{\sigma}}$ are events, $(\mathfrak{d}(x) \in \mathcal{F}$ and $[x]_{\sim_{\sigma}} \in \mathcal{F}, \forall x \in \mathbf{X})$ . + +In the rest of this section we will assume that the classification problem with $m \geq 2$ labels is well defined and let $\Omega$ be a perceptually regular classifier; we will reserve the notation $R$ to denote any classifier ( $R$ may or may not be perceptually regular) such that $R_{i} \cap \Omega_{i}$ , $i = 1, \ldots, m$ are the true positives (of class $i$ ). To define accuracy we will focus only on regular models and classifiers s.t., $\Omega_{i} \in \mathcal{F}$ and $R_{i} \in \mathcal{F}$ , for all $i = 1, \ldots, m$ . The accuracy of the classifier $R$ defined as + +$$ +\operatorname {a c c u r a c y} _ {\Omega} (R) = \mu \left(R _ {1} \cap \Omega_ {1}\right) + \dots + \mu \left(R _ {m} \cap \Omega_ {m}\right). \tag {6} +$$ + +Furthermore let us assume that $\mu (\Omega_i) > 0$ for every $i = 1,\ldots ,m$ and thus we can define recall rates + +$$ +\rho_ {i} = \frac {\mu \left(R _ {i} \cap \Omega_ {i}\right)}{\mu \left(\Omega_ {i}\right)}, \quad i = 1, \dots , m \tag {7} +$$ + +Bounds on the recall rates imply bounds on the accuracy. Namely if + +$$ +\underline {{\rho}} \leq \rho_ {i} \leq \bar {\rho}, \quad i = 1, \dots , m +$$ + +then since $\mu$ is a probability measure on $X$ , and + +$$ +\mu \left(R _ {i} \cap \Omega_ {i}\right) = \rho_ {i} \mu \left(\Omega_ {i}\right), \quad i = 1, \dots , m \tag {8} +$$ + +we get + +$$ +\underline {{\rho}} \leq \operatorname {a c c u r a c y} _ {\Omega} (R) = \sum_ {i = 1} ^ {m} \rho_ {i} \mu (\Omega_ {i}) \leq \bar {\rho}. +$$ + +# 5.2. Are Trade-Offs Possible? + +Let $i(x) \in \{1, \ldots, m\}$ be the object class label of $x \in \mathbf{X}$ i.e., $x \in \Omega_{i(x)}$ and + +$$ +\bar {k} (\Omega) = \sup _ {x \in \mathbf {X}} \left(\frac {\mu \left(\Omega_ {i (x)}\right)}{\mu (\mathfrak {d} (x))}\right). \tag {9} +$$ + +Every classifier whose recall rates do not exceed $1 / \bar{k} (\Omega)$ is totally unsafe in the sense that every correctly classified input admits adversarial Doppelgangers. Specifically: + +Observation 3. Suppose that the sets of Doppelgangers are not negligible and $\inf_{x\in \mathbf{X}}\mu (\mathfrak{d}(x)) > 0$ , and let $\Omega = \{\Omega_1,\dots ,\Omega_m\}$ be a regular world model and let $R = \{R_{1},\ldots ,R_{m}\}$ be a classifier whose recall rates are strictly smaller than $1 / \bar{k} (\Omega)$ and so + +$$ +\frac {\mu \left(R _ {i (x)} \cap \mathfrak {d} (x)\right)}{\mu (\mathfrak {d} (x))} \leq \bar {\rho} \bar {k} (\Omega) < 1. \tag {10} +$$ + +Thus every correctly classified input $x$ has adversarial Doppelgängers. + +Observation 3 shows that sacrificing accuracy may lead to increasing the probability of encountering adversarial Doppelgangers and in fact that there is no trade off for accuracies that are sufficiently low provided that all sets of Doppelgangers have positive measure. + +The lack of an opportunity for a trade-off is even more striking when one tries to improve high recall rate (and hence high accuracy) classifiers. + +Observation 4. $^{15}$ If $\inf_{x\in \mathbf{X}}\mu (\mathfrak{d}(x)) > 0$ , and let $R =$ $\{R_1,\ldots ,R_m\}$ be a classifier whose recall rates are sufficiently high so that $\underline{\rho} >1 - 1 / \bar{k} (\Omega)$ . i.e., + +$$ +\left(1 - \underline {{\rho}}\right) \bar {k} (\Omega) < 1. \tag {11} +$$ + +Then every misclassified input $x$ is an adversarial Doppelgänger. + +Definition 6. We say that a classifier has a hypersensitive behavior if every misclassified input is an adversarial Doppelgänger. + +For classifiers with hypersensitive behavior adversarial robustness can only be improved by improving accuracies, i.e., by eliminating misclassification. Observation 4 shows that if $\mu (\mathfrak{d}(x)) > 0$ , for all inputs $x\in \mathbf{X}$ , then all classifiers with sufficiently high accuracy are either regular (when accuracy equals to one) or have hypersensitive behavior. + +In summary adversarial Doppelgänger robustness - accuracy trade-off may happen for classifiers with middling accuracy rates or when there are inputs whose Doppelgängers are negligible in measure. + +# 6. Life without borders. + +A important property of the perceptual topology is that the if a classifier $\Omega$ is perceptually regular, then it "imposes an open [topological] borders policy", that is, $\partial \Omega_{c} = \emptyset$ for every label $c$ . Linguists and psychologists, have observed and postulated that natural perceptual and semantic categories are borderless. See for example [64]. Class/decision boundaries are studied and exploited in many works on classifiers, and in particular on adversarial robustness. These boundaries are artifacts of the metric topology used by the researchers, they are not perceptual phenomena. + +However, we are all familiar with the idea that some stimuli are more intrinsic/representative to/of a given class and at the same time frequently there are objects/stimuli/Input that, while they are firmly with in the class, are less representative/share few(er) features compared to the rest of the elements in the class, that is, they "are/belong to the fringe" of the class. + +Definition 7. Let $s$ be a similarity scale, i.e., a function $s: \mathbf{X} \times \mathbf{X} \to \mathbb{R}$ such that $s(x, x) \geq s(x, y), \forall x, y \in \mathbf{X}$ as in [79], [45] (measuring similarity within a fixed context) and [40]. + +The values $s(x, x)$ can be and sometimes are used to represent the salience or equivalently importance of the input within $\mathbf{X}$ , see for example [79]. The similarity scale provides a method to quantify the affinity of a input/stimulus to a given measurable subset $D \subset \mathbf{X}$ and the notions of prototype and fringe. The (s-)affinity of $x$ with a measurable set $D$ is defined as + +$$ +P (x, D) = \int_ {D} s (x, y) \tag {12} +$$ + +This is a straightforward generalization of the notion of prototypicality defined in [79]. + +Definition 8. $x \in D$ is called a prototype of $D$ (with respect to the integrable similarity scale $s: \mathbf{X} \times \mathbf{X} \to \mathbb{R}$ ) if + +$$ +P (x, D) = \sup _ {z \in D} P (z, D) \tag {13} +$$ + +$x \in D$ is called a fringe element of $D$ (with respect to the integrable similarity scale $s: \mathbf{X} \times \mathbf{X} \to \mathbb{R}$ ) if + +$$ +P (x, D) = \inf _ {z \in D} P (z, D) \tag {14} +$$ + +One is tempted to think of prototypes as the stimuli that are "clearest cases, best examples". "easy to tell apart" and to "be a good representative" and hence that optimize salience [13, 31, 63, 64, 79]. However, clearly there is no reason to expect that a stimulus that is unlikely to be observed/encountered would be selected as a prototype. The examples below show that prototypes and fringes are obtained by optimizing a mixture of the frequency of appearance (likelihood to encounter) and salience. + +The core and fringe sets may be empty. In practice one is often satisfied with finding elements that may not be true prototypes but belong to the $M$ -core, i.e., have affinity exceeding a fixed threshold $M$ . Similarly, elements whose affinity is below a given threshold $\tau$ belong to the $\tau$ -fringe. + +Many perceptual and cognitive processes exploit the intelligent agents' abilities to measure and/or compare the salience/importance of features. In some simple cases it is expected and it even might be true that salience/importance is a probability measure $f_{\Phi}$ on the space of (all possible) features $\Phi$ and the feature representations $\Phi_x$ are measurable subsets of $\Phi$ . In many accounts including Tversky's feature contrast model [79] the salience scale $f_{\Phi}$ is a context dependent, nonnegative function defined on a collection $\Upsilon(\Phi)$ of subsets of $\Phi$ which is closed under finite unions and intersections, and set differences. Furthermore, $\Phi_x \in \Upsilon(\Phi), \forall x \in \mathbf{X}$ ; and the non-negative function $f_{\Phi}$ is feature additive, i.e., $f_{\Phi}(A \cup B) = f_{\Phi}(A) +$ + +$f_{\Phi}(B)$ , if $A \cap B = \emptyset$ . The value $f(x) = f_{\phi}(\Phi_x), \forall x \in \mathbf{X}$ is called the salience/prominence of the input $x$ , [79]. We will call a (Tversky) salience scale $f_{\Phi}$ perceptually regular if the prominence/salience function $f(x) = f_{\phi}(\Phi_x)$ is perceptually regular. We will call a perceptually regular salience scale $f_{\Phi}$ fully deployable if it can be used to judge the distinguishing features of inputs/stimuli. In particular, we can use $f_{\Phi}(\Phi_x \setminus \Phi_y)$ to discriminate $x$ from $y$ . Thus if $f_{\Phi}$ is fully deployable, then $f_{\Phi}(\Phi_x \setminus \Phi_y) = 0$ and $f_{\Phi}(\Phi_y \setminus \Phi_x) = 0$ whenever $x \stackrel{\alpha \delta}{\approx} y$ or equivalently $f_{\Phi}(\Phi_x \cap \Phi_y) = f(x)$ if $x \stackrel{\alpha \delta}{\approx} y$ . + +The following special case provides a particularly useful insight into the nature of prototypes and the possible internal structure of categories as discussed in [64]. + +Example 4: If $s$ is a contrast similarity [79] such that $f_{\Phi}(\Phi_x) = s(x,x) / \theta$ is fully deployable, $\theta \in (0, + \infty)$ , and furthermore, $f_{\Phi}(\Phi_x \cap \Phi_y) = f_{\Phi}(\Phi_x)$ for every pair $x \sim_{\sigma} y$ and $\Phi_x \cap \Phi_y = \emptyset$ if $x \not\sim_{\sigma} y$ , then for every regular class $D \subset \mathbf{X}$ , and every $x \in D$ we get + +$$ +P (x, D) = \Theta \mu (D) \left(\frac {\mu ([ x ] _ {\sim_ {\sigma}})}{\mu (D)} - \frac {\alpha}{\Theta}\right) f _ {\Phi} \left(\Phi_ {x}\right) - \beta I _ {\Phi} (D), \tag {15} +$$ + +where $\Theta = (\alpha +\beta +\theta)$ and $\alpha ,\beta$ ,and $\theta$ are non-negative constants. In particular, if $D$ is a finite union of equivalence classes $\zeta_j\in \mathbf{X} / \sim_\sigma$ , then both prototypes and fringe elements exist. Furthermore, if $x$ is prototype/fringe then so are all stimuli in its component $\zeta = [x]_{\sim_{\sigma}}$ . Similar statements hold for M-core and $\tau$ -fringe elements. + +If the class $D$ is fixed, then $\frac{\mu([x]_{\sim_{\sigma}})}{\mu(D)}$ is just the probability to encounter (and learn) a set of stimuli, and $f_{\Phi}(\Phi_x)$ can be interpreted as the level of prominence. So the optimization process involves learning prominent examples that can be encountered reasonably often. Thus in this case the prototype does not fall in either of the two main branches of prototypes, that is inputs that represent the central tendency in the regular class vs. prototypes as highly "representative exemplar(s) of a category", see [15], page 52. + +More generally, when the similarity scale is bounded, for example, this is true for real similarity measures deployable by humans, then, as predicted by modern prototype theory, each perceptually regular subset $D \subset \mathbf{X}$ corresponding to real (natural) categories created and analyzed by real intelligent agents consists of core elements (possibly M-core), a layer of fringe (possibly $\tau$ -fringe) elements, and layers of elements of various levels of intermediate affinity with $D$ . In Part I of the Appendix we introduce class invariants of regular classifiers including structural entropy, expected index of coincidence, and importance. + +# 7. Quantifying Doppelgänger Vulnerability. + +By definition if a classifier $R$ is not regular, then it is vulnerable to adversarial Doppelgangers attacks, that is, for some input $x$ there exists $a(x) \stackrel{\alpha \delta}{\approx} x$ and such that $\operatorname{label}_R(x) \neq \operatorname{label}_R(a(x))$ . In particular, we say that $R$ is conceptually ambiguous at $x$ and we call the set + +$$ +\mathrm {A} (R) = \left\{x \in \mathbf {X}: \exists y \stackrel {\alpha \delta} {\approx} x \text {a n d} \operatorname {l a b e l} _ {R} (x) \neq \operatorname {l a b e l} _ {R} (y) \right\} \tag {16} +$$ + +the region of conceptual ambiguity. When $(X, \mu)$ is a probability measure space and $\mu(\mathfrak{d}(x)) > 0$ , we use the probability distribution of labels at $x$ : + +$$ +p _ {j} (x) = \frac {\mu \left(R _ {j} \bigcap \mathfrak {d} (x)\right)}{\mu (\mathfrak {d} (x))}, \quad j = 1, \dots , m \tag {17} +$$ + +and the conceptual entropy of $R$ at $x$ defined as + +$$ +H _ {R} (x) = - \sum_ {j = 1} ^ {m} p _ {j} (x) \log \left(p _ {j} (x)\right) \tag {18} +$$ + +to detect whether $R$ is conceptually ambiguous at $x$ (i.e., $H_{R}(x) > 0$ ), and to quantify the likelihoods of various adversarial Doppelgänger attacks. + +Definition 9. Let $R$ be a classifier and $\hat{a} : \mathbf{X} \to \mathbf{X}$ , we will call the inner measure of the set $\{x : \operatorname{label}_R(\hat{a}(x)) \neq \operatorname{label}_R(x)\}$ the $R$ -fooling rate of the mapping $\hat{a}$ , and we will denote it by $F_R(\hat{a})$ . A mapping $\hat{a} : \mathbf{X} \to \mathbf{X}$ is called an adversarial Doppelgänger attack to a classifier $R$ if and only if $\hat{a}(x) \stackrel{\alpha \delta}{\approx} x, \forall x \in \mathbf{X}$ , and the $R$ -fooling rate $F_R(\hat{a})$ is positive. + +The set $\{x:\mathrm{label}_R(\hat{a} (x))\neq \mathrm{label}_R(x)\}$ is a subset of the region of conceptual ambiguity of $R$ , which yields an upper bound on the $R$ -fooling rate by the outer measure of $A(R)$ .. + +$$ +F _ {R} (\hat {a}) \leq \mu_ {*} (A (R)). \tag {19} +$$ + +In specific scenarios it is possible to get an upper bound on the size, possibly the outer measure, of $A(R)$ which in turn shows that the $R$ -fooling rates are bounded away from one. See Example 11 in Part J in the Appendix. + +Adversarial Doppelgänger attacks are distinct from the adversarial attacks studied to date. The universal adversarial attacks, [47], can achieve fooling rates as close to one as one desires. As illustrated above, adversarial Doppelgänger attacks may not be able to reach fooling rates that are too high. On the other hand in some cases, the optimal fooling rate of one can be achieved. + +Observation 5. If $R$ is conceptually ambiguous at every $x \in \mathbf{X}$ , e.g., when $H_{R}(x) > 0$ for every $x \in \mathbf{X}$ , then there exists an adversarial Doppelgänger attack with $R$ -fooling rate equal to one. + +Indeed, if $R$ is conceptually ambiguous at every $x \in \mathbf{X}$ , then $\{y \in \mathfrak{d}(x) : \operatorname{label}_R(y) \neq \operatorname{label}_R(x)\} \neq \emptyset$ , for every $x$ , and therefore the axiom of choice implies that there exists a map $\hat{a} : \mathbf{X} \to \mathbf{X}$ such that $\hat{a}(x) \in \{y \in \mathfrak{d}(x) : \operatorname{label}_R(y) \neq \operatorname{label}_R(x)\}$ , for every $x \in X$ . It turns out that in practice there may be many classifiers that are conceptually ambiguous at every input. + +Example 5: Consider the case when $\stackrel{\alpha\delta}{\approx}$ is transitive, i.e., $\stackrel{\alpha\delta}{\approx}$ equals its transitive closure $\sim_{\sigma}$ , e.g., when $\mathbf{X}$ is equipped with a Grassmann structure as in [39], and there exist at least two different equivalence classes. Thus binary classification is a well defined problem. There exist at least $\prod_{\zeta \in \mathbf{X} / \sim_{\sigma}} \{U \subset \zeta : 0 < \mu(U) < \mu(\zeta)\}$ worth of fully-populated binary classifiers which are conceptually ambiguous at every input $x \in \mathbf{X}$ . + +The last example and Observation 5 show that even high accuracy classifiers can be vulnerable to adversarial Doppelgänger attacks with fooling rate equal to one. + +We conclude this section with a warning that popular methods to deal with unseen data, including marking missing data and imputation, may introduce conceptual ambiguity. For example, if a model is trained on a data set $T \subset \mathbf{X}$ that includes only parts of some elementary sets, then adding a class label NA to label unseen data can compromise adversarial Doppelgänger robustness. Indeed, let $S$ be a nonempty set of training data such that $S \subsetneq \zeta \in \mathbf{X} / \sim_{\sigma}$ . There exists, $x \in S$ such that $\mathfrak{d}(x) \setminus S \neq \emptyset$ . Every $z \in \mathfrak{d}(x) \setminus S$ , labeled as NA, is an adversarial Doppelgänger of $x$ . + +# 8. Discussion and Conclusions. + +A central focus of this paper is the adversarial Doppelgangers phenomenon, where classifiers assign different labels to inputs that humans cannot discriminate. Until now, this phenomenon has not been well understood, possibly due to the limitations of the distance-based analysis that has dominated the field. In the "absence of a distance measure that accurately captures the perceptual differences between a source and adversarial example many researchers have decided to use the $\ell_p$ distance", [29]. The available empirical observations and models - both perceptual and cognitive, including those based on just noticeable differences - provide no evidence that biologically plausible perceptual topologies are metric. This paper advances the understanding of context-related perceptual topologies in input spaces, which are rarely metric. Our investigation shows that adversarial Doppelgangers are very close to each other with respect to the context-relevant perceptual metric $d_w$ , this metric is not a manifold metric and does not generate the perceptual topology. This distinction highlights the shortcomings of traditional, purely manifold metric-based representations and analysis of perceptual spaces. + +The machine learning community has expended significant efforts aimed to build adversarially robust classifiers. This may be a march towards a bridge too far. Philosophers, experimental psychologists, and linguists, are well aware that many classification problems are not well defined due to perceptual ambiguities. Any fully populated classifier for a classification problem that is not well defined is doomed to be a victim of the adversarial Doppelgängers phenomenon. Our results reveal the structure of adversarial Doppelgänger-robust classifiers, regular classifiers, and criteria and methods to establish whether a classification problem is well defined or not. The new understanding of the structure of regular classifiers, the analysis of zones of ambiguity, and the methods to measure and bound the fooling rates of adversarial Doppelgänger attacks provide guidance on how to design adversarially robust training to improve classifiers that are not regular. In addition to revealing the impossibility to use accuracy-robustness trade-offs in many scenarios, including robustifying hypersensitive classifiers, our analysis indicates that marking unseen data can jeopardize robustness if the training data contains only a proper subset of an elementary set. + +We explore feature representations, the related concept of indiscernibility introduced by Leibniz, and their connection to indiscriminability. This investigation reveals the nature of class prototypes and fringe inputs, and how the size of a discriminative feature representation can be used to determine whether a classification problem is not well defined. Indiscernibility and indiscriminability, are often conflated in the machine learning literature. Elucidating the distinction between them is vital for understanding the limitations of current classifiers and addressing the shortcomings in their design. + +Our discussion of the Doppelgängers phenomenon brings to light a significant divergence between human perception and artificial neural network models, including feedforward models, RNN models and ResNet. The indiscriminability relations of these artificial neural network models, studied in [18], are transitive16 while it is well accepted, that, in many contexts, the human indiscriminability relation is not transitive. + +The results and insights gained from this investigation point to concrete warnings and actionable steps for improving the training and testing of classifiers. + +# References + +[1] Zeyuan Allen-Zhu and Yuanzhi Li. Feature purification: How adversarial training performs robust deep learning. In 2021 IEEE 62nd Annual Symposium on Foundations of Computer Science (FOCS), pages 977-988. IEEE, 2022. 2 +[2] Valtteri Arstila. Why the transitivity of perceptual simultaneity should be taken seriously. Frontiers in integrative neuroscience, 6:3, 2012. 3 +[3] Benjamin Balas, Lisa Nakano, and Ruth Rosenholtz. A summary-statistic representation in peripheral vision explains visual crowding. Journal of vision, 9(12):13-13, 2009. 2, 3 +[4] Max Black. The identity of indiscernibles. Mind, 61(242): 153-164, 1952. 1 +[5] Nicholas Boucher, Ilia Shumailov, Ross Anderson, and Nicolas Papernot. Bad characters: Imperceptible nlp attacks. In 2022 IEEE Symposium on Security and Privacy (SP), pages 1987-2004. IEEE, 2022. 2 +[6] William F Broderick, Gizem Rufo, Jonathan Winawer, and Eero P Simoncelli. Foveated metamers of the early visual system. bioRxiv, pages 2023-05, 2023. 1, 2, 3, 6 +[7] Fan RK Chung. Lectures on spectral graph theory. CBMS Lectures, Fresno, 6(92):17-21, 1996. 6 +[8] Samuel Clarke. A Collection of Papers, which Passed Between the Late Learned Mr. Leibnitz, and Dr. Clarke, in the Years 1715 and 1716: Relating to the Principles of Natural Philosophy and Religion. With an Appendix. To which are Added, Letters to Dr. Clarke Concerning Liberty and Necessity; from a Gentleman of the University of Cambridge: with the Doctor's Answers to Them. Also Remarks Upon a Book, Entitled, A Philosophical Enquiry Concerning Human Liberty. James Knapton, at the Crown in St. Paul's Church Yard., 1717. 3 +[9] Jozef B Cohen and William E Kappauf. Color mixture and fundamental metamers: Theory, algebra, geometry, application. The American journal of psychology, pages 171-259, 1985. 2, 3 +[10] Rafael De Clercq and Leon Horsten. Perceptual indiscriminability: In defence of Wright's proof. The Philosophical Quarterly, 54(216):439-444, 2004. 3 +[11] Lieven Decock and Igor Douven. Similarity after Goodman. Review of philosophy and psychology, 2:61-75, 2011. 2 +[12] Kaize Ding, Albert Jiongqian Liang, Bryan Perrozi, Ting Chen, Ruoxi Wang, Lichan Hong, Ed H. Chi, Huan Liu, and Derek Zhiyuan Cheng. Hyperformer: Learning expressive sparse feature representations via hypergraph transformer, 2023. 2 +[13] Igor Douven. Putting prototypes in place. Cognition, 193: 104007, 2019. 6 +[14] Michael Elad. Sparse and redundant representations: from theory to applications in signal and image processing. Springer Science & Business Media, 2010. 2 +[15] William K Estes. Classification and cognition. Oxford University Press, 1994. 7 +[16] Katalin Farkas. The subject's point of view. OUP Oxford, 2010. 2, 3 + +[17] Alhussein Fawzi, Hamza Fawzi, and Omar Fawzi. Adversarial vulnerability for any classifier. Advances in neural information processing systems, 31, 2018. 2 +[18] Jenelle Feather, Alex Durango, Ray Gonzalez, and Josh McDermott. Metamers of neural networks reveal divergence from human perceptual systems. Advances in Neural Information Processing Systems, 32, 2019. 2, 3, 8, 7 +[19] Gustav Theodor Fechner. Elemente der psychophysik. Breitkopf u. Härtel, 1860. 3 +[20] Gustav Theodor Fechner. Revision der hauptpunkte der Psychophysik. Breitkopf und Härtel, 1882. 3 +[21] Peter Forrest. The Identity of Indiscernibles. In *The Stanford Encyclopedia of Philosophy*. Metaphysics Research Lab, Stanford University, Winter 2020 edition, 2020. 3 +[22] Jeremy Freeman and Eero P Simoncelli. Metamers of the ventral stream. Nature neuroscience, 14(9):1195-1201, 2011. 2, 3 +[23] Peter Gardenfors. Conceptual spaces: The geometry of thought. MIT press, 2004. 2 +[24] Amir Globerson and Sam Roweis. Nightmare at test time: robust learning by feature deletion. In Proceedings of the 23rd international conference on Machine learning, pages 353-360, 2006. 2 +[25] Ian J Goodfellow, Jonathon Shlens, and Christian Szegedy. Explaining and harnessing adversarial examples. arXiv preprint arXiv:1412.6572, 2014. 2 +[26] Nelson Goodman. The structure of appearance. Harvard Univ. Press, 1951. 3 +[27] Delia Graff. Phenomenal continua and the sorites. Mind, 110 (440):905-936, 2001. 3 +[28] Hermann Grassmann. Zur theorie der farbenmischung. Annalen der Physik, 165(5):69-84, 1853. 3 +[29] Jamie Hayes and George Danezis. Learning universal adversarial perturbations with generative models. In 2018 IEEE Security and Privacy Workshops (SPW), pages 43-49. IEEE, 2018. 8 +[30] Martin N Hebart, Charles Y Zheng, Francisco Pereira, and Chris I Baker. Revealing the multidimensional mental representations of natural objects underlying human similarity judgements. Nature human behaviour, 4(11):1173-1185, 2020. 2 +[31] Eleanor R Heider. "Focal" color areas and the development of color names. Developmental psychology, 4(3):447, 1971. 6 +[32] Benj Hellie. Noise and perceptual indiscriminability. Mind, 114(455):481-508, 2005. 3 +[33] Christopher James Henry. Near sets: theory and applications. University of Manitoba (Canada), 2011. 2 +[34] Felix Hovsepian. A metalogical analysis of vagueness: an exploratory study into the geometry of logic. PhD thesis, University of Warwick, 1992. 1 +[35] Ke Huang and Selin Aviyente. Sparse representation for signal classification. Advances in neural information processing systems, 19, 2006. 2 +[36] Andrew Ilyas, Shibani Santurkar, Dimitris Tsipras, Logan Engstrom, Brandon Tran, and Aleksander Madry. Adversarial examples are not bugs, they are features. Advances in neural information processing systems, 32, 2019. 2 + +[37] Akshay V Jagadeesh and Justin L Gardner. Texture-like representation of objects in human visual cortex. Proceedings of the National Academy of Sciences, 119(17):e2115302119, 2022. 1, 2, 3 +[38] Ralph Kopperman. All topologies come from generalized metrics. The American Mathematical Monthly, 95(2):89-97, 1988. 1 +[39] David H Krantz. Color measurement and color theory: I. representation theorem for grassmann structures. Journal of Mathematical Psychology, 12(3):283-303, 1975. 2, 3, 8 +[40] Dekang Lin et al. An information-theoretic definition of similarity. In ICML, pages 296-304, 1998. 6 +[41] Alexander D Logvinenko. Object-colour manifold. International journal of computer vision, 101:143-160, 2013. 2, 3 +[42] Alexander D. Logvinenko and Eugene Demidenko. On counting metamers. IEEE Transactions on Image Processing, 25(2):770-775, 2016. 2, 3 +[43] Aleksander Madry, Aleksandar Makelov, Ludwig Schmidt, Dimitris Tsipras, and Adrian Vladu. Towards deep learning models resistant to adversarial attacks. arXiv preprint arXiv:1706.06083, 2017. 2 +[44] Saeed Mahloujifar, Dimitrios I Diochnos, and Mohammad Mahmoody. The curse of concentration in robust learning: Evasion and poisoning attacks from concentration of measure. In Proceedings of the AAAI Conference on Artificial Intelligence, pages 4536-4543, 2019. 2 +[45] Douglas L Medin, Robert L Goldstone, and Dedre Gentner. Respects for similarity. Psychological review, 100(2):254, 1993. 2, 6 +[46] Alexius Meinong. Über die Bedeutung des Weber'schen Gesetzes: Beiträge zur Psychologie des Vergleichens und Messens. L. Voss, 1896. 3 +[47] Seyed-Mohsen Moosavi-Dezfooli, Alhussein Fawzi, Omar Fawzi, and Pascal Frossard. Universal adversarial perturbations. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 1765-1773, 2017. 2, 7 +[48] Weili Nie, Brandon Guo, Yujia Huang, Chaowei Xiao, Arash Vahdat, and Anima Anandkumar. Diffusion models for adversarial purification. arXiv preprint arXiv:2205.07460, 2022. 5 +[49] Bruno A Olshausen and David J Field. Sparse coding with an overcomplete basis set: A strategy employed by V1? Vision research, 37(23):3311-3325, 1997. 2 +[50] Marwan Omar, Soohyeon Choi, DaeHun Nyang, and David Mohaisen. Robust natural language processing: Recent advances, challenges, and future directions. IEEE Access, 2022. 5 +[51] Wilhelm Ostwald. Physikalische Farbenlehre. Verlag Unesma, 1919. 2, 3 +[52] Nicolas Papernot, Patrick McDaniel, Somesh Jha, Matt Fredrikson, Z Berkay Celik, and Ananthram Swami. The limitations of deep learning in adversarial settings. In 2016 IEEE European symposium on security and privacy (EuroS&P), pages 372-387. IEEE, 2016. 2, 5 +[53] Zdislaw Pawlak. Classification of objects by means of attributes. Polish Academy of Sciences [PAS]. Institute of Computer Science, 1981. 3 + +[54] Charles Pelling. Exactness, inexactness, and the nontransitivity of perceptual indiscriminability. Synthese, 164 (2):289-312, 2008. 3 +[55] Mika Perälä. Aristotle on perceptual discrimination. *Phronesis*, 63(3):257–292, 2018. 1 +[56] James F Peters. Near sets. Special theory about nearness of objects. Fundamenta Informaticae, 75(1-4):407-433, 2007. 2 +[57] James F Peters and Piotr Wasilewski. Tolerance spaces: Origins, theoretical aspects and applications. Information Sciences, 195:211-225, 2012. 1, 2 +[58] Henri Poincaré. Dernières pensées. Flammarion, 1930. 2 +[59] Tim Poston. Fuzzy geometry. PhD thesis, University of Warwick, 1971. 2, 3, 1 +[60] Diana Raffman. Is perceptual indiscriminability nontransitive? Philosophical Topics, 28(1):153-175, 2000. 2 +[61] Diana Raffman. Indiscriminability and phenomenal continua. Philosophical perspectives, 26:309-322, 2012. 3 +[62] Fred S Roberts. Tolerance geometry. Notre Dame Journal of Formal Logic, 14(1):68-76, 1973. 2 +[63] Eleanor Rosch. Cognitive reference points. Cognitive Psychology, 7(4):532-547, 1975. 6 +[64] Eleanor H Rosch. On the internal structure of perceptual and semantic categories. In Cognitive development and acquisition of language, pages 111-144. Elsevier, 1973. 6, 7 +[65] Bertrand Russell. The principles of mathematics. Routledge, 2020. 3 +[66] Julius A Schreider. Equality, Resemblance and Order. Mir Publishers, Moscow, 1975. 2, 4 +[67] Ali Shafahi, W Ronny Huang, Christoph Studer, Soheil Feizi, and Tom Goldstein. Are adversarial examples inevitable? arXiv preprint arXiv:1809.02104, 2018. 2 +[68] Herbert A Simon. Models of man; social and rational. Wiley, 1957. 3 +[69] Alexei B Sossinsky. Tolerance space theory and some applications. Acta Applicandae Mathematica, 5:137-167, 1986. 6 +[70] Michael W Spratling. Classification using sparse representations: a biologically plausible approach. Biological cybernetics, 108:61-73, 2014. 2 +[71] Dong Su, Huan Zhang, Hongge Chen, Jinfeng Yi, Pin-Yu Chen, and Yupeng Gao. Is robustness the cost of accuracy? a comprehensive study on the robustness of 18 deep image classification models. In Proceedings of the European Conference on Computer Vision (ECCV), pages 631-648, 2018. 2, 5 +[72] Jeremias Sulam, Ramchandran Muthukumar, and Raman Arora. Adversarial robustness of supervised sparse coding. Advances in Neural Information Processing Systems, 33:2110-2121, 2020. 2 +[73] Christian Szegedy, Wojciech Zaremba, Ilya Sutskever, Joan Bruna, Dumitru Erhan, Ian Goodfellow, and Rob Fergus. Intriguing properties of neural networks. arXiv preprint arXiv:1312.6199, 2013. 1 +[74] Thomas Tanay and Lewis Griffin. A boundary tilting perspective on the phenomenon of adversarial examples. arXiv preprint arXiv:1608.07690, 2016. 2 + +[75] Yee Whye Teh, Max Welling, Simon Osindero, and Geoffrey E Hinton. Energy-based models for sparse overcomplete representations. Journal of Machine Learning Research, 4 (Dec):1235-1260, 2003. 2 +[76] William A Thornton. Matching lights, metamers and human visual response. The Journal of Color and Appearance, 2: 23-29, 1973. 2, 3 +[77] Edward Chace Tolman. *Purposive behavior in animals and men*. Univ of California Press, 1932. 2 +[78] Dimitris Tsipras, Shibani Santurkar, Logan Engstrom, Alexander Turner, and Aleksander Madry. Robustness may be at odds with accuracy. arXiv preprint arXiv:1805.12152, 2018. 2, 5 +[79] Amos Tversky. Features of similarity. Psychological review, 84(4):327, 1977. 2, 6, 7 +[80] Gottfried Wilhelm Freiherr von Leibniz. Leibniz: Discourse on Metaphysics. Open Court Publishing Company, 1902. 3, 2 +[81] Stella Vosniadou and Andrew Ortony. Similarity and analogical reasoning. Cambridge University Press, 1989. 2 +[82] Ernst Heinrich Weber. De Pulsu, resorptione, auditu et tactu: Annotationes anatomicae et physiologicae... CF Koehler, 1831. 3 +[83] Ernst Heinrich Weber. Tastsinn und gemeinschaft. Handworterbuch der Physiologie, 1846. 3 +[84] Timothy Williamson. Identity and Discrimination. Wiley-Blackwell, 1990. 2, 3 +[85] Crispin Wright. On the coherence of vague predicates. Synthese, pages 325-365, 1975. 3 +[86] Günter Wyszecki. Valenzmetrische untersuchung des zusammenhanges zwischen normaler und anomaler trichromasia. 1953. 2, 3 +[87] Gunther Wyszecki and Walter Stanley Stiles. Color science: concepts and methods, quantitative data and formulae. John Wiley & Sons, 2000. 3 +[88] Erik C. Zeeman. The topology of the brain and visual perception. In The Topology of 3-Manifolds. Prentice Hall, NJ, 1962. 2 +[89] E. Christopher Zeeman. Topology of the brain. In Math and Computer Science in Biology and Medicine. H.M. Stationary Office, London, United Kingdom, 1965. 2 +[90] Hongyang Zhang, Yaodong Yu, Jiantao Jiao, Eric Xing, Laurent El Ghaoui, and Michael Jordan. Theoretically principled trade-off between robustness and accuracy. In International conference on machine learning, pages 7472-7482. PMLR, 2019. 2, 5 +[91] Zheng Zhang, Yong Xu, Jian Yang, Xuelong Li, and David Zhang. A survey of sparse representation: algorithms and applications. IEEE access, 3:490-530, 2015. 2 +[92] Li Zhaoping and Keith A May. Psychophysical tests of the hypothesis of a bottom-up saliency map in primary visual cortex. PLoS Computational Biology, 3(4):e62, 2007. 3 \ No newline at end of file diff --git a/2025/Doppelgangers and Adversarial Vulnerability/images.zip b/2025/Doppelgangers and Adversarial Vulnerability/images.zip new file mode 100644 index 0000000000000000000000000000000000000000..80f71b65b51befd8e829c82c0ca25ae56b279af0 --- /dev/null +++ b/2025/Doppelgangers and Adversarial Vulnerability/images.zip @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:6f113fadff0a3844aad9d2db6d2e5c3bfd28c3cc31b5e8ff31c6700787c4951f +size 172250 diff --git a/2025/Doppelgangers and Adversarial Vulnerability/layout.json b/2025/Doppelgangers and Adversarial Vulnerability/layout.json new file mode 100644 index 0000000000000000000000000000000000000000..ca8fc0a5cb2368a7cdfad4e8940d92338911df38 --- /dev/null +++ b/2025/Doppelgangers and Adversarial Vulnerability/layout.json @@ -0,0 +1,16336 @@ +{ + "pdf_info": [ + { + "para_blocks": [ + { + "bbox": [ + 164, + 103, + 446, + 121 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 164, + 103, + 446, + 121 + ], + "spans": [ + { + "bbox": [ + 164, + 103, + 446, + 121 + ], + "type": "text", + "content": "Doppelgängers and Adversarial Vulnerability" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 227, + 144, + 384, + 200 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 227, + 144, + 384, + 200 + ], + "spans": [ + { + "bbox": [ + 227, + 144, + 384, + 200 + ], + "type": "text", + "content": "George Kamberov* \nUniversity of Alaska Anchorage \nAnchorage, AK 99508 \ngkamberov@alaska.edu" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 152, + 227, + 200, + 239 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 152, + 227, + 200, + 239 + ], + "spans": [ + { + "bbox": [ + 152, + 227, + 200, + 239 + ], + "type": "text", + "content": "Abstract" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 55, + 251, + 296, + 336 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 251, + 296, + 336 + ], + "spans": [ + { + "bbox": [ + 55, + 251, + 296, + 336 + ], + "type": "text", + "content": "Machine learning (ML) classifiers can make mistakes that are perceptually and cognitively disturbing to humans. The most notorious examples of such errors are adversarial visual metamers. This paper investigates the phenomenon of adversarial Doppelgängers (AD), which encompasses adversarial visual metamers, and compares the performance and robustness of ML classifiers to human performance." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 56, + 336, + 296, + 574 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 336, + 296, + 574 + ], + "spans": [ + { + "bbox": [ + 56, + 336, + 296, + 574 + ], + "type": "text", + "content": "We find that ADs are inputs that are close to each other with respect to a perceptual metric defined in this paper, and show that ADs are qualitatively different from the usual adversarial examples. The vast majority of classifiers are vulnerable to ADs and robustness-accuracy trade-offs may not improve them. Some classification problems do not admit any AD-robust classifiers because the underlying classes are ambiguous. We provide criteria to determine whether a classification problem is well defined; describe the structure and attributes of AD-robust classifiers; introduce and explore the notions of conceptual entropy and regions of conceptual ambiguity for classifiers that are vulnerable to AD attacks; and discuss methods to bound the AD fooling rate of an attack. We define the notion of classifiers that exhibit hypersensitive behavior, that is, classifiers whose only mistakes are adversarial Doppelgangers. Improving the AD robustness of hypersensitive classifiers is equivalent to improving accuracy. We identify conditions guaranteeing that all classifiers with sufficiently high accuracy are hypersensitive." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 56, + 598, + 135, + 611 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 598, + 135, + 611 + ], + "spans": [ + { + "bbox": [ + 56, + 598, + 135, + 611 + ], + "type": "text", + "content": "1. Introduction" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 55, + 619, + 295, + 667 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 619, + 295, + 667 + ], + "spans": [ + { + "bbox": [ + 55, + 619, + 295, + 667 + ], + "type": "text", + "content": "Perceptual metamers1 are the most striking adversarial examples studied by the machine learning community. Two perceptual metamers are shown in Figure 1. The phenomenon of metamersim studied in the visual domain, in" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 313, + 228, + 555, + 336 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 228, + 555, + 336 + ], + "spans": [ + { + "bbox": [ + 313, + 228, + 555, + 336 + ], + "type": "text", + "content": "cluding perceptual metamers, is a manifestation of the existence of Doppelgängers: different inputs or stimuli that are perceptually indiscriminable. The research community has engaged in active studies of adversarial vulnerability ever since the publication of [73]. Adversarial Doppelgängers, that is, adversarial examples which are Doppelgängers, are qualitatively different from the vast majority of known adversarial examples which humans readily discriminate from correctly classified input samples (Figure 2). There is no" + } + ] + } + ], + "index": 9 + }, + { + "type": "image", + "bbox": [ + 334, + 352, + 433, + 447 + ], + "blocks": [ + { + "bbox": [ + 334, + 352, + 433, + 447 + ], + "lines": [ + { + "bbox": [ + 334, + 352, + 433, + 447 + ], + "spans": [ + { + "bbox": [ + 334, + 352, + 433, + 447 + ], + "type": "image", + "image_path": "62ed6623050a90fa01673828c79577cb3c0d76de6010d623657b9cecd15d7a81.jpg" + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 377, + 452, + 388, + 460 + ], + "lines": [ + { + "bbox": [ + 377, + 452, + 388, + 460 + ], + "spans": [ + { + "bbox": [ + 377, + 452, + 388, + 460 + ], + "type": "text", + "content": "(a)" + } + ] + } + ], + "index": 11, + "angle": 0, + "type": "image_caption" + } + ], + "index": 10 + }, + { + "type": "image", + "bbox": [ + 440, + 352, + 533, + 447 + ], + "blocks": [ + { + "bbox": [ + 440, + 352, + 533, + 447 + ], + "lines": [ + { + "bbox": [ + 440, + 352, + 533, + 447 + ], + "spans": [ + { + "bbox": [ + 440, + 352, + 533, + 447 + ], + "type": "image", + "image_path": "d40457b559e82150da3a0cecb420f2a4a7324d6ee1438310d9d407d5b8f97774.jpg" + } + ] + } + ], + "index": 12, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 476, + 452, + 487, + 460 + ], + "lines": [ + { + "bbox": [ + 476, + 452, + 487, + 460 + ], + "spans": [ + { + "bbox": [ + 476, + 452, + 487, + 460 + ], + "type": "text", + "content": "(b)" + } + ] + } + ], + "index": 13, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 313, + 471, + 555, + 503 + ], + "lines": [ + { + "bbox": [ + 313, + 471, + 555, + 503 + ], + "spans": [ + { + "bbox": [ + 313, + 471, + 555, + 503 + ], + "type": "text", + "content": "Figure 1. Most people cannot discriminate image (a) from image (b). MobileNetV2 classifies the later image as \"persian\" and the former picture as \"taby\"." + } + ] + } + ], + "index": 14, + "angle": 0, + "type": "image_caption" + } + ], + "index": 12 + }, + { + "type": "image", + "bbox": [ + 334, + 548, + 433, + 644 + ], + "blocks": [ + { + "bbox": [ + 334, + 548, + 433, + 644 + ], + "lines": [ + { + "bbox": [ + 334, + 548, + 433, + 644 + ], + "spans": [ + { + "bbox": [ + 334, + 548, + 433, + 644 + ], + "type": "image", + "image_path": "a667691483f835cef81b7244c3cfb8676b8ca325f9fcaa6b04d3b02471edcc62.jpg" + } + ] + } + ], + "index": 15, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 363, + 647, + 405, + 657 + ], + "lines": [ + { + "bbox": [ + 363, + 647, + 405, + 657 + ], + "spans": [ + { + "bbox": [ + 363, + 647, + 405, + 657 + ], + "type": "text", + "content": "(a) Labrador" + } + ] + } + ], + "index": 16, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 313, + 666, + 555, + 700 + ], + "lines": [ + { + "bbox": [ + 313, + 666, + 555, + 700 + ], + "spans": [ + { + "bbox": [ + 313, + 666, + 555, + 700 + ], + "type": "text", + "content": "Figure 2. Applying a Fast Signed Gradient perturbation to the image (a) classified by MobileNetV2 as Labrador yields the image (b) which is classified by MobileNetV2 as Weimeraner." + } + ] + } + ], + "index": 19, + "angle": 0, + "type": "image_caption" + } + ], + "index": 15 + }, + { + "type": "image", + "bbox": [ + 436, + 548, + 533, + 644 + ], + "blocks": [ + { + "bbox": [ + 436, + 548, + 533, + 644 + ], + "lines": [ + { + "bbox": [ + 436, + 548, + 533, + 644 + ], + "spans": [ + { + "bbox": [ + 436, + 548, + 533, + 644 + ], + "type": "image", + "image_path": "d6fc1eea66ea1f1c66e3f3c7e073f39f21c9fec269a3cf921e69cfe48859c98a.jpg" + } + ] + } + ], + "index": 17, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 457, + 647, + 509, + 657 + ], + "lines": [ + { + "bbox": [ + 457, + 647, + 509, + 657 + ], + "spans": [ + { + "bbox": [ + 457, + 647, + 509, + 657 + ], + "type": "text", + "content": "(b) Weimeraner" + } + ] + } + ], + "index": 18, + "angle": 0, + "type": "image_caption" + } + ], + "index": 17 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "spans": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "type": "text", + "content": "CVF" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 145, + 0, + 496, + 37 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 145, + 0, + 496, + 37 + ], + "spans": [ + { + "bbox": [ + 145, + 0, + 496, + 37 + ], + "type": "text", + "content": "This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 55, + 673, + 295, + 713 + ], + "type": "page_footnote", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 673, + 295, + 713 + ], + "spans": [ + { + "bbox": [ + 55, + 673, + 295, + 713 + ], + "type": "text", + "content": "*This material is based upon work supported by the National Science Foundation under Grant No. 2433241 \n1“images that are physically distinct but perceptually indistinguishable”, [6]. See also “metameric images”, [37]" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "type": "text", + "content": "10244" + } + ] + } + ], + "index": 21 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 0 + }, + { + "para_blocks": [ + { + "bbox": [ + 54, + 72, + 294, + 179 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 54, + 72, + 294, + 179 + ], + "spans": [ + { + "bbox": [ + 54, + 72, + 294, + 179 + ], + "type": "text", + "content": "evidence that Doppelgängers can be studied and understood completely using the " + }, + { + "bbox": [ + 54, + 72, + 294, + 179 + ], + "type": "inline_equation", + "content": "\\ell_p" + }, + { + "bbox": [ + 54, + 72, + 294, + 179 + ], + "type": "text", + "content": " norms or more general geodesic distances on manifolds, which have been employed to quantify sample differences and to investigate adversarial examples. Perception and context impose topology on a space of inputs, but it rarely aligns with a manifold topology. We will denote this context-relative topology by " + }, + { + "bbox": [ + 54, + 72, + 294, + 179 + ], + "type": "inline_equation", + "content": "\\tau_{\\delta}" + }, + { + "bbox": [ + 54, + 72, + 294, + 179 + ], + "type": "text", + "content": ". It is defined by the context-relative ability to acquire and deploy knowledge." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 55, + 180, + 295, + 275 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 180, + 295, + 275 + ], + "spans": [ + { + "bbox": [ + 55, + 180, + 295, + 275 + ], + "type": "text", + "content": "In this paper, we explore the context-relative perceptual topologies on a space of inputs " + }, + { + "bbox": [ + 55, + 180, + 295, + 275 + ], + "type": "inline_equation", + "content": "\\mathbf{X}" + }, + { + "bbox": [ + 55, + 180, + 295, + 275 + ], + "type": "text", + "content": " and examine the vulnerability and robustness of machine learning classifiers to adversarial Doppelgangers. We show that, while the majority of classifiers are vulnerable to adversarial Doppelgangers, safe (Doppelganger robust) classifiers do exist if the classification problem is well defined. However, these robust classifiers may be very rare." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 55, + 275, + 295, + 443 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 275, + 295, + 443 + ], + "spans": [ + { + "bbox": [ + 55, + 275, + 295, + 443 + ], + "type": "text", + "content": "In Section 3, we discuss the context-relative notion of (active) indiscriminability and the topology " + }, + { + "bbox": [ + 55, + 275, + 295, + 443 + ], + "type": "inline_equation", + "content": "\\tau_{\\delta}" + }, + { + "bbox": [ + 55, + 275, + 295, + 443 + ], + "type": "text", + "content": " that it induces on a space of inputs. The separability and metric properties of various motivating examples of " + }, + { + "bbox": [ + 55, + 275, + 295, + 443 + ], + "type": "inline_equation", + "content": "\\tau_{\\delta}" + }, + { + "bbox": [ + 55, + 275, + 295, + 443 + ], + "type": "text", + "content": ". Additionally, we show that the distances between Doppelgängers are small if measured by a perceptually-based context-relative metric. We also examine the relation between indiscriminability, IDSicriminability, and feature representations. The existence and structure of Doppelgänger robust classifiers is discussed in Section 4. In Section 5, we investigate the relationship between Doppelgängers and misclassified input samples, define the notion of hypersensitive behavior, and show that improved adversarial Doppelgänger robustness does not have to lead to a reduction in accuracy." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 55, + 443, + 295, + 539 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 443, + 295, + 539 + ], + "spans": [ + { + "bbox": [ + 55, + 443, + 295, + 539 + ], + "type": "text", + "content": "The structure of perceptually regular, i.e., AD robust classifiers are discussed in Section 6. By definition, a classifier is not regular if and only if some inputs can be attacked by adversarial Doppelgangers. Not surprisingly, it turns out that some inputs are more vulnerable than others. In Section 7, we provide measures of adversarial Doppelgangers vulnerability and upper bounds on the fooling rate of an adversarial Doppelganger attack." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 55, + 549, + 139, + 561 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 549, + 139, + 561 + ], + "spans": [ + { + "bbox": [ + 55, + 549, + 139, + 561 + ], + "type": "text", + "content": "2. Related work" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 55, + 568, + 295, + 640 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 568, + 295, + 640 + ], + "spans": [ + { + "bbox": [ + 55, + 568, + 295, + 640 + ], + "type": "text", + "content": "The pair " + }, + { + "bbox": [ + 55, + 568, + 295, + 640 + ], + "type": "inline_equation", + "content": "(\\mathbf{X}, \\text{indiscriminability relation})" + }, + { + "bbox": [ + 55, + 568, + 295, + 640 + ], + "type": "text", + "content": " is a tolerance space. Tolerance spaces, rough sets, and granular computing have been discussed extensively. See [59, 62, 66, 88, 89]. The color and image metamers studied by many authors including [3, 6, 9, 18, 22, 37, 39, 41, 42, 51, 76, 86] are Doppelgangers." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 55, + 641, + 295, + 677 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 641, + 295, + 677 + ], + "spans": [ + { + "bbox": [ + 55, + 641, + 295, + 677 + ], + "type": "text", + "content": "The research on adversarial examples to date builds on the hypothesis that the space of input samples is a metric space " + }, + { + "bbox": [ + 55, + 641, + 295, + 677 + ], + "type": "inline_equation", + "content": "(\\mathbf{X},\\mathrm{dist}_{\\mathbf{X}})" + }, + { + "bbox": [ + 55, + 641, + 295, + 677 + ], + "type": "text", + "content": ". A misclassified input " + }, + { + "bbox": [ + 55, + 641, + 295, + 677 + ], + "type": "inline_equation", + "content": "x^{*}" + }, + { + "bbox": [ + 55, + 641, + 295, + 677 + ], + "type": "text", + "content": " is consid" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 313, + 72, + 553, + 179 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 72, + 553, + 179 + ], + "spans": [ + { + "bbox": [ + 313, + 72, + 553, + 179 + ], + "type": "text", + "content": "ered an adversarial example if it is nearby a correctly classified input sample " + }, + { + "bbox": [ + 313, + 72, + 553, + 179 + ], + "type": "inline_equation", + "content": "x" + }, + { + "bbox": [ + 313, + 72, + 553, + 179 + ], + "type": "text", + "content": ", i.e., " + }, + { + "bbox": [ + 313, + 72, + 553, + 179 + ], + "type": "inline_equation", + "content": "\\mathrm{dist}_{\\mathbf{X}}(x, x^{*})" + }, + { + "bbox": [ + 313, + 72, + 553, + 179 + ], + "type": "text", + "content": " is small. Usually " + }, + { + "bbox": [ + 313, + 72, + 553, + 179 + ], + "type": "inline_equation", + "content": "\\mathbf{X}" + }, + { + "bbox": [ + 313, + 72, + 553, + 179 + ], + "type": "text", + "content": " is assumed to be " + }, + { + "bbox": [ + 313, + 72, + 553, + 179 + ], + "type": "inline_equation", + "content": "\\mathbb{R}^n" + }, + { + "bbox": [ + 313, + 72, + 553, + 179 + ], + "type": "text", + "content": ", endowed with the " + }, + { + "bbox": [ + 313, + 72, + 553, + 179 + ], + "type": "inline_equation", + "content": "\\ell_p" + }, + { + "bbox": [ + 313, + 72, + 553, + 179 + ], + "type": "text", + "content": ", norm, " + }, + { + "bbox": [ + 313, + 72, + 553, + 179 + ], + "type": "inline_equation", + "content": "p = 1, 2, \\ldots, \\infty" + }, + { + "bbox": [ + 313, + 72, + 553, + 179 + ], + "type": "text", + "content": " or at least locally homeomorphic to " + }, + { + "bbox": [ + 313, + 72, + 553, + 179 + ], + "type": "inline_equation", + "content": "\\mathbb{R}^n" + }, + { + "bbox": [ + 313, + 72, + 553, + 179 + ], + "type": "text", + "content": ", i.e., a manifold, equipped with some geodesic distance. Somewhat non surprisingly many authors have shown that every classifier can be attacked with such adversarial examples [5, 17, 25, 52] or at least that this is true in many contexts, [44, 47, 67]." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 313, + 180, + 553, + 300 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 180, + 553, + 300 + ], + "spans": [ + { + "bbox": [ + 313, + 180, + 553, + 300 + ], + "type": "text", + "content": "Other papers indicate that there are paths toward eliminating adversarial examples completely, i.e., it is possible to achieve provable \"adversarial robustness\" by fixing/retraining the classifier [1, 24, 36, 43, 72, 74]. A widely accepted tenet is that \"there is a clear trade-off between accuracy and [adversarial] robustness, and a better performance in testing accuracy in general reduces [adversarial] robustness\", [71]. For empirical evidence for this trade-off and some attempts to explain this phenomenon see [71, 78, 90]." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 313, + 310, + 434, + 324 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 310, + 434, + 324 + ], + "spans": [ + { + "bbox": [ + 313, + 310, + 434, + 324 + ], + "type": "text", + "content": "3. Perceptual Topology" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 313, + 331, + 469, + 342 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 331, + 469, + 342 + ], + "spans": [ + { + "bbox": [ + 313, + 331, + 469, + 342 + ], + "type": "text", + "content": "3.1 Indiscriminability and Topology" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 313, + 342, + 553, + 430 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 342, + 553, + 430 + ], + "spans": [ + { + "bbox": [ + 313, + 342, + 553, + 430 + ], + "type": "text", + "content": "The ability to decide whether one stimulus/input is distinct from another is essential for adaptation, survival, and intelligent life. Intelligent agents are uniquely capable to activate knowledge to judge distinction. Williamson calls this context-relative process discrimination ([84]) and defines a context-relative symmetric and reflexive binary relation denoted by " + }, + { + "bbox": [ + 313, + 342, + 553, + 430 + ], + "type": "inline_equation", + "content": "\\stackrel{\\alpha\\delta}{\\approx}" + }, + { + "bbox": [ + 313, + 342, + 553, + 430 + ], + "type": "text", + "content": " and called indiscriminability:" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 313, + 437, + 553, + 486 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 437, + 553, + 486 + ], + "spans": [ + { + "bbox": [ + 313, + 437, + 553, + 486 + ], + "type": "text", + "content": "Definition 1 ([84]). Two inputs " + }, + { + "bbox": [ + 313, + 437, + 553, + 486 + ], + "type": "inline_equation", + "content": "x" + }, + { + "bbox": [ + 313, + 437, + 553, + 486 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 313, + 437, + 553, + 486 + ], + "type": "inline_equation", + "content": "y" + }, + { + "bbox": [ + 313, + 437, + 553, + 486 + ], + "type": "text", + "content": " are called indiscriminable to a subject at a time " + }, + { + "bbox": [ + 313, + 437, + 553, + 486 + ], + "type": "inline_equation", + "content": "t" + }, + { + "bbox": [ + 313, + 437, + 553, + 486 + ], + "type": "text", + "content": " if and only if at time " + }, + { + "bbox": [ + 313, + 437, + 553, + 486 + ], + "type": "inline_equation", + "content": "t" + }, + { + "bbox": [ + 313, + 437, + 553, + 486 + ], + "type": "text", + "content": " the subject is not able to activate (acquire or employ) the relevant kind of knowledge that " + }, + { + "bbox": [ + 313, + 437, + 553, + 486 + ], + "type": "inline_equation", + "content": "x" + }, + { + "bbox": [ + 313, + 437, + 553, + 486 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 313, + 437, + 553, + 486 + ], + "type": "inline_equation", + "content": "y" + }, + { + "bbox": [ + 313, + 437, + 553, + 486 + ], + "type": "text", + "content": " are distinct." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 313, + 494, + 553, + 517 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 494, + 553, + 517 + ], + "spans": [ + { + "bbox": [ + 313, + 494, + 553, + 517 + ], + "type": "text", + "content": "Indiscriminability generates a context-relative topology on the set of inputs " + }, + { + "bbox": [ + 313, + 494, + 553, + 517 + ], + "type": "inline_equation", + "content": "\\mathbf{X}" + }, + { + "bbox": [ + 313, + 494, + 553, + 517 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 313, + 525, + 553, + 599 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 525, + 553, + 599 + ], + "spans": [ + { + "bbox": [ + 313, + 525, + 553, + 599 + ], + "type": "text", + "content": "Definition 2. The the phenomenal neighborhood of an input " + }, + { + "bbox": [ + 313, + 525, + 553, + 599 + ], + "type": "inline_equation", + "content": "x \\in \\mathbf{X}" + }, + { + "bbox": [ + 313, + 525, + 553, + 599 + ], + "type": "text", + "content": ", is the set " + }, + { + "bbox": [ + 313, + 525, + 553, + 599 + ], + "type": "inline_equation", + "content": "\\mathfrak{d}(x) = \\left\\{y \\in \\mathbf{X} : y \\stackrel{\\alpha \\delta}{\\approx} x\\right\\}" + }, + { + "bbox": [ + 313, + 525, + 553, + 599 + ], + "type": "text", + "content": ". A point " + }, + { + "bbox": [ + 313, + 525, + 553, + 599 + ], + "type": "inline_equation", + "content": "y \\in \\mathfrak{d}(x) \\setminus \\{x\\}" + }, + { + "bbox": [ + 313, + 525, + 553, + 599 + ], + "type": "text", + "content": " is called a Doppelgänger [of " + }, + { + "bbox": [ + 313, + 525, + 553, + 599 + ], + "type": "inline_equation", + "content": "x" + }, + { + "bbox": [ + 313, + 525, + 553, + 599 + ], + "type": "text", + "content": "]. The perceptual topology " + }, + { + "bbox": [ + 313, + 525, + 553, + 599 + ], + "type": "inline_equation", + "content": "\\tau_{\\delta}" + }, + { + "bbox": [ + 313, + 525, + 553, + 599 + ], + "type": "text", + "content": " is the topology generated by the sub-basis " + }, + { + "bbox": [ + 313, + 525, + 553, + 599 + ], + "type": "inline_equation", + "content": "\\mathfrak{D}_{\\alpha \\delta} = \\{\\mathfrak{d}(x)\\}_{x \\in \\mathbf{X}}" + }, + { + "bbox": [ + 313, + 525, + 553, + 599 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 313, + 605, + 553, + 629 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 605, + 553, + 629 + ], + "spans": [ + { + "bbox": [ + 313, + 605, + 553, + 629 + ], + "type": "text", + "content": "Example 0: An input " + }, + { + "bbox": [ + 313, + 605, + 553, + 629 + ], + "type": "inline_equation", + "content": "x \\in \\mathbf{X}" + }, + { + "bbox": [ + 313, + 605, + 553, + 629 + ], + "type": "text", + "content": " is called optimal if it does not have non-trivial/non-identical Doppelgangers, i.e.," + } + ] + } + ], + "index": 15 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 313, + 635, + 553, + 665 + ], + "type": "page_footnote", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 635, + 553, + 665 + ], + "spans": [ + { + "bbox": [ + 313, + 635, + 553, + 665 + ], + "type": "inline_equation", + "content": "^{3}" + }, + { + "bbox": [ + 313, + 635, + 553, + 665 + ], + "type": "text", + "content": "Context and its role in discrimination and similarity judgments have been studied extensively and by many authors including [11, 23, 30, 45, 60, 77, 79, 81]." + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 313, + 665, + 553, + 712 + ], + "type": "page_footnote", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 665, + 553, + 712 + ], + "spans": [ + { + "bbox": [ + 313, + 665, + 553, + 712 + ], + "type": "text", + "content": "4Some authors refer to indiscriminability as active indiscriminability, see [16]. Poincaré discusses indiscriminability in [58] but refers to it as indiscriminability. Similarly Poston studies indistinguishability, basing it on the \"limit of discrimination\" of the biological senses and instruments, [59]." + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 55, + 683, + 295, + 713 + ], + "type": "page_footnote", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 683, + 295, + 713 + ], + "spans": [ + { + "bbox": [ + 55, + 683, + 295, + 713 + ], + "type": "text", + "content": "2 Some metamers arising in other fields including biology and chemistry are not Doppelgängers, for example, segments in many earthworms are considered metamers but are visually discriminable." + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "text", + "content": "10245" + } + ] + } + ], + "index": 20 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 1 + }, + { + "para_blocks": [ + { + "bbox": [ + 55, + 71, + 294, + 182 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 71, + 294, + 182 + ], + "spans": [ + { + "bbox": [ + 55, + 71, + 294, + 182 + ], + "type": "text", + "content": "if " + }, + { + "bbox": [ + 55, + 71, + 294, + 182 + ], + "type": "inline_equation", + "content": "\\mathfrak{d}(x) = \\{x\\}" + }, + { + "bbox": [ + 55, + 71, + 294, + 182 + ], + "type": "text", + "content": ". In particular, if " + }, + { + "bbox": [ + 55, + 71, + 294, + 182 + ], + "type": "inline_equation", + "content": "\\stackrel{\\infty \\delta}{\\approx}" + }, + { + "bbox": [ + 55, + 71, + 294, + 182 + ], + "type": "text", + "content": " is the identity relationship " + }, + { + "bbox": [ + 55, + 71, + 294, + 182 + ], + "type": "inline_equation", + "content": "=" + }, + { + "bbox": [ + 55, + 71, + 294, + 182 + ], + "type": "text", + "content": " (i.e., all inputs are optimal within the given context), then the perceptual topology " + }, + { + "bbox": [ + 55, + 71, + 294, + 182 + ], + "type": "inline_equation", + "content": "\\tau_{\\delta}" + }, + { + "bbox": [ + 55, + 71, + 294, + 182 + ], + "type": "text", + "content": " is discrete. However, the finiteness of human observations (they are subject to finite time and finite work constraints) and the bounded rationality constraints imposed by the limitations on the availability of information and computational capabilities to humans, [68], indicate that the scenario " + }, + { + "bbox": [ + 55, + 71, + 294, + 182 + ], + "type": "inline_equation", + "content": "\\mathfrak{d}(x) = \\{x\\}" + }, + { + "bbox": [ + 55, + 71, + 294, + 182 + ], + "type": "text", + "content": " for every " + }, + { + "bbox": [ + 55, + 71, + 294, + 182 + ], + "type": "inline_equation", + "content": "x \\in \\mathbf{X}" + }, + { + "bbox": [ + 55, + 71, + 294, + 182 + ], + "type": "text", + "content": " may be highly unlikely." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 55, + 183, + 294, + 279 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 183, + 294, + 279 + ], + "spans": [ + { + "bbox": [ + 55, + 183, + 294, + 279 + ], + "type": "text", + "content": "Often reflexive binary relations are defined and discussed as coverings of the underlying space. Indeed, every reflexive binary relation (RBR) " + }, + { + "bbox": [ + 55, + 183, + 294, + 279 + ], + "type": "inline_equation", + "content": "\\approx" + }, + { + "bbox": [ + 55, + 183, + 294, + 279 + ], + "type": "text", + "content": " on " + }, + { + "bbox": [ + 55, + 183, + 294, + 279 + ], + "type": "inline_equation", + "content": "\\mathbf{X}" + }, + { + "bbox": [ + 55, + 183, + 294, + 279 + ], + "type": "text", + "content": " defines a covering " + }, + { + "bbox": [ + 55, + 183, + 294, + 279 + ], + "type": "inline_equation", + "content": "\\{\\mathfrak{g}(x) = \\{y:y\\approx x\\} \\}_{x\\in \\mathbf{X}}" + }, + { + "bbox": [ + 55, + 183, + 294, + 279 + ], + "type": "text", + "content": ", of " + }, + { + "bbox": [ + 55, + 183, + 294, + 279 + ], + "type": "inline_equation", + "content": "\\mathbf{X}" + }, + { + "bbox": [ + 55, + 183, + 294, + 279 + ], + "type": "text", + "content": " and vice versa one can define reflexive binary relationships through coverings of " + }, + { + "bbox": [ + 55, + 183, + 294, + 279 + ], + "type": "inline_equation", + "content": "\\mathbf{X}" + }, + { + "bbox": [ + 55, + 183, + 294, + 279 + ], + "type": "text", + "content": ". See Appendix, Section A. In particular, the perceptual topology, " + }, + { + "bbox": [ + 55, + 183, + 294, + 279 + ], + "type": "inline_equation", + "content": "\\tau_{\\delta}" + }, + { + "bbox": [ + 55, + 183, + 294, + 279 + ], + "type": "text", + "content": ", is a tolerable topology, cf., Definition 10 in Appendix, Section A." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 55, + 281, + 294, + 341 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 281, + 294, + 341 + ], + "spans": [ + { + "bbox": [ + 55, + 281, + 294, + 341 + ], + "type": "text", + "content": "The active psychophysics research on just noticeable difference initiated by Weber and Fechner, [19, 20, 82, 83], provides one of the few classes of examples where we have empirically supported understanding of the perceptual topology." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 55, + 353, + 295, + 389 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 353, + 295, + 389 + ], + "spans": [ + { + "bbox": [ + 55, + 353, + 295, + 389 + ], + "type": "text", + "content": "Example 1: Let " + }, + { + "bbox": [ + 55, + 353, + 295, + 389 + ], + "type": "inline_equation", + "content": "\\mathbf{X}" + }, + { + "bbox": [ + 55, + 353, + 295, + 389 + ], + "type": "text", + "content": " be the closed bounded interval " + }, + { + "bbox": [ + 55, + 353, + 295, + 389 + ], + "type": "inline_equation", + "content": "[a,b]\\subset" + }, + { + "bbox": [ + 55, + 353, + 295, + 389 + ], + "type": "inline_equation", + "content": "(0, + \\infty)" + }, + { + "bbox": [ + 55, + 353, + 295, + 389 + ], + "type": "text", + "content": ". Suppose that Weber's law holds and let " + }, + { + "bbox": [ + 55, + 353, + 295, + 389 + ], + "type": "inline_equation", + "content": "k > 0" + }, + { + "bbox": [ + 55, + 353, + 295, + 389 + ], + "type": "text", + "content": " be the Weber constant. Let " + }, + { + "bbox": [ + 55, + 353, + 295, + 389 + ], + "type": "inline_equation", + "content": "w = 1 + k" + }, + { + "bbox": [ + 55, + 353, + 295, + 389 + ], + "type": "text", + "content": ", then" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 92, + 400, + 294, + 444 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 92, + 400, + 294, + 444 + ], + "spans": [ + { + "bbox": [ + 92, + 400, + 294, + 444 + ], + "type": "interline_equation", + "content": "\\mathfrak {d} (x) = \\left\\{ \\begin{array}{l l} {[ a, x w),} & {a \\leq x < a w} \\\\ {(x / w, x w),} & {a w \\leq x \\leq b / w} \\\\ {(x / w, b ],} & {b / w < x \\leq b.} \\end{array} \\right. \\tag {1}", + "image_path": "4fb926d4e7d29e4d1743a77e64e0ed1163f2a3673684cc0b249fc436fa0e4d4d.jpg" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 55, + 455, + 295, + 502 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 455, + 295, + 502 + ], + "spans": [ + { + "bbox": [ + 55, + 455, + 295, + 502 + ], + "type": "text", + "content": "The covering " + }, + { + "bbox": [ + 55, + 455, + 295, + 502 + ], + "type": "inline_equation", + "content": "\\{\\mathfrak{d}(x)\\}_{x\\in [a,b]}" + }, + { + "bbox": [ + 55, + 455, + 295, + 502 + ], + "type": "text", + "content": " defines a symmetric RBR but the relation is not transitive. The corresponding perceptual topology is " + }, + { + "bbox": [ + 55, + 455, + 295, + 502 + ], + "type": "inline_equation", + "content": "T_{0}" + }, + { + "bbox": [ + 55, + 455, + 295, + 502 + ], + "type": "text", + "content": " but not " + }, + { + "bbox": [ + 55, + 455, + 295, + 502 + ], + "type": "inline_equation", + "content": "T_{1}" + }, + { + "bbox": [ + 55, + 455, + 295, + 502 + ], + "type": "text", + "content": ", and the topology is not pseudometric." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 55, + 504, + 295, + 556 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 504, + 295, + 556 + ], + "spans": [ + { + "bbox": [ + 55, + 504, + 295, + 556 + ], + "type": "text", + "content": "The transitivity or more more often the lack of transitivity of " + }, + { + "bbox": [ + 55, + 504, + 295, + 556 + ], + "type": "inline_equation", + "content": "\\stackrel{\\alpha\\delta}{\\approx}" + }, + { + "bbox": [ + 55, + 504, + 295, + 556 + ], + "type": "text", + "content": " have been studied extensively and proven or postulated in many human experiences, [2, 6, 10, 16, 26, 27, 32, 54, 61, 84, 85]." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 55, + 567, + 295, + 672 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 567, + 295, + 672 + ], + "spans": [ + { + "bbox": [ + 55, + 567, + 295, + 672 + ], + "type": "text", + "content": "Definition 3. We will denote by " + }, + { + "bbox": [ + 55, + 567, + 295, + 672 + ], + "type": "inline_equation", + "content": "\\sim_{\\sigma}" + }, + { + "bbox": [ + 55, + 567, + 295, + 672 + ], + "type": "text", + "content": " the transitive closure of the indiciminability relation " + }, + { + "bbox": [ + 55, + 567, + 295, + 672 + ], + "type": "inline_equation", + "content": "\\stackrel{\\alpha \\delta}{\\approx}" + }, + { + "bbox": [ + 55, + 567, + 295, + 672 + ], + "type": "text", + "content": " on " + }, + { + "bbox": [ + 55, + 567, + 295, + 672 + ], + "type": "inline_equation", + "content": "\\mathbf{X}" + }, + { + "bbox": [ + 55, + 567, + 295, + 672 + ], + "type": "text", + "content": ". It is defined explicitly as " + }, + { + "bbox": [ + 55, + 567, + 295, + 672 + ], + "type": "inline_equation", + "content": "x \\sim_{\\sigma} y" + }, + { + "bbox": [ + 55, + 567, + 295, + 672 + ], + "type": "text", + "content": " iff there exists a finite chain of Doppelgangers " + }, + { + "bbox": [ + 55, + 567, + 295, + 672 + ], + "type": "inline_equation", + "content": "x = x_0 \\stackrel{\\alpha \\delta}{\\approx} x_1 \\stackrel{\\alpha \\delta}{\\approx} x_2 \\stackrel{\\alpha \\delta}{\\approx} \\dots \\stackrel{\\alpha \\delta}{\\approx} x_n = y" + }, + { + "bbox": [ + 55, + 567, + 295, + 672 + ], + "type": "text", + "content": ". We will call the relation " + }, + { + "bbox": [ + 55, + 567, + 295, + 672 + ], + "type": "inline_equation", + "content": "\\sim_{\\sigma}" + }, + { + "bbox": [ + 55, + 567, + 295, + 672 + ], + "type": "text", + "content": " perceptual metamorphy and will refer to any two inputs " + }, + { + "bbox": [ + 55, + 567, + 295, + 672 + ], + "type": "inline_equation", + "content": "x \\sim_{\\sigma} y" + }, + { + "bbox": [ + 55, + 567, + 295, + 672 + ], + "type": "text", + "content": " as metamorphic. Extending Pawlak's terminology, [53] we call the equivalence classes in " + }, + { + "bbox": [ + 55, + 567, + 295, + 672 + ], + "type": "inline_equation", + "content": "\\mathbf{X} / \\sim_{\\sigma}" + }, + { + "bbox": [ + 55, + 567, + 295, + 672 + ], + "type": "text", + "content": " (perceptually) elementary sets." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 313, + 72, + 553, + 243 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 72, + 553, + 243 + ], + "spans": [ + { + "bbox": [ + 313, + 72, + 553, + 243 + ], + "type": "text", + "content": "Example 2: If the indiscriminability relation is transitive, then each " + }, + { + "bbox": [ + 313, + 72, + 553, + 243 + ], + "type": "inline_equation", + "content": "\\mathfrak{d}(x)" + }, + { + "bbox": [ + 313, + 72, + 553, + 243 + ], + "type": "text", + "content": " is an elementary set. The perceptual topology may be optimal (recall Example 0) or not. In the former case it is Hausdorff and in fact " + }, + { + "bbox": [ + 313, + 72, + 553, + 243 + ], + "type": "inline_equation", + "content": "(\\mathbf{X},\\tau_{\\delta})" + }, + { + "bbox": [ + 313, + 72, + 553, + 243 + ], + "type": "text", + "content": " is a discrete manifold, in the later case the topology " + }, + { + "bbox": [ + 313, + 72, + 553, + 243 + ], + "type": "inline_equation", + "content": "\\tau_{\\delta}" + }, + { + "bbox": [ + 313, + 72, + 553, + 243 + ], + "type": "text", + "content": " is not " + }, + { + "bbox": [ + 313, + 72, + 553, + 243 + ], + "type": "inline_equation", + "content": "T_0" + }, + { + "bbox": [ + 313, + 72, + 553, + 243 + ], + "type": "text", + "content": ". See Part A.1 in the Appendix. Human visual perception provides a fundamental example where " + }, + { + "bbox": [ + 313, + 72, + 553, + 243 + ], + "type": "inline_equation", + "content": "\\mathfrak{d}(x)\\neq \\{x\\}" + }, + { + "bbox": [ + 313, + 72, + 553, + 243 + ], + "type": "text", + "content": ". If two images " + }, + { + "bbox": [ + 313, + 72, + 553, + 243 + ], + "type": "inline_equation", + "content": "x" + }, + { + "bbox": [ + 313, + 72, + 553, + 243 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 313, + 72, + 553, + 243 + ], + "type": "inline_equation", + "content": "y" + }, + { + "bbox": [ + 313, + 72, + 553, + 243 + ], + "type": "text", + "content": " differ only in unattended regions for example due to low saliency values (cf. [92]), then " + }, + { + "bbox": [ + 313, + 72, + 553, + 243 + ], + "type": "inline_equation", + "content": "x\\stackrel {\\alpha \\delta}{\\approx}y" + }, + { + "bbox": [ + 313, + 72, + 553, + 243 + ], + "type": "text", + "content": ". Visual metamers have been studied by many authors, including [3, 6, 9, 18, 22, 28, 37, 39, 41, 42, 51, 76, 86]. In these studies, the input space is assumed to be endowed with Grassmann structure (see [39]), and in particular, the indiscriminability relation is transitive." + }, + { + "bbox": [ + 313, + 72, + 553, + 243 + ], + "type": "inline_equation", + "content": "^6" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 313, + 243, + 553, + 347 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 243, + 553, + 347 + ], + "spans": [ + { + "bbox": [ + 313, + 243, + 553, + 347 + ], + "type": "text", + "content": "We are not aware of perceptual topologies that are metric. Still, every Doppelgänger " + }, + { + "bbox": [ + 313, + 243, + 553, + 347 + ], + "type": "inline_equation", + "content": "y \\stackrel{\\alpha \\delta}{\\approx} x" + }, + { + "bbox": [ + 313, + 243, + 553, + 347 + ], + "type": "text", + "content": " of an input " + }, + { + "bbox": [ + 313, + 243, + 553, + 347 + ], + "type": "inline_equation", + "content": "x" + }, + { + "bbox": [ + 313, + 243, + 553, + 347 + ], + "type": "text", + "content": " is a small perturbation of " + }, + { + "bbox": [ + 313, + 243, + 553, + 347 + ], + "type": "inline_equation", + "content": "x" + }, + { + "bbox": [ + 313, + 243, + 553, + 347 + ], + "type": "text", + "content": " in the sense that the " + }, + { + "bbox": [ + 313, + 243, + 553, + 347 + ], + "type": "inline_equation", + "content": "y" + }, + { + "bbox": [ + 313, + 243, + 553, + 347 + ], + "type": "text", + "content": " is a nearest neighbor of " + }, + { + "bbox": [ + 313, + 243, + 553, + 347 + ], + "type": "inline_equation", + "content": "x" + }, + { + "bbox": [ + 313, + 243, + 553, + 347 + ], + "type": "text", + "content": " with respect to an appropriate metric " + }, + { + "bbox": [ + 313, + 243, + 553, + 347 + ], + "type": "inline_equation", + "content": "d_w(\\cdot, \\cdot)" + }, + { + "bbox": [ + 313, + 243, + 553, + 347 + ], + "type": "text", + "content": " on " + }, + { + "bbox": [ + 313, + 243, + 553, + 347 + ], + "type": "inline_equation", + "content": "\\mathbf{X}" + }, + { + "bbox": [ + 313, + 243, + 553, + 347 + ], + "type": "text", + "content": ". Indeed, let the discrimination graph " + }, + { + "bbox": [ + 313, + 243, + 553, + 347 + ], + "type": "inline_equation", + "content": "\\Gamma(\\mathbf{X}, E_{\\alpha \\delta})" + }, + { + "bbox": [ + 313, + 243, + 553, + 347 + ], + "type": "text", + "content": " be the undirected simple graph, where " + }, + { + "bbox": [ + 313, + 243, + 553, + 347 + ], + "type": "inline_equation", + "content": "\\mathbf{X}" + }, + { + "bbox": [ + 313, + 243, + 553, + 347 + ], + "type": "text", + "content": " is the set of vertices and we say that there is an edge " + }, + { + "bbox": [ + 313, + 243, + 553, + 347 + ], + "type": "inline_equation", + "content": "\\{x, y\\} \\in E_{\\alpha \\delta}" + }, + { + "bbox": [ + 313, + 243, + 553, + 347 + ], + "type": "text", + "content": " between the vertices " + }, + { + "bbox": [ + 313, + 243, + 553, + 347 + ], + "type": "inline_equation", + "content": "x, y \\in \\mathbf{X}" + }, + { + "bbox": [ + 313, + 243, + 553, + 347 + ], + "type": "text", + "content": " iff " + }, + { + "bbox": [ + 313, + 243, + 553, + 347 + ], + "type": "inline_equation", + "content": "x \\stackrel{\\alpha \\delta}{\\approx} y" + }, + { + "bbox": [ + 313, + 243, + 553, + 347 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 313, + 354, + 553, + 415 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 354, + 553, + 415 + ], + "spans": [ + { + "bbox": [ + 313, + 354, + 553, + 415 + ], + "type": "text", + "content": "Definition 4. We will call the discrimination graph distance " + }, + { + "bbox": [ + 313, + 354, + 553, + 415 + ], + "type": "inline_equation", + "content": "d_{\\infty}(x,y)" + }, + { + "bbox": [ + 313, + 354, + 553, + 415 + ], + "type": "text", + "content": " between the vertices " + }, + { + "bbox": [ + 313, + 354, + 553, + 415 + ], + "type": "inline_equation", + "content": "x" + }, + { + "bbox": [ + 313, + 354, + 553, + 415 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 313, + 354, + 553, + 415 + ], + "type": "inline_equation", + "content": "y" + }, + { + "bbox": [ + 313, + 354, + 553, + 415 + ], + "type": "text", + "content": " and, in particular, " + }, + { + "bbox": [ + 313, + 354, + 553, + 415 + ], + "type": "inline_equation", + "content": "d_{\\infty}(x,y) = \\infty" + }, + { + "bbox": [ + 313, + 354, + 553, + 415 + ], + "type": "text", + "content": " iff " + }, + { + "bbox": [ + 313, + 354, + 553, + 415 + ], + "type": "inline_equation", + "content": "x \\not\\simeq_{\\sigma} y" + }, + { + "bbox": [ + 313, + 354, + 553, + 415 + ], + "type": "text", + "content": " the extended perceptual distance. The perceptual distance is the metric " + }, + { + "bbox": [ + 313, + 354, + 553, + 415 + ], + "type": "inline_equation", + "content": "d_w: \\mathbf{X} \\times \\mathbf{X} \\to [0,1]" + }, + { + "bbox": [ + 313, + 354, + 553, + 415 + ], + "type": "text", + "content": " defined by:" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 354, + 421, + 553, + 449 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 354, + 421, + 553, + 449 + ], + "spans": [ + { + "bbox": [ + 354, + 421, + 553, + 449 + ], + "type": "interline_equation", + "content": "d _ {w} (x, y) = \\frac {d _ {\\infty} (x , y)}{1 + d _ {\\infty} (x , y)}, \\forall x, y \\in \\mathbf {X}. \\tag {2}", + "image_path": "c65fdfbd42f4d5a32fc06983ecca9c231a5fd8196af9dcfa032953d377e04f64.jpg" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 313, + 453, + 553, + 489 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 453, + 553, + 489 + ], + "spans": [ + { + "bbox": [ + 313, + 453, + 553, + 489 + ], + "type": "text", + "content": "The metric " + }, + { + "bbox": [ + 313, + 453, + 553, + 489 + ], + "type": "inline_equation", + "content": "d_w" + }, + { + "bbox": [ + 313, + 453, + 553, + 489 + ], + "type": "text", + "content": " does not generate the perceptual topology. It is certainly not the usual " + }, + { + "bbox": [ + 313, + 453, + 553, + 489 + ], + "type": "inline_equation", + "content": "l_p" + }, + { + "bbox": [ + 313, + 453, + 553, + 489 + ], + "type": "text", + "content": " or any other manifold metric used in ML." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 313, + 501, + 511, + 513 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 501, + 511, + 513 + ], + "spans": [ + { + "bbox": [ + 313, + 501, + 511, + 513 + ], + "type": "text", + "content": "3.2 Indiscriminabile may not be Indiscernible" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 313, + 514, + 553, + 589 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 514, + 553, + 589 + ], + "spans": [ + { + "bbox": [ + 313, + 514, + 553, + 589 + ], + "type": "text", + "content": "Let " + }, + { + "bbox": [ + 313, + 514, + 553, + 589 + ], + "type": "inline_equation", + "content": "\\Phi" + }, + { + "bbox": [ + 313, + 514, + 553, + 589 + ], + "type": "text", + "content": " be the space of all features of the inputs/stimuli " + }, + { + "bbox": [ + 313, + 514, + 553, + 589 + ], + "type": "inline_equation", + "content": "x\\in \\mathbf{X}" + }, + { + "bbox": [ + 313, + 514, + 553, + 589 + ], + "type": "text", + "content": " and let " + }, + { + "bbox": [ + 313, + 514, + 553, + 589 + ], + "type": "inline_equation", + "content": "\\Phi_x\\subset \\Phi" + }, + { + "bbox": [ + 313, + 514, + 553, + 589 + ], + "type": "text", + "content": " be the set of features attributed to " + }, + { + "bbox": [ + 313, + 514, + 553, + 589 + ], + "type": "inline_equation", + "content": "x" + }, + { + "bbox": [ + 313, + 514, + 553, + 589 + ], + "type": "text", + "content": " in a given context. Following [21], we say that " + }, + { + "bbox": [ + 313, + 514, + 553, + 589 + ], + "type": "inline_equation", + "content": "x" + }, + { + "bbox": [ + 313, + 514, + 553, + 589 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 313, + 514, + 553, + 589 + ], + "type": "inline_equation", + "content": "y" + }, + { + "bbox": [ + 313, + 514, + 553, + 589 + ], + "type": "text", + "content": " are indiscernible, in a given context, if " + }, + { + "bbox": [ + 313, + 514, + 553, + 589 + ], + "type": "inline_equation", + "content": "\\Phi_x = \\Phi_y" + }, + { + "bbox": [ + 313, + 514, + 553, + 589 + ], + "type": "text", + "content": ". Many researchers use the terms indiscriminability and indiscernibility as synonyms. This is only accurate when " + }, + { + "bbox": [ + 313, + 514, + 553, + 589 + ], + "type": "inline_equation", + "content": "\\stackrel{\\alpha\\delta}{\\approx}" + }, + { + "bbox": [ + 313, + 514, + 553, + 589 + ], + "type": "text", + "content": " is transi" + } + ] + } + ], + "index": 14 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 313, + 595, + 553, + 614 + ], + "type": "page_footnote", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 595, + 553, + 614 + ], + "spans": [ + { + "bbox": [ + 313, + 595, + 553, + 614 + ], + "type": "text", + "content": "6In these studies, indiscriminable inputs are referred to as \"matching\" or \"metameric\", or \"alike\"." + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 313, + 615, + 553, + 654 + ], + "type": "page_footnote", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 615, + 553, + 654 + ], + "spans": [ + { + "bbox": [ + 313, + 615, + 553, + 654 + ], + "type": "text", + "content": "The open metric ball " + }, + { + "bbox": [ + 313, + 615, + 553, + 654 + ], + "type": "inline_equation", + "content": "\\tilde{B}_{1/2}^w(x)" + }, + { + "bbox": [ + 313, + 615, + 553, + 654 + ], + "type": "text", + "content": " equals " + }, + { + "bbox": [ + 313, + 615, + 553, + 654 + ], + "type": "inline_equation", + "content": "\\{x\\}" + }, + { + "bbox": [ + 313, + 615, + 553, + 654 + ], + "type": "text", + "content": ", for all inputs " + }, + { + "bbox": [ + 313, + 615, + 553, + 654 + ], + "type": "inline_equation", + "content": "x \\in \\mathbf{X}" + }, + { + "bbox": [ + 313, + 615, + 553, + 654 + ], + "type": "text", + "content": ". On the other hand, the finiteness of human observations and the hypothesis of bounded rationality suggest that biologically plausible perceptual topologies are not discrete." + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 313, + 655, + 553, + 684 + ], + "type": "page_footnote", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 655, + 553, + 684 + ], + "spans": [ + { + "bbox": [ + 313, + 655, + 553, + 684 + ], + "type": "text", + "content": "The existence of the extended metric was hinted at in [46]; it was discussed in a related context in [65] and rediscovered and exploited in [59]. For more discussion see Part F.1 in the Appendix." + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 313, + 684, + 553, + 713 + ], + "type": "page_footnote", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 684, + 553, + 713 + ], + "spans": [ + { + "bbox": [ + 313, + 684, + 553, + 713 + ], + "type": "text", + "content": "9Leibniz discussed indiscernibility and postulated the Principle/Law of Identity of Indiscernibles, " + }, + { + "bbox": [ + 313, + 684, + 553, + 713 + ], + "type": "inline_equation", + "content": "(\\Phi_{x} = \\Phi_{y})\\Rightarrow (x = y)" + }, + { + "bbox": [ + 313, + 684, + 553, + 713 + ], + "type": "text", + "content": " , in [80] and in the third and fourth papers addressed to Samuel Clarke, [8]." + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 55, + 693, + 295, + 713 + ], + "type": "page_footnote", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 693, + 295, + 713 + ], + "spans": [ + { + "bbox": [ + 55, + 693, + 295, + 713 + ], + "type": "text", + "content": "5Optimal objects and light sources have been described and studied in colorimetry, [42, 87]." + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "text", + "content": "10246" + } + ] + } + ], + "index": 21 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 2 + }, + { + "para_blocks": [ + { + "bbox": [ + 55, + 72, + 296, + 122 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 72, + 296, + 122 + ], + "spans": [ + { + "bbox": [ + 55, + 72, + 296, + 122 + ], + "type": "text", + "content": "tive." + }, + { + "bbox": [ + 55, + 72, + 296, + 122 + ], + "type": "inline_equation", + "content": "^{10}" + }, + { + "bbox": [ + 55, + 72, + 296, + 122 + ], + "type": "text", + "content": " In general, the relationship between indiscriminability and indiscernibility is not well understood. However, indiscernibility implies indiscriminability if " + }, + { + "bbox": [ + 55, + 72, + 296, + 122 + ], + "type": "inline_equation", + "content": "\\{\\Phi_x\\}_{x\\in \\mathbf{X}}" + }, + { + "bbox": [ + 55, + 72, + 296, + 122 + ], + "type": "text", + "content": " is a perceptually discriminative feature representation, i.e.," + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 119, + 129, + 295, + 148 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 119, + 129, + 295, + 148 + ], + "spans": [ + { + "bbox": [ + 119, + 129, + 295, + 148 + ], + "type": "interline_equation", + "content": "\\Phi_ {x} \\bigcap \\Phi_ {y} \\neq \\varnothing \\Longleftrightarrow x \\stackrel {\\circledast} {\\approx} y. \\tag {3}", + "image_path": "d945577d11dd5d0cdb144a7ff913eb480c901f028008793d9898448e4c74bb88.jpg" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 55, + 155, + 296, + 240 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 155, + 296, + 240 + ], + "spans": [ + { + "bbox": [ + 55, + 155, + 296, + 240 + ], + "type": "text", + "content": "The biological plausibility of discriminative feature representations is an open question. Still, they provide insight into the structure of the perceptual topology. The attributed discriminative features represent structures of Doppelgangers. Indeed, let " + }, + { + "bbox": [ + 55, + 155, + 296, + 240 + ], + "type": "inline_equation", + "content": "\\{\\Phi_x\\}_{x\\in \\mathbf{X}}" + }, + { + "bbox": [ + 55, + 155, + 296, + 240 + ], + "type": "text", + "content": " be a feature representation and let " + }, + { + "bbox": [ + 55, + 155, + 296, + 240 + ], + "type": "inline_equation", + "content": "\\operatorname {cl}(\\xi)" + }, + { + "bbox": [ + 55, + 155, + 296, + 240 + ], + "type": "text", + "content": " be the context-dependent semantic cluster of inputs sharing the feature " + }, + { + "bbox": [ + 55, + 155, + 296, + 240 + ], + "type": "inline_equation", + "content": "\\xi \\in \\Phi" + }, + { + "bbox": [ + 55, + 155, + 296, + 240 + ], + "type": "text", + "content": ". Specifically," + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 107, + 247, + 295, + 262 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 107, + 247, + 295, + 262 + ], + "spans": [ + { + "bbox": [ + 107, + 247, + 295, + 262 + ], + "type": "interline_equation", + "content": "\\mathfrak {c l} (\\xi) = \\left\\{x \\in \\mathbf {X}, \\text {s . t . ,} \\xi \\in \\Phi_ {x} \\right\\}. ^ {1 2} \\tag {4}", + "image_path": "0e13cd7414686b3b0a62bb8da33c2b6d93174deab3c19a950151fdebafad047a.jpg" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 55, + 269, + 296, + 342 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 269, + 296, + 342 + ], + "spans": [ + { + "bbox": [ + 55, + 269, + 296, + 342 + ], + "type": "text", + "content": "In particular, if " + }, + { + "bbox": [ + 55, + 269, + 296, + 342 + ], + "type": "inline_equation", + "content": "\\{\\Phi_x\\}_{x\\in \\mathbf{X}}" + }, + { + "bbox": [ + 55, + 269, + 296, + 342 + ], + "type": "text", + "content": " is a discriminative feature representation, then every attributed discriminative feature " + }, + { + "bbox": [ + 55, + 269, + 296, + 342 + ], + "type": "inline_equation", + "content": "\\xi \\in \\Phi_{x}" + }, + { + "bbox": [ + 55, + 269, + 296, + 342 + ], + "type": "text", + "content": " is associated to, and in some way explained by, a collection of Doppelgangers since " + }, + { + "bbox": [ + 55, + 269, + 296, + 342 + ], + "type": "inline_equation", + "content": "\\mathfrak{cl}(\\xi)\\subset \\mathfrak{d}(x)" + }, + { + "bbox": [ + 55, + 269, + 296, + 342 + ], + "type": "text", + "content": ". For more detailed discussion and examples of discriminative feature representations see Part B in the Appendix." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 55, + 351, + 291, + 365 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 351, + 291, + 365 + ], + "spans": [ + { + "bbox": [ + 55, + 351, + 291, + 365 + ], + "type": "text", + "content": "4. Classifiers and Adversarial Doppelgängers." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 55, + 371, + 296, + 443 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 371, + 296, + 443 + ], + "spans": [ + { + "bbox": [ + 55, + 371, + 296, + 443 + ], + "type": "text", + "content": "A classifier " + }, + { + "bbox": [ + 55, + 371, + 296, + 443 + ], + "type": "inline_equation", + "content": "R" + }, + { + "bbox": [ + 55, + 371, + 296, + 443 + ], + "type": "text", + "content": " (with " + }, + { + "bbox": [ + 55, + 371, + 296, + 443 + ], + "type": "inline_equation", + "content": "m" + }, + { + "bbox": [ + 55, + 371, + 296, + 443 + ], + "type": "text", + "content": " labels) is called fully populated iff the labeling function " + }, + { + "bbox": [ + 55, + 371, + 296, + 443 + ], + "type": "inline_equation", + "content": "\\text{label}_R: \\mathbf{X} \\to \\{1, \\dots, m\\}" + }, + { + "bbox": [ + 55, + 371, + 296, + 443 + ], + "type": "text", + "content": " is surjective mapping onto the range of labels " + }, + { + "bbox": [ + 55, + 371, + 296, + 443 + ], + "type": "inline_equation", + "content": "\\{1, \\dots, m\\}" + }, + { + "bbox": [ + 55, + 371, + 296, + 443 + ], + "type": "text", + "content": ". For any classifier " + }, + { + "bbox": [ + 55, + 371, + 296, + 443 + ], + "type": "inline_equation", + "content": "R" + }, + { + "bbox": [ + 55, + 371, + 296, + 443 + ], + "type": "text", + "content": ", with " + }, + { + "bbox": [ + 55, + 371, + 296, + 443 + ], + "type": "inline_equation", + "content": "m" + }, + { + "bbox": [ + 55, + 371, + 296, + 443 + ], + "type": "text", + "content": " labels, we will denote by " + }, + { + "bbox": [ + 55, + 371, + 296, + 443 + ], + "type": "inline_equation", + "content": "R_c" + }, + { + "bbox": [ + 55, + 371, + 296, + 443 + ], + "type": "text", + "content": " the level set of the labeling function " + }, + { + "bbox": [ + 55, + 371, + 296, + 443 + ], + "type": "inline_equation", + "content": "\\text{label}_R" + }, + { + "bbox": [ + 55, + 371, + 296, + 443 + ], + "type": "text", + "content": " for each label " + }, + { + "bbox": [ + 55, + 371, + 296, + 443 + ], + "type": "inline_equation", + "content": "c \\in \\{1, \\dots, m\\}" + }, + { + "bbox": [ + 55, + 371, + 296, + 443 + ], + "type": "text", + "content": ".13" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 55, + 450, + 296, + 509 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 450, + 296, + 509 + ], + "spans": [ + { + "bbox": [ + 55, + 450, + 296, + 509 + ], + "type": "text", + "content": "Definition 5. We say that " + }, + { + "bbox": [ + 55, + 450, + 296, + 509 + ], + "type": "inline_equation", + "content": "x" + }, + { + "bbox": [ + 55, + 450, + 296, + 509 + ], + "type": "text", + "content": " is a Doppelgänger adversarial to the classifier " + }, + { + "bbox": [ + 55, + 450, + 296, + 509 + ], + "type": "inline_equation", + "content": "R" + }, + { + "bbox": [ + 55, + 450, + 296, + 509 + ], + "type": "text", + "content": " iff " + }, + { + "bbox": [ + 55, + 450, + 296, + 509 + ], + "type": "inline_equation", + "content": "\\exists y \\in \\mathfrak{d}(x)" + }, + { + "bbox": [ + 55, + 450, + 296, + 509 + ], + "type": "text", + "content": " such that " + }, + { + "bbox": [ + 55, + 450, + 296, + 509 + ], + "type": "inline_equation", + "content": "\\operatorname{label}_R(x) \\neq \\operatorname{label}_R(y)" + }, + { + "bbox": [ + 55, + 450, + 296, + 509 + ], + "type": "text", + "content": " and we will refer to both " + }, + { + "bbox": [ + 55, + 450, + 296, + 509 + ], + "type": "inline_equation", + "content": "x" + }, + { + "bbox": [ + 55, + 450, + 296, + 509 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 55, + 450, + 296, + 509 + ], + "type": "inline_equation", + "content": "y" + }, + { + "bbox": [ + 55, + 450, + 296, + 509 + ], + "type": "text", + "content": " as adversarial Doppelgangers when the classifier " + }, + { + "bbox": [ + 55, + 450, + 296, + 509 + ], + "type": "inline_equation", + "content": "R" + }, + { + "bbox": [ + 55, + 450, + 296, + 509 + ], + "type": "text", + "content": " is clear from the context." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 55, + 510, + 296, + 557 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 510, + 296, + 557 + ], + "spans": [ + { + "bbox": [ + 55, + 510, + 296, + 557 + ], + "type": "text", + "content": "A classifier " + }, + { + "bbox": [ + 55, + 510, + 296, + 557 + ], + "type": "inline_equation", + "content": "R" + }, + { + "bbox": [ + 55, + 510, + 296, + 557 + ], + "type": "text", + "content": " is called (perceptually) regular iff it does not admit adversarial Doppelgangers. If " + }, + { + "bbox": [ + 55, + 510, + 296, + 557 + ], + "type": "inline_equation", + "content": "R" + }, + { + "bbox": [ + 55, + 510, + 296, + 557 + ], + "type": "text", + "content": " is regular, then " + }, + { + "bbox": [ + 55, + 510, + 296, + 557 + ], + "type": "inline_equation", + "content": "\\mathfrak{D}_{\\alpha \\delta} = \\{\\mathfrak{d}(x)\\}_{x \\in \\mathbf{X}}" + }, + { + "bbox": [ + 55, + 510, + 296, + 557 + ], + "type": "text", + "content": " are " + }, + { + "bbox": [ + 55, + 510, + 296, + 557 + ], + "type": "inline_equation", + "content": "R" + }, + { + "bbox": [ + 55, + 510, + 296, + 557 + ], + "type": "text", + "content": " coherent coverings (as defined in [66])." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 55, + 558, + 296, + 616 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 558, + 296, + 616 + ], + "spans": [ + { + "bbox": [ + 55, + 558, + 296, + 616 + ], + "type": "text", + "content": "We say that the classification problem with " + }, + { + "bbox": [ + 55, + 558, + 296, + 616 + ], + "type": "inline_equation", + "content": "m" + }, + { + "bbox": [ + 55, + 558, + 296, + 616 + ], + "type": "text", + "content": "-labels is well defined if there exists a fully populated and perceptually regular classifier with " + }, + { + "bbox": [ + 55, + 558, + 296, + 616 + ], + "type": "inline_equation", + "content": "m" + }, + { + "bbox": [ + 55, + 558, + 296, + 616 + ], + "type": "text", + "content": " labels. Otherwise we say that the classification problem with " + }, + { + "bbox": [ + 55, + 558, + 296, + 616 + ], + "type": "inline_equation", + "content": "m" + }, + { + "bbox": [ + 55, + 558, + 296, + 616 + ], + "type": "text", + "content": "-labels is not well defined." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 313, + 72, + 553, + 154 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 72, + 553, + 154 + ], + "spans": [ + { + "bbox": [ + 313, + 72, + 553, + 154 + ], + "type": "text", + "content": "The labeling function " + }, + { + "bbox": [ + 313, + 72, + 553, + 154 + ], + "type": "inline_equation", + "content": "\\text{label}_R" + }, + { + "bbox": [ + 313, + 72, + 553, + 154 + ], + "type": "text", + "content": " of a regular classifier is continuous with respect to the perceptual topology " + }, + { + "bbox": [ + 313, + 72, + 553, + 154 + ], + "type": "inline_equation", + "content": "\\tau_{\\delta}" + }, + { + "bbox": [ + 313, + 72, + 553, + 154 + ], + "type": "text", + "content": ". Furthermore, if " + }, + { + "bbox": [ + 313, + 72, + 553, + 154 + ], + "type": "inline_equation", + "content": "R" + }, + { + "bbox": [ + 313, + 72, + 553, + 154 + ], + "type": "text", + "content": " is not regular and " + }, + { + "bbox": [ + 313, + 72, + 553, + 154 + ], + "type": "inline_equation", + "content": "x" + }, + { + "bbox": [ + 313, + 72, + 553, + 154 + ], + "type": "text", + "content": " is a point of discontinuity of " + }, + { + "bbox": [ + 313, + 72, + 553, + 154 + ], + "type": "inline_equation", + "content": "\\text{label}_R: (\\mathbf{X}, \\tau_{\\delta}) \\to \\{1, \\dots, m\\}" + }, + { + "bbox": [ + 313, + 72, + 553, + 154 + ], + "type": "text", + "content": ", then " + }, + { + "bbox": [ + 313, + 72, + 553, + 154 + ], + "type": "inline_equation", + "content": "x" + }, + { + "bbox": [ + 313, + 72, + 553, + 154 + ], + "type": "text", + "content": " is an adversarial Doppelgänger. The discontinuity is an indication of the cognitive disruption that occurs when one encounters some AD." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 313, + 156, + 554, + 252 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 156, + 554, + 252 + ], + "spans": [ + { + "bbox": [ + 313, + 156, + 554, + 252 + ], + "type": "text", + "content": "It is well known that in some experiences perceptually unambiguous categories and hence perceptually regular classifiers do not exist. A simple example is provided by the perceptual topology in Example 1. Indeed, in this case " + }, + { + "bbox": [ + 313, + 156, + 554, + 252 + ], + "type": "inline_equation", + "content": "\\mathbf{X} / \\sim_{\\sigma}" + }, + { + "bbox": [ + 313, + 156, + 554, + 252 + ], + "type": "text", + "content": " is a singleton, i.e., every two inputs are metamorphic and hence there is a only one elementary set which equals the whole " + }, + { + "bbox": [ + 313, + 156, + 554, + 252 + ], + "type": "inline_equation", + "content": "\\mathbf{X}" + }, + { + "bbox": [ + 313, + 156, + 554, + 252 + ], + "type": "text", + "content": ". Therefore, every classifier with two or more labels must have adversarial Doppelgängers.[14]" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 313, + 252, + 554, + 324 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 252, + 554, + 324 + ], + "spans": [ + { + "bbox": [ + 313, + 252, + 554, + 324 + ], + "type": "text", + "content": "Clearly, no amount of \"robust training\" will get rid of adversarial examples of a classifier with a surjective labeling function " + }, + { + "bbox": [ + 313, + 252, + 554, + 324 + ], + "type": "inline_equation", + "content": "\\text{label}_R: (\\mathbf{X}, \\tau_{\\delta}) \\to \\{1, \\ldots, m\\}" + }, + { + "bbox": [ + 313, + 252, + 554, + 324 + ], + "type": "text", + "content": " if " + }, + { + "bbox": [ + 313, + 252, + 554, + 324 + ], + "type": "inline_equation", + "content": "\\mathbf{X}" + }, + { + "bbox": [ + 313, + 252, + 554, + 324 + ], + "type": "text", + "content": " cannot be broken into " + }, + { + "bbox": [ + 313, + 252, + 554, + 324 + ], + "type": "inline_equation", + "content": "m" + }, + { + "bbox": [ + 313, + 252, + 554, + 324 + ], + "type": "text", + "content": " perceptually unambiguous categories. In the rest of this section we investigate the non-existence, existence and internal structure of regular classifiers." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 313, + 324, + 554, + 360 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 324, + 554, + 360 + ], + "spans": [ + { + "bbox": [ + 313, + 324, + 554, + 360 + ], + "type": "text", + "content": "In Part D of the Appendix we show a specific example of a well defined classification problem and discuss the actual regular classifier." + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 313, + 361, + 555, + 460 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 361, + 555, + 460 + ], + "spans": [ + { + "bbox": [ + 313, + 361, + 555, + 460 + ], + "type": "text", + "content": "Example 3: If " + }, + { + "bbox": [ + 313, + 361, + 555, + 460 + ], + "type": "inline_equation", + "content": "\\stackrel{\\alpha \\delta}{\\approx}" + }, + { + "bbox": [ + 313, + 361, + 555, + 460 + ], + "type": "text", + "content": " is transitive then for every number of labels " + }, + { + "bbox": [ + 313, + 361, + 555, + 460 + ], + "type": "inline_equation", + "content": "m" + }, + { + "bbox": [ + 313, + 361, + 555, + 460 + ], + "type": "text", + "content": " smaller than the number of equivalence classes card " + }, + { + "bbox": [ + 313, + 361, + 555, + 460 + ], + "type": "inline_equation", + "content": "\\left(\\mathbf{X} / \\stackrel{\\alpha \\delta}{\\approx}\\right)" + }, + { + "bbox": [ + 313, + 361, + 555, + 460 + ], + "type": "text", + "content": " there exists a fully populated regular classifier with " + }, + { + "bbox": [ + 313, + 361, + 555, + 460 + ], + "type": "inline_equation", + "content": "m" + }, + { + "bbox": [ + 313, + 361, + 555, + 460 + ], + "type": "text", + "content": " labels, but if the number of labels " + }, + { + "bbox": [ + 313, + 361, + 555, + 460 + ], + "type": "inline_equation", + "content": "m" + }, + { + "bbox": [ + 313, + 361, + 555, + 460 + ], + "type": "text", + "content": " is bigger than card " + }, + { + "bbox": [ + 313, + 361, + 555, + 460 + ], + "type": "inline_equation", + "content": "\\left(\\mathbf{X} / \\stackrel{\\alpha \\delta}{\\approx}\\right)" + }, + { + "bbox": [ + 313, + 361, + 555, + 460 + ], + "type": "text", + "content": " then every fully populated classifier with " + }, + { + "bbox": [ + 313, + 361, + 555, + 460 + ], + "type": "inline_equation", + "content": "m" + }, + { + "bbox": [ + 313, + 361, + 555, + 460 + ], + "type": "text", + "content": " labels must admit adversarial Doppelgangers." + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 313, + 471, + 554, + 495 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 471, + 554, + 495 + ], + "spans": [ + { + "bbox": [ + 313, + 471, + 554, + 495 + ], + "type": "text", + "content": "Example 3 indicates that if the transitive closure " + }, + { + "bbox": [ + 313, + 471, + 554, + 495 + ], + "type": "inline_equation", + "content": "\\sim_{\\sigma}" + }, + { + "bbox": [ + 313, + 471, + 554, + 495 + ], + "type": "text", + "content": " is trivial, i.e., " + }, + { + "bbox": [ + 313, + 471, + 554, + 495 + ], + "type": "inline_equation", + "content": "\\sim_{\\sigma} = \\mathbf{X} \\times \\mathbf{X}" + }, + { + "bbox": [ + 313, + 471, + 554, + 495 + ], + "type": "text", + "content": ", then no label is safe. Namely:" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 313, + 503, + 554, + 605 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 503, + 554, + 605 + ], + "spans": [ + { + "bbox": [ + 313, + 503, + 554, + 605 + ], + "type": "text", + "content": "Observation 1. If the transitive closure " + }, + { + "bbox": [ + 313, + 503, + 554, + 605 + ], + "type": "inline_equation", + "content": "\\sim_{\\sigma}" + }, + { + "bbox": [ + 313, + 503, + 554, + 605 + ], + "type": "text", + "content": " of the indiscriminability relation " + }, + { + "bbox": [ + 313, + 503, + 554, + 605 + ], + "type": "inline_equation", + "content": "\\stackrel{\\alpha \\delta}{\\approx}" + }, + { + "bbox": [ + 313, + 503, + 554, + 605 + ], + "type": "text", + "content": " is trivial, then every fully populated classifier with two or more classes admits adversarial Doppelgängers. In particular, let " + }, + { + "bbox": [ + 313, + 503, + 554, + 605 + ], + "type": "inline_equation", + "content": "R" + }, + { + "bbox": [ + 313, + 503, + 554, + 605 + ], + "type": "text", + "content": " be a fully populated classifier with a surjective labeling function " + }, + { + "bbox": [ + 313, + 503, + 554, + 605 + ], + "type": "inline_equation", + "content": "\\text{label}_R: X \\to \\{1, 2, \\ldots, m\\}" + }, + { + "bbox": [ + 313, + 503, + 554, + 605 + ], + "type": "text", + "content": ", then for every label " + }, + { + "bbox": [ + 313, + 503, + 554, + 605 + ], + "type": "inline_equation", + "content": "c" + }, + { + "bbox": [ + 313, + 503, + 554, + 605 + ], + "type": "text", + "content": " there exist adversarial Doppelgängers " + }, + { + "bbox": [ + 313, + 503, + 554, + 605 + ], + "type": "inline_equation", + "content": "x(c) \\in \\mathbf{X}" + }, + { + "bbox": [ + 313, + 503, + 554, + 605 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 313, + 503, + 554, + 605 + ], + "type": "inline_equation", + "content": "x^*(c) \\in \\mathfrak{d}(x(c))" + }, + { + "bbox": [ + 313, + 503, + 554, + 605 + ], + "type": "text", + "content": " such that " + }, + { + "bbox": [ + 313, + 503, + 554, + 605 + ], + "type": "inline_equation", + "content": "c = \\text{label}_R(x(c))" + }, + { + "bbox": [ + 313, + 503, + 554, + 605 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 313, + 503, + 554, + 605 + ], + "type": "inline_equation", + "content": "\\text{label}_R(x^*(c)) \\neq \\text{label}_R(x(c))" + }, + { + "bbox": [ + 313, + 503, + 554, + 605 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 313, + 612, + 553, + 660 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 612, + 553, + 660 + ], + "spans": [ + { + "bbox": [ + 313, + 612, + 553, + 660 + ], + "type": "text", + "content": "The proof follows from the fact that every finite chain of Doppelgängers connecting points that are labeled differently by a classifier must contain a pair of adversarial Doppelgängers. See Lemma 2 in Appendix Section C." + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 313, + 661, + 554, + 685 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 661, + 554, + 685 + ], + "spans": [ + { + "bbox": [ + 313, + 661, + 554, + 685 + ], + "type": "text", + "content": "A straight-forward argument shows that if " + }, + { + "bbox": [ + 313, + 661, + 554, + 685 + ], + "type": "inline_equation", + "content": "R" + }, + { + "bbox": [ + 313, + 661, + 554, + 685 + ], + "type": "text", + "content": " is a perceptually regular fully populated classifier, then " + }, + { + "bbox": [ + 313, + 661, + 554, + 685 + ], + "type": "inline_equation", + "content": "x \\in R_i" + }, + { + "bbox": [ + 313, + 661, + 554, + 685 + ], + "type": "text", + "content": " iff" + } + ] + } + ], + "index": 18 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 313, + 693, + 553, + 712 + ], + "type": "page_footnote", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 693, + 553, + 712 + ], + "spans": [ + { + "bbox": [ + 313, + 693, + 553, + 712 + ], + "type": "text", + "content": "14See Lemma 3 and the short argument that follows it in Appendix, Section C." + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 65, + 622, + 219, + 632 + ], + "type": "page_footnote", + "angle": 0, + "lines": [ + { + "bbox": [ + 65, + 622, + 219, + 632 + ], + "spans": [ + { + "bbox": [ + 65, + 622, + 219, + 632 + ], + "type": "text", + "content": "10See Observation 7, in Part B of the Appendix." + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 56, + 632, + 295, + 673 + ], + "type": "page_footnote", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 632, + 295, + 673 + ], + "spans": [ + { + "bbox": [ + 56, + 632, + 295, + 673 + ], + "type": "text", + "content": "11A feature " + }, + { + "bbox": [ + 56, + 632, + 295, + 673 + ], + "type": "inline_equation", + "content": "\\xi \\in \\Phi" + }, + { + "bbox": [ + 56, + 632, + 295, + 673 + ], + "type": "text", + "content": " is called attributed if " + }, + { + "bbox": [ + 56, + 632, + 295, + 673 + ], + "type": "inline_equation", + "content": "\\xi \\in \\Phi_x" + }, + { + "bbox": [ + 56, + 632, + 295, + 673 + ], + "type": "text", + "content": " for some input " + }, + { + "bbox": [ + 56, + 632, + 295, + 673 + ], + "type": "inline_equation", + "content": "x \\in \\mathbf{X}" + }, + { + "bbox": [ + 56, + 632, + 295, + 673 + ], + "type": "text", + "content": ". It is plausible that " + }, + { + "bbox": [ + 56, + 632, + 295, + 673 + ], + "type": "inline_equation", + "content": "\\Phi = \\bigcup_{x \\in \\mathbf{X}} \\Phi_x" + }, + { + "bbox": [ + 56, + 632, + 295, + 673 + ], + "type": "text", + "content": ", and so all features are attributed. However, many models do not preclude the existence of spurious latent traits." + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 56, + 673, + 295, + 703 + ], + "type": "page_footnote", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 673, + 295, + 703 + ], + "spans": [ + { + "bbox": [ + 56, + 673, + 295, + 703 + ], + "type": "text", + "content": "12 The semantic cluster of inputs sharing the feature " + }, + { + "bbox": [ + 56, + 673, + 295, + 703 + ], + "type": "inline_equation", + "content": "\\xi \\in \\Phi" + }, + { + "bbox": [ + 56, + 673, + 295, + 703 + ], + "type": "text", + "content": " is defined for any feature representation. A feature " + }, + { + "bbox": [ + 56, + 673, + 295, + 703 + ], + "type": "inline_equation", + "content": "\\xi" + }, + { + "bbox": [ + 56, + 673, + 295, + 703 + ], + "type": "text", + "content": " is attributed in a given context, iff " + }, + { + "bbox": [ + 56, + 673, + 295, + 703 + ], + "type": "inline_equation", + "content": "\\operatorname{cl}(\\xi) \\neq \\emptyset" + }, + { + "bbox": [ + 56, + 673, + 295, + 703 + ], + "type": "text", + "content": "; a feature is a hypothetical feature, when " + }, + { + "bbox": [ + 56, + 673, + 295, + 703 + ], + "type": "inline_equation", + "content": "\\operatorname{cl}(\\xi) = \\emptyset" + }, + { + "bbox": [ + 56, + 673, + 295, + 703 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 64, + 703, + 279, + 713 + ], + "type": "page_footnote", + "angle": 0, + "lines": [ + { + "bbox": [ + 64, + 703, + 279, + 713 + ], + "spans": [ + { + "bbox": [ + 64, + 703, + 279, + 713 + ], + "type": "inline_equation", + "content": "^{13}R" + }, + { + "bbox": [ + 64, + 703, + 279, + 713 + ], + "type": "text", + "content": " is fully populated iff " + }, + { + "bbox": [ + 64, + 703, + 279, + 713 + ], + "type": "inline_equation", + "content": "R_{c}\\neq \\emptyset" + }, + { + "bbox": [ + 64, + 703, + 279, + 713 + ], + "type": "text", + "content": " for every label " + }, + { + "bbox": [ + 64, + 703, + 279, + 713 + ], + "type": "inline_equation", + "content": "c\\in \\{1,\\ldots ,m\\}" + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "text", + "content": "10247" + } + ] + } + ], + "index": 25 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 3 + }, + { + "para_blocks": [ + { + "bbox": [ + 55, + 72, + 295, + 97 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 72, + 295, + 97 + ], + "spans": [ + { + "bbox": [ + 55, + 72, + 295, + 97 + ], + "type": "inline_equation", + "content": "[x]_{\\sim_{\\sigma}} \\subset R_i" + }, + { + "bbox": [ + 55, + 72, + 295, + 97 + ], + "type": "text", + "content": ", i.e., each level set is a disjoint union of elementary sets " + }, + { + "bbox": [ + 55, + 72, + 295, + 97 + ], + "type": "inline_equation", + "content": "[x]_{\\sim_{\\sigma}}" + }, + { + "bbox": [ + 55, + 72, + 295, + 97 + ], + "type": "text", + "content": ". The pigeonhole principle yields:" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 55, + 103, + 296, + 175 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 103, + 296, + 175 + ], + "spans": [ + { + "bbox": [ + 55, + 103, + 296, + 175 + ], + "type": "text", + "content": "Observation 2. If the number of equivalence classes card " + }, + { + "bbox": [ + 55, + 103, + 296, + 175 + ], + "type": "inline_equation", + "content": "(\\mathbf{X} / \\sim_{\\sigma}) \\geq 2" + }, + { + "bbox": [ + 55, + 103, + 296, + 175 + ], + "type": "text", + "content": ", then for every natural number " + }, + { + "bbox": [ + 55, + 103, + 296, + 175 + ], + "type": "inline_equation", + "content": "2 \\leq m \\leq" + }, + { + "bbox": [ + 55, + 103, + 296, + 175 + ], + "type": "text", + "content": " card " + }, + { + "bbox": [ + 55, + 103, + 296, + 175 + ], + "type": "inline_equation", + "content": "(\\mathbf{X} / \\sim_{\\sigma})" + }, + { + "bbox": [ + 55, + 103, + 296, + 175 + ], + "type": "text", + "content": " there exists a perceptually regular fully populated classifier with " + }, + { + "bbox": [ + 55, + 103, + 296, + 175 + ], + "type": "inline_equation", + "content": "m" + }, + { + "bbox": [ + 55, + 103, + 296, + 175 + ], + "type": "text", + "content": " labels. However, if " + }, + { + "bbox": [ + 55, + 103, + 296, + 175 + ], + "type": "inline_equation", + "content": "m >" + }, + { + "bbox": [ + 55, + 103, + 296, + 175 + ], + "type": "text", + "content": " card " + }, + { + "bbox": [ + 55, + 103, + 296, + 175 + ], + "type": "inline_equation", + "content": "(\\mathbf{X} / \\sim_{\\sigma})" + }, + { + "bbox": [ + 55, + 103, + 296, + 175 + ], + "type": "text", + "content": ", then every fully populated classifier with " + }, + { + "bbox": [ + 55, + 103, + 296, + 175 + ], + "type": "inline_equation", + "content": "m" + }, + { + "bbox": [ + 55, + 103, + 296, + 175 + ], + "type": "text", + "content": " labels must have adversarial Doppelgangers." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 54, + 175, + 296, + 223 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 54, + 175, + 296, + 223 + ], + "spans": [ + { + "bbox": [ + 54, + 175, + 296, + 223 + ], + "type": "text", + "content": "In particular, if " + }, + { + "bbox": [ + 54, + 175, + 296, + 223 + ], + "type": "inline_equation", + "content": "p = \\operatorname{card}(\\mathbf{X} / \\sim_{\\sigma}) \\geq 2" + }, + { + "bbox": [ + 54, + 175, + 296, + 223 + ], + "type": "text", + "content": " is finite, then for every natural number " + }, + { + "bbox": [ + 54, + 175, + 296, + 223 + ], + "type": "inline_equation", + "content": "m \\leq p" + }, + { + "bbox": [ + 54, + 175, + 296, + 223 + ], + "type": "text", + "content": " there are exactly " + }, + { + "bbox": [ + 54, + 175, + 296, + 223 + ], + "type": "inline_equation", + "content": "S(p, m)" + }, + { + "bbox": [ + 54, + 175, + 296, + 223 + ], + "type": "text", + "content": " regular fully populated classifiers with " + }, + { + "bbox": [ + 54, + 175, + 296, + 223 + ], + "type": "inline_equation", + "content": "m" + }, + { + "bbox": [ + 54, + 175, + 296, + 223 + ], + "type": "text", + "content": " classes. Here " + }, + { + "bbox": [ + 54, + 175, + 296, + 223 + ], + "type": "inline_equation", + "content": "S(p, m)" + }, + { + "bbox": [ + 54, + 175, + 296, + 223 + ], + "type": "text", + "content": " is the Sterling number of the second kind." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 55, + 230, + 296, + 315 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 230, + 296, + 315 + ], + "spans": [ + { + "bbox": [ + 55, + 230, + 296, + 315 + ], + "type": "text", + "content": "Observation 2 shows that the problem of finding a fully populated perceptually unambiguous classifier " + }, + { + "bbox": [ + 55, + 230, + 296, + 315 + ], + "type": "inline_equation", + "content": "R" + }, + { + "bbox": [ + 55, + 230, + 296, + 315 + ], + "type": "text", + "content": " with precisely " + }, + { + "bbox": [ + 55, + 230, + 296, + 315 + ], + "type": "inline_equation", + "content": "m" + }, + { + "bbox": [ + 55, + 230, + 296, + 315 + ], + "type": "text", + "content": " labels is not well defined if card " + }, + { + "bbox": [ + 55, + 230, + 296, + 315 + ], + "type": "inline_equation", + "content": "(\\mathbf{X} / \\sim_{\\sigma}) < m" + }, + { + "bbox": [ + 55, + 230, + 296, + 315 + ], + "type": "text", + "content": " and vice versa that the same problem is well defined for every " + }, + { + "bbox": [ + 55, + 230, + 296, + 315 + ], + "type": "inline_equation", + "content": "m" + }, + { + "bbox": [ + 55, + 230, + 296, + 315 + ], + "type": "text", + "content": " such that card " + }, + { + "bbox": [ + 55, + 230, + 296, + 315 + ], + "type": "inline_equation", + "content": "(\\mathbf{X} / \\sim_{\\sigma}) \\geq m" + }, + { + "bbox": [ + 55, + 230, + 296, + 315 + ], + "type": "text", + "content": ". In the former case solutions do not exist while in the later case solutions exist and each class segment " + }, + { + "bbox": [ + 55, + 230, + 296, + 315 + ], + "type": "inline_equation", + "content": "R_{i}" + }, + { + "bbox": [ + 55, + 230, + 296, + 315 + ], + "type": "text", + "content": " is a union of equivalence classes " + }, + { + "bbox": [ + 55, + 230, + 296, + 315 + ], + "type": "inline_equation", + "content": "[x]_{\\sim_{\\sigma}}" + }, + { + "bbox": [ + 55, + 230, + 296, + 315 + ], + "type": "text", + "content": "," + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 139, + 323, + 295, + 349 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 139, + 323, + 295, + 349 + ], + "spans": [ + { + "bbox": [ + 139, + 323, + 295, + 349 + ], + "type": "interline_equation", + "content": "R _ {i} = \\bigcup_ {x \\in R _ {i}} [ x ] _ {\\sim_ {\\sigma}}. \\tag {5}", + "image_path": "3702f70988402238084ef0595380f7cdc4c313178324efc26061e6a594115d44.jpg" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 55, + 355, + 296, + 475 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 355, + 296, + 475 + ], + "spans": [ + { + "bbox": [ + 55, + 355, + 296, + 475 + ], + "type": "text", + "content": "The existence and properties of discriminative feature representations provide insight whether a classification problem is well defined. In particular, if there exists a discriminative feature representation and the set of attributed features is finite, then every classification problem, whose number of labels exceeds the number of attributed features, is not well defined. See Observation 9 in Appendix Part B.2. Further discussion of the structure of class segments of a regular classifier including the class/category core and fringe are discussed in Section 6." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 55, + 486, + 291, + 499 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 486, + 291, + 499 + ], + "spans": [ + { + "bbox": [ + 55, + 486, + 291, + 499 + ], + "type": "text", + "content": "5. Accuracy and Adversarial Doppelgängers." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 55, + 506, + 296, + 625 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 506, + 296, + 625 + ], + "spans": [ + { + "bbox": [ + 55, + 506, + 296, + 625 + ], + "type": "text", + "content": "The accuracy-adversarial robustness trade off observed and discussed in the literature involves various measures of accuracy [48, 50, 71, 78, 90]. We will discuss classification accuracy. We will show that there is a strong relationship between classifier accuracy and vulnerability to adversarial Doppelgängers. In particular, we will identify (perceptual) scenarios in which low accuracy classifiers are critically vulnerable to adversarial Doppelgänger attacks but on the other hand all high accuracy classifiers can be fooled only by Doppelgängers." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 55, + 632, + 190, + 645 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 632, + 190, + 645 + ], + "spans": [ + { + "bbox": [ + 55, + 632, + 190, + 645 + ], + "type": "text", + "content": "5.1. The Probabilistic Setup." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 55, + 649, + 296, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 649, + 296, + 715 + ], + "spans": [ + { + "bbox": [ + 55, + 649, + 296, + 715 + ], + "type": "text", + "content": "We will assume that " + }, + { + "bbox": [ + 55, + 649, + 296, + 715 + ], + "type": "inline_equation", + "content": "(\\mathbf{X},\\mathcal{F},\\mu)" + }, + { + "bbox": [ + 55, + 649, + 296, + 715 + ], + "type": "text", + "content": " is a probability measure space equipped with perceptual topology " + }, + { + "bbox": [ + 55, + 649, + 296, + 715 + ], + "type": "inline_equation", + "content": "\\tau_{\\delta}" + }, + { + "bbox": [ + 55, + 649, + 296, + 715 + ], + "type": "text", + "content": " generated by an indsicriminability relation " + }, + { + "bbox": [ + 55, + 649, + 296, + 715 + ], + "type": "inline_equation", + "content": "\\stackrel{\\alpha \\delta}{\\approx}" + }, + { + "bbox": [ + 55, + 649, + 296, + 715 + ], + "type": "text", + "content": " such that for every " + }, + { + "bbox": [ + 55, + 649, + 296, + 715 + ], + "type": "inline_equation", + "content": "x \\in \\mathbf{X}" + }, + { + "bbox": [ + 55, + 649, + 296, + 715 + ], + "type": "text", + "content": " the set of Doppelgangers " + }, + { + "bbox": [ + 55, + 649, + 296, + 715 + ], + "type": "inline_equation", + "content": "\\mathfrak{d}(x)" + }, + { + "bbox": [ + 55, + 649, + 296, + 715 + ], + "type": "text", + "content": " and the equivalence class " + }, + { + "bbox": [ + 55, + 649, + 296, + 715 + ], + "type": "inline_equation", + "content": "[x]_{\\sim_{\\sigma}}" + }, + { + "bbox": [ + 55, + 649, + 296, + 715 + ], + "type": "text", + "content": " are events, " + }, + { + "bbox": [ + 55, + 649, + 296, + 715 + ], + "type": "inline_equation", + "content": "(\\mathfrak{d}(x) \\in \\mathcal{F}" + }, + { + "bbox": [ + 55, + 649, + 296, + 715 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 55, + 649, + 296, + 715 + ], + "type": "inline_equation", + "content": "[x]_{\\sim_{\\sigma}} \\in \\mathcal{F}, \\forall x \\in \\mathbf{X})" + }, + { + "bbox": [ + 55, + 649, + 296, + 715 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 313, + 72, + 555, + 179 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 72, + 555, + 179 + ], + "spans": [ + { + "bbox": [ + 313, + 72, + 555, + 179 + ], + "type": "text", + "content": "In the rest of this section we will assume that the classification problem with " + }, + { + "bbox": [ + 313, + 72, + 555, + 179 + ], + "type": "inline_equation", + "content": "m \\geq 2" + }, + { + "bbox": [ + 313, + 72, + 555, + 179 + ], + "type": "text", + "content": " labels is well defined and let " + }, + { + "bbox": [ + 313, + 72, + 555, + 179 + ], + "type": "inline_equation", + "content": "\\Omega" + }, + { + "bbox": [ + 313, + 72, + 555, + 179 + ], + "type": "text", + "content": " be a perceptually regular classifier; we will reserve the notation " + }, + { + "bbox": [ + 313, + 72, + 555, + 179 + ], + "type": "inline_equation", + "content": "R" + }, + { + "bbox": [ + 313, + 72, + 555, + 179 + ], + "type": "text", + "content": " to denote any classifier (" + }, + { + "bbox": [ + 313, + 72, + 555, + 179 + ], + "type": "inline_equation", + "content": "R" + }, + { + "bbox": [ + 313, + 72, + 555, + 179 + ], + "type": "text", + "content": " may or may not be perceptually regular) such that " + }, + { + "bbox": [ + 313, + 72, + 555, + 179 + ], + "type": "inline_equation", + "content": "R_{i} \\cap \\Omega_{i}" + }, + { + "bbox": [ + 313, + 72, + 555, + 179 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 313, + 72, + 555, + 179 + ], + "type": "inline_equation", + "content": "i = 1, \\ldots, m" + }, + { + "bbox": [ + 313, + 72, + 555, + 179 + ], + "type": "text", + "content": " are the true positives (of class " + }, + { + "bbox": [ + 313, + 72, + 555, + 179 + ], + "type": "inline_equation", + "content": "i" + }, + { + "bbox": [ + 313, + 72, + 555, + 179 + ], + "type": "text", + "content": "). To define accuracy we will focus only on regular models and classifiers s.t., " + }, + { + "bbox": [ + 313, + 72, + 555, + 179 + ], + "type": "inline_equation", + "content": "\\Omega_{i} \\in \\mathcal{F}" + }, + { + "bbox": [ + 313, + 72, + 555, + 179 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 313, + 72, + 555, + 179 + ], + "type": "inline_equation", + "content": "R_{i} \\in \\mathcal{F}" + }, + { + "bbox": [ + 313, + 72, + 555, + 179 + ], + "type": "text", + "content": ", for all " + }, + { + "bbox": [ + 313, + 72, + 555, + 179 + ], + "type": "inline_equation", + "content": "i = 1, \\ldots, m" + }, + { + "bbox": [ + 313, + 72, + 555, + 179 + ], + "type": "text", + "content": ". The accuracy of the classifier " + }, + { + "bbox": [ + 313, + 72, + 555, + 179 + ], + "type": "inline_equation", + "content": "R" + }, + { + "bbox": [ + 313, + 72, + 555, + 179 + ], + "type": "text", + "content": " defined as" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 320, + 190, + 555, + 205 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 320, + 190, + 555, + 205 + ], + "spans": [ + { + "bbox": [ + 320, + 190, + 555, + 205 + ], + "type": "interline_equation", + "content": "\\operatorname {a c c u r a c y} _ {\\Omega} (R) = \\mu \\left(R _ {1} \\cap \\Omega_ {1}\\right) + \\dots + \\mu \\left(R _ {m} \\cap \\Omega_ {m}\\right). \\tag {6}", + "image_path": "d4667cb6c3d13cf6fe6729660fa4e01defd0dec98db6ded28553bb8e2c13393d.jpg" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 313, + 213, + 555, + 238 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 213, + 555, + 238 + ], + "spans": [ + { + "bbox": [ + 313, + 213, + 555, + 238 + ], + "type": "text", + "content": "Furthermore let us assume that " + }, + { + "bbox": [ + 313, + 213, + 555, + 238 + ], + "type": "inline_equation", + "content": "\\mu (\\Omega_i) > 0" + }, + { + "bbox": [ + 313, + 213, + 555, + 238 + ], + "type": "text", + "content": " for every " + }, + { + "bbox": [ + 313, + 213, + 555, + 238 + ], + "type": "inline_equation", + "content": "i = 1,\\ldots ,m" + }, + { + "bbox": [ + 313, + 213, + 555, + 238 + ], + "type": "text", + "content": " and thus we can define recall rates" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 364, + 247, + 554, + 274 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 364, + 247, + 554, + 274 + ], + "spans": [ + { + "bbox": [ + 364, + 247, + 554, + 274 + ], + "type": "interline_equation", + "content": "\\rho_ {i} = \\frac {\\mu \\left(R _ {i} \\cap \\Omega_ {i}\\right)}{\\mu \\left(\\Omega_ {i}\\right)}, \\quad i = 1, \\dots , m \\tag {7}", + "image_path": "0215eb3c34607a22e381d98e8c702de5191844875a9ddc83bc9f9ef19591bd7e.jpg" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 313, + 283, + 554, + 306 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 283, + 554, + 306 + ], + "spans": [ + { + "bbox": [ + 313, + 283, + 554, + 306 + ], + "type": "text", + "content": "Bounds on the recall rates imply bounds on the accuracy. Namely if" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 375, + 308, + 492, + 322 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 375, + 308, + 492, + 322 + ], + "spans": [ + { + "bbox": [ + 375, + 308, + 492, + 322 + ], + "type": "interline_equation", + "content": "\\underline {{\\rho}} \\leq \\rho_ {i} \\leq \\bar {\\rho}, \\quad i = 1, \\dots , m", + "image_path": "a4fe55830ec39912f5ef3ce2328c8e6a0845e5a5368e41a520b40ac7a0955b47.jpg" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 313, + 327, + 508, + 340 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 327, + 508, + 340 + ], + "spans": [ + { + "bbox": [ + 313, + 327, + 508, + 340 + ], + "type": "text", + "content": "then since " + }, + { + "bbox": [ + 313, + 327, + 508, + 340 + ], + "type": "inline_equation", + "content": "\\mu" + }, + { + "bbox": [ + 313, + 327, + 508, + 340 + ], + "type": "text", + "content": " is a probability measure on " + }, + { + "bbox": [ + 313, + 327, + 508, + 340 + ], + "type": "inline_equation", + "content": "X" + }, + { + "bbox": [ + 313, + 327, + 508, + 340 + ], + "type": "text", + "content": ", and" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 352, + 350, + 555, + 364 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 352, + 350, + 555, + 364 + ], + "spans": [ + { + "bbox": [ + 352, + 350, + 555, + 364 + ], + "type": "interline_equation", + "content": "\\mu \\left(R _ {i} \\cap \\Omega_ {i}\\right) = \\rho_ {i} \\mu \\left(\\Omega_ {i}\\right), \\quad i = 1, \\dots , m \\tag {8}", + "image_path": "7d4fce5a0ecb5aa98905d417c762871ab972209ca328a732fff3c77685f2167d.jpg" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 314, + 375, + 345, + 385 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 314, + 375, + 345, + 385 + ], + "spans": [ + { + "bbox": [ + 314, + 375, + 345, + 385 + ], + "type": "text", + "content": "we get" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 353, + 396, + 515, + 427 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 353, + 396, + 515, + 427 + ], + "spans": [ + { + "bbox": [ + 353, + 396, + 515, + 427 + ], + "type": "interline_equation", + "content": "\\underline {{\\rho}} \\leq \\operatorname {a c c u r a c y} _ {\\Omega} (R) = \\sum_ {i = 1} ^ {m} \\rho_ {i} \\mu (\\Omega_ {i}) \\leq \\bar {\\rho}.", + "image_path": "80b5a1006352e7f12d81244d05f5bd8019c9855c8cfa1f94633c6a7114f94242.jpg" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 313, + 437, + 454, + 449 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 437, + 454, + 449 + ], + "spans": [ + { + "bbox": [ + 313, + 437, + 454, + 449 + ], + "type": "text", + "content": "5.2. Are Trade-Offs Possible?" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 313, + 455, + 554, + 480 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 455, + 554, + 480 + ], + "spans": [ + { + "bbox": [ + 313, + 455, + 554, + 480 + ], + "type": "text", + "content": "Let " + }, + { + "bbox": [ + 313, + 455, + 554, + 480 + ], + "type": "inline_equation", + "content": "i(x) \\in \\{1, \\ldots, m\\}" + }, + { + "bbox": [ + 313, + 455, + 554, + 480 + ], + "type": "text", + "content": " be the object class label of " + }, + { + "bbox": [ + 313, + 455, + 554, + 480 + ], + "type": "inline_equation", + "content": "x \\in \\mathbf{X}" + }, + { + "bbox": [ + 313, + 455, + 554, + 480 + ], + "type": "text", + "content": " i.e., " + }, + { + "bbox": [ + 313, + 455, + 554, + 480 + ], + "type": "inline_equation", + "content": "x \\in \\Omega_{i(x)}" + }, + { + "bbox": [ + 313, + 455, + 554, + 480 + ], + "type": "text", + "content": " and" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 375, + 490, + 554, + 522 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 375, + 490, + 554, + 522 + ], + "spans": [ + { + "bbox": [ + 375, + 490, + 554, + 522 + ], + "type": "interline_equation", + "content": "\\bar {k} (\\Omega) = \\sup _ {x \\in \\mathbf {X}} \\left(\\frac {\\mu \\left(\\Omega_ {i (x)}\\right)}{\\mu (\\mathfrak {d} (x))}\\right). \\tag {9}", + "image_path": "3c0c980f92f3259c4d536870f153511514591e26466cd3c2069df659ef9baeb6.jpg" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 313, + 533, + 555, + 570 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 533, + 555, + 570 + ], + "spans": [ + { + "bbox": [ + 313, + 533, + 555, + 570 + ], + "type": "text", + "content": "Every classifier whose recall rates do not exceed " + }, + { + "bbox": [ + 313, + 533, + 555, + 570 + ], + "type": "inline_equation", + "content": "1 / \\bar{k} (\\Omega)" + }, + { + "bbox": [ + 313, + 533, + 555, + 570 + ], + "type": "text", + "content": " is totally unsafe in the sense that every correctly classified input admits adversarial Doppelgangers. Specifically:" + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 313, + 578, + 555, + 643 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 578, + 555, + 643 + ], + "spans": [ + { + "bbox": [ + 313, + 578, + 555, + 643 + ], + "type": "text", + "content": "Observation 3. Suppose that the sets of Doppelgangers are not negligible and " + }, + { + "bbox": [ + 313, + 578, + 555, + 643 + ], + "type": "inline_equation", + "content": "\\inf_{x\\in \\mathbf{X}}\\mu (\\mathfrak{d}(x)) > 0" + }, + { + "bbox": [ + 313, + 578, + 555, + 643 + ], + "type": "text", + "content": " , and let " + }, + { + "bbox": [ + 313, + 578, + 555, + 643 + ], + "type": "inline_equation", + "content": "\\Omega = \\{\\Omega_1,\\dots ,\\Omega_m\\}" + }, + { + "bbox": [ + 313, + 578, + 555, + 643 + ], + "type": "text", + "content": " be a regular world model and let " + }, + { + "bbox": [ + 313, + 578, + 555, + 643 + ], + "type": "inline_equation", + "content": "R = \\{R_{1},\\ldots ,R_{m}\\}" + }, + { + "bbox": [ + 313, + 578, + 555, + 643 + ], + "type": "text", + "content": " be a classifier whose recall rates are strictly smaller than " + }, + { + "bbox": [ + 313, + 578, + 555, + 643 + ], + "type": "inline_equation", + "content": "1 / \\bar{k} (\\Omega)" + }, + { + "bbox": [ + 313, + 578, + 555, + 643 + ], + "type": "text", + "content": " and so" + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 367, + 653, + 553, + 681 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 367, + 653, + 553, + 681 + ], + "spans": [ + { + "bbox": [ + 367, + 653, + 553, + 681 + ], + "type": "interline_equation", + "content": "\\frac {\\mu \\left(R _ {i (x)} \\cap \\mathfrak {d} (x)\\right)}{\\mu (\\mathfrak {d} (x))} \\leq \\bar {\\rho} \\bar {k} (\\Omega) < 1. \\tag {10}", + "image_path": "a692017bd6d7816b0a80c9c671d440d862c2ddeff9221aa5b51311f496ef2ccb.jpg" + } + ] + } + ], + "index": 25 + }, + { + "bbox": [ + 313, + 689, + 554, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 689, + 554, + 714 + ], + "spans": [ + { + "bbox": [ + 313, + 689, + 554, + 714 + ], + "type": "text", + "content": "Thus every correctly classified input " + }, + { + "bbox": [ + 313, + 689, + 554, + 714 + ], + "type": "inline_equation", + "content": "x" + }, + { + "bbox": [ + 313, + 689, + 554, + 714 + ], + "type": "text", + "content": " has adversarial Doppelgängers." + } + ] + } + ], + "index": 26 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "text", + "content": "10248" + } + ] + } + ], + "index": 27 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 4 + }, + { + "para_blocks": [ + { + "bbox": [ + 55, + 72, + 295, + 133 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 72, + 295, + 133 + ], + "spans": [ + { + "bbox": [ + 55, + 72, + 295, + 133 + ], + "type": "text", + "content": "Observation 3 shows that sacrificing accuracy may lead to increasing the probability of encountering adversarial Doppelgangers and in fact that there is no trade off for accuracies that are sufficiently low provided that all sets of Doppelgangers have positive measure." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 55, + 133, + 295, + 169 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 133, + 295, + 169 + ], + "spans": [ + { + "bbox": [ + 55, + 133, + 295, + 169 + ], + "type": "text", + "content": "The lack of an opportunity for a trade-off is even more striking when one tries to improve high recall rate (and hence high accuracy) classifiers." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 55, + 178, + 295, + 220 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 178, + 295, + 220 + ], + "spans": [ + { + "bbox": [ + 55, + 178, + 295, + 220 + ], + "type": "text", + "content": "Observation 4. " + }, + { + "bbox": [ + 55, + 178, + 295, + 220 + ], + "type": "inline_equation", + "content": "^{15}" + }, + { + "bbox": [ + 55, + 178, + 295, + 220 + ], + "type": "text", + "content": " If " + }, + { + "bbox": [ + 55, + 178, + 295, + 220 + ], + "type": "inline_equation", + "content": "\\inf_{x\\in \\mathbf{X}}\\mu (\\mathfrak{d}(x)) > 0" + }, + { + "bbox": [ + 55, + 178, + 295, + 220 + ], + "type": "text", + "content": " , and let " + }, + { + "bbox": [ + 55, + 178, + 295, + 220 + ], + "type": "inline_equation", + "content": "R =" + }, + { + "bbox": [ + 55, + 178, + 295, + 220 + ], + "type": "inline_equation", + "content": "\\{R_1,\\ldots ,R_m\\}" + }, + { + "bbox": [ + 55, + 178, + 295, + 220 + ], + "type": "text", + "content": " be a classifier whose recall rates are sufficiently high so that " + }, + { + "bbox": [ + 55, + 178, + 295, + 220 + ], + "type": "inline_equation", + "content": "\\underline{\\rho} >1 - 1 / \\bar{k} (\\Omega)" + }, + { + "bbox": [ + 55, + 178, + 295, + 220 + ], + "type": "text", + "content": " . i.e.," + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 137, + 230, + 295, + 247 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 137, + 230, + 295, + 247 + ], + "spans": [ + { + "bbox": [ + 137, + 230, + 295, + 247 + ], + "type": "interline_equation", + "content": "\\left(1 - \\underline {{\\rho}}\\right) \\bar {k} (\\Omega) < 1. \\tag {11}", + "image_path": "4a828e49231b790f2bd5ab351f4e3225bda98d94e8a2acdba85f3621cbbdf9bc.jpg" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 55, + 255, + 295, + 281 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 255, + 295, + 281 + ], + "spans": [ + { + "bbox": [ + 55, + 255, + 295, + 281 + ], + "type": "text", + "content": "Then every misclassified input " + }, + { + "bbox": [ + 55, + 255, + 295, + 281 + ], + "type": "inline_equation", + "content": "x" + }, + { + "bbox": [ + 55, + 255, + 295, + 281 + ], + "type": "text", + "content": " is an adversarial Doppelgänger." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 55, + 289, + 296, + 326 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 289, + 296, + 326 + ], + "spans": [ + { + "bbox": [ + 55, + 289, + 296, + 326 + ], + "type": "text", + "content": "Definition 6. We say that a classifier has a hypersensitive behavior if every misclassified input is an adversarial Doppelgänger." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 55, + 335, + 295, + 407 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 335, + 295, + 407 + ], + "spans": [ + { + "bbox": [ + 55, + 335, + 295, + 407 + ], + "type": "text", + "content": "For classifiers with hypersensitive behavior adversarial robustness can only be improved by improving accuracies, i.e., by eliminating misclassification. Observation 4 shows that if " + }, + { + "bbox": [ + 55, + 335, + 295, + 407 + ], + "type": "inline_equation", + "content": "\\mu (\\mathfrak{d}(x)) > 0" + }, + { + "bbox": [ + 55, + 335, + 295, + 407 + ], + "type": "text", + "content": ", for all inputs " + }, + { + "bbox": [ + 55, + 335, + 295, + 407 + ], + "type": "inline_equation", + "content": "x\\in \\mathbf{X}" + }, + { + "bbox": [ + 55, + 335, + 295, + 407 + ], + "type": "text", + "content": ", then all classifiers with sufficiently high accuracy are either regular (when accuracy equals to one) or have hypersensitive behavior." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 55, + 407, + 296, + 456 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 407, + 296, + 456 + ], + "spans": [ + { + "bbox": [ + 55, + 407, + 296, + 456 + ], + "type": "text", + "content": "In summary adversarial Doppelgänger robustness - accuracy trade-off may happen for classifiers with middling accuracy rates or when there are inputs whose Doppelgängers are negligible in measure." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 55, + 468, + 180, + 480 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 468, + 180, + 480 + ], + "spans": [ + { + "bbox": [ + 55, + 468, + 180, + 480 + ], + "type": "text", + "content": "6. Life without borders." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 55, + 488, + 295, + 608 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 488, + 295, + 608 + ], + "spans": [ + { + "bbox": [ + 55, + 488, + 295, + 608 + ], + "type": "text", + "content": "A important property of the perceptual topology is that the if a classifier " + }, + { + "bbox": [ + 55, + 488, + 295, + 608 + ], + "type": "inline_equation", + "content": "\\Omega" + }, + { + "bbox": [ + 55, + 488, + 295, + 608 + ], + "type": "text", + "content": " is perceptually regular, then it \"imposes an open [topological] borders policy\", that is, " + }, + { + "bbox": [ + 55, + 488, + 295, + 608 + ], + "type": "inline_equation", + "content": "\\partial \\Omega_{c} = \\emptyset" + }, + { + "bbox": [ + 55, + 488, + 295, + 608 + ], + "type": "text", + "content": " for every label " + }, + { + "bbox": [ + 55, + 488, + 295, + 608 + ], + "type": "inline_equation", + "content": "c" + }, + { + "bbox": [ + 55, + 488, + 295, + 608 + ], + "type": "text", + "content": ". Linguists and psychologists, have observed and postulated that natural perceptual and semantic categories are borderless. See for example [64]. Class/decision boundaries are studied and exploited in many works on classifiers, and in particular on adversarial robustness. These boundaries are artifacts of the metric topology used by the researchers, they are not perceptual phenomena." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 55, + 609, + 295, + 693 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 609, + 295, + 693 + ], + "spans": [ + { + "bbox": [ + 55, + 609, + 295, + 693 + ], + "type": "text", + "content": "However, we are all familiar with the idea that some stimuli are more intrinsic/representative to/of a given class and at the same time frequently there are objects/stimuli/Input that, while they are firmly with in the class, are less representative/share few(er) features compared to the rest of the elements in the class, that is, they \"are/belong to the fringe\" of the class." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 313, + 72, + 553, + 120 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 72, + 553, + 120 + ], + "spans": [ + { + "bbox": [ + 313, + 72, + 553, + 120 + ], + "type": "text", + "content": "Definition 7. Let " + }, + { + "bbox": [ + 313, + 72, + 553, + 120 + ], + "type": "inline_equation", + "content": "s" + }, + { + "bbox": [ + 313, + 72, + 553, + 120 + ], + "type": "text", + "content": " be a similarity scale, i.e., a function " + }, + { + "bbox": [ + 313, + 72, + 553, + 120 + ], + "type": "inline_equation", + "content": "s: \\mathbf{X} \\times \\mathbf{X} \\to \\mathbb{R}" + }, + { + "bbox": [ + 313, + 72, + 553, + 120 + ], + "type": "text", + "content": " such that " + }, + { + "bbox": [ + 313, + 72, + 553, + 120 + ], + "type": "inline_equation", + "content": "s(x, x) \\geq s(x, y), \\forall x, y \\in \\mathbf{X}" + }, + { + "bbox": [ + 313, + 72, + 553, + 120 + ], + "type": "text", + "content": " as in [79], [45] (measuring similarity within a fixed context) and [40]." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 313, + 126, + 555, + 210 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 126, + 555, + 210 + ], + "spans": [ + { + "bbox": [ + 313, + 126, + 555, + 210 + ], + "type": "text", + "content": "The values " + }, + { + "bbox": [ + 313, + 126, + 555, + 210 + ], + "type": "inline_equation", + "content": "s(x, x)" + }, + { + "bbox": [ + 313, + 126, + 555, + 210 + ], + "type": "text", + "content": " can be and sometimes are used to represent the salience or equivalently importance of the input within " + }, + { + "bbox": [ + 313, + 126, + 555, + 210 + ], + "type": "inline_equation", + "content": "\\mathbf{X}" + }, + { + "bbox": [ + 313, + 126, + 555, + 210 + ], + "type": "text", + "content": ", see for example [79]. The similarity scale provides a method to quantify the affinity of a input/stimulus to a given measurable subset " + }, + { + "bbox": [ + 313, + 126, + 555, + 210 + ], + "type": "inline_equation", + "content": "D \\subset \\mathbf{X}" + }, + { + "bbox": [ + 313, + 126, + 555, + 210 + ], + "type": "text", + "content": " and the notions of prototype and fringe. The (s-)affinity of " + }, + { + "bbox": [ + 313, + 126, + 555, + 210 + ], + "type": "inline_equation", + "content": "x" + }, + { + "bbox": [ + 313, + 126, + 555, + 210 + ], + "type": "text", + "content": " with a measurable set " + }, + { + "bbox": [ + 313, + 126, + 555, + 210 + ], + "type": "inline_equation", + "content": "D" + }, + { + "bbox": [ + 313, + 126, + 555, + 210 + ], + "type": "text", + "content": " is defined as" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 387, + 214, + 553, + 241 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 387, + 214, + 553, + 241 + ], + "spans": [ + { + "bbox": [ + 387, + 214, + 553, + 241 + ], + "type": "interline_equation", + "content": "P (x, D) = \\int_ {D} s (x, y) \\tag {12}", + "image_path": "de305b9ab587928205d398a79da3812b532979ca3f05b64b0f035563f57a7a78.jpg" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 313, + 246, + 555, + 270 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 246, + 555, + 270 + ], + "spans": [ + { + "bbox": [ + 313, + 246, + 555, + 270 + ], + "type": "text", + "content": "This is a straightforward generalization of the notion of prototypicality defined in [79]." + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 313, + 276, + 555, + 300 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 276, + 555, + 300 + ], + "spans": [ + { + "bbox": [ + 313, + 276, + 555, + 300 + ], + "type": "text", + "content": "Definition 8. " + }, + { + "bbox": [ + 313, + 276, + 555, + 300 + ], + "type": "inline_equation", + "content": "x \\in D" + }, + { + "bbox": [ + 313, + 276, + 555, + 300 + ], + "type": "text", + "content": " is called a prototype of " + }, + { + "bbox": [ + 313, + 276, + 555, + 300 + ], + "type": "inline_equation", + "content": "D" + }, + { + "bbox": [ + 313, + 276, + 555, + 300 + ], + "type": "text", + "content": " (with respect to the integrable similarity scale " + }, + { + "bbox": [ + 313, + 276, + 555, + 300 + ], + "type": "inline_equation", + "content": "s: \\mathbf{X} \\times \\mathbf{X} \\to \\mathbb{R}" + }, + { + "bbox": [ + 313, + 276, + 555, + 300 + ], + "type": "text", + "content": ") if" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 381, + 312, + 553, + 331 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 381, + 312, + 553, + 331 + ], + "spans": [ + { + "bbox": [ + 381, + 312, + 553, + 331 + ], + "type": "interline_equation", + "content": "P (x, D) = \\sup _ {z \\in D} P (z, D) \\tag {13}", + "image_path": "4800fa036fe644f2525382542e917060a29246bdbece35e2ed2b5e07d0e31462.jpg" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 313, + 335, + 553, + 358 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 335, + 553, + 358 + ], + "spans": [ + { + "bbox": [ + 313, + 335, + 553, + 358 + ], + "type": "inline_equation", + "content": "x \\in D" + }, + { + "bbox": [ + 313, + 335, + 553, + 358 + ], + "type": "text", + "content": " is called a fringe element of " + }, + { + "bbox": [ + 313, + 335, + 553, + 358 + ], + "type": "inline_equation", + "content": "D" + }, + { + "bbox": [ + 313, + 335, + 553, + 358 + ], + "type": "text", + "content": " (with respect to the integrable similarity scale " + }, + { + "bbox": [ + 313, + 335, + 553, + 358 + ], + "type": "inline_equation", + "content": "s: \\mathbf{X} \\times \\mathbf{X} \\to \\mathbb{R}" + }, + { + "bbox": [ + 313, + 335, + 553, + 358 + ], + "type": "text", + "content": ") if" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 381, + 366, + 553, + 384 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 381, + 366, + 553, + 384 + ], + "spans": [ + { + "bbox": [ + 381, + 366, + 553, + 384 + ], + "type": "interline_equation", + "content": "P (x, D) = \\inf _ {z \\in D} P (z, D) \\tag {14}", + "image_path": "c4a11d3e38f61dcd06a485930b4c6124cab54c92a1db4d57e7c841140e3790fe.jpg" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 313, + 391, + 554, + 498 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 391, + 554, + 498 + ], + "spans": [ + { + "bbox": [ + 313, + 391, + 554, + 498 + ], + "type": "text", + "content": "One is tempted to think of prototypes as the stimuli that are \"clearest cases, best examples\". \"easy to tell apart\" and to \"be a good representative\" and hence that optimize salience [13, 31, 63, 64, 79]. However, clearly there is no reason to expect that a stimulus that is unlikely to be observed/encountered would be selected as a prototype. The examples below show that prototypes and fringes are obtained by optimizing a mixture of the frequency of appearance (likelihood to encounter) and salience." + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 313, + 498, + 554, + 558 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 498, + 554, + 558 + ], + "spans": [ + { + "bbox": [ + 313, + 498, + 554, + 558 + ], + "type": "text", + "content": "The core and fringe sets may be empty. In practice one is often satisfied with finding elements that may not be true prototypes but belong to the " + }, + { + "bbox": [ + 313, + 498, + 554, + 558 + ], + "type": "inline_equation", + "content": "M" + }, + { + "bbox": [ + 313, + 498, + 554, + 558 + ], + "type": "text", + "content": "-core, i.e., have affinity exceeding a fixed threshold " + }, + { + "bbox": [ + 313, + 498, + 554, + 558 + ], + "type": "inline_equation", + "content": "M" + }, + { + "bbox": [ + 313, + 498, + 554, + 558 + ], + "type": "text", + "content": ". Similarly, elements whose affinity is below a given threshold " + }, + { + "bbox": [ + 313, + 498, + 554, + 558 + ], + "type": "inline_equation", + "content": "\\tau" + }, + { + "bbox": [ + 313, + 498, + 554, + 558 + ], + "type": "text", + "content": " belong to the " + }, + { + "bbox": [ + 313, + 498, + 554, + 558 + ], + "type": "inline_equation", + "content": "\\tau" + }, + { + "bbox": [ + 313, + 498, + 554, + 558 + ], + "type": "text", + "content": "-fringe." + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 313, + 559, + 555, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 559, + 555, + 714 + ], + "spans": [ + { + "bbox": [ + 313, + 559, + 555, + 714 + ], + "type": "text", + "content": "Many perceptual and cognitive processes exploit the intelligent agents' abilities to measure and/or compare the salience/importance of features. In some simple cases it is expected and it even might be true that salience/importance is a probability measure " + }, + { + "bbox": [ + 313, + 559, + 555, + 714 + ], + "type": "inline_equation", + "content": "f_{\\Phi}" + }, + { + "bbox": [ + 313, + 559, + 555, + 714 + ], + "type": "text", + "content": " on the space of (all possible) features " + }, + { + "bbox": [ + 313, + 559, + 555, + 714 + ], + "type": "inline_equation", + "content": "\\Phi" + }, + { + "bbox": [ + 313, + 559, + 555, + 714 + ], + "type": "text", + "content": " and the feature representations " + }, + { + "bbox": [ + 313, + 559, + 555, + 714 + ], + "type": "inline_equation", + "content": "\\Phi_x" + }, + { + "bbox": [ + 313, + 559, + 555, + 714 + ], + "type": "text", + "content": " are measurable subsets of " + }, + { + "bbox": [ + 313, + 559, + 555, + 714 + ], + "type": "inline_equation", + "content": "\\Phi" + }, + { + "bbox": [ + 313, + 559, + 555, + 714 + ], + "type": "text", + "content": ". In many accounts including Tversky's feature contrast model [79] the salience scale " + }, + { + "bbox": [ + 313, + 559, + 555, + 714 + ], + "type": "inline_equation", + "content": "f_{\\Phi}" + }, + { + "bbox": [ + 313, + 559, + 555, + 714 + ], + "type": "text", + "content": " is a context dependent, nonnegative function defined on a collection " + }, + { + "bbox": [ + 313, + 559, + 555, + 714 + ], + "type": "inline_equation", + "content": "\\Upsilon(\\Phi)" + }, + { + "bbox": [ + 313, + 559, + 555, + 714 + ], + "type": "text", + "content": " of subsets of " + }, + { + "bbox": [ + 313, + 559, + 555, + 714 + ], + "type": "inline_equation", + "content": "\\Phi" + }, + { + "bbox": [ + 313, + 559, + 555, + 714 + ], + "type": "text", + "content": " which is closed under finite unions and intersections, and set differences. Furthermore, " + }, + { + "bbox": [ + 313, + 559, + 555, + 714 + ], + "type": "inline_equation", + "content": "\\Phi_x \\in \\Upsilon(\\Phi), \\forall x \\in \\mathbf{X}" + }, + { + "bbox": [ + 313, + 559, + 555, + 714 + ], + "type": "text", + "content": "; and the non-negative function " + }, + { + "bbox": [ + 313, + 559, + 555, + 714 + ], + "type": "inline_equation", + "content": "f_{\\Phi}" + }, + { + "bbox": [ + 313, + 559, + 555, + 714 + ], + "type": "text", + "content": " is feature additive, i.e., " + }, + { + "bbox": [ + 313, + 559, + 555, + 714 + ], + "type": "inline_equation", + "content": "f_{\\Phi}(A \\cup B) = f_{\\Phi}(A) +" + } + ] + } + ], + "index": 21 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 64, + 702, + 177, + 713 + ], + "type": "page_footnote", + "angle": 0, + "lines": [ + { + "bbox": [ + 64, + 702, + 177, + 713 + ], + "spans": [ + { + "bbox": [ + 64, + 702, + 177, + 713 + ], + "type": "text", + "content": "15The proof is in Appendix Part E." + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "text", + "content": "10249" + } + ] + } + ], + "index": 23 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 5 + }, + { + "para_blocks": [ + { + "bbox": [ + 55, + 72, + 296, + 214 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 72, + 296, + 214 + ], + "spans": [ + { + "bbox": [ + 55, + 72, + 296, + 214 + ], + "type": "inline_equation", + "content": "f_{\\Phi}(B)" + }, + { + "bbox": [ + 55, + 72, + 296, + 214 + ], + "type": "text", + "content": ", if " + }, + { + "bbox": [ + 55, + 72, + 296, + 214 + ], + "type": "inline_equation", + "content": "A \\cap B = \\emptyset" + }, + { + "bbox": [ + 55, + 72, + 296, + 214 + ], + "type": "text", + "content": ". The value " + }, + { + "bbox": [ + 55, + 72, + 296, + 214 + ], + "type": "inline_equation", + "content": "f(x) = f_{\\phi}(\\Phi_x), \\forall x \\in \\mathbf{X}" + }, + { + "bbox": [ + 55, + 72, + 296, + 214 + ], + "type": "text", + "content": " is called the salience/prominence of the input " + }, + { + "bbox": [ + 55, + 72, + 296, + 214 + ], + "type": "inline_equation", + "content": "x" + }, + { + "bbox": [ + 55, + 72, + 296, + 214 + ], + "type": "text", + "content": ", [79]. We will call a (Tversky) salience scale " + }, + { + "bbox": [ + 55, + 72, + 296, + 214 + ], + "type": "inline_equation", + "content": "f_{\\Phi}" + }, + { + "bbox": [ + 55, + 72, + 296, + 214 + ], + "type": "text", + "content": " perceptually regular if the prominence/salience function " + }, + { + "bbox": [ + 55, + 72, + 296, + 214 + ], + "type": "inline_equation", + "content": "f(x) = f_{\\phi}(\\Phi_x)" + }, + { + "bbox": [ + 55, + 72, + 296, + 214 + ], + "type": "text", + "content": " is perceptually regular. We will call a perceptually regular salience scale " + }, + { + "bbox": [ + 55, + 72, + 296, + 214 + ], + "type": "inline_equation", + "content": "f_{\\Phi}" + }, + { + "bbox": [ + 55, + 72, + 296, + 214 + ], + "type": "text", + "content": " fully deployable if it can be used to judge the distinguishing features of inputs/stimuli. In particular, we can use " + }, + { + "bbox": [ + 55, + 72, + 296, + 214 + ], + "type": "inline_equation", + "content": "f_{\\Phi}(\\Phi_x \\setminus \\Phi_y)" + }, + { + "bbox": [ + 55, + 72, + 296, + 214 + ], + "type": "text", + "content": " to discriminate " + }, + { + "bbox": [ + 55, + 72, + 296, + 214 + ], + "type": "inline_equation", + "content": "x" + }, + { + "bbox": [ + 55, + 72, + 296, + 214 + ], + "type": "text", + "content": " from " + }, + { + "bbox": [ + 55, + 72, + 296, + 214 + ], + "type": "inline_equation", + "content": "y" + }, + { + "bbox": [ + 55, + 72, + 296, + 214 + ], + "type": "text", + "content": ". Thus if " + }, + { + "bbox": [ + 55, + 72, + 296, + 214 + ], + "type": "inline_equation", + "content": "f_{\\Phi}" + }, + { + "bbox": [ + 55, + 72, + 296, + 214 + ], + "type": "text", + "content": " is fully deployable, then " + }, + { + "bbox": [ + 55, + 72, + 296, + 214 + ], + "type": "inline_equation", + "content": "f_{\\Phi}(\\Phi_x \\setminus \\Phi_y) = 0" + }, + { + "bbox": [ + 55, + 72, + 296, + 214 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 55, + 72, + 296, + 214 + ], + "type": "inline_equation", + "content": "f_{\\Phi}(\\Phi_y \\setminus \\Phi_x) = 0" + }, + { + "bbox": [ + 55, + 72, + 296, + 214 + ], + "type": "text", + "content": " whenever " + }, + { + "bbox": [ + 55, + 72, + 296, + 214 + ], + "type": "inline_equation", + "content": "x \\stackrel{\\alpha \\delta}{\\approx} y" + }, + { + "bbox": [ + 55, + 72, + 296, + 214 + ], + "type": "text", + "content": " or equivalently " + }, + { + "bbox": [ + 55, + 72, + 296, + 214 + ], + "type": "inline_equation", + "content": "f_{\\Phi}(\\Phi_x \\cap \\Phi_y) = f(x)" + }, + { + "bbox": [ + 55, + 72, + 296, + 214 + ], + "type": "text", + "content": " if " + }, + { + "bbox": [ + 55, + 72, + 296, + 214 + ], + "type": "inline_equation", + "content": "x \\stackrel{\\alpha \\delta}{\\approx} y" + }, + { + "bbox": [ + 55, + 72, + 296, + 214 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 55, + 217, + 296, + 254 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 217, + 296, + 254 + ], + "spans": [ + { + "bbox": [ + 55, + 217, + 296, + 254 + ], + "type": "text", + "content": "The following special case provides a particularly useful insight into the nature of prototypes and the possible internal structure of categories as discussed in [64]." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 55, + 269, + 296, + 330 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 269, + 296, + 330 + ], + "spans": [ + { + "bbox": [ + 55, + 269, + 296, + 330 + ], + "type": "text", + "content": "Example 4: If " + }, + { + "bbox": [ + 55, + 269, + 296, + 330 + ], + "type": "inline_equation", + "content": "s" + }, + { + "bbox": [ + 55, + 269, + 296, + 330 + ], + "type": "text", + "content": " is a contrast similarity [79] such that " + }, + { + "bbox": [ + 55, + 269, + 296, + 330 + ], + "type": "inline_equation", + "content": "f_{\\Phi}(\\Phi_x) = s(x,x) / \\theta" + }, + { + "bbox": [ + 55, + 269, + 296, + 330 + ], + "type": "text", + "content": " is fully deployable, " + }, + { + "bbox": [ + 55, + 269, + 296, + 330 + ], + "type": "inline_equation", + "content": "\\theta \\in (0, + \\infty)" + }, + { + "bbox": [ + 55, + 269, + 296, + 330 + ], + "type": "text", + "content": ", and furthermore, " + }, + { + "bbox": [ + 55, + 269, + 296, + 330 + ], + "type": "inline_equation", + "content": "f_{\\Phi}(\\Phi_x \\cap \\Phi_y) = f_{\\Phi}(\\Phi_x)" + }, + { + "bbox": [ + 55, + 269, + 296, + 330 + ], + "type": "text", + "content": " for every pair " + }, + { + "bbox": [ + 55, + 269, + 296, + 330 + ], + "type": "inline_equation", + "content": "x \\sim_{\\sigma} y" + }, + { + "bbox": [ + 55, + 269, + 296, + 330 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 55, + 269, + 296, + 330 + ], + "type": "inline_equation", + "content": "\\Phi_x \\cap \\Phi_y = \\emptyset" + }, + { + "bbox": [ + 55, + 269, + 296, + 330 + ], + "type": "text", + "content": " if " + }, + { + "bbox": [ + 55, + 269, + 296, + 330 + ], + "type": "inline_equation", + "content": "x \\not\\sim_{\\sigma} y" + }, + { + "bbox": [ + 55, + 269, + 296, + 330 + ], + "type": "text", + "content": ", then for every regular class " + }, + { + "bbox": [ + 55, + 269, + 296, + 330 + ], + "type": "inline_equation", + "content": "D \\subset \\mathbf{X}" + }, + { + "bbox": [ + 55, + 269, + 296, + 330 + ], + "type": "text", + "content": ", and every " + }, + { + "bbox": [ + 55, + 269, + 296, + 330 + ], + "type": "inline_equation", + "content": "x \\in D" + }, + { + "bbox": [ + 55, + 269, + 296, + 330 + ], + "type": "text", + "content": " we get" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 55, + 346, + 296, + 381 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 346, + 296, + 381 + ], + "spans": [ + { + "bbox": [ + 55, + 346, + 296, + 381 + ], + "type": "interline_equation", + "content": "P (x, D) = \\Theta \\mu (D) \\left(\\frac {\\mu ([ x ] _ {\\sim_ {\\sigma}})}{\\mu (D)} - \\frac {\\alpha}{\\Theta}\\right) f _ {\\Phi} \\left(\\Phi_ {x}\\right) - \\beta I _ {\\Phi} (D), \\tag {15}", + "image_path": "2971deee2269fe6fca19606ba9ad01845fc56e2a9463cd6f65a08e239618cbbf.jpg" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 55, + 383, + 296, + 454 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 383, + 296, + 454 + ], + "spans": [ + { + "bbox": [ + 55, + 383, + 296, + 454 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 55, + 383, + 296, + 454 + ], + "type": "inline_equation", + "content": "\\Theta = (\\alpha +\\beta +\\theta)" + }, + { + "bbox": [ + 55, + 383, + 296, + 454 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 55, + 383, + 296, + 454 + ], + "type": "inline_equation", + "content": "\\alpha ,\\beta" + }, + { + "bbox": [ + 55, + 383, + 296, + 454 + ], + "type": "text", + "content": " ,and " + }, + { + "bbox": [ + 55, + 383, + 296, + 454 + ], + "type": "inline_equation", + "content": "\\theta" + }, + { + "bbox": [ + 55, + 383, + 296, + 454 + ], + "type": "text", + "content": " are non-negative constants. In particular, if " + }, + { + "bbox": [ + 55, + 383, + 296, + 454 + ], + "type": "inline_equation", + "content": "D" + }, + { + "bbox": [ + 55, + 383, + 296, + 454 + ], + "type": "text", + "content": " is a finite union of equivalence classes " + }, + { + "bbox": [ + 55, + 383, + 296, + 454 + ], + "type": "inline_equation", + "content": "\\zeta_j\\in \\mathbf{X} / \\sim_\\sigma" + }, + { + "bbox": [ + 55, + 383, + 296, + 454 + ], + "type": "text", + "content": " , then both prototypes and fringe elements exist. Furthermore, if " + }, + { + "bbox": [ + 55, + 383, + 296, + 454 + ], + "type": "inline_equation", + "content": "x" + }, + { + "bbox": [ + 55, + 383, + 296, + 454 + ], + "type": "text", + "content": " is prototype/fringe then so are all stimuli in its component " + }, + { + "bbox": [ + 55, + 383, + 296, + 454 + ], + "type": "inline_equation", + "content": "\\zeta = [x]_{\\sim_{\\sigma}}" + }, + { + "bbox": [ + 55, + 383, + 296, + 454 + ], + "type": "text", + "content": " . Similar statements hold for M-core and " + }, + { + "bbox": [ + 55, + 383, + 296, + 454 + ], + "type": "inline_equation", + "content": "\\tau" + }, + { + "bbox": [ + 55, + 383, + 296, + 454 + ], + "type": "text", + "content": " -fringe elements." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 55, + 458, + 296, + 579 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 458, + 296, + 579 + ], + "spans": [ + { + "bbox": [ + 55, + 458, + 296, + 579 + ], + "type": "text", + "content": "If the class " + }, + { + "bbox": [ + 55, + 458, + 296, + 579 + ], + "type": "inline_equation", + "content": "D" + }, + { + "bbox": [ + 55, + 458, + 296, + 579 + ], + "type": "text", + "content": " is fixed, then " + }, + { + "bbox": [ + 55, + 458, + 296, + 579 + ], + "type": "inline_equation", + "content": "\\frac{\\mu([x]_{\\sim_{\\sigma}})}{\\mu(D)}" + }, + { + "bbox": [ + 55, + 458, + 296, + 579 + ], + "type": "text", + "content": " is just the probability to encounter (and learn) a set of stimuli, and " + }, + { + "bbox": [ + 55, + 458, + 296, + 579 + ], + "type": "inline_equation", + "content": "f_{\\Phi}(\\Phi_x)" + }, + { + "bbox": [ + 55, + 458, + 296, + 579 + ], + "type": "text", + "content": " can be interpreted as the level of prominence. So the optimization process involves learning prominent examples that can be encountered reasonably often. Thus in this case the prototype does not fall in either of the two main branches of prototypes, that is inputs that represent the central tendency in the regular class vs. prototypes as highly \"representative exemplar(s) of a category\", see [15], page 52." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 55, + 582, + 296, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 582, + 296, + 713 + ], + "spans": [ + { + "bbox": [ + 55, + 582, + 296, + 713 + ], + "type": "text", + "content": "More generally, when the similarity scale is bounded, for example, this is true for real similarity measures deployable by humans, then, as predicted by modern prototype theory, each perceptually regular subset " + }, + { + "bbox": [ + 55, + 582, + 296, + 713 + ], + "type": "inline_equation", + "content": "D \\subset \\mathbf{X}" + }, + { + "bbox": [ + 55, + 582, + 296, + 713 + ], + "type": "text", + "content": " corresponding to real (natural) categories created and analyzed by real intelligent agents consists of core elements (possibly M-core), a layer of fringe (possibly " + }, + { + "bbox": [ + 55, + 582, + 296, + 713 + ], + "type": "inline_equation", + "content": "\\tau" + }, + { + "bbox": [ + 55, + 582, + 296, + 713 + ], + "type": "text", + "content": "-fringe) elements, and layers of elements of various levels of intermediate affinity with " + }, + { + "bbox": [ + 55, + 582, + 296, + 713 + ], + "type": "inline_equation", + "content": "D" + }, + { + "bbox": [ + 55, + 582, + 296, + 713 + ], + "type": "text", + "content": ". In Part I of the Appendix we introduce class invariants of regular classifiers including structural entropy, expected index of coincidence, and importance." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 313, + 71, + 538, + 86 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 71, + 538, + 86 + ], + "spans": [ + { + "bbox": [ + 313, + 71, + 538, + 86 + ], + "type": "text", + "content": "7. Quantifying Doppelgänger Vulnerability." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 313, + 91, + 554, + 156 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 91, + 554, + 156 + ], + "spans": [ + { + "bbox": [ + 313, + 91, + 554, + 156 + ], + "type": "text", + "content": "By definition if a classifier " + }, + { + "bbox": [ + 313, + 91, + 554, + 156 + ], + "type": "inline_equation", + "content": "R" + }, + { + "bbox": [ + 313, + 91, + 554, + 156 + ], + "type": "text", + "content": " is not regular, then it is vulnerable to adversarial Doppelgangers attacks, that is, for some input " + }, + { + "bbox": [ + 313, + 91, + 554, + 156 + ], + "type": "inline_equation", + "content": "x" + }, + { + "bbox": [ + 313, + 91, + 554, + 156 + ], + "type": "text", + "content": " there exists " + }, + { + "bbox": [ + 313, + 91, + 554, + 156 + ], + "type": "inline_equation", + "content": "a(x) \\stackrel{\\alpha \\delta}{\\approx} x" + }, + { + "bbox": [ + 313, + 91, + 554, + 156 + ], + "type": "text", + "content": " and such that " + }, + { + "bbox": [ + 313, + 91, + 554, + 156 + ], + "type": "inline_equation", + "content": "\\operatorname{label}_R(x) \\neq \\operatorname{label}_R(a(x))" + }, + { + "bbox": [ + 313, + 91, + 554, + 156 + ], + "type": "text", + "content": ". In particular, we say that " + }, + { + "bbox": [ + 313, + 91, + 554, + 156 + ], + "type": "inline_equation", + "content": "R" + }, + { + "bbox": [ + 313, + 91, + 554, + 156 + ], + "type": "text", + "content": " is conceptually ambiguous at " + }, + { + "bbox": [ + 313, + 91, + 554, + 156 + ], + "type": "inline_equation", + "content": "x" + }, + { + "bbox": [ + 313, + 91, + 554, + 156 + ], + "type": "text", + "content": " and we call the set" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 318, + 163, + 553, + 190 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 318, + 163, + 553, + 190 + ], + "spans": [ + { + "bbox": [ + 318, + 163, + 553, + 190 + ], + "type": "interline_equation", + "content": "\\mathrm {A} (R) = \\left\\{x \\in \\mathbf {X}: \\exists y \\stackrel {\\alpha \\delta} {\\approx} x \\text {a n d} \\operatorname {l a b e l} _ {R} (x) \\neq \\operatorname {l a b e l} _ {R} (y) \\right\\} \\tag {16}", + "image_path": "1541d0507861b2afe71392b703294fa73d0a233f858d08727615847ce1d5c756.jpg" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 313, + 191, + 554, + 227 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 191, + 554, + 227 + ], + "spans": [ + { + "bbox": [ + 313, + 191, + 554, + 227 + ], + "type": "text", + "content": "the region of conceptual ambiguity. When " + }, + { + "bbox": [ + 313, + 191, + 554, + 227 + ], + "type": "inline_equation", + "content": "(X, \\mu)" + }, + { + "bbox": [ + 313, + 191, + 554, + 227 + ], + "type": "text", + "content": " is a probability measure space and " + }, + { + "bbox": [ + 313, + 191, + 554, + 227 + ], + "type": "inline_equation", + "content": "\\mu(\\mathfrak{d}(x)) > 0" + }, + { + "bbox": [ + 313, + 191, + 554, + 227 + ], + "type": "text", + "content": ", we use the probability distribution of labels at " + }, + { + "bbox": [ + 313, + 191, + 554, + 227 + ], + "type": "inline_equation", + "content": "x" + }, + { + "bbox": [ + 313, + 191, + 554, + 227 + ], + "type": "text", + "content": ":" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 349, + 233, + 553, + 261 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 349, + 233, + 553, + 261 + ], + "spans": [ + { + "bbox": [ + 349, + 233, + 553, + 261 + ], + "type": "interline_equation", + "content": "p _ {j} (x) = \\frac {\\mu \\left(R _ {j} \\bigcap \\mathfrak {d} (x)\\right)}{\\mu (\\mathfrak {d} (x))}, \\quad j = 1, \\dots , m \\tag {17}", + "image_path": "ee62465168d9f1e7dea710413337bbcf642998e7aec6cd81ab7184b7d52c1891.jpg" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 313, + 266, + 512, + 278 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 266, + 512, + 278 + ], + "spans": [ + { + "bbox": [ + 313, + 266, + 512, + 278 + ], + "type": "text", + "content": "and the conceptual entropy of " + }, + { + "bbox": [ + 313, + 266, + 512, + 278 + ], + "type": "inline_equation", + "content": "R" + }, + { + "bbox": [ + 313, + 266, + 512, + 278 + ], + "type": "text", + "content": " at " + }, + { + "bbox": [ + 313, + 266, + 512, + 278 + ], + "type": "inline_equation", + "content": "x" + }, + { + "bbox": [ + 313, + 266, + 512, + 278 + ], + "type": "text", + "content": " defined as" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 365, + 285, + 553, + 318 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 365, + 285, + 553, + 318 + ], + "spans": [ + { + "bbox": [ + 365, + 285, + 553, + 318 + ], + "type": "interline_equation", + "content": "H _ {R} (x) = - \\sum_ {j = 1} ^ {m} p _ {j} (x) \\log \\left(p _ {j} (x)\\right) \\tag {18}", + "image_path": "dca283d60d71df39a93829d8c4ad4d407c06e46b93658a779e486aff566554dd.jpg" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 313, + 325, + 554, + 361 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 325, + 554, + 361 + ], + "spans": [ + { + "bbox": [ + 313, + 325, + 554, + 361 + ], + "type": "text", + "content": "to detect whether " + }, + { + "bbox": [ + 313, + 325, + 554, + 361 + ], + "type": "inline_equation", + "content": "R" + }, + { + "bbox": [ + 313, + 325, + 554, + 361 + ], + "type": "text", + "content": " is conceptually ambiguous at " + }, + { + "bbox": [ + 313, + 325, + 554, + 361 + ], + "type": "inline_equation", + "content": "x" + }, + { + "bbox": [ + 313, + 325, + 554, + 361 + ], + "type": "text", + "content": " (i.e., " + }, + { + "bbox": [ + 313, + 325, + 554, + 361 + ], + "type": "inline_equation", + "content": "H_{R}(x) > 0" + }, + { + "bbox": [ + 313, + 325, + 554, + 361 + ], + "type": "text", + "content": "), and to quantify the likelihoods of various adversarial Doppelgänger attacks." + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 313, + 368, + 554, + 456 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 368, + 554, + 456 + ], + "spans": [ + { + "bbox": [ + 313, + 368, + 554, + 456 + ], + "type": "text", + "content": "Definition 9. Let " + }, + { + "bbox": [ + 313, + 368, + 554, + 456 + ], + "type": "inline_equation", + "content": "R" + }, + { + "bbox": [ + 313, + 368, + 554, + 456 + ], + "type": "text", + "content": " be a classifier and " + }, + { + "bbox": [ + 313, + 368, + 554, + 456 + ], + "type": "inline_equation", + "content": "\\hat{a} : \\mathbf{X} \\to \\mathbf{X}" + }, + { + "bbox": [ + 313, + 368, + 554, + 456 + ], + "type": "text", + "content": ", we will call the inner measure of the set " + }, + { + "bbox": [ + 313, + 368, + 554, + 456 + ], + "type": "inline_equation", + "content": "\\{x : \\operatorname{label}_R(\\hat{a}(x)) \\neq \\operatorname{label}_R(x)\\}" + }, + { + "bbox": [ + 313, + 368, + 554, + 456 + ], + "type": "text", + "content": " the " + }, + { + "bbox": [ + 313, + 368, + 554, + 456 + ], + "type": "inline_equation", + "content": "R" + }, + { + "bbox": [ + 313, + 368, + 554, + 456 + ], + "type": "text", + "content": "-fooling rate of the mapping " + }, + { + "bbox": [ + 313, + 368, + 554, + 456 + ], + "type": "inline_equation", + "content": "\\hat{a}" + }, + { + "bbox": [ + 313, + 368, + 554, + 456 + ], + "type": "text", + "content": ", and we will denote it by " + }, + { + "bbox": [ + 313, + 368, + 554, + 456 + ], + "type": "inline_equation", + "content": "F_R(\\hat{a})" + }, + { + "bbox": [ + 313, + 368, + 554, + 456 + ], + "type": "text", + "content": ". A mapping " + }, + { + "bbox": [ + 313, + 368, + 554, + 456 + ], + "type": "inline_equation", + "content": "\\hat{a} : \\mathbf{X} \\to \\mathbf{X}" + }, + { + "bbox": [ + 313, + 368, + 554, + 456 + ], + "type": "text", + "content": " is called an adversarial Doppelgänger attack to a classifier " + }, + { + "bbox": [ + 313, + 368, + 554, + 456 + ], + "type": "inline_equation", + "content": "R" + }, + { + "bbox": [ + 313, + 368, + 554, + 456 + ], + "type": "text", + "content": " if and only if " + }, + { + "bbox": [ + 313, + 368, + 554, + 456 + ], + "type": "inline_equation", + "content": "\\hat{a}(x) \\stackrel{\\alpha \\delta}{\\approx} x, \\forall x \\in \\mathbf{X}" + }, + { + "bbox": [ + 313, + 368, + 554, + 456 + ], + "type": "text", + "content": ", and the " + }, + { + "bbox": [ + 313, + 368, + 554, + 456 + ], + "type": "inline_equation", + "content": "R" + }, + { + "bbox": [ + 313, + 368, + 554, + 456 + ], + "type": "text", + "content": "-fooling rate " + }, + { + "bbox": [ + 313, + 368, + 554, + 456 + ], + "type": "inline_equation", + "content": "F_R(\\hat{a})" + }, + { + "bbox": [ + 313, + 368, + 554, + 456 + ], + "type": "text", + "content": " is positive." + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 313, + 462, + 554, + 509 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 462, + 554, + 509 + ], + "spans": [ + { + "bbox": [ + 313, + 462, + 554, + 509 + ], + "type": "text", + "content": "The set " + }, + { + "bbox": [ + 313, + 462, + 554, + 509 + ], + "type": "inline_equation", + "content": "\\{x:\\mathrm{label}_R(\\hat{a} (x))\\neq \\mathrm{label}_R(x)\\}" + }, + { + "bbox": [ + 313, + 462, + 554, + 509 + ], + "type": "text", + "content": " is a subset of the region of conceptual ambiguity of " + }, + { + "bbox": [ + 313, + 462, + 554, + 509 + ], + "type": "inline_equation", + "content": "R" + }, + { + "bbox": [ + 313, + 462, + 554, + 509 + ], + "type": "text", + "content": " , which yields an upper bound on the " + }, + { + "bbox": [ + 313, + 462, + 554, + 509 + ], + "type": "inline_equation", + "content": "R" + }, + { + "bbox": [ + 313, + 462, + 554, + 509 + ], + "type": "text", + "content": " -fooling rate by the outer measure of " + }, + { + "bbox": [ + 313, + 462, + 554, + 509 + ], + "type": "inline_equation", + "content": "A(R)" + }, + { + "bbox": [ + 313, + 462, + 554, + 509 + ], + "type": "text", + "content": " .." + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 391, + 510, + 553, + 523 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 391, + 510, + 553, + 523 + ], + "spans": [ + { + "bbox": [ + 391, + 510, + 553, + 523 + ], + "type": "interline_equation", + "content": "F _ {R} (\\hat {a}) \\leq \\mu_ {*} (A (R)). \\tag {19}", + "image_path": "53b91661eabcd6abce0247ca3df50d881613b97f3be092d2992623bba7ffe283.jpg" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 313, + 527, + 554, + 574 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 527, + 554, + 574 + ], + "spans": [ + { + "bbox": [ + 313, + 527, + 554, + 574 + ], + "type": "text", + "content": "In specific scenarios it is possible to get an upper bound on the size, possibly the outer measure, of " + }, + { + "bbox": [ + 313, + 527, + 554, + 574 + ], + "type": "inline_equation", + "content": "A(R)" + }, + { + "bbox": [ + 313, + 527, + 554, + 574 + ], + "type": "text", + "content": " which in turn shows that the " + }, + { + "bbox": [ + 313, + 527, + 554, + 574 + ], + "type": "inline_equation", + "content": "R" + }, + { + "bbox": [ + 313, + 527, + 554, + 574 + ], + "type": "text", + "content": "-fooling rates are bounded away from one. See Example 11 in Part J in the Appendix." + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 313, + 575, + 554, + 658 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 575, + 554, + 658 + ], + "spans": [ + { + "bbox": [ + 313, + 575, + 554, + 658 + ], + "type": "text", + "content": "Adversarial Doppelgänger attacks are distinct from the adversarial attacks studied to date. The universal adversarial attacks, [47], can achieve fooling rates as close to one as one desires. As illustrated above, adversarial Doppelgänger attacks may not be able to reach fooling rates that are too high. On the other hand in some cases, the optimal fooling rate of one can be achieved." + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 313, + 665, + 554, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 665, + 554, + 713 + ], + "spans": [ + { + "bbox": [ + 313, + 665, + 554, + 713 + ], + "type": "text", + "content": "Observation 5. If " + }, + { + "bbox": [ + 313, + 665, + 554, + 713 + ], + "type": "inline_equation", + "content": "R" + }, + { + "bbox": [ + 313, + 665, + 554, + 713 + ], + "type": "text", + "content": " is conceptually ambiguous at every " + }, + { + "bbox": [ + 313, + 665, + 554, + 713 + ], + "type": "inline_equation", + "content": "x \\in \\mathbf{X}" + }, + { + "bbox": [ + 313, + 665, + 554, + 713 + ], + "type": "text", + "content": ", e.g., when " + }, + { + "bbox": [ + 313, + 665, + 554, + 713 + ], + "type": "inline_equation", + "content": "H_{R}(x) > 0" + }, + { + "bbox": [ + 313, + 665, + 554, + 713 + ], + "type": "text", + "content": " for every " + }, + { + "bbox": [ + 313, + 665, + 554, + 713 + ], + "type": "inline_equation", + "content": "x \\in \\mathbf{X}" + }, + { + "bbox": [ + 313, + 665, + 554, + 713 + ], + "type": "text", + "content": ", then there exists an adversarial Doppelgänger attack with " + }, + { + "bbox": [ + 313, + 665, + 554, + 713 + ], + "type": "inline_equation", + "content": "R" + }, + { + "bbox": [ + 313, + 665, + 554, + 713 + ], + "type": "text", + "content": "-fooling rate equal to one." + } + ] + } + ], + "index": 20 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "text", + "content": "10250" + } + ] + } + ], + "index": 21 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 6 + }, + { + "para_blocks": [ + { + "bbox": [ + 55, + 72, + 294, + 156 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 72, + 294, + 156 + ], + "spans": [ + { + "bbox": [ + 55, + 72, + 294, + 156 + ], + "type": "text", + "content": "Indeed, if " + }, + { + "bbox": [ + 55, + 72, + 294, + 156 + ], + "type": "inline_equation", + "content": "R" + }, + { + "bbox": [ + 55, + 72, + 294, + 156 + ], + "type": "text", + "content": " is conceptually ambiguous at every " + }, + { + "bbox": [ + 55, + 72, + 294, + 156 + ], + "type": "inline_equation", + "content": "x \\in \\mathbf{X}" + }, + { + "bbox": [ + 55, + 72, + 294, + 156 + ], + "type": "text", + "content": ", then " + }, + { + "bbox": [ + 55, + 72, + 294, + 156 + ], + "type": "inline_equation", + "content": "\\{y \\in \\mathfrak{d}(x) : \\operatorname{label}_R(y) \\neq \\operatorname{label}_R(x)\\} \\neq \\emptyset" + }, + { + "bbox": [ + 55, + 72, + 294, + 156 + ], + "type": "text", + "content": ", for every " + }, + { + "bbox": [ + 55, + 72, + 294, + 156 + ], + "type": "inline_equation", + "content": "x" + }, + { + "bbox": [ + 55, + 72, + 294, + 156 + ], + "type": "text", + "content": ", and therefore the axiom of choice implies that there exists a map " + }, + { + "bbox": [ + 55, + 72, + 294, + 156 + ], + "type": "inline_equation", + "content": "\\hat{a} : \\mathbf{X} \\to \\mathbf{X}" + }, + { + "bbox": [ + 55, + 72, + 294, + 156 + ], + "type": "text", + "content": " such that " + }, + { + "bbox": [ + 55, + 72, + 294, + 156 + ], + "type": "inline_equation", + "content": "\\hat{a}(x) \\in \\{y \\in \\mathfrak{d}(x) : \\operatorname{label}_R(y) \\neq \\operatorname{label}_R(x)\\}" + }, + { + "bbox": [ + 55, + 72, + 294, + 156 + ], + "type": "text", + "content": ", for every " + }, + { + "bbox": [ + 55, + 72, + 294, + 156 + ], + "type": "inline_equation", + "content": "x \\in X" + }, + { + "bbox": [ + 55, + 72, + 294, + 156 + ], + "type": "text", + "content": ". It turns out that in practice there may be many classifiers that are conceptually ambiguous at every input." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 55, + 156, + 294, + 262 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 156, + 294, + 262 + ], + "spans": [ + { + "bbox": [ + 55, + 156, + 294, + 262 + ], + "type": "text", + "content": "Example 5: Consider the case when " + }, + { + "bbox": [ + 55, + 156, + 294, + 262 + ], + "type": "inline_equation", + "content": "\\stackrel{\\alpha\\delta}{\\approx}" + }, + { + "bbox": [ + 55, + 156, + 294, + 262 + ], + "type": "text", + "content": " is transitive, i.e., " + }, + { + "bbox": [ + 55, + 156, + 294, + 262 + ], + "type": "inline_equation", + "content": "\\stackrel{\\alpha\\delta}{\\approx}" + }, + { + "bbox": [ + 55, + 156, + 294, + 262 + ], + "type": "text", + "content": " equals its transitive closure " + }, + { + "bbox": [ + 55, + 156, + 294, + 262 + ], + "type": "inline_equation", + "content": "\\sim_{\\sigma}" + }, + { + "bbox": [ + 55, + 156, + 294, + 262 + ], + "type": "text", + "content": ", e.g., when " + }, + { + "bbox": [ + 55, + 156, + 294, + 262 + ], + "type": "inline_equation", + "content": "\\mathbf{X}" + }, + { + "bbox": [ + 55, + 156, + 294, + 262 + ], + "type": "text", + "content": " is equipped with a Grassmann structure as in [39], and there exist at least two different equivalence classes. Thus binary classification is a well defined problem. There exist at least " + }, + { + "bbox": [ + 55, + 156, + 294, + 262 + ], + "type": "inline_equation", + "content": "\\prod_{\\zeta \\in \\mathbf{X} / \\sim_{\\sigma}} \\{U \\subset \\zeta : 0 < \\mu(U) < \\mu(\\zeta)\\}" + }, + { + "bbox": [ + 55, + 156, + 294, + 262 + ], + "type": "text", + "content": " worth of fully-populated binary classifiers which are conceptually ambiguous at every input " + }, + { + "bbox": [ + 55, + 156, + 294, + 262 + ], + "type": "inline_equation", + "content": "x \\in \\mathbf{X}" + }, + { + "bbox": [ + 55, + 156, + 294, + 262 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 55, + 263, + 294, + 300 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 263, + 294, + 300 + ], + "spans": [ + { + "bbox": [ + 55, + 263, + 294, + 300 + ], + "type": "text", + "content": "The last example and Observation 5 show that even high accuracy classifiers can be vulnerable to adversarial Doppelgänger attacks with fooling rate equal to one." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 55, + 300, + 294, + 430 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 300, + 294, + 430 + ], + "spans": [ + { + "bbox": [ + 55, + 300, + 294, + 430 + ], + "type": "text", + "content": "We conclude this section with a warning that popular methods to deal with unseen data, including marking missing data and imputation, may introduce conceptual ambiguity. For example, if a model is trained on a data set " + }, + { + "bbox": [ + 55, + 300, + 294, + 430 + ], + "type": "inline_equation", + "content": "T \\subset \\mathbf{X}" + }, + { + "bbox": [ + 55, + 300, + 294, + 430 + ], + "type": "text", + "content": " that includes only parts of some elementary sets, then adding a class label NA to label unseen data can compromise adversarial Doppelgänger robustness. Indeed, let " + }, + { + "bbox": [ + 55, + 300, + 294, + 430 + ], + "type": "inline_equation", + "content": "S" + }, + { + "bbox": [ + 55, + 300, + 294, + 430 + ], + "type": "text", + "content": " be a nonempty set of training data such that " + }, + { + "bbox": [ + 55, + 300, + 294, + 430 + ], + "type": "inline_equation", + "content": "S \\subsetneq \\zeta \\in \\mathbf{X} / \\sim_{\\sigma}" + }, + { + "bbox": [ + 55, + 300, + 294, + 430 + ], + "type": "text", + "content": ". There exists, " + }, + { + "bbox": [ + 55, + 300, + 294, + 430 + ], + "type": "inline_equation", + "content": "x \\in S" + }, + { + "bbox": [ + 55, + 300, + 294, + 430 + ], + "type": "text", + "content": " such that " + }, + { + "bbox": [ + 55, + 300, + 294, + 430 + ], + "type": "inline_equation", + "content": "\\mathfrak{d}(x) \\setminus S \\neq \\emptyset" + }, + { + "bbox": [ + 55, + 300, + 294, + 430 + ], + "type": "text", + "content": ". Every " + }, + { + "bbox": [ + 55, + 300, + 294, + 430 + ], + "type": "inline_equation", + "content": "z \\in \\mathfrak{d}(x) \\setminus S" + }, + { + "bbox": [ + 55, + 300, + 294, + 430 + ], + "type": "text", + "content": ", labeled as NA, is an adversarial Doppelgänger of " + }, + { + "bbox": [ + 55, + 300, + 294, + 430 + ], + "type": "inline_equation", + "content": "x" + }, + { + "bbox": [ + 55, + 300, + 294, + 430 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 55, + 442, + 214, + 454 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 442, + 214, + 454 + ], + "spans": [ + { + "bbox": [ + 55, + 442, + 214, + 454 + ], + "type": "text", + "content": "8. Discussion and Conclusions." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 55, + 462, + 295, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 462, + 295, + 713 + ], + "spans": [ + { + "bbox": [ + 55, + 462, + 295, + 713 + ], + "type": "text", + "content": "A central focus of this paper is the adversarial Doppelgangers phenomenon, where classifiers assign different labels to inputs that humans cannot discriminate. Until now, this phenomenon has not been well understood, possibly due to the limitations of the distance-based analysis that has dominated the field. In the \"absence of a distance measure that accurately captures the perceptual differences between a source and adversarial example many researchers have decided to use the " + }, + { + "bbox": [ + 55, + 462, + 295, + 713 + ], + "type": "inline_equation", + "content": "\\ell_p" + }, + { + "bbox": [ + 55, + 462, + 295, + 713 + ], + "type": "text", + "content": " distance\", [29]. The available empirical observations and models - both perceptual and cognitive, including those based on just noticeable differences - provide no evidence that biologically plausible perceptual topologies are metric. This paper advances the understanding of context-related perceptual topologies in input spaces, which are rarely metric. Our investigation shows that adversarial Doppelgangers are very close to each other with respect to the context-relevant perceptual metric " + }, + { + "bbox": [ + 55, + 462, + 295, + 713 + ], + "type": "inline_equation", + "content": "d_w" + }, + { + "bbox": [ + 55, + 462, + 295, + 713 + ], + "type": "text", + "content": ", this metric is not a manifold metric and does not generate the perceptual topology. This distinction highlights the shortcomings of traditional, purely manifold metric-based representations and analysis of perceptual spaces." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 313, + 72, + 553, + 335 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 72, + 553, + 335 + ], + "spans": [ + { + "bbox": [ + 313, + 72, + 553, + 335 + ], + "type": "text", + "content": "The machine learning community has expended significant efforts aimed to build adversarially robust classifiers. This may be a march towards a bridge too far. Philosophers, experimental psychologists, and linguists, are well aware that many classification problems are not well defined due to perceptual ambiguities. Any fully populated classifier for a classification problem that is not well defined is doomed to be a victim of the adversarial Doppelgängers phenomenon. Our results reveal the structure of adversarial Doppelgänger-robust classifiers, regular classifiers, and criteria and methods to establish whether a classification problem is well defined or not. The new understanding of the structure of regular classifiers, the analysis of zones of ambiguity, and the methods to measure and bound the fooling rates of adversarial Doppelgänger attacks provide guidance on how to design adversarially robust training to improve classifiers that are not regular. In addition to revealing the impossibility to use accuracy-robustness trade-offs in many scenarios, including robustifying hypersensitive classifiers, our analysis indicates that marking unseen data can jeopardize robustness if the training data contains only a proper subset of an elementary set." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 313, + 335, + 553, + 466 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 335, + 553, + 466 + ], + "spans": [ + { + "bbox": [ + 313, + 335, + 553, + 466 + ], + "type": "text", + "content": "We explore feature representations, the related concept of indiscernibility introduced by Leibniz, and their connection to indiscriminability. This investigation reveals the nature of class prototypes and fringe inputs, and how the size of a discriminative feature representation can be used to determine whether a classification problem is not well defined. Indiscernibility and indiscriminability, are often conflated in the machine learning literature. Elucidating the distinction between them is vital for understanding the limitations of current classifiers and addressing the shortcomings in their design." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 313, + 467, + 553, + 561 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 467, + 553, + 561 + ], + "spans": [ + { + "bbox": [ + 313, + 467, + 553, + 561 + ], + "type": "text", + "content": "Our discussion of the Doppelgängers phenomenon brings to light a significant divergence between human perception and artificial neural network models, including feedforward models, RNN models and ResNet. The indiscriminability relations of these artificial neural network models, studied in [18], are transitive16 while it is well accepted, that, in many contexts, the human indiscriminability relation is not transitive." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 313, + 562, + 553, + 598 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 562, + 553, + 598 + ], + "spans": [ + { + "bbox": [ + 313, + 562, + 553, + 598 + ], + "type": "text", + "content": "The results and insights gained from this investigation point to concrete warnings and actionable steps for improving the training and testing of classifiers." + } + ] + } + ], + "index": 9 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 323, + 702, + 421, + 712 + ], + "type": "page_footnote", + "angle": 0, + "lines": [ + { + "bbox": [ + 323, + 702, + 421, + 712 + ], + "spans": [ + { + "bbox": [ + 323, + 702, + 421, + 712 + ], + "type": "text", + "content": "16See Part K in the Appendix." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "text", + "content": "10251" + } + ] + } + ], + "index": 11 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 7 + }, + { + "para_blocks": [ + { + "bbox": [ + 56, + 71, + 115, + 83 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 71, + 115, + 83 + ], + "spans": [ + { + "bbox": [ + 56, + 71, + 115, + 83 + ], + "type": "text", + "content": "References" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 57, + 91, + 296, + 712 + ], + "type": "list", + "angle": 0, + "index": 17, + "blocks": [ + { + "bbox": [ + 61, + 91, + 296, + 135 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 61, + 91, + 296, + 135 + ], + "spans": [ + { + "bbox": [ + 61, + 91, + 296, + 135 + ], + "type": "text", + "content": "[1] Zeyuan Allen-Zhu and Yuanzhi Li. Feature purification: How adversarial training performs robust deep learning. In 2021 IEEE 62nd Annual Symposium on Foundations of Computer Science (FOCS), pages 977-988. IEEE, 2022. 2" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 61, + 137, + 296, + 168 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 61, + 137, + 296, + 168 + ], + "spans": [ + { + "bbox": [ + 61, + 137, + 296, + 168 + ], + "type": "text", + "content": "[2] Valtteri Arstila. Why the transitivity of perceptual simultaneity should be taken seriously. Frontiers in integrative neuroscience, 6:3, 2012. 3" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 62, + 171, + 296, + 213 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 171, + 296, + 213 + ], + "spans": [ + { + "bbox": [ + 62, + 171, + 296, + 213 + ], + "type": "text", + "content": "[3] Benjamin Balas, Lisa Nakano, and Ruth Rosenholtz. A summary-statistic representation in peripheral vision explains visual crowding. Journal of vision, 9(12):13-13, 2009. 2, 3" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 62, + 216, + 294, + 236 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 216, + 294, + 236 + ], + "spans": [ + { + "bbox": [ + 62, + 216, + 294, + 236 + ], + "type": "text", + "content": "[4] Max Black. The identity of indiscernibles. Mind, 61(242): 153-164, 1952. 1" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 62, + 239, + 295, + 282 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 239, + 295, + 282 + ], + "spans": [ + { + "bbox": [ + 62, + 239, + 295, + 282 + ], + "type": "text", + "content": "[5] Nicholas Boucher, Ilia Shumailov, Ross Anderson, and Nicolas Papernot. Bad characters: Imperceptible nlp attacks. In 2022 IEEE Symposium on Security and Privacy (SP), pages 1987-2004. IEEE, 2022. 2" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 62, + 285, + 295, + 317 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 285, + 295, + 317 + ], + "spans": [ + { + "bbox": [ + 62, + 285, + 295, + 317 + ], + "type": "text", + "content": "[6] William F Broderick, Gizem Rufo, Jonathan Winawer, and Eero P Simoncelli. Foveated metamers of the early visual system. bioRxiv, pages 2023-05, 2023. 1, 2, 3, 6" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 62, + 319, + 295, + 339 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 319, + 295, + 339 + ], + "spans": [ + { + "bbox": [ + 62, + 319, + 295, + 339 + ], + "type": "text", + "content": "[7] Fan RK Chung. Lectures on spectral graph theory. CBMS Lectures, Fresno, 6(92):17-21, 1996. 6" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 62, + 342, + 296, + 449 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 342, + 296, + 449 + ], + "spans": [ + { + "bbox": [ + 62, + 342, + 296, + 449 + ], + "type": "text", + "content": "[8] Samuel Clarke. A Collection of Papers, which Passed Between the Late Learned Mr. Leibnitz, and Dr. Clarke, in the Years 1715 and 1716: Relating to the Principles of Natural Philosophy and Religion. With an Appendix. To which are Added, Letters to Dr. Clarke Concerning Liberty and Necessity; from a Gentleman of the University of Cambridge: with the Doctor's Answers to Them. Also Remarks Upon a Book, Entitled, A Philosophical Enquiry Concerning Human Liberty. James Knapton, at the Crown in St. Paul's Church Yard., 1717. 3" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 62, + 453, + 295, + 495 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 453, + 295, + 495 + ], + "spans": [ + { + "bbox": [ + 62, + 453, + 295, + 495 + ], + "type": "text", + "content": "[9] Jozef B Cohen and William E Kappauf. Color mixture and fundamental metamers: Theory, algebra, geometry, application. The American journal of psychology, pages 171-259, 1985. 2, 3" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 57, + 498, + 295, + 529 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 498, + 295, + 529 + ], + "spans": [ + { + "bbox": [ + 57, + 498, + 295, + 529 + ], + "type": "text", + "content": "[10] Rafael De Clercq and Leon Horsten. Perceptual indiscriminability: In defence of Wright's proof. The Philosophical Quarterly, 54(216):439-444, 2004. 3" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 57, + 532, + 294, + 553 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 532, + 294, + 553 + ], + "spans": [ + { + "bbox": [ + 57, + 532, + 294, + 553 + ], + "type": "text", + "content": "[11] Lieven Decock and Igor Douven. Similarity after Goodman. Review of philosophy and psychology, 2:61-75, 2011. 2" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 57, + 555, + 295, + 608 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 555, + 295, + 608 + ], + "spans": [ + { + "bbox": [ + 57, + 555, + 295, + 608 + ], + "type": "text", + "content": "[12] Kaize Ding, Albert Jiongqian Liang, Bryan Perrozi, Ting Chen, Ruoxi Wang, Lichan Hong, Ed H. Chi, Huan Liu, and Derek Zhiyuan Cheng. Hyperformer: Learning expressive sparse feature representations via hypergraph transformer, 2023. 2" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 57, + 611, + 294, + 632 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 611, + 294, + 632 + ], + "spans": [ + { + "bbox": [ + 57, + 611, + 294, + 632 + ], + "type": "text", + "content": "[13] Igor Douven. Putting prototypes in place. Cognition, 193: 104007, 2019. 6" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 57, + 635, + 295, + 666 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 635, + 295, + 666 + ], + "spans": [ + { + "bbox": [ + 57, + 635, + 295, + 666 + ], + "type": "text", + "content": "[14] Michael Elad. Sparse and redundant representations: from theory to applications in signal and image processing. Springer Science & Business Media, 2010. 2" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 57, + 669, + 294, + 689 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 669, + 294, + 689 + ], + "spans": [ + { + "bbox": [ + 57, + 669, + 294, + 689 + ], + "type": "text", + "content": "[15] William K Estes. Classification and cognition. Oxford University Press, 1994. 7" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 57, + 692, + 294, + 712 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 692, + 294, + 712 + ], + "spans": [ + { + "bbox": [ + 57, + 692, + 294, + 712 + ], + "type": "text", + "content": "[16] Katalin Farkas. The subject's point of view. OUP Oxford, 2010. 2, 3" + } + ] + } + ], + "index": 16 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 316, + 73, + 555, + 713 + ], + "type": "list", + "angle": 0, + "index": 38, + "blocks": [ + { + "bbox": [ + 316, + 73, + 553, + 106 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 73, + 553, + 106 + ], + "spans": [ + { + "bbox": [ + 316, + 73, + 553, + 106 + ], + "type": "text", + "content": "[17] Alhussein Fawzi, Hamza Fawzi, and Omar Fawzi. Adversarial vulnerability for any classifier. Advances in neural information processing systems, 31, 2018. 2" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 317, + 107, + 553, + 149 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 107, + 553, + 149 + ], + "spans": [ + { + "bbox": [ + 317, + 107, + 553, + 149 + ], + "type": "text", + "content": "[18] Jenelle Feather, Alex Durango, Ray Gonzalez, and Josh McDermott. Metamers of neural networks reveal divergence from human perceptual systems. Advances in Neural Information Processing Systems, 32, 2019. 2, 3, 8, 7" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 317, + 151, + 553, + 171 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 151, + 553, + 171 + ], + "spans": [ + { + "bbox": [ + 317, + 151, + 553, + 171 + ], + "type": "text", + "content": "[19] Gustav Theodor Fechner. Elemente der psychophysik. Breitkopf u. Härtel, 1860. 3" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 317, + 173, + 553, + 194 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 173, + 553, + 194 + ], + "spans": [ + { + "bbox": [ + 317, + 173, + 553, + 194 + ], + "type": "text", + "content": "[20] Gustav Theodor Fechner. Revision der hauptpunkte der Psychophysik. Breitkopf und Härtel, 1882. 3" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 317, + 195, + 553, + 227 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 195, + 553, + 227 + ], + "spans": [ + { + "bbox": [ + 317, + 195, + 553, + 227 + ], + "type": "text", + "content": "[21] Peter Forrest. The Identity of Indiscernibles. In *The Stanford Encyclopedia of Philosophy*. Metaphysics Research Lab, Stanford University, Winter 2020 edition, 2020. 3" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 317, + 228, + 553, + 258 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 228, + 553, + 258 + ], + "spans": [ + { + "bbox": [ + 317, + 228, + 553, + 258 + ], + "type": "text", + "content": "[22] Jeremy Freeman and Eero P Simoncelli. Metamers of the ventral stream. Nature neuroscience, 14(9):1195-1201, 2011. 2, 3" + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 316, + 262, + 555, + 281 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 262, + 555, + 281 + ], + "spans": [ + { + "bbox": [ + 316, + 262, + 555, + 281 + ], + "type": "text", + "content": "[23] Peter Gardenfors. Conceptual spaces: The geometry of thought. MIT press, 2004. 2" + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 317, + 284, + 553, + 326 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 284, + 553, + 326 + ], + "spans": [ + { + "bbox": [ + 317, + 284, + 553, + 326 + ], + "type": "text", + "content": "[24] Amir Globerson and Sam Roweis. Nightmare at test time: robust learning by feature deletion. In Proceedings of the 23rd international conference on Machine learning, pages 353-360, 2006. 2" + } + ] + } + ], + "index": 25 + }, + { + "bbox": [ + 317, + 327, + 553, + 360 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 327, + 553, + 360 + ], + "spans": [ + { + "bbox": [ + 317, + 327, + 553, + 360 + ], + "type": "text", + "content": "[25] Ian J Goodfellow, Jonathon Shlens, and Christian Szegedy. Explaining and harnessing adversarial examples. arXiv preprint arXiv:1412.6572, 2014. 2" + } + ] + } + ], + "index": 26 + }, + { + "bbox": [ + 317, + 361, + 553, + 380 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 361, + 553, + 380 + ], + "spans": [ + { + "bbox": [ + 317, + 361, + 553, + 380 + ], + "type": "text", + "content": "[26] Nelson Goodman. The structure of appearance. Harvard Univ. Press, 1951. 3" + } + ] + } + ], + "index": 27 + }, + { + "bbox": [ + 317, + 383, + 553, + 403 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 383, + 553, + 403 + ], + "spans": [ + { + "bbox": [ + 317, + 383, + 553, + 403 + ], + "type": "text", + "content": "[27] Delia Graff. Phenomenal continua and the sorites. Mind, 110 (440):905-936, 2001. 3" + } + ] + } + ], + "index": 28 + }, + { + "bbox": [ + 317, + 405, + 553, + 425 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 405, + 553, + 425 + ], + "spans": [ + { + "bbox": [ + 317, + 405, + 553, + 425 + ], + "type": "text", + "content": "[28] Hermann Grassmann. Zur theorie der farbenmischung. Annalen der Physik, 165(5):69-84, 1853. 3" + } + ] + } + ], + "index": 29 + }, + { + "bbox": [ + 317, + 427, + 553, + 469 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 427, + 553, + 469 + ], + "spans": [ + { + "bbox": [ + 317, + 427, + 553, + 469 + ], + "type": "text", + "content": "[29] Jamie Hayes and George Danezis. Learning universal adversarial perturbations with generative models. In 2018 IEEE Security and Privacy Workshops (SPW), pages 43-49. IEEE, 2018. 8" + } + ] + } + ], + "index": 30 + }, + { + "bbox": [ + 317, + 472, + 553, + 524 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 472, + 553, + 524 + ], + "spans": [ + { + "bbox": [ + 317, + 472, + 553, + 524 + ], + "type": "text", + "content": "[30] Martin N Hebart, Charles Y Zheng, Francisco Pereira, and Chris I Baker. Revealing the multidimensional mental representations of natural objects underlying human similarity judgements. Nature human behaviour, 4(11):1173-1185, 2020. 2" + } + ] + } + ], + "index": 31 + }, + { + "bbox": [ + 317, + 526, + 553, + 557 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 526, + 553, + 557 + ], + "spans": [ + { + "bbox": [ + 317, + 526, + 553, + 557 + ], + "type": "text", + "content": "[31] Eleanor R Heider. \"Focal\" color areas and the development of color names. Developmental psychology, 4(3):447, 1971. 6" + } + ] + } + ], + "index": 32 + }, + { + "bbox": [ + 317, + 559, + 553, + 579 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 559, + 553, + 579 + ], + "spans": [ + { + "bbox": [ + 317, + 559, + 553, + 579 + ], + "type": "text", + "content": "[32] Benj Hellie. Noise and perceptual indiscriminability. Mind, 114(455):481-508, 2005. 3" + } + ] + } + ], + "index": 33 + }, + { + "bbox": [ + 317, + 582, + 553, + 602 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 582, + 553, + 602 + ], + "spans": [ + { + "bbox": [ + 317, + 582, + 553, + 602 + ], + "type": "text", + "content": "[33] Christopher James Henry. Near sets: theory and applications. University of Manitoba (Canada), 2011. 2" + } + ] + } + ], + "index": 34 + }, + { + "bbox": [ + 317, + 604, + 553, + 635 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 604, + 553, + 635 + ], + "spans": [ + { + "bbox": [ + 317, + 604, + 553, + 635 + ], + "type": "text", + "content": "[34] Felix Hovsepian. A metalogical analysis of vagueness: an exploratory study into the geometry of logic. PhD thesis, University of Warwick, 1992. 1" + } + ] + } + ], + "index": 35 + }, + { + "bbox": [ + 317, + 637, + 553, + 668 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 637, + 553, + 668 + ], + "spans": [ + { + "bbox": [ + 317, + 637, + 553, + 668 + ], + "type": "text", + "content": "[35] Ke Huang and Selin Aviyente. Sparse representation for signal classification. Advances in neural information processing systems, 19, 2006. 2" + } + ] + } + ], + "index": 36 + }, + { + "bbox": [ + 317, + 670, + 553, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 670, + 553, + 713 + ], + "spans": [ + { + "bbox": [ + 317, + 670, + 553, + 713 + ], + "type": "text", + "content": "[36] Andrew Ilyas, Shibani Santurkar, Dimitris Tsipras, Logan Engstrom, Brandon Tran, and Aleksander Madry. Adversarial examples are not bugs, they are features. Advances in neural information processing systems, 32, 2019. 2" + } + ] + } + ], + "index": 37 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "text", + "content": "10252" + } + ] + } + ], + "index": 39 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 8 + }, + { + "para_blocks": [ + { + "bbox": [ + 56, + 72, + 296, + 713 + ], + "type": "list", + "angle": 0, + "index": 17, + "blocks": [ + { + "bbox": [ + 56, + 72, + 294, + 115 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 72, + 294, + 115 + ], + "spans": [ + { + "bbox": [ + 56, + 72, + 294, + 115 + ], + "type": "text", + "content": "[37] Akshay V Jagadeesh and Justin L Gardner. Texture-like representation of objects in human visual cortex. Proceedings of the National Academy of Sciences, 119(17):e2115302119, 2022. 1, 2, 3" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 56, + 117, + 294, + 148 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 117, + 294, + 148 + ], + "spans": [ + { + "bbox": [ + 56, + 117, + 294, + 148 + ], + "type": "text", + "content": "[38] Ralph Kopperman. All topologies come from generalized metrics. The American Mathematical Monthly, 95(2):89-97, 1988. 1" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 56, + 150, + 296, + 183 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 150, + 296, + 183 + ], + "spans": [ + { + "bbox": [ + 56, + 150, + 296, + 183 + ], + "type": "text", + "content": "[39] David H Krantz. Color measurement and color theory: I. representation theorem for grassmann structures. Journal of Mathematical Psychology, 12(3):283-303, 1975. 2, 3, 8" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 56, + 184, + 294, + 205 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 184, + 294, + 205 + ], + "spans": [ + { + "bbox": [ + 56, + 184, + 294, + 205 + ], + "type": "text", + "content": "[40] Dekang Lin et al. An information-theoretic definition of similarity. In ICML, pages 296-304, 1998. 6" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 56, + 206, + 294, + 237 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 206, + 294, + 237 + ], + "spans": [ + { + "bbox": [ + 56, + 206, + 294, + 237 + ], + "type": "text", + "content": "[41] Alexander D Logvinenko. Object-colour manifold. International journal of computer vision, 101:143-160, 2013. 2, 3" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 56, + 239, + 294, + 271 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 239, + 294, + 271 + ], + "spans": [ + { + "bbox": [ + 56, + 239, + 294, + 271 + ], + "type": "text", + "content": "[42] Alexander D. Logvinenko and Eugene Demidenko. On counting metamers. IEEE Transactions on Image Processing, 25(2):770-775, 2016. 2, 3" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 56, + 272, + 294, + 316 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 272, + 294, + 316 + ], + "spans": [ + { + "bbox": [ + 56, + 272, + 294, + 316 + ], + "type": "text", + "content": "[43] Aleksander Madry, Aleksandar Makelov, Ludwig Schmidt, Dimitris Tsipras, and Adrian Vladu. Towards deep learning models resistant to adversarial attacks. arXiv preprint arXiv:1706.06083, 2017. 2" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 56, + 317, + 294, + 370 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 317, + 294, + 370 + ], + "spans": [ + { + "bbox": [ + 56, + 317, + 294, + 370 + ], + "type": "text", + "content": "[44] Saeed Mahloujifar, Dimitrios I Diochnos, and Mohammad Mahmoody. The curse of concentration in robust learning: Evasion and poisoning attacks from concentration of measure. In Proceedings of the AAAI Conference on Artificial Intelligence, pages 4536-4543, 2019. 2" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 56, + 372, + 294, + 403 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 372, + 294, + 403 + ], + "spans": [ + { + "bbox": [ + 56, + 372, + 294, + 403 + ], + "type": "text", + "content": "[45] Douglas L Medin, Robert L Goldstone, and Dedre Gentner. Respects for similarity. Psychological review, 100(2):254, 1993. 2, 6" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 56, + 404, + 294, + 437 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 404, + 294, + 437 + ], + "spans": [ + { + "bbox": [ + 56, + 404, + 294, + 437 + ], + "type": "text", + "content": "[46] Alexius Meinong. Über die Bedeutung des Weber'schen Gesetzes: Beiträge zur Psychologie des Vergleichens und Messens. L. Voss, 1896. 3" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 56, + 437, + 294, + 481 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 437, + 294, + 481 + ], + "spans": [ + { + "bbox": [ + 56, + 437, + 294, + 481 + ], + "type": "text", + "content": "[47] Seyed-Mohsen Moosavi-Dezfooli, Alhussein Fawzi, Omar Fawzi, and Pascal Frossard. Universal adversarial perturbations. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 1765-1773, 2017. 2, 7" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 56, + 482, + 294, + 525 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 482, + 294, + 525 + ], + "spans": [ + { + "bbox": [ + 56, + 482, + 294, + 525 + ], + "type": "text", + "content": "[48] Weili Nie, Brandon Guo, Yujia Huang, Chaowei Xiao, Arash Vahdat, and Anima Anandkumar. Diffusion models for adversarial purification. arXiv preprint arXiv:2205.07460, 2022. 5" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 56, + 526, + 294, + 558 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 526, + 294, + 558 + ], + "spans": [ + { + "bbox": [ + 56, + 526, + 294, + 558 + ], + "type": "text", + "content": "[49] Bruno A Olshausen and David J Field. Sparse coding with an overcomplete basis set: A strategy employed by V1? Vision research, 37(23):3311-3325, 1997. 2" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 56, + 559, + 294, + 602 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 559, + 294, + 602 + ], + "spans": [ + { + "bbox": [ + 56, + 559, + 294, + 602 + ], + "type": "text", + "content": "[50] Marwan Omar, Soohyeon Choi, DaeHun Nyang, and David Mohaisen. Robust natural language processing: Recent advances, challenges, and future directions. IEEE Access, 2022. 5" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 56, + 603, + 294, + 624 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 603, + 294, + 624 + ], + "spans": [ + { + "bbox": [ + 56, + 603, + 294, + 624 + ], + "type": "text", + "content": "[51] Wilhelm Ostwald. Physikalische Farbenlehre. Verlag Unesma, 1919. 2, 3" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 56, + 625, + 294, + 680 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 625, + 294, + 680 + ], + "spans": [ + { + "bbox": [ + 56, + 625, + 294, + 680 + ], + "type": "text", + "content": "[52] Nicolas Papernot, Patrick McDaniel, Somesh Jha, Matt Fredrikson, Z Berkay Celik, and Ananthram Swami. The limitations of deep learning in adversarial settings. In 2016 IEEE European symposium on security and privacy (EuroS&P), pages 372-387. IEEE, 2016. 2, 5" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 56, + 681, + 294, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 681, + 294, + 713 + ], + "spans": [ + { + "bbox": [ + 56, + 681, + 294, + 713 + ], + "type": "text", + "content": "[53] Zdislaw Pawlak. Classification of objects by means of attributes. Polish Academy of Sciences [PAS]. Institute of Computer Science, 1981. 3" + } + ] + } + ], + "index": 16 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 316, + 72, + 555, + 713 + ], + "type": "list", + "angle": 0, + "index": 39, + "blocks": [ + { + "bbox": [ + 316, + 72, + 553, + 106 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 72, + 553, + 106 + ], + "spans": [ + { + "bbox": [ + 316, + 72, + 553, + 106 + ], + "type": "text", + "content": "[54] Charles Pelling. Exactness, inexactness, and the nontransitivity of perceptual indiscriminability. Synthese, 164 (2):289-312, 2008. 3" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 316, + 107, + 553, + 129 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 107, + 553, + 129 + ], + "spans": [ + { + "bbox": [ + 316, + 107, + 553, + 129 + ], + "type": "text", + "content": "[55] Mika Perälä. Aristotle on perceptual discrimination. *Phronesis*, 63(3):257–292, 2018. 1" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 316, + 129, + 553, + 161 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 129, + 553, + 161 + ], + "spans": [ + { + "bbox": [ + 316, + 129, + 553, + 161 + ], + "type": "text", + "content": "[56] James F Peters. Near sets. Special theory about nearness of objects. Fundamenta Informaticae, 75(1-4):407-433, 2007. 2" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 316, + 163, + 553, + 195 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 163, + 553, + 195 + ], + "spans": [ + { + "bbox": [ + 316, + 163, + 553, + 195 + ], + "type": "text", + "content": "[57] James F Peters and Piotr Wasilewski. Tolerance spaces: Origins, theoretical aspects and applications. Information Sciences, 195:211-225, 2012. 1, 2" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 316, + 197, + 544, + 208 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 197, + 544, + 208 + ], + "spans": [ + { + "bbox": [ + 316, + 197, + 544, + 208 + ], + "type": "text", + "content": "[58] Henri Poincaré. Dernières pensées. Flammarion, 1930. 2" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 316, + 209, + 553, + 229 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 209, + 553, + 229 + ], + "spans": [ + { + "bbox": [ + 316, + 209, + 553, + 229 + ], + "type": "text", + "content": "[59] Tim Poston. Fuzzy geometry. PhD thesis, University of Warwick, 1971. 2, 3, 1" + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 316, + 231, + 553, + 253 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 231, + 553, + 253 + ], + "spans": [ + { + "bbox": [ + 316, + 231, + 553, + 253 + ], + "type": "text", + "content": "[60] Diana Raffman. Is perceptual indiscriminability nontransitive? Philosophical Topics, 28(1):153-175, 2000. 2" + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 316, + 254, + 553, + 276 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 254, + 553, + 276 + ], + "spans": [ + { + "bbox": [ + 316, + 254, + 553, + 276 + ], + "type": "text", + "content": "[61] Diana Raffman. Indiscriminability and phenomenal continua. Philosophical perspectives, 26:309-322, 2012. 3" + } + ] + } + ], + "index": 25 + }, + { + "bbox": [ + 316, + 277, + 555, + 298 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 277, + 555, + 298 + ], + "spans": [ + { + "bbox": [ + 316, + 277, + 555, + 298 + ], + "type": "text", + "content": "[62] Fred S Roberts. Tolerance geometry. Notre Dame Journal of Formal Logic, 14(1):68-76, 1973. 2" + } + ] + } + ], + "index": 26 + }, + { + "bbox": [ + 316, + 300, + 553, + 320 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 300, + 553, + 320 + ], + "spans": [ + { + "bbox": [ + 316, + 300, + 553, + 320 + ], + "type": "text", + "content": "[63] Eleanor Rosch. Cognitive reference points. Cognitive Psychology, 7(4):532-547, 1975. 6" + } + ] + } + ], + "index": 27 + }, + { + "bbox": [ + 316, + 322, + 553, + 355 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 322, + 553, + 355 + ], + "spans": [ + { + "bbox": [ + 316, + 322, + 553, + 355 + ], + "type": "text", + "content": "[64] Eleanor H Rosch. On the internal structure of perceptual and semantic categories. In Cognitive development and acquisition of language, pages 111-144. Elsevier, 1973. 6, 7" + } + ] + } + ], + "index": 28 + }, + { + "bbox": [ + 316, + 355, + 553, + 376 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 355, + 553, + 376 + ], + "spans": [ + { + "bbox": [ + 316, + 355, + 553, + 376 + ], + "type": "text", + "content": "[65] Bertrand Russell. The principles of mathematics. Routledge, 2020. 3" + } + ] + } + ], + "index": 29 + }, + { + "bbox": [ + 316, + 378, + 553, + 399 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 378, + 553, + 399 + ], + "spans": [ + { + "bbox": [ + 316, + 378, + 553, + 399 + ], + "type": "text", + "content": "[66] Julius A Schreider. Equality, Resemblance and Order. Mir Publishers, Moscow, 1975. 2, 4" + } + ] + } + ], + "index": 30 + }, + { + "bbox": [ + 316, + 401, + 553, + 434 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 401, + 553, + 434 + ], + "spans": [ + { + "bbox": [ + 316, + 401, + 553, + 434 + ], + "type": "text", + "content": "[67] Ali Shafahi, W Ronny Huang, Christoph Studer, Soheil Feizi, and Tom Goldstein. Are adversarial examples inevitable? arXiv preprint arXiv:1809.02104, 2018. 2" + } + ] + } + ], + "index": 31 + }, + { + "bbox": [ + 316, + 434, + 553, + 456 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 434, + 553, + 456 + ], + "spans": [ + { + "bbox": [ + 316, + 434, + 553, + 456 + ], + "type": "text", + "content": "[68] Herbert A Simon. Models of man; social and rational. Wiley, 1957. 3" + } + ] + } + ], + "index": 32 + }, + { + "bbox": [ + 316, + 457, + 553, + 488 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 457, + 553, + 488 + ], + "spans": [ + { + "bbox": [ + 316, + 457, + 553, + 488 + ], + "type": "text", + "content": "[69] Alexei B Sossinsky. Tolerance space theory and some applications. Acta Applicandae Mathematica, 5:137-167, 1986. 6" + } + ] + } + ], + "index": 33 + }, + { + "bbox": [ + 316, + 491, + 553, + 523 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 491, + 553, + 523 + ], + "spans": [ + { + "bbox": [ + 316, + 491, + 553, + 523 + ], + "type": "text", + "content": "[70] Michael W Spratling. Classification using sparse representations: a biologically plausible approach. Biological cybernetics, 108:61-73, 2014. 2" + } + ] + } + ], + "index": 34 + }, + { + "bbox": [ + 316, + 525, + 553, + 590 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 525, + 553, + 590 + ], + "spans": [ + { + "bbox": [ + 316, + 525, + 553, + 590 + ], + "type": "text", + "content": "[71] Dong Su, Huan Zhang, Hongge Chen, Jinfeng Yi, Pin-Yu Chen, and Yupeng Gao. Is robustness the cost of accuracy? a comprehensive study on the robustness of 18 deep image classification models. In Proceedings of the European Conference on Computer Vision (ECCV), pages 631-648, 2018. 2, 5" + } + ] + } + ], + "index": 35 + }, + { + "bbox": [ + 316, + 591, + 553, + 634 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 591, + 553, + 634 + ], + "spans": [ + { + "bbox": [ + 316, + 591, + 553, + 634 + ], + "type": "text", + "content": "[72] Jeremias Sulam, Ramchandran Muthukumar, and Raman Arora. Adversarial robustness of supervised sparse coding. Advances in Neural Information Processing Systems, 33:2110-2121, 2020. 2" + } + ] + } + ], + "index": 36 + }, + { + "bbox": [ + 316, + 635, + 553, + 679 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 635, + 553, + 679 + ], + "spans": [ + { + "bbox": [ + 316, + 635, + 553, + 679 + ], + "type": "text", + "content": "[73] Christian Szegedy, Wojciech Zaremba, Ilya Sutskever, Joan Bruna, Dumitru Erhan, Ian Goodfellow, and Rob Fergus. Intriguing properties of neural networks. arXiv preprint arXiv:1312.6199, 2013. 1" + } + ] + } + ], + "index": 37 + }, + { + "bbox": [ + 316, + 681, + 553, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 681, + 553, + 713 + ], + "spans": [ + { + "bbox": [ + 316, + 681, + 553, + 713 + ], + "type": "text", + "content": "[74] Thomas Tanay and Lewis Griffin. A boundary tilting perspective on the phenomenon of adversarial examples. arXiv preprint arXiv:1608.07690, 2016. 2" + } + ] + } + ], + "index": 38 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "text", + "content": "10253" + } + ] + } + ], + "index": 40 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 9 + }, + { + "para_blocks": [ + { + "bbox": [ + 56, + 72, + 295, + 660 + ], + "type": "list", + "angle": 0, + "index": 18, + "blocks": [ + { + "bbox": [ + 56, + 72, + 294, + 116 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 72, + 294, + 116 + ], + "spans": [ + { + "bbox": [ + 56, + 72, + 294, + 116 + ], + "type": "text", + "content": "[75] Yee Whye Teh, Max Welling, Simon Osindero, and Geoffrey E Hinton. Energy-based models for sparse overcomplete representations. Journal of Machine Learning Research, 4 (Dec):1235-1260, 2003. 2" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 56, + 118, + 294, + 150 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 118, + 294, + 150 + ], + "spans": [ + { + "bbox": [ + 56, + 118, + 294, + 150 + ], + "type": "text", + "content": "[76] William A Thornton. Matching lights, metamers and human visual response. The Journal of Color and Appearance, 2: 23-29, 1973. 2, 3" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 57, + 152, + 295, + 173 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 152, + 295, + 173 + ], + "spans": [ + { + "bbox": [ + 57, + 152, + 295, + 173 + ], + "type": "text", + "content": "[77] Edward Chace Tolman. *Purposive behavior in animals and men*. Univ of California Press, 1932. 2" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 56, + 175, + 294, + 217 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 175, + 294, + 217 + ], + "spans": [ + { + "bbox": [ + 56, + 175, + 294, + 217 + ], + "type": "text", + "content": "[78] Dimitris Tsipras, Shibani Santurkar, Logan Engstrom, Alexander Turner, and Aleksander Madry. Robustness may be at odds with accuracy. arXiv preprint arXiv:1805.12152, 2018. 2, 5" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 57, + 220, + 294, + 240 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 220, + 294, + 240 + ], + "spans": [ + { + "bbox": [ + 57, + 220, + 294, + 240 + ], + "type": "text", + "content": "[79] Amos Tversky. Features of similarity. Psychological review, 84(4):327, 1977. 2, 6, 7" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 57, + 243, + 294, + 274 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 243, + 294, + 274 + ], + "spans": [ + { + "bbox": [ + 57, + 243, + 294, + 274 + ], + "type": "text", + "content": "[80] Gottfried Wilhelm Freiherr von Leibniz. Leibniz: Discourse on Metaphysics. Open Court Publishing Company, 1902. 3, 2" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 57, + 276, + 294, + 298 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 276, + 294, + 298 + ], + "spans": [ + { + "bbox": [ + 57, + 276, + 294, + 298 + ], + "type": "text", + "content": "[81] Stella Vosniadou and Andrew Ortony. Similarity and analogical reasoning. Cambridge University Press, 1989. 2" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 57, + 300, + 294, + 331 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 300, + 294, + 331 + ], + "spans": [ + { + "bbox": [ + 57, + 300, + 294, + 331 + ], + "type": "text", + "content": "[82] Ernst Heinrich Weber. De Pulsu, resorptione, auditu et tactu: Annotationes anatomicae et physiologicae... CF Koehler, 1831. 3" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 57, + 333, + 294, + 355 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 333, + 294, + 355 + ], + "spans": [ + { + "bbox": [ + 57, + 333, + 294, + 355 + ], + "type": "text", + "content": "[83] Ernst Heinrich Weber. Tastsinn und gemeinschaft. Handworterbuch der Physiologie, 1846. 3" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 57, + 357, + 294, + 377 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 357, + 294, + 377 + ], + "spans": [ + { + "bbox": [ + 57, + 357, + 294, + 377 + ], + "type": "text", + "content": "[84] Timothy Williamson. Identity and Discrimination. Wiley-Blackwell, 1990. 2, 3" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 57, + 379, + 294, + 400 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 379, + 294, + 400 + ], + "spans": [ + { + "bbox": [ + 57, + 379, + 294, + 400 + ], + "type": "text", + "content": "[85] Crispin Wright. On the coherence of vague predicates. Synthese, pages 325-365, 1975. 3" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 57, + 402, + 294, + 434 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 402, + 294, + 434 + ], + "spans": [ + { + "bbox": [ + 57, + 402, + 294, + 434 + ], + "type": "text", + "content": "[86] Günter Wyszecki. Valenzmetrische untersuchung des zusammenhanges zwischen normaler und anomaler trichromasia. 1953. 2, 3" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 57, + 436, + 294, + 468 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 436, + 294, + 468 + ], + "spans": [ + { + "bbox": [ + 57, + 436, + 294, + 468 + ], + "type": "text", + "content": "[87] Gunther Wyszecki and Walter Stanley Stiles. Color science: concepts and methods, quantitative data and formulae. John Wiley & Sons, 2000. 3" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 57, + 470, + 294, + 502 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 470, + 294, + 502 + ], + "spans": [ + { + "bbox": [ + 57, + 470, + 294, + 502 + ], + "type": "text", + "content": "[88] Erik C. Zeeman. The topology of the brain and visual perception. In The Topology of 3-Manifolds. Prentice Hall, NJ, 1962. 2" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 57, + 503, + 294, + 536 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 503, + 294, + 536 + ], + "spans": [ + { + "bbox": [ + 57, + 503, + 294, + 536 + ], + "type": "text", + "content": "[89] E. Christopher Zeeman. Topology of the brain. In Math and Computer Science in Biology and Medicine. H.M. Stationary Office, London, United Kingdom, 1965. 2" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 57, + 537, + 294, + 591 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 537, + 294, + 591 + ], + "spans": [ + { + "bbox": [ + 57, + 537, + 294, + 591 + ], + "type": "text", + "content": "[90] Hongyang Zhang, Yaodong Yu, Jiantao Jiao, Eric Xing, Laurent El Ghaoui, and Michael Jordan. Theoretically principled trade-off between robustness and accuracy. In International conference on machine learning, pages 7472-7482. PMLR, 2019. 2, 5" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 57, + 593, + 294, + 625 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 593, + 294, + 625 + ], + "spans": [ + { + "bbox": [ + 57, + 593, + 294, + 625 + ], + "type": "text", + "content": "[91] Zheng Zhang, Yong Xu, Jian Yang, Xuelong Li, and David Zhang. A survey of sparse representation: algorithms and applications. IEEE access, 3:490-530, 2015. 2" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 57, + 627, + 294, + 660 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 627, + 294, + 660 + ], + "spans": [ + { + "bbox": [ + 57, + 627, + 294, + 660 + ], + "type": "text", + "content": "[92] Li Zhaoping and Keith A May. Psychophysical tests of the hypothesis of a bottom-up saliency map in primary visual cortex. PLoS Computational Biology, 3(4):e62, 2007. 3" + } + ] + } + ], + "index": 17 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "text", + "content": "10254" + } + ] + } + ], + "index": 19 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 10 + } + ], + "_backend": "vlm", + "_version_name": "2.6.4" +} \ No newline at end of file diff --git a/2025/Doppelgangers++_ Improved Visual Disambiguation with Geometric 3D Features/a4009708-9357-4078-b5aa-d3dd7d42ccfe_content_list.json b/2025/Doppelgangers++_ Improved Visual Disambiguation with Geometric 3D Features/a4009708-9357-4078-b5aa-d3dd7d42ccfe_content_list.json new file mode 100644 index 0000000000000000000000000000000000000000..85364bf5dcc3cbc01ff481f25fb1f7ad36334cf8 --- /dev/null +++ b/2025/Doppelgangers++_ Improved Visual Disambiguation with Geometric 3D Features/a4009708-9357-4078-b5aa-d3dd7d42ccfe_content_list.json @@ -0,0 +1,1618 @@ +[ + { + "type": "text", + "text": "Doppelgangers++: Improved Visual Disambiguation with Geometric 3D Features", + "text_level": 1, + "bbox": [ + 91, + 130, + 906, + 154 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Yuanbo Xiangli $^{1}$ Ruojin Cai $^{1}$ Hanyu Chen $^{1}$ Jeffrey Byrne $^{2}$ Noah Snavely $^{1}$ $^{1}$ Cornell University, $^{2}$ Visym Labs", + "bbox": [ + 166, + 179, + 830, + 217 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/e85edf79ca55a027c11e983dc0b5bda38223da24909f140773a47b053f71298d.jpg", + "image_caption": [ + "Ours +MASt3R-SfM" + ], + "image_footnote": [], + "bbox": [ + 140, + 232, + 285, + 296 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/4e5fb53f3d1dcbfeba8393412990b6ff3f14326b3a0753ea848f21bfe1df9771.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 143, + 306, + 285, + 359 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/e4e598dc67129d2fb7db66cfac2036a5526427222bd768459652bcf03d2848b2.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 287, + 231, + 434, + 297 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/a775f643c3ca9e057108e5669dccc344dd60c4ae376da5f03ff7ea558e0df5b8.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 289, + 297, + 434, + 361 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/f27f03a87bdd3540e70c2ef6fd0d9d4cb1596908e4bf0307f96ddbc5067d71ba.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 437, + 231, + 584, + 297 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/d8d43376eb2bf896d3073d92be4a6b4bb91c3f7465b7bc0a37d6dabb87dfc630.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 441, + 297, + 583, + 361 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/9bbd5704ad5623e79dcabb1dc7a6dd141ee34783f30333be0fb46dc74aaca4e1.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 586, + 231, + 733, + 297 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/b4cefdff827cef6d22b8bb2825c6a9d1851e8173c0ff1d20bad5d04f2f39a698.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 588, + 297, + 733, + 361 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/2a29b123581a9ed8c5064120e1e7b2f70829a9df2b3f68657a422658cbb3fa4e.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 736, + 231, + 885, + 297 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/82245e7a93305f40b2317a4271b3bf36d115e47e10a2060cfd43ba16f2511a63.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 738, + 297, + 885, + 361 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/be983518280b036162859f3bf77d472732559cbe51f366de7f5f07024d35107a.jpg", + "image_caption": [ + ":0.235/Ours:0.982" + ], + "image_footnote": [], + "bbox": [ + 192, + 371, + 290, + 470 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/c965f37db1a78807f9458f024cb65e0de1cc2d90526d02366bd5366958e70bc7.jpg", + "image_caption": [ + "GT:1/DG-OG:0.190/Ours:0.971" + ], + "image_footnote": [], + "bbox": [ + 292, + 372, + 437, + 470 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/8e44a4681b9be46df7d208feb798186e47a1f2235dbe54a6ccb05945f1e966c9.jpg", + "image_caption": [ + "GT:1 / DG-OG:0.378 / Ours:0.952" + ], + "image_footnote": [], + "bbox": [ + 439, + 372, + 586, + 470 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/e64a8c415f17ac4ac9028ffa78be166e7b522693c7a24c16cd7f1967857a1b92.jpg", + "image_caption": [ + "GT:0/DG-OG:0.993/Ours:0.025" + ], + "image_footnote": [], + "bbox": [ + 589, + 372, + 733, + 470 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/5c5cbc9397bdc7b929332a81bc456c37df569d27cece026bb5740eb307fb0c37.jpg", + "image_caption": [ + "GT:0/DG-OG:0.994/Ours:0.031" + ], + "image_footnote": [], + "bbox": [ + 736, + 372, + 885, + 470 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/924a8f102cd310d18cf5fcf83ce9b420a38a052ca7731dc5e3fdd37678d905a1.jpg", + "image_caption": [ + "Figure 1. Visual aliasing, or doppelgangers, poses severe challenges to 3D reconstruction. We propose Doppelganger++, an enhanced pairwise image classifier that excels in visual disambiguation across diverse and challenging scenes. (Top) We seamlessly integrate Doppelganger++ into SfM, successfully disambiguating each scene. (Middle) Compared to prior work (which we refer to as DG-OG [3]), Doppelgangers++ is more robust for everyday scenes, showing improved accuracy and robustness. We show pairs that DG-OG classify incorrectly and ours gets correct. (Bottom) Our new VisymScenes dataset, featuring complex daily scenes, is particularly challenging for COLMAP and DG-OG, but our method can achieve correct and complete reconstructions." + ], + "image_footnote": [], + "bbox": [ + 140, + 491, + 410, + 613 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/b53bc3b3b17778f7d7362d72012a622a2214797fd2f2d21ad6eb272cfedace40.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 411, + 491, + 570, + 550 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/84d7daad6b17fa4886892beac441f3543e868663fcea211d1b2ecb1055f228e6.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 411, + 551, + 570, + 611 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/d2f9bc25d12efb4ec378d61525efbd7b1d03b725f414aee84c5ad4351ae296ab.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 571, + 491, + 888, + 612 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Abstract", + "text_level": 1, + "bbox": [ + 248, + 698, + 326, + 713 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Accurate 3D reconstruction is frequently hindered by visual aliasing, where visually similar but distinct surfaces (aka, doppelgangers), are incorrectly matched. These spurious matches distort the structure-from-motion (SfM) process, leading to misplaced model elements and reduced accuracy. Prior efforts addressed this with CNN classifiers trained on curated datasets, but these approaches struggle to generalize across diverse real-world scenes and can require extensive parameter tuning. In this work, we present Doppelgangers++, a method to enhance doppelganger detection and improve 3D reconstruction accuracy. Our contribu", + "bbox": [ + 89, + 734, + 485, + 900 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "tions include a diversified training dataset that incorporates geo-tagged images from everyday scenes to expand robustness beyond landmark-based datasets. We further propose a Transformer-based classifier that leverages 3D-aware features from the MAST3R model, achieving superior precision and recall across both in-domain and out-of-domain tests. Doppelgangers++ integrates seamlessly into standard SfM and MAST3R-SfM pipelines, offering efficiency and adaptability across varied scenes. To evaluate SfM accuracy, we introduce an automated, geotag-based method for validating reconstructed models, eliminating the need for manual inspection. Through extensive experiments, we demonstrate that Doppelgangers++ significantly enhances pairwise vi", + "bbox": [ + 511, + 699, + 908, + 896 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "CVF", + "bbox": [ + 106, + 2, + 181, + 42 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore.", + "bbox": [ + 236, + 0, + 810, + 46 + ], + "page_idx": 0 + }, + { + "type": "aside_text", + "text": "(sneepnep)", + "bbox": [ + 99, + 253, + 130, + 335 + ], + "page_idx": 0 + }, + { + "type": "page_number", + "text": "27166", + "bbox": [ + 478, + 945, + 519, + 955 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "sual disambiguation and improves 3D reconstruction quality in complex and diverse scenarios.", + "bbox": [ + 89, + 90, + 485, + 122 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "1. Introduction", + "text_level": 1, + "bbox": [ + 91, + 150, + 225, + 166 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Visual aliasing—confusing two surfaces that look the same but are nonetheless distinct—is an pernicious problem in 3D reconstruction and SLAM systems. Pairs of images that depict visually similar yet distinct 3D surfaces (called doppelgangers by Cai et al. [3]) can generate spurious correspondence at the feature matching stage of 3D reconstruction, leading to erroneous downstream reconstructions that feature distorted geometry or incorrectly fused elements. Therefore, to ensure the accuracy of 3D reconstruction, it is critical to distinguish truly matching images from illusory matches arising from doppelganger pairs.", + "bbox": [ + 88, + 176, + 485, + 340 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "In recent work, this visual disambiguation problem was formulated as a binary classification task on image pairs [3]. The authors collected a Internet dataset of visually similar doppelganger pairs (as well as truly matching images) from Wikipedia Commons, then trained a CNN to classify image pairs as correct or incorrect matches. This classifier can be used to take a feature match graph computed from a set of photos, remove incorrect edges between doppelganger image pairs, then reconstruct a correct model using a structure from motion (SfM) pipeline like COLMAP [23]. While that work shows promising improvements on reconstruction problems, we find that it can still be brittle: First, the 3D reconstruction task demands that the classifier have has precision—even a few bad edges remaining in the image match graph can lead to an incorrect 3D model. Second, the reconstruction task is quite sensitive to the threshold on the classifier score used to prune edges from the match graph. Finally, their method was trained solely on landmark Internet photos and does not reliably generalize to new scenarios, such as more structured captures of everyday scenes, like streets and office buildings.", + "bbox": [ + 91, + 342, + 485, + 643 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "In this paper, we aim to address these issues and improve doppelganger classification in several key ways:", + "bbox": [ + 89, + 643, + 483, + 674 + ], + "page_idx": 1 + }, + { + "type": "list", + "sub_type": "text", + "list_items": [ + "1. We identify ways to expand and diversify doppelgangers training data. In particular, we leverage semi-structured image data with geographic annotations (rough GPS position and orientation), captured from everyday scenes with the Visym Collector platform [2].", + "2. We switch from a CNN-based classifier to leveraging features from MASt3R [7], a recent transformer-based geometric model that computes point clouds from two input views. Specifically, we feed an image pair into a pre-trained MASt3R model, collect the intermediate features decoded by MASt3R, and train a classification head to map these features to a doppelganger score." + ], + "bbox": [ + 89, + 674, + 483, + 854 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "We call our method Doppelgangers++. Our enhanced model achieves higher precision and recall in pairwise classification, and generalizes better across a broader range of scenes and", + "bbox": [ + 89, + 854, + 483, + 900 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "capture scenarios. Doppelganger++ integrates seamlessly into SfM pipelines for 3D reconstruction disambiguation. We further propose to leverage geo-tagged map images to quantitatively evaluate the correctness and completeness of reconstructed models, in comparison to manual inspection as required in previous approaches. Through extensive experiments we show that our model leads to more accurate and complete reconstruction results, and is less sensitive to the threshold used for pruning doppelganger matches.", + "bbox": [ + 511, + 90, + 908, + 229 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "2. Related Work", + "text_level": 1, + "bbox": [ + 511, + 243, + 658, + 258 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Local feature matching and 3D learning. Local feature matching methods have proven effective at establishing correspondences between pairs of images in SfM pipelines. Classic methods like COLMAP [23] rely on the tried-and-true SIFT features [17] to find correspondence. Modern learning-based, data-driven approaches have improved the quality of local feature matching [6, 21, 25, 33]. More recently, the DUSt3R [29] framework has proven to excel in a variety of 3D reconstruction tasks by estimating dense 3D point clouds from pairs of images. MASt3R [16] was proposed as an extension to DUSt3R that specifically targets the task of predicting dense feature matches between image pairs. All of these local feature matching methods are traditionally optimized to maximize the number of correspondences they find between similar image regions. While their ability to identify feature matches has greatly improved, they generally lack the ability to incorporate negative evidence into their predictions, and so they often find matches between regions that do not actually correspond to the same 3D surface, particularly within doppelganger image pairs. However, we show that features internal to MASt3R can be repurposed for doppelganger detection.", + "bbox": [ + 511, + 270, + 908, + 603 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Disambiguation in SfM and image matching. Disambiguating similar structures and repeated patterns in SfM is a long-standing challenge. Prior work has mainly relied on heuristics-based analysis to detect conflicting relations and ambiguities in the structure of the underlying scene graph [10, 14, 18, 19, 24, 27, 31, 32, 35] or among image-level correspondence (or the lack thereof) [4, 11-13, 20, 34]. While such methods have shown some success in detecting and resolving ambiguities in SfM pipelines, one fundamental limitation that they do not consider the rich information contained in the images themselves. In contrast, Doppelgangers [3] avoids handcrafted heuristics and takes a data-driven approach to the basic problem of identifying doppelganger image pairs. They train a CNN that classifies a pair of images as either positive or negative, with LoFTR [25]-extracted keypoints and match masks passed as auxiliary input. During inference, the binary classification results are used to pre-process the scene graph obtained from COLMAP to filter out spurious matches prior to running SfM. How", + "bbox": [ + 511, + 613, + 910, + 901 + ], + "page_idx": 1 + }, + { + "type": "page_number", + "text": "27167", + "bbox": [ + 478, + 944, + 519, + 957 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "ever, the method's reliance on COLMAP feature extraction and matching and the need for auxiliary mask information from LoFTR introduces significant overhead and complexity to the overall pipeline, making it difficult to scale to larger scenes. In addition, we find that this prior method is brittle: it can fail to generalize to domains beyond landmarks (e.g., to office buildings), and it can require parameter tuning to get good results.", + "bbox": [ + 89, + 90, + 480, + 210 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Differentiable SfM. To improve upon traditional SfM, several differentiable SfM pipelines have been proposed to optimize the entire 3D reconstruction process. These pipelines often include one or more components that are trained from large datasets. MAST3R-SfM [7] proposes an SfM pipeline that uses MAST3R features for scene graph construction and coarse-to-fine alignment. VGGSfM [28] decomposes the problem into four stages: point tracking, initial camera estimation, triangulation, and bundle adjustment, each of which is differentiable and can be optimized in an end-to-end manner. ACE0 [1] uses a scene coordinate regression network to progressively register new images to a global scene, while the network itself is iteratively trained and refined from the registered images. These methods have shown promising results in optimizing the SfM pipeline, but are not specifically designed to address the doppelganger problem, and are still prone to producing incorrect reconstructions when faced with scenes with repeated patterns or similar structures.", + "bbox": [ + 91, + 220, + 486, + 492 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3. Method", + "text_level": 1, + "bbox": [ + 89, + 508, + 181, + 523 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "We refer to a \"doppelganger\" as a case where distinct objects or surfaces look almost identical, leading algorithms to confuse one for the other [3]. In this work, we aim to improve the doppelganger classifier via expanded and diversified training data (Sec. 3.1), and leveraging geometric 3D features learned from pairwise reconstruction [16] (Sec. 3.2). We further propose an approach to quantitatively evaluate the correctness of SfM results in terms of doppelgangers (Sec. 3.3), instead of manual inspection adopted in previous work.", + "bbox": [ + 89, + 534, + 482, + 672 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3.1. The VisymScenes dataset", + "text_level": 1, + "bbox": [ + 89, + 681, + 318, + 696 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Cai et al. [3] introduced the Doppelganger dataset, built on global landmarks photos sourced from Wikipedia Commons, with viewing direction (e.g., North, South) to identify doppelganger image pairs. While the classifier trained on this dataset proved effective, we find that it struggles to generalize well to scenes beyond the dataset's domain.", + "bbox": [ + 89, + 703, + 482, + 794 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "We enhance and expand their training set to improve the robustness of doppelganger classification, by incorporating casually captured images from a wider range of scenes. To ensure sufficient diversity in the dataset, we introduce the VisymScenes dataset. VisymScenes consists of 258K ground images with GPS/IMU metadata, recorded at 149 sites in 42 cities and 15 countries, collected with the Visym Collector", + "bbox": [ + 89, + 795, + 482, + 900 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Platform [2] (details in supplementary). Each image is accompanied by metadata, such as GPS coordinates, device compass direction, and intrinsic camera calibration. While this metadata can be noisy, it still provides valuable information for identifying potential doppelganger pairs. For example, if two images exhibit numerous geometrically consistent local feature matches yet were taken from distant locations, this serves as strong evidence that the pair is likely to be a doppelganger. Examples are shown in Fig. 2.", + "bbox": [ + 511, + 90, + 906, + 226 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "We develop a series of filtering rules applied to all matched image pairs identified by the COLMAP feature matching module [23]. Note that these rules are designed according to the capture style of the Visym Collector, where cameras consistently focus on the scene of interest and maintain a reasonable distance from it. For a given pair of images, we use the (metric) distance between the camera centers $r$ , angle between their viewing directions $\\theta$ , and camera intrinsics $\\mathcal{K}$ , all derived from camera metadata, to identify confident negative (doppelganger) and positive (correct match) pairs. To identify confident negative pairs, we first identify spatially distant matching pairs, and classify them as doppelgangers. Then, we classify the remaining (nearby) pairs into three cases: the intersection point of the viewing directions is 1) in front of both cameras, 2) behind both cameras; or 3) in front of one camera and behind the other. For case 1), we label pairs with very large view angles (e.g., $>160^{\\circ}$ ) as negative, since they likely capture different 3D surfaces or mostly non-overlapping content. For case 2), if their view angle exceeds the diagonal field of view, then their view frustums are unlikely to overlap, so we label them as doppelgangers. For case 3), we check for frustum overlap using camera intrinsics; if no intersection is detected, the pair is considered a doppelganger. A similar series of rules are designed for mining positive pairs (i.e., true matches). Details of computing view intersection and filtering algorithms for both negative and positive pairs can be found in the supplementary. With these rules, we mined in total 53K positive and negative pairs across 33 sites for our doppelganger task.", + "bbox": [ + 511, + 228, + 908, + 667 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3.2. Improved Doppelganger Classifier", + "text_level": 1, + "bbox": [ + 511, + 680, + 815, + 696 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Recent advances in data-driven models [16, 29] have demonstrated impressive results in geometric vision tasks. In our work, we leverage the multi-level, 3D-aware features extracted from the MAST3R model [16] to train a doppelganger classifier on labeled image pairs, as demonstrated in Fig. 3.", + "bbox": [ + 511, + 703, + 906, + 777 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Equipped with large-scale training data and a ViT backbone, the MASt3R model captures an image representation that encodes rich 3D geometric information between paired images. However, as MASt3R was originally trained to detect correspondences and similarities, its point maps often conflate doppelganger pairs, yielding incorrect point clouds and poses for these challenging pairs (examples in Fig. 7). Despite this limitation, we find that the internal", + "bbox": [ + 511, + 780, + 908, + 900 + ], + "page_idx": 2 + }, + { + "type": "page_number", + "text": "27168", + "bbox": [ + 478, + 944, + 517, + 955 + ], + "page_idx": 2 + }, + { + "type": "image", + "img_path": "images/bb0850eb322419a465b9ad04e192ebca37cf0e16bf5b0bad9472f617b33cd9d7.jpg", + "image_caption": [ + "Figure 2. VisymScenes examples. This new dataset includes residential areas, landmarks, historical sites, business districts, and more. Here, we present four example sites. The top row shows subsets of images captured within each site. The bottom row displays pairs of visually similar but geographically distinct images from each site along with their recorded geolocations on a map. These examples demonstrate that doppelganger issues are prevalent in everyday scenes, presenting significant challenges for reliable 3D reconstruction and image matching." + ], + "image_footnote": [], + "bbox": [ + 96, + 89, + 297, + 217 + ], + "page_idx": 3 + }, + { + "type": "image", + "img_path": "images/deed301b54334765ad2b9ba161e9c30a6f3c597194057601f07c0103a4e974ff.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 297, + 90, + 493, + 215 + ], + "page_idx": 3 + }, + { + "type": "image", + "img_path": "images/58cd38fd1978cd5e8d91d604f59f38c8dddfbebcb144dbd2172707c65e7e708b.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 500, + 90, + 696, + 215 + ], + "page_idx": 3 + }, + { + "type": "image", + "img_path": "images/a9ec786564a9c2398f43f04595dccc55d98bd1dffabd857bad716672c03cc9ee.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 699, + 90, + 898, + 215 + ], + "page_idx": 3 + }, + { + "type": "image", + "img_path": "images/2f09ee2e9191f482c02336660540159a793554b71811d74fd1d530dae681cd7c.jpg", + "image_caption": [ + "Figure 3. Model design. (Left) Given an image pair, we first create a symmetrized version of the pair and feed it into the frozen MASt3R model. Multi-layer features are extracted from each decoder branch, concatenated, and fed into two learnable doppelganger classification heads. Each head generates predictions $(\\mathrm{pred}_{pq}^{v}, \\mathrm{pred}_{qp}^{v})$ , $v \\in \\{1,2\\}$ (where $pq$ and $qp$ denote the symmetrized image pair), supervised by cross-entropy loss. (Right) We use multi-layer decoder features and a Transformer-based classifier head for doppelganger prediction." + ], + "image_footnote": [], + "bbox": [ + 94, + 292, + 903, + 398 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "features learned by MASt3R contain sufficient information for doppelganger classification without the need to fine-tune MASt3R's weights. Moreover, given that the amount of available training data for the doppelganger task is dwarfed by the amount of data used to train MASt3R, we opt not to fine-tune the entire model, but instead repurpose these internal features by training an additional output head to predict a doppelganger classification score. In Sec. 4.5, we empirically demonstrate that this choice achieves comparable or better visual disambiguation performance, especially on out-of-domain scenes.", + "bbox": [ + 88, + 470, + 483, + 636 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Multi-level decoder features. For an image pair $(I_p, I_q)$ , we extract multi-level decoder features from MASt3R's decoder branches. MASt3R takes two images as input, and uses two intertwined decoders $\\mathrm{DecBlk}^1$ and $\\mathrm{DecBlk}^2$ to decode encoded features $\\mathcal{H}^1 = \\mathrm{Enc}^1(I_p)$ and $\\mathcal{H}^2 = \\mathrm{Enc}^2(I_q)$ . Each decoder has $B$ blocks, and attends to tokens from the other branch:", + "bbox": [ + 89, + 643, + 483, + 748 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\nf _ {i} ^ {1} = \\operatorname {D e c B l k} _ {i} ^ {1} \\left(f _ {i - 1} ^ {1}, f _ {i - 1} ^ {2}\\right), f _ {i} ^ {2} = \\operatorname {D e c B l k} _ {i} ^ {2} \\left(f _ {i - 1} ^ {2}, f _ {i - 1} ^ {1}\\right), \\tag {1}\n$$\n", + "text_format": "latex", + "bbox": [ + 101, + 753, + 483, + 773 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "where $f_{i}^{v}$ denotes the output tokens from the $i$ -th block of the $v$ -th branch. These two branches exchange information to capture the spatial relationships between viewpoints and the global 3D structure of the scene. For each branch $v$ , we concatenate the encoder feature $\\mathcal{H}^v$ and the tokens from the decoder blocks into $\\mathcal{F}^v = \\mathrm{concat}(\\mathcal{H}^v, \\{f_i^v\\}_{i=0}^{B-1})$ , which captures rich, multi-level spatial correspondence information between image pairs.", + "bbox": [ + 89, + 779, + 483, + 901 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Transformer-based classification heads. MAsst3R treats one image as the reference frame and projects the other image into that reference frame. A consequence of this design is that the reconstructed 3D scene for input pair $(I_p, I_q)$ is distinct (or at least, must be in a different coordinate frame) from that of $(I_q, I_p)$ . Inspired by the asymmetric design, we propose to 1) use separate Transformer heads: We introduce two independent, Transformer-based classification heads $\\mathrm{Head}_{dopp}^1$ and $\\mathrm{Head}_{dopp}^2$ to process $\\mathcal{F}^1$ and $\\mathcal{F}^2$ respectively; and 2) symmetrize image pairs such that the model evaluates both $(I_p, I_q)$ and $(I_q, I_p)$ to decide whether the given pair of images are true match or not. Thus, we end up with four scores for each image pair:", + "bbox": [ + 511, + 470, + 908, + 669 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\operatorname {p r e d} _ {u} ^ {v} = \\operatorname {H e a d} _ {\\text {d o p p}} ^ {v} \\left(\\mathcal {F} _ {u} ^ {v}\\right), u \\in \\{p q, q p \\}, v \\in \\{1, 2 \\}. \\tag {2}\n$$\n", + "text_format": "latex", + "bbox": [ + 532, + 680, + 906, + 699 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Essentially, the two heads serve as expert evaluators, each examining how likely the image pair is to be a true match (or doppelganger) from different aspects, and switching the input order allows the model to analyze the spatial relationships from both directions. We empirically show the effectiveness of this design in Sec. 4.5. Both $\\mathrm{Head}_{dopp}^{1}$ and $\\mathrm{Head}_{dopp}^{2}$ are supervised by the cross-entropy loss, encouraging $S = \\{\\mathrm{pred}_u^v\\}$ to all give high probabilities for positive matches and low probabilities for negative ones.", + "bbox": [ + 511, + 709, + 908, + 847 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Test-time voting. At test time, we combine the four scores $S$ via a voting mechanism to make a final decision. Specifically, if the majority of the heads predict that the pair is a positive", + "bbox": [ + 511, + 854, + 908, + 901 + ], + "page_idx": 3 + }, + { + "type": "page_number", + "text": "27169", + "bbox": [ + 478, + 944, + 519, + 957 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "match, we take $\\max(S)$ as the final score for the image pair. Conversely, if the majority of the heads vote for a negative match, $\\min(S)$ is used as the final score. Otherwise, we average the scores from the four heads. This approach ensures that the final decision reflects the strongest evidence supporting the consensus among the heads and thereby improves the reliability of the classifier. Eq. 3 elaborates on this voting mechanism:", + "bbox": [ + 89, + 90, + 483, + 210 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\nS _ {\\text {f i n a l}} = \\left\\{ \\begin{array}{l l} \\max (\\mathcal {S}) & \\text {i f} \\sum_ {s \\in \\mathcal {S}} \\mathbf {1} (s > 0. 5) > \\sum_ {s \\in \\mathcal {S}} \\mathbf {1} (s < 0. 5), \\\\ \\min (\\mathcal {S}) & \\text {i f} \\sum_ {s \\in \\mathcal {S}} \\mathbf {1} (s > 0. 5) < \\sum_ {s \\in \\mathcal {S}} \\mathbf {1} (s < 0. 5), \\\\ \\operatorname {m e a n} (\\mathcal {S}) & \\text {i f} \\sum_ {s \\in \\mathcal {S}} \\mathbf {1} (s > 0. 5) = \\sum_ {s \\in \\mathcal {S}} \\mathbf {1} (s < 0. 5). \\end{array} \\right. \\tag {3}\n$$\n", + "text_format": "latex", + "bbox": [ + 99, + 214, + 483, + 268 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "3.3. Evaluating Doppelganger correction in SfM", + "text_level": 1, + "bbox": [ + 89, + 277, + 464, + 292 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "There is currently no reliable benchmarking method for evaluating the accuracy and correctness of SfM reconstructions specifically in terms of how well they address the doppelganger issue. Prior work [3] relied on manual inspection of each out model to assess the effectiveness of their approach—a time-consuming, unquantifiable, and potentially error-prone process. In our work, we propose a benchmarking method for qualitatively evaluating reconstructed models with respect to doppelgangers. We leverage mapping sites like Mapillary, which provide images with location metadata that can serve as probes for validating a 3D model. Note that our method targets common scenarios where geotags are unavailable, or so noisy as to not be useful. Therefore, we do not consider the use of geotags for the reconstruction task itself—we explore pure visual disambiguation—but instead gather them from specialized sources for ground truth evaluation.", + "bbox": [ + 89, + 299, + 483, + 555 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Specifically, for a given scene with known geolocation, we acquire from Mapillary a set of nearby geo-tagged images and register them to the reconstructed model with COLMAP [23]. We then apply RANSAC to robustly estimate a similarity transformation between the registered camera positions and their corresponding image geolocation metadata (converted to ECEF coordinates). We use the resulting RANSAC inlier ratio as an indicator of the model's correctness. As an example, in Fig. 4, we can observe that before correction, the registered cameras and their geolocations show significant misalignment, with the cameras collapsing to one side due to similar-looking surfaces. After correction, the registered cameras align closely to their true geolocations, indicating that doppelganger pairs are correctly separated in the model. If the model is split into multiple components, we calculate the weighted average of inliers ratios among the components:", + "bbox": [ + 89, + 556, + 483, + 813 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\operatorname {I R} = \\sum_ {i = 1} ^ {N} \\left(\\frac {I _ {i}}{T _ {i}} \\cdot \\frac {T _ {i}}{\\sum_ {j = 1} ^ {N} T _ {j}}\\right) = \\frac {\\sum_ {i = 1} ^ {N} I _ {i}}{\\sum_ {i = 1} ^ {N} T _ {i}}, \\tag {4}\n$$\n", + "text_format": "latex", + "bbox": [ + 147, + 821, + 483, + 863 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "where $I_{i}$ is the number of RANSAC inliers and $T_{i}$ is the number of registered Mapillary images in the $i$ -th compo", + "bbox": [ + 89, + 869, + 483, + 901 + ], + "page_idx": 4 + }, + { + "type": "image", + "img_path": "images/62b80a4d91920e948489fbc03315ecd65e5a6cf816cfbc816692fd5ebf430d4e.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 535, + 90, + 880, + 188 + ], + "page_idx": 4 + }, + { + "type": "image", + "img_path": "images/91870a5b6242f432624630df8ba963a4b79e7a1846cd3951dd4191b06f7075c8.jpg", + "image_caption": [ + "Figure 4. Evaluation of doppelganger correction in SfM. (Top) We first collect sequences of geo-tagged Mapillary images around the target location and register them to the SfM model. Then, we use RANSAC to align the registered cameras and their geolocations. The inlier ratio is computed as an indicator of model accuracy. (Bottom) In the model corrupted by doppelganger pairs, the registered cameras all collapse to one side. We see that the camera poses estimated with COLMAP (right, in red) do not align well with the geotags (green), leading to a low inlier ratio, but our method leads to a much closer alignment." + ], + "image_footnote": [], + "bbox": [ + 535, + 190, + 883, + 373 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "nent. While alignment error could also serve as a metric, we opt to use the inlier ratio to reduce sensitivity to geolocation inaccuracies, as image geolocations may not always be precise. This strategy can also be used to detect broken reconstructions from datasets like MegaScenes [26], which contains over 100K SfM results from Internet photos around the world.", + "bbox": [ + 511, + 529, + 906, + 633 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "4. Experiments", + "text_level": 1, + "bbox": [ + 511, + 652, + 648, + 669 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "4.1. Experimental setup", + "text_level": 1, + "bbox": [ + 511, + 679, + 700, + 695 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Dataset. Our training set is comprised of the Doppelganger dataset [3] (dubbed DG) and the new VisymScenes dataset. Specifically, the DG training set includes 73K image pairs, nearly evenly split between positive and negative pairs. From VisymScenes, we use 47K image pairs across 26 sites as training data, evenly split into positive and negative pairs.", + "bbox": [ + 511, + 702, + 906, + 792 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "We evaluate on two tasks, 1) pairwise image visual disambiguation and 2) scene reconstruction by integrating our classifier into COLMAP [23]. For pairwise image visual disambiguation, we utilize the DG test set, and mined an additional 3,180 pairs from 7 Visym sites (distinct from the 26 training sites), equally divided into positive and negative pairs. Additionally, we created a test set from the Mapillary", + "bbox": [ + 511, + 795, + 908, + 902 + ], + "page_idx": 4 + }, + { + "type": "page_number", + "text": "27170", + "bbox": [ + 478, + 944, + 519, + 957 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Street-Level Sequences dataset [30]. This Mapillary dataset spans diverse urban and suburban environments and captures a wide range of seasons, weather conditions, cameras, lighting, and structural settings, with each image geo-tagged by GPS and compass angle. Using a similar filtering approach to the VisymScenes dataset, we mined 1,500 positive and 1,500 negative pairs as an out-of-domain test set.", + "bbox": [ + 89, + 90, + 480, + 196 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "For the scene reconstruction task, we evaluate on 16 scenes sampled from Heinly et al. [12], Wilson et al. [31], MegaScenes [26] and 5 VisymScenes test scenes. These scenes are challenging for conventional SfM pipelines as well as prior disambiguation methods due to subtle differences between distinct surfaces and repetitive patterns.", + "bbox": [ + 89, + 196, + 480, + 287 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Metrics. For pairwise visual disambiguation evaluation, we report Average Precision (AP) and ROC AUC scores, following the protocol in [3]. Additionally, we report precision at a fixed recall and recall at a fixed precision. These statistics help characterize a model's trade-off between achieving high precision and potentially missing true positives.", + "bbox": [ + 89, + 292, + 480, + 383 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "To comprehensively evaluate SfM reconstructions, we report the number of images in the SfM results, and the geo-alignment inlier ratio described in Sec. 3.3. Respectively, more images in the SfM model imply less false pair pruning, i.e., better class separation; and a higher geo-alignment inlier ratio suggests that the reconstructed model is more likely to be accurate and complete. Note that the inlier ratio is also influenced by the accuracy of the geo-tagged images collected from Mapillary. Thus, if one test scene has more accurate geo-tagged images than another, it is likely to yield a higher inlier ratio. Therefore, rather than comparing inlier ratios across different scenes, this metric is more useful for evaluating different reconstruction methods on the same scene, as the same set of geo-tagged images is used.", + "bbox": [ + 89, + 383, + 482, + 594 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Baselines. Our primary baseline is [3], as their learning-based approach consistently outperforms heuristic-based methods like thresholding on the number of post-RANSAC inliers (e.g. SIFT [17] + RANSAC [9], D2-Net [8] + RANSAC, SuperPoint [6] + SuperGlue [22]) for pairwise visual disambiguation; and is also superior than previous approaches [5, 12, 31, 32] for SfM disambiguation.", + "bbox": [ + 89, + 599, + 482, + 705 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "4.2. Implementation", + "text_level": 1, + "bbox": [ + 89, + 712, + 248, + 728 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Model details. Our doppelganger classifier head $\\mathrm{Head}_{\\mathrm{dopp}}$ is implemented with a Transformer encoder comprised of 3 layers, each with input and output dimension of 768. Each layer has 8 attention heads and a two-layer feed-forward network with a hidden dimension of 2048. Pre-layer normalization and residual connections are applied within each layer. The transformed tokens are aggregated via max pooling and linearly projected into $\\mathrm{pred}^v$ . We freeze the weights of the public MAST3R [16] model and only train the two doppelganger classification heads, supervised by a cross-entropy loss. We train for 5 epochs with a batch size of 8", + "bbox": [ + 89, + 734, + 482, + 900 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "using Adam [15] with a learning rate of $1 \\times 10^{-4}$ . More details can be found in the supplementary.", + "bbox": [ + 511, + 90, + 903, + 121 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Integration with SfM. Following [3], we integrate Doppelgangers++ into SfM to enhance its disambiguation ability. SfM takes a collection of images $\\mathcal{I} = \\{I_i\\}_{i=0}^n$ and generates geometrically verified image pairs $\\mathcal{P} = \\{(I_p, I_q)\\}$ , forming a scene graph $\\mathcal{G} = (\\mathcal{I}, \\mathcal{P})$ with images as nodes and pairs as edges. Using $\\mathcal{G}$ , SfM computes camera poses and reconstructs a 3D point cloud. Our doppelganger classifier acts as a filter on edges in $\\mathcal{G}$ , removing pairs below a probability threshold $\\tau$ to eliminate spurious matches due to repeated or symmetric structures. We integrate Doppelgangers++ into both COLMAP [23] and MASt3R-SfM [7] to showcase the effectiveness of the method. Notably, with COLMAP, SIFT features are computed for scene graph construction and MASt3R features for pruning; whereas with MASt3R-SfM, MASt3R features serve for both scene graph construction and pruning, forming a more efficient pipeline.", + "bbox": [ + 511, + 128, + 906, + 371 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "4.3. Pairwise Visual Disambiguation", + "text_level": 1, + "bbox": [ + 511, + 380, + 794, + 396 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "We compare our approach with the original Doppelganger work [3] (dubbed $DG-OG$ ) under two experimental settings: 1) we use the same training data as in [3], i.e. the DG training set, to train our model. Both models are evaluated on three benchmark test sets: DG, VisymScenes, and Mapillary. Under this configuration, the DG test set is in-domain, while VisymScenes and Mapillary are out-of-domain scenarios; 2) we expand the training data by including the new VisymScenes training set and retrain both DG-OG and our model. Evaluation is conducted on the same three test sets, with only the Mapillary test set being out-of-domain this time.", + "bbox": [ + 511, + 402, + 906, + 568 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Quantitative results are in Tab. 1. Under the first setting, where both methods are trained on DG, our model shows clear improvements across all three test sets. Specifically, on the in-domain DG test set, our model achieves higher AP and ROC AUC, indicating that it maintains high precision across all recall levels, while also being less sensitive to the decision threshold $\\tau$ with better class separation. On the out-of-domain test sets (VisymScenes and Mapillary), we observe $25\\%$ to $65\\%$ improvements in both AP and ROC AUC, highlighting the generalizability of our method.", + "bbox": [ + 511, + 568, + 905, + 718 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Although our model demonstrates improved performance even when trained solely on the DG dataset, its precision outside of the training domain is suboptimal for practical use in SfM, where high precision is essential. In the second setting, we expand the training set by including VisymScenes data. Both methods maintain similar performance on the DG test set, with our model experiencing a slight drop in recall when precision is set to 0.99. On the VisymScenes test set, both models show improvements across all metrics, with ours reaching $99\\%$ in both AP and ROC AUC. Notably, adding VisymScenes to the training set does not enhance DG-OG's performance on the out-of-domain Mapillary test", + "bbox": [ + 511, + 719, + 906, + 900 + ], + "page_idx": 5 + }, + { + "type": "page_number", + "text": "27171", + "bbox": [ + 478, + 944, + 517, + 955 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/4f0612e82fbcaa2337ccce1cdf0c08c41df84037639aa9dfccb0179ee5051fb2.jpg", + "image_caption": [ + "Figure 5. SfM Disambiguation on MegaScenes [26]. (White background) SfM results from DG-OG [3] and ours. (Black background) Verification using geo-tagged images, red points represent registered cameras and green points represent geolocations, inlier ratio (IR) is labeled on the bottom right. DG-OG fails to disambiguate this scene, predicting incorrect scores for image pairs. Our method correctly splits the model into two clean components." + ], + "image_footnote": [], + "bbox": [ + 96, + 88, + 300, + 282 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/5ef5e6ed4c8f5c689e62b63d73aa78bb99b183eaa71deaf1c989fc812d448771.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 313, + 88, + 483, + 282 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/56284431ec2eb48111d90ea0674c2dd93c3616c37e3d1cc1fb00f06f8571dc78.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 493, + 88, + 903, + 282 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/e25ab08f7ec9f681f89a95e8c932bc732295996420d7fbc44e3795fef35161d0.jpg", + "image_caption": [ + "Figure 6. SfM disambiguation on VisymScenes. We show that our classifier is more robust than DG-OG [3] on test scenes from new domains, like everyday street scenes. DG-OG has difficulty disambiguating such scenes, leading to incorrect geometry." + ], + "image_footnote": [], + "bbox": [ + 91, + 345, + 480, + 496 + ], + "page_idx": 6 + }, + { + "type": "table", + "img_path": "images/3ac1602181dcb48c73917e2775afdd1b1703fb4d60406debdf259f41d644ee20.jpg", + "table_caption": [], + "table_footnote": [], + "table_body": "
Test DataMethodMetrics (trained on DG / trained on DG + VisymScenes)
AP↑ROC AUC↑Prec@Recall=0.85↑Recall@Prec=0.99↑
DGDG-OG0.954 / 0.9560.944 / 0.9470.901 / 0.9100.611 / 0.614
Ours0.980 / 0.9810.981 / 0.9810.972 / 0.9820.690 / 0.642
VisymScenesDG-OG0.816 / 0.9380.726 / 0.9210.498 / 0.8310.340 / 0.623
Ours0.936 / 0.9910.909 / 0.9900.892 / 0.9990.542 / 0.901
MapillaryDG-OG0.566 / 0.6920.581 / 0.7010.523 / 0.5720.003 / 0.000
Ours0.950 / 0.9680.944 / 0.9580.927 / 0.9420.310 / 0.736
", + "bbox": [ + 93, + 563, + 482, + 651 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Table 1. Evaluation of pairwise disambiguation. We evaluate DG-OG [3] and our method trained on DG [3] only and DG +VisymScenes (two numbers per cell) on three test sets. Both DG-OG and ours benefit from dataset expansion, whereas ours gained more generalizability on out-of-domain test set (Mapillary) after training on both. Our classifier constantly demonstrates better precision, recall across all settings.", + "bbox": [ + 89, + 656, + 483, + 753 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "set, while ours continues to benefit from increased training diversity. These results indicate that the prior approach struggles to generalize effectively, whereas ours shows greater generalization with more varied training data.", + "bbox": [ + 89, + 762, + 483, + 824 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "4.4. Structure from Motion disambiguation", + "text_level": 1, + "bbox": [ + 89, + 832, + 426, + 849 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "We integrate our classifier trained on DG and VisymScenes into COLMAP's SfM pipeline [23], and evaluate its performance on reconstructing scenes with duplicated and symmet", + "bbox": [ + 89, + 854, + 485, + 900 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "ric structures. We compare the results with vanilla COLMAP and Cai et al. [3] (DG-OG). Quantitative results are presented in Tab. 2. Our approach registers more images across all scenes than DG-OG and operates without threshold tuning. In contrast, DG-OG relies on scene-specific thresholds, such as $\\tau = 0.97$ for the Church on Spilled Blood to correctly separate both sides of the church, and $\\tau = 0.6$ for Ponte di Rialto [26] to avoid over-segmentation and maintain completeness. Notably, DG-OG fails to disambiguate Belvedere (Vienna), while our method succeeds with a consistent threshold ( $\\tau = 0.8$ across all scenes). Our approach also consistently achieves a higher inlier ratio than the baselines, indicating greater accuracy and completeness in the reconstructed models. As an example, we qualitatively show the results of reconstructing Belvedere (Vienna) in Fig. 5, along with verification results using geo-tagged images.", + "bbox": [ + 511, + 349, + 906, + 590 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "We also evaluate on scenes from the VisymScenes test set. Since VisymScenes images come with geolocation metadata, we do not need additional Mapillary images; instead, we directly compute the inlier ratio between SfM camera positions and the geolocation metadata after RANSAC. Results show that our model effectively prunes spurious pairs and improves the reconstructed model quality, whereas DG-OG sometimes fails to distinguish doppelganger pairs. Fig. 6 shows examples where DG-OG encounters difficulties.", + "bbox": [ + 511, + 592, + 908, + 727 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "In Fig. 7, we show that while MASt3R's features are powerful, MASt3R-SfM is not free from doppelganger issues. Although our classifier was trained on image pairs mined through COLMAP's feature matching module (i.e. w/ SIFT features), it effectively prunes incorrect matches generated by MASt3R, restoring accurate reconstruction results.", + "bbox": [ + 511, + 729, + 908, + 819 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "4.5. Ablation", + "text_level": 1, + "bbox": [ + 513, + 830, + 617, + 847 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "In this section, we study the effectiveness of our designs from the following aspects: 1) fine-tuning the entire model (vs. only new heads), 2) one classification head (vs. two", + "bbox": [ + 511, + 854, + 906, + 900 + ], + "page_idx": 6 + }, + { + "type": "page_number", + "text": "27172", + "bbox": [ + 478, + 944, + 519, + 957 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/d651cbb028479382932dbdcc6c8108c6e8313fe0eb7776ee25109157f3140bd0.jpg", + "image_caption": [ + "Figure 7. MATLAB's Doppelgangers++. MATLAB also suffers from doppelganger issues. Our classifier effectively prunes false positive pairs and correctly reconstructs challenging scenes." + ], + "image_footnote": [], + "bbox": [ + 93, + 88, + 480, + 200 + ], + "page_idx": 7 + }, + { + "type": "table", + "img_path": "images/31f0079bb6304c685ccdc64f2e75fd7327aa50f90d59ad4c9e8db3f3921ac0e9.jpg", + "table_caption": [], + "table_footnote": [], + "table_body": "
Test Scenes# SFM-registered ImagesInlier Ratio (Sec. 3.3)
COLMAPDG-OGOursCOLMAPDG-OGOurs
Alexander Nevsky Cathedral [12]4474444470.5651.01.0
Arc de Triomphé [12]4243844230.6970.9660.990
Berliner Dom [12]1603158816060.7090.9870.992
Big Ben [12]3983793940.5640.8270.831
Church on Spilled Blood [12]27384+94 (τ=0.97)157+1060.5420.8810.962
Radcliffe Camera [12]28191+8494+1860.4950.9550.970
Seville [12]1498585+272+515615+303+5520.4500.7720.854
Brandenburg Gate [26]21371361+5701398+6030.4400.9000.909
Palacio de Comunicaciones [26]727307+80 (τ=0.6)308+840.2290.8230.934
Ponte di Rialto [26]652538+101 (τ=0.6)540+1070.6270.8340.844
York Minster [26]636200+362206+2840.7270.8580.901
Town Hall Tower, Kraków [26]2982552800.6090.7310.838
Belvedere (Vienna) [26]1038851 (fail)457+5000.5210.4510.874
Reichstag (building) [26]1504997+3101024+3560.4690.8040.862
St. Vitus Cathedral [26]7526736920.8530.9090.933
Royal Liver Building [26]2121711800.71.01.0
VisymSite00101544520+29014460.7700.8200.913
VisymSite0023849471+81566+820.8670.8480.942
VisymSite0028450238+179267+1200.8180.7110.909
VisymSite0042540206+2074670.8630.9240.963
VisymSite01091245237+458+78239+612+1270.8570.8620.927
", + "bbox": [ + 94, + 265, + 483, + 449 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Table 2. Evaluation of SfM results. We compare reconstructions from COLMAP [23], DG-OG [3], and our method. $\\tau = 0.8$ is used unless otherwise stated. The $^+$ indicates split reconstruction components; e.g., DG-OG and our method split the Radcliffe Camera reconstruction into two components. Because VisymScenes scenes are large, we report statistics on the largest reconstruction component produced by COLMAP, and identify the corresponding (split) components in the results of DG-OG and ours.", + "bbox": [ + 89, + 459, + 483, + 569 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "separate heads), 3) the architecture of our classification head, and 4) final-layer decoder features (vs. multi-layer decoder features). We evaluate models trained on DG and Visym-Scenes datasets and show PR curves in Fig. 8 on the three test sets with respect to the pairwise classification task. A full ablation table can be found in the supplementary.", + "bbox": [ + 89, + 570, + 483, + 661 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Fine-tuning MASt3R. As discussed in Sec.3.2, we choose to train a lighter output head on top of MASt3R features rather than fine-tuning the entire model weights. Curves in Fig. 8 show that training only the head achieves comparable performance to full model fine-tuning. The drop in recall with full model fine-tuning suggests potential overfitting to training data, making the model more conservative and less generalizable to unseen inputs. This tendency also holds on models trained solely on DG data: on the VisymScenes test set, our approach achieved 0.03 higher AP than fine-tuning. On Mapillary test set, ours achieved comparable AP but with 0.2 higher recall at precision 0.99.", + "bbox": [ + 89, + 667, + 483, + 849 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Single classification head. Our design uses two classification heads to process features from the two branches separately. An alternative approach would be to combine the", + "bbox": [ + 89, + 854, + 483, + 900 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/07045ff87fba6e6f207b98a39d7be224e82c47778c57e18f5ba83a9633309a9d.jpg", + "image_caption": [ + "Figure 8. Precision-Recall curves of ablation studies. Metrics are evaluated on models trained with DG and VisymScenes. Full tables can be found in the supplementary." + ], + "image_footnote": [], + "bbox": [ + 519, + 90, + 903, + 200 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "features from both decoder branches and use a single head to predict the doppelganger score. As shown in Fig. 8, this alternative degrades performance consistently. This may be due to the asymmetric nature of the MAST3R design, where each branch captures different kinds of information that are better to be processed separately. On the other hand, our voting mechanism between the two heads helps enhance class distinction, outputting lower scores for negative samples and higher scores for positive samples.", + "bbox": [ + 511, + 253, + 906, + 391 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Head architecture. While the decoder features of MASt3R retain rich geometric 3D information, designing a classifier with sufficient capacity to effectively leverage this information for the doppelganger prediction task is non-trivial. We compare the performance of our Transformer-based classifier with that of an MLP. Fig. 8 shows that our Transformer model outperforms the MLP across all metrics and test sets.", + "bbox": [ + 511, + 398, + 908, + 505 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Single-layer decoder feature. Our classifier head takes multi-layer decoder features as input. Here we substitute with only the final-layer decoder feature. Results in Fig. 8 show that the multi-layer decoder features are consistently superior to the single-layer ones, because the classifier is able to analyze doppelganger factors from different aspects and levels. See supplementary for more detailed analysis.", + "bbox": [ + 511, + 512, + 908, + 619 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "5. Conclusion", + "text_level": 1, + "bbox": [ + 513, + 633, + 633, + 648 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "We propose Doppelgangers++ as an effective approach to handling visual aliasing in 3D reconstruction. We introduce a new VisymScenes dataset, featuring images from diverse daily scenes, and develop rules to mine doppelganger data from the dataset to enrich our training data. We train a Transformer-based classifier that leverages geometric 3D features to classify image pairs with high precision and recall across various scenes. Doppelgangers++ integrates seamlessly into existing SfM pipelines, enhancing reconstruction accuracy without extensive parameter tuning. Further, we propose a validation method using geo-tagged map images, offering a more comprehensive and automatic way to assess SfM accuracy and model completeness. Extensive experiments show that Doppelgangers++ significantly improves visual disambiguation and 3D reconstruction quality in complex scenes.", + "bbox": [ + 511, + 659, + 908, + 900 + ], + "page_idx": 7 + }, + { + "type": "page_number", + "text": "27173", + "bbox": [ + 478, + 944, + 519, + 957 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "6. Acknowledgment", + "text_level": 1, + "bbox": [ + 89, + 90, + 263, + 107 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "We thank Brandon Li and Joseph Tung for their valuable discussion and help with the webviewer. This work was supported by the Intelligence Advanced Research Projects Activity (IARPA) via Department of Interior/Interior Business Center (DOI/IBC) contract number 140D0423C0035. The U.S. Government is authorized to reproduce and distribute reprints for governmental purposes notwithstanding any copyright annotation thereon. Disclaimer: The views and conclusions contained herein are those of the authors and should not be interpreted as necessarily representing the official policies or endorsements, either expressed or implied, of IARPA, DOI/IBC, or the U.S. Government. This research was also supported in part by the National Science Foundation under award IIS-2212084.", + "bbox": [ + 89, + 114, + 485, + 327 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "References", + "text_level": 1, + "bbox": [ + 91, + 344, + 187, + 359 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[1] Eric Brachmann, Jamie Wynn, Shuai Chen, Tommaso Cavallari, Aron Monszpart, Daniyar Turmukhambetov, and Victor Adrian Prisacariu. Scene coordinate reconstruction: posing of image collections via incremental learning of a relocalizer. In Eur. Conf. Comput. Vis., pages 421-440. Springer, 2025. 3", + "[2] J. Byrne, G. Castanon, Z. Li, and G. Ettinger. Fine-grained activities of people worldwide. In IEEE/CVF Winter Conference on Applications of Computer Vision, 2023. 2, 3", + "[3] Ruojin Cai, Joseph Tung, Qianqian Wang, Hadar Averbuch-Elor, Bharath Hariharan, and Noah Snavely. Doppelgangers: Learning to disambiguate images of similar structures. In Int. Conf. Comput. Vis., pages 34-44, 2023. 1, 2, 3, 5, 6, 7, 8", + "[4] Duygu Ceylan, Niloy Jyoti Mitra, Youyi Zheng, and Mark Pauly. Coupled structure-from-motion and 3d symmetry detection for urban facades. ACM Transactions on Graphics (TOG), 33:1-15, 2014. 2", + "[5] Zhaopeng Cui and Ping Tan. Global structure-from-motion by similarity averaging. 2015 IEEE International Conference on Computer Vision (ICCV), pages 864-872, 2015. 6", + "[6] Daniel DeTone, Tomasz Malisiewicz, and Andrew Rabinovich. Superpoint: Self-supervised interest point detection and description. In Proc. CVPR Workshops, pages 224-236, 2018. 2, 6", + "[7] Bardienus Pieter Duisterhof, Lojze Žust, Philippe Weinzaepfel, Vincent Leroy, Yohann Cabon, and Jérôme Revaud. Mast3r-sfm: a fully-integrated solution for unconstrained structure-from-motion. ArXiv, abs/2409.19152, 2024. 2, 3, 6", + "[8] Mihai Dusmanu, Ignacio Rocco, Tomás Pajdla, Marc Pollefeys, Josef Sivic, Akihiko Torii, and Torsten Sattler. D2-net: A trainable cnn for joint description and detection of local features. 2019 IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 8084-8093, 2019. 6", + "[9] Martin A. Fischler and Robert C. Bolles. Random sample consensus: a paradigm for model fitting with applications to image analysis and automated cartography. Commun. ACM, 24:381-395, 1981. 6" + ], + "bbox": [ + 99, + 369, + 485, + 898 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[10] Ye Gong, Pengwei Zhou, Changfeng Liu, Yan Yu, Jian Yao, Wei Yuan, and Li Li. A cluster-based disambiguation method using pose consistency verification for structure from motion. ISPRS Journal of Photogrammetry and Remote Sensing, 2024. 2", + "[11] Jared Heinly, Enrique Dunn, and Jan-Michael Frahm. Recovering correct reconstructions from indistinguishable geometry. 2014 2nd International Conference on 3D Vision, 1:377-384, 2014. 2", + "[12] Jared Heinly, Enrique Dunn, and Jan-Michael Frahm. Correcting for duplicate scene structure in sparse 3d reconstruction. In Eur. Conf. Comput. Vis., pages 780-795. Springer, 2014. 6, 8", + "[13] Nianjuan Jiang, Ping Tan, and Loong-Fah Cheong. Seeing double without confusion: Structure-from-motion in highly ambiguous scenes. In Conf. Comput. Vis. Pattern Recog., pages 1458-1465, 2012. 2", + "[14] Rajbir Kataria, Joseph DeGol, and Derek Hoiem. Improving structure from motion with reliable resectioning. 2020 International Conference on 3D Vision (3DV), pages 41-50, 2020. 2", + "[15] Diederik P. Kingma and Jimmy Ba. Adam: A method for stochastic optimization. CoRR, abs/1412.6980, 2014. 6", + "[16] Vincent Leroy, Yohann Cabon, and Jérôme Revaud. Grounding image matching in 3d with mast3r. ArXiv, abs/2406.09756, 2024. 2, 3, 6", + "[17] David G Lowe. Distinctive image features from scale-invariant keypoints. Int. J. Comput. Vis., 60:91-110, 2004. 2, 6", + "[18] Lalit Manam and Venu Madhav Govindu. Leveraging camera triplets for efficient and accurate structure-from-motion. 2024 IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 4959-4968, 2024. 2", + "[19] Yang Peng, Shen Yan, Yuxiang Liu, Y. Liu, and Maojun Zhang. View graph construction for scenes with duplicate structures via graph convolutional network. IET Comput. Vis., 16:389-402, 2022. 2", + "[20] Richard Roberts, Sudipta N. Sinha, Richard Szeliski, and Drew Steedly. Structure from motion for scenes with large duplicate structures. CVPR 2011, pages 3137-3144, 2011. 2", + "[21] Ignacio Rocco, Mircea Cimpoi, Relja Arandjelović, Akihiko Torii, Tomas Pajdla, and Josef Sivic. Neighbourhood consensus networks. In Adv. Neural Inform. Process. Syst., 2018. 2", + "[22] Paul-Edouard Sarlin, Daniel DeTone, Tomasz Malisiewicz, and Andrew Rabinovich. Superglue: Learning feature matching with graph neural networks. 2020 IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 4937–4946, 2019. 6", + "[23] Johannes L Schonberger and Jan-Michael Frahm. Structure-from-motion revisited. In Conf. Comput. Vis. Pattern Recog., pages 4104-4113, 2016. 2, 3, 5, 6, 7, 8", + "[24] Tianwei Shen, Siyu Zhu, Tian Fang, Runze Zhang, and Long Quan. Graph-based consistent matching for structure-from-motion. In European Conference on Computer Vision, 2016. 2" + ], + "bbox": [ + 516, + 92, + 906, + 898 + ], + "page_idx": 8 + }, + { + "type": "page_number", + "text": "27174", + "bbox": [ + 478, + 944, + 519, + 955 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[25] Jiaming Sun, Zehong Shen, Yuang Wang, Hujun Bao, and Xiaowei Zhou. Loftr: Detector-free local feature matching with transformers. In Conf. Comput. Vis. Pattern Recog., pages 8922-8931, 2021. 2", + "[26] Joseph Tung, Gene Chou, Ruojin Cai, Guandao Yang, Kai Zhang, Gordon Wetzstein, Bharath Hariharan, and Noah Snavely. Megascenes: Scene-level view synthesis at scale. ArXiv, abs/2406.11819, 2024. 5, 6, 7, 8", + "[27] Fan Wang, Aditi Nayak, Yogesh Agrawal, and Roy Shilkrot. Hierarchical image link selection scheme for duplicate structure disambiguation. In British Machine Vision Conference, 2018. 2", + "[28] Jianyuan Wang, Nikita Karaev, Christian Rupprecht, and David Novotny. Vggsfm: Visual geometry grounded deep structure from motion. In Conf. Comput. Vis. Pattern Recog., pages 21686-21697, 2024. 3", + "[29] Shuzhe Wang, Vincent Leroy, Yohann Cabon, Boris Chidlovskii, and Jérôme Revaud. DUS3R: Geometric 3d vision made easy. In Conf. Comput. Vis. Pattern Recog., pages 20697-20709, 2023. 2, 3", + "[30] Frederik Warburg, Søren Hauberg, Manuel López-Antequera, Pau Gargallo, Yubin Kuang, and Javier Civera. Mapillary street-level sequences: A dataset for lifelong place recognition. In Conf. Comput. Vis. Pattern Recog., pages 2623-2632, 2020. 6", + "[31] Kyle Wilson and Noah Snavely. Network principles for sfm: Disambiguating repeated structures with local context. In Int. Conf. Comput. Vis., pages 513-520, 2013. 2, 6, 8", + "[32] Qingan Yan, Long Yang, Ling Zhang, and Chunxia Xiao. Distinguishing the indistinguishable: Exploring structural ambiguities via geodesic context. In Conf. Comput. Vis. Pattern Recog., pages 3836-3844, 2017. 2, 6", + "[33] Kwang Moo Yi, Eduard Trulls, Vincent Lepetit, and Pascal Fua. Lift: Learned invariant feature transform. In Eur. Conf. Comput. Vis., pages 467-483, 2016. 2", + "[34] Christopher Zach, Arnold Irschara, and Horst Bischof. What can missing correspondences tell us about 3d structure and motion? In Conf. Comput. Vis. Pattern Recog., pages 1-8, 2008. 2", + "[35] Christopher Zach, Manfred Klopschitz, and Marc Pollefeys. Disambiguating visual relations using loop constraints. In Conf. Comput. Vis. Pattern Recog., pages 1426-1433, 2010." + ], + "bbox": [ + 91, + 92, + 483, + 696 + ], + "page_idx": 9 + }, + { + "type": "page_number", + "text": "27175", + "bbox": [ + 478, + 944, + 517, + 955 + ], + "page_idx": 9 + } +] \ No newline at end of file diff --git a/2025/Doppelgangers++_ Improved Visual Disambiguation with Geometric 3D Features/a4009708-9357-4078-b5aa-d3dd7d42ccfe_model.json b/2025/Doppelgangers++_ Improved Visual Disambiguation with Geometric 3D Features/a4009708-9357-4078-b5aa-d3dd7d42ccfe_model.json new file mode 100644 index 0000000000000000000000000000000000000000..27ea89f8ea529c31ab1c67dab1d2fff2d04254ef --- /dev/null +++ b/2025/Doppelgangers++_ Improved Visual Disambiguation with Geometric 3D Features/a4009708-9357-4078-b5aa-d3dd7d42ccfe_model.json @@ -0,0 +1,2035 @@ +[ + [ + { + "type": "header", + "bbox": [ + 0.107, + 0.003, + 0.182, + 0.043 + ], + "angle": 0, + "content": "CVF" + }, + { + "type": "header", + "bbox": [ + 0.237, + 0.001, + 0.812, + 0.047 + ], + "angle": 0, + "content": "This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore." + }, + { + "type": "title", + "bbox": [ + 0.092, + 0.131, + 0.908, + 0.155 + ], + "angle": 0, + "content": "Doppelgangers++: Improved Visual Disambiguation with Geometric 3D Features" + }, + { + "type": "text", + "bbox": [ + 0.167, + 0.18, + 0.831, + 0.218 + ], + "angle": 0, + "content": "Yuanbo Xiangli\\(^{1}\\) Ruojin Cai\\(^{1}\\) Hanyu Chen\\(^{1}\\) Jeffrey Byrne\\(^{2}\\) Noah Snavely\\(^{1}\\) \n\\(^{1}\\)Cornell University, \\(^{2}\\)Visym Labs" + }, + { + "type": "aside_text", + "bbox": [ + 0.1, + 0.255, + 0.131, + 0.337 + ], + "angle": 90, + "content": "(sneepnep)" + }, + { + "type": "image", + "bbox": [ + 0.142, + 0.233, + 0.286, + 0.297 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.145, + 0.298, + 0.215, + 0.306 + ], + "angle": 0, + "content": "Ours +MASt3R-SfM" + }, + { + "type": "image", + "bbox": [ + 0.144, + 0.308, + 0.286, + 0.361 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.289, + 0.232, + 0.436, + 0.298 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.29, + 0.298, + 0.436, + 0.362 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.439, + 0.232, + 0.585, + 0.298 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.442, + 0.298, + 0.584, + 0.362 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.587, + 0.232, + 0.735, + 0.298 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.589, + 0.298, + 0.735, + 0.362 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.738, + 0.232, + 0.886, + 0.298 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.739, + 0.298, + 0.886, + 0.362 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.194, + 0.372, + 0.291, + 0.472 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.195, + 0.474, + 0.285, + 0.485 + ], + "angle": 0, + "content": ":0.235/Ours:0.982" + }, + { + "type": "image", + "bbox": [ + 0.293, + 0.373, + 0.439, + 0.472 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.294, + 0.474, + 0.435, + 0.485 + ], + "angle": 0, + "content": "GT:1/DG-OG:0.190/Ours:0.971" + }, + { + "type": "image", + "bbox": [ + 0.441, + 0.373, + 0.588, + 0.472 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.442, + 0.474, + 0.583, + 0.484 + ], + "angle": 0, + "content": "GT:1 / DG-OG:0.378 / Ours:0.952" + }, + { + "type": "image", + "bbox": [ + 0.59, + 0.373, + 0.735, + 0.472 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.591, + 0.474, + 0.732, + 0.484 + ], + "angle": 0, + "content": "GT:0/DG-OG:0.993/Ours:0.025" + }, + { + "type": "image", + "bbox": [ + 0.738, + 0.373, + 0.886, + 0.472 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.741, + 0.474, + 0.882, + 0.484 + ], + "angle": 0, + "content": "GT:0/DG-OG:0.994/Ours:0.031" + }, + { + "type": "image", + "bbox": [ + 0.142, + 0.492, + 0.411, + 0.614 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.412, + 0.492, + 0.571, + 0.551 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.413, + 0.552, + 0.571, + 0.612 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.572, + 0.492, + 0.89, + 0.613 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.09, + 0.625, + 0.908, + 0.687 + ], + "angle": 0, + "content": "Figure 1. Visual aliasing, or doppelgangers, poses severe challenges to 3D reconstruction. We propose Doppelganger++, an enhanced pairwise image classifier that excels in visual disambiguation across diverse and challenging scenes. (Top) We seamlessly integrate Doppelganger++ into SfM, successfully disambiguating each scene. (Middle) Compared to prior work (which we refer to as DG-OG [3]), Doppelgangers++ is more robust for everyday scenes, showing improved accuracy and robustness. We show pairs that DG-OG classify incorrectly and ours gets correct. (Bottom) Our new VisymScenes dataset, featuring complex daily scenes, is particularly challenging for COLMAP and DG-OG, but our method can achieve correct and complete reconstructions." + }, + { + "type": "title", + "bbox": [ + 0.249, + 0.699, + 0.327, + 0.714 + ], + "angle": 0, + "content": "Abstract" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.735, + 0.486, + 0.901 + ], + "angle": 0, + "content": "Accurate 3D reconstruction is frequently hindered by visual aliasing, where visually similar but distinct surfaces (aka, doppelgangers), are incorrectly matched. These spurious matches distort the structure-from-motion (SfM) process, leading to misplaced model elements and reduced accuracy. Prior efforts addressed this with CNN classifiers trained on curated datasets, but these approaches struggle to generalize across diverse real-world scenes and can require extensive parameter tuning. In this work, we present Doppelgangers++, a method to enhance doppelganger detection and improve 3D reconstruction accuracy. Our contribu" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.7, + 0.909, + 0.897 + ], + "angle": 0, + "content": "tions include a diversified training dataset that incorporates geo-tagged images from everyday scenes to expand robustness beyond landmark-based datasets. We further propose a Transformer-based classifier that leverages 3D-aware features from the MAST3R model, achieving superior precision and recall across both in-domain and out-of-domain tests. Doppelgangers++ integrates seamlessly into standard SfM and MAST3R-SfM pipelines, offering efficiency and adaptability across varied scenes. To evaluate SfM accuracy, we introduce an automated, geotag-based method for validating reconstructed models, eliminating the need for manual inspection. Through extensive experiments, we demonstrate that Doppelgangers++ significantly enhances pairwise vi" + }, + { + "type": "page_number", + "bbox": [ + 0.479, + 0.946, + 0.521, + 0.957 + ], + "angle": 0, + "content": "27166" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.091, + 0.092, + 0.486, + 0.123 + ], + "angle": 0, + "content": "sual disambiguation and improves 3D reconstruction quality in complex and diverse scenarios." + }, + { + "type": "title", + "bbox": [ + 0.092, + 0.151, + 0.226, + 0.167 + ], + "angle": 0, + "content": "1. Introduction" + }, + { + "type": "text", + "bbox": [ + 0.089, + 0.177, + 0.486, + 0.341 + ], + "angle": 0, + "content": "Visual aliasing—confusing two surfaces that look the same but are nonetheless distinct—is an pernicious problem in 3D reconstruction and SLAM systems. Pairs of images that depict visually similar yet distinct 3D surfaces (called doppelgangers by Cai et al. [3]) can generate spurious correspondence at the feature matching stage of 3D reconstruction, leading to erroneous downstream reconstructions that feature distorted geometry or incorrectly fused elements. Therefore, to ensure the accuracy of 3D reconstruction, it is critical to distinguish truly matching images from illusory matches arising from doppelganger pairs." + }, + { + "type": "text", + "bbox": [ + 0.093, + 0.343, + 0.486, + 0.644 + ], + "angle": 0, + "content": "In recent work, this visual disambiguation problem was formulated as a binary classification task on image pairs [3]. The authors collected a Internet dataset of visually similar doppelganger pairs (as well as truly matching images) from Wikipedia Commons, then trained a CNN to classify image pairs as correct or incorrect matches. This classifier can be used to take a feature match graph computed from a set of photos, remove incorrect edges between doppelganger image pairs, then reconstruct a correct model using a structure from motion (SfM) pipeline like COLMAP [23]. While that work shows promising improvements on reconstruction problems, we find that it can still be brittle: First, the 3D reconstruction task demands that the classifier have has precision—even a few bad edges remaining in the image match graph can lead to an incorrect 3D model. Second, the reconstruction task is quite sensitive to the threshold on the classifier score used to prune edges from the match graph. Finally, their method was trained solely on landmark Internet photos and does not reliably generalize to new scenarios, such as more structured captures of everyday scenes, like streets and office buildings." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.645, + 0.485, + 0.675 + ], + "angle": 0, + "content": "In this paper, we aim to address these issues and improve doppelganger classification in several key ways:" + }, + { + "type": "text", + "bbox": [ + 0.092, + 0.675, + 0.485, + 0.749 + ], + "angle": 0, + "content": "1. We identify ways to expand and diversify doppelgangers training data. In particular, we leverage semi-structured image data with geographic annotations (rough GPS position and orientation), captured from everyday scenes with the Visym Collector platform [2]." + }, + { + "type": "text", + "bbox": [ + 0.091, + 0.75, + 0.485, + 0.856 + ], + "angle": 0, + "content": "2. We switch from a CNN-based classifier to leveraging features from MASt3R [7], a recent transformer-based geometric model that computes point clouds from two input views. Specifically, we feed an image pair into a pre-trained MASt3R model, collect the intermediate features decoded by MASt3R, and train a classification head to map these features to a doppelganger score." + }, + { + "type": "list", + "bbox": [ + 0.091, + 0.675, + 0.485, + 0.856 + ], + "angle": 0, + "content": null + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.856, + 0.485, + 0.901 + ], + "angle": 0, + "content": "We call our method Doppelgangers++. Our enhanced model achieves higher precision and recall in pairwise classification, and generalizes better across a broader range of scenes and" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.092, + 0.909, + 0.23 + ], + "angle": 0, + "content": "capture scenarios. Doppelganger++ integrates seamlessly into SfM pipelines for 3D reconstruction disambiguation. We further propose to leverage geo-tagged map images to quantitatively evaluate the correctness and completeness of reconstructed models, in comparison to manual inspection as required in previous approaches. Through extensive experiments we show that our model leads to more accurate and complete reconstruction results, and is less sensitive to the threshold used for pruning doppelganger matches." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.244, + 0.659, + 0.26 + ], + "angle": 0, + "content": "2. Related Work" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.271, + 0.909, + 0.604 + ], + "angle": 0, + "content": "Local feature matching and 3D learning. Local feature matching methods have proven effective at establishing correspondences between pairs of images in SfM pipelines. Classic methods like COLMAP [23] rely on the tried-and-true SIFT features [17] to find correspondence. Modern learning-based, data-driven approaches have improved the quality of local feature matching [6, 21, 25, 33]. More recently, the DUSt3R [29] framework has proven to excel in a variety of 3D reconstruction tasks by estimating dense 3D point clouds from pairs of images. MASt3R [16] was proposed as an extension to DUSt3R that specifically targets the task of predicting dense feature matches between image pairs. All of these local feature matching methods are traditionally optimized to maximize the number of correspondences they find between similar image regions. While their ability to identify feature matches has greatly improved, they generally lack the ability to incorporate negative evidence into their predictions, and so they often find matches between regions that do not actually correspond to the same 3D surface, particularly within doppelganger image pairs. However, we show that features internal to MASt3R can be repurposed for doppelganger detection." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.614, + 0.911, + 0.902 + ], + "angle": 0, + "content": "Disambiguation in SfM and image matching. Disambiguating similar structures and repeated patterns in SfM is a long-standing challenge. Prior work has mainly relied on heuristics-based analysis to detect conflicting relations and ambiguities in the structure of the underlying scene graph [10, 14, 18, 19, 24, 27, 31, 32, 35] or among image-level correspondence (or the lack thereof) [4, 11-13, 20, 34]. While such methods have shown some success in detecting and resolving ambiguities in SfM pipelines, one fundamental limitation that they do not consider the rich information contained in the images themselves. In contrast, Doppelgangers [3] avoids handcrafted heuristics and takes a data-driven approach to the basic problem of identifying doppelganger image pairs. They train a CNN that classifies a pair of images as either positive or negative, with LoFTR [25]-extracted keypoints and match masks passed as auxiliary input. During inference, the binary classification results are used to pre-process the scene graph obtained from COLMAP to filter out spurious matches prior to running SfM. How" + }, + { + "type": "page_number", + "bbox": [ + 0.479, + 0.945, + 0.521, + 0.958 + ], + "angle": 0, + "content": "27167" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.09, + 0.092, + 0.482, + 0.212 + ], + "angle": 0, + "content": "ever, the method's reliance on COLMAP feature extraction and matching and the need for auxiliary mask information from LoFTR introduces significant overhead and complexity to the overall pipeline, making it difficult to scale to larger scenes. In addition, we find that this prior method is brittle: it can fail to generalize to domains beyond landmarks (e.g., to office buildings), and it can require parameter tuning to get good results." + }, + { + "type": "text", + "bbox": [ + 0.093, + 0.221, + 0.487, + 0.493 + ], + "angle": 0, + "content": "Differentiable SfM. To improve upon traditional SfM, several differentiable SfM pipelines have been proposed to optimize the entire 3D reconstruction process. These pipelines often include one or more components that are trained from large datasets. MAST3R-SfM [7] proposes an SfM pipeline that uses MAST3R features for scene graph construction and coarse-to-fine alignment. VGGSfM [28] decomposes the problem into four stages: point tracking, initial camera estimation, triangulation, and bundle adjustment, each of which is differentiable and can be optimized in an end-to-end manner. ACE0 [1] uses a scene coordinate regression network to progressively register new images to a global scene, while the network itself is iteratively trained and refined from the registered images. These methods have shown promising results in optimizing the SfM pipeline, but are not specifically designed to address the doppelganger problem, and are still prone to producing incorrect reconstructions when faced with scenes with repeated patterns or similar structures." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.509, + 0.182, + 0.525 + ], + "angle": 0, + "content": "3. Method" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.535, + 0.483, + 0.673 + ], + "angle": 0, + "content": "We refer to a \"doppelganger\" as a case where distinct objects or surfaces look almost identical, leading algorithms to confuse one for the other [3]. In this work, we aim to improve the doppelganger classifier via expanded and diversified training data (Sec. 3.1), and leveraging geometric 3D features learned from pairwise reconstruction [16] (Sec. 3.2). We further propose an approach to quantitatively evaluate the correctness of SfM results in terms of doppelgangers (Sec. 3.3), instead of manual inspection adopted in previous work." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.682, + 0.32, + 0.697 + ], + "angle": 0, + "content": "3.1. The VisymScenes dataset" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.704, + 0.483, + 0.795 + ], + "angle": 0, + "content": "Cai et al. [3] introduced the Doppelganger dataset, built on global landmarks photos sourced from Wikipedia Commons, with viewing direction (e.g., North, South) to identify doppelganger image pairs. While the classifier trained on this dataset proved effective, we find that it struggles to generalize well to scenes beyond the dataset's domain." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.796, + 0.483, + 0.901 + ], + "angle": 0, + "content": "We enhance and expand their training set to improve the robustness of doppelganger classification, by incorporating casually captured images from a wider range of scenes. To ensure sufficient diversity in the dataset, we introduce the VisymScenes dataset. VisymScenes consists of 258K ground images with GPS/IMU metadata, recorded at 149 sites in 42 cities and 15 countries, collected with the Visym Collector" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.092, + 0.908, + 0.227 + ], + "angle": 0, + "content": "Platform [2] (details in supplementary). Each image is accompanied by metadata, such as GPS coordinates, device compass direction, and intrinsic camera calibration. While this metadata can be noisy, it still provides valuable information for identifying potential doppelganger pairs. For example, if two images exhibit numerous geometrically consistent local feature matches yet were taken from distant locations, this serves as strong evidence that the pair is likely to be a doppelganger. Examples are shown in Fig. 2." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.229, + 0.909, + 0.668 + ], + "angle": 0, + "content": "We develop a series of filtering rules applied to all matched image pairs identified by the COLMAP feature matching module [23]. Note that these rules are designed according to the capture style of the Visym Collector, where cameras consistently focus on the scene of interest and maintain a reasonable distance from it. For a given pair of images, we use the (metric) distance between the camera centers \\( r \\), angle between their viewing directions \\( \\theta \\), and camera intrinsics \\( \\mathcal{K} \\), all derived from camera metadata, to identify confident negative (doppelganger) and positive (correct match) pairs. To identify confident negative pairs, we first identify spatially distant matching pairs, and classify them as doppelgangers. Then, we classify the remaining (nearby) pairs into three cases: the intersection point of the viewing directions is 1) in front of both cameras, 2) behind both cameras; or 3) in front of one camera and behind the other. For case 1), we label pairs with very large view angles (e.g., \\( >160^{\\circ} \\)) as negative, since they likely capture different 3D surfaces or mostly non-overlapping content. For case 2), if their view angle exceeds the diagonal field of view, then their view frustums are unlikely to overlap, so we label them as doppelgangers. For case 3), we check for frustum overlap using camera intrinsics; if no intersection is detected, the pair is considered a doppelganger. A similar series of rules are designed for mining positive pairs (i.e., true matches). Details of computing view intersection and filtering algorithms for both negative and positive pairs can be found in the supplementary. With these rules, we mined in total 53K positive and negative pairs across 33 sites for our doppelganger task." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.681, + 0.816, + 0.697 + ], + "angle": 0, + "content": "3.2. Improved Doppelganger Classifier" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.704, + 0.908, + 0.779 + ], + "angle": 0, + "content": "Recent advances in data-driven models [16, 29] have demonstrated impressive results in geometric vision tasks. In our work, we leverage the multi-level, 3D-aware features extracted from the MAST3R model [16] to train a doppelganger classifier on labeled image pairs, as demonstrated in Fig. 3." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.781, + 0.909, + 0.901 + ], + "angle": 0, + "content": "Equipped with large-scale training data and a ViT backbone, the MASt3R model captures an image representation that encodes rich 3D geometric information between paired images. However, as MASt3R was originally trained to detect correspondences and similarities, its point maps often conflate doppelganger pairs, yielding incorrect point clouds and poses for these challenging pairs (examples in Fig. 7). Despite this limitation, we find that the internal" + }, + { + "type": "page_number", + "bbox": [ + 0.48, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "27168" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.098, + 0.09, + 0.298, + 0.218 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.299, + 0.091, + 0.495, + 0.217 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.5, + 0.091, + 0.697, + 0.217 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.7, + 0.091, + 0.899, + 0.217 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.09, + 0.227, + 0.908, + 0.285 + ], + "angle": 0, + "content": "Figure 2. VisymScenes examples. This new dataset includes residential areas, landmarks, historical sites, business districts, and more. Here, we present four example sites. The top row shows subsets of images captured within each site. The bottom row displays pairs of visually similar but geographically distinct images from each site along with their recorded geolocations on a map. These examples demonstrate that doppelganger issues are prevalent in everyday scenes, presenting significant challenges for reliable 3D reconstruction and image matching." + }, + { + "type": "image", + "bbox": [ + 0.095, + 0.293, + 0.905, + 0.4 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.09, + 0.402, + 0.908, + 0.46 + ], + "angle": 0, + "content": "Figure 3. Model design. (Left) Given an image pair, we first create a symmetrized version of the pair and feed it into the frozen MASt3R model. Multi-layer features are extracted from each decoder branch, concatenated, and fed into two learnable doppelganger classification heads. Each head generates predictions \\((\\mathrm{pred}_{pq}^{v}, \\mathrm{pred}_{qp}^{v})\\), \\(v \\in \\{1,2\\}\\) (where \\(pq\\) and \\(qp\\) denote the symmetrized image pair), supervised by cross-entropy loss. (Right) We use multi-layer decoder features and a Transformer-based classifier head for doppelganger prediction." + }, + { + "type": "text", + "bbox": [ + 0.089, + 0.472, + 0.484, + 0.637 + ], + "angle": 0, + "content": "features learned by MASt3R contain sufficient information for doppelganger classification without the need to fine-tune MASt3R's weights. Moreover, given that the amount of available training data for the doppelganger task is dwarfed by the amount of data used to train MASt3R, we opt not to fine-tune the entire model, but instead repurpose these internal features by training an additional output head to predict a doppelganger classification score. In Sec. 4.5, we empirically demonstrate that this choice achieves comparable or better visual disambiguation performance, especially on out-of-domain scenes." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.644, + 0.484, + 0.749 + ], + "angle": 0, + "content": "Multi-level decoder features. For an image pair \\((I_p, I_q)\\), we extract multi-level decoder features from MASt3R's decoder branches. MASt3R takes two images as input, and uses two intertwined decoders \\(\\mathrm{DecBlk}^1\\) and \\(\\mathrm{DecBlk}^2\\) to decode encoded features \\(\\mathcal{H}^1 = \\mathrm{Enc}^1(I_p)\\) and \\(\\mathcal{H}^2 = \\mathrm{Enc}^2(I_q)\\). Each decoder has \\(B\\) blocks, and attends to tokens from the other branch:" + }, + { + "type": "equation", + "bbox": [ + 0.102, + 0.755, + 0.484, + 0.775 + ], + "angle": 0, + "content": "\\[\nf _ {i} ^ {1} = \\operatorname {D e c B l k} _ {i} ^ {1} \\left(f _ {i - 1} ^ {1}, f _ {i - 1} ^ {2}\\right), f _ {i} ^ {2} = \\operatorname {D e c B l k} _ {i} ^ {2} \\left(f _ {i - 1} ^ {2}, f _ {i - 1} ^ {1}\\right), \\tag {1}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.78, + 0.484, + 0.902 + ], + "angle": 0, + "content": "where \\( f_{i}^{v} \\) denotes the output tokens from the \\( i \\)-th block of the \\( v \\)-th branch. These two branches exchange information to capture the spatial relationships between viewpoints and the global 3D structure of the scene. For each branch \\( v \\), we concatenate the encoder feature \\( \\mathcal{H}^v \\) and the tokens from the decoder blocks into \\( \\mathcal{F}^v = \\mathrm{concat}(\\mathcal{H}^v, \\{f_i^v\\}_{i=0}^{B-1}) \\), which captures rich, multi-level spatial correspondence information between image pairs." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.472, + 0.909, + 0.67 + ], + "angle": 0, + "content": "Transformer-based classification heads. MAsst3R treats one image as the reference frame and projects the other image into that reference frame. A consequence of this design is that the reconstructed 3D scene for input pair \\((I_p, I_q)\\) is distinct (or at least, must be in a different coordinate frame) from that of \\((I_q, I_p)\\). Inspired by the asymmetric design, we propose to 1) use separate Transformer heads: We introduce two independent, Transformer-based classification heads \\(\\mathrm{Head}_{dopp}^1\\) and \\(\\mathrm{Head}_{dopp}^2\\) to process \\(\\mathcal{F}^1\\) and \\(\\mathcal{F}^2\\) respectively; and 2) symmetrize image pairs such that the model evaluates both \\((I_p, I_q)\\) and \\((I_q, I_p)\\) to decide whether the given pair of images are true match or not. Thus, we end up with four scores for each image pair:" + }, + { + "type": "equation", + "bbox": [ + 0.533, + 0.681, + 0.907, + 0.7 + ], + "angle": 0, + "content": "\\[\n\\operatorname {p r e d} _ {u} ^ {v} = \\operatorname {H e a d} _ {\\text {d o p p}} ^ {v} \\left(\\mathcal {F} _ {u} ^ {v}\\right), u \\in \\{p q, q p \\}, v \\in \\{1, 2 \\}. \\tag {2}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.71, + 0.909, + 0.848 + ], + "angle": 0, + "content": "Essentially, the two heads serve as expert evaluators, each examining how likely the image pair is to be a true match (or doppelganger) from different aspects, and switching the input order allows the model to analyze the spatial relationships from both directions. We empirically show the effectiveness of this design in Sec. 4.5. Both \\(\\mathrm{Head}_{dopp}^{1}\\) and \\(\\mathrm{Head}_{dopp}^{2}\\) are supervised by the cross-entropy loss, encouraging \\(S = \\{\\mathrm{pred}_u^v\\}\\) to all give high probabilities for positive matches and low probabilities for negative ones." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.856, + 0.909, + 0.902 + ], + "angle": 0, + "content": "Test-time voting. At test time, we combine the four scores \\( S \\) via a voting mechanism to make a final decision. Specifically, if the majority of the heads predict that the pair is a positive" + }, + { + "type": "page_number", + "bbox": [ + 0.479, + 0.945, + 0.521, + 0.958 + ], + "angle": 0, + "content": "27169" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.09, + 0.092, + 0.485, + 0.211 + ], + "angle": 0, + "content": "match, we take \\(\\max(S)\\) as the final score for the image pair. Conversely, if the majority of the heads vote for a negative match, \\(\\min(S)\\) is used as the final score. Otherwise, we average the scores from the four heads. This approach ensures that the final decision reflects the strongest evidence supporting the consensus among the heads and thereby improves the reliability of the classifier. Eq. 3 elaborates on this voting mechanism:" + }, + { + "type": "equation", + "bbox": [ + 0.101, + 0.215, + 0.484, + 0.27 + ], + "angle": 0, + "content": "\\[\nS _ {\\text {f i n a l}} = \\left\\{ \\begin{array}{l l} \\max (\\mathcal {S}) & \\text {i f} \\sum_ {s \\in \\mathcal {S}} \\mathbf {1} (s > 0. 5) > \\sum_ {s \\in \\mathcal {S}} \\mathbf {1} (s < 0. 5), \\\\ \\min (\\mathcal {S}) & \\text {i f} \\sum_ {s \\in \\mathcal {S}} \\mathbf {1} (s > 0. 5) < \\sum_ {s \\in \\mathcal {S}} \\mathbf {1} (s < 0. 5), \\\\ \\operatorname {m e a n} (\\mathcal {S}) & \\text {i f} \\sum_ {s \\in \\mathcal {S}} \\mathbf {1} (s > 0. 5) = \\sum_ {s \\in \\mathcal {S}} \\mathbf {1} (s < 0. 5). \\end{array} \\right. \\tag {3}\n\\]" + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.278, + 0.465, + 0.294 + ], + "angle": 0, + "content": "3.3. Evaluating Doppelganger correction in SfM" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.3, + 0.485, + 0.556 + ], + "angle": 0, + "content": "There is currently no reliable benchmarking method for evaluating the accuracy and correctness of SfM reconstructions specifically in terms of how well they address the doppelganger issue. Prior work [3] relied on manual inspection of each out model to assess the effectiveness of their approach—a time-consuming, unquantifiable, and potentially error-prone process. In our work, we propose a benchmarking method for qualitatively evaluating reconstructed models with respect to doppelgangers. We leverage mapping sites like Mapillary, which provide images with location metadata that can serve as probes for validating a 3D model. Note that our method targets common scenarios where geotags are unavailable, or so noisy as to not be useful. Therefore, we do not consider the use of geotags for the reconstruction task itself—we explore pure visual disambiguation—but instead gather them from specialized sources for ground truth evaluation." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.557, + 0.484, + 0.814 + ], + "angle": 0, + "content": "Specifically, for a given scene with known geolocation, we acquire from Mapillary a set of nearby geo-tagged images and register them to the reconstructed model with COLMAP [23]. We then apply RANSAC to robustly estimate a similarity transformation between the registered camera positions and their corresponding image geolocation metadata (converted to ECEF coordinates). We use the resulting RANSAC inlier ratio as an indicator of the model's correctness. As an example, in Fig. 4, we can observe that before correction, the registered cameras and their geolocations show significant misalignment, with the cameras collapsing to one side due to similar-looking surfaces. After correction, the registered cameras align closely to their true geolocations, indicating that doppelganger pairs are correctly separated in the model. If the model is split into multiple components, we calculate the weighted average of inliers ratios among the components:" + }, + { + "type": "equation", + "bbox": [ + 0.148, + 0.822, + 0.484, + 0.864 + ], + "angle": 0, + "content": "\\[\n\\operatorname {I R} = \\sum_ {i = 1} ^ {N} \\left(\\frac {I _ {i}}{T _ {i}} \\cdot \\frac {T _ {i}}{\\sum_ {j = 1} ^ {N} T _ {j}}\\right) = \\frac {\\sum_ {i = 1} ^ {N} I _ {i}}{\\sum_ {i = 1} ^ {N} T _ {i}}, \\tag {4}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.871, + 0.485, + 0.902 + ], + "angle": 0, + "content": "where \\(I_{i}\\) is the number of RANSAC inliers and \\(T_{i}\\) is the number of registered Mapillary images in the \\(i\\)-th compo" + }, + { + "type": "image", + "bbox": [ + 0.537, + 0.092, + 0.882, + 0.189 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.537, + 0.191, + 0.885, + 0.374 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.512, + 0.38, + 0.908, + 0.518 + ], + "angle": 0, + "content": "Figure 4. Evaluation of doppelganger correction in SfM. (Top) We first collect sequences of geo-tagged Mapillary images around the target location and register them to the SfM model. Then, we use RANSAC to align the registered cameras and their geolocations. The inlier ratio is computed as an indicator of model accuracy. (Bottom) In the model corrupted by doppelganger pairs, the registered cameras all collapse to one side. We see that the camera poses estimated with COLMAP (right, in red) do not align well with the geotags (green), leading to a low inlier ratio, but our method leads to a much closer alignment." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.53, + 0.907, + 0.635 + ], + "angle": 0, + "content": "nent. While alignment error could also serve as a metric, we opt to use the inlier ratio to reduce sensitivity to geolocation inaccuracies, as image geolocations may not always be precise. This strategy can also be used to detect broken reconstructions from datasets like MegaScenes [26], which contains over 100K SfM results from Internet photos around the world." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.654, + 0.649, + 0.67 + ], + "angle": 0, + "content": "4. Experiments" + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.68, + 0.702, + 0.696 + ], + "angle": 0, + "content": "4.1. Experimental setup" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.703, + 0.907, + 0.794 + ], + "angle": 0, + "content": "Dataset. Our training set is comprised of the Doppelganger dataset [3] (dubbed DG) and the new VisymScenes dataset. Specifically, the DG training set includes 73K image pairs, nearly evenly split between positive and negative pairs. From VisymScenes, we use 47K image pairs across 26 sites as training data, evenly split into positive and negative pairs." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.796, + 0.909, + 0.903 + ], + "angle": 0, + "content": "We evaluate on two tasks, 1) pairwise image visual disambiguation and 2) scene reconstruction by integrating our classifier into COLMAP [23]. For pairwise image visual disambiguation, we utilize the DG test set, and mined an additional 3,180 pairs from 7 Visym sites (distinct from the 26 training sites), equally divided into positive and negative pairs. Additionally, we created a test set from the Mapillary" + }, + { + "type": "page_number", + "bbox": [ + 0.479, + 0.945, + 0.521, + 0.958 + ], + "angle": 0, + "content": "27170" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.09, + 0.092, + 0.482, + 0.198 + ], + "angle": 0, + "content": "Street-Level Sequences dataset [30]. This Mapillary dataset spans diverse urban and suburban environments and captures a wide range of seasons, weather conditions, cameras, lighting, and structural settings, with each image geo-tagged by GPS and compass angle. Using a similar filtering approach to the VisymScenes dataset, we mined 1,500 positive and 1,500 negative pairs as an out-of-domain test set." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.198, + 0.482, + 0.288 + ], + "angle": 0, + "content": "For the scene reconstruction task, we evaluate on 16 scenes sampled from Heinly et al. [12], Wilson et al. [31], MegaScenes [26] and 5 VisymScenes test scenes. These scenes are challenging for conventional SfM pipelines as well as prior disambiguation methods due to subtle differences between distinct surfaces and repetitive patterns." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.293, + 0.482, + 0.384 + ], + "angle": 0, + "content": "Metrics. For pairwise visual disambiguation evaluation, we report Average Precision (AP) and ROC AUC scores, following the protocol in [3]. Additionally, we report precision at a fixed recall and recall at a fixed precision. These statistics help characterize a model's trade-off between achieving high precision and potentially missing true positives." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.385, + 0.483, + 0.595 + ], + "angle": 0, + "content": "To comprehensively evaluate SfM reconstructions, we report the number of images in the SfM results, and the geo-alignment inlier ratio described in Sec. 3.3. Respectively, more images in the SfM model imply less false pair pruning, i.e., better class separation; and a higher geo-alignment inlier ratio suggests that the reconstructed model is more likely to be accurate and complete. Note that the inlier ratio is also influenced by the accuracy of the geo-tagged images collected from Mapillary. Thus, if one test scene has more accurate geo-tagged images than another, it is likely to yield a higher inlier ratio. Therefore, rather than comparing inlier ratios across different scenes, this metric is more useful for evaluating different reconstruction methods on the same scene, as the same set of geo-tagged images is used." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.6, + 0.483, + 0.706 + ], + "angle": 0, + "content": "Baselines. Our primary baseline is [3], as their learning-based approach consistently outperforms heuristic-based methods like thresholding on the number of post-RANSAC inliers (e.g. SIFT [17] + RANSAC [9], D2-Net [8] + RANSAC, SuperPoint [6] + SuperGlue [22]) for pairwise visual disambiguation; and is also superior than previous approaches [5, 12, 31, 32] for SfM disambiguation." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.713, + 0.249, + 0.729 + ], + "angle": 0, + "content": "4.2. Implementation" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.735, + 0.483, + 0.901 + ], + "angle": 0, + "content": "Model details. Our doppelganger classifier head \\(\\mathrm{Head}_{\\mathrm{dopp}}\\) is implemented with a Transformer encoder comprised of 3 layers, each with input and output dimension of 768. Each layer has 8 attention heads and a two-layer feed-forward network with a hidden dimension of 2048. Pre-layer normalization and residual connections are applied within each layer. The transformed tokens are aggregated via max pooling and linearly projected into \\(\\mathrm{pred}^v\\). We freeze the weights of the public MAST3R [16] model and only train the two doppelganger classification heads, supervised by a cross-entropy loss. We train for 5 epochs with a batch size of 8" + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.092, + 0.905, + 0.122 + ], + "angle": 0, + "content": "using Adam [15] with a learning rate of \\( 1 \\times 10^{-4} \\). More details can be found in the supplementary." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.129, + 0.907, + 0.372 + ], + "angle": 0, + "content": "Integration with SfM. Following [3], we integrate Doppelgangers++ into SfM to enhance its disambiguation ability. SfM takes a collection of images \\(\\mathcal{I} = \\{I_i\\}_{i=0}^n\\) and generates geometrically verified image pairs \\(\\mathcal{P} = \\{(I_p, I_q)\\}\\), forming a scene graph \\(\\mathcal{G} = (\\mathcal{I}, \\mathcal{P})\\) with images as nodes and pairs as edges. Using \\(\\mathcal{G}\\), SfM computes camera poses and reconstructs a 3D point cloud. Our doppelganger classifier acts as a filter on edges in \\(\\mathcal{G}\\), removing pairs below a probability threshold \\(\\tau\\) to eliminate spurious matches due to repeated or symmetric structures. We integrate Doppelgangers++ into both COLMAP [23] and MASt3R-SfM [7] to showcase the effectiveness of the method. Notably, with COLMAP, SIFT features are computed for scene graph construction and MASt3R features for pruning; whereas with MASt3R-SfM, MASt3R features serve for both scene graph construction and pruning, forming a more efficient pipeline." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.381, + 0.795, + 0.397 + ], + "angle": 0, + "content": "4.3. Pairwise Visual Disambiguation" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.403, + 0.907, + 0.569 + ], + "angle": 0, + "content": "We compare our approach with the original Doppelganger work [3] (dubbed \\( DG-OG \\)) under two experimental settings: 1) we use the same training data as in [3], i.e. the DG training set, to train our model. Both models are evaluated on three benchmark test sets: DG, VisymScenes, and Mapillary. Under this configuration, the DG test set is in-domain, while VisymScenes and Mapillary are out-of-domain scenarios; 2) we expand the training data by including the new VisymScenes training set and retrain both DG-OG and our model. Evaluation is conducted on the same three test sets, with only the Mapillary test set being out-of-domain this time." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.569, + 0.906, + 0.719 + ], + "angle": 0, + "content": "Quantitative results are in Tab. 1. Under the first setting, where both methods are trained on DG, our model shows clear improvements across all three test sets. Specifically, on the in-domain DG test set, our model achieves higher AP and ROC AUC, indicating that it maintains high precision across all recall levels, while also being less sensitive to the decision threshold \\(\\tau\\) with better class separation. On the out-of-domain test sets (VisymScenes and Mapillary), we observe \\(25\\%\\) to \\(65\\%\\) improvements in both AP and ROC AUC, highlighting the generalizability of our method." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.72, + 0.907, + 0.901 + ], + "angle": 0, + "content": "Although our model demonstrates improved performance even when trained solely on the DG dataset, its precision outside of the training domain is suboptimal for practical use in SfM, where high precision is essential. In the second setting, we expand the training set by including VisymScenes data. Both methods maintain similar performance on the DG test set, with our model experiencing a slight drop in recall when precision is set to 0.99. On the VisymScenes test set, both models show improvements across all metrics, with ours reaching \\(99\\%\\) in both AP and ROC AUC. Notably, adding VisymScenes to the training set does not enhance DG-OG's performance on the out-of-domain Mapillary test" + }, + { + "type": "page_number", + "bbox": [ + 0.479, + 0.945, + 0.518, + 0.957 + ], + "angle": 0, + "content": "27171" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.097, + 0.089, + 0.302, + 0.284 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.315, + 0.089, + 0.484, + 0.284 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.495, + 0.089, + 0.904, + 0.284 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.09, + 0.287, + 0.907, + 0.342 + ], + "angle": 0, + "content": "Figure 5. SfM Disambiguation on MegaScenes [26]. (White background) SfM results from DG-OG [3] and ours. (Black background) Verification using geo-tagged images, red points represent registered cameras and green points represent geolocations, inlier ratio (IR) is labeled on the bottom right. DG-OG fails to disambiguate this scene, predicting incorrect scores for image pairs. Our method correctly splits the model into two clean components." + }, + { + "type": "image", + "bbox": [ + 0.093, + 0.347, + 0.482, + 0.497 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.09, + 0.5, + 0.483, + 0.556 + ], + "angle": 0, + "content": "Figure 6. SfM disambiguation on VisymScenes. We show that our classifier is more robust than DG-OG [3] on test scenes from new domains, like everyday street scenes. DG-OG has difficulty disambiguating such scenes, leading to incorrect geometry." + }, + { + "type": "table", + "bbox": [ + 0.094, + 0.564, + 0.483, + 0.652 + ], + "angle": 0, + "content": "
Test DataMethodMetrics (trained on DG / trained on DG + VisymScenes)
AP↑ROC AUC↑Prec@Recall=0.85↑Recall@Prec=0.99↑
DGDG-OG0.954 / 0.9560.944 / 0.9470.901 / 0.9100.611 / 0.614
Ours0.980 / 0.9810.981 / 0.9810.972 / 0.9820.690 / 0.642
VisymScenesDG-OG0.816 / 0.9380.726 / 0.9210.498 / 0.8310.340 / 0.623
Ours0.936 / 0.9910.909 / 0.9900.892 / 0.9990.542 / 0.901
MapillaryDG-OG0.566 / 0.6920.581 / 0.7010.523 / 0.5720.003 / 0.000
Ours0.950 / 0.9680.944 / 0.9580.927 / 0.9420.310 / 0.736
" + }, + { + "type": "table_caption", + "bbox": [ + 0.09, + 0.657, + 0.485, + 0.754 + ], + "angle": 0, + "content": "Table 1. Evaluation of pairwise disambiguation. We evaluate DG-OG [3] and our method trained on DG [3] only and DG +VisymScenes (two numbers per cell) on three test sets. Both DG-OG and ours benefit from dataset expansion, whereas ours gained more generalizability on out-of-domain test set (Mapillary) after training on both. Our classifier constantly demonstrates better precision, recall across all settings." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.763, + 0.484, + 0.825 + ], + "angle": 0, + "content": "set, while ours continues to benefit from increased training diversity. These results indicate that the prior approach struggles to generalize effectively, whereas ours shows greater generalization with more varied training data." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.833, + 0.427, + 0.851 + ], + "angle": 0, + "content": "4.4. Structure from Motion disambiguation" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.856, + 0.486, + 0.901 + ], + "angle": 0, + "content": "We integrate our classifier trained on DG and VisymScenes into COLMAP's SfM pipeline [23], and evaluate its performance on reconstructing scenes with duplicated and symmet" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.35, + 0.908, + 0.592 + ], + "angle": 0, + "content": "ric structures. We compare the results with vanilla COLMAP and Cai et al. [3] (DG-OG). Quantitative results are presented in Tab. 2. Our approach registers more images across all scenes than DG-OG and operates without threshold tuning. In contrast, DG-OG relies on scene-specific thresholds, such as \\(\\tau = 0.97\\) for the Church on Spilled Blood to correctly separate both sides of the church, and \\(\\tau = 0.6\\) for Ponte di Rialto [26] to avoid over-segmentation and maintain completeness. Notably, DG-OG fails to disambiguate Belvedere (Vienna), while our method succeeds with a consistent threshold (\\(\\tau = 0.8\\) across all scenes). Our approach also consistently achieves a higher inlier ratio than the baselines, indicating greater accuracy and completeness in the reconstructed models. As an example, we qualitatively show the results of reconstructing Belvedere (Vienna) in Fig. 5, along with verification results using geo-tagged images." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.593, + 0.909, + 0.728 + ], + "angle": 0, + "content": "We also evaluate on scenes from the VisymScenes test set. Since VisymScenes images come with geolocation metadata, we do not need additional Mapillary images; instead, we directly compute the inlier ratio between SfM camera positions and the geolocation metadata after RANSAC. Results show that our model effectively prunes spurious pairs and improves the reconstructed model quality, whereas DG-OG sometimes fails to distinguish doppelganger pairs. Fig. 6 shows examples where DG-OG encounters difficulties." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.73, + 0.909, + 0.82 + ], + "angle": 0, + "content": "In Fig. 7, we show that while MASt3R's features are powerful, MASt3R-SfM is not free from doppelganger issues. Although our classifier was trained on image pairs mined through COLMAP's feature matching module (i.e. w/ SIFT features), it effectively prunes incorrect matches generated by MASt3R, restoring accurate reconstruction results." + }, + { + "type": "title", + "bbox": [ + 0.514, + 0.832, + 0.618, + 0.848 + ], + "angle": 0, + "content": "4.5. Ablation" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.856, + 0.908, + 0.901 + ], + "angle": 0, + "content": "In this section, we study the effectiveness of our designs from the following aspects: 1) fine-tuning the entire model (vs. only new heads), 2) one classification head (vs. two" + }, + { + "type": "page_number", + "bbox": [ + 0.479, + 0.945, + 0.521, + 0.958 + ], + "angle": 0, + "content": "27172" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.094, + 0.089, + 0.482, + 0.202 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.09, + 0.204, + 0.483, + 0.259 + ], + "angle": 0, + "content": "Figure 7. MATLAB's Doppelgangers++. MATLAB also suffers from doppelganger issues. Our classifier effectively prunes false positive pairs and correctly reconstructs challenging scenes." + }, + { + "type": "table", + "bbox": [ + 0.095, + 0.266, + 0.485, + 0.45 + ], + "angle": 0, + "content": "
Test Scenes# SFM-registered ImagesInlier Ratio (Sec. 3.3)
COLMAPDG-OGOursCOLMAPDG-OGOurs
Alexander Nevsky Cathedral [12]4474444470.5651.01.0
Arc de Triomphé [12]4243844230.6970.9660.990
Berliner Dom [12]1603158816060.7090.9870.992
Big Ben [12]3983793940.5640.8270.831
Church on Spilled Blood [12]27384+94 (τ=0.97)157+1060.5420.8810.962
Radcliffe Camera [12]28191+8494+1860.4950.9550.970
Seville [12]1498585+272+515615+303+5520.4500.7720.854
Brandenburg Gate [26]21371361+5701398+6030.4400.9000.909
Palacio de Comunicaciones [26]727307+80 (τ=0.6)308+840.2290.8230.934
Ponte di Rialto [26]652538+101 (τ=0.6)540+1070.6270.8340.844
York Minster [26]636200+362206+2840.7270.8580.901
Town Hall Tower, Kraków [26]2982552800.6090.7310.838
Belvedere (Vienna) [26]1038851 (fail)457+5000.5210.4510.874
Reichstag (building) [26]1504997+3101024+3560.4690.8040.862
St. Vitus Cathedral [26]7526736920.8530.9090.933
Royal Liver Building [26]2121711800.71.01.0
VisymSite00101544520+29014460.7700.8200.913
VisymSite0023849471+81566+820.8670.8480.942
VisymSite0028450238+179267+1200.8180.7110.909
VisymSite0042540206+2074670.8630.9240.963
VisymSite01091245237+458+78239+612+1270.8570.8620.927
" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.46, + 0.484, + 0.57 + ], + "angle": 0, + "content": "Table 2. Evaluation of SfM results. We compare reconstructions from COLMAP [23], DG-OG [3], and our method. \\(\\tau = 0.8\\) is used unless otherwise stated. The \\(^+\\) indicates split reconstruction components; e.g., DG-OG and our method split the Radcliffe Camera reconstruction into two components. Because VisymScenes scenes are large, we report statistics on the largest reconstruction component produced by COLMAP, and identify the corresponding (split) components in the results of DG-OG and ours." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.571, + 0.484, + 0.662 + ], + "angle": 0, + "content": "separate heads), 3) the architecture of our classification head, and 4) final-layer decoder features (vs. multi-layer decoder features). We evaluate models trained on DG and Visym-Scenes datasets and show PR curves in Fig. 8 on the three test sets with respect to the pairwise classification task. A full ablation table can be found in the supplementary." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.668, + 0.484, + 0.851 + ], + "angle": 0, + "content": "Fine-tuning MASt3R. As discussed in Sec.3.2, we choose to train a lighter output head on top of MASt3R features rather than fine-tuning the entire model weights. Curves in Fig. 8 show that training only the head achieves comparable performance to full model fine-tuning. The drop in recall with full model fine-tuning suggests potential overfitting to training data, making the model more conservative and less generalizable to unseen inputs. This tendency also holds on models trained solely on DG data: on the VisymScenes test set, our approach achieved 0.03 higher AP than fine-tuning. On Mapillary test set, ours achieved comparable AP but with 0.2 higher recall at precision 0.99." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.856, + 0.484, + 0.901 + ], + "angle": 0, + "content": "Single classification head. Our design uses two classification heads to process features from the two branches separately. An alternative approach would be to combine the" + }, + { + "type": "image", + "bbox": [ + 0.52, + 0.091, + 0.905, + 0.201 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.513, + 0.205, + 0.907, + 0.248 + ], + "angle": 0, + "content": "Figure 8. Precision-Recall curves of ablation studies. Metrics are evaluated on models trained with DG and VisymScenes. Full tables can be found in the supplementary." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.255, + 0.907, + 0.392 + ], + "angle": 0, + "content": "features from both decoder branches and use a single head to predict the doppelganger score. As shown in Fig. 8, this alternative degrades performance consistently. This may be due to the asymmetric nature of the MAST3R design, where each branch captures different kinds of information that are better to be processed separately. On the other hand, our voting mechanism between the two heads helps enhance class distinction, outputting lower scores for negative samples and higher scores for positive samples." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.399, + 0.909, + 0.506 + ], + "angle": 0, + "content": "Head architecture. While the decoder features of MASt3R retain rich geometric 3D information, designing a classifier with sufficient capacity to effectively leverage this information for the doppelganger prediction task is non-trivial. We compare the performance of our Transformer-based classifier with that of an MLP. Fig. 8 shows that our Transformer model outperforms the MLP across all metrics and test sets." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.513, + 0.909, + 0.62 + ], + "angle": 0, + "content": "Single-layer decoder feature. Our classifier head takes multi-layer decoder features as input. Here we substitute with only the final-layer decoder feature. Results in Fig. 8 show that the multi-layer decoder features are consistently superior to the single-layer ones, because the classifier is able to analyze doppelganger factors from different aspects and levels. See supplementary for more detailed analysis." + }, + { + "type": "title", + "bbox": [ + 0.514, + 0.634, + 0.634, + 0.65 + ], + "angle": 0, + "content": "5. Conclusion" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.66, + 0.909, + 0.901 + ], + "angle": 0, + "content": "We propose Doppelgangers++ as an effective approach to handling visual aliasing in 3D reconstruction. We introduce a new VisymScenes dataset, featuring images from diverse daily scenes, and develop rules to mine doppelganger data from the dataset to enrich our training data. We train a Transformer-based classifier that leverages geometric 3D features to classify image pairs with high precision and recall across various scenes. Doppelgangers++ integrates seamlessly into existing SfM pipelines, enhancing reconstruction accuracy without extensive parameter tuning. Further, we propose a validation method using geo-tagged map images, offering a more comprehensive and automatic way to assess SfM accuracy and model completeness. Extensive experiments show that Doppelgangers++ significantly improves visual disambiguation and 3D reconstruction quality in complex scenes." + }, + { + "type": "page_number", + "bbox": [ + 0.479, + 0.945, + 0.521, + 0.958 + ], + "angle": 0, + "content": "27173" + } + ], + [ + { + "type": "title", + "bbox": [ + 0.091, + 0.091, + 0.264, + 0.108 + ], + "angle": 0, + "content": "6. Acknowledgment" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.116, + 0.486, + 0.328 + ], + "angle": 0, + "content": "We thank Brandon Li and Joseph Tung for their valuable discussion and help with the webviewer. This work was supported by the Intelligence Advanced Research Projects Activity (IARPA) via Department of Interior/Interior Business Center (DOI/IBC) contract number 140D0423C0035. The U.S. Government is authorized to reproduce and distribute reprints for governmental purposes notwithstanding any copyright annotation thereon. Disclaimer: The views and conclusions contained herein are those of the authors and should not be interpreted as necessarily representing the official policies or endorsements, either expressed or implied, of IARPA, DOI/IBC, or the U.S. Government. This research was also supported in part by the National Science Foundation under award IIS-2212084." + }, + { + "type": "title", + "bbox": [ + 0.093, + 0.345, + 0.188, + 0.36 + ], + "angle": 0, + "content": "References" + }, + { + "type": "ref_text", + "bbox": [ + 0.1, + 0.37, + 0.486, + 0.452 + ], + "angle": 0, + "content": "[1] Eric Brachmann, Jamie Wynn, Shuai Chen, Tommaso Cavallari, Aron Monszpart, Daniyar Turmukhambetov, and Victor Adrian Prisacariu. Scene coordinate reconstruction: posing of image collections via incremental learning of a relocalizer. In Eur. Conf. Comput. Vis., pages 421-440. Springer, 2025. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.101, + 0.456, + 0.485, + 0.497 + ], + "angle": 0, + "content": "[2] J. Byrne, G. Castanon, Z. Li, and G. Ettinger. Fine-grained activities of people worldwide. In IEEE/CVF Winter Conference on Applications of Computer Vision, 2023. 2, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.499, + 0.485, + 0.554 + ], + "angle": 0, + "content": "[3] Ruojin Cai, Joseph Tung, Qianqian Wang, Hadar Averbuch-Elor, Bharath Hariharan, and Noah Snavely. Doppelgangers: Learning to disambiguate images of similar structures. In Int. Conf. Comput. Vis., pages 34-44, 2023. 1, 2, 3, 5, 6, 7, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.557, + 0.484, + 0.611 + ], + "angle": 0, + "content": "[4] Duygu Ceylan, Niloy Jyoti Mitra, Youyi Zheng, and Mark Pauly. Coupled structure-from-motion and 3d symmetry detection for urban facades. ACM Transactions on Graphics (TOG), 33:1-15, 2014. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.615, + 0.484, + 0.656 + ], + "angle": 0, + "content": "[5] Zhaopeng Cui and Ping Tan. Global structure-from-motion by similarity averaging. 2015 IEEE International Conference on Computer Vision (ICCV), pages 864-872, 2015. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.659, + 0.484, + 0.713 + ], + "angle": 0, + "content": "[6] Daniel DeTone, Tomasz Malisiewicz, and Andrew Rabinovich. Superpoint: Self-supervised interest point detection and description. In Proc. CVPR Workshops, pages 224-236, 2018. 2, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.716, + 0.484, + 0.771 + ], + "angle": 0, + "content": "[7] Bardienus Pieter Duisterhof, Lojze Žust, Philippe Weinzaepfel, Vincent Leroy, Yohann Cabon, and Jérôme Revaud. Mast3r-sfm: a fully-integrated solution for unconstrained structure-from-motion. ArXiv, abs/2409.19152, 2024. 2, 3, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.774, + 0.484, + 0.843 + ], + "angle": 0, + "content": "[8] Mihai Dusmanu, Ignacio Rocco, Tomás Pajdla, Marc Pollefeys, Josef Sivic, Akihiko Torii, and Torsten Sattler. D2-net: A trainable cnn for joint description and detection of local features. 2019 IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 8084-8093, 2019. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.846, + 0.484, + 0.9 + ], + "angle": 0, + "content": "[9] Martin A. Fischler and Robert C. Bolles. Random sample consensus: a paradigm for model fitting with applications to image analysis and automated cartography. Commun. ACM, 24:381-395, 1981. 6" + }, + { + "type": "list", + "bbox": [ + 0.1, + 0.37, + 0.486, + 0.9 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.093, + 0.908, + 0.161 + ], + "angle": 0, + "content": "[10] Ye Gong, Pengwei Zhou, Changfeng Liu, Yan Yu, Jian Yao, Wei Yuan, and Li Li. A cluster-based disambiguation method using pose consistency verification for structure from motion. ISPRS Journal of Photogrammetry and Remote Sensing, 2024. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.165, + 0.908, + 0.219 + ], + "angle": 0, + "content": "[11] Jared Heinly, Enrique Dunn, and Jan-Michael Frahm. Recovering correct reconstructions from indistinguishable geometry. 2014 2nd International Conference on 3D Vision, 1:377-384, 2014. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.222, + 0.908, + 0.276 + ], + "angle": 0, + "content": "[12] Jared Heinly, Enrique Dunn, and Jan-Michael Frahm. Correcting for duplicate scene structure in sparse 3d reconstruction. In Eur. Conf. Comput. Vis., pages 780-795. Springer, 2014. 6, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.28, + 0.907, + 0.335 + ], + "angle": 0, + "content": "[13] Nianjuan Jiang, Ping Tan, and Loong-Fah Cheong. Seeing double without confusion: Structure-from-motion in highly ambiguous scenes. In Conf. Comput. Vis. Pattern Recog., pages 1458-1465, 2012. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.338, + 0.908, + 0.392 + ], + "angle": 0, + "content": "[14] Rajbir Kataria, Joseph DeGol, and Derek Hoiem. Improving structure from motion with reliable resectioning. 2020 International Conference on 3D Vision (3DV), pages 41-50, 2020. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.396, + 0.907, + 0.422 + ], + "angle": 0, + "content": "[15] Diederik P. Kingma and Jimmy Ba. Adam: A method for stochastic optimization. CoRR, abs/1412.6980, 2014. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.425, + 0.908, + 0.466 + ], + "angle": 0, + "content": "[16] Vincent Leroy, Yohann Cabon, and Jérôme Revaud. Grounding image matching in 3d with mast3r. ArXiv, abs/2406.09756, 2024. 2, 3, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.47, + 0.908, + 0.509 + ], + "angle": 0, + "content": "[17] David G Lowe. Distinctive image features from scale-invariant keypoints. Int. J. Comput. Vis., 60:91-110, 2004. 2, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.513, + 0.908, + 0.569 + ], + "angle": 0, + "content": "[18] Lalit Manam and Venu Madhav Govindu. Leveraging camera triplets for efficient and accurate structure-from-motion. 2024 IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 4959-4968, 2024. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.571, + 0.907, + 0.625 + ], + "angle": 0, + "content": "[19] Yang Peng, Shen Yan, Yuxiang Liu, Y. Liu, and Maojun Zhang. View graph construction for scenes with duplicate structures via graph convolutional network. IET Comput. Vis., 16:389-402, 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.628, + 0.907, + 0.67 + ], + "angle": 0, + "content": "[20] Richard Roberts, Sudipta N. Sinha, Richard Szeliski, and Drew Steedly. Structure from motion for scenes with large duplicate structures. CVPR 2011, pages 3137-3144, 2011. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.673, + 0.908, + 0.727 + ], + "angle": 0, + "content": "[21] Ignacio Rocco, Mircea Cimpoi, Relja Arandjelović, Akihiko Torii, Tomas Pajdla, and Josef Sivic. Neighbourhood consensus networks. In Adv. Neural Inform. Process. Syst., 2018. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.731, + 0.908, + 0.798 + ], + "angle": 0, + "content": "[22] Paul-Edouard Sarlin, Daniel DeTone, Tomasz Malisiewicz, and Andrew Rabinovich. Superglue: Learning feature matching with graph neural networks. 2020 IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 4937–4946, 2019. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.802, + 0.908, + 0.843 + ], + "angle": 0, + "content": "[23] Johannes L Schonberger and Jan-Michael Frahm. Structure-from-motion revisited. In Conf. Comput. Vis. Pattern Recog., pages 4104-4113, 2016. 2, 3, 5, 6, 7, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.846, + 0.908, + 0.899 + ], + "angle": 0, + "content": "[24] Tianwei Shen, Siyu Zhu, Tian Fang, Runze Zhang, and Long Quan. Graph-based consistent matching for structure-from-motion. In European Conference on Computer Vision, 2016. 2" + }, + { + "type": "list", + "bbox": [ + 0.517, + 0.093, + 0.908, + 0.899 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.48, + 0.945, + 0.521, + 0.957 + ], + "angle": 0, + "content": "27174" + } + ], + [ + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.093, + 0.484, + 0.147 + ], + "angle": 0, + "content": "[25] Jiaming Sun, Zehong Shen, Yuang Wang, Hujun Bao, and Xiaowei Zhou. Loftr: Detector-free local feature matching with transformers. In Conf. Comput. Vis. Pattern Recog., pages 8922-8931, 2021. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.149, + 0.484, + 0.204 + ], + "angle": 0, + "content": "[26] Joseph Tung, Gene Chou, Ruojin Cai, Guandao Yang, Kai Zhang, Gordon Wetzstein, Bharath Hariharan, and Noah Snavely. Megascenes: Scene-level view synthesis at scale. ArXiv, abs/2406.11819, 2024. 5, 6, 7, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.206, + 0.484, + 0.26 + ], + "angle": 0, + "content": "[27] Fan Wang, Aditi Nayak, Yogesh Agrawal, and Roy Shilkrot. Hierarchical image link selection scheme for duplicate structure disambiguation. In British Machine Vision Conference, 2018. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.262, + 0.484, + 0.318 + ], + "angle": 0, + "content": "[28] Jianyuan Wang, Nikita Karaev, Christian Rupprecht, and David Novotny. Vggsfm: Visual geometry grounded deep structure from motion. In Conf. Comput. Vis. Pattern Recog., pages 21686-21697, 2024. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.32, + 0.484, + 0.374 + ], + "angle": 0, + "content": "[29] Shuzhe Wang, Vincent Leroy, Yohann Cabon, Boris Chidlovskii, and Jérôme Revaud. DUS3R: Geometric 3d vision made easy. In Conf. Comput. Vis. Pattern Recog., pages 20697-20709, 2023. 2, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.376, + 0.484, + 0.444 + ], + "angle": 0, + "content": "[30] Frederik Warburg, Søren Hauberg, Manuel López-Antequera, Pau Gargallo, Yubin Kuang, and Javier Civera. Mapillary street-level sequences: A dataset for lifelong place recognition. In Conf. Comput. Vis. Pattern Recog., pages 2623-2632, 2020. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.446, + 0.484, + 0.488 + ], + "angle": 0, + "content": "[31] Kyle Wilson and Noah Snavely. Network principles for sfm: Disambiguating repeated structures with local context. In Int. Conf. Comput. Vis., pages 513-520, 2013. 2, 6, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.489, + 0.484, + 0.544 + ], + "angle": 0, + "content": "[32] Qingan Yan, Long Yang, Ling Zhang, and Chunxia Xiao. Distinguishing the indistinguishable: Exploring structural ambiguities via geodesic context. In Conf. Comput. Vis. Pattern Recog., pages 3836-3844, 2017. 2, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.546, + 0.484, + 0.587 + ], + "angle": 0, + "content": "[33] Kwang Moo Yi, Eduard Trulls, Vincent Lepetit, and Pascal Fua. Lift: Learned invariant feature transform. In Eur. Conf. Comput. Vis., pages 467-483, 2016. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.589, + 0.484, + 0.642 + ], + "angle": 0, + "content": "[34] Christopher Zach, Arnold Irschara, and Horst Bischof. What can missing correspondences tell us about 3d structure and motion? In Conf. Comput. Vis. Pattern Recog., pages 1-8, 2008. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.645, + 0.484, + 0.698 + ], + "angle": 0, + "content": "[35] Christopher Zach, Manfred Klopschitz, and Marc Pollefeys. Disambiguating visual relations using loop constraints. In Conf. Comput. Vis. Pattern Recog., pages 1426-1433, 2010." + }, + { + "type": "list", + "bbox": [ + 0.093, + 0.093, + 0.484, + 0.698 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.48, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "27175" + } + ] +] \ No newline at end of file diff --git a/2025/Doppelgangers++_ Improved Visual Disambiguation with Geometric 3D Features/a4009708-9357-4078-b5aa-d3dd7d42ccfe_origin.pdf b/2025/Doppelgangers++_ Improved Visual Disambiguation with Geometric 3D Features/a4009708-9357-4078-b5aa-d3dd7d42ccfe_origin.pdf new file mode 100644 index 0000000000000000000000000000000000000000..41fa0b396b8513a0bae58aa91796bd90652e9eff --- /dev/null +++ b/2025/Doppelgangers++_ Improved Visual Disambiguation with Geometric 3D Features/a4009708-9357-4078-b5aa-d3dd7d42ccfe_origin.pdf @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:627fcc37ad6726557825b323b81d561af6d2f9318ab6cae650f6810e88733dfd +size 9633587 diff --git a/2025/Doppelgangers++_ Improved Visual Disambiguation with Geometric 3D Features/full.md b/2025/Doppelgangers++_ Improved Visual Disambiguation with Geometric 3D Features/full.md new file mode 100644 index 0000000000000000000000000000000000000000..3254a65b01d6af119082af4a526450264618ed35 --- /dev/null +++ b/2025/Doppelgangers++_ Improved Visual Disambiguation with Geometric 3D Features/full.md @@ -0,0 +1,292 @@ +# Doppelgangers++: Improved Visual Disambiguation with Geometric 3D Features + +Yuanbo Xiangli $^{1}$ Ruojin Cai $^{1}$ Hanyu Chen $^{1}$ Jeffrey Byrne $^{2}$ Noah Snavely $^{1}$ $^{1}$ Cornell University, $^{2}$ Visym Labs + +![](images/e85edf79ca55a027c11e983dc0b5bda38223da24909f140773a47b053f71298d.jpg) +Ours +MASt3R-SfM + +![](images/4e5fb53f3d1dcbfeba8393412990b6ff3f14326b3a0753ea848f21bfe1df9771.jpg) + +![](images/e4e598dc67129d2fb7db66cfac2036a5526427222bd768459652bcf03d2848b2.jpg) + +![](images/a775f643c3ca9e057108e5669dccc344dd60c4ae376da5f03ff7ea558e0df5b8.jpg) + +![](images/f27f03a87bdd3540e70c2ef6fd0d9d4cb1596908e4bf0307f96ddbc5067d71ba.jpg) + +![](images/d8d43376eb2bf896d3073d92be4a6b4bb91c3f7465b7bc0a37d6dabb87dfc630.jpg) + +![](images/9bbd5704ad5623e79dcabb1dc7a6dd141ee34783f30333be0fb46dc74aaca4e1.jpg) + +![](images/b4cefdff827cef6d22b8bb2825c6a9d1851e8173c0ff1d20bad5d04f2f39a698.jpg) + +![](images/2a29b123581a9ed8c5064120e1e7b2f70829a9df2b3f68657a422658cbb3fa4e.jpg) + +![](images/82245e7a93305f40b2317a4271b3bf36d115e47e10a2060cfd43ba16f2511a63.jpg) + +![](images/be983518280b036162859f3bf77d472732559cbe51f366de7f5f07024d35107a.jpg) +:0.235/Ours:0.982 + +![](images/c965f37db1a78807f9458f024cb65e0de1cc2d90526d02366bd5366958e70bc7.jpg) +GT:1/DG-OG:0.190/Ours:0.971 + +![](images/8e44a4681b9be46df7d208feb798186e47a1f2235dbe54a6ccb05945f1e966c9.jpg) +GT:1 / DG-OG:0.378 / Ours:0.952 + +![](images/e64a8c415f17ac4ac9028ffa78be166e7b522693c7a24c16cd7f1967857a1b92.jpg) +GT:0/DG-OG:0.993/Ours:0.025 + +![](images/5c5cbc9397bdc7b929332a81bc456c37df569d27cece026bb5740eb307fb0c37.jpg) +GT:0/DG-OG:0.994/Ours:0.031 + +![](images/924a8f102cd310d18cf5fcf83ce9b420a38a052ca7731dc5e3fdd37678d905a1.jpg) +Figure 1. Visual aliasing, or doppelgangers, poses severe challenges to 3D reconstruction. We propose Doppelganger++, an enhanced pairwise image classifier that excels in visual disambiguation across diverse and challenging scenes. (Top) We seamlessly integrate Doppelganger++ into SfM, successfully disambiguating each scene. (Middle) Compared to prior work (which we refer to as DG-OG [3]), Doppelgangers++ is more robust for everyday scenes, showing improved accuracy and robustness. We show pairs that DG-OG classify incorrectly and ours gets correct. (Bottom) Our new VisymScenes dataset, featuring complex daily scenes, is particularly challenging for COLMAP and DG-OG, but our method can achieve correct and complete reconstructions. + +![](images/b53bc3b3b17778f7d7362d72012a622a2214797fd2f2d21ad6eb272cfedace40.jpg) + +![](images/84d7daad6b17fa4886892beac441f3543e868663fcea211d1b2ecb1055f228e6.jpg) + +![](images/d2f9bc25d12efb4ec378d61525efbd7b1d03b725f414aee84c5ad4351ae296ab.jpg) + +# Abstract + +Accurate 3D reconstruction is frequently hindered by visual aliasing, where visually similar but distinct surfaces (aka, doppelgangers), are incorrectly matched. These spurious matches distort the structure-from-motion (SfM) process, leading to misplaced model elements and reduced accuracy. Prior efforts addressed this with CNN classifiers trained on curated datasets, but these approaches struggle to generalize across diverse real-world scenes and can require extensive parameter tuning. In this work, we present Doppelgangers++, a method to enhance doppelganger detection and improve 3D reconstruction accuracy. Our contribu + +tions include a diversified training dataset that incorporates geo-tagged images from everyday scenes to expand robustness beyond landmark-based datasets. We further propose a Transformer-based classifier that leverages 3D-aware features from the MAST3R model, achieving superior precision and recall across both in-domain and out-of-domain tests. Doppelgangers++ integrates seamlessly into standard SfM and MAST3R-SfM pipelines, offering efficiency and adaptability across varied scenes. To evaluate SfM accuracy, we introduce an automated, geotag-based method for validating reconstructed models, eliminating the need for manual inspection. Through extensive experiments, we demonstrate that Doppelgangers++ significantly enhances pairwise vi + +sual disambiguation and improves 3D reconstruction quality in complex and diverse scenarios. + +# 1. Introduction + +Visual aliasing—confusing two surfaces that look the same but are nonetheless distinct—is an pernicious problem in 3D reconstruction and SLAM systems. Pairs of images that depict visually similar yet distinct 3D surfaces (called doppelgangers by Cai et al. [3]) can generate spurious correspondence at the feature matching stage of 3D reconstruction, leading to erroneous downstream reconstructions that feature distorted geometry or incorrectly fused elements. Therefore, to ensure the accuracy of 3D reconstruction, it is critical to distinguish truly matching images from illusory matches arising from doppelganger pairs. + +In recent work, this visual disambiguation problem was formulated as a binary classification task on image pairs [3]. The authors collected a Internet dataset of visually similar doppelganger pairs (as well as truly matching images) from Wikipedia Commons, then trained a CNN to classify image pairs as correct or incorrect matches. This classifier can be used to take a feature match graph computed from a set of photos, remove incorrect edges between doppelganger image pairs, then reconstruct a correct model using a structure from motion (SfM) pipeline like COLMAP [23]. While that work shows promising improvements on reconstruction problems, we find that it can still be brittle: First, the 3D reconstruction task demands that the classifier have has precision—even a few bad edges remaining in the image match graph can lead to an incorrect 3D model. Second, the reconstruction task is quite sensitive to the threshold on the classifier score used to prune edges from the match graph. Finally, their method was trained solely on landmark Internet photos and does not reliably generalize to new scenarios, such as more structured captures of everyday scenes, like streets and office buildings. + +In this paper, we aim to address these issues and improve doppelganger classification in several key ways: + +1. We identify ways to expand and diversify doppelgangers training data. In particular, we leverage semi-structured image data with geographic annotations (rough GPS position and orientation), captured from everyday scenes with the Visym Collector platform [2]. +2. We switch from a CNN-based classifier to leveraging features from MASt3R [7], a recent transformer-based geometric model that computes point clouds from two input views. Specifically, we feed an image pair into a pre-trained MASt3R model, collect the intermediate features decoded by MASt3R, and train a classification head to map these features to a doppelganger score. + +We call our method Doppelgangers++. Our enhanced model achieves higher precision and recall in pairwise classification, and generalizes better across a broader range of scenes and + +capture scenarios. Doppelganger++ integrates seamlessly into SfM pipelines for 3D reconstruction disambiguation. We further propose to leverage geo-tagged map images to quantitatively evaluate the correctness and completeness of reconstructed models, in comparison to manual inspection as required in previous approaches. Through extensive experiments we show that our model leads to more accurate and complete reconstruction results, and is less sensitive to the threshold used for pruning doppelganger matches. + +# 2. Related Work + +Local feature matching and 3D learning. Local feature matching methods have proven effective at establishing correspondences between pairs of images in SfM pipelines. Classic methods like COLMAP [23] rely on the tried-and-true SIFT features [17] to find correspondence. Modern learning-based, data-driven approaches have improved the quality of local feature matching [6, 21, 25, 33]. More recently, the DUSt3R [29] framework has proven to excel in a variety of 3D reconstruction tasks by estimating dense 3D point clouds from pairs of images. MASt3R [16] was proposed as an extension to DUSt3R that specifically targets the task of predicting dense feature matches between image pairs. All of these local feature matching methods are traditionally optimized to maximize the number of correspondences they find between similar image regions. While their ability to identify feature matches has greatly improved, they generally lack the ability to incorporate negative evidence into their predictions, and so they often find matches between regions that do not actually correspond to the same 3D surface, particularly within doppelganger image pairs. However, we show that features internal to MASt3R can be repurposed for doppelganger detection. + +Disambiguation in SfM and image matching. Disambiguating similar structures and repeated patterns in SfM is a long-standing challenge. Prior work has mainly relied on heuristics-based analysis to detect conflicting relations and ambiguities in the structure of the underlying scene graph [10, 14, 18, 19, 24, 27, 31, 32, 35] or among image-level correspondence (or the lack thereof) [4, 11-13, 20, 34]. While such methods have shown some success in detecting and resolving ambiguities in SfM pipelines, one fundamental limitation that they do not consider the rich information contained in the images themselves. In contrast, Doppelgangers [3] avoids handcrafted heuristics and takes a data-driven approach to the basic problem of identifying doppelganger image pairs. They train a CNN that classifies a pair of images as either positive or negative, with LoFTR [25]-extracted keypoints and match masks passed as auxiliary input. During inference, the binary classification results are used to pre-process the scene graph obtained from COLMAP to filter out spurious matches prior to running SfM. How + +ever, the method's reliance on COLMAP feature extraction and matching and the need for auxiliary mask information from LoFTR introduces significant overhead and complexity to the overall pipeline, making it difficult to scale to larger scenes. In addition, we find that this prior method is brittle: it can fail to generalize to domains beyond landmarks (e.g., to office buildings), and it can require parameter tuning to get good results. + +Differentiable SfM. To improve upon traditional SfM, several differentiable SfM pipelines have been proposed to optimize the entire 3D reconstruction process. These pipelines often include one or more components that are trained from large datasets. MAST3R-SfM [7] proposes an SfM pipeline that uses MAST3R features for scene graph construction and coarse-to-fine alignment. VGGSfM [28] decomposes the problem into four stages: point tracking, initial camera estimation, triangulation, and bundle adjustment, each of which is differentiable and can be optimized in an end-to-end manner. ACE0 [1] uses a scene coordinate regression network to progressively register new images to a global scene, while the network itself is iteratively trained and refined from the registered images. These methods have shown promising results in optimizing the SfM pipeline, but are not specifically designed to address the doppelganger problem, and are still prone to producing incorrect reconstructions when faced with scenes with repeated patterns or similar structures. + +# 3. Method + +We refer to a "doppelganger" as a case where distinct objects or surfaces look almost identical, leading algorithms to confuse one for the other [3]. In this work, we aim to improve the doppelganger classifier via expanded and diversified training data (Sec. 3.1), and leveraging geometric 3D features learned from pairwise reconstruction [16] (Sec. 3.2). We further propose an approach to quantitatively evaluate the correctness of SfM results in terms of doppelgangers (Sec. 3.3), instead of manual inspection adopted in previous work. + +# 3.1. The VisymScenes dataset + +Cai et al. [3] introduced the Doppelganger dataset, built on global landmarks photos sourced from Wikipedia Commons, with viewing direction (e.g., North, South) to identify doppelganger image pairs. While the classifier trained on this dataset proved effective, we find that it struggles to generalize well to scenes beyond the dataset's domain. + +We enhance and expand their training set to improve the robustness of doppelganger classification, by incorporating casually captured images from a wider range of scenes. To ensure sufficient diversity in the dataset, we introduce the VisymScenes dataset. VisymScenes consists of 258K ground images with GPS/IMU metadata, recorded at 149 sites in 42 cities and 15 countries, collected with the Visym Collector + +Platform [2] (details in supplementary). Each image is accompanied by metadata, such as GPS coordinates, device compass direction, and intrinsic camera calibration. While this metadata can be noisy, it still provides valuable information for identifying potential doppelganger pairs. For example, if two images exhibit numerous geometrically consistent local feature matches yet were taken from distant locations, this serves as strong evidence that the pair is likely to be a doppelganger. Examples are shown in Fig. 2. + +We develop a series of filtering rules applied to all matched image pairs identified by the COLMAP feature matching module [23]. Note that these rules are designed according to the capture style of the Visym Collector, where cameras consistently focus on the scene of interest and maintain a reasonable distance from it. For a given pair of images, we use the (metric) distance between the camera centers $r$ , angle between their viewing directions $\theta$ , and camera intrinsics $\mathcal{K}$ , all derived from camera metadata, to identify confident negative (doppelganger) and positive (correct match) pairs. To identify confident negative pairs, we first identify spatially distant matching pairs, and classify them as doppelgangers. Then, we classify the remaining (nearby) pairs into three cases: the intersection point of the viewing directions is 1) in front of both cameras, 2) behind both cameras; or 3) in front of one camera and behind the other. For case 1), we label pairs with very large view angles (e.g., $>160^{\circ}$ ) as negative, since they likely capture different 3D surfaces or mostly non-overlapping content. For case 2), if their view angle exceeds the diagonal field of view, then their view frustums are unlikely to overlap, so we label them as doppelgangers. For case 3), we check for frustum overlap using camera intrinsics; if no intersection is detected, the pair is considered a doppelganger. A similar series of rules are designed for mining positive pairs (i.e., true matches). Details of computing view intersection and filtering algorithms for both negative and positive pairs can be found in the supplementary. With these rules, we mined in total 53K positive and negative pairs across 33 sites for our doppelganger task. + +# 3.2. Improved Doppelganger Classifier + +Recent advances in data-driven models [16, 29] have demonstrated impressive results in geometric vision tasks. In our work, we leverage the multi-level, 3D-aware features extracted from the MAST3R model [16] to train a doppelganger classifier on labeled image pairs, as demonstrated in Fig. 3. + +Equipped with large-scale training data and a ViT backbone, the MASt3R model captures an image representation that encodes rich 3D geometric information between paired images. However, as MASt3R was originally trained to detect correspondences and similarities, its point maps often conflate doppelganger pairs, yielding incorrect point clouds and poses for these challenging pairs (examples in Fig. 7). Despite this limitation, we find that the internal + +![](images/bb0850eb322419a465b9ad04e192ebca37cf0e16bf5b0bad9472f617b33cd9d7.jpg) +Figure 2. VisymScenes examples. This new dataset includes residential areas, landmarks, historical sites, business districts, and more. Here, we present four example sites. The top row shows subsets of images captured within each site. The bottom row displays pairs of visually similar but geographically distinct images from each site along with their recorded geolocations on a map. These examples demonstrate that doppelganger issues are prevalent in everyday scenes, presenting significant challenges for reliable 3D reconstruction and image matching. + +![](images/deed301b54334765ad2b9ba161e9c30a6f3c597194057601f07c0103a4e974ff.jpg) + +![](images/58cd38fd1978cd5e8d91d604f59f38c8dddfbebcb144dbd2172707c65e7e708b.jpg) + +![](images/a9ec786564a9c2398f43f04595dccc55d98bd1dffabd857bad716672c03cc9ee.jpg) + +![](images/2f09ee2e9191f482c02336660540159a793554b71811d74fd1d530dae681cd7c.jpg) +Figure 3. Model design. (Left) Given an image pair, we first create a symmetrized version of the pair and feed it into the frozen MASt3R model. Multi-layer features are extracted from each decoder branch, concatenated, and fed into two learnable doppelganger classification heads. Each head generates predictions $(\mathrm{pred}_{pq}^{v}, \mathrm{pred}_{qp}^{v})$ , $v \in \{1,2\}$ (where $pq$ and $qp$ denote the symmetrized image pair), supervised by cross-entropy loss. (Right) We use multi-layer decoder features and a Transformer-based classifier head for doppelganger prediction. + +features learned by MASt3R contain sufficient information for doppelganger classification without the need to fine-tune MASt3R's weights. Moreover, given that the amount of available training data for the doppelganger task is dwarfed by the amount of data used to train MASt3R, we opt not to fine-tune the entire model, but instead repurpose these internal features by training an additional output head to predict a doppelganger classification score. In Sec. 4.5, we empirically demonstrate that this choice achieves comparable or better visual disambiguation performance, especially on out-of-domain scenes. + +Multi-level decoder features. For an image pair $(I_p, I_q)$ , we extract multi-level decoder features from MASt3R's decoder branches. MASt3R takes two images as input, and uses two intertwined decoders $\mathrm{DecBlk}^1$ and $\mathrm{DecBlk}^2$ to decode encoded features $\mathcal{H}^1 = \mathrm{Enc}^1(I_p)$ and $\mathcal{H}^2 = \mathrm{Enc}^2(I_q)$ . Each decoder has $B$ blocks, and attends to tokens from the other branch: + +$$ +f _ {i} ^ {1} = \operatorname {D e c B l k} _ {i} ^ {1} \left(f _ {i - 1} ^ {1}, f _ {i - 1} ^ {2}\right), f _ {i} ^ {2} = \operatorname {D e c B l k} _ {i} ^ {2} \left(f _ {i - 1} ^ {2}, f _ {i - 1} ^ {1}\right), \tag {1} +$$ + +where $f_{i}^{v}$ denotes the output tokens from the $i$ -th block of the $v$ -th branch. These two branches exchange information to capture the spatial relationships between viewpoints and the global 3D structure of the scene. For each branch $v$ , we concatenate the encoder feature $\mathcal{H}^v$ and the tokens from the decoder blocks into $\mathcal{F}^v = \mathrm{concat}(\mathcal{H}^v, \{f_i^v\}_{i=0}^{B-1})$ , which captures rich, multi-level spatial correspondence information between image pairs. + +Transformer-based classification heads. MAsst3R treats one image as the reference frame and projects the other image into that reference frame. A consequence of this design is that the reconstructed 3D scene for input pair $(I_p, I_q)$ is distinct (or at least, must be in a different coordinate frame) from that of $(I_q, I_p)$ . Inspired by the asymmetric design, we propose to 1) use separate Transformer heads: We introduce two independent, Transformer-based classification heads $\mathrm{Head}_{dopp}^1$ and $\mathrm{Head}_{dopp}^2$ to process $\mathcal{F}^1$ and $\mathcal{F}^2$ respectively; and 2) symmetrize image pairs such that the model evaluates both $(I_p, I_q)$ and $(I_q, I_p)$ to decide whether the given pair of images are true match or not. Thus, we end up with four scores for each image pair: + +$$ +\operatorname {p r e d} _ {u} ^ {v} = \operatorname {H e a d} _ {\text {d o p p}} ^ {v} \left(\mathcal {F} _ {u} ^ {v}\right), u \in \{p q, q p \}, v \in \{1, 2 \}. \tag {2} +$$ + +Essentially, the two heads serve as expert evaluators, each examining how likely the image pair is to be a true match (or doppelganger) from different aspects, and switching the input order allows the model to analyze the spatial relationships from both directions. We empirically show the effectiveness of this design in Sec. 4.5. Both $\mathrm{Head}_{dopp}^{1}$ and $\mathrm{Head}_{dopp}^{2}$ are supervised by the cross-entropy loss, encouraging $S = \{\mathrm{pred}_u^v\}$ to all give high probabilities for positive matches and low probabilities for negative ones. + +Test-time voting. At test time, we combine the four scores $S$ via a voting mechanism to make a final decision. Specifically, if the majority of the heads predict that the pair is a positive + +match, we take $\max(S)$ as the final score for the image pair. Conversely, if the majority of the heads vote for a negative match, $\min(S)$ is used as the final score. Otherwise, we average the scores from the four heads. This approach ensures that the final decision reflects the strongest evidence supporting the consensus among the heads and thereby improves the reliability of the classifier. Eq. 3 elaborates on this voting mechanism: + +$$ +S _ {\text {f i n a l}} = \left\{ \begin{array}{l l} \max (\mathcal {S}) & \text {i f} \sum_ {s \in \mathcal {S}} \mathbf {1} (s > 0. 5) > \sum_ {s \in \mathcal {S}} \mathbf {1} (s < 0. 5), \\ \min (\mathcal {S}) & \text {i f} \sum_ {s \in \mathcal {S}} \mathbf {1} (s > 0. 5) < \sum_ {s \in \mathcal {S}} \mathbf {1} (s < 0. 5), \\ \operatorname {m e a n} (\mathcal {S}) & \text {i f} \sum_ {s \in \mathcal {S}} \mathbf {1} (s > 0. 5) = \sum_ {s \in \mathcal {S}} \mathbf {1} (s < 0. 5). \end{array} \right. \tag {3} +$$ + +# 3.3. Evaluating Doppelganger correction in SfM + +There is currently no reliable benchmarking method for evaluating the accuracy and correctness of SfM reconstructions specifically in terms of how well they address the doppelganger issue. Prior work [3] relied on manual inspection of each out model to assess the effectiveness of their approach—a time-consuming, unquantifiable, and potentially error-prone process. In our work, we propose a benchmarking method for qualitatively evaluating reconstructed models with respect to doppelgangers. We leverage mapping sites like Mapillary, which provide images with location metadata that can serve as probes for validating a 3D model. Note that our method targets common scenarios where geotags are unavailable, or so noisy as to not be useful. Therefore, we do not consider the use of geotags for the reconstruction task itself—we explore pure visual disambiguation—but instead gather them from specialized sources for ground truth evaluation. + +Specifically, for a given scene with known geolocation, we acquire from Mapillary a set of nearby geo-tagged images and register them to the reconstructed model with COLMAP [23]. We then apply RANSAC to robustly estimate a similarity transformation between the registered camera positions and their corresponding image geolocation metadata (converted to ECEF coordinates). We use the resulting RANSAC inlier ratio as an indicator of the model's correctness. As an example, in Fig. 4, we can observe that before correction, the registered cameras and their geolocations show significant misalignment, with the cameras collapsing to one side due to similar-looking surfaces. After correction, the registered cameras align closely to their true geolocations, indicating that doppelganger pairs are correctly separated in the model. If the model is split into multiple components, we calculate the weighted average of inliers ratios among the components: + +$$ +\operatorname {I R} = \sum_ {i = 1} ^ {N} \left(\frac {I _ {i}}{T _ {i}} \cdot \frac {T _ {i}}{\sum_ {j = 1} ^ {N} T _ {j}}\right) = \frac {\sum_ {i = 1} ^ {N} I _ {i}}{\sum_ {i = 1} ^ {N} T _ {i}}, \tag {4} +$$ + +where $I_{i}$ is the number of RANSAC inliers and $T_{i}$ is the number of registered Mapillary images in the $i$ -th compo + +![](images/62b80a4d91920e948489fbc03315ecd65e5a6cf816cfbc816692fd5ebf430d4e.jpg) + +![](images/91870a5b6242f432624630df8ba963a4b79e7a1846cd3951dd4191b06f7075c8.jpg) +Figure 4. Evaluation of doppelganger correction in SfM. (Top) We first collect sequences of geo-tagged Mapillary images around the target location and register them to the SfM model. Then, we use RANSAC to align the registered cameras and their geolocations. The inlier ratio is computed as an indicator of model accuracy. (Bottom) In the model corrupted by doppelganger pairs, the registered cameras all collapse to one side. We see that the camera poses estimated with COLMAP (right, in red) do not align well with the geotags (green), leading to a low inlier ratio, but our method leads to a much closer alignment. + +nent. While alignment error could also serve as a metric, we opt to use the inlier ratio to reduce sensitivity to geolocation inaccuracies, as image geolocations may not always be precise. This strategy can also be used to detect broken reconstructions from datasets like MegaScenes [26], which contains over 100K SfM results from Internet photos around the world. + +# 4. Experiments + +# 4.1. Experimental setup + +Dataset. Our training set is comprised of the Doppelganger dataset [3] (dubbed DG) and the new VisymScenes dataset. Specifically, the DG training set includes 73K image pairs, nearly evenly split between positive and negative pairs. From VisymScenes, we use 47K image pairs across 26 sites as training data, evenly split into positive and negative pairs. + +We evaluate on two tasks, 1) pairwise image visual disambiguation and 2) scene reconstruction by integrating our classifier into COLMAP [23]. For pairwise image visual disambiguation, we utilize the DG test set, and mined an additional 3,180 pairs from 7 Visym sites (distinct from the 26 training sites), equally divided into positive and negative pairs. Additionally, we created a test set from the Mapillary + +Street-Level Sequences dataset [30]. This Mapillary dataset spans diverse urban and suburban environments and captures a wide range of seasons, weather conditions, cameras, lighting, and structural settings, with each image geo-tagged by GPS and compass angle. Using a similar filtering approach to the VisymScenes dataset, we mined 1,500 positive and 1,500 negative pairs as an out-of-domain test set. + +For the scene reconstruction task, we evaluate on 16 scenes sampled from Heinly et al. [12], Wilson et al. [31], MegaScenes [26] and 5 VisymScenes test scenes. These scenes are challenging for conventional SfM pipelines as well as prior disambiguation methods due to subtle differences between distinct surfaces and repetitive patterns. + +Metrics. For pairwise visual disambiguation evaluation, we report Average Precision (AP) and ROC AUC scores, following the protocol in [3]. Additionally, we report precision at a fixed recall and recall at a fixed precision. These statistics help characterize a model's trade-off between achieving high precision and potentially missing true positives. + +To comprehensively evaluate SfM reconstructions, we report the number of images in the SfM results, and the geo-alignment inlier ratio described in Sec. 3.3. Respectively, more images in the SfM model imply less false pair pruning, i.e., better class separation; and a higher geo-alignment inlier ratio suggests that the reconstructed model is more likely to be accurate and complete. Note that the inlier ratio is also influenced by the accuracy of the geo-tagged images collected from Mapillary. Thus, if one test scene has more accurate geo-tagged images than another, it is likely to yield a higher inlier ratio. Therefore, rather than comparing inlier ratios across different scenes, this metric is more useful for evaluating different reconstruction methods on the same scene, as the same set of geo-tagged images is used. + +Baselines. Our primary baseline is [3], as their learning-based approach consistently outperforms heuristic-based methods like thresholding on the number of post-RANSAC inliers (e.g. SIFT [17] + RANSAC [9], D2-Net [8] + RANSAC, SuperPoint [6] + SuperGlue [22]) for pairwise visual disambiguation; and is also superior than previous approaches [5, 12, 31, 32] for SfM disambiguation. + +# 4.2. Implementation + +Model details. Our doppelganger classifier head $\mathrm{Head}_{\mathrm{dopp}}$ is implemented with a Transformer encoder comprised of 3 layers, each with input and output dimension of 768. Each layer has 8 attention heads and a two-layer feed-forward network with a hidden dimension of 2048. Pre-layer normalization and residual connections are applied within each layer. The transformed tokens are aggregated via max pooling and linearly projected into $\mathrm{pred}^v$ . We freeze the weights of the public MAST3R [16] model and only train the two doppelganger classification heads, supervised by a cross-entropy loss. We train for 5 epochs with a batch size of 8 + +using Adam [15] with a learning rate of $1 \times 10^{-4}$ . More details can be found in the supplementary. + +Integration with SfM. Following [3], we integrate Doppelgangers++ into SfM to enhance its disambiguation ability. SfM takes a collection of images $\mathcal{I} = \{I_i\}_{i=0}^n$ and generates geometrically verified image pairs $\mathcal{P} = \{(I_p, I_q)\}$ , forming a scene graph $\mathcal{G} = (\mathcal{I}, \mathcal{P})$ with images as nodes and pairs as edges. Using $\mathcal{G}$ , SfM computes camera poses and reconstructs a 3D point cloud. Our doppelganger classifier acts as a filter on edges in $\mathcal{G}$ , removing pairs below a probability threshold $\tau$ to eliminate spurious matches due to repeated or symmetric structures. We integrate Doppelgangers++ into both COLMAP [23] and MASt3R-SfM [7] to showcase the effectiveness of the method. Notably, with COLMAP, SIFT features are computed for scene graph construction and MASt3R features for pruning; whereas with MASt3R-SfM, MASt3R features serve for both scene graph construction and pruning, forming a more efficient pipeline. + +# 4.3. Pairwise Visual Disambiguation + +We compare our approach with the original Doppelganger work [3] (dubbed $DG-OG$ ) under two experimental settings: 1) we use the same training data as in [3], i.e. the DG training set, to train our model. Both models are evaluated on three benchmark test sets: DG, VisymScenes, and Mapillary. Under this configuration, the DG test set is in-domain, while VisymScenes and Mapillary are out-of-domain scenarios; 2) we expand the training data by including the new VisymScenes training set and retrain both DG-OG and our model. Evaluation is conducted on the same three test sets, with only the Mapillary test set being out-of-domain this time. + +Quantitative results are in Tab. 1. Under the first setting, where both methods are trained on DG, our model shows clear improvements across all three test sets. Specifically, on the in-domain DG test set, our model achieves higher AP and ROC AUC, indicating that it maintains high precision across all recall levels, while also being less sensitive to the decision threshold $\tau$ with better class separation. On the out-of-domain test sets (VisymScenes and Mapillary), we observe $25\%$ to $65\%$ improvements in both AP and ROC AUC, highlighting the generalizability of our method. + +Although our model demonstrates improved performance even when trained solely on the DG dataset, its precision outside of the training domain is suboptimal for practical use in SfM, where high precision is essential. In the second setting, we expand the training set by including VisymScenes data. Both methods maintain similar performance on the DG test set, with our model experiencing a slight drop in recall when precision is set to 0.99. On the VisymScenes test set, both models show improvements across all metrics, with ours reaching $99\%$ in both AP and ROC AUC. Notably, adding VisymScenes to the training set does not enhance DG-OG's performance on the out-of-domain Mapillary test + +![](images/4f0612e82fbcaa2337ccce1cdf0c08c41df84037639aa9dfccb0179ee5051fb2.jpg) +Figure 5. SfM Disambiguation on MegaScenes [26]. (White background) SfM results from DG-OG [3] and ours. (Black background) Verification using geo-tagged images, red points represent registered cameras and green points represent geolocations, inlier ratio (IR) is labeled on the bottom right. DG-OG fails to disambiguate this scene, predicting incorrect scores for image pairs. Our method correctly splits the model into two clean components. + +![](images/5ef5e6ed4c8f5c689e62b63d73aa78bb99b183eaa71deaf1c989fc812d448771.jpg) + +![](images/56284431ec2eb48111d90ea0674c2dd93c3616c37e3d1cc1fb00f06f8571dc78.jpg) + +![](images/e25ab08f7ec9f681f89a95e8c932bc732295996420d7fbc44e3795fef35161d0.jpg) +Figure 6. SfM disambiguation on VisymScenes. We show that our classifier is more robust than DG-OG [3] on test scenes from new domains, like everyday street scenes. DG-OG has difficulty disambiguating such scenes, leading to incorrect geometry. + +
Test DataMethodMetrics (trained on DG / trained on DG + VisymScenes)
AP↑ROC AUC↑Prec@Recall=0.85↑Recall@Prec=0.99↑
DGDG-OG0.954 / 0.9560.944 / 0.9470.901 / 0.9100.611 / 0.614
Ours0.980 / 0.9810.981 / 0.9810.972 / 0.9820.690 / 0.642
VisymScenesDG-OG0.816 / 0.9380.726 / 0.9210.498 / 0.8310.340 / 0.623
Ours0.936 / 0.9910.909 / 0.9900.892 / 0.9990.542 / 0.901
MapillaryDG-OG0.566 / 0.6920.581 / 0.7010.523 / 0.5720.003 / 0.000
Ours0.950 / 0.9680.944 / 0.9580.927 / 0.9420.310 / 0.736
+ +Table 1. Evaluation of pairwise disambiguation. We evaluate DG-OG [3] and our method trained on DG [3] only and DG +VisymScenes (two numbers per cell) on three test sets. Both DG-OG and ours benefit from dataset expansion, whereas ours gained more generalizability on out-of-domain test set (Mapillary) after training on both. Our classifier constantly demonstrates better precision, recall across all settings. + +set, while ours continues to benefit from increased training diversity. These results indicate that the prior approach struggles to generalize effectively, whereas ours shows greater generalization with more varied training data. + +# 4.4. Structure from Motion disambiguation + +We integrate our classifier trained on DG and VisymScenes into COLMAP's SfM pipeline [23], and evaluate its performance on reconstructing scenes with duplicated and symmet + +ric structures. We compare the results with vanilla COLMAP and Cai et al. [3] (DG-OG). Quantitative results are presented in Tab. 2. Our approach registers more images across all scenes than DG-OG and operates without threshold tuning. In contrast, DG-OG relies on scene-specific thresholds, such as $\tau = 0.97$ for the Church on Spilled Blood to correctly separate both sides of the church, and $\tau = 0.6$ for Ponte di Rialto [26] to avoid over-segmentation and maintain completeness. Notably, DG-OG fails to disambiguate Belvedere (Vienna), while our method succeeds with a consistent threshold ( $\tau = 0.8$ across all scenes). Our approach also consistently achieves a higher inlier ratio than the baselines, indicating greater accuracy and completeness in the reconstructed models. As an example, we qualitatively show the results of reconstructing Belvedere (Vienna) in Fig. 5, along with verification results using geo-tagged images. + +We also evaluate on scenes from the VisymScenes test set. Since VisymScenes images come with geolocation metadata, we do not need additional Mapillary images; instead, we directly compute the inlier ratio between SfM camera positions and the geolocation metadata after RANSAC. Results show that our model effectively prunes spurious pairs and improves the reconstructed model quality, whereas DG-OG sometimes fails to distinguish doppelganger pairs. Fig. 6 shows examples where DG-OG encounters difficulties. + +In Fig. 7, we show that while MASt3R's features are powerful, MASt3R-SfM is not free from doppelganger issues. Although our classifier was trained on image pairs mined through COLMAP's feature matching module (i.e. w/ SIFT features), it effectively prunes incorrect matches generated by MASt3R, restoring accurate reconstruction results. + +# 4.5. Ablation + +In this section, we study the effectiveness of our designs from the following aspects: 1) fine-tuning the entire model (vs. only new heads), 2) one classification head (vs. two + +![](images/d651cbb028479382932dbdcc6c8108c6e8313fe0eb7776ee25109157f3140bd0.jpg) +Figure 7. MATLAB's Doppelgangers++. MATLAB also suffers from doppelganger issues. Our classifier effectively prunes false positive pairs and correctly reconstructs challenging scenes. + +
Test Scenes# SFM-registered ImagesInlier Ratio (Sec. 3.3)
COLMAPDG-OGOursCOLMAPDG-OGOurs
Alexander Nevsky Cathedral [12]4474444470.5651.01.0
Arc de Triomphé [12]4243844230.6970.9660.990
Berliner Dom [12]1603158816060.7090.9870.992
Big Ben [12]3983793940.5640.8270.831
Church on Spilled Blood [12]27384+94 (τ=0.97)157+1060.5420.8810.962
Radcliffe Camera [12]28191+8494+1860.4950.9550.970
Seville [12]1498585+272+515615+303+5520.4500.7720.854
Brandenburg Gate [26]21371361+5701398+6030.4400.9000.909
Palacio de Comunicaciones [26]727307+80 (τ=0.6)308+840.2290.8230.934
Ponte di Rialto [26]652538+101 (τ=0.6)540+1070.6270.8340.844
York Minster [26]636200+362206+2840.7270.8580.901
Town Hall Tower, Kraków [26]2982552800.6090.7310.838
Belvedere (Vienna) [26]1038851 (fail)457+5000.5210.4510.874
Reichstag (building) [26]1504997+3101024+3560.4690.8040.862
St. Vitus Cathedral [26]7526736920.8530.9090.933
Royal Liver Building [26]2121711800.71.01.0
VisymSite00101544520+29014460.7700.8200.913
VisymSite0023849471+81566+820.8670.8480.942
VisymSite0028450238+179267+1200.8180.7110.909
VisymSite0042540206+2074670.8630.9240.963
VisymSite01091245237+458+78239+612+1270.8570.8620.927
+ +Table 2. Evaluation of SfM results. We compare reconstructions from COLMAP [23], DG-OG [3], and our method. $\tau = 0.8$ is used unless otherwise stated. The $^+$ indicates split reconstruction components; e.g., DG-OG and our method split the Radcliffe Camera reconstruction into two components. Because VisymScenes scenes are large, we report statistics on the largest reconstruction component produced by COLMAP, and identify the corresponding (split) components in the results of DG-OG and ours. + +separate heads), 3) the architecture of our classification head, and 4) final-layer decoder features (vs. multi-layer decoder features). We evaluate models trained on DG and Visym-Scenes datasets and show PR curves in Fig. 8 on the three test sets with respect to the pairwise classification task. A full ablation table can be found in the supplementary. + +Fine-tuning MASt3R. As discussed in Sec.3.2, we choose to train a lighter output head on top of MASt3R features rather than fine-tuning the entire model weights. Curves in Fig. 8 show that training only the head achieves comparable performance to full model fine-tuning. The drop in recall with full model fine-tuning suggests potential overfitting to training data, making the model more conservative and less generalizable to unseen inputs. This tendency also holds on models trained solely on DG data: on the VisymScenes test set, our approach achieved 0.03 higher AP than fine-tuning. On Mapillary test set, ours achieved comparable AP but with 0.2 higher recall at precision 0.99. + +Single classification head. Our design uses two classification heads to process features from the two branches separately. An alternative approach would be to combine the + +![](images/07045ff87fba6e6f207b98a39d7be224e82c47778c57e18f5ba83a9633309a9d.jpg) +Figure 8. Precision-Recall curves of ablation studies. Metrics are evaluated on models trained with DG and VisymScenes. Full tables can be found in the supplementary. + +features from both decoder branches and use a single head to predict the doppelganger score. As shown in Fig. 8, this alternative degrades performance consistently. This may be due to the asymmetric nature of the MAST3R design, where each branch captures different kinds of information that are better to be processed separately. On the other hand, our voting mechanism between the two heads helps enhance class distinction, outputting lower scores for negative samples and higher scores for positive samples. + +Head architecture. While the decoder features of MASt3R retain rich geometric 3D information, designing a classifier with sufficient capacity to effectively leverage this information for the doppelganger prediction task is non-trivial. We compare the performance of our Transformer-based classifier with that of an MLP. Fig. 8 shows that our Transformer model outperforms the MLP across all metrics and test sets. + +Single-layer decoder feature. Our classifier head takes multi-layer decoder features as input. Here we substitute with only the final-layer decoder feature. Results in Fig. 8 show that the multi-layer decoder features are consistently superior to the single-layer ones, because the classifier is able to analyze doppelganger factors from different aspects and levels. See supplementary for more detailed analysis. + +# 5. Conclusion + +We propose Doppelgangers++ as an effective approach to handling visual aliasing in 3D reconstruction. We introduce a new VisymScenes dataset, featuring images from diverse daily scenes, and develop rules to mine doppelganger data from the dataset to enrich our training data. We train a Transformer-based classifier that leverages geometric 3D features to classify image pairs with high precision and recall across various scenes. Doppelgangers++ integrates seamlessly into existing SfM pipelines, enhancing reconstruction accuracy without extensive parameter tuning. Further, we propose a validation method using geo-tagged map images, offering a more comprehensive and automatic way to assess SfM accuracy and model completeness. Extensive experiments show that Doppelgangers++ significantly improves visual disambiguation and 3D reconstruction quality in complex scenes. + +# 6. Acknowledgment + +We thank Brandon Li and Joseph Tung for their valuable discussion and help with the webviewer. This work was supported by the Intelligence Advanced Research Projects Activity (IARPA) via Department of Interior/Interior Business Center (DOI/IBC) contract number 140D0423C0035. The U.S. Government is authorized to reproduce and distribute reprints for governmental purposes notwithstanding any copyright annotation thereon. Disclaimer: The views and conclusions contained herein are those of the authors and should not be interpreted as necessarily representing the official policies or endorsements, either expressed or implied, of IARPA, DOI/IBC, or the U.S. Government. This research was also supported in part by the National Science Foundation under award IIS-2212084. + +# References + +[1] Eric Brachmann, Jamie Wynn, Shuai Chen, Tommaso Cavallari, Aron Monszpart, Daniyar Turmukhambetov, and Victor Adrian Prisacariu. Scene coordinate reconstruction: posing of image collections via incremental learning of a relocalizer. In Eur. Conf. Comput. Vis., pages 421-440. Springer, 2025. 3 +[2] J. Byrne, G. Castanon, Z. Li, and G. Ettinger. Fine-grained activities of people worldwide. In IEEE/CVF Winter Conference on Applications of Computer Vision, 2023. 2, 3 +[3] Ruojin Cai, Joseph Tung, Qianqian Wang, Hadar Averbuch-Elor, Bharath Hariharan, and Noah Snavely. Doppelgangers: Learning to disambiguate images of similar structures. In Int. Conf. Comput. Vis., pages 34-44, 2023. 1, 2, 3, 5, 6, 7, 8 +[4] Duygu Ceylan, Niloy Jyoti Mitra, Youyi Zheng, and Mark Pauly. Coupled structure-from-motion and 3d symmetry detection for urban facades. ACM Transactions on Graphics (TOG), 33:1-15, 2014. 2 +[5] Zhaopeng Cui and Ping Tan. Global structure-from-motion by similarity averaging. 2015 IEEE International Conference on Computer Vision (ICCV), pages 864-872, 2015. 6 +[6] Daniel DeTone, Tomasz Malisiewicz, and Andrew Rabinovich. Superpoint: Self-supervised interest point detection and description. In Proc. CVPR Workshops, pages 224-236, 2018. 2, 6 +[7] Bardienus Pieter Duisterhof, Lojze Žust, Philippe Weinzaepfel, Vincent Leroy, Yohann Cabon, and Jérôme Revaud. Mast3r-sfm: a fully-integrated solution for unconstrained structure-from-motion. ArXiv, abs/2409.19152, 2024. 2, 3, 6 +[8] Mihai Dusmanu, Ignacio Rocco, Tomás Pajdla, Marc Pollefeys, Josef Sivic, Akihiko Torii, and Torsten Sattler. D2-net: A trainable cnn for joint description and detection of local features. 2019 IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 8084-8093, 2019. 6 +[9] Martin A. Fischler and Robert C. Bolles. Random sample consensus: a paradigm for model fitting with applications to image analysis and automated cartography. Commun. ACM, 24:381-395, 1981. 6 + +[10] Ye Gong, Pengwei Zhou, Changfeng Liu, Yan Yu, Jian Yao, Wei Yuan, and Li Li. A cluster-based disambiguation method using pose consistency verification for structure from motion. ISPRS Journal of Photogrammetry and Remote Sensing, 2024. 2 +[11] Jared Heinly, Enrique Dunn, and Jan-Michael Frahm. Recovering correct reconstructions from indistinguishable geometry. 2014 2nd International Conference on 3D Vision, 1:377-384, 2014. 2 +[12] Jared Heinly, Enrique Dunn, and Jan-Michael Frahm. Correcting for duplicate scene structure in sparse 3d reconstruction. In Eur. Conf. Comput. Vis., pages 780-795. Springer, 2014. 6, 8 +[13] Nianjuan Jiang, Ping Tan, and Loong-Fah Cheong. Seeing double without confusion: Structure-from-motion in highly ambiguous scenes. In Conf. Comput. Vis. Pattern Recog., pages 1458-1465, 2012. 2 +[14] Rajbir Kataria, Joseph DeGol, and Derek Hoiem. Improving structure from motion with reliable resectioning. 2020 International Conference on 3D Vision (3DV), pages 41-50, 2020. 2 +[15] Diederik P. Kingma and Jimmy Ba. Adam: A method for stochastic optimization. CoRR, abs/1412.6980, 2014. 6 +[16] Vincent Leroy, Yohann Cabon, and Jérôme Revaud. Grounding image matching in 3d with mast3r. ArXiv, abs/2406.09756, 2024. 2, 3, 6 +[17] David G Lowe. Distinctive image features from scale-invariant keypoints. Int. J. Comput. Vis., 60:91-110, 2004. 2, 6 +[18] Lalit Manam and Venu Madhav Govindu. Leveraging camera triplets for efficient and accurate structure-from-motion. 2024 IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 4959-4968, 2024. 2 +[19] Yang Peng, Shen Yan, Yuxiang Liu, Y. Liu, and Maojun Zhang. View graph construction for scenes with duplicate structures via graph convolutional network. IET Comput. Vis., 16:389-402, 2022. 2 +[20] Richard Roberts, Sudipta N. Sinha, Richard Szeliski, and Drew Steedly. Structure from motion for scenes with large duplicate structures. CVPR 2011, pages 3137-3144, 2011. 2 +[21] Ignacio Rocco, Mircea Cimpoi, Relja Arandjelović, Akihiko Torii, Tomas Pajdla, and Josef Sivic. Neighbourhood consensus networks. In Adv. Neural Inform. Process. Syst., 2018. 2 +[22] Paul-Edouard Sarlin, Daniel DeTone, Tomasz Malisiewicz, and Andrew Rabinovich. Superglue: Learning feature matching with graph neural networks. 2020 IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 4937–4946, 2019. 6 +[23] Johannes L Schonberger and Jan-Michael Frahm. Structure-from-motion revisited. In Conf. Comput. Vis. Pattern Recog., pages 4104-4113, 2016. 2, 3, 5, 6, 7, 8 +[24] Tianwei Shen, Siyu Zhu, Tian Fang, Runze Zhang, and Long Quan. Graph-based consistent matching for structure-from-motion. In European Conference on Computer Vision, 2016. 2 + +[25] Jiaming Sun, Zehong Shen, Yuang Wang, Hujun Bao, and Xiaowei Zhou. Loftr: Detector-free local feature matching with transformers. In Conf. Comput. Vis. Pattern Recog., pages 8922-8931, 2021. 2 +[26] Joseph Tung, Gene Chou, Ruojin Cai, Guandao Yang, Kai Zhang, Gordon Wetzstein, Bharath Hariharan, and Noah Snavely. Megascenes: Scene-level view synthesis at scale. ArXiv, abs/2406.11819, 2024. 5, 6, 7, 8 +[27] Fan Wang, Aditi Nayak, Yogesh Agrawal, and Roy Shilkrot. Hierarchical image link selection scheme for duplicate structure disambiguation. In British Machine Vision Conference, 2018. 2 +[28] Jianyuan Wang, Nikita Karaev, Christian Rupprecht, and David Novotny. Vggsfm: Visual geometry grounded deep structure from motion. In Conf. Comput. Vis. Pattern Recog., pages 21686-21697, 2024. 3 +[29] Shuzhe Wang, Vincent Leroy, Yohann Cabon, Boris Chidlovskii, and Jérôme Revaud. DUS3R: Geometric 3d vision made easy. In Conf. Comput. Vis. Pattern Recog., pages 20697-20709, 2023. 2, 3 +[30] Frederik Warburg, Søren Hauberg, Manuel López-Antequera, Pau Gargallo, Yubin Kuang, and Javier Civera. Mapillary street-level sequences: A dataset for lifelong place recognition. In Conf. Comput. Vis. Pattern Recog., pages 2623-2632, 2020. 6 +[31] Kyle Wilson and Noah Snavely. Network principles for sfm: Disambiguating repeated structures with local context. In Int. Conf. Comput. Vis., pages 513-520, 2013. 2, 6, 8 +[32] Qingan Yan, Long Yang, Ling Zhang, and Chunxia Xiao. Distinguishing the indistinguishable: Exploring structural ambiguities via geodesic context. In Conf. Comput. Vis. Pattern Recog., pages 3836-3844, 2017. 2, 6 +[33] Kwang Moo Yi, Eduard Trulls, Vincent Lepetit, and Pascal Fua. Lift: Learned invariant feature transform. In Eur. Conf. Comput. Vis., pages 467-483, 2016. 2 +[34] Christopher Zach, Arnold Irschara, and Horst Bischof. What can missing correspondences tell us about 3d structure and motion? In Conf. Comput. Vis. Pattern Recog., pages 1-8, 2008. 2 +[35] Christopher Zach, Manfred Klopschitz, and Marc Pollefeys. Disambiguating visual relations using loop constraints. In Conf. Comput. Vis. Pattern Recog., pages 1426-1433, 2010. \ No newline at end of file diff --git a/2025/Doppelgangers++_ Improved Visual Disambiguation with Geometric 3D Features/images.zip b/2025/Doppelgangers++_ Improved Visual Disambiguation with Geometric 3D Features/images.zip new file mode 100644 index 0000000000000000000000000000000000000000..8597da872aa55f351a5e531cd157fbae41ef243f --- /dev/null +++ b/2025/Doppelgangers++_ Improved Visual Disambiguation with Geometric 3D Features/images.zip @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:d857339c148c6467f2e032aec5f3bd4eb2e3f1193edd447f663483c45b4e5d5a +size 677053 diff --git a/2025/Doppelgangers++_ Improved Visual Disambiguation with Geometric 3D Features/layout.json b/2025/Doppelgangers++_ Improved Visual Disambiguation with Geometric 3D Features/layout.json new file mode 100644 index 0000000000000000000000000000000000000000..449279218bae341aa6fb8c065fabf8693c513401 --- /dev/null +++ b/2025/Doppelgangers++_ Improved Visual Disambiguation with Geometric 3D Features/layout.json @@ -0,0 +1,7743 @@ +{ + "pdf_info": [ + { + "para_blocks": [ + { + "bbox": [ + 56, + 103, + 555, + 122 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 103, + 555, + 122 + ], + "spans": [ + { + "bbox": [ + 56, + 103, + 555, + 122 + ], + "type": "text", + "content": "Doppelgangers++: Improved Visual Disambiguation with Geometric 3D Features" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 102, + 142, + 508, + 172 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 102, + 142, + 508, + 172 + ], + "spans": [ + { + "bbox": [ + 102, + 142, + 508, + 172 + ], + "type": "text", + "content": "Yuanbo Xiangli" + }, + { + "bbox": [ + 102, + 142, + 508, + 172 + ], + "type": "inline_equation", + "content": "^{1}" + }, + { + "bbox": [ + 102, + 142, + 508, + 172 + ], + "type": "text", + "content": " Ruojin Cai" + }, + { + "bbox": [ + 102, + 142, + 508, + 172 + ], + "type": "inline_equation", + "content": "^{1}" + }, + { + "bbox": [ + 102, + 142, + 508, + 172 + ], + "type": "text", + "content": " Hanyu Chen" + }, + { + "bbox": [ + 102, + 142, + 508, + 172 + ], + "type": "inline_equation", + "content": "^{1}" + }, + { + "bbox": [ + 102, + 142, + 508, + 172 + ], + "type": "text", + "content": " Jeffrey Byrne" + }, + { + "bbox": [ + 102, + 142, + 508, + 172 + ], + "type": "inline_equation", + "content": "^{2}" + }, + { + "bbox": [ + 102, + 142, + 508, + 172 + ], + "type": "text", + "content": " Noah Snavely" + }, + { + "bbox": [ + 102, + 142, + 508, + 172 + ], + "type": "inline_equation", + "content": "^{1}" + }, + { + "bbox": [ + 102, + 142, + 508, + 172 + ], + "type": "inline_equation", + "content": "^{1}" + }, + { + "bbox": [ + 102, + 142, + 508, + 172 + ], + "type": "text", + "content": "Cornell University, " + }, + { + "bbox": [ + 102, + 142, + 508, + 172 + ], + "type": "inline_equation", + "content": "^{2}" + }, + { + "bbox": [ + 102, + 142, + 508, + 172 + ], + "type": "text", + "content": "Visym Labs" + } + ] + } + ], + "index": 3 + }, + { + "type": "image", + "bbox": [ + 86, + 184, + 175, + 235 + ], + "blocks": [ + { + "bbox": [ + 86, + 184, + 175, + 235 + ], + "lines": [ + { + "bbox": [ + 86, + 184, + 175, + 235 + ], + "spans": [ + { + "bbox": [ + 86, + 184, + 175, + 235 + ], + "type": "image", + "image_path": "e85edf79ca55a027c11e983dc0b5bda38223da24909f140773a47b053f71298d.jpg" + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 88, + 236, + 131, + 242 + ], + "lines": [ + { + "bbox": [ + 88, + 236, + 131, + 242 + ], + "spans": [ + { + "bbox": [ + 88, + 236, + 131, + 242 + ], + "type": "text", + "content": "Ours +MASt3R-SfM" + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "image_caption" + } + ], + "index": 5 + }, + { + "type": "image", + "bbox": [ + 88, + 243, + 175, + 285 + ], + "blocks": [ + { + "bbox": [ + 88, + 243, + 175, + 285 + ], + "lines": [ + { + "bbox": [ + 88, + 243, + 175, + 285 + ], + "spans": [ + { + "bbox": [ + 88, + 243, + 175, + 285 + ], + "type": "image", + "image_path": "4e5fb53f3d1dcbfeba8393412990b6ff3f14326b3a0753ea848f21bfe1df9771.jpg" + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "image_body" + } + ], + "index": 7 + }, + { + "type": "image", + "bbox": [ + 176, + 183, + 266, + 236 + ], + "blocks": [ + { + "bbox": [ + 176, + 183, + 266, + 236 + ], + "lines": [ + { + "bbox": [ + 176, + 183, + 266, + 236 + ], + "spans": [ + { + "bbox": [ + 176, + 183, + 266, + 236 + ], + "type": "image", + "image_path": "e4e598dc67129d2fb7db66cfac2036a5526427222bd768459652bcf03d2848b2.jpg" + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "image_body" + } + ], + "index": 8 + }, + { + "type": "image", + "bbox": [ + 177, + 236, + 266, + 286 + ], + "blocks": [ + { + "bbox": [ + 177, + 236, + 266, + 286 + ], + "lines": [ + { + "bbox": [ + 177, + 236, + 266, + 286 + ], + "spans": [ + { + "bbox": [ + 177, + 236, + 266, + 286 + ], + "type": "image", + "image_path": "a775f643c3ca9e057108e5669dccc344dd60c4ae376da5f03ff7ea558e0df5b8.jpg" + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "image_body" + } + ], + "index": 9 + }, + { + "type": "image", + "bbox": [ + 268, + 183, + 358, + 236 + ], + "blocks": [ + { + "bbox": [ + 268, + 183, + 358, + 236 + ], + "lines": [ + { + "bbox": [ + 268, + 183, + 358, + 236 + ], + "spans": [ + { + "bbox": [ + 268, + 183, + 358, + 236 + ], + "type": "image", + "image_path": "f27f03a87bdd3540e70c2ef6fd0d9d4cb1596908e4bf0307f96ddbc5067d71ba.jpg" + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "image_body" + } + ], + "index": 10 + }, + { + "type": "image", + "bbox": [ + 270, + 236, + 357, + 286 + ], + "blocks": [ + { + "bbox": [ + 270, + 236, + 357, + 286 + ], + "lines": [ + { + "bbox": [ + 270, + 236, + 357, + 286 + ], + "spans": [ + { + "bbox": [ + 270, + 236, + 357, + 286 + ], + "type": "image", + "image_path": "d8d43376eb2bf896d3073d92be4a6b4bb91c3f7465b7bc0a37d6dabb87dfc630.jpg" + } + ] + } + ], + "index": 11, + "angle": 0, + "type": "image_body" + } + ], + "index": 11 + }, + { + "type": "image", + "bbox": [ + 359, + 183, + 449, + 236 + ], + "blocks": [ + { + "bbox": [ + 359, + 183, + 449, + 236 + ], + "lines": [ + { + "bbox": [ + 359, + 183, + 449, + 236 + ], + "spans": [ + { + "bbox": [ + 359, + 183, + 449, + 236 + ], + "type": "image", + "image_path": "9bbd5704ad5623e79dcabb1dc7a6dd141ee34783f30333be0fb46dc74aaca4e1.jpg" + } + ] + } + ], + "index": 12, + "angle": 0, + "type": "image_body" + } + ], + "index": 12 + }, + { + "type": "image", + "bbox": [ + 360, + 236, + 449, + 286 + ], + "blocks": [ + { + "bbox": [ + 360, + 236, + 449, + 286 + ], + "lines": [ + { + "bbox": [ + 360, + 236, + 449, + 286 + ], + "spans": [ + { + "bbox": [ + 360, + 236, + 449, + 286 + ], + "type": "image", + "image_path": "b4cefdff827cef6d22b8bb2825c6a9d1851e8173c0ff1d20bad5d04f2f39a698.jpg" + } + ] + } + ], + "index": 13, + "angle": 0, + "type": "image_body" + } + ], + "index": 13 + }, + { + "type": "image", + "bbox": [ + 451, + 183, + 542, + 236 + ], + "blocks": [ + { + "bbox": [ + 451, + 183, + 542, + 236 + ], + "lines": [ + { + "bbox": [ + 451, + 183, + 542, + 236 + ], + "spans": [ + { + "bbox": [ + 451, + 183, + 542, + 236 + ], + "type": "image", + "image_path": "2a29b123581a9ed8c5064120e1e7b2f70829a9df2b3f68657a422658cbb3fa4e.jpg" + } + ] + } + ], + "index": 14, + "angle": 0, + "type": "image_body" + } + ], + "index": 14 + }, + { + "type": "image", + "bbox": [ + 452, + 236, + 542, + 286 + ], + "blocks": [ + { + "bbox": [ + 452, + 236, + 542, + 286 + ], + "lines": [ + { + "bbox": [ + 452, + 236, + 542, + 286 + ], + "spans": [ + { + "bbox": [ + 452, + 236, + 542, + 286 + ], + "type": "image", + "image_path": "82245e7a93305f40b2317a4271b3bf36d115e47e10a2060cfd43ba16f2511a63.jpg" + } + ] + } + ], + "index": 15, + "angle": 0, + "type": "image_body" + } + ], + "index": 15 + }, + { + "type": "image", + "bbox": [ + 118, + 294, + 178, + 373 + ], + "blocks": [ + { + "bbox": [ + 118, + 294, + 178, + 373 + ], + "lines": [ + { + "bbox": [ + 118, + 294, + 178, + 373 + ], + "spans": [ + { + "bbox": [ + 118, + 294, + 178, + 373 + ], + "type": "image", + "image_path": "be983518280b036162859f3bf77d472732559cbe51f366de7f5f07024d35107a.jpg" + } + ] + } + ], + "index": 16, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 119, + 375, + 174, + 384 + ], + "lines": [ + { + "bbox": [ + 119, + 375, + 174, + 384 + ], + "spans": [ + { + "bbox": [ + 119, + 375, + 174, + 384 + ], + "type": "text", + "content": ":0.235/Ours:0.982" + } + ] + } + ], + "index": 17, + "angle": 0, + "type": "image_caption" + } + ], + "index": 16 + }, + { + "type": "image", + "bbox": [ + 179, + 295, + 268, + 373 + ], + "blocks": [ + { + "bbox": [ + 179, + 295, + 268, + 373 + ], + "lines": [ + { + "bbox": [ + 179, + 295, + 268, + 373 + ], + "spans": [ + { + "bbox": [ + 179, + 295, + 268, + 373 + ], + "type": "image", + "image_path": "c965f37db1a78807f9458f024cb65e0de1cc2d90526d02366bd5366958e70bc7.jpg" + } + ] + } + ], + "index": 18, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 179, + 375, + 266, + 384 + ], + "lines": [ + { + "bbox": [ + 179, + 375, + 266, + 384 + ], + "spans": [ + { + "bbox": [ + 179, + 375, + 266, + 384 + ], + "type": "text", + "content": "GT:1/DG-OG:0.190/Ours:0.971" + } + ] + } + ], + "index": 19, + "angle": 0, + "type": "image_caption" + } + ], + "index": 18 + }, + { + "type": "image", + "bbox": [ + 269, + 295, + 359, + 373 + ], + "blocks": [ + { + "bbox": [ + 269, + 295, + 359, + 373 + ], + "lines": [ + { + "bbox": [ + 269, + 295, + 359, + 373 + ], + "spans": [ + { + "bbox": [ + 269, + 295, + 359, + 373 + ], + "type": "image", + "image_path": "8e44a4681b9be46df7d208feb798186e47a1f2235dbe54a6ccb05945f1e966c9.jpg" + } + ] + } + ], + "index": 20, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 270, + 375, + 356, + 383 + ], + "lines": [ + { + "bbox": [ + 270, + 375, + 356, + 383 + ], + "spans": [ + { + "bbox": [ + 270, + 375, + 356, + 383 + ], + "type": "text", + "content": "GT:1 / DG-OG:0.378 / Ours:0.952" + } + ] + } + ], + "index": 21, + "angle": 0, + "type": "image_caption" + } + ], + "index": 20 + }, + { + "type": "image", + "bbox": [ + 361, + 295, + 449, + 373 + ], + "blocks": [ + { + "bbox": [ + 361, + 295, + 449, + 373 + ], + "lines": [ + { + "bbox": [ + 361, + 295, + 449, + 373 + ], + "spans": [ + { + "bbox": [ + 361, + 295, + 449, + 373 + ], + "type": "image", + "image_path": "e64a8c415f17ac4ac9028ffa78be166e7b522693c7a24c16cd7f1967857a1b92.jpg" + } + ] + } + ], + "index": 22, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 361, + 375, + 447, + 383 + ], + "lines": [ + { + "bbox": [ + 361, + 375, + 447, + 383 + ], + "spans": [ + { + "bbox": [ + 361, + 375, + 447, + 383 + ], + "type": "text", + "content": "GT:0/DG-OG:0.993/Ours:0.025" + } + ] + } + ], + "index": 23, + "angle": 0, + "type": "image_caption" + } + ], + "index": 22 + }, + { + "type": "image", + "bbox": [ + 451, + 295, + 542, + 373 + ], + "blocks": [ + { + "bbox": [ + 451, + 295, + 542, + 373 + ], + "lines": [ + { + "bbox": [ + 451, + 295, + 542, + 373 + ], + "spans": [ + { + "bbox": [ + 451, + 295, + 542, + 373 + ], + "type": "image", + "image_path": "5c5cbc9397bdc7b929332a81bc456c37df569d27cece026bb5740eb307fb0c37.jpg" + } + ] + } + ], + "index": 24, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 453, + 375, + 539, + 383 + ], + "lines": [ + { + "bbox": [ + 453, + 375, + 539, + 383 + ], + "spans": [ + { + "bbox": [ + 453, + 375, + 539, + 383 + ], + "type": "text", + "content": "GT:0/DG-OG:0.994/Ours:0.031" + } + ] + } + ], + "index": 25, + "angle": 0, + "type": "image_caption" + } + ], + "index": 24 + }, + { + "type": "image", + "bbox": [ + 86, + 389, + 251, + 486 + ], + "blocks": [ + { + "bbox": [ + 86, + 389, + 251, + 486 + ], + "lines": [ + { + "bbox": [ + 86, + 389, + 251, + 486 + ], + "spans": [ + { + "bbox": [ + 86, + 389, + 251, + 486 + ], + "type": "image", + "image_path": "924a8f102cd310d18cf5fcf83ce9b420a38a052ca7731dc5e3fdd37678d905a1.jpg" + } + ] + } + ], + "index": 26, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 55, + 495, + 555, + 544 + ], + "lines": [ + { + "bbox": [ + 55, + 495, + 555, + 544 + ], + "spans": [ + { + "bbox": [ + 55, + 495, + 555, + 544 + ], + "type": "text", + "content": "Figure 1. Visual aliasing, or doppelgangers, poses severe challenges to 3D reconstruction. We propose Doppelganger++, an enhanced pairwise image classifier that excels in visual disambiguation across diverse and challenging scenes. (Top) We seamlessly integrate Doppelganger++ into SfM, successfully disambiguating each scene. (Middle) Compared to prior work (which we refer to as DG-OG [3]), Doppelgangers++ is more robust for everyday scenes, showing improved accuracy and robustness. We show pairs that DG-OG classify incorrectly and ours gets correct. (Bottom) Our new VisymScenes dataset, featuring complex daily scenes, is particularly challenging for COLMAP and DG-OG, but our method can achieve correct and complete reconstructions." + } + ] + } + ], + "index": 30, + "angle": 0, + "type": "image_caption" + } + ], + "index": 26 + }, + { + "type": "image", + "bbox": [ + 252, + 389, + 349, + 436 + ], + "blocks": [ + { + "bbox": [ + 252, + 389, + 349, + 436 + ], + "lines": [ + { + "bbox": [ + 252, + 389, + 349, + 436 + ], + "spans": [ + { + "bbox": [ + 252, + 389, + 349, + 436 + ], + "type": "image", + "image_path": "b53bc3b3b17778f7d7362d72012a622a2214797fd2f2d21ad6eb272cfedace40.jpg" + } + ] + } + ], + "index": 27, + "angle": 0, + "type": "image_body" + } + ], + "index": 27 + }, + { + "type": "image", + "bbox": [ + 252, + 437, + 349, + 484 + ], + "blocks": [ + { + "bbox": [ + 252, + 437, + 349, + 484 + ], + "lines": [ + { + "bbox": [ + 252, + 437, + 349, + 484 + ], + "spans": [ + { + "bbox": [ + 252, + 437, + 349, + 484 + ], + "type": "image", + "image_path": "84d7daad6b17fa4886892beac441f3543e868663fcea211d1b2ecb1055f228e6.jpg" + } + ] + } + ], + "index": 28, + "angle": 0, + "type": "image_body" + } + ], + "index": 28 + }, + { + "type": "image", + "bbox": [ + 350, + 389, + 544, + 485 + ], + "blocks": [ + { + "bbox": [ + 350, + 389, + 544, + 485 + ], + "lines": [ + { + "bbox": [ + 350, + 389, + 544, + 485 + ], + "spans": [ + { + "bbox": [ + 350, + 389, + 544, + 485 + ], + "type": "image", + "image_path": "d2f9bc25d12efb4ec378d61525efbd7b1d03b725f414aee84c5ad4351ae296ab.jpg" + } + ] + } + ], + "index": 29, + "angle": 0, + "type": "image_body" + } + ], + "index": 29 + }, + { + "bbox": [ + 152, + 553, + 200, + 565 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 152, + 553, + 200, + 565 + ], + "spans": [ + { + "bbox": [ + 152, + 553, + 200, + 565 + ], + "type": "text", + "content": "Abstract" + } + ] + } + ], + "index": 31 + }, + { + "bbox": [ + 55, + 582, + 297, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 582, + 297, + 713 + ], + "spans": [ + { + "bbox": [ + 55, + 582, + 297, + 713 + ], + "type": "text", + "content": "Accurate 3D reconstruction is frequently hindered by visual aliasing, where visually similar but distinct surfaces (aka, doppelgangers), are incorrectly matched. These spurious matches distort the structure-from-motion (SfM) process, leading to misplaced model elements and reduced accuracy. Prior efforts addressed this with CNN classifiers trained on curated datasets, but these approaches struggle to generalize across diverse real-world scenes and can require extensive parameter tuning. In this work, we present Doppelgangers++, a method to enhance doppelganger detection and improve 3D reconstruction accuracy. Our contribu" + } + ] + } + ], + "index": 32 + }, + { + "bbox": [ + 313, + 554, + 556, + 710 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 554, + 556, + 710 + ], + "spans": [ + { + "bbox": [ + 313, + 554, + 556, + 710 + ], + "type": "text", + "content": "tions include a diversified training dataset that incorporates geo-tagged images from everyday scenes to expand robustness beyond landmark-based datasets. We further propose a Transformer-based classifier that leverages 3D-aware features from the MAST3R model, achieving superior precision and recall across both in-domain and out-of-domain tests. Doppelgangers++ integrates seamlessly into standard SfM and MAST3R-SfM pipelines, offering efficiency and adaptability across varied scenes. To evaluate SfM accuracy, we introduce an automated, geotag-based method for validating reconstructed models, eliminating the need for manual inspection. Through extensive experiments, we demonstrate that Doppelgangers++ significantly enhances pairwise vi" + } + ] + } + ], + "index": 33 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "spans": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "type": "text", + "content": "CVF" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 145, + 0, + 496, + 37 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 145, + 0, + 496, + 37 + ], + "spans": [ + { + "bbox": [ + 145, + 0, + 496, + 37 + ], + "type": "text", + "content": "This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 61, + 201, + 80, + 266 + ], + "type": "aside_text", + "angle": 90, + "lines": [ + { + "bbox": [ + 61, + 201, + 80, + 266 + ], + "spans": [ + { + "bbox": [ + 61, + 201, + 80, + 266 + ], + "type": "text", + "content": "(sneepnep)" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 293, + 749, + 318, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 749, + 318, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 749, + 318, + 757 + ], + "type": "text", + "content": "27166" + } + ] + } + ], + "index": 34 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 0 + }, + { + "para_blocks": [ + { + "bbox": [ + 55, + 72, + 297, + 97 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 72, + 297, + 97 + ], + "spans": [ + { + "bbox": [ + 55, + 72, + 297, + 97 + ], + "type": "text", + "content": "sual disambiguation and improves 3D reconstruction quality in complex and diverse scenarios." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 56, + 119, + 138, + 132 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 119, + 138, + 132 + ], + "spans": [ + { + "bbox": [ + 56, + 119, + 138, + 132 + ], + "type": "text", + "content": "1. Introduction" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 54, + 140, + 297, + 270 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 54, + 140, + 297, + 270 + ], + "spans": [ + { + "bbox": [ + 54, + 140, + 297, + 270 + ], + "type": "text", + "content": "Visual aliasing—confusing two surfaces that look the same but are nonetheless distinct—is an pernicious problem in 3D reconstruction and SLAM systems. Pairs of images that depict visually similar yet distinct 3D surfaces (called doppelgangers by Cai et al. [3]) can generate spurious correspondence at the feature matching stage of 3D reconstruction, leading to erroneous downstream reconstructions that feature distorted geometry or incorrectly fused elements. Therefore, to ensure the accuracy of 3D reconstruction, it is critical to distinguish truly matching images from illusory matches arising from doppelganger pairs." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 56, + 271, + 297, + 510 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 271, + 297, + 510 + ], + "spans": [ + { + "bbox": [ + 56, + 271, + 297, + 510 + ], + "type": "text", + "content": "In recent work, this visual disambiguation problem was formulated as a binary classification task on image pairs [3]. The authors collected a Internet dataset of visually similar doppelganger pairs (as well as truly matching images) from Wikipedia Commons, then trained a CNN to classify image pairs as correct or incorrect matches. This classifier can be used to take a feature match graph computed from a set of photos, remove incorrect edges between doppelganger image pairs, then reconstruct a correct model using a structure from motion (SfM) pipeline like COLMAP [23]. While that work shows promising improvements on reconstruction problems, we find that it can still be brittle: First, the 3D reconstruction task demands that the classifier have has precision—even a few bad edges remaining in the image match graph can lead to an incorrect 3D model. Second, the reconstruction task is quite sensitive to the threshold on the classifier score used to prune edges from the match graph. Finally, their method was trained solely on landmark Internet photos and does not reliably generalize to new scenarios, such as more structured captures of everyday scenes, like streets and office buildings." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 55, + 510, + 296, + 534 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 510, + 296, + 534 + ], + "spans": [ + { + "bbox": [ + 55, + 510, + 296, + 534 + ], + "type": "text", + "content": "In this paper, we aim to address these issues and improve doppelganger classification in several key ways:" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 55, + 534, + 296, + 677 + ], + "type": "list", + "angle": 0, + "index": 7, + "blocks": [ + { + "bbox": [ + 56, + 534, + 296, + 593 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 534, + 296, + 593 + ], + "spans": [ + { + "bbox": [ + 56, + 534, + 296, + 593 + ], + "type": "text", + "content": "1. We identify ways to expand and diversify doppelgangers training data. In particular, we leverage semi-structured image data with geographic annotations (rough GPS position and orientation), captured from everyday scenes with the Visym Collector platform [2]." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 55, + 594, + 296, + 677 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 594, + 296, + 677 + ], + "spans": [ + { + "bbox": [ + 55, + 594, + 296, + 677 + ], + "type": "text", + "content": "2. We switch from a CNN-based classifier to leveraging features from MASt3R [7], a recent transformer-based geometric model that computes point clouds from two input views. Specifically, we feed an image pair into a pre-trained MASt3R model, collect the intermediate features decoded by MASt3R, and train a classification head to map these features to a doppelganger score." + } + ] + } + ], + "index": 6 + } + ], + "sub_type": "text" + }, + { + "bbox": [ + 55, + 677, + 296, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 677, + 296, + 713 + ], + "spans": [ + { + "bbox": [ + 55, + 677, + 296, + 713 + ], + "type": "text", + "content": "We call our method Doppelgangers++. Our enhanced model achieves higher precision and recall in pairwise classification, and generalizes better across a broader range of scenes and" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 313, + 72, + 556, + 182 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 72, + 556, + 182 + ], + "spans": [ + { + "bbox": [ + 313, + 72, + 556, + 182 + ], + "type": "text", + "content": "capture scenarios. Doppelganger++ integrates seamlessly into SfM pipelines for 3D reconstruction disambiguation. We further propose to leverage geo-tagged map images to quantitatively evaluate the correctness and completeness of reconstructed models, in comparison to manual inspection as required in previous approaches. Through extensive experiments we show that our model leads to more accurate and complete reconstruction results, and is less sensitive to the threshold used for pruning doppelganger matches." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 313, + 193, + 403, + 205 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 193, + 403, + 205 + ], + "spans": [ + { + "bbox": [ + 313, + 193, + 403, + 205 + ], + "type": "text", + "content": "2. Related Work" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 313, + 214, + 556, + 478 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 214, + 556, + 478 + ], + "spans": [ + { + "bbox": [ + 313, + 214, + 556, + 478 + ], + "type": "text", + "content": "Local feature matching and 3D learning. Local feature matching methods have proven effective at establishing correspondences between pairs of images in SfM pipelines. Classic methods like COLMAP [23] rely on the tried-and-true SIFT features [17] to find correspondence. Modern learning-based, data-driven approaches have improved the quality of local feature matching [6, 21, 25, 33]. More recently, the DUSt3R [29] framework has proven to excel in a variety of 3D reconstruction tasks by estimating dense 3D point clouds from pairs of images. MASt3R [16] was proposed as an extension to DUSt3R that specifically targets the task of predicting dense feature matches between image pairs. All of these local feature matching methods are traditionally optimized to maximize the number of correspondences they find between similar image regions. While their ability to identify feature matches has greatly improved, they generally lack the ability to incorporate negative evidence into their predictions, and so they often find matches between regions that do not actually correspond to the same 3D surface, particularly within doppelganger image pairs. However, we show that features internal to MASt3R can be repurposed for doppelganger detection." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 313, + 486, + 557, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 486, + 557, + 714 + ], + "spans": [ + { + "bbox": [ + 313, + 486, + 557, + 714 + ], + "type": "text", + "content": "Disambiguation in SfM and image matching. Disambiguating similar structures and repeated patterns in SfM is a long-standing challenge. Prior work has mainly relied on heuristics-based analysis to detect conflicting relations and ambiguities in the structure of the underlying scene graph [10, 14, 18, 19, 24, 27, 31, 32, 35] or among image-level correspondence (or the lack thereof) [4, 11-13, 20, 34]. While such methods have shown some success in detecting and resolving ambiguities in SfM pipelines, one fundamental limitation that they do not consider the rich information contained in the images themselves. In contrast, Doppelgangers [3] avoids handcrafted heuristics and takes a data-driven approach to the basic problem of identifying doppelganger image pairs. They train a CNN that classifies a pair of images as either positive or negative, with LoFTR [25]-extracted keypoints and match masks passed as auxiliary input. During inference, the binary classification results are used to pre-process the scene graph obtained from COLMAP to filter out spurious matches prior to running SfM. How" + } + ] + } + ], + "index": 12 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "type": "text", + "content": "27167" + } + ] + } + ], + "index": 13 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 1 + }, + { + "para_blocks": [ + { + "bbox": [ + 55, + 72, + 294, + 167 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 72, + 294, + 167 + ], + "spans": [ + { + "bbox": [ + 55, + 72, + 294, + 167 + ], + "type": "text", + "content": "ever, the method's reliance on COLMAP feature extraction and matching and the need for auxiliary mask information from LoFTR introduces significant overhead and complexity to the overall pipeline, making it difficult to scale to larger scenes. In addition, we find that this prior method is brittle: it can fail to generalize to domains beyond landmarks (e.g., to office buildings), and it can require parameter tuning to get good results." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 56, + 175, + 298, + 390 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 175, + 298, + 390 + ], + "spans": [ + { + "bbox": [ + 56, + 175, + 298, + 390 + ], + "type": "text", + "content": "Differentiable SfM. To improve upon traditional SfM, several differentiable SfM pipelines have been proposed to optimize the entire 3D reconstruction process. These pipelines often include one or more components that are trained from large datasets. MAST3R-SfM [7] proposes an SfM pipeline that uses MAST3R features for scene graph construction and coarse-to-fine alignment. VGGSfM [28] decomposes the problem into four stages: point tracking, initial camera estimation, triangulation, and bundle adjustment, each of which is differentiable and can be optimized in an end-to-end manner. ACE0 [1] uses a scene coordinate regression network to progressively register new images to a global scene, while the network itself is iteratively trained and refined from the registered images. These methods have shown promising results in optimizing the SfM pipeline, but are not specifically designed to address the doppelganger problem, and are still prone to producing incorrect reconstructions when faced with scenes with repeated patterns or similar structures." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 55, + 403, + 111, + 415 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 403, + 111, + 415 + ], + "spans": [ + { + "bbox": [ + 55, + 403, + 111, + 415 + ], + "type": "text", + "content": "3. Method" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 55, + 423, + 295, + 533 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 423, + 295, + 533 + ], + "spans": [ + { + "bbox": [ + 55, + 423, + 295, + 533 + ], + "type": "text", + "content": "We refer to a \"doppelganger\" as a case where distinct objects or surfaces look almost identical, leading algorithms to confuse one for the other [3]. In this work, we aim to improve the doppelganger classifier via expanded and diversified training data (Sec. 3.1), and leveraging geometric 3D features learned from pairwise reconstruction [16] (Sec. 3.2). We further propose an approach to quantitatively evaluate the correctness of SfM results in terms of doppelgangers (Sec. 3.3), instead of manual inspection adopted in previous work." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 55, + 540, + 195, + 552 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 540, + 195, + 552 + ], + "spans": [ + { + "bbox": [ + 55, + 540, + 195, + 552 + ], + "type": "text", + "content": "3.1. The VisymScenes dataset" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 55, + 557, + 295, + 629 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 557, + 295, + 629 + ], + "spans": [ + { + "bbox": [ + 55, + 557, + 295, + 629 + ], + "type": "text", + "content": "Cai et al. [3] introduced the Doppelganger dataset, built on global landmarks photos sourced from Wikipedia Commons, with viewing direction (e.g., North, South) to identify doppelganger image pairs. While the classifier trained on this dataset proved effective, we find that it struggles to generalize well to scenes beyond the dataset's domain." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 55, + 630, + 295, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 630, + 295, + 713 + ], + "spans": [ + { + "bbox": [ + 55, + 630, + 295, + 713 + ], + "type": "text", + "content": "We enhance and expand their training set to improve the robustness of doppelganger classification, by incorporating casually captured images from a wider range of scenes. To ensure sufficient diversity in the dataset, we introduce the VisymScenes dataset. VisymScenes consists of 258K ground images with GPS/IMU metadata, recorded at 149 sites in 42 cities and 15 countries, collected with the Visym Collector" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 313, + 72, + 555, + 179 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 72, + 555, + 179 + ], + "spans": [ + { + "bbox": [ + 313, + 72, + 555, + 179 + ], + "type": "text", + "content": "Platform [2] (details in supplementary). Each image is accompanied by metadata, such as GPS coordinates, device compass direction, and intrinsic camera calibration. While this metadata can be noisy, it still provides valuable information for identifying potential doppelganger pairs. For example, if two images exhibit numerous geometrically consistent local feature matches yet were taken from distant locations, this serves as strong evidence that the pair is likely to be a doppelganger. Examples are shown in Fig. 2." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 313, + 181, + 556, + 529 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 181, + 556, + 529 + ], + "spans": [ + { + "bbox": [ + 313, + 181, + 556, + 529 + ], + "type": "text", + "content": "We develop a series of filtering rules applied to all matched image pairs identified by the COLMAP feature matching module [23]. Note that these rules are designed according to the capture style of the Visym Collector, where cameras consistently focus on the scene of interest and maintain a reasonable distance from it. For a given pair of images, we use the (metric) distance between the camera centers " + }, + { + "bbox": [ + 313, + 181, + 556, + 529 + ], + "type": "inline_equation", + "content": "r" + }, + { + "bbox": [ + 313, + 181, + 556, + 529 + ], + "type": "text", + "content": ", angle between their viewing directions " + }, + { + "bbox": [ + 313, + 181, + 556, + 529 + ], + "type": "inline_equation", + "content": "\\theta" + }, + { + "bbox": [ + 313, + 181, + 556, + 529 + ], + "type": "text", + "content": ", and camera intrinsics " + }, + { + "bbox": [ + 313, + 181, + 556, + 529 + ], + "type": "inline_equation", + "content": "\\mathcal{K}" + }, + { + "bbox": [ + 313, + 181, + 556, + 529 + ], + "type": "text", + "content": ", all derived from camera metadata, to identify confident negative (doppelganger) and positive (correct match) pairs. To identify confident negative pairs, we first identify spatially distant matching pairs, and classify them as doppelgangers. Then, we classify the remaining (nearby) pairs into three cases: the intersection point of the viewing directions is 1) in front of both cameras, 2) behind both cameras; or 3) in front of one camera and behind the other. For case 1), we label pairs with very large view angles (e.g., " + }, + { + "bbox": [ + 313, + 181, + 556, + 529 + ], + "type": "inline_equation", + "content": ">160^{\\circ}" + }, + { + "bbox": [ + 313, + 181, + 556, + 529 + ], + "type": "text", + "content": ") as negative, since they likely capture different 3D surfaces or mostly non-overlapping content. For case 2), if their view angle exceeds the diagonal field of view, then their view frustums are unlikely to overlap, so we label them as doppelgangers. For case 3), we check for frustum overlap using camera intrinsics; if no intersection is detected, the pair is considered a doppelganger. A similar series of rules are designed for mining positive pairs (i.e., true matches). Details of computing view intersection and filtering algorithms for both negative and positive pairs can be found in the supplementary. With these rules, we mined in total 53K positive and negative pairs across 33 sites for our doppelganger task." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 313, + 539, + 499, + 552 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 539, + 499, + 552 + ], + "spans": [ + { + "bbox": [ + 313, + 539, + 499, + 552 + ], + "type": "text", + "content": "3.2. Improved Doppelganger Classifier" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 313, + 557, + 555, + 616 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 557, + 555, + 616 + ], + "spans": [ + { + "bbox": [ + 313, + 557, + 555, + 616 + ], + "type": "text", + "content": "Recent advances in data-driven models [16, 29] have demonstrated impressive results in geometric vision tasks. In our work, we leverage the multi-level, 3D-aware features extracted from the MAST3R model [16] to train a doppelganger classifier on labeled image pairs, as demonstrated in Fig. 3." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 313, + 618, + 556, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 618, + 556, + 713 + ], + "spans": [ + { + "bbox": [ + 313, + 618, + 556, + 713 + ], + "type": "text", + "content": "Equipped with large-scale training data and a ViT backbone, the MASt3R model captures an image representation that encodes rich 3D geometric information between paired images. However, as MASt3R was originally trained to detect correspondences and similarities, its point maps often conflate doppelganger pairs, yielding incorrect point clouds and poses for these challenging pairs (examples in Fig. 7). Despite this limitation, we find that the internal" + } + ] + } + ], + "index": 11 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "text", + "content": "27168" + } + ] + } + ], + "index": 12 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 2 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 59, + 71, + 182, + 172 + ], + "blocks": [ + { + "bbox": [ + 59, + 71, + 182, + 172 + ], + "lines": [ + { + "bbox": [ + 59, + 71, + 182, + 172 + ], + "spans": [ + { + "bbox": [ + 59, + 71, + 182, + 172 + ], + "type": "image", + "image_path": "bb0850eb322419a465b9ad04e192ebca37cf0e16bf5b0bad9472f617b33cd9d7.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 55, + 179, + 555, + 225 + ], + "lines": [ + { + "bbox": [ + 55, + 179, + 555, + 225 + ], + "spans": [ + { + "bbox": [ + 55, + 179, + 555, + 225 + ], + "type": "text", + "content": "Figure 2. VisymScenes examples. This new dataset includes residential areas, landmarks, historical sites, business districts, and more. Here, we present four example sites. The top row shows subsets of images captured within each site. The bottom row displays pairs of visually similar but geographically distinct images from each site along with their recorded geolocations on a map. These examples demonstrate that doppelganger issues are prevalent in everyday scenes, presenting significant challenges for reliable 3D reconstruction and image matching." + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "type": "image", + "bbox": [ + 182, + 72, + 302, + 171 + ], + "blocks": [ + { + "bbox": [ + 182, + 72, + 302, + 171 + ], + "lines": [ + { + "bbox": [ + 182, + 72, + 302, + 171 + ], + "spans": [ + { + "bbox": [ + 182, + 72, + 302, + 171 + ], + "type": "image", + "image_path": "deed301b54334765ad2b9ba161e9c30a6f3c597194057601f07c0103a4e974ff.jpg" + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_body" + } + ], + "index": 1 + }, + { + "type": "image", + "bbox": [ + 306, + 72, + 426, + 171 + ], + "blocks": [ + { + "bbox": [ + 306, + 72, + 426, + 171 + ], + "lines": [ + { + "bbox": [ + 306, + 72, + 426, + 171 + ], + "spans": [ + { + "bbox": [ + 306, + 72, + 426, + 171 + ], + "type": "image", + "image_path": "58cd38fd1978cd5e8d91d604f59f38c8dddfbebcb144dbd2172707c65e7e708b.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + } + ], + "index": 2 + }, + { + "type": "image", + "bbox": [ + 428, + 72, + 550, + 171 + ], + "blocks": [ + { + "bbox": [ + 428, + 72, + 550, + 171 + ], + "lines": [ + { + "bbox": [ + 428, + 72, + 550, + 171 + ], + "spans": [ + { + "bbox": [ + 428, + 72, + 550, + 171 + ], + "type": "image", + "image_path": "a9ec786564a9c2398f43f04595dccc55d98bd1dffabd857bad716672c03cc9ee.jpg" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_body" + } + ], + "index": 3 + }, + { + "type": "image", + "bbox": [ + 58, + 232, + 553, + 316 + ], + "blocks": [ + { + "bbox": [ + 58, + 232, + 553, + 316 + ], + "lines": [ + { + "bbox": [ + 58, + 232, + 553, + 316 + ], + "spans": [ + { + "bbox": [ + 58, + 232, + 553, + 316 + ], + "type": "image", + "image_path": "2f09ee2e9191f482c02336660540159a793554b71811d74fd1d530dae681cd7c.jpg" + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 55, + 318, + 555, + 364 + ], + "lines": [ + { + "bbox": [ + 55, + 318, + 555, + 364 + ], + "spans": [ + { + "bbox": [ + 55, + 318, + 555, + 364 + ], + "type": "text", + "content": "Figure 3. Model design. (Left) Given an image pair, we first create a symmetrized version of the pair and feed it into the frozen MASt3R model. Multi-layer features are extracted from each decoder branch, concatenated, and fed into two learnable doppelganger classification heads. Each head generates predictions " + }, + { + "bbox": [ + 55, + 318, + 555, + 364 + ], + "type": "inline_equation", + "content": "(\\mathrm{pred}_{pq}^{v}, \\mathrm{pred}_{qp}^{v})" + }, + { + "bbox": [ + 55, + 318, + 555, + 364 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 55, + 318, + 555, + 364 + ], + "type": "inline_equation", + "content": "v \\in \\{1,2\\}" + }, + { + "bbox": [ + 55, + 318, + 555, + 364 + ], + "type": "text", + "content": " (where " + }, + { + "bbox": [ + 55, + 318, + 555, + 364 + ], + "type": "inline_equation", + "content": "pq" + }, + { + "bbox": [ + 55, + 318, + 555, + 364 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 55, + 318, + 555, + 364 + ], + "type": "inline_equation", + "content": "qp" + }, + { + "bbox": [ + 55, + 318, + 555, + 364 + ], + "type": "text", + "content": " denote the symmetrized image pair), supervised by cross-entropy loss. (Right) We use multi-layer decoder features and a Transformer-based classifier head for doppelganger prediction." + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "image_caption" + } + ], + "index": 5 + }, + { + "bbox": [ + 54, + 373, + 296, + 504 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 54, + 373, + 296, + 504 + ], + "spans": [ + { + "bbox": [ + 54, + 373, + 296, + 504 + ], + "type": "text", + "content": "features learned by MASt3R contain sufficient information for doppelganger classification without the need to fine-tune MASt3R's weights. Moreover, given that the amount of available training data for the doppelganger task is dwarfed by the amount of data used to train MASt3R, we opt not to fine-tune the entire model, but instead repurpose these internal features by training an additional output head to predict a doppelganger classification score. In Sec. 4.5, we empirically demonstrate that this choice achieves comparable or better visual disambiguation performance, especially on out-of-domain scenes." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 55, + 510, + 296, + 593 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 510, + 296, + 593 + ], + "spans": [ + { + "bbox": [ + 55, + 510, + 296, + 593 + ], + "type": "text", + "content": "Multi-level decoder features. For an image pair " + }, + { + "bbox": [ + 55, + 510, + 296, + 593 + ], + "type": "inline_equation", + "content": "(I_p, I_q)" + }, + { + "bbox": [ + 55, + 510, + 296, + 593 + ], + "type": "text", + "content": ", we extract multi-level decoder features from MASt3R's decoder branches. MASt3R takes two images as input, and uses two intertwined decoders " + }, + { + "bbox": [ + 55, + 510, + 296, + 593 + ], + "type": "inline_equation", + "content": "\\mathrm{DecBlk}^1" + }, + { + "bbox": [ + 55, + 510, + 296, + 593 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 55, + 510, + 296, + 593 + ], + "type": "inline_equation", + "content": "\\mathrm{DecBlk}^2" + }, + { + "bbox": [ + 55, + 510, + 296, + 593 + ], + "type": "text", + "content": " to decode encoded features " + }, + { + "bbox": [ + 55, + 510, + 296, + 593 + ], + "type": "inline_equation", + "content": "\\mathcal{H}^1 = \\mathrm{Enc}^1(I_p)" + }, + { + "bbox": [ + 55, + 510, + 296, + 593 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 55, + 510, + 296, + 593 + ], + "type": "inline_equation", + "content": "\\mathcal{H}^2 = \\mathrm{Enc}^2(I_q)" + }, + { + "bbox": [ + 55, + 510, + 296, + 593 + ], + "type": "text", + "content": ". Each decoder has " + }, + { + "bbox": [ + 55, + 510, + 296, + 593 + ], + "type": "inline_equation", + "content": "B" + }, + { + "bbox": [ + 55, + 510, + 296, + 593 + ], + "type": "text", + "content": " blocks, and attends to tokens from the other branch:" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 62, + 597, + 296, + 613 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 597, + 296, + 613 + ], + "spans": [ + { + "bbox": [ + 62, + 597, + 296, + 613 + ], + "type": "interline_equation", + "content": "f _ {i} ^ {1} = \\operatorname {D e c B l k} _ {i} ^ {1} \\left(f _ {i - 1} ^ {1}, f _ {i - 1} ^ {2}\\right), f _ {i} ^ {2} = \\operatorname {D e c B l k} _ {i} ^ {2} \\left(f _ {i - 1} ^ {2}, f _ {i - 1} ^ {1}\\right), \\tag {1}", + "image_path": "9d12cdd64fc5feb2c4e04ee2bdef42bfb183151c48f8ec0b408a1ceb3ca67fe1.jpg" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 55, + 617, + 296, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 617, + 296, + 714 + ], + "spans": [ + { + "bbox": [ + 55, + 617, + 296, + 714 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 55, + 617, + 296, + 714 + ], + "type": "inline_equation", + "content": "f_{i}^{v}" + }, + { + "bbox": [ + 55, + 617, + 296, + 714 + ], + "type": "text", + "content": " denotes the output tokens from the " + }, + { + "bbox": [ + 55, + 617, + 296, + 714 + ], + "type": "inline_equation", + "content": "i" + }, + { + "bbox": [ + 55, + 617, + 296, + 714 + ], + "type": "text", + "content": "-th block of the " + }, + { + "bbox": [ + 55, + 617, + 296, + 714 + ], + "type": "inline_equation", + "content": "v" + }, + { + "bbox": [ + 55, + 617, + 296, + 714 + ], + "type": "text", + "content": "-th branch. These two branches exchange information to capture the spatial relationships between viewpoints and the global 3D structure of the scene. For each branch " + }, + { + "bbox": [ + 55, + 617, + 296, + 714 + ], + "type": "inline_equation", + "content": "v" + }, + { + "bbox": [ + 55, + 617, + 296, + 714 + ], + "type": "text", + "content": ", we concatenate the encoder feature " + }, + { + "bbox": [ + 55, + 617, + 296, + 714 + ], + "type": "inline_equation", + "content": "\\mathcal{H}^v" + }, + { + "bbox": [ + 55, + 617, + 296, + 714 + ], + "type": "text", + "content": " and the tokens from the decoder blocks into " + }, + { + "bbox": [ + 55, + 617, + 296, + 714 + ], + "type": "inline_equation", + "content": "\\mathcal{F}^v = \\mathrm{concat}(\\mathcal{H}^v, \\{f_i^v\\}_{i=0}^{B-1})" + }, + { + "bbox": [ + 55, + 617, + 296, + 714 + ], + "type": "text", + "content": ", which captures rich, multi-level spatial correspondence information between image pairs." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 313, + 373, + 556, + 530 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 373, + 556, + 530 + ], + "spans": [ + { + "bbox": [ + 313, + 373, + 556, + 530 + ], + "type": "text", + "content": "Transformer-based classification heads. MAsst3R treats one image as the reference frame and projects the other image into that reference frame. A consequence of this design is that the reconstructed 3D scene for input pair " + }, + { + "bbox": [ + 313, + 373, + 556, + 530 + ], + "type": "inline_equation", + "content": "(I_p, I_q)" + }, + { + "bbox": [ + 313, + 373, + 556, + 530 + ], + "type": "text", + "content": " is distinct (or at least, must be in a different coordinate frame) from that of " + }, + { + "bbox": [ + 313, + 373, + 556, + 530 + ], + "type": "inline_equation", + "content": "(I_q, I_p)" + }, + { + "bbox": [ + 313, + 373, + 556, + 530 + ], + "type": "text", + "content": ". Inspired by the asymmetric design, we propose to 1) use separate Transformer heads: We introduce two independent, Transformer-based classification heads " + }, + { + "bbox": [ + 313, + 373, + 556, + 530 + ], + "type": "inline_equation", + "content": "\\mathrm{Head}_{dopp}^1" + }, + { + "bbox": [ + 313, + 373, + 556, + 530 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 313, + 373, + 556, + 530 + ], + "type": "inline_equation", + "content": "\\mathrm{Head}_{dopp}^2" + }, + { + "bbox": [ + 313, + 373, + 556, + 530 + ], + "type": "text", + "content": " to process " + }, + { + "bbox": [ + 313, + 373, + 556, + 530 + ], + "type": "inline_equation", + "content": "\\mathcal{F}^1" + }, + { + "bbox": [ + 313, + 373, + 556, + 530 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 313, + 373, + 556, + 530 + ], + "type": "inline_equation", + "content": "\\mathcal{F}^2" + }, + { + "bbox": [ + 313, + 373, + 556, + 530 + ], + "type": "text", + "content": " respectively; and 2) symmetrize image pairs such that the model evaluates both " + }, + { + "bbox": [ + 313, + 373, + 556, + 530 + ], + "type": "inline_equation", + "content": "(I_p, I_q)" + }, + { + "bbox": [ + 313, + 373, + 556, + 530 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 313, + 373, + 556, + 530 + ], + "type": "inline_equation", + "content": "(I_q, I_p)" + }, + { + "bbox": [ + 313, + 373, + 556, + 530 + ], + "type": "text", + "content": " to decide whether the given pair of images are true match or not. Thus, we end up with four scores for each image pair:" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 326, + 539, + 555, + 554 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 326, + 539, + 555, + 554 + ], + "spans": [ + { + "bbox": [ + 326, + 539, + 555, + 554 + ], + "type": "interline_equation", + "content": "\\operatorname {p r e d} _ {u} ^ {v} = \\operatorname {H e a d} _ {\\text {d o p p}} ^ {v} \\left(\\mathcal {F} _ {u} ^ {v}\\right), u \\in \\{p q, q p \\}, v \\in \\{1, 2 \\}. \\tag {2}", + "image_path": "827a0cb774edf6c9b854592016742815b6a63136650c60666851f3832b1c6ead.jpg" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 313, + 562, + 556, + 671 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 562, + 556, + 671 + ], + "spans": [ + { + "bbox": [ + 313, + 562, + 556, + 671 + ], + "type": "text", + "content": "Essentially, the two heads serve as expert evaluators, each examining how likely the image pair is to be a true match (or doppelganger) from different aspects, and switching the input order allows the model to analyze the spatial relationships from both directions. We empirically show the effectiveness of this design in Sec. 4.5. Both " + }, + { + "bbox": [ + 313, + 562, + 556, + 671 + ], + "type": "inline_equation", + "content": "\\mathrm{Head}_{dopp}^{1}" + }, + { + "bbox": [ + 313, + 562, + 556, + 671 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 313, + 562, + 556, + 671 + ], + "type": "inline_equation", + "content": "\\mathrm{Head}_{dopp}^{2}" + }, + { + "bbox": [ + 313, + 562, + 556, + 671 + ], + "type": "text", + "content": " are supervised by the cross-entropy loss, encouraging " + }, + { + "bbox": [ + 313, + 562, + 556, + 671 + ], + "type": "inline_equation", + "content": "S = \\{\\mathrm{pred}_u^v\\}" + }, + { + "bbox": [ + 313, + 562, + 556, + 671 + ], + "type": "text", + "content": " to all give high probabilities for positive matches and low probabilities for negative ones." + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 313, + 677, + 556, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 677, + 556, + 714 + ], + "spans": [ + { + "bbox": [ + 313, + 677, + 556, + 714 + ], + "type": "text", + "content": "Test-time voting. At test time, we combine the four scores " + }, + { + "bbox": [ + 313, + 677, + 556, + 714 + ], + "type": "inline_equation", + "content": "S" + }, + { + "bbox": [ + 313, + 677, + 556, + 714 + ], + "type": "text", + "content": " via a voting mechanism to make a final decision. Specifically, if the majority of the heads predict that the pair is a positive" + } + ] + } + ], + "index": 14 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "type": "text", + "content": "27169" + } + ] + } + ], + "index": 15 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 3 + }, + { + "para_blocks": [ + { + "bbox": [ + 55, + 72, + 296, + 167 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 72, + 296, + 167 + ], + "spans": [ + { + "bbox": [ + 55, + 72, + 296, + 167 + ], + "type": "text", + "content": "match, we take " + }, + { + "bbox": [ + 55, + 72, + 296, + 167 + ], + "type": "inline_equation", + "content": "\\max(S)" + }, + { + "bbox": [ + 55, + 72, + 296, + 167 + ], + "type": "text", + "content": " as the final score for the image pair. Conversely, if the majority of the heads vote for a negative match, " + }, + { + "bbox": [ + 55, + 72, + 296, + 167 + ], + "type": "inline_equation", + "content": "\\min(S)" + }, + { + "bbox": [ + 55, + 72, + 296, + 167 + ], + "type": "text", + "content": " is used as the final score. Otherwise, we average the scores from the four heads. This approach ensures that the final decision reflects the strongest evidence supporting the consensus among the heads and thereby improves the reliability of the classifier. Eq. 3 elaborates on this voting mechanism:" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 61, + 170, + 296, + 213 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 61, + 170, + 296, + 213 + ], + "spans": [ + { + "bbox": [ + 61, + 170, + 296, + 213 + ], + "type": "interline_equation", + "content": "S _ {\\text {f i n a l}} = \\left\\{ \\begin{array}{l l} \\max (\\mathcal {S}) & \\text {i f} \\sum_ {s \\in \\mathcal {S}} \\mathbf {1} (s > 0. 5) > \\sum_ {s \\in \\mathcal {S}} \\mathbf {1} (s < 0. 5), \\\\ \\min (\\mathcal {S}) & \\text {i f} \\sum_ {s \\in \\mathcal {S}} \\mathbf {1} (s > 0. 5) < \\sum_ {s \\in \\mathcal {S}} \\mathbf {1} (s < 0. 5), \\\\ \\operatorname {m e a n} (\\mathcal {S}) & \\text {i f} \\sum_ {s \\in \\mathcal {S}} \\mathbf {1} (s > 0. 5) = \\sum_ {s \\in \\mathcal {S}} \\mathbf {1} (s < 0. 5). \\end{array} \\right. \\tag {3}", + "image_path": "5f779b4b151828888c4c68c5b3e6caaa384161cd034d2bd89904639d85f35971.jpg" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 55, + 220, + 284, + 232 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 220, + 284, + 232 + ], + "spans": [ + { + "bbox": [ + 55, + 220, + 284, + 232 + ], + "type": "text", + "content": "3.3. Evaluating Doppelganger correction in SfM" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 55, + 237, + 296, + 440 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 237, + 296, + 440 + ], + "spans": [ + { + "bbox": [ + 55, + 237, + 296, + 440 + ], + "type": "text", + "content": "There is currently no reliable benchmarking method for evaluating the accuracy and correctness of SfM reconstructions specifically in terms of how well they address the doppelganger issue. Prior work [3] relied on manual inspection of each out model to assess the effectiveness of their approach—a time-consuming, unquantifiable, and potentially error-prone process. In our work, we propose a benchmarking method for qualitatively evaluating reconstructed models with respect to doppelgangers. We leverage mapping sites like Mapillary, which provide images with location metadata that can serve as probes for validating a 3D model. Note that our method targets common scenarios where geotags are unavailable, or so noisy as to not be useful. Therefore, we do not consider the use of geotags for the reconstruction task itself—we explore pure visual disambiguation—but instead gather them from specialized sources for ground truth evaluation." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 55, + 441, + 296, + 644 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 441, + 296, + 644 + ], + "spans": [ + { + "bbox": [ + 55, + 441, + 296, + 644 + ], + "type": "text", + "content": "Specifically, for a given scene with known geolocation, we acquire from Mapillary a set of nearby geo-tagged images and register them to the reconstructed model with COLMAP [23]. We then apply RANSAC to robustly estimate a similarity transformation between the registered camera positions and their corresponding image geolocation metadata (converted to ECEF coordinates). We use the resulting RANSAC inlier ratio as an indicator of the model's correctness. As an example, in Fig. 4, we can observe that before correction, the registered cameras and their geolocations show significant misalignment, with the cameras collapsing to one side due to similar-looking surfaces. After correction, the registered cameras align closely to their true geolocations, indicating that doppelganger pairs are correctly separated in the model. If the model is split into multiple components, we calculate the weighted average of inliers ratios among the components:" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 90, + 651, + 296, + 684 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 90, + 651, + 296, + 684 + ], + "spans": [ + { + "bbox": [ + 90, + 651, + 296, + 684 + ], + "type": "interline_equation", + "content": "\\operatorname {I R} = \\sum_ {i = 1} ^ {N} \\left(\\frac {I _ {i}}{T _ {i}} \\cdot \\frac {T _ {i}}{\\sum_ {j = 1} ^ {N} T _ {j}}\\right) = \\frac {\\sum_ {i = 1} ^ {N} I _ {i}}{\\sum_ {i = 1} ^ {N} T _ {i}}, \\tag {4}", + "image_path": "6a9c893fdb3f83ae20f0104e226caa4767d7a3a2f8e60ea407ae65a1250472b2.jpg" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 55, + 689, + 296, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 689, + 296, + 714 + ], + "spans": [ + { + "bbox": [ + 55, + 689, + 296, + 714 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 55, + 689, + 296, + 714 + ], + "type": "inline_equation", + "content": "I_{i}" + }, + { + "bbox": [ + 55, + 689, + 296, + 714 + ], + "type": "text", + "content": " is the number of RANSAC inliers and " + }, + { + "bbox": [ + 55, + 689, + 296, + 714 + ], + "type": "inline_equation", + "content": "T_{i}" + }, + { + "bbox": [ + 55, + 689, + 296, + 714 + ], + "type": "text", + "content": " is the number of registered Mapillary images in the " + }, + { + "bbox": [ + 55, + 689, + 296, + 714 + ], + "type": "inline_equation", + "content": "i" + }, + { + "bbox": [ + 55, + 689, + 296, + 714 + ], + "type": "text", + "content": "-th compo" + } + ] + } + ], + "index": 6 + }, + { + "type": "image", + "bbox": [ + 328, + 72, + 539, + 149 + ], + "blocks": [ + { + "bbox": [ + 328, + 72, + 539, + 149 + ], + "lines": [ + { + "bbox": [ + 328, + 72, + 539, + 149 + ], + "spans": [ + { + "bbox": [ + 328, + 72, + 539, + 149 + ], + "type": "image", + "image_path": "62b80a4d91920e948489fbc03315ecd65e5a6cf816cfbc816692fd5ebf430d4e.jpg" + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "image_body" + } + ], + "index": 7 + }, + { + "type": "image", + "bbox": [ + 328, + 151, + 541, + 296 + ], + "blocks": [ + { + "bbox": [ + 328, + 151, + 541, + 296 + ], + "lines": [ + { + "bbox": [ + 328, + 151, + 541, + 296 + ], + "spans": [ + { + "bbox": [ + 328, + 151, + 541, + 296 + ], + "type": "image", + "image_path": "91870a5b6242f432624630df8ba963a4b79e7a1846cd3951dd4191b06f7075c8.jpg" + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 313, + 300, + 555, + 410 + ], + "lines": [ + { + "bbox": [ + 313, + 300, + 555, + 410 + ], + "spans": [ + { + "bbox": [ + 313, + 300, + 555, + 410 + ], + "type": "text", + "content": "Figure 4. Evaluation of doppelganger correction in SfM. (Top) We first collect sequences of geo-tagged Mapillary images around the target location and register them to the SfM model. Then, we use RANSAC to align the registered cameras and their geolocations. The inlier ratio is computed as an indicator of model accuracy. (Bottom) In the model corrupted by doppelganger pairs, the registered cameras all collapse to one side. We see that the camera poses estimated with COLMAP (right, in red) do not align well with the geotags (green), leading to a low inlier ratio, but our method leads to a much closer alignment." + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "image_caption" + } + ], + "index": 8 + }, + { + "bbox": [ + 313, + 419, + 555, + 502 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 419, + 555, + 502 + ], + "spans": [ + { + "bbox": [ + 313, + 419, + 555, + 502 + ], + "type": "text", + "content": "nent. While alignment error could also serve as a metric, we opt to use the inlier ratio to reduce sensitivity to geolocation inaccuracies, as image geolocations may not always be precise. This strategy can also be used to detect broken reconstructions from datasets like MegaScenes [26], which contains over 100K SfM results from Internet photos around the world." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 313, + 517, + 397, + 530 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 517, + 397, + 530 + ], + "spans": [ + { + "bbox": [ + 313, + 517, + 397, + 530 + ], + "type": "text", + "content": "4. Experiments" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 313, + 538, + 429, + 551 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 538, + 429, + 551 + ], + "spans": [ + { + "bbox": [ + 313, + 538, + 429, + 551 + ], + "type": "text", + "content": "4.1. Experimental setup" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 313, + 556, + 555, + 628 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 556, + 555, + 628 + ], + "spans": [ + { + "bbox": [ + 313, + 556, + 555, + 628 + ], + "type": "text", + "content": "Dataset. Our training set is comprised of the Doppelganger dataset [3] (dubbed DG) and the new VisymScenes dataset. Specifically, the DG training set includes 73K image pairs, nearly evenly split between positive and negative pairs. From VisymScenes, we use 47K image pairs across 26 sites as training data, evenly split into positive and negative pairs." + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 313, + 630, + 556, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 630, + 556, + 715 + ], + "spans": [ + { + "bbox": [ + 313, + 630, + 556, + 715 + ], + "type": "text", + "content": "We evaluate on two tasks, 1) pairwise image visual disambiguation and 2) scene reconstruction by integrating our classifier into COLMAP [23]. For pairwise image visual disambiguation, we utilize the DG test set, and mined an additional 3,180 pairs from 7 Visym sites (distinct from the 26 training sites), equally divided into positive and negative pairs. Additionally, we created a test set from the Mapillary" + } + ] + } + ], + "index": 14 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "type": "text", + "content": "27170" + } + ] + } + ], + "index": 15 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 4 + }, + { + "para_blocks": [ + { + "bbox": [ + 55, + 72, + 294, + 156 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 72, + 294, + 156 + ], + "spans": [ + { + "bbox": [ + 55, + 72, + 294, + 156 + ], + "type": "text", + "content": "Street-Level Sequences dataset [30]. This Mapillary dataset spans diverse urban and suburban environments and captures a wide range of seasons, weather conditions, cameras, lighting, and structural settings, with each image geo-tagged by GPS and compass angle. Using a similar filtering approach to the VisymScenes dataset, we mined 1,500 positive and 1,500 negative pairs as an out-of-domain test set." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 55, + 156, + 294, + 228 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 156, + 294, + 228 + ], + "spans": [ + { + "bbox": [ + 55, + 156, + 294, + 228 + ], + "type": "text", + "content": "For the scene reconstruction task, we evaluate on 16 scenes sampled from Heinly et al. [12], Wilson et al. [31], MegaScenes [26] and 5 VisymScenes test scenes. These scenes are challenging for conventional SfM pipelines as well as prior disambiguation methods due to subtle differences between distinct surfaces and repetitive patterns." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 55, + 232, + 294, + 304 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 232, + 294, + 304 + ], + "spans": [ + { + "bbox": [ + 55, + 232, + 294, + 304 + ], + "type": "text", + "content": "Metrics. For pairwise visual disambiguation evaluation, we report Average Precision (AP) and ROC AUC scores, following the protocol in [3]. Additionally, we report precision at a fixed recall and recall at a fixed precision. These statistics help characterize a model's trade-off between achieving high precision and potentially missing true positives." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 55, + 304, + 295, + 471 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 304, + 295, + 471 + ], + "spans": [ + { + "bbox": [ + 55, + 304, + 295, + 471 + ], + "type": "text", + "content": "To comprehensively evaluate SfM reconstructions, we report the number of images in the SfM results, and the geo-alignment inlier ratio described in Sec. 3.3. Respectively, more images in the SfM model imply less false pair pruning, i.e., better class separation; and a higher geo-alignment inlier ratio suggests that the reconstructed model is more likely to be accurate and complete. Note that the inlier ratio is also influenced by the accuracy of the geo-tagged images collected from Mapillary. Thus, if one test scene has more accurate geo-tagged images than another, it is likely to yield a higher inlier ratio. Therefore, rather than comparing inlier ratios across different scenes, this metric is more useful for evaluating different reconstruction methods on the same scene, as the same set of geo-tagged images is used." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 55, + 475, + 295, + 559 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 475, + 295, + 559 + ], + "spans": [ + { + "bbox": [ + 55, + 475, + 295, + 559 + ], + "type": "text", + "content": "Baselines. Our primary baseline is [3], as their learning-based approach consistently outperforms heuristic-based methods like thresholding on the number of post-RANSAC inliers (e.g. SIFT [17] + RANSAC [9], D2-Net [8] + RANSAC, SuperPoint [6] + SuperGlue [22]) for pairwise visual disambiguation; and is also superior than previous approaches [5, 12, 31, 32] for SfM disambiguation." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 55, + 564, + 152, + 577 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 564, + 152, + 577 + ], + "spans": [ + { + "bbox": [ + 55, + 564, + 152, + 577 + ], + "type": "text", + "content": "4.2. Implementation" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 55, + 582, + 295, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 582, + 295, + 713 + ], + "spans": [ + { + "bbox": [ + 55, + 582, + 295, + 713 + ], + "type": "text", + "content": "Model details. Our doppelganger classifier head " + }, + { + "bbox": [ + 55, + 582, + 295, + 713 + ], + "type": "inline_equation", + "content": "\\mathrm{Head}_{\\mathrm{dopp}}" + }, + { + "bbox": [ + 55, + 582, + 295, + 713 + ], + "type": "text", + "content": " is implemented with a Transformer encoder comprised of 3 layers, each with input and output dimension of 768. Each layer has 8 attention heads and a two-layer feed-forward network with a hidden dimension of 2048. Pre-layer normalization and residual connections are applied within each layer. The transformed tokens are aggregated via max pooling and linearly projected into " + }, + { + "bbox": [ + 55, + 582, + 295, + 713 + ], + "type": "inline_equation", + "content": "\\mathrm{pred}^v" + }, + { + "bbox": [ + 55, + 582, + 295, + 713 + ], + "type": "text", + "content": ". We freeze the weights of the public MAST3R [16] model and only train the two doppelganger classification heads, supervised by a cross-entropy loss. We train for 5 epochs with a batch size of 8" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 313, + 72, + 553, + 96 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 72, + 553, + 96 + ], + "spans": [ + { + "bbox": [ + 313, + 72, + 553, + 96 + ], + "type": "text", + "content": "using Adam [15] with a learning rate of " + }, + { + "bbox": [ + 313, + 72, + 553, + 96 + ], + "type": "inline_equation", + "content": "1 \\times 10^{-4}" + }, + { + "bbox": [ + 313, + 72, + 553, + 96 + ], + "type": "text", + "content": ". More details can be found in the supplementary." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 313, + 102, + 555, + 294 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 102, + 555, + 294 + ], + "spans": [ + { + "bbox": [ + 313, + 102, + 555, + 294 + ], + "type": "text", + "content": "Integration with SfM. Following [3], we integrate Doppelgangers++ into SfM to enhance its disambiguation ability. SfM takes a collection of images " + }, + { + "bbox": [ + 313, + 102, + 555, + 294 + ], + "type": "inline_equation", + "content": "\\mathcal{I} = \\{I_i\\}_{i=0}^n" + }, + { + "bbox": [ + 313, + 102, + 555, + 294 + ], + "type": "text", + "content": " and generates geometrically verified image pairs " + }, + { + "bbox": [ + 313, + 102, + 555, + 294 + ], + "type": "inline_equation", + "content": "\\mathcal{P} = \\{(I_p, I_q)\\}" + }, + { + "bbox": [ + 313, + 102, + 555, + 294 + ], + "type": "text", + "content": ", forming a scene graph " + }, + { + "bbox": [ + 313, + 102, + 555, + 294 + ], + "type": "inline_equation", + "content": "\\mathcal{G} = (\\mathcal{I}, \\mathcal{P})" + }, + { + "bbox": [ + 313, + 102, + 555, + 294 + ], + "type": "text", + "content": " with images as nodes and pairs as edges. Using " + }, + { + "bbox": [ + 313, + 102, + 555, + 294 + ], + "type": "inline_equation", + "content": "\\mathcal{G}" + }, + { + "bbox": [ + 313, + 102, + 555, + 294 + ], + "type": "text", + "content": ", SfM computes camera poses and reconstructs a 3D point cloud. Our doppelganger classifier acts as a filter on edges in " + }, + { + "bbox": [ + 313, + 102, + 555, + 294 + ], + "type": "inline_equation", + "content": "\\mathcal{G}" + }, + { + "bbox": [ + 313, + 102, + 555, + 294 + ], + "type": "text", + "content": ", removing pairs below a probability threshold " + }, + { + "bbox": [ + 313, + 102, + 555, + 294 + ], + "type": "inline_equation", + "content": "\\tau" + }, + { + "bbox": [ + 313, + 102, + 555, + 294 + ], + "type": "text", + "content": " to eliminate spurious matches due to repeated or symmetric structures. We integrate Doppelgangers++ into both COLMAP [23] and MASt3R-SfM [7] to showcase the effectiveness of the method. Notably, with COLMAP, SIFT features are computed for scene graph construction and MASt3R features for pruning; whereas with MASt3R-SfM, MASt3R features serve for both scene graph construction and pruning, forming a more efficient pipeline." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 313, + 301, + 486, + 314 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 301, + 486, + 314 + ], + "spans": [ + { + "bbox": [ + 313, + 301, + 486, + 314 + ], + "type": "text", + "content": "4.3. Pairwise Visual Disambiguation" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 313, + 319, + 555, + 450 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 319, + 555, + 450 + ], + "spans": [ + { + "bbox": [ + 313, + 319, + 555, + 450 + ], + "type": "text", + "content": "We compare our approach with the original Doppelganger work [3] (dubbed " + }, + { + "bbox": [ + 313, + 319, + 555, + 450 + ], + "type": "inline_equation", + "content": "DG-OG" + }, + { + "bbox": [ + 313, + 319, + 555, + 450 + ], + "type": "text", + "content": ") under two experimental settings: 1) we use the same training data as in [3], i.e. the DG training set, to train our model. Both models are evaluated on three benchmark test sets: DG, VisymScenes, and Mapillary. Under this configuration, the DG test set is in-domain, while VisymScenes and Mapillary are out-of-domain scenarios; 2) we expand the training data by including the new VisymScenes training set and retrain both DG-OG and our model. Evaluation is conducted on the same three test sets, with only the Mapillary test set being out-of-domain this time." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 313, + 450, + 554, + 569 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 450, + 554, + 569 + ], + "spans": [ + { + "bbox": [ + 313, + 450, + 554, + 569 + ], + "type": "text", + "content": "Quantitative results are in Tab. 1. Under the first setting, where both methods are trained on DG, our model shows clear improvements across all three test sets. Specifically, on the in-domain DG test set, our model achieves higher AP and ROC AUC, indicating that it maintains high precision across all recall levels, while also being less sensitive to the decision threshold " + }, + { + "bbox": [ + 313, + 450, + 554, + 569 + ], + "type": "inline_equation", + "content": "\\tau" + }, + { + "bbox": [ + 313, + 450, + 554, + 569 + ], + "type": "text", + "content": " with better class separation. On the out-of-domain test sets (VisymScenes and Mapillary), we observe " + }, + { + "bbox": [ + 313, + 450, + 554, + 569 + ], + "type": "inline_equation", + "content": "25\\%" + }, + { + "bbox": [ + 313, + 450, + 554, + 569 + ], + "type": "text", + "content": " to " + }, + { + "bbox": [ + 313, + 450, + 554, + 569 + ], + "type": "inline_equation", + "content": "65\\%" + }, + { + "bbox": [ + 313, + 450, + 554, + 569 + ], + "type": "text", + "content": " improvements in both AP and ROC AUC, highlighting the generalizability of our method." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 313, + 570, + 555, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 570, + 555, + 713 + ], + "spans": [ + { + "bbox": [ + 313, + 570, + 555, + 713 + ], + "type": "text", + "content": "Although our model demonstrates improved performance even when trained solely on the DG dataset, its precision outside of the training domain is suboptimal for practical use in SfM, where high precision is essential. In the second setting, we expand the training set by including VisymScenes data. Both methods maintain similar performance on the DG test set, with our model experiencing a slight drop in recall when precision is set to 0.99. On the VisymScenes test set, both models show improvements across all metrics, with ours reaching " + }, + { + "bbox": [ + 313, + 570, + 555, + 713 + ], + "type": "inline_equation", + "content": "99\\%" + }, + { + "bbox": [ + 313, + 570, + 555, + 713 + ], + "type": "text", + "content": " in both AP and ROC AUC. Notably, adding VisymScenes to the training set does not enhance DG-OG's performance on the out-of-domain Mapillary test" + } + ] + } + ], + "index": 12 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "text", + "content": "27171" + } + ] + } + ], + "index": 13 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 5 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 59, + 70, + 184, + 224 + ], + "blocks": [ + { + "bbox": [ + 59, + 70, + 184, + 224 + ], + "lines": [ + { + "bbox": [ + 59, + 70, + 184, + 224 + ], + "spans": [ + { + "bbox": [ + 59, + 70, + 184, + 224 + ], + "type": "image", + "image_path": "4f0612e82fbcaa2337ccce1cdf0c08c41df84037639aa9dfccb0179ee5051fb2.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 55, + 227, + 555, + 270 + ], + "lines": [ + { + "bbox": [ + 55, + 227, + 555, + 270 + ], + "spans": [ + { + "bbox": [ + 55, + 227, + 555, + 270 + ], + "type": "text", + "content": "Figure 5. SfM Disambiguation on MegaScenes [26]. (White background) SfM results from DG-OG [3] and ours. (Black background) Verification using geo-tagged images, red points represent registered cameras and green points represent geolocations, inlier ratio (IR) is labeled on the bottom right. DG-OG fails to disambiguate this scene, predicting incorrect scores for image pairs. Our method correctly splits the model into two clean components." + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "type": "image", + "bbox": [ + 192, + 70, + 296, + 224 + ], + "blocks": [ + { + "bbox": [ + 192, + 70, + 296, + 224 + ], + "lines": [ + { + "bbox": [ + 192, + 70, + 296, + 224 + ], + "spans": [ + { + "bbox": [ + 192, + 70, + 296, + 224 + ], + "type": "image", + "image_path": "5ef5e6ed4c8f5c689e62b63d73aa78bb99b183eaa71deaf1c989fc812d448771.jpg" + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_body" + } + ], + "index": 1 + }, + { + "type": "image", + "bbox": [ + 302, + 70, + 553, + 224 + ], + "blocks": [ + { + "bbox": [ + 302, + 70, + 553, + 224 + ], + "lines": [ + { + "bbox": [ + 302, + 70, + 553, + 224 + ], + "spans": [ + { + "bbox": [ + 302, + 70, + 553, + 224 + ], + "type": "image", + "image_path": "56284431ec2eb48111d90ea0674c2dd93c3616c37e3d1cc1fb00f06f8571dc78.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + } + ], + "index": 2 + }, + { + "type": "image", + "bbox": [ + 56, + 274, + 294, + 393 + ], + "blocks": [ + { + "bbox": [ + 56, + 274, + 294, + 393 + ], + "lines": [ + { + "bbox": [ + 56, + 274, + 294, + 393 + ], + "spans": [ + { + "bbox": [ + 56, + 274, + 294, + 393 + ], + "type": "image", + "image_path": "e25ab08f7ec9f681f89a95e8c932bc732295996420d7fbc44e3795fef35161d0.jpg" + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 55, + 396, + 295, + 440 + ], + "lines": [ + { + "bbox": [ + 55, + 396, + 295, + 440 + ], + "spans": [ + { + "bbox": [ + 55, + 396, + 295, + 440 + ], + "type": "text", + "content": "Figure 6. SfM disambiguation on VisymScenes. We show that our classifier is more robust than DG-OG [3] on test scenes from new domains, like everyday street scenes. DG-OG has difficulty disambiguating such scenes, leading to incorrect geometry." + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "image_caption" + } + ], + "index": 4 + }, + { + "type": "table", + "bbox": [ + 57, + 446, + 295, + 516 + ], + "blocks": [ + { + "bbox": [ + 57, + 446, + 295, + 516 + ], + "lines": [ + { + "bbox": [ + 57, + 446, + 295, + 516 + ], + "spans": [ + { + "bbox": [ + 57, + 446, + 295, + 516 + ], + "type": "table", + "html": "
Test DataMethodMetrics (trained on DG / trained on DG + VisymScenes)
AP↑ROC AUC↑Prec@Recall=0.85↑Recall@Prec=0.99↑
DGDG-OG0.954 / 0.9560.944 / 0.9470.901 / 0.9100.611 / 0.614
Ours0.980 / 0.9810.981 / 0.9810.972 / 0.9820.690 / 0.642
VisymScenesDG-OG0.816 / 0.9380.726 / 0.9210.498 / 0.8310.340 / 0.623
Ours0.936 / 0.9910.909 / 0.9900.892 / 0.9990.542 / 0.901
MapillaryDG-OG0.566 / 0.6920.581 / 0.7010.523 / 0.5720.003 / 0.000
Ours0.950 / 0.9680.944 / 0.9580.927 / 0.9420.310 / 0.736
", + "image_path": "3ac1602181dcb48c73917e2775afdd1b1703fb4d60406debdf259f41d644ee20.jpg" + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "table_body" + } + ], + "index": 6 + }, + { + "bbox": [ + 55, + 520, + 296, + 597 + ], + "lines": [ + { + "bbox": [ + 55, + 520, + 296, + 597 + ], + "spans": [ + { + "bbox": [ + 55, + 520, + 296, + 597 + ], + "type": "text", + "content": "Table 1. Evaluation of pairwise disambiguation. We evaluate DG-OG [3] and our method trained on DG [3] only and DG +VisymScenes (two numbers per cell) on three test sets. Both DG-OG and ours benefit from dataset expansion, whereas ours gained more generalizability on out-of-domain test set (Mapillary) after training on both. Our classifier constantly demonstrates better precision, recall across all settings." + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "text" + }, + { + "bbox": [ + 55, + 604, + 296, + 653 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 604, + 296, + 653 + ], + "spans": [ + { + "bbox": [ + 55, + 604, + 296, + 653 + ], + "type": "text", + "content": "set, while ours continues to benefit from increased training diversity. These results indicate that the prior approach struggles to generalize effectively, whereas ours shows greater generalization with more varied training data." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 55, + 659, + 261, + 673 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 659, + 261, + 673 + ], + "spans": [ + { + "bbox": [ + 55, + 659, + 261, + 673 + ], + "type": "text", + "content": "4.4. Structure from Motion disambiguation" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 55, + 677, + 297, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 677, + 297, + 713 + ], + "spans": [ + { + "bbox": [ + 55, + 677, + 297, + 713 + ], + "type": "text", + "content": "We integrate our classifier trained on DG and VisymScenes into COLMAP's SfM pipeline [23], and evaluate its performance on reconstructing scenes with duplicated and symmet" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 313, + 277, + 555, + 468 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 277, + 555, + 468 + ], + "spans": [ + { + "bbox": [ + 313, + 277, + 555, + 468 + ], + "type": "text", + "content": "ric structures. We compare the results with vanilla COLMAP and Cai et al. [3] (DG-OG). Quantitative results are presented in Tab. 2. Our approach registers more images across all scenes than DG-OG and operates without threshold tuning. In contrast, DG-OG relies on scene-specific thresholds, such as " + }, + { + "bbox": [ + 313, + 277, + 555, + 468 + ], + "type": "inline_equation", + "content": "\\tau = 0.97" + }, + { + "bbox": [ + 313, + 277, + 555, + 468 + ], + "type": "text", + "content": " for the Church on Spilled Blood to correctly separate both sides of the church, and " + }, + { + "bbox": [ + 313, + 277, + 555, + 468 + ], + "type": "inline_equation", + "content": "\\tau = 0.6" + }, + { + "bbox": [ + 313, + 277, + 555, + 468 + ], + "type": "text", + "content": " for Ponte di Rialto [26] to avoid over-segmentation and maintain completeness. Notably, DG-OG fails to disambiguate Belvedere (Vienna), while our method succeeds with a consistent threshold (" + }, + { + "bbox": [ + 313, + 277, + 555, + 468 + ], + "type": "inline_equation", + "content": "\\tau = 0.8" + }, + { + "bbox": [ + 313, + 277, + 555, + 468 + ], + "type": "text", + "content": " across all scenes). Our approach also consistently achieves a higher inlier ratio than the baselines, indicating greater accuracy and completeness in the reconstructed models. As an example, we qualitatively show the results of reconstructing Belvedere (Vienna) in Fig. 5, along with verification results using geo-tagged images." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 313, + 469, + 556, + 576 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 469, + 556, + 576 + ], + "spans": [ + { + "bbox": [ + 313, + 469, + 556, + 576 + ], + "type": "text", + "content": "We also evaluate on scenes from the VisymScenes test set. Since VisymScenes images come with geolocation metadata, we do not need additional Mapillary images; instead, we directly compute the inlier ratio between SfM camera positions and the geolocation metadata after RANSAC. Results show that our model effectively prunes spurious pairs and improves the reconstructed model quality, whereas DG-OG sometimes fails to distinguish doppelganger pairs. Fig. 6 shows examples where DG-OG encounters difficulties." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 313, + 578, + 556, + 649 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 578, + 556, + 649 + ], + "spans": [ + { + "bbox": [ + 313, + 578, + 556, + 649 + ], + "type": "text", + "content": "In Fig. 7, we show that while MASt3R's features are powerful, MASt3R-SfM is not free from doppelganger issues. Although our classifier was trained on image pairs mined through COLMAP's feature matching module (i.e. w/ SIFT features), it effectively prunes incorrect matches generated by MASt3R, restoring accurate reconstruction results." + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 314, + 658, + 378, + 671 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 314, + 658, + 378, + 671 + ], + "spans": [ + { + "bbox": [ + 314, + 658, + 378, + 671 + ], + "type": "text", + "content": "4.5. Ablation" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 313, + 677, + 555, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 677, + 555, + 713 + ], + "spans": [ + { + "bbox": [ + 313, + 677, + 555, + 713 + ], + "type": "text", + "content": "In this section, we study the effectiveness of our designs from the following aspects: 1) fine-tuning the entire model (vs. only new heads), 2) one classification head (vs. two" + } + ] + } + ], + "index": 15 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "type": "text", + "content": "27172" + } + ] + } + ], + "index": 16 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 6 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 57, + 70, + 294, + 159 + ], + "blocks": [ + { + "bbox": [ + 57, + 70, + 294, + 159 + ], + "lines": [ + { + "bbox": [ + 57, + 70, + 294, + 159 + ], + "spans": [ + { + "bbox": [ + 57, + 70, + 294, + 159 + ], + "type": "image", + "image_path": "d651cbb028479382932dbdcc6c8108c6e8313fe0eb7776ee25109157f3140bd0.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 55, + 161, + 295, + 205 + ], + "lines": [ + { + "bbox": [ + 55, + 161, + 295, + 205 + ], + "spans": [ + { + "bbox": [ + 55, + 161, + 295, + 205 + ], + "type": "text", + "content": "Figure 7. MATLAB's Doppelgangers++. MATLAB also suffers from doppelganger issues. Our classifier effectively prunes false positive pairs and correctly reconstructs challenging scenes." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "type": "table", + "bbox": [ + 58, + 210, + 296, + 356 + ], + "blocks": [ + { + "bbox": [ + 58, + 210, + 296, + 356 + ], + "lines": [ + { + "bbox": [ + 58, + 210, + 296, + 356 + ], + "spans": [ + { + "bbox": [ + 58, + 210, + 296, + 356 + ], + "type": "table", + "html": "
Test Scenes# SFM-registered ImagesInlier Ratio (Sec. 3.3)
COLMAPDG-OGOursCOLMAPDG-OGOurs
Alexander Nevsky Cathedral [12]4474444470.5651.01.0
Arc de Triomphé [12]4243844230.6970.9660.990
Berliner Dom [12]1603158816060.7090.9870.992
Big Ben [12]3983793940.5640.8270.831
Church on Spilled Blood [12]27384+94 (τ=0.97)157+1060.5420.8810.962
Radcliffe Camera [12]28191+8494+1860.4950.9550.970
Seville [12]1498585+272+515615+303+5520.4500.7720.854
Brandenburg Gate [26]21371361+5701398+6030.4400.9000.909
Palacio de Comunicaciones [26]727307+80 (τ=0.6)308+840.2290.8230.934
Ponte di Rialto [26]652538+101 (τ=0.6)540+1070.6270.8340.844
York Minster [26]636200+362206+2840.7270.8580.901
Town Hall Tower, Kraków [26]2982552800.6090.7310.838
Belvedere (Vienna) [26]1038851 (fail)457+5000.5210.4510.874
Reichstag (building) [26]1504997+3101024+3560.4690.8040.862
St. Vitus Cathedral [26]7526736920.8530.9090.933
Royal Liver Building [26]2121711800.71.01.0
VisymSite00101544520+29014460.7700.8200.913
VisymSite0023849471+81566+820.8670.8480.942
VisymSite0028450238+179267+1200.8180.7110.909
VisymSite0042540206+2074670.8630.9240.963
VisymSite01091245237+458+78239+612+1270.8570.8620.927
", + "image_path": "31f0079bb6304c685ccdc64f2e75fd7327aa50f90d59ad4c9e8db3f3921ac0e9.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "table_body" + } + ], + "index": 2 + }, + { + "bbox": [ + 55, + 364, + 296, + 451 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 364, + 296, + 451 + ], + "spans": [ + { + "bbox": [ + 55, + 364, + 296, + 451 + ], + "type": "text", + "content": "Table 2. Evaluation of SfM results. We compare reconstructions from COLMAP [23], DG-OG [3], and our method. " + }, + { + "bbox": [ + 55, + 364, + 296, + 451 + ], + "type": "inline_equation", + "content": "\\tau = 0.8" + }, + { + "bbox": [ + 55, + 364, + 296, + 451 + ], + "type": "text", + "content": " is used unless otherwise stated. The " + }, + { + "bbox": [ + 55, + 364, + 296, + 451 + ], + "type": "inline_equation", + "content": "^+" + }, + { + "bbox": [ + 55, + 364, + 296, + 451 + ], + "type": "text", + "content": " indicates split reconstruction components; e.g., DG-OG and our method split the Radcliffe Camera reconstruction into two components. Because VisymScenes scenes are large, we report statistics on the largest reconstruction component produced by COLMAP, and identify the corresponding (split) components in the results of DG-OG and ours." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 55, + 452, + 296, + 524 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 452, + 296, + 524 + ], + "spans": [ + { + "bbox": [ + 55, + 452, + 296, + 524 + ], + "type": "text", + "content": "separate heads), 3) the architecture of our classification head, and 4) final-layer decoder features (vs. multi-layer decoder features). We evaluate models trained on DG and Visym-Scenes datasets and show PR curves in Fig. 8 on the three test sets with respect to the pairwise classification task. A full ablation table can be found in the supplementary." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 55, + 529, + 296, + 673 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 529, + 296, + 673 + ], + "spans": [ + { + "bbox": [ + 55, + 529, + 296, + 673 + ], + "type": "text", + "content": "Fine-tuning MASt3R. As discussed in Sec.3.2, we choose to train a lighter output head on top of MASt3R features rather than fine-tuning the entire model weights. Curves in Fig. 8 show that training only the head achieves comparable performance to full model fine-tuning. The drop in recall with full model fine-tuning suggests potential overfitting to training data, making the model more conservative and less generalizable to unseen inputs. This tendency also holds on models trained solely on DG data: on the VisymScenes test set, our approach achieved 0.03 higher AP than fine-tuning. On Mapillary test set, ours achieved comparable AP but with 0.2 higher recall at precision 0.99." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 55, + 677, + 296, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 677, + 296, + 713 + ], + "spans": [ + { + "bbox": [ + 55, + 677, + 296, + 713 + ], + "type": "text", + "content": "Single classification head. Our design uses two classification heads to process features from the two branches separately. An alternative approach would be to combine the" + } + ] + } + ], + "index": 6 + }, + { + "type": "image", + "bbox": [ + 318, + 72, + 553, + 159 + ], + "blocks": [ + { + "bbox": [ + 318, + 72, + 553, + 159 + ], + "lines": [ + { + "bbox": [ + 318, + 72, + 553, + 159 + ], + "spans": [ + { + "bbox": [ + 318, + 72, + 553, + 159 + ], + "type": "image", + "image_path": "07045ff87fba6e6f207b98a39d7be224e82c47778c57e18f5ba83a9633309a9d.jpg" + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 313, + 162, + 555, + 196 + ], + "lines": [ + { + "bbox": [ + 313, + 162, + 555, + 196 + ], + "spans": [ + { + "bbox": [ + 313, + 162, + 555, + 196 + ], + "type": "text", + "content": "Figure 8. Precision-Recall curves of ablation studies. Metrics are evaluated on models trained with DG and VisymScenes. Full tables can be found in the supplementary." + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "image_caption" + } + ], + "index": 7 + }, + { + "bbox": [ + 313, + 201, + 555, + 310 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 201, + 555, + 310 + ], + "spans": [ + { + "bbox": [ + 313, + 201, + 555, + 310 + ], + "type": "text", + "content": "features from both decoder branches and use a single head to predict the doppelganger score. As shown in Fig. 8, this alternative degrades performance consistently. This may be due to the asymmetric nature of the MAST3R design, where each branch captures different kinds of information that are better to be processed separately. On the other hand, our voting mechanism between the two heads helps enhance class distinction, outputting lower scores for negative samples and higher scores for positive samples." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 313, + 316, + 556, + 400 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 316, + 556, + 400 + ], + "spans": [ + { + "bbox": [ + 313, + 316, + 556, + 400 + ], + "type": "text", + "content": "Head architecture. While the decoder features of MASt3R retain rich geometric 3D information, designing a classifier with sufficient capacity to effectively leverage this information for the doppelganger prediction task is non-trivial. We compare the performance of our Transformer-based classifier with that of an MLP. Fig. 8 shows that our Transformer model outperforms the MLP across all metrics and test sets." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 313, + 406, + 556, + 491 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 406, + 556, + 491 + ], + "spans": [ + { + "bbox": [ + 313, + 406, + 556, + 491 + ], + "type": "text", + "content": "Single-layer decoder feature. Our classifier head takes multi-layer decoder features as input. Here we substitute with only the final-layer decoder feature. Results in Fig. 8 show that the multi-layer decoder features are consistently superior to the single-layer ones, because the classifier is able to analyze doppelganger factors from different aspects and levels. See supplementary for more detailed analysis." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 314, + 502, + 388, + 514 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 314, + 502, + 388, + 514 + ], + "spans": [ + { + "bbox": [ + 314, + 502, + 388, + 514 + ], + "type": "text", + "content": "5. Conclusion" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 313, + 522, + 556, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 522, + 556, + 713 + ], + "spans": [ + { + "bbox": [ + 313, + 522, + 556, + 713 + ], + "type": "text", + "content": "We propose Doppelgangers++ as an effective approach to handling visual aliasing in 3D reconstruction. We introduce a new VisymScenes dataset, featuring images from diverse daily scenes, and develop rules to mine doppelganger data from the dataset to enrich our training data. We train a Transformer-based classifier that leverages geometric 3D features to classify image pairs with high precision and recall across various scenes. Doppelgangers++ integrates seamlessly into existing SfM pipelines, enhancing reconstruction accuracy without extensive parameter tuning. Further, we propose a validation method using geo-tagged map images, offering a more comprehensive and automatic way to assess SfM accuracy and model completeness. Extensive experiments show that Doppelgangers++ significantly improves visual disambiguation and 3D reconstruction quality in complex scenes." + } + ] + } + ], + "index": 13 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "type": "text", + "content": "27173" + } + ] + } + ], + "index": 14 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 7 + }, + { + "para_blocks": [ + { + "bbox": [ + 55, + 72, + 161, + 85 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 72, + 161, + 85 + ], + "spans": [ + { + "bbox": [ + 55, + 72, + 161, + 85 + ], + "type": "text", + "content": "6. Acknowledgment" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 55, + 91, + 297, + 259 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 91, + 297, + 259 + ], + "spans": [ + { + "bbox": [ + 55, + 91, + 297, + 259 + ], + "type": "text", + "content": "We thank Brandon Li and Joseph Tung for their valuable discussion and help with the webviewer. This work was supported by the Intelligence Advanced Research Projects Activity (IARPA) via Department of Interior/Interior Business Center (DOI/IBC) contract number 140D0423C0035. The U.S. Government is authorized to reproduce and distribute reprints for governmental purposes notwithstanding any copyright annotation thereon. Disclaimer: The views and conclusions contained herein are those of the authors and should not be interpreted as necessarily representing the official policies or endorsements, either expressed or implied, of IARPA, DOI/IBC, or the U.S. Government. This research was also supported in part by the National Science Foundation under award IIS-2212084." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 56, + 273, + 115, + 285 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 273, + 115, + 285 + ], + "spans": [ + { + "bbox": [ + 56, + 273, + 115, + 285 + ], + "type": "text", + "content": "References" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 61, + 293, + 297, + 712 + ], + "type": "list", + "angle": 0, + "index": 12, + "blocks": [ + { + "bbox": [ + 61, + 293, + 297, + 357 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 61, + 293, + 297, + 357 + ], + "spans": [ + { + "bbox": [ + 61, + 293, + 297, + 357 + ], + "type": "text", + "content": "[1] Eric Brachmann, Jamie Wynn, Shuai Chen, Tommaso Cavallari, Aron Monszpart, Daniyar Turmukhambetov, and Victor Adrian Prisacariu. Scene coordinate reconstruction: posing of image collections via incremental learning of a relocalizer. In Eur. Conf. Comput. Vis., pages 421-440. Springer, 2025. 3" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 61, + 361, + 296, + 393 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 61, + 361, + 296, + 393 + ], + "spans": [ + { + "bbox": [ + 61, + 361, + 296, + 393 + ], + "type": "text", + "content": "[2] J. Byrne, G. Castanon, Z. Li, and G. Ettinger. Fine-grained activities of people worldwide. In IEEE/CVF Winter Conference on Applications of Computer Vision, 2023. 2, 3" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 62, + 395, + 296, + 438 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 395, + 296, + 438 + ], + "spans": [ + { + "bbox": [ + 62, + 395, + 296, + 438 + ], + "type": "text", + "content": "[3] Ruojin Cai, Joseph Tung, Qianqian Wang, Hadar Averbuch-Elor, Bharath Hariharan, and Noah Snavely. Doppelgangers: Learning to disambiguate images of similar structures. In Int. Conf. Comput. Vis., pages 34-44, 2023. 1, 2, 3, 5, 6, 7, 8" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 62, + 441, + 296, + 483 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 441, + 296, + 483 + ], + "spans": [ + { + "bbox": [ + 62, + 441, + 296, + 483 + ], + "type": "text", + "content": "[4] Duygu Ceylan, Niloy Jyoti Mitra, Youyi Zheng, and Mark Pauly. Coupled structure-from-motion and 3d symmetry detection for urban facades. ACM Transactions on Graphics (TOG), 33:1-15, 2014. 2" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 62, + 487, + 296, + 519 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 487, + 296, + 519 + ], + "spans": [ + { + "bbox": [ + 62, + 487, + 296, + 519 + ], + "type": "text", + "content": "[5] Zhaopeng Cui and Ping Tan. Global structure-from-motion by similarity averaging. 2015 IEEE International Conference on Computer Vision (ICCV), pages 864-872, 2015. 6" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 62, + 521, + 296, + 564 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 521, + 296, + 564 + ], + "spans": [ + { + "bbox": [ + 62, + 521, + 296, + 564 + ], + "type": "text", + "content": "[6] Daniel DeTone, Tomasz Malisiewicz, and Andrew Rabinovich. Superpoint: Self-supervised interest point detection and description. In Proc. CVPR Workshops, pages 224-236, 2018. 2, 6" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 62, + 567, + 296, + 610 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 567, + 296, + 610 + ], + "spans": [ + { + "bbox": [ + 62, + 567, + 296, + 610 + ], + "type": "text", + "content": "[7] Bardienus Pieter Duisterhof, Lojze Žust, Philippe Weinzaepfel, Vincent Leroy, Yohann Cabon, and Jérôme Revaud. Mast3r-sfm: a fully-integrated solution for unconstrained structure-from-motion. ArXiv, abs/2409.19152, 2024. 2, 3, 6" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 62, + 613, + 296, + 667 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 613, + 296, + 667 + ], + "spans": [ + { + "bbox": [ + 62, + 613, + 296, + 667 + ], + "type": "text", + "content": "[8] Mihai Dusmanu, Ignacio Rocco, Tomás Pajdla, Marc Pollefeys, Josef Sivic, Akihiko Torii, and Torsten Sattler. D2-net: A trainable cnn for joint description and detection of local features. 2019 IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 8084-8093, 2019. 6" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 62, + 670, + 296, + 712 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 670, + 296, + 712 + ], + "spans": [ + { + "bbox": [ + 62, + 670, + 296, + 712 + ], + "type": "text", + "content": "[9] Martin A. Fischler and Robert C. Bolles. Random sample consensus: a paradigm for model fitting with applications to image analysis and automated cartography. Commun. ACM, 24:381-395, 1981. 6" + } + ] + } + ], + "index": 11 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 316, + 73, + 555, + 712 + ], + "type": "list", + "angle": 0, + "index": 28, + "blocks": [ + { + "bbox": [ + 316, + 73, + 555, + 127 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 73, + 555, + 127 + ], + "spans": [ + { + "bbox": [ + 316, + 73, + 555, + 127 + ], + "type": "text", + "content": "[10] Ye Gong, Pengwei Zhou, Changfeng Liu, Yan Yu, Jian Yao, Wei Yuan, and Li Li. A cluster-based disambiguation method using pose consistency verification for structure from motion. ISPRS Journal of Photogrammetry and Remote Sensing, 2024. 2" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 316, + 130, + 555, + 173 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 130, + 555, + 173 + ], + "spans": [ + { + "bbox": [ + 316, + 130, + 555, + 173 + ], + "type": "text", + "content": "[11] Jared Heinly, Enrique Dunn, and Jan-Michael Frahm. Recovering correct reconstructions from indistinguishable geometry. 2014 2nd International Conference on 3D Vision, 1:377-384, 2014. 2" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 316, + 175, + 555, + 218 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 175, + 555, + 218 + ], + "spans": [ + { + "bbox": [ + 316, + 175, + 555, + 218 + ], + "type": "text", + "content": "[12] Jared Heinly, Enrique Dunn, and Jan-Michael Frahm. Correcting for duplicate scene structure in sparse 3d reconstruction. In Eur. Conf. Comput. Vis., pages 780-795. Springer, 2014. 6, 8" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 316, + 221, + 555, + 265 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 221, + 555, + 265 + ], + "spans": [ + { + "bbox": [ + 316, + 221, + 555, + 265 + ], + "type": "text", + "content": "[13] Nianjuan Jiang, Ping Tan, and Loong-Fah Cheong. Seeing double without confusion: Structure-from-motion in highly ambiguous scenes. In Conf. Comput. Vis. Pattern Recog., pages 1458-1465, 2012. 2" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 316, + 267, + 555, + 310 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 267, + 555, + 310 + ], + "spans": [ + { + "bbox": [ + 316, + 267, + 555, + 310 + ], + "type": "text", + "content": "[14] Rajbir Kataria, Joseph DeGol, and Derek Hoiem. Improving structure from motion with reliable resectioning. 2020 International Conference on 3D Vision (3DV), pages 41-50, 2020. 2" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 316, + 313, + 555, + 334 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 313, + 555, + 334 + ], + "spans": [ + { + "bbox": [ + 316, + 313, + 555, + 334 + ], + "type": "text", + "content": "[15] Diederik P. Kingma and Jimmy Ba. Adam: A method for stochastic optimization. CoRR, abs/1412.6980, 2014. 6" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 316, + 336, + 555, + 369 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 336, + 555, + 369 + ], + "spans": [ + { + "bbox": [ + 316, + 336, + 555, + 369 + ], + "type": "text", + "content": "[16] Vincent Leroy, Yohann Cabon, and Jérôme Revaud. Grounding image matching in 3d with mast3r. ArXiv, abs/2406.09756, 2024. 2, 3, 6" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 316, + 372, + 555, + 403 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 372, + 555, + 403 + ], + "spans": [ + { + "bbox": [ + 316, + 372, + 555, + 403 + ], + "type": "text", + "content": "[17] David G Lowe. Distinctive image features from scale-invariant keypoints. Int. J. Comput. Vis., 60:91-110, 2004. 2, 6" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 316, + 406, + 555, + 450 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 406, + 555, + 450 + ], + "spans": [ + { + "bbox": [ + 316, + 406, + 555, + 450 + ], + "type": "text", + "content": "[18] Lalit Manam and Venu Madhav Govindu. Leveraging camera triplets for efficient and accurate structure-from-motion. 2024 IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 4959-4968, 2024. 2" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 316, + 452, + 555, + 495 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 452, + 555, + 495 + ], + "spans": [ + { + "bbox": [ + 316, + 452, + 555, + 495 + ], + "type": "text", + "content": "[19] Yang Peng, Shen Yan, Yuxiang Liu, Y. Liu, and Maojun Zhang. View graph construction for scenes with duplicate structures via graph convolutional network. IET Comput. Vis., 16:389-402, 2022. 2" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 316, + 497, + 555, + 530 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 497, + 555, + 530 + ], + "spans": [ + { + "bbox": [ + 316, + 497, + 555, + 530 + ], + "type": "text", + "content": "[20] Richard Roberts, Sudipta N. Sinha, Richard Szeliski, and Drew Steedly. Structure from motion for scenes with large duplicate structures. CVPR 2011, pages 3137-3144, 2011. 2" + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 316, + 533, + 555, + 575 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 533, + 555, + 575 + ], + "spans": [ + { + "bbox": [ + 316, + 533, + 555, + 575 + ], + "type": "text", + "content": "[21] Ignacio Rocco, Mircea Cimpoi, Relja Arandjelović, Akihiko Torii, Tomas Pajdla, and Josef Sivic. Neighbourhood consensus networks. In Adv. Neural Inform. Process. Syst., 2018. 2" + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 316, + 578, + 555, + 632 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 578, + 555, + 632 + ], + "spans": [ + { + "bbox": [ + 316, + 578, + 555, + 632 + ], + "type": "text", + "content": "[22] Paul-Edouard Sarlin, Daniel DeTone, Tomasz Malisiewicz, and Andrew Rabinovich. Superglue: Learning feature matching with graph neural networks. 2020 IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 4937–4946, 2019. 6" + } + ] + } + ], + "index": 25 + }, + { + "bbox": [ + 316, + 635, + 555, + 667 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 635, + 555, + 667 + ], + "spans": [ + { + "bbox": [ + 316, + 635, + 555, + 667 + ], + "type": "text", + "content": "[23] Johannes L Schonberger and Jan-Michael Frahm. Structure-from-motion revisited. In Conf. Comput. Vis. Pattern Recog., pages 4104-4113, 2016. 2, 3, 5, 6, 7, 8" + } + ] + } + ], + "index": 26 + }, + { + "bbox": [ + 316, + 670, + 555, + 712 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 670, + 555, + 712 + ], + "spans": [ + { + "bbox": [ + 316, + 670, + 555, + 712 + ], + "type": "text", + "content": "[24] Tianwei Shen, Siyu Zhu, Tian Fang, Runze Zhang, and Long Quan. Graph-based consistent matching for structure-from-motion. In European Conference on Computer Vision, 2016. 2" + } + ] + } + ], + "index": 27 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 318, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 318, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 318, + 757 + ], + "type": "text", + "content": "27174" + } + ] + } + ], + "index": 29 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 8 + }, + { + "para_blocks": [ + { + "bbox": [ + 56, + 73, + 296, + 552 + ], + "type": "list", + "angle": 0, + "index": 11, + "blocks": [ + { + "bbox": [ + 56, + 73, + 296, + 116 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 73, + 296, + 116 + ], + "spans": [ + { + "bbox": [ + 56, + 73, + 296, + 116 + ], + "type": "text", + "content": "[25] Jiaming Sun, Zehong Shen, Yuang Wang, Hujun Bao, and Xiaowei Zhou. Loftr: Detector-free local feature matching with transformers. In Conf. Comput. Vis. Pattern Recog., pages 8922-8931, 2021. 2" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 56, + 118, + 296, + 161 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 118, + 296, + 161 + ], + "spans": [ + { + "bbox": [ + 56, + 118, + 296, + 161 + ], + "type": "text", + "content": "[26] Joseph Tung, Gene Chou, Ruojin Cai, Guandao Yang, Kai Zhang, Gordon Wetzstein, Bharath Hariharan, and Noah Snavely. Megascenes: Scene-level view synthesis at scale. ArXiv, abs/2406.11819, 2024. 5, 6, 7, 8" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 56, + 163, + 296, + 205 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 163, + 296, + 205 + ], + "spans": [ + { + "bbox": [ + 56, + 163, + 296, + 205 + ], + "type": "text", + "content": "[27] Fan Wang, Aditi Nayak, Yogesh Agrawal, and Roy Shilkrot. Hierarchical image link selection scheme for duplicate structure disambiguation. In British Machine Vision Conference, 2018. 2" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 56, + 207, + 296, + 251 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 207, + 296, + 251 + ], + "spans": [ + { + "bbox": [ + 56, + 207, + 296, + 251 + ], + "type": "text", + "content": "[28] Jianyuan Wang, Nikita Karaev, Christian Rupprecht, and David Novotny. Vggsfm: Visual geometry grounded deep structure from motion. In Conf. Comput. Vis. Pattern Recog., pages 21686-21697, 2024. 3" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 56, + 253, + 296, + 296 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 253, + 296, + 296 + ], + "spans": [ + { + "bbox": [ + 56, + 253, + 296, + 296 + ], + "type": "text", + "content": "[29] Shuzhe Wang, Vincent Leroy, Yohann Cabon, Boris Chidlovskii, and Jérôme Revaud. DUS3R: Geometric 3d vision made easy. In Conf. Comput. Vis. Pattern Recog., pages 20697-20709, 2023. 2, 3" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 56, + 297, + 296, + 351 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 297, + 296, + 351 + ], + "spans": [ + { + "bbox": [ + 56, + 297, + 296, + 351 + ], + "type": "text", + "content": "[30] Frederik Warburg, Søren Hauberg, Manuel López-Antequera, Pau Gargallo, Yubin Kuang, and Javier Civera. Mapillary street-level sequences: A dataset for lifelong place recognition. In Conf. Comput. Vis. Pattern Recog., pages 2623-2632, 2020. 6" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 56, + 353, + 296, + 386 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 353, + 296, + 386 + ], + "spans": [ + { + "bbox": [ + 56, + 353, + 296, + 386 + ], + "type": "text", + "content": "[31] Kyle Wilson and Noah Snavely. Network principles for sfm: Disambiguating repeated structures with local context. In Int. Conf. Comput. Vis., pages 513-520, 2013. 2, 6, 8" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 56, + 387, + 296, + 430 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 387, + 296, + 430 + ], + "spans": [ + { + "bbox": [ + 56, + 387, + 296, + 430 + ], + "type": "text", + "content": "[32] Qingan Yan, Long Yang, Ling Zhang, and Chunxia Xiao. Distinguishing the indistinguishable: Exploring structural ambiguities via geodesic context. In Conf. Comput. Vis. Pattern Recog., pages 3836-3844, 2017. 2, 6" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 56, + 432, + 296, + 464 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 432, + 296, + 464 + ], + "spans": [ + { + "bbox": [ + 56, + 432, + 296, + 464 + ], + "type": "text", + "content": "[33] Kwang Moo Yi, Eduard Trulls, Vincent Lepetit, and Pascal Fua. Lift: Learned invariant feature transform. In Eur. Conf. Comput. Vis., pages 467-483, 2016. 2" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 56, + 466, + 296, + 508 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 466, + 296, + 508 + ], + "spans": [ + { + "bbox": [ + 56, + 466, + 296, + 508 + ], + "type": "text", + "content": "[34] Christopher Zach, Arnold Irschara, and Horst Bischof. What can missing correspondences tell us about 3d structure and motion? In Conf. Comput. Vis. Pattern Recog., pages 1-8, 2008. 2" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 56, + 510, + 296, + 552 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 510, + 296, + 552 + ], + "spans": [ + { + "bbox": [ + 56, + 510, + 296, + 552 + ], + "type": "text", + "content": "[35] Christopher Zach, Manfred Klopschitz, and Marc Pollefeys. Disambiguating visual relations using loop constraints. In Conf. Comput. Vis. Pattern Recog., pages 1426-1433, 2010." + } + ] + } + ], + "index": 10 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "text", + "content": "27175" + } + ] + } + ], + "index": 12 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 9 + } + ], + "_backend": "vlm", + "_version_name": "2.6.4" +} \ No newline at end of file diff --git a/2025/DoraCycle_ Domain-Oriented Adaptation of Unified Generative Model in Multimodal Cycles/fb2a505a-21d1-4afa-86a3-162bb4136d23_content_list.json b/2025/DoraCycle_ Domain-Oriented Adaptation of Unified Generative Model in Multimodal Cycles/fb2a505a-21d1-4afa-86a3-162bb4136d23_content_list.json new file mode 100644 index 0000000000000000000000000000000000000000..4d342b7fd058e34a4bbded414c5836332f545595 --- /dev/null +++ b/2025/DoraCycle_ Domain-Oriented Adaptation of Unified Generative Model in Multimodal Cycles/fb2a505a-21d1-4afa-86a3-162bb4136d23_content_list.json @@ -0,0 +1,2232 @@ +[ + { + "type": "text", + "text": "DoraCycle: Domain-Oriented Adaptation of Unified Generative Model in Multimodal Cycles", + "text_level": 1, + "bbox": [ + 142, + 128, + 854, + 176 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Rui Zhao, Weijia Mao, Mike Zheng Shou* Show Lab, National University of Singapore", + "bbox": [ + 318, + 203, + 671, + 239 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Abstract", + "text_level": 1, + "bbox": [ + 246, + 273, + 326, + 289 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Adapting generative models to specific domains presents an effective solution for satisfying specialized requirements. However, adapting to some complex domains remains challenging, especially when these domains require substantial paired data to capture the targeted distributions. Since unpaired data from a single modality, such as vision or language, is more readily available, we utilize the bidirectional mappings between vision and language learned by the unified generative model to enable training on unpaired data for domain adaptation. Specifically, we propose DoraCycle, which integrates two multimodal cycles: text-to-image-to-text and image-to-text-to-image. The model is optimized through cross-entropy loss computed at the cycle endpoints, where both endpoints share the same modality. This facilitates self-evolution of the model without reliance on annotated text-image pairs. Experimental results demonstrate that for tasks independent of paired knowledge, such as stylization, DoraCycle can effectively adapt the unified model using only unpaired data. For tasks involving new paired knowledge, such as specific identities, a combination of a small set of paired image-text examples and larger-scale unpaired data is sufficient for effective domain-oriented adaptation. The code will be released at https://github.com/showlab/DoraCycle.", + "bbox": [ + 89, + 306, + 483, + 670 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "1. Introduction", + "text_level": 1, + "bbox": [ + 91, + 698, + 220, + 714 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "The adaptation of pre-trained generative models to specific domains is an important aspect of advancing personalized content creation, from stylized media outputs to customized identity generation [47, 62]. However, effectively adapting generative models to complex domains remains challenging, particularly when these domains require extensive amounts of paired data to accurately capture the desired distributions. For instance, learning the visual styles and character identities across a unique movie, which involves understanding multiple characters, their relationships, and di", + "bbox": [ + 89, + 724, + 483, + 876 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/d3956901487918a2e4270e7958c4ead203a6602070952914a9d1371d9b938ffd.jpg", + "image_caption": [ + "(a) Training on paired data", + "Figure 1. Training paradigms for unified generative models. (a) Traditional training involves using paired image-text data and optimizing the unified model with paired losses for both image-to-text (I2T) and text-to-image (T2I) generation tasks. (b) In contrast, the proposed multimodal cycle training framework leverages unpaired images and texts. By using cycle-consistency losses, the unified model learns to maintain consistency between input and output across modalities, enabling adaptation without the need for extensive paired datasets." + ], + "image_footnote": [], + "bbox": [ + 519, + 273, + 687, + 459 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/6f1ddd95a07228c8cf7903e1ad20354fa39c4018ee5ae15b67d161cfd0657ef9.jpg", + "image_caption": [ + "(b) Training in Multimodal Cycles" + ], + "image_footnote": [], + "bbox": [ + 723, + 273, + 890, + 459 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "verse settings, is a highly complex task that demands vast amounts of paired frames and captions data. Collecting such paired data, especially for multimodal tasks involving vision and language, is often laborious and impractical, limiting the potential to adapt generative models at scale.", + "bbox": [ + 511, + 642, + 906, + 718 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "High-quality image-text paired data is relatively rare and scarce, but unpaired images and texts are readily available in our daily lives, such as video websites, image platforms, and content from novel websites. Therefore, we aim to explore whether it is possible to adapt generative models to target domains based on unpaired data. To achieve this goal, it is crucial for the model to have an internal capability to align the two modalities to some extent. Fortunately, recent advancements in unified generative models have encouraged us to pursue this direction.", + "bbox": [ + 511, + 719, + 908, + 869 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Recent advanced unified generative models [14, 18, 58, 67, 77] have shown great potential in unifying multimodal", + "bbox": [ + 511, + 869, + 906, + 901 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "CVF", + "bbox": [ + 106, + 2, + 181, + 42 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore.", + "bbox": [ + 236, + 0, + 810, + 46 + ], + "page_idx": 0 + }, + { + "type": "page_footnote", + "text": "*Corresponding Author.", + "bbox": [ + 109, + 887, + 238, + 898 + ], + "page_idx": 0 + }, + { + "type": "page_number", + "text": "2835", + "bbox": [ + 482, + 944, + 514, + 955 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "understanding and generation within a single model. These unified models are capable of processing and generating content across different modalities, i.e. vision and language, within a shared framework. By leveraging the bidirectional mappings between vision and language, which are learned by the unified generative model in its pre-training stage, we can map each data from one modality to another and then back to the original modality, as shown in Fig. 1. Through these two mappings, data can be maintained within the same modality, thereby imposing constraints on the deviation introduced in the process. This only requires computing the cross-entropy loss between the data in the same modality, without any paired data supervision.", + "bbox": [ + 89, + 90, + 480, + 287 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "To this end, we introduce DoraCycle, a framework for adapting unified multimodal generative models to target domains, through cycle-consistent learning with unpaired data. Unlike previous adaptation methods that heavily rely on paired text-image data, the proposed framework leverages the shared latent space of unified models to learn consistent transformations between modalities without requiring paired training examples. Specifically, we design two cycles, i.e. text-to-image-to-text (T cycle) and image-to-text-to-image (I cycle). As shown by Fig. 1 (b), leveraging the pre-trained vision-language alignment of the unified model, each multimodal cycle involves two cross-modality mappings while optimizing in the same modality. This enables calculating loss on unpaired data while implicitly refining cross-modal alignment through the intermediate step.", + "bbox": [ + 89, + 290, + 482, + 517 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "In practice, since there is no labeled ground-truth, mapping data to another modality requires multi-step inference, such as predicting the next token multiple times for text generation. Allowing all forward steps to participate in gradient backpropagation can lead to a catastrophic gradient explosion. Therefore, we stop gradients during multi-step inference and use the generated data as pseudo labels for the model to forward once again, allowing gradients to propagate. Moreover, we found that since a complete cycle requires the model to forward twice, it can lead to training instability, with the quality of pseudo data generated in the middle being compromised. To enhance the stability of pseudo data generation, we maintain a slowly updated EMA (Exponential Moving Average) [57] model, which is used for inference to generate pseudo data. Additionally, we employ gradient clipping techniques to avoid conflicts in the optimization directions of the two cycles, further increasing training stability.", + "bbox": [ + 89, + 518, + 482, + 792 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "The experiments indicate that for tasks independent of paired knowledge, such as stylization and domain-specific adaptation, DoraCycle can adapt the unified model with only unpaired data, which is both more practical and scalable. For tasks that require new paired knowledge, such as identity-specific adaptation, DoraCycle effectively utilizes small amounts of paired data along with larger-scale", + "bbox": [ + 89, + 795, + 482, + 900 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "unpaired data, making it a flexible solution for generative adaptation challenges. We conduct extensive experiments that compare DoraCycle to existing methods, showing that our approach achieves comparable or superior results while significantly reducing the need for paired data. This ability to harness large-scale unpaired data, combined with strategic usage of small paired datasets, makes DoraCycle a feasible solution for personalized multimodal content generation across a wide range of applications.", + "bbox": [ + 511, + 90, + 903, + 227 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "2. Related Works", + "text_level": 1, + "bbox": [ + 511, + 242, + 663, + 258 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "2.1. Multimodal Generation and Understanding", + "text_level": 1, + "bbox": [ + 511, + 268, + 883, + 284 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Generating visual contents from text and describing them through natural language have been extensively studied as core multimodal tasks. Advanced generative models [3, 5, 6, 8, 12, 13, 16, 19, 36, 37, 42, 44, 45, 49, 74-76], such as DALL-E [41, 43], Stable Diffusion [46], demonstrate remarkable generation capabilities, producing high-quality and diverse contents from textual prompts. Meanwhile, image captioning models [24, 26, 27, 60, 61, 63], such as mPLUG [31], and BLIP [33], push the boundaries of visual understanding, generating accurate and context-aware descriptions. Additionally, recent advancements in multimodal large language models [32], such as LLaVA [35], MiniGPT-4 [78], and InstructBLIP [11], have significantly improved the ability to understand and reason about visual content.", + "bbox": [ + 511, + 291, + 903, + 517 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Besides the powerful foundational generative models, adapting or customizing them attracts increasing interest, which enables more personalized and specific outputs based on user preferences [7, 9, 17, 20, 29, 30, 38, 68]. Approaches like DreamBooth [47] enable user-specific customization by fine-tuning generative models with personal data, allowing the generation of content tailored to individual needs or preferences.", + "bbox": [ + 511, + 518, + 905, + 640 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "2.2. Unified Multimodal Generative Models", + "text_level": 1, + "bbox": [ + 511, + 651, + 849, + 666 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Unified multimodal generative models aim to bridge the gap between understanding and generation tasks, and integrate vision and language into a single framework, enabling the model to learn shared representations across modalities [1, 2, 14, 18, 53, 56, 64, 65, 67, 69-71, 77]. SEED-X [18] utilizes a unified architecture where visual features extracted from the CLIP ViT encoder [40] are combined with text tokens and fed into a large language model to enable both next-word prediction and image regression tasks. DreamLLM [14] extends the generative capability of large language models by combining multimodal inputs directly into LLMs. Chameleon [58] employs a discrete tokenization approach for both visual and textual inputs, converting all modalities into a unified token space that is processed by a transformer-based architecture. Transfusion [77] in", + "bbox": [ + 511, + 674, + 905, + 900 + ], + "page_idx": 1 + }, + { + "type": "page_number", + "text": "2836", + "bbox": [ + 482, + 945, + 514, + 955 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "produces an advanced integration mechanism that focuses on directly fusing visual encoding with language tokens, allowing the model to effectively translate visual information into textual formats while maintaining the semantic integrity of both modalities. Show-o [67] combines autoregressive modeling with a discrete diffusion process, enabling the generation of high-quality outputs that are aligned across modalities. Our work leverages the advancements made by these foundational models and explores how to adapt the foundational model to specific domains.", + "bbox": [ + 89, + 90, + 480, + 242 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "2.3. Cycle Consistency", + "text_level": 1, + "bbox": [ + 89, + 266, + 264, + 282 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Cycle consistency has been utilized in computer vision and natural language processing as a means to enhance model robustness and consistency [10, 15, 21, 23, 51, 51, 52, 73]. CycleGAN [79] introduced cycle consistency loss to align unpaired image domains, ensuring that mappings between domains (e.g., $\\mathrm{A} \\to \\mathrm{B} \\to \\mathrm{A}$ ) remain consistent. In the field of natural language processing, back-translation employs similar ideas by translating sentences between languages in both directions to improve translation quality [50]. However, the cycle consistency in these works is in a single modality, i.e. vision or language. Recently, ITIT [34] was proposed to utilize cycle consistency to train vision-language generative models. ITIT takes in a mixture of unpaired data and paired data to pre-train the foundational generative model. It is constructed with one image-text encoder and two modality-specific decoders, which operate on the encoded image-text features to generate either text or image tokens. In contrast, we utilized a single unified transformer to parse and predict text and image tokens together. Besides, we focus on adapting pre-trained foundational models to new domains efficiently rather than retraining new foundation models.", + "bbox": [ + 91, + 292, + 483, + 625 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3. Method", + "text_level": 1, + "bbox": [ + 89, + 654, + 181, + 669 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "The proposed DoraCycle framework, as shown in Fig. 2, is built upon the unified generative model designed for multimodal tasks involving both vision and language [58, 67, 77]. The unified model uses a single transformer to learn bidirectional mappings between vision and language, providing a powerful backbone capable of processing and generating different modalities [67]. For captioning, the model takes in image tokens and predicts corresponding text tokens, while for image generation, it takes in text tokens and predicts image tokens. This versatility makes the unified model well-suited as a base for our proposed framework.", + "bbox": [ + 89, + 683, + 482, + 849 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "In the following sections, we first introduce the design of multimodal cycles, and then discuss the stabilization of optimization, and the balance of two cycles.", + "bbox": [ + 89, + 854, + 483, + 901 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3.1. Multimodal Cycles", + "text_level": 1, + "bbox": [ + 513, + 90, + 694, + 107 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "To adapt the unified model for domain-oriented adaptations using unpaired data, we design two multimodal cycles: the Image-to-Text-to-Image Cycle (I Cycle) and the Text-to-Image-to-Text Cycle (T Cycle). Each cycle utilizes data from a single modality, allowing the model to adapt without relying on paired data.", + "bbox": [ + 511, + 112, + 903, + 203 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "T Cycle: The T cycle training involves transforming an input textual sequence into an image representation and then back into a textual sequence, enforcing consistency between the generated and original text. Specifically, at each training iteration, we begin with an input text sequence $T = [t_l]_{l=1}^L$ . Conditioned on this, the unified model generates pseudo-paired image tokens $I'$ , representing the visual interpretation of the input text. The generated image tokens are then subjected to a random masking operation, denoted as $I_M'$ , where a subset of the tokens is masked. The unified model is then called to reconstruct these masked tokens to form the complete synthesized image $\\tilde{I}$ .", + "bbox": [ + 511, + 203, + 903, + 383 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "In the second half of the cycle, conditioning on image $\\tilde{I}$ , the model generates the corresponding text sequence. The objective of the T cycle is to enforce cycle consistency between the generated text and the original input text $T$ . The T cycle consistency loss is defined as follows:", + "bbox": [ + 511, + 383, + 903, + 460 + ], + "page_idx": 2 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {L} _ {T C} = - \\mathbb {E} _ {T \\in \\mathcal {D} t e x t} \\left[ \\sum_ {l = 1} ^ {L} \\log p \\left(t _ {l} \\mid \\tilde {I}, t _ {0}, \\dots , t _ {l - 1}\\right) \\right], \\tag {1}\n$$\n", + "text_format": "latex", + "bbox": [ + 529, + 484, + 906, + 525 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "where the $\\mathcal{D}$ text is the set of text samples from the target domain.", + "bbox": [ + 511, + 536, + 903, + 564 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "I Cycle: The I cycle training begins with an input image, which is subsequently mapped to a textual representation and then transformed back to an image, enforcing consistency between the generated image and the original image tokens. At each training iteration, we start with input image tokens $I$ . The unified model is used to synthesize pseudopaired text tokens $T'$ , representing the textual description of the image. We then use $T'$ in conjunction with the input image token $I$ to predict the reconstructed text tokens $\\tilde{T}$ .", + "bbox": [ + 511, + 566, + 906, + 702 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "In the second half of the cycle, we pass the masked image tokens $I_M$ and the text $\\tilde{T}$ through the model to regenerate the masked image tokens. The cycle enforces consistency between the reconstructed and the original image tokens. The loss for enforcing I cycle consistency is given by:", + "bbox": [ + 511, + 702, + 903, + 779 + ], + "page_idx": 2 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {L} _ {I C} = - \\mathbb {E} _ {I \\in \\mathcal {D} _ {i m a g e}} \\left[ \\sum_ {\\forall k: m _ {k} = 1} \\log p (i _ {k} | I _ {M}, \\tilde {T}) \\right], \\tag {2}\n$$\n", + "text_format": "latex", + "bbox": [ + 539, + 801, + 906, + 835 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "where the $D$ image is the set of image samples from the target domain, which are unpaired with the text samples.", + "bbox": [ + 511, + 840, + 903, + 869 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "By leveraging these two cycles, our framework forces the model to refine its generative understanding of both image", + "bbox": [ + 511, + 869, + 903, + 901 + ], + "page_idx": 2 + }, + { + "type": "page_number", + "text": "2837", + "bbox": [ + 482, + 944, + 514, + 955 + ], + "page_idx": 2 + }, + { + "type": "image", + "img_path": "images/167d82eea73a700f58885a073203595172afb40ea966537b2946d4ac868e18dd.jpg", + "image_caption": [ + "Figure 2. The overview of T cycle (text-to-image-to-text) of the proposed DoraCycle. The I cycle is similar and is omitted in the figure for brevity." + ], + "image_footnote": [], + "bbox": [ + 98, + 92, + 903, + 281 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "and text representations, ensuring consistency between the input and output while effectively leveraging unpaired data to adapt the unified towards the target domain.", + "bbox": [ + 88, + 347, + 482, + 393 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Efficient Training: In the intermediate steps of both cycles, generating the middle representation (i.e., captions or images) requires multiple forward passes. This is because the generation process involves either predicting the next tokens or the masked tokens multiple times. Backpropagating gradients through all these steps is computationally prohibitive. Thus, we first generate the intermediate results using the model in inference mode as pseudo-paired data, which are then used as the ground truth in the teacher-forcing scheme [54, 55] for the first half of the cycles. In this way, we reduce the number of forward passes to two, i.e. one for generating the middle result and one for the final output, thus making the overall training process more memory efficient.", + "bbox": [ + 88, + 393, + 482, + 604 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Token Differentiability: Since the intermediate outputs in each cycle are discrete tokens, which can not directly propagate gradients, we employ the Gumbel-Softmax [28] to make these token representations differentiable.", + "bbox": [ + 89, + 606, + 483, + 667 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "3.2. Stabilizing Optimization", + "text_level": 1, + "bbox": [ + 89, + 680, + 316, + 696 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Each cycle involves the same unified model twice in the forward pass, which leads to optimization instabilities. To stabilize the training process, we adopt the Exponential Moving Average (EMA) training technique [57]. Specifically, we maintain a shadow version of the model, referred to as the EMA model, which is updated using an exponentially decaying average of the parameters of the main model.", + "bbox": [ + 89, + 703, + 482, + 809 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\theta_ {\\mathrm {E M A}} \\leftarrow \\alpha \\theta_ {\\mathrm {E M A}} + (1 - \\alpha) \\theta_ {\\text {m i n}}, \\tag {3}\n$$\n", + "text_format": "latex", + "bbox": [ + 178, + 829, + 482, + 844 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "where $\\alpha$ is a decay factor (set to 0.999) that controls the update rate, and $\\theta_{\\mathrm{main}}$ represents the parameters of the main model.", + "bbox": [ + 89, + 854, + 483, + 898 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "In each training step, the EMA version of the model is used to generate the intermediate representation tokens (e.g., pseudo image or text tokens) which serve as pseudo ground truth during training. By using these stable targets from the slower-evolving EMA model, we can mitigate the risks of optimization instability. The main model is thus able to learn from more consistent and reliable intermediate targets, rather than being affected by fluctuations during the early stages of training.", + "bbox": [ + 511, + 347, + 906, + 484 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "3.3. Balancing Two Cycles", + "text_level": 1, + "bbox": [ + 511, + 496, + 718, + 512 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "We observe that the T cycle tends to converge faster than the I cycle, primarily because textual data is inherently one-dimensional and simpler to learn compared to images. This imbalance in optimization leads to a kind of collapse of the model, where it tends to generate irrelevant but self-consistent captions for images, ultimately degrading the image-text alignment capability.", + "bbox": [ + 511, + 518, + 905, + 626 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "To address this problem, we make the gradients of the T cycle orthogonal to those of the I cycle, thus preventing interference. This is achieved by modifying the gradients using gradient surgery [72]. Let $g_{T}$ and $g_{I}$ represent the gradients of the T cycle and the I cycle, respectively. We project $g_{T}$ onto the orthogonal complement of $g_{I}$ to obtain the modified gradient $g_{T}^{\\prime}$ , which is defined as:", + "bbox": [ + 511, + 627, + 905, + 734 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\ng _ {T} ^ {\\prime} = g _ {T} - \\frac {g _ {T} \\cdot g _ {I}}{g _ {I} \\cdot g _ {I}} g _ {I}, \\tag {4}\n$$\n", + "text_format": "latex", + "bbox": [ + 632, + 747, + 906, + 777 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "where $g_{T} \\cdot g_{I}$ denotes the dot product between the gradients of the T and I cycle.", + "bbox": [ + 511, + 785, + 905, + 814 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Additionally, we reweight the losses to further balance the learning between the I and T cycles. The final loss function is as,", + "bbox": [ + 511, + 816, + 905, + 859 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {L} = \\mathcal {L} _ {I C} + \\beta \\mathcal {L} _ {T C}, \\tag {5}\n$$\n", + "text_format": "latex", + "bbox": [ + 642, + 864, + 903, + 880 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "where the $\\beta$ is the weight of the T cycle loss.", + "bbox": [ + 511, + 885, + 812, + 901 + ], + "page_idx": 3 + }, + { + "type": "page_number", + "text": "2838", + "bbox": [ + 482, + 944, + 514, + 955 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "4. Experiments", + "text_level": 1, + "bbox": [ + 89, + 89, + 223, + 107 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "4.1. Implementation Details", + "text_level": 1, + "bbox": [ + 89, + 114, + 307, + 131 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "To the best of our knowledge, Show-o [67] is currently the only fully open-source unified generative model with complete pre-trained weights and training code, including both its understanding and generation capabilities. Therefore, we base DoraCycle on Show-o and conduct experiments accordingly. The base model is a unified transformer model that performs understanding and image generation by predicting discrete textual and visual tokens. We insert trainable low-rank adaptations (LoRA) [25] modules into the Q projection and V projection of the attention layers from layers 7 to 24. The LoRA rank is set to 32. The $\\beta$ is set to 0.1 to balance the optimization of two cycles.", + "bbox": [ + 89, + 138, + 483, + 319 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "The training of DoraCycle is performed on 8 NVIDIA H100 GPUs with mixed precision enabled for memory efficiency. We set the batch size to 32, with each cycle taking half of the batch when both cycles are being optimized simultaneously. The learning rate is set to $1e^{-4}$ with a cosine annealing schedule. The optimizer is AdamW with weight decay of $1e^{-2}$ . Additionally, EMA is employed to stabilize the training process, as described in Section 3.2.", + "bbox": [ + 89, + 320, + 483, + 441 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "4.2. Domain-Oriented Adaptations", + "text_level": 1, + "bbox": [ + 89, + 454, + 361, + 469 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Unpaired Training: For tasks that do not require strongly related paired knowledge, our DoraCycle can fully learn the target domain using unpaired data. For example, to learn the cyberpunk style, we collected 300 cyberpunk-style images as input for the I cycle, and used the text data from the base model pre-training dataset [4] for the T cycle, with the keyword \"cyberpunk style\" automatically injected into text, prompting the model about the target style we want.", + "bbox": [ + 89, + 476, + 482, + 598 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "The experimental results are shown in Fig. 3. Given the same text prompt to generate cyberpunk-style images, Fig. 3 (a) shows the images generated by the base model without additional training. It can be observed that the base model adds some cyberpunk elements, such as neon lights, but the overall atmosphere does not align well with the desired style. Fig. 3 (d) shows the images generated by the adapted model trained with DoraCycle, which aligns well with the target style. Traditional text-to-image customization or adaptation methods, such as DreamBooth [48], rely on paired data for training. Therefore, we simulate user-created paired data by annotating the collected images with captions, and split them into two groups. One group contained only 10 paired examples, which is an acceptable workload for users, while the other group contained captions for all 300 images, which would be labor-intensive and impractical for users. The images generated by the model that trained on 10 paired examples are shown in Fig. 3 (b). It struggled to produce good stylized images, likely because the combination of indoor bookshelves with the cyberpunk", + "bbox": [ + 89, + 598, + 483, + 900 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": " The image presents a scene of a room with a brick wall and a large bookshelf filled with various books and electronics. The bookshelf, which is black, stands against the brick wall and spans almost the entire height of the wall. It is divided into three shelves, each holding different items. On the top shelf, there are several books neatly arranged. In front of the bookshelf, there's a black leather couch positioned on one side of the room. A potted plant can be seen as well.", + "bbox": [ + 517, + 89, + 900, + 118 + ], + "page_idx": 4 + }, + { + "type": "image", + "img_path": "images/2eca2e870a5dda13242b0e49d00f3556dc8424685c88ef5f06fa73192142eba8.jpg", + "image_caption": [ + "(a) Base model (w/o training)" + ], + "image_footnote": [], + "bbox": [ + 517, + 132, + 612, + 207 + ], + "page_idx": 4 + }, + { + "type": "image", + "img_path": "images/6ae94d2a6438dbf1d6ddfcbfd6156f15ffed35e7eb9eece5f0896d9cf26e1e94.jpg", + "image_caption": [ + "Image-to-Text" + ], + "image_footnote": [], + "bbox": [ + 614, + 132, + 705, + 207 + ], + "page_idx": 4 + }, + { + "type": "image", + "img_path": "images/870c95c63c517b35aa72bdf34ec2fed87f0ce48eac5e8556631ef648fa06113f.jpg", + "image_caption": [ + "(c) Paired trained (300 pairs)" + ], + "image_footnote": [], + "bbox": [ + 707, + 132, + 805, + 207 + ], + "page_idx": 4 + }, + { + "type": "image", + "img_path": "images/23baec640737327651913a6e88984467ed8188cca2fc347ddeac4dac0e94a5a9.jpg", + "image_caption": [ + "(d) Unpaired trained (DoraCycle)" + ], + "image_footnote": [], + "bbox": [ + 807, + 132, + 903, + 207 + ], + "page_idx": 4 + }, + { + "type": "image", + "img_path": "images/f89508f9d5cf179e58c291eb433cd782af6c8cd0cdaebd46e1a82c880c87961e.jpg", + "image_caption": [ + "(b) Paired trained (10 pairs)" + ], + "image_footnote": [], + "bbox": [ + 516, + 243, + 629, + 330 + ], + "page_idx": 4 + }, + { + "type": "image", + "img_path": "images/f9169b06b29a9f339c4934169bdad9889bb1b4442ce66eac6e94d26ef9b7f909.jpg", + "image_caption": [ + "Image-to-Text" + ], + "image_footnote": [], + "bbox": [ + 593, + 332, + 633, + 362 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": " The image captures a scene on a city street. Dominating the foreground are two mannequins, each adorned in full Batman and Iron Man costumes. The Mannequin on the left is dressed in a black Batman costume, complete with a cape and mask, exuding an air of mystery and power. On the right, the other Mannequin is clad in a red Iron Man costume, featuring a helmet with a faceplate and red gloves, embodying the iconic superhero's persona. These mannequins stand on a gray sidewalk, their vibrant costumes contrasting against the monotone backdrop. Behind them, the city continues its usual hustle and bustle. Cars are parked along the street, their metallic bodies gleaming under the city lights. Buildings rise in the background, their windows reflecting the life of the city.", + "bbox": [ + 635, + 243, + 779, + 364 + ], + "page_idx": 4 + }, + { + "type": "image", + "img_path": "images/65ea5a22b8d591bb2d788c1ec0cb374a0747b1588d5d4cefef16b99599175495.jpg", + "image_caption": [ + "Figure 3. Domain-oriented adaptation with different training setups. (a) Image generated by the base model without training for adoption. (b) Image generated by the model trained with 10 paired image-text samples. (c) Image generated by the model trained with 300 paired image-text samples. (d) Image generated by the model trained by DoraCycle on only unpaired data. (e) Image-to-Text-to-Image translation performed by the adapted model trained by DoraCycle." + ], + "image_footnote": [], + "bbox": [ + 790, + 243, + 903, + 330 + ], + "page_idx": 4 + }, + { + "type": "image", + "img_path": "images/1bd13033b221a6a82e57404b8b11a3178f44a8db4074f9bd7fad93b6f784cc5b.jpg", + "image_caption": [ + "(e) Image-to-Text-to-Image translation by the adapted unified model." + ], + "image_footnote": [], + "bbox": [ + 790, + 332, + 890, + 362 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "style is too novel for the model to generalize well from limited paired data. The images generated by the model trained on 300 paired examples are shown in Fig. 3 (c), which have better outputs. In contrast, the model trained using DoraCycle does not require manual captioning, significantly reducing the workload for users.", + "bbox": [ + 511, + 535, + 905, + 626 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Fig. 3 (e) illustrates the adapted model trained by DoraCyle maintains semantic consistency through image-to-text-to-image translation. The input image is transformed into a textual description and then reconstructed into an image. The result shows that the adapted model successfully captures and retains the key visual components in the original image throughout the multimodal cycle. Notably, the identity of the characters and the details of the environment are all preserved, indicating effective bidirectional understanding and generation capabilities in the target domain. Furthermore, the newly generated image incorporates styles learned from the target domain, demonstrating the generalization of the learned knowledge to images in the wild.", + "bbox": [ + 511, + 627, + 905, + 823 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Learning Paired Knowledge: For tasks that require learning some paired knowledge, such as associating an identity name with its visual appearance, DoraCycle can incorporate a small amount of paired data to learn such associations while leveraging a large amount of unpaired", + "bbox": [ + 511, + 825, + 905, + 900 + ], + "page_idx": 4 + }, + { + "type": "page_number", + "text": "2839", + "bbox": [ + 482, + 944, + 514, + 955 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Domain 1: Black Myth Wukong", + "text_level": 1, + "bbox": [ + 406, + 92, + 598, + 106 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Text-to-Image", + "text_level": 1, + "bbox": [ + 106, + 127, + 122, + 224 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/3c4f64bbba733eb1cfc9c7f0e24405ab2a625f1866662e47be5685bffa574996.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 127, + 108, + 251, + 205 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Destined One, wears an intricately detailed golden armor with elaborate ornaments and engravings. A golden crown-like adornment sits atop his head. In the background, distant mountains and a serene sky can be seen.", + "bbox": [ + 127, + 209, + 248, + 257 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/92f2055398260c2f3c80e6ade333da387b1ef74b3ef192a8826c47608c0e5d7d.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 254, + 108, + 380, + 205 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Lingxuizi's glowing red eyes and bared fangs indicate its aggressive nature. It crouches low to the ground, its large hand-like paw spread over ancient stone paving surrounded by patches of grass.", + "bbox": [ + 256, + 208, + 375, + 251 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/f57acc46503ff6b286c6582823f92d79962e1e9a3896ba3e16f6faa381106488.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 383, + 108, + 509, + 205 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Keeper grips a gnarled wooden staff for support as he walks through a wooded area. His posture is stooped, and his facial features are mostly obscured by his thick beard. The scene is set in a forest, with a textured ground of leaves and dirt.", + "bbox": [ + 385, + 208, + 504, + 258 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/e56ab250c53c79fd423d31f583501da649fb5fbdfd242f88529ab40aac4a9a71.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 511, + 108, + 637, + 205 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Zhu Bajie stands confidently in the snow, gripping his staff with a sly smile on his boat-like face, beads hanging around his neck. Beside him lies the majestic, wounded form of Kang-Jin Loong, a silver-scaled dragon with intricate, ethereal horns and whiskers.", + "bbox": [ + 514, + 208, + 632, + 263 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/ff73fe5e4e66bf09fe7a83881410a8e0ffae1b43695f0bdf0c062549c9a3084f.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 638, + 108, + 764, + 205 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Tiger Vanguard charges through the blood-red waters towards Destined One. The massive stone statue behind him looms ominously, adding to the intensity of the scene. Destined One stands firm, his staff held at the ready, his gaze unwavering as he prepares for the oncoming assault. The water ripples around them, reflecting the dark hues of the battle setting.", + "bbox": [ + 642, + 208, + 761, + 277 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/d8f24403546657cbf631464bf9a63f5cc75c38ac27f953351b65d80eaa34ad1e.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 767, + 108, + 893, + 205 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Destined One stands tall, his gaze cast downward, his weathered yellow robe flowing with subtle wear, reflecting the trials he has endured. Beside him, Zhu Majie looks up with a playful expression, his boar-like face filled with curiosity and mischief. The two figures share a quiet moment, framed against an open, clear sky dotted with drifting embers.", + "bbox": [ + 769, + 208, + 888, + 277 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Image-to-Text", + "text_level": 1, + "bbox": [ + 106, + 308, + 122, + 364 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/c8cfb3108961accebf9fcff37647a316572990ec319ee49156e18605f7acdb0c.jpg", + "image_caption": [ + "(a)" + ], + "image_footnote": [], + "bbox": [ + 127, + 287, + 251, + 385 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Destined One stands tall, wielding a massive golden stick that dominates the frame. The individual exudes power, strength, and determination, clad in ornate armor that gleams with gold. A hat adds a touch of regality. The background is dramatic, with shadows cast by trees hinting at a mysterious atmosphere. The grass beneath the figure is lush and green, contrasting with the ruggedness of the ground.", + "bbox": [ + 127, + 388, + 251, + 465 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/704e383d3f5d55ae6966d12ab4faef5a2416475478a563593dfee90f7ebda39f.jpg", + "image_caption": [ + "(b)" + ], + "image_footnote": [], + "bbox": [ + 254, + 287, + 380, + 386 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Keeper, an elderly, hunched individual with a bald, heavily wrinkled head covered in spots. Keeper has a long, unkempt beard and is dressed in a simple, tattered blue robe. He grips a gnarled wooden staff for support as he walks through a forest fire. His posture is stooped, and his facial features are mostly obscured by his thick beard.", + "bbox": [ + 258, + 388, + 380, + 450 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/aadf39eb053a5ed03c1e774206f6facdabcdbf7ddb9b1f67cef4f56ed0c8aa91.jpg", + "image_caption": [ + "(c)" + ], + "image_footnote": [], + "bbox": [ + 382, + 287, + 509, + 386 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "The image depicts a serene, snowy landscape with a large, ancient-looking wooden bridge spanning across a frozen river. The bridge has a traditional, Asian design, with intricate carvings and a slatted roof. Bare trees with frost-covered branches surround the scene, casting a calm, wintery atmosphere. Snowflakes gently fall, enhancing the tranquil, mystical ambiance", + "bbox": [ + 383, + 388, + 504, + 464 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/e4a5e9c242fca82aa9cc3e2dd5692905c86c225cebe3c161578f9c0933a0630d.jpg", + "image_caption": [ + "(d)" + ], + "image_footnote": [], + "bbox": [ + 511, + 287, + 635, + 386 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Lingxuzi is depicted as a large, menacing fox-like creature with thick, flowing white fur. Its red eyes glow intensely, giving an eerie, supernatural appearance. The creature is in a crouching stance, with its body slightly arched and its tails fanned out behind it, showing both power and readiness to strike. The dark background adds to the ominous atmosphere, contrasting with the brightness of its fur.", + "bbox": [ + 511, + 388, + 635, + 464 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/b7f74ba3fcf5b04cea895ebbb2b4dcbe59456f4a51fe61891b691dd6768f720e.jpg", + "image_caption": [ + "(e)" + ], + "image_footnote": [], + "bbox": [ + 638, + 287, + 764, + 386 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Zhu Bajie stands menacingly in a murky, mist-covered swamp, his humanoid body combined with the head of a fierce, snarling boar. Her red, glowing eyes pierce through the fog, radiating a dark energy. Dressed in a tattered green robe, he wears a string of large beads around his neck, adding a sense of mysticism. His left hand clutches a spiked weapon, while shadows and wisps of smoke swirl around him, amplifying his sinister presence.", + "bbox": [ + 640, + 388, + 761, + 477 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/f9ecdabaabd601729254048c6fb77491b9dafde4b0029f5982d36e14a2494d62.jpg", + "image_caption": [ + "(f)" + ], + "image_footnote": [], + "bbox": [ + 767, + 287, + 893, + 386 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "A female character with a unique appearance, having a combination of human and demonic features. She has a short, black, straight hairstyle and a pair of horns protruding from her head. Her skin is pale, and she has a few dots on her forehead, giving her a mystical look. She is wearing a black, low-cut top adorned with white flowers, and her right arm is covered in red fabric. She has a focused and intense expression on her face, as if she is concentrating on something important. (D)", + "bbox": [ + 769, + 388, + 890, + 482 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Text-to-Image", + "text_level": 1, + "bbox": [ + 106, + 531, + 122, + 585 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/c3a1c8d8eb39e13c9dd4470d110fcdaf3399a6aaabea6d7afd73fe7258e8663d.jpg", + "image_caption": [ + "(g)" + ], + "image_footnote": [], + "bbox": [ + 127, + 511, + 254, + 590 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/6f2633121fd8eebbf802b5716e36b442c498538e94e037ccac324b5c8c7e3f07.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 127, + 592, + 246, + 607 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Nobita is sitting under a tree, trying to finish his math homework. He looks stressed, with one hand on his chin and scratching his head with the other.", + "bbox": [ + 127, + 612, + 246, + 640 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/ab5c77f33e601cc1c923973cc010f58c95788b20ab83699c9f4b4cce12df1178.jpg", + "image_caption": [ + "(h)" + ], + "image_footnote": [], + "bbox": [ + 256, + 511, + 382, + 607 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Suneo and Nobita are at a fishing pond, trying to catch a fish. Nobita, in his bright yellow fishing hat and a sharp outfit, confidently casts his line into the water. Suneo watches Nobita's struggle with a smile, offering some tips to help him out. The two boys share a laugh as they wait for a bite.", + "bbox": [ + 256, + 611, + 375, + 672 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/f13264c30b6a0da6e01d4025bac2335db2d2f8274eb375b9cf1362accf377f43.jpg", + "image_caption": [ + "(i)", + "Domain 2: Doraemon" + ], + "image_footnote": [], + "bbox": [ + 385, + 511, + 509, + 607 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Doraemon is cooking in the kitchen, while holding a bowl. He looks excited, flipping a pancake into the air. Nobita is standing beside him, smiling gently, watching Doraemon.", + "bbox": [ + 385, + 611, + 503, + 646 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/b8241675d95fba7f82713398267b5bbdeb136894452978f0de260cebadd351d8.jpg", + "image_caption": [ + "(j)" + ], + "image_footnote": [], + "bbox": [ + 511, + 511, + 637, + 607 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Shizuka is sitting by a piano, composing a new tune. She looks thoughtful, her fingers delicately pressing the keys as she tries different melodies. A notebook rests on the piano stand, filled with her notes and musical ideas, and she appears content in the process.", + "bbox": [ + 513, + 611, + 632, + 666 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/c2a30f70ce7eab0d82a13395f7cdd31e3e5739853fc70813e09bd9fb30c7d45e.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 638, + 511, + 763, + 608 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Shizuka and Nobita are making flower crowns at the park. Shizuka, in her soft pink dress, weaves flowers together. Nobita, in his casual shorts, tries his best to copy her.", + "bbox": [ + 640, + 611, + 759, + 646 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/ce3544a651f6016e9caeb2a81d4c904b4e05405714243cb8dae911ef4a8a14a7.jpg", + "image_caption": [ + "(1)" + ], + "image_footnote": [], + "bbox": [ + 767, + 511, + 890, + 607 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Doraemon is standing in front of the white cat with an excited expression. He is holding a bouquet of flowers, in front of the cat. The white cat, sitting on a small cushion, is looking at Doraemon with a curious or intrigued expression. The setting appears to be indoors, with a pink carpet, a scratching post in the background, and a framed picture on the wall.", + "bbox": [ + 769, + 611, + 888, + 679 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Image-to-Text", + "text_level": 1, + "bbox": [ + 106, + 712, + 122, + 766 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/66cbdf51c43a6bb3ef2b83bf353c163add826a4f6fc7f26e0d28fc812a5da9f2.jpg", + "image_caption": [ + "(s)", + "Figure 4. Image-to-text and text-to-image generation by the unified models that adapted for two domains. The special tokens are omitted." + ], + "image_footnote": [], + "bbox": [ + 127, + 680, + 254, + 789 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Doraemon is standing on the shore, looking out at the sunset. He has a calm expression, and the sun is setting over the water, casting a warm glow. The sky is filled with clouds, and there are silhouettes of buildings in the distance.", + "bbox": [ + 127, + 792, + 246, + 840 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/b3928faa52e349ecd559c44612e7824d6c7c710385ea9e3880a29c332b0ffeb3.jpg", + "image_caption": [ + "(n)", + "(t)" + ], + "image_footnote": [], + "bbox": [ + 256, + 691, + 382, + 787 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Doraemon is standing in the grass, looking surprised as he sees a group of cats. The cats are gathered around him, with one of them appearing to be a cat with a surprised expression. The background shows a fence and some bushes.", + "bbox": [ + 256, + 792, + 375, + 840 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/849259d7fbc5ab22955acb39c7b11b335ce7135ac7e17dbbcbe6776882e7b8da.jpg", + "image_caption": [ + "(u)" + ], + "image_footnote": [], + "bbox": [ + 385, + 691, + 509, + 787 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Doradamon is flying through the air, looking down at the city below. Nobita is also flying through the air, looking down at the city.", + "bbox": [ + 385, + 792, + 503, + 820 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/5fea7abcdc30d6b74bcfcd1eb4ba39cf92839d1563150bf49669380f3b02be92.jpg", + "image_caption": [ + "(p)", + "(v)" + ], + "image_footnote": [], + "bbox": [ + 511, + 691, + 637, + 787 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "The image depicts a two-story house with a large garden in front of it. The house has a tan exterior and is surrounded by a variety of plants, including palm trees and bushes. The garden is well-maintained. The house features a balcony on the second floor, adding to its charm.", + "bbox": [ + 513, + 792, + 632, + 847 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/8245120e916c34348b7d1725f64102d1ac7f99d8a264daa5af3c8f223ad0377f.jpg", + "image_caption": [ + "(q)", + "(w)" + ], + "image_footnote": [], + "bbox": [ + 638, + 691, + 764, + 787 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Nobita is standing in the doorway, holding a plate with a doughnut on it. He looks excited, and the doughnut is on a white plate. The door is open, and the room is visible in the background.", + "bbox": [ + 640, + 792, + 759, + 834 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/67ecfcbba9ae96f1be693b1efe68514ba38a55c3d9198ab7f8e1b1c19a3d13c6.jpg", + "image_caption": [ + "(r)", + "(x)" + ], + "image_footnote": [], + "bbox": [ + 767, + 691, + 890, + 787 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "White cat and black cat are sitting on the grass, looking at each other with affection, with a blue sky in the background, and a few clouds scattered across the sky.", + "bbox": [ + 769, + 792, + 888, + 827 + ], + "page_idx": 5 + }, + { + "type": "page_number", + "text": "2840", + "bbox": [ + 482, + 944, + 514, + 955 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/47a1a143534d9ec5df110d474acfeb5385995373562f237710a5036fc5032de7.jpg", + "image_caption": [ + "Figure 5. Effect of special tokens on character learning. (a) Base model without training. (b) Model trained without using special tokens, showing attribute confusion among characters. (c) Model trained with special tokens, improving character attribute alignment and reducing confusion." + ], + "image_footnote": [], + "bbox": [ + 94, + 88, + 480, + 324 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "data to comprehensively learn general aspects of the target domain. Specifically, in each batch, for data with paired ground truth, we compute the token prediction loss and also include it in the cycle, use ground truth as the pseudo middle generations, and compute the cycle loss. For unpaired data, we compute the unpaired cycle loss.", + "bbox": [ + 89, + 439, + 483, + 530 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "For example, when adapting the model to Domain 1: Black Myth Wukong and Domain 2: Doraemon, we annotate 1-3 images per unique identity with captions that specify the name of the identity. For each domain, we collect $2\\mathrm{k}$ images, which are mostly sampled from online videos, and independently collect text descriptions, which are further expanded to $1\\mathrm{k}$ by ChatGPT [39]. The final adapted model trained with DoraCycle demonstrates strong performance in both text-to-image generation and image-to-text generation, as shown in Fig. 4.", + "bbox": [ + 89, + 532, + 483, + 683 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "In terms of text-to-image results, the model trained with DoraCycle effectively generated images that aligned well with the target domains. In domain 1 (Black Myth Wukong), the generated images accurately depicted domain-specific visual elements, such as the intricate details of character appearances and the overall fantasy-like atmosphere. This indicates that the model successfully learned to generalize the visual features from text prompts to realistic images within the target domain. Similarly, in domain 2 (Doraemon), the generated images preserve the iconic cartoonish aesthetics and capture key visual details of the characters and settings, demonstrating effective domain adaptation.", + "bbox": [ + 89, + 686, + 482, + 881 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "For the image-to-text task, the model performs well in", + "bbox": [ + 109, + 885, + 482, + 900 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "generating contextually accurate captions. In domain 1, the generated captions provide rich descriptions of the characters, their attributes, and the context, effectively mirroring the visual elements present in the input images. In domain 2, the captions correctly describe the characters, their actions, and their environments concisely, maintaining consistency with the visual style. The ability of the model to generate accurate descriptions highlights its robust understanding of the visual components of the domain.", + "bbox": [ + 511, + 90, + 903, + 226 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Additionally, an interesting phenomenon can be observed in how the model handles the visual elements that are not annotated with paired data. For instance, in Fig. 4 (w), the dorayaki (a type of sweet bean-filled pancake) was described by the model as a \"doughnut\". This may be due to the fact that the anime-style representation of the dorayaki is novel, and neither the base model nor the unpaired training provided specific textual-visual pairing knowledge about it. On the other hand, in the example shown in Fig. 4 (x), we annotate the white cat as a character with paired textual and visual data, using a special token for its name: \" white cat black cat \" during the caption generation. This suggests that the model autonomously categorized the black cat as a character when learning the target domain, indicating that it may have attempted to generalize learned knowledge from one type of entity to similar ones.", + "bbox": [ + 511, + 228, + 906, + 513 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Enhanced Learning with Special Tokens: As shown in Fig. 5, we experimentally find that the model often confused multiple novel concepts in the target domain. Fig. 5 (a) shows the image generated by the base model without training taking in the name of characters. Fig. 5 (b) shows the characters generated by the trained model. During training, the names of characters are directly included in the text without special treatment, leading to attribute confusion between characters. The varying lengths of the tokenized character names also make learning difficult. To solve this problem, we introduce a simple yet efficient solution: adding special tokens around character names. We introduced start of character $(\\langle \\text{soc}\\rangle)$ and end of character $(\\langle \\text{soc}\\rangle)$ tokens to enclose character names, which significantly enhance the learning of novel concepts. As shown in Fig. 5 (c), involving special tokens improves the alignments between characters and their names.", + "bbox": [ + 511, + 516, + 908, + 772 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "4.3. Comparisons", + "text_level": 1, + "bbox": [ + 511, + 786, + 653, + 801 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "In this section, we use the Storyboard20K [66] dataset to conduct the quantitative comparison experiments. The storyboards originating from the same data source are grouped to form a domain, consisting of images and descriptive text. The data are used under three different settings, i.e. totally unpaired, only paired, and paired plus unpaired data,", + "bbox": [ + 511, + 809, + 906, + 902 + ], + "page_idx": 6 + }, + { + "type": "page_number", + "text": "2841", + "bbox": [ + 482, + 944, + 513, + 955 + ], + "page_idx": 6 + }, + { + "type": "table", + "img_path": "images/46b62c72ec46647e2c1bcad4b076d087a4640742a46022a03e236a52ee9a9773.jpg", + "table_caption": [ + "Table 1. Comparison of different training methods under various data settings. The best value is highlighted in blue , and the second-best value is highlighted in green . “P” indicates paired data, and “U” indicates unpaired data." + ], + "table_footnote": [], + "table_body": "
T cycleI cycleT DataI DataFID-1K ↓CIDEr ↑Human Eval
T2I Align ↑I2T Align ↑
DreamBooth [48]--10% P10% P33.2232.743.251.83
--100% P100% P24.9341.554.133.96
DoraCycleX✓X100% U28.9330.543.381.62
✓X100% UX36.6335.703.262.17
✓✓100% U100% U27.4438.173.843.42
✓✓10% P + 90% U10% P + 90% U25.3740.904.123.81
ITIT [34]✓✓10% P + 90% U10% P + 90% U27.5038.623.853.52
", + "bbox": [ + 94, + 132, + 906, + 292 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "as shown in Table 1.", + "bbox": [ + 89, + 316, + 230, + 330 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "The compared methods include DreamBooth [48] and ITIT [34]. We implement DreamBooth as a paired-training baseline by applying LoRA fine-tuning on the unified model. The original design of ITIT is different, in which the image and text decoders are separate models, and its code has not been released. We adjusted and re-implemented it to be suitable for our unified model architecture.", + "bbox": [ + 89, + 334, + 482, + 439 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "We use both automatic and human evaluations to compare the performance of different methods. For automatic evaluation, we use FID to measure the distribution differences between the generated images and the target domain images [22], and CIDEr to compute the error between the generated text and the ground truth [59]. For human evaluation, we create 100 questions for the generated results of models, each rated by three different human raters. The raters are asked to evaluate the alignment between the image and text on a scale from 1 to 5, where 1 indicates no relevance and 5 indicates complete alignment.", + "bbox": [ + 88, + 441, + 482, + 609 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "The experimental results in Table 1 demonstrate that the proposed DoraCycle performs competitively under several data settings. Specifically, when using a combination of paired and unpaired data, DoraCycle outperforms ITIT. Compared to DreamBooth, which heavily relies on paired data, DoraCycle outperforms it when using the same scale of paired data, i.e. $10\\%$ paired data, indicating the benefits brought by $90\\%$ unpaired data. While Dreambooth with $100\\%$ achieves the best evaluation scores, the scores of the DoraCycle with $10\\%$ paired and $90\\%$ unpaired data are comparable with them.", + "bbox": [ + 88, + 611, + 482, + 777 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Table 1 also shows the difference in the performance of DoraCyle under different cycle settings. It is shown that without the T cycle and with only the I cycle, the captioning ability of the adapted model degrades more significantly. In contrast, if only the T cycle is used and without the I cycle, the FID score increases substantially, indicating that the generated image distribution mismatches with the target distribution.", + "bbox": [ + 88, + 779, + 482, + 900 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "4.4. Ablation Studies", + "text_level": 1, + "bbox": [ + 513, + 316, + 676, + 330 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Table 2 shows that removing key components from DoraCycle significantly impacts performance. Without EMA, the FID score increases from 25.37 to 27.19, indicating lower image quality", + "bbox": [ + 511, + 345, + 668, + 482 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Table 2. Ablation Studies. EMA refers to the exponential moving average. GS refers to gradient surgery.", + "bbox": [ + 696, + 345, + 905, + 402 + ], + "page_idx": 7 + }, + { + "type": "table", + "img_path": "images/6af9013b4ab42cc8621e701c3d9aacb98875a37a45e7a8ac434a20ebdf2643f1.jpg", + "table_caption": [], + "table_footnote": [], + "table_body": "
FID-1K ↓CIDEr ↑
w/o EMA27.1938.85
w/o GS25.5439.98
DoraCycle25.3740.90
", + "bbox": [ + 700, + 412, + 903, + 474 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "due to less stable training. Removing Gradient Surgery (GS) will reduce the CIDEr score and increase the FID, indicating a worse performance. This demonstrates the importance of mitigating the interference between the optimization directions of two cycles. The complete DoraCycle framework, with both EMA and GS, has the best performance across all metrics, demonstrating the importance of these components in achieving better optimization.", + "bbox": [ + 511, + 483, + 906, + 618 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "5. Conclusion", + "text_level": 1, + "bbox": [ + 513, + 655, + 633, + 671 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "We propose the DoraCycle to adapt the unified generative model to target domains within multimodal cycles. By leveraging both image-to-text-to-image and text-to-image-to-text cycles, DoraCycle changes the learning objectives into the same modality, allowing for effective optimization using unpaired data. Our experiments show that DoraCycle can adapt the unified model to target domains using only unpaired data, or involving a small amount of paired data when necessary to learn specific concepts. Experimental results demonstrate that DoraCycle achieves advanced or comparable performance across various settings. Leveraging unpaired data broadens the application potential of DoraCycle, making it ideally suited for domain adaptation tasks where paired data is scarce or challenging to collect.", + "bbox": [ + 509, + 689, + 906, + 902 + ], + "page_idx": 7 + }, + { + "type": "page_number", + "text": "2842", + "bbox": [ + 482, + 944, + 514, + 955 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "6. Acknowledgments", + "text_level": 1, + "bbox": [ + 91, + 90, + 269, + 107 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "This research is supported by the National Research Foundation, Singapore under its AI Singapore Programme (AISG Award No: AISG3-RP-2022-030).", + "bbox": [ + 89, + 114, + 483, + 160 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "References", + "text_level": 1, + "bbox": [ + 91, + 176, + 187, + 193 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[1] Armen Aghajanyan, Bernie Huang, Candace Ross, Vladimir Karpukhin, Hu Xu, Naman Goyal, Dmytro Okhonko, Mandar Joshi, Gargi Ghosh, Mike Lewis, et al. Cm3: A causal masked multimodal model of the internet. arXiv preprint arXiv:2201.07520, 2022. 2", + "[2] Emanuele Aiello, LILI YU, Yixin Nie, Armen Aghajanyan, and Barlas Oguz. Jointly training large autoregressive multimodal models. In ICLR, 2024. 2", + "[3] Yogesh Balaji, Seungjun Nah, Xun Huang, Arash Vahdat, Jiaming Song, Karsten Kreis, Miika Aittala, Timo Aila, Samuli Laine, Bryan Catanzaro, et al. ediffi: Text-to-image diffusion models with an ensemble of expert denoisers. arXiv preprint arXiv:2211.01324, 2022. 2", + "[4] Soravit Changpinyo, Piyush Sharma, Nan Ding, and Radu Soricut. Conceptual 12m: Pushing web-scale image-text pretraining to recognize long-tail visual concepts. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 3558-3568, 2021. 5", + "[5] Junsong Chen, Chongjian Ge, Enze Xie, Yue Wu, Lewei Yao, Xiaozhe Ren, Zhongdao Wang, Ping Luo, Huchuan Lu, and Zhenguo Li. Pixart $\\backslash$ sigma: Weak-to-strong training of diffusion transformer for 4k text-to-image generation. arXiv preprint arXiv:2403.04692, 2024. 2", + "[6] Junsong Chen, Jincheng Yu, Chongjian Ge, Lewei Yao, Enze Xie, Zhongdao Wang, James T. Kwok, Ping Luo, Huchuan Lu, and Zhenguo Li. Pixart-α: Fast training of diffusion transformer for photorealistic text-to-image synthesis. In ICLR. OpenReview.net, 2024. 2", + "[7] Li Chen, Mengyi Zhao, Yiheng Liu, Mingxu Ding, Yangyang Song, Shizun Wang, Xu Wang, Hao Yang, Jing Liu, Kang Du, et al. Photovsere: Tuning-free image customization with text-to-image diffusion models. arXiv preprint arXiv:2309.05793, 2023. 2", + "[8] Mark Chen, Alec Radford, Rewon Child, Jeffrey Wu, Heewoo Jun, David Luan, and Ilya Sutskever. Generative pretraining from pixels. In ICML, pages 1691-1703, 2020. 2", + "[9] Xi Chen, Lianghua Huang, Yu Liu, Yujun Shen, Deli Zhao, and Hengshuang Zhao. Anydoor: Zero-shot object-level image customization. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 6593-6602, 2024. 2", + "[10] Yunjey Choi, Minje Choi, Munyoung Kim, Jung-Woo Ha, Sunghun Kim, and Jaegul Choo. Stargan: Unified generative adversarial networks for multi-domain image-to-image translation. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition (CVPR), 2018. 3", + "[11] Wenliang Dai, Junnan Li, Dongxu Li, Anthony Meng Huat Tiong, Junqi Zhao, Weisheng Wang, Boyang Li, Pascale Fung, and Steven Hoi. Instructlip: Towards general-" + ], + "bbox": [ + 93, + 202, + 483, + 900 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "purpose vision-language models with instruction tuning, 2023. 2", + "[12] Xiaoliang Dai, Ji Hou, Chih-Yao Ma, Sam Tsai, Jialiang Wang, Rui Wang, Peizhao Zhang, Simon Vandenhende, Xiaofang Wang, Abhimanyu Dubey, et al. Emu: Enhancing image generation models using photogenic needles in a haystack. arXiv preprint arXiv:2309.15807, 2023. 2", + "[13] Prafulla Dhariwal and Alexander Nichol. Diffusion models beat gans on image synthesis. Advances in Neural Information Processing Systems, 34:8780-8794, 2021. 2", + "[14] Runpei Dong, Chunrui Han, Yuang Peng, Zekun Qi, Zheng Ge, Jinrong Yang, Liang Zhao, Jianjian Sun, Hongyu Zhou, Haoran Wei, Xiangwen Kong, Xiangyu Zhang, Kaisheng Ma, and Li Yi. DreamLLM: Synergistic multimodal comprehension and creation. In ICLR, 2024. 1, 2", + "[15] Debidatta Dwibedi, Yusuf Aytar, Jonathan Tompson, Pierre Sermanet, and Andrew Zisserman. Temporal cycle-consistency learning. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 1801-1810, 2019. 3", + "[16] Patrick Esser, Sumith Kulal, Andreas Blattmann, Rahim Entezari, Jonas Müller, Harry Saini, Yam Levi, Dominik Lorenz, Axel Sauer, Frederic Boesel, et al. Scaling rectified flow transformers for high-resolution image synthesis. arXiv preprint arXiv:2403.03206, 2024. 2", + "[17] Rinon Gal, Yuval Alaluf, Yuval Atzmon, Or Patashnik, Amit H Bermano, Gal Chechik, and Daniel Cohen-Or. An image is worth one word: Personalizing text-to-image generation using textual inversion. arXiv preprint arXiv:2208.01618, 2022. 2", + "[18] Yuying Ge, Sijie Zhao, Jinguo Zhu, Yixiao Ge, Kun Yi, Lin Song, Chen Li, Xiaohan Ding, and Ying Shan. Seed-x: Multimodal models with unified multi-granularity comprehension and generation. arXiv preprint arXiv:2404.14396, 2024. 1, 2", + "[19] Yuchao Gu, Xintao Wang, Jay Zhangjie Wu, Yujun Shi, Yunpeng Chen, Zihan Fan, Wuyou Xiao, Rui Zhao, Shuning Chang, Weijia Wu, et al. Mix-of-show: Decentralized low-rank adaptation for multi-concept customization of diffusion models. Advances in Neural Information Processing Systems, 36:15890-15902, 2023. 2", + "[20] Jiayi Guo, Chaofei Wang, You Wu, Eric Zhang, Kai Wang, Xingqian Xu, Shiji Song, Humphrey Shi, and Gao Huang. Zero-shot generative model adaptation via image-specific prompt learning. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 11494-11503, 2023. 2", + "[21] Di He, Yingce Xia, Tao Qin, Liwei Wang, Nenghai Yu, TieYan Liu, and Wei-Ying Ma. Dual learning for machine translation. In Advances in Neural Information Processing Systems (NeurIPS), 2016. 3", + "[22] Martin Heusel, Hubert Ramsauer, Thomas Unterthiner, Bernhard Nessler, and Sepp Hochreiter. Gans trained by a two time-scale update rule converge to a local nash equilibrium. In Proceedings of the 31st International Conference on Neural Information Processing Systems, pages 6626-6637, 2017. 8" + ], + "bbox": [ + 516, + 92, + 903, + 898 + ], + "page_idx": 8 + }, + { + "type": "page_number", + "text": "2843", + "bbox": [ + 482, + 944, + 514, + 955 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[23] Judy Hoffman, Eric Tzeng, Taesung Park, Jun-Yan Zhu, Phillip Isola, Kate Saenko, Alexei A Efros, and Trevor Darryll. Cycada: Cycle-consistent adversarial domain adaptation. In Proceedings of the International Conference on Machine Learning (ICML), 2018. 3", + "[24] MD Zakir Hossain, Ferdous Sohel, Mohd Fairuz Shiratuddin, and Hamid Laga. A comprehensive survey of deep learning for image captioning. ACM Computing Surveys (CsUR), 51(6):1-36, 2019. 2", + "[25] Edward J. Hu, Yelong Shen, Phillip Wallis, Zeyuan Allen-Zhu, Yuanzhi Li, Shean Wang, and Weizhu Chen. Lora: Low-rank adaptation of large language models. In International Conference on Learning Representations, 2022. 5", + "[26] Jia Cheng Hu, Roberto Cavicchioli, and Alessandro Capontodi. Exploiting multiple sequence lengths in fast end to end training for image captioning. In 2023 IEEE International Conference on Big Data (BigData), pages 2173-2182. IEEE, 2023. 2", + "[27] Xiaowei Hu, Zhe Gan, Jianfeng Wang, Zhengyuan Yang, Zicheng Liu, Yumao Lu, and Lijuan Wang. Scaling up vision-language pre-training for image captioning. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 17980-17989, 2022. 2", + "[28] Eric Jang, Shixiang Gu, and Ben Poole. Categorical reparameterization with gumbel-softmax. arXiv preprint arXiv:1611.01144, 2016. 4", + "[29] Xuhui Jia, Yang Zhao, Kelvin CK Chan, Yandong Li, Han Zhang, Boqing Gong, Tingbo Hou, Huisheng Wang, and Yu-Chuan Su. Taming encoder for zero fine-tuning image customization with text-to-image diffusion models. arXiv preprint arXiv:2304.02642, 2023. 2", + "[30] Nupur Kumari, Bingliang Zhang, Richard Zhang, Eli Shechtman, and Jun-Yan Zhu. Multi-concept customization of text-to-image diffusion. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 1931-1941, 2023. 2", + "[31] Chenliang Li, Haiyang Xu, Junfeng Tian, Wei Wang, Ming Yan, Bin Bi, Jiabo Ye, Hehong Chen, Guohai Xu, Zheng Cao, et al. mplug: Effective and efficient vision-language learning by cross-modal skip-connections. arXiv preprint arXiv:2205.12005, 2022. 2", + "[32] Chunyuan Li, Zhe Gan, Zhengyuan Yang, Jianwei Yang, Linjie Li, Lijuan Wang, Jianfeng Gao, et al. Multimodal foundation models: From specialists to general-purpose assistants. Foundations and Trends® in Computer Graphics and Vision, 16(1-2):1-214, 2024. 2", + "[33] Junnan Li, Dongxu Li, Silvio Savarese, and Steven Hoi. Blip-2: Bootstrapping language-image pre-training with frozen image encoders and large language models. In International conference on machine learning, pages 19730-19742. PMLR, 2023. 2", + "[34] Tianhong Li, Sangnie Bhardwaj, Yonglong Tian, Han Zhang, Jarred Barber, Dina Katabi, Guillaume Lajoie, Huiwen Chang, and Dilip Krishnan. Leveraging unpaired data for vision-language generative models via cycle consistency. arXiv preprint arXiv:2310.03734, 2023. 3, 8", + "[35] Haotian Liu, Chunyuan Li, Qingyang Wu, and Yong Jae Lee. Visual instruction tuning. NeurIPS, 36, 2024. 2" + ], + "bbox": [ + 91, + 90, + 480, + 900 + ], + "page_idx": 9 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[36] Nanye Ma, Mark Goldstein, Michael S Albergo, Nicholas M Boffi, Eric Vanden-Eijnden, and Saining Xie. Sit: Exploring flow and diffusion-based generative models with scalable interpolant transformers. arXiv preprint arXiv:2401.08740, 2024. 2", + "[37] Xin Ma, Yaohui Wang, Gengyun Jia, Xinyuan Chen, Zwei Liu, Yuan-Fang Li, Cunjian Chen, and Yu Qiao. Latte: Latent diffusion transformer for video generation. arXiv preprint arXiv:2401.03048, 2024. 2", + "[38] Ron Mokady, Amir Hertz, Kfir Aberman, Yael Pritch, and Daniel Cohen-Or. Null-text inversion for editing real images using guided diffusion models. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 6038–6047, 2023. 2", + "[39] OpenAI. Chatgpt. https://chatgpt.com/, 2024.7", + "[40] Alec Radford, Jong Wook Kim, Chris Hallacy, Aditya Ramesh, Gabriel Goh, Sandhini Agarwal, Girish Sastry, Amanda Askell, Pamela Mishkin, Jack Clark, et al. Learning transferable visual models from natural language supervision. In International conference on machine learning, pages 8748-8763. PMLR, 2021. 2", + "[41] Aditya Ramesh, Mikhail Pavlov, Gabriel Goh, Scott Gray, Chelsea Voss, Alec Radford, Mark Chen, and Ilya Sutskever. Zero-shot text-to-image generation. In ICML, pages 8821-8831. Pmlr, 2021. 2", + "[42] Aditya Ramesh, Prafulla Dhariwal, Alex Nichol, Casey Chu, and Mark Chen. Hierarchical text-conditional image generation with CLIP latents. CoRR, abs/2204.06125, 2022. 2", + "[43] Aditya Ramesh, Prafulla Dhariwal, Alex Nichol, Casey Chu, and Mark Chen. Hierarchical diffusion models for text-to-image generation. arXiv preprint arXiv:2204.06125, 2022. 2", + "[44] Aditya Ramesh, Prafulla Dhariwal, Alex Nichol, Casey Chu, and Mark Chen. Hierarchical text-conditional image generation with clip latents. arXiv preprint arXiv:2204.06125, 1 (2):3, 2022. 2", + "[45] Robin Rombach, Andreas Blattmann, Dominik Lorenz, Patrick Esser, and Björn Ommer. High-resolution image synthesis with latent diffusion models. In CVPR, pages 10684-10695, 2022. 2", + "[46] Robin Rombach, Andreas Blattmann, Dominik Lorenz, Patrick Esser, and Björn Ommer. High-resolution image synthesis with latent diffusion models. arXiv preprint arXiv:2112.10752, 2022. 2", + "[47] Nataniel Ruiz, Yuanzhen Li, Varun Jampani, Yael Pritch, Michael Rubinstein, and Kfir Aberman. Dreambooth: Fine tuning text-to-image diffusion models for subject-driven generation. arXiv preprint arXiv:2208.12242, 2022. 1, 2", + "[48] Nataniel Ruiz, Yuanzhen Li, Varun Jampani, Yael Pritch, Michael Rubinstein, and Kfir Aberman. Dreambooth: Fine tuning text-to-image diffusion models for subject-driven generation. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 22500-22510, 2023. 5, 8", + "[49] Chitwan Sahara, William Chan, Saurabh Saxena, Lala Li, Jay Whang, Emily L Denton, Kamyar Ghasemipour, Raphael Gontijo Lopes, Burcu Karagol Ayan, Tim Salimans," + ], + "bbox": [ + 516, + 90, + 906, + 900 + ], + "page_idx": 9 + }, + { + "type": "page_number", + "text": "2844", + "bbox": [ + 482, + 944, + 514, + 955 + ], + "page_idx": 9 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "et al. Photorealistic text-to-image diffusion models with deep language understanding. Advances in Neural Information Processing Systems, 35:36479-36494, 2022. 2", + "[50] Rico Sennrich, Barry Haddow, and Alexandra Birch. Improving neural machine translation models with monolingual data. In Proceedings of the 54th Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers), 2016. 3", + "[51] Meet Shah, Xinlei Chen, Marcus Rohrbach, and Devi Parikh. Cycle-consistency for robust visual question answering. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 6649-6658, 2019. 3", + "[52] Sameer Shah, Siddharth Bharadwaj, Devi Parikh, and Dhruv Batra. Cycle-consistency for robust visual question answering. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition (CVPR), 2019. 3", + "[53] Quan Sun, Yufeng Cui, Xiaosong Zhang, Fan Zhang, Qiying Yu, Yueze Wang, Yongming Rao, Jingjing Liu, Tiejun Huang, and Xinlong Wang. Generative multimodal models are in-context learners. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 14398-14409, 2024. 2", + "[54] I Sutskever. Sequence to sequence learning with neural networks. arXiv preprint arXiv:1409.3215, 2014. 4", + "[55] Richard S Sutton. Learning to predict by the methods of temporal differences. Machine learning, 3:9-44, 1988. 4", + "[56] Zineng Tang, Ziyi Yang, Chenguang Zhu, Michael Zeng, and Mohit Bansal. Any-to-any generation via composable diffusion. NeurIPS, 36, 2024. 2", + "[57] Antti Tarvainen and Harri Valpola. Mean teachers are better role models: Weight-averaged consistency targets improve semi-supervised deep learning results. In Advances in Neural Information Processing Systems, pages 1195–1204, 2017. 2, 4", + "[58] Chameleon Team. Chameleon: Mixed-modal early-fusion foundation models. arXiv preprint arXiv:2405.09818, 2024. 1, 2, 3", + "[59] Ramakrishna Vedantam, C. Lawrence Zitnick, and Devi Parikh. Cider: Consensus-based image description evaluation. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition (CVPR), pages 4566-4575, 2015. 8", + "[60] Jianfeng Wang, Zhengyuan Yang, Xiaowei Hu, Linjie Li, Kevin Lin, Zhe Gan, Zicheng Liu, Ce Liu, and Lijuan Wang. Git: A generative image-to-text transformer for vision and language. arXiv preprint arXiv:2205.14100, 2022. 2", + "[61] Peng Wang, An Yang, Rui Men, Junyang Lin, Shuai Bai, Zhikang Li, Jianxin Ma, Chang Zhou, Jingren Zhou, and Hongxia Yang. Ofa: Unifying architectures, tasks, and modalities through a simple sequence-to-sequence learning framework. In International conference on machine learning, pages 23318-23340. PMLR, 2022. 2", + "[62] Qixun Wang, Xu Bai, Haofan Wang, Zekui Qin, Anthony Chen, Huaxia Li, Xu Tang, and Yao Hu. Instantid: Zero-shot identity-preserving generation in seconds. arXiv preprint arXiv:2401.07519, 2024. 1", + "[63] Yuchi Wang, Shuhuai Ren, Rundong Gao, Linli Yao, Qingyan Guo, Kaikai An, Jianhong Bai, and Xu Sun." + ], + "bbox": [ + 91, + 90, + 482, + 900 + ], + "page_idx": 10 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "Ladic: Are diffusion models really inferior to autoregressive counterparts for image-to-text generation? arXiv preprint arXiv:2404.10763, 2024. 2", + "[64] Chengyue Wu, Xiaokang Chen, Zhiyu Wu, Yiyang Ma, Xingchao Liu, Zizheng Pan, Wen Liu, Zhenda Xie, Xingkai Yu, Chong Ruan, et al. Janus: Decoupling visual encoding for unified multimodal understanding and generation. arXiv preprint arXiv:2410.13848, 2024. 2", + "[65] Shengqiong Wu, Hao Fei, Leigang Qu, Wei Ji, and Tat-Seng Chua. Next-gpt: Any-to-any multimodal llm. arXiv preprint arXiv:2309.05519, 2023. 2", + "[66] Jinheng Xie, Jiajun Feng, Zhaoxu Tian, Kevin Qinghong Lin, Yawen Huang, Xi Xia, Nanxu Gong, Xu Zuo, Ji-aqi Yang, Yefeng Zheng, et al. Learning long-form video prior via generative pre-training. arXiv preprint arXiv:2404.15909, 2024. 7", + "[67] Jinheng Xie, Weijia Mao, Zechen Bai, David Junhao Zhang, Weihao Wang, Kevin Qinghong Lin, Yuchao Gu, Zhijie Chen, Zhenheng Yang, and Mike Zheng Shou. Show-o: One single transformer to unify multimodal understanding and generation. arXiv preprint arXiv:2408.12528, 2024. 1, 2, 3, 5", + "[68] Ceyuan Yang, Yujun Shen, Zhiyi Zhang, Yinghao Xu, Jiapeng Zhu, Zhirong Wu, and Bolei Zhou. One-shot generative domain adaptation. In Proceedings of the IEEE/cvf international conference on computer vision, pages 7733-7742, 2023. 2", + "[69] Hanrong Ye, De-An Huang, Yao Lu, Zhiding Yu, Wei Ping, Andrew Tao, Jan Kautz, Song Han, Dan Xu, Pavlo Molchanov, et al. X-vila: Cross-modality alignment for large language model. arXiv preprint arXiv:2405.19335, 2024. 2", + "[70] Haoxuan You, Mandy Guo, Zhecan Wang, Kai-Wei Chang, Jason Baldridge, and Jiahui Yu. Cobit: A contrastive bidirectional image-text generation model. arXiv preprint arXiv:2303.13455, 2023.", + "[71] Lili Yu, Bowen Shi, Ramakanth Pasunuru, Benjamin Muller, Olga Golovneva, Tianlu Wang, Arun Babu, Binh Tang, Brian Karrer, Shelly Sheynin, et al. Scaling autoregressive multimodal models: Pretraining and instruction tuning. arXiv preprint arXiv:2309.02591, 2(3), 2023. 2", + "[72] Tianhe Yu, Saurabh Kumar, Abhishek Gupta, Sergey Levine, Karol Hausman, and Chelsea Finn. Gradient surgery for multi-task learning. In Advances in Neural Information Processing Systems, pages 5824-5836, 2020. 4", + "[73] Gengwei Zhang, Guoliang Kang, Yi Yang, and Yunchao Wei. Few-shot segmentation via cycle-consistent transformer. Advances in Neural Information Processing Systems, 34:21984-21996, 2021. 3", + "[74] Rui Zhao, Wei Li, Zhipeng Hu, Lincheng Li, Zhengxia Zou, Zhenwei Shi, and Changjie Fan. Zero-shot text-to-parameter translation for game character auto-creation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 21013–21023, 2023. 2", + "[75] Rui Zhao, Yuchao Gu, Jay Zhangjie Wu, David Junhao Zhang, Jia-Wei Liu, Weijia Wu, Jussi Keppo, and Mike Zheng Shou. Motiondirector: Motion customization of text-to-video diffusion models. In European Conference on Computer Vision, pages 273–290. Springer, 2024." + ], + "bbox": [ + 516, + 92, + 903, + 900 + ], + "page_idx": 10 + }, + { + "type": "page_number", + "text": "2845", + "bbox": [ + 482, + 944, + 514, + 955 + ], + "page_idx": 10 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[76] Rui Zhao, Hangjie Yuan, Yujie Wei, Shiwei Zhang, Yuchao Gu, Lingmin Ran, Xiang Wang, Jay Zhangjie Wu, David Junhao Zhang, Yingya Zhang, et al. Evolvedirector: Approaching advanced text-to-image generation with large vision-language models. Advances in Neural Information Processing Systems, 37:122104-122129, 2025. 2", + "[77] Chunting Zhou, Lili Yu, Arun Babu, Kushal Tirumala, Michihiro Yasunaga, Leonid Shamis, Jacob Kahn, Xuezhe Ma, Luke Zettlemoyer, and Omer Levy. Transfusion: Predict the next token and diffuse images with one multi-modal model. arXiv preprint arXiv:2408.11039, 2024. 1, 2, 3", + "[78] Deyao Zhu, Jun Chen, Xiaoqian Shen, Xiang Li, and Mohamed Elhoseiny. Minigpt-4: Enhancing vision-language understanding with advanced large language models. CoRR, abs/2304.10592, 2023. 2", + "[79] Jun-Yan Zhu, Taesung Park, Phillip Isola, and Alexei A. Efros. Unpaired image-to-image translation using cycle-consistent adversarial networks. In Proceedings of the IEEE International Conference on Computer Vision (ICCV), 2017. 3" + ], + "bbox": [ + 91, + 92, + 480, + 369 + ], + "page_idx": 11 + }, + { + "type": "page_number", + "text": "2846", + "bbox": [ + 482, + 945, + 516, + 955 + ], + "page_idx": 11 + } +] \ No newline at end of file diff --git a/2025/DoraCycle_ Domain-Oriented Adaptation of Unified Generative Model in Multimodal Cycles/fb2a505a-21d1-4afa-86a3-162bb4136d23_model.json b/2025/DoraCycle_ Domain-Oriented Adaptation of Unified Generative Model in Multimodal Cycles/fb2a505a-21d1-4afa-86a3-162bb4136d23_model.json new file mode 100644 index 0000000000000000000000000000000000000000..147bc3bfed38abdd03f7d2200d2ae841022f88c9 --- /dev/null +++ b/2025/DoraCycle_ Domain-Oriented Adaptation of Unified Generative Model in Multimodal Cycles/fb2a505a-21d1-4afa-86a3-162bb4136d23_model.json @@ -0,0 +1,3293 @@ +[ + [ + { + "type": "header", + "bbox": [ + 0.107, + 0.003, + 0.182, + 0.044 + ], + "angle": 0, + "content": "CVF" + }, + { + "type": "header", + "bbox": [ + 0.237, + 0.001, + 0.812, + 0.047 + ], + "angle": 0, + "content": "This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore." + }, + { + "type": "title", + "bbox": [ + 0.143, + 0.13, + 0.856, + 0.177 + ], + "angle": 0, + "content": "DoraCycle: Domain-Oriented Adaptation of Unified Generative Model in Multimodal Cycles" + }, + { + "type": "text", + "bbox": [ + 0.32, + 0.204, + 0.673, + 0.241 + ], + "angle": 0, + "content": "Rui Zhao, Weijia Mao, Mike Zheng Shou* Show Lab, National University of Singapore" + }, + { + "type": "title", + "bbox": [ + 0.248, + 0.274, + 0.327, + 0.29 + ], + "angle": 0, + "content": "Abstract" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.307, + 0.485, + 0.671 + ], + "angle": 0, + "content": "Adapting generative models to specific domains presents an effective solution for satisfying specialized requirements. However, adapting to some complex domains remains challenging, especially when these domains require substantial paired data to capture the targeted distributions. Since unpaired data from a single modality, such as vision or language, is more readily available, we utilize the bidirectional mappings between vision and language learned by the unified generative model to enable training on unpaired data for domain adaptation. Specifically, we propose DoraCycle, which integrates two multimodal cycles: text-to-image-to-text and image-to-text-to-image. The model is optimized through cross-entropy loss computed at the cycle endpoints, where both endpoints share the same modality. This facilitates self-evolution of the model without reliance on annotated text-image pairs. Experimental results demonstrate that for tasks independent of paired knowledge, such as stylization, DoraCycle can effectively adapt the unified model using only unpaired data. For tasks involving new paired knowledge, such as specific identities, a combination of a small set of paired image-text examples and larger-scale unpaired data is sufficient for effective domain-oriented adaptation. The code will be released at https://github.com/showlab/DoraCycle." + }, + { + "type": "title", + "bbox": [ + 0.092, + 0.699, + 0.222, + 0.715 + ], + "angle": 0, + "content": "1. Introduction" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.725, + 0.484, + 0.877 + ], + "angle": 0, + "content": "The adaptation of pre-trained generative models to specific domains is an important aspect of advancing personalized content creation, from stylized media outputs to customized identity generation [47, 62]. However, effectively adapting generative models to complex domains remains challenging, particularly when these domains require extensive amounts of paired data to accurately capture the desired distributions. For instance, learning the visual styles and character identities across a unique movie, which involves understanding multiple characters, their relationships, and di" + }, + { + "type": "image", + "bbox": [ + 0.52, + 0.275, + 0.688, + 0.46 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.532, + 0.463, + 0.671, + 0.475 + ], + "angle": 0, + "content": "(a) Training on paired data" + }, + { + "type": "image", + "bbox": [ + 0.724, + 0.275, + 0.891, + 0.46 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.722, + 0.462, + 0.898, + 0.475 + ], + "angle": 0, + "content": "(b) Training in Multimodal Cycles" + }, + { + "type": "image_caption", + "bbox": [ + 0.512, + 0.491, + 0.907, + 0.617 + ], + "angle": 0, + "content": "Figure 1. Training paradigms for unified generative models. (a) Traditional training involves using paired image-text data and optimizing the unified model with paired losses for both image-to-text (I2T) and text-to-image (T2I) generation tasks. (b) In contrast, the proposed multimodal cycle training framework leverages unpaired images and texts. By using cycle-consistency losses, the unified model learns to maintain consistency between input and output across modalities, enabling adaptation without the need for extensive paired datasets." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.643, + 0.907, + 0.719 + ], + "angle": 0, + "content": "verse settings, is a highly complex task that demands vast amounts of paired frames and captions data. Collecting such paired data, especially for multimodal tasks involving vision and language, is often laborious and impractical, limiting the potential to adapt generative models at scale." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.72, + 0.909, + 0.87 + ], + "angle": 0, + "content": "High-quality image-text paired data is relatively rare and scarce, but unpaired images and texts are readily available in our daily lives, such as video websites, image platforms, and content from novel websites. Therefore, we aim to explore whether it is possible to adapt generative models to target domains based on unpaired data. To achieve this goal, it is crucial for the model to have an internal capability to align the two modalities to some extent. Fortunately, recent advancements in unified generative models have encouraged us to pursue this direction." + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.871, + 0.908, + 0.902 + ], + "angle": 0, + "content": "Recent advanced unified generative models [14, 18, 58, 67, 77] have shown great potential in unifying multimodal" + }, + { + "type": "page_footnote", + "bbox": [ + 0.11, + 0.888, + 0.24, + 0.9 + ], + "angle": 0, + "content": "*Corresponding Author." + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.516, + 0.957 + ], + "angle": 0, + "content": "2835" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.09, + 0.092, + 0.482, + 0.288 + ], + "angle": 0, + "content": "understanding and generation within a single model. These unified models are capable of processing and generating content across different modalities, i.e. vision and language, within a shared framework. By leveraging the bidirectional mappings between vision and language, which are learned by the unified generative model in its pre-training stage, we can map each data from one modality to another and then back to the original modality, as shown in Fig. 1. Through these two mappings, data can be maintained within the same modality, thereby imposing constraints on the deviation introduced in the process. This only requires computing the cross-entropy loss between the data in the same modality, without any paired data supervision." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.291, + 0.483, + 0.518 + ], + "angle": 0, + "content": "To this end, we introduce DoraCycle, a framework for adapting unified multimodal generative models to target domains, through cycle-consistent learning with unpaired data. Unlike previous adaptation methods that heavily rely on paired text-image data, the proposed framework leverages the shared latent space of unified models to learn consistent transformations between modalities without requiring paired training examples. Specifically, we design two cycles, i.e. text-to-image-to-text (T cycle) and image-to-text-to-image (I cycle). As shown by Fig. 1 (b), leveraging the pre-trained vision-language alignment of the unified model, each multimodal cycle involves two cross-modality mappings while optimizing in the same modality. This enables calculating loss on unpaired data while implicitly refining cross-modal alignment through the intermediate step." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.52, + 0.483, + 0.793 + ], + "angle": 0, + "content": "In practice, since there is no labeled ground-truth, mapping data to another modality requires multi-step inference, such as predicting the next token multiple times for text generation. Allowing all forward steps to participate in gradient backpropagation can lead to a catastrophic gradient explosion. Therefore, we stop gradients during multi-step inference and use the generated data as pseudo labels for the model to forward once again, allowing gradients to propagate. Moreover, we found that since a complete cycle requires the model to forward twice, it can lead to training instability, with the quality of pseudo data generated in the middle being compromised. To enhance the stability of pseudo data generation, we maintain a slowly updated EMA (Exponential Moving Average) [57] model, which is used for inference to generate pseudo data. Additionally, we employ gradient clipping techniques to avoid conflicts in the optimization directions of the two cycles, further increasing training stability." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.796, + 0.483, + 0.901 + ], + "angle": 0, + "content": "The experiments indicate that for tasks independent of paired knowledge, such as stylization and domain-specific adaptation, DoraCycle can adapt the unified model with only unpaired data, which is both more practical and scalable. For tasks that require new paired knowledge, such as identity-specific adaptation, DoraCycle effectively utilizes small amounts of paired data along with larger-scale" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.092, + 0.905, + 0.228 + ], + "angle": 0, + "content": "unpaired data, making it a flexible solution for generative adaptation challenges. We conduct extensive experiments that compare DoraCycle to existing methods, showing that our approach achieves comparable or superior results while significantly reducing the need for paired data. This ability to harness large-scale unpaired data, combined with strategic usage of small paired datasets, makes DoraCycle a feasible solution for personalized multimodal content generation across a wide range of applications." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.243, + 0.665, + 0.259 + ], + "angle": 0, + "content": "2. Related Works" + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.269, + 0.885, + 0.285 + ], + "angle": 0, + "content": "2.1. Multimodal Generation and Understanding" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.292, + 0.905, + 0.518 + ], + "angle": 0, + "content": "Generating visual contents from text and describing them through natural language have been extensively studied as core multimodal tasks. Advanced generative models [3, 5, 6, 8, 12, 13, 16, 19, 36, 37, 42, 44, 45, 49, 74-76], such as DALL-E [41, 43], Stable Diffusion [46], demonstrate remarkable generation capabilities, producing high-quality and diverse contents from textual prompts. Meanwhile, image captioning models [24, 26, 27, 60, 61, 63], such as mPLUG [31], and BLIP [33], push the boundaries of visual understanding, generating accurate and context-aware descriptions. Additionally, recent advancements in multimodal large language models [32], such as LLaVA [35], MiniGPT-4 [78], and InstructBLIP [11], have significantly improved the ability to understand and reason about visual content." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.52, + 0.906, + 0.641 + ], + "angle": 0, + "content": "Besides the powerful foundational generative models, adapting or customizing them attracts increasing interest, which enables more personalized and specific outputs based on user preferences [7, 9, 17, 20, 29, 30, 38, 68]. Approaches like DreamBooth [47] enable user-specific customization by fine-tuning generative models with personal data, allowing the generation of content tailored to individual needs or preferences." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.652, + 0.851, + 0.667 + ], + "angle": 0, + "content": "2.2. Unified Multimodal Generative Models" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.675, + 0.906, + 0.901 + ], + "angle": 0, + "content": "Unified multimodal generative models aim to bridge the gap between understanding and generation tasks, and integrate vision and language into a single framework, enabling the model to learn shared representations across modalities [1, 2, 14, 18, 53, 56, 64, 65, 67, 69-71, 77]. SEED-X [18] utilizes a unified architecture where visual features extracted from the CLIP ViT encoder [40] are combined with text tokens and fed into a large language model to enable both next-word prediction and image regression tasks. DreamLLM [14] extends the generative capability of large language models by combining multimodal inputs directly into LLMs. Chameleon [58] employs a discrete tokenization approach for both visual and textual inputs, converting all modalities into a unified token space that is processed by a transformer-based architecture. Transfusion [77] in" + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.946, + 0.516, + 0.956 + ], + "angle": 0, + "content": "2836" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.09, + 0.092, + 0.482, + 0.243 + ], + "angle": 0, + "content": "produces an advanced integration mechanism that focuses on directly fusing visual encoding with language tokens, allowing the model to effectively translate visual information into textual formats while maintaining the semantic integrity of both modalities. Show-o [67] combines autoregressive modeling with a discrete diffusion process, enabling the generation of high-quality outputs that are aligned across modalities. Our work leverages the advancements made by these foundational models and explores how to adapt the foundational model to specific domains." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.267, + 0.266, + 0.284 + ], + "angle": 0, + "content": "2.3. Cycle Consistency" + }, + { + "type": "text", + "bbox": [ + 0.093, + 0.294, + 0.485, + 0.626 + ], + "angle": 0, + "content": "Cycle consistency has been utilized in computer vision and natural language processing as a means to enhance model robustness and consistency [10, 15, 21, 23, 51, 51, 52, 73]. CycleGAN [79] introduced cycle consistency loss to align unpaired image domains, ensuring that mappings between domains (e.g., \\(\\mathrm{A} \\to \\mathrm{B} \\to \\mathrm{A}\\)) remain consistent. In the field of natural language processing, back-translation employs similar ideas by translating sentences between languages in both directions to improve translation quality [50]. However, the cycle consistency in these works is in a single modality, i.e. vision or language. Recently, ITIT [34] was proposed to utilize cycle consistency to train vision-language generative models. ITIT takes in a mixture of unpaired data and paired data to pre-train the foundational generative model. It is constructed with one image-text encoder and two modality-specific decoders, which operate on the encoded image-text features to generate either text or image tokens. In contrast, we utilized a single unified transformer to parse and predict text and image tokens together. Besides, we focus on adapting pre-trained foundational models to new domains efficiently rather than retraining new foundation models." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.655, + 0.182, + 0.67 + ], + "angle": 0, + "content": "3. Method" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.684, + 0.483, + 0.851 + ], + "angle": 0, + "content": "The proposed DoraCycle framework, as shown in Fig. 2, is built upon the unified generative model designed for multimodal tasks involving both vision and language [58, 67, 77]. The unified model uses a single transformer to learn bidirectional mappings between vision and language, providing a powerful backbone capable of processing and generating different modalities [67]. For captioning, the model takes in image tokens and predicts corresponding text tokens, while for image generation, it takes in text tokens and predicts image tokens. This versatility makes the unified model well-suited as a base for our proposed framework." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.856, + 0.484, + 0.902 + ], + "angle": 0, + "content": "In the following sections, we first introduce the design of multimodal cycles, and then discuss the stabilization of optimization, and the balance of two cycles." + }, + { + "type": "title", + "bbox": [ + 0.514, + 0.091, + 0.696, + 0.108 + ], + "angle": 0, + "content": "3.1. Multimodal Cycles" + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.113, + 0.905, + 0.204 + ], + "angle": 0, + "content": "To adapt the unified model for domain-oriented adaptations using unpaired data, we design two multimodal cycles: the Image-to-Text-to-Image Cycle (I Cycle) and the Text-to-Image-to-Text Cycle (T Cycle). Each cycle utilizes data from a single modality, allowing the model to adapt without relying on paired data." + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.204, + 0.905, + 0.385 + ], + "angle": 0, + "content": "T Cycle: The T cycle training involves transforming an input textual sequence into an image representation and then back into a textual sequence, enforcing consistency between the generated and original text. Specifically, at each training iteration, we begin with an input text sequence \\( T = [t_l]_{l=1}^L \\). Conditioned on this, the unified model generates pseudo-paired image tokens \\( I' \\), representing the visual interpretation of the input text. The generated image tokens are then subjected to a random masking operation, denoted as \\( I_M' \\), where a subset of the tokens is masked. The unified model is then called to reconstruct these masked tokens to form the complete synthesized image \\( \\tilde{I} \\)." + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.385, + 0.905, + 0.461 + ], + "angle": 0, + "content": "In the second half of the cycle, conditioning on image \\(\\tilde{I}\\), the model generates the corresponding text sequence. The objective of the T cycle is to enforce cycle consistency between the generated text and the original input text \\(T\\). The T cycle consistency loss is defined as follows:" + }, + { + "type": "equation", + "bbox": [ + 0.531, + 0.485, + 0.907, + 0.526 + ], + "angle": 0, + "content": "\\[\n\\mathcal {L} _ {T C} = - \\mathbb {E} _ {T \\in \\mathcal {D} t e x t} \\left[ \\sum_ {l = 1} ^ {L} \\log p \\left(t _ {l} \\mid \\tilde {I}, t _ {0}, \\dots , t _ {l - 1}\\right) \\right], \\tag {1}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.537, + 0.905, + 0.565 + ], + "angle": 0, + "content": "where the \\( \\mathcal{D} \\) text is the set of text samples from the target domain." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.567, + 0.907, + 0.703 + ], + "angle": 0, + "content": "I Cycle: The I cycle training begins with an input image, which is subsequently mapped to a textual representation and then transformed back to an image, enforcing consistency between the generated image and the original image tokens. At each training iteration, we start with input image tokens \\( I \\). The unified model is used to synthesize pseudopaired text tokens \\( T' \\), representing the textual description of the image. We then use \\( T' \\) in conjunction with the input image token \\( I \\) to predict the reconstructed text tokens \\( \\tilde{T} \\)." + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.703, + 0.905, + 0.78 + ], + "angle": 0, + "content": "In the second half of the cycle, we pass the masked image tokens \\( I_M \\) and the text \\( \\tilde{T} \\) through the model to regenerate the masked image tokens. The cycle enforces consistency between the reconstructed and the original image tokens. The loss for enforcing I cycle consistency is given by:" + }, + { + "type": "equation", + "bbox": [ + 0.54, + 0.803, + 0.907, + 0.837 + ], + "angle": 0, + "content": "\\[\n\\mathcal {L} _ {I C} = - \\mathbb {E} _ {I \\in \\mathcal {D} _ {i m a g e}} \\left[ \\sum_ {\\forall k: m _ {k} = 1} \\log p (i _ {k} | I _ {M}, \\tilde {T}) \\right], \\tag {2}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.841, + 0.905, + 0.871 + ], + "angle": 0, + "content": "where the \\(D\\) image is the set of image samples from the target domain, which are unpaired with the text samples." + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.871, + 0.905, + 0.902 + ], + "angle": 0, + "content": "By leveraging these two cycles, our framework forces the model to refine its generative understanding of both image" + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.516, + 0.957 + ], + "angle": 0, + "content": "2837" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.099, + 0.093, + 0.904, + 0.282 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.09, + 0.293, + 0.908, + 0.322 + ], + "angle": 0, + "content": "Figure 2. The overview of T cycle (text-to-image-to-text) of the proposed DoraCycle. The I cycle is similar and is omitted in the figure for brevity." + }, + { + "type": "text", + "bbox": [ + 0.089, + 0.348, + 0.483, + 0.394 + ], + "angle": 0, + "content": "and text representations, ensuring consistency between the input and output while effectively leveraging unpaired data to adapt the unified towards the target domain." + }, + { + "type": "text", + "bbox": [ + 0.089, + 0.395, + 0.483, + 0.605 + ], + "angle": 0, + "content": "Efficient Training: In the intermediate steps of both cycles, generating the middle representation (i.e., captions or images) requires multiple forward passes. This is because the generation process involves either predicting the next tokens or the masked tokens multiple times. Backpropagating gradients through all these steps is computationally prohibitive. Thus, we first generate the intermediate results using the model in inference mode as pseudo-paired data, which are then used as the ground truth in the teacher-forcing scheme [54, 55] for the first half of the cycles. In this way, we reduce the number of forward passes to two, i.e. one for generating the middle result and one for the final output, thus making the overall training process more memory efficient." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.607, + 0.484, + 0.669 + ], + "angle": 0, + "content": "Token Differentiability: Since the intermediate outputs in each cycle are discrete tokens, which can not directly propagate gradients, we employ the Gumbel-Softmax [28] to make these token representations differentiable." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.681, + 0.317, + 0.697 + ], + "angle": 0, + "content": "3.2. Stabilizing Optimization" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.704, + 0.483, + 0.81 + ], + "angle": 0, + "content": "Each cycle involves the same unified model twice in the forward pass, which leads to optimization instabilities. To stabilize the training process, we adopt the Exponential Moving Average (EMA) training technique [57]. Specifically, we maintain a shadow version of the model, referred to as the EMA model, which is updated using an exponentially decaying average of the parameters of the main model." + }, + { + "type": "equation", + "bbox": [ + 0.179, + 0.83, + 0.483, + 0.845 + ], + "angle": 0, + "content": "\\[\n\\theta_ {\\mathrm {E M A}} \\leftarrow \\alpha \\theta_ {\\mathrm {E M A}} + (1 - \\alpha) \\theta_ {\\text {m i n}}, \\tag {3}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.856, + 0.484, + 0.9 + ], + "angle": 0, + "content": "where \\(\\alpha\\) is a decay factor (set to 0.999) that controls the update rate, and \\(\\theta_{\\mathrm{main}}\\) represents the parameters of the main model." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.348, + 0.907, + 0.485 + ], + "angle": 0, + "content": "In each training step, the EMA version of the model is used to generate the intermediate representation tokens (e.g., pseudo image or text tokens) which serve as pseudo ground truth during training. By using these stable targets from the slower-evolving EMA model, we can mitigate the risks of optimization instability. The main model is thus able to learn from more consistent and reliable intermediate targets, rather than being affected by fluctuations during the early stages of training." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.497, + 0.72, + 0.513 + ], + "angle": 0, + "content": "3.3. Balancing Two Cycles" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.52, + 0.906, + 0.627 + ], + "angle": 0, + "content": "We observe that the T cycle tends to converge faster than the I cycle, primarily because textual data is inherently one-dimensional and simpler to learn compared to images. This imbalance in optimization leads to a kind of collapse of the model, where it tends to generate irrelevant but self-consistent captions for images, ultimately degrading the image-text alignment capability." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.628, + 0.906, + 0.735 + ], + "angle": 0, + "content": "To address this problem, we make the gradients of the T cycle orthogonal to those of the I cycle, thus preventing interference. This is achieved by modifying the gradients using gradient surgery [72]. Let \\( g_{T} \\) and \\( g_{I} \\) represent the gradients of the T cycle and the I cycle, respectively. We project \\( g_{T} \\) onto the orthogonal complement of \\( g_{I} \\) to obtain the modified gradient \\( g_{T}^{\\prime} \\), which is defined as:" + }, + { + "type": "equation", + "bbox": [ + 0.633, + 0.748, + 0.907, + 0.779 + ], + "angle": 0, + "content": "\\[\ng _ {T} ^ {\\prime} = g _ {T} - \\frac {g _ {T} \\cdot g _ {I}}{g _ {I} \\cdot g _ {I}} g _ {I}, \\tag {4}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.786, + 0.906, + 0.815 + ], + "angle": 0, + "content": "where \\( g_{T} \\cdot g_{I} \\) denotes the dot product between the gradients of the T and I cycle." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.817, + 0.906, + 0.861 + ], + "angle": 0, + "content": "Additionally, we reweight the losses to further balance the learning between the I and T cycles. The final loss function is as," + }, + { + "type": "equation", + "bbox": [ + 0.643, + 0.866, + 0.905, + 0.881 + ], + "angle": 0, + "content": "\\[\n\\mathcal {L} = \\mathcal {L} _ {I C} + \\beta \\mathcal {L} _ {T C}, \\tag {5}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.886, + 0.813, + 0.902 + ], + "angle": 0, + "content": "where the \\(\\beta\\) is the weight of the T cycle loss." + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.516, + 0.956 + ], + "angle": 0, + "content": "2838" + } + ], + [ + { + "type": "title", + "bbox": [ + 0.091, + 0.09, + 0.224, + 0.108 + ], + "angle": 0, + "content": "4. Experiments" + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.115, + 0.308, + 0.132 + ], + "angle": 0, + "content": "4.1. Implementation Details" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.139, + 0.484, + 0.32 + ], + "angle": 0, + "content": "To the best of our knowledge, Show-o [67] is currently the only fully open-source unified generative model with complete pre-trained weights and training code, including both its understanding and generation capabilities. Therefore, we base DoraCycle on Show-o and conduct experiments accordingly. The base model is a unified transformer model that performs understanding and image generation by predicting discrete textual and visual tokens. We insert trainable low-rank adaptations (LoRA) [25] modules into the Q projection and V projection of the attention layers from layers 7 to 24. The LoRA rank is set to 32. The \\(\\beta\\) is set to 0.1 to balance the optimization of two cycles." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.321, + 0.484, + 0.442 + ], + "angle": 0, + "content": "The training of DoraCycle is performed on 8 NVIDIA H100 GPUs with mixed precision enabled for memory efficiency. We set the batch size to 32, with each cycle taking half of the batch when both cycles are being optimized simultaneously. The learning rate is set to \\(1e^{-4}\\) with a cosine annealing schedule. The optimizer is AdamW with weight decay of \\(1e^{-2}\\). Additionally, EMA is employed to stabilize the training process, as described in Section 3.2." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.455, + 0.362, + 0.47 + ], + "angle": 0, + "content": "4.2. Domain-Oriented Adaptations" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.477, + 0.483, + 0.599 + ], + "angle": 0, + "content": "Unpaired Training: For tasks that do not require strongly related paired knowledge, our DoraCycle can fully learn the target domain using unpaired data. For example, to learn the cyberpunk style, we collected 300 cyberpunk-style images as input for the I cycle, and used the text data from the base model pre-training dataset [4] for the T cycle, with the keyword \"cyberpunk style\" automatically injected into text, prompting the model about the target style we want." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.599, + 0.484, + 0.901 + ], + "angle": 0, + "content": "The experimental results are shown in Fig. 3. Given the same text prompt to generate cyberpunk-style images, Fig. 3 (a) shows the images generated by the base model without additional training. It can be observed that the base model adds some cyberpunk elements, such as neon lights, but the overall atmosphere does not align well with the desired style. Fig. 3 (d) shows the images generated by the adapted model trained with DoraCycle, which aligns well with the target style. Traditional text-to-image customization or adaptation methods, such as DreamBooth [48], rely on paired data for training. Therefore, we simulate user-created paired data by annotating the collected images with captions, and split them into two groups. One group contained only 10 paired examples, which is an acceptable workload for users, while the other group contained captions for all 300 images, which would be labor-intensive and impractical for users. The images generated by the model that trained on 10 paired examples are shown in Fig. 3 (b). It struggled to produce good stylized images, likely because the combination of indoor bookshelves with the cyberpunk" + }, + { + "type": "text", + "bbox": [ + 0.518, + 0.09, + 0.901, + 0.119 + ], + "angle": 0, + "content": " The image presents a scene of a room with a brick wall and a large bookshelf filled with various books and electronics. The bookshelf, which is black, stands against the brick wall and spans almost the entire height of the wall. It is divided into three shelves, each holding different items. On the top shelf, there are several books neatly arranged. In front of the bookshelf, there's a black leather couch positioned on one side of the room. A potted plant can be seen as well." + }, + { + "type": "image_caption", + "bbox": [ + 0.703, + 0.12, + 0.783, + 0.13 + ], + "angle": 0, + "content": "Image-to-Text" + }, + { + "type": "image", + "bbox": [ + 0.518, + 0.133, + 0.614, + 0.208 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.53, + 0.212, + 0.602, + 0.235 + ], + "angle": 0, + "content": "(a) Base model (w/o training)" + }, + { + "type": "image", + "bbox": [ + 0.616, + 0.133, + 0.707, + 0.208 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.619, + 0.213, + 0.703, + 0.235 + ], + "angle": 0, + "content": "(b) Paired trained (10 pairs)" + }, + { + "type": "image", + "bbox": [ + 0.709, + 0.133, + 0.807, + 0.208 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.717, + 0.213, + 0.799, + 0.235 + ], + "angle": 0, + "content": "(c) Paired trained (300 pairs)" + }, + { + "type": "image", + "bbox": [ + 0.808, + 0.133, + 0.905, + 0.208 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.8, + 0.213, + 0.903, + 0.235 + ], + "angle": 0, + "content": "(d) Unpaired trained (DoraCycle)" + }, + { + "type": "image", + "bbox": [ + 0.517, + 0.244, + 0.63, + 0.331 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.528, + 0.348, + 0.588, + 0.358 + ], + "angle": 0, + "content": "Image-to-Text" + }, + { + "type": "image", + "bbox": [ + 0.594, + 0.333, + 0.634, + 0.363 + ], + "angle": 0, + "content": null + }, + { + "type": "text", + "bbox": [ + 0.637, + 0.244, + 0.781, + 0.365 + ], + "angle": 0, + "content": " The image captures a scene on a city street. Dominating the foreground are two mannequins, each adorned in full Batman and Iron Man costumes. The Mannequin on the left is dressed in a black Batman costume, complete with a cape and mask, exuding an air of mystery and power. On the right, the other Mannequin is clad in a red Iron Man costume, featuring a helmet with a faceplate and red gloves, embodying the iconic superhero's persona. These mannequins stand on a gray sidewalk, their vibrant costumes contrasting against the monotone backdrop. Behind them, the city continues its usual hustle and bustle. Cars are parked along the street, their metallic bodies gleaming under the city lights. Buildings rise in the background, their windows reflecting the life of the city." + }, + { + "type": "image", + "bbox": [ + 0.792, + 0.244, + 0.904, + 0.331 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.791, + 0.333, + 0.892, + 0.363 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.545, + 0.372, + 0.868, + 0.384 + ], + "angle": 0, + "content": "(e) Image-to-Text-to-Image translation by the adapted unified model." + }, + { + "type": "image_caption", + "bbox": [ + 0.512, + 0.396, + 0.907, + 0.507 + ], + "angle": 0, + "content": "Figure 3. Domain-oriented adaptation with different training setups. (a) Image generated by the base model without training for adoption. (b) Image generated by the model trained with 10 paired image-text samples. (c) Image generated by the model trained with 300 paired image-text samples. (d) Image generated by the model trained by DoraCycle on only unpaired data. (e) Image-to-Text-to-Image translation performed by the adapted model trained by DoraCycle." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.536, + 0.906, + 0.627 + ], + "angle": 0, + "content": "style is too novel for the model to generalize well from limited paired data. The images generated by the model trained on 300 paired examples are shown in Fig. 3 (c), which have better outputs. In contrast, the model trained using DoraCycle does not require manual captioning, significantly reducing the workload for users." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.628, + 0.906, + 0.824 + ], + "angle": 0, + "content": "Fig. 3 (e) illustrates the adapted model trained by DoraCyle maintains semantic consistency through image-to-text-to-image translation. The input image is transformed into a textual description and then reconstructed into an image. The result shows that the adapted model successfully captures and retains the key visual components in the original image throughout the multimodal cycle. Notably, the identity of the characters and the details of the environment are all preserved, indicating effective bidirectional understanding and generation capabilities in the target domain. Furthermore, the newly generated image incorporates styles learned from the target domain, demonstrating the generalization of the learned knowledge to images in the wild." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.826, + 0.906, + 0.901 + ], + "angle": 0, + "content": "Learning Paired Knowledge: For tasks that require learning some paired knowledge, such as associating an identity name with its visual appearance, DoraCycle can incorporate a small amount of paired data to learn such associations while leveraging a large amount of unpaired" + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.516, + 0.957 + ], + "angle": 0, + "content": "2839" + } + ], + [ + { + "type": "title", + "bbox": [ + 0.407, + 0.093, + 0.599, + 0.107 + ], + "angle": 0, + "content": "Domain 1: Black Myth Wukong" + }, + { + "type": "title", + "bbox": [ + 0.107, + 0.128, + 0.124, + 0.226 + ], + "angle": 270, + "content": "Text-to-Image" + }, + { + "type": "image", + "bbox": [ + 0.128, + 0.109, + 0.253, + 0.207 + ], + "angle": 0, + "content": null + }, + { + "type": "text", + "bbox": [ + 0.129, + 0.21, + 0.249, + 0.258 + ], + "angle": 0, + "content": "Destined One, wears an intricately detailed golden armor with elaborate ornaments and engravings. A golden crown-like adornment sits atop his head. In the background, distant mountains and a serene sky can be seen." + }, + { + "type": "image", + "bbox": [ + 0.256, + 0.109, + 0.382, + 0.207 + ], + "angle": 0, + "content": null + }, + { + "type": "text", + "bbox": [ + 0.258, + 0.209, + 0.377, + 0.252 + ], + "angle": 0, + "content": "Lingxuizi's glowing red eyes and bared fangs indicate its aggressive nature. It crouches low to the ground, its large hand-like paw spread over ancient stone paving surrounded by patches of grass." + }, + { + "type": "image", + "bbox": [ + 0.384, + 0.109, + 0.51, + 0.207 + ], + "angle": 0, + "content": null + }, + { + "type": "text", + "bbox": [ + 0.386, + 0.209, + 0.505, + 0.26 + ], + "angle": 0, + "content": "Keeper grips a gnarled wooden staff for support as he walks through a wooded area. His posture is stooped, and his facial features are mostly obscured by his thick beard. The scene is set in a forest, with a textured ground of leaves and dirt." + }, + { + "type": "image", + "bbox": [ + 0.512, + 0.109, + 0.638, + 0.207 + ], + "angle": 0, + "content": null + }, + { + "type": "text", + "bbox": [ + 0.515, + 0.209, + 0.633, + 0.265 + ], + "angle": 0, + "content": "Zhu Bajie stands confidently in the snow, gripping his staff with a sly smile on his boat-like face, beads hanging around his neck. Beside him lies the majestic, wounded form of Kang-Jin Loong, a silver-scaled dragon with intricate, ethereal horns and whiskers." + }, + { + "type": "image", + "bbox": [ + 0.64, + 0.109, + 0.766, + 0.207 + ], + "angle": 0, + "content": null + }, + { + "type": "text", + "bbox": [ + 0.643, + 0.209, + 0.762, + 0.278 + ], + "angle": 0, + "content": "Tiger Vanguard charges through the blood-red waters towards Destined One. The massive stone statue behind him looms ominously, adding to the intensity of the scene. Destined One stands firm, his staff held at the ready, his gaze unwavering as he prepares for the oncoming assault. The water ripples around them, reflecting the dark hues of the battle setting." + }, + { + "type": "image", + "bbox": [ + 0.768, + 0.109, + 0.895, + 0.207 + ], + "angle": 0, + "content": null + }, + { + "type": "text", + "bbox": [ + 0.77, + 0.209, + 0.89, + 0.278 + ], + "angle": 0, + "content": "Destined One stands tall, his gaze cast downward, his weathered yellow robe flowing with subtle wear, reflecting the trials he has endured. Beside him, Zhu Majie looks up with a playful expression, his boar-like face filled with curiosity and mischief. The two figures share a quiet moment, framed against an open, clear sky dotted with drifting embers." + }, + { + "type": "image_caption", + "bbox": [ + 0.182, + 0.279, + 0.197, + 0.288 + ], + "angle": 0, + "content": "(a)" + }, + { + "type": "title", + "bbox": [ + 0.107, + 0.309, + 0.124, + 0.365 + ], + "angle": 270, + "content": "Image-to-Text" + }, + { + "type": "image", + "bbox": [ + 0.129, + 0.289, + 0.253, + 0.386 + ], + "angle": 0, + "content": null + }, + { + "type": "text", + "bbox": [ + 0.129, + 0.39, + 0.252, + 0.466 + ], + "angle": 0, + "content": "Destined One stands tall, wielding a massive golden stick that dominates the frame. The individual exudes power, strength, and determination, clad in ornate armor that gleams with gold. A hat adds a touch of regality. The background is dramatic, with shadows cast by trees hinting at a mysterious atmosphere. The grass beneath the figure is lush and green, contrasting with the ruggedness of the ground." + }, + { + "type": "image_caption", + "bbox": [ + 0.31, + 0.279, + 0.326, + 0.288 + ], + "angle": 0, + "content": "(b)" + }, + { + "type": "image", + "bbox": [ + 0.256, + 0.289, + 0.382, + 0.387 + ], + "angle": 0, + "content": null + }, + { + "type": "text", + "bbox": [ + 0.259, + 0.39, + 0.381, + 0.452 + ], + "angle": 0, + "content": "Keeper, an elderly, hunched individual with a bald, heavily wrinkled head covered in spots. Keeper has a long, unkempt beard and is dressed in a simple, tattered blue robe. He grips a gnarled wooden staff for support as he walks through a forest fire. His posture is stooped, and his facial features are mostly obscured by his thick beard." + }, + { + "type": "image_caption", + "bbox": [ + 0.439, + 0.279, + 0.455, + 0.288 + ], + "angle": 0, + "content": "(c)" + }, + { + "type": "image", + "bbox": [ + 0.383, + 0.289, + 0.511, + 0.387 + ], + "angle": 0, + "content": null + }, + { + "type": "text", + "bbox": [ + 0.385, + 0.39, + 0.506, + 0.465 + ], + "angle": 0, + "content": "The image depicts a serene, snowy landscape with a large, ancient-looking wooden bridge spanning across a frozen river. The bridge has a traditional, Asian design, with intricate carvings and a slatted roof. Bare trees with frost-covered branches surround the scene, casting a calm, wintery atmosphere. Snowflakes gently fall, enhancing the tranquil, mystical ambiance" + }, + { + "type": "image_caption", + "bbox": [ + 0.568, + 0.279, + 0.583, + 0.288 + ], + "angle": 0, + "content": "(d)" + }, + { + "type": "image", + "bbox": [ + 0.513, + 0.289, + 0.637, + 0.387 + ], + "angle": 0, + "content": null + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.39, + 0.637, + 0.465 + ], + "angle": 0, + "content": "Lingxuzi is depicted as a large, menacing fox-like creature with thick, flowing white fur. Its red eyes glow intensely, giving an eerie, supernatural appearance. The creature is in a crouching stance, with its body slightly arched and its tails fanned out behind it, showing both power and readiness to strike. The dark background adds to the ominous atmosphere, contrasting with the brightness of its fur." + }, + { + "type": "image_caption", + "bbox": [ + 0.696, + 0.279, + 0.711, + 0.288 + ], + "angle": 0, + "content": "(e)" + }, + { + "type": "image", + "bbox": [ + 0.64, + 0.289, + 0.766, + 0.387 + ], + "angle": 0, + "content": null + }, + { + "type": "text", + "bbox": [ + 0.642, + 0.39, + 0.763, + 0.478 + ], + "angle": 0, + "content": "Zhu Bajie stands menacingly in a murky, mist-covered swamp, his humanoid body combined with the head of a fierce, snarling boar. Her red, glowing eyes pierce through the fog, radiating a dark energy. Dressed in a tattered green robe, he wears a string of large beads around his neck, adding a sense of mysticism. His left hand clutches a spiked weapon, while shadows and wisps of smoke swirl around him, amplifying his sinister presence." + }, + { + "type": "image_caption", + "bbox": [ + 0.824, + 0.279, + 0.838, + 0.288 + ], + "angle": 0, + "content": "(f)" + }, + { + "type": "image", + "bbox": [ + 0.769, + 0.289, + 0.895, + 0.387 + ], + "angle": 0, + "content": null + }, + { + "type": "text", + "bbox": [ + 0.77, + 0.39, + 0.891, + 0.483 + ], + "angle": 0, + "content": "A female character with a unique appearance, having a combination of human and demonic features. She has a short, black, straight hairstyle and a pair of horns protruding from her head. Her skin is pale, and she has a few dots on her forehead, giving her a mystical look. She is wearing a black, low-cut top adorned with white flowers, and her right arm is covered in red fabric. She has a focused and intense expression on her face, as if she is concentrating on something important. (D)" + }, + { + "type": "image_caption", + "bbox": [ + 0.825, + 0.482, + 0.838, + 0.49 + ], + "angle": 0, + "content": "(1)" + }, + { + "type": "image_caption", + "bbox": [ + 0.182, + 0.48, + 0.197, + 0.49 + ], + "angle": 0, + "content": "(g)" + }, + { + "type": "image_caption", + "bbox": [ + 0.31, + 0.48, + 0.326, + 0.49 + ], + "angle": 0, + "content": "(h)" + }, + { + "type": "image_caption", + "bbox": [ + 0.439, + 0.48, + 0.454, + 0.49 + ], + "angle": 0, + "content": "(i)" + }, + { + "type": "image_caption", + "bbox": [ + 0.568, + 0.48, + 0.583, + 0.49 + ], + "angle": 0, + "content": "(j)" + }, + { + "type": "image_caption", + "bbox": [ + 0.433, + 0.496, + 0.566, + 0.507 + ], + "angle": 0, + "content": "Domain 2: Doraemon" + }, + { + "type": "title", + "bbox": [ + 0.107, + 0.532, + 0.124, + 0.586 + ], + "angle": 270, + "content": "Text-to-Image" + }, + { + "type": "image", + "bbox": [ + 0.129, + 0.512, + 0.255, + 0.592 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.129, + 0.593, + 0.248, + 0.608 + ], + "angle": 0, + "content": null + }, + { + "type": "text", + "bbox": [ + 0.129, + 0.613, + 0.248, + 0.641 + ], + "angle": 0, + "content": "Nobita is sitting under a tree, trying to finish his math homework. He looks stressed, with one hand on his chin and scratching his head with the other." + }, + { + "type": "image", + "bbox": [ + 0.257, + 0.512, + 0.383, + 0.608 + ], + "angle": 0, + "content": null + }, + { + "type": "text", + "bbox": [ + 0.258, + 0.612, + 0.377, + 0.674 + ], + "angle": 0, + "content": "Suneo and Nobita are at a fishing pond, trying to catch a fish. Nobita, in his bright yellow fishing hat and a sharp outfit, confidently casts his line into the water. Suneo watches Nobita's struggle with a smile, offering some tips to help him out. The two boys share a laugh as they wait for a bite." + }, + { + "type": "image", + "bbox": [ + 0.386, + 0.512, + 0.511, + 0.608 + ], + "angle": 0, + "content": null + }, + { + "type": "text", + "bbox": [ + 0.386, + 0.612, + 0.504, + 0.647 + ], + "angle": 0, + "content": "Doraemon is cooking in the kitchen, while holding a bowl. He looks excited, flipping a pancake into the air. Nobita is standing beside him, smiling gently, watching Doraemon." + }, + { + "type": "image", + "bbox": [ + 0.513, + 0.512, + 0.638, + 0.608 + ], + "angle": 0, + "content": null + }, + { + "type": "text", + "bbox": [ + 0.514, + 0.612, + 0.633, + 0.667 + ], + "angle": 0, + "content": "Shizuka is sitting by a piano, composing a new tune. She looks thoughtful, her fingers delicately pressing the keys as she tries different melodies. A notebook rests on the piano stand, filled with her notes and musical ideas, and she appears content in the process." + }, + { + "type": "image", + "bbox": [ + 0.64, + 0.512, + 0.764, + 0.609 + ], + "angle": 0, + "content": null + }, + { + "type": "text", + "bbox": [ + 0.642, + 0.612, + 0.761, + 0.647 + ], + "angle": 0, + "content": "Shizuka and Nobita are making flower crowns at the park. Shizuka, in her soft pink dress, weaves flowers together. Nobita, in his casual shorts, tries his best to copy her." + }, + { + "type": "image", + "bbox": [ + 0.768, + 0.512, + 0.892, + 0.608 + ], + "angle": 0, + "content": null + }, + { + "type": "text", + "bbox": [ + 0.77, + 0.612, + 0.89, + 0.68 + ], + "angle": 0, + "content": "Doraemon is standing in front of the white cat with an excited expression. He is holding a bouquet of flowers, in front of the cat. The white cat, sitting on a small cushion, is looking at Doraemon with a curious or intrigued expression. The setting appears to be indoors, with a pink carpet, a scratching post in the background, and a framed picture on the wall." + }, + { + "type": "title", + "bbox": [ + 0.107, + 0.713, + 0.124, + 0.767 + ], + "angle": 270, + "content": "Image-to-Text" + }, + { + "type": "image", + "bbox": [ + 0.129, + 0.681, + 0.255, + 0.79 + ], + "angle": 0, + "content": null + }, + { + "type": "text", + "bbox": [ + 0.129, + 0.794, + 0.248, + 0.842 + ], + "angle": 0, + "content": "Doraemon is standing on the shore, looking out at the sunset. He has a calm expression, and the sun is setting over the water, casting a warm glow. The sky is filled with clouds, and there are silhouettes of buildings in the distance." + }, + { + "type": "image_caption", + "bbox": [ + 0.31, + 0.681, + 0.326, + 0.691 + ], + "angle": 0, + "content": "(n)" + }, + { + "type": "image", + "bbox": [ + 0.257, + 0.693, + 0.383, + 0.789 + ], + "angle": 0, + "content": null + }, + { + "type": "text", + "bbox": [ + 0.258, + 0.794, + 0.377, + 0.841 + ], + "angle": 0, + "content": "Doraemon is standing in the grass, looking surprised as he sees a group of cats. The cats are gathered around him, with one of them appearing to be a cat with a surprised expression. The background shows a fence and some bushes." + }, + { + "type": "image", + "bbox": [ + 0.386, + 0.693, + 0.511, + 0.789 + ], + "angle": 0, + "content": null + }, + { + "type": "text", + "bbox": [ + 0.386, + 0.793, + 0.504, + 0.821 + ], + "angle": 0, + "content": "Doradamon is flying through the air, looking down at the city below. Nobita is also flying through the air, looking down at the city." + }, + { + "type": "image_caption", + "bbox": [ + 0.567, + 0.681, + 0.583, + 0.691 + ], + "angle": 0, + "content": "(p)" + }, + { + "type": "image", + "bbox": [ + 0.513, + 0.693, + 0.638, + 0.789 + ], + "angle": 0, + "content": null + }, + { + "type": "text", + "bbox": [ + 0.514, + 0.793, + 0.633, + 0.848 + ], + "angle": 0, + "content": "The image depicts a two-story house with a large garden in front of it. The house has a tan exterior and is surrounded by a variety of plants, including palm trees and bushes. The garden is well-maintained. The house features a balcony on the second floor, adding to its charm." + }, + { + "type": "image_caption", + "bbox": [ + 0.695, + 0.681, + 0.711, + 0.691 + ], + "angle": 0, + "content": "(q)" + }, + { + "type": "image", + "bbox": [ + 0.64, + 0.693, + 0.765, + 0.789 + ], + "angle": 0, + "content": null + }, + { + "type": "text", + "bbox": [ + 0.642, + 0.793, + 0.761, + 0.835 + ], + "angle": 0, + "content": "Nobita is standing in the doorway, holding a plate with a doughnut on it. He looks excited, and the doughnut is on a white plate. The door is open, and the room is visible in the background." + }, + { + "type": "image_caption", + "bbox": [ + 0.824, + 0.682, + 0.838, + 0.691 + ], + "angle": 0, + "content": "(r)" + }, + { + "type": "image", + "bbox": [ + 0.768, + 0.692, + 0.892, + 0.789 + ], + "angle": 0, + "content": null + }, + { + "type": "text", + "bbox": [ + 0.77, + 0.793, + 0.89, + 0.828 + ], + "angle": 0, + "content": "White cat and black cat are sitting on the grass, looking at each other with affection, with a blue sky in the background, and a few clouds scattered across the sky." + }, + { + "type": "image_caption", + "bbox": [ + 0.182, + 0.848, + 0.197, + 0.858 + ], + "angle": 0, + "content": "(s)" + }, + { + "type": "image_caption", + "bbox": [ + 0.31, + 0.848, + 0.325, + 0.858 + ], + "angle": 0, + "content": "(t)" + }, + { + "type": "image_caption", + "bbox": [ + 0.438, + 0.848, + 0.454, + 0.858 + ], + "angle": 0, + "content": "(u)" + }, + { + "type": "image_caption", + "bbox": [ + 0.567, + 0.848, + 0.583, + 0.859 + ], + "angle": 0, + "content": "(v)" + }, + { + "type": "image_caption", + "bbox": [ + 0.693, + 0.848, + 0.712, + 0.858 + ], + "angle": 0, + "content": "(w)" + }, + { + "type": "image_caption", + "bbox": [ + 0.824, + 0.848, + 0.839, + 0.858 + ], + "angle": 0, + "content": "(x)" + }, + { + "type": "image_caption", + "bbox": [ + 0.092, + 0.876, + 0.903, + 0.891 + ], + "angle": 0, + "content": "Figure 4. Image-to-text and text-to-image generation by the unified models that adapted for two domains. The special tokens are omitted." + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.516, + 0.957 + ], + "angle": 0, + "content": "2840" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.095, + 0.089, + 0.481, + 0.325 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.09, + 0.337, + 0.483, + 0.407 + ], + "angle": 0, + "content": "Figure 5. Effect of special tokens on character learning. (a) Base model without training. (b) Model trained without using special tokens, showing attribute confusion among characters. (c) Model trained with special tokens, improving character attribute alignment and reducing confusion." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.44, + 0.484, + 0.531 + ], + "angle": 0, + "content": "data to comprehensively learn general aspects of the target domain. Specifically, in each batch, for data with paired ground truth, we compute the token prediction loss and also include it in the cycle, use ground truth as the pseudo middle generations, and compute the cycle loss. For unpaired data, we compute the unpaired cycle loss." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.533, + 0.484, + 0.684 + ], + "angle": 0, + "content": "For example, when adapting the model to Domain 1: Black Myth Wukong and Domain 2: Doraemon, we annotate 1-3 images per unique identity with captions that specify the name of the identity. For each domain, we collect \\(2\\mathrm{k}\\) images, which are mostly sampled from online videos, and independently collect text descriptions, which are further expanded to \\(1\\mathrm{k}\\) by ChatGPT [39]. The final adapted model trained with DoraCycle demonstrates strong performance in both text-to-image generation and image-to-text generation, as shown in Fig. 4." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.687, + 0.483, + 0.882 + ], + "angle": 0, + "content": "In terms of text-to-image results, the model trained with DoraCycle effectively generated images that aligned well with the target domains. In domain 1 (Black Myth Wukong), the generated images accurately depicted domain-specific visual elements, such as the intricate details of character appearances and the overall fantasy-like atmosphere. This indicates that the model successfully learned to generalize the visual features from text prompts to realistic images within the target domain. Similarly, in domain 2 (Doraemon), the generated images preserve the iconic cartoonish aesthetics and capture key visual details of the characters and settings, demonstrating effective domain adaptation." + }, + { + "type": "text", + "bbox": [ + 0.11, + 0.886, + 0.483, + 0.901 + ], + "angle": 0, + "content": "For the image-to-text task, the model performs well in" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.092, + 0.905, + 0.227 + ], + "angle": 0, + "content": "generating contextually accurate captions. In domain 1, the generated captions provide rich descriptions of the characters, their attributes, and the context, effectively mirroring the visual elements present in the input images. In domain 2, the captions correctly describe the characters, their actions, and their environments concisely, maintaining consistency with the visual style. The ability of the model to generate accurate descriptions highlights its robust understanding of the visual components of the domain." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.229, + 0.907, + 0.515 + ], + "angle": 0, + "content": "Additionally, an interesting phenomenon can be observed in how the model handles the visual elements that are not annotated with paired data. For instance, in Fig. 4 (w), the dorayaki (a type of sweet bean-filled pancake) was described by the model as a \"doughnut\". This may be due to the fact that the anime-style representation of the dorayaki is novel, and neither the base model nor the unpaired training provided specific textual-visual pairing knowledge about it. On the other hand, in the example shown in Fig. 4 (x), we annotate the white cat as a character with paired textual and visual data, using a special token for its name: \" white cat black cat \" during the caption generation. This suggests that the model autonomously categorized the black cat as a character when learning the target domain, indicating that it may have attempted to generalize learned knowledge from one type of entity to similar ones." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.517, + 0.909, + 0.773 + ], + "angle": 0, + "content": "Enhanced Learning with Special Tokens: As shown in Fig. 5, we experimentally find that the model often confused multiple novel concepts in the target domain. Fig. 5 (a) shows the image generated by the base model without training taking in the name of characters. Fig. 5 (b) shows the characters generated by the trained model. During training, the names of characters are directly included in the text without special treatment, leading to attribute confusion between characters. The varying lengths of the tokenized character names also make learning difficult. To solve this problem, we introduce a simple yet efficient solution: adding special tokens around character names. We introduced start of character \\((\\langle \\text{soc}\\rangle)\\) and end of character \\((\\langle \\text{soc}\\rangle)\\) tokens to enclose character names, which significantly enhance the learning of novel concepts. As shown in Fig. 5 (c), involving special tokens improves the alignments between characters and their names." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.787, + 0.655, + 0.803 + ], + "angle": 0, + "content": "4.3. Comparisons" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.81, + 0.907, + 0.903 + ], + "angle": 0, + "content": "In this section, we use the Storyboard20K [66] dataset to conduct the quantitative comparison experiments. The storyboards originating from the same data source are grouped to form a domain, consisting of images and descriptive text. The data are used under three different settings, i.e. totally unpaired, only paired, and paired plus unpaired data," + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.514, + 0.957 + ], + "angle": 0, + "content": "2841" + } + ], + [ + { + "type": "table_caption", + "bbox": [ + 0.091, + 0.09, + 0.908, + 0.121 + ], + "angle": 0, + "content": "Table 1. Comparison of different training methods under various data settings. The best value is highlighted in blue , and the second-best value is highlighted in green . “P” indicates paired data, and “U” indicates unpaired data." + }, + { + "type": "table", + "bbox": [ + 0.095, + 0.133, + 0.907, + 0.293 + ], + "angle": 0, + "content": "
T cycleI cycleT DataI DataFID-1K ↓CIDEr ↑Human Eval
T2I Align ↑I2T Align ↑
DreamBooth [48]--10% P10% P33.2232.743.251.83
--100% P100% P24.9341.554.133.96
DoraCycleX✓X100% U28.9330.543.381.62
✓X100% UX36.6335.703.262.17
✓✓100% U100% U27.4438.173.843.42
✓✓10% P + 90% U10% P + 90% U25.3740.904.123.81
ITIT [34]✓✓10% P + 90% U10% P + 90% U27.5038.623.853.52
" + }, + { + "type": "text", + "bbox": [ + 0.091, + 0.318, + 0.232, + 0.332 + ], + "angle": 0, + "content": "as shown in Table 1." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.335, + 0.483, + 0.44 + ], + "angle": 0, + "content": "The compared methods include DreamBooth [48] and ITIT [34]. We implement DreamBooth as a paired-training baseline by applying LoRA fine-tuning on the unified model. The original design of ITIT is different, in which the image and text decoders are separate models, and its code has not been released. We adjusted and re-implemented it to be suitable for our unified model architecture." + }, + { + "type": "text", + "bbox": [ + 0.089, + 0.443, + 0.483, + 0.61 + ], + "angle": 0, + "content": "We use both automatic and human evaluations to compare the performance of different methods. For automatic evaluation, we use FID to measure the distribution differences between the generated images and the target domain images [22], and CIDEr to compute the error between the generated text and the ground truth [59]. For human evaluation, we create 100 questions for the generated results of models, each rated by three different human raters. The raters are asked to evaluate the alignment between the image and text on a scale from 1 to 5, where 1 indicates no relevance and 5 indicates complete alignment." + }, + { + "type": "text", + "bbox": [ + 0.089, + 0.612, + 0.483, + 0.778 + ], + "angle": 0, + "content": "The experimental results in Table 1 demonstrate that the proposed DoraCycle performs competitively under several data settings. Specifically, when using a combination of paired and unpaired data, DoraCycle outperforms ITIT. Compared to DreamBooth, which heavily relies on paired data, DoraCycle outperforms it when using the same scale of paired data, i.e. \\(10\\%\\) paired data, indicating the benefits brought by \\(90\\%\\) unpaired data. While Dreambooth with \\(100\\%\\) achieves the best evaluation scores, the scores of the DoraCycle with \\(10\\%\\) paired and \\(90\\%\\) unpaired data are comparable with them." + }, + { + "type": "text", + "bbox": [ + 0.089, + 0.78, + 0.483, + 0.901 + ], + "angle": 0, + "content": "Table 1 also shows the difference in the performance of DoraCyle under different cycle settings. It is shown that without the T cycle and with only the I cycle, the captioning ability of the adapted model degrades more significantly. In contrast, if only the T cycle is used and without the I cycle, the FID score increases substantially, indicating that the generated image distribution mismatches with the target distribution." + }, + { + "type": "title", + "bbox": [ + 0.514, + 0.317, + 0.678, + 0.332 + ], + "angle": 0, + "content": "4.4. Ablation Studies" + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.347, + 0.669, + 0.483 + ], + "angle": 0, + "content": "Table 2 shows that removing key components from DoraCycle significantly impacts performance. Without EMA, the FID score increases from 25.37 to 27.19, indicating lower image quality" + }, + { + "type": "text", + "bbox": [ + 0.697, + 0.347, + 0.906, + 0.404 + ], + "angle": 0, + "content": "Table 2. Ablation Studies. EMA refers to the exponential moving average. GS refers to gradient surgery." + }, + { + "type": "table", + "bbox": [ + 0.702, + 0.414, + 0.905, + 0.476 + ], + "angle": 0, + "content": "
FID-1K ↓CIDEr ↑
w/o EMA27.1938.85
w/o GS25.5439.98
DoraCycle25.3740.90
" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.484, + 0.907, + 0.619 + ], + "angle": 0, + "content": "due to less stable training. Removing Gradient Surgery (GS) will reduce the CIDEr score and increase the FID, indicating a worse performance. This demonstrates the importance of mitigating the interference between the optimization directions of two cycles. The complete DoraCycle framework, with both EMA and GS, has the best performance across all metrics, demonstrating the importance of these components in achieving better optimization." + }, + { + "type": "title", + "bbox": [ + 0.514, + 0.656, + 0.634, + 0.672 + ], + "angle": 0, + "content": "5. Conclusion" + }, + { + "type": "text", + "bbox": [ + 0.511, + 0.69, + 0.907, + 0.903 + ], + "angle": 0, + "content": "We propose the DoraCycle to adapt the unified generative model to target domains within multimodal cycles. By leveraging both image-to-text-to-image and text-to-image-to-text cycles, DoraCycle changes the learning objectives into the same modality, allowing for effective optimization using unpaired data. Our experiments show that DoraCycle can adapt the unified model to target domains using only unpaired data, or involving a small amount of paired data when necessary to learn specific concepts. Experimental results demonstrate that DoraCycle achieves advanced or comparable performance across various settings. Leveraging unpaired data broadens the application potential of DoraCycle, making it ideally suited for domain adaptation tasks where paired data is scarce or challenging to collect." + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.516, + 0.957 + ], + "angle": 0, + "content": "2842" + } + ], + [ + { + "type": "title", + "bbox": [ + 0.092, + 0.091, + 0.27, + 0.108 + ], + "angle": 0, + "content": "6. Acknowledgments" + }, + { + "type": "text", + "bbox": [ + 0.091, + 0.116, + 0.484, + 0.161 + ], + "angle": 0, + "content": "This research is supported by the National Research Foundation, Singapore under its AI Singapore Programme (AISG Award No: AISG3-RP-2022-030)." + }, + { + "type": "title", + "bbox": [ + 0.093, + 0.178, + 0.188, + 0.194 + ], + "angle": 0, + "content": "References" + }, + { + "type": "ref_text", + "bbox": [ + 0.101, + 0.203, + 0.484, + 0.271 + ], + "angle": 0, + "content": "[1] Armen Aghajanyan, Bernie Huang, Candace Ross, Vladimir Karpukhin, Hu Xu, Naman Goyal, Dmytro Okhonko, Mandar Joshi, Gargi Ghosh, Mike Lewis, et al. Cm3: A causal masked multimodal model of the internet. arXiv preprint arXiv:2201.07520, 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.1, + 0.274, + 0.484, + 0.315 + ], + "angle": 0, + "content": "[2] Emanuele Aiello, LILI YU, Yixin Nie, Armen Aghajanyan, and Barlas Oguz. Jointly training large autoregressive multimodal models. In ICLR, 2024. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.101, + 0.317, + 0.484, + 0.386 + ], + "angle": 0, + "content": "[3] Yogesh Balaji, Seungjun Nah, Xun Huang, Arash Vahdat, Jiaming Song, Karsten Kreis, Miika Aittala, Timo Aila, Samuli Laine, Bryan Catanzaro, et al. ediffi: Text-to-image diffusion models with an ensemble of expert denoisers. arXiv preprint arXiv:2211.01324, 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.389, + 0.484, + 0.457 + ], + "angle": 0, + "content": "[4] Soravit Changpinyo, Piyush Sharma, Nan Ding, and Radu Soricut. Conceptual 12m: Pushing web-scale image-text pretraining to recognize long-tail visual concepts. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 3558-3568, 2021. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.46, + 0.484, + 0.529 + ], + "angle": 0, + "content": "[5] Junsong Chen, Chongjian Ge, Enze Xie, Yue Wu, Lewei Yao, Xiaozhe Ren, Zhongdao Wang, Ping Luo, Huchuan Lu, and Zhenguo Li. Pixart\\(\\backslash\\)sigma: Weak-to-strong training of diffusion transformer for 4k text-to-image generation. arXiv preprint arXiv:2403.04692, 2024. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.532, + 0.484, + 0.6 + ], + "angle": 0, + "content": "[6] Junsong Chen, Jincheng Yu, Chongjian Ge, Lewei Yao, Enze Xie, Zhongdao Wang, James T. Kwok, Ping Luo, Huchuan Lu, and Zhenguo Li. Pixart-α: Fast training of diffusion transformer for photorealistic text-to-image synthesis. In ICLR. OpenReview.net, 2024. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.602, + 0.484, + 0.671 + ], + "angle": 0, + "content": "[7] Li Chen, Mengyi Zhao, Yiheng Liu, Mingxu Ding, Yangyang Song, Shizun Wang, Xu Wang, Hao Yang, Jing Liu, Kang Du, et al. Photovsere: Tuning-free image customization with text-to-image diffusion models. arXiv preprint arXiv:2309.05793, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.673, + 0.484, + 0.715 + ], + "angle": 0, + "content": "[8] Mark Chen, Alec Radford, Rewon Child, Jeffrey Wu, Heewoo Jun, David Luan, and Ilya Sutskever. Generative pretraining from pixels. In ICML, pages 1691-1703, 2020. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.717, + 0.484, + 0.785 + ], + "angle": 0, + "content": "[9] Xi Chen, Lianghua Huang, Yu Liu, Yujun Shen, Deli Zhao, and Hengshuang Zhao. Anydoor: Zero-shot object-level image customization. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 6593-6602, 2024. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.788, + 0.484, + 0.857 + ], + "angle": 0, + "content": "[10] Yunjey Choi, Minje Choi, Munyoung Kim, Jung-Woo Ha, Sunghun Kim, and Jaegul Choo. Stargan: Unified generative adversarial networks for multi-domain image-to-image translation. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition (CVPR), 2018. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.859, + 0.484, + 0.901 + ], + "angle": 0, + "content": "[11] Wenliang Dai, Junnan Li, Dongxu Li, Anthony Meng Huat Tiong, Junqi Zhao, Weisheng Wang, Boyang Li, Pascale Fung, and Steven Hoi. Instructlip: Towards general-" + }, + { + "type": "list", + "bbox": [ + 0.094, + 0.203, + 0.484, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.545, + 0.093, + 0.905, + 0.119 + ], + "angle": 0, + "content": "purpose vision-language models with instruction tuning, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.122, + 0.905, + 0.191 + ], + "angle": 0, + "content": "[12] Xiaoliang Dai, Ji Hou, Chih-Yao Ma, Sam Tsai, Jialiang Wang, Rui Wang, Peizhao Zhang, Simon Vandenhende, Xiaofang Wang, Abhimanyu Dubey, et al. Emu: Enhancing image generation models using photogenic needles in a haystack. arXiv preprint arXiv:2309.15807, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.193, + 0.905, + 0.234 + ], + "angle": 0, + "content": "[13] Prafulla Dhariwal and Alexander Nichol. Diffusion models beat gans on image synthesis. Advances in Neural Information Processing Systems, 34:8780-8794, 2021. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.236, + 0.905, + 0.305 + ], + "angle": 0, + "content": "[14] Runpei Dong, Chunrui Han, Yuang Peng, Zekun Qi, Zheng Ge, Jinrong Yang, Liang Zhao, Jianjian Sun, Hongyu Zhou, Haoran Wei, Xiangwen Kong, Xiangyu Zhang, Kaisheng Ma, and Li Yi. DreamLLM: Synergistic multimodal comprehension and creation. In ICLR, 2024. 1, 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.307, + 0.905, + 0.376 + ], + "angle": 0, + "content": "[15] Debidatta Dwibedi, Yusuf Aytar, Jonathan Tompson, Pierre Sermanet, and Andrew Zisserman. Temporal cycle-consistency learning. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 1801-1810, 2019. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.379, + 0.905, + 0.447 + ], + "angle": 0, + "content": "[16] Patrick Esser, Sumith Kulal, Andreas Blattmann, Rahim Entezari, Jonas Müller, Harry Saini, Yam Levi, Dominik Lorenz, Axel Sauer, Frederic Boesel, et al. Scaling rectified flow transformers for high-resolution image synthesis. arXiv preprint arXiv:2403.03206, 2024. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.449, + 0.905, + 0.517 + ], + "angle": 0, + "content": "[17] Rinon Gal, Yuval Alaluf, Yuval Atzmon, Or Patashnik, Amit H Bermano, Gal Chechik, and Daniel Cohen-Or. An image is worth one word: Personalizing text-to-image generation using textual inversion. arXiv preprint arXiv:2208.01618, 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.52, + 0.905, + 0.588 + ], + "angle": 0, + "content": "[18] Yuying Ge, Sijie Zhao, Jinguo Zhu, Yixiao Ge, Kun Yi, Lin Song, Chen Li, Xiaohan Ding, and Ying Shan. Seed-x: Multimodal models with unified multi-granularity comprehension and generation. arXiv preprint arXiv:2404.14396, 2024. 1, 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.591, + 0.905, + 0.673 + ], + "angle": 0, + "content": "[19] Yuchao Gu, Xintao Wang, Jay Zhangjie Wu, Yujun Shi, Yunpeng Chen, Zihan Fan, Wuyou Xiao, Rui Zhao, Shuning Chang, Weijia Wu, et al. Mix-of-show: Decentralized low-rank adaptation for multi-concept customization of diffusion models. Advances in Neural Information Processing Systems, 36:15890-15902, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.676, + 0.905, + 0.758 + ], + "angle": 0, + "content": "[20] Jiayi Guo, Chaofei Wang, You Wu, Eric Zhang, Kai Wang, Xingqian Xu, Shiji Song, Humphrey Shi, and Gao Huang. Zero-shot generative model adaptation via image-specific prompt learning. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 11494-11503, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.761, + 0.905, + 0.815 + ], + "angle": 0, + "content": "[21] Di He, Yingce Xia, Tao Qin, Liwei Wang, Nenghai Yu, TieYan Liu, and Wei-Ying Ma. Dual learning for machine translation. In Advances in Neural Information Processing Systems (NeurIPS), 2016. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.818, + 0.905, + 0.899 + ], + "angle": 0, + "content": "[22] Martin Heusel, Hubert Ramsauer, Thomas Unterthiner, Bernhard Nessler, and Sepp Hochreiter. Gans trained by a two time-scale update rule converge to a local nash equilibrium. In Proceedings of the 31st International Conference on Neural Information Processing Systems, pages 6626-6637, 2017. 8" + }, + { + "type": "list", + "bbox": [ + 0.517, + 0.093, + 0.905, + 0.899 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.516, + 0.957 + ], + "angle": 0, + "content": "2843" + } + ], + [ + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.092, + 0.482, + 0.161 + ], + "angle": 0, + "content": "[23] Judy Hoffman, Eric Tzeng, Taesung Park, Jun-Yan Zhu, Phillip Isola, Kate Saenko, Alexei A Efros, and Trevor Darryll. Cycada: Cycle-consistent adversarial domain adaptation. In Proceedings of the International Conference on Machine Learning (ICML), 2018. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.162, + 0.482, + 0.218 + ], + "angle": 0, + "content": "[24] MD Zakir Hossain, Ferdous Sohel, Mohd Fairuz Shiratuddin, and Hamid Laga. A comprehensive survey of deep learning for image captioning. ACM Computing Surveys (CsUR), 51(6):1-36, 2019. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.218, + 0.482, + 0.273 + ], + "angle": 0, + "content": "[25] Edward J. Hu, Yelong Shen, Phillip Wallis, Zeyuan Allen-Zhu, Yuanzhi Li, Shean Wang, and Weizhu Chen. Lora: Low-rank adaptation of large language models. In International Conference on Learning Representations, 2022. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.274, + 0.482, + 0.342 + ], + "angle": 0, + "content": "[26] Jia Cheng Hu, Roberto Cavicchioli, and Alessandro Capontodi. Exploiting multiple sequence lengths in fast end to end training for image captioning. In 2023 IEEE International Conference on Big Data (BigData), pages 2173-2182. IEEE, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.344, + 0.482, + 0.412 + ], + "angle": 0, + "content": "[27] Xiaowei Hu, Zhe Gan, Jianfeng Wang, Zhengyuan Yang, Zicheng Liu, Yumao Lu, and Lijuan Wang. Scaling up vision-language pre-training for image captioning. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 17980-17989, 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.413, + 0.482, + 0.453 + ], + "angle": 0, + "content": "[28] Eric Jang, Shixiang Gu, and Ben Poole. Categorical reparameterization with gumbel-softmax. arXiv preprint arXiv:1611.01144, 2016. 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.455, + 0.482, + 0.524 + ], + "angle": 0, + "content": "[29] Xuhui Jia, Yang Zhao, Kelvin CK Chan, Yandong Li, Han Zhang, Boqing Gong, Tingbo Hou, Huisheng Wang, and Yu-Chuan Su. Taming encoder for zero fine-tuning image customization with text-to-image diffusion models. arXiv preprint arXiv:2304.02642, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.525, + 0.482, + 0.594 + ], + "angle": 0, + "content": "[30] Nupur Kumari, Bingliang Zhang, Richard Zhang, Eli Shechtman, and Jun-Yan Zhu. Multi-concept customization of text-to-image diffusion. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 1931-1941, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.595, + 0.482, + 0.663 + ], + "angle": 0, + "content": "[31] Chenliang Li, Haiyang Xu, Junfeng Tian, Wei Wang, Ming Yan, Bin Bi, Jiabo Ye, Hehong Chen, Guohai Xu, Zheng Cao, et al. mplug: Effective and efficient vision-language learning by cross-modal skip-connections. arXiv preprint arXiv:2205.12005, 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.664, + 0.482, + 0.733 + ], + "angle": 0, + "content": "[32] Chunyuan Li, Zhe Gan, Zhengyuan Yang, Jianwei Yang, Linjie Li, Lijuan Wang, Jianfeng Gao, et al. Multimodal foundation models: From specialists to general-purpose assistants. Foundations and Trends® in Computer Graphics and Vision, 16(1-2):1-214, 2024. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.734, + 0.482, + 0.802 + ], + "angle": 0, + "content": "[33] Junnan Li, Dongxu Li, Silvio Savarese, and Steven Hoi. Blip-2: Bootstrapping language-image pre-training with frozen image encoders and large language models. In International conference on machine learning, pages 19730-19742. PMLR, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.803, + 0.482, + 0.872 + ], + "angle": 0, + "content": "[34] Tianhong Li, Sangnie Bhardwaj, Yonglong Tian, Han Zhang, Jarred Barber, Dina Katabi, Guillaume Lajoie, Huiwen Chang, and Dilip Krishnan. Leveraging unpaired data for vision-language generative models via cycle consistency. arXiv preprint arXiv:2310.03734, 2023. 3, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.873, + 0.482, + 0.901 + ], + "angle": 0, + "content": "[35] Haotian Liu, Chunyuan Li, Qingyang Wu, and Yong Jae Lee. Visual instruction tuning. NeurIPS, 36, 2024. 2" + }, + { + "type": "list", + "bbox": [ + 0.093, + 0.092, + 0.482, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.092, + 0.905, + 0.16 + ], + "angle": 0, + "content": "[36] Nanye Ma, Mark Goldstein, Michael S Albergo, Nicholas M Boffi, Eric Vanden-Eijnden, and Saining Xie. Sit: Exploring flow and diffusion-based generative models with scalable interpolant transformers. arXiv preprint arXiv:2401.08740, 2024. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.162, + 0.905, + 0.219 + ], + "angle": 0, + "content": "[37] Xin Ma, Yaohui Wang, Gengyun Jia, Xinyuan Chen, Zwei Liu, Yuan-Fang Li, Cunjian Chen, and Yu Qiao. Latte: Latent diffusion transformer for video generation. arXiv preprint arXiv:2401.03048, 2024. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.22, + 0.907, + 0.29 + ], + "angle": 0, + "content": "[38] Ron Mokady, Amir Hertz, Kfir Aberman, Yael Pritch, and Daniel Cohen-Or. Null-text inversion for editing real images using guided diffusion models. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 6038–6047, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.291, + 0.892, + 0.305 + ], + "angle": 0, + "content": "[39] OpenAI. Chatgpt. https://chatgpt.com/, 2024.7" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.306, + 0.905, + 0.388 + ], + "angle": 0, + "content": "[40] Alec Radford, Jong Wook Kim, Chris Hallacy, Aditya Ramesh, Gabriel Goh, Sandhini Agarwal, Girish Sastry, Amanda Askell, Pamela Mishkin, Jack Clark, et al. Learning transferable visual models from natural language supervision. In International conference on machine learning, pages 8748-8763. PMLR, 2021. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.39, + 0.905, + 0.446 + ], + "angle": 0, + "content": "[41] Aditya Ramesh, Mikhail Pavlov, Gabriel Goh, Scott Gray, Chelsea Voss, Alec Radford, Mark Chen, and Ilya Sutskever. Zero-shot text-to-image generation. In ICML, pages 8821-8831. Pmlr, 2021. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.447, + 0.904, + 0.489 + ], + "angle": 0, + "content": "[42] Aditya Ramesh, Prafulla Dhariwal, Alex Nichol, Casey Chu, and Mark Chen. Hierarchical text-conditional image generation with CLIP latents. CoRR, abs/2204.06125, 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.49, + 0.905, + 0.544 + ], + "angle": 0, + "content": "[43] Aditya Ramesh, Prafulla Dhariwal, Alex Nichol, Casey Chu, and Mark Chen. Hierarchical diffusion models for text-to-image generation. arXiv preprint arXiv:2204.06125, 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.547, + 0.905, + 0.603 + ], + "angle": 0, + "content": "[44] Aditya Ramesh, Prafulla Dhariwal, Alex Nichol, Casey Chu, and Mark Chen. Hierarchical text-conditional image generation with clip latents. arXiv preprint arXiv:2204.06125, 1 (2):3, 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.604, + 0.905, + 0.659 + ], + "angle": 0, + "content": "[45] Robin Rombach, Andreas Blattmann, Dominik Lorenz, Patrick Esser, and Björn Ommer. High-resolution image synthesis with latent diffusion models. In CVPR, pages 10684-10695, 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.661, + 0.905, + 0.716 + ], + "angle": 0, + "content": "[46] Robin Rombach, Andreas Blattmann, Dominik Lorenz, Patrick Esser, and Björn Ommer. High-resolution image synthesis with latent diffusion models. arXiv preprint arXiv:2112.10752, 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.718, + 0.905, + 0.773 + ], + "angle": 0, + "content": "[47] Nataniel Ruiz, Yuanzhen Li, Varun Jampani, Yael Pritch, Michael Rubinstein, and Kfir Aberman. Dreambooth: Fine tuning text-to-image diffusion models for subject-driven generation. arXiv preprint arXiv:2208.12242, 2022. 1, 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.775, + 0.905, + 0.857 + ], + "angle": 0, + "content": "[48] Nataniel Ruiz, Yuanzhen Li, Varun Jampani, Yael Pritch, Michael Rubinstein, and Kfir Aberman. Dreambooth: Fine tuning text-to-image diffusion models for subject-driven generation. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 22500-22510, 2023. 5, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.858, + 0.905, + 0.901 + ], + "angle": 0, + "content": "[49] Chitwan Sahara, William Chan, Saurabh Saxena, Lala Li, Jay Whang, Emily L Denton, Kamyar Ghasemipour, Raphael Gontijo Lopes, Burcu Karagol Ayan, Tim Salimans," + }, + { + "type": "list", + "bbox": [ + 0.517, + 0.092, + 0.907, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.516, + 0.956 + ], + "angle": 0, + "content": "2844" + } + ], + [ + { + "type": "ref_text", + "bbox": [ + 0.125, + 0.092, + 0.482, + 0.134 + ], + "angle": 0, + "content": "et al. Photorealistic text-to-image diffusion models with deep language understanding. Advances in Neural Information Processing Systems, 35:36479-36494, 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.135, + 0.483, + 0.205 + ], + "angle": 0, + "content": "[50] Rico Sennrich, Barry Haddow, and Alexandra Birch. Improving neural machine translation models with monolingual data. In Proceedings of the 54th Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers), 2016. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.205, + 0.482, + 0.26 + ], + "angle": 0, + "content": "[51] Meet Shah, Xinlei Chen, Marcus Rohrbach, and Devi Parikh. Cycle-consistency for robust visual question answering. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 6649-6658, 2019. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.261, + 0.482, + 0.316 + ], + "angle": 0, + "content": "[52] Sameer Shah, Siddharth Bharadwaj, Devi Parikh, and Dhruv Batra. Cycle-consistency for robust visual question answering. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition (CVPR), 2019. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.316, + 0.482, + 0.399 + ], + "angle": 0, + "content": "[53] Quan Sun, Yufeng Cui, Xiaosong Zhang, Fan Zhang, Qiying Yu, Yueze Wang, Yongming Rao, Jingjing Liu, Tiejun Huang, and Xinlong Wang. Generative multimodal models are in-context learners. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 14398-14409, 2024. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.4, + 0.482, + 0.427 + ], + "angle": 0, + "content": "[54] I Sutskever. Sequence to sequence learning with neural networks. arXiv preprint arXiv:1409.3215, 2014. 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.428, + 0.482, + 0.455 + ], + "angle": 0, + "content": "[55] Richard S Sutton. Learning to predict by the methods of temporal differences. Machine learning, 3:9-44, 1988. 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.456, + 0.482, + 0.496 + ], + "angle": 0, + "content": "[56] Zineng Tang, Ziyi Yang, Chenguang Zhu, Michael Zeng, and Mohit Bansal. Any-to-any generation via composable diffusion. NeurIPS, 36, 2024. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.497, + 0.482, + 0.564 + ], + "angle": 0, + "content": "[57] Antti Tarvainen and Harri Valpola. Mean teachers are better role models: Weight-averaged consistency targets improve semi-supervised deep learning results. In Advances in Neural Information Processing Systems, pages 1195–1204, 2017. 2, 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.567, + 0.482, + 0.607 + ], + "angle": 0, + "content": "[58] Chameleon Team. Chameleon: Mixed-modal early-fusion foundation models. arXiv preprint arXiv:2405.09818, 2024. 1, 2, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.609, + 0.482, + 0.676 + ], + "angle": 0, + "content": "[59] Ramakrishna Vedantam, C. Lawrence Zitnick, and Devi Parikh. Cider: Consensus-based image description evaluation. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition (CVPR), pages 4566-4575, 2015. 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.679, + 0.482, + 0.734 + ], + "angle": 0, + "content": "[60] Jianfeng Wang, Zhengyuan Yang, Xiaowei Hu, Linjie Li, Kevin Lin, Zhe Gan, Zicheng Liu, Ce Liu, and Lijuan Wang. Git: A generative image-to-text transformer for vision and language. arXiv preprint arXiv:2205.14100, 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.735, + 0.482, + 0.817 + ], + "angle": 0, + "content": "[61] Peng Wang, An Yang, Rui Men, Junyang Lin, Shuai Bai, Zhikang Li, Jianxin Ma, Chang Zhou, Jingren Zhou, and Hongxia Yang. Ofa: Unifying architectures, tasks, and modalities through a simple sequence-to-sequence learning framework. In International conference on machine learning, pages 23318-23340. PMLR, 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.818, + 0.482, + 0.872 + ], + "angle": 0, + "content": "[62] Qixun Wang, Xu Bai, Haofan Wang, Zekui Qin, Anthony Chen, Huaxia Li, Xu Tang, and Yao Hu. Instantid: Zero-shot identity-preserving generation in seconds. arXiv preprint arXiv:2401.07519, 2024. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.873, + 0.482, + 0.901 + ], + "angle": 0, + "content": "[63] Yuchi Wang, Shuhuai Ren, Rundong Gao, Linli Yao, Qingyan Guo, Kaikai An, Jianhong Bai, and Xu Sun." + }, + { + "type": "list", + "bbox": [ + 0.093, + 0.092, + 0.483, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.545, + 0.093, + 0.905, + 0.134 + ], + "angle": 0, + "content": "Ladic: Are diffusion models really inferior to autoregressive counterparts for image-to-text generation? arXiv preprint arXiv:2404.10763, 2024. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.135, + 0.905, + 0.204 + ], + "angle": 0, + "content": "[64] Chengyue Wu, Xiaokang Chen, Zhiyu Wu, Yiyang Ma, Xingchao Liu, Zizheng Pan, Wen Liu, Zhenda Xie, Xingkai Yu, Chong Ruan, et al. Janus: Decoupling visual encoding for unified multimodal understanding and generation. arXiv preprint arXiv:2410.13848, 2024. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.205, + 0.905, + 0.245 + ], + "angle": 0, + "content": "[65] Shengqiong Wu, Hao Fei, Leigang Qu, Wei Ji, and Tat-Seng Chua. Next-gpt: Any-to-any multimodal llm. arXiv preprint arXiv:2309.05519, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.247, + 0.905, + 0.315 + ], + "angle": 0, + "content": "[66] Jinheng Xie, Jiajun Feng, Zhaoxu Tian, Kevin Qinghong Lin, Yawen Huang, Xi Xia, Nanxu Gong, Xu Zuo, Ji-aqi Yang, Yefeng Zheng, et al. Learning long-form video prior via generative pre-training. arXiv preprint arXiv:2404.15909, 2024. 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.316, + 0.905, + 0.398 + ], + "angle": 0, + "content": "[67] Jinheng Xie, Weijia Mao, Zechen Bai, David Junhao Zhang, Weihao Wang, Kevin Qinghong Lin, Yuchao Gu, Zhijie Chen, Zhenheng Yang, and Mike Zheng Shou. Show-o: One single transformer to unify multimodal understanding and generation. arXiv preprint arXiv:2408.12528, 2024. 1, 2, 3, 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.4, + 0.905, + 0.467 + ], + "angle": 0, + "content": "[68] Ceyuan Yang, Yujun Shen, Zhiyi Zhang, Yinghao Xu, Jiapeng Zhu, Zhirong Wu, and Bolei Zhou. One-shot generative domain adaptation. In Proceedings of the IEEE/cvf international conference on computer vision, pages 7733-7742, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.469, + 0.905, + 0.525 + ], + "angle": 0, + "content": "[69] Hanrong Ye, De-An Huang, Yao Lu, Zhiding Yu, Wei Ping, Andrew Tao, Jan Kautz, Song Han, Dan Xu, Pavlo Molchanov, et al. X-vila: Cross-modality alignment for large language model. arXiv preprint arXiv:2405.19335, 2024. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.526, + 0.905, + 0.58 + ], + "angle": 0, + "content": "[70] Haoxuan You, Mandy Guo, Zhecan Wang, Kai-Wei Chang, Jason Baldridge, and Jiahui Yu. Cobit: A contrastive bidirectional image-text generation model. arXiv preprint arXiv:2303.13455, 2023." + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.581, + 0.905, + 0.65 + ], + "angle": 0, + "content": "[71] Lili Yu, Bowen Shi, Ramakanth Pasunuru, Benjamin Muller, Olga Golovneva, Tianlu Wang, Arun Babu, Binh Tang, Brian Karrer, Shelly Sheynin, et al. Scaling autoregressive multimodal models: Pretraining and instruction tuning. arXiv preprint arXiv:2309.02591, 2(3), 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.651, + 0.905, + 0.706 + ], + "angle": 0, + "content": "[72] Tianhe Yu, Saurabh Kumar, Abhishek Gupta, Sergey Levine, Karol Hausman, and Chelsea Finn. Gradient surgery for multi-task learning. In Advances in Neural Information Processing Systems, pages 5824-5836, 2020. 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.707, + 0.905, + 0.761 + ], + "angle": 0, + "content": "[73] Gengwei Zhang, Guoliang Kang, Yi Yang, and Yunchao Wei. Few-shot segmentation via cycle-consistent transformer. Advances in Neural Information Processing Systems, 34:21984-21996, 2021. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.762, + 0.905, + 0.831 + ], + "angle": 0, + "content": "[74] Rui Zhao, Wei Li, Zhipeng Hu, Lincheng Li, Zhengxia Zou, Zhenwei Shi, and Changjie Fan. Zero-shot text-to-parameter translation for game character auto-creation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 21013–21023, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.832, + 0.905, + 0.901 + ], + "angle": 0, + "content": "[75] Rui Zhao, Yuchao Gu, Jay Zhangjie Wu, David Junhao Zhang, Jia-Wei Liu, Weijia Wu, Jussi Keppo, and Mike Zheng Shou. Motiondirector: Motion customization of text-to-video diffusion models. In European Conference on Computer Vision, pages 273–290. Springer, 2024." + }, + { + "type": "list", + "bbox": [ + 0.517, + 0.093, + 0.905, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.516, + 0.956 + ], + "angle": 0, + "content": "2845" + } + ], + [ + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.093, + 0.482, + 0.176 + ], + "angle": 0, + "content": "[76] Rui Zhao, Hangjie Yuan, Yujie Wei, Shiwei Zhang, Yuchao Gu, Lingmin Ran, Xiang Wang, Jay Zhangjie Wu, David Junhao Zhang, Yingya Zhang, et al. Evolvedirector: Approaching advanced text-to-image generation with large vision-language models. Advances in Neural Information Processing Systems, 37:122104-122129, 2025. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.178, + 0.482, + 0.247 + ], + "angle": 0, + "content": "[77] Chunting Zhou, Lili Yu, Arun Babu, Kushal Tirumala, Michihiro Yasunaga, Leonid Shamis, Jacob Kahn, Xuezhe Ma, Luke Zettlemoyer, and Omer Levy. Transfusion: Predict the next token and diffuse images with one multi-modal model. arXiv preprint arXiv:2408.11039, 2024. 1, 2, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.249, + 0.482, + 0.302 + ], + "angle": 0, + "content": "[78] Deyao Zhu, Jun Chen, Xiaoqian Shen, Xiang Li, and Mohamed Elhoseiny. Minigpt-4: Enhancing vision-language understanding with advanced large language models. CoRR, abs/2304.10592, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.305, + 0.482, + 0.371 + ], + "angle": 0, + "content": "[79] Jun-Yan Zhu, Taesung Park, Phillip Isola, and Alexei A. Efros. Unpaired image-to-image translation using cycle-consistent adversarial networks. In Proceedings of the IEEE International Conference on Computer Vision (ICCV), 2017. 3" + }, + { + "type": "list", + "bbox": [ + 0.093, + 0.093, + 0.482, + 0.371 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.946, + 0.517, + 0.956 + ], + "angle": 0, + "content": "2846" + } + ] +] \ No newline at end of file diff --git a/2025/DoraCycle_ Domain-Oriented Adaptation of Unified Generative Model in Multimodal Cycles/fb2a505a-21d1-4afa-86a3-162bb4136d23_origin.pdf b/2025/DoraCycle_ Domain-Oriented Adaptation of Unified Generative Model in Multimodal Cycles/fb2a505a-21d1-4afa-86a3-162bb4136d23_origin.pdf new file mode 100644 index 0000000000000000000000000000000000000000..0a8e8f1aacc34e02419ebafc6eb67c2963b8040a --- /dev/null +++ b/2025/DoraCycle_ Domain-Oriented Adaptation of Unified Generative Model in Multimodal Cycles/fb2a505a-21d1-4afa-86a3-162bb4136d23_origin.pdf @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:72c7e11bd859f90bbd362b30adcd51242b79711260770c574fcb28ed78be61e0 +size 1000505 diff --git a/2025/DoraCycle_ Domain-Oriented Adaptation of Unified Generative Model in Multimodal Cycles/full.md b/2025/DoraCycle_ Domain-Oriented Adaptation of Unified Generative Model in Multimodal Cycles/full.md new file mode 100644 index 0000000000000000000000000000000000000000..14489de0901a4a27bac9f6f970ae980658bcb804 --- /dev/null +++ b/2025/DoraCycle_ Domain-Oriented Adaptation of Unified Generative Model in Multimodal Cycles/full.md @@ -0,0 +1,448 @@ +# DoraCycle: Domain-Oriented Adaptation of Unified Generative Model in Multimodal Cycles + +Rui Zhao, Weijia Mao, Mike Zheng Shou* Show Lab, National University of Singapore + +# Abstract + +Adapting generative models to specific domains presents an effective solution for satisfying specialized requirements. However, adapting to some complex domains remains challenging, especially when these domains require substantial paired data to capture the targeted distributions. Since unpaired data from a single modality, such as vision or language, is more readily available, we utilize the bidirectional mappings between vision and language learned by the unified generative model to enable training on unpaired data for domain adaptation. Specifically, we propose DoraCycle, which integrates two multimodal cycles: text-to-image-to-text and image-to-text-to-image. The model is optimized through cross-entropy loss computed at the cycle endpoints, where both endpoints share the same modality. This facilitates self-evolution of the model without reliance on annotated text-image pairs. Experimental results demonstrate that for tasks independent of paired knowledge, such as stylization, DoraCycle can effectively adapt the unified model using only unpaired data. For tasks involving new paired knowledge, such as specific identities, a combination of a small set of paired image-text examples and larger-scale unpaired data is sufficient for effective domain-oriented adaptation. The code will be released at https://github.com/showlab/DoraCycle. + +# 1. Introduction + +The adaptation of pre-trained generative models to specific domains is an important aspect of advancing personalized content creation, from stylized media outputs to customized identity generation [47, 62]. However, effectively adapting generative models to complex domains remains challenging, particularly when these domains require extensive amounts of paired data to accurately capture the desired distributions. For instance, learning the visual styles and character identities across a unique movie, which involves understanding multiple characters, their relationships, and di + +![](images/d3956901487918a2e4270e7958c4ead203a6602070952914a9d1371d9b938ffd.jpg) +(a) Training on paired data +Figure 1. Training paradigms for unified generative models. (a) Traditional training involves using paired image-text data and optimizing the unified model with paired losses for both image-to-text (I2T) and text-to-image (T2I) generation tasks. (b) In contrast, the proposed multimodal cycle training framework leverages unpaired images and texts. By using cycle-consistency losses, the unified model learns to maintain consistency between input and output across modalities, enabling adaptation without the need for extensive paired datasets. + +![](images/6f1ddd95a07228c8cf7903e1ad20354fa39c4018ee5ae15b67d161cfd0657ef9.jpg) +(b) Training in Multimodal Cycles + +verse settings, is a highly complex task that demands vast amounts of paired frames and captions data. Collecting such paired data, especially for multimodal tasks involving vision and language, is often laborious and impractical, limiting the potential to adapt generative models at scale. + +High-quality image-text paired data is relatively rare and scarce, but unpaired images and texts are readily available in our daily lives, such as video websites, image platforms, and content from novel websites. Therefore, we aim to explore whether it is possible to adapt generative models to target domains based on unpaired data. To achieve this goal, it is crucial for the model to have an internal capability to align the two modalities to some extent. Fortunately, recent advancements in unified generative models have encouraged us to pursue this direction. + +Recent advanced unified generative models [14, 18, 58, 67, 77] have shown great potential in unifying multimodal + +understanding and generation within a single model. These unified models are capable of processing and generating content across different modalities, i.e. vision and language, within a shared framework. By leveraging the bidirectional mappings between vision and language, which are learned by the unified generative model in its pre-training stage, we can map each data from one modality to another and then back to the original modality, as shown in Fig. 1. Through these two mappings, data can be maintained within the same modality, thereby imposing constraints on the deviation introduced in the process. This only requires computing the cross-entropy loss between the data in the same modality, without any paired data supervision. + +To this end, we introduce DoraCycle, a framework for adapting unified multimodal generative models to target domains, through cycle-consistent learning with unpaired data. Unlike previous adaptation methods that heavily rely on paired text-image data, the proposed framework leverages the shared latent space of unified models to learn consistent transformations between modalities without requiring paired training examples. Specifically, we design two cycles, i.e. text-to-image-to-text (T cycle) and image-to-text-to-image (I cycle). As shown by Fig. 1 (b), leveraging the pre-trained vision-language alignment of the unified model, each multimodal cycle involves two cross-modality mappings while optimizing in the same modality. This enables calculating loss on unpaired data while implicitly refining cross-modal alignment through the intermediate step. + +In practice, since there is no labeled ground-truth, mapping data to another modality requires multi-step inference, such as predicting the next token multiple times for text generation. Allowing all forward steps to participate in gradient backpropagation can lead to a catastrophic gradient explosion. Therefore, we stop gradients during multi-step inference and use the generated data as pseudo labels for the model to forward once again, allowing gradients to propagate. Moreover, we found that since a complete cycle requires the model to forward twice, it can lead to training instability, with the quality of pseudo data generated in the middle being compromised. To enhance the stability of pseudo data generation, we maintain a slowly updated EMA (Exponential Moving Average) [57] model, which is used for inference to generate pseudo data. Additionally, we employ gradient clipping techniques to avoid conflicts in the optimization directions of the two cycles, further increasing training stability. + +The experiments indicate that for tasks independent of paired knowledge, such as stylization and domain-specific adaptation, DoraCycle can adapt the unified model with only unpaired data, which is both more practical and scalable. For tasks that require new paired knowledge, such as identity-specific adaptation, DoraCycle effectively utilizes small amounts of paired data along with larger-scale + +unpaired data, making it a flexible solution for generative adaptation challenges. We conduct extensive experiments that compare DoraCycle to existing methods, showing that our approach achieves comparable or superior results while significantly reducing the need for paired data. This ability to harness large-scale unpaired data, combined with strategic usage of small paired datasets, makes DoraCycle a feasible solution for personalized multimodal content generation across a wide range of applications. + +# 2. Related Works + +# 2.1. Multimodal Generation and Understanding + +Generating visual contents from text and describing them through natural language have been extensively studied as core multimodal tasks. Advanced generative models [3, 5, 6, 8, 12, 13, 16, 19, 36, 37, 42, 44, 45, 49, 74-76], such as DALL-E [41, 43], Stable Diffusion [46], demonstrate remarkable generation capabilities, producing high-quality and diverse contents from textual prompts. Meanwhile, image captioning models [24, 26, 27, 60, 61, 63], such as mPLUG [31], and BLIP [33], push the boundaries of visual understanding, generating accurate and context-aware descriptions. Additionally, recent advancements in multimodal large language models [32], such as LLaVA [35], MiniGPT-4 [78], and InstructBLIP [11], have significantly improved the ability to understand and reason about visual content. + +Besides the powerful foundational generative models, adapting or customizing them attracts increasing interest, which enables more personalized and specific outputs based on user preferences [7, 9, 17, 20, 29, 30, 38, 68]. Approaches like DreamBooth [47] enable user-specific customization by fine-tuning generative models with personal data, allowing the generation of content tailored to individual needs or preferences. + +# 2.2. Unified Multimodal Generative Models + +Unified multimodal generative models aim to bridge the gap between understanding and generation tasks, and integrate vision and language into a single framework, enabling the model to learn shared representations across modalities [1, 2, 14, 18, 53, 56, 64, 65, 67, 69-71, 77]. SEED-X [18] utilizes a unified architecture where visual features extracted from the CLIP ViT encoder [40] are combined with text tokens and fed into a large language model to enable both next-word prediction and image regression tasks. DreamLLM [14] extends the generative capability of large language models by combining multimodal inputs directly into LLMs. Chameleon [58] employs a discrete tokenization approach for both visual and textual inputs, converting all modalities into a unified token space that is processed by a transformer-based architecture. Transfusion [77] in + +produces an advanced integration mechanism that focuses on directly fusing visual encoding with language tokens, allowing the model to effectively translate visual information into textual formats while maintaining the semantic integrity of both modalities. Show-o [67] combines autoregressive modeling with a discrete diffusion process, enabling the generation of high-quality outputs that are aligned across modalities. Our work leverages the advancements made by these foundational models and explores how to adapt the foundational model to specific domains. + +# 2.3. Cycle Consistency + +Cycle consistency has been utilized in computer vision and natural language processing as a means to enhance model robustness and consistency [10, 15, 21, 23, 51, 51, 52, 73]. CycleGAN [79] introduced cycle consistency loss to align unpaired image domains, ensuring that mappings between domains (e.g., $\mathrm{A} \to \mathrm{B} \to \mathrm{A}$ ) remain consistent. In the field of natural language processing, back-translation employs similar ideas by translating sentences between languages in both directions to improve translation quality [50]. However, the cycle consistency in these works is in a single modality, i.e. vision or language. Recently, ITIT [34] was proposed to utilize cycle consistency to train vision-language generative models. ITIT takes in a mixture of unpaired data and paired data to pre-train the foundational generative model. It is constructed with one image-text encoder and two modality-specific decoders, which operate on the encoded image-text features to generate either text or image tokens. In contrast, we utilized a single unified transformer to parse and predict text and image tokens together. Besides, we focus on adapting pre-trained foundational models to new domains efficiently rather than retraining new foundation models. + +# 3. Method + +The proposed DoraCycle framework, as shown in Fig. 2, is built upon the unified generative model designed for multimodal tasks involving both vision and language [58, 67, 77]. The unified model uses a single transformer to learn bidirectional mappings between vision and language, providing a powerful backbone capable of processing and generating different modalities [67]. For captioning, the model takes in image tokens and predicts corresponding text tokens, while for image generation, it takes in text tokens and predicts image tokens. This versatility makes the unified model well-suited as a base for our proposed framework. + +In the following sections, we first introduce the design of multimodal cycles, and then discuss the stabilization of optimization, and the balance of two cycles. + +# 3.1. Multimodal Cycles + +To adapt the unified model for domain-oriented adaptations using unpaired data, we design two multimodal cycles: the Image-to-Text-to-Image Cycle (I Cycle) and the Text-to-Image-to-Text Cycle (T Cycle). Each cycle utilizes data from a single modality, allowing the model to adapt without relying on paired data. + +T Cycle: The T cycle training involves transforming an input textual sequence into an image representation and then back into a textual sequence, enforcing consistency between the generated and original text. Specifically, at each training iteration, we begin with an input text sequence $T = [t_l]_{l=1}^L$ . Conditioned on this, the unified model generates pseudo-paired image tokens $I'$ , representing the visual interpretation of the input text. The generated image tokens are then subjected to a random masking operation, denoted as $I_M'$ , where a subset of the tokens is masked. The unified model is then called to reconstruct these masked tokens to form the complete synthesized image $\tilde{I}$ . + +In the second half of the cycle, conditioning on image $\tilde{I}$ , the model generates the corresponding text sequence. The objective of the T cycle is to enforce cycle consistency between the generated text and the original input text $T$ . The T cycle consistency loss is defined as follows: + +$$ +\mathcal {L} _ {T C} = - \mathbb {E} _ {T \in \mathcal {D} t e x t} \left[ \sum_ {l = 1} ^ {L} \log p \left(t _ {l} \mid \tilde {I}, t _ {0}, \dots , t _ {l - 1}\right) \right], \tag {1} +$$ + +where the $\mathcal{D}$ text is the set of text samples from the target domain. + +I Cycle: The I cycle training begins with an input image, which is subsequently mapped to a textual representation and then transformed back to an image, enforcing consistency between the generated image and the original image tokens. At each training iteration, we start with input image tokens $I$ . The unified model is used to synthesize pseudopaired text tokens $T'$ , representing the textual description of the image. We then use $T'$ in conjunction with the input image token $I$ to predict the reconstructed text tokens $\tilde{T}$ . + +In the second half of the cycle, we pass the masked image tokens $I_M$ and the text $\tilde{T}$ through the model to regenerate the masked image tokens. The cycle enforces consistency between the reconstructed and the original image tokens. The loss for enforcing I cycle consistency is given by: + +$$ +\mathcal {L} _ {I C} = - \mathbb {E} _ {I \in \mathcal {D} _ {i m a g e}} \left[ \sum_ {\forall k: m _ {k} = 1} \log p (i _ {k} | I _ {M}, \tilde {T}) \right], \tag {2} +$$ + +where the $D$ image is the set of image samples from the target domain, which are unpaired with the text samples. + +By leveraging these two cycles, our framework forces the model to refine its generative understanding of both image + +![](images/167d82eea73a700f58885a073203595172afb40ea966537b2946d4ac868e18dd.jpg) +Figure 2. The overview of T cycle (text-to-image-to-text) of the proposed DoraCycle. The I cycle is similar and is omitted in the figure for brevity. + +and text representations, ensuring consistency between the input and output while effectively leveraging unpaired data to adapt the unified towards the target domain. + +Efficient Training: In the intermediate steps of both cycles, generating the middle representation (i.e., captions or images) requires multiple forward passes. This is because the generation process involves either predicting the next tokens or the masked tokens multiple times. Backpropagating gradients through all these steps is computationally prohibitive. Thus, we first generate the intermediate results using the model in inference mode as pseudo-paired data, which are then used as the ground truth in the teacher-forcing scheme [54, 55] for the first half of the cycles. In this way, we reduce the number of forward passes to two, i.e. one for generating the middle result and one for the final output, thus making the overall training process more memory efficient. + +Token Differentiability: Since the intermediate outputs in each cycle are discrete tokens, which can not directly propagate gradients, we employ the Gumbel-Softmax [28] to make these token representations differentiable. + +# 3.2. Stabilizing Optimization + +Each cycle involves the same unified model twice in the forward pass, which leads to optimization instabilities. To stabilize the training process, we adopt the Exponential Moving Average (EMA) training technique [57]. Specifically, we maintain a shadow version of the model, referred to as the EMA model, which is updated using an exponentially decaying average of the parameters of the main model. + +$$ +\theta_ {\mathrm {E M A}} \leftarrow \alpha \theta_ {\mathrm {E M A}} + (1 - \alpha) \theta_ {\text {m i n}}, \tag {3} +$$ + +where $\alpha$ is a decay factor (set to 0.999) that controls the update rate, and $\theta_{\mathrm{main}}$ represents the parameters of the main model. + +In each training step, the EMA version of the model is used to generate the intermediate representation tokens (e.g., pseudo image or text tokens) which serve as pseudo ground truth during training. By using these stable targets from the slower-evolving EMA model, we can mitigate the risks of optimization instability. The main model is thus able to learn from more consistent and reliable intermediate targets, rather than being affected by fluctuations during the early stages of training. + +# 3.3. Balancing Two Cycles + +We observe that the T cycle tends to converge faster than the I cycle, primarily because textual data is inherently one-dimensional and simpler to learn compared to images. This imbalance in optimization leads to a kind of collapse of the model, where it tends to generate irrelevant but self-consistent captions for images, ultimately degrading the image-text alignment capability. + +To address this problem, we make the gradients of the T cycle orthogonal to those of the I cycle, thus preventing interference. This is achieved by modifying the gradients using gradient surgery [72]. Let $g_{T}$ and $g_{I}$ represent the gradients of the T cycle and the I cycle, respectively. We project $g_{T}$ onto the orthogonal complement of $g_{I}$ to obtain the modified gradient $g_{T}^{\prime}$ , which is defined as: + +$$ +g _ {T} ^ {\prime} = g _ {T} - \frac {g _ {T} \cdot g _ {I}}{g _ {I} \cdot g _ {I}} g _ {I}, \tag {4} +$$ + +where $g_{T} \cdot g_{I}$ denotes the dot product between the gradients of the T and I cycle. + +Additionally, we reweight the losses to further balance the learning between the I and T cycles. The final loss function is as, + +$$ +\mathcal {L} = \mathcal {L} _ {I C} + \beta \mathcal {L} _ {T C}, \tag {5} +$$ + +where the $\beta$ is the weight of the T cycle loss. + +# 4. Experiments + +# 4.1. Implementation Details + +To the best of our knowledge, Show-o [67] is currently the only fully open-source unified generative model with complete pre-trained weights and training code, including both its understanding and generation capabilities. Therefore, we base DoraCycle on Show-o and conduct experiments accordingly. The base model is a unified transformer model that performs understanding and image generation by predicting discrete textual and visual tokens. We insert trainable low-rank adaptations (LoRA) [25] modules into the Q projection and V projection of the attention layers from layers 7 to 24. The LoRA rank is set to 32. The $\beta$ is set to 0.1 to balance the optimization of two cycles. + +The training of DoraCycle is performed on 8 NVIDIA H100 GPUs with mixed precision enabled for memory efficiency. We set the batch size to 32, with each cycle taking half of the batch when both cycles are being optimized simultaneously. The learning rate is set to $1e^{-4}$ with a cosine annealing schedule. The optimizer is AdamW with weight decay of $1e^{-2}$ . Additionally, EMA is employed to stabilize the training process, as described in Section 3.2. + +# 4.2. Domain-Oriented Adaptations + +Unpaired Training: For tasks that do not require strongly related paired knowledge, our DoraCycle can fully learn the target domain using unpaired data. For example, to learn the cyberpunk style, we collected 300 cyberpunk-style images as input for the I cycle, and used the text data from the base model pre-training dataset [4] for the T cycle, with the keyword "cyberpunk style" automatically injected into text, prompting the model about the target style we want. + +The experimental results are shown in Fig. 3. Given the same text prompt to generate cyberpunk-style images, Fig. 3 (a) shows the images generated by the base model without additional training. It can be observed that the base model adds some cyberpunk elements, such as neon lights, but the overall atmosphere does not align well with the desired style. Fig. 3 (d) shows the images generated by the adapted model trained with DoraCycle, which aligns well with the target style. Traditional text-to-image customization or adaptation methods, such as DreamBooth [48], rely on paired data for training. Therefore, we simulate user-created paired data by annotating the collected images with captions, and split them into two groups. One group contained only 10 paired examples, which is an acceptable workload for users, while the other group contained captions for all 300 images, which would be labor-intensive and impractical for users. The images generated by the model that trained on 10 paired examples are shown in Fig. 3 (b). It struggled to produce good stylized images, likely because the combination of indoor bookshelves with the cyberpunk + + The image presents a scene of a room with a brick wall and a large bookshelf filled with various books and electronics. The bookshelf, which is black, stands against the brick wall and spans almost the entire height of the wall. It is divided into three shelves, each holding different items. On the top shelf, there are several books neatly arranged. In front of the bookshelf, there's a black leather couch positioned on one side of the room. A potted plant can be seen as well. + +![](images/2eca2e870a5dda13242b0e49d00f3556dc8424685c88ef5f06fa73192142eba8.jpg) +(a) Base model (w/o training) + +![](images/6ae94d2a6438dbf1d6ddfcbfd6156f15ffed35e7eb9eece5f0896d9cf26e1e94.jpg) +Image-to-Text + +![](images/870c95c63c517b35aa72bdf34ec2fed87f0ce48eac5e8556631ef648fa06113f.jpg) +(c) Paired trained (300 pairs) + +![](images/23baec640737327651913a6e88984467ed8188cca2fc347ddeac4dac0e94a5a9.jpg) +(d) Unpaired trained (DoraCycle) + +![](images/f89508f9d5cf179e58c291eb433cd782af6c8cd0cdaebd46e1a82c880c87961e.jpg) +(b) Paired trained (10 pairs) + +![](images/f9169b06b29a9f339c4934169bdad9889bb1b4442ce66eac6e94d26ef9b7f909.jpg) +Image-to-Text + + The image captures a scene on a city street. Dominating the foreground are two mannequins, each adorned in full Batman and Iron Man costumes. The Mannequin on the left is dressed in a black Batman costume, complete with a cape and mask, exuding an air of mystery and power. On the right, the other Mannequin is clad in a red Iron Man costume, featuring a helmet with a faceplate and red gloves, embodying the iconic superhero's persona. These mannequins stand on a gray sidewalk, their vibrant costumes contrasting against the monotone backdrop. Behind them, the city continues its usual hustle and bustle. Cars are parked along the street, their metallic bodies gleaming under the city lights. Buildings rise in the background, their windows reflecting the life of the city. + +![](images/65ea5a22b8d591bb2d788c1ec0cb374a0747b1588d5d4cefef16b99599175495.jpg) +Figure 3. Domain-oriented adaptation with different training setups. (a) Image generated by the base model without training for adoption. (b) Image generated by the model trained with 10 paired image-text samples. (c) Image generated by the model trained with 300 paired image-text samples. (d) Image generated by the model trained by DoraCycle on only unpaired data. (e) Image-to-Text-to-Image translation performed by the adapted model trained by DoraCycle. + +![](images/1bd13033b221a6a82e57404b8b11a3178f44a8db4074f9bd7fad93b6f784cc5b.jpg) +(e) Image-to-Text-to-Image translation by the adapted unified model. + +style is too novel for the model to generalize well from limited paired data. The images generated by the model trained on 300 paired examples are shown in Fig. 3 (c), which have better outputs. In contrast, the model trained using DoraCycle does not require manual captioning, significantly reducing the workload for users. + +Fig. 3 (e) illustrates the adapted model trained by DoraCyle maintains semantic consistency through image-to-text-to-image translation. The input image is transformed into a textual description and then reconstructed into an image. The result shows that the adapted model successfully captures and retains the key visual components in the original image throughout the multimodal cycle. Notably, the identity of the characters and the details of the environment are all preserved, indicating effective bidirectional understanding and generation capabilities in the target domain. Furthermore, the newly generated image incorporates styles learned from the target domain, demonstrating the generalization of the learned knowledge to images in the wild. + +Learning Paired Knowledge: For tasks that require learning some paired knowledge, such as associating an identity name with its visual appearance, DoraCycle can incorporate a small amount of paired data to learn such associations while leveraging a large amount of unpaired + +# Domain 1: Black Myth Wukong + +# Text-to-Image + +![](images/3c4f64bbba733eb1cfc9c7f0e24405ab2a625f1866662e47be5685bffa574996.jpg) + +Destined One, wears an intricately detailed golden armor with elaborate ornaments and engravings. A golden crown-like adornment sits atop his head. In the background, distant mountains and a serene sky can be seen. + +![](images/92f2055398260c2f3c80e6ade333da387b1ef74b3ef192a8826c47608c0e5d7d.jpg) + +Lingxuizi's glowing red eyes and bared fangs indicate its aggressive nature. It crouches low to the ground, its large hand-like paw spread over ancient stone paving surrounded by patches of grass. + +![](images/f57acc46503ff6b286c6582823f92d79962e1e9a3896ba3e16f6faa381106488.jpg) + +Keeper grips a gnarled wooden staff for support as he walks through a wooded area. His posture is stooped, and his facial features are mostly obscured by his thick beard. The scene is set in a forest, with a textured ground of leaves and dirt. + +![](images/e56ab250c53c79fd423d31f583501da649fb5fbdfd242f88529ab40aac4a9a71.jpg) + +Zhu Bajie stands confidently in the snow, gripping his staff with a sly smile on his boat-like face, beads hanging around his neck. Beside him lies the majestic, wounded form of Kang-Jin Loong, a silver-scaled dragon with intricate, ethereal horns and whiskers. + +![](images/ff73fe5e4e66bf09fe7a83881410a8e0ffae1b43695f0bdf0c062549c9a3084f.jpg) + +Tiger Vanguard charges through the blood-red waters towards Destined One. The massive stone statue behind him looms ominously, adding to the intensity of the scene. Destined One stands firm, his staff held at the ready, his gaze unwavering as he prepares for the oncoming assault. The water ripples around them, reflecting the dark hues of the battle setting. + +![](images/d8f24403546657cbf631464bf9a63f5cc75c38ac27f953351b65d80eaa34ad1e.jpg) + +Destined One stands tall, his gaze cast downward, his weathered yellow robe flowing with subtle wear, reflecting the trials he has endured. Beside him, Zhu Majie looks up with a playful expression, his boar-like face filled with curiosity and mischief. The two figures share a quiet moment, framed against an open, clear sky dotted with drifting embers. + +# Image-to-Text + +![](images/c8cfb3108961accebf9fcff37647a316572990ec319ee49156e18605f7acdb0c.jpg) +(a) + +Destined One stands tall, wielding a massive golden stick that dominates the frame. The individual exudes power, strength, and determination, clad in ornate armor that gleams with gold. A hat adds a touch of regality. The background is dramatic, with shadows cast by trees hinting at a mysterious atmosphere. The grass beneath the figure is lush and green, contrasting with the ruggedness of the ground. + +![](images/704e383d3f5d55ae6966d12ab4faef5a2416475478a563593dfee90f7ebda39f.jpg) +(b) + +Keeper, an elderly, hunched individual with a bald, heavily wrinkled head covered in spots. Keeper has a long, unkempt beard and is dressed in a simple, tattered blue robe. He grips a gnarled wooden staff for support as he walks through a forest fire. His posture is stooped, and his facial features are mostly obscured by his thick beard. + +![](images/aadf39eb053a5ed03c1e774206f6facdabcdbf7ddb9b1f67cef4f56ed0c8aa91.jpg) +(c) + +The image depicts a serene, snowy landscape with a large, ancient-looking wooden bridge spanning across a frozen river. The bridge has a traditional, Asian design, with intricate carvings and a slatted roof. Bare trees with frost-covered branches surround the scene, casting a calm, wintery atmosphere. Snowflakes gently fall, enhancing the tranquil, mystical ambiance + +![](images/e4a5e9c242fca82aa9cc3e2dd5692905c86c225cebe3c161578f9c0933a0630d.jpg) +(d) + +Lingxuzi is depicted as a large, menacing fox-like creature with thick, flowing white fur. Its red eyes glow intensely, giving an eerie, supernatural appearance. The creature is in a crouching stance, with its body slightly arched and its tails fanned out behind it, showing both power and readiness to strike. The dark background adds to the ominous atmosphere, contrasting with the brightness of its fur. + +![](images/b7f74ba3fcf5b04cea895ebbb2b4dcbe59456f4a51fe61891b691dd6768f720e.jpg) +(e) + +Zhu Bajie stands menacingly in a murky, mist-covered swamp, his humanoid body combined with the head of a fierce, snarling boar. Her red, glowing eyes pierce through the fog, radiating a dark energy. Dressed in a tattered green robe, he wears a string of large beads around his neck, adding a sense of mysticism. His left hand clutches a spiked weapon, while shadows and wisps of smoke swirl around him, amplifying his sinister presence. + +![](images/f9ecdabaabd601729254048c6fb77491b9dafde4b0029f5982d36e14a2494d62.jpg) +(f) + +A female character with a unique appearance, having a combination of human and demonic features. She has a short, black, straight hairstyle and a pair of horns protruding from her head. Her skin is pale, and she has a few dots on her forehead, giving her a mystical look. She is wearing a black, low-cut top adorned with white flowers, and her right arm is covered in red fabric. She has a focused and intense expression on her face, as if she is concentrating on something important. (D) + +# Text-to-Image + +![](images/c3a1c8d8eb39e13c9dd4470d110fcdaf3399a6aaabea6d7afd73fe7258e8663d.jpg) +(g) + +![](images/6f2633121fd8eebbf802b5716e36b442c498538e94e037ccac324b5c8c7e3f07.jpg) + +Nobita is sitting under a tree, trying to finish his math homework. He looks stressed, with one hand on his chin and scratching his head with the other. + +![](images/ab5c77f33e601cc1c923973cc010f58c95788b20ab83699c9f4b4cce12df1178.jpg) +(h) + +Suneo and Nobita are at a fishing pond, trying to catch a fish. Nobita, in his bright yellow fishing hat and a sharp outfit, confidently casts his line into the water. Suneo watches Nobita's struggle with a smile, offering some tips to help him out. The two boys share a laugh as they wait for a bite. + +![](images/f13264c30b6a0da6e01d4025bac2335db2d2f8274eb375b9cf1362accf377f43.jpg) +(i) +Domain 2: Doraemon + +Doraemon is cooking in the kitchen, while holding a bowl. He looks excited, flipping a pancake into the air. Nobita is standing beside him, smiling gently, watching Doraemon. + +![](images/b8241675d95fba7f82713398267b5bbdeb136894452978f0de260cebadd351d8.jpg) +(j) + +Shizuka is sitting by a piano, composing a new tune. She looks thoughtful, her fingers delicately pressing the keys as she tries different melodies. A notebook rests on the piano stand, filled with her notes and musical ideas, and she appears content in the process. + +![](images/c2a30f70ce7eab0d82a13395f7cdd31e3e5739853fc70813e09bd9fb30c7d45e.jpg) + +Shizuka and Nobita are making flower crowns at the park. Shizuka, in her soft pink dress, weaves flowers together. Nobita, in his casual shorts, tries his best to copy her. + +![](images/ce3544a651f6016e9caeb2a81d4c904b4e05405714243cb8dae911ef4a8a14a7.jpg) +(1) + +Doraemon is standing in front of the white cat with an excited expression. He is holding a bouquet of flowers, in front of the cat. The white cat, sitting on a small cushion, is looking at Doraemon with a curious or intrigued expression. The setting appears to be indoors, with a pink carpet, a scratching post in the background, and a framed picture on the wall. + +# Image-to-Text + +![](images/66cbdf51c43a6bb3ef2b83bf353c163add826a4f6fc7f26e0d28fc812a5da9f2.jpg) +(s) +Figure 4. Image-to-text and text-to-image generation by the unified models that adapted for two domains. The special tokens are omitted. + +Doraemon is standing on the shore, looking out at the sunset. He has a calm expression, and the sun is setting over the water, casting a warm glow. The sky is filled with clouds, and there are silhouettes of buildings in the distance. + +![](images/b3928faa52e349ecd559c44612e7824d6c7c710385ea9e3880a29c332b0ffeb3.jpg) +(n) +(t) + +Doraemon is standing in the grass, looking surprised as he sees a group of cats. The cats are gathered around him, with one of them appearing to be a cat with a surprised expression. The background shows a fence and some bushes. + +![](images/849259d7fbc5ab22955acb39c7b11b335ce7135ac7e17dbbcbe6776882e7b8da.jpg) +(u) + +Doradamon is flying through the air, looking down at the city below. Nobita is also flying through the air, looking down at the city. + +![](images/5fea7abcdc30d6b74bcfcd1eb4ba39cf92839d1563150bf49669380f3b02be92.jpg) +(p) +(v) + +The image depicts a two-story house with a large garden in front of it. The house has a tan exterior and is surrounded by a variety of plants, including palm trees and bushes. The garden is well-maintained. The house features a balcony on the second floor, adding to its charm. + +![](images/8245120e916c34348b7d1725f64102d1ac7f99d8a264daa5af3c8f223ad0377f.jpg) +(q) +(w) + +Nobita is standing in the doorway, holding a plate with a doughnut on it. He looks excited, and the doughnut is on a white plate. The door is open, and the room is visible in the background. + +![](images/67ecfcbba9ae96f1be693b1efe68514ba38a55c3d9198ab7f8e1b1c19a3d13c6.jpg) +(r) +(x) + +White cat and black cat are sitting on the grass, looking at each other with affection, with a blue sky in the background, and a few clouds scattered across the sky. + +![](images/47a1a143534d9ec5df110d474acfeb5385995373562f237710a5036fc5032de7.jpg) +Figure 5. Effect of special tokens on character learning. (a) Base model without training. (b) Model trained without using special tokens, showing attribute confusion among characters. (c) Model trained with special tokens, improving character attribute alignment and reducing confusion. + +data to comprehensively learn general aspects of the target domain. Specifically, in each batch, for data with paired ground truth, we compute the token prediction loss and also include it in the cycle, use ground truth as the pseudo middle generations, and compute the cycle loss. For unpaired data, we compute the unpaired cycle loss. + +For example, when adapting the model to Domain 1: Black Myth Wukong and Domain 2: Doraemon, we annotate 1-3 images per unique identity with captions that specify the name of the identity. For each domain, we collect $2\mathrm{k}$ images, which are mostly sampled from online videos, and independently collect text descriptions, which are further expanded to $1\mathrm{k}$ by ChatGPT [39]. The final adapted model trained with DoraCycle demonstrates strong performance in both text-to-image generation and image-to-text generation, as shown in Fig. 4. + +In terms of text-to-image results, the model trained with DoraCycle effectively generated images that aligned well with the target domains. In domain 1 (Black Myth Wukong), the generated images accurately depicted domain-specific visual elements, such as the intricate details of character appearances and the overall fantasy-like atmosphere. This indicates that the model successfully learned to generalize the visual features from text prompts to realistic images within the target domain. Similarly, in domain 2 (Doraemon), the generated images preserve the iconic cartoonish aesthetics and capture key visual details of the characters and settings, demonstrating effective domain adaptation. + +For the image-to-text task, the model performs well in + +generating contextually accurate captions. In domain 1, the generated captions provide rich descriptions of the characters, their attributes, and the context, effectively mirroring the visual elements present in the input images. In domain 2, the captions correctly describe the characters, their actions, and their environments concisely, maintaining consistency with the visual style. The ability of the model to generate accurate descriptions highlights its robust understanding of the visual components of the domain. + +Additionally, an interesting phenomenon can be observed in how the model handles the visual elements that are not annotated with paired data. For instance, in Fig. 4 (w), the dorayaki (a type of sweet bean-filled pancake) was described by the model as a "doughnut". This may be due to the fact that the anime-style representation of the dorayaki is novel, and neither the base model nor the unpaired training provided specific textual-visual pairing knowledge about it. On the other hand, in the example shown in Fig. 4 (x), we annotate the white cat as a character with paired textual and visual data, using a special token for its name: " white cat black cat " during the caption generation. This suggests that the model autonomously categorized the black cat as a character when learning the target domain, indicating that it may have attempted to generalize learned knowledge from one type of entity to similar ones. + +Enhanced Learning with Special Tokens: As shown in Fig. 5, we experimentally find that the model often confused multiple novel concepts in the target domain. Fig. 5 (a) shows the image generated by the base model without training taking in the name of characters. Fig. 5 (b) shows the characters generated by the trained model. During training, the names of characters are directly included in the text without special treatment, leading to attribute confusion between characters. The varying lengths of the tokenized character names also make learning difficult. To solve this problem, we introduce a simple yet efficient solution: adding special tokens around character names. We introduced start of character $(\langle \text{soc}\rangle)$ and end of character $(\langle \text{soc}\rangle)$ tokens to enclose character names, which significantly enhance the learning of novel concepts. As shown in Fig. 5 (c), involving special tokens improves the alignments between characters and their names. + +# 4.3. Comparisons + +In this section, we use the Storyboard20K [66] dataset to conduct the quantitative comparison experiments. The storyboards originating from the same data source are grouped to form a domain, consisting of images and descriptive text. The data are used under three different settings, i.e. totally unpaired, only paired, and paired plus unpaired data, + +Table 1. Comparison of different training methods under various data settings. The best value is highlighted in blue , and the second-best value is highlighted in green . “P” indicates paired data, and “U” indicates unpaired data. + +
T cycleI cycleT DataI DataFID-1K ↓CIDEr ↑Human Eval
T2I Align ↑I2T Align ↑
DreamBooth [48]--10% P10% P33.2232.743.251.83
--100% P100% P24.9341.554.133.96
DoraCycleX✓X100% U28.9330.543.381.62
✓X100% UX36.6335.703.262.17
✓✓100% U100% U27.4438.173.843.42
✓✓10% P + 90% U10% P + 90% U25.3740.904.123.81
ITIT [34]✓✓10% P + 90% U10% P + 90% U27.5038.623.853.52
+ +as shown in Table 1. + +The compared methods include DreamBooth [48] and ITIT [34]. We implement DreamBooth as a paired-training baseline by applying LoRA fine-tuning on the unified model. The original design of ITIT is different, in which the image and text decoders are separate models, and its code has not been released. We adjusted and re-implemented it to be suitable for our unified model architecture. + +We use both automatic and human evaluations to compare the performance of different methods. For automatic evaluation, we use FID to measure the distribution differences between the generated images and the target domain images [22], and CIDEr to compute the error between the generated text and the ground truth [59]. For human evaluation, we create 100 questions for the generated results of models, each rated by three different human raters. The raters are asked to evaluate the alignment between the image and text on a scale from 1 to 5, where 1 indicates no relevance and 5 indicates complete alignment. + +The experimental results in Table 1 demonstrate that the proposed DoraCycle performs competitively under several data settings. Specifically, when using a combination of paired and unpaired data, DoraCycle outperforms ITIT. Compared to DreamBooth, which heavily relies on paired data, DoraCycle outperforms it when using the same scale of paired data, i.e. $10\%$ paired data, indicating the benefits brought by $90\%$ unpaired data. While Dreambooth with $100\%$ achieves the best evaluation scores, the scores of the DoraCycle with $10\%$ paired and $90\%$ unpaired data are comparable with them. + +Table 1 also shows the difference in the performance of DoraCyle under different cycle settings. It is shown that without the T cycle and with only the I cycle, the captioning ability of the adapted model degrades more significantly. In contrast, if only the T cycle is used and without the I cycle, the FID score increases substantially, indicating that the generated image distribution mismatches with the target distribution. + +# 4.4. Ablation Studies + +Table 2 shows that removing key components from DoraCycle significantly impacts performance. Without EMA, the FID score increases from 25.37 to 27.19, indicating lower image quality + +Table 2. Ablation Studies. EMA refers to the exponential moving average. GS refers to gradient surgery. + +
FID-1K ↓CIDEr ↑
w/o EMA27.1938.85
w/o GS25.5439.98
DoraCycle25.3740.90
+ +due to less stable training. Removing Gradient Surgery (GS) will reduce the CIDEr score and increase the FID, indicating a worse performance. This demonstrates the importance of mitigating the interference between the optimization directions of two cycles. The complete DoraCycle framework, with both EMA and GS, has the best performance across all metrics, demonstrating the importance of these components in achieving better optimization. + +# 5. Conclusion + +We propose the DoraCycle to adapt the unified generative model to target domains within multimodal cycles. By leveraging both image-to-text-to-image and text-to-image-to-text cycles, DoraCycle changes the learning objectives into the same modality, allowing for effective optimization using unpaired data. Our experiments show that DoraCycle can adapt the unified model to target domains using only unpaired data, or involving a small amount of paired data when necessary to learn specific concepts. Experimental results demonstrate that DoraCycle achieves advanced or comparable performance across various settings. Leveraging unpaired data broadens the application potential of DoraCycle, making it ideally suited for domain adaptation tasks where paired data is scarce or challenging to collect. + +# 6. Acknowledgments + +This research is supported by the National Research Foundation, Singapore under its AI Singapore Programme (AISG Award No: AISG3-RP-2022-030). + +# References + +[1] Armen Aghajanyan, Bernie Huang, Candace Ross, Vladimir Karpukhin, Hu Xu, Naman Goyal, Dmytro Okhonko, Mandar Joshi, Gargi Ghosh, Mike Lewis, et al. Cm3: A causal masked multimodal model of the internet. arXiv preprint arXiv:2201.07520, 2022. 2 +[2] Emanuele Aiello, LILI YU, Yixin Nie, Armen Aghajanyan, and Barlas Oguz. Jointly training large autoregressive multimodal models. In ICLR, 2024. 2 +[3] Yogesh Balaji, Seungjun Nah, Xun Huang, Arash Vahdat, Jiaming Song, Karsten Kreis, Miika Aittala, Timo Aila, Samuli Laine, Bryan Catanzaro, et al. ediffi: Text-to-image diffusion models with an ensemble of expert denoisers. arXiv preprint arXiv:2211.01324, 2022. 2 +[4] Soravit Changpinyo, Piyush Sharma, Nan Ding, and Radu Soricut. Conceptual 12m: Pushing web-scale image-text pretraining to recognize long-tail visual concepts. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 3558-3568, 2021. 5 +[5] Junsong Chen, Chongjian Ge, Enze Xie, Yue Wu, Lewei Yao, Xiaozhe Ren, Zhongdao Wang, Ping Luo, Huchuan Lu, and Zhenguo Li. Pixart $\backslash$ sigma: Weak-to-strong training of diffusion transformer for 4k text-to-image generation. arXiv preprint arXiv:2403.04692, 2024. 2 +[6] Junsong Chen, Jincheng Yu, Chongjian Ge, Lewei Yao, Enze Xie, Zhongdao Wang, James T. Kwok, Ping Luo, Huchuan Lu, and Zhenguo Li. Pixart-α: Fast training of diffusion transformer for photorealistic text-to-image synthesis. In ICLR. OpenReview.net, 2024. 2 +[7] Li Chen, Mengyi Zhao, Yiheng Liu, Mingxu Ding, Yangyang Song, Shizun Wang, Xu Wang, Hao Yang, Jing Liu, Kang Du, et al. Photovsere: Tuning-free image customization with text-to-image diffusion models. arXiv preprint arXiv:2309.05793, 2023. 2 +[8] Mark Chen, Alec Radford, Rewon Child, Jeffrey Wu, Heewoo Jun, David Luan, and Ilya Sutskever. Generative pretraining from pixels. In ICML, pages 1691-1703, 2020. 2 +[9] Xi Chen, Lianghua Huang, Yu Liu, Yujun Shen, Deli Zhao, and Hengshuang Zhao. Anydoor: Zero-shot object-level image customization. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 6593-6602, 2024. 2 +[10] Yunjey Choi, Minje Choi, Munyoung Kim, Jung-Woo Ha, Sunghun Kim, and Jaegul Choo. Stargan: Unified generative adversarial networks for multi-domain image-to-image translation. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition (CVPR), 2018. 3 +[11] Wenliang Dai, Junnan Li, Dongxu Li, Anthony Meng Huat Tiong, Junqi Zhao, Weisheng Wang, Boyang Li, Pascale Fung, and Steven Hoi. Instructlip: Towards general- + +purpose vision-language models with instruction tuning, 2023. 2 +[12] Xiaoliang Dai, Ji Hou, Chih-Yao Ma, Sam Tsai, Jialiang Wang, Rui Wang, Peizhao Zhang, Simon Vandenhende, Xiaofang Wang, Abhimanyu Dubey, et al. Emu: Enhancing image generation models using photogenic needles in a haystack. arXiv preprint arXiv:2309.15807, 2023. 2 +[13] Prafulla Dhariwal and Alexander Nichol. Diffusion models beat gans on image synthesis. Advances in Neural Information Processing Systems, 34:8780-8794, 2021. 2 +[14] Runpei Dong, Chunrui Han, Yuang Peng, Zekun Qi, Zheng Ge, Jinrong Yang, Liang Zhao, Jianjian Sun, Hongyu Zhou, Haoran Wei, Xiangwen Kong, Xiangyu Zhang, Kaisheng Ma, and Li Yi. DreamLLM: Synergistic multimodal comprehension and creation. In ICLR, 2024. 1, 2 +[15] Debidatta Dwibedi, Yusuf Aytar, Jonathan Tompson, Pierre Sermanet, and Andrew Zisserman. Temporal cycle-consistency learning. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 1801-1810, 2019. 3 +[16] Patrick Esser, Sumith Kulal, Andreas Blattmann, Rahim Entezari, Jonas Müller, Harry Saini, Yam Levi, Dominik Lorenz, Axel Sauer, Frederic Boesel, et al. Scaling rectified flow transformers for high-resolution image synthesis. arXiv preprint arXiv:2403.03206, 2024. 2 +[17] Rinon Gal, Yuval Alaluf, Yuval Atzmon, Or Patashnik, Amit H Bermano, Gal Chechik, and Daniel Cohen-Or. An image is worth one word: Personalizing text-to-image generation using textual inversion. arXiv preprint arXiv:2208.01618, 2022. 2 +[18] Yuying Ge, Sijie Zhao, Jinguo Zhu, Yixiao Ge, Kun Yi, Lin Song, Chen Li, Xiaohan Ding, and Ying Shan. Seed-x: Multimodal models with unified multi-granularity comprehension and generation. arXiv preprint arXiv:2404.14396, 2024. 1, 2 +[19] Yuchao Gu, Xintao Wang, Jay Zhangjie Wu, Yujun Shi, Yunpeng Chen, Zihan Fan, Wuyou Xiao, Rui Zhao, Shuning Chang, Weijia Wu, et al. Mix-of-show: Decentralized low-rank adaptation for multi-concept customization of diffusion models. Advances in Neural Information Processing Systems, 36:15890-15902, 2023. 2 +[20] Jiayi Guo, Chaofei Wang, You Wu, Eric Zhang, Kai Wang, Xingqian Xu, Shiji Song, Humphrey Shi, and Gao Huang. Zero-shot generative model adaptation via image-specific prompt learning. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 11494-11503, 2023. 2 +[21] Di He, Yingce Xia, Tao Qin, Liwei Wang, Nenghai Yu, TieYan Liu, and Wei-Ying Ma. Dual learning for machine translation. In Advances in Neural Information Processing Systems (NeurIPS), 2016. 3 +[22] Martin Heusel, Hubert Ramsauer, Thomas Unterthiner, Bernhard Nessler, and Sepp Hochreiter. Gans trained by a two time-scale update rule converge to a local nash equilibrium. In Proceedings of the 31st International Conference on Neural Information Processing Systems, pages 6626-6637, 2017. 8 + +[23] Judy Hoffman, Eric Tzeng, Taesung Park, Jun-Yan Zhu, Phillip Isola, Kate Saenko, Alexei A Efros, and Trevor Darryll. Cycada: Cycle-consistent adversarial domain adaptation. In Proceedings of the International Conference on Machine Learning (ICML), 2018. 3 +[24] MD Zakir Hossain, Ferdous Sohel, Mohd Fairuz Shiratuddin, and Hamid Laga. A comprehensive survey of deep learning for image captioning. ACM Computing Surveys (CsUR), 51(6):1-36, 2019. 2 +[25] Edward J. Hu, Yelong Shen, Phillip Wallis, Zeyuan Allen-Zhu, Yuanzhi Li, Shean Wang, and Weizhu Chen. Lora: Low-rank adaptation of large language models. In International Conference on Learning Representations, 2022. 5 +[26] Jia Cheng Hu, Roberto Cavicchioli, and Alessandro Capontodi. Exploiting multiple sequence lengths in fast end to end training for image captioning. In 2023 IEEE International Conference on Big Data (BigData), pages 2173-2182. IEEE, 2023. 2 +[27] Xiaowei Hu, Zhe Gan, Jianfeng Wang, Zhengyuan Yang, Zicheng Liu, Yumao Lu, and Lijuan Wang. Scaling up vision-language pre-training for image captioning. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 17980-17989, 2022. 2 +[28] Eric Jang, Shixiang Gu, and Ben Poole. Categorical reparameterization with gumbel-softmax. arXiv preprint arXiv:1611.01144, 2016. 4 +[29] Xuhui Jia, Yang Zhao, Kelvin CK Chan, Yandong Li, Han Zhang, Boqing Gong, Tingbo Hou, Huisheng Wang, and Yu-Chuan Su. Taming encoder for zero fine-tuning image customization with text-to-image diffusion models. arXiv preprint arXiv:2304.02642, 2023. 2 +[30] Nupur Kumari, Bingliang Zhang, Richard Zhang, Eli Shechtman, and Jun-Yan Zhu. Multi-concept customization of text-to-image diffusion. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 1931-1941, 2023. 2 +[31] Chenliang Li, Haiyang Xu, Junfeng Tian, Wei Wang, Ming Yan, Bin Bi, Jiabo Ye, Hehong Chen, Guohai Xu, Zheng Cao, et al. mplug: Effective and efficient vision-language learning by cross-modal skip-connections. arXiv preprint arXiv:2205.12005, 2022. 2 +[32] Chunyuan Li, Zhe Gan, Zhengyuan Yang, Jianwei Yang, Linjie Li, Lijuan Wang, Jianfeng Gao, et al. Multimodal foundation models: From specialists to general-purpose assistants. Foundations and Trends® in Computer Graphics and Vision, 16(1-2):1-214, 2024. 2 +[33] Junnan Li, Dongxu Li, Silvio Savarese, and Steven Hoi. Blip-2: Bootstrapping language-image pre-training with frozen image encoders and large language models. In International conference on machine learning, pages 19730-19742. PMLR, 2023. 2 +[34] Tianhong Li, Sangnie Bhardwaj, Yonglong Tian, Han Zhang, Jarred Barber, Dina Katabi, Guillaume Lajoie, Huiwen Chang, and Dilip Krishnan. Leveraging unpaired data for vision-language generative models via cycle consistency. arXiv preprint arXiv:2310.03734, 2023. 3, 8 +[35] Haotian Liu, Chunyuan Li, Qingyang Wu, and Yong Jae Lee. Visual instruction tuning. NeurIPS, 36, 2024. 2 + +[36] Nanye Ma, Mark Goldstein, Michael S Albergo, Nicholas M Boffi, Eric Vanden-Eijnden, and Saining Xie. Sit: Exploring flow and diffusion-based generative models with scalable interpolant transformers. arXiv preprint arXiv:2401.08740, 2024. 2 +[37] Xin Ma, Yaohui Wang, Gengyun Jia, Xinyuan Chen, Zwei Liu, Yuan-Fang Li, Cunjian Chen, and Yu Qiao. Latte: Latent diffusion transformer for video generation. arXiv preprint arXiv:2401.03048, 2024. 2 +[38] Ron Mokady, Amir Hertz, Kfir Aberman, Yael Pritch, and Daniel Cohen-Or. Null-text inversion for editing real images using guided diffusion models. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 6038–6047, 2023. 2 +[39] OpenAI. Chatgpt. https://chatgpt.com/, 2024.7 +[40] Alec Radford, Jong Wook Kim, Chris Hallacy, Aditya Ramesh, Gabriel Goh, Sandhini Agarwal, Girish Sastry, Amanda Askell, Pamela Mishkin, Jack Clark, et al. Learning transferable visual models from natural language supervision. In International conference on machine learning, pages 8748-8763. PMLR, 2021. 2 +[41] Aditya Ramesh, Mikhail Pavlov, Gabriel Goh, Scott Gray, Chelsea Voss, Alec Radford, Mark Chen, and Ilya Sutskever. Zero-shot text-to-image generation. In ICML, pages 8821-8831. Pmlr, 2021. 2 +[42] Aditya Ramesh, Prafulla Dhariwal, Alex Nichol, Casey Chu, and Mark Chen. Hierarchical text-conditional image generation with CLIP latents. CoRR, abs/2204.06125, 2022. 2 +[43] Aditya Ramesh, Prafulla Dhariwal, Alex Nichol, Casey Chu, and Mark Chen. Hierarchical diffusion models for text-to-image generation. arXiv preprint arXiv:2204.06125, 2022. 2 +[44] Aditya Ramesh, Prafulla Dhariwal, Alex Nichol, Casey Chu, and Mark Chen. Hierarchical text-conditional image generation with clip latents. arXiv preprint arXiv:2204.06125, 1 (2):3, 2022. 2 +[45] Robin Rombach, Andreas Blattmann, Dominik Lorenz, Patrick Esser, and Björn Ommer. High-resolution image synthesis with latent diffusion models. In CVPR, pages 10684-10695, 2022. 2 +[46] Robin Rombach, Andreas Blattmann, Dominik Lorenz, Patrick Esser, and Björn Ommer. High-resolution image synthesis with latent diffusion models. arXiv preprint arXiv:2112.10752, 2022. 2 +[47] Nataniel Ruiz, Yuanzhen Li, Varun Jampani, Yael Pritch, Michael Rubinstein, and Kfir Aberman. Dreambooth: Fine tuning text-to-image diffusion models for subject-driven generation. arXiv preprint arXiv:2208.12242, 2022. 1, 2 +[48] Nataniel Ruiz, Yuanzhen Li, Varun Jampani, Yael Pritch, Michael Rubinstein, and Kfir Aberman. Dreambooth: Fine tuning text-to-image diffusion models for subject-driven generation. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 22500-22510, 2023. 5, 8 +[49] Chitwan Sahara, William Chan, Saurabh Saxena, Lala Li, Jay Whang, Emily L Denton, Kamyar Ghasemipour, Raphael Gontijo Lopes, Burcu Karagol Ayan, Tim Salimans, + +et al. Photorealistic text-to-image diffusion models with deep language understanding. Advances in Neural Information Processing Systems, 35:36479-36494, 2022. 2 +[50] Rico Sennrich, Barry Haddow, and Alexandra Birch. Improving neural machine translation models with monolingual data. In Proceedings of the 54th Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers), 2016. 3 +[51] Meet Shah, Xinlei Chen, Marcus Rohrbach, and Devi Parikh. Cycle-consistency for robust visual question answering. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 6649-6658, 2019. 3 +[52] Sameer Shah, Siddharth Bharadwaj, Devi Parikh, and Dhruv Batra. Cycle-consistency for robust visual question answering. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition (CVPR), 2019. 3 +[53] Quan Sun, Yufeng Cui, Xiaosong Zhang, Fan Zhang, Qiying Yu, Yueze Wang, Yongming Rao, Jingjing Liu, Tiejun Huang, and Xinlong Wang. Generative multimodal models are in-context learners. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 14398-14409, 2024. 2 +[54] I Sutskever. Sequence to sequence learning with neural networks. arXiv preprint arXiv:1409.3215, 2014. 4 +[55] Richard S Sutton. Learning to predict by the methods of temporal differences. Machine learning, 3:9-44, 1988. 4 +[56] Zineng Tang, Ziyi Yang, Chenguang Zhu, Michael Zeng, and Mohit Bansal. Any-to-any generation via composable diffusion. NeurIPS, 36, 2024. 2 +[57] Antti Tarvainen and Harri Valpola. Mean teachers are better role models: Weight-averaged consistency targets improve semi-supervised deep learning results. In Advances in Neural Information Processing Systems, pages 1195–1204, 2017. 2, 4 +[58] Chameleon Team. Chameleon: Mixed-modal early-fusion foundation models. arXiv preprint arXiv:2405.09818, 2024. 1, 2, 3 +[59] Ramakrishna Vedantam, C. Lawrence Zitnick, and Devi Parikh. Cider: Consensus-based image description evaluation. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition (CVPR), pages 4566-4575, 2015. 8 +[60] Jianfeng Wang, Zhengyuan Yang, Xiaowei Hu, Linjie Li, Kevin Lin, Zhe Gan, Zicheng Liu, Ce Liu, and Lijuan Wang. Git: A generative image-to-text transformer for vision and language. arXiv preprint arXiv:2205.14100, 2022. 2 +[61] Peng Wang, An Yang, Rui Men, Junyang Lin, Shuai Bai, Zhikang Li, Jianxin Ma, Chang Zhou, Jingren Zhou, and Hongxia Yang. Ofa: Unifying architectures, tasks, and modalities through a simple sequence-to-sequence learning framework. In International conference on machine learning, pages 23318-23340. PMLR, 2022. 2 +[62] Qixun Wang, Xu Bai, Haofan Wang, Zekui Qin, Anthony Chen, Huaxia Li, Xu Tang, and Yao Hu. Instantid: Zero-shot identity-preserving generation in seconds. arXiv preprint arXiv:2401.07519, 2024. 1 +[63] Yuchi Wang, Shuhuai Ren, Rundong Gao, Linli Yao, Qingyan Guo, Kaikai An, Jianhong Bai, and Xu Sun. + +Ladic: Are diffusion models really inferior to autoregressive counterparts for image-to-text generation? arXiv preprint arXiv:2404.10763, 2024. 2 +[64] Chengyue Wu, Xiaokang Chen, Zhiyu Wu, Yiyang Ma, Xingchao Liu, Zizheng Pan, Wen Liu, Zhenda Xie, Xingkai Yu, Chong Ruan, et al. Janus: Decoupling visual encoding for unified multimodal understanding and generation. arXiv preprint arXiv:2410.13848, 2024. 2 +[65] Shengqiong Wu, Hao Fei, Leigang Qu, Wei Ji, and Tat-Seng Chua. Next-gpt: Any-to-any multimodal llm. arXiv preprint arXiv:2309.05519, 2023. 2 +[66] Jinheng Xie, Jiajun Feng, Zhaoxu Tian, Kevin Qinghong Lin, Yawen Huang, Xi Xia, Nanxu Gong, Xu Zuo, Ji-aqi Yang, Yefeng Zheng, et al. Learning long-form video prior via generative pre-training. arXiv preprint arXiv:2404.15909, 2024. 7 +[67] Jinheng Xie, Weijia Mao, Zechen Bai, David Junhao Zhang, Weihao Wang, Kevin Qinghong Lin, Yuchao Gu, Zhijie Chen, Zhenheng Yang, and Mike Zheng Shou. Show-o: One single transformer to unify multimodal understanding and generation. arXiv preprint arXiv:2408.12528, 2024. 1, 2, 3, 5 +[68] Ceyuan Yang, Yujun Shen, Zhiyi Zhang, Yinghao Xu, Jiapeng Zhu, Zhirong Wu, and Bolei Zhou. One-shot generative domain adaptation. In Proceedings of the IEEE/cvf international conference on computer vision, pages 7733-7742, 2023. 2 +[69] Hanrong Ye, De-An Huang, Yao Lu, Zhiding Yu, Wei Ping, Andrew Tao, Jan Kautz, Song Han, Dan Xu, Pavlo Molchanov, et al. X-vila: Cross-modality alignment for large language model. arXiv preprint arXiv:2405.19335, 2024. 2 +[70] Haoxuan You, Mandy Guo, Zhecan Wang, Kai-Wei Chang, Jason Baldridge, and Jiahui Yu. Cobit: A contrastive bidirectional image-text generation model. arXiv preprint arXiv:2303.13455, 2023. +[71] Lili Yu, Bowen Shi, Ramakanth Pasunuru, Benjamin Muller, Olga Golovneva, Tianlu Wang, Arun Babu, Binh Tang, Brian Karrer, Shelly Sheynin, et al. Scaling autoregressive multimodal models: Pretraining and instruction tuning. arXiv preprint arXiv:2309.02591, 2(3), 2023. 2 +[72] Tianhe Yu, Saurabh Kumar, Abhishek Gupta, Sergey Levine, Karol Hausman, and Chelsea Finn. Gradient surgery for multi-task learning. In Advances in Neural Information Processing Systems, pages 5824-5836, 2020. 4 +[73] Gengwei Zhang, Guoliang Kang, Yi Yang, and Yunchao Wei. Few-shot segmentation via cycle-consistent transformer. Advances in Neural Information Processing Systems, 34:21984-21996, 2021. 3 +[74] Rui Zhao, Wei Li, Zhipeng Hu, Lincheng Li, Zhengxia Zou, Zhenwei Shi, and Changjie Fan. Zero-shot text-to-parameter translation for game character auto-creation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 21013–21023, 2023. 2 +[75] Rui Zhao, Yuchao Gu, Jay Zhangjie Wu, David Junhao Zhang, Jia-Wei Liu, Weijia Wu, Jussi Keppo, and Mike Zheng Shou. Motiondirector: Motion customization of text-to-video diffusion models. In European Conference on Computer Vision, pages 273–290. Springer, 2024. + +[76] Rui Zhao, Hangjie Yuan, Yujie Wei, Shiwei Zhang, Yuchao Gu, Lingmin Ran, Xiang Wang, Jay Zhangjie Wu, David Junhao Zhang, Yingya Zhang, et al. Evolvedirector: Approaching advanced text-to-image generation with large vision-language models. Advances in Neural Information Processing Systems, 37:122104-122129, 2025. 2 +[77] Chunting Zhou, Lili Yu, Arun Babu, Kushal Tirumala, Michihiro Yasunaga, Leonid Shamis, Jacob Kahn, Xuezhe Ma, Luke Zettlemoyer, and Omer Levy. Transfusion: Predict the next token and diffuse images with one multi-modal model. arXiv preprint arXiv:2408.11039, 2024. 1, 2, 3 +[78] Deyao Zhu, Jun Chen, Xiaoqian Shen, Xiang Li, and Mohamed Elhoseiny. Minigpt-4: Enhancing vision-language understanding with advanced large language models. CoRR, abs/2304.10592, 2023. 2 +[79] Jun-Yan Zhu, Taesung Park, Phillip Isola, and Alexei A. Efros. Unpaired image-to-image translation using cycle-consistent adversarial networks. In Proceedings of the IEEE International Conference on Computer Vision (ICCV), 2017. 3 \ No newline at end of file diff --git a/2025/DoraCycle_ Domain-Oriented Adaptation of Unified Generative Model in Multimodal Cycles/images.zip b/2025/DoraCycle_ Domain-Oriented Adaptation of Unified Generative Model in Multimodal Cycles/images.zip new file mode 100644 index 0000000000000000000000000000000000000000..8cb3595ea65571092222b377116bcaa872d37bfb --- /dev/null +++ b/2025/DoraCycle_ Domain-Oriented Adaptation of Unified Generative Model in Multimodal Cycles/images.zip @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:dc51ff0e41fecef90a53d5cca413f95144ab743b3d48f2310b7ed3623b7642f1 +size 480181 diff --git a/2025/DoraCycle_ Domain-Oriented Adaptation of Unified Generative Model in Multimodal Cycles/layout.json b/2025/DoraCycle_ Domain-Oriented Adaptation of Unified Generative Model in Multimodal Cycles/layout.json new file mode 100644 index 0000000000000000000000000000000000000000..7271f41d1af86f170018fed4c7ff9b2e85ab5254 --- /dev/null +++ b/2025/DoraCycle_ Domain-Oriented Adaptation of Unified Generative Model in Multimodal Cycles/layout.json @@ -0,0 +1,11041 @@ +{ + "pdf_info": [ + { + "para_blocks": [ + { + "bbox": [ + 87, + 102, + 523, + 140 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 87, + 102, + 523, + 140 + ], + "spans": [ + { + "bbox": [ + 87, + 102, + 523, + 140 + ], + "type": "text", + "content": "DoraCycle: Domain-Oriented Adaptation of Unified Generative Model in Multimodal Cycles" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 195, + 161, + 411, + 190 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 195, + 161, + 411, + 190 + ], + "spans": [ + { + "bbox": [ + 195, + 161, + 411, + 190 + ], + "type": "text", + "content": "Rui Zhao, Weijia Mao, Mike Zheng Shou* Show Lab, National University of Singapore" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 151, + 217, + 200, + 229 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 151, + 217, + 200, + 229 + ], + "spans": [ + { + "bbox": [ + 151, + 217, + 200, + 229 + ], + "type": "text", + "content": "Abstract" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 55, + 243, + 296, + 531 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 243, + 296, + 531 + ], + "spans": [ + { + "bbox": [ + 55, + 243, + 296, + 531 + ], + "type": "text", + "content": "Adapting generative models to specific domains presents an effective solution for satisfying specialized requirements. However, adapting to some complex domains remains challenging, especially when these domains require substantial paired data to capture the targeted distributions. Since unpaired data from a single modality, such as vision or language, is more readily available, we utilize the bidirectional mappings between vision and language learned by the unified generative model to enable training on unpaired data for domain adaptation. Specifically, we propose DoraCycle, which integrates two multimodal cycles: text-to-image-to-text and image-to-text-to-image. The model is optimized through cross-entropy loss computed at the cycle endpoints, where both endpoints share the same modality. This facilitates self-evolution of the model without reliance on annotated text-image pairs. Experimental results demonstrate that for tasks independent of paired knowledge, such as stylization, DoraCycle can effectively adapt the unified model using only unpaired data. For tasks involving new paired knowledge, such as specific identities, a combination of a small set of paired image-text examples and larger-scale unpaired data is sufficient for effective domain-oriented adaptation. The code will be released at https://github.com/showlab/DoraCycle." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 56, + 553, + 135, + 566 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 553, + 135, + 566 + ], + "spans": [ + { + "bbox": [ + 56, + 553, + 135, + 566 + ], + "type": "text", + "content": "1. Introduction" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 55, + 574, + 296, + 694 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 574, + 296, + 694 + ], + "spans": [ + { + "bbox": [ + 55, + 574, + 296, + 694 + ], + "type": "text", + "content": "The adaptation of pre-trained generative models to specific domains is an important aspect of advancing personalized content creation, from stylized media outputs to customized identity generation [47, 62]. However, effectively adapting generative models to complex domains remains challenging, particularly when these domains require extensive amounts of paired data to accurately capture the desired distributions. For instance, learning the visual styles and character identities across a unique movie, which involves understanding multiple characters, their relationships, and di" + } + ] + } + ], + "index": 7 + }, + { + "type": "image", + "bbox": [ + 318, + 217, + 421, + 364 + ], + "blocks": [ + { + "bbox": [ + 318, + 217, + 421, + 364 + ], + "lines": [ + { + "bbox": [ + 318, + 217, + 421, + 364 + ], + "spans": [ + { + "bbox": [ + 318, + 217, + 421, + 364 + ], + "type": "image", + "image_path": "d3956901487918a2e4270e7958c4ead203a6602070952914a9d1371d9b938ffd.jpg" + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 325, + 366, + 410, + 376 + ], + "lines": [ + { + "bbox": [ + 325, + 366, + 410, + 376 + ], + "spans": [ + { + "bbox": [ + 325, + 366, + 410, + 376 + ], + "type": "text", + "content": "(a) Training on paired data" + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 313, + 388, + 555, + 488 + ], + "lines": [ + { + "bbox": [ + 313, + 388, + 555, + 488 + ], + "spans": [ + { + "bbox": [ + 313, + 388, + 555, + 488 + ], + "type": "text", + "content": "Figure 1. Training paradigms for unified generative models. (a) Traditional training involves using paired image-text data and optimizing the unified model with paired losses for both image-to-text (I2T) and text-to-image (T2I) generation tasks. (b) In contrast, the proposed multimodal cycle training framework leverages unpaired images and texts. By using cycle-consistency losses, the unified model learns to maintain consistency between input and output across modalities, enabling adaptation without the need for extensive paired datasets." + } + ] + } + ], + "index": 12, + "angle": 0, + "type": "image_caption" + } + ], + "index": 8 + }, + { + "type": "image", + "bbox": [ + 443, + 217, + 545, + 364 + ], + "blocks": [ + { + "bbox": [ + 443, + 217, + 545, + 364 + ], + "lines": [ + { + "bbox": [ + 443, + 217, + 545, + 364 + ], + "spans": [ + { + "bbox": [ + 443, + 217, + 545, + 364 + ], + "type": "image", + "image_path": "6f1ddd95a07228c8cf7903e1ad20354fa39c4018ee5ae15b67d161cfd0657ef9.jpg" + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 441, + 365, + 549, + 376 + ], + "lines": [ + { + "bbox": [ + 441, + 365, + 549, + 376 + ], + "spans": [ + { + "bbox": [ + 441, + 365, + 549, + 376 + ], + "type": "text", + "content": "(b) Training in Multimodal Cycles" + } + ] + } + ], + "index": 11, + "angle": 0, + "type": "image_caption" + } + ], + "index": 10 + }, + { + "bbox": [ + 313, + 509, + 555, + 569 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 509, + 555, + 569 + ], + "spans": [ + { + "bbox": [ + 313, + 509, + 555, + 569 + ], + "type": "text", + "content": "verse settings, is a highly complex task that demands vast amounts of paired frames and captions data. Collecting such paired data, especially for multimodal tasks involving vision and language, is often laborious and impractical, limiting the potential to adapt generative models at scale." + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 313, + 570, + 556, + 689 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 570, + 556, + 689 + ], + "spans": [ + { + "bbox": [ + 313, + 570, + 556, + 689 + ], + "type": "text", + "content": "High-quality image-text paired data is relatively rare and scarce, but unpaired images and texts are readily available in our daily lives, such as video websites, image platforms, and content from novel websites. Therefore, we aim to explore whether it is possible to adapt generative models to target domains based on unpaired data. To achieve this goal, it is crucial for the model to have an internal capability to align the two modalities to some extent. Fortunately, recent advancements in unified generative models have encouraged us to pursue this direction." + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 313, + 689, + 555, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 689, + 555, + 714 + ], + "spans": [ + { + "bbox": [ + 313, + 689, + 555, + 714 + ], + "type": "text", + "content": "Recent advanced unified generative models [14, 18, 58, 67, 77] have shown great potential in unifying multimodal" + } + ] + } + ], + "index": 15 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "spans": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "type": "text", + "content": "CVF" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 145, + 0, + 496, + 37 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 145, + 0, + 496, + 37 + ], + "spans": [ + { + "bbox": [ + 145, + 0, + 496, + 37 + ], + "type": "text", + "content": "This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 67, + 703, + 146, + 712 + ], + "type": "page_footnote", + "angle": 0, + "lines": [ + { + "bbox": [ + 67, + 703, + 146, + 712 + ], + "spans": [ + { + "bbox": [ + 67, + 703, + 146, + 712 + ], + "type": "text", + "content": "*Corresponding Author." + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "text", + "content": "2835" + } + ] + } + ], + "index": 17 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 0 + }, + { + "para_blocks": [ + { + "bbox": [ + 55, + 72, + 294, + 228 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 72, + 294, + 228 + ], + "spans": [ + { + "bbox": [ + 55, + 72, + 294, + 228 + ], + "type": "text", + "content": "understanding and generation within a single model. These unified models are capable of processing and generating content across different modalities, i.e. vision and language, within a shared framework. By leveraging the bidirectional mappings between vision and language, which are learned by the unified generative model in its pre-training stage, we can map each data from one modality to another and then back to the original modality, as shown in Fig. 1. Through these two mappings, data can be maintained within the same modality, thereby imposing constraints on the deviation introduced in the process. This only requires computing the cross-entropy loss between the data in the same modality, without any paired data supervision." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 55, + 230, + 295, + 410 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 230, + 295, + 410 + ], + "spans": [ + { + "bbox": [ + 55, + 230, + 295, + 410 + ], + "type": "text", + "content": "To this end, we introduce DoraCycle, a framework for adapting unified multimodal generative models to target domains, through cycle-consistent learning with unpaired data. Unlike previous adaptation methods that heavily rely on paired text-image data, the proposed framework leverages the shared latent space of unified models to learn consistent transformations between modalities without requiring paired training examples. Specifically, we design two cycles, i.e. text-to-image-to-text (T cycle) and image-to-text-to-image (I cycle). As shown by Fig. 1 (b), leveraging the pre-trained vision-language alignment of the unified model, each multimodal cycle involves two cross-modality mappings while optimizing in the same modality. This enables calculating loss on unpaired data while implicitly refining cross-modal alignment through the intermediate step." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 55, + 411, + 295, + 628 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 411, + 295, + 628 + ], + "spans": [ + { + "bbox": [ + 55, + 411, + 295, + 628 + ], + "type": "text", + "content": "In practice, since there is no labeled ground-truth, mapping data to another modality requires multi-step inference, such as predicting the next token multiple times for text generation. Allowing all forward steps to participate in gradient backpropagation can lead to a catastrophic gradient explosion. Therefore, we stop gradients during multi-step inference and use the generated data as pseudo labels for the model to forward once again, allowing gradients to propagate. Moreover, we found that since a complete cycle requires the model to forward twice, it can lead to training instability, with the quality of pseudo data generated in the middle being compromised. To enhance the stability of pseudo data generation, we maintain a slowly updated EMA (Exponential Moving Average) [57] model, which is used for inference to generate pseudo data. Additionally, we employ gradient clipping techniques to avoid conflicts in the optimization directions of the two cycles, further increasing training stability." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 55, + 630, + 295, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 630, + 295, + 713 + ], + "spans": [ + { + "bbox": [ + 55, + 630, + 295, + 713 + ], + "type": "text", + "content": "The experiments indicate that for tasks independent of paired knowledge, such as stylization and domain-specific adaptation, DoraCycle can adapt the unified model with only unpaired data, which is both more practical and scalable. For tasks that require new paired knowledge, such as identity-specific adaptation, DoraCycle effectively utilizes small amounts of paired data along with larger-scale" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 313, + 72, + 553, + 180 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 72, + 553, + 180 + ], + "spans": [ + { + "bbox": [ + 313, + 72, + 553, + 180 + ], + "type": "text", + "content": "unpaired data, making it a flexible solution for generative adaptation challenges. We conduct extensive experiments that compare DoraCycle to existing methods, showing that our approach achieves comparable or superior results while significantly reducing the need for paired data. This ability to harness large-scale unpaired data, combined with strategic usage of small paired datasets, makes DoraCycle a feasible solution for personalized multimodal content generation across a wide range of applications." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 313, + 192, + 406, + 205 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 192, + 406, + 205 + ], + "spans": [ + { + "bbox": [ + 313, + 192, + 406, + 205 + ], + "type": "text", + "content": "2. Related Works" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 313, + 213, + 541, + 225 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 213, + 541, + 225 + ], + "spans": [ + { + "bbox": [ + 313, + 213, + 541, + 225 + ], + "type": "text", + "content": "2.1. Multimodal Generation and Understanding" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 313, + 231, + 553, + 410 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 231, + 553, + 410 + ], + "spans": [ + { + "bbox": [ + 313, + 231, + 553, + 410 + ], + "type": "text", + "content": "Generating visual contents from text and describing them through natural language have been extensively studied as core multimodal tasks. Advanced generative models [3, 5, 6, 8, 12, 13, 16, 19, 36, 37, 42, 44, 45, 49, 74-76], such as DALL-E [41, 43], Stable Diffusion [46], demonstrate remarkable generation capabilities, producing high-quality and diverse contents from textual prompts. Meanwhile, image captioning models [24, 26, 27, 60, 61, 63], such as mPLUG [31], and BLIP [33], push the boundaries of visual understanding, generating accurate and context-aware descriptions. Additionally, recent advancements in multimodal large language models [32], such as LLaVA [35], MiniGPT-4 [78], and InstructBLIP [11], have significantly improved the ability to understand and reason about visual content." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 313, + 411, + 554, + 507 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 411, + 554, + 507 + ], + "spans": [ + { + "bbox": [ + 313, + 411, + 554, + 507 + ], + "type": "text", + "content": "Besides the powerful foundational generative models, adapting or customizing them attracts increasing interest, which enables more personalized and specific outputs based on user preferences [7, 9, 17, 20, 29, 30, 38, 68]. Approaches like DreamBooth [47] enable user-specific customization by fine-tuning generative models with personal data, allowing the generation of content tailored to individual needs or preferences." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 313, + 516, + 520, + 528 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 516, + 520, + 528 + ], + "spans": [ + { + "bbox": [ + 313, + 516, + 520, + 528 + ], + "type": "text", + "content": "2.2. Unified Multimodal Generative Models" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 313, + 534, + 554, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 534, + 554, + 713 + ], + "spans": [ + { + "bbox": [ + 313, + 534, + 554, + 713 + ], + "type": "text", + "content": "Unified multimodal generative models aim to bridge the gap between understanding and generation tasks, and integrate vision and language into a single framework, enabling the model to learn shared representations across modalities [1, 2, 14, 18, 53, 56, 64, 65, 67, 69-71, 77]. SEED-X [18] utilizes a unified architecture where visual features extracted from the CLIP ViT encoder [40] are combined with text tokens and fed into a large language model to enable both next-word prediction and image regression tasks. DreamLLM [14] extends the generative capability of large language models by combining multimodal inputs directly into LLMs. Chameleon [58] employs a discrete tokenization approach for both visual and textual inputs, converting all modalities into a unified token space that is processed by a transformer-based architecture. Transfusion [77] in" + } + ] + } + ], + "index": 10 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 749, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 749, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 749, + 315, + 757 + ], + "type": "text", + "content": "2836" + } + ] + } + ], + "index": 11 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 1 + }, + { + "para_blocks": [ + { + "bbox": [ + 55, + 72, + 294, + 192 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 72, + 294, + 192 + ], + "spans": [ + { + "bbox": [ + 55, + 72, + 294, + 192 + ], + "type": "text", + "content": "produces an advanced integration mechanism that focuses on directly fusing visual encoding with language tokens, allowing the model to effectively translate visual information into textual formats while maintaining the semantic integrity of both modalities. Show-o [67] combines autoregressive modeling with a discrete diffusion process, enabling the generation of high-quality outputs that are aligned across modalities. Our work leverages the advancements made by these foundational models and explores how to adapt the foundational model to specific domains." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 55, + 211, + 162, + 224 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 211, + 162, + 224 + ], + "spans": [ + { + "bbox": [ + 55, + 211, + 162, + 224 + ], + "type": "text", + "content": "2.3. Cycle Consistency" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 56, + 232, + 296, + 495 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 232, + 296, + 495 + ], + "spans": [ + { + "bbox": [ + 56, + 232, + 296, + 495 + ], + "type": "text", + "content": "Cycle consistency has been utilized in computer vision and natural language processing as a means to enhance model robustness and consistency [10, 15, 21, 23, 51, 51, 52, 73]. CycleGAN [79] introduced cycle consistency loss to align unpaired image domains, ensuring that mappings between domains (e.g., " + }, + { + "bbox": [ + 56, + 232, + 296, + 495 + ], + "type": "inline_equation", + "content": "\\mathrm{A} \\to \\mathrm{B} \\to \\mathrm{A}" + }, + { + "bbox": [ + 56, + 232, + 296, + 495 + ], + "type": "text", + "content": ") remain consistent. In the field of natural language processing, back-translation employs similar ideas by translating sentences between languages in both directions to improve translation quality [50]. However, the cycle consistency in these works is in a single modality, i.e. vision or language. Recently, ITIT [34] was proposed to utilize cycle consistency to train vision-language generative models. ITIT takes in a mixture of unpaired data and paired data to pre-train the foundational generative model. It is constructed with one image-text encoder and two modality-specific decoders, which operate on the encoded image-text features to generate either text or image tokens. In contrast, we utilized a single unified transformer to parse and predict text and image tokens together. Besides, we focus on adapting pre-trained foundational models to new domains efficiently rather than retraining new foundation models." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 55, + 518, + 111, + 530 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 518, + 111, + 530 + ], + "spans": [ + { + "bbox": [ + 55, + 518, + 111, + 530 + ], + "type": "text", + "content": "3. Method" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 55, + 541, + 295, + 673 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 541, + 295, + 673 + ], + "spans": [ + { + "bbox": [ + 55, + 541, + 295, + 673 + ], + "type": "text", + "content": "The proposed DoraCycle framework, as shown in Fig. 2, is built upon the unified generative model designed for multimodal tasks involving both vision and language [58, 67, 77]. The unified model uses a single transformer to learn bidirectional mappings between vision and language, providing a powerful backbone capable of processing and generating different modalities [67]. For captioning, the model takes in image tokens and predicts corresponding text tokens, while for image generation, it takes in text tokens and predicts image tokens. This versatility makes the unified model well-suited as a base for our proposed framework." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 55, + 677, + 296, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 677, + 296, + 714 + ], + "spans": [ + { + "bbox": [ + 55, + 677, + 296, + 714 + ], + "type": "text", + "content": "In the following sections, we first introduce the design of multimodal cycles, and then discuss the stabilization of optimization, and the balance of two cycles." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 314, + 72, + 425, + 85 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 314, + 72, + 425, + 85 + ], + "spans": [ + { + "bbox": [ + 314, + 72, + 425, + 85 + ], + "type": "text", + "content": "3.1. Multimodal Cycles" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 313, + 89, + 553, + 161 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 89, + 553, + 161 + ], + "spans": [ + { + "bbox": [ + 313, + 89, + 553, + 161 + ], + "type": "text", + "content": "To adapt the unified model for domain-oriented adaptations using unpaired data, we design two multimodal cycles: the Image-to-Text-to-Image Cycle (I Cycle) and the Text-to-Image-to-Text Cycle (T Cycle). Each cycle utilizes data from a single modality, allowing the model to adapt without relying on paired data." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 313, + 161, + 553, + 304 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 161, + 553, + 304 + ], + "spans": [ + { + "bbox": [ + 313, + 161, + 553, + 304 + ], + "type": "text", + "content": "T Cycle: The T cycle training involves transforming an input textual sequence into an image representation and then back into a textual sequence, enforcing consistency between the generated and original text. Specifically, at each training iteration, we begin with an input text sequence " + }, + { + "bbox": [ + 313, + 161, + 553, + 304 + ], + "type": "inline_equation", + "content": "T = [t_l]_{l=1}^L" + }, + { + "bbox": [ + 313, + 161, + 553, + 304 + ], + "type": "text", + "content": ". Conditioned on this, the unified model generates pseudo-paired image tokens " + }, + { + "bbox": [ + 313, + 161, + 553, + 304 + ], + "type": "inline_equation", + "content": "I'" + }, + { + "bbox": [ + 313, + 161, + 553, + 304 + ], + "type": "text", + "content": ", representing the visual interpretation of the input text. The generated image tokens are then subjected to a random masking operation, denoted as " + }, + { + "bbox": [ + 313, + 161, + 553, + 304 + ], + "type": "inline_equation", + "content": "I_M'" + }, + { + "bbox": [ + 313, + 161, + 553, + 304 + ], + "type": "text", + "content": ", where a subset of the tokens is masked. The unified model is then called to reconstruct these masked tokens to form the complete synthesized image " + }, + { + "bbox": [ + 313, + 161, + 553, + 304 + ], + "type": "inline_equation", + "content": "\\tilde{I}" + }, + { + "bbox": [ + 313, + 161, + 553, + 304 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 313, + 304, + 553, + 365 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 304, + 553, + 365 + ], + "spans": [ + { + "bbox": [ + 313, + 304, + 553, + 365 + ], + "type": "text", + "content": "In the second half of the cycle, conditioning on image " + }, + { + "bbox": [ + 313, + 304, + 553, + 365 + ], + "type": "inline_equation", + "content": "\\tilde{I}" + }, + { + "bbox": [ + 313, + 304, + 553, + 365 + ], + "type": "text", + "content": ", the model generates the corresponding text sequence. The objective of the T cycle is to enforce cycle consistency between the generated text and the original input text " + }, + { + "bbox": [ + 313, + 304, + 553, + 365 + ], + "type": "inline_equation", + "content": "T" + }, + { + "bbox": [ + 313, + 304, + 553, + 365 + ], + "type": "text", + "content": ". The T cycle consistency loss is defined as follows:" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 324, + 384, + 555, + 416 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 324, + 384, + 555, + 416 + ], + "spans": [ + { + "bbox": [ + 324, + 384, + 555, + 416 + ], + "type": "interline_equation", + "content": "\\mathcal {L} _ {T C} = - \\mathbb {E} _ {T \\in \\mathcal {D} t e x t} \\left[ \\sum_ {l = 1} ^ {L} \\log p \\left(t _ {l} \\mid \\tilde {I}, t _ {0}, \\dots , t _ {l - 1}\\right) \\right], \\tag {1}", + "image_path": "19968ff7e242efc73386fe9c4ffa20c3011ba9eb09223adba261b01c81202a4a.jpg" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 313, + 425, + 553, + 447 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 425, + 553, + 447 + ], + "spans": [ + { + "bbox": [ + 313, + 425, + 553, + 447 + ], + "type": "text", + "content": "where the " + }, + { + "bbox": [ + 313, + 425, + 553, + 447 + ], + "type": "inline_equation", + "content": "\\mathcal{D}" + }, + { + "bbox": [ + 313, + 425, + 553, + 447 + ], + "type": "text", + "content": " text is the set of text samples from the target domain." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 313, + 449, + 555, + 556 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 449, + 555, + 556 + ], + "spans": [ + { + "bbox": [ + 313, + 449, + 555, + 556 + ], + "type": "text", + "content": "I Cycle: The I cycle training begins with an input image, which is subsequently mapped to a textual representation and then transformed back to an image, enforcing consistency between the generated image and the original image tokens. At each training iteration, we start with input image tokens " + }, + { + "bbox": [ + 313, + 449, + 555, + 556 + ], + "type": "inline_equation", + "content": "I" + }, + { + "bbox": [ + 313, + 449, + 555, + 556 + ], + "type": "text", + "content": ". The unified model is used to synthesize pseudopaired text tokens " + }, + { + "bbox": [ + 313, + 449, + 555, + 556 + ], + "type": "inline_equation", + "content": "T'" + }, + { + "bbox": [ + 313, + 449, + 555, + 556 + ], + "type": "text", + "content": ", representing the textual description of the image. We then use " + }, + { + "bbox": [ + 313, + 449, + 555, + 556 + ], + "type": "inline_equation", + "content": "T'" + }, + { + "bbox": [ + 313, + 449, + 555, + 556 + ], + "type": "text", + "content": " in conjunction with the input image token " + }, + { + "bbox": [ + 313, + 449, + 555, + 556 + ], + "type": "inline_equation", + "content": "I" + }, + { + "bbox": [ + 313, + 449, + 555, + 556 + ], + "type": "text", + "content": " to predict the reconstructed text tokens " + }, + { + "bbox": [ + 313, + 449, + 555, + 556 + ], + "type": "inline_equation", + "content": "\\tilde{T}" + }, + { + "bbox": [ + 313, + 449, + 555, + 556 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 313, + 556, + 553, + 617 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 556, + 553, + 617 + ], + "spans": [ + { + "bbox": [ + 313, + 556, + 553, + 617 + ], + "type": "text", + "content": "In the second half of the cycle, we pass the masked image tokens " + }, + { + "bbox": [ + 313, + 556, + 553, + 617 + ], + "type": "inline_equation", + "content": "I_M" + }, + { + "bbox": [ + 313, + 556, + 553, + 617 + ], + "type": "text", + "content": " and the text " + }, + { + "bbox": [ + 313, + 556, + 553, + 617 + ], + "type": "inline_equation", + "content": "\\tilde{T}" + }, + { + "bbox": [ + 313, + 556, + 553, + 617 + ], + "type": "text", + "content": " through the model to regenerate the masked image tokens. The cycle enforces consistency between the reconstructed and the original image tokens. The loss for enforcing I cycle consistency is given by:" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 330, + 635, + 555, + 662 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 330, + 635, + 555, + 662 + ], + "spans": [ + { + "bbox": [ + 330, + 635, + 555, + 662 + ], + "type": "interline_equation", + "content": "\\mathcal {L} _ {I C} = - \\mathbb {E} _ {I \\in \\mathcal {D} _ {i m a g e}} \\left[ \\sum_ {\\forall k: m _ {k} = 1} \\log p (i _ {k} | I _ {M}, \\tilde {T}) \\right], \\tag {2}", + "image_path": "bacb7466f4171b9448c0ae41c9623dfa6213cf2ced928452e244648e4e0636d5.jpg" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 313, + 666, + 553, + 689 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 666, + 553, + 689 + ], + "spans": [ + { + "bbox": [ + 313, + 666, + 553, + 689 + ], + "type": "text", + "content": "where the " + }, + { + "bbox": [ + 313, + 666, + 553, + 689 + ], + "type": "inline_equation", + "content": "D" + }, + { + "bbox": [ + 313, + 666, + 553, + 689 + ], + "type": "text", + "content": " image is the set of image samples from the target domain, which are unpaired with the text samples." + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 313, + 689, + 553, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 689, + 553, + 714 + ], + "spans": [ + { + "bbox": [ + 313, + 689, + 553, + 714 + ], + "type": "text", + "content": "By leveraging these two cycles, our framework forces the model to refine its generative understanding of both image" + } + ] + } + ], + "index": 16 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "text", + "content": "2837" + } + ] + } + ], + "index": 17 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 2 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 60, + 73, + 553, + 223 + ], + "blocks": [ + { + "bbox": [ + 60, + 73, + 553, + 223 + ], + "lines": [ + { + "bbox": [ + 60, + 73, + 553, + 223 + ], + "spans": [ + { + "bbox": [ + 60, + 73, + 553, + 223 + ], + "type": "image", + "image_path": "167d82eea73a700f58885a073203595172afb40ea966537b2946d4ac868e18dd.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 55, + 232, + 555, + 255 + ], + "lines": [ + { + "bbox": [ + 55, + 232, + 555, + 255 + ], + "spans": [ + { + "bbox": [ + 55, + 232, + 555, + 255 + ], + "type": "text", + "content": "Figure 2. The overview of T cycle (text-to-image-to-text) of the proposed DoraCycle. The I cycle is similar and is omitted in the figure for brevity." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "bbox": [ + 54, + 275, + 295, + 312 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 54, + 275, + 295, + 312 + ], + "spans": [ + { + "bbox": [ + 54, + 275, + 295, + 312 + ], + "type": "text", + "content": "and text representations, ensuring consistency between the input and output while effectively leveraging unpaired data to adapt the unified towards the target domain." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 54, + 312, + 295, + 479 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 54, + 312, + 295, + 479 + ], + "spans": [ + { + "bbox": [ + 54, + 312, + 295, + 479 + ], + "type": "text", + "content": "Efficient Training: In the intermediate steps of both cycles, generating the middle representation (i.e., captions or images) requires multiple forward passes. This is because the generation process involves either predicting the next tokens or the masked tokens multiple times. Backpropagating gradients through all these steps is computationally prohibitive. Thus, we first generate the intermediate results using the model in inference mode as pseudo-paired data, which are then used as the ground truth in the teacher-forcing scheme [54, 55] for the first half of the cycles. In this way, we reduce the number of forward passes to two, i.e. one for generating the middle result and one for the final output, thus making the overall training process more memory efficient." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 55, + 480, + 296, + 529 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 480, + 296, + 529 + ], + "spans": [ + { + "bbox": [ + 55, + 480, + 296, + 529 + ], + "type": "text", + "content": "Token Differentiability: Since the intermediate outputs in each cycle are discrete tokens, which can not directly propagate gradients, we employ the Gumbel-Softmax [28] to make these token representations differentiable." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 55, + 539, + 194, + 552 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 539, + 194, + 552 + ], + "spans": [ + { + "bbox": [ + 55, + 539, + 194, + 552 + ], + "type": "text", + "content": "3.2. Stabilizing Optimization" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 55, + 557, + 295, + 641 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 557, + 295, + 641 + ], + "spans": [ + { + "bbox": [ + 55, + 557, + 295, + 641 + ], + "type": "text", + "content": "Each cycle involves the same unified model twice in the forward pass, which leads to optimization instabilities. To stabilize the training process, we adopt the Exponential Moving Average (EMA) training technique [57]. Specifically, we maintain a shadow version of the model, referred to as the EMA model, which is updated using an exponentially decaying average of the parameters of the main model." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 109, + 657, + 295, + 669 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 109, + 657, + 295, + 669 + ], + "spans": [ + { + "bbox": [ + 109, + 657, + 295, + 669 + ], + "type": "interline_equation", + "content": "\\theta_ {\\mathrm {E M A}} \\leftarrow \\alpha \\theta_ {\\mathrm {E M A}} + (1 - \\alpha) \\theta_ {\\text {m i n}}, \\tag {3}", + "image_path": "23fe6454057edf761702bbe305e329953546eb254f0a5cd30bfc2303a909de66.jpg" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 55, + 677, + 296, + 712 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 677, + 296, + 712 + ], + "spans": [ + { + "bbox": [ + 55, + 677, + 296, + 712 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 55, + 677, + 296, + 712 + ], + "type": "inline_equation", + "content": "\\alpha" + }, + { + "bbox": [ + 55, + 677, + 296, + 712 + ], + "type": "text", + "content": " is a decay factor (set to 0.999) that controls the update rate, and " + }, + { + "bbox": [ + 55, + 677, + 296, + 712 + ], + "type": "inline_equation", + "content": "\\theta_{\\mathrm{main}}" + }, + { + "bbox": [ + 55, + 677, + 296, + 712 + ], + "type": "text", + "content": " represents the parameters of the main model." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 313, + 275, + 555, + 384 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 275, + 555, + 384 + ], + "spans": [ + { + "bbox": [ + 313, + 275, + 555, + 384 + ], + "type": "text", + "content": "In each training step, the EMA version of the model is used to generate the intermediate representation tokens (e.g., pseudo image or text tokens) which serve as pseudo ground truth during training. By using these stable targets from the slower-evolving EMA model, we can mitigate the risks of optimization instability. The main model is thus able to learn from more consistent and reliable intermediate targets, rather than being affected by fluctuations during the early stages of training." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 313, + 393, + 440, + 406 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 393, + 440, + 406 + ], + "spans": [ + { + "bbox": [ + 313, + 393, + 440, + 406 + ], + "type": "text", + "content": "3.3. Balancing Two Cycles" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 313, + 411, + 554, + 496 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 411, + 554, + 496 + ], + "spans": [ + { + "bbox": [ + 313, + 411, + 554, + 496 + ], + "type": "text", + "content": "We observe that the T cycle tends to converge faster than the I cycle, primarily because textual data is inherently one-dimensional and simpler to learn compared to images. This imbalance in optimization leads to a kind of collapse of the model, where it tends to generate irrelevant but self-consistent captions for images, ultimately degrading the image-text alignment capability." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 313, + 497, + 554, + 582 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 497, + 554, + 582 + ], + "spans": [ + { + "bbox": [ + 313, + 497, + 554, + 582 + ], + "type": "text", + "content": "To address this problem, we make the gradients of the T cycle orthogonal to those of the I cycle, thus preventing interference. This is achieved by modifying the gradients using gradient surgery [72]. Let " + }, + { + "bbox": [ + 313, + 497, + 554, + 582 + ], + "type": "inline_equation", + "content": "g_{T}" + }, + { + "bbox": [ + 313, + 497, + 554, + 582 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 313, + 497, + 554, + 582 + ], + "type": "inline_equation", + "content": "g_{I}" + }, + { + "bbox": [ + 313, + 497, + 554, + 582 + ], + "type": "text", + "content": " represent the gradients of the T cycle and the I cycle, respectively. We project " + }, + { + "bbox": [ + 313, + 497, + 554, + 582 + ], + "type": "inline_equation", + "content": "g_{T}" + }, + { + "bbox": [ + 313, + 497, + 554, + 582 + ], + "type": "text", + "content": " onto the orthogonal complement of " + }, + { + "bbox": [ + 313, + 497, + 554, + 582 + ], + "type": "inline_equation", + "content": "g_{I}" + }, + { + "bbox": [ + 313, + 497, + 554, + 582 + ], + "type": "text", + "content": " to obtain the modified gradient " + }, + { + "bbox": [ + 313, + 497, + 554, + 582 + ], + "type": "inline_equation", + "content": "g_{T}^{\\prime}" + }, + { + "bbox": [ + 313, + 497, + 554, + 582 + ], + "type": "text", + "content": ", which is defined as:" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 387, + 592, + 555, + 616 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 387, + 592, + 555, + 616 + ], + "spans": [ + { + "bbox": [ + 387, + 592, + 555, + 616 + ], + "type": "interline_equation", + "content": "g _ {T} ^ {\\prime} = g _ {T} - \\frac {g _ {T} \\cdot g _ {I}}{g _ {I} \\cdot g _ {I}} g _ {I}, \\tag {4}", + "image_path": "8ab27cacd6ee81fed0ac7d2e80c0a90cc802b58bd8ab06aa993e35b30e9276a7.jpg" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 313, + 622, + 554, + 645 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 622, + 554, + 645 + ], + "spans": [ + { + "bbox": [ + 313, + 622, + 554, + 645 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 313, + 622, + 554, + 645 + ], + "type": "inline_equation", + "content": "g_{T} \\cdot g_{I}" + }, + { + "bbox": [ + 313, + 622, + 554, + 645 + ], + "type": "text", + "content": " denotes the dot product between the gradients of the T and I cycle." + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 313, + 647, + 554, + 681 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 647, + 554, + 681 + ], + "spans": [ + { + "bbox": [ + 313, + 647, + 554, + 681 + ], + "type": "text", + "content": "Additionally, we reweight the losses to further balance the learning between the I and T cycles. The final loss function is as," + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 393, + 685, + 553, + 697 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 393, + 685, + 553, + 697 + ], + "spans": [ + { + "bbox": [ + 393, + 685, + 553, + 697 + ], + "type": "interline_equation", + "content": "\\mathcal {L} = \\mathcal {L} _ {I C} + \\beta \\mathcal {L} _ {T C}, \\tag {5}", + "image_path": "6082e14eebedb86b226259c15795e4f6b76fa7b8efb56c5b8635a4ab4936338e.jpg" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 313, + 701, + 497, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 701, + 497, + 714 + ], + "spans": [ + { + "bbox": [ + 313, + 701, + 497, + 714 + ], + "type": "text", + "content": "where the " + }, + { + "bbox": [ + 313, + 701, + 497, + 714 + ], + "type": "inline_equation", + "content": "\\beta" + }, + { + "bbox": [ + 313, + 701, + 497, + 714 + ], + "type": "text", + "content": " is the weight of the T cycle loss." + } + ] + } + ], + "index": 17 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "text", + "content": "2838" + } + ] + } + ], + "index": 18 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 3 + }, + { + "para_blocks": [ + { + "bbox": [ + 55, + 71, + 137, + 85 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 71, + 137, + 85 + ], + "spans": [ + { + "bbox": [ + 55, + 71, + 137, + 85 + ], + "type": "text", + "content": "4. Experiments" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 55, + 91, + 188, + 104 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 91, + 188, + 104 + ], + "spans": [ + { + "bbox": [ + 55, + 91, + 188, + 104 + ], + "type": "text", + "content": "4.1. Implementation Details" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 55, + 110, + 296, + 253 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 110, + 296, + 253 + ], + "spans": [ + { + "bbox": [ + 55, + 110, + 296, + 253 + ], + "type": "text", + "content": "To the best of our knowledge, Show-o [67] is currently the only fully open-source unified generative model with complete pre-trained weights and training code, including both its understanding and generation capabilities. Therefore, we base DoraCycle on Show-o and conduct experiments accordingly. The base model is a unified transformer model that performs understanding and image generation by predicting discrete textual and visual tokens. We insert trainable low-rank adaptations (LoRA) [25] modules into the Q projection and V projection of the attention layers from layers 7 to 24. The LoRA rank is set to 32. The " + }, + { + "bbox": [ + 55, + 110, + 296, + 253 + ], + "type": "inline_equation", + "content": "\\beta" + }, + { + "bbox": [ + 55, + 110, + 296, + 253 + ], + "type": "text", + "content": " is set to 0.1 to balance the optimization of two cycles." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 55, + 254, + 296, + 350 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 254, + 296, + 350 + ], + "spans": [ + { + "bbox": [ + 55, + 254, + 296, + 350 + ], + "type": "text", + "content": "The training of DoraCycle is performed on 8 NVIDIA H100 GPUs with mixed precision enabled for memory efficiency. We set the batch size to 32, with each cycle taking half of the batch when both cycles are being optimized simultaneously. The learning rate is set to " + }, + { + "bbox": [ + 55, + 254, + 296, + 350 + ], + "type": "inline_equation", + "content": "1e^{-4}" + }, + { + "bbox": [ + 55, + 254, + 296, + 350 + ], + "type": "text", + "content": " with a cosine annealing schedule. The optimizer is AdamW with weight decay of " + }, + { + "bbox": [ + 55, + 254, + 296, + 350 + ], + "type": "inline_equation", + "content": "1e^{-2}" + }, + { + "bbox": [ + 55, + 254, + 296, + 350 + ], + "type": "text", + "content": ". Additionally, EMA is employed to stabilize the training process, as described in Section 3.2." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 55, + 360, + 221, + 372 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 360, + 221, + 372 + ], + "spans": [ + { + "bbox": [ + 55, + 360, + 221, + 372 + ], + "type": "text", + "content": "4.2. Domain-Oriented Adaptations" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 55, + 377, + 295, + 474 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 377, + 295, + 474 + ], + "spans": [ + { + "bbox": [ + 55, + 377, + 295, + 474 + ], + "type": "text", + "content": "Unpaired Training: For tasks that do not require strongly related paired knowledge, our DoraCycle can fully learn the target domain using unpaired data. For example, to learn the cyberpunk style, we collected 300 cyberpunk-style images as input for the I cycle, and used the text data from the base model pre-training dataset [4] for the T cycle, with the keyword \"cyberpunk style\" automatically injected into text, prompting the model about the target style we want." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 55, + 474, + 296, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 474, + 296, + 713 + ], + "spans": [ + { + "bbox": [ + 55, + 474, + 296, + 713 + ], + "type": "text", + "content": "The experimental results are shown in Fig. 3. Given the same text prompt to generate cyberpunk-style images, Fig. 3 (a) shows the images generated by the base model without additional training. It can be observed that the base model adds some cyberpunk elements, such as neon lights, but the overall atmosphere does not align well with the desired style. Fig. 3 (d) shows the images generated by the adapted model trained with DoraCycle, which aligns well with the target style. Traditional text-to-image customization or adaptation methods, such as DreamBooth [48], rely on paired data for training. Therefore, we simulate user-created paired data by annotating the collected images with captions, and split them into two groups. One group contained only 10 paired examples, which is an acceptable workload for users, while the other group contained captions for all 300 images, which would be labor-intensive and impractical for users. The images generated by the model that trained on 10 paired examples are shown in Fig. 3 (b). It struggled to produce good stylized images, likely because the combination of indoor bookshelves with the cyberpunk" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 317, + 71, + 551, + 94 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 71, + 551, + 94 + ], + "spans": [ + { + "bbox": [ + 317, + 71, + 551, + 94 + ], + "type": "text", + "content": " The image presents a scene of a room with a brick wall and a large bookshelf filled with various books and electronics. The bookshelf, which is black, stands against the brick wall and spans almost the entire height of the wall. It is divided into three shelves, each holding different items. On the top shelf, there are several books neatly arranged. In front of the bookshelf, there's a black leather couch positioned on one side of the room. A potted plant can be seen as well." + } + ] + } + ], + "index": 7 + }, + { + "type": "image", + "bbox": [ + 317, + 105, + 375, + 164 + ], + "blocks": [ + { + "bbox": [ + 317, + 105, + 375, + 164 + ], + "lines": [ + { + "bbox": [ + 317, + 105, + 375, + 164 + ], + "spans": [ + { + "bbox": [ + 317, + 105, + 375, + 164 + ], + "type": "image", + "image_path": "2eca2e870a5dda13242b0e49d00f3556dc8424685c88ef5f06fa73192142eba8.jpg" + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 324, + 167, + 368, + 186 + ], + "lines": [ + { + "bbox": [ + 324, + 167, + 368, + 186 + ], + "spans": [ + { + "bbox": [ + 324, + 167, + 368, + 186 + ], + "type": "text", + "content": "(a) Base model (w/o training)" + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "image_caption" + } + ], + "index": 9 + }, + { + "type": "image", + "bbox": [ + 376, + 105, + 432, + 164 + ], + "blocks": [ + { + "bbox": [ + 430, + 95, + 479, + 102 + ], + "lines": [ + { + "bbox": [ + 430, + 95, + 479, + 102 + ], + "spans": [ + { + "bbox": [ + 430, + 95, + 479, + 102 + ], + "type": "text", + "content": "Image-to-Text" + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 376, + 105, + 432, + 164 + ], + "lines": [ + { + "bbox": [ + 376, + 105, + 432, + 164 + ], + "spans": [ + { + "bbox": [ + 376, + 105, + 432, + 164 + ], + "type": "image", + "image_path": "6ae94d2a6438dbf1d6ddfcbfd6156f15ffed35e7eb9eece5f0896d9cf26e1e94.jpg" + } + ] + } + ], + "index": 11, + "angle": 0, + "type": "image_body" + } + ], + "index": 11 + }, + { + "type": "image", + "bbox": [ + 433, + 105, + 493, + 164 + ], + "blocks": [ + { + "bbox": [ + 433, + 105, + 493, + 164 + ], + "lines": [ + { + "bbox": [ + 433, + 105, + 493, + 164 + ], + "spans": [ + { + "bbox": [ + 433, + 105, + 493, + 164 + ], + "type": "image", + "image_path": "870c95c63c517b35aa72bdf34ec2fed87f0ce48eac5e8556631ef648fa06113f.jpg" + } + ] + } + ], + "index": 13, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 438, + 168, + 488, + 186 + ], + "lines": [ + { + "bbox": [ + 438, + 168, + 488, + 186 + ], + "spans": [ + { + "bbox": [ + 438, + 168, + 488, + 186 + ], + "type": "text", + "content": "(c) Paired trained (300 pairs)" + } + ] + } + ], + "index": 14, + "angle": 0, + "type": "image_caption" + } + ], + "index": 13 + }, + { + "type": "image", + "bbox": [ + 494, + 105, + 553, + 164 + ], + "blocks": [ + { + "bbox": [ + 494, + 105, + 553, + 164 + ], + "lines": [ + { + "bbox": [ + 494, + 105, + 553, + 164 + ], + "spans": [ + { + "bbox": [ + 494, + 105, + 553, + 164 + ], + "type": "image", + "image_path": "23baec640737327651913a6e88984467ed8188cca2fc347ddeac4dac0e94a5a9.jpg" + } + ] + } + ], + "index": 15, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 489, + 168, + 552, + 186 + ], + "lines": [ + { + "bbox": [ + 489, + 168, + 552, + 186 + ], + "spans": [ + { + "bbox": [ + 489, + 168, + 552, + 186 + ], + "type": "text", + "content": "(d) Unpaired trained (DoraCycle)" + } + ] + } + ], + "index": 16, + "angle": 0, + "type": "image_caption" + } + ], + "index": 15 + }, + { + "type": "image", + "bbox": [ + 316, + 193, + 385, + 262 + ], + "blocks": [ + { + "bbox": [ + 378, + 168, + 430, + 186 + ], + "lines": [ + { + "bbox": [ + 378, + 168, + 430, + 186 + ], + "spans": [ + { + "bbox": [ + 378, + 168, + 430, + 186 + ], + "type": "text", + "content": "(b) Paired trained (10 pairs)" + } + ] + } + ], + "index": 12, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 316, + 193, + 385, + 262 + ], + "lines": [ + { + "bbox": [ + 316, + 193, + 385, + 262 + ], + "spans": [ + { + "bbox": [ + 316, + 193, + 385, + 262 + ], + "type": "image", + "image_path": "f89508f9d5cf179e58c291eb433cd782af6c8cd0cdaebd46e1a82c880c87961e.jpg" + } + ] + } + ], + "index": 17, + "angle": 0, + "type": "image_body" + } + ], + "index": 17 + }, + { + "type": "image", + "bbox": [ + 363, + 263, + 388, + 287 + ], + "blocks": [ + { + "bbox": [ + 323, + 275, + 359, + 283 + ], + "lines": [ + { + "bbox": [ + 323, + 275, + 359, + 283 + ], + "spans": [ + { + "bbox": [ + 323, + 275, + 359, + 283 + ], + "type": "text", + "content": "Image-to-Text" + } + ] + } + ], + "index": 18, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 363, + 263, + 388, + 287 + ], + "lines": [ + { + "bbox": [ + 363, + 263, + 388, + 287 + ], + "spans": [ + { + "bbox": [ + 363, + 263, + 388, + 287 + ], + "type": "image", + "image_path": "f9169b06b29a9f339c4934169bdad9889bb1b4442ce66eac6e94d26ef9b7f909.jpg" + } + ] + } + ], + "index": 19, + "angle": 0, + "type": "image_body" + } + ], + "index": 19 + }, + { + "bbox": [ + 389, + 193, + 477, + 289 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 389, + 193, + 477, + 289 + ], + "spans": [ + { + "bbox": [ + 389, + 193, + 477, + 289 + ], + "type": "text", + "content": " The image captures a scene on a city street. Dominating the foreground are two mannequins, each adorned in full Batman and Iron Man costumes. The Mannequin on the left is dressed in a black Batman costume, complete with a cape and mask, exuding an air of mystery and power. On the right, the other Mannequin is clad in a red Iron Man costume, featuring a helmet with a faceplate and red gloves, embodying the iconic superhero's persona. These mannequins stand on a gray sidewalk, their vibrant costumes contrasting against the monotone backdrop. Behind them, the city continues its usual hustle and bustle. Cars are parked along the street, their metallic bodies gleaming under the city lights. Buildings rise in the background, their windows reflecting the life of the city." + } + ] + } + ], + "index": 20 + }, + { + "type": "image", + "bbox": [ + 484, + 193, + 553, + 262 + ], + "blocks": [ + { + "bbox": [ + 484, + 193, + 553, + 262 + ], + "lines": [ + { + "bbox": [ + 484, + 193, + 553, + 262 + ], + "spans": [ + { + "bbox": [ + 484, + 193, + 553, + 262 + ], + "type": "image", + "image_path": "65ea5a22b8d591bb2d788c1ec0cb374a0747b1588d5d4cefef16b99599175495.jpg" + } + ] + } + ], + "index": 21, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 313, + 313, + 555, + 401 + ], + "lines": [ + { + "bbox": [ + 313, + 313, + 555, + 401 + ], + "spans": [ + { + "bbox": [ + 313, + 313, + 555, + 401 + ], + "type": "text", + "content": "Figure 3. Domain-oriented adaptation with different training setups. (a) Image generated by the base model without training for adoption. (b) Image generated by the model trained with 10 paired image-text samples. (c) Image generated by the model trained with 300 paired image-text samples. (d) Image generated by the model trained by DoraCycle on only unpaired data. (e) Image-to-Text-to-Image translation performed by the adapted model trained by DoraCycle." + } + ] + } + ], + "index": 24, + "angle": 0, + "type": "image_caption" + } + ], + "index": 21 + }, + { + "type": "image", + "bbox": [ + 484, + 263, + 545, + 287 + ], + "blocks": [ + { + "bbox": [ + 484, + 263, + 545, + 287 + ], + "lines": [ + { + "bbox": [ + 484, + 263, + 545, + 287 + ], + "spans": [ + { + "bbox": [ + 484, + 263, + 545, + 287 + ], + "type": "image", + "image_path": "1bd13033b221a6a82e57404b8b11a3178f44a8db4074f9bd7fad93b6f784cc5b.jpg" + } + ] + } + ], + "index": 22, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 333, + 294, + 531, + 304 + ], + "lines": [ + { + "bbox": [ + 333, + 294, + 531, + 304 + ], + "spans": [ + { + "bbox": [ + 333, + 294, + 531, + 304 + ], + "type": "text", + "content": "(e) Image-to-Text-to-Image translation by the adapted unified model." + } + ] + } + ], + "index": 23, + "angle": 0, + "type": "image_caption" + } + ], + "index": 22 + }, + { + "bbox": [ + 313, + 424, + 554, + 496 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 424, + 554, + 496 + ], + "spans": [ + { + "bbox": [ + 313, + 424, + 554, + 496 + ], + "type": "text", + "content": "style is too novel for the model to generalize well from limited paired data. The images generated by the model trained on 300 paired examples are shown in Fig. 3 (c), which have better outputs. In contrast, the model trained using DoraCycle does not require manual captioning, significantly reducing the workload for users." + } + ] + } + ], + "index": 25 + }, + { + "bbox": [ + 313, + 497, + 554, + 652 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 497, + 554, + 652 + ], + "spans": [ + { + "bbox": [ + 313, + 497, + 554, + 652 + ], + "type": "text", + "content": "Fig. 3 (e) illustrates the adapted model trained by DoraCyle maintains semantic consistency through image-to-text-to-image translation. The input image is transformed into a textual description and then reconstructed into an image. The result shows that the adapted model successfully captures and retains the key visual components in the original image throughout the multimodal cycle. Notably, the identity of the characters and the details of the environment are all preserved, indicating effective bidirectional understanding and generation capabilities in the target domain. Furthermore, the newly generated image incorporates styles learned from the target domain, demonstrating the generalization of the learned knowledge to images in the wild." + } + ] + } + ], + "index": 26 + }, + { + "bbox": [ + 313, + 654, + 554, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 654, + 554, + 713 + ], + "spans": [ + { + "bbox": [ + 313, + 654, + 554, + 713 + ], + "type": "text", + "content": "Learning Paired Knowledge: For tasks that require learning some paired knowledge, such as associating an identity name with its visual appearance, DoraCycle can incorporate a small amount of paired data to learn such associations while leveraging a large amount of unpaired" + } + ] + } + ], + "index": 27 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "text", + "content": "2839" + } + ] + } + ], + "index": 28 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 4 + }, + { + "para_blocks": [ + { + "bbox": [ + 249, + 73, + 366, + 84 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 249, + 73, + 366, + 84 + ], + "spans": [ + { + "bbox": [ + 249, + 73, + 366, + 84 + ], + "type": "text", + "content": "Domain 1: Black Myth Wukong" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 65, + 101, + 75, + 178 + ], + "type": "title", + "angle": 270, + "lines": [ + { + "bbox": [ + 65, + 101, + 75, + 178 + ], + "spans": [ + { + "bbox": [ + 65, + 101, + 75, + 178 + ], + "type": "text", + "content": "Text-to-Image" + } + ] + } + ], + "index": 1 + }, + { + "type": "image", + "bbox": [ + 78, + 86, + 154, + 163 + ], + "blocks": [ + { + "bbox": [ + 78, + 86, + 154, + 163 + ], + "lines": [ + { + "bbox": [ + 78, + 86, + 154, + 163 + ], + "spans": [ + { + "bbox": [ + 78, + 86, + 154, + 163 + ], + "type": "image", + "image_path": "3c4f64bbba733eb1cfc9c7f0e24405ab2a625f1866662e47be5685bffa574996.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + } + ], + "index": 2 + }, + { + "bbox": [ + 78, + 166, + 152, + 204 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 78, + 166, + 152, + 204 + ], + "spans": [ + { + "bbox": [ + 78, + 166, + 152, + 204 + ], + "type": "text", + "content": "Destined One, wears an intricately detailed golden armor with elaborate ornaments and engravings. A golden crown-like adornment sits atop his head. In the background, distant mountains and a serene sky can be seen." + } + ] + } + ], + "index": 3 + }, + { + "type": "image", + "bbox": [ + 156, + 86, + 233, + 163 + ], + "blocks": [ + { + "bbox": [ + 156, + 86, + 233, + 163 + ], + "lines": [ + { + "bbox": [ + 156, + 86, + 233, + 163 + ], + "spans": [ + { + "bbox": [ + 156, + 86, + 233, + 163 + ], + "type": "image", + "image_path": "92f2055398260c2f3c80e6ade333da387b1ef74b3ef192a8826c47608c0e5d7d.jpg" + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_body" + } + ], + "index": 4 + }, + { + "bbox": [ + 157, + 165, + 230, + 199 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 157, + 165, + 230, + 199 + ], + "spans": [ + { + "bbox": [ + 157, + 165, + 230, + 199 + ], + "type": "text", + "content": "Lingxuizi's glowing red eyes and bared fangs indicate its aggressive nature. It crouches low to the ground, its large hand-like paw spread over ancient stone paving surrounded by patches of grass." + } + ] + } + ], + "index": 5 + }, + { + "type": "image", + "bbox": [ + 235, + 86, + 312, + 163 + ], + "blocks": [ + { + "bbox": [ + 235, + 86, + 312, + 163 + ], + "lines": [ + { + "bbox": [ + 235, + 86, + 312, + 163 + ], + "spans": [ + { + "bbox": [ + 235, + 86, + 312, + 163 + ], + "type": "image", + "image_path": "f57acc46503ff6b286c6582823f92d79962e1e9a3896ba3e16f6faa381106488.jpg" + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "image_body" + } + ], + "index": 6 + }, + { + "bbox": [ + 236, + 165, + 309, + 205 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 236, + 165, + 309, + 205 + ], + "spans": [ + { + "bbox": [ + 236, + 165, + 309, + 205 + ], + "type": "text", + "content": "Keeper grips a gnarled wooden staff for support as he walks through a wooded area. His posture is stooped, and his facial features are mostly obscured by his thick beard. The scene is set in a forest, with a textured ground of leaves and dirt." + } + ] + } + ], + "index": 7 + }, + { + "type": "image", + "bbox": [ + 313, + 86, + 390, + 163 + ], + "blocks": [ + { + "bbox": [ + 313, + 86, + 390, + 163 + ], + "lines": [ + { + "bbox": [ + 313, + 86, + 390, + 163 + ], + "spans": [ + { + "bbox": [ + 313, + 86, + 390, + 163 + ], + "type": "image", + "image_path": "e56ab250c53c79fd423d31f583501da649fb5fbdfd242f88529ab40aac4a9a71.jpg" + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "image_body" + } + ], + "index": 8 + }, + { + "bbox": [ + 315, + 165, + 387, + 209 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 315, + 165, + 387, + 209 + ], + "spans": [ + { + "bbox": [ + 315, + 165, + 387, + 209 + ], + "type": "text", + "content": "Zhu Bajie stands confidently in the snow, gripping his staff with a sly smile on his boat-like face, beads hanging around his neck. Beside him lies the majestic, wounded form of Kang-Jin Loong, a silver-scaled dragon with intricate, ethereal horns and whiskers." + } + ] + } + ], + "index": 9 + }, + { + "type": "image", + "bbox": [ + 391, + 86, + 468, + 163 + ], + "blocks": [ + { + "bbox": [ + 391, + 86, + 468, + 163 + ], + "lines": [ + { + "bbox": [ + 391, + 86, + 468, + 163 + ], + "spans": [ + { + "bbox": [ + 391, + 86, + 468, + 163 + ], + "type": "image", + "image_path": "ff73fe5e4e66bf09fe7a83881410a8e0ffae1b43695f0bdf0c062549c9a3084f.jpg" + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "image_body" + } + ], + "index": 10 + }, + { + "bbox": [ + 393, + 165, + 466, + 220 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 393, + 165, + 466, + 220 + ], + "spans": [ + { + "bbox": [ + 393, + 165, + 466, + 220 + ], + "type": "text", + "content": "Tiger Vanguard charges through the blood-red waters towards Destined One. The massive stone statue behind him looms ominously, adding to the intensity of the scene. Destined One stands firm, his staff held at the ready, his gaze unwavering as he prepares for the oncoming assault. The water ripples around them, reflecting the dark hues of the battle setting." + } + ] + } + ], + "index": 11 + }, + { + "type": "image", + "bbox": [ + 470, + 86, + 547, + 163 + ], + "blocks": [ + { + "bbox": [ + 470, + 86, + 547, + 163 + ], + "lines": [ + { + "bbox": [ + 470, + 86, + 547, + 163 + ], + "spans": [ + { + "bbox": [ + 470, + 86, + 547, + 163 + ], + "type": "image", + "image_path": "d8f24403546657cbf631464bf9a63f5cc75c38ac27f953351b65d80eaa34ad1e.jpg" + } + ] + } + ], + "index": 12, + "angle": 0, + "type": "image_body" + } + ], + "index": 12 + }, + { + "bbox": [ + 471, + 165, + 544, + 220 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 471, + 165, + 544, + 220 + ], + "spans": [ + { + "bbox": [ + 471, + 165, + 544, + 220 + ], + "type": "text", + "content": "Destined One stands tall, his gaze cast downward, his weathered yellow robe flowing with subtle wear, reflecting the trials he has endured. Beside him, Zhu Majie looks up with a playful expression, his boar-like face filled with curiosity and mischief. The two figures share a quiet moment, framed against an open, clear sky dotted with drifting embers." + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 65, + 244, + 75, + 289 + ], + "type": "title", + "angle": 270, + "lines": [ + { + "bbox": [ + 65, + 244, + 75, + 289 + ], + "spans": [ + { + "bbox": [ + 65, + 244, + 75, + 289 + ], + "type": "text", + "content": "Image-to-Text" + } + ] + } + ], + "index": 15 + }, + { + "type": "image", + "bbox": [ + 78, + 228, + 154, + 305 + ], + "blocks": [ + { + "bbox": [ + 111, + 220, + 120, + 228 + ], + "lines": [ + { + "bbox": [ + 111, + 220, + 120, + 228 + ], + "spans": [ + { + "bbox": [ + 111, + 220, + 120, + 228 + ], + "type": "text", + "content": "(a)" + } + ] + } + ], + "index": 14, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 78, + 228, + 154, + 305 + ], + "lines": [ + { + "bbox": [ + 78, + 228, + 154, + 305 + ], + "spans": [ + { + "bbox": [ + 78, + 228, + 154, + 305 + ], + "type": "image", + "image_path": "c8cfb3108961accebf9fcff37647a316572990ec319ee49156e18605f7acdb0c.jpg" + } + ] + } + ], + "index": 16, + "angle": 0, + "type": "image_body" + } + ], + "index": 16 + }, + { + "bbox": [ + 78, + 308, + 154, + 369 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 78, + 308, + 154, + 369 + ], + "spans": [ + { + "bbox": [ + 78, + 308, + 154, + 369 + ], + "type": "text", + "content": "Destined One stands tall, wielding a massive golden stick that dominates the frame. The individual exudes power, strength, and determination, clad in ornate armor that gleams with gold. A hat adds a touch of regality. The background is dramatic, with shadows cast by trees hinting at a mysterious atmosphere. The grass beneath the figure is lush and green, contrasting with the ruggedness of the ground." + } + ] + } + ], + "index": 17 + }, + { + "type": "image", + "bbox": [ + 156, + 228, + 233, + 306 + ], + "blocks": [ + { + "bbox": [ + 189, + 220, + 199, + 228 + ], + "lines": [ + { + "bbox": [ + 189, + 220, + 199, + 228 + ], + "spans": [ + { + "bbox": [ + 189, + 220, + 199, + 228 + ], + "type": "text", + "content": "(b)" + } + ] + } + ], + "index": 18, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 156, + 228, + 233, + 306 + ], + "lines": [ + { + "bbox": [ + 156, + 228, + 233, + 306 + ], + "spans": [ + { + "bbox": [ + 156, + 228, + 233, + 306 + ], + "type": "image", + "image_path": "704e383d3f5d55ae6966d12ab4faef5a2416475478a563593dfee90f7ebda39f.jpg" + } + ] + } + ], + "index": 19, + "angle": 0, + "type": "image_body" + } + ], + "index": 19 + }, + { + "bbox": [ + 158, + 308, + 233, + 357 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 158, + 308, + 233, + 357 + ], + "spans": [ + { + "bbox": [ + 158, + 308, + 233, + 357 + ], + "type": "text", + "content": "Keeper, an elderly, hunched individual with a bald, heavily wrinkled head covered in spots. Keeper has a long, unkempt beard and is dressed in a simple, tattered blue robe. He grips a gnarled wooden staff for support as he walks through a forest fire. His posture is stooped, and his facial features are mostly obscured by his thick beard." + } + ] + } + ], + "index": 20 + }, + { + "type": "image", + "bbox": [ + 234, + 228, + 312, + 306 + ], + "blocks": [ + { + "bbox": [ + 268, + 220, + 278, + 228 + ], + "lines": [ + { + "bbox": [ + 268, + 220, + 278, + 228 + ], + "spans": [ + { + "bbox": [ + 268, + 220, + 278, + 228 + ], + "type": "text", + "content": "(c)" + } + ] + } + ], + "index": 21, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 234, + 228, + 312, + 306 + ], + "lines": [ + { + "bbox": [ + 234, + 228, + 312, + 306 + ], + "spans": [ + { + "bbox": [ + 234, + 228, + 312, + 306 + ], + "type": "image", + "image_path": "aadf39eb053a5ed03c1e774206f6facdabcdbf7ddb9b1f67cef4f56ed0c8aa91.jpg" + } + ] + } + ], + "index": 22, + "angle": 0, + "type": "image_body" + } + ], + "index": 22 + }, + { + "bbox": [ + 235, + 308, + 309, + 368 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 235, + 308, + 309, + 368 + ], + "spans": [ + { + "bbox": [ + 235, + 308, + 309, + 368 + ], + "type": "text", + "content": "The image depicts a serene, snowy landscape with a large, ancient-looking wooden bridge spanning across a frozen river. The bridge has a traditional, Asian design, with intricate carvings and a slatted roof. Bare trees with frost-covered branches surround the scene, casting a calm, wintery atmosphere. Snowflakes gently fall, enhancing the tranquil, mystical ambiance" + } + ] + } + ], + "index": 23 + }, + { + "type": "image", + "bbox": [ + 313, + 228, + 389, + 306 + ], + "blocks": [ + { + "bbox": [ + 347, + 220, + 356, + 228 + ], + "lines": [ + { + "bbox": [ + 347, + 220, + 356, + 228 + ], + "spans": [ + { + "bbox": [ + 347, + 220, + 356, + 228 + ], + "type": "text", + "content": "(d)" + } + ] + } + ], + "index": 24, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 313, + 228, + 389, + 306 + ], + "lines": [ + { + "bbox": [ + 313, + 228, + 389, + 306 + ], + "spans": [ + { + "bbox": [ + 313, + 228, + 389, + 306 + ], + "type": "image", + "image_path": "e4a5e9c242fca82aa9cc3e2dd5692905c86c225cebe3c161578f9c0933a0630d.jpg" + } + ] + } + ], + "index": 25, + "angle": 0, + "type": "image_body" + } + ], + "index": 25 + }, + { + "bbox": [ + 313, + 308, + 389, + 368 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 308, + 389, + 368 + ], + "spans": [ + { + "bbox": [ + 313, + 308, + 389, + 368 + ], + "type": "text", + "content": "Lingxuzi is depicted as a large, menacing fox-like creature with thick, flowing white fur. Its red eyes glow intensely, giving an eerie, supernatural appearance. The creature is in a crouching stance, with its body slightly arched and its tails fanned out behind it, showing both power and readiness to strike. The dark background adds to the ominous atmosphere, contrasting with the brightness of its fur." + } + ] + } + ], + "index": 26 + }, + { + "type": "image", + "bbox": [ + 391, + 228, + 468, + 306 + ], + "blocks": [ + { + "bbox": [ + 425, + 220, + 435, + 228 + ], + "lines": [ + { + "bbox": [ + 425, + 220, + 435, + 228 + ], + "spans": [ + { + "bbox": [ + 425, + 220, + 435, + 228 + ], + "type": "text", + "content": "(e)" + } + ] + } + ], + "index": 27, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 391, + 228, + 468, + 306 + ], + "lines": [ + { + "bbox": [ + 391, + 228, + 468, + 306 + ], + "spans": [ + { + "bbox": [ + 391, + 228, + 468, + 306 + ], + "type": "image", + "image_path": "b7f74ba3fcf5b04cea895ebbb2b4dcbe59456f4a51fe61891b691dd6768f720e.jpg" + } + ] + } + ], + "index": 28, + "angle": 0, + "type": "image_body" + } + ], + "index": 28 + }, + { + "bbox": [ + 392, + 308, + 466, + 378 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 392, + 308, + 466, + 378 + ], + "spans": [ + { + "bbox": [ + 392, + 308, + 466, + 378 + ], + "type": "text", + "content": "Zhu Bajie stands menacingly in a murky, mist-covered swamp, his humanoid body combined with the head of a fierce, snarling boar. Her red, glowing eyes pierce through the fog, radiating a dark energy. Dressed in a tattered green robe, he wears a string of large beads around his neck, adding a sense of mysticism. His left hand clutches a spiked weapon, while shadows and wisps of smoke swirl around him, amplifying his sinister presence." + } + ] + } + ], + "index": 29 + }, + { + "type": "image", + "bbox": [ + 470, + 228, + 547, + 306 + ], + "blocks": [ + { + "bbox": [ + 504, + 220, + 512, + 228 + ], + "lines": [ + { + "bbox": [ + 504, + 220, + 512, + 228 + ], + "spans": [ + { + "bbox": [ + 504, + 220, + 512, + 228 + ], + "type": "text", + "content": "(f)" + } + ] + } + ], + "index": 30, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 470, + 228, + 547, + 306 + ], + "lines": [ + { + "bbox": [ + 470, + 228, + 547, + 306 + ], + "spans": [ + { + "bbox": [ + 470, + 228, + 547, + 306 + ], + "type": "image", + "image_path": "f9ecdabaabd601729254048c6fb77491b9dafde4b0029f5982d36e14a2494d62.jpg" + } + ] + } + ], + "index": 31, + "angle": 0, + "type": "image_body" + } + ], + "index": 31 + }, + { + "bbox": [ + 471, + 308, + 545, + 382 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 471, + 308, + 545, + 382 + ], + "spans": [ + { + "bbox": [ + 471, + 308, + 545, + 382 + ], + "type": "text", + "content": "A female character with a unique appearance, having a combination of human and demonic features. She has a short, black, straight hairstyle and a pair of horns protruding from her head. Her skin is pale, and she has a few dots on her forehead, giving her a mystical look. She is wearing a black, low-cut top adorned with white flowers, and her right arm is covered in red fabric. She has a focused and intense expression on her face, as if she is concentrating on something important. (D)" + } + ] + } + ], + "index": 32 + }, + { + "bbox": [ + 65, + 421, + 75, + 464 + ], + "type": "title", + "angle": 270, + "lines": [ + { + "bbox": [ + 65, + 421, + 75, + 464 + ], + "spans": [ + { + "bbox": [ + 65, + 421, + 75, + 464 + ], + "type": "text", + "content": "Text-to-Image" + } + ] + } + ], + "index": 39 + }, + { + "type": "image", + "bbox": [ + 78, + 405, + 156, + 468 + ], + "blocks": [ + { + "bbox": [ + 111, + 380, + 120, + 388 + ], + "lines": [ + { + "bbox": [ + 111, + 380, + 120, + 388 + ], + "spans": [ + { + "bbox": [ + 111, + 380, + 120, + 388 + ], + "type": "text", + "content": "(g)" + } + ] + } + ], + "index": 34, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 78, + 405, + 156, + 468 + ], + "lines": [ + { + "bbox": [ + 78, + 405, + 156, + 468 + ], + "spans": [ + { + "bbox": [ + 78, + 405, + 156, + 468 + ], + "type": "image", + "image_path": "c3a1c8d8eb39e13c9dd4470d110fcdaf3399a6aaabea6d7afd73fe7258e8663d.jpg" + } + ] + } + ], + "index": 40, + "angle": 0, + "type": "image_body" + } + ], + "index": 40 + }, + { + "type": "image", + "bbox": [ + 78, + 469, + 151, + 481 + ], + "blocks": [ + { + "bbox": [ + 78, + 469, + 151, + 481 + ], + "lines": [ + { + "bbox": [ + 78, + 469, + 151, + 481 + ], + "spans": [ + { + "bbox": [ + 78, + 469, + 151, + 481 + ], + "type": "image", + "image_path": "6f2633121fd8eebbf802b5716e36b442c498538e94e037ccac324b5c8c7e3f07.jpg" + } + ] + } + ], + "index": 41, + "angle": 0, + "type": "image_body" + } + ], + "index": 41 + }, + { + "bbox": [ + 78, + 485, + 151, + 507 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 78, + 485, + 151, + 507 + ], + "spans": [ + { + "bbox": [ + 78, + 485, + 151, + 507 + ], + "type": "text", + "content": "Nobita is sitting under a tree, trying to finish his math homework. He looks stressed, with one hand on his chin and scratching his head with the other." + } + ] + } + ], + "index": 42 + }, + { + "type": "image", + "bbox": [ + 157, + 405, + 234, + 481 + ], + "blocks": [ + { + "bbox": [ + 189, + 380, + 199, + 388 + ], + "lines": [ + { + "bbox": [ + 189, + 380, + 199, + 388 + ], + "spans": [ + { + "bbox": [ + 189, + 380, + 199, + 388 + ], + "type": "text", + "content": "(h)" + } + ] + } + ], + "index": 35, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 157, + 405, + 234, + 481 + ], + "lines": [ + { + "bbox": [ + 157, + 405, + 234, + 481 + ], + "spans": [ + { + "bbox": [ + 157, + 405, + 234, + 481 + ], + "type": "image", + "image_path": "ab5c77f33e601cc1c923973cc010f58c95788b20ab83699c9f4b4cce12df1178.jpg" + } + ] + } + ], + "index": 43, + "angle": 0, + "type": "image_body" + } + ], + "index": 43 + }, + { + "bbox": [ + 157, + 484, + 230, + 533 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 157, + 484, + 230, + 533 + ], + "spans": [ + { + "bbox": [ + 157, + 484, + 230, + 533 + ], + "type": "text", + "content": "Suneo and Nobita are at a fishing pond, trying to catch a fish. Nobita, in his bright yellow fishing hat and a sharp outfit, confidently casts his line into the water. Suneo watches Nobita's struggle with a smile, offering some tips to help him out. The two boys share a laugh as they wait for a bite." + } + ] + } + ], + "index": 44 + }, + { + "type": "image", + "bbox": [ + 236, + 405, + 312, + 481 + ], + "blocks": [ + { + "bbox": [ + 268, + 380, + 277, + 388 + ], + "lines": [ + { + "bbox": [ + 268, + 380, + 277, + 388 + ], + "spans": [ + { + "bbox": [ + 268, + 380, + 277, + 388 + ], + "type": "text", + "content": "(i)" + } + ] + } + ], + "index": 36, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 264, + 392, + 346, + 401 + ], + "lines": [ + { + "bbox": [ + 264, + 392, + 346, + 401 + ], + "spans": [ + { + "bbox": [ + 264, + 392, + 346, + 401 + ], + "type": "text", + "content": "Domain 2: Doraemon" + } + ] + } + ], + "index": 38, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 236, + 405, + 312, + 481 + ], + "lines": [ + { + "bbox": [ + 236, + 405, + 312, + 481 + ], + "spans": [ + { + "bbox": [ + 236, + 405, + 312, + 481 + ], + "type": "image", + "image_path": "f13264c30b6a0da6e01d4025bac2335db2d2f8274eb375b9cf1362accf377f43.jpg" + } + ] + } + ], + "index": 45, + "angle": 0, + "type": "image_body" + } + ], + "index": 45 + }, + { + "bbox": [ + 236, + 484, + 308, + 512 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 236, + 484, + 308, + 512 + ], + "spans": [ + { + "bbox": [ + 236, + 484, + 308, + 512 + ], + "type": "text", + "content": "Doraemon is cooking in the kitchen, while holding a bowl. He looks excited, flipping a pancake into the air. Nobita is standing beside him, smiling gently, watching Doraemon." + } + ] + } + ], + "index": 46 + }, + { + "type": "image", + "bbox": [ + 313, + 405, + 390, + 481 + ], + "blocks": [ + { + "bbox": [ + 347, + 380, + 356, + 388 + ], + "lines": [ + { + "bbox": [ + 347, + 380, + 356, + 388 + ], + "spans": [ + { + "bbox": [ + 347, + 380, + 356, + 388 + ], + "type": "text", + "content": "(j)" + } + ] + } + ], + "index": 37, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 313, + 405, + 390, + 481 + ], + "lines": [ + { + "bbox": [ + 313, + 405, + 390, + 481 + ], + "spans": [ + { + "bbox": [ + 313, + 405, + 390, + 481 + ], + "type": "image", + "image_path": "b8241675d95fba7f82713398267b5bbdeb136894452978f0de260cebadd351d8.jpg" + } + ] + } + ], + "index": 47, + "angle": 0, + "type": "image_body" + } + ], + "index": 47 + }, + { + "bbox": [ + 314, + 484, + 387, + 528 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 314, + 484, + 387, + 528 + ], + "spans": [ + { + "bbox": [ + 314, + 484, + 387, + 528 + ], + "type": "text", + "content": "Shizuka is sitting by a piano, composing a new tune. She looks thoughtful, her fingers delicately pressing the keys as she tries different melodies. A notebook rests on the piano stand, filled with her notes and musical ideas, and she appears content in the process." + } + ] + } + ], + "index": 48 + }, + { + "type": "image", + "bbox": [ + 391, + 405, + 467, + 482 + ], + "blocks": [ + { + "bbox": [ + 391, + 405, + 467, + 482 + ], + "lines": [ + { + "bbox": [ + 391, + 405, + 467, + 482 + ], + "spans": [ + { + "bbox": [ + 391, + 405, + 467, + 482 + ], + "type": "image", + "image_path": "c2a30f70ce7eab0d82a13395f7cdd31e3e5739853fc70813e09bd9fb30c7d45e.jpg" + } + ] + } + ], + "index": 49, + "angle": 0, + "type": "image_body" + } + ], + "index": 49 + }, + { + "bbox": [ + 392, + 484, + 465, + 512 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 392, + 484, + 465, + 512 + ], + "spans": [ + { + "bbox": [ + 392, + 484, + 465, + 512 + ], + "type": "text", + "content": "Shizuka and Nobita are making flower crowns at the park. Shizuka, in her soft pink dress, weaves flowers together. Nobita, in his casual shorts, tries his best to copy her." + } + ] + } + ], + "index": 50 + }, + { + "type": "image", + "bbox": [ + 470, + 405, + 545, + 481 + ], + "blocks": [ + { + "bbox": [ + 504, + 381, + 512, + 388 + ], + "lines": [ + { + "bbox": [ + 504, + 381, + 512, + 388 + ], + "spans": [ + { + "bbox": [ + 504, + 381, + 512, + 388 + ], + "type": "text", + "content": "(1)" + } + ] + } + ], + "index": 33, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 470, + 405, + 545, + 481 + ], + "lines": [ + { + "bbox": [ + 470, + 405, + 545, + 481 + ], + "spans": [ + { + "bbox": [ + 470, + 405, + 545, + 481 + ], + "type": "image", + "image_path": "ce3544a651f6016e9caeb2a81d4c904b4e05405714243cb8dae911ef4a8a14a7.jpg" + } + ] + } + ], + "index": 51, + "angle": 0, + "type": "image_body" + } + ], + "index": 51 + }, + { + "bbox": [ + 471, + 484, + 544, + 538 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 471, + 484, + 544, + 538 + ], + "spans": [ + { + "bbox": [ + 471, + 484, + 544, + 538 + ], + "type": "text", + "content": "Doraemon is standing in front of the white cat with an excited expression. He is holding a bouquet of flowers, in front of the cat. The white cat, sitting on a small cushion, is looking at Doraemon with a curious or intrigued expression. The setting appears to be indoors, with a pink carpet, a scratching post in the background, and a framed picture on the wall." + } + ] + } + ], + "index": 52 + }, + { + "bbox": [ + 65, + 564, + 75, + 607 + ], + "type": "title", + "angle": 270, + "lines": [ + { + "bbox": [ + 65, + 564, + 75, + 607 + ], + "spans": [ + { + "bbox": [ + 65, + 564, + 75, + 607 + ], + "type": "text", + "content": "Image-to-Text" + } + ] + } + ], + "index": 53 + }, + { + "type": "image", + "bbox": [ + 78, + 539, + 156, + 625 + ], + "blocks": [ + { + "bbox": [ + 78, + 539, + 156, + 625 + ], + "lines": [ + { + "bbox": [ + 78, + 539, + 156, + 625 + ], + "spans": [ + { + "bbox": [ + 78, + 539, + 156, + 625 + ], + "type": "image", + "image_path": "66cbdf51c43a6bb3ef2b83bf353c163add826a4f6fc7f26e0d28fc812a5da9f2.jpg" + } + ] + } + ], + "index": 54, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 111, + 671, + 120, + 679 + ], + "lines": [ + { + "bbox": [ + 111, + 671, + 120, + 679 + ], + "spans": [ + { + "bbox": [ + 111, + 671, + 120, + 679 + ], + "type": "text", + "content": "(s)" + } + ] + } + ], + "index": 70, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 56, + 693, + 552, + 705 + ], + "lines": [ + { + "bbox": [ + 56, + 693, + 552, + 705 + ], + "spans": [ + { + "bbox": [ + 56, + 693, + 552, + 705 + ], + "type": "text", + "content": "Figure 4. Image-to-text and text-to-image generation by the unified models that adapted for two domains. The special tokens are omitted." + } + ] + } + ], + "index": 76, + "angle": 0, + "type": "image_caption" + } + ], + "index": 54 + }, + { + "bbox": [ + 78, + 628, + 151, + 666 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 78, + 628, + 151, + 666 + ], + "spans": [ + { + "bbox": [ + 78, + 628, + 151, + 666 + ], + "type": "text", + "content": "Doraemon is standing on the shore, looking out at the sunset. He has a calm expression, and the sun is setting over the water, casting a warm glow. The sky is filled with clouds, and there are silhouettes of buildings in the distance." + } + ] + } + ], + "index": 55 + }, + { + "type": "image", + "bbox": [ + 157, + 548, + 234, + 624 + ], + "blocks": [ + { + "bbox": [ + 189, + 539, + 199, + 547 + ], + "lines": [ + { + "bbox": [ + 189, + 539, + 199, + 547 + ], + "spans": [ + { + "bbox": [ + 189, + 539, + 199, + 547 + ], + "type": "text", + "content": "(n)" + } + ] + } + ], + "index": 56, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 157, + 548, + 234, + 624 + ], + "lines": [ + { + "bbox": [ + 157, + 548, + 234, + 624 + ], + "spans": [ + { + "bbox": [ + 157, + 548, + 234, + 624 + ], + "type": "image", + "image_path": "b3928faa52e349ecd559c44612e7824d6c7c710385ea9e3880a29c332b0ffeb3.jpg" + } + ] + } + ], + "index": 57, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 189, + 671, + 198, + 679 + ], + "lines": [ + { + "bbox": [ + 189, + 671, + 198, + 679 + ], + "spans": [ + { + "bbox": [ + 189, + 671, + 198, + 679 + ], + "type": "text", + "content": "(t)" + } + ] + } + ], + "index": 71, + "angle": 0, + "type": "image_caption" + } + ], + "index": 57 + }, + { + "bbox": [ + 157, + 628, + 230, + 666 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 157, + 628, + 230, + 666 + ], + "spans": [ + { + "bbox": [ + 157, + 628, + 230, + 666 + ], + "type": "text", + "content": "Doraemon is standing in the grass, looking surprised as he sees a group of cats. The cats are gathered around him, with one of them appearing to be a cat with a surprised expression. The background shows a fence and some bushes." + } + ] + } + ], + "index": 58 + }, + { + "type": "image", + "bbox": [ + 236, + 548, + 312, + 624 + ], + "blocks": [ + { + "bbox": [ + 236, + 548, + 312, + 624 + ], + "lines": [ + { + "bbox": [ + 236, + 548, + 312, + 624 + ], + "spans": [ + { + "bbox": [ + 236, + 548, + 312, + 624 + ], + "type": "image", + "image_path": "849259d7fbc5ab22955acb39c7b11b335ce7135ac7e17dbbcbe6776882e7b8da.jpg" + } + ] + } + ], + "index": 59, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 268, + 671, + 277, + 679 + ], + "lines": [ + { + "bbox": [ + 268, + 671, + 277, + 679 + ], + "spans": [ + { + "bbox": [ + 268, + 671, + 277, + 679 + ], + "type": "text", + "content": "(u)" + } + ] + } + ], + "index": 72, + "angle": 0, + "type": "image_caption" + } + ], + "index": 59 + }, + { + "bbox": [ + 236, + 628, + 308, + 650 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 236, + 628, + 308, + 650 + ], + "spans": [ + { + "bbox": [ + 236, + 628, + 308, + 650 + ], + "type": "text", + "content": "Doradamon is flying through the air, looking down at the city below. Nobita is also flying through the air, looking down at the city." + } + ] + } + ], + "index": 60 + }, + { + "type": "image", + "bbox": [ + 313, + 548, + 390, + 624 + ], + "blocks": [ + { + "bbox": [ + 347, + 539, + 356, + 547 + ], + "lines": [ + { + "bbox": [ + 347, + 539, + 356, + 547 + ], + "spans": [ + { + "bbox": [ + 347, + 539, + 356, + 547 + ], + "type": "text", + "content": "(p)" + } + ] + } + ], + "index": 61, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 313, + 548, + 390, + 624 + ], + "lines": [ + { + "bbox": [ + 313, + 548, + 390, + 624 + ], + "spans": [ + { + "bbox": [ + 313, + 548, + 390, + 624 + ], + "type": "image", + "image_path": "5fea7abcdc30d6b74bcfcd1eb4ba39cf92839d1563150bf49669380f3b02be92.jpg" + } + ] + } + ], + "index": 62, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 347, + 671, + 356, + 680 + ], + "lines": [ + { + "bbox": [ + 347, + 671, + 356, + 680 + ], + "spans": [ + { + "bbox": [ + 347, + 671, + 356, + 680 + ], + "type": "text", + "content": "(v)" + } + ] + } + ], + "index": 73, + "angle": 0, + "type": "image_caption" + } + ], + "index": 62 + }, + { + "bbox": [ + 314, + 628, + 387, + 671 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 314, + 628, + 387, + 671 + ], + "spans": [ + { + "bbox": [ + 314, + 628, + 387, + 671 + ], + "type": "text", + "content": "The image depicts a two-story house with a large garden in front of it. The house has a tan exterior and is surrounded by a variety of plants, including palm trees and bushes. The garden is well-maintained. The house features a balcony on the second floor, adding to its charm." + } + ] + } + ], + "index": 63 + }, + { + "type": "image", + "bbox": [ + 391, + 548, + 468, + 624 + ], + "blocks": [ + { + "bbox": [ + 425, + 539, + 435, + 547 + ], + "lines": [ + { + "bbox": [ + 425, + 539, + 435, + 547 + ], + "spans": [ + { + "bbox": [ + 425, + 539, + 435, + 547 + ], + "type": "text", + "content": "(q)" + } + ] + } + ], + "index": 64, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 391, + 548, + 468, + 624 + ], + "lines": [ + { + "bbox": [ + 391, + 548, + 468, + 624 + ], + "spans": [ + { + "bbox": [ + 391, + 548, + 468, + 624 + ], + "type": "image", + "image_path": "8245120e916c34348b7d1725f64102d1ac7f99d8a264daa5af3c8f223ad0377f.jpg" + } + ] + } + ], + "index": 65, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 424, + 671, + 435, + 679 + ], + "lines": [ + { + "bbox": [ + 424, + 671, + 435, + 679 + ], + "spans": [ + { + "bbox": [ + 424, + 671, + 435, + 679 + ], + "type": "text", + "content": "(w)" + } + ] + } + ], + "index": 74, + "angle": 0, + "type": "image_caption" + } + ], + "index": 65 + }, + { + "bbox": [ + 392, + 628, + 465, + 661 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 392, + 628, + 465, + 661 + ], + "spans": [ + { + "bbox": [ + 392, + 628, + 465, + 661 + ], + "type": "text", + "content": "Nobita is standing in the doorway, holding a plate with a doughnut on it. He looks excited, and the doughnut is on a white plate. The door is open, and the room is visible in the background." + } + ] + } + ], + "index": 66 + }, + { + "type": "image", + "bbox": [ + 470, + 548, + 545, + 624 + ], + "blocks": [ + { + "bbox": [ + 504, + 540, + 512, + 547 + ], + "lines": [ + { + "bbox": [ + 504, + 540, + 512, + 547 + ], + "spans": [ + { + "bbox": [ + 504, + 540, + 512, + 547 + ], + "type": "text", + "content": "(r)" + } + ] + } + ], + "index": 67, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 470, + 548, + 545, + 624 + ], + "lines": [ + { + "bbox": [ + 470, + 548, + 545, + 624 + ], + "spans": [ + { + "bbox": [ + 470, + 548, + 545, + 624 + ], + "type": "image", + "image_path": "67ecfcbba9ae96f1be693b1efe68514ba38a55c3d9198ab7f8e1b1c19a3d13c6.jpg" + } + ] + } + ], + "index": 68, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 504, + 671, + 513, + 679 + ], + "lines": [ + { + "bbox": [ + 504, + 671, + 513, + 679 + ], + "spans": [ + { + "bbox": [ + 504, + 671, + 513, + 679 + ], + "type": "text", + "content": "(x)" + } + ] + } + ], + "index": 75, + "angle": 0, + "type": "image_caption" + } + ], + "index": 68 + }, + { + "bbox": [ + 471, + 628, + 544, + 655 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 471, + 628, + 544, + 655 + ], + "spans": [ + { + "bbox": [ + 471, + 628, + 544, + 655 + ], + "type": "text", + "content": "White cat and black cat are sitting on the grass, looking at each other with affection, with a blue sky in the background, and a few clouds scattered across the sky." + } + ] + } + ], + "index": 69 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "text", + "content": "2840" + } + ] + } + ], + "index": 77 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 5 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 58, + 70, + 294, + 257 + ], + "blocks": [ + { + "bbox": [ + 58, + 70, + 294, + 257 + ], + "lines": [ + { + "bbox": [ + 58, + 70, + 294, + 257 + ], + "spans": [ + { + "bbox": [ + 58, + 70, + 294, + 257 + ], + "type": "image", + "image_path": "47a1a143534d9ec5df110d474acfeb5385995373562f237710a5036fc5032de7.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 55, + 266, + 295, + 322 + ], + "lines": [ + { + "bbox": [ + 55, + 266, + 295, + 322 + ], + "spans": [ + { + "bbox": [ + 55, + 266, + 295, + 322 + ], + "type": "text", + "content": "Figure 5. Effect of special tokens on character learning. (a) Base model without training. (b) Model trained without using special tokens, showing attribute confusion among characters. (c) Model trained with special tokens, improving character attribute alignment and reducing confusion." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "bbox": [ + 55, + 348, + 296, + 420 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 348, + 296, + 420 + ], + "spans": [ + { + "bbox": [ + 55, + 348, + 296, + 420 + ], + "type": "text", + "content": "data to comprehensively learn general aspects of the target domain. Specifically, in each batch, for data with paired ground truth, we compute the token prediction loss and also include it in the cycle, use ground truth as the pseudo middle generations, and compute the cycle loss. For unpaired data, we compute the unpaired cycle loss." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 55, + 422, + 296, + 541 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 422, + 296, + 541 + ], + "spans": [ + { + "bbox": [ + 55, + 422, + 296, + 541 + ], + "type": "text", + "content": "For example, when adapting the model to Domain 1: Black Myth Wukong and Domain 2: Doraemon, we annotate 1-3 images per unique identity with captions that specify the name of the identity. For each domain, we collect " + }, + { + "bbox": [ + 55, + 422, + 296, + 541 + ], + "type": "inline_equation", + "content": "2\\mathrm{k}" + }, + { + "bbox": [ + 55, + 422, + 296, + 541 + ], + "type": "text", + "content": " images, which are mostly sampled from online videos, and independently collect text descriptions, which are further expanded to " + }, + { + "bbox": [ + 55, + 422, + 296, + 541 + ], + "type": "inline_equation", + "content": "1\\mathrm{k}" + }, + { + "bbox": [ + 55, + 422, + 296, + 541 + ], + "type": "text", + "content": " by ChatGPT [39]. The final adapted model trained with DoraCycle demonstrates strong performance in both text-to-image generation and image-to-text generation, as shown in Fig. 4." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 55, + 544, + 295, + 698 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 544, + 295, + 698 + ], + "spans": [ + { + "bbox": [ + 55, + 544, + 295, + 698 + ], + "type": "text", + "content": "In terms of text-to-image results, the model trained with DoraCycle effectively generated images that aligned well with the target domains. In domain 1 (Black Myth Wukong), the generated images accurately depicted domain-specific visual elements, such as the intricate details of character appearances and the overall fantasy-like atmosphere. This indicates that the model successfully learned to generalize the visual features from text prompts to realistic images within the target domain. Similarly, in domain 2 (Doraemon), the generated images preserve the iconic cartoonish aesthetics and capture key visual details of the characters and settings, demonstrating effective domain adaptation." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 67, + 701, + 295, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 67, + 701, + 295, + 713 + ], + "spans": [ + { + "bbox": [ + 67, + 701, + 295, + 713 + ], + "type": "text", + "content": "For the image-to-text task, the model performs well in" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 313, + 72, + 553, + 179 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 72, + 553, + 179 + ], + "spans": [ + { + "bbox": [ + 313, + 72, + 553, + 179 + ], + "type": "text", + "content": "generating contextually accurate captions. In domain 1, the generated captions provide rich descriptions of the characters, their attributes, and the context, effectively mirroring the visual elements present in the input images. In domain 2, the captions correctly describe the characters, their actions, and their environments concisely, maintaining consistency with the visual style. The ability of the model to generate accurate descriptions highlights its robust understanding of the visual components of the domain." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 313, + 181, + 555, + 407 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 181, + 555, + 407 + ], + "spans": [ + { + "bbox": [ + 313, + 181, + 555, + 407 + ], + "type": "text", + "content": "Additionally, an interesting phenomenon can be observed in how the model handles the visual elements that are not annotated with paired data. For instance, in Fig. 4 (w), the dorayaki (a type of sweet bean-filled pancake) was described by the model as a \"doughnut\". This may be due to the fact that the anime-style representation of the dorayaki is novel, and neither the base model nor the unpaired training provided specific textual-visual pairing knowledge about it. On the other hand, in the example shown in Fig. 4 (x), we annotate the white cat as a character with paired textual and visual data, using a special token for its name: \" white cat black cat \" during the caption generation. This suggests that the model autonomously categorized the black cat as a character when learning the target domain, indicating that it may have attempted to generalize learned knowledge from one type of entity to similar ones." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 313, + 409, + 556, + 612 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 409, + 556, + 612 + ], + "spans": [ + { + "bbox": [ + 313, + 409, + 556, + 612 + ], + "type": "text", + "content": "Enhanced Learning with Special Tokens: As shown in Fig. 5, we experimentally find that the model often confused multiple novel concepts in the target domain. Fig. 5 (a) shows the image generated by the base model without training taking in the name of characters. Fig. 5 (b) shows the characters generated by the trained model. During training, the names of characters are directly included in the text without special treatment, leading to attribute confusion between characters. The varying lengths of the tokenized character names also make learning difficult. To solve this problem, we introduce a simple yet efficient solution: adding special tokens around character names. We introduced start of character " + }, + { + "bbox": [ + 313, + 409, + 556, + 612 + ], + "type": "inline_equation", + "content": "(\\langle \\text{soc}\\rangle)" + }, + { + "bbox": [ + 313, + 409, + 556, + 612 + ], + "type": "text", + "content": " and end of character " + }, + { + "bbox": [ + 313, + 409, + 556, + 612 + ], + "type": "inline_equation", + "content": "(\\langle \\text{soc}\\rangle)" + }, + { + "bbox": [ + 313, + 409, + 556, + 612 + ], + "type": "text", + "content": " tokens to enclose character names, which significantly enhance the learning of novel concepts. As shown in Fig. 5 (c), involving special tokens improves the alignments between characters and their names." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 313, + 623, + 400, + 635 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 623, + 400, + 635 + ], + "spans": [ + { + "bbox": [ + 313, + 623, + 400, + 635 + ], + "type": "text", + "content": "4.3. Comparisons" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 313, + 641, + 555, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 641, + 555, + 715 + ], + "spans": [ + { + "bbox": [ + 313, + 641, + 555, + 715 + ], + "type": "text", + "content": "In this section, we use the Storyboard20K [66] dataset to conduct the quantitative comparison experiments. The storyboards originating from the same data source are grouped to form a domain, consisting of images and descriptive text. The data are used under three different settings, i.e. totally unpaired, only paired, and paired plus unpaired data," + } + ] + } + ], + "index": 10 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 748, + 314, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 314, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 314, + 757 + ], + "type": "text", + "content": "2841" + } + ] + } + ], + "index": 11 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 6 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 58, + 105, + 555, + 232 + ], + "blocks": [ + { + "bbox": [ + 55, + 71, + 555, + 95 + ], + "lines": [ + { + "bbox": [ + 55, + 71, + 555, + 95 + ], + "spans": [ + { + "bbox": [ + 55, + 71, + 555, + 95 + ], + "type": "text", + "content": "Table 1. Comparison of different training methods under various data settings. The best value is highlighted in blue , and the second-best value is highlighted in green . “P” indicates paired data, and “U” indicates unpaired data." + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 58, + 105, + 555, + 232 + ], + "lines": [ + { + "bbox": [ + 58, + 105, + 555, + 232 + ], + "spans": [ + { + "bbox": [ + 58, + 105, + 555, + 232 + ], + "type": "table", + "html": "
T cycleI cycleT DataI DataFID-1K ↓CIDEr ↑Human Eval
T2I Align ↑I2T Align ↑
DreamBooth [48]--10% P10% P33.2232.743.251.83
--100% P100% P24.9341.554.133.96
DoraCycleX✓X100% U28.9330.543.381.62
✓X100% UX36.6335.703.262.17
✓✓100% U100% U27.4438.173.843.42
✓✓10% P + 90% U10% P + 90% U25.3740.904.123.81
ITIT [34]✓✓10% P + 90% U10% P + 90% U27.5038.623.853.52
", + "image_path": "46b62c72ec46647e2c1bcad4b076d087a4640742a46022a03e236a52ee9a9773.jpg" + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "table_body" + } + ], + "index": 1 + }, + { + "bbox": [ + 55, + 251, + 141, + 262 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 251, + 141, + 262 + ], + "spans": [ + { + "bbox": [ + 55, + 251, + 141, + 262 + ], + "type": "text", + "content": "as shown in Table 1." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 55, + 265, + 295, + 348 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 265, + 295, + 348 + ], + "spans": [ + { + "bbox": [ + 55, + 265, + 295, + 348 + ], + "type": "text", + "content": "The compared methods include DreamBooth [48] and ITIT [34]. We implement DreamBooth as a paired-training baseline by applying LoRA fine-tuning on the unified model. The original design of ITIT is different, in which the image and text decoders are separate models, and its code has not been released. We adjusted and re-implemented it to be suitable for our unified model architecture." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 54, + 350, + 295, + 483 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 54, + 350, + 295, + 483 + ], + "spans": [ + { + "bbox": [ + 54, + 350, + 295, + 483 + ], + "type": "text", + "content": "We use both automatic and human evaluations to compare the performance of different methods. For automatic evaluation, we use FID to measure the distribution differences between the generated images and the target domain images [22], and CIDEr to compute the error between the generated text and the ground truth [59]. For human evaluation, we create 100 questions for the generated results of models, each rated by three different human raters. The raters are asked to evaluate the alignment between the image and text on a scale from 1 to 5, where 1 indicates no relevance and 5 indicates complete alignment." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 54, + 484, + 295, + 616 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 54, + 484, + 295, + 616 + ], + "spans": [ + { + "bbox": [ + 54, + 484, + 295, + 616 + ], + "type": "text", + "content": "The experimental results in Table 1 demonstrate that the proposed DoraCycle performs competitively under several data settings. Specifically, when using a combination of paired and unpaired data, DoraCycle outperforms ITIT. Compared to DreamBooth, which heavily relies on paired data, DoraCycle outperforms it when using the same scale of paired data, i.e. " + }, + { + "bbox": [ + 54, + 484, + 295, + 616 + ], + "type": "inline_equation", + "content": "10\\%" + }, + { + "bbox": [ + 54, + 484, + 295, + 616 + ], + "type": "text", + "content": " paired data, indicating the benefits brought by " + }, + { + "bbox": [ + 54, + 484, + 295, + 616 + ], + "type": "inline_equation", + "content": "90\\%" + }, + { + "bbox": [ + 54, + 484, + 295, + 616 + ], + "type": "text", + "content": " unpaired data. While Dreambooth with " + }, + { + "bbox": [ + 54, + 484, + 295, + 616 + ], + "type": "inline_equation", + "content": "100\\%" + }, + { + "bbox": [ + 54, + 484, + 295, + 616 + ], + "type": "text", + "content": " achieves the best evaluation scores, the scores of the DoraCycle with " + }, + { + "bbox": [ + 54, + 484, + 295, + 616 + ], + "type": "inline_equation", + "content": "10\\%" + }, + { + "bbox": [ + 54, + 484, + 295, + 616 + ], + "type": "text", + "content": " paired and " + }, + { + "bbox": [ + 54, + 484, + 295, + 616 + ], + "type": "inline_equation", + "content": "90\\%" + }, + { + "bbox": [ + 54, + 484, + 295, + 616 + ], + "type": "text", + "content": " unpaired data are comparable with them." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 54, + 617, + 295, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 54, + 617, + 295, + 713 + ], + "spans": [ + { + "bbox": [ + 54, + 617, + 295, + 713 + ], + "type": "text", + "content": "Table 1 also shows the difference in the performance of DoraCyle under different cycle settings. It is shown that without the T cycle and with only the I cycle, the captioning ability of the adapted model degrades more significantly. In contrast, if only the T cycle is used and without the I cycle, the FID score increases substantially, indicating that the generated image distribution mismatches with the target distribution." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 314, + 251, + 414, + 262 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 314, + 251, + 414, + 262 + ], + "spans": [ + { + "bbox": [ + 314, + 251, + 414, + 262 + ], + "type": "text", + "content": "4.4. Ablation Studies" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 313, + 274, + 409, + 382 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 274, + 409, + 382 + ], + "spans": [ + { + "bbox": [ + 313, + 274, + 409, + 382 + ], + "type": "text", + "content": "Table 2 shows that removing key components from DoraCycle significantly impacts performance. Without EMA, the FID score increases from 25.37 to 27.19, indicating lower image quality" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 426, + 274, + 554, + 319 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 426, + 274, + 554, + 319 + ], + "spans": [ + { + "bbox": [ + 426, + 274, + 554, + 319 + ], + "type": "text", + "content": "Table 2. Ablation Studies. EMA refers to the exponential moving average. GS refers to gradient surgery." + } + ] + } + ], + "index": 9 + }, + { + "type": "table", + "bbox": [ + 429, + 327, + 553, + 376 + ], + "blocks": [ + { + "bbox": [ + 429, + 327, + 553, + 376 + ], + "lines": [ + { + "bbox": [ + 429, + 327, + 553, + 376 + ], + "spans": [ + { + "bbox": [ + 429, + 327, + 553, + 376 + ], + "type": "table", + "html": "
FID-1K ↓CIDEr ↑
w/o EMA27.1938.85
w/o GS25.5439.98
DoraCycle25.3740.90
", + "image_path": "6af9013b4ab42cc8621e701c3d9aacb98875a37a45e7a8ac434a20ebdf2643f1.jpg" + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "table_body" + } + ], + "index": 10 + }, + { + "bbox": [ + 313, + 383, + 555, + 490 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 383, + 555, + 490 + ], + "spans": [ + { + "bbox": [ + 313, + 383, + 555, + 490 + ], + "type": "text", + "content": "due to less stable training. Removing Gradient Surgery (GS) will reduce the CIDEr score and increase the FID, indicating a worse performance. This demonstrates the importance of mitigating the interference between the optimization directions of two cycles. The complete DoraCycle framework, with both EMA and GS, has the best performance across all metrics, demonstrating the importance of these components in achieving better optimization." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 314, + 519, + 388, + 532 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 314, + 519, + 388, + 532 + ], + "spans": [ + { + "bbox": [ + 314, + 519, + 388, + 532 + ], + "type": "text", + "content": "5. Conclusion" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 312, + 546, + 555, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 312, + 546, + 555, + 715 + ], + "spans": [ + { + "bbox": [ + 312, + 546, + 555, + 715 + ], + "type": "text", + "content": "We propose the DoraCycle to adapt the unified generative model to target domains within multimodal cycles. By leveraging both image-to-text-to-image and text-to-image-to-text cycles, DoraCycle changes the learning objectives into the same modality, allowing for effective optimization using unpaired data. Our experiments show that DoraCycle can adapt the unified model to target domains using only unpaired data, or involving a small amount of paired data when necessary to learn specific concepts. Experimental results demonstrate that DoraCycle achieves advanced or comparable performance across various settings. Leveraging unpaired data broadens the application potential of DoraCycle, making it ideally suited for domain adaptation tasks where paired data is scarce or challenging to collect." + } + ] + } + ], + "index": 13 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "text", + "content": "2842" + } + ] + } + ], + "index": 14 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 7 + }, + { + "para_blocks": [ + { + "bbox": [ + 56, + 72, + 165, + 85 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 72, + 165, + 85 + ], + "spans": [ + { + "bbox": [ + 56, + 72, + 165, + 85 + ], + "type": "text", + "content": "6. Acknowledgments" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 55, + 91, + 296, + 127 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 91, + 296, + 127 + ], + "spans": [ + { + "bbox": [ + 55, + 91, + 296, + 127 + ], + "type": "text", + "content": "This research is supported by the National Research Foundation, Singapore under its AI Singapore Programme (AISG Award No: AISG3-RP-2022-030)." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 56, + 140, + 115, + 153 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 140, + 115, + 153 + ], + "spans": [ + { + "bbox": [ + 56, + 140, + 115, + 153 + ], + "type": "text", + "content": "References" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 57, + 160, + 296, + 713 + ], + "type": "list", + "angle": 0, + "index": 14, + "blocks": [ + { + "bbox": [ + 61, + 160, + 296, + 214 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 61, + 160, + 296, + 214 + ], + "spans": [ + { + "bbox": [ + 61, + 160, + 296, + 214 + ], + "type": "text", + "content": "[1] Armen Aghajanyan, Bernie Huang, Candace Ross, Vladimir Karpukhin, Hu Xu, Naman Goyal, Dmytro Okhonko, Mandar Joshi, Gargi Ghosh, Mike Lewis, et al. Cm3: A causal masked multimodal model of the internet. arXiv preprint arXiv:2201.07520, 2022. 2" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 61, + 217, + 296, + 249 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 61, + 217, + 296, + 249 + ], + "spans": [ + { + "bbox": [ + 61, + 217, + 296, + 249 + ], + "type": "text", + "content": "[2] Emanuele Aiello, LILI YU, Yixin Nie, Armen Aghajanyan, and Barlas Oguz. Jointly training large autoregressive multimodal models. In ICLR, 2024. 2" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 61, + 251, + 296, + 305 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 61, + 251, + 296, + 305 + ], + "spans": [ + { + "bbox": [ + 61, + 251, + 296, + 305 + ], + "type": "text", + "content": "[3] Yogesh Balaji, Seungjun Nah, Xun Huang, Arash Vahdat, Jiaming Song, Karsten Kreis, Miika Aittala, Timo Aila, Samuli Laine, Bryan Catanzaro, et al. ediffi: Text-to-image diffusion models with an ensemble of expert denoisers. arXiv preprint arXiv:2211.01324, 2022. 2" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 62, + 308, + 296, + 361 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 308, + 296, + 361 + ], + "spans": [ + { + "bbox": [ + 62, + 308, + 296, + 361 + ], + "type": "text", + "content": "[4] Soravit Changpinyo, Piyush Sharma, Nan Ding, and Radu Soricut. Conceptual 12m: Pushing web-scale image-text pretraining to recognize long-tail visual concepts. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 3558-3568, 2021. 5" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 62, + 364, + 296, + 418 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 364, + 296, + 418 + ], + "spans": [ + { + "bbox": [ + 62, + 364, + 296, + 418 + ], + "type": "text", + "content": "[5] Junsong Chen, Chongjian Ge, Enze Xie, Yue Wu, Lewei Yao, Xiaozhe Ren, Zhongdao Wang, Ping Luo, Huchuan Lu, and Zhenguo Li. Pixart" + }, + { + "bbox": [ + 62, + 364, + 296, + 418 + ], + "type": "inline_equation", + "content": "\\backslash" + }, + { + "bbox": [ + 62, + 364, + 296, + 418 + ], + "type": "text", + "content": "sigma: Weak-to-strong training of diffusion transformer for 4k text-to-image generation. arXiv preprint arXiv:2403.04692, 2024. 2" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 62, + 421, + 296, + 475 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 421, + 296, + 475 + ], + "spans": [ + { + "bbox": [ + 62, + 421, + 296, + 475 + ], + "type": "text", + "content": "[6] Junsong Chen, Jincheng Yu, Chongjian Ge, Lewei Yao, Enze Xie, Zhongdao Wang, James T. Kwok, Ping Luo, Huchuan Lu, and Zhenguo Li. Pixart-α: Fast training of diffusion transformer for photorealistic text-to-image synthesis. In ICLR. OpenReview.net, 2024. 2" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 62, + 476, + 296, + 531 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 476, + 296, + 531 + ], + "spans": [ + { + "bbox": [ + 62, + 476, + 296, + 531 + ], + "type": "text", + "content": "[7] Li Chen, Mengyi Zhao, Yiheng Liu, Mingxu Ding, Yangyang Song, Shizun Wang, Xu Wang, Hao Yang, Jing Liu, Kang Du, et al. Photovsere: Tuning-free image customization with text-to-image diffusion models. arXiv preprint arXiv:2309.05793, 2023. 2" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 62, + 533, + 296, + 566 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 533, + 296, + 566 + ], + "spans": [ + { + "bbox": [ + 62, + 533, + 296, + 566 + ], + "type": "text", + "content": "[8] Mark Chen, Alec Radford, Rewon Child, Jeffrey Wu, Heewoo Jun, David Luan, and Ilya Sutskever. Generative pretraining from pixels. In ICML, pages 1691-1703, 2020. 2" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 62, + 567, + 296, + 621 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 567, + 296, + 621 + ], + "spans": [ + { + "bbox": [ + 62, + 567, + 296, + 621 + ], + "type": "text", + "content": "[9] Xi Chen, Lianghua Huang, Yu Liu, Yujun Shen, Deli Zhao, and Hengshuang Zhao. Anydoor: Zero-shot object-level image customization. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 6593-6602, 2024. 2" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 57, + 624, + 296, + 678 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 624, + 296, + 678 + ], + "spans": [ + { + "bbox": [ + 57, + 624, + 296, + 678 + ], + "type": "text", + "content": "[10] Yunjey Choi, Minje Choi, Munyoung Kim, Jung-Woo Ha, Sunghun Kim, and Jaegul Choo. Stargan: Unified generative adversarial networks for multi-domain image-to-image translation. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition (CVPR), 2018. 3" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 57, + 680, + 296, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 680, + 296, + 713 + ], + "spans": [ + { + "bbox": [ + 57, + 680, + 296, + 713 + ], + "type": "text", + "content": "[11] Wenliang Dai, Junnan Li, Dongxu Li, Anthony Meng Huat Tiong, Junqi Zhao, Weisheng Wang, Boyang Li, Pascale Fung, and Steven Hoi. Instructlip: Towards general-" + } + ] + } + ], + "index": 13 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 316, + 73, + 553, + 712 + ], + "type": "list", + "angle": 0, + "index": 27, + "blocks": [ + { + "bbox": [ + 333, + 73, + 553, + 94 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 333, + 73, + 553, + 94 + ], + "spans": [ + { + "bbox": [ + 333, + 73, + 553, + 94 + ], + "type": "text", + "content": "purpose vision-language models with instruction tuning, 2023. 2" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 316, + 96, + 553, + 151 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 96, + 553, + 151 + ], + "spans": [ + { + "bbox": [ + 316, + 96, + 553, + 151 + ], + "type": "text", + "content": "[12] Xiaoliang Dai, Ji Hou, Chih-Yao Ma, Sam Tsai, Jialiang Wang, Rui Wang, Peizhao Zhang, Simon Vandenhende, Xiaofang Wang, Abhimanyu Dubey, et al. Emu: Enhancing image generation models using photogenic needles in a haystack. arXiv preprint arXiv:2309.15807, 2023. 2" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 316, + 152, + 553, + 185 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 152, + 553, + 185 + ], + "spans": [ + { + "bbox": [ + 316, + 152, + 553, + 185 + ], + "type": "text", + "content": "[13] Prafulla Dhariwal and Alexander Nichol. Diffusion models beat gans on image synthesis. Advances in Neural Information Processing Systems, 34:8780-8794, 2021. 2" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 317, + 186, + 553, + 241 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 186, + 553, + 241 + ], + "spans": [ + { + "bbox": [ + 317, + 186, + 553, + 241 + ], + "type": "text", + "content": "[14] Runpei Dong, Chunrui Han, Yuang Peng, Zekun Qi, Zheng Ge, Jinrong Yang, Liang Zhao, Jianjian Sun, Hongyu Zhou, Haoran Wei, Xiangwen Kong, Xiangyu Zhang, Kaisheng Ma, and Li Yi. DreamLLM: Synergistic multimodal comprehension and creation. In ICLR, 2024. 1, 2" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 316, + 243, + 553, + 297 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 243, + 553, + 297 + ], + "spans": [ + { + "bbox": [ + 316, + 243, + 553, + 297 + ], + "type": "text", + "content": "[15] Debidatta Dwibedi, Yusuf Aytar, Jonathan Tompson, Pierre Sermanet, and Andrew Zisserman. Temporal cycle-consistency learning. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 1801-1810, 2019. 3" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 317, + 300, + 553, + 354 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 300, + 553, + 354 + ], + "spans": [ + { + "bbox": [ + 317, + 300, + 553, + 354 + ], + "type": "text", + "content": "[16] Patrick Esser, Sumith Kulal, Andreas Blattmann, Rahim Entezari, Jonas Müller, Harry Saini, Yam Levi, Dominik Lorenz, Axel Sauer, Frederic Boesel, et al. Scaling rectified flow transformers for high-resolution image synthesis. arXiv preprint arXiv:2403.03206, 2024. 2" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 317, + 355, + 553, + 409 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 355, + 553, + 409 + ], + "spans": [ + { + "bbox": [ + 317, + 355, + 553, + 409 + ], + "type": "text", + "content": "[17] Rinon Gal, Yuval Alaluf, Yuval Atzmon, Or Patashnik, Amit H Bermano, Gal Chechik, and Daniel Cohen-Or. An image is worth one word: Personalizing text-to-image generation using textual inversion. arXiv preprint arXiv:2208.01618, 2022. 2" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 317, + 411, + 553, + 465 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 411, + 553, + 465 + ], + "spans": [ + { + "bbox": [ + 317, + 411, + 553, + 465 + ], + "type": "text", + "content": "[18] Yuying Ge, Sijie Zhao, Jinguo Zhu, Yixiao Ge, Kun Yi, Lin Song, Chen Li, Xiaohan Ding, and Ying Shan. Seed-x: Multimodal models with unified multi-granularity comprehension and generation. arXiv preprint arXiv:2404.14396, 2024. 1, 2" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 317, + 468, + 553, + 533 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 468, + 553, + 533 + ], + "spans": [ + { + "bbox": [ + 317, + 468, + 553, + 533 + ], + "type": "text", + "content": "[19] Yuchao Gu, Xintao Wang, Jay Zhangjie Wu, Yujun Shi, Yunpeng Chen, Zihan Fan, Wuyou Xiao, Rui Zhao, Shuning Chang, Weijia Wu, et al. Mix-of-show: Decentralized low-rank adaptation for multi-concept customization of diffusion models. Advances in Neural Information Processing Systems, 36:15890-15902, 2023. 2" + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 317, + 535, + 553, + 600 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 535, + 553, + 600 + ], + "spans": [ + { + "bbox": [ + 317, + 535, + 553, + 600 + ], + "type": "text", + "content": "[20] Jiayi Guo, Chaofei Wang, You Wu, Eric Zhang, Kai Wang, Xingqian Xu, Shiji Song, Humphrey Shi, and Gao Huang. Zero-shot generative model adaptation via image-specific prompt learning. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 11494-11503, 2023. 2" + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 316, + 602, + 553, + 645 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 602, + 553, + 645 + ], + "spans": [ + { + "bbox": [ + 316, + 602, + 553, + 645 + ], + "type": "text", + "content": "[21] Di He, Yingce Xia, Tao Qin, Liwei Wang, Nenghai Yu, TieYan Liu, and Wei-Ying Ma. Dual learning for machine translation. In Advances in Neural Information Processing Systems (NeurIPS), 2016. 3" + } + ] + } + ], + "index": 25 + }, + { + "bbox": [ + 316, + 647, + 553, + 712 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 647, + 553, + 712 + ], + "spans": [ + { + "bbox": [ + 316, + 647, + 553, + 712 + ], + "type": "text", + "content": "[22] Martin Heusel, Hubert Ramsauer, Thomas Unterthiner, Bernhard Nessler, and Sepp Hochreiter. Gans trained by a two time-scale update rule converge to a local nash equilibrium. In Proceedings of the 31st International Conference on Neural Information Processing Systems, pages 6626-6637, 2017. 8" + } + ] + } + ], + "index": 26 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "text", + "content": "2843" + } + ] + } + ], + "index": 28 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 8 + }, + { + "para_blocks": [ + { + "bbox": [ + 56, + 72, + 294, + 713 + ], + "type": "list", + "angle": 0, + "index": 13, + "blocks": [ + { + "bbox": [ + 56, + 72, + 294, + 127 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 72, + 294, + 127 + ], + "spans": [ + { + "bbox": [ + 56, + 72, + 294, + 127 + ], + "type": "text", + "content": "[23] Judy Hoffman, Eric Tzeng, Taesung Park, Jun-Yan Zhu, Phillip Isola, Kate Saenko, Alexei A Efros, and Trevor Darryll. Cycada: Cycle-consistent adversarial domain adaptation. In Proceedings of the International Conference on Machine Learning (ICML), 2018. 3" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 56, + 128, + 294, + 172 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 128, + 294, + 172 + ], + "spans": [ + { + "bbox": [ + 56, + 128, + 294, + 172 + ], + "type": "text", + "content": "[24] MD Zakir Hossain, Ferdous Sohel, Mohd Fairuz Shiratuddin, and Hamid Laga. A comprehensive survey of deep learning for image captioning. ACM Computing Surveys (CsUR), 51(6):1-36, 2019. 2" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 56, + 172, + 294, + 216 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 172, + 294, + 216 + ], + "spans": [ + { + "bbox": [ + 56, + 172, + 294, + 216 + ], + "type": "text", + "content": "[25] Edward J. Hu, Yelong Shen, Phillip Wallis, Zeyuan Allen-Zhu, Yuanzhi Li, Shean Wang, and Weizhu Chen. Lora: Low-rank adaptation of large language models. In International Conference on Learning Representations, 2022. 5" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 56, + 217, + 294, + 270 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 217, + 294, + 270 + ], + "spans": [ + { + "bbox": [ + 56, + 217, + 294, + 270 + ], + "type": "text", + "content": "[26] Jia Cheng Hu, Roberto Cavicchioli, and Alessandro Capontodi. Exploiting multiple sequence lengths in fast end to end training for image captioning. In 2023 IEEE International Conference on Big Data (BigData), pages 2173-2182. IEEE, 2023. 2" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 56, + 272, + 294, + 326 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 272, + 294, + 326 + ], + "spans": [ + { + "bbox": [ + 56, + 272, + 294, + 326 + ], + "type": "text", + "content": "[27] Xiaowei Hu, Zhe Gan, Jianfeng Wang, Zhengyuan Yang, Zicheng Liu, Yumao Lu, and Lijuan Wang. Scaling up vision-language pre-training for image captioning. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 17980-17989, 2022. 2" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 56, + 327, + 294, + 358 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 327, + 294, + 358 + ], + "spans": [ + { + "bbox": [ + 56, + 327, + 294, + 358 + ], + "type": "text", + "content": "[28] Eric Jang, Shixiang Gu, and Ben Poole. Categorical reparameterization with gumbel-softmax. arXiv preprint arXiv:1611.01144, 2016. 4" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 56, + 360, + 294, + 415 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 360, + 294, + 415 + ], + "spans": [ + { + "bbox": [ + 56, + 360, + 294, + 415 + ], + "type": "text", + "content": "[29] Xuhui Jia, Yang Zhao, Kelvin CK Chan, Yandong Li, Han Zhang, Boqing Gong, Tingbo Hou, Huisheng Wang, and Yu-Chuan Su. Taming encoder for zero fine-tuning image customization with text-to-image diffusion models. arXiv preprint arXiv:2304.02642, 2023. 2" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 56, + 415, + 294, + 470 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 415, + 294, + 470 + ], + "spans": [ + { + "bbox": [ + 56, + 415, + 294, + 470 + ], + "type": "text", + "content": "[30] Nupur Kumari, Bingliang Zhang, Richard Zhang, Eli Shechtman, and Jun-Yan Zhu. Multi-concept customization of text-to-image diffusion. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 1931-1941, 2023. 2" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 56, + 471, + 294, + 525 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 471, + 294, + 525 + ], + "spans": [ + { + "bbox": [ + 56, + 471, + 294, + 525 + ], + "type": "text", + "content": "[31] Chenliang Li, Haiyang Xu, Junfeng Tian, Wei Wang, Ming Yan, Bin Bi, Jiabo Ye, Hehong Chen, Guohai Xu, Zheng Cao, et al. mplug: Effective and efficient vision-language learning by cross-modal skip-connections. arXiv preprint arXiv:2205.12005, 2022. 2" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 56, + 525, + 294, + 580 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 525, + 294, + 580 + ], + "spans": [ + { + "bbox": [ + 56, + 525, + 294, + 580 + ], + "type": "text", + "content": "[32] Chunyuan Li, Zhe Gan, Zhengyuan Yang, Jianwei Yang, Linjie Li, Lijuan Wang, Jianfeng Gao, et al. Multimodal foundation models: From specialists to general-purpose assistants. Foundations and Trends® in Computer Graphics and Vision, 16(1-2):1-214, 2024. 2" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 56, + 581, + 294, + 635 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 581, + 294, + 635 + ], + "spans": [ + { + "bbox": [ + 56, + 581, + 294, + 635 + ], + "type": "text", + "content": "[33] Junnan Li, Dongxu Li, Silvio Savarese, and Steven Hoi. Blip-2: Bootstrapping language-image pre-training with frozen image encoders and large language models. In International conference on machine learning, pages 19730-19742. PMLR, 2023. 2" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 56, + 635, + 294, + 690 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 635, + 294, + 690 + ], + "spans": [ + { + "bbox": [ + 56, + 635, + 294, + 690 + ], + "type": "text", + "content": "[34] Tianhong Li, Sangnie Bhardwaj, Yonglong Tian, Han Zhang, Jarred Barber, Dina Katabi, Guillaume Lajoie, Huiwen Chang, and Dilip Krishnan. Leveraging unpaired data for vision-language generative models via cycle consistency. arXiv preprint arXiv:2310.03734, 2023. 3, 8" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 56, + 691, + 294, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 691, + 294, + 713 + ], + "spans": [ + { + "bbox": [ + 56, + 691, + 294, + 713 + ], + "type": "text", + "content": "[35] Haotian Liu, Chunyuan Li, Qingyang Wu, and Yong Jae Lee. Visual instruction tuning. NeurIPS, 36, 2024. 2" + } + ] + } + ], + "index": 12 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 316, + 72, + 555, + 713 + ], + "type": "list", + "angle": 0, + "index": 28, + "blocks": [ + { + "bbox": [ + 316, + 72, + 553, + 126 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 72, + 553, + 126 + ], + "spans": [ + { + "bbox": [ + 316, + 72, + 553, + 126 + ], + "type": "text", + "content": "[36] Nanye Ma, Mark Goldstein, Michael S Albergo, Nicholas M Boffi, Eric Vanden-Eijnden, and Saining Xie. Sit: Exploring flow and diffusion-based generative models with scalable interpolant transformers. arXiv preprint arXiv:2401.08740, 2024. 2" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 316, + 128, + 553, + 173 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 128, + 553, + 173 + ], + "spans": [ + { + "bbox": [ + 316, + 128, + 553, + 173 + ], + "type": "text", + "content": "[37] Xin Ma, Yaohui Wang, Gengyun Jia, Xinyuan Chen, Zwei Liu, Yuan-Fang Li, Cunjian Chen, and Yu Qiao. Latte: Latent diffusion transformer for video generation. arXiv preprint arXiv:2401.03048, 2024. 2" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 316, + 174, + 555, + 229 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 174, + 555, + 229 + ], + "spans": [ + { + "bbox": [ + 316, + 174, + 555, + 229 + ], + "type": "text", + "content": "[38] Ron Mokady, Amir Hertz, Kfir Aberman, Yael Pritch, and Daniel Cohen-Or. Null-text inversion for editing real images using guided diffusion models. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 6038–6047, 2023. 2" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 316, + 230, + 545, + 241 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 230, + 545, + 241 + ], + "spans": [ + { + "bbox": [ + 316, + 230, + 545, + 241 + ], + "type": "text", + "content": "[39] OpenAI. Chatgpt. https://chatgpt.com/, 2024.7" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 316, + 242, + 553, + 307 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 242, + 553, + 307 + ], + "spans": [ + { + "bbox": [ + 316, + 242, + 553, + 307 + ], + "type": "text", + "content": "[40] Alec Radford, Jong Wook Kim, Chris Hallacy, Aditya Ramesh, Gabriel Goh, Sandhini Agarwal, Girish Sastry, Amanda Askell, Pamela Mishkin, Jack Clark, et al. Learning transferable visual models from natural language supervision. In International conference on machine learning, pages 8748-8763. PMLR, 2021. 2" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 316, + 308, + 553, + 353 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 308, + 553, + 353 + ], + "spans": [ + { + "bbox": [ + 316, + 308, + 553, + 353 + ], + "type": "text", + "content": "[41] Aditya Ramesh, Mikhail Pavlov, Gabriel Goh, Scott Gray, Chelsea Voss, Alec Radford, Mark Chen, and Ilya Sutskever. Zero-shot text-to-image generation. In ICML, pages 8821-8831. Pmlr, 2021. 2" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 316, + 354, + 553, + 387 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 354, + 553, + 387 + ], + "spans": [ + { + "bbox": [ + 316, + 354, + 553, + 387 + ], + "type": "text", + "content": "[42] Aditya Ramesh, Prafulla Dhariwal, Alex Nichol, Casey Chu, and Mark Chen. Hierarchical text-conditional image generation with CLIP latents. CoRR, abs/2204.06125, 2022. 2" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 316, + 388, + 553, + 430 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 388, + 553, + 430 + ], + "spans": [ + { + "bbox": [ + 316, + 388, + 553, + 430 + ], + "type": "text", + "content": "[43] Aditya Ramesh, Prafulla Dhariwal, Alex Nichol, Casey Chu, and Mark Chen. Hierarchical diffusion models for text-to-image generation. arXiv preprint arXiv:2204.06125, 2022. 2" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 316, + 433, + 553, + 477 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 433, + 553, + 477 + ], + "spans": [ + { + "bbox": [ + 316, + 433, + 553, + 477 + ], + "type": "text", + "content": "[44] Aditya Ramesh, Prafulla Dhariwal, Alex Nichol, Casey Chu, and Mark Chen. Hierarchical text-conditional image generation with clip latents. arXiv preprint arXiv:2204.06125, 1 (2):3, 2022. 2" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 316, + 478, + 553, + 521 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 478, + 553, + 521 + ], + "spans": [ + { + "bbox": [ + 316, + 478, + 553, + 521 + ], + "type": "text", + "content": "[45] Robin Rombach, Andreas Blattmann, Dominik Lorenz, Patrick Esser, and Björn Ommer. High-resolution image synthesis with latent diffusion models. In CVPR, pages 10684-10695, 2022. 2" + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 316, + 523, + 553, + 567 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 523, + 553, + 567 + ], + "spans": [ + { + "bbox": [ + 316, + 523, + 553, + 567 + ], + "type": "text", + "content": "[46] Robin Rombach, Andreas Blattmann, Dominik Lorenz, Patrick Esser, and Björn Ommer. High-resolution image synthesis with latent diffusion models. arXiv preprint arXiv:2112.10752, 2022. 2" + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 316, + 568, + 553, + 612 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 568, + 553, + 612 + ], + "spans": [ + { + "bbox": [ + 316, + 568, + 553, + 612 + ], + "type": "text", + "content": "[47] Nataniel Ruiz, Yuanzhen Li, Varun Jampani, Yael Pritch, Michael Rubinstein, and Kfir Aberman. Dreambooth: Fine tuning text-to-image diffusion models for subject-driven generation. arXiv preprint arXiv:2208.12242, 2022. 1, 2" + } + ] + } + ], + "index": 25 + }, + { + "bbox": [ + 316, + 613, + 553, + 678 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 613, + 553, + 678 + ], + "spans": [ + { + "bbox": [ + 316, + 613, + 553, + 678 + ], + "type": "text", + "content": "[48] Nataniel Ruiz, Yuanzhen Li, Varun Jampani, Yael Pritch, Michael Rubinstein, and Kfir Aberman. Dreambooth: Fine tuning text-to-image diffusion models for subject-driven generation. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 22500-22510, 2023. 5, 8" + } + ] + } + ], + "index": 26 + }, + { + "bbox": [ + 316, + 679, + 553, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 679, + 553, + 713 + ], + "spans": [ + { + "bbox": [ + 316, + 679, + 553, + 713 + ], + "type": "text", + "content": "[49] Chitwan Sahara, William Chan, Saurabh Saxena, Lala Li, Jay Whang, Emily L Denton, Kamyar Ghasemipour, Raphael Gontijo Lopes, Burcu Karagol Ayan, Tim Salimans," + } + ] + } + ], + "index": 27 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "text", + "content": "2844" + } + ] + } + ], + "index": 29 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 9 + }, + { + "para_blocks": [ + { + "bbox": [ + 56, + 72, + 295, + 713 + ], + "type": "list", + "angle": 0, + "index": 15, + "blocks": [ + { + "bbox": [ + 76, + 72, + 294, + 106 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 76, + 72, + 294, + 106 + ], + "spans": [ + { + "bbox": [ + 76, + 72, + 294, + 106 + ], + "type": "text", + "content": "et al. Photorealistic text-to-image diffusion models with deep language understanding. Advances in Neural Information Processing Systems, 35:36479-36494, 2022. 2" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 56, + 106, + 295, + 162 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 106, + 295, + 162 + ], + "spans": [ + { + "bbox": [ + 56, + 106, + 295, + 162 + ], + "type": "text", + "content": "[50] Rico Sennrich, Barry Haddow, and Alexandra Birch. Improving neural machine translation models with monolingual data. In Proceedings of the 54th Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers), 2016. 3" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 56, + 162, + 294, + 205 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 162, + 294, + 205 + ], + "spans": [ + { + "bbox": [ + 56, + 162, + 294, + 205 + ], + "type": "text", + "content": "[51] Meet Shah, Xinlei Chen, Marcus Rohrbach, and Devi Parikh. Cycle-consistency for robust visual question answering. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 6649-6658, 2019. 3" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 56, + 206, + 294, + 250 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 206, + 294, + 250 + ], + "spans": [ + { + "bbox": [ + 56, + 206, + 294, + 250 + ], + "type": "text", + "content": "[52] Sameer Shah, Siddharth Bharadwaj, Devi Parikh, and Dhruv Batra. Cycle-consistency for robust visual question answering. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition (CVPR), 2019. 3" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 56, + 250, + 294, + 316 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 250, + 294, + 316 + ], + "spans": [ + { + "bbox": [ + 56, + 250, + 294, + 316 + ], + "type": "text", + "content": "[53] Quan Sun, Yufeng Cui, Xiaosong Zhang, Fan Zhang, Qiying Yu, Yueze Wang, Yongming Rao, Jingjing Liu, Tiejun Huang, and Xinlong Wang. Generative multimodal models are in-context learners. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 14398-14409, 2024. 2" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 56, + 316, + 294, + 338 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 316, + 294, + 338 + ], + "spans": [ + { + "bbox": [ + 56, + 316, + 294, + 338 + ], + "type": "text", + "content": "[54] I Sutskever. Sequence to sequence learning with neural networks. arXiv preprint arXiv:1409.3215, 2014. 4" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 56, + 338, + 294, + 360 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 338, + 294, + 360 + ], + "spans": [ + { + "bbox": [ + 56, + 338, + 294, + 360 + ], + "type": "text", + "content": "[55] Richard S Sutton. Learning to predict by the methods of temporal differences. Machine learning, 3:9-44, 1988. 4" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 56, + 361, + 294, + 392 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 361, + 294, + 392 + ], + "spans": [ + { + "bbox": [ + 56, + 361, + 294, + 392 + ], + "type": "text", + "content": "[56] Zineng Tang, Ziyi Yang, Chenguang Zhu, Michael Zeng, and Mohit Bansal. Any-to-any generation via composable diffusion. NeurIPS, 36, 2024. 2" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 56, + 393, + 294, + 446 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 393, + 294, + 446 + ], + "spans": [ + { + "bbox": [ + 56, + 393, + 294, + 446 + ], + "type": "text", + "content": "[57] Antti Tarvainen and Harri Valpola. Mean teachers are better role models: Weight-averaged consistency targets improve semi-supervised deep learning results. In Advances in Neural Information Processing Systems, pages 1195–1204, 2017. 2, 4" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 56, + 449, + 294, + 480 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 449, + 294, + 480 + ], + "spans": [ + { + "bbox": [ + 56, + 449, + 294, + 480 + ], + "type": "text", + "content": "[58] Chameleon Team. Chameleon: Mixed-modal early-fusion foundation models. arXiv preprint arXiv:2405.09818, 2024. 1, 2, 3" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 56, + 482, + 294, + 535 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 482, + 294, + 535 + ], + "spans": [ + { + "bbox": [ + 56, + 482, + 294, + 535 + ], + "type": "text", + "content": "[59] Ramakrishna Vedantam, C. Lawrence Zitnick, and Devi Parikh. Cider: Consensus-based image description evaluation. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition (CVPR), pages 4566-4575, 2015. 8" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 56, + 537, + 294, + 581 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 537, + 294, + 581 + ], + "spans": [ + { + "bbox": [ + 56, + 537, + 294, + 581 + ], + "type": "text", + "content": "[60] Jianfeng Wang, Zhengyuan Yang, Xiaowei Hu, Linjie Li, Kevin Lin, Zhe Gan, Zicheng Liu, Ce Liu, and Lijuan Wang. Git: A generative image-to-text transformer for vision and language. arXiv preprint arXiv:2205.14100, 2022. 2" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 56, + 582, + 294, + 647 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 582, + 294, + 647 + ], + "spans": [ + { + "bbox": [ + 56, + 582, + 294, + 647 + ], + "type": "text", + "content": "[61] Peng Wang, An Yang, Rui Men, Junyang Lin, Shuai Bai, Zhikang Li, Jianxin Ma, Chang Zhou, Jingren Zhou, and Hongxia Yang. Ofa: Unifying architectures, tasks, and modalities through a simple sequence-to-sequence learning framework. In International conference on machine learning, pages 23318-23340. PMLR, 2022. 2" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 56, + 647, + 294, + 690 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 647, + 294, + 690 + ], + "spans": [ + { + "bbox": [ + 56, + 647, + 294, + 690 + ], + "type": "text", + "content": "[62] Qixun Wang, Xu Bai, Haofan Wang, Zekui Qin, Anthony Chen, Huaxia Li, Xu Tang, and Yao Hu. Instantid: Zero-shot identity-preserving generation in seconds. arXiv preprint arXiv:2401.07519, 2024. 1" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 56, + 691, + 294, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 691, + 294, + 713 + ], + "spans": [ + { + "bbox": [ + 56, + 691, + 294, + 713 + ], + "type": "text", + "content": "[63] Yuchi Wang, Shuhuai Ren, Rundong Gao, Linli Yao, Qingyan Guo, Kaikai An, Jianhong Bai, and Xu Sun." + } + ] + } + ], + "index": 14 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 316, + 73, + 553, + 713 + ], + "type": "list", + "angle": 0, + "index": 29, + "blocks": [ + { + "bbox": [ + 333, + 73, + 553, + 106 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 333, + 73, + 553, + 106 + ], + "spans": [ + { + "bbox": [ + 333, + 73, + 553, + 106 + ], + "type": "text", + "content": "Ladic: Are diffusion models really inferior to autoregressive counterparts for image-to-text generation? arXiv preprint arXiv:2404.10763, 2024. 2" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 316, + 106, + 553, + 161 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 106, + 553, + 161 + ], + "spans": [ + { + "bbox": [ + 316, + 106, + 553, + 161 + ], + "type": "text", + "content": "[64] Chengyue Wu, Xiaokang Chen, Zhiyu Wu, Yiyang Ma, Xingchao Liu, Zizheng Pan, Wen Liu, Zhenda Xie, Xingkai Yu, Chong Ruan, et al. Janus: Decoupling visual encoding for unified multimodal understanding and generation. arXiv preprint arXiv:2410.13848, 2024. 2" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 316, + 162, + 553, + 194 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 162, + 553, + 194 + ], + "spans": [ + { + "bbox": [ + 316, + 162, + 553, + 194 + ], + "type": "text", + "content": "[65] Shengqiong Wu, Hao Fei, Leigang Qu, Wei Ji, and Tat-Seng Chua. Next-gpt: Any-to-any multimodal llm. arXiv preprint arXiv:2309.05519, 2023. 2" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 316, + 195, + 553, + 249 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 195, + 553, + 249 + ], + "spans": [ + { + "bbox": [ + 316, + 195, + 553, + 249 + ], + "type": "text", + "content": "[66] Jinheng Xie, Jiajun Feng, Zhaoxu Tian, Kevin Qinghong Lin, Yawen Huang, Xi Xia, Nanxu Gong, Xu Zuo, Ji-aqi Yang, Yefeng Zheng, et al. Learning long-form video prior via generative pre-training. arXiv preprint arXiv:2404.15909, 2024. 7" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 316, + 250, + 553, + 315 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 250, + 553, + 315 + ], + "spans": [ + { + "bbox": [ + 316, + 250, + 553, + 315 + ], + "type": "text", + "content": "[67] Jinheng Xie, Weijia Mao, Zechen Bai, David Junhao Zhang, Weihao Wang, Kevin Qinghong Lin, Yuchao Gu, Zhijie Chen, Zhenheng Yang, and Mike Zheng Shou. Show-o: One single transformer to unify multimodal understanding and generation. arXiv preprint arXiv:2408.12528, 2024. 1, 2, 3, 5" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 316, + 316, + 553, + 369 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 316, + 553, + 369 + ], + "spans": [ + { + "bbox": [ + 316, + 316, + 553, + 369 + ], + "type": "text", + "content": "[68] Ceyuan Yang, Yujun Shen, Zhiyi Zhang, Yinghao Xu, Jiapeng Zhu, Zhirong Wu, and Bolei Zhou. One-shot generative domain adaptation. In Proceedings of the IEEE/cvf international conference on computer vision, pages 7733-7742, 2023. 2" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 316, + 371, + 553, + 415 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 371, + 553, + 415 + ], + "spans": [ + { + "bbox": [ + 316, + 371, + 553, + 415 + ], + "type": "text", + "content": "[69] Hanrong Ye, De-An Huang, Yao Lu, Zhiding Yu, Wei Ping, Andrew Tao, Jan Kautz, Song Han, Dan Xu, Pavlo Molchanov, et al. X-vila: Cross-modality alignment for large language model. arXiv preprint arXiv:2405.19335, 2024. 2" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 316, + 416, + 553, + 459 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 416, + 553, + 459 + ], + "spans": [ + { + "bbox": [ + 316, + 416, + 553, + 459 + ], + "type": "text", + "content": "[70] Haoxuan You, Mandy Guo, Zhecan Wang, Kai-Wei Chang, Jason Baldridge, and Jiahui Yu. Cobit: A contrastive bidirectional image-text generation model. arXiv preprint arXiv:2303.13455, 2023." + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 316, + 460, + 553, + 514 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 460, + 553, + 514 + ], + "spans": [ + { + "bbox": [ + 316, + 460, + 553, + 514 + ], + "type": "text", + "content": "[71] Lili Yu, Bowen Shi, Ramakanth Pasunuru, Benjamin Muller, Olga Golovneva, Tianlu Wang, Arun Babu, Binh Tang, Brian Karrer, Shelly Sheynin, et al. Scaling autoregressive multimodal models: Pretraining and instruction tuning. arXiv preprint arXiv:2309.02591, 2(3), 2023. 2" + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 316, + 515, + 553, + 559 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 515, + 553, + 559 + ], + "spans": [ + { + "bbox": [ + 316, + 515, + 553, + 559 + ], + "type": "text", + "content": "[72] Tianhe Yu, Saurabh Kumar, Abhishek Gupta, Sergey Levine, Karol Hausman, and Chelsea Finn. Gradient surgery for multi-task learning. In Advances in Neural Information Processing Systems, pages 5824-5836, 2020. 4" + } + ] + } + ], + "index": 25 + }, + { + "bbox": [ + 316, + 559, + 553, + 602 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 559, + 553, + 602 + ], + "spans": [ + { + "bbox": [ + 316, + 559, + 553, + 602 + ], + "type": "text", + "content": "[73] Gengwei Zhang, Guoliang Kang, Yi Yang, and Yunchao Wei. Few-shot segmentation via cycle-consistent transformer. Advances in Neural Information Processing Systems, 34:21984-21996, 2021. 3" + } + ] + } + ], + "index": 26 + }, + { + "bbox": [ + 316, + 603, + 553, + 658 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 603, + 553, + 658 + ], + "spans": [ + { + "bbox": [ + 316, + 603, + 553, + 658 + ], + "type": "text", + "content": "[74] Rui Zhao, Wei Li, Zhipeng Hu, Lincheng Li, Zhengxia Zou, Zhenwei Shi, and Changjie Fan. Zero-shot text-to-parameter translation for game character auto-creation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 21013–21023, 2023. 2" + } + ] + } + ], + "index": 27 + }, + { + "bbox": [ + 316, + 658, + 553, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 658, + 553, + 713 + ], + "spans": [ + { + "bbox": [ + 316, + 658, + 553, + 713 + ], + "type": "text", + "content": "[75] Rui Zhao, Yuchao Gu, Jay Zhangjie Wu, David Junhao Zhang, Jia-Wei Liu, Weijia Wu, Jussi Keppo, and Mike Zheng Shou. Motiondirector: Motion customization of text-to-video diffusion models. In European Conference on Computer Vision, pages 273–290. Springer, 2024." + } + ] + } + ], + "index": 28 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "text", + "content": "2845" + } + ] + } + ], + "index": 30 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 10 + }, + { + "para_blocks": [ + { + "bbox": [ + 56, + 73, + 294, + 293 + ], + "type": "list", + "angle": 0, + "index": 4, + "blocks": [ + { + "bbox": [ + 56, + 73, + 294, + 139 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 73, + 294, + 139 + ], + "spans": [ + { + "bbox": [ + 56, + 73, + 294, + 139 + ], + "type": "text", + "content": "[76] Rui Zhao, Hangjie Yuan, Yujie Wei, Shiwei Zhang, Yuchao Gu, Lingmin Ran, Xiang Wang, Jay Zhangjie Wu, David Junhao Zhang, Yingya Zhang, et al. Evolvedirector: Approaching advanced text-to-image generation with large vision-language models. Advances in Neural Information Processing Systems, 37:122104-122129, 2025. 2" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 56, + 140, + 294, + 195 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 140, + 294, + 195 + ], + "spans": [ + { + "bbox": [ + 56, + 140, + 294, + 195 + ], + "type": "text", + "content": "[77] Chunting Zhou, Lili Yu, Arun Babu, Kushal Tirumala, Michihiro Yasunaga, Leonid Shamis, Jacob Kahn, Xuezhe Ma, Luke Zettlemoyer, and Omer Levy. Transfusion: Predict the next token and diffuse images with one multi-modal model. arXiv preprint arXiv:2408.11039, 2024. 1, 2, 3" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 56, + 197, + 294, + 239 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 197, + 294, + 239 + ], + "spans": [ + { + "bbox": [ + 56, + 197, + 294, + 239 + ], + "type": "text", + "content": "[78] Deyao Zhu, Jun Chen, Xiaoqian Shen, Xiang Li, and Mohamed Elhoseiny. Minigpt-4: Enhancing vision-language understanding with advanced large language models. CoRR, abs/2304.10592, 2023. 2" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 56, + 241, + 294, + 293 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 241, + 294, + 293 + ], + "spans": [ + { + "bbox": [ + 56, + 241, + 294, + 293 + ], + "type": "text", + "content": "[79] Jun-Yan Zhu, Taesung Park, Phillip Isola, and Alexei A. Efros. Unpaired image-to-image translation using cycle-consistent adversarial networks. In Proceedings of the IEEE International Conference on Computer Vision (ICCV), 2017. 3" + } + ] + } + ], + "index": 3 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 749, + 316, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 749, + 316, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 749, + 316, + 757 + ], + "type": "text", + "content": "2846" + } + ] + } + ], + "index": 5 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 11 + } + ], + "_backend": "vlm", + "_version_name": "2.6.4" +} \ No newline at end of file diff --git a/2025/Dora_ Sampling and Benchmarking for 3D Shape Variational Auto-Encoders/6b478709-2cef-4587-8775-92fc50ace470_content_list.json b/2025/Dora_ Sampling and Benchmarking for 3D Shape Variational Auto-Encoders/6b478709-2cef-4587-8775-92fc50ace470_content_list.json new file mode 100644 index 0000000000000000000000000000000000000000..f56cbbc42399b9f5002f0d8b44f18d5375cd761b --- /dev/null +++ b/2025/Dora_ Sampling and Benchmarking for 3D Shape Variational Auto-Encoders/6b478709-2cef-4587-8775-92fc50ace470_content_list.json @@ -0,0 +1,1890 @@ +[ + { + "type": "text", + "text": "Dora: Sampling and Benchmarking for 3D Shape Variational Auto-Encoders", + "text_level": 1, + "bbox": [ + 109, + 130, + 885, + 152 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Rui Chen $^{1,2}$ Jianfeng Zhang $^{2\\dagger}$ Yixun Liang $^{1,3}$ Guan Luo $^{2,4}$ Weiyu Li $^{1,3}$ \nJiarui Liu $^{1,3}$ Xiu Li $^{2}$ Xiaoxiao Long $^{1,3}$ Jiashi Feng $^{2}$ Ping Tan $^{1,3\\dagger}$ \n†Corresponding authors \n $^{1}$ The Hong Kong University of Science and Technology $^{2}$ ByteDance Seed \n $^{3}$ LightIllusions $^{4}$ Tsinghua University \nhttps://aruichen.github.io/Dora", + "bbox": [ + 192, + 178, + 803, + 287 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Abstract", + "text_level": 1, + "bbox": [ + 246, + 321, + 326, + 338 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Recent 3D content generation pipelines commonly employ Variational Autoencoders (VAEs) to encode shapes into compact latent representations for diffusion-based generation. However, the widely adopted uniform point sampling strategy in Shape VAE training often leads to a significant loss of geometric details, limiting the quality of shape reconstruction and downstream generation tasks. We present Dora-VAE, a novel approach that enhances VAE reconstruction through our proposed sharp edge sampling strategy and a dual cross-attention mechanism. By identifying and prioritizing regions with high geometric complexity during training, our method significantly improves the preservation of fine-grained shape features. Such sampling strategy and the dual attention mechanism enable the VAE to focus on crucial geometric details that are typically missed by uniform sampling approaches. To systematically evaluate VAE reconstruction quality, we additionally propose Dora-bench, a benchmark that quantifies shape complexity through the density of sharp edges, introducing a new metric focused on reconstruction accuracy at these salient geometric features. Extensive experiments on the Dora-bench demonstrate that Dora-VAE achieves comparable reconstruction quality to the state-of-the-art dense XCube-VAE while requiring a latent space at least $8 \\times$ smaller (1,280 vs. $>10,000$ codes). Project page: https://aruichen.github.io/Dora.", + "bbox": [ + 89, + 357, + 485, + 750 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "1. Introduction", + "text_level": 1, + "bbox": [ + 91, + 782, + 220, + 797 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "3D content creation is vital to delivering realistic and immersive experiences in various industries, including games, movies, and AR/VR. However, traditional 3D modeling typically demands significant expertise and manual effort, making it time-consuming and challenging, especially for non-expert users. Recent advances in AI-powered 3D con", + "bbox": [ + 89, + 809, + 483, + 901 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "tent generation methods [2, 3, 7, 9, 17, 21, 25, 28, 30, 42, 44, 57] have transformed the field, making it more accessible to many more users.", + "bbox": [ + 511, + 323, + 903, + 367 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Following the success of text-to-image generation models [4, 5, 15, 37, 55], recent 3D content creation approaches [36, 49, 56] adopt a two-stage pipeline: encoding 3D shapes into a latent space using variational autoencoders (VAEs), followed by training a latent diffusion model. The performance of such a generative pipeline heavily relies on the VAE's capability to faithfully encode and reconstruct 3D shapes.", + "bbox": [ + 511, + 369, + 906, + 489 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Existing 3D VAEs operate by sampling points on mesh surfaces for shape encoding and then reconstructing the original 3D meshes by its decoder. This process faces unique challenges compared to 2D image VAEs where the input image is fully observable. In comparison, the sampled point cloud often cannot capture all the necessary shape information, which could harm the performance of 3D VAEs.", + "bbox": [ + 511, + 489, + 905, + 597 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Volume-based method [36] leverages sparse convolution [47] to process millions of voxelized points for high-fidelity reconstruction. Its dense sampling captures precise shape information. However, this method produces large latent codes (commonly $>10,000$ tokens), which significantly complicate the training of diffusion models. On the other hand, vector-set (Vecset) methods [24, 54, 56, 58] use transformers to achieve compact latent representations (hundreds to thousands of tokens), enabling efficient diffusion [24, 56]. However, due to the quadratic complexity of transformer networks, it often only samples a few thousand points to represent a 3D shape, which leads to information loss and performance degradation. Therefore, we seek to improve the reconstruction quality of Vecset-based VAEs and maintain their compact representation.", + "bbox": [ + 511, + 598, + 906, + 824 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "We begin by analyzing the shape reconstruction capability of Vecset-based VAEs. Through careful analysis, we find these methods have limited reconstruction performance, which stems from their commonly used uniform sampling. When computational constraints limit the total", + "bbox": [ + 511, + 825, + 903, + 900 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "CVF", + "bbox": [ + 106, + 2, + 181, + 42 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore.", + "bbox": [ + 236, + 0, + 810, + 46 + ], + "page_idx": 0 + }, + { + "type": "page_number", + "text": "16251", + "bbox": [ + 480, + 944, + 517, + 955 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/1d7091c2be5c9b7d56c50afe9c7af043c387fd6b81d9f0d95246d92e7a0806c3.jpg", + "image_caption": [ + "(a) Ground Truth(GT)" + ], + "image_footnote": [], + "bbox": [ + 122, + 90, + 259, + 185 + ], + "page_idx": 1 + }, + { + "type": "image", + "img_path": "images/3723c8ddf5058b50cb5b8b944adf45a354d3034414836c04f599aac1b01dbed1.jpg", + "image_caption": [ + "(d) Reconstruction accuracy", + "Figure 1. Sampling strategy comparison. Given the ground truth mesh shown in (a), we visualize point clouds produced by uniform sampling in (b) and those generated by our proposed Sharp Edge Sampling (SES) in (c), at various sampling rates. In (d), we compare the reconstruction accuracy trained with our SES and the uniform sampling using the F-score metric. The comparison demonstrates SES consistently outperforms uniform sampling under varying sampling rates, as the point clouds generated by SES are more effective in capturing the salient features of the object." + ], + "image_footnote": [], + "bbox": [ + 102, + 196, + 279, + 334 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "point number", + "bbox": [ + 287, + 94, + 323, + 112 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "4096", + "bbox": [ + 343, + 98, + 370, + 107 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "8192", + "bbox": [ + 465, + 98, + 493, + 107 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "16384", + "bbox": [ + 583, + 98, + 616, + 107 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "20480", + "bbox": [ + 696, + 98, + 730, + 107 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "40960", + "bbox": [ + 821, + 98, + 854, + 107 + ], + "page_idx": 1 + }, + { + "type": "image", + "img_path": "images/61b761c8c0524fdbefc7fee85173068b8de6ee8ecbadde13ff5b5c01b104cf6f.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 305, + 126, + 421, + 209 + ], + "page_idx": 1 + }, + { + "type": "image", + "img_path": "images/3791fd82f0a53b4151db302506296faedd3f58194ee58670b9ea10fce597f1a0.jpg", + "image_caption": [ + "(b) Visual results of the uniform sampling" + ], + "image_footnote": [], + "bbox": [ + 426, + 127, + 540, + 209 + ], + "page_idx": 1 + }, + { + "type": "image", + "img_path": "images/718df31d5a83af4156943caa575e9190d455c9c0fc75e022af8f1e707517dd3e.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 547, + 127, + 661, + 209 + ], + "page_idx": 1 + }, + { + "type": "image", + "img_path": "images/d194ca7c9733c32f31613aae2706321ba6bdb020abe917ab69ac1f73fb0c1b16.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 666, + 127, + 782, + 209 + ], + "page_idx": 1 + }, + { + "type": "image", + "img_path": "images/c8b578ccaeb00d7c59a2bb47d98b19908a0c59fa44b82f4a0efdd5cab138d2e2.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 787, + 128, + 901, + 209 + ], + "page_idx": 1 + }, + { + "type": "image", + "img_path": "images/10420fe341601ac28b74e94a6b7974bfb0b6d3ebb3f785d8eac49363d427b2be.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 305, + 237, + 421, + 319 + ], + "page_idx": 1 + }, + { + "type": "image", + "img_path": "images/0e5defacb8d30c2a7f52511540f4ac6ba76050768d5784e1bbb978b333d5414e.jpg", + "image_caption": [ + "(c) Visual results of our sharp edge sampling (SES)" + ], + "image_footnote": [], + "bbox": [ + 426, + 237, + 542, + 319 + ], + "page_idx": 1 + }, + { + "type": "image", + "img_path": "images/18a1a5cb08e34fdaad47c637593d26cf9f5228855c12451dd979ef4bfc1f4d50.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 547, + 237, + 661, + 319 + ], + "page_idx": 1 + }, + { + "type": "image", + "img_path": "images/cb2c82564119ae842784851ef1448d769987a3ad74a4674936b0276f111ec738.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 666, + 237, + 782, + 319 + ], + "page_idx": 1 + }, + { + "type": "image", + "img_path": "images/da169309bca4faf0378c628b5eaf7703bd0c38df8c5f59d04f8753c253e7449c.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 787, + 237, + 901, + 319 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "number of sampled points, uniform sampling fails to prioritize geometrically salient regions, leading to the loss of fine details. To validate this observation, we experiment on a 3D mesh with intricate geometric details (e.g., keyboard buttons) shown in Figure 1 (a). We visualize the point cloud with different sampling strategies at various sampling densities in (b) and (c). As demonstrated in Figure 1 (b), even with increasing sampling rates, uniform sampling fails to preserve sharp features like keyboard buttons. This simple experiment confirms that uniform sampling fundamentally limits the capturing of geometric details, which in turn affects the VAE's reconstruction capability and the result quality of the learned diffusion models.", + "bbox": [ + 88, + 458, + 482, + 654 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Inspired by the success of importance sampling in various geometric processing tasks [14, 48], we introduce a similar strategy for 3D VAE training. While existing importance sampling methods focus on down-sampling point clouds, our task requires sampling points directly from mesh surfaces for shape VAE training [49, 56]. This fundamental difference necessitates a new sampling approach specifically designed for preserving geometric features from mesh representations.", + "bbox": [ + 88, + 656, + 482, + 792 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "To address this challenge, we propose a Sharp Edge Sampling (SES) algorithm that adaptively samples points based on geometric saliency. Specifically, SES first identifies edges with significant dihedral angles on the mesh, indicating regions of high geometric complexity. It then samples points along these salient regions while maintaining a balance with uniformly sampled points to capture the over", + "bbox": [ + 88, + 795, + 482, + 900 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "all structure. This approach ensures comprehensive coverage of both fine details and global geometry.", + "bbox": [ + 511, + 458, + 903, + 488 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Building upon SES, we present Dora-VAE, a novel method achieving high-fidelity reconstruction while maintaining compact latent representations. To fully leverage these detail-rich point clouds sampled by our SES, we design a dual cross-attention architecture that effectively processes both salient and uniform regions during encoding. As shown in Figure 1, our method significantly outperforms uniform sampling in preserving shape details on the keyboard (c), with consistent improvements in F-Score across different sampling rates (d).", + "bbox": [ + 511, + 489, + 905, + 641 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "The common evaluation protocol for 3D VAEs is also biased. It typically uses a set of randomly selected 3D shapes, and employs general metrics (e.g., F-score, Chamfer distance) to measure the shape reconstruction quality. However, we argue it is necessary to divide the test shapes into sub-classes of different shape complexity to better evaluate these 3D VAEs. To facilitate the evaluation of 3D VAEs, we further introduce Dora-Bench, a new benchmark based on existing datasets with our novel Sharp Normal Error (SNE) metric. Dora-Bench categorizes test shapes based on their geometric complexity and SNE focuses on measuring the reconstruction quality of salient geometric features. This combination enables a more rigorous assessment of 3D VAEs.", + "bbox": [ + 511, + 642, + 905, + 852 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Extensive experiments demonstrate that our Dora-VAE achieves superior results. When integrated into downstream 3D diffusion models, it significantly enhances the qual", + "bbox": [ + 511, + 854, + 905, + 900 + ], + "page_idx": 1 + }, + { + "type": "page_number", + "text": "16252", + "bbox": [ + 480, + 944, + 517, + 955 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "ity of generated 3D shapes, which validates that our novel sampling strategy and dual cross-attention architecture effectively preserve geometric details with compact latent spaces. To summarize, our main contributions include: 1) We propose Dora-VAE, a novel 3D VAE model for high-quality reconstruction with compact latent representations, accompanied by Dora-Bench, a comprehensive benchmark for evaluating 3D VAEs. 2) We introduce, for the first time, importance sampling to the task of 3D VAE learning and propose a Sharp Edge Sampling (SES) algorithm to prioritize geometrically salient regions. Building on SES, we design a dual cross-attention architecture to effectively encode these detail-rich point clouds. 3) We develop a systematic evaluation benchmark based on existing datasets with our novel Sharp Normal Error (SNE) metric that specifically assesses reconstruction accuracy of fine geometric details, enabling more rigorous evaluation than conventional random sampling approaches.", + "bbox": [ + 89, + 90, + 483, + 363 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "2. Related work", + "text_level": 1, + "bbox": [ + 89, + 376, + 228, + 392 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Importance Sampling in Point Clouds. Importance sampling techniques have been widely used in point cloud processing tasks [14, 48]. For instance, APES [48] proposes attention-based sampling for point cloud classification and segmentation. However, these methods operate directly on point clouds rather than meshes, making them less suitable for VAE-based shape representation where preserving complete geometric information is crucial.", + "bbox": [ + 89, + 402, + 483, + 521 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3D Shape VAEs. Recent 3D shape VAEs follow two main approaches: volume-based and vector set-based. Volume-based methods [36, 50] like XCube [36] use sparse convolution to encode voxelized surfaces, achieving high reconstruction quality but requiring large latent codes ( $>10,000$ tokens). While these methods excel at preserving geometric details, their large latent spaces pose significant challenges for downstream diffusion model training. In contrast, vector set-based approaches [24, 49, 54, 56, 58] encode uniformly sampled surface points using transformers, producing highly compact latent spaces that are particularly suitable for diffusion models. However, these methods often struggle with geometric detail preservation, especially in regions with complex surface features. Our analysis reveals that this limitation primarily stems from their uniform sampling strategy: when computational constraints restrict the total number of processable points, uniform sampling fails to prioritize geometrically significant regions, leading to insufficient capture of fine details. This information loss at the sampling stage fundamentally limits these methods' ability to learn and preserve intricate geometric features.", + "bbox": [ + 89, + 522, + 483, + 839 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3D Content Creation. Current 3D generation methods can be categorized into three groups. Optimization-based methods [6, 23, 26, 27, 30, 34, 35, 39, 41, 45], pioneered by DreamFusion [34] utilize score distillation sam", + "bbox": [ + 89, + 839, + 483, + 901 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "pling (SDS) to optimize 3D representations [19, 31, 38] using 2D diffusion model priors. While these methods can achieve photorealistic results, they suffer from slow generation speed, training instability, and often struggle to maintain geometric consistency. Large reconstruction models, like LRM [16] and follow-up works [22, 29, 40, 43, 46, 51] employ large-scale sparse-view reconstruction for efficient 3D generation. However, their lack of explicit geometric priors often leads to compromised geometric fidelity and inconsistent surface details. 3D native generative models [24, 49, 54, 56, 58], represented by 3DShape2VecSet [54], adopt a two-stage approach: first training a 3D VAE to encode shapes into latent space, then training a conditional latent diffusion model for generation. This approach ensures better geometric consistency through the VAE's built-in geometric constraints. However, the quality of generated shapes is fundamentally limited by the VAE's reconstruction capability. Recent works [24, 53] have shown that improving VAE reconstruction directly enhances downstream generation quality, which motivates our focus on advancing VAE design.", + "bbox": [ + 511, + 90, + 906, + 409 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3. Method", + "text_level": 1, + "bbox": [ + 513, + 422, + 604, + 438 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "In this section, we present Dora-VAE for high-quality 3D reconstruction, and Dora-Bench for 3D VAE evaluation. We first briefly review 3DShape2VecSet [54] in Section 3.1, the foundation of our method and then detail our key innovations in Section 3.2 and Section 3.3.", + "bbox": [ + 511, + 449, + 906, + 523 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3.1. Preliminary: 3DShape2VecSet", + "text_level": 1, + "bbox": [ + 511, + 535, + 784, + 551 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3DShape2VecSet [54] introduces a transformer-based 3D VAE that encodes uniformly sampled surface points into compact latent codes. Given a 3D surface S, their pipeline consists of three key steps:", + "bbox": [ + 511, + 556, + 905, + 618 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "- Surface Sampling: Uniformly sample $N_{d}$ points on the surface S using Poisson disk sampling [52] to obtain a dense point cloud $P_{d}$ , then downsample it to $N_{s}$ points via Farthest Point Sampling (FPS) [32] to get a sparse point cloud $P_{s}$ :", + "bbox": [ + 513, + 619, + 905, + 694 + ], + "page_idx": 2 + }, + { + "type": "equation", + "text": "\n$$\nP _ {d} = \\left\\{p _ {d} ^ {i} \\in S \\mid i = 1, \\dots , N _ {d} \\right\\}, P _ {s} = \\operatorname {F P S} \\left(P _ {d}, N _ {s}\\right). \\tag {1}\n$$\n", + "text_format": "latex", + "bbox": [ + 535, + 707, + 906, + 724 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "- Feature Encoding: Compute the point cloud feature $C$ via the cross-attention between $P_{s}$ and $P_{d}$ , followed by some self-attention layers to generate the latent code $z$ :", + "bbox": [ + 513, + 737, + 905, + 782 + ], + "page_idx": 2 + }, + { + "type": "equation", + "text": "\n$$\nC = \\operatorname {C r o s s A t t n} \\left(P _ {s}, P _ {d}, P _ {d}\\right), z = \\operatorname {S e l f A t t n} (C). \\tag {2}\n$$\n", + "text_format": "latex", + "bbox": [ + 550, + 795, + 906, + 811 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "- Geometry Decoding: Further decode $z$ through selfattention layers and predict occupancy values using randomly sampled spatial query points $Q_{space} \\in \\mathbb{R}^3$ :", + "bbox": [ + 511, + 824, + 905, + 871 + ], + "page_idx": 2 + }, + { + "type": "equation", + "text": "\n$$\n\\hat {O} = \\operatorname {C r o s s A t t n} \\left(Q _ {\\text {s p a c e}}, \\operatorname {S e l f A t t n} (z)\\right). \\tag {3}\n$$\n", + "text_format": "latex", + "bbox": [ + 589, + 883, + 906, + 902 + ], + "page_idx": 2 + }, + { + "type": "page_number", + "text": "16253", + "bbox": [ + 480, + 944, + 517, + 955 + ], + "page_idx": 2 + }, + { + "type": "image", + "img_path": "images/3086c3de39cf007f9521440ba59a4bf290529273b569fcbfb1d3a64390923a11.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 94, + 89, + 478, + 198 + ], + "page_idx": 3 + }, + { + "type": "image", + "img_path": "images/6260bff6b48d558afdc9f3e2d77308aca2788d2a8507c69e764d197bfc799384.jpg", + "image_caption": [ + "(a) Sharp edge sampling", + "Figure 2. Overview of Dora-VAE. (a) We utilize the proposed sharp edge sampling technique to extract both salient and uniform points from the input mesh. These points are then combined with dense points, effectively capturing both salient regions and smooth areas. (b) To enhance the encoding of point clouds sampled through sharp edge sampling, we design a dual cross-attention architecture." + ], + "image_footnote": [], + "bbox": [ + 91, + 213, + 480, + 339 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "While this method generates compact latent codes for diffusion, the uniform sampling limits its ability to capture fine geometric details. Our work addresses this limitation by carefully designed sampling and encoding strategies.", + "bbox": [ + 89, + 460, + 483, + 522 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "3.2. Dora-VAE", + "text_level": 1, + "bbox": [ + 89, + 532, + 207, + 546 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Figure 2 gives an overview of our pipeline. For each input mesh, we augment the uniformly sampled point cloud $P_{u}$ with more important points $P_{a}$ sampled by our proposed sharp edge sampling strategy, which forms the dense point cloud $P_{d}$ . During the encoding process, we compute the attention for $P_{u}$ and $P_{a}$ separately via a simple-yet-effective dual cross attention mechanism and sum the results for self-attention to compute the latent code $z$ . Our VAE training largely follows that of 3DShape2VecSet [54], which is supervised by a loss evaluated on the occupancy field.", + "bbox": [ + 89, + 554, + 482, + 705 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "3.2.1. Sharp Edge Sampling (SES)", + "text_level": 1, + "bbox": [ + 89, + 714, + 333, + 729 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "We propose SES to effectively sample points from geometrically salient regions. To ensure surface coverage, we also sample points uniformly. Our final sampled point cloud $P_{d}$ combines uniformly sampled points $P_{u}$ with points specifically sampled from salient regions $P_{a}$ as $P_{d} = P_{u} \\cup P_{a}$ . Our method computes salient points $P_{a}$ through two steps: detecting salient edges and sampling points from these regions.", + "bbox": [ + 89, + 734, + 482, + 853 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Salient Edges Detection. Given a triangular mesh, we identify a set of salient edges $\\Gamma$ by analyzing dihedral angles between adjacent faces, which calculates the angle be", + "bbox": [ + 89, + 854, + 483, + 900 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "tween the normal vectors of adjacent faces, providing a direct measure of surface curvature at mesh edges. For each edge $e$ shared by adjacent faces $f_{1}$ and $f_{2}$ , we compute the dihedral angle $\\theta_{e}$ as:", + "bbox": [ + 511, + 90, + 905, + 151 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\theta_ {e} = \\operatorname {a r c c o s} \\left(\\frac {\\mathbf {n} _ {f _ {1}} \\cdot \\mathbf {n} _ {f _ {2}}}{\\| \\mathbf {n} _ {f _ {1}} \\| \\| \\mathbf {n} _ {f _ {2}} \\|}\\right), \\tag {4}\n$$\n", + "text_format": "latex", + "bbox": [ + 606, + 157, + 906, + 191 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "where $\\mathbf{n}_{f_1}$ and $\\mathbf{n}_{f_2}$ are the normals of $f_{1}$ and $f_{2}$ . The salient edge set $\\Gamma$ contains all edges with a dihedral angle exceeding a predefined threshold $\\tau$ :", + "bbox": [ + 511, + 196, + 906, + 241 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\Gamma = \\left\\{e \\mid \\theta_ {e} > \\tau \\right\\} \\tag {5}\n$$\n", + "text_format": "latex", + "bbox": [ + 648, + 256, + 903, + 272 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Let $N_{\\Gamma} = |\\Gamma|$ represent the number of the salient edges.", + "bbox": [ + 511, + 276, + 880, + 290 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Salient Points Sampling. For each salient edge $e \\in \\Gamma$ , we collect its two vertices $v_{e,1}$ and $v_{e,2}$ into a salient vertex set $P_{\\Gamma}$ :", + "bbox": [ + 511, + 291, + 905, + 335 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\nP _ {\\Gamma} = \\left\\{v _ {e, 1}, v _ {e, 2} \\mid e \\in \\Gamma \\right\\}, \\tag {6}\n$$\n", + "text_format": "latex", + "bbox": [ + 620, + 335, + 903, + 352 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "where duplicate vertices from connecting edges are included only once. Let $N_V = |P_{\\Gamma}|$ denote the number of unique vertices in $P_{\\Gamma}$ .", + "bbox": [ + 511, + 356, + 905, + 400 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Given a target number of salient points $N_{\\mathrm{desired}}$ , we generate the salient point set $P_{a}$ based on the available salient vertices:", + "bbox": [ + 511, + 402, + 905, + 445 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\nP _ {a} = \\left\\{ \\begin{array}{l l} \\operatorname {F P S} \\left(P _ {\\Gamma}, N _ {\\text {d e s i r e d}}\\right), & \\text {i f} N _ {\\text {d e s i r e d}} < = N _ {V}, \\\\ P _ {\\Gamma} \\cup P _ {\\text {i n t e r p o l a t e d}}, & \\text {i f} 0 < N _ {V} < N _ {\\text {d e s i r e d}}, \\\\ \\emptyset , & \\text {i f} N _ {V} = 0. \\end{array} \\right. \\tag {7}\n$$\n", + "text_format": "latex", + "bbox": [ + 531, + 465, + 903, + 521 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "When we have excess salient vertices ( $N_{\\text{desired}} \\leq N_V$ ), we use FPS to downsample $P_{\\Gamma}$ to obtain $P_a$ . For cases with insufficient salient vertices ( $N_V < N_{\\text{desired}}$ ), we include all vertices from $P_{\\Gamma}$ and supplement with additional points $P_{\\text{interpolated}}$ . These additional points are generated by uniformly sampling $(N_{\\text{desired}} - N_V) / N_{\\Gamma}$ points along each salient edge in $\\Gamma$ , ensuring comprehensive coverage of salient features. When no salient edges are detected ( $N_V = 0$ ), $P_a$ remains empty.", + "bbox": [ + 511, + 527, + 905, + 664 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "3.2.2. Dual Cross Attention", + "text_level": 1, + "bbox": [ + 511, + 669, + 705, + 681 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Given the point clouds $P_{d}$ produced by our SES strategy, we design a dual cross-attention architecture to effectively encode both uniform and salient regions. Following 3DShape2VecSet [54], we first downsample $P_{u}$ and $P_{a}$ separately using FPS:", + "bbox": [ + 511, + 686, + 905, + 763 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\nP _ {s} = \\operatorname {F P S} \\left(P _ {u}, N _ {s, 1}\\right) \\cup \\operatorname {F P S} \\left(P _ {a}, N _ {s, 2}\\right), \\tag {8}\n$$\n", + "text_format": "latex", + "bbox": [ + 578, + 771, + 903, + 787 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "where $N_{s,1}$ and $N_{s,2}$ is the number of downsampled point clouds from $P_u$ and $P_a$ , respectively. We then compute cross-attention features separately for uniform and salient points as follows:", + "bbox": [ + 511, + 792, + 905, + 853 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\nC _ {u} = \\operatorname {C r o s s A t t n} \\left(P _ {s}, P _ {u}, P _ {u}\\right) \\tag {9}\n$$\n", + "text_format": "latex", + "bbox": [ + 611, + 862, + 903, + 878 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\nC _ {a} = \\operatorname {C r o s s A t t n} \\left(P _ {s}, P _ {a}, P _ {a}\\right) \\tag {10}\n$$\n", + "text_format": "latex", + "bbox": [ + 612, + 881, + 903, + 896 + ], + "page_idx": 3 + }, + { + "type": "page_number", + "text": "16254", + "bbox": [ + 480, + 944, + 517, + 955 + ], + "page_idx": 3 + }, + { + "type": "image", + "img_path": "images/762d6fdac16e892ef128d95b0fbfb7f3d7afc4834a9dd2ac14d2d5ea6fad7bdf.jpg", + "image_caption": [ + "(a) Counts of Different Datasets Across Detail Levels" + ], + "image_footnote": [], + "bbox": [ + 96, + 88, + 478, + 176 + ], + "page_idx": 4 + }, + { + "type": "image", + "img_path": "images/0e35f9650b8f748045dad41a39219dfa06a62f94199fe9252f06feb1b9d6333d.jpg", + "image_caption": [ + "(b) Total Counts by Detail Levels" + ], + "image_footnote": [], + "bbox": [ + 106, + 195, + 256, + 272 + ], + "page_idx": 4 + }, + { + "type": "image", + "img_path": "images/0426c7405dd740c90ea7c6ed7b3f5db4a0dca48fa80ee10cf5f8aa39af21ef5e.jpg", + "image_caption": [ + "(c) Visualization of detail levels", + "Figure 3. Our proposed benchmark include 3D shapes from the ABO [10], GSO [13], Meta [1], and Objaverse [12] datasets. (a) The histogram of different datasets across different shape complexities. (b) The pie chart of the total counts by shape complexities. (c) Sample shapes of different shape complexities." + ], + "image_footnote": [], + "bbox": [ + 267, + 186, + 468, + 277 + ], + "page_idx": 4 + }, + { + "type": "image", + "img_path": "images/9a18b1c24ae4dcfc95938048cfe55d2e56a7e6b0c465994029eeab5eccc3b6fe.jpg", + "image_caption": [ + "Figure 4. The process of computing sharp normal errors (SNE). We compute MSE loss in the sharp regions of the normal." + ], + "image_footnote": [], + "bbox": [ + 99, + 396, + 478, + 589 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "The final point cloud feature $C$ combines both attention results:", + "bbox": [ + 89, + 665, + 482, + 694 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\nC = C _ {u} + C _ {a}. \\tag {11}\n$$\n", + "text_format": "latex", + "bbox": [ + 233, + 707, + 482, + 723 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "This dual attention design enables separate focus on uniform and salient regions during feature extraction. Following 3DShape2VecSet [54], we use $C$ to predict the occupancy field $\\hat{O}$ through self-attention blocks. The whole model with parameters $\\psi$ are optimized using MSE loss:", + "bbox": [ + 89, + 742, + 483, + 818 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\nabla_ {\\psi} \\mathcal {L} _ {\\mathrm {M S E}} (\\hat {O}, O) = \\mathbb {E} \\left[ 2 (\\hat {O} - O) \\frac {\\partial \\hat {O}}{\\partial \\psi} \\right]. \\tag {12}\n$$\n", + "text_format": "latex", + "bbox": [ + 153, + 835, + 482, + 878 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "3.3. Dora-Bench", + "text_level": 1, + "bbox": [ + 513, + 90, + 643, + 104 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "3.3.1. Geometric Complexity-based Evaluation", + "text_level": 1, + "bbox": [ + 511, + 112, + 841, + 127 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "To enable a more rigorous evaluation of VAE performance, we propose Dora-bench, a benchmark that systematically categorizes test shapes based on their geometric complexity. Unlike previous methods that use randomly selected test sets, we measure shape complexity using the number of salient edges $N_{\\Gamma}$ (Section 3.2.1) and classify shapes into four levels:", + "bbox": [ + 511, + 131, + 903, + 234 + ], + "page_idx": 4 + }, + { + "type": "list", + "sub_type": "text", + "list_items": [ + "- Level 1 (Less Detail): $0 < N_{\\Gamma} \\leq 5000$ ;", + "- Level 2 (Moderate Detail): $5000 < N_{\\Gamma} \\leq 10000$ ;", + "- Level 3 (Rich Detail): $10000 < N_{\\Gamma} \\leq 50000$ ;", + "- Level 4 (Very Rich Detail): $N_{\\Gamma} > 50000$" + ], + "bbox": [ + 511, + 237, + 862, + 297 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "We curate test shapes from multiple public datasets including GSO [13], ABO [10], Meta [1], and Objaverse [12] to ensure diverse geometric complexities. Figure 3 shows the distribution of shapes across complexity levels (a,b) and example meshes from each level (c). Please refer to our supplementary materials for more examples.", + "bbox": [ + 511, + 297, + 903, + 388 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "3.3.2. Sharp Normal Error (SNE)", + "text_level": 1, + "bbox": [ + 511, + 393, + 751, + 407 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Building on our Dora-bench, we further introduce Sharp Normal Error (SNE) to evaluate reconstruction quality in salient regions. While existing metrics like Chamfer Distance and F-Score capture overall shape similarity, they fail to specifically assess the preservation of fine geometric details. SNE addresses this limitation by measuring normal map differences between reconstructed and ground truth shapes in geometrically significant areas. As illustrated in Figure 4, we render normal maps of the ground truth shape from multiple viewpoints and identify salient regions using Canny edge detection. These regions are dilated to create evaluation masks. The final SNE metric is computed as the Mean Squared Error between ground truth and reconstructed normal maps within the masked areas. This process enables focused evaluation of how well VAEs preserve sharp geometric features during reconstruction.", + "bbox": [ + 511, + 412, + 906, + 654 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "4. Experiments", + "text_level": 1, + "bbox": [ + 511, + 665, + 645, + 681 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "We conducted intensive experiments to validate the effectiveness of our proposed 3D VAE and compare it with other state-of-the-art methods.", + "bbox": [ + 511, + 690, + 905, + 734 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "4.1. Implementation Details", + "text_level": 1, + "bbox": [ + 511, + 742, + 730, + 758 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "We follow CLAY [56] for mesh preprocessing to ensure watertight 3D models. Our VAE is trained on a subset filtered from Objaverse [12], containing approximately 400,000 3D meshes. We filter out low-quality meshes with missing faces or severe self-intersections to ensure training stability. Our training is conducted on 32 A100 GPUs for two days using a batch size of 2048 and a learning rate of 5e-5. We employ Flash-Attention-v2 [11], mixed-precision training with FP16 and gradient checkpointing [8] to optimize", + "bbox": [ + 511, + 763, + 906, + 902 + ], + "page_idx": 4 + }, + { + "type": "page_number", + "text": "16255", + "bbox": [ + 480, + 944, + 517, + 955 + ], + "page_idx": 4 + }, + { + "type": "image", + "img_path": "images/c29adb258d92ddda61c839c850298bfd3c87a122d627da02089f38748083b368.jpg", + "image_caption": [ + "Figure 5. Qualitative comparison of the VAE reconstruction results. $\\dagger$ indicates the fine-tuning model that uses the same training data as ours." + ], + "image_footnote": [], + "bbox": [ + 91, + 87, + 903, + 854 + ], + "page_idx": 5 + }, + { + "type": "page_number", + "text": "16256", + "bbox": [ + 480, + 945, + 519, + 955 + ], + "page_idx": 5 + }, + { + "type": "table", + "img_path": "images/c61ef96b3fe20e54924317f44b2efa23caceccb877dce1ff3931f518e1af30ad.jpg", + "table_caption": [], + "table_footnote": [], + "table_body": "
MethodsLCL↑ F-score(0.01) × 100↑ F-score(0.005) × 100↓ CD × 10000↓ SNE × 100
L1L2L3L4L1L2L3L4L1L2L3L4L1L2L3L4
Xcube [36]>1000098.96898.79998.61598.22695.52593.87292.32285.3656.3156.2887.9359.9261.5791.4321.4301.679
Xcube† [36]>1000099.39399.79499.82499.07996.75395.53593.42287.3654.0154.1425.7407.6271.5431.4081.2591.639
Craftsman [24]25698.01695.87491.75681.73987.99482.54973.00057.3794.3899.12914.53033.4411.9061.8732.1913.933
Ours w/o DCA128099.96499.92599.67897.89096.56195.97591.61883.1242.2362.5064.4446.4321.4481.2151.2051.828
Ours w/o SES,DCA128099.94499.81497.29496.77995.97794.62388.40679.2402.4222.9833.9806.1961.4961.3131.3522.207
Ours full25699.50798.98696.66989.57793.27290.46682.38668.6693.3565.20210.27624.5271.5551.4101.6183.035
128099.98899.95599.88099.17097.03896.83193.45887.4732.0972.5003.9455.2651.4331.1861.1371.579
", + "bbox": [ + 107, + 88, + 893, + 243 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Table 1. Quantitative comparison in Dora-bench. ${}^{ \\dagger }$ indicates the fine-tuning model that uses the same training data as ours.", + "bbox": [ + 135, + 247, + 859, + 261 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "memory usage and training efficiency. For the parameters in sharp edge sampling, we set $N_{\\mathrm{desired}} = 16384$ and $\\tau = 30$ . We set the low threshold to 20 and the high threshold to 200 for the canny edge detection.", + "bbox": [ + 89, + 287, + 482, + 349 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "4.2. Evaluation Setting", + "text_level": 1, + "bbox": [ + 89, + 359, + 272, + 376 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Metrics. We evaluate reconstruction quality by comparing input meshes with their decoded counterparts from different 3D VAEs using 1M sampled points under three metrics: 1) F-score $(r)$ [20], which reconstruction accuracy by computing precision and recall of point correspondences within distance threshold $r$ . Specifically, we report F-score (0.01) and F-score (0.005) with shapes normalized $[-1,1]$ . 2) Chamfer Distance (CD), which computes the average distance between each reconstructed point and its nearest ground truth point. 3) Sharp Normal Errors (SNE) as proposed in Section 3.3.2, which evaluate normal map differences in salient regions. For fair comparison, we also report Latent Code Length (LCL) as longer codes typically enable better reconstruction.", + "bbox": [ + 89, + 382, + 483, + 593 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Baselines. We compare Dora-VAE with state-of-the-art approaches, including: 1) XCube-VAE [36], a volumetric method with larger latent codes; 2) XCube-VAE† [36], our fine-tuned version of the original XCube-VAE on the same dataset; 3) Craftsman-VAE[24], which fine-tune the 3DShape2VecSet [54] with shorter latent codes on Objverse. We exclude VAE models from Direct3D [49] and CLAY[56] as as their implementations were not publicly available at submission time.", + "bbox": [ + 89, + 594, + 483, + 729 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "4.3. Qualitative Comparison", + "text_level": 1, + "bbox": [ + 89, + 741, + 312, + 758 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Figure 5 shows visual comparisons of different methods across different complexity levels from our Dora-bench dataset. We visualize both ground truth and reconstructed meshes using surface normal coloring to highlight geometric details. For shapes with lower complexity (L1 and L2), all methods achieve comparable reconstruction quality. However, when dealing with shapes of higher complexity (L3 and L4), the advantages of our method become obvious. While XCube-VAE achieves similar visual quality", + "bbox": [ + 89, + 763, + 482, + 902 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "to ours, it requires a significantly larger latent space - more than $8 \\times$ the size of ours ( $>10,000$ dimensions vs. 1,280). This substantial reduction in latent code length, while maintaining high reconstruction fidelity, makes our method particularly suitable for training 3D diffusion models. In contrast, Craftsman-VAE shows a noticeable degradation in reconstruction quality for complex shapes, failing to capture fine geometric details. Additional visual comparisons are provided in the appendix.", + "bbox": [ + 511, + 287, + 906, + 425 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "4.4. Quantitative Comparison", + "text_level": 1, + "bbox": [ + 511, + 435, + 746, + 450 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Table 1 presents quantitative results of different methods across different complexity levels of Dora-bench. Our method consistently outperforms baselines across all levels, with larger margins on more complex shapes (L3 and L4). The advantage is particularly evident in CD metrics, where our method with only 256 latent codes surpasses even our fine-tuned version of XCube-VAE (3.356 vs. 4.015). When using 1280 latent codes, our method further decreases CD to 2.097, achieving a $47.77\\%$ improvement over XCube-VAE†. We attribute XCube-VAE's lower performance to its use of NKSR [18] for mesh extraction, which introduces additional quantization errors.", + "bbox": [ + 511, + 457, + 906, + 638 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Notably, our method demonstrates superior performance in preserving geometric details, as reflected by the SNE metric. For example, in L4 shapes where geometric complexity is highest, our method achieves an SNE of 1.579 compared to 1.639 from XCube-VAE†, representing a $3.7\\%$ improvement. This significant gain in SNE aligns with our qualitative observations in Figure 5, where our method better preserves fine details such as sharp edges and complex surface variations, demonstrating the effectiveness of our sharp edge sampling strategy.", + "bbox": [ + 511, + 638, + 908, + 791 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "4.5. Ablation Studies", + "text_level": 1, + "bbox": [ + 511, + 801, + 679, + 816 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "To evaluate the contribution of each component, we compare our full model with two variants under the same training conditions:", + "bbox": [ + 511, + 824, + 903, + 868 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "- Ours w/o SES, DCA. This variant removes both sharp edge sampling (SES) and dual cross attention (DCA), i.e.", + "bbox": [ + 511, + 869, + 906, + 901 + ], + "page_idx": 6 + }, + { + "type": "page_number", + "text": "16257", + "bbox": [ + 480, + 944, + 519, + 957 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/a725a621e317993e35a9a1c648a7adc36716ce5021113677bdd3e1b0e6ac6125.jpg", + "image_caption": [ + "GT" + ], + "image_footnote": [], + "bbox": [ + 127, + 107, + 256, + 219 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/5fe5ec243e59f522a4a135eaa1a908510eb65606e4581824486aa11e830c8ea2.jpg", + "image_caption": [ + "Full", + "w/o DAC" + ], + "image_footnote": [], + "bbox": [ + 307, + 107, + 433, + 219 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/45c477727be1322becd3da2860816f627377b230cfe97a126222cb7586a5f408.jpg", + "image_caption": [ + "w/o SES, DCA", + "Figure 6. Ablation studies of our method. Given the ground truth of mesh, we employ both our full model and its variations to reconstruct the ground truth mesh, highlighting significant reconstruction discrepancies with red boxes." + ], + "image_footnote": [], + "bbox": [ + 130, + 242, + 256, + 354 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/92f385a5f1b1973336a224a2a8c63ee85ae383a0f4a3a91f2532a6ec85601df8.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 307, + 241, + 434, + 354 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "using only uniformly sampled point clouds with Poisson disk sampling [52], while maintaining an equal $N_{d}$ .", + "bbox": [ + 102, + 453, + 482, + 483 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "- Ours w/o DCA. This variant retains SES but removes DCA, i.e. using a single cross attention adopted by [54]. As shown in Figure 6 and Table 1, our full model consistently outperforms these variants, validating the effectiveness of both components.", + "bbox": [ + 89, + 483, + 482, + 559 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "5. Application: Single Image to 3D", + "text_level": 1, + "bbox": [ + 89, + 571, + 383, + 590 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "We demonstrate the effectiveness of our VAE by applying it to single-image 3D generation through diffusion models. Following CLAY [56], we implement a latent diffusion model based on the DiT [33] architecture. For a fair comparison, we fine-tune Craftsman-VAE on our dataset (denoted as Craftsman-VAE†) since both Craftsman-VAE and 3DShape2VecSet were originally trained on smaller datasets. Note that XCube-VAE is excluded from comparison due to its 10,000-dimensional latent codes being impractical for diffusion model training. Figure 7 shows some generation results from diffusion models trained with our Dora-VAE and Craftsman-VAE†. Both models share identical architecture (0.39B parameters) and training conditions (same dataset, 32 A100 GPUs, 3 days). Our Dora-VAE demonstrates significantly better preservation of geometric details in the generated shapes, validating its effectiveness as a foundation for 3D generation tasks. While further improvements could be achieved with more extensive training data and computational resources, we focus on validating the VAE's capabilities in this work and leave such exten", + "bbox": [ + 89, + 598, + 482, + 900 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/d526132c9f29f10e84fa0ffa2c3e54ffed14ca3f54d575be6896746582f43c94.jpg", + "image_caption": [ + "Input Image" + ], + "image_footnote": [], + "bbox": [ + 526, + 119, + 635, + 220 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/a481d437a55e51a85aa0457a52c71c839638cd48e407d959e40364ce87d3b6ce.jpg", + "image_caption": [ + "Ours" + ], + "image_footnote": [], + "bbox": [ + 653, + 117, + 767, + 220 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/670442b42bbf9d85c9ecec5f31388d75e27fae45204cd0d6eab806a79ee4b776.jpg", + "image_caption": [ + "Craftsman" + ], + "image_footnote": [], + "bbox": [ + 774, + 117, + 901, + 220 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/d17e93218df0ade95f0621be106e9dc8d1e6304447621c1f07fe0389f4deb311.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 522, + 229, + 627, + 333 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/e15693c532de6a739e4452ddd77ea78328d3dc7ad71ea47c61cdc43a4948231a.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 651, + 229, + 758, + 329 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/09f313a12a61443cfb6b32a96dbaf984a6564a0029c8c124a1ef86939b10d737.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 782, + 228, + 890, + 333 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/d1d0e13606feabfc15e3f1121b3ded86c314dba301781fed07a821d9a34517a2.jpg", + "image_caption": [ + "Figure 7. The diffusion results of the single image to 3D generation trained on our Dora-VAE and Craftsman†. The 3D geometry generated by the diffusion model trained on our proposed Dora-VAE has more details under the same experimental environment." + ], + "image_footnote": [], + "bbox": [ + 524, + 340, + 627, + 446 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/ea2c630b92887b7c8376bc23c0c84accb0065c03c74bbd2df2e63347c830857a.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 651, + 338, + 758, + 445 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/f85fb20808f7c52a8c63c90913f1a53d230dcbbd22d4180cf275b7ce2090701a.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 779, + 340, + 883, + 445 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "sions for future work.", + "bbox": [ + 513, + 527, + 660, + 541 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "6. Conclusion", + "text_level": 1, + "bbox": [ + 511, + 556, + 633, + 571 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "In this work, we introduce Dora-VAE, a novel VAE designed for high-quality 3D shape compression and reconstruction. At its core, Dora-VAE introduces sharp edge sampling to effectively capture salient geometric features, complemented by a dual cross-attention architecture that enhances the encoding of these detail-rich point clouds. To enable more rigorous evaluation of VAE performance, we develop Dora-bench, which systematically categorizes shapes based on geometric complexity and introduces the Sharp Normal Error (SNE) metric for specifically assessing the preservation of fine geometric details. Our comprehensive experiments demonstrate that Dora-VAE significantly outperforms existing methods across varying levels of shape complexity. Furthermore, we show that the improved reconstruction capability of Dora-VAE directly enhances the quality of downstream tasks by applying it to single-image 3D generation. The superior performance in generating geometric details validates our approach of focusing on salient region sampling and encoding for 3D VAE design.", + "bbox": [ + 511, + 580, + 906, + 868 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Acknowledgements. This work is partially supported by the project L0751 between Bytedance and HKUST.", + "bbox": [ + 511, + 868, + 905, + 898 + ], + "page_idx": 7 + }, + { + "type": "page_number", + "text": "16258", + "bbox": [ + 480, + 944, + 517, + 955 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "References", + "text_level": 1, + "bbox": [ + 91, + 89, + 187, + 104 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[1] Digital twin catalog. META, 2024. https://www.projectaria.com/datasets/dtc/.5", + "[2] Yukang Cao, Yan-Pei Cao, Kai Han, Ying Shan, and KwanYee K. Wong. Dreamavatar: Text-and-shape guided 3d human avatar generation via diffusion models. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 958-968, 2024. 1", + "[3] Yukang Cao, Liang Pan, Kai Han, Kwan-Yee K Wong, and Ziwei Liu. Avatargo: Zero-shot 4d human-object interaction generation and animation. arXiv preprint arXiv:2410.07164, 2024. 1", + "[4] Junsong Chen, Jincheng Yu, Chongjian Ge, Lewei Yao, Enze Xie, Yue Wu, Zhongdao Wang, James Kwok, Ping Luo, Huchuan Lu, and Zhenguo Li. Pixart-α: Fast training of diffusion transformer for photorealistic text-to-image synthesis, 2023. 1", + "[5] Junsong Chen, Yue Wu, Simian Luo, Enze Xie, Sayak Paul, Ping Luo, Hang Zhao, and Zhenguo Li. Pixart-δ: Fast and controllable image generation with latent consistency models, 2024. 1", + "[6] Rui Chen, Yongwei Chen, Ningxin Jiao, and Kui Jia. Fantasia3d: Disentangling geometry and appearance for high-quality text-to-3d content creation. In Proceedings of the IEEE/CVF International Conference on Computer Vision (ICCV), 2023. 3", + "[7] Rui Chen, Mingyi Shi, Shaoli Huang, Ping Tan, Taku Komura, and Xuelin Chen. Taming diffusion probabilistic models for character control. In ACM SIGGRAPH 2024 Conference Papers, New York, NY, USA, 2024. Association for Computing Machinery. 1", + "[8] Tianqi Chen, Bing Xu, Chiyuan Zhang, and Carlos Guestrin. Training deep nets with sublinear memory cost. CoRR, abs/1604.06174, 2016. 5", + "[9] Yongwei Chen, Rui Chen, Jiabao Lei, Yabin Zhang, and Kui Jia. Tango: Text-driven photorealistic and robust 3d stylization via lighting decomposition. Advances in Neural Information Processing Systems, 35:30923-30936, 2022. 1", + "[10] Jasmine Collins, Shubham Goel, Kenan Deng, Achleshwar Luthra, Leon Xu, Erhan Gundogdu, Xi Zhang, Tomas F Yago Vicente, Thomas Dideriksen, Himanshu Arora, et al. Abo: Dataset and benchmarks for real-world 3d object understanding. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 21126-21136, 2022. 5", + "[11] Tri Dao. Flashattention-2: Faster attention with better parallelism and work partitioning. In The Twelfth International Conference on Learning Representations, ICLR 2024, Vienna, Austria, May 7-11, 2024. OpenReview.net, 2024. 5", + "[12] Matt Deitke, Dustin Schwenk, Jordi Salvador, Luca Weihs, Oscar Michel, Eli VanderBilt, Ludwig Schmidt, Kiana Ehsani, Aniruddha Kembhavi, and Ali Farhadi. Objverse: A universe of annotated 3d objects. arXiv preprint arXiv:2212.08051, 2022. 5", + "[13] Laura Downs, Anthony Francis, Nate Koenig, Brandon Kinman, Ryan Hickman, Krista Reymann, Thomas B McHugh," + ], + "bbox": [ + 93, + 114, + 482, + 900 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "and Vincent Vanhoucke. Google scanned objects: A high-quality dataset of 3d scanned household items. In 2022 International Conference on Robotics and Automation (ICRA), pages 2553-2560. IEEE, 2022. 5", + "[14] Fabian Groh, Patrick Wieschollek, and Hendrik P. A. Lensch. Flex-convolution (million-scale point-cloud learning beyond grid-worlds). In *Asian Conference on Computer Vision (ACCV)*, 2018. 2, 3", + "[15] Jonathan Ho, Ajay Jain, and Pieter Abbeel. Denoising diffusion probabilistic models. arXiv preprint arxiv:2006.11239, 2020. 1", + "[16] Yicong Hong, Kai Zhang, Juixiang Gu, Sai Bi, Yang Zhou, Difan Liu, Feng Liu, Kalyan Sunkavalli, Trung Bui, and Hao Tan. LRM: large reconstruction model for single image to 3d. In The Twelfth International Conference on Learning Representations, ICLR 2024, Vienna, Austria, May 7-11, 2024. OpenReview.net, 2024. 3", + "[17] Yicong Hong, Kai Zhang, Juixiang Gu, Sai Bi, Yang Zhou, Difan Liu, Feng Liu, Kalyan Sunkavalli, Trung Bui, and Hao Tan. Lrm: Large reconstruction model for single image to 3d, 2024. 1", + "[18] Jiahui Huang, Zan Gojcic, Matan Atzmon, Or Litany, Sanja Fidler, and Francis Williams. Neural kernel surface reconstruction. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 4369-4379, 2023. 7", + "[19] Bernhard Kerbl, Georgios Kopanas, Thomas Leimkuhler, and George Drettakis. 3d gaussian splatting for real-time radiance field rendering. ACM Trans. Graph., 42(4):139-1, 2023. 3", + "[20] Arno Knapitsch, Jaesik Park, Qian-Yi Zhou, and Vladlen Koltun. Tanks and temples: benchmarking large-scale scene reconstruction. ACM Trans. Graph., 36(4), 2017. 7", + "[21] Yushi Lan, Fangzhou Hong, Shuai Yang, Shangchen Zhou, Xuyi Meng, Bo Dai, Xingang Pan, and Chen Change Loy. Ln3diff: Scalable latent neural fields diffusion for speedy 3d generation. In ECCV, 2024. 1", + "[22] Jiahao Li, Hao Tan, Kai Zhang, Zexiang Xu, Fujun Luan, Yinghao Xu, Yicong Hong, Kalyan Sunkavalli, Greg Shakhnarovich, and Sai Bi. Instant3d: Fast text-to-3d with sparse-view generation and large reconstruction model. arXiv preprint arXiv:2311.06214, 2023. 3", + "[23] Weiyu Li, Rui Chen, Xuelin Chen, and Ping Tan. Sweetdreamer: Aligning geometric priors in 2d diffusion for consistent text-to-3d. International Conference on Learning Representations (ICLR), 2024. 3", + "[24] Weiyu Li, Jiarui Liu, Rui Chen, Yixun Liang, Xuelin Chen, Ping Tan, and Xiaoxiao Long. Craftsman: High-fidelity mesh generation with 3d native generation and interactive geometry refiner. arXiv preprint arXiv:2405.14979, 2024. 1, 3, 7", + "[25] Yixun Liang, Xin Yang, Jiantao Lin, Haodong Li, Xiaogang Xu, and Yingcong Chen. Luciddreamer: Towards high-fidelity text-to-3d generation via interval score matching. arXiv preprint arXiv:2311.11284, 2023. 1", + "[26] Chen-Hsuan Lin, Jun Gao, Luming Tang, Towaki Takikawa, Xiaohui Zeng, Xun Huang, Karsten Kreis, Sanja Fidler," + ], + "bbox": [ + 517, + 92, + 903, + 900 + ], + "page_idx": 8 + }, + { + "type": "page_number", + "text": "16259", + "bbox": [ + 480, + 944, + 517, + 955 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "Ming-Yu Liu, and Tsung-Yi Lin. Magic3d: High-resolution text-to-3d content creation. In IEEE Conference on Computer Vision and Pattern Recognition (CVPR), 2023. 3", + "[27] Fangfu Liu, Diankun Wu, Yi Wei, Yongming Rao, and Yueqi Duan. Sherpa3d: Boosting high-fidelity text-to-3d generation via coarse 3d prior, 2023. 3", + "[28] Minghua Liu, Chao Xu, Haian Jin, Linghao Chen, Mukund Varma T, Zexiang Xu, and Hao Su. One-2-3-45: Any single image to 3d mesh in 45 seconds without per-shape optimization. 2024. 1", + "[29] Minghua Liu, Chong Zeng, Xinyue Wei, Ruoxi Shi, Linghao Chen, Chao Xu, Mengqi Zhang, Zhaoning Wang, Xiaoshuai Zhang, Isabella Liu, Hongzhi Wu, and Hao Su. Meshformer: High-quality mesh generation with 3d-guided reconstruction model. arXiv preprint arXiv:2408.10198, 2024. 3", + "[30] Yuan Liu, Cheng Lin, Zijiao Zeng, Xiaoxiao Long, Lingjie Liu, Taku Komura, and Wenping Wang. Syncdreamer: Generating multiview-consistent images from a single-view image. arXiv preprint arXiv:2309.03453, 2023. 1, 3", + "[31] Ben Mildenhall, Pratul P. Srinivasan, Matthew Tancik, Jonathan T. Barron, Ravi Ramamoorthi, and Ren Ng. Nerf: Representing scenes as neural radiance fields for view synthesis. In European Conference on Computer Vision (ECCV), 2020. 3", + "[32] Carsten Moenning and Neil A Dodgson. Fast marching farthest point sampling. Technical report, University of Cambridge, Computer Laboratory, 2003. 3", + "[33] William Peebles and Saining Xie. Scalable diffusion models with transformers. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 4195-4205, 2023. 8", + "[34] Ben Poole, Ajay Jain, Jonathan T. Barron, and Ben Mildenhall. Dreamfusion: Text-to-3d using 2d diffusion. arXiv, 2022. 3", + "[35] Lingteng Qiu, Guanying Chen, Xiaodong Gu, Qi Zuo, Mutian Xu, Yushuang Wu, Weihao Yuan, Zilong Dong, Liefeng Bo, and Xiaoguang Han. Richdreamer: A generalizable normal-depth diffusion model for detail richness in text-to-3d. In IEEE/CVF Conference on Computer Vision and Pattern Recognition, CVPR 2024, Seattle, WA, USA, June 16-22, 2024, pages 9914-9925. IEEE, 2024. 3", + "[36] Xuanchi Ren, Jiahui Huang, Xiaohui Zeng, Ken Museth, Sanja Fidler, and Francis Williams. Xcube: Large-scale 3d generative modeling using sparse voxel hierarchies. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 4209-4219, 2024. 1, 3, 7", + "[37] Robin Rombach, Andreas Blattmann, Dominik Lorenz, Patrick Esser, and Björn Ommer. High-resolution image synthesis with latent diffusion models. In Conference on Computer Vision and Pattern Recognition (CVPR), pages 10684-10695, 2022. 1", + "[38] Tianchang Shen, Jun Gao, Kangxue Yin, Ming-Yu Liu, and Sanja Fidler. Deep marching tetrahedra: a hybrid representation for high-resolution 3d shape synthesis. Advances in Neural Information Processing Systems, 34:6087-6101, 2021. 3" + ], + "bbox": [ + 91, + 90, + 482, + 898 + ], + "page_idx": 9 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[39] Yichun Shi, Peng Wang, Jianglong Ye, Long Mai, Kejie Li, and Xiao Yang. Mvdream: Multi-view diffusion for 3d generation. arXiv:2308.16512, 2023. 3", + "[40] Jiaxiang Tang, Zhaoxi Chen, Xiaokang Chen, Tengfei Wang, Gang Zeng, and Ziwei Liu. Lgm: Large multi-view gaussian model for high-resolution 3d content creation. arXiv preprint arXiv:2402.05054, 2024. 3", + "[41] Jiaxiang Tang, Jiawei Ren, Hang Zhou, Ziwei Liu, and Gang Zeng. Dreamgaussian: Generative gaussian splatting for efficient 3d content creation. In The Twelfth International Conference on Learning Representations, ICLR 2024, Vienna, Austria, May 7-11, 2024. OpenReview.net, 2024. 3", + "[42] Dmitry Tochilkin, David Pankratz, Zexiang Liu, Zixuan Huang, , Adam Letts, Yangguang Li, Ding Liang, Christian Laforte, Varun Jampani, and Yan-Pei Cao. Triposr: Fast 3d object reconstruction from a single image. arXiv preprint arXiv:2403.02151, 2024. 1", + "[43] Peng Wang, Hao Tan, Sai Bi, Yinghao Xu, Fujun Luan, Kalyan Sunkavalli, Wenping Wang, Zexiang Xu, and Kai Zhang. Pf-lrm: Pose-free large reconstruction model for joint pose and shape prediction. arXiv preprint arXiv:2311.12024, 2023. 3", + "[44] Tengfei Wang, Bo Zhang, Ting Zhang, Shuyang Gu, Jianmin Bao, Tadas Baltrusaitis, Jingjing Shen, Dong Chen, Fang Wen, Qifeng Chen, and Baining Guo. Rodin: A generative model for sculpting 3d digital avatars using diffusion, 2022. 1", + "[45] Zhengyi Wang, Cheng Lu, Yikai Wang, Fan Bao, Chongxuan Li, Hang Su, and Jun Zhu. Prolificdreamer: High-fidelity and diverse text-to-3d generation with variational score distillation. arXiv preprint arXiv:2305.16213, 2023. 3", + "[46] Zhengyi Wang, Yikai Wang, Yifei Chen, Chengdong Xiang, Shuo Chen, Dajiang Yu, Chongxuan Li, Hang Su, and Jun Zhu. Crm: Single image to 3d textured mesh with convolutional reconstruction model. arXiv preprint arXiv:2403.05034, 2024. 3", + "[47] Francis Williams, Jiahui Huang, Jonathan Swartz, Gergely Klar, Vijay Thakkar, Matthew Cong, Xuanchi Ren, Rui-long Li, Clement Fuji-Tsang, Sanja Fidler, Eftychios Sifakis, and Ken Museth. fvdb: A deep-learning framework for sparse, large-scale, and high-performance spatial intelligence. ACM Transactions on Graphics (TOG), 43(4):133:1-133:15, 2024. 1", + "[48] Chengzhi Wu, Junwei Zheng, Julius Pfrommer, and Jürgen Beyerer. Attention-based point cloud edge sampling. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), 2023. 2, 3", + "[49] Shuang Wu, Youtian Lin, Feihu Zhang, Yifei Zeng, Jingxi Xu, Philip Torr, Xun Cao, and Yao Yao. Direct3d: Scalable image-to-3d generation via 3d latent diffusion transformer. arXiv:2405.14832, 2024. 1, 2, 3, 7", + "[50] Bojun Xiong, Si-Tong Wei, Xin-Yang Zheng, Yan-Pei Cao, Zhouhui Lian, and Peng-Shuai Wang. OctFusion: Octree-based diffusion models for 3d shape generation. arXiv, 2024. 3", + "[51] Jiale Xu, Weihao Cheng, Yiming Gao, Xintao Wang, Shenghua Gao, and Ying Shan. Instantmesh: Efficient 3d" + ], + "bbox": [ + 516, + 90, + 903, + 900 + ], + "page_idx": 9 + }, + { + "type": "page_number", + "text": "16260", + "bbox": [ + 480, + 944, + 519, + 955 + ], + "page_idx": 9 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "mesh generation from a single image with sparse-view large reconstruction models. arXiv preprint arXiv:2404.07191, 2024.3", + "[52] Cem Yuksel. Sample elimination for generating poisson disk sample sets. The Eurographs Association & John Wiley & Sons, Ltd., 2015. 3, 8", + "[53] Biao Zhang and Peter Wonka. Lagem: A large geometry model for 3d representation learning and diffusion. arXiv preprint arXiv:2410.01295, 2024. 3", + "[54] Biao Zhang, Jiapeng Tang, Matthias Nießner, and Peter Wonka. 3dshape2vecset: A 3d shape representation for neural fields and generative diffusion models. ACM Transactions on Graphics (SIGGRAPH), 42(4), 2023. 1, 3, 4, 5, 7, 8", + "[55] Lvmin Zhang, Anyi Rao, and Maneesh Agrawala. Adding conditional control to text-to-image diffusion models, 2023. 1", + "[56] Longwen Zhang, Ziyu Wang, Qixuan Zhang, Qiwei Qiu, Anqi Pang, Haoran Jiang, Wei Yang, Lan Xu, and Jingyi Yu. Clay: A controllable large-scale generative model for creating high-quality 3d assets. ACM Transactions on Graphics (TOG), 43(4):1-20, 2024. 1, 2, 3, 5, 7, 8", + "[57] Xuying Zhang, Bo-Wen Yin, Yuming Chen, Zheng Lin, Yunheng Li, Qibin Hou, and Ming-Ming Cheng. Temo: Towards text-driven 3d stylization for multi-object meshes. arXiv preprint arXiv:2312.04248, 2023. 1", + "[58] Zibo Zhao, Wen Liu, Xin Chen, Xianfang Zeng, Rui Wang, Pei Cheng, BIN FU, Tao Chen, Gang YU, and Shenghua Gao. Michelangelo: Conditional 3d shape generation based on shape-image-text aligned latent representation. In Advances in Neural Information Processing Systems (NeurIPS), 2023. 1, 3" + ], + "bbox": [ + 91, + 90, + 482, + 527 + ], + "page_idx": 10 + }, + { + "type": "page_number", + "text": "16261", + "bbox": [ + 480, + 945, + 517, + 955 + ], + "page_idx": 10 + } +] \ No newline at end of file diff --git a/2025/Dora_ Sampling and Benchmarking for 3D Shape Variational Auto-Encoders/6b478709-2cef-4587-8775-92fc50ace470_model.json b/2025/Dora_ Sampling and Benchmarking for 3D Shape Variational Auto-Encoders/6b478709-2cef-4587-8775-92fc50ace470_model.json new file mode 100644 index 0000000000000000000000000000000000000000..053edfc68f0937a00f9540ce481f457d378d958d --- /dev/null +++ b/2025/Dora_ Sampling and Benchmarking for 3D Shape Variational Auto-Encoders/6b478709-2cef-4587-8775-92fc50ace470_model.json @@ -0,0 +1,2653 @@ +[ + [ + { + "type": "header", + "bbox": [ + 0.107, + 0.003, + 0.182, + 0.043 + ], + "angle": 0, + "content": "CVF" + }, + { + "type": "header", + "bbox": [ + 0.237, + 0.001, + 0.812, + 0.047 + ], + "angle": 0, + "content": "This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore." + }, + { + "type": "title", + "bbox": [ + 0.111, + 0.131, + 0.887, + 0.153 + ], + "angle": 0, + "content": "Dora: Sampling and Benchmarking for 3D Shape Variational Auto-Encoders" + }, + { + "type": "text", + "bbox": [ + 0.193, + 0.179, + 0.805, + 0.288 + ], + "angle": 0, + "content": "Rui Chen\\(^{1,2}\\) Jianfeng Zhang\\(^{2\\dagger}\\) Yixun Liang\\(^{1,3}\\) Guan Luo\\(^{2,4}\\) Weiyu Li\\(^{1,3}\\) \nJiarui Liu\\(^{1,3}\\) Xiu Li\\(^{2}\\) Xiaoxiao Long\\(^{1,3}\\) Jiashi Feng\\(^{2}\\) Ping Tan\\(^{1,3\\dagger}\\) \n†Corresponding authors \n\\(^{1}\\)The Hong Kong University of Science and Technology \\(^{2}\\)ByteDance Seed \n\\(^{3}\\)LightIllusions \\(^{4}\\)Tsinghua University \nhttps://aruichen.github.io/Dora" + }, + { + "type": "title", + "bbox": [ + 0.248, + 0.322, + 0.327, + 0.339 + ], + "angle": 0, + "content": "Abstract" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.358, + 0.486, + 0.751 + ], + "angle": 0, + "content": "Recent 3D content generation pipelines commonly employ Variational Autoencoders (VAEs) to encode shapes into compact latent representations for diffusion-based generation. However, the widely adopted uniform point sampling strategy in Shape VAE training often leads to a significant loss of geometric details, limiting the quality of shape reconstruction and downstream generation tasks. We present Dora-VAE, a novel approach that enhances VAE reconstruction through our proposed sharp edge sampling strategy and a dual cross-attention mechanism. By identifying and prioritizing regions with high geometric complexity during training, our method significantly improves the preservation of fine-grained shape features. Such sampling strategy and the dual attention mechanism enable the VAE to focus on crucial geometric details that are typically missed by uniform sampling approaches. To systematically evaluate VAE reconstruction quality, we additionally propose Dora-bench, a benchmark that quantifies shape complexity through the density of sharp edges, introducing a new metric focused on reconstruction accuracy at these salient geometric features. Extensive experiments on the Dora-bench demonstrate that Dora-VAE achieves comparable reconstruction quality to the state-of-the-art dense XCube-VAE while requiring a latent space at least \\(8 \\times\\) smaller (1,280 vs. \\(>10,000\\) codes). Project page: https://aruichen.github.io/Dora." + }, + { + "type": "title", + "bbox": [ + 0.092, + 0.784, + 0.222, + 0.799 + ], + "angle": 0, + "content": "1. Introduction" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.81, + 0.485, + 0.902 + ], + "angle": 0, + "content": "3D content creation is vital to delivering realistic and immersive experiences in various industries, including games, movies, and AR/VR. However, traditional 3D modeling typically demands significant expertise and manual effort, making it time-consuming and challenging, especially for non-expert users. Recent advances in AI-powered 3D con" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.324, + 0.905, + 0.368 + ], + "angle": 0, + "content": "tent generation methods [2, 3, 7, 9, 17, 21, 25, 28, 30, 42, 44, 57] have transformed the field, making it more accessible to many more users." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.37, + 0.907, + 0.49 + ], + "angle": 0, + "content": "Following the success of text-to-image generation models [4, 5, 15, 37, 55], recent 3D content creation approaches [36, 49, 56] adopt a two-stage pipeline: encoding 3D shapes into a latent space using variational autoencoders (VAEs), followed by training a latent diffusion model. The performance of such a generative pipeline heavily relies on the VAE's capability to faithfully encode and reconstruct 3D shapes." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.491, + 0.906, + 0.598 + ], + "angle": 0, + "content": "Existing 3D VAEs operate by sampling points on mesh surfaces for shape encoding and then reconstructing the original 3D meshes by its decoder. This process faces unique challenges compared to 2D image VAEs where the input image is fully observable. In comparison, the sampled point cloud often cannot capture all the necessary shape information, which could harm the performance of 3D VAEs." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.599, + 0.907, + 0.825 + ], + "angle": 0, + "content": "Volume-based method [36] leverages sparse convolution [47] to process millions of voxelized points for high-fidelity reconstruction. Its dense sampling captures precise shape information. However, this method produces large latent codes (commonly \\(>10,000\\) tokens), which significantly complicate the training of diffusion models. On the other hand, vector-set (Vecset) methods [24, 54, 56, 58] use transformers to achieve compact latent representations (hundreds to thousands of tokens), enabling efficient diffusion [24, 56]. However, due to the quadratic complexity of transformer networks, it often only samples a few thousand points to represent a 3D shape, which leads to information loss and performance degradation. Therefore, we seek to improve the reconstruction quality of Vecset-based VAEs and maintain their compact representation." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.826, + 0.905, + 0.901 + ], + "angle": 0, + "content": "We begin by analyzing the shape reconstruction capability of Vecset-based VAEs. Through careful analysis, we find these methods have limited reconstruction performance, which stems from their commonly used uniform sampling. When computational constraints limit the total" + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.518, + 0.957 + ], + "angle": 0, + "content": "16251" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.123, + 0.091, + 0.26, + 0.186 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.128, + 0.185, + 0.242, + 0.196 + ], + "angle": 0, + "content": "(a) Ground Truth(GT)" + }, + { + "type": "image", + "bbox": [ + 0.104, + 0.197, + 0.281, + 0.335 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.13, + 0.337, + 0.273, + 0.349 + ], + "angle": 0, + "content": "(d) Reconstruction accuracy" + }, + { + "type": "text", + "bbox": [ + 0.288, + 0.095, + 0.325, + 0.113 + ], + "angle": 0, + "content": "point number" + }, + { + "type": "text", + "bbox": [ + 0.344, + 0.099, + 0.372, + 0.108 + ], + "angle": 0, + "content": "4096" + }, + { + "type": "text", + "bbox": [ + 0.466, + 0.099, + 0.494, + 0.108 + ], + "angle": 0, + "content": "8192" + }, + { + "type": "text", + "bbox": [ + 0.584, + 0.099, + 0.617, + 0.108 + ], + "angle": 0, + "content": "16384" + }, + { + "type": "text", + "bbox": [ + 0.697, + 0.099, + 0.732, + 0.108 + ], + "angle": 0, + "content": "20480" + }, + { + "type": "text", + "bbox": [ + 0.823, + 0.099, + 0.856, + 0.108 + ], + "angle": 0, + "content": "40960" + }, + { + "type": "image", + "bbox": [ + 0.306, + 0.127, + 0.422, + 0.21 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.428, + 0.128, + 0.542, + 0.21 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.548, + 0.128, + 0.663, + 0.21 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.668, + 0.128, + 0.783, + 0.21 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.788, + 0.129, + 0.902, + 0.21 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.476, + 0.212, + 0.682, + 0.224 + ], + "angle": 0, + "content": "(b) Visual results of the uniform sampling" + }, + { + "type": "image", + "bbox": [ + 0.307, + 0.238, + 0.422, + 0.32 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.428, + 0.238, + 0.543, + 0.32 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.548, + 0.238, + 0.662, + 0.32 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.668, + 0.238, + 0.783, + 0.32 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.789, + 0.238, + 0.903, + 0.32 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.451, + 0.335, + 0.712, + 0.347 + ], + "angle": 0, + "content": "(c) Visual results of our sharp edge sampling (SES)" + }, + { + "type": "image_caption", + "bbox": [ + 0.089, + 0.363, + 0.907, + 0.435 + ], + "angle": 0, + "content": "Figure 1. Sampling strategy comparison. Given the ground truth mesh shown in (a), we visualize point clouds produced by uniform sampling in (b) and those generated by our proposed Sharp Edge Sampling (SES) in (c), at various sampling rates. In (d), we compare the reconstruction accuracy trained with our SES and the uniform sampling using the F-score metric. The comparison demonstrates SES consistently outperforms uniform sampling under varying sampling rates, as the point clouds generated by SES are more effective in capturing the salient features of the object." + }, + { + "type": "text", + "bbox": [ + 0.089, + 0.459, + 0.483, + 0.655 + ], + "angle": 0, + "content": "number of sampled points, uniform sampling fails to prioritize geometrically salient regions, leading to the loss of fine details. To validate this observation, we experiment on a 3D mesh with intricate geometric details (e.g., keyboard buttons) shown in Figure 1 (a). We visualize the point cloud with different sampling strategies at various sampling densities in (b) and (c). As demonstrated in Figure 1 (b), even with increasing sampling rates, uniform sampling fails to preserve sharp features like keyboard buttons. This simple experiment confirms that uniform sampling fundamentally limits the capturing of geometric details, which in turn affects the VAE's reconstruction capability and the result quality of the learned diffusion models." + }, + { + "type": "text", + "bbox": [ + 0.089, + 0.657, + 0.483, + 0.793 + ], + "angle": 0, + "content": "Inspired by the success of importance sampling in various geometric processing tasks [14, 48], we introduce a similar strategy for 3D VAE training. While existing importance sampling methods focus on down-sampling point clouds, our task requires sampling points directly from mesh surfaces for shape VAE training [49, 56]. This fundamental difference necessitates a new sampling approach specifically designed for preserving geometric features from mesh representations." + }, + { + "type": "text", + "bbox": [ + 0.089, + 0.796, + 0.483, + 0.901 + ], + "angle": 0, + "content": "To address this challenge, we propose a Sharp Edge Sampling (SES) algorithm that adaptively samples points based on geometric saliency. Specifically, SES first identifies edges with significant dihedral angles on the mesh, indicating regions of high geometric complexity. It then samples points along these salient regions while maintaining a balance with uniformly sampled points to capture the over" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.459, + 0.905, + 0.489 + ], + "angle": 0, + "content": "all structure. This approach ensures comprehensive coverage of both fine details and global geometry." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.49, + 0.906, + 0.642 + ], + "angle": 0, + "content": "Building upon SES, we present Dora-VAE, a novel method achieving high-fidelity reconstruction while maintaining compact latent representations. To fully leverage these detail-rich point clouds sampled by our SES, we design a dual cross-attention architecture that effectively processes both salient and uniform regions during encoding. As shown in Figure 1, our method significantly outperforms uniform sampling in preserving shape details on the keyboard (c), with consistent improvements in F-Score across different sampling rates (d)." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.643, + 0.906, + 0.853 + ], + "angle": 0, + "content": "The common evaluation protocol for 3D VAEs is also biased. It typically uses a set of randomly selected 3D shapes, and employs general metrics (e.g., F-score, Chamfer distance) to measure the shape reconstruction quality. However, we argue it is necessary to divide the test shapes into sub-classes of different shape complexity to better evaluate these 3D VAEs. To facilitate the evaluation of 3D VAEs, we further introduce Dora-Bench, a new benchmark based on existing datasets with our novel Sharp Normal Error (SNE) metric. Dora-Bench categorizes test shapes based on their geometric complexity and SNE focuses on measuring the reconstruction quality of salient geometric features. This combination enables a more rigorous assessment of 3D VAEs." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.856, + 0.906, + 0.901 + ], + "angle": 0, + "content": "Extensive experiments demonstrate that our Dora-VAE achieves superior results. When integrated into downstream 3D diffusion models, it significantly enhances the qual" + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "16252" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.09, + 0.091, + 0.485, + 0.364 + ], + "angle": 0, + "content": "ity of generated 3D shapes, which validates that our novel sampling strategy and dual cross-attention architecture effectively preserve geometric details with compact latent spaces. To summarize, our main contributions include: 1) We propose Dora-VAE, a novel 3D VAE model for high-quality reconstruction with compact latent representations, accompanied by Dora-Bench, a comprehensive benchmark for evaluating 3D VAEs. 2) We introduce, for the first time, importance sampling to the task of 3D VAE learning and propose a Sharp Edge Sampling (SES) algorithm to prioritize geometrically salient regions. Building on SES, we design a dual cross-attention architecture to effectively encode these detail-rich point clouds. 3) We develop a systematic evaluation benchmark based on existing datasets with our novel Sharp Normal Error (SNE) metric that specifically assesses reconstruction accuracy of fine geometric details, enabling more rigorous evaluation than conventional random sampling approaches." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.377, + 0.229, + 0.393 + ], + "angle": 0, + "content": "2. Related work" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.403, + 0.484, + 0.522 + ], + "angle": 0, + "content": "Importance Sampling in Point Clouds. Importance sampling techniques have been widely used in point cloud processing tasks [14, 48]. For instance, APES [48] proposes attention-based sampling for point cloud classification and segmentation. However, these methods operate directly on point clouds rather than meshes, making them less suitable for VAE-based shape representation where preserving complete geometric information is crucial." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.523, + 0.484, + 0.84 + ], + "angle": 0, + "content": "3D Shape VAEs. Recent 3D shape VAEs follow two main approaches: volume-based and vector set-based. Volume-based methods [36, 50] like XCube [36] use sparse convolution to encode voxelized surfaces, achieving high reconstruction quality but requiring large latent codes (\\(>10,000\\) tokens). While these methods excel at preserving geometric details, their large latent spaces pose significant challenges for downstream diffusion model training. In contrast, vector set-based approaches [24, 49, 54, 56, 58] encode uniformly sampled surface points using transformers, producing highly compact latent spaces that are particularly suitable for diffusion models. However, these methods often struggle with geometric detail preservation, especially in regions with complex surface features. Our analysis reveals that this limitation primarily stems from their uniform sampling strategy: when computational constraints restrict the total number of processable points, uniform sampling fails to prioritize geometrically significant regions, leading to insufficient capture of fine details. This information loss at the sampling stage fundamentally limits these methods' ability to learn and preserve intricate geometric features." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.84, + 0.484, + 0.902 + ], + "angle": 0, + "content": "3D Content Creation. Current 3D generation methods can be categorized into three groups. Optimization-based methods [6, 23, 26, 27, 30, 34, 35, 39, 41, 45], pioneered by DreamFusion [34] utilize score distillation sam" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.091, + 0.907, + 0.41 + ], + "angle": 0, + "content": "pling (SDS) to optimize 3D representations [19, 31, 38] using 2D diffusion model priors. While these methods can achieve photorealistic results, they suffer from slow generation speed, training instability, and often struggle to maintain geometric consistency. Large reconstruction models, like LRM [16] and follow-up works [22, 29, 40, 43, 46, 51] employ large-scale sparse-view reconstruction for efficient 3D generation. However, their lack of explicit geometric priors often leads to compromised geometric fidelity and inconsistent surface details. 3D native generative models [24, 49, 54, 56, 58], represented by 3DShape2VecSet [54], adopt a two-stage approach: first training a 3D VAE to encode shapes into latent space, then training a conditional latent diffusion model for generation. This approach ensures better geometric consistency through the VAE's built-in geometric constraints. However, the quality of generated shapes is fundamentally limited by the VAE's reconstruction capability. Recent works [24, 53] have shown that improving VAE reconstruction directly enhances downstream generation quality, which motivates our focus on advancing VAE design." + }, + { + "type": "title", + "bbox": [ + 0.514, + 0.424, + 0.605, + 0.439 + ], + "angle": 0, + "content": "3. Method" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.45, + 0.907, + 0.525 + ], + "angle": 0, + "content": "In this section, we present Dora-VAE for high-quality 3D reconstruction, and Dora-Bench for 3D VAE evaluation. We first briefly review 3DShape2VecSet [54] in Section 3.1, the foundation of our method and then detail our key innovations in Section 3.2 and Section 3.3." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.536, + 0.785, + 0.552 + ], + "angle": 0, + "content": "3.1. Preliminary: 3DShape2VecSet" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.558, + 0.906, + 0.619 + ], + "angle": 0, + "content": "3DShape2VecSet [54] introduces a transformer-based 3D VAE that encodes uniformly sampled surface points into compact latent codes. Given a 3D surface S, their pipeline consists of three key steps:" + }, + { + "type": "text", + "bbox": [ + 0.514, + 0.62, + 0.906, + 0.695 + ], + "angle": 0, + "content": "- Surface Sampling: Uniformly sample \\( N_{d} \\) points on the surface S using Poisson disk sampling [52] to obtain a dense point cloud \\( P_{d} \\), then downsample it to \\( N_{s} \\) points via Farthest Point Sampling (FPS) [32] to get a sparse point cloud \\( P_{s} \\):" + }, + { + "type": "equation", + "bbox": [ + 0.536, + 0.708, + 0.907, + 0.726 + ], + "angle": 0, + "content": "\\[\nP _ {d} = \\left\\{p _ {d} ^ {i} \\in S \\mid i = 1, \\dots , N _ {d} \\right\\}, P _ {s} = \\operatorname {F P S} \\left(P _ {d}, N _ {s}\\right). \\tag {1}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.514, + 0.738, + 0.906, + 0.783 + ], + "angle": 0, + "content": "- Feature Encoding: Compute the point cloud feature \\( C \\) via the cross-attention between \\( P_{s} \\) and \\( P_{d} \\), followed by some self-attention layers to generate the latent code \\( z \\):" + }, + { + "type": "equation", + "bbox": [ + 0.552, + 0.796, + 0.907, + 0.813 + ], + "angle": 0, + "content": "\\[\nC = \\operatorname {C r o s s A t t n} \\left(P _ {s}, P _ {d}, P _ {d}\\right), z = \\operatorname {S e l f A t t n} (C). \\tag {2}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.825, + 0.906, + 0.872 + ], + "angle": 0, + "content": "- Geometry Decoding: Further decode \\( z \\) through selfattention layers and predict occupancy values using randomly sampled spatial query points \\( Q_{space} \\in \\mathbb{R}^3 \\):" + }, + { + "type": "equation", + "bbox": [ + 0.59, + 0.884, + 0.907, + 0.903 + ], + "angle": 0, + "content": "\\[\n\\hat {O} = \\operatorname {C r o s s A t t n} \\left(Q _ {\\text {s p a c e}}, \\operatorname {S e l f A t t n} (z)\\right). \\tag {3}\n\\]" + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "16253" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.095, + 0.09, + 0.48, + 0.199 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.199, + 0.201, + 0.34, + 0.214 + ], + "angle": 0, + "content": "(a) Sharp edge sampling" + }, + { + "type": "image", + "bbox": [ + 0.092, + 0.214, + 0.482, + 0.34 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.09, + 0.351, + 0.483, + 0.447 + ], + "angle": 0, + "content": "Figure 2. Overview of Dora-VAE. (a) We utilize the proposed sharp edge sampling technique to extract both salient and uniform points from the input mesh. These points are then combined with dense points, effectively capturing both salient regions and smooth areas. (b) To enhance the encoding of point clouds sampled through sharp edge sampling, we design a dual cross-attention architecture." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.462, + 0.484, + 0.523 + ], + "angle": 0, + "content": "While this method generates compact latent codes for diffusion, the uniform sampling limits its ability to capture fine geometric details. Our work addresses this limitation by carefully designed sampling and encoding strategies." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.533, + 0.208, + 0.547 + ], + "angle": 0, + "content": "3.2. Dora-VAE" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.555, + 0.483, + 0.707 + ], + "angle": 0, + "content": "Figure 2 gives an overview of our pipeline. For each input mesh, we augment the uniformly sampled point cloud \\( P_{u} \\) with more important points \\( P_{a} \\) sampled by our proposed sharp edge sampling strategy, which forms the dense point cloud \\( P_{d} \\). During the encoding process, we compute the attention for \\( P_{u} \\) and \\( P_{a} \\) separately via a simple-yet-effective dual cross attention mechanism and sum the results for self-attention to compute the latent code \\( z \\). Our VAE training largely follows that of 3DShape2VecSet [54], which is supervised by a loss evaluated on the occupancy field." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.715, + 0.334, + 0.73 + ], + "angle": 0, + "content": "3.2.1. Sharp Edge Sampling (SES)" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.735, + 0.483, + 0.854 + ], + "angle": 0, + "content": "We propose SES to effectively sample points from geometrically salient regions. To ensure surface coverage, we also sample points uniformly. Our final sampled point cloud \\( P_{d} \\) combines uniformly sampled points \\( P_{u} \\) with points specifically sampled from salient regions \\( P_{a} \\) as \\( P_{d} = P_{u} \\cup P_{a} \\). Our method computes salient points \\( P_{a} \\) through two steps: detecting salient edges and sampling points from these regions." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.856, + 0.484, + 0.901 + ], + "angle": 0, + "content": "Salient Edges Detection. Given a triangular mesh, we identify a set of salient edges \\(\\Gamma\\) by analyzing dihedral angles between adjacent faces, which calculates the angle be" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.092, + 0.906, + 0.152 + ], + "angle": 0, + "content": "tween the normal vectors of adjacent faces, providing a direct measure of surface curvature at mesh edges. For each edge \\( e \\) shared by adjacent faces \\( f_{1} \\) and \\( f_{2} \\), we compute the dihedral angle \\( \\theta_{e} \\) as:" + }, + { + "type": "equation", + "bbox": [ + 0.607, + 0.158, + 0.907, + 0.193 + ], + "angle": 0, + "content": "\\[\n\\theta_ {e} = \\operatorname {a r c c o s} \\left(\\frac {\\mathbf {n} _ {f _ {1}} \\cdot \\mathbf {n} _ {f _ {2}}}{\\| \\mathbf {n} _ {f _ {1}} \\| \\| \\mathbf {n} _ {f _ {2}} \\|}\\right), \\tag {4}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.197, + 0.907, + 0.242 + ], + "angle": 0, + "content": "where \\(\\mathbf{n}_{f_1}\\) and \\(\\mathbf{n}_{f_2}\\) are the normals of \\(f_{1}\\) and \\(f_{2}\\). The salient edge set \\(\\Gamma\\) contains all edges with a dihedral angle exceeding a predefined threshold \\(\\tau\\):" + }, + { + "type": "equation", + "bbox": [ + 0.649, + 0.257, + 0.905, + 0.273 + ], + "angle": 0, + "content": "\\[\n\\Gamma = \\left\\{e \\mid \\theta_ {e} > \\tau \\right\\} \\tag {5}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.277, + 0.882, + 0.291 + ], + "angle": 0, + "content": "Let \\(N_{\\Gamma} = |\\Gamma|\\) represent the number of the salient edges." + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.292, + 0.906, + 0.336 + ], + "angle": 0, + "content": "Salient Points Sampling. For each salient edge \\( e \\in \\Gamma \\), we collect its two vertices \\( v_{e,1} \\) and \\( v_{e,2} \\) into a salient vertex set \\( P_{\\Gamma} \\):" + }, + { + "type": "equation", + "bbox": [ + 0.622, + 0.337, + 0.905, + 0.353 + ], + "angle": 0, + "content": "\\[\nP _ {\\Gamma} = \\left\\{v _ {e, 1}, v _ {e, 2} \\mid e \\in \\Gamma \\right\\}, \\tag {6}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.357, + 0.906, + 0.401 + ], + "angle": 0, + "content": "where duplicate vertices from connecting edges are included only once. Let \\( N_V = |P_{\\Gamma}| \\) denote the number of unique vertices in \\( P_{\\Gamma} \\)." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.403, + 0.906, + 0.446 + ], + "angle": 0, + "content": "Given a target number of salient points \\( N_{\\mathrm{desired}} \\), we generate the salient point set \\( P_{a} \\) based on the available salient vertices:" + }, + { + "type": "equation", + "bbox": [ + 0.532, + 0.467, + 0.905, + 0.522 + ], + "angle": 0, + "content": "\\[\nP _ {a} = \\left\\{ \\begin{array}{l l} \\operatorname {F P S} \\left(P _ {\\Gamma}, N _ {\\text {d e s i r e d}}\\right), & \\text {i f} N _ {\\text {d e s i r e d}} < = N _ {V}, \\\\ P _ {\\Gamma} \\cup P _ {\\text {i n t e r p o l a t e d}}, & \\text {i f} 0 < N _ {V} < N _ {\\text {d e s i r e d}}, \\\\ \\emptyset , & \\text {i f} N _ {V} = 0. \\end{array} \\right. \\tag {7}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.528, + 0.906, + 0.665 + ], + "angle": 0, + "content": "When we have excess salient vertices (\\(N_{\\text{desired}} \\leq N_V\\)), we use FPS to downsample \\(P_{\\Gamma}\\) to obtain \\(P_a\\). For cases with insufficient salient vertices (\\(N_V < N_{\\text{desired}}\\)), we include all vertices from \\(P_{\\Gamma}\\) and supplement with additional points \\(P_{\\text{interpolated}}\\). These additional points are generated by uniformly sampling \\((N_{\\text{desired}} - N_V) / N_{\\Gamma}\\) points along each salient edge in \\(\\Gamma\\), ensuring comprehensive coverage of salient features. When no salient edges are detected (\\(N_V = 0\\)), \\(P_a\\) remains empty." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.67, + 0.707, + 0.683 + ], + "angle": 0, + "content": "3.2.2. Dual Cross Attention" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.688, + 0.906, + 0.764 + ], + "angle": 0, + "content": "Given the point clouds \\( P_{d} \\) produced by our SES strategy, we design a dual cross-attention architecture to effectively encode both uniform and salient regions. Following 3DShape2VecSet [54], we first downsample \\( P_{u} \\) and \\( P_{a} \\) separately using FPS:" + }, + { + "type": "equation", + "bbox": [ + 0.58, + 0.772, + 0.905, + 0.788 + ], + "angle": 0, + "content": "\\[\nP _ {s} = \\operatorname {F P S} \\left(P _ {u}, N _ {s, 1}\\right) \\cup \\operatorname {F P S} \\left(P _ {a}, N _ {s, 2}\\right), \\tag {8}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.794, + 0.906, + 0.854 + ], + "angle": 0, + "content": "where \\(N_{s,1}\\) and \\(N_{s,2}\\) is the number of downsampled point clouds from \\(P_u\\) and \\(P_a\\), respectively. We then compute cross-attention features separately for uniform and salient points as follows:" + }, + { + "type": "equation", + "bbox": [ + 0.612, + 0.863, + 0.905, + 0.879 + ], + "angle": 0, + "content": "\\[\nC _ {u} = \\operatorname {C r o s s A t t n} \\left(P _ {s}, P _ {u}, P _ {u}\\right) \\tag {9}\n\\]" + }, + { + "type": "equation", + "bbox": [ + 0.614, + 0.882, + 0.905, + 0.897 + ], + "angle": 0, + "content": "\\[\nC _ {a} = \\operatorname {C r o s s A t t n} \\left(P _ {s}, P _ {a}, P _ {a}\\right) \\tag {10}\n\\]" + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "16254" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.097, + 0.089, + 0.48, + 0.177 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.182, + 0.177, + 0.416, + 0.186 + ], + "angle": 0, + "content": "(a) Counts of Different Datasets Across Detail Levels" + }, + { + "type": "image", + "bbox": [ + 0.107, + 0.196, + 0.258, + 0.273 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.108, + 0.276, + 0.256, + 0.287 + ], + "angle": 0, + "content": "(b) Total Counts by Detail Levels" + }, + { + "type": "image", + "bbox": [ + 0.268, + 0.188, + 0.47, + 0.278 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.29, + 0.278, + 0.43, + 0.287 + ], + "angle": 0, + "content": "(c) Visualization of detail levels" + }, + { + "type": "image_caption", + "bbox": [ + 0.09, + 0.3, + 0.484, + 0.371 + ], + "angle": 0, + "content": "Figure 3. Our proposed benchmark include 3D shapes from the ABO [10], GSO [13], Meta [1], and Objaverse [12] datasets. (a) The histogram of different datasets across different shape complexities. (b) The pie chart of the total counts by shape complexities. (c) Sample shapes of different shape complexities." + }, + { + "type": "image", + "bbox": [ + 0.101, + 0.397, + 0.48, + 0.59 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.09, + 0.604, + 0.482, + 0.633 + ], + "angle": 0, + "content": "Figure 4. The process of computing sharp normal errors (SNE). We compute MSE loss in the sharp regions of the normal." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.666, + 0.483, + 0.695 + ], + "angle": 0, + "content": "The final point cloud feature \\( C \\) combines both attention results:" + }, + { + "type": "equation", + "bbox": [ + 0.235, + 0.708, + 0.483, + 0.724 + ], + "angle": 0, + "content": "\\[\nC = C _ {u} + C _ {a}. \\tag {11}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.743, + 0.484, + 0.819 + ], + "angle": 0, + "content": "This dual attention design enables separate focus on uniform and salient regions during feature extraction. Following 3DShape2VecSet [54], we use \\(C\\) to predict the occupancy field \\(\\hat{O}\\) through self-attention blocks. The whole model with parameters \\(\\psi\\) are optimized using MSE loss:" + }, + { + "type": "equation", + "bbox": [ + 0.154, + 0.836, + 0.483, + 0.879 + ], + "angle": 0, + "content": "\\[\n\\nabla_ {\\psi} \\mathcal {L} _ {\\mathrm {M S E}} (\\hat {O}, O) = \\mathbb {E} \\left[ 2 (\\hat {O} - O) \\frac {\\partial \\hat {O}}{\\partial \\psi} \\right]. \\tag {12}\n\\]" + }, + { + "type": "title", + "bbox": [ + 0.514, + 0.091, + 0.644, + 0.105 + ], + "angle": 0, + "content": "3.3. Dora-Bench" + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.113, + 0.843, + 0.128 + ], + "angle": 0, + "content": "3.3.1. Geometric Complexity-based Evaluation" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.132, + 0.905, + 0.235 + ], + "angle": 0, + "content": "To enable a more rigorous evaluation of VAE performance, we propose Dora-bench, a benchmark that systematically categorizes test shapes based on their geometric complexity. Unlike previous methods that use randomly selected test sets, we measure shape complexity using the number of salient edges \\( N_{\\Gamma} \\) (Section 3.2.1) and classify shapes into four levels:" + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.238, + 0.797, + 0.252 + ], + "angle": 0, + "content": "- Level 1 (Less Detail): \\(0 < N_{\\Gamma} \\leq 5000\\);" + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.253, + 0.864, + 0.267 + ], + "angle": 0, + "content": "- Level 2 (Moderate Detail): \\(5000 < N_{\\Gamma} \\leq 10000\\);" + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.268, + 0.837, + 0.282 + ], + "angle": 0, + "content": "- Level 3 (Rich Detail): \\(10000 < N_{\\Gamma} \\leq 50000\\);" + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.283, + 0.81, + 0.298 + ], + "angle": 0, + "content": "- Level 4 (Very Rich Detail): \\(N_{\\Gamma} > 50000\\)" + }, + { + "type": "list", + "bbox": [ + 0.513, + 0.238, + 0.864, + 0.298 + ], + "angle": 0, + "content": null + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.299, + 0.905, + 0.389 + ], + "angle": 0, + "content": "We curate test shapes from multiple public datasets including GSO [13], ABO [10], Meta [1], and Objaverse [12] to ensure diverse geometric complexities. Figure 3 shows the distribution of shapes across complexity levels (a,b) and example meshes from each level (c). Please refer to our supplementary materials for more examples." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.394, + 0.753, + 0.409 + ], + "angle": 0, + "content": "3.3.2. Sharp Normal Error (SNE)" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.413, + 0.907, + 0.655 + ], + "angle": 0, + "content": "Building on our Dora-bench, we further introduce Sharp Normal Error (SNE) to evaluate reconstruction quality in salient regions. While existing metrics like Chamfer Distance and F-Score capture overall shape similarity, they fail to specifically assess the preservation of fine geometric details. SNE addresses this limitation by measuring normal map differences between reconstructed and ground truth shapes in geometrically significant areas. As illustrated in Figure 4, we render normal maps of the ground truth shape from multiple viewpoints and identify salient regions using Canny edge detection. These regions are dilated to create evaluation masks. The final SNE metric is computed as the Mean Squared Error between ground truth and reconstructed normal maps within the masked areas. This process enables focused evaluation of how well VAEs preserve sharp geometric features during reconstruction." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.666, + 0.646, + 0.683 + ], + "angle": 0, + "content": "4. Experiments" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.691, + 0.906, + 0.735 + ], + "angle": 0, + "content": "We conducted intensive experiments to validate the effectiveness of our proposed 3D VAE and compare it with other state-of-the-art methods." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.743, + 0.731, + 0.759 + ], + "angle": 0, + "content": "4.1. Implementation Details" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.765, + 0.907, + 0.903 + ], + "angle": 0, + "content": "We follow CLAY [56] for mesh preprocessing to ensure watertight 3D models. Our VAE is trained on a subset filtered from Objaverse [12], containing approximately 400,000 3D meshes. We filter out low-quality meshes with missing faces or severe self-intersections to ensure training stability. Our training is conducted on 32 A100 GPUs for two days using a batch size of 2048 and a learning rate of 5e-5. We employ Flash-Attention-v2 [11], mixed-precision training with FP16 and gradient checkpointing [8] to optimize" + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "16255" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.092, + 0.088, + 0.905, + 0.856 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.09, + 0.867, + 0.907, + 0.896 + ], + "angle": 0, + "content": "Figure 5. Qualitative comparison of the VAE reconstruction results. \\(\\dagger\\) indicates the fine-tuning model that uses the same training data as ours." + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.946, + 0.52, + 0.957 + ], + "angle": 0, + "content": "16256" + } + ], + [ + { + "type": "table", + "bbox": [ + 0.109, + 0.089, + 0.895, + 0.244 + ], + "angle": 0, + "content": "
MethodsLCL↑ F-score(0.01) × 100↑ F-score(0.005) × 100↓ CD × 10000↓ SNE × 100
L1L2L3L4L1L2L3L4L1L2L3L4L1L2L3L4
Xcube [36]>1000098.96898.79998.61598.22695.52593.87292.32285.3656.3156.2887.9359.9261.5791.4321.4301.679
Xcube† [36]>1000099.39399.79499.82499.07996.75395.53593.42287.3654.0154.1425.7407.6271.5431.4081.2591.639
Craftsman [24]25698.01695.87491.75681.73987.99482.54973.00057.3794.3899.12914.53033.4411.9061.8732.1913.933
Ours w/o DCA128099.96499.92599.67897.89096.56195.97591.61883.1242.2362.5064.4446.4321.4481.2151.2051.828
Ours w/o SES,DCA128099.94499.81497.29496.77995.97794.62388.40679.2402.4222.9833.9806.1961.4961.3131.3522.207
Ours full25699.50798.98696.66989.57793.27290.46682.38668.6693.3565.20210.27624.5271.5551.4101.6183.035
128099.98899.95599.88099.17097.03896.83193.45887.4732.0972.5003.9455.2651.4331.1861.1371.579
" + }, + { + "type": "table_caption", + "bbox": [ + 0.136, + 0.248, + 0.86, + 0.262 + ], + "angle": 0, + "content": "Table 1. Quantitative comparison in Dora-bench. \\( {}^{ \\dagger } \\) indicates the fine-tuning model that uses the same training data as ours." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.289, + 0.483, + 0.35 + ], + "angle": 0, + "content": "memory usage and training efficiency. For the parameters in sharp edge sampling, we set \\( N_{\\mathrm{desired}} = 16384 \\) and \\( \\tau = 30 \\). We set the low threshold to 20 and the high threshold to 200 for the canny edge detection." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.36, + 0.273, + 0.377 + ], + "angle": 0, + "content": "4.2. Evaluation Setting" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.383, + 0.484, + 0.594 + ], + "angle": 0, + "content": "Metrics. We evaluate reconstruction quality by comparing input meshes with their decoded counterparts from different 3D VAEs using 1M sampled points under three metrics: 1) F-score \\((r)\\) [20], which reconstruction accuracy by computing precision and recall of point correspondences within distance threshold \\(r\\). Specifically, we report F-score (0.01) and F-score (0.005) with shapes normalized \\([-1,1]\\). 2) Chamfer Distance (CD), which computes the average distance between each reconstructed point and its nearest ground truth point. 3) Sharp Normal Errors (SNE) as proposed in Section 3.3.2, which evaluate normal map differences in salient regions. For fair comparison, we also report Latent Code Length (LCL) as longer codes typically enable better reconstruction." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.595, + 0.484, + 0.731 + ], + "angle": 0, + "content": "Baselines. We compare Dora-VAE with state-of-the-art approaches, including: 1) XCube-VAE [36], a volumetric method with larger latent codes; 2) XCube-VAE† [36], our fine-tuned version of the original XCube-VAE on the same dataset; 3) Craftsman-VAE[24], which fine-tune the 3DShape2VecSet [54] with shorter latent codes on Objverse. We exclude VAE models from Direct3D [49] and CLAY[56] as as their implementations were not publicly available at submission time." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.742, + 0.313, + 0.759 + ], + "angle": 0, + "content": "4.3. Qualitative Comparison" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.765, + 0.483, + 0.903 + ], + "angle": 0, + "content": "Figure 5 shows visual comparisons of different methods across different complexity levels from our Dora-bench dataset. We visualize both ground truth and reconstructed meshes using surface normal coloring to highlight geometric details. For shapes with lower complexity (L1 and L2), all methods achieve comparable reconstruction quality. However, when dealing with shapes of higher complexity (L3 and L4), the advantages of our method become obvious. While XCube-VAE achieves similar visual quality" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.289, + 0.907, + 0.426 + ], + "angle": 0, + "content": "to ours, it requires a significantly larger latent space - more than \\(8 \\times\\) the size of ours (\\(>10,000\\) dimensions vs. 1,280). This substantial reduction in latent code length, while maintaining high reconstruction fidelity, makes our method particularly suitable for training 3D diffusion models. In contrast, Craftsman-VAE shows a noticeable degradation in reconstruction quality for complex shapes, failing to capture fine geometric details. Additional visual comparisons are provided in the appendix." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.436, + 0.747, + 0.452 + ], + "angle": 0, + "content": "4.4. Quantitative Comparison" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.458, + 0.907, + 0.639 + ], + "angle": 0, + "content": "Table 1 presents quantitative results of different methods across different complexity levels of Dora-bench. Our method consistently outperforms baselines across all levels, with larger margins on more complex shapes (L3 and L4). The advantage is particularly evident in CD metrics, where our method with only 256 latent codes surpasses even our fine-tuned version of XCube-VAE (3.356 vs. 4.015). When using 1280 latent codes, our method further decreases CD to 2.097, achieving a \\(47.77\\%\\) improvement over XCube-VAE†. We attribute XCube-VAE's lower performance to its use of NKSR [18] for mesh extraction, which introduces additional quantization errors." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.64, + 0.909, + 0.792 + ], + "angle": 0, + "content": "Notably, our method demonstrates superior performance in preserving geometric details, as reflected by the SNE metric. For example, in L4 shapes where geometric complexity is highest, our method achieves an SNE of 1.579 compared to 1.639 from XCube-VAE†, representing a \\(3.7\\%\\) improvement. This significant gain in SNE aligns with our qualitative observations in Figure 5, where our method better preserves fine details such as sharp edges and complex surface variations, demonstrating the effectiveness of our sharp edge sampling strategy." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.802, + 0.681, + 0.817 + ], + "angle": 0, + "content": "4.5. Ablation Studies" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.825, + 0.905, + 0.869 + ], + "angle": 0, + "content": "To evaluate the contribution of each component, we compare our full model with two variants under the same training conditions:" + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.871, + 0.907, + 0.902 + ], + "angle": 0, + "content": "- Ours w/o SES, DCA. This variant removes both sharp edge sampling (SES) and dual cross attention (DCA), i.e." + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.52, + 0.958 + ], + "angle": 0, + "content": "16257" + } + ], + [ + { + "type": "image_caption", + "bbox": [ + 0.182, + 0.094, + 0.203, + 0.105 + ], + "angle": 0, + "content": "GT" + }, + { + "type": "image", + "bbox": [ + 0.129, + 0.108, + 0.257, + 0.22 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.147, + 0.225, + 0.238, + 0.237 + ], + "angle": 0, + "content": "w/o SES, DCA" + }, + { + "type": "image_caption", + "bbox": [ + 0.359, + 0.093, + 0.385, + 0.104 + ], + "angle": 0, + "content": "Full" + }, + { + "type": "image", + "bbox": [ + 0.308, + 0.108, + 0.434, + 0.22 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.345, + 0.223, + 0.404, + 0.235 + ], + "angle": 0, + "content": "w/o DAC" + }, + { + "type": "image", + "bbox": [ + 0.131, + 0.243, + 0.258, + 0.356 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.308, + 0.242, + 0.436, + 0.356 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.09, + 0.371, + 0.483, + 0.426 + ], + "angle": 0, + "content": "Figure 6. Ablation studies of our method. Given the ground truth of mesh, we employ both our full model and its variations to reconstruct the ground truth mesh, highlighting significant reconstruction discrepancies with red boxes." + }, + { + "type": "text", + "bbox": [ + 0.104, + 0.454, + 0.483, + 0.484 + ], + "angle": 0, + "content": "using only uniformly sampled point clouds with Poisson disk sampling [52], while maintaining an equal \\( N_{d} \\)." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.484, + 0.483, + 0.56 + ], + "angle": 0, + "content": "- Ours w/o DCA. This variant retains SES but removes DCA, i.e. using a single cross attention adopted by [54]. As shown in Figure 6 and Table 1, our full model consistently outperforms these variants, validating the effectiveness of both components." + }, + { + "type": "title", + "bbox": [ + 0.09, + 0.573, + 0.384, + 0.591 + ], + "angle": 0, + "content": "5. Application: Single Image to 3D" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.599, + 0.483, + 0.901 + ], + "angle": 0, + "content": "We demonstrate the effectiveness of our VAE by applying it to single-image 3D generation through diffusion models. Following CLAY [56], we implement a latent diffusion model based on the DiT [33] architecture. For a fair comparison, we fine-tune Craftsman-VAE on our dataset (denoted as Craftsman-VAE†) since both Craftsman-VAE and 3DShape2VecSet were originally trained on smaller datasets. Note that XCube-VAE is excluded from comparison due to its 10,000-dimensional latent codes being impractical for diffusion model training. Figure 7 shows some generation results from diffusion models trained with our Dora-VAE and Craftsman-VAE†. Both models share identical architecture (0.39B parameters) and training conditions (same dataset, 32 A100 GPUs, 3 days). Our Dora-VAE demonstrates significantly better preservation of geometric details in the generated shapes, validating its effectiveness as a foundation for 3D generation tasks. While further improvements could be achieved with more extensive training data and computational resources, we focus on validating the VAE's capabilities in this work and leave such exten" + }, + { + "type": "image_caption", + "bbox": [ + 0.536, + 0.093, + 0.62, + 0.108 + ], + "angle": 0, + "content": "Input Image" + }, + { + "type": "image", + "bbox": [ + 0.527, + 0.121, + 0.637, + 0.221 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.686, + 0.093, + 0.72, + 0.107 + ], + "angle": 0, + "content": "Ours" + }, + { + "type": "image", + "bbox": [ + 0.655, + 0.118, + 0.768, + 0.221 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.791, + 0.092, + 0.866, + 0.106 + ], + "angle": 0, + "content": "Craftsman" + }, + { + "type": "image", + "bbox": [ + 0.775, + 0.118, + 0.903, + 0.221 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.523, + 0.23, + 0.629, + 0.334 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.653, + 0.23, + 0.759, + 0.33 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.783, + 0.229, + 0.892, + 0.334 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.526, + 0.342, + 0.628, + 0.448 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.653, + 0.339, + 0.759, + 0.446 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.78, + 0.342, + 0.884, + 0.446 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.512, + 0.467, + 0.907, + 0.524 + ], + "angle": 0, + "content": "Figure 7. The diffusion results of the single image to 3D generation trained on our Dora-VAE and Craftsman†. The 3D geometry generated by the diffusion model trained on our proposed Dora-VAE has more details under the same experimental environment." + }, + { + "type": "text", + "bbox": [ + 0.514, + 0.528, + 0.661, + 0.542 + ], + "angle": 0, + "content": "sions for future work." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.557, + 0.634, + 0.573 + ], + "angle": 0, + "content": "6. Conclusion" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.581, + 0.907, + 0.869 + ], + "angle": 0, + "content": "In this work, we introduce Dora-VAE, a novel VAE designed for high-quality 3D shape compression and reconstruction. At its core, Dora-VAE introduces sharp edge sampling to effectively capture salient geometric features, complemented by a dual cross-attention architecture that enhances the encoding of these detail-rich point clouds. To enable more rigorous evaluation of VAE performance, we develop Dora-bench, which systematically categorizes shapes based on geometric complexity and introduces the Sharp Normal Error (SNE) metric for specifically assessing the preservation of fine geometric details. Our comprehensive experiments demonstrate that Dora-VAE significantly outperforms existing methods across varying levels of shape complexity. Furthermore, we show that the improved reconstruction capability of Dora-VAE directly enhances the quality of downstream tasks by applying it to single-image 3D generation. The superior performance in generating geometric details validates our approach of focusing on salient region sampling and encoding for 3D VAE design." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.869, + 0.906, + 0.899 + ], + "angle": 0, + "content": "Acknowledgements. This work is partially supported by the project L0751 between Bytedance and HKUST." + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "16258" + } + ], + [ + { + "type": "title", + "bbox": [ + 0.093, + 0.09, + 0.188, + 0.106 + ], + "angle": 0, + "content": "References" + }, + { + "type": "ref_text", + "bbox": [ + 0.101, + 0.115, + 0.482, + 0.143 + ], + "angle": 0, + "content": "[1] Digital twin catalog. META, 2024. https://www.projectaria.com/datasets/dtc/.5" + }, + { + "type": "ref_text", + "bbox": [ + 0.101, + 0.145, + 0.483, + 0.215 + ], + "angle": 0, + "content": "[2] Yukang Cao, Yan-Pei Cao, Kai Han, Ying Shan, and KwanYee K. Wong. Dreamavatar: Text-and-shape guided 3d human avatar generation via diffusion models. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 958-968, 2024. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.217, + 0.483, + 0.271 + ], + "angle": 0, + "content": "[3] Yukang Cao, Liang Pan, Kai Han, Kwan-Yee K Wong, and Ziwei Liu. Avatargo: Zero-shot 4d human-object interaction generation and animation. arXiv preprint arXiv:2410.07164, 2024. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.274, + 0.482, + 0.342 + ], + "angle": 0, + "content": "[4] Junsong Chen, Jincheng Yu, Chongjian Ge, Lewei Yao, Enze Xie, Yue Wu, Zhongdao Wang, James Kwok, Ping Luo, Huchuan Lu, and Zhenguo Li. Pixart-α: Fast training of diffusion transformer for photorealistic text-to-image synthesis, 2023. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.345, + 0.482, + 0.4 + ], + "angle": 0, + "content": "[5] Junsong Chen, Yue Wu, Simian Luo, Enze Xie, Sayak Paul, Ping Luo, Hang Zhao, and Zhenguo Li. Pixart-δ: Fast and controllable image generation with latent consistency models, 2024. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.403, + 0.482, + 0.47 + ], + "angle": 0, + "content": "[6] Rui Chen, Yongwei Chen, Ningxin Jiao, and Kui Jia. Fantasia3d: Disentangling geometry and appearance for high-quality text-to-3d content creation. In Proceedings of the IEEE/CVF International Conference on Computer Vision (ICCV), 2023. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.473, + 0.482, + 0.543 + ], + "angle": 0, + "content": "[7] Rui Chen, Mingyi Shi, Shaoli Huang, Ping Tan, Taku Komura, and Xuelin Chen. Taming diffusion probabilistic models for character control. In ACM SIGGRAPH 2024 Conference Papers, New York, NY, USA, 2024. Association for Computing Machinery. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.545, + 0.482, + 0.585 + ], + "angle": 0, + "content": "[8] Tianqi Chen, Bing Xu, Chiyuan Zhang, and Carlos Guestrin. Training deep nets with sublinear memory cost. CoRR, abs/1604.06174, 2016. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.588, + 0.482, + 0.643 + ], + "angle": 0, + "content": "[9] Yongwei Chen, Rui Chen, Jiabao Lei, Yabin Zhang, and Kui Jia. Tango: Text-driven photorealistic and robust 3d stylization via lighting decomposition. Advances in Neural Information Processing Systems, 35:30923-30936, 2022. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.646, + 0.482, + 0.741 + ], + "angle": 0, + "content": "[10] Jasmine Collins, Shubham Goel, Kenan Deng, Achleshwar Luthra, Leon Xu, Erhan Gundogdu, Xi Zhang, Tomas F Yago Vicente, Thomas Dideriksen, Himanshu Arora, et al. Abo: Dataset and benchmarks for real-world 3d object understanding. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 21126-21136, 2022. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.745, + 0.482, + 0.8 + ], + "angle": 0, + "content": "[11] Tri Dao. Flashattention-2: Faster attention with better parallelism and work partitioning. In The Twelfth International Conference on Learning Representations, ICLR 2024, Vienna, Austria, May 7-11, 2024. OpenReview.net, 2024. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.802, + 0.482, + 0.87 + ], + "angle": 0, + "content": "[12] Matt Deitke, Dustin Schwenk, Jordi Salvador, Luca Weihs, Oscar Michel, Eli VanderBilt, Ludwig Schmidt, Kiana Ehsani, Aniruddha Kembhavi, and Ali Farhadi. Objverse: A universe of annotated 3d objects. arXiv preprint arXiv:2212.08051, 2022. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.873, + 0.482, + 0.901 + ], + "angle": 0, + "content": "[13] Laura Downs, Anthony Francis, Nate Koenig, Brandon Kinman, Ryan Hickman, Krista Reymann, Thomas B McHugh," + }, + { + "type": "list", + "bbox": [ + 0.094, + 0.115, + 0.483, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.548, + 0.093, + 0.905, + 0.148 + ], + "angle": 0, + "content": "and Vincent Vanhoucke. Google scanned objects: A high-quality dataset of 3d scanned household items. In 2022 International Conference on Robotics and Automation (ICRA), pages 2553-2560. IEEE, 2022. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.151, + 0.905, + 0.204 + ], + "angle": 0, + "content": "[14] Fabian Groh, Patrick Wieschollek, and Hendrik P. A. Lensch. Flex-convolution (million-scale point-cloud learning beyond grid-worlds). In *Asian Conference on Computer Vision (ACCV)*, 2018. 2, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.207, + 0.905, + 0.247 + ], + "angle": 0, + "content": "[15] Jonathan Ho, Ajay Jain, and Pieter Abbeel. Denoising diffusion probabilistic models. arXiv preprint arxiv:2006.11239, 2020. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.25, + 0.905, + 0.332 + ], + "angle": 0, + "content": "[16] Yicong Hong, Kai Zhang, Juixiang Gu, Sai Bi, Yang Zhou, Difan Liu, Feng Liu, Kalyan Sunkavalli, Trung Bui, and Hao Tan. LRM: large reconstruction model for single image to 3d. In The Twelfth International Conference on Learning Representations, ICLR 2024, Vienna, Austria, May 7-11, 2024. OpenReview.net, 2024. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.334, + 0.905, + 0.388 + ], + "angle": 0, + "content": "[17] Yicong Hong, Kai Zhang, Juixiang Gu, Sai Bi, Yang Zhou, Difan Liu, Feng Liu, Kalyan Sunkavalli, Trung Bui, and Hao Tan. Lrm: Large reconstruction model for single image to 3d, 2024. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.391, + 0.905, + 0.459 + ], + "angle": 0, + "content": "[18] Jiahui Huang, Zan Gojcic, Matan Atzmon, Or Litany, Sanja Fidler, and Francis Williams. Neural kernel surface reconstruction. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 4369-4379, 2023. 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.462, + 0.905, + 0.515 + ], + "angle": 0, + "content": "[19] Bernhard Kerbl, Georgios Kopanas, Thomas Leimkuhler, and George Drettakis. 3d gaussian splatting for real-time radiance field rendering. ACM Trans. Graph., 42(4):139-1, 2023. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.519, + 0.905, + 0.559 + ], + "angle": 0, + "content": "[20] Arno Knapitsch, Jaesik Park, Qian-Yi Zhou, and Vladlen Koltun. Tanks and temples: benchmarking large-scale scene reconstruction. ACM Trans. Graph., 36(4), 2017. 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.562, + 0.905, + 0.616 + ], + "angle": 0, + "content": "[21] Yushi Lan, Fangzhou Hong, Shuai Yang, Shangchen Zhou, Xuyi Meng, Bo Dai, Xingang Pan, and Chen Change Loy. Ln3diff: Scalable latent neural fields diffusion for speedy 3d generation. In ECCV, 2024. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.619, + 0.905, + 0.687 + ], + "angle": 0, + "content": "[22] Jiahao Li, Hao Tan, Kai Zhang, Zexiang Xu, Fujun Luan, Yinghao Xu, Yicong Hong, Kalyan Sunkavalli, Greg Shakhnarovich, and Sai Bi. Instant3d: Fast text-to-3d with sparse-view generation and large reconstruction model. arXiv preprint arXiv:2311.06214, 2023. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.689, + 0.905, + 0.744 + ], + "angle": 0, + "content": "[23] Weiyu Li, Rui Chen, Xuelin Chen, and Ping Tan. Sweetdreamer: Aligning geometric priors in 2d diffusion for consistent text-to-3d. International Conference on Learning Representations (ICLR), 2024. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.746, + 0.905, + 0.814 + ], + "angle": 0, + "content": "[24] Weiyu Li, Jiarui Liu, Rui Chen, Yixun Liang, Xuelin Chen, Ping Tan, and Xiaoxiao Long. Craftsman: High-fidelity mesh generation with 3d native generation and interactive geometry refiner. arXiv preprint arXiv:2405.14979, 2024. 1, 3, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.817, + 0.905, + 0.871 + ], + "angle": 0, + "content": "[25] Yixun Liang, Xin Yang, Jiantao Lin, Haodong Li, Xiaogang Xu, and Yingcong Chen. Luciddreamer: Towards high-fidelity text-to-3d generation via interval score matching. arXiv preprint arXiv:2311.11284, 2023. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.874, + 0.905, + 0.901 + ], + "angle": 0, + "content": "[26] Chen-Hsuan Lin, Jun Gao, Luming Tang, Towaki Takikawa, Xiaohui Zeng, Xun Huang, Karsten Kreis, Sanja Fidler," + }, + { + "type": "list", + "bbox": [ + 0.518, + 0.093, + 0.905, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "16259" + } + ], + [ + { + "type": "ref_text", + "bbox": [ + 0.125, + 0.092, + 0.482, + 0.135 + ], + "angle": 0, + "content": "Ming-Yu Liu, and Tsung-Yi Lin. Magic3d: High-resolution text-to-3d content creation. In IEEE Conference on Computer Vision and Pattern Recognition (CVPR), 2023. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.137, + 0.483, + 0.178 + ], + "angle": 0, + "content": "[27] Fangfu Liu, Diankun Wu, Yi Wei, Yongming Rao, and Yueqi Duan. Sherpa3d: Boosting high-fidelity text-to-3d generation via coarse 3d prior, 2023. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.181, + 0.482, + 0.235 + ], + "angle": 0, + "content": "[28] Minghua Liu, Chao Xu, Haian Jin, Linghao Chen, Mukund Varma T, Zexiang Xu, and Hao Su. One-2-3-45: Any single image to 3d mesh in 45 seconds without per-shape optimization. 2024. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.239, + 0.482, + 0.308 + ], + "angle": 0, + "content": "[29] Minghua Liu, Chong Zeng, Xinyue Wei, Ruoxi Shi, Linghao Chen, Chao Xu, Mengqi Zhang, Zhaoning Wang, Xiaoshuai Zhang, Isabella Liu, Hongzhi Wu, and Hao Su. Meshformer: High-quality mesh generation with 3d-guided reconstruction model. arXiv preprint arXiv:2408.10198, 2024. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.311, + 0.482, + 0.367 + ], + "angle": 0, + "content": "[30] Yuan Liu, Cheng Lin, Zijiao Zeng, Xiaoxiao Long, Lingjie Liu, Taku Komura, and Wenping Wang. Syncdreamer: Generating multiview-consistent images from a single-view image. arXiv preprint arXiv:2309.03453, 2023. 1, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.369, + 0.482, + 0.438 + ], + "angle": 0, + "content": "[31] Ben Mildenhall, Pratul P. Srinivasan, Matthew Tancik, Jonathan T. Barron, Ravi Ramamoorthi, and Ren Ng. Nerf: Representing scenes as neural radiance fields for view synthesis. In European Conference on Computer Vision (ECCV), 2020. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.442, + 0.482, + 0.483 + ], + "angle": 0, + "content": "[32] Carsten Moenning and Neil A Dodgson. Fast marching farthest point sampling. Technical report, University of Cambridge, Computer Laboratory, 2003. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.486, + 0.482, + 0.54 + ], + "angle": 0, + "content": "[33] William Peebles and Saining Xie. Scalable diffusion models with transformers. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 4195-4205, 2023. 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.544, + 0.482, + 0.584 + ], + "angle": 0, + "content": "[34] Ben Poole, Ajay Jain, Jonathan T. Barron, and Ben Mildenhall. Dreamfusion: Text-to-3d using 2d diffusion. arXiv, 2022. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.588, + 0.482, + 0.685 + ], + "angle": 0, + "content": "[35] Lingteng Qiu, Guanying Chen, Xiaodong Gu, Qi Zuo, Mutian Xu, Yushuang Wu, Weihao Yuan, Zilong Dong, Liefeng Bo, and Xiaoguang Han. Richdreamer: A generalizable normal-depth diffusion model for detail richness in text-to-3d. In IEEE/CVF Conference on Computer Vision and Pattern Recognition, CVPR 2024, Seattle, WA, USA, June 16-22, 2024, pages 9914-9925. IEEE, 2024. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.688, + 0.482, + 0.757 + ], + "angle": 0, + "content": "[36] Xuanchi Ren, Jiahui Huang, Xiaohui Zeng, Ken Museth, Sanja Fidler, and Francis Williams. Xcube: Large-scale 3d generative modeling using sparse voxel hierarchies. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 4209-4219, 2024. 1, 3, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.76, + 0.482, + 0.828 + ], + "angle": 0, + "content": "[37] Robin Rombach, Andreas Blattmann, Dominik Lorenz, Patrick Esser, and Björn Ommer. High-resolution image synthesis with latent diffusion models. In Conference on Computer Vision and Pattern Recognition (CVPR), pages 10684-10695, 2022. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.832, + 0.482, + 0.899 + ], + "angle": 0, + "content": "[38] Tianchang Shen, Jun Gao, Kangxue Yin, Ming-Yu Liu, and Sanja Fidler. Deep marching tetrahedra: a hybrid representation for high-resolution 3d shape synthesis. Advances in Neural Information Processing Systems, 34:6087-6101, 2021. 3" + }, + { + "type": "list", + "bbox": [ + 0.093, + 0.092, + 0.483, + 0.899 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.092, + 0.905, + 0.134 + ], + "angle": 0, + "content": "[39] Yichun Shi, Peng Wang, Jianglong Ye, Long Mai, Kejie Li, and Xiao Yang. Mvdream: Multi-view diffusion for 3d generation. arXiv:2308.16512, 2023. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.137, + 0.905, + 0.19 + ], + "angle": 0, + "content": "[40] Jiaxiang Tang, Zhaoxi Chen, Xiaokang Chen, Tengfei Wang, Gang Zeng, and Ziwei Liu. Lgm: Large multi-view gaussian model for high-resolution 3d content creation. arXiv preprint arXiv:2402.05054, 2024. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.193, + 0.905, + 0.262 + ], + "angle": 0, + "content": "[41] Jiaxiang Tang, Jiawei Ren, Hang Zhou, Ziwei Liu, and Gang Zeng. Dreamgaussian: Generative gaussian splatting for efficient 3d content creation. In The Twelfth International Conference on Learning Representations, ICLR 2024, Vienna, Austria, May 7-11, 2024. OpenReview.net, 2024. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.264, + 0.905, + 0.331 + ], + "angle": 0, + "content": "[42] Dmitry Tochilkin, David Pankratz, Zexiang Liu, Zixuan Huang, , Adam Letts, Yangguang Li, Ding Liang, Christian Laforte, Varun Jampani, and Yan-Pei Cao. Triposr: Fast 3d object reconstruction from a single image. arXiv preprint arXiv:2403.02151, 2024. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.334, + 0.905, + 0.403 + ], + "angle": 0, + "content": "[43] Peng Wang, Hao Tan, Sai Bi, Yinghao Xu, Fujun Luan, Kalyan Sunkavalli, Wenping Wang, Zexiang Xu, and Kai Zhang. Pf-lrm: Pose-free large reconstruction model for joint pose and shape prediction. arXiv preprint arXiv:2311.12024, 2023. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.406, + 0.905, + 0.473 + ], + "angle": 0, + "content": "[44] Tengfei Wang, Bo Zhang, Ting Zhang, Shuyang Gu, Jianmin Bao, Tadas Baltrusaitis, Jingjing Shen, Dong Chen, Fang Wen, Qifeng Chen, and Baining Guo. Rodin: A generative model for sculpting 3d digital avatars using diffusion, 2022. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.476, + 0.905, + 0.532 + ], + "angle": 0, + "content": "[45] Zhengyi Wang, Cheng Lu, Yikai Wang, Fan Bao, Chongxuan Li, Hang Su, and Jun Zhu. Prolificdreamer: High-fidelity and diverse text-to-3d generation with variational score distillation. arXiv preprint arXiv:2305.16213, 2023. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.533, + 0.905, + 0.601 + ], + "angle": 0, + "content": "[46] Zhengyi Wang, Yikai Wang, Yifei Chen, Chengdong Xiang, Shuo Chen, Dajiang Yu, Chongxuan Li, Hang Su, and Jun Zhu. Crm: Single image to 3d textured mesh with convolutional reconstruction model. arXiv preprint arXiv:2403.05034, 2024. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.604, + 0.905, + 0.699 + ], + "angle": 0, + "content": "[47] Francis Williams, Jiahui Huang, Jonathan Swartz, Gergely Klar, Vijay Thakkar, Matthew Cong, Xuanchi Ren, Rui-long Li, Clement Fuji-Tsang, Sanja Fidler, Eftychios Sifakis, and Ken Museth. fvdb: A deep-learning framework for sparse, large-scale, and high-performance spatial intelligence. ACM Transactions on Graphics (TOG), 43(4):133:1-133:15, 2024. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.703, + 0.905, + 0.757 + ], + "angle": 0, + "content": "[48] Chengzhi Wu, Junwei Zheng, Julius Pfrommer, and Jürgen Beyerer. Attention-based point cloud edge sampling. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), 2023. 2, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.759, + 0.905, + 0.814 + ], + "angle": 0, + "content": "[49] Shuang Wu, Youtian Lin, Feihu Zhang, Yifei Zeng, Jingxi Xu, Philip Torr, Xun Cao, and Yao Yao. Direct3d: Scalable image-to-3d generation via 3d latent diffusion transformer. arXiv:2405.14832, 2024. 1, 2, 3, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.817, + 0.905, + 0.87 + ], + "angle": 0, + "content": "[50] Bojun Xiong, Si-Tong Wei, Xin-Yang Zheng, Yan-Pei Cao, Zhouhui Lian, and Peng-Shuai Wang. OctFusion: Octree-based diffusion models for 3d shape generation. arXiv, 2024. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.873, + 0.905, + 0.901 + ], + "angle": 0, + "content": "[51] Jiale Xu, Weihao Cheng, Yiming Gao, Xintao Wang, Shenghua Gao, and Ying Shan. Instantmesh: Efficient 3d" + }, + { + "type": "list", + "bbox": [ + 0.517, + 0.092, + 0.905, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.52, + 0.957 + ], + "angle": 0, + "content": "16260" + } + ], + [ + { + "type": "ref_text", + "bbox": [ + 0.124, + 0.092, + 0.482, + 0.133 + ], + "angle": 0, + "content": "mesh generation from a single image with sparse-view large reconstruction models. arXiv preprint arXiv:2404.07191, 2024.3" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.135, + 0.483, + 0.177 + ], + "angle": 0, + "content": "[52] Cem Yuksel. Sample elimination for generating poisson disk sample sets. The Eurographs Association & John Wiley & Sons, Ltd., 2015. 3, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.179, + 0.482, + 0.22 + ], + "angle": 0, + "content": "[53] Biao Zhang and Peter Wonka. Lagem: A large geometry model for 3d representation learning and diffusion. arXiv preprint arXiv:2410.01295, 2024. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.222, + 0.482, + 0.276 + ], + "angle": 0, + "content": "[54] Biao Zhang, Jiapeng Tang, Matthias Nießner, and Peter Wonka. 3dshape2vecset: A 3d shape representation for neural fields and generative diffusion models. ACM Transactions on Graphics (SIGGRAPH), 42(4), 2023. 1, 3, 4, 5, 7, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.278, + 0.482, + 0.317 + ], + "angle": 0, + "content": "[55] Lvmin Zhang, Anyi Rao, and Maneesh Agrawala. Adding conditional control to text-to-image diffusion models, 2023. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.321, + 0.482, + 0.389 + ], + "angle": 0, + "content": "[56] Longwen Zhang, Ziyu Wang, Qixuan Zhang, Qiwei Qiu, Anqi Pang, Haoran Jiang, Wei Yang, Lan Xu, and Jingyi Yu. Clay: A controllable large-scale generative model for creating high-quality 3d assets. ACM Transactions on Graphics (TOG), 43(4):1-20, 2024. 1, 2, 3, 5, 7, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.391, + 0.482, + 0.446 + ], + "angle": 0, + "content": "[57] Xuying Zhang, Bo-Wen Yin, Yuming Chen, Zheng Lin, Yunheng Li, Qibin Hou, and Ming-Ming Cheng. Temo: Towards text-driven 3d stylization for multi-object meshes. arXiv preprint arXiv:2312.04248, 2023. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.448, + 0.482, + 0.529 + ], + "angle": 0, + "content": "[58] Zibo Zhao, Wen Liu, Xin Chen, Xianfang Zeng, Rui Wang, Pei Cheng, BIN FU, Tao Chen, Gang YU, and Shenghua Gao. Michelangelo: Conditional 3d shape generation based on shape-image-text aligned latent representation. In Advances in Neural Information Processing Systems (NeurIPS), 2023. 1, 3" + }, + { + "type": "list", + "bbox": [ + 0.093, + 0.092, + 0.483, + 0.529 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.946, + 0.518, + 0.956 + ], + "angle": 0, + "content": "16261" + } + ] +] \ No newline at end of file diff --git a/2025/Dora_ Sampling and Benchmarking for 3D Shape Variational Auto-Encoders/6b478709-2cef-4587-8775-92fc50ace470_origin.pdf b/2025/Dora_ Sampling and Benchmarking for 3D Shape Variational Auto-Encoders/6b478709-2cef-4587-8775-92fc50ace470_origin.pdf new file mode 100644 index 0000000000000000000000000000000000000000..5ff5585da9fd1489ac5d932ad3c743adbcc18c22 --- /dev/null +++ b/2025/Dora_ Sampling and Benchmarking for 3D Shape Variational Auto-Encoders/6b478709-2cef-4587-8775-92fc50ace470_origin.pdf @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:d05a39301525e6a0a4c4f45d7c66539fe05eaa2bf859f661ac54b126e3881459 +size 9045128 diff --git a/2025/Dora_ Sampling and Benchmarking for 3D Shape Variational Auto-Encoders/full.md b/2025/Dora_ Sampling and Benchmarking for 3D Shape Variational Auto-Encoders/full.md new file mode 100644 index 0000000000000000000000000000000000000000..4398a1523671378385879a80b8e326c6b182924c --- /dev/null +++ b/2025/Dora_ Sampling and Benchmarking for 3D Shape Variational Auto-Encoders/full.md @@ -0,0 +1,387 @@ +# Dora: Sampling and Benchmarking for 3D Shape Variational Auto-Encoders + +Rui Chen $^{1,2}$ Jianfeng Zhang $^{2\dagger}$ Yixun Liang $^{1,3}$ Guan Luo $^{2,4}$ Weiyu Li $^{1,3}$ +Jiarui Liu $^{1,3}$ Xiu Li $^{2}$ Xiaoxiao Long $^{1,3}$ Jiashi Feng $^{2}$ Ping Tan $^{1,3\dagger}$ +†Corresponding authors + $^{1}$ The Hong Kong University of Science and Technology $^{2}$ ByteDance Seed + $^{3}$ LightIllusions $^{4}$ Tsinghua University +https://aruichen.github.io/Dora + +# Abstract + +Recent 3D content generation pipelines commonly employ Variational Autoencoders (VAEs) to encode shapes into compact latent representations for diffusion-based generation. However, the widely adopted uniform point sampling strategy in Shape VAE training often leads to a significant loss of geometric details, limiting the quality of shape reconstruction and downstream generation tasks. We present Dora-VAE, a novel approach that enhances VAE reconstruction through our proposed sharp edge sampling strategy and a dual cross-attention mechanism. By identifying and prioritizing regions with high geometric complexity during training, our method significantly improves the preservation of fine-grained shape features. Such sampling strategy and the dual attention mechanism enable the VAE to focus on crucial geometric details that are typically missed by uniform sampling approaches. To systematically evaluate VAE reconstruction quality, we additionally propose Dora-bench, a benchmark that quantifies shape complexity through the density of sharp edges, introducing a new metric focused on reconstruction accuracy at these salient geometric features. Extensive experiments on the Dora-bench demonstrate that Dora-VAE achieves comparable reconstruction quality to the state-of-the-art dense XCube-VAE while requiring a latent space at least $8 \times$ smaller (1,280 vs. $>10,000$ codes). Project page: https://aruichen.github.io/Dora. + +# 1. Introduction + +3D content creation is vital to delivering realistic and immersive experiences in various industries, including games, movies, and AR/VR. However, traditional 3D modeling typically demands significant expertise and manual effort, making it time-consuming and challenging, especially for non-expert users. Recent advances in AI-powered 3D con + +tent generation methods [2, 3, 7, 9, 17, 21, 25, 28, 30, 42, 44, 57] have transformed the field, making it more accessible to many more users. + +Following the success of text-to-image generation models [4, 5, 15, 37, 55], recent 3D content creation approaches [36, 49, 56] adopt a two-stage pipeline: encoding 3D shapes into a latent space using variational autoencoders (VAEs), followed by training a latent diffusion model. The performance of such a generative pipeline heavily relies on the VAE's capability to faithfully encode and reconstruct 3D shapes. + +Existing 3D VAEs operate by sampling points on mesh surfaces for shape encoding and then reconstructing the original 3D meshes by its decoder. This process faces unique challenges compared to 2D image VAEs where the input image is fully observable. In comparison, the sampled point cloud often cannot capture all the necessary shape information, which could harm the performance of 3D VAEs. + +Volume-based method [36] leverages sparse convolution [47] to process millions of voxelized points for high-fidelity reconstruction. Its dense sampling captures precise shape information. However, this method produces large latent codes (commonly $>10,000$ tokens), which significantly complicate the training of diffusion models. On the other hand, vector-set (Vecset) methods [24, 54, 56, 58] use transformers to achieve compact latent representations (hundreds to thousands of tokens), enabling efficient diffusion [24, 56]. However, due to the quadratic complexity of transformer networks, it often only samples a few thousand points to represent a 3D shape, which leads to information loss and performance degradation. Therefore, we seek to improve the reconstruction quality of Vecset-based VAEs and maintain their compact representation. + +We begin by analyzing the shape reconstruction capability of Vecset-based VAEs. Through careful analysis, we find these methods have limited reconstruction performance, which stems from their commonly used uniform sampling. When computational constraints limit the total + +![](images/1d7091c2be5c9b7d56c50afe9c7af043c387fd6b81d9f0d95246d92e7a0806c3.jpg) +(a) Ground Truth(GT) + +![](images/3723c8ddf5058b50cb5b8b944adf45a354d3034414836c04f599aac1b01dbed1.jpg) +(d) Reconstruction accuracy +Figure 1. Sampling strategy comparison. Given the ground truth mesh shown in (a), we visualize point clouds produced by uniform sampling in (b) and those generated by our proposed Sharp Edge Sampling (SES) in (c), at various sampling rates. In (d), we compare the reconstruction accuracy trained with our SES and the uniform sampling using the F-score metric. The comparison demonstrates SES consistently outperforms uniform sampling under varying sampling rates, as the point clouds generated by SES are more effective in capturing the salient features of the object. + +point number + +4096 + +8192 + +16384 + +20480 + +40960 + +![](images/61b761c8c0524fdbefc7fee85173068b8de6ee8ecbadde13ff5b5c01b104cf6f.jpg) + +![](images/3791fd82f0a53b4151db302506296faedd3f58194ee58670b9ea10fce597f1a0.jpg) +(b) Visual results of the uniform sampling + +![](images/718df31d5a83af4156943caa575e9190d455c9c0fc75e022af8f1e707517dd3e.jpg) + +![](images/d194ca7c9733c32f31613aae2706321ba6bdb020abe917ab69ac1f73fb0c1b16.jpg) + +![](images/c8b578ccaeb00d7c59a2bb47d98b19908a0c59fa44b82f4a0efdd5cab138d2e2.jpg) + +![](images/10420fe341601ac28b74e94a6b7974bfb0b6d3ebb3f785d8eac49363d427b2be.jpg) + +![](images/0e5defacb8d30c2a7f52511540f4ac6ba76050768d5784e1bbb978b333d5414e.jpg) +(c) Visual results of our sharp edge sampling (SES) + +![](images/18a1a5cb08e34fdaad47c637593d26cf9f5228855c12451dd979ef4bfc1f4d50.jpg) + +![](images/cb2c82564119ae842784851ef1448d769987a3ad74a4674936b0276f111ec738.jpg) + +![](images/da169309bca4faf0378c628b5eaf7703bd0c38df8c5f59d04f8753c253e7449c.jpg) + +number of sampled points, uniform sampling fails to prioritize geometrically salient regions, leading to the loss of fine details. To validate this observation, we experiment on a 3D mesh with intricate geometric details (e.g., keyboard buttons) shown in Figure 1 (a). We visualize the point cloud with different sampling strategies at various sampling densities in (b) and (c). As demonstrated in Figure 1 (b), even with increasing sampling rates, uniform sampling fails to preserve sharp features like keyboard buttons. This simple experiment confirms that uniform sampling fundamentally limits the capturing of geometric details, which in turn affects the VAE's reconstruction capability and the result quality of the learned diffusion models. + +Inspired by the success of importance sampling in various geometric processing tasks [14, 48], we introduce a similar strategy for 3D VAE training. While existing importance sampling methods focus on down-sampling point clouds, our task requires sampling points directly from mesh surfaces for shape VAE training [49, 56]. This fundamental difference necessitates a new sampling approach specifically designed for preserving geometric features from mesh representations. + +To address this challenge, we propose a Sharp Edge Sampling (SES) algorithm that adaptively samples points based on geometric saliency. Specifically, SES first identifies edges with significant dihedral angles on the mesh, indicating regions of high geometric complexity. It then samples points along these salient regions while maintaining a balance with uniformly sampled points to capture the over + +all structure. This approach ensures comprehensive coverage of both fine details and global geometry. + +Building upon SES, we present Dora-VAE, a novel method achieving high-fidelity reconstruction while maintaining compact latent representations. To fully leverage these detail-rich point clouds sampled by our SES, we design a dual cross-attention architecture that effectively processes both salient and uniform regions during encoding. As shown in Figure 1, our method significantly outperforms uniform sampling in preserving shape details on the keyboard (c), with consistent improvements in F-Score across different sampling rates (d). + +The common evaluation protocol for 3D VAEs is also biased. It typically uses a set of randomly selected 3D shapes, and employs general metrics (e.g., F-score, Chamfer distance) to measure the shape reconstruction quality. However, we argue it is necessary to divide the test shapes into sub-classes of different shape complexity to better evaluate these 3D VAEs. To facilitate the evaluation of 3D VAEs, we further introduce Dora-Bench, a new benchmark based on existing datasets with our novel Sharp Normal Error (SNE) metric. Dora-Bench categorizes test shapes based on their geometric complexity and SNE focuses on measuring the reconstruction quality of salient geometric features. This combination enables a more rigorous assessment of 3D VAEs. + +Extensive experiments demonstrate that our Dora-VAE achieves superior results. When integrated into downstream 3D diffusion models, it significantly enhances the qual + +ity of generated 3D shapes, which validates that our novel sampling strategy and dual cross-attention architecture effectively preserve geometric details with compact latent spaces. To summarize, our main contributions include: 1) We propose Dora-VAE, a novel 3D VAE model for high-quality reconstruction with compact latent representations, accompanied by Dora-Bench, a comprehensive benchmark for evaluating 3D VAEs. 2) We introduce, for the first time, importance sampling to the task of 3D VAE learning and propose a Sharp Edge Sampling (SES) algorithm to prioritize geometrically salient regions. Building on SES, we design a dual cross-attention architecture to effectively encode these detail-rich point clouds. 3) We develop a systematic evaluation benchmark based on existing datasets with our novel Sharp Normal Error (SNE) metric that specifically assesses reconstruction accuracy of fine geometric details, enabling more rigorous evaluation than conventional random sampling approaches. + +# 2. Related work + +Importance Sampling in Point Clouds. Importance sampling techniques have been widely used in point cloud processing tasks [14, 48]. For instance, APES [48] proposes attention-based sampling for point cloud classification and segmentation. However, these methods operate directly on point clouds rather than meshes, making them less suitable for VAE-based shape representation where preserving complete geometric information is crucial. + +3D Shape VAEs. Recent 3D shape VAEs follow two main approaches: volume-based and vector set-based. Volume-based methods [36, 50] like XCube [36] use sparse convolution to encode voxelized surfaces, achieving high reconstruction quality but requiring large latent codes ( $>10,000$ tokens). While these methods excel at preserving geometric details, their large latent spaces pose significant challenges for downstream diffusion model training. In contrast, vector set-based approaches [24, 49, 54, 56, 58] encode uniformly sampled surface points using transformers, producing highly compact latent spaces that are particularly suitable for diffusion models. However, these methods often struggle with geometric detail preservation, especially in regions with complex surface features. Our analysis reveals that this limitation primarily stems from their uniform sampling strategy: when computational constraints restrict the total number of processable points, uniform sampling fails to prioritize geometrically significant regions, leading to insufficient capture of fine details. This information loss at the sampling stage fundamentally limits these methods' ability to learn and preserve intricate geometric features. + +3D Content Creation. Current 3D generation methods can be categorized into three groups. Optimization-based methods [6, 23, 26, 27, 30, 34, 35, 39, 41, 45], pioneered by DreamFusion [34] utilize score distillation sam + +pling (SDS) to optimize 3D representations [19, 31, 38] using 2D diffusion model priors. While these methods can achieve photorealistic results, they suffer from slow generation speed, training instability, and often struggle to maintain geometric consistency. Large reconstruction models, like LRM [16] and follow-up works [22, 29, 40, 43, 46, 51] employ large-scale sparse-view reconstruction for efficient 3D generation. However, their lack of explicit geometric priors often leads to compromised geometric fidelity and inconsistent surface details. 3D native generative models [24, 49, 54, 56, 58], represented by 3DShape2VecSet [54], adopt a two-stage approach: first training a 3D VAE to encode shapes into latent space, then training a conditional latent diffusion model for generation. This approach ensures better geometric consistency through the VAE's built-in geometric constraints. However, the quality of generated shapes is fundamentally limited by the VAE's reconstruction capability. Recent works [24, 53] have shown that improving VAE reconstruction directly enhances downstream generation quality, which motivates our focus on advancing VAE design. + +# 3. Method + +In this section, we present Dora-VAE for high-quality 3D reconstruction, and Dora-Bench for 3D VAE evaluation. We first briefly review 3DShape2VecSet [54] in Section 3.1, the foundation of our method and then detail our key innovations in Section 3.2 and Section 3.3. + +# 3.1. Preliminary: 3DShape2VecSet + +3DShape2VecSet [54] introduces a transformer-based 3D VAE that encodes uniformly sampled surface points into compact latent codes. Given a 3D surface S, their pipeline consists of three key steps: + +- Surface Sampling: Uniformly sample $N_{d}$ points on the surface S using Poisson disk sampling [52] to obtain a dense point cloud $P_{d}$ , then downsample it to $N_{s}$ points via Farthest Point Sampling (FPS) [32] to get a sparse point cloud $P_{s}$ : + +$$ +P _ {d} = \left\{p _ {d} ^ {i} \in S \mid i = 1, \dots , N _ {d} \right\}, P _ {s} = \operatorname {F P S} \left(P _ {d}, N _ {s}\right). \tag {1} +$$ + +- Feature Encoding: Compute the point cloud feature $C$ via the cross-attention between $P_{s}$ and $P_{d}$ , followed by some self-attention layers to generate the latent code $z$ : + +$$ +C = \operatorname {C r o s s A t t n} \left(P _ {s}, P _ {d}, P _ {d}\right), z = \operatorname {S e l f A t t n} (C). \tag {2} +$$ + +- Geometry Decoding: Further decode $z$ through selfattention layers and predict occupancy values using randomly sampled spatial query points $Q_{space} \in \mathbb{R}^3$ : + +$$ +\hat {O} = \operatorname {C r o s s A t t n} \left(Q _ {\text {s p a c e}}, \operatorname {S e l f A t t n} (z)\right). \tag {3} +$$ + +![](images/3086c3de39cf007f9521440ba59a4bf290529273b569fcbfb1d3a64390923a11.jpg) + +![](images/6260bff6b48d558afdc9f3e2d77308aca2788d2a8507c69e764d197bfc799384.jpg) +(a) Sharp edge sampling +Figure 2. Overview of Dora-VAE. (a) We utilize the proposed sharp edge sampling technique to extract both salient and uniform points from the input mesh. These points are then combined with dense points, effectively capturing both salient regions and smooth areas. (b) To enhance the encoding of point clouds sampled through sharp edge sampling, we design a dual cross-attention architecture. + +While this method generates compact latent codes for diffusion, the uniform sampling limits its ability to capture fine geometric details. Our work addresses this limitation by carefully designed sampling and encoding strategies. + +# 3.2. Dora-VAE + +Figure 2 gives an overview of our pipeline. For each input mesh, we augment the uniformly sampled point cloud $P_{u}$ with more important points $P_{a}$ sampled by our proposed sharp edge sampling strategy, which forms the dense point cloud $P_{d}$ . During the encoding process, we compute the attention for $P_{u}$ and $P_{a}$ separately via a simple-yet-effective dual cross attention mechanism and sum the results for self-attention to compute the latent code $z$ . Our VAE training largely follows that of 3DShape2VecSet [54], which is supervised by a loss evaluated on the occupancy field. + +# 3.2.1. Sharp Edge Sampling (SES) + +We propose SES to effectively sample points from geometrically salient regions. To ensure surface coverage, we also sample points uniformly. Our final sampled point cloud $P_{d}$ combines uniformly sampled points $P_{u}$ with points specifically sampled from salient regions $P_{a}$ as $P_{d} = P_{u} \cup P_{a}$ . Our method computes salient points $P_{a}$ through two steps: detecting salient edges and sampling points from these regions. + +Salient Edges Detection. Given a triangular mesh, we identify a set of salient edges $\Gamma$ by analyzing dihedral angles between adjacent faces, which calculates the angle be + +tween the normal vectors of adjacent faces, providing a direct measure of surface curvature at mesh edges. For each edge $e$ shared by adjacent faces $f_{1}$ and $f_{2}$ , we compute the dihedral angle $\theta_{e}$ as: + +$$ +\theta_ {e} = \operatorname {a r c c o s} \left(\frac {\mathbf {n} _ {f _ {1}} \cdot \mathbf {n} _ {f _ {2}}}{\| \mathbf {n} _ {f _ {1}} \| \| \mathbf {n} _ {f _ {2}} \|}\right), \tag {4} +$$ + +where $\mathbf{n}_{f_1}$ and $\mathbf{n}_{f_2}$ are the normals of $f_{1}$ and $f_{2}$ . The salient edge set $\Gamma$ contains all edges with a dihedral angle exceeding a predefined threshold $\tau$ : + +$$ +\Gamma = \left\{e \mid \theta_ {e} > \tau \right\} \tag {5} +$$ + +Let $N_{\Gamma} = |\Gamma|$ represent the number of the salient edges. + +Salient Points Sampling. For each salient edge $e \in \Gamma$ , we collect its two vertices $v_{e,1}$ and $v_{e,2}$ into a salient vertex set $P_{\Gamma}$ : + +$$ +P _ {\Gamma} = \left\{v _ {e, 1}, v _ {e, 2} \mid e \in \Gamma \right\}, \tag {6} +$$ + +where duplicate vertices from connecting edges are included only once. Let $N_V = |P_{\Gamma}|$ denote the number of unique vertices in $P_{\Gamma}$ . + +Given a target number of salient points $N_{\mathrm{desired}}$ , we generate the salient point set $P_{a}$ based on the available salient vertices: + +$$ +P _ {a} = \left\{ \begin{array}{l l} \operatorname {F P S} \left(P _ {\Gamma}, N _ {\text {d e s i r e d}}\right), & \text {i f} N _ {\text {d e s i r e d}} < = N _ {V}, \\ P _ {\Gamma} \cup P _ {\text {i n t e r p o l a t e d}}, & \text {i f} 0 < N _ {V} < N _ {\text {d e s i r e d}}, \\ \emptyset , & \text {i f} N _ {V} = 0. \end{array} \right. \tag {7} +$$ + +When we have excess salient vertices ( $N_{\text{desired}} \leq N_V$ ), we use FPS to downsample $P_{\Gamma}$ to obtain $P_a$ . For cases with insufficient salient vertices ( $N_V < N_{\text{desired}}$ ), we include all vertices from $P_{\Gamma}$ and supplement with additional points $P_{\text{interpolated}}$ . These additional points are generated by uniformly sampling $(N_{\text{desired}} - N_V) / N_{\Gamma}$ points along each salient edge in $\Gamma$ , ensuring comprehensive coverage of salient features. When no salient edges are detected ( $N_V = 0$ ), $P_a$ remains empty. + +# 3.2.2. Dual Cross Attention + +Given the point clouds $P_{d}$ produced by our SES strategy, we design a dual cross-attention architecture to effectively encode both uniform and salient regions. Following 3DShape2VecSet [54], we first downsample $P_{u}$ and $P_{a}$ separately using FPS: + +$$ +P _ {s} = \operatorname {F P S} \left(P _ {u}, N _ {s, 1}\right) \cup \operatorname {F P S} \left(P _ {a}, N _ {s, 2}\right), \tag {8} +$$ + +where $N_{s,1}$ and $N_{s,2}$ is the number of downsampled point clouds from $P_u$ and $P_a$ , respectively. We then compute cross-attention features separately for uniform and salient points as follows: + +$$ +C _ {u} = \operatorname {C r o s s A t t n} \left(P _ {s}, P _ {u}, P _ {u}\right) \tag {9} +$$ + +$$ +C _ {a} = \operatorname {C r o s s A t t n} \left(P _ {s}, P _ {a}, P _ {a}\right) \tag {10} +$$ + +![](images/762d6fdac16e892ef128d95b0fbfb7f3d7afc4834a9dd2ac14d2d5ea6fad7bdf.jpg) +(a) Counts of Different Datasets Across Detail Levels + +![](images/0e35f9650b8f748045dad41a39219dfa06a62f94199fe9252f06feb1b9d6333d.jpg) +(b) Total Counts by Detail Levels + +![](images/0426c7405dd740c90ea7c6ed7b3f5db4a0dca48fa80ee10cf5f8aa39af21ef5e.jpg) +(c) Visualization of detail levels +Figure 3. Our proposed benchmark include 3D shapes from the ABO [10], GSO [13], Meta [1], and Objaverse [12] datasets. (a) The histogram of different datasets across different shape complexities. (b) The pie chart of the total counts by shape complexities. (c) Sample shapes of different shape complexities. + +![](images/9a18b1c24ae4dcfc95938048cfe55d2e56a7e6b0c465994029eeab5eccc3b6fe.jpg) +Figure 4. The process of computing sharp normal errors (SNE). We compute MSE loss in the sharp regions of the normal. + +The final point cloud feature $C$ combines both attention results: + +$$ +C = C _ {u} + C _ {a}. \tag {11} +$$ + +This dual attention design enables separate focus on uniform and salient regions during feature extraction. Following 3DShape2VecSet [54], we use $C$ to predict the occupancy field $\hat{O}$ through self-attention blocks. The whole model with parameters $\psi$ are optimized using MSE loss: + +$$ +\nabla_ {\psi} \mathcal {L} _ {\mathrm {M S E}} (\hat {O}, O) = \mathbb {E} \left[ 2 (\hat {O} - O) \frac {\partial \hat {O}}{\partial \psi} \right]. \tag {12} +$$ + +# 3.3. Dora-Bench + +# 3.3.1. Geometric Complexity-based Evaluation + +To enable a more rigorous evaluation of VAE performance, we propose Dora-bench, a benchmark that systematically categorizes test shapes based on their geometric complexity. Unlike previous methods that use randomly selected test sets, we measure shape complexity using the number of salient edges $N_{\Gamma}$ (Section 3.2.1) and classify shapes into four levels: + +- Level 1 (Less Detail): $0 < N_{\Gamma} \leq 5000$ ; +- Level 2 (Moderate Detail): $5000 < N_{\Gamma} \leq 10000$ ; +- Level 3 (Rich Detail): $10000 < N_{\Gamma} \leq 50000$ ; +- Level 4 (Very Rich Detail): $N_{\Gamma} > 50000$ + +We curate test shapes from multiple public datasets including GSO [13], ABO [10], Meta [1], and Objaverse [12] to ensure diverse geometric complexities. Figure 3 shows the distribution of shapes across complexity levels (a,b) and example meshes from each level (c). Please refer to our supplementary materials for more examples. + +# 3.3.2. Sharp Normal Error (SNE) + +Building on our Dora-bench, we further introduce Sharp Normal Error (SNE) to evaluate reconstruction quality in salient regions. While existing metrics like Chamfer Distance and F-Score capture overall shape similarity, they fail to specifically assess the preservation of fine geometric details. SNE addresses this limitation by measuring normal map differences between reconstructed and ground truth shapes in geometrically significant areas. As illustrated in Figure 4, we render normal maps of the ground truth shape from multiple viewpoints and identify salient regions using Canny edge detection. These regions are dilated to create evaluation masks. The final SNE metric is computed as the Mean Squared Error between ground truth and reconstructed normal maps within the masked areas. This process enables focused evaluation of how well VAEs preserve sharp geometric features during reconstruction. + +# 4. Experiments + +We conducted intensive experiments to validate the effectiveness of our proposed 3D VAE and compare it with other state-of-the-art methods. + +# 4.1. Implementation Details + +We follow CLAY [56] for mesh preprocessing to ensure watertight 3D models. Our VAE is trained on a subset filtered from Objaverse [12], containing approximately 400,000 3D meshes. We filter out low-quality meshes with missing faces or severe self-intersections to ensure training stability. Our training is conducted on 32 A100 GPUs for two days using a batch size of 2048 and a learning rate of 5e-5. We employ Flash-Attention-v2 [11], mixed-precision training with FP16 and gradient checkpointing [8] to optimize + +![](images/c29adb258d92ddda61c839c850298bfd3c87a122d627da02089f38748083b368.jpg) +Figure 5. Qualitative comparison of the VAE reconstruction results. $\dagger$ indicates the fine-tuning model that uses the same training data as ours. + +
MethodsLCL↑ F-score(0.01) × 100↑ F-score(0.005) × 100↓ CD × 10000↓ SNE × 100
L1L2L3L4L1L2L3L4L1L2L3L4L1L2L3L4
Xcube [36]>1000098.96898.79998.61598.22695.52593.87292.32285.3656.3156.2887.9359.9261.5791.4321.4301.679
Xcube† [36]>1000099.39399.79499.82499.07996.75395.53593.42287.3654.0154.1425.7407.6271.5431.4081.2591.639
Craftsman [24]25698.01695.87491.75681.73987.99482.54973.00057.3794.3899.12914.53033.4411.9061.8732.1913.933
Ours w/o DCA128099.96499.92599.67897.89096.56195.97591.61883.1242.2362.5064.4446.4321.4481.2151.2051.828
Ours w/o SES,DCA128099.94499.81497.29496.77995.97794.62388.40679.2402.4222.9833.9806.1961.4961.3131.3522.207
Ours full25699.50798.98696.66989.57793.27290.46682.38668.6693.3565.20210.27624.5271.5551.4101.6183.035
128099.98899.95599.88099.17097.03896.83193.45887.4732.0972.5003.9455.2651.4331.1861.1371.579
+ +Table 1. Quantitative comparison in Dora-bench. ${}^{ \dagger }$ indicates the fine-tuning model that uses the same training data as ours. + +memory usage and training efficiency. For the parameters in sharp edge sampling, we set $N_{\mathrm{desired}} = 16384$ and $\tau = 30$ . We set the low threshold to 20 and the high threshold to 200 for the canny edge detection. + +# 4.2. Evaluation Setting + +Metrics. We evaluate reconstruction quality by comparing input meshes with their decoded counterparts from different 3D VAEs using 1M sampled points under three metrics: 1) F-score $(r)$ [20], which reconstruction accuracy by computing precision and recall of point correspondences within distance threshold $r$ . Specifically, we report F-score (0.01) and F-score (0.005) with shapes normalized $[-1,1]$ . 2) Chamfer Distance (CD), which computes the average distance between each reconstructed point and its nearest ground truth point. 3) Sharp Normal Errors (SNE) as proposed in Section 3.3.2, which evaluate normal map differences in salient regions. For fair comparison, we also report Latent Code Length (LCL) as longer codes typically enable better reconstruction. + +Baselines. We compare Dora-VAE with state-of-the-art approaches, including: 1) XCube-VAE [36], a volumetric method with larger latent codes; 2) XCube-VAE† [36], our fine-tuned version of the original XCube-VAE on the same dataset; 3) Craftsman-VAE[24], which fine-tune the 3DShape2VecSet [54] with shorter latent codes on Objverse. We exclude VAE models from Direct3D [49] and CLAY[56] as as their implementations were not publicly available at submission time. + +# 4.3. Qualitative Comparison + +Figure 5 shows visual comparisons of different methods across different complexity levels from our Dora-bench dataset. We visualize both ground truth and reconstructed meshes using surface normal coloring to highlight geometric details. For shapes with lower complexity (L1 and L2), all methods achieve comparable reconstruction quality. However, when dealing with shapes of higher complexity (L3 and L4), the advantages of our method become obvious. While XCube-VAE achieves similar visual quality + +to ours, it requires a significantly larger latent space - more than $8 \times$ the size of ours ( $>10,000$ dimensions vs. 1,280). This substantial reduction in latent code length, while maintaining high reconstruction fidelity, makes our method particularly suitable for training 3D diffusion models. In contrast, Craftsman-VAE shows a noticeable degradation in reconstruction quality for complex shapes, failing to capture fine geometric details. Additional visual comparisons are provided in the appendix. + +# 4.4. Quantitative Comparison + +Table 1 presents quantitative results of different methods across different complexity levels of Dora-bench. Our method consistently outperforms baselines across all levels, with larger margins on more complex shapes (L3 and L4). The advantage is particularly evident in CD metrics, where our method with only 256 latent codes surpasses even our fine-tuned version of XCube-VAE (3.356 vs. 4.015). When using 1280 latent codes, our method further decreases CD to 2.097, achieving a $47.77\%$ improvement over XCube-VAE†. We attribute XCube-VAE's lower performance to its use of NKSR [18] for mesh extraction, which introduces additional quantization errors. + +Notably, our method demonstrates superior performance in preserving geometric details, as reflected by the SNE metric. For example, in L4 shapes where geometric complexity is highest, our method achieves an SNE of 1.579 compared to 1.639 from XCube-VAE†, representing a $3.7\%$ improvement. This significant gain in SNE aligns with our qualitative observations in Figure 5, where our method better preserves fine details such as sharp edges and complex surface variations, demonstrating the effectiveness of our sharp edge sampling strategy. + +# 4.5. Ablation Studies + +To evaluate the contribution of each component, we compare our full model with two variants under the same training conditions: + +- Ours w/o SES, DCA. This variant removes both sharp edge sampling (SES) and dual cross attention (DCA), i.e. + +![](images/a725a621e317993e35a9a1c648a7adc36716ce5021113677bdd3e1b0e6ac6125.jpg) +GT + +![](images/5fe5ec243e59f522a4a135eaa1a908510eb65606e4581824486aa11e830c8ea2.jpg) +Full +w/o DAC + +![](images/45c477727be1322becd3da2860816f627377b230cfe97a126222cb7586a5f408.jpg) +w/o SES, DCA +Figure 6. Ablation studies of our method. Given the ground truth of mesh, we employ both our full model and its variations to reconstruct the ground truth mesh, highlighting significant reconstruction discrepancies with red boxes. + +![](images/92f385a5f1b1973336a224a2a8c63ee85ae383a0f4a3a91f2532a6ec85601df8.jpg) + +using only uniformly sampled point clouds with Poisson disk sampling [52], while maintaining an equal $N_{d}$ . + +- Ours w/o DCA. This variant retains SES but removes DCA, i.e. using a single cross attention adopted by [54]. As shown in Figure 6 and Table 1, our full model consistently outperforms these variants, validating the effectiveness of both components. + +# 5. Application: Single Image to 3D + +We demonstrate the effectiveness of our VAE by applying it to single-image 3D generation through diffusion models. Following CLAY [56], we implement a latent diffusion model based on the DiT [33] architecture. For a fair comparison, we fine-tune Craftsman-VAE on our dataset (denoted as Craftsman-VAE†) since both Craftsman-VAE and 3DShape2VecSet were originally trained on smaller datasets. Note that XCube-VAE is excluded from comparison due to its 10,000-dimensional latent codes being impractical for diffusion model training. Figure 7 shows some generation results from diffusion models trained with our Dora-VAE and Craftsman-VAE†. Both models share identical architecture (0.39B parameters) and training conditions (same dataset, 32 A100 GPUs, 3 days). Our Dora-VAE demonstrates significantly better preservation of geometric details in the generated shapes, validating its effectiveness as a foundation for 3D generation tasks. While further improvements could be achieved with more extensive training data and computational resources, we focus on validating the VAE's capabilities in this work and leave such exten + +![](images/d526132c9f29f10e84fa0ffa2c3e54ffed14ca3f54d575be6896746582f43c94.jpg) +Input Image + +![](images/a481d437a55e51a85aa0457a52c71c839638cd48e407d959e40364ce87d3b6ce.jpg) +Ours + +![](images/670442b42bbf9d85c9ecec5f31388d75e27fae45204cd0d6eab806a79ee4b776.jpg) +Craftsman + +![](images/d17e93218df0ade95f0621be106e9dc8d1e6304447621c1f07fe0389f4deb311.jpg) + +![](images/e15693c532de6a739e4452ddd77ea78328d3dc7ad71ea47c61cdc43a4948231a.jpg) + +![](images/09f313a12a61443cfb6b32a96dbaf984a6564a0029c8c124a1ef86939b10d737.jpg) + +![](images/d1d0e13606feabfc15e3f1121b3ded86c314dba301781fed07a821d9a34517a2.jpg) +Figure 7. The diffusion results of the single image to 3D generation trained on our Dora-VAE and Craftsman†. The 3D geometry generated by the diffusion model trained on our proposed Dora-VAE has more details under the same experimental environment. + +![](images/ea2c630b92887b7c8376bc23c0c84accb0065c03c74bbd2df2e63347c830857a.jpg) + +![](images/f85fb20808f7c52a8c63c90913f1a53d230dcbbd22d4180cf275b7ce2090701a.jpg) + +sions for future work. + +# 6. Conclusion + +In this work, we introduce Dora-VAE, a novel VAE designed for high-quality 3D shape compression and reconstruction. At its core, Dora-VAE introduces sharp edge sampling to effectively capture salient geometric features, complemented by a dual cross-attention architecture that enhances the encoding of these detail-rich point clouds. To enable more rigorous evaluation of VAE performance, we develop Dora-bench, which systematically categorizes shapes based on geometric complexity and introduces the Sharp Normal Error (SNE) metric for specifically assessing the preservation of fine geometric details. Our comprehensive experiments demonstrate that Dora-VAE significantly outperforms existing methods across varying levels of shape complexity. Furthermore, we show that the improved reconstruction capability of Dora-VAE directly enhances the quality of downstream tasks by applying it to single-image 3D generation. The superior performance in generating geometric details validates our approach of focusing on salient region sampling and encoding for 3D VAE design. + +Acknowledgements. This work is partially supported by the project L0751 between Bytedance and HKUST. + +# References + +[1] Digital twin catalog. META, 2024. https://www.projectaria.com/datasets/dtc/.5 +[2] Yukang Cao, Yan-Pei Cao, Kai Han, Ying Shan, and KwanYee K. Wong. Dreamavatar: Text-and-shape guided 3d human avatar generation via diffusion models. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 958-968, 2024. 1 +[3] Yukang Cao, Liang Pan, Kai Han, Kwan-Yee K Wong, and Ziwei Liu. Avatargo: Zero-shot 4d human-object interaction generation and animation. arXiv preprint arXiv:2410.07164, 2024. 1 +[4] Junsong Chen, Jincheng Yu, Chongjian Ge, Lewei Yao, Enze Xie, Yue Wu, Zhongdao Wang, James Kwok, Ping Luo, Huchuan Lu, and Zhenguo Li. Pixart-α: Fast training of diffusion transformer for photorealistic text-to-image synthesis, 2023. 1 +[5] Junsong Chen, Yue Wu, Simian Luo, Enze Xie, Sayak Paul, Ping Luo, Hang Zhao, and Zhenguo Li. Pixart-δ: Fast and controllable image generation with latent consistency models, 2024. 1 +[6] Rui Chen, Yongwei Chen, Ningxin Jiao, and Kui Jia. Fantasia3d: Disentangling geometry and appearance for high-quality text-to-3d content creation. In Proceedings of the IEEE/CVF International Conference on Computer Vision (ICCV), 2023. 3 +[7] Rui Chen, Mingyi Shi, Shaoli Huang, Ping Tan, Taku Komura, and Xuelin Chen. Taming diffusion probabilistic models for character control. In ACM SIGGRAPH 2024 Conference Papers, New York, NY, USA, 2024. Association for Computing Machinery. 1 +[8] Tianqi Chen, Bing Xu, Chiyuan Zhang, and Carlos Guestrin. Training deep nets with sublinear memory cost. CoRR, abs/1604.06174, 2016. 5 +[9] Yongwei Chen, Rui Chen, Jiabao Lei, Yabin Zhang, and Kui Jia. Tango: Text-driven photorealistic and robust 3d stylization via lighting decomposition. Advances in Neural Information Processing Systems, 35:30923-30936, 2022. 1 +[10] Jasmine Collins, Shubham Goel, Kenan Deng, Achleshwar Luthra, Leon Xu, Erhan Gundogdu, Xi Zhang, Tomas F Yago Vicente, Thomas Dideriksen, Himanshu Arora, et al. Abo: Dataset and benchmarks for real-world 3d object understanding. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 21126-21136, 2022. 5 +[11] Tri Dao. Flashattention-2: Faster attention with better parallelism and work partitioning. In The Twelfth International Conference on Learning Representations, ICLR 2024, Vienna, Austria, May 7-11, 2024. OpenReview.net, 2024. 5 +[12] Matt Deitke, Dustin Schwenk, Jordi Salvador, Luca Weihs, Oscar Michel, Eli VanderBilt, Ludwig Schmidt, Kiana Ehsani, Aniruddha Kembhavi, and Ali Farhadi. Objverse: A universe of annotated 3d objects. arXiv preprint arXiv:2212.08051, 2022. 5 +[13] Laura Downs, Anthony Francis, Nate Koenig, Brandon Kinman, Ryan Hickman, Krista Reymann, Thomas B McHugh, + +and Vincent Vanhoucke. Google scanned objects: A high-quality dataset of 3d scanned household items. In 2022 International Conference on Robotics and Automation (ICRA), pages 2553-2560. IEEE, 2022. 5 +[14] Fabian Groh, Patrick Wieschollek, and Hendrik P. A. Lensch. Flex-convolution (million-scale point-cloud learning beyond grid-worlds). In *Asian Conference on Computer Vision (ACCV)*, 2018. 2, 3 +[15] Jonathan Ho, Ajay Jain, and Pieter Abbeel. Denoising diffusion probabilistic models. arXiv preprint arxiv:2006.11239, 2020. 1 +[16] Yicong Hong, Kai Zhang, Juixiang Gu, Sai Bi, Yang Zhou, Difan Liu, Feng Liu, Kalyan Sunkavalli, Trung Bui, and Hao Tan. LRM: large reconstruction model for single image to 3d. In The Twelfth International Conference on Learning Representations, ICLR 2024, Vienna, Austria, May 7-11, 2024. OpenReview.net, 2024. 3 +[17] Yicong Hong, Kai Zhang, Juixiang Gu, Sai Bi, Yang Zhou, Difan Liu, Feng Liu, Kalyan Sunkavalli, Trung Bui, and Hao Tan. Lrm: Large reconstruction model for single image to 3d, 2024. 1 +[18] Jiahui Huang, Zan Gojcic, Matan Atzmon, Or Litany, Sanja Fidler, and Francis Williams. Neural kernel surface reconstruction. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 4369-4379, 2023. 7 +[19] Bernhard Kerbl, Georgios Kopanas, Thomas Leimkuhler, and George Drettakis. 3d gaussian splatting for real-time radiance field rendering. ACM Trans. Graph., 42(4):139-1, 2023. 3 +[20] Arno Knapitsch, Jaesik Park, Qian-Yi Zhou, and Vladlen Koltun. Tanks and temples: benchmarking large-scale scene reconstruction. ACM Trans. Graph., 36(4), 2017. 7 +[21] Yushi Lan, Fangzhou Hong, Shuai Yang, Shangchen Zhou, Xuyi Meng, Bo Dai, Xingang Pan, and Chen Change Loy. Ln3diff: Scalable latent neural fields diffusion for speedy 3d generation. In ECCV, 2024. 1 +[22] Jiahao Li, Hao Tan, Kai Zhang, Zexiang Xu, Fujun Luan, Yinghao Xu, Yicong Hong, Kalyan Sunkavalli, Greg Shakhnarovich, and Sai Bi. Instant3d: Fast text-to-3d with sparse-view generation and large reconstruction model. arXiv preprint arXiv:2311.06214, 2023. 3 +[23] Weiyu Li, Rui Chen, Xuelin Chen, and Ping Tan. Sweetdreamer: Aligning geometric priors in 2d diffusion for consistent text-to-3d. International Conference on Learning Representations (ICLR), 2024. 3 +[24] Weiyu Li, Jiarui Liu, Rui Chen, Yixun Liang, Xuelin Chen, Ping Tan, and Xiaoxiao Long. Craftsman: High-fidelity mesh generation with 3d native generation and interactive geometry refiner. arXiv preprint arXiv:2405.14979, 2024. 1, 3, 7 +[25] Yixun Liang, Xin Yang, Jiantao Lin, Haodong Li, Xiaogang Xu, and Yingcong Chen. Luciddreamer: Towards high-fidelity text-to-3d generation via interval score matching. arXiv preprint arXiv:2311.11284, 2023. 1 +[26] Chen-Hsuan Lin, Jun Gao, Luming Tang, Towaki Takikawa, Xiaohui Zeng, Xun Huang, Karsten Kreis, Sanja Fidler, + +Ming-Yu Liu, and Tsung-Yi Lin. Magic3d: High-resolution text-to-3d content creation. In IEEE Conference on Computer Vision and Pattern Recognition (CVPR), 2023. 3 +[27] Fangfu Liu, Diankun Wu, Yi Wei, Yongming Rao, and Yueqi Duan. Sherpa3d: Boosting high-fidelity text-to-3d generation via coarse 3d prior, 2023. 3 +[28] Minghua Liu, Chao Xu, Haian Jin, Linghao Chen, Mukund Varma T, Zexiang Xu, and Hao Su. One-2-3-45: Any single image to 3d mesh in 45 seconds without per-shape optimization. 2024. 1 +[29] Minghua Liu, Chong Zeng, Xinyue Wei, Ruoxi Shi, Linghao Chen, Chao Xu, Mengqi Zhang, Zhaoning Wang, Xiaoshuai Zhang, Isabella Liu, Hongzhi Wu, and Hao Su. Meshformer: High-quality mesh generation with 3d-guided reconstruction model. arXiv preprint arXiv:2408.10198, 2024. 3 +[30] Yuan Liu, Cheng Lin, Zijiao Zeng, Xiaoxiao Long, Lingjie Liu, Taku Komura, and Wenping Wang. Syncdreamer: Generating multiview-consistent images from a single-view image. arXiv preprint arXiv:2309.03453, 2023. 1, 3 +[31] Ben Mildenhall, Pratul P. Srinivasan, Matthew Tancik, Jonathan T. Barron, Ravi Ramamoorthi, and Ren Ng. Nerf: Representing scenes as neural radiance fields for view synthesis. In European Conference on Computer Vision (ECCV), 2020. 3 +[32] Carsten Moenning and Neil A Dodgson. Fast marching farthest point sampling. Technical report, University of Cambridge, Computer Laboratory, 2003. 3 +[33] William Peebles and Saining Xie. Scalable diffusion models with transformers. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 4195-4205, 2023. 8 +[34] Ben Poole, Ajay Jain, Jonathan T. Barron, and Ben Mildenhall. Dreamfusion: Text-to-3d using 2d diffusion. arXiv, 2022. 3 +[35] Lingteng Qiu, Guanying Chen, Xiaodong Gu, Qi Zuo, Mutian Xu, Yushuang Wu, Weihao Yuan, Zilong Dong, Liefeng Bo, and Xiaoguang Han. Richdreamer: A generalizable normal-depth diffusion model for detail richness in text-to-3d. In IEEE/CVF Conference on Computer Vision and Pattern Recognition, CVPR 2024, Seattle, WA, USA, June 16-22, 2024, pages 9914-9925. IEEE, 2024. 3 +[36] Xuanchi Ren, Jiahui Huang, Xiaohui Zeng, Ken Museth, Sanja Fidler, and Francis Williams. Xcube: Large-scale 3d generative modeling using sparse voxel hierarchies. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 4209-4219, 2024. 1, 3, 7 +[37] Robin Rombach, Andreas Blattmann, Dominik Lorenz, Patrick Esser, and Björn Ommer. High-resolution image synthesis with latent diffusion models. In Conference on Computer Vision and Pattern Recognition (CVPR), pages 10684-10695, 2022. 1 +[38] Tianchang Shen, Jun Gao, Kangxue Yin, Ming-Yu Liu, and Sanja Fidler. Deep marching tetrahedra: a hybrid representation for high-resolution 3d shape synthesis. Advances in Neural Information Processing Systems, 34:6087-6101, 2021. 3 + +[39] Yichun Shi, Peng Wang, Jianglong Ye, Long Mai, Kejie Li, and Xiao Yang. Mvdream: Multi-view diffusion for 3d generation. arXiv:2308.16512, 2023. 3 +[40] Jiaxiang Tang, Zhaoxi Chen, Xiaokang Chen, Tengfei Wang, Gang Zeng, and Ziwei Liu. Lgm: Large multi-view gaussian model for high-resolution 3d content creation. arXiv preprint arXiv:2402.05054, 2024. 3 +[41] Jiaxiang Tang, Jiawei Ren, Hang Zhou, Ziwei Liu, and Gang Zeng. Dreamgaussian: Generative gaussian splatting for efficient 3d content creation. In The Twelfth International Conference on Learning Representations, ICLR 2024, Vienna, Austria, May 7-11, 2024. OpenReview.net, 2024. 3 +[42] Dmitry Tochilkin, David Pankratz, Zexiang Liu, Zixuan Huang, , Adam Letts, Yangguang Li, Ding Liang, Christian Laforte, Varun Jampani, and Yan-Pei Cao. Triposr: Fast 3d object reconstruction from a single image. arXiv preprint arXiv:2403.02151, 2024. 1 +[43] Peng Wang, Hao Tan, Sai Bi, Yinghao Xu, Fujun Luan, Kalyan Sunkavalli, Wenping Wang, Zexiang Xu, and Kai Zhang. Pf-lrm: Pose-free large reconstruction model for joint pose and shape prediction. arXiv preprint arXiv:2311.12024, 2023. 3 +[44] Tengfei Wang, Bo Zhang, Ting Zhang, Shuyang Gu, Jianmin Bao, Tadas Baltrusaitis, Jingjing Shen, Dong Chen, Fang Wen, Qifeng Chen, and Baining Guo. Rodin: A generative model for sculpting 3d digital avatars using diffusion, 2022. 1 +[45] Zhengyi Wang, Cheng Lu, Yikai Wang, Fan Bao, Chongxuan Li, Hang Su, and Jun Zhu. Prolificdreamer: High-fidelity and diverse text-to-3d generation with variational score distillation. arXiv preprint arXiv:2305.16213, 2023. 3 +[46] Zhengyi Wang, Yikai Wang, Yifei Chen, Chengdong Xiang, Shuo Chen, Dajiang Yu, Chongxuan Li, Hang Su, and Jun Zhu. Crm: Single image to 3d textured mesh with convolutional reconstruction model. arXiv preprint arXiv:2403.05034, 2024. 3 +[47] Francis Williams, Jiahui Huang, Jonathan Swartz, Gergely Klar, Vijay Thakkar, Matthew Cong, Xuanchi Ren, Rui-long Li, Clement Fuji-Tsang, Sanja Fidler, Eftychios Sifakis, and Ken Museth. fvdb: A deep-learning framework for sparse, large-scale, and high-performance spatial intelligence. ACM Transactions on Graphics (TOG), 43(4):133:1-133:15, 2024. 1 +[48] Chengzhi Wu, Junwei Zheng, Julius Pfrommer, and Jürgen Beyerer. Attention-based point cloud edge sampling. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), 2023. 2, 3 +[49] Shuang Wu, Youtian Lin, Feihu Zhang, Yifei Zeng, Jingxi Xu, Philip Torr, Xun Cao, and Yao Yao. Direct3d: Scalable image-to-3d generation via 3d latent diffusion transformer. arXiv:2405.14832, 2024. 1, 2, 3, 7 +[50] Bojun Xiong, Si-Tong Wei, Xin-Yang Zheng, Yan-Pei Cao, Zhouhui Lian, and Peng-Shuai Wang. OctFusion: Octree-based diffusion models for 3d shape generation. arXiv, 2024. 3 +[51] Jiale Xu, Weihao Cheng, Yiming Gao, Xintao Wang, Shenghua Gao, and Ying Shan. Instantmesh: Efficient 3d + +mesh generation from a single image with sparse-view large reconstruction models. arXiv preprint arXiv:2404.07191, 2024.3 +[52] Cem Yuksel. Sample elimination for generating poisson disk sample sets. The Eurographs Association & John Wiley & Sons, Ltd., 2015. 3, 8 +[53] Biao Zhang and Peter Wonka. Lagem: A large geometry model for 3d representation learning and diffusion. arXiv preprint arXiv:2410.01295, 2024. 3 +[54] Biao Zhang, Jiapeng Tang, Matthias Nießner, and Peter Wonka. 3dshape2vecset: A 3d shape representation for neural fields and generative diffusion models. ACM Transactions on Graphics (SIGGRAPH), 42(4), 2023. 1, 3, 4, 5, 7, 8 +[55] Lvmin Zhang, Anyi Rao, and Maneesh Agrawala. Adding conditional control to text-to-image diffusion models, 2023. 1 +[56] Longwen Zhang, Ziyu Wang, Qixuan Zhang, Qiwei Qiu, Anqi Pang, Haoran Jiang, Wei Yang, Lan Xu, and Jingyi Yu. Clay: A controllable large-scale generative model for creating high-quality 3d assets. ACM Transactions on Graphics (TOG), 43(4):1-20, 2024. 1, 2, 3, 5, 7, 8 +[57] Xuying Zhang, Bo-Wen Yin, Yuming Chen, Zheng Lin, Yunheng Li, Qibin Hou, and Ming-Ming Cheng. Temo: Towards text-driven 3d stylization for multi-object meshes. arXiv preprint arXiv:2312.04248, 2023. 1 +[58] Zibo Zhao, Wen Liu, Xin Chen, Xianfang Zeng, Rui Wang, Pei Cheng, BIN FU, Tao Chen, Gang YU, and Shenghua Gao. Michelangelo: Conditional 3d shape generation based on shape-image-text aligned latent representation. In Advances in Neural Information Processing Systems (NeurIPS), 2023. 1, 3 \ No newline at end of file diff --git a/2025/Dora_ Sampling and Benchmarking for 3D Shape Variational Auto-Encoders/images.zip b/2025/Dora_ Sampling and Benchmarking for 3D Shape Variational Auto-Encoders/images.zip new file mode 100644 index 0000000000000000000000000000000000000000..45b9b2c14e2f66a1f829c533e7fe26900cff3953 --- /dev/null +++ b/2025/Dora_ Sampling and Benchmarking for 3D Shape Variational Auto-Encoders/images.zip @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:c7abc0a0723ef1ff0022f3b163a4a6933568af4e902a0fed56c415eed366101a +size 871701 diff --git a/2025/Dora_ Sampling and Benchmarking for 3D Shape Variational Auto-Encoders/layout.json b/2025/Dora_ Sampling and Benchmarking for 3D Shape Variational Auto-Encoders/layout.json new file mode 100644 index 0000000000000000000000000000000000000000..bc2b896be5676c8cb636e86251317be4ac1eb963 --- /dev/null +++ b/2025/Dora_ Sampling and Benchmarking for 3D Shape Variational Auto-Encoders/layout.json @@ -0,0 +1,10239 @@ +{ + "pdf_info": [ + { + "para_blocks": [ + { + "bbox": [ + 67, + 103, + 542, + 121 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 67, + 103, + 542, + 121 + ], + "spans": [ + { + "bbox": [ + 67, + 103, + 542, + 121 + ], + "type": "text", + "content": "Dora: Sampling and Benchmarking for 3D Shape Variational Auto-Encoders" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 118, + 141, + 492, + 228 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 118, + 141, + 492, + 228 + ], + "spans": [ + { + "bbox": [ + 118, + 141, + 492, + 228 + ], + "type": "text", + "content": "Rui Chen" + }, + { + "bbox": [ + 118, + 141, + 492, + 228 + ], + "type": "inline_equation", + "content": "^{1,2}" + }, + { + "bbox": [ + 118, + 141, + 492, + 228 + ], + "type": "text", + "content": " Jianfeng Zhang" + }, + { + "bbox": [ + 118, + 141, + 492, + 228 + ], + "type": "inline_equation", + "content": "^{2\\dagger}" + }, + { + "bbox": [ + 118, + 141, + 492, + 228 + ], + "type": "text", + "content": " Yixun Liang" + }, + { + "bbox": [ + 118, + 141, + 492, + 228 + ], + "type": "inline_equation", + "content": "^{1,3}" + }, + { + "bbox": [ + 118, + 141, + 492, + 228 + ], + "type": "text", + "content": " Guan Luo" + }, + { + "bbox": [ + 118, + 141, + 492, + 228 + ], + "type": "inline_equation", + "content": "^{2,4}" + }, + { + "bbox": [ + 118, + 141, + 492, + 228 + ], + "type": "text", + "content": " Weiyu Li" + }, + { + "bbox": [ + 118, + 141, + 492, + 228 + ], + "type": "inline_equation", + "content": "^{1,3}" + }, + { + "bbox": [ + 118, + 141, + 492, + 228 + ], + "type": "text", + "content": " \nJiarui Liu" + }, + { + "bbox": [ + 118, + 141, + 492, + 228 + ], + "type": "inline_equation", + "content": "^{1,3}" + }, + { + "bbox": [ + 118, + 141, + 492, + 228 + ], + "type": "text", + "content": " Xiu Li" + }, + { + "bbox": [ + 118, + 141, + 492, + 228 + ], + "type": "inline_equation", + "content": "^{2}" + }, + { + "bbox": [ + 118, + 141, + 492, + 228 + ], + "type": "text", + "content": " Xiaoxiao Long" + }, + { + "bbox": [ + 118, + 141, + 492, + 228 + ], + "type": "inline_equation", + "content": "^{1,3}" + }, + { + "bbox": [ + 118, + 141, + 492, + 228 + ], + "type": "text", + "content": " Jiashi Feng" + }, + { + "bbox": [ + 118, + 141, + 492, + 228 + ], + "type": "inline_equation", + "content": "^{2}" + }, + { + "bbox": [ + 118, + 141, + 492, + 228 + ], + "type": "text", + "content": " Ping Tan" + }, + { + "bbox": [ + 118, + 141, + 492, + 228 + ], + "type": "inline_equation", + "content": "^{1,3\\dagger}" + }, + { + "bbox": [ + 118, + 141, + 492, + 228 + ], + "type": "text", + "content": " \n†Corresponding authors \n" + }, + { + "bbox": [ + 118, + 141, + 492, + 228 + ], + "type": "inline_equation", + "content": "^{1}" + }, + { + "bbox": [ + 118, + 141, + 492, + 228 + ], + "type": "text", + "content": "The Hong Kong University of Science and Technology " + }, + { + "bbox": [ + 118, + 141, + 492, + 228 + ], + "type": "inline_equation", + "content": "^{2}" + }, + { + "bbox": [ + 118, + 141, + 492, + 228 + ], + "type": "text", + "content": "ByteDance Seed \n" + }, + { + "bbox": [ + 118, + 141, + 492, + 228 + ], + "type": "inline_equation", + "content": "^{3}" + }, + { + "bbox": [ + 118, + 141, + 492, + 228 + ], + "type": "text", + "content": "LightIllusions " + }, + { + "bbox": [ + 118, + 141, + 492, + 228 + ], + "type": "inline_equation", + "content": "^{4}" + }, + { + "bbox": [ + 118, + 141, + 492, + 228 + ], + "type": "text", + "content": "Tsinghua University \nhttps://aruichen.github.io/Dora" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 151, + 255, + 200, + 268 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 151, + 255, + 200, + 268 + ], + "spans": [ + { + "bbox": [ + 151, + 255, + 200, + 268 + ], + "type": "text", + "content": "Abstract" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 55, + 283, + 297, + 594 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 283, + 297, + 594 + ], + "spans": [ + { + "bbox": [ + 55, + 283, + 297, + 594 + ], + "type": "text", + "content": "Recent 3D content generation pipelines commonly employ Variational Autoencoders (VAEs) to encode shapes into compact latent representations for diffusion-based generation. However, the widely adopted uniform point sampling strategy in Shape VAE training often leads to a significant loss of geometric details, limiting the quality of shape reconstruction and downstream generation tasks. We present Dora-VAE, a novel approach that enhances VAE reconstruction through our proposed sharp edge sampling strategy and a dual cross-attention mechanism. By identifying and prioritizing regions with high geometric complexity during training, our method significantly improves the preservation of fine-grained shape features. Such sampling strategy and the dual attention mechanism enable the VAE to focus on crucial geometric details that are typically missed by uniform sampling approaches. To systematically evaluate VAE reconstruction quality, we additionally propose Dora-bench, a benchmark that quantifies shape complexity through the density of sharp edges, introducing a new metric focused on reconstruction accuracy at these salient geometric features. Extensive experiments on the Dora-bench demonstrate that Dora-VAE achieves comparable reconstruction quality to the state-of-the-art dense XCube-VAE while requiring a latent space at least " + }, + { + "bbox": [ + 55, + 283, + 297, + 594 + ], + "type": "inline_equation", + "content": "8 \\times" + }, + { + "bbox": [ + 55, + 283, + 297, + 594 + ], + "type": "text", + "content": " smaller (1,280 vs. " + }, + { + "bbox": [ + 55, + 283, + 297, + 594 + ], + "type": "inline_equation", + "content": ">10,000" + }, + { + "bbox": [ + 55, + 283, + 297, + 594 + ], + "type": "text", + "content": " codes). Project page: https://aruichen.github.io/Dora." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 56, + 620, + 135, + 632 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 620, + 135, + 632 + ], + "spans": [ + { + "bbox": [ + 56, + 620, + 135, + 632 + ], + "type": "text", + "content": "1. Introduction" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 55, + 641, + 296, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 641, + 296, + 714 + ], + "spans": [ + { + "bbox": [ + 55, + 641, + 296, + 714 + ], + "type": "text", + "content": "3D content creation is vital to delivering realistic and immersive experiences in various industries, including games, movies, and AR/VR. However, traditional 3D modeling typically demands significant expertise and manual effort, making it time-consuming and challenging, especially for non-expert users. Recent advances in AI-powered 3D con" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 313, + 256, + 553, + 291 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 256, + 553, + 291 + ], + "spans": [ + { + "bbox": [ + 313, + 256, + 553, + 291 + ], + "type": "text", + "content": "tent generation methods [2, 3, 7, 9, 17, 21, 25, 28, 30, 42, 44, 57] have transformed the field, making it more accessible to many more users." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 313, + 293, + 555, + 388 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 293, + 555, + 388 + ], + "spans": [ + { + "bbox": [ + 313, + 293, + 555, + 388 + ], + "type": "text", + "content": "Following the success of text-to-image generation models [4, 5, 15, 37, 55], recent 3D content creation approaches [36, 49, 56] adopt a two-stage pipeline: encoding 3D shapes into a latent space using variational autoencoders (VAEs), followed by training a latent diffusion model. The performance of such a generative pipeline heavily relies on the VAE's capability to faithfully encode and reconstruct 3D shapes." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 313, + 388, + 554, + 473 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 388, + 554, + 473 + ], + "spans": [ + { + "bbox": [ + 313, + 388, + 554, + 473 + ], + "type": "text", + "content": "Existing 3D VAEs operate by sampling points on mesh surfaces for shape encoding and then reconstructing the original 3D meshes by its decoder. This process faces unique challenges compared to 2D image VAEs where the input image is fully observable. In comparison, the sampled point cloud often cannot capture all the necessary shape information, which could harm the performance of 3D VAEs." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 313, + 474, + 555, + 653 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 474, + 555, + 653 + ], + "spans": [ + { + "bbox": [ + 313, + 474, + 555, + 653 + ], + "type": "text", + "content": "Volume-based method [36] leverages sparse convolution [47] to process millions of voxelized points for high-fidelity reconstruction. Its dense sampling captures precise shape information. However, this method produces large latent codes (commonly " + }, + { + "bbox": [ + 313, + 474, + 555, + 653 + ], + "type": "inline_equation", + "content": ">10,000" + }, + { + "bbox": [ + 313, + 474, + 555, + 653 + ], + "type": "text", + "content": " tokens), which significantly complicate the training of diffusion models. On the other hand, vector-set (Vecset) methods [24, 54, 56, 58] use transformers to achieve compact latent representations (hundreds to thousands of tokens), enabling efficient diffusion [24, 56]. However, due to the quadratic complexity of transformer networks, it often only samples a few thousand points to represent a 3D shape, which leads to information loss and performance degradation. Therefore, we seek to improve the reconstruction quality of Vecset-based VAEs and maintain their compact representation." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 313, + 654, + 553, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 654, + 553, + 713 + ], + "spans": [ + { + "bbox": [ + 313, + 654, + 553, + 713 + ], + "type": "text", + "content": "We begin by analyzing the shape reconstruction capability of Vecset-based VAEs. Through careful analysis, we find these methods have limited reconstruction performance, which stems from their commonly used uniform sampling. When computational constraints limit the total" + } + ] + } + ], + "index": 12 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "spans": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "type": "text", + "content": "CVF" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 145, + 0, + 496, + 37 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 145, + 0, + 496, + 37 + ], + "spans": [ + { + "bbox": [ + 145, + 0, + 496, + 37 + ], + "type": "text", + "content": "This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "text", + "content": "16251" + } + ] + } + ], + "index": 13 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 0 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 75, + 72, + 159, + 147 + ], + "blocks": [ + { + "bbox": [ + 75, + 72, + 159, + 147 + ], + "lines": [ + { + "bbox": [ + 75, + 72, + 159, + 147 + ], + "spans": [ + { + "bbox": [ + 75, + 72, + 159, + 147 + ], + "type": "image", + "image_path": "1d7091c2be5c9b7d56c50afe9c7af043c387fd6b81d9f0d95246d92e7a0806c3.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 78, + 146, + 148, + 155 + ], + "lines": [ + { + "bbox": [ + 78, + 146, + 148, + 155 + ], + "spans": [ + { + "bbox": [ + 78, + 146, + 148, + 155 + ], + "type": "text", + "content": "(a) Ground Truth(GT)" + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "type": "image", + "bbox": [ + 63, + 156, + 171, + 265 + ], + "blocks": [ + { + "bbox": [ + 63, + 156, + 171, + 265 + ], + "lines": [ + { + "bbox": [ + 63, + 156, + 171, + 265 + ], + "spans": [ + { + "bbox": [ + 63, + 156, + 171, + 265 + ], + "type": "image", + "image_path": "3723c8ddf5058b50cb5b8b944adf45a354d3034414836c04f599aac1b01dbed1.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 79, + 266, + 167, + 276 + ], + "lines": [ + { + "bbox": [ + 79, + 266, + 167, + 276 + ], + "spans": [ + { + "bbox": [ + 79, + 266, + 167, + 276 + ], + "type": "text", + "content": "(d) Reconstruction accuracy" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 54, + 287, + 555, + 344 + ], + "lines": [ + { + "bbox": [ + 54, + 287, + 555, + 344 + ], + "spans": [ + { + "bbox": [ + 54, + 287, + 555, + 344 + ], + "type": "text", + "content": "Figure 1. Sampling strategy comparison. Given the ground truth mesh shown in (a), we visualize point clouds produced by uniform sampling in (b) and those generated by our proposed Sharp Edge Sampling (SES) in (c), at various sampling rates. In (d), we compare the reconstruction accuracy trained with our SES and the uniform sampling using the F-score metric. The comparison demonstrates SES consistently outperforms uniform sampling under varying sampling rates, as the point clouds generated by SES are more effective in capturing the salient features of the object." + } + ] + } + ], + "index": 22, + "angle": 0, + "type": "image_caption" + } + ], + "index": 2 + }, + { + "bbox": [ + 176, + 75, + 198, + 89 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 176, + 75, + 198, + 89 + ], + "spans": [ + { + "bbox": [ + 176, + 75, + 198, + 89 + ], + "type": "text", + "content": "point number" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 210, + 78, + 227, + 85 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 210, + 78, + 227, + 85 + ], + "spans": [ + { + "bbox": [ + 210, + 78, + 227, + 85 + ], + "type": "text", + "content": "4096" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 285, + 78, + 302, + 85 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 285, + 78, + 302, + 85 + ], + "spans": [ + { + "bbox": [ + 285, + 78, + 302, + 85 + ], + "type": "text", + "content": "8192" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 357, + 78, + 377, + 85 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 357, + 78, + 377, + 85 + ], + "spans": [ + { + "bbox": [ + 357, + 78, + 377, + 85 + ], + "type": "text", + "content": "16384" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 426, + 78, + 447, + 85 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 426, + 78, + 447, + 85 + ], + "spans": [ + { + "bbox": [ + 426, + 78, + 447, + 85 + ], + "type": "text", + "content": "20480" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 503, + 78, + 523, + 85 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 503, + 78, + 523, + 85 + ], + "spans": [ + { + "bbox": [ + 503, + 78, + 523, + 85 + ], + "type": "text", + "content": "40960" + } + ] + } + ], + "index": 9 + }, + { + "type": "image", + "bbox": [ + 187, + 100, + 258, + 166 + ], + "blocks": [ + { + "bbox": [ + 187, + 100, + 258, + 166 + ], + "lines": [ + { + "bbox": [ + 187, + 100, + 258, + 166 + ], + "spans": [ + { + "bbox": [ + 187, + 100, + 258, + 166 + ], + "type": "image", + "image_path": "61b761c8c0524fdbefc7fee85173068b8de6ee8ecbadde13ff5b5c01b104cf6f.jpg" + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "image_body" + } + ], + "index": 10 + }, + { + "type": "image", + "bbox": [ + 261, + 101, + 331, + 166 + ], + "blocks": [ + { + "bbox": [ + 261, + 101, + 331, + 166 + ], + "lines": [ + { + "bbox": [ + 261, + 101, + 331, + 166 + ], + "spans": [ + { + "bbox": [ + 261, + 101, + 331, + 166 + ], + "type": "image", + "image_path": "3791fd82f0a53b4151db302506296faedd3f58194ee58670b9ea10fce597f1a0.jpg" + } + ] + } + ], + "index": 11, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 291, + 167, + 417, + 177 + ], + "lines": [ + { + "bbox": [ + 291, + 167, + 417, + 177 + ], + "spans": [ + { + "bbox": [ + 291, + 167, + 417, + 177 + ], + "type": "text", + "content": "(b) Visual results of the uniform sampling" + } + ] + } + ], + "index": 15, + "angle": 0, + "type": "image_caption" + } + ], + "index": 11 + }, + { + "type": "image", + "bbox": [ + 335, + 101, + 405, + 166 + ], + "blocks": [ + { + "bbox": [ + 335, + 101, + 405, + 166 + ], + "lines": [ + { + "bbox": [ + 335, + 101, + 405, + 166 + ], + "spans": [ + { + "bbox": [ + 335, + 101, + 405, + 166 + ], + "type": "image", + "image_path": "718df31d5a83af4156943caa575e9190d455c9c0fc75e022af8f1e707517dd3e.jpg" + } + ] + } + ], + "index": 12, + "angle": 0, + "type": "image_body" + } + ], + "index": 12 + }, + { + "type": "image", + "bbox": [ + 408, + 101, + 479, + 166 + ], + "blocks": [ + { + "bbox": [ + 408, + 101, + 479, + 166 + ], + "lines": [ + { + "bbox": [ + 408, + 101, + 479, + 166 + ], + "spans": [ + { + "bbox": [ + 408, + 101, + 479, + 166 + ], + "type": "image", + "image_path": "d194ca7c9733c32f31613aae2706321ba6bdb020abe917ab69ac1f73fb0c1b16.jpg" + } + ] + } + ], + "index": 13, + "angle": 0, + "type": "image_body" + } + ], + "index": 13 + }, + { + "type": "image", + "bbox": [ + 482, + 102, + 552, + 166 + ], + "blocks": [ + { + "bbox": [ + 482, + 102, + 552, + 166 + ], + "lines": [ + { + "bbox": [ + 482, + 102, + 552, + 166 + ], + "spans": [ + { + "bbox": [ + 482, + 102, + 552, + 166 + ], + "type": "image", + "image_path": "c8b578ccaeb00d7c59a2bb47d98b19908a0c59fa44b82f4a0efdd5cab138d2e2.jpg" + } + ] + } + ], + "index": 14, + "angle": 0, + "type": "image_body" + } + ], + "index": 14 + }, + { + "type": "image", + "bbox": [ + 187, + 188, + 258, + 253 + ], + "blocks": [ + { + "bbox": [ + 187, + 188, + 258, + 253 + ], + "lines": [ + { + "bbox": [ + 187, + 188, + 258, + 253 + ], + "spans": [ + { + "bbox": [ + 187, + 188, + 258, + 253 + ], + "type": "image", + "image_path": "10420fe341601ac28b74e94a6b7974bfb0b6d3ebb3f785d8eac49363d427b2be.jpg" + } + ] + } + ], + "index": 16, + "angle": 0, + "type": "image_body" + } + ], + "index": 16 + }, + { + "type": "image", + "bbox": [ + 261, + 188, + 332, + 253 + ], + "blocks": [ + { + "bbox": [ + 261, + 188, + 332, + 253 + ], + "lines": [ + { + "bbox": [ + 261, + 188, + 332, + 253 + ], + "spans": [ + { + "bbox": [ + 261, + 188, + 332, + 253 + ], + "type": "image", + "image_path": "0e5defacb8d30c2a7f52511540f4ac6ba76050768d5784e1bbb978b333d5414e.jpg" + } + ] + } + ], + "index": 17, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 276, + 265, + 435, + 274 + ], + "lines": [ + { + "bbox": [ + 276, + 265, + 435, + 274 + ], + "spans": [ + { + "bbox": [ + 276, + 265, + 435, + 274 + ], + "type": "text", + "content": "(c) Visual results of our sharp edge sampling (SES)" + } + ] + } + ], + "index": 21, + "angle": 0, + "type": "image_caption" + } + ], + "index": 17 + }, + { + "type": "image", + "bbox": [ + 335, + 188, + 405, + 253 + ], + "blocks": [ + { + "bbox": [ + 335, + 188, + 405, + 253 + ], + "lines": [ + { + "bbox": [ + 335, + 188, + 405, + 253 + ], + "spans": [ + { + "bbox": [ + 335, + 188, + 405, + 253 + ], + "type": "image", + "image_path": "18a1a5cb08e34fdaad47c637593d26cf9f5228855c12451dd979ef4bfc1f4d50.jpg" + } + ] + } + ], + "index": 18, + "angle": 0, + "type": "image_body" + } + ], + "index": 18 + }, + { + "type": "image", + "bbox": [ + 408, + 188, + 479, + 253 + ], + "blocks": [ + { + "bbox": [ + 408, + 188, + 479, + 253 + ], + "lines": [ + { + "bbox": [ + 408, + 188, + 479, + 253 + ], + "spans": [ + { + "bbox": [ + 408, + 188, + 479, + 253 + ], + "type": "image", + "image_path": "cb2c82564119ae842784851ef1448d769987a3ad74a4674936b0276f111ec738.jpg" + } + ] + } + ], + "index": 19, + "angle": 0, + "type": "image_body" + } + ], + "index": 19 + }, + { + "type": "image", + "bbox": [ + 482, + 188, + 552, + 253 + ], + "blocks": [ + { + "bbox": [ + 482, + 188, + 552, + 253 + ], + "lines": [ + { + "bbox": [ + 482, + 188, + 552, + 253 + ], + "spans": [ + { + "bbox": [ + 482, + 188, + 552, + 253 + ], + "type": "image", + "image_path": "da169309bca4faf0378c628b5eaf7703bd0c38df8c5f59d04f8753c253e7449c.jpg" + } + ] + } + ], + "index": 20, + "angle": 0, + "type": "image_body" + } + ], + "index": 20 + }, + { + "bbox": [ + 54, + 363, + 295, + 518 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 54, + 363, + 295, + 518 + ], + "spans": [ + { + "bbox": [ + 54, + 363, + 295, + 518 + ], + "type": "text", + "content": "number of sampled points, uniform sampling fails to prioritize geometrically salient regions, leading to the loss of fine details. To validate this observation, we experiment on a 3D mesh with intricate geometric details (e.g., keyboard buttons) shown in Figure 1 (a). We visualize the point cloud with different sampling strategies at various sampling densities in (b) and (c). As demonstrated in Figure 1 (b), even with increasing sampling rates, uniform sampling fails to preserve sharp features like keyboard buttons. This simple experiment confirms that uniform sampling fundamentally limits the capturing of geometric details, which in turn affects the VAE's reconstruction capability and the result quality of the learned diffusion models." + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 54, + 520, + 295, + 628 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 54, + 520, + 295, + 628 + ], + "spans": [ + { + "bbox": [ + 54, + 520, + 295, + 628 + ], + "type": "text", + "content": "Inspired by the success of importance sampling in various geometric processing tasks [14, 48], we introduce a similar strategy for 3D VAE training. While existing importance sampling methods focus on down-sampling point clouds, our task requires sampling points directly from mesh surfaces for shape VAE training [49, 56]. This fundamental difference necessitates a new sampling approach specifically designed for preserving geometric features from mesh representations." + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 54, + 630, + 295, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 54, + 630, + 295, + 713 + ], + "spans": [ + { + "bbox": [ + 54, + 630, + 295, + 713 + ], + "type": "text", + "content": "To address this challenge, we propose a Sharp Edge Sampling (SES) algorithm that adaptively samples points based on geometric saliency. Specifically, SES first identifies edges with significant dihedral angles on the mesh, indicating regions of high geometric complexity. It then samples points along these salient regions while maintaining a balance with uniformly sampled points to capture the over" + } + ] + } + ], + "index": 25 + }, + { + "bbox": [ + 313, + 363, + 553, + 387 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 363, + 553, + 387 + ], + "spans": [ + { + "bbox": [ + 313, + 363, + 553, + 387 + ], + "type": "text", + "content": "all structure. This approach ensures comprehensive coverage of both fine details and global geometry." + } + ] + } + ], + "index": 26 + }, + { + "bbox": [ + 313, + 388, + 554, + 508 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 388, + 554, + 508 + ], + "spans": [ + { + "bbox": [ + 313, + 388, + 554, + 508 + ], + "type": "text", + "content": "Building upon SES, we present Dora-VAE, a novel method achieving high-fidelity reconstruction while maintaining compact latent representations. To fully leverage these detail-rich point clouds sampled by our SES, we design a dual cross-attention architecture that effectively processes both salient and uniform regions during encoding. As shown in Figure 1, our method significantly outperforms uniform sampling in preserving shape details on the keyboard (c), with consistent improvements in F-Score across different sampling rates (d)." + } + ] + } + ], + "index": 27 + }, + { + "bbox": [ + 313, + 509, + 554, + 675 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 509, + 554, + 675 + ], + "spans": [ + { + "bbox": [ + 313, + 509, + 554, + 675 + ], + "type": "text", + "content": "The common evaluation protocol for 3D VAEs is also biased. It typically uses a set of randomly selected 3D shapes, and employs general metrics (e.g., F-score, Chamfer distance) to measure the shape reconstruction quality. However, we argue it is necessary to divide the test shapes into sub-classes of different shape complexity to better evaluate these 3D VAEs. To facilitate the evaluation of 3D VAEs, we further introduce Dora-Bench, a new benchmark based on existing datasets with our novel Sharp Normal Error (SNE) metric. Dora-Bench categorizes test shapes based on their geometric complexity and SNE focuses on measuring the reconstruction quality of salient geometric features. This combination enables a more rigorous assessment of 3D VAEs." + } + ] + } + ], + "index": 28 + }, + { + "bbox": [ + 313, + 677, + 554, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 677, + 554, + 713 + ], + "spans": [ + { + "bbox": [ + 313, + 677, + 554, + 713 + ], + "type": "text", + "content": "Extensive experiments demonstrate that our Dora-VAE achieves superior results. When integrated into downstream 3D diffusion models, it significantly enhances the qual" + } + ] + } + ], + "index": 29 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "text", + "content": "16252" + } + ] + } + ], + "index": 30 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 1 + }, + { + "para_blocks": [ + { + "bbox": [ + 55, + 72, + 296, + 288 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 72, + 296, + 288 + ], + "spans": [ + { + "bbox": [ + 55, + 72, + 296, + 288 + ], + "type": "text", + "content": "ity of generated 3D shapes, which validates that our novel sampling strategy and dual cross-attention architecture effectively preserve geometric details with compact latent spaces. To summarize, our main contributions include: 1) We propose Dora-VAE, a novel 3D VAE model for high-quality reconstruction with compact latent representations, accompanied by Dora-Bench, a comprehensive benchmark for evaluating 3D VAEs. 2) We introduce, for the first time, importance sampling to the task of 3D VAE learning and propose a Sharp Edge Sampling (SES) algorithm to prioritize geometrically salient regions. Building on SES, we design a dual cross-attention architecture to effectively encode these detail-rich point clouds. 3) We develop a systematic evaluation benchmark based on existing datasets with our novel Sharp Normal Error (SNE) metric that specifically assesses reconstruction accuracy of fine geometric details, enabling more rigorous evaluation than conventional random sampling approaches." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 55, + 298, + 140, + 311 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 298, + 140, + 311 + ], + "spans": [ + { + "bbox": [ + 55, + 298, + 140, + 311 + ], + "type": "text", + "content": "2. Related work" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 55, + 319, + 296, + 413 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 319, + 296, + 413 + ], + "spans": [ + { + "bbox": [ + 55, + 319, + 296, + 413 + ], + "type": "text", + "content": "Importance Sampling in Point Clouds. Importance sampling techniques have been widely used in point cloud processing tasks [14, 48]. For instance, APES [48] proposes attention-based sampling for point cloud classification and segmentation. However, these methods operate directly on point clouds rather than meshes, making them less suitable for VAE-based shape representation where preserving complete geometric information is crucial." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 55, + 414, + 296, + 665 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 414, + 296, + 665 + ], + "spans": [ + { + "bbox": [ + 55, + 414, + 296, + 665 + ], + "type": "text", + "content": "3D Shape VAEs. Recent 3D shape VAEs follow two main approaches: volume-based and vector set-based. Volume-based methods [36, 50] like XCube [36] use sparse convolution to encode voxelized surfaces, achieving high reconstruction quality but requiring large latent codes (" + }, + { + "bbox": [ + 55, + 414, + 296, + 665 + ], + "type": "inline_equation", + "content": ">10,000" + }, + { + "bbox": [ + 55, + 414, + 296, + 665 + ], + "type": "text", + "content": " tokens). While these methods excel at preserving geometric details, their large latent spaces pose significant challenges for downstream diffusion model training. In contrast, vector set-based approaches [24, 49, 54, 56, 58] encode uniformly sampled surface points using transformers, producing highly compact latent spaces that are particularly suitable for diffusion models. However, these methods often struggle with geometric detail preservation, especially in regions with complex surface features. Our analysis reveals that this limitation primarily stems from their uniform sampling strategy: when computational constraints restrict the total number of processable points, uniform sampling fails to prioritize geometrically significant regions, leading to insufficient capture of fine details. This information loss at the sampling stage fundamentally limits these methods' ability to learn and preserve intricate geometric features." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 55, + 665, + 296, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 665, + 296, + 714 + ], + "spans": [ + { + "bbox": [ + 55, + 665, + 296, + 714 + ], + "type": "text", + "content": "3D Content Creation. Current 3D generation methods can be categorized into three groups. Optimization-based methods [6, 23, 26, 27, 30, 34, 35, 39, 41, 45], pioneered by DreamFusion [34] utilize score distillation sam" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 313, + 72, + 555, + 324 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 72, + 555, + 324 + ], + "spans": [ + { + "bbox": [ + 313, + 72, + 555, + 324 + ], + "type": "text", + "content": "pling (SDS) to optimize 3D representations [19, 31, 38] using 2D diffusion model priors. While these methods can achieve photorealistic results, they suffer from slow generation speed, training instability, and often struggle to maintain geometric consistency. Large reconstruction models, like LRM [16] and follow-up works [22, 29, 40, 43, 46, 51] employ large-scale sparse-view reconstruction for efficient 3D generation. However, their lack of explicit geometric priors often leads to compromised geometric fidelity and inconsistent surface details. 3D native generative models [24, 49, 54, 56, 58], represented by 3DShape2VecSet [54], adopt a two-stage approach: first training a 3D VAE to encode shapes into latent space, then training a conditional latent diffusion model for generation. This approach ensures better geometric consistency through the VAE's built-in geometric constraints. However, the quality of generated shapes is fundamentally limited by the VAE's reconstruction capability. Recent works [24, 53] have shown that improving VAE reconstruction directly enhances downstream generation quality, which motivates our focus on advancing VAE design." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 314, + 335, + 370, + 347 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 314, + 335, + 370, + 347 + ], + "spans": [ + { + "bbox": [ + 314, + 335, + 370, + 347 + ], + "type": "text", + "content": "3. Method" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 313, + 356, + 555, + 415 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 356, + 555, + 415 + ], + "spans": [ + { + "bbox": [ + 313, + 356, + 555, + 415 + ], + "type": "text", + "content": "In this section, we present Dora-VAE for high-quality 3D reconstruction, and Dora-Bench for 3D VAE evaluation. We first briefly review 3DShape2VecSet [54] in Section 3.1, the foundation of our method and then detail our key innovations in Section 3.2 and Section 3.3." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 313, + 424, + 480, + 437 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 424, + 480, + 437 + ], + "spans": [ + { + "bbox": [ + 313, + 424, + 480, + 437 + ], + "type": "text", + "content": "3.1. Preliminary: 3DShape2VecSet" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 313, + 441, + 554, + 490 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 441, + 554, + 490 + ], + "spans": [ + { + "bbox": [ + 313, + 441, + 554, + 490 + ], + "type": "text", + "content": "3DShape2VecSet [54] introduces a transformer-based 3D VAE that encodes uniformly sampled surface points into compact latent codes. Given a 3D surface S, their pipeline consists of three key steps:" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 314, + 491, + 554, + 550 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 314, + 491, + 554, + 550 + ], + "spans": [ + { + "bbox": [ + 314, + 491, + 554, + 550 + ], + "type": "text", + "content": "- Surface Sampling: Uniformly sample " + }, + { + "bbox": [ + 314, + 491, + 554, + 550 + ], + "type": "inline_equation", + "content": "N_{d}" + }, + { + "bbox": [ + 314, + 491, + 554, + 550 + ], + "type": "text", + "content": " points on the surface S using Poisson disk sampling [52] to obtain a dense point cloud " + }, + { + "bbox": [ + 314, + 491, + 554, + 550 + ], + "type": "inline_equation", + "content": "P_{d}" + }, + { + "bbox": [ + 314, + 491, + 554, + 550 + ], + "type": "text", + "content": ", then downsample it to " + }, + { + "bbox": [ + 314, + 491, + 554, + 550 + ], + "type": "inline_equation", + "content": "N_{s}" + }, + { + "bbox": [ + 314, + 491, + 554, + 550 + ], + "type": "text", + "content": " points via Farthest Point Sampling (FPS) [32] to get a sparse point cloud " + }, + { + "bbox": [ + 314, + 491, + 554, + 550 + ], + "type": "inline_equation", + "content": "P_{s}" + }, + { + "bbox": [ + 314, + 491, + 554, + 550 + ], + "type": "text", + "content": ":" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 328, + 560, + 555, + 574 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 328, + 560, + 555, + 574 + ], + "spans": [ + { + "bbox": [ + 328, + 560, + 555, + 574 + ], + "type": "interline_equation", + "content": "P _ {d} = \\left\\{p _ {d} ^ {i} \\in S \\mid i = 1, \\dots , N _ {d} \\right\\}, P _ {s} = \\operatorname {F P S} \\left(P _ {d}, N _ {s}\\right). \\tag {1}", + "image_path": "a3ccee68497407cc9f6ae817327f91bb1572dc3de062b83f0252a66105e5ff12.jpg" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 314, + 584, + 554, + 620 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 314, + 584, + 554, + 620 + ], + "spans": [ + { + "bbox": [ + 314, + 584, + 554, + 620 + ], + "type": "text", + "content": "- Feature Encoding: Compute the point cloud feature " + }, + { + "bbox": [ + 314, + 584, + 554, + 620 + ], + "type": "inline_equation", + "content": "C" + }, + { + "bbox": [ + 314, + 584, + 554, + 620 + ], + "type": "text", + "content": " via the cross-attention between " + }, + { + "bbox": [ + 314, + 584, + 554, + 620 + ], + "type": "inline_equation", + "content": "P_{s}" + }, + { + "bbox": [ + 314, + 584, + 554, + 620 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 314, + 584, + 554, + 620 + ], + "type": "inline_equation", + "content": "P_{d}" + }, + { + "bbox": [ + 314, + 584, + 554, + 620 + ], + "type": "text", + "content": ", followed by some self-attention layers to generate the latent code " + }, + { + "bbox": [ + 314, + 584, + 554, + 620 + ], + "type": "inline_equation", + "content": "z" + }, + { + "bbox": [ + 314, + 584, + 554, + 620 + ], + "type": "text", + "content": ":" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 337, + 630, + 555, + 643 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 337, + 630, + 555, + 643 + ], + "spans": [ + { + "bbox": [ + 337, + 630, + 555, + 643 + ], + "type": "interline_equation", + "content": "C = \\operatorname {C r o s s A t t n} \\left(P _ {s}, P _ {d}, P _ {d}\\right), z = \\operatorname {S e l f A t t n} (C). \\tag {2}", + "image_path": "6854deda4d0060bab6748c9ca0ace7a088a5b7fabf794961cd8b3dbac44241ce.jpg" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 313, + 653, + 554, + 690 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 653, + 554, + 690 + ], + "spans": [ + { + "bbox": [ + 313, + 653, + 554, + 690 + ], + "type": "text", + "content": "- Geometry Decoding: Further decode " + }, + { + "bbox": [ + 313, + 653, + 554, + 690 + ], + "type": "inline_equation", + "content": "z" + }, + { + "bbox": [ + 313, + 653, + 554, + 690 + ], + "type": "text", + "content": " through selfattention layers and predict occupancy values using randomly sampled spatial query points " + }, + { + "bbox": [ + 313, + 653, + 554, + 690 + ], + "type": "inline_equation", + "content": "Q_{space} \\in \\mathbb{R}^3" + }, + { + "bbox": [ + 313, + 653, + 554, + 690 + ], + "type": "text", + "content": ":" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 361, + 700, + 555, + 715 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 361, + 700, + 555, + 715 + ], + "spans": [ + { + "bbox": [ + 361, + 700, + 555, + 715 + ], + "type": "interline_equation", + "content": "\\hat {O} = \\operatorname {C r o s s A t t n} \\left(Q _ {\\text {s p a c e}}, \\operatorname {S e l f A t t n} (z)\\right). \\tag {3}", + "image_path": "1b47253716077769bfcde1f8a53342683058f2fd91e20b230aac8cfa185e5ad3.jpg" + } + ] + } + ], + "index": 15 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "text", + "content": "16253" + } + ] + } + ], + "index": 16 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 2 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 58, + 71, + 293, + 157 + ], + "blocks": [ + { + "bbox": [ + 58, + 71, + 293, + 157 + ], + "lines": [ + { + "bbox": [ + 58, + 71, + 293, + 157 + ], + "spans": [ + { + "bbox": [ + 58, + 71, + 293, + 157 + ], + "type": "image", + "image_path": "3086c3de39cf007f9521440ba59a4bf290529273b569fcbfb1d3a64390923a11.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + } + ], + "index": 0 + }, + { + "type": "image", + "bbox": [ + 56, + 169, + 294, + 269 + ], + "blocks": [ + { + "bbox": [ + 121, + 159, + 208, + 169 + ], + "lines": [ + { + "bbox": [ + 121, + 159, + 208, + 169 + ], + "spans": [ + { + "bbox": [ + 121, + 159, + 208, + 169 + ], + "type": "text", + "content": "(a) Sharp edge sampling" + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 56, + 169, + 294, + 269 + ], + "lines": [ + { + "bbox": [ + 56, + 169, + 294, + 269 + ], + "spans": [ + { + "bbox": [ + 56, + 169, + 294, + 269 + ], + "type": "image", + "image_path": "6260bff6b48d558afdc9f3e2d77308aca2788d2a8507c69e764d197bfc799384.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 55, + 277, + 295, + 354 + ], + "lines": [ + { + "bbox": [ + 55, + 277, + 295, + 354 + ], + "spans": [ + { + "bbox": [ + 55, + 277, + 295, + 354 + ], + "type": "text", + "content": "Figure 2. Overview of Dora-VAE. (a) We utilize the proposed sharp edge sampling technique to extract both salient and uniform points from the input mesh. These points are then combined with dense points, effectively capturing both salient regions and smooth areas. (b) To enhance the encoding of point clouds sampled through sharp edge sampling, we design a dual cross-attention architecture." + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_caption" + } + ], + "index": 2 + }, + { + "bbox": [ + 55, + 365, + 296, + 414 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 365, + 296, + 414 + ], + "spans": [ + { + "bbox": [ + 55, + 365, + 296, + 414 + ], + "type": "text", + "content": "While this method generates compact latent codes for diffusion, the uniform sampling limits its ability to capture fine geometric details. Our work addresses this limitation by carefully designed sampling and encoding strategies." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 55, + 422, + 127, + 433 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 422, + 127, + 433 + ], + "spans": [ + { + "bbox": [ + 55, + 422, + 127, + 433 + ], + "type": "text", + "content": "3.2. Dora-VAE" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 55, + 439, + 295, + 559 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 439, + 295, + 559 + ], + "spans": [ + { + "bbox": [ + 55, + 439, + 295, + 559 + ], + "type": "text", + "content": "Figure 2 gives an overview of our pipeline. For each input mesh, we augment the uniformly sampled point cloud " + }, + { + "bbox": [ + 55, + 439, + 295, + 559 + ], + "type": "inline_equation", + "content": "P_{u}" + }, + { + "bbox": [ + 55, + 439, + 295, + 559 + ], + "type": "text", + "content": " with more important points " + }, + { + "bbox": [ + 55, + 439, + 295, + 559 + ], + "type": "inline_equation", + "content": "P_{a}" + }, + { + "bbox": [ + 55, + 439, + 295, + 559 + ], + "type": "text", + "content": " sampled by our proposed sharp edge sampling strategy, which forms the dense point cloud " + }, + { + "bbox": [ + 55, + 439, + 295, + 559 + ], + "type": "inline_equation", + "content": "P_{d}" + }, + { + "bbox": [ + 55, + 439, + 295, + 559 + ], + "type": "text", + "content": ". During the encoding process, we compute the attention for " + }, + { + "bbox": [ + 55, + 439, + 295, + 559 + ], + "type": "inline_equation", + "content": "P_{u}" + }, + { + "bbox": [ + 55, + 439, + 295, + 559 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 55, + 439, + 295, + 559 + ], + "type": "inline_equation", + "content": "P_{a}" + }, + { + "bbox": [ + 55, + 439, + 295, + 559 + ], + "type": "text", + "content": " separately via a simple-yet-effective dual cross attention mechanism and sum the results for self-attention to compute the latent code " + }, + { + "bbox": [ + 55, + 439, + 295, + 559 + ], + "type": "inline_equation", + "content": "z" + }, + { + "bbox": [ + 55, + 439, + 295, + 559 + ], + "type": "text", + "content": ". Our VAE training largely follows that of 3DShape2VecSet [54], which is supervised by a loss evaluated on the occupancy field." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 55, + 566, + 204, + 578 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 566, + 204, + 578 + ], + "spans": [ + { + "bbox": [ + 55, + 566, + 204, + 578 + ], + "type": "text", + "content": "3.2.1. Sharp Edge Sampling (SES)" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 55, + 582, + 295, + 676 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 582, + 295, + 676 + ], + "spans": [ + { + "bbox": [ + 55, + 582, + 295, + 676 + ], + "type": "text", + "content": "We propose SES to effectively sample points from geometrically salient regions. To ensure surface coverage, we also sample points uniformly. Our final sampled point cloud " + }, + { + "bbox": [ + 55, + 582, + 295, + 676 + ], + "type": "inline_equation", + "content": "P_{d}" + }, + { + "bbox": [ + 55, + 582, + 295, + 676 + ], + "type": "text", + "content": " combines uniformly sampled points " + }, + { + "bbox": [ + 55, + 582, + 295, + 676 + ], + "type": "inline_equation", + "content": "P_{u}" + }, + { + "bbox": [ + 55, + 582, + 295, + 676 + ], + "type": "text", + "content": " with points specifically sampled from salient regions " + }, + { + "bbox": [ + 55, + 582, + 295, + 676 + ], + "type": "inline_equation", + "content": "P_{a}" + }, + { + "bbox": [ + 55, + 582, + 295, + 676 + ], + "type": "text", + "content": " as " + }, + { + "bbox": [ + 55, + 582, + 295, + 676 + ], + "type": "inline_equation", + "content": "P_{d} = P_{u} \\cup P_{a}" + }, + { + "bbox": [ + 55, + 582, + 295, + 676 + ], + "type": "text", + "content": ". Our method computes salient points " + }, + { + "bbox": [ + 55, + 582, + 295, + 676 + ], + "type": "inline_equation", + "content": "P_{a}" + }, + { + "bbox": [ + 55, + 582, + 295, + 676 + ], + "type": "text", + "content": " through two steps: detecting salient edges and sampling points from these regions." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 55, + 677, + 296, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 677, + 296, + 713 + ], + "spans": [ + { + "bbox": [ + 55, + 677, + 296, + 713 + ], + "type": "text", + "content": "Salient Edges Detection. Given a triangular mesh, we identify a set of salient edges " + }, + { + "bbox": [ + 55, + 677, + 296, + 713 + ], + "type": "inline_equation", + "content": "\\Gamma" + }, + { + "bbox": [ + 55, + 677, + 296, + 713 + ], + "type": "text", + "content": " by analyzing dihedral angles between adjacent faces, which calculates the angle be" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 313, + 72, + 554, + 120 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 72, + 554, + 120 + ], + "spans": [ + { + "bbox": [ + 313, + 72, + 554, + 120 + ], + "type": "text", + "content": "tween the normal vectors of adjacent faces, providing a direct measure of surface curvature at mesh edges. For each edge " + }, + { + "bbox": [ + 313, + 72, + 554, + 120 + ], + "type": "inline_equation", + "content": "e" + }, + { + "bbox": [ + 313, + 72, + 554, + 120 + ], + "type": "text", + "content": " shared by adjacent faces " + }, + { + "bbox": [ + 313, + 72, + 554, + 120 + ], + "type": "inline_equation", + "content": "f_{1}" + }, + { + "bbox": [ + 313, + 72, + 554, + 120 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 313, + 72, + 554, + 120 + ], + "type": "inline_equation", + "content": "f_{2}" + }, + { + "bbox": [ + 313, + 72, + 554, + 120 + ], + "type": "text", + "content": ", we compute the dihedral angle " + }, + { + "bbox": [ + 313, + 72, + 554, + 120 + ], + "type": "inline_equation", + "content": "\\theta_{e}" + }, + { + "bbox": [ + 313, + 72, + 554, + 120 + ], + "type": "text", + "content": " as:" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 371, + 125, + 555, + 152 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 371, + 125, + 555, + 152 + ], + "spans": [ + { + "bbox": [ + 371, + 125, + 555, + 152 + ], + "type": "interline_equation", + "content": "\\theta_ {e} = \\operatorname {a r c c o s} \\left(\\frac {\\mathbf {n} _ {f _ {1}} \\cdot \\mathbf {n} _ {f _ {2}}}{\\| \\mathbf {n} _ {f _ {1}} \\| \\| \\mathbf {n} _ {f _ {2}} \\|}\\right), \\tag {4}", + "image_path": "e7b2329ef7d2428975d3ae1ff40d5a8a3036e936dda2a3aaa187a446a09d4524.jpg" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 313, + 156, + 555, + 191 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 156, + 555, + 191 + ], + "spans": [ + { + "bbox": [ + 313, + 156, + 555, + 191 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 313, + 156, + 555, + 191 + ], + "type": "inline_equation", + "content": "\\mathbf{n}_{f_1}" + }, + { + "bbox": [ + 313, + 156, + 555, + 191 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 313, + 156, + 555, + 191 + ], + "type": "inline_equation", + "content": "\\mathbf{n}_{f_2}" + }, + { + "bbox": [ + 313, + 156, + 555, + 191 + ], + "type": "text", + "content": " are the normals of " + }, + { + "bbox": [ + 313, + 156, + 555, + 191 + ], + "type": "inline_equation", + "content": "f_{1}" + }, + { + "bbox": [ + 313, + 156, + 555, + 191 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 313, + 156, + 555, + 191 + ], + "type": "inline_equation", + "content": "f_{2}" + }, + { + "bbox": [ + 313, + 156, + 555, + 191 + ], + "type": "text", + "content": ". The salient edge set " + }, + { + "bbox": [ + 313, + 156, + 555, + 191 + ], + "type": "inline_equation", + "content": "\\Gamma" + }, + { + "bbox": [ + 313, + 156, + 555, + 191 + ], + "type": "text", + "content": " contains all edges with a dihedral angle exceeding a predefined threshold " + }, + { + "bbox": [ + 313, + 156, + 555, + 191 + ], + "type": "inline_equation", + "content": "\\tau" + }, + { + "bbox": [ + 313, + 156, + 555, + 191 + ], + "type": "text", + "content": ":" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 397, + 203, + 553, + 216 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 397, + 203, + 553, + 216 + ], + "spans": [ + { + "bbox": [ + 397, + 203, + 553, + 216 + ], + "type": "interline_equation", + "content": "\\Gamma = \\left\\{e \\mid \\theta_ {e} > \\tau \\right\\} \\tag {5}", + "image_path": "224622a4eb2ccf06d9550d1cb68e770d992950c28cbd947ffa49f5f7239a4538.jpg" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 313, + 219, + 539, + 230 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 219, + 539, + 230 + ], + "spans": [ + { + "bbox": [ + 313, + 219, + 539, + 230 + ], + "type": "text", + "content": "Let " + }, + { + "bbox": [ + 313, + 219, + 539, + 230 + ], + "type": "inline_equation", + "content": "N_{\\Gamma} = |\\Gamma|" + }, + { + "bbox": [ + 313, + 219, + 539, + 230 + ], + "type": "text", + "content": " represent the number of the salient edges." + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 313, + 231, + 554, + 266 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 231, + 554, + 266 + ], + "spans": [ + { + "bbox": [ + 313, + 231, + 554, + 266 + ], + "type": "text", + "content": "Salient Points Sampling. For each salient edge " + }, + { + "bbox": [ + 313, + 231, + 554, + 266 + ], + "type": "inline_equation", + "content": "e \\in \\Gamma" + }, + { + "bbox": [ + 313, + 231, + 554, + 266 + ], + "type": "text", + "content": ", we collect its two vertices " + }, + { + "bbox": [ + 313, + 231, + 554, + 266 + ], + "type": "inline_equation", + "content": "v_{e,1}" + }, + { + "bbox": [ + 313, + 231, + 554, + 266 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 313, + 231, + 554, + 266 + ], + "type": "inline_equation", + "content": "v_{e,2}" + }, + { + "bbox": [ + 313, + 231, + 554, + 266 + ], + "type": "text", + "content": " into a salient vertex set " + }, + { + "bbox": [ + 313, + 231, + 554, + 266 + ], + "type": "inline_equation", + "content": "P_{\\Gamma}" + }, + { + "bbox": [ + 313, + 231, + 554, + 266 + ], + "type": "text", + "content": ":" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 380, + 266, + 553, + 279 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 380, + 266, + 553, + 279 + ], + "spans": [ + { + "bbox": [ + 380, + 266, + 553, + 279 + ], + "type": "interline_equation", + "content": "P _ {\\Gamma} = \\left\\{v _ {e, 1}, v _ {e, 2} \\mid e \\in \\Gamma \\right\\}, \\tag {6}", + "image_path": "9ae474596eeadc9703f05292be263c99f72832bd557012eedc6f51a8d898d960.jpg" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 313, + 282, + 554, + 317 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 282, + 554, + 317 + ], + "spans": [ + { + "bbox": [ + 313, + 282, + 554, + 317 + ], + "type": "text", + "content": "where duplicate vertices from connecting edges are included only once. Let " + }, + { + "bbox": [ + 313, + 282, + 554, + 317 + ], + "type": "inline_equation", + "content": "N_V = |P_{\\Gamma}|" + }, + { + "bbox": [ + 313, + 282, + 554, + 317 + ], + "type": "text", + "content": " denote the number of unique vertices in " + }, + { + "bbox": [ + 313, + 282, + 554, + 317 + ], + "type": "inline_equation", + "content": "P_{\\Gamma}" + }, + { + "bbox": [ + 313, + 282, + 554, + 317 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 313, + 319, + 554, + 353 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 319, + 554, + 353 + ], + "spans": [ + { + "bbox": [ + 313, + 319, + 554, + 353 + ], + "type": "text", + "content": "Given a target number of salient points " + }, + { + "bbox": [ + 313, + 319, + 554, + 353 + ], + "type": "inline_equation", + "content": "N_{\\mathrm{desired}}" + }, + { + "bbox": [ + 313, + 319, + 554, + 353 + ], + "type": "text", + "content": ", we generate the salient point set " + }, + { + "bbox": [ + 313, + 319, + 554, + 353 + ], + "type": "inline_equation", + "content": "P_{a}" + }, + { + "bbox": [ + 313, + 319, + 554, + 353 + ], + "type": "text", + "content": " based on the available salient vertices:" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 325, + 369, + 553, + 413 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 325, + 369, + 553, + 413 + ], + "spans": [ + { + "bbox": [ + 325, + 369, + 553, + 413 + ], + "type": "interline_equation", + "content": "P _ {a} = \\left\\{ \\begin{array}{l l} \\operatorname {F P S} \\left(P _ {\\Gamma}, N _ {\\text {d e s i r e d}}\\right), & \\text {i f} N _ {\\text {d e s i r e d}} < = N _ {V}, \\\\ P _ {\\Gamma} \\cup P _ {\\text {i n t e r p o l a t e d}}, & \\text {i f} 0 < N _ {V} < N _ {\\text {d e s i r e d}}, \\\\ \\emptyset , & \\text {i f} N _ {V} = 0. \\end{array} \\right. \\tag {7}", + "image_path": "00fa56aa6fae3e5d9dd61e259c596c1299e1ed540a3516d0adf66018c16bb4dc.jpg" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 313, + 418, + 554, + 526 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 418, + 554, + 526 + ], + "spans": [ + { + "bbox": [ + 313, + 418, + 554, + 526 + ], + "type": "text", + "content": "When we have excess salient vertices (" + }, + { + "bbox": [ + 313, + 418, + 554, + 526 + ], + "type": "inline_equation", + "content": "N_{\\text{desired}} \\leq N_V" + }, + { + "bbox": [ + 313, + 418, + 554, + 526 + ], + "type": "text", + "content": "), we use FPS to downsample " + }, + { + "bbox": [ + 313, + 418, + 554, + 526 + ], + "type": "inline_equation", + "content": "P_{\\Gamma}" + }, + { + "bbox": [ + 313, + 418, + 554, + 526 + ], + "type": "text", + "content": " to obtain " + }, + { + "bbox": [ + 313, + 418, + 554, + 526 + ], + "type": "inline_equation", + "content": "P_a" + }, + { + "bbox": [ + 313, + 418, + 554, + 526 + ], + "type": "text", + "content": ". For cases with insufficient salient vertices (" + }, + { + "bbox": [ + 313, + 418, + 554, + 526 + ], + "type": "inline_equation", + "content": "N_V < N_{\\text{desired}}" + }, + { + "bbox": [ + 313, + 418, + 554, + 526 + ], + "type": "text", + "content": "), we include all vertices from " + }, + { + "bbox": [ + 313, + 418, + 554, + 526 + ], + "type": "inline_equation", + "content": "P_{\\Gamma}" + }, + { + "bbox": [ + 313, + 418, + 554, + 526 + ], + "type": "text", + "content": " and supplement with additional points " + }, + { + "bbox": [ + 313, + 418, + 554, + 526 + ], + "type": "inline_equation", + "content": "P_{\\text{interpolated}}" + }, + { + "bbox": [ + 313, + 418, + 554, + 526 + ], + "type": "text", + "content": ". These additional points are generated by uniformly sampling " + }, + { + "bbox": [ + 313, + 418, + 554, + 526 + ], + "type": "inline_equation", + "content": "(N_{\\text{desired}} - N_V) / N_{\\Gamma}" + }, + { + "bbox": [ + 313, + 418, + 554, + 526 + ], + "type": "text", + "content": " points along each salient edge in " + }, + { + "bbox": [ + 313, + 418, + 554, + 526 + ], + "type": "inline_equation", + "content": "\\Gamma" + }, + { + "bbox": [ + 313, + 418, + 554, + 526 + ], + "type": "text", + "content": ", ensuring comprehensive coverage of salient features. When no salient edges are detected (" + }, + { + "bbox": [ + 313, + 418, + 554, + 526 + ], + "type": "inline_equation", + "content": "N_V = 0" + }, + { + "bbox": [ + 313, + 418, + 554, + 526 + ], + "type": "text", + "content": "), " + }, + { + "bbox": [ + 313, + 418, + 554, + 526 + ], + "type": "inline_equation", + "content": "P_a" + }, + { + "bbox": [ + 313, + 418, + 554, + 526 + ], + "type": "text", + "content": " remains empty." + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 313, + 530, + 432, + 540 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 530, + 432, + 540 + ], + "spans": [ + { + "bbox": [ + 313, + 530, + 432, + 540 + ], + "type": "text", + "content": "3.2.2. Dual Cross Attention" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 313, + 544, + 554, + 605 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 544, + 554, + 605 + ], + "spans": [ + { + "bbox": [ + 313, + 544, + 554, + 605 + ], + "type": "text", + "content": "Given the point clouds " + }, + { + "bbox": [ + 313, + 544, + 554, + 605 + ], + "type": "inline_equation", + "content": "P_{d}" + }, + { + "bbox": [ + 313, + 544, + 554, + 605 + ], + "type": "text", + "content": " produced by our SES strategy, we design a dual cross-attention architecture to effectively encode both uniform and salient regions. Following 3DShape2VecSet [54], we first downsample " + }, + { + "bbox": [ + 313, + 544, + 554, + 605 + ], + "type": "inline_equation", + "content": "P_{u}" + }, + { + "bbox": [ + 313, + 544, + 554, + 605 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 313, + 544, + 554, + 605 + ], + "type": "inline_equation", + "content": "P_{a}" + }, + { + "bbox": [ + 313, + 544, + 554, + 605 + ], + "type": "text", + "content": " separately using FPS:" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 354, + 611, + 553, + 624 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 354, + 611, + 553, + 624 + ], + "spans": [ + { + "bbox": [ + 354, + 611, + 553, + 624 + ], + "type": "interline_equation", + "content": "P _ {s} = \\operatorname {F P S} \\left(P _ {u}, N _ {s, 1}\\right) \\cup \\operatorname {F P S} \\left(P _ {a}, N _ {s, 2}\\right), \\tag {8}", + "image_path": "ae3f89c86965804a94a8eca81c868f92ed188c293c1ae33082f8fad236d609de.jpg" + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 313, + 628, + 554, + 676 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 628, + 554, + 676 + ], + "spans": [ + { + "bbox": [ + 313, + 628, + 554, + 676 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 313, + 628, + 554, + 676 + ], + "type": "inline_equation", + "content": "N_{s,1}" + }, + { + "bbox": [ + 313, + 628, + 554, + 676 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 313, + 628, + 554, + 676 + ], + "type": "inline_equation", + "content": "N_{s,2}" + }, + { + "bbox": [ + 313, + 628, + 554, + 676 + ], + "type": "text", + "content": " is the number of downsampled point clouds from " + }, + { + "bbox": [ + 313, + 628, + 554, + 676 + ], + "type": "inline_equation", + "content": "P_u" + }, + { + "bbox": [ + 313, + 628, + 554, + 676 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 313, + 628, + 554, + 676 + ], + "type": "inline_equation", + "content": "P_a" + }, + { + "bbox": [ + 313, + 628, + 554, + 676 + ], + "type": "text", + "content": ", respectively. We then compute cross-attention features separately for uniform and salient points as follows:" + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 374, + 683, + 553, + 696 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 374, + 683, + 553, + 696 + ], + "spans": [ + { + "bbox": [ + 374, + 683, + 553, + 696 + ], + "type": "interline_equation", + "content": "C _ {u} = \\operatorname {C r o s s A t t n} \\left(P _ {s}, P _ {u}, P _ {u}\\right) \\tag {9}", + "image_path": "5ce001937df9f23d1284ef04d2bdd3cf575c83f37d74c8255e40d850aae3279d.jpg" + } + ] + } + ], + "index": 25 + }, + { + "bbox": [ + 375, + 698, + 553, + 710 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 375, + 698, + 553, + 710 + ], + "spans": [ + { + "bbox": [ + 375, + 698, + 553, + 710 + ], + "type": "interline_equation", + "content": "C _ {a} = \\operatorname {C r o s s A t t n} \\left(P _ {s}, P _ {a}, P _ {a}\\right) \\tag {10}", + "image_path": "6ca5320315906497fa96c44a89b7c133e6e9a859edf5e7c3e0f785f22ca1a121.jpg" + } + ] + } + ], + "index": 26 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "text", + "content": "16254" + } + ] + } + ], + "index": 27 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 3 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 59, + 70, + 293, + 140 + ], + "blocks": [ + { + "bbox": [ + 59, + 70, + 293, + 140 + ], + "lines": [ + { + "bbox": [ + 59, + 70, + 293, + 140 + ], + "spans": [ + { + "bbox": [ + 59, + 70, + 293, + 140 + ], + "type": "image", + "image_path": "762d6fdac16e892ef128d95b0fbfb7f3d7afc4834a9dd2ac14d2d5ea6fad7bdf.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 111, + 140, + 254, + 147 + ], + "lines": [ + { + "bbox": [ + 111, + 140, + 254, + 147 + ], + "spans": [ + { + "bbox": [ + 111, + 140, + 254, + 147 + ], + "type": "text", + "content": "(a) Counts of Different Datasets Across Detail Levels" + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "type": "image", + "bbox": [ + 65, + 155, + 157, + 216 + ], + "blocks": [ + { + "bbox": [ + 65, + 155, + 157, + 216 + ], + "lines": [ + { + "bbox": [ + 65, + 155, + 157, + 216 + ], + "spans": [ + { + "bbox": [ + 65, + 155, + 157, + 216 + ], + "type": "image", + "image_path": "0e35f9650b8f748045dad41a39219dfa06a62f94199fe9252f06feb1b9d6333d.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 66, + 218, + 156, + 227 + ], + "lines": [ + { + "bbox": [ + 66, + 218, + 156, + 227 + ], + "spans": [ + { + "bbox": [ + 66, + 218, + 156, + 227 + ], + "type": "text", + "content": "(b) Total Counts by Detail Levels" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_caption" + } + ], + "index": 2 + }, + { + "type": "image", + "bbox": [ + 164, + 148, + 287, + 220 + ], + "blocks": [ + { + "bbox": [ + 164, + 148, + 287, + 220 + ], + "lines": [ + { + "bbox": [ + 164, + 148, + 287, + 220 + ], + "spans": [ + { + "bbox": [ + 164, + 148, + 287, + 220 + ], + "type": "image", + "image_path": "0426c7405dd740c90ea7c6ed7b3f5db4a0dca48fa80ee10cf5f8aa39af21ef5e.jpg" + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 177, + 220, + 263, + 227 + ], + "lines": [ + { + "bbox": [ + 177, + 220, + 263, + 227 + ], + "spans": [ + { + "bbox": [ + 177, + 220, + 263, + 227 + ], + "type": "text", + "content": "(c) Visualization of detail levels" + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 55, + 237, + 296, + 293 + ], + "lines": [ + { + "bbox": [ + 55, + 237, + 296, + 293 + ], + "spans": [ + { + "bbox": [ + 55, + 237, + 296, + 293 + ], + "type": "text", + "content": "Figure 3. Our proposed benchmark include 3D shapes from the ABO [10], GSO [13], Meta [1], and Objaverse [12] datasets. (a) The histogram of different datasets across different shape complexities. (b) The pie chart of the total counts by shape complexities. (c) Sample shapes of different shape complexities." + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "image_caption" + } + ], + "index": 4 + }, + { + "type": "image", + "bbox": [ + 61, + 314, + 293, + 467 + ], + "blocks": [ + { + "bbox": [ + 61, + 314, + 293, + 467 + ], + "lines": [ + { + "bbox": [ + 61, + 314, + 293, + 467 + ], + "spans": [ + { + "bbox": [ + 61, + 314, + 293, + 467 + ], + "type": "image", + "image_path": "9a18b1c24ae4dcfc95938048cfe55d2e56a7e6b0c465994029eeab5eccc3b6fe.jpg" + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 55, + 478, + 294, + 501 + ], + "lines": [ + { + "bbox": [ + 55, + 478, + 294, + 501 + ], + "spans": [ + { + "bbox": [ + 55, + 478, + 294, + 501 + ], + "type": "text", + "content": "Figure 4. The process of computing sharp normal errors (SNE). We compute MSE loss in the sharp regions of the normal." + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "image_caption" + } + ], + "index": 7 + }, + { + "bbox": [ + 55, + 527, + 295, + 550 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 527, + 295, + 550 + ], + "spans": [ + { + "bbox": [ + 55, + 527, + 295, + 550 + ], + "type": "text", + "content": "The final point cloud feature " + }, + { + "bbox": [ + 55, + 527, + 295, + 550 + ], + "type": "inline_equation", + "content": "C" + }, + { + "bbox": [ + 55, + 527, + 295, + 550 + ], + "type": "text", + "content": " combines both attention results:" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 143, + 560, + 295, + 573 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 143, + 560, + 295, + 573 + ], + "spans": [ + { + "bbox": [ + 143, + 560, + 295, + 573 + ], + "type": "interline_equation", + "content": "C = C _ {u} + C _ {a}. \\tag {11}", + "image_path": "1d1d06b6e1746e56a8c53fd170cb1bfa0609b19700f15276921a8ac6cc100a58.jpg" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 55, + 588, + 296, + 648 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 588, + 296, + 648 + ], + "spans": [ + { + "bbox": [ + 55, + 588, + 296, + 648 + ], + "type": "text", + "content": "This dual attention design enables separate focus on uniform and salient regions during feature extraction. Following 3DShape2VecSet [54], we use " + }, + { + "bbox": [ + 55, + 588, + 296, + 648 + ], + "type": "inline_equation", + "content": "C" + }, + { + "bbox": [ + 55, + 588, + 296, + 648 + ], + "type": "text", + "content": " to predict the occupancy field " + }, + { + "bbox": [ + 55, + 588, + 296, + 648 + ], + "type": "inline_equation", + "content": "\\hat{O}" + }, + { + "bbox": [ + 55, + 588, + 296, + 648 + ], + "type": "text", + "content": " through self-attention blocks. The whole model with parameters " + }, + { + "bbox": [ + 55, + 588, + 296, + 648 + ], + "type": "inline_equation", + "content": "\\psi" + }, + { + "bbox": [ + 55, + 588, + 296, + 648 + ], + "type": "text", + "content": " are optimized using MSE loss:" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 94, + 662, + 295, + 696 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 94, + 662, + 295, + 696 + ], + "spans": [ + { + "bbox": [ + 94, + 662, + 295, + 696 + ], + "type": "interline_equation", + "content": "\\nabla_ {\\psi} \\mathcal {L} _ {\\mathrm {M S E}} (\\hat {O}, O) = \\mathbb {E} \\left[ 2 (\\hat {O} - O) \\frac {\\partial \\hat {O}}{\\partial \\psi} \\right]. \\tag {12}", + "image_path": "f0ca547003d59c8a6675d956d95156742c13a552bace7efcf021b78bf3160841.jpg" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 314, + 72, + 394, + 83 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 314, + 72, + 394, + 83 + ], + "spans": [ + { + "bbox": [ + 314, + 72, + 394, + 83 + ], + "type": "text", + "content": "3.3. Dora-Bench" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 313, + 89, + 515, + 101 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 89, + 515, + 101 + ], + "spans": [ + { + "bbox": [ + 313, + 89, + 515, + 101 + ], + "type": "text", + "content": "3.3.1. Geometric Complexity-based Evaluation" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 313, + 104, + 553, + 186 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 104, + 553, + 186 + ], + "spans": [ + { + "bbox": [ + 313, + 104, + 553, + 186 + ], + "type": "text", + "content": "To enable a more rigorous evaluation of VAE performance, we propose Dora-bench, a benchmark that systematically categorizes test shapes based on their geometric complexity. Unlike previous methods that use randomly selected test sets, we measure shape complexity using the number of salient edges " + }, + { + "bbox": [ + 313, + 104, + 553, + 186 + ], + "type": "inline_equation", + "content": "N_{\\Gamma}" + }, + { + "bbox": [ + 313, + 104, + 553, + 186 + ], + "type": "text", + "content": " (Section 3.2.1) and classify shapes into four levels:" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 313, + 188, + 528, + 236 + ], + "type": "list", + "angle": 0, + "index": 20, + "blocks": [ + { + "bbox": [ + 313, + 188, + 487, + 199 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 188, + 487, + 199 + ], + "spans": [ + { + "bbox": [ + 313, + 188, + 487, + 199 + ], + "type": "text", + "content": "- Level 1 (Less Detail): " + }, + { + "bbox": [ + 313, + 188, + 487, + 199 + ], + "type": "inline_equation", + "content": "0 < N_{\\Gamma} \\leq 5000" + }, + { + "bbox": [ + 313, + 188, + 487, + 199 + ], + "type": "text", + "content": ";" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 313, + 200, + 528, + 211 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 200, + 528, + 211 + ], + "spans": [ + { + "bbox": [ + 313, + 200, + 528, + 211 + ], + "type": "text", + "content": "- Level 2 (Moderate Detail): " + }, + { + "bbox": [ + 313, + 200, + 528, + 211 + ], + "type": "inline_equation", + "content": "5000 < N_{\\Gamma} \\leq 10000" + }, + { + "bbox": [ + 313, + 200, + 528, + 211 + ], + "type": "text", + "content": ";" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 313, + 212, + 512, + 223 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 212, + 512, + 223 + ], + "spans": [ + { + "bbox": [ + 313, + 212, + 512, + 223 + ], + "type": "text", + "content": "- Level 3 (Rich Detail): " + }, + { + "bbox": [ + 313, + 212, + 512, + 223 + ], + "type": "inline_equation", + "content": "10000 < N_{\\Gamma} \\leq 50000" + }, + { + "bbox": [ + 313, + 212, + 512, + 223 + ], + "type": "text", + "content": ";" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 313, + 224, + 495, + 236 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 224, + 495, + 236 + ], + "spans": [ + { + "bbox": [ + 313, + 224, + 495, + 236 + ], + "type": "text", + "content": "- Level 4 (Very Rich Detail): " + }, + { + "bbox": [ + 313, + 224, + 495, + 236 + ], + "type": "inline_equation", + "content": "N_{\\Gamma} > 50000" + } + ] + } + ], + "index": 19 + } + ], + "sub_type": "text" + }, + { + "bbox": [ + 313, + 236, + 553, + 308 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 236, + 553, + 308 + ], + "spans": [ + { + "bbox": [ + 313, + 236, + 553, + 308 + ], + "type": "text", + "content": "We curate test shapes from multiple public datasets including GSO [13], ABO [10], Meta [1], and Objaverse [12] to ensure diverse geometric complexities. Figure 3 shows the distribution of shapes across complexity levels (a,b) and example meshes from each level (c). Please refer to our supplementary materials for more examples." + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 313, + 312, + 460, + 323 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 312, + 460, + 323 + ], + "spans": [ + { + "bbox": [ + 313, + 312, + 460, + 323 + ], + "type": "text", + "content": "3.3.2. Sharp Normal Error (SNE)" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 313, + 327, + 555, + 518 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 327, + 555, + 518 + ], + "spans": [ + { + "bbox": [ + 313, + 327, + 555, + 518 + ], + "type": "text", + "content": "Building on our Dora-bench, we further introduce Sharp Normal Error (SNE) to evaluate reconstruction quality in salient regions. While existing metrics like Chamfer Distance and F-Score capture overall shape similarity, they fail to specifically assess the preservation of fine geometric details. SNE addresses this limitation by measuring normal map differences between reconstructed and ground truth shapes in geometrically significant areas. As illustrated in Figure 4, we render normal maps of the ground truth shape from multiple viewpoints and identify salient regions using Canny edge detection. These regions are dilated to create evaluation masks. The final SNE metric is computed as the Mean Squared Error between ground truth and reconstructed normal maps within the masked areas. This process enables focused evaluation of how well VAEs preserve sharp geometric features during reconstruction." + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 313, + 527, + 395, + 540 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 527, + 395, + 540 + ], + "spans": [ + { + "bbox": [ + 313, + 527, + 395, + 540 + ], + "type": "text", + "content": "4. Experiments" + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 313, + 547, + 554, + 582 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 547, + 554, + 582 + ], + "spans": [ + { + "bbox": [ + 313, + 547, + 554, + 582 + ], + "type": "text", + "content": "We conducted intensive experiments to validate the effectiveness of our proposed 3D VAE and compare it with other state-of-the-art methods." + } + ] + } + ], + "index": 25 + }, + { + "bbox": [ + 313, + 588, + 447, + 601 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 588, + 447, + 601 + ], + "spans": [ + { + "bbox": [ + 313, + 588, + 447, + 601 + ], + "type": "text", + "content": "4.1. Implementation Details" + } + ] + } + ], + "index": 26 + }, + { + "bbox": [ + 313, + 605, + 555, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 605, + 555, + 715 + ], + "spans": [ + { + "bbox": [ + 313, + 605, + 555, + 715 + ], + "type": "text", + "content": "We follow CLAY [56] for mesh preprocessing to ensure watertight 3D models. Our VAE is trained on a subset filtered from Objaverse [12], containing approximately 400,000 3D meshes. We filter out low-quality meshes with missing faces or severe self-intersections to ensure training stability. Our training is conducted on 32 A100 GPUs for two days using a batch size of 2048 and a learning rate of 5e-5. We employ Flash-Attention-v2 [11], mixed-precision training with FP16 and gradient checkpointing [8] to optimize" + } + ] + } + ], + "index": 27 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "text", + "content": "16255" + } + ] + } + ], + "index": 28 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 4 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 56, + 69, + 553, + 677 + ], + "blocks": [ + { + "bbox": [ + 56, + 69, + 553, + 677 + ], + "lines": [ + { + "bbox": [ + 56, + 69, + 553, + 677 + ], + "spans": [ + { + "bbox": [ + 56, + 69, + 553, + 677 + ], + "type": "image", + "image_path": "c29adb258d92ddda61c839c850298bfd3c87a122d627da02089f38748083b368.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 55, + 686, + 555, + 709 + ], + "lines": [ + { + "bbox": [ + 55, + 686, + 555, + 709 + ], + "spans": [ + { + "bbox": [ + 55, + 686, + 555, + 709 + ], + "type": "text", + "content": "Figure 5. Qualitative comparison of the VAE reconstruction results. " + }, + { + "bbox": [ + 55, + 686, + 555, + 709 + ], + "type": "inline_equation", + "content": "\\dagger" + }, + { + "bbox": [ + 55, + 686, + 555, + 709 + ], + "type": "text", + "content": " indicates the fine-tuning model that uses the same training data as ours." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 749, + 318, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 749, + 318, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 749, + 318, + 757 + ], + "type": "text", + "content": "16256" + } + ] + } + ], + "index": 2 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 5 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 66, + 70, + 547, + 193 + ], + "blocks": [ + { + "bbox": [ + 66, + 70, + 547, + 193 + ], + "lines": [ + { + "bbox": [ + 66, + 70, + 547, + 193 + ], + "spans": [ + { + "bbox": [ + 66, + 70, + 547, + 193 + ], + "type": "table", + "html": "
MethodsLCL↑ F-score(0.01) × 100↑ F-score(0.005) × 100↓ CD × 10000↓ SNE × 100
L1L2L3L4L1L2L3L4L1L2L3L4L1L2L3L4
Xcube [36]>1000098.96898.79998.61598.22695.52593.87292.32285.3656.3156.2887.9359.9261.5791.4321.4301.679
Xcube† [36]>1000099.39399.79499.82499.07996.75395.53593.42287.3654.0154.1425.7407.6271.5431.4081.2591.639
Craftsman [24]25698.01695.87491.75681.73987.99482.54973.00057.3794.3899.12914.53033.4411.9061.8732.1913.933
Ours w/o DCA128099.96499.92599.67897.89096.56195.97591.61883.1242.2362.5064.4446.4321.4481.2151.2051.828
Ours w/o SES,DCA128099.94499.81497.29496.77995.97794.62388.40679.2402.4222.9833.9806.1961.4961.3131.3522.207
Ours full25699.50798.98696.66989.57793.27290.46682.38668.6693.3565.20210.27624.5271.5551.4101.6183.035
128099.98899.95599.88099.17097.03896.83193.45887.4732.0972.5003.9455.2651.4331.1861.1371.579
", + "image_path": "c61ef96b3fe20e54924317f44b2efa23caceccb877dce1ff3931f518e1af30ad.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "table_body" + } + ], + "index": 0 + }, + { + "bbox": [ + 83, + 196, + 526, + 207 + ], + "lines": [ + { + "bbox": [ + 83, + 196, + 526, + 207 + ], + "spans": [ + { + "bbox": [ + 83, + 196, + 526, + 207 + ], + "type": "text", + "content": "Table 1. Quantitative comparison in Dora-bench. " + }, + { + "bbox": [ + 83, + 196, + 526, + 207 + ], + "type": "inline_equation", + "content": "{}^{ \\dagger }" + }, + { + "bbox": [ + 83, + 196, + 526, + 207 + ], + "type": "text", + "content": " indicates the fine-tuning model that uses the same training data as ours." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "text" + }, + { + "bbox": [ + 55, + 228, + 295, + 277 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 228, + 295, + 277 + ], + "spans": [ + { + "bbox": [ + 55, + 228, + 295, + 277 + ], + "type": "text", + "content": "memory usage and training efficiency. For the parameters in sharp edge sampling, we set " + }, + { + "bbox": [ + 55, + 228, + 295, + 277 + ], + "type": "inline_equation", + "content": "N_{\\mathrm{desired}} = 16384" + }, + { + "bbox": [ + 55, + 228, + 295, + 277 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 55, + 228, + 295, + 277 + ], + "type": "inline_equation", + "content": "\\tau = 30" + }, + { + "bbox": [ + 55, + 228, + 295, + 277 + ], + "type": "text", + "content": ". We set the low threshold to 20 and the high threshold to 200 for the canny edge detection." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 55, + 285, + 167, + 298 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 285, + 167, + 298 + ], + "spans": [ + { + "bbox": [ + 55, + 285, + 167, + 298 + ], + "type": "text", + "content": "4.2. Evaluation Setting" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 55, + 303, + 296, + 470 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 303, + 296, + 470 + ], + "spans": [ + { + "bbox": [ + 55, + 303, + 296, + 470 + ], + "type": "text", + "content": "Metrics. We evaluate reconstruction quality by comparing input meshes with their decoded counterparts from different 3D VAEs using 1M sampled points under three metrics: 1) F-score " + }, + { + "bbox": [ + 55, + 303, + 296, + 470 + ], + "type": "inline_equation", + "content": "(r)" + }, + { + "bbox": [ + 55, + 303, + 296, + 470 + ], + "type": "text", + "content": " [20], which reconstruction accuracy by computing precision and recall of point correspondences within distance threshold " + }, + { + "bbox": [ + 55, + 303, + 296, + 470 + ], + "type": "inline_equation", + "content": "r" + }, + { + "bbox": [ + 55, + 303, + 296, + 470 + ], + "type": "text", + "content": ". Specifically, we report F-score (0.01) and F-score (0.005) with shapes normalized " + }, + { + "bbox": [ + 55, + 303, + 296, + 470 + ], + "type": "inline_equation", + "content": "[-1,1]" + }, + { + "bbox": [ + 55, + 303, + 296, + 470 + ], + "type": "text", + "content": ". 2) Chamfer Distance (CD), which computes the average distance between each reconstructed point and its nearest ground truth point. 3) Sharp Normal Errors (SNE) as proposed in Section 3.3.2, which evaluate normal map differences in salient regions. For fair comparison, we also report Latent Code Length (LCL) as longer codes typically enable better reconstruction." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 55, + 471, + 296, + 578 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 471, + 296, + 578 + ], + "spans": [ + { + "bbox": [ + 55, + 471, + 296, + 578 + ], + "type": "text", + "content": "Baselines. We compare Dora-VAE with state-of-the-art approaches, including: 1) XCube-VAE [36], a volumetric method with larger latent codes; 2) XCube-VAE† [36], our fine-tuned version of the original XCube-VAE on the same dataset; 3) Craftsman-VAE[24], which fine-tune the 3DShape2VecSet [54] with shorter latent codes on Objverse. We exclude VAE models from Direct3D [49] and CLAY[56] as as their implementations were not publicly available at submission time." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 55, + 587, + 191, + 601 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 587, + 191, + 601 + ], + "spans": [ + { + "bbox": [ + 55, + 587, + 191, + 601 + ], + "type": "text", + "content": "4.3. Qualitative Comparison" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 55, + 605, + 295, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 605, + 295, + 715 + ], + "spans": [ + { + "bbox": [ + 55, + 605, + 295, + 715 + ], + "type": "text", + "content": "Figure 5 shows visual comparisons of different methods across different complexity levels from our Dora-bench dataset. We visualize both ground truth and reconstructed meshes using surface normal coloring to highlight geometric details. For shapes with lower complexity (L1 and L2), all methods achieve comparable reconstruction quality. However, when dealing with shapes of higher complexity (L3 and L4), the advantages of our method become obvious. While XCube-VAE achieves similar visual quality" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 313, + 228, + 555, + 337 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 228, + 555, + 337 + ], + "spans": [ + { + "bbox": [ + 313, + 228, + 555, + 337 + ], + "type": "text", + "content": "to ours, it requires a significantly larger latent space - more than " + }, + { + "bbox": [ + 313, + 228, + 555, + 337 + ], + "type": "inline_equation", + "content": "8 \\times" + }, + { + "bbox": [ + 313, + 228, + 555, + 337 + ], + "type": "text", + "content": " the size of ours (" + }, + { + "bbox": [ + 313, + 228, + 555, + 337 + ], + "type": "inline_equation", + "content": ">10,000" + }, + { + "bbox": [ + 313, + 228, + 555, + 337 + ], + "type": "text", + "content": " dimensions vs. 1,280). This substantial reduction in latent code length, while maintaining high reconstruction fidelity, makes our method particularly suitable for training 3D diffusion models. In contrast, Craftsman-VAE shows a noticeable degradation in reconstruction quality for complex shapes, failing to capture fine geometric details. Additional visual comparisons are provided in the appendix." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 313, + 345, + 457, + 357 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 345, + 457, + 357 + ], + "spans": [ + { + "bbox": [ + 313, + 345, + 457, + 357 + ], + "type": "text", + "content": "4.4. Quantitative Comparison" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 313, + 362, + 555, + 506 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 362, + 555, + 506 + ], + "spans": [ + { + "bbox": [ + 313, + 362, + 555, + 506 + ], + "type": "text", + "content": "Table 1 presents quantitative results of different methods across different complexity levels of Dora-bench. Our method consistently outperforms baselines across all levels, with larger margins on more complex shapes (L3 and L4). The advantage is particularly evident in CD metrics, where our method with only 256 latent codes surpasses even our fine-tuned version of XCube-VAE (3.356 vs. 4.015). When using 1280 latent codes, our method further decreases CD to 2.097, achieving a " + }, + { + "bbox": [ + 313, + 362, + 555, + 506 + ], + "type": "inline_equation", + "content": "47.77\\%" + }, + { + "bbox": [ + 313, + 362, + 555, + 506 + ], + "type": "text", + "content": " improvement over XCube-VAE†. We attribute XCube-VAE's lower performance to its use of NKSR [18] for mesh extraction, which introduces additional quantization errors." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 313, + 506, + 556, + 627 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 506, + 556, + 627 + ], + "spans": [ + { + "bbox": [ + 313, + 506, + 556, + 627 + ], + "type": "text", + "content": "Notably, our method demonstrates superior performance in preserving geometric details, as reflected by the SNE metric. For example, in L4 shapes where geometric complexity is highest, our method achieves an SNE of 1.579 compared to 1.639 from XCube-VAE†, representing a " + }, + { + "bbox": [ + 313, + 506, + 556, + 627 + ], + "type": "inline_equation", + "content": "3.7\\%" + }, + { + "bbox": [ + 313, + 506, + 556, + 627 + ], + "type": "text", + "content": " improvement. This significant gain in SNE aligns with our qualitative observations in Figure 5, where our method better preserves fine details such as sharp edges and complex surface variations, demonstrating the effectiveness of our sharp edge sampling strategy." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 313, + 635, + 416, + 647 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 635, + 416, + 647 + ], + "spans": [ + { + "bbox": [ + 313, + 635, + 416, + 647 + ], + "type": "text", + "content": "4.5. Ablation Studies" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 313, + 653, + 553, + 688 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 653, + 553, + 688 + ], + "spans": [ + { + "bbox": [ + 313, + 653, + 553, + 688 + ], + "type": "text", + "content": "To evaluate the contribution of each component, we compare our full model with two variants under the same training conditions:" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 313, + 689, + 555, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 689, + 555, + 714 + ], + "spans": [ + { + "bbox": [ + 313, + 689, + 555, + 714 + ], + "type": "text", + "content": "- Ours w/o SES, DCA. This variant removes both sharp edge sampling (SES) and dual cross attention (DCA), i.e." + } + ] + } + ], + "index": 14 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 318, + 758 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 318, + 758 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 318, + 758 + ], + "type": "text", + "content": "16257" + } + ] + } + ], + "index": 15 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 6 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 78, + 85, + 157, + 174 + ], + "blocks": [ + { + "bbox": [ + 111, + 74, + 124, + 83 + ], + "lines": [ + { + "bbox": [ + 111, + 74, + 124, + 83 + ], + "spans": [ + { + "bbox": [ + 111, + 74, + 124, + 83 + ], + "type": "text", + "content": "GT" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 78, + 85, + 157, + 174 + ], + "lines": [ + { + "bbox": [ + 78, + 85, + 157, + 174 + ], + "spans": [ + { + "bbox": [ + 78, + 85, + 157, + 174 + ], + "type": "image", + "image_path": "a725a621e317993e35a9a1c648a7adc36716ce5021113677bdd3e1b0e6ac6125.jpg" + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_body" + } + ], + "index": 1 + }, + { + "type": "image", + "bbox": [ + 188, + 85, + 265, + 174 + ], + "blocks": [ + { + "bbox": [ + 219, + 73, + 235, + 82 + ], + "lines": [ + { + "bbox": [ + 219, + 73, + 235, + 82 + ], + "spans": [ + { + "bbox": [ + 219, + 73, + 235, + 82 + ], + "type": "text", + "content": "Full" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 188, + 85, + 265, + 174 + ], + "lines": [ + { + "bbox": [ + 188, + 85, + 265, + 174 + ], + "spans": [ + { + "bbox": [ + 188, + 85, + 265, + 174 + ], + "type": "image", + "image_path": "5fe5ec243e59f522a4a135eaa1a908510eb65606e4581824486aa11e830c8ea2.jpg" + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 211, + 176, + 247, + 186 + ], + "lines": [ + { + "bbox": [ + 211, + 176, + 247, + 186 + ], + "spans": [ + { + "bbox": [ + 211, + 176, + 247, + 186 + ], + "type": "text", + "content": "w/o DAC" + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "image_caption" + } + ], + "index": 4 + }, + { + "type": "image", + "bbox": [ + 80, + 192, + 157, + 281 + ], + "blocks": [ + { + "bbox": [ + 89, + 178, + 145, + 187 + ], + "lines": [ + { + "bbox": [ + 89, + 178, + 145, + 187 + ], + "spans": [ + { + "bbox": [ + 89, + 178, + 145, + 187 + ], + "type": "text", + "content": "w/o SES, DCA" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 80, + 192, + 157, + 281 + ], + "lines": [ + { + "bbox": [ + 80, + 192, + 157, + 281 + ], + "spans": [ + { + "bbox": [ + 80, + 192, + 157, + 281 + ], + "type": "image", + "image_path": "45c477727be1322becd3da2860816f627377b230cfe97a126222cb7586a5f408.jpg" + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 55, + 293, + 295, + 337 + ], + "lines": [ + { + "bbox": [ + 55, + 293, + 295, + 337 + ], + "spans": [ + { + "bbox": [ + 55, + 293, + 295, + 337 + ], + "type": "text", + "content": "Figure 6. Ablation studies of our method. Given the ground truth of mesh, we employ both our full model and its variations to reconstruct the ground truth mesh, highlighting significant reconstruction discrepancies with red boxes." + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "image_caption" + } + ], + "index": 6 + }, + { + "type": "image", + "bbox": [ + 188, + 191, + 266, + 281 + ], + "blocks": [ + { + "bbox": [ + 188, + 191, + 266, + 281 + ], + "lines": [ + { + "bbox": [ + 188, + 191, + 266, + 281 + ], + "spans": [ + { + "bbox": [ + 188, + 191, + 266, + 281 + ], + "type": "image", + "image_path": "92f385a5f1b1973336a224a2a8c63ee85ae383a0f4a3a91f2532a6ec85601df8.jpg" + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "image_body" + } + ], + "index": 7 + }, + { + "bbox": [ + 63, + 359, + 295, + 383 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 63, + 359, + 295, + 383 + ], + "spans": [ + { + "bbox": [ + 63, + 359, + 295, + 383 + ], + "type": "text", + "content": "using only uniformly sampled point clouds with Poisson disk sampling [52], while maintaining an equal " + }, + { + "bbox": [ + 63, + 359, + 295, + 383 + ], + "type": "inline_equation", + "content": "N_{d}" + }, + { + "bbox": [ + 63, + 359, + 295, + 383 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 55, + 383, + 295, + 443 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 383, + 295, + 443 + ], + "spans": [ + { + "bbox": [ + 55, + 383, + 295, + 443 + ], + "type": "text", + "content": "- Ours w/o DCA. This variant retains SES but removes DCA, i.e. using a single cross attention adopted by [54]. As shown in Figure 6 and Table 1, our full model consistently outperforms these variants, validating the effectiveness of both components." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 55, + 453, + 235, + 468 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 453, + 235, + 468 + ], + "spans": [ + { + "bbox": [ + 55, + 453, + 235, + 468 + ], + "type": "text", + "content": "5. Application: Single Image to 3D" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 55, + 474, + 295, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 474, + 295, + 713 + ], + "spans": [ + { + "bbox": [ + 55, + 474, + 295, + 713 + ], + "type": "text", + "content": "We demonstrate the effectiveness of our VAE by applying it to single-image 3D generation through diffusion models. Following CLAY [56], we implement a latent diffusion model based on the DiT [33] architecture. For a fair comparison, we fine-tune Craftsman-VAE on our dataset (denoted as Craftsman-VAE†) since both Craftsman-VAE and 3DShape2VecSet were originally trained on smaller datasets. Note that XCube-VAE is excluded from comparison due to its 10,000-dimensional latent codes being impractical for diffusion model training. Figure 7 shows some generation results from diffusion models trained with our Dora-VAE and Craftsman-VAE†. Both models share identical architecture (0.39B parameters) and training conditions (same dataset, 32 A100 GPUs, 3 days). Our Dora-VAE demonstrates significantly better preservation of geometric details in the generated shapes, validating its effectiveness as a foundation for 3D generation tasks. While further improvements could be achieved with more extensive training data and computational resources, we focus on validating the VAE's capabilities in this work and leave such exten" + } + ] + } + ], + "index": 12 + }, + { + "type": "image", + "bbox": [ + 322, + 95, + 389, + 175 + ], + "blocks": [ + { + "bbox": [ + 328, + 73, + 379, + 85 + ], + "lines": [ + { + "bbox": [ + 328, + 73, + 379, + 85 + ], + "spans": [ + { + "bbox": [ + 328, + 73, + 379, + 85 + ], + "type": "text", + "content": "Input Image" + } + ] + } + ], + "index": 13, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 322, + 95, + 389, + 175 + ], + "lines": [ + { + "bbox": [ + 322, + 95, + 389, + 175 + ], + "spans": [ + { + "bbox": [ + 322, + 95, + 389, + 175 + ], + "type": "image", + "image_path": "d526132c9f29f10e84fa0ffa2c3e54ffed14ca3f54d575be6896746582f43c94.jpg" + } + ] + } + ], + "index": 14, + "angle": 0, + "type": "image_body" + } + ], + "index": 14 + }, + { + "type": "image", + "bbox": [ + 400, + 93, + 470, + 175 + ], + "blocks": [ + { + "bbox": [ + 419, + 73, + 440, + 84 + ], + "lines": [ + { + "bbox": [ + 419, + 73, + 440, + 84 + ], + "spans": [ + { + "bbox": [ + 419, + 73, + 440, + 84 + ], + "type": "text", + "content": "Ours" + } + ] + } + ], + "index": 15, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 400, + 93, + 470, + 175 + ], + "lines": [ + { + "bbox": [ + 400, + 93, + 470, + 175 + ], + "spans": [ + { + "bbox": [ + 400, + 93, + 470, + 175 + ], + "type": "image", + "image_path": "a481d437a55e51a85aa0457a52c71c839638cd48e407d959e40364ce87d3b6ce.jpg" + } + ] + } + ], + "index": 16, + "angle": 0, + "type": "image_body" + } + ], + "index": 16 + }, + { + "type": "image", + "bbox": [ + 474, + 93, + 552, + 175 + ], + "blocks": [ + { + "bbox": [ + 484, + 72, + 529, + 83 + ], + "lines": [ + { + "bbox": [ + 484, + 72, + 529, + 83 + ], + "spans": [ + { + "bbox": [ + 484, + 72, + 529, + 83 + ], + "type": "text", + "content": "Craftsman" + } + ] + } + ], + "index": 17, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 474, + 93, + 552, + 175 + ], + "lines": [ + { + "bbox": [ + 474, + 93, + 552, + 175 + ], + "spans": [ + { + "bbox": [ + 474, + 93, + 552, + 175 + ], + "type": "image", + "image_path": "670442b42bbf9d85c9ecec5f31388d75e27fae45204cd0d6eab806a79ee4b776.jpg" + } + ] + } + ], + "index": 18, + "angle": 0, + "type": "image_body" + } + ], + "index": 18 + }, + { + "type": "image", + "bbox": [ + 320, + 182, + 384, + 264 + ], + "blocks": [ + { + "bbox": [ + 320, + 182, + 384, + 264 + ], + "lines": [ + { + "bbox": [ + 320, + 182, + 384, + 264 + ], + "spans": [ + { + "bbox": [ + 320, + 182, + 384, + 264 + ], + "type": "image", + "image_path": "d17e93218df0ade95f0621be106e9dc8d1e6304447621c1f07fe0389f4deb311.jpg" + } + ] + } + ], + "index": 19, + "angle": 0, + "type": "image_body" + } + ], + "index": 19 + }, + { + "type": "image", + "bbox": [ + 399, + 182, + 464, + 261 + ], + "blocks": [ + { + "bbox": [ + 399, + 182, + 464, + 261 + ], + "lines": [ + { + "bbox": [ + 399, + 182, + 464, + 261 + ], + "spans": [ + { + "bbox": [ + 399, + 182, + 464, + 261 + ], + "type": "image", + "image_path": "e15693c532de6a739e4452ddd77ea78328d3dc7ad71ea47c61cdc43a4948231a.jpg" + } + ] + } + ], + "index": 20, + "angle": 0, + "type": "image_body" + } + ], + "index": 20 + }, + { + "type": "image", + "bbox": [ + 479, + 181, + 545, + 264 + ], + "blocks": [ + { + "bbox": [ + 479, + 181, + 545, + 264 + ], + "lines": [ + { + "bbox": [ + 479, + 181, + 545, + 264 + ], + "spans": [ + { + "bbox": [ + 479, + 181, + 545, + 264 + ], + "type": "image", + "image_path": "09f313a12a61443cfb6b32a96dbaf984a6564a0029c8c124a1ef86939b10d737.jpg" + } + ] + } + ], + "index": 21, + "angle": 0, + "type": "image_body" + } + ], + "index": 21 + }, + { + "type": "image", + "bbox": [ + 321, + 270, + 384, + 354 + ], + "blocks": [ + { + "bbox": [ + 321, + 270, + 384, + 354 + ], + "lines": [ + { + "bbox": [ + 321, + 270, + 384, + 354 + ], + "spans": [ + { + "bbox": [ + 321, + 270, + 384, + 354 + ], + "type": "image", + "image_path": "d1d0e13606feabfc15e3f1121b3ded86c314dba301781fed07a821d9a34517a2.jpg" + } + ] + } + ], + "index": 22, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 313, + 369, + 555, + 415 + ], + "lines": [ + { + "bbox": [ + 313, + 369, + 555, + 415 + ], + "spans": [ + { + "bbox": [ + 313, + 369, + 555, + 415 + ], + "type": "text", + "content": "Figure 7. The diffusion results of the single image to 3D generation trained on our Dora-VAE and Craftsman†. The 3D geometry generated by the diffusion model trained on our proposed Dora-VAE has more details under the same experimental environment." + } + ] + } + ], + "index": 25, + "angle": 0, + "type": "image_caption" + } + ], + "index": 22 + }, + { + "type": "image", + "bbox": [ + 399, + 268, + 464, + 353 + ], + "blocks": [ + { + "bbox": [ + 399, + 268, + 464, + 353 + ], + "lines": [ + { + "bbox": [ + 399, + 268, + 464, + 353 + ], + "spans": [ + { + "bbox": [ + 399, + 268, + 464, + 353 + ], + "type": "image", + "image_path": "ea2c630b92887b7c8376bc23c0c84accb0065c03c74bbd2df2e63347c830857a.jpg" + } + ] + } + ], + "index": 23, + "angle": 0, + "type": "image_body" + } + ], + "index": 23 + }, + { + "type": "image", + "bbox": [ + 477, + 270, + 541, + 353 + ], + "blocks": [ + { + "bbox": [ + 477, + 270, + 541, + 353 + ], + "lines": [ + { + "bbox": [ + 477, + 270, + 541, + 353 + ], + "spans": [ + { + "bbox": [ + 477, + 270, + 541, + 353 + ], + "type": "image", + "image_path": "f85fb20808f7c52a8c63c90913f1a53d230dcbbd22d4180cf275b7ce2090701a.jpg" + } + ] + } + ], + "index": 24, + "angle": 0, + "type": "image_body" + } + ], + "index": 24 + }, + { + "bbox": [ + 314, + 418, + 404, + 429 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 314, + 418, + 404, + 429 + ], + "spans": [ + { + "bbox": [ + 314, + 418, + 404, + 429 + ], + "type": "text", + "content": "sions for future work." + } + ] + } + ], + "index": 26 + }, + { + "bbox": [ + 313, + 441, + 388, + 453 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 441, + 388, + 453 + ], + "spans": [ + { + "bbox": [ + 313, + 441, + 388, + 453 + ], + "type": "text", + "content": "6. Conclusion" + } + ] + } + ], + "index": 27 + }, + { + "bbox": [ + 313, + 460, + 555, + 688 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 460, + 555, + 688 + ], + "spans": [ + { + "bbox": [ + 313, + 460, + 555, + 688 + ], + "type": "text", + "content": "In this work, we introduce Dora-VAE, a novel VAE designed for high-quality 3D shape compression and reconstruction. At its core, Dora-VAE introduces sharp edge sampling to effectively capture salient geometric features, complemented by a dual cross-attention architecture that enhances the encoding of these detail-rich point clouds. To enable more rigorous evaluation of VAE performance, we develop Dora-bench, which systematically categorizes shapes based on geometric complexity and introduces the Sharp Normal Error (SNE) metric for specifically assessing the preservation of fine geometric details. Our comprehensive experiments demonstrate that Dora-VAE significantly outperforms existing methods across varying levels of shape complexity. Furthermore, we show that the improved reconstruction capability of Dora-VAE directly enhances the quality of downstream tasks by applying it to single-image 3D generation. The superior performance in generating geometric details validates our approach of focusing on salient region sampling and encoding for 3D VAE design." + } + ] + } + ], + "index": 28 + }, + { + "bbox": [ + 313, + 688, + 554, + 712 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 688, + 554, + 712 + ], + "spans": [ + { + "bbox": [ + 313, + 688, + 554, + 712 + ], + "type": "text", + "content": "Acknowledgements. This work is partially supported by the project L0751 between Bytedance and HKUST." + } + ] + } + ], + "index": 29 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "text", + "content": "16258" + } + ] + } + ], + "index": 30 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 7 + }, + { + "para_blocks": [ + { + "bbox": [ + 56, + 71, + 115, + 83 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 71, + 115, + 83 + ], + "spans": [ + { + "bbox": [ + 56, + 71, + 115, + 83 + ], + "type": "text", + "content": "References" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 57, + 91, + 295, + 713 + ], + "type": "list", + "angle": 0, + "index": 14, + "blocks": [ + { + "bbox": [ + 61, + 91, + 294, + 113 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 61, + 91, + 294, + 113 + ], + "spans": [ + { + "bbox": [ + 61, + 91, + 294, + 113 + ], + "type": "text", + "content": "[1] Digital twin catalog. META, 2024. https://www.projectaria.com/datasets/dtc/.5" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 61, + 114, + 295, + 170 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 61, + 114, + 295, + 170 + ], + "spans": [ + { + "bbox": [ + 61, + 114, + 295, + 170 + ], + "type": "text", + "content": "[2] Yukang Cao, Yan-Pei Cao, Kai Han, Ying Shan, and KwanYee K. Wong. Dreamavatar: Text-and-shape guided 3d human avatar generation via diffusion models. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 958-968, 2024. 1" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 62, + 171, + 295, + 214 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 171, + 295, + 214 + ], + "spans": [ + { + "bbox": [ + 62, + 171, + 295, + 214 + ], + "type": "text", + "content": "[3] Yukang Cao, Liang Pan, Kai Han, Kwan-Yee K Wong, and Ziwei Liu. Avatargo: Zero-shot 4d human-object interaction generation and animation. arXiv preprint arXiv:2410.07164, 2024. 1" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 62, + 217, + 294, + 270 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 217, + 294, + 270 + ], + "spans": [ + { + "bbox": [ + 62, + 217, + 294, + 270 + ], + "type": "text", + "content": "[4] Junsong Chen, Jincheng Yu, Chongjian Ge, Lewei Yao, Enze Xie, Yue Wu, Zhongdao Wang, James Kwok, Ping Luo, Huchuan Lu, and Zhenguo Li. Pixart-α: Fast training of diffusion transformer for photorealistic text-to-image synthesis, 2023. 1" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 62, + 273, + 294, + 316 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 273, + 294, + 316 + ], + "spans": [ + { + "bbox": [ + 62, + 273, + 294, + 316 + ], + "type": "text", + "content": "[5] Junsong Chen, Yue Wu, Simian Luo, Enze Xie, Sayak Paul, Ping Luo, Hang Zhao, and Zhenguo Li. Pixart-δ: Fast and controllable image generation with latent consistency models, 2024. 1" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 62, + 319, + 294, + 372 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 319, + 294, + 372 + ], + "spans": [ + { + "bbox": [ + 62, + 319, + 294, + 372 + ], + "type": "text", + "content": "[6] Rui Chen, Yongwei Chen, Ningxin Jiao, and Kui Jia. Fantasia3d: Disentangling geometry and appearance for high-quality text-to-3d content creation. In Proceedings of the IEEE/CVF International Conference on Computer Vision (ICCV), 2023. 3" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 62, + 374, + 294, + 430 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 374, + 294, + 430 + ], + "spans": [ + { + "bbox": [ + 62, + 374, + 294, + 430 + ], + "type": "text", + "content": "[7] Rui Chen, Mingyi Shi, Shaoli Huang, Ping Tan, Taku Komura, and Xuelin Chen. Taming diffusion probabilistic models for character control. In ACM SIGGRAPH 2024 Conference Papers, New York, NY, USA, 2024. Association for Computing Machinery. 1" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 62, + 431, + 294, + 463 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 431, + 294, + 463 + ], + "spans": [ + { + "bbox": [ + 62, + 431, + 294, + 463 + ], + "type": "text", + "content": "[8] Tianqi Chen, Bing Xu, Chiyuan Zhang, and Carlos Guestrin. Training deep nets with sublinear memory cost. CoRR, abs/1604.06174, 2016. 5" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 62, + 465, + 294, + 509 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 465, + 294, + 509 + ], + "spans": [ + { + "bbox": [ + 62, + 465, + 294, + 509 + ], + "type": "text", + "content": "[9] Yongwei Chen, Rui Chen, Jiabao Lei, Yabin Zhang, and Kui Jia. Tango: Text-driven photorealistic and robust 3d stylization via lighting decomposition. Advances in Neural Information Processing Systems, 35:30923-30936, 2022. 1" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 57, + 511, + 294, + 586 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 511, + 294, + 586 + ], + "spans": [ + { + "bbox": [ + 57, + 511, + 294, + 586 + ], + "type": "text", + "content": "[10] Jasmine Collins, Shubham Goel, Kenan Deng, Achleshwar Luthra, Leon Xu, Erhan Gundogdu, Xi Zhang, Tomas F Yago Vicente, Thomas Dideriksen, Himanshu Arora, et al. Abo: Dataset and benchmarks for real-world 3d object understanding. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 21126-21136, 2022. 5" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 57, + 590, + 294, + 633 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 590, + 294, + 633 + ], + "spans": [ + { + "bbox": [ + 57, + 590, + 294, + 633 + ], + "type": "text", + "content": "[11] Tri Dao. Flashattention-2: Faster attention with better parallelism and work partitioning. In The Twelfth International Conference on Learning Representations, ICLR 2024, Vienna, Austria, May 7-11, 2024. OpenReview.net, 2024. 5" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 57, + 635, + 294, + 689 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 635, + 294, + 689 + ], + "spans": [ + { + "bbox": [ + 57, + 635, + 294, + 689 + ], + "type": "text", + "content": "[12] Matt Deitke, Dustin Schwenk, Jordi Salvador, Luca Weihs, Oscar Michel, Eli VanderBilt, Ludwig Schmidt, Kiana Ehsani, Aniruddha Kembhavi, and Ali Farhadi. Objverse: A universe of annotated 3d objects. arXiv preprint arXiv:2212.08051, 2022. 5" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 57, + 691, + 294, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 691, + 294, + 713 + ], + "spans": [ + { + "bbox": [ + 57, + 691, + 294, + 713 + ], + "type": "text", + "content": "[13] Laura Downs, Anthony Francis, Nate Koenig, Brandon Kinman, Ryan Hickman, Krista Reymann, Thomas B McHugh," + } + ] + } + ], + "index": 13 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 317, + 73, + 553, + 713 + ], + "type": "list", + "angle": 0, + "index": 29, + "blocks": [ + { + "bbox": [ + 335, + 73, + 553, + 117 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 335, + 73, + 553, + 117 + ], + "spans": [ + { + "bbox": [ + 335, + 73, + 553, + 117 + ], + "type": "text", + "content": "and Vincent Vanhoucke. Google scanned objects: A high-quality dataset of 3d scanned household items. In 2022 International Conference on Robotics and Automation (ICRA), pages 2553-2560. IEEE, 2022. 5" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 317, + 119, + 553, + 161 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 119, + 553, + 161 + ], + "spans": [ + { + "bbox": [ + 317, + 119, + 553, + 161 + ], + "type": "text", + "content": "[14] Fabian Groh, Patrick Wieschollek, and Hendrik P. A. Lensch. Flex-convolution (million-scale point-cloud learning beyond grid-worlds). In *Asian Conference on Computer Vision (ACCV)*, 2018. 2, 3" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 317, + 163, + 553, + 195 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 163, + 553, + 195 + ], + "spans": [ + { + "bbox": [ + 317, + 163, + 553, + 195 + ], + "type": "text", + "content": "[15] Jonathan Ho, Ajay Jain, and Pieter Abbeel. Denoising diffusion probabilistic models. arXiv preprint arxiv:2006.11239, 2020. 1" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 317, + 198, + 553, + 262 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 198, + 553, + 262 + ], + "spans": [ + { + "bbox": [ + 317, + 198, + 553, + 262 + ], + "type": "text", + "content": "[16] Yicong Hong, Kai Zhang, Juixiang Gu, Sai Bi, Yang Zhou, Difan Liu, Feng Liu, Kalyan Sunkavalli, Trung Bui, and Hao Tan. LRM: large reconstruction model for single image to 3d. In The Twelfth International Conference on Learning Representations, ICLR 2024, Vienna, Austria, May 7-11, 2024. OpenReview.net, 2024. 3" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 317, + 264, + 553, + 307 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 264, + 553, + 307 + ], + "spans": [ + { + "bbox": [ + 317, + 264, + 553, + 307 + ], + "type": "text", + "content": "[17] Yicong Hong, Kai Zhang, Juixiang Gu, Sai Bi, Yang Zhou, Difan Liu, Feng Liu, Kalyan Sunkavalli, Trung Bui, and Hao Tan. Lrm: Large reconstruction model for single image to 3d, 2024. 1" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 317, + 309, + 553, + 363 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 309, + 553, + 363 + ], + "spans": [ + { + "bbox": [ + 317, + 309, + 553, + 363 + ], + "type": "text", + "content": "[18] Jiahui Huang, Zan Gojcic, Matan Atzmon, Or Litany, Sanja Fidler, and Francis Williams. Neural kernel surface reconstruction. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 4369-4379, 2023. 7" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 317, + 365, + 553, + 407 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 365, + 553, + 407 + ], + "spans": [ + { + "bbox": [ + 317, + 365, + 553, + 407 + ], + "type": "text", + "content": "[19] Bernhard Kerbl, Georgios Kopanas, Thomas Leimkuhler, and George Drettakis. 3d gaussian splatting for real-time radiance field rendering. ACM Trans. Graph., 42(4):139-1, 2023. 3" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 317, + 411, + 553, + 442 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 411, + 553, + 442 + ], + "spans": [ + { + "bbox": [ + 317, + 411, + 553, + 442 + ], + "type": "text", + "content": "[20] Arno Knapitsch, Jaesik Park, Qian-Yi Zhou, and Vladlen Koltun. Tanks and temples: benchmarking large-scale scene reconstruction. ACM Trans. Graph., 36(4), 2017. 7" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 317, + 445, + 553, + 487 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 445, + 553, + 487 + ], + "spans": [ + { + "bbox": [ + 317, + 445, + 553, + 487 + ], + "type": "text", + "content": "[21] Yushi Lan, Fangzhou Hong, Shuai Yang, Shangchen Zhou, Xuyi Meng, Bo Dai, Xingang Pan, and Chen Change Loy. Ln3diff: Scalable latent neural fields diffusion for speedy 3d generation. In ECCV, 2024. 1" + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 317, + 490, + 553, + 544 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 490, + 553, + 544 + ], + "spans": [ + { + "bbox": [ + 317, + 490, + 553, + 544 + ], + "type": "text", + "content": "[22] Jiahao Li, Hao Tan, Kai Zhang, Zexiang Xu, Fujun Luan, Yinghao Xu, Yicong Hong, Kalyan Sunkavalli, Greg Shakhnarovich, and Sai Bi. Instant3d: Fast text-to-3d with sparse-view generation and large reconstruction model. arXiv preprint arXiv:2311.06214, 2023. 3" + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 317, + 545, + 553, + 589 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 545, + 553, + 589 + ], + "spans": [ + { + "bbox": [ + 317, + 545, + 553, + 589 + ], + "type": "text", + "content": "[23] Weiyu Li, Rui Chen, Xuelin Chen, and Ping Tan. Sweetdreamer: Aligning geometric priors in 2d diffusion for consistent text-to-3d. International Conference on Learning Representations (ICLR), 2024. 3" + } + ] + } + ], + "index": 25 + }, + { + "bbox": [ + 317, + 590, + 553, + 644 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 590, + 553, + 644 + ], + "spans": [ + { + "bbox": [ + 317, + 590, + 553, + 644 + ], + "type": "text", + "content": "[24] Weiyu Li, Jiarui Liu, Rui Chen, Yixun Liang, Xuelin Chen, Ping Tan, and Xiaoxiao Long. Craftsman: High-fidelity mesh generation with 3d native generation and interactive geometry refiner. arXiv preprint arXiv:2405.14979, 2024. 1, 3, 7" + } + ] + } + ], + "index": 26 + }, + { + "bbox": [ + 317, + 647, + 553, + 689 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 647, + 553, + 689 + ], + "spans": [ + { + "bbox": [ + 317, + 647, + 553, + 689 + ], + "type": "text", + "content": "[25] Yixun Liang, Xin Yang, Jiantao Lin, Haodong Li, Xiaogang Xu, and Yingcong Chen. Luciddreamer: Towards high-fidelity text-to-3d generation via interval score matching. arXiv preprint arXiv:2311.11284, 2023. 1" + } + ] + } + ], + "index": 27 + }, + { + "bbox": [ + 317, + 692, + 553, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 692, + 553, + 713 + ], + "spans": [ + { + "bbox": [ + 317, + 692, + 553, + 713 + ], + "type": "text", + "content": "[26] Chen-Hsuan Lin, Jun Gao, Luming Tang, Towaki Takikawa, Xiaohui Zeng, Xun Huang, Karsten Kreis, Sanja Fidler," + } + ] + } + ], + "index": 28 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "text", + "content": "16259" + } + ] + } + ], + "index": 30 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 8 + }, + { + "para_blocks": [ + { + "bbox": [ + 56, + 72, + 295, + 712 + ], + "type": "list", + "angle": 0, + "index": 13, + "blocks": [ + { + "bbox": [ + 76, + 72, + 294, + 106 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 76, + 72, + 294, + 106 + ], + "spans": [ + { + "bbox": [ + 76, + 72, + 294, + 106 + ], + "type": "text", + "content": "Ming-Yu Liu, and Tsung-Yi Lin. Magic3d: High-resolution text-to-3d content creation. In IEEE Conference on Computer Vision and Pattern Recognition (CVPR), 2023. 3" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 56, + 108, + 295, + 140 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 108, + 295, + 140 + ], + "spans": [ + { + "bbox": [ + 56, + 108, + 295, + 140 + ], + "type": "text", + "content": "[27] Fangfu Liu, Diankun Wu, Yi Wei, Yongming Rao, and Yueqi Duan. Sherpa3d: Boosting high-fidelity text-to-3d generation via coarse 3d prior, 2023. 3" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 56, + 143, + 294, + 186 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 143, + 294, + 186 + ], + "spans": [ + { + "bbox": [ + 56, + 143, + 294, + 186 + ], + "type": "text", + "content": "[28] Minghua Liu, Chao Xu, Haian Jin, Linghao Chen, Mukund Varma T, Zexiang Xu, and Hao Su. One-2-3-45: Any single image to 3d mesh in 45 seconds without per-shape optimization. 2024. 1" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 56, + 189, + 294, + 243 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 189, + 294, + 243 + ], + "spans": [ + { + "bbox": [ + 56, + 189, + 294, + 243 + ], + "type": "text", + "content": "[29] Minghua Liu, Chong Zeng, Xinyue Wei, Ruoxi Shi, Linghao Chen, Chao Xu, Mengqi Zhang, Zhaoning Wang, Xiaoshuai Zhang, Isabella Liu, Hongzhi Wu, and Hao Su. Meshformer: High-quality mesh generation with 3d-guided reconstruction model. arXiv preprint arXiv:2408.10198, 2024. 3" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 56, + 246, + 294, + 290 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 246, + 294, + 290 + ], + "spans": [ + { + "bbox": [ + 56, + 246, + 294, + 290 + ], + "type": "text", + "content": "[30] Yuan Liu, Cheng Lin, Zijiao Zeng, Xiaoxiao Long, Lingjie Liu, Taku Komura, and Wenping Wang. Syncdreamer: Generating multiview-consistent images from a single-view image. arXiv preprint arXiv:2309.03453, 2023. 1, 3" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 56, + 292, + 294, + 346 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 292, + 294, + 346 + ], + "spans": [ + { + "bbox": [ + 56, + 292, + 294, + 346 + ], + "type": "text", + "content": "[31] Ben Mildenhall, Pratul P. Srinivasan, Matthew Tancik, Jonathan T. Barron, Ravi Ramamoorthi, and Ren Ng. Nerf: Representing scenes as neural radiance fields for view synthesis. In European Conference on Computer Vision (ECCV), 2020. 3" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 56, + 350, + 294, + 382 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 350, + 294, + 382 + ], + "spans": [ + { + "bbox": [ + 56, + 350, + 294, + 382 + ], + "type": "text", + "content": "[32] Carsten Moenning and Neil A Dodgson. Fast marching farthest point sampling. Technical report, University of Cambridge, Computer Laboratory, 2003. 3" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 56, + 384, + 294, + 427 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 384, + 294, + 427 + ], + "spans": [ + { + "bbox": [ + 56, + 384, + 294, + 427 + ], + "type": "text", + "content": "[33] William Peebles and Saining Xie. Scalable diffusion models with transformers. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 4195-4205, 2023. 8" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 56, + 430, + 294, + 462 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 430, + 294, + 462 + ], + "spans": [ + { + "bbox": [ + 56, + 430, + 294, + 462 + ], + "type": "text", + "content": "[34] Ben Poole, Ajay Jain, Jonathan T. Barron, and Ben Mildenhall. Dreamfusion: Text-to-3d using 2d diffusion. arXiv, 2022. 3" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 56, + 465, + 294, + 542 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 465, + 294, + 542 + ], + "spans": [ + { + "bbox": [ + 56, + 465, + 294, + 542 + ], + "type": "text", + "content": "[35] Lingteng Qiu, Guanying Chen, Xiaodong Gu, Qi Zuo, Mutian Xu, Yushuang Wu, Weihao Yuan, Zilong Dong, Liefeng Bo, and Xiaoguang Han. Richdreamer: A generalizable normal-depth diffusion model for detail richness in text-to-3d. In IEEE/CVF Conference on Computer Vision and Pattern Recognition, CVPR 2024, Seattle, WA, USA, June 16-22, 2024, pages 9914-9925. IEEE, 2024. 3" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 56, + 544, + 294, + 599 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 544, + 294, + 599 + ], + "spans": [ + { + "bbox": [ + 56, + 544, + 294, + 599 + ], + "type": "text", + "content": "[36] Xuanchi Ren, Jiahui Huang, Xiaohui Zeng, Ken Museth, Sanja Fidler, and Francis Williams. Xcube: Large-scale 3d generative modeling using sparse voxel hierarchies. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 4209-4219, 2024. 1, 3, 7" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 56, + 601, + 294, + 655 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 601, + 294, + 655 + ], + "spans": [ + { + "bbox": [ + 56, + 601, + 294, + 655 + ], + "type": "text", + "content": "[37] Robin Rombach, Andreas Blattmann, Dominik Lorenz, Patrick Esser, and Björn Ommer. High-resolution image synthesis with latent diffusion models. In Conference on Computer Vision and Pattern Recognition (CVPR), pages 10684-10695, 2022. 1" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 56, + 658, + 294, + 712 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 658, + 294, + 712 + ], + "spans": [ + { + "bbox": [ + 56, + 658, + 294, + 712 + ], + "type": "text", + "content": "[38] Tianchang Shen, Jun Gao, Kangxue Yin, Ming-Yu Liu, and Sanja Fidler. Deep marching tetrahedra: a hybrid representation for high-resolution 3d shape synthesis. Advances in Neural Information Processing Systems, 34:6087-6101, 2021. 3" + } + ] + } + ], + "index": 12 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 316, + 72, + 553, + 713 + ], + "type": "list", + "angle": 0, + "index": 27, + "blocks": [ + { + "bbox": [ + 316, + 72, + 553, + 106 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 72, + 553, + 106 + ], + "spans": [ + { + "bbox": [ + 316, + 72, + 553, + 106 + ], + "type": "text", + "content": "[39] Yichun Shi, Peng Wang, Jianglong Ye, Long Mai, Kejie Li, and Xiao Yang. Mvdream: Multi-view diffusion for 3d generation. arXiv:2308.16512, 2023. 3" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 316, + 108, + 553, + 150 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 108, + 553, + 150 + ], + "spans": [ + { + "bbox": [ + 316, + 108, + 553, + 150 + ], + "type": "text", + "content": "[40] Jiaxiang Tang, Zhaoxi Chen, Xiaokang Chen, Tengfei Wang, Gang Zeng, and Ziwei Liu. Lgm: Large multi-view gaussian model for high-resolution 3d content creation. arXiv preprint arXiv:2402.05054, 2024. 3" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 316, + 152, + 553, + 207 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 152, + 553, + 207 + ], + "spans": [ + { + "bbox": [ + 316, + 152, + 553, + 207 + ], + "type": "text", + "content": "[41] Jiaxiang Tang, Jiawei Ren, Hang Zhou, Ziwei Liu, and Gang Zeng. Dreamgaussian: Generative gaussian splatting for efficient 3d content creation. In The Twelfth International Conference on Learning Representations, ICLR 2024, Vienna, Austria, May 7-11, 2024. OpenReview.net, 2024. 3" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 316, + 209, + 553, + 262 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 209, + 553, + 262 + ], + "spans": [ + { + "bbox": [ + 316, + 209, + 553, + 262 + ], + "type": "text", + "content": "[42] Dmitry Tochilkin, David Pankratz, Zexiang Liu, Zixuan Huang, , Adam Letts, Yangguang Li, Ding Liang, Christian Laforte, Varun Jampani, and Yan-Pei Cao. Triposr: Fast 3d object reconstruction from a single image. arXiv preprint arXiv:2403.02151, 2024. 1" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 316, + 264, + 553, + 319 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 264, + 553, + 319 + ], + "spans": [ + { + "bbox": [ + 316, + 264, + 553, + 319 + ], + "type": "text", + "content": "[43] Peng Wang, Hao Tan, Sai Bi, Yinghao Xu, Fujun Luan, Kalyan Sunkavalli, Wenping Wang, Zexiang Xu, and Kai Zhang. Pf-lrm: Pose-free large reconstruction model for joint pose and shape prediction. arXiv preprint arXiv:2311.12024, 2023. 3" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 316, + 321, + 553, + 374 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 321, + 553, + 374 + ], + "spans": [ + { + "bbox": [ + 316, + 321, + 553, + 374 + ], + "type": "text", + "content": "[44] Tengfei Wang, Bo Zhang, Ting Zhang, Shuyang Gu, Jianmin Bao, Tadas Baltrusaitis, Jingjing Shen, Dong Chen, Fang Wen, Qifeng Chen, and Baining Guo. Rodin: A generative model for sculpting 3d digital avatars using diffusion, 2022. 1" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 316, + 376, + 553, + 421 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 376, + 553, + 421 + ], + "spans": [ + { + "bbox": [ + 316, + 376, + 553, + 421 + ], + "type": "text", + "content": "[45] Zhengyi Wang, Cheng Lu, Yikai Wang, Fan Bao, Chongxuan Li, Hang Su, and Jun Zhu. Prolificdreamer: High-fidelity and diverse text-to-3d generation with variational score distillation. arXiv preprint arXiv:2305.16213, 2023. 3" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 316, + 422, + 553, + 475 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 422, + 553, + 475 + ], + "spans": [ + { + "bbox": [ + 316, + 422, + 553, + 475 + ], + "type": "text", + "content": "[46] Zhengyi Wang, Yikai Wang, Yifei Chen, Chengdong Xiang, Shuo Chen, Dajiang Yu, Chongxuan Li, Hang Su, and Jun Zhu. Crm: Single image to 3d textured mesh with convolutional reconstruction model. arXiv preprint arXiv:2403.05034, 2024. 3" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 316, + 478, + 553, + 553 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 478, + 553, + 553 + ], + "spans": [ + { + "bbox": [ + 316, + 478, + 553, + 553 + ], + "type": "text", + "content": "[47] Francis Williams, Jiahui Huang, Jonathan Swartz, Gergely Klar, Vijay Thakkar, Matthew Cong, Xuanchi Ren, Rui-long Li, Clement Fuji-Tsang, Sanja Fidler, Eftychios Sifakis, and Ken Museth. fvdb: A deep-learning framework for sparse, large-scale, and high-performance spatial intelligence. ACM Transactions on Graphics (TOG), 43(4):133:1-133:15, 2024. 1" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 316, + 556, + 553, + 599 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 556, + 553, + 599 + ], + "spans": [ + { + "bbox": [ + 316, + 556, + 553, + 599 + ], + "type": "text", + "content": "[48] Chengzhi Wu, Junwei Zheng, Julius Pfrommer, and Jürgen Beyerer. Attention-based point cloud edge sampling. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), 2023. 2, 3" + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 316, + 601, + 553, + 644 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 601, + 553, + 644 + ], + "spans": [ + { + "bbox": [ + 316, + 601, + 553, + 644 + ], + "type": "text", + "content": "[49] Shuang Wu, Youtian Lin, Feihu Zhang, Yifei Zeng, Jingxi Xu, Philip Torr, Xun Cao, and Yao Yao. Direct3d: Scalable image-to-3d generation via 3d latent diffusion transformer. arXiv:2405.14832, 2024. 1, 2, 3, 7" + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 316, + 647, + 553, + 689 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 647, + 553, + 689 + ], + "spans": [ + { + "bbox": [ + 316, + 647, + 553, + 689 + ], + "type": "text", + "content": "[50] Bojun Xiong, Si-Tong Wei, Xin-Yang Zheng, Yan-Pei Cao, Zhouhui Lian, and Peng-Shuai Wang. OctFusion: Octree-based diffusion models for 3d shape generation. arXiv, 2024. 3" + } + ] + } + ], + "index": 25 + }, + { + "bbox": [ + 316, + 691, + 553, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 691, + 553, + 713 + ], + "spans": [ + { + "bbox": [ + 316, + 691, + 553, + 713 + ], + "type": "text", + "content": "[51] Jiale Xu, Weihao Cheng, Yiming Gao, Xintao Wang, Shenghua Gao, and Ying Shan. Instantmesh: Efficient 3d" + } + ] + } + ], + "index": 26 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "type": "text", + "content": "16260" + } + ] + } + ], + "index": 28 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 9 + }, + { + "para_blocks": [ + { + "bbox": [ + 56, + 72, + 295, + 418 + ], + "type": "list", + "angle": 0, + "index": 8, + "blocks": [ + { + "bbox": [ + 75, + 72, + 294, + 105 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 75, + 72, + 294, + 105 + ], + "spans": [ + { + "bbox": [ + 75, + 72, + 294, + 105 + ], + "type": "text", + "content": "mesh generation from a single image with sparse-view large reconstruction models. arXiv preprint arXiv:2404.07191, 2024.3" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 56, + 106, + 295, + 140 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 106, + 295, + 140 + ], + "spans": [ + { + "bbox": [ + 56, + 106, + 295, + 140 + ], + "type": "text", + "content": "[52] Cem Yuksel. Sample elimination for generating poisson disk sample sets. The Eurographs Association & John Wiley & Sons, Ltd., 2015. 3, 8" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 56, + 141, + 294, + 174 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 141, + 294, + 174 + ], + "spans": [ + { + "bbox": [ + 56, + 141, + 294, + 174 + ], + "type": "text", + "content": "[53] Biao Zhang and Peter Wonka. Lagem: A large geometry model for 3d representation learning and diffusion. arXiv preprint arXiv:2410.01295, 2024. 3" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 56, + 175, + 294, + 218 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 175, + 294, + 218 + ], + "spans": [ + { + "bbox": [ + 56, + 175, + 294, + 218 + ], + "type": "text", + "content": "[54] Biao Zhang, Jiapeng Tang, Matthias Nießner, and Peter Wonka. 3dshape2vecset: A 3d shape representation for neural fields and generative diffusion models. ACM Transactions on Graphics (SIGGRAPH), 42(4), 2023. 1, 3, 4, 5, 7, 8" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 56, + 220, + 294, + 251 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 220, + 294, + 251 + ], + "spans": [ + { + "bbox": [ + 56, + 220, + 294, + 251 + ], + "type": "text", + "content": "[55] Lvmin Zhang, Anyi Rao, and Maneesh Agrawala. Adding conditional control to text-to-image diffusion models, 2023. 1" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 56, + 254, + 294, + 308 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 254, + 294, + 308 + ], + "spans": [ + { + "bbox": [ + 56, + 254, + 294, + 308 + ], + "type": "text", + "content": "[56] Longwen Zhang, Ziyu Wang, Qixuan Zhang, Qiwei Qiu, Anqi Pang, Haoran Jiang, Wei Yang, Lan Xu, and Jingyi Yu. Clay: A controllable large-scale generative model for creating high-quality 3d assets. ACM Transactions on Graphics (TOG), 43(4):1-20, 2024. 1, 2, 3, 5, 7, 8" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 56, + 309, + 294, + 353 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 309, + 294, + 353 + ], + "spans": [ + { + "bbox": [ + 56, + 309, + 294, + 353 + ], + "type": "text", + "content": "[57] Xuying Zhang, Bo-Wen Yin, Yuming Chen, Zheng Lin, Yunheng Li, Qibin Hou, and Ming-Ming Cheng. Temo: Towards text-driven 3d stylization for multi-object meshes. arXiv preprint arXiv:2312.04248, 2023. 1" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 56, + 354, + 294, + 418 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 354, + 294, + 418 + ], + "spans": [ + { + "bbox": [ + 56, + 354, + 294, + 418 + ], + "type": "text", + "content": "[58] Zibo Zhao, Wen Liu, Xin Chen, Xianfang Zeng, Rui Wang, Pei Cheng, BIN FU, Tao Chen, Gang YU, and Shenghua Gao. Michelangelo: Conditional 3d shape generation based on shape-image-text aligned latent representation. In Advances in Neural Information Processing Systems (NeurIPS), 2023. 1, 3" + } + ] + } + ], + "index": 7 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 749, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 749, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 749, + 317, + 757 + ], + "type": "text", + "content": "16261" + } + ] + } + ], + "index": 9 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 10 + } + ], + "_backend": "vlm", + "_version_name": "2.6.4" +} \ No newline at end of file diff --git a/2025/Dr. Splat_ Directly Referring 3D Gaussian Splatting via Direct Language Embedding Registration/04973834-dfc5-4d84-8238-2be21678c244_content_list.json b/2025/Dr. Splat_ Directly Referring 3D Gaussian Splatting via Direct Language Embedding Registration/04973834-dfc5-4d84-8238-2be21678c244_content_list.json new file mode 100644 index 0000000000000000000000000000000000000000..7fc7818c042a6b1f0d63c8da7e34c29a5e7532ba --- /dev/null +++ b/2025/Dr. Splat_ Directly Referring 3D Gaussian Splatting via Direct Language Embedding Registration/04973834-dfc5-4d84-8238-2be21678c244_content_list.json @@ -0,0 +1,1772 @@ +[ + { + "type": "text", + "text": "Dr. Splat: Directly Referring 3D Gaussian Splatting via Direct Language Embedding Registration", + "text_level": 1, + "bbox": [ + 236, + 130, + 759, + 176 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Kim Jun-Seong1", + "bbox": [ + 173, + 202, + 305, + 220 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "GeonU Kim", + "bbox": [ + 344, + 203, + 454, + 220 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Kim Yu-Ji2", + "bbox": [ + 495, + 203, + 586, + 220 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Yu-Chiang Frank Wang3", + "bbox": [ + 627, + 203, + 823, + 220 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Jaesung Choe $^{3*}$", + "bbox": [ + 344, + 220, + 470, + 238 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Tae-Hyun Oh $^{1,2,4*}$", + "bbox": [ + 504, + 220, + 650, + 238 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "$^{1}$ Department of Electrical Engineering and $^{2}$ Graduate School of AI, POSTECH", + "bbox": [ + 184, + 238, + 813, + 256 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "3NVIDIA", + "bbox": [ + 313, + 256, + 393, + 272 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "$^{4}$ School of Computing, KAIST", + "bbox": [ + 436, + 257, + 681, + 273 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "{junseong.kim, gukim, ugkim}@postech.ac.kr", + "bbox": [ + 313, + 277, + 681, + 291 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "{frankwang, jchoe}@nvidia.com taehyun@kaist.ac.kr", + "bbox": [ + 269, + 295, + 728, + 308 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Abstract", + "text_level": 1, + "bbox": [ + 246, + 325, + 325, + 340 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "We introduce Dr. Splat, a novel approach for open-vocabulary 3D scene understanding leveraging 3D Gaussian Splatting. Unlike existing language-embedded 3DGS methods, which rely on a rendering process, our method directly associates language-aligned CLIP embeddings with 3D Gaussians for holistic 3D scene understanding. The key of our method is a language feature registration technique where CLIP embeddings are assigned to the dominant Gaussians intersected by each pixel-ray. Moreover, we integrate Product Quantization (PQ) trained on general large-scale image data to compactly represent embeddings without per-scene optimization. Experiments demonstrate that our approach significantly outperforms existing approaches in 3D perception benchmarks, such as open-vocabulary 3D semantic segmentation, 3D object localization, and 3D object selection tasks. For video results, please visit: https://drsplat.github.io/", + "bbox": [ + 89, + 357, + 483, + 613 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "1. Introduction", + "text_level": 1, + "bbox": [ + 89, + 642, + 220, + 657 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Open-vocabulary 3D scene understanding represents a significant challenge in the field of computer vision, with applications spanning autonomous navigation, robotics, and augmented reality. This approach aims to enable the interpretation and referencing of 3D spatial information through natural language, allowing for applicability beyond a restricted set of predefined categories [2, 3, 28, 29, 33, 36, 40]. Previously, open-vocabulary 3D scene understanding has been explored using point-cloud-based methods [11, 15, 17, 27, 30, 35, 38]. Recently, the 3D Gaussian Splitting (3DGS) [16] has introduced a continuous representation integrated on explicit 3D Gaussians, which differs from traditional point-cloud approaches, enabling rapid progress in practical applications [41]. Current research has begun to explore", + "bbox": [ + 89, + 667, + 483, + 878 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/723f26eaf0279c7ee09a4c6a34fe5122c8fe1a6bf384a89d446ad64dff1aa6dc.jpg", + "image_caption": [ + "Render 2D images" + ], + "image_footnote": [], + "bbox": [ + 514, + 325, + 705, + 441 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/623dc797a4b1fa4dcb95bf70256b6dbedfe49900678e1c6327cee3ea495a5102.jpg", + "image_caption": [ + "B number of images" + ], + "image_footnote": [], + "bbox": [ + 516, + 452, + 705, + 506 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/7cf17d34a06bf835f2ec62c6165e1ed103649f856fa6babc777fcb4cd5454c9e.jpg", + "image_caption": [ + "3D search scheme (ours)", + "'Direct' 3D Search" + ], + "image_footnote": [], + "bbox": [ + 717, + 340, + 901, + 444 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/8f7765ced90b2d82f0f6133bc1fd5eab763b1a5ef1795b1d37fe67a51e9672ad.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 759, + 452, + 861, + 473 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/da76d519fcbeba25815c37e1aea704ca110e3d069c2533019f8af079e618a080.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 779, + 473, + 841, + 512 + ], + "page_idx": 0 + }, + { + "type": "table", + "img_path": "images/f90ce695a0150e2c33afbcbeebc1196415a9955e108061843afad0635b72368c.jpg", + "table_caption": [], + "table_footnote": [], + "table_body": "
Search domainper-scene codebookFeature distill.SearchDB size
LERF [17]2Drequired~ 24hslowlarge
LangSplat [30]2Drequired~ 4hslowlarge
LEGaussians [34]2Drequired~ 4hslowlarge
OpenGaussian [37]3Drequired~ 1hfastsmall
Dr. Splat (Ours)3Dnone~ 10mfastsmall
", + "bbox": [ + 517, + 522, + 903, + 588 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Figure 1. Comparison of 2D (left) vs. our direct 3D search (right) for open-vocabulary 3D scene understanding. The 2D approach relies on multiview rendering, incurring high computational costs. Our method directly links language features to 3D Gaussians, enabling efficient and complete spatial coverage. The table highlights Dr. Splat's superior efficiency over related methods.", + "bbox": [ + 511, + 593, + 906, + 676 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "methods for associating language-based features with 3D Gaussian splats to enhance scene understanding capabilities.", + "bbox": [ + 511, + 689, + 906, + 719 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Several recent approaches [30, 34, 43] introduce 3D Gaussian representation [16] into the open-vocabulary scene understanding. This unique representation uses 3D Gaussians to achieve high-quality scene rendering, offering a more structured representation that addresses some limitations of point clouds. Building on this, these methods employ 2D vision-language models to transfer language knowledge to 3D Gaussians \"via rendered feature maps\".", + "bbox": [ + 511, + 719, + 906, + 839 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Despite its promise, such rendering-based distillation methods [30, 34] share two limitations. First, we found that there is a discrepancy between optimized embeddings in 3D Gaussians and 2D language-aligned embeddings. This gap", + "bbox": [ + 511, + 839, + 908, + 901 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "CVF", + "bbox": [ + 106, + 2, + 181, + 42 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore.", + "bbox": [ + 236, + 0, + 810, + 46 + ], + "page_idx": 0 + }, + { + "type": "page_footnote", + "text": "*Corresponding Authors", + "bbox": [ + 107, + 886, + 241, + 898 + ], + "page_idx": 0 + }, + { + "type": "page_number", + "text": "14137", + "bbox": [ + 480, + 944, + 517, + 955 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "arises mainly from an intermediate rendering step that may distort CLIP embeddings during training. Then, the reliance on rendering impedes holistic 3D scene understanding, additional task-processing such as 3D semantic segmentation and 3D object localization, and making full spatial coverage calculations less efficient than direct 3D Gaussian methods [37] including ours as illustrated in Fig. 1.", + "bbox": [ + 89, + 90, + 483, + 196 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "To address this issue, this work proposes Dr. Splat. Our method bypasses the rendering stage, enabling direct interaction with 3D Gaussians for registering and referring the well-preserved language-aligned CLIP embeddings in the 3D space. This makes our Dr. Splat clearly distinguishable from prior works, facilitating a seamless integration of representative embeddings from 2D vision language models into the 3D spatial structure without compromising exhaustive rendering process that has been exploited [14, 30, 34, 41-43]. Moreover, we propose to use a Product Quantization (PQ) feature encoding method to represent embeddings compactly and efficiently without any per-scene optimization. Rather than storing full-length feature vectors or per-scene specifically compressed embeddings [14, 30, 34, 41-43], each Gaussian in our Dr. Splat stores an index from a pre-trained PQ, significantly reducing memory usage up to $6.25\\%$ compression ratio. By preserving the richness of embeddings while reducing memory usage, PQ is integral to our framework's high scalability and its ability to perform 3D perception tasks, such as open-vocabulary 3D object localization, 3D object selection, and 3D semantic segmentation. Our contributions are summarized as follows:", + "bbox": [ + 91, + 198, + 486, + 527 + ], + "page_idx": 1 + }, + { + "type": "list", + "sub_type": "text", + "list_items": [ + "- We propose Dr. Splat, direct registration and referencing of language-aligned features in 3D Gaussians, bypassing intermediate rendering and preserving feature accuracy.", + "- We introduce the PQ encoding method for compact feature representation, reducing memory usage while maintaining essential 3D feature properties.", + "- We present a novel evaluation protocol to assess accuracy of 3D localization and segmentation for 3D Gaussians, with pseudo-labeling methods and volume-aware metrics." + ], + "bbox": [ + 89, + 532, + 483, + 676 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "2. Related Work and Motivation", + "text_level": 1, + "bbox": [ + 89, + 689, + 366, + 705 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Language-based 3D scene understanding. Open-set 3D scene understanding has seen considerable advancements, with a focus on methods that leverage language knowledge into 3D representation such as point clouds, neural radiance fields (NeRF) [26], and Gaussian Splatting [16] for 3D comprehension. Point-based methods [5, 12, 15, 25, 27, 38, 39] in open-vocabulary settings process point cloud data trained from language embeddings [23, 31] for open-set categories.", + "bbox": [ + 89, + 719, + 483, + 839 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "NeRF-based approaches [6, 17, 18, 21, 24, 32] leverage semantic embeddings from 2D foundation models, such as CLIP [31], LSeg [23] and DINO [1] for open-vocabulary understanding. While the rendering process enhances 2D", + "bbox": [ + 89, + 839, + 483, + 901 + ], + "page_idx": 1 + }, + { + "type": "image", + "img_path": "images/2fdf2de7718c75f0ce02338691c88a40088d2c4d6c2ecbccebfa85da3bf4c2b5.jpg", + "image_caption": [ + "Input Image", + "Figure 2. Visualization of discrepancy in rendered 2D features and 3D features. Color indicates a cosine similarity score between query features from a text query and either (a) 3D features distilled by 2D rendering [30] or (b) directly registered 3D features." + ], + "image_footnote": [], + "bbox": [ + 516, + 90, + 638, + 172 + ], + "page_idx": 1 + }, + { + "type": "image", + "img_path": "images/f0afa253f25e50670761bad93d295c56c96e5835de9bf78550bebad21f8d2a3f.jpg", + "image_caption": [ + "(a) Rendering" + ], + "image_footnote": [], + "bbox": [ + 638, + 90, + 758, + 171 + ], + "page_idx": 1 + }, + { + "type": "image", + "img_path": "images/a727fe67c7febb26a65755ae330ab24c80f1d7a693cd511d07889dcf75e8a6ab.jpg", + "image_caption": [ + "(b) Registration" + ], + "image_footnote": [], + "bbox": [ + 759, + 101, + 880, + 171 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "perception tasks, the implicit nature of NeRF constrains the holistic understanding of 3D structures and dominantly provides 'rendered' feature maps.", + "bbox": [ + 511, + 265, + 906, + 311 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "3D Gaussian Splatting (3DGS) [16] has emerged as a promising rendering method, as well as a novel representation for open-vocabulary 3D scene understanding. Since this research is the close related work with our work, we first elucidate the preliminary of 3DGS, followed by focusing on language embedded 3DGS as follows.", + "bbox": [ + 511, + 311, + 906, + 402 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Preliminary of 3D Gaussian Splatting. 3DGS [16] encodes appearance and geometry of the target scene into the 3D Gaussian representation. Each 3D primitive representation is expressed as a 3D Gaussian distribution having mean $\\pmb{\\mu} = [x_{\\mu},y_{\\mu},z_{\\mu}]^{\\top}$ for 3D position and covariance matrix $\\Sigma_{\\mathrm{3D}}\\in \\mathbb{R}^{3\\times 3}$ for 3D volume, as well as the opacity value $\\alpha$ and the color c. In particular, the covariance matrix is decomposed into the scale matrix $S\\in \\mathbb{R}^{3\\times 3}$ and the rotation matrix $R\\in SO(3)$ , $\\Sigma_{\\mathrm{3D}} = RSS^{\\top}R^{\\top}$ . In brief, $N$ numbers of 3D Gaussians can be parametrized as $\\Theta = \\{\\pmb {\\theta}_i\\}_{i = 1}^N = \\{\\pmb {\\mu}_i,S_i,R_i,\\alpha_i,\\mathbf{c}_i\\}_{i = 1}^N$ . 3D Gaussians $\\Theta$ are used to render a 2D pixel color c computed as:", + "bbox": [ + 511, + 406, + 908, + 589 + ], + "page_idx": 1 + }, + { + "type": "equation", + "text": "\n$$\n\\hat {\\mathbf {c}} (\\theta) = \\sum_ {i = 1} ^ {N} T _ {i} \\tilde {\\alpha} _ {i} \\mathbf {c} _ {i}, \\text {s . t .} \\tilde {\\alpha} _ {i} = \\alpha_ {i} \\exp \\left(- \\frac {1}{2} \\mathbf {d} ^ {\\top} \\Sigma_ {2 D} ^ {- 1} \\mathbf {d}\\right), \\tag {1}\n$$\n", + "text_format": "latex", + "bbox": [ + 521, + 599, + 906, + 628 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "$T_{i}$ is a transmittance, $\\tilde{\\alpha}_{i}$ is an effective opacity value computed from the Gaussian's opacity $\\alpha$ , the pixel distance $\\mathbf{d} \\in \\mathbb{R}^{2 \\times 1}$ from the target pixel to the projected center location of the Gaussian in pixel, and $\\Sigma_{2D}$ is the 2D covariance matrix in the image domain obtained from the splatting algorithm [16, 44]. The 3D Gaussian parameters $\\Theta$ of a scene are optimized by minimizing the rendering loss between the input image color $\\mathbf{c}$ and the rendered color $\\hat{\\mathbf{c}}(\\theta)$ in Eq. (1) as $\\arg \\min_{\\theta} \\| \\mathbf{c} - \\hat{\\mathbf{c}}(\\theta) \\|_F^2$ .", + "bbox": [ + 511, + 638, + 908, + 777 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Language embedded 3D Gaussian Splitting. The basic idea of the language embedded Gaussian representation [9, 14, 22, 30, 34, 41-43] is to replace the color rendering to language embedding rendering. Language embedded 3D Gaussians are parameterized as $\\Phi = \\{\\theta_i,\\tilde{\\mathbf{f}}_i\\}_{i = 1}^N = \\{\\pmb {\\mu}_i,S_i,R_i,\\alpha_i,\\mathbf{c}_i,\\tilde{\\mathbf{f}}_i\\}_{i = 1}^N$ where $\\tilde{\\mathbf{f}}_i$ denotes Gaussian-registered language embeddings across $N$ numbers 3D Gaussians which will be discussed soon. Then,", + "bbox": [ + 511, + 779, + 910, + 900 + ], + "page_idx": 1 + }, + { + "type": "page_number", + "text": "14138", + "bbox": [ + 480, + 944, + 519, + 955 + ], + "page_idx": 1 + }, + { + "type": "image", + "img_path": "images/bc16a901451bbd7d47666029bb23cb06177bb3d61c9169e04f64d082bd270c42.jpg", + "image_caption": [ + "(a) Preprocessing stage" + ], + "image_footnote": [], + "bbox": [ + 112, + 89, + 289, + 226 + ], + "page_idx": 2 + }, + { + "type": "image", + "img_path": "images/aca0342daac3b053a078af98d85ea144768656025c72e696d9355079d86dd4ea.jpg", + "image_caption": [ + "Figure 3. Overview of Dr. Splat. (a) In the preprocessing stage, we compute optimized 3D Gaussians [16] and Product Quantization (PQ) codebook construction. (b) During training, we extract CLIP embeddings from given images $\\{\\mathbf{I}\\}$ , and then proceed feature registration process (Sec. 3.1). Finally, we obtain 3D Gaussians $\\Phi^{\\mathrm{ours}}$ with PQ indices $\\{j\\}$ (Sec. 3.2)." + ], + "image_footnote": [], + "bbox": [ + 326, + 92, + 653, + 227 + ], + "page_idx": 2 + }, + { + "type": "image", + "img_path": "images/f54340f180fe2b3c7e15a18582f5fd220e0f09d837291846a941ec1fdcc95234.jpg", + "image_caption": [ + "(b) Training stage" + ], + "image_footnote": [], + "bbox": [ + 656, + 93, + 875, + 227 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "analogous to the color rendering Eq. (1), the language embedding rendering is expressed as:", + "bbox": [ + 89, + 323, + 485, + 354 + ], + "page_idx": 2 + }, + { + "type": "equation", + "text": "\n$$\n\\hat {\\mathbf {f}} = \\sum_ {i = 1} ^ {N} T _ {i} \\tilde {\\alpha} _ {i} \\tilde {\\mathbf {f}} _ {i}, \\tag {2}\n$$\n", + "text_format": "latex", + "bbox": [ + 220, + 361, + 483, + 388 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "where $\\hat{\\mathbf{f}}$ denotes a rendered language embedding. Likewise, the Gaussian-registered language embeddings $\\{\\tilde{\\mathbf{f}}\\}$ are optimized by minimizing the rendering loss between the 2D language embedding $\\mathbf{f}$ extracted from an input image and a rendered language embedding map $\\hat{\\mathbf{f}}$ as $\\arg \\min_{\\{\\tilde{\\mathbf{f}}\\}}\\| \\mathbf{f} - \\hat{\\mathbf{f}} \\|_F^2$ at each corresponding pixel. This can be regarded as distilling vision language models into Gaussian-registered language embedding $\\tilde{\\mathbf{f}}$ through volume rendering Eq. (2). The Gaussian-registered language embeddings are separately trained after pre-training and fixing the pre-trained 3DGS $\\Theta$ for a scene. The language embeddings to be distilled are typically obtained from CLIP [31]. Since storing 32-bit 512-D CLIP features $\\mathbf{f}$ in every 3D Gaussians is memory-expensive, one can use a compressed feature per scene depending on the needs [14, 19, 34, 41-43].", + "bbox": [ + 88, + 397, + 483, + 625 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Motivation. Such language-embedded radiance fields provide useful representation and language interfaces for many practical and crucial applications. While most of existing works focus on the training efficiency, the complexity in inference time has barely been discussed. Considering a scenario to text-query a 3D location of the language-embedded Gaussians, i.e., 3D localization, the aforementioned methods first require rendering a 2D language embedding map at each specific camera pose. We cannot directly retrieve over the distributed embeddings $\\{\\tilde{\\mathbf{f}}_i\\}$ in 3D Gaussians, because the embeddings do not carry language information directly, but their weighted summed (rendered) features $\\hat{\\mathbf{f}}$ do. This issue becomes even severer with compressed features as in [30]: their decompression decoders are not designed for and incompatible with directly applying to the distributed compressed language embeddings in each 3D Gaussian, yielding degenerated CLIP decoding (refer to Fig. 2).", + "bbox": [ + 89, + 628, + 483, + 885 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "This introduces multiple challenges and hassles. First, it", + "bbox": [ + 109, + 885, + 482, + 900 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "is challenging to find the best or proper camera rendering views that contain the object to find. One may attempt to pre-compute the minimal number of cameras and their camera poses that cover all the 3D Gaussians in a scene with proper resolutions, similarly by point-based approach [11]. However, this is a well-known set covering problem [7] with constraints which is known to be an NP-hard problem.", + "bbox": [ + 511, + 323, + 906, + 429 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Second, even with pre-computed rendered views, the retrieval complexity over the rendered images remains substantial [8]. Suppose a scene consisting of one million Gaussians, but just a single rendered language embedding map in pixel domain already has nearly a million pixels; thus, we need a dedicated system to efficiently retrieve over all the views. Third, since the retrieval is conducted in the 2D space, to find a 3D location, we need a separate mechanism to lift the localization to the 3D space, i.e., increasing the system complexity. In addition, 32-bit floating 512-Dimension CLIP features for millions of Gaussian are memory intensive, which is often not manageable. To reduce this burden, the existing methods [37] apply compressions with per-scene optimized codebooks, which hinders extension or generalization to other scenes.", + "bbox": [ + 511, + 431, + 908, + 656 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "To overcome these, we propose a training-free algorithm for the direct allocation of language embeddings to 3D Gaussians, allowing efficient computation and interaction within the 3D space. As a concurrent work, OpenGaussian [37] tackles a similar challenge with our work, but still requires per-scene codebook construction Fig. 1.", + "bbox": [ + 511, + 659, + 908, + 750 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3. Dr. Splat", + "text_level": 1, + "bbox": [ + 511, + 767, + 612, + 785 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "This section provides details of our method. We first explain how we directly register CLIP embeddings into Gaussian-registered language embeddings, Sec. 3.1. Then, we introduce Product Quantization (PQ) into our framework to efficiently store Gaussian-registered language embeddings, Sec. 3.2. Lastly, we describe the inference stage for text query-based 3D Gaussian localization, Sec. 3.3.", + "bbox": [ + 511, + 794, + 908, + 900 + ], + "page_idx": 2 + }, + { + "type": "page_number", + "text": "14139", + "bbox": [ + 480, + 944, + 517, + 955 + ], + "page_idx": 2 + }, + { + "type": "image", + "img_path": "images/6a58cf2fe02beb59208dadb777373d4f5453e42fdd45b6e7b9a238b3f49badaf.jpg", + "image_caption": [ + "(a) Map CLIP features to Gaussians", + "Figure 4. Feature registration process in Dr. Splat. (a) We first map per-pixel CLIP embeddings $\\{\\mathbf{f}^{\\mathrm{map}}\\}$ to Gaussians. Here, we only map dominant $k$ Gaussians along pixel ray $r$ , named Top- $k$ Gaussians. (b) After collecting embeddings, we compute aggregated features (Eq. (6)). (c) Finally, we re-use PQ to obtain the PQ indices $j$ of aggregated features and update Gaussian parameters $\\Phi^{\\mathrm{ours}}$ ." + ], + "image_footnote": [], + "bbox": [ + 140, + 88, + 344, + 234 + ], + "page_idx": 3 + }, + { + "type": "image", + "img_path": "images/4971ba003bc5b4ef9be830de69b10f0accca954ab1a688cecfcc1fc394f5d2b0.jpg", + "image_caption": [ + "(b) Aggregate multiview features", + "(c) Register feature to $\\Phi^{\\text{ours}}$" + ], + "image_footnote": [], + "bbox": [ + 359, + 87, + 880, + 233 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "3.1. Feature registration process", + "text_level": 1, + "bbox": [ + 89, + 323, + 341, + 339 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Our goal is to reconstruct a language embedded 3D space represented by 3D Gaussians $\\Phi$ , which we can directly interact in 3D space without feature rendering Eq. (2). For that, following LangSplat [30], we begin by extracting per-pixel CLIP embedding maps $\\mathbf{F}^{\\mathrm{map}} \\in \\mathbb{R}^{D \\times H \\times W}$ from training images of the target scenes, where $D$ is the dimension of CLIP embeddings, $H$ and $W$ are the height and width of the training images. Given training images, we extracts a dictionary of binary masks and language embeddings extracted from the images as: $\\mathcal{F}^{\\mathrm{map}} = \\{\\mathbf{M}_j : \\mathbf{f}_j^{\\mathrm{map}} \\mid j = 1, \\dots, M\\}$ , where $\\mathbf{M}_j \\in \\mathbb{R}^{H \\times W}$ is a binary mask extracted using SAM [20] and $\\mathbf{f}_j^{\\mathrm{map}} \\in \\mathbb{R}^D$ is a corresponding CLIP embedding from a cropped image with $\\mathbf{M}_j$ . Each mask $\\mathbf{M}_j$ belongs to an image, and the masks are not overlapped to each other. With this dictionary, a CLIP embedding map $\\mathbf{F}^{\\mathrm{map}}(\\mathbf{I}, \\mathbf{r})$ at a pixel $\\mathbf{r}$ in a training image $\\mathbf{I}$ is computed as:", + "bbox": [ + 89, + 345, + 483, + 590 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\mathbf {F} ^ {\\text {m a p}} (\\mathbf {I}, \\mathbf {r}) = \\sum_ {j = 1} ^ {M} \\mathbf {M} _ {j} (\\mathbf {I}, \\mathbf {r}) \\cdot \\mathbf {f} _ {j} ^ {\\text {m a p}}, \\tag {3}\n$$\n", + "text_format": "latex", + "bbox": [ + 161, + 602, + 483, + 631 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "where $\\mathbf{M}_j(\\mathbf{I},\\mathbf{r})\\in \\{0,1\\}$ indicates whether the mask $\\mathbf{M}_j$ contains the pixel $\\mathbf{r}$ in the image $\\mathbf{I}$ . Using $\\mathbf{F}^{\\mathrm{map}}$ , we reconstruct language embedded 3D Gaussians via a novel feature registration process as visualized in Fig. 3.", + "bbox": [ + 89, + 643, + 483, + 703 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "During the feature registration process, our algorithm iterates through training images of the scene. Using projection relation, we link 3D Gaussians $\\Phi$ to CLIP embeddings. Each Gaussian can link to multiple CLIP embeddings derived from different images. Then we aggregate collected embeddings to a single embedding to be assigned to each Gaussian. To ensure a consistent aggregation of the embeddings from multi-view images, we first compute a weight $w_{i}(\\mathbf{I},\\mathbf{r})$ representing the contribution of $\\theta_{i}$ to construct each pixel $\\mathbf{r}$ in a training image $\\mathbf{I}$ . The weights are computed with the volume rendering equation Eq. (1) as:", + "bbox": [ + 89, + 705, + 483, + 871 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\nw _ {i} (\\mathbf {I}, \\mathbf {r}) = T _ {i} (\\mathbf {I}, \\mathbf {r}) \\cdot \\tilde {\\alpha} _ {i} (\\mathbf {I}, \\mathbf {r}), \\tag {4}\n$$\n", + "text_format": "latex", + "bbox": [ + 187, + 885, + 483, + 901 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "where $T_{i}(\\mathbf{I},\\mathbf{r})$ and $\\tilde{\\alpha}_{i}(\\mathbf{I},\\mathbf{r})$ are the transmittance and the effective opacity value of $\\theta_{i}$ for a pixel $\\mathbf{r}$ in an image $\\mathbf{I}$ , stated in Eq. (1). With the per-pixel weights, we calculate $w_{ij}$ representing a weight between each Gaussian $\\theta_{i}$ and corresponding language embedding maps $\\mathbf{f}_j^{\\mathrm{map}}$ , which is for aggregating CLIP embeddings from $\\mathbf{F}^{\\mathrm{map}}$ and register the embedding to each Gaussian. The weights are computed as:", + "bbox": [ + 511, + 323, + 906, + 429 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\nw _ {i j} = \\sum_ {\\mathbf {I} \\in \\mathcal {I}} \\sum_ {\\mathbf {r} \\in \\mathbf {I}} \\mathbf {M} _ {j} (\\mathbf {I}, \\mathbf {r}) \\cdot w _ {i} (\\mathbf {I}, \\mathbf {r}), \\tag {5}\n$$\n", + "text_format": "latex", + "bbox": [ + 571, + 440, + 906, + 465 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "where $\\mathcal{I}$ is the set of the training images. In this iterative process, we aggregate weights only for Top- $k$ Gaussians with the highest weights $w_{i}(\\mathbf{I},\\mathbf{r})$ , along the ray of each pixel ray $\\mathbf{r}$ (see Fig. 4). After aggregation, we prune the Gaussians which are not assigned any weight, i.e., $\\sum_{j=1}^{M} w_{ij} = 0$ . This summation aggregates weights between Gaussians and the CLIP embeddings by linking per-pixel weights $w_{i}(\\mathbf{I},\\mathbf{r})$ of each Gaussian to its corresponding CLIP embeddings. With the obtained weights, we register an aggregated feature $\\dot{\\mathbf{f}}_i$ to each Gaussian with weighted-averaging as:", + "bbox": [ + 511, + 473, + 906, + 625 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\dot {\\mathbf {f}} _ {i} = \\mathbf {f} _ {i} / | | \\mathbf {f} _ {i} | | _ {2}, \\text {w h e r e} \\quad \\mathbf {f} _ {i} = \\sum_ {j = 1} ^ {M} \\frac {w _ {i j}}{\\sum_ {k = 1} ^ {M} w _ {i k}} \\mathbf {f} _ {j} ^ {\\text {m a p}}. \\tag {6}\n$$\n", + "text_format": "latex", + "bbox": [ + 529, + 635, + 906, + 664 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "This process enables 3D-aware feature registration to be consistent across various viewpoints, by aggregating features in the original high-dimensional feature space. The proposed process can be interpreted as an inverse volume rendering without gradient-based optimization, which enables our method to be faster than the prior methods requiring perscene gradient-based optimization [27, 30, 34] for feature registration in 3D space.", + "bbox": [ + 511, + 672, + 908, + 792 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "3.2. Product-Quantized CLIP embeddings", + "text_level": 1, + "bbox": [ + 511, + 801, + 841, + 819 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Memory efficiency is a challenge in 3D scene representations, especially when associating Gaussians with high-dimensional feature vectors. LangSplat [30] addresses this by introducing an encoder-decoder network, while LeGaussian [34] and OpenGaussian [37] utilize codebook construc", + "bbox": [ + 511, + 824, + 906, + 900 + ], + "page_idx": 3 + }, + { + "type": "page_number", + "text": "14140", + "bbox": [ + 480, + 944, + 517, + 955 + ], + "page_idx": 3 + }, + { + "type": "table", + "img_path": "images/2c0571696a15a9a53254393435519e76a5a4e247f8fa9cc87f1782e132578b50.jpg", + "table_caption": [], + "table_footnote": [], + "table_body": "
MethodsmIoUmAcc @ 0.25
waldo_kitchenramenfigurinesteatimeMeanwaldo_kitchenramenfigurinesteatimeMean
LangSplat-m [30]8.296.118.3316.589.8313.6414.088.9327.1215.94
OpenGaussian [37]34.6023.8759.3354.4443.0650.0035.2180.3672.8859.61
Ours (Top-10)37.0524.3354.4257.3543.2963.6435.2180.3677.9764.30
Ours (Top-20)38.3324.5853.9456.1943.2663.6435.2182.1476.2764.32
Ours (Top-40)39.0724.7053.3657.2043.5863.6435.2180.3676.2763.87
", + "bbox": [ + 114, + 88, + 883, + 193 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Table 1. 3D object selection results on the LeRF-OVS dataset [17]. To measure 3D object selection performance, we calculate 2D segmentation accuracy on rendering of selected 3D Gaussians. Note that our model does not require per-scene optimization, demonstrating its robustness across diverse scenes. Bold and Underline stand for first and second best performance.", + "bbox": [ + 89, + 203, + 906, + 246 + ], + "page_idx": 4 + }, + { + "type": "image", + "img_path": "images/42dd55066e5ac092d686ad46cf83089ae2f133b774de38c9a7f70912a0fdcd22.jpg", + "image_caption": [ + "Figure 5. Qualitative results of the object selection on the LeRF-OVS dataset [17]. We visualize rendering of selected 3D Gaussians for LangSplat [30], OpenGaussian [37], and ours. For LangSplat, activations are often distributed randomly, fail to localize the target. OpenGaussian often struggles to distinguish closely situated objects. In contrast, our model shows activations precisely limited to the queried object regions, effectively localizing only the relevant areas." + ], + "image_footnote": [], + "bbox": [ + 114, + 271, + 251, + 484 + ], + "page_idx": 4 + }, + { + "type": "image", + "img_path": "images/5cfd6f2ed8ab85d8fd3a5e46d87974d73f21e5beebe0ef40a115c4d6b44692bc.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 251, + 271, + 372, + 484 + ], + "page_idx": 4 + }, + { + "type": "image", + "img_path": "images/23d2d3d1f429017f8d554d145f66b3475175ec479fc75eebbbfac5f2953c0d18.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 375, + 271, + 495, + 484 + ], + "page_idx": 4 + }, + { + "type": "image", + "img_path": "images/0245c63cd0c27778f78a287b354a76abb1e7514fe05af7fa14b11d54ebd7d9f3.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 501, + 271, + 637, + 484 + ], + "page_idx": 4 + }, + { + "type": "image", + "img_path": "images/7abf6229bb01b5699460772cb48c1322fec2e565a156a66d0108d37f81152783.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 638, + 271, + 758, + 484 + ], + "page_idx": 4 + }, + { + "type": "image", + "img_path": "images/1dd7bc29c4e690359fc8c42a213d96ef133a8b2c20cd2abd3d025e8c16a17924.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 761, + 271, + 882, + 484 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "tion. However, these approaches introduce additional per-scene computational costs for scene-specific parameter tuning of neural networks or codebooks (see Fig. 1). In contrast, we propose to use Product Quantization (PQ) on a large-scale image dataset, eliminating per-scene training.", + "bbox": [ + 89, + 571, + 485, + 647 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Product Quantization. PQ [13] is a widely used technique for efficient embedding compression, particularly valuable in large-scale applications. The PQ process begins by dividing the original $D$ -dimensional feature vector $\\mathbf{v}$ into $L$ sub-vectors: $\\mathbf{v} = [\\mathbf{v}_1,\\mathbf{v}_2,\\dots ,\\mathbf{v}_L]$ . Each sub-vector $\\mathbf{v}_i$ is then independently quantized to a predefined number of centroids $s_{ij}$ in a predefined codebook $S_{i}$ for that sub-vector. These centroids are learned via clustering, creating a codebook for each subspace. Once the centroids are established, each sub-vector is replaced by the index of the nearest centroid in its respective codebook. The centroid indices $j_{i} = [j_{i1},j_{i2},\\ldots ,j_{iL}]$ are optimized by minimizing $\\arg \\min_k\\| \\mathbf{v}_i - s_{ik}\\|$ to quantize a given vector $\\mathbf{v}_i$ where $j_{ik}$ is an 8-bit unsigned integer.", + "bbox": [ + 89, + 654, + 482, + 864 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Then, we can measure the distance between the query and data by adding distances between coarse centroids. Once the", + "bbox": [ + 89, + 869, + 483, + 900 + ], + "page_idx": 4 + }, + { + "type": "image", + "img_path": "images/0e0078a6532ba08089c97e86c554134716d13da880e39ddeae2e1e7291cd23ce.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 526, + 570, + 645, + 765 + ], + "page_idx": 4 + }, + { + "type": "image", + "img_path": "images/18a40acb4c8461f487d030403c978efc5baa11ca1b4f93f66726976cdcd8f91a.jpg", + "image_caption": [ + "Figure 6. Limitations of point-based IoU measurement. This figure shows the effect of removing the top and bottom $30\\%$ of Gaussians according to the proposed significant score, implying that volume differences significantly impact 3D accuracy. The results highlight the need for the proposed IoU metric for 3D Gaussians." + ], + "image_footnote": [], + "bbox": [ + 645, + 570, + 764, + 771 + ], + "page_idx": 4 + }, + { + "type": "image", + "img_path": "images/01353414c5ff054c5b8723a00c26086fb62fc46477629e77e814574b061054fe.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 767, + 571, + 883, + 771 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "distances between centroids are computed as a lookup table, the computation shifts to simple indexing, which reduces the search complexity from $\\mathcal{O}(D)$ to $\\mathcal{O}(1)$ for a $D$ dimen", + "bbox": [ + 511, + 854, + 908, + 902 + ], + "page_idx": 4 + }, + { + "type": "page_number", + "text": "14141", + "bbox": [ + 480, + 944, + 517, + 955 + ], + "page_idx": 4 + }, + { + "type": "table", + "img_path": "images/6e943fd223922385345560d8d15ca6e8f9ab4f2851957f6a7005e8e90a26eb4d.jpg", + "table_caption": [], + "table_footnote": [], + "table_body": "
3D\nmIoUIoU > 0.1519 classes\nIoU > 0.3IoU > 0.45
LangSplat-m [30]8.017.17.82.9
LEGaussians-m [34]9.519.18.97.3
OpenGaussian [37]25.259.538.018.3
Ours (Top-20)25.060.740.320.0
Ours (Top-40)25.460.740.325.6
", + "bbox": [ + 101, + 88, + 480, + 183 + ], + "page_idx": 5 + }, + { + "type": "table", + "img_path": "images/f75d749f04c7944c007e27d41792d902ea94081d3613d217b6a82c927f29a9c4.jpg", + "table_caption": [ + "(a) 3D object localization task." + ], + "table_footnote": [], + "table_body": "
19 classes15 classes10 classes
mIoUmAcc.mIoUmAcc.mIoUmAcc.
LangSplat-m [30]2.09.24.914.68.023.9
LEGaussians-m [34]1.67.94.616.17.724.9
OpenGaussian [37]30.146.538.156.849.771.4
Ours (Top-20)28.044.638.260.447.268.9
Ours (Top-40)29.647.738.260.450.273.5
", + "bbox": [ + 493, + 88, + 893, + 183 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "(b) Open-vocabulary 3D semantic segmentation task.", + "bbox": [ + 550, + 184, + 831, + 196 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Table 2. Quantitative comparison in the ScanNet dataset [4]. Left: Localization prediction is defined as 3D regions with a text similarity score above threshold. Right: We assign segmentation labels by finding max activations among all classes. Note that Bold and Underline stand for first and second best performance, respectively.", + "bbox": [ + 89, + 200, + 906, + 244 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/e7360054e021d98a71a00ff79e35a322409c81ec51cf9f1ae80953595a74b077.jpg", + "image_caption": [ + "Figure 7. Qualitative results of 3D object localization. We visualize 3D localization activations (yellow) for \"chair\" and \"desk\" in the ScanNet dataset, comparing our method with others. It turns out that LangSplat-m and LEGaussians-m fail to localize objects accurately, while OpenGaussian struggles with object correspondence. Our model delivers precise and consistent localization across diverse queries." + ], + "image_footnote": [], + "bbox": [ + 109, + 253, + 883, + 392 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "sion sample. This approach notably reduces computational complexity, making it suitable for large-scale search.", + "bbox": [ + 89, + 463, + 482, + 493 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "In our setup for language-based 3D scene understanding, we build PQ centroids based on CLIP embeddings using a large-scale image dataset, the LVIS dataset [10], that contains over 1.2M instances covering various long-tail classes and ground truth segmentation. We extract instance patches from images and collect patch-wise CLIP embeddings. After we build this CLIP embedding database, we proceed with the construction of the centroid codebook for our PQ. Once PQ is trained, any query embedding can be approximated by assigning the closest centroid for each subvector. This is a one-time procedure; once we determine the codebook, we can use it for any scene generally. In our setup, each embedding is represented as a sequence of centroid indices rather than a high-dimensional vector. Accordingly, our language embedded Gaussians are parametrized as $\\Phi^{\\mathrm{ours}} = \\{\\phi_i^{\\mathrm{ours}}\\}_{i=1}^N = \\{\\theta_i, j_i\\}_{i=1}^N$ . where the aggregated feature $\\hat{\\mathbf{f}}_i$ are converted as a quantized feature $\\bar{\\mathbf{f}}_i$ by the corresponding PQ index $j_i$ .", + "bbox": [ + 88, + 496, + 485, + 770 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "3.3. Text-query based 3D localization", + "text_level": 1, + "bbox": [ + 89, + 785, + 379, + 801 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "After training 3D Gaussians $\\Phi^{\\mathrm{ours}}$ with our feature registration process and PQ, we describe the details of an inference mode that facilitates direct interaction with 3DGS upon receiving input queries, such as text. This is related to similarity score computation between a query and sources, i.e. Gaussian embeddings. Given a text, we first extract a", + "bbox": [ + 89, + 809, + 485, + 901 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "query feature $\\mathbf{q}$ using CLIP text encoder [31]. We reconstruct the quantized features $\\{\\bar{\\mathbf{f}}_i\\}_{i=1}^N$ from the stored PQ indices $\\{j_i\\}_{i=1}^N$ . Then, we compute a cosine similarity score between the query feature $\\mathbf{q}$ and all quantized features.", + "bbox": [ + 511, + 463, + 906, + 523 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Despite its simplicity, solely relying on the cosine similarity may result in diminished discriminability across certain similarity scores. To address this limitation, we incorporate a re-ranking process based on relative activation with respect to the canonical feature. For this process, we adopt the relevancy scoring method proposed in LeRF [17], which enables more precise similarity analysis for a query. Specifically, each rendered language embedding, $\\mathbf{f}^{\\mathrm{map}}$ and a text query feature $\\mathbf{q}$ , yield a relevance score determined by $\\min_{i}\\frac{\\exp(\\mathbf{f}^{\\mathrm{map}}\\cdot\\mathbf{q})}{\\exp(\\mathbf{f}^{\\mathrm{map}}\\cdot\\mathbf{q}) + \\exp(\\mathbf{f}^{\\mathrm{map}}\\cdot\\mathbf{f}^{\\mathrm{canon,i}})}$ , where $(\\cdot)$ is an element-wise dot product operator and $\\mathbf{f}^{\\mathrm{canon,i}}$ indicates CLIP embeddings of a designated canonical term selected from a set of \"object,\" \"things,\" \"stuff,\" and \"texture\". Then, we sample 3D Gaussians based on the relevance score for downstream tasks.", + "bbox": [ + 511, + 523, + 908, + 739 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "4. Experiments", + "text_level": 1, + "bbox": [ + 511, + 753, + 645, + 771 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Dataset. We use two datasets to evaluate the 3D scene understanding performance. For the 3D object selection task (Sec. 4.1), we use the LERF [17] dataset annotated by LangSplat [30], which consists of several multi-view images of 3D scenes containing long-tail objects and includes ground truth 2D ground truth annotations for texture queries. For 3D object localization Sec. 4.2 and 3D semantic segmentation Sec. 4.3 task, we employ the ScanNet [4] dataset.", + "bbox": [ + 511, + 779, + 908, + 902 + ], + "page_idx": 5 + }, + { + "type": "page_number", + "text": "14142", + "bbox": [ + 480, + 944, + 519, + 955 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/74bf00ac799dd062d56c084505c13cea6436e7b9a368cf80cb7eab240c38994b.jpg", + "image_caption": [ + "Figure 8. Visualization of open-vocabulary 3D semantic segmentation on the ScanNet dataset [4]. We visualize 3D Gaussian splat-based semantic segmentation using language features allocation of OpenGaussian [37] and Dr. Splat (ours) model on the same RGB-pretrained 3DGS. Note that, not specifically designed for segmentation, it achieves high performance as a result of language-based Gaussian updates." + ], + "image_footnote": [], + "bbox": [ + 107, + 88, + 883, + 228 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "ScanNet is a large-scale benchmark that provides data on indoor scenes, including calibrated RGBD images and 3D point clouds with ground-truth semantic labels. We randomly select eight scenes from ScanNet for the experiments.", + "bbox": [ + 88, + 299, + 485, + 359 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Competing methods. The only method available for a fair comparison with our method is the concurrent work, Open-Gaussian [37]. To study the various aspects of our method, we introduce baseline methods modified from rasterization-based ones [30, 34], for direct 3D referring operation, denoted as LangSplat-m and LEGaussians-m. As discussed in Sec. 2, without modification, global search over a whole scene is quite demanding. To ensure fair evaluation, we use the same initial 3D Gaussians being trained only using RGB inputs for all comparing methods, and freeze the Gaussians during the language feature allocation process. Also, the per-pixel CLIP [31] embedding maps are unified for SAM-based [20] methods [30, 37] including ours. We follow the hyperparameter settings favorable to each respective paper.", + "bbox": [ + 88, + 361, + 483, + 574 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "4.1.3D object selection", + "text_level": 1, + "bbox": [ + 89, + 587, + 272, + 603 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Settings. We first extract text features from an open-vocabulary text query using the CLIP model. Next, we compare text features to the 3D features embedded in each Gaussian using cosine similarity. By thresholding the similarity, we identify the 3D Gaussians that are relevant to the given text query. The selected 3D points are subsequently rendered into multi-view images using the 3DGS rasterization pipeline.", + "bbox": [ + 88, + 609, + 485, + 731 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Results. We compare our model quantitatively with 3DGS-based language-embedded models as shown in Table 1. The results demonstrate that our method performs better object selection in most scenes, showing an improvement of over 0.5 in mIoU and more than 4.5 in mAcc compared to counterpart models. Notably, the rasterization-based method, LangSplat-m, often underperforms in most scenes.", + "bbox": [ + 88, + 732, + 485, + 838 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Qualitative results are shown in Fig. 5. For LangSplat-m, the activations often shows random 3D Gaussians or fail to localize entirely (e.g., see \"coffee mug\"), highlighting the limitations of rasterization-based methods and their un", + "bbox": [ + 88, + 839, + 485, + 898 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "suitability for 3D understanding, aligning the observation from Fig. 2. OpenGaussian frequently exhibits false activations with incorrect text-object pairs (e.g., \"apple\" and \"tea in a glass\") and struggles to distinguish between nearby objects (e.g., \"waldo,\" \"rubik's cube\"). This artifacts can be attributed to use of spatial clustering and limited encoder capacity.", + "bbox": [ + 511, + 299, + 906, + 405 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "In contrast, our model leverages general image features thanks to the general PQ, maintaining feature distinctiveness regardless of scene complexity. Our feature registration considers the 3D geometry of the 3D Gaussians, which results in superior performance in 3D scene understanding tasks.", + "bbox": [ + 511, + 406, + 906, + 482 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "4.2. 3D object localization", + "text_level": 1, + "bbox": [ + 511, + 496, + 718, + 512 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Settings. Similar to the 3D object selection task, we calculate the cosine similarity between text query and 3D features embedded in each Gaussian. By thresholding the similarity, we identify the 3D Gaussians relevant to the given text query. To measure volume-aware localization evaluation, we propose a protocol to measure the IoU of 3D Gaussians that expands the traditional metric of point cloud-based approaches by incorporating volumetric information of 3D Gaussians.", + "bbox": [ + 511, + 518, + 908, + 654 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Novel evaluation protocol for 3D localization in 3DGS. Unlike conventional evaluation protocol for the 3D localization task in point clouds, it is tricky to evaluate 3D localization performance in 3D Gaussians [16]. This is primarily due to the un-deterministic structure of Gaussian distribution. To address this issue, we compute 3DGS pseudo-labels for evaluating the 3DGS localization in a volume-aware way. The details can be found in the supplementary material.", + "bbox": [ + 511, + 657, + 908, + 777 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Given the ground truth, we measure IoU considering the spatial significance of each Gaussian and define a significant score $d_{i}$ for each Gaussian $\\theta_{i}$ with its scale $\\mathbf{s}_i = [s_{ix}, s_{iy}, s_{iy}]$ and opacity $\\alpha_{i}$ as $d_{i} = s_{ix}s_{iy}s_{iz}\\alpha_{i}$ , where $s_{ix}s_{iy}s_{iz}$ denotes a relative ellipsoid volume of a Gaussian $\\theta_{i}$ . With the obtained significant scores $\\mathbf{d} = [d_1, d_2, \\dots, d_N]$ , we compute weighted IoU of 3D Gaussians to approximate volumes. The proposed metric is designed to assign a larger", + "bbox": [ + 511, + 780, + 910, + 901 + ], + "page_idx": 6 + }, + { + "type": "page_number", + "text": "14143", + "bbox": [ + 480, + 944, + 519, + 955 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "weight to the Gaussians with higher significant scores, when measuring IoU. Figure 6 shows that the impact of each Gaussian on the scene extremely varies depending on their significant scores, which demonstrates the necessity of the proposed IoU metric on 3D Gaussians that regards unequal contributions of each Gaussian.", + "bbox": [ + 89, + 90, + 483, + 180 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Results. We report the 3D localization performance on the Scannet dataset in Table 2a. The 2D rasterization-based methods [30, 34] struggle to achieve precise activations for 3D localization. They inherently face challenges when applying for 3D tasks because they need to render 2D images for the scene interaction. Even with the 3D space search method, OpenGaussian [37], our model consistently demonstrates superior performance and achieves higher accuracy in localization. Figure 7 also shows that LangSplat-m and LEGaussians-m fail to properly localize the objects, and OpenGaussian misses queried objects in the scene.", + "bbox": [ + 89, + 181, + 483, + 348 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "4.3. 3D semantic segmentation", + "text_level": 1, + "bbox": [ + 89, + 359, + 328, + 376 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Settings. For a given set of open-vocabulary text labels, we perform segmentation by assigning each Gaussian a label having the highest activation among the known label set.", + "bbox": [ + 89, + 382, + 483, + 428 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Results. The numerical comparison is presented in Table 2b. Although not explicitly designed for semantic segmentation, our model achieves notable performance in this task as a result of accurately updating each Gaussian with language features. Consistent with previous observations, rasterization-based 3DGS models exhibit lower segmentation performance. While OpenGaussian performs position-based clustering, our model demonstrates comparable performance, surpassing the baseline as the Top- $k$ value increases. Our model also achieves better segmentation results, with a visual comparison of the segmented scene shown in Fig. 8.", + "bbox": [ + 89, + 429, + 483, + 594 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "4.4. Ablation study", + "text_level": 1, + "bbox": [ + 89, + 606, + 240, + 622 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "We conduct an ablation study using the ScanNet dataset on different hyper-parameters of Dr. Splat to measure the contribution of each component.", + "bbox": [ + 89, + 628, + 482, + 672 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Product Quantization. PQ introduces a trade-off between memory usage, computational efficiency, and accuracy. To better understand the balance between computational cost and localization quality, we conduct an ablation study by varying the number of sub-vectors. We evaluate performance at sub-vector sizes of 64, 128, and 256. Notably, these settings correspond to bit-size reductions of 1/32, 1/16, and 1/8 of the original CLIP feature, respectively. We measure the query distance computation time for one million data points, averaging results over 100 iterations for efficiency measure. Our findings reveal a favorable trade-off between quantization performance and accuracy (see Fig. 9-(b)) in the Pareto front with our PQ configurations. This achieves a balance that maximizes memory and computational efficiency while minimizing any loss in accuracy.", + "bbox": [ + 89, + 674, + 483, + 900 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/976c9798fa322047f804bddbd485f7e83bc57c08b9c0475f944ca46d55412b45.jpg", + "image_caption": [ + "(a) Ablation on PQ parameters." + ], + "image_footnote": [], + "bbox": [ + 522, + 89, + 705, + 178 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/00be046b6c22477fe8a63532227a14e3f6fde73bd86141153e372d00f2e190a9.jpg", + "image_caption": [ + "(b) Ablation on Top- $k$", + "Figure 9. Ablation study on (a) PQ and (b) Top- $k$ Gaussians." + ], + "image_footnote": [], + "bbox": [ + 714, + 89, + 898, + 176 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Top- $k$ Gaussians. We examine the influence of the number of Gaussians assigned per ray. This parameter affects both memory requirements and computation, serving as a critical factor in overall performance. The ratio of pruned Gaussians and the mIoU results from different $k$ are presented in Fig. 9a. We observe that increasing the aggregating number of Gaussians per ray improves localization performance; however, it results in higher memory consumption and the number of occupied Gaussians, indicating a clear trade-off.", + "bbox": [ + 511, + 241, + 906, + 377 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "5. Discussion and Conclusion", + "text_level": 1, + "bbox": [ + 511, + 393, + 761, + 409 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "We present Dr. Splat, which is a novel approach for open-vocabulary 3D scene understanding by directly registering language embeddings to 3D Gaussians, eliminating the need for an intermediate rendering process. Compared to the previous 2D rendering-based methods [30, 34], which have limited search domain and capacity, our method directly searches 3D space while preserving the fidelity of language embeddings. This operation is further accelerated by the integration of Product Quantization (PQ)", + "bbox": [ + 511, + 419, + 906, + 555 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Experimental results validate Dr. Splat's superior performance across various 3D scene understanding tasks, including open-vocabulary 3D object selection, 3D object localization, and 3D semantic segmentation. These findings highlight Dr. Splat's ability to transform 3D scene understanding by achieving a balance between highly representative quality and computational efficiency. This breakthrough paves the way for advanced applications in robotics, autonomous navigation, and augmented reality.", + "bbox": [ + 511, + 556, + 908, + 691 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Acknowledgment", + "text_level": 1, + "bbox": [ + 513, + 709, + 663, + 724 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "This work was supported by Institute of Information & communications Technology Planning & Evaluation (IITP) (No.RS-2021-II212068, Artificial Intelligence Innovation Hub, No.RS-2024-00457882, National AI Research Lab Project, No.RS-2023-00225630, Development of Artificial Intelligence for Text-based 3D Movie Generation, No.RS-2019-II191906, Artificial Intelligence Graduate School Program(POSTECH)) and by the National Research Foundation of Korea (NRF) (No.RS-2024-00358135, Corner Vision: Learning to Look Around the Corner through Multi-modal Signals) grant funded by the Korea government (MSIT).", + "bbox": [ + 511, + 734, + 908, + 900 + ], + "page_idx": 7 + }, + { + "type": "page_number", + "text": "14144", + "bbox": [ + 480, + 944, + 517, + 955 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "References", + "text_level": 1, + "bbox": [ + 91, + 89, + 187, + 104 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[1] Mathilde Caron, Hugo Touvron, Ishan Misra, Hervé Jégou, Julien Mairal, Piotr Bojanowski, and Armand Joulin. Emerging properties in self-supervised vision transformers. In Proceedings of the IEEE/CVF international conference on computer vision, pages 9650-9660, 2021. 2", + "[2] Jaesung Choe, Chunghyun Park, Francois Rameau, Jaesik Park, and In So Kweon. Pointmixer: Mlp-mixer for point cloud understanding. In European Conference on Computer Vision, pages 620-640. Springer, 2022. 1", + "[3] Christopher Choy, JunYoung Gwak, and Silvio Savarese. 4d spatio-temporal convnets: Minkowski convolutional neural networks. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 3075-3084, 2019. 1", + "[4] Angela Dai, Angel X Chang, Manolis Savva, Maciej Halber, Thomas Funkhouser, and Matthias Nießner. Scannet: Richly-annotated 3d reconstructions of indoor scenes. In CVPR, pages 5828-5839, 2017. 6, 7", + "[5] Runyu Ding, Jihan Yang, Chuhui Xue, Wenqing Zhang, Song Bai, and Xiaojuan Qi. Pla: Language-driven open-vocabulary 3d scene understanding. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 7010-7019, 2023. 2", + "[6] Francis Engelmann, Fabian Manhardt, Michael Niemeyer, Keisuke Tateno, Marc Pollefeys, and Federico Tombari. OpenNerf: Open Set 3D Neural Scene Segmentation with PixelWise Features and Rendered Novel Views. In ICLR, 2024. 2", + "[7] Michael R. Garey and David S. Johnson. Computers and intractability. a guide to the theory of np-completeness. W. H. Freeman and company, 174, 1979. 3", + "[8] Antoine Guédon, Tom Monnier, Pascal Monasse, and Vincent Lepetit. Macarons: Mapping and coverage anticipation with rgb online self-supervision. In CVPR, 2023. 3", + "[9] Jun Guo, Xiaojian Ma, Yue Fan, Huaping Liu, and Qing Li. Semantic gaussians: Open-vocabulary scene understanding with 3d gaussian splatting. arXiv preprint arXiv:2403.15624, 2024. 2", + "[10] Agrim Gupta, Piotr Dollar, and Ross Girshick. Lvis: A dataset for large vocabulary instance segmentation. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 5356-5364, 2019. 6", + "[11] Zhening Huang, Xiaoyang Wu, Xi Chen, Hengshuang Zhao, Lei Zhu, and Joan Lasenby. Openins3d: Snap and lookup for 3d open-vocabulary instance segmentation. In European Conference on Computer Vision, pages 169-185. Springer, 2025. 1, 3", + "[12] Krishna Murthy Jatavallabhula, Alihusein Kuwajerwala, Qiao Gu, Mohd Omama, Tao Chen, Alaa Maalouf, Shuang Li, Ganesh Iyer, Soroush Saryazdi, Nikhil Keetha, et al. Conceptfusion: Open-set multimodal 3d mapping. arXiv preprint arXiv:2302.07241, 2023. 2", + "[13] Herve Jegou, Matthijs Douze, and Cordelia Schmid. Product quantization for nearest neighbor search. IEEE transactions on pattern analysis and machine intelligence, 33(1):117-128, 2010. 5" + ], + "bbox": [ + 93, + 114, + 485, + 898 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[14] Yuzhou Ji, He Zhu, Junshu Tang, Wuyi Liu, Zhizhong Zhang, Yuan Xie, and Xin Tan. Fastlgs: Speeding up language embedded gaussians with feature grid mapping. arXiv preprint arXiv:2406.01916, 2024. 2, 3", + "[15] Li Jiang, Shaoshuai Shi, and Bernt Schiele. Open-vocabulary 3d semantic segmentation with foundation models. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 21284-21294, 2024. 1, 2", + "[16] Bernhard Kerbl, Georgios Kopanas, Thomas Leimkuhler, and George Drettakis. 3d gaussian splatting for real-time radiance field rendering. ACM TOG, 2023. 1, 2, 3, 7", + "[17] Justin Kerr, Chung Min Kim, Ken Goldberg, Angjoo Kanazawa, and Matthew Tancik. Lerf: Language embedded radiance fields. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 19729-19739, 2023. 1, 2, 5, 6", + "[18] GeonU Kim, Kim Youwang, and Tae-Hyun Oh. FPRF: Feedforward photorealistic style transfer of large-scale 3D neural radiance fields. In AAAI, 2024. 2", + "[19] GeonU Kim, Kim Youwang, Lee Hyoseok, and Tae-Hyun Oh. Fpgs: Feed-forward semantic-aware photorealistic style transfer of large-scale gaussian splatting, 2025. 3", + "[20] Alexander Kirillov, Eric Mintun, Nikhila Ravi, Hanzi Mao, Chloe Rolland, Laura Gustafson, Tete Xiao, Spencer Whitehead, Alexander C. Berg, Wan-Yen Lo, Piotr Dólar, and Ross B. Girshick. Segment anything. In ICCV, 2023. 4, 7", + "[21] Sosuke Kobayashi, Eiichi Matsumoto, and Vincent Sitzmann. Decomposing nerf for editing via feature field distillation. Advances in Neural Information Processing Systems, 35:23311-23330, 2022. 2", + "[22] Hyunjee Lee, Youngsik Yun, Jeongmin Bae, Seoha Kim, and Youngjung Uh. Rethinking open-vocabulary segmentation of radiance fields in 3d space, 2024. 2", + "[23] Boyi Li, Kilian Q Weinberger, Serge Belongie, Vladlen Koltun, and René Ranftl. Language-driven semantic segmentation. arXiv preprint arXiv:2201.03546, 2022. 2", + "[24] Kunhao Liu, Fangneng Zhan, Jiahui Zhang, Muyu Xu, Yingchen Yu, Abdulmotaleb El Saddik, Christian Theobalt, Eric Xing, and Shijian Lu. Weakly supervised 3d open-vocabulary segmentation. Advances in Neural Information Processing Systems, 36:53433-53456, 2023. 2", + "[25] Minghua Liu, Yinhao Zhu, Hong Cai, Shizhong Han, Zhan Ling, Fatih Porikli, and Hao Su. Partslip: Low-shot part segmentation for 3d point clouds via pretrained image-language models. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 21736-21746, 2023. 2", + "[26] Ben Mildenhall, Pratul P. Srinivasan, Matthew Tancik, Jonathan T. Barron, Ravi Ramamoorthi, and Ren Ng. Nerf: Representing scenes as neural radiance fields for view synthesis. In ECCV, 2020. 2", + "[27] Songyou Peng, Kyle Genova, Chiyu Jiang, Andrea Tagliasacchi, Marc Pollefeys, Thomas Funkhouser, et al. Openscene: 3d scene understanding with open vocabularies. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 815-824, 2023. 1, 2, 4", + "[28] Charles R Qi, Hao Su, Kaichun Mo, and Leonidas J Guibas. Pointnet: Deep learning on point sets for 3d classification" + ], + "bbox": [ + 516, + 92, + 906, + 900 + ], + "page_idx": 8 + }, + { + "type": "page_number", + "text": "14145", + "bbox": [ + 480, + 944, + 517, + 955 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "and segmentation. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 652-660, 2017. 1", + "[29] Guocheng Qian, Yuchen Li, Houwen Peng, Jinjie Mai, Hasan Hammoud, Mohamed Elhoseiny, and Bernard Ghanem. Pointnext: Revisiting pointnet++ with improved training and scaling strategies. Advances in neural information processing systems, 35:23192-23204, 2022. 1", + "[30] Minghan Qin, Wanhua Li, Jiawei Zhou, Haoqian Wang, and Hanspeter Pfister. Langsplat: 3d language gaussian splatting. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 20051-20060, 2024. 1, 2, 3, 4, 5, 6, 7, 8", + "[31] Alec Radford, Jong Wook Kim, Chris Hallacy, Aditya Ramesh, Gabriel Goh, Sandhini Agarwal, Girish Sastry, Amanda Askell, Pamela Mishkin, Jack Clark, et al. Learning transferable visual models from natural language supervision. In International conference on machine learning, pages 8748-8763. PMLR, 2021. 2, 3, 6, 7", + "[32] Adam Rashid, Satvik Sharma, Chung Min Kim, Justin Kerr, Lawrence Yunliang Chen, Angjoo Kanazawa, and Ken Goldberg. Language embedded radiance fields for zero-shot task-oriented grasping. In 7th Annual Conference on Robot Learning, 2023. 2", + "[33] Damien Robert, Hugo Raguet, and Loic Landrieu. Efficient 3d semantic segmentation with superpoint transformer. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 17195-17204, 2023. 1", + "[34] Jin-Chuan Shi, Miao Wang, Hao-Bin Duan, and Shao-Hua Guan. Language embedded 3d gaussians for open-vocabulary scene understanding. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 5333-5343, 2024. 1, 2, 3, 4, 6, 7, 8", + "[35] Ayca Takmaz, Elisabetta Fedele, Robert W Sumner, Marc Pollefeys, Federico Tombari, and Francis Engelmann. Openmask3d: Open-vocabulary 3d instance segmentation. arXiv preprint arXiv:2306.13631, 2023. 1", + "[36] Xiaoyang Wu, Li Jiang, Peng-Shuai Wang, Zhijian Liu, Xihui Liu, Yu Qiao, Wanli Ouyang, Tong He, and Hengshuang Zhao. Point transformer v3: Simpler faster stronger. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 4840-4851, 2024. 1", + "[37] Yanmin Wu, Jiarui Meng, Hajjie Li, Chenming Wu, Yahao Shi, Xinhua Cheng, Chen Zhao, Haocheng Feng, Errui Ding, Jingdong Wang, et al. Opengaussian: Towards point-level 3d gaussian-based open vocabulary understanding. arXiv preprint arXiv:2406.02058, 2024. 1, 2, 3, 4, 5, 6, 7, 8", + "[38] Jihan Yang, Runyu Ding, Weipeng Deng, Zhe Wang, and Xiaojuan Qi. Regionplc: Regional point-language contrastive learning for open-world 3d scene understanding. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 19823-19832, 2024. 1, 2", + "[39] Junbo Zhang, Runpei Dong, and Kaisheng Ma. Clip-for3d: Learning free open-world 3d scene representations from 2d dense clip. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 2048-2059, 2023. 2", + "[40] Hengshuang Zhao, Li Jiang, Jiaya Jia, Philip HS Torr, and Vladlen Koltun. Point transformer. In Proceedings of" + ], + "bbox": [ + 91, + 90, + 483, + 900 + ], + "page_idx": 9 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "the IEEE/CVF international conference on computer vision, pages 16259-16268, 2021. 1", + "[41] Yuhang Zheng, Xiangyu Chen, Yupeng Zheng, Songen Gu, Runyi Yang, Bu Jin, Pengfei Li, Chengliang Zhong, Zeng-mao Wang, Lina Liu, Chao Yang, Dawei Wang, Zhen Chen, Xiaoxiao Long, and Meiqing Wang. Gaussiangrasper: 3d language gaussian splatting for open-vocabulary robotic grasping. IEEE Robotics and Automation Letters, 2024. 1, 2, 3", + "[42] Shijie Zhou, Haoran Chang, Sicheng Jiang, Zhiwen Fan, Zehao Zhu, Dejia Xu, Pradyumna Chari, Suya You, Zhangyang Wang, and Achuta Kadambi. Feature 3dgs: Supercharging 3d gaussian splattering to enable distilled feature fields. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 21676-21685, 2024.", + "[43] Xingxing Zuo, Pouya Samangouei, Yunwen Zhou, Yan Di, and Mingyang Li. Fmgs: Foundation model embedded 3d gaussian splatting for holistic 3d scene understanding. IJCV, 2024. 1, 2, 3", + "[44] Matthias Zwicker, Hanspeter Pfister, Jeroen van Baar, and Markus H. Gross. EWA volume splatting. In Visualization Conference, 2001. 2" + ], + "bbox": [ + 516, + 92, + 906, + 401 + ], + "page_idx": 9 + }, + { + "type": "page_number", + "text": "14146", + "bbox": [ + 480, + 945, + 519, + 955 + ], + "page_idx": 9 + } +] \ No newline at end of file diff --git a/2025/Dr. Splat_ Directly Referring 3D Gaussian Splatting via Direct Language Embedding Registration/04973834-dfc5-4d84-8238-2be21678c244_model.json b/2025/Dr. Splat_ Directly Referring 3D Gaussian Splatting via Direct Language Embedding Registration/04973834-dfc5-4d84-8238-2be21678c244_model.json new file mode 100644 index 0000000000000000000000000000000000000000..9840145d52f486622e1658bd0bac136135b3e2ce --- /dev/null +++ b/2025/Dr. Splat_ Directly Referring 3D Gaussian Splatting via Direct Language Embedding Registration/04973834-dfc5-4d84-8238-2be21678c244_model.json @@ -0,0 +1,2398 @@ +[ + [ + { + "type": "header", + "bbox": [ + 0.107, + 0.003, + 0.182, + 0.044 + ], + "angle": 0, + "content": "CVF" + }, + { + "type": "header", + "bbox": [ + 0.237, + 0.001, + 0.812, + 0.047 + ], + "angle": 0, + "content": "This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore." + }, + { + "type": "title", + "bbox": [ + 0.237, + 0.131, + 0.761, + 0.177 + ], + "angle": 0, + "content": "Dr. Splat: Directly Referring 3D Gaussian Splatting via Direct Language Embedding Registration" + }, + { + "type": "text", + "bbox": [ + 0.174, + 0.203, + 0.306, + 0.222 + ], + "angle": 0, + "content": "Kim Jun-Seong1" + }, + { + "type": "text", + "bbox": [ + 0.346, + 0.204, + 0.455, + 0.221 + ], + "angle": 0, + "content": "GeonU Kim" + }, + { + "type": "text", + "bbox": [ + 0.496, + 0.204, + 0.587, + 0.222 + ], + "angle": 0, + "content": "Kim Yu-Ji2" + }, + { + "type": "text", + "bbox": [ + 0.629, + 0.204, + 0.824, + 0.222 + ], + "angle": 0, + "content": "Yu-Chiang Frank Wang3" + }, + { + "type": "text", + "bbox": [ + 0.346, + 0.222, + 0.472, + 0.239 + ], + "angle": 0, + "content": "Jaesung Choe\\(^{3*}\\)" + }, + { + "type": "text", + "bbox": [ + 0.505, + 0.222, + 0.651, + 0.239 + ], + "angle": 0, + "content": "Tae-Hyun Oh\\(^{1,2,4*}\\)" + }, + { + "type": "text", + "bbox": [ + 0.185, + 0.239, + 0.815, + 0.257 + ], + "angle": 0, + "content": "\\(^{1}\\)Department of Electrical Engineering and \\(^{2}\\)Graduate School of AI, POSTECH" + }, + { + "type": "text", + "bbox": [ + 0.315, + 0.257, + 0.395, + 0.273 + ], + "angle": 0, + "content": "3NVIDIA" + }, + { + "type": "text", + "bbox": [ + 0.437, + 0.258, + 0.683, + 0.275 + ], + "angle": 0, + "content": "\\(^{4}\\)School of Computing, KAIST" + }, + { + "type": "text", + "bbox": [ + 0.315, + 0.278, + 0.683, + 0.292 + ], + "angle": 0, + "content": "{junseong.kim, gukim, ugkim}@postech.ac.kr" + }, + { + "type": "text", + "bbox": [ + 0.27, + 0.296, + 0.729, + 0.309 + ], + "angle": 0, + "content": "{frankwang, jchoe}@nvidia.com taehyun@kaist.ac.kr" + }, + { + "type": "title", + "bbox": [ + 0.248, + 0.327, + 0.326, + 0.342 + ], + "angle": 0, + "content": "Abstract" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.358, + 0.485, + 0.614 + ], + "angle": 0, + "content": "We introduce Dr. Splat, a novel approach for open-vocabulary 3D scene understanding leveraging 3D Gaussian Splatting. Unlike existing language-embedded 3DGS methods, which rely on a rendering process, our method directly associates language-aligned CLIP embeddings with 3D Gaussians for holistic 3D scene understanding. The key of our method is a language feature registration technique where CLIP embeddings are assigned to the dominant Gaussians intersected by each pixel-ray. Moreover, we integrate Product Quantization (PQ) trained on general large-scale image data to compactly represent embeddings without per-scene optimization. Experiments demonstrate that our approach significantly outperforms existing approaches in 3D perception benchmarks, such as open-vocabulary 3D semantic segmentation, 3D object localization, and 3D object selection tasks. For video results, please visit: https://drsplat.github.io/" + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.643, + 0.222, + 0.658 + ], + "angle": 0, + "content": "1. Introduction" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.668, + 0.484, + 0.88 + ], + "angle": 0, + "content": "Open-vocabulary 3D scene understanding represents a significant challenge in the field of computer vision, with applications spanning autonomous navigation, robotics, and augmented reality. This approach aims to enable the interpretation and referencing of 3D spatial information through natural language, allowing for applicability beyond a restricted set of predefined categories [2, 3, 28, 29, 33, 36, 40]. Previously, open-vocabulary 3D scene understanding has been explored using point-cloud-based methods [11, 15, 17, 27, 30, 35, 38]. Recently, the 3D Gaussian Splitting (3DGS) [16] has introduced a continuous representation integrated on explicit 3D Gaussians, which differs from traditional point-cloud approaches, enabling rapid progress in practical applications [41]. Current research has begun to explore" + }, + { + "type": "image", + "bbox": [ + 0.516, + 0.327, + 0.707, + 0.443 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.564, + 0.443, + 0.663, + 0.453 + ], + "angle": 0, + "content": "Render 2D images" + }, + { + "type": "image", + "bbox": [ + 0.517, + 0.453, + 0.707, + 0.507 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.58, + 0.508, + 0.651, + 0.516 + ], + "angle": 0, + "content": "B number of images" + }, + { + "type": "image_caption", + "bbox": [ + 0.731, + 0.327, + 0.892, + 0.341 + ], + "angle": 0, + "content": "3D search scheme (ours)" + }, + { + "type": "image", + "bbox": [ + 0.718, + 0.342, + 0.902, + 0.445 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.765, + 0.444, + 0.859, + 0.453 + ], + "angle": 0, + "content": "'Direct' 3D Search" + }, + { + "type": "image", + "bbox": [ + 0.761, + 0.453, + 0.862, + 0.474 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.781, + 0.474, + 0.842, + 0.513 + ], + "angle": 0, + "content": null + }, + { + "type": "table", + "bbox": [ + 0.518, + 0.523, + 0.905, + 0.589 + ], + "angle": 0, + "content": "
Search domainper-scene codebookFeature distill.SearchDB size
LERF [17]2Drequired~ 24hslowlarge
LangSplat [30]2Drequired~ 4hslowlarge
LEGaussians [34]2Drequired~ 4hslowlarge
OpenGaussian [37]3Drequired~ 1hfastsmall
Dr. Splat (Ours)3Dnone~ 10mfastsmall
" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.594, + 0.907, + 0.677 + ], + "angle": 0, + "content": "Figure 1. Comparison of 2D (left) vs. our direct 3D search (right) for open-vocabulary 3D scene understanding. The 2D approach relies on multiview rendering, incurring high computational costs. Our method directly links language features to 3D Gaussians, enabling efficient and complete spatial coverage. The table highlights Dr. Splat's superior efficiency over related methods." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.69, + 0.907, + 0.72 + ], + "angle": 0, + "content": "methods for associating language-based features with 3D Gaussian splats to enhance scene understanding capabilities." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.72, + 0.907, + 0.84 + ], + "angle": 0, + "content": "Several recent approaches [30, 34, 43] introduce 3D Gaussian representation [16] into the open-vocabulary scene understanding. This unique representation uses 3D Gaussians to achieve high-quality scene rendering, offering a more structured representation that addresses some limitations of point clouds. Building on this, these methods employ 2D vision-language models to transfer language knowledge to 3D Gaussians \"via rendered feature maps\"." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.84, + 0.909, + 0.902 + ], + "angle": 0, + "content": "Despite its promise, such rendering-based distillation methods [30, 34] share two limitations. First, we found that there is a discrepancy between optimized embeddings in 3D Gaussians and 2D language-aligned embeddings. This gap" + }, + { + "type": "page_footnote", + "bbox": [ + 0.109, + 0.887, + 0.242, + 0.9 + ], + "angle": 0, + "content": "*Corresponding Authors" + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "14137" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.09, + 0.092, + 0.485, + 0.198 + ], + "angle": 0, + "content": "arises mainly from an intermediate rendering step that may distort CLIP embeddings during training. Then, the reliance on rendering impedes holistic 3D scene understanding, additional task-processing such as 3D semantic segmentation and 3D object localization, and making full spatial coverage calculations less efficient than direct 3D Gaussian methods [37] including ours as illustrated in Fig. 1." + }, + { + "type": "text", + "bbox": [ + 0.093, + 0.199, + 0.487, + 0.529 + ], + "angle": 0, + "content": "To address this issue, this work proposes Dr. Splat. Our method bypasses the rendering stage, enabling direct interaction with 3D Gaussians for registering and referring the well-preserved language-aligned CLIP embeddings in the 3D space. This makes our Dr. Splat clearly distinguishable from prior works, facilitating a seamless integration of representative embeddings from 2D vision language models into the 3D spatial structure without compromising exhaustive rendering process that has been exploited [14, 30, 34, 41-43]. Moreover, we propose to use a Product Quantization (PQ) feature encoding method to represent embeddings compactly and efficiently without any per-scene optimization. Rather than storing full-length feature vectors or per-scene specifically compressed embeddings [14, 30, 34, 41-43], each Gaussian in our Dr. Splat stores an index from a pre-trained PQ, significantly reducing memory usage up to \\(6.25\\%\\) compression ratio. By preserving the richness of embeddings while reducing memory usage, PQ is integral to our framework's high scalability and its ability to perform 3D perception tasks, such as open-vocabulary 3D object localization, 3D object selection, and 3D semantic segmentation. Our contributions are summarized as follows:" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.533, + 0.483, + 0.579 + ], + "angle": 0, + "content": "- We propose Dr. Splat, direct registration and referencing of language-aligned features in 3D Gaussians, bypassing intermediate rendering and preserving feature accuracy." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.583, + 0.483, + 0.628 + ], + "angle": 0, + "content": "- We introduce the PQ encoding method for compact feature representation, reducing memory usage while maintaining essential 3D feature properties." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.631, + 0.484, + 0.677 + ], + "angle": 0, + "content": "- We present a novel evaluation protocol to assess accuracy of 3D localization and segmentation for 3D Gaussians, with pseudo-labeling methods and volume-aware metrics." + }, + { + "type": "list", + "bbox": [ + 0.09, + 0.533, + 0.484, + 0.677 + ], + "angle": 0, + "content": null + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.69, + 0.367, + 0.706 + ], + "angle": 0, + "content": "2. Related Work and Motivation" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.72, + 0.485, + 0.84 + ], + "angle": 0, + "content": "Language-based 3D scene understanding. Open-set 3D scene understanding has seen considerable advancements, with a focus on methods that leverage language knowledge into 3D representation such as point clouds, neural radiance fields (NeRF) [26], and Gaussian Splatting [16] for 3D comprehension. Point-based methods [5, 12, 15, 25, 27, 38, 39] in open-vocabulary settings process point cloud data trained from language embeddings [23, 31] for open-set categories." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.84, + 0.484, + 0.902 + ], + "angle": 0, + "content": "NeRF-based approaches [6, 17, 18, 21, 24, 32] leverage semantic embeddings from 2D foundation models, such as CLIP [31], LSeg [23] and DINO [1] for open-vocabulary understanding. While the rendering process enhances 2D" + }, + { + "type": "image", + "bbox": [ + 0.517, + 0.092, + 0.639, + 0.173 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.543, + 0.174, + 0.612, + 0.185 + ], + "angle": 0, + "content": "Input Image" + }, + { + "type": "image", + "bbox": [ + 0.64, + 0.092, + 0.759, + 0.172 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.661, + 0.174, + 0.738, + 0.185 + ], + "angle": 0, + "content": "(a) Rendering" + }, + { + "type": "image", + "bbox": [ + 0.76, + 0.102, + 0.882, + 0.172 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.777, + 0.174, + 0.864, + 0.185 + ], + "angle": 0, + "content": "(b) Registration" + }, + { + "type": "image_caption", + "bbox": [ + 0.512, + 0.196, + 0.907, + 0.252 + ], + "angle": 0, + "content": "Figure 2. Visualization of discrepancy in rendered 2D features and 3D features. Color indicates a cosine similarity score between query features from a text query and either (a) 3D features distilled by 2D rendering [30] or (b) directly registered 3D features." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.266, + 0.907, + 0.312 + ], + "angle": 0, + "content": "perception tasks, the implicit nature of NeRF constrains the holistic understanding of 3D structures and dominantly provides 'rendered' feature maps." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.312, + 0.908, + 0.403 + ], + "angle": 0, + "content": "3D Gaussian Splatting (3DGS) [16] has emerged as a promising rendering method, as well as a novel representation for open-vocabulary 3D scene understanding. Since this research is the close related work with our work, we first elucidate the preliminary of 3DGS, followed by focusing on language embedded 3DGS as follows." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.407, + 0.909, + 0.59 + ], + "angle": 0, + "content": "Preliminary of 3D Gaussian Splatting. 3DGS [16] encodes appearance and geometry of the target scene into the 3D Gaussian representation. Each 3D primitive representation is expressed as a 3D Gaussian distribution having mean \\(\\pmb{\\mu} = [x_{\\mu},y_{\\mu},z_{\\mu}]^{\\top}\\) for 3D position and covariance matrix \\(\\Sigma_{\\mathrm{3D}}\\in \\mathbb{R}^{3\\times 3}\\) for 3D volume, as well as the opacity value \\(\\alpha\\) and the color c. In particular, the covariance matrix is decomposed into the scale matrix \\(S\\in \\mathbb{R}^{3\\times 3}\\) and the rotation matrix \\(R\\in SO(3)\\), \\(\\Sigma_{\\mathrm{3D}} = RSS^{\\top}R^{\\top}\\). In brief, \\(N\\) numbers of 3D Gaussians can be parametrized as \\(\\Theta = \\{\\pmb {\\theta}_i\\}_{i = 1}^N = \\{\\pmb {\\mu}_i,S_i,R_i,\\alpha_i,\\mathbf{c}_i\\}_{i = 1}^N\\). 3D Gaussians \\(\\Theta\\) are used to render a 2D pixel color c computed as:" + }, + { + "type": "equation", + "bbox": [ + 0.522, + 0.601, + 0.907, + 0.629 + ], + "angle": 0, + "content": "\\[\n\\hat {\\mathbf {c}} (\\theta) = \\sum_ {i = 1} ^ {N} T _ {i} \\tilde {\\alpha} _ {i} \\mathbf {c} _ {i}, \\text {s . t .} \\tilde {\\alpha} _ {i} = \\alpha_ {i} \\exp \\left(- \\frac {1}{2} \\mathbf {d} ^ {\\top} \\Sigma_ {2 D} ^ {- 1} \\mathbf {d}\\right), \\tag {1}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.64, + 0.909, + 0.778 + ], + "angle": 0, + "content": "\\(T_{i}\\) is a transmittance, \\(\\tilde{\\alpha}_{i}\\) is an effective opacity value computed from the Gaussian's opacity \\(\\alpha\\), the pixel distance \\(\\mathbf{d} \\in \\mathbb{R}^{2 \\times 1}\\) from the target pixel to the projected center location of the Gaussian in pixel, and \\(\\Sigma_{2D}\\) is the 2D covariance matrix in the image domain obtained from the splatting algorithm [16, 44]. The 3D Gaussian parameters \\(\\Theta\\) of a scene are optimized by minimizing the rendering loss between the input image color \\(\\mathbf{c}\\) and the rendered color \\(\\hat{\\mathbf{c}}(\\theta)\\) in Eq. (1) as \\(\\arg \\min_{\\theta} \\| \\mathbf{c} - \\hat{\\mathbf{c}}(\\theta) \\|_F^2\\)." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.78, + 0.911, + 0.901 + ], + "angle": 0, + "content": "Language embedded 3D Gaussian Splitting. The basic idea of the language embedded Gaussian representation [9, 14, 22, 30, 34, 41-43] is to replace the color rendering to language embedding rendering. Language embedded 3D Gaussians are parameterized as \\(\\Phi = \\{\\theta_i,\\tilde{\\mathbf{f}}_i\\}_{i = 1}^N = \\{\\pmb {\\mu}_i,S_i,R_i,\\alpha_i,\\mathbf{c}_i,\\tilde{\\mathbf{f}}_i\\}_{i = 1}^N\\) where \\(\\tilde{\\mathbf{f}}_i\\) denotes Gaussian-registered language embeddings across \\(N\\) numbers 3D Gaussians which will be discussed soon. Then," + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.52, + 0.957 + ], + "angle": 0, + "content": "14138" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.113, + 0.09, + 0.29, + 0.227 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.121, + 0.237, + 0.294, + 0.252 + ], + "angle": 0, + "content": "(a) Preprocessing stage" + }, + { + "type": "image", + "bbox": [ + 0.328, + 0.093, + 0.655, + 0.228 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.526, + 0.237, + 0.656, + 0.252 + ], + "angle": 0, + "content": "(b) Training stage" + }, + { + "type": "image", + "bbox": [ + 0.658, + 0.094, + 0.876, + 0.228 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.09, + 0.256, + 0.908, + 0.298 + ], + "angle": 0, + "content": "Figure 3. Overview of Dr. Splat. (a) In the preprocessing stage, we compute optimized 3D Gaussians [16] and Product Quantization (PQ) codebook construction. (b) During training, we extract CLIP embeddings from given images \\(\\{\\mathbf{I}\\}\\), and then proceed feature registration process (Sec. 3.1). Finally, we obtain 3D Gaussians \\(\\Phi^{\\mathrm{ours}}\\) with PQ indices \\(\\{j\\}\\) (Sec. 3.2)." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.324, + 0.486, + 0.355 + ], + "angle": 0, + "content": "analogous to the color rendering Eq. (1), the language embedding rendering is expressed as:" + }, + { + "type": "equation", + "bbox": [ + 0.221, + 0.362, + 0.484, + 0.39 + ], + "angle": 0, + "content": "\\[\n\\hat {\\mathbf {f}} = \\sum_ {i = 1} ^ {N} T _ {i} \\tilde {\\alpha} _ {i} \\tilde {\\mathbf {f}} _ {i}, \\tag {2}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.089, + 0.398, + 0.485, + 0.626 + ], + "angle": 0, + "content": "where \\(\\hat{\\mathbf{f}}\\) denotes a rendered language embedding. Likewise, the Gaussian-registered language embeddings \\(\\{\\tilde{\\mathbf{f}}\\}\\) are optimized by minimizing the rendering loss between the 2D language embedding \\(\\mathbf{f}\\) extracted from an input image and a rendered language embedding map \\(\\hat{\\mathbf{f}}\\) as \\(\\arg \\min_{\\{\\tilde{\\mathbf{f}}\\}}\\| \\mathbf{f} - \\hat{\\mathbf{f}} \\|_F^2\\) at each corresponding pixel. This can be regarded as distilling vision language models into Gaussian-registered language embedding \\(\\tilde{\\mathbf{f}}\\) through volume rendering Eq. (2). The Gaussian-registered language embeddings are separately trained after pre-training and fixing the pre-trained 3DGS \\(\\Theta\\) for a scene. The language embeddings to be distilled are typically obtained from CLIP [31]. Since storing 32-bit 512-D CLIP features \\(\\mathbf{f}\\) in every 3D Gaussians is memory-expensive, one can use a compressed feature per scene depending on the needs [14, 19, 34, 41-43]." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.629, + 0.484, + 0.886 + ], + "angle": 0, + "content": "Motivation. Such language-embedded radiance fields provide useful representation and language interfaces for many practical and crucial applications. While most of existing works focus on the training efficiency, the complexity in inference time has barely been discussed. Considering a scenario to text-query a 3D location of the language-embedded Gaussians, i.e., 3D localization, the aforementioned methods first require rendering a 2D language embedding map at each specific camera pose. We cannot directly retrieve over the distributed embeddings \\(\\{\\tilde{\\mathbf{f}}_i\\}\\) in 3D Gaussians, because the embeddings do not carry language information directly, but their weighted summed (rendered) features \\(\\hat{\\mathbf{f}}\\) do. This issue becomes even severer with compressed features as in [30]: their decompression decoders are not designed for and incompatible with directly applying to the distributed compressed language embeddings in each 3D Gaussian, yielding degenerated CLIP decoding (refer to Fig. 2)." + }, + { + "type": "text", + "bbox": [ + 0.11, + 0.886, + 0.483, + 0.901 + ], + "angle": 0, + "content": "This introduces multiple challenges and hassles. First, it" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.324, + 0.907, + 0.43 + ], + "angle": 0, + "content": "is challenging to find the best or proper camera rendering views that contain the object to find. One may attempt to pre-compute the minimal number of cameras and their camera poses that cover all the 3D Gaussians in a scene with proper resolutions, similarly by point-based approach [11]. However, this is a well-known set covering problem [7] with constraints which is known to be an NP-hard problem." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.432, + 0.909, + 0.657 + ], + "angle": 0, + "content": "Second, even with pre-computed rendered views, the retrieval complexity over the rendered images remains substantial [8]. Suppose a scene consisting of one million Gaussians, but just a single rendered language embedding map in pixel domain already has nearly a million pixels; thus, we need a dedicated system to efficiently retrieve over all the views. Third, since the retrieval is conducted in the 2D space, to find a 3D location, we need a separate mechanism to lift the localization to the 3D space, i.e., increasing the system complexity. In addition, 32-bit floating 512-Dimension CLIP features for millions of Gaussian are memory intensive, which is often not manageable. To reduce this burden, the existing methods [37] apply compressions with per-scene optimized codebooks, which hinders extension or generalization to other scenes." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.66, + 0.909, + 0.751 + ], + "angle": 0, + "content": "To overcome these, we propose a training-free algorithm for the direct allocation of language embeddings to 3D Gaussians, allowing efficient computation and interaction within the 3D space. As a concurrent work, OpenGaussian [37] tackles a similar challenge with our work, but still requires per-scene codebook construction Fig. 1." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.768, + 0.614, + 0.786 + ], + "angle": 0, + "content": "3. Dr. Splat" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.795, + 0.909, + 0.901 + ], + "angle": 0, + "content": "This section provides details of our method. We first explain how we directly register CLIP embeddings into Gaussian-registered language embeddings, Sec. 3.1. Then, we introduce Product Quantization (PQ) into our framework to efficiently store Gaussian-registered language embeddings, Sec. 3.2. Lastly, we describe the inference stage for text query-based 3D Gaussian localization, Sec. 3.3." + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "14139" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.141, + 0.089, + 0.345, + 0.235 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.11, + 0.236, + 0.366, + 0.251 + ], + "angle": 0, + "content": "(a) Map CLIP features to Gaussians" + }, + { + "type": "image", + "bbox": [ + 0.36, + 0.088, + 0.882, + 0.234 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.383, + 0.236, + 0.625, + 0.252 + ], + "angle": 0, + "content": "(b) Aggregate multiview features" + }, + { + "type": "image_caption", + "bbox": [ + 0.676, + 0.236, + 0.882, + 0.251 + ], + "angle": 0, + "content": "(c) Register feature to \\(\\Phi^{\\text{ours}}\\)" + }, + { + "type": "image_caption", + "bbox": [ + 0.09, + 0.255, + 0.908, + 0.298 + ], + "angle": 0, + "content": "Figure 4. Feature registration process in Dr. Splat. (a) We first map per-pixel CLIP embeddings \\(\\{\\mathbf{f}^{\\mathrm{map}}\\}\\) to Gaussians. Here, we only map dominant \\(k\\) Gaussians along pixel ray \\(r\\), named Top-\\(k\\) Gaussians. (b) After collecting embeddings, we compute aggregated features (Eq. (6)). (c) Finally, we re-use PQ to obtain the PQ indices \\(j\\) of aggregated features and update Gaussian parameters \\(\\Phi^{\\mathrm{ours}}\\)." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.324, + 0.342, + 0.34 + ], + "angle": 0, + "content": "3.1. Feature registration process" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.346, + 0.485, + 0.591 + ], + "angle": 0, + "content": "Our goal is to reconstruct a language embedded 3D space represented by 3D Gaussians \\(\\Phi\\), which we can directly interact in 3D space without feature rendering Eq. (2). For that, following LangSplat [30], we begin by extracting per-pixel CLIP embedding maps \\(\\mathbf{F}^{\\mathrm{map}} \\in \\mathbb{R}^{D \\times H \\times W}\\) from training images of the target scenes, where \\(D\\) is the dimension of CLIP embeddings, \\(H\\) and \\(W\\) are the height and width of the training images. Given training images, we extracts a dictionary of binary masks and language embeddings extracted from the images as: \\(\\mathcal{F}^{\\mathrm{map}} = \\{\\mathbf{M}_j : \\mathbf{f}_j^{\\mathrm{map}} \\mid j = 1, \\dots, M\\}\\), where \\(\\mathbf{M}_j \\in \\mathbb{R}^{H \\times W}\\) is a binary mask extracted using SAM [20] and \\(\\mathbf{f}_j^{\\mathrm{map}} \\in \\mathbb{R}^D\\) is a corresponding CLIP embedding from a cropped image with \\(\\mathbf{M}_j\\). Each mask \\(\\mathbf{M}_j\\) belongs to an image, and the masks are not overlapped to each other. With this dictionary, a CLIP embedding map \\(\\mathbf{F}^{\\mathrm{map}}(\\mathbf{I}, \\mathbf{r})\\) at a pixel \\(\\mathbf{r}\\) in a training image \\(\\mathbf{I}\\) is computed as:" + }, + { + "type": "equation", + "bbox": [ + 0.163, + 0.603, + 0.484, + 0.632 + ], + "angle": 0, + "content": "\\[\n\\mathbf {F} ^ {\\text {m a p}} (\\mathbf {I}, \\mathbf {r}) = \\sum_ {j = 1} ^ {M} \\mathbf {M} _ {j} (\\mathbf {I}, \\mathbf {r}) \\cdot \\mathbf {f} _ {j} ^ {\\text {m a p}}, \\tag {3}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.644, + 0.484, + 0.704 + ], + "angle": 0, + "content": "where \\(\\mathbf{M}_j(\\mathbf{I},\\mathbf{r})\\in \\{0,1\\}\\) indicates whether the mask \\(\\mathbf{M}_j\\) contains the pixel \\(\\mathbf{r}\\) in the image \\(\\mathbf{I}\\). Using \\(\\mathbf{F}^{\\mathrm{map}}\\), we reconstruct language embedded 3D Gaussians via a novel feature registration process as visualized in Fig. 3." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.706, + 0.484, + 0.872 + ], + "angle": 0, + "content": "During the feature registration process, our algorithm iterates through training images of the scene. Using projection relation, we link 3D Gaussians \\(\\Phi\\) to CLIP embeddings. Each Gaussian can link to multiple CLIP embeddings derived from different images. Then we aggregate collected embeddings to a single embedding to be assigned to each Gaussian. To ensure a consistent aggregation of the embeddings from multi-view images, we first compute a weight \\(w_{i}(\\mathbf{I},\\mathbf{r})\\) representing the contribution of \\(\\theta_{i}\\) to construct each pixel \\(\\mathbf{r}\\) in a training image \\(\\mathbf{I}\\). The weights are computed with the volume rendering equation Eq. (1) as:" + }, + { + "type": "equation", + "bbox": [ + 0.189, + 0.886, + 0.484, + 0.902 + ], + "angle": 0, + "content": "\\[\nw _ {i} (\\mathbf {I}, \\mathbf {r}) = T _ {i} (\\mathbf {I}, \\mathbf {r}) \\cdot \\tilde {\\alpha} _ {i} (\\mathbf {I}, \\mathbf {r}), \\tag {4}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.324, + 0.907, + 0.43 + ], + "angle": 0, + "content": "where \\( T_{i}(\\mathbf{I},\\mathbf{r}) \\) and \\( \\tilde{\\alpha}_{i}(\\mathbf{I},\\mathbf{r}) \\) are the transmittance and the effective opacity value of \\( \\theta_{i} \\) for a pixel \\( \\mathbf{r} \\) in an image \\( \\mathbf{I} \\), stated in Eq. (1). With the per-pixel weights, we calculate \\( w_{ij} \\) representing a weight between each Gaussian \\( \\theta_{i} \\) and corresponding language embedding maps \\( \\mathbf{f}_j^{\\mathrm{map}} \\), which is for aggregating CLIP embeddings from \\( \\mathbf{F}^{\\mathrm{map}} \\) and register the embedding to each Gaussian. The weights are computed as:" + }, + { + "type": "equation", + "bbox": [ + 0.572, + 0.441, + 0.907, + 0.466 + ], + "angle": 0, + "content": "\\[\nw _ {i j} = \\sum_ {\\mathbf {I} \\in \\mathcal {I}} \\sum_ {\\mathbf {r} \\in \\mathbf {I}} \\mathbf {M} _ {j} (\\mathbf {I}, \\mathbf {r}) \\cdot w _ {i} (\\mathbf {I}, \\mathbf {r}), \\tag {5}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.474, + 0.907, + 0.626 + ], + "angle": 0, + "content": "where \\(\\mathcal{I}\\) is the set of the training images. In this iterative process, we aggregate weights only for Top-\\(k\\) Gaussians with the highest weights \\(w_{i}(\\mathbf{I},\\mathbf{r})\\), along the ray of each pixel ray \\(\\mathbf{r}\\) (see Fig. 4). After aggregation, we prune the Gaussians which are not assigned any weight, i.e., \\(\\sum_{j=1}^{M} w_{ij} = 0\\). This summation aggregates weights between Gaussians and the CLIP embeddings by linking per-pixel weights \\(w_{i}(\\mathbf{I},\\mathbf{r})\\) of each Gaussian to its corresponding CLIP embeddings. With the obtained weights, we register an aggregated feature \\(\\dot{\\mathbf{f}}_i\\) to each Gaussian with weighted-averaging as:" + }, + { + "type": "equation", + "bbox": [ + 0.53, + 0.636, + 0.907, + 0.665 + ], + "angle": 0, + "content": "\\[\n\\dot {\\mathbf {f}} _ {i} = \\mathbf {f} _ {i} / | | \\mathbf {f} _ {i} | | _ {2}, \\text {w h e r e} \\quad \\mathbf {f} _ {i} = \\sum_ {j = 1} ^ {M} \\frac {w _ {i j}}{\\sum_ {k = 1} ^ {M} w _ {i k}} \\mathbf {f} _ {j} ^ {\\text {m a p}}. \\tag {6}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.673, + 0.909, + 0.794 + ], + "angle": 0, + "content": "This process enables 3D-aware feature registration to be consistent across various viewpoints, by aggregating features in the original high-dimensional feature space. The proposed process can be interpreted as an inverse volume rendering without gradient-based optimization, which enables our method to be faster than the prior methods requiring perscene gradient-based optimization [27, 30, 34] for feature registration in 3D space." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.803, + 0.842, + 0.82 + ], + "angle": 0, + "content": "3.2. Product-Quantized CLIP embeddings" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.825, + 0.908, + 0.901 + ], + "angle": 0, + "content": "Memory efficiency is a challenge in 3D scene representations, especially when associating Gaussians with high-dimensional feature vectors. LangSplat [30] addresses this by introducing an encoder-decoder network, while LeGaussian [34] and OpenGaussian [37] utilize codebook construc" + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "14140" + } + ], + [ + { + "type": "table", + "bbox": [ + 0.116, + 0.089, + 0.885, + 0.194 + ], + "angle": 0, + "content": "
MethodsmIoUmAcc @ 0.25
waldo_kitchenramenfigurinesteatimeMeanwaldo_kitchenramenfigurinesteatimeMean
LangSplat-m [30]8.296.118.3316.589.8313.6414.088.9327.1215.94
OpenGaussian [37]34.6023.8759.3354.4443.0650.0035.2180.3672.8859.61
Ours (Top-10)37.0524.3354.4257.3543.2963.6435.2180.3677.9764.30
Ours (Top-20)38.3324.5853.9456.1943.2663.6435.2182.1476.2764.32
Ours (Top-40)39.0724.7053.3657.2043.5863.6435.2180.3676.2763.87
" + }, + { + "type": "table_caption", + "bbox": [ + 0.09, + 0.204, + 0.907, + 0.247 + ], + "angle": 0, + "content": "Table 1. 3D object selection results on the LeRF-OVS dataset [17]. To measure 3D object selection performance, we calculate 2D segmentation accuracy on rendering of selected 3D Gaussians. Note that our model does not require per-scene optimization, demonstrating its robustness across diverse scenes. Bold and Underline stand for first and second best performance." + }, + { + "type": "image", + "bbox": [ + 0.115, + 0.272, + 0.252, + 0.485 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.253, + 0.272, + 0.373, + 0.485 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.376, + 0.272, + 0.496, + 0.485 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.503, + 0.272, + 0.638, + 0.485 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.64, + 0.272, + 0.759, + 0.485 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.763, + 0.272, + 0.883, + 0.485 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.09, + 0.489, + 0.907, + 0.545 + ], + "angle": 0, + "content": "Figure 5. Qualitative results of the object selection on the LeRF-OVS dataset [17]. We visualize rendering of selected 3D Gaussians for LangSplat [30], OpenGaussian [37], and ours. For LangSplat, activations are often distributed randomly, fail to localize the target. OpenGaussian often struggles to distinguish closely situated objects. In contrast, our model shows activations precisely limited to the queried object regions, effectively localizing only the relevant areas." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.572, + 0.486, + 0.648 + ], + "angle": 0, + "content": "tion. However, these approaches introduce additional per-scene computational costs for scene-specific parameter tuning of neural networks or codebooks (see Fig. 1). In contrast, we propose to use Product Quantization (PQ) on a large-scale image dataset, eliminating per-scene training." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.655, + 0.483, + 0.866 + ], + "angle": 0, + "content": "Product Quantization. PQ [13] is a widely used technique for efficient embedding compression, particularly valuable in large-scale applications. The PQ process begins by dividing the original \\(D\\)-dimensional feature vector \\(\\mathbf{v}\\) into \\(L\\) sub-vectors: \\(\\mathbf{v} = [\\mathbf{v}_1,\\mathbf{v}_2,\\dots ,\\mathbf{v}_L]\\). Each sub-vector \\(\\mathbf{v}_i\\) is then independently quantized to a predefined number of centroids \\(s_{ij}\\) in a predefined codebook \\(S_{i}\\) for that sub-vector. These centroids are learned via clustering, creating a codebook for each subspace. Once the centroids are established, each sub-vector is replaced by the index of the nearest centroid in its respective codebook. The centroid indices \\(j_{i} = [j_{i1},j_{i2},\\ldots ,j_{iL}]\\) are optimized by minimizing \\(\\arg \\min_k\\| \\mathbf{v}_i - s_{ik}\\|\\) to quantize a given vector \\(\\mathbf{v}_i\\) where \\(j_{ik}\\) is an 8-bit unsigned integer." + }, + { + "type": "text", + "bbox": [ + 0.091, + 0.871, + 0.484, + 0.901 + ], + "angle": 0, + "content": "Then, we can measure the distance between the query and data by adding distances between coarse centroids. Once the" + }, + { + "type": "image", + "bbox": [ + 0.527, + 0.571, + 0.646, + 0.766 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.647, + 0.571, + 0.766, + 0.772 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.768, + 0.573, + 0.885, + 0.772 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.513, + 0.774, + 0.907, + 0.843 + ], + "angle": 0, + "content": "Figure 6. Limitations of point-based IoU measurement. This figure shows the effect of removing the top and bottom \\(30\\%\\) of Gaussians according to the proposed significant score, implying that volume differences significantly impact 3D accuracy. The results highlight the need for the proposed IoU metric for 3D Gaussians." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.856, + 0.909, + 0.903 + ], + "angle": 0, + "content": "distances between centroids are computed as a lookup table, the computation shifts to simple indexing, which reduces the search complexity from \\(\\mathcal{O}(D)\\) to \\(\\mathcal{O}(1)\\) for a \\(D\\) dimen" + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.518, + 0.957 + ], + "angle": 0, + "content": "14141" + } + ], + [ + { + "type": "table", + "bbox": [ + 0.102, + 0.089, + 0.482, + 0.184 + ], + "angle": 0, + "content": "
3D\nmIoUIoU > 0.1519 classes\nIoU > 0.3IoU > 0.45
LangSplat-m [30]8.017.17.82.9
LEGaussians-m [34]9.519.18.97.3
OpenGaussian [37]25.259.538.018.3
Ours (Top-20)25.060.740.320.0
Ours (Top-40)25.460.740.325.6
" + }, + { + "type": "table_caption", + "bbox": [ + 0.208, + 0.185, + 0.372, + 0.198 + ], + "angle": 0, + "content": "(a) 3D object localization task." + }, + { + "type": "table", + "bbox": [ + 0.495, + 0.089, + 0.895, + 0.184 + ], + "angle": 0, + "content": "
19 classes15 classes10 classes
mIoUmAcc.mIoUmAcc.mIoUmAcc.
LangSplat-m [30]2.09.24.914.68.023.9
LEGaussians-m [34]1.67.94.616.17.724.9
OpenGaussian [37]30.146.538.156.849.771.4
Ours (Top-20)28.044.638.260.447.268.9
Ours (Top-40)29.647.738.260.450.273.5
" + }, + { + "type": "table_caption", + "bbox": [ + 0.552, + 0.185, + 0.833, + 0.197 + ], + "angle": 0, + "content": "(b) Open-vocabulary 3D semantic segmentation task." + }, + { + "type": "table_caption", + "bbox": [ + 0.09, + 0.201, + 0.907, + 0.245 + ], + "angle": 0, + "content": "Table 2. Quantitative comparison in the ScanNet dataset [4]. Left: Localization prediction is defined as 3D regions with a text similarity score above threshold. Right: We assign segmentation labels by finding max activations among all classes. Note that Bold and Underline stand for first and second best performance, respectively." + }, + { + "type": "image", + "bbox": [ + 0.11, + 0.254, + 0.885, + 0.393 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.09, + 0.396, + 0.908, + 0.439 + ], + "angle": 0, + "content": "Figure 7. Qualitative results of 3D object localization. We visualize 3D localization activations (yellow) for \"chair\" and \"desk\" in the ScanNet dataset, comparing our method with others. It turns out that LangSplat-m and LEGaussians-m fail to localize objects accurately, while OpenGaussian struggles with object correspondence. Our model delivers precise and consistent localization across diverse queries." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.464, + 0.483, + 0.494 + ], + "angle": 0, + "content": "sion sample. This approach notably reduces computational complexity, making it suitable for large-scale search." + }, + { + "type": "text", + "bbox": [ + 0.089, + 0.497, + 0.486, + 0.771 + ], + "angle": 0, + "content": "In our setup for language-based 3D scene understanding, we build PQ centroids based on CLIP embeddings using a large-scale image dataset, the LVIS dataset [10], that contains over 1.2M instances covering various long-tail classes and ground truth segmentation. We extract instance patches from images and collect patch-wise CLIP embeddings. After we build this CLIP embedding database, we proceed with the construction of the centroid codebook for our PQ. Once PQ is trained, any query embedding can be approximated by assigning the closest centroid for each subvector. This is a one-time procedure; once we determine the codebook, we can use it for any scene generally. In our setup, each embedding is represented as a sequence of centroid indices rather than a high-dimensional vector. Accordingly, our language embedded Gaussians are parametrized as \\(\\Phi^{\\mathrm{ours}} = \\{\\phi_i^{\\mathrm{ours}}\\}_{i=1}^N = \\{\\theta_i, j_i\\}_{i=1}^N\\). where the aggregated feature \\(\\hat{\\mathbf{f}}_i\\) are converted as a quantized feature \\(\\bar{\\mathbf{f}}_i\\) by the corresponding PQ index \\(j_i\\)." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.786, + 0.38, + 0.803 + ], + "angle": 0, + "content": "3.3. Text-query based 3D localization" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.81, + 0.486, + 0.902 + ], + "angle": 0, + "content": "After training 3D Gaussians \\(\\Phi^{\\mathrm{ours}}\\) with our feature registration process and PQ, we describe the details of an inference mode that facilitates direct interaction with 3DGS upon receiving input queries, such as text. This is related to similarity score computation between a query and sources, i.e. Gaussian embeddings. Given a text, we first extract a" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.464, + 0.907, + 0.525 + ], + "angle": 0, + "content": "query feature \\(\\mathbf{q}\\) using CLIP text encoder [31]. We reconstruct the quantized features \\(\\{\\bar{\\mathbf{f}}_i\\}_{i=1}^N\\) from the stored PQ indices \\(\\{j_i\\}_{i=1}^N\\). Then, we compute a cosine similarity score between the query feature \\(\\mathbf{q}\\) and all quantized features." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.525, + 0.909, + 0.741 + ], + "angle": 0, + "content": "Despite its simplicity, solely relying on the cosine similarity may result in diminished discriminability across certain similarity scores. To address this limitation, we incorporate a re-ranking process based on relative activation with respect to the canonical feature. For this process, we adopt the relevancy scoring method proposed in LeRF [17], which enables more precise similarity analysis for a query. Specifically, each rendered language embedding, \\(\\mathbf{f}^{\\mathrm{map}}\\) and a text query feature \\(\\mathbf{q}\\), yield a relevance score determined by \\(\\min_{i}\\frac{\\exp(\\mathbf{f}^{\\mathrm{map}}\\cdot\\mathbf{q})}{\\exp(\\mathbf{f}^{\\mathrm{map}}\\cdot\\mathbf{q}) + \\exp(\\mathbf{f}^{\\mathrm{map}}\\cdot\\mathbf{f}^{\\mathrm{canon,i}})}\\), where \\((\\cdot)\\) is an element-wise dot product operator and \\(\\mathbf{f}^{\\mathrm{canon,i}}\\) indicates CLIP embeddings of a designated canonical term selected from a set of \"object,\" \"things,\" \"stuff,\" and \"texture\". Then, we sample 3D Gaussians based on the relevance score for downstream tasks." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.755, + 0.646, + 0.772 + ], + "angle": 0, + "content": "4. Experiments" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.78, + 0.909, + 0.903 + ], + "angle": 0, + "content": "Dataset. We use two datasets to evaluate the 3D scene understanding performance. For the 3D object selection task (Sec. 4.1), we use the LERF [17] dataset annotated by LangSplat [30], which consists of several multi-view images of 3D scenes containing long-tail objects and includes ground truth 2D ground truth annotations for texture queries. For 3D object localization Sec. 4.2 and 3D semantic segmentation Sec. 4.3 task, we employ the ScanNet [4] dataset." + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.52, + 0.957 + ], + "angle": 0, + "content": "14142" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.108, + 0.089, + 0.885, + 0.229 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.09, + 0.231, + 0.907, + 0.275 + ], + "angle": 0, + "content": "Figure 8. Visualization of open-vocabulary 3D semantic segmentation on the ScanNet dataset [4]. We visualize 3D Gaussian splat-based semantic segmentation using language features allocation of OpenGaussian [37] and Dr. Splat (ours) model on the same RGB-pretrained 3DGS. Note that, not specifically designed for segmentation, it achieves high performance as a result of language-based Gaussian updates." + }, + { + "type": "text", + "bbox": [ + 0.089, + 0.3, + 0.486, + 0.361 + ], + "angle": 0, + "content": "ScanNet is a large-scale benchmark that provides data on indoor scenes, including calibrated RGBD images and 3D point clouds with ground-truth semantic labels. We randomly select eight scenes from ScanNet for the experiments." + }, + { + "type": "text", + "bbox": [ + 0.089, + 0.362, + 0.485, + 0.575 + ], + "angle": 0, + "content": "Competing methods. The only method available for a fair comparison with our method is the concurrent work, Open-Gaussian [37]. To study the various aspects of our method, we introduce baseline methods modified from rasterization-based ones [30, 34], for direct 3D referring operation, denoted as LangSplat-m and LEGaussians-m. As discussed in Sec. 2, without modification, global search over a whole scene is quite demanding. To ensure fair evaluation, we use the same initial 3D Gaussians being trained only using RGB inputs for all comparing methods, and freeze the Gaussians during the language feature allocation process. Also, the per-pixel CLIP [31] embedding maps are unified for SAM-based [20] methods [30, 37] including ours. We follow the hyperparameter settings favorable to each respective paper." + }, + { + "type": "title", + "bbox": [ + 0.09, + 0.588, + 0.273, + 0.604 + ], + "angle": 0, + "content": "4.1.3D object selection" + }, + { + "type": "text", + "bbox": [ + 0.089, + 0.611, + 0.486, + 0.732 + ], + "angle": 0, + "content": "Settings. We first extract text features from an open-vocabulary text query using the CLIP model. Next, we compare text features to the 3D features embedded in each Gaussian using cosine similarity. By thresholding the similarity, we identify the 3D Gaussians that are relevant to the given text query. The selected 3D points are subsequently rendered into multi-view images using the 3DGS rasterization pipeline." + }, + { + "type": "text", + "bbox": [ + 0.089, + 0.733, + 0.486, + 0.839 + ], + "angle": 0, + "content": "Results. We compare our model quantitatively with 3DGS-based language-embedded models as shown in Table 1. The results demonstrate that our method performs better object selection in most scenes, showing an improvement of over 0.5 in mIoU and more than 4.5 in mAcc compared to counterpart models. Notably, the rasterization-based method, LangSplat-m, often underperforms in most scenes." + }, + { + "type": "text", + "bbox": [ + 0.089, + 0.84, + 0.486, + 0.9 + ], + "angle": 0, + "content": "Qualitative results are shown in Fig. 5. For LangSplat-m, the activations often shows random 3D Gaussians or fail to localize entirely (e.g., see \"coffee mug\"), highlighting the limitations of rasterization-based methods and their un" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.3, + 0.907, + 0.406 + ], + "angle": 0, + "content": "suitability for 3D understanding, aligning the observation from Fig. 2. OpenGaussian frequently exhibits false activations with incorrect text-object pairs (e.g., \"apple\" and \"tea in a glass\") and struggles to distinguish between nearby objects (e.g., \"waldo,\" \"rubik's cube\"). This artifacts can be attributed to use of spatial clustering and limited encoder capacity." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.407, + 0.907, + 0.483 + ], + "angle": 0, + "content": "In contrast, our model leverages general image features thanks to the general PQ, maintaining feature distinctiveness regardless of scene complexity. Our feature registration considers the 3D geometry of the 3D Gaussians, which results in superior performance in 3D scene understanding tasks." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.497, + 0.719, + 0.513 + ], + "angle": 0, + "content": "4.2. 3D object localization" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.52, + 0.909, + 0.655 + ], + "angle": 0, + "content": "Settings. Similar to the 3D object selection task, we calculate the cosine similarity between text query and 3D features embedded in each Gaussian. By thresholding the similarity, we identify the 3D Gaussians relevant to the given text query. To measure volume-aware localization evaluation, we propose a protocol to measure the IoU of 3D Gaussians that expands the traditional metric of point cloud-based approaches by incorporating volumetric information of 3D Gaussians." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.658, + 0.909, + 0.779 + ], + "angle": 0, + "content": "Novel evaluation protocol for 3D localization in 3DGS. Unlike conventional evaluation protocol for the 3D localization task in point clouds, it is tricky to evaluate 3D localization performance in 3D Gaussians [16]. This is primarily due to the un-deterministic structure of Gaussian distribution. To address this issue, we compute 3DGS pseudo-labels for evaluating the 3DGS localization in a volume-aware way. The details can be found in the supplementary material." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.781, + 0.911, + 0.902 + ], + "angle": 0, + "content": "Given the ground truth, we measure IoU considering the spatial significance of each Gaussian and define a significant score \\( d_{i} \\) for each Gaussian \\( \\theta_{i} \\) with its scale \\( \\mathbf{s}_i = [s_{ix}, s_{iy}, s_{iy}] \\) and opacity \\( \\alpha_{i} \\) as \\( d_{i} = s_{ix}s_{iy}s_{iz}\\alpha_{i} \\), where \\( s_{ix}s_{iy}s_{iz} \\) denotes a relative ellipsoid volume of a Gaussian \\( \\theta_{i} \\). With the obtained significant scores \\( \\mathbf{d} = [d_1, d_2, \\dots, d_N] \\), we compute weighted IoU of 3D Gaussians to approximate volumes. The proposed metric is designed to assign a larger" + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.52, + 0.957 + ], + "angle": 0, + "content": "14143" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.09, + 0.092, + 0.485, + 0.181 + ], + "angle": 0, + "content": "weight to the Gaussians with higher significant scores, when measuring IoU. Figure 6 shows that the impact of each Gaussian on the scene extremely varies depending on their significant scores, which demonstrates the necessity of the proposed IoU metric on 3D Gaussians that regards unequal contributions of each Gaussian." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.183, + 0.484, + 0.349 + ], + "angle": 0, + "content": "Results. We report the 3D localization performance on the Scannet dataset in Table 2a. The 2D rasterization-based methods [30, 34] struggle to achieve precise activations for 3D localization. They inherently face challenges when applying for 3D tasks because they need to render 2D images for the scene interaction. Even with the 3D space search method, OpenGaussian [37], our model consistently demonstrates superior performance and achieves higher accuracy in localization. Figure 7 also shows that LangSplat-m and LEGaussians-m fail to properly localize the objects, and OpenGaussian misses queried objects in the scene." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.36, + 0.329, + 0.377 + ], + "angle": 0, + "content": "4.3. 3D semantic segmentation" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.383, + 0.484, + 0.429 + ], + "angle": 0, + "content": "Settings. For a given set of open-vocabulary text labels, we perform segmentation by assigning each Gaussian a label having the highest activation among the known label set." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.43, + 0.484, + 0.595 + ], + "angle": 0, + "content": "Results. The numerical comparison is presented in Table 2b. Although not explicitly designed for semantic segmentation, our model achieves notable performance in this task as a result of accurately updating each Gaussian with language features. Consistent with previous observations, rasterization-based 3DGS models exhibit lower segmentation performance. While OpenGaussian performs position-based clustering, our model demonstrates comparable performance, surpassing the baseline as the Top-\\(k\\) value increases. Our model also achieves better segmentation results, with a visual comparison of the segmented scene shown in Fig. 8." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.607, + 0.241, + 0.623 + ], + "angle": 0, + "content": "4.4. Ablation study" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.629, + 0.483, + 0.674 + ], + "angle": 0, + "content": "We conduct an ablation study using the ScanNet dataset on different hyper-parameters of Dr. Splat to measure the contribution of each component." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.675, + 0.484, + 0.901 + ], + "angle": 0, + "content": "Product Quantization. PQ introduces a trade-off between memory usage, computational efficiency, and accuracy. To better understand the balance between computational cost and localization quality, we conduct an ablation study by varying the number of sub-vectors. We evaluate performance at sub-vector sizes of 64, 128, and 256. Notably, these settings correspond to bit-size reductions of 1/32, 1/16, and 1/8 of the original CLIP feature, respectively. We measure the query distance computation time for one million data points, averaging results over 100 iterations for efficiency measure. Our findings reveal a favorable trade-off between quantization performance and accuracy (see Fig. 9-(b)) in the Pareto front with our PQ configurations. This achieves a balance that maximizes memory and computational efficiency while minimizing any loss in accuracy." + }, + { + "type": "image", + "bbox": [ + 0.523, + 0.09, + 0.707, + 0.179 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.529, + 0.182, + 0.694, + 0.194 + ], + "angle": 0, + "content": "(a) Ablation on PQ parameters." + }, + { + "type": "image", + "bbox": [ + 0.715, + 0.09, + 0.899, + 0.178 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.745, + 0.182, + 0.865, + 0.194 + ], + "angle": 0, + "content": "(b) Ablation on Top- \\(k\\)" + }, + { + "type": "image_caption", + "bbox": [ + 0.529, + 0.2, + 0.89, + 0.214 + ], + "angle": 0, + "content": "Figure 9. Ablation study on (a) PQ and (b) Top- \\(k\\) Gaussians." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.242, + 0.907, + 0.378 + ], + "angle": 0, + "content": "Top-\\(k\\) Gaussians. We examine the influence of the number of Gaussians assigned per ray. This parameter affects both memory requirements and computation, serving as a critical factor in overall performance. The ratio of pruned Gaussians and the mIoU results from different \\(k\\) are presented in Fig. 9a. We observe that increasing the aggregating number of Gaussians per ray improves localization performance; however, it results in higher memory consumption and the number of occupied Gaussians, indicating a clear trade-off." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.394, + 0.763, + 0.41 + ], + "angle": 0, + "content": "5. Discussion and Conclusion" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.42, + 0.907, + 0.556 + ], + "angle": 0, + "content": "We present Dr. Splat, which is a novel approach for open-vocabulary 3D scene understanding by directly registering language embeddings to 3D Gaussians, eliminating the need for an intermediate rendering process. Compared to the previous 2D rendering-based methods [30, 34], which have limited search domain and capacity, our method directly searches 3D space while preserving the fidelity of language embeddings. This operation is further accelerated by the integration of Product Quantization (PQ)" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.557, + 0.909, + 0.693 + ], + "angle": 0, + "content": "Experimental results validate Dr. Splat's superior performance across various 3D scene understanding tasks, including open-vocabulary 3D object selection, 3D object localization, and 3D semantic segmentation. These findings highlight Dr. Splat's ability to transform 3D scene understanding by achieving a balance between highly representative quality and computational efficiency. This breakthrough paves the way for advanced applications in robotics, autonomous navigation, and augmented reality." + }, + { + "type": "title", + "bbox": [ + 0.514, + 0.71, + 0.665, + 0.726 + ], + "angle": 0, + "content": "Acknowledgment" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.735, + 0.909, + 0.901 + ], + "angle": 0, + "content": "This work was supported by Institute of Information & communications Technology Planning & Evaluation (IITP) (No.RS-2021-II212068, Artificial Intelligence Innovation Hub, No.RS-2024-00457882, National AI Research Lab Project, No.RS-2023-00225630, Development of Artificial Intelligence for Text-based 3D Movie Generation, No.RS-2019-II191906, Artificial Intelligence Graduate School Program(POSTECH)) and by the National Research Foundation of Korea (NRF) (No.RS-2024-00358135, Corner Vision: Learning to Look Around the Corner through Multi-modal Signals) grant funded by the Korea government (MSIT)." + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "14144" + } + ], + [ + { + "type": "title", + "bbox": [ + 0.093, + 0.09, + 0.188, + 0.106 + ], + "angle": 0, + "content": "References" + }, + { + "type": "ref_text", + "bbox": [ + 0.101, + 0.115, + 0.486, + 0.184 + ], + "angle": 0, + "content": "[1] Mathilde Caron, Hugo Touvron, Ishan Misra, Hervé Jégou, Julien Mairal, Piotr Bojanowski, and Armand Joulin. Emerging properties in self-supervised vision transformers. In Proceedings of the IEEE/CVF international conference on computer vision, pages 9650-9660, 2021. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.101, + 0.186, + 0.484, + 0.241 + ], + "angle": 0, + "content": "[2] Jaesung Choe, Chunghyun Park, Francois Rameau, Jaesik Park, and In So Kweon. Pointmixer: Mlp-mixer for point cloud understanding. In European Conference on Computer Vision, pages 620-640. Springer, 2022. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.242, + 0.484, + 0.31 + ], + "angle": 0, + "content": "[3] Christopher Choy, JunYoung Gwak, and Silvio Savarese. 4d spatio-temporal convnets: Minkowski convolutional neural networks. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 3075-3084, 2019. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.312, + 0.484, + 0.367 + ], + "angle": 0, + "content": "[4] Angela Dai, Angel X Chang, Manolis Savva, Maciej Halber, Thomas Funkhouser, and Matthias Nießner. Scannet: Richly-annotated 3d reconstructions of indoor scenes. In CVPR, pages 5828-5839, 2017. 6, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.368, + 0.484, + 0.436 + ], + "angle": 0, + "content": "[5] Runyu Ding, Jihan Yang, Chuhui Xue, Wenqing Zhang, Song Bai, and Xiaojuan Qi. Pla: Language-driven open-vocabulary 3d scene understanding. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 7010-7019, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.438, + 0.484, + 0.505 + ], + "angle": 0, + "content": "[6] Francis Engelmann, Fabian Manhardt, Michael Niemeyer, Keisuke Tateno, Marc Pollefeys, and Federico Tombari. OpenNerf: Open Set 3D Neural Scene Segmentation with PixelWise Features and Rendered Novel Views. In ICLR, 2024. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.508, + 0.484, + 0.55 + ], + "angle": 0, + "content": "[7] Michael R. Garey and David S. Johnson. Computers and intractability. a guide to the theory of np-completeness. W. H. Freeman and company, 174, 1979. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.551, + 0.484, + 0.592 + ], + "angle": 0, + "content": "[8] Antoine Guédon, Tom Monnier, Pascal Monasse, and Vincent Lepetit. Macarons: Mapping and coverage anticipation with rgb online self-supervision. In CVPR, 2023. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.594, + 0.484, + 0.647 + ], + "angle": 0, + "content": "[9] Jun Guo, Xiaojian Ma, Yue Fan, Huaping Liu, and Qing Li. Semantic gaussians: Open-vocabulary scene understanding with 3d gaussian splatting. arXiv preprint arXiv:2403.15624, 2024. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.649, + 0.483, + 0.704 + ], + "angle": 0, + "content": "[10] Agrim Gupta, Piotr Dollar, and Ross Girshick. Lvis: A dataset for large vocabulary instance segmentation. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 5356-5364, 2019. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.705, + 0.484, + 0.773 + ], + "angle": 0, + "content": "[11] Zhening Huang, Xiaoyang Wu, Xi Chen, Hengshuang Zhao, Lei Zhu, and Joan Lasenby. Openins3d: Snap and lookup for 3d open-vocabulary instance segmentation. In European Conference on Computer Vision, pages 169-185. Springer, 2025. 1, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.776, + 0.484, + 0.844 + ], + "angle": 0, + "content": "[12] Krishna Murthy Jatavallabhula, Alihusein Kuwajerwala, Qiao Gu, Mohd Omama, Tao Chen, Alaa Maalouf, Shuang Li, Ganesh Iyer, Soroush Saryazdi, Nikhil Keetha, et al. Conceptfusion: Open-set multimodal 3d mapping. arXiv preprint arXiv:2302.07241, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.846, + 0.484, + 0.9 + ], + "angle": 0, + "content": "[13] Herve Jegou, Matthijs Douze, and Cordelia Schmid. Product quantization for nearest neighbor search. IEEE transactions on pattern analysis and machine intelligence, 33(1):117-128, 2010. 5" + }, + { + "type": "list", + "bbox": [ + 0.094, + 0.115, + 0.486, + 0.9 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.093, + 0.908, + 0.147 + ], + "angle": 0, + "content": "[14] Yuzhou Ji, He Zhu, Junshu Tang, Wuyi Liu, Zhizhong Zhang, Yuan Xie, and Xin Tan. Fastlgs: Speeding up language embedded gaussians with feature grid mapping. arXiv preprint arXiv:2406.01916, 2024. 2, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.149, + 0.908, + 0.204 + ], + "angle": 0, + "content": "[15] Li Jiang, Shaoshuai Shi, and Bernt Schiele. Open-vocabulary 3d semantic segmentation with foundation models. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 21284-21294, 2024. 1, 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.205, + 0.905, + 0.245 + ], + "angle": 0, + "content": "[16] Bernhard Kerbl, Georgios Kopanas, Thomas Leimkuhler, and George Drettakis. 3d gaussian splatting for real-time radiance field rendering. ACM TOG, 2023. 1, 2, 3, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.247, + 0.907, + 0.314 + ], + "angle": 0, + "content": "[17] Justin Kerr, Chung Min Kim, Ken Goldberg, Angjoo Kanazawa, and Matthew Tancik. Lerf: Language embedded radiance fields. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 19729-19739, 2023. 1, 2, 5, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.316, + 0.907, + 0.357 + ], + "angle": 0, + "content": "[18] GeonU Kim, Kim Youwang, and Tae-Hyun Oh. FPRF: Feedforward photorealistic style transfer of large-scale 3D neural radiance fields. In AAAI, 2024. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.358, + 0.905, + 0.399 + ], + "angle": 0, + "content": "[19] GeonU Kim, Kim Youwang, Lee Hyoseok, and Tae-Hyun Oh. Fpgs: Feed-forward semantic-aware photorealistic style transfer of large-scale gaussian splatting, 2025. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.401, + 0.907, + 0.455 + ], + "angle": 0, + "content": "[20] Alexander Kirillov, Eric Mintun, Nikhila Ravi, Hanzi Mao, Chloe Rolland, Laura Gustafson, Tete Xiao, Spencer Whitehead, Alexander C. Berg, Wan-Yen Lo, Piotr Dólar, and Ross B. Girshick. Segment anything. In ICCV, 2023. 4, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.456, + 0.908, + 0.51 + ], + "angle": 0, + "content": "[21] Sosuke Kobayashi, Eiichi Matsumoto, and Vincent Sitzmann. Decomposing nerf for editing via feature field distillation. Advances in Neural Information Processing Systems, 35:23311-23330, 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.511, + 0.905, + 0.552 + ], + "angle": 0, + "content": "[22] Hyunjee Lee, Youngsik Yun, Jeongmin Bae, Seoha Kim, and Youngjung Uh. Rethinking open-vocabulary segmentation of radiance fields in 3d space, 2024. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.554, + 0.905, + 0.594 + ], + "angle": 0, + "content": "[23] Boyi Li, Kilian Q Weinberger, Serge Belongie, Vladlen Koltun, and René Ranftl. Language-driven semantic segmentation. arXiv preprint arXiv:2201.03546, 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.596, + 0.908, + 0.663 + ], + "angle": 0, + "content": "[24] Kunhao Liu, Fangneng Zhan, Jiahui Zhang, Muyu Xu, Yingchen Yu, Abdulmotaleb El Saddik, Christian Theobalt, Eric Xing, and Shijian Lu. Weakly supervised 3d open-vocabulary segmentation. Advances in Neural Information Processing Systems, 36:53433-53456, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.665, + 0.908, + 0.746 + ], + "angle": 0, + "content": "[25] Minghua Liu, Yinhao Zhu, Hong Cai, Shizhong Han, Zhan Ling, Fatih Porikli, and Hao Su. Partslip: Low-shot part segmentation for 3d point clouds via pretrained image-language models. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 21736-21746, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.748, + 0.908, + 0.802 + ], + "angle": 0, + "content": "[26] Ben Mildenhall, Pratul P. Srinivasan, Matthew Tancik, Jonathan T. Barron, Ravi Ramamoorthi, and Ren Ng. Nerf: Representing scenes as neural radiance fields for view synthesis. In ECCV, 2020. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.804, + 0.908, + 0.873 + ], + "angle": 0, + "content": "[27] Songyou Peng, Kyle Genova, Chiyu Jiang, Andrea Tagliasacchi, Marc Pollefeys, Thomas Funkhouser, et al. Openscene: 3d scene understanding with open vocabularies. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 815-824, 2023. 1, 2, 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.874, + 0.905, + 0.901 + ], + "angle": 0, + "content": "[28] Charles R Qi, Hao Su, Kaichun Mo, and Leonidas J Guibas. Pointnet: Deep learning on point sets for 3d classification" + }, + { + "type": "list", + "bbox": [ + 0.517, + 0.093, + 0.908, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "14145" + } + ], + [ + { + "type": "ref_text", + "bbox": [ + 0.125, + 0.092, + 0.484, + 0.133 + ], + "angle": 0, + "content": "and segmentation. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 652-660, 2017. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.134, + 0.484, + 0.203 + ], + "angle": 0, + "content": "[29] Guocheng Qian, Yuchen Li, Houwen Peng, Jinjie Mai, Hasan Hammoud, Mohamed Elhoseiny, and Bernard Ghanem. Pointnext: Revisiting pointnet++ with improved training and scaling strategies. Advances in neural information processing systems, 35:23192-23204, 2022. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.204, + 0.484, + 0.272 + ], + "angle": 0, + "content": "[30] Minghan Qin, Wanhua Li, Jiawei Zhou, Haoqian Wang, and Hanspeter Pfister. Langsplat: 3d language gaussian splatting. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 20051-20060, 2024. 1, 2, 3, 4, 5, 6, 7, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.274, + 0.484, + 0.356 + ], + "angle": 0, + "content": "[31] Alec Radford, Jong Wook Kim, Chris Hallacy, Aditya Ramesh, Gabriel Goh, Sandhini Agarwal, Girish Sastry, Amanda Askell, Pamela Mishkin, Jack Clark, et al. Learning transferable visual models from natural language supervision. In International conference on machine learning, pages 8748-8763. PMLR, 2021. 2, 3, 6, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.357, + 0.484, + 0.426 + ], + "angle": 0, + "content": "[32] Adam Rashid, Satvik Sharma, Chung Min Kim, Justin Kerr, Lawrence Yunliang Chen, Angjoo Kanazawa, and Ken Goldberg. Language embedded radiance fields for zero-shot task-oriented grasping. In 7th Annual Conference on Robot Learning, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.427, + 0.484, + 0.482 + ], + "angle": 0, + "content": "[33] Damien Robert, Hugo Raguet, and Loic Landrieu. Efficient 3d semantic segmentation with superpoint transformer. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 17195-17204, 2023. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.483, + 0.484, + 0.55 + ], + "angle": 0, + "content": "[34] Jin-Chuan Shi, Miao Wang, Hao-Bin Duan, and Shao-Hua Guan. Language embedded 3d gaussians for open-vocabulary scene understanding. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 5333-5343, 2024. 1, 2, 3, 4, 6, 7, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.552, + 0.484, + 0.607 + ], + "angle": 0, + "content": "[35] Ayca Takmaz, Elisabetta Fedele, Robert W Sumner, Marc Pollefeys, Federico Tombari, and Francis Engelmann. Openmask3d: Open-vocabulary 3d instance segmentation. arXiv preprint arXiv:2306.13631, 2023. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.608, + 0.484, + 0.677 + ], + "angle": 0, + "content": "[36] Xiaoyang Wu, Li Jiang, Peng-Shuai Wang, Zhijian Liu, Xihui Liu, Yu Qiao, Wanli Ouyang, Tong He, and Hengshuang Zhao. Point transformer v3: Simpler faster stronger. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 4840-4851, 2024. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.679, + 0.484, + 0.747 + ], + "angle": 0, + "content": "[37] Yanmin Wu, Jiarui Meng, Hajjie Li, Chenming Wu, Yahao Shi, Xinhua Cheng, Chen Zhao, Haocheng Feng, Errui Ding, Jingdong Wang, et al. Opengaussian: Towards point-level 3d gaussian-based open vocabulary understanding. arXiv preprint arXiv:2406.02058, 2024. 1, 2, 3, 4, 5, 6, 7, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.748, + 0.484, + 0.816 + ], + "angle": 0, + "content": "[38] Jihan Yang, Runyu Ding, Weipeng Deng, Zhe Wang, and Xiaojuan Qi. Regionplc: Regional point-language contrastive learning for open-world 3d scene understanding. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 19823-19832, 2024. 1, 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.817, + 0.484, + 0.873 + ], + "angle": 0, + "content": "[39] Junbo Zhang, Runpei Dong, and Kaisheng Ma. Clip-for3d: Learning free open-world 3d scene representations from 2d dense clip. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 2048-2059, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.873, + 0.484, + 0.901 + ], + "angle": 0, + "content": "[40] Hengshuang Zhao, Li Jiang, Jiaya Jia, Philip HS Torr, and Vladlen Koltun. Point transformer. In Proceedings of" + }, + { + "type": "list", + "bbox": [ + 0.093, + 0.092, + 0.484, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.549, + 0.093, + 0.907, + 0.12 + ], + "angle": 0, + "content": "the IEEE/CVF international conference on computer vision, pages 16259-16268, 2021. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.122, + 0.907, + 0.217 + ], + "angle": 0, + "content": "[41] Yuhang Zheng, Xiangyu Chen, Yupeng Zheng, Songen Gu, Runyi Yang, Bu Jin, Pengfei Li, Chengliang Zhong, Zeng-mao Wang, Lina Liu, Chao Yang, Dawei Wang, Zhen Chen, Xiaoxiao Long, and Meiqing Wang. Gaussiangrasper: 3d language gaussian splatting for open-vocabulary robotic grasping. IEEE Robotics and Automation Letters, 2024. 1, 2, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.219, + 0.907, + 0.303 + ], + "angle": 0, + "content": "[42] Shijie Zhou, Haoran Chang, Sicheng Jiang, Zhiwen Fan, Zehao Zhu, Dejia Xu, Pradyumna Chari, Suya You, Zhangyang Wang, and Achuta Kadambi. Feature 3dgs: Supercharging 3d gaussian splattering to enable distilled feature fields. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 21676-21685, 2024." + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.304, + 0.907, + 0.358 + ], + "angle": 0, + "content": "[43] Xingxing Zuo, Pouya Samangouei, Yunwen Zhou, Yan Di, and Mingyang Li. Fmgs: Foundation model embedded 3d gaussian splatting for holistic 3d scene understanding. IJCV, 2024. 1, 2, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.36, + 0.906, + 0.402 + ], + "angle": 0, + "content": "[44] Matthias Zwicker, Hanspeter Pfister, Jeroen van Baar, and Markus H. Gross. EWA volume splatting. In Visualization Conference, 2001. 2" + }, + { + "type": "list", + "bbox": [ + 0.517, + 0.093, + 0.907, + 0.402 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.946, + 0.52, + 0.957 + ], + "angle": 0, + "content": "14146" + } + ] +] \ No newline at end of file diff --git a/2025/Dr. Splat_ Directly Referring 3D Gaussian Splatting via Direct Language Embedding Registration/04973834-dfc5-4d84-8238-2be21678c244_origin.pdf b/2025/Dr. Splat_ Directly Referring 3D Gaussian Splatting via Direct Language Embedding Registration/04973834-dfc5-4d84-8238-2be21678c244_origin.pdf new file mode 100644 index 0000000000000000000000000000000000000000..cd9b8d7d1b96064056766b2c2365c6e88eb546ca --- /dev/null +++ b/2025/Dr. Splat_ Directly Referring 3D Gaussian Splatting via Direct Language Embedding Registration/04973834-dfc5-4d84-8238-2be21678c244_origin.pdf @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:23eceb462a559e32a2acdc7a549c770f9c58b7f5d1527226e7c638faa0de5c0e +size 3414051 diff --git a/2025/Dr. Splat_ Directly Referring 3D Gaussian Splatting via Direct Language Embedding Registration/full.md b/2025/Dr. Splat_ Directly Referring 3D Gaussian Splatting via Direct Language Embedding Registration/full.md new file mode 100644 index 0000000000000000000000000000000000000000..adaee459d84084505008842e576b62799c173fde --- /dev/null +++ b/2025/Dr. Splat_ Directly Referring 3D Gaussian Splatting via Direct Language Embedding Registration/full.md @@ -0,0 +1,343 @@ +# Dr. Splat: Directly Referring 3D Gaussian Splatting via Direct Language Embedding Registration + +Kim Jun-Seong1 + +GeonU Kim + +Kim Yu-Ji2 + +Yu-Chiang Frank Wang3 + +Jaesung Choe $^{3*}$ + +Tae-Hyun Oh $^{1,2,4*}$ + +$^{1}$ Department of Electrical Engineering and $^{2}$ Graduate School of AI, POSTECH + +3NVIDIA + +$^{4}$ School of Computing, KAIST + +{junseong.kim, gukim, ugkim}@postech.ac.kr + +{frankwang, jchoe}@nvidia.com taehyun@kaist.ac.kr + +# Abstract + +We introduce Dr. Splat, a novel approach for open-vocabulary 3D scene understanding leveraging 3D Gaussian Splatting. Unlike existing language-embedded 3DGS methods, which rely on a rendering process, our method directly associates language-aligned CLIP embeddings with 3D Gaussians for holistic 3D scene understanding. The key of our method is a language feature registration technique where CLIP embeddings are assigned to the dominant Gaussians intersected by each pixel-ray. Moreover, we integrate Product Quantization (PQ) trained on general large-scale image data to compactly represent embeddings without per-scene optimization. Experiments demonstrate that our approach significantly outperforms existing approaches in 3D perception benchmarks, such as open-vocabulary 3D semantic segmentation, 3D object localization, and 3D object selection tasks. For video results, please visit: https://drsplat.github.io/ + +# 1. Introduction + +Open-vocabulary 3D scene understanding represents a significant challenge in the field of computer vision, with applications spanning autonomous navigation, robotics, and augmented reality. This approach aims to enable the interpretation and referencing of 3D spatial information through natural language, allowing for applicability beyond a restricted set of predefined categories [2, 3, 28, 29, 33, 36, 40]. Previously, open-vocabulary 3D scene understanding has been explored using point-cloud-based methods [11, 15, 17, 27, 30, 35, 38]. Recently, the 3D Gaussian Splitting (3DGS) [16] has introduced a continuous representation integrated on explicit 3D Gaussians, which differs from traditional point-cloud approaches, enabling rapid progress in practical applications [41]. Current research has begun to explore + +![](images/723f26eaf0279c7ee09a4c6a34fe5122c8fe1a6bf384a89d446ad64dff1aa6dc.jpg) +Render 2D images + +![](images/623dc797a4b1fa4dcb95bf70256b6dbedfe49900678e1c6327cee3ea495a5102.jpg) +B number of images + +![](images/7cf17d34a06bf835f2ec62c6165e1ed103649f856fa6babc777fcb4cd5454c9e.jpg) +3D search scheme (ours) +'Direct' 3D Search + +![](images/8f7765ced90b2d82f0f6133bc1fd5eab763b1a5ef1795b1d37fe67a51e9672ad.jpg) + +![](images/da76d519fcbeba25815c37e1aea704ca110e3d069c2533019f8af079e618a080.jpg) + +
Search domainper-scene codebookFeature distill.SearchDB size
LERF [17]2Drequired~ 24hslowlarge
LangSplat [30]2Drequired~ 4hslowlarge
LEGaussians [34]2Drequired~ 4hslowlarge
OpenGaussian [37]3Drequired~ 1hfastsmall
Dr. Splat (Ours)3Dnone~ 10mfastsmall
+ +Figure 1. Comparison of 2D (left) vs. our direct 3D search (right) for open-vocabulary 3D scene understanding. The 2D approach relies on multiview rendering, incurring high computational costs. Our method directly links language features to 3D Gaussians, enabling efficient and complete spatial coverage. The table highlights Dr. Splat's superior efficiency over related methods. + +methods for associating language-based features with 3D Gaussian splats to enhance scene understanding capabilities. + +Several recent approaches [30, 34, 43] introduce 3D Gaussian representation [16] into the open-vocabulary scene understanding. This unique representation uses 3D Gaussians to achieve high-quality scene rendering, offering a more structured representation that addresses some limitations of point clouds. Building on this, these methods employ 2D vision-language models to transfer language knowledge to 3D Gaussians "via rendered feature maps". + +Despite its promise, such rendering-based distillation methods [30, 34] share two limitations. First, we found that there is a discrepancy between optimized embeddings in 3D Gaussians and 2D language-aligned embeddings. This gap + +arises mainly from an intermediate rendering step that may distort CLIP embeddings during training. Then, the reliance on rendering impedes holistic 3D scene understanding, additional task-processing such as 3D semantic segmentation and 3D object localization, and making full spatial coverage calculations less efficient than direct 3D Gaussian methods [37] including ours as illustrated in Fig. 1. + +To address this issue, this work proposes Dr. Splat. Our method bypasses the rendering stage, enabling direct interaction with 3D Gaussians for registering and referring the well-preserved language-aligned CLIP embeddings in the 3D space. This makes our Dr. Splat clearly distinguishable from prior works, facilitating a seamless integration of representative embeddings from 2D vision language models into the 3D spatial structure without compromising exhaustive rendering process that has been exploited [14, 30, 34, 41-43]. Moreover, we propose to use a Product Quantization (PQ) feature encoding method to represent embeddings compactly and efficiently without any per-scene optimization. Rather than storing full-length feature vectors or per-scene specifically compressed embeddings [14, 30, 34, 41-43], each Gaussian in our Dr. Splat stores an index from a pre-trained PQ, significantly reducing memory usage up to $6.25\%$ compression ratio. By preserving the richness of embeddings while reducing memory usage, PQ is integral to our framework's high scalability and its ability to perform 3D perception tasks, such as open-vocabulary 3D object localization, 3D object selection, and 3D semantic segmentation. Our contributions are summarized as follows: + +- We propose Dr. Splat, direct registration and referencing of language-aligned features in 3D Gaussians, bypassing intermediate rendering and preserving feature accuracy. +- We introduce the PQ encoding method for compact feature representation, reducing memory usage while maintaining essential 3D feature properties. +- We present a novel evaluation protocol to assess accuracy of 3D localization and segmentation for 3D Gaussians, with pseudo-labeling methods and volume-aware metrics. + +# 2. Related Work and Motivation + +Language-based 3D scene understanding. Open-set 3D scene understanding has seen considerable advancements, with a focus on methods that leverage language knowledge into 3D representation such as point clouds, neural radiance fields (NeRF) [26], and Gaussian Splatting [16] for 3D comprehension. Point-based methods [5, 12, 15, 25, 27, 38, 39] in open-vocabulary settings process point cloud data trained from language embeddings [23, 31] for open-set categories. + +NeRF-based approaches [6, 17, 18, 21, 24, 32] leverage semantic embeddings from 2D foundation models, such as CLIP [31], LSeg [23] and DINO [1] for open-vocabulary understanding. While the rendering process enhances 2D + +![](images/2fdf2de7718c75f0ce02338691c88a40088d2c4d6c2ecbccebfa85da3bf4c2b5.jpg) +Input Image +Figure 2. Visualization of discrepancy in rendered 2D features and 3D features. Color indicates a cosine similarity score between query features from a text query and either (a) 3D features distilled by 2D rendering [30] or (b) directly registered 3D features. + +![](images/f0afa253f25e50670761bad93d295c56c96e5835de9bf78550bebad21f8d2a3f.jpg) +(a) Rendering + +![](images/a727fe67c7febb26a65755ae330ab24c80f1d7a693cd511d07889dcf75e8a6ab.jpg) +(b) Registration + +perception tasks, the implicit nature of NeRF constrains the holistic understanding of 3D structures and dominantly provides 'rendered' feature maps. + +3D Gaussian Splatting (3DGS) [16] has emerged as a promising rendering method, as well as a novel representation for open-vocabulary 3D scene understanding. Since this research is the close related work with our work, we first elucidate the preliminary of 3DGS, followed by focusing on language embedded 3DGS as follows. + +Preliminary of 3D Gaussian Splatting. 3DGS [16] encodes appearance and geometry of the target scene into the 3D Gaussian representation. Each 3D primitive representation is expressed as a 3D Gaussian distribution having mean $\pmb{\mu} = [x_{\mu},y_{\mu},z_{\mu}]^{\top}$ for 3D position and covariance matrix $\Sigma_{\mathrm{3D}}\in \mathbb{R}^{3\times 3}$ for 3D volume, as well as the opacity value $\alpha$ and the color c. In particular, the covariance matrix is decomposed into the scale matrix $S\in \mathbb{R}^{3\times 3}$ and the rotation matrix $R\in SO(3)$ , $\Sigma_{\mathrm{3D}} = RSS^{\top}R^{\top}$ . In brief, $N$ numbers of 3D Gaussians can be parametrized as $\Theta = \{\pmb {\theta}_i\}_{i = 1}^N = \{\pmb {\mu}_i,S_i,R_i,\alpha_i,\mathbf{c}_i\}_{i = 1}^N$ . 3D Gaussians $\Theta$ are used to render a 2D pixel color c computed as: + +$$ +\hat {\mathbf {c}} (\theta) = \sum_ {i = 1} ^ {N} T _ {i} \tilde {\alpha} _ {i} \mathbf {c} _ {i}, \text {s . t .} \tilde {\alpha} _ {i} = \alpha_ {i} \exp \left(- \frac {1}{2} \mathbf {d} ^ {\top} \Sigma_ {2 D} ^ {- 1} \mathbf {d}\right), \tag {1} +$$ + +$T_{i}$ is a transmittance, $\tilde{\alpha}_{i}$ is an effective opacity value computed from the Gaussian's opacity $\alpha$ , the pixel distance $\mathbf{d} \in \mathbb{R}^{2 \times 1}$ from the target pixel to the projected center location of the Gaussian in pixel, and $\Sigma_{2D}$ is the 2D covariance matrix in the image domain obtained from the splatting algorithm [16, 44]. The 3D Gaussian parameters $\Theta$ of a scene are optimized by minimizing the rendering loss between the input image color $\mathbf{c}$ and the rendered color $\hat{\mathbf{c}}(\theta)$ in Eq. (1) as $\arg \min_{\theta} \| \mathbf{c} - \hat{\mathbf{c}}(\theta) \|_F^2$ . + +Language embedded 3D Gaussian Splitting. The basic idea of the language embedded Gaussian representation [9, 14, 22, 30, 34, 41-43] is to replace the color rendering to language embedding rendering. Language embedded 3D Gaussians are parameterized as $\Phi = \{\theta_i,\tilde{\mathbf{f}}_i\}_{i = 1}^N = \{\pmb {\mu}_i,S_i,R_i,\alpha_i,\mathbf{c}_i,\tilde{\mathbf{f}}_i\}_{i = 1}^N$ where $\tilde{\mathbf{f}}_i$ denotes Gaussian-registered language embeddings across $N$ numbers 3D Gaussians which will be discussed soon. Then, + +![](images/bc16a901451bbd7d47666029bb23cb06177bb3d61c9169e04f64d082bd270c42.jpg) +(a) Preprocessing stage + +![](images/aca0342daac3b053a078af98d85ea144768656025c72e696d9355079d86dd4ea.jpg) +Figure 3. Overview of Dr. Splat. (a) In the preprocessing stage, we compute optimized 3D Gaussians [16] and Product Quantization (PQ) codebook construction. (b) During training, we extract CLIP embeddings from given images $\{\mathbf{I}\}$ , and then proceed feature registration process (Sec. 3.1). Finally, we obtain 3D Gaussians $\Phi^{\mathrm{ours}}$ with PQ indices $\{j\}$ (Sec. 3.2). + +![](images/f54340f180fe2b3c7e15a18582f5fd220e0f09d837291846a941ec1fdcc95234.jpg) +(b) Training stage + +analogous to the color rendering Eq. (1), the language embedding rendering is expressed as: + +$$ +\hat {\mathbf {f}} = \sum_ {i = 1} ^ {N} T _ {i} \tilde {\alpha} _ {i} \tilde {\mathbf {f}} _ {i}, \tag {2} +$$ + +where $\hat{\mathbf{f}}$ denotes a rendered language embedding. Likewise, the Gaussian-registered language embeddings $\{\tilde{\mathbf{f}}\}$ are optimized by minimizing the rendering loss between the 2D language embedding $\mathbf{f}$ extracted from an input image and a rendered language embedding map $\hat{\mathbf{f}}$ as $\arg \min_{\{\tilde{\mathbf{f}}\}}\| \mathbf{f} - \hat{\mathbf{f}} \|_F^2$ at each corresponding pixel. This can be regarded as distilling vision language models into Gaussian-registered language embedding $\tilde{\mathbf{f}}$ through volume rendering Eq. (2). The Gaussian-registered language embeddings are separately trained after pre-training and fixing the pre-trained 3DGS $\Theta$ for a scene. The language embeddings to be distilled are typically obtained from CLIP [31]. Since storing 32-bit 512-D CLIP features $\mathbf{f}$ in every 3D Gaussians is memory-expensive, one can use a compressed feature per scene depending on the needs [14, 19, 34, 41-43]. + +Motivation. Such language-embedded radiance fields provide useful representation and language interfaces for many practical and crucial applications. While most of existing works focus on the training efficiency, the complexity in inference time has barely been discussed. Considering a scenario to text-query a 3D location of the language-embedded Gaussians, i.e., 3D localization, the aforementioned methods first require rendering a 2D language embedding map at each specific camera pose. We cannot directly retrieve over the distributed embeddings $\{\tilde{\mathbf{f}}_i\}$ in 3D Gaussians, because the embeddings do not carry language information directly, but their weighted summed (rendered) features $\hat{\mathbf{f}}$ do. This issue becomes even severer with compressed features as in [30]: their decompression decoders are not designed for and incompatible with directly applying to the distributed compressed language embeddings in each 3D Gaussian, yielding degenerated CLIP decoding (refer to Fig. 2). + +This introduces multiple challenges and hassles. First, it + +is challenging to find the best or proper camera rendering views that contain the object to find. One may attempt to pre-compute the minimal number of cameras and their camera poses that cover all the 3D Gaussians in a scene with proper resolutions, similarly by point-based approach [11]. However, this is a well-known set covering problem [7] with constraints which is known to be an NP-hard problem. + +Second, even with pre-computed rendered views, the retrieval complexity over the rendered images remains substantial [8]. Suppose a scene consisting of one million Gaussians, but just a single rendered language embedding map in pixel domain already has nearly a million pixels; thus, we need a dedicated system to efficiently retrieve over all the views. Third, since the retrieval is conducted in the 2D space, to find a 3D location, we need a separate mechanism to lift the localization to the 3D space, i.e., increasing the system complexity. In addition, 32-bit floating 512-Dimension CLIP features for millions of Gaussian are memory intensive, which is often not manageable. To reduce this burden, the existing methods [37] apply compressions with per-scene optimized codebooks, which hinders extension or generalization to other scenes. + +To overcome these, we propose a training-free algorithm for the direct allocation of language embeddings to 3D Gaussians, allowing efficient computation and interaction within the 3D space. As a concurrent work, OpenGaussian [37] tackles a similar challenge with our work, but still requires per-scene codebook construction Fig. 1. + +# 3. Dr. Splat + +This section provides details of our method. We first explain how we directly register CLIP embeddings into Gaussian-registered language embeddings, Sec. 3.1. Then, we introduce Product Quantization (PQ) into our framework to efficiently store Gaussian-registered language embeddings, Sec. 3.2. Lastly, we describe the inference stage for text query-based 3D Gaussian localization, Sec. 3.3. + +![](images/6a58cf2fe02beb59208dadb777373d4f5453e42fdd45b6e7b9a238b3f49badaf.jpg) +(a) Map CLIP features to Gaussians +Figure 4. Feature registration process in Dr. Splat. (a) We first map per-pixel CLIP embeddings $\{\mathbf{f}^{\mathrm{map}}\}$ to Gaussians. Here, we only map dominant $k$ Gaussians along pixel ray $r$ , named Top- $k$ Gaussians. (b) After collecting embeddings, we compute aggregated features (Eq. (6)). (c) Finally, we re-use PQ to obtain the PQ indices $j$ of aggregated features and update Gaussian parameters $\Phi^{\mathrm{ours}}$ . + +![](images/4971ba003bc5b4ef9be830de69b10f0accca954ab1a688cecfcc1fc394f5d2b0.jpg) +(b) Aggregate multiview features +(c) Register feature to $\Phi^{\text{ours}}$ + +# 3.1. Feature registration process + +Our goal is to reconstruct a language embedded 3D space represented by 3D Gaussians $\Phi$ , which we can directly interact in 3D space without feature rendering Eq. (2). For that, following LangSplat [30], we begin by extracting per-pixel CLIP embedding maps $\mathbf{F}^{\mathrm{map}} \in \mathbb{R}^{D \times H \times W}$ from training images of the target scenes, where $D$ is the dimension of CLIP embeddings, $H$ and $W$ are the height and width of the training images. Given training images, we extracts a dictionary of binary masks and language embeddings extracted from the images as: $\mathcal{F}^{\mathrm{map}} = \{\mathbf{M}_j : \mathbf{f}_j^{\mathrm{map}} \mid j = 1, \dots, M\}$ , where $\mathbf{M}_j \in \mathbb{R}^{H \times W}$ is a binary mask extracted using SAM [20] and $\mathbf{f}_j^{\mathrm{map}} \in \mathbb{R}^D$ is a corresponding CLIP embedding from a cropped image with $\mathbf{M}_j$ . Each mask $\mathbf{M}_j$ belongs to an image, and the masks are not overlapped to each other. With this dictionary, a CLIP embedding map $\mathbf{F}^{\mathrm{map}}(\mathbf{I}, \mathbf{r})$ at a pixel $\mathbf{r}$ in a training image $\mathbf{I}$ is computed as: + +$$ +\mathbf {F} ^ {\text {m a p}} (\mathbf {I}, \mathbf {r}) = \sum_ {j = 1} ^ {M} \mathbf {M} _ {j} (\mathbf {I}, \mathbf {r}) \cdot \mathbf {f} _ {j} ^ {\text {m a p}}, \tag {3} +$$ + +where $\mathbf{M}_j(\mathbf{I},\mathbf{r})\in \{0,1\}$ indicates whether the mask $\mathbf{M}_j$ contains the pixel $\mathbf{r}$ in the image $\mathbf{I}$ . Using $\mathbf{F}^{\mathrm{map}}$ , we reconstruct language embedded 3D Gaussians via a novel feature registration process as visualized in Fig. 3. + +During the feature registration process, our algorithm iterates through training images of the scene. Using projection relation, we link 3D Gaussians $\Phi$ to CLIP embeddings. Each Gaussian can link to multiple CLIP embeddings derived from different images. Then we aggregate collected embeddings to a single embedding to be assigned to each Gaussian. To ensure a consistent aggregation of the embeddings from multi-view images, we first compute a weight $w_{i}(\mathbf{I},\mathbf{r})$ representing the contribution of $\theta_{i}$ to construct each pixel $\mathbf{r}$ in a training image $\mathbf{I}$ . The weights are computed with the volume rendering equation Eq. (1) as: + +$$ +w _ {i} (\mathbf {I}, \mathbf {r}) = T _ {i} (\mathbf {I}, \mathbf {r}) \cdot \tilde {\alpha} _ {i} (\mathbf {I}, \mathbf {r}), \tag {4} +$$ + +where $T_{i}(\mathbf{I},\mathbf{r})$ and $\tilde{\alpha}_{i}(\mathbf{I},\mathbf{r})$ are the transmittance and the effective opacity value of $\theta_{i}$ for a pixel $\mathbf{r}$ in an image $\mathbf{I}$ , stated in Eq. (1). With the per-pixel weights, we calculate $w_{ij}$ representing a weight between each Gaussian $\theta_{i}$ and corresponding language embedding maps $\mathbf{f}_j^{\mathrm{map}}$ , which is for aggregating CLIP embeddings from $\mathbf{F}^{\mathrm{map}}$ and register the embedding to each Gaussian. The weights are computed as: + +$$ +w _ {i j} = \sum_ {\mathbf {I} \in \mathcal {I}} \sum_ {\mathbf {r} \in \mathbf {I}} \mathbf {M} _ {j} (\mathbf {I}, \mathbf {r}) \cdot w _ {i} (\mathbf {I}, \mathbf {r}), \tag {5} +$$ + +where $\mathcal{I}$ is the set of the training images. In this iterative process, we aggregate weights only for Top- $k$ Gaussians with the highest weights $w_{i}(\mathbf{I},\mathbf{r})$ , along the ray of each pixel ray $\mathbf{r}$ (see Fig. 4). After aggregation, we prune the Gaussians which are not assigned any weight, i.e., $\sum_{j=1}^{M} w_{ij} = 0$ . This summation aggregates weights between Gaussians and the CLIP embeddings by linking per-pixel weights $w_{i}(\mathbf{I},\mathbf{r})$ of each Gaussian to its corresponding CLIP embeddings. With the obtained weights, we register an aggregated feature $\dot{\mathbf{f}}_i$ to each Gaussian with weighted-averaging as: + +$$ +\dot {\mathbf {f}} _ {i} = \mathbf {f} _ {i} / | | \mathbf {f} _ {i} | | _ {2}, \text {w h e r e} \quad \mathbf {f} _ {i} = \sum_ {j = 1} ^ {M} \frac {w _ {i j}}{\sum_ {k = 1} ^ {M} w _ {i k}} \mathbf {f} _ {j} ^ {\text {m a p}}. \tag {6} +$$ + +This process enables 3D-aware feature registration to be consistent across various viewpoints, by aggregating features in the original high-dimensional feature space. The proposed process can be interpreted as an inverse volume rendering without gradient-based optimization, which enables our method to be faster than the prior methods requiring perscene gradient-based optimization [27, 30, 34] for feature registration in 3D space. + +# 3.2. Product-Quantized CLIP embeddings + +Memory efficiency is a challenge in 3D scene representations, especially when associating Gaussians with high-dimensional feature vectors. LangSplat [30] addresses this by introducing an encoder-decoder network, while LeGaussian [34] and OpenGaussian [37] utilize codebook construc + +
MethodsmIoUmAcc @ 0.25
waldo_kitchenramenfigurinesteatimeMeanwaldo_kitchenramenfigurinesteatimeMean
LangSplat-m [30]8.296.118.3316.589.8313.6414.088.9327.1215.94
OpenGaussian [37]34.6023.8759.3354.4443.0650.0035.2180.3672.8859.61
Ours (Top-10)37.0524.3354.4257.3543.2963.6435.2180.3677.9764.30
Ours (Top-20)38.3324.5853.9456.1943.2663.6435.2182.1476.2764.32
Ours (Top-40)39.0724.7053.3657.2043.5863.6435.2180.3676.2763.87
+ +Table 1. 3D object selection results on the LeRF-OVS dataset [17]. To measure 3D object selection performance, we calculate 2D segmentation accuracy on rendering of selected 3D Gaussians. Note that our model does not require per-scene optimization, demonstrating its robustness across diverse scenes. Bold and Underline stand for first and second best performance. + +![](images/42dd55066e5ac092d686ad46cf83089ae2f133b774de38c9a7f70912a0fdcd22.jpg) +Figure 5. Qualitative results of the object selection on the LeRF-OVS dataset [17]. We visualize rendering of selected 3D Gaussians for LangSplat [30], OpenGaussian [37], and ours. For LangSplat, activations are often distributed randomly, fail to localize the target. OpenGaussian often struggles to distinguish closely situated objects. In contrast, our model shows activations precisely limited to the queried object regions, effectively localizing only the relevant areas. + +![](images/5cfd6f2ed8ab85d8fd3a5e46d87974d73f21e5beebe0ef40a115c4d6b44692bc.jpg) + +![](images/23d2d3d1f429017f8d554d145f66b3475175ec479fc75eebbbfac5f2953c0d18.jpg) + +![](images/0245c63cd0c27778f78a287b354a76abb1e7514fe05af7fa14b11d54ebd7d9f3.jpg) + +![](images/7abf6229bb01b5699460772cb48c1322fec2e565a156a66d0108d37f81152783.jpg) + +![](images/1dd7bc29c4e690359fc8c42a213d96ef133a8b2c20cd2abd3d025e8c16a17924.jpg) + +tion. However, these approaches introduce additional per-scene computational costs for scene-specific parameter tuning of neural networks or codebooks (see Fig. 1). In contrast, we propose to use Product Quantization (PQ) on a large-scale image dataset, eliminating per-scene training. + +Product Quantization. PQ [13] is a widely used technique for efficient embedding compression, particularly valuable in large-scale applications. The PQ process begins by dividing the original $D$ -dimensional feature vector $\mathbf{v}$ into $L$ sub-vectors: $\mathbf{v} = [\mathbf{v}_1,\mathbf{v}_2,\dots ,\mathbf{v}_L]$ . Each sub-vector $\mathbf{v}_i$ is then independently quantized to a predefined number of centroids $s_{ij}$ in a predefined codebook $S_{i}$ for that sub-vector. These centroids are learned via clustering, creating a codebook for each subspace. Once the centroids are established, each sub-vector is replaced by the index of the nearest centroid in its respective codebook. The centroid indices $j_{i} = [j_{i1},j_{i2},\ldots ,j_{iL}]$ are optimized by minimizing $\arg \min_k\| \mathbf{v}_i - s_{ik}\|$ to quantize a given vector $\mathbf{v}_i$ where $j_{ik}$ is an 8-bit unsigned integer. + +Then, we can measure the distance between the query and data by adding distances between coarse centroids. Once the + +![](images/0e0078a6532ba08089c97e86c554134716d13da880e39ddeae2e1e7291cd23ce.jpg) + +![](images/18a40acb4c8461f487d030403c978efc5baa11ca1b4f93f66726976cdcd8f91a.jpg) +Figure 6. Limitations of point-based IoU measurement. This figure shows the effect of removing the top and bottom $30\%$ of Gaussians according to the proposed significant score, implying that volume differences significantly impact 3D accuracy. The results highlight the need for the proposed IoU metric for 3D Gaussians. + +![](images/01353414c5ff054c5b8723a00c26086fb62fc46477629e77e814574b061054fe.jpg) + +distances between centroids are computed as a lookup table, the computation shifts to simple indexing, which reduces the search complexity from $\mathcal{O}(D)$ to $\mathcal{O}(1)$ for a $D$ dimen + +
3D +mIoUIoU > 0.1519 classes +IoU > 0.3IoU > 0.45
LangSplat-m [30]8.017.17.82.9
LEGaussians-m [34]9.519.18.97.3
OpenGaussian [37]25.259.538.018.3
Ours (Top-20)25.060.740.320.0
Ours (Top-40)25.460.740.325.6
+ +(a) 3D object localization task. + +
19 classes15 classes10 classes
mIoUmAcc.mIoUmAcc.mIoUmAcc.
LangSplat-m [30]2.09.24.914.68.023.9
LEGaussians-m [34]1.67.94.616.17.724.9
OpenGaussian [37]30.146.538.156.849.771.4
Ours (Top-20)28.044.638.260.447.268.9
Ours (Top-40)29.647.738.260.450.273.5
+ +(b) Open-vocabulary 3D semantic segmentation task. + +Table 2. Quantitative comparison in the ScanNet dataset [4]. Left: Localization prediction is defined as 3D regions with a text similarity score above threshold. Right: We assign segmentation labels by finding max activations among all classes. Note that Bold and Underline stand for first and second best performance, respectively. + +![](images/e7360054e021d98a71a00ff79e35a322409c81ec51cf9f1ae80953595a74b077.jpg) +Figure 7. Qualitative results of 3D object localization. We visualize 3D localization activations (yellow) for "chair" and "desk" in the ScanNet dataset, comparing our method with others. It turns out that LangSplat-m and LEGaussians-m fail to localize objects accurately, while OpenGaussian struggles with object correspondence. Our model delivers precise and consistent localization across diverse queries. + +sion sample. This approach notably reduces computational complexity, making it suitable for large-scale search. + +In our setup for language-based 3D scene understanding, we build PQ centroids based on CLIP embeddings using a large-scale image dataset, the LVIS dataset [10], that contains over 1.2M instances covering various long-tail classes and ground truth segmentation. We extract instance patches from images and collect patch-wise CLIP embeddings. After we build this CLIP embedding database, we proceed with the construction of the centroid codebook for our PQ. Once PQ is trained, any query embedding can be approximated by assigning the closest centroid for each subvector. This is a one-time procedure; once we determine the codebook, we can use it for any scene generally. In our setup, each embedding is represented as a sequence of centroid indices rather than a high-dimensional vector. Accordingly, our language embedded Gaussians are parametrized as $\Phi^{\mathrm{ours}} = \{\phi_i^{\mathrm{ours}}\}_{i=1}^N = \{\theta_i, j_i\}_{i=1}^N$ . where the aggregated feature $\hat{\mathbf{f}}_i$ are converted as a quantized feature $\bar{\mathbf{f}}_i$ by the corresponding PQ index $j_i$ . + +# 3.3. Text-query based 3D localization + +After training 3D Gaussians $\Phi^{\mathrm{ours}}$ with our feature registration process and PQ, we describe the details of an inference mode that facilitates direct interaction with 3DGS upon receiving input queries, such as text. This is related to similarity score computation between a query and sources, i.e. Gaussian embeddings. Given a text, we first extract a + +query feature $\mathbf{q}$ using CLIP text encoder [31]. We reconstruct the quantized features $\{\bar{\mathbf{f}}_i\}_{i=1}^N$ from the stored PQ indices $\{j_i\}_{i=1}^N$ . Then, we compute a cosine similarity score between the query feature $\mathbf{q}$ and all quantized features. + +Despite its simplicity, solely relying on the cosine similarity may result in diminished discriminability across certain similarity scores. To address this limitation, we incorporate a re-ranking process based on relative activation with respect to the canonical feature. For this process, we adopt the relevancy scoring method proposed in LeRF [17], which enables more precise similarity analysis for a query. Specifically, each rendered language embedding, $\mathbf{f}^{\mathrm{map}}$ and a text query feature $\mathbf{q}$ , yield a relevance score determined by $\min_{i}\frac{\exp(\mathbf{f}^{\mathrm{map}}\cdot\mathbf{q})}{\exp(\mathbf{f}^{\mathrm{map}}\cdot\mathbf{q}) + \exp(\mathbf{f}^{\mathrm{map}}\cdot\mathbf{f}^{\mathrm{canon,i}})}$ , where $(\cdot)$ is an element-wise dot product operator and $\mathbf{f}^{\mathrm{canon,i}}$ indicates CLIP embeddings of a designated canonical term selected from a set of "object," "things," "stuff," and "texture". Then, we sample 3D Gaussians based on the relevance score for downstream tasks. + +# 4. Experiments + +Dataset. We use two datasets to evaluate the 3D scene understanding performance. For the 3D object selection task (Sec. 4.1), we use the LERF [17] dataset annotated by LangSplat [30], which consists of several multi-view images of 3D scenes containing long-tail objects and includes ground truth 2D ground truth annotations for texture queries. For 3D object localization Sec. 4.2 and 3D semantic segmentation Sec. 4.3 task, we employ the ScanNet [4] dataset. + +![](images/74bf00ac799dd062d56c084505c13cea6436e7b9a368cf80cb7eab240c38994b.jpg) +Figure 8. Visualization of open-vocabulary 3D semantic segmentation on the ScanNet dataset [4]. We visualize 3D Gaussian splat-based semantic segmentation using language features allocation of OpenGaussian [37] and Dr. Splat (ours) model on the same RGB-pretrained 3DGS. Note that, not specifically designed for segmentation, it achieves high performance as a result of language-based Gaussian updates. + +ScanNet is a large-scale benchmark that provides data on indoor scenes, including calibrated RGBD images and 3D point clouds with ground-truth semantic labels. We randomly select eight scenes from ScanNet for the experiments. + +Competing methods. The only method available for a fair comparison with our method is the concurrent work, Open-Gaussian [37]. To study the various aspects of our method, we introduce baseline methods modified from rasterization-based ones [30, 34], for direct 3D referring operation, denoted as LangSplat-m and LEGaussians-m. As discussed in Sec. 2, without modification, global search over a whole scene is quite demanding. To ensure fair evaluation, we use the same initial 3D Gaussians being trained only using RGB inputs for all comparing methods, and freeze the Gaussians during the language feature allocation process. Also, the per-pixel CLIP [31] embedding maps are unified for SAM-based [20] methods [30, 37] including ours. We follow the hyperparameter settings favorable to each respective paper. + +# 4.1.3D object selection + +Settings. We first extract text features from an open-vocabulary text query using the CLIP model. Next, we compare text features to the 3D features embedded in each Gaussian using cosine similarity. By thresholding the similarity, we identify the 3D Gaussians that are relevant to the given text query. The selected 3D points are subsequently rendered into multi-view images using the 3DGS rasterization pipeline. + +Results. We compare our model quantitatively with 3DGS-based language-embedded models as shown in Table 1. The results demonstrate that our method performs better object selection in most scenes, showing an improvement of over 0.5 in mIoU and more than 4.5 in mAcc compared to counterpart models. Notably, the rasterization-based method, LangSplat-m, often underperforms in most scenes. + +Qualitative results are shown in Fig. 5. For LangSplat-m, the activations often shows random 3D Gaussians or fail to localize entirely (e.g., see "coffee mug"), highlighting the limitations of rasterization-based methods and their un + +suitability for 3D understanding, aligning the observation from Fig. 2. OpenGaussian frequently exhibits false activations with incorrect text-object pairs (e.g., "apple" and "tea in a glass") and struggles to distinguish between nearby objects (e.g., "waldo," "rubik's cube"). This artifacts can be attributed to use of spatial clustering and limited encoder capacity. + +In contrast, our model leverages general image features thanks to the general PQ, maintaining feature distinctiveness regardless of scene complexity. Our feature registration considers the 3D geometry of the 3D Gaussians, which results in superior performance in 3D scene understanding tasks. + +# 4.2. 3D object localization + +Settings. Similar to the 3D object selection task, we calculate the cosine similarity between text query and 3D features embedded in each Gaussian. By thresholding the similarity, we identify the 3D Gaussians relevant to the given text query. To measure volume-aware localization evaluation, we propose a protocol to measure the IoU of 3D Gaussians that expands the traditional metric of point cloud-based approaches by incorporating volumetric information of 3D Gaussians. + +Novel evaluation protocol for 3D localization in 3DGS. Unlike conventional evaluation protocol for the 3D localization task in point clouds, it is tricky to evaluate 3D localization performance in 3D Gaussians [16]. This is primarily due to the un-deterministic structure of Gaussian distribution. To address this issue, we compute 3DGS pseudo-labels for evaluating the 3DGS localization in a volume-aware way. The details can be found in the supplementary material. + +Given the ground truth, we measure IoU considering the spatial significance of each Gaussian and define a significant score $d_{i}$ for each Gaussian $\theta_{i}$ with its scale $\mathbf{s}_i = [s_{ix}, s_{iy}, s_{iy}]$ and opacity $\alpha_{i}$ as $d_{i} = s_{ix}s_{iy}s_{iz}\alpha_{i}$ , where $s_{ix}s_{iy}s_{iz}$ denotes a relative ellipsoid volume of a Gaussian $\theta_{i}$ . With the obtained significant scores $\mathbf{d} = [d_1, d_2, \dots, d_N]$ , we compute weighted IoU of 3D Gaussians to approximate volumes. The proposed metric is designed to assign a larger + +weight to the Gaussians with higher significant scores, when measuring IoU. Figure 6 shows that the impact of each Gaussian on the scene extremely varies depending on their significant scores, which demonstrates the necessity of the proposed IoU metric on 3D Gaussians that regards unequal contributions of each Gaussian. + +Results. We report the 3D localization performance on the Scannet dataset in Table 2a. The 2D rasterization-based methods [30, 34] struggle to achieve precise activations for 3D localization. They inherently face challenges when applying for 3D tasks because they need to render 2D images for the scene interaction. Even with the 3D space search method, OpenGaussian [37], our model consistently demonstrates superior performance and achieves higher accuracy in localization. Figure 7 also shows that LangSplat-m and LEGaussians-m fail to properly localize the objects, and OpenGaussian misses queried objects in the scene. + +# 4.3. 3D semantic segmentation + +Settings. For a given set of open-vocabulary text labels, we perform segmentation by assigning each Gaussian a label having the highest activation among the known label set. + +Results. The numerical comparison is presented in Table 2b. Although not explicitly designed for semantic segmentation, our model achieves notable performance in this task as a result of accurately updating each Gaussian with language features. Consistent with previous observations, rasterization-based 3DGS models exhibit lower segmentation performance. While OpenGaussian performs position-based clustering, our model demonstrates comparable performance, surpassing the baseline as the Top- $k$ value increases. Our model also achieves better segmentation results, with a visual comparison of the segmented scene shown in Fig. 8. + +# 4.4. Ablation study + +We conduct an ablation study using the ScanNet dataset on different hyper-parameters of Dr. Splat to measure the contribution of each component. + +Product Quantization. PQ introduces a trade-off between memory usage, computational efficiency, and accuracy. To better understand the balance between computational cost and localization quality, we conduct an ablation study by varying the number of sub-vectors. We evaluate performance at sub-vector sizes of 64, 128, and 256. Notably, these settings correspond to bit-size reductions of 1/32, 1/16, and 1/8 of the original CLIP feature, respectively. We measure the query distance computation time for one million data points, averaging results over 100 iterations for efficiency measure. Our findings reveal a favorable trade-off between quantization performance and accuracy (see Fig. 9-(b)) in the Pareto front with our PQ configurations. This achieves a balance that maximizes memory and computational efficiency while minimizing any loss in accuracy. + +![](images/976c9798fa322047f804bddbd485f7e83bc57c08b9c0475f944ca46d55412b45.jpg) +(a) Ablation on PQ parameters. + +![](images/00be046b6c22477fe8a63532227a14e3f6fde73bd86141153e372d00f2e190a9.jpg) +(b) Ablation on Top- $k$ +Figure 9. Ablation study on (a) PQ and (b) Top- $k$ Gaussians. + +Top- $k$ Gaussians. We examine the influence of the number of Gaussians assigned per ray. This parameter affects both memory requirements and computation, serving as a critical factor in overall performance. The ratio of pruned Gaussians and the mIoU results from different $k$ are presented in Fig. 9a. We observe that increasing the aggregating number of Gaussians per ray improves localization performance; however, it results in higher memory consumption and the number of occupied Gaussians, indicating a clear trade-off. + +# 5. Discussion and Conclusion + +We present Dr. Splat, which is a novel approach for open-vocabulary 3D scene understanding by directly registering language embeddings to 3D Gaussians, eliminating the need for an intermediate rendering process. Compared to the previous 2D rendering-based methods [30, 34], which have limited search domain and capacity, our method directly searches 3D space while preserving the fidelity of language embeddings. This operation is further accelerated by the integration of Product Quantization (PQ) + +Experimental results validate Dr. Splat's superior performance across various 3D scene understanding tasks, including open-vocabulary 3D object selection, 3D object localization, and 3D semantic segmentation. These findings highlight Dr. Splat's ability to transform 3D scene understanding by achieving a balance between highly representative quality and computational efficiency. This breakthrough paves the way for advanced applications in robotics, autonomous navigation, and augmented reality. + +# Acknowledgment + +This work was supported by Institute of Information & communications Technology Planning & Evaluation (IITP) (No.RS-2021-II212068, Artificial Intelligence Innovation Hub, No.RS-2024-00457882, National AI Research Lab Project, No.RS-2023-00225630, Development of Artificial Intelligence for Text-based 3D Movie Generation, No.RS-2019-II191906, Artificial Intelligence Graduate School Program(POSTECH)) and by the National Research Foundation of Korea (NRF) (No.RS-2024-00358135, Corner Vision: Learning to Look Around the Corner through Multi-modal Signals) grant funded by the Korea government (MSIT). + +# References + +[1] Mathilde Caron, Hugo Touvron, Ishan Misra, Hervé Jégou, Julien Mairal, Piotr Bojanowski, and Armand Joulin. Emerging properties in self-supervised vision transformers. In Proceedings of the IEEE/CVF international conference on computer vision, pages 9650-9660, 2021. 2 +[2] Jaesung Choe, Chunghyun Park, Francois Rameau, Jaesik Park, and In So Kweon. Pointmixer: Mlp-mixer for point cloud understanding. In European Conference on Computer Vision, pages 620-640. Springer, 2022. 1 +[3] Christopher Choy, JunYoung Gwak, and Silvio Savarese. 4d spatio-temporal convnets: Minkowski convolutional neural networks. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 3075-3084, 2019. 1 +[4] Angela Dai, Angel X Chang, Manolis Savva, Maciej Halber, Thomas Funkhouser, and Matthias Nießner. Scannet: Richly-annotated 3d reconstructions of indoor scenes. In CVPR, pages 5828-5839, 2017. 6, 7 +[5] Runyu Ding, Jihan Yang, Chuhui Xue, Wenqing Zhang, Song Bai, and Xiaojuan Qi. Pla: Language-driven open-vocabulary 3d scene understanding. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 7010-7019, 2023. 2 +[6] Francis Engelmann, Fabian Manhardt, Michael Niemeyer, Keisuke Tateno, Marc Pollefeys, and Federico Tombari. OpenNerf: Open Set 3D Neural Scene Segmentation with PixelWise Features and Rendered Novel Views. In ICLR, 2024. 2 +[7] Michael R. Garey and David S. Johnson. Computers and intractability. a guide to the theory of np-completeness. W. H. Freeman and company, 174, 1979. 3 +[8] Antoine Guédon, Tom Monnier, Pascal Monasse, and Vincent Lepetit. Macarons: Mapping and coverage anticipation with rgb online self-supervision. In CVPR, 2023. 3 +[9] Jun Guo, Xiaojian Ma, Yue Fan, Huaping Liu, and Qing Li. Semantic gaussians: Open-vocabulary scene understanding with 3d gaussian splatting. arXiv preprint arXiv:2403.15624, 2024. 2 +[10] Agrim Gupta, Piotr Dollar, and Ross Girshick. Lvis: A dataset for large vocabulary instance segmentation. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 5356-5364, 2019. 6 +[11] Zhening Huang, Xiaoyang Wu, Xi Chen, Hengshuang Zhao, Lei Zhu, and Joan Lasenby. Openins3d: Snap and lookup for 3d open-vocabulary instance segmentation. In European Conference on Computer Vision, pages 169-185. Springer, 2025. 1, 3 +[12] Krishna Murthy Jatavallabhula, Alihusein Kuwajerwala, Qiao Gu, Mohd Omama, Tao Chen, Alaa Maalouf, Shuang Li, Ganesh Iyer, Soroush Saryazdi, Nikhil Keetha, et al. Conceptfusion: Open-set multimodal 3d mapping. arXiv preprint arXiv:2302.07241, 2023. 2 +[13] Herve Jegou, Matthijs Douze, and Cordelia Schmid. Product quantization for nearest neighbor search. IEEE transactions on pattern analysis and machine intelligence, 33(1):117-128, 2010. 5 + +[14] Yuzhou Ji, He Zhu, Junshu Tang, Wuyi Liu, Zhizhong Zhang, Yuan Xie, and Xin Tan. Fastlgs: Speeding up language embedded gaussians with feature grid mapping. arXiv preprint arXiv:2406.01916, 2024. 2, 3 +[15] Li Jiang, Shaoshuai Shi, and Bernt Schiele. Open-vocabulary 3d semantic segmentation with foundation models. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 21284-21294, 2024. 1, 2 +[16] Bernhard Kerbl, Georgios Kopanas, Thomas Leimkuhler, and George Drettakis. 3d gaussian splatting for real-time radiance field rendering. ACM TOG, 2023. 1, 2, 3, 7 +[17] Justin Kerr, Chung Min Kim, Ken Goldberg, Angjoo Kanazawa, and Matthew Tancik. Lerf: Language embedded radiance fields. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 19729-19739, 2023. 1, 2, 5, 6 +[18] GeonU Kim, Kim Youwang, and Tae-Hyun Oh. FPRF: Feedforward photorealistic style transfer of large-scale 3D neural radiance fields. In AAAI, 2024. 2 +[19] GeonU Kim, Kim Youwang, Lee Hyoseok, and Tae-Hyun Oh. Fpgs: Feed-forward semantic-aware photorealistic style transfer of large-scale gaussian splatting, 2025. 3 +[20] Alexander Kirillov, Eric Mintun, Nikhila Ravi, Hanzi Mao, Chloe Rolland, Laura Gustafson, Tete Xiao, Spencer Whitehead, Alexander C. Berg, Wan-Yen Lo, Piotr Dólar, and Ross B. Girshick. Segment anything. In ICCV, 2023. 4, 7 +[21] Sosuke Kobayashi, Eiichi Matsumoto, and Vincent Sitzmann. Decomposing nerf for editing via feature field distillation. Advances in Neural Information Processing Systems, 35:23311-23330, 2022. 2 +[22] Hyunjee Lee, Youngsik Yun, Jeongmin Bae, Seoha Kim, and Youngjung Uh. Rethinking open-vocabulary segmentation of radiance fields in 3d space, 2024. 2 +[23] Boyi Li, Kilian Q Weinberger, Serge Belongie, Vladlen Koltun, and René Ranftl. Language-driven semantic segmentation. arXiv preprint arXiv:2201.03546, 2022. 2 +[24] Kunhao Liu, Fangneng Zhan, Jiahui Zhang, Muyu Xu, Yingchen Yu, Abdulmotaleb El Saddik, Christian Theobalt, Eric Xing, and Shijian Lu. Weakly supervised 3d open-vocabulary segmentation. Advances in Neural Information Processing Systems, 36:53433-53456, 2023. 2 +[25] Minghua Liu, Yinhao Zhu, Hong Cai, Shizhong Han, Zhan Ling, Fatih Porikli, and Hao Su. Partslip: Low-shot part segmentation for 3d point clouds via pretrained image-language models. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 21736-21746, 2023. 2 +[26] Ben Mildenhall, Pratul P. Srinivasan, Matthew Tancik, Jonathan T. Barron, Ravi Ramamoorthi, and Ren Ng. Nerf: Representing scenes as neural radiance fields for view synthesis. In ECCV, 2020. 2 +[27] Songyou Peng, Kyle Genova, Chiyu Jiang, Andrea Tagliasacchi, Marc Pollefeys, Thomas Funkhouser, et al. Openscene: 3d scene understanding with open vocabularies. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 815-824, 2023. 1, 2, 4 +[28] Charles R Qi, Hao Su, Kaichun Mo, and Leonidas J Guibas. Pointnet: Deep learning on point sets for 3d classification + +and segmentation. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 652-660, 2017. 1 +[29] Guocheng Qian, Yuchen Li, Houwen Peng, Jinjie Mai, Hasan Hammoud, Mohamed Elhoseiny, and Bernard Ghanem. Pointnext: Revisiting pointnet++ with improved training and scaling strategies. Advances in neural information processing systems, 35:23192-23204, 2022. 1 +[30] Minghan Qin, Wanhua Li, Jiawei Zhou, Haoqian Wang, and Hanspeter Pfister. Langsplat: 3d language gaussian splatting. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 20051-20060, 2024. 1, 2, 3, 4, 5, 6, 7, 8 +[31] Alec Radford, Jong Wook Kim, Chris Hallacy, Aditya Ramesh, Gabriel Goh, Sandhini Agarwal, Girish Sastry, Amanda Askell, Pamela Mishkin, Jack Clark, et al. Learning transferable visual models from natural language supervision. In International conference on machine learning, pages 8748-8763. PMLR, 2021. 2, 3, 6, 7 +[32] Adam Rashid, Satvik Sharma, Chung Min Kim, Justin Kerr, Lawrence Yunliang Chen, Angjoo Kanazawa, and Ken Goldberg. Language embedded radiance fields for zero-shot task-oriented grasping. In 7th Annual Conference on Robot Learning, 2023. 2 +[33] Damien Robert, Hugo Raguet, and Loic Landrieu. Efficient 3d semantic segmentation with superpoint transformer. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 17195-17204, 2023. 1 +[34] Jin-Chuan Shi, Miao Wang, Hao-Bin Duan, and Shao-Hua Guan. Language embedded 3d gaussians for open-vocabulary scene understanding. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 5333-5343, 2024. 1, 2, 3, 4, 6, 7, 8 +[35] Ayca Takmaz, Elisabetta Fedele, Robert W Sumner, Marc Pollefeys, Federico Tombari, and Francis Engelmann. Openmask3d: Open-vocabulary 3d instance segmentation. arXiv preprint arXiv:2306.13631, 2023. 1 +[36] Xiaoyang Wu, Li Jiang, Peng-Shuai Wang, Zhijian Liu, Xihui Liu, Yu Qiao, Wanli Ouyang, Tong He, and Hengshuang Zhao. Point transformer v3: Simpler faster stronger. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 4840-4851, 2024. 1 +[37] Yanmin Wu, Jiarui Meng, Hajjie Li, Chenming Wu, Yahao Shi, Xinhua Cheng, Chen Zhao, Haocheng Feng, Errui Ding, Jingdong Wang, et al. Opengaussian: Towards point-level 3d gaussian-based open vocabulary understanding. arXiv preprint arXiv:2406.02058, 2024. 1, 2, 3, 4, 5, 6, 7, 8 +[38] Jihan Yang, Runyu Ding, Weipeng Deng, Zhe Wang, and Xiaojuan Qi. Regionplc: Regional point-language contrastive learning for open-world 3d scene understanding. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 19823-19832, 2024. 1, 2 +[39] Junbo Zhang, Runpei Dong, and Kaisheng Ma. Clip-for3d: Learning free open-world 3d scene representations from 2d dense clip. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 2048-2059, 2023. 2 +[40] Hengshuang Zhao, Li Jiang, Jiaya Jia, Philip HS Torr, and Vladlen Koltun. Point transformer. In Proceedings of + +the IEEE/CVF international conference on computer vision, pages 16259-16268, 2021. 1 +[41] Yuhang Zheng, Xiangyu Chen, Yupeng Zheng, Songen Gu, Runyi Yang, Bu Jin, Pengfei Li, Chengliang Zhong, Zeng-mao Wang, Lina Liu, Chao Yang, Dawei Wang, Zhen Chen, Xiaoxiao Long, and Meiqing Wang. Gaussiangrasper: 3d language gaussian splatting for open-vocabulary robotic grasping. IEEE Robotics and Automation Letters, 2024. 1, 2, 3 +[42] Shijie Zhou, Haoran Chang, Sicheng Jiang, Zhiwen Fan, Zehao Zhu, Dejia Xu, Pradyumna Chari, Suya You, Zhangyang Wang, and Achuta Kadambi. Feature 3dgs: Supercharging 3d gaussian splattering to enable distilled feature fields. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 21676-21685, 2024. +[43] Xingxing Zuo, Pouya Samangouei, Yunwen Zhou, Yan Di, and Mingyang Li. Fmgs: Foundation model embedded 3d gaussian splatting for holistic 3d scene understanding. IJCV, 2024. 1, 2, 3 +[44] Matthias Zwicker, Hanspeter Pfister, Jeroen van Baar, and Markus H. Gross. EWA volume splatting. In Visualization Conference, 2001. 2 \ No newline at end of file diff --git a/2025/Dr. Splat_ Directly Referring 3D Gaussian Splatting via Direct Language Embedding Registration/images.zip b/2025/Dr. Splat_ Directly Referring 3D Gaussian Splatting via Direct Language Embedding Registration/images.zip new file mode 100644 index 0000000000000000000000000000000000000000..707371c123d2947f112247f2e114290b88c2184a --- /dev/null +++ b/2025/Dr. Splat_ Directly Referring 3D Gaussian Splatting via Direct Language Embedding Registration/images.zip @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:385b3c3234f7cf6046eac9ca461d803105a4f6353b8f9f6d8708706216788f88 +size 654864 diff --git a/2025/Dr. Splat_ Directly Referring 3D Gaussian Splatting via Direct Language Embedding Registration/layout.json b/2025/Dr. Splat_ Directly Referring 3D Gaussian Splatting via Direct Language Embedding Registration/layout.json new file mode 100644 index 0000000000000000000000000000000000000000..7cdbcda9aa065e08bc8dc9e0ef3b1d0bf54ac067 --- /dev/null +++ b/2025/Dr. Splat_ Directly Referring 3D Gaussian Splatting via Direct Language Embedding Registration/layout.json @@ -0,0 +1,10009 @@ +{ + "pdf_info": [ + { + "para_blocks": [ + { + "bbox": [ + 145, + 103, + 465, + 140 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 145, + 103, + 465, + 140 + ], + "spans": [ + { + "bbox": [ + 145, + 103, + 465, + 140 + ], + "type": "text", + "content": "Dr. Splat: Directly Referring 3D Gaussian Splatting via Direct Language Embedding Registration" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 106, + 160, + 187, + 175 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 106, + 160, + 187, + 175 + ], + "spans": [ + { + "bbox": [ + 106, + 160, + 187, + 175 + ], + "type": "text", + "content": "Kim Jun-Seong1" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 211, + 161, + 278, + 175 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 211, + 161, + 278, + 175 + ], + "spans": [ + { + "bbox": [ + 211, + 161, + 278, + 175 + ], + "type": "text", + "content": "GeonU Kim" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 303, + 161, + 359, + 175 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 303, + 161, + 359, + 175 + ], + "spans": [ + { + "bbox": [ + 303, + 161, + 359, + 175 + ], + "type": "text", + "content": "Kim Yu-Ji2" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 384, + 161, + 504, + 175 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 384, + 161, + 504, + 175 + ], + "spans": [ + { + "bbox": [ + 384, + 161, + 504, + 175 + ], + "type": "text", + "content": "Yu-Chiang Frank Wang3" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 211, + 175, + 288, + 189 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 211, + 175, + 288, + 189 + ], + "spans": [ + { + "bbox": [ + 211, + 175, + 288, + 189 + ], + "type": "text", + "content": "Jaesung Choe" + }, + { + "bbox": [ + 211, + 175, + 288, + 189 + ], + "type": "inline_equation", + "content": "^{3*}" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 309, + 175, + 398, + 189 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 309, + 175, + 398, + 189 + ], + "spans": [ + { + "bbox": [ + 309, + 175, + 398, + 189 + ], + "type": "text", + "content": "Tae-Hyun Oh" + }, + { + "bbox": [ + 309, + 175, + 398, + 189 + ], + "type": "inline_equation", + "content": "^{1,2,4*}" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 113, + 189, + 498, + 203 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 113, + 189, + 498, + 203 + ], + "spans": [ + { + "bbox": [ + 113, + 189, + 498, + 203 + ], + "type": "inline_equation", + "content": "^{1}" + }, + { + "bbox": [ + 113, + 189, + 498, + 203 + ], + "type": "text", + "content": "Department of Electrical Engineering and " + }, + { + "bbox": [ + 113, + 189, + 498, + 203 + ], + "type": "inline_equation", + "content": "^{2}" + }, + { + "bbox": [ + 113, + 189, + 498, + 203 + ], + "type": "text", + "content": "Graduate School of AI, POSTECH" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 192, + 203, + 241, + 216 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 192, + 203, + 241, + 216 + ], + "spans": [ + { + "bbox": [ + 192, + 203, + 241, + 216 + ], + "type": "text", + "content": "3NVIDIA" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 267, + 204, + 417, + 217 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 267, + 204, + 417, + 217 + ], + "spans": [ + { + "bbox": [ + 267, + 204, + 417, + 217 + ], + "type": "inline_equation", + "content": "^{4}" + }, + { + "bbox": [ + 267, + 204, + 417, + 217 + ], + "type": "text", + "content": "School of Computing, KAIST" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 192, + 220, + 417, + 231 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 192, + 220, + 417, + 231 + ], + "spans": [ + { + "bbox": [ + 192, + 220, + 417, + 231 + ], + "type": "text", + "content": "{junseong.kim, gukim, ugkim}@postech.ac.kr" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 165, + 234, + 446, + 244 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 165, + 234, + 446, + 244 + ], + "spans": [ + { + "bbox": [ + 165, + 234, + 446, + 244 + ], + "type": "text", + "content": "{frankwang, jchoe}@nvidia.com taehyun@kaist.ac.kr" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 151, + 258, + 199, + 270 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 151, + 258, + 199, + 270 + ], + "spans": [ + { + "bbox": [ + 151, + 258, + 199, + 270 + ], + "type": "text", + "content": "Abstract" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 55, + 283, + 296, + 486 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 283, + 296, + 486 + ], + "spans": [ + { + "bbox": [ + 55, + 283, + 296, + 486 + ], + "type": "text", + "content": "We introduce Dr. Splat, a novel approach for open-vocabulary 3D scene understanding leveraging 3D Gaussian Splatting. Unlike existing language-embedded 3DGS methods, which rely on a rendering process, our method directly associates language-aligned CLIP embeddings with 3D Gaussians for holistic 3D scene understanding. The key of our method is a language feature registration technique where CLIP embeddings are assigned to the dominant Gaussians intersected by each pixel-ray. Moreover, we integrate Product Quantization (PQ) trained on general large-scale image data to compactly represent embeddings without per-scene optimization. Experiments demonstrate that our approach significantly outperforms existing approaches in 3D perception benchmarks, such as open-vocabulary 3D semantic segmentation, 3D object localization, and 3D object selection tasks. For video results, please visit: https://drsplat.github.io/" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 55, + 509, + 135, + 521 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 509, + 135, + 521 + ], + "spans": [ + { + "bbox": [ + 55, + 509, + 135, + 521 + ], + "type": "text", + "content": "1. Introduction" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 55, + 529, + 296, + 696 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 529, + 296, + 696 + ], + "spans": [ + { + "bbox": [ + 55, + 529, + 296, + 696 + ], + "type": "text", + "content": "Open-vocabulary 3D scene understanding represents a significant challenge in the field of computer vision, with applications spanning autonomous navigation, robotics, and augmented reality. This approach aims to enable the interpretation and referencing of 3D spatial information through natural language, allowing for applicability beyond a restricted set of predefined categories [2, 3, 28, 29, 33, 36, 40]. Previously, open-vocabulary 3D scene understanding has been explored using point-cloud-based methods [11, 15, 17, 27, 30, 35, 38]. Recently, the 3D Gaussian Splitting (3DGS) [16] has introduced a continuous representation integrated on explicit 3D Gaussians, which differs from traditional point-cloud approaches, enabling rapid progress in practical applications [41]. Current research has begun to explore" + } + ] + } + ], + "index": 17 + }, + { + "type": "image", + "bbox": [ + 315, + 258, + 432, + 350 + ], + "blocks": [ + { + "bbox": [ + 315, + 258, + 432, + 350 + ], + "lines": [ + { + "bbox": [ + 315, + 258, + 432, + 350 + ], + "spans": [ + { + "bbox": [ + 315, + 258, + 432, + 350 + ], + "type": "image", + "image_path": "723f26eaf0279c7ee09a4c6a34fe5122c8fe1a6bf384a89d446ad64dff1aa6dc.jpg" + } + ] + } + ], + "index": 18, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 345, + 350, + 405, + 358 + ], + "lines": [ + { + "bbox": [ + 345, + 350, + 405, + 358 + ], + "spans": [ + { + "bbox": [ + 345, + 350, + 405, + 358 + ], + "type": "text", + "content": "Render 2D images" + } + ] + } + ], + "index": 19, + "angle": 0, + "type": "image_caption" + } + ], + "index": 18 + }, + { + "type": "image", + "bbox": [ + 316, + 358, + 432, + 401 + ], + "blocks": [ + { + "bbox": [ + 316, + 358, + 432, + 401 + ], + "lines": [ + { + "bbox": [ + 316, + 358, + 432, + 401 + ], + "spans": [ + { + "bbox": [ + 316, + 358, + 432, + 401 + ], + "type": "image", + "image_path": "623dc797a4b1fa4dcb95bf70256b6dbedfe49900678e1c6327cee3ea495a5102.jpg" + } + ] + } + ], + "index": 20, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 354, + 402, + 398, + 408 + ], + "lines": [ + { + "bbox": [ + 354, + 402, + 398, + 408 + ], + "spans": [ + { + "bbox": [ + 354, + 402, + 398, + 408 + ], + "type": "text", + "content": "B number of images" + } + ] + } + ], + "index": 21, + "angle": 0, + "type": "image_caption" + } + ], + "index": 20 + }, + { + "type": "image", + "bbox": [ + 439, + 270, + 552, + 352 + ], + "blocks": [ + { + "bbox": [ + 447, + 258, + 545, + 270 + ], + "lines": [ + { + "bbox": [ + 447, + 258, + 545, + 270 + ], + "spans": [ + { + "bbox": [ + 447, + 258, + 545, + 270 + ], + "type": "text", + "content": "3D search scheme (ours)" + } + ] + } + ], + "index": 22, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 439, + 270, + 552, + 352 + ], + "lines": [ + { + "bbox": [ + 439, + 270, + 552, + 352 + ], + "spans": [ + { + "bbox": [ + 439, + 270, + 552, + 352 + ], + "type": "image", + "image_path": "7cf17d34a06bf835f2ec62c6165e1ed103649f856fa6babc777fcb4cd5454c9e.jpg" + } + ] + } + ], + "index": 23, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 468, + 351, + 525, + 358 + ], + "lines": [ + { + "bbox": [ + 468, + 351, + 525, + 358 + ], + "spans": [ + { + "bbox": [ + 468, + 351, + 525, + 358 + ], + "type": "text", + "content": "'Direct' 3D Search" + } + ] + } + ], + "index": 24, + "angle": 0, + "type": "image_caption" + } + ], + "index": 23 + }, + { + "type": "image", + "bbox": [ + 465, + 358, + 527, + 375 + ], + "blocks": [ + { + "bbox": [ + 465, + 358, + 527, + 375 + ], + "lines": [ + { + "bbox": [ + 465, + 358, + 527, + 375 + ], + "spans": [ + { + "bbox": [ + 465, + 358, + 527, + 375 + ], + "type": "image", + "image_path": "8f7765ced90b2d82f0f6133bc1fd5eab763b1a5ef1795b1d37fe67a51e9672ad.jpg" + } + ] + } + ], + "index": 25, + "angle": 0, + "type": "image_body" + } + ], + "index": 25 + }, + { + "type": "image", + "bbox": [ + 477, + 375, + 515, + 406 + ], + "blocks": [ + { + "bbox": [ + 477, + 375, + 515, + 406 + ], + "lines": [ + { + "bbox": [ + 477, + 375, + 515, + 406 + ], + "spans": [ + { + "bbox": [ + 477, + 375, + 515, + 406 + ], + "type": "image", + "image_path": "da76d519fcbeba25815c37e1aea704ca110e3d069c2533019f8af079e618a080.jpg" + } + ] + } + ], + "index": 26, + "angle": 0, + "type": "image_body" + } + ], + "index": 26 + }, + { + "type": "table", + "bbox": [ + 317, + 414, + 553, + 466 + ], + "blocks": [ + { + "bbox": [ + 317, + 414, + 553, + 466 + ], + "lines": [ + { + "bbox": [ + 317, + 414, + 553, + 466 + ], + "spans": [ + { + "bbox": [ + 317, + 414, + 553, + 466 + ], + "type": "table", + "html": "
Search domainper-scene codebookFeature distill.SearchDB size
LERF [17]2Drequired~ 24hslowlarge
LangSplat [30]2Drequired~ 4hslowlarge
LEGaussians [34]2Drequired~ 4hslowlarge
OpenGaussian [37]3Drequired~ 1hfastsmall
Dr. Splat (Ours)3Dnone~ 10mfastsmall
", + "image_path": "f90ce695a0150e2c33afbcbeebc1196415a9955e108061843afad0635b72368c.jpg" + } + ] + } + ], + "index": 27, + "angle": 0, + "type": "table_body" + } + ], + "index": 27 + }, + { + "bbox": [ + 313, + 470, + 555, + 536 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 470, + 555, + 536 + ], + "spans": [ + { + "bbox": [ + 313, + 470, + 555, + 536 + ], + "type": "text", + "content": "Figure 1. Comparison of 2D (left) vs. our direct 3D search (right) for open-vocabulary 3D scene understanding. The 2D approach relies on multiview rendering, incurring high computational costs. Our method directly links language features to 3D Gaussians, enabling efficient and complete spatial coverage. The table highlights Dr. Splat's superior efficiency over related methods." + } + ] + } + ], + "index": 28 + }, + { + "bbox": [ + 313, + 546, + 555, + 570 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 546, + 555, + 570 + ], + "spans": [ + { + "bbox": [ + 313, + 546, + 555, + 570 + ], + "type": "text", + "content": "methods for associating language-based features with 3D Gaussian splats to enhance scene understanding capabilities." + } + ] + } + ], + "index": 29 + }, + { + "bbox": [ + 313, + 570, + 555, + 665 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 570, + 555, + 665 + ], + "spans": [ + { + "bbox": [ + 313, + 570, + 555, + 665 + ], + "type": "text", + "content": "Several recent approaches [30, 34, 43] introduce 3D Gaussian representation [16] into the open-vocabulary scene understanding. This unique representation uses 3D Gaussians to achieve high-quality scene rendering, offering a more structured representation that addresses some limitations of point clouds. Building on this, these methods employ 2D vision-language models to transfer language knowledge to 3D Gaussians \"via rendered feature maps\"." + } + ] + } + ], + "index": 30 + }, + { + "bbox": [ + 313, + 665, + 556, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 665, + 556, + 714 + ], + "spans": [ + { + "bbox": [ + 313, + 665, + 556, + 714 + ], + "type": "text", + "content": "Despite its promise, such rendering-based distillation methods [30, 34] share two limitations. First, we found that there is a discrepancy between optimized embeddings in 3D Gaussians and 2D language-aligned embeddings. This gap" + } + ] + } + ], + "index": 31 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "spans": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "type": "text", + "content": "CVF" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 145, + 0, + 496, + 37 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 145, + 0, + 496, + 37 + ], + "spans": [ + { + "bbox": [ + 145, + 0, + 496, + 37 + ], + "type": "text", + "content": "This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 66, + 702, + 148, + 712 + ], + "type": "page_footnote", + "angle": 0, + "lines": [ + { + "bbox": [ + 66, + 702, + 148, + 712 + ], + "spans": [ + { + "bbox": [ + 66, + 702, + 148, + 712 + ], + "type": "text", + "content": "*Corresponding Authors" + } + ] + } + ], + "index": 32 + }, + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "text", + "content": "14137" + } + ] + } + ], + "index": 33 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 0 + }, + { + "para_blocks": [ + { + "bbox": [ + 55, + 72, + 296, + 156 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 72, + 296, + 156 + ], + "spans": [ + { + "bbox": [ + 55, + 72, + 296, + 156 + ], + "type": "text", + "content": "arises mainly from an intermediate rendering step that may distort CLIP embeddings during training. Then, the reliance on rendering impedes holistic 3D scene understanding, additional task-processing such as 3D semantic segmentation and 3D object localization, and making full spatial coverage calculations less efficient than direct 3D Gaussian methods [37] including ours as illustrated in Fig. 1." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 56, + 157, + 298, + 418 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 157, + 298, + 418 + ], + "spans": [ + { + "bbox": [ + 56, + 157, + 298, + 418 + ], + "type": "text", + "content": "To address this issue, this work proposes Dr. Splat. Our method bypasses the rendering stage, enabling direct interaction with 3D Gaussians for registering and referring the well-preserved language-aligned CLIP embeddings in the 3D space. This makes our Dr. Splat clearly distinguishable from prior works, facilitating a seamless integration of representative embeddings from 2D vision language models into the 3D spatial structure without compromising exhaustive rendering process that has been exploited [14, 30, 34, 41-43]. Moreover, we propose to use a Product Quantization (PQ) feature encoding method to represent embeddings compactly and efficiently without any per-scene optimization. Rather than storing full-length feature vectors or per-scene specifically compressed embeddings [14, 30, 34, 41-43], each Gaussian in our Dr. Splat stores an index from a pre-trained PQ, significantly reducing memory usage up to " + }, + { + "bbox": [ + 56, + 157, + 298, + 418 + ], + "type": "inline_equation", + "content": "6.25\\%" + }, + { + "bbox": [ + 56, + 157, + 298, + 418 + ], + "type": "text", + "content": " compression ratio. By preserving the richness of embeddings while reducing memory usage, PQ is integral to our framework's high scalability and its ability to perform 3D perception tasks, such as open-vocabulary 3D object localization, 3D object selection, and 3D semantic segmentation. Our contributions are summarized as follows:" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 55, + 422, + 296, + 536 + ], + "type": "list", + "angle": 0, + "index": 5, + "blocks": [ + { + "bbox": [ + 55, + 422, + 295, + 458 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 422, + 295, + 458 + ], + "spans": [ + { + "bbox": [ + 55, + 422, + 295, + 458 + ], + "type": "text", + "content": "- We propose Dr. Splat, direct registration and referencing of language-aligned features in 3D Gaussians, bypassing intermediate rendering and preserving feature accuracy." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 55, + 461, + 295, + 497 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 461, + 295, + 497 + ], + "spans": [ + { + "bbox": [ + 55, + 461, + 295, + 497 + ], + "type": "text", + "content": "- We introduce the PQ encoding method for compact feature representation, reducing memory usage while maintaining essential 3D feature properties." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 55, + 499, + 296, + 536 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 499, + 296, + 536 + ], + "spans": [ + { + "bbox": [ + 55, + 499, + 296, + 536 + ], + "type": "text", + "content": "- We present a novel evaluation protocol to assess accuracy of 3D localization and segmentation for 3D Gaussians, with pseudo-labeling methods and volume-aware metrics." + } + ] + } + ], + "index": 4 + } + ], + "sub_type": "text" + }, + { + "bbox": [ + 55, + 546, + 224, + 559 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 546, + 224, + 559 + ], + "spans": [ + { + "bbox": [ + 55, + 546, + 224, + 559 + ], + "type": "text", + "content": "2. Related Work and Motivation" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 55, + 570, + 296, + 665 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 570, + 296, + 665 + ], + "spans": [ + { + "bbox": [ + 55, + 570, + 296, + 665 + ], + "type": "text", + "content": "Language-based 3D scene understanding. Open-set 3D scene understanding has seen considerable advancements, with a focus on methods that leverage language knowledge into 3D representation such as point clouds, neural radiance fields (NeRF) [26], and Gaussian Splatting [16] for 3D comprehension. Point-based methods [5, 12, 15, 25, 27, 38, 39] in open-vocabulary settings process point cloud data trained from language embeddings [23, 31] for open-set categories." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 55, + 665, + 296, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 665, + 296, + 714 + ], + "spans": [ + { + "bbox": [ + 55, + 665, + 296, + 714 + ], + "type": "text", + "content": "NeRF-based approaches [6, 17, 18, 21, 24, 32] leverage semantic embeddings from 2D foundation models, such as CLIP [31], LSeg [23] and DINO [1] for open-vocabulary understanding. While the rendering process enhances 2D" + } + ] + } + ], + "index": 8 + }, + { + "type": "image", + "bbox": [ + 316, + 72, + 391, + 137 + ], + "blocks": [ + { + "bbox": [ + 316, + 72, + 391, + 137 + ], + "lines": [ + { + "bbox": [ + 316, + 72, + 391, + 137 + ], + "spans": [ + { + "bbox": [ + 316, + 72, + 391, + 137 + ], + "type": "image", + "image_path": "2fdf2de7718c75f0ce02338691c88a40088d2c4d6c2ecbccebfa85da3bf4c2b5.jpg" + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 332, + 137, + 374, + 146 + ], + "lines": [ + { + "bbox": [ + 332, + 137, + 374, + 146 + ], + "spans": [ + { + "bbox": [ + 332, + 137, + 374, + 146 + ], + "type": "text", + "content": "Input Image" + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 313, + 155, + 555, + 199 + ], + "lines": [ + { + "bbox": [ + 313, + 155, + 555, + 199 + ], + "spans": [ + { + "bbox": [ + 313, + 155, + 555, + 199 + ], + "type": "text", + "content": "Figure 2. Visualization of discrepancy in rendered 2D features and 3D features. Color indicates a cosine similarity score between query features from a text query and either (a) 3D features distilled by 2D rendering [30] or (b) directly registered 3D features." + } + ] + } + ], + "index": 15, + "angle": 0, + "type": "image_caption" + } + ], + "index": 9 + }, + { + "type": "image", + "bbox": [ + 391, + 72, + 464, + 136 + ], + "blocks": [ + { + "bbox": [ + 391, + 72, + 464, + 136 + ], + "lines": [ + { + "bbox": [ + 391, + 72, + 464, + 136 + ], + "spans": [ + { + "bbox": [ + 391, + 72, + 464, + 136 + ], + "type": "image", + "image_path": "f0afa253f25e50670761bad93d295c56c96e5835de9bf78550bebad21f8d2a3f.jpg" + } + ] + } + ], + "index": 11, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 404, + 137, + 451, + 146 + ], + "lines": [ + { + "bbox": [ + 404, + 137, + 451, + 146 + ], + "spans": [ + { + "bbox": [ + 404, + 137, + 451, + 146 + ], + "type": "text", + "content": "(a) Rendering" + } + ] + } + ], + "index": 12, + "angle": 0, + "type": "image_caption" + } + ], + "index": 11 + }, + { + "type": "image", + "bbox": [ + 465, + 80, + 539, + 136 + ], + "blocks": [ + { + "bbox": [ + 465, + 80, + 539, + 136 + ], + "lines": [ + { + "bbox": [ + 465, + 80, + 539, + 136 + ], + "spans": [ + { + "bbox": [ + 465, + 80, + 539, + 136 + ], + "type": "image", + "image_path": "a727fe67c7febb26a65755ae330ab24c80f1d7a693cd511d07889dcf75e8a6ab.jpg" + } + ] + } + ], + "index": 13, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 475, + 137, + 528, + 146 + ], + "lines": [ + { + "bbox": [ + 475, + 137, + 528, + 146 + ], + "spans": [ + { + "bbox": [ + 475, + 137, + 528, + 146 + ], + "type": "text", + "content": "(b) Registration" + } + ] + } + ], + "index": 14, + "angle": 0, + "type": "image_caption" + } + ], + "index": 13 + }, + { + "bbox": [ + 313, + 210, + 555, + 247 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 210, + 555, + 247 + ], + "spans": [ + { + "bbox": [ + 313, + 210, + 555, + 247 + ], + "type": "text", + "content": "perception tasks, the implicit nature of NeRF constrains the holistic understanding of 3D structures and dominantly provides 'rendered' feature maps." + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 313, + 247, + 555, + 319 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 247, + 555, + 319 + ], + "spans": [ + { + "bbox": [ + 313, + 247, + 555, + 319 + ], + "type": "text", + "content": "3D Gaussian Splatting (3DGS) [16] has emerged as a promising rendering method, as well as a novel representation for open-vocabulary 3D scene understanding. Since this research is the close related work with our work, we first elucidate the preliminary of 3DGS, followed by focusing on language embedded 3DGS as follows." + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 313, + 322, + 556, + 467 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 322, + 556, + 467 + ], + "spans": [ + { + "bbox": [ + 313, + 322, + 556, + 467 + ], + "type": "text", + "content": "Preliminary of 3D Gaussian Splatting. 3DGS [16] encodes appearance and geometry of the target scene into the 3D Gaussian representation. Each 3D primitive representation is expressed as a 3D Gaussian distribution having mean " + }, + { + "bbox": [ + 313, + 322, + 556, + 467 + ], + "type": "inline_equation", + "content": "\\pmb{\\mu} = [x_{\\mu},y_{\\mu},z_{\\mu}]^{\\top}" + }, + { + "bbox": [ + 313, + 322, + 556, + 467 + ], + "type": "text", + "content": " for 3D position and covariance matrix " + }, + { + "bbox": [ + 313, + 322, + 556, + 467 + ], + "type": "inline_equation", + "content": "\\Sigma_{\\mathrm{3D}}\\in \\mathbb{R}^{3\\times 3}" + }, + { + "bbox": [ + 313, + 322, + 556, + 467 + ], + "type": "text", + "content": " for 3D volume, as well as the opacity value " + }, + { + "bbox": [ + 313, + 322, + 556, + 467 + ], + "type": "inline_equation", + "content": "\\alpha" + }, + { + "bbox": [ + 313, + 322, + 556, + 467 + ], + "type": "text", + "content": " and the color c. In particular, the covariance matrix is decomposed into the scale matrix " + }, + { + "bbox": [ + 313, + 322, + 556, + 467 + ], + "type": "inline_equation", + "content": "S\\in \\mathbb{R}^{3\\times 3}" + }, + { + "bbox": [ + 313, + 322, + 556, + 467 + ], + "type": "text", + "content": " and the rotation matrix " + }, + { + "bbox": [ + 313, + 322, + 556, + 467 + ], + "type": "inline_equation", + "content": "R\\in SO(3)" + }, + { + "bbox": [ + 313, + 322, + 556, + 467 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 313, + 322, + 556, + 467 + ], + "type": "inline_equation", + "content": "\\Sigma_{\\mathrm{3D}} = RSS^{\\top}R^{\\top}" + }, + { + "bbox": [ + 313, + 322, + 556, + 467 + ], + "type": "text", + "content": ". In brief, " + }, + { + "bbox": [ + 313, + 322, + 556, + 467 + ], + "type": "inline_equation", + "content": "N" + }, + { + "bbox": [ + 313, + 322, + 556, + 467 + ], + "type": "text", + "content": " numbers of 3D Gaussians can be parametrized as " + }, + { + "bbox": [ + 313, + 322, + 556, + 467 + ], + "type": "inline_equation", + "content": "\\Theta = \\{\\pmb {\\theta}_i\\}_{i = 1}^N = \\{\\pmb {\\mu}_i,S_i,R_i,\\alpha_i,\\mathbf{c}_i\\}_{i = 1}^N" + }, + { + "bbox": [ + 313, + 322, + 556, + 467 + ], + "type": "text", + "content": ". 3D Gaussians " + }, + { + "bbox": [ + 313, + 322, + 556, + 467 + ], + "type": "inline_equation", + "content": "\\Theta" + }, + { + "bbox": [ + 313, + 322, + 556, + 467 + ], + "type": "text", + "content": " are used to render a 2D pixel color c computed as:" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 319, + 475, + 555, + 498 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 319, + 475, + 555, + 498 + ], + "spans": [ + { + "bbox": [ + 319, + 475, + 555, + 498 + ], + "type": "interline_equation", + "content": "\\hat {\\mathbf {c}} (\\theta) = \\sum_ {i = 1} ^ {N} T _ {i} \\tilde {\\alpha} _ {i} \\mathbf {c} _ {i}, \\text {s . t .} \\tilde {\\alpha} _ {i} = \\alpha_ {i} \\exp \\left(- \\frac {1}{2} \\mathbf {d} ^ {\\top} \\Sigma_ {2 D} ^ {- 1} \\mathbf {d}\\right), \\tag {1}", + "image_path": "d15d44a89ea3bf782aab3b2b5b061f6007b1c37fa3a439ea7b61f516492a8fad.jpg" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 313, + 506, + 556, + 616 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 506, + 556, + 616 + ], + "spans": [ + { + "bbox": [ + 313, + 506, + 556, + 616 + ], + "type": "inline_equation", + "content": "T_{i}" + }, + { + "bbox": [ + 313, + 506, + 556, + 616 + ], + "type": "text", + "content": " is a transmittance, " + }, + { + "bbox": [ + 313, + 506, + 556, + 616 + ], + "type": "inline_equation", + "content": "\\tilde{\\alpha}_{i}" + }, + { + "bbox": [ + 313, + 506, + 556, + 616 + ], + "type": "text", + "content": " is an effective opacity value computed from the Gaussian's opacity " + }, + { + "bbox": [ + 313, + 506, + 556, + 616 + ], + "type": "inline_equation", + "content": "\\alpha" + }, + { + "bbox": [ + 313, + 506, + 556, + 616 + ], + "type": "text", + "content": ", the pixel distance " + }, + { + "bbox": [ + 313, + 506, + 556, + 616 + ], + "type": "inline_equation", + "content": "\\mathbf{d} \\in \\mathbb{R}^{2 \\times 1}" + }, + { + "bbox": [ + 313, + 506, + 556, + 616 + ], + "type": "text", + "content": " from the target pixel to the projected center location of the Gaussian in pixel, and " + }, + { + "bbox": [ + 313, + 506, + 556, + 616 + ], + "type": "inline_equation", + "content": "\\Sigma_{2D}" + }, + { + "bbox": [ + 313, + 506, + 556, + 616 + ], + "type": "text", + "content": " is the 2D covariance matrix in the image domain obtained from the splatting algorithm [16, 44]. The 3D Gaussian parameters " + }, + { + "bbox": [ + 313, + 506, + 556, + 616 + ], + "type": "inline_equation", + "content": "\\Theta" + }, + { + "bbox": [ + 313, + 506, + 556, + 616 + ], + "type": "text", + "content": " of a scene are optimized by minimizing the rendering loss between the input image color " + }, + { + "bbox": [ + 313, + 506, + 556, + 616 + ], + "type": "inline_equation", + "content": "\\mathbf{c}" + }, + { + "bbox": [ + 313, + 506, + 556, + 616 + ], + "type": "text", + "content": " and the rendered color " + }, + { + "bbox": [ + 313, + 506, + 556, + 616 + ], + "type": "inline_equation", + "content": "\\hat{\\mathbf{c}}(\\theta)" + }, + { + "bbox": [ + 313, + 506, + 556, + 616 + ], + "type": "text", + "content": " in Eq. (1) as " + }, + { + "bbox": [ + 313, + 506, + 556, + 616 + ], + "type": "inline_equation", + "content": "\\arg \\min_{\\theta} \\| \\mathbf{c} - \\hat{\\mathbf{c}}(\\theta) \\|_F^2" + }, + { + "bbox": [ + 313, + 506, + 556, + 616 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 313, + 617, + 557, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 617, + 557, + 713 + ], + "spans": [ + { + "bbox": [ + 313, + 617, + 557, + 713 + ], + "type": "text", + "content": "Language embedded 3D Gaussian Splitting. The basic idea of the language embedded Gaussian representation [9, 14, 22, 30, 34, 41-43] is to replace the color rendering to language embedding rendering. Language embedded 3D Gaussians are parameterized as " + }, + { + "bbox": [ + 313, + 617, + 557, + 713 + ], + "type": "inline_equation", + "content": "\\Phi = \\{\\theta_i,\\tilde{\\mathbf{f}}_i\\}_{i = 1}^N = \\{\\pmb {\\mu}_i,S_i,R_i,\\alpha_i,\\mathbf{c}_i,\\tilde{\\mathbf{f}}_i\\}_{i = 1}^N" + }, + { + "bbox": [ + 313, + 617, + 557, + 713 + ], + "type": "text", + "content": " where " + }, + { + "bbox": [ + 313, + 617, + 557, + 713 + ], + "type": "inline_equation", + "content": "\\tilde{\\mathbf{f}}_i" + }, + { + "bbox": [ + 313, + 617, + 557, + 713 + ], + "type": "text", + "content": " denotes Gaussian-registered language embeddings across " + }, + { + "bbox": [ + 313, + 617, + 557, + 713 + ], + "type": "inline_equation", + "content": "N" + }, + { + "bbox": [ + 313, + 617, + 557, + 713 + ], + "type": "text", + "content": " numbers 3D Gaussians which will be discussed soon. Then," + } + ] + } + ], + "index": 21 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "type": "text", + "content": "14138" + } + ] + } + ], + "index": 22 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 1 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 69, + 71, + 177, + 179 + ], + "blocks": [ + { + "bbox": [ + 69, + 71, + 177, + 179 + ], + "lines": [ + { + "bbox": [ + 69, + 71, + 177, + 179 + ], + "spans": [ + { + "bbox": [ + 69, + 71, + 177, + 179 + ], + "type": "image", + "image_path": "bc16a901451bbd7d47666029bb23cb06177bb3d61c9169e04f64d082bd270c42.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 74, + 187, + 179, + 199 + ], + "lines": [ + { + "bbox": [ + 74, + 187, + 179, + 199 + ], + "spans": [ + { + "bbox": [ + 74, + 187, + 179, + 199 + ], + "type": "text", + "content": "(a) Preprocessing stage" + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "type": "image", + "bbox": [ + 200, + 73, + 400, + 180 + ], + "blocks": [ + { + "bbox": [ + 200, + 73, + 400, + 180 + ], + "lines": [ + { + "bbox": [ + 200, + 73, + 400, + 180 + ], + "spans": [ + { + "bbox": [ + 200, + 73, + 400, + 180 + ], + "type": "image", + "image_path": "aca0342daac3b053a078af98d85ea144768656025c72e696d9355079d86dd4ea.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 55, + 202, + 555, + 236 + ], + "lines": [ + { + "bbox": [ + 55, + 202, + 555, + 236 + ], + "spans": [ + { + "bbox": [ + 55, + 202, + 555, + 236 + ], + "type": "text", + "content": "Figure 3. Overview of Dr. Splat. (a) In the preprocessing stage, we compute optimized 3D Gaussians [16] and Product Quantization (PQ) codebook construction. (b) During training, we extract CLIP embeddings from given images " + }, + { + "bbox": [ + 55, + 202, + 555, + 236 + ], + "type": "inline_equation", + "content": "\\{\\mathbf{I}\\}" + }, + { + "bbox": [ + 55, + 202, + 555, + 236 + ], + "type": "text", + "content": ", and then proceed feature registration process (Sec. 3.1). Finally, we obtain 3D Gaussians " + }, + { + "bbox": [ + 55, + 202, + 555, + 236 + ], + "type": "inline_equation", + "content": "\\Phi^{\\mathrm{ours}}" + }, + { + "bbox": [ + 55, + 202, + 555, + 236 + ], + "type": "text", + "content": " with PQ indices " + }, + { + "bbox": [ + 55, + 202, + 555, + 236 + ], + "type": "inline_equation", + "content": "\\{j\\}" + }, + { + "bbox": [ + 55, + 202, + 555, + 236 + ], + "type": "text", + "content": " (Sec. 3.2)." + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "image_caption" + } + ], + "index": 2 + }, + { + "type": "image", + "bbox": [ + 402, + 74, + 536, + 180 + ], + "blocks": [ + { + "bbox": [ + 321, + 187, + 401, + 199 + ], + "lines": [ + { + "bbox": [ + 321, + 187, + 401, + 199 + ], + "spans": [ + { + "bbox": [ + 321, + 187, + 401, + 199 + ], + "type": "text", + "content": "(b) Training stage" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 402, + 74, + 536, + 180 + ], + "lines": [ + { + "bbox": [ + 402, + 74, + 536, + 180 + ], + "spans": [ + { + "bbox": [ + 402, + 74, + 536, + 180 + ], + "type": "image", + "image_path": "f54340f180fe2b3c7e15a18582f5fd220e0f09d837291846a941ec1fdcc95234.jpg" + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_body" + } + ], + "index": 4 + }, + { + "bbox": [ + 55, + 256, + 297, + 281 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 256, + 297, + 281 + ], + "spans": [ + { + "bbox": [ + 55, + 256, + 297, + 281 + ], + "type": "text", + "content": "analogous to the color rendering Eq. (1), the language embedding rendering is expressed as:" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 135, + 286, + 296, + 308 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 135, + 286, + 296, + 308 + ], + "spans": [ + { + "bbox": [ + 135, + 286, + 296, + 308 + ], + "type": "interline_equation", + "content": "\\hat {\\mathbf {f}} = \\sum_ {i = 1} ^ {N} T _ {i} \\tilde {\\alpha} _ {i} \\tilde {\\mathbf {f}} _ {i}, \\tag {2}", + "image_path": "d84218af6f27fd891ecd6dd804d9b05501e1d67d0d38d9ae9bc3b4a44f004992.jpg" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 54, + 315, + 296, + 495 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 54, + 315, + 296, + 495 + ], + "spans": [ + { + "bbox": [ + 54, + 315, + 296, + 495 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 54, + 315, + 296, + 495 + ], + "type": "inline_equation", + "content": "\\hat{\\mathbf{f}}" + }, + { + "bbox": [ + 54, + 315, + 296, + 495 + ], + "type": "text", + "content": " denotes a rendered language embedding. Likewise, the Gaussian-registered language embeddings " + }, + { + "bbox": [ + 54, + 315, + 296, + 495 + ], + "type": "inline_equation", + "content": "\\{\\tilde{\\mathbf{f}}\\}" + }, + { + "bbox": [ + 54, + 315, + 296, + 495 + ], + "type": "text", + "content": " are optimized by minimizing the rendering loss between the 2D language embedding " + }, + { + "bbox": [ + 54, + 315, + 296, + 495 + ], + "type": "inline_equation", + "content": "\\mathbf{f}" + }, + { + "bbox": [ + 54, + 315, + 296, + 495 + ], + "type": "text", + "content": " extracted from an input image and a rendered language embedding map " + }, + { + "bbox": [ + 54, + 315, + 296, + 495 + ], + "type": "inline_equation", + "content": "\\hat{\\mathbf{f}}" + }, + { + "bbox": [ + 54, + 315, + 296, + 495 + ], + "type": "text", + "content": " as " + }, + { + "bbox": [ + 54, + 315, + 296, + 495 + ], + "type": "inline_equation", + "content": "\\arg \\min_{\\{\\tilde{\\mathbf{f}}\\}}\\| \\mathbf{f} - \\hat{\\mathbf{f}} \\|_F^2" + }, + { + "bbox": [ + 54, + 315, + 296, + 495 + ], + "type": "text", + "content": " at each corresponding pixel. This can be regarded as distilling vision language models into Gaussian-registered language embedding " + }, + { + "bbox": [ + 54, + 315, + 296, + 495 + ], + "type": "inline_equation", + "content": "\\tilde{\\mathbf{f}}" + }, + { + "bbox": [ + 54, + 315, + 296, + 495 + ], + "type": "text", + "content": " through volume rendering Eq. (2). The Gaussian-registered language embeddings are separately trained after pre-training and fixing the pre-trained 3DGS " + }, + { + "bbox": [ + 54, + 315, + 296, + 495 + ], + "type": "inline_equation", + "content": "\\Theta" + }, + { + "bbox": [ + 54, + 315, + 296, + 495 + ], + "type": "text", + "content": " for a scene. The language embeddings to be distilled are typically obtained from CLIP [31]. Since storing 32-bit 512-D CLIP features " + }, + { + "bbox": [ + 54, + 315, + 296, + 495 + ], + "type": "inline_equation", + "content": "\\mathbf{f}" + }, + { + "bbox": [ + 54, + 315, + 296, + 495 + ], + "type": "text", + "content": " in every 3D Gaussians is memory-expensive, one can use a compressed feature per scene depending on the needs [14, 19, 34, 41-43]." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 55, + 498, + 296, + 701 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 498, + 296, + 701 + ], + "spans": [ + { + "bbox": [ + 55, + 498, + 296, + 701 + ], + "type": "text", + "content": "Motivation. Such language-embedded radiance fields provide useful representation and language interfaces for many practical and crucial applications. While most of existing works focus on the training efficiency, the complexity in inference time has barely been discussed. Considering a scenario to text-query a 3D location of the language-embedded Gaussians, i.e., 3D localization, the aforementioned methods first require rendering a 2D language embedding map at each specific camera pose. We cannot directly retrieve over the distributed embeddings " + }, + { + "bbox": [ + 55, + 498, + 296, + 701 + ], + "type": "inline_equation", + "content": "\\{\\tilde{\\mathbf{f}}_i\\}" + }, + { + "bbox": [ + 55, + 498, + 296, + 701 + ], + "type": "text", + "content": " in 3D Gaussians, because the embeddings do not carry language information directly, but their weighted summed (rendered) features " + }, + { + "bbox": [ + 55, + 498, + 296, + 701 + ], + "type": "inline_equation", + "content": "\\hat{\\mathbf{f}}" + }, + { + "bbox": [ + 55, + 498, + 296, + 701 + ], + "type": "text", + "content": " do. This issue becomes even severer with compressed features as in [30]: their decompression decoders are not designed for and incompatible with directly applying to the distributed compressed language embeddings in each 3D Gaussian, yielding degenerated CLIP decoding (refer to Fig. 2)." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 67, + 701, + 295, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 67, + 701, + 295, + 713 + ], + "spans": [ + { + "bbox": [ + 67, + 701, + 295, + 713 + ], + "type": "text", + "content": "This introduces multiple challenges and hassles. First, it" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 313, + 256, + 555, + 340 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 256, + 555, + 340 + ], + "spans": [ + { + "bbox": [ + 313, + 256, + 555, + 340 + ], + "type": "text", + "content": "is challenging to find the best or proper camera rendering views that contain the object to find. One may attempt to pre-compute the minimal number of cameras and their camera poses that cover all the 3D Gaussians in a scene with proper resolutions, similarly by point-based approach [11]. However, this is a well-known set covering problem [7] with constraints which is known to be an NP-hard problem." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 313, + 342, + 556, + 520 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 342, + 556, + 520 + ], + "spans": [ + { + "bbox": [ + 313, + 342, + 556, + 520 + ], + "type": "text", + "content": "Second, even with pre-computed rendered views, the retrieval complexity over the rendered images remains substantial [8]. Suppose a scene consisting of one million Gaussians, but just a single rendered language embedding map in pixel domain already has nearly a million pixels; thus, we need a dedicated system to efficiently retrieve over all the views. Third, since the retrieval is conducted in the 2D space, to find a 3D location, we need a separate mechanism to lift the localization to the 3D space, i.e., increasing the system complexity. In addition, 32-bit floating 512-Dimension CLIP features for millions of Gaussian are memory intensive, which is often not manageable. To reduce this burden, the existing methods [37] apply compressions with per-scene optimized codebooks, which hinders extension or generalization to other scenes." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 313, + 522, + 556, + 594 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 522, + 556, + 594 + ], + "spans": [ + { + "bbox": [ + 313, + 522, + 556, + 594 + ], + "type": "text", + "content": "To overcome these, we propose a training-free algorithm for the direct allocation of language embeddings to 3D Gaussians, allowing efficient computation and interaction within the 3D space. As a concurrent work, OpenGaussian [37] tackles a similar challenge with our work, but still requires per-scene codebook construction Fig. 1." + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 313, + 608, + 375, + 622 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 608, + 375, + 622 + ], + "spans": [ + { + "bbox": [ + 313, + 608, + 375, + 622 + ], + "type": "text", + "content": "3. Dr. Splat" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 313, + 629, + 556, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 629, + 556, + 713 + ], + "spans": [ + { + "bbox": [ + 313, + 629, + 556, + 713 + ], + "type": "text", + "content": "This section provides details of our method. We first explain how we directly register CLIP embeddings into Gaussian-registered language embeddings, Sec. 3.1. Then, we introduce Product Quantization (PQ) into our framework to efficiently store Gaussian-registered language embeddings, Sec. 3.2. Lastly, we describe the inference stage for text query-based 3D Gaussian localization, Sec. 3.3." + } + ] + } + ], + "index": 15 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "text", + "content": "14139" + } + ] + } + ], + "index": 16 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 2 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 86, + 70, + 211, + 186 + ], + "blocks": [ + { + "bbox": [ + 86, + 70, + 211, + 186 + ], + "lines": [ + { + "bbox": [ + 86, + 70, + 211, + 186 + ], + "spans": [ + { + "bbox": [ + 86, + 70, + 211, + 186 + ], + "type": "image", + "image_path": "6a58cf2fe02beb59208dadb777373d4f5453e42fdd45b6e7b9a238b3f49badaf.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 67, + 186, + 223, + 198 + ], + "lines": [ + { + "bbox": [ + 67, + 186, + 223, + 198 + ], + "spans": [ + { + "bbox": [ + 67, + 186, + 223, + 198 + ], + "type": "text", + "content": "(a) Map CLIP features to Gaussians" + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 55, + 201, + 555, + 236 + ], + "lines": [ + { + "bbox": [ + 55, + 201, + 555, + 236 + ], + "spans": [ + { + "bbox": [ + 55, + 201, + 555, + 236 + ], + "type": "text", + "content": "Figure 4. Feature registration process in Dr. Splat. (a) We first map per-pixel CLIP embeddings " + }, + { + "bbox": [ + 55, + 201, + 555, + 236 + ], + "type": "inline_equation", + "content": "\\{\\mathbf{f}^{\\mathrm{map}}\\}" + }, + { + "bbox": [ + 55, + 201, + 555, + 236 + ], + "type": "text", + "content": " to Gaussians. Here, we only map dominant " + }, + { + "bbox": [ + 55, + 201, + 555, + 236 + ], + "type": "inline_equation", + "content": "k" + }, + { + "bbox": [ + 55, + 201, + 555, + 236 + ], + "type": "text", + "content": " Gaussians along pixel ray " + }, + { + "bbox": [ + 55, + 201, + 555, + 236 + ], + "type": "inline_equation", + "content": "r" + }, + { + "bbox": [ + 55, + 201, + 555, + 236 + ], + "type": "text", + "content": ", named Top-" + }, + { + "bbox": [ + 55, + 201, + 555, + 236 + ], + "type": "inline_equation", + "content": "k" + }, + { + "bbox": [ + 55, + 201, + 555, + 236 + ], + "type": "text", + "content": " Gaussians. (b) After collecting embeddings, we compute aggregated features (Eq. (6)). (c) Finally, we re-use PQ to obtain the PQ indices " + }, + { + "bbox": [ + 55, + 201, + 555, + 236 + ], + "type": "inline_equation", + "content": "j" + }, + { + "bbox": [ + 55, + 201, + 555, + 236 + ], + "type": "text", + "content": " of aggregated features and update Gaussian parameters " + }, + { + "bbox": [ + 55, + 201, + 555, + 236 + ], + "type": "inline_equation", + "content": "\\Phi^{\\mathrm{ours}}" + }, + { + "bbox": [ + 55, + 201, + 555, + 236 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "type": "image", + "bbox": [ + 220, + 69, + 539, + 185 + ], + "blocks": [ + { + "bbox": [ + 220, + 69, + 539, + 185 + ], + "lines": [ + { + "bbox": [ + 220, + 69, + 539, + 185 + ], + "spans": [ + { + "bbox": [ + 220, + 69, + 539, + 185 + ], + "type": "image", + "image_path": "4971ba003bc5b4ef9be830de69b10f0accca954ab1a688cecfcc1fc394f5d2b0.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 234, + 186, + 382, + 199 + ], + "lines": [ + { + "bbox": [ + 234, + 186, + 382, + 199 + ], + "spans": [ + { + "bbox": [ + 234, + 186, + 382, + 199 + ], + "type": "text", + "content": "(b) Aggregate multiview features" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 413, + 186, + 539, + 198 + ], + "lines": [ + { + "bbox": [ + 413, + 186, + 539, + 198 + ], + "spans": [ + { + "bbox": [ + 413, + 186, + 539, + 198 + ], + "type": "text", + "content": "(c) Register feature to " + }, + { + "bbox": [ + 413, + 186, + 539, + 198 + ], + "type": "inline_equation", + "content": "\\Phi^{\\text{ours}}" + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_caption" + } + ], + "index": 2 + }, + { + "bbox": [ + 55, + 256, + 209, + 269 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 256, + 209, + 269 + ], + "spans": [ + { + "bbox": [ + 55, + 256, + 209, + 269 + ], + "type": "text", + "content": "3.1. Feature registration process" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 55, + 274, + 296, + 468 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 274, + 296, + 468 + ], + "spans": [ + { + "bbox": [ + 55, + 274, + 296, + 468 + ], + "type": "text", + "content": "Our goal is to reconstruct a language embedded 3D space represented by 3D Gaussians " + }, + { + "bbox": [ + 55, + 274, + 296, + 468 + ], + "type": "inline_equation", + "content": "\\Phi" + }, + { + "bbox": [ + 55, + 274, + 296, + 468 + ], + "type": "text", + "content": ", which we can directly interact in 3D space without feature rendering Eq. (2). For that, following LangSplat [30], we begin by extracting per-pixel CLIP embedding maps " + }, + { + "bbox": [ + 55, + 274, + 296, + 468 + ], + "type": "inline_equation", + "content": "\\mathbf{F}^{\\mathrm{map}} \\in \\mathbb{R}^{D \\times H \\times W}" + }, + { + "bbox": [ + 55, + 274, + 296, + 468 + ], + "type": "text", + "content": " from training images of the target scenes, where " + }, + { + "bbox": [ + 55, + 274, + 296, + 468 + ], + "type": "inline_equation", + "content": "D" + }, + { + "bbox": [ + 55, + 274, + 296, + 468 + ], + "type": "text", + "content": " is the dimension of CLIP embeddings, " + }, + { + "bbox": [ + 55, + 274, + 296, + 468 + ], + "type": "inline_equation", + "content": "H" + }, + { + "bbox": [ + 55, + 274, + 296, + 468 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 55, + 274, + 296, + 468 + ], + "type": "inline_equation", + "content": "W" + }, + { + "bbox": [ + 55, + 274, + 296, + 468 + ], + "type": "text", + "content": " are the height and width of the training images. Given training images, we extracts a dictionary of binary masks and language embeddings extracted from the images as: " + }, + { + "bbox": [ + 55, + 274, + 296, + 468 + ], + "type": "inline_equation", + "content": "\\mathcal{F}^{\\mathrm{map}} = \\{\\mathbf{M}_j : \\mathbf{f}_j^{\\mathrm{map}} \\mid j = 1, \\dots, M\\}" + }, + { + "bbox": [ + 55, + 274, + 296, + 468 + ], + "type": "text", + "content": ", where " + }, + { + "bbox": [ + 55, + 274, + 296, + 468 + ], + "type": "inline_equation", + "content": "\\mathbf{M}_j \\in \\mathbb{R}^{H \\times W}" + }, + { + "bbox": [ + 55, + 274, + 296, + 468 + ], + "type": "text", + "content": " is a binary mask extracted using SAM [20] and " + }, + { + "bbox": [ + 55, + 274, + 296, + 468 + ], + "type": "inline_equation", + "content": "\\mathbf{f}_j^{\\mathrm{map}} \\in \\mathbb{R}^D" + }, + { + "bbox": [ + 55, + 274, + 296, + 468 + ], + "type": "text", + "content": " is a corresponding CLIP embedding from a cropped image with " + }, + { + "bbox": [ + 55, + 274, + 296, + 468 + ], + "type": "inline_equation", + "content": "\\mathbf{M}_j" + }, + { + "bbox": [ + 55, + 274, + 296, + 468 + ], + "type": "text", + "content": ". Each mask " + }, + { + "bbox": [ + 55, + 274, + 296, + 468 + ], + "type": "inline_equation", + "content": "\\mathbf{M}_j" + }, + { + "bbox": [ + 55, + 274, + 296, + 468 + ], + "type": "text", + "content": " belongs to an image, and the masks are not overlapped to each other. With this dictionary, a CLIP embedding map " + }, + { + "bbox": [ + 55, + 274, + 296, + 468 + ], + "type": "inline_equation", + "content": "\\mathbf{F}^{\\mathrm{map}}(\\mathbf{I}, \\mathbf{r})" + }, + { + "bbox": [ + 55, + 274, + 296, + 468 + ], + "type": "text", + "content": " at a pixel " + }, + { + "bbox": [ + 55, + 274, + 296, + 468 + ], + "type": "inline_equation", + "content": "\\mathbf{r}" + }, + { + "bbox": [ + 55, + 274, + 296, + 468 + ], + "type": "text", + "content": " in a training image " + }, + { + "bbox": [ + 55, + 274, + 296, + 468 + ], + "type": "inline_equation", + "content": "\\mathbf{I}" + }, + { + "bbox": [ + 55, + 274, + 296, + 468 + ], + "type": "text", + "content": " is computed as:" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 99, + 477, + 296, + 500 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 99, + 477, + 296, + 500 + ], + "spans": [ + { + "bbox": [ + 99, + 477, + 296, + 500 + ], + "type": "interline_equation", + "content": "\\mathbf {F} ^ {\\text {m a p}} (\\mathbf {I}, \\mathbf {r}) = \\sum_ {j = 1} ^ {M} \\mathbf {M} _ {j} (\\mathbf {I}, \\mathbf {r}) \\cdot \\mathbf {f} _ {j} ^ {\\text {m a p}}, \\tag {3}", + "image_path": "d08173a415543e29b3d5c8ec9b45095435e379e6538a7411c31290aac58e3992.jpg" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 55, + 510, + 296, + 557 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 510, + 296, + 557 + ], + "spans": [ + { + "bbox": [ + 55, + 510, + 296, + 557 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 55, + 510, + 296, + 557 + ], + "type": "inline_equation", + "content": "\\mathbf{M}_j(\\mathbf{I},\\mathbf{r})\\in \\{0,1\\}" + }, + { + "bbox": [ + 55, + 510, + 296, + 557 + ], + "type": "text", + "content": " indicates whether the mask " + }, + { + "bbox": [ + 55, + 510, + 296, + 557 + ], + "type": "inline_equation", + "content": "\\mathbf{M}_j" + }, + { + "bbox": [ + 55, + 510, + 296, + 557 + ], + "type": "text", + "content": " contains the pixel " + }, + { + "bbox": [ + 55, + 510, + 296, + 557 + ], + "type": "inline_equation", + "content": "\\mathbf{r}" + }, + { + "bbox": [ + 55, + 510, + 296, + 557 + ], + "type": "text", + "content": " in the image " + }, + { + "bbox": [ + 55, + 510, + 296, + 557 + ], + "type": "inline_equation", + "content": "\\mathbf{I}" + }, + { + "bbox": [ + 55, + 510, + 296, + 557 + ], + "type": "text", + "content": ". Using " + }, + { + "bbox": [ + 55, + 510, + 296, + 557 + ], + "type": "inline_equation", + "content": "\\mathbf{F}^{\\mathrm{map}}" + }, + { + "bbox": [ + 55, + 510, + 296, + 557 + ], + "type": "text", + "content": ", we reconstruct language embedded 3D Gaussians via a novel feature registration process as visualized in Fig. 3." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 55, + 559, + 296, + 690 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 559, + 296, + 690 + ], + "spans": [ + { + "bbox": [ + 55, + 559, + 296, + 690 + ], + "type": "text", + "content": "During the feature registration process, our algorithm iterates through training images of the scene. Using projection relation, we link 3D Gaussians " + }, + { + "bbox": [ + 55, + 559, + 296, + 690 + ], + "type": "inline_equation", + "content": "\\Phi" + }, + { + "bbox": [ + 55, + 559, + 296, + 690 + ], + "type": "text", + "content": " to CLIP embeddings. Each Gaussian can link to multiple CLIP embeddings derived from different images. Then we aggregate collected embeddings to a single embedding to be assigned to each Gaussian. To ensure a consistent aggregation of the embeddings from multi-view images, we first compute a weight " + }, + { + "bbox": [ + 55, + 559, + 296, + 690 + ], + "type": "inline_equation", + "content": "w_{i}(\\mathbf{I},\\mathbf{r})" + }, + { + "bbox": [ + 55, + 559, + 296, + 690 + ], + "type": "text", + "content": " representing the contribution of " + }, + { + "bbox": [ + 55, + 559, + 296, + 690 + ], + "type": "inline_equation", + "content": "\\theta_{i}" + }, + { + "bbox": [ + 55, + 559, + 296, + 690 + ], + "type": "text", + "content": " to construct each pixel " + }, + { + "bbox": [ + 55, + 559, + 296, + 690 + ], + "type": "inline_equation", + "content": "\\mathbf{r}" + }, + { + "bbox": [ + 55, + 559, + 296, + 690 + ], + "type": "text", + "content": " in a training image " + }, + { + "bbox": [ + 55, + 559, + 296, + 690 + ], + "type": "inline_equation", + "content": "\\mathbf{I}" + }, + { + "bbox": [ + 55, + 559, + 296, + 690 + ], + "type": "text", + "content": ". The weights are computed with the volume rendering equation Eq. (1) as:" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 115, + 701, + 296, + 714 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 115, + 701, + 296, + 714 + ], + "spans": [ + { + "bbox": [ + 115, + 701, + 296, + 714 + ], + "type": "interline_equation", + "content": "w _ {i} (\\mathbf {I}, \\mathbf {r}) = T _ {i} (\\mathbf {I}, \\mathbf {r}) \\cdot \\tilde {\\alpha} _ {i} (\\mathbf {I}, \\mathbf {r}), \\tag {4}", + "image_path": "ac5df0d8282b0204ce8b7a98c4d6957c15beaa427b08153e1747d9a06c4b9220.jpg" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 313, + 256, + 555, + 340 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 256, + 555, + 340 + ], + "spans": [ + { + "bbox": [ + 313, + 256, + 555, + 340 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 313, + 256, + 555, + 340 + ], + "type": "inline_equation", + "content": "T_{i}(\\mathbf{I},\\mathbf{r})" + }, + { + "bbox": [ + 313, + 256, + 555, + 340 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 313, + 256, + 555, + 340 + ], + "type": "inline_equation", + "content": "\\tilde{\\alpha}_{i}(\\mathbf{I},\\mathbf{r})" + }, + { + "bbox": [ + 313, + 256, + 555, + 340 + ], + "type": "text", + "content": " are the transmittance and the effective opacity value of " + }, + { + "bbox": [ + 313, + 256, + 555, + 340 + ], + "type": "inline_equation", + "content": "\\theta_{i}" + }, + { + "bbox": [ + 313, + 256, + 555, + 340 + ], + "type": "text", + "content": " for a pixel " + }, + { + "bbox": [ + 313, + 256, + 555, + 340 + ], + "type": "inline_equation", + "content": "\\mathbf{r}" + }, + { + "bbox": [ + 313, + 256, + 555, + 340 + ], + "type": "text", + "content": " in an image " + }, + { + "bbox": [ + 313, + 256, + 555, + 340 + ], + "type": "inline_equation", + "content": "\\mathbf{I}" + }, + { + "bbox": [ + 313, + 256, + 555, + 340 + ], + "type": "text", + "content": ", stated in Eq. (1). With the per-pixel weights, we calculate " + }, + { + "bbox": [ + 313, + 256, + 555, + 340 + ], + "type": "inline_equation", + "content": "w_{ij}" + }, + { + "bbox": [ + 313, + 256, + 555, + 340 + ], + "type": "text", + "content": " representing a weight between each Gaussian " + }, + { + "bbox": [ + 313, + 256, + 555, + 340 + ], + "type": "inline_equation", + "content": "\\theta_{i}" + }, + { + "bbox": [ + 313, + 256, + 555, + 340 + ], + "type": "text", + "content": " and corresponding language embedding maps " + }, + { + "bbox": [ + 313, + 256, + 555, + 340 + ], + "type": "inline_equation", + "content": "\\mathbf{f}_j^{\\mathrm{map}}" + }, + { + "bbox": [ + 313, + 256, + 555, + 340 + ], + "type": "text", + "content": ", which is for aggregating CLIP embeddings from " + }, + { + "bbox": [ + 313, + 256, + 555, + 340 + ], + "type": "inline_equation", + "content": "\\mathbf{F}^{\\mathrm{map}}" + }, + { + "bbox": [ + 313, + 256, + 555, + 340 + ], + "type": "text", + "content": " and register the embedding to each Gaussian. The weights are computed as:" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 350, + 349, + 555, + 369 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 350, + 349, + 555, + 369 + ], + "spans": [ + { + "bbox": [ + 350, + 349, + 555, + 369 + ], + "type": "interline_equation", + "content": "w _ {i j} = \\sum_ {\\mathbf {I} \\in \\mathcal {I}} \\sum_ {\\mathbf {r} \\in \\mathbf {I}} \\mathbf {M} _ {j} (\\mathbf {I}, \\mathbf {r}) \\cdot w _ {i} (\\mathbf {I}, \\mathbf {r}), \\tag {5}", + "image_path": "6466e36a6315320bda57a48628668a65ceda771e56e12b9ffa444c2819bd850d.jpg" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 313, + 375, + 555, + 495 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 375, + 555, + 495 + ], + "spans": [ + { + "bbox": [ + 313, + 375, + 555, + 495 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 313, + 375, + 555, + 495 + ], + "type": "inline_equation", + "content": "\\mathcal{I}" + }, + { + "bbox": [ + 313, + 375, + 555, + 495 + ], + "type": "text", + "content": " is the set of the training images. In this iterative process, we aggregate weights only for Top-" + }, + { + "bbox": [ + 313, + 375, + 555, + 495 + ], + "type": "inline_equation", + "content": "k" + }, + { + "bbox": [ + 313, + 375, + 555, + 495 + ], + "type": "text", + "content": " Gaussians with the highest weights " + }, + { + "bbox": [ + 313, + 375, + 555, + 495 + ], + "type": "inline_equation", + "content": "w_{i}(\\mathbf{I},\\mathbf{r})" + }, + { + "bbox": [ + 313, + 375, + 555, + 495 + ], + "type": "text", + "content": ", along the ray of each pixel ray " + }, + { + "bbox": [ + 313, + 375, + 555, + 495 + ], + "type": "inline_equation", + "content": "\\mathbf{r}" + }, + { + "bbox": [ + 313, + 375, + 555, + 495 + ], + "type": "text", + "content": " (see Fig. 4). After aggregation, we prune the Gaussians which are not assigned any weight, i.e., " + }, + { + "bbox": [ + 313, + 375, + 555, + 495 + ], + "type": "inline_equation", + "content": "\\sum_{j=1}^{M} w_{ij} = 0" + }, + { + "bbox": [ + 313, + 375, + 555, + 495 + ], + "type": "text", + "content": ". This summation aggregates weights between Gaussians and the CLIP embeddings by linking per-pixel weights " + }, + { + "bbox": [ + 313, + 375, + 555, + 495 + ], + "type": "inline_equation", + "content": "w_{i}(\\mathbf{I},\\mathbf{r})" + }, + { + "bbox": [ + 313, + 375, + 555, + 495 + ], + "type": "text", + "content": " of each Gaussian to its corresponding CLIP embeddings. With the obtained weights, we register an aggregated feature " + }, + { + "bbox": [ + 313, + 375, + 555, + 495 + ], + "type": "inline_equation", + "content": "\\dot{\\mathbf{f}}_i" + }, + { + "bbox": [ + 313, + 375, + 555, + 495 + ], + "type": "text", + "content": " to each Gaussian with weighted-averaging as:" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 324, + 503, + 555, + 526 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 324, + 503, + 555, + 526 + ], + "spans": [ + { + "bbox": [ + 324, + 503, + 555, + 526 + ], + "type": "interline_equation", + "content": "\\dot {\\mathbf {f}} _ {i} = \\mathbf {f} _ {i} / | | \\mathbf {f} _ {i} | | _ {2}, \\text {w h e r e} \\quad \\mathbf {f} _ {i} = \\sum_ {j = 1} ^ {M} \\frac {w _ {i j}}{\\sum_ {k = 1} ^ {M} w _ {i k}} \\mathbf {f} _ {j} ^ {\\text {m a p}}. \\tag {6}", + "image_path": "34d2b851ecff8d5f671312838d03ff460b5480dbd6a2a7d15a0854acd2ff614f.jpg" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 313, + 533, + 556, + 628 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 533, + 556, + 628 + ], + "spans": [ + { + "bbox": [ + 313, + 533, + 556, + 628 + ], + "type": "text", + "content": "This process enables 3D-aware feature registration to be consistent across various viewpoints, by aggregating features in the original high-dimensional feature space. The proposed process can be interpreted as an inverse volume rendering without gradient-based optimization, which enables our method to be faster than the prior methods requiring perscene gradient-based optimization [27, 30, 34] for feature registration in 3D space." + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 313, + 635, + 515, + 649 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 635, + 515, + 649 + ], + "spans": [ + { + "bbox": [ + 313, + 635, + 515, + 649 + ], + "type": "text", + "content": "3.2. Product-Quantized CLIP embeddings" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 313, + 653, + 555, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 653, + 555, + 713 + ], + "spans": [ + { + "bbox": [ + 313, + 653, + 555, + 713 + ], + "type": "text", + "content": "Memory efficiency is a challenge in 3D scene representations, especially when associating Gaussians with high-dimensional feature vectors. LangSplat [30] addresses this by introducing an encoder-decoder network, while LeGaussian [34] and OpenGaussian [37] utilize codebook construc" + } + ] + } + ], + "index": 18 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "text", + "content": "14140" + } + ] + } + ], + "index": 19 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 3 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 70, + 70, + 541, + 153 + ], + "blocks": [ + { + "bbox": [ + 70, + 70, + 541, + 153 + ], + "lines": [ + { + "bbox": [ + 70, + 70, + 541, + 153 + ], + "spans": [ + { + "bbox": [ + 70, + 70, + 541, + 153 + ], + "type": "table", + "html": "
MethodsmIoUmAcc @ 0.25
waldo_kitchenramenfigurinesteatimeMeanwaldo_kitchenramenfigurinesteatimeMean
LangSplat-m [30]8.296.118.3316.589.8313.6414.088.9327.1215.94
OpenGaussian [37]34.6023.8759.3354.4443.0650.0035.2180.3672.8859.61
Ours (Top-10)37.0524.3354.4257.3543.2963.6435.2180.3677.9764.30
Ours (Top-20)38.3324.5853.9456.1943.2663.6435.2182.1476.2764.32
Ours (Top-40)39.0724.7053.3657.2043.5863.6435.2180.3676.2763.87
", + "image_path": "2c0571696a15a9a53254393435519e76a5a4e247f8fa9cc87f1782e132578b50.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "table_body" + } + ], + "index": 0 + }, + { + "bbox": [ + 55, + 161, + 555, + 195 + ], + "lines": [ + { + "bbox": [ + 55, + 161, + 555, + 195 + ], + "spans": [ + { + "bbox": [ + 55, + 161, + 555, + 195 + ], + "type": "text", + "content": "Table 1. 3D object selection results on the LeRF-OVS dataset [17]. To measure 3D object selection performance, we calculate 2D segmentation accuracy on rendering of selected 3D Gaussians. Note that our model does not require per-scene optimization, demonstrating its robustness across diverse scenes. Bold and Underline stand for first and second best performance." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "text" + }, + { + "type": "image", + "bbox": [ + 70, + 215, + 154, + 384 + ], + "blocks": [ + { + "bbox": [ + 70, + 215, + 154, + 384 + ], + "lines": [ + { + "bbox": [ + 70, + 215, + 154, + 384 + ], + "spans": [ + { + "bbox": [ + 70, + 215, + 154, + 384 + ], + "type": "image", + "image_path": "42dd55066e5ac092d686ad46cf83089ae2f133b774de38c9a7f70912a0fdcd22.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 55, + 387, + 555, + 431 + ], + "lines": [ + { + "bbox": [ + 55, + 387, + 555, + 431 + ], + "spans": [ + { + "bbox": [ + 55, + 387, + 555, + 431 + ], + "type": "text", + "content": "Figure 5. Qualitative results of the object selection on the LeRF-OVS dataset [17]. We visualize rendering of selected 3D Gaussians for LangSplat [30], OpenGaussian [37], and ours. For LangSplat, activations are often distributed randomly, fail to localize the target. OpenGaussian often struggles to distinguish closely situated objects. In contrast, our model shows activations precisely limited to the queried object regions, effectively localizing only the relevant areas." + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "image_caption" + } + ], + "index": 2 + }, + { + "type": "image", + "bbox": [ + 154, + 215, + 228, + 384 + ], + "blocks": [ + { + "bbox": [ + 154, + 215, + 228, + 384 + ], + "lines": [ + { + "bbox": [ + 154, + 215, + 228, + 384 + ], + "spans": [ + { + "bbox": [ + 154, + 215, + 228, + 384 + ], + "type": "image", + "image_path": "5cfd6f2ed8ab85d8fd3a5e46d87974d73f21e5beebe0ef40a115c4d6b44692bc.jpg" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_body" + } + ], + "index": 3 + }, + { + "type": "image", + "bbox": [ + 230, + 215, + 303, + 384 + ], + "blocks": [ + { + "bbox": [ + 230, + 215, + 303, + 384 + ], + "lines": [ + { + "bbox": [ + 230, + 215, + 303, + 384 + ], + "spans": [ + { + "bbox": [ + 230, + 215, + 303, + 384 + ], + "type": "image", + "image_path": "23d2d3d1f429017f8d554d145f66b3475175ec479fc75eebbbfac5f2953c0d18.jpg" + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_body" + } + ], + "index": 4 + }, + { + "type": "image", + "bbox": [ + 307, + 215, + 390, + 384 + ], + "blocks": [ + { + "bbox": [ + 307, + 215, + 390, + 384 + ], + "lines": [ + { + "bbox": [ + 307, + 215, + 390, + 384 + ], + "spans": [ + { + "bbox": [ + 307, + 215, + 390, + 384 + ], + "type": "image", + "image_path": "0245c63cd0c27778f78a287b354a76abb1e7514fe05af7fa14b11d54ebd7d9f3.jpg" + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "image_body" + } + ], + "index": 5 + }, + { + "type": "image", + "bbox": [ + 391, + 215, + 464, + 384 + ], + "blocks": [ + { + "bbox": [ + 391, + 215, + 464, + 384 + ], + "lines": [ + { + "bbox": [ + 391, + 215, + 464, + 384 + ], + "spans": [ + { + "bbox": [ + 391, + 215, + 464, + 384 + ], + "type": "image", + "image_path": "7abf6229bb01b5699460772cb48c1322fec2e565a156a66d0108d37f81152783.jpg" + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "image_body" + } + ], + "index": 6 + }, + { + "type": "image", + "bbox": [ + 466, + 215, + 540, + 384 + ], + "blocks": [ + { + "bbox": [ + 466, + 215, + 540, + 384 + ], + "lines": [ + { + "bbox": [ + 466, + 215, + 540, + 384 + ], + "spans": [ + { + "bbox": [ + 466, + 215, + 540, + 384 + ], + "type": "image", + "image_path": "1dd7bc29c4e690359fc8c42a213d96ef133a8b2c20cd2abd3d025e8c16a17924.jpg" + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "image_body" + } + ], + "index": 7 + }, + { + "bbox": [ + 55, + 453, + 297, + 513 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 453, + 297, + 513 + ], + "spans": [ + { + "bbox": [ + 55, + 453, + 297, + 513 + ], + "type": "text", + "content": "tion. However, these approaches introduce additional per-scene computational costs for scene-specific parameter tuning of neural networks or codebooks (see Fig. 1). In contrast, we propose to use Product Quantization (PQ) on a large-scale image dataset, eliminating per-scene training." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 55, + 518, + 295, + 685 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 518, + 295, + 685 + ], + "spans": [ + { + "bbox": [ + 55, + 518, + 295, + 685 + ], + "type": "text", + "content": "Product Quantization. PQ [13] is a widely used technique for efficient embedding compression, particularly valuable in large-scale applications. The PQ process begins by dividing the original " + }, + { + "bbox": [ + 55, + 518, + 295, + 685 + ], + "type": "inline_equation", + "content": "D" + }, + { + "bbox": [ + 55, + 518, + 295, + 685 + ], + "type": "text", + "content": "-dimensional feature vector " + }, + { + "bbox": [ + 55, + 518, + 295, + 685 + ], + "type": "inline_equation", + "content": "\\mathbf{v}" + }, + { + "bbox": [ + 55, + 518, + 295, + 685 + ], + "type": "text", + "content": " into " + }, + { + "bbox": [ + 55, + 518, + 295, + 685 + ], + "type": "inline_equation", + "content": "L" + }, + { + "bbox": [ + 55, + 518, + 295, + 685 + ], + "type": "text", + "content": " sub-vectors: " + }, + { + "bbox": [ + 55, + 518, + 295, + 685 + ], + "type": "inline_equation", + "content": "\\mathbf{v} = [\\mathbf{v}_1,\\mathbf{v}_2,\\dots ,\\mathbf{v}_L]" + }, + { + "bbox": [ + 55, + 518, + 295, + 685 + ], + "type": "text", + "content": ". Each sub-vector " + }, + { + "bbox": [ + 55, + 518, + 295, + 685 + ], + "type": "inline_equation", + "content": "\\mathbf{v}_i" + }, + { + "bbox": [ + 55, + 518, + 295, + 685 + ], + "type": "text", + "content": " is then independently quantized to a predefined number of centroids " + }, + { + "bbox": [ + 55, + 518, + 295, + 685 + ], + "type": "inline_equation", + "content": "s_{ij}" + }, + { + "bbox": [ + 55, + 518, + 295, + 685 + ], + "type": "text", + "content": " in a predefined codebook " + }, + { + "bbox": [ + 55, + 518, + 295, + 685 + ], + "type": "inline_equation", + "content": "S_{i}" + }, + { + "bbox": [ + 55, + 518, + 295, + 685 + ], + "type": "text", + "content": " for that sub-vector. These centroids are learned via clustering, creating a codebook for each subspace. Once the centroids are established, each sub-vector is replaced by the index of the nearest centroid in its respective codebook. The centroid indices " + }, + { + "bbox": [ + 55, + 518, + 295, + 685 + ], + "type": "inline_equation", + "content": "j_{i} = [j_{i1},j_{i2},\\ldots ,j_{iL}]" + }, + { + "bbox": [ + 55, + 518, + 295, + 685 + ], + "type": "text", + "content": " are optimized by minimizing " + }, + { + "bbox": [ + 55, + 518, + 295, + 685 + ], + "type": "inline_equation", + "content": "\\arg \\min_k\\| \\mathbf{v}_i - s_{ik}\\|" + }, + { + "bbox": [ + 55, + 518, + 295, + 685 + ], + "type": "text", + "content": " to quantize a given vector " + }, + { + "bbox": [ + 55, + 518, + 295, + 685 + ], + "type": "inline_equation", + "content": "\\mathbf{v}_i" + }, + { + "bbox": [ + 55, + 518, + 295, + 685 + ], + "type": "text", + "content": " where " + }, + { + "bbox": [ + 55, + 518, + 295, + 685 + ], + "type": "inline_equation", + "content": "j_{ik}" + }, + { + "bbox": [ + 55, + 518, + 295, + 685 + ], + "type": "text", + "content": " is an 8-bit unsigned integer." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 55, + 689, + 296, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 689, + 296, + 713 + ], + "spans": [ + { + "bbox": [ + 55, + 689, + 296, + 713 + ], + "type": "text", + "content": "Then, we can measure the distance between the query and data by adding distances between coarse centroids. Once the" + } + ] + } + ], + "index": 11 + }, + { + "type": "image", + "bbox": [ + 322, + 452, + 395, + 606 + ], + "blocks": [ + { + "bbox": [ + 322, + 452, + 395, + 606 + ], + "lines": [ + { + "bbox": [ + 322, + 452, + 395, + 606 + ], + "spans": [ + { + "bbox": [ + 322, + 452, + 395, + 606 + ], + "type": "image", + "image_path": "0e0078a6532ba08089c97e86c554134716d13da880e39ddeae2e1e7291cd23ce.jpg" + } + ] + } + ], + "index": 12, + "angle": 0, + "type": "image_body" + } + ], + "index": 12 + }, + { + "type": "image", + "bbox": [ + 395, + 452, + 468, + 611 + ], + "blocks": [ + { + "bbox": [ + 395, + 452, + 468, + 611 + ], + "lines": [ + { + "bbox": [ + 395, + 452, + 468, + 611 + ], + "spans": [ + { + "bbox": [ + 395, + 452, + 468, + 611 + ], + "type": "image", + "image_path": "18a40acb4c8461f487d030403c978efc5baa11ca1b4f93f66726976cdcd8f91a.jpg" + } + ] + } + ], + "index": 13, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 313, + 613, + 555, + 667 + ], + "lines": [ + { + "bbox": [ + 313, + 613, + 555, + 667 + ], + "spans": [ + { + "bbox": [ + 313, + 613, + 555, + 667 + ], + "type": "text", + "content": "Figure 6. Limitations of point-based IoU measurement. This figure shows the effect of removing the top and bottom " + }, + { + "bbox": [ + 313, + 613, + 555, + 667 + ], + "type": "inline_equation", + "content": "30\\%" + }, + { + "bbox": [ + 313, + 613, + 555, + 667 + ], + "type": "text", + "content": " of Gaussians according to the proposed significant score, implying that volume differences significantly impact 3D accuracy. The results highlight the need for the proposed IoU metric for 3D Gaussians." + } + ] + } + ], + "index": 15, + "angle": 0, + "type": "image_caption" + } + ], + "index": 13 + }, + { + "type": "image", + "bbox": [ + 470, + 453, + 541, + 611 + ], + "blocks": [ + { + "bbox": [ + 470, + 453, + 541, + 611 + ], + "lines": [ + { + "bbox": [ + 470, + 453, + 541, + 611 + ], + "spans": [ + { + "bbox": [ + 470, + 453, + 541, + 611 + ], + "type": "image", + "image_path": "01353414c5ff054c5b8723a00c26086fb62fc46477629e77e814574b061054fe.jpg" + } + ] + } + ], + "index": 14, + "angle": 0, + "type": "image_body" + } + ], + "index": 14 + }, + { + "bbox": [ + 313, + 677, + 556, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 677, + 556, + 715 + ], + "spans": [ + { + "bbox": [ + 313, + 677, + 556, + 715 + ], + "type": "text", + "content": "distances between centroids are computed as a lookup table, the computation shifts to simple indexing, which reduces the search complexity from " + }, + { + "bbox": [ + 313, + 677, + 556, + 715 + ], + "type": "inline_equation", + "content": "\\mathcal{O}(D)" + }, + { + "bbox": [ + 313, + 677, + 556, + 715 + ], + "type": "text", + "content": " to " + }, + { + "bbox": [ + 313, + 677, + 556, + 715 + ], + "type": "inline_equation", + "content": "\\mathcal{O}(1)" + }, + { + "bbox": [ + 313, + 677, + 556, + 715 + ], + "type": "text", + "content": " for a " + }, + { + "bbox": [ + 313, + 677, + 556, + 715 + ], + "type": "inline_equation", + "content": "D" + }, + { + "bbox": [ + 313, + 677, + 556, + 715 + ], + "type": "text", + "content": " dimen" + } + ] + } + ], + "index": 16 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "text", + "content": "14141" + } + ] + } + ], + "index": 17 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 4 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 62, + 70, + 294, + 145 + ], + "blocks": [ + { + "bbox": [ + 62, + 70, + 294, + 145 + ], + "lines": [ + { + "bbox": [ + 62, + 70, + 294, + 145 + ], + "spans": [ + { + "bbox": [ + 62, + 70, + 294, + 145 + ], + "type": "table", + "html": "
3D\nmIoUIoU > 0.1519 classes\nIoU > 0.3IoU > 0.45
LangSplat-m [30]8.017.17.82.9
LEGaussians-m [34]9.519.18.97.3
OpenGaussian [37]25.259.538.018.3
Ours (Top-20)25.060.740.320.0
Ours (Top-40)25.460.740.325.6
", + "image_path": "6e943fd223922385345560d8d15ca6e8f9ab4f2851957f6a7005e8e90a26eb4d.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "table_body" + } + ], + "index": 0 + }, + { + "type": "table", + "bbox": [ + 302, + 70, + 547, + 145 + ], + "blocks": [ + { + "bbox": [ + 127, + 146, + 227, + 156 + ], + "lines": [ + { + "bbox": [ + 127, + 146, + 227, + 156 + ], + "spans": [ + { + "bbox": [ + 127, + 146, + 227, + 156 + ], + "type": "text", + "content": "(a) 3D object localization task." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 302, + 70, + 547, + 145 + ], + "lines": [ + { + "bbox": [ + 302, + 70, + 547, + 145 + ], + "spans": [ + { + "bbox": [ + 302, + 70, + 547, + 145 + ], + "type": "table", + "html": "
19 classes15 classes10 classes
mIoUmAcc.mIoUmAcc.mIoUmAcc.
LangSplat-m [30]2.09.24.914.68.023.9
LEGaussians-m [34]1.67.94.616.17.724.9
OpenGaussian [37]30.146.538.156.849.771.4
Ours (Top-20)28.044.638.260.447.268.9
Ours (Top-40)29.647.738.260.450.273.5
", + "image_path": "f75d749f04c7944c007e27d41792d902ea94081d3613d217b6a82c927f29a9c4.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "table_body" + } + ], + "index": 2 + }, + { + "bbox": [ + 337, + 146, + 509, + 156 + ], + "lines": [ + { + "bbox": [ + 337, + 146, + 509, + 156 + ], + "spans": [ + { + "bbox": [ + 337, + 146, + 509, + 156 + ], + "type": "text", + "content": "(b) Open-vocabulary 3D semantic segmentation task." + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "text" + }, + { + "bbox": [ + 55, + 159, + 555, + 194 + ], + "lines": [ + { + "bbox": [ + 55, + 159, + 555, + 194 + ], + "spans": [ + { + "bbox": [ + 55, + 159, + 555, + 194 + ], + "type": "text", + "content": "Table 2. Quantitative comparison in the ScanNet dataset [4]. Left: Localization prediction is defined as 3D regions with a text similarity score above threshold. Right: We assign segmentation labels by finding max activations among all classes. Note that Bold and Underline stand for first and second best performance, respectively." + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "text" + }, + { + "type": "image", + "bbox": [ + 67, + 201, + 541, + 311 + ], + "blocks": [ + { + "bbox": [ + 67, + 201, + 541, + 311 + ], + "lines": [ + { + "bbox": [ + 67, + 201, + 541, + 311 + ], + "spans": [ + { + "bbox": [ + 67, + 201, + 541, + 311 + ], + "type": "image", + "image_path": "e7360054e021d98a71a00ff79e35a322409c81ec51cf9f1ae80953595a74b077.jpg" + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 55, + 313, + 555, + 347 + ], + "lines": [ + { + "bbox": [ + 55, + 313, + 555, + 347 + ], + "spans": [ + { + "bbox": [ + 55, + 313, + 555, + 347 + ], + "type": "text", + "content": "Figure 7. Qualitative results of 3D object localization. We visualize 3D localization activations (yellow) for \"chair\" and \"desk\" in the ScanNet dataset, comparing our method with others. It turns out that LangSplat-m and LEGaussians-m fail to localize objects accurately, while OpenGaussian struggles with object correspondence. Our model delivers precise and consistent localization across diverse queries." + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "image_caption" + } + ], + "index": 5 + }, + { + "bbox": [ + 55, + 367, + 295, + 391 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 367, + 295, + 391 + ], + "spans": [ + { + "bbox": [ + 55, + 367, + 295, + 391 + ], + "type": "text", + "content": "sion sample. This approach notably reduces computational complexity, making it suitable for large-scale search." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 54, + 393, + 297, + 610 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 54, + 393, + 297, + 610 + ], + "spans": [ + { + "bbox": [ + 54, + 393, + 297, + 610 + ], + "type": "text", + "content": "In our setup for language-based 3D scene understanding, we build PQ centroids based on CLIP embeddings using a large-scale image dataset, the LVIS dataset [10], that contains over 1.2M instances covering various long-tail classes and ground truth segmentation. We extract instance patches from images and collect patch-wise CLIP embeddings. After we build this CLIP embedding database, we proceed with the construction of the centroid codebook for our PQ. Once PQ is trained, any query embedding can be approximated by assigning the closest centroid for each subvector. This is a one-time procedure; once we determine the codebook, we can use it for any scene generally. In our setup, each embedding is represented as a sequence of centroid indices rather than a high-dimensional vector. Accordingly, our language embedded Gaussians are parametrized as " + }, + { + "bbox": [ + 54, + 393, + 297, + 610 + ], + "type": "inline_equation", + "content": "\\Phi^{\\mathrm{ours}} = \\{\\phi_i^{\\mathrm{ours}}\\}_{i=1}^N = \\{\\theta_i, j_i\\}_{i=1}^N" + }, + { + "bbox": [ + 54, + 393, + 297, + 610 + ], + "type": "text", + "content": ". where the aggregated feature " + }, + { + "bbox": [ + 54, + 393, + 297, + 610 + ], + "type": "inline_equation", + "content": "\\hat{\\mathbf{f}}_i" + }, + { + "bbox": [ + 54, + 393, + 297, + 610 + ], + "type": "text", + "content": " are converted as a quantized feature " + }, + { + "bbox": [ + 54, + 393, + 297, + 610 + ], + "type": "inline_equation", + "content": "\\bar{\\mathbf{f}}_i" + }, + { + "bbox": [ + 54, + 393, + 297, + 610 + ], + "type": "text", + "content": " by the corresponding PQ index " + }, + { + "bbox": [ + 54, + 393, + 297, + 610 + ], + "type": "inline_equation", + "content": "j_i" + }, + { + "bbox": [ + 54, + 393, + 297, + 610 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 55, + 622, + 232, + 635 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 622, + 232, + 635 + ], + "spans": [ + { + "bbox": [ + 55, + 622, + 232, + 635 + ], + "type": "text", + "content": "3.3. Text-query based 3D localization" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 55, + 641, + 297, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 641, + 297, + 714 + ], + "spans": [ + { + "bbox": [ + 55, + 641, + 297, + 714 + ], + "type": "text", + "content": "After training 3D Gaussians " + }, + { + "bbox": [ + 55, + 641, + 297, + 714 + ], + "type": "inline_equation", + "content": "\\Phi^{\\mathrm{ours}}" + }, + { + "bbox": [ + 55, + 641, + 297, + 714 + ], + "type": "text", + "content": " with our feature registration process and PQ, we describe the details of an inference mode that facilitates direct interaction with 3DGS upon receiving input queries, such as text. This is related to similarity score computation between a query and sources, i.e. Gaussian embeddings. Given a text, we first extract a" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 313, + 367, + 555, + 415 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 367, + 555, + 415 + ], + "spans": [ + { + "bbox": [ + 313, + 367, + 555, + 415 + ], + "type": "text", + "content": "query feature " + }, + { + "bbox": [ + 313, + 367, + 555, + 415 + ], + "type": "inline_equation", + "content": "\\mathbf{q}" + }, + { + "bbox": [ + 313, + 367, + 555, + 415 + ], + "type": "text", + "content": " using CLIP text encoder [31]. We reconstruct the quantized features " + }, + { + "bbox": [ + 313, + 367, + 555, + 415 + ], + "type": "inline_equation", + "content": "\\{\\bar{\\mathbf{f}}_i\\}_{i=1}^N" + }, + { + "bbox": [ + 313, + 367, + 555, + 415 + ], + "type": "text", + "content": " from the stored PQ indices " + }, + { + "bbox": [ + 313, + 367, + 555, + 415 + ], + "type": "inline_equation", + "content": "\\{j_i\\}_{i=1}^N" + }, + { + "bbox": [ + 313, + 367, + 555, + 415 + ], + "type": "text", + "content": ". Then, we compute a cosine similarity score between the query feature " + }, + { + "bbox": [ + 313, + 367, + 555, + 415 + ], + "type": "inline_equation", + "content": "\\mathbf{q}" + }, + { + "bbox": [ + 313, + 367, + 555, + 415 + ], + "type": "text", + "content": " and all quantized features." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 313, + 415, + 556, + 586 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 415, + 556, + 586 + ], + "spans": [ + { + "bbox": [ + 313, + 415, + 556, + 586 + ], + "type": "text", + "content": "Despite its simplicity, solely relying on the cosine similarity may result in diminished discriminability across certain similarity scores. To address this limitation, we incorporate a re-ranking process based on relative activation with respect to the canonical feature. For this process, we adopt the relevancy scoring method proposed in LeRF [17], which enables more precise similarity analysis for a query. Specifically, each rendered language embedding, " + }, + { + "bbox": [ + 313, + 415, + 556, + 586 + ], + "type": "inline_equation", + "content": "\\mathbf{f}^{\\mathrm{map}}" + }, + { + "bbox": [ + 313, + 415, + 556, + 586 + ], + "type": "text", + "content": " and a text query feature " + }, + { + "bbox": [ + 313, + 415, + 556, + 586 + ], + "type": "inline_equation", + "content": "\\mathbf{q}" + }, + { + "bbox": [ + 313, + 415, + 556, + 586 + ], + "type": "text", + "content": ", yield a relevance score determined by " + }, + { + "bbox": [ + 313, + 415, + 556, + 586 + ], + "type": "inline_equation", + "content": "\\min_{i}\\frac{\\exp(\\mathbf{f}^{\\mathrm{map}}\\cdot\\mathbf{q})}{\\exp(\\mathbf{f}^{\\mathrm{map}}\\cdot\\mathbf{q}) + \\exp(\\mathbf{f}^{\\mathrm{map}}\\cdot\\mathbf{f}^{\\mathrm{canon,i}})}" + }, + { + "bbox": [ + 313, + 415, + 556, + 586 + ], + "type": "text", + "content": ", where " + }, + { + "bbox": [ + 313, + 415, + 556, + 586 + ], + "type": "inline_equation", + "content": "(\\cdot)" + }, + { + "bbox": [ + 313, + 415, + 556, + 586 + ], + "type": "text", + "content": " is an element-wise dot product operator and " + }, + { + "bbox": [ + 313, + 415, + 556, + 586 + ], + "type": "inline_equation", + "content": "\\mathbf{f}^{\\mathrm{canon,i}}" + }, + { + "bbox": [ + 313, + 415, + 556, + 586 + ], + "type": "text", + "content": " indicates CLIP embeddings of a designated canonical term selected from a set of \"object,\" \"things,\" \"stuff,\" and \"texture\". Then, we sample 3D Gaussians based on the relevance score for downstream tasks." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 313, + 597, + 395, + 611 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 597, + 395, + 611 + ], + "spans": [ + { + "bbox": [ + 313, + 597, + 395, + 611 + ], + "type": "text", + "content": "4. Experiments" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 313, + 617, + 556, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 617, + 556, + 715 + ], + "spans": [ + { + "bbox": [ + 313, + 617, + 556, + 715 + ], + "type": "text", + "content": "Dataset. We use two datasets to evaluate the 3D scene understanding performance. For the 3D object selection task (Sec. 4.1), we use the LERF [17] dataset annotated by LangSplat [30], which consists of several multi-view images of 3D scenes containing long-tail objects and includes ground truth 2D ground truth annotations for texture queries. For 3D object localization Sec. 4.2 and 3D semantic segmentation Sec. 4.3 task, we employ the ScanNet [4] dataset." + } + ] + } + ], + "index": 14 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "type": "text", + "content": "14142" + } + ] + } + ], + "index": 15 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 5 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 66, + 70, + 541, + 181 + ], + "blocks": [ + { + "bbox": [ + 66, + 70, + 541, + 181 + ], + "lines": [ + { + "bbox": [ + 66, + 70, + 541, + 181 + ], + "spans": [ + { + "bbox": [ + 66, + 70, + 541, + 181 + ], + "type": "image", + "image_path": "74bf00ac799dd062d56c084505c13cea6436e7b9a368cf80cb7eab240c38994b.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 55, + 182, + 555, + 217 + ], + "lines": [ + { + "bbox": [ + 55, + 182, + 555, + 217 + ], + "spans": [ + { + "bbox": [ + 55, + 182, + 555, + 217 + ], + "type": "text", + "content": "Figure 8. Visualization of open-vocabulary 3D semantic segmentation on the ScanNet dataset [4]. We visualize 3D Gaussian splat-based semantic segmentation using language features allocation of OpenGaussian [37] and Dr. Splat (ours) model on the same RGB-pretrained 3DGS. Note that, not specifically designed for segmentation, it achieves high performance as a result of language-based Gaussian updates." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "bbox": [ + 54, + 237, + 297, + 285 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 54, + 237, + 297, + 285 + ], + "spans": [ + { + "bbox": [ + 54, + 237, + 297, + 285 + ], + "type": "text", + "content": "ScanNet is a large-scale benchmark that provides data on indoor scenes, including calibrated RGBD images and 3D point clouds with ground-truth semantic labels. We randomly select eight scenes from ScanNet for the experiments." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 54, + 286, + 296, + 455 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 54, + 286, + 296, + 455 + ], + "spans": [ + { + "bbox": [ + 54, + 286, + 296, + 455 + ], + "type": "text", + "content": "Competing methods. The only method available for a fair comparison with our method is the concurrent work, Open-Gaussian [37]. To study the various aspects of our method, we introduce baseline methods modified from rasterization-based ones [30, 34], for direct 3D referring operation, denoted as LangSplat-m and LEGaussians-m. As discussed in Sec. 2, without modification, global search over a whole scene is quite demanding. To ensure fair evaluation, we use the same initial 3D Gaussians being trained only using RGB inputs for all comparing methods, and freeze the Gaussians during the language feature allocation process. Also, the per-pixel CLIP [31] embedding maps are unified for SAM-based [20] methods [30, 37] including ours. We follow the hyperparameter settings favorable to each respective paper." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 55, + 465, + 167, + 478 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 465, + 167, + 478 + ], + "spans": [ + { + "bbox": [ + 55, + 465, + 167, + 478 + ], + "type": "text", + "content": "4.1.3D object selection" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 54, + 483, + 297, + 579 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 54, + 483, + 297, + 579 + ], + "spans": [ + { + "bbox": [ + 54, + 483, + 297, + 579 + ], + "type": "text", + "content": "Settings. We first extract text features from an open-vocabulary text query using the CLIP model. Next, we compare text features to the 3D features embedded in each Gaussian using cosine similarity. By thresholding the similarity, we identify the 3D Gaussians that are relevant to the given text query. The selected 3D points are subsequently rendered into multi-view images using the 3DGS rasterization pipeline." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 54, + 580, + 297, + 664 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 54, + 580, + 297, + 664 + ], + "spans": [ + { + "bbox": [ + 54, + 580, + 297, + 664 + ], + "type": "text", + "content": "Results. We compare our model quantitatively with 3DGS-based language-embedded models as shown in Table 1. The results demonstrate that our method performs better object selection in most scenes, showing an improvement of over 0.5 in mIoU and more than 4.5 in mAcc compared to counterpart models. Notably, the rasterization-based method, LangSplat-m, often underperforms in most scenes." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 54, + 665, + 297, + 712 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 54, + 665, + 297, + 712 + ], + "spans": [ + { + "bbox": [ + 54, + 665, + 297, + 712 + ], + "type": "text", + "content": "Qualitative results are shown in Fig. 5. For LangSplat-m, the activations often shows random 3D Gaussians or fail to localize entirely (e.g., see \"coffee mug\"), highlighting the limitations of rasterization-based methods and their un" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 313, + 237, + 555, + 321 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 237, + 555, + 321 + ], + "spans": [ + { + "bbox": [ + 313, + 237, + 555, + 321 + ], + "type": "text", + "content": "suitability for 3D understanding, aligning the observation from Fig. 2. OpenGaussian frequently exhibits false activations with incorrect text-object pairs (e.g., \"apple\" and \"tea in a glass\") and struggles to distinguish between nearby objects (e.g., \"waldo,\" \"rubik's cube\"). This artifacts can be attributed to use of spatial clustering and limited encoder capacity." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 313, + 322, + 555, + 382 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 322, + 555, + 382 + ], + "spans": [ + { + "bbox": [ + 313, + 322, + 555, + 382 + ], + "type": "text", + "content": "In contrast, our model leverages general image features thanks to the general PQ, maintaining feature distinctiveness regardless of scene complexity. Our feature registration considers the 3D geometry of the 3D Gaussians, which results in superior performance in 3D scene understanding tasks." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 313, + 393, + 440, + 406 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 393, + 440, + 406 + ], + "spans": [ + { + "bbox": [ + 313, + 393, + 440, + 406 + ], + "type": "text", + "content": "4.2. 3D object localization" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 313, + 411, + 556, + 518 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 411, + 556, + 518 + ], + "spans": [ + { + "bbox": [ + 313, + 411, + 556, + 518 + ], + "type": "text", + "content": "Settings. Similar to the 3D object selection task, we calculate the cosine similarity between text query and 3D features embedded in each Gaussian. By thresholding the similarity, we identify the 3D Gaussians relevant to the given text query. To measure volume-aware localization evaluation, we propose a protocol to measure the IoU of 3D Gaussians that expands the traditional metric of point cloud-based approaches by incorporating volumetric information of 3D Gaussians." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 313, + 521, + 556, + 616 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 521, + 556, + 616 + ], + "spans": [ + { + "bbox": [ + 313, + 521, + 556, + 616 + ], + "type": "text", + "content": "Novel evaluation protocol for 3D localization in 3DGS. Unlike conventional evaluation protocol for the 3D localization task in point clouds, it is tricky to evaluate 3D localization performance in 3D Gaussians [16]. This is primarily due to the un-deterministic structure of Gaussian distribution. To address this issue, we compute 3DGS pseudo-labels for evaluating the 3DGS localization in a volume-aware way. The details can be found in the supplementary material." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 313, + 618, + 557, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 618, + 557, + 714 + ], + "spans": [ + { + "bbox": [ + 313, + 618, + 557, + 714 + ], + "type": "text", + "content": "Given the ground truth, we measure IoU considering the spatial significance of each Gaussian and define a significant score " + }, + { + "bbox": [ + 313, + 618, + 557, + 714 + ], + "type": "inline_equation", + "content": "d_{i}" + }, + { + "bbox": [ + 313, + 618, + 557, + 714 + ], + "type": "text", + "content": " for each Gaussian " + }, + { + "bbox": [ + 313, + 618, + 557, + 714 + ], + "type": "inline_equation", + "content": "\\theta_{i}" + }, + { + "bbox": [ + 313, + 618, + 557, + 714 + ], + "type": "text", + "content": " with its scale " + }, + { + "bbox": [ + 313, + 618, + 557, + 714 + ], + "type": "inline_equation", + "content": "\\mathbf{s}_i = [s_{ix}, s_{iy}, s_{iy}]" + }, + { + "bbox": [ + 313, + 618, + 557, + 714 + ], + "type": "text", + "content": " and opacity " + }, + { + "bbox": [ + 313, + 618, + 557, + 714 + ], + "type": "inline_equation", + "content": "\\alpha_{i}" + }, + { + "bbox": [ + 313, + 618, + 557, + 714 + ], + "type": "text", + "content": " as " + }, + { + "bbox": [ + 313, + 618, + 557, + 714 + ], + "type": "inline_equation", + "content": "d_{i} = s_{ix}s_{iy}s_{iz}\\alpha_{i}" + }, + { + "bbox": [ + 313, + 618, + 557, + 714 + ], + "type": "text", + "content": ", where " + }, + { + "bbox": [ + 313, + 618, + 557, + 714 + ], + "type": "inline_equation", + "content": "s_{ix}s_{iy}s_{iz}" + }, + { + "bbox": [ + 313, + 618, + 557, + 714 + ], + "type": "text", + "content": " denotes a relative ellipsoid volume of a Gaussian " + }, + { + "bbox": [ + 313, + 618, + 557, + 714 + ], + "type": "inline_equation", + "content": "\\theta_{i}" + }, + { + "bbox": [ + 313, + 618, + 557, + 714 + ], + "type": "text", + "content": ". With the obtained significant scores " + }, + { + "bbox": [ + 313, + 618, + 557, + 714 + ], + "type": "inline_equation", + "content": "\\mathbf{d} = [d_1, d_2, \\dots, d_N]" + }, + { + "bbox": [ + 313, + 618, + 557, + 714 + ], + "type": "text", + "content": ", we compute weighted IoU of 3D Gaussians to approximate volumes. The proposed metric is designed to assign a larger" + } + ] + } + ], + "index": 13 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "type": "text", + "content": "14143" + } + ] + } + ], + "index": 14 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 6 + }, + { + "para_blocks": [ + { + "bbox": [ + 55, + 72, + 296, + 143 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 72, + 296, + 143 + ], + "spans": [ + { + "bbox": [ + 55, + 72, + 296, + 143 + ], + "type": "text", + "content": "weight to the Gaussians with higher significant scores, when measuring IoU. Figure 6 shows that the impact of each Gaussian on the scene extremely varies depending on their significant scores, which demonstrates the necessity of the proposed IoU metric on 3D Gaussians that regards unequal contributions of each Gaussian." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 55, + 144, + 296, + 276 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 144, + 296, + 276 + ], + "spans": [ + { + "bbox": [ + 55, + 144, + 296, + 276 + ], + "type": "text", + "content": "Results. We report the 3D localization performance on the Scannet dataset in Table 2a. The 2D rasterization-based methods [30, 34] struggle to achieve precise activations for 3D localization. They inherently face challenges when applying for 3D tasks because they need to render 2D images for the scene interaction. Even with the 3D space search method, OpenGaussian [37], our model consistently demonstrates superior performance and achieves higher accuracy in localization. Figure 7 also shows that LangSplat-m and LEGaussians-m fail to properly localize the objects, and OpenGaussian misses queried objects in the scene." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 55, + 285, + 201, + 298 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 285, + 201, + 298 + ], + "spans": [ + { + "bbox": [ + 55, + 285, + 201, + 298 + ], + "type": "text", + "content": "4.3. 3D semantic segmentation" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 55, + 303, + 296, + 339 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 303, + 296, + 339 + ], + "spans": [ + { + "bbox": [ + 55, + 303, + 296, + 339 + ], + "type": "text", + "content": "Settings. For a given set of open-vocabulary text labels, we perform segmentation by assigning each Gaussian a label having the highest activation among the known label set." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 55, + 340, + 296, + 471 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 340, + 296, + 471 + ], + "spans": [ + { + "bbox": [ + 55, + 340, + 296, + 471 + ], + "type": "text", + "content": "Results. The numerical comparison is presented in Table 2b. Although not explicitly designed for semantic segmentation, our model achieves notable performance in this task as a result of accurately updating each Gaussian with language features. Consistent with previous observations, rasterization-based 3DGS models exhibit lower segmentation performance. While OpenGaussian performs position-based clustering, our model demonstrates comparable performance, surpassing the baseline as the Top-" + }, + { + "bbox": [ + 55, + 340, + 296, + 471 + ], + "type": "inline_equation", + "content": "k" + }, + { + "bbox": [ + 55, + 340, + 296, + 471 + ], + "type": "text", + "content": " value increases. Our model also achieves better segmentation results, with a visual comparison of the segmented scene shown in Fig. 8." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 55, + 480, + 147, + 493 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 480, + 147, + 493 + ], + "spans": [ + { + "bbox": [ + 55, + 480, + 147, + 493 + ], + "type": "text", + "content": "4.4. Ablation study" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 55, + 498, + 295, + 533 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 498, + 295, + 533 + ], + "spans": [ + { + "bbox": [ + 55, + 498, + 295, + 533 + ], + "type": "text", + "content": "We conduct an ablation study using the ScanNet dataset on different hyper-parameters of Dr. Splat to measure the contribution of each component." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 55, + 534, + 296, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 534, + 296, + 713 + ], + "spans": [ + { + "bbox": [ + 55, + 534, + 296, + 713 + ], + "type": "text", + "content": "Product Quantization. PQ introduces a trade-off between memory usage, computational efficiency, and accuracy. To better understand the balance between computational cost and localization quality, we conduct an ablation study by varying the number of sub-vectors. We evaluate performance at sub-vector sizes of 64, 128, and 256. Notably, these settings correspond to bit-size reductions of 1/32, 1/16, and 1/8 of the original CLIP feature, respectively. We measure the query distance computation time for one million data points, averaging results over 100 iterations for efficiency measure. Our findings reveal a favorable trade-off between quantization performance and accuracy (see Fig. 9-(b)) in the Pareto front with our PQ configurations. This achieves a balance that maximizes memory and computational efficiency while minimizing any loss in accuracy." + } + ] + } + ], + "index": 7 + }, + { + "type": "image", + "bbox": [ + 320, + 71, + 432, + 141 + ], + "blocks": [ + { + "bbox": [ + 320, + 71, + 432, + 141 + ], + "lines": [ + { + "bbox": [ + 320, + 71, + 432, + 141 + ], + "spans": [ + { + "bbox": [ + 320, + 71, + 432, + 141 + ], + "type": "image", + "image_path": "976c9798fa322047f804bddbd485f7e83bc57c08b9c0475f944ca46d55412b45.jpg" + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 323, + 144, + 424, + 153 + ], + "lines": [ + { + "bbox": [ + 323, + 144, + 424, + 153 + ], + "spans": [ + { + "bbox": [ + 323, + 144, + 424, + 153 + ], + "type": "text", + "content": "(a) Ablation on PQ parameters." + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "image_caption" + } + ], + "index": 8 + }, + { + "type": "image", + "bbox": [ + 437, + 71, + 550, + 140 + ], + "blocks": [ + { + "bbox": [ + 437, + 71, + 550, + 140 + ], + "lines": [ + { + "bbox": [ + 437, + 71, + 550, + 140 + ], + "spans": [ + { + "bbox": [ + 437, + 71, + 550, + 140 + ], + "type": "image", + "image_path": "00be046b6c22477fe8a63532227a14e3f6fde73bd86141153e372d00f2e190a9.jpg" + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 455, + 144, + 529, + 153 + ], + "lines": [ + { + "bbox": [ + 455, + 144, + 529, + 153 + ], + "spans": [ + { + "bbox": [ + 455, + 144, + 529, + 153 + ], + "type": "text", + "content": "(b) Ablation on Top- " + }, + { + "bbox": [ + 455, + 144, + 529, + 153 + ], + "type": "inline_equation", + "content": "k" + } + ] + } + ], + "index": 11, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 323, + 158, + 544, + 169 + ], + "lines": [ + { + "bbox": [ + 323, + 158, + 544, + 169 + ], + "spans": [ + { + "bbox": [ + 323, + 158, + 544, + 169 + ], + "type": "text", + "content": "Figure 9. Ablation study on (a) PQ and (b) Top- " + }, + { + "bbox": [ + 323, + 158, + 544, + 169 + ], + "type": "inline_equation", + "content": "k" + }, + { + "bbox": [ + 323, + 158, + 544, + 169 + ], + "type": "text", + "content": " Gaussians." + } + ] + } + ], + "index": 12, + "angle": 0, + "type": "image_caption" + } + ], + "index": 10 + }, + { + "bbox": [ + 313, + 191, + 555, + 299 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 191, + 555, + 299 + ], + "spans": [ + { + "bbox": [ + 313, + 191, + 555, + 299 + ], + "type": "text", + "content": "Top-" + }, + { + "bbox": [ + 313, + 191, + 555, + 299 + ], + "type": "inline_equation", + "content": "k" + }, + { + "bbox": [ + 313, + 191, + 555, + 299 + ], + "type": "text", + "content": " Gaussians. We examine the influence of the number of Gaussians assigned per ray. This parameter affects both memory requirements and computation, serving as a critical factor in overall performance. The ratio of pruned Gaussians and the mIoU results from different " + }, + { + "bbox": [ + 313, + 191, + 555, + 299 + ], + "type": "inline_equation", + "content": "k" + }, + { + "bbox": [ + 313, + 191, + 555, + 299 + ], + "type": "text", + "content": " are presented in Fig. 9a. We observe that increasing the aggregating number of Gaussians per ray improves localization performance; however, it results in higher memory consumption and the number of occupied Gaussians, indicating a clear trade-off." + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 313, + 312, + 466, + 324 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 312, + 466, + 324 + ], + "spans": [ + { + "bbox": [ + 313, + 312, + 466, + 324 + ], + "type": "text", + "content": "5. Discussion and Conclusion" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 313, + 332, + 555, + 440 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 332, + 555, + 440 + ], + "spans": [ + { + "bbox": [ + 313, + 332, + 555, + 440 + ], + "type": "text", + "content": "We present Dr. Splat, which is a novel approach for open-vocabulary 3D scene understanding by directly registering language embeddings to 3D Gaussians, eliminating the need for an intermediate rendering process. Compared to the previous 2D rendering-based methods [30, 34], which have limited search domain and capacity, our method directly searches 3D space while preserving the fidelity of language embeddings. This operation is further accelerated by the integration of Product Quantization (PQ)" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 313, + 441, + 556, + 548 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 441, + 556, + 548 + ], + "spans": [ + { + "bbox": [ + 313, + 441, + 556, + 548 + ], + "type": "text", + "content": "Experimental results validate Dr. Splat's superior performance across various 3D scene understanding tasks, including open-vocabulary 3D object selection, 3D object localization, and 3D semantic segmentation. These findings highlight Dr. Splat's ability to transform 3D scene understanding by achieving a balance between highly representative quality and computational efficiency. This breakthrough paves the way for advanced applications in robotics, autonomous navigation, and augmented reality." + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 314, + 562, + 406, + 574 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 314, + 562, + 406, + 574 + ], + "spans": [ + { + "bbox": [ + 314, + 562, + 406, + 574 + ], + "type": "text", + "content": "Acknowledgment" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 313, + 582, + 556, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 582, + 556, + 713 + ], + "spans": [ + { + "bbox": [ + 313, + 582, + 556, + 713 + ], + "type": "text", + "content": "This work was supported by Institute of Information & communications Technology Planning & Evaluation (IITP) (No.RS-2021-II212068, Artificial Intelligence Innovation Hub, No.RS-2024-00457882, National AI Research Lab Project, No.RS-2023-00225630, Development of Artificial Intelligence for Text-based 3D Movie Generation, No.RS-2019-II191906, Artificial Intelligence Graduate School Program(POSTECH)) and by the National Research Foundation of Korea (NRF) (No.RS-2024-00358135, Corner Vision: Learning to Look Around the Corner through Multi-modal Signals) grant funded by the Korea government (MSIT)." + } + ] + } + ], + "index": 18 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "text", + "content": "14144" + } + ] + } + ], + "index": 19 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 7 + }, + { + "para_blocks": [ + { + "bbox": [ + 56, + 71, + 115, + 83 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 71, + 115, + 83 + ], + "spans": [ + { + "bbox": [ + 56, + 71, + 115, + 83 + ], + "type": "text", + "content": "References" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 57, + 91, + 297, + 712 + ], + "type": "list", + "angle": 0, + "index": 14, + "blocks": [ + { + "bbox": [ + 61, + 91, + 297, + 145 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 61, + 91, + 297, + 145 + ], + "spans": [ + { + "bbox": [ + 61, + 91, + 297, + 145 + ], + "type": "text", + "content": "[1] Mathilde Caron, Hugo Touvron, Ishan Misra, Hervé Jégou, Julien Mairal, Piotr Bojanowski, and Armand Joulin. Emerging properties in self-supervised vision transformers. In Proceedings of the IEEE/CVF international conference on computer vision, pages 9650-9660, 2021. 2" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 61, + 147, + 296, + 190 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 61, + 147, + 296, + 190 + ], + "spans": [ + { + "bbox": [ + 61, + 147, + 296, + 190 + ], + "type": "text", + "content": "[2] Jaesung Choe, Chunghyun Park, Francois Rameau, Jaesik Park, and In So Kweon. Pointmixer: Mlp-mixer for point cloud understanding. In European Conference on Computer Vision, pages 620-640. Springer, 2022. 1" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 62, + 191, + 296, + 245 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 191, + 296, + 245 + ], + "spans": [ + { + "bbox": [ + 62, + 191, + 296, + 245 + ], + "type": "text", + "content": "[3] Christopher Choy, JunYoung Gwak, and Silvio Savarese. 4d spatio-temporal convnets: Minkowski convolutional neural networks. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 3075-3084, 2019. 1" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 62, + 247, + 296, + 290 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 247, + 296, + 290 + ], + "spans": [ + { + "bbox": [ + 62, + 247, + 296, + 290 + ], + "type": "text", + "content": "[4] Angela Dai, Angel X Chang, Manolis Savva, Maciej Halber, Thomas Funkhouser, and Matthias Nießner. Scannet: Richly-annotated 3d reconstructions of indoor scenes. In CVPR, pages 5828-5839, 2017. 6, 7" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 62, + 291, + 296, + 345 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 291, + 296, + 345 + ], + "spans": [ + { + "bbox": [ + 62, + 291, + 296, + 345 + ], + "type": "text", + "content": "[5] Runyu Ding, Jihan Yang, Chuhui Xue, Wenqing Zhang, Song Bai, and Xiaojuan Qi. Pla: Language-driven open-vocabulary 3d scene understanding. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 7010-7019, 2023. 2" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 62, + 346, + 296, + 399 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 346, + 296, + 399 + ], + "spans": [ + { + "bbox": [ + 62, + 346, + 296, + 399 + ], + "type": "text", + "content": "[6] Francis Engelmann, Fabian Manhardt, Michael Niemeyer, Keisuke Tateno, Marc Pollefeys, and Federico Tombari. OpenNerf: Open Set 3D Neural Scene Segmentation with PixelWise Features and Rendered Novel Views. In ICLR, 2024. 2" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 62, + 402, + 296, + 435 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 402, + 296, + 435 + ], + "spans": [ + { + "bbox": [ + 62, + 402, + 296, + 435 + ], + "type": "text", + "content": "[7] Michael R. Garey and David S. Johnson. Computers and intractability. a guide to the theory of np-completeness. W. H. Freeman and company, 174, 1979. 3" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 62, + 436, + 296, + 468 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 436, + 296, + 468 + ], + "spans": [ + { + "bbox": [ + 62, + 436, + 296, + 468 + ], + "type": "text", + "content": "[8] Antoine Guédon, Tom Monnier, Pascal Monasse, and Vincent Lepetit. Macarons: Mapping and coverage anticipation with rgb online self-supervision. In CVPR, 2023. 3" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 62, + 470, + 296, + 512 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 470, + 296, + 512 + ], + "spans": [ + { + "bbox": [ + 62, + 470, + 296, + 512 + ], + "type": "text", + "content": "[9] Jun Guo, Xiaojian Ma, Yue Fan, Huaping Liu, and Qing Li. Semantic gaussians: Open-vocabulary scene understanding with 3d gaussian splatting. arXiv preprint arXiv:2403.15624, 2024. 2" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 57, + 514, + 295, + 557 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 514, + 295, + 557 + ], + "spans": [ + { + "bbox": [ + 57, + 514, + 295, + 557 + ], + "type": "text", + "content": "[10] Agrim Gupta, Piotr Dollar, and Ross Girshick. Lvis: A dataset for large vocabulary instance segmentation. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 5356-5364, 2019. 6" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 57, + 558, + 296, + 612 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 558, + 296, + 612 + ], + "spans": [ + { + "bbox": [ + 57, + 558, + 296, + 612 + ], + "type": "text", + "content": "[11] Zhening Huang, Xiaoyang Wu, Xi Chen, Hengshuang Zhao, Lei Zhu, and Joan Lasenby. Openins3d: Snap and lookup for 3d open-vocabulary instance segmentation. In European Conference on Computer Vision, pages 169-185. Springer, 2025. 1, 3" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 57, + 614, + 296, + 668 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 614, + 296, + 668 + ], + "spans": [ + { + "bbox": [ + 57, + 614, + 296, + 668 + ], + "type": "text", + "content": "[12] Krishna Murthy Jatavallabhula, Alihusein Kuwajerwala, Qiao Gu, Mohd Omama, Tao Chen, Alaa Maalouf, Shuang Li, Ganesh Iyer, Soroush Saryazdi, Nikhil Keetha, et al. Conceptfusion: Open-set multimodal 3d mapping. arXiv preprint arXiv:2302.07241, 2023. 2" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 57, + 670, + 296, + 712 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 670, + 296, + 712 + ], + "spans": [ + { + "bbox": [ + 57, + 670, + 296, + 712 + ], + "type": "text", + "content": "[13] Herve Jegou, Matthijs Douze, and Cordelia Schmid. Product quantization for nearest neighbor search. IEEE transactions on pattern analysis and machine intelligence, 33(1):117-128, 2010. 5" + } + ] + } + ], + "index": 13 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 316, + 73, + 555, + 713 + ], + "type": "list", + "angle": 0, + "index": 30, + "blocks": [ + { + "bbox": [ + 316, + 73, + 555, + 116 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 73, + 555, + 116 + ], + "spans": [ + { + "bbox": [ + 316, + 73, + 555, + 116 + ], + "type": "text", + "content": "[14] Yuzhou Ji, He Zhu, Junshu Tang, Wuyi Liu, Zhizhong Zhang, Yuan Xie, and Xin Tan. Fastlgs: Speeding up language embedded gaussians with feature grid mapping. arXiv preprint arXiv:2406.01916, 2024. 2, 3" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 317, + 118, + 555, + 161 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 118, + 555, + 161 + ], + "spans": [ + { + "bbox": [ + 317, + 118, + 555, + 161 + ], + "type": "text", + "content": "[15] Li Jiang, Shaoshuai Shi, and Bernt Schiele. Open-vocabulary 3d semantic segmentation with foundation models. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 21284-21294, 2024. 1, 2" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 317, + 162, + 553, + 194 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 162, + 553, + 194 + ], + "spans": [ + { + "bbox": [ + 317, + 162, + 553, + 194 + ], + "type": "text", + "content": "[16] Bernhard Kerbl, Georgios Kopanas, Thomas Leimkuhler, and George Drettakis. 3d gaussian splatting for real-time radiance field rendering. ACM TOG, 2023. 1, 2, 3, 7" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 316, + 195, + 555, + 248 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 195, + 555, + 248 + ], + "spans": [ + { + "bbox": [ + 316, + 195, + 555, + 248 + ], + "type": "text", + "content": "[17] Justin Kerr, Chung Min Kim, Ken Goldberg, Angjoo Kanazawa, and Matthew Tancik. Lerf: Language embedded radiance fields. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 19729-19739, 2023. 1, 2, 5, 6" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 316, + 250, + 555, + 282 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 250, + 555, + 282 + ], + "spans": [ + { + "bbox": [ + 316, + 250, + 555, + 282 + ], + "type": "text", + "content": "[18] GeonU Kim, Kim Youwang, and Tae-Hyun Oh. FPRF: Feedforward photorealistic style transfer of large-scale 3D neural radiance fields. In AAAI, 2024. 2" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 316, + 283, + 553, + 316 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 283, + 553, + 316 + ], + "spans": [ + { + "bbox": [ + 316, + 283, + 553, + 316 + ], + "type": "text", + "content": "[19] GeonU Kim, Kim Youwang, Lee Hyoseok, and Tae-Hyun Oh. Fpgs: Feed-forward semantic-aware photorealistic style transfer of large-scale gaussian splatting, 2025. 3" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 316, + 317, + 555, + 360 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 317, + 555, + 360 + ], + "spans": [ + { + "bbox": [ + 316, + 317, + 555, + 360 + ], + "type": "text", + "content": "[20] Alexander Kirillov, Eric Mintun, Nikhila Ravi, Hanzi Mao, Chloe Rolland, Laura Gustafson, Tete Xiao, Spencer Whitehead, Alexander C. Berg, Wan-Yen Lo, Piotr Dólar, and Ross B. Girshick. Segment anything. In ICCV, 2023. 4, 7" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 316, + 361, + 555, + 403 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 361, + 555, + 403 + ], + "spans": [ + { + "bbox": [ + 316, + 361, + 555, + 403 + ], + "type": "text", + "content": "[21] Sosuke Kobayashi, Eiichi Matsumoto, and Vincent Sitzmann. Decomposing nerf for editing via feature field distillation. Advances in Neural Information Processing Systems, 35:23311-23330, 2022. 2" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 316, + 404, + 553, + 437 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 404, + 553, + 437 + ], + "spans": [ + { + "bbox": [ + 316, + 404, + 553, + 437 + ], + "type": "text", + "content": "[22] Hyunjee Lee, Youngsik Yun, Jeongmin Bae, Seoha Kim, and Youngjung Uh. Rethinking open-vocabulary segmentation of radiance fields in 3d space, 2024. 2" + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 316, + 438, + 553, + 470 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 438, + 553, + 470 + ], + "spans": [ + { + "bbox": [ + 316, + 438, + 553, + 470 + ], + "type": "text", + "content": "[23] Boyi Li, Kilian Q Weinberger, Serge Belongie, Vladlen Koltun, and René Ranftl. Language-driven semantic segmentation. arXiv preprint arXiv:2201.03546, 2022. 2" + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 317, + 472, + 555, + 525 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 472, + 555, + 525 + ], + "spans": [ + { + "bbox": [ + 317, + 472, + 555, + 525 + ], + "type": "text", + "content": "[24] Kunhao Liu, Fangneng Zhan, Jiahui Zhang, Muyu Xu, Yingchen Yu, Abdulmotaleb El Saddik, Christian Theobalt, Eric Xing, and Shijian Lu. Weakly supervised 3d open-vocabulary segmentation. Advances in Neural Information Processing Systems, 36:53433-53456, 2023. 2" + } + ] + } + ], + "index": 25 + }, + { + "bbox": [ + 317, + 526, + 555, + 590 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 526, + 555, + 590 + ], + "spans": [ + { + "bbox": [ + 317, + 526, + 555, + 590 + ], + "type": "text", + "content": "[25] Minghua Liu, Yinhao Zhu, Hong Cai, Shizhong Han, Zhan Ling, Fatih Porikli, and Hao Su. Partslip: Low-shot part segmentation for 3d point clouds via pretrained image-language models. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 21736-21746, 2023. 2" + } + ] + } + ], + "index": 26 + }, + { + "bbox": [ + 317, + 592, + 555, + 635 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 592, + 555, + 635 + ], + "spans": [ + { + "bbox": [ + 317, + 592, + 555, + 635 + ], + "type": "text", + "content": "[26] Ben Mildenhall, Pratul P. Srinivasan, Matthew Tancik, Jonathan T. Barron, Ravi Ramamoorthi, and Ren Ng. Nerf: Representing scenes as neural radiance fields for view synthesis. In ECCV, 2020. 2" + } + ] + } + ], + "index": 27 + }, + { + "bbox": [ + 317, + 636, + 555, + 691 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 636, + 555, + 691 + ], + "spans": [ + { + "bbox": [ + 317, + 636, + 555, + 691 + ], + "type": "text", + "content": "[27] Songyou Peng, Kyle Genova, Chiyu Jiang, Andrea Tagliasacchi, Marc Pollefeys, Thomas Funkhouser, et al. Openscene: 3d scene understanding with open vocabularies. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 815-824, 2023. 1, 2, 4" + } + ] + } + ], + "index": 28 + }, + { + "bbox": [ + 316, + 692, + 553, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 692, + 553, + 713 + ], + "spans": [ + { + "bbox": [ + 316, + 692, + 553, + 713 + ], + "type": "text", + "content": "[28] Charles R Qi, Hao Su, Kaichun Mo, and Leonidas J Guibas. Pointnet: Deep learning on point sets for 3d classification" + } + ] + } + ], + "index": 29 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "text", + "content": "14145" + } + ] + } + ], + "index": 31 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 8 + }, + { + "para_blocks": [ + { + "bbox": [ + 56, + 72, + 296, + 713 + ], + "type": "list", + "angle": 0, + "index": 13, + "blocks": [ + { + "bbox": [ + 76, + 72, + 296, + 105 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 76, + 72, + 296, + 105 + ], + "spans": [ + { + "bbox": [ + 76, + 72, + 296, + 105 + ], + "type": "text", + "content": "and segmentation. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 652-660, 2017. 1" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 56, + 106, + 296, + 160 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 106, + 296, + 160 + ], + "spans": [ + { + "bbox": [ + 56, + 106, + 296, + 160 + ], + "type": "text", + "content": "[29] Guocheng Qian, Yuchen Li, Houwen Peng, Jinjie Mai, Hasan Hammoud, Mohamed Elhoseiny, and Bernard Ghanem. Pointnext: Revisiting pointnet++ with improved training and scaling strategies. Advances in neural information processing systems, 35:23192-23204, 2022. 1" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 56, + 161, + 296, + 215 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 161, + 296, + 215 + ], + "spans": [ + { + "bbox": [ + 56, + 161, + 296, + 215 + ], + "type": "text", + "content": "[30] Minghan Qin, Wanhua Li, Jiawei Zhou, Haoqian Wang, and Hanspeter Pfister. Langsplat: 3d language gaussian splatting. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 20051-20060, 2024. 1, 2, 3, 4, 5, 6, 7, 8" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 56, + 217, + 296, + 281 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 217, + 296, + 281 + ], + "spans": [ + { + "bbox": [ + 56, + 217, + 296, + 281 + ], + "type": "text", + "content": "[31] Alec Radford, Jong Wook Kim, Chris Hallacy, Aditya Ramesh, Gabriel Goh, Sandhini Agarwal, Girish Sastry, Amanda Askell, Pamela Mishkin, Jack Clark, et al. Learning transferable visual models from natural language supervision. In International conference on machine learning, pages 8748-8763. PMLR, 2021. 2, 3, 6, 7" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 56, + 282, + 296, + 337 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 282, + 296, + 337 + ], + "spans": [ + { + "bbox": [ + 56, + 282, + 296, + 337 + ], + "type": "text", + "content": "[32] Adam Rashid, Satvik Sharma, Chung Min Kim, Justin Kerr, Lawrence Yunliang Chen, Angjoo Kanazawa, and Ken Goldberg. Language embedded radiance fields for zero-shot task-oriented grasping. In 7th Annual Conference on Robot Learning, 2023. 2" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 56, + 338, + 296, + 381 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 338, + 296, + 381 + ], + "spans": [ + { + "bbox": [ + 56, + 338, + 296, + 381 + ], + "type": "text", + "content": "[33] Damien Robert, Hugo Raguet, and Loic Landrieu. Efficient 3d semantic segmentation with superpoint transformer. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 17195-17204, 2023. 1" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 56, + 382, + 296, + 435 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 382, + 296, + 435 + ], + "spans": [ + { + "bbox": [ + 56, + 382, + 296, + 435 + ], + "type": "text", + "content": "[34] Jin-Chuan Shi, Miao Wang, Hao-Bin Duan, and Shao-Hua Guan. Language embedded 3d gaussians for open-vocabulary scene understanding. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 5333-5343, 2024. 1, 2, 3, 4, 6, 7, 8" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 56, + 437, + 296, + 480 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 437, + 296, + 480 + ], + "spans": [ + { + "bbox": [ + 56, + 437, + 296, + 480 + ], + "type": "text", + "content": "[35] Ayca Takmaz, Elisabetta Fedele, Robert W Sumner, Marc Pollefeys, Federico Tombari, and Francis Engelmann. Openmask3d: Open-vocabulary 3d instance segmentation. arXiv preprint arXiv:2306.13631, 2023. 1" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 56, + 481, + 296, + 536 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 481, + 296, + 536 + ], + "spans": [ + { + "bbox": [ + 56, + 481, + 296, + 536 + ], + "type": "text", + "content": "[36] Xiaoyang Wu, Li Jiang, Peng-Shuai Wang, Zhijian Liu, Xihui Liu, Yu Qiao, Wanli Ouyang, Tong He, and Hengshuang Zhao. Point transformer v3: Simpler faster stronger. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 4840-4851, 2024. 1" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 56, + 537, + 296, + 591 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 537, + 296, + 591 + ], + "spans": [ + { + "bbox": [ + 56, + 537, + 296, + 591 + ], + "type": "text", + "content": "[37] Yanmin Wu, Jiarui Meng, Hajjie Li, Chenming Wu, Yahao Shi, Xinhua Cheng, Chen Zhao, Haocheng Feng, Errui Ding, Jingdong Wang, et al. Opengaussian: Towards point-level 3d gaussian-based open vocabulary understanding. arXiv preprint arXiv:2406.02058, 2024. 1, 2, 3, 4, 5, 6, 7, 8" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 56, + 592, + 296, + 646 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 592, + 296, + 646 + ], + "spans": [ + { + "bbox": [ + 56, + 592, + 296, + 646 + ], + "type": "text", + "content": "[38] Jihan Yang, Runyu Ding, Weipeng Deng, Zhe Wang, and Xiaojuan Qi. Regionplc: Regional point-language contrastive learning for open-world 3d scene understanding. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 19823-19832, 2024. 1, 2" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 56, + 647, + 296, + 691 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 647, + 296, + 691 + ], + "spans": [ + { + "bbox": [ + 56, + 647, + 296, + 691 + ], + "type": "text", + "content": "[39] Junbo Zhang, Runpei Dong, and Kaisheng Ma. Clip-for3d: Learning free open-world 3d scene representations from 2d dense clip. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 2048-2059, 2023. 2" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 56, + 691, + 296, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 691, + 296, + 713 + ], + "spans": [ + { + "bbox": [ + 56, + 691, + 296, + 713 + ], + "type": "text", + "content": "[40] Hengshuang Zhao, Li Jiang, Jiaya Jia, Philip HS Torr, and Vladlen Koltun. Point transformer. In Proceedings of" + } + ] + } + ], + "index": 12 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 316, + 73, + 555, + 318 + ], + "type": "list", + "angle": 0, + "index": 19, + "blocks": [ + { + "bbox": [ + 335, + 73, + 555, + 95 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 335, + 73, + 555, + 95 + ], + "spans": [ + { + "bbox": [ + 335, + 73, + 555, + 95 + ], + "type": "text", + "content": "the IEEE/CVF international conference on computer vision, pages 16259-16268, 2021. 1" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 316, + 96, + 555, + 171 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 96, + 555, + 171 + ], + "spans": [ + { + "bbox": [ + 316, + 96, + 555, + 171 + ], + "type": "text", + "content": "[41] Yuhang Zheng, Xiangyu Chen, Yupeng Zheng, Songen Gu, Runyi Yang, Bu Jin, Pengfei Li, Chengliang Zhong, Zeng-mao Wang, Lina Liu, Chao Yang, Dawei Wang, Zhen Chen, Xiaoxiao Long, and Meiqing Wang. Gaussiangrasper: 3d language gaussian splatting for open-vocabulary robotic grasping. IEEE Robotics and Automation Letters, 2024. 1, 2, 3" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 316, + 173, + 555, + 239 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 173, + 555, + 239 + ], + "spans": [ + { + "bbox": [ + 316, + 173, + 555, + 239 + ], + "type": "text", + "content": "[42] Shijie Zhou, Haoran Chang, Sicheng Jiang, Zhiwen Fan, Zehao Zhu, Dejia Xu, Pradyumna Chari, Suya You, Zhangyang Wang, and Achuta Kadambi. Feature 3dgs: Supercharging 3d gaussian splattering to enable distilled feature fields. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 21676-21685, 2024." + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 316, + 240, + 555, + 283 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 240, + 555, + 283 + ], + "spans": [ + { + "bbox": [ + 316, + 240, + 555, + 283 + ], + "type": "text", + "content": "[43] Xingxing Zuo, Pouya Samangouei, Yunwen Zhou, Yan Di, and Mingyang Li. Fmgs: Foundation model embedded 3d gaussian splatting for holistic 3d scene understanding. IJCV, 2024. 1, 2, 3" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 316, + 285, + 554, + 318 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 285, + 554, + 318 + ], + "spans": [ + { + "bbox": [ + 316, + 285, + 554, + 318 + ], + "type": "text", + "content": "[44] Matthias Zwicker, Hanspeter Pfister, Jeroen van Baar, and Markus H. Gross. EWA volume splatting. In Visualization Conference, 2001. 2" + } + ] + } + ], + "index": 18 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 749, + 318, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 749, + 318, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 749, + 318, + 757 + ], + "type": "text", + "content": "14146" + } + ] + } + ], + "index": 20 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 9 + } + ], + "_backend": "vlm", + "_version_name": "2.6.4" +} \ No newline at end of file diff --git a/2025/DrVideo_ Document Retrieval Based Long Video Understanding/81160fea-84bd-416d-98f4-7b4c205b85c0_content_list.json b/2025/DrVideo_ Document Retrieval Based Long Video Understanding/81160fea-84bd-416d-98f4-7b4c205b85c0_content_list.json new file mode 100644 index 0000000000000000000000000000000000000000..d966fa8cba9fbe5c380eb55a18e5ec6ffb318ca9 --- /dev/null +++ b/2025/DrVideo_ Document Retrieval Based Long Video Understanding/81160fea-84bd-416d-98f4-7b4c205b85c0_content_list.json @@ -0,0 +1,1516 @@ +[ + { + "type": "text", + "text": "DrVideo: Document Retrieval Based Long Video Understanding", + "text_level": 1, + "bbox": [ + 174, + 130, + 823, + 152 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Ziyu Ma $^{1,2}$ , Chenhui Gou $^{2}$ , Hengcan Shi $^{1,2}$ , Bin Sun $^{1}$ , Shutao Li $^{1}$ , Hamid Rezatofighi $^{2}$ , Jianfei Cai $^{2}$", + "bbox": [ + 236, + 179, + 756, + 215 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "1 College of Electrical and Information Engineering, Hunan University,", + "bbox": [ + 217, + 215, + 779, + 234 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "2 Data Science & AI Department, Faculty of IT, Monash University", + "bbox": [ + 233, + 234, + 764, + 252 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "{maziyu, sunbin611, shutao_li}@hnu.edu.cn", + "bbox": [ + 318, + 253, + 676, + 268 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "{chenhui.gou, hengcan.shi, hamid.rezatofighi, jianfei.cai}@monash.edu", + "bbox": [ + 191, + 272, + 800, + 287 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Abstract", + "text_level": 1, + "bbox": [ + 246, + 321, + 326, + 338 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Most of the existing methods for video understanding primarily focus on videos only lasting tens of seconds, with limited exploration of techniques for handling long videos. The increased number of frames in long videos poses two main challenges: difficulty in locating key information and performing long-range reasoning. Thus, we propose DrVideo, a document-retrieval-based system designed for long video understanding. Our key idea is to convert the long-video understanding problem into a long-document understanding task so as to effectively leverage the power of large language models. Specifically, DrVideo first transforms a long video into a coarse text-based long document to initially retrieve key frames and then updates the documents with the augmented key frame information. It then employs an agent-based iterative loop to continuously search for missing information and augment the document until sufficient question-related information is gathered for making the final predictions in a chain-of-thought manner. Extensive experiments on long video benchmarks confirm the effectiveness of our method. DrVideo significantly outperforms existing LLM-based state-of-the-art methods on EgoSchema benchmark (3 minutes), MovieChat-1K benchmark (10 minutes), and the long split of Video-MME benchmark (average of 44 minutes). Code is available at https://github.com/Upper9527/DrVideo.", + "bbox": [ + 88, + 353, + 485, + 731 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "1. Introduction", + "text_level": 1, + "bbox": [ + 89, + 762, + 220, + 777 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Video understanding is a challenging task in computer vision, requiring the processing of spatio-temporal information and advanced reasoning abilities. Previous works have successfully processed short videos lasting around tens of seconds [8, 27, 32, 53, 69, 70]. However, how to deal with", + "bbox": [ + 89, + 787, + 483, + 864 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "long video understanding remains unclear. Recent advancements in large language models (LLMs) [6, 10, 34, 41, 45, 48, 56] demonstrate strong abilities in language understanding and reasoning across long text sequences. These advancements have inspired the development of videolanguage models (Video-LMs) [25, 38, 68, 77] to address long video understanding issues.", + "bbox": [ + 511, + 323, + 905, + 429 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Video-LMs typically encode a video as a sequence of visual tokens, which are then concatenated with language tokens into a long sequence and LLM is then used to understand this long sequence. Although Video-LMs improve video understanding, they have the following limitations. (i) They cannot process an entire video as input to a VideoLM (e.g., the visual encoder OpenAI's CLIP-L-14 outputs $24 \\times 24$ tokens for each image, while the max length of LLaVA-NeXT-Video [77] is only 8192). (ii) They typically perform uniform or random frame sampling at large strides to handle long videos without considering the content (e.g., 16 for PLLaVA [68] and 6 for IG-VLM [25]), leading to potential key information loss. (iii) The simple concatenation of visual tokens increases the difficulty for the LLM to locate question-related (key) visual information within the long video sequence, complicating long-range reasoning across the vision token sequences.", + "bbox": [ + 511, + 430, + 906, + 686 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Apart from the studies of encoding a video as a sequence of visual tokens, another line of research converts raw videos into captions and then utilities the long-range reasoning abilities of LLMs to predict the final answer. A pioneering work - LLoVi [74] uses a short-term video model [78] in conjunction with an LLM [40] to solve the long video understanding task. Given a long video, LLoVi first divides it into multiple short clips and converts them into short textual descriptions. Afterward, these short textual descriptions are summarized by the LLM, and the summary is finally used by LLoVi to answer the given question. Inspired by LLoVi, VideoAgent [61] designs an agent-based system to further leverage the reasoning abilities of LLMs and locate question-related video clips or images. Initially,", + "bbox": [ + 511, + 688, + 908, + 902 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "CVF", + "bbox": [ + 106, + 2, + 181, + 42 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore.", + "bbox": [ + 236, + 0, + 810, + 46 + ], + "page_idx": 0 + }, + { + "type": "page_footnote", + "text": "*Equal contribution.", + "bbox": [ + 107, + 875, + 220, + 887 + ], + "page_idx": 0 + }, + { + "type": "page_footnote", + "text": "† Corresponding author.", + "bbox": [ + 109, + 887, + 235, + 898 + ], + "page_idx": 0 + }, + { + "type": "page_number", + "text": "18936", + "bbox": [ + 480, + 944, + 519, + 955 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/aef6f5c7e3fbd886231fae7a678d4a9ff61203dc2838a65af6ffb065fa5d0703.jpg", + "image_caption": [ + "Figure 1. Overview of our DrVideo framework. It comprises five components: a video-document conversion module, a retrieval module, a document augmentation module, a multi-stage agent interaction loop and an answering module." + ], + "image_footnote": [], + "bbox": [ + 135, + 85, + 861, + 295 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "VideoAgent samples only a minimal number of frames from the video (i.e., 5 frames), converting them into textual descriptions as input for the LLM. Then, through continuous feedback and interaction with the LLM [1], VLM [78], and CLIP [51], it gradually finds question-related frames in a coarse-to-fine manner via image-text similarity. Finally, VideoAgent answers the question based on these question-related frames. Despite these LLM-based methods overcoming the challenge of performing long-range reasoning and achieving promising results, they still suffer from the following problems.", + "bbox": [ + 88, + 347, + 485, + 513 + ], + "page_idx": 1 + }, + { + "type": "list", + "sub_type": "text", + "list_items": [ + "- These methods struggle to accurately identify question-related frames. For instance, VideoAgent [61] performs sparse sampling on long videos (e.g., selecting 5 frames) and subsequently searches for keyframes from the unsampled frames in a coarse-to-fine manner based on the initial sparse space via CLIP image-text similarly given the missing question-related information inferred by the LLM, which improves inference efficiency but unavoidably at the cost of lacking a holistic grasp of video content, particularly when a long video consists of multiple events (e.g., over 100 events) with significant content variations, and since the CLIP text encoder is trained with simple language structures and limited context (the max length is 77), it struggles to understand complex video semantics, limiting its retrieval accuracy.", + "- These methods convert videos into captions, addressing long-range reasoning but overlooking critical information loss. Specifically, even if question-related frames are identified, the generated captions cannot cover all the essential information in the key frames. For example, the caption \"a woman is looking in the mirror\" does not contribute to answering the question \"what is the woman wearing when standing in front of the mirror in the video\". In this case, LLM can only make an aimless and biased surmise to answer the question." + ], + "bbox": [ + 89, + 523, + 485, + 901 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "To solve the above problem, we turn to how human understands a long video. In particular, a human needs to gain a general understanding of the video, identifying events across segments. Then, the human locates question-related segments, examines them in detail, and answers the question based on the extracted information. Inspired by this, in this paper, we propose DrVideo, a document retrieval-based system for long video understanding that progressively finds key frames and generates a question-related augmented document for LLMs to make confident predictions. Our key idea is to convert long video understanding into long document understanding and leverage the impressive capabilities of LLMs in key information location [46] and long-range reasoning [12] on various NLP tasks. Note that long-document retrieval in our paper means passage retrieval within a single long document, not retrieval from a corpus consisting of long documents.", + "bbox": [ + 511, + 347, + 908, + 604 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Specifically, DrVideo transforms raw long videos into long documents using a pretrained frame/clip-level visual captioner (e.g., LLaVA-NeXT [33]) to make the LLM have a holistic grasp of video content. Different from VideoAgent [61], which performs sparse sampling and identifies key frames via image-text similarity without information augmentation, DrVideo performs dense sampling and proposes a retrieval module and a multi-stage agent interaction loop that dynamically finds the potential missing information through sequential steps of reasoning and then interacts with a document augmentation module to request augmentation of the information within language space. Once sufficient question-related information is gathered, it feeds into the answering module to get the final prediction.", + "bbox": [ + 511, + 608, + 908, + 821 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "DrVideo achieves significant performance gains (3.8% on EgoSchema subset [36], 8.3% on MovieChat-1K global mode, 24.8% on MovieChat-1K breakpoint mode [49], and 6.3% on the long split of Video-MME [15]) over the existing LLM-based state-of-the-art methods [14, 37, 61, 74]", + "bbox": [ + 511, + 824, + 910, + 902 + ], + "page_idx": 1 + }, + { + "type": "page_number", + "text": "18937", + "bbox": [ + 480, + 944, + 519, + 955 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "across different long video benchmarks, from 3 minutes to 10 minutes and longer than 1 hour. Surprisingly, combining the subtitles on the long split of Video-MME benchmark, DrVideo obtains $71.7\\%$ accuracy, outperforming many heavily-engineered large-scale proprietary models (e.g., Gemini 1.5 Flash [55], GPT-4o mini [44], and GPT-4V [42]). This highlights the great potential of DrVideo. Moreover, DrVideo is a training-free framework and is researcher-friendly, as our results can be replicated on an RTX 4090 with a reasonable number of GPT accesses.", + "bbox": [ + 89, + 90, + 480, + 241 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "The contribution of this work can be summarized as:", + "bbox": [ + 109, + 242, + 459, + 256 + ], + "page_idx": 2 + }, + { + "type": "list", + "sub_type": "text", + "list_items": [ + "- We propose DrVideo, the first document-retrieval-based system designed for the long video understanding task by converting it into a long-document retrieval and understanding task to effectively leverage the power of large language models.", + "- Different from previous works, we propose a retrieval module and a new multi-stage agent interaction loop that dynamically finds the potential missing information and augments these information within language space.", + "- DrVideo outperforms the existing LLM-based state-of-the-art methods across different long video benchmarks (i.e., EgoSchema [36], MovieChat-1K [49], and the long split of Video-MME [15]) by a margin." + ], + "bbox": [ + 89, + 257, + 482, + 454 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "2. Related Work", + "text_level": 1, + "bbox": [ + 89, + 467, + 232, + 483 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Long Video Understanding. Modeling long videos, which are several minutes or more in length, generally requires advanced temporal modeling, resulting in complex model designs. LF-VILA [52] introduces a Temporal Window Attention (HTWA) mechanism to capture long-range dependencies in long videos. MeMViT [67] and MovieChat [49] employ a memory-based design to save question-related information from previous video segments. Some other approaches employ space-time graphs [21, 62] or relational space-time modules [71] to capture spatio-temporal dependencies from the raw long videos. Recently, S4ND [39], ViS4mer [22], and S5 [58] have used Structured State-Space Sequence (S4) [18] layers to capture long-range dependencies in videos. Different from these methods, our DrVideo does not design a complex module to perform long video understanding. Instead, we develop a document retrieval-based system with LLM for zero-shot long video understanding.", + "bbox": [ + 89, + 492, + 482, + 763 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "LLMs for Video Understanding. The recent rise of LLMs [73] and VideoChat [30] align the visual features extracted by pretrained visual models to LLMs and apply them to video understanding. Video ChatCaptioner [7] and ChatVideo [57] utilize LLMs to represent videos and engage users through dialogues, respectively. VidIL [64] applies the image-level models to video understanding tasks via few-shot learning. In addition to short-term video understanding, recent studies [4, 11, 32] have explored LLMs", + "bbox": [ + 89, + 765, + 482, + 901 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "for long-range video modeling. For instance, GPT-4 is applied in various long-range video modeling tasks in [32], though quantitative evaluation is limited. Meanwhile, the research in [11] focuses on movie datasets with minimal visual analysis [36], relying largely on speech and subtitles. In contrast, DrVideo focuses on vision modality in multiple benchmarks.", + "bbox": [ + 511, + 90, + 903, + 196 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "LLM Agents. In parallel, the computer vision community has started exploring the use of LLMs as agents in various vision tasks such as GUI understanding and robot navigation [5, 13, 20, 53]. In the realm of long video comprehension, initial efforts have employed an agent-like approach, where LLMs interact with external tools or integrate additional functionalities [14, 17, 53, 61, 65, 72]. In contrast to these approaches, our method, DrVideo, reconceptualizes long video understanding as a process of document retrieval, augmentation, and understanding, to leverage the strong capability of LLMs.", + "bbox": [ + 511, + 200, + 906, + 368 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3. Methodology", + "text_level": 1, + "bbox": [ + 511, + 397, + 648, + 414 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "In this section, we detail our proposed document retrieval-based system for long video understanding. As illustrated in Figure 1, given a long video and a question about the video, DrVideo first translates the long video into a long document, referred to as the initial video document. Then, the retrieval module identifies the top $K$ key frames by calculating the similarity between the question and the initial video document. The document augmentation module enriches the information of these key frames and adds it to the initial video document, creating a new video document (named the updated video document), which is used as the starting point for the subsequent multi-stage agent interaction loop.", + "bbox": [ + 511, + 426, + 906, + 625 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "The loop contains two distinct agents: the planning agent and the interaction agent. The planning agent judges whether the updated video document is sufficient for answering the question. If not, the updated video document is fed into the interaction agent to dynamically find missing key frames. The interaction agent interacts with the document augmentation module to request the required information of these key frames. The output of the document augmentation module is added to the current document to obtain the latest updated video document. The new video document is then sent to the planning agent again to further judge whether the current information is sufficient. This searching and interaction loop continues until the planning agent considers the current information sufficient for answering the question or the maximum iteration is reached. After the loop ends, the final video document and the related question are given to the answering module to get the final prediction.", + "bbox": [ + 511, + 628, + 906, + 900 + ], + "page_idx": 2 + }, + { + "type": "page_number", + "text": "18938", + "bbox": [ + 480, + 945, + 517, + 955 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3.1. Video-Document Conversion Module", + "text_level": 1, + "bbox": [ + 89, + 90, + 410, + 104 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "By translating each single video frame $V_{t}$ (or a short video clip) into a short description $S_{V_t}$ , we convert an input long video into the initial video document $Doc_{init}$ :", + "bbox": [ + 89, + 112, + 482, + 157 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\nD o c _ {i n i t} = \\left\\{\\left\\{1, S _ {V _ {1}} \\right\\}, \\left\\{2, S _ {V _ {2}} \\right\\}, \\dots , \\left\\{T, S _ {V _ {T}} \\right\\} \\right\\} \\tag {1}\n$$\n", + "text_format": "latex", + "bbox": [ + 129, + 167, + 482, + 186 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "where $T$ is the length of the video. Specifically, $\\mathrm{S}_{V_t}$ is generated by a large vision-language captioning model $\\phi_{vlm}$ (e.g., LLaVA-NeXT [33]), i.e., $\\mathrm{S}_{V_t} = \\phi_{vlm}(\\mathcal{P}, V_t)$ , where $\\mathcal{P}$ is the prompt for requesting a short description about the image, e.g., describe the picture in no more than 50 words. The captioning model used in DrVideo can be replaced by other captioning models (e.g., LaViLa [78], BLIP-2 [29]). The experimental results with different captioning models can be found in Table 6.", + "bbox": [ + 89, + 195, + 483, + 330 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "3.2. Document (Frames) Retrieval Module", + "text_level": 1, + "bbox": [ + 89, + 340, + 418, + 356 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "After obtaining $Doc_{init}$ , we introduce a document retrieval module to identify the top question-related frames by calculating the similarity between the question and the whole document. Specifically, we use the OpenAI embedding model $\\phi_{emb}$ [43] to obtain the vector representation of the initial document, i.e., $\\mathcal{E}_{\\mathrm{doc}} = \\phi_{emb}(Doc_{init})$ . Given a specific retrieval text $\\mathcal{RT}$ , i.e., the question $Q$ , the retrieval module first computes the embedding $\\mathcal{E}_{\\mathcal{RT}} = \\phi_{emb}(\\mathcal{RT})$ and then retrieves the top $K$ frames based on the cosine similarity between $\\mathcal{E}_{\\mathcal{RT}}$ and $\\mathcal{E}_{\\mathrm{doc}}$ :", + "bbox": [ + 89, + 362, + 483, + 513 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\operatorname {t o p k} _ {\\text {一}} \\operatorname {d o c} = \\underset {t} {\\arg \\operatorname {t o p}} K \\cos \\left(\\mathcal {E} _ {\\mathcal {R T}}, \\mathcal {E} _ {\\mathrm {d o c} _ {t}}\\right) \\tag {2}\n$$\n", + "text_format": "latex", + "bbox": [ + 148, + 516, + 482, + 541 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "where $\\mathcal{E}_{\\mathrm{doc}_t}$ is the embedding of the $t$ -th description $S_{V_t}$ .", + "bbox": [ + 89, + 544, + 465, + 560 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "3.3. Document Augmentation Module", + "text_level": 1, + "bbox": [ + 89, + 568, + 382, + 583 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "For each $t' \\in \\text{topk\\_doc}$ , we use the LLaVA-NeXT model [33] with different prompts $\\mathcal{AP}$ (augmented prompts) to generate a detailed description $L_{V_{t'}}$ , i.e., $L_{V_{t'}} = \\phi_{vlm}(\\mathcal{AP}_{t'}, V_{t'})$ . The initial $\\mathcal{AP}_{t'}$ is a general prompt: If there are factual errors in the question, provide a precise description of the image; if not, proceed to answer the question: $\\{Q\\}$ . The updated video document $\\mathcal{AD}$ becomes", + "bbox": [ + 89, + 590, + 483, + 709 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {A D} = \\left\\{\\left\\{ \\begin{array}{l l} \\left\\{t, S _ {V _ {t}}, L _ {V _ {t} ^ {\\prime}} \\right\\} & \\text {i f} t = t ^ {\\prime} \\\\ \\left\\{t, S _ {V _ {t}} \\right\\} & \\text {o t h e r w i s e} \\end{array} \\right| t = 1, \\dots , T \\right\\}. \\tag {3}\n$$\n", + "text_format": "latex", + "bbox": [ + 116, + 718, + 482, + 755 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Note that $K$ is significantly smaller than the total number of frames (e.g., $K = 5$ compared to $T = 90$ ), so the additional descriptions will not substantially increase the overall length of the document.", + "bbox": [ + 89, + 763, + 482, + 824 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "3.4. Multi-Stage Agent Interaction Loop", + "text_level": 1, + "bbox": [ + 89, + 833, + 403, + 849 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Apart from the retrieved key frames, we further introduce a multi-stage agent interaction loop to dynamically find other potential key frames and augment them with different types", + "bbox": [ + 89, + 854, + 483, + 901 + ], + "page_idx": 3 + }, + { + "type": "image", + "img_path": "images/bf3ad29ab9cee039b2f4a4123239ff00ee6e390e8ed9cd05704e9e3713fdadcd.jpg", + "image_caption": [ + "Figure 2. Illustration of the multi-stage agent interaction loop and answering module. There are two agents in the multi-stage agent interaction loop: a planning agent to plan the next step and an interaction agent to dynamically find missing information and interact with the document augmentation module." + ], + "image_footnote": [], + "bbox": [ + 517, + 90, + 895, + 311 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "of information for the final answer. We employ two distinct agents: planning and interaction agents, each tailored for a specific task. Both agents utilize the same LLM, i.e., GPT-3.5, but are given different prompts and reason varying types of information. The detail is shown in Fig. 2", + "bbox": [ + 511, + 405, + 905, + 479 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Planning Agent. Given the question $Q$ and the updated video document $\\mathcal{AD}_i$ (where $i$ denotes the iteration step), along with the analysis history $\\mathcal{H}$ from all previous steps (initialized as empty $\\{\\}$ ), the planning agent first determines whether the updated video document is sufficient to generate a confident answer. If the information is sufficient, the process moves to the answering module to get the final answer. If not, the agent provides an analysis of why the updated video document $\\mathcal{AD}_i$ is inadequate. The agent then updates $\\mathcal{H}_i$ with its analysis and passes the updated $\\mathcal{H}_i$ to the interaction agent.", + "bbox": [ + 511, + 482, + 905, + 647 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Interaction Agent. Given the current video document $\\mathcal{AD}_i$ and the updated $\\mathcal{H}_i$ , the interaction agent finds $N$ frames with missing details crucial for answering $Q$ :", + "bbox": [ + 511, + 648, + 905, + 694 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\forall n \\in N, n \\in T, n \\notin \\text {t o p k} _ {-} \\text {d o c}, \\text {a n d} N < K. \\tag {4}\n$$\n", + "text_format": "latex", + "bbox": [ + 537, + 705, + 903, + 722 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "In addition to identifying the missing $N$ key frames, the interaction agent also determines the augmented information type for each $n \\in N$ . To achieve it, we give a task-specific prompt to the interaction agent, i.e., Your task is to determine which frame needs which type of information and can answer this question accurately, reasonably, and without contradiction. The two types of information are as follows: A: Given an image, get a detailed description of the image (image caption, just like what is shown in this image?) B: Given an image, get a response to the above question (visual question answering). After obtaining $N$", + "bbox": [ + 511, + 734, + 906, + 901 + ], + "page_idx": 3 + }, + { + "type": "page_number", + "text": "18939", + "bbox": [ + 480, + 944, + 517, + 955 + ], + "page_idx": 3 + }, + { + "type": "table", + "img_path": "images/c9f019dbf2a337df8f50ac7d0555789fedf56a26c8761b78c5731cf0203c3875.jpg", + "table_caption": [], + "table_footnote": [], + "table_body": "
Method(M)LLMSubsetFullset
MoReVQA [37]PaLM-2-51.7
Vamos [60]GPT-451.248.3
ProViQ [9]GPT-3.557.1-
LLoVi [74]GPT-3.557.650.3
IG-VLM [25]GPT-4V59.8-
VideoAgent [61]GPT-460.254.1
MVU [47]Mistral-13B60.337.6
LLoVi [74]GPT-461.2-
VideoAgent [14]GPT-462.860.2
DrVideo (ours)GPT-3.562.6-
DrVideo (ours)GPT-466.461.0
", + "bbox": [ + 125, + 88, + 450, + 275 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "frames along with their required information type, the agent interacts with the document augmentation module to enrich the information of these frames to update the current video document for the next step.", + "bbox": [ + 89, + 325, + 483, + 387 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "3.5. Answering Module", + "text_level": 1, + "bbox": [ + 89, + 397, + 272, + 412 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Given the final video document $\\mathcal{AD}_{final}$ , we employ another agent to provide a prediction using a zero-shot chain-of-thought (CoT) approach [26, 66]. The answering module outputs the corresponding answer, the confidence score, and the reasoning behind the answer. Besides improving prediction accuracy, the CoT approach allows us to trace inference steps, ensuring transparency and explainability in the decision-making process.", + "bbox": [ + 89, + 417, + 483, + 541 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "4. Experiments", + "text_level": 1, + "bbox": [ + 89, + 554, + 223, + 571 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "In this section, we first introduce the datasets and implementation details and then present the results and ablations of our proposed DrVideo.", + "bbox": [ + 89, + 580, + 482, + 626 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "4.1. Datasets and Metrics", + "text_level": 1, + "bbox": [ + 89, + 636, + 290, + 651 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "In our experiments, we evaluate our DrVideo using three well-established datasets, emphasizing its zero-shot long video understanding capabilities.", + "bbox": [ + 89, + 657, + 482, + 703 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "EgoSchema. EgoSchema dataset [36] consists of 5000 multiple-choice questions sourced from 5000 three-minute egocentric videos. This dataset only has a subset of 500 questions with publicly accessible labels, while the full set is evaluated on the leaderboard.", + "bbox": [ + 89, + 704, + 482, + 777 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "MovieChat-1K. MovieChat-1K [49] is a longer video understanding benchmark, which contains 1000 videos from movies and TV shows, each video lasting approximately 10 minutes. This dataset has two modes: global mode and breakpoint mode. The global mode refers to analyzing the entire video to understand its overall content and context, while the breakpoint mode focuses on analyzing specific frames or scenes.", + "bbox": [ + 89, + 779, + 482, + 901 + ], + "page_idx": 4 + }, + { + "type": "table", + "img_path": "images/0e10d8cf3c024e274e533aa8c9baf1a45884cdfab3080503a3955f07d2d32164.jpg", + "table_caption": [ + "Table 1. Results on EgoSchema compared to existing LLM-based state-of-the-art methods." + ], + "table_footnote": [], + "table_body": "
MethodGlobalBreakpoint
Acc.ScoreAcc.Score
Open-Source MLLM
Video LLaMA [75]51.72.7239.12.11
Video Chat [30]57.83.0846.12.32
Video-ChatGPT [35]47.62.8948.02.43
MovieChat [49]68.33.1548.32.46
MovieChat+ [50]71.23.5149.62.62
Based on Open-source Captioners and Proprietary LLMs
MM-VID [32]58.62.8610.40.56
LLoVi [74]58.32.8717.81.03
\\( VideoAgent^* \\) [61]65.4*3.12*31.6*2.05*
Sullam Jeoung, et al [23]84.8---
DrVideo (ours)93.14.4156.42.75
", + "bbox": [ + 517, + 88, + 903, + 295 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Table 2. Performance comparison on the MoiveChat-1K [49] benchmark against state-of-the-art methods. * represents re-implemented results, implementation details refer to appendix.", + "bbox": [ + 511, + 305, + 903, + 348 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Video-MME. Video-MME [15] is a recently proposed benchmark for comprehensive video analysis evaluation. We evaluate DrVideo on the \"long-term videos\" split of this dataset (long split), where video lengths vary from 30 to 60 minutes, with an average duration of 44 minutes.", + "bbox": [ + 511, + 356, + 906, + 431 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Since EgoSchema and Video-MME are multi-choice tasks, we use accuracy as the evaluation metric for both datasets. MovieChat-1K focuses on open-ended questions, we use GPT-assisted evaluation to assess both the accuracy (true/false) and quality (score 0-5) of the models. We select Gemini-Pro [54] as the evaluation assistant and use the same prompt [38] to conduct a fair comparison.", + "bbox": [ + 511, + 431, + 905, + 537 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "4.2. Implementation Details", + "text_level": 1, + "bbox": [ + 511, + 546, + 730, + 561 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "For EgoSchema [36] dataset, we choose LaViLa [78] as the captioning model to convert videos into documents. Note that the training data of our LaViLa model does not include EgoSchema videos, which is the same as LLoVi [74]. For MovieChat-1K and Video-MME benchmarks, LLaVA-NeXT [33] is used as our captioning model to generate brief descriptions for each frame. We preprocess videos by simply sampling them at 0.5 FPS for EgoSchema [36] and MovieChat-1K [49] and 0.2 FPS for Video-MME [15]. LLaVA-NeXT [33] is chosen to augment key frames in both datasets via different prompts. In the comparison experiments, GPT-4 [1], i.e., gpt-4-1106-preview, is used as the agent to evaluate the performance of our DrVideo on EgoSchema and MovieChat-1K. DeepSeek [19], i.e., DeepSeek V2.5, is used as the agent to evaluate the performance on Video-MME. For the ablation study, we use GPT-3.5 [40], i.e., gpt-3.5-turbo-1106, to evaluate the effectiveness of our DrVideo due to the API cost.", + "bbox": [ + 511, + 568, + 906, + 840 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "4.3. Main Results", + "text_level": 1, + "bbox": [ + 511, + 848, + 651, + 863 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Comparative results on EgoSchema. Table 1 compares our DrVideo with existing LLM-based state-of-the", + "bbox": [ + 511, + 869, + 903, + 901 + ], + "page_idx": 4 + }, + { + "type": "page_number", + "text": "18940", + "bbox": [ + 480, + 944, + 517, + 955 + ], + "page_idx": 4 + }, + { + "type": "table", + "img_path": "images/b13dff7a51ebb95c2eb807ffc938bf49cb469b51d1c3d4677848f13e682b6d70.jpg", + "table_caption": [], + "table_footnote": [], + "table_body": "
MethodFramesw/o Subsw Subs
Proprietary MLLM
Claude 3.5 Sonnet [2]2051.254.7
GPT-4V [42]1053.556.9
GPT-4o mini [44]25058.663.4
Gemini 1.5 Flash [55]0.5 FPS61.168.8
GPT-4o [44]38465.372.1
Gemini 1.5 Pro [55]0.5 FPS67.477.4
Open-Source MLLM
LongVA [76]12846.247.6
VITA-8x7B [16]3248.650.9
LLaVA-OneVision-72B [28]3260.062.4
Qwen2-VL-72B [3]76862.274.3
Based on Open-source Captioners and Proprietary LLMs
\\( VideoAgent^* \\)[61]7.640.2*44.4*
\\( LLoVi^* \\)[74]0.2 FPS45.4*67.7*
DrVideo (ours)0.2 FPS51.771.7
-only subs0.2 FPS-68.5
", + "bbox": [ + 106, + 88, + 470, + 369 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "art methods [9, 14, 25, 37, 47, 60, 61, 74] on EgoSchema benchmark [36]. Specifically, compared with LLoVi [74] and VideoAgent [61] that leverage the same visual captioner (LaViLa [78]) and LLM (GPT-4), DrVideo significantly outperforms these methods by $5.2\\%$ and $6.2\\%$ respectively. Compared with VideoAgent [14] which uses videospecific models (ViCLIP from InternVid [63] and VideoLLaVA [31]), DrVideo still performs better and achieves $3.6\\%$ higher accuracy on the subset evaluation. Moreover, DrVideo significantly outperforms IG-VLM [25] by $6.6\\%$ , which uses strong MLLM, i.e., GPT-4V. These results confirm our idea of converting the long-video understanding task into a long-document understanding task is feasible and our DrVideo can find the question-related key frames more accurately.", + "bbox": [ + 88, + 491, + 482, + 718 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Comparative results on longer video benchmarks. To further emphasize the advantage of our method on longer videos, we evaluate our DrVideo on the other two video benchmarks, i.e., MoiveChat-1K (10 minutes) [49], and the long split of Video-MME (average 44 minutes) [15]. Table 2 presents the comparative results of our DrVideo and the current state-of-the-art methods [30, 38, 49, 74, 75] on the MovieChat-1K benchmark. Table 3 shows the comparative results of our DrVideo on the Video-MME benchmark, with comparisons against three types of methods: proprietary MLLMs [2, 42, 44, 55], open-source MLLMs [3, 16, 28, 76], and methods based on open-source caption", + "bbox": [ + 89, + 719, + 482, + 901 + ], + "page_idx": 5 + }, + { + "type": "table", + "img_path": "images/31cd407c979a5d19cec95d3c01e57ec5d6d7c86413205a6af5f531545fd940f8.jpg", + "table_caption": [ + "Table 3. Video-MME long split results. $w/o$ subs represents the model without any subtitles. $w$ subs represents that the subtitles corresponding to the sampled frames are both fed into the LLM to evaluate the performance. Our DrVideo outperforms a strong proprietary MLLM (Claude 3.5 Sonnet) and some open-source MLLMs (e.g., VITA-8x7B). * represents re-implemented results, implementation details refer to appendix." + ], + "table_footnote": [], + "table_body": "
RMMSAILCoTAcc.(%)
✓✓✓62.6
✓✓X62.2
✓X✓60.6
X✓✓59.4
XX✓57.4
", + "bbox": [ + 547, + 88, + 874, + 191 + ], + "page_idx": 5 + }, + { + "type": "table", + "img_path": "images/204b50764ed18a1eb37988d7f528c0366c19eddebee202289d8a53ad1927970c.jpg", + "table_caption": [ + "Table 4. Ablation results on different combinations of the retrieval module (RM), the multi-stage agent interaction loop (MSAIL), and the CoT on EgoSchema." + ], + "table_footnote": [ + "(a) Performance with different types of information" + ], + "table_body": "
VQACaptionAcc. (%)
✓✓62.6
×✓60.4
✓×61.8
", + "bbox": [ + 529, + 255, + 722, + 330 + ], + "page_idx": 5 + }, + { + "type": "table", + "img_path": "images/c9e13824dd2e56e1f7c5fcb434e08100a539d6a44015002d0e9acb66289f0942.jpg", + "table_caption": [], + "table_footnote": [ + "(b) Performance with different Top- $K$ frames" + ], + "table_body": "
Top-KAcc. (%)
562.6
1061.4
2060.6
", + "bbox": [ + 728, + 255, + 864, + 330 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Table 5. Ablation results with different settings on EgoSchema.", + "bbox": [ + 519, + 366, + 897, + 380 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "ers and proprietary LLMs [61, 74].", + "bbox": [ + 511, + 397, + 746, + 412 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Compared to methods like LLoVi [74] and VideoAgent [61] that are based on open-source captioners (e.g., LLaVA-NeXT [33]) and proprietary LLMs (e.g., GPT-4 [1]), DrVideo achieves a $8.3\\%$ improvement in global mode and a $24.8\\%$ improvement in breakpoint mode on the MovieChat-1K benchmark in Table 2. It also achieves at least $6.3\\%$ improvement under the w/o subtitle setting on the Video-MME benchmark in Table 3. This indicates that DrVideo can accurately identify question-related frames and avoid critical information loss through augmentation, highlighting the potential of document retrieval methods for processing longer videos. Compared to proprietary MLLMs and open-source MLLMs, DrVideo outperforms Claude 3.5 Sonnet [2] by $0.5\\%$ and some open-source MLLMs (e.g., VITA-8x7B [16]) under the w/o subtitle setting on Video-MME benchmark. It also achieves a $21.9\\%$ improvement in global mode and a $6.8\\%$ improvement in breakpoint mode on MovieChat-1K benchmark.", + "bbox": [ + 511, + 414, + 906, + 686 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Moreover, combining the subtitle on the Video-MME benchmark, DrVideo achieves $71.7\\%$ accuracy and outperforms many heavily-engineered large-scale proprietary models (e.g., Gemini 1.5 Flash [55], GPT-4o mini [44], and GPT-4V [42]). To better understand this significant improvement, we conduct an experiment where only the subtitle, question and options are input into the LLM (i.e., DeepSeek [19]) to predict the answer and the accuracy is $68.5\\%$ . This suggests: (i) subtitles contain extensive question-related information, playing a crucial role in understanding long videos; (ii) compared with LLoVi and VideoAgent, DrVideo enhances visual information via document retrieval and augmentation, effectively integrating subtitles and resulting in a $3.2\\%$ improvement.", + "bbox": [ + 511, + 688, + 908, + 902 + ], + "page_idx": 5 + }, + { + "type": "page_number", + "text": "18941", + "bbox": [ + 480, + 944, + 517, + 955 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/4dfa8b54a6ab788efd253eb7da839e72d5dd4331d464ab23df0e11ba1fdb0505.jpg", + "image_caption": [ + "Figure 3. Performance of different rounds with DrVideo and VideoAgent [61] on EgoSchema. To align with VideoAgent, GPT-4, i.e., gpt-4-turbo-1106-preview, is used as the LLM agents." + ], + "image_footnote": [], + "bbox": [ + 112, + 95, + 464, + 252 + ], + "page_idx": 6 + }, + { + "type": "table", + "img_path": "images/cb0e9c83de53f3fa0ecd17d823af36c9e031688171199099d621325aa6d95e80.jpg", + "table_caption": [], + "table_footnote": [], + "table_body": "
Captioning ModelTypeAcc.(%)
LaViLa [78]Clip-based62.6
LLaVA-NeXT [33]Frame-based61.2
BLIP-2 [29]Frame-based59.6
", + "bbox": [ + 114, + 328, + 460, + 401 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "4.4. Ablation Studies", + "text_level": 1, + "bbox": [ + 89, + 438, + 254, + 452 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Unless otherwise specified, we use GPT-3.5 as the default setting in the below experiments on EgoSchema subset [36]. Effects of different components of DrVideo. We conduct experiments with different combinations of the individual components in DrVideo, including the retrieval module, the multi-stage agent interaction loop, and the CoT used in the answering module. The results are presented in Table 4. We can see that: (i) without the retrieval module and the agent loop, the performance of our DrVideo is similar to LLoVi [74]; (ii) By adding the retrieval module and then the agent loop, the performance of DrVideo improves from $57.4\\%$ to $60.6\\%$ and then to $62.6\\%$ . This demonstrates that the retrieval module can identify key frames via measuring semantic similarity, while the agent loop can locate additional key frames through contextual inference; (iii) the additional CoT contributes to the accuracy improvement from $62.2\\%$ to $62.6\\%$ , demonstrating its effectiveness.", + "bbox": [ + 89, + 460, + 482, + 715 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Effects of different types of augment information. To better enhance the required information for each potential key frame, we define two types of augment information, namely VQA (question-related information) and caption (general detailed information). Table 5a presents the performance of DrVideo with different types of information. We can see that when only one type of information (VQA or caption) is used for augmentation, the model's performance declines. This suggests the necessity of adaptively augmenting different types of information for each potential key frame.", + "bbox": [ + 89, + 718, + 482, + 883 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Effects of different initial top $K$ key frames. By default,", + "bbox": [ + 89, + 885, + 482, + 900 + ], + "page_idx": 6 + }, + { + "type": "table", + "img_path": "images/71b531860e2c2b075d2f8fb8f9add18afeafb254fd32575f4c04e5b58c036ef3.jpg", + "table_caption": [ + "Table 6. Performance of different VLMs on EgoSchema." + ], + "table_footnote": [], + "table_body": "
LLMsTypeSizeAcc.(%)
Mistral-8x7B [24]Open-Source8x7B47.6
DeepSeek [19]ProprietaryN/A61.2
GPT-3.5 [40]ProprietaryN/A62.6
GPT-4 [1]ProprietaryN/A66.4
", + "bbox": [ + 537, + 89, + 883, + 160 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Table 7. Performance of different LLMs on EgoSchema.", + "bbox": [ + 540, + 170, + 877, + 185 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "the number of the initial key frames is set as $K = 5$ . Intuitively, more frames should lead to better performance at the cost of increasing the complexity. However, our ablation in Table 5b shows an opposite trend, i.e., more key frames resulting in worse performance. This suggests that enhancing more frames is not necessarily beneficial, as it can introduce a large amount of noisy and irrelevant information that interferes with the LLM's judgment.", + "bbox": [ + 511, + 200, + 903, + 321 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Effects of iterative rounds. We also examined the impact of the iterative rounds $I$ on model performance. Fig. 3 shows the performance of different iterative rounds with DrVideo and VideoAgent [61]. We have the following observations. (i) DrVideo consistently outperforms VideoAgent, reaching the peak performance at $I = 2$ , while VideoAgent peaks at $I = 3$ . This indicates that DrVideo can locate the information needed to answer questions in fewer interactions and with higher quality, underscoring the superiority of our document retrieval-based framework. (ii) When the number of interactions exceeds two, DrVideo's performance declines. This is likely due to more noisy and irrelevant information being introduced into the LLM with the increase of the iterations. This further confirms that excessive information is not necessarily beneficial.", + "bbox": [ + 511, + 323, + 905, + 549 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Effects of different foundation models. DrVideo incorporates two types of foundational models — large language model (LLM), and visual language model (VLM), where VLM is used as the frame captioner while LLM is used as the planning agent, the interaction agent, and the answering agent. To assess the impact of the captions produced by different VLMs, we examine three state-of-the-art VLMs: frame-based BLIP-2 [29], LLaVA-NeXT [33], and clip-based LaViLa [78], as presented in Table 6. We observe that LaViLa achieves the best performance, while BLIP-2 is the worst. It indicates the superiority of the clip-based model which can capture motion information in videos and is more suited for video understanding.", + "bbox": [ + 511, + 551, + 905, + 747 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "To evaluate the effectiveness of DrVideo with different LLMs, Table 7 shows the results of DrVideo using open-source LLM (Mistral-8x7B [24]) or Proprietary LLM (DeepSeek [19], GPT-3.5 [40], or GPT-4 [1]) as the planning, interaction and answering agents. From the results, we can see that: (i) different LLMs in DrVideo exhibit significant performance gaps and the proprietary LLM significantly outperforms the open-source LLM, indicating DrVideo's dependency on the choice of LLM; (ii) as the LLM's capability increases, the performance of DrVideo", + "bbox": [ + 511, + 750, + 905, + 900 + ], + "page_idx": 6 + }, + { + "type": "page_number", + "text": "18942", + "bbox": [ + 480, + 944, + 517, + 955 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Question: By analyzing c's activities in this video, determine their overall focus or intention, and describe the progression of tasks leading to its fulfillment.", + "bbox": [ + 96, + 92, + 883, + 116 + ], + "page_idx": 7 + }, + { + "type": "list", + "sub_type": "text", + "list_items": [ + "A. C's overall focus or intention in this video is to set up their work station and start working on their computer.", + "B. The central objective in this video, c's overall focus or intention, is to engage in conversation with the person.", + "C. C's overall focus or intention in this video is to eat the wrapper.", + "D. C's overall focus or primary intention in this specific video content is to effectively play the guitar", + "E. The primary objective and intention in this particular c's video is to demonstrate how to effectively clean the cloth." + ], + "bbox": [ + 99, + 117, + 808, + 175 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/50c85371c322e3c19aa59852da71c6437cf3c180a65671e89e2cf99d44417dea.jpg", + "image_caption": [ + "Figure 4. Case study on an instance from EgoSchema. DrVideo accurately identifies key frames and chooses the correct answer." + ], + "image_footnote": [], + "bbox": [ + 91, + 176, + 919, + 549 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "improves, demonstrating the great potential of DrVideo. It's worth noting that our main contribution is the introduction of a framework for translating the long video understanding into a document retrieval and understanding task to effectively leverage the power of large language models, rather than the employment of any specific models. Our DrVideo can be further improved by combining with better LLM and VLM models (e.g., tarsier [59]).", + "bbox": [ + 89, + 592, + 483, + 713 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Case study. Here we give a detailed example from EgoSchema [36] in Fig. 4 to qualitatively demonstrate the capability of our DrVideo. Specifically, the question is to determine the overall focus or intention and describe the progression of tasks leading to its fulfillment. DrVideo first converts the raw video into an initial video document, which is then processed by the frame retrieval module to obtain key frames (frames 15, 46, 61, 84, and 86) and is augmented with detailed key frame descriptions. The updated video document is then fed into the planning agent to assess if sufficient information is available for answering, i.e., the provided information is sufficient to infer character C's", + "bbox": [ + 89, + 719, + 483, + 900 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "main focus or intent. This updated document, along with the rationale, is input to the interaction agent to identify the additional key frames (frames 4, 5, 6, 14, and 19 here). After further augmentation, the updated document is sent back to the planning agent, for which the agent confirms its sufficiency. Finally, the answering module takes in the final video document and predicts the correct answer.", + "bbox": [ + 511, + 592, + 906, + 698 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "5. Conclusion", + "text_level": 1, + "bbox": [ + 511, + 709, + 633, + 724 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "We have proposed DrVideo, a document retrieval-based system for long video understanding. Different from previous LLM-based methods, DrVideo proposes to adapt long-video understanding to long-document understanding and searches for missing information via document retrieval and multi-stage agent loop. Extensive comparative experiments and ablation studies on the three challenging datasets, i.e., EgoSchema, MovieChat-1K, and Video-MME, demonstrate the effectiveness of DrVideo. We believe our work can serve as a strong baseline to stimulate further developments on the topic of long video understanding.", + "bbox": [ + 511, + 734, + 906, + 902 + ], + "page_idx": 7 + }, + { + "type": "page_number", + "text": "18943", + "bbox": [ + 480, + 944, + 517, + 955 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Acknowledgment", + "text_level": 1, + "bbox": [ + 91, + 90, + 240, + 107 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "References", + "text_level": 1, + "bbox": [ + 91, + 205, + 186, + 222 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[1] Josh Achiam, Steven Adler, Sandhini Agarwal, Lama Ahmad, Ilge Akkaya, Florencia Leoni Aleman, Diogo Almeida, Janko Altenschmidt, Sam Altman, Shyamal Anadkat, et al. Gpt-4 technical report. arXiv preprint arXiv:2303.08774, 2023. 2, 5, 6, 7", + "[2] Anthropic. Claude 3.5 Sonnet: Enhanced Conversational AI. https://www.anthropic.com/index/claude-3.5-sonnet/, 2024.6", + "[3] Jinze Bai, Shuai Bai, Shusheng Yang, Shijie Wang, Sinan Tan, Peng Wang, Junyang Lin, Chang Zhou, and Jingren Zhou. Qwen-vl: A versatile vision-language model for understanding, localization, text reading, and beyond. arXiv preprint arXiv:2308.12966, 2023. 6", + "[4] Aanisha Bhattacharya, Yaman K Singla, Balaji Krishnamurthy, Rajiv Ratn Shah, and Changyou Chen. A video is worth 4096 tokens: Verbalize story videos to understand them in zero shot. arXiv preprint arXiv:2305.09758, 2023. 3", + "[5] Anthony Brohan, Noah Brown, Justice Carbajal, Yevgen Chebotar, Xi Chen, Krzysztof Choromanski, Tianli Ding, Danny Driess, Avinava Dubey, Chelsea Finn, et al. Rt-2: Vision-language-action models transfer web knowledge to robotic control. arXiv preprint arXiv:2307.15818, 2023. 3", + "[6] Tom Brown, Benjamin Mann, Nick Ryder, Melanie Subbiah, Jared D Kaplan, Prafulla Dhariwal, Arvind Neelakantan, Pranav Shyam, Girish Sastry, Amanda Askell, et al. Language models are few-shot learners. Adv. Neural Inform. Process. Syst., pages 1877-1901, 2020. 1", + "[7] Jun Chen, Deyao Zhu, Kilichbek Haydarov, Xiang Li, and Mohamed Elhoseiny. Video chatcaptioner: Towards the enriched spatiotemporal descriptions. arXiv preprint arXiv:2304.04227, 2023. 3", + "[8] Feng Cheng, Xizi Wang, Jie Lei, David Crandall, Mohit Bansal, and Gedas Bertasius. Vindlu: A recipe for effective video-and-language pretraining. In IEEE/CVF Conf. Comput. Vis. Pattern Recog., pages 10739-10750, 2023. 1", + "[9] Rohan Choudhury, Koichiro Niinuma, Kris M Kitani, and László A Jeni. Zero-shot video question answering with procedural programs. arXiv preprint arXiv:2312.00937, 2023. 5, 6", + "[10] Aakanksha Chowdhery, Sharan Narang, Jacob Devlin, Maarten Bosma, Gaurav Mishra, Adam Roberts, Paul Barham, Hyung Won Chung, Charles Sutton, Sebastian Gehrmann, et al. Palm: Scaling language modeling with pathways. arXiv preprint arXiv:2204.02311, 2022. 1", + "[11] Jiwan Chung and Youngjae Yu. Long story short: a summarize-then-search method for long video question answering. In Brit. Mach. Vis. Conf., 2023. 3" + ], + "bbox": [ + 93, + 231, + 482, + 900 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "This work is supported by the National Natural Science Foundation of China (62221002, 62171183, U24A20327, 62401202), the Joint Fund of the Ministry of Education of China, and the China Scholarship Council (CSC) under Grant No. 202306130148.", + "[12] Yiran Ding, Li Lyna Zhang, Chengruidong Zhang, Yuanyuan Xu, Ning Shang, Jiahang Xu, Fan Yang, and Mao Yang. Longgrope: Extending llm context window beyond 2 million tokens. arXiv preprint arXiv:2402.13753, 2024. 2", + "[13] Danny Driess, Fei Xia, Mehdi SM Sajjadi, Corey Lynch, Aakanksha Chowdhery, Brian Ichter, Ayzaan Wahid, Jonathan Tompson, Quan Vuong, Tianhe Yu, et al. Palme: An embodied multimodal language model. arXiv preprint arXiv:2303.03378, 2023. 3", + "[14] Yue Fan, Xiaojian Ma, Rujie Wu, Yuntao Du, Jiaqi Li, Zhi Gao, and Qing Li. Videoagent: A memory-augmented multimodal agent for video understanding. arXiv preprint arXiv:2403.11481, 2024. 2, 3, 5, 6", + "[15] Chaoyou Fu, Yuhan Dai, Yondong Luo, Lei Li, Shuhuai Ren, Renrui Zhang, Zihan Wang, Chenyu Zhou, Yunhang Shen, Mengdan Zhang, et al. Video-mme: The first-ever comprehensive evaluation benchmark of multi-modal llms in video analysis. arXiv preprint arXiv:2405.21075, 2024. 2, 3, 5, 6", + "[16] Chaoyou Fu, Haojia Lin, Zuwei Long, Yunhang Shen, Meng Zhao, Yifan Zhang, Xiong Wang, Di Yin, Long Ma, Xiawu Zheng, et al. Vita: Towards open-source interactive omni multimodal llm. arXiv preprint arXiv:2408.05211, 2024. 6", + "[17] Difei Gao, Lei Ji, Luowei Zhou, Kevin Qinghong Lin, Joya Chen, Zihan Fan, and Mike Zheng Shou. Assistgpt: A general multi-modal assistant that can plan, execute, inspect, and learn. arXiv preprint arXiv:2306.08640, 2023. 3", + "[18] Albert Gu, Karan Goel, and Christopher Ré. Efficiently modeling long sequences with structured state spaces. arXiv preprint arXiv:2111.00396, 2021. 3", + "[19] Daya Guo, Qihao Zhu, Dejian Yang, Zhenda Xie, Kai Dong, Wentao Zhang, Guanting Chen, Xiao Bi, Yu Wu, YK Li, et al. Deepseek-coder: When the large language model meets programming-the rise of code intelligence. arXiv preprint arXiv:2401.14196, 2024. 5, 6, 7", + "[20] Wenyi Hong, Weihan Wang, Qingsong Lv, Jiazheng Xu, Wenmeng Yu, Junhui Ji, Yan Wang, Zihan Wang, Yuxiao Dong, Ming Ding, et al. Cogagent: A visual language model for gui agents. arXiv preprint arXiv:2312.08914, 2023. 3", + "[21] Noureldien Hussein, Efstratios Gavves, and Arnold WM Smeulders. Videograph: Recognizing minutes-long human activities in videos. arXiv preprint arXiv:1905.05143, 2019. 3", + "[22] Md Mohaiminul Islam and Gedas Bertasius. Long movie clip classification with state-space video models. In Eur. Conf. Comput. Vis., pages 87-104, 2022. 3", + "[23] Sullam Jeoung, Goeric Huybrechts, Bhavana Ganesh, Aram Galstyan, and Sravan Bodapati. Adaptive video understanding agent: Enhancing efficiency with dynamic frame sampling and feedback-driven reasoning. arXiv preprint arXiv:2410.20252, 2024. 5", + "[24] Albert Q Jiang, Alexandre Sablayrolles, Arthur Mensch, Chris Bamford, Devendra Singh Chaplot, Diego de las Casas, Florian Bressand, Gianna Lengyel, Guillaume Lample, Lucile Saulnier, et al. Mistral 7b. arXiv preprint arXiv:2310.06825, 2023. 7", + "[25] Wonkyun Kim, Changin Choi, Wonseok Lee, and Wonjong Rhee. An image grid can be worth a video: Zero" + ], + "bbox": [ + 93, + 92, + 903, + 900 + ], + "page_idx": 8 + }, + { + "type": "page_number", + "text": "18944", + "bbox": [ + 480, + 945, + 517, + 955 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "shot video question answering using a vlm. arXiv preprint arXiv:2403.18406, 2024. 1, 5, 6", + "[26] Takeshi Kojima, Shixiang Shane Gu, Machel Reid, Yutaka Matsuo, and Yusuke Iwasawa. Large language models are zero-shot reasoners. Adv. Neural Inform. Process. Syst., pages 22199-22213, 2022. 5", + "[27] Jie Lei, Linjie Li, Luowei Zhou, Zhe Gan, Tamara L Berg, Mohit Bansal, and Jingjing Liu. Less is more: Clipbert for video-and-language learning via sparse sampling. In IEEE/CVF Conf. Comput. Vis. Pattern Recog., pages 7331-7341, 2021. 1", + "[28] Bo Li, Yuanhan Zhang, Dong Guo, Renrui Zhang, Feng Li, Hao Zhang, Kaichen Zhang, Yanwei Li, Ziwei Liu, and Chunyuan Li. Llava-onevision: Easy visual task transfer. arXiv preprint arXiv:2408.03326, 2024. 6", + "[29] Junnan Li, Dongxu Li, Silvio Savarese, and Steven Hoi. BLIP-2: bootstrapping language-image pre-training with frozen image encoders and large language models. In Proc. Int. Conf. Mach. Learn., 2023. 4, 7", + "[30] KunChang Li, Yinan He, Yi Wang, Yizhuo Li, Wenhai Wang, Ping Luo, Yali Wang, Limin Wang, and Yu Qiao. Videochat: Chat-centric video understanding. arXiv preprint arXiv:2305.06355, 2023. 3, 5, 6", + "[31] Bin Lin, Bin Zhu, Yang Ye, Munan Ning, Peng Jin, and Li Yuan. Video-llava: Learning united visual representation by alignment before projection. arXiv preprint arXiv:2311.10122, 2023. 6", + "[32] Kevin Lin, Faisal Ahmed, Linjie Li, Chung-Ching Lin, Ehsan Azarnasab, Zhengyuan Yang, Jianfeng Wang, Lin Liang, Zicheng Liu, Yumao Lu, et al. Mm-vid: Advancing video understanding with gpt-4v (ision). arXiv preprint arXiv:2310.19773, 2023. 1, 3, 5", + "[33] Haotian Liu, Chunyuan Li, Yuheng Li, Bo Li, Yuanhan Zhang, Sheng Shen, and Yong Jae Lee. Llava next: Improved reasoning,OCR,and world knowledge.https://llava-v1.github.io/blog/2024-01-30-llava NEXT/,2024.2,4,5,6,7", + "[34] Ziyu Ma, Shutao Li, Bin Sun, Jianfei Cai, Zuxiang Long, and Fuyan Ma. Gerea: Question-aware prompt captions for knowledge-based visual question answering. arXiv preprint arXiv:2402.02503, 2024. 1", + "[35] Muhammad Maaz, Hanoona Rasheed, Salman Khan, and Fahad Shahbaz Khan. Video-chatgpt: Towards detailed video understanding via large vision and language models. arXiv preprint arXiv:2306.05424, 2023. 5", + "[36] Karttikeya Mangalam, Raiymbek Akshulakov, and Jitendra Malik. Egoschema: A diagnostic benchmark for very long-form video language understanding. arXiv preprint arXiv:2308.09126, 2023. 2, 3, 5, 6, 7, 8", + "[37] Juhong Min, Shyamal Buch, Arsha Nagrani, Minsu Cho, and Cordelia Schmid. Morevqa: Exploring modular reasoning models for video question answering. In IEEE/CVF Conf. Comput. Vis. Pattern Recog., pages 13235-13245, 2024. 2, 5, 6", + "[38] Salman Khan Muhammad Maaz, Hanoona Rasheed and Fahad Khan. Video-chatgpt: Towards detailed video understanding via large vision and language models. ArXiv 2306.05424, 2023. 1, 5, 6" + ], + "bbox": [ + 91, + 92, + 483, + 898 + ], + "page_idx": 9 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[39] Eric Nguyen, Karan Goel, Albert Gu, Gordon Downs, Preey Shah, Tri Dao, Stephen Baccus, and Christopher Ré. S4nd: Modeling images and videos as multidimensional signals with state spaces. Adv. Neural Inform. Process. Syst., pages 2846-2861, 2022. 3", + "[40] OpenAI. GPT3.5. https://platform.openai.com/docs/models/gpt-3-5, 2021.1, 5, 7", + "[41] OpenAI. Introducing chatgpt. https://openai.com/blog/chatgpt, 2022.1", + "[42] OpenAI. GPT-4V(ison) System Card. https://openai.com/index/gpt-4v-system-card/, 2023.3,6", + "[43] OpenAI. New embedding models and api updates. https://openai.com/index/new-embedding-models-and-api-updates/, 2024.4", + "[44] OpenAI. GPT-4o Mini: Advancing Cost-Efficient Intelligence. https://openai.com/index/gpt-4o-mini-advancing-cost-efficient-intelligence/, 2024.3,6", + "[45] Long Ouyang, Jeffrey Wu, Xu Jiang, Diogo Almeida, Carroll Wainwright, Pamela Mishkin, Chong Zhang, Sandhini Agarwal, Katarina Slama, Alex Ray, et al. Training language models to follow instructions with human feedback. Adv. Neural Inform. Process. Syst., pages 27730-27744, 2022. 1", + "[46] Vincent Perot, Kai Kang, Florian Luisier, Guolong Su, Xiaoyu Sun, Ramya Sree Boppana, Zilong Wang, Zifeng Wang, Jiaqi Mu, Hao Zhang, et al. Lmdx: Language model-based document information extraction and localization. arXiv preprint arXiv:2309.10952, 2023. 2", + "[47] Kanchana Ranasinghe, Xiang Li, Kumara Kahatapitiya, and Michael S Ryoo. Understanding long videos in one multimodal language model pass. arXiv preprint arXiv:2403.16998, 2024. 5, 6", + "[48] Teven Le Scao, Angela Fan, Christopher Akiki, Ellie Pavlick, Suzana Ilic, Daniel Hesslow, Roman Castagné, Alexandra Sasha Luccioni, François Yvon, Matthias Galle, et al. Bloom: A 176b-parameter open-access multilingual language model. arXiv preprint arXiv:2211.05100, 2022. 1", + "[49] Enxin Song, Wenhao Chai, Guanhong Wang, Yucheng Zhang, Haoyang Zhou, Feiyang Wu, Xun Guo, Tian Ye, Yan Lu, Jenq-Neng Hwang, et al. Moviechat: From dense token to sparse memory for long video understanding. arXiv preprint arXiv:2307.16449, 2023. 2, 3, 5, 6", + "[50] Enxin Song, Wenhao Chai, Tian Ye, Jenq-Neng Hwang, Xi Li, and Gaoang Wang. Moviechat+: Question-aware sparse memory for long video question answering. arXiv preprint arXiv:2404.17176, 2024. 5", + "[51] Quan Sun, Jinsheng Wang, Qiying Yu, Yufeng Cui, Fan Zhang, Xiaosong Zhang, and Xinlong Wang. Eva-clip-18b: Scaling clip to 18 billion parameters. arXiv preprint arXiv:2402.04252, 2024. 2", + "[52] Yuchong Sun, Hongwei Xue, Ruihua Song, Bei Liu, Huan Yang, and Jianlong Fu. Long-form video-language pretraining with multimodal temporal contrastive learning. Adv. Neural Inform. Process. Syst., 35:38032-38045, 2022. 3", + "[53] Dídac Surís, Sachit Menon, and Carl Vondrick. Vipergpt: Visual inference via python execution for reasoning. Int. Conf. Comput. Vis., pages 11888-11898, 2023. 1, 3" + ], + "bbox": [ + 516, + 92, + 903, + 900 + ], + "page_idx": 9 + }, + { + "type": "page_number", + "text": "18945", + "bbox": [ + 480, + 945, + 517, + 955 + ], + "page_idx": 9 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[54] Gemini Team, Rohan Anil, Sebastian Borgeaud, Yonghui Wu, Jean-Baptiste Alayrac, Jiahui Yu, Radu Soricut, Johan Schalkwyk, Andrew M Dai, Anja Hauth, et al. Gemini: a family of highly capable multimodal models. arXiv preprint arXiv:2312.11805, 2023. 5", + "[55] Gemini Team, Petko Georgiev, Ving Ian Lei, Ryan Burnell, Libin Bai, Anmol Gulati, Garrett Tanzer, Damien Vincent, Zhufeng Pan, Shibo Wang, et al. Gemini 1.5: Unlocking multimodal understanding across millions of tokens of context. arXiv preprint arXiv:2403.05530, 2024. 3, 6", + "[56] Hugo Touvron, Thibaut Lavril, Gautier Izacard, Xavier Martinet, Marie-Anne Lachaux, Timothée Lacroix, Baptiste Rozière, Naman Goyal, Eric Hambro, Faisal Azhar, et al. Llama: Open and efficient foundation language models. arXiv preprint arXiv:2302.13971, 2023. 1", + "[57] Junke Wang, Dongdong Chen, Chong Luo, Xiyang Dai, Lu Yuan, Zuxuan Wu, and Yu-Gang Jiang. Chatvideo: A tracklet-centric multimodal and versatile video understanding system. arXiv preprint arXiv:2304.14407, 2023. 3", + "[58] Jue Wang, Wentao Zhu, Pichao Wang, Xiang Yu, Linda Liu, Mohamed Omar, and Raffay Hamid. Selective structured state-spaces for long-form video understanding. In IEEE/CVF Conf. Comput. Vis. Pattern Recog., pages 6387-6397, 2023. 3", + "[59] Jiawei Wang, Liping Yuan, and Yuchen Zhang. Tarsier: Recipes for training and evaluating large video description models. arXiv preprint arXiv:2407.00634, 2024. 8", + "[60] Shijie Wang, Qi Zhao, Minh Quan Do, Nakul Agarwal, Kwonjoon Lee, and Chen Sun. Vamos: Versatile action models for video understanding. arXiv preprint arXiv:2311.13627, 2023. 5, 6", + "[61] Xiaohan Wang, Yuhui Zhang, Orr Zohar, and Serena Yeung-Levy. Videoagent: Long-form video understanding with large language model as agent. arXiv preprint arXiv:2403.10517, 2024. 1, 2, 3, 5, 6, 7", + "[62] Yang Wang, Gedas Bertasius, Tae-Hyun Oh, Abhinav Gupta, Minh Hoai, and Lorenzo Torresani. Supervoxel attention graphs for long-range video modeling. In Proc. IEEE/CVF Winter Conf. Appl. Comput. Vis., pages 155-166, 2021. 3", + "[63] Yi Wang, Yinan He, Yizhuo Li, Kunchang Li, Jiashuo Yu, Xin Ma, Xinhao Li, Guo Chen, Xinyuan Chen, Yaohui Wang, et al. Intervid: A large-scale video-text dataset for multimodal understanding and generation. arXiv preprint arXiv:2307.06942, 2023. 6", + "[64] Zhenhailong Wang, Manling Li, Ruochen Xu, Luowei Zhou, Jie Lei, Xudong Lin, Shuohang Wang, Ziyi Yang, Chenguang Zhu, Derek Hoiem, et al. Language models with image descriptors are strong few-shot video-language learners. Adv. Neural Inform. Process. Syst., pages 8483-8497, 2022. 3", + "[65] Ziyang Wang, Shoubin Yu, Elias Stengel-Eskin, Jaehong Yoon, Feng Cheng, Gedas Bertasius, and Mohit Bansal. Videotree: Adaptive tree-based video representation for llm reasoning on long videos. arXiv preprint arXiv:2405.19209, 2024.3", + "[66] Jason Wei, Xuezhi Wang, Dale Schuurmans, Maarten Bosma, Fei Xia, Ed Chi, Quoc V Le, Denny Zhou, et al." + ], + "bbox": [ + 91, + 90, + 485, + 900 + ], + "page_idx": 10 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "Chain-of-thought prompting elicits reasoning in large language models. Adv. Neural Inform. Process. Syst., pages 24824-24837, 2022. 5", + "[67] Chao-Yuan Wu, Yanghao Li, Karttikeya Mangalam, Haoqi Fan, Bo Xiong, Jitendra Malik, and Christoph Feichtenhofer. Memvit: Memory-augmented multiscale vision transformer for efficient long-term video recognition. In IEEE/CVF Conf. Comput. Vis. Pattern Recog., pages 13587-13597, 2022. 3", + "[68] Lin Xu, Yilin Zhao, Daquan Zhou, Zhijie Lin, See Kiong Ng, and Jiashi Feng. Pllava: Parameter-free llava extension from images to videos for video dense captioning. arXiv preprint arXiv:2404.16994, 2024. 1", + "[69] Antoine Yang, Antoine Miech, Josef Sivic, Ivan Laptev, and Cordelia Schmid. Just ask: Learning to answer questions from millions of narrated videos. In Int. Conf. Comput. Vis., pages 1686-1697, 2021. 1", + "[70] Antoine Yang, Antoine Miech, Josef Sivic, Ivan Laptev, and Cordelia Schmid. Zero-shot video question answering via frozen bidirectional language models. Adv. Neural Inform. Process. Syst., 35:124-141, 2022. 1", + "[71] Xitong Yang, Fu-Jen Chu, Matt Feiszli, Raghav Goyal, Lorenzo Torresani, and Du Tran. Relational space-time query in long-form videos. In IEEE/CVF Conf. Comput. Vis. Pattern Recog., pages 6398-6408, 2023. 3", + "[72] Zongxin Yang, Guikun Chen, Xiaodi Li, Wenguan Wang, and Yi Yang. Doraemongpt: Toward understanding dynamic scenes with large language models. arXiv preprint arXiv:2401.08392, 2024. 3", + "[73] Andy Zeng, Maria Attarian, Brian Ichter, Krzysztof Choromanski, Adrian Wong, Stefan Welker, Federico Tombari, Aveek Purohit, Michael Ryoo, Vikas Sindhwani, Johnny Lee, Vincent Vanhoucke, and Pete Florence. Socratic models: Composing zero-shot multimodal reasoning with language. arXiv preprint arXiv:2204.00598, 2022. 3", + "[74] Ce Zhang, Taixi Lu, Md Mohaiminul Islam, Ziyang Wang, Shoubin Yu, Mohit Bansal, and Gedas Bertasius. A simple llm framework for long-range video question-answering. arXiv preprint arXiv:2312.17235, 2023. 1, 2, 5, 6, 7", + "[75] Hang Zhang, Xin Li, and Lidong Bing. Video-llama: An instruction-tuned audio-visual language model for video understanding. arXiv preprint arXiv:2306.02858, 2023. 5, 6", + "[76] Peiyuan Zhang, Kaichen Zhang, Bo Li, Guangtao Zeng, Jingkang Yang, Yuanhan Zhang, Ziyue Wang, Haoran Tan, Chunyuan Li, and Ziwei Liu. Long context transfer from language to vision. arXiv preprint arXiv:2406.16852, 2024. 6", + "[77] Yuanhan Zhang, Bo Li, Haotian Liu, Yong Jae Lee, Liangke Gui, Di Fu, Jiashi Feng, Ziwei Liu, and Chunyuan Li. Llava-last: A strong zero-shot video understanding model. https://llava-v1.github.io/blog/2024-04-30-llava-last-video/, 2024.1", + "[78] Yue Zhao, Ishan Misra, Philipp Krahenbuhl, and Rohit Girdhar. Learning video representations from large language models. In IEEE/CVF Conf. Comput. Vis. Pattern Recog., pages 6586-6597, 2023. 1, 2, 4, 5, 6, 7" + ], + "bbox": [ + 516, + 92, + 903, + 869 + ], + "page_idx": 10 + }, + { + "type": "page_number", + "text": "18946", + "bbox": [ + 480, + 945, + 519, + 955 + ], + "page_idx": 10 + } +] \ No newline at end of file diff --git a/2025/DrVideo_ Document Retrieval Based Long Video Understanding/81160fea-84bd-416d-98f4-7b4c205b85c0_model.json b/2025/DrVideo_ Document Retrieval Based Long Video Understanding/81160fea-84bd-416d-98f4-7b4c205b85c0_model.json new file mode 100644 index 0000000000000000000000000000000000000000..63f174b5fd4b1b40782715cb1aa9f8d35e1776a5 --- /dev/null +++ b/2025/DrVideo_ Document Retrieval Based Long Video Understanding/81160fea-84bd-416d-98f4-7b4c205b85c0_model.json @@ -0,0 +1,2466 @@ +[ + [ + { + "type": "header", + "bbox": [ + 0.107, + 0.003, + 0.182, + 0.043 + ], + "angle": 0, + "content": "CVF" + }, + { + "type": "header", + "bbox": [ + 0.237, + 0.001, + 0.812, + 0.047 + ], + "angle": 0, + "content": "This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore." + }, + { + "type": "title", + "bbox": [ + 0.175, + 0.131, + 0.825, + 0.154 + ], + "angle": 0, + "content": "DrVideo: Document Retrieval Based Long Video Understanding" + }, + { + "type": "text", + "bbox": [ + 0.238, + 0.18, + 0.757, + 0.217 + ], + "angle": 0, + "content": "Ziyu Ma\\(^{1,2}\\), Chenhui Gou\\(^{2}\\), Hengcan Shi\\(^{1,2}\\), Bin Sun\\(^{1}\\), Shutao Li\\(^{1}\\), Hamid Rezatofighi\\(^{2}\\), Jianfei Cai\\(^{2}\\)" + }, + { + "type": "text", + "bbox": [ + 0.218, + 0.217, + 0.78, + 0.235 + ], + "angle": 0, + "content": "1 College of Electrical and Information Engineering, Hunan University," + }, + { + "type": "text", + "bbox": [ + 0.234, + 0.235, + 0.765, + 0.253 + ], + "angle": 0, + "content": "2 Data Science & AI Department, Faculty of IT, Monash University" + }, + { + "type": "text", + "bbox": [ + 0.32, + 0.255, + 0.678, + 0.269 + ], + "angle": 0, + "content": "{maziyu, sunbin611, shutao_li}@hnu.edu.cn" + }, + { + "type": "text", + "bbox": [ + 0.192, + 0.273, + 0.802, + 0.288 + ], + "angle": 0, + "content": "{chenhui.gou, hengcan.shi, hamid.rezatofighi, jianfei.cai}@monash.edu" + }, + { + "type": "title", + "bbox": [ + 0.248, + 0.322, + 0.327, + 0.339 + ], + "angle": 0, + "content": "Abstract" + }, + { + "type": "text", + "bbox": [ + 0.089, + 0.354, + 0.486, + 0.732 + ], + "angle": 0, + "content": "Most of the existing methods for video understanding primarily focus on videos only lasting tens of seconds, with limited exploration of techniques for handling long videos. The increased number of frames in long videos poses two main challenges: difficulty in locating key information and performing long-range reasoning. Thus, we propose DrVideo, a document-retrieval-based system designed for long video understanding. Our key idea is to convert the long-video understanding problem into a long-document understanding task so as to effectively leverage the power of large language models. Specifically, DrVideo first transforms a long video into a coarse text-based long document to initially retrieve key frames and then updates the documents with the augmented key frame information. It then employs an agent-based iterative loop to continuously search for missing information and augment the document until sufficient question-related information is gathered for making the final predictions in a chain-of-thought manner. Extensive experiments on long video benchmarks confirm the effectiveness of our method. DrVideo significantly outperforms existing LLM-based state-of-the-art methods on EgoSchema benchmark (3 minutes), MovieChat-1K benchmark (10 minutes), and the long split of Video-MME benchmark (average of 44 minutes). Code is available at https://github.com/Upper9527/DrVideo." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.763, + 0.222, + 0.779 + ], + "angle": 0, + "content": "1. Introduction" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.788, + 0.485, + 0.866 + ], + "angle": 0, + "content": "Video understanding is a challenging task in computer vision, requiring the processing of spatio-temporal information and advanced reasoning abilities. Previous works have successfully processed short videos lasting around tens of seconds [8, 27, 32, 53, 69, 70]. However, how to deal with" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.324, + 0.906, + 0.43 + ], + "angle": 0, + "content": "long video understanding remains unclear. Recent advancements in large language models (LLMs) [6, 10, 34, 41, 45, 48, 56] demonstrate strong abilities in language understanding and reasoning across long text sequences. These advancements have inspired the development of videolanguage models (Video-LMs) [25, 38, 68, 77] to address long video understanding issues." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.431, + 0.907, + 0.688 + ], + "angle": 0, + "content": "Video-LMs typically encode a video as a sequence of visual tokens, which are then concatenated with language tokens into a long sequence and LLM is then used to understand this long sequence. Although Video-LMs improve video understanding, they have the following limitations. (i) They cannot process an entire video as input to a VideoLM (e.g., the visual encoder OpenAI's CLIP-L-14 outputs \\(24 \\times 24\\) tokens for each image, while the max length of LLaVA-NeXT-Video [77] is only 8192). (ii) They typically perform uniform or random frame sampling at large strides to handle long videos without considering the content (e.g., 16 for PLLaVA [68] and 6 for IG-VLM [25]), leading to potential key information loss. (iii) The simple concatenation of visual tokens increases the difficulty for the LLM to locate question-related (key) visual information within the long video sequence, complicating long-range reasoning across the vision token sequences." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.689, + 0.909, + 0.903 + ], + "angle": 0, + "content": "Apart from the studies of encoding a video as a sequence of visual tokens, another line of research converts raw videos into captions and then utilities the long-range reasoning abilities of LLMs to predict the final answer. A pioneering work - LLoVi [74] uses a short-term video model [78] in conjunction with an LLM [40] to solve the long video understanding task. Given a long video, LLoVi first divides it into multiple short clips and converts them into short textual descriptions. Afterward, these short textual descriptions are summarized by the LLM, and the summary is finally used by LLoVi to answer the given question. Inspired by LLoVi, VideoAgent [61] designs an agent-based system to further leverage the reasoning abilities of LLMs and locate question-related video clips or images. Initially," + }, + { + "type": "page_footnote", + "bbox": [ + 0.109, + 0.875, + 0.221, + 0.888 + ], + "angle": 0, + "content": "*Equal contribution." + }, + { + "type": "page_footnote", + "bbox": [ + 0.111, + 0.888, + 0.236, + 0.9 + ], + "angle": 0, + "content": "† Corresponding author." + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.52, + 0.957 + ], + "angle": 0, + "content": "18936" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.137, + 0.087, + 0.862, + 0.296 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.09, + 0.308, + 0.908, + 0.339 + ], + "angle": 0, + "content": "Figure 1. Overview of our DrVideo framework. It comprises five components: a video-document conversion module, a retrieval module, a document augmentation module, a multi-stage agent interaction loop and an answering module." + }, + { + "type": "text", + "bbox": [ + 0.089, + 0.348, + 0.486, + 0.515 + ], + "angle": 0, + "content": "VideoAgent samples only a minimal number of frames from the video (i.e., 5 frames), converting them into textual descriptions as input for the LLM. Then, through continuous feedback and interaction with the LLM [1], VLM [78], and CLIP [51], it gradually finds question-related frames in a coarse-to-fine manner via image-text similarity. Finally, VideoAgent answers the question based on these question-related frames. Despite these LLM-based methods overcoming the challenge of performing long-range reasoning and achieving promising results, they still suffer from the following problems." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.524, + 0.486, + 0.75 + ], + "angle": 0, + "content": "- These methods struggle to accurately identify question-related frames. For instance, VideoAgent [61] performs sparse sampling on long videos (e.g., selecting 5 frames) and subsequently searches for keyframes from the unsampled frames in a coarse-to-fine manner based on the initial sparse space via CLIP image-text similarly given the missing question-related information inferred by the LLM, which improves inference efficiency but unavoidably at the cost of lacking a holistic grasp of video content, particularly when a long video consists of multiple events (e.g., over 100 events) with significant content variations, and since the CLIP text encoder is trained with simple language structures and limited context (the max length is 77), it struggles to understand complex video semantics, limiting its retrieval accuracy." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.75, + 0.486, + 0.902 + ], + "angle": 0, + "content": "- These methods convert videos into captions, addressing long-range reasoning but overlooking critical information loss. Specifically, even if question-related frames are identified, the generated captions cannot cover all the essential information in the key frames. For example, the caption \"a woman is looking in the mirror\" does not contribute to answering the question \"what is the woman wearing when standing in front of the mirror in the video\". In this case, LLM can only make an aimless and biased surmise to answer the question." + }, + { + "type": "list", + "bbox": [ + 0.09, + 0.524, + 0.486, + 0.902 + ], + "angle": 0, + "content": null + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.348, + 0.909, + 0.606 + ], + "angle": 0, + "content": "To solve the above problem, we turn to how human understands a long video. In particular, a human needs to gain a general understanding of the video, identifying events across segments. Then, the human locates question-related segments, examines them in detail, and answers the question based on the extracted information. Inspired by this, in this paper, we propose DrVideo, a document retrieval-based system for long video understanding that progressively finds key frames and generates a question-related augmented document for LLMs to make confident predictions. Our key idea is to convert long video understanding into long document understanding and leverage the impressive capabilities of LLMs in key information location [46] and long-range reasoning [12] on various NLP tasks. Note that long-document retrieval in our paper means passage retrieval within a single long document, not retrieval from a corpus consisting of long documents." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.609, + 0.909, + 0.822 + ], + "angle": 0, + "content": "Specifically, DrVideo transforms raw long videos into long documents using a pretrained frame/clip-level visual captioner (e.g., LLaVA-NeXT [33]) to make the LLM have a holistic grasp of video content. Different from VideoAgent [61], which performs sparse sampling and identifies key frames via image-text similarity without information augmentation, DrVideo performs dense sampling and proposes a retrieval module and a multi-stage agent interaction loop that dynamically finds the potential missing information through sequential steps of reasoning and then interacts with a document augmentation module to request augmentation of the information within language space. Once sufficient question-related information is gathered, it feeds into the answering module to get the final prediction." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.825, + 0.911, + 0.903 + ], + "angle": 0, + "content": "DrVideo achieves significant performance gains (3.8% on EgoSchema subset [36], 8.3% on MovieChat-1K global mode, 24.8% on MovieChat-1K breakpoint mode [49], and 6.3% on the long split of Video-MME [15]) over the existing LLM-based state-of-the-art methods [14, 37, 61, 74]" + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.52, + 0.957 + ], + "angle": 0, + "content": "18937" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.09, + 0.092, + 0.482, + 0.242 + ], + "angle": 0, + "content": "across different long video benchmarks, from 3 minutes to 10 minutes and longer than 1 hour. Surprisingly, combining the subtitles on the long split of Video-MME benchmark, DrVideo obtains \\(71.7\\%\\) accuracy, outperforming many heavily-engineered large-scale proprietary models (e.g., Gemini 1.5 Flash [55], GPT-4o mini [44], and GPT-4V [42]). This highlights the great potential of DrVideo. Moreover, DrVideo is a training-free framework and is researcher-friendly, as our results can be replicated on an RTX 4090 with a reasonable number of GPT accesses." + }, + { + "type": "text", + "bbox": [ + 0.11, + 0.243, + 0.46, + 0.257 + ], + "angle": 0, + "content": "The contribution of this work can be summarized as:" + }, + { + "type": "text", + "bbox": [ + 0.091, + 0.258, + 0.483, + 0.332 + ], + "angle": 0, + "content": "- We propose DrVideo, the first document-retrieval-based system designed for the long video understanding task by converting it into a long-document retrieval and understanding task to effectively leverage the power of large language models." + }, + { + "type": "text", + "bbox": [ + 0.091, + 0.334, + 0.483, + 0.393 + ], + "angle": 0, + "content": "- Different from previous works, we propose a retrieval module and a new multi-stage agent interaction loop that dynamically finds the potential missing information and augments these information within language space." + }, + { + "type": "text", + "bbox": [ + 0.091, + 0.394, + 0.483, + 0.455 + ], + "angle": 0, + "content": "- DrVideo outperforms the existing LLM-based state-of-the-art methods across different long video benchmarks (i.e., EgoSchema [36], MovieChat-1K [49], and the long split of Video-MME [15]) by a margin." + }, + { + "type": "list", + "bbox": [ + 0.091, + 0.258, + 0.483, + 0.455 + ], + "angle": 0, + "content": null + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.468, + 0.233, + 0.484 + ], + "angle": 0, + "content": "2. Related Work" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.493, + 0.483, + 0.765 + ], + "angle": 0, + "content": "Long Video Understanding. Modeling long videos, which are several minutes or more in length, generally requires advanced temporal modeling, resulting in complex model designs. LF-VILA [52] introduces a Temporal Window Attention (HTWA) mechanism to capture long-range dependencies in long videos. MeMViT [67] and MovieChat [49] employ a memory-based design to save question-related information from previous video segments. Some other approaches employ space-time graphs [21, 62] or relational space-time modules [71] to capture spatio-temporal dependencies from the raw long videos. Recently, S4ND [39], ViS4mer [22], and S5 [58] have used Structured State-Space Sequence (S4) [18] layers to capture long-range dependencies in videos. Different from these methods, our DrVideo does not design a complex module to perform long video understanding. Instead, we develop a document retrieval-based system with LLM for zero-shot long video understanding." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.766, + 0.483, + 0.902 + ], + "angle": 0, + "content": "LLMs for Video Understanding. The recent rise of LLMs [73] and VideoChat [30] align the visual features extracted by pretrained visual models to LLMs and apply them to video understanding. Video ChatCaptioner [7] and ChatVideo [57] utilize LLMs to represent videos and engage users through dialogues, respectively. VidIL [64] applies the image-level models to video understanding tasks via few-shot learning. In addition to short-term video understanding, recent studies [4, 11, 32] have explored LLMs" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.092, + 0.905, + 0.198 + ], + "angle": 0, + "content": "for long-range video modeling. For instance, GPT-4 is applied in various long-range video modeling tasks in [32], though quantitative evaluation is limited. Meanwhile, the research in [11] focuses on movie datasets with minimal visual analysis [36], relying largely on speech and subtitles. In contrast, DrVideo focuses on vision modality in multiple benchmarks." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.202, + 0.907, + 0.369 + ], + "angle": 0, + "content": "LLM Agents. In parallel, the computer vision community has started exploring the use of LLMs as agents in various vision tasks such as GUI understanding and robot navigation [5, 13, 20, 53]. In the realm of long video comprehension, initial efforts have employed an agent-like approach, where LLMs interact with external tools or integrate additional functionalities [14, 17, 53, 61, 65, 72]. In contrast to these approaches, our method, DrVideo, reconceptualizes long video understanding as a process of document retrieval, augmentation, and understanding, to leverage the strong capability of LLMs." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.398, + 0.649, + 0.415 + ], + "angle": 0, + "content": "3. Methodology" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.428, + 0.907, + 0.625 + ], + "angle": 0, + "content": "In this section, we detail our proposed document retrieval-based system for long video understanding. As illustrated in Figure 1, given a long video and a question about the video, DrVideo first translates the long video into a long document, referred to as the initial video document. Then, the retrieval module identifies the top \\( K \\) key frames by calculating the similarity between the question and the initial video document. The document augmentation module enriches the information of these key frames and adds it to the initial video document, creating a new video document (named the updated video document), which is used as the starting point for the subsequent multi-stage agent interaction loop." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.629, + 0.907, + 0.901 + ], + "angle": 0, + "content": "The loop contains two distinct agents: the planning agent and the interaction agent. The planning agent judges whether the updated video document is sufficient for answering the question. If not, the updated video document is fed into the interaction agent to dynamically find missing key frames. The interaction agent interacts with the document augmentation module to request the required information of these key frames. The output of the document augmentation module is added to the current document to obtain the latest updated video document. The new video document is then sent to the planning agent again to further judge whether the current information is sufficient. This searching and interaction loop continues until the planning agent considers the current information sufficient for answering the question or the maximum iteration is reached. After the loop ends, the final video document and the related question are given to the answering module to get the final prediction." + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.946, + 0.519, + 0.956 + ], + "angle": 0, + "content": "18938" + } + ], + [ + { + "type": "title", + "bbox": [ + 0.091, + 0.091, + 0.411, + 0.106 + ], + "angle": 0, + "content": "3.1. Video-Document Conversion Module" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.113, + 0.483, + 0.158 + ], + "angle": 0, + "content": "By translating each single video frame \\( V_{t} \\) (or a short video clip) into a short description \\( S_{V_t} \\), we convert an input long video into the initial video document \\( Doc_{init} \\):" + }, + { + "type": "equation", + "bbox": [ + 0.13, + 0.169, + 0.483, + 0.187 + ], + "angle": 0, + "content": "\\[\nD o c _ {i n i t} = \\left\\{\\left\\{1, S _ {V _ {1}} \\right\\}, \\left\\{2, S _ {V _ {2}} \\right\\}, \\dots , \\left\\{T, S _ {V _ {T}} \\right\\} \\right\\} \\tag {1}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.196, + 0.484, + 0.332 + ], + "angle": 0, + "content": "where \\(T\\) is the length of the video. Specifically, \\(\\mathrm{S}_{V_t}\\) is generated by a large vision-language captioning model \\(\\phi_{vlm}\\) (e.g., LLaVA-NeXT [33]), i.e., \\(\\mathrm{S}_{V_t} = \\phi_{vlm}(\\mathcal{P}, V_t)\\), where \\(\\mathcal{P}\\) is the prompt for requesting a short description about the image, e.g., describe the picture in no more than 50 words. The captioning model used in DrVideo can be replaced by other captioning models (e.g., LaViLa [78], BLIP-2 [29]). The experimental results with different captioning models can be found in Table 6." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.341, + 0.419, + 0.357 + ], + "angle": 0, + "content": "3.2. Document (Frames) Retrieval Module" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.363, + 0.484, + 0.514 + ], + "angle": 0, + "content": "After obtaining \\(Doc_{init}\\), we introduce a document retrieval module to identify the top question-related frames by calculating the similarity between the question and the whole document. Specifically, we use the OpenAI embedding model \\(\\phi_{emb}\\) [43] to obtain the vector representation of the initial document, i.e., \\(\\mathcal{E}_{\\mathrm{doc}} = \\phi_{emb}(Doc_{init})\\). Given a specific retrieval text \\(\\mathcal{RT}\\), i.e., the question \\(Q\\), the retrieval module first computes the embedding \\(\\mathcal{E}_{\\mathcal{RT}} = \\phi_{emb}(\\mathcal{RT})\\) and then retrieves the top \\(K\\) frames based on the cosine similarity between \\(\\mathcal{E}_{\\mathcal{RT}}\\) and \\(\\mathcal{E}_{\\mathrm{doc}}\\):" + }, + { + "type": "equation", + "bbox": [ + 0.149, + 0.517, + 0.483, + 0.542 + ], + "angle": 0, + "content": "\\[\n\\operatorname {t o p k} _ {\\text {一}} \\operatorname {d o c} = \\underset {t} {\\arg \\operatorname {t o p}} K \\cos \\left(\\mathcal {E} _ {\\mathcal {R T}}, \\mathcal {E} _ {\\mathrm {d o c} _ {t}}\\right) \\tag {2}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.545, + 0.466, + 0.561 + ], + "angle": 0, + "content": "where \\(\\mathcal{E}_{\\mathrm{doc}_t}\\) is the embedding of the \\(t\\)-th description \\(S_{V_t}\\)." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.569, + 0.383, + 0.584 + ], + "angle": 0, + "content": "3.3. Document Augmentation Module" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.591, + 0.484, + 0.71 + ], + "angle": 0, + "content": "For each \\(t' \\in \\text{topk\\_doc}\\), we use the LLaVA-NeXT model [33] with different prompts \\(\\mathcal{AP}\\) (augmented prompts) to generate a detailed description \\(L_{V_{t'}}\\), i.e., \\(L_{V_{t'}} = \\phi_{vlm}(\\mathcal{AP}_{t'}, V_{t'})\\). The initial \\(\\mathcal{AP}_{t'}\\) is a general prompt: If there are factual errors in the question, provide a precise description of the image; if not, proceed to answer the question: \\(\\{Q\\}\\). The updated video document \\(\\mathcal{AD}\\) becomes" + }, + { + "type": "equation", + "bbox": [ + 0.117, + 0.719, + 0.483, + 0.756 + ], + "angle": 0, + "content": "\\[\n\\mathcal {A D} = \\left\\{\\left\\{ \\begin{array}{l l} \\left\\{t, S _ {V _ {t}}, L _ {V _ {t} ^ {\\prime}} \\right\\} & \\text {i f} t = t ^ {\\prime} \\\\ \\left\\{t, S _ {V _ {t}} \\right\\} & \\text {o t h e r w i s e} \\end{array} \\right| t = 1, \\dots , T \\right\\}. \\tag {3}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.765, + 0.483, + 0.825 + ], + "angle": 0, + "content": "Note that \\( K \\) is significantly smaller than the total number of frames (e.g., \\( K = 5 \\) compared to \\( T = 90 \\)), so the additional descriptions will not substantially increase the overall length of the document." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.834, + 0.405, + 0.851 + ], + "angle": 0, + "content": "3.4. Multi-Stage Agent Interaction Loop" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.856, + 0.484, + 0.902 + ], + "angle": 0, + "content": "Apart from the retrieved key frames, we further introduce a multi-stage agent interaction loop to dynamically find other potential key frames and augment them with different types" + }, + { + "type": "image", + "bbox": [ + 0.518, + 0.091, + 0.897, + 0.313 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.513, + 0.325, + 0.907, + 0.395 + ], + "angle": 0, + "content": "Figure 2. Illustration of the multi-stage agent interaction loop and answering module. There are two agents in the multi-stage agent interaction loop: a planning agent to plan the next step and an interaction agent to dynamically find missing information and interact with the document augmentation module." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.406, + 0.906, + 0.481 + ], + "angle": 0, + "content": "of information for the final answer. We employ two distinct agents: planning and interaction agents, each tailored for a specific task. Both agents utilize the same LLM, i.e., GPT-3.5, but are given different prompts and reason varying types of information. The detail is shown in Fig. 2" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.483, + 0.906, + 0.648 + ], + "angle": 0, + "content": "Planning Agent. Given the question \\( Q \\) and the updated video document \\( \\mathcal{AD}_i \\) (where \\( i \\) denotes the iteration step), along with the analysis history \\( \\mathcal{H} \\) from all previous steps (initialized as empty \\( \\{\\} \\)), the planning agent first determines whether the updated video document is sufficient to generate a confident answer. If the information is sufficient, the process moves to the answering module to get the final answer. If not, the agent provides an analysis of why the updated video document \\( \\mathcal{AD}_i \\) is inadequate. The agent then updates \\( \\mathcal{H}_i \\) with its analysis and passes the updated \\( \\mathcal{H}_i \\) to the interaction agent." + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.649, + 0.906, + 0.695 + ], + "angle": 0, + "content": "Interaction Agent. Given the current video document \\(\\mathcal{AD}_i\\) and the updated \\(\\mathcal{H}_i\\), the interaction agent finds \\(N\\) frames with missing details crucial for answering \\(Q\\):" + }, + { + "type": "equation", + "bbox": [ + 0.539, + 0.706, + 0.905, + 0.723 + ], + "angle": 0, + "content": "\\[\n\\forall n \\in N, n \\in T, n \\notin \\text {t o p k} _ {-} \\text {d o c}, \\text {a n d} N < K. \\tag {4}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.735, + 0.907, + 0.902 + ], + "angle": 0, + "content": "In addition to identifying the missing \\(N\\) key frames, the interaction agent also determines the augmented information type for each \\(n \\in N\\). To achieve it, we give a task-specific prompt to the interaction agent, i.e., Your task is to determine which frame needs which type of information and can answer this question accurately, reasonably, and without contradiction. The two types of information are as follows: A: Given an image, get a detailed description of the image (image caption, just like what is shown in this image?) B: Given an image, get a response to the above question (visual question answering). After obtaining \\(N\\)" + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "18939" + } + ], + [ + { + "type": "table", + "bbox": [ + 0.126, + 0.089, + 0.451, + 0.276 + ], + "angle": 0, + "content": "
Method(M)LLMSubsetFullset
MoReVQA [37]PaLM-2-51.7
Vamos [60]GPT-451.248.3
ProViQ [9]GPT-3.557.1-
LLoVi [74]GPT-3.557.650.3
IG-VLM [25]GPT-4V59.8-
VideoAgent [61]GPT-460.254.1
MVU [47]Mistral-13B60.337.6
LLoVi [74]GPT-461.2-
VideoAgent [14]GPT-462.860.2
DrVideo (ours)GPT-3.562.6-
DrVideo (ours)GPT-466.461.0
" + }, + { + "type": "table_caption", + "bbox": [ + 0.09, + 0.286, + 0.483, + 0.315 + ], + "angle": 0, + "content": "Table 1. Results on EgoSchema compared to existing LLM-based state-of-the-art methods." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.327, + 0.484, + 0.388 + ], + "angle": 0, + "content": "frames along with their required information type, the agent interacts with the document augmentation module to enrich the information of these frames to update the current video document for the next step." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.398, + 0.274, + 0.414 + ], + "angle": 0, + "content": "3.5. Answering Module" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.419, + 0.484, + 0.542 + ], + "angle": 0, + "content": "Given the final video document \\(\\mathcal{AD}_{final}\\), we employ another agent to provide a prediction using a zero-shot chain-of-thought (CoT) approach [26, 66]. The answering module outputs the corresponding answer, the confidence score, and the reasoning behind the answer. Besides improving prediction accuracy, the CoT approach allows us to trace inference steps, ensuring transparency and explainability in the decision-making process." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.555, + 0.224, + 0.572 + ], + "angle": 0, + "content": "4. Experiments" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.581, + 0.483, + 0.627 + ], + "angle": 0, + "content": "In this section, we first introduce the datasets and implementation details and then present the results and ablations of our proposed DrVideo." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.637, + 0.291, + 0.652 + ], + "angle": 0, + "content": "4.1. Datasets and Metrics" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.659, + 0.483, + 0.704 + ], + "angle": 0, + "content": "In our experiments, we evaluate our DrVideo using three well-established datasets, emphasizing its zero-shot long video understanding capabilities." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.705, + 0.483, + 0.779 + ], + "angle": 0, + "content": "EgoSchema. EgoSchema dataset [36] consists of 5000 multiple-choice questions sourced from 5000 three-minute egocentric videos. This dataset only has a subset of 500 questions with publicly accessible labels, while the full set is evaluated on the leaderboard." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.78, + 0.483, + 0.902 + ], + "angle": 0, + "content": "MovieChat-1K. MovieChat-1K [49] is a longer video understanding benchmark, which contains 1000 videos from movies and TV shows, each video lasting approximately 10 minutes. This dataset has two modes: global mode and breakpoint mode. The global mode refers to analyzing the entire video to understand its overall content and context, while the breakpoint mode focuses on analyzing specific frames or scenes." + }, + { + "type": "table", + "bbox": [ + 0.518, + 0.089, + 0.905, + 0.296 + ], + "angle": 0, + "content": "
MethodGlobalBreakpoint
Acc.ScoreAcc.Score
Open-Source MLLM
Video LLaMA [75]51.72.7239.12.11
Video Chat [30]57.83.0846.12.32
Video-ChatGPT [35]47.62.8948.02.43
MovieChat [49]68.33.1548.32.46
MovieChat+ [50]71.23.5149.62.62
Based on Open-source Captioners and Proprietary LLMs
MM-VID [32]58.62.8610.40.56
LLoVi [74]58.32.8717.81.03
\\( VideoAgent^* \\) [61]65.4*3.12*31.6*2.05*
Sullam Jeoung, et al [23]84.8---
DrVideo (ours)93.14.4156.42.75
" + }, + { + "type": "table_caption", + "bbox": [ + 0.513, + 0.306, + 0.905, + 0.349 + ], + "angle": 0, + "content": "Table 2. Performance comparison on the MoiveChat-1K [49] benchmark against state-of-the-art methods. * represents re-implemented results, implementation details refer to appendix." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.357, + 0.907, + 0.432 + ], + "angle": 0, + "content": "Video-MME. Video-MME [15] is a recently proposed benchmark for comprehensive video analysis evaluation. We evaluate DrVideo on the \"long-term videos\" split of this dataset (long split), where video lengths vary from 30 to 60 minutes, with an average duration of 44 minutes." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.433, + 0.906, + 0.538 + ], + "angle": 0, + "content": "Since EgoSchema and Video-MME are multi-choice tasks, we use accuracy as the evaluation metric for both datasets. MovieChat-1K focuses on open-ended questions, we use GPT-assisted evaluation to assess both the accuracy (true/false) and quality (score 0-5) of the models. We select Gemini-Pro [54] as the evaluation assistant and use the same prompt [38] to conduct a fair comparison." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.547, + 0.731, + 0.563 + ], + "angle": 0, + "content": "4.2. Implementation Details" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.569, + 0.907, + 0.841 + ], + "angle": 0, + "content": "For EgoSchema [36] dataset, we choose LaViLa [78] as the captioning model to convert videos into documents. Note that the training data of our LaViLa model does not include EgoSchema videos, which is the same as LLoVi [74]. For MovieChat-1K and Video-MME benchmarks, LLaVA-NeXT [33] is used as our captioning model to generate brief descriptions for each frame. We preprocess videos by simply sampling them at 0.5 FPS for EgoSchema [36] and MovieChat-1K [49] and 0.2 FPS for Video-MME [15]. LLaVA-NeXT [33] is chosen to augment key frames in both datasets via different prompts. In the comparison experiments, GPT-4 [1], i.e., gpt-4-1106-preview, is used as the agent to evaluate the performance of our DrVideo on EgoSchema and MovieChat-1K. DeepSeek [19], i.e., DeepSeek V2.5, is used as the agent to evaluate the performance on Video-MME. For the ablation study, we use GPT-3.5 [40], i.e., gpt-3.5-turbo-1106, to evaluate the effectiveness of our DrVideo due to the API cost." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.849, + 0.653, + 0.864 + ], + "angle": 0, + "content": "4.3. Main Results" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.871, + 0.905, + 0.902 + ], + "angle": 0, + "content": "Comparative results on EgoSchema. Table 1 compares our DrVideo with existing LLM-based state-of-the" + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "18940" + } + ], + [ + { + "type": "table", + "bbox": [ + 0.107, + 0.089, + 0.472, + 0.37 + ], + "angle": 0, + "content": "
MethodFramesw/o Subsw Subs
Proprietary MLLM
Claude 3.5 Sonnet [2]2051.254.7
GPT-4V [42]1053.556.9
GPT-4o mini [44]25058.663.4
Gemini 1.5 Flash [55]0.5 FPS61.168.8
GPT-4o [44]38465.372.1
Gemini 1.5 Pro [55]0.5 FPS67.477.4
Open-Source MLLM
LongVA [76]12846.247.6
VITA-8x7B [16]3248.650.9
LLaVA-OneVision-72B [28]3260.062.4
Qwen2-VL-72B [3]76862.274.3
Based on Open-source Captioners and Proprietary LLMs
\\( VideoAgent^* \\)[61]7.640.2*44.4*
\\( LLoVi^* \\)[74]0.2 FPS45.4*67.7*
DrVideo (ours)0.2 FPS51.771.7
-only subs0.2 FPS-68.5
" + }, + { + "type": "table_caption", + "bbox": [ + 0.09, + 0.38, + 0.484, + 0.478 + ], + "angle": 0, + "content": "Table 3. Video-MME long split results. \\( w/o \\) subs represents the model without any subtitles. \\( w \\) subs represents that the subtitles corresponding to the sampled frames are both fed into the LLM to evaluate the performance. Our DrVideo outperforms a strong proprietary MLLM (Claude 3.5 Sonnet) and some open-source MLLMs (e.g., VITA-8x7B). * represents re-implemented results, implementation details refer to appendix." + }, + { + "type": "text", + "bbox": [ + 0.089, + 0.492, + 0.483, + 0.719 + ], + "angle": 0, + "content": "art methods [9, 14, 25, 37, 47, 60, 61, 74] on EgoSchema benchmark [36]. Specifically, compared with LLoVi [74] and VideoAgent [61] that leverage the same visual captioner (LaViLa [78]) and LLM (GPT-4), DrVideo significantly outperforms these methods by \\(5.2\\%\\) and \\(6.2\\%\\) respectively. Compared with VideoAgent [14] which uses videospecific models (ViCLIP from InternVid [63] and VideoLLaVA [31]), DrVideo still performs better and achieves \\(3.6\\%\\) higher accuracy on the subset evaluation. Moreover, DrVideo significantly outperforms IG-VLM [25] by \\(6.6\\%\\), which uses strong MLLM, i.e., GPT-4V. These results confirm our idea of converting the long-video understanding task into a long-document understanding task is feasible and our DrVideo can find the question-related key frames more accurately." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.72, + 0.483, + 0.902 + ], + "angle": 0, + "content": "Comparative results on longer video benchmarks. To further emphasize the advantage of our method on longer videos, we evaluate our DrVideo on the other two video benchmarks, i.e., MoiveChat-1K (10 minutes) [49], and the long split of Video-MME (average 44 minutes) [15]. Table 2 presents the comparative results of our DrVideo and the current state-of-the-art methods [30, 38, 49, 74, 75] on the MovieChat-1K benchmark. Table 3 shows the comparative results of our DrVideo on the Video-MME benchmark, with comparisons against three types of methods: proprietary MLLMs [2, 42, 44, 55], open-source MLLMs [3, 16, 28, 76], and methods based on open-source caption" + }, + { + "type": "table", + "bbox": [ + 0.548, + 0.089, + 0.875, + 0.193 + ], + "angle": 0, + "content": "
RMMSAILCoTAcc.(%)
✓✓✓62.6
✓✓X62.2
✓X✓60.6
X✓✓59.4
XX✓57.4
" + }, + { + "type": "table_caption", + "bbox": [ + 0.513, + 0.203, + 0.907, + 0.245 + ], + "angle": 0, + "content": "Table 4. Ablation results on different combinations of the retrieval module (RM), the multi-stage agent interaction loop (MSAIL), and the CoT on EgoSchema." + }, + { + "type": "table", + "bbox": [ + 0.53, + 0.256, + 0.723, + 0.332 + ], + "angle": 0, + "content": "
VQACaptionAcc. (%)
✓✓62.6
×✓60.4
✓×61.8
" + }, + { + "type": "table_footnote", + "bbox": [ + 0.53, + 0.334, + 0.684, + 0.356 + ], + "angle": 0, + "content": "(a) Performance with different types of information" + }, + { + "type": "table", + "bbox": [ + 0.729, + 0.256, + 0.865, + 0.332 + ], + "angle": 0, + "content": "
Top-KAcc. (%)
562.6
1061.4
2060.6
" + }, + { + "type": "table_footnote", + "bbox": [ + 0.704, + 0.334, + 0.89, + 0.355 + ], + "angle": 0, + "content": "(b) Performance with different Top- \\(K\\) frames" + }, + { + "type": "table_caption", + "bbox": [ + 0.521, + 0.367, + 0.898, + 0.381 + ], + "angle": 0, + "content": "Table 5. Ablation results with different settings on EgoSchema." + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.398, + 0.747, + 0.413 + ], + "angle": 0, + "content": "ers and proprietary LLMs [61, 74]." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.415, + 0.907, + 0.687 + ], + "angle": 0, + "content": "Compared to methods like LLoVi [74] and VideoAgent [61] that are based on open-source captioners (e.g., LLaVA-NeXT [33]) and proprietary LLMs (e.g., GPT-4 [1]), DrVideo achieves a \\(8.3\\%\\) improvement in global mode and a \\(24.8\\%\\) improvement in breakpoint mode on the MovieChat-1K benchmark in Table 2. It also achieves at least \\(6.3\\%\\) improvement under the w/o subtitle setting on the Video-MME benchmark in Table 3. This indicates that DrVideo can accurately identify question-related frames and avoid critical information loss through augmentation, highlighting the potential of document retrieval methods for processing longer videos. Compared to proprietary MLLMs and open-source MLLMs, DrVideo outperforms Claude 3.5 Sonnet [2] by \\(0.5\\%\\) and some open-source MLLMs (e.g., VITA-8x7B [16]) under the w/o subtitle setting on Video-MME benchmark. It also achieves a \\(21.9\\%\\) improvement in global mode and a \\(6.8\\%\\) improvement in breakpoint mode on MovieChat-1K benchmark." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.689, + 0.909, + 0.903 + ], + "angle": 0, + "content": "Moreover, combining the subtitle on the Video-MME benchmark, DrVideo achieves \\(71.7\\%\\) accuracy and outperforms many heavily-engineered large-scale proprietary models (e.g., Gemini 1.5 Flash [55], GPT-4o mini [44], and GPT-4V [42]). To better understand this significant improvement, we conduct an experiment where only the subtitle, question and options are input into the LLM (i.e., DeepSeek [19]) to predict the answer and the accuracy is \\(68.5\\%\\). This suggests: (i) subtitles contain extensive question-related information, playing a crucial role in understanding long videos; (ii) compared with LLoVi and VideoAgent, DrVideo enhances visual information via document retrieval and augmentation, effectively integrating subtitles and resulting in a \\(3.2\\%\\) improvement." + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.518, + 0.957 + ], + "angle": 0, + "content": "18941" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.113, + 0.096, + 0.465, + 0.253 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.09, + 0.272, + 0.483, + 0.318 + ], + "angle": 0, + "content": "Figure 3. Performance of different rounds with DrVideo and VideoAgent [61] on EgoSchema. To align with VideoAgent, GPT-4, i.e., gpt-4-turbo-1106-preview, is used as the LLM agents." + }, + { + "type": "table", + "bbox": [ + 0.115, + 0.329, + 0.462, + 0.402 + ], + "angle": 0, + "content": "
Captioning ModelTypeAcc.(%)
LaViLa [78]Clip-based62.6
LLaVA-NeXT [33]Frame-based61.2
BLIP-2 [29]Frame-based59.6
" + }, + { + "type": "table_caption", + "bbox": [ + 0.117, + 0.412, + 0.456, + 0.426 + ], + "angle": 0, + "content": "Table 6. Performance of different VLMs on EgoSchema." + }, + { + "type": "title", + "bbox": [ + 0.09, + 0.439, + 0.255, + 0.453 + ], + "angle": 0, + "content": "4.4. Ablation Studies" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.461, + 0.483, + 0.717 + ], + "angle": 0, + "content": "Unless otherwise specified, we use GPT-3.5 as the default setting in the below experiments on EgoSchema subset [36]. Effects of different components of DrVideo. We conduct experiments with different combinations of the individual components in DrVideo, including the retrieval module, the multi-stage agent interaction loop, and the CoT used in the answering module. The results are presented in Table 4. We can see that: (i) without the retrieval module and the agent loop, the performance of our DrVideo is similar to LLoVi [74]; (ii) By adding the retrieval module and then the agent loop, the performance of DrVideo improves from \\(57.4\\%\\) to \\(60.6\\%\\) and then to \\(62.6\\%\\). This demonstrates that the retrieval module can identify key frames via measuring semantic similarity, while the agent loop can locate additional key frames through contextual inference; (iii) the additional CoT contributes to the accuracy improvement from \\(62.2\\%\\) to \\(62.6\\%\\), demonstrating its effectiveness." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.719, + 0.483, + 0.884 + ], + "angle": 0, + "content": "Effects of different types of augment information. To better enhance the required information for each potential key frame, we define two types of augment information, namely VQA (question-related information) and caption (general detailed information). Table 5a presents the performance of DrVideo with different types of information. We can see that when only one type of information (VQA or caption) is used for augmentation, the model's performance declines. This suggests the necessity of adaptively augmenting different types of information for each potential key frame." + }, + { + "type": "text", + "bbox": [ + 0.091, + 0.886, + 0.483, + 0.901 + ], + "angle": 0, + "content": "Effects of different initial top \\(K\\) key frames. By default," + }, + { + "type": "table", + "bbox": [ + 0.538, + 0.09, + 0.885, + 0.161 + ], + "angle": 0, + "content": "
LLMsTypeSizeAcc.(%)
Mistral-8x7B [24]Open-Source8x7B47.6
DeepSeek [19]ProprietaryN/A61.2
GPT-3.5 [40]ProprietaryN/A62.6
GPT-4 [1]ProprietaryN/A66.4
" + }, + { + "type": "table_caption", + "bbox": [ + 0.541, + 0.171, + 0.878, + 0.186 + ], + "angle": 0, + "content": "Table 7. Performance of different LLMs on EgoSchema." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.201, + 0.905, + 0.322 + ], + "angle": 0, + "content": "the number of the initial key frames is set as \\( K = 5 \\). Intuitively, more frames should lead to better performance at the cost of increasing the complexity. However, our ablation in Table 5b shows an opposite trend, i.e., more key frames resulting in worse performance. This suggests that enhancing more frames is not necessarily beneficial, as it can introduce a large amount of noisy and irrelevant information that interferes with the LLM's judgment." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.324, + 0.906, + 0.55 + ], + "angle": 0, + "content": "Effects of iterative rounds. We also examined the impact of the iterative rounds \\( I \\) on model performance. Fig. 3 shows the performance of different iterative rounds with DrVideo and VideoAgent [61]. We have the following observations. (i) DrVideo consistently outperforms VideoAgent, reaching the peak performance at \\( I = 2 \\), while VideoAgent peaks at \\( I = 3 \\). This indicates that DrVideo can locate the information needed to answer questions in fewer interactions and with higher quality, underscoring the superiority of our document retrieval-based framework. (ii) When the number of interactions exceeds two, DrVideo's performance declines. This is likely due to more noisy and irrelevant information being introduced into the LLM with the increase of the iterations. This further confirms that excessive information is not necessarily beneficial." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.552, + 0.906, + 0.748 + ], + "angle": 0, + "content": "Effects of different foundation models. DrVideo incorporates two types of foundational models — large language model (LLM), and visual language model (VLM), where VLM is used as the frame captioner while LLM is used as the planning agent, the interaction agent, and the answering agent. To assess the impact of the captions produced by different VLMs, we examine three state-of-the-art VLMs: frame-based BLIP-2 [29], LLaVA-NeXT [33], and clip-based LaViLa [78], as presented in Table 6. We observe that LaViLa achieves the best performance, while BLIP-2 is the worst. It indicates the superiority of the clip-based model which can capture motion information in videos and is more suited for video understanding." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.75, + 0.906, + 0.901 + ], + "angle": 0, + "content": "To evaluate the effectiveness of DrVideo with different LLMs, Table 7 shows the results of DrVideo using open-source LLM (Mistral-8x7B [24]) or Proprietary LLM (DeepSeek [19], GPT-3.5 [40], or GPT-4 [1]) as the planning, interaction and answering agents. From the results, we can see that: (i) different LLMs in DrVideo exhibit significant performance gaps and the proprietary LLM significantly outperforms the open-source LLM, indicating DrVideo's dependency on the choice of LLM; (ii) as the LLM's capability increases, the performance of DrVideo" + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.519, + 0.956 + ], + "angle": 0, + "content": "18942" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.098, + 0.093, + 0.885, + 0.117 + ], + "angle": 0, + "content": "Question: By analyzing c's activities in this video, determine their overall focus or intention, and describe the progression of tasks leading to its fulfillment." + }, + { + "type": "text", + "bbox": [ + 0.101, + 0.118, + 0.775, + 0.129 + ], + "angle": 0, + "content": "A. C's overall focus or intention in this video is to set up their work station and start working on their computer." + }, + { + "type": "text", + "bbox": [ + 0.102, + 0.13, + 0.78, + 0.141 + ], + "angle": 0, + "content": "B. The central objective in this video, c's overall focus or intention, is to engage in conversation with the person." + }, + { + "type": "text", + "bbox": [ + 0.102, + 0.142, + 0.502, + 0.152 + ], + "angle": 0, + "content": "C. C's overall focus or intention in this video is to eat the wrapper." + }, + { + "type": "text", + "bbox": [ + 0.103, + 0.153, + 0.704, + 0.164 + ], + "angle": 0, + "content": "D. C's overall focus or primary intention in this specific video content is to effectively play the guitar" + }, + { + "type": "text", + "bbox": [ + 0.103, + 0.165, + 0.81, + 0.176 + ], + "angle": 0, + "content": "E. The primary objective and intention in this particular c's video is to demonstrate how to effectively clean the cloth." + }, + { + "type": "list", + "bbox": [ + 0.101, + 0.118, + 0.81, + 0.176 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.092, + 0.178, + 0.92, + 0.55 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.119, + 0.559, + 0.877, + 0.573 + ], + "angle": 0, + "content": "Figure 4. Case study on an instance from EgoSchema. DrVideo accurately identifies key frames and chooses the correct answer." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.593, + 0.484, + 0.714 + ], + "angle": 0, + "content": "improves, demonstrating the great potential of DrVideo. It's worth noting that our main contribution is the introduction of a framework for translating the long video understanding into a document retrieval and understanding task to effectively leverage the power of large language models, rather than the employment of any specific models. Our DrVideo can be further improved by combining with better LLM and VLM models (e.g., tarsier [59])." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.72, + 0.484, + 0.901 + ], + "angle": 0, + "content": "Case study. Here we give a detailed example from EgoSchema [36] in Fig. 4 to qualitatively demonstrate the capability of our DrVideo. Specifically, the question is to determine the overall focus or intention and describe the progression of tasks leading to its fulfillment. DrVideo first converts the raw video into an initial video document, which is then processed by the frame retrieval module to obtain key frames (frames 15, 46, 61, 84, and 86) and is augmented with detailed key frame descriptions. The updated video document is then fed into the planning agent to assess if sufficient information is available for answering, i.e., the provided information is sufficient to infer character C's" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.593, + 0.907, + 0.699 + ], + "angle": 0, + "content": "main focus or intent. This updated document, along with the rationale, is input to the interaction agent to identify the additional key frames (frames 4, 5, 6, 14, and 19 here). After further augmentation, the updated document is sent back to the planning agent, for which the agent confirms its sufficiency. Finally, the answering module takes in the final video document and predicts the correct answer." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.71, + 0.634, + 0.725 + ], + "angle": 0, + "content": "5. Conclusion" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.735, + 0.907, + 0.903 + ], + "angle": 0, + "content": "We have proposed DrVideo, a document retrieval-based system for long video understanding. Different from previous LLM-based methods, DrVideo proposes to adapt long-video understanding to long-document understanding and searches for missing information via document retrieval and multi-stage agent loop. Extensive comparative experiments and ablation studies on the three challenging datasets, i.e., EgoSchema, MovieChat-1K, and Video-MME, demonstrate the effectiveness of DrVideo. We believe our work can serve as a strong baseline to stimulate further developments on the topic of long video understanding." + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "18943" + } + ], + [ + { + "type": "title", + "bbox": [ + 0.093, + 0.091, + 0.241, + 0.108 + ], + "angle": 0, + "content": "Acknowledgment" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.116, + 0.483, + 0.191 + ], + "angle": 0, + "content": "This work is supported by the National Natural Science Foundation of China (62221002, 62171183, U24A20327, 62401202), the Joint Fund of the Ministry of Education of China, and the China Scholarship Council (CSC) under Grant No. 202306130148." + }, + { + "type": "title", + "bbox": [ + 0.093, + 0.207, + 0.187, + 0.223 + ], + "angle": 0, + "content": "References" + }, + { + "type": "ref_text", + "bbox": [ + 0.101, + 0.232, + 0.482, + 0.3 + ], + "angle": 0, + "content": "[1] Josh Achiam, Steven Adler, Sandhini Agarwal, Lama Ahmad, Ilge Akkaya, Florencia Leoni Aleman, Diogo Almeida, Janko Altenschmidt, Sam Altman, Shyamal Anadkat, et al. Gpt-4 technical report. arXiv preprint arXiv:2303.08774, 2023. 2, 5, 6, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.303, + 0.483, + 0.344 + ], + "angle": 0, + "content": "[2] Anthropic. Claude 3.5 Sonnet: Enhanced Conversational AI. https://www.anthropic.com/index/claude-3.5-sonnet/, 2024.6" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.346, + 0.483, + 0.415 + ], + "angle": 0, + "content": "[3] Jinze Bai, Shuai Bai, Shusheng Yang, Shijie Wang, Sinan Tan, Peng Wang, Junyang Lin, Chang Zhou, and Jingren Zhou. Qwen-vl: A versatile vision-language model for understanding, localization, text reading, and beyond. arXiv preprint arXiv:2308.12966, 2023. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.418, + 0.483, + 0.473 + ], + "angle": 0, + "content": "[4] Aanisha Bhattacharya, Yaman K Singla, Balaji Krishnamurthy, Rajiv Ratn Shah, and Changyou Chen. A video is worth 4096 tokens: Verbalize story videos to understand them in zero shot. arXiv preprint arXiv:2305.09758, 2023. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.476, + 0.482, + 0.544 + ], + "angle": 0, + "content": "[5] Anthony Brohan, Noah Brown, Justice Carbajal, Yevgen Chebotar, Xi Chen, Krzysztof Choromanski, Tianli Ding, Danny Driess, Avinava Dubey, Chelsea Finn, et al. Rt-2: Vision-language-action models transfer web knowledge to robotic control. arXiv preprint arXiv:2307.15818, 2023. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.546, + 0.482, + 0.614 + ], + "angle": 0, + "content": "[6] Tom Brown, Benjamin Mann, Nick Ryder, Melanie Subbiah, Jared D Kaplan, Prafulla Dhariwal, Arvind Neelakantan, Pranav Shyam, Girish Sastry, Amanda Askell, et al. Language models are few-shot learners. Adv. Neural Inform. Process. Syst., pages 1877-1901, 2020. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.617, + 0.482, + 0.671 + ], + "angle": 0, + "content": "[7] Jun Chen, Deyao Zhu, Kilichbek Haydarov, Xiang Li, and Mohamed Elhoseiny. Video chatcaptioner: Towards the enriched spatiotemporal descriptions. arXiv preprint arXiv:2304.04227, 2023. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.674, + 0.482, + 0.73 + ], + "angle": 0, + "content": "[8] Feng Cheng, Xizi Wang, Jie Lei, David Crandall, Mohit Bansal, and Gedas Bertasius. Vindlu: A recipe for effective video-and-language pretraining. In IEEE/CVF Conf. Comput. Vis. Pattern Recog., pages 10739-10750, 2023. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.732, + 0.482, + 0.785 + ], + "angle": 0, + "content": "[9] Rohan Choudhury, Koichiro Niinuma, Kris M Kitani, and László A Jeni. Zero-shot video question answering with procedural programs. arXiv preprint arXiv:2312.00937, 2023. 5, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.789, + 0.482, + 0.857 + ], + "angle": 0, + "content": "[10] Aakanksha Chowdhery, Sharan Narang, Jacob Devlin, Maarten Bosma, Gaurav Mishra, Adam Roberts, Paul Barham, Hyung Won Chung, Charles Sutton, Sebastian Gehrmann, et al. Palm: Scaling language modeling with pathways. arXiv preprint arXiv:2204.02311, 2022. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.86, + 0.482, + 0.901 + ], + "angle": 0, + "content": "[11] Jiwan Chung and Youngjae Yu. Long story short: a summarize-then-search method for long video question answering. In Brit. Mach. Vis. Conf., 2023. 3" + }, + { + "type": "list", + "bbox": [ + 0.094, + 0.232, + 0.483, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.093, + 0.905, + 0.147 + ], + "angle": 0, + "content": "[12] Yiran Ding, Li Lyna Zhang, Chengruidong Zhang, Yuanyuan Xu, Ning Shang, Jiahang Xu, Fan Yang, and Mao Yang. Longgrope: Extending llm context window beyond 2 million tokens. arXiv preprint arXiv:2402.13753, 2024. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.151, + 0.905, + 0.218 + ], + "angle": 0, + "content": "[13] Danny Driess, Fei Xia, Mehdi SM Sajjadi, Corey Lynch, Aakanksha Chowdhery, Brian Ichter, Ayzaan Wahid, Jonathan Tompson, Quan Vuong, Tianhe Yu, et al. Palme: An embodied multimodal language model. arXiv preprint arXiv:2303.03378, 2023. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.221, + 0.905, + 0.275 + ], + "angle": 0, + "content": "[14] Yue Fan, Xiaojian Ma, Rujie Wu, Yuntao Du, Jiaqi Li, Zhi Gao, and Qing Li. Videoagent: A memory-augmented multimodal agent for video understanding. arXiv preprint arXiv:2403.11481, 2024. 2, 3, 5, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.278, + 0.904, + 0.346 + ], + "angle": 0, + "content": "[15] Chaoyou Fu, Yuhan Dai, Yondong Luo, Lei Li, Shuhuai Ren, Renrui Zhang, Zihan Wang, Chenyu Zhou, Yunhang Shen, Mengdan Zhang, et al. Video-mme: The first-ever comprehensive evaluation benchmark of multi-modal llms in video analysis. arXiv preprint arXiv:2405.21075, 2024. 2, 3, 5, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.349, + 0.904, + 0.403 + ], + "angle": 0, + "content": "[16] Chaoyou Fu, Haojia Lin, Zuwei Long, Yunhang Shen, Meng Zhao, Yifan Zhang, Xiong Wang, Di Yin, Long Ma, Xiawu Zheng, et al. Vita: Towards open-source interactive omni multimodal llm. arXiv preprint arXiv:2408.05211, 2024. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.406, + 0.904, + 0.46 + ], + "angle": 0, + "content": "[17] Difei Gao, Lei Ji, Luowei Zhou, Kevin Qinghong Lin, Joya Chen, Zihan Fan, and Mike Zheng Shou. Assistgpt: A general multi-modal assistant that can plan, execute, inspect, and learn. arXiv preprint arXiv:2306.08640, 2023. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.462, + 0.904, + 0.503 + ], + "angle": 0, + "content": "[18] Albert Gu, Karan Goel, and Christopher Ré. Efficiently modeling long sequences with structured state spaces. arXiv preprint arXiv:2111.00396, 2021. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.505, + 0.905, + 0.572 + ], + "angle": 0, + "content": "[19] Daya Guo, Qihao Zhu, Dejian Yang, Zhenda Xie, Kai Dong, Wentao Zhang, Guanting Chen, Xiao Bi, Yu Wu, YK Li, et al. Deepseek-coder: When the large language model meets programming-the rise of code intelligence. arXiv preprint arXiv:2401.14196, 2024. 5, 6, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.576, + 0.905, + 0.63 + ], + "angle": 0, + "content": "[20] Wenyi Hong, Weihan Wang, Qingsong Lv, Jiazheng Xu, Wenmeng Yu, Junhui Ji, Yan Wang, Zihan Wang, Yuxiao Dong, Ming Ding, et al. Cogagent: A visual language model for gui agents. arXiv preprint arXiv:2312.08914, 2023. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.633, + 0.905, + 0.685 + ], + "angle": 0, + "content": "[21] Noureldien Hussein, Efstratios Gavves, and Arnold WM Smeulders. Videograph: Recognizing minutes-long human activities in videos. arXiv preprint arXiv:1905.05143, 2019. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.689, + 0.905, + 0.731 + ], + "angle": 0, + "content": "[22] Md Mohaiminul Islam and Gedas Bertasius. Long movie clip classification with state-space video models. In Eur. Conf. Comput. Vis., pages 87-104, 2022. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.733, + 0.905, + 0.8 + ], + "angle": 0, + "content": "[23] Sullam Jeoung, Goeric Huybrechts, Bhavana Ganesh, Aram Galstyan, and Sravan Bodapati. Adaptive video understanding agent: Enhancing efficiency with dynamic frame sampling and feedback-driven reasoning. arXiv preprint arXiv:2410.20252, 2024. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.803, + 0.905, + 0.87 + ], + "angle": 0, + "content": "[24] Albert Q Jiang, Alexandre Sablayrolles, Arthur Mensch, Chris Bamford, Devendra Singh Chaplot, Diego de las Casas, Florian Bressand, Gianna Lengyel, Guillaume Lample, Lucile Saulnier, et al. Mistral 7b. arXiv preprint arXiv:2310.06825, 2023. 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.873, + 0.905, + 0.901 + ], + "angle": 0, + "content": "[25] Wonkyun Kim, Changin Choi, Wonseok Lee, and Wonjong Rhee. An image grid can be worth a video: Zero" + }, + { + "type": "list", + "bbox": [ + 0.094, + 0.093, + 0.905, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.946, + 0.519, + 0.957 + ], + "angle": 0, + "content": "18944" + } + ], + [ + { + "type": "ref_text", + "bbox": [ + 0.126, + 0.093, + 0.483, + 0.12 + ], + "angle": 0, + "content": "shot video question answering using a vlm. arXiv preprint arXiv:2403.18406, 2024. 1, 5, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.121, + 0.484, + 0.176 + ], + "angle": 0, + "content": "[26] Takeshi Kojima, Shixiang Shane Gu, Machel Reid, Yutaka Matsuo, and Yusuke Iwasawa. Large language models are zero-shot reasoners. Adv. Neural Inform. Process. Syst., pages 22199-22213, 2022. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.177, + 0.484, + 0.245 + ], + "angle": 0, + "content": "[27] Jie Lei, Linjie Li, Luowei Zhou, Zhe Gan, Tamara L Berg, Mohit Bansal, and Jingjing Liu. Less is more: Clipbert for video-and-language learning via sparse sampling. In IEEE/CVF Conf. Comput. Vis. Pattern Recog., pages 7331-7341, 2021. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.246, + 0.483, + 0.302 + ], + "angle": 0, + "content": "[28] Bo Li, Yuanhan Zhang, Dong Guo, Renrui Zhang, Feng Li, Hao Zhang, Kaichen Zhang, Yanwei Li, Ziwei Liu, and Chunyuan Li. Llava-onevision: Easy visual task transfer. arXiv preprint arXiv:2408.03326, 2024. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.302, + 0.483, + 0.357 + ], + "angle": 0, + "content": "[29] Junnan Li, Dongxu Li, Silvio Savarese, and Steven Hoi. BLIP-2: bootstrapping language-image pre-training with frozen image encoders and large language models. In Proc. Int. Conf. Mach. Learn., 2023. 4, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.359, + 0.483, + 0.413 + ], + "angle": 0, + "content": "[30] KunChang Li, Yinan He, Yi Wang, Yizhuo Li, Wenhai Wang, Ping Luo, Yali Wang, Limin Wang, and Yu Qiao. Videochat: Chat-centric video understanding. arXiv preprint arXiv:2305.06355, 2023. 3, 5, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.414, + 0.483, + 0.468 + ], + "angle": 0, + "content": "[31] Bin Lin, Bin Zhu, Yang Ye, Munan Ning, Peng Jin, and Li Yuan. Video-llava: Learning united visual representation by alignment before projection. arXiv preprint arXiv:2311.10122, 2023. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.469, + 0.483, + 0.538 + ], + "angle": 0, + "content": "[32] Kevin Lin, Faisal Ahmed, Linjie Li, Chung-Ching Lin, Ehsan Azarnasab, Zhengyuan Yang, Jianfeng Wang, Lin Liang, Zicheng Liu, Yumao Lu, et al. Mm-vid: Advancing video understanding with gpt-4v (ision). arXiv preprint arXiv:2310.19773, 2023. 1, 3, 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.539, + 0.483, + 0.607 + ], + "angle": 0, + "content": "[33] Haotian Liu, Chunyuan Li, Yuheng Li, Bo Li, Yuanhan Zhang, Sheng Shen, and Yong Jae Lee. Llava next: Improved reasoning,OCR,and world knowledge.https://llava-v1.github.io/blog/2024-01-30-llava NEXT/,2024.2,4,5,6,7" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.608, + 0.483, + 0.663 + ], + "angle": 0, + "content": "[34] Ziyu Ma, Shutao Li, Bin Sun, Jianfei Cai, Zuxiang Long, and Fuyan Ma. Gerea: Question-aware prompt captions for knowledge-based visual question answering. arXiv preprint arXiv:2402.02503, 2024. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.664, + 0.483, + 0.72 + ], + "angle": 0, + "content": "[35] Muhammad Maaz, Hanoona Rasheed, Salman Khan, and Fahad Shahbaz Khan. Video-chatgpt: Towards detailed video understanding via large vision and language models. arXiv preprint arXiv:2306.05424, 2023. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.721, + 0.483, + 0.775 + ], + "angle": 0, + "content": "[36] Karttikeya Mangalam, Raiymbek Akshulakov, and Jitendra Malik. Egoschema: A diagnostic benchmark for very long-form video language understanding. arXiv preprint arXiv:2308.09126, 2023. 2, 3, 5, 6, 7, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.776, + 0.483, + 0.844 + ], + "angle": 0, + "content": "[37] Juhong Min, Shyamal Buch, Arsha Nagrani, Minsu Cho, and Cordelia Schmid. Morevqa: Exploring modular reasoning models for video question answering. In IEEE/CVF Conf. Comput. Vis. Pattern Recog., pages 13235-13245, 2024. 2, 5, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.845, + 0.483, + 0.9 + ], + "angle": 0, + "content": "[38] Salman Khan Muhammad Maaz, Hanoona Rasheed and Fahad Khan. Video-chatgpt: Towards detailed video understanding via large vision and language models. ArXiv 2306.05424, 2023. 1, 5, 6" + }, + { + "type": "list", + "bbox": [ + 0.093, + 0.093, + 0.484, + 0.9 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.093, + 0.905, + 0.161 + ], + "angle": 0, + "content": "[39] Eric Nguyen, Karan Goel, Albert Gu, Gordon Downs, Preey Shah, Tri Dao, Stephen Baccus, and Christopher Ré. S4nd: Modeling images and videos as multidimensional signals with state spaces. Adv. Neural Inform. Process. Syst., pages 2846-2861, 2022. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.162, + 0.905, + 0.19 + ], + "angle": 0, + "content": "[40] OpenAI. GPT3.5. https://platform.openai.com/docs/models/gpt-3-5, 2021.1, 5, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.191, + 0.905, + 0.218 + ], + "angle": 0, + "content": "[41] OpenAI. Introducing chatgpt. https://openai.com/blog/chatgpt, 2022.1" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.219, + 0.905, + 0.259 + ], + "angle": 0, + "content": "[42] OpenAI. GPT-4V(ison) System Card. https://openai.com/index/gpt-4v-system-card/, 2023.3,6" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.26, + 0.905, + 0.302 + ], + "angle": 0, + "content": "[43] OpenAI. New embedding models and api updates. https://openai.com/index/new-embedding-models-and-api-updates/, 2024.4" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.302, + 0.905, + 0.357 + ], + "angle": 0, + "content": "[44] OpenAI. GPT-4o Mini: Advancing Cost-Efficient Intelligence. https://openai.com/index/gpt-4o-mini-advancing-cost-efficient-intelligence/, 2024.3,6" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.359, + 0.905, + 0.428 + ], + "angle": 0, + "content": "[45] Long Ouyang, Jeffrey Wu, Xu Jiang, Diogo Almeida, Carroll Wainwright, Pamela Mishkin, Chong Zhang, Sandhini Agarwal, Katarina Slama, Alex Ray, et al. Training language models to follow instructions with human feedback. Adv. Neural Inform. Process. Syst., pages 27730-27744, 2022. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.429, + 0.905, + 0.496 + ], + "angle": 0, + "content": "[46] Vincent Perot, Kai Kang, Florian Luisier, Guolong Su, Xiaoyu Sun, Ramya Sree Boppana, Zilong Wang, Zifeng Wang, Jiaqi Mu, Hao Zhang, et al. Lmdx: Language model-based document information extraction and localization. arXiv preprint arXiv:2309.10952, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.497, + 0.905, + 0.552 + ], + "angle": 0, + "content": "[47] Kanchana Ranasinghe, Xiang Li, Kumara Kahatapitiya, and Michael S Ryoo. Understanding long videos in one multimodal language model pass. arXiv preprint arXiv:2403.16998, 2024. 5, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.553, + 0.905, + 0.622 + ], + "angle": 0, + "content": "[48] Teven Le Scao, Angela Fan, Christopher Akiki, Ellie Pavlick, Suzana Ilic, Daniel Hesslow, Roman Castagné, Alexandra Sasha Luccioni, François Yvon, Matthias Galle, et al. Bloom: A 176b-parameter open-access multilingual language model. arXiv preprint arXiv:2211.05100, 2022. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.622, + 0.905, + 0.692 + ], + "angle": 0, + "content": "[49] Enxin Song, Wenhao Chai, Guanhong Wang, Yucheng Zhang, Haoyang Zhou, Feiyang Wu, Xun Guo, Tian Ye, Yan Lu, Jenq-Neng Hwang, et al. Moviechat: From dense token to sparse memory for long video understanding. arXiv preprint arXiv:2307.16449, 2023. 2, 3, 5, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.692, + 0.905, + 0.747 + ], + "angle": 0, + "content": "[50] Enxin Song, Wenhao Chai, Tian Ye, Jenq-Neng Hwang, Xi Li, and Gaoang Wang. Moviechat+: Question-aware sparse memory for long video question answering. arXiv preprint arXiv:2404.17176, 2024. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.748, + 0.905, + 0.802 + ], + "angle": 0, + "content": "[51] Quan Sun, Jinsheng Wang, Qiying Yu, Yufeng Cui, Fan Zhang, Xiaosong Zhang, and Xinlong Wang. Eva-clip-18b: Scaling clip to 18 billion parameters. arXiv preprint arXiv:2402.04252, 2024. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.803, + 0.905, + 0.859 + ], + "angle": 0, + "content": "[52] Yuchong Sun, Hongwei Xue, Ruihua Song, Bei Liu, Huan Yang, and Jianlong Fu. Long-form video-language pretraining with multimodal temporal contrastive learning. Adv. Neural Inform. Process. Syst., 35:38032-38045, 2022. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.86, + 0.905, + 0.901 + ], + "angle": 0, + "content": "[53] Dídac Surís, Sachit Menon, and Carl Vondrick. Vipergpt: Visual inference via python execution for reasoning. Int. Conf. Comput. Vis., pages 11888-11898, 2023. 1, 3" + }, + { + "type": "list", + "bbox": [ + 0.517, + 0.093, + 0.905, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.946, + 0.519, + 0.957 + ], + "angle": 0, + "content": "18945" + } + ], + [ + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.092, + 0.486, + 0.161 + ], + "angle": 0, + "content": "[54] Gemini Team, Rohan Anil, Sebastian Borgeaud, Yonghui Wu, Jean-Baptiste Alayrac, Jiahui Yu, Radu Soricut, Johan Schalkwyk, Andrew M Dai, Anja Hauth, et al. Gemini: a family of highly capable multimodal models. arXiv preprint arXiv:2312.11805, 2023. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.163, + 0.483, + 0.234 + ], + "angle": 0, + "content": "[55] Gemini Team, Petko Georgiev, Ving Ian Lei, Ryan Burnell, Libin Bai, Anmol Gulati, Garrett Tanzer, Damien Vincent, Zhufeng Pan, Shibo Wang, et al. Gemini 1.5: Unlocking multimodal understanding across millions of tokens of context. arXiv preprint arXiv:2403.05530, 2024. 3, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.234, + 0.483, + 0.303 + ], + "angle": 0, + "content": "[56] Hugo Touvron, Thibaut Lavril, Gautier Izacard, Xavier Martinet, Marie-Anne Lachaux, Timothée Lacroix, Baptiste Rozière, Naman Goyal, Eric Hambro, Faisal Azhar, et al. Llama: Open and efficient foundation language models. arXiv preprint arXiv:2302.13971, 2023. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.305, + 0.483, + 0.36 + ], + "angle": 0, + "content": "[57] Junke Wang, Dongdong Chen, Chong Luo, Xiyang Dai, Lu Yuan, Zuxuan Wu, and Yu-Gang Jiang. Chatvideo: A tracklet-centric multimodal and versatile video understanding system. arXiv preprint arXiv:2304.14407, 2023. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.362, + 0.483, + 0.431 + ], + "angle": 0, + "content": "[58] Jue Wang, Wentao Zhu, Pichao Wang, Xiang Yu, Linda Liu, Mohamed Omar, and Raffay Hamid. Selective structured state-spaces for long-form video understanding. In IEEE/CVF Conf. Comput. Vis. Pattern Recog., pages 6387-6397, 2023. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.433, + 0.483, + 0.475 + ], + "angle": 0, + "content": "[59] Jiawei Wang, Liping Yuan, and Yuchen Zhang. Tarsier: Recipes for training and evaluating large video description models. arXiv preprint arXiv:2407.00634, 2024. 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.476, + 0.483, + 0.532 + ], + "angle": 0, + "content": "[60] Shijie Wang, Qi Zhao, Minh Quan Do, Nakul Agarwal, Kwonjoon Lee, and Chen Sun. Vamos: Versatile action models for video understanding. arXiv preprint arXiv:2311.13627, 2023. 5, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.533, + 0.483, + 0.588 + ], + "angle": 0, + "content": "[61] Xiaohan Wang, Yuhui Zhang, Orr Zohar, and Serena Yeung-Levy. Videoagent: Long-form video understanding with large language model as agent. arXiv preprint arXiv:2403.10517, 2024. 1, 2, 3, 5, 6, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.59, + 0.483, + 0.646 + ], + "angle": 0, + "content": "[62] Yang Wang, Gedas Bertasius, Tae-Hyun Oh, Abhinav Gupta, Minh Hoai, and Lorenzo Torresani. Supervoxel attention graphs for long-range video modeling. In Proc. IEEE/CVF Winter Conf. Appl. Comput. Vis., pages 155-166, 2021. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.647, + 0.483, + 0.716 + ], + "angle": 0, + "content": "[63] Yi Wang, Yinan He, Yizhuo Li, Kunchang Li, Jiashuo Yu, Xin Ma, Xinhao Li, Guo Chen, Xinyuan Chen, Yaohui Wang, et al. Intervid: A large-scale video-text dataset for multimodal understanding and generation. arXiv preprint arXiv:2307.06942, 2023. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.718, + 0.483, + 0.8 + ], + "angle": 0, + "content": "[64] Zhenhailong Wang, Manling Li, Ruochen Xu, Luowei Zhou, Jie Lei, Xudong Lin, Shuohang Wang, Ziyi Yang, Chenguang Zhu, Derek Hoiem, et al. Language models with image descriptors are strong few-shot video-language learners. Adv. Neural Inform. Process. Syst., pages 8483-8497, 2022. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.802, + 0.483, + 0.871 + ], + "angle": 0, + "content": "[65] Ziyang Wang, Shoubin Yu, Elias Stengel-Eskin, Jaehong Yoon, Feng Cheng, Gedas Bertasius, and Mohit Bansal. Videotree: Adaptive tree-based video representation for llm reasoning on long videos. arXiv preprint arXiv:2405.19209, 2024.3" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.873, + 0.483, + 0.901 + ], + "angle": 0, + "content": "[66] Jason Wei, Xuezhi Wang, Dale Schuurmans, Maarten Bosma, Fei Xia, Ed Chi, Quoc V Le, Denny Zhou, et al." + }, + { + "type": "list", + "bbox": [ + 0.093, + 0.092, + 0.486, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.545, + 0.093, + 0.905, + 0.134 + ], + "angle": 0, + "content": "Chain-of-thought prompting elicits reasoning in large language models. Adv. Neural Inform. Process. Syst., pages 24824-24837, 2022. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.136, + 0.905, + 0.205 + ], + "angle": 0, + "content": "[67] Chao-Yuan Wu, Yanghao Li, Karttikeya Mangalam, Haoqi Fan, Bo Xiong, Jitendra Malik, and Christoph Feichtenhofer. Memvit: Memory-augmented multiscale vision transformer for efficient long-term video recognition. In IEEE/CVF Conf. Comput. Vis. Pattern Recog., pages 13587-13597, 2022. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.207, + 0.905, + 0.261 + ], + "angle": 0, + "content": "[68] Lin Xu, Yilin Zhao, Daquan Zhou, Zhijie Lin, See Kiong Ng, and Jiashi Feng. Pllava: Parameter-free llava extension from images to videos for video dense captioning. arXiv preprint arXiv:2404.16994, 2024. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.263, + 0.905, + 0.319 + ], + "angle": 0, + "content": "[69] Antoine Yang, Antoine Miech, Josef Sivic, Ivan Laptev, and Cordelia Schmid. Just ask: Learning to answer questions from millions of narrated videos. In Int. Conf. Comput. Vis., pages 1686-1697, 2021. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.32, + 0.905, + 0.374 + ], + "angle": 0, + "content": "[70] Antoine Yang, Antoine Miech, Josef Sivic, Ivan Laptev, and Cordelia Schmid. Zero-shot video question answering via frozen bidirectional language models. Adv. Neural Inform. Process. Syst., 35:124-141, 2022. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.375, + 0.905, + 0.432 + ], + "angle": 0, + "content": "[71] Xitong Yang, Fu-Jen Chu, Matt Feiszli, Raghav Goyal, Lorenzo Torresani, and Du Tran. Relational space-time query in long-form videos. In IEEE/CVF Conf. Comput. Vis. Pattern Recog., pages 6398-6408, 2023. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.433, + 0.905, + 0.488 + ], + "angle": 0, + "content": "[72] Zongxin Yang, Guikun Chen, Xiaodi Li, Wenguan Wang, and Yi Yang. Doraemongpt: Toward understanding dynamic scenes with large language models. arXiv preprint arXiv:2401.08392, 2024. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.489, + 0.905, + 0.572 + ], + "angle": 0, + "content": "[73] Andy Zeng, Maria Attarian, Brian Ichter, Krzysztof Choromanski, Adrian Wong, Stefan Welker, Federico Tombari, Aveek Purohit, Michael Ryoo, Vikas Sindhwani, Johnny Lee, Vincent Vanhoucke, and Pete Florence. Socratic models: Composing zero-shot multimodal reasoning with language. arXiv preprint arXiv:2204.00598, 2022. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.574, + 0.905, + 0.628 + ], + "angle": 0, + "content": "[74] Ce Zhang, Taixi Lu, Md Mohaiminul Islam, Ziyang Wang, Shoubin Yu, Mohit Bansal, and Gedas Bertasius. A simple llm framework for long-range video question-answering. arXiv preprint arXiv:2312.17235, 2023. 1, 2, 5, 6, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.63, + 0.905, + 0.672 + ], + "angle": 0, + "content": "[75] Hang Zhang, Xin Li, and Lidong Bing. Video-llama: An instruction-tuned audio-visual language model for video understanding. arXiv preprint arXiv:2306.02858, 2023. 5, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.673, + 0.905, + 0.74 + ], + "angle": 0, + "content": "[76] Peiyuan Zhang, Kaichen Zhang, Bo Li, Guangtao Zeng, Jingkang Yang, Yuanhan Zhang, Ziyue Wang, Haoran Tan, Chunyuan Li, and Ziwei Liu. Long context transfer from language to vision. arXiv preprint arXiv:2406.16852, 2024. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.743, + 0.905, + 0.812 + ], + "angle": 0, + "content": "[77] Yuanhan Zhang, Bo Li, Haotian Liu, Yong Jae Lee, Liangke Gui, Di Fu, Jiashi Feng, Ziwei Liu, and Chunyuan Li. Llava-last: A strong zero-shot video understanding model. https://llava-v1.github.io/blog/2024-04-30-llava-last-video/, 2024.1" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.814, + 0.905, + 0.871 + ], + "angle": 0, + "content": "[78] Yue Zhao, Ishan Misra, Philipp Krahenbuhl, and Rohit Girdhar. Learning video representations from large language models. In IEEE/CVF Conf. Comput. Vis. Pattern Recog., pages 6586-6597, 2023. 1, 2, 4, 5, 6, 7" + }, + { + "type": "list", + "bbox": [ + 0.517, + 0.093, + 0.905, + 0.871 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.946, + 0.52, + 0.957 + ], + "angle": 0, + "content": "18946" + } + ] +] \ No newline at end of file diff --git a/2025/DrVideo_ Document Retrieval Based Long Video Understanding/81160fea-84bd-416d-98f4-7b4c205b85c0_origin.pdf b/2025/DrVideo_ Document Retrieval Based Long Video Understanding/81160fea-84bd-416d-98f4-7b4c205b85c0_origin.pdf new file mode 100644 index 0000000000000000000000000000000000000000..9e4ea294beaf4c66da9fcb4feba74aaa7bb8d2b0 --- /dev/null +++ b/2025/DrVideo_ Document Retrieval Based Long Video Understanding/81160fea-84bd-416d-98f4-7b4c205b85c0_origin.pdf @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:f65a01e6aeb82af75dca11d6c638c95d6045cb949fffb11f5344f38e2762b25b +size 1685202 diff --git a/2025/DrVideo_ Document Retrieval Based Long Video Understanding/full.md b/2025/DrVideo_ Document Retrieval Based Long Video Understanding/full.md new file mode 100644 index 0000000000000000000000000000000000000000..6718bac63ed18ca56e7de79b24fd633aeb9f6c53 --- /dev/null +++ b/2025/DrVideo_ Document Retrieval Based Long Video Understanding/full.md @@ -0,0 +1,317 @@ +# DrVideo: Document Retrieval Based Long Video Understanding + +Ziyu Ma $^{1,2}$ , Chenhui Gou $^{2}$ , Hengcan Shi $^{1,2}$ , Bin Sun $^{1}$ , Shutao Li $^{1}$ , Hamid Rezatofighi $^{2}$ , Jianfei Cai $^{2}$ + +1 College of Electrical and Information Engineering, Hunan University, + +2 Data Science & AI Department, Faculty of IT, Monash University + +{maziyu, sunbin611, shutao_li}@hnu.edu.cn + +{chenhui.gou, hengcan.shi, hamid.rezatofighi, jianfei.cai}@monash.edu + +# Abstract + +Most of the existing methods for video understanding primarily focus on videos only lasting tens of seconds, with limited exploration of techniques for handling long videos. The increased number of frames in long videos poses two main challenges: difficulty in locating key information and performing long-range reasoning. Thus, we propose DrVideo, a document-retrieval-based system designed for long video understanding. Our key idea is to convert the long-video understanding problem into a long-document understanding task so as to effectively leverage the power of large language models. Specifically, DrVideo first transforms a long video into a coarse text-based long document to initially retrieve key frames and then updates the documents with the augmented key frame information. It then employs an agent-based iterative loop to continuously search for missing information and augment the document until sufficient question-related information is gathered for making the final predictions in a chain-of-thought manner. Extensive experiments on long video benchmarks confirm the effectiveness of our method. DrVideo significantly outperforms existing LLM-based state-of-the-art methods on EgoSchema benchmark (3 minutes), MovieChat-1K benchmark (10 minutes), and the long split of Video-MME benchmark (average of 44 minutes). Code is available at https://github.com/Upper9527/DrVideo. + +# 1. Introduction + +Video understanding is a challenging task in computer vision, requiring the processing of spatio-temporal information and advanced reasoning abilities. Previous works have successfully processed short videos lasting around tens of seconds [8, 27, 32, 53, 69, 70]. However, how to deal with + +long video understanding remains unclear. Recent advancements in large language models (LLMs) [6, 10, 34, 41, 45, 48, 56] demonstrate strong abilities in language understanding and reasoning across long text sequences. These advancements have inspired the development of videolanguage models (Video-LMs) [25, 38, 68, 77] to address long video understanding issues. + +Video-LMs typically encode a video as a sequence of visual tokens, which are then concatenated with language tokens into a long sequence and LLM is then used to understand this long sequence. Although Video-LMs improve video understanding, they have the following limitations. (i) They cannot process an entire video as input to a VideoLM (e.g., the visual encoder OpenAI's CLIP-L-14 outputs $24 \times 24$ tokens for each image, while the max length of LLaVA-NeXT-Video [77] is only 8192). (ii) They typically perform uniform or random frame sampling at large strides to handle long videos without considering the content (e.g., 16 for PLLaVA [68] and 6 for IG-VLM [25]), leading to potential key information loss. (iii) The simple concatenation of visual tokens increases the difficulty for the LLM to locate question-related (key) visual information within the long video sequence, complicating long-range reasoning across the vision token sequences. + +Apart from the studies of encoding a video as a sequence of visual tokens, another line of research converts raw videos into captions and then utilities the long-range reasoning abilities of LLMs to predict the final answer. A pioneering work - LLoVi [74] uses a short-term video model [78] in conjunction with an LLM [40] to solve the long video understanding task. Given a long video, LLoVi first divides it into multiple short clips and converts them into short textual descriptions. Afterward, these short textual descriptions are summarized by the LLM, and the summary is finally used by LLoVi to answer the given question. Inspired by LLoVi, VideoAgent [61] designs an agent-based system to further leverage the reasoning abilities of LLMs and locate question-related video clips or images. Initially, + +![](images/aef6f5c7e3fbd886231fae7a678d4a9ff61203dc2838a65af6ffb065fa5d0703.jpg) +Figure 1. Overview of our DrVideo framework. It comprises five components: a video-document conversion module, a retrieval module, a document augmentation module, a multi-stage agent interaction loop and an answering module. + +VideoAgent samples only a minimal number of frames from the video (i.e., 5 frames), converting them into textual descriptions as input for the LLM. Then, through continuous feedback and interaction with the LLM [1], VLM [78], and CLIP [51], it gradually finds question-related frames in a coarse-to-fine manner via image-text similarity. Finally, VideoAgent answers the question based on these question-related frames. Despite these LLM-based methods overcoming the challenge of performing long-range reasoning and achieving promising results, they still suffer from the following problems. + +- These methods struggle to accurately identify question-related frames. For instance, VideoAgent [61] performs sparse sampling on long videos (e.g., selecting 5 frames) and subsequently searches for keyframes from the unsampled frames in a coarse-to-fine manner based on the initial sparse space via CLIP image-text similarly given the missing question-related information inferred by the LLM, which improves inference efficiency but unavoidably at the cost of lacking a holistic grasp of video content, particularly when a long video consists of multiple events (e.g., over 100 events) with significant content variations, and since the CLIP text encoder is trained with simple language structures and limited context (the max length is 77), it struggles to understand complex video semantics, limiting its retrieval accuracy. +- These methods convert videos into captions, addressing long-range reasoning but overlooking critical information loss. Specifically, even if question-related frames are identified, the generated captions cannot cover all the essential information in the key frames. For example, the caption "a woman is looking in the mirror" does not contribute to answering the question "what is the woman wearing when standing in front of the mirror in the video". In this case, LLM can only make an aimless and biased surmise to answer the question. + +To solve the above problem, we turn to how human understands a long video. In particular, a human needs to gain a general understanding of the video, identifying events across segments. Then, the human locates question-related segments, examines them in detail, and answers the question based on the extracted information. Inspired by this, in this paper, we propose DrVideo, a document retrieval-based system for long video understanding that progressively finds key frames and generates a question-related augmented document for LLMs to make confident predictions. Our key idea is to convert long video understanding into long document understanding and leverage the impressive capabilities of LLMs in key information location [46] and long-range reasoning [12] on various NLP tasks. Note that long-document retrieval in our paper means passage retrieval within a single long document, not retrieval from a corpus consisting of long documents. + +Specifically, DrVideo transforms raw long videos into long documents using a pretrained frame/clip-level visual captioner (e.g., LLaVA-NeXT [33]) to make the LLM have a holistic grasp of video content. Different from VideoAgent [61], which performs sparse sampling and identifies key frames via image-text similarity without information augmentation, DrVideo performs dense sampling and proposes a retrieval module and a multi-stage agent interaction loop that dynamically finds the potential missing information through sequential steps of reasoning and then interacts with a document augmentation module to request augmentation of the information within language space. Once sufficient question-related information is gathered, it feeds into the answering module to get the final prediction. + +DrVideo achieves significant performance gains (3.8% on EgoSchema subset [36], 8.3% on MovieChat-1K global mode, 24.8% on MovieChat-1K breakpoint mode [49], and 6.3% on the long split of Video-MME [15]) over the existing LLM-based state-of-the-art methods [14, 37, 61, 74] + +across different long video benchmarks, from 3 minutes to 10 minutes and longer than 1 hour. Surprisingly, combining the subtitles on the long split of Video-MME benchmark, DrVideo obtains $71.7\%$ accuracy, outperforming many heavily-engineered large-scale proprietary models (e.g., Gemini 1.5 Flash [55], GPT-4o mini [44], and GPT-4V [42]). This highlights the great potential of DrVideo. Moreover, DrVideo is a training-free framework and is researcher-friendly, as our results can be replicated on an RTX 4090 with a reasonable number of GPT accesses. + +The contribution of this work can be summarized as: + +- We propose DrVideo, the first document-retrieval-based system designed for the long video understanding task by converting it into a long-document retrieval and understanding task to effectively leverage the power of large language models. +- Different from previous works, we propose a retrieval module and a new multi-stage agent interaction loop that dynamically finds the potential missing information and augments these information within language space. +- DrVideo outperforms the existing LLM-based state-of-the-art methods across different long video benchmarks (i.e., EgoSchema [36], MovieChat-1K [49], and the long split of Video-MME [15]) by a margin. + +# 2. Related Work + +Long Video Understanding. Modeling long videos, which are several minutes or more in length, generally requires advanced temporal modeling, resulting in complex model designs. LF-VILA [52] introduces a Temporal Window Attention (HTWA) mechanism to capture long-range dependencies in long videos. MeMViT [67] and MovieChat [49] employ a memory-based design to save question-related information from previous video segments. Some other approaches employ space-time graphs [21, 62] or relational space-time modules [71] to capture spatio-temporal dependencies from the raw long videos. Recently, S4ND [39], ViS4mer [22], and S5 [58] have used Structured State-Space Sequence (S4) [18] layers to capture long-range dependencies in videos. Different from these methods, our DrVideo does not design a complex module to perform long video understanding. Instead, we develop a document retrieval-based system with LLM for zero-shot long video understanding. + +LLMs for Video Understanding. The recent rise of LLMs [73] and VideoChat [30] align the visual features extracted by pretrained visual models to LLMs and apply them to video understanding. Video ChatCaptioner [7] and ChatVideo [57] utilize LLMs to represent videos and engage users through dialogues, respectively. VidIL [64] applies the image-level models to video understanding tasks via few-shot learning. In addition to short-term video understanding, recent studies [4, 11, 32] have explored LLMs + +for long-range video modeling. For instance, GPT-4 is applied in various long-range video modeling tasks in [32], though quantitative evaluation is limited. Meanwhile, the research in [11] focuses on movie datasets with minimal visual analysis [36], relying largely on speech and subtitles. In contrast, DrVideo focuses on vision modality in multiple benchmarks. + +LLM Agents. In parallel, the computer vision community has started exploring the use of LLMs as agents in various vision tasks such as GUI understanding and robot navigation [5, 13, 20, 53]. In the realm of long video comprehension, initial efforts have employed an agent-like approach, where LLMs interact with external tools or integrate additional functionalities [14, 17, 53, 61, 65, 72]. In contrast to these approaches, our method, DrVideo, reconceptualizes long video understanding as a process of document retrieval, augmentation, and understanding, to leverage the strong capability of LLMs. + +# 3. Methodology + +In this section, we detail our proposed document retrieval-based system for long video understanding. As illustrated in Figure 1, given a long video and a question about the video, DrVideo first translates the long video into a long document, referred to as the initial video document. Then, the retrieval module identifies the top $K$ key frames by calculating the similarity between the question and the initial video document. The document augmentation module enriches the information of these key frames and adds it to the initial video document, creating a new video document (named the updated video document), which is used as the starting point for the subsequent multi-stage agent interaction loop. + +The loop contains two distinct agents: the planning agent and the interaction agent. The planning agent judges whether the updated video document is sufficient for answering the question. If not, the updated video document is fed into the interaction agent to dynamically find missing key frames. The interaction agent interacts with the document augmentation module to request the required information of these key frames. The output of the document augmentation module is added to the current document to obtain the latest updated video document. The new video document is then sent to the planning agent again to further judge whether the current information is sufficient. This searching and interaction loop continues until the planning agent considers the current information sufficient for answering the question or the maximum iteration is reached. After the loop ends, the final video document and the related question are given to the answering module to get the final prediction. + +# 3.1. Video-Document Conversion Module + +By translating each single video frame $V_{t}$ (or a short video clip) into a short description $S_{V_t}$ , we convert an input long video into the initial video document $Doc_{init}$ : + +$$ +D o c _ {i n i t} = \left\{\left\{1, S _ {V _ {1}} \right\}, \left\{2, S _ {V _ {2}} \right\}, \dots , \left\{T, S _ {V _ {T}} \right\} \right\} \tag {1} +$$ + +where $T$ is the length of the video. Specifically, $\mathrm{S}_{V_t}$ is generated by a large vision-language captioning model $\phi_{vlm}$ (e.g., LLaVA-NeXT [33]), i.e., $\mathrm{S}_{V_t} = \phi_{vlm}(\mathcal{P}, V_t)$ , where $\mathcal{P}$ is the prompt for requesting a short description about the image, e.g., describe the picture in no more than 50 words. The captioning model used in DrVideo can be replaced by other captioning models (e.g., LaViLa [78], BLIP-2 [29]). The experimental results with different captioning models can be found in Table 6. + +# 3.2. Document (Frames) Retrieval Module + +After obtaining $Doc_{init}$ , we introduce a document retrieval module to identify the top question-related frames by calculating the similarity between the question and the whole document. Specifically, we use the OpenAI embedding model $\phi_{emb}$ [43] to obtain the vector representation of the initial document, i.e., $\mathcal{E}_{\mathrm{doc}} = \phi_{emb}(Doc_{init})$ . Given a specific retrieval text $\mathcal{RT}$ , i.e., the question $Q$ , the retrieval module first computes the embedding $\mathcal{E}_{\mathcal{RT}} = \phi_{emb}(\mathcal{RT})$ and then retrieves the top $K$ frames based on the cosine similarity between $\mathcal{E}_{\mathcal{RT}}$ and $\mathcal{E}_{\mathrm{doc}}$ : + +$$ +\operatorname {t o p k} _ {\text {一}} \operatorname {d o c} = \underset {t} {\arg \operatorname {t o p}} K \cos \left(\mathcal {E} _ {\mathcal {R T}}, \mathcal {E} _ {\mathrm {d o c} _ {t}}\right) \tag {2} +$$ + +where $\mathcal{E}_{\mathrm{doc}_t}$ is the embedding of the $t$ -th description $S_{V_t}$ . + +# 3.3. Document Augmentation Module + +For each $t' \in \text{topk\_doc}$ , we use the LLaVA-NeXT model [33] with different prompts $\mathcal{AP}$ (augmented prompts) to generate a detailed description $L_{V_{t'}}$ , i.e., $L_{V_{t'}} = \phi_{vlm}(\mathcal{AP}_{t'}, V_{t'})$ . The initial $\mathcal{AP}_{t'}$ is a general prompt: If there are factual errors in the question, provide a precise description of the image; if not, proceed to answer the question: $\{Q\}$ . The updated video document $\mathcal{AD}$ becomes + +$$ +\mathcal {A D} = \left\{\left\{ \begin{array}{l l} \left\{t, S _ {V _ {t}}, L _ {V _ {t} ^ {\prime}} \right\} & \text {i f} t = t ^ {\prime} \\ \left\{t, S _ {V _ {t}} \right\} & \text {o t h e r w i s e} \end{array} \right| t = 1, \dots , T \right\}. \tag {3} +$$ + +Note that $K$ is significantly smaller than the total number of frames (e.g., $K = 5$ compared to $T = 90$ ), so the additional descriptions will not substantially increase the overall length of the document. + +# 3.4. Multi-Stage Agent Interaction Loop + +Apart from the retrieved key frames, we further introduce a multi-stage agent interaction loop to dynamically find other potential key frames and augment them with different types + +![](images/bf3ad29ab9cee039b2f4a4123239ff00ee6e390e8ed9cd05704e9e3713fdadcd.jpg) +Figure 2. Illustration of the multi-stage agent interaction loop and answering module. There are two agents in the multi-stage agent interaction loop: a planning agent to plan the next step and an interaction agent to dynamically find missing information and interact with the document augmentation module. + +of information for the final answer. We employ two distinct agents: planning and interaction agents, each tailored for a specific task. Both agents utilize the same LLM, i.e., GPT-3.5, but are given different prompts and reason varying types of information. The detail is shown in Fig. 2 + +Planning Agent. Given the question $Q$ and the updated video document $\mathcal{AD}_i$ (where $i$ denotes the iteration step), along with the analysis history $\mathcal{H}$ from all previous steps (initialized as empty $\{\}$ ), the planning agent first determines whether the updated video document is sufficient to generate a confident answer. If the information is sufficient, the process moves to the answering module to get the final answer. If not, the agent provides an analysis of why the updated video document $\mathcal{AD}_i$ is inadequate. The agent then updates $\mathcal{H}_i$ with its analysis and passes the updated $\mathcal{H}_i$ to the interaction agent. + +Interaction Agent. Given the current video document $\mathcal{AD}_i$ and the updated $\mathcal{H}_i$ , the interaction agent finds $N$ frames with missing details crucial for answering $Q$ : + +$$ +\forall n \in N, n \in T, n \notin \text {t o p k} _ {-} \text {d o c}, \text {a n d} N < K. \tag {4} +$$ + +In addition to identifying the missing $N$ key frames, the interaction agent also determines the augmented information type for each $n \in N$ . To achieve it, we give a task-specific prompt to the interaction agent, i.e., Your task is to determine which frame needs which type of information and can answer this question accurately, reasonably, and without contradiction. The two types of information are as follows: A: Given an image, get a detailed description of the image (image caption, just like what is shown in this image?) B: Given an image, get a response to the above question (visual question answering). After obtaining $N$ + +
Method(M)LLMSubsetFullset
MoReVQA [37]PaLM-2-51.7
Vamos [60]GPT-451.248.3
ProViQ [9]GPT-3.557.1-
LLoVi [74]GPT-3.557.650.3
IG-VLM [25]GPT-4V59.8-
VideoAgent [61]GPT-460.254.1
MVU [47]Mistral-13B60.337.6
LLoVi [74]GPT-461.2-
VideoAgent [14]GPT-462.860.2
DrVideo (ours)GPT-3.562.6-
DrVideo (ours)GPT-466.461.0
+ +frames along with their required information type, the agent interacts with the document augmentation module to enrich the information of these frames to update the current video document for the next step. + +# 3.5. Answering Module + +Given the final video document $\mathcal{AD}_{final}$ , we employ another agent to provide a prediction using a zero-shot chain-of-thought (CoT) approach [26, 66]. The answering module outputs the corresponding answer, the confidence score, and the reasoning behind the answer. Besides improving prediction accuracy, the CoT approach allows us to trace inference steps, ensuring transparency and explainability in the decision-making process. + +# 4. Experiments + +In this section, we first introduce the datasets and implementation details and then present the results and ablations of our proposed DrVideo. + +# 4.1. Datasets and Metrics + +In our experiments, we evaluate our DrVideo using three well-established datasets, emphasizing its zero-shot long video understanding capabilities. + +EgoSchema. EgoSchema dataset [36] consists of 5000 multiple-choice questions sourced from 5000 three-minute egocentric videos. This dataset only has a subset of 500 questions with publicly accessible labels, while the full set is evaluated on the leaderboard. + +MovieChat-1K. MovieChat-1K [49] is a longer video understanding benchmark, which contains 1000 videos from movies and TV shows, each video lasting approximately 10 minutes. This dataset has two modes: global mode and breakpoint mode. The global mode refers to analyzing the entire video to understand its overall content and context, while the breakpoint mode focuses on analyzing specific frames or scenes. + +Table 1. Results on EgoSchema compared to existing LLM-based state-of-the-art methods. + +
MethodGlobalBreakpoint
Acc.ScoreAcc.Score
Open-Source MLLM
Video LLaMA [75]51.72.7239.12.11
Video Chat [30]57.83.0846.12.32
Video-ChatGPT [35]47.62.8948.02.43
MovieChat [49]68.33.1548.32.46
MovieChat+ [50]71.23.5149.62.62
Based on Open-source Captioners and Proprietary LLMs
MM-VID [32]58.62.8610.40.56
LLoVi [74]58.32.8717.81.03
\( VideoAgent^* \) [61]65.4*3.12*31.6*2.05*
Sullam Jeoung, et al [23]84.8---
DrVideo (ours)93.14.4156.42.75
+ +Table 2. Performance comparison on the MoiveChat-1K [49] benchmark against state-of-the-art methods. * represents re-implemented results, implementation details refer to appendix. + +Video-MME. Video-MME [15] is a recently proposed benchmark for comprehensive video analysis evaluation. We evaluate DrVideo on the "long-term videos" split of this dataset (long split), where video lengths vary from 30 to 60 minutes, with an average duration of 44 minutes. + +Since EgoSchema and Video-MME are multi-choice tasks, we use accuracy as the evaluation metric for both datasets. MovieChat-1K focuses on open-ended questions, we use GPT-assisted evaluation to assess both the accuracy (true/false) and quality (score 0-5) of the models. We select Gemini-Pro [54] as the evaluation assistant and use the same prompt [38] to conduct a fair comparison. + +# 4.2. Implementation Details + +For EgoSchema [36] dataset, we choose LaViLa [78] as the captioning model to convert videos into documents. Note that the training data of our LaViLa model does not include EgoSchema videos, which is the same as LLoVi [74]. For MovieChat-1K and Video-MME benchmarks, LLaVA-NeXT [33] is used as our captioning model to generate brief descriptions for each frame. We preprocess videos by simply sampling them at 0.5 FPS for EgoSchema [36] and MovieChat-1K [49] and 0.2 FPS for Video-MME [15]. LLaVA-NeXT [33] is chosen to augment key frames in both datasets via different prompts. In the comparison experiments, GPT-4 [1], i.e., gpt-4-1106-preview, is used as the agent to evaluate the performance of our DrVideo on EgoSchema and MovieChat-1K. DeepSeek [19], i.e., DeepSeek V2.5, is used as the agent to evaluate the performance on Video-MME. For the ablation study, we use GPT-3.5 [40], i.e., gpt-3.5-turbo-1106, to evaluate the effectiveness of our DrVideo due to the API cost. + +# 4.3. Main Results + +Comparative results on EgoSchema. Table 1 compares our DrVideo with existing LLM-based state-of-the + +
MethodFramesw/o Subsw Subs
Proprietary MLLM
Claude 3.5 Sonnet [2]2051.254.7
GPT-4V [42]1053.556.9
GPT-4o mini [44]25058.663.4
Gemini 1.5 Flash [55]0.5 FPS61.168.8
GPT-4o [44]38465.372.1
Gemini 1.5 Pro [55]0.5 FPS67.477.4
Open-Source MLLM
LongVA [76]12846.247.6
VITA-8x7B [16]3248.650.9
LLaVA-OneVision-72B [28]3260.062.4
Qwen2-VL-72B [3]76862.274.3
Based on Open-source Captioners and Proprietary LLMs
\( VideoAgent^* \)[61]7.640.2*44.4*
\( LLoVi^* \)[74]0.2 FPS45.4*67.7*
DrVideo (ours)0.2 FPS51.771.7
-only subs0.2 FPS-68.5
+ +art methods [9, 14, 25, 37, 47, 60, 61, 74] on EgoSchema benchmark [36]. Specifically, compared with LLoVi [74] and VideoAgent [61] that leverage the same visual captioner (LaViLa [78]) and LLM (GPT-4), DrVideo significantly outperforms these methods by $5.2\%$ and $6.2\%$ respectively. Compared with VideoAgent [14] which uses videospecific models (ViCLIP from InternVid [63] and VideoLLaVA [31]), DrVideo still performs better and achieves $3.6\%$ higher accuracy on the subset evaluation. Moreover, DrVideo significantly outperforms IG-VLM [25] by $6.6\%$ , which uses strong MLLM, i.e., GPT-4V. These results confirm our idea of converting the long-video understanding task into a long-document understanding task is feasible and our DrVideo can find the question-related key frames more accurately. + +Comparative results on longer video benchmarks. To further emphasize the advantage of our method on longer videos, we evaluate our DrVideo on the other two video benchmarks, i.e., MoiveChat-1K (10 minutes) [49], and the long split of Video-MME (average 44 minutes) [15]. Table 2 presents the comparative results of our DrVideo and the current state-of-the-art methods [30, 38, 49, 74, 75] on the MovieChat-1K benchmark. Table 3 shows the comparative results of our DrVideo on the Video-MME benchmark, with comparisons against three types of methods: proprietary MLLMs [2, 42, 44, 55], open-source MLLMs [3, 16, 28, 76], and methods based on open-source caption + +Table 3. Video-MME long split results. $w/o$ subs represents the model without any subtitles. $w$ subs represents that the subtitles corresponding to the sampled frames are both fed into the LLM to evaluate the performance. Our DrVideo outperforms a strong proprietary MLLM (Claude 3.5 Sonnet) and some open-source MLLMs (e.g., VITA-8x7B). * represents re-implemented results, implementation details refer to appendix. + +
RMMSAILCoTAcc.(%)
✓✓✓62.6
✓✓X62.2
✓X✓60.6
X✓✓59.4
XX✓57.4
+ +Table 4. Ablation results on different combinations of the retrieval module (RM), the multi-stage agent interaction loop (MSAIL), and the CoT on EgoSchema. + +
VQACaptionAcc. (%)
✓✓62.6
×✓60.4
✓×61.8
+ +(a) Performance with different types of information + +
Top-KAcc. (%)
562.6
1061.4
2060.6
+ +(b) Performance with different Top- $K$ frames + +Table 5. Ablation results with different settings on EgoSchema. + +ers and proprietary LLMs [61, 74]. + +Compared to methods like LLoVi [74] and VideoAgent [61] that are based on open-source captioners (e.g., LLaVA-NeXT [33]) and proprietary LLMs (e.g., GPT-4 [1]), DrVideo achieves a $8.3\%$ improvement in global mode and a $24.8\%$ improvement in breakpoint mode on the MovieChat-1K benchmark in Table 2. It also achieves at least $6.3\%$ improvement under the w/o subtitle setting on the Video-MME benchmark in Table 3. This indicates that DrVideo can accurately identify question-related frames and avoid critical information loss through augmentation, highlighting the potential of document retrieval methods for processing longer videos. Compared to proprietary MLLMs and open-source MLLMs, DrVideo outperforms Claude 3.5 Sonnet [2] by $0.5\%$ and some open-source MLLMs (e.g., VITA-8x7B [16]) under the w/o subtitle setting on Video-MME benchmark. It also achieves a $21.9\%$ improvement in global mode and a $6.8\%$ improvement in breakpoint mode on MovieChat-1K benchmark. + +Moreover, combining the subtitle on the Video-MME benchmark, DrVideo achieves $71.7\%$ accuracy and outperforms many heavily-engineered large-scale proprietary models (e.g., Gemini 1.5 Flash [55], GPT-4o mini [44], and GPT-4V [42]). To better understand this significant improvement, we conduct an experiment where only the subtitle, question and options are input into the LLM (i.e., DeepSeek [19]) to predict the answer and the accuracy is $68.5\%$ . This suggests: (i) subtitles contain extensive question-related information, playing a crucial role in understanding long videos; (ii) compared with LLoVi and VideoAgent, DrVideo enhances visual information via document retrieval and augmentation, effectively integrating subtitles and resulting in a $3.2\%$ improvement. + +![](images/4dfa8b54a6ab788efd253eb7da839e72d5dd4331d464ab23df0e11ba1fdb0505.jpg) +Figure 3. Performance of different rounds with DrVideo and VideoAgent [61] on EgoSchema. To align with VideoAgent, GPT-4, i.e., gpt-4-turbo-1106-preview, is used as the LLM agents. + +
Captioning ModelTypeAcc.(%)
LaViLa [78]Clip-based62.6
LLaVA-NeXT [33]Frame-based61.2
BLIP-2 [29]Frame-based59.6
+ +# 4.4. Ablation Studies + +Unless otherwise specified, we use GPT-3.5 as the default setting in the below experiments on EgoSchema subset [36]. Effects of different components of DrVideo. We conduct experiments with different combinations of the individual components in DrVideo, including the retrieval module, the multi-stage agent interaction loop, and the CoT used in the answering module. The results are presented in Table 4. We can see that: (i) without the retrieval module and the agent loop, the performance of our DrVideo is similar to LLoVi [74]; (ii) By adding the retrieval module and then the agent loop, the performance of DrVideo improves from $57.4\%$ to $60.6\%$ and then to $62.6\%$ . This demonstrates that the retrieval module can identify key frames via measuring semantic similarity, while the agent loop can locate additional key frames through contextual inference; (iii) the additional CoT contributes to the accuracy improvement from $62.2\%$ to $62.6\%$ , demonstrating its effectiveness. + +Effects of different types of augment information. To better enhance the required information for each potential key frame, we define two types of augment information, namely VQA (question-related information) and caption (general detailed information). Table 5a presents the performance of DrVideo with different types of information. We can see that when only one type of information (VQA or caption) is used for augmentation, the model's performance declines. This suggests the necessity of adaptively augmenting different types of information for each potential key frame. + +Effects of different initial top $K$ key frames. By default, + +Table 6. Performance of different VLMs on EgoSchema. + +
LLMsTypeSizeAcc.(%)
Mistral-8x7B [24]Open-Source8x7B47.6
DeepSeek [19]ProprietaryN/A61.2
GPT-3.5 [40]ProprietaryN/A62.6
GPT-4 [1]ProprietaryN/A66.4
+ +Table 7. Performance of different LLMs on EgoSchema. + +the number of the initial key frames is set as $K = 5$ . Intuitively, more frames should lead to better performance at the cost of increasing the complexity. However, our ablation in Table 5b shows an opposite trend, i.e., more key frames resulting in worse performance. This suggests that enhancing more frames is not necessarily beneficial, as it can introduce a large amount of noisy and irrelevant information that interferes with the LLM's judgment. + +Effects of iterative rounds. We also examined the impact of the iterative rounds $I$ on model performance. Fig. 3 shows the performance of different iterative rounds with DrVideo and VideoAgent [61]. We have the following observations. (i) DrVideo consistently outperforms VideoAgent, reaching the peak performance at $I = 2$ , while VideoAgent peaks at $I = 3$ . This indicates that DrVideo can locate the information needed to answer questions in fewer interactions and with higher quality, underscoring the superiority of our document retrieval-based framework. (ii) When the number of interactions exceeds two, DrVideo's performance declines. This is likely due to more noisy and irrelevant information being introduced into the LLM with the increase of the iterations. This further confirms that excessive information is not necessarily beneficial. + +Effects of different foundation models. DrVideo incorporates two types of foundational models — large language model (LLM), and visual language model (VLM), where VLM is used as the frame captioner while LLM is used as the planning agent, the interaction agent, and the answering agent. To assess the impact of the captions produced by different VLMs, we examine three state-of-the-art VLMs: frame-based BLIP-2 [29], LLaVA-NeXT [33], and clip-based LaViLa [78], as presented in Table 6. We observe that LaViLa achieves the best performance, while BLIP-2 is the worst. It indicates the superiority of the clip-based model which can capture motion information in videos and is more suited for video understanding. + +To evaluate the effectiveness of DrVideo with different LLMs, Table 7 shows the results of DrVideo using open-source LLM (Mistral-8x7B [24]) or Proprietary LLM (DeepSeek [19], GPT-3.5 [40], or GPT-4 [1]) as the planning, interaction and answering agents. From the results, we can see that: (i) different LLMs in DrVideo exhibit significant performance gaps and the proprietary LLM significantly outperforms the open-source LLM, indicating DrVideo's dependency on the choice of LLM; (ii) as the LLM's capability increases, the performance of DrVideo + +Question: By analyzing c's activities in this video, determine their overall focus or intention, and describe the progression of tasks leading to its fulfillment. + +A. C's overall focus or intention in this video is to set up their work station and start working on their computer. +B. The central objective in this video, c's overall focus or intention, is to engage in conversation with the person. +C. C's overall focus or intention in this video is to eat the wrapper. +D. C's overall focus or primary intention in this specific video content is to effectively play the guitar +E. The primary objective and intention in this particular c's video is to demonstrate how to effectively clean the cloth. + +![](images/50c85371c322e3c19aa59852da71c6437cf3c180a65671e89e2cf99d44417dea.jpg) +Figure 4. Case study on an instance from EgoSchema. DrVideo accurately identifies key frames and chooses the correct answer. + +improves, demonstrating the great potential of DrVideo. It's worth noting that our main contribution is the introduction of a framework for translating the long video understanding into a document retrieval and understanding task to effectively leverage the power of large language models, rather than the employment of any specific models. Our DrVideo can be further improved by combining with better LLM and VLM models (e.g., tarsier [59]). + +Case study. Here we give a detailed example from EgoSchema [36] in Fig. 4 to qualitatively demonstrate the capability of our DrVideo. Specifically, the question is to determine the overall focus or intention and describe the progression of tasks leading to its fulfillment. DrVideo first converts the raw video into an initial video document, which is then processed by the frame retrieval module to obtain key frames (frames 15, 46, 61, 84, and 86) and is augmented with detailed key frame descriptions. The updated video document is then fed into the planning agent to assess if sufficient information is available for answering, i.e., the provided information is sufficient to infer character C's + +main focus or intent. This updated document, along with the rationale, is input to the interaction agent to identify the additional key frames (frames 4, 5, 6, 14, and 19 here). After further augmentation, the updated document is sent back to the planning agent, for which the agent confirms its sufficiency. Finally, the answering module takes in the final video document and predicts the correct answer. + +# 5. Conclusion + +We have proposed DrVideo, a document retrieval-based system for long video understanding. Different from previous LLM-based methods, DrVideo proposes to adapt long-video understanding to long-document understanding and searches for missing information via document retrieval and multi-stage agent loop. Extensive comparative experiments and ablation studies on the three challenging datasets, i.e., EgoSchema, MovieChat-1K, and Video-MME, demonstrate the effectiveness of DrVideo. We believe our work can serve as a strong baseline to stimulate further developments on the topic of long video understanding. + +# Acknowledgment + +# References + +[1] Josh Achiam, Steven Adler, Sandhini Agarwal, Lama Ahmad, Ilge Akkaya, Florencia Leoni Aleman, Diogo Almeida, Janko Altenschmidt, Sam Altman, Shyamal Anadkat, et al. Gpt-4 technical report. arXiv preprint arXiv:2303.08774, 2023. 2, 5, 6, 7 +[2] Anthropic. Claude 3.5 Sonnet: Enhanced Conversational AI. https://www.anthropic.com/index/claude-3.5-sonnet/, 2024.6 +[3] Jinze Bai, Shuai Bai, Shusheng Yang, Shijie Wang, Sinan Tan, Peng Wang, Junyang Lin, Chang Zhou, and Jingren Zhou. Qwen-vl: A versatile vision-language model for understanding, localization, text reading, and beyond. arXiv preprint arXiv:2308.12966, 2023. 6 +[4] Aanisha Bhattacharya, Yaman K Singla, Balaji Krishnamurthy, Rajiv Ratn Shah, and Changyou Chen. A video is worth 4096 tokens: Verbalize story videos to understand them in zero shot. arXiv preprint arXiv:2305.09758, 2023. 3 +[5] Anthony Brohan, Noah Brown, Justice Carbajal, Yevgen Chebotar, Xi Chen, Krzysztof Choromanski, Tianli Ding, Danny Driess, Avinava Dubey, Chelsea Finn, et al. Rt-2: Vision-language-action models transfer web knowledge to robotic control. arXiv preprint arXiv:2307.15818, 2023. 3 +[6] Tom Brown, Benjamin Mann, Nick Ryder, Melanie Subbiah, Jared D Kaplan, Prafulla Dhariwal, Arvind Neelakantan, Pranav Shyam, Girish Sastry, Amanda Askell, et al. Language models are few-shot learners. Adv. Neural Inform. Process. Syst., pages 1877-1901, 2020. 1 +[7] Jun Chen, Deyao Zhu, Kilichbek Haydarov, Xiang Li, and Mohamed Elhoseiny. Video chatcaptioner: Towards the enriched spatiotemporal descriptions. arXiv preprint arXiv:2304.04227, 2023. 3 +[8] Feng Cheng, Xizi Wang, Jie Lei, David Crandall, Mohit Bansal, and Gedas Bertasius. Vindlu: A recipe for effective video-and-language pretraining. In IEEE/CVF Conf. Comput. Vis. Pattern Recog., pages 10739-10750, 2023. 1 +[9] Rohan Choudhury, Koichiro Niinuma, Kris M Kitani, and László A Jeni. Zero-shot video question answering with procedural programs. arXiv preprint arXiv:2312.00937, 2023. 5, 6 +[10] Aakanksha Chowdhery, Sharan Narang, Jacob Devlin, Maarten Bosma, Gaurav Mishra, Adam Roberts, Paul Barham, Hyung Won Chung, Charles Sutton, Sebastian Gehrmann, et al. Palm: Scaling language modeling with pathways. arXiv preprint arXiv:2204.02311, 2022. 1 +[11] Jiwan Chung and Youngjae Yu. Long story short: a summarize-then-search method for long video question answering. In Brit. Mach. Vis. Conf., 2023. 3 + +This work is supported by the National Natural Science Foundation of China (62221002, 62171183, U24A20327, 62401202), the Joint Fund of the Ministry of Education of China, and the China Scholarship Council (CSC) under Grant No. 202306130148. +[12] Yiran Ding, Li Lyna Zhang, Chengruidong Zhang, Yuanyuan Xu, Ning Shang, Jiahang Xu, Fan Yang, and Mao Yang. Longgrope: Extending llm context window beyond 2 million tokens. arXiv preprint arXiv:2402.13753, 2024. 2 +[13] Danny Driess, Fei Xia, Mehdi SM Sajjadi, Corey Lynch, Aakanksha Chowdhery, Brian Ichter, Ayzaan Wahid, Jonathan Tompson, Quan Vuong, Tianhe Yu, et al. Palme: An embodied multimodal language model. arXiv preprint arXiv:2303.03378, 2023. 3 +[14] Yue Fan, Xiaojian Ma, Rujie Wu, Yuntao Du, Jiaqi Li, Zhi Gao, and Qing Li. Videoagent: A memory-augmented multimodal agent for video understanding. arXiv preprint arXiv:2403.11481, 2024. 2, 3, 5, 6 +[15] Chaoyou Fu, Yuhan Dai, Yondong Luo, Lei Li, Shuhuai Ren, Renrui Zhang, Zihan Wang, Chenyu Zhou, Yunhang Shen, Mengdan Zhang, et al. Video-mme: The first-ever comprehensive evaluation benchmark of multi-modal llms in video analysis. arXiv preprint arXiv:2405.21075, 2024. 2, 3, 5, 6 +[16] Chaoyou Fu, Haojia Lin, Zuwei Long, Yunhang Shen, Meng Zhao, Yifan Zhang, Xiong Wang, Di Yin, Long Ma, Xiawu Zheng, et al. Vita: Towards open-source interactive omni multimodal llm. arXiv preprint arXiv:2408.05211, 2024. 6 +[17] Difei Gao, Lei Ji, Luowei Zhou, Kevin Qinghong Lin, Joya Chen, Zihan Fan, and Mike Zheng Shou. Assistgpt: A general multi-modal assistant that can plan, execute, inspect, and learn. arXiv preprint arXiv:2306.08640, 2023. 3 +[18] Albert Gu, Karan Goel, and Christopher Ré. Efficiently modeling long sequences with structured state spaces. arXiv preprint arXiv:2111.00396, 2021. 3 +[19] Daya Guo, Qihao Zhu, Dejian Yang, Zhenda Xie, Kai Dong, Wentao Zhang, Guanting Chen, Xiao Bi, Yu Wu, YK Li, et al. Deepseek-coder: When the large language model meets programming-the rise of code intelligence. arXiv preprint arXiv:2401.14196, 2024. 5, 6, 7 +[20] Wenyi Hong, Weihan Wang, Qingsong Lv, Jiazheng Xu, Wenmeng Yu, Junhui Ji, Yan Wang, Zihan Wang, Yuxiao Dong, Ming Ding, et al. Cogagent: A visual language model for gui agents. arXiv preprint arXiv:2312.08914, 2023. 3 +[21] Noureldien Hussein, Efstratios Gavves, and Arnold WM Smeulders. Videograph: Recognizing minutes-long human activities in videos. arXiv preprint arXiv:1905.05143, 2019. 3 +[22] Md Mohaiminul Islam and Gedas Bertasius. Long movie clip classification with state-space video models. In Eur. Conf. Comput. Vis., pages 87-104, 2022. 3 +[23] Sullam Jeoung, Goeric Huybrechts, Bhavana Ganesh, Aram Galstyan, and Sravan Bodapati. Adaptive video understanding agent: Enhancing efficiency with dynamic frame sampling and feedback-driven reasoning. arXiv preprint arXiv:2410.20252, 2024. 5 +[24] Albert Q Jiang, Alexandre Sablayrolles, Arthur Mensch, Chris Bamford, Devendra Singh Chaplot, Diego de las Casas, Florian Bressand, Gianna Lengyel, Guillaume Lample, Lucile Saulnier, et al. Mistral 7b. arXiv preprint arXiv:2310.06825, 2023. 7 +[25] Wonkyun Kim, Changin Choi, Wonseok Lee, and Wonjong Rhee. An image grid can be worth a video: Zero + +shot video question answering using a vlm. arXiv preprint arXiv:2403.18406, 2024. 1, 5, 6 +[26] Takeshi Kojima, Shixiang Shane Gu, Machel Reid, Yutaka Matsuo, and Yusuke Iwasawa. Large language models are zero-shot reasoners. Adv. Neural Inform. Process. Syst., pages 22199-22213, 2022. 5 +[27] Jie Lei, Linjie Li, Luowei Zhou, Zhe Gan, Tamara L Berg, Mohit Bansal, and Jingjing Liu. Less is more: Clipbert for video-and-language learning via sparse sampling. In IEEE/CVF Conf. Comput. Vis. Pattern Recog., pages 7331-7341, 2021. 1 +[28] Bo Li, Yuanhan Zhang, Dong Guo, Renrui Zhang, Feng Li, Hao Zhang, Kaichen Zhang, Yanwei Li, Ziwei Liu, and Chunyuan Li. Llava-onevision: Easy visual task transfer. arXiv preprint arXiv:2408.03326, 2024. 6 +[29] Junnan Li, Dongxu Li, Silvio Savarese, and Steven Hoi. BLIP-2: bootstrapping language-image pre-training with frozen image encoders and large language models. In Proc. Int. Conf. Mach. Learn., 2023. 4, 7 +[30] KunChang Li, Yinan He, Yi Wang, Yizhuo Li, Wenhai Wang, Ping Luo, Yali Wang, Limin Wang, and Yu Qiao. Videochat: Chat-centric video understanding. arXiv preprint arXiv:2305.06355, 2023. 3, 5, 6 +[31] Bin Lin, Bin Zhu, Yang Ye, Munan Ning, Peng Jin, and Li Yuan. Video-llava: Learning united visual representation by alignment before projection. arXiv preprint arXiv:2311.10122, 2023. 6 +[32] Kevin Lin, Faisal Ahmed, Linjie Li, Chung-Ching Lin, Ehsan Azarnasab, Zhengyuan Yang, Jianfeng Wang, Lin Liang, Zicheng Liu, Yumao Lu, et al. Mm-vid: Advancing video understanding with gpt-4v (ision). arXiv preprint arXiv:2310.19773, 2023. 1, 3, 5 +[33] Haotian Liu, Chunyuan Li, Yuheng Li, Bo Li, Yuanhan Zhang, Sheng Shen, and Yong Jae Lee. Llava next: Improved reasoning,OCR,and world knowledge.https://llava-v1.github.io/blog/2024-01-30-llava NEXT/,2024.2,4,5,6,7 +[34] Ziyu Ma, Shutao Li, Bin Sun, Jianfei Cai, Zuxiang Long, and Fuyan Ma. Gerea: Question-aware prompt captions for knowledge-based visual question answering. arXiv preprint arXiv:2402.02503, 2024. 1 +[35] Muhammad Maaz, Hanoona Rasheed, Salman Khan, and Fahad Shahbaz Khan. Video-chatgpt: Towards detailed video understanding via large vision and language models. arXiv preprint arXiv:2306.05424, 2023. 5 +[36] Karttikeya Mangalam, Raiymbek Akshulakov, and Jitendra Malik. Egoschema: A diagnostic benchmark for very long-form video language understanding. arXiv preprint arXiv:2308.09126, 2023. 2, 3, 5, 6, 7, 8 +[37] Juhong Min, Shyamal Buch, Arsha Nagrani, Minsu Cho, and Cordelia Schmid. Morevqa: Exploring modular reasoning models for video question answering. In IEEE/CVF Conf. Comput. Vis. Pattern Recog., pages 13235-13245, 2024. 2, 5, 6 +[38] Salman Khan Muhammad Maaz, Hanoona Rasheed and Fahad Khan. Video-chatgpt: Towards detailed video understanding via large vision and language models. ArXiv 2306.05424, 2023. 1, 5, 6 + +[39] Eric Nguyen, Karan Goel, Albert Gu, Gordon Downs, Preey Shah, Tri Dao, Stephen Baccus, and Christopher Ré. S4nd: Modeling images and videos as multidimensional signals with state spaces. Adv. Neural Inform. Process. Syst., pages 2846-2861, 2022. 3 +[40] OpenAI. GPT3.5. https://platform.openai.com/docs/models/gpt-3-5, 2021.1, 5, 7 +[41] OpenAI. Introducing chatgpt. https://openai.com/blog/chatgpt, 2022.1 +[42] OpenAI. GPT-4V(ison) System Card. https://openai.com/index/gpt-4v-system-card/, 2023.3,6 +[43] OpenAI. New embedding models and api updates. https://openai.com/index/new-embedding-models-and-api-updates/, 2024.4 +[44] OpenAI. GPT-4o Mini: Advancing Cost-Efficient Intelligence. https://openai.com/index/gpt-4o-mini-advancing-cost-efficient-intelligence/, 2024.3,6 +[45] Long Ouyang, Jeffrey Wu, Xu Jiang, Diogo Almeida, Carroll Wainwright, Pamela Mishkin, Chong Zhang, Sandhini Agarwal, Katarina Slama, Alex Ray, et al. Training language models to follow instructions with human feedback. Adv. Neural Inform. Process. Syst., pages 27730-27744, 2022. 1 +[46] Vincent Perot, Kai Kang, Florian Luisier, Guolong Su, Xiaoyu Sun, Ramya Sree Boppana, Zilong Wang, Zifeng Wang, Jiaqi Mu, Hao Zhang, et al. Lmdx: Language model-based document information extraction and localization. arXiv preprint arXiv:2309.10952, 2023. 2 +[47] Kanchana Ranasinghe, Xiang Li, Kumara Kahatapitiya, and Michael S Ryoo. Understanding long videos in one multimodal language model pass. arXiv preprint arXiv:2403.16998, 2024. 5, 6 +[48] Teven Le Scao, Angela Fan, Christopher Akiki, Ellie Pavlick, Suzana Ilic, Daniel Hesslow, Roman Castagné, Alexandra Sasha Luccioni, François Yvon, Matthias Galle, et al. Bloom: A 176b-parameter open-access multilingual language model. arXiv preprint arXiv:2211.05100, 2022. 1 +[49] Enxin Song, Wenhao Chai, Guanhong Wang, Yucheng Zhang, Haoyang Zhou, Feiyang Wu, Xun Guo, Tian Ye, Yan Lu, Jenq-Neng Hwang, et al. Moviechat: From dense token to sparse memory for long video understanding. arXiv preprint arXiv:2307.16449, 2023. 2, 3, 5, 6 +[50] Enxin Song, Wenhao Chai, Tian Ye, Jenq-Neng Hwang, Xi Li, and Gaoang Wang. Moviechat+: Question-aware sparse memory for long video question answering. arXiv preprint arXiv:2404.17176, 2024. 5 +[51] Quan Sun, Jinsheng Wang, Qiying Yu, Yufeng Cui, Fan Zhang, Xiaosong Zhang, and Xinlong Wang. Eva-clip-18b: Scaling clip to 18 billion parameters. arXiv preprint arXiv:2402.04252, 2024. 2 +[52] Yuchong Sun, Hongwei Xue, Ruihua Song, Bei Liu, Huan Yang, and Jianlong Fu. Long-form video-language pretraining with multimodal temporal contrastive learning. Adv. Neural Inform. Process. Syst., 35:38032-38045, 2022. 3 +[53] Dídac Surís, Sachit Menon, and Carl Vondrick. Vipergpt: Visual inference via python execution for reasoning. Int. Conf. Comput. Vis., pages 11888-11898, 2023. 1, 3 + +[54] Gemini Team, Rohan Anil, Sebastian Borgeaud, Yonghui Wu, Jean-Baptiste Alayrac, Jiahui Yu, Radu Soricut, Johan Schalkwyk, Andrew M Dai, Anja Hauth, et al. Gemini: a family of highly capable multimodal models. arXiv preprint arXiv:2312.11805, 2023. 5 +[55] Gemini Team, Petko Georgiev, Ving Ian Lei, Ryan Burnell, Libin Bai, Anmol Gulati, Garrett Tanzer, Damien Vincent, Zhufeng Pan, Shibo Wang, et al. Gemini 1.5: Unlocking multimodal understanding across millions of tokens of context. arXiv preprint arXiv:2403.05530, 2024. 3, 6 +[56] Hugo Touvron, Thibaut Lavril, Gautier Izacard, Xavier Martinet, Marie-Anne Lachaux, Timothée Lacroix, Baptiste Rozière, Naman Goyal, Eric Hambro, Faisal Azhar, et al. Llama: Open and efficient foundation language models. arXiv preprint arXiv:2302.13971, 2023. 1 +[57] Junke Wang, Dongdong Chen, Chong Luo, Xiyang Dai, Lu Yuan, Zuxuan Wu, and Yu-Gang Jiang. Chatvideo: A tracklet-centric multimodal and versatile video understanding system. arXiv preprint arXiv:2304.14407, 2023. 3 +[58] Jue Wang, Wentao Zhu, Pichao Wang, Xiang Yu, Linda Liu, Mohamed Omar, and Raffay Hamid. Selective structured state-spaces for long-form video understanding. In IEEE/CVF Conf. Comput. Vis. Pattern Recog., pages 6387-6397, 2023. 3 +[59] Jiawei Wang, Liping Yuan, and Yuchen Zhang. Tarsier: Recipes for training and evaluating large video description models. arXiv preprint arXiv:2407.00634, 2024. 8 +[60] Shijie Wang, Qi Zhao, Minh Quan Do, Nakul Agarwal, Kwonjoon Lee, and Chen Sun. Vamos: Versatile action models for video understanding. arXiv preprint arXiv:2311.13627, 2023. 5, 6 +[61] Xiaohan Wang, Yuhui Zhang, Orr Zohar, and Serena Yeung-Levy. Videoagent: Long-form video understanding with large language model as agent. arXiv preprint arXiv:2403.10517, 2024. 1, 2, 3, 5, 6, 7 +[62] Yang Wang, Gedas Bertasius, Tae-Hyun Oh, Abhinav Gupta, Minh Hoai, and Lorenzo Torresani. Supervoxel attention graphs for long-range video modeling. In Proc. IEEE/CVF Winter Conf. Appl. Comput. Vis., pages 155-166, 2021. 3 +[63] Yi Wang, Yinan He, Yizhuo Li, Kunchang Li, Jiashuo Yu, Xin Ma, Xinhao Li, Guo Chen, Xinyuan Chen, Yaohui Wang, et al. Intervid: A large-scale video-text dataset for multimodal understanding and generation. arXiv preprint arXiv:2307.06942, 2023. 6 +[64] Zhenhailong Wang, Manling Li, Ruochen Xu, Luowei Zhou, Jie Lei, Xudong Lin, Shuohang Wang, Ziyi Yang, Chenguang Zhu, Derek Hoiem, et al. Language models with image descriptors are strong few-shot video-language learners. Adv. Neural Inform. Process. Syst., pages 8483-8497, 2022. 3 +[65] Ziyang Wang, Shoubin Yu, Elias Stengel-Eskin, Jaehong Yoon, Feng Cheng, Gedas Bertasius, and Mohit Bansal. Videotree: Adaptive tree-based video representation for llm reasoning on long videos. arXiv preprint arXiv:2405.19209, 2024.3 +[66] Jason Wei, Xuezhi Wang, Dale Schuurmans, Maarten Bosma, Fei Xia, Ed Chi, Quoc V Le, Denny Zhou, et al. + +Chain-of-thought prompting elicits reasoning in large language models. Adv. Neural Inform. Process. Syst., pages 24824-24837, 2022. 5 +[67] Chao-Yuan Wu, Yanghao Li, Karttikeya Mangalam, Haoqi Fan, Bo Xiong, Jitendra Malik, and Christoph Feichtenhofer. Memvit: Memory-augmented multiscale vision transformer for efficient long-term video recognition. In IEEE/CVF Conf. Comput. Vis. Pattern Recog., pages 13587-13597, 2022. 3 +[68] Lin Xu, Yilin Zhao, Daquan Zhou, Zhijie Lin, See Kiong Ng, and Jiashi Feng. Pllava: Parameter-free llava extension from images to videos for video dense captioning. arXiv preprint arXiv:2404.16994, 2024. 1 +[69] Antoine Yang, Antoine Miech, Josef Sivic, Ivan Laptev, and Cordelia Schmid. Just ask: Learning to answer questions from millions of narrated videos. In Int. Conf. Comput. Vis., pages 1686-1697, 2021. 1 +[70] Antoine Yang, Antoine Miech, Josef Sivic, Ivan Laptev, and Cordelia Schmid. Zero-shot video question answering via frozen bidirectional language models. Adv. Neural Inform. Process. Syst., 35:124-141, 2022. 1 +[71] Xitong Yang, Fu-Jen Chu, Matt Feiszli, Raghav Goyal, Lorenzo Torresani, and Du Tran. Relational space-time query in long-form videos. In IEEE/CVF Conf. Comput. Vis. Pattern Recog., pages 6398-6408, 2023. 3 +[72] Zongxin Yang, Guikun Chen, Xiaodi Li, Wenguan Wang, and Yi Yang. Doraemongpt: Toward understanding dynamic scenes with large language models. arXiv preprint arXiv:2401.08392, 2024. 3 +[73] Andy Zeng, Maria Attarian, Brian Ichter, Krzysztof Choromanski, Adrian Wong, Stefan Welker, Federico Tombari, Aveek Purohit, Michael Ryoo, Vikas Sindhwani, Johnny Lee, Vincent Vanhoucke, and Pete Florence. Socratic models: Composing zero-shot multimodal reasoning with language. arXiv preprint arXiv:2204.00598, 2022. 3 +[74] Ce Zhang, Taixi Lu, Md Mohaiminul Islam, Ziyang Wang, Shoubin Yu, Mohit Bansal, and Gedas Bertasius. A simple llm framework for long-range video question-answering. arXiv preprint arXiv:2312.17235, 2023. 1, 2, 5, 6, 7 +[75] Hang Zhang, Xin Li, and Lidong Bing. Video-llama: An instruction-tuned audio-visual language model for video understanding. arXiv preprint arXiv:2306.02858, 2023. 5, 6 +[76] Peiyuan Zhang, Kaichen Zhang, Bo Li, Guangtao Zeng, Jingkang Yang, Yuanhan Zhang, Ziyue Wang, Haoran Tan, Chunyuan Li, and Ziwei Liu. Long context transfer from language to vision. arXiv preprint arXiv:2406.16852, 2024. 6 +[77] Yuanhan Zhang, Bo Li, Haotian Liu, Yong Jae Lee, Liangke Gui, Di Fu, Jiashi Feng, Ziwei Liu, and Chunyuan Li. Llava-last: A strong zero-shot video understanding model. https://llava-v1.github.io/blog/2024-04-30-llava-last-video/, 2024.1 +[78] Yue Zhao, Ishan Misra, Philipp Krahenbuhl, and Rohit Girdhar. Learning video representations from large language models. In IEEE/CVF Conf. Comput. Vis. Pattern Recog., pages 6586-6597, 2023. 1, 2, 4, 5, 6, 7 \ No newline at end of file diff --git a/2025/DrVideo_ Document Retrieval Based Long Video Understanding/images.zip b/2025/DrVideo_ Document Retrieval Based Long Video Understanding/images.zip new file mode 100644 index 0000000000000000000000000000000000000000..b8b572132f7b5299f8a6e50af70ffc45b8bf82a6 --- /dev/null +++ b/2025/DrVideo_ Document Retrieval Based Long Video Understanding/images.zip @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:301fb12e31f37e53860e051e5cca187962f0e97e3505de9ba642f0e2443231d1 +size 587250 diff --git a/2025/DrVideo_ Document Retrieval Based Long Video Understanding/layout.json b/2025/DrVideo_ Document Retrieval Based Long Video Understanding/layout.json new file mode 100644 index 0000000000000000000000000000000000000000..2b36eaa458243eead5ec48c666f9b6945b7da747 --- /dev/null +++ b/2025/DrVideo_ Document Retrieval Based Long Video Understanding/layout.json @@ -0,0 +1,9058 @@ +{ + "pdf_info": [ + { + "para_blocks": [ + { + "bbox": [ + 107, + 103, + 504, + 121 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 107, + 103, + 504, + 121 + ], + "spans": [ + { + "bbox": [ + 107, + 103, + 504, + 121 + ], + "type": "text", + "content": "DrVideo: Document Retrieval Based Long Video Understanding" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 145, + 142, + 463, + 171 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 145, + 142, + 463, + 171 + ], + "spans": [ + { + "bbox": [ + 145, + 142, + 463, + 171 + ], + "type": "text", + "content": "Ziyu Ma" + }, + { + "bbox": [ + 145, + 142, + 463, + 171 + ], + "type": "inline_equation", + "content": "^{1,2}" + }, + { + "bbox": [ + 145, + 142, + 463, + 171 + ], + "type": "text", + "content": ", Chenhui Gou" + }, + { + "bbox": [ + 145, + 142, + 463, + 171 + ], + "type": "inline_equation", + "content": "^{2}" + }, + { + "bbox": [ + 145, + 142, + 463, + 171 + ], + "type": "text", + "content": ", Hengcan Shi" + }, + { + "bbox": [ + 145, + 142, + 463, + 171 + ], + "type": "inline_equation", + "content": "^{1,2}" + }, + { + "bbox": [ + 145, + 142, + 463, + 171 + ], + "type": "text", + "content": ", Bin Sun" + }, + { + "bbox": [ + 145, + 142, + 463, + 171 + ], + "type": "inline_equation", + "content": "^{1}" + }, + { + "bbox": [ + 145, + 142, + 463, + 171 + ], + "type": "text", + "content": ", Shutao Li" + }, + { + "bbox": [ + 145, + 142, + 463, + 171 + ], + "type": "inline_equation", + "content": "^{1}" + }, + { + "bbox": [ + 145, + 142, + 463, + 171 + ], + "type": "text", + "content": ", Hamid Rezatofighi" + }, + { + "bbox": [ + 145, + 142, + 463, + 171 + ], + "type": "inline_equation", + "content": "^{2}" + }, + { + "bbox": [ + 145, + 142, + 463, + 171 + ], + "type": "text", + "content": ", Jianfei Cai" + }, + { + "bbox": [ + 145, + 142, + 463, + 171 + ], + "type": "inline_equation", + "content": "^{2}" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 133, + 171, + 477, + 186 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 171, + 477, + 186 + ], + "spans": [ + { + "bbox": [ + 133, + 171, + 477, + 186 + ], + "type": "text", + "content": "1 College of Electrical and Information Engineering, Hunan University," + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 143, + 186, + 468, + 200 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 143, + 186, + 468, + 200 + ], + "spans": [ + { + "bbox": [ + 143, + 186, + 468, + 200 + ], + "type": "text", + "content": "2 Data Science & AI Department, Faculty of IT, Monash University" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 195, + 201, + 414, + 213 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 195, + 201, + 414, + 213 + ], + "spans": [ + { + "bbox": [ + 195, + 201, + 414, + 213 + ], + "type": "text", + "content": "{maziyu, sunbin611, shutao_li}@hnu.edu.cn" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 117, + 216, + 490, + 228 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 117, + 216, + 490, + 228 + ], + "spans": [ + { + "bbox": [ + 117, + 216, + 490, + 228 + ], + "type": "text", + "content": "{chenhui.gou, hengcan.shi, hamid.rezatofighi, jianfei.cai}@monash.edu" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 151, + 255, + 200, + 268 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 151, + 255, + 200, + 268 + ], + "spans": [ + { + "bbox": [ + 151, + 255, + 200, + 268 + ], + "type": "text", + "content": "Abstract" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 54, + 280, + 297, + 579 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 54, + 280, + 297, + 579 + ], + "spans": [ + { + "bbox": [ + 54, + 280, + 297, + 579 + ], + "type": "text", + "content": "Most of the existing methods for video understanding primarily focus on videos only lasting tens of seconds, with limited exploration of techniques for handling long videos. The increased number of frames in long videos poses two main challenges: difficulty in locating key information and performing long-range reasoning. Thus, we propose DrVideo, a document-retrieval-based system designed for long video understanding. Our key idea is to convert the long-video understanding problem into a long-document understanding task so as to effectively leverage the power of large language models. Specifically, DrVideo first transforms a long video into a coarse text-based long document to initially retrieve key frames and then updates the documents with the augmented key frame information. It then employs an agent-based iterative loop to continuously search for missing information and augment the document until sufficient question-related information is gathered for making the final predictions in a chain-of-thought manner. Extensive experiments on long video benchmarks confirm the effectiveness of our method. DrVideo significantly outperforms existing LLM-based state-of-the-art methods on EgoSchema benchmark (3 minutes), MovieChat-1K benchmark (10 minutes), and the long split of Video-MME benchmark (average of 44 minutes). Code is available at https://github.com/Upper9527/DrVideo." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 55, + 604, + 135, + 616 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 604, + 135, + 616 + ], + "spans": [ + { + "bbox": [ + 55, + 604, + 135, + 616 + ], + "type": "text", + "content": "1. Introduction" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 55, + 624, + 296, + 685 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 624, + 296, + 685 + ], + "spans": [ + { + "bbox": [ + 55, + 624, + 296, + 685 + ], + "type": "text", + "content": "Video understanding is a challenging task in computer vision, requiring the processing of spatio-temporal information and advanced reasoning abilities. Previous works have successfully processed short videos lasting around tens of seconds [8, 27, 32, 53, 69, 70]. However, how to deal with" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 313, + 256, + 554, + 340 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 256, + 554, + 340 + ], + "spans": [ + { + "bbox": [ + 313, + 256, + 554, + 340 + ], + "type": "text", + "content": "long video understanding remains unclear. Recent advancements in large language models (LLMs) [6, 10, 34, 41, 45, 48, 56] demonstrate strong abilities in language understanding and reasoning across long text sequences. These advancements have inspired the development of videolanguage models (Video-LMs) [25, 38, 68, 77] to address long video understanding issues." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 313, + 341, + 555, + 544 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 341, + 555, + 544 + ], + "spans": [ + { + "bbox": [ + 313, + 341, + 555, + 544 + ], + "type": "text", + "content": "Video-LMs typically encode a video as a sequence of visual tokens, which are then concatenated with language tokens into a long sequence and LLM is then used to understand this long sequence. Although Video-LMs improve video understanding, they have the following limitations. (i) They cannot process an entire video as input to a VideoLM (e.g., the visual encoder OpenAI's CLIP-L-14 outputs " + }, + { + "bbox": [ + 313, + 341, + 555, + 544 + ], + "type": "inline_equation", + "content": "24 \\times 24" + }, + { + "bbox": [ + 313, + 341, + 555, + 544 + ], + "type": "text", + "content": " tokens for each image, while the max length of LLaVA-NeXT-Video [77] is only 8192). (ii) They typically perform uniform or random frame sampling at large strides to handle long videos without considering the content (e.g., 16 for PLLaVA [68] and 6 for IG-VLM [25]), leading to potential key information loss. (iii) The simple concatenation of visual tokens increases the difficulty for the LLM to locate question-related (key) visual information within the long video sequence, complicating long-range reasoning across the vision token sequences." + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 313, + 545, + 556, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 545, + 556, + 715 + ], + "spans": [ + { + "bbox": [ + 313, + 545, + 556, + 715 + ], + "type": "text", + "content": "Apart from the studies of encoding a video as a sequence of visual tokens, another line of research converts raw videos into captions and then utilities the long-range reasoning abilities of LLMs to predict the final answer. A pioneering work - LLoVi [74] uses a short-term video model [78] in conjunction with an LLM [40] to solve the long video understanding task. Given a long video, LLoVi first divides it into multiple short clips and converts them into short textual descriptions. Afterward, these short textual descriptions are summarized by the LLM, and the summary is finally used by LLoVi to answer the given question. Inspired by LLoVi, VideoAgent [61] designs an agent-based system to further leverage the reasoning abilities of LLMs and locate question-related video clips or images. Initially," + } + ] + } + ], + "index": 14 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "spans": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "type": "text", + "content": "CVF" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 145, + 0, + 496, + 37 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 145, + 0, + 496, + 37 + ], + "spans": [ + { + "bbox": [ + 145, + 0, + 496, + 37 + ], + "type": "text", + "content": "This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 66, + 693, + 135, + 703 + ], + "type": "page_footnote", + "angle": 0, + "lines": [ + { + "bbox": [ + 66, + 693, + 135, + 703 + ], + "spans": [ + { + "bbox": [ + 66, + 693, + 135, + 703 + ], + "type": "text", + "content": "*Equal contribution." + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 67, + 703, + 144, + 712 + ], + "type": "page_footnote", + "angle": 0, + "lines": [ + { + "bbox": [ + 67, + 703, + 144, + 712 + ], + "spans": [ + { + "bbox": [ + 67, + 703, + 144, + 712 + ], + "type": "text", + "content": "† Corresponding author." + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "type": "text", + "content": "18936" + } + ] + } + ], + "index": 17 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 0 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 83, + 68, + 527, + 234 + ], + "blocks": [ + { + "bbox": [ + 83, + 68, + 527, + 234 + ], + "lines": [ + { + "bbox": [ + 83, + 68, + 527, + 234 + ], + "spans": [ + { + "bbox": [ + 83, + 68, + 527, + 234 + ], + "type": "image", + "image_path": "aef6f5c7e3fbd886231fae7a678d4a9ff61203dc2838a65af6ffb065fa5d0703.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 55, + 243, + 555, + 268 + ], + "lines": [ + { + "bbox": [ + 55, + 243, + 555, + 268 + ], + "spans": [ + { + "bbox": [ + 55, + 243, + 555, + 268 + ], + "type": "text", + "content": "Figure 1. Overview of our DrVideo framework. It comprises five components: a video-document conversion module, a retrieval module, a document augmentation module, a multi-stage agent interaction loop and an answering module." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "bbox": [ + 54, + 275, + 297, + 407 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 54, + 275, + 297, + 407 + ], + "spans": [ + { + "bbox": [ + 54, + 275, + 297, + 407 + ], + "type": "text", + "content": "VideoAgent samples only a minimal number of frames from the video (i.e., 5 frames), converting them into textual descriptions as input for the LLM. Then, through continuous feedback and interaction with the LLM [1], VLM [78], and CLIP [51], it gradually finds question-related frames in a coarse-to-fine manner via image-text similarity. Finally, VideoAgent answers the question based on these question-related frames. Despite these LLM-based methods overcoming the challenge of performing long-range reasoning and achieving promising results, they still suffer from the following problems." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 55, + 415, + 297, + 714 + ], + "type": "list", + "angle": 0, + "index": 5, + "blocks": [ + { + "bbox": [ + 55, + 415, + 297, + 594 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 415, + 297, + 594 + ], + "spans": [ + { + "bbox": [ + 55, + 415, + 297, + 594 + ], + "type": "text", + "content": "- These methods struggle to accurately identify question-related frames. For instance, VideoAgent [61] performs sparse sampling on long videos (e.g., selecting 5 frames) and subsequently searches for keyframes from the unsampled frames in a coarse-to-fine manner based on the initial sparse space via CLIP image-text similarly given the missing question-related information inferred by the LLM, which improves inference efficiency but unavoidably at the cost of lacking a holistic grasp of video content, particularly when a long video consists of multiple events (e.g., over 100 events) with significant content variations, and since the CLIP text encoder is trained with simple language structures and limited context (the max length is 77), it struggles to understand complex video semantics, limiting its retrieval accuracy." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 55, + 594, + 297, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 594, + 297, + 714 + ], + "spans": [ + { + "bbox": [ + 55, + 594, + 297, + 714 + ], + "type": "text", + "content": "- These methods convert videos into captions, addressing long-range reasoning but overlooking critical information loss. Specifically, even if question-related frames are identified, the generated captions cannot cover all the essential information in the key frames. For example, the caption \"a woman is looking in the mirror\" does not contribute to answering the question \"what is the woman wearing when standing in front of the mirror in the video\". In this case, LLM can only make an aimless and biased surmise to answer the question." + } + ] + } + ], + "index": 4 + } + ], + "sub_type": "text" + }, + { + "bbox": [ + 313, + 275, + 556, + 479 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 275, + 556, + 479 + ], + "spans": [ + { + "bbox": [ + 313, + 275, + 556, + 479 + ], + "type": "text", + "content": "To solve the above problem, we turn to how human understands a long video. In particular, a human needs to gain a general understanding of the video, identifying events across segments. Then, the human locates question-related segments, examines them in detail, and answers the question based on the extracted information. Inspired by this, in this paper, we propose DrVideo, a document retrieval-based system for long video understanding that progressively finds key frames and generates a question-related augmented document for LLMs to make confident predictions. Our key idea is to convert long video understanding into long document understanding and leverage the impressive capabilities of LLMs in key information location [46] and long-range reasoning [12] on various NLP tasks. Note that long-document retrieval in our paper means passage retrieval within a single long document, not retrieval from a corpus consisting of long documents." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 313, + 482, + 556, + 651 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 482, + 556, + 651 + ], + "spans": [ + { + "bbox": [ + 313, + 482, + 556, + 651 + ], + "type": "text", + "content": "Specifically, DrVideo transforms raw long videos into long documents using a pretrained frame/clip-level visual captioner (e.g., LLaVA-NeXT [33]) to make the LLM have a holistic grasp of video content. Different from VideoAgent [61], which performs sparse sampling and identifies key frames via image-text similarity without information augmentation, DrVideo performs dense sampling and proposes a retrieval module and a multi-stage agent interaction loop that dynamically finds the potential missing information through sequential steps of reasoning and then interacts with a document augmentation module to request augmentation of the information within language space. Once sufficient question-related information is gathered, it feeds into the answering module to get the final prediction." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 313, + 653, + 557, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 653, + 557, + 715 + ], + "spans": [ + { + "bbox": [ + 313, + 653, + 557, + 715 + ], + "type": "text", + "content": "DrVideo achieves significant performance gains (3.8% on EgoSchema subset [36], 8.3% on MovieChat-1K global mode, 24.8% on MovieChat-1K breakpoint mode [49], and 6.3% on the long split of Video-MME [15]) over the existing LLM-based state-of-the-art methods [14, 37, 61, 74]" + } + ] + } + ], + "index": 8 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "type": "text", + "content": "18937" + } + ] + } + ], + "index": 9 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 1 + }, + { + "para_blocks": [ + { + "bbox": [ + 55, + 72, + 294, + 191 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 72, + 294, + 191 + ], + "spans": [ + { + "bbox": [ + 55, + 72, + 294, + 191 + ], + "type": "text", + "content": "across different long video benchmarks, from 3 minutes to 10 minutes and longer than 1 hour. Surprisingly, combining the subtitles on the long split of Video-MME benchmark, DrVideo obtains " + }, + { + "bbox": [ + 55, + 72, + 294, + 191 + ], + "type": "inline_equation", + "content": "71.7\\%" + }, + { + "bbox": [ + 55, + 72, + 294, + 191 + ], + "type": "text", + "content": " accuracy, outperforming many heavily-engineered large-scale proprietary models (e.g., Gemini 1.5 Flash [55], GPT-4o mini [44], and GPT-4V [42]). This highlights the great potential of DrVideo. Moreover, DrVideo is a training-free framework and is researcher-friendly, as our results can be replicated on an RTX 4090 with a reasonable number of GPT accesses." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 67, + 192, + 281, + 203 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 67, + 192, + 281, + 203 + ], + "spans": [ + { + "bbox": [ + 67, + 192, + 281, + 203 + ], + "type": "text", + "content": "The contribution of this work can be summarized as:" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 55, + 204, + 295, + 360 + ], + "type": "list", + "angle": 0, + "index": 5, + "blocks": [ + { + "bbox": [ + 55, + 204, + 295, + 262 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 204, + 295, + 262 + ], + "spans": [ + { + "bbox": [ + 55, + 204, + 295, + 262 + ], + "type": "text", + "content": "- We propose DrVideo, the first document-retrieval-based system designed for the long video understanding task by converting it into a long-document retrieval and understanding task to effectively leverage the power of large language models." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 55, + 264, + 295, + 311 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 264, + 295, + 311 + ], + "spans": [ + { + "bbox": [ + 55, + 264, + 295, + 311 + ], + "type": "text", + "content": "- Different from previous works, we propose a retrieval module and a new multi-stage agent interaction loop that dynamically finds the potential missing information and augments these information within language space." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 55, + 312, + 295, + 360 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 312, + 295, + 360 + ], + "spans": [ + { + "bbox": [ + 55, + 312, + 295, + 360 + ], + "type": "text", + "content": "- DrVideo outperforms the existing LLM-based state-of-the-art methods across different long video benchmarks (i.e., EgoSchema [36], MovieChat-1K [49], and the long split of Video-MME [15]) by a margin." + } + ] + } + ], + "index": 4 + } + ], + "sub_type": "text" + }, + { + "bbox": [ + 55, + 370, + 142, + 383 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 370, + 142, + 383 + ], + "spans": [ + { + "bbox": [ + 55, + 370, + 142, + 383 + ], + "type": "text", + "content": "2. Related Work" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 55, + 390, + 295, + 605 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 390, + 295, + 605 + ], + "spans": [ + { + "bbox": [ + 55, + 390, + 295, + 605 + ], + "type": "text", + "content": "Long Video Understanding. Modeling long videos, which are several minutes or more in length, generally requires advanced temporal modeling, resulting in complex model designs. LF-VILA [52] introduces a Temporal Window Attention (HTWA) mechanism to capture long-range dependencies in long videos. MeMViT [67] and MovieChat [49] employ a memory-based design to save question-related information from previous video segments. Some other approaches employ space-time graphs [21, 62] or relational space-time modules [71] to capture spatio-temporal dependencies from the raw long videos. Recently, S4ND [39], ViS4mer [22], and S5 [58] have used Structured State-Space Sequence (S4) [18] layers to capture long-range dependencies in videos. Different from these methods, our DrVideo does not design a complex module to perform long video understanding. Instead, we develop a document retrieval-based system with LLM for zero-shot long video understanding." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 55, + 606, + 295, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 606, + 295, + 714 + ], + "spans": [ + { + "bbox": [ + 55, + 606, + 295, + 714 + ], + "type": "text", + "content": "LLMs for Video Understanding. The recent rise of LLMs [73] and VideoChat [30] align the visual features extracted by pretrained visual models to LLMs and apply them to video understanding. Video ChatCaptioner [7] and ChatVideo [57] utilize LLMs to represent videos and engage users through dialogues, respectively. VidIL [64] applies the image-level models to video understanding tasks via few-shot learning. In addition to short-term video understanding, recent studies [4, 11, 32] have explored LLMs" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 313, + 72, + 553, + 156 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 72, + 553, + 156 + ], + "spans": [ + { + "bbox": [ + 313, + 72, + 553, + 156 + ], + "type": "text", + "content": "for long-range video modeling. For instance, GPT-4 is applied in various long-range video modeling tasks in [32], though quantitative evaluation is limited. Meanwhile, the research in [11] focuses on movie datasets with minimal visual analysis [36], relying largely on speech and subtitles. In contrast, DrVideo focuses on vision modality in multiple benchmarks." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 313, + 159, + 555, + 292 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 159, + 555, + 292 + ], + "spans": [ + { + "bbox": [ + 313, + 159, + 555, + 292 + ], + "type": "text", + "content": "LLM Agents. In parallel, the computer vision community has started exploring the use of LLMs as agents in various vision tasks such as GUI understanding and robot navigation [5, 13, 20, 53]. In the realm of long video comprehension, initial efforts have employed an agent-like approach, where LLMs interact with external tools or integrate additional functionalities [14, 17, 53, 61, 65, 72]. In contrast to these approaches, our method, DrVideo, reconceptualizes long video understanding as a process of document retrieval, augmentation, and understanding, to leverage the strong capability of LLMs." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 313, + 315, + 397, + 328 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 315, + 397, + 328 + ], + "spans": [ + { + "bbox": [ + 313, + 315, + 397, + 328 + ], + "type": "text", + "content": "3. Methodology" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 313, + 338, + 555, + 495 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 338, + 555, + 495 + ], + "spans": [ + { + "bbox": [ + 313, + 338, + 555, + 495 + ], + "type": "text", + "content": "In this section, we detail our proposed document retrieval-based system for long video understanding. As illustrated in Figure 1, given a long video and a question about the video, DrVideo first translates the long video into a long document, referred to as the initial video document. Then, the retrieval module identifies the top " + }, + { + "bbox": [ + 313, + 338, + 555, + 495 + ], + "type": "inline_equation", + "content": "K" + }, + { + "bbox": [ + 313, + 338, + 555, + 495 + ], + "type": "text", + "content": " key frames by calculating the similarity between the question and the initial video document. The document augmentation module enriches the information of these key frames and adds it to the initial video document, creating a new video document (named the updated video document), which is used as the starting point for the subsequent multi-stage agent interaction loop." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 313, + 498, + 555, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 498, + 555, + 713 + ], + "spans": [ + { + "bbox": [ + 313, + 498, + 555, + 713 + ], + "type": "text", + "content": "The loop contains two distinct agents: the planning agent and the interaction agent. The planning agent judges whether the updated video document is sufficient for answering the question. If not, the updated video document is fed into the interaction agent to dynamically find missing key frames. The interaction agent interacts with the document augmentation module to request the required information of these key frames. The output of the document augmentation module is added to the current document to obtain the latest updated video document. The new video document is then sent to the planning agent again to further judge whether the current information is sufficient. This searching and interaction loop continues until the planning agent considers the current information sufficient for answering the question or the maximum iteration is reached. After the loop ends, the final video document and the related question are given to the answering module to get the final prediction." + } + ] + } + ], + "index": 13 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 749, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 749, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 749, + 317, + 757 + ], + "type": "text", + "content": "18938" + } + ] + } + ], + "index": 14 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 2 + }, + { + "para_blocks": [ + { + "bbox": [ + 55, + 72, + 251, + 83 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 72, + 251, + 83 + ], + "spans": [ + { + "bbox": [ + 55, + 72, + 251, + 83 + ], + "type": "text", + "content": "3.1. Video-Document Conversion Module" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 55, + 89, + 295, + 125 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 89, + 295, + 125 + ], + "spans": [ + { + "bbox": [ + 55, + 89, + 295, + 125 + ], + "type": "text", + "content": "By translating each single video frame " + }, + { + "bbox": [ + 55, + 89, + 295, + 125 + ], + "type": "inline_equation", + "content": "V_{t}" + }, + { + "bbox": [ + 55, + 89, + 295, + 125 + ], + "type": "text", + "content": " (or a short video clip) into a short description " + }, + { + "bbox": [ + 55, + 89, + 295, + 125 + ], + "type": "inline_equation", + "content": "S_{V_t}" + }, + { + "bbox": [ + 55, + 89, + 295, + 125 + ], + "type": "text", + "content": ", we convert an input long video into the initial video document " + }, + { + "bbox": [ + 55, + 89, + 295, + 125 + ], + "type": "inline_equation", + "content": "Doc_{init}" + }, + { + "bbox": [ + 55, + 89, + 295, + 125 + ], + "type": "text", + "content": ":" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 79, + 133, + 295, + 148 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 79, + 133, + 295, + 148 + ], + "spans": [ + { + "bbox": [ + 79, + 133, + 295, + 148 + ], + "type": "interline_equation", + "content": "D o c _ {i n i t} = \\left\\{\\left\\{1, S _ {V _ {1}} \\right\\}, \\left\\{2, S _ {V _ {2}} \\right\\}, \\dots , \\left\\{T, S _ {V _ {T}} \\right\\} \\right\\} \\tag {1}", + "image_path": "91337ba9c7e641c2c5e5874597c842d55233d5a4899e749c4261c3e9f227e100.jpg" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 55, + 155, + 296, + 262 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 155, + 296, + 262 + ], + "spans": [ + { + "bbox": [ + 55, + 155, + 296, + 262 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 55, + 155, + 296, + 262 + ], + "type": "inline_equation", + "content": "T" + }, + { + "bbox": [ + 55, + 155, + 296, + 262 + ], + "type": "text", + "content": " is the length of the video. Specifically, " + }, + { + "bbox": [ + 55, + 155, + 296, + 262 + ], + "type": "inline_equation", + "content": "\\mathrm{S}_{V_t}" + }, + { + "bbox": [ + 55, + 155, + 296, + 262 + ], + "type": "text", + "content": " is generated by a large vision-language captioning model " + }, + { + "bbox": [ + 55, + 155, + 296, + 262 + ], + "type": "inline_equation", + "content": "\\phi_{vlm}" + }, + { + "bbox": [ + 55, + 155, + 296, + 262 + ], + "type": "text", + "content": " (e.g., LLaVA-NeXT [33]), i.e., " + }, + { + "bbox": [ + 55, + 155, + 296, + 262 + ], + "type": "inline_equation", + "content": "\\mathrm{S}_{V_t} = \\phi_{vlm}(\\mathcal{P}, V_t)" + }, + { + "bbox": [ + 55, + 155, + 296, + 262 + ], + "type": "text", + "content": ", where " + }, + { + "bbox": [ + 55, + 155, + 296, + 262 + ], + "type": "inline_equation", + "content": "\\mathcal{P}" + }, + { + "bbox": [ + 55, + 155, + 296, + 262 + ], + "type": "text", + "content": " is the prompt for requesting a short description about the image, e.g., describe the picture in no more than 50 words. The captioning model used in DrVideo can be replaced by other captioning models (e.g., LaViLa [78], BLIP-2 [29]). The experimental results with different captioning models can be found in Table 6." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 55, + 270, + 256, + 282 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 270, + 256, + 282 + ], + "spans": [ + { + "bbox": [ + 55, + 270, + 256, + 282 + ], + "type": "text", + "content": "3.2. Document (Frames) Retrieval Module" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 55, + 287, + 296, + 407 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 287, + 296, + 407 + ], + "spans": [ + { + "bbox": [ + 55, + 287, + 296, + 407 + ], + "type": "text", + "content": "After obtaining " + }, + { + "bbox": [ + 55, + 287, + 296, + 407 + ], + "type": "inline_equation", + "content": "Doc_{init}" + }, + { + "bbox": [ + 55, + 287, + 296, + 407 + ], + "type": "text", + "content": ", we introduce a document retrieval module to identify the top question-related frames by calculating the similarity between the question and the whole document. Specifically, we use the OpenAI embedding model " + }, + { + "bbox": [ + 55, + 287, + 296, + 407 + ], + "type": "inline_equation", + "content": "\\phi_{emb}" + }, + { + "bbox": [ + 55, + 287, + 296, + 407 + ], + "type": "text", + "content": " [43] to obtain the vector representation of the initial document, i.e., " + }, + { + "bbox": [ + 55, + 287, + 296, + 407 + ], + "type": "inline_equation", + "content": "\\mathcal{E}_{\\mathrm{doc}} = \\phi_{emb}(Doc_{init})" + }, + { + "bbox": [ + 55, + 287, + 296, + 407 + ], + "type": "text", + "content": ". Given a specific retrieval text " + }, + { + "bbox": [ + 55, + 287, + 296, + 407 + ], + "type": "inline_equation", + "content": "\\mathcal{RT}" + }, + { + "bbox": [ + 55, + 287, + 296, + 407 + ], + "type": "text", + "content": ", i.e., the question " + }, + { + "bbox": [ + 55, + 287, + 296, + 407 + ], + "type": "inline_equation", + "content": "Q" + }, + { + "bbox": [ + 55, + 287, + 296, + 407 + ], + "type": "text", + "content": ", the retrieval module first computes the embedding " + }, + { + "bbox": [ + 55, + 287, + 296, + 407 + ], + "type": "inline_equation", + "content": "\\mathcal{E}_{\\mathcal{RT}} = \\phi_{emb}(\\mathcal{RT})" + }, + { + "bbox": [ + 55, + 287, + 296, + 407 + ], + "type": "text", + "content": " and then retrieves the top " + }, + { + "bbox": [ + 55, + 287, + 296, + 407 + ], + "type": "inline_equation", + "content": "K" + }, + { + "bbox": [ + 55, + 287, + 296, + 407 + ], + "type": "text", + "content": " frames based on the cosine similarity between " + }, + { + "bbox": [ + 55, + 287, + 296, + 407 + ], + "type": "inline_equation", + "content": "\\mathcal{E}_{\\mathcal{RT}}" + }, + { + "bbox": [ + 55, + 287, + 296, + 407 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 55, + 287, + 296, + 407 + ], + "type": "inline_equation", + "content": "\\mathcal{E}_{\\mathrm{doc}}" + }, + { + "bbox": [ + 55, + 287, + 296, + 407 + ], + "type": "text", + "content": ":" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 91, + 409, + 295, + 429 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 91, + 409, + 295, + 429 + ], + "spans": [ + { + "bbox": [ + 91, + 409, + 295, + 429 + ], + "type": "interline_equation", + "content": "\\operatorname {t o p k} _ {\\text {一}} \\operatorname {d o c} = \\underset {t} {\\arg \\operatorname {t o p}} K \\cos \\left(\\mathcal {E} _ {\\mathcal {R T}}, \\mathcal {E} _ {\\mathrm {d o c} _ {t}}\\right) \\tag {2}", + "image_path": "440c2aebc81c6d147c38ecfde942fa685e1ffb5f4172c7a25268216b338a92d2.jpg" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 55, + 431, + 285, + 444 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 431, + 285, + 444 + ], + "spans": [ + { + "bbox": [ + 55, + 431, + 285, + 444 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 55, + 431, + 285, + 444 + ], + "type": "inline_equation", + "content": "\\mathcal{E}_{\\mathrm{doc}_t}" + }, + { + "bbox": [ + 55, + 431, + 285, + 444 + ], + "type": "text", + "content": " is the embedding of the " + }, + { + "bbox": [ + 55, + 431, + 285, + 444 + ], + "type": "inline_equation", + "content": "t" + }, + { + "bbox": [ + 55, + 431, + 285, + 444 + ], + "type": "text", + "content": "-th description " + }, + { + "bbox": [ + 55, + 431, + 285, + 444 + ], + "type": "inline_equation", + "content": "S_{V_t}" + }, + { + "bbox": [ + 55, + 431, + 285, + 444 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 55, + 450, + 234, + 462 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 450, + 234, + 462 + ], + "spans": [ + { + "bbox": [ + 55, + 450, + 234, + 462 + ], + "type": "text", + "content": "3.3. Document Augmentation Module" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 55, + 468, + 296, + 562 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 468, + 296, + 562 + ], + "spans": [ + { + "bbox": [ + 55, + 468, + 296, + 562 + ], + "type": "text", + "content": "For each " + }, + { + "bbox": [ + 55, + 468, + 296, + 562 + ], + "type": "inline_equation", + "content": "t' \\in \\text{topk\\_doc}" + }, + { + "bbox": [ + 55, + 468, + 296, + 562 + ], + "type": "text", + "content": ", we use the LLaVA-NeXT model [33] with different prompts " + }, + { + "bbox": [ + 55, + 468, + 296, + 562 + ], + "type": "inline_equation", + "content": "\\mathcal{AP}" + }, + { + "bbox": [ + 55, + 468, + 296, + 562 + ], + "type": "text", + "content": " (augmented prompts) to generate a detailed description " + }, + { + "bbox": [ + 55, + 468, + 296, + 562 + ], + "type": "inline_equation", + "content": "L_{V_{t'}}" + }, + { + "bbox": [ + 55, + 468, + 296, + 562 + ], + "type": "text", + "content": ", i.e., " + }, + { + "bbox": [ + 55, + 468, + 296, + 562 + ], + "type": "inline_equation", + "content": "L_{V_{t'}} = \\phi_{vlm}(\\mathcal{AP}_{t'}, V_{t'})" + }, + { + "bbox": [ + 55, + 468, + 296, + 562 + ], + "type": "text", + "content": ". The initial " + }, + { + "bbox": [ + 55, + 468, + 296, + 562 + ], + "type": "inline_equation", + "content": "\\mathcal{AP}_{t'}" + }, + { + "bbox": [ + 55, + 468, + 296, + 562 + ], + "type": "text", + "content": " is a general prompt: If there are factual errors in the question, provide a precise description of the image; if not, proceed to answer the question: " + }, + { + "bbox": [ + 55, + 468, + 296, + 562 + ], + "type": "inline_equation", + "content": "\\{Q\\}" + }, + { + "bbox": [ + 55, + 468, + 296, + 562 + ], + "type": "text", + "content": ". The updated video document " + }, + { + "bbox": [ + 55, + 468, + 296, + 562 + ], + "type": "inline_equation", + "content": "\\mathcal{AD}" + }, + { + "bbox": [ + 55, + 468, + 296, + 562 + ], + "type": "text", + "content": " becomes" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 71, + 569, + 295, + 598 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 71, + 569, + 295, + 598 + ], + "spans": [ + { + "bbox": [ + 71, + 569, + 295, + 598 + ], + "type": "interline_equation", + "content": "\\mathcal {A D} = \\left\\{\\left\\{ \\begin{array}{l l} \\left\\{t, S _ {V _ {t}}, L _ {V _ {t} ^ {\\prime}} \\right\\} & \\text {i f} t = t ^ {\\prime} \\\\ \\left\\{t, S _ {V _ {t}} \\right\\} & \\text {o t h e r w i s e} \\end{array} \\right| t = 1, \\dots , T \\right\\}. \\tag {3}", + "image_path": "7874a75e280503b53266984ea03070b6c6b885e7735cfb00966a7da75f836e0a.jpg" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 55, + 605, + 295, + 653 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 605, + 295, + 653 + ], + "spans": [ + { + "bbox": [ + 55, + 605, + 295, + 653 + ], + "type": "text", + "content": "Note that " + }, + { + "bbox": [ + 55, + 605, + 295, + 653 + ], + "type": "inline_equation", + "content": "K" + }, + { + "bbox": [ + 55, + 605, + 295, + 653 + ], + "type": "text", + "content": " is significantly smaller than the total number of frames (e.g., " + }, + { + "bbox": [ + 55, + 605, + 295, + 653 + ], + "type": "inline_equation", + "content": "K = 5" + }, + { + "bbox": [ + 55, + 605, + 295, + 653 + ], + "type": "text", + "content": " compared to " + }, + { + "bbox": [ + 55, + 605, + 295, + 653 + ], + "type": "inline_equation", + "content": "T = 90" + }, + { + "bbox": [ + 55, + 605, + 295, + 653 + ], + "type": "text", + "content": "), so the additional descriptions will not substantially increase the overall length of the document." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 55, + 660, + 247, + 673 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 660, + 247, + 673 + ], + "spans": [ + { + "bbox": [ + 55, + 660, + 247, + 673 + ], + "type": "text", + "content": "3.4. Multi-Stage Agent Interaction Loop" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 55, + 677, + 296, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 677, + 296, + 714 + ], + "spans": [ + { + "bbox": [ + 55, + 677, + 296, + 714 + ], + "type": "text", + "content": "Apart from the retrieved key frames, we further introduce a multi-stage agent interaction loop to dynamically find other potential key frames and augment them with different types" + } + ] + } + ], + "index": 13 + }, + { + "type": "image", + "bbox": [ + 317, + 72, + 548, + 247 + ], + "blocks": [ + { + "bbox": [ + 317, + 72, + 548, + 247 + ], + "lines": [ + { + "bbox": [ + 317, + 72, + 548, + 247 + ], + "spans": [ + { + "bbox": [ + 317, + 72, + 548, + 247 + ], + "type": "image", + "image_path": "bf3ad29ab9cee039b2f4a4123239ff00ee6e390e8ed9cd05704e9e3713fdadcd.jpg" + } + ] + } + ], + "index": 14, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 313, + 257, + 555, + 312 + ], + "lines": [ + { + "bbox": [ + 313, + 257, + 555, + 312 + ], + "spans": [ + { + "bbox": [ + 313, + 257, + 555, + 312 + ], + "type": "text", + "content": "Figure 2. Illustration of the multi-stage agent interaction loop and answering module. There are two agents in the multi-stage agent interaction loop: a planning agent to plan the next step and an interaction agent to dynamically find missing information and interact with the document augmentation module." + } + ] + } + ], + "index": 15, + "angle": 0, + "type": "image_caption" + } + ], + "index": 14 + }, + { + "bbox": [ + 313, + 321, + 554, + 380 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 321, + 554, + 380 + ], + "spans": [ + { + "bbox": [ + 313, + 321, + 554, + 380 + ], + "type": "text", + "content": "of information for the final answer. We employ two distinct agents: planning and interaction agents, each tailored for a specific task. Both agents utilize the same LLM, i.e., GPT-3.5, but are given different prompts and reason varying types of information. The detail is shown in Fig. 2" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 313, + 382, + 554, + 513 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 382, + 554, + 513 + ], + "spans": [ + { + "bbox": [ + 313, + 382, + 554, + 513 + ], + "type": "text", + "content": "Planning Agent. Given the question " + }, + { + "bbox": [ + 313, + 382, + 554, + 513 + ], + "type": "inline_equation", + "content": "Q" + }, + { + "bbox": [ + 313, + 382, + 554, + 513 + ], + "type": "text", + "content": " and the updated video document " + }, + { + "bbox": [ + 313, + 382, + 554, + 513 + ], + "type": "inline_equation", + "content": "\\mathcal{AD}_i" + }, + { + "bbox": [ + 313, + 382, + 554, + 513 + ], + "type": "text", + "content": " (where " + }, + { + "bbox": [ + 313, + 382, + 554, + 513 + ], + "type": "inline_equation", + "content": "i" + }, + { + "bbox": [ + 313, + 382, + 554, + 513 + ], + "type": "text", + "content": " denotes the iteration step), along with the analysis history " + }, + { + "bbox": [ + 313, + 382, + 554, + 513 + ], + "type": "inline_equation", + "content": "\\mathcal{H}" + }, + { + "bbox": [ + 313, + 382, + 554, + 513 + ], + "type": "text", + "content": " from all previous steps (initialized as empty " + }, + { + "bbox": [ + 313, + 382, + 554, + 513 + ], + "type": "inline_equation", + "content": "\\{\\}" + }, + { + "bbox": [ + 313, + 382, + 554, + 513 + ], + "type": "text", + "content": "), the planning agent first determines whether the updated video document is sufficient to generate a confident answer. If the information is sufficient, the process moves to the answering module to get the final answer. If not, the agent provides an analysis of why the updated video document " + }, + { + "bbox": [ + 313, + 382, + 554, + 513 + ], + "type": "inline_equation", + "content": "\\mathcal{AD}_i" + }, + { + "bbox": [ + 313, + 382, + 554, + 513 + ], + "type": "text", + "content": " is inadequate. The agent then updates " + }, + { + "bbox": [ + 313, + 382, + 554, + 513 + ], + "type": "inline_equation", + "content": "\\mathcal{H}_i" + }, + { + "bbox": [ + 313, + 382, + 554, + 513 + ], + "type": "text", + "content": " with its analysis and passes the updated " + }, + { + "bbox": [ + 313, + 382, + 554, + 513 + ], + "type": "inline_equation", + "content": "\\mathcal{H}_i" + }, + { + "bbox": [ + 313, + 382, + 554, + 513 + ], + "type": "text", + "content": " to the interaction agent." + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 313, + 514, + 554, + 550 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 514, + 554, + 550 + ], + "spans": [ + { + "bbox": [ + 313, + 514, + 554, + 550 + ], + "type": "text", + "content": "Interaction Agent. Given the current video document " + }, + { + "bbox": [ + 313, + 514, + 554, + 550 + ], + "type": "inline_equation", + "content": "\\mathcal{AD}_i" + }, + { + "bbox": [ + 313, + 514, + 554, + 550 + ], + "type": "text", + "content": " and the updated " + }, + { + "bbox": [ + 313, + 514, + 554, + 550 + ], + "type": "inline_equation", + "content": "\\mathcal{H}_i" + }, + { + "bbox": [ + 313, + 514, + 554, + 550 + ], + "type": "text", + "content": ", the interaction agent finds " + }, + { + "bbox": [ + 313, + 514, + 554, + 550 + ], + "type": "inline_equation", + "content": "N" + }, + { + "bbox": [ + 313, + 514, + 554, + 550 + ], + "type": "text", + "content": " frames with missing details crucial for answering " + }, + { + "bbox": [ + 313, + 514, + 554, + 550 + ], + "type": "inline_equation", + "content": "Q" + }, + { + "bbox": [ + 313, + 514, + 554, + 550 + ], + "type": "text", + "content": ":" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 329, + 559, + 553, + 572 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 329, + 559, + 553, + 572 + ], + "spans": [ + { + "bbox": [ + 329, + 559, + 553, + 572 + ], + "type": "interline_equation", + "content": "\\forall n \\in N, n \\in T, n \\notin \\text {t o p k} _ {-} \\text {d o c}, \\text {a n d} N < K. \\tag {4}", + "image_path": "32b37159b6ad79bfdd7d99db5e619b9da11c4d7c1f718a5d44a29098cdaae868.jpg" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 313, + 582, + 555, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 582, + 555, + 714 + ], + "spans": [ + { + "bbox": [ + 313, + 582, + 555, + 714 + ], + "type": "text", + "content": "In addition to identifying the missing " + }, + { + "bbox": [ + 313, + 582, + 555, + 714 + ], + "type": "inline_equation", + "content": "N" + }, + { + "bbox": [ + 313, + 582, + 555, + 714 + ], + "type": "text", + "content": " key frames, the interaction agent also determines the augmented information type for each " + }, + { + "bbox": [ + 313, + 582, + 555, + 714 + ], + "type": "inline_equation", + "content": "n \\in N" + }, + { + "bbox": [ + 313, + 582, + 555, + 714 + ], + "type": "text", + "content": ". To achieve it, we give a task-specific prompt to the interaction agent, i.e., Your task is to determine which frame needs which type of information and can answer this question accurately, reasonably, and without contradiction. The two types of information are as follows: A: Given an image, get a detailed description of the image (image caption, just like what is shown in this image?) B: Given an image, get a response to the above question (visual question answering). After obtaining " + }, + { + "bbox": [ + 313, + 582, + 555, + 714 + ], + "type": "inline_equation", + "content": "N" + } + ] + } + ], + "index": 20 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "text", + "content": "18939" + } + ] + } + ], + "index": 21 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 3 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 77, + 70, + 276, + 218 + ], + "blocks": [ + { + "bbox": [ + 77, + 70, + 276, + 218 + ], + "lines": [ + { + "bbox": [ + 77, + 70, + 276, + 218 + ], + "spans": [ + { + "bbox": [ + 77, + 70, + 276, + 218 + ], + "type": "table", + "html": "
Method(M)LLMSubsetFullset
MoReVQA [37]PaLM-2-51.7
Vamos [60]GPT-451.248.3
ProViQ [9]GPT-3.557.1-
LLoVi [74]GPT-3.557.650.3
IG-VLM [25]GPT-4V59.8-
VideoAgent [61]GPT-460.254.1
MVU [47]Mistral-13B60.337.6
LLoVi [74]GPT-461.2-
VideoAgent [14]GPT-462.860.2
DrVideo (ours)GPT-3.562.6-
DrVideo (ours)GPT-466.461.0
", + "image_path": "c9f019dbf2a337df8f50ac7d0555789fedf56a26c8761b78c5731cf0203c3875.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "table_body" + } + ], + "index": 0 + }, + { + "bbox": [ + 55, + 258, + 296, + 307 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 258, + 296, + 307 + ], + "spans": [ + { + "bbox": [ + 55, + 258, + 296, + 307 + ], + "type": "text", + "content": "frames along with their required information type, the agent interacts with the document augmentation module to enrich the information of these frames to update the current video document for the next step." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 55, + 315, + 167, + 327 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 315, + 167, + 327 + ], + "spans": [ + { + "bbox": [ + 55, + 315, + 167, + 327 + ], + "type": "text", + "content": "3.5. Answering Module" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 55, + 331, + 296, + 429 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 331, + 296, + 429 + ], + "spans": [ + { + "bbox": [ + 55, + 331, + 296, + 429 + ], + "type": "text", + "content": "Given the final video document " + }, + { + "bbox": [ + 55, + 331, + 296, + 429 + ], + "type": "inline_equation", + "content": "\\mathcal{AD}_{final}" + }, + { + "bbox": [ + 55, + 331, + 296, + 429 + ], + "type": "text", + "content": ", we employ another agent to provide a prediction using a zero-shot chain-of-thought (CoT) approach [26, 66]. The answering module outputs the corresponding answer, the confidence score, and the reasoning behind the answer. Besides improving prediction accuracy, the CoT approach allows us to trace inference steps, ensuring transparency and explainability in the decision-making process." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 55, + 439, + 137, + 453 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 439, + 137, + 453 + ], + "spans": [ + { + "bbox": [ + 55, + 439, + 137, + 453 + ], + "type": "text", + "content": "4. Experiments" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 55, + 460, + 295, + 496 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 460, + 295, + 496 + ], + "spans": [ + { + "bbox": [ + 55, + 460, + 295, + 496 + ], + "type": "text", + "content": "In this section, we first introduce the datasets and implementation details and then present the results and ablations of our proposed DrVideo." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 55, + 504, + 178, + 516 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 504, + 178, + 516 + ], + "spans": [ + { + "bbox": [ + 55, + 504, + 178, + 516 + ], + "type": "text", + "content": "4.1. Datasets and Metrics" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 55, + 521, + 295, + 557 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 521, + 295, + 557 + ], + "spans": [ + { + "bbox": [ + 55, + 521, + 295, + 557 + ], + "type": "text", + "content": "In our experiments, we evaluate our DrVideo using three well-established datasets, emphasizing its zero-shot long video understanding capabilities." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 55, + 558, + 295, + 616 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 558, + 295, + 616 + ], + "spans": [ + { + "bbox": [ + 55, + 558, + 295, + 616 + ], + "type": "text", + "content": "EgoSchema. EgoSchema dataset [36] consists of 5000 multiple-choice questions sourced from 5000 three-minute egocentric videos. This dataset only has a subset of 500 questions with publicly accessible labels, while the full set is evaluated on the leaderboard." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 55, + 617, + 295, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 617, + 295, + 714 + ], + "spans": [ + { + "bbox": [ + 55, + 617, + 295, + 714 + ], + "type": "text", + "content": "MovieChat-1K. MovieChat-1K [49] is a longer video understanding benchmark, which contains 1000 videos from movies and TV shows, each video lasting approximately 10 minutes. This dataset has two modes: global mode and breakpoint mode. The global mode refers to analyzing the entire video to understand its overall content and context, while the breakpoint mode focuses on analyzing specific frames or scenes." + } + ] + } + ], + "index": 10 + }, + { + "type": "table", + "bbox": [ + 317, + 70, + 553, + 234 + ], + "blocks": [ + { + "bbox": [ + 55, + 226, + 295, + 249 + ], + "lines": [ + { + "bbox": [ + 55, + 226, + 295, + 249 + ], + "spans": [ + { + "bbox": [ + 55, + 226, + 295, + 249 + ], + "type": "text", + "content": "Table 1. Results on EgoSchema compared to existing LLM-based state-of-the-art methods." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 317, + 70, + 553, + 234 + ], + "lines": [ + { + "bbox": [ + 317, + 70, + 553, + 234 + ], + "spans": [ + { + "bbox": [ + 317, + 70, + 553, + 234 + ], + "type": "table", + "html": "
MethodGlobalBreakpoint
Acc.ScoreAcc.Score
Open-Source MLLM
Video LLaMA [75]51.72.7239.12.11
Video Chat [30]57.83.0846.12.32
Video-ChatGPT [35]47.62.8948.02.43
MovieChat [49]68.33.1548.32.46
MovieChat+ [50]71.23.5149.62.62
Based on Open-source Captioners and Proprietary LLMs
MM-VID [32]58.62.8610.40.56
LLoVi [74]58.32.8717.81.03
\\( VideoAgent^* \\) [61]65.4*3.12*31.6*2.05*
Sullam Jeoung, et al [23]84.8---
DrVideo (ours)93.14.4156.42.75
", + "image_path": "0e10d8cf3c024e274e533aa8c9baf1a45884cdfab3080503a3955f07d2d32164.jpg" + } + ] + } + ], + "index": 11, + "angle": 0, + "type": "table_body" + } + ], + "index": 11 + }, + { + "bbox": [ + 313, + 242, + 553, + 276 + ], + "lines": [ + { + "bbox": [ + 313, + 242, + 553, + 276 + ], + "spans": [ + { + "bbox": [ + 313, + 242, + 553, + 276 + ], + "type": "text", + "content": "Table 2. Performance comparison on the MoiveChat-1K [49] benchmark against state-of-the-art methods. * represents re-implemented results, implementation details refer to appendix." + } + ] + } + ], + "index": 12, + "angle": 0, + "type": "text" + }, + { + "bbox": [ + 313, + 282, + 555, + 342 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 282, + 555, + 342 + ], + "spans": [ + { + "bbox": [ + 313, + 282, + 555, + 342 + ], + "type": "text", + "content": "Video-MME. Video-MME [15] is a recently proposed benchmark for comprehensive video analysis evaluation. We evaluate DrVideo on the \"long-term videos\" split of this dataset (long split), where video lengths vary from 30 to 60 minutes, with an average duration of 44 minutes." + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 313, + 342, + 554, + 426 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 342, + 554, + 426 + ], + "spans": [ + { + "bbox": [ + 313, + 342, + 554, + 426 + ], + "type": "text", + "content": "Since EgoSchema and Video-MME are multi-choice tasks, we use accuracy as the evaluation metric for both datasets. MovieChat-1K focuses on open-ended questions, we use GPT-assisted evaluation to assess both the accuracy (true/false) and quality (score 0-5) of the models. We select Gemini-Pro [54] as the evaluation assistant and use the same prompt [38] to conduct a fair comparison." + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 313, + 433, + 447, + 445 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 433, + 447, + 445 + ], + "spans": [ + { + "bbox": [ + 313, + 433, + 447, + 445 + ], + "type": "text", + "content": "4.2. Implementation Details" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 313, + 450, + 555, + 666 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 450, + 555, + 666 + ], + "spans": [ + { + "bbox": [ + 313, + 450, + 555, + 666 + ], + "type": "text", + "content": "For EgoSchema [36] dataset, we choose LaViLa [78] as the captioning model to convert videos into documents. Note that the training data of our LaViLa model does not include EgoSchema videos, which is the same as LLoVi [74]. For MovieChat-1K and Video-MME benchmarks, LLaVA-NeXT [33] is used as our captioning model to generate brief descriptions for each frame. We preprocess videos by simply sampling them at 0.5 FPS for EgoSchema [36] and MovieChat-1K [49] and 0.2 FPS for Video-MME [15]. LLaVA-NeXT [33] is chosen to augment key frames in both datasets via different prompts. In the comparison experiments, GPT-4 [1], i.e., gpt-4-1106-preview, is used as the agent to evaluate the performance of our DrVideo on EgoSchema and MovieChat-1K. DeepSeek [19], i.e., DeepSeek V2.5, is used as the agent to evaluate the performance on Video-MME. For the ablation study, we use GPT-3.5 [40], i.e., gpt-3.5-turbo-1106, to evaluate the effectiveness of our DrVideo due to the API cost." + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 313, + 672, + 399, + 684 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 672, + 399, + 684 + ], + "spans": [ + { + "bbox": [ + 313, + 672, + 399, + 684 + ], + "type": "text", + "content": "4.3. Main Results" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 313, + 689, + 553, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 689, + 553, + 714 + ], + "spans": [ + { + "bbox": [ + 313, + 689, + 553, + 714 + ], + "type": "text", + "content": "Comparative results on EgoSchema. Table 1 compares our DrVideo with existing LLM-based state-of-the" + } + ] + } + ], + "index": 18 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "text", + "content": "18940" + } + ] + } + ], + "index": 19 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 4 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 65, + 70, + 288, + 293 + ], + "blocks": [ + { + "bbox": [ + 65, + 70, + 288, + 293 + ], + "lines": [ + { + "bbox": [ + 65, + 70, + 288, + 293 + ], + "spans": [ + { + "bbox": [ + 65, + 70, + 288, + 293 + ], + "type": "table", + "html": "
MethodFramesw/o Subsw Subs
Proprietary MLLM
Claude 3.5 Sonnet [2]2051.254.7
GPT-4V [42]1053.556.9
GPT-4o mini [44]25058.663.4
Gemini 1.5 Flash [55]0.5 FPS61.168.8
GPT-4o [44]38465.372.1
Gemini 1.5 Pro [55]0.5 FPS67.477.4
Open-Source MLLM
LongVA [76]12846.247.6
VITA-8x7B [16]3248.650.9
LLaVA-OneVision-72B [28]3260.062.4
Qwen2-VL-72B [3]76862.274.3
Based on Open-source Captioners and Proprietary LLMs
\\( VideoAgent^* \\)[61]7.640.2*44.4*
\\( LLoVi^* \\)[74]0.2 FPS45.4*67.7*
DrVideo (ours)0.2 FPS51.771.7
-only subs0.2 FPS-68.5
", + "image_path": "b13dff7a51ebb95c2eb807ffc938bf49cb469b51d1c3d4677848f13e682b6d70.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "table_body" + } + ], + "index": 0 + }, + { + "bbox": [ + 54, + 389, + 295, + 569 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 54, + 389, + 295, + 569 + ], + "spans": [ + { + "bbox": [ + 54, + 389, + 295, + 569 + ], + "type": "text", + "content": "art methods [9, 14, 25, 37, 47, 60, 61, 74] on EgoSchema benchmark [36]. Specifically, compared with LLoVi [74] and VideoAgent [61] that leverage the same visual captioner (LaViLa [78]) and LLM (GPT-4), DrVideo significantly outperforms these methods by " + }, + { + "bbox": [ + 54, + 389, + 295, + 569 + ], + "type": "inline_equation", + "content": "5.2\\%" + }, + { + "bbox": [ + 54, + 389, + 295, + 569 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 54, + 389, + 295, + 569 + ], + "type": "inline_equation", + "content": "6.2\\%" + }, + { + "bbox": [ + 54, + 389, + 295, + 569 + ], + "type": "text", + "content": " respectively. Compared with VideoAgent [14] which uses videospecific models (ViCLIP from InternVid [63] and VideoLLaVA [31]), DrVideo still performs better and achieves " + }, + { + "bbox": [ + 54, + 389, + 295, + 569 + ], + "type": "inline_equation", + "content": "3.6\\%" + }, + { + "bbox": [ + 54, + 389, + 295, + 569 + ], + "type": "text", + "content": " higher accuracy on the subset evaluation. Moreover, DrVideo significantly outperforms IG-VLM [25] by " + }, + { + "bbox": [ + 54, + 389, + 295, + 569 + ], + "type": "inline_equation", + "content": "6.6\\%" + }, + { + "bbox": [ + 54, + 389, + 295, + 569 + ], + "type": "text", + "content": ", which uses strong MLLM, i.e., GPT-4V. These results confirm our idea of converting the long-video understanding task into a long-document understanding task is feasible and our DrVideo can find the question-related key frames more accurately." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 55, + 570, + 295, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 570, + 295, + 714 + ], + "spans": [ + { + "bbox": [ + 55, + 570, + 295, + 714 + ], + "type": "text", + "content": "Comparative results on longer video benchmarks. To further emphasize the advantage of our method on longer videos, we evaluate our DrVideo on the other two video benchmarks, i.e., MoiveChat-1K (10 minutes) [49], and the long split of Video-MME (average 44 minutes) [15]. Table 2 presents the comparative results of our DrVideo and the current state-of-the-art methods [30, 38, 49, 74, 75] on the MovieChat-1K benchmark. Table 3 shows the comparative results of our DrVideo on the Video-MME benchmark, with comparisons against three types of methods: proprietary MLLMs [2, 42, 44, 55], open-source MLLMs [3, 16, 28, 76], and methods based on open-source caption" + } + ] + } + ], + "index": 3 + }, + { + "type": "table", + "bbox": [ + 335, + 70, + 535, + 152 + ], + "blocks": [ + { + "bbox": [ + 55, + 300, + 296, + 378 + ], + "lines": [ + { + "bbox": [ + 55, + 300, + 296, + 378 + ], + "spans": [ + { + "bbox": [ + 55, + 300, + 296, + 378 + ], + "type": "text", + "content": "Table 3. Video-MME long split results. " + }, + { + "bbox": [ + 55, + 300, + 296, + 378 + ], + "type": "inline_equation", + "content": "w/o" + }, + { + "bbox": [ + 55, + 300, + 296, + 378 + ], + "type": "text", + "content": " subs represents the model without any subtitles. " + }, + { + "bbox": [ + 55, + 300, + 296, + 378 + ], + "type": "inline_equation", + "content": "w" + }, + { + "bbox": [ + 55, + 300, + 296, + 378 + ], + "type": "text", + "content": " subs represents that the subtitles corresponding to the sampled frames are both fed into the LLM to evaluate the performance. Our DrVideo outperforms a strong proprietary MLLM (Claude 3.5 Sonnet) and some open-source MLLMs (e.g., VITA-8x7B). * represents re-implemented results, implementation details refer to appendix." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 335, + 70, + 535, + 152 + ], + "lines": [ + { + "bbox": [ + 335, + 70, + 535, + 152 + ], + "spans": [ + { + "bbox": [ + 335, + 70, + 535, + 152 + ], + "type": "table", + "html": "
RMMSAILCoTAcc.(%)
✓✓✓62.6
✓✓X62.2
✓X✓60.6
X✓✓59.4
XX✓57.4
", + "image_path": "31cd407c979a5d19cec95d3c01e57ec5d6d7c86413205a6af5f531545fd940f8.jpg" + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "table_body" + } + ], + "index": 4 + }, + { + "type": "table", + "bbox": [ + 324, + 202, + 442, + 262 + ], + "blocks": [ + { + "bbox": [ + 313, + 160, + 555, + 194 + ], + "lines": [ + { + "bbox": [ + 313, + 160, + 555, + 194 + ], + "spans": [ + { + "bbox": [ + 313, + 160, + 555, + 194 + ], + "type": "text", + "content": "Table 4. Ablation results on different combinations of the retrieval module (RM), the multi-stage agent interaction loop (MSAIL), and the CoT on EgoSchema." + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 324, + 202, + 442, + 262 + ], + "lines": [ + { + "bbox": [ + 324, + 202, + 442, + 262 + ], + "spans": [ + { + "bbox": [ + 324, + 202, + 442, + 262 + ], + "type": "table", + "html": "
VQACaptionAcc. (%)
✓✓62.6
×✓60.4
✓×61.8
", + "image_path": "204b50764ed18a1eb37988d7f528c0366c19eddebee202289d8a53ad1927970c.jpg" + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "table_body" + }, + { + "bbox": [ + 324, + 264, + 418, + 281 + ], + "lines": [ + { + "bbox": [ + 324, + 264, + 418, + 281 + ], + "spans": [ + { + "bbox": [ + 324, + 264, + 418, + 281 + ], + "type": "text", + "content": "(a) Performance with different types of information" + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "table_footnote" + } + ], + "index": 6 + }, + { + "type": "table", + "bbox": [ + 446, + 202, + 529, + 262 + ], + "blocks": [ + { + "bbox": [ + 446, + 202, + 529, + 262 + ], + "lines": [ + { + "bbox": [ + 446, + 202, + 529, + 262 + ], + "spans": [ + { + "bbox": [ + 446, + 202, + 529, + 262 + ], + "type": "table", + "html": "
Top-KAcc. (%)
562.6
1061.4
2060.6
", + "image_path": "c9e13824dd2e56e1f7c5fcb434e08100a539d6a44015002d0e9acb66289f0942.jpg" + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "table_body" + }, + { + "bbox": [ + 430, + 264, + 544, + 281 + ], + "lines": [ + { + "bbox": [ + 430, + 264, + 544, + 281 + ], + "spans": [ + { + "bbox": [ + 430, + 264, + 544, + 281 + ], + "type": "text", + "content": "(b) Performance with different Top- " + }, + { + "bbox": [ + 430, + 264, + 544, + 281 + ], + "type": "inline_equation", + "content": "K" + }, + { + "bbox": [ + 430, + 264, + 544, + 281 + ], + "type": "text", + "content": " frames" + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "table_footnote" + } + ], + "index": 8 + }, + { + "bbox": [ + 318, + 290, + 549, + 301 + ], + "lines": [ + { + "bbox": [ + 318, + 290, + 549, + 301 + ], + "spans": [ + { + "bbox": [ + 318, + 290, + 549, + 301 + ], + "type": "text", + "content": "Table 5. Ablation results with different settings on EgoSchema." + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "text" + }, + { + "bbox": [ + 313, + 315, + 457, + 327 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 315, + 457, + 327 + ], + "spans": [ + { + "bbox": [ + 313, + 315, + 457, + 327 + ], + "type": "text", + "content": "ers and proprietary LLMs [61, 74]." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 313, + 328, + 555, + 544 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 328, + 555, + 544 + ], + "spans": [ + { + "bbox": [ + 313, + 328, + 555, + 544 + ], + "type": "text", + "content": "Compared to methods like LLoVi [74] and VideoAgent [61] that are based on open-source captioners (e.g., LLaVA-NeXT [33]) and proprietary LLMs (e.g., GPT-4 [1]), DrVideo achieves a " + }, + { + "bbox": [ + 313, + 328, + 555, + 544 + ], + "type": "inline_equation", + "content": "8.3\\%" + }, + { + "bbox": [ + 313, + 328, + 555, + 544 + ], + "type": "text", + "content": " improvement in global mode and a " + }, + { + "bbox": [ + 313, + 328, + 555, + 544 + ], + "type": "inline_equation", + "content": "24.8\\%" + }, + { + "bbox": [ + 313, + 328, + 555, + 544 + ], + "type": "text", + "content": " improvement in breakpoint mode on the MovieChat-1K benchmark in Table 2. It also achieves at least " + }, + { + "bbox": [ + 313, + 328, + 555, + 544 + ], + "type": "inline_equation", + "content": "6.3\\%" + }, + { + "bbox": [ + 313, + 328, + 555, + 544 + ], + "type": "text", + "content": " improvement under the w/o subtitle setting on the Video-MME benchmark in Table 3. This indicates that DrVideo can accurately identify question-related frames and avoid critical information loss through augmentation, highlighting the potential of document retrieval methods for processing longer videos. Compared to proprietary MLLMs and open-source MLLMs, DrVideo outperforms Claude 3.5 Sonnet [2] by " + }, + { + "bbox": [ + 313, + 328, + 555, + 544 + ], + "type": "inline_equation", + "content": "0.5\\%" + }, + { + "bbox": [ + 313, + 328, + 555, + 544 + ], + "type": "text", + "content": " and some open-source MLLMs (e.g., VITA-8x7B [16]) under the w/o subtitle setting on Video-MME benchmark. It also achieves a " + }, + { + "bbox": [ + 313, + 328, + 555, + 544 + ], + "type": "inline_equation", + "content": "21.9\\%" + }, + { + "bbox": [ + 313, + 328, + 555, + 544 + ], + "type": "text", + "content": " improvement in global mode and a " + }, + { + "bbox": [ + 313, + 328, + 555, + 544 + ], + "type": "inline_equation", + "content": "6.8\\%" + }, + { + "bbox": [ + 313, + 328, + 555, + 544 + ], + "type": "text", + "content": " improvement in breakpoint mode on MovieChat-1K benchmark." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 313, + 545, + 556, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 545, + 556, + 715 + ], + "spans": [ + { + "bbox": [ + 313, + 545, + 556, + 715 + ], + "type": "text", + "content": "Moreover, combining the subtitle on the Video-MME benchmark, DrVideo achieves " + }, + { + "bbox": [ + 313, + 545, + 556, + 715 + ], + "type": "inline_equation", + "content": "71.7\\%" + }, + { + "bbox": [ + 313, + 545, + 556, + 715 + ], + "type": "text", + "content": " accuracy and outperforms many heavily-engineered large-scale proprietary models (e.g., Gemini 1.5 Flash [55], GPT-4o mini [44], and GPT-4V [42]). To better understand this significant improvement, we conduct an experiment where only the subtitle, question and options are input into the LLM (i.e., DeepSeek [19]) to predict the answer and the accuracy is " + }, + { + "bbox": [ + 313, + 545, + 556, + 715 + ], + "type": "inline_equation", + "content": "68.5\\%" + }, + { + "bbox": [ + 313, + 545, + 556, + 715 + ], + "type": "text", + "content": ". This suggests: (i) subtitles contain extensive question-related information, playing a crucial role in understanding long videos; (ii) compared with LLoVi and VideoAgent, DrVideo enhances visual information via document retrieval and augmentation, effectively integrating subtitles and resulting in a " + }, + { + "bbox": [ + 313, + 545, + 556, + 715 + ], + "type": "inline_equation", + "content": "3.2\\%" + }, + { + "bbox": [ + 313, + 545, + 556, + 715 + ], + "type": "text", + "content": " improvement." + } + ] + } + ], + "index": 13 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "text", + "content": "18941" + } + ] + } + ], + "index": 14 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 5 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 69, + 76, + 284, + 200 + ], + "blocks": [ + { + "bbox": [ + 69, + 76, + 284, + 200 + ], + "lines": [ + { + "bbox": [ + 69, + 76, + 284, + 200 + ], + "spans": [ + { + "bbox": [ + 69, + 76, + 284, + 200 + ], + "type": "image", + "image_path": "4dfa8b54a6ab788efd253eb7da839e72d5dd4331d464ab23df0e11ba1fdb0505.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 55, + 215, + 295, + 251 + ], + "lines": [ + { + "bbox": [ + 55, + 215, + 295, + 251 + ], + "spans": [ + { + "bbox": [ + 55, + 215, + 295, + 251 + ], + "type": "text", + "content": "Figure 3. Performance of different rounds with DrVideo and VideoAgent [61] on EgoSchema. To align with VideoAgent, GPT-4, i.e., gpt-4-turbo-1106-preview, is used as the LLM agents." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "type": "table", + "bbox": [ + 70, + 260, + 282, + 318 + ], + "blocks": [ + { + "bbox": [ + 70, + 260, + 282, + 318 + ], + "lines": [ + { + "bbox": [ + 70, + 260, + 282, + 318 + ], + "spans": [ + { + "bbox": [ + 70, + 260, + 282, + 318 + ], + "type": "table", + "html": "
Captioning ModelTypeAcc.(%)
LaViLa [78]Clip-based62.6
LLaVA-NeXT [33]Frame-based61.2
BLIP-2 [29]Frame-based59.6
", + "image_path": "cb0e9c83de53f3fa0ecd17d823af36c9e031688171199099d621325aa6d95e80.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "table_body" + } + ], + "index": 2 + }, + { + "bbox": [ + 55, + 347, + 156, + 358 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 347, + 156, + 358 + ], + "spans": [ + { + "bbox": [ + 55, + 347, + 156, + 358 + ], + "type": "text", + "content": "4.4. Ablation Studies" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 55, + 365, + 295, + 567 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 365, + 295, + 567 + ], + "spans": [ + { + "bbox": [ + 55, + 365, + 295, + 567 + ], + "type": "text", + "content": "Unless otherwise specified, we use GPT-3.5 as the default setting in the below experiments on EgoSchema subset [36]. Effects of different components of DrVideo. We conduct experiments with different combinations of the individual components in DrVideo, including the retrieval module, the multi-stage agent interaction loop, and the CoT used in the answering module. The results are presented in Table 4. We can see that: (i) without the retrieval module and the agent loop, the performance of our DrVideo is similar to LLoVi [74]; (ii) By adding the retrieval module and then the agent loop, the performance of DrVideo improves from " + }, + { + "bbox": [ + 55, + 365, + 295, + 567 + ], + "type": "inline_equation", + "content": "57.4\\%" + }, + { + "bbox": [ + 55, + 365, + 295, + 567 + ], + "type": "text", + "content": " to " + }, + { + "bbox": [ + 55, + 365, + 295, + 567 + ], + "type": "inline_equation", + "content": "60.6\\%" + }, + { + "bbox": [ + 55, + 365, + 295, + 567 + ], + "type": "text", + "content": " and then to " + }, + { + "bbox": [ + 55, + 365, + 295, + 567 + ], + "type": "inline_equation", + "content": "62.6\\%" + }, + { + "bbox": [ + 55, + 365, + 295, + 567 + ], + "type": "text", + "content": ". This demonstrates that the retrieval module can identify key frames via measuring semantic similarity, while the agent loop can locate additional key frames through contextual inference; (iii) the additional CoT contributes to the accuracy improvement from " + }, + { + "bbox": [ + 55, + 365, + 295, + 567 + ], + "type": "inline_equation", + "content": "62.2\\%" + }, + { + "bbox": [ + 55, + 365, + 295, + 567 + ], + "type": "text", + "content": " to " + }, + { + "bbox": [ + 55, + 365, + 295, + 567 + ], + "type": "inline_equation", + "content": "62.6\\%" + }, + { + "bbox": [ + 55, + 365, + 295, + 567 + ], + "type": "text", + "content": ", demonstrating its effectiveness." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 55, + 569, + 295, + 700 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 569, + 295, + 700 + ], + "spans": [ + { + "bbox": [ + 55, + 569, + 295, + 700 + ], + "type": "text", + "content": "Effects of different types of augment information. To better enhance the required information for each potential key frame, we define two types of augment information, namely VQA (question-related information) and caption (general detailed information). Table 5a presents the performance of DrVideo with different types of information. We can see that when only one type of information (VQA or caption) is used for augmentation, the model's performance declines. This suggests the necessity of adaptively augmenting different types of information for each potential key frame." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 55, + 701, + 295, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 701, + 295, + 713 + ], + "spans": [ + { + "bbox": [ + 55, + 701, + 295, + 713 + ], + "type": "text", + "content": "Effects of different initial top " + }, + { + "bbox": [ + 55, + 701, + 295, + 713 + ], + "type": "inline_equation", + "content": "K" + }, + { + "bbox": [ + 55, + 701, + 295, + 713 + ], + "type": "text", + "content": " key frames. By default," + } + ] + } + ], + "index": 7 + }, + { + "type": "table", + "bbox": [ + 329, + 71, + 541, + 127 + ], + "blocks": [ + { + "bbox": [ + 71, + 326, + 279, + 337 + ], + "lines": [ + { + "bbox": [ + 71, + 326, + 279, + 337 + ], + "spans": [ + { + "bbox": [ + 71, + 326, + 279, + 337 + ], + "type": "text", + "content": "Table 6. Performance of different VLMs on EgoSchema." + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 329, + 71, + 541, + 127 + ], + "lines": [ + { + "bbox": [ + 329, + 71, + 541, + 127 + ], + "spans": [ + { + "bbox": [ + 329, + 71, + 541, + 127 + ], + "type": "table", + "html": "
LLMsTypeSizeAcc.(%)
Mistral-8x7B [24]Open-Source8x7B47.6
DeepSeek [19]ProprietaryN/A61.2
GPT-3.5 [40]ProprietaryN/A62.6
GPT-4 [1]ProprietaryN/A66.4
", + "image_path": "71b531860e2c2b075d2f8fb8f9add18afeafb254fd32575f4c04e5b58c036ef3.jpg" + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "table_body" + } + ], + "index": 8 + }, + { + "bbox": [ + 331, + 135, + 537, + 147 + ], + "lines": [ + { + "bbox": [ + 331, + 135, + 537, + 147 + ], + "spans": [ + { + "bbox": [ + 331, + 135, + 537, + 147 + ], + "type": "text", + "content": "Table 7. Performance of different LLMs on EgoSchema." + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "text" + }, + { + "bbox": [ + 313, + 159, + 553, + 255 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 159, + 553, + 255 + ], + "spans": [ + { + "bbox": [ + 313, + 159, + 553, + 255 + ], + "type": "text", + "content": "the number of the initial key frames is set as " + }, + { + "bbox": [ + 313, + 159, + 553, + 255 + ], + "type": "inline_equation", + "content": "K = 5" + }, + { + "bbox": [ + 313, + 159, + 553, + 255 + ], + "type": "text", + "content": ". Intuitively, more frames should lead to better performance at the cost of increasing the complexity. However, our ablation in Table 5b shows an opposite trend, i.e., more key frames resulting in worse performance. This suggests that enhancing more frames is not necessarily beneficial, as it can introduce a large amount of noisy and irrelevant information that interferes with the LLM's judgment." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 313, + 256, + 554, + 435 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 256, + 554, + 435 + ], + "spans": [ + { + "bbox": [ + 313, + 256, + 554, + 435 + ], + "type": "text", + "content": "Effects of iterative rounds. We also examined the impact of the iterative rounds " + }, + { + "bbox": [ + 313, + 256, + 554, + 435 + ], + "type": "inline_equation", + "content": "I" + }, + { + "bbox": [ + 313, + 256, + 554, + 435 + ], + "type": "text", + "content": " on model performance. Fig. 3 shows the performance of different iterative rounds with DrVideo and VideoAgent [61]. We have the following observations. (i) DrVideo consistently outperforms VideoAgent, reaching the peak performance at " + }, + { + "bbox": [ + 313, + 256, + 554, + 435 + ], + "type": "inline_equation", + "content": "I = 2" + }, + { + "bbox": [ + 313, + 256, + 554, + 435 + ], + "type": "text", + "content": ", while VideoAgent peaks at " + }, + { + "bbox": [ + 313, + 256, + 554, + 435 + ], + "type": "inline_equation", + "content": "I = 3" + }, + { + "bbox": [ + 313, + 256, + 554, + 435 + ], + "type": "text", + "content": ". This indicates that DrVideo can locate the information needed to answer questions in fewer interactions and with higher quality, underscoring the superiority of our document retrieval-based framework. (ii) When the number of interactions exceeds two, DrVideo's performance declines. This is likely due to more noisy and irrelevant information being introduced into the LLM with the increase of the iterations. This further confirms that excessive information is not necessarily beneficial." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 313, + 437, + 554, + 592 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 437, + 554, + 592 + ], + "spans": [ + { + "bbox": [ + 313, + 437, + 554, + 592 + ], + "type": "text", + "content": "Effects of different foundation models. DrVideo incorporates two types of foundational models — large language model (LLM), and visual language model (VLM), where VLM is used as the frame captioner while LLM is used as the planning agent, the interaction agent, and the answering agent. To assess the impact of the captions produced by different VLMs, we examine three state-of-the-art VLMs: frame-based BLIP-2 [29], LLaVA-NeXT [33], and clip-based LaViLa [78], as presented in Table 6. We observe that LaViLa achieves the best performance, while BLIP-2 is the worst. It indicates the superiority of the clip-based model which can capture motion information in videos and is more suited for video understanding." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 313, + 594, + 554, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 594, + 554, + 713 + ], + "spans": [ + { + "bbox": [ + 313, + 594, + 554, + 713 + ], + "type": "text", + "content": "To evaluate the effectiveness of DrVideo with different LLMs, Table 7 shows the results of DrVideo using open-source LLM (Mistral-8x7B [24]) or Proprietary LLM (DeepSeek [19], GPT-3.5 [40], or GPT-4 [1]) as the planning, interaction and answering agents. From the results, we can see that: (i) different LLMs in DrVideo exhibit significant performance gaps and the proprietary LLM significantly outperforms the open-source LLM, indicating DrVideo's dependency on the choice of LLM; (ii) as the LLM's capability increases, the performance of DrVideo" + } + ] + } + ], + "index": 13 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "text", + "content": "18942" + } + ] + } + ], + "index": 14 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 6 + }, + { + "para_blocks": [ + { + "bbox": [ + 59, + 73, + 541, + 92 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 59, + 73, + 541, + 92 + ], + "spans": [ + { + "bbox": [ + 59, + 73, + 541, + 92 + ], + "type": "text", + "content": "Question: By analyzing c's activities in this video, determine their overall focus or intention, and describe the progression of tasks leading to its fulfillment." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 61, + 93, + 495, + 139 + ], + "type": "list", + "angle": 0, + "index": 6, + "blocks": [ + { + "bbox": [ + 61, + 93, + 474, + 102 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 61, + 93, + 474, + 102 + ], + "spans": [ + { + "bbox": [ + 61, + 93, + 474, + 102 + ], + "type": "text", + "content": "A. C's overall focus or intention in this video is to set up their work station and start working on their computer." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 62, + 102, + 477, + 111 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 102, + 477, + 111 + ], + "spans": [ + { + "bbox": [ + 62, + 102, + 477, + 111 + ], + "type": "text", + "content": "B. The central objective in this video, c's overall focus or intention, is to engage in conversation with the person." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 62, + 112, + 307, + 120 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 112, + 307, + 120 + ], + "spans": [ + { + "bbox": [ + 62, + 112, + 307, + 120 + ], + "type": "text", + "content": "C. C's overall focus or intention in this video is to eat the wrapper." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 63, + 121, + 430, + 129 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 63, + 121, + 430, + 129 + ], + "spans": [ + { + "bbox": [ + 63, + 121, + 430, + 129 + ], + "type": "text", + "content": "D. C's overall focus or primary intention in this specific video content is to effectively play the guitar" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 63, + 130, + 495, + 139 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 63, + 130, + 495, + 139 + ], + "spans": [ + { + "bbox": [ + 63, + 130, + 495, + 139 + ], + "type": "text", + "content": "E. The primary objective and intention in this particular c's video is to demonstrate how to effectively clean the cloth." + } + ] + } + ], + "index": 5 + } + ], + "sub_type": "text" + }, + { + "type": "image", + "bbox": [ + 56, + 140, + 563, + 435 + ], + "blocks": [ + { + "bbox": [ + 56, + 140, + 563, + 435 + ], + "lines": [ + { + "bbox": [ + 56, + 140, + 563, + 435 + ], + "spans": [ + { + "bbox": [ + 56, + 140, + 563, + 435 + ], + "type": "image", + "image_path": "50c85371c322e3c19aa59852da71c6437cf3c180a65671e89e2cf99d44417dea.jpg" + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 72, + 442, + 536, + 453 + ], + "lines": [ + { + "bbox": [ + 72, + 442, + 536, + 453 + ], + "spans": [ + { + "bbox": [ + 72, + 442, + 536, + 453 + ], + "type": "text", + "content": "Figure 4. Case study on an instance from EgoSchema. DrVideo accurately identifies key frames and chooses the correct answer." + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "image_caption" + } + ], + "index": 7 + }, + { + "bbox": [ + 55, + 469, + 296, + 565 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 469, + 296, + 565 + ], + "spans": [ + { + "bbox": [ + 55, + 469, + 296, + 565 + ], + "type": "text", + "content": "improves, demonstrating the great potential of DrVideo. It's worth noting that our main contribution is the introduction of a framework for translating the long video understanding into a document retrieval and understanding task to effectively leverage the power of large language models, rather than the employment of any specific models. Our DrVideo can be further improved by combining with better LLM and VLM models (e.g., tarsier [59])." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 55, + 570, + 296, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 570, + 296, + 713 + ], + "spans": [ + { + "bbox": [ + 55, + 570, + 296, + 713 + ], + "type": "text", + "content": "Case study. Here we give a detailed example from EgoSchema [36] in Fig. 4 to qualitatively demonstrate the capability of our DrVideo. Specifically, the question is to determine the overall focus or intention and describe the progression of tasks leading to its fulfillment. DrVideo first converts the raw video into an initial video document, which is then processed by the frame retrieval module to obtain key frames (frames 15, 46, 61, 84, and 86) and is augmented with detailed key frame descriptions. The updated video document is then fed into the planning agent to assess if sufficient information is available for answering, i.e., the provided information is sufficient to infer character C's" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 313, + 469, + 555, + 553 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 469, + 555, + 553 + ], + "spans": [ + { + "bbox": [ + 313, + 469, + 555, + 553 + ], + "type": "text", + "content": "main focus or intent. This updated document, along with the rationale, is input to the interaction agent to identify the additional key frames (frames 4, 5, 6, 14, and 19 here). After further augmentation, the updated document is sent back to the planning agent, for which the agent confirms its sufficiency. Finally, the answering module takes in the final video document and predicts the correct answer." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 313, + 562, + 388, + 574 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 562, + 388, + 574 + ], + "spans": [ + { + "bbox": [ + 313, + 562, + 388, + 574 + ], + "type": "text", + "content": "5. Conclusion" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 313, + 582, + 555, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 582, + 555, + 715 + ], + "spans": [ + { + "bbox": [ + 313, + 582, + 555, + 715 + ], + "type": "text", + "content": "We have proposed DrVideo, a document retrieval-based system for long video understanding. Different from previous LLM-based methods, DrVideo proposes to adapt long-video understanding to long-document understanding and searches for missing information via document retrieval and multi-stage agent loop. Extensive comparative experiments and ablation studies on the three challenging datasets, i.e., EgoSchema, MovieChat-1K, and Video-MME, demonstrate the effectiveness of DrVideo. We believe our work can serve as a strong baseline to stimulate further developments on the topic of long video understanding." + } + ] + } + ], + "index": 13 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "text", + "content": "18943" + } + ] + } + ], + "index": 14 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 7 + }, + { + "para_blocks": [ + { + "bbox": [ + 56, + 72, + 147, + 85 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 72, + 147, + 85 + ], + "spans": [ + { + "bbox": [ + 56, + 72, + 147, + 85 + ], + "type": "text", + "content": "Acknowledgment" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 56, + 163, + 114, + 176 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 163, + 114, + 176 + ], + "spans": [ + { + "bbox": [ + 56, + 163, + 114, + 176 + ], + "type": "text", + "content": "References" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 57, + 183, + 295, + 713 + ], + "type": "list", + "angle": 0, + "index": 14, + "blocks": [ + { + "bbox": [ + 61, + 183, + 294, + 237 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 61, + 183, + 294, + 237 + ], + "spans": [ + { + "bbox": [ + 61, + 183, + 294, + 237 + ], + "type": "text", + "content": "[1] Josh Achiam, Steven Adler, Sandhini Agarwal, Lama Ahmad, Ilge Akkaya, Florencia Leoni Aleman, Diogo Almeida, Janko Altenschmidt, Sam Altman, Shyamal Anadkat, et al. Gpt-4 technical report. arXiv preprint arXiv:2303.08774, 2023. 2, 5, 6, 7" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 62, + 239, + 295, + 272 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 239, + 295, + 272 + ], + "spans": [ + { + "bbox": [ + 62, + 239, + 295, + 272 + ], + "type": "text", + "content": "[2] Anthropic. Claude 3.5 Sonnet: Enhanced Conversational AI. https://www.anthropic.com/index/claude-3.5-sonnet/, 2024.6" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 62, + 274, + 295, + 328 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 274, + 295, + 328 + ], + "spans": [ + { + "bbox": [ + 62, + 274, + 295, + 328 + ], + "type": "text", + "content": "[3] Jinze Bai, Shuai Bai, Shusheng Yang, Shijie Wang, Sinan Tan, Peng Wang, Junyang Lin, Chang Zhou, and Jingren Zhou. Qwen-vl: A versatile vision-language model for understanding, localization, text reading, and beyond. arXiv preprint arXiv:2308.12966, 2023. 6" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 62, + 331, + 295, + 374 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 331, + 295, + 374 + ], + "spans": [ + { + "bbox": [ + 62, + 331, + 295, + 374 + ], + "type": "text", + "content": "[4] Aanisha Bhattacharya, Yaman K Singla, Balaji Krishnamurthy, Rajiv Ratn Shah, and Changyou Chen. A video is worth 4096 tokens: Verbalize story videos to understand them in zero shot. arXiv preprint arXiv:2305.09758, 2023. 3" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 62, + 376, + 294, + 430 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 376, + 294, + 430 + ], + "spans": [ + { + "bbox": [ + 62, + 376, + 294, + 430 + ], + "type": "text", + "content": "[5] Anthony Brohan, Noah Brown, Justice Carbajal, Yevgen Chebotar, Xi Chen, Krzysztof Choromanski, Tianli Ding, Danny Driess, Avinava Dubey, Chelsea Finn, et al. Rt-2: Vision-language-action models transfer web knowledge to robotic control. arXiv preprint arXiv:2307.15818, 2023. 3" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 62, + 432, + 294, + 486 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 432, + 294, + 486 + ], + "spans": [ + { + "bbox": [ + 62, + 432, + 294, + 486 + ], + "type": "text", + "content": "[6] Tom Brown, Benjamin Mann, Nick Ryder, Melanie Subbiah, Jared D Kaplan, Prafulla Dhariwal, Arvind Neelakantan, Pranav Shyam, Girish Sastry, Amanda Askell, et al. Language models are few-shot learners. Adv. Neural Inform. Process. Syst., pages 1877-1901, 2020. 1" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 62, + 488, + 294, + 531 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 488, + 294, + 531 + ], + "spans": [ + { + "bbox": [ + 62, + 488, + 294, + 531 + ], + "type": "text", + "content": "[7] Jun Chen, Deyao Zhu, Kilichbek Haydarov, Xiang Li, and Mohamed Elhoseiny. Video chatcaptioner: Towards the enriched spatiotemporal descriptions. arXiv preprint arXiv:2304.04227, 2023. 3" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 62, + 533, + 294, + 578 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 533, + 294, + 578 + ], + "spans": [ + { + "bbox": [ + 62, + 533, + 294, + 578 + ], + "type": "text", + "content": "[8] Feng Cheng, Xizi Wang, Jie Lei, David Crandall, Mohit Bansal, and Gedas Bertasius. Vindlu: A recipe for effective video-and-language pretraining. In IEEE/CVF Conf. Comput. Vis. Pattern Recog., pages 10739-10750, 2023. 1" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 62, + 579, + 294, + 621 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 579, + 294, + 621 + ], + "spans": [ + { + "bbox": [ + 62, + 579, + 294, + 621 + ], + "type": "text", + "content": "[9] Rohan Choudhury, Koichiro Niinuma, Kris M Kitani, and László A Jeni. Zero-shot video question answering with procedural programs. arXiv preprint arXiv:2312.00937, 2023. 5, 6" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 57, + 624, + 294, + 678 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 624, + 294, + 678 + ], + "spans": [ + { + "bbox": [ + 57, + 624, + 294, + 678 + ], + "type": "text", + "content": "[10] Aakanksha Chowdhery, Sharan Narang, Jacob Devlin, Maarten Bosma, Gaurav Mishra, Adam Roberts, Paul Barham, Hyung Won Chung, Charles Sutton, Sebastian Gehrmann, et al. Palm: Scaling language modeling with pathways. arXiv preprint arXiv:2204.02311, 2022. 1" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 57, + 681, + 294, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 681, + 294, + 713 + ], + "spans": [ + { + "bbox": [ + 57, + 681, + 294, + 713 + ], + "type": "text", + "content": "[11] Jiwan Chung and Youngjae Yu. Long story short: a summarize-then-search method for long video question answering. In Brit. Mach. Vis. Conf., 2023. 3" + } + ] + } + ], + "index": 13 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 57, + 73, + 553, + 713 + ], + "type": "list", + "angle": 0, + "index": 29, + "blocks": [ + { + "bbox": [ + 55, + 91, + 295, + 151 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 91, + 295, + 151 + ], + "spans": [ + { + "bbox": [ + 55, + 91, + 295, + 151 + ], + "type": "text", + "content": "This work is supported by the National Natural Science Foundation of China (62221002, 62171183, U24A20327, 62401202), the Joint Fund of the Ministry of Education of China, and the China Scholarship Council (CSC) under Grant No. 202306130148." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 316, + 73, + 553, + 116 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 73, + 553, + 116 + ], + "spans": [ + { + "bbox": [ + 316, + 73, + 553, + 116 + ], + "type": "text", + "content": "[12] Yiran Ding, Li Lyna Zhang, Chengruidong Zhang, Yuanyuan Xu, Ning Shang, Jiahang Xu, Fan Yang, and Mao Yang. Longgrope: Extending llm context window beyond 2 million tokens. arXiv preprint arXiv:2402.13753, 2024. 2" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 317, + 119, + 553, + 172 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 119, + 553, + 172 + ], + "spans": [ + { + "bbox": [ + 317, + 119, + 553, + 172 + ], + "type": "text", + "content": "[13] Danny Driess, Fei Xia, Mehdi SM Sajjadi, Corey Lynch, Aakanksha Chowdhery, Brian Ichter, Ayzaan Wahid, Jonathan Tompson, Quan Vuong, Tianhe Yu, et al. Palme: An embodied multimodal language model. arXiv preprint arXiv:2303.03378, 2023. 3" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 317, + 175, + 553, + 217 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 175, + 553, + 217 + ], + "spans": [ + { + "bbox": [ + 317, + 175, + 553, + 217 + ], + "type": "text", + "content": "[14] Yue Fan, Xiaojian Ma, Rujie Wu, Yuntao Du, Jiaqi Li, Zhi Gao, and Qing Li. Videoagent: A memory-augmented multimodal agent for video understanding. arXiv preprint arXiv:2403.11481, 2024. 2, 3, 5, 6" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 317, + 220, + 553, + 274 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 220, + 553, + 274 + ], + "spans": [ + { + "bbox": [ + 317, + 220, + 553, + 274 + ], + "type": "text", + "content": "[15] Chaoyou Fu, Yuhan Dai, Yondong Luo, Lei Li, Shuhuai Ren, Renrui Zhang, Zihan Wang, Chenyu Zhou, Yunhang Shen, Mengdan Zhang, et al. Video-mme: The first-ever comprehensive evaluation benchmark of multi-modal llms in video analysis. arXiv preprint arXiv:2405.21075, 2024. 2, 3, 5, 6" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 317, + 276, + 553, + 319 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 276, + 553, + 319 + ], + "spans": [ + { + "bbox": [ + 317, + 276, + 553, + 319 + ], + "type": "text", + "content": "[16] Chaoyou Fu, Haojia Lin, Zuwei Long, Yunhang Shen, Meng Zhao, Yifan Zhang, Xiong Wang, Di Yin, Long Ma, Xiawu Zheng, et al. Vita: Towards open-source interactive omni multimodal llm. arXiv preprint arXiv:2408.05211, 2024. 6" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 317, + 321, + 553, + 364 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 321, + 553, + 364 + ], + "spans": [ + { + "bbox": [ + 317, + 321, + 553, + 364 + ], + "type": "text", + "content": "[17] Difei Gao, Lei Ji, Luowei Zhou, Kevin Qinghong Lin, Joya Chen, Zihan Fan, and Mike Zheng Shou. Assistgpt: A general multi-modal assistant that can plan, execute, inspect, and learn. arXiv preprint arXiv:2306.08640, 2023. 3" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 317, + 365, + 553, + 398 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 365, + 553, + 398 + ], + "spans": [ + { + "bbox": [ + 317, + 365, + 553, + 398 + ], + "type": "text", + "content": "[18] Albert Gu, Karan Goel, and Christopher Ré. Efficiently modeling long sequences with structured state spaces. arXiv preprint arXiv:2111.00396, 2021. 3" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 317, + 399, + 553, + 453 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 399, + 553, + 453 + ], + "spans": [ + { + "bbox": [ + 317, + 399, + 553, + 453 + ], + "type": "text", + "content": "[19] Daya Guo, Qihao Zhu, Dejian Yang, Zhenda Xie, Kai Dong, Wentao Zhang, Guanting Chen, Xiao Bi, Yu Wu, YK Li, et al. Deepseek-coder: When the large language model meets programming-the rise of code intelligence. arXiv preprint arXiv:2401.14196, 2024. 5, 6, 7" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 317, + 456, + 553, + 498 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 456, + 553, + 498 + ], + "spans": [ + { + "bbox": [ + 317, + 456, + 553, + 498 + ], + "type": "text", + "content": "[20] Wenyi Hong, Weihan Wang, Qingsong Lv, Jiazheng Xu, Wenmeng Yu, Junhui Ji, Yan Wang, Zihan Wang, Yuxiao Dong, Ming Ding, et al. Cogagent: A visual language model for gui agents. arXiv preprint arXiv:2312.08914, 2023. 3" + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 317, + 501, + 553, + 542 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 501, + 553, + 542 + ], + "spans": [ + { + "bbox": [ + 317, + 501, + 553, + 542 + ], + "type": "text", + "content": "[21] Noureldien Hussein, Efstratios Gavves, and Arnold WM Smeulders. Videograph: Recognizing minutes-long human activities in videos. arXiv preprint arXiv:1905.05143, 2019. 3" + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 317, + 545, + 553, + 578 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 545, + 553, + 578 + ], + "spans": [ + { + "bbox": [ + 317, + 545, + 553, + 578 + ], + "type": "text", + "content": "[22] Md Mohaiminul Islam and Gedas Bertasius. Long movie clip classification with state-space video models. In Eur. Conf. Comput. Vis., pages 87-104, 2022. 3" + } + ] + } + ], + "index": 25 + }, + { + "bbox": [ + 317, + 580, + 553, + 633 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 580, + 553, + 633 + ], + "spans": [ + { + "bbox": [ + 317, + 580, + 553, + 633 + ], + "type": "text", + "content": "[23] Sullam Jeoung, Goeric Huybrechts, Bhavana Ganesh, Aram Galstyan, and Sravan Bodapati. Adaptive video understanding agent: Enhancing efficiency with dynamic frame sampling and feedback-driven reasoning. arXiv preprint arXiv:2410.20252, 2024. 5" + } + ] + } + ], + "index": 26 + }, + { + "bbox": [ + 317, + 635, + 553, + 689 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 635, + 553, + 689 + ], + "spans": [ + { + "bbox": [ + 317, + 635, + 553, + 689 + ], + "type": "text", + "content": "[24] Albert Q Jiang, Alexandre Sablayrolles, Arthur Mensch, Chris Bamford, Devendra Singh Chaplot, Diego de las Casas, Florian Bressand, Gianna Lengyel, Guillaume Lample, Lucile Saulnier, et al. Mistral 7b. arXiv preprint arXiv:2310.06825, 2023. 7" + } + ] + } + ], + "index": 27 + }, + { + "bbox": [ + 316, + 691, + 553, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 691, + 553, + 713 + ], + "spans": [ + { + "bbox": [ + 316, + 691, + 553, + 713 + ], + "type": "text", + "content": "[25] Wonkyun Kim, Changin Choi, Wonseok Lee, and Wonjong Rhee. An image grid can be worth a video: Zero" + } + ] + } + ], + "index": 28 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 749, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 749, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 749, + 317, + 757 + ], + "type": "text", + "content": "18944" + } + ] + } + ], + "index": 30 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 8 + }, + { + "para_blocks": [ + { + "bbox": [ + 56, + 73, + 296, + 712 + ], + "type": "list", + "angle": 0, + "index": 14, + "blocks": [ + { + "bbox": [ + 77, + 73, + 295, + 95 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 77, + 73, + 295, + 95 + ], + "spans": [ + { + "bbox": [ + 77, + 73, + 295, + 95 + ], + "type": "text", + "content": "shot video question answering using a vlm. arXiv preprint arXiv:2403.18406, 2024. 1, 5, 6" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 56, + 95, + 296, + 139 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 95, + 296, + 139 + ], + "spans": [ + { + "bbox": [ + 56, + 95, + 296, + 139 + ], + "type": "text", + "content": "[26] Takeshi Kojima, Shixiang Shane Gu, Machel Reid, Yutaka Matsuo, and Yusuke Iwasawa. Large language models are zero-shot reasoners. Adv. Neural Inform. Process. Syst., pages 22199-22213, 2022. 5" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 56, + 140, + 296, + 194 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 140, + 296, + 194 + ], + "spans": [ + { + "bbox": [ + 56, + 140, + 296, + 194 + ], + "type": "text", + "content": "[27] Jie Lei, Linjie Li, Luowei Zhou, Zhe Gan, Tamara L Berg, Mohit Bansal, and Jingjing Liu. Less is more: Clipbert for video-and-language learning via sparse sampling. In IEEE/CVF Conf. Comput. Vis. Pattern Recog., pages 7331-7341, 2021. 1" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 56, + 194, + 295, + 239 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 194, + 295, + 239 + ], + "spans": [ + { + "bbox": [ + 56, + 194, + 295, + 239 + ], + "type": "text", + "content": "[28] Bo Li, Yuanhan Zhang, Dong Guo, Renrui Zhang, Feng Li, Hao Zhang, Kaichen Zhang, Yanwei Li, Ziwei Liu, and Chunyuan Li. Llava-onevision: Easy visual task transfer. arXiv preprint arXiv:2408.03326, 2024. 6" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 56, + 239, + 295, + 282 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 239, + 295, + 282 + ], + "spans": [ + { + "bbox": [ + 56, + 239, + 295, + 282 + ], + "type": "text", + "content": "[29] Junnan Li, Dongxu Li, Silvio Savarese, and Steven Hoi. BLIP-2: bootstrapping language-image pre-training with frozen image encoders and large language models. In Proc. Int. Conf. Mach. Learn., 2023. 4, 7" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 56, + 284, + 295, + 327 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 284, + 295, + 327 + ], + "spans": [ + { + "bbox": [ + 56, + 284, + 295, + 327 + ], + "type": "text", + "content": "[30] KunChang Li, Yinan He, Yi Wang, Yizhuo Li, Wenhai Wang, Ping Luo, Yali Wang, Limin Wang, and Yu Qiao. Videochat: Chat-centric video understanding. arXiv preprint arXiv:2305.06355, 2023. 3, 5, 6" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 56, + 327, + 295, + 370 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 327, + 295, + 370 + ], + "spans": [ + { + "bbox": [ + 56, + 327, + 295, + 370 + ], + "type": "text", + "content": "[31] Bin Lin, Bin Zhu, Yang Ye, Munan Ning, Peng Jin, and Li Yuan. Video-llava: Learning united visual representation by alignment before projection. arXiv preprint arXiv:2311.10122, 2023. 6" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 56, + 371, + 295, + 426 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 371, + 295, + 426 + ], + "spans": [ + { + "bbox": [ + 56, + 371, + 295, + 426 + ], + "type": "text", + "content": "[32] Kevin Lin, Faisal Ahmed, Linjie Li, Chung-Ching Lin, Ehsan Azarnasab, Zhengyuan Yang, Jianfeng Wang, Lin Liang, Zicheng Liu, Yumao Lu, et al. Mm-vid: Advancing video understanding with gpt-4v (ision). arXiv preprint arXiv:2310.19773, 2023. 1, 3, 5" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 56, + 426, + 295, + 480 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 426, + 295, + 480 + ], + "spans": [ + { + "bbox": [ + 56, + 426, + 295, + 480 + ], + "type": "text", + "content": "[33] Haotian Liu, Chunyuan Li, Yuheng Li, Bo Li, Yuanhan Zhang, Sheng Shen, and Yong Jae Lee. Llava next: Improved reasoning,OCR,and world knowledge.https://llava-v1.github.io/blog/2024-01-30-llava NEXT/,2024.2,4,5,6,7" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 56, + 481, + 295, + 525 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 481, + 295, + 525 + ], + "spans": [ + { + "bbox": [ + 56, + 481, + 295, + 525 + ], + "type": "text", + "content": "[34] Ziyu Ma, Shutao Li, Bin Sun, Jianfei Cai, Zuxiang Long, and Fuyan Ma. Gerea: Question-aware prompt captions for knowledge-based visual question answering. arXiv preprint arXiv:2402.02503, 2024. 1" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 56, + 525, + 295, + 570 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 525, + 295, + 570 + ], + "spans": [ + { + "bbox": [ + 56, + 525, + 295, + 570 + ], + "type": "text", + "content": "[35] Muhammad Maaz, Hanoona Rasheed, Salman Khan, and Fahad Shahbaz Khan. Video-chatgpt: Towards detailed video understanding via large vision and language models. arXiv preprint arXiv:2306.05424, 2023. 5" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 56, + 571, + 295, + 613 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 571, + 295, + 613 + ], + "spans": [ + { + "bbox": [ + 56, + 571, + 295, + 613 + ], + "type": "text", + "content": "[36] Karttikeya Mangalam, Raiymbek Akshulakov, and Jitendra Malik. Egoschema: A diagnostic benchmark for very long-form video language understanding. arXiv preprint arXiv:2308.09126, 2023. 2, 3, 5, 6, 7, 8" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 56, + 614, + 295, + 668 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 614, + 295, + 668 + ], + "spans": [ + { + "bbox": [ + 56, + 614, + 295, + 668 + ], + "type": "text", + "content": "[37] Juhong Min, Shyamal Buch, Arsha Nagrani, Minsu Cho, and Cordelia Schmid. Morevqa: Exploring modular reasoning models for video question answering. In IEEE/CVF Conf. Comput. Vis. Pattern Recog., pages 13235-13245, 2024. 2, 5, 6" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 56, + 669, + 295, + 712 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 669, + 295, + 712 + ], + "spans": [ + { + "bbox": [ + 56, + 669, + 295, + 712 + ], + "type": "text", + "content": "[38] Salman Khan Muhammad Maaz, Hanoona Rasheed and Fahad Khan. Video-chatgpt: Towards detailed video understanding via large vision and language models. ArXiv 2306.05424, 2023. 1, 5, 6" + } + ] + } + ], + "index": 13 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 316, + 73, + 553, + 713 + ], + "type": "list", + "angle": 0, + "index": 30, + "blocks": [ + { + "bbox": [ + 316, + 73, + 553, + 127 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 73, + 553, + 127 + ], + "spans": [ + { + "bbox": [ + 316, + 73, + 553, + 127 + ], + "type": "text", + "content": "[39] Eric Nguyen, Karan Goel, Albert Gu, Gordon Downs, Preey Shah, Tri Dao, Stephen Baccus, and Christopher Ré. S4nd: Modeling images and videos as multidimensional signals with state spaces. Adv. Neural Inform. Process. Syst., pages 2846-2861, 2022. 3" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 316, + 128, + 553, + 150 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 128, + 553, + 150 + ], + "spans": [ + { + "bbox": [ + 316, + 128, + 553, + 150 + ], + "type": "text", + "content": "[40] OpenAI. GPT3.5. https://platform.openai.com/docs/models/gpt-3-5, 2021.1, 5, 7" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 316, + 151, + 553, + 172 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 151, + 553, + 172 + ], + "spans": [ + { + "bbox": [ + 316, + 151, + 553, + 172 + ], + "type": "text", + "content": "[41] OpenAI. Introducing chatgpt. https://openai.com/blog/chatgpt, 2022.1" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 316, + 173, + 553, + 205 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 173, + 553, + 205 + ], + "spans": [ + { + "bbox": [ + 316, + 173, + 553, + 205 + ], + "type": "text", + "content": "[42] OpenAI. GPT-4V(ison) System Card. https://openai.com/index/gpt-4v-system-card/, 2023.3,6" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 316, + 205, + 553, + 239 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 205, + 553, + 239 + ], + "spans": [ + { + "bbox": [ + 316, + 205, + 553, + 239 + ], + "type": "text", + "content": "[43] OpenAI. New embedding models and api updates. https://openai.com/index/new-embedding-models-and-api-updates/, 2024.4" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 316, + 239, + 553, + 282 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 239, + 553, + 282 + ], + "spans": [ + { + "bbox": [ + 316, + 239, + 553, + 282 + ], + "type": "text", + "content": "[44] OpenAI. GPT-4o Mini: Advancing Cost-Efficient Intelligence. https://openai.com/index/gpt-4o-mini-advancing-cost-efficient-intelligence/, 2024.3,6" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 316, + 284, + 553, + 338 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 284, + 553, + 338 + ], + "spans": [ + { + "bbox": [ + 316, + 284, + 553, + 338 + ], + "type": "text", + "content": "[45] Long Ouyang, Jeffrey Wu, Xu Jiang, Diogo Almeida, Carroll Wainwright, Pamela Mishkin, Chong Zhang, Sandhini Agarwal, Katarina Slama, Alex Ray, et al. Training language models to follow instructions with human feedback. Adv. Neural Inform. Process. Syst., pages 27730-27744, 2022. 1" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 316, + 339, + 553, + 392 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 339, + 553, + 392 + ], + "spans": [ + { + "bbox": [ + 316, + 339, + 553, + 392 + ], + "type": "text", + "content": "[46] Vincent Perot, Kai Kang, Florian Luisier, Guolong Su, Xiaoyu Sun, Ramya Sree Boppana, Zilong Wang, Zifeng Wang, Jiaqi Mu, Hao Zhang, et al. Lmdx: Language model-based document information extraction and localization. arXiv preprint arXiv:2309.10952, 2023. 2" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 316, + 393, + 553, + 437 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 393, + 553, + 437 + ], + "spans": [ + { + "bbox": [ + 316, + 393, + 553, + 437 + ], + "type": "text", + "content": "[47] Kanchana Ranasinghe, Xiang Li, Kumara Kahatapitiya, and Michael S Ryoo. Understanding long videos in one multimodal language model pass. arXiv preprint arXiv:2403.16998, 2024. 5, 6" + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 316, + 437, + 553, + 492 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 437, + 553, + 492 + ], + "spans": [ + { + "bbox": [ + 316, + 437, + 553, + 492 + ], + "type": "text", + "content": "[48] Teven Le Scao, Angela Fan, Christopher Akiki, Ellie Pavlick, Suzana Ilic, Daniel Hesslow, Roman Castagné, Alexandra Sasha Luccioni, François Yvon, Matthias Galle, et al. Bloom: A 176b-parameter open-access multilingual language model. arXiv preprint arXiv:2211.05100, 2022. 1" + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 316, + 492, + 553, + 548 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 492, + 553, + 548 + ], + "spans": [ + { + "bbox": [ + 316, + 492, + 553, + 548 + ], + "type": "text", + "content": "[49] Enxin Song, Wenhao Chai, Guanhong Wang, Yucheng Zhang, Haoyang Zhou, Feiyang Wu, Xun Guo, Tian Ye, Yan Lu, Jenq-Neng Hwang, et al. Moviechat: From dense token to sparse memory for long video understanding. arXiv preprint arXiv:2307.16449, 2023. 2, 3, 5, 6" + } + ] + } + ], + "index": 25 + }, + { + "bbox": [ + 316, + 548, + 553, + 591 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 548, + 553, + 591 + ], + "spans": [ + { + "bbox": [ + 316, + 548, + 553, + 591 + ], + "type": "text", + "content": "[50] Enxin Song, Wenhao Chai, Tian Ye, Jenq-Neng Hwang, Xi Li, and Gaoang Wang. Moviechat+: Question-aware sparse memory for long video question answering. arXiv preprint arXiv:2404.17176, 2024. 5" + } + ] + } + ], + "index": 26 + }, + { + "bbox": [ + 316, + 592, + 553, + 635 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 592, + 553, + 635 + ], + "spans": [ + { + "bbox": [ + 316, + 592, + 553, + 635 + ], + "type": "text", + "content": "[51] Quan Sun, Jinsheng Wang, Qiying Yu, Yufeng Cui, Fan Zhang, Xiaosong Zhang, and Xinlong Wang. Eva-clip-18b: Scaling clip to 18 billion parameters. arXiv preprint arXiv:2402.04252, 2024. 2" + } + ] + } + ], + "index": 27 + }, + { + "bbox": [ + 316, + 635, + 553, + 680 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 635, + 553, + 680 + ], + "spans": [ + { + "bbox": [ + 316, + 635, + 553, + 680 + ], + "type": "text", + "content": "[52] Yuchong Sun, Hongwei Xue, Ruihua Song, Bei Liu, Huan Yang, and Jianlong Fu. Long-form video-language pretraining with multimodal temporal contrastive learning. Adv. Neural Inform. Process. Syst., 35:38032-38045, 2022. 3" + } + ] + } + ], + "index": 28 + }, + { + "bbox": [ + 316, + 681, + 553, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 681, + 553, + 713 + ], + "spans": [ + { + "bbox": [ + 316, + 681, + 553, + 713 + ], + "type": "text", + "content": "[53] Dídac Surís, Sachit Menon, and Carl Vondrick. Vipergpt: Visual inference via python execution for reasoning. Int. Conf. Comput. Vis., pages 11888-11898, 2023. 1, 3" + } + ] + } + ], + "index": 29 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 749, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 749, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 749, + 317, + 757 + ], + "type": "text", + "content": "18945" + } + ] + } + ], + "index": 31 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 9 + }, + { + "para_blocks": [ + { + "bbox": [ + 56, + 72, + 297, + 713 + ], + "type": "list", + "angle": 0, + "index": 13, + "blocks": [ + { + "bbox": [ + 56, + 72, + 297, + 127 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 72, + 297, + 127 + ], + "spans": [ + { + "bbox": [ + 56, + 72, + 297, + 127 + ], + "type": "text", + "content": "[54] Gemini Team, Rohan Anil, Sebastian Borgeaud, Yonghui Wu, Jean-Baptiste Alayrac, Jiahui Yu, Radu Soricut, Johan Schalkwyk, Andrew M Dai, Anja Hauth, et al. Gemini: a family of highly capable multimodal models. arXiv preprint arXiv:2312.11805, 2023. 5" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 56, + 129, + 295, + 185 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 129, + 295, + 185 + ], + "spans": [ + { + "bbox": [ + 56, + 129, + 295, + 185 + ], + "type": "text", + "content": "[55] Gemini Team, Petko Georgiev, Ving Ian Lei, Ryan Burnell, Libin Bai, Anmol Gulati, Garrett Tanzer, Damien Vincent, Zhufeng Pan, Shibo Wang, et al. Gemini 1.5: Unlocking multimodal understanding across millions of tokens of context. arXiv preprint arXiv:2403.05530, 2024. 3, 6" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 56, + 185, + 295, + 239 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 185, + 295, + 239 + ], + "spans": [ + { + "bbox": [ + 56, + 185, + 295, + 239 + ], + "type": "text", + "content": "[56] Hugo Touvron, Thibaut Lavril, Gautier Izacard, Xavier Martinet, Marie-Anne Lachaux, Timothée Lacroix, Baptiste Rozière, Naman Goyal, Eric Hambro, Faisal Azhar, et al. Llama: Open and efficient foundation language models. arXiv preprint arXiv:2302.13971, 2023. 1" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 56, + 241, + 295, + 285 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 241, + 295, + 285 + ], + "spans": [ + { + "bbox": [ + 56, + 241, + 295, + 285 + ], + "type": "text", + "content": "[57] Junke Wang, Dongdong Chen, Chong Luo, Xiyang Dai, Lu Yuan, Zuxuan Wu, and Yu-Gang Jiang. Chatvideo: A tracklet-centric multimodal and versatile video understanding system. arXiv preprint arXiv:2304.14407, 2023. 3" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 56, + 286, + 295, + 341 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 286, + 295, + 341 + ], + "spans": [ + { + "bbox": [ + 56, + 286, + 295, + 341 + ], + "type": "text", + "content": "[58] Jue Wang, Wentao Zhu, Pichao Wang, Xiang Yu, Linda Liu, Mohamed Omar, and Raffay Hamid. Selective structured state-spaces for long-form video understanding. In IEEE/CVF Conf. Comput. Vis. Pattern Recog., pages 6387-6397, 2023. 3" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 56, + 342, + 295, + 376 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 342, + 295, + 376 + ], + "spans": [ + { + "bbox": [ + 56, + 342, + 295, + 376 + ], + "type": "text", + "content": "[59] Jiawei Wang, Liping Yuan, and Yuchen Zhang. Tarsier: Recipes for training and evaluating large video description models. arXiv preprint arXiv:2407.00634, 2024. 8" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 56, + 376, + 295, + 421 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 376, + 295, + 421 + ], + "spans": [ + { + "bbox": [ + 56, + 376, + 295, + 421 + ], + "type": "text", + "content": "[60] Shijie Wang, Qi Zhao, Minh Quan Do, Nakul Agarwal, Kwonjoon Lee, and Chen Sun. Vamos: Versatile action models for video understanding. arXiv preprint arXiv:2311.13627, 2023. 5, 6" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 56, + 422, + 295, + 465 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 422, + 295, + 465 + ], + "spans": [ + { + "bbox": [ + 56, + 422, + 295, + 465 + ], + "type": "text", + "content": "[61] Xiaohan Wang, Yuhui Zhang, Orr Zohar, and Serena Yeung-Levy. Videoagent: Long-form video understanding with large language model as agent. arXiv preprint arXiv:2403.10517, 2024. 1, 2, 3, 5, 6, 7" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 56, + 467, + 295, + 511 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 467, + 295, + 511 + ], + "spans": [ + { + "bbox": [ + 56, + 467, + 295, + 511 + ], + "type": "text", + "content": "[62] Yang Wang, Gedas Bertasius, Tae-Hyun Oh, Abhinav Gupta, Minh Hoai, and Lorenzo Torresani. Supervoxel attention graphs for long-range video modeling. In Proc. IEEE/CVF Winter Conf. Appl. Comput. Vis., pages 155-166, 2021. 3" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 56, + 512, + 295, + 567 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 512, + 295, + 567 + ], + "spans": [ + { + "bbox": [ + 56, + 512, + 295, + 567 + ], + "type": "text", + "content": "[63] Yi Wang, Yinan He, Yizhuo Li, Kunchang Li, Jiashuo Yu, Xin Ma, Xinhao Li, Guo Chen, Xinyuan Chen, Yaohui Wang, et al. Intervid: A large-scale video-text dataset for multimodal understanding and generation. arXiv preprint arXiv:2307.06942, 2023. 6" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 56, + 568, + 295, + 633 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 568, + 295, + 633 + ], + "spans": [ + { + "bbox": [ + 56, + 568, + 295, + 633 + ], + "type": "text", + "content": "[64] Zhenhailong Wang, Manling Li, Ruochen Xu, Luowei Zhou, Jie Lei, Xudong Lin, Shuohang Wang, Ziyi Yang, Chenguang Zhu, Derek Hoiem, et al. Language models with image descriptors are strong few-shot video-language learners. Adv. Neural Inform. Process. Syst., pages 8483-8497, 2022. 3" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 56, + 635, + 295, + 689 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 635, + 295, + 689 + ], + "spans": [ + { + "bbox": [ + 56, + 635, + 295, + 689 + ], + "type": "text", + "content": "[65] Ziyang Wang, Shoubin Yu, Elias Stengel-Eskin, Jaehong Yoon, Feng Cheng, Gedas Bertasius, and Mohit Bansal. Videotree: Adaptive tree-based video representation for llm reasoning on long videos. arXiv preprint arXiv:2405.19209, 2024.3" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 56, + 691, + 295, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 691, + 295, + 713 + ], + "spans": [ + { + "bbox": [ + 56, + 691, + 295, + 713 + ], + "type": "text", + "content": "[66] Jason Wei, Xuezhi Wang, Dale Schuurmans, Maarten Bosma, Fei Xia, Ed Chi, Quoc V Le, Denny Zhou, et al." + } + ] + } + ], + "index": 12 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 316, + 73, + 553, + 689 + ], + "type": "list", + "angle": 0, + "index": 27, + "blocks": [ + { + "bbox": [ + 333, + 73, + 553, + 106 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 333, + 73, + 553, + 106 + ], + "spans": [ + { + "bbox": [ + 333, + 73, + 553, + 106 + ], + "type": "text", + "content": "Chain-of-thought prompting elicits reasoning in large language models. Adv. Neural Inform. Process. Syst., pages 24824-24837, 2022. 5" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 316, + 107, + 553, + 162 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 107, + 553, + 162 + ], + "spans": [ + { + "bbox": [ + 316, + 107, + 553, + 162 + ], + "type": "text", + "content": "[67] Chao-Yuan Wu, Yanghao Li, Karttikeya Mangalam, Haoqi Fan, Bo Xiong, Jitendra Malik, and Christoph Feichtenhofer. Memvit: Memory-augmented multiscale vision transformer for efficient long-term video recognition. In IEEE/CVF Conf. Comput. Vis. Pattern Recog., pages 13587-13597, 2022. 3" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 316, + 163, + 553, + 206 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 163, + 553, + 206 + ], + "spans": [ + { + "bbox": [ + 316, + 163, + 553, + 206 + ], + "type": "text", + "content": "[68] Lin Xu, Yilin Zhao, Daquan Zhou, Zhijie Lin, See Kiong Ng, and Jiashi Feng. Pllava: Parameter-free llava extension from images to videos for video dense captioning. arXiv preprint arXiv:2404.16994, 2024. 1" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 316, + 208, + 553, + 252 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 208, + 553, + 252 + ], + "spans": [ + { + "bbox": [ + 316, + 208, + 553, + 252 + ], + "type": "text", + "content": "[69] Antoine Yang, Antoine Miech, Josef Sivic, Ivan Laptev, and Cordelia Schmid. Just ask: Learning to answer questions from millions of narrated videos. In Int. Conf. Comput. Vis., pages 1686-1697, 2021. 1" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 316, + 253, + 553, + 296 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 253, + 553, + 296 + ], + "spans": [ + { + "bbox": [ + 316, + 253, + 553, + 296 + ], + "type": "text", + "content": "[70] Antoine Yang, Antoine Miech, Josef Sivic, Ivan Laptev, and Cordelia Schmid. Zero-shot video question answering via frozen bidirectional language models. Adv. Neural Inform. Process. Syst., 35:124-141, 2022. 1" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 316, + 297, + 553, + 342 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 297, + 553, + 342 + ], + "spans": [ + { + "bbox": [ + 316, + 297, + 553, + 342 + ], + "type": "text", + "content": "[71] Xitong Yang, Fu-Jen Chu, Matt Feiszli, Raghav Goyal, Lorenzo Torresani, and Du Tran. Relational space-time query in long-form videos. In IEEE/CVF Conf. Comput. Vis. Pattern Recog., pages 6398-6408, 2023. 3" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 316, + 342, + 553, + 386 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 342, + 553, + 386 + ], + "spans": [ + { + "bbox": [ + 316, + 342, + 553, + 386 + ], + "type": "text", + "content": "[72] Zongxin Yang, Guikun Chen, Xiaodi Li, Wenguan Wang, and Yi Yang. Doraemongpt: Toward understanding dynamic scenes with large language models. arXiv preprint arXiv:2401.08392, 2024. 3" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 316, + 387, + 553, + 453 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 387, + 553, + 453 + ], + "spans": [ + { + "bbox": [ + 316, + 387, + 553, + 453 + ], + "type": "text", + "content": "[73] Andy Zeng, Maria Attarian, Brian Ichter, Krzysztof Choromanski, Adrian Wong, Stefan Welker, Federico Tombari, Aveek Purohit, Michael Ryoo, Vikas Sindhwani, Johnny Lee, Vincent Vanhoucke, and Pete Florence. Socratic models: Composing zero-shot multimodal reasoning with language. arXiv preprint arXiv:2204.00598, 2022. 3" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 316, + 454, + 553, + 497 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 454, + 553, + 497 + ], + "spans": [ + { + "bbox": [ + 316, + 454, + 553, + 497 + ], + "type": "text", + "content": "[74] Ce Zhang, Taixi Lu, Md Mohaiminul Islam, Ziyang Wang, Shoubin Yu, Mohit Bansal, and Gedas Bertasius. A simple llm framework for long-range video question-answering. arXiv preprint arXiv:2312.17235, 2023. 1, 2, 5, 6, 7" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 316, + 498, + 553, + 532 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 498, + 553, + 532 + ], + "spans": [ + { + "bbox": [ + 316, + 498, + 553, + 532 + ], + "type": "text", + "content": "[75] Hang Zhang, Xin Li, and Lidong Bing. Video-llama: An instruction-tuned audio-visual language model for video understanding. arXiv preprint arXiv:2306.02858, 2023. 5, 6" + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 316, + 533, + 553, + 586 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 533, + 553, + 586 + ], + "spans": [ + { + "bbox": [ + 316, + 533, + 553, + 586 + ], + "type": "text", + "content": "[76] Peiyuan Zhang, Kaichen Zhang, Bo Li, Guangtao Zeng, Jingkang Yang, Yuanhan Zhang, Ziyue Wang, Haoran Tan, Chunyuan Li, and Ziwei Liu. Long context transfer from language to vision. arXiv preprint arXiv:2406.16852, 2024. 6" + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 316, + 588, + 553, + 643 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 588, + 553, + 643 + ], + "spans": [ + { + "bbox": [ + 316, + 588, + 553, + 643 + ], + "type": "text", + "content": "[77] Yuanhan Zhang, Bo Li, Haotian Liu, Yong Jae Lee, Liangke Gui, Di Fu, Jiashi Feng, Ziwei Liu, and Chunyuan Li. Llava-last: A strong zero-shot video understanding model. https://llava-v1.github.io/blog/2024-04-30-llava-last-video/, 2024.1" + } + ] + } + ], + "index": 25 + }, + { + "bbox": [ + 316, + 644, + 553, + 689 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 644, + 553, + 689 + ], + "spans": [ + { + "bbox": [ + 316, + 644, + 553, + 689 + ], + "type": "text", + "content": "[78] Yue Zhao, Ishan Misra, Philipp Krahenbuhl, and Rohit Girdhar. Learning video representations from large language models. In IEEE/CVF Conf. Comput. Vis. Pattern Recog., pages 6586-6597, 2023. 1, 2, 4, 5, 6, 7" + } + ] + } + ], + "index": 26 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 749, + 318, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 749, + 318, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 749, + 318, + 757 + ], + "type": "text", + "content": "18946" + } + ] + } + ], + "index": 28 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 10 + } + ], + "_backend": "vlm", + "_version_name": "2.6.4" +} \ No newline at end of file diff --git a/2025/Dragin3D_ Image Editing by Dragging in 3D Space/63363afe-acf7-462c-bf4b-9a484c6ae8ae_content_list.json b/2025/Dragin3D_ Image Editing by Dragging in 3D Space/63363afe-acf7-462c-bf4b-9a484c6ae8ae_content_list.json new file mode 100644 index 0000000000000000000000000000000000000000..3d866320ba32e630661fe1ffbf1373b1dbd0b666 --- /dev/null +++ b/2025/Dragin3D_ Image Editing by Dragging in 3D Space/63363afe-acf7-462c-bf4b-9a484c6ae8ae_content_list.json @@ -0,0 +1,1487 @@ +[ + { + "type": "text", + "text": "Dragin3D: Image Editing by Dragging in 3D Space", + "text_level": 1, + "bbox": [ + 241, + 130, + 756, + 152 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Weiran Guang $^{12}$ Xiaoguang Gu $^{2*}$ Mengqi Huang $^{1}$ Zhendong Mao $^{1}$", + "bbox": [ + 205, + 179, + 787, + 198 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "1University of Science and Technology of China", + "bbox": [ + 305, + 198, + 691, + 217 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "$^{2}$ State Key Laboratory of Communication Content Cognition, People's Daily Online", + "bbox": [ + 163, + 215, + 828, + 234 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "guangwr, huangmq}@mail.ustc.edu.cn, guxiaoguang@people.cn, zdmao@ustc.edu.cn", + "bbox": [ + 153, + 236, + 831, + 251 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Abstract", + "text_level": 1, + "bbox": [ + 246, + 286, + 326, + 301 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Interactive drag editing of images is a valuable task that has gained considerable attention for its precision and controllability. However, existing approaches have primarily focused on manipulating the shape or movement of objects in 2D plane. We propose to extend this drag-based editing task to 3D space. Firstly, we utilize the trajectory of two points to represent the rotational trajectory of the object. Gaussian maps of a circle and a square are centered at these two points, respectively. We use distinct shapes to ensure that symmetric views produce different object representations. Secondly, we introduce a lightweight mapping network to embed the object features into two Gaussian maps to obtain a continuous control condition that guides the model in learning the correspondence between the trajectory and the object. Finally, to overcome the limitations of current 3D object reconstruction datasets, which typically consist of object maps with transparent backgrounds, we affix random backgrounds to them. This modification helps improve the model's ability to ignore background interference when editing real images with complex backgrounds. Experiments demonstrate that our approach successfully achieves object rotation within the drag framework and demonstrates strong generalization to real-world images.", + "bbox": [ + 89, + 318, + 485, + 667 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "1. Introduction", + "text_level": 1, + "bbox": [ + 91, + 696, + 220, + 713 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Drag-based image editing has garnered substantial research interest in recent years, aiming to enable the manipulation of objects via user-defined handle and target points [30, 36]. Most existing image editing works are text-based, which heavily depends on textual and semantic alignment. Sometimes, it is challenging to articulate user-desired edits in words. Drag-based editing offers a more intuitive, interactive, and granular editing signal, allowing for precise manipulation at the pixel level. Users specify the start and end points based on intended edits, eliminating the need for a", + "bbox": [ + 89, + 723, + 483, + 875 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "proper prompt. This enables user-friendly image modification and offers more precise control over positioning, making it valuable for practical applications.", + "bbox": [ + 511, + 287, + 903, + 333 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Existing drag-based editing methods can be categorized into optimized-based and training-based. DragGAN [30] is an early representative of optimized-based methods, encompassing two key components: point tracking and motion supervision. The latent code of StyleGAN [13] is optimized while updating the position of the handle point, ensuring that the features near the target points align with those near the handle points. Since DragGAN was proposed, both drag-based image and video editing techniques have been significantly advanced. Subsequently, researchers extended this process to diffusion models by optimizing the features of UNet to break through the inherent limitation of the GAN model [27, 36]. Recent studies have focused on addressing issues related to inaccurate point tracking and incomplete motion supervision, enabling more stable dragging by formulating better point tracking and motion supervision protocols [7, 19]. Furthermore, to overcome the limitations posed by variations in object motion patterns, the training-based method [43, 45] has been introduced using trajectory guidance, where the motion trajectory of the object is encoded as a control condition. In conclusion, ongoing efforts aim to enhance editing controllability and fidelity.", + "bbox": [ + 511, + 337, + 906, + 684 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "However, existing drag-based methods are limited to 2D drag editing and fail to achieve drag in 3D space, which significantly limits their practical application in the real 3D world. 3D generation has significantly expanded the application in games, animation, virtual production, etc. It enables more realistic creation of 3D visual spaces that closely resemble the physical world. We consider whether it is possible to generate novel view images of the object through simple point-drag interactions. The two drag paradigms discussed above are inadequate for solving this task. The optimization-based approach primarily seeks to supervise feature consistency between the handle point and the target point, which cannot well control the substantial changes in features that arise due to variations in spatial perspective, as", + "bbox": [ + 511, + 688, + 908, + 902 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "CVF", + "bbox": [ + 106, + 2, + 181, + 42 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore.", + "bbox": [ + 236, + 0, + 807, + 46 + ], + "page_idx": 0 + }, + { + "type": "page_footnote", + "text": "*Xiaoguang Gu is the Corresponding author.", + "bbox": [ + 107, + 886, + 346, + 900 + ], + "page_idx": 0 + }, + { + "type": "page_number", + "text": "21502", + "bbox": [ + 478, + 944, + 519, + 957 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/91d03067ba77be0f3c380cf0c1ce35f1e9c1dc5859c593528bbd75c7f4777e22.jpg", + "image_caption": [ + "Figure 1. Existing drag method for object rotation representation. (a) The optimization-based method. The sculpture makes tiny rotations, and changes in the position of the target point still tend to maintain a frontal perspective, which fails to achieve 3D editing with varying amplitudes based solely on the dragging intensity. (b) The training-based method. In the upper figure, the object moves along the trajectory of the dragging point. In the lower figure, the object rotates in the direction of the trajectory of the dragging point. The trajectory is inclined upwards, with the ship exhibiting a tendency to rotate towards the target position. Arguably, whether the ship can achieve 3D rotation is entirely random." + ], + "image_footnote": [], + "bbox": [ + 96, + 88, + 478, + 220 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "illustrated in Fig. 1(a). On the other hand, the trajectory-based methods introduce ambiguity in the trend of object motion, as demonstrated in Fig. 1(b). Therefore, in this study, we introduce a new drag-based object rotation task.", + "bbox": [ + 89, + 411, + 482, + 470 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "The key challenges of this task lie in how to accurately represent and track the desired 3D editing trajectory from a 2D image, and how to establish a correspondence between object motion trajectories and their respective representations. Due to the ambiguity in existing training-based methods for single-point trajectory representation, combined with the feature alignment principle of optimization-based methods, we propose Dragin3D, which extends drag-based editing to 3D space for the first time. We introduce a double-point representation approach that associates object trajectories with image features. Given an image of an object and a double-point rotation trajectory, Dragin3D generates a novel view image of the object from the corresponding trajectory, as shown in Fig. 2. The core concept of Dragin3D is to precisely capture the direction and magnitude of the object's rotation using the double-point trajectory. Specifically, our method consists of two main components: (1) Drag Trajectory Representation (DTR): we map the rotational trajectory of the object onto the rotational trajectory of a line segment, track the two endpoints of this segment, and project them into Gaussian maps with different shapes; (2) Control Condition Inject Module (CCIM): to enhance fidelity to the object, we extract image features through a lightweight mapping network and inject these features into the corresponding Gaussian maps.", + "bbox": [ + 89, + 473, + 482, + 849 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Our contributions are summarized as follows:", + "bbox": [ + 109, + 851, + 413, + 864 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "- We propose an innovative interactive image editing method Dragin3D. To the best of our knowledge, we are", + "bbox": [ + 89, + 869, + 482, + 901 + ], + "page_idx": 1 + }, + { + "type": "image", + "img_path": "images/c22e3b4194ee56c01cc287d1fc6892a7586d85a68b7d4953b53ca8781b6f5721.jpg", + "image_caption": [ + "Figure 2. Drag results of our proposed paradigm Dragin3D. Given an input image and coordinates of two representation points, i.e., $[P_{11}, P_{21}],[P_{12}, P_{22}],\\ldots,[P_{1n}, P_{2n}]$ . Dragin3D can generate a novel view image of the object for the corresponding trajectory." + ], + "image_footnote": [], + "bbox": [ + 519, + 88, + 903, + 196 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "the first to achieve multi-view image generation of objects by dragging them in 3D space based on a diffusion model.", + "bbox": [ + 524, + 299, + 903, + 329 + ], + "page_idx": 1 + }, + { + "type": "list", + "sub_type": "text", + "list_items": [ + "- We propose to represent the object rotation trajectory by a double-point motion trajectory, which maps 3D object motion onto a 2D plane and eliminates the ambiguity of the object motion mode.", + "- We introduce a tracking point projection method, which projects the tracking points onto Gaussian maps and aligns them with the object's feature information.", + "- Experimental results demonstrate that despite being trained on a synthetic 3D dataset, our method generalizes well beyond the dataset's distribution." + ], + "bbox": [ + 513, + 330, + 903, + 479 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "2. Related Work", + "text_level": 1, + "bbox": [ + 513, + 508, + 653, + 523 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "2.1. Drag-based controllable image editing", + "text_level": 1, + "bbox": [ + 511, + 539, + 841, + 556 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "In the field of image editing, several established works have focused on tasks such as image customization [16, 18, 33, 41], image editing [1, 4, 6, 9, 11, 14], which modify object attributes, quantity, etc. through text-based operations. Due to the limitations of text-based editing in controlling local changes at a fine-grained level, the Drag approach introduces a novel approach to image editing [7, 8, 19, 23, 24, 27-30, 36]. This interactive manipulation allows users to precisely designate both the editing area and control mode with a simple click. [30, 36] propose a point-based editing technique that integrates point tracking and motion supervision. [19] enhances drag editing stability and reliability without requiring precise point tracking by introducing adaptively updated template features, fuzzy localization, and linear search techniques. [7] introduces a discriminative point tracking method and a confidence-based latent space enhancement strategy, resulting in more stable and accurate drag performance. These works primarily focus on enhancing source and target feature optimization. However, when extended to object rotation manipulation, significant feature differences across views pose a challenge that pure feature alignment cannot resolve.", + "bbox": [ + 511, + 568, + 906, + 900 + ], + "page_idx": 1 + }, + { + "type": "page_number", + "text": "21503", + "bbox": [ + 478, + 944, + 517, + 955 + ], + "page_idx": 1 + }, + { + "type": "image", + "img_path": "images/0c05505f6758dfeddf97b8869d416d05b027f8d35278371b092d2e3c2d852e77.jpg", + "image_caption": [ + "Figure 3. Dragin3D Architecture. We track the representation points by calculating the transfer matrix and plot circle and square Gaussian maps centered on two representation points, respectively. We encode the image to be edited to extract the image features, which are injected into the original encoded Gaussian map features to obtain new Gaussian map features. The original Gaussian map features are added to the newly injected Gaussian map features as an input control condition for ControlNet." + ], + "image_footnote": [], + "bbox": [ + 117, + 89, + 880, + 314 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "2.2. Trajectory-based controllable image editing", + "text_level": 1, + "bbox": [ + 89, + 406, + 464, + 422 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Recently, trajectory-based motion control [10, 25, 35, 39, 40, 43, 45, 49] has emerged as an efficient and user-friendly method for controllable video generation [5, 12, 15, 48], with drag editing serving as a trajectory-based task. The gradual movement of handle points forms a continuous trajectory, providing a more precise guidance signal in contrast to defined start and end points. [45] introduces a unified control framework that integrates text, image, and trajectory data, utilizing a trajectory sampler to extract trajectories from open-domain video streams. [43] extracts features from the diffusion model as entity representations, and incorporates them into a denoised 3D U-Net with a noisy 2D Gaussian representation. [10] focuses on specific camera motion, generating videos by adding temporal constraints to a T2I model to capture motion dynamics, leveraging eight independent LoRA models for basic camera control. [40] differentiates between camera and object motions, enabling independent or combined control of the camera pose matrix and object trajectory to produce personalized videos. However, these motions are limited to 2D translations, making it difficult to generate novel views of objects in 3D space.", + "bbox": [ + 91, + 429, + 483, + 748 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "2.3. 3D generation Models", + "text_level": 1, + "bbox": [ + 89, + 756, + 294, + 772 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Based on the impressive generative effect of diffusion models on 2D images, the approach of 3D generation via diffusion models has emerged [17, 20-22, 31, 34, 44, 50]. [31] is one of the pioneering methods, introduces the SDS loss function, which enables the updating of NeRF [26] weights by minimizing the discrepancy between noise levels. Several 3D image generation methods using enhanced SDS loss functions have been proposed, employing a 2D", + "bbox": [ + 89, + 779, + 483, + 902 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "generative model to create 3D structures from textual descriptions without relying on 3D data, thus overcoming the challenge of limited 3D data availability. [44] integrates the diffusion process of image rendering and synthesis using NeRF, ensuring that the generated images remain semantically aligned with the text and geometrically consistent with the depth map during training. [38, 46] directly embed camera parameters as control conditions in the model to realize the camera pose control. However, the direct embedding of discrete values, such as camera parameters, as control conditions often lack the necessary precision.", + "bbox": [ + 511, + 407, + 908, + 575 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3. Methods", + "text_level": 1, + "bbox": [ + 513, + 590, + 612, + 606 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3.1. Task Definition & Preliminaries", + "text_level": 1, + "bbox": [ + 511, + 617, + 795, + 632 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3D Drag Editing Definition. We observed that previous drag tasks lacked a unified framework for moving objects in 3D space by drag interaction, for example, object rotation, so we propose to extend the task to 3D space. This task marks a significant advancement beyond merely editing the appearance of objects in a 2D image. Given a reference object image, our approach allows users to obtain a new perspective of the object through a simple drag interaction.", + "bbox": [ + 511, + 642, + 906, + 763 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Preliminary of Stable Video Diffusion. Our paradigm is based on Stable Video Diffusion (SVD) [2] due to its robust predictive capabilities for dynamically moving objects. The key modification from Stable Diffusion (SD) [3] is the integration of a temporal layer into the denoising model's U-Net architecture, resulting in a 3D-UNet. Compared to SD trained on image datasets, video frames capture more dynamic information about objects, leading to more sensitivity to morphological changes of the same objects, ensur", + "bbox": [ + 511, + 763, + 908, + 902 + ], + "page_idx": 2 + }, + { + "type": "page_number", + "text": "21504", + "bbox": [ + 478, + 944, + 519, + 957 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "ing perspective and temporal consistency of rotation.", + "bbox": [ + 89, + 90, + 439, + 104 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Preliminary of ControlNet. ControlNet [47] is one of the representative works in the field of controlled image generation, which supports a variety of inputs such as edge maps, segmentation maps, keypoints, etc., as conditional inputs to guide the image generation process. $\\mathcal{F}(\\cdot ;\\Theta)$ is a neural network block with parameters $\\Theta$ . ControlNet locks all parameters in $\\Theta$ and clones them into a trainable copy $\\Theta_{c}$ . The output of ControlNet $y_{c}$ can be expressed as:", + "bbox": [ + 89, + 106, + 483, + 227 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\boldsymbol {y} _ {\\mathbf {c}} = \\mathcal {F} (\\boldsymbol {x}; \\Theta) + \\mathcal {Z} \\left(\\mathcal {F} \\left(\\boldsymbol {x} + \\mathcal {Z} \\left(\\boldsymbol {c}; \\Theta_ {\\mathrm {z} 1}\\right); \\Theta_ {\\mathrm {c}}\\right); \\Theta_ {\\mathrm {z} 2}\\right), \\tag {1}\n$$\n", + "text_format": "latex", + "bbox": [ + 99, + 253, + 482, + 271 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "where $c$ is the condition input, $\\mathcal{Z}(\\cdot ;\\cdot)$ is a $1\\times 1$ convolutional layer initialized with zero weights and bias. When ControlNet is used in Stable Diffusion, the training objective is:", + "bbox": [ + 89, + 281, + 483, + 328 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {L} = \\mathbb {E} _ {\\boldsymbol {z} _ {0}, t, \\boldsymbol {c} _ {t}, \\boldsymbol {c} _ {\\mathrm {f}}, \\epsilon \\sim \\mathcal {N} (0, 1)} \\left[ \\| \\epsilon - \\epsilon_ {\\theta} \\left(z _ {t}, t, \\boldsymbol {c} _ {t}, \\boldsymbol {c} _ {\\mathrm {f}}\\right)) \\| _ {2} ^ {2} \\right], \\tag {2}\n$$\n", + "text_format": "latex", + "bbox": [ + 104, + 353, + 482, + 372 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "where $z_0$ denotes the input image, $z_t$ denotes the noisy image in timestep $t$ , $c_t$ and $c_f$ denote text prompts and the task-specific conditions, respectively.", + "bbox": [ + 89, + 382, + 483, + 426 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Preliminary of Gaussian heatmap. Gaussian heatmaps are commonly used for keypoint detection tasks. A heatmap is a feature map of the same size as the input, with a value of 1 at a specified coordinate (i.e., the representation points in our task) and 0 at all other positions. Additionally, to align with inputs like mask or depth maps used in ControlNet, a Gaussian heatmap is a suitable input, which effectively converts discrete point signals into continuous image signals.", + "bbox": [ + 89, + 428, + 483, + 547 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "3.2. Overall Framework", + "text_level": 1, + "bbox": [ + 89, + 558, + 279, + 571 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "For this novel dragging in 3D space task, we propose a framework named Dragin3D, which incorporates two key innovations: 3D drag trajectory representation and control condition injection. The overall workflow of our model is illustrated in Fig. 3. We employ an architecture that combines SVD and ControlNet. Firstly, we propose to describe the motion trajectory of the object with two points, and compute to track the positions of the representation points in different views. Next, we construct Gaussian maps of different shapes centered on these two points. These maps are then encoded, and image feature representations are injected into the Gaussian maps to align them with the corresponding point locations. Finally, the encoded Gaussian map features are linked to new Gaussian map features that incorporate object-specific information, which are used as control condition inputs for ControlNet.", + "bbox": [ + 88, + 580, + 482, + 823 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "3.3. 3D Drag Trajectory Representation (DTR)", + "text_level": 1, + "bbox": [ + 89, + 830, + 452, + 848 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "To extend the editing task to 3D space, we propose a 3D Drag Trajectory Representation (DTR) method, which explicitly represents the object's trajectory using a rotational", + "bbox": [ + 89, + 854, + 483, + 901 + ], + "page_idx": 3 + }, + { + "type": "image", + "img_path": "images/0d4c1ccf8db16bad9ec345f478fd54deda6c06588911a99a6d8955dcfeaa3f0e.jpg", + "image_caption": [ + "#" + ], + "image_footnote": [], + "bbox": [ + 535, + 88, + 612, + 155 + ], + "page_idx": 3 + }, + { + "type": "image", + "img_path": "images/7f33791087c9f9c7b1c483af2abc33c60a6ac27b798a947b6cca073d3314783d.jpg", + "image_caption": [ + "#", + "Figure 4. Our 3D drag trajectory representation method. The computed tracking points are projected into Gaussian maps, ensuring the rotational trajectory of the object is visually aligned with the trajectory of the line segment formed by the representation points." + ], + "image_footnote": [], + "bbox": [ + 633, + 89, + 705, + 155 + ], + "page_idx": 3 + }, + { + "type": "image", + "img_path": "images/104c231831aee4d0864d1d1bd7bb9e5ce9c25f1fb0e8078ca19af6ec8996d5a3.jpg", + "image_caption": [ + "#" + ], + "image_footnote": [], + "bbox": [ + 725, + 89, + 800, + 155 + ], + "page_idx": 3 + }, + { + "type": "image", + "img_path": "images/494d1e506e3b8aefecb56841ab7e6c9231e4718119908cd6eac2e54b3308e229.jpg", + "image_caption": [ + "#" + ], + "image_footnote": [], + "bbox": [ + 823, + 89, + 900, + 155 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "line segment defined by two points. In general, an object's rotation in 3D space involves variations around three coordinate axes across three planes. Since left-right rotation is more intuitive and commonly used at the operational level, our approach and experiments focus on the z-axis rotation.", + "bbox": [ + 511, + 261, + 905, + 335 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "The selection and stable tracking of the representation points are critical, as they directly affect the accuracy of the subsequent results. To ensure that the motion trajectory of the line segment is coherent and clearly defined, we propose a method for calculating the position of a specified coordinate point in images captured from different views. Then we project the computed points onto the 2D plane, as shown in Fig. 4. The required camera intrinsic and extrinsic parameters are provided by the 3D dataset. Extrinsic parameter calibration is performed by solving the transfer matrix using the real-world coordinates and pixel coordinates of fixed points. Conversely, when the camera extrinsic parameter is available, we can calculate the pixel coordinates of a fixed point by maintaining the constancy of its world coordinates. To simplify the calculation of 3D coordinates, we assume that the z-axis coordinate of each pixel point is fixed to 1. While this assumption introduces some deviation, the resulting calculation provides a reasonable approximation of the trajectory and magnitude of the object's movement. The relevant calculation formulas are as follows.", + "bbox": [ + 511, + 337, + 906, + 637 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\nx _ {c 1} = \\left(x _ {1} - c _ {x}\\right) * z _ {c 1} / f _ {x}, y _ {c 1} = \\left(y _ {1} - c _ {y}\\right) * z _ {c 1} / f _ {y}, \\tag {3}\n$$\n", + "text_format": "latex", + "bbox": [ + 522, + 664, + 903, + 681 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\begin{array}{l} \\left[ \\begin{array}{l l l} x _ {w} & y _ {w} & z _ {w} \\end{array} \\right] ^ {- 1} = T _ {1} ^ {- 1} \\cdot \\left[ \\begin{array}{l l l} x _ {c 1} & y _ {c 1} & z _ {c 1} \\end{array} \\right] ^ {- 1}, \\\\ \\left[ \\begin{array}{c} x _ {c 2} y _ {c 2} z _ {c 2} 1 \\end{array} \\right] ^ {- 1} = T _ {2} \\cdot \\left[ \\begin{array}{c} x _ {w} y _ {w} z _ {w} 1 \\end{array} \\right] ^ {- 1}, \\\\ \\end{array}\n$$\n", + "text_format": "latex", + "bbox": [ + 532, + 713, + 867, + 757 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\nx _ {2} = f _ {x} * x _ {c 2} / z _ {c 2} + c _ {x}, y _ {2} = f _ {y} * y _ {c 2} / z _ {c 2} + c _ {y}, \\tag {5}\n$$\n", + "text_format": "latex", + "bbox": [ + 535, + 782, + 903, + 799 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "where the intrinsic $(f_{x},f_{y},c_{x},c_{y})$ and extrinsic $(T_{1},T_{2})$ of the camera are labeled in the dataset. We define $P_{1}$ $(x_{1},y_{1})$ as the pixel coordinates of the specified point in the first image, $P_{c1}$ $(x_{c1},y_{c1},z_{c1})$ is the camera coordinates of the specified point in the first image, $P_{w}$ $(x_{w},y_{w},z_{w})$ is the world coordinates, $P_{c2}$ $(x_{c2},y_{c2},z_{c2})$ and $P_{2}$ $(x_{2},y_{2})$ are", + "bbox": [ + 511, + 809, + 906, + 901 + ], + "page_idx": 3 + }, + { + "type": "page_number", + "text": "21505", + "bbox": [ + 478, + 944, + 517, + 955 + ], + "page_idx": 3 + }, + { + "type": "image", + "img_path": "images/acfcb0464ad9529309c26ef4376e3998770dcf4a098a096aac88af7eda717cb1.jpg", + "image_caption": [ + "Figure 5. Dragin3D Visualization. Given an input image and a set of trajectory point pairs, we obtained Gaussian maps of the trajectories as input (first column), which is then used to generate the corresponding objects at the specified trajectory points (subsequent columns). The arrows indicate rotation directions, the red line is the current view of the object, the gray dotted line shows the trajectory that has already been traversed, and the gray solid line is the future trajectory." + ], + "image_footnote": [], + "bbox": [ + 151, + 89, + 851, + 321 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "the camera coordinates and pixel coordinates of the specified point in the second image, respectively.", + "bbox": [ + 89, + 414, + 480, + 445 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "3.4. Control Condition Inject Module (CCIM)", + "text_level": 1, + "bbox": [ + 89, + 454, + 447, + 470 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "In order to process the existing input image and the computed trajectory representations into suitable input conditions for the model, we proposed a Control Condition Inject Module (CCIM). Drawing inspiration from DragAnything [43], which uses discrete pixel points as input control conditions by mapping them into a Gaussian map. We propose to use two representation points. However, we observed that the two representation points of the object are nearly identical in the front and rear views, as well as in the symmetric left and right side views. We propose a Shape DistinGuish Strategy (SDGS) to distinguish the two points by their respective shapes. Specifically, we create circular and square Gaussian maps centered on the two points, respectively, as shown in Fig. 9 (green boxes). Then the Gaussian map images $x \\in \\mathbb{R}^{F \\times H \\times W \\times 3}$ are encoded into image features $f_{gi} \\in \\mathbb{R}^{F \\times H \\times W \\times C}$ , where $F$ denotes the input frames and $C$ denotes the channel size.", + "bbox": [ + 91, + 478, + 482, + 733 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "CCIM. Since multi-view image generation requires the guidance of input object representation information, the Gaussian map that only indicates the location and trajectory of the object is insufficient. To establish a correspondence between the location of the representation points and the image features, we propose to use a trainable image prompt mapping network, which allows the image to function as an image prompt similar to the textual guidance condition. In order to ensure consistency between training and inference, we use only the first frame image $x_{1}$ during training. The first frame image embedding is extracted by the pre", + "bbox": [ + 89, + 734, + 482, + 901 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "trained CLIP [32] image encoder, and our image prompt mapping network projects this CLIP embedding into an image prompt feature $f_{ip} \\in \\mathbb{R}^{1 \\times C}$ . Then we replace the features at the representation points' locations in the Gaussian map image features $f_{gi}$ with this sequence of image prompt features. The resulting new Gaussian map image features are denoted as $f_{ngi} \\in \\mathbb{R}^{F \\times H \\times W \\times C}$ . It is worth noting that since the start and end frames are randomly selected during training, all images in the training set with different views have the potential to be encoded as image prompts. This ensures diversity and richness in the model's ability to learn from varying perspectives.", + "bbox": [ + 511, + 414, + 906, + 595 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Through the above analysis, the input condition of ControlNet is the sum of $f_{gi}$ and $f_{ngi}$ , that is $f_{input} = f_{gi} + f_{ngi}$ . During training, we only optimize parameters of the image prompt mapping network and ControlNet, and the final optimization objective is:", + "bbox": [ + 511, + 595, + 906, + 672 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\min _ {\\theta} E _ {x _ {0}, \\epsilon , t} \\| \\epsilon - \\epsilon_ {\\theta} (x _ {t}, t, c (G, G _ {n e w})) \\| ^ {2}, \\tag {6}\n$$\n", + "text_format": "latex", + "bbox": [ + 566, + 688, + 906, + 707 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "where $\\epsilon \\sim \\mathcal{N}(0,I)$ , $t\\in [0,T]$ denotes the time step of diffusion process, $c(\\cdot)$ is the condition mechanism with explicit controls. $G$ and $G_{new}$ denote the initial and new Gaussian map injected with image features, respectively.", + "bbox": [ + 511, + 713, + 906, + 776 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "4. Experiments", + "text_level": 1, + "bbox": [ + 511, + 790, + 648, + 808 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "4.1. Experimental Setup", + "text_level": 1, + "bbox": [ + 511, + 816, + 702, + 832 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Datasets. We utilize the widely adopted OmniObject3D dataset [42], containing approximately 6k 3D object models in real-world across more than 200 categories. We selected 2.4k 3D objects for training and 200 for testing. To", + "bbox": [ + 511, + 839, + 906, + 902 + ], + "page_idx": 4 + }, + { + "type": "page_number", + "text": "21506", + "bbox": [ + 478, + 944, + 519, + 957 + ], + "page_idx": 4 + }, + { + "type": "image", + "img_path": "images/ed21d843db3387f12549fc69d3736aa8f519fc02b25d89a252b3ce517a7dc5ad.jpg", + "image_caption": [ + "Figure 6. Comparison with existing methods. We compare existing methods relevant to our task. For 3D dataset images with pose labels, we directly compute the representation points, with the first row representing the ground truth image and its corresponding rotation. For real images, we utilize the interaction interface to define the desired drag target points (\"Our drag points\" in the figure). Additionally, we specify a drag trajectory for DragAnything that closely resembles our trajectory. DragAnything interprets the trajectory as the movement of the object or camera along it but cannot handle rotations. SV3D struggles with generalization to real-world images and lacks sufficient detail and 3D consistency. Wonder3D generates higher-quality 3D images but is limited to six fixed views." + ], + "image_footnote": [], + "bbox": [ + 109, + 85, + 372, + 344 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/ef6f849b22b045db15210aa77528e21409ec34fc9c849efbf1a77354d023229c.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 383, + 85, + 632, + 343 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/94053b6f4b9348ff728fb0f35ca8673251199aab5fc1cc9161cd539fb8f349d7.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 635, + 85, + 888, + 343 + ], + "page_idx": 5 + }, + { + "type": "table", + "img_path": "images/076828d6b42cf57dedc1590424cb8eecb31d8d1590f87a92651d2b09dbf93134.jpg", + "table_caption": [], + "table_footnote": [], + "table_body": "
MethodsMulti-viewBackgroundTime
DragAnything×✓~40s
SV3D✓×~2min
Wonder3D✓×~2min
Ours✓✓~25s
", + "bbox": [ + 102, + 460, + 470, + 554 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Table 1. Functionality comparisons. Our proposed Dragin3D achieves multi-view generation and supports editing images with backgrounds, ensuring efficient and rapid editing with minimal time consumption. In contrast, methods from different areas are unable to perform these functions simultaneously.", + "bbox": [ + 89, + 564, + 482, + 636 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "better simulate real-world scenes, we incorporate diverse background images into the dataset. Additionally, the image resolution is resized to $256*256$ to optimize computational efficiency during training.", + "bbox": [ + 89, + 667, + 482, + 728 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Baselines. Since Dragin3D integrates drag with multiview tasks, we use DragAnything [43], SV3D [38], and Wonder3D [22] as baselines. DragAnything and SV3D are trained on video framework, Wonder3D is trained on image framework. DragAnything is a 2D interaction editing method trained on video datasets, SV3D and Wonder3D are 3D generation methods trained on multi-view datasets.", + "bbox": [ + 89, + 731, + 482, + 837 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Evaluation Metrics. For multi-view generation evaluation, we use the following three metrics for quantitative evaluation: For the quality of generated images, we employ Fréchet Inception Distance (FID), which calculates the", + "bbox": [ + 89, + 839, + 483, + 901 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Frechet distance between the feature vectors of the generated image and ground truth. For the perceptual similarity of the 3D effects, we use Learned Perceptual Image Patch Similarity (LPIPS), which evaluates the perceptual similarity between the generated image and the ground truth by learning the inverse mapping of the generated image. For the quality of the generated motion sequences, we apply Frechet Video Distance (FVD), which leverages a 3D feature extractor to compare video sequences. For drag trajectory accuracy evaluation, we adopt the MD metric in DragDiffusion, which employs DIFT [37] to identify the points in the images dragged by specified trajectories, corresponding to the starting representation points. The average Euclidean distance between the identified points and the reference trajectory points is then computed.", + "bbox": [ + 511, + 464, + 906, + 691 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Implements Details. We randomly select 8 coherent object rotation images from the OmniObject3D dataset, which are used as the individual frames of video input to the SVD architecture. Eight A800 GPUs are used to train 20k steps with a batch size of 6, the learning rate is set to 2e-5.", + "bbox": [ + 511, + 693, + 906, + 768 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "4.2. Qualitative Results", + "text_level": 1, + "bbox": [ + 511, + 784, + 694, + 800 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "The visualization results of Dragin3D are shown in Fig. 5. Given an image with a corresponding trajectory, Dragin3D generates a high-fidelity image of the object from a new perspective, maintaining strong consistency with the original image. In addition, Dragin3D demonstrates superior generalization performance, which successfully rotates ob", + "bbox": [ + 511, + 809, + 906, + 902 + ], + "page_idx": 5 + }, + { + "type": "page_number", + "text": "21507", + "bbox": [ + 478, + 944, + 517, + 955 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/1f0343b0fc7610f981d1a355897bd884ff149e2a267a918b014d68aa57fb61f5.jpg", + "image_caption": [ + "Figure 7. Comparison with 3D methods. We use the image without background as input to 3D methods and compare the results with ground truth, demonstrating that our method is comparable." + ], + "image_footnote": [], + "bbox": [ + 96, + 87, + 478, + 256 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "jects by dragging the representation points on real images that lie outside the training dataset distribution.", + "bbox": [ + 88, + 338, + 482, + 367 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "As shown in Fig. 7, we input the image into SV3D and Wonder3D after background removal and compare the results with the ground truth from the corresponding viewpoints. Our methods demonstrate comparable performance in terms of rotation accuracy and 3D consistency. Since our main contribution lies in the generalizability to real-world images, we include the background to reduce the gap between the ideal 3D dataset and the real-world scene. As illustrated in Fig. 6, we present the visualization results of each method with background using both 3D datasets and real images. The real images we provided have homogeneous backgrounds, minimizing their impact on the object. The results show that the performance of the other methods is significantly lower, limiting their practical applicability, while our method demonstrates a one-stage process that eliminates background interference during the image processing stage, performs well in detail preservation and coherent rotation. Additionally, we found that DragDiffusion [36] can achieve partial rotation in certain cases of Drag-Bench. However, the rotation is minimal and random, lacking precise control over the direction and angle of rotation as specified by the instructions, as illustrated in Fig. 8.", + "bbox": [ + 91, + 369, + 483, + 700 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "4.3. Quantitative Results", + "text_level": 1, + "bbox": [ + 89, + 710, + 284, + 726 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Since our proposed task diverges from traditional drag-based editing and produces significantly different outcomes, and 3D generation models like Wonder3D focus on generating images without background from partial views, we do not include them in a quantitative comparison. Our Dragin3D shares a similar goal with SV3D, which aims to generate images from new perspectives based on a single input image. Therefore, we chose to make a comparison with SV3D. Due to the absence of ground truth in the real-world data, we randomly selected 200 subsets from the OmniObject3D dataset as the testset for computation, which were", + "bbox": [ + 88, + 734, + 482, + 900 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/861afd2f4b5a8d5651b02fab108b6df5ca28e4ae4d7a86e3527af9502c7e2315.jpg", + "image_caption": [ + "Figure 8. Comparison of rotation cases with DragDiffusion. DragDiffusion primarily manifests as tiny rotations induced by movements in the plane towards the target point, while Dragin3D enables more accurate control of stereoscopic rotation." + ], + "image_footnote": [], + "bbox": [ + 531, + 87, + 890, + 207 + ], + "page_idx": 6 + }, + { + "type": "table", + "img_path": "images/2c72541890901a89f435718cb217fe150252e3803f4c81005deb3ae39cf0cc0b.jpg", + "table_caption": [], + "table_footnote": [], + "table_body": "
MethodsFID↓LPIPS↓FVD↓
SV3D [38]52.50.410944.6
Dragin3D (Ours)43.40.179584.6
", + "bbox": [ + 553, + 287, + 864, + 351 + ], + "page_idx": 6 + }, + { + "type": "table", + "img_path": "images/0f6510974506bd3733a2227833ce27661006a1499174bab768fb1110b486856d.jpg", + "table_caption": [ + "Table 2. Quantitative comparisons. Our proposed Dragin3D outperforms SV3D in all metrics." + ], + "table_footnote": [], + "table_body": "
MethodsDragDiffusion [36]Dragin3D (Ours)
MD↓33.019.38
", + "bbox": [ + 535, + 404, + 883, + 452 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Table 3. Quantitative comparisons. Our proposed Dragin3D outperforms DragDiffusion in control accuracy evaluation.", + "bbox": [ + 511, + 460, + 903, + 489 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "not visible during training. We compared the similarity between the generated images and ground truth across the corresponding viewpoints from three perspectives: image, 3D perceptual, and video. As shown in Tab. 2, Dragin3D outperforms SV3D in all metrics. To evaluate the accuracy of trajectory control, we selected object rotation cases in Drag-Bench including animal, human, and sculpture heads. For each set of desired rotation trajectories, we computed the MDs for both Dragdiffusion and Dragin3D, with Dragin3D outperforming Dragdiffusion, as shown in Tab. 3. Additionally, we conducted a user study in Supp. 7.2.", + "bbox": [ + 511, + 517, + 906, + 684 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "4.4. Ablations", + "text_level": 1, + "bbox": [ + 511, + 695, + 624, + 709 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Effectiveness of shape distinguish strategy (SDGS). To demonstrate the effectiveness of SDGS, we designed experiments in which the Gaussian maps for two points were constrained to the same shape, as shown in Tab. 4. We specifically focused on the symmetric rotation of objects, as illustrated in Fig. 9. Without distinguishing between the two points, the object in the subsequent stage of the rotation becomes difficult to distinguish from its symmetrical configuration in the preliminary stage. In contrast, the generated object preserves structural integrity and details when SDGS is applied. More examples can be found in Supp. 7.3.", + "bbox": [ + 511, + 718, + 906, + 883 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Effectiveness of control condition inject module", + "bbox": [ + 532, + 885, + 903, + 900 + ], + "page_idx": 6 + }, + { + "type": "page_number", + "text": "21508", + "bbox": [ + 478, + 944, + 517, + 955 + ], + "page_idx": 6 + }, + { + "type": "table", + "img_path": "images/a9acd42514dc7774dc047bf7d54083249a8ea71e7bf6b9e56af627876c119dcc.jpg", + "table_caption": [], + "table_footnote": [], + "table_body": "
MethodsFID↓LPIPS↓FVD↓
Dragin3D w/o SDGS44.10.182676.8
Dragin3D w/o CCIM45.20.194694.1
Dragin3D43.40.179584.6
", + "bbox": [ + 117, + 88, + 455, + 165 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Table 4. Ablation study results. All metrics have declined without SDGS and CCIM, respectively.", + "bbox": [ + 89, + 176, + 482, + 203 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/31c4543317ea4d6d144fbba6b8bc5abfcdf2a0fbceea39a314241bde2c7daa47.jpg", + "image_caption": [ + "Figure 9. Effectiveness of shape distinguish strategy (SDGS). We present the resulting images for two cases. The first row displays the trajectory points projected onto Gaussian maps, while the second and third rows show the objects from the corresponding view of these trajectory points, generated with and without SDGS, respectively. The positions of the two trajectory points in the first frame closely correspond to those in the last frame (green box). In the latter half of the sequence, significant misalignment occurs due to interference from the Gaussian representation (brown box)." + ], + "image_footnote": [], + "bbox": [ + 101, + 220, + 472, + 479 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "(CCIM). As illustrated in Tab. 4, to assess the effectiveness of CCIM, we conducted experiments using only Gaussian map image features $f_{gi}$ derived from the tracking points as the control input. The results reveal a marked enhancement across all evaluation metrics upon the integration of CCIM. Visual comparisons of intermediate results are presented in Fig. 10. Without CCIM, the object's rotation initially exhibits low-amplitude fluctuations, followed by a sudden and drastic change. Specifically, the early frames retain the original image, while later frames generate objects with significant distortion and poor fidelity. In contrast, the rotation transition is smooth and stable, and the object's details are preserved throughout. These results demonstrate that CCIM effectively aligns the rotation trajectory with image features by incorporating object-specific information, ensuring that the generated perspectives remain faithful to the original image content. More examples can be found in Supp. 7.3.", + "bbox": [ + 89, + 643, + 482, + 900 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/bd5319896f6c8f5d3d35247090c8b82f3ce1f284acefe9341986e29df39a9ed7.jpg", + "image_caption": [ + "Figure 10. Effectiveness of control condition inject module (CCIM). We present the resulting images for two cases. The first row displays the trajectory points projected onto Gaussian maps, while the second and third rows show the objects from the corresponding view of these trajectory points, generated with and without CCIM, respectively. Without CCIM, the image remains largely unchanged in the early stages (red box), but later undergoes a sudden distortion, causing significant fidelity loss (yellow box)." + ], + "image_footnote": [], + "bbox": [ + 526, + 90, + 893, + 420 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "5. Conclusion and Future Works", + "text_level": 1, + "bbox": [ + 511, + 571, + 790, + 585 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "In this paper, we present Dragin3D, which extends the interactive trajectory-based dragging task to 3D space. To represent the rotation trajectory of an object from the starting view to the terminating view, we introduce a double-point representation approach, and map the rotation of the object to a line segment formed by these two points. Our approach introduces control conditions through ControlNet architecture based on the large-scale video generation pretrained model SVD. In addition, we randomly add different backgrounds to the 3D object dataset, enabling the model to directly handle images with backgrounds. Through extensive experiments, Dragin3D successfully generates novel view images of the object based on the trajectories of two points, minimizing background interference while preserving the original object features, and demonstrating exceptional generalization on real-world images. For future work, since our approach currently supports object rotation in only one plane, we plan to expand our trajectory-based editing task to encompass a broader range of views, enabling arbitrary rotations in 3D space.", + "bbox": [ + 511, + 598, + 906, + 900 + ], + "page_idx": 7 + }, + { + "type": "page_number", + "text": "21509", + "bbox": [ + 478, + 944, + 517, + 955 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "6. Acknowledgments", + "text_level": 1, + "bbox": [ + 89, + 90, + 269, + 107 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "This research is supported by Artificial Intelligence-National Science and Technology Major Project 2023ZD0121200, the National Science Fund for Excellent Young Scholars under Grant 62222212, and National Natural Science Foundation of China under Grant 623B2094.", + "bbox": [ + 89, + 114, + 483, + 205 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "References", + "text_level": 1, + "bbox": [ + 91, + 220, + 187, + 236 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[1] Omer Bar-Tal, Dolev Ofri-Amar, Rafail Fridman, Yoni Kasten, and Tali Dekel. Text2live: Text-driven layered image and video editing. In European conference on computer vision, pages 707-723. Springer, 2022. 2", + "[2] Andreas Blattmann, Tim Dockhorn, Sumith Kulal, Daniel Mendelevitch, Maciej Kilian, Dominik Lorenz, Yam Levi, Zion English, Vikram Voleti, Adam Letts, et al. Stable video diffusion: Scaling latent video diffusion models to large datasets. arXiv preprint arXiv:2311.15127, 2023. 3", + "[3] Andreas Blattmann, Robin Rombach, Huan Ling, Tim Dockhorn, Seung Wook Kim, Sanja Fidler, and Karsten Kreis. Align your latents: High-resolution video synthesis with latent diffusion models. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 22563-22575, 2023. 3", + "[4] Tim Brooks, Aleksander Holynski, and Alexei A Efros. Instructpix2pix: Learning to follow image editing instructions. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 18392-18402, 2023. 2", + "[5] Weifeng Chen, Yatai Ji, Jie Wu, Hefeng Wu, Pan Xie, Jiashi Li, Xin Xia, Xuefeng Xiao, and Liang Lin. Control-a-video: Controllable text-to-video generation with diffusion models. arXiv preprint arXiv:2305.13840, 2023. 3", + "[6] Guillaume Couairon, Jakob Verbeek, Holger Schwenk, and Matthieu Cord. Diffedit: Diffusion-based semantic image editing with mask guidance. arXiv preprint arXiv:2210.11427, 2022. 2", + "[7] Yutao Cui, Xiaotong Zhao, Guozhen Zhang, Shengming Cao, Kai Ma, and Limin Wang. Stabledrag: Stable dragging for point-based image editing. arXiv preprint arXiv:2403.04437, 2024.1, 2", + "[8] Yuki Endo. User-controllable latent transformer for stylegan image layout editing. In Computer Graphics Forum, pages 395-406. Wiley Online Library, 2022. 2", + "[9] Dave Epstein, Allan Jabri, Ben Poole, Alexei Efros, and Aleksander Holynski. Diffusion self-guidance for controllable image generation. Advances in Neural Information Processing Systems, 36:16222-16239, 2023. 2", + "[10] Yuwei Guo, Ceyuan Yang, Anyi Rao, Zhengyang Liang, Yaohui Wang, Yu Qiao, Maneesh Agrawala, Dahua Lin, and Bo Dai. Animatediff: Animate your personalized text-to-image diffusion models without specific tuning. arXiv preprint arXiv:2307.04725, 2023. 3", + "[11] Amir Hertz, Ron Mokady, Jay Tenenbaum, Kfir Aberman, Yael Pritch, and Daniel Cohen-Or. Prompt-to-prompt im" + ], + "bbox": [ + 93, + 246, + 483, + 900 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "age editing with cross attention control. arXiv preprint arXiv:2208.01626, 2022. 2", + "[12] Yaosi Hu, Chong Luo, and Zhenzhong Chen. Make it move: controllable image-to-video generation with text descriptions. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 18219-18228, 2022. 3", + "[13] Tero Karras, Samuli Laine, Miika Aittala, Janne Hellsten, Jaakko Lehtinen, and Timo Aila. Analyzing and improving the image quality of stylegan. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 8110-8119, 2020. 1", + "[14] Bahjat Kawar, Shiran Zada, Oran Lang, Omer Tov, Huiwen Chang, Tali Dekel, Inbar Mosseri, and Michal Irani. Imagc: Text-based real image editing with diffusion models. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 6007-6017, 2023. 2", + "[15] Ali Köksal, Kenan E Ak, Ying Sun, Deepu Rajan, and Joo Hwee Lim. Controllable video generation with text-based instructions. IEEE transactions on multimedia, 26: 190-201, 2023. 3", + "[16] Nupur Kumari, Bingliang Zhang, Richard Zhang, Eli Shechtman, and Jun-Yan Zhu. Multi-concept customization of text-to-image diffusion. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 1931-1941, 2023. 2", + "[17] Jeong-gi Kwak, Erqun Dong, Yuhe Jin, Hanseok Ko, Shweta Mahajan, and Kwang Moo Yi. Vivid-1-to-3: Novel view synthesis with video diffusion models. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 6775–6785, 2024. 3", + "[18] Dongxu Li, Junnan Li, and Steven Hoi. Blip-diffusion: Pretrained subject representation for controllable text-to-image generation and editing. Advances in Neural Information Processing Systems, 36, 2024. 2", + "[19] Pengyang Ling, Lin Chen, Pan Zhang, Huaian Chen, Yi Jin, and Jinjin Zheng. Freedrag: Feature dragging for reliable point-based image editing. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 6860-6870, 2024. 1, 2", + "[20] Minghua Liu, Chao Xu, Haian Jin, Linghao Chen, Mukund Varma T, Zexiang Xu, and Hao Su. One-2-3-45: Any single image to 3d mesh in 45 seconds without per-shape optimization. Advances in Neural Information Processing Systems, 36, 2024. 3", + "[21] Yuan Liu, Cheng Lin, Zijiao Zeng, Xiaoxiao Long, Lingjie Liu, Taku Komura, and Wenping Wang. Syncdreamer: Generating multiview-consistent images from a single-view image. arXiv preprint arXiv:2309.03453, 2023.", + "[22] Xiaoxiao Long, Yuan-Chen Guo, Cheng Lin, Yuan Liu, Zhiyang Dou, Lingjie Liu, Yuexin Ma, Song-Hai Zhang, Marc Habermann, Christian Theobalt, et al. Wonder3d: Single image to 3d using cross-domain diffusion. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 9970-9980, 2024. 3, 6", + "[23] Jingyi Lu, Xinghui Li, and Kai Han. Regiondrag: Fast region-based image editing with diffusion models. arXiv preprint arXiv:2407.18247, 2024. 2" + ], + "bbox": [ + 516, + 92, + 906, + 900 + ], + "page_idx": 8 + }, + { + "type": "page_number", + "text": "21510", + "bbox": [ + 478, + 944, + 519, + 955 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[24] Minxing Luo, Wentao Cheng, and Jian Yang. Rotationdrag: Point-based image editing with rotated diffusion features. arXiv preprint arXiv:2401.06442, 2024. 2", + "[25] Wan-Duo Kurt Ma, John P Lewis, and W Bastiaan Kleijn. Trailblazer: Trajectory control for diffusion-based video generation. arXiv preprint arXiv:2401.00896, 2023. 3", + "[26] Ben Mildenhall, Pratul P Srinivasan, Matthew Tancik, Jonathan T Barron, Ravi Ramamoorthi, and Ren Ng. Nerf: Representing scenes as neural radiance fields for view synthesis. Communications of the ACM, 65(1):99-106, 2021. 3", + "[27] Chong Mou, Xintao Wang, Jiechong Song, Ying Shan, and Jian Zhang. Dragondiffusion: Enabling drag-style manipulation on diffusion models. arXiv preprint arXiv:2307.02421, 2023. 1, 2", + "[28] Chong Mou, Xintao Wang, Jiechong Song, Ying Shan, and Jian Zhang. Diffeditor: Boosting accuracy and flexibility on diffusion-based image editing. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 8488-8497, 2024.", + "[29] Shen Nie, Hanzhong Allan Guo, Cheng Lu, Yuhao Zhou, Chenyu Zheng, and Chongxuan Li. The blessing of randomness: Sde beats ode in general diffusion-based image editing. arXiv preprint arXiv:2311.01410, 2023.", + "[30] Xingang Pan, Ayush Tewari, Thomas Leimkuhler, Lingjie Liu, Abhimitra Meka, and Christian Theobalt. Drag your gan: Interactive point-based manipulation on the generative image manifold. In ACM SIGGRAPH 2023 Conference Proceedings, pages 1-11, 2023. 1, 2", + "[31] Ben Poole, Ajay Jain, Jonathan T Barron, and Ben Mildenhall. Dreamfusion: Text-to-3d using 2d diffusion. arXiv preprint arXiv:2209.14988, 2022. 3", + "[32] Alec Radford, Jong Wook Kim, Chris Hallacy, Aditya Ramesh, Gabriel Goh, Sandhini Agarwal, Girish Sastry, Amanda Askell, Pamela Mishkin, Jack Clark, et al. Learning transferable visual models from natural language supervision. In International conference on machine learning, pages 8748-8763. PMLR, 2021. 5", + "[33] Nataniel Ruiz, Yuanzhen Li, Varun Jampani, Yael Pritch, Michael Rubinstein, and Kfir Aberman. Dreambooth: Fine tuning text-to-image diffusion models for subject-driven generation. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 22500-22510, 2023. 2", + "[34] Ruoxi Shi, Hansheng Chen, Zhuoyang Zhang, Minghua Liu, Chao Xu, Xinyue Wei, Linghao Chen, Chong Zeng, and Hao Su. Zero123++: a single image to consistent multi-view diffusion base model. arXiv preprint arXiv:2310.15110, 2023. 3", + "[35] Xiaoyu Shi, Zhaoyang Huang, Fu-Yun Wang, Weikang Bian, Dasong Li, Yi Zhang, Manyuan Zhang, Ka Chun Cheung, Simon See, Hongwei Qin, et al. Motion-i2v: Consistent and controllable image-to-video generation with explicit motion modeling. In ACM SIGGRAPH 2024 Conference Papers, pages 1-11, 2024. 3", + "[36] Yujun Shi, Chuhui Xue, Jun Hao Liew, Jiachun Pan, Hanshu Yan, Wenqing Zhang, Vincent YF Tan, and Song Bai." + ], + "bbox": [ + 91, + 92, + 482, + 901 + ], + "page_idx": 9 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "Dragdiffusion: Harnessing diffusion models for interactive point-based image editing. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 8839-8849, 2024. 1, 2, 7", + "[37] Luming Tang, Menglin Jia, Qianqian Wang, Cheng Perng Phoo, and Bharath Hariharan. Emergent correspondence from image diffusion. Advances in Neural Information Processing Systems, 36:1363-1389, 2023. 6", + "[38] Vikram Voleti, Chun-Han Yao, Mark Boss, Adam Letts, David Pankratz, Dmitry Tochilkin, Christian Laforte, Robin Rombach, and Varun Jampani. Sv3d: Novel multi-view synthesis and 3d generation from a single image using latent video diffusion. In European Conference on Computer Vision, pages 439-457. Springer, 2025. 3, 6, 7", + "[39] Xiang Wang, Hangjie Yuan, Shiwei Zhang, Dayou Chen, Jiuniu Wang, Yingya Zhang, Yujun Shen, Deli Zhao, and Jingren Zhou. Videocomposer: Compositional video synthesis with motion controllability. Advances in Neural Information Processing Systems, 36, 2024. 3", + "[40] Zhouxia Wang, Ziyang Yuan, Xintao Wang, Yaowei Li, Tianshui Chen, Menghan Xia, Ping Luo, and Ying Shan. Motionctrl: A unified and flexible motion controller for video generation. In ACM SIGGRAPH 2024 Conference Papers, pages 1-11, 2024. 3", + "[41] Yuxiang Wei, Yabo Zhang, Zhilong Ji, Jinfeng Bai, Lei Zhang, and Wangmeng Zuo. Elite: Encoding visual concepts into textual embeddings for customized text-to-image generation. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 15943-15953, 2023. 2", + "[42] Tong Wu, Jiarui Zhang, Xiao Fu, Yuxin Wang, Jiawei Ren, Liang Pan, Wayne Wu, Lei Yang, Jiaqi Wang, Chen Qian, et al. Omniobject3d: Large-vocabulary 3d object dataset for realistic perception, reconstruction and generation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 803-814, 2023. 5", + "[43] Weijia Wu, Zhuang Li, Yuchao Gu, Rui Zhao, Yefei He, David Junhao Zhang, Mike Zheng Shou, Yan Li, Tingting Gao, and Di Zhang. Draganything: Motion control for anything using entity representation. In European Conference on Computer Vision, pages 331-348. Springer, 2025. 1, 3, 5, 6", + "[44] Haibo Yang, Yang Chen, Yingwei Pan, Ting Yao, Zhineng Chen, and Tao Mei. 3dstyle-diffusion: Pursuing fine-grained text-driven 3d stylization with 2d diffusion models. In Proceedings of the 31st ACM International Conference on Multimedia, pages 6860-6868, 2023. 3", + "[45] Shengming Yin, Chenfei Wu, Jian Liang, Jie Shi, Houqiang Li, Gong Ming, and Nan Duan. Dragnuwa: Fine-grained control in video generation by integrating text, image, and trajectory. arXiv preprint arXiv:2308.08089, 2023. 1, 3", + "[46] Ziyang Yuan, Mingdeng Cao, Xintao Wang, Zhongang Qi, Chun Yuan, and Ying Shan. Customnet: Zero-shot object customization with variable-viewpoints in text-to-image diffusion models. arXiv preprint arXiv:2310.19784, 2023. 3", + "[47] Lvmin Zhang, Anyi Rao, and Maneesh Agrawala. Adding conditional control to text-to-image diffusion models. In" + ], + "bbox": [ + 516, + 92, + 903, + 901 + ], + "page_idx": 9 + }, + { + "type": "page_number", + "text": "21511", + "bbox": [ + 478, + 944, + 517, + 955 + ], + "page_idx": 9 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 3836-3847, 2023. 4", + "[48] Yabo Zhang, Yuxiang Wei, Dongsheng Jiang, Xiaopeng Zhang, Wangmeng Zuo, and Qi Tian. Controlvideo: Training-free controllable text-to-video generation. arXiv preprint arXiv:2305.13077, 2023. 3", + "[49] Zhenghao Zhang, Junchao Liao, Menghao Li, Long Qin, and Weizhi Wang. Tora: Trajectory-oriented diffusion transformer for video generation. arXiv preprint arXiv:2407.21705, 2024.3", + "[50] Chuanxia Zheng and Andrea Vedaldi. Free3d: Consistent novel view synthesis without 3d representation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 9720-9731, 2024. 3" + ], + "bbox": [ + 91, + 90, + 482, + 289 + ], + "page_idx": 10 + }, + { + "type": "page_number", + "text": "21512", + "bbox": [ + 478, + 945, + 517, + 955 + ], + "page_idx": 10 + } +] \ No newline at end of file diff --git a/2025/Dragin3D_ Image Editing by Dragging in 3D Space/63363afe-acf7-462c-bf4b-9a484c6ae8ae_model.json b/2025/Dragin3D_ Image Editing by Dragging in 3D Space/63363afe-acf7-462c-bf4b-9a484c6ae8ae_model.json new file mode 100644 index 0000000000000000000000000000000000000000..fb042a9024d9ad384b51a332f892264e06358255 --- /dev/null +++ b/2025/Dragin3D_ Image Editing by Dragging in 3D Space/63363afe-acf7-462c-bf4b-9a484c6ae8ae_model.json @@ -0,0 +1,2125 @@ +[ + [ + { + "type": "header", + "bbox": [ + 0.107, + 0.003, + 0.182, + 0.043 + ], + "angle": 0, + "content": "CVF" + }, + { + "type": "header", + "bbox": [ + 0.237, + 0.001, + 0.808, + 0.047 + ], + "angle": 0, + "content": "This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore." + }, + { + "type": "title", + "bbox": [ + 0.242, + 0.131, + 0.758, + 0.154 + ], + "angle": 0, + "content": "Dragin3D: Image Editing by Dragging in 3D Space" + }, + { + "type": "text", + "bbox": [ + 0.207, + 0.18, + 0.788, + 0.199 + ], + "angle": 0, + "content": "Weiran Guang\\(^{12}\\) Xiaoguang Gu\\(^{2*}\\) Mengqi Huang\\(^{1}\\) Zhendong Mao\\(^{1}\\)" + }, + { + "type": "text", + "bbox": [ + 0.307, + 0.199, + 0.692, + 0.218 + ], + "angle": 0, + "content": "1University of Science and Technology of China" + }, + { + "type": "text", + "bbox": [ + 0.165, + 0.217, + 0.83, + 0.235 + ], + "angle": 0, + "content": "\\(^{2}\\)State Key Laboratory of Communication Content Cognition, People's Daily Online" + }, + { + "type": "text", + "bbox": [ + 0.155, + 0.237, + 0.833, + 0.252 + ], + "angle": 0, + "content": "guangwr, huangmq}@mail.ustc.edu.cn, guxiaoguang@people.cn, zdmao@ustc.edu.cn" + }, + { + "type": "title", + "bbox": [ + 0.248, + 0.287, + 0.327, + 0.303 + ], + "angle": 0, + "content": "Abstract" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.319, + 0.486, + 0.668 + ], + "angle": 0, + "content": "Interactive drag editing of images is a valuable task that has gained considerable attention for its precision and controllability. However, existing approaches have primarily focused on manipulating the shape or movement of objects in 2D plane. We propose to extend this drag-based editing task to 3D space. Firstly, we utilize the trajectory of two points to represent the rotational trajectory of the object. Gaussian maps of a circle and a square are centered at these two points, respectively. We use distinct shapes to ensure that symmetric views produce different object representations. Secondly, we introduce a lightweight mapping network to embed the object features into two Gaussian maps to obtain a continuous control condition that guides the model in learning the correspondence between the trajectory and the object. Finally, to overcome the limitations of current 3D object reconstruction datasets, which typically consist of object maps with transparent backgrounds, we affix random backgrounds to them. This modification helps improve the model's ability to ignore background interference when editing real images with complex backgrounds. Experiments demonstrate that our approach successfully achieves object rotation within the drag framework and demonstrates strong generalization to real-world images." + }, + { + "type": "title", + "bbox": [ + 0.092, + 0.698, + 0.222, + 0.714 + ], + "angle": 0, + "content": "1. Introduction" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.724, + 0.485, + 0.875 + ], + "angle": 0, + "content": "Drag-based image editing has garnered substantial research interest in recent years, aiming to enable the manipulation of objects via user-defined handle and target points [30, 36]. Most existing image editing works are text-based, which heavily depends on textual and semantic alignment. Sometimes, it is challenging to articulate user-desired edits in words. Drag-based editing offers a more intuitive, interactive, and granular editing signal, allowing for precise manipulation at the pixel level. Users specify the start and end points based on intended edits, eliminating the need for a" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.288, + 0.905, + 0.334 + ], + "angle": 0, + "content": "proper prompt. This enables user-friendly image modification and offers more precise control over positioning, making it valuable for practical applications." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.338, + 0.907, + 0.685 + ], + "angle": 0, + "content": "Existing drag-based editing methods can be categorized into optimized-based and training-based. DragGAN [30] is an early representative of optimized-based methods, encompassing two key components: point tracking and motion supervision. The latent code of StyleGAN [13] is optimized while updating the position of the handle point, ensuring that the features near the target points align with those near the handle points. Since DragGAN was proposed, both drag-based image and video editing techniques have been significantly advanced. Subsequently, researchers extended this process to diffusion models by optimizing the features of UNet to break through the inherent limitation of the GAN model [27, 36]. Recent studies have focused on addressing issues related to inaccurate point tracking and incomplete motion supervision, enabling more stable dragging by formulating better point tracking and motion supervision protocols [7, 19]. Furthermore, to overcome the limitations posed by variations in object motion patterns, the training-based method [43, 45] has been introduced using trajectory guidance, where the motion trajectory of the object is encoded as a control condition. In conclusion, ongoing efforts aim to enhance editing controllability and fidelity." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.689, + 0.909, + 0.903 + ], + "angle": 0, + "content": "However, existing drag-based methods are limited to 2D drag editing and fail to achieve drag in 3D space, which significantly limits their practical application in the real 3D world. 3D generation has significantly expanded the application in games, animation, virtual production, etc. It enables more realistic creation of 3D visual spaces that closely resemble the physical world. We consider whether it is possible to generate novel view images of the object through simple point-drag interactions. The two drag paradigms discussed above are inadequate for solving this task. The optimization-based approach primarily seeks to supervise feature consistency between the handle point and the target point, which cannot well control the substantial changes in features that arise due to variations in spatial perspective, as" + }, + { + "type": "page_footnote", + "bbox": [ + 0.109, + 0.887, + 0.348, + 0.901 + ], + "angle": 0, + "content": "*Xiaoguang Gu is the Corresponding author." + }, + { + "type": "page_number", + "bbox": [ + 0.479, + 0.945, + 0.521, + 0.958 + ], + "angle": 0, + "content": "21502" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.098, + 0.089, + 0.479, + 0.221 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.09, + 0.232, + 0.483, + 0.384 + ], + "angle": 0, + "content": "Figure 1. Existing drag method for object rotation representation. (a) The optimization-based method. The sculpture makes tiny rotations, and changes in the position of the target point still tend to maintain a frontal perspective, which fails to achieve 3D editing with varying amplitudes based solely on the dragging intensity. (b) The training-based method. In the upper figure, the object moves along the trajectory of the dragging point. In the lower figure, the object rotates in the direction of the trajectory of the dragging point. The trajectory is inclined upwards, with the ship exhibiting a tendency to rotate towards the target position. Arguably, whether the ship can achieve 3D rotation is entirely random." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.412, + 0.483, + 0.472 + ], + "angle": 0, + "content": "illustrated in Fig. 1(a). On the other hand, the trajectory-based methods introduce ambiguity in the trend of object motion, as demonstrated in Fig. 1(b). Therefore, in this study, we introduce a new drag-based object rotation task." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.474, + 0.483, + 0.851 + ], + "angle": 0, + "content": "The key challenges of this task lie in how to accurately represent and track the desired 3D editing trajectory from a 2D image, and how to establish a correspondence between object motion trajectories and their respective representations. Due to the ambiguity in existing training-based methods for single-point trajectory representation, combined with the feature alignment principle of optimization-based methods, we propose Dragin3D, which extends drag-based editing to 3D space for the first time. We introduce a double-point representation approach that associates object trajectories with image features. Given an image of an object and a double-point rotation trajectory, Dragin3D generates a novel view image of the object from the corresponding trajectory, as shown in Fig. 2. The core concept of Dragin3D is to precisely capture the direction and magnitude of the object's rotation using the double-point trajectory. Specifically, our method consists of two main components: (1) Drag Trajectory Representation (DTR): we map the rotational trajectory of the object onto the rotational trajectory of a line segment, track the two endpoints of this segment, and project them into Gaussian maps with different shapes; (2) Control Condition Inject Module (CCIM): to enhance fidelity to the object, we extract image features through a lightweight mapping network and inject these features into the corresponding Gaussian maps." + }, + { + "type": "text", + "bbox": [ + 0.11, + 0.852, + 0.414, + 0.866 + ], + "angle": 0, + "content": "Our contributions are summarized as follows:" + }, + { + "type": "text", + "bbox": [ + 0.091, + 0.871, + 0.483, + 0.902 + ], + "angle": 0, + "content": "- We propose an innovative interactive image editing method Dragin3D. To the best of our knowledge, we are" + }, + { + "type": "image", + "bbox": [ + 0.52, + 0.089, + 0.905, + 0.198 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.513, + 0.207, + 0.907, + 0.264 + ], + "angle": 0, + "content": "Figure 2. Drag results of our proposed paradigm Dragin3D. Given an input image and coordinates of two representation points, i.e., \\([P_{11}, P_{21}],[P_{12}, P_{22}],\\ldots,[P_{1n}, P_{2n}]\\). Dragin3D can generate a novel view image of the object for the corresponding trajectory." + }, + { + "type": "text", + "bbox": [ + 0.526, + 0.3, + 0.905, + 0.33 + ], + "angle": 0, + "content": "the first to achieve multi-view image generation of objects by dragging them in 3D space based on a diffusion model." + }, + { + "type": "text", + "bbox": [ + 0.514, + 0.332, + 0.905, + 0.39 + ], + "angle": 0, + "content": "- We propose to represent the object rotation trajectory by a double-point motion trajectory, which maps 3D object motion onto a 2D plane and eliminates the ambiguity of the object motion mode." + }, + { + "type": "text", + "bbox": [ + 0.514, + 0.391, + 0.905, + 0.435 + ], + "angle": 0, + "content": "- We introduce a tracking point projection method, which projects the tracking points onto Gaussian maps and aligns them with the object's feature information." + }, + { + "type": "text", + "bbox": [ + 0.514, + 0.436, + 0.905, + 0.481 + ], + "angle": 0, + "content": "- Experimental results demonstrate that despite being trained on a synthetic 3D dataset, our method generalizes well beyond the dataset's distribution." + }, + { + "type": "list", + "bbox": [ + 0.514, + 0.332, + 0.905, + 0.481 + ], + "angle": 0, + "content": null + }, + { + "type": "title", + "bbox": [ + 0.514, + 0.51, + 0.655, + 0.525 + ], + "angle": 0, + "content": "2. Related Work" + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.54, + 0.843, + 0.557 + ], + "angle": 0, + "content": "2.1. Drag-based controllable image editing" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.569, + 0.907, + 0.901 + ], + "angle": 0, + "content": "In the field of image editing, several established works have focused on tasks such as image customization [16, 18, 33, 41], image editing [1, 4, 6, 9, 11, 14], which modify object attributes, quantity, etc. through text-based operations. Due to the limitations of text-based editing in controlling local changes at a fine-grained level, the Drag approach introduces a novel approach to image editing [7, 8, 19, 23, 24, 27-30, 36]. This interactive manipulation allows users to precisely designate both the editing area and control mode with a simple click. [30, 36] propose a point-based editing technique that integrates point tracking and motion supervision. [19] enhances drag editing stability and reliability without requiring precise point tracking by introducing adaptively updated template features, fuzzy localization, and linear search techniques. [7] introduces a discriminative point tracking method and a confidence-based latent space enhancement strategy, resulting in more stable and accurate drag performance. These works primarily focus on enhancing source and target feature optimization. However, when extended to object rotation manipulation, significant feature differences across views pose a challenge that pure feature alignment cannot resolve." + }, + { + "type": "page_number", + "bbox": [ + 0.479, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "21503" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.119, + 0.09, + 0.882, + 0.315 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.089, + 0.325, + 0.908, + 0.385 + ], + "angle": 0, + "content": "Figure 3. Dragin3D Architecture. We track the representation points by calculating the transfer matrix and plot circle and square Gaussian maps centered on two representation points, respectively. We encode the image to be edited to extract the image features, which are injected into the original encoded Gaussian map features to obtain new Gaussian map features. The original Gaussian map features are added to the newly injected Gaussian map features as an input control condition for ControlNet." + }, + { + "type": "title", + "bbox": [ + 0.09, + 0.407, + 0.465, + 0.423 + ], + "angle": 0, + "content": "2.2. Trajectory-based controllable image editing" + }, + { + "type": "text", + "bbox": [ + 0.093, + 0.43, + 0.485, + 0.749 + ], + "angle": 0, + "content": "Recently, trajectory-based motion control [10, 25, 35, 39, 40, 43, 45, 49] has emerged as an efficient and user-friendly method for controllable video generation [5, 12, 15, 48], with drag editing serving as a trajectory-based task. The gradual movement of handle points forms a continuous trajectory, providing a more precise guidance signal in contrast to defined start and end points. [45] introduces a unified control framework that integrates text, image, and trajectory data, utilizing a trajectory sampler to extract trajectories from open-domain video streams. [43] extracts features from the diffusion model as entity representations, and incorporates them into a denoised 3D U-Net with a noisy 2D Gaussian representation. [10] focuses on specific camera motion, generating videos by adding temporal constraints to a T2I model to capture motion dynamics, leveraging eight independent LoRA models for basic camera control. [40] differentiates between camera and object motions, enabling independent or combined control of the camera pose matrix and object trajectory to produce personalized videos. However, these motions are limited to 2D translations, making it difficult to generate novel views of objects in 3D space." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.757, + 0.295, + 0.773 + ], + "angle": 0, + "content": "2.3. 3D generation Models" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.78, + 0.484, + 0.903 + ], + "angle": 0, + "content": "Based on the impressive generative effect of diffusion models on 2D images, the approach of 3D generation via diffusion models has emerged [17, 20-22, 31, 34, 44, 50]. [31] is one of the pioneering methods, introduces the SDS loss function, which enables the updating of NeRF [26] weights by minimizing the discrepancy between noise levels. Several 3D image generation methods using enhanced SDS loss functions have been proposed, employing a 2D" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.408, + 0.909, + 0.576 + ], + "angle": 0, + "content": "generative model to create 3D structures from textual descriptions without relying on 3D data, thus overcoming the challenge of limited 3D data availability. [44] integrates the diffusion process of image rendering and synthesis using NeRF, ensuring that the generated images remain semantically aligned with the text and geometrically consistent with the depth map during training. [38, 46] directly embed camera parameters as control conditions in the model to realize the camera pose control. However, the direct embedding of discrete values, such as camera parameters, as control conditions often lack the necessary precision." + }, + { + "type": "title", + "bbox": [ + 0.514, + 0.592, + 0.613, + 0.607 + ], + "angle": 0, + "content": "3. Methods" + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.618, + 0.796, + 0.633 + ], + "angle": 0, + "content": "3.1. Task Definition & Preliminaries" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.643, + 0.907, + 0.764 + ], + "angle": 0, + "content": "3D Drag Editing Definition. We observed that previous drag tasks lacked a unified framework for moving objects in 3D space by drag interaction, for example, object rotation, so we propose to extend the task to 3D space. This task marks a significant advancement beyond merely editing the appearance of objects in a 2D image. Given a reference object image, our approach allows users to obtain a new perspective of the object through a simple drag interaction." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.765, + 0.909, + 0.903 + ], + "angle": 0, + "content": "Preliminary of Stable Video Diffusion. Our paradigm is based on Stable Video Diffusion (SVD) [2] due to its robust predictive capabilities for dynamically moving objects. The key modification from Stable Diffusion (SD) [3] is the integration of a temporal layer into the denoising model's U-Net architecture, resulting in a 3D-UNet. Compared to SD trained on image datasets, video frames capture more dynamic information about objects, leading to more sensitivity to morphological changes of the same objects, ensur" + }, + { + "type": "page_number", + "bbox": [ + 0.479, + 0.945, + 0.521, + 0.958 + ], + "angle": 0, + "content": "21504" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.091, + 0.092, + 0.441, + 0.106 + ], + "angle": 0, + "content": "ing perspective and temporal consistency of rotation." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.107, + 0.484, + 0.228 + ], + "angle": 0, + "content": "Preliminary of ControlNet. ControlNet [47] is one of the representative works in the field of controlled image generation, which supports a variety of inputs such as edge maps, segmentation maps, keypoints, etc., as conditional inputs to guide the image generation process. \\(\\mathcal{F}(\\cdot ;\\Theta)\\) is a neural network block with parameters \\(\\Theta\\). ControlNet locks all parameters in \\(\\Theta\\) and clones them into a trainable copy \\(\\Theta_{c}\\). The output of ControlNet \\(y_{c}\\) can be expressed as:" + }, + { + "type": "equation", + "bbox": [ + 0.101, + 0.255, + 0.483, + 0.272 + ], + "angle": 0, + "content": "\\[\n\\boldsymbol {y} _ {\\mathbf {c}} = \\mathcal {F} (\\boldsymbol {x}; \\Theta) + \\mathcal {Z} \\left(\\mathcal {F} \\left(\\boldsymbol {x} + \\mathcal {Z} \\left(\\boldsymbol {c}; \\Theta_ {\\mathrm {z} 1}\\right); \\Theta_ {\\mathrm {c}}\\right); \\Theta_ {\\mathrm {z} 2}\\right), \\tag {1}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.282, + 0.484, + 0.329 + ], + "angle": 0, + "content": "where \\( c \\) is the condition input, \\( \\mathcal{Z}(\\cdot ;\\cdot) \\) is a \\( 1\\times 1 \\) convolutional layer initialized with zero weights and bias. When ControlNet is used in Stable Diffusion, the training objective is:" + }, + { + "type": "equation", + "bbox": [ + 0.106, + 0.354, + 0.483, + 0.373 + ], + "angle": 0, + "content": "\\[\n\\mathcal {L} = \\mathbb {E} _ {\\boldsymbol {z} _ {0}, t, \\boldsymbol {c} _ {t}, \\boldsymbol {c} _ {\\mathrm {f}}, \\epsilon \\sim \\mathcal {N} (0, 1)} \\left[ \\| \\epsilon - \\epsilon_ {\\theta} \\left(z _ {t}, t, \\boldsymbol {c} _ {t}, \\boldsymbol {c} _ {\\mathrm {f}}\\right)) \\| _ {2} ^ {2} \\right], \\tag {2}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.383, + 0.484, + 0.428 + ], + "angle": 0, + "content": "where \\( z_0 \\) denotes the input image, \\( z_t \\) denotes the noisy image in timestep \\( t \\), \\( c_t \\) and \\( c_f \\) denote text prompts and the task-specific conditions, respectively." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.429, + 0.484, + 0.549 + ], + "angle": 0, + "content": "Preliminary of Gaussian heatmap. Gaussian heatmaps are commonly used for keypoint detection tasks. A heatmap is a feature map of the same size as the input, with a value of 1 at a specified coordinate (i.e., the representation points in our task) and 0 at all other positions. Additionally, to align with inputs like mask or depth maps used in ControlNet, a Gaussian heatmap is a suitable input, which effectively converts discrete point signals into continuous image signals." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.559, + 0.281, + 0.573 + ], + "angle": 0, + "content": "3.2. Overall Framework" + }, + { + "type": "text", + "bbox": [ + 0.089, + 0.581, + 0.483, + 0.824 + ], + "angle": 0, + "content": "For this novel dragging in 3D space task, we propose a framework named Dragin3D, which incorporates two key innovations: 3D drag trajectory representation and control condition injection. The overall workflow of our model is illustrated in Fig. 3. We employ an architecture that combines SVD and ControlNet. Firstly, we propose to describe the motion trajectory of the object with two points, and compute to track the positions of the representation points in different views. Next, we construct Gaussian maps of different shapes centered on these two points. These maps are then encoded, and image feature representations are injected into the Gaussian maps to align them with the corresponding point locations. Finally, the encoded Gaussian map features are linked to new Gaussian map features that incorporate object-specific information, which are used as control condition inputs for ControlNet." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.832, + 0.454, + 0.849 + ], + "angle": 0, + "content": "3.3. 3D Drag Trajectory Representation (DTR)" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.856, + 0.484, + 0.902 + ], + "angle": 0, + "content": "To extend the editing task to 3D space, we propose a 3D Drag Trajectory Representation (DTR) method, which explicitly represents the object's trajectory using a rotational" + }, + { + "type": "image", + "bbox": [ + 0.536, + 0.089, + 0.613, + 0.156 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.536, + 0.157, + 0.551, + 0.166 + ], + "angle": 0, + "content": "#" + }, + { + "type": "image", + "bbox": [ + 0.634, + 0.09, + 0.707, + 0.156 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.631, + 0.157, + 0.645, + 0.167 + ], + "angle": 0, + "content": "#" + }, + { + "type": "image", + "bbox": [ + 0.727, + 0.09, + 0.802, + 0.156 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.726, + 0.157, + 0.741, + 0.167 + ], + "angle": 0, + "content": "#" + }, + { + "type": "image", + "bbox": [ + 0.825, + 0.09, + 0.901, + 0.156 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.821, + 0.157, + 0.835, + 0.167 + ], + "angle": 0, + "content": "#" + }, + { + "type": "image_caption", + "bbox": [ + 0.512, + 0.181, + 0.906, + 0.237 + ], + "angle": 0, + "content": "Figure 4. Our 3D drag trajectory representation method. The computed tracking points are projected into Gaussian maps, ensuring the rotational trajectory of the object is visually aligned with the trajectory of the line segment formed by the representation points." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.262, + 0.906, + 0.337 + ], + "angle": 0, + "content": "line segment defined by two points. In general, an object's rotation in 3D space involves variations around three coordinate axes across three planes. Since left-right rotation is more intuitive and commonly used at the operational level, our approach and experiments focus on the z-axis rotation." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.338, + 0.907, + 0.638 + ], + "angle": 0, + "content": "The selection and stable tracking of the representation points are critical, as they directly affect the accuracy of the subsequent results. To ensure that the motion trajectory of the line segment is coherent and clearly defined, we propose a method for calculating the position of a specified coordinate point in images captured from different views. Then we project the computed points onto the 2D plane, as shown in Fig. 4. The required camera intrinsic and extrinsic parameters are provided by the 3D dataset. Extrinsic parameter calibration is performed by solving the transfer matrix using the real-world coordinates and pixel coordinates of fixed points. Conversely, when the camera extrinsic parameter is available, we can calculate the pixel coordinates of a fixed point by maintaining the constancy of its world coordinates. To simplify the calculation of 3D coordinates, we assume that the z-axis coordinate of each pixel point is fixed to 1. While this assumption introduces some deviation, the resulting calculation provides a reasonable approximation of the trajectory and magnitude of the object's movement. The relevant calculation formulas are as follows." + }, + { + "type": "equation", + "bbox": [ + 0.523, + 0.665, + 0.905, + 0.682 + ], + "angle": 0, + "content": "\\[\nx _ {c 1} = \\left(x _ {1} - c _ {x}\\right) * z _ {c 1} / f _ {x}, y _ {c 1} = \\left(y _ {1} - c _ {y}\\right) * z _ {c 1} / f _ {y}, \\tag {3}\n\\]" + }, + { + "type": "equation", + "bbox": [ + 0.534, + 0.714, + 0.868, + 0.758 + ], + "angle": 0, + "content": "\\[\n\\begin{array}{l} \\left[ \\begin{array}{l l l} x _ {w} & y _ {w} & z _ {w} \\end{array} \\right] ^ {- 1} = T _ {1} ^ {- 1} \\cdot \\left[ \\begin{array}{l l l} x _ {c 1} & y _ {c 1} & z _ {c 1} \\end{array} \\right] ^ {- 1}, \\\\ \\left[ \\begin{array}{c} x _ {c 2} y _ {c 2} z _ {c 2} 1 \\end{array} \\right] ^ {- 1} = T _ {2} \\cdot \\left[ \\begin{array}{c} x _ {w} y _ {w} z _ {w} 1 \\end{array} \\right] ^ {- 1}, \\\\ \\end{array}\n\\]" + }, + { + "type": "equation", + "bbox": [ + 0.536, + 0.784, + 0.905, + 0.8 + ], + "angle": 0, + "content": "\\[\nx _ {2} = f _ {x} * x _ {c 2} / z _ {c 2} + c _ {x}, y _ {2} = f _ {y} * y _ {c 2} / z _ {c 2} + c _ {y}, \\tag {5}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.81, + 0.907, + 0.902 + ], + "angle": 0, + "content": "where the intrinsic \\((f_{x},f_{y},c_{x},c_{y})\\) and extrinsic \\((T_{1},T_{2})\\) of the camera are labeled in the dataset. We define \\(P_{1}\\) \\((x_{1},y_{1})\\) as the pixel coordinates of the specified point in the first image, \\(P_{c1}\\) \\((x_{c1},y_{c1},z_{c1})\\) is the camera coordinates of the specified point in the first image, \\(P_{w}\\) \\((x_{w},y_{w},z_{w})\\) is the world coordinates, \\(P_{c2}\\) \\((x_{c2},y_{c2},z_{c2})\\) and \\(P_{2}\\) \\((x_{2},y_{2})\\) are" + }, + { + "type": "page_number", + "bbox": [ + 0.479, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "21505" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.152, + 0.09, + 0.852, + 0.322 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.089, + 0.331, + 0.908, + 0.39 + ], + "angle": 0, + "content": "Figure 5. Dragin3D Visualization. Given an input image and a set of trajectory point pairs, we obtained Gaussian maps of the trajectories as input (first column), which is then used to generate the corresponding objects at the specified trajectory points (subsequent columns). The arrows indicate rotation directions, the red line is the current view of the object, the gray dotted line shows the trajectory that has already been traversed, and the gray solid line is the future trajectory." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.415, + 0.482, + 0.446 + ], + "angle": 0, + "content": "the camera coordinates and pixel coordinates of the specified point in the second image, respectively." + }, + { + "type": "title", + "bbox": [ + 0.09, + 0.455, + 0.449, + 0.472 + ], + "angle": 0, + "content": "3.4. Control Condition Inject Module (CCIM)" + }, + { + "type": "text", + "bbox": [ + 0.093, + 0.479, + 0.483, + 0.734 + ], + "angle": 0, + "content": "In order to process the existing input image and the computed trajectory representations into suitable input conditions for the model, we proposed a Control Condition Inject Module (CCIM). Drawing inspiration from DragAnything [43], which uses discrete pixel points as input control conditions by mapping them into a Gaussian map. We propose to use two representation points. However, we observed that the two representation points of the object are nearly identical in the front and rear views, as well as in the symmetric left and right side views. We propose a Shape DistinGuish Strategy (SDGS) to distinguish the two points by their respective shapes. Specifically, we create circular and square Gaussian maps centered on the two points, respectively, as shown in Fig. 9 (green boxes). Then the Gaussian map images \\( x \\in \\mathbb{R}^{F \\times H \\times W \\times 3} \\) are encoded into image features \\( f_{gi} \\in \\mathbb{R}^{F \\times H \\times W \\times C} \\), where \\( F \\) denotes the input frames and \\( C \\) denotes the channel size." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.735, + 0.483, + 0.902 + ], + "angle": 0, + "content": "CCIM. Since multi-view image generation requires the guidance of input object representation information, the Gaussian map that only indicates the location and trajectory of the object is insufficient. To establish a correspondence between the location of the representation points and the image features, we propose to use a trainable image prompt mapping network, which allows the image to function as an image prompt similar to the textual guidance condition. In order to ensure consistency between training and inference, we use only the first frame image \\( x_{1} \\) during training. The first frame image embedding is extracted by the pre" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.415, + 0.907, + 0.596 + ], + "angle": 0, + "content": "trained CLIP [32] image encoder, and our image prompt mapping network projects this CLIP embedding into an image prompt feature \\( f_{ip} \\in \\mathbb{R}^{1 \\times C} \\). Then we replace the features at the representation points' locations in the Gaussian map image features \\( f_{gi} \\) with this sequence of image prompt features. The resulting new Gaussian map image features are denoted as \\( f_{ngi} \\in \\mathbb{R}^{F \\times H \\times W \\times C} \\). It is worth noting that since the start and end frames are randomly selected during training, all images in the training set with different views have the potential to be encoded as image prompts. This ensures diversity and richness in the model's ability to learn from varying perspectives." + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.597, + 0.907, + 0.673 + ], + "angle": 0, + "content": "Through the above analysis, the input condition of ControlNet is the sum of \\( f_{gi} \\) and \\( f_{ngi} \\), that is \\( f_{input} = f_{gi} + f_{ngi} \\). During training, we only optimize parameters of the image prompt mapping network and ControlNet, and the final optimization objective is:" + }, + { + "type": "equation", + "bbox": [ + 0.568, + 0.689, + 0.907, + 0.708 + ], + "angle": 0, + "content": "\\[\n\\min _ {\\theta} E _ {x _ {0}, \\epsilon , t} \\| \\epsilon - \\epsilon_ {\\theta} (x _ {t}, t, c (G, G _ {n e w})) \\| ^ {2}, \\tag {6}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.714, + 0.908, + 0.777 + ], + "angle": 0, + "content": "where \\(\\epsilon \\sim \\mathcal{N}(0,I)\\), \\(t\\in [0,T]\\) denotes the time step of diffusion process, \\(c(\\cdot)\\) is the condition mechanism with explicit controls. \\(G\\) and \\(G_{new}\\) denote the initial and new Gaussian map injected with image features, respectively." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.791, + 0.649, + 0.809 + ], + "angle": 0, + "content": "4. Experiments" + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.817, + 0.704, + 0.833 + ], + "angle": 0, + "content": "4.1. Experimental Setup" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.84, + 0.908, + 0.903 + ], + "angle": 0, + "content": "Datasets. We utilize the widely adopted OmniObject3D dataset [42], containing approximately 6k 3D object models in real-world across more than 200 categories. We selected 2.4k 3D objects for training and 200 for testing. To" + }, + { + "type": "page_number", + "bbox": [ + 0.479, + 0.945, + 0.521, + 0.958 + ], + "angle": 0, + "content": "21506" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.111, + 0.087, + 0.373, + 0.345 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.384, + 0.087, + 0.633, + 0.344 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.636, + 0.087, + 0.889, + 0.344 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.089, + 0.355, + 0.908, + 0.44 + ], + "angle": 0, + "content": "Figure 6. Comparison with existing methods. We compare existing methods relevant to our task. For 3D dataset images with pose labels, we directly compute the representation points, with the first row representing the ground truth image and its corresponding rotation. For real images, we utilize the interaction interface to define the desired drag target points (\"Our drag points\" in the figure). Additionally, we specify a drag trajectory for DragAnything that closely resembles our trajectory. DragAnything interprets the trajectory as the movement of the object or camera along it but cannot handle rotations. SV3D struggles with generalization to real-world images and lacks sufficient detail and 3D consistency. Wonder3D generates higher-quality 3D images but is limited to six fixed views." + }, + { + "type": "table", + "bbox": [ + 0.104, + 0.462, + 0.472, + 0.555 + ], + "angle": 0, + "content": "
MethodsMulti-viewBackgroundTime
DragAnything×✓~40s
SV3D✓×~2min
Wonder3D✓×~2min
Ours✓✓~25s
" + }, + { + "type": "table_caption", + "bbox": [ + 0.09, + 0.565, + 0.483, + 0.637 + ], + "angle": 0, + "content": "Table 1. Functionality comparisons. Our proposed Dragin3D achieves multi-view generation and supports editing images with backgrounds, ensuring efficient and rapid editing with minimal time consumption. In contrast, methods from different areas are unable to perform these functions simultaneously." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.668, + 0.483, + 0.729 + ], + "angle": 0, + "content": "better simulate real-world scenes, we incorporate diverse background images into the dataset. Additionally, the image resolution is resized to \\(256*256\\) to optimize computational efficiency during training." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.732, + 0.483, + 0.838 + ], + "angle": 0, + "content": "Baselines. Since Dragin3D integrates drag with multiview tasks, we use DragAnything [43], SV3D [38], and Wonder3D [22] as baselines. DragAnything and SV3D are trained on video framework, Wonder3D is trained on image framework. DragAnything is a 2D interaction editing method trained on video datasets, SV3D and Wonder3D are 3D generation methods trained on multi-view datasets." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.84, + 0.484, + 0.902 + ], + "angle": 0, + "content": "Evaluation Metrics. For multi-view generation evaluation, we use the following three metrics for quantitative evaluation: For the quality of generated images, we employ Fréchet Inception Distance (FID), which calculates the" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.465, + 0.907, + 0.692 + ], + "angle": 0, + "content": "Frechet distance between the feature vectors of the generated image and ground truth. For the perceptual similarity of the 3D effects, we use Learned Perceptual Image Patch Similarity (LPIPS), which evaluates the perceptual similarity between the generated image and the ground truth by learning the inverse mapping of the generated image. For the quality of the generated motion sequences, we apply Frechet Video Distance (FVD), which leverages a 3D feature extractor to compare video sequences. For drag trajectory accuracy evaluation, we adopt the MD metric in DragDiffusion, which employs DIFT [37] to identify the points in the images dragged by specified trajectories, corresponding to the starting representation points. The average Euclidean distance between the identified points and the reference trajectory points is then computed." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.694, + 0.907, + 0.77 + ], + "angle": 0, + "content": "Implements Details. We randomly select 8 coherent object rotation images from the OmniObject3D dataset, which are used as the individual frames of video input to the SVD architecture. Eight A800 GPUs are used to train 20k steps with a batch size of 6, the learning rate is set to 2e-5." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.785, + 0.696, + 0.801 + ], + "angle": 0, + "content": "4.2. Qualitative Results" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.81, + 0.907, + 0.903 + ], + "angle": 0, + "content": "The visualization results of Dragin3D are shown in Fig. 5. Given an image with a corresponding trajectory, Dragin3D generates a high-fidelity image of the object from a new perspective, maintaining strong consistency with the original image. In addition, Dragin3D demonstrates superior generalization performance, which successfully rotates ob" + }, + { + "type": "page_number", + "bbox": [ + 0.479, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "21507" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.097, + 0.088, + 0.48, + 0.257 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.09, + 0.269, + 0.483, + 0.312 + ], + "angle": 0, + "content": "Figure 7. Comparison with 3D methods. We use the image without background as input to 3D methods and compare the results with ground truth, demonstrating that our method is comparable." + }, + { + "type": "text", + "bbox": [ + 0.089, + 0.339, + 0.483, + 0.368 + ], + "angle": 0, + "content": "jects by dragging the representation points on real images that lie outside the training dataset distribution." + }, + { + "type": "text", + "bbox": [ + 0.093, + 0.37, + 0.485, + 0.702 + ], + "angle": 0, + "content": "As shown in Fig. 7, we input the image into SV3D and Wonder3D after background removal and compare the results with the ground truth from the corresponding viewpoints. Our methods demonstrate comparable performance in terms of rotation accuracy and 3D consistency. Since our main contribution lies in the generalizability to real-world images, we include the background to reduce the gap between the ideal 3D dataset and the real-world scene. As illustrated in Fig. 6, we present the visualization results of each method with background using both 3D datasets and real images. The real images we provided have homogeneous backgrounds, minimizing their impact on the object. The results show that the performance of the other methods is significantly lower, limiting their practical applicability, while our method demonstrates a one-stage process that eliminates background interference during the image processing stage, performs well in detail preservation and coherent rotation. Additionally, we found that DragDiffusion [36] can achieve partial rotation in certain cases of Drag-Bench. However, the rotation is minimal and random, lacking precise control over the direction and angle of rotation as specified by the instructions, as illustrated in Fig. 8." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.711, + 0.285, + 0.727 + ], + "angle": 0, + "content": "4.3. Quantitative Results" + }, + { + "type": "text", + "bbox": [ + 0.089, + 0.735, + 0.483, + 0.901 + ], + "angle": 0, + "content": "Since our proposed task diverges from traditional drag-based editing and produces significantly different outcomes, and 3D generation models like Wonder3D focus on generating images without background from partial views, we do not include them in a quantitative comparison. Our Dragin3D shares a similar goal with SV3D, which aims to generate images from new perspectives based on a single input image. Therefore, we chose to make a comparison with SV3D. Due to the absence of ground truth in the real-world data, we randomly selected 200 subsets from the OmniObject3D dataset as the testset for computation, which were" + }, + { + "type": "image", + "bbox": [ + 0.532, + 0.088, + 0.891, + 0.208 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.513, + 0.218, + 0.907, + 0.275 + ], + "angle": 0, + "content": "Figure 8. Comparison of rotation cases with DragDiffusion. DragDiffusion primarily manifests as tiny rotations induced by movements in the plane towards the target point, while Dragin3D enables more accurate control of stereoscopic rotation." + }, + { + "type": "table", + "bbox": [ + 0.554, + 0.289, + 0.866, + 0.352 + ], + "angle": 0, + "content": "
MethodsFID↓LPIPS↓FVD↓
SV3D [38]52.50.410944.6
Dragin3D (Ours)43.40.179584.6
" + }, + { + "type": "table_caption", + "bbox": [ + 0.513, + 0.362, + 0.905, + 0.39 + ], + "angle": 0, + "content": "Table 2. Quantitative comparisons. Our proposed Dragin3D outperforms SV3D in all metrics." + }, + { + "type": "table", + "bbox": [ + 0.536, + 0.405, + 0.885, + 0.453 + ], + "angle": 0, + "content": "
MethodsDragDiffusion [36]Dragin3D (Ours)
MD↓33.019.38
" + }, + { + "type": "table_caption", + "bbox": [ + 0.513, + 0.462, + 0.905, + 0.491 + ], + "angle": 0, + "content": "Table 3. Quantitative comparisons. Our proposed Dragin3D outperforms DragDiffusion in control accuracy evaluation." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.518, + 0.907, + 0.685 + ], + "angle": 0, + "content": "not visible during training. We compared the similarity between the generated images and ground truth across the corresponding viewpoints from three perspectives: image, 3D perceptual, and video. As shown in Tab. 2, Dragin3D outperforms SV3D in all metrics. To evaluate the accuracy of trajectory control, we selected object rotation cases in Drag-Bench including animal, human, and sculpture heads. For each set of desired rotation trajectories, we computed the MDs for both Dragdiffusion and Dragin3D, with Dragin3D outperforming Dragdiffusion, as shown in Tab. 3. Additionally, we conducted a user study in Supp. 7.2." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.696, + 0.625, + 0.71 + ], + "angle": 0, + "content": "4.4. Ablations" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.719, + 0.907, + 0.885 + ], + "angle": 0, + "content": "Effectiveness of shape distinguish strategy (SDGS). To demonstrate the effectiveness of SDGS, we designed experiments in which the Gaussian maps for two points were constrained to the same shape, as shown in Tab. 4. We specifically focused on the symmetric rotation of objects, as illustrated in Fig. 9. Without distinguishing between the two points, the object in the subsequent stage of the rotation becomes difficult to distinguish from its symmetrical configuration in the preliminary stage. In contrast, the generated object preserves structural integrity and details when SDGS is applied. More examples can be found in Supp. 7.3." + }, + { + "type": "text", + "bbox": [ + 0.533, + 0.886, + 0.905, + 0.901 + ], + "angle": 0, + "content": "Effectiveness of control condition inject module" + }, + { + "type": "page_number", + "bbox": [ + 0.479, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "21508" + } + ], + [ + { + "type": "table", + "bbox": [ + 0.118, + 0.089, + 0.457, + 0.166 + ], + "angle": 0, + "content": "
MethodsFID↓LPIPS↓FVD↓
Dragin3D w/o SDGS44.10.182676.8
Dragin3D w/o CCIM45.20.194694.1
Dragin3D43.40.179584.6
" + }, + { + "type": "table_caption", + "bbox": [ + 0.091, + 0.177, + 0.483, + 0.204 + ], + "angle": 0, + "content": "Table 4. Ablation study results. All metrics have declined without SDGS and CCIM, respectively." + }, + { + "type": "image", + "bbox": [ + 0.102, + 0.221, + 0.473, + 0.48 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.09, + 0.491, + 0.483, + 0.616 + ], + "angle": 0, + "content": "Figure 9. Effectiveness of shape distinguish strategy (SDGS). We present the resulting images for two cases. The first row displays the trajectory points projected onto Gaussian maps, while the second and third rows show the objects from the corresponding view of these trajectory points, generated with and without SDGS, respectively. The positions of the two trajectory points in the first frame closely correspond to those in the last frame (green box). In the latter half of the sequence, significant misalignment occurs due to interference from the Gaussian representation (brown box)." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.644, + 0.483, + 0.901 + ], + "angle": 0, + "content": "(CCIM). As illustrated in Tab. 4, to assess the effectiveness of CCIM, we conducted experiments using only Gaussian map image features \\( f_{gi} \\) derived from the tracking points as the control input. The results reveal a marked enhancement across all evaluation metrics upon the integration of CCIM. Visual comparisons of intermediate results are presented in Fig. 10. Without CCIM, the object's rotation initially exhibits low-amplitude fluctuations, followed by a sudden and drastic change. Specifically, the early frames retain the original image, while later frames generate objects with significant distortion and poor fidelity. In contrast, the rotation transition is smooth and stable, and the object's details are preserved throughout. These results demonstrate that CCIM effectively aligns the rotation trajectory with image features by incorporating object-specific information, ensuring that the generated perspectives remain faithful to the original image content. More examples can be found in Supp. 7.3." + }, + { + "type": "image", + "bbox": [ + 0.527, + 0.091, + 0.895, + 0.421 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.513, + 0.432, + 0.908, + 0.543 + ], + "angle": 0, + "content": "Figure 10. Effectiveness of control condition inject module (CCIM). We present the resulting images for two cases. The first row displays the trajectory points projected onto Gaussian maps, while the second and third rows show the objects from the corresponding view of these trajectory points, generated with and without CCIM, respectively. Without CCIM, the image remains largely unchanged in the early stages (red box), but later undergoes a sudden distortion, causing significant fidelity loss (yellow box)." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.572, + 0.791, + 0.587 + ], + "angle": 0, + "content": "5. Conclusion and Future Works" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.599, + 0.907, + 0.901 + ], + "angle": 0, + "content": "In this paper, we present Dragin3D, which extends the interactive trajectory-based dragging task to 3D space. To represent the rotation trajectory of an object from the starting view to the terminating view, we introduce a double-point representation approach, and map the rotation of the object to a line segment formed by these two points. Our approach introduces control conditions through ControlNet architecture based on the large-scale video generation pretrained model SVD. In addition, we randomly add different backgrounds to the 3D object dataset, enabling the model to directly handle images with backgrounds. Through extensive experiments, Dragin3D successfully generates novel view images of the object based on the trajectories of two points, minimizing background interference while preserving the original object features, and demonstrating exceptional generalization on real-world images. For future work, since our approach currently supports object rotation in only one plane, we plan to expand our trajectory-based editing task to encompass a broader range of views, enabling arbitrary rotations in 3D space." + }, + { + "type": "page_number", + "bbox": [ + 0.479, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "21509" + } + ], + [ + { + "type": "title", + "bbox": [ + 0.091, + 0.091, + 0.27, + 0.108 + ], + "angle": 0, + "content": "6. Acknowledgments" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.115, + 0.484, + 0.207 + ], + "angle": 0, + "content": "This research is supported by Artificial Intelligence-National Science and Technology Major Project 2023ZD0121200, the National Science Fund for Excellent Young Scholars under Grant 62222212, and National Natural Science Foundation of China under Grant 623B2094." + }, + { + "type": "title", + "bbox": [ + 0.093, + 0.222, + 0.188, + 0.237 + ], + "angle": 0, + "content": "References" + }, + { + "type": "ref_text", + "bbox": [ + 0.101, + 0.247, + 0.484, + 0.302 + ], + "angle": 0, + "content": "[1] Omer Bar-Tal, Dolev Ofri-Amar, Rafail Fridman, Yoni Kasten, and Tali Dekel. Text2live: Text-driven layered image and video editing. In European conference on computer vision, pages 707-723. Springer, 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.1, + 0.304, + 0.484, + 0.373 + ], + "angle": 0, + "content": "[2] Andreas Blattmann, Tim Dockhorn, Sumith Kulal, Daniel Mendelevitch, Maciej Kilian, Dominik Lorenz, Yam Levi, Zion English, Vikram Voleti, Adam Letts, et al. Stable video diffusion: Scaling latent video diffusion models to large datasets. arXiv preprint arXiv:2311.15127, 2023. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.101, + 0.375, + 0.484, + 0.457 + ], + "angle": 0, + "content": "[3] Andreas Blattmann, Robin Rombach, Huan Ling, Tim Dockhorn, Seung Wook Kim, Sanja Fidler, and Karsten Kreis. Align your latents: High-resolution video synthesis with latent diffusion models. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 22563-22575, 2023. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.46, + 0.484, + 0.528 + ], + "angle": 0, + "content": "[4] Tim Brooks, Aleksander Holynski, and Alexei A Efros. Instructpix2pix: Learning to follow image editing instructions. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 18392-18402, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.531, + 0.482, + 0.586 + ], + "angle": 0, + "content": "[5] Weifeng Chen, Yatai Ji, Jie Wu, Hefeng Wu, Pan Xie, Jiashi Li, Xin Xia, Xuefeng Xiao, and Liang Lin. Control-a-video: Controllable text-to-video generation with diffusion models. arXiv preprint arXiv:2305.13840, 2023. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.588, + 0.484, + 0.643 + ], + "angle": 0, + "content": "[6] Guillaume Couairon, Jakob Verbeek, Holger Schwenk, and Matthieu Cord. Diffedit: Diffusion-based semantic image editing with mask guidance. arXiv preprint arXiv:2210.11427, 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.645, + 0.484, + 0.699 + ], + "angle": 0, + "content": "[7] Yutao Cui, Xiaotong Zhao, Guozhen Zhang, Shengming Cao, Kai Ma, and Limin Wang. Stabledrag: Stable dragging for point-based image editing. arXiv preprint arXiv:2403.04437, 2024.1, 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.702, + 0.484, + 0.744 + ], + "angle": 0, + "content": "[8] Yuki Endo. User-controllable latent transformer for stylegan image layout editing. In Computer Graphics Forum, pages 395-406. Wiley Online Library, 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.746, + 0.484, + 0.8 + ], + "angle": 0, + "content": "[9] Dave Epstein, Allan Jabri, Ben Poole, Alexei Efros, and Aleksander Holynski. Diffusion self-guidance for controllable image generation. Advances in Neural Information Processing Systems, 36:16222-16239, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.803, + 0.484, + 0.871 + ], + "angle": 0, + "content": "[10] Yuwei Guo, Ceyuan Yang, Anyi Rao, Zhengyang Liang, Yaohui Wang, Yu Qiao, Maneesh Agrawala, Dahua Lin, and Bo Dai. Animatediff: Animate your personalized text-to-image diffusion models without specific tuning. arXiv preprint arXiv:2307.04725, 2023. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.873, + 0.484, + 0.901 + ], + "angle": 0, + "content": "[11] Amir Hertz, Ron Mokady, Jay Tenenbaum, Kfir Aberman, Yael Pritch, and Daniel Cohen-Or. Prompt-to-prompt im" + }, + { + "type": "list", + "bbox": [ + 0.094, + 0.247, + 0.484, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.545, + 0.093, + 0.907, + 0.12 + ], + "angle": 0, + "content": "age editing with cross attention control. arXiv preprint arXiv:2208.01626, 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.122, + 0.907, + 0.188 + ], + "angle": 0, + "content": "[12] Yaosi Hu, Chong Luo, and Zhenzhong Chen. Make it move: controllable image-to-video generation with text descriptions. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 18219-18228, 2022. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.19, + 0.907, + 0.26 + ], + "angle": 0, + "content": "[13] Tero Karras, Samuli Laine, Miika Aittala, Janne Hellsten, Jaakko Lehtinen, and Timo Aila. Analyzing and improving the image quality of stylegan. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 8110-8119, 2020. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.261, + 0.906, + 0.329 + ], + "angle": 0, + "content": "[14] Bahjat Kawar, Shiran Zada, Oran Lang, Omer Tov, Huiwen Chang, Tali Dekel, Inbar Mosseri, and Michal Irani. Imagc: Text-based real image editing with diffusion models. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 6007-6017, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.33, + 0.906, + 0.385 + ], + "angle": 0, + "content": "[15] Ali Köksal, Kenan E Ak, Ying Sun, Deepu Rajan, and Joo Hwee Lim. Controllable video generation with text-based instructions. IEEE transactions on multimedia, 26: 190-201, 2023. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.386, + 0.906, + 0.455 + ], + "angle": 0, + "content": "[16] Nupur Kumari, Bingliang Zhang, Richard Zhang, Eli Shechtman, and Jun-Yan Zhu. Multi-concept customization of text-to-image diffusion. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 1931-1941, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.456, + 0.907, + 0.525 + ], + "angle": 0, + "content": "[17] Jeong-gi Kwak, Erqun Dong, Yuhe Jin, Hanseok Ko, Shweta Mahajan, and Kwang Moo Yi. Vivid-1-to-3: Novel view synthesis with video diffusion models. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 6775–6785, 2024. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.526, + 0.906, + 0.58 + ], + "angle": 0, + "content": "[18] Dongxu Li, Junnan Li, and Steven Hoi. Blip-diffusion: Pretrained subject representation for controllable text-to-image generation and editing. Advances in Neural Information Processing Systems, 36, 2024. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.581, + 0.906, + 0.65 + ], + "angle": 0, + "content": "[19] Pengyang Ling, Lin Chen, Pan Zhang, Huaian Chen, Yi Jin, and Jinjin Zheng. Freedrag: Feature dragging for reliable point-based image editing. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 6860-6870, 2024. 1, 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.651, + 0.906, + 0.718 + ], + "angle": 0, + "content": "[20] Minghua Liu, Chao Xu, Haian Jin, Linghao Chen, Mukund Varma T, Zexiang Xu, and Hao Su. One-2-3-45: Any single image to 3d mesh in 45 seconds without per-shape optimization. Advances in Neural Information Processing Systems, 36, 2024. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.72, + 0.906, + 0.775 + ], + "angle": 0, + "content": "[21] Yuan Liu, Cheng Lin, Zijiao Zeng, Xiaoxiao Long, Lingjie Liu, Taku Komura, and Wenping Wang. Syncdreamer: Generating multiview-consistent images from a single-view image. arXiv preprint arXiv:2309.03453, 2023." + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.776, + 0.906, + 0.859 + ], + "angle": 0, + "content": "[22] Xiaoxiao Long, Yuan-Chen Guo, Cheng Lin, Yuan Liu, Zhiyang Dou, Lingjie Liu, Yuexin Ma, Song-Hai Zhang, Marc Habermann, Christian Theobalt, et al. Wonder3d: Single image to 3d using cross-domain diffusion. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 9970-9980, 2024. 3, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.86, + 0.906, + 0.901 + ], + "angle": 0, + "content": "[23] Jingyi Lu, Xinghui Li, and Kai Han. Regiondrag: Fast region-based image editing with diffusion models. arXiv preprint arXiv:2407.18247, 2024. 2" + }, + { + "type": "list", + "bbox": [ + 0.517, + 0.093, + 0.907, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.479, + 0.945, + 0.521, + 0.957 + ], + "angle": 0, + "content": "21510" + } + ], + [ + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.093, + 0.482, + 0.134 + ], + "angle": 0, + "content": "[24] Minxing Luo, Wentao Cheng, and Jian Yang. Rotationdrag: Point-based image editing with rotated diffusion features. arXiv preprint arXiv:2401.06442, 2024. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.136, + 0.482, + 0.177 + ], + "angle": 0, + "content": "[25] Wan-Duo Kurt Ma, John P Lewis, and W Bastiaan Kleijn. Trailblazer: Trajectory control for diffusion-based video generation. arXiv preprint arXiv:2401.00896, 2023. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.179, + 0.482, + 0.247 + ], + "angle": 0, + "content": "[26] Ben Mildenhall, Pratul P Srinivasan, Matthew Tancik, Jonathan T Barron, Ravi Ramamoorthi, and Ren Ng. Nerf: Representing scenes as neural radiance fields for view synthesis. Communications of the ACM, 65(1):99-106, 2021. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.25, + 0.482, + 0.305 + ], + "angle": 0, + "content": "[27] Chong Mou, Xintao Wang, Jiechong Song, Ying Shan, and Jian Zhang. Dragondiffusion: Enabling drag-style manipulation on diffusion models. arXiv preprint arXiv:2307.02421, 2023. 1, 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.307, + 0.483, + 0.376 + ], + "angle": 0, + "content": "[28] Chong Mou, Xintao Wang, Jiechong Song, Ying Shan, and Jian Zhang. Diffeditor: Boosting accuracy and flexibility on diffusion-based image editing. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 8488-8497, 2024." + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.378, + 0.482, + 0.433 + ], + "angle": 0, + "content": "[29] Shen Nie, Hanzhong Allan Guo, Cheng Lu, Yuhao Zhou, Chenyu Zheng, and Chongxuan Li. The blessing of randomness: Sde beats ode in general diffusion-based image editing. arXiv preprint arXiv:2311.01410, 2023." + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.435, + 0.482, + 0.504 + ], + "angle": 0, + "content": "[30] Xingang Pan, Ayush Tewari, Thomas Leimkuhler, Lingjie Liu, Abhimitra Meka, and Christian Theobalt. Drag your gan: Interactive point-based manipulation on the generative image manifold. In ACM SIGGRAPH 2023 Conference Proceedings, pages 1-11, 2023. 1, 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.506, + 0.482, + 0.547 + ], + "angle": 0, + "content": "[31] Ben Poole, Ajay Jain, Jonathan T Barron, and Ben Mildenhall. Dreamfusion: Text-to-3d using 2d diffusion. arXiv preprint arXiv:2209.14988, 2022. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.549, + 0.482, + 0.631 + ], + "angle": 0, + "content": "[32] Alec Radford, Jong Wook Kim, Chris Hallacy, Aditya Ramesh, Gabriel Goh, Sandhini Agarwal, Girish Sastry, Amanda Askell, Pamela Mishkin, Jack Clark, et al. Learning transferable visual models from natural language supervision. In International conference on machine learning, pages 8748-8763. PMLR, 2021. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.633, + 0.482, + 0.716 + ], + "angle": 0, + "content": "[33] Nataniel Ruiz, Yuanzhen Li, Varun Jampani, Yael Pritch, Michael Rubinstein, and Kfir Aberman. Dreambooth: Fine tuning text-to-image diffusion models for subject-driven generation. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 22500-22510, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.718, + 0.482, + 0.786 + ], + "angle": 0, + "content": "[34] Ruoxi Shi, Hansheng Chen, Zhuoyang Zhang, Minghua Liu, Chao Xu, Xinyue Wei, Linghao Chen, Chong Zeng, and Hao Su. Zero123++: a single image to consistent multi-view diffusion base model. arXiv preprint arXiv:2310.15110, 2023. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.789, + 0.482, + 0.871 + ], + "angle": 0, + "content": "[35] Xiaoyu Shi, Zhaoyang Huang, Fu-Yun Wang, Weikang Bian, Dasong Li, Yi Zhang, Manyuan Zhang, Ka Chun Cheung, Simon See, Hongwei Qin, et al. Motion-i2v: Consistent and controllable image-to-video generation with explicit motion modeling. In ACM SIGGRAPH 2024 Conference Papers, pages 1-11, 2024. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.873, + 0.482, + 0.902 + ], + "angle": 0, + "content": "[36] Yujun Shi, Chuhui Xue, Jun Hao Liew, Jiachun Pan, Hanshu Yan, Wenqing Zhang, Vincent YF Tan, and Song Bai." + }, + { + "type": "list", + "bbox": [ + 0.093, + 0.093, + 0.483, + 0.902 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.545, + 0.093, + 0.905, + 0.149 + ], + "angle": 0, + "content": "Dragdiffusion: Harnessing diffusion models for interactive point-based image editing. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 8839-8849, 2024. 1, 2, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.151, + 0.905, + 0.206 + ], + "angle": 0, + "content": "[37] Luming Tang, Menglin Jia, Qianqian Wang, Cheng Perng Phoo, and Bharath Hariharan. Emergent correspondence from image diffusion. Advances in Neural Information Processing Systems, 36:1363-1389, 2023. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.208, + 0.905, + 0.29 + ], + "angle": 0, + "content": "[38] Vikram Voleti, Chun-Han Yao, Mark Boss, Adam Letts, David Pankratz, Dmitry Tochilkin, Christian Laforte, Robin Rombach, and Varun Jampani. Sv3d: Novel multi-view synthesis and 3d generation from a single image using latent video diffusion. In European Conference on Computer Vision, pages 439-457. Springer, 2025. 3, 6, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.292, + 0.905, + 0.36 + ], + "angle": 0, + "content": "[39] Xiang Wang, Hangjie Yuan, Shiwei Zhang, Dayou Chen, Jiuniu Wang, Yingya Zhang, Yujun Shen, Deli Zhao, and Jingren Zhou. Videocomposer: Compositional video synthesis with motion controllability. Advances in Neural Information Processing Systems, 36, 2024. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.363, + 0.905, + 0.432 + ], + "angle": 0, + "content": "[40] Zhouxia Wang, Ziyang Yuan, Xintao Wang, Yaowei Li, Tianshui Chen, Menghan Xia, Ping Luo, and Ying Shan. Motionctrl: A unified and flexible motion controller for video generation. In ACM SIGGRAPH 2024 Conference Papers, pages 1-11, 2024. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.434, + 0.905, + 0.515 + ], + "angle": 0, + "content": "[41] Yuxiang Wei, Yabo Zhang, Zhilong Ji, Jinfeng Bai, Lei Zhang, and Wangmeng Zuo. Elite: Encoding visual concepts into textual embeddings for customized text-to-image generation. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 15943-15953, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.519, + 0.905, + 0.602 + ], + "angle": 0, + "content": "[42] Tong Wu, Jiarui Zhang, Xiao Fu, Yuxin Wang, Jiawei Ren, Liang Pan, Wayne Wu, Lei Yang, Jiaqi Wang, Chen Qian, et al. Omniobject3d: Large-vocabulary 3d object dataset for realistic perception, reconstruction and generation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 803-814, 2023. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.604, + 0.905, + 0.685 + ], + "angle": 0, + "content": "[43] Weijia Wu, Zhuang Li, Yuchao Gu, Rui Zhao, Yefei He, David Junhao Zhang, Mike Zheng Shou, Yan Li, Tingting Gao, and Di Zhang. Draganything: Motion control for anything using entity representation. In European Conference on Computer Vision, pages 331-348. Springer, 2025. 1, 3, 5, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.688, + 0.905, + 0.757 + ], + "angle": 0, + "content": "[44] Haibo Yang, Yang Chen, Yingwei Pan, Ting Yao, Zhineng Chen, and Tao Mei. 3dstyle-diffusion: Pursuing fine-grained text-driven 3d stylization with 2d diffusion models. In Proceedings of the 31st ACM International Conference on Multimedia, pages 6860-6868, 2023. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.759, + 0.905, + 0.815 + ], + "angle": 0, + "content": "[45] Shengming Yin, Chenfei Wu, Jian Liang, Jie Shi, Houqiang Li, Gong Ming, and Nan Duan. Dragnuwa: Fine-grained control in video generation by integrating text, image, and trajectory. arXiv preprint arXiv:2308.08089, 2023. 1, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.817, + 0.905, + 0.871 + ], + "angle": 0, + "content": "[46] Ziyang Yuan, Mingdeng Cao, Xintao Wang, Zhongang Qi, Chun Yuan, and Ying Shan. Customnet: Zero-shot object customization with variable-viewpoints in text-to-image diffusion models. arXiv preprint arXiv:2310.19784, 2023. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.873, + 0.905, + 0.902 + ], + "angle": 0, + "content": "[47] Lvmin Zhang, Anyi Rao, and Maneesh Agrawala. Adding conditional control to text-to-image diffusion models. In" + }, + { + "type": "list", + "bbox": [ + 0.517, + 0.093, + 0.905, + 0.902 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.48, + 0.945, + 0.518, + 0.957 + ], + "angle": 0, + "content": "21511" + } + ], + [ + { + "type": "ref_text", + "bbox": [ + 0.125, + 0.092, + 0.482, + 0.12 + ], + "angle": 0, + "content": "Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 3836-3847, 2023. 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.122, + 0.483, + 0.177 + ], + "angle": 0, + "content": "[48] Yabo Zhang, Yuxiang Wei, Dongsheng Jiang, Xiaopeng Zhang, Wangmeng Zuo, and Qi Tian. Controlvideo: Training-free controllable text-to-video generation. arXiv preprint arXiv:2305.13077, 2023. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.179, + 0.482, + 0.233 + ], + "angle": 0, + "content": "[49] Zhenghao Zhang, Junchao Liao, Menghao Li, Long Qin, and Weizhi Wang. Tora: Trajectory-oriented diffusion transformer for video generation. arXiv preprint arXiv:2407.21705, 2024.3" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.235, + 0.482, + 0.29 + ], + "angle": 0, + "content": "[50] Chuanxia Zheng and Andrea Vedaldi. Free3d: Consistent novel view synthesis without 3d representation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 9720-9731, 2024. 3" + }, + { + "type": "list", + "bbox": [ + 0.093, + 0.092, + 0.483, + 0.29 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.48, + 0.946, + 0.519, + 0.956 + ], + "angle": 0, + "content": "21512" + } + ] +] \ No newline at end of file diff --git a/2025/Dragin3D_ Image Editing by Dragging in 3D Space/63363afe-acf7-462c-bf4b-9a484c6ae8ae_origin.pdf b/2025/Dragin3D_ Image Editing by Dragging in 3D Space/63363afe-acf7-462c-bf4b-9a484c6ae8ae_origin.pdf new file mode 100644 index 0000000000000000000000000000000000000000..7b7a89122ae08525783ab46b8603df7c8ec5dfc6 --- /dev/null +++ b/2025/Dragin3D_ Image Editing by Dragging in 3D Space/63363afe-acf7-462c-bf4b-9a484c6ae8ae_origin.pdf @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:aa07e799922b835239d5c6166ec1941ef5c94b43130af59d1315f9fda14f0e65 +size 9629465 diff --git a/2025/Dragin3D_ Image Editing by Dragging in 3D Space/full.md b/2025/Dragin3D_ Image Editing by Dragging in 3D Space/full.md new file mode 100644 index 0000000000000000000000000000000000000000..0aa5a21bdb9d6da580e64449cc76b23464012c1f --- /dev/null +++ b/2025/Dragin3D_ Image Editing by Dragging in 3D Space/full.md @@ -0,0 +1,289 @@ +# Dragin3D: Image Editing by Dragging in 3D Space + +Weiran Guang $^{12}$ Xiaoguang Gu $^{2*}$ Mengqi Huang $^{1}$ Zhendong Mao $^{1}$ + +1University of Science and Technology of China + +$^{2}$ State Key Laboratory of Communication Content Cognition, People's Daily Online + +guangwr, huangmq}@mail.ustc.edu.cn, guxiaoguang@people.cn, zdmao@ustc.edu.cn + +# Abstract + +Interactive drag editing of images is a valuable task that has gained considerable attention for its precision and controllability. However, existing approaches have primarily focused on manipulating the shape or movement of objects in 2D plane. We propose to extend this drag-based editing task to 3D space. Firstly, we utilize the trajectory of two points to represent the rotational trajectory of the object. Gaussian maps of a circle and a square are centered at these two points, respectively. We use distinct shapes to ensure that symmetric views produce different object representations. Secondly, we introduce a lightweight mapping network to embed the object features into two Gaussian maps to obtain a continuous control condition that guides the model in learning the correspondence between the trajectory and the object. Finally, to overcome the limitations of current 3D object reconstruction datasets, which typically consist of object maps with transparent backgrounds, we affix random backgrounds to them. This modification helps improve the model's ability to ignore background interference when editing real images with complex backgrounds. Experiments demonstrate that our approach successfully achieves object rotation within the drag framework and demonstrates strong generalization to real-world images. + +# 1. Introduction + +Drag-based image editing has garnered substantial research interest in recent years, aiming to enable the manipulation of objects via user-defined handle and target points [30, 36]. Most existing image editing works are text-based, which heavily depends on textual and semantic alignment. Sometimes, it is challenging to articulate user-desired edits in words. Drag-based editing offers a more intuitive, interactive, and granular editing signal, allowing for precise manipulation at the pixel level. Users specify the start and end points based on intended edits, eliminating the need for a + +proper prompt. This enables user-friendly image modification and offers more precise control over positioning, making it valuable for practical applications. + +Existing drag-based editing methods can be categorized into optimized-based and training-based. DragGAN [30] is an early representative of optimized-based methods, encompassing two key components: point tracking and motion supervision. The latent code of StyleGAN [13] is optimized while updating the position of the handle point, ensuring that the features near the target points align with those near the handle points. Since DragGAN was proposed, both drag-based image and video editing techniques have been significantly advanced. Subsequently, researchers extended this process to diffusion models by optimizing the features of UNet to break through the inherent limitation of the GAN model [27, 36]. Recent studies have focused on addressing issues related to inaccurate point tracking and incomplete motion supervision, enabling more stable dragging by formulating better point tracking and motion supervision protocols [7, 19]. Furthermore, to overcome the limitations posed by variations in object motion patterns, the training-based method [43, 45] has been introduced using trajectory guidance, where the motion trajectory of the object is encoded as a control condition. In conclusion, ongoing efforts aim to enhance editing controllability and fidelity. + +However, existing drag-based methods are limited to 2D drag editing and fail to achieve drag in 3D space, which significantly limits their practical application in the real 3D world. 3D generation has significantly expanded the application in games, animation, virtual production, etc. It enables more realistic creation of 3D visual spaces that closely resemble the physical world. We consider whether it is possible to generate novel view images of the object through simple point-drag interactions. The two drag paradigms discussed above are inadequate for solving this task. The optimization-based approach primarily seeks to supervise feature consistency between the handle point and the target point, which cannot well control the substantial changes in features that arise due to variations in spatial perspective, as + +![](images/91d03067ba77be0f3c380cf0c1ce35f1e9c1dc5859c593528bbd75c7f4777e22.jpg) +Figure 1. Existing drag method for object rotation representation. (a) The optimization-based method. The sculpture makes tiny rotations, and changes in the position of the target point still tend to maintain a frontal perspective, which fails to achieve 3D editing with varying amplitudes based solely on the dragging intensity. (b) The training-based method. In the upper figure, the object moves along the trajectory of the dragging point. In the lower figure, the object rotates in the direction of the trajectory of the dragging point. The trajectory is inclined upwards, with the ship exhibiting a tendency to rotate towards the target position. Arguably, whether the ship can achieve 3D rotation is entirely random. + +illustrated in Fig. 1(a). On the other hand, the trajectory-based methods introduce ambiguity in the trend of object motion, as demonstrated in Fig. 1(b). Therefore, in this study, we introduce a new drag-based object rotation task. + +The key challenges of this task lie in how to accurately represent and track the desired 3D editing trajectory from a 2D image, and how to establish a correspondence between object motion trajectories and their respective representations. Due to the ambiguity in existing training-based methods for single-point trajectory representation, combined with the feature alignment principle of optimization-based methods, we propose Dragin3D, which extends drag-based editing to 3D space for the first time. We introduce a double-point representation approach that associates object trajectories with image features. Given an image of an object and a double-point rotation trajectory, Dragin3D generates a novel view image of the object from the corresponding trajectory, as shown in Fig. 2. The core concept of Dragin3D is to precisely capture the direction and magnitude of the object's rotation using the double-point trajectory. Specifically, our method consists of two main components: (1) Drag Trajectory Representation (DTR): we map the rotational trajectory of the object onto the rotational trajectory of a line segment, track the two endpoints of this segment, and project them into Gaussian maps with different shapes; (2) Control Condition Inject Module (CCIM): to enhance fidelity to the object, we extract image features through a lightweight mapping network and inject these features into the corresponding Gaussian maps. + +Our contributions are summarized as follows: + +- We propose an innovative interactive image editing method Dragin3D. To the best of our knowledge, we are + +![](images/c22e3b4194ee56c01cc287d1fc6892a7586d85a68b7d4953b53ca8781b6f5721.jpg) +Figure 2. Drag results of our proposed paradigm Dragin3D. Given an input image and coordinates of two representation points, i.e., $[P_{11}, P_{21}],[P_{12}, P_{22}],\ldots,[P_{1n}, P_{2n}]$ . Dragin3D can generate a novel view image of the object for the corresponding trajectory. + +the first to achieve multi-view image generation of objects by dragging them in 3D space based on a diffusion model. + +- We propose to represent the object rotation trajectory by a double-point motion trajectory, which maps 3D object motion onto a 2D plane and eliminates the ambiguity of the object motion mode. +- We introduce a tracking point projection method, which projects the tracking points onto Gaussian maps and aligns them with the object's feature information. +- Experimental results demonstrate that despite being trained on a synthetic 3D dataset, our method generalizes well beyond the dataset's distribution. + +# 2. Related Work + +# 2.1. Drag-based controllable image editing + +In the field of image editing, several established works have focused on tasks such as image customization [16, 18, 33, 41], image editing [1, 4, 6, 9, 11, 14], which modify object attributes, quantity, etc. through text-based operations. Due to the limitations of text-based editing in controlling local changes at a fine-grained level, the Drag approach introduces a novel approach to image editing [7, 8, 19, 23, 24, 27-30, 36]. This interactive manipulation allows users to precisely designate both the editing area and control mode with a simple click. [30, 36] propose a point-based editing technique that integrates point tracking and motion supervision. [19] enhances drag editing stability and reliability without requiring precise point tracking by introducing adaptively updated template features, fuzzy localization, and linear search techniques. [7] introduces a discriminative point tracking method and a confidence-based latent space enhancement strategy, resulting in more stable and accurate drag performance. These works primarily focus on enhancing source and target feature optimization. However, when extended to object rotation manipulation, significant feature differences across views pose a challenge that pure feature alignment cannot resolve. + +![](images/0c05505f6758dfeddf97b8869d416d05b027f8d35278371b092d2e3c2d852e77.jpg) +Figure 3. Dragin3D Architecture. We track the representation points by calculating the transfer matrix and plot circle and square Gaussian maps centered on two representation points, respectively. We encode the image to be edited to extract the image features, which are injected into the original encoded Gaussian map features to obtain new Gaussian map features. The original Gaussian map features are added to the newly injected Gaussian map features as an input control condition for ControlNet. + +# 2.2. Trajectory-based controllable image editing + +Recently, trajectory-based motion control [10, 25, 35, 39, 40, 43, 45, 49] has emerged as an efficient and user-friendly method for controllable video generation [5, 12, 15, 48], with drag editing serving as a trajectory-based task. The gradual movement of handle points forms a continuous trajectory, providing a more precise guidance signal in contrast to defined start and end points. [45] introduces a unified control framework that integrates text, image, and trajectory data, utilizing a trajectory sampler to extract trajectories from open-domain video streams. [43] extracts features from the diffusion model as entity representations, and incorporates them into a denoised 3D U-Net with a noisy 2D Gaussian representation. [10] focuses on specific camera motion, generating videos by adding temporal constraints to a T2I model to capture motion dynamics, leveraging eight independent LoRA models for basic camera control. [40] differentiates between camera and object motions, enabling independent or combined control of the camera pose matrix and object trajectory to produce personalized videos. However, these motions are limited to 2D translations, making it difficult to generate novel views of objects in 3D space. + +# 2.3. 3D generation Models + +Based on the impressive generative effect of diffusion models on 2D images, the approach of 3D generation via diffusion models has emerged [17, 20-22, 31, 34, 44, 50]. [31] is one of the pioneering methods, introduces the SDS loss function, which enables the updating of NeRF [26] weights by minimizing the discrepancy between noise levels. Several 3D image generation methods using enhanced SDS loss functions have been proposed, employing a 2D + +generative model to create 3D structures from textual descriptions without relying on 3D data, thus overcoming the challenge of limited 3D data availability. [44] integrates the diffusion process of image rendering and synthesis using NeRF, ensuring that the generated images remain semantically aligned with the text and geometrically consistent with the depth map during training. [38, 46] directly embed camera parameters as control conditions in the model to realize the camera pose control. However, the direct embedding of discrete values, such as camera parameters, as control conditions often lack the necessary precision. + +# 3. Methods + +# 3.1. Task Definition & Preliminaries + +3D Drag Editing Definition. We observed that previous drag tasks lacked a unified framework for moving objects in 3D space by drag interaction, for example, object rotation, so we propose to extend the task to 3D space. This task marks a significant advancement beyond merely editing the appearance of objects in a 2D image. Given a reference object image, our approach allows users to obtain a new perspective of the object through a simple drag interaction. + +Preliminary of Stable Video Diffusion. Our paradigm is based on Stable Video Diffusion (SVD) [2] due to its robust predictive capabilities for dynamically moving objects. The key modification from Stable Diffusion (SD) [3] is the integration of a temporal layer into the denoising model's U-Net architecture, resulting in a 3D-UNet. Compared to SD trained on image datasets, video frames capture more dynamic information about objects, leading to more sensitivity to morphological changes of the same objects, ensur + +ing perspective and temporal consistency of rotation. + +Preliminary of ControlNet. ControlNet [47] is one of the representative works in the field of controlled image generation, which supports a variety of inputs such as edge maps, segmentation maps, keypoints, etc., as conditional inputs to guide the image generation process. $\mathcal{F}(\cdot ;\Theta)$ is a neural network block with parameters $\Theta$ . ControlNet locks all parameters in $\Theta$ and clones them into a trainable copy $\Theta_{c}$ . The output of ControlNet $y_{c}$ can be expressed as: + +$$ +\boldsymbol {y} _ {\mathbf {c}} = \mathcal {F} (\boldsymbol {x}; \Theta) + \mathcal {Z} \left(\mathcal {F} \left(\boldsymbol {x} + \mathcal {Z} \left(\boldsymbol {c}; \Theta_ {\mathrm {z} 1}\right); \Theta_ {\mathrm {c}}\right); \Theta_ {\mathrm {z} 2}\right), \tag {1} +$$ + +where $c$ is the condition input, $\mathcal{Z}(\cdot ;\cdot)$ is a $1\times 1$ convolutional layer initialized with zero weights and bias. When ControlNet is used in Stable Diffusion, the training objective is: + +$$ +\mathcal {L} = \mathbb {E} _ {\boldsymbol {z} _ {0}, t, \boldsymbol {c} _ {t}, \boldsymbol {c} _ {\mathrm {f}}, \epsilon \sim \mathcal {N} (0, 1)} \left[ \| \epsilon - \epsilon_ {\theta} \left(z _ {t}, t, \boldsymbol {c} _ {t}, \boldsymbol {c} _ {\mathrm {f}}\right)) \| _ {2} ^ {2} \right], \tag {2} +$$ + +where $z_0$ denotes the input image, $z_t$ denotes the noisy image in timestep $t$ , $c_t$ and $c_f$ denote text prompts and the task-specific conditions, respectively. + +Preliminary of Gaussian heatmap. Gaussian heatmaps are commonly used for keypoint detection tasks. A heatmap is a feature map of the same size as the input, with a value of 1 at a specified coordinate (i.e., the representation points in our task) and 0 at all other positions. Additionally, to align with inputs like mask or depth maps used in ControlNet, a Gaussian heatmap is a suitable input, which effectively converts discrete point signals into continuous image signals. + +# 3.2. Overall Framework + +For this novel dragging in 3D space task, we propose a framework named Dragin3D, which incorporates two key innovations: 3D drag trajectory representation and control condition injection. The overall workflow of our model is illustrated in Fig. 3. We employ an architecture that combines SVD and ControlNet. Firstly, we propose to describe the motion trajectory of the object with two points, and compute to track the positions of the representation points in different views. Next, we construct Gaussian maps of different shapes centered on these two points. These maps are then encoded, and image feature representations are injected into the Gaussian maps to align them with the corresponding point locations. Finally, the encoded Gaussian map features are linked to new Gaussian map features that incorporate object-specific information, which are used as control condition inputs for ControlNet. + +# 3.3. 3D Drag Trajectory Representation (DTR) + +To extend the editing task to 3D space, we propose a 3D Drag Trajectory Representation (DTR) method, which explicitly represents the object's trajectory using a rotational + +![](images/0d4c1ccf8db16bad9ec345f478fd54deda6c06588911a99a6d8955dcfeaa3f0e.jpg) +# + +![](images/7f33791087c9f9c7b1c483af2abc33c60a6ac27b798a947b6cca073d3314783d.jpg) +# +Figure 4. Our 3D drag trajectory representation method. The computed tracking points are projected into Gaussian maps, ensuring the rotational trajectory of the object is visually aligned with the trajectory of the line segment formed by the representation points. + +![](images/104c231831aee4d0864d1d1bd7bb9e5ce9c25f1fb0e8078ca19af6ec8996d5a3.jpg) +# + +![](images/494d1e506e3b8aefecb56841ab7e6c9231e4718119908cd6eac2e54b3308e229.jpg) +# + +line segment defined by two points. In general, an object's rotation in 3D space involves variations around three coordinate axes across three planes. Since left-right rotation is more intuitive and commonly used at the operational level, our approach and experiments focus on the z-axis rotation. + +The selection and stable tracking of the representation points are critical, as they directly affect the accuracy of the subsequent results. To ensure that the motion trajectory of the line segment is coherent and clearly defined, we propose a method for calculating the position of a specified coordinate point in images captured from different views. Then we project the computed points onto the 2D plane, as shown in Fig. 4. The required camera intrinsic and extrinsic parameters are provided by the 3D dataset. Extrinsic parameter calibration is performed by solving the transfer matrix using the real-world coordinates and pixel coordinates of fixed points. Conversely, when the camera extrinsic parameter is available, we can calculate the pixel coordinates of a fixed point by maintaining the constancy of its world coordinates. To simplify the calculation of 3D coordinates, we assume that the z-axis coordinate of each pixel point is fixed to 1. While this assumption introduces some deviation, the resulting calculation provides a reasonable approximation of the trajectory and magnitude of the object's movement. The relevant calculation formulas are as follows. + +$$ +x _ {c 1} = \left(x _ {1} - c _ {x}\right) * z _ {c 1} / f _ {x}, y _ {c 1} = \left(y _ {1} - c _ {y}\right) * z _ {c 1} / f _ {y}, \tag {3} +$$ + +$$ +\begin{array}{l} \left[ \begin{array}{l l l} x _ {w} & y _ {w} & z _ {w} \end{array} \right] ^ {- 1} = T _ {1} ^ {- 1} \cdot \left[ \begin{array}{l l l} x _ {c 1} & y _ {c 1} & z _ {c 1} \end{array} \right] ^ {- 1}, \\ \left[ \begin{array}{c} x _ {c 2} y _ {c 2} z _ {c 2} 1 \end{array} \right] ^ {- 1} = T _ {2} \cdot \left[ \begin{array}{c} x _ {w} y _ {w} z _ {w} 1 \end{array} \right] ^ {- 1}, \\ \end{array} +$$ + +$$ +x _ {2} = f _ {x} * x _ {c 2} / z _ {c 2} + c _ {x}, y _ {2} = f _ {y} * y _ {c 2} / z _ {c 2} + c _ {y}, \tag {5} +$$ + +where the intrinsic $(f_{x},f_{y},c_{x},c_{y})$ and extrinsic $(T_{1},T_{2})$ of the camera are labeled in the dataset. We define $P_{1}$ $(x_{1},y_{1})$ as the pixel coordinates of the specified point in the first image, $P_{c1}$ $(x_{c1},y_{c1},z_{c1})$ is the camera coordinates of the specified point in the first image, $P_{w}$ $(x_{w},y_{w},z_{w})$ is the world coordinates, $P_{c2}$ $(x_{c2},y_{c2},z_{c2})$ and $P_{2}$ $(x_{2},y_{2})$ are + +![](images/acfcb0464ad9529309c26ef4376e3998770dcf4a098a096aac88af7eda717cb1.jpg) +Figure 5. Dragin3D Visualization. Given an input image and a set of trajectory point pairs, we obtained Gaussian maps of the trajectories as input (first column), which is then used to generate the corresponding objects at the specified trajectory points (subsequent columns). The arrows indicate rotation directions, the red line is the current view of the object, the gray dotted line shows the trajectory that has already been traversed, and the gray solid line is the future trajectory. + +the camera coordinates and pixel coordinates of the specified point in the second image, respectively. + +# 3.4. Control Condition Inject Module (CCIM) + +In order to process the existing input image and the computed trajectory representations into suitable input conditions for the model, we proposed a Control Condition Inject Module (CCIM). Drawing inspiration from DragAnything [43], which uses discrete pixel points as input control conditions by mapping them into a Gaussian map. We propose to use two representation points. However, we observed that the two representation points of the object are nearly identical in the front and rear views, as well as in the symmetric left and right side views. We propose a Shape DistinGuish Strategy (SDGS) to distinguish the two points by their respective shapes. Specifically, we create circular and square Gaussian maps centered on the two points, respectively, as shown in Fig. 9 (green boxes). Then the Gaussian map images $x \in \mathbb{R}^{F \times H \times W \times 3}$ are encoded into image features $f_{gi} \in \mathbb{R}^{F \times H \times W \times C}$ , where $F$ denotes the input frames and $C$ denotes the channel size. + +CCIM. Since multi-view image generation requires the guidance of input object representation information, the Gaussian map that only indicates the location and trajectory of the object is insufficient. To establish a correspondence between the location of the representation points and the image features, we propose to use a trainable image prompt mapping network, which allows the image to function as an image prompt similar to the textual guidance condition. In order to ensure consistency between training and inference, we use only the first frame image $x_{1}$ during training. The first frame image embedding is extracted by the pre + +trained CLIP [32] image encoder, and our image prompt mapping network projects this CLIP embedding into an image prompt feature $f_{ip} \in \mathbb{R}^{1 \times C}$ . Then we replace the features at the representation points' locations in the Gaussian map image features $f_{gi}$ with this sequence of image prompt features. The resulting new Gaussian map image features are denoted as $f_{ngi} \in \mathbb{R}^{F \times H \times W \times C}$ . It is worth noting that since the start and end frames are randomly selected during training, all images in the training set with different views have the potential to be encoded as image prompts. This ensures diversity and richness in the model's ability to learn from varying perspectives. + +Through the above analysis, the input condition of ControlNet is the sum of $f_{gi}$ and $f_{ngi}$ , that is $f_{input} = f_{gi} + f_{ngi}$ . During training, we only optimize parameters of the image prompt mapping network and ControlNet, and the final optimization objective is: + +$$ +\min _ {\theta} E _ {x _ {0}, \epsilon , t} \| \epsilon - \epsilon_ {\theta} (x _ {t}, t, c (G, G _ {n e w})) \| ^ {2}, \tag {6} +$$ + +where $\epsilon \sim \mathcal{N}(0,I)$ , $t\in [0,T]$ denotes the time step of diffusion process, $c(\cdot)$ is the condition mechanism with explicit controls. $G$ and $G_{new}$ denote the initial and new Gaussian map injected with image features, respectively. + +# 4. Experiments + +# 4.1. Experimental Setup + +Datasets. We utilize the widely adopted OmniObject3D dataset [42], containing approximately 6k 3D object models in real-world across more than 200 categories. We selected 2.4k 3D objects for training and 200 for testing. To + +![](images/ed21d843db3387f12549fc69d3736aa8f519fc02b25d89a252b3ce517a7dc5ad.jpg) +Figure 6. Comparison with existing methods. We compare existing methods relevant to our task. For 3D dataset images with pose labels, we directly compute the representation points, with the first row representing the ground truth image and its corresponding rotation. For real images, we utilize the interaction interface to define the desired drag target points ("Our drag points" in the figure). Additionally, we specify a drag trajectory for DragAnything that closely resembles our trajectory. DragAnything interprets the trajectory as the movement of the object or camera along it but cannot handle rotations. SV3D struggles with generalization to real-world images and lacks sufficient detail and 3D consistency. Wonder3D generates higher-quality 3D images but is limited to six fixed views. + +![](images/ef6f849b22b045db15210aa77528e21409ec34fc9c849efbf1a77354d023229c.jpg) + +![](images/94053b6f4b9348ff728fb0f35ca8673251199aab5fc1cc9161cd539fb8f349d7.jpg) + +
MethodsMulti-viewBackgroundTime
DragAnything×✓~40s
SV3D✓×~2min
Wonder3D✓×~2min
Ours✓✓~25s
+ +Table 1. Functionality comparisons. Our proposed Dragin3D achieves multi-view generation and supports editing images with backgrounds, ensuring efficient and rapid editing with minimal time consumption. In contrast, methods from different areas are unable to perform these functions simultaneously. + +better simulate real-world scenes, we incorporate diverse background images into the dataset. Additionally, the image resolution is resized to $256*256$ to optimize computational efficiency during training. + +Baselines. Since Dragin3D integrates drag with multiview tasks, we use DragAnything [43], SV3D [38], and Wonder3D [22] as baselines. DragAnything and SV3D are trained on video framework, Wonder3D is trained on image framework. DragAnything is a 2D interaction editing method trained on video datasets, SV3D and Wonder3D are 3D generation methods trained on multi-view datasets. + +Evaluation Metrics. For multi-view generation evaluation, we use the following three metrics for quantitative evaluation: For the quality of generated images, we employ Fréchet Inception Distance (FID), which calculates the + +Frechet distance between the feature vectors of the generated image and ground truth. For the perceptual similarity of the 3D effects, we use Learned Perceptual Image Patch Similarity (LPIPS), which evaluates the perceptual similarity between the generated image and the ground truth by learning the inverse mapping of the generated image. For the quality of the generated motion sequences, we apply Frechet Video Distance (FVD), which leverages a 3D feature extractor to compare video sequences. For drag trajectory accuracy evaluation, we adopt the MD metric in DragDiffusion, which employs DIFT [37] to identify the points in the images dragged by specified trajectories, corresponding to the starting representation points. The average Euclidean distance between the identified points and the reference trajectory points is then computed. + +Implements Details. We randomly select 8 coherent object rotation images from the OmniObject3D dataset, which are used as the individual frames of video input to the SVD architecture. Eight A800 GPUs are used to train 20k steps with a batch size of 6, the learning rate is set to 2e-5. + +# 4.2. Qualitative Results + +The visualization results of Dragin3D are shown in Fig. 5. Given an image with a corresponding trajectory, Dragin3D generates a high-fidelity image of the object from a new perspective, maintaining strong consistency with the original image. In addition, Dragin3D demonstrates superior generalization performance, which successfully rotates ob + +![](images/1f0343b0fc7610f981d1a355897bd884ff149e2a267a918b014d68aa57fb61f5.jpg) +Figure 7. Comparison with 3D methods. We use the image without background as input to 3D methods and compare the results with ground truth, demonstrating that our method is comparable. + +jects by dragging the representation points on real images that lie outside the training dataset distribution. + +As shown in Fig. 7, we input the image into SV3D and Wonder3D after background removal and compare the results with the ground truth from the corresponding viewpoints. Our methods demonstrate comparable performance in terms of rotation accuracy and 3D consistency. Since our main contribution lies in the generalizability to real-world images, we include the background to reduce the gap between the ideal 3D dataset and the real-world scene. As illustrated in Fig. 6, we present the visualization results of each method with background using both 3D datasets and real images. The real images we provided have homogeneous backgrounds, minimizing their impact on the object. The results show that the performance of the other methods is significantly lower, limiting their practical applicability, while our method demonstrates a one-stage process that eliminates background interference during the image processing stage, performs well in detail preservation and coherent rotation. Additionally, we found that DragDiffusion [36] can achieve partial rotation in certain cases of Drag-Bench. However, the rotation is minimal and random, lacking precise control over the direction and angle of rotation as specified by the instructions, as illustrated in Fig. 8. + +# 4.3. Quantitative Results + +Since our proposed task diverges from traditional drag-based editing and produces significantly different outcomes, and 3D generation models like Wonder3D focus on generating images without background from partial views, we do not include them in a quantitative comparison. Our Dragin3D shares a similar goal with SV3D, which aims to generate images from new perspectives based on a single input image. Therefore, we chose to make a comparison with SV3D. Due to the absence of ground truth in the real-world data, we randomly selected 200 subsets from the OmniObject3D dataset as the testset for computation, which were + +![](images/861afd2f4b5a8d5651b02fab108b6df5ca28e4ae4d7a86e3527af9502c7e2315.jpg) +Figure 8. Comparison of rotation cases with DragDiffusion. DragDiffusion primarily manifests as tiny rotations induced by movements in the plane towards the target point, while Dragin3D enables more accurate control of stereoscopic rotation. + +
MethodsFID↓LPIPS↓FVD↓
SV3D [38]52.50.410944.6
Dragin3D (Ours)43.40.179584.6
+ +Table 2. Quantitative comparisons. Our proposed Dragin3D outperforms SV3D in all metrics. + +
MethodsDragDiffusion [36]Dragin3D (Ours)
MD↓33.019.38
+ +Table 3. Quantitative comparisons. Our proposed Dragin3D outperforms DragDiffusion in control accuracy evaluation. + +not visible during training. We compared the similarity between the generated images and ground truth across the corresponding viewpoints from three perspectives: image, 3D perceptual, and video. As shown in Tab. 2, Dragin3D outperforms SV3D in all metrics. To evaluate the accuracy of trajectory control, we selected object rotation cases in Drag-Bench including animal, human, and sculpture heads. For each set of desired rotation trajectories, we computed the MDs for both Dragdiffusion and Dragin3D, with Dragin3D outperforming Dragdiffusion, as shown in Tab. 3. Additionally, we conducted a user study in Supp. 7.2. + +# 4.4. Ablations + +Effectiveness of shape distinguish strategy (SDGS). To demonstrate the effectiveness of SDGS, we designed experiments in which the Gaussian maps for two points were constrained to the same shape, as shown in Tab. 4. We specifically focused on the symmetric rotation of objects, as illustrated in Fig. 9. Without distinguishing between the two points, the object in the subsequent stage of the rotation becomes difficult to distinguish from its symmetrical configuration in the preliminary stage. In contrast, the generated object preserves structural integrity and details when SDGS is applied. More examples can be found in Supp. 7.3. + +Effectiveness of control condition inject module + +
MethodsFID↓LPIPS↓FVD↓
Dragin3D w/o SDGS44.10.182676.8
Dragin3D w/o CCIM45.20.194694.1
Dragin3D43.40.179584.6
+ +Table 4. Ablation study results. All metrics have declined without SDGS and CCIM, respectively. + +![](images/31c4543317ea4d6d144fbba6b8bc5abfcdf2a0fbceea39a314241bde2c7daa47.jpg) +Figure 9. Effectiveness of shape distinguish strategy (SDGS). We present the resulting images for two cases. The first row displays the trajectory points projected onto Gaussian maps, while the second and third rows show the objects from the corresponding view of these trajectory points, generated with and without SDGS, respectively. The positions of the two trajectory points in the first frame closely correspond to those in the last frame (green box). In the latter half of the sequence, significant misalignment occurs due to interference from the Gaussian representation (brown box). + +(CCIM). As illustrated in Tab. 4, to assess the effectiveness of CCIM, we conducted experiments using only Gaussian map image features $f_{gi}$ derived from the tracking points as the control input. The results reveal a marked enhancement across all evaluation metrics upon the integration of CCIM. Visual comparisons of intermediate results are presented in Fig. 10. Without CCIM, the object's rotation initially exhibits low-amplitude fluctuations, followed by a sudden and drastic change. Specifically, the early frames retain the original image, while later frames generate objects with significant distortion and poor fidelity. In contrast, the rotation transition is smooth and stable, and the object's details are preserved throughout. These results demonstrate that CCIM effectively aligns the rotation trajectory with image features by incorporating object-specific information, ensuring that the generated perspectives remain faithful to the original image content. More examples can be found in Supp. 7.3. + +![](images/bd5319896f6c8f5d3d35247090c8b82f3ce1f284acefe9341986e29df39a9ed7.jpg) +Figure 10. Effectiveness of control condition inject module (CCIM). We present the resulting images for two cases. The first row displays the trajectory points projected onto Gaussian maps, while the second and third rows show the objects from the corresponding view of these trajectory points, generated with and without CCIM, respectively. Without CCIM, the image remains largely unchanged in the early stages (red box), but later undergoes a sudden distortion, causing significant fidelity loss (yellow box). + +# 5. Conclusion and Future Works + +In this paper, we present Dragin3D, which extends the interactive trajectory-based dragging task to 3D space. To represent the rotation trajectory of an object from the starting view to the terminating view, we introduce a double-point representation approach, and map the rotation of the object to a line segment formed by these two points. Our approach introduces control conditions through ControlNet architecture based on the large-scale video generation pretrained model SVD. In addition, we randomly add different backgrounds to the 3D object dataset, enabling the model to directly handle images with backgrounds. Through extensive experiments, Dragin3D successfully generates novel view images of the object based on the trajectories of two points, minimizing background interference while preserving the original object features, and demonstrating exceptional generalization on real-world images. For future work, since our approach currently supports object rotation in only one plane, we plan to expand our trajectory-based editing task to encompass a broader range of views, enabling arbitrary rotations in 3D space. + +# 6. Acknowledgments + +This research is supported by Artificial Intelligence-National Science and Technology Major Project 2023ZD0121200, the National Science Fund for Excellent Young Scholars under Grant 62222212, and National Natural Science Foundation of China under Grant 623B2094. + +# References + +[1] Omer Bar-Tal, Dolev Ofri-Amar, Rafail Fridman, Yoni Kasten, and Tali Dekel. Text2live: Text-driven layered image and video editing. In European conference on computer vision, pages 707-723. Springer, 2022. 2 +[2] Andreas Blattmann, Tim Dockhorn, Sumith Kulal, Daniel Mendelevitch, Maciej Kilian, Dominik Lorenz, Yam Levi, Zion English, Vikram Voleti, Adam Letts, et al. Stable video diffusion: Scaling latent video diffusion models to large datasets. arXiv preprint arXiv:2311.15127, 2023. 3 +[3] Andreas Blattmann, Robin Rombach, Huan Ling, Tim Dockhorn, Seung Wook Kim, Sanja Fidler, and Karsten Kreis. Align your latents: High-resolution video synthesis with latent diffusion models. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 22563-22575, 2023. 3 +[4] Tim Brooks, Aleksander Holynski, and Alexei A Efros. Instructpix2pix: Learning to follow image editing instructions. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 18392-18402, 2023. 2 +[5] Weifeng Chen, Yatai Ji, Jie Wu, Hefeng Wu, Pan Xie, Jiashi Li, Xin Xia, Xuefeng Xiao, and Liang Lin. Control-a-video: Controllable text-to-video generation with diffusion models. arXiv preprint arXiv:2305.13840, 2023. 3 +[6] Guillaume Couairon, Jakob Verbeek, Holger Schwenk, and Matthieu Cord. Diffedit: Diffusion-based semantic image editing with mask guidance. arXiv preprint arXiv:2210.11427, 2022. 2 +[7] Yutao Cui, Xiaotong Zhao, Guozhen Zhang, Shengming Cao, Kai Ma, and Limin Wang. Stabledrag: Stable dragging for point-based image editing. arXiv preprint arXiv:2403.04437, 2024.1, 2 +[8] Yuki Endo. User-controllable latent transformer for stylegan image layout editing. In Computer Graphics Forum, pages 395-406. Wiley Online Library, 2022. 2 +[9] Dave Epstein, Allan Jabri, Ben Poole, Alexei Efros, and Aleksander Holynski. Diffusion self-guidance for controllable image generation. Advances in Neural Information Processing Systems, 36:16222-16239, 2023. 2 +[10] Yuwei Guo, Ceyuan Yang, Anyi Rao, Zhengyang Liang, Yaohui Wang, Yu Qiao, Maneesh Agrawala, Dahua Lin, and Bo Dai. Animatediff: Animate your personalized text-to-image diffusion models without specific tuning. arXiv preprint arXiv:2307.04725, 2023. 3 +[11] Amir Hertz, Ron Mokady, Jay Tenenbaum, Kfir Aberman, Yael Pritch, and Daniel Cohen-Or. Prompt-to-prompt im + +age editing with cross attention control. arXiv preprint arXiv:2208.01626, 2022. 2 +[12] Yaosi Hu, Chong Luo, and Zhenzhong Chen. Make it move: controllable image-to-video generation with text descriptions. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 18219-18228, 2022. 3 +[13] Tero Karras, Samuli Laine, Miika Aittala, Janne Hellsten, Jaakko Lehtinen, and Timo Aila. Analyzing and improving the image quality of stylegan. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 8110-8119, 2020. 1 +[14] Bahjat Kawar, Shiran Zada, Oran Lang, Omer Tov, Huiwen Chang, Tali Dekel, Inbar Mosseri, and Michal Irani. Imagc: Text-based real image editing with diffusion models. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 6007-6017, 2023. 2 +[15] Ali Köksal, Kenan E Ak, Ying Sun, Deepu Rajan, and Joo Hwee Lim. Controllable video generation with text-based instructions. IEEE transactions on multimedia, 26: 190-201, 2023. 3 +[16] Nupur Kumari, Bingliang Zhang, Richard Zhang, Eli Shechtman, and Jun-Yan Zhu. Multi-concept customization of text-to-image diffusion. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 1931-1941, 2023. 2 +[17] Jeong-gi Kwak, Erqun Dong, Yuhe Jin, Hanseok Ko, Shweta Mahajan, and Kwang Moo Yi. Vivid-1-to-3: Novel view synthesis with video diffusion models. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 6775–6785, 2024. 3 +[18] Dongxu Li, Junnan Li, and Steven Hoi. Blip-diffusion: Pretrained subject representation for controllable text-to-image generation and editing. Advances in Neural Information Processing Systems, 36, 2024. 2 +[19] Pengyang Ling, Lin Chen, Pan Zhang, Huaian Chen, Yi Jin, and Jinjin Zheng. Freedrag: Feature dragging for reliable point-based image editing. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 6860-6870, 2024. 1, 2 +[20] Minghua Liu, Chao Xu, Haian Jin, Linghao Chen, Mukund Varma T, Zexiang Xu, and Hao Su. One-2-3-45: Any single image to 3d mesh in 45 seconds without per-shape optimization. Advances in Neural Information Processing Systems, 36, 2024. 3 +[21] Yuan Liu, Cheng Lin, Zijiao Zeng, Xiaoxiao Long, Lingjie Liu, Taku Komura, and Wenping Wang. Syncdreamer: Generating multiview-consistent images from a single-view image. arXiv preprint arXiv:2309.03453, 2023. +[22] Xiaoxiao Long, Yuan-Chen Guo, Cheng Lin, Yuan Liu, Zhiyang Dou, Lingjie Liu, Yuexin Ma, Song-Hai Zhang, Marc Habermann, Christian Theobalt, et al. Wonder3d: Single image to 3d using cross-domain diffusion. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 9970-9980, 2024. 3, 6 +[23] Jingyi Lu, Xinghui Li, and Kai Han. Regiondrag: Fast region-based image editing with diffusion models. arXiv preprint arXiv:2407.18247, 2024. 2 + +[24] Minxing Luo, Wentao Cheng, and Jian Yang. Rotationdrag: Point-based image editing with rotated diffusion features. arXiv preprint arXiv:2401.06442, 2024. 2 +[25] Wan-Duo Kurt Ma, John P Lewis, and W Bastiaan Kleijn. Trailblazer: Trajectory control for diffusion-based video generation. arXiv preprint arXiv:2401.00896, 2023. 3 +[26] Ben Mildenhall, Pratul P Srinivasan, Matthew Tancik, Jonathan T Barron, Ravi Ramamoorthi, and Ren Ng. Nerf: Representing scenes as neural radiance fields for view synthesis. Communications of the ACM, 65(1):99-106, 2021. 3 +[27] Chong Mou, Xintao Wang, Jiechong Song, Ying Shan, and Jian Zhang. Dragondiffusion: Enabling drag-style manipulation on diffusion models. arXiv preprint arXiv:2307.02421, 2023. 1, 2 +[28] Chong Mou, Xintao Wang, Jiechong Song, Ying Shan, and Jian Zhang. Diffeditor: Boosting accuracy and flexibility on diffusion-based image editing. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 8488-8497, 2024. +[29] Shen Nie, Hanzhong Allan Guo, Cheng Lu, Yuhao Zhou, Chenyu Zheng, and Chongxuan Li. The blessing of randomness: Sde beats ode in general diffusion-based image editing. arXiv preprint arXiv:2311.01410, 2023. +[30] Xingang Pan, Ayush Tewari, Thomas Leimkuhler, Lingjie Liu, Abhimitra Meka, and Christian Theobalt. Drag your gan: Interactive point-based manipulation on the generative image manifold. In ACM SIGGRAPH 2023 Conference Proceedings, pages 1-11, 2023. 1, 2 +[31] Ben Poole, Ajay Jain, Jonathan T Barron, and Ben Mildenhall. Dreamfusion: Text-to-3d using 2d diffusion. arXiv preprint arXiv:2209.14988, 2022. 3 +[32] Alec Radford, Jong Wook Kim, Chris Hallacy, Aditya Ramesh, Gabriel Goh, Sandhini Agarwal, Girish Sastry, Amanda Askell, Pamela Mishkin, Jack Clark, et al. Learning transferable visual models from natural language supervision. In International conference on machine learning, pages 8748-8763. PMLR, 2021. 5 +[33] Nataniel Ruiz, Yuanzhen Li, Varun Jampani, Yael Pritch, Michael Rubinstein, and Kfir Aberman. Dreambooth: Fine tuning text-to-image diffusion models for subject-driven generation. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 22500-22510, 2023. 2 +[34] Ruoxi Shi, Hansheng Chen, Zhuoyang Zhang, Minghua Liu, Chao Xu, Xinyue Wei, Linghao Chen, Chong Zeng, and Hao Su. Zero123++: a single image to consistent multi-view diffusion base model. arXiv preprint arXiv:2310.15110, 2023. 3 +[35] Xiaoyu Shi, Zhaoyang Huang, Fu-Yun Wang, Weikang Bian, Dasong Li, Yi Zhang, Manyuan Zhang, Ka Chun Cheung, Simon See, Hongwei Qin, et al. Motion-i2v: Consistent and controllable image-to-video generation with explicit motion modeling. In ACM SIGGRAPH 2024 Conference Papers, pages 1-11, 2024. 3 +[36] Yujun Shi, Chuhui Xue, Jun Hao Liew, Jiachun Pan, Hanshu Yan, Wenqing Zhang, Vincent YF Tan, and Song Bai. + +Dragdiffusion: Harnessing diffusion models for interactive point-based image editing. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 8839-8849, 2024. 1, 2, 7 +[37] Luming Tang, Menglin Jia, Qianqian Wang, Cheng Perng Phoo, and Bharath Hariharan. Emergent correspondence from image diffusion. Advances in Neural Information Processing Systems, 36:1363-1389, 2023. 6 +[38] Vikram Voleti, Chun-Han Yao, Mark Boss, Adam Letts, David Pankratz, Dmitry Tochilkin, Christian Laforte, Robin Rombach, and Varun Jampani. Sv3d: Novel multi-view synthesis and 3d generation from a single image using latent video diffusion. In European Conference on Computer Vision, pages 439-457. Springer, 2025. 3, 6, 7 +[39] Xiang Wang, Hangjie Yuan, Shiwei Zhang, Dayou Chen, Jiuniu Wang, Yingya Zhang, Yujun Shen, Deli Zhao, and Jingren Zhou. Videocomposer: Compositional video synthesis with motion controllability. Advances in Neural Information Processing Systems, 36, 2024. 3 +[40] Zhouxia Wang, Ziyang Yuan, Xintao Wang, Yaowei Li, Tianshui Chen, Menghan Xia, Ping Luo, and Ying Shan. Motionctrl: A unified and flexible motion controller for video generation. In ACM SIGGRAPH 2024 Conference Papers, pages 1-11, 2024. 3 +[41] Yuxiang Wei, Yabo Zhang, Zhilong Ji, Jinfeng Bai, Lei Zhang, and Wangmeng Zuo. Elite: Encoding visual concepts into textual embeddings for customized text-to-image generation. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 15943-15953, 2023. 2 +[42] Tong Wu, Jiarui Zhang, Xiao Fu, Yuxin Wang, Jiawei Ren, Liang Pan, Wayne Wu, Lei Yang, Jiaqi Wang, Chen Qian, et al. Omniobject3d: Large-vocabulary 3d object dataset for realistic perception, reconstruction and generation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 803-814, 2023. 5 +[43] Weijia Wu, Zhuang Li, Yuchao Gu, Rui Zhao, Yefei He, David Junhao Zhang, Mike Zheng Shou, Yan Li, Tingting Gao, and Di Zhang. Draganything: Motion control for anything using entity representation. In European Conference on Computer Vision, pages 331-348. Springer, 2025. 1, 3, 5, 6 +[44] Haibo Yang, Yang Chen, Yingwei Pan, Ting Yao, Zhineng Chen, and Tao Mei. 3dstyle-diffusion: Pursuing fine-grained text-driven 3d stylization with 2d diffusion models. In Proceedings of the 31st ACM International Conference on Multimedia, pages 6860-6868, 2023. 3 +[45] Shengming Yin, Chenfei Wu, Jian Liang, Jie Shi, Houqiang Li, Gong Ming, and Nan Duan. Dragnuwa: Fine-grained control in video generation by integrating text, image, and trajectory. arXiv preprint arXiv:2308.08089, 2023. 1, 3 +[46] Ziyang Yuan, Mingdeng Cao, Xintao Wang, Zhongang Qi, Chun Yuan, and Ying Shan. Customnet: Zero-shot object customization with variable-viewpoints in text-to-image diffusion models. arXiv preprint arXiv:2310.19784, 2023. 3 +[47] Lvmin Zhang, Anyi Rao, and Maneesh Agrawala. Adding conditional control to text-to-image diffusion models. In + +Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 3836-3847, 2023. 4 +[48] Yabo Zhang, Yuxiang Wei, Dongsheng Jiang, Xiaopeng Zhang, Wangmeng Zuo, and Qi Tian. Controlvideo: Training-free controllable text-to-video generation. arXiv preprint arXiv:2305.13077, 2023. 3 +[49] Zhenghao Zhang, Junchao Liao, Menghao Li, Long Qin, and Weizhi Wang. Tora: Trajectory-oriented diffusion transformer for video generation. arXiv preprint arXiv:2407.21705, 2024.3 +[50] Chuanxia Zheng and Andrea Vedaldi. Free3d: Consistent novel view synthesis without 3d representation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 9720-9731, 2024. 3 \ No newline at end of file diff --git a/2025/Dragin3D_ Image Editing by Dragging in 3D Space/images.zip b/2025/Dragin3D_ Image Editing by Dragging in 3D Space/images.zip new file mode 100644 index 0000000000000000000000000000000000000000..eeda7772a49d13d9e06bc68374cf37dae4aaa32e --- /dev/null +++ b/2025/Dragin3D_ Image Editing by Dragging in 3D Space/images.zip @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:f2f3b3049e365aae26ffa2931502a2e8c4a56f133a3ee10f0c8d1a4b02610682 +size 670393 diff --git a/2025/Dragin3D_ Image Editing by Dragging in 3D Space/layout.json b/2025/Dragin3D_ Image Editing by Dragging in 3D Space/layout.json new file mode 100644 index 0000000000000000000000000000000000000000..5dec639e74033f8283d790a36bc16e9b182df6ba --- /dev/null +++ b/2025/Dragin3D_ Image Editing by Dragging in 3D Space/layout.json @@ -0,0 +1,7464 @@ +{ + "pdf_info": [ + { + "para_blocks": [ + { + "bbox": [ + 148, + 103, + 463, + 121 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 148, + 103, + 463, + 121 + ], + "spans": [ + { + "bbox": [ + 148, + 103, + 463, + 121 + ], + "type": "text", + "content": "Dragin3D: Image Editing by Dragging in 3D Space" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 126, + 142, + 482, + 157 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 126, + 142, + 482, + 157 + ], + "spans": [ + { + "bbox": [ + 126, + 142, + 482, + 157 + ], + "type": "text", + "content": "Weiran Guang" + }, + { + "bbox": [ + 126, + 142, + 482, + 157 + ], + "type": "inline_equation", + "content": "^{12}" + }, + { + "bbox": [ + 126, + 142, + 482, + 157 + ], + "type": "text", + "content": " Xiaoguang Gu" + }, + { + "bbox": [ + 126, + 142, + 482, + 157 + ], + "type": "inline_equation", + "content": "^{2*}" + }, + { + "bbox": [ + 126, + 142, + 482, + 157 + ], + "type": "text", + "content": " Mengqi Huang" + }, + { + "bbox": [ + 126, + 142, + 482, + 157 + ], + "type": "inline_equation", + "content": "^{1}" + }, + { + "bbox": [ + 126, + 142, + 482, + 157 + ], + "type": "text", + "content": " Zhendong Mao" + }, + { + "bbox": [ + 126, + 142, + 482, + 157 + ], + "type": "inline_equation", + "content": "^{1}" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 187, + 157, + 423, + 172 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 187, + 157, + 423, + 172 + ], + "spans": [ + { + "bbox": [ + 187, + 157, + 423, + 172 + ], + "type": "text", + "content": "1University of Science and Technology of China" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 100, + 171, + 507, + 186 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 100, + 171, + 507, + 186 + ], + "spans": [ + { + "bbox": [ + 100, + 171, + 507, + 186 + ], + "type": "inline_equation", + "content": "^{2}" + }, + { + "bbox": [ + 100, + 171, + 507, + 186 + ], + "type": "text", + "content": "State Key Laboratory of Communication Content Cognition, People's Daily Online" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 94, + 187, + 509, + 199 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 94, + 187, + 509, + 199 + ], + "spans": [ + { + "bbox": [ + 94, + 187, + 509, + 199 + ], + "type": "text", + "content": "guangwr, huangmq}@mail.ustc.edu.cn, guxiaoguang@people.cn, zdmao@ustc.edu.cn" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 151, + 227, + 200, + 239 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 151, + 227, + 200, + 239 + ], + "spans": [ + { + "bbox": [ + 151, + 227, + 200, + 239 + ], + "type": "text", + "content": "Abstract" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 55, + 252, + 297, + 529 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 252, + 297, + 529 + ], + "spans": [ + { + "bbox": [ + 55, + 252, + 297, + 529 + ], + "type": "text", + "content": "Interactive drag editing of images is a valuable task that has gained considerable attention for its precision and controllability. However, existing approaches have primarily focused on manipulating the shape or movement of objects in 2D plane. We propose to extend this drag-based editing task to 3D space. Firstly, we utilize the trajectory of two points to represent the rotational trajectory of the object. Gaussian maps of a circle and a square are centered at these two points, respectively. We use distinct shapes to ensure that symmetric views produce different object representations. Secondly, we introduce a lightweight mapping network to embed the object features into two Gaussian maps to obtain a continuous control condition that guides the model in learning the correspondence between the trajectory and the object. Finally, to overcome the limitations of current 3D object reconstruction datasets, which typically consist of object maps with transparent backgrounds, we affix random backgrounds to them. This modification helps improve the model's ability to ignore background interference when editing real images with complex backgrounds. Experiments demonstrate that our approach successfully achieves object rotation within the drag framework and demonstrates strong generalization to real-world images." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 56, + 552, + 135, + 565 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 552, + 135, + 565 + ], + "spans": [ + { + "bbox": [ + 56, + 552, + 135, + 565 + ], + "type": "text", + "content": "1. Introduction" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 55, + 573, + 296, + 693 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 573, + 296, + 693 + ], + "spans": [ + { + "bbox": [ + 55, + 573, + 296, + 693 + ], + "type": "text", + "content": "Drag-based image editing has garnered substantial research interest in recent years, aiming to enable the manipulation of objects via user-defined handle and target points [30, 36]. Most existing image editing works are text-based, which heavily depends on textual and semantic alignment. Sometimes, it is challenging to articulate user-desired edits in words. Drag-based editing offers a more intuitive, interactive, and granular editing signal, allowing for precise manipulation at the pixel level. Users specify the start and end points based on intended edits, eliminating the need for a" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 313, + 228, + 553, + 264 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 228, + 553, + 264 + ], + "spans": [ + { + "bbox": [ + 313, + 228, + 553, + 264 + ], + "type": "text", + "content": "proper prompt. This enables user-friendly image modification and offers more precise control over positioning, making it valuable for practical applications." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 313, + 267, + 555, + 542 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 267, + 555, + 542 + ], + "spans": [ + { + "bbox": [ + 313, + 267, + 555, + 542 + ], + "type": "text", + "content": "Existing drag-based editing methods can be categorized into optimized-based and training-based. DragGAN [30] is an early representative of optimized-based methods, encompassing two key components: point tracking and motion supervision. The latent code of StyleGAN [13] is optimized while updating the position of the handle point, ensuring that the features near the target points align with those near the handle points. Since DragGAN was proposed, both drag-based image and video editing techniques have been significantly advanced. Subsequently, researchers extended this process to diffusion models by optimizing the features of UNet to break through the inherent limitation of the GAN model [27, 36]. Recent studies have focused on addressing issues related to inaccurate point tracking and incomplete motion supervision, enabling more stable dragging by formulating better point tracking and motion supervision protocols [7, 19]. Furthermore, to overcome the limitations posed by variations in object motion patterns, the training-based method [43, 45] has been introduced using trajectory guidance, where the motion trajectory of the object is encoded as a control condition. In conclusion, ongoing efforts aim to enhance editing controllability and fidelity." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 313, + 545, + 556, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 545, + 556, + 715 + ], + "spans": [ + { + "bbox": [ + 313, + 545, + 556, + 715 + ], + "type": "text", + "content": "However, existing drag-based methods are limited to 2D drag editing and fail to achieve drag in 3D space, which significantly limits their practical application in the real 3D world. 3D generation has significantly expanded the application in games, animation, virtual production, etc. It enables more realistic creation of 3D visual spaces that closely resemble the physical world. We consider whether it is possible to generate novel view images of the object through simple point-drag interactions. The two drag paradigms discussed above are inadequate for solving this task. The optimization-based approach primarily seeks to supervise feature consistency between the handle point and the target point, which cannot well control the substantial changes in features that arise due to variations in spatial perspective, as" + } + ] + } + ], + "index": 13 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "spans": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "type": "text", + "content": "CVF" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 145, + 0, + 494, + 37 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 145, + 0, + 494, + 37 + ], + "spans": [ + { + "bbox": [ + 145, + 0, + 494, + 37 + ], + "type": "text", + "content": "This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 66, + 702, + 212, + 713 + ], + "type": "page_footnote", + "angle": 0, + "lines": [ + { + "bbox": [ + 66, + 702, + 212, + 713 + ], + "spans": [ + { + "bbox": [ + 66, + 702, + 212, + 713 + ], + "type": "text", + "content": "*Xiaoguang Gu is the Corresponding author." + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "type": "text", + "content": "21502" + } + ] + } + ], + "index": 15 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 0 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 59, + 70, + 293, + 175 + ], + "blocks": [ + { + "bbox": [ + 59, + 70, + 293, + 175 + ], + "lines": [ + { + "bbox": [ + 59, + 70, + 293, + 175 + ], + "spans": [ + { + "bbox": [ + 59, + 70, + 293, + 175 + ], + "type": "image", + "image_path": "91d03067ba77be0f3c380cf0c1ce35f1e9c1dc5859c593528bbd75c7f4777e22.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 55, + 183, + 295, + 304 + ], + "lines": [ + { + "bbox": [ + 55, + 183, + 295, + 304 + ], + "spans": [ + { + "bbox": [ + 55, + 183, + 295, + 304 + ], + "type": "text", + "content": "Figure 1. Existing drag method for object rotation representation. (a) The optimization-based method. The sculpture makes tiny rotations, and changes in the position of the target point still tend to maintain a frontal perspective, which fails to achieve 3D editing with varying amplitudes based solely on the dragging intensity. (b) The training-based method. In the upper figure, the object moves along the trajectory of the dragging point. In the lower figure, the object rotates in the direction of the trajectory of the dragging point. The trajectory is inclined upwards, with the ship exhibiting a tendency to rotate towards the target position. Arguably, whether the ship can achieve 3D rotation is entirely random." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "bbox": [ + 55, + 326, + 295, + 373 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 326, + 295, + 373 + ], + "spans": [ + { + "bbox": [ + 55, + 326, + 295, + 373 + ], + "type": "text", + "content": "illustrated in Fig. 1(a). On the other hand, the trajectory-based methods introduce ambiguity in the trend of object motion, as demonstrated in Fig. 1(b). Therefore, in this study, we introduce a new drag-based object rotation task." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 55, + 375, + 295, + 673 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 375, + 295, + 673 + ], + "spans": [ + { + "bbox": [ + 55, + 375, + 295, + 673 + ], + "type": "text", + "content": "The key challenges of this task lie in how to accurately represent and track the desired 3D editing trajectory from a 2D image, and how to establish a correspondence between object motion trajectories and their respective representations. Due to the ambiguity in existing training-based methods for single-point trajectory representation, combined with the feature alignment principle of optimization-based methods, we propose Dragin3D, which extends drag-based editing to 3D space for the first time. We introduce a double-point representation approach that associates object trajectories with image features. Given an image of an object and a double-point rotation trajectory, Dragin3D generates a novel view image of the object from the corresponding trajectory, as shown in Fig. 2. The core concept of Dragin3D is to precisely capture the direction and magnitude of the object's rotation using the double-point trajectory. Specifically, our method consists of two main components: (1) Drag Trajectory Representation (DTR): we map the rotational trajectory of the object onto the rotational trajectory of a line segment, track the two endpoints of this segment, and project them into Gaussian maps with different shapes; (2) Control Condition Inject Module (CCIM): to enhance fidelity to the object, we extract image features through a lightweight mapping network and inject these features into the corresponding Gaussian maps." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 67, + 674, + 253, + 685 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 67, + 674, + 253, + 685 + ], + "spans": [ + { + "bbox": [ + 67, + 674, + 253, + 685 + ], + "type": "text", + "content": "Our contributions are summarized as follows:" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 55, + 689, + 295, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 689, + 295, + 714 + ], + "spans": [ + { + "bbox": [ + 55, + 689, + 295, + 714 + ], + "type": "text", + "content": "- We propose an innovative interactive image editing method Dragin3D. To the best of our knowledge, we are" + } + ] + } + ], + "index": 5 + }, + { + "type": "image", + "bbox": [ + 318, + 70, + 553, + 156 + ], + "blocks": [ + { + "bbox": [ + 318, + 70, + 553, + 156 + ], + "lines": [ + { + "bbox": [ + 318, + 70, + 553, + 156 + ], + "spans": [ + { + "bbox": [ + 318, + 70, + 553, + 156 + ], + "type": "image", + "image_path": "c22e3b4194ee56c01cc287d1fc6892a7586d85a68b7d4953b53ca8781b6f5721.jpg" + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 313, + 163, + 555, + 209 + ], + "lines": [ + { + "bbox": [ + 313, + 163, + 555, + 209 + ], + "spans": [ + { + "bbox": [ + 313, + 163, + 555, + 209 + ], + "type": "text", + "content": "Figure 2. Drag results of our proposed paradigm Dragin3D. Given an input image and coordinates of two representation points, i.e., " + }, + { + "bbox": [ + 313, + 163, + 555, + 209 + ], + "type": "inline_equation", + "content": "[P_{11}, P_{21}],[P_{12}, P_{22}],\\ldots,[P_{1n}, P_{2n}]" + }, + { + "bbox": [ + 313, + 163, + 555, + 209 + ], + "type": "text", + "content": ". Dragin3D can generate a novel view image of the object for the corresponding trajectory." + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "image_caption" + } + ], + "index": 6 + }, + { + "bbox": [ + 321, + 237, + 553, + 261 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 321, + 237, + 553, + 261 + ], + "spans": [ + { + "bbox": [ + 321, + 237, + 553, + 261 + ], + "type": "text", + "content": "the first to achieve multi-view image generation of objects by dragging them in 3D space based on a diffusion model." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 314, + 262, + 553, + 380 + ], + "type": "list", + "angle": 0, + "index": 12, + "blocks": [ + { + "bbox": [ + 314, + 262, + 553, + 308 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 314, + 262, + 553, + 308 + ], + "spans": [ + { + "bbox": [ + 314, + 262, + 553, + 308 + ], + "type": "text", + "content": "- We propose to represent the object rotation trajectory by a double-point motion trajectory, which maps 3D object motion onto a 2D plane and eliminates the ambiguity of the object motion mode." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 314, + 309, + 553, + 344 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 314, + 309, + 553, + 344 + ], + "spans": [ + { + "bbox": [ + 314, + 309, + 553, + 344 + ], + "type": "text", + "content": "- We introduce a tracking point projection method, which projects the tracking points onto Gaussian maps and aligns them with the object's feature information." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 314, + 345, + 553, + 380 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 314, + 345, + 553, + 380 + ], + "spans": [ + { + "bbox": [ + 314, + 345, + 553, + 380 + ], + "type": "text", + "content": "- Experimental results demonstrate that despite being trained on a synthetic 3D dataset, our method generalizes well beyond the dataset's distribution." + } + ] + } + ], + "index": 11 + } + ], + "sub_type": "text" + }, + { + "bbox": [ + 314, + 403, + 400, + 415 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 314, + 403, + 400, + 415 + ], + "spans": [ + { + "bbox": [ + 314, + 403, + 400, + 415 + ], + "type": "text", + "content": "2. Related Work" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 313, + 427, + 515, + 441 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 427, + 515, + 441 + ], + "spans": [ + { + "bbox": [ + 313, + 427, + 515, + 441 + ], + "type": "text", + "content": "2.1. Drag-based controllable image editing" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 313, + 450, + 555, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 450, + 555, + 713 + ], + "spans": [ + { + "bbox": [ + 313, + 450, + 555, + 713 + ], + "type": "text", + "content": "In the field of image editing, several established works have focused on tasks such as image customization [16, 18, 33, 41], image editing [1, 4, 6, 9, 11, 14], which modify object attributes, quantity, etc. through text-based operations. Due to the limitations of text-based editing in controlling local changes at a fine-grained level, the Drag approach introduces a novel approach to image editing [7, 8, 19, 23, 24, 27-30, 36]. This interactive manipulation allows users to precisely designate both the editing area and control mode with a simple click. [30, 36] propose a point-based editing technique that integrates point tracking and motion supervision. [19] enhances drag editing stability and reliability without requiring precise point tracking by introducing adaptively updated template features, fuzzy localization, and linear search techniques. [7] introduces a discriminative point tracking method and a confidence-based latent space enhancement strategy, resulting in more stable and accurate drag performance. These works primarily focus on enhancing source and target feature optimization. However, when extended to object rotation manipulation, significant feature differences across views pose a challenge that pure feature alignment cannot resolve." + } + ] + } + ], + "index": 15 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "text", + "content": "21503" + } + ] + } + ], + "index": 16 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 1 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 72, + 71, + 539, + 249 + ], + "blocks": [ + { + "bbox": [ + 72, + 71, + 539, + 249 + ], + "lines": [ + { + "bbox": [ + 72, + 71, + 539, + 249 + ], + "spans": [ + { + "bbox": [ + 72, + 71, + 539, + 249 + ], + "type": "image", + "image_path": "0c05505f6758dfeddf97b8869d416d05b027f8d35278371b092d2e3c2d852e77.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 54, + 257, + 555, + 304 + ], + "lines": [ + { + "bbox": [ + 54, + 257, + 555, + 304 + ], + "spans": [ + { + "bbox": [ + 54, + 257, + 555, + 304 + ], + "type": "text", + "content": "Figure 3. Dragin3D Architecture. We track the representation points by calculating the transfer matrix and plot circle and square Gaussian maps centered on two representation points, respectively. We encode the image to be edited to extract the image features, which are injected into the original encoded Gaussian map features to obtain new Gaussian map features. The original Gaussian map features are added to the newly injected Gaussian map features as an input control condition for ControlNet." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "bbox": [ + 55, + 322, + 284, + 335 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 322, + 284, + 335 + ], + "spans": [ + { + "bbox": [ + 55, + 322, + 284, + 335 + ], + "type": "text", + "content": "2.2. Trajectory-based controllable image editing" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 56, + 340, + 296, + 593 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 340, + 296, + 593 + ], + "spans": [ + { + "bbox": [ + 56, + 340, + 296, + 593 + ], + "type": "text", + "content": "Recently, trajectory-based motion control [10, 25, 35, 39, 40, 43, 45, 49] has emerged as an efficient and user-friendly method for controllable video generation [5, 12, 15, 48], with drag editing serving as a trajectory-based task. The gradual movement of handle points forms a continuous trajectory, providing a more precise guidance signal in contrast to defined start and end points. [45] introduces a unified control framework that integrates text, image, and trajectory data, utilizing a trajectory sampler to extract trajectories from open-domain video streams. [43] extracts features from the diffusion model as entity representations, and incorporates them into a denoised 3D U-Net with a noisy 2D Gaussian representation. [10] focuses on specific camera motion, generating videos by adding temporal constraints to a T2I model to capture motion dynamics, leveraging eight independent LoRA models for basic camera control. [40] differentiates between camera and object motions, enabling independent or combined control of the camera pose matrix and object trajectory to produce personalized videos. However, these motions are limited to 2D translations, making it difficult to generate novel views of objects in 3D space." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 55, + 599, + 180, + 612 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 599, + 180, + 612 + ], + "spans": [ + { + "bbox": [ + 55, + 599, + 180, + 612 + ], + "type": "text", + "content": "2.3. 3D generation Models" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 55, + 617, + 296, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 617, + 296, + 715 + ], + "spans": [ + { + "bbox": [ + 55, + 617, + 296, + 715 + ], + "type": "text", + "content": "Based on the impressive generative effect of diffusion models on 2D images, the approach of 3D generation via diffusion models has emerged [17, 20-22, 31, 34, 44, 50]. [31] is one of the pioneering methods, introduces the SDS loss function, which enables the updating of NeRF [26] weights by minimizing the discrepancy between noise levels. Several 3D image generation methods using enhanced SDS loss functions have been proposed, employing a 2D" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 313, + 323, + 556, + 456 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 323, + 556, + 456 + ], + "spans": [ + { + "bbox": [ + 313, + 323, + 556, + 456 + ], + "type": "text", + "content": "generative model to create 3D structures from textual descriptions without relying on 3D data, thus overcoming the challenge of limited 3D data availability. [44] integrates the diffusion process of image rendering and synthesis using NeRF, ensuring that the generated images remain semantically aligned with the text and geometrically consistent with the depth map during training. [38, 46] directly embed camera parameters as control conditions in the model to realize the camera pose control. However, the direct embedding of discrete values, such as camera parameters, as control conditions often lack the necessary precision." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 314, + 468, + 375, + 480 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 314, + 468, + 375, + 480 + ], + "spans": [ + { + "bbox": [ + 314, + 468, + 375, + 480 + ], + "type": "text", + "content": "3. Methods" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 313, + 489, + 487, + 501 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 489, + 487, + 501 + ], + "spans": [ + { + "bbox": [ + 313, + 489, + 487, + 501 + ], + "type": "text", + "content": "3.1. Task Definition & Preliminaries" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 313, + 509, + 555, + 605 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 509, + 555, + 605 + ], + "spans": [ + { + "bbox": [ + 313, + 509, + 555, + 605 + ], + "type": "text", + "content": "3D Drag Editing Definition. We observed that previous drag tasks lacked a unified framework for moving objects in 3D space by drag interaction, for example, object rotation, so we propose to extend the task to 3D space. This task marks a significant advancement beyond merely editing the appearance of objects in a 2D image. Given a reference object image, our approach allows users to obtain a new perspective of the object through a simple drag interaction." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 313, + 605, + 556, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 605, + 556, + 715 + ], + "spans": [ + { + "bbox": [ + 313, + 605, + 556, + 715 + ], + "type": "text", + "content": "Preliminary of Stable Video Diffusion. Our paradigm is based on Stable Video Diffusion (SVD) [2] due to its robust predictive capabilities for dynamically moving objects. The key modification from Stable Diffusion (SD) [3] is the integration of a temporal layer into the denoising model's U-Net architecture, resulting in a 3D-UNet. Compared to SD trained on image datasets, video frames capture more dynamic information about objects, leading to more sensitivity to morphological changes of the same objects, ensur" + } + ] + } + ], + "index": 10 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "type": "text", + "content": "21504" + } + ] + } + ], + "index": 11 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 2 + }, + { + "para_blocks": [ + { + "bbox": [ + 55, + 72, + 269, + 83 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 72, + 269, + 83 + ], + "spans": [ + { + "bbox": [ + 55, + 72, + 269, + 83 + ], + "type": "text", + "content": "ing perspective and temporal consistency of rotation." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 55, + 84, + 296, + 180 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 84, + 296, + 180 + ], + "spans": [ + { + "bbox": [ + 55, + 84, + 296, + 180 + ], + "type": "text", + "content": "Preliminary of ControlNet. ControlNet [47] is one of the representative works in the field of controlled image generation, which supports a variety of inputs such as edge maps, segmentation maps, keypoints, etc., as conditional inputs to guide the image generation process. " + }, + { + "bbox": [ + 55, + 84, + 296, + 180 + ], + "type": "inline_equation", + "content": "\\mathcal{F}(\\cdot ;\\Theta)" + }, + { + "bbox": [ + 55, + 84, + 296, + 180 + ], + "type": "text", + "content": " is a neural network block with parameters " + }, + { + "bbox": [ + 55, + 84, + 296, + 180 + ], + "type": "inline_equation", + "content": "\\Theta" + }, + { + "bbox": [ + 55, + 84, + 296, + 180 + ], + "type": "text", + "content": ". ControlNet locks all parameters in " + }, + { + "bbox": [ + 55, + 84, + 296, + 180 + ], + "type": "inline_equation", + "content": "\\Theta" + }, + { + "bbox": [ + 55, + 84, + 296, + 180 + ], + "type": "text", + "content": " and clones them into a trainable copy " + }, + { + "bbox": [ + 55, + 84, + 296, + 180 + ], + "type": "inline_equation", + "content": "\\Theta_{c}" + }, + { + "bbox": [ + 55, + 84, + 296, + 180 + ], + "type": "text", + "content": ". The output of ControlNet " + }, + { + "bbox": [ + 55, + 84, + 296, + 180 + ], + "type": "inline_equation", + "content": "y_{c}" + }, + { + "bbox": [ + 55, + 84, + 296, + 180 + ], + "type": "text", + "content": " can be expressed as:" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 61, + 201, + 295, + 215 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 61, + 201, + 295, + 215 + ], + "spans": [ + { + "bbox": [ + 61, + 201, + 295, + 215 + ], + "type": "interline_equation", + "content": "\\boldsymbol {y} _ {\\mathbf {c}} = \\mathcal {F} (\\boldsymbol {x}; \\Theta) + \\mathcal {Z} \\left(\\mathcal {F} \\left(\\boldsymbol {x} + \\mathcal {Z} \\left(\\boldsymbol {c}; \\Theta_ {\\mathrm {z} 1}\\right); \\Theta_ {\\mathrm {c}}\\right); \\Theta_ {\\mathrm {z} 2}\\right), \\tag {1}", + "image_path": "bc4a41ac69b77884eb721b134c6b2c41954606266fe61bb8ebc7ab3be7584c2a.jpg" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 55, + 223, + 296, + 260 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 223, + 296, + 260 + ], + "spans": [ + { + "bbox": [ + 55, + 223, + 296, + 260 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 55, + 223, + 296, + 260 + ], + "type": "inline_equation", + "content": "c" + }, + { + "bbox": [ + 55, + 223, + 296, + 260 + ], + "type": "text", + "content": " is the condition input, " + }, + { + "bbox": [ + 55, + 223, + 296, + 260 + ], + "type": "inline_equation", + "content": "\\mathcal{Z}(\\cdot ;\\cdot)" + }, + { + "bbox": [ + 55, + 223, + 296, + 260 + ], + "type": "text", + "content": " is a " + }, + { + "bbox": [ + 55, + 223, + 296, + 260 + ], + "type": "inline_equation", + "content": "1\\times 1" + }, + { + "bbox": [ + 55, + 223, + 296, + 260 + ], + "type": "text", + "content": " convolutional layer initialized with zero weights and bias. When ControlNet is used in Stable Diffusion, the training objective is:" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 64, + 280, + 295, + 295 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 64, + 280, + 295, + 295 + ], + "spans": [ + { + "bbox": [ + 64, + 280, + 295, + 295 + ], + "type": "interline_equation", + "content": "\\mathcal {L} = \\mathbb {E} _ {\\boldsymbol {z} _ {0}, t, \\boldsymbol {c} _ {t}, \\boldsymbol {c} _ {\\mathrm {f}}, \\epsilon \\sim \\mathcal {N} (0, 1)} \\left[ \\| \\epsilon - \\epsilon_ {\\theta} \\left(z _ {t}, t, \\boldsymbol {c} _ {t}, \\boldsymbol {c} _ {\\mathrm {f}}\\right)) \\| _ {2} ^ {2} \\right], \\tag {2}", + "image_path": "b4b335a150cfed625798c8f6b77a66a7a1b051d2a6f05d5d61ee315846ab45c1.jpg" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 55, + 303, + 296, + 338 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 303, + 296, + 338 + ], + "spans": [ + { + "bbox": [ + 55, + 303, + 296, + 338 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 55, + 303, + 296, + 338 + ], + "type": "inline_equation", + "content": "z_0" + }, + { + "bbox": [ + 55, + 303, + 296, + 338 + ], + "type": "text", + "content": " denotes the input image, " + }, + { + "bbox": [ + 55, + 303, + 296, + 338 + ], + "type": "inline_equation", + "content": "z_t" + }, + { + "bbox": [ + 55, + 303, + 296, + 338 + ], + "type": "text", + "content": " denotes the noisy image in timestep " + }, + { + "bbox": [ + 55, + 303, + 296, + 338 + ], + "type": "inline_equation", + "content": "t" + }, + { + "bbox": [ + 55, + 303, + 296, + 338 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 55, + 303, + 296, + 338 + ], + "type": "inline_equation", + "content": "c_t" + }, + { + "bbox": [ + 55, + 303, + 296, + 338 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 55, + 303, + 296, + 338 + ], + "type": "inline_equation", + "content": "c_f" + }, + { + "bbox": [ + 55, + 303, + 296, + 338 + ], + "type": "text", + "content": " denote text prompts and the task-specific conditions, respectively." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 55, + 339, + 296, + 434 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 339, + 296, + 434 + ], + "spans": [ + { + "bbox": [ + 55, + 339, + 296, + 434 + ], + "type": "text", + "content": "Preliminary of Gaussian heatmap. Gaussian heatmaps are commonly used for keypoint detection tasks. A heatmap is a feature map of the same size as the input, with a value of 1 at a specified coordinate (i.e., the representation points in our task) and 0 at all other positions. Additionally, to align with inputs like mask or depth maps used in ControlNet, a Gaussian heatmap is a suitable input, which effectively converts discrete point signals into continuous image signals." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 55, + 442, + 171, + 453 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 442, + 171, + 453 + ], + "spans": [ + { + "bbox": [ + 55, + 442, + 171, + 453 + ], + "type": "text", + "content": "3.2. Overall Framework" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 54, + 460, + 295, + 652 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 54, + 460, + 295, + 652 + ], + "spans": [ + { + "bbox": [ + 54, + 460, + 295, + 652 + ], + "type": "text", + "content": "For this novel dragging in 3D space task, we propose a framework named Dragin3D, which incorporates two key innovations: 3D drag trajectory representation and control condition injection. The overall workflow of our model is illustrated in Fig. 3. We employ an architecture that combines SVD and ControlNet. Firstly, we propose to describe the motion trajectory of the object with two points, and compute to track the positions of the representation points in different views. Next, we construct Gaussian maps of different shapes centered on these two points. These maps are then encoded, and image feature representations are injected into the Gaussian maps to align them with the corresponding point locations. Finally, the encoded Gaussian map features are linked to new Gaussian map features that incorporate object-specific information, which are used as control condition inputs for ControlNet." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 55, + 658, + 277, + 672 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 658, + 277, + 672 + ], + "spans": [ + { + "bbox": [ + 55, + 658, + 277, + 672 + ], + "type": "text", + "content": "3.3. 3D Drag Trajectory Representation (DTR)" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 55, + 677, + 296, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 677, + 296, + 714 + ], + "spans": [ + { + "bbox": [ + 55, + 677, + 296, + 714 + ], + "type": "text", + "content": "To extend the editing task to 3D space, we propose a 3D Drag Trajectory Representation (DTR) method, which explicitly represents the object's trajectory using a rotational" + } + ] + } + ], + "index": 10 + }, + { + "type": "image", + "bbox": [ + 328, + 70, + 375, + 123 + ], + "blocks": [ + { + "bbox": [ + 328, + 70, + 375, + 123 + ], + "lines": [ + { + "bbox": [ + 328, + 70, + 375, + 123 + ], + "spans": [ + { + "bbox": [ + 328, + 70, + 375, + 123 + ], + "type": "image", + "image_path": "0d4c1ccf8db16bad9ec345f478fd54deda6c06588911a99a6d8955dcfeaa3f0e.jpg" + } + ] + } + ], + "index": 11, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 328, + 124, + 337, + 131 + ], + "lines": [ + { + "bbox": [ + 328, + 124, + 337, + 131 + ], + "spans": [ + { + "bbox": [ + 328, + 124, + 337, + 131 + ], + "type": "text", + "content": "#" + } + ] + } + ], + "index": 12, + "angle": 0, + "type": "image_caption" + } + ], + "index": 11 + }, + { + "type": "image", + "bbox": [ + 388, + 71, + 432, + 123 + ], + "blocks": [ + { + "bbox": [ + 388, + 71, + 432, + 123 + ], + "lines": [ + { + "bbox": [ + 388, + 71, + 432, + 123 + ], + "spans": [ + { + "bbox": [ + 388, + 71, + 432, + 123 + ], + "type": "image", + "image_path": "7f33791087c9f9c7b1c483af2abc33c60a6ac27b798a947b6cca073d3314783d.jpg" + } + ] + } + ], + "index": 13, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 386, + 124, + 394, + 132 + ], + "lines": [ + { + "bbox": [ + 386, + 124, + 394, + 132 + ], + "spans": [ + { + "bbox": [ + 386, + 124, + 394, + 132 + ], + "type": "text", + "content": "#" + } + ] + } + ], + "index": 14, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 313, + 143, + 554, + 187 + ], + "lines": [ + { + "bbox": [ + 313, + 143, + 554, + 187 + ], + "spans": [ + { + "bbox": [ + 313, + 143, + 554, + 187 + ], + "type": "text", + "content": "Figure 4. Our 3D drag trajectory representation method. The computed tracking points are projected into Gaussian maps, ensuring the rotational trajectory of the object is visually aligned with the trajectory of the line segment formed by the representation points." + } + ] + } + ], + "index": 19, + "angle": 0, + "type": "image_caption" + } + ], + "index": 13 + }, + { + "type": "image", + "bbox": [ + 444, + 71, + 490, + 123 + ], + "blocks": [ + { + "bbox": [ + 444, + 71, + 490, + 123 + ], + "lines": [ + { + "bbox": [ + 444, + 71, + 490, + 123 + ], + "spans": [ + { + "bbox": [ + 444, + 71, + 490, + 123 + ], + "type": "image", + "image_path": "104c231831aee4d0864d1d1bd7bb9e5ce9c25f1fb0e8078ca19af6ec8996d5a3.jpg" + } + ] + } + ], + "index": 15, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 444, + 124, + 453, + 132 + ], + "lines": [ + { + "bbox": [ + 444, + 124, + 453, + 132 + ], + "spans": [ + { + "bbox": [ + 444, + 124, + 453, + 132 + ], + "type": "text", + "content": "#" + } + ] + } + ], + "index": 16, + "angle": 0, + "type": "image_caption" + } + ], + "index": 15 + }, + { + "type": "image", + "bbox": [ + 504, + 71, + 551, + 123 + ], + "blocks": [ + { + "bbox": [ + 504, + 71, + 551, + 123 + ], + "lines": [ + { + "bbox": [ + 504, + 71, + 551, + 123 + ], + "spans": [ + { + "bbox": [ + 504, + 71, + 551, + 123 + ], + "type": "image", + "image_path": "494d1e506e3b8aefecb56841ab7e6c9231e4718119908cd6eac2e54b3308e229.jpg" + } + ] + } + ], + "index": 17, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 502, + 124, + 511, + 132 + ], + "lines": [ + { + "bbox": [ + 502, + 124, + 511, + 132 + ], + "spans": [ + { + "bbox": [ + 502, + 124, + 511, + 132 + ], + "type": "text", + "content": "#" + } + ] + } + ], + "index": 18, + "angle": 0, + "type": "image_caption" + } + ], + "index": 17 + }, + { + "bbox": [ + 313, + 207, + 554, + 266 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 207, + 554, + 266 + ], + "spans": [ + { + "bbox": [ + 313, + 207, + 554, + 266 + ], + "type": "text", + "content": "line segment defined by two points. In general, an object's rotation in 3D space involves variations around three coordinate axes across three planes. Since left-right rotation is more intuitive and commonly used at the operational level, our approach and experiments focus on the z-axis rotation." + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 313, + 267, + 555, + 505 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 267, + 555, + 505 + ], + "spans": [ + { + "bbox": [ + 313, + 267, + 555, + 505 + ], + "type": "text", + "content": "The selection and stable tracking of the representation points are critical, as they directly affect the accuracy of the subsequent results. To ensure that the motion trajectory of the line segment is coherent and clearly defined, we propose a method for calculating the position of a specified coordinate point in images captured from different views. Then we project the computed points onto the 2D plane, as shown in Fig. 4. The required camera intrinsic and extrinsic parameters are provided by the 3D dataset. Extrinsic parameter calibration is performed by solving the transfer matrix using the real-world coordinates and pixel coordinates of fixed points. Conversely, when the camera extrinsic parameter is available, we can calculate the pixel coordinates of a fixed point by maintaining the constancy of its world coordinates. To simplify the calculation of 3D coordinates, we assume that the z-axis coordinate of each pixel point is fixed to 1. While this assumption introduces some deviation, the resulting calculation provides a reasonable approximation of the trajectory and magnitude of the object's movement. The relevant calculation formulas are as follows." + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 320, + 526, + 553, + 540 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 320, + 526, + 553, + 540 + ], + "spans": [ + { + "bbox": [ + 320, + 526, + 553, + 540 + ], + "type": "interline_equation", + "content": "x _ {c 1} = \\left(x _ {1} - c _ {x}\\right) * z _ {c 1} / f _ {x}, y _ {c 1} = \\left(y _ {1} - c _ {y}\\right) * z _ {c 1} / f _ {y}, \\tag {3}", + "image_path": "01ab700f9a0bf9f8f7d86395d1430bee058271373f1f18c7734440b17a6e4d41.jpg" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 326, + 565, + 531, + 600 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 326, + 565, + 531, + 600 + ], + "spans": [ + { + "bbox": [ + 326, + 565, + 531, + 600 + ], + "type": "interline_equation", + "content": "\\begin{array}{l} \\left[ \\begin{array}{l l l} x _ {w} & y _ {w} & z _ {w} \\end{array} \\right] ^ {- 1} = T _ {1} ^ {- 1} \\cdot \\left[ \\begin{array}{l l l} x _ {c 1} & y _ {c 1} & z _ {c 1} \\end{array} \\right] ^ {- 1}, \\\\ \\left[ \\begin{array}{c} x _ {c 2} y _ {c 2} z _ {c 2} 1 \\end{array} \\right] ^ {- 1} = T _ {2} \\cdot \\left[ \\begin{array}{c} x _ {w} y _ {w} z _ {w} 1 \\end{array} \\right] ^ {- 1}, \\\\ \\end{array}", + "image_path": "d2993ae2b1f35fe56e01201c4f1eed5498d5fcfeafeccfc3349537ca5c60d526.jpg" + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 328, + 620, + 553, + 633 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 328, + 620, + 553, + 633 + ], + "spans": [ + { + "bbox": [ + 328, + 620, + 553, + 633 + ], + "type": "interline_equation", + "content": "x _ {2} = f _ {x} * x _ {c 2} / z _ {c 2} + c _ {x}, y _ {2} = f _ {y} * y _ {c 2} / z _ {c 2} + c _ {y}, \\tag {5}", + "image_path": "1bfc5f98013bb9c4c383d5bab302fadf5bc489104b98cf0ae2b41fbf209dde28.jpg" + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 313, + 641, + 555, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 641, + 555, + 714 + ], + "spans": [ + { + "bbox": [ + 313, + 641, + 555, + 714 + ], + "type": "text", + "content": "where the intrinsic " + }, + { + "bbox": [ + 313, + 641, + 555, + 714 + ], + "type": "inline_equation", + "content": "(f_{x},f_{y},c_{x},c_{y})" + }, + { + "bbox": [ + 313, + 641, + 555, + 714 + ], + "type": "text", + "content": " and extrinsic " + }, + { + "bbox": [ + 313, + 641, + 555, + 714 + ], + "type": "inline_equation", + "content": "(T_{1},T_{2})" + }, + { + "bbox": [ + 313, + 641, + 555, + 714 + ], + "type": "text", + "content": " of the camera are labeled in the dataset. We define " + }, + { + "bbox": [ + 313, + 641, + 555, + 714 + ], + "type": "inline_equation", + "content": "P_{1}" + }, + { + "bbox": [ + 313, + 641, + 555, + 714 + ], + "type": "inline_equation", + "content": "(x_{1},y_{1})" + }, + { + "bbox": [ + 313, + 641, + 555, + 714 + ], + "type": "text", + "content": " as the pixel coordinates of the specified point in the first image, " + }, + { + "bbox": [ + 313, + 641, + 555, + 714 + ], + "type": "inline_equation", + "content": "P_{c1}" + }, + { + "bbox": [ + 313, + 641, + 555, + 714 + ], + "type": "inline_equation", + "content": "(x_{c1},y_{c1},z_{c1})" + }, + { + "bbox": [ + 313, + 641, + 555, + 714 + ], + "type": "text", + "content": " is the camera coordinates of the specified point in the first image, " + }, + { + "bbox": [ + 313, + 641, + 555, + 714 + ], + "type": "inline_equation", + "content": "P_{w}" + }, + { + "bbox": [ + 313, + 641, + 555, + 714 + ], + "type": "inline_equation", + "content": "(x_{w},y_{w},z_{w})" + }, + { + "bbox": [ + 313, + 641, + 555, + 714 + ], + "type": "text", + "content": " is the world coordinates, " + }, + { + "bbox": [ + 313, + 641, + 555, + 714 + ], + "type": "inline_equation", + "content": "P_{c2}" + }, + { + "bbox": [ + 313, + 641, + 555, + 714 + ], + "type": "inline_equation", + "content": "(x_{c2},y_{c2},z_{c2})" + }, + { + "bbox": [ + 313, + 641, + 555, + 714 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 313, + 641, + 555, + 714 + ], + "type": "inline_equation", + "content": "P_{2}" + }, + { + "bbox": [ + 313, + 641, + 555, + 714 + ], + "type": "inline_equation", + "content": "(x_{2},y_{2})" + }, + { + "bbox": [ + 313, + 641, + 555, + 714 + ], + "type": "text", + "content": " are" + } + ] + } + ], + "index": 25 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "text", + "content": "21505" + } + ] + } + ], + "index": 26 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 3 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 93, + 71, + 521, + 255 + ], + "blocks": [ + { + "bbox": [ + 93, + 71, + 521, + 255 + ], + "lines": [ + { + "bbox": [ + 93, + 71, + 521, + 255 + ], + "spans": [ + { + "bbox": [ + 93, + 71, + 521, + 255 + ], + "type": "image", + "image_path": "acfcb0464ad9529309c26ef4376e3998770dcf4a098a096aac88af7eda717cb1.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 54, + 262, + 555, + 308 + ], + "lines": [ + { + "bbox": [ + 54, + 262, + 555, + 308 + ], + "spans": [ + { + "bbox": [ + 54, + 262, + 555, + 308 + ], + "type": "text", + "content": "Figure 5. Dragin3D Visualization. Given an input image and a set of trajectory point pairs, we obtained Gaussian maps of the trajectories as input (first column), which is then used to generate the corresponding objects at the specified trajectory points (subsequent columns). The arrows indicate rotation directions, the red line is the current view of the object, the gray dotted line shows the trajectory that has already been traversed, and the gray solid line is the future trajectory." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "bbox": [ + 55, + 328, + 294, + 353 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 328, + 294, + 353 + ], + "spans": [ + { + "bbox": [ + 55, + 328, + 294, + 353 + ], + "type": "text", + "content": "the camera coordinates and pixel coordinates of the specified point in the second image, respectively." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 55, + 360, + 274, + 373 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 360, + 274, + 373 + ], + "spans": [ + { + "bbox": [ + 55, + 360, + 274, + 373 + ], + "type": "text", + "content": "3.4. Control Condition Inject Module (CCIM)" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 56, + 379, + 295, + 581 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 379, + 295, + 581 + ], + "spans": [ + { + "bbox": [ + 56, + 379, + 295, + 581 + ], + "type": "text", + "content": "In order to process the existing input image and the computed trajectory representations into suitable input conditions for the model, we proposed a Control Condition Inject Module (CCIM). Drawing inspiration from DragAnything [43], which uses discrete pixel points as input control conditions by mapping them into a Gaussian map. We propose to use two representation points. However, we observed that the two representation points of the object are nearly identical in the front and rear views, as well as in the symmetric left and right side views. We propose a Shape DistinGuish Strategy (SDGS) to distinguish the two points by their respective shapes. Specifically, we create circular and square Gaussian maps centered on the two points, respectively, as shown in Fig. 9 (green boxes). Then the Gaussian map images " + }, + { + "bbox": [ + 56, + 379, + 295, + 581 + ], + "type": "inline_equation", + "content": "x \\in \\mathbb{R}^{F \\times H \\times W \\times 3}" + }, + { + "bbox": [ + 56, + 379, + 295, + 581 + ], + "type": "text", + "content": " are encoded into image features " + }, + { + "bbox": [ + 56, + 379, + 295, + 581 + ], + "type": "inline_equation", + "content": "f_{gi} \\in \\mathbb{R}^{F \\times H \\times W \\times C}" + }, + { + "bbox": [ + 56, + 379, + 295, + 581 + ], + "type": "text", + "content": ", where " + }, + { + "bbox": [ + 56, + 379, + 295, + 581 + ], + "type": "inline_equation", + "content": "F" + }, + { + "bbox": [ + 56, + 379, + 295, + 581 + ], + "type": "text", + "content": " denotes the input frames and " + }, + { + "bbox": [ + 56, + 379, + 295, + 581 + ], + "type": "inline_equation", + "content": "C" + }, + { + "bbox": [ + 56, + 379, + 295, + 581 + ], + "type": "text", + "content": " denotes the channel size." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 55, + 582, + 295, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 582, + 295, + 714 + ], + "spans": [ + { + "bbox": [ + 55, + 582, + 295, + 714 + ], + "type": "text", + "content": "CCIM. Since multi-view image generation requires the guidance of input object representation information, the Gaussian map that only indicates the location and trajectory of the object is insufficient. To establish a correspondence between the location of the representation points and the image features, we propose to use a trainable image prompt mapping network, which allows the image to function as an image prompt similar to the textual guidance condition. In order to ensure consistency between training and inference, we use only the first frame image " + }, + { + "bbox": [ + 55, + 582, + 295, + 714 + ], + "type": "inline_equation", + "content": "x_{1}" + }, + { + "bbox": [ + 55, + 582, + 295, + 714 + ], + "type": "text", + "content": " during training. The first frame image embedding is extracted by the pre" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 313, + 328, + 555, + 472 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 328, + 555, + 472 + ], + "spans": [ + { + "bbox": [ + 313, + 328, + 555, + 472 + ], + "type": "text", + "content": "trained CLIP [32] image encoder, and our image prompt mapping network projects this CLIP embedding into an image prompt feature " + }, + { + "bbox": [ + 313, + 328, + 555, + 472 + ], + "type": "inline_equation", + "content": "f_{ip} \\in \\mathbb{R}^{1 \\times C}" + }, + { + "bbox": [ + 313, + 328, + 555, + 472 + ], + "type": "text", + "content": ". Then we replace the features at the representation points' locations in the Gaussian map image features " + }, + { + "bbox": [ + 313, + 328, + 555, + 472 + ], + "type": "inline_equation", + "content": "f_{gi}" + }, + { + "bbox": [ + 313, + 328, + 555, + 472 + ], + "type": "text", + "content": " with this sequence of image prompt features. The resulting new Gaussian map image features are denoted as " + }, + { + "bbox": [ + 313, + 328, + 555, + 472 + ], + "type": "inline_equation", + "content": "f_{ngi} \\in \\mathbb{R}^{F \\times H \\times W \\times C}" + }, + { + "bbox": [ + 313, + 328, + 555, + 472 + ], + "type": "text", + "content": ". It is worth noting that since the start and end frames are randomly selected during training, all images in the training set with different views have the potential to be encoded as image prompts. This ensures diversity and richness in the model's ability to learn from varying perspectives." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 313, + 472, + 555, + 533 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 472, + 555, + 533 + ], + "spans": [ + { + "bbox": [ + 313, + 472, + 555, + 533 + ], + "type": "text", + "content": "Through the above analysis, the input condition of ControlNet is the sum of " + }, + { + "bbox": [ + 313, + 472, + 555, + 533 + ], + "type": "inline_equation", + "content": "f_{gi}" + }, + { + "bbox": [ + 313, + 472, + 555, + 533 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 313, + 472, + 555, + 533 + ], + "type": "inline_equation", + "content": "f_{ngi}" + }, + { + "bbox": [ + 313, + 472, + 555, + 533 + ], + "type": "text", + "content": ", that is " + }, + { + "bbox": [ + 313, + 472, + 555, + 533 + ], + "type": "inline_equation", + "content": "f_{input} = f_{gi} + f_{ngi}" + }, + { + "bbox": [ + 313, + 472, + 555, + 533 + ], + "type": "text", + "content": ". During training, we only optimize parameters of the image prompt mapping network and ControlNet, and the final optimization objective is:" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 347, + 545, + 555, + 560 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 347, + 545, + 555, + 560 + ], + "spans": [ + { + "bbox": [ + 347, + 545, + 555, + 560 + ], + "type": "interline_equation", + "content": "\\min _ {\\theta} E _ {x _ {0}, \\epsilon , t} \\| \\epsilon - \\epsilon_ {\\theta} (x _ {t}, t, c (G, G _ {n e w})) \\| ^ {2}, \\tag {6}", + "image_path": "11a27c08f73cf2c8b48fb29cb2433f73ac810465380ee58430e6f7894a51b502.jpg" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 313, + 565, + 555, + 615 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 565, + 555, + 615 + ], + "spans": [ + { + "bbox": [ + 313, + 565, + 555, + 615 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 313, + 565, + 555, + 615 + ], + "type": "inline_equation", + "content": "\\epsilon \\sim \\mathcal{N}(0,I)" + }, + { + "bbox": [ + 313, + 565, + 555, + 615 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 313, + 565, + 555, + 615 + ], + "type": "inline_equation", + "content": "t\\in [0,T]" + }, + { + "bbox": [ + 313, + 565, + 555, + 615 + ], + "type": "text", + "content": " denotes the time step of diffusion process, " + }, + { + "bbox": [ + 313, + 565, + 555, + 615 + ], + "type": "inline_equation", + "content": "c(\\cdot)" + }, + { + "bbox": [ + 313, + 565, + 555, + 615 + ], + "type": "text", + "content": " is the condition mechanism with explicit controls. " + }, + { + "bbox": [ + 313, + 565, + 555, + 615 + ], + "type": "inline_equation", + "content": "G" + }, + { + "bbox": [ + 313, + 565, + 555, + 615 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 313, + 565, + 555, + 615 + ], + "type": "inline_equation", + "content": "G_{new}" + }, + { + "bbox": [ + 313, + 565, + 555, + 615 + ], + "type": "text", + "content": " denote the initial and new Gaussian map injected with image features, respectively." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 313, + 626, + 397, + 640 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 626, + 397, + 640 + ], + "spans": [ + { + "bbox": [ + 313, + 626, + 397, + 640 + ], + "type": "text", + "content": "4. Experiments" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 313, + 647, + 430, + 659 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 647, + 430, + 659 + ], + "spans": [ + { + "bbox": [ + 313, + 647, + 430, + 659 + ], + "type": "text", + "content": "4.1. Experimental Setup" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 313, + 665, + 555, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 665, + 555, + 715 + ], + "spans": [ + { + "bbox": [ + 313, + 665, + 555, + 715 + ], + "type": "text", + "content": "Datasets. We utilize the widely adopted OmniObject3D dataset [42], containing approximately 6k 3D object models in real-world across more than 200 categories. We selected 2.4k 3D objects for training and 200 for testing. To" + } + ] + } + ], + "index": 12 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "type": "text", + "content": "21506" + } + ] + } + ], + "index": 13 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 4 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 67, + 68, + 228, + 273 + ], + "blocks": [ + { + "bbox": [ + 67, + 68, + 228, + 273 + ], + "lines": [ + { + "bbox": [ + 67, + 68, + 228, + 273 + ], + "spans": [ + { + "bbox": [ + 67, + 68, + 228, + 273 + ], + "type": "image", + "image_path": "ed21d843db3387f12549fc69d3736aa8f519fc02b25d89a252b3ce517a7dc5ad.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 54, + 281, + 555, + 348 + ], + "lines": [ + { + "bbox": [ + 54, + 281, + 555, + 348 + ], + "spans": [ + { + "bbox": [ + 54, + 281, + 555, + 348 + ], + "type": "text", + "content": "Figure 6. Comparison with existing methods. We compare existing methods relevant to our task. For 3D dataset images with pose labels, we directly compute the representation points, with the first row representing the ground truth image and its corresponding rotation. For real images, we utilize the interaction interface to define the desired drag target points (\"Our drag points\" in the figure). Additionally, we specify a drag trajectory for DragAnything that closely resembles our trajectory. DragAnything interprets the trajectory as the movement of the object or camera along it but cannot handle rotations. SV3D struggles with generalization to real-world images and lacks sufficient detail and 3D consistency. Wonder3D generates higher-quality 3D images but is limited to six fixed views." + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "type": "image", + "bbox": [ + 235, + 68, + 387, + 272 + ], + "blocks": [ + { + "bbox": [ + 235, + 68, + 387, + 272 + ], + "lines": [ + { + "bbox": [ + 235, + 68, + 387, + 272 + ], + "spans": [ + { + "bbox": [ + 235, + 68, + 387, + 272 + ], + "type": "image", + "image_path": "ef6f849b22b045db15210aa77528e21409ec34fc9c849efbf1a77354d023229c.jpg" + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_body" + } + ], + "index": 1 + }, + { + "type": "image", + "bbox": [ + 389, + 68, + 544, + 272 + ], + "blocks": [ + { + "bbox": [ + 389, + 68, + 544, + 272 + ], + "lines": [ + { + "bbox": [ + 389, + 68, + 544, + 272 + ], + "spans": [ + { + "bbox": [ + 389, + 68, + 544, + 272 + ], + "type": "image", + "image_path": "94053b6f4b9348ff728fb0f35ca8673251199aab5fc1cc9161cd539fb8f349d7.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + } + ], + "index": 2 + }, + { + "type": "table", + "bbox": [ + 63, + 365, + 288, + 439 + ], + "blocks": [ + { + "bbox": [ + 63, + 365, + 288, + 439 + ], + "lines": [ + { + "bbox": [ + 63, + 365, + 288, + 439 + ], + "spans": [ + { + "bbox": [ + 63, + 365, + 288, + 439 + ], + "type": "table", + "html": "
MethodsMulti-viewBackgroundTime
DragAnything×✓~40s
SV3D✓×~2min
Wonder3D✓×~2min
Ours✓✓~25s
", + "image_path": "076828d6b42cf57dedc1590424cb8eecb31d8d1590f87a92651d2b09dbf93134.jpg" + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "table_body" + } + ], + "index": 4 + }, + { + "bbox": [ + 55, + 447, + 295, + 504 + ], + "lines": [ + { + "bbox": [ + 55, + 447, + 295, + 504 + ], + "spans": [ + { + "bbox": [ + 55, + 447, + 295, + 504 + ], + "type": "text", + "content": "Table 1. Functionality comparisons. Our proposed Dragin3D achieves multi-view generation and supports editing images with backgrounds, ensuring efficient and rapid editing with minimal time consumption. In contrast, methods from different areas are unable to perform these functions simultaneously." + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "text" + }, + { + "bbox": [ + 55, + 529, + 295, + 577 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 529, + 295, + 577 + ], + "spans": [ + { + "bbox": [ + 55, + 529, + 295, + 577 + ], + "type": "text", + "content": "better simulate real-world scenes, we incorporate diverse background images into the dataset. Additionally, the image resolution is resized to " + }, + { + "bbox": [ + 55, + 529, + 295, + 577 + ], + "type": "inline_equation", + "content": "256*256" + }, + { + "bbox": [ + 55, + 529, + 295, + 577 + ], + "type": "text", + "content": " to optimize computational efficiency during training." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 55, + 579, + 295, + 663 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 579, + 295, + 663 + ], + "spans": [ + { + "bbox": [ + 55, + 579, + 295, + 663 + ], + "type": "text", + "content": "Baselines. Since Dragin3D integrates drag with multiview tasks, we use DragAnything [43], SV3D [38], and Wonder3D [22] as baselines. DragAnything and SV3D are trained on video framework, Wonder3D is trained on image framework. DragAnything is a 2D interaction editing method trained on video datasets, SV3D and Wonder3D are 3D generation methods trained on multi-view datasets." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 55, + 665, + 296, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 665, + 296, + 714 + ], + "spans": [ + { + "bbox": [ + 55, + 665, + 296, + 714 + ], + "type": "text", + "content": "Evaluation Metrics. For multi-view generation evaluation, we use the following three metrics for quantitative evaluation: For the quality of generated images, we employ Fréchet Inception Distance (FID), which calculates the" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 313, + 368, + 555, + 548 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 368, + 555, + 548 + ], + "spans": [ + { + "bbox": [ + 313, + 368, + 555, + 548 + ], + "type": "text", + "content": "Frechet distance between the feature vectors of the generated image and ground truth. For the perceptual similarity of the 3D effects, we use Learned Perceptual Image Patch Similarity (LPIPS), which evaluates the perceptual similarity between the generated image and the ground truth by learning the inverse mapping of the generated image. For the quality of the generated motion sequences, we apply Frechet Video Distance (FVD), which leverages a 3D feature extractor to compare video sequences. For drag trajectory accuracy evaluation, we adopt the MD metric in DragDiffusion, which employs DIFT [37] to identify the points in the images dragged by specified trajectories, corresponding to the starting representation points. The average Euclidean distance between the identified points and the reference trajectory points is then computed." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 313, + 549, + 555, + 609 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 549, + 555, + 609 + ], + "spans": [ + { + "bbox": [ + 313, + 549, + 555, + 609 + ], + "type": "text", + "content": "Implements Details. We randomly select 8 coherent object rotation images from the OmniObject3D dataset, which are used as the individual frames of video input to the SVD architecture. Eight A800 GPUs are used to train 20k steps with a batch size of 6, the learning rate is set to 2e-5." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 313, + 621, + 425, + 634 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 621, + 425, + 634 + ], + "spans": [ + { + "bbox": [ + 313, + 621, + 425, + 634 + ], + "type": "text", + "content": "4.2. Qualitative Results" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 313, + 641, + 555, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 641, + 555, + 715 + ], + "spans": [ + { + "bbox": [ + 313, + 641, + 555, + 715 + ], + "type": "text", + "content": "The visualization results of Dragin3D are shown in Fig. 5. Given an image with a corresponding trajectory, Dragin3D generates a high-fidelity image of the object from a new perspective, maintaining strong consistency with the original image. In addition, Dragin3D demonstrates superior generalization performance, which successfully rotates ob" + } + ] + } + ], + "index": 12 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "text", + "content": "21507" + } + ] + } + ], + "index": 13 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 5 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 59, + 69, + 293, + 203 + ], + "blocks": [ + { + "bbox": [ + 59, + 69, + 293, + 203 + ], + "lines": [ + { + "bbox": [ + 59, + 69, + 293, + 203 + ], + "spans": [ + { + "bbox": [ + 59, + 69, + 293, + 203 + ], + "type": "image", + "image_path": "1f0343b0fc7610f981d1a355897bd884ff149e2a267a918b014d68aa57fb61f5.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 55, + 213, + 295, + 247 + ], + "lines": [ + { + "bbox": [ + 55, + 213, + 295, + 247 + ], + "spans": [ + { + "bbox": [ + 55, + 213, + 295, + 247 + ], + "type": "text", + "content": "Figure 7. Comparison with 3D methods. We use the image without background as input to 3D methods and compare the results with ground truth, demonstrating that our method is comparable." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "bbox": [ + 54, + 268, + 295, + 291 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 54, + 268, + 295, + 291 + ], + "spans": [ + { + "bbox": [ + 54, + 268, + 295, + 291 + ], + "type": "text", + "content": "jects by dragging the representation points on real images that lie outside the training dataset distribution." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 56, + 293, + 296, + 555 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 293, + 296, + 555 + ], + "spans": [ + { + "bbox": [ + 56, + 293, + 296, + 555 + ], + "type": "text", + "content": "As shown in Fig. 7, we input the image into SV3D and Wonder3D after background removal and compare the results with the ground truth from the corresponding viewpoints. Our methods demonstrate comparable performance in terms of rotation accuracy and 3D consistency. Since our main contribution lies in the generalizability to real-world images, we include the background to reduce the gap between the ideal 3D dataset and the real-world scene. As illustrated in Fig. 6, we present the visualization results of each method with background using both 3D datasets and real images. The real images we provided have homogeneous backgrounds, minimizing their impact on the object. The results show that the performance of the other methods is significantly lower, limiting their practical applicability, while our method demonstrates a one-stage process that eliminates background interference during the image processing stage, performs well in detail preservation and coherent rotation. Additionally, we found that DragDiffusion [36] can achieve partial rotation in certain cases of Drag-Bench. However, the rotation is minimal and random, lacking precise control over the direction and angle of rotation as specified by the instructions, as illustrated in Fig. 8." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 55, + 563, + 174, + 575 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 563, + 174, + 575 + ], + "spans": [ + { + "bbox": [ + 55, + 563, + 174, + 575 + ], + "type": "text", + "content": "4.3. Quantitative Results" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 54, + 582, + 295, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 54, + 582, + 295, + 713 + ], + "spans": [ + { + "bbox": [ + 54, + 582, + 295, + 713 + ], + "type": "text", + "content": "Since our proposed task diverges from traditional drag-based editing and produces significantly different outcomes, and 3D generation models like Wonder3D focus on generating images without background from partial views, we do not include them in a quantitative comparison. Our Dragin3D shares a similar goal with SV3D, which aims to generate images from new perspectives based on a single input image. Therefore, we chose to make a comparison with SV3D. Due to the absence of ground truth in the real-world data, we randomly selected 200 subsets from the OmniObject3D dataset as the testset for computation, which were" + } + ] + } + ], + "index": 5 + }, + { + "type": "image", + "bbox": [ + 325, + 69, + 545, + 164 + ], + "blocks": [ + { + "bbox": [ + 325, + 69, + 545, + 164 + ], + "lines": [ + { + "bbox": [ + 325, + 69, + 545, + 164 + ], + "spans": [ + { + "bbox": [ + 325, + 69, + 545, + 164 + ], + "type": "image", + "image_path": "861afd2f4b5a8d5651b02fab108b6df5ca28e4ae4d7a86e3527af9502c7e2315.jpg" + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 313, + 172, + 555, + 217 + ], + "lines": [ + { + "bbox": [ + 313, + 172, + 555, + 217 + ], + "spans": [ + { + "bbox": [ + 313, + 172, + 555, + 217 + ], + "type": "text", + "content": "Figure 8. Comparison of rotation cases with DragDiffusion. DragDiffusion primarily manifests as tiny rotations induced by movements in the plane towards the target point, while Dragin3D enables more accurate control of stereoscopic rotation." + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "image_caption" + } + ], + "index": 6 + }, + { + "type": "table", + "bbox": [ + 339, + 228, + 529, + 278 + ], + "blocks": [ + { + "bbox": [ + 339, + 228, + 529, + 278 + ], + "lines": [ + { + "bbox": [ + 339, + 228, + 529, + 278 + ], + "spans": [ + { + "bbox": [ + 339, + 228, + 529, + 278 + ], + "type": "table", + "html": "
MethodsFID↓LPIPS↓FVD↓
SV3D [38]52.50.410944.6
Dragin3D (Ours)43.40.179584.6
", + "image_path": "2c72541890901a89f435718cb217fe150252e3803f4c81005deb3ae39cf0cc0b.jpg" + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "table_body" + } + ], + "index": 8 + }, + { + "type": "table", + "bbox": [ + 328, + 320, + 541, + 358 + ], + "blocks": [ + { + "bbox": [ + 313, + 286, + 553, + 308 + ], + "lines": [ + { + "bbox": [ + 313, + 286, + 553, + 308 + ], + "spans": [ + { + "bbox": [ + 313, + 286, + 553, + 308 + ], + "type": "text", + "content": "Table 2. Quantitative comparisons. Our proposed Dragin3D outperforms SV3D in all metrics." + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 328, + 320, + 541, + 358 + ], + "lines": [ + { + "bbox": [ + 328, + 320, + 541, + 358 + ], + "spans": [ + { + "bbox": [ + 328, + 320, + 541, + 358 + ], + "type": "table", + "html": "
MethodsDragDiffusion [36]Dragin3D (Ours)
MD↓33.019.38
", + "image_path": "0f6510974506bd3733a2227833ce27661006a1499174bab768fb1110b486856d.jpg" + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "table_body" + } + ], + "index": 10 + }, + { + "bbox": [ + 313, + 365, + 553, + 388 + ], + "lines": [ + { + "bbox": [ + 313, + 365, + 553, + 388 + ], + "spans": [ + { + "bbox": [ + 313, + 365, + 553, + 388 + ], + "type": "text", + "content": "Table 3. Quantitative comparisons. Our proposed Dragin3D outperforms DragDiffusion in control accuracy evaluation." + } + ] + } + ], + "index": 11, + "angle": 0, + "type": "text" + }, + { + "bbox": [ + 313, + 410, + 555, + 542 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 410, + 555, + 542 + ], + "spans": [ + { + "bbox": [ + 313, + 410, + 555, + 542 + ], + "type": "text", + "content": "not visible during training. We compared the similarity between the generated images and ground truth across the corresponding viewpoints from three perspectives: image, 3D perceptual, and video. As shown in Tab. 2, Dragin3D outperforms SV3D in all metrics. To evaluate the accuracy of trajectory control, we selected object rotation cases in Drag-Bench including animal, human, and sculpture heads. For each set of desired rotation trajectories, we computed the MDs for both Dragdiffusion and Dragin3D, with Dragin3D outperforming Dragdiffusion, as shown in Tab. 3. Additionally, we conducted a user study in Supp. 7.2." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 313, + 551, + 382, + 562 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 551, + 382, + 562 + ], + "spans": [ + { + "bbox": [ + 313, + 551, + 382, + 562 + ], + "type": "text", + "content": "4.4. Ablations" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 313, + 569, + 555, + 700 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 569, + 555, + 700 + ], + "spans": [ + { + "bbox": [ + 313, + 569, + 555, + 700 + ], + "type": "text", + "content": "Effectiveness of shape distinguish strategy (SDGS). To demonstrate the effectiveness of SDGS, we designed experiments in which the Gaussian maps for two points were constrained to the same shape, as shown in Tab. 4. We specifically focused on the symmetric rotation of objects, as illustrated in Fig. 9. Without distinguishing between the two points, the object in the subsequent stage of the rotation becomes difficult to distinguish from its symmetrical configuration in the preliminary stage. In contrast, the generated object preserves structural integrity and details when SDGS is applied. More examples can be found in Supp. 7.3." + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 326, + 701, + 553, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 326, + 701, + 553, + 713 + ], + "spans": [ + { + "bbox": [ + 326, + 701, + 553, + 713 + ], + "type": "text", + "content": "Effectiveness of control condition inject module" + } + ] + } + ], + "index": 15 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "text", + "content": "21508" + } + ] + } + ], + "index": 16 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 6 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 72, + 70, + 279, + 131 + ], + "blocks": [ + { + "bbox": [ + 72, + 70, + 279, + 131 + ], + "lines": [ + { + "bbox": [ + 72, + 70, + 279, + 131 + ], + "spans": [ + { + "bbox": [ + 72, + 70, + 279, + 131 + ], + "type": "table", + "html": "
MethodsFID↓LPIPS↓FVD↓
Dragin3D w/o SDGS44.10.182676.8
Dragin3D w/o CCIM45.20.194694.1
Dragin3D43.40.179584.6
", + "image_path": "a9acd42514dc7774dc047bf7d54083249a8ea71e7bf6b9e56af627876c119dcc.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "table_body" + } + ], + "index": 0 + }, + { + "bbox": [ + 55, + 140, + 295, + 161 + ], + "lines": [ + { + "bbox": [ + 55, + 140, + 295, + 161 + ], + "spans": [ + { + "bbox": [ + 55, + 140, + 295, + 161 + ], + "type": "text", + "content": "Table 4. Ablation study results. All metrics have declined without SDGS and CCIM, respectively." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "text" + }, + { + "type": "image", + "bbox": [ + 62, + 175, + 289, + 380 + ], + "blocks": [ + { + "bbox": [ + 62, + 175, + 289, + 380 + ], + "lines": [ + { + "bbox": [ + 62, + 175, + 289, + 380 + ], + "spans": [ + { + "bbox": [ + 62, + 175, + 289, + 380 + ], + "type": "image", + "image_path": "31c4543317ea4d6d144fbba6b8bc5abfcdf2a0fbceea39a314241bde2c7daa47.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 55, + 388, + 295, + 487 + ], + "lines": [ + { + "bbox": [ + 55, + 388, + 295, + 487 + ], + "spans": [ + { + "bbox": [ + 55, + 388, + 295, + 487 + ], + "type": "text", + "content": "Figure 9. Effectiveness of shape distinguish strategy (SDGS). We present the resulting images for two cases. The first row displays the trajectory points projected onto Gaussian maps, while the second and third rows show the objects from the corresponding view of these trajectory points, generated with and without SDGS, respectively. The positions of the two trajectory points in the first frame closely correspond to those in the last frame (green box). In the latter half of the sequence, significant misalignment occurs due to interference from the Gaussian representation (brown box)." + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_caption" + } + ], + "index": 2 + }, + { + "bbox": [ + 55, + 510, + 295, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 510, + 295, + 713 + ], + "spans": [ + { + "bbox": [ + 55, + 510, + 295, + 713 + ], + "type": "text", + "content": "(CCIM). As illustrated in Tab. 4, to assess the effectiveness of CCIM, we conducted experiments using only Gaussian map image features " + }, + { + "bbox": [ + 55, + 510, + 295, + 713 + ], + "type": "inline_equation", + "content": "f_{gi}" + }, + { + "bbox": [ + 55, + 510, + 295, + 713 + ], + "type": "text", + "content": " derived from the tracking points as the control input. The results reveal a marked enhancement across all evaluation metrics upon the integration of CCIM. Visual comparisons of intermediate results are presented in Fig. 10. Without CCIM, the object's rotation initially exhibits low-amplitude fluctuations, followed by a sudden and drastic change. Specifically, the early frames retain the original image, while later frames generate objects with significant distortion and poor fidelity. In contrast, the rotation transition is smooth and stable, and the object's details are preserved throughout. These results demonstrate that CCIM effectively aligns the rotation trajectory with image features by incorporating object-specific information, ensuring that the generated perspectives remain faithful to the original image content. More examples can be found in Supp. 7.3." + } + ] + } + ], + "index": 4 + }, + { + "type": "image", + "bbox": [ + 322, + 72, + 547, + 333 + ], + "blocks": [ + { + "bbox": [ + 322, + 72, + 547, + 333 + ], + "lines": [ + { + "bbox": [ + 322, + 72, + 547, + 333 + ], + "spans": [ + { + "bbox": [ + 322, + 72, + 547, + 333 + ], + "type": "image", + "image_path": "bd5319896f6c8f5d3d35247090c8b82f3ce1f284acefe9341986e29df39a9ed7.jpg" + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 313, + 342, + 555, + 430 + ], + "lines": [ + { + "bbox": [ + 313, + 342, + 555, + 430 + ], + "spans": [ + { + "bbox": [ + 313, + 342, + 555, + 430 + ], + "type": "text", + "content": "Figure 10. Effectiveness of control condition inject module (CCIM). We present the resulting images for two cases. The first row displays the trajectory points projected onto Gaussian maps, while the second and third rows show the objects from the corresponding view of these trajectory points, generated with and without CCIM, respectively. Without CCIM, the image remains largely unchanged in the early stages (red box), but later undergoes a sudden distortion, causing significant fidelity loss (yellow box)." + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "image_caption" + } + ], + "index": 5 + }, + { + "bbox": [ + 313, + 453, + 484, + 464 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 453, + 484, + 464 + ], + "spans": [ + { + "bbox": [ + 313, + 453, + 484, + 464 + ], + "type": "text", + "content": "5. Conclusion and Future Works" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 313, + 474, + 555, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 474, + 555, + 713 + ], + "spans": [ + { + "bbox": [ + 313, + 474, + 555, + 713 + ], + "type": "text", + "content": "In this paper, we present Dragin3D, which extends the interactive trajectory-based dragging task to 3D space. To represent the rotation trajectory of an object from the starting view to the terminating view, we introduce a double-point representation approach, and map the rotation of the object to a line segment formed by these two points. Our approach introduces control conditions through ControlNet architecture based on the large-scale video generation pretrained model SVD. In addition, we randomly add different backgrounds to the 3D object dataset, enabling the model to directly handle images with backgrounds. Through extensive experiments, Dragin3D successfully generates novel view images of the object based on the trajectories of two points, minimizing background interference while preserving the original object features, and demonstrating exceptional generalization on real-world images. For future work, since our approach currently supports object rotation in only one plane, we plan to expand our trajectory-based editing task to encompass a broader range of views, enabling arbitrary rotations in 3D space." + } + ] + } + ], + "index": 8 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "text", + "content": "21509" + } + ] + } + ], + "index": 9 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 7 + }, + { + "para_blocks": [ + { + "bbox": [ + 55, + 72, + 165, + 85 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 72, + 165, + 85 + ], + "spans": [ + { + "bbox": [ + 55, + 72, + 165, + 85 + ], + "type": "text", + "content": "6. Acknowledgments" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 55, + 91, + 296, + 163 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 91, + 296, + 163 + ], + "spans": [ + { + "bbox": [ + 55, + 91, + 296, + 163 + ], + "type": "text", + "content": "This research is supported by Artificial Intelligence-National Science and Technology Major Project 2023ZD0121200, the National Science Fund for Excellent Young Scholars under Grant 62222212, and National Natural Science Foundation of China under Grant 623B2094." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 56, + 175, + 115, + 187 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 175, + 115, + 187 + ], + "spans": [ + { + "bbox": [ + 56, + 175, + 115, + 187 + ], + "type": "text", + "content": "References" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 57, + 195, + 296, + 713 + ], + "type": "list", + "angle": 0, + "index": 14, + "blocks": [ + { + "bbox": [ + 61, + 195, + 296, + 239 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 61, + 195, + 296, + 239 + ], + "spans": [ + { + "bbox": [ + 61, + 195, + 296, + 239 + ], + "type": "text", + "content": "[1] Omer Bar-Tal, Dolev Ofri-Amar, Rafail Fridman, Yoni Kasten, and Tali Dekel. Text2live: Text-driven layered image and video editing. In European conference on computer vision, pages 707-723. Springer, 2022. 2" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 61, + 240, + 296, + 295 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 61, + 240, + 296, + 295 + ], + "spans": [ + { + "bbox": [ + 61, + 240, + 296, + 295 + ], + "type": "text", + "content": "[2] Andreas Blattmann, Tim Dockhorn, Sumith Kulal, Daniel Mendelevitch, Maciej Kilian, Dominik Lorenz, Yam Levi, Zion English, Vikram Voleti, Adam Letts, et al. Stable video diffusion: Scaling latent video diffusion models to large datasets. arXiv preprint arXiv:2311.15127, 2023. 3" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 61, + 297, + 296, + 361 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 61, + 297, + 296, + 361 + ], + "spans": [ + { + "bbox": [ + 61, + 297, + 296, + 361 + ], + "type": "text", + "content": "[3] Andreas Blattmann, Robin Rombach, Huan Ling, Tim Dockhorn, Seung Wook Kim, Sanja Fidler, and Karsten Kreis. Align your latents: High-resolution video synthesis with latent diffusion models. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 22563-22575, 2023. 3" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 62, + 364, + 296, + 418 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 364, + 296, + 418 + ], + "spans": [ + { + "bbox": [ + 62, + 364, + 296, + 418 + ], + "type": "text", + "content": "[4] Tim Brooks, Aleksander Holynski, and Alexei A Efros. Instructpix2pix: Learning to follow image editing instructions. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 18392-18402, 2023. 2" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 62, + 420, + 294, + 464 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 420, + 294, + 464 + ], + "spans": [ + { + "bbox": [ + 62, + 420, + 294, + 464 + ], + "type": "text", + "content": "[5] Weifeng Chen, Yatai Ji, Jie Wu, Hefeng Wu, Pan Xie, Jiashi Li, Xin Xia, Xuefeng Xiao, and Liang Lin. Control-a-video: Controllable text-to-video generation with diffusion models. arXiv preprint arXiv:2305.13840, 2023. 3" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 62, + 465, + 296, + 509 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 465, + 296, + 509 + ], + "spans": [ + { + "bbox": [ + 62, + 465, + 296, + 509 + ], + "type": "text", + "content": "[6] Guillaume Couairon, Jakob Verbeek, Holger Schwenk, and Matthieu Cord. Diffedit: Diffusion-based semantic image editing with mask guidance. arXiv preprint arXiv:2210.11427, 2022. 2" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 62, + 510, + 296, + 553 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 510, + 296, + 553 + ], + "spans": [ + { + "bbox": [ + 62, + 510, + 296, + 553 + ], + "type": "text", + "content": "[7] Yutao Cui, Xiaotong Zhao, Guozhen Zhang, Shengming Cao, Kai Ma, and Limin Wang. Stabledrag: Stable dragging for point-based image editing. arXiv preprint arXiv:2403.04437, 2024.1, 2" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 62, + 555, + 296, + 589 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 555, + 296, + 589 + ], + "spans": [ + { + "bbox": [ + 62, + 555, + 296, + 589 + ], + "type": "text", + "content": "[8] Yuki Endo. User-controllable latent transformer for stylegan image layout editing. In Computer Graphics Forum, pages 395-406. Wiley Online Library, 2022. 2" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 62, + 590, + 296, + 633 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 590, + 296, + 633 + ], + "spans": [ + { + "bbox": [ + 62, + 590, + 296, + 633 + ], + "type": "text", + "content": "[9] Dave Epstein, Allan Jabri, Ben Poole, Alexei Efros, and Aleksander Holynski. Diffusion self-guidance for controllable image generation. Advances in Neural Information Processing Systems, 36:16222-16239, 2023. 2" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 57, + 635, + 296, + 689 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 635, + 296, + 689 + ], + "spans": [ + { + "bbox": [ + 57, + 635, + 296, + 689 + ], + "type": "text", + "content": "[10] Yuwei Guo, Ceyuan Yang, Anyi Rao, Zhengyang Liang, Yaohui Wang, Yu Qiao, Maneesh Agrawala, Dahua Lin, and Bo Dai. Animatediff: Animate your personalized text-to-image diffusion models without specific tuning. arXiv preprint arXiv:2307.04725, 2023. 3" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 57, + 691, + 296, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 691, + 296, + 713 + ], + "spans": [ + { + "bbox": [ + 57, + 691, + 296, + 713 + ], + "type": "text", + "content": "[11] Amir Hertz, Ron Mokady, Jay Tenenbaum, Kfir Aberman, Yael Pritch, and Daniel Cohen-Or. Prompt-to-prompt im" + } + ] + } + ], + "index": 13 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 316, + 73, + 555, + 713 + ], + "type": "list", + "angle": 0, + "index": 28, + "blocks": [ + { + "bbox": [ + 333, + 73, + 555, + 95 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 333, + 73, + 555, + 95 + ], + "spans": [ + { + "bbox": [ + 333, + 73, + 555, + 95 + ], + "type": "text", + "content": "age editing with cross attention control. arXiv preprint arXiv:2208.01626, 2022. 2" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 316, + 96, + 555, + 148 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 96, + 555, + 148 + ], + "spans": [ + { + "bbox": [ + 316, + 96, + 555, + 148 + ], + "type": "text", + "content": "[12] Yaosi Hu, Chong Luo, and Zhenzhong Chen. Make it move: controllable image-to-video generation with text descriptions. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 18219-18228, 2022. 3" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 316, + 150, + 555, + 205 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 150, + 555, + 205 + ], + "spans": [ + { + "bbox": [ + 316, + 150, + 555, + 205 + ], + "type": "text", + "content": "[13] Tero Karras, Samuli Laine, Miika Aittala, Janne Hellsten, Jaakko Lehtinen, and Timo Aila. Analyzing and improving the image quality of stylegan. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 8110-8119, 2020. 1" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 316, + 206, + 554, + 260 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 206, + 554, + 260 + ], + "spans": [ + { + "bbox": [ + 316, + 206, + 554, + 260 + ], + "type": "text", + "content": "[14] Bahjat Kawar, Shiran Zada, Oran Lang, Omer Tov, Huiwen Chang, Tali Dekel, Inbar Mosseri, and Michal Irani. Imagc: Text-based real image editing with diffusion models. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 6007-6017, 2023. 2" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 316, + 261, + 554, + 304 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 261, + 554, + 304 + ], + "spans": [ + { + "bbox": [ + 316, + 261, + 554, + 304 + ], + "type": "text", + "content": "[15] Ali Köksal, Kenan E Ak, Ying Sun, Deepu Rajan, and Joo Hwee Lim. Controllable video generation with text-based instructions. IEEE transactions on multimedia, 26: 190-201, 2023. 3" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 316, + 305, + 554, + 360 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 305, + 554, + 360 + ], + "spans": [ + { + "bbox": [ + 316, + 305, + 554, + 360 + ], + "type": "text", + "content": "[16] Nupur Kumari, Bingliang Zhang, Richard Zhang, Eli Shechtman, and Jun-Yan Zhu. Multi-concept customization of text-to-image diffusion. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 1931-1941, 2023. 2" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 316, + 361, + 555, + 415 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 361, + 555, + 415 + ], + "spans": [ + { + "bbox": [ + 316, + 361, + 555, + 415 + ], + "type": "text", + "content": "[17] Jeong-gi Kwak, Erqun Dong, Yuhe Jin, Hanseok Ko, Shweta Mahajan, and Kwang Moo Yi. Vivid-1-to-3: Novel view synthesis with video diffusion models. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 6775–6785, 2024. 3" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 316, + 416, + 554, + 459 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 416, + 554, + 459 + ], + "spans": [ + { + "bbox": [ + 316, + 416, + 554, + 459 + ], + "type": "text", + "content": "[18] Dongxu Li, Junnan Li, and Steven Hoi. Blip-diffusion: Pretrained subject representation for controllable text-to-image generation and editing. Advances in Neural Information Processing Systems, 36, 2024. 2" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 317, + 460, + 554, + 514 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 460, + 554, + 514 + ], + "spans": [ + { + "bbox": [ + 317, + 460, + 554, + 514 + ], + "type": "text", + "content": "[19] Pengyang Ling, Lin Chen, Pan Zhang, Huaian Chen, Yi Jin, and Jinjin Zheng. Freedrag: Feature dragging for reliable point-based image editing. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 6860-6870, 2024. 1, 2" + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 316, + 515, + 554, + 568 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 515, + 554, + 568 + ], + "spans": [ + { + "bbox": [ + 316, + 515, + 554, + 568 + ], + "type": "text", + "content": "[20] Minghua Liu, Chao Xu, Haian Jin, Linghao Chen, Mukund Varma T, Zexiang Xu, and Hao Su. One-2-3-45: Any single image to 3d mesh in 45 seconds without per-shape optimization. Advances in Neural Information Processing Systems, 36, 2024. 3" + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 316, + 570, + 554, + 613 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 570, + 554, + 613 + ], + "spans": [ + { + "bbox": [ + 316, + 570, + 554, + 613 + ], + "type": "text", + "content": "[21] Yuan Liu, Cheng Lin, Zijiao Zeng, Xiaoxiao Long, Lingjie Liu, Taku Komura, and Wenping Wang. Syncdreamer: Generating multiview-consistent images from a single-view image. arXiv preprint arXiv:2309.03453, 2023." + } + ] + } + ], + "index": 25 + }, + { + "bbox": [ + 317, + 614, + 554, + 680 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 614, + 554, + 680 + ], + "spans": [ + { + "bbox": [ + 317, + 614, + 554, + 680 + ], + "type": "text", + "content": "[22] Xiaoxiao Long, Yuan-Chen Guo, Cheng Lin, Yuan Liu, Zhiyang Dou, Lingjie Liu, Yuexin Ma, Song-Hai Zhang, Marc Habermann, Christian Theobalt, et al. Wonder3d: Single image to 3d using cross-domain diffusion. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 9970-9980, 2024. 3, 6" + } + ] + } + ], + "index": 26 + }, + { + "bbox": [ + 316, + 681, + 554, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 681, + 554, + 713 + ], + "spans": [ + { + "bbox": [ + 316, + 681, + 554, + 713 + ], + "type": "text", + "content": "[23] Jingyi Lu, Xinghui Li, and Kai Han. Regiondrag: Fast region-based image editing with diffusion models. arXiv preprint arXiv:2407.18247, 2024. 2" + } + ] + } + ], + "index": 27 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 318, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 318, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 318, + 757 + ], + "type": "text", + "content": "21510" + } + ] + } + ], + "index": 29 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 8 + }, + { + "para_blocks": [ + { + "bbox": [ + 56, + 73, + 295, + 714 + ], + "type": "list", + "angle": 0, + "index": 13, + "blocks": [ + { + "bbox": [ + 56, + 73, + 294, + 106 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 73, + 294, + 106 + ], + "spans": [ + { + "bbox": [ + 56, + 73, + 294, + 106 + ], + "type": "text", + "content": "[24] Minxing Luo, Wentao Cheng, and Jian Yang. Rotationdrag: Point-based image editing with rotated diffusion features. arXiv preprint arXiv:2401.06442, 2024. 2" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 56, + 107, + 294, + 140 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 107, + 294, + 140 + ], + "spans": [ + { + "bbox": [ + 56, + 107, + 294, + 140 + ], + "type": "text", + "content": "[25] Wan-Duo Kurt Ma, John P Lewis, and W Bastiaan Kleijn. Trailblazer: Trajectory control for diffusion-based video generation. arXiv preprint arXiv:2401.00896, 2023. 3" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 56, + 141, + 294, + 195 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 141, + 294, + 195 + ], + "spans": [ + { + "bbox": [ + 56, + 141, + 294, + 195 + ], + "type": "text", + "content": "[26] Ben Mildenhall, Pratul P Srinivasan, Matthew Tancik, Jonathan T Barron, Ravi Ramamoorthi, and Ren Ng. Nerf: Representing scenes as neural radiance fields for view synthesis. Communications of the ACM, 65(1):99-106, 2021. 3" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 56, + 198, + 294, + 241 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 198, + 294, + 241 + ], + "spans": [ + { + "bbox": [ + 56, + 198, + 294, + 241 + ], + "type": "text", + "content": "[27] Chong Mou, Xintao Wang, Jiechong Song, Ying Shan, and Jian Zhang. Dragondiffusion: Enabling drag-style manipulation on diffusion models. arXiv preprint arXiv:2307.02421, 2023. 1, 2" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 56, + 243, + 295, + 297 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 243, + 295, + 297 + ], + "spans": [ + { + "bbox": [ + 56, + 243, + 295, + 297 + ], + "type": "text", + "content": "[28] Chong Mou, Xintao Wang, Jiechong Song, Ying Shan, and Jian Zhang. Diffeditor: Boosting accuracy and flexibility on diffusion-based image editing. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 8488-8497, 2024." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 56, + 299, + 294, + 342 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 299, + 294, + 342 + ], + "spans": [ + { + "bbox": [ + 56, + 299, + 294, + 342 + ], + "type": "text", + "content": "[29] Shen Nie, Hanzhong Allan Guo, Cheng Lu, Yuhao Zhou, Chenyu Zheng, and Chongxuan Li. The blessing of randomness: Sde beats ode in general diffusion-based image editing. arXiv preprint arXiv:2311.01410, 2023." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 56, + 344, + 294, + 399 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 344, + 294, + 399 + ], + "spans": [ + { + "bbox": [ + 56, + 344, + 294, + 399 + ], + "type": "text", + "content": "[30] Xingang Pan, Ayush Tewari, Thomas Leimkuhler, Lingjie Liu, Abhimitra Meka, and Christian Theobalt. Drag your gan: Interactive point-based manipulation on the generative image manifold. In ACM SIGGRAPH 2023 Conference Proceedings, pages 1-11, 2023. 1, 2" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 56, + 400, + 294, + 433 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 400, + 294, + 433 + ], + "spans": [ + { + "bbox": [ + 56, + 400, + 294, + 433 + ], + "type": "text", + "content": "[31] Ben Poole, Ajay Jain, Jonathan T Barron, and Ben Mildenhall. Dreamfusion: Text-to-3d using 2d diffusion. arXiv preprint arXiv:2209.14988, 2022. 3" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 56, + 434, + 294, + 499 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 434, + 294, + 499 + ], + "spans": [ + { + "bbox": [ + 56, + 434, + 294, + 499 + ], + "type": "text", + "content": "[32] Alec Radford, Jong Wook Kim, Chris Hallacy, Aditya Ramesh, Gabriel Goh, Sandhini Agarwal, Girish Sastry, Amanda Askell, Pamela Mishkin, Jack Clark, et al. Learning transferable visual models from natural language supervision. In International conference on machine learning, pages 8748-8763. PMLR, 2021. 5" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 56, + 501, + 294, + 567 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 501, + 294, + 567 + ], + "spans": [ + { + "bbox": [ + 56, + 501, + 294, + 567 + ], + "type": "text", + "content": "[33] Nataniel Ruiz, Yuanzhen Li, Varun Jampani, Yael Pritch, Michael Rubinstein, and Kfir Aberman. Dreambooth: Fine tuning text-to-image diffusion models for subject-driven generation. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 22500-22510, 2023. 2" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 56, + 568, + 294, + 622 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 568, + 294, + 622 + ], + "spans": [ + { + "bbox": [ + 56, + 568, + 294, + 622 + ], + "type": "text", + "content": "[34] Ruoxi Shi, Hansheng Chen, Zhuoyang Zhang, Minghua Liu, Chao Xu, Xinyue Wei, Linghao Chen, Chong Zeng, and Hao Su. Zero123++: a single image to consistent multi-view diffusion base model. arXiv preprint arXiv:2310.15110, 2023. 3" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 56, + 624, + 294, + 689 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 624, + 294, + 689 + ], + "spans": [ + { + "bbox": [ + 56, + 624, + 294, + 689 + ], + "type": "text", + "content": "[35] Xiaoyu Shi, Zhaoyang Huang, Fu-Yun Wang, Weikang Bian, Dasong Li, Yi Zhang, Manyuan Zhang, Ka Chun Cheung, Simon See, Hongwei Qin, et al. Motion-i2v: Consistent and controllable image-to-video generation with explicit motion modeling. In ACM SIGGRAPH 2024 Conference Papers, pages 1-11, 2024. 3" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 56, + 691, + 294, + 714 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 691, + 294, + 714 + ], + "spans": [ + { + "bbox": [ + 56, + 691, + 294, + 714 + ], + "type": "text", + "content": "[36] Yujun Shi, Chuhui Xue, Jun Hao Liew, Jiachun Pan, Hanshu Yan, Wenqing Zhang, Vincent YF Tan, and Song Bai." + } + ] + } + ], + "index": 12 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 316, + 73, + 553, + 714 + ], + "type": "list", + "angle": 0, + "index": 26, + "blocks": [ + { + "bbox": [ + 333, + 73, + 553, + 118 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 333, + 73, + 553, + 118 + ], + "spans": [ + { + "bbox": [ + 333, + 73, + 553, + 118 + ], + "type": "text", + "content": "Dragdiffusion: Harnessing diffusion models for interactive point-based image editing. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 8839-8849, 2024. 1, 2, 7" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 316, + 119, + 553, + 163 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 119, + 553, + 163 + ], + "spans": [ + { + "bbox": [ + 316, + 119, + 553, + 163 + ], + "type": "text", + "content": "[37] Luming Tang, Menglin Jia, Qianqian Wang, Cheng Perng Phoo, and Bharath Hariharan. Emergent correspondence from image diffusion. Advances in Neural Information Processing Systems, 36:1363-1389, 2023. 6" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 316, + 164, + 553, + 229 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 164, + 553, + 229 + ], + "spans": [ + { + "bbox": [ + 316, + 164, + 553, + 229 + ], + "type": "text", + "content": "[38] Vikram Voleti, Chun-Han Yao, Mark Boss, Adam Letts, David Pankratz, Dmitry Tochilkin, Christian Laforte, Robin Rombach, and Varun Jampani. Sv3d: Novel multi-view synthesis and 3d generation from a single image using latent video diffusion. In European Conference on Computer Vision, pages 439-457. Springer, 2025. 3, 6, 7" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 316, + 231, + 553, + 285 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 231, + 553, + 285 + ], + "spans": [ + { + "bbox": [ + 316, + 231, + 553, + 285 + ], + "type": "text", + "content": "[39] Xiang Wang, Hangjie Yuan, Shiwei Zhang, Dayou Chen, Jiuniu Wang, Yingya Zhang, Yujun Shen, Deli Zhao, and Jingren Zhou. Videocomposer: Compositional video synthesis with motion controllability. Advances in Neural Information Processing Systems, 36, 2024. 3" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 316, + 287, + 553, + 342 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 287, + 553, + 342 + ], + "spans": [ + { + "bbox": [ + 316, + 287, + 553, + 342 + ], + "type": "text", + "content": "[40] Zhouxia Wang, Ziyang Yuan, Xintao Wang, Yaowei Li, Tianshui Chen, Menghan Xia, Ping Luo, and Ying Shan. Motionctrl: A unified and flexible motion controller for video generation. In ACM SIGGRAPH 2024 Conference Papers, pages 1-11, 2024. 3" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 316, + 343, + 553, + 407 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 343, + 553, + 407 + ], + "spans": [ + { + "bbox": [ + 316, + 343, + 553, + 407 + ], + "type": "text", + "content": "[41] Yuxiang Wei, Yabo Zhang, Zhilong Ji, Jinfeng Bai, Lei Zhang, and Wangmeng Zuo. Elite: Encoding visual concepts into textual embeddings for customized text-to-image generation. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 15943-15953, 2023. 2" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 316, + 411, + 553, + 476 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 411, + 553, + 476 + ], + "spans": [ + { + "bbox": [ + 316, + 411, + 553, + 476 + ], + "type": "text", + "content": "[42] Tong Wu, Jiarui Zhang, Xiao Fu, Yuxin Wang, Jiawei Ren, Liang Pan, Wayne Wu, Lei Yang, Jiaqi Wang, Chen Qian, et al. Omniobject3d: Large-vocabulary 3d object dataset for realistic perception, reconstruction and generation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 803-814, 2023. 5" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 316, + 478, + 553, + 542 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 478, + 553, + 542 + ], + "spans": [ + { + "bbox": [ + 316, + 478, + 553, + 542 + ], + "type": "text", + "content": "[43] Weijia Wu, Zhuang Li, Yuchao Gu, Rui Zhao, Yefei He, David Junhao Zhang, Mike Zheng Shou, Yan Li, Tingting Gao, and Di Zhang. Draganything: Motion control for anything using entity representation. In European Conference on Computer Vision, pages 331-348. Springer, 2025. 1, 3, 5, 6" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 316, + 544, + 553, + 599 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 544, + 553, + 599 + ], + "spans": [ + { + "bbox": [ + 316, + 544, + 553, + 599 + ], + "type": "text", + "content": "[44] Haibo Yang, Yang Chen, Yingwei Pan, Ting Yao, Zhineng Chen, and Tao Mei. 3dstyle-diffusion: Pursuing fine-grained text-driven 3d stylization with 2d diffusion models. In Proceedings of the 31st ACM International Conference on Multimedia, pages 6860-6868, 2023. 3" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 316, + 601, + 553, + 645 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 601, + 553, + 645 + ], + "spans": [ + { + "bbox": [ + 316, + 601, + 553, + 645 + ], + "type": "text", + "content": "[45] Shengming Yin, Chenfei Wu, Jian Liang, Jie Shi, Houqiang Li, Gong Ming, and Nan Duan. Dragnuwa: Fine-grained control in video generation by integrating text, image, and trajectory. arXiv preprint arXiv:2308.08089, 2023. 1, 3" + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 316, + 647, + 553, + 689 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 647, + 553, + 689 + ], + "spans": [ + { + "bbox": [ + 316, + 647, + 553, + 689 + ], + "type": "text", + "content": "[46] Ziyang Yuan, Mingdeng Cao, Xintao Wang, Zhongang Qi, Chun Yuan, and Ying Shan. Customnet: Zero-shot object customization with variable-viewpoints in text-to-image diffusion models. arXiv preprint arXiv:2310.19784, 2023. 3" + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 316, + 691, + 553, + 714 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 691, + 553, + 714 + ], + "spans": [ + { + "bbox": [ + 316, + 691, + 553, + 714 + ], + "type": "text", + "content": "[47] Lvmin Zhang, Anyi Rao, and Maneesh Agrawala. Adding conditional control to text-to-image diffusion models. In" + } + ] + } + ], + "index": 25 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "text", + "content": "21511" + } + ] + } + ], + "index": 27 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 9 + }, + { + "para_blocks": [ + { + "bbox": [ + 56, + 72, + 295, + 229 + ], + "type": "list", + "angle": 0, + "index": 4, + "blocks": [ + { + "bbox": [ + 76, + 72, + 294, + 95 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 76, + 72, + 294, + 95 + ], + "spans": [ + { + "bbox": [ + 76, + 72, + 294, + 95 + ], + "type": "text", + "content": "Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 3836-3847, 2023. 4" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 56, + 96, + 295, + 140 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 96, + 295, + 140 + ], + "spans": [ + { + "bbox": [ + 56, + 96, + 295, + 140 + ], + "type": "text", + "content": "[48] Yabo Zhang, Yuxiang Wei, Dongsheng Jiang, Xiaopeng Zhang, Wangmeng Zuo, and Qi Tian. Controlvideo: Training-free controllable text-to-video generation. arXiv preprint arXiv:2305.13077, 2023. 3" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 56, + 141, + 294, + 184 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 141, + 294, + 184 + ], + "spans": [ + { + "bbox": [ + 56, + 141, + 294, + 184 + ], + "type": "text", + "content": "[49] Zhenghao Zhang, Junchao Liao, Menghao Li, Long Qin, and Weizhi Wang. Tora: Trajectory-oriented diffusion transformer for video generation. arXiv preprint arXiv:2407.21705, 2024.3" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 56, + 186, + 294, + 229 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 186, + 294, + 229 + ], + "spans": [ + { + "bbox": [ + 56, + 186, + 294, + 229 + ], + "type": "text", + "content": "[50] Chuanxia Zheng and Andrea Vedaldi. Free3d: Consistent novel view synthesis without 3d representation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 9720-9731, 2024. 3" + } + ] + } + ], + "index": 3 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 749, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 749, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 749, + 317, + 757 + ], + "type": "text", + "content": "21512" + } + ] + } + ], + "index": 5 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 10 + } + ], + "_backend": "vlm", + "_version_name": "2.6.4" +} \ No newline at end of file diff --git a/2025/DreamCache_ Finetuning-Free Lightweight Personalized Image Generation via Feature Caching/484a2287-cf26-4bf8-9e41-034501dc6452_content_list.json b/2025/DreamCache_ Finetuning-Free Lightweight Personalized Image Generation via Feature Caching/484a2287-cf26-4bf8-9e41-034501dc6452_content_list.json new file mode 100644 index 0000000000000000000000000000000000000000..e1e5f6f3309bfd478453bf6380773f5a349b095b --- /dev/null +++ b/2025/DreamCache_ Finetuning-Free Lightweight Personalized Image Generation via Feature Caching/484a2287-cf26-4bf8-9e41-034501dc6452_content_list.json @@ -0,0 +1,1737 @@ +[ + { + "type": "text", + "text": "DreamCache: Finetuning-Free Lightweight Personalized Image Generation via Feature Caching", + "text_level": 1, + "bbox": [ + 246, + 147, + 751, + 191 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Emanuele Aiello Umberto Michieli Diego Valsesia Mete Ozay Enrico Magli Politecnico di Torino Samsung R&D Institute UK", + "bbox": [ + 145, + 215, + 849, + 252 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Abstract", + "text_level": 1, + "bbox": [ + 259, + 282, + 333, + 299 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Personalized image generation requires text-to-image generative models that capture the core features of a reference subject to allow for controlled generation across different contexts. Existing methods face challenges due to complex training requirements, high inference costs, limited flexibility, or a combination of these issues. In this paper, we introduce DreamCache, a scalable approach for efficient and high-quality personalized image generation. By caching a small number of reference image features from a subset of layers and a single timestep of the pretrained diffusion denoiser, DreamCache enables dynamic modulation of the generated image features through lightweight, trained conditioning adapters. DreamCache achieves state-of-the-art image and text alignment, utilizing an order of magnitude fewer extra parameters, and is both more computationally effective and versatile than existing models.", + "bbox": [ + 107, + 316, + 485, + 546 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "1. Introduction", + "text_level": 1, + "bbox": [ + 109, + 566, + 235, + 580 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Recent advancements in text-to-image generation, fueled by the development of diffusion models [10, 28], have enabled high-quality and diverse image generation from textual descriptions. Diffusion models [23, 25] gradually transform random noise into images through a sequence of denoising steps, conditioned on the input text prompt.", + "bbox": [ + 107, + 585, + 483, + 672 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "An active area of research is personalizing these models, enabling the generation of novel images of a reference subject in various contexts, while maintaining flexibility for text-based editing. Early personalization techniques [1, 6, 8, 24, 30, 32, 35, 35], such as the seminal Dream-Booth [24] relied on fine-tuning (FT) the generative model for each reference subject. However, these approaches are often impractical for many use cases due to costly test-time FT, which can take several minutes per subject. To address this, FT-free (i.e., zero-shot) personalized image generation methods have emerged to eliminate test-time optimization. These FT-free approaches can be broadly categorized into", + "bbox": [ + 107, + 672, + 485, + 847 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/2ff59b4bdbab438456d466a87c6e85d3cb828d258db8bfe36cf3490f74168362.jpg", + "image_caption": [ + "Figure 1. DreamCache is a finetuning-free personalized image generation method that achieves an optimal balance between subject fidelity, memory efficiency, and adherence to text prompts." + ], + "image_footnote": [], + "bbox": [ + 581, + 285, + 818, + 467 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "two families: encoder-based methods and reference-based methods, each with distinct drawbacks.", + "bbox": [ + 511, + 547, + 885, + 575 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Encoder-based methods [7, 15, 20, 33, 36] utilize dedicated image encoders, such as CLIP [22] or DINO [2], to extract relevant features from reference images. While these encoders can produce high-quality results, they are often large, require extensive training to align text and image features, and reduce the model's flexibility [13, 15, 19, 36].", + "bbox": [ + 509, + 577, + 888, + 662 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "In contrast, reference-based methods [21, 37] condition the diffusion model directly on reference features drawn from the U-Net denoiser, integrating these features at each denoising step. While effective, these methods require feature extraction at every generation step, leading to higher computational costs and memory demands. Additionally, they often require an input textual caption for conditioning, which introduces variability and can decrease output precision.", + "bbox": [ + 509, + 665, + 888, + 794 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Some recent works have proposed to finetune the U-Net backbone itself [21, 37, 39]. However, this hinders the model's ability to switch between personalized and nonpersonalized tasks and risks inducing the \"language drift\" phenomenon, where the personalization training degrades the model's linguistic comprehension [12, 24].", + "bbox": [ + 509, + 795, + 888, + 882 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "CVF", + "bbox": [ + 116, + 16, + 189, + 54 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore.", + "bbox": [ + 243, + 13, + 799, + 56 + ], + "page_idx": 0 + }, + { + "type": "page_footnote", + "text": "$^{1}$ Code, model, and synthetic dataset can be found at https://github.com/Emanuele97x/DreamCache", + "bbox": [ + 109, + 857, + 485, + 881 + ], + "page_idx": 0 + }, + { + "type": "page_number", + "text": "12480", + "bbox": [ + 480, + 922, + 517, + 934 + ], + "page_idx": 0 + }, + { + "type": "table", + "img_path": "images/6e415f08761cafc5535bfeb4985c595b0b5ee0160e328b3c0a90082a61e0db16.jpg", + "table_caption": [ + "Table 1. Methods overview. Our DreamCache achieves state-of-the-art generation quality at reduced computational costs. *: value refers to the personalization stage for each personal subject." + ], + "table_footnote": [], + "table_body": "
MethodFT-freeEnc-freePlug&PlayRef-UNet-freeExtra ParamsTrain Params# DatasetTrain Time
Textual Inversion [6]X✓✓✓768*768*3-5*50 min*
DreamBooth [24]X✓X✓-0.9B*3-5*10 min*
Custom Diffusion [12]X✓X✓-57M*3-5*10 min*
ELITE [33]✓XX✓457M77M125K14 days
BLIP-Diffusion [13]✓X✓✓380M1.5B129M96 days
IP-Adapter [36]✓X✓✓402M22M10M28 days
Kosmos-G [19]✓X✓✓1.6B1.6B9M-
JeDi [37]✓✓XX-0.9B3M48 days
SuTI [4]✓XX✓400M2.5B500K-
Subject-Diffusion [15]✓X✓X700M700M76M-
BootPig [21]✓✓XX0.95B0.95B200K18 hours
ToffeeNet [39]✓XX✓632M0.9B5M-
CAFE [38]✓XX✓14B1B355K-
DreamCache (ours)✓✓✓✓25M25M400K40 hours
", + "bbox": [ + 122, + 145, + 874, + 383 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "In this work, we propose DreamCache, a novel finetuning-free approach to personalized image generation, that overcomes the limitations of existing methods (see Fig. 1) by using a feature caching mechanism that enables text-free encoding and efficient conditioning during personalization. Specifically, we first create a synthetic dataset [21] containing triplets of captions, target images, and reference subjects to capture subjects in various contexts. Next, we pretrain lightweight attention-based conditioning adapters to inject subject-specific features into the image generation process. During personalization, the reference image is processed through the pretrained denoiser of the base diffusion model without text conditioning, thus eliminating the need for user-generated captions, while caching features from a small subset of layers at a single timestep. For personalized sampling, these cached features are injected into the denoiser through the pretrained conditioning adapters.", + "bbox": [ + 112, + 407, + 483, + 652 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Table 1 summarizes the key properties of existing methods and illustrates how DreamCache fits within the current landscape; further details are explored in Sec. 2. As an encoder-free approach, DreamCache introduces only a small number of additional parameters, making it significantly lightweight and suitable for deployment on resource-constrained devices. For example, methods like [33] and [13] introduce 380M parameters due to their reliance on CLIP encoders, whereas DreamCache requires only 25M additional parameters. Moreover, caching features from a few selected U-Net layers at a single preprocessing timestep bypasses the need for full U-Net reference processing during generation, leading to substantial computational and memory savings that enable real-time, high-quality personalized generation. Another key advantage of DreamCache is its plug-and-play", + "bbox": [ + 112, + 652, + 483, + 869 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "nature, allowing concurrent generation of personalized and non-personalized content without altering original U-Net weights, thus preserving the integrity of the original model and enabling a wider range of deployment scenarios, especially on mobile platforms.", + "bbox": [ + 514, + 407, + 885, + 479 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "In summary, DreamCache represents a significant step toward practical, and scalable personalized image generation, with the following contributions:", + "bbox": [ + 514, + 481, + 885, + 523 + ], + "page_idx": 1 + }, + { + "type": "list", + "sub_type": "text", + "list_items": [ + "- We propose a feature caching approach that creates multi-resolution representations of the reference image in a caption-free and efficient manner.", + "- We design an attention-based conditioning mechanism that leverages the cached features for personalized image generation, achieving computational- and memory- efficient personalized sampling.", + "- Our approach achieves state-of-the-art quality in personalized image generation at substantially lower computational and data costs compared to existing methods." + ], + "bbox": [ + 511, + 527, + 885, + 669 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "2. Background and Related Work", + "text_level": 1, + "bbox": [ + 514, + 686, + 784, + 700 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Personalized image generation aims at generating images containing a specific subject. This task has been widely studied, resulting in two main approaches: fine-tuning methods, which require test-time finetuning on multiple subject reference images, and finetuning-free (zero-shot) methods, which learn a generalizable conditioning mechanism to generate reference subjects without the need for further optimization.", + "bbox": [ + 514, + 710, + 885, + 810 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Finetuning-based Personalization DreamBooth [24] finetunes the entire U-Net with reference images while introducing a regularization loss to mitigate overfitting. On the other hand, Custom Diffusion [12] only finetunes the $K$ and", + "bbox": [ + 514, + 825, + 885, + 880 + ], + "page_idx": 1 + }, + { + "type": "page_number", + "text": "12481", + "bbox": [ + 480, + 924, + 516, + 934 + ], + "page_idx": 1 + }, + { + "type": "image", + "img_path": "images/72e59097773136a3ca128e6e35d632f15e0757d7e32b35195cf460eb0822589d.jpg", + "image_caption": [ + "\"A dragon...\"" + ], + "image_footnote": [], + "bbox": [ + 114, + 122, + 264, + 237 + ], + "page_idx": 2 + }, + { + "type": "image", + "img_path": "images/7063e425d55245aac90538df6ccdc4c59e0aa1dc7fdda3987894df887d553467.jpg", + "image_caption": [ + "\"as street graffiti\"" + ], + "image_footnote": [], + "bbox": [ + 266, + 125, + 415, + 237 + ], + "page_idx": 2 + }, + { + "type": "image", + "img_path": "images/a8d50836b16284dc7b021527485e2a63d2f65a1c505a1c98940637cd5f9f63ac.jpg", + "image_caption": [ + "\"playing with fire\"" + ], + "image_footnote": [], + "bbox": [ + 418, + 122, + 566, + 237 + ], + "page_idx": 2 + }, + { + "type": "image", + "img_path": "images/3288fe34e86e7c2c5690b5f415cf6ce77dd50e11d50a4ce0fc0845a094bada51.jpg", + "image_caption": [ + "\"as a plushie\"" + ], + "image_footnote": [], + "bbox": [ + 570, + 123, + 718, + 237 + ], + "page_idx": 2 + }, + { + "type": "image", + "img_path": "images/42cba27336adc0a3a020f273eee03a02b1786e4b1912d7211f0edb071baf6f56.jpg", + "image_caption": [ + "\"working as a barista\"" + ], + "image_footnote": [], + "bbox": [ + 725, + 122, + 875, + 237 + ], + "page_idx": 2 + }, + { + "type": "image", + "img_path": "images/b8fd579d3cf15ec40ec046824de141036ba2da37a0d3e1c0c12a7a18972c5a68.jpg", + "image_caption": [ + "\"A cat...\"" + ], + "image_footnote": [], + "bbox": [ + 114, + 253, + 263, + 368 + ], + "page_idx": 2 + }, + { + "type": "image", + "img_path": "images/00db6d3885758325f7772d3f86d0e88bf4b11f9b17c537c25770b5e27d7f2d12.jpg", + "image_caption": [ + "\"in Ukiyo-e style\"" + ], + "image_footnote": [], + "bbox": [ + 264, + 255, + 415, + 367 + ], + "page_idx": 2 + }, + { + "type": "image", + "img_path": "images/957f74af1e220a4a34c67d7f3ce19eda399f8a1118a1aa1270ce965557b246c9.jpg", + "image_caption": [ + "\"with a rainbow scarf\"", + "Figure 2. Personalized generations by DreamCache. The first column contains reference images. The generated images correspond to the text prompts above each column." + ], + "image_footnote": [], + "bbox": [ + 416, + 253, + 568, + 369 + ], + "page_idx": 2 + }, + { + "type": "image", + "img_path": "images/969b036c282fc912b595c03d260274a3a16a411b9f73873cf36260c6eae53b59.jpg", + "image_caption": [ + "\"Van Gogh painting\"" + ], + "image_footnote": [], + "bbox": [ + 570, + 253, + 718, + 369 + ], + "page_idx": 2 + }, + { + "type": "image", + "img_path": "images/5a3086329e7c27ee0b1766b5132b3ec444206b07702c156a24d8bbb408726be4.jpg", + "image_caption": [ + "\"wearing a diploma hat\"" + ], + "image_footnote": [], + "bbox": [ + 725, + 253, + 875, + 369 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "$V$ projections for the cross-attention blocks of the U-Net. Text-based personalization methods optimize single (like Textual Inversion [6]) or multiple (like in $\\mathrm{P + }$ [32]) input token embeddings. Later methods [1, 8, 30, 35] build on these, with innovations like Perfusion [30] using dynamic rank-1 updates to prevent overfitting while keeping encodings lightweight. However, all finetuning-based methods are computationally intensive, often requiring minutes of finetuning per reference subject at test time.", + "bbox": [ + 107, + 436, + 485, + 566 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Finetuning-Free Personalization To reduce computational demands, recent research has shifted toward zero-shot personalization methods that eliminate subject-specific finetuning, typically employing image encoders to condition the generation process via the features extracted from the reference images [7, 15, 20, 33, 36]. Examples include BLIP-Diffusion [13], which pretrains a Q-Former to learn image features aligned with text, and IP-Adapter, which uses a frozen CLIP encoder to extract text-aligned visual features that modulate the cross-attention layers of the generative model. Other approaches, like Kosmos-G [19] and CAFE [38], connect large language models (LLMs) with diffusion models to condition generation on personalized concepts. SuTI [4] takes a different approach by training millions of subject-specific experts and subsequently training a model via apprenticeship learning, enabling effective zero-shot personalized generation at test time.", + "bbox": [ + 107, + 579, + 485, + 823 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Alternatively, encoder-free methods such as JeDi [37] and BootPIG [21] use features from the generative model's backbone to guide the generation. JeDi creates a multi-view synthetic dataset and modifies spatial self-attention to jointly", + "bbox": [ + 107, + 824, + 483, + 882 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "attend to images of the same concept in a batch. BootPIG retains a trainable copy of the original U-Net, adding reference self-attention layers to enable adaptation of the personalized model for reference features. While these methods remove the need for an additional encoder, they still require computationally expensive inference, as reference images must be processed in parallel during generation—a cost that accumulates due to the iterative nature of the diffusion process.", + "bbox": [ + 511, + 436, + 888, + 551 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "In contrast, our method, DreamCache, caches a subset of reference features from the U-Net without text conditioning, eliminating the need for parallel inference and reducing memory overhead by avoiding separate model loading. This results in a test-time computational efficiency similar to encoder-based methods while offering the flexibility of encoder-free feature extraction and injection.", + "bbox": [ + 511, + 551, + 888, + 652 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Recent works such as BootPIG [21] and Toffee-5M [39] emphasize the importance of synthetic data that explicitly decouples the subject from the background, reporting improved performance. Inspired by these approaches, we adopt a similar generative pipeline to create a synthetic dataset we use to train DreamCache.", + "bbox": [ + 511, + 652, + 888, + 738 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Moreover, since our method is plug-and-play and keeps the U-Net frozen, we reduce the high cost of training faced by other approaches [15, 19, 21, 37-39]. A detailed overview of current methods, the number of trained parameters, and their training cost can be found in Table 1.", + "bbox": [ + 511, + 739, + 887, + 811 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Feature Caching Feature caching has been explored to reduce generation time in diffusion models by caching intermediate activations. Some studies [17, 34] exploit temporal redundancy during the training process to cache ac", + "bbox": [ + 511, + 824, + 888, + 882 + ], + "page_idx": 2 + }, + { + "type": "page_number", + "text": "12482", + "bbox": [ + 480, + 922, + 517, + 934 + ], + "page_idx": 2 + }, + { + "type": "image", + "img_path": "images/cc90f61ffcc2fdd5a17e9ca033fa3489b1b277e98fc320d9267a8e3a9e803009.jpg", + "image_caption": [ + "Figure 3. Overview of DreamCache. Original U-Net layers are shown in violet, while the novel components introduced by DreamCache are highlighted in green. During personalization, features from selected layers of the diffusion denoiser are cached from a single timestep, using a null text prompt. These cached features serve as reference-specific information. During generation, conditioning adapters inject the cached features into the denoiser, modulating the features of the generated image to create a personalized output." + ], + "image_footnote": [], + "bbox": [ + 133, + 116, + 870, + 344 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "tivations across timesteps, reducing computational load at later timesteps. Other works focus on caching layer activations within the diffusion framework, avoiding redundant computations. Learning-to-Cache [16] introduces a dynamic caching mechanism that learns to skip computation for selected layers of the diffusion model. In contrast to those works, which generally cache intra-model features for some layers to save computations, we utilize feature caching to encode multi-resolution features of a reference image from a few selected layers to condition the generation process of a new personalized image. Our approach recalls successful few-shot learners for discriminative problems [18, 27, 31] and extends them to personalized image generation.", + "bbox": [ + 107, + 441, + 485, + 628 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "3. Method", + "text_level": 1, + "bbox": [ + 109, + 641, + 194, + 657 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Given a pretrained text-to-image generative model $\\epsilon_{\\theta}$ and an image containing a reference subject $\\mathbf{I}_{\\mathrm{ref}}$ , the goal of personalized sampling is to generate novel images containing the reference subject in various contexts while maintaining textual control. We propose DreamCache, a novel approach for extracting conditioning signals from $\\mathbf{I}_{\\mathrm{ref}}$ and guiding of the image generation process. This method leverages a pretrained diffusion model, conditioning adapters that are pretrained with a synthetic dataset, and a feature cache from the reference image. Sample outputs generated by DreamCache are shown in Fig. 2, with a method overview in Fig. 3.", + "bbox": [ + 107, + 665, + 485, + 823 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "At the core of DreamCache, we utilize the denoiser in the pretrained diffusion model to extract multi-resolution features from $\\mathbf{I}_{\\mathrm{ref}}$ by caching the activations of a few selected layers. To improve generalization, we cache features", + "bbox": [ + 107, + 824, + 485, + 882 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "using a forward pass with a null text prompt. When personalized sampling is performed, the cached features are processed by adapters to act as conditioning signals, modulating the denoiser features of the image under generation at corresponding layers. These adapters, once pretrained on a synthetic dataset, enable zero-shot personalized generation with any new reference image, requiring no further finetuning once its features are cached.", + "bbox": [ + 511, + 441, + 888, + 556 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "In the following, we detail the three main aspects of DreamCache, namely i) how to cache reference features (Sec. 3.1); ii) how to condition the diffusion model on the cached features for personalized sampling (Sec. 3.2); iii) how to train the adapters used for model conditioning (Sec. 3.3).", + "bbox": [ + 511, + 559, + 888, + 631 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "3.1. Caching Reference Features", + "text_level": 1, + "bbox": [ + 511, + 641, + 754, + 657 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "To extract information from the reference image for personalized sampling, we perform a forward pass through the denoiser of the diffusion model at a single timestep. We select $t = 1$ , the least noisy timestep, to obtain clean features that are optimal for conditioning the personalized generation process. Additionally, we remove the text conditioning to decouple visual content of the reference image from the text caption, thus also eliminating the need for user-provided captions for reference images. This contrasts with methods such as JeDi [37], which are sensitive to caption content. During the forward pass, activations are computed for all layers of the denoiser, but only a subset is cached. Based on our experiments with the Stable Diffusion U-Net, we find that caching features from a middle bottleneck layer and every second layer in the decoder offers the best balance between", + "bbox": [ + 509, + 666, + 888, + 882 + ], + "page_idx": 3 + }, + { + "type": "page_number", + "text": "12483", + "bbox": [ + 480, + 922, + 517, + 934 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "generation quality and caching efficiency (see Sec. 4.3).", + "bbox": [ + 109, + 109, + 460, + 125 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Formally, the feature cache $\\mathcal{H}_{\\mathrm{FC}}$ consists of the activations of the denoiser $\\epsilon_{\\theta}$ from the selected layers $\\mathcal{L}$ at timestep $t = 1$ , using a null text prompt $\\varnothing$ and noisy reference image $\\mathbf{I}_{\\mathrm{ref}} + \\mathbf{n}_t$ , with noise realization $\\mathbf{n}_t$ , expressed as:", + "bbox": [ + 109, + 125, + 483, + 183 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {H} _ {\\mathrm {F C}} = \\left\\{\\mathrm {h} _ {\\text {r e f}, L}: L \\in \\mathcal {L} \\right\\} \\tag {1}\n$$\n", + "text_format": "latex", + "bbox": [ + 184, + 196, + 483, + 212 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\mathbf {h} _ {\\text {r e f}, L} = \\epsilon_ {\\theta} \\left(\\mathbf {I} _ {\\text {r e f}} + \\mathbf {n} _ {t}, \\varnothing , t; l) \\right| _ {t = 1, l = L}. \\tag {2}\n$$\n", + "text_format": "latex", + "bbox": [ + 176, + 214, + 483, + 229 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Notice that the cached features have different spatial resolutions, from the low-resolution bottleneck layer to the higher-resolution decoder layers, allowing a multi-resolution representation of the reference image. This is particularly useful for enabling both global semantics and fine-grained detail guidance. As in prior work [15, 21, 33], we foreground-segment the reference image to isolate the subject from the background before caching its features.", + "bbox": [ + 107, + 244, + 485, + 359 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "3.2. Conditioning on Cached Reference Features", + "text_level": 1, + "bbox": [ + 109, + 371, + 467, + 386 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "We propose a novel conditioning adapter mechanism composed of: i) a cross-attention block between the cached features and the features of the image under generation; ii) a concatenation operation between the features from the self-attention block of the original U-Net denoising backbone and those of the cross-attention block; and iii) a projection layer. A block diagram is shown in Fig. 3 (right).", + "bbox": [ + 107, + 393, + 483, + 494 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Omitting the layer subscript for clarity, the conditioning adapter mechanism is expressed mathematically as follows:", + "bbox": [ + 109, + 496, + 483, + 525 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\mathbf {q} = \\mathbf {W} _ {Q} \\mathbf {h}, \\quad \\mathbf {k} _ {c} = \\mathbf {W} _ {K} \\mathbf {h} _ {\\text {r e f}}, \\quad \\mathbf {v} _ {c} = \\mathbf {W} _ {V} \\mathbf {h} _ {\\text {r e f}}, \\tag {3}\n$$\n", + "text_format": "latex", + "bbox": [ + 148, + 553, + 483, + 568 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\mathbf {a} _ {c} = \\operatorname {s o f t m a x} \\left(\\frac {\\mathbf {q k} _ {c} ^ {T}}{\\sqrt {d}}\\right) \\mathbf {v} _ {c}, \\tag {4}\n$$\n", + "text_format": "latex", + "bbox": [ + 147, + 571, + 483, + 603 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\mathbf {a} = \\mathbf {W} _ {\\text {p r o j}} \\left(\\left[ \\mathbf {a}; \\mathbf {a} _ {c} \\right]\\right), \\tag {5}\n$$\n", + "text_format": "latex", + "bbox": [ + 147, + 607, + 483, + 622 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "where $\\mathbf{h} \\in \\mathbb{R}^{N \\times d}$ are the current $d$ -dimensional features of the $N$ -pixel image under generation, and $\\mathbf{h}_{\\mathrm{ref}}$ are the cached reference features from Eq. (2). $\\mathbf{W}_Q$ , $\\mathbf{W}_K$ , $\\mathbf{W}_V$ , and $\\mathbf{W}_{\\mathrm{proj}}$ are learnable projection matrices, whose training is described in Sec. 3.3. The concatenation $[\\mathbf{a}; \\mathbf{a}_c] \\in \\mathbb{R}^{N \\times 2d}$ combines the output of self-attention $\\mathbf{a} \\in \\mathbb{R}^{N \\times d}$ and cross-attention $\\mathbf{a}_c \\in \\mathbb{R}^{N \\times d}$ . The concatenation operation allows a flexible information fusion without explicit alignment constraints, compared to other approaches in similar works (see Sec. 4.3). The learnable projection matrix $\\mathbf{W}_{\\mathrm{proj}}$ reduces the dimensionality of the concatenated features back to $\\mathbb{R}^{N \\times d}$ to maintain compatibility with the original backbone.", + "bbox": [ + 107, + 635, + 485, + 808 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Overall, the approach proposed for the adapter enriches feature representations used in the diffusion process of the image under generation by allowing the model to leverage both primary and conditioning-based contextual information from the cache.", + "bbox": [ + 107, + 810, + 483, + 881 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "3.3. Training the Conditioning Adapters", + "text_level": 1, + "bbox": [ + 511, + 109, + 808, + 125 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "The additional parameters introduced in Sec. 3.2 to process the cached features must be trained on a large and varied dataset to ensure they generalize for any reference subject.", + "bbox": [ + 511, + 131, + 885, + 175 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Collecting paired data for this training process would be prohibitively expensive, as it requires multiple images of the same subject in different contexts. To address this, we draw inspiration from the recently proposed synthetic data generation pipeline in BootPIG [21] to construct our training data. First, we utilize a large language model (Llama 3.2 [5]) to generate captions for potential target images. Each caption is used to generate an image via Stable Diffusion [23]. We then use the Segment Anything Model (SAM) [11] and Grounding DINO [14] to accurately segment the reference subject based on the text caption and generate a foreground mask of the main object in the caption.", + "bbox": [ + 511, + 175, + 887, + 347 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "We treat the Stable Diffusion-generated image as the target image, the foreground object pasted on a white background as the reference image, and the LLM-generated caption as the textual prompt during our training pipeline. Compared to BootPIG, our pipeline employs open-source models, making it more accessible. We will release our synthetic dataset to facilitate reproducibility and further research, since similar datasets, including BootPIG's [21], have not been publicly released. Additional details on the dataset and its statistics can be found in the Supp. Mat.", + "bbox": [ + 511, + 347, + 888, + 491 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "We train the newly introduced adapters' parameters $(\\mathbf{W}_Q, \\mathbf{W}_K, \\mathbf{W}_V, \\text{and} \\mathbf{W}_{\\mathrm{proj}})$ with the standard score matching loss [29] using both the text-conditioned noisy input and the cached reference features:", + "bbox": [ + 511, + 491, + 888, + 547 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {L} _ {\\text {d i f f u s i o n}} = \\mathbb {E} _ {\\mathbf {x} _ {0}, \\epsilon , \\mathbf {c} _ {T}, \\mathbf {I} _ {\\text {r e f}}, t} \\left[ \\| \\epsilon - \\epsilon_ {\\theta} ^ {\\prime} \\left(\\mathbf {x} _ {t}, \\mathbf {c} _ {T}, \\mathcal {H} _ {\\mathrm {F C}}, t\\right) \\| _ {2} ^ {2} \\right], \\tag {6}\n$$\n", + "text_format": "latex", + "bbox": [ + 522, + 559, + 887, + 578 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "where $\\mathbf{x}_0$ is the target image, $\\mathbf{c}_T$ is the text prompt generated by the large language model, $\\epsilon$ is Gaussian noise, and $t$ is the diffusion timestep sampled uniformly from $1,\\dots ,T$ . The noisy image at timestep $t$ , $\\mathbf{x}_t$ , is obtained by gradually adding noise to $\\mathbf{x}_0$ during the forward diffusion process. The function $\\epsilon_{\\theta}^{\\prime}$ represents the modified denoising model that incorporates the conditioning adapters.", + "bbox": [ + 511, + 588, + 888, + 689 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "4. Experimental Results", + "text_level": 1, + "bbox": [ + 511, + 700, + 707, + 717 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "In this section, we present our experimental results, including quantitative and qualitative comparisons, an ablation study, and an analysis section that visualizes the behavior of the newly introduced cross-attention mechanism in the adapters.", + "bbox": [ + 511, + 726, + 887, + 784 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Implementation Details We evaluate our method on two versions of Stable Diffusion (SD) [23], specifically versions 1.5 and 2.1, to ensure fair comparison with state-of-the-art methods across different backbones. As described in the ablation study, our caching and conditioning mechanism is applied to the middle layer and every second layer of the", + "bbox": [ + 511, + 795, + 887, + 882 + ], + "page_idx": 4 + }, + { + "type": "page_number", + "text": "12484", + "bbox": [ + 480, + 924, + 517, + 934 + ], + "page_idx": 4 + }, + { + "type": "table", + "img_path": "images/3b23b4af1301842b793a74c656f712576ebde622d56f6548b0d7b62a3893b3a3.jpg", + "table_caption": [ + "Table 2. Quantitative results on DreamBooth. DreamCache obtains a better balance between DINO score and CLIP-T compared to all baselines, while also offering a more efficient computational tradeoff (see Table 1)." + ], + "table_footnote": [], + "table_body": "
MethodBackbone#RefDINO (↑)CLIP-I (↑)CLIP-T (↑)
test-time finetuningDreamBooth [24]Imagen3-50.6960.8120.306
DreamBooth [24]SD 1.53-50.6680.8030.305
Textual Inversion [6]SD 1.53-50.5690.7800.255
Custom Diffusion [12]SD 1.53-50.6430.7900.305
BLIP-Diffusion (FT) [13]SD 1.53-50.6700.8050.302
finetuning freeELITE [33]SD 1.510.6210.7710.293
BLIP-Diffusion [13]SD 1.510.5940.7790.300
IP-Adapter [36]SD 1.510.6670.8130.289
Kosmos-G [19]SD 1.510.6940.8470.287
Jedi [37]SD 1.510.6190.7820.304
DreamCache (ours)SD 1.510.7130.8100.298
Re-Imagen [3]Imagen1-30.6000.7400.270
SuTI [4]Imagen1-30.7410.8190.304
Subject-Diffusion [15]SD 2.110.7710.7790.293
BootPig [21]SD 2.130.6740.7970.311
ToffeeNet [39]SD 2.110.7280.8170.306
CAFE [38]SD 2.110.7150.8270.294
DreamCache (ours)SD 2.110.7670.8160.301
", + "bbox": [ + 215, + 143, + 779, + 441 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "decoder. The total number of trainable parameters for DreamCache is 25M. We use the original SD codebase and train the model on $4 \\times 80\\mathrm{GB}$ A100 GPUs for 25k steps with a batch size of 128, using the AdamW optimizer with a learning rate of $10^{-5}$ . Input images are resized to $512 \\times 512$ , with scale, shift, and resize augmentations applied to reference images to enhance model robustness to perturbations. Ablations are conducted on SD 1.5. We generate images with 50 sampling steps, employing classifier-free guidance for image and text conditioning, using a guidance scale of 7.5.", + "bbox": [ + 107, + 465, + 485, + 609 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Evaluation Quantitative evaluations are conducted on the DreamBooth dataset [24], following prior approaches. DreamBooth consists of 30 subjects, each with 25 text prompts. We use a single input image per subject and generate 4 images per subject-prompt combination, resulting in 3,000 generated images. We use pretrained DINO ViT-S/16 and CLIP ViT-B/32 models to calculate the average cosine similarity of global image embeddings between generated and reference images, with metrics denoted as DINO and CLIP-I, respectively. To assess text alignment, we calculate the cosine similarity between embeddings from generated images and text prompts using CLIP's image and text encoders [9], with the corresponding score denoted as CLIP-T.", + "bbox": [ + 107, + 619, + 485, + 808 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "4.1. Zero-Shot Personalization", + "text_level": 1, + "bbox": [ + 109, + 816, + 334, + 830 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "We compare DreamCache against state-of-the-art methods for finetuning-based and zero-shot personalization. Table 2 presents quantitative results, indicating the diffusion back", + "bbox": [ + 107, + 839, + 485, + 882 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "bone and the number of reference images for each method. Our approach achieves competitive or superior performance compared to other computationally-intensive state-of-the-art methods, which are trained on larger datasets and with significantly more parameters. We refer the reader to Table 1 for the data requirements, training time, and parameter count of the various methods. We remark that generally DINO is a preferred metric for image similarity with respect to CLIP-I, as it is more sensitive to the appearance and fine-grained details of the subjects.", + "bbox": [ + 511, + 465, + 888, + 609 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "We also present qualitative comparisons with Kosmos-G [19] and BLIP-Diffusion [13]. We remark that several other methods are not reproducible due to the lack of code, datasets, or trained checkpoints. As seen in Fig. 4, our method excels in subject preservation and textual alignment, producing visually superior results. We also notice that Kosmos-G reports a high CLIP-I score, but after inspecting the generated images, it is clear that the score does not entirely reflect the preservation of the reference subject in generated images. In fact, Kosmos-G presents a high degree of background interference, where the partial replication of the reference background boosts the alignment score. For this reason, we also report foreground-masked metrics on the subjects, like MCLIP-I and MDINO, in the Supp. Mat.", + "bbox": [ + 511, + 609, + 888, + 811 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "4.2. Inference Time Evaluation", + "text_level": 1, + "bbox": [ + 511, + 816, + 743, + 830 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "The computational efficiency of our method is compared to the reference-based method BootPIG [21] and encoder-based approaches such as Kosmos-G [19] and Subject", + "bbox": [ + 511, + 839, + 888, + 882 + ], + "page_idx": 5 + }, + { + "type": "page_number", + "text": "12485", + "bbox": [ + 480, + 924, + 517, + 934 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/be29f76c7d4ede8064e27ff6766020940a0dab6031ab99231271016101e5e9bb.jpg", + "image_caption": [ + "\"A dog\"" + ], + "image_footnote": [], + "bbox": [ + 116, + 122, + 223, + 204 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/82158a728d40bcdcdae325a08b9c4c5fe2f83fac63acfe6e09d69d7d5cade368.jpg", + "image_caption": [ + "\"wearing a Santa hat\"" + ], + "image_footnote": [], + "bbox": [ + 228, + 122, + 336, + 200 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/ebb431d605f744ceba52ff9c87c1fd831101d5964b8fe8a94d2a1213ebfc960e.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 338, + 122, + 444, + 204 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/0364081437e14de23d98385fd1a18defe231a5d0b002fa36707819391bd60757.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 447, + 122, + 553, + 203 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/ced4cdce5491e89288ba7db0672f261bef4d985f78cd71337119c205cc195a7a.jpg", + "image_caption": [ + "\"with the Eiffel Tower in the background\"" + ], + "image_footnote": [], + "bbox": [ + 555, + 122, + 661, + 204 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/a5b006650f369bfea8d691cc7ef75697044ec7a13823eadf9ce38f8b55a71f67.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 665, + 122, + 772, + 204 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/e16785f60eae1f9af2092ee8fa023a215bc9e4ea4bca676702e36cafae9ed04f.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 774, + 122, + 880, + 204 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/2507e224b7535a30b0c5d314074d3ebb7113126ee2677b8a2f15bad3d167e2fe.jpg", + "image_caption": [ + "\"A can\"" + ], + "image_footnote": [], + "bbox": [ + 116, + 220, + 223, + 303 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/96dc5471e153bef8c50cff411a269645b40cb182a8103679ba5204127d1f51ca.jpg", + "image_caption": [ + "\"floating on top of water\"" + ], + "image_footnote": [], + "bbox": [ + 228, + 220, + 336, + 304 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/a9c98199e7c76e5c44e19e87094fb5d7a876d874100fa4a0b1ea4f51d7a70810.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 339, + 220, + 444, + 304 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/eabc7d6929f8ca80374bd68054aaae9c72a1218e62a43e5faa83a80aa9c0d98a.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 447, + 220, + 553, + 304 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/8c97f3fa5025ee58cc1860117f3a538453dbfbfe57950ba3c1611db97a727ea9.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 555, + 220, + 663, + 304 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/5894285b043526c36b61b0ed750d470915bdddb7307918789218439515c7a83c.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 665, + 220, + 772, + 304 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/6a02427b8d3f7feb761e3e47665eaf364a38b8df2c53f230257eab10a96c8f75.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 774, + 220, + 880, + 304 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/96160c2ac6ad20efef76172f424b0d6b532a8e97c03f7b2a71539b3518f85516.jpg", + "image_caption": [ + "\"A toy\"", + "Reference", + "Figure 4. Visual comparison. Personalized generations on sample concepts. DreamCache preserves reference concept appearance and does not suffer from background interference. BLIP-D [13] and Kosmos-G [19] cannot faithfully preserve visual details from the reference." + ], + "image_footnote": [], + "bbox": [ + 116, + 320, + 223, + 402 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/4a3ed0a138fb260e32faf41fd198449b4fc83a1804119156e3cb4cde4e750018.jpg", + "image_caption": [ + "BLIP-D" + ], + "image_footnote": [], + "bbox": [ + 228, + 320, + 336, + 402 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/dc4e18dcb953af79fe1e5ddb1f9a4844b854051d85e344a7410a2e28448e84f5.jpg", + "image_caption": [ + "\"on the beach\"", + "Kosmos-G" + ], + "image_footnote": [], + "bbox": [ + 339, + 320, + 444, + 402 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/1a327d304d745d5ced93dd3492b05fb1a99ea13a8dc777607d2957c6a641507e.jpg", + "image_caption": [ + "DreamCache" + ], + "image_footnote": [], + "bbox": [ + 447, + 320, + 553, + 402 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/2dc8043099673fd79307756f9fcbe82bbfc3c1fb2d46fca4bdec5a801817a14b.jpg", + "image_caption": [ + "BLIP-D" + ], + "image_footnote": [], + "bbox": [ + 555, + 320, + 663, + 402 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/3208ce0f09f90f0927ec307c2ddb01da153c1ffd72eb7ef6af8cf8a16870360a.jpg", + "image_caption": [ + "Kosmos-G" + ], + "image_footnote": [], + "bbox": [ + 665, + 320, + 772, + 402 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/9c0c65754d7fddb6c2b4a3c01bb56a035ab730190c3e7132aa3155679e98f25f.jpg", + "image_caption": [ + "DreamCache" + ], + "image_footnote": [], + "bbox": [ + 774, + 320, + 880, + 402 + ], + "page_idx": 6 + }, + { + "type": "table", + "img_path": "images/ae89d8b9290f227f45398655acd7a9ee079e06c60bec37d662686a8e128f2efc.jpg", + "table_caption": [ + "Table 3. Computational comparison.*: time to generate an image with 100 timesteps, evaluated on a single NVIDIA A100 GPU." + ], + "table_footnote": [], + "table_body": "
MethodInference Time*Extra Params Size
ELITE [33]6.24 s914 MB
BLIP-Diffusion [13]3.92 s760 MB
BootPig [21]7.55 s1900 MB
DreamCache (ours)3.88 s42 MB
", + "bbox": [ + 114, + 518, + 478, + 608 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Diffusion [15]. Table 3 provides a detailed comparison of inference time, accounting for both personalization time (e.g., the time to generate the cache for DreamCache) and the time to sample the personalized image. We also report the increase in model size, i.e., the storage (in FP16 precision) required for the extra parameters to allow for personalization, showing that DreamCache is one order of magnitude more compact than the state of the art. Overall, DreamCache offers a lightweight solution that achieves state-of-the-art performance with faster inference and reduced computational overhead.", + "bbox": [ + 107, + 633, + 483, + 790 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "4.3. Ablation Studies", + "text_level": 1, + "bbox": [ + 109, + 801, + 264, + 815 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "We validate our design choices through a series of studies, examining different conditioning mechanisms, evaluating our feature caching approach, and analyzing the impact of synthetic dataset scaling.", + "bbox": [ + 107, + 824, + 483, + 882 + ], + "page_idx": 6 + }, + { + "type": "table", + "img_path": "images/e9b619f27d9a9a4b577709ddc1fe8ffd455fdbd18b54fad041a6fce5c076252d.jpg", + "table_caption": [ + "Table 4. Reference feature integration. DreamCache uses the best tradeoff between accuracy and complexity." + ], + "table_footnote": [], + "table_body": "
MethodCLIP-I (↑)CLIP-T (↑)Params
Textual Sum [36]0.7880,28219M
Spatial Sum0.8120.29316M
Decoupled Blocks [8]0.8080.30061M
Spatial Concat (ours)0.8100.29825M
", + "bbox": [ + 517, + 518, + 883, + 609 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Reference Feature Integration We compare various conditioning strategies to integrate reference features in Table 4. Our spatial cross-attention block with concatenation between the output of self- and cross-attention (\"Spatial Concat\") was evaluated against different alternatives, including IP-Adapter's conditioning mechanism [36] (\"Textual Sum\"), which sums the decoupled cross-attention output with that from textual cross-attention. We also tested a variant (\"Spatial Sum\") where self- and cross-attention conditioning outputs are summed. Additionally, we also assessed an alternative conditioning procedure inspired by ViCo [8] (\"Decoupled Blocks\"), involving independent and interleaved cross-attention blocks. Results in Table 4 indicate that the proposed \"Spatial Concat\" offers the best balance of text alignment and parameter efficiency.", + "bbox": [ + 511, + 635, + 888, + 851 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "We further explored optimal conditioning insertion within the U-Net backbone in Table 5, determining that applying", + "bbox": [ + 511, + 852, + 887, + 882 + ], + "page_idx": 6 + }, + { + "type": "page_number", + "text": "12486", + "bbox": [ + 480, + 924, + 517, + 934 + ], + "page_idx": 6 + }, + { + "type": "table", + "img_path": "images/4e46f7f99d2d6af168c097f2393a24d13ad334df9edd5e7dc6f91ffc68347f8d.jpg", + "table_caption": [ + "Table 5. Cache positioning in the U-Net backbone offers a further tradeoff between accuracy and complexity." + ], + "table_footnote": [], + "table_body": "
EncoderMiddleDecoderCLIP-I (↑)CLIP-T (↑)Params
✓XX0.7210.30311M
✓✓X0.7490.30619M
X✓X0.7160.3028M
XX✓0.7990.29617M
X✓✓0.8100.29825M
✓✓✓0.8130.29736M
", + "bbox": [ + 109, + 145, + 483, + 263 + ], + "page_idx": 7 + }, + { + "type": "table", + "img_path": "images/d27e0ac50d4246f5c89f22880691d6fd930631df4aee284788b4ca86b97ceca0.jpg", + "table_caption": [ + "Table 6. Caching with text is not influential and adds complexity." + ], + "table_footnote": [], + "table_body": "
Text-FreeCLIP-I (↑)CLIP-T (↑)
×0.8110.295
✓0.8100.298
", + "bbox": [ + 181, + 301, + 411, + 362 + ], + "page_idx": 7 + }, + { + "type": "table", + "img_path": "images/386934df26c286e8e9a670e75a33b5f999ff89eadb67720d228bc5d2e0e2d967.jpg", + "table_caption": [ + "Table 7. Dataset impact for both synthetic and real data." + ], + "table_footnote": [], + "table_body": "
DatasetCLIP-I (↑)CLIP-T (↑)
Synthetic-50K0.7810.304
Synthetic-200K0.7970,301
Synthetic-400K0.8100.298
LAION-5M0.8140.242
", + "bbox": [ + 163, + 400, + 431, + 489 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "conditioning (and therefore feature caching) at the middle layer and every second layer of the decoder achieved the best tradeoff between performance and parameter count.", + "bbox": [ + 107, + 505, + 483, + 549 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Text Input for Cached Features Our feature caching procedure is designed to be text-free, leveraging the classifier-free guidance used during pretraining where captions were occasionally omitted. We compare this approach with a version including textual inputs during caching (e.g., 'A photo of...'). Table 6 shows that adding text conditioning slightly reduces text alignment while increasing complexity and potentially introducing noise in cases of inaccurate captions.", + "bbox": [ + 107, + 556, + 485, + 672 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Dataset Impact We demonstrate the importance of our synthetic dataset to train the conditioning adapters and the effect of scaling its size. For this purpose, we created synthetic datasets according to the procedure in Sec. 3.3 of sizes 50K, 200K, and 400K samples. We also tested the real-world 5M samples from the LAION [26] dataset, which, lacking target-caption-reference triplets, required reuse of target images as reference images too. Table 7 shows that increasing dataset size improves image alignment, though slightly reduces textual alignment. Notably, LAION improves image alignment but struggles with textual alignment. This highlights the importance of triplet data (target image, reference image, and caption) for effective zero-shot personalization, ensuring both subject preservation and textual editability.", + "bbox": [ + 107, + 680, + 485, + 883 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/8f6e6e1d1e4b6a2758dcd4909a5593d41acc79da2a621ce9c161ce2cef600754.jpg", + "image_caption": [ + "Figure 5. Visualization of reference image impact. Cross-attention maps between cached reference features and features of the image under generation. Left: attention map at layers at $16 \\times 16$ resolution (left reference, right generated). Right: $32 \\times 32$ . Attention values are highly localized in the region of interest." + ], + "image_footnote": [], + "bbox": [ + 519, + 104, + 883, + 256 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "4.4. Visualizing Reference Impact", + "text_level": 1, + "bbox": [ + 511, + 357, + 764, + 373 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Finally, we analyze how the cross-attention mechanism in DreamCache impacts image generation by visualizing cached reference feature influence. Attention map visualizations at different resolutions are provided in Fig. 5. Specifically, attention maps between the query from the current generation and the key derived from reference features reveal a highly localized focus on the subject, without interference from background elements. This mechanism models correspondences effectively, integrating reference information into the generated image.", + "bbox": [ + 511, + 380, + 888, + 523 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "5. Discussion and Conclusions", + "text_level": 1, + "bbox": [ + 511, + 530, + 756, + 544 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "In this paper, we proposed DreamCache, a novel approach to personalized text-to-image generation that uses feature caching to overcome the limitations of existing methods. By caching reference features from a small subset of layers of the U-Net only once, our method significantly reduces both computational and memory demands, enabling efficient, real-time personalized image generation. Unlike previous approaches, DreamCache avoids the need for costly finetuning, external image encoders, or parallel reference processing, making it lightweight and suitable for plug-and-play deployment. Our experiments demonstrate that DreamCache achieves state-of-the-art zero-shot personalization with only 25M additional parameters and a fast training process. While DreamCache is a promising direction towards more efficient personalized generation, it has some limitations. Although effective for single-subject personalization, our approach may require adaptation for complex multi-subject generation where feature interference can occur. Additionally, certain edge cases, such as highly abstract or stylistic images, may challenge the caching mechanism's capacity to accurately preserve subject details. To address these challenges, future work may explore adaptive caching techniques or multi-reference feature integration.", + "bbox": [ + 509, + 550, + 888, + 882 + ], + "page_idx": 7 + }, + { + "type": "page_number", + "text": "12487", + "bbox": [ + 480, + 922, + 517, + 934 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "References", + "text_level": 1, + "bbox": [ + 109, + 108, + 200, + 123 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[1] Moab Arar, Rinon Gal, Yuval Atzmon, Gal Chechik, Daniel Cohen-Or, Ariel Shamir, and Amit H. Bermano. Domain-agnostic tuning-encoder for fast personalization of text-to-image models. In SIGGRAPH Asia 2023 Conference Papers, pages 1–10, 2023. 1, 3", + "[2] Mathilde Caron, Hugo Touvron, Ishan Misra, Hervé Jégou, Julien Mairal, Piotr Bojanowski, and Armand Joulin. Emerging properties in self-supervised vision transformers. In Proceedings of the IEEE/CVF international conference on computer vision, pages 9650-9660, 2021. 1", + "[3] Wenhu Chen, Hexiang Hu, Chitwan Sahara, and William W Cohen. Re-imagen: Retrieval-augmented text-to-image generator. In The Eleventh International Conference on Learning Representations, 2022. 6", + "[4] Wenhu Chen, Hexiang Hu, Yandong Li, Nataniel Rui, Xuhui Jia, Ming-Wei Chang, and William W Cohen. Subject-driven text-to-image generation via apprenticeship learning. arXiv preprint arXiv:2304.00186, 2023. 2, 3, 6", + "[5] Abhimanyu Dubey, Abhinav Jauhri, Abhinav Pandey, Abhishek Kadian, Ahmad Al-Dahle, Aiesha Letman, Akhil Mathur, Alan Schelten, Amy Yang, Angela Fan, et al. The llama 3 herd of models. arXiv preprint arXiv:2407.21783, 2024. 5", + "[6] Rinon Gal, Yuval Alaluf, Yuval Atzmon, Or Patashnik, Amit Haim Bermano, Gal Chechik, and Daniel Cohen-or. An image is worth one word: Personalizing text-to-image generation using textual inversion. In The Eleventh International Conference on Learning Representations, 2022. 1, 2, 3, 6", + "[7] Rinon Gal, Moab Arar, Yuval Atzmon, Amit H Bermano, Gal Chechik, and Daniel Cohen-Or. Encoder-based domain tuning for fast personalization of text-to-image models. ACM Transactions on Graphics (TOG), 42(4):1-13, 2023. 1, 3", + "[8] Shaozhe Hao, Kai Han, Shihao Zhao, and Kwan-Yee K Wong. Vico: Plug-and-play visual condition for personalized text-to-image generation. arXiv preprint arXiv:2306.00971, 2023. 1, 3, 7", + "[9] Jack Hessel, Ari Holtzman, Maxwell Forbes, Ronan Le Bras, and Yejin Choi. Clipscore: A reference-free evaluation metric for image captioning. In Proceedings of the 2021 Conference on Empirical Methods in Natural Language Processing, pages 7514-7528, 2021. 6", + "[10] Jonathan Ho, Ajay Jain, and Pieter Abbeel. Denoising diffusion probabilistic models. Advances in neural information processing systems, 33:6840-6851, 2020. 1", + "[11] Alexander Kirillov, Eric Mintun, Nikhila Ravi, Hanzi Mao, Chloe Rolland, Laura Gustafson, Tete Xiao, Spencer Whitehead, Alexander C Berg, Wan-Yen Lo, et al. Segment anything. arXiv preprint arXiv:2304.02643, 2023. 5", + "[12] Nupur Kumari, Bingliang Zhang, Richard Zhang, Eli Shechtman, and Jun-Yan Zhu. Multi-concept customization of text-to-image diffusion. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 1931-1941, 2023. 1, 2, 6", + "[13] Dongxu Li, Junnan Li, and Steven CH Hoi. Blip-diffusion: Pre-trained subject representation for control" + ], + "bbox": [ + 109, + 132, + 485, + 881 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "lable text-to-image generation and editing. arXiv preprint arXiv:2305.14720, 2023. 1, 2, 3, 6, 7", + "[14] Shilong Liu, Zhaoyang Zeng, Tianhe Ren, Feng Li, Hao Zhang, Jie Yang, Chunyuan Li, Jianwei Yang, Hang Su, Jun Zhu, et al. Grounding dino: Marrying dino with grounded pre-training for open-set object detection. arXiv preprint arXiv:2303.05499, 2023. 5", + "[15] Jian Ma, Junhao Liang, Chen Chen, and Haonan Lu. Subject-diffusion: Open domain personalized text-to-image generation without test-time fine-tuning. arXiv preprint arXiv:2307.11410, 2023. 1, 2, 3, 5, 6, 7", + "[16] Xinyin Ma, Gongfan Fang, Michael Bi Mi, and Xinchao Wang. Learning-to-cache: Accelerating diffusion transformer via layer caching. arXiv preprint arXiv:2406.01733, 2024. 4", + "[17] Xinyin Ma, Gongfan Fang, and Xinchao Wang. Deepcache: Accelerating diffusion models for free. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 15762-15772, 2024. 3", + "[18] Boris Oreshkin, Pau Rodríguez López, and Alexandre Lacoste. Tadam: Task dependent adaptive metric for improved few-shot learning. Advances in neural information processing systems, 31, 2018. 4", + "[19] Xichen Pan, Li Dong, Shaohan Huang, Zhiliang Peng, Wenhu Chen, and Furu Wei. Kosmos-g: Generating images in context with multimodal large language models. arXiv preprint arXiv:2310.02992, 2023. 1, 2, 3, 6, 7", + "[20] Maitreya Patel, Sangmin Jung, Chitta Baral, and Yezhou Yang. λ-eclipse: Multi-concept personalized text-to-image diffusion models by leveraging clip latent space. arXiv preprint arXiv:2402.05195, 2024. 1, 3", + "[21] Senthil Purushwalkam, Akash Gokul, Shafiq Joty, and Nikhil Naik. Bootpig: Bootstrapping zero-shot personalized image generation capabilities in pretrained diffusion models. arXiv preprint arXiv:2401.13974, 2024. 1, 2, 3, 5, 6, 7", + "[22] Alec Radford, Jong Wook Kim, Chris Hallacy, Aditya Ramesh, Gabriel Goh, Sandhini Agarwal, Girish Sastry, Amanda Askell, Pamela Mishkin, Jack Clark, et al. Learning transferable visual models from natural language supervision. In International conference on machine learning, pages 8748-8763. PMLR, 2021. 1", + "[23] Robin Rombach, Andreas Blattmann, Dominik Lorenz, Patrick Esser, and Björn Ommer. High-resolution image synthesis with latent diffusion models. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 10684-10695, 2022. 1, 5", + "[24] Nataniel Ruiz, Yuanzhen Li, Varun Jampani, Yael Pritch, Michael Rubinstein, and Kfir Aberman. Dreambooth: Fine tuning text-to-image diffusion models for subject-driven generation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 22500-22510, 2023. 1, 2, 6", + "[25] Chitwan Sahara, William Chan, Saurabh Saxena, Lala Li, Jay Whang, Emily L Denton, Kamyar Ghasemipour, Raphael Gontijo Lopes, Burcu Karagol Ayan, Tim Salimans, et al. Photorealistic text-to-image diffusion models with deep language understanding. Advances in Neural Information Processing Systems, 35:36479-36494, 2022. 1" + ], + "bbox": [ + 514, + 111, + 885, + 881 + ], + "page_idx": 8 + }, + { + "type": "page_number", + "text": "12488", + "bbox": [ + 480, + 924, + 517, + 934 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[26] Christoph Schuhmann, Romain Beaumont, Richard Vencu, Cade Gordon, Ross Wightman, Mehdi Cherti, Theo Coombes, Aarush Katta, Clayton Mullis, Mitchell Wortsman, et al. Laion-5b: An open large-scale dataset for training next generation image-text models. Advances in Neural Information Processing Systems, 35:25278–25294, 2022. 8", + "[27] Jake Snell, Kevin Swersky, and Richard Zemel. Prototypical networks for few-shot learning. Advances in neural information processing systems, 30, 2017. 4", + "[28] Jascha Sohl-Dickstein, Eric Weiss, Niru Maheswaranathan, and Surya Ganguli. Deep unsupervised learning using nonequilibrium thermodynamics. In International conference on machine learning, pages 2256-2265. PMLR, 2015. 1", + "[29] Yang Song and Stefano Ermon. Generative modeling by estimating gradients of the data distribution. Advances in neural information processing systems, 32, 2019. 5", + "[30] Yoad Tewel, Rinon Gal, Gal Chechik, and Yuval Atzmon. Key-locked rank one editing for text-to-image personalization. In ACM SIGGRAPH 2023 Conference Proceedings, pages 1-11, 2023. 1, 3", + "[31] Eleni Triantafillou, Hugo Larochelle, Jake Snell, Josh Tenenbaum, Kevin Jordan Swersky, Mengye Ren, Richard Zemel, and Sachin Ravi. Meta-learning for semi-supervised few-shot classification. In International Conference on Learning Representations, 2018. 4", + "[32] Andrey Voynov, Qinghao Chu, Daniel Cohen-Or, and Kfir Aberman. p+: Extended textual conditioning in text-to-image generation. arXiv preprint arXiv:2303.09522, 2023. 1, 3", + "[33] Yuxiang Wei, Yabo Zhang, Zhilong Ji, Jinfeng Bai, Lei Zhang, and Wangmeng Zuo. Elite: Encoding visual concepts into textual embeddings for customized text-to-image generation. 2023 IEEE/CVF International Conference on Computer Vision (ICCV), pages 15897-15907, 2023. 1, 2, 3, 5, 6, 7", + "[34] Felix Wimbauer, Bichen Wu, Edgar Schoenfeld, Xiaoliang Dai, Ji Hou, Zijian He, Artsiom Sanakoyeu, Peizhao Zhang, Sam Tsai, Jonas Kohler, et al. Cache me if you can: Accelerating diffusion models through block caching. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 6211-6220, 2024. 3", + "[35] Jianan Yang, Haobo Wang, Yanming Zhang, Ruixuan Xiao, Sai Wu, Gang Chen, and Junbo Zhao. Controllable textual inversion for personalized text-to-image generation. arXiv preprint arXiv:2304.05265, 2023. 1, 3", + "[36] Hu Ye, Jun Zhang, Sibo Liu, Xiao Han, and Wei Yang. Ip-adapter: Text compatible image prompt adapter for text-to-image diffusion models. arXiv preprint arXiv:2308.06721, 2023. 1, 2, 3, 6, 7", + "[37] Yu Zeng, Vishal M Patel, Haochen Wang, Xun Huang, Ting-Chun Wang, Ming-Yu Liu, and Yogesh Balaji. Jedi: Joint-image diffusion models for finetuning-free personalized text-to-image generation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 6786-6795, 2024. 1, 2, 3, 4, 6", + "[38] Yufan Zhou, Ruiyi Zhang, Jiumiang Gu, and Tong Sun. Customization assistant for text-to-image generation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 9182-9191, 2024. 2, 3, 6" + ], + "bbox": [ + 109, + 109, + 485, + 881 + ], + "page_idx": 9 + }, + { + "type": "ref_text", + "text": "[39] Yufan Zhou, Ruiyi Zhang, Kaizhi Zheng, Nanxuan Zhao, Jiumiang Gu, Zichao Wang, Xin Eric Wang, and Tong Sun. Toffee: Efficient million-scale dataset construction for subject-driven text-to-image generation. arXiv preprint arXiv:2406.09305, 2024.1,2,3,6", + "bbox": [ + 514, + 109, + 888, + 176 + ], + "page_idx": 9 + }, + { + "type": "page_number", + "text": "12489", + "bbox": [ + 480, + 924, + 517, + 934 + ], + "page_idx": 9 + } +] \ No newline at end of file diff --git a/2025/DreamCache_ Finetuning-Free Lightweight Personalized Image Generation via Feature Caching/484a2287-cf26-4bf8-9e41-034501dc6452_model.json b/2025/DreamCache_ Finetuning-Free Lightweight Personalized Image Generation via Feature Caching/484a2287-cf26-4bf8-9e41-034501dc6452_model.json new file mode 100644 index 0000000000000000000000000000000000000000..65f0911f4694dbf44dd4851d504fdad00841dadf --- /dev/null +++ b/2025/DreamCache_ Finetuning-Free Lightweight Personalized Image Generation via Feature Caching/484a2287-cf26-4bf8-9e41-034501dc6452_model.json @@ -0,0 +1,2387 @@ +[ + [ + { + "type": "header", + "bbox": [ + 0.117, + 0.017, + 0.191, + 0.055 + ], + "angle": 0, + "content": "CVF" + }, + { + "type": "header", + "bbox": [ + 0.244, + 0.015, + 0.8, + 0.058 + ], + "angle": 0, + "content": "This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore." + }, + { + "type": "title", + "bbox": [ + 0.247, + 0.148, + 0.753, + 0.193 + ], + "angle": 0, + "content": "DreamCache: Finetuning-Free Lightweight Personalized Image Generation via Feature Caching" + }, + { + "type": "text", + "bbox": [ + 0.146, + 0.217, + 0.851, + 0.253 + ], + "angle": 0, + "content": "Emanuele Aiello Umberto Michieli Diego Valsesia Mete Ozay Enrico Magli Politecnico di Torino Samsung R&D Institute UK" + }, + { + "type": "title", + "bbox": [ + 0.26, + 0.284, + 0.334, + 0.3 + ], + "angle": 0, + "content": "Abstract" + }, + { + "type": "text", + "bbox": [ + 0.108, + 0.317, + 0.486, + 0.547 + ], + "angle": 0, + "content": "Personalized image generation requires text-to-image generative models that capture the core features of a reference subject to allow for controlled generation across different contexts. Existing methods face challenges due to complex training requirements, high inference costs, limited flexibility, or a combination of these issues. In this paper, we introduce DreamCache, a scalable approach for efficient and high-quality personalized image generation. By caching a small number of reference image features from a subset of layers and a single timestep of the pretrained diffusion denoiser, DreamCache enables dynamic modulation of the generated image features through lightweight, trained conditioning adapters. DreamCache achieves state-of-the-art image and text alignment, utilizing an order of magnitude fewer extra parameters, and is both more computationally effective and versatile than existing models." + }, + { + "type": "title", + "bbox": [ + 0.11, + 0.567, + 0.236, + 0.582 + ], + "angle": 0, + "content": "1. Introduction" + }, + { + "type": "text", + "bbox": [ + 0.108, + 0.587, + 0.485, + 0.673 + ], + "angle": 0, + "content": "Recent advancements in text-to-image generation, fueled by the development of diffusion models [10, 28], have enabled high-quality and diverse image generation from textual descriptions. Diffusion models [23, 25] gradually transform random noise into images through a sequence of denoising steps, conditioned on the input text prompt." + }, + { + "type": "text", + "bbox": [ + 0.108, + 0.674, + 0.486, + 0.848 + ], + "angle": 0, + "content": "An active area of research is personalizing these models, enabling the generation of novel images of a reference subject in various contexts, while maintaining flexibility for text-based editing. Early personalization techniques [1, 6, 8, 24, 30, 32, 35, 35], such as the seminal Dream-Booth [24] relied on fine-tuning (FT) the generative model for each reference subject. However, these approaches are often impractical for many use cases due to costly test-time FT, which can take several minutes per subject. To address this, FT-free (i.e., zero-shot) personalized image generation methods have emerged to eliminate test-time optimization. These FT-free approaches can be broadly categorized into" + }, + { + "type": "image", + "bbox": [ + 0.582, + 0.286, + 0.819, + 0.468 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.512, + 0.481, + 0.89, + 0.522 + ], + "angle": 0, + "content": "Figure 1. DreamCache is a finetuning-free personalized image generation method that achieves an optimal balance between subject fidelity, memory efficiency, and adherence to text prompts." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.548, + 0.887, + 0.576 + ], + "angle": 0, + "content": "two families: encoder-based methods and reference-based methods, each with distinct drawbacks." + }, + { + "type": "text", + "bbox": [ + 0.511, + 0.578, + 0.889, + 0.664 + ], + "angle": 0, + "content": "Encoder-based methods [7, 15, 20, 33, 36] utilize dedicated image encoders, such as CLIP [22] or DINO [2], to extract relevant features from reference images. While these encoders can produce high-quality results, they are often large, require extensive training to align text and image features, and reduce the model's flexibility [13, 15, 19, 36]." + }, + { + "type": "text", + "bbox": [ + 0.511, + 0.666, + 0.889, + 0.795 + ], + "angle": 0, + "content": "In contrast, reference-based methods [21, 37] condition the diffusion model directly on reference features drawn from the U-Net denoiser, integrating these features at each denoising step. While effective, these methods require feature extraction at every generation step, leading to higher computational costs and memory demands. Additionally, they often require an input textual caption for conditioning, which introduces variability and can decrease output precision." + }, + { + "type": "text", + "bbox": [ + 0.511, + 0.796, + 0.889, + 0.883 + ], + "angle": 0, + "content": "Some recent works have proposed to finetune the U-Net backbone itself [21, 37, 39]. However, this hinders the model's ability to switch between personalized and nonpersonalized tasks and risks inducing the \"language drift\" phenomenon, where the personalization training degrades the model's linguistic comprehension [12, 24]." + }, + { + "type": "page_footnote", + "bbox": [ + 0.11, + 0.858, + 0.486, + 0.882 + ], + "angle": 0, + "content": "\\(^{1}\\)Code, model, and synthetic dataset can be found at https://github.com/Emanuele97x/DreamCache" + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.924, + 0.519, + 0.935 + ], + "angle": 0, + "content": "12480" + } + ], + [ + { + "type": "table_caption", + "bbox": [ + 0.11, + 0.109, + 0.888, + 0.135 + ], + "angle": 0, + "content": "Table 1. Methods overview. Our DreamCache achieves state-of-the-art generation quality at reduced computational costs. *: value refers to the personalization stage for each personal subject." + }, + { + "type": "table", + "bbox": [ + 0.124, + 0.146, + 0.875, + 0.384 + ], + "angle": 0, + "content": "
MethodFT-freeEnc-freePlug&PlayRef-UNet-freeExtra ParamsTrain Params# DatasetTrain Time
Textual Inversion [6]X✓✓✓768*768*3-5*50 min*
DreamBooth [24]X✓X✓-0.9B*3-5*10 min*
Custom Diffusion [12]X✓X✓-57M*3-5*10 min*
ELITE [33]✓XX✓457M77M125K14 days
BLIP-Diffusion [13]✓X✓✓380M1.5B129M96 days
IP-Adapter [36]✓X✓✓402M22M10M28 days
Kosmos-G [19]✓X✓✓1.6B1.6B9M-
JeDi [37]✓✓XX-0.9B3M48 days
SuTI [4]✓XX✓400M2.5B500K-
Subject-Diffusion [15]✓X✓X700M700M76M-
BootPig [21]✓✓XX0.95B0.95B200K18 hours
ToffeeNet [39]✓XX✓632M0.9B5M-
CAFE [38]✓XX✓14B1B355K-
DreamCache (ours)✓✓✓✓25M25M400K40 hours
" + }, + { + "type": "text", + "bbox": [ + 0.113, + 0.409, + 0.485, + 0.653 + ], + "angle": 0, + "content": "In this work, we propose DreamCache, a novel finetuning-free approach to personalized image generation, that overcomes the limitations of existing methods (see Fig. 1) by using a feature caching mechanism that enables text-free encoding and efficient conditioning during personalization. Specifically, we first create a synthetic dataset [21] containing triplets of captions, target images, and reference subjects to capture subjects in various contexts. Next, we pretrain lightweight attention-based conditioning adapters to inject subject-specific features into the image generation process. During personalization, the reference image is processed through the pretrained denoiser of the base diffusion model without text conditioning, thus eliminating the need for user-generated captions, while caching features from a small subset of layers at a single timestep. For personalized sampling, these cached features are injected into the denoiser through the pretrained conditioning adapters." + }, + { + "type": "text", + "bbox": [ + 0.113, + 0.654, + 0.485, + 0.87 + ], + "angle": 0, + "content": "Table 1 summarizes the key properties of existing methods and illustrates how DreamCache fits within the current landscape; further details are explored in Sec. 2. As an encoder-free approach, DreamCache introduces only a small number of additional parameters, making it significantly lightweight and suitable for deployment on resource-constrained devices. For example, methods like [33] and [13] introduce 380M parameters due to their reliance on CLIP encoders, whereas DreamCache requires only 25M additional parameters. Moreover, caching features from a few selected U-Net layers at a single preprocessing timestep bypasses the need for full U-Net reference processing during generation, leading to substantial computational and memory savings that enable real-time, high-quality personalized generation. Another key advantage of DreamCache is its plug-and-play" + }, + { + "type": "text", + "bbox": [ + 0.516, + 0.409, + 0.887, + 0.48 + ], + "angle": 0, + "content": "nature, allowing concurrent generation of personalized and non-personalized content without altering original U-Net weights, thus preserving the integrity of the original model and enabling a wider range of deployment scenarios, especially on mobile platforms." + }, + { + "type": "text", + "bbox": [ + 0.516, + 0.482, + 0.887, + 0.524 + ], + "angle": 0, + "content": "In summary, DreamCache represents a significant step toward practical, and scalable personalized image generation, with the following contributions:" + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.528, + 0.886, + 0.569 + ], + "angle": 0, + "content": "- We propose a feature caching approach that creates multi-resolution representations of the reference image in a caption-free and efficient manner." + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.571, + 0.886, + 0.627 + ], + "angle": 0, + "content": "- We design an attention-based conditioning mechanism that leverages the cached features for personalized image generation, achieving computational- and memory- efficient personalized sampling." + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.629, + 0.886, + 0.67 + ], + "angle": 0, + "content": "- Our approach achieves state-of-the-art quality in personalized image generation at substantially lower computational and data costs compared to existing methods." + }, + { + "type": "list", + "bbox": [ + 0.513, + 0.528, + 0.886, + 0.67 + ], + "angle": 0, + "content": null + }, + { + "type": "title", + "bbox": [ + 0.516, + 0.687, + 0.785, + 0.702 + ], + "angle": 0, + "content": "2. Background and Related Work" + }, + { + "type": "text", + "bbox": [ + 0.515, + 0.712, + 0.887, + 0.811 + ], + "angle": 0, + "content": "Personalized image generation aims at generating images containing a specific subject. This task has been widely studied, resulting in two main approaches: fine-tuning methods, which require test-time finetuning on multiple subject reference images, and finetuning-free (zero-shot) methods, which learn a generalizable conditioning mechanism to generate reference subjects without the need for further optimization." + }, + { + "type": "text", + "bbox": [ + 0.516, + 0.826, + 0.887, + 0.881 + ], + "angle": 0, + "content": "Finetuning-based Personalization DreamBooth [24] finetunes the entire U-Net with reference images while introducing a regularization loss to mitigate overfitting. On the other hand, Custom Diffusion [12] only finetunes the \\(K\\) and" + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.925, + 0.517, + 0.935 + ], + "angle": 0, + "content": "12481" + } + ], + [ + { + "type": "image_caption", + "bbox": [ + 0.148, + 0.109, + 0.234, + 0.122 + ], + "angle": 0, + "content": "\"A dragon...\"" + }, + { + "type": "image", + "bbox": [ + 0.116, + 0.123, + 0.265, + 0.238 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.161, + 0.241, + 0.221, + 0.253 + ], + "angle": 0, + "content": "\"A cat...\"" + }, + { + "type": "image_caption", + "bbox": [ + 0.284, + 0.109, + 0.4, + 0.123 + ], + "angle": 0, + "content": "\"as street graffiti\"" + }, + { + "type": "image", + "bbox": [ + 0.267, + 0.125, + 0.416, + 0.238 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.283, + 0.241, + 0.4, + 0.255 + ], + "angle": 0, + "content": "\"in Ukiyo-e style\"" + }, + { + "type": "image_caption", + "bbox": [ + 0.434, + 0.109, + 0.554, + 0.123 + ], + "angle": 0, + "content": "\"playing with fire\"" + }, + { + "type": "image", + "bbox": [ + 0.419, + 0.123, + 0.568, + 0.238 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.419, + 0.241, + 0.568, + 0.255 + ], + "angle": 0, + "content": "\"with a rainbow scarf\"" + }, + { + "type": "image_caption", + "bbox": [ + 0.599, + 0.109, + 0.692, + 0.123 + ], + "angle": 0, + "content": "\"as a plushie\"" + }, + { + "type": "image", + "bbox": [ + 0.571, + 0.124, + 0.72, + 0.238 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.577, + 0.241, + 0.714, + 0.255 + ], + "angle": 0, + "content": "\"Van Gogh painting\"" + }, + { + "type": "image_caption", + "bbox": [ + 0.728, + 0.109, + 0.876, + 0.122 + ], + "angle": 0, + "content": "\"working as a barista\"" + }, + { + "type": "image", + "bbox": [ + 0.727, + 0.123, + 0.877, + 0.238 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.725, + 0.241, + 0.882, + 0.255 + ], + "angle": 0, + "content": "\"wearing a diploma hat\"" + }, + { + "type": "image", + "bbox": [ + 0.115, + 0.255, + 0.264, + 0.369 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.265, + 0.256, + 0.416, + 0.368 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.418, + 0.255, + 0.569, + 0.37 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.571, + 0.255, + 0.72, + 0.37 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.727, + 0.255, + 0.877, + 0.37 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.109, + 0.385, + 0.888, + 0.412 + ], + "angle": 0, + "content": "Figure 2. Personalized generations by DreamCache. The first column contains reference images. The generated images correspond to the text prompts above each column." + }, + { + "type": "text", + "bbox": [ + 0.109, + 0.437, + 0.486, + 0.567 + ], + "angle": 0, + "content": "\\(V\\) projections for the cross-attention blocks of the U-Net. Text-based personalization methods optimize single (like Textual Inversion [6]) or multiple (like in \\(\\mathrm{P + }\\) [32]) input token embeddings. Later methods [1, 8, 30, 35] build on these, with innovations like Perfusion [30] using dynamic rank-1 updates to prevent overfitting while keeping encodings lightweight. However, all finetuning-based methods are computationally intensive, often requiring minutes of finetuning per reference subject at test time." + }, + { + "type": "text", + "bbox": [ + 0.109, + 0.58, + 0.486, + 0.824 + ], + "angle": 0, + "content": "Finetuning-Free Personalization To reduce computational demands, recent research has shifted toward zero-shot personalization methods that eliminate subject-specific finetuning, typically employing image encoders to condition the generation process via the features extracted from the reference images [7, 15, 20, 33, 36]. Examples include BLIP-Diffusion [13], which pretrains a Q-Former to learn image features aligned with text, and IP-Adapter, which uses a frozen CLIP encoder to extract text-aligned visual features that modulate the cross-attention layers of the generative model. Other approaches, like Kosmos-G [19] and CAFE [38], connect large language models (LLMs) with diffusion models to condition generation on personalized concepts. SuTI [4] takes a different approach by training millions of subject-specific experts and subsequently training a model via apprenticeship learning, enabling effective zero-shot personalized generation at test time." + }, + { + "type": "text", + "bbox": [ + 0.109, + 0.825, + 0.485, + 0.883 + ], + "angle": 0, + "content": "Alternatively, encoder-free methods such as JeDi [37] and BootPIG [21] use features from the generative model's backbone to guide the generation. JeDi creates a multi-view synthetic dataset and modifies spatial self-attention to jointly" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.437, + 0.889, + 0.552 + ], + "angle": 0, + "content": "attend to images of the same concept in a batch. BootPIG retains a trainable copy of the original U-Net, adding reference self-attention layers to enable adaptation of the personalized model for reference features. While these methods remove the need for an additional encoder, they still require computationally expensive inference, as reference images must be processed in parallel during generation—a cost that accumulates due to the iterative nature of the diffusion process." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.553, + 0.889, + 0.653 + ], + "angle": 0, + "content": "In contrast, our method, DreamCache, caches a subset of reference features from the U-Net without text conditioning, eliminating the need for parallel inference and reducing memory overhead by avoiding separate model loading. This results in a test-time computational efficiency similar to encoder-based methods while offering the flexibility of encoder-free feature extraction and injection." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.653, + 0.889, + 0.739 + ], + "angle": 0, + "content": "Recent works such as BootPIG [21] and Toffee-5M [39] emphasize the importance of synthetic data that explicitly decouples the subject from the background, reporting improved performance. Inspired by these approaches, we adopt a similar generative pipeline to create a synthetic dataset we use to train DreamCache." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.74, + 0.888, + 0.812 + ], + "angle": 0, + "content": "Moreover, since our method is plug-and-play and keeps the U-Net frozen, we reduce the high cost of training faced by other approaches [15, 19, 21, 37-39]. A detailed overview of current methods, the number of trained parameters, and their training cost can be found in Table 1." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.825, + 0.89, + 0.883 + ], + "angle": 0, + "content": "Feature Caching Feature caching has been explored to reduce generation time in diffusion models by caching intermediate activations. Some studies [17, 34] exploit temporal redundancy during the training process to cache ac" + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.924, + 0.519, + 0.935 + ], + "angle": 0, + "content": "12482" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.135, + 0.117, + 0.872, + 0.345 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.108, + 0.363, + 0.89, + 0.418 + ], + "angle": 0, + "content": "Figure 3. Overview of DreamCache. Original U-Net layers are shown in violet, while the novel components introduced by DreamCache are highlighted in green. During personalization, features from selected layers of the diffusion denoiser are cached from a single timestep, using a null text prompt. These cached features serve as reference-specific information. During generation, conditioning adapters inject the cached features into the denoiser, modulating the features of the generated image to create a personalized output." + }, + { + "type": "text", + "bbox": [ + 0.109, + 0.442, + 0.486, + 0.63 + ], + "angle": 0, + "content": "tivations across timesteps, reducing computational load at later timesteps. Other works focus on caching layer activations within the diffusion framework, avoiding redundant computations. Learning-to-Cache [16] introduces a dynamic caching mechanism that learns to skip computation for selected layers of the diffusion model. In contrast to those works, which generally cache intra-model features for some layers to save computations, we utilize feature caching to encode multi-resolution features of a reference image from a few selected layers to condition the generation process of a new personalized image. Our approach recalls successful few-shot learners for discriminative problems [18, 27, 31] and extends them to personalized image generation." + }, + { + "type": "title", + "bbox": [ + 0.11, + 0.642, + 0.196, + 0.658 + ], + "angle": 0, + "content": "3. Method" + }, + { + "type": "text", + "bbox": [ + 0.109, + 0.666, + 0.486, + 0.824 + ], + "angle": 0, + "content": "Given a pretrained text-to-image generative model \\(\\epsilon_{\\theta}\\) and an image containing a reference subject \\(\\mathbf{I}_{\\mathrm{ref}}\\), the goal of personalized sampling is to generate novel images containing the reference subject in various contexts while maintaining textual control. We propose DreamCache, a novel approach for extracting conditioning signals from \\(\\mathbf{I}_{\\mathrm{ref}}\\) and guiding of the image generation process. This method leverages a pretrained diffusion model, conditioning adapters that are pretrained with a synthetic dataset, and a feature cache from the reference image. Sample outputs generated by DreamCache are shown in Fig. 2, with a method overview in Fig. 3." + }, + { + "type": "text", + "bbox": [ + 0.109, + 0.825, + 0.486, + 0.883 + ], + "angle": 0, + "content": "At the core of DreamCache, we utilize the denoiser in the pretrained diffusion model to extract multi-resolution features from \\(\\mathbf{I}_{\\mathrm{ref}}\\) by caching the activations of a few selected layers. To improve generalization, we cache features" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.442, + 0.889, + 0.557 + ], + "angle": 0, + "content": "using a forward pass with a null text prompt. When personalized sampling is performed, the cached features are processed by adapters to act as conditioning signals, modulating the denoiser features of the image under generation at corresponding layers. These adapters, once pretrained on a synthetic dataset, enable zero-shot personalized generation with any new reference image, requiring no further finetuning once its features are cached." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.56, + 0.89, + 0.632 + ], + "angle": 0, + "content": "In the following, we detail the three main aspects of DreamCache, namely i) how to cache reference features (Sec. 3.1); ii) how to condition the diffusion model on the cached features for personalized sampling (Sec. 3.2); iii) how to train the adapters used for model conditioning (Sec. 3.3)." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.642, + 0.755, + 0.658 + ], + "angle": 0, + "content": "3.1. Caching Reference Features" + }, + { + "type": "text", + "bbox": [ + 0.511, + 0.667, + 0.889, + 0.883 + ], + "angle": 0, + "content": "To extract information from the reference image for personalized sampling, we perform a forward pass through the denoiser of the diffusion model at a single timestep. We select \\( t = 1 \\), the least noisy timestep, to obtain clean features that are optimal for conditioning the personalized generation process. Additionally, we remove the text conditioning to decouple visual content of the reference image from the text caption, thus also eliminating the need for user-provided captions for reference images. This contrasts with methods such as JeDi [37], which are sensitive to caption content. During the forward pass, activations are computed for all layers of the denoiser, but only a subset is cached. Based on our experiments with the Stable Diffusion U-Net, we find that caching features from a middle bottleneck layer and every second layer in the decoder offers the best balance between" + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.924, + 0.518, + 0.935 + ], + "angle": 0, + "content": "12483" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.11, + 0.111, + 0.462, + 0.125 + ], + "angle": 0, + "content": "generation quality and caching efficiency (see Sec. 4.3)." + }, + { + "type": "text", + "bbox": [ + 0.11, + 0.126, + 0.484, + 0.184 + ], + "angle": 0, + "content": "Formally, the feature cache \\(\\mathcal{H}_{\\mathrm{FC}}\\) consists of the activations of the denoiser \\(\\epsilon_{\\theta}\\) from the selected layers \\(\\mathcal{L}\\) at timestep \\(t = 1\\), using a null text prompt \\(\\varnothing\\) and noisy reference image \\(\\mathbf{I}_{\\mathrm{ref}} + \\mathbf{n}_t\\), with noise realization \\(\\mathbf{n}_t\\), expressed as:" + }, + { + "type": "equation", + "bbox": [ + 0.186, + 0.197, + 0.484, + 0.213 + ], + "angle": 0, + "content": "\\[\n\\mathcal {H} _ {\\mathrm {F C}} = \\left\\{\\mathrm {h} _ {\\text {r e f}, L}: L \\in \\mathcal {L} \\right\\} \\tag {1}\n\\]" + }, + { + "type": "equation", + "bbox": [ + 0.178, + 0.215, + 0.484, + 0.231 + ], + "angle": 0, + "content": "\\[\n\\mathbf {h} _ {\\text {r e f}, L} = \\epsilon_ {\\theta} \\left(\\mathbf {I} _ {\\text {r e f}} + \\mathbf {n} _ {t}, \\varnothing , t; l) \\right| _ {t = 1, l = L}. \\tag {2}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.109, + 0.245, + 0.486, + 0.36 + ], + "angle": 0, + "content": "Notice that the cached features have different spatial resolutions, from the low-resolution bottleneck layer to the higher-resolution decoder layers, allowing a multi-resolution representation of the reference image. This is particularly useful for enabling both global semantics and fine-grained detail guidance. As in prior work [15, 21, 33], we foreground-segment the reference image to isolate the subject from the background before caching its features." + }, + { + "type": "title", + "bbox": [ + 0.11, + 0.372, + 0.468, + 0.387 + ], + "angle": 0, + "content": "3.2. Conditioning on Cached Reference Features" + }, + { + "type": "text", + "bbox": [ + 0.109, + 0.395, + 0.484, + 0.495 + ], + "angle": 0, + "content": "We propose a novel conditioning adapter mechanism composed of: i) a cross-attention block between the cached features and the features of the image under generation; ii) a concatenation operation between the features from the self-attention block of the original U-Net denoising backbone and those of the cross-attention block; and iii) a projection layer. A block diagram is shown in Fig. 3 (right)." + }, + { + "type": "text", + "bbox": [ + 0.11, + 0.497, + 0.484, + 0.526 + ], + "angle": 0, + "content": "Omitting the layer subscript for clarity, the conditioning adapter mechanism is expressed mathematically as follows:" + }, + { + "type": "equation", + "bbox": [ + 0.149, + 0.554, + 0.484, + 0.569 + ], + "angle": 0, + "content": "\\[\n\\mathbf {q} = \\mathbf {W} _ {Q} \\mathbf {h}, \\quad \\mathbf {k} _ {c} = \\mathbf {W} _ {K} \\mathbf {h} _ {\\text {r e f}}, \\quad \\mathbf {v} _ {c} = \\mathbf {W} _ {V} \\mathbf {h} _ {\\text {r e f}}, \\tag {3}\n\\]" + }, + { + "type": "equation", + "bbox": [ + 0.148, + 0.572, + 0.484, + 0.604 + ], + "angle": 0, + "content": "\\[\n\\mathbf {a} _ {c} = \\operatorname {s o f t m a x} \\left(\\frac {\\mathbf {q k} _ {c} ^ {T}}{\\sqrt {d}}\\right) \\mathbf {v} _ {c}, \\tag {4}\n\\]" + }, + { + "type": "equation", + "bbox": [ + 0.148, + 0.608, + 0.484, + 0.623 + ], + "angle": 0, + "content": "\\[\n\\mathbf {a} = \\mathbf {W} _ {\\text {p r o j}} \\left(\\left[ \\mathbf {a}; \\mathbf {a} _ {c} \\right]\\right), \\tag {5}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.109, + 0.636, + 0.486, + 0.809 + ], + "angle": 0, + "content": "where \\(\\mathbf{h} \\in \\mathbb{R}^{N \\times d}\\) are the current \\(d\\)-dimensional features of the \\(N\\)-pixel image under generation, and \\(\\mathbf{h}_{\\mathrm{ref}}\\) are the cached reference features from Eq. (2). \\(\\mathbf{W}_Q\\), \\(\\mathbf{W}_K\\), \\(\\mathbf{W}_V\\), and \\(\\mathbf{W}_{\\mathrm{proj}}\\) are learnable projection matrices, whose training is described in Sec. 3.3. The concatenation \\([\\mathbf{a}; \\mathbf{a}_c] \\in \\mathbb{R}^{N \\times 2d}\\) combines the output of self-attention \\(\\mathbf{a} \\in \\mathbb{R}^{N \\times d}\\) and cross-attention \\(\\mathbf{a}_c \\in \\mathbb{R}^{N \\times d}\\). The concatenation operation allows a flexible information fusion without explicit alignment constraints, compared to other approaches in similar works (see Sec. 4.3). The learnable projection matrix \\(\\mathbf{W}_{\\mathrm{proj}}\\) reduces the dimensionality of the concatenated features back to \\(\\mathbb{R}^{N \\times d}\\) to maintain compatibility with the original backbone." + }, + { + "type": "text", + "bbox": [ + 0.109, + 0.811, + 0.484, + 0.882 + ], + "angle": 0, + "content": "Overall, the approach proposed for the adapter enriches feature representations used in the diffusion process of the image under generation by allowing the model to leverage both primary and conditioning-based contextual information from the cache." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.11, + 0.81, + 0.126 + ], + "angle": 0, + "content": "3.3. Training the Conditioning Adapters" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.132, + 0.887, + 0.176 + ], + "angle": 0, + "content": "The additional parameters introduced in Sec. 3.2 to process the cached features must be trained on a large and varied dataset to ensure they generalize for any reference subject." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.176, + 0.888, + 0.348 + ], + "angle": 0, + "content": "Collecting paired data for this training process would be prohibitively expensive, as it requires multiple images of the same subject in different contexts. To address this, we draw inspiration from the recently proposed synthetic data generation pipeline in BootPIG [21] to construct our training data. First, we utilize a large language model (Llama 3.2 [5]) to generate captions for potential target images. Each caption is used to generate an image via Stable Diffusion [23]. We then use the Segment Anything Model (SAM) [11] and Grounding DINO [14] to accurately segment the reference subject based on the text caption and generate a foreground mask of the main object in the caption." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.348, + 0.889, + 0.492 + ], + "angle": 0, + "content": "We treat the Stable Diffusion-generated image as the target image, the foreground object pasted on a white background as the reference image, and the LLM-generated caption as the textual prompt during our training pipeline. Compared to BootPIG, our pipeline employs open-source models, making it more accessible. We will release our synthetic dataset to facilitate reproducibility and further research, since similar datasets, including BootPIG's [21], have not been publicly released. Additional details on the dataset and its statistics can be found in the Supp. Mat." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.492, + 0.889, + 0.549 + ], + "angle": 0, + "content": "We train the newly introduced adapters' parameters \\((\\mathbf{W}_Q, \\mathbf{W}_K, \\mathbf{W}_V, \\text{and} \\mathbf{W}_{\\mathrm{proj}})\\) with the standard score matching loss [29] using both the text-conditioned noisy input and the cached reference features:" + }, + { + "type": "equation", + "bbox": [ + 0.523, + 0.56, + 0.888, + 0.579 + ], + "angle": 0, + "content": "\\[\n\\mathcal {L} _ {\\text {d i f f u s i o n}} = \\mathbb {E} _ {\\mathbf {x} _ {0}, \\epsilon , \\mathbf {c} _ {T}, \\mathbf {I} _ {\\text {r e f}}, t} \\left[ \\| \\epsilon - \\epsilon_ {\\theta} ^ {\\prime} \\left(\\mathbf {x} _ {t}, \\mathbf {c} _ {T}, \\mathcal {H} _ {\\mathrm {F C}}, t\\right) \\| _ {2} ^ {2} \\right], \\tag {6}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.589, + 0.889, + 0.69 + ], + "angle": 0, + "content": "where \\(\\mathbf{x}_0\\) is the target image, \\(\\mathbf{c}_T\\) is the text prompt generated by the large language model, \\(\\epsilon\\) is Gaussian noise, and \\(t\\) is the diffusion timestep sampled uniformly from \\(1,\\dots ,T\\). The noisy image at timestep \\(t\\), \\(\\mathbf{x}_t\\), is obtained by gradually adding noise to \\(\\mathbf{x}_0\\) during the forward diffusion process. The function \\(\\epsilon_{\\theta}^{\\prime}\\) represents the modified denoising model that incorporates the conditioning adapters." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.702, + 0.709, + 0.718 + ], + "angle": 0, + "content": "4. Experimental Results" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.727, + 0.888, + 0.785 + ], + "angle": 0, + "content": "In this section, we present our experimental results, including quantitative and qualitative comparisons, an ablation study, and an analysis section that visualizes the behavior of the newly introduced cross-attention mechanism in the adapters." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.796, + 0.888, + 0.883 + ], + "angle": 0, + "content": "Implementation Details We evaluate our method on two versions of Stable Diffusion (SD) [23], specifically versions 1.5 and 2.1, to ensure fair comparison with state-of-the-art methods across different backbones. As described in the ablation study, our caching and conditioning mechanism is applied to the middle layer and every second layer of the" + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.925, + 0.518, + 0.935 + ], + "angle": 0, + "content": "12484" + } + ], + [ + { + "type": "table_caption", + "bbox": [ + 0.11, + 0.108, + 0.887, + 0.135 + ], + "angle": 0, + "content": "Table 2. Quantitative results on DreamBooth. DreamCache obtains a better balance between DINO score and CLIP-T compared to all baselines, while also offering a more efficient computational tradeoff (see Table 1)." + }, + { + "type": "table", + "bbox": [ + 0.216, + 0.145, + 0.78, + 0.442 + ], + "angle": 0, + "content": "
MethodBackbone#RefDINO (↑)CLIP-I (↑)CLIP-T (↑)
test-time finetuningDreamBooth [24]Imagen3-50.6960.8120.306
DreamBooth [24]SD 1.53-50.6680.8030.305
Textual Inversion [6]SD 1.53-50.5690.7800.255
Custom Diffusion [12]SD 1.53-50.6430.7900.305
BLIP-Diffusion (FT) [13]SD 1.53-50.6700.8050.302
finetuning freeELITE [33]SD 1.510.6210.7710.293
BLIP-Diffusion [13]SD 1.510.5940.7790.300
IP-Adapter [36]SD 1.510.6670.8130.289
Kosmos-G [19]SD 1.510.6940.8470.287
Jedi [37]SD 1.510.6190.7820.304
DreamCache (ours)SD 1.510.7130.8100.298
Re-Imagen [3]Imagen1-30.6000.7400.270
SuTI [4]Imagen1-30.7410.8190.304
Subject-Diffusion [15]SD 2.110.7710.7790.293
BootPig [21]SD 2.130.6740.7970.311
ToffeeNet [39]SD 2.110.7280.8170.306
CAFE [38]SD 2.110.7150.8270.294
DreamCache (ours)SD 2.110.7670.8160.301
" + }, + { + "type": "text", + "bbox": [ + 0.109, + 0.466, + 0.486, + 0.61 + ], + "angle": 0, + "content": "decoder. The total number of trainable parameters for DreamCache is 25M. We use the original SD codebase and train the model on \\(4 \\times 80\\mathrm{GB}\\) A100 GPUs for 25k steps with a batch size of 128, using the AdamW optimizer with a learning rate of \\(10^{-5}\\). Input images are resized to \\(512 \\times 512\\), with scale, shift, and resize augmentations applied to reference images to enhance model robustness to perturbations. Ablations are conducted on SD 1.5. We generate images with 50 sampling steps, employing classifier-free guidance for image and text conditioning, using a guidance scale of 7.5." + }, + { + "type": "text", + "bbox": [ + 0.109, + 0.621, + 0.486, + 0.809 + ], + "angle": 0, + "content": "Evaluation Quantitative evaluations are conducted on the DreamBooth dataset [24], following prior approaches. DreamBooth consists of 30 subjects, each with 25 text prompts. We use a single input image per subject and generate 4 images per subject-prompt combination, resulting in 3,000 generated images. We use pretrained DINO ViT-S/16 and CLIP ViT-B/32 models to calculate the average cosine similarity of global image embeddings between generated and reference images, with metrics denoted as DINO and CLIP-I, respectively. To assess text alignment, we calculate the cosine similarity between embeddings from generated images and text prompts using CLIP's image and text encoders [9], with the corresponding score denoted as CLIP-T." + }, + { + "type": "title", + "bbox": [ + 0.11, + 0.817, + 0.336, + 0.832 + ], + "angle": 0, + "content": "4.1. Zero-Shot Personalization" + }, + { + "type": "text", + "bbox": [ + 0.109, + 0.84, + 0.486, + 0.883 + ], + "angle": 0, + "content": "We compare DreamCache against state-of-the-art methods for finetuning-based and zero-shot personalization. Table 2 presents quantitative results, indicating the diffusion back" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.466, + 0.889, + 0.61 + ], + "angle": 0, + "content": "bone and the number of reference images for each method. Our approach achieves competitive or superior performance compared to other computationally-intensive state-of-the-art methods, which are trained on larger datasets and with significantly more parameters. We refer the reader to Table 1 for the data requirements, training time, and parameter count of the various methods. We remark that generally DINO is a preferred metric for image similarity with respect to CLIP-I, as it is more sensitive to the appearance and fine-grained details of the subjects." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.61, + 0.889, + 0.812 + ], + "angle": 0, + "content": "We also present qualitative comparisons with Kosmos-G [19] and BLIP-Diffusion [13]. We remark that several other methods are not reproducible due to the lack of code, datasets, or trained checkpoints. As seen in Fig. 4, our method excels in subject preservation and textual alignment, producing visually superior results. We also notice that Kosmos-G reports a high CLIP-I score, but after inspecting the generated images, it is clear that the score does not entirely reflect the preservation of the reference subject in generated images. In fact, Kosmos-G presents a high degree of background interference, where the partial replication of the reference background boosts the alignment score. For this reason, we also report foreground-masked metrics on the subjects, like MCLIP-I and MDINO, in the Supp. Mat." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.817, + 0.744, + 0.832 + ], + "angle": 0, + "content": "4.2. Inference Time Evaluation" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.84, + 0.889, + 0.883 + ], + "angle": 0, + "content": "The computational efficiency of our method is compared to the reference-based method BootPIG [21] and encoder-based approaches such as Kosmos-G [19] and Subject" + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.925, + 0.518, + 0.935 + ], + "angle": 0, + "content": "12485" + } + ], + [ + { + "type": "image_caption", + "bbox": [ + 0.142, + 0.108, + 0.198, + 0.122 + ], + "angle": 0, + "content": "\"A dog\"" + }, + { + "type": "image", + "bbox": [ + 0.117, + 0.123, + 0.224, + 0.205 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.144, + 0.209, + 0.197, + 0.22 + ], + "angle": 0, + "content": "\"A can\"" + }, + { + "type": "image", + "bbox": [ + 0.23, + 0.123, + 0.337, + 0.201 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.323, + 0.109, + 0.464, + 0.122 + ], + "angle": 0, + "content": "\"wearing a Santa hat\"" + }, + { + "type": "image", + "bbox": [ + 0.339, + 0.123, + 0.446, + 0.205 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.448, + 0.123, + 0.555, + 0.204 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.588, + 0.109, + 0.851, + 0.122 + ], + "angle": 0, + "content": "\"with the Eiffel Tower in the background\"" + }, + { + "type": "image", + "bbox": [ + 0.557, + 0.123, + 0.663, + 0.205 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.666, + 0.123, + 0.773, + 0.205 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.775, + 0.123, + 0.882, + 0.205 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.31, + 0.208, + 0.475, + 0.222 + ], + "angle": 0, + "content": "\"floating on top of water\"" + }, + { + "type": "image", + "bbox": [ + 0.117, + 0.222, + 0.224, + 0.304 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.146, + 0.307, + 0.196, + 0.32 + ], + "angle": 0, + "content": "\"A toy\"" + }, + { + "type": "image", + "bbox": [ + 0.23, + 0.222, + 0.337, + 0.305 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.34, + 0.222, + 0.446, + 0.305 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.449, + 0.222, + 0.555, + 0.305 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.557, + 0.222, + 0.664, + 0.305 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.666, + 0.222, + 0.773, + 0.305 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.775, + 0.222, + 0.882, + 0.305 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.345, + 0.308, + 0.442, + 0.32 + ], + "angle": 0, + "content": "\"on the beach\"" + }, + { + "type": "image", + "bbox": [ + 0.117, + 0.321, + 0.224, + 0.404 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.136, + 0.407, + 0.205, + 0.419 + ], + "angle": 0, + "content": "Reference" + }, + { + "type": "image", + "bbox": [ + 0.23, + 0.321, + 0.337, + 0.404 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.256, + 0.407, + 0.312, + 0.419 + ], + "angle": 0, + "content": "BLIP-D" + }, + { + "type": "image", + "bbox": [ + 0.34, + 0.321, + 0.446, + 0.404 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.357, + 0.407, + 0.428, + 0.419 + ], + "angle": 0, + "content": "Kosmos-G" + }, + { + "type": "image", + "bbox": [ + 0.449, + 0.321, + 0.555, + 0.404 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.455, + 0.407, + 0.544, + 0.419 + ], + "angle": 0, + "content": "DreamCache" + }, + { + "type": "image", + "bbox": [ + 0.557, + 0.321, + 0.664, + 0.404 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.583, + 0.407, + 0.638, + 0.419 + ], + "angle": 0, + "content": "BLIP-D" + }, + { + "type": "image", + "bbox": [ + 0.666, + 0.321, + 0.773, + 0.404 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.683, + 0.407, + 0.755, + 0.419 + ], + "angle": 0, + "content": "Kosmos-G" + }, + { + "type": "image", + "bbox": [ + 0.775, + 0.321, + 0.882, + 0.404 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.782, + 0.407, + 0.871, + 0.419 + ], + "angle": 0, + "content": "DreamCache" + }, + { + "type": "image_caption", + "bbox": [ + 0.11, + 0.433, + 0.887, + 0.46 + ], + "angle": 0, + "content": "Figure 4. Visual comparison. Personalized generations on sample concepts. DreamCache preserves reference concept appearance and does not suffer from background interference. BLIP-D [13] and Kosmos-G [19] cannot faithfully preserve visual details from the reference." + }, + { + "type": "table_caption", + "bbox": [ + 0.11, + 0.483, + 0.483, + 0.51 + ], + "angle": 0, + "content": "Table 3. Computational comparison.*: time to generate an image with 100 timesteps, evaluated on a single NVIDIA A100 GPU." + }, + { + "type": "table", + "bbox": [ + 0.115, + 0.52, + 0.48, + 0.609 + ], + "angle": 0, + "content": "
MethodInference Time*Extra Params Size
ELITE [33]6.24 s914 MB
BLIP-Diffusion [13]3.92 s760 MB
BootPig [21]7.55 s1900 MB
DreamCache (ours)3.88 s42 MB
" + }, + { + "type": "text", + "bbox": [ + 0.109, + 0.634, + 0.485, + 0.791 + ], + "angle": 0, + "content": "Diffusion [15]. Table 3 provides a detailed comparison of inference time, accounting for both personalization time (e.g., the time to generate the cache for DreamCache) and the time to sample the personalized image. We also report the increase in model size, i.e., the storage (in FP16 precision) required for the extra parameters to allow for personalization, showing that DreamCache is one order of magnitude more compact than the state of the art. Overall, DreamCache offers a lightweight solution that achieves state-of-the-art performance with faster inference and reduced computational overhead." + }, + { + "type": "title", + "bbox": [ + 0.11, + 0.802, + 0.266, + 0.816 + ], + "angle": 0, + "content": "4.3. Ablation Studies" + }, + { + "type": "text", + "bbox": [ + 0.109, + 0.825, + 0.484, + 0.883 + ], + "angle": 0, + "content": "We validate our design choices through a series of studies, examining different conditioning mechanisms, evaluating our feature caching approach, and analyzing the impact of synthetic dataset scaling." + }, + { + "type": "table_caption", + "bbox": [ + 0.512, + 0.483, + 0.887, + 0.51 + ], + "angle": 0, + "content": "Table 4. Reference feature integration. DreamCache uses the best tradeoff between accuracy and complexity." + }, + { + "type": "table", + "bbox": [ + 0.518, + 0.52, + 0.884, + 0.61 + ], + "angle": 0, + "content": "
MethodCLIP-I (↑)CLIP-T (↑)Params
Textual Sum [36]0.7880,28219M
Spatial Sum0.8120.29316M
Decoupled Blocks [8]0.8080.30061M
Spatial Concat (ours)0.8100.29825M
" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.636, + 0.889, + 0.852 + ], + "angle": 0, + "content": "Reference Feature Integration We compare various conditioning strategies to integrate reference features in Table 4. Our spatial cross-attention block with concatenation between the output of self- and cross-attention (\"Spatial Concat\") was evaluated against different alternatives, including IP-Adapter's conditioning mechanism [36] (\"Textual Sum\"), which sums the decoupled cross-attention output with that from textual cross-attention. We also tested a variant (\"Spatial Sum\") where self- and cross-attention conditioning outputs are summed. Additionally, we also assessed an alternative conditioning procedure inspired by ViCo [8] (\"Decoupled Blocks\"), involving independent and interleaved cross-attention blocks. Results in Table 4 indicate that the proposed \"Spatial Concat\" offers the best balance of text alignment and parameter efficiency." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.853, + 0.888, + 0.883 + ], + "angle": 0, + "content": "We further explored optimal conditioning insertion within the U-Net backbone in Table 5, determining that applying" + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.925, + 0.518, + 0.935 + ], + "angle": 0, + "content": "12486" + } + ], + [ + { + "type": "table_caption", + "bbox": [ + 0.11, + 0.109, + 0.486, + 0.135 + ], + "angle": 0, + "content": "Table 5. Cache positioning in the U-Net backbone offers a further tradeoff between accuracy and complexity." + }, + { + "type": "table", + "bbox": [ + 0.111, + 0.146, + 0.484, + 0.264 + ], + "angle": 0, + "content": "
EncoderMiddleDecoderCLIP-I (↑)CLIP-T (↑)Params
✓XX0.7210.30311M
✓✓X0.7490.30619M
X✓X0.7160.3028M
XX✓0.7990.29617M
X✓✓0.8100.29825M
✓✓✓0.8130.29736M
" + }, + { + "type": "table_caption", + "bbox": [ + 0.11, + 0.278, + 0.483, + 0.293 + ], + "angle": 0, + "content": "Table 6. Caching with text is not influential and adds complexity." + }, + { + "type": "table", + "bbox": [ + 0.183, + 0.303, + 0.413, + 0.363 + ], + "angle": 0, + "content": "
Text-FreeCLIP-I (↑)CLIP-T (↑)
×0.8110.295
✓0.8100.298
" + }, + { + "type": "table_caption", + "bbox": [ + 0.135, + 0.377, + 0.458, + 0.391 + ], + "angle": 0, + "content": "Table 7. Dataset impact for both synthetic and real data." + }, + { + "type": "table", + "bbox": [ + 0.165, + 0.401, + 0.432, + 0.49 + ], + "angle": 0, + "content": "
DatasetCLIP-I (↑)CLIP-T (↑)
Synthetic-50K0.7810.304
Synthetic-200K0.7970,301
Synthetic-400K0.8100.298
LAION-5M0.8140.242
" + }, + { + "type": "text", + "bbox": [ + 0.109, + 0.506, + 0.484, + 0.55 + ], + "angle": 0, + "content": "conditioning (and therefore feature caching) at the middle layer and every second layer of the decoder achieved the best tradeoff between performance and parameter count." + }, + { + "type": "text", + "bbox": [ + 0.109, + 0.557, + 0.486, + 0.673 + ], + "angle": 0, + "content": "Text Input for Cached Features Our feature caching procedure is designed to be text-free, leveraging the classifier-free guidance used during pretraining where captions were occasionally omitted. We compare this approach with a version including textual inputs during caching (e.g., 'A photo of...'). Table 6 shows that adding text conditioning slightly reduces text alignment while increasing complexity and potentially introducing noise in cases of inaccurate captions." + }, + { + "type": "text", + "bbox": [ + 0.109, + 0.681, + 0.486, + 0.884 + ], + "angle": 0, + "content": "Dataset Impact We demonstrate the importance of our synthetic dataset to train the conditioning adapters and the effect of scaling its size. For this purpose, we created synthetic datasets according to the procedure in Sec. 3.3 of sizes 50K, 200K, and 400K samples. We also tested the real-world 5M samples from the LAION [26] dataset, which, lacking target-caption-reference triplets, required reuse of target images as reference images too. Table 7 shows that increasing dataset size improves image alignment, though slightly reduces textual alignment. Notably, LAION improves image alignment but struggles with textual alignment. This highlights the importance of triplet data (target image, reference image, and caption) for effective zero-shot personalization, ensuring both subject preservation and textual editability." + }, + { + "type": "image", + "bbox": [ + 0.52, + 0.106, + 0.885, + 0.257 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.512, + 0.277, + 0.89, + 0.345 + ], + "angle": 0, + "content": "Figure 5. Visualization of reference image impact. Cross-attention maps between cached reference features and features of the image under generation. Left: attention map at layers at \\(16 \\times 16\\) resolution (left reference, right generated). Right: \\(32 \\times 32\\). Attention values are highly localized in the region of interest." + }, + { + "type": "title", + "bbox": [ + 0.512, + 0.358, + 0.766, + 0.374 + ], + "angle": 0, + "content": "4.4. Visualizing Reference Impact" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.381, + 0.889, + 0.525 + ], + "angle": 0, + "content": "Finally, we analyze how the cross-attention mechanism in DreamCache impacts image generation by visualizing cached reference feature influence. Attention map visualizations at different resolutions are provided in Fig. 5. Specifically, attention maps between the query from the current generation and the key derived from reference features reveal a highly localized focus on the subject, without interference from background elements. This mechanism models correspondences effectively, integrating reference information into the generated image." + }, + { + "type": "title", + "bbox": [ + 0.512, + 0.531, + 0.757, + 0.545 + ], + "angle": 0, + "content": "5. Discussion and Conclusions" + }, + { + "type": "text", + "bbox": [ + 0.511, + 0.551, + 0.889, + 0.883 + ], + "angle": 0, + "content": "In this paper, we proposed DreamCache, a novel approach to personalized text-to-image generation that uses feature caching to overcome the limitations of existing methods. By caching reference features from a small subset of layers of the U-Net only once, our method significantly reduces both computational and memory demands, enabling efficient, real-time personalized image generation. Unlike previous approaches, DreamCache avoids the need for costly finetuning, external image encoders, or parallel reference processing, making it lightweight and suitable for plug-and-play deployment. Our experiments demonstrate that DreamCache achieves state-of-the-art zero-shot personalization with only 25M additional parameters and a fast training process. While DreamCache is a promising direction towards more efficient personalized generation, it has some limitations. Although effective for single-subject personalization, our approach may require adaptation for complex multi-subject generation where feature interference can occur. Additionally, certain edge cases, such as highly abstract or stylistic images, may challenge the caching mechanism's capacity to accurately preserve subject details. To address these challenges, future work may explore adaptive caching techniques or multi-reference feature integration." + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.924, + 0.518, + 0.935 + ], + "angle": 0, + "content": "12487" + } + ], + [ + { + "type": "title", + "bbox": [ + 0.111, + 0.109, + 0.201, + 0.124 + ], + "angle": 0, + "content": "References" + }, + { + "type": "ref_text", + "bbox": [ + 0.119, + 0.133, + 0.486, + 0.198 + ], + "angle": 0, + "content": "[1] Moab Arar, Rinon Gal, Yuval Atzmon, Gal Chechik, Daniel Cohen-Or, Ariel Shamir, and Amit H. Bermano. Domain-agnostic tuning-encoder for fast personalization of text-to-image models. In SIGGRAPH Asia 2023 Conference Papers, pages 1–10, 2023. 1, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.119, + 0.2, + 0.486, + 0.265 + ], + "angle": 0, + "content": "[2] Mathilde Caron, Hugo Touvron, Ishan Misra, Hervé Jégou, Julien Mairal, Piotr Bojanowski, and Armand Joulin. Emerging properties in self-supervised vision transformers. In Proceedings of the IEEE/CVF international conference on computer vision, pages 9650-9660, 2021. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.119, + 0.267, + 0.485, + 0.319 + ], + "angle": 0, + "content": "[3] Wenhu Chen, Hexiang Hu, Chitwan Sahara, and William W Cohen. Re-imagen: Retrieval-augmented text-to-image generator. In The Eleventh International Conference on Learning Representations, 2022. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.119, + 0.321, + 0.484, + 0.374 + ], + "angle": 0, + "content": "[4] Wenhu Chen, Hexiang Hu, Yandong Li, Nataniel Rui, Xuhui Jia, Ming-Wei Chang, and William W Cohen. Subject-driven text-to-image generation via apprenticeship learning. arXiv preprint arXiv:2304.00186, 2023. 2, 3, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.119, + 0.375, + 0.485, + 0.438 + ], + "angle": 0, + "content": "[5] Abhimanyu Dubey, Abhinav Jauhri, Abhinav Pandey, Abhishek Kadian, Ahmad Al-Dahle, Aiesha Letman, Akhil Mathur, Alan Schelten, Amy Yang, Angela Fan, et al. The llama 3 herd of models. arXiv preprint arXiv:2407.21783, 2024. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.119, + 0.441, + 0.486, + 0.519 + ], + "angle": 0, + "content": "[6] Rinon Gal, Yuval Alaluf, Yuval Atzmon, Or Patashnik, Amit Haim Bermano, Gal Chechik, and Daniel Cohen-or. An image is worth one word: Personalizing text-to-image generation using textual inversion. In The Eleventh International Conference on Learning Representations, 2022. 1, 2, 3, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.119, + 0.521, + 0.485, + 0.574 + ], + "angle": 0, + "content": "[7] Rinon Gal, Moab Arar, Yuval Atzmon, Amit H Bermano, Gal Chechik, and Daniel Cohen-Or. Encoder-based domain tuning for fast personalization of text-to-image models. ACM Transactions on Graphics (TOG), 42(4):1-13, 2023. 1, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.119, + 0.576, + 0.485, + 0.626 + ], + "angle": 0, + "content": "[8] Shaozhe Hao, Kai Han, Shihao Zhao, and Kwan-Yee K Wong. Vico: Plug-and-play visual condition for personalized text-to-image generation. arXiv preprint arXiv:2306.00971, 2023. 1, 3, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.119, + 0.629, + 0.485, + 0.693 + ], + "angle": 0, + "content": "[9] Jack Hessel, Ari Holtzman, Maxwell Forbes, Ronan Le Bras, and Yejin Choi. Clipscore: A reference-free evaluation metric for image captioning. In Proceedings of the 2021 Conference on Empirical Methods in Natural Language Processing, pages 7514-7528, 2021. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.111, + 0.695, + 0.485, + 0.734 + ], + "angle": 0, + "content": "[10] Jonathan Ho, Ajay Jain, and Pieter Abbeel. Denoising diffusion probabilistic models. Advances in neural information processing systems, 33:6840-6851, 2020. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.111, + 0.736, + 0.485, + 0.788 + ], + "angle": 0, + "content": "[11] Alexander Kirillov, Eric Mintun, Nikhila Ravi, Hanzi Mao, Chloe Rolland, Laura Gustafson, Tete Xiao, Spencer Whitehead, Alexander C Berg, Wan-Yen Lo, et al. Segment anything. arXiv preprint arXiv:2304.02643, 2023. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.111, + 0.79, + 0.485, + 0.853 + ], + "angle": 0, + "content": "[12] Nupur Kumari, Bingliang Zhang, Richard Zhang, Eli Shechtman, and Jun-Yan Zhu. Multi-concept customization of text-to-image diffusion. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 1931-1941, 2023. 1, 2, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.111, + 0.856, + 0.485, + 0.882 + ], + "angle": 0, + "content": "[13] Dongxu Li, Junnan Li, and Steven CH Hoi. Blip-diffusion: Pre-trained subject representation for control" + }, + { + "type": "list", + "bbox": [ + 0.111, + 0.133, + 0.486, + 0.882 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.545, + 0.112, + 0.887, + 0.137 + ], + "angle": 0, + "content": "lable text-to-image generation and editing. arXiv preprint arXiv:2305.14720, 2023. 1, 2, 3, 6, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.516, + 0.139, + 0.887, + 0.204 + ], + "angle": 0, + "content": "[14] Shilong Liu, Zhaoyang Zeng, Tianhe Ren, Feng Li, Hao Zhang, Jie Yang, Chunyuan Li, Jianwei Yang, Hang Su, Jun Zhu, et al. Grounding dino: Marrying dino with grounded pre-training for open-set object detection. arXiv preprint arXiv:2303.05499, 2023. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.515, + 0.207, + 0.887, + 0.259 + ], + "angle": 0, + "content": "[15] Jian Ma, Junhao Liang, Chen Chen, and Haonan Lu. Subject-diffusion: Open domain personalized text-to-image generation without test-time fine-tuning. arXiv preprint arXiv:2307.11410, 2023. 1, 2, 3, 5, 6, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.515, + 0.261, + 0.887, + 0.3 + ], + "angle": 0, + "content": "[16] Xinyin Ma, Gongfan Fang, Michael Bi Mi, and Xinchao Wang. Learning-to-cache: Accelerating diffusion transformer via layer caching. arXiv preprint arXiv:2406.01733, 2024. 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.515, + 0.302, + 0.887, + 0.355 + ], + "angle": 0, + "content": "[17] Xinyin Ma, Gongfan Fang, and Xinchao Wang. Deepcache: Accelerating diffusion models for free. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 15762-15772, 2024. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.515, + 0.357, + 0.887, + 0.408 + ], + "angle": 0, + "content": "[18] Boris Oreshkin, Pau Rodríguez López, and Alexandre Lacoste. Tadam: Task dependent adaptive metric for improved few-shot learning. Advances in neural information processing systems, 31, 2018. 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.515, + 0.411, + 0.887, + 0.462 + ], + "angle": 0, + "content": "[19] Xichen Pan, Li Dong, Shaohan Huang, Zhiliang Peng, Wenhu Chen, and Furu Wei. Kosmos-g: Generating images in context with multimodal large language models. arXiv preprint arXiv:2310.02992, 2023. 1, 2, 3, 6, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.515, + 0.466, + 0.887, + 0.517 + ], + "angle": 0, + "content": "[20] Maitreya Patel, Sangmin Jung, Chitta Baral, and Yezhou Yang. λ-eclipse: Multi-concept personalized text-to-image diffusion models by leveraging clip latent space. arXiv preprint arXiv:2402.05195, 2024. 1, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.515, + 0.52, + 0.887, + 0.572 + ], + "angle": 0, + "content": "[21] Senthil Purushwalkam, Akash Gokul, Shafiq Joty, and Nikhil Naik. Bootpig: Bootstrapping zero-shot personalized image generation capabilities in pretrained diffusion models. arXiv preprint arXiv:2401.13974, 2024. 1, 2, 3, 5, 6, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.515, + 0.575, + 0.887, + 0.652 + ], + "angle": 0, + "content": "[22] Alec Radford, Jong Wook Kim, Chris Hallacy, Aditya Ramesh, Gabriel Goh, Sandhini Agarwal, Girish Sastry, Amanda Askell, Pamela Mishkin, Jack Clark, et al. Learning transferable visual models from natural language supervision. In International conference on machine learning, pages 8748-8763. PMLR, 2021. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.515, + 0.655, + 0.887, + 0.72 + ], + "angle": 0, + "content": "[23] Robin Rombach, Andreas Blattmann, Dominik Lorenz, Patrick Esser, and Björn Ommer. High-resolution image synthesis with latent diffusion models. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 10684-10695, 2022. 1, 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.515, + 0.723, + 0.887, + 0.8 + ], + "angle": 0, + "content": "[24] Nataniel Ruiz, Yuanzhen Li, Varun Jampani, Yael Pritch, Michael Rubinstein, and Kfir Aberman. Dreambooth: Fine tuning text-to-image diffusion models for subject-driven generation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 22500-22510, 2023. 1, 2, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.515, + 0.803, + 0.887, + 0.882 + ], + "angle": 0, + "content": "[25] Chitwan Sahara, William Chan, Saurabh Saxena, Lala Li, Jay Whang, Emily L Denton, Kamyar Ghasemipour, Raphael Gontijo Lopes, Burcu Karagol Ayan, Tim Salimans, et al. Photorealistic text-to-image diffusion models with deep language understanding. Advances in Neural Information Processing Systems, 35:36479-36494, 2022. 1" + }, + { + "type": "list", + "bbox": [ + 0.515, + 0.112, + 0.887, + 0.882 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.925, + 0.518, + 0.935 + ], + "angle": 0, + "content": "12488" + } + ], + [ + { + "type": "ref_text", + "bbox": [ + 0.111, + 0.111, + 0.486, + 0.19 + ], + "angle": 0, + "content": "[26] Christoph Schuhmann, Romain Beaumont, Richard Vencu, Cade Gordon, Ross Wightman, Mehdi Cherti, Theo Coombes, Aarush Katta, Clayton Mullis, Mitchell Wortsman, et al. Laion-5b: An open large-scale dataset for training next generation image-text models. Advances in Neural Information Processing Systems, 35:25278–25294, 2022. 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.111, + 0.191, + 0.486, + 0.231 + ], + "angle": 0, + "content": "[27] Jake Snell, Kevin Swersky, and Richard Zemel. Prototypical networks for few-shot learning. Advances in neural information processing systems, 30, 2017. 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.113, + 0.232, + 0.485, + 0.295 + ], + "angle": 0, + "content": "[28] Jascha Sohl-Dickstein, Eric Weiss, Niru Maheswaranathan, and Surya Ganguli. Deep unsupervised learning using nonequilibrium thermodynamics. In International conference on machine learning, pages 2256-2265. PMLR, 2015. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.113, + 0.298, + 0.483, + 0.337 + ], + "angle": 0, + "content": "[29] Yang Song and Stefano Ermon. Generative modeling by estimating gradients of the data distribution. Advances in neural information processing systems, 32, 2019. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.113, + 0.338, + 0.484, + 0.389 + ], + "angle": 0, + "content": "[30] Yoad Tewel, Rinon Gal, Gal Chechik, and Yuval Atzmon. Key-locked rank one editing for text-to-image personalization. In ACM SIGGRAPH 2023 Conference Proceedings, pages 1-11, 2023. 1, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.113, + 0.391, + 0.485, + 0.457 + ], + "angle": 0, + "content": "[31] Eleni Triantafillou, Hugo Larochelle, Jake Snell, Josh Tenenbaum, Kevin Jordan Swersky, Mengye Ren, Richard Zemel, and Sachin Ravi. Meta-learning for semi-supervised few-shot classification. In International Conference on Learning Representations, 2018. 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.113, + 0.458, + 0.483, + 0.497 + ], + "angle": 0, + "content": "[32] Andrey Voynov, Qinghao Chu, Daniel Cohen-Or, and Kfir Aberman. p+: Extended textual conditioning in text-to-image generation. arXiv preprint arXiv:2303.09522, 2023. 1, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.113, + 0.498, + 0.485, + 0.563 + ], + "angle": 0, + "content": "[33] Yuxiang Wei, Yabo Zhang, Zhilong Ji, Jinfeng Bai, Lei Zhang, and Wangmeng Zuo. Elite: Encoding visual concepts into textual embeddings for customized text-to-image generation. 2023 IEEE/CVF International Conference on Computer Vision (ICCV), pages 15897-15907, 2023. 1, 2, 3, 5, 6, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.113, + 0.565, + 0.484, + 0.643 + ], + "angle": 0, + "content": "[34] Felix Wimbauer, Bichen Wu, Edgar Schoenfeld, Xiaoliang Dai, Ji Hou, Zijian He, Artsiom Sanakoyeu, Peizhao Zhang, Sam Tsai, Jonas Kohler, et al. Cache me if you can: Accelerating diffusion models through block caching. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 6211-6220, 2024. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.113, + 0.645, + 0.484, + 0.696 + ], + "angle": 0, + "content": "[35] Jianan Yang, Haobo Wang, Yanming Zhang, Ruixuan Xiao, Sai Wu, Gang Chen, and Junbo Zhao. Controllable textual inversion for personalized text-to-image generation. arXiv preprint arXiv:2304.05265, 2023. 1, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.113, + 0.697, + 0.485, + 0.748 + ], + "angle": 0, + "content": "[36] Hu Ye, Jun Zhang, Sibo Liu, Xiao Han, and Wei Yang. Ip-adapter: Text compatible image prompt adapter for text-to-image diffusion models. arXiv preprint arXiv:2308.06721, 2023. 1, 2, 3, 6, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.113, + 0.751, + 0.484, + 0.829 + ], + "angle": 0, + "content": "[37] Yu Zeng, Vishal M Patel, Haochen Wang, Xun Huang, Ting-Chun Wang, Ming-Yu Liu, and Yogesh Balaji. Jedi: Joint-image diffusion models for finetuning-free personalized text-to-image generation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 6786-6795, 2024. 1, 2, 3, 4, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.113, + 0.831, + 0.484, + 0.882 + ], + "angle": 0, + "content": "[38] Yufan Zhou, Ruiyi Zhang, Jiumiang Gu, and Tong Sun. Customization assistant for text-to-image generation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 9182-9191, 2024. 2, 3, 6" + }, + { + "type": "list", + "bbox": [ + 0.111, + 0.111, + 0.486, + 0.882 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.515, + 0.111, + 0.89, + 0.177 + ], + "angle": 0, + "content": "[39] Yufan Zhou, Ruiyi Zhang, Kaizhi Zheng, Nanxuan Zhao, Jiumiang Gu, Zichao Wang, Xin Eric Wang, and Tong Sun. Toffee: Efficient million-scale dataset construction for subject-driven text-to-image generation. arXiv preprint arXiv:2406.09305, 2024.1,2,3,6" + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.925, + 0.518, + 0.935 + ], + "angle": 0, + "content": "12489" + } + ] +] \ No newline at end of file diff --git a/2025/DreamCache_ Finetuning-Free Lightweight Personalized Image Generation via Feature Caching/484a2287-cf26-4bf8-9e41-034501dc6452_origin.pdf b/2025/DreamCache_ Finetuning-Free Lightweight Personalized Image Generation via Feature Caching/484a2287-cf26-4bf8-9e41-034501dc6452_origin.pdf new file mode 100644 index 0000000000000000000000000000000000000000..652d986b1dda158a1a83c6589e96d35a508ca682 --- /dev/null +++ b/2025/DreamCache_ Finetuning-Free Lightweight Personalized Image Generation via Feature Caching/484a2287-cf26-4bf8-9e41-034501dc6452_origin.pdf @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:b4869aeba633890ef6ae3c75228bb2547fab93f0196c69d5490839219b041b09 +size 1135826 diff --git a/2025/DreamCache_ Finetuning-Free Lightweight Personalized Image Generation via Feature Caching/full.md b/2025/DreamCache_ Finetuning-Free Lightweight Personalized Image Generation via Feature Caching/full.md new file mode 100644 index 0000000000000000000000000000000000000000..b8a1444663693ff8a19812748c97dcb959b30802 --- /dev/null +++ b/2025/DreamCache_ Finetuning-Free Lightweight Personalized Image Generation via Feature Caching/full.md @@ -0,0 +1,340 @@ +# DreamCache: Finetuning-Free Lightweight Personalized Image Generation via Feature Caching + +Emanuele Aiello Umberto Michieli Diego Valsesia Mete Ozay Enrico Magli Politecnico di Torino Samsung R&D Institute UK + +# Abstract + +Personalized image generation requires text-to-image generative models that capture the core features of a reference subject to allow for controlled generation across different contexts. Existing methods face challenges due to complex training requirements, high inference costs, limited flexibility, or a combination of these issues. In this paper, we introduce DreamCache, a scalable approach for efficient and high-quality personalized image generation. By caching a small number of reference image features from a subset of layers and a single timestep of the pretrained diffusion denoiser, DreamCache enables dynamic modulation of the generated image features through lightweight, trained conditioning adapters. DreamCache achieves state-of-the-art image and text alignment, utilizing an order of magnitude fewer extra parameters, and is both more computationally effective and versatile than existing models. + +# 1. Introduction + +Recent advancements in text-to-image generation, fueled by the development of diffusion models [10, 28], have enabled high-quality and diverse image generation from textual descriptions. Diffusion models [23, 25] gradually transform random noise into images through a sequence of denoising steps, conditioned on the input text prompt. + +An active area of research is personalizing these models, enabling the generation of novel images of a reference subject in various contexts, while maintaining flexibility for text-based editing. Early personalization techniques [1, 6, 8, 24, 30, 32, 35, 35], such as the seminal Dream-Booth [24] relied on fine-tuning (FT) the generative model for each reference subject. However, these approaches are often impractical for many use cases due to costly test-time FT, which can take several minutes per subject. To address this, FT-free (i.e., zero-shot) personalized image generation methods have emerged to eliminate test-time optimization. These FT-free approaches can be broadly categorized into + +![](images/2ff59b4bdbab438456d466a87c6e85d3cb828d258db8bfe36cf3490f74168362.jpg) +Figure 1. DreamCache is a finetuning-free personalized image generation method that achieves an optimal balance between subject fidelity, memory efficiency, and adherence to text prompts. + +two families: encoder-based methods and reference-based methods, each with distinct drawbacks. + +Encoder-based methods [7, 15, 20, 33, 36] utilize dedicated image encoders, such as CLIP [22] or DINO [2], to extract relevant features from reference images. While these encoders can produce high-quality results, they are often large, require extensive training to align text and image features, and reduce the model's flexibility [13, 15, 19, 36]. + +In contrast, reference-based methods [21, 37] condition the diffusion model directly on reference features drawn from the U-Net denoiser, integrating these features at each denoising step. While effective, these methods require feature extraction at every generation step, leading to higher computational costs and memory demands. Additionally, they often require an input textual caption for conditioning, which introduces variability and can decrease output precision. + +Some recent works have proposed to finetune the U-Net backbone itself [21, 37, 39]. However, this hinders the model's ability to switch between personalized and nonpersonalized tasks and risks inducing the "language drift" phenomenon, where the personalization training degrades the model's linguistic comprehension [12, 24]. + +Table 1. Methods overview. Our DreamCache achieves state-of-the-art generation quality at reduced computational costs. *: value refers to the personalization stage for each personal subject. + +
MethodFT-freeEnc-freePlug&PlayRef-UNet-freeExtra ParamsTrain Params# DatasetTrain Time
Textual Inversion [6]X✓✓✓768*768*3-5*50 min*
DreamBooth [24]X✓X✓-0.9B*3-5*10 min*
Custom Diffusion [12]X✓X✓-57M*3-5*10 min*
ELITE [33]✓XX✓457M77M125K14 days
BLIP-Diffusion [13]✓X✓✓380M1.5B129M96 days
IP-Adapter [36]✓X✓✓402M22M10M28 days
Kosmos-G [19]✓X✓✓1.6B1.6B9M-
JeDi [37]✓✓XX-0.9B3M48 days
SuTI [4]✓XX✓400M2.5B500K-
Subject-Diffusion [15]✓X✓X700M700M76M-
BootPig [21]✓✓XX0.95B0.95B200K18 hours
ToffeeNet [39]✓XX✓632M0.9B5M-
CAFE [38]✓XX✓14B1B355K-
DreamCache (ours)✓✓✓✓25M25M400K40 hours
+ +In this work, we propose DreamCache, a novel finetuning-free approach to personalized image generation, that overcomes the limitations of existing methods (see Fig. 1) by using a feature caching mechanism that enables text-free encoding and efficient conditioning during personalization. Specifically, we first create a synthetic dataset [21] containing triplets of captions, target images, and reference subjects to capture subjects in various contexts. Next, we pretrain lightweight attention-based conditioning adapters to inject subject-specific features into the image generation process. During personalization, the reference image is processed through the pretrained denoiser of the base diffusion model without text conditioning, thus eliminating the need for user-generated captions, while caching features from a small subset of layers at a single timestep. For personalized sampling, these cached features are injected into the denoiser through the pretrained conditioning adapters. + +Table 1 summarizes the key properties of existing methods and illustrates how DreamCache fits within the current landscape; further details are explored in Sec. 2. As an encoder-free approach, DreamCache introduces only a small number of additional parameters, making it significantly lightweight and suitable for deployment on resource-constrained devices. For example, methods like [33] and [13] introduce 380M parameters due to their reliance on CLIP encoders, whereas DreamCache requires only 25M additional parameters. Moreover, caching features from a few selected U-Net layers at a single preprocessing timestep bypasses the need for full U-Net reference processing during generation, leading to substantial computational and memory savings that enable real-time, high-quality personalized generation. Another key advantage of DreamCache is its plug-and-play + +nature, allowing concurrent generation of personalized and non-personalized content without altering original U-Net weights, thus preserving the integrity of the original model and enabling a wider range of deployment scenarios, especially on mobile platforms. + +In summary, DreamCache represents a significant step toward practical, and scalable personalized image generation, with the following contributions: + +- We propose a feature caching approach that creates multi-resolution representations of the reference image in a caption-free and efficient manner. +- We design an attention-based conditioning mechanism that leverages the cached features for personalized image generation, achieving computational- and memory- efficient personalized sampling. +- Our approach achieves state-of-the-art quality in personalized image generation at substantially lower computational and data costs compared to existing methods. + +# 2. Background and Related Work + +Personalized image generation aims at generating images containing a specific subject. This task has been widely studied, resulting in two main approaches: fine-tuning methods, which require test-time finetuning on multiple subject reference images, and finetuning-free (zero-shot) methods, which learn a generalizable conditioning mechanism to generate reference subjects without the need for further optimization. + +Finetuning-based Personalization DreamBooth [24] finetunes the entire U-Net with reference images while introducing a regularization loss to mitigate overfitting. On the other hand, Custom Diffusion [12] only finetunes the $K$ and + +![](images/72e59097773136a3ca128e6e35d632f15e0757d7e32b35195cf460eb0822589d.jpg) +"A dragon..." + +![](images/7063e425d55245aac90538df6ccdc4c59e0aa1dc7fdda3987894df887d553467.jpg) +"as street graffiti" + +![](images/a8d50836b16284dc7b021527485e2a63d2f65a1c505a1c98940637cd5f9f63ac.jpg) +"playing with fire" + +![](images/3288fe34e86e7c2c5690b5f415cf6ce77dd50e11d50a4ce0fc0845a094bada51.jpg) +"as a plushie" + +![](images/42cba27336adc0a3a020f273eee03a02b1786e4b1912d7211f0edb071baf6f56.jpg) +"working as a barista" + +![](images/b8fd579d3cf15ec40ec046824de141036ba2da37a0d3e1c0c12a7a18972c5a68.jpg) +"A cat..." + +![](images/00db6d3885758325f7772d3f86d0e88bf4b11f9b17c537c25770b5e27d7f2d12.jpg) +"in Ukiyo-e style" + +![](images/957f74af1e220a4a34c67d7f3ce19eda399f8a1118a1aa1270ce965557b246c9.jpg) +"with a rainbow scarf" +Figure 2. Personalized generations by DreamCache. The first column contains reference images. The generated images correspond to the text prompts above each column. + +![](images/969b036c282fc912b595c03d260274a3a16a411b9f73873cf36260c6eae53b59.jpg) +"Van Gogh painting" + +![](images/5a3086329e7c27ee0b1766b5132b3ec444206b07702c156a24d8bbb408726be4.jpg) +"wearing a diploma hat" + +$V$ projections for the cross-attention blocks of the U-Net. Text-based personalization methods optimize single (like Textual Inversion [6]) or multiple (like in $\mathrm{P + }$ [32]) input token embeddings. Later methods [1, 8, 30, 35] build on these, with innovations like Perfusion [30] using dynamic rank-1 updates to prevent overfitting while keeping encodings lightweight. However, all finetuning-based methods are computationally intensive, often requiring minutes of finetuning per reference subject at test time. + +Finetuning-Free Personalization To reduce computational demands, recent research has shifted toward zero-shot personalization methods that eliminate subject-specific finetuning, typically employing image encoders to condition the generation process via the features extracted from the reference images [7, 15, 20, 33, 36]. Examples include BLIP-Diffusion [13], which pretrains a Q-Former to learn image features aligned with text, and IP-Adapter, which uses a frozen CLIP encoder to extract text-aligned visual features that modulate the cross-attention layers of the generative model. Other approaches, like Kosmos-G [19] and CAFE [38], connect large language models (LLMs) with diffusion models to condition generation on personalized concepts. SuTI [4] takes a different approach by training millions of subject-specific experts and subsequently training a model via apprenticeship learning, enabling effective zero-shot personalized generation at test time. + +Alternatively, encoder-free methods such as JeDi [37] and BootPIG [21] use features from the generative model's backbone to guide the generation. JeDi creates a multi-view synthetic dataset and modifies spatial self-attention to jointly + +attend to images of the same concept in a batch. BootPIG retains a trainable copy of the original U-Net, adding reference self-attention layers to enable adaptation of the personalized model for reference features. While these methods remove the need for an additional encoder, they still require computationally expensive inference, as reference images must be processed in parallel during generation—a cost that accumulates due to the iterative nature of the diffusion process. + +In contrast, our method, DreamCache, caches a subset of reference features from the U-Net without text conditioning, eliminating the need for parallel inference and reducing memory overhead by avoiding separate model loading. This results in a test-time computational efficiency similar to encoder-based methods while offering the flexibility of encoder-free feature extraction and injection. + +Recent works such as BootPIG [21] and Toffee-5M [39] emphasize the importance of synthetic data that explicitly decouples the subject from the background, reporting improved performance. Inspired by these approaches, we adopt a similar generative pipeline to create a synthetic dataset we use to train DreamCache. + +Moreover, since our method is plug-and-play and keeps the U-Net frozen, we reduce the high cost of training faced by other approaches [15, 19, 21, 37-39]. A detailed overview of current methods, the number of trained parameters, and their training cost can be found in Table 1. + +Feature Caching Feature caching has been explored to reduce generation time in diffusion models by caching intermediate activations. Some studies [17, 34] exploit temporal redundancy during the training process to cache ac + +![](images/cc90f61ffcc2fdd5a17e9ca033fa3489b1b277e98fc320d9267a8e3a9e803009.jpg) +Figure 3. Overview of DreamCache. Original U-Net layers are shown in violet, while the novel components introduced by DreamCache are highlighted in green. During personalization, features from selected layers of the diffusion denoiser are cached from a single timestep, using a null text prompt. These cached features serve as reference-specific information. During generation, conditioning adapters inject the cached features into the denoiser, modulating the features of the generated image to create a personalized output. + +tivations across timesteps, reducing computational load at later timesteps. Other works focus on caching layer activations within the diffusion framework, avoiding redundant computations. Learning-to-Cache [16] introduces a dynamic caching mechanism that learns to skip computation for selected layers of the diffusion model. In contrast to those works, which generally cache intra-model features for some layers to save computations, we utilize feature caching to encode multi-resolution features of a reference image from a few selected layers to condition the generation process of a new personalized image. Our approach recalls successful few-shot learners for discriminative problems [18, 27, 31] and extends them to personalized image generation. + +# 3. Method + +Given a pretrained text-to-image generative model $\epsilon_{\theta}$ and an image containing a reference subject $\mathbf{I}_{\mathrm{ref}}$ , the goal of personalized sampling is to generate novel images containing the reference subject in various contexts while maintaining textual control. We propose DreamCache, a novel approach for extracting conditioning signals from $\mathbf{I}_{\mathrm{ref}}$ and guiding of the image generation process. This method leverages a pretrained diffusion model, conditioning adapters that are pretrained with a synthetic dataset, and a feature cache from the reference image. Sample outputs generated by DreamCache are shown in Fig. 2, with a method overview in Fig. 3. + +At the core of DreamCache, we utilize the denoiser in the pretrained diffusion model to extract multi-resolution features from $\mathbf{I}_{\mathrm{ref}}$ by caching the activations of a few selected layers. To improve generalization, we cache features + +using a forward pass with a null text prompt. When personalized sampling is performed, the cached features are processed by adapters to act as conditioning signals, modulating the denoiser features of the image under generation at corresponding layers. These adapters, once pretrained on a synthetic dataset, enable zero-shot personalized generation with any new reference image, requiring no further finetuning once its features are cached. + +In the following, we detail the three main aspects of DreamCache, namely i) how to cache reference features (Sec. 3.1); ii) how to condition the diffusion model on the cached features for personalized sampling (Sec. 3.2); iii) how to train the adapters used for model conditioning (Sec. 3.3). + +# 3.1. Caching Reference Features + +To extract information from the reference image for personalized sampling, we perform a forward pass through the denoiser of the diffusion model at a single timestep. We select $t = 1$ , the least noisy timestep, to obtain clean features that are optimal for conditioning the personalized generation process. Additionally, we remove the text conditioning to decouple visual content of the reference image from the text caption, thus also eliminating the need for user-provided captions for reference images. This contrasts with methods such as JeDi [37], which are sensitive to caption content. During the forward pass, activations are computed for all layers of the denoiser, but only a subset is cached. Based on our experiments with the Stable Diffusion U-Net, we find that caching features from a middle bottleneck layer and every second layer in the decoder offers the best balance between + +generation quality and caching efficiency (see Sec. 4.3). + +Formally, the feature cache $\mathcal{H}_{\mathrm{FC}}$ consists of the activations of the denoiser $\epsilon_{\theta}$ from the selected layers $\mathcal{L}$ at timestep $t = 1$ , using a null text prompt $\varnothing$ and noisy reference image $\mathbf{I}_{\mathrm{ref}} + \mathbf{n}_t$ , with noise realization $\mathbf{n}_t$ , expressed as: + +$$ +\mathcal {H} _ {\mathrm {F C}} = \left\{\mathrm {h} _ {\text {r e f}, L}: L \in \mathcal {L} \right\} \tag {1} +$$ + +$$ +\mathbf {h} _ {\text {r e f}, L} = \epsilon_ {\theta} \left(\mathbf {I} _ {\text {r e f}} + \mathbf {n} _ {t}, \varnothing , t; l) \right| _ {t = 1, l = L}. \tag {2} +$$ + +Notice that the cached features have different spatial resolutions, from the low-resolution bottleneck layer to the higher-resolution decoder layers, allowing a multi-resolution representation of the reference image. This is particularly useful for enabling both global semantics and fine-grained detail guidance. As in prior work [15, 21, 33], we foreground-segment the reference image to isolate the subject from the background before caching its features. + +# 3.2. Conditioning on Cached Reference Features + +We propose a novel conditioning adapter mechanism composed of: i) a cross-attention block between the cached features and the features of the image under generation; ii) a concatenation operation between the features from the self-attention block of the original U-Net denoising backbone and those of the cross-attention block; and iii) a projection layer. A block diagram is shown in Fig. 3 (right). + +Omitting the layer subscript for clarity, the conditioning adapter mechanism is expressed mathematically as follows: + +$$ +\mathbf {q} = \mathbf {W} _ {Q} \mathbf {h}, \quad \mathbf {k} _ {c} = \mathbf {W} _ {K} \mathbf {h} _ {\text {r e f}}, \quad \mathbf {v} _ {c} = \mathbf {W} _ {V} \mathbf {h} _ {\text {r e f}}, \tag {3} +$$ + +$$ +\mathbf {a} _ {c} = \operatorname {s o f t m a x} \left(\frac {\mathbf {q k} _ {c} ^ {T}}{\sqrt {d}}\right) \mathbf {v} _ {c}, \tag {4} +$$ + +$$ +\mathbf {a} = \mathbf {W} _ {\text {p r o j}} \left(\left[ \mathbf {a}; \mathbf {a} _ {c} \right]\right), \tag {5} +$$ + +where $\mathbf{h} \in \mathbb{R}^{N \times d}$ are the current $d$ -dimensional features of the $N$ -pixel image under generation, and $\mathbf{h}_{\mathrm{ref}}$ are the cached reference features from Eq. (2). $\mathbf{W}_Q$ , $\mathbf{W}_K$ , $\mathbf{W}_V$ , and $\mathbf{W}_{\mathrm{proj}}$ are learnable projection matrices, whose training is described in Sec. 3.3. The concatenation $[\mathbf{a}; \mathbf{a}_c] \in \mathbb{R}^{N \times 2d}$ combines the output of self-attention $\mathbf{a} \in \mathbb{R}^{N \times d}$ and cross-attention $\mathbf{a}_c \in \mathbb{R}^{N \times d}$ . The concatenation operation allows a flexible information fusion without explicit alignment constraints, compared to other approaches in similar works (see Sec. 4.3). The learnable projection matrix $\mathbf{W}_{\mathrm{proj}}$ reduces the dimensionality of the concatenated features back to $\mathbb{R}^{N \times d}$ to maintain compatibility with the original backbone. + +Overall, the approach proposed for the adapter enriches feature representations used in the diffusion process of the image under generation by allowing the model to leverage both primary and conditioning-based contextual information from the cache. + +# 3.3. Training the Conditioning Adapters + +The additional parameters introduced in Sec. 3.2 to process the cached features must be trained on a large and varied dataset to ensure they generalize for any reference subject. + +Collecting paired data for this training process would be prohibitively expensive, as it requires multiple images of the same subject in different contexts. To address this, we draw inspiration from the recently proposed synthetic data generation pipeline in BootPIG [21] to construct our training data. First, we utilize a large language model (Llama 3.2 [5]) to generate captions for potential target images. Each caption is used to generate an image via Stable Diffusion [23]. We then use the Segment Anything Model (SAM) [11] and Grounding DINO [14] to accurately segment the reference subject based on the text caption and generate a foreground mask of the main object in the caption. + +We treat the Stable Diffusion-generated image as the target image, the foreground object pasted on a white background as the reference image, and the LLM-generated caption as the textual prompt during our training pipeline. Compared to BootPIG, our pipeline employs open-source models, making it more accessible. We will release our synthetic dataset to facilitate reproducibility and further research, since similar datasets, including BootPIG's [21], have not been publicly released. Additional details on the dataset and its statistics can be found in the Supp. Mat. + +We train the newly introduced adapters' parameters $(\mathbf{W}_Q, \mathbf{W}_K, \mathbf{W}_V, \text{and} \mathbf{W}_{\mathrm{proj}})$ with the standard score matching loss [29] using both the text-conditioned noisy input and the cached reference features: + +$$ +\mathcal {L} _ {\text {d i f f u s i o n}} = \mathbb {E} _ {\mathbf {x} _ {0}, \epsilon , \mathbf {c} _ {T}, \mathbf {I} _ {\text {r e f}}, t} \left[ \| \epsilon - \epsilon_ {\theta} ^ {\prime} \left(\mathbf {x} _ {t}, \mathbf {c} _ {T}, \mathcal {H} _ {\mathrm {F C}}, t\right) \| _ {2} ^ {2} \right], \tag {6} +$$ + +where $\mathbf{x}_0$ is the target image, $\mathbf{c}_T$ is the text prompt generated by the large language model, $\epsilon$ is Gaussian noise, and $t$ is the diffusion timestep sampled uniformly from $1,\dots ,T$ . The noisy image at timestep $t$ , $\mathbf{x}_t$ , is obtained by gradually adding noise to $\mathbf{x}_0$ during the forward diffusion process. The function $\epsilon_{\theta}^{\prime}$ represents the modified denoising model that incorporates the conditioning adapters. + +# 4. Experimental Results + +In this section, we present our experimental results, including quantitative and qualitative comparisons, an ablation study, and an analysis section that visualizes the behavior of the newly introduced cross-attention mechanism in the adapters. + +Implementation Details We evaluate our method on two versions of Stable Diffusion (SD) [23], specifically versions 1.5 and 2.1, to ensure fair comparison with state-of-the-art methods across different backbones. As described in the ablation study, our caching and conditioning mechanism is applied to the middle layer and every second layer of the + +Table 2. Quantitative results on DreamBooth. DreamCache obtains a better balance between DINO score and CLIP-T compared to all baselines, while also offering a more efficient computational tradeoff (see Table 1). + +
MethodBackbone#RefDINO (↑)CLIP-I (↑)CLIP-T (↑)
test-time finetuningDreamBooth [24]Imagen3-50.6960.8120.306
DreamBooth [24]SD 1.53-50.6680.8030.305
Textual Inversion [6]SD 1.53-50.5690.7800.255
Custom Diffusion [12]SD 1.53-50.6430.7900.305
BLIP-Diffusion (FT) [13]SD 1.53-50.6700.8050.302
finetuning freeELITE [33]SD 1.510.6210.7710.293
BLIP-Diffusion [13]SD 1.510.5940.7790.300
IP-Adapter [36]SD 1.510.6670.8130.289
Kosmos-G [19]SD 1.510.6940.8470.287
Jedi [37]SD 1.510.6190.7820.304
DreamCache (ours)SD 1.510.7130.8100.298
Re-Imagen [3]Imagen1-30.6000.7400.270
SuTI [4]Imagen1-30.7410.8190.304
Subject-Diffusion [15]SD 2.110.7710.7790.293
BootPig [21]SD 2.130.6740.7970.311
ToffeeNet [39]SD 2.110.7280.8170.306
CAFE [38]SD 2.110.7150.8270.294
DreamCache (ours)SD 2.110.7670.8160.301
+ +decoder. The total number of trainable parameters for DreamCache is 25M. We use the original SD codebase and train the model on $4 \times 80\mathrm{GB}$ A100 GPUs for 25k steps with a batch size of 128, using the AdamW optimizer with a learning rate of $10^{-5}$ . Input images are resized to $512 \times 512$ , with scale, shift, and resize augmentations applied to reference images to enhance model robustness to perturbations. Ablations are conducted on SD 1.5. We generate images with 50 sampling steps, employing classifier-free guidance for image and text conditioning, using a guidance scale of 7.5. + +Evaluation Quantitative evaluations are conducted on the DreamBooth dataset [24], following prior approaches. DreamBooth consists of 30 subjects, each with 25 text prompts. We use a single input image per subject and generate 4 images per subject-prompt combination, resulting in 3,000 generated images. We use pretrained DINO ViT-S/16 and CLIP ViT-B/32 models to calculate the average cosine similarity of global image embeddings between generated and reference images, with metrics denoted as DINO and CLIP-I, respectively. To assess text alignment, we calculate the cosine similarity between embeddings from generated images and text prompts using CLIP's image and text encoders [9], with the corresponding score denoted as CLIP-T. + +# 4.1. Zero-Shot Personalization + +We compare DreamCache against state-of-the-art methods for finetuning-based and zero-shot personalization. Table 2 presents quantitative results, indicating the diffusion back + +bone and the number of reference images for each method. Our approach achieves competitive or superior performance compared to other computationally-intensive state-of-the-art methods, which are trained on larger datasets and with significantly more parameters. We refer the reader to Table 1 for the data requirements, training time, and parameter count of the various methods. We remark that generally DINO is a preferred metric for image similarity with respect to CLIP-I, as it is more sensitive to the appearance and fine-grained details of the subjects. + +We also present qualitative comparisons with Kosmos-G [19] and BLIP-Diffusion [13]. We remark that several other methods are not reproducible due to the lack of code, datasets, or trained checkpoints. As seen in Fig. 4, our method excels in subject preservation and textual alignment, producing visually superior results. We also notice that Kosmos-G reports a high CLIP-I score, but after inspecting the generated images, it is clear that the score does not entirely reflect the preservation of the reference subject in generated images. In fact, Kosmos-G presents a high degree of background interference, where the partial replication of the reference background boosts the alignment score. For this reason, we also report foreground-masked metrics on the subjects, like MCLIP-I and MDINO, in the Supp. Mat. + +# 4.2. Inference Time Evaluation + +The computational efficiency of our method is compared to the reference-based method BootPIG [21] and encoder-based approaches such as Kosmos-G [19] and Subject + +![](images/be29f76c7d4ede8064e27ff6766020940a0dab6031ab99231271016101e5e9bb.jpg) +"A dog" + +![](images/82158a728d40bcdcdae325a08b9c4c5fe2f83fac63acfe6e09d69d7d5cade368.jpg) +"wearing a Santa hat" + +![](images/ebb431d605f744ceba52ff9c87c1fd831101d5964b8fe8a94d2a1213ebfc960e.jpg) + +![](images/0364081437e14de23d98385fd1a18defe231a5d0b002fa36707819391bd60757.jpg) + +![](images/ced4cdce5491e89288ba7db0672f261bef4d985f78cd71337119c205cc195a7a.jpg) +"with the Eiffel Tower in the background" + +![](images/a5b006650f369bfea8d691cc7ef75697044ec7a13823eadf9ce38f8b55a71f67.jpg) + +![](images/e16785f60eae1f9af2092ee8fa023a215bc9e4ea4bca676702e36cafae9ed04f.jpg) + +![](images/2507e224b7535a30b0c5d314074d3ebb7113126ee2677b8a2f15bad3d167e2fe.jpg) +"A can" + +![](images/96dc5471e153bef8c50cff411a269645b40cb182a8103679ba5204127d1f51ca.jpg) +"floating on top of water" + +![](images/a9c98199e7c76e5c44e19e87094fb5d7a876d874100fa4a0b1ea4f51d7a70810.jpg) + +![](images/eabc7d6929f8ca80374bd68054aaae9c72a1218e62a43e5faa83a80aa9c0d98a.jpg) + +![](images/8c97f3fa5025ee58cc1860117f3a538453dbfbfe57950ba3c1611db97a727ea9.jpg) + +![](images/5894285b043526c36b61b0ed750d470915bdddb7307918789218439515c7a83c.jpg) + +![](images/6a02427b8d3f7feb761e3e47665eaf364a38b8df2c53f230257eab10a96c8f75.jpg) + +![](images/96160c2ac6ad20efef76172f424b0d6b532a8e97c03f7b2a71539b3518f85516.jpg) +"A toy" +Reference +Figure 4. Visual comparison. Personalized generations on sample concepts. DreamCache preserves reference concept appearance and does not suffer from background interference. BLIP-D [13] and Kosmos-G [19] cannot faithfully preserve visual details from the reference. + +![](images/4a3ed0a138fb260e32faf41fd198449b4fc83a1804119156e3cb4cde4e750018.jpg) +BLIP-D + +![](images/dc4e18dcb953af79fe1e5ddb1f9a4844b854051d85e344a7410a2e28448e84f5.jpg) +"on the beach" +Kosmos-G + +![](images/1a327d304d745d5ced93dd3492b05fb1a99ea13a8dc777607d2957c6a641507e.jpg) +DreamCache + +![](images/2dc8043099673fd79307756f9fcbe82bbfc3c1fb2d46fca4bdec5a801817a14b.jpg) +BLIP-D + +![](images/3208ce0f09f90f0927ec307c2ddb01da153c1ffd72eb7ef6af8cf8a16870360a.jpg) +Kosmos-G + +![](images/9c0c65754d7fddb6c2b4a3c01bb56a035ab730190c3e7132aa3155679e98f25f.jpg) +DreamCache + +Table 3. Computational comparison.*: time to generate an image with 100 timesteps, evaluated on a single NVIDIA A100 GPU. + +
MethodInference Time*Extra Params Size
ELITE [33]6.24 s914 MB
BLIP-Diffusion [13]3.92 s760 MB
BootPig [21]7.55 s1900 MB
DreamCache (ours)3.88 s42 MB
+ +Diffusion [15]. Table 3 provides a detailed comparison of inference time, accounting for both personalization time (e.g., the time to generate the cache for DreamCache) and the time to sample the personalized image. We also report the increase in model size, i.e., the storage (in FP16 precision) required for the extra parameters to allow for personalization, showing that DreamCache is one order of magnitude more compact than the state of the art. Overall, DreamCache offers a lightweight solution that achieves state-of-the-art performance with faster inference and reduced computational overhead. + +# 4.3. Ablation Studies + +We validate our design choices through a series of studies, examining different conditioning mechanisms, evaluating our feature caching approach, and analyzing the impact of synthetic dataset scaling. + +Table 4. Reference feature integration. DreamCache uses the best tradeoff between accuracy and complexity. + +
MethodCLIP-I (↑)CLIP-T (↑)Params
Textual Sum [36]0.7880,28219M
Spatial Sum0.8120.29316M
Decoupled Blocks [8]0.8080.30061M
Spatial Concat (ours)0.8100.29825M
+ +Reference Feature Integration We compare various conditioning strategies to integrate reference features in Table 4. Our spatial cross-attention block with concatenation between the output of self- and cross-attention ("Spatial Concat") was evaluated against different alternatives, including IP-Adapter's conditioning mechanism [36] ("Textual Sum"), which sums the decoupled cross-attention output with that from textual cross-attention. We also tested a variant ("Spatial Sum") where self- and cross-attention conditioning outputs are summed. Additionally, we also assessed an alternative conditioning procedure inspired by ViCo [8] ("Decoupled Blocks"), involving independent and interleaved cross-attention blocks. Results in Table 4 indicate that the proposed "Spatial Concat" offers the best balance of text alignment and parameter efficiency. + +We further explored optimal conditioning insertion within the U-Net backbone in Table 5, determining that applying + +Table 5. Cache positioning in the U-Net backbone offers a further tradeoff between accuracy and complexity. + +
EncoderMiddleDecoderCLIP-I (↑)CLIP-T (↑)Params
✓XX0.7210.30311M
✓✓X0.7490.30619M
X✓X0.7160.3028M
XX✓0.7990.29617M
X✓✓0.8100.29825M
✓✓✓0.8130.29736M
+ +Table 6. Caching with text is not influential and adds complexity. + +
Text-FreeCLIP-I (↑)CLIP-T (↑)
×0.8110.295
✓0.8100.298
+ +Table 7. Dataset impact for both synthetic and real data. + +
DatasetCLIP-I (↑)CLIP-T (↑)
Synthetic-50K0.7810.304
Synthetic-200K0.7970,301
Synthetic-400K0.8100.298
LAION-5M0.8140.242
+ +conditioning (and therefore feature caching) at the middle layer and every second layer of the decoder achieved the best tradeoff between performance and parameter count. + +Text Input for Cached Features Our feature caching procedure is designed to be text-free, leveraging the classifier-free guidance used during pretraining where captions were occasionally omitted. We compare this approach with a version including textual inputs during caching (e.g., 'A photo of...'). Table 6 shows that adding text conditioning slightly reduces text alignment while increasing complexity and potentially introducing noise in cases of inaccurate captions. + +Dataset Impact We demonstrate the importance of our synthetic dataset to train the conditioning adapters and the effect of scaling its size. For this purpose, we created synthetic datasets according to the procedure in Sec. 3.3 of sizes 50K, 200K, and 400K samples. We also tested the real-world 5M samples from the LAION [26] dataset, which, lacking target-caption-reference triplets, required reuse of target images as reference images too. Table 7 shows that increasing dataset size improves image alignment, though slightly reduces textual alignment. Notably, LAION improves image alignment but struggles with textual alignment. This highlights the importance of triplet data (target image, reference image, and caption) for effective zero-shot personalization, ensuring both subject preservation and textual editability. + +![](images/8f6e6e1d1e4b6a2758dcd4909a5593d41acc79da2a621ce9c161ce2cef600754.jpg) +Figure 5. Visualization of reference image impact. Cross-attention maps between cached reference features and features of the image under generation. Left: attention map at layers at $16 \times 16$ resolution (left reference, right generated). Right: $32 \times 32$ . Attention values are highly localized in the region of interest. + +# 4.4. Visualizing Reference Impact + +Finally, we analyze how the cross-attention mechanism in DreamCache impacts image generation by visualizing cached reference feature influence. Attention map visualizations at different resolutions are provided in Fig. 5. Specifically, attention maps between the query from the current generation and the key derived from reference features reveal a highly localized focus on the subject, without interference from background elements. This mechanism models correspondences effectively, integrating reference information into the generated image. + +# 5. Discussion and Conclusions + +In this paper, we proposed DreamCache, a novel approach to personalized text-to-image generation that uses feature caching to overcome the limitations of existing methods. By caching reference features from a small subset of layers of the U-Net only once, our method significantly reduces both computational and memory demands, enabling efficient, real-time personalized image generation. Unlike previous approaches, DreamCache avoids the need for costly finetuning, external image encoders, or parallel reference processing, making it lightweight and suitable for plug-and-play deployment. Our experiments demonstrate that DreamCache achieves state-of-the-art zero-shot personalization with only 25M additional parameters and a fast training process. While DreamCache is a promising direction towards more efficient personalized generation, it has some limitations. Although effective for single-subject personalization, our approach may require adaptation for complex multi-subject generation where feature interference can occur. Additionally, certain edge cases, such as highly abstract or stylistic images, may challenge the caching mechanism's capacity to accurately preserve subject details. To address these challenges, future work may explore adaptive caching techniques or multi-reference feature integration. + +# References + +[1] Moab Arar, Rinon Gal, Yuval Atzmon, Gal Chechik, Daniel Cohen-Or, Ariel Shamir, and Amit H. Bermano. Domain-agnostic tuning-encoder for fast personalization of text-to-image models. In SIGGRAPH Asia 2023 Conference Papers, pages 1–10, 2023. 1, 3 +[2] Mathilde Caron, Hugo Touvron, Ishan Misra, Hervé Jégou, Julien Mairal, Piotr Bojanowski, and Armand Joulin. Emerging properties in self-supervised vision transformers. In Proceedings of the IEEE/CVF international conference on computer vision, pages 9650-9660, 2021. 1 +[3] Wenhu Chen, Hexiang Hu, Chitwan Sahara, and William W Cohen. Re-imagen: Retrieval-augmented text-to-image generator. In The Eleventh International Conference on Learning Representations, 2022. 6 +[4] Wenhu Chen, Hexiang Hu, Yandong Li, Nataniel Rui, Xuhui Jia, Ming-Wei Chang, and William W Cohen. Subject-driven text-to-image generation via apprenticeship learning. arXiv preprint arXiv:2304.00186, 2023. 2, 3, 6 +[5] Abhimanyu Dubey, Abhinav Jauhri, Abhinav Pandey, Abhishek Kadian, Ahmad Al-Dahle, Aiesha Letman, Akhil Mathur, Alan Schelten, Amy Yang, Angela Fan, et al. The llama 3 herd of models. arXiv preprint arXiv:2407.21783, 2024. 5 +[6] Rinon Gal, Yuval Alaluf, Yuval Atzmon, Or Patashnik, Amit Haim Bermano, Gal Chechik, and Daniel Cohen-or. An image is worth one word: Personalizing text-to-image generation using textual inversion. In The Eleventh International Conference on Learning Representations, 2022. 1, 2, 3, 6 +[7] Rinon Gal, Moab Arar, Yuval Atzmon, Amit H Bermano, Gal Chechik, and Daniel Cohen-Or. Encoder-based domain tuning for fast personalization of text-to-image models. ACM Transactions on Graphics (TOG), 42(4):1-13, 2023. 1, 3 +[8] Shaozhe Hao, Kai Han, Shihao Zhao, and Kwan-Yee K Wong. Vico: Plug-and-play visual condition for personalized text-to-image generation. arXiv preprint arXiv:2306.00971, 2023. 1, 3, 7 +[9] Jack Hessel, Ari Holtzman, Maxwell Forbes, Ronan Le Bras, and Yejin Choi. Clipscore: A reference-free evaluation metric for image captioning. In Proceedings of the 2021 Conference on Empirical Methods in Natural Language Processing, pages 7514-7528, 2021. 6 +[10] Jonathan Ho, Ajay Jain, and Pieter Abbeel. Denoising diffusion probabilistic models. Advances in neural information processing systems, 33:6840-6851, 2020. 1 +[11] Alexander Kirillov, Eric Mintun, Nikhila Ravi, Hanzi Mao, Chloe Rolland, Laura Gustafson, Tete Xiao, Spencer Whitehead, Alexander C Berg, Wan-Yen Lo, et al. Segment anything. arXiv preprint arXiv:2304.02643, 2023. 5 +[12] Nupur Kumari, Bingliang Zhang, Richard Zhang, Eli Shechtman, and Jun-Yan Zhu. Multi-concept customization of text-to-image diffusion. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 1931-1941, 2023. 1, 2, 6 +[13] Dongxu Li, Junnan Li, and Steven CH Hoi. Blip-diffusion: Pre-trained subject representation for control + +lable text-to-image generation and editing. arXiv preprint arXiv:2305.14720, 2023. 1, 2, 3, 6, 7 +[14] Shilong Liu, Zhaoyang Zeng, Tianhe Ren, Feng Li, Hao Zhang, Jie Yang, Chunyuan Li, Jianwei Yang, Hang Su, Jun Zhu, et al. Grounding dino: Marrying dino with grounded pre-training for open-set object detection. arXiv preprint arXiv:2303.05499, 2023. 5 +[15] Jian Ma, Junhao Liang, Chen Chen, and Haonan Lu. Subject-diffusion: Open domain personalized text-to-image generation without test-time fine-tuning. arXiv preprint arXiv:2307.11410, 2023. 1, 2, 3, 5, 6, 7 +[16] Xinyin Ma, Gongfan Fang, Michael Bi Mi, and Xinchao Wang. Learning-to-cache: Accelerating diffusion transformer via layer caching. arXiv preprint arXiv:2406.01733, 2024. 4 +[17] Xinyin Ma, Gongfan Fang, and Xinchao Wang. Deepcache: Accelerating diffusion models for free. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 15762-15772, 2024. 3 +[18] Boris Oreshkin, Pau Rodríguez López, and Alexandre Lacoste. Tadam: Task dependent adaptive metric for improved few-shot learning. Advances in neural information processing systems, 31, 2018. 4 +[19] Xichen Pan, Li Dong, Shaohan Huang, Zhiliang Peng, Wenhu Chen, and Furu Wei. Kosmos-g: Generating images in context with multimodal large language models. arXiv preprint arXiv:2310.02992, 2023. 1, 2, 3, 6, 7 +[20] Maitreya Patel, Sangmin Jung, Chitta Baral, and Yezhou Yang. λ-eclipse: Multi-concept personalized text-to-image diffusion models by leveraging clip latent space. arXiv preprint arXiv:2402.05195, 2024. 1, 3 +[21] Senthil Purushwalkam, Akash Gokul, Shafiq Joty, and Nikhil Naik. Bootpig: Bootstrapping zero-shot personalized image generation capabilities in pretrained diffusion models. arXiv preprint arXiv:2401.13974, 2024. 1, 2, 3, 5, 6, 7 +[22] Alec Radford, Jong Wook Kim, Chris Hallacy, Aditya Ramesh, Gabriel Goh, Sandhini Agarwal, Girish Sastry, Amanda Askell, Pamela Mishkin, Jack Clark, et al. Learning transferable visual models from natural language supervision. In International conference on machine learning, pages 8748-8763. PMLR, 2021. 1 +[23] Robin Rombach, Andreas Blattmann, Dominik Lorenz, Patrick Esser, and Björn Ommer. High-resolution image synthesis with latent diffusion models. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 10684-10695, 2022. 1, 5 +[24] Nataniel Ruiz, Yuanzhen Li, Varun Jampani, Yael Pritch, Michael Rubinstein, and Kfir Aberman. Dreambooth: Fine tuning text-to-image diffusion models for subject-driven generation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 22500-22510, 2023. 1, 2, 6 +[25] Chitwan Sahara, William Chan, Saurabh Saxena, Lala Li, Jay Whang, Emily L Denton, Kamyar Ghasemipour, Raphael Gontijo Lopes, Burcu Karagol Ayan, Tim Salimans, et al. Photorealistic text-to-image diffusion models with deep language understanding. Advances in Neural Information Processing Systems, 35:36479-36494, 2022. 1 + +[26] Christoph Schuhmann, Romain Beaumont, Richard Vencu, Cade Gordon, Ross Wightman, Mehdi Cherti, Theo Coombes, Aarush Katta, Clayton Mullis, Mitchell Wortsman, et al. Laion-5b: An open large-scale dataset for training next generation image-text models. Advances in Neural Information Processing Systems, 35:25278–25294, 2022. 8 +[27] Jake Snell, Kevin Swersky, and Richard Zemel. Prototypical networks for few-shot learning. Advances in neural information processing systems, 30, 2017. 4 +[28] Jascha Sohl-Dickstein, Eric Weiss, Niru Maheswaranathan, and Surya Ganguli. Deep unsupervised learning using nonequilibrium thermodynamics. In International conference on machine learning, pages 2256-2265. PMLR, 2015. 1 +[29] Yang Song and Stefano Ermon. Generative modeling by estimating gradients of the data distribution. Advances in neural information processing systems, 32, 2019. 5 +[30] Yoad Tewel, Rinon Gal, Gal Chechik, and Yuval Atzmon. Key-locked rank one editing for text-to-image personalization. In ACM SIGGRAPH 2023 Conference Proceedings, pages 1-11, 2023. 1, 3 +[31] Eleni Triantafillou, Hugo Larochelle, Jake Snell, Josh Tenenbaum, Kevin Jordan Swersky, Mengye Ren, Richard Zemel, and Sachin Ravi. Meta-learning for semi-supervised few-shot classification. In International Conference on Learning Representations, 2018. 4 +[32] Andrey Voynov, Qinghao Chu, Daniel Cohen-Or, and Kfir Aberman. p+: Extended textual conditioning in text-to-image generation. arXiv preprint arXiv:2303.09522, 2023. 1, 3 +[33] Yuxiang Wei, Yabo Zhang, Zhilong Ji, Jinfeng Bai, Lei Zhang, and Wangmeng Zuo. Elite: Encoding visual concepts into textual embeddings for customized text-to-image generation. 2023 IEEE/CVF International Conference on Computer Vision (ICCV), pages 15897-15907, 2023. 1, 2, 3, 5, 6, 7 +[34] Felix Wimbauer, Bichen Wu, Edgar Schoenfeld, Xiaoliang Dai, Ji Hou, Zijian He, Artsiom Sanakoyeu, Peizhao Zhang, Sam Tsai, Jonas Kohler, et al. Cache me if you can: Accelerating diffusion models through block caching. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 6211-6220, 2024. 3 +[35] Jianan Yang, Haobo Wang, Yanming Zhang, Ruixuan Xiao, Sai Wu, Gang Chen, and Junbo Zhao. Controllable textual inversion for personalized text-to-image generation. arXiv preprint arXiv:2304.05265, 2023. 1, 3 +[36] Hu Ye, Jun Zhang, Sibo Liu, Xiao Han, and Wei Yang. Ip-adapter: Text compatible image prompt adapter for text-to-image diffusion models. arXiv preprint arXiv:2308.06721, 2023. 1, 2, 3, 6, 7 +[37] Yu Zeng, Vishal M Patel, Haochen Wang, Xun Huang, Ting-Chun Wang, Ming-Yu Liu, and Yogesh Balaji. Jedi: Joint-image diffusion models for finetuning-free personalized text-to-image generation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 6786-6795, 2024. 1, 2, 3, 4, 6 +[38] Yufan Zhou, Ruiyi Zhang, Jiumiang Gu, and Tong Sun. Customization assistant for text-to-image generation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 9182-9191, 2024. 2, 3, 6 + +[39] Yufan Zhou, Ruiyi Zhang, Kaizhi Zheng, Nanxuan Zhao, Jiumiang Gu, Zichao Wang, Xin Eric Wang, and Tong Sun. Toffee: Efficient million-scale dataset construction for subject-driven text-to-image generation. arXiv preprint arXiv:2406.09305, 2024.1,2,3,6 \ No newline at end of file diff --git a/2025/DreamCache_ Finetuning-Free Lightweight Personalized Image Generation via Feature Caching/images.zip b/2025/DreamCache_ Finetuning-Free Lightweight Personalized Image Generation via Feature Caching/images.zip new file mode 100644 index 0000000000000000000000000000000000000000..911fc1810ab115c8ba2ae18be42de131ddcb787f --- /dev/null +++ b/2025/DreamCache_ Finetuning-Free Lightweight Personalized Image Generation via Feature Caching/images.zip @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:8e343ce8e6a4cc1fe699e97d8021dbcd241efa48c184b570fc350c3195023d39 +size 631912 diff --git a/2025/DreamCache_ Finetuning-Free Lightweight Personalized Image Generation via Feature Caching/layout.json b/2025/DreamCache_ Finetuning-Free Lightweight Personalized Image Generation via Feature Caching/layout.json new file mode 100644 index 0000000000000000000000000000000000000000..8cf1acb42b87b1391875d499f320294c230f0bdc --- /dev/null +++ b/2025/DreamCache_ Finetuning-Free Lightweight Personalized Image Generation via Feature Caching/layout.json @@ -0,0 +1,8500 @@ +{ + "pdf_info": [ + { + "para_blocks": [ + { + "bbox": [ + 151, + 117, + 460, + 152 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 151, + 117, + 460, + 152 + ], + "spans": [ + { + "bbox": [ + 151, + 117, + 460, + 152 + ], + "type": "text", + "content": "DreamCache: Finetuning-Free Lightweight Personalized Image Generation via Feature Caching" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 89, + 171, + 520, + 200 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 89, + 171, + 520, + 200 + ], + "spans": [ + { + "bbox": [ + 89, + 171, + 520, + 200 + ], + "type": "text", + "content": "Emanuele Aiello Umberto Michieli Diego Valsesia Mete Ozay Enrico Magli Politecnico di Torino Samsung R&D Institute UK" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 159, + 224, + 204, + 237 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 159, + 224, + 204, + 237 + ], + "spans": [ + { + "bbox": [ + 159, + 224, + 204, + 237 + ], + "type": "text", + "content": "Abstract" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 66, + 251, + 297, + 433 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 66, + 251, + 297, + 433 + ], + "spans": [ + { + "bbox": [ + 66, + 251, + 297, + 433 + ], + "type": "text", + "content": "Personalized image generation requires text-to-image generative models that capture the core features of a reference subject to allow for controlled generation across different contexts. Existing methods face challenges due to complex training requirements, high inference costs, limited flexibility, or a combination of these issues. In this paper, we introduce DreamCache, a scalable approach for efficient and high-quality personalized image generation. By caching a small number of reference image features from a subset of layers and a single timestep of the pretrained diffusion denoiser, DreamCache enables dynamic modulation of the generated image features through lightweight, trained conditioning adapters. DreamCache achieves state-of-the-art image and text alignment, utilizing an order of magnitude fewer extra parameters, and is both more computationally effective and versatile than existing models." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 67, + 449, + 144, + 460 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 67, + 449, + 144, + 460 + ], + "spans": [ + { + "bbox": [ + 67, + 449, + 144, + 460 + ], + "type": "text", + "content": "1. Introduction" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 66, + 464, + 296, + 533 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 66, + 464, + 296, + 533 + ], + "spans": [ + { + "bbox": [ + 66, + 464, + 296, + 533 + ], + "type": "text", + "content": "Recent advancements in text-to-image generation, fueled by the development of diffusion models [10, 28], have enabled high-quality and diverse image generation from textual descriptions. Diffusion models [23, 25] gradually transform random noise into images through a sequence of denoising steps, conditioned on the input text prompt." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 66, + 533, + 297, + 671 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 66, + 533, + 297, + 671 + ], + "spans": [ + { + "bbox": [ + 66, + 533, + 297, + 671 + ], + "type": "text", + "content": "An active area of research is personalizing these models, enabling the generation of novel images of a reference subject in various contexts, while maintaining flexibility for text-based editing. Early personalization techniques [1, 6, 8, 24, 30, 32, 35, 35], such as the seminal Dream-Booth [24] relied on fine-tuning (FT) the generative model for each reference subject. However, these approaches are often impractical for many use cases due to costly test-time FT, which can take several minutes per subject. To address this, FT-free (i.e., zero-shot) personalized image generation methods have emerged to eliminate test-time optimization. These FT-free approaches can be broadly categorized into" + } + ] + } + ], + "index": 8 + }, + { + "type": "image", + "bbox": [ + 356, + 226, + 501, + 370 + ], + "blocks": [ + { + "bbox": [ + 356, + 226, + 501, + 370 + ], + "lines": [ + { + "bbox": [ + 356, + 226, + 501, + 370 + ], + "spans": [ + { + "bbox": [ + 356, + 226, + 501, + 370 + ], + "type": "image", + "image_path": "2ff59b4bdbab438456d466a87c6e85d3cb828d258db8bfe36cf3490f74168362.jpg" + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 313, + 380, + 544, + 413 + ], + "lines": [ + { + "bbox": [ + 313, + 380, + 544, + 413 + ], + "spans": [ + { + "bbox": [ + 313, + 380, + 544, + 413 + ], + "type": "text", + "content": "Figure 1. DreamCache is a finetuning-free personalized image generation method that achieves an optimal balance between subject fidelity, memory efficiency, and adherence to text prompts." + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "image_caption" + } + ], + "index": 9 + }, + { + "bbox": [ + 313, + 434, + 542, + 456 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 434, + 542, + 456 + ], + "spans": [ + { + "bbox": [ + 313, + 434, + 542, + 456 + ], + "type": "text", + "content": "two families: encoder-based methods and reference-based methods, each with distinct drawbacks." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 312, + 457, + 544, + 525 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 312, + 457, + 544, + 525 + ], + "spans": [ + { + "bbox": [ + 312, + 457, + 544, + 525 + ], + "type": "text", + "content": "Encoder-based methods [7, 15, 20, 33, 36] utilize dedicated image encoders, such as CLIP [22] or DINO [2], to extract relevant features from reference images. While these encoders can produce high-quality results, they are often large, require extensive training to align text and image features, and reduce the model's flexibility [13, 15, 19, 36]." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 312, + 527, + 544, + 629 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 312, + 527, + 544, + 629 + ], + "spans": [ + { + "bbox": [ + 312, + 527, + 544, + 629 + ], + "type": "text", + "content": "In contrast, reference-based methods [21, 37] condition the diffusion model directly on reference features drawn from the U-Net denoiser, integrating these features at each denoising step. While effective, these methods require feature extraction at every generation step, leading to higher computational costs and memory demands. Additionally, they often require an input textual caption for conditioning, which introduces variability and can decrease output precision." + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 312, + 630, + 544, + 699 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 312, + 630, + 544, + 699 + ], + "spans": [ + { + "bbox": [ + 312, + 630, + 544, + 699 + ], + "type": "text", + "content": "Some recent works have proposed to finetune the U-Net backbone itself [21, 37, 39]. However, this hinders the model's ability to switch between personalized and nonpersonalized tasks and risks inducing the \"language drift\" phenomenon, where the personalization training degrades the model's linguistic comprehension [12, 24]." + } + ] + } + ], + "index": 14 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 71, + 13, + 116, + 43 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 71, + 13, + 116, + 43 + ], + "spans": [ + { + "bbox": [ + 71, + 13, + 116, + 43 + ], + "type": "text", + "content": "CVF" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 149, + 11, + 489, + 45 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 149, + 11, + 489, + 45 + ], + "spans": [ + { + "bbox": [ + 149, + 11, + 489, + 45 + ], + "type": "text", + "content": "This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 67, + 679, + 297, + 698 + ], + "type": "page_footnote", + "angle": 0, + "lines": [ + { + "bbox": [ + 67, + 679, + 297, + 698 + ], + "spans": [ + { + "bbox": [ + 67, + 679, + 297, + 698 + ], + "type": "inline_equation", + "content": "^{1}" + }, + { + "bbox": [ + 67, + 679, + 297, + 698 + ], + "type": "text", + "content": "Code, model, and synthetic dataset can be found at https://github.com/Emanuele97x/DreamCache" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 294, + 731, + 317, + 740 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 731, + 317, + 740 + ], + "spans": [ + { + "bbox": [ + 294, + 731, + 317, + 740 + ], + "type": "text", + "content": "12480" + } + ] + } + ], + "index": 16 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 0 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 75, + 115, + 535, + 304 + ], + "blocks": [ + { + "bbox": [ + 67, + 86, + 543, + 106 + ], + "lines": [ + { + "bbox": [ + 67, + 86, + 543, + 106 + ], + "spans": [ + { + "bbox": [ + 67, + 86, + 543, + 106 + ], + "type": "text", + "content": "Table 1. Methods overview. Our DreamCache achieves state-of-the-art generation quality at reduced computational costs. *: value refers to the personalization stage for each personal subject." + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 75, + 115, + 535, + 304 + ], + "lines": [ + { + "bbox": [ + 75, + 115, + 535, + 304 + ], + "spans": [ + { + "bbox": [ + 75, + 115, + 535, + 304 + ], + "type": "table", + "html": "
MethodFT-freeEnc-freePlug&PlayRef-UNet-freeExtra ParamsTrain Params# DatasetTrain Time
Textual Inversion [6]X✓✓✓768*768*3-5*50 min*
DreamBooth [24]X✓X✓-0.9B*3-5*10 min*
Custom Diffusion [12]X✓X✓-57M*3-5*10 min*
ELITE [33]✓XX✓457M77M125K14 days
BLIP-Diffusion [13]✓X✓✓380M1.5B129M96 days
IP-Adapter [36]✓X✓✓402M22M10M28 days
Kosmos-G [19]✓X✓✓1.6B1.6B9M-
JeDi [37]✓✓XX-0.9B3M48 days
SuTI [4]✓XX✓400M2.5B500K-
Subject-Diffusion [15]✓X✓X700M700M76M-
BootPig [21]✓✓XX0.95B0.95B200K18 hours
ToffeeNet [39]✓XX✓632M0.9B5M-
CAFE [38]✓XX✓14B1B355K-
DreamCache (ours)✓✓✓✓25M25M400K40 hours
", + "image_path": "6e415f08761cafc5535bfeb4985c595b0b5ee0160e328b3c0a90082a61e0db16.jpg" + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "table_body" + } + ], + "index": 1 + }, + { + "bbox": [ + 69, + 323, + 296, + 517 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 69, + 323, + 296, + 517 + ], + "spans": [ + { + "bbox": [ + 69, + 323, + 296, + 517 + ], + "type": "text", + "content": "In this work, we propose DreamCache, a novel finetuning-free approach to personalized image generation, that overcomes the limitations of existing methods (see Fig. 1) by using a feature caching mechanism that enables text-free encoding and efficient conditioning during personalization. Specifically, we first create a synthetic dataset [21] containing triplets of captions, target images, and reference subjects to capture subjects in various contexts. Next, we pretrain lightweight attention-based conditioning adapters to inject subject-specific features into the image generation process. During personalization, the reference image is processed through the pretrained denoiser of the base diffusion model without text conditioning, thus eliminating the need for user-generated captions, while caching features from a small subset of layers at a single timestep. For personalized sampling, these cached features are injected into the denoiser through the pretrained conditioning adapters." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 69, + 517, + 296, + 689 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 69, + 517, + 296, + 689 + ], + "spans": [ + { + "bbox": [ + 69, + 517, + 296, + 689 + ], + "type": "text", + "content": "Table 1 summarizes the key properties of existing methods and illustrates how DreamCache fits within the current landscape; further details are explored in Sec. 2. As an encoder-free approach, DreamCache introduces only a small number of additional parameters, making it significantly lightweight and suitable for deployment on resource-constrained devices. For example, methods like [33] and [13] introduce 380M parameters due to their reliance on CLIP encoders, whereas DreamCache requires only 25M additional parameters. Moreover, caching features from a few selected U-Net layers at a single preprocessing timestep bypasses the need for full U-Net reference processing during generation, leading to substantial computational and memory savings that enable real-time, high-quality personalized generation. Another key advantage of DreamCache is its plug-and-play" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 315, + 323, + 542, + 380 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 315, + 323, + 542, + 380 + ], + "spans": [ + { + "bbox": [ + 315, + 323, + 542, + 380 + ], + "type": "text", + "content": "nature, allowing concurrent generation of personalized and non-personalized content without altering original U-Net weights, thus preserving the integrity of the original model and enabling a wider range of deployment scenarios, especially on mobile platforms." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 315, + 381, + 542, + 415 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 315, + 381, + 542, + 415 + ], + "spans": [ + { + "bbox": [ + 315, + 381, + 542, + 415 + ], + "type": "text", + "content": "In summary, DreamCache represents a significant step toward practical, and scalable personalized image generation, with the following contributions:" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 313, + 418, + 542, + 530 + ], + "type": "list", + "angle": 0, + "index": 9, + "blocks": [ + { + "bbox": [ + 313, + 418, + 542, + 450 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 418, + 542, + 450 + ], + "spans": [ + { + "bbox": [ + 313, + 418, + 542, + 450 + ], + "type": "text", + "content": "- We propose a feature caching approach that creates multi-resolution representations of the reference image in a caption-free and efficient manner." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 313, + 452, + 542, + 496 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 452, + 542, + 496 + ], + "spans": [ + { + "bbox": [ + 313, + 452, + 542, + 496 + ], + "type": "text", + "content": "- We design an attention-based conditioning mechanism that leverages the cached features for personalized image generation, achieving computational- and memory- efficient personalized sampling." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 313, + 498, + 542, + 530 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 498, + 542, + 530 + ], + "spans": [ + { + "bbox": [ + 313, + 498, + 542, + 530 + ], + "type": "text", + "content": "- Our approach achieves state-of-the-art quality in personalized image generation at substantially lower computational and data costs compared to existing methods." + } + ] + } + ], + "index": 8 + } + ], + "sub_type": "text" + }, + { + "bbox": [ + 315, + 544, + 480, + 555 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 315, + 544, + 480, + 555 + ], + "spans": [ + { + "bbox": [ + 315, + 544, + 480, + 555 + ], + "type": "text", + "content": "2. Background and Related Work" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 315, + 563, + 542, + 642 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 315, + 563, + 542, + 642 + ], + "spans": [ + { + "bbox": [ + 315, + 563, + 542, + 642 + ], + "type": "text", + "content": "Personalized image generation aims at generating images containing a specific subject. This task has been widely studied, resulting in two main approaches: fine-tuning methods, which require test-time finetuning on multiple subject reference images, and finetuning-free (zero-shot) methods, which learn a generalizable conditioning mechanism to generate reference subjects without the need for further optimization." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 315, + 654, + 542, + 697 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 315, + 654, + 542, + 697 + ], + "spans": [ + { + "bbox": [ + 315, + 654, + 542, + 697 + ], + "type": "text", + "content": "Finetuning-based Personalization DreamBooth [24] finetunes the entire U-Net with reference images while introducing a regularization loss to mitigate overfitting. On the other hand, Custom Diffusion [12] only finetunes the " + }, + { + "bbox": [ + 315, + 654, + 542, + 697 + ], + "type": "inline_equation", + "content": "K" + }, + { + "bbox": [ + 315, + 654, + 542, + 697 + ], + "type": "text", + "content": " and" + } + ] + } + ], + "index": 12 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 732, + 316, + 740 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 732, + 316, + 740 + ], + "spans": [ + { + "bbox": [ + 294, + 732, + 316, + 740 + ], + "type": "text", + "content": "12481" + } + ] + } + ], + "index": 13 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 1 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 70, + 97, + 162, + 188 + ], + "blocks": [ + { + "bbox": [ + 90, + 86, + 143, + 96 + ], + "lines": [ + { + "bbox": [ + 90, + 86, + 143, + 96 + ], + "spans": [ + { + "bbox": [ + 90, + 86, + 143, + 96 + ], + "type": "text", + "content": "\"A dragon...\"" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 70, + 97, + 162, + 188 + ], + "lines": [ + { + "bbox": [ + 70, + 97, + 162, + 188 + ], + "spans": [ + { + "bbox": [ + 70, + 97, + 162, + 188 + ], + "type": "image", + "image_path": "72e59097773136a3ca128e6e35d632f15e0757d7e32b35195cf460eb0822589d.jpg" + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_body" + } + ], + "index": 1 + }, + { + "type": "image", + "bbox": [ + 163, + 99, + 254, + 188 + ], + "blocks": [ + { + "bbox": [ + 173, + 86, + 244, + 97 + ], + "lines": [ + { + "bbox": [ + 173, + 86, + 244, + 97 + ], + "spans": [ + { + "bbox": [ + 173, + 86, + 244, + 97 + ], + "type": "text", + "content": "\"as street graffiti\"" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 163, + 99, + 254, + 188 + ], + "lines": [ + { + "bbox": [ + 163, + 99, + 254, + 188 + ], + "spans": [ + { + "bbox": [ + 163, + 99, + 254, + 188 + ], + "type": "image", + "image_path": "7063e425d55245aac90538df6ccdc4c59e0aa1dc7fdda3987894df887d553467.jpg" + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_body" + } + ], + "index": 4 + }, + { + "type": "image", + "bbox": [ + 256, + 97, + 347, + 188 + ], + "blocks": [ + { + "bbox": [ + 265, + 86, + 339, + 97 + ], + "lines": [ + { + "bbox": [ + 265, + 86, + 339, + 97 + ], + "spans": [ + { + "bbox": [ + 265, + 86, + 339, + 97 + ], + "type": "text", + "content": "\"playing with fire\"" + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 256, + 97, + 347, + 188 + ], + "lines": [ + { + "bbox": [ + 256, + 97, + 347, + 188 + ], + "spans": [ + { + "bbox": [ + 256, + 97, + 347, + 188 + ], + "type": "image", + "image_path": "a8d50836b16284dc7b021527485e2a63d2f65a1c505a1c98940637cd5f9f63ac.jpg" + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "image_body" + } + ], + "index": 7 + }, + { + "type": "image", + "bbox": [ + 349, + 98, + 440, + 188 + ], + "blocks": [ + { + "bbox": [ + 366, + 86, + 423, + 97 + ], + "lines": [ + { + "bbox": [ + 366, + 86, + 423, + 97 + ], + "spans": [ + { + "bbox": [ + 366, + 86, + 423, + 97 + ], + "type": "text", + "content": "\"as a plushie\"" + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 349, + 98, + 440, + 188 + ], + "lines": [ + { + "bbox": [ + 349, + 98, + 440, + 188 + ], + "spans": [ + { + "bbox": [ + 349, + 98, + 440, + 188 + ], + "type": "image", + "image_path": "3288fe34e86e7c2c5690b5f415cf6ce77dd50e11d50a4ce0fc0845a094bada51.jpg" + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "image_body" + } + ], + "index": 10 + }, + { + "type": "image", + "bbox": [ + 444, + 97, + 536, + 188 + ], + "blocks": [ + { + "bbox": [ + 445, + 86, + 536, + 96 + ], + "lines": [ + { + "bbox": [ + 445, + 86, + 536, + 96 + ], + "spans": [ + { + "bbox": [ + 445, + 86, + 536, + 96 + ], + "type": "text", + "content": "\"working as a barista\"" + } + ] + } + ], + "index": 12, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 444, + 97, + 536, + 188 + ], + "lines": [ + { + "bbox": [ + 444, + 97, + 536, + 188 + ], + "spans": [ + { + "bbox": [ + 444, + 97, + 536, + 188 + ], + "type": "image", + "image_path": "42cba27336adc0a3a020f273eee03a02b1786e4b1912d7211f0edb071baf6f56.jpg" + } + ] + } + ], + "index": 13, + "angle": 0, + "type": "image_body" + } + ], + "index": 13 + }, + { + "type": "image", + "bbox": [ + 70, + 201, + 161, + 292 + ], + "blocks": [ + { + "bbox": [ + 98, + 190, + 135, + 200 + ], + "lines": [ + { + "bbox": [ + 98, + 190, + 135, + 200 + ], + "spans": [ + { + "bbox": [ + 98, + 190, + 135, + 200 + ], + "type": "text", + "content": "\"A cat...\"" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 70, + 201, + 161, + 292 + ], + "lines": [ + { + "bbox": [ + 70, + 201, + 161, + 292 + ], + "spans": [ + { + "bbox": [ + 70, + 201, + 161, + 292 + ], + "type": "image", + "image_path": "b8fd579d3cf15ec40ec046824de141036ba2da37a0d3e1c0c12a7a18972c5a68.jpg" + } + ] + } + ], + "index": 15, + "angle": 0, + "type": "image_body" + } + ], + "index": 15 + }, + { + "type": "image", + "bbox": [ + 162, + 202, + 254, + 291 + ], + "blocks": [ + { + "bbox": [ + 173, + 190, + 244, + 201 + ], + "lines": [ + { + "bbox": [ + 173, + 190, + 244, + 201 + ], + "spans": [ + { + "bbox": [ + 173, + 190, + 244, + 201 + ], + "type": "text", + "content": "\"in Ukiyo-e style\"" + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 162, + 202, + 254, + 291 + ], + "lines": [ + { + "bbox": [ + 162, + 202, + 254, + 291 + ], + "spans": [ + { + "bbox": [ + 162, + 202, + 254, + 291 + ], + "type": "image", + "image_path": "00db6d3885758325f7772d3f86d0e88bf4b11f9b17c537c25770b5e27d7f2d12.jpg" + } + ] + } + ], + "index": 16, + "angle": 0, + "type": "image_body" + } + ], + "index": 16 + }, + { + "type": "image", + "bbox": [ + 255, + 201, + 348, + 293 + ], + "blocks": [ + { + "bbox": [ + 256, + 190, + 347, + 201 + ], + "lines": [ + { + "bbox": [ + 256, + 190, + 347, + 201 + ], + "spans": [ + { + "bbox": [ + 256, + 190, + 347, + 201 + ], + "type": "text", + "content": "\"with a rainbow scarf\"" + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 255, + 201, + 348, + 293 + ], + "lines": [ + { + "bbox": [ + 255, + 201, + 348, + 293 + ], + "spans": [ + { + "bbox": [ + 255, + 201, + 348, + 293 + ], + "type": "image", + "image_path": "957f74af1e220a4a34c67d7f3ce19eda399f8a1118a1aa1270ce965557b246c9.jpg" + } + ] + } + ], + "index": 17, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 66, + 304, + 543, + 326 + ], + "lines": [ + { + "bbox": [ + 66, + 304, + 543, + 326 + ], + "spans": [ + { + "bbox": [ + 66, + 304, + 543, + 326 + ], + "type": "text", + "content": "Figure 2. Personalized generations by DreamCache. The first column contains reference images. The generated images correspond to the text prompts above each column." + } + ] + } + ], + "index": 20, + "angle": 0, + "type": "image_caption" + } + ], + "index": 17 + }, + { + "type": "image", + "bbox": [ + 349, + 201, + 440, + 293 + ], + "blocks": [ + { + "bbox": [ + 353, + 190, + 436, + 201 + ], + "lines": [ + { + "bbox": [ + 353, + 190, + 436, + 201 + ], + "spans": [ + { + "bbox": [ + 353, + 190, + 436, + 201 + ], + "type": "text", + "content": "\"Van Gogh painting\"" + } + ] + } + ], + "index": 11, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 349, + 201, + 440, + 293 + ], + "lines": [ + { + "bbox": [ + 349, + 201, + 440, + 293 + ], + "spans": [ + { + "bbox": [ + 349, + 201, + 440, + 293 + ], + "type": "image", + "image_path": "969b036c282fc912b595c03d260274a3a16a411b9f73873cf36260c6eae53b59.jpg" + } + ] + } + ], + "index": 18, + "angle": 0, + "type": "image_body" + } + ], + "index": 18 + }, + { + "type": "image", + "bbox": [ + 444, + 201, + 536, + 293 + ], + "blocks": [ + { + "bbox": [ + 443, + 190, + 539, + 201 + ], + "lines": [ + { + "bbox": [ + 443, + 190, + 539, + 201 + ], + "spans": [ + { + "bbox": [ + 443, + 190, + 539, + 201 + ], + "type": "text", + "content": "\"wearing a diploma hat\"" + } + ] + } + ], + "index": 14, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 444, + 201, + 536, + 293 + ], + "lines": [ + { + "bbox": [ + 444, + 201, + 536, + 293 + ], + "spans": [ + { + "bbox": [ + 444, + 201, + 536, + 293 + ], + "type": "image", + "image_path": "5a3086329e7c27ee0b1766b5132b3ec444206b07702c156a24d8bbb408726be4.jpg" + } + ] + } + ], + "index": 19, + "angle": 0, + "type": "image_body" + } + ], + "index": 19 + }, + { + "bbox": [ + 66, + 346, + 297, + 449 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 66, + 346, + 297, + 449 + ], + "spans": [ + { + "bbox": [ + 66, + 346, + 297, + 449 + ], + "type": "inline_equation", + "content": "V" + }, + { + "bbox": [ + 66, + 346, + 297, + 449 + ], + "type": "text", + "content": " projections for the cross-attention blocks of the U-Net. Text-based personalization methods optimize single (like Textual Inversion [6]) or multiple (like in " + }, + { + "bbox": [ + 66, + 346, + 297, + 449 + ], + "type": "inline_equation", + "content": "\\mathrm{P + }" + }, + { + "bbox": [ + 66, + 346, + 297, + 449 + ], + "type": "text", + "content": " [32]) input token embeddings. Later methods [1, 8, 30, 35] build on these, with innovations like Perfusion [30] using dynamic rank-1 updates to prevent overfitting while keeping encodings lightweight. However, all finetuning-based methods are computationally intensive, often requiring minutes of finetuning per reference subject at test time." + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 66, + 459, + 297, + 652 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 66, + 459, + 297, + 652 + ], + "spans": [ + { + "bbox": [ + 66, + 459, + 297, + 652 + ], + "type": "text", + "content": "Finetuning-Free Personalization To reduce computational demands, recent research has shifted toward zero-shot personalization methods that eliminate subject-specific finetuning, typically employing image encoders to condition the generation process via the features extracted from the reference images [7, 15, 20, 33, 36]. Examples include BLIP-Diffusion [13], which pretrains a Q-Former to learn image features aligned with text, and IP-Adapter, which uses a frozen CLIP encoder to extract text-aligned visual features that modulate the cross-attention layers of the generative model. Other approaches, like Kosmos-G [19] and CAFE [38], connect large language models (LLMs) with diffusion models to condition generation on personalized concepts. SuTI [4] takes a different approach by training millions of subject-specific experts and subsequently training a model via apprenticeship learning, enabling effective zero-shot personalized generation at test time." + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 66, + 653, + 296, + 699 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 66, + 653, + 296, + 699 + ], + "spans": [ + { + "bbox": [ + 66, + 653, + 296, + 699 + ], + "type": "text", + "content": "Alternatively, encoder-free methods such as JeDi [37] and BootPIG [21] use features from the generative model's backbone to guide the generation. JeDi creates a multi-view synthetic dataset and modifies spatial self-attention to jointly" + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 313, + 346, + 544, + 437 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 346, + 544, + 437 + ], + "spans": [ + { + "bbox": [ + 313, + 346, + 544, + 437 + ], + "type": "text", + "content": "attend to images of the same concept in a batch. BootPIG retains a trainable copy of the original U-Net, adding reference self-attention layers to enable adaptation of the personalized model for reference features. While these methods remove the need for an additional encoder, they still require computationally expensive inference, as reference images must be processed in parallel during generation—a cost that accumulates due to the iterative nature of the diffusion process." + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 313, + 437, + 544, + 517 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 437, + 544, + 517 + ], + "spans": [ + { + "bbox": [ + 313, + 437, + 544, + 517 + ], + "type": "text", + "content": "In contrast, our method, DreamCache, caches a subset of reference features from the U-Net without text conditioning, eliminating the need for parallel inference and reducing memory overhead by avoiding separate model loading. This results in a test-time computational efficiency similar to encoder-based methods while offering the flexibility of encoder-free feature extraction and injection." + } + ] + } + ], + "index": 25 + }, + { + "bbox": [ + 313, + 517, + 544, + 585 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 517, + 544, + 585 + ], + "spans": [ + { + "bbox": [ + 313, + 517, + 544, + 585 + ], + "type": "text", + "content": "Recent works such as BootPIG [21] and Toffee-5M [39] emphasize the importance of synthetic data that explicitly decouples the subject from the background, reporting improved performance. Inspired by these approaches, we adopt a similar generative pipeline to create a synthetic dataset we use to train DreamCache." + } + ] + } + ], + "index": 26 + }, + { + "bbox": [ + 313, + 586, + 543, + 643 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 586, + 543, + 643 + ], + "spans": [ + { + "bbox": [ + 313, + 586, + 543, + 643 + ], + "type": "text", + "content": "Moreover, since our method is plug-and-play and keeps the U-Net frozen, we reduce the high cost of training faced by other approaches [15, 19, 21, 37-39]. A detailed overview of current methods, the number of trained parameters, and their training cost can be found in Table 1." + } + ] + } + ], + "index": 27 + }, + { + "bbox": [ + 313, + 653, + 544, + 699 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 653, + 544, + 699 + ], + "spans": [ + { + "bbox": [ + 313, + 653, + 544, + 699 + ], + "type": "text", + "content": "Feature Caching Feature caching has been explored to reduce generation time in diffusion models by caching intermediate activations. Some studies [17, 34] exploit temporal redundancy during the training process to cache ac" + } + ] + } + ], + "index": 28 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 731, + 317, + 740 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 731, + 317, + 740 + ], + "spans": [ + { + "bbox": [ + 294, + 731, + 317, + 740 + ], + "type": "text", + "content": "12482" + } + ] + } + ], + "index": 29 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 2 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 82, + 92, + 533, + 273 + ], + "blocks": [ + { + "bbox": [ + 82, + 92, + 533, + 273 + ], + "lines": [ + { + "bbox": [ + 82, + 92, + 533, + 273 + ], + "spans": [ + { + "bbox": [ + 82, + 92, + 533, + 273 + ], + "type": "image", + "image_path": "cc90f61ffcc2fdd5a17e9ca033fa3489b1b277e98fc320d9267a8e3a9e803009.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 66, + 287, + 544, + 331 + ], + "lines": [ + { + "bbox": [ + 66, + 287, + 544, + 331 + ], + "spans": [ + { + "bbox": [ + 66, + 287, + 544, + 331 + ], + "type": "text", + "content": "Figure 3. Overview of DreamCache. Original U-Net layers are shown in violet, while the novel components introduced by DreamCache are highlighted in green. During personalization, features from selected layers of the diffusion denoiser are cached from a single timestep, using a null text prompt. These cached features serve as reference-specific information. During generation, conditioning adapters inject the cached features into the denoiser, modulating the features of the generated image to create a personalized output." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "bbox": [ + 66, + 350, + 297, + 498 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 66, + 350, + 297, + 498 + ], + "spans": [ + { + "bbox": [ + 66, + 350, + 297, + 498 + ], + "type": "text", + "content": "tivations across timesteps, reducing computational load at later timesteps. Other works focus on caching layer activations within the diffusion framework, avoiding redundant computations. Learning-to-Cache [16] introduces a dynamic caching mechanism that learns to skip computation for selected layers of the diffusion model. In contrast to those works, which generally cache intra-model features for some layers to save computations, we utilize feature caching to encode multi-resolution features of a reference image from a few selected layers to condition the generation process of a new personalized image. Our approach recalls successful few-shot learners for discriminative problems [18, 27, 31] and extends them to personalized image generation." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 67, + 508, + 119, + 521 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 67, + 508, + 119, + 521 + ], + "spans": [ + { + "bbox": [ + 67, + 508, + 119, + 521 + ], + "type": "text", + "content": "3. Method" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 66, + 527, + 297, + 652 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 66, + 527, + 297, + 652 + ], + "spans": [ + { + "bbox": [ + 66, + 527, + 297, + 652 + ], + "type": "text", + "content": "Given a pretrained text-to-image generative model " + }, + { + "bbox": [ + 66, + 527, + 297, + 652 + ], + "type": "inline_equation", + "content": "\\epsilon_{\\theta}" + }, + { + "bbox": [ + 66, + 527, + 297, + 652 + ], + "type": "text", + "content": " and an image containing a reference subject " + }, + { + "bbox": [ + 66, + 527, + 297, + 652 + ], + "type": "inline_equation", + "content": "\\mathbf{I}_{\\mathrm{ref}}" + }, + { + "bbox": [ + 66, + 527, + 297, + 652 + ], + "type": "text", + "content": ", the goal of personalized sampling is to generate novel images containing the reference subject in various contexts while maintaining textual control. We propose DreamCache, a novel approach for extracting conditioning signals from " + }, + { + "bbox": [ + 66, + 527, + 297, + 652 + ], + "type": "inline_equation", + "content": "\\mathbf{I}_{\\mathrm{ref}}" + }, + { + "bbox": [ + 66, + 527, + 297, + 652 + ], + "type": "text", + "content": " and guiding of the image generation process. This method leverages a pretrained diffusion model, conditioning adapters that are pretrained with a synthetic dataset, and a feature cache from the reference image. Sample outputs generated by DreamCache are shown in Fig. 2, with a method overview in Fig. 3." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 66, + 653, + 297, + 699 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 66, + 653, + 297, + 699 + ], + "spans": [ + { + "bbox": [ + 66, + 653, + 297, + 699 + ], + "type": "text", + "content": "At the core of DreamCache, we utilize the denoiser in the pretrained diffusion model to extract multi-resolution features from " + }, + { + "bbox": [ + 66, + 653, + 297, + 699 + ], + "type": "inline_equation", + "content": "\\mathbf{I}_{\\mathrm{ref}}" + }, + { + "bbox": [ + 66, + 653, + 297, + 699 + ], + "type": "text", + "content": " by caching the activations of a few selected layers. To improve generalization, we cache features" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 313, + 350, + 544, + 441 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 350, + 544, + 441 + ], + "spans": [ + { + "bbox": [ + 313, + 350, + 544, + 441 + ], + "type": "text", + "content": "using a forward pass with a null text prompt. When personalized sampling is performed, the cached features are processed by adapters to act as conditioning signals, modulating the denoiser features of the image under generation at corresponding layers. These adapters, once pretrained on a synthetic dataset, enable zero-shot personalized generation with any new reference image, requiring no further finetuning once its features are cached." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 313, + 443, + 544, + 500 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 443, + 544, + 500 + ], + "spans": [ + { + "bbox": [ + 313, + 443, + 544, + 500 + ], + "type": "text", + "content": "In the following, we detail the three main aspects of DreamCache, namely i) how to cache reference features (Sec. 3.1); ii) how to condition the diffusion model on the cached features for personalized sampling (Sec. 3.2); iii) how to train the adapters used for model conditioning (Sec. 3.3)." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 313, + 508, + 462, + 521 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 508, + 462, + 521 + ], + "spans": [ + { + "bbox": [ + 313, + 508, + 462, + 521 + ], + "type": "text", + "content": "3.1. Caching Reference Features" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 312, + 528, + 544, + 699 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 312, + 528, + 544, + 699 + ], + "spans": [ + { + "bbox": [ + 312, + 528, + 544, + 699 + ], + "type": "text", + "content": "To extract information from the reference image for personalized sampling, we perform a forward pass through the denoiser of the diffusion model at a single timestep. We select " + }, + { + "bbox": [ + 312, + 528, + 544, + 699 + ], + "type": "inline_equation", + "content": "t = 1" + }, + { + "bbox": [ + 312, + 528, + 544, + 699 + ], + "type": "text", + "content": ", the least noisy timestep, to obtain clean features that are optimal for conditioning the personalized generation process. Additionally, we remove the text conditioning to decouple visual content of the reference image from the text caption, thus also eliminating the need for user-provided captions for reference images. This contrasts with methods such as JeDi [37], which are sensitive to caption content. During the forward pass, activations are computed for all layers of the denoiser, but only a subset is cached. Based on our experiments with the Stable Diffusion U-Net, we find that caching features from a middle bottleneck layer and every second layer in the decoder offers the best balance between" + } + ] + } + ], + "index": 9 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 731, + 317, + 740 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 731, + 317, + 740 + ], + "spans": [ + { + "bbox": [ + 294, + 731, + 317, + 740 + ], + "type": "text", + "content": "12483" + } + ] + } + ], + "index": 10 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 3 + }, + { + "para_blocks": [ + { + "bbox": [ + 67, + 87, + 282, + 99 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 67, + 87, + 282, + 99 + ], + "spans": [ + { + "bbox": [ + 67, + 87, + 282, + 99 + ], + "type": "text", + "content": "generation quality and caching efficiency (see Sec. 4.3)." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 67, + 99, + 296, + 145 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 67, + 99, + 296, + 145 + ], + "spans": [ + { + "bbox": [ + 67, + 99, + 296, + 145 + ], + "type": "text", + "content": "Formally, the feature cache " + }, + { + "bbox": [ + 67, + 99, + 296, + 145 + ], + "type": "inline_equation", + "content": "\\mathcal{H}_{\\mathrm{FC}}" + }, + { + "bbox": [ + 67, + 99, + 296, + 145 + ], + "type": "text", + "content": " consists of the activations of the denoiser " + }, + { + "bbox": [ + 67, + 99, + 296, + 145 + ], + "type": "inline_equation", + "content": "\\epsilon_{\\theta}" + }, + { + "bbox": [ + 67, + 99, + 296, + 145 + ], + "type": "text", + "content": " from the selected layers " + }, + { + "bbox": [ + 67, + 99, + 296, + 145 + ], + "type": "inline_equation", + "content": "\\mathcal{L}" + }, + { + "bbox": [ + 67, + 99, + 296, + 145 + ], + "type": "text", + "content": " at timestep " + }, + { + "bbox": [ + 67, + 99, + 296, + 145 + ], + "type": "inline_equation", + "content": "t = 1" + }, + { + "bbox": [ + 67, + 99, + 296, + 145 + ], + "type": "text", + "content": ", using a null text prompt " + }, + { + "bbox": [ + 67, + 99, + 296, + 145 + ], + "type": "inline_equation", + "content": "\\varnothing" + }, + { + "bbox": [ + 67, + 99, + 296, + 145 + ], + "type": "text", + "content": " and noisy reference image " + }, + { + "bbox": [ + 67, + 99, + 296, + 145 + ], + "type": "inline_equation", + "content": "\\mathbf{I}_{\\mathrm{ref}} + \\mathbf{n}_t" + }, + { + "bbox": [ + 67, + 99, + 296, + 145 + ], + "type": "text", + "content": ", with noise realization " + }, + { + "bbox": [ + 67, + 99, + 296, + 145 + ], + "type": "inline_equation", + "content": "\\mathbf{n}_t" + }, + { + "bbox": [ + 67, + 99, + 296, + 145 + ], + "type": "text", + "content": ", expressed as:" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 113, + 156, + 296, + 168 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 113, + 156, + 296, + 168 + ], + "spans": [ + { + "bbox": [ + 113, + 156, + 296, + 168 + ], + "type": "interline_equation", + "content": "\\mathcal {H} _ {\\mathrm {F C}} = \\left\\{\\mathrm {h} _ {\\text {r e f}, L}: L \\in \\mathcal {L} \\right\\} \\tag {1}", + "image_path": "ad0c7da64790cb535d1f3a52858affe490d75604289803b38860428d9ebed236.jpg" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 108, + 170, + 296, + 182 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 108, + 170, + 296, + 182 + ], + "spans": [ + { + "bbox": [ + 108, + 170, + 296, + 182 + ], + "type": "interline_equation", + "content": "\\mathbf {h} _ {\\text {r e f}, L} = \\epsilon_ {\\theta} \\left(\\mathbf {I} _ {\\text {r e f}} + \\mathbf {n} _ {t}, \\varnothing , t; l) \\right| _ {t = 1, l = L}. \\tag {2}", + "image_path": "02201a6f900ca5bf686c3254a113654114a614e2dfece55bd1a4fc74516801fe.jpg" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 66, + 194, + 297, + 285 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 66, + 194, + 297, + 285 + ], + "spans": [ + { + "bbox": [ + 66, + 194, + 297, + 285 + ], + "type": "text", + "content": "Notice that the cached features have different spatial resolutions, from the low-resolution bottleneck layer to the higher-resolution decoder layers, allowing a multi-resolution representation of the reference image. This is particularly useful for enabling both global semantics and fine-grained detail guidance. As in prior work [15, 21, 33], we foreground-segment the reference image to isolate the subject from the background before caching its features." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 67, + 294, + 286, + 306 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 67, + 294, + 286, + 306 + ], + "spans": [ + { + "bbox": [ + 67, + 294, + 286, + 306 + ], + "type": "text", + "content": "3.2. Conditioning on Cached Reference Features" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 66, + 312, + 296, + 392 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 66, + 312, + 296, + 392 + ], + "spans": [ + { + "bbox": [ + 66, + 312, + 296, + 392 + ], + "type": "text", + "content": "We propose a novel conditioning adapter mechanism composed of: i) a cross-attention block between the cached features and the features of the image under generation; ii) a concatenation operation between the features from the self-attention block of the original U-Net denoising backbone and those of the cross-attention block; and iii) a projection layer. A block diagram is shown in Fig. 3 (right)." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 67, + 393, + 296, + 416 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 67, + 393, + 296, + 416 + ], + "spans": [ + { + "bbox": [ + 67, + 393, + 296, + 416 + ], + "type": "text", + "content": "Omitting the layer subscript for clarity, the conditioning adapter mechanism is expressed mathematically as follows:" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 91, + 438, + 296, + 450 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 91, + 438, + 296, + 450 + ], + "spans": [ + { + "bbox": [ + 91, + 438, + 296, + 450 + ], + "type": "interline_equation", + "content": "\\mathbf {q} = \\mathbf {W} _ {Q} \\mathbf {h}, \\quad \\mathbf {k} _ {c} = \\mathbf {W} _ {K} \\mathbf {h} _ {\\text {r e f}}, \\quad \\mathbf {v} _ {c} = \\mathbf {W} _ {V} \\mathbf {h} _ {\\text {r e f}}, \\tag {3}", + "image_path": "ca7a3b69c32b6aa2e3333c22ca2d775bca8742fa7a6b61d3a3ab6dbd58a69a0a.jpg" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 90, + 453, + 296, + 478 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 90, + 453, + 296, + 478 + ], + "spans": [ + { + "bbox": [ + 90, + 453, + 296, + 478 + ], + "type": "interline_equation", + "content": "\\mathbf {a} _ {c} = \\operatorname {s o f t m a x} \\left(\\frac {\\mathbf {q k} _ {c} ^ {T}}{\\sqrt {d}}\\right) \\mathbf {v} _ {c}, \\tag {4}", + "image_path": "8d9df13ad86ae3905016e038c7776c6cae4c7bfbcee9d64cdb54f86cc9a0ff65.jpg" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 90, + 481, + 296, + 493 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 90, + 481, + 296, + 493 + ], + "spans": [ + { + "bbox": [ + 90, + 481, + 296, + 493 + ], + "type": "interline_equation", + "content": "\\mathbf {a} = \\mathbf {W} _ {\\text {p r o j}} \\left(\\left[ \\mathbf {a}; \\mathbf {a} _ {c} \\right]\\right), \\tag {5}", + "image_path": "af21cdb14691a4d65204fbb4e7c8af3f8bb8f853c75506b36309b102c0d13c80.jpg" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 66, + 503, + 297, + 640 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 66, + 503, + 297, + 640 + ], + "spans": [ + { + "bbox": [ + 66, + 503, + 297, + 640 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 66, + 503, + 297, + 640 + ], + "type": "inline_equation", + "content": "\\mathbf{h} \\in \\mathbb{R}^{N \\times d}" + }, + { + "bbox": [ + 66, + 503, + 297, + 640 + ], + "type": "text", + "content": " are the current " + }, + { + "bbox": [ + 66, + 503, + 297, + 640 + ], + "type": "inline_equation", + "content": "d" + }, + { + "bbox": [ + 66, + 503, + 297, + 640 + ], + "type": "text", + "content": "-dimensional features of the " + }, + { + "bbox": [ + 66, + 503, + 297, + 640 + ], + "type": "inline_equation", + "content": "N" + }, + { + "bbox": [ + 66, + 503, + 297, + 640 + ], + "type": "text", + "content": "-pixel image under generation, and " + }, + { + "bbox": [ + 66, + 503, + 297, + 640 + ], + "type": "inline_equation", + "content": "\\mathbf{h}_{\\mathrm{ref}}" + }, + { + "bbox": [ + 66, + 503, + 297, + 640 + ], + "type": "text", + "content": " are the cached reference features from Eq. (2). " + }, + { + "bbox": [ + 66, + 503, + 297, + 640 + ], + "type": "inline_equation", + "content": "\\mathbf{W}_Q" + }, + { + "bbox": [ + 66, + 503, + 297, + 640 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 66, + 503, + 297, + 640 + ], + "type": "inline_equation", + "content": "\\mathbf{W}_K" + }, + { + "bbox": [ + 66, + 503, + 297, + 640 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 66, + 503, + 297, + 640 + ], + "type": "inline_equation", + "content": "\\mathbf{W}_V" + }, + { + "bbox": [ + 66, + 503, + 297, + 640 + ], + "type": "text", + "content": ", and " + }, + { + "bbox": [ + 66, + 503, + 297, + 640 + ], + "type": "inline_equation", + "content": "\\mathbf{W}_{\\mathrm{proj}}" + }, + { + "bbox": [ + 66, + 503, + 297, + 640 + ], + "type": "text", + "content": " are learnable projection matrices, whose training is described in Sec. 3.3. The concatenation " + }, + { + "bbox": [ + 66, + 503, + 297, + 640 + ], + "type": "inline_equation", + "content": "[\\mathbf{a}; \\mathbf{a}_c] \\in \\mathbb{R}^{N \\times 2d}" + }, + { + "bbox": [ + 66, + 503, + 297, + 640 + ], + "type": "text", + "content": " combines the output of self-attention " + }, + { + "bbox": [ + 66, + 503, + 297, + 640 + ], + "type": "inline_equation", + "content": "\\mathbf{a} \\in \\mathbb{R}^{N \\times d}" + }, + { + "bbox": [ + 66, + 503, + 297, + 640 + ], + "type": "text", + "content": " and cross-attention " + }, + { + "bbox": [ + 66, + 503, + 297, + 640 + ], + "type": "inline_equation", + "content": "\\mathbf{a}_c \\in \\mathbb{R}^{N \\times d}" + }, + { + "bbox": [ + 66, + 503, + 297, + 640 + ], + "type": "text", + "content": ". The concatenation operation allows a flexible information fusion without explicit alignment constraints, compared to other approaches in similar works (see Sec. 4.3). The learnable projection matrix " + }, + { + "bbox": [ + 66, + 503, + 297, + 640 + ], + "type": "inline_equation", + "content": "\\mathbf{W}_{\\mathrm{proj}}" + }, + { + "bbox": [ + 66, + 503, + 297, + 640 + ], + "type": "text", + "content": " reduces the dimensionality of the concatenated features back to " + }, + { + "bbox": [ + 66, + 503, + 297, + 640 + ], + "type": "inline_equation", + "content": "\\mathbb{R}^{N \\times d}" + }, + { + "bbox": [ + 66, + 503, + 297, + 640 + ], + "type": "text", + "content": " to maintain compatibility with the original backbone." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 66, + 642, + 296, + 698 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 66, + 642, + 296, + 698 + ], + "spans": [ + { + "bbox": [ + 66, + 642, + 296, + 698 + ], + "type": "text", + "content": "Overall, the approach proposed for the adapter enriches feature representations used in the diffusion process of the image under generation by allowing the model to leverage both primary and conditioning-based contextual information from the cache." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 313, + 87, + 495, + 99 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 87, + 495, + 99 + ], + "spans": [ + { + "bbox": [ + 313, + 87, + 495, + 99 + ], + "type": "text", + "content": "3.3. Training the Conditioning Adapters" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 313, + 104, + 542, + 139 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 104, + 542, + 139 + ], + "spans": [ + { + "bbox": [ + 313, + 104, + 542, + 139 + ], + "type": "text", + "content": "The additional parameters introduced in Sec. 3.2 to process the cached features must be trained on a large and varied dataset to ensure they generalize for any reference subject." + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 313, + 139, + 543, + 275 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 139, + 543, + 275 + ], + "spans": [ + { + "bbox": [ + 313, + 139, + 543, + 275 + ], + "type": "text", + "content": "Collecting paired data for this training process would be prohibitively expensive, as it requires multiple images of the same subject in different contexts. To address this, we draw inspiration from the recently proposed synthetic data generation pipeline in BootPIG [21] to construct our training data. First, we utilize a large language model (Llama 3.2 [5]) to generate captions for potential target images. Each caption is used to generate an image via Stable Diffusion [23]. We then use the Segment Anything Model (SAM) [11] and Grounding DINO [14] to accurately segment the reference subject based on the text caption and generate a foreground mask of the main object in the caption." + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 313, + 275, + 544, + 389 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 275, + 544, + 389 + ], + "spans": [ + { + "bbox": [ + 313, + 275, + 544, + 389 + ], + "type": "text", + "content": "We treat the Stable Diffusion-generated image as the target image, the foreground object pasted on a white background as the reference image, and the LLM-generated caption as the textual prompt during our training pipeline. Compared to BootPIG, our pipeline employs open-source models, making it more accessible. We will release our synthetic dataset to facilitate reproducibility and further research, since similar datasets, including BootPIG's [21], have not been publicly released. Additional details on the dataset and its statistics can be found in the Supp. Mat." + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 313, + 389, + 544, + 434 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 389, + 544, + 434 + ], + "spans": [ + { + "bbox": [ + 313, + 389, + 544, + 434 + ], + "type": "text", + "content": "We train the newly introduced adapters' parameters " + }, + { + "bbox": [ + 313, + 389, + 544, + 434 + ], + "type": "inline_equation", + "content": "(\\mathbf{W}_Q, \\mathbf{W}_K, \\mathbf{W}_V, \\text{and} \\mathbf{W}_{\\mathrm{proj}})" + }, + { + "bbox": [ + 313, + 389, + 544, + 434 + ], + "type": "text", + "content": " with the standard score matching loss [29] using both the text-conditioned noisy input and the cached reference features:" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 320, + 443, + 543, + 458 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 320, + 443, + 543, + 458 + ], + "spans": [ + { + "bbox": [ + 320, + 443, + 543, + 458 + ], + "type": "interline_equation", + "content": "\\mathcal {L} _ {\\text {d i f f u s i o n}} = \\mathbb {E} _ {\\mathbf {x} _ {0}, \\epsilon , \\mathbf {c} _ {T}, \\mathbf {I} _ {\\text {r e f}}, t} \\left[ \\| \\epsilon - \\epsilon_ {\\theta} ^ {\\prime} \\left(\\mathbf {x} _ {t}, \\mathbf {c} _ {T}, \\mathcal {H} _ {\\mathrm {F C}}, t\\right) \\| _ {2} ^ {2} \\right], \\tag {6}", + "image_path": "b015e9cffbeab39bded98154026eb7c0b82e73df0bf0def6346dcfff876b599c.jpg" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 313, + 466, + 544, + 546 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 466, + 544, + 546 + ], + "spans": [ + { + "bbox": [ + 313, + 466, + 544, + 546 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 313, + 466, + 544, + 546 + ], + "type": "inline_equation", + "content": "\\mathbf{x}_0" + }, + { + "bbox": [ + 313, + 466, + 544, + 546 + ], + "type": "text", + "content": " is the target image, " + }, + { + "bbox": [ + 313, + 466, + 544, + 546 + ], + "type": "inline_equation", + "content": "\\mathbf{c}_T" + }, + { + "bbox": [ + 313, + 466, + 544, + 546 + ], + "type": "text", + "content": " is the text prompt generated by the large language model, " + }, + { + "bbox": [ + 313, + 466, + 544, + 546 + ], + "type": "inline_equation", + "content": "\\epsilon" + }, + { + "bbox": [ + 313, + 466, + 544, + 546 + ], + "type": "text", + "content": " is Gaussian noise, and " + }, + { + "bbox": [ + 313, + 466, + 544, + 546 + ], + "type": "inline_equation", + "content": "t" + }, + { + "bbox": [ + 313, + 466, + 544, + 546 + ], + "type": "text", + "content": " is the diffusion timestep sampled uniformly from " + }, + { + "bbox": [ + 313, + 466, + 544, + 546 + ], + "type": "inline_equation", + "content": "1,\\dots ,T" + }, + { + "bbox": [ + 313, + 466, + 544, + 546 + ], + "type": "text", + "content": ". The noisy image at timestep " + }, + { + "bbox": [ + 313, + 466, + 544, + 546 + ], + "type": "inline_equation", + "content": "t" + }, + { + "bbox": [ + 313, + 466, + 544, + 546 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 313, + 466, + 544, + 546 + ], + "type": "inline_equation", + "content": "\\mathbf{x}_t" + }, + { + "bbox": [ + 313, + 466, + 544, + 546 + ], + "type": "text", + "content": ", is obtained by gradually adding noise to " + }, + { + "bbox": [ + 313, + 466, + 544, + 546 + ], + "type": "inline_equation", + "content": "\\mathbf{x}_0" + }, + { + "bbox": [ + 313, + 466, + 544, + 546 + ], + "type": "text", + "content": " during the forward diffusion process. The function " + }, + { + "bbox": [ + 313, + 466, + 544, + 546 + ], + "type": "inline_equation", + "content": "\\epsilon_{\\theta}^{\\prime}" + }, + { + "bbox": [ + 313, + 466, + 544, + 546 + ], + "type": "text", + "content": " represents the modified denoising model that incorporates the conditioning adapters." + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 313, + 555, + 433, + 568 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 555, + 433, + 568 + ], + "spans": [ + { + "bbox": [ + 313, + 555, + 433, + 568 + ], + "type": "text", + "content": "4. Experimental Results" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 313, + 575, + 543, + 621 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 575, + 543, + 621 + ], + "spans": [ + { + "bbox": [ + 313, + 575, + 543, + 621 + ], + "type": "text", + "content": "In this section, we present our experimental results, including quantitative and qualitative comparisons, an ablation study, and an analysis section that visualizes the behavior of the newly introduced cross-attention mechanism in the adapters." + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 313, + 630, + 543, + 699 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 630, + 543, + 699 + ], + "spans": [ + { + "bbox": [ + 313, + 630, + 543, + 699 + ], + "type": "text", + "content": "Implementation Details We evaluate our method on two versions of Stable Diffusion (SD) [23], specifically versions 1.5 and 2.1, to ensure fair comparison with state-of-the-art methods across different backbones. As described in the ablation study, our caching and conditioning mechanism is applied to the middle layer and every second layer of the" + } + ] + } + ], + "index": 22 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 732, + 317, + 740 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 732, + 317, + 740 + ], + "spans": [ + { + "bbox": [ + 294, + 732, + 317, + 740 + ], + "type": "text", + "content": "12484" + } + ] + } + ], + "index": 23 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 4 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 132, + 114, + 477, + 350 + ], + "blocks": [ + { + "bbox": [ + 67, + 85, + 542, + 106 + ], + "lines": [ + { + "bbox": [ + 67, + 85, + 542, + 106 + ], + "spans": [ + { + "bbox": [ + 67, + 85, + 542, + 106 + ], + "type": "text", + "content": "Table 2. Quantitative results on DreamBooth. DreamCache obtains a better balance between DINO score and CLIP-T compared to all baselines, while also offering a more efficient computational tradeoff (see Table 1)." + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 132, + 114, + 477, + 350 + ], + "lines": [ + { + "bbox": [ + 132, + 114, + 477, + 350 + ], + "spans": [ + { + "bbox": [ + 132, + 114, + 477, + 350 + ], + "type": "table", + "html": "
MethodBackbone#RefDINO (↑)CLIP-I (↑)CLIP-T (↑)
test-time finetuningDreamBooth [24]Imagen3-50.6960.8120.306
DreamBooth [24]SD 1.53-50.6680.8030.305
Textual Inversion [6]SD 1.53-50.5690.7800.255
Custom Diffusion [12]SD 1.53-50.6430.7900.305
BLIP-Diffusion (FT) [13]SD 1.53-50.6700.8050.302
finetuning freeELITE [33]SD 1.510.6210.7710.293
BLIP-Diffusion [13]SD 1.510.5940.7790.300
IP-Adapter [36]SD 1.510.6670.8130.289
Kosmos-G [19]SD 1.510.6940.8470.287
Jedi [37]SD 1.510.6190.7820.304
DreamCache (ours)SD 1.510.7130.8100.298
Re-Imagen [3]Imagen1-30.6000.7400.270
SuTI [4]Imagen1-30.7410.8190.304
Subject-Diffusion [15]SD 2.110.7710.7790.293
BootPig [21]SD 2.130.6740.7970.311
ToffeeNet [39]SD 2.110.7280.8170.306
CAFE [38]SD 2.110.7150.8270.294
DreamCache (ours)SD 2.110.7670.8160.301
", + "image_path": "3b23b4af1301842b793a74c656f712576ebde622d56f6548b0d7b62a3893b3a3.jpg" + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "table_body" + } + ], + "index": 1 + }, + { + "bbox": [ + 66, + 369, + 297, + 483 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 66, + 369, + 297, + 483 + ], + "spans": [ + { + "bbox": [ + 66, + 369, + 297, + 483 + ], + "type": "text", + "content": "decoder. The total number of trainable parameters for DreamCache is 25M. We use the original SD codebase and train the model on " + }, + { + "bbox": [ + 66, + 369, + 297, + 483 + ], + "type": "inline_equation", + "content": "4 \\times 80\\mathrm{GB}" + }, + { + "bbox": [ + 66, + 369, + 297, + 483 + ], + "type": "text", + "content": " A100 GPUs for 25k steps with a batch size of 128, using the AdamW optimizer with a learning rate of " + }, + { + "bbox": [ + 66, + 369, + 297, + 483 + ], + "type": "inline_equation", + "content": "10^{-5}" + }, + { + "bbox": [ + 66, + 369, + 297, + 483 + ], + "type": "text", + "content": ". Input images are resized to " + }, + { + "bbox": [ + 66, + 369, + 297, + 483 + ], + "type": "inline_equation", + "content": "512 \\times 512" + }, + { + "bbox": [ + 66, + 369, + 297, + 483 + ], + "type": "text", + "content": ", with scale, shift, and resize augmentations applied to reference images to enhance model robustness to perturbations. Ablations are conducted on SD 1.5. We generate images with 50 sampling steps, employing classifier-free guidance for image and text conditioning, using a guidance scale of 7.5." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 66, + 491, + 297, + 640 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 66, + 491, + 297, + 640 + ], + "spans": [ + { + "bbox": [ + 66, + 491, + 297, + 640 + ], + "type": "text", + "content": "Evaluation Quantitative evaluations are conducted on the DreamBooth dataset [24], following prior approaches. DreamBooth consists of 30 subjects, each with 25 text prompts. We use a single input image per subject and generate 4 images per subject-prompt combination, resulting in 3,000 generated images. We use pretrained DINO ViT-S/16 and CLIP ViT-B/32 models to calculate the average cosine similarity of global image embeddings between generated and reference images, with metrics denoted as DINO and CLIP-I, respectively. To assess text alignment, we calculate the cosine similarity between embeddings from generated images and text prompts using CLIP's image and text encoders [9], with the corresponding score denoted as CLIP-T." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 67, + 647, + 205, + 658 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 67, + 647, + 205, + 658 + ], + "spans": [ + { + "bbox": [ + 67, + 647, + 205, + 658 + ], + "type": "text", + "content": "4.1. Zero-Shot Personalization" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 66, + 665, + 297, + 699 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 66, + 665, + 297, + 699 + ], + "spans": [ + { + "bbox": [ + 66, + 665, + 297, + 699 + ], + "type": "text", + "content": "We compare DreamCache against state-of-the-art methods for finetuning-based and zero-shot personalization. Table 2 presents quantitative results, indicating the diffusion back" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 313, + 369, + 544, + 483 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 369, + 544, + 483 + ], + "spans": [ + { + "bbox": [ + 313, + 369, + 544, + 483 + ], + "type": "text", + "content": "bone and the number of reference images for each method. Our approach achieves competitive or superior performance compared to other computationally-intensive state-of-the-art methods, which are trained on larger datasets and with significantly more parameters. We refer the reader to Table 1 for the data requirements, training time, and parameter count of the various methods. We remark that generally DINO is a preferred metric for image similarity with respect to CLIP-I, as it is more sensitive to the appearance and fine-grained details of the subjects." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 313, + 483, + 544, + 643 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 483, + 544, + 643 + ], + "spans": [ + { + "bbox": [ + 313, + 483, + 544, + 643 + ], + "type": "text", + "content": "We also present qualitative comparisons with Kosmos-G [19] and BLIP-Diffusion [13]. We remark that several other methods are not reproducible due to the lack of code, datasets, or trained checkpoints. As seen in Fig. 4, our method excels in subject preservation and textual alignment, producing visually superior results. We also notice that Kosmos-G reports a high CLIP-I score, but after inspecting the generated images, it is clear that the score does not entirely reflect the preservation of the reference subject in generated images. In fact, Kosmos-G presents a high degree of background interference, where the partial replication of the reference background boosts the alignment score. For this reason, we also report foreground-masked metrics on the subjects, like MCLIP-I and MDINO, in the Supp. Mat." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 313, + 647, + 455, + 658 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 647, + 455, + 658 + ], + "spans": [ + { + "bbox": [ + 313, + 647, + 455, + 658 + ], + "type": "text", + "content": "4.2. Inference Time Evaluation" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 313, + 665, + 544, + 699 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 665, + 544, + 699 + ], + "spans": [ + { + "bbox": [ + 313, + 665, + 544, + 699 + ], + "type": "text", + "content": "The computational efficiency of our method is compared to the reference-based method BootPIG [21] and encoder-based approaches such as Kosmos-G [19] and Subject" + } + ] + } + ], + "index": 9 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 732, + 317, + 740 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 732, + 317, + 740 + ], + "spans": [ + { + "bbox": [ + 294, + 732, + 317, + 740 + ], + "type": "text", + "content": "12485" + } + ] + } + ], + "index": 10 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 5 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 71, + 97, + 137, + 162 + ], + "blocks": [ + { + "bbox": [ + 86, + 85, + 121, + 96 + ], + "lines": [ + { + "bbox": [ + 86, + 85, + 121, + 96 + ], + "spans": [ + { + "bbox": [ + 86, + 85, + 121, + 96 + ], + "type": "text", + "content": "\"A dog\"" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 71, + 97, + 137, + 162 + ], + "lines": [ + { + "bbox": [ + 71, + 97, + 137, + 162 + ], + "spans": [ + { + "bbox": [ + 71, + 97, + 137, + 162 + ], + "type": "image", + "image_path": "be29f76c7d4ede8064e27ff6766020940a0dab6031ab99231271016101e5e9bb.jpg" + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_body" + } + ], + "index": 1 + }, + { + "type": "image", + "bbox": [ + 140, + 97, + 206, + 159 + ], + "blocks": [ + { + "bbox": [ + 140, + 97, + 206, + 159 + ], + "lines": [ + { + "bbox": [ + 140, + 97, + 206, + 159 + ], + "spans": [ + { + "bbox": [ + 140, + 97, + 206, + 159 + ], + "type": "image", + "image_path": "82158a728d40bcdcdae325a08b9c4c5fe2f83fac63acfe6e09d69d7d5cade368.jpg" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 197, + 86, + 283, + 96 + ], + "lines": [ + { + "bbox": [ + 197, + 86, + 283, + 96 + ], + "spans": [ + { + "bbox": [ + 197, + 86, + 283, + 96 + ], + "type": "text", + "content": "\"wearing a Santa hat\"" + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_caption" + } + ], + "index": 3 + }, + { + "type": "image", + "bbox": [ + 207, + 97, + 272, + 162 + ], + "blocks": [ + { + "bbox": [ + 207, + 97, + 272, + 162 + ], + "lines": [ + { + "bbox": [ + 207, + 97, + 272, + 162 + ], + "spans": [ + { + "bbox": [ + 207, + 97, + 272, + 162 + ], + "type": "image", + "image_path": "ebb431d605f744ceba52ff9c87c1fd831101d5964b8fe8a94d2a1213ebfc960e.jpg" + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "image_body" + } + ], + "index": 5 + }, + { + "type": "image", + "bbox": [ + 274, + 97, + 339, + 161 + ], + "blocks": [ + { + "bbox": [ + 274, + 97, + 339, + 161 + ], + "lines": [ + { + "bbox": [ + 274, + 97, + 339, + 161 + ], + "spans": [ + { + "bbox": [ + 274, + 97, + 339, + 161 + ], + "type": "image", + "image_path": "0364081437e14de23d98385fd1a18defe231a5d0b002fa36707819391bd60757.jpg" + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "image_body" + } + ], + "index": 6 + }, + { + "type": "image", + "bbox": [ + 340, + 97, + 405, + 162 + ], + "blocks": [ + { + "bbox": [ + 359, + 86, + 520, + 96 + ], + "lines": [ + { + "bbox": [ + 359, + 86, + 520, + 96 + ], + "spans": [ + { + "bbox": [ + 359, + 86, + 520, + 96 + ], + "type": "text", + "content": "\"with the Eiffel Tower in the background\"" + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 340, + 97, + 405, + 162 + ], + "lines": [ + { + "bbox": [ + 340, + 97, + 405, + 162 + ], + "spans": [ + { + "bbox": [ + 340, + 97, + 405, + 162 + ], + "type": "image", + "image_path": "ced4cdce5491e89288ba7db0672f261bef4d985f78cd71337119c205cc195a7a.jpg" + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "image_body" + } + ], + "index": 8 + }, + { + "type": "image", + "bbox": [ + 407, + 97, + 473, + 162 + ], + "blocks": [ + { + "bbox": [ + 407, + 97, + 473, + 162 + ], + "lines": [ + { + "bbox": [ + 407, + 97, + 473, + 162 + ], + "spans": [ + { + "bbox": [ + 407, + 97, + 473, + 162 + ], + "type": "image", + "image_path": "a5b006650f369bfea8d691cc7ef75697044ec7a13823eadf9ce38f8b55a71f67.jpg" + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "image_body" + } + ], + "index": 9 + }, + { + "type": "image", + "bbox": [ + 474, + 97, + 539, + 162 + ], + "blocks": [ + { + "bbox": [ + 474, + 97, + 539, + 162 + ], + "lines": [ + { + "bbox": [ + 474, + 97, + 539, + 162 + ], + "spans": [ + { + "bbox": [ + 474, + 97, + 539, + 162 + ], + "type": "image", + "image_path": "e16785f60eae1f9af2092ee8fa023a215bc9e4ea4bca676702e36cafae9ed04f.jpg" + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "image_body" + } + ], + "index": 10 + }, + { + "type": "image", + "bbox": [ + 71, + 175, + 137, + 240 + ], + "blocks": [ + { + "bbox": [ + 88, + 165, + 120, + 174 + ], + "lines": [ + { + "bbox": [ + 88, + 165, + 120, + 174 + ], + "spans": [ + { + "bbox": [ + 88, + 165, + 120, + 174 + ], + "type": "text", + "content": "\"A can\"" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 71, + 175, + 137, + 240 + ], + "lines": [ + { + "bbox": [ + 71, + 175, + 137, + 240 + ], + "spans": [ + { + "bbox": [ + 71, + 175, + 137, + 240 + ], + "type": "image", + "image_path": "2507e224b7535a30b0c5d314074d3ebb7113126ee2677b8a2f15bad3d167e2fe.jpg" + } + ] + } + ], + "index": 12, + "angle": 0, + "type": "image_body" + } + ], + "index": 12 + }, + { + "type": "image", + "bbox": [ + 140, + 175, + 206, + 241 + ], + "blocks": [ + { + "bbox": [ + 189, + 164, + 290, + 175 + ], + "lines": [ + { + "bbox": [ + 189, + 164, + 290, + 175 + ], + "spans": [ + { + "bbox": [ + 189, + 164, + 290, + 175 + ], + "type": "text", + "content": "\"floating on top of water\"" + } + ] + } + ], + "index": 11, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 140, + 175, + 206, + 241 + ], + "lines": [ + { + "bbox": [ + 140, + 175, + 206, + 241 + ], + "spans": [ + { + "bbox": [ + 140, + 175, + 206, + 241 + ], + "type": "image", + "image_path": "96dc5471e153bef8c50cff411a269645b40cb182a8103679ba5204127d1f51ca.jpg" + } + ] + } + ], + "index": 14, + "angle": 0, + "type": "image_body" + } + ], + "index": 14 + }, + { + "type": "image", + "bbox": [ + 208, + 175, + 272, + 241 + ], + "blocks": [ + { + "bbox": [ + 208, + 175, + 272, + 241 + ], + "lines": [ + { + "bbox": [ + 208, + 175, + 272, + 241 + ], + "spans": [ + { + "bbox": [ + 208, + 175, + 272, + 241 + ], + "type": "image", + "image_path": "a9c98199e7c76e5c44e19e87094fb5d7a876d874100fa4a0b1ea4f51d7a70810.jpg" + } + ] + } + ], + "index": 15, + "angle": 0, + "type": "image_body" + } + ], + "index": 15 + }, + { + "type": "image", + "bbox": [ + 274, + 175, + 339, + 241 + ], + "blocks": [ + { + "bbox": [ + 274, + 175, + 339, + 241 + ], + "lines": [ + { + "bbox": [ + 274, + 175, + 339, + 241 + ], + "spans": [ + { + "bbox": [ + 274, + 175, + 339, + 241 + ], + "type": "image", + "image_path": "eabc7d6929f8ca80374bd68054aaae9c72a1218e62a43e5faa83a80aa9c0d98a.jpg" + } + ] + } + ], + "index": 16, + "angle": 0, + "type": "image_body" + } + ], + "index": 16 + }, + { + "type": "image", + "bbox": [ + 340, + 175, + 406, + 241 + ], + "blocks": [ + { + "bbox": [ + 340, + 175, + 406, + 241 + ], + "lines": [ + { + "bbox": [ + 340, + 175, + 406, + 241 + ], + "spans": [ + { + "bbox": [ + 340, + 175, + 406, + 241 + ], + "type": "image", + "image_path": "8c97f3fa5025ee58cc1860117f3a538453dbfbfe57950ba3c1611db97a727ea9.jpg" + } + ] + } + ], + "index": 17, + "angle": 0, + "type": "image_body" + } + ], + "index": 17 + }, + { + "type": "image", + "bbox": [ + 407, + 175, + 473, + 241 + ], + "blocks": [ + { + "bbox": [ + 407, + 175, + 473, + 241 + ], + "lines": [ + { + "bbox": [ + 407, + 175, + 473, + 241 + ], + "spans": [ + { + "bbox": [ + 407, + 175, + 473, + 241 + ], + "type": "image", + "image_path": "5894285b043526c36b61b0ed750d470915bdddb7307918789218439515c7a83c.jpg" + } + ] + } + ], + "index": 18, + "angle": 0, + "type": "image_body" + } + ], + "index": 18 + }, + { + "type": "image", + "bbox": [ + 474, + 175, + 539, + 241 + ], + "blocks": [ + { + "bbox": [ + 474, + 175, + 539, + 241 + ], + "lines": [ + { + "bbox": [ + 474, + 175, + 539, + 241 + ], + "spans": [ + { + "bbox": [ + 474, + 175, + 539, + 241 + ], + "type": "image", + "image_path": "6a02427b8d3f7feb761e3e47665eaf364a38b8df2c53f230257eab10a96c8f75.jpg" + } + ] + } + ], + "index": 19, + "angle": 0, + "type": "image_body" + } + ], + "index": 19 + }, + { + "type": "image", + "bbox": [ + 71, + 254, + 137, + 319 + ], + "blocks": [ + { + "bbox": [ + 89, + 243, + 119, + 253 + ], + "lines": [ + { + "bbox": [ + 89, + 243, + 119, + 253 + ], + "spans": [ + { + "bbox": [ + 89, + 243, + 119, + 253 + ], + "type": "text", + "content": "\"A toy\"" + } + ] + } + ], + "index": 13, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 71, + 254, + 137, + 319 + ], + "lines": [ + { + "bbox": [ + 71, + 254, + 137, + 319 + ], + "spans": [ + { + "bbox": [ + 71, + 254, + 137, + 319 + ], + "type": "image", + "image_path": "96160c2ac6ad20efef76172f424b0d6b532a8e97c03f7b2a71539b3518f85516.jpg" + } + ] + } + ], + "index": 21, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 83, + 322, + 125, + 331 + ], + "lines": [ + { + "bbox": [ + 83, + 322, + 125, + 331 + ], + "spans": [ + { + "bbox": [ + 83, + 322, + 125, + 331 + ], + "type": "text", + "content": "Reference" + } + ] + } + ], + "index": 22, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 67, + 342, + 542, + 364 + ], + "lines": [ + { + "bbox": [ + 67, + 342, + 542, + 364 + ], + "spans": [ + { + "bbox": [ + 67, + 342, + 542, + 364 + ], + "type": "text", + "content": "Figure 4. Visual comparison. Personalized generations on sample concepts. DreamCache preserves reference concept appearance and does not suffer from background interference. BLIP-D [13] and Kosmos-G [19] cannot faithfully preserve visual details from the reference." + } + ] + } + ], + "index": 35, + "angle": 0, + "type": "image_caption" + } + ], + "index": 21 + }, + { + "type": "image", + "bbox": [ + 140, + 254, + 206, + 319 + ], + "blocks": [ + { + "bbox": [ + 140, + 254, + 206, + 319 + ], + "lines": [ + { + "bbox": [ + 140, + 254, + 206, + 319 + ], + "spans": [ + { + "bbox": [ + 140, + 254, + 206, + 319 + ], + "type": "image", + "image_path": "4a3ed0a138fb260e32faf41fd198449b4fc83a1804119156e3cb4cde4e750018.jpg" + } + ] + } + ], + "index": 23, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 156, + 322, + 190, + 331 + ], + "lines": [ + { + "bbox": [ + 156, + 322, + 190, + 331 + ], + "spans": [ + { + "bbox": [ + 156, + 322, + 190, + 331 + ], + "type": "text", + "content": "BLIP-D" + } + ] + } + ], + "index": 24, + "angle": 0, + "type": "image_caption" + } + ], + "index": 23 + }, + { + "type": "image", + "bbox": [ + 208, + 254, + 272, + 319 + ], + "blocks": [ + { + "bbox": [ + 211, + 243, + 270, + 253 + ], + "lines": [ + { + "bbox": [ + 211, + 243, + 270, + 253 + ], + "spans": [ + { + "bbox": [ + 211, + 243, + 270, + 253 + ], + "type": "text", + "content": "\"on the beach\"" + } + ] + } + ], + "index": 20, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 208, + 254, + 272, + 319 + ], + "lines": [ + { + "bbox": [ + 208, + 254, + 272, + 319 + ], + "spans": [ + { + "bbox": [ + 208, + 254, + 272, + 319 + ], + "type": "image", + "image_path": "dc4e18dcb953af79fe1e5ddb1f9a4844b854051d85e344a7410a2e28448e84f5.jpg" + } + ] + } + ], + "index": 25, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 218, + 322, + 261, + 331 + ], + "lines": [ + { + "bbox": [ + 218, + 322, + 261, + 331 + ], + "spans": [ + { + "bbox": [ + 218, + 322, + 261, + 331 + ], + "type": "text", + "content": "Kosmos-G" + } + ] + } + ], + "index": 26, + "angle": 0, + "type": "image_caption" + } + ], + "index": 25 + }, + { + "type": "image", + "bbox": [ + 274, + 254, + 339, + 319 + ], + "blocks": [ + { + "bbox": [ + 274, + 254, + 339, + 319 + ], + "lines": [ + { + "bbox": [ + 274, + 254, + 339, + 319 + ], + "spans": [ + { + "bbox": [ + 274, + 254, + 339, + 319 + ], + "type": "image", + "image_path": "1a327d304d745d5ced93dd3492b05fb1a99ea13a8dc777607d2957c6a641507e.jpg" + } + ] + } + ], + "index": 27, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 278, + 322, + 332, + 331 + ], + "lines": [ + { + "bbox": [ + 278, + 322, + 332, + 331 + ], + "spans": [ + { + "bbox": [ + 278, + 322, + 332, + 331 + ], + "type": "text", + "content": "DreamCache" + } + ] + } + ], + "index": 28, + "angle": 0, + "type": "image_caption" + } + ], + "index": 27 + }, + { + "type": "image", + "bbox": [ + 340, + 254, + 406, + 319 + ], + "blocks": [ + { + "bbox": [ + 340, + 254, + 406, + 319 + ], + "lines": [ + { + "bbox": [ + 340, + 254, + 406, + 319 + ], + "spans": [ + { + "bbox": [ + 340, + 254, + 406, + 319 + ], + "type": "image", + "image_path": "2dc8043099673fd79307756f9fcbe82bbfc3c1fb2d46fca4bdec5a801817a14b.jpg" + } + ] + } + ], + "index": 29, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 356, + 322, + 390, + 331 + ], + "lines": [ + { + "bbox": [ + 356, + 322, + 390, + 331 + ], + "spans": [ + { + "bbox": [ + 356, + 322, + 390, + 331 + ], + "type": "text", + "content": "BLIP-D" + } + ] + } + ], + "index": 30, + "angle": 0, + "type": "image_caption" + } + ], + "index": 29 + }, + { + "type": "image", + "bbox": [ + 407, + 254, + 473, + 319 + ], + "blocks": [ + { + "bbox": [ + 407, + 254, + 473, + 319 + ], + "lines": [ + { + "bbox": [ + 407, + 254, + 473, + 319 + ], + "spans": [ + { + "bbox": [ + 407, + 254, + 473, + 319 + ], + "type": "image", + "image_path": "3208ce0f09f90f0927ec307c2ddb01da153c1ffd72eb7ef6af8cf8a16870360a.jpg" + } + ] + } + ], + "index": 31, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 417, + 322, + 462, + 331 + ], + "lines": [ + { + "bbox": [ + 417, + 322, + 462, + 331 + ], + "spans": [ + { + "bbox": [ + 417, + 322, + 462, + 331 + ], + "type": "text", + "content": "Kosmos-G" + } + ] + } + ], + "index": 32, + "angle": 0, + "type": "image_caption" + } + ], + "index": 31 + }, + { + "type": "image", + "bbox": [ + 474, + 254, + 539, + 319 + ], + "blocks": [ + { + "bbox": [ + 474, + 254, + 539, + 319 + ], + "lines": [ + { + "bbox": [ + 474, + 254, + 539, + 319 + ], + "spans": [ + { + "bbox": [ + 474, + 254, + 539, + 319 + ], + "type": "image", + "image_path": "9c0c65754d7fddb6c2b4a3c01bb56a035ab730190c3e7132aa3155679e98f25f.jpg" + } + ] + } + ], + "index": 33, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 478, + 322, + 533, + 331 + ], + "lines": [ + { + "bbox": [ + 478, + 322, + 533, + 331 + ], + "spans": [ + { + "bbox": [ + 478, + 322, + 533, + 331 + ], + "type": "text", + "content": "DreamCache" + } + ] + } + ], + "index": 34, + "angle": 0, + "type": "image_caption" + } + ], + "index": 33 + }, + { + "type": "table", + "bbox": [ + 70, + 411, + 293, + 482 + ], + "blocks": [ + { + "bbox": [ + 67, + 382, + 295, + 403 + ], + "lines": [ + { + "bbox": [ + 67, + 382, + 295, + 403 + ], + "spans": [ + { + "bbox": [ + 67, + 382, + 295, + 403 + ], + "type": "text", + "content": "Table 3. Computational comparison.*: time to generate an image with 100 timesteps, evaluated on a single NVIDIA A100 GPU." + } + ] + } + ], + "index": 36, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 70, + 411, + 293, + 482 + ], + "lines": [ + { + "bbox": [ + 70, + 411, + 293, + 482 + ], + "spans": [ + { + "bbox": [ + 70, + 411, + 293, + 482 + ], + "type": "table", + "html": "
MethodInference Time*Extra Params Size
ELITE [33]6.24 s914 MB
BLIP-Diffusion [13]3.92 s760 MB
BootPig [21]7.55 s1900 MB
DreamCache (ours)3.88 s42 MB
", + "image_path": "ae89d8b9290f227f45398655acd7a9ee079e06c60bec37d662686a8e128f2efc.jpg" + } + ] + } + ], + "index": 37, + "angle": 0, + "type": "table_body" + } + ], + "index": 37 + }, + { + "bbox": [ + 66, + 502, + 296, + 626 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 66, + 502, + 296, + 626 + ], + "spans": [ + { + "bbox": [ + 66, + 502, + 296, + 626 + ], + "type": "text", + "content": "Diffusion [15]. Table 3 provides a detailed comparison of inference time, accounting for both personalization time (e.g., the time to generate the cache for DreamCache) and the time to sample the personalized image. We also report the increase in model size, i.e., the storage (in FP16 precision) required for the extra parameters to allow for personalization, showing that DreamCache is one order of magnitude more compact than the state of the art. Overall, DreamCache offers a lightweight solution that achieves state-of-the-art performance with faster inference and reduced computational overhead." + } + ] + } + ], + "index": 38 + }, + { + "bbox": [ + 67, + 635, + 162, + 646 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 67, + 635, + 162, + 646 + ], + "spans": [ + { + "bbox": [ + 67, + 635, + 162, + 646 + ], + "type": "text", + "content": "4.3. Ablation Studies" + } + ] + } + ], + "index": 39 + }, + { + "bbox": [ + 66, + 653, + 296, + 699 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 66, + 653, + 296, + 699 + ], + "spans": [ + { + "bbox": [ + 66, + 653, + 296, + 699 + ], + "type": "text", + "content": "We validate our design choices through a series of studies, examining different conditioning mechanisms, evaluating our feature caching approach, and analyzing the impact of synthetic dataset scaling." + } + ] + } + ], + "index": 40 + }, + { + "type": "table", + "bbox": [ + 317, + 411, + 541, + 483 + ], + "blocks": [ + { + "bbox": [ + 313, + 382, + 542, + 403 + ], + "lines": [ + { + "bbox": [ + 313, + 382, + 542, + 403 + ], + "spans": [ + { + "bbox": [ + 313, + 382, + 542, + 403 + ], + "type": "text", + "content": "Table 4. Reference feature integration. DreamCache uses the best tradeoff between accuracy and complexity." + } + ] + } + ], + "index": 41, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 317, + 411, + 541, + 483 + ], + "lines": [ + { + "bbox": [ + 317, + 411, + 541, + 483 + ], + "spans": [ + { + "bbox": [ + 317, + 411, + 541, + 483 + ], + "type": "table", + "html": "
MethodCLIP-I (↑)CLIP-T (↑)Params
Textual Sum [36]0.7880,28219M
Spatial Sum0.8120.29316M
Decoupled Blocks [8]0.8080.30061M
Spatial Concat (ours)0.8100.29825M
", + "image_path": "e9b619f27d9a9a4b577709ddc1fe8ffd455fdbd18b54fad041a6fce5c076252d.jpg" + } + ] + } + ], + "index": 42, + "angle": 0, + "type": "table_body" + } + ], + "index": 42 + }, + { + "bbox": [ + 313, + 503, + 544, + 674 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 503, + 544, + 674 + ], + "spans": [ + { + "bbox": [ + 313, + 503, + 544, + 674 + ], + "type": "text", + "content": "Reference Feature Integration We compare various conditioning strategies to integrate reference features in Table 4. Our spatial cross-attention block with concatenation between the output of self- and cross-attention (\"Spatial Concat\") was evaluated against different alternatives, including IP-Adapter's conditioning mechanism [36] (\"Textual Sum\"), which sums the decoupled cross-attention output with that from textual cross-attention. We also tested a variant (\"Spatial Sum\") where self- and cross-attention conditioning outputs are summed. Additionally, we also assessed an alternative conditioning procedure inspired by ViCo [8] (\"Decoupled Blocks\"), involving independent and interleaved cross-attention blocks. Results in Table 4 indicate that the proposed \"Spatial Concat\" offers the best balance of text alignment and parameter efficiency." + } + ] + } + ], + "index": 43 + }, + { + "bbox": [ + 313, + 675, + 543, + 699 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 675, + 543, + 699 + ], + "spans": [ + { + "bbox": [ + 313, + 675, + 543, + 699 + ], + "type": "text", + "content": "We further explored optimal conditioning insertion within the U-Net backbone in Table 5, determining that applying" + } + ] + } + ], + "index": 44 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 732, + 317, + 740 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 732, + 317, + 740 + ], + "spans": [ + { + "bbox": [ + 294, + 732, + 317, + 740 + ], + "type": "text", + "content": "12486" + } + ] + } + ], + "index": 45 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 6 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 67, + 115, + 296, + 209 + ], + "blocks": [ + { + "bbox": [ + 67, + 86, + 297, + 106 + ], + "lines": [ + { + "bbox": [ + 67, + 86, + 297, + 106 + ], + "spans": [ + { + "bbox": [ + 67, + 86, + 297, + 106 + ], + "type": "text", + "content": "Table 5. Cache positioning in the U-Net backbone offers a further tradeoff between accuracy and complexity." + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 67, + 115, + 296, + 209 + ], + "lines": [ + { + "bbox": [ + 67, + 115, + 296, + 209 + ], + "spans": [ + { + "bbox": [ + 67, + 115, + 296, + 209 + ], + "type": "table", + "html": "
EncoderMiddleDecoderCLIP-I (↑)CLIP-T (↑)Params
✓XX0.7210.30311M
✓✓X0.7490.30619M
X✓X0.7160.3028M
XX✓0.7990.29617M
X✓✓0.8100.29825M
✓✓✓0.8130.29736M
", + "image_path": "4e46f7f99d2d6af168c097f2393a24d13ad334df9edd5e7dc6f91ffc68347f8d.jpg" + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "table_body" + } + ], + "index": 1 + }, + { + "type": "table", + "bbox": [ + 111, + 239, + 252, + 287 + ], + "blocks": [ + { + "bbox": [ + 67, + 220, + 295, + 232 + ], + "lines": [ + { + "bbox": [ + 67, + 220, + 295, + 232 + ], + "spans": [ + { + "bbox": [ + 67, + 220, + 295, + 232 + ], + "type": "text", + "content": "Table 6. Caching with text is not influential and adds complexity." + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 111, + 239, + 252, + 287 + ], + "lines": [ + { + "bbox": [ + 111, + 239, + 252, + 287 + ], + "spans": [ + { + "bbox": [ + 111, + 239, + 252, + 287 + ], + "type": "table", + "html": "
Text-FreeCLIP-I (↑)CLIP-T (↑)
×0.8110.295
✓0.8100.298
", + "image_path": "d27e0ac50d4246f5c89f22880691d6fd930631df4aee284788b4ca86b97ceca0.jpg" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "table_body" + } + ], + "index": 3 + }, + { + "type": "table", + "bbox": [ + 100, + 317, + 264, + 388 + ], + "blocks": [ + { + "bbox": [ + 82, + 298, + 280, + 309 + ], + "lines": [ + { + "bbox": [ + 82, + 298, + 280, + 309 + ], + "spans": [ + { + "bbox": [ + 82, + 298, + 280, + 309 + ], + "type": "text", + "content": "Table 7. Dataset impact for both synthetic and real data." + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 100, + 317, + 264, + 388 + ], + "lines": [ + { + "bbox": [ + 100, + 317, + 264, + 388 + ], + "spans": [ + { + "bbox": [ + 100, + 317, + 264, + 388 + ], + "type": "table", + "html": "
DatasetCLIP-I (↑)CLIP-T (↑)
Synthetic-50K0.7810.304
Synthetic-200K0.7970,301
Synthetic-400K0.8100.298
LAION-5M0.8140.242
", + "image_path": "386934df26c286e8e9a670e75a33b5f999ff89eadb67720d228bc5d2e0e2d967.jpg" + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "table_body" + } + ], + "index": 5 + }, + { + "bbox": [ + 66, + 400, + 296, + 435 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 66, + 400, + 296, + 435 + ], + "spans": [ + { + "bbox": [ + 66, + 400, + 296, + 435 + ], + "type": "text", + "content": "conditioning (and therefore feature caching) at the middle layer and every second layer of the decoder achieved the best tradeoff between performance and parameter count." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 66, + 441, + 297, + 533 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 66, + 441, + 297, + 533 + ], + "spans": [ + { + "bbox": [ + 66, + 441, + 297, + 533 + ], + "type": "text", + "content": "Text Input for Cached Features Our feature caching procedure is designed to be text-free, leveraging the classifier-free guidance used during pretraining where captions were occasionally omitted. We compare this approach with a version including textual inputs during caching (e.g., 'A photo of...'). Table 6 shows that adding text conditioning slightly reduces text alignment while increasing complexity and potentially introducing noise in cases of inaccurate captions." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 66, + 539, + 297, + 700 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 66, + 539, + 297, + 700 + ], + "spans": [ + { + "bbox": [ + 66, + 539, + 297, + 700 + ], + "type": "text", + "content": "Dataset Impact We demonstrate the importance of our synthetic dataset to train the conditioning adapters and the effect of scaling its size. For this purpose, we created synthetic datasets according to the procedure in Sec. 3.3 of sizes 50K, 200K, and 400K samples. We also tested the real-world 5M samples from the LAION [26] dataset, which, lacking target-caption-reference triplets, required reuse of target images as reference images too. Table 7 shows that increasing dataset size improves image alignment, though slightly reduces textual alignment. Notably, LAION improves image alignment but struggles with textual alignment. This highlights the importance of triplet data (target image, reference image, and caption) for effective zero-shot personalization, ensuring both subject preservation and textual editability." + } + ] + } + ], + "index": 8 + }, + { + "type": "image", + "bbox": [ + 318, + 83, + 541, + 203 + ], + "blocks": [ + { + "bbox": [ + 318, + 83, + 541, + 203 + ], + "lines": [ + { + "bbox": [ + 318, + 83, + 541, + 203 + ], + "spans": [ + { + "bbox": [ + 318, + 83, + 541, + 203 + ], + "type": "image", + "image_path": "8f6e6e1d1e4b6a2758dcd4909a5593d41acc79da2a621ce9c161ce2cef600754.jpg" + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 313, + 219, + 544, + 273 + ], + "lines": [ + { + "bbox": [ + 313, + 219, + 544, + 273 + ], + "spans": [ + { + "bbox": [ + 313, + 219, + 544, + 273 + ], + "type": "text", + "content": "Figure 5. Visualization of reference image impact. Cross-attention maps between cached reference features and features of the image under generation. Left: attention map at layers at " + }, + { + "bbox": [ + 313, + 219, + 544, + 273 + ], + "type": "inline_equation", + "content": "16 \\times 16" + }, + { + "bbox": [ + 313, + 219, + 544, + 273 + ], + "type": "text", + "content": " resolution (left reference, right generated). Right: " + }, + { + "bbox": [ + 313, + 219, + 544, + 273 + ], + "type": "inline_equation", + "content": "32 \\times 32" + }, + { + "bbox": [ + 313, + 219, + 544, + 273 + ], + "type": "text", + "content": ". Attention values are highly localized in the region of interest." + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "image_caption" + } + ], + "index": 9 + }, + { + "bbox": [ + 313, + 283, + 468, + 296 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 283, + 468, + 296 + ], + "spans": [ + { + "bbox": [ + 313, + 283, + 468, + 296 + ], + "type": "text", + "content": "4.4. Visualizing Reference Impact" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 313, + 301, + 544, + 415 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 301, + 544, + 415 + ], + "spans": [ + { + "bbox": [ + 313, + 301, + 544, + 415 + ], + "type": "text", + "content": "Finally, we analyze how the cross-attention mechanism in DreamCache impacts image generation by visualizing cached reference feature influence. Attention map visualizations at different resolutions are provided in Fig. 5. Specifically, attention maps between the query from the current generation and the key derived from reference features reveal a highly localized focus on the subject, without interference from background elements. This mechanism models correspondences effectively, integrating reference information into the generated image." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 313, + 420, + 463, + 431 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 420, + 463, + 431 + ], + "spans": [ + { + "bbox": [ + 313, + 420, + 463, + 431 + ], + "type": "text", + "content": "5. Discussion and Conclusions" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 312, + 436, + 544, + 699 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 312, + 436, + 544, + 699 + ], + "spans": [ + { + "bbox": [ + 312, + 436, + 544, + 699 + ], + "type": "text", + "content": "In this paper, we proposed DreamCache, a novel approach to personalized text-to-image generation that uses feature caching to overcome the limitations of existing methods. By caching reference features from a small subset of layers of the U-Net only once, our method significantly reduces both computational and memory demands, enabling efficient, real-time personalized image generation. Unlike previous approaches, DreamCache avoids the need for costly finetuning, external image encoders, or parallel reference processing, making it lightweight and suitable for plug-and-play deployment. Our experiments demonstrate that DreamCache achieves state-of-the-art zero-shot personalization with only 25M additional parameters and a fast training process. While DreamCache is a promising direction towards more efficient personalized generation, it has some limitations. Although effective for single-subject personalization, our approach may require adaptation for complex multi-subject generation where feature interference can occur. Additionally, certain edge cases, such as highly abstract or stylistic images, may challenge the caching mechanism's capacity to accurately preserve subject details. To address these challenges, future work may explore adaptive caching techniques or multi-reference feature integration." + } + ] + } + ], + "index": 14 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 731, + 317, + 740 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 731, + 317, + 740 + ], + "spans": [ + { + "bbox": [ + 294, + 731, + 317, + 740 + ], + "type": "text", + "content": "12487" + } + ] + } + ], + "index": 15 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 7 + }, + { + "para_blocks": [ + { + "bbox": [ + 67, + 86, + 123, + 98 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 67, + 86, + 123, + 98 + ], + "spans": [ + { + "bbox": [ + 67, + 86, + 123, + 98 + ], + "type": "text", + "content": "References" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 67, + 105, + 297, + 698 + ], + "type": "list", + "angle": 0, + "index": 14, + "blocks": [ + { + "bbox": [ + 72, + 105, + 297, + 156 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 72, + 105, + 297, + 156 + ], + "spans": [ + { + "bbox": [ + 72, + 105, + 297, + 156 + ], + "type": "text", + "content": "[1] Moab Arar, Rinon Gal, Yuval Atzmon, Gal Chechik, Daniel Cohen-Or, Ariel Shamir, and Amit H. Bermano. Domain-agnostic tuning-encoder for fast personalization of text-to-image models. In SIGGRAPH Asia 2023 Conference Papers, pages 1–10, 2023. 1, 3" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 72, + 158, + 297, + 209 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 72, + 158, + 297, + 209 + ], + "spans": [ + { + "bbox": [ + 72, + 158, + 297, + 209 + ], + "type": "text", + "content": "[2] Mathilde Caron, Hugo Touvron, Ishan Misra, Hervé Jégou, Julien Mairal, Piotr Bojanowski, and Armand Joulin. Emerging properties in self-supervised vision transformers. In Proceedings of the IEEE/CVF international conference on computer vision, pages 9650-9660, 2021. 1" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 72, + 211, + 296, + 252 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 72, + 211, + 296, + 252 + ], + "spans": [ + { + "bbox": [ + 72, + 211, + 296, + 252 + ], + "type": "text", + "content": "[3] Wenhu Chen, Hexiang Hu, Chitwan Sahara, and William W Cohen. Re-imagen: Retrieval-augmented text-to-image generator. In The Eleventh International Conference on Learning Representations, 2022. 6" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 72, + 254, + 296, + 296 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 72, + 254, + 296, + 296 + ], + "spans": [ + { + "bbox": [ + 72, + 254, + 296, + 296 + ], + "type": "text", + "content": "[4] Wenhu Chen, Hexiang Hu, Yandong Li, Nataniel Rui, Xuhui Jia, Ming-Wei Chang, and William W Cohen. Subject-driven text-to-image generation via apprenticeship learning. arXiv preprint arXiv:2304.00186, 2023. 2, 3, 6" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 72, + 297, + 296, + 346 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 72, + 297, + 296, + 346 + ], + "spans": [ + { + "bbox": [ + 72, + 297, + 296, + 346 + ], + "type": "text", + "content": "[5] Abhimanyu Dubey, Abhinav Jauhri, Abhinav Pandey, Abhishek Kadian, Ahmad Al-Dahle, Aiesha Letman, Akhil Mathur, Alan Schelten, Amy Yang, Angela Fan, et al. The llama 3 herd of models. arXiv preprint arXiv:2407.21783, 2024. 5" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 72, + 349, + 297, + 411 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 72, + 349, + 297, + 411 + ], + "spans": [ + { + "bbox": [ + 72, + 349, + 297, + 411 + ], + "type": "text", + "content": "[6] Rinon Gal, Yuval Alaluf, Yuval Atzmon, Or Patashnik, Amit Haim Bermano, Gal Chechik, and Daniel Cohen-or. An image is worth one word: Personalizing text-to-image generation using textual inversion. In The Eleventh International Conference on Learning Representations, 2022. 1, 2, 3, 6" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 72, + 412, + 296, + 454 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 72, + 412, + 296, + 454 + ], + "spans": [ + { + "bbox": [ + 72, + 412, + 296, + 454 + ], + "type": "text", + "content": "[7] Rinon Gal, Moab Arar, Yuval Atzmon, Amit H Bermano, Gal Chechik, and Daniel Cohen-Or. Encoder-based domain tuning for fast personalization of text-to-image models. ACM Transactions on Graphics (TOG), 42(4):1-13, 2023. 1, 3" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 72, + 456, + 296, + 495 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 72, + 456, + 296, + 495 + ], + "spans": [ + { + "bbox": [ + 72, + 456, + 296, + 495 + ], + "type": "text", + "content": "[8] Shaozhe Hao, Kai Han, Shihao Zhao, and Kwan-Yee K Wong. Vico: Plug-and-play visual condition for personalized text-to-image generation. arXiv preprint arXiv:2306.00971, 2023. 1, 3, 7" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 72, + 498, + 296, + 548 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 72, + 498, + 296, + 548 + ], + "spans": [ + { + "bbox": [ + 72, + 498, + 296, + 548 + ], + "type": "text", + "content": "[9] Jack Hessel, Ari Holtzman, Maxwell Forbes, Ronan Le Bras, and Yejin Choi. Clipscore: A reference-free evaluation metric for image captioning. In Proceedings of the 2021 Conference on Empirical Methods in Natural Language Processing, pages 7514-7528, 2021. 6" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 67, + 550, + 296, + 581 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 67, + 550, + 296, + 581 + ], + "spans": [ + { + "bbox": [ + 67, + 550, + 296, + 581 + ], + "type": "text", + "content": "[10] Jonathan Ho, Ajay Jain, and Pieter Abbeel. Denoising diffusion probabilistic models. Advances in neural information processing systems, 33:6840-6851, 2020. 1" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 67, + 582, + 296, + 624 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 67, + 582, + 296, + 624 + ], + "spans": [ + { + "bbox": [ + 67, + 582, + 296, + 624 + ], + "type": "text", + "content": "[11] Alexander Kirillov, Eric Mintun, Nikhila Ravi, Hanzi Mao, Chloe Rolland, Laura Gustafson, Tete Xiao, Spencer Whitehead, Alexander C Berg, Wan-Yen Lo, et al. Segment anything. arXiv preprint arXiv:2304.02643, 2023. 5" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 67, + 625, + 296, + 675 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 67, + 625, + 296, + 675 + ], + "spans": [ + { + "bbox": [ + 67, + 625, + 296, + 675 + ], + "type": "text", + "content": "[12] Nupur Kumari, Bingliang Zhang, Richard Zhang, Eli Shechtman, and Jun-Yan Zhu. Multi-concept customization of text-to-image diffusion. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 1931-1941, 2023. 1, 2, 6" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 67, + 677, + 296, + 698 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 67, + 677, + 296, + 698 + ], + "spans": [ + { + "bbox": [ + 67, + 677, + 296, + 698 + ], + "type": "text", + "content": "[13] Dongxu Li, Junnan Li, and Steven CH Hoi. Blip-diffusion: Pre-trained subject representation for control" + } + ] + } + ], + "index": 13 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 315, + 88, + 542, + 698 + ], + "type": "list", + "angle": 0, + "index": 28, + "blocks": [ + { + "bbox": [ + 333, + 88, + 542, + 108 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 333, + 88, + 542, + 108 + ], + "spans": [ + { + "bbox": [ + 333, + 88, + 542, + 108 + ], + "type": "text", + "content": "lable text-to-image generation and editing. arXiv preprint arXiv:2305.14720, 2023. 1, 2, 3, 6, 7" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 315, + 110, + 542, + 161 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 315, + 110, + 542, + 161 + ], + "spans": [ + { + "bbox": [ + 315, + 110, + 542, + 161 + ], + "type": "text", + "content": "[14] Shilong Liu, Zhaoyang Zeng, Tianhe Ren, Feng Li, Hao Zhang, Jie Yang, Chunyuan Li, Jianwei Yang, Hang Su, Jun Zhu, et al. Grounding dino: Marrying dino with grounded pre-training for open-set object detection. arXiv preprint arXiv:2303.05499, 2023. 5" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 315, + 163, + 542, + 205 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 315, + 163, + 542, + 205 + ], + "spans": [ + { + "bbox": [ + 315, + 163, + 542, + 205 + ], + "type": "text", + "content": "[15] Jian Ma, Junhao Liang, Chen Chen, and Haonan Lu. Subject-diffusion: Open domain personalized text-to-image generation without test-time fine-tuning. arXiv preprint arXiv:2307.11410, 2023. 1, 2, 3, 5, 6, 7" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 315, + 206, + 542, + 237 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 315, + 206, + 542, + 237 + ], + "spans": [ + { + "bbox": [ + 315, + 206, + 542, + 237 + ], + "type": "text", + "content": "[16] Xinyin Ma, Gongfan Fang, Michael Bi Mi, and Xinchao Wang. Learning-to-cache: Accelerating diffusion transformer via layer caching. arXiv preprint arXiv:2406.01733, 2024. 4" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 315, + 239, + 542, + 281 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 315, + 239, + 542, + 281 + ], + "spans": [ + { + "bbox": [ + 315, + 239, + 542, + 281 + ], + "type": "text", + "content": "[17] Xinyin Ma, Gongfan Fang, and Xinchao Wang. Deepcache: Accelerating diffusion models for free. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 15762-15772, 2024. 3" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 315, + 282, + 542, + 323 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 315, + 282, + 542, + 323 + ], + "spans": [ + { + "bbox": [ + 315, + 282, + 542, + 323 + ], + "type": "text", + "content": "[18] Boris Oreshkin, Pau Rodríguez López, and Alexandre Lacoste. Tadam: Task dependent adaptive metric for improved few-shot learning. Advances in neural information processing systems, 31, 2018. 4" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 315, + 325, + 542, + 365 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 315, + 325, + 542, + 365 + ], + "spans": [ + { + "bbox": [ + 315, + 325, + 542, + 365 + ], + "type": "text", + "content": "[19] Xichen Pan, Li Dong, Shaohan Huang, Zhiliang Peng, Wenhu Chen, and Furu Wei. Kosmos-g: Generating images in context with multimodal large language models. arXiv preprint arXiv:2310.02992, 2023. 1, 2, 3, 6, 7" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 315, + 369, + 542, + 409 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 315, + 369, + 542, + 409 + ], + "spans": [ + { + "bbox": [ + 315, + 369, + 542, + 409 + ], + "type": "text", + "content": "[20] Maitreya Patel, Sangmin Jung, Chitta Baral, and Yezhou Yang. λ-eclipse: Multi-concept personalized text-to-image diffusion models by leveraging clip latent space. arXiv preprint arXiv:2402.05195, 2024. 1, 3" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 315, + 411, + 542, + 453 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 315, + 411, + 542, + 453 + ], + "spans": [ + { + "bbox": [ + 315, + 411, + 542, + 453 + ], + "type": "text", + "content": "[21] Senthil Purushwalkam, Akash Gokul, Shafiq Joty, and Nikhil Naik. Bootpig: Bootstrapping zero-shot personalized image generation capabilities in pretrained diffusion models. arXiv preprint arXiv:2401.13974, 2024. 1, 2, 3, 5, 6, 7" + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 315, + 455, + 542, + 516 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 315, + 455, + 542, + 516 + ], + "spans": [ + { + "bbox": [ + 315, + 455, + 542, + 516 + ], + "type": "text", + "content": "[22] Alec Radford, Jong Wook Kim, Chris Hallacy, Aditya Ramesh, Gabriel Goh, Sandhini Agarwal, Girish Sastry, Amanda Askell, Pamela Mishkin, Jack Clark, et al. Learning transferable visual models from natural language supervision. In International conference on machine learning, pages 8748-8763. PMLR, 2021. 1" + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 315, + 518, + 542, + 570 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 315, + 518, + 542, + 570 + ], + "spans": [ + { + "bbox": [ + 315, + 518, + 542, + 570 + ], + "type": "text", + "content": "[23] Robin Rombach, Andreas Blattmann, Dominik Lorenz, Patrick Esser, and Björn Ommer. High-resolution image synthesis with latent diffusion models. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 10684-10695, 2022. 1, 5" + } + ] + } + ], + "index": 25 + }, + { + "bbox": [ + 315, + 572, + 542, + 633 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 315, + 572, + 542, + 633 + ], + "spans": [ + { + "bbox": [ + 315, + 572, + 542, + 633 + ], + "type": "text", + "content": "[24] Nataniel Ruiz, Yuanzhen Li, Varun Jampani, Yael Pritch, Michael Rubinstein, and Kfir Aberman. Dreambooth: Fine tuning text-to-image diffusion models for subject-driven generation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 22500-22510, 2023. 1, 2, 6" + } + ] + } + ], + "index": 26 + }, + { + "bbox": [ + 315, + 635, + 542, + 698 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 315, + 635, + 542, + 698 + ], + "spans": [ + { + "bbox": [ + 315, + 635, + 542, + 698 + ], + "type": "text", + "content": "[25] Chitwan Sahara, William Chan, Saurabh Saxena, Lala Li, Jay Whang, Emily L Denton, Kamyar Ghasemipour, Raphael Gontijo Lopes, Burcu Karagol Ayan, Tim Salimans, et al. Photorealistic text-to-image diffusion models with deep language understanding. Advances in Neural Information Processing Systems, 35:36479-36494, 2022. 1" + } + ] + } + ], + "index": 27 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 732, + 317, + 740 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 732, + 317, + 740 + ], + "spans": [ + { + "bbox": [ + 294, + 732, + 317, + 740 + ], + "type": "text", + "content": "12488" + } + ] + } + ], + "index": 29 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 8 + }, + { + "para_blocks": [ + { + "bbox": [ + 67, + 87, + 297, + 698 + ], + "type": "list", + "angle": 0, + "index": 13, + "blocks": [ + { + "bbox": [ + 67, + 87, + 297, + 150 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 67, + 87, + 297, + 150 + ], + "spans": [ + { + "bbox": [ + 67, + 87, + 297, + 150 + ], + "type": "text", + "content": "[26] Christoph Schuhmann, Romain Beaumont, Richard Vencu, Cade Gordon, Ross Wightman, Mehdi Cherti, Theo Coombes, Aarush Katta, Clayton Mullis, Mitchell Wortsman, et al. Laion-5b: An open large-scale dataset for training next generation image-text models. Advances in Neural Information Processing Systems, 35:25278–25294, 2022. 8" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 67, + 151, + 297, + 182 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 67, + 151, + 297, + 182 + ], + "spans": [ + { + "bbox": [ + 67, + 151, + 297, + 182 + ], + "type": "text", + "content": "[27] Jake Snell, Kevin Swersky, and Richard Zemel. Prototypical networks for few-shot learning. Advances in neural information processing systems, 30, 2017. 4" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 69, + 183, + 296, + 233 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 69, + 183, + 296, + 233 + ], + "spans": [ + { + "bbox": [ + 69, + 183, + 296, + 233 + ], + "type": "text", + "content": "[28] Jascha Sohl-Dickstein, Eric Weiss, Niru Maheswaranathan, and Surya Ganguli. Deep unsupervised learning using nonequilibrium thermodynamics. In International conference on machine learning, pages 2256-2265. PMLR, 2015. 1" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 69, + 236, + 295, + 266 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 69, + 236, + 295, + 266 + ], + "spans": [ + { + "bbox": [ + 69, + 236, + 295, + 266 + ], + "type": "text", + "content": "[29] Yang Song and Stefano Ermon. Generative modeling by estimating gradients of the data distribution. Advances in neural information processing systems, 32, 2019. 5" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 69, + 267, + 296, + 308 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 69, + 267, + 296, + 308 + ], + "spans": [ + { + "bbox": [ + 69, + 267, + 296, + 308 + ], + "type": "text", + "content": "[30] Yoad Tewel, Rinon Gal, Gal Chechik, and Yuval Atzmon. Key-locked rank one editing for text-to-image personalization. In ACM SIGGRAPH 2023 Conference Proceedings, pages 1-11, 2023. 1, 3" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 69, + 309, + 296, + 361 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 69, + 309, + 296, + 361 + ], + "spans": [ + { + "bbox": [ + 69, + 309, + 296, + 361 + ], + "type": "text", + "content": "[31] Eleni Triantafillou, Hugo Larochelle, Jake Snell, Josh Tenenbaum, Kevin Jordan Swersky, Mengye Ren, Richard Zemel, and Sachin Ravi. Meta-learning for semi-supervised few-shot classification. In International Conference on Learning Representations, 2018. 4" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 69, + 362, + 295, + 393 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 69, + 362, + 295, + 393 + ], + "spans": [ + { + "bbox": [ + 69, + 362, + 295, + 393 + ], + "type": "text", + "content": "[32] Andrey Voynov, Qinghao Chu, Daniel Cohen-Or, and Kfir Aberman. p+: Extended textual conditioning in text-to-image generation. arXiv preprint arXiv:2303.09522, 2023. 1, 3" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 69, + 394, + 296, + 445 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 69, + 394, + 296, + 445 + ], + "spans": [ + { + "bbox": [ + 69, + 394, + 296, + 445 + ], + "type": "text", + "content": "[33] Yuxiang Wei, Yabo Zhang, Zhilong Ji, Jinfeng Bai, Lei Zhang, and Wangmeng Zuo. Elite: Encoding visual concepts into textual embeddings for customized text-to-image generation. 2023 IEEE/CVF International Conference on Computer Vision (ICCV), pages 15897-15907, 2023. 1, 2, 3, 5, 6, 7" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 69, + 447, + 296, + 509 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 69, + 447, + 296, + 509 + ], + "spans": [ + { + "bbox": [ + 69, + 447, + 296, + 509 + ], + "type": "text", + "content": "[34] Felix Wimbauer, Bichen Wu, Edgar Schoenfeld, Xiaoliang Dai, Ji Hou, Zijian He, Artsiom Sanakoyeu, Peizhao Zhang, Sam Tsai, Jonas Kohler, et al. Cache me if you can: Accelerating diffusion models through block caching. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 6211-6220, 2024. 3" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 69, + 510, + 296, + 551 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 69, + 510, + 296, + 551 + ], + "spans": [ + { + "bbox": [ + 69, + 510, + 296, + 551 + ], + "type": "text", + "content": "[35] Jianan Yang, Haobo Wang, Yanming Zhang, Ruixuan Xiao, Sai Wu, Gang Chen, and Junbo Zhao. Controllable textual inversion for personalized text-to-image generation. arXiv preprint arXiv:2304.05265, 2023. 1, 3" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 69, + 552, + 296, + 592 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 69, + 552, + 296, + 592 + ], + "spans": [ + { + "bbox": [ + 69, + 552, + 296, + 592 + ], + "type": "text", + "content": "[36] Hu Ye, Jun Zhang, Sibo Liu, Xiao Han, and Wei Yang. Ip-adapter: Text compatible image prompt adapter for text-to-image diffusion models. arXiv preprint arXiv:2308.06721, 2023. 1, 2, 3, 6, 7" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 69, + 594, + 296, + 656 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 69, + 594, + 296, + 656 + ], + "spans": [ + { + "bbox": [ + 69, + 594, + 296, + 656 + ], + "type": "text", + "content": "[37] Yu Zeng, Vishal M Patel, Haochen Wang, Xun Huang, Ting-Chun Wang, Ming-Yu Liu, and Yogesh Balaji. Jedi: Joint-image diffusion models for finetuning-free personalized text-to-image generation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 6786-6795, 2024. 1, 2, 3, 4, 6" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 69, + 658, + 296, + 698 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 69, + 658, + 296, + 698 + ], + "spans": [ + { + "bbox": [ + 69, + 658, + 296, + 698 + ], + "type": "text", + "content": "[38] Yufan Zhou, Ruiyi Zhang, Jiumiang Gu, and Tong Sun. Customization assistant for text-to-image generation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 9182-9191, 2024. 2, 3, 6" + } + ] + } + ], + "index": 12 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 315, + 87, + 544, + 140 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 315, + 87, + 544, + 140 + ], + "spans": [ + { + "bbox": [ + 315, + 87, + 544, + 140 + ], + "type": "text", + "content": "[39] Yufan Zhou, Ruiyi Zhang, Kaizhi Zheng, Nanxuan Zhao, Jiumiang Gu, Zichao Wang, Xin Eric Wang, and Tong Sun. Toffee: Efficient million-scale dataset construction for subject-driven text-to-image generation. arXiv preprint arXiv:2406.09305, 2024.1,2,3,6" + } + ] + } + ], + "index": 14 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 732, + 317, + 740 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 732, + 317, + 740 + ], + "spans": [ + { + "bbox": [ + 294, + 732, + 317, + 740 + ], + "type": "text", + "content": "12489" + } + ] + } + ], + "index": 15 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 9 + } + ], + "_backend": "vlm", + "_version_name": "2.6.4" +} \ No newline at end of file diff --git a/2025/DreamOmni_ Unified Image Generation and Editing/ff84d3e1-2391-4334-90b8-2fe54fd66a09_content_list.json b/2025/DreamOmni_ Unified Image Generation and Editing/ff84d3e1-2391-4334-90b8-2fe54fd66a09_content_list.json new file mode 100644 index 0000000000000000000000000000000000000000..b64102ab78ef33bb6aee6b821d0aa3b70221827a --- /dev/null +++ b/2025/DreamOmni_ Unified Image Generation and Editing/ff84d3e1-2391-4334-90b8-2fe54fd66a09_content_list.json @@ -0,0 +1,2258 @@ +[ + { + "type": "text", + "text": "DreamOmni: Unified Image Generation and Editing", + "text_level": 1, + "bbox": [ + 218, + 130, + 751, + 152 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Bin Xia1, Yuechen Zhang1, Jingyao Li1, Chengyao Wang1, Yitong Wang2, Xinglong Wu2, Bei Yu1, and Jiaya Jia3 \n1 CUHK, 2 ByteDance Inc, 3 HKUST \nhttps://zj-binxia.github.io/DreamOmni-ProjectPage/", + "bbox": [ + 243, + 179, + 725, + 252 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/bf70f440b4a435429a3f193d5c3eef55613b916e010b16063c62fdd196b50cd5.jpg", + "image_caption": [ + "Figure 1. The gallery of DreamOmni. DreamOmni, as a native unified image generation and editing model, can handle various tasks." + ], + "image_footnote": [], + "bbox": [ + 83, + 260, + 885, + 594 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Abstract", + "text_level": 1, + "bbox": [ + 233, + 632, + 313, + 647 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "projectpagepCurrently, the success of large language models (LLMs) illustrates that a unified multitasking approach can significantly enhance model usability, streamline deployment, and foster synergistic benefits across different tasks. However, in computer vision, while text-to-image (T2I) models have significantly improved generation quality through scaling up, their framework design did not initially consider how to unify with downstream tasks, such as various types of editing. To address this, we introduce DreamOmni, a unified model for image generation and editing. We begin by analyzing existing frameworks and the requirements of downstream tasks, proposing a unified framework that integrates both T2I models and various editing tasks. Furthermore, another key challenge is the efficient creation of high-quality editing data, particularly for instruction-based and drag-based editing. To this", + "bbox": [ + 73, + 659, + 472, + 901 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "end, we develop a synthetic data pipeline using sticker-like elements to synthesize accurate, high-quality datasets efficiently, which enables editing data scaling up for unified model training. For training, DreamOmni jointly trains T2I generation and downstream tasks. T2I training enhances the model's understanding of specific concepts and improves generation quality, while editing training helps the model grasp the nuances of the editing task. This collaboration significantly boosts editing performance. Extensive experiments confirm the effectiveness of DreamOmni. The code and model will be released.", + "bbox": [ + 496, + 633, + 893, + 799 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "1. Introductian", + "text_level": 1, + "bbox": [ + 500, + 811, + 632, + 828 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Recently, text-to-image (T2I) generative foundation models [6, 15, 27, 39, 44, 47] have made remarkable progress, driving the development of various downstream applications such as image editing, video generation, and more.", + "bbox": [ + 496, + 839, + 893, + 902 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "CVF", + "bbox": [ + 106, + 2, + 181, + 42 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore.", + "bbox": [ + 236, + 0, + 807, + 46 + ], + "page_idx": 0 + }, + { + "type": "page_number", + "text": "28533", + "bbox": [ + 478, + 944, + 519, + 957 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Currently, T2I foundation models face two significant challenges: (1) Adapting these models for downstream applications often requires the integration of various plugins (such as ControlNet [62] and IP-adapter [60]) in different ways, or the extension of input channels (e.g., SD-inpainting [44], InstructP2P [8]). This reliance on specialized frameworks hampers multi-task generalization and complicates deployment. (2) High-quality and accurate editing data is difficult to obtain, including instruction-based editing, drag editing, and subject-driven generation data. In this paper, we propose to unify T2I models with multiple editing tasks, such as instruction-based editing, inpainting & outpainting, drag editing, and reference image generation within a single framework. Additionally, we introduce an efficient synthetic data pipeline for efficiently and accurately constructing editing data, facilitating the training of native unified generation and editing model.", + "bbox": [ + 75, + 90, + 472, + 349 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "We began by exploring fundamental model frameworks to develop a unified framework that is both highly effective and efficient, achieving rapid training convergence. (1) We aligned certain frameworks, such as Unet [44] and DIT [38], to have similar parameter sizes and runtime settings for fair comparison (Fig. 3). Our analysis showed that DIT's effectiveness comes from concentrating most of its computational and parameter load on latent at $2 \\times$ downsampling size, whereas Unet distributes more of its operations at smaller scales. This allocation makes DIT comparatively superior, as concentrating computations at the $2 \\times$ downsampled latent achieves a better trade-off. However, we observed that Unet's residual connections significantly improve the model's training convergence speed. Inspired by the efficiency of high-resolution latent computations, we also incorporated additional residual convolutional networks for the input-scale latent. (2) We replaced the original text encoder with a vision-language model (VLM) to unify visual-language prompt encoding. This encoded prompt was then concatenated with the noisy latent, allowing for integrated computation within the DIT framework.", + "bbox": [ + 75, + 352, + 472, + 670 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Next, we need to obtain accurate training data efficiently for our unified model. For various types of editing, such as instruction-based editing, it's challenging to create and filter pairs of data that precisely match the instructions while avoiding detail corruption. We have discovered that the key to effective editing lies in helping the model understand the meaning of the editing operations, rather than learning specific concepts (a capability the model already possesses in T2I training). Therefore, we propose a synthetic collage data pipeline that easily scales up to generate accurate and diverse edited data, enabling models to better understand the nuances of specific editing tasks (Fig. 2 (b)). (1) For instruction-based editing, we constructed synthetic data for three main categories: addition, removal, and replacement. (2) For drag editing, we utilized synthetic data to cre", + "bbox": [ + 75, + 674, + 472, + 902 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "ate scaling, translating, and rotating editing data. (3) For inpainting and outpainting, we randomly generate masks for each given image. (4) For reference image generation, We divided the tasks into subject-driven generation and image-conditioned generation (i.e., ControlNet-like generation) tasks. For subject-driven generation, we used a canvas filled with stickers, allowing the model to reference a specific sticker for generation, which helps train the model to extract the demanded objects or details from given images for generation. For the image-conditioned generation, we created canny, depth, and segment maps, similar to ControlNet. (5) We use synthetic data to enhance the model's T2I accuracy in responding to attributes such as quantity, position, relationships, color, shape, and text. Concretely, we randomly arrange different quantities of stickers, text, and geometric shapes on a canvas, and obtain precise descriptions of their quantity, position, relationships, and colors based on their exact coordinates.", + "bbox": [ + 496, + 90, + 893, + 362 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "During training, thanks to our unified framework that integrates T2I with multiple editing tasks, we can easily train T2I data alongside various types of editing data, obtaining a native unified image generation and editing model, called DreamOmni. The editing tasks benefit from the T2I data, which helps prevent the model from forgetting specific concepts and generation quality decline, while the T2I task can leverage synthetic data to enhance instruction following.", + "bbox": [ + 496, + 363, + 893, + 484 + ], + "page_idx": 1 + }, + { + "type": "list", + "sub_type": "text", + "list_items": [ + "- We conducted an analysis of existing model frameworks under fair settings and, drawing on the characteristics of different tasks, proposed an efficient and powerful unified image generation and editing framework, DreamOmni.", + "- We introduced a synthetic collage data pipeline to tackle the current inefficiency and difficulty in creating and filtering high-quality editing data. Furthermore, we utilized the synthetic collage data pipeline to enhance the accuracy of the T2I model's output. Experimental results demonstrate that synthetic data is an effective, high-quality, and cost-efficient method to scale up data for achieving unified image generation and editing training.", + "- After unified training with T2I data and various synthetic datasets, our DreamOmni showcases competitive performance on T2I generation and various editing tasks." + ], + "bbox": [ + 500, + 486, + 890, + 712 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "2. Related Work", + "text_level": 1, + "bbox": [ + 500, + 724, + 640, + 739 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Text-to-image Diffusion Model. Diffusion models [1, 5, 14, 21, 23, 24, 35, 52, 53] have emerged as highly effective methods for image generation, outperforming previous generative models like GANs [20] and VAEs [26]. The latent diffusion model [44], also known as Stable Diffusion, enhances the diffusion process within the image latent space and utilizes cross-attention to guide image generation through text input. This approach enhances the usability of the image generation model, driving its widespread adoption. Subsequently, more advanced text-to", + "bbox": [ + 496, + 750, + 893, + 902 + ], + "page_idx": 1 + }, + { + "type": "page_number", + "text": "28534", + "bbox": [ + 478, + 944, + 519, + 957 + ], + "page_idx": 1 + }, + { + "type": "image", + "img_path": "images/79c9ce9b16fe1fcd7f8e050f03f30082678e6dd7c75e3b2c475751e3f4660543.jpg", + "image_caption": [ + "(a) The framework of DreamOmni", + "Figure 2. The overview of DreamOmni. (a) The DreamOmni framework supports unified image generation and editing, with fast training convergence and powerful performance. (b) To overcome the difficulty and inefficiency in data creation and filtering for image editing, we propose a collage-based synthetic data pipeline. This pipeline enables the efficient creation of data for various editing tasks, such as adding, deleting, and replacement operations in instruction-based editing, as well as translation, scaling, and rotation in drag editing. Additionally, it supports reference image generation and segmentation & detection. Furthermore, our synthetic data pipeline enhances the accuracy of T2I generation. Due to space limitations, we have optionally shown the corresponding prompts or instructions for these cases." + ], + "image_footnote": [], + "bbox": [ + 86, + 89, + 398, + 378 + ], + "page_idx": 2 + }, + { + "type": "image", + "img_path": "images/c8f3b7938d763df3eae9a1281a2554c79ef16137b19a0612ebc783d7f4712413.jpg", + "image_caption": [ + "(b) The examples of synthetic data" + ], + "image_footnote": [], + "bbox": [ + 406, + 90, + 887, + 373 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "image (T2I) models were developed, such as Imagen [47], DALLE [43], and others [3, 6, 13, 16, 39]. Recently, the DIT frameworks [4, 38] has become increasingly popular as more models [10, 15, 31] transition from the Unet framework [45], thanks to its straightforward structure and outstanding compatibility with multimodal inputs. However, the current T2I model fails to account for the design of downstream editing tasks, necessitating numerous specialized adaptations. For instance, Instructp2p [8] requires an expansion of input channels, while ControlNet [62] and IP-adapter [60] rely on additional plugins and various methods for injecting reference image information. These fragmented designs are inconsistent with the trend toward unified models, complicating deployment and impeding the joint training of T2I and editing tasks. In this paper, we will analyze the existing T2I frameworks and propose a more efficient and powerful unified generation and editing model.", + "bbox": [ + 75, + 491, + 468, + 747 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Multimodal-guided Image Editing and Generation offers richer guiding information and has broader applications [2, 7, 9, 12, 37, 54, 55] than T2I models. It encompasses instruction-based editing [8, 18, 25, 49, 61], inpainting & outpainting [34, 57, 59], drag editing [32, 51], and reference image generation [28, 29, 40, 62, 63]. A significant challenge in editing, especially in instruction-based and drag editing, lies in creating and filtering accurately paired editing data. For example, InstructP2P [8] utilizes finetuned GPT-3 and a Prompt-to-Prompt strategy [22] to", + "bbox": [ + 75, + 750, + 470, + 901 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "generate instruction-based editing datasets, but it achieves a success rate of under $15\\%$ with frequent artifacts and detail loss. Additionally, filtering high-quality editing data is a persistent challenge, complicating scaling up. Magicbrush [61] addresses this by employing people to create and filter editing data, but this manual method limits data volume. Furthermore, for drag editing, most methods [32, 51] are training-free, which restricts their performance and efficiency. To this end, InstaDrag [50] leverages video data to construct drag editing data. However, InstaDrag still faces challenges with inefficiency in filtering valid data from video. Moreover, for subject-driven generation [28, 36], DreamBooth [46] requires training on several subject-related images before each inference, which is inconvenient and inefficient. Some works [30, 58] collect multiple photos of the same person from the internet to build datasets for training more efficient and powerful subject-driven generation models. However, the process of gathering and filtering such data still presents a challenge, limiting the scaling up of the dataset. In this paper, we propose a simple, efficient, and accurate synthetic collage data pipeline to create quite a few data for unified training.", + "bbox": [ + 496, + 491, + 890, + 824 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3. Methodology", + "text_level": 1, + "bbox": [ + 500, + 842, + 633, + 859 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "The unification of multiple tasks is a trend and pursuit in the field of computer vision and AI, which not only en", + "bbox": [ + 498, + 869, + 890, + 900 + ], + "page_idx": 2 + }, + { + "type": "page_number", + "text": "28535", + "bbox": [ + 478, + 944, + 517, + 955 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "hances model usability and reduces deployment complexity but also enables collaborative training that fosters synergies between tasks. However, current T2I foundation models are primarily designed specifically for T2I and often overlook the potential for integration with other tasks, such as various image editing tasks. To this end, we propose DreamOmni, a unified model for image generation and editing. We design and train DreamOmni from three aspects. (1) In Sec. 3.1, we compare various frameworks in a fair setting and design a powerful and fast training converging framework that supports unified multi-tasking based on the characteristics of different tasks. (2) In Sec. 3.2, we aim to propose a convenient, efficient, and accurate synthetic data pipeline for data scaling up to facilitate multi-task unified training and enhance the model's instruction-following ability. (3) In Sec. 3.3, we introduce the training scheme for DreamOmni. Native unified training of T2I and various editing tasks prevents concept forgetting and generation quality decline, while enhancing the model's editing and prompt-following capabilities.", + "bbox": [ + 75, + 90, + 472, + 395 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "3.1. Framework", + "text_level": 1, + "bbox": [ + 76, + 406, + 205, + 420 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "In this part, we aim to design a unified and powerful image generation and editing framework. Currently, different editing models often have distinct structure designs. For example, IP-adapter [60] and BLIP-Diffusion [28] inject information through cross-attention to maintain subjects. In contrast, InstructP2P [8] achieves editing consistency by adding different numbers of input channels for models. These structures are tailored for specific tasks and lack generalizability. To this end, as shown in Fig. 2 (a), we concatenate VLM features with noisy latent and input them into a DIT Block for joint multi-head self-attention operations. After that, the VLM features and noisy latent are processed by the FeedForward modules. This allows the model to autonomously learn any level features (from overall consistency to subject consistency) for editing and generation. Notably, for FeedForward modules, we separate the VLM features and noisy latent, passing them through two distinct FeedForward modules with the same network structure. Additionally, instead of using CLIP [41] or T5 [42] as the text encoder, we introduce a Vision-Language Model (VLM) that enables joint understanding and encoding of both image and text prompts. For tasks that require high consistency, such as instruction-based editing and drag editing, we input the VAE encoding of the source image into the DIT model to ensure strong consistency between the non-edited regions of the output image and the source image.", + "bbox": [ + 75, + 431, + 472, + 824 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "In the current framework design, some works, such as DIT [38], are compared under label-conditioned generation rather than T2I. However, T2I is inherently more complex than label-based generation, as it requires the integration and understanding of complex prompts. Additionally, many", + "bbox": [ + 75, + 825, + 472, + 902 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "T2I models [15, 44], such as SDXL [39], are trained with different model sizes, datasets, and training settings. This variability makes it challenging to assess the impact of different model components on overall performance. Moreover, SDXL incorporates many Transformer blocks within its UNet structure. So, why does DIT outperform SDXL? To address this, we conducted extensive experiments as shown in Fig. 3. We observe that DIT surpasses Unet because it allocates most of its computation to $2 \\times$ downsampled latent, whereas Unet allocates a greater proportion to $4 \\times$ downsampled latent. Since attention operations on $1 \\times$ latent cause memory burden, we further employ residual Conv blocks to refine generation details for $1 \\times$ latent.", + "bbox": [ + 496, + 90, + 893, + 287 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Additionally, we observed that the use of long connections in the UNet framework can significantly accelerate the model's training convergence without compromising performance. As shown in Fig. 2 (a), we concatenate early and later features along the channel dimension and apply a linear layer to combine the two features. Notably, linear layers used for VLM features and noisy latent are different.", + "bbox": [ + 496, + 287, + 893, + 395 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "3.2. Synthetic Data", + "text_level": 1, + "bbox": [ + 500, + 407, + 650, + 422 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "In addition to the unified framework, we also require quite a few data to support joint training. While T2I data is readily available, creating and filtering accurate, high-quality data for tasks like instruction-based editing is far more challenging. To address this, we introduce a synthetic collage data pipeline that efficiently and accurately generates the required editing data. As shown in Fig. 2 (b), our pipeline encompasses six tasks. Notably, this is not the full extent of our synthetic pipeline's capabilities; it is also capable of handling more complex task combinations.", + "bbox": [ + 496, + 431, + 893, + 582 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "- T2I generation. As shown in Fig. 2 (b), in addition to the conventional T2I data, we further enhance the model's performance on the T2I by incorporating synthetic data, specifically focusing on improving the text, shape, position, quantity, and color generation. Specifically, for the text, we randomly generate words or short phrases on a blank canvas using a variety of fonts, colors, thicknesses, and sizes. For the shape and quantity, we randomly create geometric shapes with different quantities, colors, and sizes, and arrange them on the canvas. Based on these attributes and their positions, we generate accurate prompts, which are then refined using a LLM. Furthermore, we use a diverse set of stickers and segmentation data for synthesis, placing them on the canvas and calculating their precise spatial relationships. These prompts are then created and further polished by the LLM to produce descriptions that are more natural.", + "bbox": [ + 500, + 583, + 893, + 839 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "- Inpainting & Outpainting. We randomly generate masks for smearing, blocks, and image edges. Notably, during training, in addition to feeding the masked image and its corresponding mask into the VLM for encoding, we also", + "bbox": [ + 500, + 840, + 893, + 900 + ], + "page_idx": 3 + }, + { + "type": "page_number", + "text": "28536", + "bbox": [ + 478, + 945, + 519, + 955 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "include the image caption with a $50\\%$ probability.", + "bbox": [ + 89, + 90, + 421, + 106 + ], + "page_idx": 4 + }, + { + "type": "list", + "sub_type": "text", + "list_items": [ + "- Instruction-based editing. We categorize the task into three operations: addition, removal, and replacement. For both removal and replacement, we randomly select a background image and an object image to create the source image. In the removal case, the target image is simply the background image. For replacement, the target image is generated by replacing the object in the same location with a different object. Notably, for the addition, due to the need for the added object to be placed in a contextually appropriate position relative to the background, we use a blank background in the paper.", + "- Drag editing. We categorize the data into three types: translation, scaling, and rotation. Notably, Instadrag [50] treats each pair of drag points as a separate image, which is sparse and impractical due to the fixed number of drag points required. Therefore, we represent each drag point using the format $(x,y,dx,dy)$ as a prompt input, where $x$ and $y$ denote the coordinates of the drag points in the source image, and $dx$ and $dy$ represent the translation vector. Furthermore, we normalize these coordinates by dividing them by the image's width or height.", + "- Reference image generation. We categorize the data into two types: image-conditioned generation, akin to ControlNet [62], and subject-driven generation. For image-conditioned generation, we begin by selecting high-quality images and creating corresponding canny maps, depth maps, and segmentation masks as source images for training. For subject-driven generation, we synthesize source images and then randomly select objects within these images to create target images. The model is trained to generate new content based on specific attributes referenced from the source image, allowing for the flexible generation of varied scenes and subjects.", + "- Segmentation & detection. We randomly selected a background image and an object image. These images are then composited together to create a source image. Following this, we apply color manipulation or draw a bounding box around the object region based on the alpha channel of the object image, obtaining the target image." + ], + "bbox": [ + 76, + 107, + 470, + 696 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Our synthetic collage data pipeline is efficient and precise, enabling the generation of billions of diverse images for large-scale pretraining and fine-tuning of DreamOmni.", + "bbox": [ + 75, + 696, + 468, + 742 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "3.3. Model Training", + "text_level": 1, + "bbox": [ + 76, + 755, + 235, + 771 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "After careful consideration, we chose a $2.5B$ parameter size for DreamOmni's DIT model. This size strikes a good balance between being user-friendly and ensuring powerful performance. For the VLM encoder, we directly adopt the Qwen2-VL [56] 7B model, based on three reasons: (1) It supports image inputs of arbitrary resolution, (2) It delivers strong model performance, and (3) It is released under a permissive open-source license. The VLM features are derived", + "bbox": [ + 75, + 780, + 470, + 902 + ], + "page_idx": 4 + }, + { + "type": "image", + "img_path": "images/44c58641be4599ab2285e3cabd82a6d9c636faa9477dc1d5d4fc07044562d205.jpg", + "image_caption": [ + "Figure 3. Comparison of different frameworks. The left figure shows the FID comparison among different frameworks, while the right table shows their number of parameters and runtime." + ], + "image_footnote": [], + "bbox": [ + 501, + 89, + 696, + 202 + ], + "page_idx": 4 + }, + { + "type": "table", + "img_path": "images/29467ed51cde2f280c59cddb5041660156f912761b6b9cf218c86bf7ade5ab5b.jpg", + "table_caption": [], + "table_footnote": [], + "table_body": "
FrameworkParam (M)Runtime (ms)
SDXL [39]83828.59
Pixart [10]86234.61
SD3 [15]89625.73
DreamOmni-V182828.32
DreamOmni-V283625.14
DreamOmni-V3 (Ours)82727.04
", + "bbox": [ + 710, + 117, + 887, + 178 + ], + "page_idx": 4 + }, + { + "type": "table", + "img_path": "images/2d6f645ad4eab1bcdd06cca8d1288b62a8e0360986c5c23f599b2cee74af7274.jpg", + "table_caption": [ + "Table 1. Comparisons on GenEval [19]. Our model outperforms all current open models (including SOTA SD3-Medium [15])." + ], + "table_footnote": [], + "table_body": "
ModelObjectsCountingColorsPositionColor Attribution
OverallSingleTwo
SD 1.5 [44]0.430.970.380.350.760.040.06
PixArt-alpha [10]0.480.980.50.440.80.080.07
SDv2.10.50.980.510.440.850.070.17
SDXL [39]0.550.980.740.390.850.150.23
SD-cascade0.520.980.570.470.870.080.15
SD3-Medium [15]0.700.990.840.630.880.280.55
DreamOmni (Ours)0.700.990.810.650.880.340.54
", + "bbox": [ + 503, + 287, + 892, + 393 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "from the penultimate layer of Qwen2-VL. Additionally, we use the FLUX-schnell's VAE as DreamOmni's VAE, which retains more latent channels, enabling the model to capture finer image details. Furthermore, we optimize DreamOmni using Rectified Flow [33], which performs the forward process by linearly interpolating between noise and data along a straight trajectory. we train DreamOmni using loss $\\mathcal{L}$ :", + "bbox": [ + 496, + 412, + 892, + 518 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {L} = \\mathbb {E} \\left(\\left\\| (\\mathbf {z} - \\epsilon) - v _ {\\theta} \\left(\\mathbf {z} _ {t}, \\mathbf {c}, t\\right) \\right\\| _ {2} ^ {2}\\right), \\tag {1}\n$$\n", + "text_format": "latex", + "bbox": [ + 573, + 523, + 890, + 549 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "where $\\mathbf{z}_t = t\\mathbf{z} + (1 - t)\\epsilon$ represents the noised feature map at timestep $t$ . Ground truth images are encoded by VAE into latent space to derive $\\mathbf{z}$ . Here, $\\epsilon \\in \\mathcal{N}(0,\\mathbf{I})$ represents Gaussian noise, and $v_{\\theta}$ refers to the DIT model. $\\mathbf{c}$ indicates the conditional information.", + "bbox": [ + 496, + 551, + 893, + 627 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "The T2I training dataset consists of $125M$ images, including both the LAION dataset $(103M)$ [48] and our collected data $(22M)$ . These images have been annotated using InternVL2 [11]. For synthetic data, such as T2I, instruction editing, inpainting & outpainting, drag editing, and reference image generation, we generated $12M$ images respectively, bringing the total to approximately $60M$ images. Additionally, for segmentation & detection, we created $8M$ images. Prior to training, to improve training speed usage, we first encoded prompts into VLM features.", + "bbox": [ + 496, + 628, + 893, + 779 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Our training process is divided into three stages. In the first stage, we train on images with $256 \\times 256$ size, a batch size of 2048, a learning rate of $1 \\times 10^{-4}$ , and $377K$ iterations. In the second stage, we train on images with $512 \\times 512$ size, a batch size of 1024, a learning rate of $5 \\times 10^{-5}$ , and $189K$ iterations. In the final stage, we select 12M high-quality T2I data and randomly sample 1M high-quality images from each type of synthetic data for training.", + "bbox": [ + 496, + 780, + 893, + 902 + ], + "page_idx": 4 + }, + { + "type": "page_number", + "text": "28537", + "bbox": [ + 478, + 944, + 519, + 957 + ], + "page_idx": 4 + }, + { + "type": "image", + "img_path": "images/9c60bf850c8a95d2f6eb26717927c3f612a9e003f8a092bb85b7ba8a5b6bd738.jpg", + "image_caption": [], + "image_footnote": [ + "A heartwarming 3D rendered scene of an elderly farmer and a tiny orange kitten dressed in matching denim overalls. The farmer, with a gentle smile, walks alongside the kitten in a lush, green garden filled with thriving plants, showcasing a fruitful harvest. The intricate details of the overalls and the farmer's worn, weathered face tell a story of years spent tending to the land." + ], + "bbox": [ + 250, + 94, + 367, + 185 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/a36710f3359dedcc844ff90df08ef9716fb7c4f2dbeab84cdf8898a79eef3233.jpg", + "image_caption": [], + "image_footnote": [ + "Deep in the enchanted forest lives a woman who is the moon fairy. Her long blonde hair shines in the starlight, tangled with her flowers that glow with a soft blue glow. Her eyes are the color of the night and shine with the magic of the night. The fairy wears a dress made of moon petals, woven with threads of moonlight that shine with an iridescent glow, a crown of stars adorns her head, shining with the light of the full moon that illuminates the forest. Her wings are translucent like glass, with a pale glow reminiscent of the glow of the moon. HD, 16K, photo, cinematic, poster" + ], + "bbox": [ + 250, + 188, + 367, + 280 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/eabff75628630a192c7f6174f4723c1a442c0339e4cd675ded4fada74f41ca56.jpg", + "image_caption": [ + "SD1.5" + ], + "image_footnote": [ + "A unique, intricately detailed creature resembling a reptile, possibly a lizard or a gecko. It has a vibrant blue and green scaled body, with large, round, and expressive eyes that are a deep shade of blue. The creature is adorned with a floral crown made of various flowers, including pink roses and blue blossoms, as well as pinecones. The backdrop is a soft, blurred forest setting, suggesting a serene and mystical ambiance." + ], + "bbox": [ + 250, + 284, + 367, + 376 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/d98b52f9a32a072e3fe5d58538399420e9ee0919efb756ae921862b99d4ed7de.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 374, + 94, + 491, + 185 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/19c02e74acd35645c68fb7aefd3a7876c7c76afe4ccbbcd12d4280232a0ae5f4.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 374, + 188, + 491, + 279 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/c593cd696b0d44bb9efa3739f3c8506d12b7e4b16e25b1c29222a505527892fb.jpg", + "image_caption": [ + "SD-Cascade" + ], + "image_footnote": [], + "bbox": [ + 374, + 284, + 491, + 375 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/86862a86d59f2f5fecf5efbecb8420c04196a9cae40f48d221840437e745dc73.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 500, + 94, + 617, + 184 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/630c98528f6cef16ce49b32151d93dee68848f9d2b70a97f2a8c3f2c3c1cd14b.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 500, + 188, + 616, + 279 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/64de70a1c52d4e34de7c27a9ee6e2c76bf8b2d75e6bb502d3fb4be18c0d23f14.jpg", + "image_caption": [ + "SDXL" + ], + "image_footnote": [], + "bbox": [ + 500, + 284, + 617, + 375 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/28ebbdf421691bb0c275448e82a3881af5f7991bf577b9a896562c4fc993a8c7.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 625, + 94, + 743, + 184 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/b3704c9f83d07fdaa18dfbaf01952ad70ba88df830b969a5053dd7583cb86dbc.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 625, + 189, + 741, + 279 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/a5905d27e5255b2c57cc24d52e74b989416583e20f2408ea5b75ad98290d5ba5.jpg", + "image_caption": [ + "SD3-Medium" + ], + "image_footnote": [], + "bbox": [ + 625, + 284, + 741, + 375 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/5241c5395538de762335a1ceb954535adad35a385efa25ae6a3e343218157c4b.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 750, + 94, + 867, + 184 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/d50cfe5b1598ca98a282138c143de6f5ed3c709ca3bcc97e484d4f0a6078bc38.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 750, + 189, + 867, + 279 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/b146b33a9948c008b9eede92c808b61ea54b4987bd4cadc1358f34c9021d493e.jpg", + "image_caption": [ + "DreamOmni (Ours)" + ], + "image_footnote": [], + "bbox": [ + 750, + 284, + 867, + 375 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/fc372607486dac8f2444b979e802e1229e914be714e81550d6d0d2a681023a11.jpg", + "image_caption": [ + "Prompt" + ], + "image_footnote": [], + "bbox": [ + 86, + 464, + 184, + 540 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/df26f9c61d877b45f3074187783722826d34dc331107cd69c99e4ed30e708f71.jpg", + "image_caption": [ + "Masked Image" + ], + "image_footnote": [], + "bbox": [ + 88, + 542, + 184, + 618 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/44a3c553e8aa42dbaec3f56e1ebbce95dac706a3a457cd10e93bb049c016cf0a.jpg", + "image_caption": [ + "Figure 4. Visual comparison on T2I generation. Compared to other competitive methods (including SD3-Medium [15], SDXL [39], SD-Cascade, and SD1.5 [44]), our DreamOmni not only better adheres to user prompts but also generate more visually appealing results with delicate details, elegant composition, and so on." + ], + "image_footnote": [], + "bbox": [ + 186, + 465, + 282, + 540 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/a6348fcb43f06e64f55e50418c9c155719d8c67a855950446df76add04a405bd.jpg", + "image_caption": [ + "ControlNet Inpainting" + ], + "image_footnote": [], + "bbox": [ + 186, + 542, + 282, + 618 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/63f07d40f1c7c588f59765f17eaf0603bc21f25f56016545d4397e43575423a8.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 285, + 465, + 380, + 540 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/1d484fd1dacff2233c71294d957c63143aeb78a78a36d3a50c8d8e5327b4e6cd.jpg", + "image_caption": [ + "SD Inpainting" + ], + "image_footnote": [], + "bbox": [ + 285, + 542, + 382, + 618 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/7944ebbe070f47a98f8fe25d76bd2c8a624b665b21c92239a06b17ae05d4e025.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 383, + 465, + 480, + 540 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/55c1a929f02a64bfdb5f5b747b74085b99d7a886ca8f41d8d9b0c5f2be9837a4.jpg", + "image_caption": [ + "DreamOmni" + ], + "image_footnote": [], + "bbox": [ + 385, + 542, + 480, + 618 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/c3f8a16228c0d87658cdc71237cdf9aa8639df763950c1df5712eccbc56fda42.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 491, + 465, + 586, + 540 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/75e8040ab91ac2c4c85a4201768648c99b9057cc04444c34dc6ffde6fea71e47.jpg", + "image_caption": [ + "Masked Image", + "Figure 5. Visual comparison on inpainting & outpainting between DreamOmni, ControlNet-Inpainting [62] and SD-inpainting [44]." + ], + "image_footnote": [], + "bbox": [ + 491, + 542, + 586, + 618 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/8bb81f02630bd43ec637e10d158244f669982a9a47c77ed4809f936eeca84a9d.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 589, + 465, + 684, + 540 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/51ec26871e80adffb8ce709f78b0a8c022c985b30f9dc915f34a17d734d956b9.jpg", + "image_caption": [ + "ControlNet Inpainting" + ], + "image_footnote": [], + "bbox": [ + 589, + 542, + 684, + 618 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/a5e8e081ae3e2ce2b20a07485c536e0988ed2d0286267adf01c053bc378f722d.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 687, + 465, + 782, + 540 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/00311f0d0da94a5007156fcc982ee548bc98a4cb4daf938907b3c1add935292a.jpg", + "image_caption": [ + "SD Inpainting" + ], + "image_footnote": [], + "bbox": [ + 687, + 542, + 782, + 618 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/df38ee661153399d24cdbbade590ea3dec06c0e5b769793eda28fa91fd424280.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 785, + 465, + 882, + 540 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/3faede984f7722c407634dbfae9c4391608161546e5dfc9968ba4f3db535fdb0.jpg", + "image_caption": [ + "DreamOmni" + ], + "image_footnote": [], + "bbox": [ + 785, + 542, + 882, + 618 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "We train models with $1024 \\times 1024$ size, a batch size of 256, a learning rate of $2 \\times 10^{-5}$ , and $140K$ iterations. All experiments are conducted on 64 A100 GPUs. Additionally, to enable the model to generate images at varying resolutions, similar to the approach used in SDXL [39], we divide the images into 31 buckets based on their aspect ratios, ranging from $4:1$ to $1:4$ , during training.", + "bbox": [ + 75, + 686, + 468, + 792 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "4. Experiments", + "text_level": 1, + "bbox": [ + 76, + 800, + 209, + 816 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Evaluation on Framework. We compared several T2I model frameworks under similar settings to identify effective components. Using the same VAE, CLIP text encoder, parameters, runtime, and LAION training/testing datasets, we evaluated the Unet-based SDXL [39], DIT-", + "bbox": [ + 75, + 824, + 468, + 898 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "based Pixart [10], SD3-Medium [15], and our developed DreamOmni variations. Notably, as shown in Fig. 3, to facilitate comparison, we do not use the full $2.5B$ parameters of DreamOmni, but instead adjust the parameters of all models to $0.85B$ . DreamOmni-V1 has two downsampling layers ( $2\\times$ and $4\\times$ ) like SDXL but lacks an Unet connection. DreamOmni-V2 builds on DreamOmni-V1 by adding the Unet connection, while DreamOmni-V3 takes DreamOmni-V2 further by focusing all DIT operations on $2\\times$ downsampled latent. (1) Models with the Unet connection (SDXL, DreamOmni-V2, DreamOmni-V3) showed significantly faster convergence than those without (SD3-Medium, DreamOmni-V1). DreamOmni-V3, notably, converges four times faster than the SD3-Medium, enhancing", + "bbox": [ + 496, + 686, + 890, + 898 + ], + "page_idx": 5 + }, + { + "type": "page_number", + "text": "28538", + "bbox": [ + 478, + 944, + 517, + 955 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/828f4d95ba889ae4b9d04ba9e6f79da3c1518fc6e78677b058d2a28f2d84c97c.jpg", + "image_caption": [ + "A humanoid cat from the race american shorthair with black fur. Standing as a barbarian tabaxi with a greatsword. Dressed in light clothing, something like the ancient geeks weared, photorealistic. standing in front of an oasis in the dessert." + ], + "image_footnote": [], + "bbox": [ + 191, + 92, + 285, + 166 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/26dcdb6d7cfc5e275ac15e558032eb3569c54ae11a605d06bebdeb82c5f1188a.jpg", + "image_caption": [ + "elf girl travelling alongside a majestRyuic dragon, close-up view, anime style, high fantasy, digital art, masterpiece painting", + "Image", + "Condition" + ], + "image_footnote": [], + "bbox": [ + 191, + 170, + 285, + 244 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/58f6e08643e10e515a0b16dcc3836f09d102f5ba42f2f583e250ecdee69c9ec6.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 289, + 92, + 383, + 166 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/d7cc60c44731c166bae4251f1a3dc113b280698fdca951d03280dade47264e1e.jpg", + "image_caption": [ + "ControlNet" + ], + "image_footnote": [], + "bbox": [ + 289, + 170, + 383, + 244 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/f1125c81febc4efb17f152ce3de1408a5b9369fbad67999ae27170671a1ef680.jpg", + "image_caption": [ + "A cute teenage-child fairy with blue hair and butterflies in her ears, holding balloons and a birthday cake with a candle lit on it in her other hand. Her hand is holding the plate. The artwork is in the style pixar anime. her facial expression is very happy and excited but in a cute way" + ], + "image_footnote": [], + "bbox": [ + 385, + 93, + 483, + 166 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/9dd97af1fa68a90233a7857ba19ae9d301934d958c03038bc8630d1c1eae324c.jpg", + "image_caption": [ + "DreamOmni", + "(Ours)", + "Prompt" + ], + "image_footnote": [], + "bbox": [ + 385, + 170, + 483, + 244 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/6a28139b11cfc28452b984717b712e5a436905c3c254b81eed52622643c97945.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 586, + 92, + 681, + 166 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/5f236f35dd1ff78053a54c74cd5172dffa72abcf20bbbd0e4750a8608ac84027.jpg", + "image_caption": [ + "ironman is riding a white little pony", + "Image", + "Condition" + ], + "image_footnote": [], + "bbox": [ + 586, + 170, + 681, + 244 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/b86641f75a31a3ea4edb1d74175b54aa1e84d4c66c25335a837126047a6d3752.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 684, + 92, + 777, + 166 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/9fff3d28baf0402e201d513cdec8bc8d0439c2f79d63ae299bf85a10326b51d2.jpg", + "image_caption": [ + "ControlNet" + ], + "image_footnote": [], + "bbox": [ + 684, + 170, + 777, + 244 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/8c3e9d93f8ffec9dc5f980640ac62d4d8280c232679ade8c4ad77b61f0fb0f81.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 781, + 92, + 877, + 166 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/6316ebfc0ed010e05c69008c1d5a9e033a7937675262fcba8f771ebdf77b37b8.jpg", + "image_caption": [ + "DreamOmni", + "(Ours)" + ], + "image_footnote": [], + "bbox": [ + 781, + 170, + 877, + 244 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/ce23874b6bc3822c78a52cf1f92f4c5442e5394d7e0835485f2e51433bf03367.jpg", + "image_caption": [ + "Prompt", + "A girl (in the given image) is playing a red electric guitar. She is wearing a black jacket and a banie, and she has headphones. The background features autumn leaves, suggesting a seasonal setting.", + "Generate an image using the character from the supplied image as reference. A girl is holding a lantern. She is surrounded by vibrant flowers and a whimsical, magical atmosphere. The girl is wearing a blue and white outfit with a ruffled collar, and she has a thoughtful expression." + ], + "image_footnote": [], + "bbox": [ + 176, + 335, + 230, + 398 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/71d6401569b886c3669726ff3af16b5a21a45d87a364d700d75c1175a4fad111.jpg", + "image_caption": [ + "Prompt", + "Image condition", + "Figure 7. Visual comparison on subject-driven generation. Compared with BLIP-Diffusion [28] and IP-adapter [60], DreamOmni excels at accurately following user prompts while preserving the specified subject. Furthermore, its success across both photo and anime cases underscores DreamOmni's generalization capabilities." + ], + "image_footnote": [], + "bbox": [ + 179, + 402, + 228, + 468 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/16ae7a1dd5f5fb65f7045c28e903c928bc95e9a424f69eaad6309229bd461333.jpg", + "image_caption": [ + "Figure 6. Visual comparison on image-conditioned generation. Our DreamOmni excels at using canny, depth, and segmentation maps as conditions for image generation. Compared to the classic ControlNet [62], DreamOmni not only follows the user's prompts and image conditions more accurately but also generates content and color schemes that are visually more pleasing." + ], + "image_footnote": [], + "bbox": [ + 243, + 335, + 323, + 398 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/29f2aa46a034fd5b8519eeadea1b1b6647695637649b4b09bc429ef85c335a59.jpg", + "image_caption": [ + "BLIP-Diffusion" + ], + "image_footnote": [], + "bbox": [ + 243, + 402, + 323, + 468 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/5a999d537b19673d8251107ad71754ae76bffbc83a62e3bfa3ace6c256273ef4.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 326, + 335, + 406, + 398 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/55b9589255c538909e1ce964bcf9e95e60c83551b93b2619aed607cf1bbbd2cf.jpg", + "image_caption": [ + "IP-adapter" + ], + "image_footnote": [], + "bbox": [ + 326, + 402, + 408, + 468 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/17fbb8252ee01bae475fa6c3dda431ef06b63f3009f8de0afdc9fabf40238dca.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 408, + 335, + 491, + 398 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/1c302bb80fc9aa2d8507c0b7e2497803e94efc2126323b5e19cd89aea8723c3b.jpg", + "image_caption": [ + "DreamOmni", + "(Ours)" + ], + "image_footnote": [], + "bbox": [ + 408, + 402, + 491, + 468 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/2e3935f859f63a69b1b29076674734db78a7a57aff7772ecdf3bf5e8d1666adf.jpg", + "image_caption": [ + "A man (in the given image) is wearing a white t-shirt and white pants. He is smiling." + ], + "image_footnote": [], + "bbox": [ + 560, + 335, + 625, + 398 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/e33c2d93ffff5a82d96ea400c5aeae17ddd9374a2323de34025f89940de84673.jpg", + "image_caption": [ + "A close-up of a woman (in the given image) with her head tilted slightly to the side.", + "Prompt", + "Image Condition" + ], + "image_footnote": [], + "bbox": [ + 560, + 402, + 622, + 468 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/e7d1a78cf261fcb7e43495e79540c1b3fb005e50ec5df9693ac8a76d247346b9.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 635, + 335, + 715, + 398 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/45833f4e0aeb40caf13eab52f03c1d41ccecffd7f1a7477913bd21031894b971.jpg", + "image_caption": [ + "BLIP-Diffusion" + ], + "image_footnote": [], + "bbox": [ + 635, + 402, + 715, + 467 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/266f72eaee2b4599407b8fd9ee640e181ca837262dc687317c3aad7da75b7a05.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 718, + 335, + 797, + 398 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/08e9a6fce616c8cfcf902f6e6fe893cc7e1261092c035db857ace89280b2edb8.jpg", + "image_caption": [ + "IP-adapter" + ], + "image_footnote": [], + "bbox": [ + 718, + 402, + 797, + 467 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/86df605e36e088a9b243442809bbb836755f08b9c5295b24865ca53516c645a5.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 800, + 335, + 883, + 398 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/d29fd01cf1cc4418b0cce699238d9dfb532335ea6f8cfbad690beeba0e7be888.jpg", + "image_caption": [ + "DreamOmni", + "(Ours)" + ], + "image_footnote": [], + "bbox": [ + 800, + 402, + 883, + 467 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "both training and fine-tuning efficiency. (2) Comparing DreamOmni-V3 and DreamOmni-V2, we found that concentrating DIT block computations on higher-resolution latent $(2\\times)$ is more cost-effective.", + "bbox": [ + 75, + 561, + 468, + 619 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Evaluation on T2I generation. As shown in Tab. 1, our synthetic data significantly enhances DreamOmni's T2I generation capabilities in aspects such as quantity, color, and position, enabling our model to achieve SOTA results on GenEval [19]. Notably, SD3-Medium [15] is a $2B$ open-source SOTA T2I model, which has a similar parameters as our DreamOmni. Furthermore, qualitative results are shown in Fig. 4. We can see that DreamOmni's outputs are not only more visually pleasing but also align more accurately with the given prompts.", + "bbox": [ + 75, + 625, + 468, + 776 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Evaluation on Inpainting. We compare our DreamOmni with ControlNet-Inpainting [62] and SD-Inpainting [44] on our high-quality evaluation datasets to evaluate its performance. The quantitative results, presented in Tab. 2, demonstrate that DreamOmni significantly outperforms both ControlNet-Inpainting and SD-Inpainting, highlighting its superior generation quality and coherence. Visual results (Fig. 5) further emphasize that DreamOmni ex", + "bbox": [ + 75, + 779, + 468, + 900 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "cels in generating fine details, surpassing both ControlNet-Inpainting and SD-Inpainting. Additionally, DreamOmni effectively handles large mask holes, generating realistic content instead of blurry and incoherent outputs.", + "bbox": [ + 496, + 561, + 890, + 621 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Evaluation on Reference Image Generation. For image-conditioned generation, we conducted a comparison with ControlNet [62] on canny, depth, and segmentation image conditions. Visual results are presented in Fig. 6. These results reveal that, across all tested conditions, our DreamOmni significantly outperforms ControlNet. Our method not only adheres more faithfully to both the image condition and the prompt but also demonstrates enhanced visual quality with better composition and richer detail.", + "bbox": [ + 496, + 625, + 890, + 758 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "For subject-driven image generation, we compared our method with competitive approaches, including BLIP-Diffusion [28] and IP-Adapter [60]. To demonstrate DreamOmni's powerful generalization, we validated their performance on both anime and photographic images. Visual results are shown in Fig. 7. Compared to other methods, our DreamOmni not only preserves the specified subject effectively but also follows the prompt well.", + "bbox": [ + 496, + 762, + 890, + 882 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Evaluation on Instruction Editing. We compare", + "bbox": [ + 500, + 885, + 890, + 901 + ], + "page_idx": 6 + }, + { + "type": "page_number", + "text": "28539", + "bbox": [ + 478, + 944, + 517, + 955 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/26ac1993b18c5d5fde0b92a4e9abcdedef56978bce85191dbf35d6605508dcec.jpg", + "image_caption": [ + "Remove the star ornament from the snow-covered tree." + ], + "image_footnote": [], + "bbox": [ + 151, + 93, + 235, + 157 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/54c14c2dc65bf48785e3855b8126a9d12e462c98130142bc0feba4b036062dfa.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 235, + 93, + 316, + 157 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/6ce6d807399980032fd04191a4813c2c4e47a85f36f070cd15743a33360f89f5.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 316, + 93, + 400, + 157 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/b2055dde6260cd53640cad3ed66a63efab8d44c2af4ce21c54d7e4f4747e5545.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 400, + 93, + 483, + 157 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/f574abba04f76f8f6d4d4d2e623b9ab8c62cb7f14eaedee413ffe33bdebc875f.jpg", + "image_caption": [ + "Add a straw hat on the person's head.", + "Instruction", + "Input Image" + ], + "image_footnote": [], + "bbox": [ + 153, + 160, + 235, + 224 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/80aaffea2850bbcc03f2b6acc282035b9ad1aee9cae49ffc1ae4ec74be7dbfbb.jpg", + "image_caption": [ + "InstructP2P" + ], + "image_footnote": [], + "bbox": [ + 235, + 160, + 316, + 224 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/96fca43af04d42f865e53bed58ec7814453234d70bfff082d2f8e0fc820a9ded.jpg", + "image_caption": [ + "MGIE" + ], + "image_footnote": [], + "bbox": [ + 316, + 160, + 400, + 224 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/dabcef71389c39f00a3f4ce172b9f164d67b1bf7d5f91f8ba3f6be55465727d4.jpg", + "image_caption": [ + "DreamOmni" + ], + "image_footnote": [], + "bbox": [ + 401, + 160, + 485, + 224 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/d465152c882e8b47a9990ec8678da46c0e9df0bd956a71cd60b4381b5cd1a816.jpg", + "image_caption": [ + "Replace the owl with a hawk" + ], + "image_footnote": [], + "bbox": [ + 550, + 93, + 633, + 157 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/b5e033c73ce20add6b6c4dcd2a1e6ed22923116fa2f039751487eecee7182d65.jpg", + "image_caption": [ + "Remove the green leafy plant near the yellow flower.", + "Instruction", + "Input Image" + ], + "image_footnote": [], + "bbox": [ + 550, + 160, + 633, + 224 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/a1975dc3ff5eefccd8213013d3e20438d3c6877a160e00f0a7b351bf6e2abd8f.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 635, + 93, + 715, + 157 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/707c2c486db7c59b0c69672c49d4a71abc1b0ecd4d5c893846e20ef80e927956.jpg", + "image_caption": [ + "InstructP2P" + ], + "image_footnote": [], + "bbox": [ + 635, + 160, + 715, + 224 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/e0578c815ace48acd27602d24a95d2e3188024e4a4031131a2a9d622096fde7c.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 717, + 93, + 797, + 157 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/65fb4c5bfdd36d542acf03c452d98ca0975af18d059f9cf80fa87d58af0b58be.jpg", + "image_caption": [ + "MGIE" + ], + "image_footnote": [], + "bbox": [ + 717, + 160, + 797, + 224 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/390b6b6a83e771d22d678eb0f70598a295e599a3d1a29b5e0f8a04c3ea1f2934.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 800, + 93, + 883, + 157 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/7422a2c7d86afda30228ef2d1353bc5d923439e58f7998aa2469725f842b4284.jpg", + "image_caption": [ + "DreamOmni" + ], + "image_footnote": [], + "bbox": [ + 800, + 160, + 883, + 224 + ], + "page_idx": 7 + }, + { + "type": "table", + "img_path": "images/66b9019baefe3beed15b0c539d69e73d6c0621b28a2f8d5276a03d47a8cd75a7.jpg", + "table_caption": [ + "Table 2. Quantitative comparison for inpainting & outpainting." + ], + "table_footnote": [], + "table_body": "
ModelInpaintingOutpainting
FID↓LPIPS↓FID↓LPIPS↓
SD-inpainting [44]1.35220.15602.91790.2475
ControlNet-inpainting [62]1.83930.15944.23370.2521
DreamOmni (Ours)0.83710.12031.69260.1995
", + "bbox": [ + 81, + 311, + 468, + 398 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "DreamOmni with competitive methods such as MGIE [17] and InstructP2P [8]. The visual results are presented in Fig. 8. We can see that DreamOmni performs more accurate edits, including addition, removal, and replacement. Specifically, our editing results exhibit superior consistency with minimal changes to non-edited areas and higher-quality generation of the edited content. This further validates that our synthetic data pipeline is an efficient and effective method for creating instruction-based editing datasets, enabling models to learn precise instruction-based editing. Moreover, the efficiency of our synthetic data pipeline allows for models to easily scale up diverse training data.", + "bbox": [ + 75, + 412, + 467, + 593 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Evaluation on Drag Editing. We evaluate DreamOmni on our synthetic evaluation dataset. The visual results are shown in Fig. 9. (1) Compared to the target image, we can see that DreamOmni can accurately perform translation, rotation, and scaling drag edits. (2) for translation and scaling, DreamOmni can maintain the integrity of the dragged object. However, large-scale rotation operations are more challenging for DreamOmni, as they involve complex transformations of the object itself, which may lead to deformations of the edited object. (3) The results demonstrate the effectiveness of our synthetic data pipeline for drag editing, and encoding the drag point positions and displacement information as instruction inputs allows the model to learn precise drag edits (shown in Fig. 2).", + "bbox": [ + 75, + 594, + 467, + 805 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "5. Conclusion", + "text_level": 1, + "bbox": [ + 76, + 814, + 194, + 830 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Current T2I foundation models lack a unified framework and training for downstream tasks, such as image editing. To address this, we introduce DreamOmni, a unified model for T2I generation and editing. We evaluate the frameworks", + "bbox": [ + 75, + 839, + 468, + 900 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/41a1ec74bd953ea13832c6550d8426b71c73ee5cd26514ff84cd2862f3753e5d.jpg", + "image_caption": [ + "Figure 8. Visual comparison on instruction-based editing. Our DreamOmni achieves more precise editing (including addition, removal, and replacement) compared to competitive methods such as MGIE [17] and InstructP2P [8]." + ], + "image_footnote": [], + "bbox": [ + 513, + 295, + 633, + 388 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/8e8d804f5243b71054d8655b1877e6915f5509e60e4e67efa14fce1c78aa32b9.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 513, + 388, + 633, + 482 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/90ab57500dc7cda0239a1c665654686bb9a9580257c2426083497beb015c1e8a.jpg", + "image_caption": [ + "Input Image & Drag Points", + "Figure 9. Visual results of drag editing: DreamOmni accurately performs translation, rotation, and scaling edits." + ], + "image_footnote": [], + "bbox": [ + 513, + 483, + 633, + 578 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/590e66926b4764db811df766b8c8c1994e8dfc0af06f1942ce54ae5705e3bb37.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 637, + 295, + 756, + 388 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/21f7d314731317b03a2de6b89db9fbc21a33fc259b328262bf75da13053fe3fa.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 637, + 388, + 756, + 482 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/b8f4b69ef02823a76d1457bddcd9fa9cfae7920ed9369078c48777bf8d3a4861.jpg", + "image_caption": [ + "Target Image" + ], + "image_footnote": [], + "bbox": [ + 637, + 483, + 756, + 578 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/788919fb008bd5ac4aa475ecb20d22edeb03f5ac6f180a53bdb647f3db096eba.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 759, + 295, + 879, + 388 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/f9933cfd1d5171916f164c753451183f97ca130d1ad5fa1c5eab1465cc77ea67.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 759, + 388, + 879, + 482 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/cc1058d78ae81e8c3a819a8878400c262743be1a8b43c95238b75c87350cb4a0.jpg", + "image_caption": [ + "DreamOmni" + ], + "image_footnote": [], + "bbox": [ + 759, + 483, + 879, + 578 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "of existing models under fair settings and consider the specific needs of various editing tasks. From this analysis, we develop a framework that integrates T2I with various editing tasks. Besides, a challenge in training editing models is the creation of high-quality, large-scale editing data, which is inefficient. To overcome this, we designed a synthetic collage data pipeline capable of efficiently generating quite a few precise, high-quality editing data. Moreover, the pipeline enhances the model's generation accuracy in text, position, quantity, color, and geometry. By jointly training on T2I and multi-task synthetic data, we develop a native, unified model for both image generation and editing. T2I training strengthens the model's grasp of specific concepts and improves generation quality, while editing training enables it to handle the requirements of editing tasks.", + "bbox": [ + 496, + 672, + 890, + 900 + ], + "page_idx": 7 + }, + { + "type": "page_number", + "text": "28540", + "bbox": [ + 478, + 944, + 517, + 955 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Acknowledgements. This work was supported in part by the Research Grants Council under the Areas of Excellence scheme grant AoE/E-601/22-R.", + "bbox": [ + 76, + 90, + 472, + 137 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "References", + "text_level": 1, + "bbox": [ + 78, + 150, + 174, + 166 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[1] Jacob Austin, Daniel D Johnson, Jonathan Ho, Daniel Tarlow, and Rianne van den Berg. Structured denoising diffusion models in discrete state-spaces. NeurIPS, 2021. 2", + "[2] Omri Avrahami, Dani Lischinski, and Ohad Fried. Blended diffusion for text-driven editing of natural images. In CVPR, 2022. 3", + "[3] Yogesh Balaji, Seungjun Nah, Xun Huang, Arash Vahdat, Jiaming Song, Qinsheng Zhang, Karsten Kreis, Miika Aittala, Timo Aila, Samuli Laine, et al. ediff-i: Text-to-image diffusion models with an ensemble of expert denoisers. arXiv preprint arXiv:2211.01324, 2022. 3", + "[4] Fan Bao, Shen Nie, Kaiwen Xue, Yue Cao, Chongxuan Li, Hang Su, and Jun Zhu. All are worth words: A vit backbone for diffusion models. In CVPR, 2023. 3", + "[5] Georgios Batzolis, Jan Stanczuk, Carola-Bibiane Schonlieb, and Christian Etmann. Conditional image generation with score-based diffusion models. arXiv preprint arXiv:2111.13606, 2021. 2", + "[6] James Betker, Gabriel Goh, Li Jing, Tim Brooks, Jianfeng Wang, Linjie Li, Long Ouyang, Juntang Zhuang, Joyce Lee, Yufei Guo, et al. Improving image generation with better captions. Computer Science. https://cdn.openai.com/papers/dall-e-3.pdf, 2023. 1, 3", + "[7] Manuel Brack, Felix Friedrich, Katharina Kornmeier, Linoy Tsaban, Patrick Schramowski, Kristian Kersting, and Apolinário Passos. *Ledits++: Limitless image editing using text-to-image models.* arXiv preprint arXiv:2311.16711, 2023. 3", + "[8] Tim Brooks, Aleksander Holynski, and Alexei A Efros. Instructpix2pix: Learning to follow image editing instructions. In CVPR, 2023. 2, 3, 4, 8", + "[9] Mingdeng Cao, Xintao Wang, Zhongang Qi, Ying Shan, Xiaohu Qie, and Yinqiang Zheng. Masactrl: Tuning-free mutual self-attention control for consistent image synthesis and editing. In ICCV, 2023. 3", + "[10] Junsong Chen, Jincheng Yu, Chongjian Ge, Lewei Yao, Enze Xie, Yue Wu, Zhongdao Wang, James Kwok, Ping Luo, Huchuan Lu, et al. Pixart-α: Fast training of diffusion transformer for photorealistic text-to-image synthesis. ICLR, 2024. 3, 5, 6", + "[11] Zhe Chen, Weiyun Wang, Hao Tian, Shenglong Ye, Zhang-wei Gao, Erfei Cui, Wenwen Tong, Kongzhi Hu, Jiapeng Luo, Zheng Ma, et al. How far are we to gpt-4v? closing the gap to commercial multimodal models with open-source suites. arXiv preprint arXiv:2404.16821, 2024. 5", + "[12] Guillaume Couairon, Jakob Verbeek, Holger Schwenk, and Matthieu Cord. Diffedit: Diffusion-based semantic image editing with mask guidance. arXiv preprint arXiv:2210.11427, 2022.3", + "[13] Xiaoliang Dai, Ji Hou, Chih-Yao Ma, Sam Tsai, Jialiang Wang, Rui Wang, Peizhao Zhang, Simon Vandenhende, Xi" + ], + "bbox": [ + 76, + 175, + 470, + 900 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "aofang Wang, Abhimanyu Dubey, et al. Emu: Enhancing image generation models using photogenic needles in a haystack. arXiv preprint arXiv:2309.15807, 2023. 3", + "[14] Prafulla Dhariwal and Alexander Nichol. Diffusion models beat gans on image synthesis. NeurIPS, 2021. 2", + "[15] Patrick Esser, Sumith Kulal, Andreas Blattmann, Rahim Entezari, Jonas Müller, Harry Saini, Yam Levi, Dominik Lorenz, Axel Sauer, Frederic Boessel, et al. Scaling rectified flow transformers for high-resolution image synthesis. In ICML, 2024. 1, 3, 4, 5, 6, 7", + "[16] Zhida Feng, Zhenyu Zhang, Xintong Yu, Yewei Fang, Lanxin Li, Xuyi Chen, Yuxiang Lu, Jiaxiang Liu, Weichong Yin, Shikun Feng, et al. Ernie-vilg 2.0: Improving text-to-image diffusion model with knowledge-enhanced mixture-of-denoising-experts. In CVPR, 2023. 3", + "[17] Tsu-Jui Fu, Wenze Hu, Xianzhi Du, William Yang Wang, Yinfei Yang, and Zhe Gan. Guiding instruction-based image editing via multimodal large language models. *ICLR*, 2024. 8", + "[18] Zigang Geng, Binxin Yang, Tiankai Hang, Chen Li, Shuyang Gu, Ting Zhang, Jianmin Bao, Zheng Zhang, Han Hu, Dong Chen, et al. Instructdiffusion: A generalist modeling interface for vision tasks. arXiv preprint arXiv:2309.03895, 2023. 3", + "[19] Dhruba Ghosh, Hannaneh Hajishirzi, and Ludwig Schmidt. Geneval: An object-focused framework for evaluating text-to-image alignment. NeurIPS, 36, 2024. 5, 7", + "[20] Ian Goodfellow, Jean Pouget-Abadie, Mehdi Mirza, Bing Xu, David Warde-Farley, Sherjil Ozair, Aaron Courville, and Yoshua Bengio. Generative adversarial nets. NeurIPS, 2014. 2", + "[21] Shuyang Gu, Dong Chen, Jianmin Bao, Fang Wen, Bo Zhang, Dongdong Chen, Lu Yuan, and Baining Guo. Vector quantized diffusion model for text-to-image synthesis. In CVPR, 2022. 2", + "[22] Amir Hertz, Ron Mokady, Jay Tenenbaum, Kfir Aberman, Yael Pritch, and Daniel Cohen-Or. Prompt-to-prompt image editing with cross attention control. arXiv preprint arXiv:2208.01626, 2022. 3", + "[23] Jonathan Ho, Ajay Jain, and Pieter Abbeel. Denoising diffusion probabilistic models. NeurIPS, 2020. 2", + "[24] Jonathan Ho, Chitwan Sahara, William Chan, David J Fleet, Mohammad Norouzi, and Tim Salimans. Cascaded diffusion models for high fidelity image generation. JMLR, 2022. 2", + "[25] Yuzhou Huang, Liangbin Xie, Xintao Wang, Ziyang Yuan, Xiaodong Cun, Yixiao Ge, Jiantao Zhou, Chao Dong, Rui Huang, Ruimao Zhang, et al. Smartedit: Exploring complex instruction-based image editing with multimodal large language models. CVPR, 2024. 3", + "[26] Diederik P Kingma. Auto-encoding variational bayes. arXiv preprint arXiv:1312.6114, 2013. 2", + "[27] Daiqing Li, Aleks Kamko, Ehsan Akhgari, Ali Sabet, Linmiao Xu, and Suhail Doshi. Playground v2. 5: Three insights towards enhancing aesthetic quality in text-to-image generation. arXiv preprint arXiv:2402.17245, 2024. 1", + "[28] Dongxu Li, Junnan Li, and Steven Hoi. Blip-diffusion: Pretrained subject representation for controllable text-to-image generation and editing. *NeurIPS*, 2024. 3, 4, 7" + ], + "bbox": [ + 503, + 92, + 890, + 900 + ], + "page_idx": 8 + }, + { + "type": "page_number", + "text": "28541", + "bbox": [ + 478, + 944, + 517, + 955 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[29] Ming Li, Taojiannan Yang, Huafeng Kuang, Jie Wu, Zhaoning Wang, Xuefeng Xiao, and Chen Chen. Controlnet ++: Improving conditional controls with efficient consistency feedback. In ECCV, 2025.3", + "[30] Zhen Li, Mingdeng Cao, Xintao Wang, Zhongang Qi, MingMing Cheng, and Ying Shan. Photomaker: Customizing realistic human photos via stacked id embedding. In CVPR, 2024. 3", + "[31] Zhimin Li, Jianwei Zhang, Qin Lin, Jiangfeng Xiong, Yanxin Long, Xinchi Deng, Yingfang Zhang, Xingchao Liu, Minbin Huang, Zedong Xiao, et al. Hunyuan-dit: A powerful multi-resolution diffusion transformer with fine-grained chinese understanding. arXiv preprint arXiv:2405.08748, 2024. 3", + "[32] Pengyang Ling, Lin Chen, Pan Zhang, Huaian Chen, Yi Jin, and Jinjin Zheng. Freedrag: Feature dragging for reliable point-based image editing. In CVPR, 2024. 3", + "[33] Xingchao Liu, Chengyue Gong, and Qiang Liu. Flow straight and fast: Learning to generate and transfer data with rectified flow. arXiv preprint arXiv:2209.03003, 2022. 5", + "[34] Hayk Manukyan, Andranik Sargsyan, Barsegh Atanyan, Zhangyang Wang, Shant Navasardyan, and Humphrey Shi. Hd-painter: high-resolution and prompt-faithful text-guided image inpainting with diffusion models. arXiv preprint arXiv:2312.14091, 2023. 3", + "[35] Alex Nichol, Prafulla Dhariwal, Aditya Ramesh, Pranav Shyam, Pamela Mishkin, Bob McGrew, Ilya Sutskever, and Mark Chen. Glide: Towards photorealistic image generation and editing with text-guided diffusion models. arXiv preprint arXiv:2112.10741, 2021. 2", + "[36] Xichen Pan, Li Dong, Shaohan Huang, Zhiliang Peng, Wenhu Chen, and Furu Wei. Kosmos-g: Generating images in context with multimodal large language models. arXiv preprint arXiv:2310.02992, 2023. 3", + "[37] Gaurav Parmar, Krishna Kumar Singh, Richard Zhang, Yijun Li, Jingwan Lu, and Jun-Yan Zhu. Zero-shot image-to-image translation. In ACM SIGGRAPH, 2023. 3", + "[38] William Peebles and Saining Xie. Scalable diffusion models with transformers. In ICCV, 2023. 2, 3, 4", + "[39] Dustin Podell, Zion English, Kyle Lacey, Andreas Blattmann, Tim Dockhorn, Jonas Müller, Joe Penna, and Robin Rombach. Sdxl: Improving latent diffusion models for high-resolution image synthesis. NeurIPS, 2023. 1, 3, 4, 5, 6", + "[40] Can Qin, Shu Zhang, Ning Yu, Yihao Feng, Xinyi Yang, Yingbo Zhou, Huan Wang, Juan Carlos Niebles, Caiming Xiong, Silvio Savarese, et al. Unicontrol: A unified diffusion model for controllable visual generation in the wild. arXiv preprint arXiv:2305.11147, 2023. 3", + "[41] Alec Radford, Jong Wook Kim, Chris Hallacy, Aditya Ramesh, Gabriel Goh, Sandhini Agarwal, Girish Sastry, Amanda Askell, Pamela Mishkin, Jack Clark, et al. Learning transferable visual models from natural language supervision. In ICML, 2021. 4", + "[42] Colin Raffel, Noam Shazeer, Adam Roberts, Katherine Lee, Sharan Narang, Michael Matena, Yanqi Zhou, Wei Li, and Peter J Liu. Exploring the limits of transfer learning with a" + ], + "bbox": [ + 78, + 90, + 468, + 900 + ], + "page_idx": 9 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "unified text-to-text transformer. Journal of machine learning research, 2020. 4", + "[43] Aditya Ramesh, Mikhail Pavlov, Gabriel Goh, Scott Gray, Chelsea Voss, Alec Radford, Mark Chen, and Ilya Sutskever. Zero-shot text-to-image generation. In ICML, 2021. 3", + "[44] Robin Rombach, Andreas Blattmann, Dominik Lorenz, Patrick Esser, and Björn Ommer. High-resolution image synthesis with latent diffusion models. In CVPR, 2022. 1, 2, 4, 5, 6, 7, 8", + "[45] Olaf Ronneberger, Philipp Fischer, and Thomas Brox. U-net: Convolutional networks for biomedical image segmentation. In MICCAI, 2015. 3", + "[46] Nataniel Ruiz, Yuanzhen Li, Varun Jampani, Yael Pritch, Michael Rubinstein, and Kfir Aberman. Dreambooth: Fine tuning text-to-image diffusion models for subject-driven generation. In CVPR, 2023. 3", + "[47] Chitwan Sahara, William Chan, Saurabh Saxena, Lala Li, Jay Whang, Emily L Denton, Kamyar Ghasemipour, Raphael Gontijo Lopes, Burcu Karagol Ayan, Tim Salimans, et al. Photorealistic text-to-image diffusion models with deep language understanding. NeurIPS, 2022. 1, 3", + "[48] Christoph Schuhmann, Romain Beaumont, Richard Vencu, Cade Gordon, Ross Wightman, Mehdi Cherti, Theo Coombes, Aarush Katta, Clayton Mullis, Mitchell Wortsman, et al. Laion-5b: An open large-scale dataset for training next generation image-text models. NeurIPS, 2022. 5", + "[49] Shelly Sheynin, Adam Polyak, Uriel Singer, Yuval Kirstain, Amit Zohar, Oron Ashual, Devi Parikh, and Yaniv Taigman. Emu edit: Precise image editing via recognition and generation tasks. arXiv preprint arXiv:2311.10089, 2023. 3", + "[50] Yujun Shi, Jun Hao Liew, Hanshu Yan, Vincent YF Tan, and Jiashi Feng. Instadrag: Lightning fast and accurate drag-based image editing emerging from videos. arXiv preprint arXiv:2405.13722, 2024. 3, 5", + "[51] Yujun Shi, Chuhui Xue, Jun Hao Liew, Jiachun Pan, Hanshu Yan, Wenqing Zhang, Vincent YF Tan, and Song Bai. Dragdiffusion: Harnessing diffusion models for interactive point-based image editing. In CVPR, 2024. 3", + "[52] Jascha Sohl-Dickstein, Eric Weiss, Niru Maheswaranathan, and Surya Ganguli. Deep unsupervised learning using nonequilibrium thermodynamics. In ICML, 2015. 2", + "[53] Yang Song, Jascha Sohl-Dickstein, Diederik P Kingma, Abhishek Kumar, Stefano Ermon, and Ben Poole. Score-based generative modeling through stochastic differential equations. arXiv preprint arXiv:2011.13456, 2020. 2", + "[54] Narek Tumanyan, Michal Geyer, Shai Bagon, and Tali Dekel. Plug-and-play diffusion features for text-driven image-to-image translation. In CVPR, 2023. 3", + "[55] Bram Wallace, Akash Gokul, and Nikhil Naik. Edict: Exact diffusion inversion via coupled transformations. In CVPR, 2023. 3", + "[56] Peng Wang, Shuai Bai, Sinan Tan, Shijie Wang, Zhihao Fan, Jinze Bai, Keqin Chen, Xuejing Liu, Jialin Wang, Wenbin Ge, et al. Qwen2-vl: Enhancing vision-language model's perception of the world at any resolution. arXiv preprint arXiv:2409.12191, 2024. 5" + ], + "bbox": [ + 501, + 92, + 890, + 898 + ], + "page_idx": 9 + }, + { + "type": "page_number", + "text": "28542", + "bbox": [ + 478, + 945, + 519, + 955 + ], + "page_idx": 9 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[57] Bin Xia, Shi Yin Wang, Yingfan Tao, Yitong Wang, and Jiaya Jia. Llmga: Multimodal large language model based generation assistant. In European Conference on Computer Vision, 2025. 3", + "[58] Shitao Xiao, Yueze Wang, Junjie Zhou, Huaying Yuan, Xingrun Xing, Ruiran Yan, Shuting Wang, Tiejun Huang, and Zheng Liu. Omnigen: Unified image generation. arXiv preprint arXiv:2409.11340, 2024. 3", + "[59] Shaoan Xie, Zhifei Zhang, Zhe Lin, Tobias Hinz, and Kun Zhang. Smartbrush: Text and shape guided object inpainting with diffusion model. In CVPR, 2023. 3", + "[60] Hu Ye, Jun Zhang, Sibo Liu, Xiao Han, and Wei Yang. Ip-adapter: Text compatible image prompt adapter for text-to-image diffusion models. arXiv preprint arXiv:2308.06721, 2023. 2, 3, 4, 7", + "[61] Kai Zhang, Lingbo Mo, Wenhu Chen, Huan Sun, and Yu Su. Magicbrush: A manually annotated dataset for instruction-guided image editing. NeurIPS, 2024. 3", + "[62] Lvmin Zhang, Anyi Rao, and Maneesh Agrawala. Adding conditional control to text-to-image diffusion models. In ICCV, 2023. 2, 3, 5, 6, 7, 8", + "[63] Shihao Zhao, Dongdong Chen, Yen-Chun Chen, Jianmin Bao, Shaozhe Hao, Lu Yuan, and Kwan-Yee K. Wong. Uni-controlnet: All-in-one control to text-to-image diffusion models. NeurIPS, 2023. 3" + ], + "bbox": [ + 78, + 90, + 468, + 444 + ], + "page_idx": 10 + }, + { + "type": "page_number", + "text": "28543", + "bbox": [ + 478, + 945, + 517, + 955 + ], + "page_idx": 10 + } +] \ No newline at end of file diff --git a/2025/DreamOmni_ Unified Image Generation and Editing/ff84d3e1-2391-4334-90b8-2fe54fd66a09_model.json b/2025/DreamOmni_ Unified Image Generation and Editing/ff84d3e1-2391-4334-90b8-2fe54fd66a09_model.json new file mode 100644 index 0000000000000000000000000000000000000000..2f28f5839c15187e8f632a2f1ac9b2498209e47a --- /dev/null +++ b/2025/DreamOmni_ Unified Image Generation and Editing/ff84d3e1-2391-4334-90b8-2fe54fd66a09_model.json @@ -0,0 +1,3522 @@ +[ + [ + { + "type": "header", + "bbox": [ + 0.107, + 0.003, + 0.182, + 0.043 + ], + "angle": 0, + "content": "CVF" + }, + { + "type": "header", + "bbox": [ + 0.237, + 0.001, + 0.808, + 0.047 + ], + "angle": 0, + "content": "This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore." + }, + { + "type": "title", + "bbox": [ + 0.22, + 0.131, + 0.753, + 0.154 + ], + "angle": 0, + "content": "DreamOmni: Unified Image Generation and Editing" + }, + { + "type": "text", + "bbox": [ + 0.245, + 0.18, + 0.726, + 0.253 + ], + "angle": 0, + "content": "Bin Xia1, Yuechen Zhang1, Jingyao Li1, Chengyao Wang1, Yitong Wang2, Xinglong Wu2, Bei Yu1, and Jiaya Jia3 \n1 CUHK, 2 ByteDance Inc, 3 HKUST \nhttps://zj-binxia.github.io/DreamOmni-ProjectPage/" + }, + { + "type": "image", + "bbox": [ + 0.084, + 0.261, + 0.887, + 0.595 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.091, + 0.602, + 0.877, + 0.617 + ], + "angle": 0, + "content": "Figure 1. The gallery of DreamOmni. DreamOmni, as a native unified image generation and editing model, can handle various tasks." + }, + { + "type": "title", + "bbox": [ + 0.235, + 0.633, + 0.314, + 0.648 + ], + "angle": 0, + "content": "Abstract" + }, + { + "type": "text", + "bbox": [ + 0.075, + 0.66, + 0.473, + 0.902 + ], + "angle": 0, + "content": "projectpagepCurrently, the success of large language models (LLMs) illustrates that a unified multitasking approach can significantly enhance model usability, streamline deployment, and foster synergistic benefits across different tasks. However, in computer vision, while text-to-image (T2I) models have significantly improved generation quality through scaling up, their framework design did not initially consider how to unify with downstream tasks, such as various types of editing. To address this, we introduce DreamOmni, a unified model for image generation and editing. We begin by analyzing existing frameworks and the requirements of downstream tasks, proposing a unified framework that integrates both T2I models and various editing tasks. Furthermore, another key challenge is the efficient creation of high-quality editing data, particularly for instruction-based and drag-based editing. To this" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.634, + 0.895, + 0.8 + ], + "angle": 0, + "content": "end, we develop a synthetic data pipeline using sticker-like elements to synthesize accurate, high-quality datasets efficiently, which enables editing data scaling up for unified model training. For training, DreamOmni jointly trains T2I generation and downstream tasks. T2I training enhances the model's understanding of specific concepts and improves generation quality, while editing training helps the model grasp the nuances of the editing task. This collaboration significantly boosts editing performance. Extensive experiments confirm the effectiveness of DreamOmni. The code and model will be released." + }, + { + "type": "title", + "bbox": [ + 0.501, + 0.813, + 0.633, + 0.829 + ], + "angle": 0, + "content": "1. Introductian" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.84, + 0.894, + 0.903 + ], + "angle": 0, + "content": "Recently, text-to-image (T2I) generative foundation models [6, 15, 27, 39, 44, 47] have made remarkable progress, driving the development of various downstream applications such as image editing, video generation, and more." + }, + { + "type": "page_number", + "bbox": [ + 0.479, + 0.945, + 0.52, + 0.958 + ], + "angle": 0, + "content": "28533" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.076, + 0.092, + 0.473, + 0.35 + ], + "angle": 0, + "content": "Currently, T2I foundation models face two significant challenges: (1) Adapting these models for downstream applications often requires the integration of various plugins (such as ControlNet [62] and IP-adapter [60]) in different ways, or the extension of input channels (e.g., SD-inpainting [44], InstructP2P [8]). This reliance on specialized frameworks hampers multi-task generalization and complicates deployment. (2) High-quality and accurate editing data is difficult to obtain, including instruction-based editing, drag editing, and subject-driven generation data. In this paper, we propose to unify T2I models with multiple editing tasks, such as instruction-based editing, inpainting & outpainting, drag editing, and reference image generation within a single framework. Additionally, we introduce an efficient synthetic data pipeline for efficiently and accurately constructing editing data, facilitating the training of native unified generation and editing model." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.353, + 0.473, + 0.671 + ], + "angle": 0, + "content": "We began by exploring fundamental model frameworks to develop a unified framework that is both highly effective and efficient, achieving rapid training convergence. (1) We aligned certain frameworks, such as Unet [44] and DIT [38], to have similar parameter sizes and runtime settings for fair comparison (Fig. 3). Our analysis showed that DIT's effectiveness comes from concentrating most of its computational and parameter load on latent at \\(2 \\times\\) downsampling size, whereas Unet distributes more of its operations at smaller scales. This allocation makes DIT comparatively superior, as concentrating computations at the \\(2 \\times\\) downsampled latent achieves a better trade-off. However, we observed that Unet's residual connections significantly improve the model's training convergence speed. Inspired by the efficiency of high-resolution latent computations, we also incorporated additional residual convolutional networks for the input-scale latent. (2) We replaced the original text encoder with a vision-language model (VLM) to unify visual-language prompt encoding. This encoded prompt was then concatenated with the noisy latent, allowing for integrated computation within the DIT framework." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.675, + 0.473, + 0.903 + ], + "angle": 0, + "content": "Next, we need to obtain accurate training data efficiently for our unified model. For various types of editing, such as instruction-based editing, it's challenging to create and filter pairs of data that precisely match the instructions while avoiding detail corruption. We have discovered that the key to effective editing lies in helping the model understand the meaning of the editing operations, rather than learning specific concepts (a capability the model already possesses in T2I training). Therefore, we propose a synthetic collage data pipeline that easily scales up to generate accurate and diverse edited data, enabling models to better understand the nuances of specific editing tasks (Fig. 2 (b)). (1) For instruction-based editing, we constructed synthetic data for three main categories: addition, removal, and replacement. (2) For drag editing, we utilized synthetic data to cre" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.092, + 0.895, + 0.363 + ], + "angle": 0, + "content": "ate scaling, translating, and rotating editing data. (3) For inpainting and outpainting, we randomly generate masks for each given image. (4) For reference image generation, We divided the tasks into subject-driven generation and image-conditioned generation (i.e., ControlNet-like generation) tasks. For subject-driven generation, we used a canvas filled with stickers, allowing the model to reference a specific sticker for generation, which helps train the model to extract the demanded objects or details from given images for generation. For the image-conditioned generation, we created canny, depth, and segment maps, similar to ControlNet. (5) We use synthetic data to enhance the model's T2I accuracy in responding to attributes such as quantity, position, relationships, color, shape, and text. Concretely, we randomly arrange different quantities of stickers, text, and geometric shapes on a canvas, and obtain precise descriptions of their quantity, position, relationships, and colors based on their exact coordinates." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.364, + 0.895, + 0.485 + ], + "angle": 0, + "content": "During training, thanks to our unified framework that integrates T2I with multiple editing tasks, we can easily train T2I data alongside various types of editing data, obtaining a native unified image generation and editing model, called DreamOmni. The editing tasks benefit from the T2I data, which helps prevent the model from forgetting specific concepts and generation quality decline, while the T2I task can leverage synthetic data to enhance instruction following." + }, + { + "type": "text", + "bbox": [ + 0.5, + 0.487, + 0.892, + 0.547 + ], + "angle": 0, + "content": "- We conducted an analysis of existing model frameworks under fair settings and, drawing on the characteristics of different tasks, proposed an efficient and powerful unified image generation and editing framework, DreamOmni." + }, + { + "type": "text", + "bbox": [ + 0.5, + 0.548, + 0.892, + 0.668 + ], + "angle": 0, + "content": "- We introduced a synthetic collage data pipeline to tackle the current inefficiency and difficulty in creating and filtering high-quality editing data. Furthermore, we utilized the synthetic collage data pipeline to enhance the accuracy of the T2I model's output. Experimental results demonstrate that synthetic data is an effective, high-quality, and cost-efficient method to scale up data for achieving unified image generation and editing training." + }, + { + "type": "text", + "bbox": [ + 0.5, + 0.668, + 0.892, + 0.713 + ], + "angle": 0, + "content": "- After unified training with T2I data and various synthetic datasets, our DreamOmni showcases competitive performance on T2I generation and various editing tasks." + }, + { + "type": "list", + "bbox": [ + 0.5, + 0.487, + 0.892, + 0.713 + ], + "angle": 0, + "content": null + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.725, + 0.642, + 0.74 + ], + "angle": 0, + "content": "2. Related Work" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.75, + 0.894, + 0.903 + ], + "angle": 0, + "content": "Text-to-image Diffusion Model. Diffusion models [1, 5, 14, 21, 23, 24, 35, 52, 53] have emerged as highly effective methods for image generation, outperforming previous generative models like GANs [20] and VAEs [26]. The latent diffusion model [44], also known as Stable Diffusion, enhances the diffusion process within the image latent space and utilizes cross-attention to guide image generation through text input. This approach enhances the usability of the image generation model, driving its widespread adoption. Subsequently, more advanced text-to" + }, + { + "type": "page_number", + "bbox": [ + 0.479, + 0.945, + 0.521, + 0.958 + ], + "angle": 0, + "content": "28534" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.087, + 0.09, + 0.4, + 0.38 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.157, + 0.382, + 0.356, + 0.392 + ], + "angle": 0, + "content": "(a) The framework of DreamOmni" + }, + { + "type": "image", + "bbox": [ + 0.408, + 0.091, + 0.888, + 0.374 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.556, + 0.38, + 0.753, + 0.392 + ], + "angle": 0, + "content": "(b) The examples of synthetic data" + }, + { + "type": "image_caption", + "bbox": [ + 0.076, + 0.396, + 0.892, + 0.48 + ], + "angle": 0, + "content": "Figure 2. The overview of DreamOmni. (a) The DreamOmni framework supports unified image generation and editing, with fast training convergence and powerful performance. (b) To overcome the difficulty and inefficiency in data creation and filtering for image editing, we propose a collage-based synthetic data pipeline. This pipeline enables the efficient creation of data for various editing tasks, such as adding, deleting, and replacement operations in instruction-based editing, as well as translation, scaling, and rotation in drag editing. Additionally, it supports reference image generation and segmentation & detection. Furthermore, our synthetic data pipeline enhances the accuracy of T2I generation. Due to space limitations, we have optionally shown the corresponding prompts or instructions for these cases." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.492, + 0.47, + 0.748 + ], + "angle": 0, + "content": "image (T2I) models were developed, such as Imagen [47], DALLE [43], and others [3, 6, 13, 16, 39]. Recently, the DIT frameworks [4, 38] has become increasingly popular as more models [10, 15, 31] transition from the Unet framework [45], thanks to its straightforward structure and outstanding compatibility with multimodal inputs. However, the current T2I model fails to account for the design of downstream editing tasks, necessitating numerous specialized adaptations. For instance, Instructp2p [8] requires an expansion of input channels, while ControlNet [62] and IP-adapter [60] rely on additional plugins and various methods for injecting reference image information. These fragmented designs are inconsistent with the trend toward unified models, complicating deployment and impeding the joint training of T2I and editing tasks. In this paper, we will analyze the existing T2I frameworks and propose a more efficient and powerful unified generation and editing model." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.75, + 0.471, + 0.902 + ], + "angle": 0, + "content": "Multimodal-guided Image Editing and Generation offers richer guiding information and has broader applications [2, 7, 9, 12, 37, 54, 55] than T2I models. It encompasses instruction-based editing [8, 18, 25, 49, 61], inpainting & outpainting [34, 57, 59], drag editing [32, 51], and reference image generation [28, 29, 40, 62, 63]. A significant challenge in editing, especially in instruction-based and drag editing, lies in creating and filtering accurately paired editing data. For example, InstructP2P [8] utilizes finetuned GPT-3 and a Prompt-to-Prompt strategy [22] to" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.492, + 0.892, + 0.825 + ], + "angle": 0, + "content": "generate instruction-based editing datasets, but it achieves a success rate of under \\(15\\%\\) with frequent artifacts and detail loss. Additionally, filtering high-quality editing data is a persistent challenge, complicating scaling up. Magicbrush [61] addresses this by employing people to create and filter editing data, but this manual method limits data volume. Furthermore, for drag editing, most methods [32, 51] are training-free, which restricts their performance and efficiency. To this end, InstaDrag [50] leverages video data to construct drag editing data. However, InstaDrag still faces challenges with inefficiency in filtering valid data from video. Moreover, for subject-driven generation [28, 36], DreamBooth [46] requires training on several subject-related images before each inference, which is inconvenient and inefficient. Some works [30, 58] collect multiple photos of the same person from the internet to build datasets for training more efficient and powerful subject-driven generation models. However, the process of gathering and filtering such data still presents a challenge, limiting the scaling up of the dataset. In this paper, we propose a simple, efficient, and accurate synthetic collage data pipeline to create quite a few data for unified training." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.843, + 0.634, + 0.86 + ], + "angle": 0, + "content": "3. Methodology" + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.871, + 0.892, + 0.901 + ], + "angle": 0, + "content": "The unification of multiple tasks is a trend and pursuit in the field of computer vision and AI, which not only en" + }, + { + "type": "page_number", + "bbox": [ + 0.48, + 0.945, + 0.518, + 0.957 + ], + "angle": 0, + "content": "28535" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.076, + 0.092, + 0.473, + 0.396 + ], + "angle": 0, + "content": "hances model usability and reduces deployment complexity but also enables collaborative training that fosters synergies between tasks. However, current T2I foundation models are primarily designed specifically for T2I and often overlook the potential for integration with other tasks, such as various image editing tasks. To this end, we propose DreamOmni, a unified model for image generation and editing. We design and train DreamOmni from three aspects. (1) In Sec. 3.1, we compare various frameworks in a fair setting and design a powerful and fast training converging framework that supports unified multi-tasking based on the characteristics of different tasks. (2) In Sec. 3.2, we aim to propose a convenient, efficient, and accurate synthetic data pipeline for data scaling up to facilitate multi-task unified training and enhance the model's instruction-following ability. (3) In Sec. 3.3, we introduce the training scheme for DreamOmni. Native unified training of T2I and various editing tasks prevents concept forgetting and generation quality decline, while enhancing the model's editing and prompt-following capabilities." + }, + { + "type": "title", + "bbox": [ + 0.078, + 0.407, + 0.206, + 0.421 + ], + "angle": 0, + "content": "3.1. Framework" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.432, + 0.473, + 0.825 + ], + "angle": 0, + "content": "In this part, we aim to design a unified and powerful image generation and editing framework. Currently, different editing models often have distinct structure designs. For example, IP-adapter [60] and BLIP-Diffusion [28] inject information through cross-attention to maintain subjects. In contrast, InstructP2P [8] achieves editing consistency by adding different numbers of input channels for models. These structures are tailored for specific tasks and lack generalizability. To this end, as shown in Fig. 2 (a), we concatenate VLM features with noisy latent and input them into a DIT Block for joint multi-head self-attention operations. After that, the VLM features and noisy latent are processed by the FeedForward modules. This allows the model to autonomously learn any level features (from overall consistency to subject consistency) for editing and generation. Notably, for FeedForward modules, we separate the VLM features and noisy latent, passing them through two distinct FeedForward modules with the same network structure. Additionally, instead of using CLIP [41] or T5 [42] as the text encoder, we introduce a Vision-Language Model (VLM) that enables joint understanding and encoding of both image and text prompts. For tasks that require high consistency, such as instruction-based editing and drag editing, we input the VAE encoding of the source image into the DIT model to ensure strong consistency between the non-edited regions of the output image and the source image." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.826, + 0.473, + 0.903 + ], + "angle": 0, + "content": "In the current framework design, some works, such as DIT [38], are compared under label-conditioned generation rather than T2I. However, T2I is inherently more complex than label-based generation, as it requires the integration and understanding of complex prompts. Additionally, many" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.092, + 0.895, + 0.288 + ], + "angle": 0, + "content": "T2I models [15, 44], such as SDXL [39], are trained with different model sizes, datasets, and training settings. This variability makes it challenging to assess the impact of different model components on overall performance. Moreover, SDXL incorporates many Transformer blocks within its UNet structure. So, why does DIT outperform SDXL? To address this, we conducted extensive experiments as shown in Fig. 3. We observe that DIT surpasses Unet because it allocates most of its computation to \\(2 \\times\\) downsampled latent, whereas Unet allocates a greater proportion to \\(4 \\times\\) downsampled latent. Since attention operations on \\(1 \\times\\) latent cause memory burden, we further employ residual Conv blocks to refine generation details for \\(1 \\times\\) latent." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.289, + 0.895, + 0.396 + ], + "angle": 0, + "content": "Additionally, we observed that the use of long connections in the UNet framework can significantly accelerate the model's training convergence without compromising performance. As shown in Fig. 2 (a), we concatenate early and later features along the channel dimension and apply a linear layer to combine the two features. Notably, linear layers used for VLM features and noisy latent are different." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.408, + 0.651, + 0.423 + ], + "angle": 0, + "content": "3.2. Synthetic Data" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.432, + 0.895, + 0.583 + ], + "angle": 0, + "content": "In addition to the unified framework, we also require quite a few data to support joint training. While T2I data is readily available, creating and filtering accurate, high-quality data for tasks like instruction-based editing is far more challenging. To address this, we introduce a synthetic collage data pipeline that efficiently and accurately generates the required editing data. As shown in Fig. 2 (b), our pipeline encompasses six tasks. Notably, this is not the full extent of our synthetic pipeline's capabilities; it is also capable of handling more complex task combinations." + }, + { + "type": "text", + "bbox": [ + 0.5, + 0.584, + 0.895, + 0.84 + ], + "angle": 0, + "content": "- T2I generation. As shown in Fig. 2 (b), in addition to the conventional T2I data, we further enhance the model's performance on the T2I by incorporating synthetic data, specifically focusing on improving the text, shape, position, quantity, and color generation. Specifically, for the text, we randomly generate words or short phrases on a blank canvas using a variety of fonts, colors, thicknesses, and sizes. For the shape and quantity, we randomly create geometric shapes with different quantities, colors, and sizes, and arrange them on the canvas. Based on these attributes and their positions, we generate accurate prompts, which are then refined using a LLM. Furthermore, we use a diverse set of stickers and segmentation data for synthesis, placing them on the canvas and calculating their precise spatial relationships. These prompts are then created and further polished by the LLM to produce descriptions that are more natural." + }, + { + "type": "text", + "bbox": [ + 0.5, + 0.841, + 0.895, + 0.901 + ], + "angle": 0, + "content": "- Inpainting & Outpainting. We randomly generate masks for smearing, blocks, and image edges. Notably, during training, in addition to feeding the masked image and its corresponding mask into the VLM for encoding, we also" + }, + { + "type": "page_number", + "bbox": [ + 0.48, + 0.946, + 0.521, + 0.957 + ], + "angle": 0, + "content": "28536" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.091, + 0.092, + 0.423, + 0.107 + ], + "angle": 0, + "content": "include the image caption with a \\(50\\%\\) probability." + }, + { + "type": "text", + "bbox": [ + 0.078, + 0.108, + 0.47, + 0.273 + ], + "angle": 0, + "content": "- Instruction-based editing. We categorize the task into three operations: addition, removal, and replacement. For both removal and replacement, we randomly select a background image and an object image to create the source image. In the removal case, the target image is simply the background image. For replacement, the target image is generated by replacing the object in the same location with a different object. Notably, for the addition, due to the need for the added object to be placed in a contextually appropriate position relative to the background, we use a blank background in the paper." + }, + { + "type": "text", + "bbox": [ + 0.078, + 0.273, + 0.471, + 0.424 + ], + "angle": 0, + "content": "- Drag editing. We categorize the data into three types: translation, scaling, and rotation. Notably, Instadrag [50] treats each pair of drag points as a separate image, which is sparse and impractical due to the fixed number of drag points required. Therefore, we represent each drag point using the format \\((x,y,dx,dy)\\) as a prompt input, where \\(x\\) and \\(y\\) denote the coordinates of the drag points in the source image, and \\(dx\\) and \\(dy\\) represent the translation vector. Furthermore, we normalize these coordinates by dividing them by the image's width or height." + }, + { + "type": "text", + "bbox": [ + 0.078, + 0.424, + 0.47, + 0.606 + ], + "angle": 0, + "content": "- Reference image generation. We categorize the data into two types: image-conditioned generation, akin to ControlNet [62], and subject-driven generation. For image-conditioned generation, we begin by selecting high-quality images and creating corresponding canny maps, depth maps, and segmentation masks as source images for training. For subject-driven generation, we synthesize source images and then randomly select objects within these images to create target images. The model is trained to generate new content based on specific attributes referenced from the source image, allowing for the flexible generation of varied scenes and subjects." + }, + { + "type": "text", + "bbox": [ + 0.078, + 0.606, + 0.47, + 0.697 + ], + "angle": 0, + "content": "- Segmentation & detection. We randomly selected a background image and an object image. These images are then composited together to create a source image. Following this, we apply color manipulation or draw a bounding box around the object region based on the alpha channel of the object image, obtaining the target image." + }, + { + "type": "list", + "bbox": [ + 0.078, + 0.108, + 0.471, + 0.697 + ], + "angle": 0, + "content": null + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.697, + 0.47, + 0.743 + ], + "angle": 0, + "content": "Our synthetic collage data pipeline is efficient and precise, enabling the generation of billions of diverse images for large-scale pretraining and fine-tuning of DreamOmni." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.756, + 0.236, + 0.772 + ], + "angle": 0, + "content": "3.3. Model Training" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.781, + 0.471, + 0.903 + ], + "angle": 0, + "content": "After careful consideration, we chose a \\(2.5B\\) parameter size for DreamOmni's DIT model. This size strikes a good balance between being user-friendly and ensuring powerful performance. For the VLM encoder, we directly adopt the Qwen2-VL [56] 7B model, based on three reasons: (1) It supports image inputs of arbitrary resolution, (2) It delivers strong model performance, and (3) It is released under a permissive open-source license. The VLM features are derived" + }, + { + "type": "image", + "bbox": [ + 0.503, + 0.09, + 0.697, + 0.203 + ], + "angle": 0, + "content": null + }, + { + "type": "table", + "bbox": [ + 0.711, + 0.118, + 0.888, + 0.179 + ], + "angle": 0, + "content": "
FrameworkParam (M)Runtime (ms)
SDXL [39]83828.59
Pixart [10]86234.61
SD3 [15]89625.73
DreamOmni-V182828.32
DreamOmni-V283625.14
DreamOmni-V3 (Ours)82727.04
" + }, + { + "type": "image_caption", + "bbox": [ + 0.499, + 0.207, + 0.894, + 0.25 + ], + "angle": 0, + "content": "Figure 3. Comparison of different frameworks. The left figure shows the FID comparison among different frameworks, while the right table shows their number of parameters and runtime." + }, + { + "type": "table_caption", + "bbox": [ + 0.5, + 0.257, + 0.893, + 0.286 + ], + "angle": 0, + "content": "Table 1. Comparisons on GenEval [19]. Our model outperforms all current open models (including SOTA SD3-Medium [15])." + }, + { + "type": "table", + "bbox": [ + 0.504, + 0.289, + 0.893, + 0.395 + ], + "angle": 0, + "content": "
ModelObjectsCountingColorsPositionColor Attribution
OverallSingleTwo
SD 1.5 [44]0.430.970.380.350.760.040.06
PixArt-alpha [10]0.480.980.50.440.80.080.07
SDv2.10.50.980.510.440.850.070.17
SDXL [39]0.550.980.740.390.850.150.23
SD-cascade0.520.980.570.470.870.080.15
SD3-Medium [15]0.700.990.840.630.880.280.55
DreamOmni (Ours)0.700.990.810.650.880.340.54
" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.414, + 0.893, + 0.52 + ], + "angle": 0, + "content": "from the penultimate layer of Qwen2-VL. Additionally, we use the FLUX-schnell's VAE as DreamOmni's VAE, which retains more latent channels, enabling the model to capture finer image details. Furthermore, we optimize DreamOmni using Rectified Flow [33], which performs the forward process by linearly interpolating between noise and data along a straight trajectory. we train DreamOmni using loss \\(\\mathcal{L}\\):" + }, + { + "type": "equation", + "bbox": [ + 0.574, + 0.524, + 0.892, + 0.55 + ], + "angle": 0, + "content": "\\[\n\\mathcal {L} = \\mathbb {E} \\left(\\left\\| (\\mathbf {z} - \\epsilon) - v _ {\\theta} \\left(\\mathbf {z} _ {t}, \\mathbf {c}, t\\right) \\right\\| _ {2} ^ {2}\\right), \\tag {1}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.553, + 0.894, + 0.628 + ], + "angle": 0, + "content": "where \\(\\mathbf{z}_t = t\\mathbf{z} + (1 - t)\\epsilon\\) represents the noised feature map at timestep \\(t\\). Ground truth images are encoded by VAE into latent space to derive \\(\\mathbf{z}\\). Here, \\(\\epsilon \\in \\mathcal{N}(0,\\mathbf{I})\\) represents Gaussian noise, and \\(v_{\\theta}\\) refers to the DIT model. \\(\\mathbf{c}\\) indicates the conditional information." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.629, + 0.894, + 0.78 + ], + "angle": 0, + "content": "The T2I training dataset consists of \\(125M\\) images, including both the LAION dataset \\((103M)\\) [48] and our collected data \\((22M)\\). These images have been annotated using InternVL2 [11]. For synthetic data, such as T2I, instruction editing, inpainting & outpainting, drag editing, and reference image generation, we generated \\(12M\\) images respectively, bringing the total to approximately \\(60M\\) images. Additionally, for segmentation & detection, we created \\(8M\\) images. Prior to training, to improve training speed usage, we first encoded prompts into VLM features." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.781, + 0.895, + 0.903 + ], + "angle": 0, + "content": "Our training process is divided into three stages. In the first stage, we train on images with \\(256 \\times 256\\) size, a batch size of 2048, a learning rate of \\(1 \\times 10^{-4}\\), and \\(377K\\) iterations. In the second stage, we train on images with \\(512 \\times 512\\) size, a batch size of 1024, a learning rate of \\(5 \\times 10^{-5}\\), and \\(189K\\) iterations. In the final stage, we select 12M high-quality T2I data and randomly sample 1M high-quality images from each type of synthetic data for training." + }, + { + "type": "page_number", + "bbox": [ + 0.479, + 0.945, + 0.521, + 0.958 + ], + "angle": 0, + "content": "28537" + } + ], + [ + { + "type": "image_footnote", + "bbox": [ + 0.1, + 0.102, + 0.241, + 0.18 + ], + "angle": 0, + "content": "A heartwarming 3D rendered scene of an elderly farmer and a tiny orange kitten dressed in matching denim overalls. The farmer, with a gentle smile, walks alongside the kitten in a lush, green garden filled with thriving plants, showcasing a fruitful harvest. The intricate details of the overalls and the farmer's worn, weathered face tell a story of years spent tending to the land." + }, + { + "type": "image_footnote", + "bbox": [ + 0.096, + 0.191, + 0.246, + 0.284 + ], + "angle": 0, + "content": "Deep in the enchanted forest lives a woman who is the moon fairy. Her long blonde hair shines in the starlight, tangled with her flowers that glow with a soft blue glow. Her eyes are the color of the night and shine with the magic of the night. The fairy wears a dress made of moon petals, woven with threads of moonlight that shine with an iridescent glow, a crown of stars adorns her head, shining with the light of the full moon that illuminates the forest. Her wings are translucent like glass, with a pale glow reminiscent of the glow of the moon. HD, 16K, photo, cinematic, poster" + }, + { + "type": "image_footnote", + "bbox": [ + 0.099, + 0.291, + 0.241, + 0.375 + ], + "angle": 0, + "content": "A unique, intricately detailed creature resembling a reptile, possibly a lizard or a gecko. It has a vibrant blue and green scaled body, with large, round, and expressive eyes that are a deep shade of blue. The creature is adorned with a floral crown made of various flowers, including pink roses and blue blossoms, as well as pinecones. The backdrop is a soft, blurred forest setting, suggesting a serene and mystical ambiance." + }, + { + "type": "image_caption", + "bbox": [ + 0.141, + 0.384, + 0.193, + 0.398 + ], + "angle": 0, + "content": "Prompt" + }, + { + "type": "image", + "bbox": [ + 0.25, + 0.095, + 0.368, + 0.186 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.25, + 0.189, + 0.368, + 0.281 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.25, + 0.285, + 0.368, + 0.377 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.286, + 0.383, + 0.33, + 0.396 + ], + "angle": 0, + "content": "SD1.5" + }, + { + "type": "image", + "bbox": [ + 0.375, + 0.095, + 0.492, + 0.186 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.375, + 0.189, + 0.492, + 0.28 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.375, + 0.285, + 0.492, + 0.376 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.399, + 0.383, + 0.481, + 0.396 + ], + "angle": 0, + "content": "SD-Cascade" + }, + { + "type": "image", + "bbox": [ + 0.501, + 0.095, + 0.618, + 0.185 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.501, + 0.189, + 0.617, + 0.28 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.5, + 0.285, + 0.618, + 0.376 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.541, + 0.383, + 0.585, + 0.396 + ], + "angle": 0, + "content": "SDXL" + }, + { + "type": "image", + "bbox": [ + 0.627, + 0.095, + 0.744, + 0.185 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.627, + 0.19, + 0.743, + 0.28 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.627, + 0.285, + 0.743, + 0.376 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.639, + 0.383, + 0.73, + 0.396 + ], + "angle": 0, + "content": "SD3-Medium" + }, + { + "type": "image", + "bbox": [ + 0.751, + 0.095, + 0.869, + 0.185 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.751, + 0.19, + 0.869, + 0.28 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.751, + 0.285, + 0.869, + 0.376 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.749, + 0.383, + 0.878, + 0.397 + ], + "angle": 0, + "content": "DreamOmni (Ours)" + }, + { + "type": "image_caption", + "bbox": [ + 0.076, + 0.408, + 0.892, + 0.45 + ], + "angle": 0, + "content": "Figure 4. Visual comparison on T2I generation. Compared to other competitive methods (including SD3-Medium [15], SDXL [39], SD-Cascade, and SD1.5 [44]), our DreamOmni not only better adheres to user prompts but also generate more visually appealing results with delicate details, elegant composition, and so on." + }, + { + "type": "image", + "bbox": [ + 0.088, + 0.465, + 0.185, + 0.541 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.089, + 0.544, + 0.185, + 0.619 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.095, + 0.627, + 0.177, + 0.639 + ], + "angle": 0, + "content": "Masked Image" + }, + { + "type": "image", + "bbox": [ + 0.187, + 0.466, + 0.284, + 0.541 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.187, + 0.544, + 0.284, + 0.619 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.207, + 0.622, + 0.268, + 0.646 + ], + "angle": 0, + "content": "ControlNet Inpainting" + }, + { + "type": "image", + "bbox": [ + 0.286, + 0.466, + 0.382, + 0.541 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.286, + 0.544, + 0.383, + 0.619 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.295, + 0.627, + 0.374, + 0.639 + ], + "angle": 0, + "content": "SD Inpainting" + }, + { + "type": "image", + "bbox": [ + 0.385, + 0.466, + 0.482, + 0.541 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.386, + 0.544, + 0.482, + 0.619 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.398, + 0.628, + 0.467, + 0.639 + ], + "angle": 0, + "content": "DreamOmni" + }, + { + "type": "image", + "bbox": [ + 0.492, + 0.466, + 0.588, + 0.541 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.492, + 0.544, + 0.588, + 0.619 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.5, + 0.627, + 0.582, + 0.64 + ], + "angle": 0, + "content": "Masked Image" + }, + { + "type": "image", + "bbox": [ + 0.591, + 0.466, + 0.686, + 0.541 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.591, + 0.544, + 0.686, + 0.619 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.611, + 0.623, + 0.674, + 0.646 + ], + "angle": 0, + "content": "ControlNet Inpainting" + }, + { + "type": "image", + "bbox": [ + 0.688, + 0.466, + 0.784, + 0.541 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.688, + 0.544, + 0.784, + 0.619 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.7, + 0.627, + 0.778, + 0.639 + ], + "angle": 0, + "content": "SD Inpainting" + }, + { + "type": "image", + "bbox": [ + 0.787, + 0.466, + 0.883, + 0.541 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.787, + 0.544, + 0.883, + 0.619 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.803, + 0.628, + 0.872, + 0.639 + ], + "angle": 0, + "content": "DreamOmni" + }, + { + "type": "image_caption", + "bbox": [ + 0.089, + 0.653, + 0.878, + 0.668 + ], + "angle": 0, + "content": "Figure 5. Visual comparison on inpainting & outpainting between DreamOmni, ControlNet-Inpainting [62] and SD-inpainting [44]." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.687, + 0.47, + 0.793 + ], + "angle": 0, + "content": "We train models with \\(1024 \\times 1024\\) size, a batch size of 256, a learning rate of \\(2 \\times 10^{-5}\\), and \\(140K\\) iterations. All experiments are conducted on 64 A100 GPUs. Additionally, to enable the model to generate images at varying resolutions, similar to the approach used in SDXL [39], we divide the images into 31 buckets based on their aspect ratios, ranging from \\(4:1\\) to \\(1:4\\), during training." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.801, + 0.21, + 0.817 + ], + "angle": 0, + "content": "4. Experiments" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.825, + 0.47, + 0.9 + ], + "angle": 0, + "content": "Evaluation on Framework. We compared several T2I model frameworks under similar settings to identify effective components. Using the same VAE, CLIP text encoder, parameters, runtime, and LAION training/testing datasets, we evaluated the Unet-based SDXL [39], DIT-" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.687, + 0.892, + 0.899 + ], + "angle": 0, + "content": "based Pixart [10], SD3-Medium [15], and our developed DreamOmni variations. Notably, as shown in Fig. 3, to facilitate comparison, we do not use the full \\(2.5B\\) parameters of DreamOmni, but instead adjust the parameters of all models to \\(0.85B\\). DreamOmni-V1 has two downsampling layers (\\(2\\times\\) and \\(4\\times\\)) like SDXL but lacks an Unet connection. DreamOmni-V2 builds on DreamOmni-V1 by adding the Unet connection, while DreamOmni-V3 takes DreamOmni-V2 further by focusing all DIT operations on \\(2\\times\\) downsampled latent. (1) Models with the Unet connection (SDXL, DreamOmni-V2, DreamOmni-V3) showed significantly faster convergence than those without (SD3-Medium, DreamOmni-V1). DreamOmni-V3, notably, converges four times faster than the SD3-Medium, enhancing" + }, + { + "type": "page_number", + "bbox": [ + 0.479, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "28538" + } + ], + [ + { + "type": "image_caption", + "bbox": [ + 0.092, + 0.097, + 0.186, + 0.16 + ], + "angle": 0, + "content": "A humanoid cat from the race american shorthair with black fur. Standing as a barbarian tabaxi with a greatsword. Dressed in light clothing, something like the ancient geeks weared, photorealistic. standing in front of an oasis in the dessert." + }, + { + "type": "image_caption", + "bbox": [ + 0.092, + 0.196, + 0.186, + 0.232 + ], + "angle": 0, + "content": "elf girl travelling alongside a majestRyuic dragon, close-up view, anime style, high fantasy, digital art, masterpiece painting" + }, + { + "type": "image_caption", + "bbox": [ + 0.108, + 0.251, + 0.151, + 0.264 + ], + "angle": 0, + "content": "Prompt" + }, + { + "type": "image", + "bbox": [ + 0.192, + 0.093, + 0.287, + 0.167 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.192, + 0.171, + 0.286, + 0.245 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.218, + 0.247, + 0.256, + 0.258 + ], + "angle": 0, + "content": "Image" + }, + { + "type": "image_caption", + "bbox": [ + 0.211, + 0.26, + 0.266, + 0.269 + ], + "angle": 0, + "content": "Condition" + }, + { + "type": "image", + "bbox": [ + 0.29, + 0.093, + 0.385, + 0.167 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.29, + 0.171, + 0.385, + 0.245 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.306, + 0.251, + 0.368, + 0.263 + ], + "angle": 0, + "content": "ControlNet" + }, + { + "type": "image", + "bbox": [ + 0.387, + 0.094, + 0.484, + 0.167 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.387, + 0.171, + 0.485, + 0.245 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.4, + 0.247, + 0.468, + 0.258 + ], + "angle": 0, + "content": "DreamOmni" + }, + { + "type": "image_caption", + "bbox": [ + 0.418, + 0.259, + 0.454, + 0.27 + ], + "angle": 0, + "content": "(Ours)" + }, + { + "type": "image_caption", + "bbox": [ + 0.49, + 0.097, + 0.582, + 0.166 + ], + "angle": 0, + "content": "A cute teenage-child fairy with blue hair and butterflies in her ears, holding balloons and a birthday cake with a candle lit on it in her other hand. Her hand is holding the plate. The artwork is in the style pixar anime. her facial expression is very happy and excited but in a cute way" + }, + { + "type": "image_caption", + "bbox": [ + 0.493, + 0.199, + 0.581, + 0.214 + ], + "angle": 0, + "content": "ironman is riding a white little pony" + }, + { + "type": "image_caption", + "bbox": [ + 0.506, + 0.252, + 0.546, + 0.264 + ], + "angle": 0, + "content": "Prompt" + }, + { + "type": "image", + "bbox": [ + 0.587, + 0.093, + 0.683, + 0.167 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.588, + 0.171, + 0.682, + 0.245 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.618, + 0.247, + 0.652, + 0.258 + ], + "angle": 0, + "content": "Image" + }, + { + "type": "image_caption", + "bbox": [ + 0.607, + 0.259, + 0.662, + 0.269 + ], + "angle": 0, + "content": "Condition" + }, + { + "type": "image", + "bbox": [ + 0.685, + 0.093, + 0.779, + 0.167 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.685, + 0.171, + 0.779, + 0.245 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.702, + 0.252, + 0.764, + 0.262 + ], + "angle": 0, + "content": "ControlNet" + }, + { + "type": "image", + "bbox": [ + 0.782, + 0.093, + 0.878, + 0.167 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.782, + 0.171, + 0.878, + 0.245 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.797, + 0.247, + 0.864, + 0.258 + ], + "angle": 0, + "content": "DreamOmni" + }, + { + "type": "image_caption", + "bbox": [ + 0.814, + 0.26, + 0.85, + 0.272 + ], + "angle": 0, + "content": "(Ours)" + }, + { + "type": "image_caption", + "bbox": [ + 0.076, + 0.28, + 0.893, + 0.322 + ], + "angle": 0, + "content": "Figure 6. Visual comparison on image-conditioned generation. Our DreamOmni excels at using canny, depth, and segmentation maps as conditions for image generation. Compared to the classic ControlNet [62], DreamOmni not only follows the user's prompts and image conditions more accurately but also generates content and color schemes that are visually more pleasing." + }, + { + "type": "image_caption", + "bbox": [ + 0.088, + 0.338, + 0.174, + 0.395 + ], + "angle": 0, + "content": "A girl (in the given image) is playing a red electric guitar. She is wearing a black jacket and a banie, and she has headphones. The background features autumn leaves, suggesting a seasonal setting." + }, + { + "type": "image_caption", + "bbox": [ + 0.088, + 0.401, + 0.172, + 0.471 + ], + "angle": 0, + "content": "Generate an image using the character from the supplied image as reference. A girl is holding a lantern. She is surrounded by vibrant flowers and a whimsical, magical atmosphere. The girl is wearing a blue and white outfit with a ruffled collar, and she has a thoughtful expression." + }, + { + "type": "image_caption", + "bbox": [ + 0.109, + 0.477, + 0.146, + 0.488 + ], + "angle": 0, + "content": "Prompt" + }, + { + "type": "image", + "bbox": [ + 0.177, + 0.336, + 0.231, + 0.4 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.18, + 0.403, + 0.23, + 0.469 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.188, + 0.474, + 0.227, + 0.493 + ], + "angle": 0, + "content": "Image condition" + }, + { + "type": "image", + "bbox": [ + 0.244, + 0.336, + 0.325, + 0.4 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.244, + 0.403, + 0.325, + 0.469 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.25, + 0.478, + 0.324, + 0.488 + ], + "angle": 0, + "content": "BLIP-Diffusion" + }, + { + "type": "image", + "bbox": [ + 0.327, + 0.336, + 0.408, + 0.4 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.327, + 0.403, + 0.409, + 0.469 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.346, + 0.478, + 0.395, + 0.488 + ], + "angle": 0, + "content": "IP-adapter" + }, + { + "type": "image", + "bbox": [ + 0.41, + 0.336, + 0.493, + 0.4 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.41, + 0.403, + 0.493, + 0.469 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.423, + 0.473, + 0.481, + 0.483 + ], + "angle": 0, + "content": "DreamOmni" + }, + { + "type": "image_caption", + "bbox": [ + 0.438, + 0.484, + 0.469, + 0.494 + ], + "angle": 0, + "content": "(Ours)" + }, + { + "type": "image_caption", + "bbox": [ + 0.501, + 0.346, + 0.557, + 0.389 + ], + "angle": 0, + "content": "A man (in the given image) is wearing a white t-shirt and white pants. He is smiling." + }, + { + "type": "image_caption", + "bbox": [ + 0.503, + 0.416, + 0.557, + 0.458 + ], + "angle": 0, + "content": "A close-up of a woman (in the given image) with her head tilted slightly to the side." + }, + { + "type": "image_caption", + "bbox": [ + 0.512, + 0.478, + 0.548, + 0.489 + ], + "angle": 0, + "content": "Prompt" + }, + { + "type": "image", + "bbox": [ + 0.562, + 0.336, + 0.627, + 0.4 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.562, + 0.403, + 0.624, + 0.469 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.572, + 0.474, + 0.619, + 0.493 + ], + "angle": 0, + "content": "Image Condition" + }, + { + "type": "image", + "bbox": [ + 0.636, + 0.336, + 0.717, + 0.4 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.636, + 0.403, + 0.717, + 0.468 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.642, + 0.478, + 0.716, + 0.488 + ], + "angle": 0, + "content": "BLIP-Diffusion" + }, + { + "type": "image", + "bbox": [ + 0.72, + 0.336, + 0.799, + 0.4 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.719, + 0.403, + 0.799, + 0.468 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.738, + 0.478, + 0.787, + 0.488 + ], + "angle": 0, + "content": "IP-adapter" + }, + { + "type": "image", + "bbox": [ + 0.802, + 0.336, + 0.884, + 0.4 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.802, + 0.403, + 0.884, + 0.468 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.815, + 0.473, + 0.873, + 0.483 + ], + "angle": 0, + "content": "DreamOmni" + }, + { + "type": "image_caption", + "bbox": [ + 0.83, + 0.484, + 0.86, + 0.494 + ], + "angle": 0, + "content": "(Ours)" + }, + { + "type": "image_caption", + "bbox": [ + 0.076, + 0.504, + 0.893, + 0.546 + ], + "angle": 0, + "content": "Figure 7. Visual comparison on subject-driven generation. Compared with BLIP-Diffusion [28] and IP-adapter [60], DreamOmni excels at accurately following user prompts while preserving the specified subject. Furthermore, its success across both photo and anime cases underscores DreamOmni's generalization capabilities." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.562, + 0.469, + 0.621 + ], + "angle": 0, + "content": "both training and fine-tuning efficiency. (2) Comparing DreamOmni-V3 and DreamOmni-V2, we found that concentrating DIT block computations on higher-resolution latent \\((2\\times)\\) is more cost-effective." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.625, + 0.469, + 0.777 + ], + "angle": 0, + "content": "Evaluation on T2I generation. As shown in Tab. 1, our synthetic data significantly enhances DreamOmni's T2I generation capabilities in aspects such as quantity, color, and position, enabling our model to achieve SOTA results on GenEval [19]. Notably, SD3-Medium [15] is a \\(2B\\) open-source SOTA T2I model, which has a similar parameters as our DreamOmni. Furthermore, qualitative results are shown in Fig. 4. We can see that DreamOmni's outputs are not only more visually pleasing but also align more accurately with the given prompts." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.78, + 0.469, + 0.901 + ], + "angle": 0, + "content": "Evaluation on Inpainting. We compare our DreamOmni with ControlNet-Inpainting [62] and SD-Inpainting [44] on our high-quality evaluation datasets to evaluate its performance. The quantitative results, presented in Tab. 2, demonstrate that DreamOmni significantly outperforms both ControlNet-Inpainting and SD-Inpainting, highlighting its superior generation quality and coherence. Visual results (Fig. 5) further emphasize that DreamOmni ex" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.562, + 0.892, + 0.622 + ], + "angle": 0, + "content": "cels in generating fine details, surpassing both ControlNet-Inpainting and SD-Inpainting. Additionally, DreamOmni effectively handles large mask holes, generating realistic content instead of blurry and incoherent outputs." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.625, + 0.892, + 0.76 + ], + "angle": 0, + "content": "Evaluation on Reference Image Generation. For image-conditioned generation, we conducted a comparison with ControlNet [62] on canny, depth, and segmentation image conditions. Visual results are presented in Fig. 6. These results reveal that, across all tested conditions, our DreamOmni significantly outperforms ControlNet. Our method not only adheres more faithfully to both the image condition and the prompt but also demonstrates enhanced visual quality with better composition and richer detail." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.763, + 0.892, + 0.883 + ], + "angle": 0, + "content": "For subject-driven image generation, we compared our method with competitive approaches, including BLIP-Diffusion [28] and IP-Adapter [60]. To demonstrate DreamOmni's powerful generalization, we validated their performance on both anime and photographic images. Visual results are shown in Fig. 7. Compared to other methods, our DreamOmni not only preserves the specified subject effectively but also follows the prompt well." + }, + { + "type": "text", + "bbox": [ + 0.5, + 0.886, + 0.892, + 0.902 + ], + "angle": 0, + "content": "Evaluation on Instruction Editing. We compare" + }, + { + "type": "page_number", + "bbox": [ + 0.48, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "28539" + } + ], + [ + { + "type": "image_caption", + "bbox": [ + 0.086, + 0.113, + 0.149, + 0.147 + ], + "angle": 0, + "content": "Remove the star ornament from the snow-covered tree." + }, + { + "type": "image_caption", + "bbox": [ + 0.088, + 0.181, + 0.148, + 0.206 + ], + "angle": 0, + "content": "Add a straw hat on the person's head." + }, + { + "type": "image_caption", + "bbox": [ + 0.093, + 0.228, + 0.145, + 0.238 + ], + "angle": 0, + "content": "Instruction" + }, + { + "type": "image", + "bbox": [ + 0.152, + 0.094, + 0.236, + 0.159 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.236, + 0.094, + 0.318, + 0.159 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.318, + 0.094, + 0.401, + 0.159 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.401, + 0.094, + 0.485, + 0.159 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.154, + 0.161, + 0.236, + 0.225 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.167, + 0.228, + 0.225, + 0.24 + ], + "angle": 0, + "content": "Input Image" + }, + { + "type": "image", + "bbox": [ + 0.236, + 0.161, + 0.318, + 0.225 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.252, + 0.228, + 0.306, + 0.239 + ], + "angle": 0, + "content": "InstructP2P" + }, + { + "type": "image", + "bbox": [ + 0.318, + 0.161, + 0.401, + 0.225 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.345, + 0.228, + 0.375, + 0.239 + ], + "angle": 0, + "content": "MGIE" + }, + { + "type": "image", + "bbox": [ + 0.402, + 0.161, + 0.486, + 0.225 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.413, + 0.228, + 0.472, + 0.239 + ], + "angle": 0, + "content": "DreamOmni" + }, + { + "type": "image_caption", + "bbox": [ + 0.498, + 0.115, + 0.546, + 0.14 + ], + "angle": 0, + "content": "Replace the owl with a hawk" + }, + { + "type": "image_caption", + "bbox": [ + 0.498, + 0.175, + 0.546, + 0.216 + ], + "angle": 0, + "content": "Remove the green leafy plant near the yellow flower." + }, + { + "type": "image_caption", + "bbox": [ + 0.495, + 0.228, + 0.547, + 0.239 + ], + "angle": 0, + "content": "Instruction" + }, + { + "type": "image", + "bbox": [ + 0.552, + 0.094, + 0.634, + 0.159 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.552, + 0.161, + 0.634, + 0.225 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.567, + 0.228, + 0.624, + 0.239 + ], + "angle": 0, + "content": "Input Image" + }, + { + "type": "image", + "bbox": [ + 0.636, + 0.094, + 0.716, + 0.159 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.636, + 0.161, + 0.717, + 0.225 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.654, + 0.228, + 0.708, + 0.239 + ], + "angle": 0, + "content": "InstructP2P" + }, + { + "type": "image", + "bbox": [ + 0.718, + 0.094, + 0.799, + 0.159 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.718, + 0.161, + 0.799, + 0.225 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.747, + 0.228, + 0.777, + 0.238 + ], + "angle": 0, + "content": "MGIE" + }, + { + "type": "image", + "bbox": [ + 0.802, + 0.094, + 0.885, + 0.159 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.802, + 0.161, + 0.885, + 0.225 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.816, + 0.228, + 0.875, + 0.238 + ], + "angle": 0, + "content": "DreamOmni" + }, + { + "type": "image_caption", + "bbox": [ + 0.076, + 0.25, + 0.892, + 0.279 + ], + "angle": 0, + "content": "Figure 8. Visual comparison on instruction-based editing. Our DreamOmni achieves more precise editing (including addition, removal, and replacement) compared to competitive methods such as MGIE [17] and InstructP2P [8]." + }, + { + "type": "table_caption", + "bbox": [ + 0.086, + 0.293, + 0.46, + 0.307 + ], + "angle": 0, + "content": "Table 2. Quantitative comparison for inpainting & outpainting." + }, + { + "type": "table", + "bbox": [ + 0.082, + 0.313, + 0.47, + 0.399 + ], + "angle": 0, + "content": "
ModelInpaintingOutpainting
FID↓LPIPS↓FID↓LPIPS↓
SD-inpainting [44]1.35220.15602.91790.2475
ControlNet-inpainting [62]1.83930.15944.23370.2521
DreamOmni (Ours)0.83710.12031.69260.1995
" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.413, + 0.468, + 0.594 + ], + "angle": 0, + "content": "DreamOmni with competitive methods such as MGIE [17] and InstructP2P [8]. The visual results are presented in Fig. 8. We can see that DreamOmni performs more accurate edits, including addition, removal, and replacement. Specifically, our editing results exhibit superior consistency with minimal changes to non-edited areas and higher-quality generation of the edited content. This further validates that our synthetic data pipeline is an efficient and effective method for creating instruction-based editing datasets, enabling models to learn precise instruction-based editing. Moreover, the efficiency of our synthetic data pipeline allows for models to easily scale up diverse training data." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.595, + 0.468, + 0.806 + ], + "angle": 0, + "content": "Evaluation on Drag Editing. We evaluate DreamOmni on our synthetic evaluation dataset. The visual results are shown in Fig. 9. (1) Compared to the target image, we can see that DreamOmni can accurately perform translation, rotation, and scaling drag edits. (2) for translation and scaling, DreamOmni can maintain the integrity of the dragged object. However, large-scale rotation operations are more challenging for DreamOmni, as they involve complex transformations of the object itself, which may lead to deformations of the edited object. (3) The results demonstrate the effectiveness of our synthetic data pipeline for drag editing, and encoding the drag point positions and displacement information as instruction inputs allows the model to learn precise drag edits (shown in Fig. 2)." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.815, + 0.196, + 0.831 + ], + "angle": 0, + "content": "5. Conclusion" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.84, + 0.469, + 0.901 + ], + "angle": 0, + "content": "Current T2I foundation models lack a unified framework and training for downstream tasks, such as image editing. To address this, we introduce DreamOmni, a unified model for T2I generation and editing. We evaluate the frameworks" + }, + { + "type": "image", + "bbox": [ + 0.514, + 0.296, + 0.635, + 0.389 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.514, + 0.39, + 0.635, + 0.483 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.514, + 0.484, + 0.635, + 0.579 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.518, + 0.583, + 0.619, + 0.613 + ], + "angle": 0, + "content": "Input Image & Drag Points" + }, + { + "type": "image", + "bbox": [ + 0.638, + 0.296, + 0.757, + 0.389 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.638, + 0.39, + 0.757, + 0.483 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.638, + 0.484, + 0.757, + 0.579 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.657, + 0.59, + 0.75, + 0.605 + ], + "angle": 0, + "content": "Target Image" + }, + { + "type": "image", + "bbox": [ + 0.76, + 0.296, + 0.88, + 0.389 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.76, + 0.39, + 0.88, + 0.483 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.76, + 0.484, + 0.88, + 0.579 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.777, + 0.59, + 0.863, + 0.604 + ], + "angle": 0, + "content": "DreamOmni" + }, + { + "type": "image_caption", + "bbox": [ + 0.499, + 0.624, + 0.892, + 0.652 + ], + "angle": 0, + "content": "Figure 9. Visual results of drag editing: DreamOmni accurately performs translation, rotation, and scaling edits." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.674, + 0.892, + 0.901 + ], + "angle": 0, + "content": "of existing models under fair settings and consider the specific needs of various editing tasks. From this analysis, we develop a framework that integrates T2I with various editing tasks. Besides, a challenge in training editing models is the creation of high-quality, large-scale editing data, which is inefficient. To overcome this, we designed a synthetic collage data pipeline capable of efficiently generating quite a few precise, high-quality editing data. Moreover, the pipeline enhances the model's generation accuracy in text, position, quantity, color, and geometry. By jointly training on T2I and multi-task synthetic data, we develop a native, unified model for both image generation and editing. T2I training strengthens the model's grasp of specific concepts and improves generation quality, while editing training enables it to handle the requirements of editing tasks." + }, + { + "type": "page_number", + "bbox": [ + 0.479, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "28540" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.078, + 0.092, + 0.473, + 0.138 + ], + "angle": 0, + "content": "Acknowledgements. This work was supported in part by the Research Grants Council under the Areas of Excellence scheme grant AoE/E-601/22-R." + }, + { + "type": "title", + "bbox": [ + 0.08, + 0.151, + 0.175, + 0.167 + ], + "angle": 0, + "content": "References" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.176, + 0.47, + 0.218 + ], + "angle": 0, + "content": "[1] Jacob Austin, Daniel D Johnson, Jonathan Ho, Daniel Tarlow, and Rianne van den Berg. Structured denoising diffusion models in discrete state-spaces. NeurIPS, 2021. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.219, + 0.472, + 0.26 + ], + "angle": 0, + "content": "[2] Omri Avrahami, Dani Lischinski, and Ohad Fried. Blended diffusion for text-driven editing of natural images. In CVPR, 2022. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.263, + 0.472, + 0.331 + ], + "angle": 0, + "content": "[3] Yogesh Balaji, Seungjun Nah, Xun Huang, Arash Vahdat, Jiaming Song, Qinsheng Zhang, Karsten Kreis, Miika Aittala, Timo Aila, Samuli Laine, et al. ediff-i: Text-to-image diffusion models with an ensemble of expert denoisers. arXiv preprint arXiv:2211.01324, 2022. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.333, + 0.47, + 0.374 + ], + "angle": 0, + "content": "[4] Fan Bao, Shen Nie, Kaiwen Xue, Yue Cao, Chongxuan Li, Hang Su, and Jun Zhu. All are worth words: A vit backbone for diffusion models. In CVPR, 2023. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.376, + 0.47, + 0.431 + ], + "angle": 0, + "content": "[5] Georgios Batzolis, Jan Stanczuk, Carola-Bibiane Schonlieb, and Christian Etmann. Conditional image generation with score-based diffusion models. arXiv preprint arXiv:2111.13606, 2021. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.433, + 0.47, + 0.502 + ], + "angle": 0, + "content": "[6] James Betker, Gabriel Goh, Li Jing, Tim Brooks, Jianfeng Wang, Linjie Li, Long Ouyang, Juntang Zhuang, Joyce Lee, Yufei Guo, et al. Improving image generation with better captions. Computer Science. https://cdn.openai.com/papers/dall-e-3.pdf, 2023. 1, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.504, + 0.47, + 0.572 + ], + "angle": 0, + "content": "[7] Manuel Brack, Felix Friedrich, Katharina Kornmeier, Linoy Tsaban, Patrick Schramowski, Kristian Kersting, and Apolinário Passos. *Ledits++: Limitless image editing using text-to-image models.* arXiv preprint arXiv:2311.16711, 2023. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.575, + 0.469, + 0.616 + ], + "angle": 0, + "content": "[8] Tim Brooks, Aleksander Holynski, and Alexei A Efros. Instructpix2pix: Learning to follow image editing instructions. In CVPR, 2023. 2, 3, 4, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.618, + 0.47, + 0.673 + ], + "angle": 0, + "content": "[9] Mingdeng Cao, Xintao Wang, Zhongang Qi, Ying Shan, Xiaohu Qie, and Yinqiang Zheng. Masactrl: Tuning-free mutual self-attention control for consistent image synthesis and editing. In ICCV, 2023. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.078, + 0.675, + 0.47, + 0.743 + ], + "angle": 0, + "content": "[10] Junsong Chen, Jincheng Yu, Chongjian Ge, Lewei Yao, Enze Xie, Yue Wu, Zhongdao Wang, James Kwok, Ping Luo, Huchuan Lu, et al. Pixart-α: Fast training of diffusion transformer for photorealistic text-to-image synthesis. ICLR, 2024. 3, 5, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.078, + 0.746, + 0.47, + 0.815 + ], + "angle": 0, + "content": "[11] Zhe Chen, Weiyun Wang, Hao Tian, Shenglong Ye, Zhang-wei Gao, Erfei Cui, Wenwen Tong, Kongzhi Hu, Jiapeng Luo, Zheng Ma, et al. How far are we to gpt-4v? closing the gap to commercial multimodal models with open-source suites. arXiv preprint arXiv:2404.16821, 2024. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.078, + 0.817, + 0.47, + 0.871 + ], + "angle": 0, + "content": "[12] Guillaume Couairon, Jakob Verbeek, Holger Schwenk, and Matthieu Cord. Diffedit: Diffusion-based semantic image editing with mask guidance. arXiv preprint arXiv:2210.11427, 2022.3" + }, + { + "type": "ref_text", + "bbox": [ + 0.078, + 0.873, + 0.47, + 0.901 + ], + "angle": 0, + "content": "[13] Xiaoliang Dai, Ji Hou, Chih-Yao Ma, Sam Tsai, Jialiang Wang, Rui Wang, Peizhao Zhang, Simon Vandenhende, Xi" + }, + { + "type": "list", + "bbox": [ + 0.078, + 0.176, + 0.472, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.533, + 0.093, + 0.892, + 0.134 + ], + "angle": 0, + "content": "aofang Wang, Abhimanyu Dubey, et al. Emu: Enhancing image generation models using photogenic needles in a haystack. arXiv preprint arXiv:2309.15807, 2023. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.135, + 0.892, + 0.162 + ], + "angle": 0, + "content": "[14] Prafulla Dhariwal and Alexander Nichol. Diffusion models beat gans on image synthesis. NeurIPS, 2021. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.163, + 0.892, + 0.231 + ], + "angle": 0, + "content": "[15] Patrick Esser, Sumith Kulal, Andreas Blattmann, Rahim Entezari, Jonas Müller, Harry Saini, Yam Levi, Dominik Lorenz, Axel Sauer, Frederic Boessel, et al. Scaling rectified flow transformers for high-resolution image synthesis. In ICML, 2024. 1, 3, 4, 5, 6, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.233, + 0.892, + 0.301 + ], + "angle": 0, + "content": "[16] Zhida Feng, Zhenyu Zhang, Xintong Yu, Yewei Fang, Lanxin Li, Xuyi Chen, Yuxiang Lu, Jiaxiang Liu, Weichong Yin, Shikun Feng, et al. Ernie-vilg 2.0: Improving text-to-image diffusion model with knowledge-enhanced mixture-of-denoising-experts. In CVPR, 2023. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.302, + 0.892, + 0.356 + ], + "angle": 0, + "content": "[17] Tsu-Jui Fu, Wenze Hu, Xianzhi Du, William Yang Wang, Yinfei Yang, and Zhe Gan. Guiding instruction-based image editing via multimodal large language models. *ICLR*, 2024. 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.358, + 0.892, + 0.425 + ], + "angle": 0, + "content": "[18] Zigang Geng, Binxin Yang, Tiankai Hang, Chen Li, Shuyang Gu, Ting Zhang, Jianmin Bao, Zheng Zhang, Han Hu, Dong Chen, et al. Instructdiffusion: A generalist modeling interface for vision tasks. arXiv preprint arXiv:2309.03895, 2023. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.427, + 0.892, + 0.468 + ], + "angle": 0, + "content": "[19] Dhruba Ghosh, Hannaneh Hajishirzi, and Ludwig Schmidt. Geneval: An object-focused framework for evaluating text-to-image alignment. NeurIPS, 36, 2024. 5, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.469, + 0.892, + 0.522 + ], + "angle": 0, + "content": "[20] Ian Goodfellow, Jean Pouget-Abadie, Mehdi Mirza, Bing Xu, David Warde-Farley, Sherjil Ozair, Aaron Courville, and Yoshua Bengio. Generative adversarial nets. NeurIPS, 2014. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.525, + 0.892, + 0.579 + ], + "angle": 0, + "content": "[21] Shuyang Gu, Dong Chen, Jianmin Bao, Fang Wen, Bo Zhang, Dongdong Chen, Lu Yuan, and Baining Guo. Vector quantized diffusion model for text-to-image synthesis. In CVPR, 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.581, + 0.892, + 0.635 + ], + "angle": 0, + "content": "[22] Amir Hertz, Ron Mokady, Jay Tenenbaum, Kfir Aberman, Yael Pritch, and Daniel Cohen-Or. Prompt-to-prompt image editing with cross attention control. arXiv preprint arXiv:2208.01626, 2022. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.637, + 0.892, + 0.663 + ], + "angle": 0, + "content": "[23] Jonathan Ho, Ajay Jain, and Pieter Abbeel. Denoising diffusion probabilistic models. NeurIPS, 2020. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.665, + 0.892, + 0.706 + ], + "angle": 0, + "content": "[24] Jonathan Ho, Chitwan Sahara, William Chan, David J Fleet, Mohammad Norouzi, and Tim Salimans. Cascaded diffusion models for high fidelity image generation. JMLR, 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.707, + 0.892, + 0.775 + ], + "angle": 0, + "content": "[25] Yuzhou Huang, Liangbin Xie, Xintao Wang, Ziyang Yuan, Xiaodong Cun, Yixiao Ge, Jiantao Zhou, Chao Dong, Rui Huang, Ruimao Zhang, et al. Smartedit: Exploring complex instruction-based image editing with multimodal large language models. CVPR, 2024. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.776, + 0.892, + 0.803 + ], + "angle": 0, + "content": "[26] Diederik P Kingma. Auto-encoding variational bayes. arXiv preprint arXiv:1312.6114, 2013. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.804, + 0.892, + 0.859 + ], + "angle": 0, + "content": "[27] Daiqing Li, Aleks Kamko, Ehsan Akhgari, Ali Sabet, Linmiao Xu, and Suhail Doshi. Playground v2. 5: Three insights towards enhancing aesthetic quality in text-to-image generation. arXiv preprint arXiv:2402.17245, 2024. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.86, + 0.892, + 0.901 + ], + "angle": 0, + "content": "[28] Dongxu Li, Junnan Li, and Steven Hoi. Blip-diffusion: Pretrained subject representation for controllable text-to-image generation and editing. *NeurIPS*, 2024. 3, 4, 7" + }, + { + "type": "list", + "bbox": [ + 0.504, + 0.093, + 0.892, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.48, + 0.945, + 0.518, + 0.957 + ], + "angle": 0, + "content": "28541" + } + ], + [ + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.092, + 0.47, + 0.147 + ], + "angle": 0, + "content": "[29] Ming Li, Taojiannan Yang, Huafeng Kuang, Jie Wu, Zhaoning Wang, Xuefeng Xiao, and Chen Chen. Controlnet ++: Improving conditional controls with efficient consistency feedback. In ECCV, 2025.3" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.149, + 0.47, + 0.204 + ], + "angle": 0, + "content": "[30] Zhen Li, Mingdeng Cao, Xintao Wang, Zhongang Qi, MingMing Cheng, and Ying Shan. Photomaker: Customizing realistic human photos via stacked id embedding. In CVPR, 2024. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.207, + 0.469, + 0.288 + ], + "angle": 0, + "content": "[31] Zhimin Li, Jianwei Zhang, Qin Lin, Jiangfeng Xiong, Yanxin Long, Xinchi Deng, Yingfang Zhang, Xingchao Liu, Minbin Huang, Zedong Xiao, et al. Hunyuan-dit: A powerful multi-resolution diffusion transformer with fine-grained chinese understanding. arXiv preprint arXiv:2405.08748, 2024. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.291, + 0.469, + 0.332 + ], + "angle": 0, + "content": "[32] Pengyang Ling, Lin Chen, Pan Zhang, Huaian Chen, Yi Jin, and Jinjin Zheng. Freedrag: Feature dragging for reliable point-based image editing. In CVPR, 2024. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.334, + 0.469, + 0.375 + ], + "angle": 0, + "content": "[33] Xingchao Liu, Chengyue Gong, and Qiang Liu. Flow straight and fast: Learning to generate and transfer data with rectified flow. arXiv preprint arXiv:2209.03003, 2022. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.377, + 0.469, + 0.445 + ], + "angle": 0, + "content": "[34] Hayk Manukyan, Andranik Sargsyan, Barsegh Atanyan, Zhangyang Wang, Shant Navasardyan, and Humphrey Shi. Hd-painter: high-resolution and prompt-faithful text-guided image inpainting with diffusion models. arXiv preprint arXiv:2312.14091, 2023. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.448, + 0.469, + 0.516 + ], + "angle": 0, + "content": "[35] Alex Nichol, Prafulla Dhariwal, Aditya Ramesh, Pranav Shyam, Pamela Mishkin, Bob McGrew, Ilya Sutskever, and Mark Chen. Glide: Towards photorealistic image generation and editing with text-guided diffusion models. arXiv preprint arXiv:2112.10741, 2021. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.519, + 0.469, + 0.573 + ], + "angle": 0, + "content": "[36] Xichen Pan, Li Dong, Shaohan Huang, Zhiliang Peng, Wenhu Chen, and Furu Wei. Kosmos-g: Generating images in context with multimodal large language models. arXiv preprint arXiv:2310.02992, 2023. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.575, + 0.469, + 0.616 + ], + "angle": 0, + "content": "[37] Gaurav Parmar, Krishna Kumar Singh, Richard Zhang, Yijun Li, Jingwan Lu, and Jun-Yan Zhu. Zero-shot image-to-image translation. In ACM SIGGRAPH, 2023. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.618, + 0.469, + 0.645 + ], + "angle": 0, + "content": "[38] William Peebles and Saining Xie. Scalable diffusion models with transformers. In ICCV, 2023. 2, 3, 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.648, + 0.469, + 0.715 + ], + "angle": 0, + "content": "[39] Dustin Podell, Zion English, Kyle Lacey, Andreas Blattmann, Tim Dockhorn, Jonas Müller, Joe Penna, and Robin Rombach. Sdxl: Improving latent diffusion models for high-resolution image synthesis. NeurIPS, 2023. 1, 3, 4, 5, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.718, + 0.469, + 0.786 + ], + "angle": 0, + "content": "[40] Can Qin, Shu Zhang, Ning Yu, Yihao Feng, Xinyi Yang, Yingbo Zhou, Huan Wang, Juan Carlos Niebles, Caiming Xiong, Silvio Savarese, et al. Unicontrol: A unified diffusion model for controllable visual generation in the wild. arXiv preprint arXiv:2305.11147, 2023. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.789, + 0.469, + 0.856 + ], + "angle": 0, + "content": "[41] Alec Radford, Jong Wook Kim, Chris Hallacy, Aditya Ramesh, Gabriel Goh, Sandhini Agarwal, Girish Sastry, Amanda Askell, Pamela Mishkin, Jack Clark, et al. Learning transferable visual models from natural language supervision. In ICML, 2021. 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.859, + 0.469, + 0.901 + ], + "angle": 0, + "content": "[42] Colin Raffel, Noam Shazeer, Adam Roberts, Katherine Lee, Sharan Narang, Michael Matena, Yanqi Zhou, Wei Li, and Peter J Liu. Exploring the limits of transfer learning with a" + }, + { + "type": "list", + "bbox": [ + 0.08, + 0.092, + 0.47, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.535, + 0.093, + 0.892, + 0.12 + ], + "angle": 0, + "content": "unified text-to-text transformer. Journal of machine learning research, 2020. 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.123, + 0.892, + 0.164 + ], + "angle": 0, + "content": "[43] Aditya Ramesh, Mikhail Pavlov, Gabriel Goh, Scott Gray, Chelsea Voss, Alec Radford, Mark Chen, and Ilya Sutskever. Zero-shot text-to-image generation. In ICML, 2021. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.167, + 0.892, + 0.221 + ], + "angle": 0, + "content": "[44] Robin Rombach, Andreas Blattmann, Dominik Lorenz, Patrick Esser, and Björn Ommer. High-resolution image synthesis with latent diffusion models. In CVPR, 2022. 1, 2, 4, 5, 6, 7, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.224, + 0.892, + 0.264 + ], + "angle": 0, + "content": "[45] Olaf Ronneberger, Philipp Fischer, and Thomas Brox. U-net: Convolutional networks for biomedical image segmentation. In MICCAI, 2015. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.268, + 0.892, + 0.323 + ], + "angle": 0, + "content": "[46] Nataniel Ruiz, Yuanzhen Li, Varun Jampani, Yael Pritch, Michael Rubinstein, and Kfir Aberman. Dreambooth: Fine tuning text-to-image diffusion models for subject-driven generation. In CVPR, 2023. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.326, + 0.892, + 0.395 + ], + "angle": 0, + "content": "[47] Chitwan Sahara, William Chan, Saurabh Saxena, Lala Li, Jay Whang, Emily L Denton, Kamyar Ghasemipour, Raphael Gontijo Lopes, Burcu Karagol Ayan, Tim Salimans, et al. Photorealistic text-to-image diffusion models with deep language understanding. NeurIPS, 2022. 1, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.398, + 0.892, + 0.466 + ], + "angle": 0, + "content": "[48] Christoph Schuhmann, Romain Beaumont, Richard Vencu, Cade Gordon, Ross Wightman, Mehdi Cherti, Theo Coombes, Aarush Katta, Clayton Mullis, Mitchell Wortsman, et al. Laion-5b: An open large-scale dataset for training next generation image-text models. NeurIPS, 2022. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.469, + 0.892, + 0.524 + ], + "angle": 0, + "content": "[49] Shelly Sheynin, Adam Polyak, Uriel Singer, Yuval Kirstain, Amit Zohar, Oron Ashual, Devi Parikh, and Yaniv Taigman. Emu edit: Precise image editing via recognition and generation tasks. arXiv preprint arXiv:2311.10089, 2023. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.527, + 0.892, + 0.58 + ], + "angle": 0, + "content": "[50] Yujun Shi, Jun Hao Liew, Hanshu Yan, Vincent YF Tan, and Jiashi Feng. Instadrag: Lightning fast and accurate drag-based image editing emerging from videos. arXiv preprint arXiv:2405.13722, 2024. 3, 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.584, + 0.892, + 0.64 + ], + "angle": 0, + "content": "[51] Yujun Shi, Chuhui Xue, Jun Hao Liew, Jiachun Pan, Hanshu Yan, Wenqing Zhang, Vincent YF Tan, and Song Bai. Dragdiffusion: Harnessing diffusion models for interactive point-based image editing. In CVPR, 2024. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.642, + 0.892, + 0.684 + ], + "angle": 0, + "content": "[52] Jascha Sohl-Dickstein, Eric Weiss, Niru Maheswaranathan, and Surya Ganguli. Deep unsupervised learning using nonequilibrium thermodynamics. In ICML, 2015. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.686, + 0.892, + 0.741 + ], + "angle": 0, + "content": "[53] Yang Song, Jascha Sohl-Dickstein, Diederik P Kingma, Abhishek Kumar, Stefano Ermon, and Ben Poole. Score-based generative modeling through stochastic differential equations. arXiv preprint arXiv:2011.13456, 2020. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.744, + 0.892, + 0.785 + ], + "angle": 0, + "content": "[54] Narek Tumanyan, Michal Geyer, Shai Bagon, and Tali Dekel. Plug-and-play diffusion features for text-driven image-to-image translation. In CVPR, 2023. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.788, + 0.892, + 0.828 + ], + "angle": 0, + "content": "[55] Bram Wallace, Akash Gokul, and Nikhil Naik. Edict: Exact diffusion inversion via coupled transformations. In CVPR, 2023. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.831, + 0.892, + 0.9 + ], + "angle": 0, + "content": "[56] Peng Wang, Shuai Bai, Sinan Tan, Shijie Wang, Zhihao Fan, Jinze Bai, Keqin Chen, Xuejing Liu, Jialin Wang, Wenbin Ge, et al. Qwen2-vl: Enhancing vision-language model's perception of the world at any resolution. arXiv preprint arXiv:2409.12191, 2024. 5" + }, + { + "type": "list", + "bbox": [ + 0.503, + 0.093, + 0.892, + 0.9 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.48, + 0.946, + 0.52, + 0.957 + ], + "angle": 0, + "content": "28542" + } + ], + [ + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.092, + 0.47, + 0.146 + ], + "angle": 0, + "content": "[57] Bin Xia, Shi Yin Wang, Yingfan Tao, Yitong Wang, and Jiaya Jia. Llmga: Multimodal large language model based generation assistant. In European Conference on Computer Vision, 2025. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.081, + 0.149, + 0.47, + 0.205 + ], + "angle": 0, + "content": "[58] Shitao Xiao, Yueze Wang, Junjie Zhou, Huaying Yuan, Xingrun Xing, Ruiran Yan, Shuting Wang, Tiejun Huang, and Zheng Liu. Omnigen: Unified image generation. arXiv preprint arXiv:2409.11340, 2024. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.081, + 0.207, + 0.469, + 0.247 + ], + "angle": 0, + "content": "[59] Shaoan Xie, Zhifei Zhang, Zhe Lin, Tobias Hinz, and Kun Zhang. Smartbrush: Text and shape guided object inpainting with diffusion model. In CVPR, 2023. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.081, + 0.249, + 0.469, + 0.303 + ], + "angle": 0, + "content": "[60] Hu Ye, Jun Zhang, Sibo Liu, Xiao Han, and Wei Yang. Ip-adapter: Text compatible image prompt adapter for text-to-image diffusion models. arXiv preprint arXiv:2308.06721, 2023. 2, 3, 4, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.081, + 0.305, + 0.468, + 0.347 + ], + "angle": 0, + "content": "[61] Kai Zhang, Lingbo Mo, Wenhu Chen, Huan Sun, and Yu Su. Magicbrush: A manually annotated dataset for instruction-guided image editing. NeurIPS, 2024. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.081, + 0.349, + 0.468, + 0.389 + ], + "angle": 0, + "content": "[62] Lvmin Zhang, Anyi Rao, and Maneesh Agrawala. Adding conditional control to text-to-image diffusion models. In ICCV, 2023. 2, 3, 5, 6, 7, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.081, + 0.391, + 0.468, + 0.445 + ], + "angle": 0, + "content": "[63] Shihao Zhao, Dongdong Chen, Yen-Chun Chen, Jianmin Bao, Shaozhe Hao, Lu Yuan, and Kwan-Yee K. Wong. Uni-controlnet: All-in-one control to text-to-image diffusion models. NeurIPS, 2023. 3" + }, + { + "type": "list", + "bbox": [ + 0.08, + 0.092, + 0.47, + 0.445 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.48, + 0.946, + 0.518, + 0.956 + ], + "angle": 0, + "content": "28543" + } + ] +] \ No newline at end of file diff --git a/2025/DreamOmni_ Unified Image Generation and Editing/ff84d3e1-2391-4334-90b8-2fe54fd66a09_origin.pdf b/2025/DreamOmni_ Unified Image Generation and Editing/ff84d3e1-2391-4334-90b8-2fe54fd66a09_origin.pdf new file mode 100644 index 0000000000000000000000000000000000000000..a848be4d7191d8283fcf5b467d19c4ecb4fb1376 --- /dev/null +++ b/2025/DreamOmni_ Unified Image Generation and Editing/ff84d3e1-2391-4334-90b8-2fe54fd66a09_origin.pdf @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:2d2eee03be7929ae6b1cebefae0e4e5a4589df6c987f483ee77bcffdde35f49d +size 9572828 diff --git a/2025/DreamOmni_ Unified Image Generation and Editing/full.md b/2025/DreamOmni_ Unified Image Generation and Editing/full.md new file mode 100644 index 0000000000000000000000000000000000000000..a75e2e75f6e3bb371aecae57eca464afa9db1094 --- /dev/null +++ b/2025/DreamOmni_ Unified Image Generation and Editing/full.md @@ -0,0 +1,457 @@ +# DreamOmni: Unified Image Generation and Editing + +Bin Xia1, Yuechen Zhang1, Jingyao Li1, Chengyao Wang1, Yitong Wang2, Xinglong Wu2, Bei Yu1, and Jiaya Jia3 +1 CUHK, 2 ByteDance Inc, 3 HKUST +https://zj-binxia.github.io/DreamOmni-ProjectPage/ + +![](images/bf70f440b4a435429a3f193d5c3eef55613b916e010b16063c62fdd196b50cd5.jpg) +Figure 1. The gallery of DreamOmni. DreamOmni, as a native unified image generation and editing model, can handle various tasks. + +# Abstract + +projectpagepCurrently, the success of large language models (LLMs) illustrates that a unified multitasking approach can significantly enhance model usability, streamline deployment, and foster synergistic benefits across different tasks. However, in computer vision, while text-to-image (T2I) models have significantly improved generation quality through scaling up, their framework design did not initially consider how to unify with downstream tasks, such as various types of editing. To address this, we introduce DreamOmni, a unified model for image generation and editing. We begin by analyzing existing frameworks and the requirements of downstream tasks, proposing a unified framework that integrates both T2I models and various editing tasks. Furthermore, another key challenge is the efficient creation of high-quality editing data, particularly for instruction-based and drag-based editing. To this + +end, we develop a synthetic data pipeline using sticker-like elements to synthesize accurate, high-quality datasets efficiently, which enables editing data scaling up for unified model training. For training, DreamOmni jointly trains T2I generation and downstream tasks. T2I training enhances the model's understanding of specific concepts and improves generation quality, while editing training helps the model grasp the nuances of the editing task. This collaboration significantly boosts editing performance. Extensive experiments confirm the effectiveness of DreamOmni. The code and model will be released. + +# 1. Introductian + +Recently, text-to-image (T2I) generative foundation models [6, 15, 27, 39, 44, 47] have made remarkable progress, driving the development of various downstream applications such as image editing, video generation, and more. + +Currently, T2I foundation models face two significant challenges: (1) Adapting these models for downstream applications often requires the integration of various plugins (such as ControlNet [62] and IP-adapter [60]) in different ways, or the extension of input channels (e.g., SD-inpainting [44], InstructP2P [8]). This reliance on specialized frameworks hampers multi-task generalization and complicates deployment. (2) High-quality and accurate editing data is difficult to obtain, including instruction-based editing, drag editing, and subject-driven generation data. In this paper, we propose to unify T2I models with multiple editing tasks, such as instruction-based editing, inpainting & outpainting, drag editing, and reference image generation within a single framework. Additionally, we introduce an efficient synthetic data pipeline for efficiently and accurately constructing editing data, facilitating the training of native unified generation and editing model. + +We began by exploring fundamental model frameworks to develop a unified framework that is both highly effective and efficient, achieving rapid training convergence. (1) We aligned certain frameworks, such as Unet [44] and DIT [38], to have similar parameter sizes and runtime settings for fair comparison (Fig. 3). Our analysis showed that DIT's effectiveness comes from concentrating most of its computational and parameter load on latent at $2 \times$ downsampling size, whereas Unet distributes more of its operations at smaller scales. This allocation makes DIT comparatively superior, as concentrating computations at the $2 \times$ downsampled latent achieves a better trade-off. However, we observed that Unet's residual connections significantly improve the model's training convergence speed. Inspired by the efficiency of high-resolution latent computations, we also incorporated additional residual convolutional networks for the input-scale latent. (2) We replaced the original text encoder with a vision-language model (VLM) to unify visual-language prompt encoding. This encoded prompt was then concatenated with the noisy latent, allowing for integrated computation within the DIT framework. + +Next, we need to obtain accurate training data efficiently for our unified model. For various types of editing, such as instruction-based editing, it's challenging to create and filter pairs of data that precisely match the instructions while avoiding detail corruption. We have discovered that the key to effective editing lies in helping the model understand the meaning of the editing operations, rather than learning specific concepts (a capability the model already possesses in T2I training). Therefore, we propose a synthetic collage data pipeline that easily scales up to generate accurate and diverse edited data, enabling models to better understand the nuances of specific editing tasks (Fig. 2 (b)). (1) For instruction-based editing, we constructed synthetic data for three main categories: addition, removal, and replacement. (2) For drag editing, we utilized synthetic data to cre + +ate scaling, translating, and rotating editing data. (3) For inpainting and outpainting, we randomly generate masks for each given image. (4) For reference image generation, We divided the tasks into subject-driven generation and image-conditioned generation (i.e., ControlNet-like generation) tasks. For subject-driven generation, we used a canvas filled with stickers, allowing the model to reference a specific sticker for generation, which helps train the model to extract the demanded objects or details from given images for generation. For the image-conditioned generation, we created canny, depth, and segment maps, similar to ControlNet. (5) We use synthetic data to enhance the model's T2I accuracy in responding to attributes such as quantity, position, relationships, color, shape, and text. Concretely, we randomly arrange different quantities of stickers, text, and geometric shapes on a canvas, and obtain precise descriptions of their quantity, position, relationships, and colors based on their exact coordinates. + +During training, thanks to our unified framework that integrates T2I with multiple editing tasks, we can easily train T2I data alongside various types of editing data, obtaining a native unified image generation and editing model, called DreamOmni. The editing tasks benefit from the T2I data, which helps prevent the model from forgetting specific concepts and generation quality decline, while the T2I task can leverage synthetic data to enhance instruction following. + +- We conducted an analysis of existing model frameworks under fair settings and, drawing on the characteristics of different tasks, proposed an efficient and powerful unified image generation and editing framework, DreamOmni. +- We introduced a synthetic collage data pipeline to tackle the current inefficiency and difficulty in creating and filtering high-quality editing data. Furthermore, we utilized the synthetic collage data pipeline to enhance the accuracy of the T2I model's output. Experimental results demonstrate that synthetic data is an effective, high-quality, and cost-efficient method to scale up data for achieving unified image generation and editing training. +- After unified training with T2I data and various synthetic datasets, our DreamOmni showcases competitive performance on T2I generation and various editing tasks. + +# 2. Related Work + +Text-to-image Diffusion Model. Diffusion models [1, 5, 14, 21, 23, 24, 35, 52, 53] have emerged as highly effective methods for image generation, outperforming previous generative models like GANs [20] and VAEs [26]. The latent diffusion model [44], also known as Stable Diffusion, enhances the diffusion process within the image latent space and utilizes cross-attention to guide image generation through text input. This approach enhances the usability of the image generation model, driving its widespread adoption. Subsequently, more advanced text-to + +![](images/79c9ce9b16fe1fcd7f8e050f03f30082678e6dd7c75e3b2c475751e3f4660543.jpg) +(a) The framework of DreamOmni +Figure 2. The overview of DreamOmni. (a) The DreamOmni framework supports unified image generation and editing, with fast training convergence and powerful performance. (b) To overcome the difficulty and inefficiency in data creation and filtering for image editing, we propose a collage-based synthetic data pipeline. This pipeline enables the efficient creation of data for various editing tasks, such as adding, deleting, and replacement operations in instruction-based editing, as well as translation, scaling, and rotation in drag editing. Additionally, it supports reference image generation and segmentation & detection. Furthermore, our synthetic data pipeline enhances the accuracy of T2I generation. Due to space limitations, we have optionally shown the corresponding prompts or instructions for these cases. + +![](images/c8f3b7938d763df3eae9a1281a2554c79ef16137b19a0612ebc783d7f4712413.jpg) +(b) The examples of synthetic data + +image (T2I) models were developed, such as Imagen [47], DALLE [43], and others [3, 6, 13, 16, 39]. Recently, the DIT frameworks [4, 38] has become increasingly popular as more models [10, 15, 31] transition from the Unet framework [45], thanks to its straightforward structure and outstanding compatibility with multimodal inputs. However, the current T2I model fails to account for the design of downstream editing tasks, necessitating numerous specialized adaptations. For instance, Instructp2p [8] requires an expansion of input channels, while ControlNet [62] and IP-adapter [60] rely on additional plugins and various methods for injecting reference image information. These fragmented designs are inconsistent with the trend toward unified models, complicating deployment and impeding the joint training of T2I and editing tasks. In this paper, we will analyze the existing T2I frameworks and propose a more efficient and powerful unified generation and editing model. + +Multimodal-guided Image Editing and Generation offers richer guiding information and has broader applications [2, 7, 9, 12, 37, 54, 55] than T2I models. It encompasses instruction-based editing [8, 18, 25, 49, 61], inpainting & outpainting [34, 57, 59], drag editing [32, 51], and reference image generation [28, 29, 40, 62, 63]. A significant challenge in editing, especially in instruction-based and drag editing, lies in creating and filtering accurately paired editing data. For example, InstructP2P [8] utilizes finetuned GPT-3 and a Prompt-to-Prompt strategy [22] to + +generate instruction-based editing datasets, but it achieves a success rate of under $15\%$ with frequent artifacts and detail loss. Additionally, filtering high-quality editing data is a persistent challenge, complicating scaling up. Magicbrush [61] addresses this by employing people to create and filter editing data, but this manual method limits data volume. Furthermore, for drag editing, most methods [32, 51] are training-free, which restricts their performance and efficiency. To this end, InstaDrag [50] leverages video data to construct drag editing data. However, InstaDrag still faces challenges with inefficiency in filtering valid data from video. Moreover, for subject-driven generation [28, 36], DreamBooth [46] requires training on several subject-related images before each inference, which is inconvenient and inefficient. Some works [30, 58] collect multiple photos of the same person from the internet to build datasets for training more efficient and powerful subject-driven generation models. However, the process of gathering and filtering such data still presents a challenge, limiting the scaling up of the dataset. In this paper, we propose a simple, efficient, and accurate synthetic collage data pipeline to create quite a few data for unified training. + +# 3. Methodology + +The unification of multiple tasks is a trend and pursuit in the field of computer vision and AI, which not only en + +hances model usability and reduces deployment complexity but also enables collaborative training that fosters synergies between tasks. However, current T2I foundation models are primarily designed specifically for T2I and often overlook the potential for integration with other tasks, such as various image editing tasks. To this end, we propose DreamOmni, a unified model for image generation and editing. We design and train DreamOmni from three aspects. (1) In Sec. 3.1, we compare various frameworks in a fair setting and design a powerful and fast training converging framework that supports unified multi-tasking based on the characteristics of different tasks. (2) In Sec. 3.2, we aim to propose a convenient, efficient, and accurate synthetic data pipeline for data scaling up to facilitate multi-task unified training and enhance the model's instruction-following ability. (3) In Sec. 3.3, we introduce the training scheme for DreamOmni. Native unified training of T2I and various editing tasks prevents concept forgetting and generation quality decline, while enhancing the model's editing and prompt-following capabilities. + +# 3.1. Framework + +In this part, we aim to design a unified and powerful image generation and editing framework. Currently, different editing models often have distinct structure designs. For example, IP-adapter [60] and BLIP-Diffusion [28] inject information through cross-attention to maintain subjects. In contrast, InstructP2P [8] achieves editing consistency by adding different numbers of input channels for models. These structures are tailored for specific tasks and lack generalizability. To this end, as shown in Fig. 2 (a), we concatenate VLM features with noisy latent and input them into a DIT Block for joint multi-head self-attention operations. After that, the VLM features and noisy latent are processed by the FeedForward modules. This allows the model to autonomously learn any level features (from overall consistency to subject consistency) for editing and generation. Notably, for FeedForward modules, we separate the VLM features and noisy latent, passing them through two distinct FeedForward modules with the same network structure. Additionally, instead of using CLIP [41] or T5 [42] as the text encoder, we introduce a Vision-Language Model (VLM) that enables joint understanding and encoding of both image and text prompts. For tasks that require high consistency, such as instruction-based editing and drag editing, we input the VAE encoding of the source image into the DIT model to ensure strong consistency between the non-edited regions of the output image and the source image. + +In the current framework design, some works, such as DIT [38], are compared under label-conditioned generation rather than T2I. However, T2I is inherently more complex than label-based generation, as it requires the integration and understanding of complex prompts. Additionally, many + +T2I models [15, 44], such as SDXL [39], are trained with different model sizes, datasets, and training settings. This variability makes it challenging to assess the impact of different model components on overall performance. Moreover, SDXL incorporates many Transformer blocks within its UNet structure. So, why does DIT outperform SDXL? To address this, we conducted extensive experiments as shown in Fig. 3. We observe that DIT surpasses Unet because it allocates most of its computation to $2 \times$ downsampled latent, whereas Unet allocates a greater proportion to $4 \times$ downsampled latent. Since attention operations on $1 \times$ latent cause memory burden, we further employ residual Conv blocks to refine generation details for $1 \times$ latent. + +Additionally, we observed that the use of long connections in the UNet framework can significantly accelerate the model's training convergence without compromising performance. As shown in Fig. 2 (a), we concatenate early and later features along the channel dimension and apply a linear layer to combine the two features. Notably, linear layers used for VLM features and noisy latent are different. + +# 3.2. Synthetic Data + +In addition to the unified framework, we also require quite a few data to support joint training. While T2I data is readily available, creating and filtering accurate, high-quality data for tasks like instruction-based editing is far more challenging. To address this, we introduce a synthetic collage data pipeline that efficiently and accurately generates the required editing data. As shown in Fig. 2 (b), our pipeline encompasses six tasks. Notably, this is not the full extent of our synthetic pipeline's capabilities; it is also capable of handling more complex task combinations. + +- T2I generation. As shown in Fig. 2 (b), in addition to the conventional T2I data, we further enhance the model's performance on the T2I by incorporating synthetic data, specifically focusing on improving the text, shape, position, quantity, and color generation. Specifically, for the text, we randomly generate words or short phrases on a blank canvas using a variety of fonts, colors, thicknesses, and sizes. For the shape and quantity, we randomly create geometric shapes with different quantities, colors, and sizes, and arrange them on the canvas. Based on these attributes and their positions, we generate accurate prompts, which are then refined using a LLM. Furthermore, we use a diverse set of stickers and segmentation data for synthesis, placing them on the canvas and calculating their precise spatial relationships. These prompts are then created and further polished by the LLM to produce descriptions that are more natural. + +- Inpainting & Outpainting. We randomly generate masks for smearing, blocks, and image edges. Notably, during training, in addition to feeding the masked image and its corresponding mask into the VLM for encoding, we also + +include the image caption with a $50\%$ probability. + +- Instruction-based editing. We categorize the task into three operations: addition, removal, and replacement. For both removal and replacement, we randomly select a background image and an object image to create the source image. In the removal case, the target image is simply the background image. For replacement, the target image is generated by replacing the object in the same location with a different object. Notably, for the addition, due to the need for the added object to be placed in a contextually appropriate position relative to the background, we use a blank background in the paper. +- Drag editing. We categorize the data into three types: translation, scaling, and rotation. Notably, Instadrag [50] treats each pair of drag points as a separate image, which is sparse and impractical due to the fixed number of drag points required. Therefore, we represent each drag point using the format $(x,y,dx,dy)$ as a prompt input, where $x$ and $y$ denote the coordinates of the drag points in the source image, and $dx$ and $dy$ represent the translation vector. Furthermore, we normalize these coordinates by dividing them by the image's width or height. +- Reference image generation. We categorize the data into two types: image-conditioned generation, akin to ControlNet [62], and subject-driven generation. For image-conditioned generation, we begin by selecting high-quality images and creating corresponding canny maps, depth maps, and segmentation masks as source images for training. For subject-driven generation, we synthesize source images and then randomly select objects within these images to create target images. The model is trained to generate new content based on specific attributes referenced from the source image, allowing for the flexible generation of varied scenes and subjects. +- Segmentation & detection. We randomly selected a background image and an object image. These images are then composited together to create a source image. Following this, we apply color manipulation or draw a bounding box around the object region based on the alpha channel of the object image, obtaining the target image. + +Our synthetic collage data pipeline is efficient and precise, enabling the generation of billions of diverse images for large-scale pretraining and fine-tuning of DreamOmni. + +# 3.3. Model Training + +After careful consideration, we chose a $2.5B$ parameter size for DreamOmni's DIT model. This size strikes a good balance between being user-friendly and ensuring powerful performance. For the VLM encoder, we directly adopt the Qwen2-VL [56] 7B model, based on three reasons: (1) It supports image inputs of arbitrary resolution, (2) It delivers strong model performance, and (3) It is released under a permissive open-source license. The VLM features are derived + +![](images/44c58641be4599ab2285e3cabd82a6d9c636faa9477dc1d5d4fc07044562d205.jpg) +Figure 3. Comparison of different frameworks. The left figure shows the FID comparison among different frameworks, while the right table shows their number of parameters and runtime. + +
FrameworkParam (M)Runtime (ms)
SDXL [39]83828.59
Pixart [10]86234.61
SD3 [15]89625.73
DreamOmni-V182828.32
DreamOmni-V283625.14
DreamOmni-V3 (Ours)82727.04
+ +Table 1. Comparisons on GenEval [19]. Our model outperforms all current open models (including SOTA SD3-Medium [15]). + +
ModelObjectsCountingColorsPositionColor Attribution
OverallSingleTwo
SD 1.5 [44]0.430.970.380.350.760.040.06
PixArt-alpha [10]0.480.980.50.440.80.080.07
SDv2.10.50.980.510.440.850.070.17
SDXL [39]0.550.980.740.390.850.150.23
SD-cascade0.520.980.570.470.870.080.15
SD3-Medium [15]0.700.990.840.630.880.280.55
DreamOmni (Ours)0.700.990.810.650.880.340.54
+ +from the penultimate layer of Qwen2-VL. Additionally, we use the FLUX-schnell's VAE as DreamOmni's VAE, which retains more latent channels, enabling the model to capture finer image details. Furthermore, we optimize DreamOmni using Rectified Flow [33], which performs the forward process by linearly interpolating between noise and data along a straight trajectory. we train DreamOmni using loss $\mathcal{L}$ : + +$$ +\mathcal {L} = \mathbb {E} \left(\left\| (\mathbf {z} - \epsilon) - v _ {\theta} \left(\mathbf {z} _ {t}, \mathbf {c}, t\right) \right\| _ {2} ^ {2}\right), \tag {1} +$$ + +where $\mathbf{z}_t = t\mathbf{z} + (1 - t)\epsilon$ represents the noised feature map at timestep $t$ . Ground truth images are encoded by VAE into latent space to derive $\mathbf{z}$ . Here, $\epsilon \in \mathcal{N}(0,\mathbf{I})$ represents Gaussian noise, and $v_{\theta}$ refers to the DIT model. $\mathbf{c}$ indicates the conditional information. + +The T2I training dataset consists of $125M$ images, including both the LAION dataset $(103M)$ [48] and our collected data $(22M)$ . These images have been annotated using InternVL2 [11]. For synthetic data, such as T2I, instruction editing, inpainting & outpainting, drag editing, and reference image generation, we generated $12M$ images respectively, bringing the total to approximately $60M$ images. Additionally, for segmentation & detection, we created $8M$ images. Prior to training, to improve training speed usage, we first encoded prompts into VLM features. + +Our training process is divided into three stages. In the first stage, we train on images with $256 \times 256$ size, a batch size of 2048, a learning rate of $1 \times 10^{-4}$ , and $377K$ iterations. In the second stage, we train on images with $512 \times 512$ size, a batch size of 1024, a learning rate of $5 \times 10^{-5}$ , and $189K$ iterations. In the final stage, we select 12M high-quality T2I data and randomly sample 1M high-quality images from each type of synthetic data for training. + +![](images/9c60bf850c8a95d2f6eb26717927c3f612a9e003f8a092bb85b7ba8a5b6bd738.jpg) +A heartwarming 3D rendered scene of an elderly farmer and a tiny orange kitten dressed in matching denim overalls. The farmer, with a gentle smile, walks alongside the kitten in a lush, green garden filled with thriving plants, showcasing a fruitful harvest. The intricate details of the overalls and the farmer's worn, weathered face tell a story of years spent tending to the land. + +![](images/a36710f3359dedcc844ff90df08ef9716fb7c4f2dbeab84cdf8898a79eef3233.jpg) +Deep in the enchanted forest lives a woman who is the moon fairy. Her long blonde hair shines in the starlight, tangled with her flowers that glow with a soft blue glow. Her eyes are the color of the night and shine with the magic of the night. The fairy wears a dress made of moon petals, woven with threads of moonlight that shine with an iridescent glow, a crown of stars adorns her head, shining with the light of the full moon that illuminates the forest. Her wings are translucent like glass, with a pale glow reminiscent of the glow of the moon. HD, 16K, photo, cinematic, poster + +SD1.5 +![](images/eabff75628630a192c7f6174f4723c1a442c0339e4cd675ded4fada74f41ca56.jpg) +A unique, intricately detailed creature resembling a reptile, possibly a lizard or a gecko. It has a vibrant blue and green scaled body, with large, round, and expressive eyes that are a deep shade of blue. The creature is adorned with a floral crown made of various flowers, including pink roses and blue blossoms, as well as pinecones. The backdrop is a soft, blurred forest setting, suggesting a serene and mystical ambiance. + +![](images/d98b52f9a32a072e3fe5d58538399420e9ee0919efb756ae921862b99d4ed7de.jpg) + +![](images/19c02e74acd35645c68fb7aefd3a7876c7c76afe4ccbbcd12d4280232a0ae5f4.jpg) + +![](images/c593cd696b0d44bb9efa3739f3c8506d12b7e4b16e25b1c29222a505527892fb.jpg) +SD-Cascade + +![](images/86862a86d59f2f5fecf5efbecb8420c04196a9cae40f48d221840437e745dc73.jpg) + +![](images/630c98528f6cef16ce49b32151d93dee68848f9d2b70a97f2a8c3f2c3c1cd14b.jpg) + +![](images/64de70a1c52d4e34de7c27a9ee6e2c76bf8b2d75e6bb502d3fb4be18c0d23f14.jpg) +SDXL + +![](images/28ebbdf421691bb0c275448e82a3881af5f7991bf577b9a896562c4fc993a8c7.jpg) + +![](images/b3704c9f83d07fdaa18dfbaf01952ad70ba88df830b969a5053dd7583cb86dbc.jpg) + +![](images/a5905d27e5255b2c57cc24d52e74b989416583e20f2408ea5b75ad98290d5ba5.jpg) +SD3-Medium + +![](images/5241c5395538de762335a1ceb954535adad35a385efa25ae6a3e343218157c4b.jpg) + +![](images/d50cfe5b1598ca98a282138c143de6f5ed3c709ca3bcc97e484d4f0a6078bc38.jpg) + +![](images/b146b33a9948c008b9eede92c808b61ea54b4987bd4cadc1358f34c9021d493e.jpg) +DreamOmni (Ours) + +![](images/fc372607486dac8f2444b979e802e1229e914be714e81550d6d0d2a681023a11.jpg) +Prompt + +![](images/df26f9c61d877b45f3074187783722826d34dc331107cd69c99e4ed30e708f71.jpg) +Masked Image + +![](images/44a3c553e8aa42dbaec3f56e1ebbce95dac706a3a457cd10e93bb049c016cf0a.jpg) +Figure 4. Visual comparison on T2I generation. Compared to other competitive methods (including SD3-Medium [15], SDXL [39], SD-Cascade, and SD1.5 [44]), our DreamOmni not only better adheres to user prompts but also generate more visually appealing results with delicate details, elegant composition, and so on. + +![](images/a6348fcb43f06e64f55e50418c9c155719d8c67a855950446df76add04a405bd.jpg) +ControlNet Inpainting + +![](images/63f07d40f1c7c588f59765f17eaf0603bc21f25f56016545d4397e43575423a8.jpg) + +![](images/1d484fd1dacff2233c71294d957c63143aeb78a78a36d3a50c8d8e5327b4e6cd.jpg) +SD Inpainting + +![](images/7944ebbe070f47a98f8fe25d76bd2c8a624b665b21c92239a06b17ae05d4e025.jpg) + +![](images/55c1a929f02a64bfdb5f5b747b74085b99d7a886ca8f41d8d9b0c5f2be9837a4.jpg) +DreamOmni + +![](images/c3f8a16228c0d87658cdc71237cdf9aa8639df763950c1df5712eccbc56fda42.jpg) + +![](images/75e8040ab91ac2c4c85a4201768648c99b9057cc04444c34dc6ffde6fea71e47.jpg) +Masked Image +Figure 5. Visual comparison on inpainting & outpainting between DreamOmni, ControlNet-Inpainting [62] and SD-inpainting [44]. + +![](images/8bb81f02630bd43ec637e10d158244f669982a9a47c77ed4809f936eeca84a9d.jpg) + +![](images/51ec26871e80adffb8ce709f78b0a8c022c985b30f9dc915f34a17d734d956b9.jpg) +ControlNet Inpainting + +![](images/a5e8e081ae3e2ce2b20a07485c536e0988ed2d0286267adf01c053bc378f722d.jpg) + +![](images/00311f0d0da94a5007156fcc982ee548bc98a4cb4daf938907b3c1add935292a.jpg) +SD Inpainting + +![](images/df38ee661153399d24cdbbade590ea3dec06c0e5b769793eda28fa91fd424280.jpg) + +![](images/3faede984f7722c407634dbfae9c4391608161546e5dfc9968ba4f3db535fdb0.jpg) +DreamOmni + +We train models with $1024 \times 1024$ size, a batch size of 256, a learning rate of $2 \times 10^{-5}$ , and $140K$ iterations. All experiments are conducted on 64 A100 GPUs. Additionally, to enable the model to generate images at varying resolutions, similar to the approach used in SDXL [39], we divide the images into 31 buckets based on their aspect ratios, ranging from $4:1$ to $1:4$ , during training. + +# 4. Experiments + +Evaluation on Framework. We compared several T2I model frameworks under similar settings to identify effective components. Using the same VAE, CLIP text encoder, parameters, runtime, and LAION training/testing datasets, we evaluated the Unet-based SDXL [39], DIT- + +based Pixart [10], SD3-Medium [15], and our developed DreamOmni variations. Notably, as shown in Fig. 3, to facilitate comparison, we do not use the full $2.5B$ parameters of DreamOmni, but instead adjust the parameters of all models to $0.85B$ . DreamOmni-V1 has two downsampling layers ( $2\times$ and $4\times$ ) like SDXL but lacks an Unet connection. DreamOmni-V2 builds on DreamOmni-V1 by adding the Unet connection, while DreamOmni-V3 takes DreamOmni-V2 further by focusing all DIT operations on $2\times$ downsampled latent. (1) Models with the Unet connection (SDXL, DreamOmni-V2, DreamOmni-V3) showed significantly faster convergence than those without (SD3-Medium, DreamOmni-V1). DreamOmni-V3, notably, converges four times faster than the SD3-Medium, enhancing + +![](images/828f4d95ba889ae4b9d04ba9e6f79da3c1518fc6e78677b058d2a28f2d84c97c.jpg) +A humanoid cat from the race american shorthair with black fur. Standing as a barbarian tabaxi with a greatsword. Dressed in light clothing, something like the ancient geeks weared, photorealistic. standing in front of an oasis in the dessert. + +![](images/26dcdb6d7cfc5e275ac15e558032eb3569c54ae11a605d06bebdeb82c5f1188a.jpg) +elf girl travelling alongside a majestRyuic dragon, close-up view, anime style, high fantasy, digital art, masterpiece painting +Image +Condition + +![](images/58f6e08643e10e515a0b16dcc3836f09d102f5ba42f2f583e250ecdee69c9ec6.jpg) + +![](images/d7cc60c44731c166bae4251f1a3dc113b280698fdca951d03280dade47264e1e.jpg) +ControlNet + +![](images/f1125c81febc4efb17f152ce3de1408a5b9369fbad67999ae27170671a1ef680.jpg) +A cute teenage-child fairy with blue hair and butterflies in her ears, holding balloons and a birthday cake with a candle lit on it in her other hand. Her hand is holding the plate. The artwork is in the style pixar anime. her facial expression is very happy and excited but in a cute way + +![](images/9dd97af1fa68a90233a7857ba19ae9d301934d958c03038bc8630d1c1eae324c.jpg) +DreamOmni +(Ours) +Prompt + +![](images/6a28139b11cfc28452b984717b712e5a436905c3c254b81eed52622643c97945.jpg) + +![](images/5f236f35dd1ff78053a54c74cd5172dffa72abcf20bbbd0e4750a8608ac84027.jpg) +ironman is riding a white little pony +Image +Condition + +![](images/b86641f75a31a3ea4edb1d74175b54aa1e84d4c66c25335a837126047a6d3752.jpg) + +![](images/9fff3d28baf0402e201d513cdec8bc8d0439c2f79d63ae299bf85a10326b51d2.jpg) +ControlNet + +![](images/8c3e9d93f8ffec9dc5f980640ac62d4d8280c232679ade8c4ad77b61f0fb0f81.jpg) + +![](images/6316ebfc0ed010e05c69008c1d5a9e033a7937675262fcba8f771ebdf77b37b8.jpg) +DreamOmni +(Ours) + +![](images/ce23874b6bc3822c78a52cf1f92f4c5442e5394d7e0835485f2e51433bf03367.jpg) +Prompt +A girl (in the given image) is playing a red electric guitar. She is wearing a black jacket and a banie, and she has headphones. The background features autumn leaves, suggesting a seasonal setting. +Generate an image using the character from the supplied image as reference. A girl is holding a lantern. She is surrounded by vibrant flowers and a whimsical, magical atmosphere. The girl is wearing a blue and white outfit with a ruffled collar, and she has a thoughtful expression. + +![](images/71d6401569b886c3669726ff3af16b5a21a45d87a364d700d75c1175a4fad111.jpg) +Prompt +Image condition +Figure 7. Visual comparison on subject-driven generation. Compared with BLIP-Diffusion [28] and IP-adapter [60], DreamOmni excels at accurately following user prompts while preserving the specified subject. Furthermore, its success across both photo and anime cases underscores DreamOmni's generalization capabilities. + +![](images/16ae7a1dd5f5fb65f7045c28e903c928bc95e9a424f69eaad6309229bd461333.jpg) +Figure 6. Visual comparison on image-conditioned generation. Our DreamOmni excels at using canny, depth, and segmentation maps as conditions for image generation. Compared to the classic ControlNet [62], DreamOmni not only follows the user's prompts and image conditions more accurately but also generates content and color schemes that are visually more pleasing. + +![](images/29f2aa46a034fd5b8519eeadea1b1b6647695637649b4b09bc429ef85c335a59.jpg) +BLIP-Diffusion + +![](images/5a999d537b19673d8251107ad71754ae76bffbc83a62e3bfa3ace6c256273ef4.jpg) + +![](images/55b9589255c538909e1ce964bcf9e95e60c83551b93b2619aed607cf1bbbd2cf.jpg) +IP-adapter + +![](images/17fbb8252ee01bae475fa6c3dda431ef06b63f3009f8de0afdc9fabf40238dca.jpg) + +![](images/1c302bb80fc9aa2d8507c0b7e2497803e94efc2126323b5e19cd89aea8723c3b.jpg) +DreamOmni +(Ours) + +![](images/2e3935f859f63a69b1b29076674734db78a7a57aff7772ecdf3bf5e8d1666adf.jpg) +A man (in the given image) is wearing a white t-shirt and white pants. He is smiling. + +![](images/e33c2d93ffff5a82d96ea400c5aeae17ddd9374a2323de34025f89940de84673.jpg) +A close-up of a woman (in the given image) with her head tilted slightly to the side. +Prompt +Image Condition + +![](images/e7d1a78cf261fcb7e43495e79540c1b3fb005e50ec5df9693ac8a76d247346b9.jpg) + +![](images/45833f4e0aeb40caf13eab52f03c1d41ccecffd7f1a7477913bd21031894b971.jpg) +BLIP-Diffusion + +![](images/266f72eaee2b4599407b8fd9ee640e181ca837262dc687317c3aad7da75b7a05.jpg) + +![](images/08e9a6fce616c8cfcf902f6e6fe893cc7e1261092c035db857ace89280b2edb8.jpg) +IP-adapter + +![](images/86df605e36e088a9b243442809bbb836755f08b9c5295b24865ca53516c645a5.jpg) + +![](images/d29fd01cf1cc4418b0cce699238d9dfb532335ea6f8cfbad690beeba0e7be888.jpg) +DreamOmni +(Ours) + +both training and fine-tuning efficiency. (2) Comparing DreamOmni-V3 and DreamOmni-V2, we found that concentrating DIT block computations on higher-resolution latent $(2\times)$ is more cost-effective. + +Evaluation on T2I generation. As shown in Tab. 1, our synthetic data significantly enhances DreamOmni's T2I generation capabilities in aspects such as quantity, color, and position, enabling our model to achieve SOTA results on GenEval [19]. Notably, SD3-Medium [15] is a $2B$ open-source SOTA T2I model, which has a similar parameters as our DreamOmni. Furthermore, qualitative results are shown in Fig. 4. We can see that DreamOmni's outputs are not only more visually pleasing but also align more accurately with the given prompts. + +Evaluation on Inpainting. We compare our DreamOmni with ControlNet-Inpainting [62] and SD-Inpainting [44] on our high-quality evaluation datasets to evaluate its performance. The quantitative results, presented in Tab. 2, demonstrate that DreamOmni significantly outperforms both ControlNet-Inpainting and SD-Inpainting, highlighting its superior generation quality and coherence. Visual results (Fig. 5) further emphasize that DreamOmni ex + +cels in generating fine details, surpassing both ControlNet-Inpainting and SD-Inpainting. Additionally, DreamOmni effectively handles large mask holes, generating realistic content instead of blurry and incoherent outputs. + +Evaluation on Reference Image Generation. For image-conditioned generation, we conducted a comparison with ControlNet [62] on canny, depth, and segmentation image conditions. Visual results are presented in Fig. 6. These results reveal that, across all tested conditions, our DreamOmni significantly outperforms ControlNet. Our method not only adheres more faithfully to both the image condition and the prompt but also demonstrates enhanced visual quality with better composition and richer detail. + +For subject-driven image generation, we compared our method with competitive approaches, including BLIP-Diffusion [28] and IP-Adapter [60]. To demonstrate DreamOmni's powerful generalization, we validated their performance on both anime and photographic images. Visual results are shown in Fig. 7. Compared to other methods, our DreamOmni not only preserves the specified subject effectively but also follows the prompt well. + +Evaluation on Instruction Editing. We compare + +![](images/26ac1993b18c5d5fde0b92a4e9abcdedef56978bce85191dbf35d6605508dcec.jpg) +Remove the star ornament from the snow-covered tree. + +![](images/54c14c2dc65bf48785e3855b8126a9d12e462c98130142bc0feba4b036062dfa.jpg) + +![](images/6ce6d807399980032fd04191a4813c2c4e47a85f36f070cd15743a33360f89f5.jpg) + +![](images/b2055dde6260cd53640cad3ed66a63efab8d44c2af4ce21c54d7e4f4747e5545.jpg) + +![](images/f574abba04f76f8f6d4d4d2e623b9ab8c62cb7f14eaedee413ffe33bdebc875f.jpg) +Add a straw hat on the person's head. +Instruction +Input Image + +![](images/80aaffea2850bbcc03f2b6acc282035b9ad1aee9cae49ffc1ae4ec74be7dbfbb.jpg) +InstructP2P + +![](images/96fca43af04d42f865e53bed58ec7814453234d70bfff082d2f8e0fc820a9ded.jpg) +MGIE + +![](images/dabcef71389c39f00a3f4ce172b9f164d67b1bf7d5f91f8ba3f6be55465727d4.jpg) +DreamOmni + +![](images/d465152c882e8b47a9990ec8678da46c0e9df0bd956a71cd60b4381b5cd1a816.jpg) +Replace the owl with a hawk + +![](images/b5e033c73ce20add6b6c4dcd2a1e6ed22923116fa2f039751487eecee7182d65.jpg) +Remove the green leafy plant near the yellow flower. +Instruction +Input Image + +![](images/a1975dc3ff5eefccd8213013d3e20438d3c6877a160e00f0a7b351bf6e2abd8f.jpg) + +![](images/707c2c486db7c59b0c69672c49d4a71abc1b0ecd4d5c893846e20ef80e927956.jpg) +InstructP2P + +![](images/e0578c815ace48acd27602d24a95d2e3188024e4a4031131a2a9d622096fde7c.jpg) + +![](images/65fb4c5bfdd36d542acf03c452d98ca0975af18d059f9cf80fa87d58af0b58be.jpg) +MGIE + +![](images/390b6b6a83e771d22d678eb0f70598a295e599a3d1a29b5e0f8a04c3ea1f2934.jpg) + +![](images/7422a2c7d86afda30228ef2d1353bc5d923439e58f7998aa2469725f842b4284.jpg) +DreamOmni + +Table 2. Quantitative comparison for inpainting & outpainting. + +
ModelInpaintingOutpainting
FID↓LPIPS↓FID↓LPIPS↓
SD-inpainting [44]1.35220.15602.91790.2475
ControlNet-inpainting [62]1.83930.15944.23370.2521
DreamOmni (Ours)0.83710.12031.69260.1995
+ +DreamOmni with competitive methods such as MGIE [17] and InstructP2P [8]. The visual results are presented in Fig. 8. We can see that DreamOmni performs more accurate edits, including addition, removal, and replacement. Specifically, our editing results exhibit superior consistency with minimal changes to non-edited areas and higher-quality generation of the edited content. This further validates that our synthetic data pipeline is an efficient and effective method for creating instruction-based editing datasets, enabling models to learn precise instruction-based editing. Moreover, the efficiency of our synthetic data pipeline allows for models to easily scale up diverse training data. + +Evaluation on Drag Editing. We evaluate DreamOmni on our synthetic evaluation dataset. The visual results are shown in Fig. 9. (1) Compared to the target image, we can see that DreamOmni can accurately perform translation, rotation, and scaling drag edits. (2) for translation and scaling, DreamOmni can maintain the integrity of the dragged object. However, large-scale rotation operations are more challenging for DreamOmni, as they involve complex transformations of the object itself, which may lead to deformations of the edited object. (3) The results demonstrate the effectiveness of our synthetic data pipeline for drag editing, and encoding the drag point positions and displacement information as instruction inputs allows the model to learn precise drag edits (shown in Fig. 2). + +# 5. Conclusion + +Current T2I foundation models lack a unified framework and training for downstream tasks, such as image editing. To address this, we introduce DreamOmni, a unified model for T2I generation and editing. We evaluate the frameworks + +![](images/41a1ec74bd953ea13832c6550d8426b71c73ee5cd26514ff84cd2862f3753e5d.jpg) +Figure 8. Visual comparison on instruction-based editing. Our DreamOmni achieves more precise editing (including addition, removal, and replacement) compared to competitive methods such as MGIE [17] and InstructP2P [8]. + +![](images/8e8d804f5243b71054d8655b1877e6915f5509e60e4e67efa14fce1c78aa32b9.jpg) + +![](images/90ab57500dc7cda0239a1c665654686bb9a9580257c2426083497beb015c1e8a.jpg) +Input Image & Drag Points +Figure 9. Visual results of drag editing: DreamOmni accurately performs translation, rotation, and scaling edits. + +![](images/590e66926b4764db811df766b8c8c1994e8dfc0af06f1942ce54ae5705e3bb37.jpg) + +![](images/21f7d314731317b03a2de6b89db9fbc21a33fc259b328262bf75da13053fe3fa.jpg) + +![](images/b8f4b69ef02823a76d1457bddcd9fa9cfae7920ed9369078c48777bf8d3a4861.jpg) +Target Image + +![](images/788919fb008bd5ac4aa475ecb20d22edeb03f5ac6f180a53bdb647f3db096eba.jpg) + +![](images/f9933cfd1d5171916f164c753451183f97ca130d1ad5fa1c5eab1465cc77ea67.jpg) + +![](images/cc1058d78ae81e8c3a819a8878400c262743be1a8b43c95238b75c87350cb4a0.jpg) +DreamOmni + +of existing models under fair settings and consider the specific needs of various editing tasks. From this analysis, we develop a framework that integrates T2I with various editing tasks. Besides, a challenge in training editing models is the creation of high-quality, large-scale editing data, which is inefficient. To overcome this, we designed a synthetic collage data pipeline capable of efficiently generating quite a few precise, high-quality editing data. Moreover, the pipeline enhances the model's generation accuracy in text, position, quantity, color, and geometry. By jointly training on T2I and multi-task synthetic data, we develop a native, unified model for both image generation and editing. T2I training strengthens the model's grasp of specific concepts and improves generation quality, while editing training enables it to handle the requirements of editing tasks. + +Acknowledgements. This work was supported in part by the Research Grants Council under the Areas of Excellence scheme grant AoE/E-601/22-R. + +# References + +[1] Jacob Austin, Daniel D Johnson, Jonathan Ho, Daniel Tarlow, and Rianne van den Berg. Structured denoising diffusion models in discrete state-spaces. NeurIPS, 2021. 2 +[2] Omri Avrahami, Dani Lischinski, and Ohad Fried. Blended diffusion for text-driven editing of natural images. In CVPR, 2022. 3 +[3] Yogesh Balaji, Seungjun Nah, Xun Huang, Arash Vahdat, Jiaming Song, Qinsheng Zhang, Karsten Kreis, Miika Aittala, Timo Aila, Samuli Laine, et al. ediff-i: Text-to-image diffusion models with an ensemble of expert denoisers. arXiv preprint arXiv:2211.01324, 2022. 3 +[4] Fan Bao, Shen Nie, Kaiwen Xue, Yue Cao, Chongxuan Li, Hang Su, and Jun Zhu. All are worth words: A vit backbone for diffusion models. In CVPR, 2023. 3 +[5] Georgios Batzolis, Jan Stanczuk, Carola-Bibiane Schonlieb, and Christian Etmann. Conditional image generation with score-based diffusion models. arXiv preprint arXiv:2111.13606, 2021. 2 +[6] James Betker, Gabriel Goh, Li Jing, Tim Brooks, Jianfeng Wang, Linjie Li, Long Ouyang, Juntang Zhuang, Joyce Lee, Yufei Guo, et al. Improving image generation with better captions. Computer Science. https://cdn.openai.com/papers/dall-e-3.pdf, 2023. 1, 3 +[7] Manuel Brack, Felix Friedrich, Katharina Kornmeier, Linoy Tsaban, Patrick Schramowski, Kristian Kersting, and Apolinário Passos. *Ledits++: Limitless image editing using text-to-image models.* arXiv preprint arXiv:2311.16711, 2023. 3 +[8] Tim Brooks, Aleksander Holynski, and Alexei A Efros. Instructpix2pix: Learning to follow image editing instructions. In CVPR, 2023. 2, 3, 4, 8 +[9] Mingdeng Cao, Xintao Wang, Zhongang Qi, Ying Shan, Xiaohu Qie, and Yinqiang Zheng. Masactrl: Tuning-free mutual self-attention control for consistent image synthesis and editing. In ICCV, 2023. 3 +[10] Junsong Chen, Jincheng Yu, Chongjian Ge, Lewei Yao, Enze Xie, Yue Wu, Zhongdao Wang, James Kwok, Ping Luo, Huchuan Lu, et al. Pixart-α: Fast training of diffusion transformer for photorealistic text-to-image synthesis. ICLR, 2024. 3, 5, 6 +[11] Zhe Chen, Weiyun Wang, Hao Tian, Shenglong Ye, Zhang-wei Gao, Erfei Cui, Wenwen Tong, Kongzhi Hu, Jiapeng Luo, Zheng Ma, et al. How far are we to gpt-4v? closing the gap to commercial multimodal models with open-source suites. arXiv preprint arXiv:2404.16821, 2024. 5 +[12] Guillaume Couairon, Jakob Verbeek, Holger Schwenk, and Matthieu Cord. Diffedit: Diffusion-based semantic image editing with mask guidance. arXiv preprint arXiv:2210.11427, 2022.3 +[13] Xiaoliang Dai, Ji Hou, Chih-Yao Ma, Sam Tsai, Jialiang Wang, Rui Wang, Peizhao Zhang, Simon Vandenhende, Xi + +aofang Wang, Abhimanyu Dubey, et al. Emu: Enhancing image generation models using photogenic needles in a haystack. arXiv preprint arXiv:2309.15807, 2023. 3 +[14] Prafulla Dhariwal and Alexander Nichol. Diffusion models beat gans on image synthesis. NeurIPS, 2021. 2 +[15] Patrick Esser, Sumith Kulal, Andreas Blattmann, Rahim Entezari, Jonas Müller, Harry Saini, Yam Levi, Dominik Lorenz, Axel Sauer, Frederic Boessel, et al. Scaling rectified flow transformers for high-resolution image synthesis. In ICML, 2024. 1, 3, 4, 5, 6, 7 +[16] Zhida Feng, Zhenyu Zhang, Xintong Yu, Yewei Fang, Lanxin Li, Xuyi Chen, Yuxiang Lu, Jiaxiang Liu, Weichong Yin, Shikun Feng, et al. Ernie-vilg 2.0: Improving text-to-image diffusion model with knowledge-enhanced mixture-of-denoising-experts. In CVPR, 2023. 3 +[17] Tsu-Jui Fu, Wenze Hu, Xianzhi Du, William Yang Wang, Yinfei Yang, and Zhe Gan. Guiding instruction-based image editing via multimodal large language models. *ICLR*, 2024. 8 +[18] Zigang Geng, Binxin Yang, Tiankai Hang, Chen Li, Shuyang Gu, Ting Zhang, Jianmin Bao, Zheng Zhang, Han Hu, Dong Chen, et al. Instructdiffusion: A generalist modeling interface for vision tasks. arXiv preprint arXiv:2309.03895, 2023. 3 +[19] Dhruba Ghosh, Hannaneh Hajishirzi, and Ludwig Schmidt. Geneval: An object-focused framework for evaluating text-to-image alignment. NeurIPS, 36, 2024. 5, 7 +[20] Ian Goodfellow, Jean Pouget-Abadie, Mehdi Mirza, Bing Xu, David Warde-Farley, Sherjil Ozair, Aaron Courville, and Yoshua Bengio. Generative adversarial nets. NeurIPS, 2014. 2 +[21] Shuyang Gu, Dong Chen, Jianmin Bao, Fang Wen, Bo Zhang, Dongdong Chen, Lu Yuan, and Baining Guo. Vector quantized diffusion model for text-to-image synthesis. In CVPR, 2022. 2 +[22] Amir Hertz, Ron Mokady, Jay Tenenbaum, Kfir Aberman, Yael Pritch, and Daniel Cohen-Or. Prompt-to-prompt image editing with cross attention control. arXiv preprint arXiv:2208.01626, 2022. 3 +[23] Jonathan Ho, Ajay Jain, and Pieter Abbeel. Denoising diffusion probabilistic models. NeurIPS, 2020. 2 +[24] Jonathan Ho, Chitwan Sahara, William Chan, David J Fleet, Mohammad Norouzi, and Tim Salimans. Cascaded diffusion models for high fidelity image generation. JMLR, 2022. 2 +[25] Yuzhou Huang, Liangbin Xie, Xintao Wang, Ziyang Yuan, Xiaodong Cun, Yixiao Ge, Jiantao Zhou, Chao Dong, Rui Huang, Ruimao Zhang, et al. Smartedit: Exploring complex instruction-based image editing with multimodal large language models. CVPR, 2024. 3 +[26] Diederik P Kingma. Auto-encoding variational bayes. arXiv preprint arXiv:1312.6114, 2013. 2 +[27] Daiqing Li, Aleks Kamko, Ehsan Akhgari, Ali Sabet, Linmiao Xu, and Suhail Doshi. Playground v2. 5: Three insights towards enhancing aesthetic quality in text-to-image generation. arXiv preprint arXiv:2402.17245, 2024. 1 +[28] Dongxu Li, Junnan Li, and Steven Hoi. Blip-diffusion: Pretrained subject representation for controllable text-to-image generation and editing. *NeurIPS*, 2024. 3, 4, 7 + +[29] Ming Li, Taojiannan Yang, Huafeng Kuang, Jie Wu, Zhaoning Wang, Xuefeng Xiao, and Chen Chen. Controlnet ++: Improving conditional controls with efficient consistency feedback. In ECCV, 2025.3 +[30] Zhen Li, Mingdeng Cao, Xintao Wang, Zhongang Qi, MingMing Cheng, and Ying Shan. Photomaker: Customizing realistic human photos via stacked id embedding. In CVPR, 2024. 3 +[31] Zhimin Li, Jianwei Zhang, Qin Lin, Jiangfeng Xiong, Yanxin Long, Xinchi Deng, Yingfang Zhang, Xingchao Liu, Minbin Huang, Zedong Xiao, et al. Hunyuan-dit: A powerful multi-resolution diffusion transformer with fine-grained chinese understanding. arXiv preprint arXiv:2405.08748, 2024. 3 +[32] Pengyang Ling, Lin Chen, Pan Zhang, Huaian Chen, Yi Jin, and Jinjin Zheng. Freedrag: Feature dragging for reliable point-based image editing. In CVPR, 2024. 3 +[33] Xingchao Liu, Chengyue Gong, and Qiang Liu. Flow straight and fast: Learning to generate and transfer data with rectified flow. arXiv preprint arXiv:2209.03003, 2022. 5 +[34] Hayk Manukyan, Andranik Sargsyan, Barsegh Atanyan, Zhangyang Wang, Shant Navasardyan, and Humphrey Shi. Hd-painter: high-resolution and prompt-faithful text-guided image inpainting with diffusion models. arXiv preprint arXiv:2312.14091, 2023. 3 +[35] Alex Nichol, Prafulla Dhariwal, Aditya Ramesh, Pranav Shyam, Pamela Mishkin, Bob McGrew, Ilya Sutskever, and Mark Chen. Glide: Towards photorealistic image generation and editing with text-guided diffusion models. arXiv preprint arXiv:2112.10741, 2021. 2 +[36] Xichen Pan, Li Dong, Shaohan Huang, Zhiliang Peng, Wenhu Chen, and Furu Wei. Kosmos-g: Generating images in context with multimodal large language models. arXiv preprint arXiv:2310.02992, 2023. 3 +[37] Gaurav Parmar, Krishna Kumar Singh, Richard Zhang, Yijun Li, Jingwan Lu, and Jun-Yan Zhu. Zero-shot image-to-image translation. In ACM SIGGRAPH, 2023. 3 +[38] William Peebles and Saining Xie. Scalable diffusion models with transformers. In ICCV, 2023. 2, 3, 4 +[39] Dustin Podell, Zion English, Kyle Lacey, Andreas Blattmann, Tim Dockhorn, Jonas Müller, Joe Penna, and Robin Rombach. Sdxl: Improving latent diffusion models for high-resolution image synthesis. NeurIPS, 2023. 1, 3, 4, 5, 6 +[40] Can Qin, Shu Zhang, Ning Yu, Yihao Feng, Xinyi Yang, Yingbo Zhou, Huan Wang, Juan Carlos Niebles, Caiming Xiong, Silvio Savarese, et al. Unicontrol: A unified diffusion model for controllable visual generation in the wild. arXiv preprint arXiv:2305.11147, 2023. 3 +[41] Alec Radford, Jong Wook Kim, Chris Hallacy, Aditya Ramesh, Gabriel Goh, Sandhini Agarwal, Girish Sastry, Amanda Askell, Pamela Mishkin, Jack Clark, et al. Learning transferable visual models from natural language supervision. In ICML, 2021. 4 +[42] Colin Raffel, Noam Shazeer, Adam Roberts, Katherine Lee, Sharan Narang, Michael Matena, Yanqi Zhou, Wei Li, and Peter J Liu. Exploring the limits of transfer learning with a + +unified text-to-text transformer. Journal of machine learning research, 2020. 4 +[43] Aditya Ramesh, Mikhail Pavlov, Gabriel Goh, Scott Gray, Chelsea Voss, Alec Radford, Mark Chen, and Ilya Sutskever. Zero-shot text-to-image generation. In ICML, 2021. 3 +[44] Robin Rombach, Andreas Blattmann, Dominik Lorenz, Patrick Esser, and Björn Ommer. High-resolution image synthesis with latent diffusion models. In CVPR, 2022. 1, 2, 4, 5, 6, 7, 8 +[45] Olaf Ronneberger, Philipp Fischer, and Thomas Brox. U-net: Convolutional networks for biomedical image segmentation. In MICCAI, 2015. 3 +[46] Nataniel Ruiz, Yuanzhen Li, Varun Jampani, Yael Pritch, Michael Rubinstein, and Kfir Aberman. Dreambooth: Fine tuning text-to-image diffusion models for subject-driven generation. In CVPR, 2023. 3 +[47] Chitwan Sahara, William Chan, Saurabh Saxena, Lala Li, Jay Whang, Emily L Denton, Kamyar Ghasemipour, Raphael Gontijo Lopes, Burcu Karagol Ayan, Tim Salimans, et al. Photorealistic text-to-image diffusion models with deep language understanding. NeurIPS, 2022. 1, 3 +[48] Christoph Schuhmann, Romain Beaumont, Richard Vencu, Cade Gordon, Ross Wightman, Mehdi Cherti, Theo Coombes, Aarush Katta, Clayton Mullis, Mitchell Wortsman, et al. Laion-5b: An open large-scale dataset for training next generation image-text models. NeurIPS, 2022. 5 +[49] Shelly Sheynin, Adam Polyak, Uriel Singer, Yuval Kirstain, Amit Zohar, Oron Ashual, Devi Parikh, and Yaniv Taigman. Emu edit: Precise image editing via recognition and generation tasks. arXiv preprint arXiv:2311.10089, 2023. 3 +[50] Yujun Shi, Jun Hao Liew, Hanshu Yan, Vincent YF Tan, and Jiashi Feng. Instadrag: Lightning fast and accurate drag-based image editing emerging from videos. arXiv preprint arXiv:2405.13722, 2024. 3, 5 +[51] Yujun Shi, Chuhui Xue, Jun Hao Liew, Jiachun Pan, Hanshu Yan, Wenqing Zhang, Vincent YF Tan, and Song Bai. Dragdiffusion: Harnessing diffusion models for interactive point-based image editing. In CVPR, 2024. 3 +[52] Jascha Sohl-Dickstein, Eric Weiss, Niru Maheswaranathan, and Surya Ganguli. Deep unsupervised learning using nonequilibrium thermodynamics. In ICML, 2015. 2 +[53] Yang Song, Jascha Sohl-Dickstein, Diederik P Kingma, Abhishek Kumar, Stefano Ermon, and Ben Poole. Score-based generative modeling through stochastic differential equations. arXiv preprint arXiv:2011.13456, 2020. 2 +[54] Narek Tumanyan, Michal Geyer, Shai Bagon, and Tali Dekel. Plug-and-play diffusion features for text-driven image-to-image translation. In CVPR, 2023. 3 +[55] Bram Wallace, Akash Gokul, and Nikhil Naik. Edict: Exact diffusion inversion via coupled transformations. In CVPR, 2023. 3 +[56] Peng Wang, Shuai Bai, Sinan Tan, Shijie Wang, Zhihao Fan, Jinze Bai, Keqin Chen, Xuejing Liu, Jialin Wang, Wenbin Ge, et al. Qwen2-vl: Enhancing vision-language model's perception of the world at any resolution. arXiv preprint arXiv:2409.12191, 2024. 5 + +[57] Bin Xia, Shi Yin Wang, Yingfan Tao, Yitong Wang, and Jiaya Jia. Llmga: Multimodal large language model based generation assistant. In European Conference on Computer Vision, 2025. 3 +[58] Shitao Xiao, Yueze Wang, Junjie Zhou, Huaying Yuan, Xingrun Xing, Ruiran Yan, Shuting Wang, Tiejun Huang, and Zheng Liu. Omnigen: Unified image generation. arXiv preprint arXiv:2409.11340, 2024. 3 +[59] Shaoan Xie, Zhifei Zhang, Zhe Lin, Tobias Hinz, and Kun Zhang. Smartbrush: Text and shape guided object inpainting with diffusion model. In CVPR, 2023. 3 +[60] Hu Ye, Jun Zhang, Sibo Liu, Xiao Han, and Wei Yang. Ip-adapter: Text compatible image prompt adapter for text-to-image diffusion models. arXiv preprint arXiv:2308.06721, 2023. 2, 3, 4, 7 +[61] Kai Zhang, Lingbo Mo, Wenhu Chen, Huan Sun, and Yu Su. Magicbrush: A manually annotated dataset for instruction-guided image editing. NeurIPS, 2024. 3 +[62] Lvmin Zhang, Anyi Rao, and Maneesh Agrawala. Adding conditional control to text-to-image diffusion models. In ICCV, 2023. 2, 3, 5, 6, 7, 8 +[63] Shihao Zhao, Dongdong Chen, Yen-Chun Chen, Jianmin Bao, Shaozhe Hao, Lu Yuan, and Kwan-Yee K. Wong. Uni-controlnet: All-in-one control to text-to-image diffusion models. NeurIPS, 2023. 3 \ No newline at end of file diff --git a/2025/DreamOmni_ Unified Image Generation and Editing/images.zip b/2025/DreamOmni_ Unified Image Generation and Editing/images.zip new file mode 100644 index 0000000000000000000000000000000000000000..32d386ffb98243203792ba21a633cfde3bbfdcb6 --- /dev/null +++ b/2025/DreamOmni_ Unified Image Generation and Editing/images.zip @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:afd84a04b38e46317215730f158c648e8ed655f031bda8bb16633b9ebe9f5e9e +size 1048644 diff --git a/2025/DreamOmni_ Unified Image Generation and Editing/layout.json b/2025/DreamOmni_ Unified Image Generation and Editing/layout.json new file mode 100644 index 0000000000000000000000000000000000000000..6e5b1c071c59e6ed74b85dc7f7561acd20d9d7df --- /dev/null +++ b/2025/DreamOmni_ Unified Image Generation and Editing/layout.json @@ -0,0 +1,12484 @@ +{ + "pdf_info": [ + { + "para_blocks": [ + { + "bbox": [ + 134, + 103, + 460, + 121 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 134, + 103, + 460, + 121 + ], + "spans": [ + { + "bbox": [ + 134, + 103, + 460, + 121 + ], + "type": "text", + "content": "DreamOmni: Unified Image Generation and Editing" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 149, + 142, + 444, + 200 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 149, + 142, + 444, + 200 + ], + "spans": [ + { + "bbox": [ + 149, + 142, + 444, + 200 + ], + "type": "text", + "content": "Bin Xia1, Yuechen Zhang1, Jingyao Li1, Chengyao Wang1, Yitong Wang2, Xinglong Wu2, Bei Yu1, and Jiaya Jia3 \n1 CUHK, 2 ByteDance Inc, 3 HKUST \nhttps://zj-binxia.github.io/DreamOmni-ProjectPage/" + } + ] + } + ], + "index": 3 + }, + { + "type": "image", + "bbox": [ + 51, + 206, + 542, + 471 + ], + "blocks": [ + { + "bbox": [ + 51, + 206, + 542, + 471 + ], + "lines": [ + { + "bbox": [ + 51, + 206, + 542, + 471 + ], + "spans": [ + { + "bbox": [ + 51, + 206, + 542, + 471 + ], + "type": "image", + "image_path": "bf70f440b4a435429a3f193d5c3eef55613b916e010b16063c62fdd196b50cd5.jpg" + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 55, + 476, + 536, + 488 + ], + "lines": [ + { + "bbox": [ + 55, + 476, + 536, + 488 + ], + "spans": [ + { + "bbox": [ + 55, + 476, + 536, + 488 + ], + "type": "text", + "content": "Figure 1. The gallery of DreamOmni. DreamOmni, as a native unified image generation and editing model, can handle various tasks." + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "image_caption" + } + ], + "index": 4 + }, + { + "bbox": [ + 143, + 501, + 192, + 513 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 143, + 501, + 192, + 513 + ], + "spans": [ + { + "bbox": [ + 143, + 501, + 192, + 513 + ], + "type": "text", + "content": "Abstract" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 45, + 522, + 289, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 45, + 522, + 289, + 714 + ], + "spans": [ + { + "bbox": [ + 45, + 522, + 289, + 714 + ], + "type": "text", + "content": "projectpagepCurrently, the success of large language models (LLMs) illustrates that a unified multitasking approach can significantly enhance model usability, streamline deployment, and foster synergistic benefits across different tasks. However, in computer vision, while text-to-image (T2I) models have significantly improved generation quality through scaling up, their framework design did not initially consider how to unify with downstream tasks, such as various types of editing. To address this, we introduce DreamOmni, a unified model for image generation and editing. We begin by analyzing existing frameworks and the requirements of downstream tasks, proposing a unified framework that integrates both T2I models and various editing tasks. Furthermore, another key challenge is the efficient creation of high-quality editing data, particularly for instruction-based and drag-based editing. To this" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 304, + 502, + 547, + 633 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 502, + 547, + 633 + ], + "spans": [ + { + "bbox": [ + 304, + 502, + 547, + 633 + ], + "type": "text", + "content": "end, we develop a synthetic data pipeline using sticker-like elements to synthesize accurate, high-quality datasets efficiently, which enables editing data scaling up for unified model training. For training, DreamOmni jointly trains T2I generation and downstream tasks. T2I training enhances the model's understanding of specific concepts and improves generation quality, while editing training helps the model grasp the nuances of the editing task. This collaboration significantly boosts editing performance. Extensive experiments confirm the effectiveness of DreamOmni. The code and model will be released." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 306, + 643, + 387, + 656 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 643, + 387, + 656 + ], + "spans": [ + { + "bbox": [ + 306, + 643, + 387, + 656 + ], + "type": "text", + "content": "1. Introductian" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 304, + 665, + 547, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 665, + 547, + 715 + ], + "spans": [ + { + "bbox": [ + 304, + 665, + 547, + 715 + ], + "type": "text", + "content": "Recently, text-to-image (T2I) generative foundation models [6, 15, 27, 39, 44, 47] have made remarkable progress, driving the development of various downstream applications such as image editing, video generation, and more." + } + ] + } + ], + "index": 10 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "spans": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "type": "text", + "content": "CVF" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 145, + 0, + 494, + 37 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 145, + 0, + 494, + 37 + ], + "spans": [ + { + "bbox": [ + 145, + 0, + 494, + 37 + ], + "type": "text", + "content": "This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "type": "text", + "content": "28533" + } + ] + } + ], + "index": 11 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 0 + }, + { + "para_blocks": [ + { + "bbox": [ + 46, + 72, + 289, + 277 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 72, + 289, + 277 + ], + "spans": [ + { + "bbox": [ + 46, + 72, + 289, + 277 + ], + "type": "text", + "content": "Currently, T2I foundation models face two significant challenges: (1) Adapting these models for downstream applications often requires the integration of various plugins (such as ControlNet [62] and IP-adapter [60]) in different ways, or the extension of input channels (e.g., SD-inpainting [44], InstructP2P [8]). This reliance on specialized frameworks hampers multi-task generalization and complicates deployment. (2) High-quality and accurate editing data is difficult to obtain, including instruction-based editing, drag editing, and subject-driven generation data. In this paper, we propose to unify T2I models with multiple editing tasks, such as instruction-based editing, inpainting & outpainting, drag editing, and reference image generation within a single framework. Additionally, we introduce an efficient synthetic data pipeline for efficiently and accurately constructing editing data, facilitating the training of native unified generation and editing model." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 46, + 279, + 289, + 531 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 279, + 289, + 531 + ], + "spans": [ + { + "bbox": [ + 46, + 279, + 289, + 531 + ], + "type": "text", + "content": "We began by exploring fundamental model frameworks to develop a unified framework that is both highly effective and efficient, achieving rapid training convergence. (1) We aligned certain frameworks, such as Unet [44] and DIT [38], to have similar parameter sizes and runtime settings for fair comparison (Fig. 3). Our analysis showed that DIT's effectiveness comes from concentrating most of its computational and parameter load on latent at " + }, + { + "bbox": [ + 46, + 279, + 289, + 531 + ], + "type": "inline_equation", + "content": "2 \\times" + }, + { + "bbox": [ + 46, + 279, + 289, + 531 + ], + "type": "text", + "content": " downsampling size, whereas Unet distributes more of its operations at smaller scales. This allocation makes DIT comparatively superior, as concentrating computations at the " + }, + { + "bbox": [ + 46, + 279, + 289, + 531 + ], + "type": "inline_equation", + "content": "2 \\times" + }, + { + "bbox": [ + 46, + 279, + 289, + 531 + ], + "type": "text", + "content": " downsampled latent achieves a better trade-off. However, we observed that Unet's residual connections significantly improve the model's training convergence speed. Inspired by the efficiency of high-resolution latent computations, we also incorporated additional residual convolutional networks for the input-scale latent. (2) We replaced the original text encoder with a vision-language model (VLM) to unify visual-language prompt encoding. This encoded prompt was then concatenated with the noisy latent, allowing for integrated computation within the DIT framework." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 46, + 534, + 289, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 534, + 289, + 715 + ], + "spans": [ + { + "bbox": [ + 46, + 534, + 289, + 715 + ], + "type": "text", + "content": "Next, we need to obtain accurate training data efficiently for our unified model. For various types of editing, such as instruction-based editing, it's challenging to create and filter pairs of data that precisely match the instructions while avoiding detail corruption. We have discovered that the key to effective editing lies in helping the model understand the meaning of the editing operations, rather than learning specific concepts (a capability the model already possesses in T2I training). Therefore, we propose a synthetic collage data pipeline that easily scales up to generate accurate and diverse edited data, enabling models to better understand the nuances of specific editing tasks (Fig. 2 (b)). (1) For instruction-based editing, we constructed synthetic data for three main categories: addition, removal, and replacement. (2) For drag editing, we utilized synthetic data to cre" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 304, + 72, + 547, + 287 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 72, + 547, + 287 + ], + "spans": [ + { + "bbox": [ + 304, + 72, + 547, + 287 + ], + "type": "text", + "content": "ate scaling, translating, and rotating editing data. (3) For inpainting and outpainting, we randomly generate masks for each given image. (4) For reference image generation, We divided the tasks into subject-driven generation and image-conditioned generation (i.e., ControlNet-like generation) tasks. For subject-driven generation, we used a canvas filled with stickers, allowing the model to reference a specific sticker for generation, which helps train the model to extract the demanded objects or details from given images for generation. For the image-conditioned generation, we created canny, depth, and segment maps, similar to ControlNet. (5) We use synthetic data to enhance the model's T2I accuracy in responding to attributes such as quantity, position, relationships, color, shape, and text. Concretely, we randomly arrange different quantities of stickers, text, and geometric shapes on a canvas, and obtain precise descriptions of their quantity, position, relationships, and colors based on their exact coordinates." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 304, + 288, + 547, + 384 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 288, + 547, + 384 + ], + "spans": [ + { + "bbox": [ + 304, + 288, + 547, + 384 + ], + "type": "text", + "content": "During training, thanks to our unified framework that integrates T2I with multiple editing tasks, we can easily train T2I data alongside various types of editing data, obtaining a native unified image generation and editing model, called DreamOmni. The editing tasks benefit from the T2I data, which helps prevent the model from forgetting specific concepts and generation quality decline, while the T2I task can leverage synthetic data to enhance instruction following." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 306, + 385, + 545, + 564 + ], + "type": "list", + "angle": 0, + "index": 8, + "blocks": [ + { + "bbox": [ + 306, + 385, + 545, + 433 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 385, + 545, + 433 + ], + "spans": [ + { + "bbox": [ + 306, + 385, + 545, + 433 + ], + "type": "text", + "content": "- We conducted an analysis of existing model frameworks under fair settings and, drawing on the characteristics of different tasks, proposed an efficient and powerful unified image generation and editing framework, DreamOmni." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 306, + 434, + 545, + 529 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 434, + 545, + 529 + ], + "spans": [ + { + "bbox": [ + 306, + 434, + 545, + 529 + ], + "type": "text", + "content": "- We introduced a synthetic collage data pipeline to tackle the current inefficiency and difficulty in creating and filtering high-quality editing data. Furthermore, we utilized the synthetic collage data pipeline to enhance the accuracy of the T2I model's output. Experimental results demonstrate that synthetic data is an effective, high-quality, and cost-efficient method to scale up data for achieving unified image generation and editing training." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 306, + 529, + 545, + 564 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 529, + 545, + 564 + ], + "spans": [ + { + "bbox": [ + 306, + 529, + 545, + 564 + ], + "type": "text", + "content": "- After unified training with T2I data and various synthetic datasets, our DreamOmni showcases competitive performance on T2I generation and various editing tasks." + } + ] + } + ], + "index": 7 + } + ], + "sub_type": "text" + }, + { + "bbox": [ + 306, + 574, + 392, + 586 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 574, + 392, + 586 + ], + "spans": [ + { + "bbox": [ + 306, + 574, + 392, + 586 + ], + "type": "text", + "content": "2. Related Work" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 304, + 594, + 547, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 594, + 547, + 715 + ], + "spans": [ + { + "bbox": [ + 304, + 594, + 547, + 715 + ], + "type": "text", + "content": "Text-to-image Diffusion Model. Diffusion models [1, 5, 14, 21, 23, 24, 35, 52, 53] have emerged as highly effective methods for image generation, outperforming previous generative models like GANs [20] and VAEs [26]. The latent diffusion model [44], also known as Stable Diffusion, enhances the diffusion process within the image latent space and utilizes cross-attention to guide image generation through text input. This approach enhances the usability of the image generation model, driving its widespread adoption. Subsequently, more advanced text-to" + } + ] + } + ], + "index": 10 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "type": "text", + "content": "28534" + } + ] + } + ], + "index": 11 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 1 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 53, + 71, + 244, + 300 + ], + "blocks": [ + { + "bbox": [ + 53, + 71, + 244, + 300 + ], + "lines": [ + { + "bbox": [ + 53, + 71, + 244, + 300 + ], + "spans": [ + { + "bbox": [ + 53, + 71, + 244, + 300 + ], + "type": "image", + "image_path": "79c9ce9b16fe1fcd7f8e050f03f30082678e6dd7c75e3b2c475751e3f4660543.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 96, + 302, + 217, + 310 + ], + "lines": [ + { + "bbox": [ + 96, + 302, + 217, + 310 + ], + "spans": [ + { + "bbox": [ + 96, + 302, + 217, + 310 + ], + "type": "text", + "content": "(a) The framework of DreamOmni" + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 46, + 313, + 545, + 380 + ], + "lines": [ + { + "bbox": [ + 46, + 313, + 545, + 380 + ], + "spans": [ + { + "bbox": [ + 46, + 313, + 545, + 380 + ], + "type": "text", + "content": "Figure 2. The overview of DreamOmni. (a) The DreamOmni framework supports unified image generation and editing, with fast training convergence and powerful performance. (b) To overcome the difficulty and inefficiency in data creation and filtering for image editing, we propose a collage-based synthetic data pipeline. This pipeline enables the efficient creation of data for various editing tasks, such as adding, deleting, and replacement operations in instruction-based editing, as well as translation, scaling, and rotation in drag editing. Additionally, it supports reference image generation and segmentation & detection. Furthermore, our synthetic data pipeline enhances the accuracy of T2I generation. Due to space limitations, we have optionally shown the corresponding prompts or instructions for these cases." + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "type": "image", + "bbox": [ + 249, + 72, + 543, + 296 + ], + "blocks": [ + { + "bbox": [ + 249, + 72, + 543, + 296 + ], + "lines": [ + { + "bbox": [ + 249, + 72, + 543, + 296 + ], + "spans": [ + { + "bbox": [ + 249, + 72, + 543, + 296 + ], + "type": "image", + "image_path": "c8f3b7938d763df3eae9a1281a2554c79ef16137b19a0612ebc783d7f4712413.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 340, + 300, + 460, + 310 + ], + "lines": [ + { + "bbox": [ + 340, + 300, + 460, + 310 + ], + "spans": [ + { + "bbox": [ + 340, + 300, + 460, + 310 + ], + "type": "text", + "content": "(b) The examples of synthetic data" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_caption" + } + ], + "index": 2 + }, + { + "bbox": [ + 46, + 389, + 287, + 592 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 389, + 287, + 592 + ], + "spans": [ + { + "bbox": [ + 46, + 389, + 287, + 592 + ], + "type": "text", + "content": "image (T2I) models were developed, such as Imagen [47], DALLE [43], and others [3, 6, 13, 16, 39]. Recently, the DIT frameworks [4, 38] has become increasingly popular as more models [10, 15, 31] transition from the Unet framework [45], thanks to its straightforward structure and outstanding compatibility with multimodal inputs. However, the current T2I model fails to account for the design of downstream editing tasks, necessitating numerous specialized adaptations. For instance, Instructp2p [8] requires an expansion of input channels, while ControlNet [62] and IP-adapter [60] rely on additional plugins and various methods for injecting reference image information. These fragmented designs are inconsistent with the trend toward unified models, complicating deployment and impeding the joint training of T2I and editing tasks. In this paper, we will analyze the existing T2I frameworks and propose a more efficient and powerful unified generation and editing model." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 46, + 594, + 288, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 594, + 288, + 714 + ], + "spans": [ + { + "bbox": [ + 46, + 594, + 288, + 714 + ], + "type": "text", + "content": "Multimodal-guided Image Editing and Generation offers richer guiding information and has broader applications [2, 7, 9, 12, 37, 54, 55] than T2I models. It encompasses instruction-based editing [8, 18, 25, 49, 61], inpainting & outpainting [34, 57, 59], drag editing [32, 51], and reference image generation [28, 29, 40, 62, 63]. A significant challenge in editing, especially in instruction-based and drag editing, lies in creating and filtering accurately paired editing data. For example, InstructP2P [8] utilizes finetuned GPT-3 and a Prompt-to-Prompt strategy [22] to" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 304, + 389, + 545, + 653 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 389, + 545, + 653 + ], + "spans": [ + { + "bbox": [ + 304, + 389, + 545, + 653 + ], + "type": "text", + "content": "generate instruction-based editing datasets, but it achieves a success rate of under " + }, + { + "bbox": [ + 304, + 389, + 545, + 653 + ], + "type": "inline_equation", + "content": "15\\%" + }, + { + "bbox": [ + 304, + 389, + 545, + 653 + ], + "type": "text", + "content": " with frequent artifacts and detail loss. Additionally, filtering high-quality editing data is a persistent challenge, complicating scaling up. Magicbrush [61] addresses this by employing people to create and filter editing data, but this manual method limits data volume. Furthermore, for drag editing, most methods [32, 51] are training-free, which restricts their performance and efficiency. To this end, InstaDrag [50] leverages video data to construct drag editing data. However, InstaDrag still faces challenges with inefficiency in filtering valid data from video. Moreover, for subject-driven generation [28, 36], DreamBooth [46] requires training on several subject-related images before each inference, which is inconvenient and inefficient. Some works [30, 58] collect multiple photos of the same person from the internet to build datasets for training more efficient and powerful subject-driven generation models. However, the process of gathering and filtering such data still presents a challenge, limiting the scaling up of the dataset. In this paper, we propose a simple, efficient, and accurate synthetic collage data pipeline to create quite a few data for unified training." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 306, + 667, + 388, + 681 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 667, + 388, + 681 + ], + "spans": [ + { + "bbox": [ + 306, + 667, + 388, + 681 + ], + "type": "text", + "content": "3. Methodology" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 305, + 689, + 545, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 689, + 545, + 713 + ], + "spans": [ + { + "bbox": [ + 305, + 689, + 545, + 713 + ], + "type": "text", + "content": "The unification of multiple tasks is a trend and pursuit in the field of computer vision and AI, which not only en" + } + ] + } + ], + "index": 9 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "text", + "content": "28535" + } + ] + } + ], + "index": 10 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 2 + }, + { + "para_blocks": [ + { + "bbox": [ + 46, + 72, + 289, + 313 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 72, + 289, + 313 + ], + "spans": [ + { + "bbox": [ + 46, + 72, + 289, + 313 + ], + "type": "text", + "content": "hances model usability and reduces deployment complexity but also enables collaborative training that fosters synergies between tasks. However, current T2I foundation models are primarily designed specifically for T2I and often overlook the potential for integration with other tasks, such as various image editing tasks. To this end, we propose DreamOmni, a unified model for image generation and editing. We design and train DreamOmni from three aspects. (1) In Sec. 3.1, we compare various frameworks in a fair setting and design a powerful and fast training converging framework that supports unified multi-tasking based on the characteristics of different tasks. (2) In Sec. 3.2, we aim to propose a convenient, efficient, and accurate synthetic data pipeline for data scaling up to facilitate multi-task unified training and enhance the model's instruction-following ability. (3) In Sec. 3.3, we introduce the training scheme for DreamOmni. Native unified training of T2I and various editing tasks prevents concept forgetting and generation quality decline, while enhancing the model's editing and prompt-following capabilities." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 47, + 322, + 126, + 333 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 322, + 126, + 333 + ], + "spans": [ + { + "bbox": [ + 47, + 322, + 126, + 333 + ], + "type": "text", + "content": "3.1. Framework" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 46, + 342, + 289, + 653 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 342, + 289, + 653 + ], + "spans": [ + { + "bbox": [ + 46, + 342, + 289, + 653 + ], + "type": "text", + "content": "In this part, we aim to design a unified and powerful image generation and editing framework. Currently, different editing models often have distinct structure designs. For example, IP-adapter [60] and BLIP-Diffusion [28] inject information through cross-attention to maintain subjects. In contrast, InstructP2P [8] achieves editing consistency by adding different numbers of input channels for models. These structures are tailored for specific tasks and lack generalizability. To this end, as shown in Fig. 2 (a), we concatenate VLM features with noisy latent and input them into a DIT Block for joint multi-head self-attention operations. After that, the VLM features and noisy latent are processed by the FeedForward modules. This allows the model to autonomously learn any level features (from overall consistency to subject consistency) for editing and generation. Notably, for FeedForward modules, we separate the VLM features and noisy latent, passing them through two distinct FeedForward modules with the same network structure. Additionally, instead of using CLIP [41] or T5 [42] as the text encoder, we introduce a Vision-Language Model (VLM) that enables joint understanding and encoding of both image and text prompts. For tasks that require high consistency, such as instruction-based editing and drag editing, we input the VAE encoding of the source image into the DIT model to ensure strong consistency between the non-edited regions of the output image and the source image." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 46, + 654, + 289, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 654, + 289, + 715 + ], + "spans": [ + { + "bbox": [ + 46, + 654, + 289, + 715 + ], + "type": "text", + "content": "In the current framework design, some works, such as DIT [38], are compared under label-conditioned generation rather than T2I. However, T2I is inherently more complex than label-based generation, as it requires the integration and understanding of complex prompts. Additionally, many" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 304, + 72, + 547, + 228 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 72, + 547, + 228 + ], + "spans": [ + { + "bbox": [ + 304, + 72, + 547, + 228 + ], + "type": "text", + "content": "T2I models [15, 44], such as SDXL [39], are trained with different model sizes, datasets, and training settings. This variability makes it challenging to assess the impact of different model components on overall performance. Moreover, SDXL incorporates many Transformer blocks within its UNet structure. So, why does DIT outperform SDXL? To address this, we conducted extensive experiments as shown in Fig. 3. We observe that DIT surpasses Unet because it allocates most of its computation to " + }, + { + "bbox": [ + 304, + 72, + 547, + 228 + ], + "type": "inline_equation", + "content": "2 \\times" + }, + { + "bbox": [ + 304, + 72, + 547, + 228 + ], + "type": "text", + "content": " downsampled latent, whereas Unet allocates a greater proportion to " + }, + { + "bbox": [ + 304, + 72, + 547, + 228 + ], + "type": "inline_equation", + "content": "4 \\times" + }, + { + "bbox": [ + 304, + 72, + 547, + 228 + ], + "type": "text", + "content": " downsampled latent. Since attention operations on " + }, + { + "bbox": [ + 304, + 72, + 547, + 228 + ], + "type": "inline_equation", + "content": "1 \\times" + }, + { + "bbox": [ + 304, + 72, + 547, + 228 + ], + "type": "text", + "content": " latent cause memory burden, we further employ residual Conv blocks to refine generation details for " + }, + { + "bbox": [ + 304, + 72, + 547, + 228 + ], + "type": "inline_equation", + "content": "1 \\times" + }, + { + "bbox": [ + 304, + 72, + 547, + 228 + ], + "type": "text", + "content": " latent." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 304, + 228, + 547, + 313 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 228, + 547, + 313 + ], + "spans": [ + { + "bbox": [ + 304, + 228, + 547, + 313 + ], + "type": "text", + "content": "Additionally, we observed that the use of long connections in the UNet framework can significantly accelerate the model's training convergence without compromising performance. As shown in Fig. 2 (a), we concatenate early and later features along the channel dimension and apply a linear layer to combine the two features. Notably, linear layers used for VLM features and noisy latent are different." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 306, + 323, + 398, + 335 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 323, + 398, + 335 + ], + "spans": [ + { + "bbox": [ + 306, + 323, + 398, + 335 + ], + "type": "text", + "content": "3.2. Synthetic Data" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 304, + 342, + 547, + 461 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 342, + 547, + 461 + ], + "spans": [ + { + "bbox": [ + 304, + 342, + 547, + 461 + ], + "type": "text", + "content": "In addition to the unified framework, we also require quite a few data to support joint training. While T2I data is readily available, creating and filtering accurate, high-quality data for tasks like instruction-based editing is far more challenging. To address this, we introduce a synthetic collage data pipeline that efficiently and accurately generates the required editing data. As shown in Fig. 2 (b), our pipeline encompasses six tasks. Notably, this is not the full extent of our synthetic pipeline's capabilities; it is also capable of handling more complex task combinations." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 306, + 462, + 547, + 665 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 462, + 547, + 665 + ], + "spans": [ + { + "bbox": [ + 306, + 462, + 547, + 665 + ], + "type": "text", + "content": "- T2I generation. As shown in Fig. 2 (b), in addition to the conventional T2I data, we further enhance the model's performance on the T2I by incorporating synthetic data, specifically focusing on improving the text, shape, position, quantity, and color generation. Specifically, for the text, we randomly generate words or short phrases on a blank canvas using a variety of fonts, colors, thicknesses, and sizes. For the shape and quantity, we randomly create geometric shapes with different quantities, colors, and sizes, and arrange them on the canvas. Based on these attributes and their positions, we generate accurate prompts, which are then refined using a LLM. Furthermore, we use a diverse set of stickers and segmentation data for synthesis, placing them on the canvas and calculating their precise spatial relationships. These prompts are then created and further polished by the LLM to produce descriptions that are more natural." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 306, + 666, + 547, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 666, + 547, + 713 + ], + "spans": [ + { + "bbox": [ + 306, + 666, + 547, + 713 + ], + "type": "text", + "content": "- Inpainting & Outpainting. We randomly generate masks for smearing, blocks, and image edges. Notably, during training, in addition to feeding the masked image and its corresponding mask into the VLM for encoding, we also" + } + ] + } + ], + "index": 9 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 749, + 318, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 749, + 318, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 749, + 318, + 757 + ], + "type": "text", + "content": "28536" + } + ] + } + ], + "index": 10 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 3 + }, + { + "para_blocks": [ + { + "bbox": [ + 55, + 72, + 258, + 84 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 72, + 258, + 84 + ], + "spans": [ + { + "bbox": [ + 55, + 72, + 258, + 84 + ], + "type": "text", + "content": "include the image caption with a " + }, + { + "bbox": [ + 55, + 72, + 258, + 84 + ], + "type": "inline_equation", + "content": "50\\%" + }, + { + "bbox": [ + 55, + 72, + 258, + 84 + ], + "type": "text", + "content": " probability." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 47, + 85, + 288, + 552 + ], + "type": "list", + "angle": 0, + "index": 5, + "blocks": [ + { + "bbox": [ + 47, + 85, + 287, + 216 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 85, + 287, + 216 + ], + "spans": [ + { + "bbox": [ + 47, + 85, + 287, + 216 + ], + "type": "text", + "content": "- Instruction-based editing. We categorize the task into three operations: addition, removal, and replacement. For both removal and replacement, we randomly select a background image and an object image to create the source image. In the removal case, the target image is simply the background image. For replacement, the target image is generated by replacing the object in the same location with a different object. Notably, for the addition, due to the need for the added object to be placed in a contextually appropriate position relative to the background, we use a blank background in the paper." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 47, + 216, + 288, + 335 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 216, + 288, + 335 + ], + "spans": [ + { + "bbox": [ + 47, + 216, + 288, + 335 + ], + "type": "text", + "content": "- Drag editing. We categorize the data into three types: translation, scaling, and rotation. Notably, Instadrag [50] treats each pair of drag points as a separate image, which is sparse and impractical due to the fixed number of drag points required. Therefore, we represent each drag point using the format " + }, + { + "bbox": [ + 47, + 216, + 288, + 335 + ], + "type": "inline_equation", + "content": "(x,y,dx,dy)" + }, + { + "bbox": [ + 47, + 216, + 288, + 335 + ], + "type": "text", + "content": " as a prompt input, where " + }, + { + "bbox": [ + 47, + 216, + 288, + 335 + ], + "type": "inline_equation", + "content": "x" + }, + { + "bbox": [ + 47, + 216, + 288, + 335 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 47, + 216, + 288, + 335 + ], + "type": "inline_equation", + "content": "y" + }, + { + "bbox": [ + 47, + 216, + 288, + 335 + ], + "type": "text", + "content": " denote the coordinates of the drag points in the source image, and " + }, + { + "bbox": [ + 47, + 216, + 288, + 335 + ], + "type": "inline_equation", + "content": "dx" + }, + { + "bbox": [ + 47, + 216, + 288, + 335 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 47, + 216, + 288, + 335 + ], + "type": "inline_equation", + "content": "dy" + }, + { + "bbox": [ + 47, + 216, + 288, + 335 + ], + "type": "text", + "content": " represent the translation vector. Furthermore, we normalize these coordinates by dividing them by the image's width or height." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 47, + 335, + 287, + 479 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 335, + 287, + 479 + ], + "spans": [ + { + "bbox": [ + 47, + 335, + 287, + 479 + ], + "type": "text", + "content": "- Reference image generation. We categorize the data into two types: image-conditioned generation, akin to ControlNet [62], and subject-driven generation. For image-conditioned generation, we begin by selecting high-quality images and creating corresponding canny maps, depth maps, and segmentation masks as source images for training. For subject-driven generation, we synthesize source images and then randomly select objects within these images to create target images. The model is trained to generate new content based on specific attributes referenced from the source image, allowing for the flexible generation of varied scenes and subjects." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 47, + 479, + 287, + 552 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 479, + 287, + 552 + ], + "spans": [ + { + "bbox": [ + 47, + 479, + 287, + 552 + ], + "type": "text", + "content": "- Segmentation & detection. We randomly selected a background image and an object image. These images are then composited together to create a source image. Following this, we apply color manipulation or draw a bounding box around the object region based on the alpha channel of the object image, obtaining the target image." + } + ] + } + ], + "index": 4 + } + ], + "sub_type": "text" + }, + { + "bbox": [ + 46, + 552, + 287, + 588 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 552, + 287, + 588 + ], + "spans": [ + { + "bbox": [ + 46, + 552, + 287, + 588 + ], + "type": "text", + "content": "Our synthetic collage data pipeline is efficient and precise, enabling the generation of billions of diverse images for large-scale pretraining and fine-tuning of DreamOmni." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 47, + 598, + 144, + 611 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 598, + 144, + 611 + ], + "spans": [ + { + "bbox": [ + 47, + 598, + 144, + 611 + ], + "type": "text", + "content": "3.3. Model Training" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 46, + 618, + 288, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 618, + 288, + 715 + ], + "spans": [ + { + "bbox": [ + 46, + 618, + 288, + 715 + ], + "type": "text", + "content": "After careful consideration, we chose a " + }, + { + "bbox": [ + 46, + 618, + 288, + 715 + ], + "type": "inline_equation", + "content": "2.5B" + }, + { + "bbox": [ + 46, + 618, + 288, + 715 + ], + "type": "text", + "content": " parameter size for DreamOmni's DIT model. This size strikes a good balance between being user-friendly and ensuring powerful performance. For the VLM encoder, we directly adopt the Qwen2-VL [56] 7B model, based on three reasons: (1) It supports image inputs of arbitrary resolution, (2) It delivers strong model performance, and (3) It is released under a permissive open-source license. The VLM features are derived" + } + ] + } + ], + "index": 8 + }, + { + "type": "image", + "bbox": [ + 307, + 71, + 426, + 160 + ], + "blocks": [ + { + "bbox": [ + 307, + 71, + 426, + 160 + ], + "lines": [ + { + "bbox": [ + 307, + 71, + 426, + 160 + ], + "spans": [ + { + "bbox": [ + 307, + 71, + 426, + 160 + ], + "type": "image", + "image_path": "44c58641be4599ab2285e3cabd82a6d9c636faa9477dc1d5d4fc07044562d205.jpg" + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 305, + 163, + 547, + 198 + ], + "lines": [ + { + "bbox": [ + 305, + 163, + 547, + 198 + ], + "spans": [ + { + "bbox": [ + 305, + 163, + 547, + 198 + ], + "type": "text", + "content": "Figure 3. Comparison of different frameworks. The left figure shows the FID comparison among different frameworks, while the right table shows their number of parameters and runtime." + } + ] + } + ], + "index": 11, + "angle": 0, + "type": "image_caption" + } + ], + "index": 9 + }, + { + "type": "table", + "bbox": [ + 435, + 93, + 543, + 141 + ], + "blocks": [ + { + "bbox": [ + 435, + 93, + 543, + 141 + ], + "lines": [ + { + "bbox": [ + 435, + 93, + 543, + 141 + ], + "spans": [ + { + "bbox": [ + 435, + 93, + 543, + 141 + ], + "type": "table", + "html": "
FrameworkParam (M)Runtime (ms)
SDXL [39]83828.59
Pixart [10]86234.61
SD3 [15]89625.73
DreamOmni-V182828.32
DreamOmni-V283625.14
DreamOmni-V3 (Ours)82727.04
", + "image_path": "29467ed51cde2f280c59cddb5041660156f912761b6b9cf218c86bf7ade5ab5b.jpg" + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "table_body" + } + ], + "index": 10 + }, + { + "type": "table", + "bbox": [ + 308, + 228, + 546, + 312 + ], + "blocks": [ + { + "bbox": [ + 306, + 203, + 546, + 226 + ], + "lines": [ + { + "bbox": [ + 306, + 203, + 546, + 226 + ], + "spans": [ + { + "bbox": [ + 306, + 203, + 546, + 226 + ], + "type": "text", + "content": "Table 1. Comparisons on GenEval [19]. Our model outperforms all current open models (including SOTA SD3-Medium [15])." + } + ] + } + ], + "index": 12, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 308, + 228, + 546, + 312 + ], + "lines": [ + { + "bbox": [ + 308, + 228, + 546, + 312 + ], + "spans": [ + { + "bbox": [ + 308, + 228, + 546, + 312 + ], + "type": "table", + "html": "
ModelObjectsCountingColorsPositionColor Attribution
OverallSingleTwo
SD 1.5 [44]0.430.970.380.350.760.040.06
PixArt-alpha [10]0.480.980.50.440.80.080.07
SDv2.10.50.980.510.440.850.070.17
SDXL [39]0.550.980.740.390.850.150.23
SD-cascade0.520.980.570.470.870.080.15
SD3-Medium [15]0.700.990.840.630.880.280.55
DreamOmni (Ours)0.700.990.810.650.880.340.54
", + "image_path": "2d6f645ad4eab1bcdd06cca8d1288b62a8e0360986c5c23f599b2cee74af7274.jpg" + } + ] + } + ], + "index": 13, + "angle": 0, + "type": "table_body" + } + ], + "index": 13 + }, + { + "bbox": [ + 304, + 327, + 546, + 411 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 327, + 546, + 411 + ], + "spans": [ + { + "bbox": [ + 304, + 327, + 546, + 411 + ], + "type": "text", + "content": "from the penultimate layer of Qwen2-VL. Additionally, we use the FLUX-schnell's VAE as DreamOmni's VAE, which retains more latent channels, enabling the model to capture finer image details. Furthermore, we optimize DreamOmni using Rectified Flow [33], which performs the forward process by linearly interpolating between noise and data along a straight trajectory. we train DreamOmni using loss " + }, + { + "bbox": [ + 304, + 327, + 546, + 411 + ], + "type": "inline_equation", + "content": "\\mathcal{L}" + }, + { + "bbox": [ + 304, + 327, + 546, + 411 + ], + "type": "text", + "content": ":" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 351, + 415, + 545, + 435 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 351, + 415, + 545, + 435 + ], + "spans": [ + { + "bbox": [ + 351, + 415, + 545, + 435 + ], + "type": "interline_equation", + "content": "\\mathcal {L} = \\mathbb {E} \\left(\\left\\| (\\mathbf {z} - \\epsilon) - v _ {\\theta} \\left(\\mathbf {z} _ {t}, \\mathbf {c}, t\\right) \\right\\| _ {2} ^ {2}\\right), \\tag {1}", + "image_path": "46123f25139daef53c2bca89ff0055355943cf19532773a7fd0c6ee8e6cfabf4.jpg" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 304, + 437, + 547, + 497 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 437, + 547, + 497 + ], + "spans": [ + { + "bbox": [ + 304, + 437, + 547, + 497 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 304, + 437, + 547, + 497 + ], + "type": "inline_equation", + "content": "\\mathbf{z}_t = t\\mathbf{z} + (1 - t)\\epsilon" + }, + { + "bbox": [ + 304, + 437, + 547, + 497 + ], + "type": "text", + "content": " represents the noised feature map at timestep " + }, + { + "bbox": [ + 304, + 437, + 547, + 497 + ], + "type": "inline_equation", + "content": "t" + }, + { + "bbox": [ + 304, + 437, + 547, + 497 + ], + "type": "text", + "content": ". Ground truth images are encoded by VAE into latent space to derive " + }, + { + "bbox": [ + 304, + 437, + 547, + 497 + ], + "type": "inline_equation", + "content": "\\mathbf{z}" + }, + { + "bbox": [ + 304, + 437, + 547, + 497 + ], + "type": "text", + "content": ". Here, " + }, + { + "bbox": [ + 304, + 437, + 547, + 497 + ], + "type": "inline_equation", + "content": "\\epsilon \\in \\mathcal{N}(0,\\mathbf{I})" + }, + { + "bbox": [ + 304, + 437, + 547, + 497 + ], + "type": "text", + "content": " represents Gaussian noise, and " + }, + { + "bbox": [ + 304, + 437, + 547, + 497 + ], + "type": "inline_equation", + "content": "v_{\\theta}" + }, + { + "bbox": [ + 304, + 437, + 547, + 497 + ], + "type": "text", + "content": " refers to the DIT model. " + }, + { + "bbox": [ + 304, + 437, + 547, + 497 + ], + "type": "inline_equation", + "content": "\\mathbf{c}" + }, + { + "bbox": [ + 304, + 437, + 547, + 497 + ], + "type": "text", + "content": " indicates the conditional information." + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 304, + 498, + 547, + 617 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 498, + 547, + 617 + ], + "spans": [ + { + "bbox": [ + 304, + 498, + 547, + 617 + ], + "type": "text", + "content": "The T2I training dataset consists of " + }, + { + "bbox": [ + 304, + 498, + 547, + 617 + ], + "type": "inline_equation", + "content": "125M" + }, + { + "bbox": [ + 304, + 498, + 547, + 617 + ], + "type": "text", + "content": " images, including both the LAION dataset " + }, + { + "bbox": [ + 304, + 498, + 547, + 617 + ], + "type": "inline_equation", + "content": "(103M)" + }, + { + "bbox": [ + 304, + 498, + 547, + 617 + ], + "type": "text", + "content": " [48] and our collected data " + }, + { + "bbox": [ + 304, + 498, + 547, + 617 + ], + "type": "inline_equation", + "content": "(22M)" + }, + { + "bbox": [ + 304, + 498, + 547, + 617 + ], + "type": "text", + "content": ". These images have been annotated using InternVL2 [11]. For synthetic data, such as T2I, instruction editing, inpainting & outpainting, drag editing, and reference image generation, we generated " + }, + { + "bbox": [ + 304, + 498, + 547, + 617 + ], + "type": "inline_equation", + "content": "12M" + }, + { + "bbox": [ + 304, + 498, + 547, + 617 + ], + "type": "text", + "content": " images respectively, bringing the total to approximately " + }, + { + "bbox": [ + 304, + 498, + 547, + 617 + ], + "type": "inline_equation", + "content": "60M" + }, + { + "bbox": [ + 304, + 498, + 547, + 617 + ], + "type": "text", + "content": " images. Additionally, for segmentation & detection, we created " + }, + { + "bbox": [ + 304, + 498, + 547, + 617 + ], + "type": "inline_equation", + "content": "8M" + }, + { + "bbox": [ + 304, + 498, + 547, + 617 + ], + "type": "text", + "content": " images. Prior to training, to improve training speed usage, we first encoded prompts into VLM features." + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 304, + 618, + 547, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 618, + 547, + 715 + ], + "spans": [ + { + "bbox": [ + 304, + 618, + 547, + 715 + ], + "type": "text", + "content": "Our training process is divided into three stages. In the first stage, we train on images with " + }, + { + "bbox": [ + 304, + 618, + 547, + 715 + ], + "type": "inline_equation", + "content": "256 \\times 256" + }, + { + "bbox": [ + 304, + 618, + 547, + 715 + ], + "type": "text", + "content": " size, a batch size of 2048, a learning rate of " + }, + { + "bbox": [ + 304, + 618, + 547, + 715 + ], + "type": "inline_equation", + "content": "1 \\times 10^{-4}" + }, + { + "bbox": [ + 304, + 618, + 547, + 715 + ], + "type": "text", + "content": ", and " + }, + { + "bbox": [ + 304, + 618, + 547, + 715 + ], + "type": "inline_equation", + "content": "377K" + }, + { + "bbox": [ + 304, + 618, + 547, + 715 + ], + "type": "text", + "content": " iterations. In the second stage, we train on images with " + }, + { + "bbox": [ + 304, + 618, + 547, + 715 + ], + "type": "inline_equation", + "content": "512 \\times 512" + }, + { + "bbox": [ + 304, + 618, + 547, + 715 + ], + "type": "text", + "content": " size, a batch size of 1024, a learning rate of " + }, + { + "bbox": [ + 304, + 618, + 547, + 715 + ], + "type": "inline_equation", + "content": "5 \\times 10^{-5}" + }, + { + "bbox": [ + 304, + 618, + 547, + 715 + ], + "type": "text", + "content": ", and " + }, + { + "bbox": [ + 304, + 618, + 547, + 715 + ], + "type": "inline_equation", + "content": "189K" + }, + { + "bbox": [ + 304, + 618, + 547, + 715 + ], + "type": "text", + "content": " iterations. In the final stage, we select 12M high-quality T2I data and randomly sample 1M high-quality images from each type of synthetic data for training." + } + ] + } + ], + "index": 18 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "type": "text", + "content": "28537" + } + ] + } + ], + "index": 19 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 4 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 153, + 75, + 225, + 147 + ], + "blocks": [ + { + "bbox": [ + 61, + 80, + 147, + 142 + ], + "lines": [ + { + "bbox": [ + 61, + 80, + 147, + 142 + ], + "spans": [ + { + "bbox": [ + 61, + 80, + 147, + 142 + ], + "type": "text", + "content": "A heartwarming 3D rendered scene of an elderly farmer and a tiny orange kitten dressed in matching denim overalls. The farmer, with a gentle smile, walks alongside the kitten in a lush, green garden filled with thriving plants, showcasing a fruitful harvest. The intricate details of the overalls and the farmer's worn, weathered face tell a story of years spent tending to the land." + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_footnote" + }, + { + "bbox": [ + 153, + 75, + 225, + 147 + ], + "lines": [ + { + "bbox": [ + 153, + 75, + 225, + 147 + ], + "spans": [ + { + "bbox": [ + 153, + 75, + 225, + 147 + ], + "type": "image", + "image_path": "9c60bf850c8a95d2f6eb26717927c3f612a9e003f8a092bb85b7ba8a5b6bd738.jpg" + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_body" + } + ], + "index": 4 + }, + { + "type": "image", + "bbox": [ + 153, + 149, + 225, + 222 + ], + "blocks": [ + { + "bbox": [ + 58, + 151, + 150, + 224 + ], + "lines": [ + { + "bbox": [ + 58, + 151, + 150, + 224 + ], + "spans": [ + { + "bbox": [ + 58, + 151, + 150, + 224 + ], + "type": "text", + "content": "Deep in the enchanted forest lives a woman who is the moon fairy. Her long blonde hair shines in the starlight, tangled with her flowers that glow with a soft blue glow. Her eyes are the color of the night and shine with the magic of the night. The fairy wears a dress made of moon petals, woven with threads of moonlight that shine with an iridescent glow, a crown of stars adorns her head, shining with the light of the full moon that illuminates the forest. Her wings are translucent like glass, with a pale glow reminiscent of the glow of the moon. HD, 16K, photo, cinematic, poster" + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_footnote" + }, + { + "bbox": [ + 153, + 149, + 225, + 222 + ], + "lines": [ + { + "bbox": [ + 153, + 149, + 225, + 222 + ], + "spans": [ + { + "bbox": [ + 153, + 149, + 225, + 222 + ], + "type": "image", + "image_path": "a36710f3359dedcc844ff90df08ef9716fb7c4f2dbeab84cdf8898a79eef3233.jpg" + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "image_body" + } + ], + "index": 5 + }, + { + "type": "image", + "bbox": [ + 153, + 225, + 225, + 298 + ], + "blocks": [ + { + "bbox": [ + 60, + 230, + 147, + 297 + ], + "lines": [ + { + "bbox": [ + 60, + 230, + 147, + 297 + ], + "spans": [ + { + "bbox": [ + 60, + 230, + 147, + 297 + ], + "type": "text", + "content": "A unique, intricately detailed creature resembling a reptile, possibly a lizard or a gecko. It has a vibrant blue and green scaled body, with large, round, and expressive eyes that are a deep shade of blue. The creature is adorned with a floral crown made of various flowers, including pink roses and blue blossoms, as well as pinecones. The backdrop is a soft, blurred forest setting, suggesting a serene and mystical ambiance." + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_footnote" + }, + { + "bbox": [ + 153, + 225, + 225, + 298 + ], + "lines": [ + { + "bbox": [ + 153, + 225, + 225, + 298 + ], + "spans": [ + { + "bbox": [ + 153, + 225, + 225, + 298 + ], + "type": "image", + "image_path": "eabff75628630a192c7f6174f4723c1a442c0339e4cd675ded4fada74f41ca56.jpg" + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 175, + 303, + 201, + 313 + ], + "lines": [ + { + "bbox": [ + 175, + 303, + 201, + 313 + ], + "spans": [ + { + "bbox": [ + 175, + 303, + 201, + 313 + ], + "type": "text", + "content": "SD1.5" + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "image_caption" + } + ], + "index": 6 + }, + { + "type": "image", + "bbox": [ + 229, + 75, + 301, + 147 + ], + "blocks": [ + { + "bbox": [ + 229, + 75, + 301, + 147 + ], + "lines": [ + { + "bbox": [ + 229, + 75, + 301, + 147 + ], + "spans": [ + { + "bbox": [ + 229, + 75, + 301, + 147 + ], + "type": "image", + "image_path": "d98b52f9a32a072e3fe5d58538399420e9ee0919efb756ae921862b99d4ed7de.jpg" + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "image_body" + } + ], + "index": 8 + }, + { + "type": "image", + "bbox": [ + 229, + 149, + 301, + 221 + ], + "blocks": [ + { + "bbox": [ + 229, + 149, + 301, + 221 + ], + "lines": [ + { + "bbox": [ + 229, + 149, + 301, + 221 + ], + "spans": [ + { + "bbox": [ + 229, + 149, + 301, + 221 + ], + "type": "image", + "image_path": "19c02e74acd35645c68fb7aefd3a7876c7c76afe4ccbbcd12d4280232a0ae5f4.jpg" + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "image_body" + } + ], + "index": 9 + }, + { + "type": "image", + "bbox": [ + 229, + 225, + 301, + 297 + ], + "blocks": [ + { + "bbox": [ + 229, + 225, + 301, + 297 + ], + "lines": [ + { + "bbox": [ + 229, + 225, + 301, + 297 + ], + "spans": [ + { + "bbox": [ + 229, + 225, + 301, + 297 + ], + "type": "image", + "image_path": "c593cd696b0d44bb9efa3739f3c8506d12b7e4b16e25b1c29222a505527892fb.jpg" + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 244, + 303, + 294, + 313 + ], + "lines": [ + { + "bbox": [ + 244, + 303, + 294, + 313 + ], + "spans": [ + { + "bbox": [ + 244, + 303, + 294, + 313 + ], + "type": "text", + "content": "SD-Cascade" + } + ] + } + ], + "index": 11, + "angle": 0, + "type": "image_caption" + } + ], + "index": 10 + }, + { + "type": "image", + "bbox": [ + 306, + 75, + 378, + 146 + ], + "blocks": [ + { + "bbox": [ + 306, + 75, + 378, + 146 + ], + "lines": [ + { + "bbox": [ + 306, + 75, + 378, + 146 + ], + "spans": [ + { + "bbox": [ + 306, + 75, + 378, + 146 + ], + "type": "image", + "image_path": "86862a86d59f2f5fecf5efbecb8420c04196a9cae40f48d221840437e745dc73.jpg" + } + ] + } + ], + "index": 12, + "angle": 0, + "type": "image_body" + } + ], + "index": 12 + }, + { + "type": "image", + "bbox": [ + 306, + 149, + 377, + 221 + ], + "blocks": [ + { + "bbox": [ + 306, + 149, + 377, + 221 + ], + "lines": [ + { + "bbox": [ + 306, + 149, + 377, + 221 + ], + "spans": [ + { + "bbox": [ + 306, + 149, + 377, + 221 + ], + "type": "image", + "image_path": "630c98528f6cef16ce49b32151d93dee68848f9d2b70a97f2a8c3f2c3c1cd14b.jpg" + } + ] + } + ], + "index": 13, + "angle": 0, + "type": "image_body" + } + ], + "index": 13 + }, + { + "type": "image", + "bbox": [ + 306, + 225, + 378, + 297 + ], + "blocks": [ + { + "bbox": [ + 306, + 225, + 378, + 297 + ], + "lines": [ + { + "bbox": [ + 306, + 225, + 378, + 297 + ], + "spans": [ + { + "bbox": [ + 306, + 225, + 378, + 297 + ], + "type": "image", + "image_path": "64de70a1c52d4e34de7c27a9ee6e2c76bf8b2d75e6bb502d3fb4be18c0d23f14.jpg" + } + ] + } + ], + "index": 14, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 331, + 303, + 358, + 313 + ], + "lines": [ + { + "bbox": [ + 331, + 303, + 358, + 313 + ], + "spans": [ + { + "bbox": [ + 331, + 303, + 358, + 313 + ], + "type": "text", + "content": "SDXL" + } + ] + } + ], + "index": 15, + "angle": 0, + "type": "image_caption" + } + ], + "index": 14 + }, + { + "type": "image", + "bbox": [ + 383, + 75, + 455, + 146 + ], + "blocks": [ + { + "bbox": [ + 383, + 75, + 455, + 146 + ], + "lines": [ + { + "bbox": [ + 383, + 75, + 455, + 146 + ], + "spans": [ + { + "bbox": [ + 383, + 75, + 455, + 146 + ], + "type": "image", + "image_path": "28ebbdf421691bb0c275448e82a3881af5f7991bf577b9a896562c4fc993a8c7.jpg" + } + ] + } + ], + "index": 16, + "angle": 0, + "type": "image_body" + } + ], + "index": 16 + }, + { + "type": "image", + "bbox": [ + 383, + 150, + 454, + 221 + ], + "blocks": [ + { + "bbox": [ + 383, + 150, + 454, + 221 + ], + "lines": [ + { + "bbox": [ + 383, + 150, + 454, + 221 + ], + "spans": [ + { + "bbox": [ + 383, + 150, + 454, + 221 + ], + "type": "image", + "image_path": "b3704c9f83d07fdaa18dfbaf01952ad70ba88df830b969a5053dd7583cb86dbc.jpg" + } + ] + } + ], + "index": 17, + "angle": 0, + "type": "image_body" + } + ], + "index": 17 + }, + { + "type": "image", + "bbox": [ + 383, + 225, + 454, + 297 + ], + "blocks": [ + { + "bbox": [ + 383, + 225, + 454, + 297 + ], + "lines": [ + { + "bbox": [ + 383, + 225, + 454, + 297 + ], + "spans": [ + { + "bbox": [ + 383, + 225, + 454, + 297 + ], + "type": "image", + "image_path": "a5905d27e5255b2c57cc24d52e74b989416583e20f2408ea5b75ad98290d5ba5.jpg" + } + ] + } + ], + "index": 18, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 391, + 303, + 446, + 313 + ], + "lines": [ + { + "bbox": [ + 391, + 303, + 446, + 313 + ], + "spans": [ + { + "bbox": [ + 391, + 303, + 446, + 313 + ], + "type": "text", + "content": "SD3-Medium" + } + ] + } + ], + "index": 19, + "angle": 0, + "type": "image_caption" + } + ], + "index": 18 + }, + { + "type": "image", + "bbox": [ + 459, + 75, + 531, + 146 + ], + "blocks": [ + { + "bbox": [ + 459, + 75, + 531, + 146 + ], + "lines": [ + { + "bbox": [ + 459, + 75, + 531, + 146 + ], + "spans": [ + { + "bbox": [ + 459, + 75, + 531, + 146 + ], + "type": "image", + "image_path": "5241c5395538de762335a1ceb954535adad35a385efa25ae6a3e343218157c4b.jpg" + } + ] + } + ], + "index": 20, + "angle": 0, + "type": "image_body" + } + ], + "index": 20 + }, + { + "type": "image", + "bbox": [ + 459, + 150, + 531, + 221 + ], + "blocks": [ + { + "bbox": [ + 459, + 150, + 531, + 221 + ], + "lines": [ + { + "bbox": [ + 459, + 150, + 531, + 221 + ], + "spans": [ + { + "bbox": [ + 459, + 150, + 531, + 221 + ], + "type": "image", + "image_path": "d50cfe5b1598ca98a282138c143de6f5ed3c709ca3bcc97e484d4f0a6078bc38.jpg" + } + ] + } + ], + "index": 21, + "angle": 0, + "type": "image_body" + } + ], + "index": 21 + }, + { + "type": "image", + "bbox": [ + 459, + 225, + 531, + 297 + ], + "blocks": [ + { + "bbox": [ + 459, + 225, + 531, + 297 + ], + "lines": [ + { + "bbox": [ + 459, + 225, + 531, + 297 + ], + "spans": [ + { + "bbox": [ + 459, + 225, + 531, + 297 + ], + "type": "image", + "image_path": "b146b33a9948c008b9eede92c808b61ea54b4987bd4cadc1358f34c9021d493e.jpg" + } + ] + } + ], + "index": 22, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 458, + 303, + 537, + 314 + ], + "lines": [ + { + "bbox": [ + 458, + 303, + 537, + 314 + ], + "spans": [ + { + "bbox": [ + 458, + 303, + 537, + 314 + ], + "type": "text", + "content": "DreamOmni (Ours)" + } + ] + } + ], + "index": 23, + "angle": 0, + "type": "image_caption" + } + ], + "index": 22 + }, + { + "type": "image", + "bbox": [ + 53, + 368, + 113, + 428 + ], + "blocks": [ + { + "bbox": [ + 86, + 304, + 118, + 315 + ], + "lines": [ + { + "bbox": [ + 86, + 304, + 118, + 315 + ], + "spans": [ + { + "bbox": [ + 86, + 304, + 118, + 315 + ], + "type": "text", + "content": "Prompt" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 53, + 368, + 113, + 428 + ], + "lines": [ + { + "bbox": [ + 53, + 368, + 113, + 428 + ], + "spans": [ + { + "bbox": [ + 53, + 368, + 113, + 428 + ], + "type": "image", + "image_path": "fc372607486dac8f2444b979e802e1229e914be714e81550d6d0d2a681023a11.jpg" + } + ] + } + ], + "index": 25, + "angle": 0, + "type": "image_body" + } + ], + "index": 25 + }, + { + "type": "image", + "bbox": [ + 54, + 430, + 113, + 490 + ], + "blocks": [ + { + "bbox": [ + 54, + 430, + 113, + 490 + ], + "lines": [ + { + "bbox": [ + 54, + 430, + 113, + 490 + ], + "spans": [ + { + "bbox": [ + 54, + 430, + 113, + 490 + ], + "type": "image", + "image_path": "df26f9c61d877b45f3074187783722826d34dc331107cd69c99e4ed30e708f71.jpg" + } + ] + } + ], + "index": 26, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 58, + 496, + 108, + 506 + ], + "lines": [ + { + "bbox": [ + 58, + 496, + 108, + 506 + ], + "spans": [ + { + "bbox": [ + 58, + 496, + 108, + 506 + ], + "type": "text", + "content": "Masked Image" + } + ] + } + ], + "index": 27, + "angle": 0, + "type": "image_caption" + } + ], + "index": 26 + }, + { + "type": "image", + "bbox": [ + 114, + 369, + 173, + 428 + ], + "blocks": [ + { + "bbox": [ + 46, + 323, + 545, + 356 + ], + "lines": [ + { + "bbox": [ + 46, + 323, + 545, + 356 + ], + "spans": [ + { + "bbox": [ + 46, + 323, + 545, + 356 + ], + "type": "text", + "content": "Figure 4. Visual comparison on T2I generation. Compared to other competitive methods (including SD3-Medium [15], SDXL [39], SD-Cascade, and SD1.5 [44]), our DreamOmni not only better adheres to user prompts but also generate more visually appealing results with delicate details, elegant composition, and so on." + } + ] + } + ], + "index": 24, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 114, + 369, + 173, + 428 + ], + "lines": [ + { + "bbox": [ + 114, + 369, + 173, + 428 + ], + "spans": [ + { + "bbox": [ + 114, + 369, + 173, + 428 + ], + "type": "image", + "image_path": "44a3c553e8aa42dbaec3f56e1ebbce95dac706a3a457cd10e93bb049c016cf0a.jpg" + } + ] + } + ], + "index": 28, + "angle": 0, + "type": "image_body" + } + ], + "index": 28 + }, + { + "type": "image", + "bbox": [ + 114, + 430, + 173, + 490 + ], + "blocks": [ + { + "bbox": [ + 114, + 430, + 173, + 490 + ], + "lines": [ + { + "bbox": [ + 114, + 430, + 173, + 490 + ], + "spans": [ + { + "bbox": [ + 114, + 430, + 173, + 490 + ], + "type": "image", + "image_path": "a6348fcb43f06e64f55e50418c9c155719d8c67a855950446df76add04a405bd.jpg" + } + ] + } + ], + "index": 29, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 126, + 492, + 164, + 511 + ], + "lines": [ + { + "bbox": [ + 126, + 492, + 164, + 511 + ], + "spans": [ + { + "bbox": [ + 126, + 492, + 164, + 511 + ], + "type": "text", + "content": "ControlNet Inpainting" + } + ] + } + ], + "index": 30, + "angle": 0, + "type": "image_caption" + } + ], + "index": 29 + }, + { + "type": "image", + "bbox": [ + 175, + 369, + 233, + 428 + ], + "blocks": [ + { + "bbox": [ + 175, + 369, + 233, + 428 + ], + "lines": [ + { + "bbox": [ + 175, + 369, + 233, + 428 + ], + "spans": [ + { + "bbox": [ + 175, + 369, + 233, + 428 + ], + "type": "image", + "image_path": "63f07d40f1c7c588f59765f17eaf0603bc21f25f56016545d4397e43575423a8.jpg" + } + ] + } + ], + "index": 31, + "angle": 0, + "type": "image_body" + } + ], + "index": 31 + }, + { + "type": "image", + "bbox": [ + 175, + 430, + 234, + 490 + ], + "blocks": [ + { + "bbox": [ + 175, + 430, + 234, + 490 + ], + "lines": [ + { + "bbox": [ + 175, + 430, + 234, + 490 + ], + "spans": [ + { + "bbox": [ + 175, + 430, + 234, + 490 + ], + "type": "image", + "image_path": "1d484fd1dacff2233c71294d957c63143aeb78a78a36d3a50c8d8e5327b4e6cd.jpg" + } + ] + } + ], + "index": 32, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 180, + 496, + 228, + 506 + ], + "lines": [ + { + "bbox": [ + 180, + 496, + 228, + 506 + ], + "spans": [ + { + "bbox": [ + 180, + 496, + 228, + 506 + ], + "type": "text", + "content": "SD Inpainting" + } + ] + } + ], + "index": 33, + "angle": 0, + "type": "image_caption" + } + ], + "index": 32 + }, + { + "type": "image", + "bbox": [ + 235, + 369, + 294, + 428 + ], + "blocks": [ + { + "bbox": [ + 235, + 369, + 294, + 428 + ], + "lines": [ + { + "bbox": [ + 235, + 369, + 294, + 428 + ], + "spans": [ + { + "bbox": [ + 235, + 369, + 294, + 428 + ], + "type": "image", + "image_path": "7944ebbe070f47a98f8fe25d76bd2c8a624b665b21c92239a06b17ae05d4e025.jpg" + } + ] + } + ], + "index": 34, + "angle": 0, + "type": "image_body" + } + ], + "index": 34 + }, + { + "type": "image", + "bbox": [ + 236, + 430, + 294, + 490 + ], + "blocks": [ + { + "bbox": [ + 236, + 430, + 294, + 490 + ], + "lines": [ + { + "bbox": [ + 236, + 430, + 294, + 490 + ], + "spans": [ + { + "bbox": [ + 236, + 430, + 294, + 490 + ], + "type": "image", + "image_path": "55c1a929f02a64bfdb5f5b747b74085b99d7a886ca8f41d8d9b0c5f2be9837a4.jpg" + } + ] + } + ], + "index": 35, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 243, + 497, + 285, + 506 + ], + "lines": [ + { + "bbox": [ + 243, + 497, + 285, + 506 + ], + "spans": [ + { + "bbox": [ + 243, + 497, + 285, + 506 + ], + "type": "text", + "content": "DreamOmni" + } + ] + } + ], + "index": 36, + "angle": 0, + "type": "image_caption" + } + ], + "index": 35 + }, + { + "type": "image", + "bbox": [ + 301, + 369, + 359, + 428 + ], + "blocks": [ + { + "bbox": [ + 301, + 369, + 359, + 428 + ], + "lines": [ + { + "bbox": [ + 301, + 369, + 359, + 428 + ], + "spans": [ + { + "bbox": [ + 301, + 369, + 359, + 428 + ], + "type": "image", + "image_path": "c3f8a16228c0d87658cdc71237cdf9aa8639df763950c1df5712eccbc56fda42.jpg" + } + ] + } + ], + "index": 37, + "angle": 0, + "type": "image_body" + } + ], + "index": 37 + }, + { + "type": "image", + "bbox": [ + 301, + 430, + 359, + 490 + ], + "blocks": [ + { + "bbox": [ + 301, + 430, + 359, + 490 + ], + "lines": [ + { + "bbox": [ + 301, + 430, + 359, + 490 + ], + "spans": [ + { + "bbox": [ + 301, + 430, + 359, + 490 + ], + "type": "image", + "image_path": "75e8040ab91ac2c4c85a4201768648c99b9057cc04444c34dc6ffde6fea71e47.jpg" + } + ] + } + ], + "index": 38, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 306, + 496, + 356, + 506 + ], + "lines": [ + { + "bbox": [ + 306, + 496, + 356, + 506 + ], + "spans": [ + { + "bbox": [ + 306, + 496, + 356, + 506 + ], + "type": "text", + "content": "Masked Image" + } + ] + } + ], + "index": 39, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 54, + 517, + 537, + 529 + ], + "lines": [ + { + "bbox": [ + 54, + 517, + 537, + 529 + ], + "spans": [ + { + "bbox": [ + 54, + 517, + 537, + 529 + ], + "type": "text", + "content": "Figure 5. Visual comparison on inpainting & outpainting between DreamOmni, ControlNet-Inpainting [62] and SD-inpainting [44]." + } + ] + } + ], + "index": 49, + "angle": 0, + "type": "image_caption" + } + ], + "index": 38 + }, + { + "type": "image", + "bbox": [ + 361, + 369, + 419, + 428 + ], + "blocks": [ + { + "bbox": [ + 361, + 369, + 419, + 428 + ], + "lines": [ + { + "bbox": [ + 361, + 369, + 419, + 428 + ], + "spans": [ + { + "bbox": [ + 361, + 369, + 419, + 428 + ], + "type": "image", + "image_path": "8bb81f02630bd43ec637e10d158244f669982a9a47c77ed4809f936eeca84a9d.jpg" + } + ] + } + ], + "index": 40, + "angle": 0, + "type": "image_body" + } + ], + "index": 40 + }, + { + "type": "image", + "bbox": [ + 361, + 430, + 419, + 490 + ], + "blocks": [ + { + "bbox": [ + 361, + 430, + 419, + 490 + ], + "lines": [ + { + "bbox": [ + 361, + 430, + 419, + 490 + ], + "spans": [ + { + "bbox": [ + 361, + 430, + 419, + 490 + ], + "type": "image", + "image_path": "51ec26871e80adffb8ce709f78b0a8c022c985b30f9dc915f34a17d734d956b9.jpg" + } + ] + } + ], + "index": 41, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 373, + 493, + 412, + 511 + ], + "lines": [ + { + "bbox": [ + 373, + 493, + 412, + 511 + ], + "spans": [ + { + "bbox": [ + 373, + 493, + 412, + 511 + ], + "type": "text", + "content": "ControlNet Inpainting" + } + ] + } + ], + "index": 42, + "angle": 0, + "type": "image_caption" + } + ], + "index": 41 + }, + { + "type": "image", + "bbox": [ + 421, + 369, + 479, + 428 + ], + "blocks": [ + { + "bbox": [ + 421, + 369, + 479, + 428 + ], + "lines": [ + { + "bbox": [ + 421, + 369, + 479, + 428 + ], + "spans": [ + { + "bbox": [ + 421, + 369, + 479, + 428 + ], + "type": "image", + "image_path": "a5e8e081ae3e2ce2b20a07485c536e0988ed2d0286267adf01c053bc378f722d.jpg" + } + ] + } + ], + "index": 43, + "angle": 0, + "type": "image_body" + } + ], + "index": 43 + }, + { + "type": "image", + "bbox": [ + 421, + 430, + 479, + 490 + ], + "blocks": [ + { + "bbox": [ + 421, + 430, + 479, + 490 + ], + "lines": [ + { + "bbox": [ + 421, + 430, + 479, + 490 + ], + "spans": [ + { + "bbox": [ + 421, + 430, + 479, + 490 + ], + "type": "image", + "image_path": "00311f0d0da94a5007156fcc982ee548bc98a4cb4daf938907b3c1add935292a.jpg" + } + ] + } + ], + "index": 44, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 428, + 496, + 476, + 506 + ], + "lines": [ + { + "bbox": [ + 428, + 496, + 476, + 506 + ], + "spans": [ + { + "bbox": [ + 428, + 496, + 476, + 506 + ], + "type": "text", + "content": "SD Inpainting" + } + ] + } + ], + "index": 45, + "angle": 0, + "type": "image_caption" + } + ], + "index": 44 + }, + { + "type": "image", + "bbox": [ + 481, + 369, + 540, + 428 + ], + "blocks": [ + { + "bbox": [ + 481, + 369, + 540, + 428 + ], + "lines": [ + { + "bbox": [ + 481, + 369, + 540, + 428 + ], + "spans": [ + { + "bbox": [ + 481, + 369, + 540, + 428 + ], + "type": "image", + "image_path": "df38ee661153399d24cdbbade590ea3dec06c0e5b769793eda28fa91fd424280.jpg" + } + ] + } + ], + "index": 46, + "angle": 0, + "type": "image_body" + } + ], + "index": 46 + }, + { + "type": "image", + "bbox": [ + 481, + 430, + 540, + 490 + ], + "blocks": [ + { + "bbox": [ + 481, + 430, + 540, + 490 + ], + "lines": [ + { + "bbox": [ + 481, + 430, + 540, + 490 + ], + "spans": [ + { + "bbox": [ + 481, + 430, + 540, + 490 + ], + "type": "image", + "image_path": "3faede984f7722c407634dbfae9c4391608161546e5dfc9968ba4f3db535fdb0.jpg" + } + ] + } + ], + "index": 47, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 491, + 497, + 533, + 506 + ], + "lines": [ + { + "bbox": [ + 491, + 497, + 533, + 506 + ], + "spans": [ + { + "bbox": [ + 491, + 497, + 533, + 506 + ], + "type": "text", + "content": "DreamOmni" + } + ] + } + ], + "index": 48, + "angle": 0, + "type": "image_caption" + } + ], + "index": 47 + }, + { + "bbox": [ + 46, + 544, + 287, + 628 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 544, + 287, + 628 + ], + "spans": [ + { + "bbox": [ + 46, + 544, + 287, + 628 + ], + "type": "text", + "content": "We train models with " + }, + { + "bbox": [ + 46, + 544, + 287, + 628 + ], + "type": "inline_equation", + "content": "1024 \\times 1024" + }, + { + "bbox": [ + 46, + 544, + 287, + 628 + ], + "type": "text", + "content": " size, a batch size of 256, a learning rate of " + }, + { + "bbox": [ + 46, + 544, + 287, + 628 + ], + "type": "inline_equation", + "content": "2 \\times 10^{-5}" + }, + { + "bbox": [ + 46, + 544, + 287, + 628 + ], + "type": "text", + "content": ", and " + }, + { + "bbox": [ + 46, + 544, + 287, + 628 + ], + "type": "inline_equation", + "content": "140K" + }, + { + "bbox": [ + 46, + 544, + 287, + 628 + ], + "type": "text", + "content": " iterations. All experiments are conducted on 64 A100 GPUs. Additionally, to enable the model to generate images at varying resolutions, similar to the approach used in SDXL [39], we divide the images into 31 buckets based on their aspect ratios, ranging from " + }, + { + "bbox": [ + 46, + 544, + 287, + 628 + ], + "type": "inline_equation", + "content": "4:1" + }, + { + "bbox": [ + 46, + 544, + 287, + 628 + ], + "type": "text", + "content": " to " + }, + { + "bbox": [ + 46, + 544, + 287, + 628 + ], + "type": "inline_equation", + "content": "1:4" + }, + { + "bbox": [ + 46, + 544, + 287, + 628 + ], + "type": "text", + "content": ", during training." + } + ] + } + ], + "index": 50 + }, + { + "bbox": [ + 47, + 634, + 128, + 647 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 634, + 128, + 647 + ], + "spans": [ + { + "bbox": [ + 47, + 634, + 128, + 647 + ], + "type": "text", + "content": "4. Experiments" + } + ] + } + ], + "index": 51 + }, + { + "bbox": [ + 46, + 653, + 287, + 712 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 653, + 287, + 712 + ], + "spans": [ + { + "bbox": [ + 46, + 653, + 287, + 712 + ], + "type": "text", + "content": "Evaluation on Framework. We compared several T2I model frameworks under similar settings to identify effective components. Using the same VAE, CLIP text encoder, parameters, runtime, and LAION training/testing datasets, we evaluated the Unet-based SDXL [39], DIT-" + } + ] + } + ], + "index": 52 + }, + { + "bbox": [ + 304, + 544, + 545, + 712 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 544, + 545, + 712 + ], + "spans": [ + { + "bbox": [ + 304, + 544, + 545, + 712 + ], + "type": "text", + "content": "based Pixart [10], SD3-Medium [15], and our developed DreamOmni variations. Notably, as shown in Fig. 3, to facilitate comparison, we do not use the full " + }, + { + "bbox": [ + 304, + 544, + 545, + 712 + ], + "type": "inline_equation", + "content": "2.5B" + }, + { + "bbox": [ + 304, + 544, + 545, + 712 + ], + "type": "text", + "content": " parameters of DreamOmni, but instead adjust the parameters of all models to " + }, + { + "bbox": [ + 304, + 544, + 545, + 712 + ], + "type": "inline_equation", + "content": "0.85B" + }, + { + "bbox": [ + 304, + 544, + 545, + 712 + ], + "type": "text", + "content": ". DreamOmni-V1 has two downsampling layers (" + }, + { + "bbox": [ + 304, + 544, + 545, + 712 + ], + "type": "inline_equation", + "content": "2\\times" + }, + { + "bbox": [ + 304, + 544, + 545, + 712 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 304, + 544, + 545, + 712 + ], + "type": "inline_equation", + "content": "4\\times" + }, + { + "bbox": [ + 304, + 544, + 545, + 712 + ], + "type": "text", + "content": ") like SDXL but lacks an Unet connection. DreamOmni-V2 builds on DreamOmni-V1 by adding the Unet connection, while DreamOmni-V3 takes DreamOmni-V2 further by focusing all DIT operations on " + }, + { + "bbox": [ + 304, + 544, + 545, + 712 + ], + "type": "inline_equation", + "content": "2\\times" + }, + { + "bbox": [ + 304, + 544, + 545, + 712 + ], + "type": "text", + "content": " downsampled latent. (1) Models with the Unet connection (SDXL, DreamOmni-V2, DreamOmni-V3) showed significantly faster convergence than those without (SD3-Medium, DreamOmni-V1). DreamOmni-V3, notably, converges four times faster than the SD3-Medium, enhancing" + } + ] + } + ], + "index": 53 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "text", + "content": "28538" + } + ] + } + ], + "index": 54 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 5 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 117, + 73, + 175, + 132 + ], + "blocks": [ + { + "bbox": [ + 56, + 76, + 113, + 126 + ], + "lines": [ + { + "bbox": [ + 56, + 76, + 113, + 126 + ], + "spans": [ + { + "bbox": [ + 56, + 76, + 113, + 126 + ], + "type": "text", + "content": "A humanoid cat from the race american shorthair with black fur. Standing as a barbarian tabaxi with a greatsword. Dressed in light clothing, something like the ancient geeks weared, photorealistic. standing in front of an oasis in the dessert." + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 117, + 73, + 175, + 132 + ], + "lines": [ + { + "bbox": [ + 117, + 73, + 175, + 132 + ], + "spans": [ + { + "bbox": [ + 117, + 73, + 175, + 132 + ], + "type": "image", + "image_path": "828f4d95ba889ae4b9d04ba9e6f79da3c1518fc6e78677b058d2a28f2d84c97c.jpg" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_body" + } + ], + "index": 3 + }, + { + "type": "image", + "bbox": [ + 117, + 135, + 175, + 194 + ], + "blocks": [ + { + "bbox": [ + 56, + 155, + 113, + 183 + ], + "lines": [ + { + "bbox": [ + 56, + 155, + 113, + 183 + ], + "spans": [ + { + "bbox": [ + 56, + 155, + 113, + 183 + ], + "type": "text", + "content": "elf girl travelling alongside a majestRyuic dragon, close-up view, anime style, high fantasy, digital art, masterpiece painting" + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 117, + 135, + 175, + 194 + ], + "lines": [ + { + "bbox": [ + 117, + 135, + 175, + 194 + ], + "spans": [ + { + "bbox": [ + 117, + 135, + 175, + 194 + ], + "type": "image", + "image_path": "26dcdb6d7cfc5e275ac15e558032eb3569c54ae11a605d06bebdeb82c5f1188a.jpg" + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 133, + 195, + 156, + 204 + ], + "lines": [ + { + "bbox": [ + 133, + 195, + 156, + 204 + ], + "spans": [ + { + "bbox": [ + 133, + 195, + 156, + 204 + ], + "type": "text", + "content": "Image" + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 129, + 205, + 162, + 213 + ], + "lines": [ + { + "bbox": [ + 129, + 205, + 162, + 213 + ], + "spans": [ + { + "bbox": [ + 129, + 205, + 162, + 213 + ], + "type": "text", + "content": "Condition" + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "image_caption" + } + ], + "index": 4 + }, + { + "type": "image", + "bbox": [ + 177, + 73, + 235, + 132 + ], + "blocks": [ + { + "bbox": [ + 177, + 73, + 235, + 132 + ], + "lines": [ + { + "bbox": [ + 177, + 73, + 235, + 132 + ], + "spans": [ + { + "bbox": [ + 177, + 73, + 235, + 132 + ], + "type": "image", + "image_path": "58f6e08643e10e515a0b16dcc3836f09d102f5ba42f2f583e250ecdee69c9ec6.jpg" + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "image_body" + } + ], + "index": 7 + }, + { + "type": "image", + "bbox": [ + 177, + 135, + 235, + 194 + ], + "blocks": [ + { + "bbox": [ + 177, + 135, + 235, + 194 + ], + "lines": [ + { + "bbox": [ + 177, + 135, + 235, + 194 + ], + "spans": [ + { + "bbox": [ + 177, + 135, + 235, + 194 + ], + "type": "image", + "image_path": "d7cc60c44731c166bae4251f1a3dc113b280698fdca951d03280dade47264e1e.jpg" + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 187, + 198, + 225, + 208 + ], + "lines": [ + { + "bbox": [ + 187, + 198, + 225, + 208 + ], + "spans": [ + { + "bbox": [ + 187, + 198, + 225, + 208 + ], + "type": "text", + "content": "ControlNet" + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "image_caption" + } + ], + "index": 8 + }, + { + "type": "image", + "bbox": [ + 236, + 74, + 296, + 132 + ], + "blocks": [ + { + "bbox": [ + 236, + 74, + 296, + 132 + ], + "lines": [ + { + "bbox": [ + 236, + 74, + 296, + 132 + ], + "spans": [ + { + "bbox": [ + 236, + 74, + 296, + 132 + ], + "type": "image", + "image_path": "f1125c81febc4efb17f152ce3de1408a5b9369fbad67999ae27170671a1ef680.jpg" + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 299, + 76, + 356, + 131 + ], + "lines": [ + { + "bbox": [ + 299, + 76, + 356, + 131 + ], + "spans": [ + { + "bbox": [ + 299, + 76, + 356, + 131 + ], + "type": "text", + "content": "A cute teenage-child fairy with blue hair and butterflies in her ears, holding balloons and a birthday cake with a candle lit on it in her other hand. Her hand is holding the plate. The artwork is in the style pixar anime. her facial expression is very happy and excited but in a cute way" + } + ] + } + ], + "index": 14, + "angle": 0, + "type": "image_caption" + } + ], + "index": 10 + }, + { + "type": "image", + "bbox": [ + 236, + 135, + 296, + 194 + ], + "blocks": [ + { + "bbox": [ + 236, + 135, + 296, + 194 + ], + "lines": [ + { + "bbox": [ + 236, + 135, + 296, + 194 + ], + "spans": [ + { + "bbox": [ + 236, + 135, + 296, + 194 + ], + "type": "image", + "image_path": "9dd97af1fa68a90233a7857ba19ae9d301934d958c03038bc8630d1c1eae324c.jpg" + } + ] + } + ], + "index": 11, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 244, + 195, + 286, + 204 + ], + "lines": [ + { + "bbox": [ + 244, + 195, + 286, + 204 + ], + "spans": [ + { + "bbox": [ + 244, + 195, + 286, + 204 + ], + "type": "text", + "content": "DreamOmni" + } + ] + } + ], + "index": 12, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 255, + 205, + 277, + 213 + ], + "lines": [ + { + "bbox": [ + 255, + 205, + 277, + 213 + ], + "spans": [ + { + "bbox": [ + 255, + 205, + 277, + 213 + ], + "type": "text", + "content": "(Ours)" + } + ] + } + ], + "index": 13, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 309, + 199, + 334, + 209 + ], + "lines": [ + { + "bbox": [ + 309, + 199, + 334, + 209 + ], + "spans": [ + { + "bbox": [ + 309, + 199, + 334, + 209 + ], + "type": "text", + "content": "Prompt" + } + ] + } + ], + "index": 16, + "angle": 0, + "type": "image_caption" + } + ], + "index": 11 + }, + { + "type": "image", + "bbox": [ + 359, + 73, + 417, + 132 + ], + "blocks": [ + { + "bbox": [ + 359, + 73, + 417, + 132 + ], + "lines": [ + { + "bbox": [ + 359, + 73, + 417, + 132 + ], + "spans": [ + { + "bbox": [ + 359, + 73, + 417, + 132 + ], + "type": "image", + "image_path": "6a28139b11cfc28452b984717b712e5a436905c3c254b81eed52622643c97945.jpg" + } + ] + } + ], + "index": 17, + "angle": 0, + "type": "image_body" + } + ], + "index": 17 + }, + { + "type": "image", + "bbox": [ + 359, + 135, + 417, + 194 + ], + "blocks": [ + { + "bbox": [ + 301, + 157, + 355, + 169 + ], + "lines": [ + { + "bbox": [ + 301, + 157, + 355, + 169 + ], + "spans": [ + { + "bbox": [ + 301, + 157, + 355, + 169 + ], + "type": "text", + "content": "ironman is riding a white little pony" + } + ] + } + ], + "index": 15, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 359, + 135, + 417, + 194 + ], + "lines": [ + { + "bbox": [ + 359, + 135, + 417, + 194 + ], + "spans": [ + { + "bbox": [ + 359, + 135, + 417, + 194 + ], + "type": "image", + "image_path": "5f236f35dd1ff78053a54c74cd5172dffa72abcf20bbbd0e4750a8608ac84027.jpg" + } + ] + } + ], + "index": 18, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 378, + 195, + 399, + 204 + ], + "lines": [ + { + "bbox": [ + 378, + 195, + 399, + 204 + ], + "spans": [ + { + "bbox": [ + 378, + 195, + 399, + 204 + ], + "type": "text", + "content": "Image" + } + ] + } + ], + "index": 19, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 371, + 205, + 405, + 213 + ], + "lines": [ + { + "bbox": [ + 371, + 205, + 405, + 213 + ], + "spans": [ + { + "bbox": [ + 371, + 205, + 405, + 213 + ], + "type": "text", + "content": "Condition" + } + ] + } + ], + "index": 20, + "angle": 0, + "type": "image_caption" + } + ], + "index": 18 + }, + { + "type": "image", + "bbox": [ + 419, + 73, + 476, + 132 + ], + "blocks": [ + { + "bbox": [ + 419, + 73, + 476, + 132 + ], + "lines": [ + { + "bbox": [ + 419, + 73, + 476, + 132 + ], + "spans": [ + { + "bbox": [ + 419, + 73, + 476, + 132 + ], + "type": "image", + "image_path": "b86641f75a31a3ea4edb1d74175b54aa1e84d4c66c25335a837126047a6d3752.jpg" + } + ] + } + ], + "index": 21, + "angle": 0, + "type": "image_body" + } + ], + "index": 21 + }, + { + "type": "image", + "bbox": [ + 419, + 135, + 476, + 194 + ], + "blocks": [ + { + "bbox": [ + 419, + 135, + 476, + 194 + ], + "lines": [ + { + "bbox": [ + 419, + 135, + 476, + 194 + ], + "spans": [ + { + "bbox": [ + 419, + 135, + 476, + 194 + ], + "type": "image", + "image_path": "9fff3d28baf0402e201d513cdec8bc8d0439c2f79d63ae299bf85a10326b51d2.jpg" + } + ] + } + ], + "index": 22, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 429, + 199, + 467, + 207 + ], + "lines": [ + { + "bbox": [ + 429, + 199, + 467, + 207 + ], + "spans": [ + { + "bbox": [ + 429, + 199, + 467, + 207 + ], + "type": "text", + "content": "ControlNet" + } + ] + } + ], + "index": 23, + "angle": 0, + "type": "image_caption" + } + ], + "index": 22 + }, + { + "type": "image", + "bbox": [ + 478, + 73, + 537, + 132 + ], + "blocks": [ + { + "bbox": [ + 478, + 73, + 537, + 132 + ], + "lines": [ + { + "bbox": [ + 478, + 73, + 537, + 132 + ], + "spans": [ + { + "bbox": [ + 478, + 73, + 537, + 132 + ], + "type": "image", + "image_path": "8c3e9d93f8ffec9dc5f980640ac62d4d8280c232679ade8c4ad77b61f0fb0f81.jpg" + } + ] + } + ], + "index": 24, + "angle": 0, + "type": "image_body" + } + ], + "index": 24 + }, + { + "type": "image", + "bbox": [ + 478, + 135, + 537, + 194 + ], + "blocks": [ + { + "bbox": [ + 478, + 135, + 537, + 194 + ], + "lines": [ + { + "bbox": [ + 478, + 135, + 537, + 194 + ], + "spans": [ + { + "bbox": [ + 478, + 135, + 537, + 194 + ], + "type": "image", + "image_path": "6316ebfc0ed010e05c69008c1d5a9e033a7937675262fcba8f771ebdf77b37b8.jpg" + } + ] + } + ], + "index": 25, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 487, + 195, + 528, + 204 + ], + "lines": [ + { + "bbox": [ + 487, + 195, + 528, + 204 + ], + "spans": [ + { + "bbox": [ + 487, + 195, + 528, + 204 + ], + "type": "text", + "content": "DreamOmni" + } + ] + } + ], + "index": 26, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 498, + 205, + 520, + 215 + ], + "lines": [ + { + "bbox": [ + 498, + 205, + 520, + 215 + ], + "spans": [ + { + "bbox": [ + 498, + 205, + 520, + 215 + ], + "type": "text", + "content": "(Ours)" + } + ] + } + ], + "index": 27, + "angle": 0, + "type": "image_caption" + } + ], + "index": 25 + }, + { + "type": "image", + "bbox": [ + 108, + 266, + 141, + 316 + ], + "blocks": [ + { + "bbox": [ + 66, + 198, + 92, + 209 + ], + "lines": [ + { + "bbox": [ + 66, + 198, + 92, + 209 + ], + "spans": [ + { + "bbox": [ + 66, + 198, + 92, + 209 + ], + "type": "text", + "content": "Prompt" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 53, + 267, + 106, + 312 + ], + "lines": [ + { + "bbox": [ + 53, + 267, + 106, + 312 + ], + "spans": [ + { + "bbox": [ + 53, + 267, + 106, + 312 + ], + "type": "text", + "content": "A girl (in the given image) is playing a red electric guitar. She is wearing a black jacket and a banie, and she has headphones. The background features autumn leaves, suggesting a seasonal setting." + } + ] + } + ], + "index": 29, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 53, + 317, + 105, + 373 + ], + "lines": [ + { + "bbox": [ + 53, + 317, + 105, + 373 + ], + "spans": [ + { + "bbox": [ + 53, + 317, + 105, + 373 + ], + "type": "text", + "content": "Generate an image using the character from the supplied image as reference. A girl is holding a lantern. She is surrounded by vibrant flowers and a whimsical, magical atmosphere. The girl is wearing a blue and white outfit with a ruffled collar, and she has a thoughtful expression." + } + ] + } + ], + "index": 30, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 108, + 266, + 141, + 316 + ], + "lines": [ + { + "bbox": [ + 108, + 266, + 141, + 316 + ], + "spans": [ + { + "bbox": [ + 108, + 266, + 141, + 316 + ], + "type": "image", + "image_path": "ce23874b6bc3822c78a52cf1f92f4c5442e5394d7e0835485f2e51433bf03367.jpg" + } + ] + } + ], + "index": 32, + "angle": 0, + "type": "image_body" + } + ], + "index": 32 + }, + { + "type": "image", + "bbox": [ + 110, + 319, + 140, + 371 + ], + "blocks": [ + { + "bbox": [ + 66, + 377, + 89, + 386 + ], + "lines": [ + { + "bbox": [ + 66, + 377, + 89, + 386 + ], + "spans": [ + { + "bbox": [ + 66, + 377, + 89, + 386 + ], + "type": "text", + "content": "Prompt" + } + ] + } + ], + "index": 31, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 110, + 319, + 140, + 371 + ], + "lines": [ + { + "bbox": [ + 110, + 319, + 140, + 371 + ], + "spans": [ + { + "bbox": [ + 110, + 319, + 140, + 371 + ], + "type": "image", + "image_path": "71d6401569b886c3669726ff3af16b5a21a45d87a364d700d75c1175a4fad111.jpg" + } + ] + } + ], + "index": 33, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 115, + 375, + 138, + 390 + ], + "lines": [ + { + "bbox": [ + 115, + 375, + 138, + 390 + ], + "spans": [ + { + "bbox": [ + 115, + 375, + 138, + 390 + ], + "type": "text", + "content": "Image condition" + } + ] + } + ], + "index": 34, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 46, + 399, + 546, + 432 + ], + "lines": [ + { + "bbox": [ + 46, + 399, + 546, + 432 + ], + "spans": [ + { + "bbox": [ + 46, + 399, + 546, + 432 + ], + "type": "text", + "content": "Figure 7. Visual comparison on subject-driven generation. Compared with BLIP-Diffusion [28] and IP-adapter [60], DreamOmni excels at accurately following user prompts while preserving the specified subject. Furthermore, its success across both photo and anime cases underscores DreamOmni's generalization capabilities." + } + ] + } + ], + "index": 61, + "angle": 0, + "type": "image_caption" + } + ], + "index": 33 + }, + { + "type": "image", + "bbox": [ + 149, + 266, + 198, + 316 + ], + "blocks": [ + { + "bbox": [ + 46, + 221, + 546, + 255 + ], + "lines": [ + { + "bbox": [ + 46, + 221, + 546, + 255 + ], + "spans": [ + { + "bbox": [ + 46, + 221, + 546, + 255 + ], + "type": "text", + "content": "Figure 6. Visual comparison on image-conditioned generation. Our DreamOmni excels at using canny, depth, and segmentation maps as conditions for image generation. Compared to the classic ControlNet [62], DreamOmni not only follows the user's prompts and image conditions more accurately but also generates content and color schemes that are visually more pleasing." + } + ] + } + ], + "index": 28, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 149, + 266, + 198, + 316 + ], + "lines": [ + { + "bbox": [ + 149, + 266, + 198, + 316 + ], + "spans": [ + { + "bbox": [ + 149, + 266, + 198, + 316 + ], + "type": "image", + "image_path": "16ae7a1dd5f5fb65f7045c28e903c928bc95e9a424f69eaad6309229bd461333.jpg" + } + ] + } + ], + "index": 35, + "angle": 0, + "type": "image_body" + } + ], + "index": 35 + }, + { + "type": "image", + "bbox": [ + 149, + 319, + 198, + 371 + ], + "blocks": [ + { + "bbox": [ + 149, + 319, + 198, + 371 + ], + "lines": [ + { + "bbox": [ + 149, + 319, + 198, + 371 + ], + "spans": [ + { + "bbox": [ + 149, + 319, + 198, + 371 + ], + "type": "image", + "image_path": "29f2aa46a034fd5b8519eeadea1b1b6647695637649b4b09bc429ef85c335a59.jpg" + } + ] + } + ], + "index": 36, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 153, + 378, + 198, + 386 + ], + "lines": [ + { + "bbox": [ + 153, + 378, + 198, + 386 + ], + "spans": [ + { + "bbox": [ + 153, + 378, + 198, + 386 + ], + "type": "text", + "content": "BLIP-Diffusion" + } + ] + } + ], + "index": 37, + "angle": 0, + "type": "image_caption" + } + ], + "index": 36 + }, + { + "type": "image", + "bbox": [ + 200, + 266, + 249, + 316 + ], + "blocks": [ + { + "bbox": [ + 200, + 266, + 249, + 316 + ], + "lines": [ + { + "bbox": [ + 200, + 266, + 249, + 316 + ], + "spans": [ + { + "bbox": [ + 200, + 266, + 249, + 316 + ], + "type": "image", + "image_path": "5a999d537b19673d8251107ad71754ae76bffbc83a62e3bfa3ace6c256273ef4.jpg" + } + ] + } + ], + "index": 38, + "angle": 0, + "type": "image_body" + } + ], + "index": 38 + }, + { + "type": "image", + "bbox": [ + 200, + 319, + 250, + 371 + ], + "blocks": [ + { + "bbox": [ + 200, + 319, + 250, + 371 + ], + "lines": [ + { + "bbox": [ + 200, + 319, + 250, + 371 + ], + "spans": [ + { + "bbox": [ + 200, + 319, + 250, + 371 + ], + "type": "image", + "image_path": "55b9589255c538909e1ce964bcf9e95e60c83551b93b2619aed607cf1bbbd2cf.jpg" + } + ] + } + ], + "index": 39, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 211, + 378, + 241, + 386 + ], + "lines": [ + { + "bbox": [ + 211, + 378, + 241, + 386 + ], + "spans": [ + { + "bbox": [ + 211, + 378, + 241, + 386 + ], + "type": "text", + "content": "IP-adapter" + } + ] + } + ], + "index": 40, + "angle": 0, + "type": "image_caption" + } + ], + "index": 39 + }, + { + "type": "image", + "bbox": [ + 250, + 266, + 301, + 316 + ], + "blocks": [ + { + "bbox": [ + 250, + 266, + 301, + 316 + ], + "lines": [ + { + "bbox": [ + 250, + 266, + 301, + 316 + ], + "spans": [ + { + "bbox": [ + 250, + 266, + 301, + 316 + ], + "type": "image", + "image_path": "17fbb8252ee01bae475fa6c3dda431ef06b63f3009f8de0afdc9fabf40238dca.jpg" + } + ] + } + ], + "index": 41, + "angle": 0, + "type": "image_body" + } + ], + "index": 41 + }, + { + "type": "image", + "bbox": [ + 250, + 319, + 301, + 371 + ], + "blocks": [ + { + "bbox": [ + 250, + 319, + 301, + 371 + ], + "lines": [ + { + "bbox": [ + 250, + 319, + 301, + 371 + ], + "spans": [ + { + "bbox": [ + 250, + 319, + 301, + 371 + ], + "type": "image", + "image_path": "1c302bb80fc9aa2d8507c0b7e2497803e94efc2126323b5e19cd89aea8723c3b.jpg" + } + ] + } + ], + "index": 42, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 258, + 374, + 294, + 382 + ], + "lines": [ + { + "bbox": [ + 258, + 374, + 294, + 382 + ], + "spans": [ + { + "bbox": [ + 258, + 374, + 294, + 382 + ], + "type": "text", + "content": "DreamOmni" + } + ] + } + ], + "index": 43, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 268, + 383, + 287, + 391 + ], + "lines": [ + { + "bbox": [ + 268, + 383, + 287, + 391 + ], + "spans": [ + { + "bbox": [ + 268, + 383, + 287, + 391 + ], + "type": "text", + "content": "(Ours)" + } + ] + } + ], + "index": 44, + "angle": 0, + "type": "image_caption" + } + ], + "index": 42 + }, + { + "type": "image", + "bbox": [ + 343, + 266, + 383, + 316 + ], + "blocks": [ + { + "bbox": [ + 306, + 274, + 340, + 308 + ], + "lines": [ + { + "bbox": [ + 306, + 274, + 340, + 308 + ], + "spans": [ + { + "bbox": [ + 306, + 274, + 340, + 308 + ], + "type": "text", + "content": "A man (in the given image) is wearing a white t-shirt and white pants. He is smiling." + } + ] + } + ], + "index": 45, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 343, + 266, + 383, + 316 + ], + "lines": [ + { + "bbox": [ + 343, + 266, + 383, + 316 + ], + "spans": [ + { + "bbox": [ + 343, + 266, + 383, + 316 + ], + "type": "image", + "image_path": "2e3935f859f63a69b1b29076674734db78a7a57aff7772ecdf3bf5e8d1666adf.jpg" + } + ] + } + ], + "index": 48, + "angle": 0, + "type": "image_body" + } + ], + "index": 48 + }, + { + "type": "image", + "bbox": [ + 343, + 319, + 381, + 371 + ], + "blocks": [ + { + "bbox": [ + 307, + 329, + 340, + 362 + ], + "lines": [ + { + "bbox": [ + 307, + 329, + 340, + 362 + ], + "spans": [ + { + "bbox": [ + 307, + 329, + 340, + 362 + ], + "type": "text", + "content": "A close-up of a woman (in the given image) with her head tilted slightly to the side." + } + ] + } + ], + "index": 46, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 313, + 378, + 335, + 387 + ], + "lines": [ + { + "bbox": [ + 313, + 378, + 335, + 387 + ], + "spans": [ + { + "bbox": [ + 313, + 378, + 335, + 387 + ], + "type": "text", + "content": "Prompt" + } + ] + } + ], + "index": 47, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 343, + 319, + 381, + 371 + ], + "lines": [ + { + "bbox": [ + 343, + 319, + 381, + 371 + ], + "spans": [ + { + "bbox": [ + 343, + 319, + 381, + 371 + ], + "type": "image", + "image_path": "e33c2d93ffff5a82d96ea400c5aeae17ddd9374a2323de34025f89940de84673.jpg" + } + ] + } + ], + "index": 49, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 350, + 375, + 378, + 390 + ], + "lines": [ + { + "bbox": [ + 350, + 375, + 378, + 390 + ], + "spans": [ + { + "bbox": [ + 350, + 375, + 378, + 390 + ], + "type": "text", + "content": "Image Condition" + } + ] + } + ], + "index": 50, + "angle": 0, + "type": "image_caption" + } + ], + "index": 49 + }, + { + "type": "image", + "bbox": [ + 389, + 266, + 438, + 316 + ], + "blocks": [ + { + "bbox": [ + 389, + 266, + 438, + 316 + ], + "lines": [ + { + "bbox": [ + 389, + 266, + 438, + 316 + ], + "spans": [ + { + "bbox": [ + 389, + 266, + 438, + 316 + ], + "type": "image", + "image_path": "e7d1a78cf261fcb7e43495e79540c1b3fb005e50ec5df9693ac8a76d247346b9.jpg" + } + ] + } + ], + "index": 51, + "angle": 0, + "type": "image_body" + } + ], + "index": 51 + }, + { + "type": "image", + "bbox": [ + 389, + 319, + 438, + 370 + ], + "blocks": [ + { + "bbox": [ + 389, + 319, + 438, + 370 + ], + "lines": [ + { + "bbox": [ + 389, + 319, + 438, + 370 + ], + "spans": [ + { + "bbox": [ + 389, + 319, + 438, + 370 + ], + "type": "image", + "image_path": "45833f4e0aeb40caf13eab52f03c1d41ccecffd7f1a7477913bd21031894b971.jpg" + } + ] + } + ], + "index": 52, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 392, + 378, + 438, + 386 + ], + "lines": [ + { + "bbox": [ + 392, + 378, + 438, + 386 + ], + "spans": [ + { + "bbox": [ + 392, + 378, + 438, + 386 + ], + "type": "text", + "content": "BLIP-Diffusion" + } + ] + } + ], + "index": 53, + "angle": 0, + "type": "image_caption" + } + ], + "index": 52 + }, + { + "type": "image", + "bbox": [ + 440, + 266, + 488, + 316 + ], + "blocks": [ + { + "bbox": [ + 440, + 266, + 488, + 316 + ], + "lines": [ + { + "bbox": [ + 440, + 266, + 488, + 316 + ], + "spans": [ + { + "bbox": [ + 440, + 266, + 488, + 316 + ], + "type": "image", + "image_path": "266f72eaee2b4599407b8fd9ee640e181ca837262dc687317c3aad7da75b7a05.jpg" + } + ] + } + ], + "index": 54, + "angle": 0, + "type": "image_body" + } + ], + "index": 54 + }, + { + "type": "image", + "bbox": [ + 440, + 319, + 488, + 370 + ], + "blocks": [ + { + "bbox": [ + 440, + 319, + 488, + 370 + ], + "lines": [ + { + "bbox": [ + 440, + 319, + 488, + 370 + ], + "spans": [ + { + "bbox": [ + 440, + 319, + 488, + 370 + ], + "type": "image", + "image_path": "08e9a6fce616c8cfcf902f6e6fe893cc7e1261092c035db857ace89280b2edb8.jpg" + } + ] + } + ], + "index": 55, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 451, + 378, + 481, + 386 + ], + "lines": [ + { + "bbox": [ + 451, + 378, + 481, + 386 + ], + "spans": [ + { + "bbox": [ + 451, + 378, + 481, + 386 + ], + "type": "text", + "content": "IP-adapter" + } + ] + } + ], + "index": 56, + "angle": 0, + "type": "image_caption" + } + ], + "index": 55 + }, + { + "type": "image", + "bbox": [ + 490, + 266, + 541, + 316 + ], + "blocks": [ + { + "bbox": [ + 490, + 266, + 541, + 316 + ], + "lines": [ + { + "bbox": [ + 490, + 266, + 541, + 316 + ], + "spans": [ + { + "bbox": [ + 490, + 266, + 541, + 316 + ], + "type": "image", + "image_path": "86df605e36e088a9b243442809bbb836755f08b9c5295b24865ca53516c645a5.jpg" + } + ] + } + ], + "index": 57, + "angle": 0, + "type": "image_body" + } + ], + "index": 57 + }, + { + "type": "image", + "bbox": [ + 490, + 319, + 541, + 370 + ], + "blocks": [ + { + "bbox": [ + 490, + 319, + 541, + 370 + ], + "lines": [ + { + "bbox": [ + 490, + 319, + 541, + 370 + ], + "spans": [ + { + "bbox": [ + 490, + 319, + 541, + 370 + ], + "type": "image", + "image_path": "d29fd01cf1cc4418b0cce699238d9dfb532335ea6f8cfbad690beeba0e7be888.jpg" + } + ] + } + ], + "index": 58, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 498, + 374, + 534, + 382 + ], + "lines": [ + { + "bbox": [ + 498, + 374, + 534, + 382 + ], + "spans": [ + { + "bbox": [ + 498, + 374, + 534, + 382 + ], + "type": "text", + "content": "DreamOmni" + } + ] + } + ], + "index": 59, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 507, + 383, + 526, + 391 + ], + "lines": [ + { + "bbox": [ + 507, + 383, + 526, + 391 + ], + "spans": [ + { + "bbox": [ + 507, + 383, + 526, + 391 + ], + "type": "text", + "content": "(Ours)" + } + ] + } + ], + "index": 60, + "angle": 0, + "type": "image_caption" + } + ], + "index": 58 + }, + { + "bbox": [ + 46, + 445, + 287, + 491 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 445, + 287, + 491 + ], + "spans": [ + { + "bbox": [ + 46, + 445, + 287, + 491 + ], + "type": "text", + "content": "both training and fine-tuning efficiency. (2) Comparing DreamOmni-V3 and DreamOmni-V2, we found that concentrating DIT block computations on higher-resolution latent " + }, + { + "bbox": [ + 46, + 445, + 287, + 491 + ], + "type": "inline_equation", + "content": "(2\\times)" + }, + { + "bbox": [ + 46, + 445, + 287, + 491 + ], + "type": "text", + "content": " is more cost-effective." + } + ] + } + ], + "index": 62 + }, + { + "bbox": [ + 46, + 495, + 287, + 615 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 495, + 287, + 615 + ], + "spans": [ + { + "bbox": [ + 46, + 495, + 287, + 615 + ], + "type": "text", + "content": "Evaluation on T2I generation. As shown in Tab. 1, our synthetic data significantly enhances DreamOmni's T2I generation capabilities in aspects such as quantity, color, and position, enabling our model to achieve SOTA results on GenEval [19]. Notably, SD3-Medium [15] is a " + }, + { + "bbox": [ + 46, + 495, + 287, + 615 + ], + "type": "inline_equation", + "content": "2B" + }, + { + "bbox": [ + 46, + 495, + 287, + 615 + ], + "type": "text", + "content": " open-source SOTA T2I model, which has a similar parameters as our DreamOmni. Furthermore, qualitative results are shown in Fig. 4. We can see that DreamOmni's outputs are not only more visually pleasing but also align more accurately with the given prompts." + } + ] + } + ], + "index": 63 + }, + { + "bbox": [ + 46, + 617, + 287, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 617, + 287, + 713 + ], + "spans": [ + { + "bbox": [ + 46, + 617, + 287, + 713 + ], + "type": "text", + "content": "Evaluation on Inpainting. We compare our DreamOmni with ControlNet-Inpainting [62] and SD-Inpainting [44] on our high-quality evaluation datasets to evaluate its performance. The quantitative results, presented in Tab. 2, demonstrate that DreamOmni significantly outperforms both ControlNet-Inpainting and SD-Inpainting, highlighting its superior generation quality and coherence. Visual results (Fig. 5) further emphasize that DreamOmni ex" + } + ] + } + ], + "index": 64 + }, + { + "bbox": [ + 304, + 445, + 545, + 492 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 445, + 545, + 492 + ], + "spans": [ + { + "bbox": [ + 304, + 445, + 545, + 492 + ], + "type": "text", + "content": "cels in generating fine details, surpassing both ControlNet-Inpainting and SD-Inpainting. Additionally, DreamOmni effectively handles large mask holes, generating realistic content instead of blurry and incoherent outputs." + } + ] + } + ], + "index": 65 + }, + { + "bbox": [ + 304, + 495, + 545, + 601 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 495, + 545, + 601 + ], + "spans": [ + { + "bbox": [ + 304, + 495, + 545, + 601 + ], + "type": "text", + "content": "Evaluation on Reference Image Generation. For image-conditioned generation, we conducted a comparison with ControlNet [62] on canny, depth, and segmentation image conditions. Visual results are presented in Fig. 6. These results reveal that, across all tested conditions, our DreamOmni significantly outperforms ControlNet. Our method not only adheres more faithfully to both the image condition and the prompt but also demonstrates enhanced visual quality with better composition and richer detail." + } + ] + } + ], + "index": 66 + }, + { + "bbox": [ + 304, + 604, + 545, + 699 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 604, + 545, + 699 + ], + "spans": [ + { + "bbox": [ + 304, + 604, + 545, + 699 + ], + "type": "text", + "content": "For subject-driven image generation, we compared our method with competitive approaches, including BLIP-Diffusion [28] and IP-Adapter [60]. To demonstrate DreamOmni's powerful generalization, we validated their performance on both anime and photographic images. Visual results are shown in Fig. 7. Compared to other methods, our DreamOmni not only preserves the specified subject effectively but also follows the prompt well." + } + ] + } + ], + "index": 67 + }, + { + "bbox": [ + 306, + 701, + 545, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 701, + 545, + 714 + ], + "spans": [ + { + "bbox": [ + 306, + 701, + 545, + 714 + ], + "type": "text", + "content": "Evaluation on Instruction Editing. We compare" + } + ] + } + ], + "index": 68 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "text", + "content": "28539" + } + ] + } + ], + "index": 69 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 6 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 93, + 74, + 144, + 125 + ], + "blocks": [ + { + "bbox": [ + 52, + 89, + 91, + 116 + ], + "lines": [ + { + "bbox": [ + 52, + 89, + 91, + 116 + ], + "spans": [ + { + "bbox": [ + 52, + 89, + 91, + 116 + ], + "type": "text", + "content": "Remove the star ornament from the snow-covered tree." + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 93, + 74, + 144, + 125 + ], + "lines": [ + { + "bbox": [ + 93, + 74, + 144, + 125 + ], + "spans": [ + { + "bbox": [ + 93, + 74, + 144, + 125 + ], + "type": "image", + "image_path": "26ac1993b18c5d5fde0b92a4e9abcdedef56978bce85191dbf35d6605508dcec.jpg" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_body" + } + ], + "index": 3 + }, + { + "type": "image", + "bbox": [ + 144, + 74, + 194, + 125 + ], + "blocks": [ + { + "bbox": [ + 144, + 74, + 194, + 125 + ], + "lines": [ + { + "bbox": [ + 144, + 74, + 194, + 125 + ], + "spans": [ + { + "bbox": [ + 144, + 74, + 194, + 125 + ], + "type": "image", + "image_path": "54c14c2dc65bf48785e3855b8126a9d12e462c98130142bc0feba4b036062dfa.jpg" + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_body" + } + ], + "index": 4 + }, + { + "type": "image", + "bbox": [ + 194, + 74, + 245, + 125 + ], + "blocks": [ + { + "bbox": [ + 194, + 74, + 245, + 125 + ], + "lines": [ + { + "bbox": [ + 194, + 74, + 245, + 125 + ], + "spans": [ + { + "bbox": [ + 194, + 74, + 245, + 125 + ], + "type": "image", + "image_path": "6ce6d807399980032fd04191a4813c2c4e47a85f36f070cd15743a33360f89f5.jpg" + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "image_body" + } + ], + "index": 5 + }, + { + "type": "image", + "bbox": [ + 245, + 74, + 296, + 125 + ], + "blocks": [ + { + "bbox": [ + 245, + 74, + 296, + 125 + ], + "lines": [ + { + "bbox": [ + 245, + 74, + 296, + 125 + ], + "spans": [ + { + "bbox": [ + 245, + 74, + 296, + 125 + ], + "type": "image", + "image_path": "b2055dde6260cd53640cad3ed66a63efab8d44c2af4ce21c54d7e4f4747e5545.jpg" + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "image_body" + } + ], + "index": 6 + }, + { + "type": "image", + "bbox": [ + 94, + 127, + 144, + 178 + ], + "blocks": [ + { + "bbox": [ + 53, + 143, + 90, + 163 + ], + "lines": [ + { + "bbox": [ + 53, + 143, + 90, + 163 + ], + "spans": [ + { + "bbox": [ + 53, + 143, + 90, + 163 + ], + "type": "text", + "content": "Add a straw hat on the person's head." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 56, + 180, + 88, + 188 + ], + "lines": [ + { + "bbox": [ + 56, + 180, + 88, + 188 + ], + "spans": [ + { + "bbox": [ + 56, + 180, + 88, + 188 + ], + "type": "text", + "content": "Instruction" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 94, + 127, + 144, + 178 + ], + "lines": [ + { + "bbox": [ + 94, + 127, + 144, + 178 + ], + "spans": [ + { + "bbox": [ + 94, + 127, + 144, + 178 + ], + "type": "image", + "image_path": "f574abba04f76f8f6d4d4d2e623b9ab8c62cb7f14eaedee413ffe33bdebc875f.jpg" + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 102, + 180, + 137, + 190 + ], + "lines": [ + { + "bbox": [ + 102, + 180, + 137, + 190 + ], + "spans": [ + { + "bbox": [ + 102, + 180, + 137, + 190 + ], + "type": "text", + "content": "Input Image" + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "image_caption" + } + ], + "index": 7 + }, + { + "type": "image", + "bbox": [ + 144, + 127, + 194, + 178 + ], + "blocks": [ + { + "bbox": [ + 144, + 127, + 194, + 178 + ], + "lines": [ + { + "bbox": [ + 144, + 127, + 194, + 178 + ], + "spans": [ + { + "bbox": [ + 144, + 127, + 194, + 178 + ], + "type": "image", + "image_path": "80aaffea2850bbcc03f2b6acc282035b9ad1aee9cae49ffc1ae4ec74be7dbfbb.jpg" + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 154, + 180, + 187, + 189 + ], + "lines": [ + { + "bbox": [ + 154, + 180, + 187, + 189 + ], + "spans": [ + { + "bbox": [ + 154, + 180, + 187, + 189 + ], + "type": "text", + "content": "InstructP2P" + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "image_caption" + } + ], + "index": 9 + }, + { + "type": "image", + "bbox": [ + 194, + 127, + 245, + 178 + ], + "blocks": [ + { + "bbox": [ + 194, + 127, + 245, + 178 + ], + "lines": [ + { + "bbox": [ + 194, + 127, + 245, + 178 + ], + "spans": [ + { + "bbox": [ + 194, + 127, + 245, + 178 + ], + "type": "image", + "image_path": "96fca43af04d42f865e53bed58ec7814453234d70bfff082d2f8e0fc820a9ded.jpg" + } + ] + } + ], + "index": 11, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 211, + 180, + 229, + 189 + ], + "lines": [ + { + "bbox": [ + 211, + 180, + 229, + 189 + ], + "spans": [ + { + "bbox": [ + 211, + 180, + 229, + 189 + ], + "type": "text", + "content": "MGIE" + } + ] + } + ], + "index": 12, + "angle": 0, + "type": "image_caption" + } + ], + "index": 11 + }, + { + "type": "image", + "bbox": [ + 246, + 127, + 297, + 178 + ], + "blocks": [ + { + "bbox": [ + 246, + 127, + 297, + 178 + ], + "lines": [ + { + "bbox": [ + 246, + 127, + 297, + 178 + ], + "spans": [ + { + "bbox": [ + 246, + 127, + 297, + 178 + ], + "type": "image", + "image_path": "dabcef71389c39f00a3f4ce172b9f164d67b1bf7d5f91f8ba3f6be55465727d4.jpg" + } + ] + } + ], + "index": 13, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 252, + 180, + 288, + 189 + ], + "lines": [ + { + "bbox": [ + 252, + 180, + 288, + 189 + ], + "spans": [ + { + "bbox": [ + 252, + 180, + 288, + 189 + ], + "type": "text", + "content": "DreamOmni" + } + ] + } + ], + "index": 14, + "angle": 0, + "type": "image_caption" + } + ], + "index": 13 + }, + { + "type": "image", + "bbox": [ + 337, + 74, + 388, + 125 + ], + "blocks": [ + { + "bbox": [ + 304, + 91, + 334, + 110 + ], + "lines": [ + { + "bbox": [ + 304, + 91, + 334, + 110 + ], + "spans": [ + { + "bbox": [ + 304, + 91, + 334, + 110 + ], + "type": "text", + "content": "Replace the owl with a hawk" + } + ] + } + ], + "index": 15, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 337, + 74, + 388, + 125 + ], + "lines": [ + { + "bbox": [ + 337, + 74, + 388, + 125 + ], + "spans": [ + { + "bbox": [ + 337, + 74, + 388, + 125 + ], + "type": "image", + "image_path": "d465152c882e8b47a9990ec8678da46c0e9df0bd956a71cd60b4381b5cd1a816.jpg" + } + ] + } + ], + "index": 18, + "angle": 0, + "type": "image_body" + } + ], + "index": 18 + }, + { + "type": "image", + "bbox": [ + 337, + 127, + 388, + 178 + ], + "blocks": [ + { + "bbox": [ + 304, + 138, + 334, + 171 + ], + "lines": [ + { + "bbox": [ + 304, + 138, + 334, + 171 + ], + "spans": [ + { + "bbox": [ + 304, + 138, + 334, + 171 + ], + "type": "text", + "content": "Remove the green leafy plant near the yellow flower." + } + ] + } + ], + "index": 16, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 302, + 180, + 334, + 189 + ], + "lines": [ + { + "bbox": [ + 302, + 180, + 334, + 189 + ], + "spans": [ + { + "bbox": [ + 302, + 180, + 334, + 189 + ], + "type": "text", + "content": "Instruction" + } + ] + } + ], + "index": 17, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 337, + 127, + 388, + 178 + ], + "lines": [ + { + "bbox": [ + 337, + 127, + 388, + 178 + ], + "spans": [ + { + "bbox": [ + 337, + 127, + 388, + 178 + ], + "type": "image", + "image_path": "b5e033c73ce20add6b6c4dcd2a1e6ed22923116fa2f039751487eecee7182d65.jpg" + } + ] + } + ], + "index": 19, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 347, + 180, + 381, + 189 + ], + "lines": [ + { + "bbox": [ + 347, + 180, + 381, + 189 + ], + "spans": [ + { + "bbox": [ + 347, + 180, + 381, + 189 + ], + "type": "text", + "content": "Input Image" + } + ] + } + ], + "index": 20, + "angle": 0, + "type": "image_caption" + } + ], + "index": 19 + }, + { + "type": "image", + "bbox": [ + 389, + 74, + 438, + 125 + ], + "blocks": [ + { + "bbox": [ + 389, + 74, + 438, + 125 + ], + "lines": [ + { + "bbox": [ + 389, + 74, + 438, + 125 + ], + "spans": [ + { + "bbox": [ + 389, + 74, + 438, + 125 + ], + "type": "image", + "image_path": "a1975dc3ff5eefccd8213013d3e20438d3c6877a160e00f0a7b351bf6e2abd8f.jpg" + } + ] + } + ], + "index": 21, + "angle": 0, + "type": "image_body" + } + ], + "index": 21 + }, + { + "type": "image", + "bbox": [ + 389, + 127, + 438, + 178 + ], + "blocks": [ + { + "bbox": [ + 389, + 127, + 438, + 178 + ], + "lines": [ + { + "bbox": [ + 389, + 127, + 438, + 178 + ], + "spans": [ + { + "bbox": [ + 389, + 127, + 438, + 178 + ], + "type": "image", + "image_path": "707c2c486db7c59b0c69672c49d4a71abc1b0ecd4d5c893846e20ef80e927956.jpg" + } + ] + } + ], + "index": 22, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 400, + 180, + 433, + 189 + ], + "lines": [ + { + "bbox": [ + 400, + 180, + 433, + 189 + ], + "spans": [ + { + "bbox": [ + 400, + 180, + 433, + 189 + ], + "type": "text", + "content": "InstructP2P" + } + ] + } + ], + "index": 23, + "angle": 0, + "type": "image_caption" + } + ], + "index": 22 + }, + { + "type": "image", + "bbox": [ + 439, + 74, + 488, + 125 + ], + "blocks": [ + { + "bbox": [ + 439, + 74, + 488, + 125 + ], + "lines": [ + { + "bbox": [ + 439, + 74, + 488, + 125 + ], + "spans": [ + { + "bbox": [ + 439, + 74, + 488, + 125 + ], + "type": "image", + "image_path": "e0578c815ace48acd27602d24a95d2e3188024e4a4031131a2a9d622096fde7c.jpg" + } + ] + } + ], + "index": 24, + "angle": 0, + "type": "image_body" + } + ], + "index": 24 + }, + { + "type": "image", + "bbox": [ + 439, + 127, + 488, + 178 + ], + "blocks": [ + { + "bbox": [ + 439, + 127, + 488, + 178 + ], + "lines": [ + { + "bbox": [ + 439, + 127, + 488, + 178 + ], + "spans": [ + { + "bbox": [ + 439, + 127, + 488, + 178 + ], + "type": "image", + "image_path": "65fb4c5bfdd36d542acf03c452d98ca0975af18d059f9cf80fa87d58af0b58be.jpg" + } + ] + } + ], + "index": 25, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 457, + 180, + 475, + 188 + ], + "lines": [ + { + "bbox": [ + 457, + 180, + 475, + 188 + ], + "spans": [ + { + "bbox": [ + 457, + 180, + 475, + 188 + ], + "type": "text", + "content": "MGIE" + } + ] + } + ], + "index": 26, + "angle": 0, + "type": "image_caption" + } + ], + "index": 25 + }, + { + "type": "image", + "bbox": [ + 490, + 74, + 541, + 125 + ], + "blocks": [ + { + "bbox": [ + 490, + 74, + 541, + 125 + ], + "lines": [ + { + "bbox": [ + 490, + 74, + 541, + 125 + ], + "spans": [ + { + "bbox": [ + 490, + 74, + 541, + 125 + ], + "type": "image", + "image_path": "390b6b6a83e771d22d678eb0f70598a295e599a3d1a29b5e0f8a04c3ea1f2934.jpg" + } + ] + } + ], + "index": 27, + "angle": 0, + "type": "image_body" + } + ], + "index": 27 + }, + { + "type": "image", + "bbox": [ + 490, + 127, + 541, + 178 + ], + "blocks": [ + { + "bbox": [ + 490, + 127, + 541, + 178 + ], + "lines": [ + { + "bbox": [ + 490, + 127, + 541, + 178 + ], + "spans": [ + { + "bbox": [ + 490, + 127, + 541, + 178 + ], + "type": "image", + "image_path": "7422a2c7d86afda30228ef2d1353bc5d923439e58f7998aa2469725f842b4284.jpg" + } + ] + } + ], + "index": 28, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 499, + 180, + 535, + 188 + ], + "lines": [ + { + "bbox": [ + 499, + 180, + 535, + 188 + ], + "spans": [ + { + "bbox": [ + 499, + 180, + 535, + 188 + ], + "type": "text", + "content": "DreamOmni" + } + ] + } + ], + "index": 29, + "angle": 0, + "type": "image_caption" + } + ], + "index": 28 + }, + { + "type": "table", + "bbox": [ + 50, + 247, + 287, + 316 + ], + "blocks": [ + { + "bbox": [ + 52, + 232, + 281, + 243 + ], + "lines": [ + { + "bbox": [ + 52, + 232, + 281, + 243 + ], + "spans": [ + { + "bbox": [ + 52, + 232, + 281, + 243 + ], + "type": "text", + "content": "Table 2. Quantitative comparison for inpainting & outpainting." + } + ] + } + ], + "index": 31, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 50, + 247, + 287, + 316 + ], + "lines": [ + { + "bbox": [ + 50, + 247, + 287, + 316 + ], + "spans": [ + { + "bbox": [ + 50, + 247, + 287, + 316 + ], + "type": "table", + "html": "
ModelInpaintingOutpainting
FID↓LPIPS↓FID↓LPIPS↓
SD-inpainting [44]1.35220.15602.91790.2475
ControlNet-inpainting [62]1.83930.15944.23370.2521
DreamOmni (Ours)0.83710.12031.69260.1995
", + "image_path": "66b9019baefe3beed15b0c539d69e73d6c0621b28a2f8d5276a03d47a8cd75a7.jpg" + } + ] + } + ], + "index": 32, + "angle": 0, + "type": "table_body" + } + ], + "index": 32 + }, + { + "bbox": [ + 46, + 327, + 286, + 470 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 327, + 286, + 470 + ], + "spans": [ + { + "bbox": [ + 46, + 327, + 286, + 470 + ], + "type": "text", + "content": "DreamOmni with competitive methods such as MGIE [17] and InstructP2P [8]. The visual results are presented in Fig. 8. We can see that DreamOmni performs more accurate edits, including addition, removal, and replacement. Specifically, our editing results exhibit superior consistency with minimal changes to non-edited areas and higher-quality generation of the edited content. This further validates that our synthetic data pipeline is an efficient and effective method for creating instruction-based editing datasets, enabling models to learn precise instruction-based editing. Moreover, the efficiency of our synthetic data pipeline allows for models to easily scale up diverse training data." + } + ] + } + ], + "index": 33 + }, + { + "bbox": [ + 46, + 471, + 286, + 638 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 471, + 286, + 638 + ], + "spans": [ + { + "bbox": [ + 46, + 471, + 286, + 638 + ], + "type": "text", + "content": "Evaluation on Drag Editing. We evaluate DreamOmni on our synthetic evaluation dataset. The visual results are shown in Fig. 9. (1) Compared to the target image, we can see that DreamOmni can accurately perform translation, rotation, and scaling drag edits. (2) for translation and scaling, DreamOmni can maintain the integrity of the dragged object. However, large-scale rotation operations are more challenging for DreamOmni, as they involve complex transformations of the object itself, which may lead to deformations of the edited object. (3) The results demonstrate the effectiveness of our synthetic data pipeline for drag editing, and encoding the drag point positions and displacement information as instruction inputs allows the model to learn precise drag edits (shown in Fig. 2)." + } + ] + } + ], + "index": 34 + }, + { + "bbox": [ + 47, + 645, + 119, + 658 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 645, + 119, + 658 + ], + "spans": [ + { + "bbox": [ + 47, + 645, + 119, + 658 + ], + "type": "text", + "content": "5. Conclusion" + } + ] + } + ], + "index": 35 + }, + { + "bbox": [ + 46, + 665, + 287, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 665, + 287, + 713 + ], + "spans": [ + { + "bbox": [ + 46, + 665, + 287, + 713 + ], + "type": "text", + "content": "Current T2I foundation models lack a unified framework and training for downstream tasks, such as image editing. To address this, we introduce DreamOmni, a unified model for T2I generation and editing. We evaluate the frameworks" + } + ] + } + ], + "index": 36 + }, + { + "type": "image", + "bbox": [ + 314, + 234, + 388, + 308 + ], + "blocks": [ + { + "bbox": [ + 46, + 198, + 545, + 220 + ], + "lines": [ + { + "bbox": [ + 46, + 198, + 545, + 220 + ], + "spans": [ + { + "bbox": [ + 46, + 198, + 545, + 220 + ], + "type": "text", + "content": "Figure 8. Visual comparison on instruction-based editing. Our DreamOmni achieves more precise editing (including addition, removal, and replacement) compared to competitive methods such as MGIE [17] and InstructP2P [8]." + } + ] + } + ], + "index": 30, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 314, + 234, + 388, + 308 + ], + "lines": [ + { + "bbox": [ + 314, + 234, + 388, + 308 + ], + "spans": [ + { + "bbox": [ + 314, + 234, + 388, + 308 + ], + "type": "image", + "image_path": "41a1ec74bd953ea13832c6550d8426b71c73ee5cd26514ff84cd2862f3753e5d.jpg" + } + ] + } + ], + "index": 37, + "angle": 0, + "type": "image_body" + } + ], + "index": 37 + }, + { + "type": "image", + "bbox": [ + 314, + 308, + 388, + 382 + ], + "blocks": [ + { + "bbox": [ + 314, + 308, + 388, + 382 + ], + "lines": [ + { + "bbox": [ + 314, + 308, + 388, + 382 + ], + "spans": [ + { + "bbox": [ + 314, + 308, + 388, + 382 + ], + "type": "image", + "image_path": "8e8d804f5243b71054d8655b1877e6915f5509e60e4e67efa14fce1c78aa32b9.jpg" + } + ] + } + ], + "index": 38, + "angle": 0, + "type": "image_body" + } + ], + "index": 38 + }, + { + "type": "image", + "bbox": [ + 314, + 383, + 388, + 458 + ], + "blocks": [ + { + "bbox": [ + 314, + 383, + 388, + 458 + ], + "lines": [ + { + "bbox": [ + 314, + 383, + 388, + 458 + ], + "spans": [ + { + "bbox": [ + 314, + 383, + 388, + 458 + ], + "type": "image", + "image_path": "90ab57500dc7cda0239a1c665654686bb9a9580257c2426083497beb015c1e8a.jpg" + } + ] + } + ], + "index": 39, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 317, + 461, + 378, + 485 + ], + "lines": [ + { + "bbox": [ + 317, + 461, + 378, + 485 + ], + "spans": [ + { + "bbox": [ + 317, + 461, + 378, + 485 + ], + "type": "text", + "content": "Input Image & Drag Points" + } + ] + } + ], + "index": 40, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 305, + 494, + 545, + 516 + ], + "lines": [ + { + "bbox": [ + 305, + 494, + 545, + 516 + ], + "spans": [ + { + "bbox": [ + 305, + 494, + 545, + 516 + ], + "type": "text", + "content": "Figure 9. Visual results of drag editing: DreamOmni accurately performs translation, rotation, and scaling edits." + } + ] + } + ], + "index": 49, + "angle": 0, + "type": "image_caption" + } + ], + "index": 39 + }, + { + "type": "image", + "bbox": [ + 390, + 234, + 463, + 308 + ], + "blocks": [ + { + "bbox": [ + 390, + 234, + 463, + 308 + ], + "lines": [ + { + "bbox": [ + 390, + 234, + 463, + 308 + ], + "spans": [ + { + "bbox": [ + 390, + 234, + 463, + 308 + ], + "type": "image", + "image_path": "590e66926b4764db811df766b8c8c1994e8dfc0af06f1942ce54ae5705e3bb37.jpg" + } + ] + } + ], + "index": 41, + "angle": 0, + "type": "image_body" + } + ], + "index": 41 + }, + { + "type": "image", + "bbox": [ + 390, + 308, + 463, + 382 + ], + "blocks": [ + { + "bbox": [ + 390, + 308, + 463, + 382 + ], + "lines": [ + { + "bbox": [ + 390, + 308, + 463, + 382 + ], + "spans": [ + { + "bbox": [ + 390, + 308, + 463, + 382 + ], + "type": "image", + "image_path": "21f7d314731317b03a2de6b89db9fbc21a33fc259b328262bf75da13053fe3fa.jpg" + } + ] + } + ], + "index": 42, + "angle": 0, + "type": "image_body" + } + ], + "index": 42 + }, + { + "type": "image", + "bbox": [ + 390, + 383, + 463, + 458 + ], + "blocks": [ + { + "bbox": [ + 390, + 383, + 463, + 458 + ], + "lines": [ + { + "bbox": [ + 390, + 383, + 463, + 458 + ], + "spans": [ + { + "bbox": [ + 390, + 383, + 463, + 458 + ], + "type": "image", + "image_path": "b8f4b69ef02823a76d1457bddcd9fa9cfae7920ed9369078c48777bf8d3a4861.jpg" + } + ] + } + ], + "index": 43, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 402, + 467, + 459, + 479 + ], + "lines": [ + { + "bbox": [ + 402, + 467, + 459, + 479 + ], + "spans": [ + { + "bbox": [ + 402, + 467, + 459, + 479 + ], + "type": "text", + "content": "Target Image" + } + ] + } + ], + "index": 44, + "angle": 0, + "type": "image_caption" + } + ], + "index": 43 + }, + { + "type": "image", + "bbox": [ + 465, + 234, + 538, + 308 + ], + "blocks": [ + { + "bbox": [ + 465, + 234, + 538, + 308 + ], + "lines": [ + { + "bbox": [ + 465, + 234, + 538, + 308 + ], + "spans": [ + { + "bbox": [ + 465, + 234, + 538, + 308 + ], + "type": "image", + "image_path": "788919fb008bd5ac4aa475ecb20d22edeb03f5ac6f180a53bdb647f3db096eba.jpg" + } + ] + } + ], + "index": 45, + "angle": 0, + "type": "image_body" + } + ], + "index": 45 + }, + { + "type": "image", + "bbox": [ + 465, + 308, + 538, + 382 + ], + "blocks": [ + { + "bbox": [ + 465, + 308, + 538, + 382 + ], + "lines": [ + { + "bbox": [ + 465, + 308, + 538, + 382 + ], + "spans": [ + { + "bbox": [ + 465, + 308, + 538, + 382 + ], + "type": "image", + "image_path": "f9933cfd1d5171916f164c753451183f97ca130d1ad5fa1c5eab1465cc77ea67.jpg" + } + ] + } + ], + "index": 46, + "angle": 0, + "type": "image_body" + } + ], + "index": 46 + }, + { + "type": "image", + "bbox": [ + 465, + 383, + 538, + 458 + ], + "blocks": [ + { + "bbox": [ + 465, + 383, + 538, + 458 + ], + "lines": [ + { + "bbox": [ + 465, + 383, + 538, + 458 + ], + "spans": [ + { + "bbox": [ + 465, + 383, + 538, + 458 + ], + "type": "image", + "image_path": "cc1058d78ae81e8c3a819a8878400c262743be1a8b43c95238b75c87350cb4a0.jpg" + } + ] + } + ], + "index": 47, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 475, + 467, + 528, + 478 + ], + "lines": [ + { + "bbox": [ + 475, + 467, + 528, + 478 + ], + "spans": [ + { + "bbox": [ + 475, + 467, + 528, + 478 + ], + "type": "text", + "content": "DreamOmni" + } + ] + } + ], + "index": 48, + "angle": 0, + "type": "image_caption" + } + ], + "index": 47 + }, + { + "bbox": [ + 304, + 533, + 545, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 533, + 545, + 713 + ], + "spans": [ + { + "bbox": [ + 304, + 533, + 545, + 713 + ], + "type": "text", + "content": "of existing models under fair settings and consider the specific needs of various editing tasks. From this analysis, we develop a framework that integrates T2I with various editing tasks. Besides, a challenge in training editing models is the creation of high-quality, large-scale editing data, which is inefficient. To overcome this, we designed a synthetic collage data pipeline capable of efficiently generating quite a few precise, high-quality editing data. Moreover, the pipeline enhances the model's generation accuracy in text, position, quantity, color, and geometry. By jointly training on T2I and multi-task synthetic data, we develop a native, unified model for both image generation and editing. T2I training strengthens the model's grasp of specific concepts and improves generation quality, while editing training enables it to handle the requirements of editing tasks." + } + ] + } + ], + "index": 50 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "text", + "content": "28540" + } + ] + } + ], + "index": 51 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 7 + }, + { + "para_blocks": [ + { + "bbox": [ + 47, + 72, + 289, + 109 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 72, + 289, + 109 + ], + "spans": [ + { + "bbox": [ + 47, + 72, + 289, + 109 + ], + "type": "text", + "content": "Acknowledgements. This work was supported in part by the Research Grants Council under the Areas of Excellence scheme grant AoE/E-601/22-R." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 48, + 119, + 107, + 132 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 119, + 107, + 132 + ], + "spans": [ + { + "bbox": [ + 48, + 119, + 107, + 132 + ], + "type": "text", + "content": "References" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 47, + 139, + 288, + 713 + ], + "type": "list", + "angle": 0, + "index": 15, + "blocks": [ + { + "bbox": [ + 53, + 139, + 287, + 172 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 139, + 287, + 172 + ], + "spans": [ + { + "bbox": [ + 53, + 139, + 287, + 172 + ], + "type": "text", + "content": "[1] Jacob Austin, Daniel D Johnson, Jonathan Ho, Daniel Tarlow, and Rianne van den Berg. Structured denoising diffusion models in discrete state-spaces. NeurIPS, 2021. 2" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 53, + 173, + 288, + 205 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 173, + 288, + 205 + ], + "spans": [ + { + "bbox": [ + 53, + 173, + 288, + 205 + ], + "type": "text", + "content": "[2] Omri Avrahami, Dani Lischinski, and Ohad Fried. Blended diffusion for text-driven editing of natural images. In CVPR, 2022. 3" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 53, + 208, + 288, + 262 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 208, + 288, + 262 + ], + "spans": [ + { + "bbox": [ + 53, + 208, + 288, + 262 + ], + "type": "text", + "content": "[3] Yogesh Balaji, Seungjun Nah, Xun Huang, Arash Vahdat, Jiaming Song, Qinsheng Zhang, Karsten Kreis, Miika Aittala, Timo Aila, Samuli Laine, et al. ediff-i: Text-to-image diffusion models with an ensemble of expert denoisers. arXiv preprint arXiv:2211.01324, 2022. 3" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 53, + 263, + 287, + 296 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 263, + 287, + 296 + ], + "spans": [ + { + "bbox": [ + 53, + 263, + 287, + 296 + ], + "type": "text", + "content": "[4] Fan Bao, Shen Nie, Kaiwen Xue, Yue Cao, Chongxuan Li, Hang Su, and Jun Zhu. All are worth words: A vit backbone for diffusion models. In CVPR, 2023. 3" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 53, + 297, + 287, + 341 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 297, + 287, + 341 + ], + "spans": [ + { + "bbox": [ + 53, + 297, + 287, + 341 + ], + "type": "text", + "content": "[5] Georgios Batzolis, Jan Stanczuk, Carola-Bibiane Schonlieb, and Christian Etmann. Conditional image generation with score-based diffusion models. arXiv preprint arXiv:2111.13606, 2021. 2" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 53, + 342, + 287, + 397 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 342, + 287, + 397 + ], + "spans": [ + { + "bbox": [ + 53, + 342, + 287, + 397 + ], + "type": "text", + "content": "[6] James Betker, Gabriel Goh, Li Jing, Tim Brooks, Jianfeng Wang, Linjie Li, Long Ouyang, Juntang Zhuang, Joyce Lee, Yufei Guo, et al. Improving image generation with better captions. Computer Science. https://cdn.openai.com/papers/dall-e-3.pdf, 2023. 1, 3" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 53, + 399, + 287, + 453 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 399, + 287, + 453 + ], + "spans": [ + { + "bbox": [ + 53, + 399, + 287, + 453 + ], + "type": "text", + "content": "[7] Manuel Brack, Felix Friedrich, Katharina Kornmeier, Linoy Tsaban, Patrick Schramowski, Kristian Kersting, and Apolinário Passos. *Ledits++: Limitless image editing using text-to-image models.* arXiv preprint arXiv:2311.16711, 2023. 3" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 53, + 455, + 287, + 487 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 455, + 287, + 487 + ], + "spans": [ + { + "bbox": [ + 53, + 455, + 287, + 487 + ], + "type": "text", + "content": "[8] Tim Brooks, Aleksander Holynski, and Alexei A Efros. Instructpix2pix: Learning to follow image editing instructions. In CVPR, 2023. 2, 3, 4, 8" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 53, + 489, + 287, + 533 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 489, + 287, + 533 + ], + "spans": [ + { + "bbox": [ + 53, + 489, + 287, + 533 + ], + "type": "text", + "content": "[9] Mingdeng Cao, Xintao Wang, Zhongang Qi, Ying Shan, Xiaohu Qie, and Yinqiang Zheng. Masactrl: Tuning-free mutual self-attention control for consistent image synthesis and editing. In ICCV, 2023. 3" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 47, + 534, + 287, + 588 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 534, + 287, + 588 + ], + "spans": [ + { + "bbox": [ + 47, + 534, + 287, + 588 + ], + "type": "text", + "content": "[10] Junsong Chen, Jincheng Yu, Chongjian Ge, Lewei Yao, Enze Xie, Yue Wu, Zhongdao Wang, James Kwok, Ping Luo, Huchuan Lu, et al. Pixart-α: Fast training of diffusion transformer for photorealistic text-to-image synthesis. ICLR, 2024. 3, 5, 6" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 47, + 590, + 287, + 645 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 590, + 287, + 645 + ], + "spans": [ + { + "bbox": [ + 47, + 590, + 287, + 645 + ], + "type": "text", + "content": "[11] Zhe Chen, Weiyun Wang, Hao Tian, Shenglong Ye, Zhang-wei Gao, Erfei Cui, Wenwen Tong, Kongzhi Hu, Jiapeng Luo, Zheng Ma, et al. How far are we to gpt-4v? closing the gap to commercial multimodal models with open-source suites. arXiv preprint arXiv:2404.16821, 2024. 5" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 47, + 647, + 287, + 689 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 647, + 287, + 689 + ], + "spans": [ + { + "bbox": [ + 47, + 647, + 287, + 689 + ], + "type": "text", + "content": "[12] Guillaume Couairon, Jakob Verbeek, Holger Schwenk, and Matthieu Cord. Diffedit: Diffusion-based semantic image editing with mask guidance. arXiv preprint arXiv:2210.11427, 2022.3" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 47, + 691, + 287, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 691, + 287, + 713 + ], + "spans": [ + { + "bbox": [ + 47, + 691, + 287, + 713 + ], + "type": "text", + "content": "[13] Xiaoliang Dai, Ji Hou, Chih-Yao Ma, Sam Tsai, Jialiang Wang, Rui Wang, Peizhao Zhang, Simon Vandenhende, Xi" + } + ] + } + ], + "index": 14 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 308, + 73, + 545, + 713 + ], + "type": "list", + "angle": 0, + "index": 32, + "blocks": [ + { + "bbox": [ + 326, + 73, + 545, + 106 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 326, + 73, + 545, + 106 + ], + "spans": [ + { + "bbox": [ + 326, + 73, + 545, + 106 + ], + "type": "text", + "content": "aofang Wang, Abhimanyu Dubey, et al. Emu: Enhancing image generation models using photogenic needles in a haystack. arXiv preprint arXiv:2309.15807, 2023. 3" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 308, + 106, + 545, + 128 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 106, + 545, + 128 + ], + "spans": [ + { + "bbox": [ + 308, + 106, + 545, + 128 + ], + "type": "text", + "content": "[14] Prafulla Dhariwal and Alexander Nichol. Diffusion models beat gans on image synthesis. NeurIPS, 2021. 2" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 308, + 129, + 545, + 182 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 129, + 545, + 182 + ], + "spans": [ + { + "bbox": [ + 308, + 129, + 545, + 182 + ], + "type": "text", + "content": "[15] Patrick Esser, Sumith Kulal, Andreas Blattmann, Rahim Entezari, Jonas Müller, Harry Saini, Yam Levi, Dominik Lorenz, Axel Sauer, Frederic Boessel, et al. Scaling rectified flow transformers for high-resolution image synthesis. In ICML, 2024. 1, 3, 4, 5, 6, 7" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 308, + 184, + 545, + 238 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 184, + 545, + 238 + ], + "spans": [ + { + "bbox": [ + 308, + 184, + 545, + 238 + ], + "type": "text", + "content": "[16] Zhida Feng, Zhenyu Zhang, Xintong Yu, Yewei Fang, Lanxin Li, Xuyi Chen, Yuxiang Lu, Jiaxiang Liu, Weichong Yin, Shikun Feng, et al. Ernie-vilg 2.0: Improving text-to-image diffusion model with knowledge-enhanced mixture-of-denoising-experts. In CVPR, 2023. 3" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 308, + 239, + 545, + 281 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 239, + 545, + 281 + ], + "spans": [ + { + "bbox": [ + 308, + 239, + 545, + 281 + ], + "type": "text", + "content": "[17] Tsu-Jui Fu, Wenze Hu, Xianzhi Du, William Yang Wang, Yinfei Yang, and Zhe Gan. Guiding instruction-based image editing via multimodal large language models. *ICLR*, 2024. 8" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 308, + 283, + 545, + 336 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 283, + 545, + 336 + ], + "spans": [ + { + "bbox": [ + 308, + 283, + 545, + 336 + ], + "type": "text", + "content": "[18] Zigang Geng, Binxin Yang, Tiankai Hang, Chen Li, Shuyang Gu, Ting Zhang, Jianmin Bao, Zheng Zhang, Han Hu, Dong Chen, et al. Instructdiffusion: A generalist modeling interface for vision tasks. arXiv preprint arXiv:2309.03895, 2023. 3" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 308, + 338, + 545, + 370 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 338, + 545, + 370 + ], + "spans": [ + { + "bbox": [ + 308, + 338, + 545, + 370 + ], + "type": "text", + "content": "[19] Dhruba Ghosh, Hannaneh Hajishirzi, and Ludwig Schmidt. Geneval: An object-focused framework for evaluating text-to-image alignment. NeurIPS, 36, 2024. 5, 7" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 308, + 371, + 545, + 413 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 371, + 545, + 413 + ], + "spans": [ + { + "bbox": [ + 308, + 371, + 545, + 413 + ], + "type": "text", + "content": "[20] Ian Goodfellow, Jean Pouget-Abadie, Mehdi Mirza, Bing Xu, David Warde-Farley, Sherjil Ozair, Aaron Courville, and Yoshua Bengio. Generative adversarial nets. NeurIPS, 2014. 2" + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 308, + 415, + 545, + 458 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 415, + 545, + 458 + ], + "spans": [ + { + "bbox": [ + 308, + 415, + 545, + 458 + ], + "type": "text", + "content": "[21] Shuyang Gu, Dong Chen, Jianmin Bao, Fang Wen, Bo Zhang, Dongdong Chen, Lu Yuan, and Baining Guo. Vector quantized diffusion model for text-to-image synthesis. In CVPR, 2022. 2" + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 308, + 460, + 545, + 502 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 460, + 545, + 502 + ], + "spans": [ + { + "bbox": [ + 308, + 460, + 545, + 502 + ], + "type": "text", + "content": "[22] Amir Hertz, Ron Mokady, Jay Tenenbaum, Kfir Aberman, Yael Pritch, and Daniel Cohen-Or. Prompt-to-prompt image editing with cross attention control. arXiv preprint arXiv:2208.01626, 2022. 3" + } + ] + } + ], + "index": 25 + }, + { + "bbox": [ + 308, + 504, + 545, + 525 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 504, + 545, + 525 + ], + "spans": [ + { + "bbox": [ + 308, + 504, + 545, + 525 + ], + "type": "text", + "content": "[23] Jonathan Ho, Ajay Jain, and Pieter Abbeel. Denoising diffusion probabilistic models. NeurIPS, 2020. 2" + } + ] + } + ], + "index": 26 + }, + { + "bbox": [ + 308, + 526, + 545, + 559 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 526, + 545, + 559 + ], + "spans": [ + { + "bbox": [ + 308, + 526, + 545, + 559 + ], + "type": "text", + "content": "[24] Jonathan Ho, Chitwan Sahara, William Chan, David J Fleet, Mohammad Norouzi, and Tim Salimans. Cascaded diffusion models for high fidelity image generation. JMLR, 2022. 2" + } + ] + } + ], + "index": 27 + }, + { + "bbox": [ + 308, + 559, + 545, + 613 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 559, + 545, + 613 + ], + "spans": [ + { + "bbox": [ + 308, + 559, + 545, + 613 + ], + "type": "text", + "content": "[25] Yuzhou Huang, Liangbin Xie, Xintao Wang, Ziyang Yuan, Xiaodong Cun, Yixiao Ge, Jiantao Zhou, Chao Dong, Rui Huang, Ruimao Zhang, et al. Smartedit: Exploring complex instruction-based image editing with multimodal large language models. CVPR, 2024. 3" + } + ] + } + ], + "index": 28 + }, + { + "bbox": [ + 308, + 614, + 545, + 635 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 614, + 545, + 635 + ], + "spans": [ + { + "bbox": [ + 308, + 614, + 545, + 635 + ], + "type": "text", + "content": "[26] Diederik P Kingma. Auto-encoding variational bayes. arXiv preprint arXiv:1312.6114, 2013. 2" + } + ] + } + ], + "index": 29 + }, + { + "bbox": [ + 308, + 636, + 545, + 680 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 636, + 545, + 680 + ], + "spans": [ + { + "bbox": [ + 308, + 636, + 545, + 680 + ], + "type": "text", + "content": "[27] Daiqing Li, Aleks Kamko, Ehsan Akhgari, Ali Sabet, Linmiao Xu, and Suhail Doshi. Playground v2. 5: Three insights towards enhancing aesthetic quality in text-to-image generation. arXiv preprint arXiv:2402.17245, 2024. 1" + } + ] + } + ], + "index": 30 + }, + { + "bbox": [ + 308, + 681, + 545, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 681, + 545, + 713 + ], + "spans": [ + { + "bbox": [ + 308, + 681, + 545, + 713 + ], + "type": "text", + "content": "[28] Dongxu Li, Junnan Li, and Steven Hoi. Blip-diffusion: Pretrained subject representation for controllable text-to-image generation and editing. *NeurIPS*, 2024. 3, 4, 7" + } + ] + } + ], + "index": 31 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "text", + "content": "28541" + } + ] + } + ], + "index": 33 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 8 + }, + { + "para_blocks": [ + { + "bbox": [ + 48, + 72, + 287, + 713 + ], + "type": "list", + "angle": 0, + "index": 14, + "blocks": [ + { + "bbox": [ + 48, + 72, + 287, + 116 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 72, + 287, + 116 + ], + "spans": [ + { + "bbox": [ + 48, + 72, + 287, + 116 + ], + "type": "text", + "content": "[29] Ming Li, Taojiannan Yang, Huafeng Kuang, Jie Wu, Zhaoning Wang, Xuefeng Xiao, and Chen Chen. Controlnet ++: Improving conditional controls with efficient consistency feedback. In ECCV, 2025.3" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 48, + 118, + 287, + 161 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 118, + 287, + 161 + ], + "spans": [ + { + "bbox": [ + 48, + 118, + 287, + 161 + ], + "type": "text", + "content": "[30] Zhen Li, Mingdeng Cao, Xintao Wang, Zhongang Qi, MingMing Cheng, and Ying Shan. Photomaker: Customizing realistic human photos via stacked id embedding. In CVPR, 2024. 3" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 48, + 163, + 287, + 228 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 163, + 287, + 228 + ], + "spans": [ + { + "bbox": [ + 48, + 163, + 287, + 228 + ], + "type": "text", + "content": "[31] Zhimin Li, Jianwei Zhang, Qin Lin, Jiangfeng Xiong, Yanxin Long, Xinchi Deng, Yingfang Zhang, Xingchao Liu, Minbin Huang, Zedong Xiao, et al. Hunyuan-dit: A powerful multi-resolution diffusion transformer with fine-grained chinese understanding. arXiv preprint arXiv:2405.08748, 2024. 3" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 48, + 230, + 287, + 262 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 230, + 287, + 262 + ], + "spans": [ + { + "bbox": [ + 48, + 230, + 287, + 262 + ], + "type": "text", + "content": "[32] Pengyang Ling, Lin Chen, Pan Zhang, Huaian Chen, Yi Jin, and Jinjin Zheng. Freedrag: Feature dragging for reliable point-based image editing. In CVPR, 2024. 3" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 48, + 264, + 287, + 297 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 264, + 287, + 297 + ], + "spans": [ + { + "bbox": [ + 48, + 264, + 287, + 297 + ], + "type": "text", + "content": "[33] Xingchao Liu, Chengyue Gong, and Qiang Liu. Flow straight and fast: Learning to generate and transfer data with rectified flow. arXiv preprint arXiv:2209.03003, 2022. 5" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 48, + 298, + 287, + 352 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 298, + 287, + 352 + ], + "spans": [ + { + "bbox": [ + 48, + 298, + 287, + 352 + ], + "type": "text", + "content": "[34] Hayk Manukyan, Andranik Sargsyan, Barsegh Atanyan, Zhangyang Wang, Shant Navasardyan, and Humphrey Shi. Hd-painter: high-resolution and prompt-faithful text-guided image inpainting with diffusion models. arXiv preprint arXiv:2312.14091, 2023. 3" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 48, + 354, + 287, + 408 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 354, + 287, + 408 + ], + "spans": [ + { + "bbox": [ + 48, + 354, + 287, + 408 + ], + "type": "text", + "content": "[35] Alex Nichol, Prafulla Dhariwal, Aditya Ramesh, Pranav Shyam, Pamela Mishkin, Bob McGrew, Ilya Sutskever, and Mark Chen. Glide: Towards photorealistic image generation and editing with text-guided diffusion models. arXiv preprint arXiv:2112.10741, 2021. 2" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 48, + 411, + 287, + 453 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 411, + 287, + 453 + ], + "spans": [ + { + "bbox": [ + 48, + 411, + 287, + 453 + ], + "type": "text", + "content": "[36] Xichen Pan, Li Dong, Shaohan Huang, Zhiliang Peng, Wenhu Chen, and Furu Wei. Kosmos-g: Generating images in context with multimodal large language models. arXiv preprint arXiv:2310.02992, 2023. 3" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 48, + 455, + 287, + 487 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 455, + 287, + 487 + ], + "spans": [ + { + "bbox": [ + 48, + 455, + 287, + 487 + ], + "type": "text", + "content": "[37] Gaurav Parmar, Krishna Kumar Singh, Richard Zhang, Yijun Li, Jingwan Lu, and Jun-Yan Zhu. Zero-shot image-to-image translation. In ACM SIGGRAPH, 2023. 3" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 48, + 489, + 287, + 510 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 489, + 287, + 510 + ], + "spans": [ + { + "bbox": [ + 48, + 489, + 287, + 510 + ], + "type": "text", + "content": "[38] William Peebles and Saining Xie. Scalable diffusion models with transformers. In ICCV, 2023. 2, 3, 4" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 48, + 513, + 287, + 566 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 513, + 287, + 566 + ], + "spans": [ + { + "bbox": [ + 48, + 513, + 287, + 566 + ], + "type": "text", + "content": "[39] Dustin Podell, Zion English, Kyle Lacey, Andreas Blattmann, Tim Dockhorn, Jonas Müller, Joe Penna, and Robin Rombach. Sdxl: Improving latent diffusion models for high-resolution image synthesis. NeurIPS, 2023. 1, 3, 4, 5, 6" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 48, + 568, + 287, + 622 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 568, + 287, + 622 + ], + "spans": [ + { + "bbox": [ + 48, + 568, + 287, + 622 + ], + "type": "text", + "content": "[40] Can Qin, Shu Zhang, Ning Yu, Yihao Feng, Xinyi Yang, Yingbo Zhou, Huan Wang, Juan Carlos Niebles, Caiming Xiong, Silvio Savarese, et al. Unicontrol: A unified diffusion model for controllable visual generation in the wild. arXiv preprint arXiv:2305.11147, 2023. 3" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 48, + 624, + 287, + 677 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 624, + 287, + 677 + ], + "spans": [ + { + "bbox": [ + 48, + 624, + 287, + 677 + ], + "type": "text", + "content": "[41] Alec Radford, Jong Wook Kim, Chris Hallacy, Aditya Ramesh, Gabriel Goh, Sandhini Agarwal, Girish Sastry, Amanda Askell, Pamela Mishkin, Jack Clark, et al. Learning transferable visual models from natural language supervision. In ICML, 2021. 4" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 48, + 680, + 287, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 680, + 287, + 713 + ], + "spans": [ + { + "bbox": [ + 48, + 680, + 287, + 713 + ], + "type": "text", + "content": "[42] Colin Raffel, Noam Shazeer, Adam Roberts, Katherine Lee, Sharan Narang, Michael Matena, Yanqi Zhou, Wei Li, and Peter J Liu. Exploring the limits of transfer learning with a" + } + ] + } + ], + "index": 13 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 307, + 73, + 545, + 712 + ], + "type": "list", + "angle": 0, + "index": 30, + "blocks": [ + { + "bbox": [ + 327, + 73, + 545, + 95 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 327, + 73, + 545, + 95 + ], + "spans": [ + { + "bbox": [ + 327, + 73, + 545, + 95 + ], + "type": "text", + "content": "unified text-to-text transformer. Journal of machine learning research, 2020. 4" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 308, + 97, + 545, + 129 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 97, + 545, + 129 + ], + "spans": [ + { + "bbox": [ + 308, + 97, + 545, + 129 + ], + "type": "text", + "content": "[43] Aditya Ramesh, Mikhail Pavlov, Gabriel Goh, Scott Gray, Chelsea Voss, Alec Radford, Mark Chen, and Ilya Sutskever. Zero-shot text-to-image generation. In ICML, 2021. 3" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 307, + 132, + 545, + 175 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 132, + 545, + 175 + ], + "spans": [ + { + "bbox": [ + 307, + 132, + 545, + 175 + ], + "type": "text", + "content": "[44] Robin Rombach, Andreas Blattmann, Dominik Lorenz, Patrick Esser, and Björn Ommer. High-resolution image synthesis with latent diffusion models. In CVPR, 2022. 1, 2, 4, 5, 6, 7, 8" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 308, + 177, + 545, + 209 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 177, + 545, + 209 + ], + "spans": [ + { + "bbox": [ + 308, + 177, + 545, + 209 + ], + "type": "text", + "content": "[45] Olaf Ronneberger, Philipp Fischer, and Thomas Brox. U-net: Convolutional networks for biomedical image segmentation. In MICCAI, 2015. 3" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 308, + 212, + 545, + 255 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 212, + 545, + 255 + ], + "spans": [ + { + "bbox": [ + 308, + 212, + 545, + 255 + ], + "type": "text", + "content": "[46] Nataniel Ruiz, Yuanzhen Li, Varun Jampani, Yael Pritch, Michael Rubinstein, and Kfir Aberman. Dreambooth: Fine tuning text-to-image diffusion models for subject-driven generation. In CVPR, 2023. 3" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 308, + 258, + 545, + 312 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 258, + 545, + 312 + ], + "spans": [ + { + "bbox": [ + 308, + 258, + 545, + 312 + ], + "type": "text", + "content": "[47] Chitwan Sahara, William Chan, Saurabh Saxena, Lala Li, Jay Whang, Emily L Denton, Kamyar Ghasemipour, Raphael Gontijo Lopes, Burcu Karagol Ayan, Tim Salimans, et al. Photorealistic text-to-image diffusion models with deep language understanding. NeurIPS, 2022. 1, 3" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 308, + 315, + 545, + 369 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 315, + 545, + 369 + ], + "spans": [ + { + "bbox": [ + 308, + 315, + 545, + 369 + ], + "type": "text", + "content": "[48] Christoph Schuhmann, Romain Beaumont, Richard Vencu, Cade Gordon, Ross Wightman, Mehdi Cherti, Theo Coombes, Aarush Katta, Clayton Mullis, Mitchell Wortsman, et al. Laion-5b: An open large-scale dataset for training next generation image-text models. NeurIPS, 2022. 5" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 308, + 371, + 545, + 415 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 371, + 545, + 415 + ], + "spans": [ + { + "bbox": [ + 308, + 371, + 545, + 415 + ], + "type": "text", + "content": "[49] Shelly Sheynin, Adam Polyak, Uriel Singer, Yuval Kirstain, Amit Zohar, Oron Ashual, Devi Parikh, and Yaniv Taigman. Emu edit: Precise image editing via recognition and generation tasks. arXiv preprint arXiv:2311.10089, 2023. 3" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 308, + 417, + 545, + 459 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 417, + 545, + 459 + ], + "spans": [ + { + "bbox": [ + 308, + 417, + 545, + 459 + ], + "type": "text", + "content": "[50] Yujun Shi, Jun Hao Liew, Hanshu Yan, Vincent YF Tan, and Jiashi Feng. Instadrag: Lightning fast and accurate drag-based image editing emerging from videos. arXiv preprint arXiv:2405.13722, 2024. 3, 5" + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 308, + 462, + 545, + 506 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 462, + 545, + 506 + ], + "spans": [ + { + "bbox": [ + 308, + 462, + 545, + 506 + ], + "type": "text", + "content": "[51] Yujun Shi, Chuhui Xue, Jun Hao Liew, Jiachun Pan, Hanshu Yan, Wenqing Zhang, Vincent YF Tan, and Song Bai. Dragdiffusion: Harnessing diffusion models for interactive point-based image editing. In CVPR, 2024. 3" + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 308, + 508, + 545, + 541 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 508, + 545, + 541 + ], + "spans": [ + { + "bbox": [ + 308, + 508, + 545, + 541 + ], + "type": "text", + "content": "[52] Jascha Sohl-Dickstein, Eric Weiss, Niru Maheswaranathan, and Surya Ganguli. Deep unsupervised learning using nonequilibrium thermodynamics. In ICML, 2015. 2" + } + ] + } + ], + "index": 25 + }, + { + "bbox": [ + 308, + 543, + 545, + 586 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 543, + 545, + 586 + ], + "spans": [ + { + "bbox": [ + 308, + 543, + 545, + 586 + ], + "type": "text", + "content": "[53] Yang Song, Jascha Sohl-Dickstein, Diederik P Kingma, Abhishek Kumar, Stefano Ermon, and Ben Poole. Score-based generative modeling through stochastic differential equations. arXiv preprint arXiv:2011.13456, 2020. 2" + } + ] + } + ], + "index": 26 + }, + { + "bbox": [ + 308, + 589, + 545, + 621 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 589, + 545, + 621 + ], + "spans": [ + { + "bbox": [ + 308, + 589, + 545, + 621 + ], + "type": "text", + "content": "[54] Narek Tumanyan, Michal Geyer, Shai Bagon, and Tali Dekel. Plug-and-play diffusion features for text-driven image-to-image translation. In CVPR, 2023. 3" + } + ] + } + ], + "index": 27 + }, + { + "bbox": [ + 308, + 624, + 545, + 655 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 624, + 545, + 655 + ], + "spans": [ + { + "bbox": [ + 308, + 624, + 545, + 655 + ], + "type": "text", + "content": "[55] Bram Wallace, Akash Gokul, and Nikhil Naik. Edict: Exact diffusion inversion via coupled transformations. In CVPR, 2023. 3" + } + ] + } + ], + "index": 28 + }, + { + "bbox": [ + 308, + 658, + 545, + 712 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 658, + 545, + 712 + ], + "spans": [ + { + "bbox": [ + 308, + 658, + 545, + 712 + ], + "type": "text", + "content": "[56] Peng Wang, Shuai Bai, Sinan Tan, Shijie Wang, Zhihao Fan, Jinze Bai, Keqin Chen, Xuejing Liu, Jialin Wang, Wenbin Ge, et al. Qwen2-vl: Enhancing vision-language model's perception of the world at any resolution. arXiv preprint arXiv:2409.12191, 2024. 5" + } + ] + } + ], + "index": 29 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 749, + 318, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 749, + 318, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 749, + 318, + 757 + ], + "type": "text", + "content": "28542" + } + ] + } + ], + "index": 31 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 9 + }, + { + "para_blocks": [ + { + "bbox": [ + 48, + 72, + 287, + 352 + ], + "type": "list", + "angle": 0, + "index": 7, + "blocks": [ + { + "bbox": [ + 48, + 72, + 287, + 115 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 72, + 287, + 115 + ], + "spans": [ + { + "bbox": [ + 48, + 72, + 287, + 115 + ], + "type": "text", + "content": "[57] Bin Xia, Shi Yin Wang, Yingfan Tao, Yitong Wang, and Jiaya Jia. Llmga: Multimodal large language model based generation assistant. In European Conference on Computer Vision, 2025. 3" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 49, + 118, + 287, + 162 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 49, + 118, + 287, + 162 + ], + "spans": [ + { + "bbox": [ + 49, + 118, + 287, + 162 + ], + "type": "text", + "content": "[58] Shitao Xiao, Yueze Wang, Junjie Zhou, Huaying Yuan, Xingrun Xing, Ruiran Yan, Shuting Wang, Tiejun Huang, and Zheng Liu. Omnigen: Unified image generation. arXiv preprint arXiv:2409.11340, 2024. 3" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 49, + 163, + 287, + 195 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 49, + 163, + 287, + 195 + ], + "spans": [ + { + "bbox": [ + 49, + 163, + 287, + 195 + ], + "type": "text", + "content": "[59] Shaoan Xie, Zhifei Zhang, Zhe Lin, Tobias Hinz, and Kun Zhang. Smartbrush: Text and shape guided object inpainting with diffusion model. In CVPR, 2023. 3" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 49, + 197, + 287, + 239 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 49, + 197, + 287, + 239 + ], + "spans": [ + { + "bbox": [ + 49, + 197, + 287, + 239 + ], + "type": "text", + "content": "[60] Hu Ye, Jun Zhang, Sibo Liu, Xiao Han, and Wei Yang. Ip-adapter: Text compatible image prompt adapter for text-to-image diffusion models. arXiv preprint arXiv:2308.06721, 2023. 2, 3, 4, 7" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 49, + 241, + 286, + 274 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 49, + 241, + 286, + 274 + ], + "spans": [ + { + "bbox": [ + 49, + 241, + 286, + 274 + ], + "type": "text", + "content": "[61] Kai Zhang, Lingbo Mo, Wenhu Chen, Huan Sun, and Yu Su. Magicbrush: A manually annotated dataset for instruction-guided image editing. NeurIPS, 2024. 3" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 49, + 276, + 286, + 308 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 49, + 276, + 286, + 308 + ], + "spans": [ + { + "bbox": [ + 49, + 276, + 286, + 308 + ], + "type": "text", + "content": "[62] Lvmin Zhang, Anyi Rao, and Maneesh Agrawala. Adding conditional control to text-to-image diffusion models. In ICCV, 2023. 2, 3, 5, 6, 7, 8" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 49, + 309, + 286, + 352 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 49, + 309, + 286, + 352 + ], + "spans": [ + { + "bbox": [ + 49, + 309, + 286, + 352 + ], + "type": "text", + "content": "[63] Shihao Zhao, Dongdong Chen, Yen-Chun Chen, Jianmin Bao, Shaozhe Hao, Lu Yuan, and Kwan-Yee K. Wong. Uni-controlnet: All-in-one control to text-to-image diffusion models. NeurIPS, 2023. 3" + } + ] + } + ], + "index": 6 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 749, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 749, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 749, + 317, + 757 + ], + "type": "text", + "content": "28543" + } + ] + } + ], + "index": 8 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 10 + } + ], + "_backend": "vlm", + "_version_name": "2.6.4" +} \ No newline at end of file diff --git a/2025/DreamRelation_ Bridging Customization and Relation Generation/dc0bd82f-ae71-4d19-a1ab-522a355a226a_content_list.json b/2025/DreamRelation_ Bridging Customization and Relation Generation/dc0bd82f-ae71-4d19-a1ab-522a355a226a_content_list.json new file mode 100644 index 0000000000000000000000000000000000000000..8274b8d4a61faad9eeae3a15c9c9dc8bc7c15698 --- /dev/null +++ b/2025/DreamRelation_ Bridging Customization and Relation Generation/dc0bd82f-ae71-4d19-a1ab-522a355a226a_content_list.json @@ -0,0 +1,1737 @@ +[ + { + "type": "text", + "text": "DreamRelation: Bridging Customization and Relation Generation", + "text_level": 1, + "bbox": [ + 165, + 130, + 831, + 152 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Qingyu Shi $^{1}$ Lu Qi $^{3*}$ Jianzong Wu $^{1}$ Jinbin Bai $^{5}$ Jingbo Wang $^{4}$ Yunhai Tong $^{1}$ Xiangtai Li $^{2,4*}$ $^{1}$ PKU NTU Insta360 Research Shanghai AI Laboratory NUS \nProject page: https://shi-qingyu.github.io/DreamRelation.github.io", + "bbox": [ + 91, + 178, + 928, + 233 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/3cc4df9f5289bb5b2e581c9f6838dcc383b548bd087fecfac7349bbcafc80669.jpg", + "image_caption": [ + "Figure 1. In our Relation-Aware Image Customization task, the generated images must accurately preserve the relationships between objects and maintain their identity. We highlight the limitations of previous approaches using three color codes: red indicates failure to capture relationships, blue marks missing objects, and orange represents object confusion. Each image is annotated to reflect its specific issue. Our results, highlighted by green boxes, demonstrate the advantages of our proposed method." + ], + "image_footnote": [], + "bbox": [ + 99, + 239, + 893, + 554 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Abstract", + "text_level": 1, + "bbox": [ + 246, + 646, + 326, + 662 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Customized image generation is essential for creating personalized content based on user prompts, allowing large-scale text-to-image diffusion models to more effectively meet individual needs. However, existing models often neglect the relationships between customized objects in generated images. In contrast, this work addresses this gap by focusing on relation-aware customized image generation, which seeks to preserve the identities from image prompts while maintaining the relationship specified in text prompts. Specifically, we introduce DreamRelation, a framework that disentangles identity and relation learning using a carefully curated dataset. Our training data consists of relation-specific images, independent object images containing identity information, and text prompts to guide relation genera", + "bbox": [ + 88, + 689, + 485, + 901 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "tion. Then, we propose two key modules to tackle the two main challenges—generating accurate and natural relationships, especially when significant pose adjustments are required, and avoiding object confusion in cases of overlap. First, we introduce a keypoint matching loss that effectively guides the model in adjusting object poses closely tied to their relationships. Second, we incorporate local features of the image prompts to better distinguish between objects, preventing confusion in overlapping cases. Extensive results on our proposed benchmarks demonstrate the superiority of DreamRelation in generating precise relations while preserving object identities across a diverse set of objects and relationships.", + "bbox": [ + 511, + 648, + 908, + 845 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "CVF", + "bbox": [ + 106, + 2, + 181, + 42 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore.", + "bbox": [ + 236, + 0, + 808, + 46 + ], + "page_idx": 0 + }, + { + "type": "page_footnote", + "text": "$^{1}$ Corresponding authors: Lu Qi, Xiangtai Li.", + "bbox": [ + 531, + 886, + 779, + 901 + ], + "page_idx": 0 + }, + { + "type": "page_number", + "text": "15723", + "bbox": [ + 480, + 944, + 517, + 955 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "1. Introduction", + "text_level": 1, + "bbox": [ + 91, + 89, + 222, + 106 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Driven by large-scale text-to-image diffusion models [26, 27, 30, 32], customized image generation has recently made significant strides [1, 14, 15, 31, 35, 36, 43]. This task focuses on generating images that preserve the identity of objects from user-provided inputs, enabling the creation of personalized and meaningful content. It has shown value in numerous applications, including personalized artwork, branding, virtual fashion try-ons, social media content creation, augmented reality experiences, and more.", + "bbox": [ + 89, + 114, + 480, + 251 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Despite the success of many methods for customizing single or multiple objects [8, 16, 18, 20, 23, 24, 35, 42], they often overlook the relationships between objects and the corresponding text prompts. For instance, when two user-provided objects are paired with a text prompt specifying a particular relationship, the generated output should not only preserve their identities but also accurately reflect the intended relationship, such as a 'hug'. This introduces new challenges in what we refer to as relation-aware customized image generation, which focuses on preserving multiple identities while adhering to relationship prompts.", + "bbox": [ + 89, + 251, + 482, + 417 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "An intuitive solution to this issue is to adapt existing relation-aware generation methods [11, 21, 22] to tuning-based or training-based customized methods for customizing objects while maintaining the relation. However, both approaches face challenges. Tuning-based methods struggle to preserve multiple identities, as they invert objects into specific tokens. While training-based schemes often fail to balance image and text prompts, frequently overlooking key textual elements and hindering the generation of relationships between objects. As illustrated in Fig. 1, previous methods fail to capture the relationships described in the text prompt and lose identity preservation.", + "bbox": [ + 89, + 417, + 482, + 598 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "We attribute this failure to two key factors: a lack of relevant data and an ineffective model design. Unlike data augmentation techniques such as flipping or rotation, commonly used to create paired training data in object customization methods [5, 40], our approach requires a triplet of images: two image prompts and one target image. The image prompts should contain similar objects but exhibit distinct poses compared to the target image. To collect these triplets, we propose a data engine to curate our finetuning set. We leverage an advanced text-to-image generation model [2] to generate triplets where the same object pair is shared across the images. Through text prompt guidance, the image prompt provides strong identity information. This enables the decoupling of the relationship in the target image, which enhances the relation learning.", + "bbox": [ + 89, + 598, + 482, + 824 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "For the model design, we propose DreamRelation, which applies the Low-Rank Adaptation (LoRA) [9] strategy to the text cross-attention layers of existing diffusion models to process user-provided text prompts. In DreamRelation, two key modules are introduced during training to", + "bbox": [ + 89, + 825, + 482, + 900 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "enhance the relation generation in customized generation. First, we introduce a keypoint matching loss (KML) as additional supervision to explicitly encourage the model to manipulate object poses, since relationships between objects are closely tied to their poses. Importantly, the KML operates on the latent representation rather than the original image space, aligning with the default diffusion loss. Second, since relation-aware customization requires local features from image prompts, such as the \"hands\" features for generating \"shaking hands\"—which are not captured by CLIP's coarse image-level features, we introduce dense features from CLIP. Through partitioning and pooling, we obtain local tokens that contain detail and local information. To further enhance the compatibility between dense features and image-level features, we employ a self-distillation method [39] to improve their alignment.", + "bbox": [ + 511, + 90, + 903, + 332 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "To more comprehensively evaluate relation-aware customized image generation, we constructed our Relation-Bench based on three established benchmarks [15, 31]. Our methods demonstrate strong performance compared to existing approaches, achieving significant improvements in visual quality and quantitative results.", + "bbox": [ + 511, + 333, + 903, + 422 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "The contributions of this work are:", + "bbox": [ + 532, + 422, + 764, + 436 + ], + "page_idx": 1 + }, + { + "type": "list", + "sub_type": "text", + "list_items": [ + "- We explore a novel task called relation-aware customized generation, which aims to preserve multiple identities from image prompts while adhering to the relationships specified in text prompts. This task can enhance various user-driven applications by enabling more control.", + "- We introduce a data engine that uses the recent text-to-image generation model to generate triplet images where the same object is present with distinct poses. This well-curated dataset enables the model to focus on relation learning during fine-tuning, leveraging the identity information provided by the image prompts.", + "- Our proposed DreamRelation has two key modules, including keypoint matching loss and local token injection, to enhance the relation generation in customized generation. The extensive experiments on three benchmarks demonstrate the effectiveness of our method." + ], + "bbox": [ + 513, + 439, + 903, + 679 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "2. Related Work", + "text_level": 1, + "bbox": [ + 513, + 693, + 653, + 709 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Diffusion-Based Text-to-Image Generation. Diffusion-based text-to-image models [30, 32] generate high-quality images based on user-provided text prompts. These works encode the text prompt through the text encoder [6, 28] and inject the text embedding into the U-Net's cross-attention layers. Some methods [26, 27, 29, 38] upscale diffusion models and incorporate additional conditions as priors to generate high-resolution images. Meanwhile, several methods [2, 17] introduce stronger text encoders or large language models (LLMs) to enhance the text comprehension capabilities of diffusion models. Recent works [4, 25] have replaced the U-Net denoiser with the Transformers [33].", + "bbox": [ + 511, + 719, + 903, + 900 + ], + "page_idx": 1 + }, + { + "type": "page_number", + "text": "15724", + "bbox": [ + 480, + 944, + 519, + 955 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "However, these models primarily focus on text conditions and struggle to handle other forms of guidance, such as user-provided images. In contrast, our work facilitates customized generation while following relation conditions given by text inputs.", + "bbox": [ + 89, + 90, + 480, + 167 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Diffusion-Based Customized Generation. Customized image generation aims to produce new images based on user-provided objects. Tuning-based methods accomplish this by fine-tuning specific parameters of diffusion models to learn new objects. Some of these methods [7, 20, 34] employ text embeddings to represent the object identity. On the other hand, DreamBooth [31] fine-tunes the entire U-Net and introduces a prior preservation loss to mitigate language drift. Additionally, several approaches [8, 15, 18, 23] integrate text embeddings and U-Net parameters for image customization. These methods fine-tune both components during object learning, achieving impressive results. Considering the substantial cost of fine-tuning for commercial uses, training-based methods [5, 16, 24, 36, 43] have been proposed. This approach primarily utilizes an encoder to extract the object's identity and inject it into the U-Net. For instance, MS-Diffusion [35] integrates grounding tokens with a Resampler to maintain detailed identity and employs layout guidance to explicitly locate the objects. Despite their efficiency during inference, training-based methods often face challenges in balancing identity preservation and text control. A common issue is that the relationships described in text prompts are frequently overlooked. To address this limitation, we introduce DreamRelation, a framework designed to bridge the customization and relation generation.", + "bbox": [ + 91, + 186, + 483, + 580 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Relation-Aware Text-to-Image Generation. Inspired by tuning-based customization methods [7], some recent works [11, 22, 37] represent a \"neglected word\" by learnable parameters. These methods fine-tune part of the parameters on the content co-existing images. For instance, Reversion [11] fine-tunes the text embedding on a set of relation co-existing images and introduces a relation-steering contrastive loss. Reversion gains a better alignment between relational words and generated images by injecting new text embedding into prompts. While effective with prepositions and adjectives, Reversion struggles with relationships that involve significant overlaps, and cannot customize user-provided objects. Moreover, this method still struggles to generate vivid relationships while maintaining the fidelity of multiple customized objects, likely due to the inherent disregard for text embeddings in customization methods. Our work bridges the gap in the predicate relation-following capability of training-based customization methods. This enables efficient inference that naturally and accurately generates relationships in the text prompts.", + "bbox": [ + 89, + 598, + 483, + 901 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3. Bridging Customization and Relation Generation", + "text_level": 1, + "bbox": [ + 513, + 89, + 903, + 122 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3.1. Problem definition", + "text_level": 1, + "bbox": [ + 513, + 131, + 692, + 146 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Different from conventional customization tasks, we explore a new setting that focuses on image generation by both image prompts $c_{i} \\in \\mathcal{R}^{N \\times 3 \\times H \\times W}$ and text prompts $c_{t}$ . We call this setting relation-aware customized image generation due to the requirements that the generated image $\\hat{x}$ should strictly preserve each identity and keep the relationship among those identities provided by $c_{i}$ and $c_{t}$ .", + "bbox": [ + 511, + 154, + 903, + 258 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "We define this task as:", + "bbox": [ + 532, + 260, + 683, + 273 + ], + "page_idx": 2 + }, + { + "type": "equation", + "text": "\n$$\n\\hat {x} = \\Phi_ {\\theta} \\left(c _ {i}, c _ {t}\\right) \\tag {1}\n$$\n", + "text_format": "latex", + "bbox": [ + 658, + 284, + 903, + 299 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "where $\\Phi_{\\theta}$ is the network parameterized by $\\theta$ . For brevity clarification, we set $N = 2$ due to the basic triplet element defined in the DreamRelation.", + "bbox": [ + 511, + 308, + 903, + 354 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3.2. Discussion and Motivation", + "text_level": 1, + "bbox": [ + 513, + 361, + 754, + 376 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "While previous works [8, 15, 43] have explored the customization of multiple objects under text control, they lack the ability to effectively manage the relationships between these objects. Specifically, the state-of-the-art multi-object customization model [35] utilizes bounding boxes to enhance multi-object generation. Given the image prompt $c_{i}$ and the text prompt $c_{t}$ , it leverages CLIP [28] and the Resampler [35] to extract text tokens $tok_{\\text{text}}$ and image tokens $tok_{\\text{image}}$ , respectively. These tokens are then injected into the U-Net $\\epsilon_{\\theta}$ through parallel cross-attention layers:", + "bbox": [ + 511, + 383, + 906, + 535 + ], + "page_idx": 2 + }, + { + "type": "equation", + "text": "\n$$\nh = \\operatorname {S o f t m a x} \\left(\\frac {Q K _ {i}}{\\sqrt {d}} + \\mathcal {M}\\right) V _ {i} + \\operatorname {S o f t m a x} \\left(\\frac {Q K _ {t}}{\\sqrt {d}}\\right) V _ {t} \\tag {2}\n$$\n", + "text_format": "latex", + "bbox": [ + 531, + 541, + 903, + 574 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "where $Q$ represents the query, while $K_{i}, V_{i}$ and $K_{t}, V_{t}$ denote the keys and values obtained from $tok_{image}$ and $tok_{text}$ , respectively. The mask $\\mathcal{M}$ , generated from the bounding boxes, assists the model in localizing objects during multi-object customization. However, it encounters difficulties when bounding boxes overlap, frequently resulting in object confusion, as shown in Fig. 1. Consequently, relation-aware generation in customized models remains a largely unexplored area of research.", + "bbox": [ + 511, + 580, + 905, + 715 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "On the other hand, we revisit the relation-aware generation task. Existing relation inversion methods [11, 22] focus on inverting a relationship into a text embedding $R^{*}$ within pre-trained text-to-image diffusion models. Given a set of images $\\{x_{k}\\}_{k = 1}^{n}$ that share the same relationship, these methods use a denoising loss $\\mathcal{L}_{RI}$ to fine-tune $R^{*}$ , ensuring alignment with the specific relationship:", + "bbox": [ + 511, + 715, + 905, + 821 + ], + "page_idx": 2 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {L} _ {R I} = \\mathbb {E} _ {z _ {t}, t, \\epsilon , c} [ \\| \\epsilon_ {t} - \\boldsymbol {\\epsilon} _ {\\boldsymbol {\\theta}} (z _ {t}, t, c _ {t}) \\| _ {2} ^ {2} ] \\tag {3}\n$$\n", + "text_format": "latex", + "bbox": [ + 584, + 830, + 903, + 847 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Directly applying relation-aware generation methods to customized models leads to suboptimal performance, as illustrated in Fig. 4. We attribute this to two main reasons.", + "bbox": [ + 511, + 854, + 905, + 900 + ], + "page_idx": 2 + }, + { + "type": "page_number", + "text": "15725", + "bbox": [ + 480, + 944, + 517, + 955 + ], + "page_idx": 2 + }, + { + "type": "image", + "img_path": "images/179ca59630421d63fad72096ed5ba6770a9ce18feecc3fef2f58a9ec048377c3.jpg", + "image_caption": [ + "Figure 2. The overview of DreamRelation. DreamRelation utilizes the off-the-shelf identity extractor to decouple the relation and identity information in relation-specific images. After getting the U-Net output $\\hat{\\epsilon}$ , we predict $\\hat{z}_0$ and calculate the keypoint matching loss. The part in the dotted box is only for training." + ], + "image_footnote": [], + "bbox": [ + 91, + 88, + 903, + 297 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "First, the balance between the control of $c_{i}$ and $c_{t}$ is not addressed in relation-aware generation tasks. During inference, the model tends to overlook the target relationship $R^{*}$ due to the influence of $c_{i}$ , resulting in incorrect relationship generation. Second, even if the shared relationship exists across $\\{x_{k}\\}_{k = 1}^{n}$ , the disentangled object identity information complicates the learning process. Fine-tuning solely on $\\{x_{k}\\}_{k = 1}^{n}$ makes it challenging for the model to effectively capture the relationship. As shown in the second row of Fig. 4, where ReVersion [11] struggles to generate the \"hugging\" relationship even in the absence of $c_{i}$ . In the following subsections, we will introduce our data collection method and DreamRelation to bridge customization and relation generation.", + "bbox": [ + 91, + 354, + 483, + 565 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "3.3. Data Collection", + "text_level": 1, + "bbox": [ + 89, + 575, + 246, + 590 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "To address the balance between $c_{i}$ and $c_{t}$ while facilitating relation learning from $\\{x_{k}\\}_{k = 1}^{n}$ , we first design a data engine for collecting high-quality tuning data. The ideal tuning data should be in the form of $\\mathcal{D} = (x_k, c_i, c_t)$ . The $c_{i}$ contain the object in $x_{k}$ while $c_{t}$ describes the relationship between objects in $x_{k}$ . The identity information in $c_{i}$ helps decouple the relationship within $x_{k}$ , enhancing relation learning while maintaining a balanced integration of $c_{i}$ and $c_{t}$ . Unfortunately, directly cropping the objects from $x_{k}$ to obtain $c_{i}$ results in a copy-and-paste effect [5], even when using data augmentation techniques like flipping and rotation. Therefore, we introduce our data engine below.", + "bbox": [ + 89, + 598, + 482, + 779 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Relation-aware data engine. Fig. 3 illustrates the difference of our data engine. Both of $x_{k}$ and $c_{i}$ are generated by the recent text-to-image generation model, rather than through cropping and augmentation. Specifically, inspired by the performance of DALL-E 3 and its multi-turn dialog capability. We use DALL-E 3 to generate $\\mathcal{D} = (x,c_i,c_t)$ . Where $x_{k}$ and $c_{i}$ share the same object identity by leveraging the prompt \"The photo of the same.\" We observe that", + "bbox": [ + 89, + 779, + 482, + 901 + ], + "page_idx": 3 + }, + { + "type": "image", + "img_path": "images/7a43ac3f81a67b732255010dc504844454122f94c365dc3bbc25b93779444dde.jpg", + "image_caption": [ + "Figure 3. The rigid object cropping data engine, which results in minimal changes to object pose, the cropped image prompts also contain relation information, leading to a copy-and-paste effect, where the text prompt is neglected. Our relation-aware data engine, on the other hand, focuses on relation learning by decoupling identity information from images." + ], + "image_footnote": [], + "bbox": [ + 519, + 351, + 923, + 599 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "prompting in this manner allows DALL-E 3 to remember and preserve identity in common categories such as \"tiger\", which is enough for relation learning. Next, we employ X-Pose [41], SAM [13], and LLaVA [19] to annotate $x$ , $c_i$ with keypoints, masks, and captions for samples.", + "bbox": [ + 511, + 707, + 906, + 784 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "3.4. DreamRelation", + "text_level": 1, + "bbox": [ + 511, + 796, + 668, + 810 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "3.4.1. Relation Learning", + "text_level": 1, + "bbox": [ + 511, + 820, + 687, + 835 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "To balance $c_{i}$ and $c_{t}$ while enhancing relation learning, we decouple the identity information and relationships in $x_{k}$ . Specifically, we leverage an off-the-shelf ID extractor from the state-of-the-art customized model [35]. The identity in-", + "bbox": [ + 511, + 840, + 906, + 901 + ], + "page_idx": 3 + }, + { + "type": "page_number", + "text": "15726", + "bbox": [ + 480, + 944, + 519, + 955 + ], + "page_idx": 3 + }, + { + "type": "table", + "img_path": "images/dd50c029d71fa0c16d4b9664b266989f8eac8675f626f1cba7165c55835ce89a.jpg", + "table_caption": [], + "table_footnote": [], + "table_body": "
MethodsSingle-Object CusMulti-Object CusNon-ConfusingRelation-Aware
Textual Inversion [7]✓✗✗✗
DreamBooth [31]✓✗✗✗
Custom Diffusion [15]✓✓✗✗
Mix-of-Show [8]✓✓✗✗
CLIF [18]✓✓✓✗
MultiBooth [45]✓✓✗✗
BLIP-Diffusion [16]✓✗✗✗
ELITE [36]✓✗✗✗
AnyDoor [5]✓✗✗✗
SSR-Encoder [43]✓✓✗✗
λ-eclipse [24]✓✓✗✗
MS-Diffusion [35]✓✓✗✗
ReVersion [11]✗✗✓✓
ADI [10]✗✗✗✓
DreamRelation✓✓✓✓
", + "bbox": [ + 96, + 88, + 480, + 232 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Table 1. Setting Comparison for Different Models. We include several representative methods here. Our proposed DreamRelation can meet the demand of the relation-aware customization task.", + "bbox": [ + 89, + 237, + 482, + 280 + ], + "page_idx": 4 + }, + { + "type": "image", + "img_path": "images/91f5be357e43116896e2acb82eb1f233545ff1bd6c4dc219ab993a11cbc320d0.jpg", + "image_caption": [ + "Figure 4. Comparing our method with ReVersion across different base models, our approach demonstrates superior performance in the relation-aware generation task." + ], + "image_footnote": [], + "bbox": [ + 101, + 297, + 468, + 476 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "formation in $c_{i}$ enables the model to focus on the relationship information in $x_{k}$ and $c_{t}$ during relation learning. To achieve this, we employ parallel cross-attention layers to process $c_{i}$ and $c_{t}$ , which can be formulated as:", + "bbox": [ + 89, + 541, + 483, + 602 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\nh = \\gamma \\cdot \\operatorname {S o f t m a x} \\left(\\frac {Q K _ {i}}{\\sqrt {d}} + \\mathcal {M}\\right) V _ {i} + \\operatorname {S o f t m a x} \\left(\\frac {Q K _ {t}}{\\sqrt {d}}\\right) V _ {t} \\tag {4}\n$$\n", + "text_format": "latex", + "bbox": [ + 98, + 619, + 482, + 655 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "where $Q = W_{q}h$ , $K_{i} = W_{k_{i}}c_{i}$ , $V_{i} = W_{v_{i}}c_{i}$ , $K_{t} = W_{k_{t}}c_{t}$ , $V_{t} = W_{v_{t}}c_{t}$ , and $\\gamma$ is a hyperparameter for scaling control from $c_{i}$ . For clarity, we omit the final linear layer $W_{out}$ . The relation information mainly depends on text prompts. Therefore, we inject LoRA layers in all of the $W_{q}$ , $W_{k_{t}}$ , $W_{v_{t}}$ , and $W_{out}$ to encourage the model to pay more attention to the relationship in $c_{t}$ . We freeze all other parameters during fine-tuning. As shown in Fig. 4, benefiting from the decoupled identity information in $c_{i}$ , our method effectively captures the relationship in $x_{k}$ and $c_{t}$ , accurately generating the corresponding images. Notably, after fine-tuning, our method can also be directly integrated into SDXL [27], effectively enabling the model to address the relation-aware generation task. The extensive results are presented in the Supplementary Material (SM).", + "bbox": [ + 89, + 672, + 482, + 901 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "3.4.2. Keypoint Matching Loss", + "text_level": 1, + "bbox": [ + 513, + 90, + 732, + 107 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Most relationships have specific requirements for the pose, such as \"hugging\" requiring the arms to cross over the object's body, and \"riding a bicycle\" requiring the feet of the object to be on the pedal plate. Therefore, it's important to manipulate the pose of objects in the right status accurately. Intuitively, we introduce explicit supervision in the diffusion latent space during the tuning process, named Keypoint Matching Loss (KML).", + "bbox": [ + 511, + 109, + 906, + 229 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Specifically, considering our task only involves common objects instead of humans, we employ X-Pose [41] as a keypoint detector because it can detect any keypoints in complex real-world scenarios. We detect 17 keypoints for each object in $x_{k}$ and $c_{i}$ , and denote the keypoint coordinates as $c_{kp}^{x_k}, c_{kp}^{c_i} \\in \\mathcal{R}^{17 \\times 2}$ respectively. To encourage the model to generate accurate poses in $\\mathcal{D}(\\hat{z}_0)$ , where $\\mathcal{D}$ is the VAE decoder. We use the U-Net's output $\\hat{\\epsilon}$ to predict $z_0$ during fine-tuning:", + "bbox": [ + 511, + 231, + 906, + 366 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\hat {z} _ {0} = \\frac {z _ {t} - \\sqrt {1 - \\bar {\\alpha} _ {t}} \\hat {\\epsilon}}{\\sqrt {\\bar {\\alpha} _ {t}}} \\tag {5}\n$$\n", + "text_format": "latex", + "bbox": [ + 637, + 364, + 905, + 398 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "We use the VAE encoder $\\mathcal{E}$ to obtain the latent representation of image prompts $\\mathcal{E}(c_i)$ . Then we calculate the MSE loss on the corresponding keypoint locations between $\\mathcal{E}(c_i)$ and $\\hat{z}_0$ :", + "bbox": [ + 511, + 402, + 906, + 462 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {L} _ {K M L} = \\frac {1}{n _ {k p}} \\mathbb {E} _ {\\boldsymbol {z} _ {t}, c _ {i}} \\left\\| \\mathcal {E} (c _ {i}) [ c _ {k p} ^ {c _ {i}} ] - \\hat {z} _ {0} [ c _ {k p} ^ {x _ {k}} ] \\right\\| _ {2} ^ {2} \\tag {6}\n$$\n", + "text_format": "latex", + "bbox": [ + 557, + 470, + 906, + 503 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {L} = \\mathcal {L} _ {\\text {d e n o i s e}} + \\lambda \\cdot \\mathcal {L} _ {K M L} \\tag {7}\n$$\n", + "text_format": "latex", + "bbox": [ + 594, + 505, + 905, + 522 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "where $n_{kp}$ denote the number of keypoint, $c_{kp}^{c_i}$ and $c_{kp}^{x_k}$ are keypoint's coordinates in $c_i$ and $x_k$ respectively. We scale the coordinates to fit the size of $\\mathcal{E}(c_i)$ , $\\hat{z}_0 \\lambda$ controls for the relative weight of KML. The part inside the dotted box in Fig. 2 illustrates the model fine-tuning with the KML. We find KML is effective in encouraging the model to manipulate the pose for relation-aware generation.", + "bbox": [ + 511, + 531, + 905, + 638 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "3.4.3. Local Token Injection", + "text_level": 1, + "bbox": [ + 513, + 643, + 712, + 659 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "When generating relationships, local and detailed features are essential—for instance, the feature of arms to accurately construct relationships like \"carrying\". However, the coarse image-level features extracted by the CLIP Image Encoder fail to provide such fine-grained details, leading to confusion between objects in Fig. 7. To address this limitation, we introduce dense features from the CLIP Image Encoder [44]. Specifically, we modify the last layer of the CLIP Image Encoder to obtain dense features: $h_{\\mathrm{dense}} =$ ModifiedAttention( $h_{\\mathrm{clip}}'$ ), where $h_{\\mathrm{clip}}'$ represents the hidden states of the second-to-last layer.", + "bbox": [ + 511, + 662, + 906, + 829 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\nh _ {\\mathrm {t m p}} = \\operatorname {P r o j} _ {v} \\left(\\operatorname {n o r m} \\left(h _ {\\mathrm {c l i p}} ^ {\\prime}\\right)\\right), \\tag {8}\n$$\n", + "text_format": "latex", + "bbox": [ + 614, + 839, + 906, + 856 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\nh _ {\\mathrm {t m p}} = h _ {\\mathrm {c l i p}} ^ {\\prime} + \\operatorname {P r o j} _ {\\text {o u t}} \\left(h _ {\\mathrm {t m p}}\\right), \\tag {9}\n$$\n", + "text_format": "latex", + "bbox": [ + 616, + 859, + 906, + 877 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\nh _ {\\text {d e n s e}} = h _ {\\text {t m p}} + \\operatorname {F F N} \\left(h _ {\\text {t m p}}\\right) \\tag {10}\n$$\n", + "text_format": "latex", + "bbox": [ + 607, + 880, + 905, + 897 + ], + "page_idx": 4 + }, + { + "type": "page_number", + "text": "15727", + "bbox": [ + 480, + 944, + 517, + 955 + ], + "page_idx": 4 + }, + { + "type": "image", + "img_path": "images/809ee864e0f3795e33b31f9468f23149cc2ed865d2a79d8a8d671053d613027d.jpg", + "image_caption": [ + "Figure 5. Additional results demonstrate the effectiveness of relation-aware generation. DreamRelation adapts the same pair of objects to different relationships in a natural and accurate manner." + ], + "image_footnote": [], + "bbox": [ + 174, + 88, + 821, + 282 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "To further enhance the compatibility between dense features and image-level features, we employ self-distillation on the CLIP Image Encoder [39] to derive our Local Image Encoder. Specifically, we align local regions of the dense features with their corresponding image-level features using cosine similarity.", + "bbox": [ + 89, + 330, + 482, + 421 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "During inference, we extract local tokens $tok_{\\mathrm{local}}$ from the dense feature $h_{\\mathrm{dense}}$ through partitioning and pooling. These local tokens are then concatenated with the image-level tokens $tok_{\\mathrm{image}}$ and passed to the ID extractor. The ID extractor, which is a transformer-based architecture as illustrated in Fig. 2 (b), can be written as:", + "bbox": [ + 89, + 422, + 483, + 513 + ], + "page_idx": 5 + }, + { + "type": "equation", + "text": "\n$$\nq = q + \\text {A t t e n t i o n} (\\text {c o n c a t} [ q, \\text {t o k} _ {\\text {i m a g e}}, \\text {t o k} _ {\\text {l o c a l}} ]) \\tag {11}\n$$\n", + "text_format": "latex", + "bbox": [ + 114, + 526, + 482, + 545 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Regarding the compatibility between CLIP dense features and the original image-level features, experimental results demonstrate that our Local Image Encoder outperforms other dense representations [3]. For a detailed discussion, we refer readers to the Supplementary Material. Additionally, the injection of local tokens enhances identity preservation, as evidenced by improvements in evaluation metrics and visual results, shown in Fig. 7 and Tab. 3.", + "bbox": [ + 89, + 556, + 483, + 678 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "4. Experiments", + "text_level": 1, + "bbox": [ + 89, + 693, + 223, + 709 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Implementation Details. We generate a tuning set for relation learning with 4-6 samples per relationship. We incorporate LoRA layers into all text cross-attention layers of the U-Net. We set the LoRA rank to $r = 4$ , the parallel cross-attention scaling factor to $\\gamma = 0.6$ , and the keypoint matching loss weight to $\\lambda = 1e - 3$ . The model is finetuned for 500 steps, using 2 A100 GPUs, with a total batch size of 8, completing the process in 10 minutes. We use the Adam optimizer [12] with a learning rate of $1e - 4$ and no weight decay, resulting in a total of 3.1M trainable parameters. Note that our learnable parameters are only in the text cross-attention layers, making it compatible with", + "bbox": [ + 89, + 719, + 483, + 900 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "any SDXL-based models. Considering the strong identity preservation capabilities of MS-Diffusion, we implement our DreamRelation on MS-Diffusion. During fine-tuning and inference, we concatenate local tokens with image tokens. We put more self-distillation details of the Local Image Encoder in the Supplementary Material.", + "bbox": [ + 511, + 330, + 906, + 421 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Evaluation. To evaluate relation-aware customized image generation, we propose RelationBench, consisting of 44 objects from DreamBench [31] and CustomConcept101 [15], along with 25 relationships. The object categories include pets, plushies, toys, people, and cartoons. Using GPT-4, we generate 100 cases for single-object and multi-object evaluations. Following previous works [31, 35], we evaluate our method on three metrics: (1) Identity Preservation, which assesses the alignment between the generated images and the image prompts, using the CLIP image score and DINO score that calculate the cosine similarity between the class embeddings of images, referred to as CLIP-I and DINO, respectively; (2) Text Alignment, which evaluates how well the generated images align with the text prompts, using the CLIP image-text score, denoted as CLIP-T; and (3) Relation Alignment. We observed that nouns in the prompts can inflate CLIP-T scores, making them less accurate for evaluating relation generation. To address this, we extract the relation word from the text prompt using spaCy and calculate the CLIP image-text score, denoted as CLIP-R.", + "bbox": [ + 511, + 422, + 908, + 726 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "4.1. Main Results", + "text_level": 1, + "bbox": [ + 511, + 739, + 653, + 753 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "We evaluate our method through both quantitative and qualitative results in single- and multi-object cases. In single-object cases, we primarily focus on the alignment of the object's pose with the relationship in the text prompt and the model's ability to preserve identity. In multi-object cases, in addition to relation generation and identity preservation, we also need to ensure no confusion between the multiple objects.", + "bbox": [ + 511, + 762, + 906, + 883 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Quantitative results. We compare our method with", + "bbox": [ + 511, + 885, + 906, + 900 + ], + "page_idx": 5 + }, + { + "type": "page_number", + "text": "15728", + "bbox": [ + 480, + 944, + 517, + 955 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/379cf19086e316ab79992f9b391916aedabc0f7ca89b83b404a29b1f4b25a375.jpg", + "image_caption": [ + "Objects", + "A cat is kissing with a dog.", + "A dog is shaking hands with a plushie bear.", + "A dog is fencing with a plushie panda." + ], + "image_footnote": [], + "bbox": [ + 96, + 104, + 143, + 175 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/e102369285e21b5cc8ae72b36faee5b2f7b6ac7514d0d985b4b00b054557487d.jpg", + "image_caption": [ + "Prompts", + "Ours" + ], + "image_footnote": [], + "bbox": [ + 230, + 104, + 321, + 174 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/e10a75568002dc11826e587a9e8eac413e349b8c993468977332e3f7949ecef8.jpg", + "image_caption": [ + "ReVersion+MS" + ], + "image_footnote": [], + "bbox": [ + 326, + 104, + 416, + 172 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/8d19a2a803bcd79ee9e96e9cb83dae13b0b3883e0cf39d8b9e7d01008fc1be9a.jpg", + "image_caption": [ + "MS-Diffusion" + ], + "image_footnote": [], + "bbox": [ + 421, + 104, + 509, + 172 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/b2589730f24b1c2f3f12031d4b1773e0ab117f2a7f7b744cc00913ab8f3ae562.jpg", + "image_caption": [ + "SSR-Encoder" + ], + "image_footnote": [], + "bbox": [ + 514, + 104, + 604, + 172 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/8bf4caad1cc5881f70b03d4d991c0d7957631c5a62c953cf9b34fa1c2e3dc06d.jpg", + "image_caption": [ + "$\\lambda$ -eclipse" + ], + "image_footnote": [], + "bbox": [ + 609, + 104, + 699, + 172 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/988c9e1efedd07f5f5656173d3d0ebd0e3fb88b7835f4af9f27664ad6a1e4ff5.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 609, + 176, + 699, + 246 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/ede574f721266b94a4be719cd335833cd807c661ecf5343115e7796ac2214c53.jpg", + "image_caption": [ + "Custom Diffusion DreamBooth (SDXL)" + ], + "image_footnote": [], + "bbox": [ + 702, + 104, + 792, + 172 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/1d6107a00e1ba9db511ff8589de8ebd015dc5b79f7e44311a182427c896f2645.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 702, + 176, + 792, + 246 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/95eec3e4856dc4ea959accb9c0817c96907d14e9dd7b83c270aec1448d66b0f1.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 797, + 104, + 885, + 172 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/d2a93e7beb1b95cff6a234f46d3fc23c03ce3861b22648b1aa57bfa2a26af9b3.jpg", + "image_caption": [ + "Figure 6. We compare multi-object performance with training-based and tuning-based methods. Additionally, we integrate ReVersion [11] into MS-Diffusion, where the learnable text token is injected into the text prompt during MS-Diffusion's inference, resulting in ReVersion+MS. Our method shows a clear advantage in relation-aware generation and avoiding object confusion in overlapping scenarios." + ], + "image_footnote": [], + "bbox": [ + 702, + 250, + 792, + 319 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/9e418c165bf878515dc4e89a4e5c8c317ba989b5ffebd56d776fa23f1ad0a435.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 797, + 176, + 885, + 246 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/82bc6c3f3d392bd90ffb7eaa46d951666f771eb021df5195fc4f4f89b0aa459f.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 797, + 250, + 885, + 319 + ], + "page_idx": 6 + }, + { + "type": "table", + "img_path": "images/c4d76e37d26021ad0d3e4ea128069e1f75f10b1dd61b42c522c80369a3ffb3fc.jpg", + "table_caption": [], + "table_footnote": [], + "table_body": "
MethodSingle-objectMulti-object
CLIP-TCLIP-RCLIP-IDINOCLIP-TCLIP-RCLIP-IDINO
DreamBooth (SDXL) [31]27.818.274.262.624.316.267.857.2
Custom Diffusion [15]26.515.273.258.820.115.464.755.3
Cones-V2 [20]24.413.572.157.221.315.264.354.2
ELITE [36]25.714.975.461.5————
AnyDoor [5]24.514.777.462.221.614.969.759.8
BLIP-Diffusion [16]26.215.777.457.7————
SSR-Encoder [43]25.515.980.459.424.214.672.156.2
MS-Diffusion [35]26.518.878.764.526.918.973.858.8
Reversion+MS [11]27.819.377.863.127.219.274.259.7
Ours30.621.477.963.428.920.475.462.1
", + "bbox": [ + 129, + 378, + 859, + 571 + ], + "page_idx": 6 + }, + { + "type": "table", + "img_path": "images/529124363985afe02154831bb0db72c73d48d300b31c18a41d74ba80825e04b6.jpg", + "table_caption": [ + "Table 2. Quantitative comparison on RelationBench, Bold and underline represent the highest and second-highest metrics." + ], + "table_footnote": [], + "table_body": "
MethodMulti-object
CLIP-TCLIP-RCLIP-IDINO
w/o Relation-aware Data27.319.475.359.8
w/o Local Token Injection28.519.575.159.9
w/o Keypoint Matching Loss27.419.275.261.2
Full Model28.920.475.462.1
", + "bbox": [ + 93, + 602, + 473, + 683 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Table 3. Ablation study on our proposed components", + "bbox": [ + 127, + 686, + 444, + 700 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "baseline models across three benchmarks: Relation-Bench, DreamBench, and multi-object cases in Custom-Concept101. As shown in Tab. 2, our method demonstrates a clear advantage on RelationBench for single-object cases, particularly in the CLIP-T and CLIP-R metrics. However, its performance is slightly lower in the CLIP-I metric, likely due to significant pose variations in the objects, which may decrease the CLIP-I score. For multi-object generation, our method outperforms other approaches across all evaluation metrics, which we attribute to its ability to effectively avoid object confusion and significantly enhance CLIP-I performance. Furthermore, as shown in Tab. 4, we evaluate", + "bbox": [ + 89, + 719, + 483, + 900 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "single-object generation performance on DreamBench. Our method outperforms others in the CLIP-T metric while delivering competitive results in CLIP-I and DINO. We conduct experiments on multi-object cases from CustomConcept101, which we denote as M-CustomConcept101. As shown in Fig. 5, our model achieves the highest performance in the DINO score and ranks second in both the CLIP-T and CLIP-I metrics.", + "bbox": [ + 511, + 604, + 906, + 724 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Qualitative Comparison. We conduct qualitative comparison experiments on RelationBench to evaluate relation-aware customized image generation. As shown in Fig. 6, our method outperforms others in generating multi-object relationships, effectively avoiding object confusion while accurately capturing the intended relationships. Additional results can be found in the SM.", + "bbox": [ + 511, + 727, + 906, + 832 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "4.2. Ablation study and Analysis", + "text_level": 1, + "bbox": [ + 511, + 845, + 764, + 863 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Relation-Aware Data Engine. Instead of fine-tuning on our curated $(x_{k},c_{i},c_{t})$ pairs. We train a text embedding $R^{*}$", + "bbox": [ + 511, + 869, + 905, + 901 + ], + "page_idx": 6 + }, + { + "type": "page_number", + "text": "15729", + "bbox": [ + 480, + 944, + 517, + 955 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/c32e4c51e7a31b10383fb3a9b685ecc9b41537c37ccecf9bf7fb6145c57a1a43.jpg", + "image_caption": [ + "a) objects" + ], + "image_footnote": [], + "bbox": [ + 107, + 104, + 179, + 215 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/0cc68667bf281a2204618555ea89da1f93e41bba7b0e13375c76845b552ad4cc.jpg", + "image_caption": [ + "wKML", + "A cat is shaking hands with a dog." + ], + "image_footnote": [], + "bbox": [ + 179, + 104, + 321, + 215 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/2cc0ebf8b7f7e3aea1fbb1cf771a434fd3853238e211b5d69d00071345cc47d3.jpg", + "image_caption": [ + "w/o KML" + ], + "image_footnote": [], + "bbox": [ + 321, + 104, + 465, + 215 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/8436a73e636a2148e8b7a4b22337942cefc829a9361b77e0a299c9032012c955.jpg", + "image_caption": [ + "b) objects" + ], + "image_footnote": [], + "bbox": [ + 107, + 251, + 181, + 308 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/585e297493747fd77bf7493be8bdea864588b9d43a1d8f9f364a84618a821200.jpg", + "image_caption": [ + "w local tokens", + "A dog is carrying a sloth." + ], + "image_footnote": [], + "bbox": [ + 181, + 251, + 323, + 362 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/48d030056b3ff97702524a71b39ca2139dcc3c05dac93881cabd07bad6aaff44.jpg", + "image_caption": [ + "w/o local tokens" + ], + "image_footnote": [], + "bbox": [ + 323, + 251, + 467, + 362 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/41e954c311fecc6c76611c1ed057e5c60c72fa2762f1539244342bd52da504b9.jpg", + "image_caption": [ + "Figure 7. Ablation study of keypoint matching loss (KML) and local tokens: The images are generated using the same random seed. We use green and red boxes to highlight the main differences." + ], + "image_footnote": [], + "bbox": [ + 102, + 426, + 130, + 494 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/bbab11eb667e11037d0aaa1da0d501ce47968c0ae8263f930f8f5c0d24132ae1.jpg", + "image_caption": [ + "A dog, a cat, a plushie bear hug with each other.", + "Figure 8. DreamRelation for three objects. Compared to the baseline model, DreamRelation shows a clear advantage." + ], + "image_footnote": [], + "bbox": [ + 130, + 428, + 183, + 496 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/f413434c347e23d2334146db8ee38434698549a26d84273dae987ff814449779.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 183, + 430, + 210, + 496 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/bdeac0a670eabb1140cb1c1fb6c118c3778e4b7f077aba905985c69174d782f6.jpg", + "image_caption": [ + "A dog, a cat, and a plushie bear are fighting with one another." + ], + "image_footnote": [], + "bbox": [ + 210, + 430, + 295, + 496 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/c6946fccc9d404af3675a38a9f8de4efe96dce4e18597be7b39406edfcc50e73.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 295, + 430, + 352, + 496 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/0a85a82b8ac6fbae2b0a50ec754485584c551cd7d3372904df6bc833592c4e57.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 352, + 430, + 457, + 496 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "using a set of relation-specific data $(x_{k},c_{i}^{\\prime},c_{t})$ , where $c_{i}^{\\prime}$ is cropped from $x_{k}$ . The well-trained text embedding $R^{*}$ is injected into the text embedding during inference. The quantitative results in Tab. 3 show that our relation-aware data engine outperforms in both CLIP-T and CLIP-R scores.", + "bbox": [ + 88, + 551, + 482, + 628 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Keypoint Matching Loss (KML). We omit KML during fine-tuning. As shown in Fig. 7 a), including the KML leads to more accurate and natural relation-aware customized images using the same random seed. Quantitative results, presented in Tab. 3, show a decline across all evaluation metrics when the KML is removed. Additionally, we experimented with different values for $\\lambda$ and found that the best performance occurs at $\\lambda = 1e - 3$ .", + "bbox": [ + 88, + 628, + 482, + 748 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Local Token Injection. For local token injection, the visual comparison in Fig. 7 b) shows severe object confusion when local tokens are omitted. The quantitative results in Tab. 3 demonstrate that local tokens improve relation-aware customization. We conducted an ablation study on the number of local tokens, as presented in Tab. 6, and determined the optimal number to be 16. Additional ablation studies are shown in the SM. Combining these methods ensures that DreamRelation achieves high-quality relation-aware customized generation.", + "bbox": [ + 88, + 750, + 482, + 901 + ], + "page_idx": 7 + }, + { + "type": "table", + "img_path": "images/238eeb2740c11e0e48e04604c379bf23c9126ca30bf212cbe022135be0102ba7.jpg", + "table_caption": [], + "table_footnote": [], + "table_body": "
MethodSingle-object
CLIP-TCLIP-IDINO
DreamBooth (SDXL) [31]31.281.569.2
Custom Diffusion [15]28.477.266.8
Cones-V2 [20]31.076.567.2
ELITE [36]29.877.462.5
AnyDoor [5]25.582.167.8
SSR-Encoder [43]30.882.161.2
MS-Diffusion [35]31.579.368.2
Ours31.681.768.4
", + "bbox": [ + 549, + 88, + 864, + 218 + ], + "page_idx": 7 + }, + { + "type": "table", + "img_path": "images/0337ebcb900b60ce15fee8b5234bcd7d2c6148199ab55049ad12caf1c9f6ac16.jpg", + "table_caption": [ + "Table 4. Evaluation Results on DreamBench." + ], + "table_footnote": [], + "table_body": "
MethodMulti-object
CLIP-TCLIP-IDINO
DreamBooth (SDXL) [31]29.567.449.2
Custom Diffusion [15]28.166.248.8
Cones-V2 [20]29.266.547.2
AnyDoor [5]20.272.151.2
SSR-Encoder [43]30.671.152.2
λ-eclipse [24]29.268.248.2
MS-Diffusion [35]28.070.251.2
Ours29.771.452.3
", + "bbox": [ + 549, + 239, + 864, + 371 + ], + "page_idx": 7 + }, + { + "type": "table", + "img_path": "images/a36c412af49e3504c03b18093270e3a83379801ffa29ebdb4acb8ce996950ed2.jpg", + "table_caption": [ + "Table 5. Evaluation Results on M-CustomConcept101." + ], + "table_footnote": [], + "table_body": "
Num Local TokensMulti-object
CLIP-TCLIP-RCLIP-IDINO
N=2×228.319.475.360.1
N=4×428.920.475.462.1
N=8×827.918.174.159.4
", + "bbox": [ + 540, + 392, + 870, + 460 + ], + "page_idx": 7 + }, + { + "type": "table", + "img_path": "images/d4100e646cefbf17bc65414a9c0b9eb42b9b3abcb5a85c69476e5d178cfdfc26.jpg", + "table_caption": [ + "Table 6. Ablation study on the number of local tokens." + ], + "table_footnote": [], + "table_body": "
LambdaMulti-object
CLIP-TCLIP-RCLIP-IDINO
λ=1e-227.519.372.659.6
λ=1e-328.920.475.462.1
λ=1e-426.818.173.960.4
", + "bbox": [ + 550, + 483, + 861, + 560 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Table 7. Ablation study on $\\lambda$ ,controlling relative weight of KML.", + "bbox": [ + 513, + 564, + 901, + 579 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "5. Conclusion", + "text_level": 1, + "bbox": [ + 513, + 592, + 633, + 607 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "In this work, we introduce a new challenging task: relation-aware customized image generation. This task aims to generate objects that adhere to the relationship specified in the text prompt and preserve the identities of the user-provided images. To support this, we propose a data engine for generating high-quality fine-tuning data. Our method, DreamRelation, incorporates Keypoint Matching Loss and Local Token Injection, effectively capturing relation information and generating natural relations between customized objects. We also present a fair comparison with other methods on a new benchmark, RelationBench. Extensive experiments demonstrate the effectiveness of our method, highlighting its potential for applications in interactive scenario generation, relation detection, and more. Our research can inspire the direction of relation-aware generation in the community.", + "bbox": [ + 511, + 617, + 906, + 844 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Acknowledgement. This work is supported by the National Key Research and Development Program of China (No. 2023YFC3807600).", + "bbox": [ + 511, + 844, + 906, + 887 + ], + "page_idx": 7 + }, + { + "type": "page_number", + "text": "15730", + "bbox": [ + 480, + 944, + 519, + 955 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "References", + "text_level": 1, + "bbox": [ + 91, + 89, + 187, + 104 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[1] Jinbin Bai, Tian Ye, Wei Chow, Enxin Song, Qing-Guo Chen, Xiangtai Li, Zhen Dong, Lei Zhu, and Shuicheng Yan. Meissonic: Revitalizing masked generative transformers for efficient high-resolution text-to-image synthesis. arXiv preprint arXiv:2410.08261, 2024. 2", + "[2] James Betker, Gabriel Goh, Li Jing, Tim Brooks, Jianfeng Wang, Linjie Li, Long Ouyang, Juntang Zhuang, Joyce Lee, Yufei Guo, et al. Improving image generation with better captions. Computer Science., 2:3, 2023. 2", + "[3] Mathilde Caron, Hugo Touvron, Ishan Misra, Herve Jégou, Julien Mairal, Piotr Bojanowski, and Armand Joulin. Emerging properties in self-supervised vision transformers. In Int. Conf. Comput. Vis., 2021. 6", + "[4] Junsong Chen, Jincheng Yu, Chongjian Ge, Lewei Yao, Enze Xie, Yue Wu, Zhongdao Wang, James Kwok, Ping Luo, Huchuan Lu, et al. Pixart-α: Fast training of diffusion transformer for photorealistic text-to-image synthesis. In ICLR, 2024. 2", + "[5] Xi Chen, Lianghua Huang, Yu Liu, Yujun Shen, Deli Zhao, and Hengshuang Zhao. Anydoor: Zero-shot object-level image customization. In CVPR, 2024. 2, 3, 4, 5, 7, 8", + "[6] Hyung Won Chung, Le Hou, Shayne Longpre, Barret Zoph, Yi Tay, William Fedus, Yunxuan Li, Xuezhi Wang, Mostafa Dehghani, Siddhartha Brahma, et al. Scaling instruction-finetuned language models. arXiv preprint arXiv:2210.11416, 2022. 2", + "[7] Rinon Gal, Yuval Alaluf, Yuval Atzmon, Or Patashnik, Amit Haim Bermano, Gal Chechik, and Daniel Cohen-Or. An image is worth one word: Personalizing text-to-image generation using textual inversion. In ICLR, 2023. 3, 5", + "[8] Yuchao Gu, Xintao Wang, Jay Zhangjie Wu, Yujun Shi, Yunpeng Chen, Zihan Fan, Wuyou Xiao, Rui Zhao, Shuning Chang, Weijia Wu, et al. Mix-of-show: Decentralized low-rank adaptation for multi-concept customization of diffusion models. NeurIPS, 2024. 2, 3, 5", + "[9] Edward J Hu, Yelong Shen, Phillip Wallis abd Zeyuan Allen-Zhu, Yuanzhi Li, Shean Wang, Lu Wang, and Weizhu Chen. Lora: Low-rank adaptation of large language models. In ICLR, 2022. 2", + "[10] Siteng Huang, Biao Gong, Yutong Feng, Xi Chen, Yuqian Fu, Yu Liu, and Donglin Wang. Learning disentangled identifiers for action-customized text-to-image generation. In CVPR, 2024. 5", + "[11] Ziqi Huang, Tianxing Wu, Yuming Jiang, Kelvin CK Chan, and Ziwei Liu. Reversion: Diffusion-based relation inversion from images. arXiv preprint arXiv:2303.13495, 2023. 2, 3, 4, 5, 7", + "[12] Diederik P Kingma and Jimmy Ba. Adam: A method for stochastic optimization. arXiv preprint arXiv:1412.6980, 2014. 6", + "[13] Alexander Kirillov, Eric Mintun, Nikhila Ravi, Hanzi Mao, Chloe Rolland, Laura Gustafson, Tete Xiao, Spencer Whitehead, Alexander C. Berg, Wan-Yen Lo, Piotr Dollar, and Ross B. Girshick. Segment anything. In ICCV, 2023. 4", + "[14] Lingjie Kong, Kai Wu, Xiaobin Hu, Wenhui Han, Jinlong Peng, Chengming Xu, Donghao Luo, Jiangning Zhang," + ], + "bbox": [ + 93, + 114, + 483, + 902 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "Chengjie Wang, and Yanwei Fu. Anymaker: Zero-shot general object customization via decoupled dual-level id injection. arXiv preprint arXiv:2406.11643, 2024. 2", + "[15] Nupur Kumari, Bingliang Zhang, Richard Zhang, Eli Shechtman, and Jun-Yan Zhu. Multi-concept customization of text-to-image diffusion. In CVPR, 2023. 2, 3, 5, 6, 7, 8", + "[16] Dongxu Li, Junnan Li, and Steven C. H. Hoi. Blip-diffusion: Pre-trained subject representation for controllable text-to-image generation and editing. In NeurIPS, 2023. 2, 3, 5, 7", + "[17] Long Lian, Boyi Li, Adam Yala, and Trevor Darrell. Llm-grounded diffusion: Enhancing prompt understanding of text-to-image diffusion models with large language models. arXiv preprint arXiv:2305.13655, 2023. 2", + "[18] Wang Lin, Jingyuan Chen, Jiaxin Shi, Yichen Zhu, Chen Liang, Junzhong Miao, Tao Jin, Zhou Zhao, Fei Wu, Shuicheng Yan, et al. Non-confusing generation of customized concepts in diffusion models. ICML, 2024. 2, 3, 5", + "[19] Haotian Liu, Chunyuan Li, Qingyang Wu, and Yong Jae Lee. Visual instruction tuning. NeurIPS, 2024. 4", + "[20] Zhiheng Liu, Yifei Zhang, Yujun Shen, Kecheng Zheng, Kai Zhu, Ruili Feng, Yu Liu, Deli Zhao, Jingren Zhou, and Yang Cao. Cones 2: Customizable image synthesis with multiple subjects. In NeurIPS, 2023. 2, 3, 7, 8", + "[21] Joanna Materzynska, Josef Sivic, Eli Shechtman, Antonio Torralba, Richard Zhang, and Bryan Russell. Customizing motion in text-to-video diffusion models. arXiv preprint arXiv:2312.04966, 2023. 2", + "[22] mengmeng Ge, Xu Jia, Takashi Isobe, Xiaomin Li, Qinghe Wang, Jing Mu, Dong Zhou, liwang Amd, Huchuan Lu, Lu Tian, Ashish Sirasao, and Emad Barsoum. Customizing text-to-image generation with inverted interaction. In ACM MM, 2024. 2, 3", + "[23] Lianyu Pang, Jian Yin, Baoquan Zhao, Feize Wu, Fu Lee Wang, Qing Li, and Xudong Mao. Attndreambooth: Towards text-aligned personalized text-to-image generation. arXiv preprint arXiv:2406.05000, 2024. 2, 3", + "[24] Maitreya Patel, Sangmin Jung, Chitta Baral, and Yezhou Yang. λ-eclipse: Multi-concept personalized text-to-image diffusion models by leveraging clip latent space. arXiv preprint arXiv:2402.05195, 2024. 2, 3, 5, 8", + "[25] William Peebles and Saining Xie. Scalable diffusion models with transformers. In ICCV, 2023. 2", + "[26] Pablo Pernias, Dominic Rampas, Mats Leon Richter, Christopher Pal, and Marc Aubreville. Würstchen: An efficient architecture for large-scale text-to-image diffusion models. In ICLR, 2023. 2", + "[27] Dustin Podell, Zion English, Kyle Lacey, Andreas Blattmann, Tim Dockhorn, Jonas Müller, Joe Penna, and Robin Rombach. Sdxl: Improving latent diffusion models for high-resolution image synthesis. arXiv preprint arXiv:2307.01952, 2023. 2, 5", + "[28] Alec Radford, Jong Wook Kim, Chris Hallacy, Aditya Ramesh, Gabriel Goh, Sandhini Agarwal, Girish Sastry, Amanda Askell, Pamela Mishkin, Jack Clark, Gretchen Krueger, and Ilya Sutskever. Learning transferable visual" + ], + "bbox": [ + 516, + 92, + 903, + 900 + ], + "page_idx": 8 + }, + { + "type": "page_number", + "text": "15731", + "bbox": [ + 480, + 944, + 517, + 955 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "models from natural language supervision. In ICML, 2021. 2,3", + "[29] Jingjing Ren, Wenbo Li, Haoyu Chen, Renjing Pei, Bin Shao, Yong Guo, Long Peng, Fenglong Song, and Lei Zhu. Ultrapixel: Advancing ultra-high-resolution image synthesis to new peaks. arXiv preprint arXiv:2407.02158, 2024. 2", + "[30] Robin Rombach, Andreas Blattmann, Dominik Lorenz, Patrick Esser, and Björn Ommer. High-resolution image synthesis with latent diffusion models. In CVPR, 2022. 2", + "[31] Nataniel Ruiz, Yuanzhen Li, Varun Jampani, Yael Pritch, Michael Rubinstein, and Kfir Aberman. Dreambooth: Fine tuning text-to-image diffusion models for subject-driven generation. In CVPR, 2023. 2, 3, 5, 6, 7, 8", + "[32] Chitwan Sahara, William Chan, Saurabh Saxena, Lala Li, Jay Whang, Emily L. Denton, Seyed Kamyar Seyed Ghasemipour, Raphael Gontijo Lopes, Burcu Karagol Ayan, Tim Salimans, Jonathan Ho, David J. Fleet, and Mohammad Norouzi. Photorealistic text-to-image diffusion models with deep language understanding. In NeurIPS, 2022. 2", + "[33] Ashish Vaswani, Noam Shazeer, Niki Parmar, Jakob Uszkoreit, Llion Jones, Aidan N Gomez, Lukasz Kaiser, and Illia Polosukhin. Attention is all you need. In NeurIPS, 2017. 2", + "[34] Andrey Voynov, Qinghao Chu, Daniel Cohen-Or, and Kfir Aberman. p+: Extended textual conditioning in text-to-image generation. arXiv preprint arXiv:2303.09522, 2023. 3", + "[35] X Wang, Siming Fu, Qihan Huang, Wanggui He, and Hao Jiang. Ms-diffusion: Multi-subject zero-shot image personalization with layout guidance. arXiv preprint arXiv:2406.07209, 2024. 2, 3, 4, 5, 6, 7, 8", + "[36] Yuxiang Wei, Yabo Zhang, Zhilong Ji, Jinfeng Bai, Lei Zhang, and Wangmeng Zuo. ELITE: encoding visual concepts into textual embeddings for customized text-to-image generation. In ICCV, 2023. 2, 3, 5, 7, 8", + "[37] Yujie Wei, Shiwei Zhang, Zhiwu Qing, Hangjie Yuan, Zhiheng Liu, Yu Liu, Yingya Zhang, Jingren Zhou, and Hongming Shan. Dreamvideo: Composing your dream videos with customized subject and motion. In CVPR, 2024. 3", + "[38] Jianzong Wu, Xiangtai Li, Yanhong Zeng, Jiangning Zhang, Qianyu Zhou, Yining Li, Yunhai Tong, and Kai Chen. Motionbooth: Motion-aware customized text-to-video generation. arXiv preprint arXiv:2406.17758, 2024. 2", + "[39] Size Wu, Wenwei Zhang, Lumin Xu, Sheng Jin, Xiangtai Li, Wentao Liu, and Chen Change Loy. Clipself: Vision transformer distills itself for open-vocabulary dense prediction. ICLR, 2024. 2, 6", + "[40] Binxin Yang, Shuyang Gu, Bo Zhang, Ting Zhang, Xuejin Chen, Xiaoyan Sun, Dong Chen, and Fang Wen. Paint by example: Exemplar-based image editing with diffusion models. In CVPR, 2023. 2", + "[41] Jie Yang, Ailing Zeng, Ruimao Zhang, and Lei Zhang. Unipose: Detecting any keypoints. ECCV, 2024. 4, 5", + "[42] Hu Ye, Jun Zhang, Sibo Liu, Xiao Han, and Wei Yang. Ip-adapter: Text compatible image prompt adapter for text-to-image diffusion models. arXiv preprint arxiv:2308.06721, 2023. 2" + ], + "bbox": [ + 91, + 90, + 482, + 898 + ], + "page_idx": 9 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[43] Yuxuan Zhang, Yiren Song, Jiaming Liu, Rui Wang, Jinpeng Yu, Hao Tang, Huaxia Li, Xu Tang, Yao Hu, Han Pan, et al. Ssr-encoder: Encoding selective subject representation for subject-driven generation. In CVPR, 2024. 2, 3, 5, 7, 8", + "[44] Chong Zhou, Chen Change Loy, and Bo Dai. Extract free dense labels from clip. In ECCV, 2022. 5", + "[45] Chenyang Zhu, Kai Li, Yue Ma, Chunming He, and Li Xiu. Multibooth: Towards generating all your concepts in an image from text. arXiv preprint arXiv:2404.14239, 2024. 5" + ], + "bbox": [ + 514, + 90, + 905, + 219 + ], + "page_idx": 9 + }, + { + "type": "page_number", + "text": "15732", + "bbox": [ + 480, + 945, + 517, + 955 + ], + "page_idx": 9 + } +] \ No newline at end of file diff --git a/2025/DreamRelation_ Bridging Customization and Relation Generation/dc0bd82f-ae71-4d19-a1ab-522a355a226a_model.json b/2025/DreamRelation_ Bridging Customization and Relation Generation/dc0bd82f-ae71-4d19-a1ab-522a355a226a_model.json new file mode 100644 index 0000000000000000000000000000000000000000..5984e49e02caada06fb0e6bd9bba8051b1516401 --- /dev/null +++ b/2025/DreamRelation_ Bridging Customization and Relation Generation/dc0bd82f-ae71-4d19-a1ab-522a355a226a_model.json @@ -0,0 +1,2442 @@ +[ + [ + { + "type": "header", + "bbox": [ + 0.107, + 0.003, + 0.182, + 0.043 + ], + "angle": 0, + "content": "CVF" + }, + { + "type": "header", + "bbox": [ + 0.237, + 0.001, + 0.809, + 0.047 + ], + "angle": 0, + "content": "This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore." + }, + { + "type": "title", + "bbox": [ + 0.166, + 0.131, + 0.833, + 0.154 + ], + "angle": 0, + "content": "DreamRelation: Bridging Customization and Relation Generation" + }, + { + "type": "text", + "bbox": [ + 0.092, + 0.179, + 0.929, + 0.234 + ], + "angle": 0, + "content": "Qingyu Shi\\(^{1}\\) Lu Qi\\(^{3*}\\) Jianzong Wu\\(^{1}\\) Jinbin Bai\\(^{5}\\) Jingbo Wang\\(^{4}\\) Yunhai Tong\\(^{1}\\) Xiangtai Li\\(^{2,4*}\\) \n\\(^{1}\\) PKU NTU Insta360 Research Shanghai AI Laboratory NUS \nProject page: https://shi-qingyu.github.io/DreamRelation.github.io" + }, + { + "type": "image", + "bbox": [ + 0.101, + 0.24, + 0.895, + 0.555 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.089, + 0.566, + 0.908, + 0.624 + ], + "angle": 0, + "content": "Figure 1. In our Relation-Aware Image Customization task, the generated images must accurately preserve the relationships between objects and maintain their identity. We highlight the limitations of previous approaches using three color codes: red indicates failure to capture relationships, blue marks missing objects, and orange represents object confusion. Each image is annotated to reflect its specific issue. Our results, highlighted by green boxes, demonstrate the advantages of our proposed method." + }, + { + "type": "title", + "bbox": [ + 0.248, + 0.647, + 0.327, + 0.663 + ], + "angle": 0, + "content": "Abstract" + }, + { + "type": "text", + "bbox": [ + 0.089, + 0.69, + 0.486, + 0.902 + ], + "angle": 0, + "content": "Customized image generation is essential for creating personalized content based on user prompts, allowing large-scale text-to-image diffusion models to more effectively meet individual needs. However, existing models often neglect the relationships between customized objects in generated images. In contrast, this work addresses this gap by focusing on relation-aware customized image generation, which seeks to preserve the identities from image prompts while maintaining the relationship specified in text prompts. Specifically, we introduce DreamRelation, a framework that disentangles identity and relation learning using a carefully curated dataset. Our training data consists of relation-specific images, independent object images containing identity information, and text prompts to guide relation genera" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.649, + 0.909, + 0.846 + ], + "angle": 0, + "content": "tion. Then, we propose two key modules to tackle the two main challenges—generating accurate and natural relationships, especially when significant pose adjustments are required, and avoiding object confusion in cases of overlap. First, we introduce a keypoint matching loss that effectively guides the model in adjusting object poses closely tied to their relationships. Second, we incorporate local features of the image prompts to better distinguish between objects, preventing confusion in overlapping cases. Extensive results on our proposed benchmarks demonstrate the superiority of DreamRelation in generating precise relations while preserving object identities across a diverse set of objects and relationships." + }, + { + "type": "page_footnote", + "bbox": [ + 0.532, + 0.887, + 0.781, + 0.902 + ], + "angle": 0, + "content": "\\(^{1}\\)Corresponding authors: Lu Qi, Xiangtai Li." + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "15723" + } + ], + [ + { + "type": "title", + "bbox": [ + 0.093, + 0.09, + 0.223, + 0.107 + ], + "angle": 0, + "content": "1. Introduction" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.115, + 0.482, + 0.252 + ], + "angle": 0, + "content": "Driven by large-scale text-to-image diffusion models [26, 27, 30, 32], customized image generation has recently made significant strides [1, 14, 15, 31, 35, 36, 43]. This task focuses on generating images that preserve the identity of objects from user-provided inputs, enabling the creation of personalized and meaningful content. It has shown value in numerous applications, including personalized artwork, branding, virtual fashion try-ons, social media content creation, augmented reality experiences, and more." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.252, + 0.483, + 0.418 + ], + "angle": 0, + "content": "Despite the success of many methods for customizing single or multiple objects [8, 16, 18, 20, 23, 24, 35, 42], they often overlook the relationships between objects and the corresponding text prompts. For instance, when two user-provided objects are paired with a text prompt specifying a particular relationship, the generated output should not only preserve their identities but also accurately reflect the intended relationship, such as a 'hug'. This introduces new challenges in what we refer to as relation-aware customized image generation, which focuses on preserving multiple identities while adhering to relationship prompts." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.418, + 0.483, + 0.599 + ], + "angle": 0, + "content": "An intuitive solution to this issue is to adapt existing relation-aware generation methods [11, 21, 22] to tuning-based or training-based customized methods for customizing objects while maintaining the relation. However, both approaches face challenges. Tuning-based methods struggle to preserve multiple identities, as they invert objects into specific tokens. While training-based schemes often fail to balance image and text prompts, frequently overlooking key textual elements and hindering the generation of relationships between objects. As illustrated in Fig. 1, previous methods fail to capture the relationships described in the text prompt and lose identity preservation." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.599, + 0.483, + 0.825 + ], + "angle": 0, + "content": "We attribute this failure to two key factors: a lack of relevant data and an ineffective model design. Unlike data augmentation techniques such as flipping or rotation, commonly used to create paired training data in object customization methods [5, 40], our approach requires a triplet of images: two image prompts and one target image. The image prompts should contain similar objects but exhibit distinct poses compared to the target image. To collect these triplets, we propose a data engine to curate our finetuning set. We leverage an advanced text-to-image generation model [2] to generate triplets where the same object pair is shared across the images. Through text prompt guidance, the image prompt provides strong identity information. This enables the decoupling of the relationship in the target image, which enhances the relation learning." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.826, + 0.483, + 0.901 + ], + "angle": 0, + "content": "For the model design, we propose DreamRelation, which applies the Low-Rank Adaptation (LoRA) [9] strategy to the text cross-attention layers of existing diffusion models to process user-provided text prompts. In DreamRelation, two key modules are introduced during training to" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.092, + 0.905, + 0.333 + ], + "angle": 0, + "content": "enhance the relation generation in customized generation. First, we introduce a keypoint matching loss (KML) as additional supervision to explicitly encourage the model to manipulate object poses, since relationships between objects are closely tied to their poses. Importantly, the KML operates on the latent representation rather than the original image space, aligning with the default diffusion loss. Second, since relation-aware customization requires local features from image prompts, such as the \"hands\" features for generating \"shaking hands\"—which are not captured by CLIP's coarse image-level features, we introduce dense features from CLIP. Through partitioning and pooling, we obtain local tokens that contain detail and local information. To further enhance the compatibility between dense features and image-level features, we employ a self-distillation method [39] to improve their alignment." + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.334, + 0.905, + 0.424 + ], + "angle": 0, + "content": "To more comprehensively evaluate relation-aware customized image generation, we constructed our Relation-Bench based on three established benchmarks [15, 31]. Our methods demonstrate strong performance compared to existing approaches, achieving significant improvements in visual quality and quantitative results." + }, + { + "type": "text", + "bbox": [ + 0.534, + 0.424, + 0.766, + 0.438 + ], + "angle": 0, + "content": "The contributions of this work are:" + }, + { + "type": "text", + "bbox": [ + 0.514, + 0.44, + 0.905, + 0.514 + ], + "angle": 0, + "content": "- We explore a novel task called relation-aware customized generation, which aims to preserve multiple identities from image prompts while adhering to the relationships specified in text prompts. This task can enhance various user-driven applications by enabling more control." + }, + { + "type": "text", + "bbox": [ + 0.514, + 0.515, + 0.905, + 0.605 + ], + "angle": 0, + "content": "- We introduce a data engine that uses the recent text-to-image generation model to generate triplet images where the same object is present with distinct poses. This well-curated dataset enables the model to focus on relation learning during fine-tuning, leveraging the identity information provided by the image prompts." + }, + { + "type": "text", + "bbox": [ + 0.514, + 0.606, + 0.905, + 0.68 + ], + "angle": 0, + "content": "- Our proposed DreamRelation has two key modules, including keypoint matching loss and local token injection, to enhance the relation generation in customized generation. The extensive experiments on three benchmarks demonstrate the effectiveness of our method." + }, + { + "type": "list", + "bbox": [ + 0.514, + 0.44, + 0.905, + 0.68 + ], + "angle": 0, + "content": null + }, + { + "type": "title", + "bbox": [ + 0.514, + 0.694, + 0.655, + 0.71 + ], + "angle": 0, + "content": "2. Related Work" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.72, + 0.905, + 0.901 + ], + "angle": 0, + "content": "Diffusion-Based Text-to-Image Generation. Diffusion-based text-to-image models [30, 32] generate high-quality images based on user-provided text prompts. These works encode the text prompt through the text encoder [6, 28] and inject the text embedding into the U-Net's cross-attention layers. Some methods [26, 27, 29, 38] upscale diffusion models and incorporate additional conditions as priors to generate high-resolution images. Meanwhile, several methods [2, 17] introduce stronger text encoders or large language models (LLMs) to enhance the text comprehension capabilities of diffusion models. Recent works [4, 25] have replaced the U-Net denoiser with the Transformers [33]." + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.52, + 0.957 + ], + "angle": 0, + "content": "15724" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.09, + 0.092, + 0.482, + 0.168 + ], + "angle": 0, + "content": "However, these models primarily focus on text conditions and struggle to handle other forms of guidance, such as user-provided images. In contrast, our work facilitates customized generation while following relation conditions given by text inputs." + }, + { + "type": "text", + "bbox": [ + 0.093, + 0.187, + 0.485, + 0.581 + ], + "angle": 0, + "content": "Diffusion-Based Customized Generation. Customized image generation aims to produce new images based on user-provided objects. Tuning-based methods accomplish this by fine-tuning specific parameters of diffusion models to learn new objects. Some of these methods [7, 20, 34] employ text embeddings to represent the object identity. On the other hand, DreamBooth [31] fine-tunes the entire U-Net and introduces a prior preservation loss to mitigate language drift. Additionally, several approaches [8, 15, 18, 23] integrate text embeddings and U-Net parameters for image customization. These methods fine-tune both components during object learning, achieving impressive results. Considering the substantial cost of fine-tuning for commercial uses, training-based methods [5, 16, 24, 36, 43] have been proposed. This approach primarily utilizes an encoder to extract the object's identity and inject it into the U-Net. For instance, MS-Diffusion [35] integrates grounding tokens with a Resampler to maintain detailed identity and employs layout guidance to explicitly locate the objects. Despite their efficiency during inference, training-based methods often face challenges in balancing identity preservation and text control. A common issue is that the relationships described in text prompts are frequently overlooked. To address this limitation, we introduce DreamRelation, a framework designed to bridge the customization and relation generation." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.599, + 0.484, + 0.902 + ], + "angle": 0, + "content": "Relation-Aware Text-to-Image Generation. Inspired by tuning-based customization methods [7], some recent works [11, 22, 37] represent a \"neglected word\" by learnable parameters. These methods fine-tune part of the parameters on the content co-existing images. For instance, Reversion [11] fine-tunes the text embedding on a set of relation co-existing images and introduces a relation-steering contrastive loss. Reversion gains a better alignment between relational words and generated images by injecting new text embedding into prompts. While effective with prepositions and adjectives, Reversion struggles with relationships that involve significant overlaps, and cannot customize user-provided objects. Moreover, this method still struggles to generate vivid relationships while maintaining the fidelity of multiple customized objects, likely due to the inherent disregard for text embeddings in customization methods. Our work bridges the gap in the predicate relation-following capability of training-based customization methods. This enables efficient inference that naturally and accurately generates relationships in the text prompts." + }, + { + "type": "title", + "bbox": [ + 0.514, + 0.09, + 0.905, + 0.123 + ], + "angle": 0, + "content": "3. Bridging Customization and Relation Generation" + }, + { + "type": "title", + "bbox": [ + 0.514, + 0.132, + 0.694, + 0.147 + ], + "angle": 0, + "content": "3.1. Problem definition" + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.155, + 0.905, + 0.26 + ], + "angle": 0, + "content": "Different from conventional customization tasks, we explore a new setting that focuses on image generation by both image prompts \\( c_{i} \\in \\mathcal{R}^{N \\times 3 \\times H \\times W} \\) and text prompts \\( c_{t} \\). We call this setting relation-aware customized image generation due to the requirements that the generated image \\( \\hat{x} \\) should strictly preserve each identity and keep the relationship among those identities provided by \\( c_{i} \\) and \\( c_{t} \\)." + }, + { + "type": "text", + "bbox": [ + 0.534, + 0.261, + 0.684, + 0.274 + ], + "angle": 0, + "content": "We define this task as:" + }, + { + "type": "equation", + "bbox": [ + 0.659, + 0.285, + 0.905, + 0.3 + ], + "angle": 0, + "content": "\\[\n\\hat {x} = \\Phi_ {\\theta} \\left(c _ {i}, c _ {t}\\right) \\tag {1}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.309, + 0.905, + 0.355 + ], + "angle": 0, + "content": "where \\(\\Phi_{\\theta}\\) is the network parameterized by \\(\\theta\\). For brevity clarification, we set \\(N = 2\\) due to the basic triplet element defined in the DreamRelation." + }, + { + "type": "title", + "bbox": [ + 0.514, + 0.362, + 0.755, + 0.377 + ], + "angle": 0, + "content": "3.2. Discussion and Motivation" + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.384, + 0.907, + 0.536 + ], + "angle": 0, + "content": "While previous works [8, 15, 43] have explored the customization of multiple objects under text control, they lack the ability to effectively manage the relationships between these objects. Specifically, the state-of-the-art multi-object customization model [35] utilizes bounding boxes to enhance multi-object generation. Given the image prompt \\( c_{i} \\) and the text prompt \\( c_{t} \\), it leverages CLIP [28] and the Resampler [35] to extract text tokens \\( tok_{\\text{text}} \\) and image tokens \\( tok_{\\text{image}} \\), respectively. These tokens are then injected into the U-Net \\( \\epsilon_{\\theta} \\) through parallel cross-attention layers:" + }, + { + "type": "equation", + "bbox": [ + 0.532, + 0.542, + 0.905, + 0.575 + ], + "angle": 0, + "content": "\\[\nh = \\operatorname {S o f t m a x} \\left(\\frac {Q K _ {i}}{\\sqrt {d}} + \\mathcal {M}\\right) V _ {i} + \\operatorname {S o f t m a x} \\left(\\frac {Q K _ {t}}{\\sqrt {d}}\\right) V _ {t} \\tag {2}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.581, + 0.906, + 0.716 + ], + "angle": 0, + "content": "where \\(Q\\) represents the query, while \\(K_{i}, V_{i}\\) and \\(K_{t}, V_{t}\\) denote the keys and values obtained from \\(tok_{image}\\) and \\(tok_{text}\\), respectively. The mask \\(\\mathcal{M}\\), generated from the bounding boxes, assists the model in localizing objects during multi-object customization. However, it encounters difficulties when bounding boxes overlap, frequently resulting in object confusion, as shown in Fig. 1. Consequently, relation-aware generation in customized models remains a largely unexplored area of research." + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.717, + 0.906, + 0.822 + ], + "angle": 0, + "content": "On the other hand, we revisit the relation-aware generation task. Existing relation inversion methods [11, 22] focus on inverting a relationship into a text embedding \\( R^{*} \\) within pre-trained text-to-image diffusion models. Given a set of images \\( \\{x_{k}\\}_{k = 1}^{n} \\) that share the same relationship, these methods use a denoising loss \\( \\mathcal{L}_{RI} \\) to fine-tune \\( R^{*} \\), ensuring alignment with the specific relationship:" + }, + { + "type": "equation", + "bbox": [ + 0.586, + 0.831, + 0.905, + 0.848 + ], + "angle": 0, + "content": "\\[\n\\mathcal {L} _ {R I} = \\mathbb {E} _ {z _ {t}, t, \\epsilon , c} [ \\| \\epsilon_ {t} - \\boldsymbol {\\epsilon} _ {\\boldsymbol {\\theta}} (z _ {t}, t, c _ {t}) \\| _ {2} ^ {2} ] \\tag {3}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.856, + 0.906, + 0.901 + ], + "angle": 0, + "content": "Directly applying relation-aware generation methods to customized models leads to suboptimal performance, as illustrated in Fig. 4. We attribute this to two main reasons." + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "15725" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.093, + 0.089, + 0.905, + 0.299 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.089, + 0.304, + 0.908, + 0.348 + ], + "angle": 0, + "content": "Figure 2. The overview of DreamRelation. DreamRelation utilizes the off-the-shelf identity extractor to decouple the relation and identity information in relation-specific images. After getting the U-Net output \\(\\hat{\\epsilon}\\), we predict \\(\\hat{z}_0\\) and calculate the keypoint matching loss. The part in the dotted box is only for training." + }, + { + "type": "text", + "bbox": [ + 0.093, + 0.355, + 0.485, + 0.566 + ], + "angle": 0, + "content": "First, the balance between the control of \\( c_{i} \\) and \\( c_{t} \\) is not addressed in relation-aware generation tasks. During inference, the model tends to overlook the target relationship \\( R^{*} \\) due to the influence of \\( c_{i} \\), resulting in incorrect relationship generation. Second, even if the shared relationship exists across \\( \\{x_{k}\\}_{k = 1}^{n} \\), the disentangled object identity information complicates the learning process. Fine-tuning solely on \\( \\{x_{k}\\}_{k = 1}^{n} \\) makes it challenging for the model to effectively capture the relationship. As shown in the second row of Fig. 4, where ReVersion [11] struggles to generate the \"hugging\" relationship even in the absence of \\( c_{i} \\). In the following subsections, we will introduce our data collection method and DreamRelation to bridge customization and relation generation." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.577, + 0.248, + 0.591 + ], + "angle": 0, + "content": "3.3. Data Collection" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.599, + 0.483, + 0.78 + ], + "angle": 0, + "content": "To address the balance between \\( c_{i} \\) and \\( c_{t} \\) while facilitating relation learning from \\( \\{x_{k}\\}_{k = 1}^{n} \\), we first design a data engine for collecting high-quality tuning data. The ideal tuning data should be in the form of \\( \\mathcal{D} = (x_k, c_i, c_t) \\). The \\( c_{i} \\) contain the object in \\( x_{k} \\) while \\( c_{t} \\) describes the relationship between objects in \\( x_{k} \\). The identity information in \\( c_{i} \\) helps decouple the relationship within \\( x_{k} \\), enhancing relation learning while maintaining a balanced integration of \\( c_{i} \\) and \\( c_{t} \\). Unfortunately, directly cropping the objects from \\( x_{k} \\) to obtain \\( c_{i} \\) results in a copy-and-paste effect [5], even when using data augmentation techniques like flipping and rotation. Therefore, we introduce our data engine below." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.78, + 0.483, + 0.902 + ], + "angle": 0, + "content": "Relation-aware data engine. Fig. 3 illustrates the difference of our data engine. Both of \\( x_{k} \\) and \\( c_{i} \\) are generated by the recent text-to-image generation model, rather than through cropping and augmentation. Specifically, inspired by the performance of DALL-E 3 and its multi-turn dialog capability. We use DALL-E 3 to generate \\( \\mathcal{D} = (x,c_i,c_t) \\). Where \\( x_{k} \\) and \\( c_{i} \\) share the same object identity by leveraging the prompt \"The photo of the same.\" We observe that" + }, + { + "type": "image", + "bbox": [ + 0.52, + 0.352, + 0.924, + 0.601 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.512, + 0.613, + 0.907, + 0.698 + ], + "angle": 0, + "content": "Figure 3. The rigid object cropping data engine, which results in minimal changes to object pose, the cropped image prompts also contain relation information, leading to a copy-and-paste effect, where the text prompt is neglected. Our relation-aware data engine, on the other hand, focuses on relation learning by decoupling identity information from images." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.708, + 0.907, + 0.785 + ], + "angle": 0, + "content": "prompting in this manner allows DALL-E 3 to remember and preserve identity in common categories such as \"tiger\", which is enough for relation learning. Next, we employ X-Pose [41], SAM [13], and LLaVA [19] to annotate \\( x \\), \\( c_i \\) with keypoints, masks, and captions for samples." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.797, + 0.669, + 0.811 + ], + "angle": 0, + "content": "3.4. DreamRelation" + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.821, + 0.688, + 0.836 + ], + "angle": 0, + "content": "3.4.1. Relation Learning" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.841, + 0.907, + 0.902 + ], + "angle": 0, + "content": "To balance \\( c_{i} \\) and \\( c_{t} \\) while enhancing relation learning, we decouple the identity information and relationships in \\( x_{k} \\). Specifically, we leverage an off-the-shelf ID extractor from the state-of-the-art customized model [35]. The identity in-" + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.52, + 0.957 + ], + "angle": 0, + "content": "15726" + } + ], + [ + { + "type": "table", + "bbox": [ + 0.098, + 0.089, + 0.482, + 0.233 + ], + "angle": 0, + "content": "
MethodsSingle-Object CusMulti-Object CusNon-ConfusingRelation-Aware
Textual Inversion [7]✓✗✗✗
DreamBooth [31]✓✗✗✗
Custom Diffusion [15]✓✓✗✗
Mix-of-Show [8]✓✓✗✗
CLIF [18]✓✓✓✗
MultiBooth [45]✓✓✗✗
BLIP-Diffusion [16]✓✗✗✗
ELITE [36]✓✗✗✗
AnyDoor [5]✓✗✗✗
SSR-Encoder [43]✓✓✗✗
λ-eclipse [24]✓✓✗✗
MS-Diffusion [35]✓✓✗✗
ReVersion [11]✗✗✓✓
ADI [10]✗✗✗✓
DreamRelation✓✓✓✓
" + }, + { + "type": "table_caption", + "bbox": [ + 0.091, + 0.238, + 0.483, + 0.281 + ], + "angle": 0, + "content": "Table 1. Setting Comparison for Different Models. We include several representative methods here. Our proposed DreamRelation can meet the demand of the relation-aware customization task." + }, + { + "type": "image", + "bbox": [ + 0.102, + 0.298, + 0.47, + 0.477 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.09, + 0.482, + 0.483, + 0.525 + ], + "angle": 0, + "content": "Figure 4. Comparing our method with ReVersion across different base models, our approach demonstrates superior performance in the relation-aware generation task." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.542, + 0.484, + 0.603 + ], + "angle": 0, + "content": "formation in \\( c_{i} \\) enables the model to focus on the relationship information in \\( x_{k} \\) and \\( c_{t} \\) during relation learning. To achieve this, we employ parallel cross-attention layers to process \\( c_{i} \\) and \\( c_{t} \\), which can be formulated as:" + }, + { + "type": "equation", + "bbox": [ + 0.099, + 0.621, + 0.483, + 0.656 + ], + "angle": 0, + "content": "\\[\nh = \\gamma \\cdot \\operatorname {S o f t m a x} \\left(\\frac {Q K _ {i}}{\\sqrt {d}} + \\mathcal {M}\\right) V _ {i} + \\operatorname {S o f t m a x} \\left(\\frac {Q K _ {t}}{\\sqrt {d}}\\right) V _ {t} \\tag {4}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.674, + 0.483, + 0.902 + ], + "angle": 0, + "content": "where \\( Q = W_{q}h \\), \\( K_{i} = W_{k_{i}}c_{i} \\), \\( V_{i} = W_{v_{i}}c_{i} \\), \\( K_{t} = W_{k_{t}}c_{t} \\), \\( V_{t} = W_{v_{t}}c_{t} \\), and \\( \\gamma \\) is a hyperparameter for scaling control from \\( c_{i} \\). For clarity, we omit the final linear layer \\( W_{out} \\). The relation information mainly depends on text prompts. Therefore, we inject LoRA layers in all of the \\( W_{q} \\), \\( W_{k_{t}} \\), \\( W_{v_{t}} \\), and \\( W_{out} \\) to encourage the model to pay more attention to the relationship in \\( c_{t} \\). We freeze all other parameters during fine-tuning. As shown in Fig. 4, benefiting from the decoupled identity information in \\( c_{i} \\), our method effectively captures the relationship in \\( x_{k} \\) and \\( c_{t} \\), accurately generating the corresponding images. Notably, after fine-tuning, our method can also be directly integrated into SDXL [27], effectively enabling the model to address the relation-aware generation task. The extensive results are presented in the Supplementary Material (SM)." + }, + { + "type": "title", + "bbox": [ + 0.514, + 0.092, + 0.733, + 0.108 + ], + "angle": 0, + "content": "3.4.2. Keypoint Matching Loss" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.111, + 0.907, + 0.231 + ], + "angle": 0, + "content": "Most relationships have specific requirements for the pose, such as \"hugging\" requiring the arms to cross over the object's body, and \"riding a bicycle\" requiring the feet of the object to be on the pedal plate. Therefore, it's important to manipulate the pose of objects in the right status accurately. Intuitively, we introduce explicit supervision in the diffusion latent space during the tuning process, named Keypoint Matching Loss (KML)." + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.232, + 0.907, + 0.367 + ], + "angle": 0, + "content": "Specifically, considering our task only involves common objects instead of humans, we employ X-Pose [41] as a keypoint detector because it can detect any keypoints in complex real-world scenarios. We detect 17 keypoints for each object in \\( x_{k} \\) and \\( c_{i} \\), and denote the keypoint coordinates as \\( c_{kp}^{x_k}, c_{kp}^{c_i} \\in \\mathcal{R}^{17 \\times 2} \\) respectively. To encourage the model to generate accurate poses in \\( \\mathcal{D}(\\hat{z}_0) \\), where \\( \\mathcal{D} \\) is the VAE decoder. We use the U-Net's output \\( \\hat{\\epsilon} \\) to predict \\( z_0 \\) during fine-tuning:" + }, + { + "type": "equation", + "bbox": [ + 0.638, + 0.366, + 0.906, + 0.4 + ], + "angle": 0, + "content": "\\[\n\\hat {z} _ {0} = \\frac {z _ {t} - \\sqrt {1 - \\bar {\\alpha} _ {t}} \\hat {\\epsilon}}{\\sqrt {\\bar {\\alpha} _ {t}}} \\tag {5}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.403, + 0.907, + 0.463 + ], + "angle": 0, + "content": "We use the VAE encoder \\(\\mathcal{E}\\) to obtain the latent representation of image prompts \\(\\mathcal{E}(c_i)\\). Then we calculate the MSE loss on the corresponding keypoint locations between \\(\\mathcal{E}(c_i)\\) and \\(\\hat{z}_0\\):" + }, + { + "type": "equation", + "bbox": [ + 0.558, + 0.471, + 0.907, + 0.504 + ], + "angle": 0, + "content": "\\[\n\\mathcal {L} _ {K M L} = \\frac {1}{n _ {k p}} \\mathbb {E} _ {\\boldsymbol {z} _ {t}, c _ {i}} \\left\\| \\mathcal {E} (c _ {i}) [ c _ {k p} ^ {c _ {i}} ] - \\hat {z} _ {0} [ c _ {k p} ^ {x _ {k}} ] \\right\\| _ {2} ^ {2} \\tag {6}\n\\]" + }, + { + "type": "equation", + "bbox": [ + 0.595, + 0.506, + 0.906, + 0.523 + ], + "angle": 0, + "content": "\\[\n\\mathcal {L} = \\mathcal {L} _ {\\text {d e n o i s e}} + \\lambda \\cdot \\mathcal {L} _ {K M L} \\tag {7}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.532, + 0.906, + 0.639 + ], + "angle": 0, + "content": "where \\( n_{kp} \\) denote the number of keypoint, \\( c_{kp}^{c_i} \\) and \\( c_{kp}^{x_k} \\) are keypoint's coordinates in \\( c_i \\) and \\( x_k \\) respectively. We scale the coordinates to fit the size of \\( \\mathcal{E}(c_i) \\), \\( \\hat{z}_0 \\lambda \\) controls for the relative weight of KML. The part inside the dotted box in Fig. 2 illustrates the model fine-tuning with the KML. We find KML is effective in encouraging the model to manipulate the pose for relation-aware generation." + }, + { + "type": "title", + "bbox": [ + 0.514, + 0.645, + 0.714, + 0.66 + ], + "angle": 0, + "content": "3.4.3. Local Token Injection" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.664, + 0.907, + 0.83 + ], + "angle": 0, + "content": "When generating relationships, local and detailed features are essential—for instance, the feature of arms to accurately construct relationships like \"carrying\". However, the coarse image-level features extracted by the CLIP Image Encoder fail to provide such fine-grained details, leading to confusion between objects in Fig. 7. To address this limitation, we introduce dense features from the CLIP Image Encoder [44]. Specifically, we modify the last layer of the CLIP Image Encoder to obtain dense features: \\( h_{\\mathrm{dense}} = \\) ModifiedAttention(\\( h_{\\mathrm{clip}}' \\)), where \\( h_{\\mathrm{clip}}' \\) represents the hidden states of the second-to-last layer." + }, + { + "type": "equation", + "bbox": [ + 0.616, + 0.84, + 0.907, + 0.857 + ], + "angle": 0, + "content": "\\[\nh _ {\\mathrm {t m p}} = \\operatorname {P r o j} _ {v} \\left(\\operatorname {n o r m} \\left(h _ {\\mathrm {c l i p}} ^ {\\prime}\\right)\\right), \\tag {8}\n\\]" + }, + { + "type": "equation", + "bbox": [ + 0.617, + 0.86, + 0.907, + 0.878 + ], + "angle": 0, + "content": "\\[\nh _ {\\mathrm {t m p}} = h _ {\\mathrm {c l i p}} ^ {\\prime} + \\operatorname {P r o j} _ {\\text {o u t}} \\left(h _ {\\mathrm {t m p}}\\right), \\tag {9}\n\\]" + }, + { + "type": "equation", + "bbox": [ + 0.609, + 0.881, + 0.906, + 0.898 + ], + "angle": 0, + "content": "\\[\nh _ {\\text {d e n s e}} = h _ {\\text {t m p}} + \\operatorname {F F N} \\left(h _ {\\text {t m p}}\\right) \\tag {10}\n\\]" + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "15727" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.175, + 0.089, + 0.823, + 0.284 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.09, + 0.29, + 0.908, + 0.318 + ], + "angle": 0, + "content": "Figure 5. Additional results demonstrate the effectiveness of relation-aware generation. DreamRelation adapts the same pair of objects to different relationships in a natural and accurate manner." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.332, + 0.483, + 0.422 + ], + "angle": 0, + "content": "To further enhance the compatibility between dense features and image-level features, we employ self-distillation on the CLIP Image Encoder [39] to derive our Local Image Encoder. Specifically, we align local regions of the dense features with their corresponding image-level features using cosine similarity." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.424, + 0.484, + 0.514 + ], + "angle": 0, + "content": "During inference, we extract local tokens \\(tok_{\\mathrm{local}}\\) from the dense feature \\(h_{\\mathrm{dense}}\\) through partitioning and pooling. These local tokens are then concatenated with the image-level tokens \\(tok_{\\mathrm{image}}\\) and passed to the ID extractor. The ID extractor, which is a transformer-based architecture as illustrated in Fig. 2 (b), can be written as:" + }, + { + "type": "equation", + "bbox": [ + 0.116, + 0.527, + 0.483, + 0.546 + ], + "angle": 0, + "content": "\\[\nq = q + \\text {A t t e n t i o n} (\\text {c o n c a t} [ q, \\text {t o k} _ {\\text {i m a g e}}, \\text {t o k} _ {\\text {l o c a l}} ]) \\tag {11}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.558, + 0.484, + 0.679 + ], + "angle": 0, + "content": "Regarding the compatibility between CLIP dense features and the original image-level features, experimental results demonstrate that our Local Image Encoder outperforms other dense representations [3]. For a detailed discussion, we refer readers to the Supplementary Material. Additionally, the injection of local tokens enhances identity preservation, as evidenced by improvements in evaluation metrics and visual results, shown in Fig. 7 and Tab. 3." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.694, + 0.224, + 0.71 + ], + "angle": 0, + "content": "4. Experiments" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.72, + 0.484, + 0.901 + ], + "angle": 0, + "content": "Implementation Details. We generate a tuning set for relation learning with 4-6 samples per relationship. We incorporate LoRA layers into all text cross-attention layers of the U-Net. We set the LoRA rank to \\( r = 4 \\), the parallel cross-attention scaling factor to \\( \\gamma = 0.6 \\), and the keypoint matching loss weight to \\( \\lambda = 1e - 3 \\). The model is finetuned for 500 steps, using 2 A100 GPUs, with a total batch size of 8, completing the process in 10 minutes. We use the Adam optimizer [12] with a learning rate of \\( 1e - 4 \\) and no weight decay, resulting in a total of 3.1M trainable parameters. Note that our learnable parameters are only in the text cross-attention layers, making it compatible with" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.332, + 0.907, + 0.422 + ], + "angle": 0, + "content": "any SDXL-based models. Considering the strong identity preservation capabilities of MS-Diffusion, we implement our DreamRelation on MS-Diffusion. During fine-tuning and inference, we concatenate local tokens with image tokens. We put more self-distillation details of the Local Image Encoder in the Supplementary Material." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.424, + 0.909, + 0.727 + ], + "angle": 0, + "content": "Evaluation. To evaluate relation-aware customized image generation, we propose RelationBench, consisting of 44 objects from DreamBench [31] and CustomConcept101 [15], along with 25 relationships. The object categories include pets, plushies, toys, people, and cartoons. Using GPT-4, we generate 100 cases for single-object and multi-object evaluations. Following previous works [31, 35], we evaluate our method on three metrics: (1) Identity Preservation, which assesses the alignment between the generated images and the image prompts, using the CLIP image score and DINO score that calculate the cosine similarity between the class embeddings of images, referred to as CLIP-I and DINO, respectively; (2) Text Alignment, which evaluates how well the generated images align with the text prompts, using the CLIP image-text score, denoted as CLIP-T; and (3) Relation Alignment. We observed that nouns in the prompts can inflate CLIP-T scores, making them less accurate for evaluating relation generation. To address this, we extract the relation word from the text prompt using spaCy and calculate the CLIP image-text score, denoted as CLIP-R." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.74, + 0.655, + 0.755 + ], + "angle": 0, + "content": "4.1. Main Results" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.763, + 0.907, + 0.884 + ], + "angle": 0, + "content": "We evaluate our method through both quantitative and qualitative results in single- and multi-object cases. In single-object cases, we primarily focus on the alignment of the object's pose with the relationship in the text prompt and the model's ability to preserve identity. In multi-object cases, in addition to relation generation and identity preservation, we also need to ensure no confusion between the multiple objects." + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.886, + 0.907, + 0.901 + ], + "angle": 0, + "content": "Quantitative results. We compare our method with" + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "15728" + } + ], + [ + { + "type": "image_caption", + "bbox": [ + 0.097, + 0.091, + 0.142, + 0.104 + ], + "angle": 0, + "content": "Objects" + }, + { + "type": "image", + "bbox": [ + 0.097, + 0.106, + 0.144, + 0.176 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.165, + 0.092, + 0.21, + 0.103 + ], + "angle": 0, + "content": "Prompts" + }, + { + "type": "image_caption", + "bbox": [ + 0.159, + 0.124, + 0.217, + 0.157 + ], + "angle": 0, + "content": "A cat is kissing with a dog." + }, + { + "type": "image_caption", + "bbox": [ + 0.156, + 0.192, + 0.221, + 0.236 + ], + "angle": 0, + "content": "A dog is shaking hands with a plushie bear." + }, + { + "type": "image_caption", + "bbox": [ + 0.156, + 0.264, + 0.22, + 0.308 + ], + "angle": 0, + "content": "A dog is fencing with a plushie panda." + }, + { + "type": "image_caption", + "bbox": [ + 0.265, + 0.092, + 0.29, + 0.102 + ], + "angle": 0, + "content": "Ours" + }, + { + "type": "image", + "bbox": [ + 0.231, + 0.106, + 0.323, + 0.175 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.333, + 0.092, + 0.412, + 0.102 + ], + "angle": 0, + "content": "ReVersion+MS" + }, + { + "type": "image", + "bbox": [ + 0.327, + 0.106, + 0.418, + 0.174 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.429, + 0.092, + 0.5, + 0.102 + ], + "angle": 0, + "content": "MS-Diffusion" + }, + { + "type": "image", + "bbox": [ + 0.422, + 0.106, + 0.511, + 0.174 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.525, + 0.092, + 0.595, + 0.102 + ], + "angle": 0, + "content": "SSR-Encoder" + }, + { + "type": "image", + "bbox": [ + 0.515, + 0.106, + 0.605, + 0.174 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.633, + 0.092, + 0.679, + 0.103 + ], + "angle": 0, + "content": "\\(\\lambda\\) -eclipse" + }, + { + "type": "image", + "bbox": [ + 0.61, + 0.106, + 0.7, + 0.174 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.702, + 0.092, + 0.903, + 0.102 + ], + "angle": 0, + "content": "Custom Diffusion DreamBooth (SDXL)" + }, + { + "type": "image", + "bbox": [ + 0.61, + 0.178, + 0.7, + 0.247 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.704, + 0.106, + 0.794, + 0.174 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.704, + 0.178, + 0.794, + 0.247 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.798, + 0.106, + 0.887, + 0.174 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.704, + 0.251, + 0.794, + 0.32 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.798, + 0.178, + 0.887, + 0.247 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.798, + 0.251, + 0.887, + 0.32 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.09, + 0.33, + 0.907, + 0.372 + ], + "angle": 0, + "content": "Figure 6. We compare multi-object performance with training-based and tuning-based methods. Additionally, we integrate ReVersion [11] into MS-Diffusion, where the learnable text token is injected into the text prompt during MS-Diffusion's inference, resulting in ReVersion+MS. Our method shows a clear advantage in relation-aware generation and avoiding object confusion in overlapping scenarios." + }, + { + "type": "table", + "bbox": [ + 0.13, + 0.379, + 0.861, + 0.572 + ], + "angle": 0, + "content": "
MethodSingle-objectMulti-object
CLIP-TCLIP-RCLIP-IDINOCLIP-TCLIP-RCLIP-IDINO
DreamBooth (SDXL) [31]27.818.274.262.624.316.267.857.2
Custom Diffusion [15]26.515.273.258.820.115.464.755.3
Cones-V2 [20]24.413.572.157.221.315.264.354.2
ELITE [36]25.714.975.461.5————
AnyDoor [5]24.514.777.462.221.614.969.759.8
BLIP-Diffusion [16]26.215.777.457.7————
SSR-Encoder [43]25.515.980.459.424.214.672.156.2
MS-Diffusion [35]26.518.878.764.526.918.973.858.8
Reversion+MS [11]27.819.377.863.127.219.274.259.7
Ours30.621.477.963.428.920.475.462.1
" + }, + { + "type": "table_caption", + "bbox": [ + 0.137, + 0.576, + 0.858, + 0.591 + ], + "angle": 0, + "content": "Table 2. Quantitative comparison on RelationBench, Bold and underline represent the highest and second-highest metrics." + }, + { + "type": "table", + "bbox": [ + 0.094, + 0.603, + 0.475, + 0.684 + ], + "angle": 0, + "content": "
MethodMulti-object
CLIP-TCLIP-RCLIP-IDINO
w/o Relation-aware Data27.319.475.359.8
w/o Local Token Injection28.519.575.159.9
w/o Keypoint Matching Loss27.419.275.261.2
Full Model28.920.475.462.1
" + }, + { + "type": "table_caption", + "bbox": [ + 0.128, + 0.688, + 0.446, + 0.702 + ], + "angle": 0, + "content": "Table 3. Ablation study on our proposed components" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.72, + 0.484, + 0.901 + ], + "angle": 0, + "content": "baseline models across three benchmarks: Relation-Bench, DreamBench, and multi-object cases in Custom-Concept101. As shown in Tab. 2, our method demonstrates a clear advantage on RelationBench for single-object cases, particularly in the CLIP-T and CLIP-R metrics. However, its performance is slightly lower in the CLIP-I metric, likely due to significant pose variations in the objects, which may decrease the CLIP-I score. For multi-object generation, our method outperforms other approaches across all evaluation metrics, which we attribute to its ability to effectively avoid object confusion and significantly enhance CLIP-I performance. Furthermore, as shown in Tab. 4, we evaluate" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.605, + 0.907, + 0.725 + ], + "angle": 0, + "content": "single-object generation performance on DreamBench. Our method outperforms others in the CLIP-T metric while delivering competitive results in CLIP-I and DINO. We conduct experiments on multi-object cases from CustomConcept101, which we denote as M-CustomConcept101. As shown in Fig. 5, our model achieves the highest performance in the DINO score and ranks second in both the CLIP-T and CLIP-I metrics." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.728, + 0.907, + 0.833 + ], + "angle": 0, + "content": "Qualitative Comparison. We conduct qualitative comparison experiments on RelationBench to evaluate relation-aware customized image generation. As shown in Fig. 6, our method outperforms others in generating multi-object relationships, effectively avoiding object confusion while accurately capturing the intended relationships. Additional results can be found in the SM." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.847, + 0.766, + 0.864 + ], + "angle": 0, + "content": "4.2. Ablation study and Analysis" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.871, + 0.906, + 0.902 + ], + "angle": 0, + "content": "Relation-Aware Data Engine. Instead of fine-tuning on our curated \\((x_{k},c_{i},c_{t})\\) pairs. We train a text embedding \\(R^{*}\\)" + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "15729" + } + ], + [ + { + "type": "image_caption", + "bbox": [ + 0.102, + 0.089, + 0.167, + 0.103 + ], + "angle": 0, + "content": "a) objects" + }, + { + "type": "image_caption", + "bbox": [ + 0.227, + 0.089, + 0.277, + 0.101 + ], + "angle": 0, + "content": "wKML" + }, + { + "type": "image_caption", + "bbox": [ + 0.368, + 0.089, + 0.428, + 0.101 + ], + "angle": 0, + "content": "w/o KML" + }, + { + "type": "image", + "bbox": [ + 0.108, + 0.106, + 0.18, + 0.216 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.181, + 0.106, + 0.323, + 0.216 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.323, + 0.106, + 0.467, + 0.217 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.186, + 0.218, + 0.389, + 0.231 + ], + "angle": 0, + "content": "A cat is shaking hands with a dog." + }, + { + "type": "image_caption", + "bbox": [ + 0.102, + 0.232, + 0.167, + 0.248 + ], + "angle": 0, + "content": "b) objects" + }, + { + "type": "image_caption", + "bbox": [ + 0.216, + 0.236, + 0.304, + 0.248 + ], + "angle": 0, + "content": "w local tokens" + }, + { + "type": "image_caption", + "bbox": [ + 0.347, + 0.236, + 0.446, + 0.247 + ], + "angle": 0, + "content": "w/o local tokens" + }, + { + "type": "image", + "bbox": [ + 0.109, + 0.252, + 0.182, + 0.309 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.182, + 0.252, + 0.325, + 0.363 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.325, + 0.252, + 0.468, + 0.363 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.216, + 0.365, + 0.366, + 0.378 + ], + "angle": 0, + "content": "A dog is carrying a sloth." + }, + { + "type": "image_caption", + "bbox": [ + 0.09, + 0.383, + 0.483, + 0.424 + ], + "angle": 0, + "content": "Figure 7. Ablation study of keypoint matching loss (KML) and local tokens: The images are generated using the same random seed. We use green and red boxes to highlight the main differences." + }, + { + "type": "image", + "bbox": [ + 0.104, + 0.428, + 0.131, + 0.496 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.131, + 0.429, + 0.184, + 0.497 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.184, + 0.431, + 0.212, + 0.497 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.212, + 0.431, + 0.296, + 0.497 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.296, + 0.431, + 0.353, + 0.497 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.353, + 0.431, + 0.459, + 0.497 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.132, + 0.499, + 0.275, + 0.505 + ], + "angle": 0, + "content": "A dog, a cat, a plushie bear hug with each other." + }, + { + "type": "image_caption", + "bbox": [ + 0.285, + 0.499, + 0.469, + 0.506 + ], + "angle": 0, + "content": "A dog, a cat, and a plushie bear are fighting with one another." + }, + { + "type": "image_caption", + "bbox": [ + 0.09, + 0.512, + 0.483, + 0.54 + ], + "angle": 0, + "content": "Figure 8. DreamRelation for three objects. Compared to the baseline model, DreamRelation shows a clear advantage." + }, + { + "type": "text", + "bbox": [ + 0.089, + 0.553, + 0.483, + 0.629 + ], + "angle": 0, + "content": "using a set of relation-specific data \\((x_{k},c_{i}^{\\prime},c_{t})\\), where \\(c_{i}^{\\prime}\\) is cropped from \\(x_{k}\\). The well-trained text embedding \\(R^{*}\\) is injected into the text embedding during inference. The quantitative results in Tab. 3 show that our relation-aware data engine outperforms in both CLIP-T and CLIP-R scores." + }, + { + "type": "text", + "bbox": [ + 0.089, + 0.63, + 0.483, + 0.749 + ], + "angle": 0, + "content": "Keypoint Matching Loss (KML). We omit KML during fine-tuning. As shown in Fig. 7 a), including the KML leads to more accurate and natural relation-aware customized images using the same random seed. Quantitative results, presented in Tab. 3, show a decline across all evaluation metrics when the KML is removed. Additionally, we experimented with different values for \\(\\lambda\\) and found that the best performance occurs at \\(\\lambda = 1e - 3\\)." + }, + { + "type": "text", + "bbox": [ + 0.089, + 0.75, + 0.483, + 0.902 + ], + "angle": 0, + "content": "Local Token Injection. For local token injection, the visual comparison in Fig. 7 b) shows severe object confusion when local tokens are omitted. The quantitative results in Tab. 3 demonstrate that local tokens improve relation-aware customization. We conducted an ablation study on the number of local tokens, as presented in Tab. 6, and determined the optimal number to be 16. Additional ablation studies are shown in the SM. Combining these methods ensures that DreamRelation achieves high-quality relation-aware customized generation." + }, + { + "type": "table", + "bbox": [ + 0.55, + 0.089, + 0.865, + 0.219 + ], + "angle": 0, + "content": "
MethodSingle-object
CLIP-TCLIP-IDINO
DreamBooth (SDXL) [31]31.281.569.2
Custom Diffusion [15]28.477.266.8
Cones-V2 [20]31.076.567.2
ELITE [36]29.877.462.5
AnyDoor [5]25.582.167.8
SSR-Encoder [43]30.882.161.2
MS-Diffusion [35]31.579.368.2
Ours31.681.768.4
" + }, + { + "type": "table_caption", + "bbox": [ + 0.576, + 0.224, + 0.844, + 0.238 + ], + "angle": 0, + "content": "Table 4. Evaluation Results on DreamBench." + }, + { + "type": "table", + "bbox": [ + 0.55, + 0.24, + 0.865, + 0.372 + ], + "angle": 0, + "content": "
MethodMulti-object
CLIP-TCLIP-IDINO
DreamBooth (SDXL) [31]29.567.449.2
Custom Diffusion [15]28.166.248.8
Cones-V2 [20]29.266.547.2
AnyDoor [5]20.272.151.2
SSR-Encoder [43]30.671.152.2
λ-eclipse [24]29.268.248.2
MS-Diffusion [35]28.070.251.2
Ours29.771.452.3
" + }, + { + "type": "table_caption", + "bbox": [ + 0.547, + 0.376, + 0.872, + 0.39 + ], + "angle": 0, + "content": "Table 5. Evaluation Results on M-CustomConcept101." + }, + { + "type": "table", + "bbox": [ + 0.542, + 0.393, + 0.872, + 0.462 + ], + "angle": 0, + "content": "
Num Local TokensMulti-object
CLIP-TCLIP-RCLIP-IDINO
N=2×228.319.475.360.1
N=4×428.920.475.462.1
N=8×827.918.174.159.4
" + }, + { + "type": "table_caption", + "bbox": [ + 0.549, + 0.466, + 0.871, + 0.481 + ], + "angle": 0, + "content": "Table 6. Ablation study on the number of local tokens." + }, + { + "type": "table", + "bbox": [ + 0.552, + 0.484, + 0.862, + 0.561 + ], + "angle": 0, + "content": "
LambdaMulti-object
CLIP-TCLIP-RCLIP-IDINO
λ=1e-227.519.372.659.6
λ=1e-328.920.475.462.1
λ=1e-426.818.173.960.4
" + }, + { + "type": "table_caption", + "bbox": [ + 0.514, + 0.565, + 0.903, + 0.58 + ], + "angle": 0, + "content": "Table 7. Ablation study on \\( \\lambda \\) ,controlling relative weight of KML." + }, + { + "type": "title", + "bbox": [ + 0.514, + 0.593, + 0.634, + 0.608 + ], + "angle": 0, + "content": "5. Conclusion" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.618, + 0.907, + 0.845 + ], + "angle": 0, + "content": "In this work, we introduce a new challenging task: relation-aware customized image generation. This task aims to generate objects that adhere to the relationship specified in the text prompt and preserve the identities of the user-provided images. To support this, we propose a data engine for generating high-quality fine-tuning data. Our method, DreamRelation, incorporates Keypoint Matching Loss and Local Token Injection, effectively capturing relation information and generating natural relations between customized objects. We also present a fair comparison with other methods on a new benchmark, RelationBench. Extensive experiments demonstrate the effectiveness of our method, highlighting its potential for applications in interactive scenario generation, relation detection, and more. Our research can inspire the direction of relation-aware generation in the community." + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.845, + 0.907, + 0.888 + ], + "angle": 0, + "content": "Acknowledgement. This work is supported by the National Key Research and Development Program of China (No. 2023YFC3807600)." + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.52, + 0.957 + ], + "angle": 0, + "content": "15730" + } + ], + [ + { + "type": "title", + "bbox": [ + 0.093, + 0.09, + 0.188, + 0.106 + ], + "angle": 0, + "content": "References" + }, + { + "type": "ref_text", + "bbox": [ + 0.101, + 0.115, + 0.484, + 0.184 + ], + "angle": 0, + "content": "[1] Jinbin Bai, Tian Ye, Wei Chow, Enxin Song, Qing-Guo Chen, Xiangtai Li, Zhen Dong, Lei Zhu, and Shuicheng Yan. Meissonic: Revitalizing masked generative transformers for efficient high-resolution text-to-image synthesis. arXiv preprint arXiv:2410.08261, 2024. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.101, + 0.185, + 0.484, + 0.242 + ], + "angle": 0, + "content": "[2] James Betker, Gabriel Goh, Li Jing, Tim Brooks, Jianfeng Wang, Linjie Li, Long Ouyang, Juntang Zhuang, Joyce Lee, Yufei Guo, et al. Improving image generation with better captions. Computer Science., 2:3, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.242, + 0.483, + 0.297 + ], + "angle": 0, + "content": "[3] Mathilde Caron, Hugo Touvron, Ishan Misra, Herve Jégou, Julien Mairal, Piotr Bojanowski, and Armand Joulin. Emerging properties in self-supervised vision transformers. In Int. Conf. Comput. Vis., 2021. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.298, + 0.483, + 0.366 + ], + "angle": 0, + "content": "[4] Junsong Chen, Jincheng Yu, Chongjian Ge, Lewei Yao, Enze Xie, Yue Wu, Zhongdao Wang, James Kwok, Ping Luo, Huchuan Lu, et al. Pixart-α: Fast training of diffusion transformer for photorealistic text-to-image synthesis. In ICLR, 2024. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.368, + 0.483, + 0.41 + ], + "angle": 0, + "content": "[5] Xi Chen, Lianghua Huang, Yu Liu, Yujun Shen, Deli Zhao, and Hengshuang Zhao. Anydoor: Zero-shot object-level image customization. In CVPR, 2024. 2, 3, 4, 5, 7, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.411, + 0.483, + 0.478 + ], + "angle": 0, + "content": "[6] Hyung Won Chung, Le Hou, Shayne Longpre, Barret Zoph, Yi Tay, William Fedus, Yunxuan Li, Xuezhi Wang, Mostafa Dehghani, Siddhartha Brahma, et al. Scaling instruction-finetuned language models. arXiv preprint arXiv:2210.11416, 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.48, + 0.483, + 0.535 + ], + "angle": 0, + "content": "[7] Rinon Gal, Yuval Alaluf, Yuval Atzmon, Or Patashnik, Amit Haim Bermano, Gal Chechik, and Daniel Cohen-Or. An image is worth one word: Personalizing text-to-image generation using textual inversion. In ICLR, 2023. 3, 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.537, + 0.483, + 0.605 + ], + "angle": 0, + "content": "[8] Yuchao Gu, Xintao Wang, Jay Zhangjie Wu, Yujun Shi, Yunpeng Chen, Zihan Fan, Wuyou Xiao, Rui Zhao, Shuning Chang, Weijia Wu, et al. Mix-of-show: Decentralized low-rank adaptation for multi-concept customization of diffusion models. NeurIPS, 2024. 2, 3, 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.607, + 0.483, + 0.661 + ], + "angle": 0, + "content": "[9] Edward J Hu, Yelong Shen, Phillip Wallis abd Zeyuan Allen-Zhu, Yuanzhi Li, Shean Wang, Lu Wang, and Weizhu Chen. Lora: Low-rank adaptation of large language models. In ICLR, 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.663, + 0.483, + 0.717 + ], + "angle": 0, + "content": "[10] Siteng Huang, Biao Gong, Yutong Feng, Xi Chen, Yuqian Fu, Yu Liu, and Donglin Wang. Learning disentangled identifiers for action-customized text-to-image generation. In CVPR, 2024. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.719, + 0.483, + 0.773 + ], + "angle": 0, + "content": "[11] Ziqi Huang, Tianxing Wu, Yuming Jiang, Kelvin CK Chan, and Ziwei Liu. Reversion: Diffusion-based relation inversion from images. arXiv preprint arXiv:2303.13495, 2023. 2, 3, 4, 5, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.775, + 0.483, + 0.815 + ], + "angle": 0, + "content": "[12] Diederik P Kingma and Jimmy Ba. Adam: A method for stochastic optimization. arXiv preprint arXiv:1412.6980, 2014. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.817, + 0.483, + 0.872 + ], + "angle": 0, + "content": "[13] Alexander Kirillov, Eric Mintun, Nikhila Ravi, Hanzi Mao, Chloe Rolland, Laura Gustafson, Tete Xiao, Spencer Whitehead, Alexander C. Berg, Wan-Yen Lo, Piotr Dollar, and Ross B. Girshick. Segment anything. In ICCV, 2023. 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.873, + 0.483, + 0.903 + ], + "angle": 0, + "content": "[14] Lingjie Kong, Kai Wu, Xiaobin Hu, Wenhui Han, Jinlong Peng, Chengming Xu, Donghao Luo, Jiangning Zhang," + }, + { + "type": "list", + "bbox": [ + 0.094, + 0.115, + 0.484, + 0.903 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.545, + 0.093, + 0.905, + 0.135 + ], + "angle": 0, + "content": "Chengjie Wang, and Yanwei Fu. Anymaker: Zero-shot general object customization via decoupled dual-level id injection. arXiv preprint arXiv:2406.11643, 2024. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.136, + 0.905, + 0.178 + ], + "angle": 0, + "content": "[15] Nupur Kumari, Bingliang Zhang, Richard Zhang, Eli Shechtman, and Jun-Yan Zhu. Multi-concept customization of text-to-image diffusion. In CVPR, 2023. 2, 3, 5, 6, 7, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.179, + 0.905, + 0.233 + ], + "angle": 0, + "content": "[16] Dongxu Li, Junnan Li, and Steven C. H. Hoi. Blip-diffusion: Pre-trained subject representation for controllable text-to-image generation and editing. In NeurIPS, 2023. 2, 3, 5, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.235, + 0.905, + 0.291 + ], + "angle": 0, + "content": "[17] Long Lian, Boyi Li, Adam Yala, and Trevor Darrell. Llm-grounded diffusion: Enhancing prompt understanding of text-to-image diffusion models with large language models. arXiv preprint arXiv:2305.13655, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.293, + 0.905, + 0.359 + ], + "angle": 0, + "content": "[18] Wang Lin, Jingyuan Chen, Jiaxin Shi, Yichen Zhu, Chen Liang, Junzhong Miao, Tao Jin, Zhou Zhao, Fei Wu, Shuicheng Yan, et al. Non-confusing generation of customized concepts in diffusion models. ICML, 2024. 2, 3, 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.362, + 0.904, + 0.39 + ], + "angle": 0, + "content": "[19] Haotian Liu, Chunyuan Li, Qingyang Wu, and Yong Jae Lee. Visual instruction tuning. NeurIPS, 2024. 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.392, + 0.905, + 0.447 + ], + "angle": 0, + "content": "[20] Zhiheng Liu, Yifei Zhang, Yujun Shen, Kecheng Zheng, Kai Zhu, Ruili Feng, Yu Liu, Deli Zhao, Jingren Zhou, and Yang Cao. Cones 2: Customizable image synthesis with multiple subjects. In NeurIPS, 2023. 2, 3, 7, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.449, + 0.905, + 0.503 + ], + "angle": 0, + "content": "[21] Joanna Materzynska, Josef Sivic, Eli Shechtman, Antonio Torralba, Richard Zhang, and Bryan Russell. Customizing motion in text-to-video diffusion models. arXiv preprint arXiv:2312.04966, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.505, + 0.905, + 0.573 + ], + "angle": 0, + "content": "[22] mengmeng Ge, Xu Jia, Takashi Isobe, Xiaomin Li, Qinghe Wang, Jing Mu, Dong Zhou, liwang Amd, Huchuan Lu, Lu Tian, Ashish Sirasao, and Emad Barsoum. Customizing text-to-image generation with inverted interaction. In ACM MM, 2024. 2, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.576, + 0.905, + 0.631 + ], + "angle": 0, + "content": "[23] Lianyu Pang, Jian Yin, Baoquan Zhao, Feize Wu, Fu Lee Wang, Qing Li, and Xudong Mao. Attndreambooth: Towards text-aligned personalized text-to-image generation. arXiv preprint arXiv:2406.05000, 2024. 2, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.633, + 0.905, + 0.688 + ], + "angle": 0, + "content": "[24] Maitreya Patel, Sangmin Jung, Chitta Baral, and Yezhou Yang. λ-eclipse: Multi-concept personalized text-to-image diffusion models by leveraging clip latent space. arXiv preprint arXiv:2402.05195, 2024. 2, 3, 5, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.689, + 0.905, + 0.716 + ], + "angle": 0, + "content": "[25] William Peebles and Saining Xie. Scalable diffusion models with transformers. In ICCV, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.718, + 0.905, + 0.773 + ], + "angle": 0, + "content": "[26] Pablo Pernias, Dominic Rampas, Mats Leon Richter, Christopher Pal, and Marc Aubreville. Würstchen: An efficient architecture for large-scale text-to-image diffusion models. In ICLR, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.775, + 0.905, + 0.843 + ], + "angle": 0, + "content": "[27] Dustin Podell, Zion English, Kyle Lacey, Andreas Blattmann, Tim Dockhorn, Jonas Müller, Joe Penna, and Robin Rombach. Sdxl: Improving latent diffusion models for high-resolution image synthesis. arXiv preprint arXiv:2307.01952, 2023. 2, 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.846, + 0.905, + 0.901 + ], + "angle": 0, + "content": "[28] Alec Radford, Jong Wook Kim, Chris Hallacy, Aditya Ramesh, Gabriel Goh, Sandhini Agarwal, Girish Sastry, Amanda Askell, Pamela Mishkin, Jack Clark, Gretchen Krueger, and Ilya Sutskever. Learning transferable visual" + }, + { + "type": "list", + "bbox": [ + 0.517, + 0.093, + 0.905, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.518, + 0.957 + ], + "angle": 0, + "content": "15731" + } + ], + [ + { + "type": "ref_text", + "bbox": [ + 0.125, + 0.092, + 0.482, + 0.119 + ], + "angle": 0, + "content": "models from natural language supervision. In ICML, 2021. 2,3" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.122, + 0.483, + 0.178 + ], + "angle": 0, + "content": "[29] Jingjing Ren, Wenbo Li, Haoyu Chen, Renjing Pei, Bin Shao, Yong Guo, Long Peng, Fenglong Song, and Lei Zhu. Ultrapixel: Advancing ultra-high-resolution image synthesis to new peaks. arXiv preprint arXiv:2407.02158, 2024. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.181, + 0.482, + 0.221 + ], + "angle": 0, + "content": "[30] Robin Rombach, Andreas Blattmann, Dominik Lorenz, Patrick Esser, and Björn Ommer. High-resolution image synthesis with latent diffusion models. In CVPR, 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.224, + 0.482, + 0.279 + ], + "angle": 0, + "content": "[31] Nataniel Ruiz, Yuanzhen Li, Varun Jampani, Yael Pritch, Michael Rubinstein, and Kfir Aberman. Dreambooth: Fine tuning text-to-image diffusion models for subject-driven generation. In CVPR, 2023. 2, 3, 5, 6, 7, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.282, + 0.482, + 0.365 + ], + "angle": 0, + "content": "[32] Chitwan Sahara, William Chan, Saurabh Saxena, Lala Li, Jay Whang, Emily L. Denton, Seyed Kamyar Seyed Ghasemipour, Raphael Gontijo Lopes, Burcu Karagol Ayan, Tim Salimans, Jonathan Ho, David J. Fleet, and Mohammad Norouzi. Photorealistic text-to-image diffusion models with deep language understanding. In NeurIPS, 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.368, + 0.482, + 0.408 + ], + "angle": 0, + "content": "[33] Ashish Vaswani, Noam Shazeer, Niki Parmar, Jakob Uszkoreit, Llion Jones, Aidan N Gomez, Lukasz Kaiser, and Illia Polosukhin. Attention is all you need. In NeurIPS, 2017. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.412, + 0.482, + 0.464 + ], + "angle": 0, + "content": "[34] Andrey Voynov, Qinghao Chu, Daniel Cohen-Or, and Kfir Aberman. p+: Extended textual conditioning in text-to-image generation. arXiv preprint arXiv:2303.09522, 2023. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.469, + 0.482, + 0.524 + ], + "angle": 0, + "content": "[35] X Wang, Siming Fu, Qihan Huang, Wanggui He, and Hao Jiang. Ms-diffusion: Multi-subject zero-shot image personalization with layout guidance. arXiv preprint arXiv:2406.07209, 2024. 2, 3, 4, 5, 6, 7, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.527, + 0.482, + 0.581 + ], + "angle": 0, + "content": "[36] Yuxiang Wei, Yabo Zhang, Zhilong Ji, Jinfeng Bai, Lei Zhang, and Wangmeng Zuo. ELITE: encoding visual concepts into textual embeddings for customized text-to-image generation. In ICCV, 2023. 2, 3, 5, 7, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.585, + 0.482, + 0.64 + ], + "angle": 0, + "content": "[37] Yujie Wei, Shiwei Zhang, Zhiwu Qing, Hangjie Yuan, Zhiheng Liu, Yu Liu, Yingya Zhang, Jingren Zhou, and Hongming Shan. Dreamvideo: Composing your dream videos with customized subject and motion. In CVPR, 2024. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.643, + 0.482, + 0.697 + ], + "angle": 0, + "content": "[38] Jianzong Wu, Xiangtai Li, Yanhong Zeng, Jiangning Zhang, Qianyu Zhou, Yining Li, Yunhai Tong, and Kai Chen. Motionbooth: Motion-aware customized text-to-video generation. arXiv preprint arXiv:2406.17758, 2024. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.7, + 0.482, + 0.754 + ], + "angle": 0, + "content": "[39] Size Wu, Wenwei Zhang, Lumin Xu, Sheng Jin, Xiangtai Li, Wentao Liu, and Chen Change Loy. Clipself: Vision transformer distills itself for open-vocabulary dense prediction. ICLR, 2024. 2, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.758, + 0.482, + 0.812 + ], + "angle": 0, + "content": "[40] Binxin Yang, Shuyang Gu, Bo Zhang, Ting Zhang, Xuejin Chen, Xiaoyan Sun, Dong Chen, and Fang Wen. Paint by example: Exemplar-based image editing with diffusion models. In CVPR, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.816, + 0.482, + 0.843 + ], + "angle": 0, + "content": "[41] Jie Yang, Ailing Zeng, Ruimao Zhang, and Lei Zhang. Unipose: Detecting any keypoints. ECCV, 2024. 4, 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.846, + 0.482, + 0.899 + ], + "angle": 0, + "content": "[42] Hu Ye, Jun Zhang, Sibo Liu, Xiao Han, and Wei Yang. Ip-adapter: Text compatible image prompt adapter for text-to-image diffusion models. arXiv preprint arxiv:2308.06721, 2023. 2" + }, + { + "type": "list", + "bbox": [ + 0.093, + 0.092, + 0.483, + 0.899 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.516, + 0.092, + 0.906, + 0.147 + ], + "angle": 0, + "content": "[43] Yuxuan Zhang, Yiren Song, Jiaming Liu, Rui Wang, Jinpeng Yu, Hao Tang, Huaxia Li, Xu Tang, Yao Hu, Han Pan, et al. Ssr-encoder: Encoding selective subject representation for subject-driven generation. In CVPR, 2024. 2, 3, 5, 7, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.516, + 0.149, + 0.905, + 0.176 + ], + "angle": 0, + "content": "[44] Chong Zhou, Chen Change Loy, and Bo Dai. Extract free dense labels from clip. In ECCV, 2022. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.179, + 0.905, + 0.22 + ], + "angle": 0, + "content": "[45] Chenyang Zhu, Kai Li, Yue Ma, Chunming He, and Li Xiu. Multibooth: Towards generating all your concepts in an image from text. arXiv preprint arXiv:2404.14239, 2024. 5" + }, + { + "type": "list", + "bbox": [ + 0.516, + 0.092, + 0.906, + 0.22 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.946, + 0.519, + 0.957 + ], + "angle": 0, + "content": "15732" + } + ] +] \ No newline at end of file diff --git a/2025/DreamRelation_ Bridging Customization and Relation Generation/dc0bd82f-ae71-4d19-a1ab-522a355a226a_origin.pdf b/2025/DreamRelation_ Bridging Customization and Relation Generation/dc0bd82f-ae71-4d19-a1ab-522a355a226a_origin.pdf new file mode 100644 index 0000000000000000000000000000000000000000..8314eccaa2c78324257171139f6fc368a7025607 --- /dev/null +++ b/2025/DreamRelation_ Bridging Customization and Relation Generation/dc0bd82f-ae71-4d19-a1ab-522a355a226a_origin.pdf @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:7769080ee3997affa7d3243568dd46ea7565440f2e29f04641a6b775b9e448b7 +size 1678870 diff --git a/2025/DreamRelation_ Bridging Customization and Relation Generation/full.md b/2025/DreamRelation_ Bridging Customization and Relation Generation/full.md new file mode 100644 index 0000000000000000000000000000000000000000..db0d8f85d52c5539326e8ac4ba9c569d4953eb17 --- /dev/null +++ b/2025/DreamRelation_ Bridging Customization and Relation Generation/full.md @@ -0,0 +1,352 @@ +# DreamRelation: Bridging Customization and Relation Generation + +Qingyu Shi $^{1}$ Lu Qi $^{3*}$ Jianzong Wu $^{1}$ Jinbin Bai $^{5}$ Jingbo Wang $^{4}$ Yunhai Tong $^{1}$ Xiangtai Li $^{2,4*}$ $^{1}$ PKU NTU Insta360 Research Shanghai AI Laboratory NUS +Project page: https://shi-qingyu.github.io/DreamRelation.github.io + +![](images/3cc4df9f5289bb5b2e581c9f6838dcc383b548bd087fecfac7349bbcafc80669.jpg) +Figure 1. In our Relation-Aware Image Customization task, the generated images must accurately preserve the relationships between objects and maintain their identity. We highlight the limitations of previous approaches using three color codes: red indicates failure to capture relationships, blue marks missing objects, and orange represents object confusion. Each image is annotated to reflect its specific issue. Our results, highlighted by green boxes, demonstrate the advantages of our proposed method. + +# Abstract + +Customized image generation is essential for creating personalized content based on user prompts, allowing large-scale text-to-image diffusion models to more effectively meet individual needs. However, existing models often neglect the relationships between customized objects in generated images. In contrast, this work addresses this gap by focusing on relation-aware customized image generation, which seeks to preserve the identities from image prompts while maintaining the relationship specified in text prompts. Specifically, we introduce DreamRelation, a framework that disentangles identity and relation learning using a carefully curated dataset. Our training data consists of relation-specific images, independent object images containing identity information, and text prompts to guide relation genera + +tion. Then, we propose two key modules to tackle the two main challenges—generating accurate and natural relationships, especially when significant pose adjustments are required, and avoiding object confusion in cases of overlap. First, we introduce a keypoint matching loss that effectively guides the model in adjusting object poses closely tied to their relationships. Second, we incorporate local features of the image prompts to better distinguish between objects, preventing confusion in overlapping cases. Extensive results on our proposed benchmarks demonstrate the superiority of DreamRelation in generating precise relations while preserving object identities across a diverse set of objects and relationships. + +# 1. Introduction + +Driven by large-scale text-to-image diffusion models [26, 27, 30, 32], customized image generation has recently made significant strides [1, 14, 15, 31, 35, 36, 43]. This task focuses on generating images that preserve the identity of objects from user-provided inputs, enabling the creation of personalized and meaningful content. It has shown value in numerous applications, including personalized artwork, branding, virtual fashion try-ons, social media content creation, augmented reality experiences, and more. + +Despite the success of many methods for customizing single or multiple objects [8, 16, 18, 20, 23, 24, 35, 42], they often overlook the relationships between objects and the corresponding text prompts. For instance, when two user-provided objects are paired with a text prompt specifying a particular relationship, the generated output should not only preserve their identities but also accurately reflect the intended relationship, such as a 'hug'. This introduces new challenges in what we refer to as relation-aware customized image generation, which focuses on preserving multiple identities while adhering to relationship prompts. + +An intuitive solution to this issue is to adapt existing relation-aware generation methods [11, 21, 22] to tuning-based or training-based customized methods for customizing objects while maintaining the relation. However, both approaches face challenges. Tuning-based methods struggle to preserve multiple identities, as they invert objects into specific tokens. While training-based schemes often fail to balance image and text prompts, frequently overlooking key textual elements and hindering the generation of relationships between objects. As illustrated in Fig. 1, previous methods fail to capture the relationships described in the text prompt and lose identity preservation. + +We attribute this failure to two key factors: a lack of relevant data and an ineffective model design. Unlike data augmentation techniques such as flipping or rotation, commonly used to create paired training data in object customization methods [5, 40], our approach requires a triplet of images: two image prompts and one target image. The image prompts should contain similar objects but exhibit distinct poses compared to the target image. To collect these triplets, we propose a data engine to curate our finetuning set. We leverage an advanced text-to-image generation model [2] to generate triplets where the same object pair is shared across the images. Through text prompt guidance, the image prompt provides strong identity information. This enables the decoupling of the relationship in the target image, which enhances the relation learning. + +For the model design, we propose DreamRelation, which applies the Low-Rank Adaptation (LoRA) [9] strategy to the text cross-attention layers of existing diffusion models to process user-provided text prompts. In DreamRelation, two key modules are introduced during training to + +enhance the relation generation in customized generation. First, we introduce a keypoint matching loss (KML) as additional supervision to explicitly encourage the model to manipulate object poses, since relationships between objects are closely tied to their poses. Importantly, the KML operates on the latent representation rather than the original image space, aligning with the default diffusion loss. Second, since relation-aware customization requires local features from image prompts, such as the "hands" features for generating "shaking hands"—which are not captured by CLIP's coarse image-level features, we introduce dense features from CLIP. Through partitioning and pooling, we obtain local tokens that contain detail and local information. To further enhance the compatibility between dense features and image-level features, we employ a self-distillation method [39] to improve their alignment. + +To more comprehensively evaluate relation-aware customized image generation, we constructed our Relation-Bench based on three established benchmarks [15, 31]. Our methods demonstrate strong performance compared to existing approaches, achieving significant improvements in visual quality and quantitative results. + +The contributions of this work are: + +- We explore a novel task called relation-aware customized generation, which aims to preserve multiple identities from image prompts while adhering to the relationships specified in text prompts. This task can enhance various user-driven applications by enabling more control. +- We introduce a data engine that uses the recent text-to-image generation model to generate triplet images where the same object is present with distinct poses. This well-curated dataset enables the model to focus on relation learning during fine-tuning, leveraging the identity information provided by the image prompts. +- Our proposed DreamRelation has two key modules, including keypoint matching loss and local token injection, to enhance the relation generation in customized generation. The extensive experiments on three benchmarks demonstrate the effectiveness of our method. + +# 2. Related Work + +Diffusion-Based Text-to-Image Generation. Diffusion-based text-to-image models [30, 32] generate high-quality images based on user-provided text prompts. These works encode the text prompt through the text encoder [6, 28] and inject the text embedding into the U-Net's cross-attention layers. Some methods [26, 27, 29, 38] upscale diffusion models and incorporate additional conditions as priors to generate high-resolution images. Meanwhile, several methods [2, 17] introduce stronger text encoders or large language models (LLMs) to enhance the text comprehension capabilities of diffusion models. Recent works [4, 25] have replaced the U-Net denoiser with the Transformers [33]. + +However, these models primarily focus on text conditions and struggle to handle other forms of guidance, such as user-provided images. In contrast, our work facilitates customized generation while following relation conditions given by text inputs. + +Diffusion-Based Customized Generation. Customized image generation aims to produce new images based on user-provided objects. Tuning-based methods accomplish this by fine-tuning specific parameters of diffusion models to learn new objects. Some of these methods [7, 20, 34] employ text embeddings to represent the object identity. On the other hand, DreamBooth [31] fine-tunes the entire U-Net and introduces a prior preservation loss to mitigate language drift. Additionally, several approaches [8, 15, 18, 23] integrate text embeddings and U-Net parameters for image customization. These methods fine-tune both components during object learning, achieving impressive results. Considering the substantial cost of fine-tuning for commercial uses, training-based methods [5, 16, 24, 36, 43] have been proposed. This approach primarily utilizes an encoder to extract the object's identity and inject it into the U-Net. For instance, MS-Diffusion [35] integrates grounding tokens with a Resampler to maintain detailed identity and employs layout guidance to explicitly locate the objects. Despite their efficiency during inference, training-based methods often face challenges in balancing identity preservation and text control. A common issue is that the relationships described in text prompts are frequently overlooked. To address this limitation, we introduce DreamRelation, a framework designed to bridge the customization and relation generation. + +Relation-Aware Text-to-Image Generation. Inspired by tuning-based customization methods [7], some recent works [11, 22, 37] represent a "neglected word" by learnable parameters. These methods fine-tune part of the parameters on the content co-existing images. For instance, Reversion [11] fine-tunes the text embedding on a set of relation co-existing images and introduces a relation-steering contrastive loss. Reversion gains a better alignment between relational words and generated images by injecting new text embedding into prompts. While effective with prepositions and adjectives, Reversion struggles with relationships that involve significant overlaps, and cannot customize user-provided objects. Moreover, this method still struggles to generate vivid relationships while maintaining the fidelity of multiple customized objects, likely due to the inherent disregard for text embeddings in customization methods. Our work bridges the gap in the predicate relation-following capability of training-based customization methods. This enables efficient inference that naturally and accurately generates relationships in the text prompts. + +# 3. Bridging Customization and Relation Generation + +# 3.1. Problem definition + +Different from conventional customization tasks, we explore a new setting that focuses on image generation by both image prompts $c_{i} \in \mathcal{R}^{N \times 3 \times H \times W}$ and text prompts $c_{t}$ . We call this setting relation-aware customized image generation due to the requirements that the generated image $\hat{x}$ should strictly preserve each identity and keep the relationship among those identities provided by $c_{i}$ and $c_{t}$ . + +We define this task as: + +$$ +\hat {x} = \Phi_ {\theta} \left(c _ {i}, c _ {t}\right) \tag {1} +$$ + +where $\Phi_{\theta}$ is the network parameterized by $\theta$ . For brevity clarification, we set $N = 2$ due to the basic triplet element defined in the DreamRelation. + +# 3.2. Discussion and Motivation + +While previous works [8, 15, 43] have explored the customization of multiple objects under text control, they lack the ability to effectively manage the relationships between these objects. Specifically, the state-of-the-art multi-object customization model [35] utilizes bounding boxes to enhance multi-object generation. Given the image prompt $c_{i}$ and the text prompt $c_{t}$ , it leverages CLIP [28] and the Resampler [35] to extract text tokens $tok_{\text{text}}$ and image tokens $tok_{\text{image}}$ , respectively. These tokens are then injected into the U-Net $\epsilon_{\theta}$ through parallel cross-attention layers: + +$$ +h = \operatorname {S o f t m a x} \left(\frac {Q K _ {i}}{\sqrt {d}} + \mathcal {M}\right) V _ {i} + \operatorname {S o f t m a x} \left(\frac {Q K _ {t}}{\sqrt {d}}\right) V _ {t} \tag {2} +$$ + +where $Q$ represents the query, while $K_{i}, V_{i}$ and $K_{t}, V_{t}$ denote the keys and values obtained from $tok_{image}$ and $tok_{text}$ , respectively. The mask $\mathcal{M}$ , generated from the bounding boxes, assists the model in localizing objects during multi-object customization. However, it encounters difficulties when bounding boxes overlap, frequently resulting in object confusion, as shown in Fig. 1. Consequently, relation-aware generation in customized models remains a largely unexplored area of research. + +On the other hand, we revisit the relation-aware generation task. Existing relation inversion methods [11, 22] focus on inverting a relationship into a text embedding $R^{*}$ within pre-trained text-to-image diffusion models. Given a set of images $\{x_{k}\}_{k = 1}^{n}$ that share the same relationship, these methods use a denoising loss $\mathcal{L}_{RI}$ to fine-tune $R^{*}$ , ensuring alignment with the specific relationship: + +$$ +\mathcal {L} _ {R I} = \mathbb {E} _ {z _ {t}, t, \epsilon , c} [ \| \epsilon_ {t} - \boldsymbol {\epsilon} _ {\boldsymbol {\theta}} (z _ {t}, t, c _ {t}) \| _ {2} ^ {2} ] \tag {3} +$$ + +Directly applying relation-aware generation methods to customized models leads to suboptimal performance, as illustrated in Fig. 4. We attribute this to two main reasons. + +![](images/179ca59630421d63fad72096ed5ba6770a9ce18feecc3fef2f58a9ec048377c3.jpg) +Figure 2. The overview of DreamRelation. DreamRelation utilizes the off-the-shelf identity extractor to decouple the relation and identity information in relation-specific images. After getting the U-Net output $\hat{\epsilon}$ , we predict $\hat{z}_0$ and calculate the keypoint matching loss. The part in the dotted box is only for training. + +First, the balance between the control of $c_{i}$ and $c_{t}$ is not addressed in relation-aware generation tasks. During inference, the model tends to overlook the target relationship $R^{*}$ due to the influence of $c_{i}$ , resulting in incorrect relationship generation. Second, even if the shared relationship exists across $\{x_{k}\}_{k = 1}^{n}$ , the disentangled object identity information complicates the learning process. Fine-tuning solely on $\{x_{k}\}_{k = 1}^{n}$ makes it challenging for the model to effectively capture the relationship. As shown in the second row of Fig. 4, where ReVersion [11] struggles to generate the "hugging" relationship even in the absence of $c_{i}$ . In the following subsections, we will introduce our data collection method and DreamRelation to bridge customization and relation generation. + +# 3.3. Data Collection + +To address the balance between $c_{i}$ and $c_{t}$ while facilitating relation learning from $\{x_{k}\}_{k = 1}^{n}$ , we first design a data engine for collecting high-quality tuning data. The ideal tuning data should be in the form of $\mathcal{D} = (x_k, c_i, c_t)$ . The $c_{i}$ contain the object in $x_{k}$ while $c_{t}$ describes the relationship between objects in $x_{k}$ . The identity information in $c_{i}$ helps decouple the relationship within $x_{k}$ , enhancing relation learning while maintaining a balanced integration of $c_{i}$ and $c_{t}$ . Unfortunately, directly cropping the objects from $x_{k}$ to obtain $c_{i}$ results in a copy-and-paste effect [5], even when using data augmentation techniques like flipping and rotation. Therefore, we introduce our data engine below. + +Relation-aware data engine. Fig. 3 illustrates the difference of our data engine. Both of $x_{k}$ and $c_{i}$ are generated by the recent text-to-image generation model, rather than through cropping and augmentation. Specifically, inspired by the performance of DALL-E 3 and its multi-turn dialog capability. We use DALL-E 3 to generate $\mathcal{D} = (x,c_i,c_t)$ . Where $x_{k}$ and $c_{i}$ share the same object identity by leveraging the prompt "The photo of the same." We observe that + +![](images/7a43ac3f81a67b732255010dc504844454122f94c365dc3bbc25b93779444dde.jpg) +Figure 3. The rigid object cropping data engine, which results in minimal changes to object pose, the cropped image prompts also contain relation information, leading to a copy-and-paste effect, where the text prompt is neglected. Our relation-aware data engine, on the other hand, focuses on relation learning by decoupling identity information from images. + +prompting in this manner allows DALL-E 3 to remember and preserve identity in common categories such as "tiger", which is enough for relation learning. Next, we employ X-Pose [41], SAM [13], and LLaVA [19] to annotate $x$ , $c_i$ with keypoints, masks, and captions for samples. + +# 3.4. DreamRelation + +# 3.4.1. Relation Learning + +To balance $c_{i}$ and $c_{t}$ while enhancing relation learning, we decouple the identity information and relationships in $x_{k}$ . Specifically, we leverage an off-the-shelf ID extractor from the state-of-the-art customized model [35]. The identity in- + +
MethodsSingle-Object CusMulti-Object CusNon-ConfusingRelation-Aware
Textual Inversion [7]✓✗✗✗
DreamBooth [31]✓✗✗✗
Custom Diffusion [15]✓✓✗✗
Mix-of-Show [8]✓✓✗✗
CLIF [18]✓✓✓✗
MultiBooth [45]✓✓✗✗
BLIP-Diffusion [16]✓✗✗✗
ELITE [36]✓✗✗✗
AnyDoor [5]✓✗✗✗
SSR-Encoder [43]✓✓✗✗
λ-eclipse [24]✓✓✗✗
MS-Diffusion [35]✓✓✗✗
ReVersion [11]✗✗✓✓
ADI [10]✗✗✗✓
DreamRelation✓✓✓✓
+ +Table 1. Setting Comparison for Different Models. We include several representative methods here. Our proposed DreamRelation can meet the demand of the relation-aware customization task. + +![](images/91f5be357e43116896e2acb82eb1f233545ff1bd6c4dc219ab993a11cbc320d0.jpg) +Figure 4. Comparing our method with ReVersion across different base models, our approach demonstrates superior performance in the relation-aware generation task. + +formation in $c_{i}$ enables the model to focus on the relationship information in $x_{k}$ and $c_{t}$ during relation learning. To achieve this, we employ parallel cross-attention layers to process $c_{i}$ and $c_{t}$ , which can be formulated as: + +$$ +h = \gamma \cdot \operatorname {S o f t m a x} \left(\frac {Q K _ {i}}{\sqrt {d}} + \mathcal {M}\right) V _ {i} + \operatorname {S o f t m a x} \left(\frac {Q K _ {t}}{\sqrt {d}}\right) V _ {t} \tag {4} +$$ + +where $Q = W_{q}h$ , $K_{i} = W_{k_{i}}c_{i}$ , $V_{i} = W_{v_{i}}c_{i}$ , $K_{t} = W_{k_{t}}c_{t}$ , $V_{t} = W_{v_{t}}c_{t}$ , and $\gamma$ is a hyperparameter for scaling control from $c_{i}$ . For clarity, we omit the final linear layer $W_{out}$ . The relation information mainly depends on text prompts. Therefore, we inject LoRA layers in all of the $W_{q}$ , $W_{k_{t}}$ , $W_{v_{t}}$ , and $W_{out}$ to encourage the model to pay more attention to the relationship in $c_{t}$ . We freeze all other parameters during fine-tuning. As shown in Fig. 4, benefiting from the decoupled identity information in $c_{i}$ , our method effectively captures the relationship in $x_{k}$ and $c_{t}$ , accurately generating the corresponding images. Notably, after fine-tuning, our method can also be directly integrated into SDXL [27], effectively enabling the model to address the relation-aware generation task. The extensive results are presented in the Supplementary Material (SM). + +# 3.4.2. Keypoint Matching Loss + +Most relationships have specific requirements for the pose, such as "hugging" requiring the arms to cross over the object's body, and "riding a bicycle" requiring the feet of the object to be on the pedal plate. Therefore, it's important to manipulate the pose of objects in the right status accurately. Intuitively, we introduce explicit supervision in the diffusion latent space during the tuning process, named Keypoint Matching Loss (KML). + +Specifically, considering our task only involves common objects instead of humans, we employ X-Pose [41] as a keypoint detector because it can detect any keypoints in complex real-world scenarios. We detect 17 keypoints for each object in $x_{k}$ and $c_{i}$ , and denote the keypoint coordinates as $c_{kp}^{x_k}, c_{kp}^{c_i} \in \mathcal{R}^{17 \times 2}$ respectively. To encourage the model to generate accurate poses in $\mathcal{D}(\hat{z}_0)$ , where $\mathcal{D}$ is the VAE decoder. We use the U-Net's output $\hat{\epsilon}$ to predict $z_0$ during fine-tuning: + +$$ +\hat {z} _ {0} = \frac {z _ {t} - \sqrt {1 - \bar {\alpha} _ {t}} \hat {\epsilon}}{\sqrt {\bar {\alpha} _ {t}}} \tag {5} +$$ + +We use the VAE encoder $\mathcal{E}$ to obtain the latent representation of image prompts $\mathcal{E}(c_i)$ . Then we calculate the MSE loss on the corresponding keypoint locations between $\mathcal{E}(c_i)$ and $\hat{z}_0$ : + +$$ +\mathcal {L} _ {K M L} = \frac {1}{n _ {k p}} \mathbb {E} _ {\boldsymbol {z} _ {t}, c _ {i}} \left\| \mathcal {E} (c _ {i}) [ c _ {k p} ^ {c _ {i}} ] - \hat {z} _ {0} [ c _ {k p} ^ {x _ {k}} ] \right\| _ {2} ^ {2} \tag {6} +$$ + +$$ +\mathcal {L} = \mathcal {L} _ {\text {d e n o i s e}} + \lambda \cdot \mathcal {L} _ {K M L} \tag {7} +$$ + +where $n_{kp}$ denote the number of keypoint, $c_{kp}^{c_i}$ and $c_{kp}^{x_k}$ are keypoint's coordinates in $c_i$ and $x_k$ respectively. We scale the coordinates to fit the size of $\mathcal{E}(c_i)$ , $\hat{z}_0 \lambda$ controls for the relative weight of KML. The part inside the dotted box in Fig. 2 illustrates the model fine-tuning with the KML. We find KML is effective in encouraging the model to manipulate the pose for relation-aware generation. + +# 3.4.3. Local Token Injection + +When generating relationships, local and detailed features are essential—for instance, the feature of arms to accurately construct relationships like "carrying". However, the coarse image-level features extracted by the CLIP Image Encoder fail to provide such fine-grained details, leading to confusion between objects in Fig. 7. To address this limitation, we introduce dense features from the CLIP Image Encoder [44]. Specifically, we modify the last layer of the CLIP Image Encoder to obtain dense features: $h_{\mathrm{dense}} =$ ModifiedAttention( $h_{\mathrm{clip}}'$ ), where $h_{\mathrm{clip}}'$ represents the hidden states of the second-to-last layer. + +$$ +h _ {\mathrm {t m p}} = \operatorname {P r o j} _ {v} \left(\operatorname {n o r m} \left(h _ {\mathrm {c l i p}} ^ {\prime}\right)\right), \tag {8} +$$ + +$$ +h _ {\mathrm {t m p}} = h _ {\mathrm {c l i p}} ^ {\prime} + \operatorname {P r o j} _ {\text {o u t}} \left(h _ {\mathrm {t m p}}\right), \tag {9} +$$ + +$$ +h _ {\text {d e n s e}} = h _ {\text {t m p}} + \operatorname {F F N} \left(h _ {\text {t m p}}\right) \tag {10} +$$ + +![](images/809ee864e0f3795e33b31f9468f23149cc2ed865d2a79d8a8d671053d613027d.jpg) +Figure 5. Additional results demonstrate the effectiveness of relation-aware generation. DreamRelation adapts the same pair of objects to different relationships in a natural and accurate manner. + +To further enhance the compatibility between dense features and image-level features, we employ self-distillation on the CLIP Image Encoder [39] to derive our Local Image Encoder. Specifically, we align local regions of the dense features with their corresponding image-level features using cosine similarity. + +During inference, we extract local tokens $tok_{\mathrm{local}}$ from the dense feature $h_{\mathrm{dense}}$ through partitioning and pooling. These local tokens are then concatenated with the image-level tokens $tok_{\mathrm{image}}$ and passed to the ID extractor. The ID extractor, which is a transformer-based architecture as illustrated in Fig. 2 (b), can be written as: + +$$ +q = q + \text {A t t e n t i o n} (\text {c o n c a t} [ q, \text {t o k} _ {\text {i m a g e}}, \text {t o k} _ {\text {l o c a l}} ]) \tag {11} +$$ + +Regarding the compatibility between CLIP dense features and the original image-level features, experimental results demonstrate that our Local Image Encoder outperforms other dense representations [3]. For a detailed discussion, we refer readers to the Supplementary Material. Additionally, the injection of local tokens enhances identity preservation, as evidenced by improvements in evaluation metrics and visual results, shown in Fig. 7 and Tab. 3. + +# 4. Experiments + +Implementation Details. We generate a tuning set for relation learning with 4-6 samples per relationship. We incorporate LoRA layers into all text cross-attention layers of the U-Net. We set the LoRA rank to $r = 4$ , the parallel cross-attention scaling factor to $\gamma = 0.6$ , and the keypoint matching loss weight to $\lambda = 1e - 3$ . The model is finetuned for 500 steps, using 2 A100 GPUs, with a total batch size of 8, completing the process in 10 minutes. We use the Adam optimizer [12] with a learning rate of $1e - 4$ and no weight decay, resulting in a total of 3.1M trainable parameters. Note that our learnable parameters are only in the text cross-attention layers, making it compatible with + +any SDXL-based models. Considering the strong identity preservation capabilities of MS-Diffusion, we implement our DreamRelation on MS-Diffusion. During fine-tuning and inference, we concatenate local tokens with image tokens. We put more self-distillation details of the Local Image Encoder in the Supplementary Material. + +Evaluation. To evaluate relation-aware customized image generation, we propose RelationBench, consisting of 44 objects from DreamBench [31] and CustomConcept101 [15], along with 25 relationships. The object categories include pets, plushies, toys, people, and cartoons. Using GPT-4, we generate 100 cases for single-object and multi-object evaluations. Following previous works [31, 35], we evaluate our method on three metrics: (1) Identity Preservation, which assesses the alignment between the generated images and the image prompts, using the CLIP image score and DINO score that calculate the cosine similarity between the class embeddings of images, referred to as CLIP-I and DINO, respectively; (2) Text Alignment, which evaluates how well the generated images align with the text prompts, using the CLIP image-text score, denoted as CLIP-T; and (3) Relation Alignment. We observed that nouns in the prompts can inflate CLIP-T scores, making them less accurate for evaluating relation generation. To address this, we extract the relation word from the text prompt using spaCy and calculate the CLIP image-text score, denoted as CLIP-R. + +# 4.1. Main Results + +We evaluate our method through both quantitative and qualitative results in single- and multi-object cases. In single-object cases, we primarily focus on the alignment of the object's pose with the relationship in the text prompt and the model's ability to preserve identity. In multi-object cases, in addition to relation generation and identity preservation, we also need to ensure no confusion between the multiple objects. + +Quantitative results. We compare our method with + +![](images/379cf19086e316ab79992f9b391916aedabc0f7ca89b83b404a29b1f4b25a375.jpg) +Objects +A cat is kissing with a dog. +A dog is shaking hands with a plushie bear. +A dog is fencing with a plushie panda. + +![](images/e102369285e21b5cc8ae72b36faee5b2f7b6ac7514d0d985b4b00b054557487d.jpg) +Prompts +Ours + +![](images/e10a75568002dc11826e587a9e8eac413e349b8c993468977332e3f7949ecef8.jpg) +ReVersion+MS + +![](images/8d19a2a803bcd79ee9e96e9cb83dae13b0b3883e0cf39d8b9e7d01008fc1be9a.jpg) +MS-Diffusion + +![](images/b2589730f24b1c2f3f12031d4b1773e0ab117f2a7f7b744cc00913ab8f3ae562.jpg) +SSR-Encoder + +![](images/8bf4caad1cc5881f70b03d4d991c0d7957631c5a62c953cf9b34fa1c2e3dc06d.jpg) +$\lambda$ -eclipse + +![](images/988c9e1efedd07f5f5656173d3d0ebd0e3fb88b7835f4af9f27664ad6a1e4ff5.jpg) + +![](images/ede574f721266b94a4be719cd335833cd807c661ecf5343115e7796ac2214c53.jpg) +Custom Diffusion DreamBooth (SDXL) + +![](images/1d6107a00e1ba9db511ff8589de8ebd015dc5b79f7e44311a182427c896f2645.jpg) + +![](images/95eec3e4856dc4ea959accb9c0817c96907d14e9dd7b83c270aec1448d66b0f1.jpg) + +![](images/d2a93e7beb1b95cff6a234f46d3fc23c03ce3861b22648b1aa57bfa2a26af9b3.jpg) +Figure 6. We compare multi-object performance with training-based and tuning-based methods. Additionally, we integrate ReVersion [11] into MS-Diffusion, where the learnable text token is injected into the text prompt during MS-Diffusion's inference, resulting in ReVersion+MS. Our method shows a clear advantage in relation-aware generation and avoiding object confusion in overlapping scenarios. + +![](images/9e418c165bf878515dc4e89a4e5c8c317ba989b5ffebd56d776fa23f1ad0a435.jpg) + +![](images/82bc6c3f3d392bd90ffb7eaa46d951666f771eb021df5195fc4f4f89b0aa459f.jpg) + +
MethodSingle-objectMulti-object
CLIP-TCLIP-RCLIP-IDINOCLIP-TCLIP-RCLIP-IDINO
DreamBooth (SDXL) [31]27.818.274.262.624.316.267.857.2
Custom Diffusion [15]26.515.273.258.820.115.464.755.3
Cones-V2 [20]24.413.572.157.221.315.264.354.2
ELITE [36]25.714.975.461.5————
AnyDoor [5]24.514.777.462.221.614.969.759.8
BLIP-Diffusion [16]26.215.777.457.7————
SSR-Encoder [43]25.515.980.459.424.214.672.156.2
MS-Diffusion [35]26.518.878.764.526.918.973.858.8
Reversion+MS [11]27.819.377.863.127.219.274.259.7
Ours30.621.477.963.428.920.475.462.1
+ +Table 2. Quantitative comparison on RelationBench, Bold and underline represent the highest and second-highest metrics. + +
MethodMulti-object
CLIP-TCLIP-RCLIP-IDINO
w/o Relation-aware Data27.319.475.359.8
w/o Local Token Injection28.519.575.159.9
w/o Keypoint Matching Loss27.419.275.261.2
Full Model28.920.475.462.1
+ +Table 3. Ablation study on our proposed components + +baseline models across three benchmarks: Relation-Bench, DreamBench, and multi-object cases in Custom-Concept101. As shown in Tab. 2, our method demonstrates a clear advantage on RelationBench for single-object cases, particularly in the CLIP-T and CLIP-R metrics. However, its performance is slightly lower in the CLIP-I metric, likely due to significant pose variations in the objects, which may decrease the CLIP-I score. For multi-object generation, our method outperforms other approaches across all evaluation metrics, which we attribute to its ability to effectively avoid object confusion and significantly enhance CLIP-I performance. Furthermore, as shown in Tab. 4, we evaluate + +single-object generation performance on DreamBench. Our method outperforms others in the CLIP-T metric while delivering competitive results in CLIP-I and DINO. We conduct experiments on multi-object cases from CustomConcept101, which we denote as M-CustomConcept101. As shown in Fig. 5, our model achieves the highest performance in the DINO score and ranks second in both the CLIP-T and CLIP-I metrics. + +Qualitative Comparison. We conduct qualitative comparison experiments on RelationBench to evaluate relation-aware customized image generation. As shown in Fig. 6, our method outperforms others in generating multi-object relationships, effectively avoiding object confusion while accurately capturing the intended relationships. Additional results can be found in the SM. + +# 4.2. Ablation study and Analysis + +Relation-Aware Data Engine. Instead of fine-tuning on our curated $(x_{k},c_{i},c_{t})$ pairs. We train a text embedding $R^{*}$ + +![](images/c32e4c51e7a31b10383fb3a9b685ecc9b41537c37ccecf9bf7fb6145c57a1a43.jpg) +a) objects + +![](images/0cc68667bf281a2204618555ea89da1f93e41bba7b0e13375c76845b552ad4cc.jpg) +wKML +A cat is shaking hands with a dog. + +![](images/2cc0ebf8b7f7e3aea1fbb1cf771a434fd3853238e211b5d69d00071345cc47d3.jpg) +w/o KML + +![](images/8436a73e636a2148e8b7a4b22337942cefc829a9361b77e0a299c9032012c955.jpg) +b) objects + +![](images/585e297493747fd77bf7493be8bdea864588b9d43a1d8f9f364a84618a821200.jpg) +w local tokens +A dog is carrying a sloth. + +![](images/48d030056b3ff97702524a71b39ca2139dcc3c05dac93881cabd07bad6aaff44.jpg) +w/o local tokens + +![](images/41e954c311fecc6c76611c1ed057e5c60c72fa2762f1539244342bd52da504b9.jpg) +Figure 7. Ablation study of keypoint matching loss (KML) and local tokens: The images are generated using the same random seed. We use green and red boxes to highlight the main differences. + +![](images/bbab11eb667e11037d0aaa1da0d501ce47968c0ae8263f930f8f5c0d24132ae1.jpg) +A dog, a cat, a plushie bear hug with each other. +Figure 8. DreamRelation for three objects. Compared to the baseline model, DreamRelation shows a clear advantage. + +![](images/f413434c347e23d2334146db8ee38434698549a26d84273dae987ff814449779.jpg) + +![](images/bdeac0a670eabb1140cb1c1fb6c118c3778e4b7f077aba905985c69174d782f6.jpg) +A dog, a cat, and a plushie bear are fighting with one another. + +![](images/c6946fccc9d404af3675a38a9f8de4efe96dce4e18597be7b39406edfcc50e73.jpg) + +![](images/0a85a82b8ac6fbae2b0a50ec754485584c551cd7d3372904df6bc833592c4e57.jpg) + +using a set of relation-specific data $(x_{k},c_{i}^{\prime},c_{t})$ , where $c_{i}^{\prime}$ is cropped from $x_{k}$ . The well-trained text embedding $R^{*}$ is injected into the text embedding during inference. The quantitative results in Tab. 3 show that our relation-aware data engine outperforms in both CLIP-T and CLIP-R scores. + +Keypoint Matching Loss (KML). We omit KML during fine-tuning. As shown in Fig. 7 a), including the KML leads to more accurate and natural relation-aware customized images using the same random seed. Quantitative results, presented in Tab. 3, show a decline across all evaluation metrics when the KML is removed. Additionally, we experimented with different values for $\lambda$ and found that the best performance occurs at $\lambda = 1e - 3$ . + +Local Token Injection. For local token injection, the visual comparison in Fig. 7 b) shows severe object confusion when local tokens are omitted. The quantitative results in Tab. 3 demonstrate that local tokens improve relation-aware customization. We conducted an ablation study on the number of local tokens, as presented in Tab. 6, and determined the optimal number to be 16. Additional ablation studies are shown in the SM. Combining these methods ensures that DreamRelation achieves high-quality relation-aware customized generation. + +
MethodSingle-object
CLIP-TCLIP-IDINO
DreamBooth (SDXL) [31]31.281.569.2
Custom Diffusion [15]28.477.266.8
Cones-V2 [20]31.076.567.2
ELITE [36]29.877.462.5
AnyDoor [5]25.582.167.8
SSR-Encoder [43]30.882.161.2
MS-Diffusion [35]31.579.368.2
Ours31.681.768.4
+ +Table 4. Evaluation Results on DreamBench. + +
MethodMulti-object
CLIP-TCLIP-IDINO
DreamBooth (SDXL) [31]29.567.449.2
Custom Diffusion [15]28.166.248.8
Cones-V2 [20]29.266.547.2
AnyDoor [5]20.272.151.2
SSR-Encoder [43]30.671.152.2
λ-eclipse [24]29.268.248.2
MS-Diffusion [35]28.070.251.2
Ours29.771.452.3
+ +Table 5. Evaluation Results on M-CustomConcept101. + +
Num Local TokensMulti-object
CLIP-TCLIP-RCLIP-IDINO
N=2×228.319.475.360.1
N=4×428.920.475.462.1
N=8×827.918.174.159.4
+ +Table 6. Ablation study on the number of local tokens. + +
LambdaMulti-object
CLIP-TCLIP-RCLIP-IDINO
λ=1e-227.519.372.659.6
λ=1e-328.920.475.462.1
λ=1e-426.818.173.960.4
+ +Table 7. Ablation study on $\lambda$ ,controlling relative weight of KML. + +# 5. Conclusion + +In this work, we introduce a new challenging task: relation-aware customized image generation. This task aims to generate objects that adhere to the relationship specified in the text prompt and preserve the identities of the user-provided images. To support this, we propose a data engine for generating high-quality fine-tuning data. Our method, DreamRelation, incorporates Keypoint Matching Loss and Local Token Injection, effectively capturing relation information and generating natural relations between customized objects. We also present a fair comparison with other methods on a new benchmark, RelationBench. Extensive experiments demonstrate the effectiveness of our method, highlighting its potential for applications in interactive scenario generation, relation detection, and more. Our research can inspire the direction of relation-aware generation in the community. + +Acknowledgement. This work is supported by the National Key Research and Development Program of China (No. 2023YFC3807600). + +# References + +[1] Jinbin Bai, Tian Ye, Wei Chow, Enxin Song, Qing-Guo Chen, Xiangtai Li, Zhen Dong, Lei Zhu, and Shuicheng Yan. Meissonic: Revitalizing masked generative transformers for efficient high-resolution text-to-image synthesis. arXiv preprint arXiv:2410.08261, 2024. 2 +[2] James Betker, Gabriel Goh, Li Jing, Tim Brooks, Jianfeng Wang, Linjie Li, Long Ouyang, Juntang Zhuang, Joyce Lee, Yufei Guo, et al. Improving image generation with better captions. Computer Science., 2:3, 2023. 2 +[3] Mathilde Caron, Hugo Touvron, Ishan Misra, Herve Jégou, Julien Mairal, Piotr Bojanowski, and Armand Joulin. Emerging properties in self-supervised vision transformers. In Int. Conf. Comput. Vis., 2021. 6 +[4] Junsong Chen, Jincheng Yu, Chongjian Ge, Lewei Yao, Enze Xie, Yue Wu, Zhongdao Wang, James Kwok, Ping Luo, Huchuan Lu, et al. Pixart-α: Fast training of diffusion transformer for photorealistic text-to-image synthesis. In ICLR, 2024. 2 +[5] Xi Chen, Lianghua Huang, Yu Liu, Yujun Shen, Deli Zhao, and Hengshuang Zhao. Anydoor: Zero-shot object-level image customization. In CVPR, 2024. 2, 3, 4, 5, 7, 8 +[6] Hyung Won Chung, Le Hou, Shayne Longpre, Barret Zoph, Yi Tay, William Fedus, Yunxuan Li, Xuezhi Wang, Mostafa Dehghani, Siddhartha Brahma, et al. Scaling instruction-finetuned language models. arXiv preprint arXiv:2210.11416, 2022. 2 +[7] Rinon Gal, Yuval Alaluf, Yuval Atzmon, Or Patashnik, Amit Haim Bermano, Gal Chechik, and Daniel Cohen-Or. An image is worth one word: Personalizing text-to-image generation using textual inversion. In ICLR, 2023. 3, 5 +[8] Yuchao Gu, Xintao Wang, Jay Zhangjie Wu, Yujun Shi, Yunpeng Chen, Zihan Fan, Wuyou Xiao, Rui Zhao, Shuning Chang, Weijia Wu, et al. Mix-of-show: Decentralized low-rank adaptation for multi-concept customization of diffusion models. NeurIPS, 2024. 2, 3, 5 +[9] Edward J Hu, Yelong Shen, Phillip Wallis abd Zeyuan Allen-Zhu, Yuanzhi Li, Shean Wang, Lu Wang, and Weizhu Chen. Lora: Low-rank adaptation of large language models. In ICLR, 2022. 2 +[10] Siteng Huang, Biao Gong, Yutong Feng, Xi Chen, Yuqian Fu, Yu Liu, and Donglin Wang. Learning disentangled identifiers for action-customized text-to-image generation. In CVPR, 2024. 5 +[11] Ziqi Huang, Tianxing Wu, Yuming Jiang, Kelvin CK Chan, and Ziwei Liu. Reversion: Diffusion-based relation inversion from images. arXiv preprint arXiv:2303.13495, 2023. 2, 3, 4, 5, 7 +[12] Diederik P Kingma and Jimmy Ba. Adam: A method for stochastic optimization. arXiv preprint arXiv:1412.6980, 2014. 6 +[13] Alexander Kirillov, Eric Mintun, Nikhila Ravi, Hanzi Mao, Chloe Rolland, Laura Gustafson, Tete Xiao, Spencer Whitehead, Alexander C. Berg, Wan-Yen Lo, Piotr Dollar, and Ross B. Girshick. Segment anything. In ICCV, 2023. 4 +[14] Lingjie Kong, Kai Wu, Xiaobin Hu, Wenhui Han, Jinlong Peng, Chengming Xu, Donghao Luo, Jiangning Zhang, + +Chengjie Wang, and Yanwei Fu. Anymaker: Zero-shot general object customization via decoupled dual-level id injection. arXiv preprint arXiv:2406.11643, 2024. 2 +[15] Nupur Kumari, Bingliang Zhang, Richard Zhang, Eli Shechtman, and Jun-Yan Zhu. Multi-concept customization of text-to-image diffusion. In CVPR, 2023. 2, 3, 5, 6, 7, 8 +[16] Dongxu Li, Junnan Li, and Steven C. H. Hoi. Blip-diffusion: Pre-trained subject representation for controllable text-to-image generation and editing. In NeurIPS, 2023. 2, 3, 5, 7 +[17] Long Lian, Boyi Li, Adam Yala, and Trevor Darrell. Llm-grounded diffusion: Enhancing prompt understanding of text-to-image diffusion models with large language models. arXiv preprint arXiv:2305.13655, 2023. 2 +[18] Wang Lin, Jingyuan Chen, Jiaxin Shi, Yichen Zhu, Chen Liang, Junzhong Miao, Tao Jin, Zhou Zhao, Fei Wu, Shuicheng Yan, et al. Non-confusing generation of customized concepts in diffusion models. ICML, 2024. 2, 3, 5 +[19] Haotian Liu, Chunyuan Li, Qingyang Wu, and Yong Jae Lee. Visual instruction tuning. NeurIPS, 2024. 4 +[20] Zhiheng Liu, Yifei Zhang, Yujun Shen, Kecheng Zheng, Kai Zhu, Ruili Feng, Yu Liu, Deli Zhao, Jingren Zhou, and Yang Cao. Cones 2: Customizable image synthesis with multiple subjects. In NeurIPS, 2023. 2, 3, 7, 8 +[21] Joanna Materzynska, Josef Sivic, Eli Shechtman, Antonio Torralba, Richard Zhang, and Bryan Russell. Customizing motion in text-to-video diffusion models. arXiv preprint arXiv:2312.04966, 2023. 2 +[22] mengmeng Ge, Xu Jia, Takashi Isobe, Xiaomin Li, Qinghe Wang, Jing Mu, Dong Zhou, liwang Amd, Huchuan Lu, Lu Tian, Ashish Sirasao, and Emad Barsoum. Customizing text-to-image generation with inverted interaction. In ACM MM, 2024. 2, 3 +[23] Lianyu Pang, Jian Yin, Baoquan Zhao, Feize Wu, Fu Lee Wang, Qing Li, and Xudong Mao. Attndreambooth: Towards text-aligned personalized text-to-image generation. arXiv preprint arXiv:2406.05000, 2024. 2, 3 +[24] Maitreya Patel, Sangmin Jung, Chitta Baral, and Yezhou Yang. λ-eclipse: Multi-concept personalized text-to-image diffusion models by leveraging clip latent space. arXiv preprint arXiv:2402.05195, 2024. 2, 3, 5, 8 +[25] William Peebles and Saining Xie. Scalable diffusion models with transformers. In ICCV, 2023. 2 +[26] Pablo Pernias, Dominic Rampas, Mats Leon Richter, Christopher Pal, and Marc Aubreville. Würstchen: An efficient architecture for large-scale text-to-image diffusion models. In ICLR, 2023. 2 +[27] Dustin Podell, Zion English, Kyle Lacey, Andreas Blattmann, Tim Dockhorn, Jonas Müller, Joe Penna, and Robin Rombach. Sdxl: Improving latent diffusion models for high-resolution image synthesis. arXiv preprint arXiv:2307.01952, 2023. 2, 5 +[28] Alec Radford, Jong Wook Kim, Chris Hallacy, Aditya Ramesh, Gabriel Goh, Sandhini Agarwal, Girish Sastry, Amanda Askell, Pamela Mishkin, Jack Clark, Gretchen Krueger, and Ilya Sutskever. Learning transferable visual + +models from natural language supervision. In ICML, 2021. 2,3 +[29] Jingjing Ren, Wenbo Li, Haoyu Chen, Renjing Pei, Bin Shao, Yong Guo, Long Peng, Fenglong Song, and Lei Zhu. Ultrapixel: Advancing ultra-high-resolution image synthesis to new peaks. arXiv preprint arXiv:2407.02158, 2024. 2 +[30] Robin Rombach, Andreas Blattmann, Dominik Lorenz, Patrick Esser, and Björn Ommer. High-resolution image synthesis with latent diffusion models. In CVPR, 2022. 2 +[31] Nataniel Ruiz, Yuanzhen Li, Varun Jampani, Yael Pritch, Michael Rubinstein, and Kfir Aberman. Dreambooth: Fine tuning text-to-image diffusion models for subject-driven generation. In CVPR, 2023. 2, 3, 5, 6, 7, 8 +[32] Chitwan Sahara, William Chan, Saurabh Saxena, Lala Li, Jay Whang, Emily L. Denton, Seyed Kamyar Seyed Ghasemipour, Raphael Gontijo Lopes, Burcu Karagol Ayan, Tim Salimans, Jonathan Ho, David J. Fleet, and Mohammad Norouzi. Photorealistic text-to-image diffusion models with deep language understanding. In NeurIPS, 2022. 2 +[33] Ashish Vaswani, Noam Shazeer, Niki Parmar, Jakob Uszkoreit, Llion Jones, Aidan N Gomez, Lukasz Kaiser, and Illia Polosukhin. Attention is all you need. In NeurIPS, 2017. 2 +[34] Andrey Voynov, Qinghao Chu, Daniel Cohen-Or, and Kfir Aberman. p+: Extended textual conditioning in text-to-image generation. arXiv preprint arXiv:2303.09522, 2023. 3 +[35] X Wang, Siming Fu, Qihan Huang, Wanggui He, and Hao Jiang. Ms-diffusion: Multi-subject zero-shot image personalization with layout guidance. arXiv preprint arXiv:2406.07209, 2024. 2, 3, 4, 5, 6, 7, 8 +[36] Yuxiang Wei, Yabo Zhang, Zhilong Ji, Jinfeng Bai, Lei Zhang, and Wangmeng Zuo. ELITE: encoding visual concepts into textual embeddings for customized text-to-image generation. In ICCV, 2023. 2, 3, 5, 7, 8 +[37] Yujie Wei, Shiwei Zhang, Zhiwu Qing, Hangjie Yuan, Zhiheng Liu, Yu Liu, Yingya Zhang, Jingren Zhou, and Hongming Shan. Dreamvideo: Composing your dream videos with customized subject and motion. In CVPR, 2024. 3 +[38] Jianzong Wu, Xiangtai Li, Yanhong Zeng, Jiangning Zhang, Qianyu Zhou, Yining Li, Yunhai Tong, and Kai Chen. Motionbooth: Motion-aware customized text-to-video generation. arXiv preprint arXiv:2406.17758, 2024. 2 +[39] Size Wu, Wenwei Zhang, Lumin Xu, Sheng Jin, Xiangtai Li, Wentao Liu, and Chen Change Loy. Clipself: Vision transformer distills itself for open-vocabulary dense prediction. ICLR, 2024. 2, 6 +[40] Binxin Yang, Shuyang Gu, Bo Zhang, Ting Zhang, Xuejin Chen, Xiaoyan Sun, Dong Chen, and Fang Wen. Paint by example: Exemplar-based image editing with diffusion models. In CVPR, 2023. 2 +[41] Jie Yang, Ailing Zeng, Ruimao Zhang, and Lei Zhang. Unipose: Detecting any keypoints. ECCV, 2024. 4, 5 +[42] Hu Ye, Jun Zhang, Sibo Liu, Xiao Han, and Wei Yang. Ip-adapter: Text compatible image prompt adapter for text-to-image diffusion models. arXiv preprint arxiv:2308.06721, 2023. 2 + +[43] Yuxuan Zhang, Yiren Song, Jiaming Liu, Rui Wang, Jinpeng Yu, Hao Tang, Huaxia Li, Xu Tang, Yao Hu, Han Pan, et al. Ssr-encoder: Encoding selective subject representation for subject-driven generation. In CVPR, 2024. 2, 3, 5, 7, 8 +[44] Chong Zhou, Chen Change Loy, and Bo Dai. Extract free dense labels from clip. In ECCV, 2022. 5 +[45] Chenyang Zhu, Kai Li, Yue Ma, Chunming He, and Li Xiu. Multibooth: Towards generating all your concepts in an image from text. arXiv preprint arXiv:2404.14239, 2024. 5 \ No newline at end of file diff --git a/2025/DreamRelation_ Bridging Customization and Relation Generation/images.zip b/2025/DreamRelation_ Bridging Customization and Relation Generation/images.zip new file mode 100644 index 0000000000000000000000000000000000000000..2309c7128990c6f54f87be494e6eb44b075b36b5 --- /dev/null +++ b/2025/DreamRelation_ Bridging Customization and Relation Generation/images.zip @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:265bb7f8aeaf979f01c04a7e921b65e2a7ef5716b3e274fe575d34af217d770f +size 940344 diff --git a/2025/DreamRelation_ Bridging Customization and Relation Generation/layout.json b/2025/DreamRelation_ Bridging Customization and Relation Generation/layout.json new file mode 100644 index 0000000000000000000000000000000000000000..7c04c80cefebfaa59172aaa6367bd88aa12b057e --- /dev/null +++ b/2025/DreamRelation_ Bridging Customization and Relation Generation/layout.json @@ -0,0 +1,10183 @@ +{ + "pdf_info": [ + { + "para_blocks": [ + { + "bbox": [ + 101, + 103, + 509, + 121 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 101, + 103, + 509, + 121 + ], + "spans": [ + { + "bbox": [ + 101, + 103, + 509, + 121 + ], + "type": "text", + "content": "DreamRelation: Bridging Customization and Relation Generation" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 56, + 141, + 568, + 185 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 141, + 568, + 185 + ], + "spans": [ + { + "bbox": [ + 56, + 141, + 568, + 185 + ], + "type": "text", + "content": "Qingyu Shi" + }, + { + "bbox": [ + 56, + 141, + 568, + 185 + ], + "type": "inline_equation", + "content": "^{1}" + }, + { + "bbox": [ + 56, + 141, + 568, + 185 + ], + "type": "text", + "content": " Lu Qi" + }, + { + "bbox": [ + 56, + 141, + 568, + 185 + ], + "type": "inline_equation", + "content": "^{3*}" + }, + { + "bbox": [ + 56, + 141, + 568, + 185 + ], + "type": "text", + "content": " Jianzong Wu" + }, + { + "bbox": [ + 56, + 141, + 568, + 185 + ], + "type": "inline_equation", + "content": "^{1}" + }, + { + "bbox": [ + 56, + 141, + 568, + 185 + ], + "type": "text", + "content": " Jinbin Bai" + }, + { + "bbox": [ + 56, + 141, + 568, + 185 + ], + "type": "inline_equation", + "content": "^{5}" + }, + { + "bbox": [ + 56, + 141, + 568, + 185 + ], + "type": "text", + "content": " Jingbo Wang" + }, + { + "bbox": [ + 56, + 141, + 568, + 185 + ], + "type": "inline_equation", + "content": "^{4}" + }, + { + "bbox": [ + 56, + 141, + 568, + 185 + ], + "type": "text", + "content": " Yunhai Tong" + }, + { + "bbox": [ + 56, + 141, + 568, + 185 + ], + "type": "inline_equation", + "content": "^{1}" + }, + { + "bbox": [ + 56, + 141, + 568, + 185 + ], + "type": "text", + "content": " Xiangtai Li" + }, + { + "bbox": [ + 56, + 141, + 568, + 185 + ], + "type": "inline_equation", + "content": "^{2,4*}" + }, + { + "bbox": [ + 56, + 141, + 568, + 185 + ], + "type": "inline_equation", + "content": "^{1}" + }, + { + "bbox": [ + 56, + 141, + 568, + 185 + ], + "type": "text", + "content": " PKU NTU Insta360 Research Shanghai AI Laboratory NUS \nProject page: https://shi-qingyu.github.io/DreamRelation.github.io" + } + ] + } + ], + "index": 3 + }, + { + "type": "image", + "bbox": [ + 61, + 190, + 547, + 439 + ], + "blocks": [ + { + "bbox": [ + 61, + 190, + 547, + 439 + ], + "lines": [ + { + "bbox": [ + 61, + 190, + 547, + 439 + ], + "spans": [ + { + "bbox": [ + 61, + 190, + 547, + 439 + ], + "type": "image", + "image_path": "3cc4df9f5289bb5b2e581c9f6838dcc383b548bd087fecfac7349bbcafc80669.jpg" + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 54, + 448, + 555, + 494 + ], + "lines": [ + { + "bbox": [ + 54, + 448, + 555, + 494 + ], + "spans": [ + { + "bbox": [ + 54, + 448, + 555, + 494 + ], + "type": "text", + "content": "Figure 1. In our Relation-Aware Image Customization task, the generated images must accurately preserve the relationships between objects and maintain their identity. We highlight the limitations of previous approaches using three color codes: red indicates failure to capture relationships, blue marks missing objects, and orange represents object confusion. Each image is annotated to reflect its specific issue. Our results, highlighted by green boxes, demonstrate the advantages of our proposed method." + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "image_caption" + } + ], + "index": 4 + }, + { + "bbox": [ + 151, + 512, + 200, + 525 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 151, + 512, + 200, + 525 + ], + "spans": [ + { + "bbox": [ + 151, + 512, + 200, + 525 + ], + "type": "text", + "content": "Abstract" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 54, + 546, + 297, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 54, + 546, + 297, + 714 + ], + "spans": [ + { + "bbox": [ + 54, + 546, + 297, + 714 + ], + "type": "text", + "content": "Customized image generation is essential for creating personalized content based on user prompts, allowing large-scale text-to-image diffusion models to more effectively meet individual needs. However, existing models often neglect the relationships between customized objects in generated images. In contrast, this work addresses this gap by focusing on relation-aware customized image generation, which seeks to preserve the identities from image prompts while maintaining the relationship specified in text prompts. Specifically, we introduce DreamRelation, a framework that disentangles identity and relation learning using a carefully curated dataset. Our training data consists of relation-specific images, independent object images containing identity information, and text prompts to guide relation genera" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 313, + 514, + 556, + 670 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 514, + 556, + 670 + ], + "spans": [ + { + "bbox": [ + 313, + 514, + 556, + 670 + ], + "type": "text", + "content": "tion. Then, we propose two key modules to tackle the two main challenges—generating accurate and natural relationships, especially when significant pose adjustments are required, and avoiding object confusion in cases of overlap. First, we introduce a keypoint matching loss that effectively guides the model in adjusting object poses closely tied to their relationships. Second, we incorporate local features of the image prompts to better distinguish between objects, preventing confusion in overlapping cases. Extensive results on our proposed benchmarks demonstrate the superiority of DreamRelation in generating precise relations while preserving object identities across a diverse set of objects and relationships." + } + ] + } + ], + "index": 8 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "spans": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "type": "text", + "content": "CVF" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 145, + 0, + 495, + 37 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 145, + 0, + 495, + 37 + ], + "spans": [ + { + "bbox": [ + 145, + 0, + 495, + 37 + ], + "type": "text", + "content": "This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 325, + 702, + 477, + 714 + ], + "type": "page_footnote", + "angle": 0, + "lines": [ + { + "bbox": [ + 325, + 702, + 477, + 714 + ], + "spans": [ + { + "bbox": [ + 325, + 702, + 477, + 714 + ], + "type": "inline_equation", + "content": "^{1}" + }, + { + "bbox": [ + 325, + 702, + 477, + 714 + ], + "type": "text", + "content": "Corresponding authors: Lu Qi, Xiangtai Li." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "text", + "content": "15723" + } + ] + } + ], + "index": 10 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 0 + }, + { + "para_blocks": [ + { + "bbox": [ + 56, + 71, + 136, + 84 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 71, + 136, + 84 + ], + "spans": [ + { + "bbox": [ + 56, + 71, + 136, + 84 + ], + "type": "text", + "content": "1. Introduction" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 55, + 91, + 294, + 199 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 91, + 294, + 199 + ], + "spans": [ + { + "bbox": [ + 55, + 91, + 294, + 199 + ], + "type": "text", + "content": "Driven by large-scale text-to-image diffusion models [26, 27, 30, 32], customized image generation has recently made significant strides [1, 14, 15, 31, 35, 36, 43]. This task focuses on generating images that preserve the identity of objects from user-provided inputs, enabling the creation of personalized and meaningful content. It has shown value in numerous applications, including personalized artwork, branding, virtual fashion try-ons, social media content creation, augmented reality experiences, and more." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 55, + 199, + 295, + 331 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 199, + 295, + 331 + ], + "spans": [ + { + "bbox": [ + 55, + 199, + 295, + 331 + ], + "type": "text", + "content": "Despite the success of many methods for customizing single or multiple objects [8, 16, 18, 20, 23, 24, 35, 42], they often overlook the relationships between objects and the corresponding text prompts. For instance, when two user-provided objects are paired with a text prompt specifying a particular relationship, the generated output should not only preserve their identities but also accurately reflect the intended relationship, such as a 'hug'. This introduces new challenges in what we refer to as relation-aware customized image generation, which focuses on preserving multiple identities while adhering to relationship prompts." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 55, + 331, + 295, + 474 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 331, + 295, + 474 + ], + "spans": [ + { + "bbox": [ + 55, + 331, + 295, + 474 + ], + "type": "text", + "content": "An intuitive solution to this issue is to adapt existing relation-aware generation methods [11, 21, 22] to tuning-based or training-based customized methods for customizing objects while maintaining the relation. However, both approaches face challenges. Tuning-based methods struggle to preserve multiple identities, as they invert objects into specific tokens. While training-based schemes often fail to balance image and text prompts, frequently overlooking key textual elements and hindering the generation of relationships between objects. As illustrated in Fig. 1, previous methods fail to capture the relationships described in the text prompt and lose identity preservation." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 55, + 474, + 295, + 653 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 474, + 295, + 653 + ], + "spans": [ + { + "bbox": [ + 55, + 474, + 295, + 653 + ], + "type": "text", + "content": "We attribute this failure to two key factors: a lack of relevant data and an ineffective model design. Unlike data augmentation techniques such as flipping or rotation, commonly used to create paired training data in object customization methods [5, 40], our approach requires a triplet of images: two image prompts and one target image. The image prompts should contain similar objects but exhibit distinct poses compared to the target image. To collect these triplets, we propose a data engine to curate our finetuning set. We leverage an advanced text-to-image generation model [2] to generate triplets where the same object pair is shared across the images. Through text prompt guidance, the image prompt provides strong identity information. This enables the decoupling of the relationship in the target image, which enhances the relation learning." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 55, + 654, + 295, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 654, + 295, + 713 + ], + "spans": [ + { + "bbox": [ + 55, + 654, + 295, + 713 + ], + "type": "text", + "content": "For the model design, we propose DreamRelation, which applies the Low-Rank Adaptation (LoRA) [9] strategy to the text cross-attention layers of existing diffusion models to process user-provided text prompts. In DreamRelation, two key modules are introduced during training to" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 313, + 72, + 553, + 263 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 72, + 553, + 263 + ], + "spans": [ + { + "bbox": [ + 313, + 72, + 553, + 263 + ], + "type": "text", + "content": "enhance the relation generation in customized generation. First, we introduce a keypoint matching loss (KML) as additional supervision to explicitly encourage the model to manipulate object poses, since relationships between objects are closely tied to their poses. Importantly, the KML operates on the latent representation rather than the original image space, aligning with the default diffusion loss. Second, since relation-aware customization requires local features from image prompts, such as the \"hands\" features for generating \"shaking hands\"—which are not captured by CLIP's coarse image-level features, we introduce dense features from CLIP. Through partitioning and pooling, we obtain local tokens that contain detail and local information. To further enhance the compatibility between dense features and image-level features, we employ a self-distillation method [39] to improve their alignment." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 313, + 264, + 553, + 335 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 264, + 553, + 335 + ], + "spans": [ + { + "bbox": [ + 313, + 264, + 553, + 335 + ], + "type": "text", + "content": "To more comprehensively evaluate relation-aware customized image generation, we constructed our Relation-Bench based on three established benchmarks [15, 31]. Our methods demonstrate strong performance compared to existing approaches, achieving significant improvements in visual quality and quantitative results." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 326, + 335, + 468, + 346 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 326, + 335, + 468, + 346 + ], + "spans": [ + { + "bbox": [ + 326, + 335, + 468, + 346 + ], + "type": "text", + "content": "The contributions of this work are:" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 314, + 348, + 553, + 538 + ], + "type": "list", + "angle": 0, + "index": 12, + "blocks": [ + { + "bbox": [ + 314, + 348, + 553, + 407 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 314, + 348, + 553, + 407 + ], + "spans": [ + { + "bbox": [ + 314, + 348, + 553, + 407 + ], + "type": "text", + "content": "- We explore a novel task called relation-aware customized generation, which aims to preserve multiple identities from image prompts while adhering to the relationships specified in text prompts. This task can enhance various user-driven applications by enabling more control." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 314, + 407, + 553, + 479 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 314, + 407, + 553, + 479 + ], + "spans": [ + { + "bbox": [ + 314, + 407, + 553, + 479 + ], + "type": "text", + "content": "- We introduce a data engine that uses the recent text-to-image generation model to generate triplet images where the same object is present with distinct poses. This well-curated dataset enables the model to focus on relation learning during fine-tuning, leveraging the identity information provided by the image prompts." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 314, + 479, + 553, + 538 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 314, + 479, + 553, + 538 + ], + "spans": [ + { + "bbox": [ + 314, + 479, + 553, + 538 + ], + "type": "text", + "content": "- Our proposed DreamRelation has two key modules, including keypoint matching loss and local token injection, to enhance the relation generation in customized generation. The extensive experiments on three benchmarks demonstrate the effectiveness of our method." + } + ] + } + ], + "index": 11 + } + ], + "sub_type": "text" + }, + { + "bbox": [ + 314, + 549, + 400, + 562 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 314, + 549, + 400, + 562 + ], + "spans": [ + { + "bbox": [ + 314, + 549, + 400, + 562 + ], + "type": "text", + "content": "2. Related Work" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 313, + 570, + 553, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 570, + 553, + 713 + ], + "spans": [ + { + "bbox": [ + 313, + 570, + 553, + 713 + ], + "type": "text", + "content": "Diffusion-Based Text-to-Image Generation. Diffusion-based text-to-image models [30, 32] generate high-quality images based on user-provided text prompts. These works encode the text prompt through the text encoder [6, 28] and inject the text embedding into the U-Net's cross-attention layers. Some methods [26, 27, 29, 38] upscale diffusion models and incorporate additional conditions as priors to generate high-resolution images. Meanwhile, several methods [2, 17] introduce stronger text encoders or large language models (LLMs) to enhance the text comprehension capabilities of diffusion models. Recent works [4, 25] have replaced the U-Net denoiser with the Transformers [33]." + } + ] + } + ], + "index": 14 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "type": "text", + "content": "15724" + } + ] + } + ], + "index": 15 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 1 + }, + { + "para_blocks": [ + { + "bbox": [ + 55, + 72, + 294, + 133 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 72, + 294, + 133 + ], + "spans": [ + { + "bbox": [ + 55, + 72, + 294, + 133 + ], + "type": "text", + "content": "However, these models primarily focus on text conditions and struggle to handle other forms of guidance, such as user-provided images. In contrast, our work facilitates customized generation while following relation conditions given by text inputs." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 56, + 148, + 296, + 460 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 148, + 296, + 460 + ], + "spans": [ + { + "bbox": [ + 56, + 148, + 296, + 460 + ], + "type": "text", + "content": "Diffusion-Based Customized Generation. Customized image generation aims to produce new images based on user-provided objects. Tuning-based methods accomplish this by fine-tuning specific parameters of diffusion models to learn new objects. Some of these methods [7, 20, 34] employ text embeddings to represent the object identity. On the other hand, DreamBooth [31] fine-tunes the entire U-Net and introduces a prior preservation loss to mitigate language drift. Additionally, several approaches [8, 15, 18, 23] integrate text embeddings and U-Net parameters for image customization. These methods fine-tune both components during object learning, achieving impressive results. Considering the substantial cost of fine-tuning for commercial uses, training-based methods [5, 16, 24, 36, 43] have been proposed. This approach primarily utilizes an encoder to extract the object's identity and inject it into the U-Net. For instance, MS-Diffusion [35] integrates grounding tokens with a Resampler to maintain detailed identity and employs layout guidance to explicitly locate the objects. Despite their efficiency during inference, training-based methods often face challenges in balancing identity preservation and text control. A common issue is that the relationships described in text prompts are frequently overlooked. To address this limitation, we introduce DreamRelation, a framework designed to bridge the customization and relation generation." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 55, + 474, + 296, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 474, + 296, + 714 + ], + "spans": [ + { + "bbox": [ + 55, + 474, + 296, + 714 + ], + "type": "text", + "content": "Relation-Aware Text-to-Image Generation. Inspired by tuning-based customization methods [7], some recent works [11, 22, 37] represent a \"neglected word\" by learnable parameters. These methods fine-tune part of the parameters on the content co-existing images. For instance, Reversion [11] fine-tunes the text embedding on a set of relation co-existing images and introduces a relation-steering contrastive loss. Reversion gains a better alignment between relational words and generated images by injecting new text embedding into prompts. While effective with prepositions and adjectives, Reversion struggles with relationships that involve significant overlaps, and cannot customize user-provided objects. Moreover, this method still struggles to generate vivid relationships while maintaining the fidelity of multiple customized objects, likely due to the inherent disregard for text embeddings in customization methods. Our work bridges the gap in the predicate relation-following capability of training-based customization methods. This enables efficient inference that naturally and accurately generates relationships in the text prompts." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 314, + 71, + 553, + 97 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 314, + 71, + 553, + 97 + ], + "spans": [ + { + "bbox": [ + 314, + 71, + 553, + 97 + ], + "type": "text", + "content": "3. Bridging Customization and Relation Generation" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 314, + 104, + 424, + 116 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 314, + 104, + 424, + 116 + ], + "spans": [ + { + "bbox": [ + 314, + 104, + 424, + 116 + ], + "type": "text", + "content": "3.1. Problem definition" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 313, + 122, + 553, + 205 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 122, + 553, + 205 + ], + "spans": [ + { + "bbox": [ + 313, + 122, + 553, + 205 + ], + "type": "text", + "content": "Different from conventional customization tasks, we explore a new setting that focuses on image generation by both image prompts " + }, + { + "bbox": [ + 313, + 122, + 553, + 205 + ], + "type": "inline_equation", + "content": "c_{i} \\in \\mathcal{R}^{N \\times 3 \\times H \\times W}" + }, + { + "bbox": [ + 313, + 122, + 553, + 205 + ], + "type": "text", + "content": " and text prompts " + }, + { + "bbox": [ + 313, + 122, + 553, + 205 + ], + "type": "inline_equation", + "content": "c_{t}" + }, + { + "bbox": [ + 313, + 122, + 553, + 205 + ], + "type": "text", + "content": ". We call this setting relation-aware customized image generation due to the requirements that the generated image " + }, + { + "bbox": [ + 313, + 122, + 553, + 205 + ], + "type": "inline_equation", + "content": "\\hat{x}" + }, + { + "bbox": [ + 313, + 122, + 553, + 205 + ], + "type": "text", + "content": " should strictly preserve each identity and keep the relationship among those identities provided by " + }, + { + "bbox": [ + 313, + 122, + 553, + 205 + ], + "type": "inline_equation", + "content": "c_{i}" + }, + { + "bbox": [ + 313, + 122, + 553, + 205 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 313, + 122, + 553, + 205 + ], + "type": "inline_equation", + "content": "c_{t}" + }, + { + "bbox": [ + 313, + 122, + 553, + 205 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 326, + 206, + 418, + 217 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 326, + 206, + 418, + 217 + ], + "spans": [ + { + "bbox": [ + 326, + 206, + 418, + 217 + ], + "type": "text", + "content": "We define this task as:" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 403, + 225, + 553, + 237 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 403, + 225, + 553, + 237 + ], + "spans": [ + { + "bbox": [ + 403, + 225, + 553, + 237 + ], + "type": "interline_equation", + "content": "\\hat {x} = \\Phi_ {\\theta} \\left(c _ {i}, c _ {t}\\right) \\tag {1}", + "image_path": "3b627d66aa5d201996f6587ab65a4532d8f8ac8f36d8e2d137a2c7579674282e.jpg" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 313, + 244, + 553, + 281 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 244, + 553, + 281 + ], + "spans": [ + { + "bbox": [ + 313, + 244, + 553, + 281 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 313, + 244, + 553, + 281 + ], + "type": "inline_equation", + "content": "\\Phi_{\\theta}" + }, + { + "bbox": [ + 313, + 244, + 553, + 281 + ], + "type": "text", + "content": " is the network parameterized by " + }, + { + "bbox": [ + 313, + 244, + 553, + 281 + ], + "type": "inline_equation", + "content": "\\theta" + }, + { + "bbox": [ + 313, + 244, + 553, + 281 + ], + "type": "text", + "content": ". For brevity clarification, we set " + }, + { + "bbox": [ + 313, + 244, + 553, + 281 + ], + "type": "inline_equation", + "content": "N = 2" + }, + { + "bbox": [ + 313, + 244, + 553, + 281 + ], + "type": "text", + "content": " due to the basic triplet element defined in the DreamRelation." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 314, + 286, + 462, + 298 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 314, + 286, + 462, + 298 + ], + "spans": [ + { + "bbox": [ + 314, + 286, + 462, + 298 + ], + "type": "text", + "content": "3.2. Discussion and Motivation" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 313, + 304, + 555, + 424 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 304, + 555, + 424 + ], + "spans": [ + { + "bbox": [ + 313, + 304, + 555, + 424 + ], + "type": "text", + "content": "While previous works [8, 15, 43] have explored the customization of multiple objects under text control, they lack the ability to effectively manage the relationships between these objects. Specifically, the state-of-the-art multi-object customization model [35] utilizes bounding boxes to enhance multi-object generation. Given the image prompt " + }, + { + "bbox": [ + 313, + 304, + 555, + 424 + ], + "type": "inline_equation", + "content": "c_{i}" + }, + { + "bbox": [ + 313, + 304, + 555, + 424 + ], + "type": "text", + "content": " and the text prompt " + }, + { + "bbox": [ + 313, + 304, + 555, + 424 + ], + "type": "inline_equation", + "content": "c_{t}" + }, + { + "bbox": [ + 313, + 304, + 555, + 424 + ], + "type": "text", + "content": ", it leverages CLIP [28] and the Resampler [35] to extract text tokens " + }, + { + "bbox": [ + 313, + 304, + 555, + 424 + ], + "type": "inline_equation", + "content": "tok_{\\text{text}}" + }, + { + "bbox": [ + 313, + 304, + 555, + 424 + ], + "type": "text", + "content": " and image tokens " + }, + { + "bbox": [ + 313, + 304, + 555, + 424 + ], + "type": "inline_equation", + "content": "tok_{\\text{image}}" + }, + { + "bbox": [ + 313, + 304, + 555, + 424 + ], + "type": "text", + "content": ", respectively. These tokens are then injected into the U-Net " + }, + { + "bbox": [ + 313, + 304, + 555, + 424 + ], + "type": "inline_equation", + "content": "\\epsilon_{\\theta}" + }, + { + "bbox": [ + 313, + 304, + 555, + 424 + ], + "type": "text", + "content": " through parallel cross-attention layers:" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 325, + 429, + 553, + 455 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 325, + 429, + 553, + 455 + ], + "spans": [ + { + "bbox": [ + 325, + 429, + 553, + 455 + ], + "type": "interline_equation", + "content": "h = \\operatorname {S o f t m a x} \\left(\\frac {Q K _ {i}}{\\sqrt {d}} + \\mathcal {M}\\right) V _ {i} + \\operatorname {S o f t m a x} \\left(\\frac {Q K _ {t}}{\\sqrt {d}}\\right) V _ {t} \\tag {2}", + "image_path": "54202a89dc7d254888b826dab1a0364f05dc9377afe326f4cdd696bb5fee89e4.jpg" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 313, + 460, + 554, + 567 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 460, + 554, + 567 + ], + "spans": [ + { + "bbox": [ + 313, + 460, + 554, + 567 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 313, + 460, + 554, + 567 + ], + "type": "inline_equation", + "content": "Q" + }, + { + "bbox": [ + 313, + 460, + 554, + 567 + ], + "type": "text", + "content": " represents the query, while " + }, + { + "bbox": [ + 313, + 460, + 554, + 567 + ], + "type": "inline_equation", + "content": "K_{i}, V_{i}" + }, + { + "bbox": [ + 313, + 460, + 554, + 567 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 313, + 460, + 554, + 567 + ], + "type": "inline_equation", + "content": "K_{t}, V_{t}" + }, + { + "bbox": [ + 313, + 460, + 554, + 567 + ], + "type": "text", + "content": " denote the keys and values obtained from " + }, + { + "bbox": [ + 313, + 460, + 554, + 567 + ], + "type": "inline_equation", + "content": "tok_{image}" + }, + { + "bbox": [ + 313, + 460, + 554, + 567 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 313, + 460, + 554, + 567 + ], + "type": "inline_equation", + "content": "tok_{text}" + }, + { + "bbox": [ + 313, + 460, + 554, + 567 + ], + "type": "text", + "content": ", respectively. The mask " + }, + { + "bbox": [ + 313, + 460, + 554, + 567 + ], + "type": "inline_equation", + "content": "\\mathcal{M}" + }, + { + "bbox": [ + 313, + 460, + 554, + 567 + ], + "type": "text", + "content": ", generated from the bounding boxes, assists the model in localizing objects during multi-object customization. However, it encounters difficulties when bounding boxes overlap, frequently resulting in object confusion, as shown in Fig. 1. Consequently, relation-aware generation in customized models remains a largely unexplored area of research." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 313, + 567, + 554, + 651 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 567, + 554, + 651 + ], + "spans": [ + { + "bbox": [ + 313, + 567, + 554, + 651 + ], + "type": "text", + "content": "On the other hand, we revisit the relation-aware generation task. Existing relation inversion methods [11, 22] focus on inverting a relationship into a text embedding " + }, + { + "bbox": [ + 313, + 567, + 554, + 651 + ], + "type": "inline_equation", + "content": "R^{*}" + }, + { + "bbox": [ + 313, + 567, + 554, + 651 + ], + "type": "text", + "content": " within pre-trained text-to-image diffusion models. Given a set of images " + }, + { + "bbox": [ + 313, + 567, + 554, + 651 + ], + "type": "inline_equation", + "content": "\\{x_{k}\\}_{k = 1}^{n}" + }, + { + "bbox": [ + 313, + 567, + 554, + 651 + ], + "type": "text", + "content": " that share the same relationship, these methods use a denoising loss " + }, + { + "bbox": [ + 313, + 567, + 554, + 651 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{RI}" + }, + { + "bbox": [ + 313, + 567, + 554, + 651 + ], + "type": "text", + "content": " to fine-tune " + }, + { + "bbox": [ + 313, + 567, + 554, + 651 + ], + "type": "inline_equation", + "content": "R^{*}" + }, + { + "bbox": [ + 313, + 567, + 554, + 651 + ], + "type": "text", + "content": ", ensuring alignment with the specific relationship:" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 358, + 658, + 553, + 671 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 358, + 658, + 553, + 671 + ], + "spans": [ + { + "bbox": [ + 358, + 658, + 553, + 671 + ], + "type": "interline_equation", + "content": "\\mathcal {L} _ {R I} = \\mathbb {E} _ {z _ {t}, t, \\epsilon , c} [ \\| \\epsilon_ {t} - \\boldsymbol {\\epsilon} _ {\\boldsymbol {\\theta}} (z _ {t}, t, c _ {t}) \\| _ {2} ^ {2} ] \\tag {3}", + "image_path": "c3cc55d3b9058a08033dff5220ad373be2e21f4bc48600a69ee0b3bb6de62588.jpg" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 313, + 677, + 554, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 677, + 554, + 713 + ], + "spans": [ + { + "bbox": [ + 313, + 677, + 554, + 713 + ], + "type": "text", + "content": "Directly applying relation-aware generation methods to customized models leads to suboptimal performance, as illustrated in Fig. 4. We attribute this to two main reasons." + } + ] + } + ], + "index": 15 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "text", + "content": "15725" + } + ] + } + ], + "index": 16 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 2 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 56, + 70, + 553, + 236 + ], + "blocks": [ + { + "bbox": [ + 56, + 70, + 553, + 236 + ], + "lines": [ + { + "bbox": [ + 56, + 70, + 553, + 236 + ], + "spans": [ + { + "bbox": [ + 56, + 70, + 553, + 236 + ], + "type": "image", + "image_path": "179ca59630421d63fad72096ed5ba6770a9ce18feecc3fef2f58a9ec048377c3.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 54, + 240, + 555, + 275 + ], + "lines": [ + { + "bbox": [ + 54, + 240, + 555, + 275 + ], + "spans": [ + { + "bbox": [ + 54, + 240, + 555, + 275 + ], + "type": "text", + "content": "Figure 2. The overview of DreamRelation. DreamRelation utilizes the off-the-shelf identity extractor to decouple the relation and identity information in relation-specific images. After getting the U-Net output " + }, + { + "bbox": [ + 54, + 240, + 555, + 275 + ], + "type": "inline_equation", + "content": "\\hat{\\epsilon}" + }, + { + "bbox": [ + 54, + 240, + 555, + 275 + ], + "type": "text", + "content": ", we predict " + }, + { + "bbox": [ + 54, + 240, + 555, + 275 + ], + "type": "inline_equation", + "content": "\\hat{z}_0" + }, + { + "bbox": [ + 54, + 240, + 555, + 275 + ], + "type": "text", + "content": " and calculate the keypoint matching loss. The part in the dotted box is only for training." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "bbox": [ + 56, + 281, + 296, + 448 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 281, + 296, + 448 + ], + "spans": [ + { + "bbox": [ + 56, + 281, + 296, + 448 + ], + "type": "text", + "content": "First, the balance between the control of " + }, + { + "bbox": [ + 56, + 281, + 296, + 448 + ], + "type": "inline_equation", + "content": "c_{i}" + }, + { + "bbox": [ + 56, + 281, + 296, + 448 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 56, + 281, + 296, + 448 + ], + "type": "inline_equation", + "content": "c_{t}" + }, + { + "bbox": [ + 56, + 281, + 296, + 448 + ], + "type": "text", + "content": " is not addressed in relation-aware generation tasks. During inference, the model tends to overlook the target relationship " + }, + { + "bbox": [ + 56, + 281, + 296, + 448 + ], + "type": "inline_equation", + "content": "R^{*}" + }, + { + "bbox": [ + 56, + 281, + 296, + 448 + ], + "type": "text", + "content": " due to the influence of " + }, + { + "bbox": [ + 56, + 281, + 296, + 448 + ], + "type": "inline_equation", + "content": "c_{i}" + }, + { + "bbox": [ + 56, + 281, + 296, + 448 + ], + "type": "text", + "content": ", resulting in incorrect relationship generation. Second, even if the shared relationship exists across " + }, + { + "bbox": [ + 56, + 281, + 296, + 448 + ], + "type": "inline_equation", + "content": "\\{x_{k}\\}_{k = 1}^{n}" + }, + { + "bbox": [ + 56, + 281, + 296, + 448 + ], + "type": "text", + "content": ", the disentangled object identity information complicates the learning process. Fine-tuning solely on " + }, + { + "bbox": [ + 56, + 281, + 296, + 448 + ], + "type": "inline_equation", + "content": "\\{x_{k}\\}_{k = 1}^{n}" + }, + { + "bbox": [ + 56, + 281, + 296, + 448 + ], + "type": "text", + "content": " makes it challenging for the model to effectively capture the relationship. As shown in the second row of Fig. 4, where ReVersion [11] struggles to generate the \"hugging\" relationship even in the absence of " + }, + { + "bbox": [ + 56, + 281, + 296, + 448 + ], + "type": "inline_equation", + "content": "c_{i}" + }, + { + "bbox": [ + 56, + 281, + 296, + 448 + ], + "type": "text", + "content": ". In the following subsections, we will introduce our data collection method and DreamRelation to bridge customization and relation generation." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 55, + 456, + 151, + 468 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 456, + 151, + 468 + ], + "spans": [ + { + "bbox": [ + 55, + 456, + 151, + 468 + ], + "type": "text", + "content": "3.3. Data Collection" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 55, + 474, + 295, + 617 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 474, + 295, + 617 + ], + "spans": [ + { + "bbox": [ + 55, + 474, + 295, + 617 + ], + "type": "text", + "content": "To address the balance between " + }, + { + "bbox": [ + 55, + 474, + 295, + 617 + ], + "type": "inline_equation", + "content": "c_{i}" + }, + { + "bbox": [ + 55, + 474, + 295, + 617 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 55, + 474, + 295, + 617 + ], + "type": "inline_equation", + "content": "c_{t}" + }, + { + "bbox": [ + 55, + 474, + 295, + 617 + ], + "type": "text", + "content": " while facilitating relation learning from " + }, + { + "bbox": [ + 55, + 474, + 295, + 617 + ], + "type": "inline_equation", + "content": "\\{x_{k}\\}_{k = 1}^{n}" + }, + { + "bbox": [ + 55, + 474, + 295, + 617 + ], + "type": "text", + "content": ", we first design a data engine for collecting high-quality tuning data. The ideal tuning data should be in the form of " + }, + { + "bbox": [ + 55, + 474, + 295, + 617 + ], + "type": "inline_equation", + "content": "\\mathcal{D} = (x_k, c_i, c_t)" + }, + { + "bbox": [ + 55, + 474, + 295, + 617 + ], + "type": "text", + "content": ". The " + }, + { + "bbox": [ + 55, + 474, + 295, + 617 + ], + "type": "inline_equation", + "content": "c_{i}" + }, + { + "bbox": [ + 55, + 474, + 295, + 617 + ], + "type": "text", + "content": " contain the object in " + }, + { + "bbox": [ + 55, + 474, + 295, + 617 + ], + "type": "inline_equation", + "content": "x_{k}" + }, + { + "bbox": [ + 55, + 474, + 295, + 617 + ], + "type": "text", + "content": " while " + }, + { + "bbox": [ + 55, + 474, + 295, + 617 + ], + "type": "inline_equation", + "content": "c_{t}" + }, + { + "bbox": [ + 55, + 474, + 295, + 617 + ], + "type": "text", + "content": " describes the relationship between objects in " + }, + { + "bbox": [ + 55, + 474, + 295, + 617 + ], + "type": "inline_equation", + "content": "x_{k}" + }, + { + "bbox": [ + 55, + 474, + 295, + 617 + ], + "type": "text", + "content": ". The identity information in " + }, + { + "bbox": [ + 55, + 474, + 295, + 617 + ], + "type": "inline_equation", + "content": "c_{i}" + }, + { + "bbox": [ + 55, + 474, + 295, + 617 + ], + "type": "text", + "content": " helps decouple the relationship within " + }, + { + "bbox": [ + 55, + 474, + 295, + 617 + ], + "type": "inline_equation", + "content": "x_{k}" + }, + { + "bbox": [ + 55, + 474, + 295, + 617 + ], + "type": "text", + "content": ", enhancing relation learning while maintaining a balanced integration of " + }, + { + "bbox": [ + 55, + 474, + 295, + 617 + ], + "type": "inline_equation", + "content": "c_{i}" + }, + { + "bbox": [ + 55, + 474, + 295, + 617 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 55, + 474, + 295, + 617 + ], + "type": "inline_equation", + "content": "c_{t}" + }, + { + "bbox": [ + 55, + 474, + 295, + 617 + ], + "type": "text", + "content": ". Unfortunately, directly cropping the objects from " + }, + { + "bbox": [ + 55, + 474, + 295, + 617 + ], + "type": "inline_equation", + "content": "x_{k}" + }, + { + "bbox": [ + 55, + 474, + 295, + 617 + ], + "type": "text", + "content": " to obtain " + }, + { + "bbox": [ + 55, + 474, + 295, + 617 + ], + "type": "inline_equation", + "content": "c_{i}" + }, + { + "bbox": [ + 55, + 474, + 295, + 617 + ], + "type": "text", + "content": " results in a copy-and-paste effect [5], even when using data augmentation techniques like flipping and rotation. Therefore, we introduce our data engine below." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 55, + 617, + 295, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 617, + 295, + 714 + ], + "spans": [ + { + "bbox": [ + 55, + 617, + 295, + 714 + ], + "type": "text", + "content": "Relation-aware data engine. Fig. 3 illustrates the difference of our data engine. Both of " + }, + { + "bbox": [ + 55, + 617, + 295, + 714 + ], + "type": "inline_equation", + "content": "x_{k}" + }, + { + "bbox": [ + 55, + 617, + 295, + 714 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 55, + 617, + 295, + 714 + ], + "type": "inline_equation", + "content": "c_{i}" + }, + { + "bbox": [ + 55, + 617, + 295, + 714 + ], + "type": "text", + "content": " are generated by the recent text-to-image generation model, rather than through cropping and augmentation. Specifically, inspired by the performance of DALL-E 3 and its multi-turn dialog capability. We use DALL-E 3 to generate " + }, + { + "bbox": [ + 55, + 617, + 295, + 714 + ], + "type": "inline_equation", + "content": "\\mathcal{D} = (x,c_i,c_t)" + }, + { + "bbox": [ + 55, + 617, + 295, + 714 + ], + "type": "text", + "content": ". Where " + }, + { + "bbox": [ + 55, + 617, + 295, + 714 + ], + "type": "inline_equation", + "content": "x_{k}" + }, + { + "bbox": [ + 55, + 617, + 295, + 714 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 55, + 617, + 295, + 714 + ], + "type": "inline_equation", + "content": "c_{i}" + }, + { + "bbox": [ + 55, + 617, + 295, + 714 + ], + "type": "text", + "content": " share the same object identity by leveraging the prompt \"The photo of the same.\" We observe that" + } + ] + } + ], + "index": 5 + }, + { + "type": "image", + "bbox": [ + 318, + 278, + 565, + 475 + ], + "blocks": [ + { + "bbox": [ + 318, + 278, + 565, + 475 + ], + "lines": [ + { + "bbox": [ + 318, + 278, + 565, + 475 + ], + "spans": [ + { + "bbox": [ + 318, + 278, + 565, + 475 + ], + "type": "image", + "image_path": "7a43ac3f81a67b732255010dc504844454122f94c365dc3bbc25b93779444dde.jpg" + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 313, + 485, + 555, + 552 + ], + "lines": [ + { + "bbox": [ + 313, + 485, + 555, + 552 + ], + "spans": [ + { + "bbox": [ + 313, + 485, + 555, + 552 + ], + "type": "text", + "content": "Figure 3. The rigid object cropping data engine, which results in minimal changes to object pose, the cropped image prompts also contain relation information, leading to a copy-and-paste effect, where the text prompt is neglected. Our relation-aware data engine, on the other hand, focuses on relation learning by decoupling identity information from images." + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "image_caption" + } + ], + "index": 6 + }, + { + "bbox": [ + 313, + 560, + 555, + 621 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 560, + 555, + 621 + ], + "spans": [ + { + "bbox": [ + 313, + 560, + 555, + 621 + ], + "type": "text", + "content": "prompting in this manner allows DALL-E 3 to remember and preserve identity in common categories such as \"tiger\", which is enough for relation learning. Next, we employ X-Pose [41], SAM [13], and LLaVA [19] to annotate " + }, + { + "bbox": [ + 313, + 560, + 555, + 621 + ], + "type": "inline_equation", + "content": "x" + }, + { + "bbox": [ + 313, + 560, + 555, + 621 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 313, + 560, + 555, + 621 + ], + "type": "inline_equation", + "content": "c_i" + }, + { + "bbox": [ + 313, + 560, + 555, + 621 + ], + "type": "text", + "content": " with keypoints, masks, and captions for samples." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 313, + 631, + 409, + 642 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 631, + 409, + 642 + ], + "spans": [ + { + "bbox": [ + 313, + 631, + 409, + 642 + ], + "type": "text", + "content": "3.4. DreamRelation" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 313, + 650, + 421, + 662 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 650, + 421, + 662 + ], + "spans": [ + { + "bbox": [ + 313, + 650, + 421, + 662 + ], + "type": "text", + "content": "3.4.1. Relation Learning" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 313, + 666, + 555, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 666, + 555, + 714 + ], + "spans": [ + { + "bbox": [ + 313, + 666, + 555, + 714 + ], + "type": "text", + "content": "To balance " + }, + { + "bbox": [ + 313, + 666, + 555, + 714 + ], + "type": "inline_equation", + "content": "c_{i}" + }, + { + "bbox": [ + 313, + 666, + 555, + 714 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 313, + 666, + 555, + 714 + ], + "type": "inline_equation", + "content": "c_{t}" + }, + { + "bbox": [ + 313, + 666, + 555, + 714 + ], + "type": "text", + "content": " while enhancing relation learning, we decouple the identity information and relationships in " + }, + { + "bbox": [ + 313, + 666, + 555, + 714 + ], + "type": "inline_equation", + "content": "x_{k}" + }, + { + "bbox": [ + 313, + 666, + 555, + 714 + ], + "type": "text", + "content": ". Specifically, we leverage an off-the-shelf ID extractor from the state-of-the-art customized model [35]. The identity in-" + } + ] + } + ], + "index": 11 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "type": "text", + "content": "15726" + } + ] + } + ], + "index": 12 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 3 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 59, + 70, + 294, + 184 + ], + "blocks": [ + { + "bbox": [ + 59, + 70, + 294, + 184 + ], + "lines": [ + { + "bbox": [ + 59, + 70, + 294, + 184 + ], + "spans": [ + { + "bbox": [ + 59, + 70, + 294, + 184 + ], + "type": "table", + "html": "
MethodsSingle-Object CusMulti-Object CusNon-ConfusingRelation-Aware
Textual Inversion [7]✓✗✗✗
DreamBooth [31]✓✗✗✗
Custom Diffusion [15]✓✓✗✗
Mix-of-Show [8]✓✓✗✗
CLIF [18]✓✓✓✗
MultiBooth [45]✓✓✗✗
BLIP-Diffusion [16]✓✗✗✗
ELITE [36]✓✗✗✗
AnyDoor [5]✓✗✗✗
SSR-Encoder [43]✓✓✗✗
λ-eclipse [24]✓✓✗✗
MS-Diffusion [35]✓✓✗✗
ReVersion [11]✗✗✓✓
ADI [10]✗✗✗✓
DreamRelation✓✓✓✓
", + "image_path": "dd50c029d71fa0c16d4b9664b266989f8eac8675f626f1cba7165c55835ce89a.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "table_body" + } + ], + "index": 0 + }, + { + "bbox": [ + 55, + 188, + 295, + 222 + ], + "lines": [ + { + "bbox": [ + 55, + 188, + 295, + 222 + ], + "spans": [ + { + "bbox": [ + 55, + 188, + 295, + 222 + ], + "type": "text", + "content": "Table 1. Setting Comparison for Different Models. We include several representative methods here. Our proposed DreamRelation can meet the demand of the relation-aware customization task." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "text" + }, + { + "type": "image", + "bbox": [ + 62, + 236, + 287, + 377 + ], + "blocks": [ + { + "bbox": [ + 62, + 236, + 287, + 377 + ], + "lines": [ + { + "bbox": [ + 62, + 236, + 287, + 377 + ], + "spans": [ + { + "bbox": [ + 62, + 236, + 287, + 377 + ], + "type": "image", + "image_path": "91f5be357e43116896e2acb82eb1f233545ff1bd6c4dc219ab993a11cbc320d0.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 55, + 381, + 295, + 415 + ], + "lines": [ + { + "bbox": [ + 55, + 381, + 295, + 415 + ], + "spans": [ + { + "bbox": [ + 55, + 381, + 295, + 415 + ], + "type": "text", + "content": "Figure 4. Comparing our method with ReVersion across different base models, our approach demonstrates superior performance in the relation-aware generation task." + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_caption" + } + ], + "index": 2 + }, + { + "bbox": [ + 55, + 429, + 296, + 477 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 429, + 296, + 477 + ], + "spans": [ + { + "bbox": [ + 55, + 429, + 296, + 477 + ], + "type": "text", + "content": "formation in " + }, + { + "bbox": [ + 55, + 429, + 296, + 477 + ], + "type": "inline_equation", + "content": "c_{i}" + }, + { + "bbox": [ + 55, + 429, + 296, + 477 + ], + "type": "text", + "content": " enables the model to focus on the relationship information in " + }, + { + "bbox": [ + 55, + 429, + 296, + 477 + ], + "type": "inline_equation", + "content": "x_{k}" + }, + { + "bbox": [ + 55, + 429, + 296, + 477 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 55, + 429, + 296, + 477 + ], + "type": "inline_equation", + "content": "c_{t}" + }, + { + "bbox": [ + 55, + 429, + 296, + 477 + ], + "type": "text", + "content": " during relation learning. To achieve this, we employ parallel cross-attention layers to process " + }, + { + "bbox": [ + 55, + 429, + 296, + 477 + ], + "type": "inline_equation", + "content": "c_{i}" + }, + { + "bbox": [ + 55, + 429, + 296, + 477 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 55, + 429, + 296, + 477 + ], + "type": "inline_equation", + "content": "c_{t}" + }, + { + "bbox": [ + 55, + 429, + 296, + 477 + ], + "type": "text", + "content": ", which can be formulated as:" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 60, + 491, + 295, + 519 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 60, + 491, + 295, + 519 + ], + "spans": [ + { + "bbox": [ + 60, + 491, + 295, + 519 + ], + "type": "interline_equation", + "content": "h = \\gamma \\cdot \\operatorname {S o f t m a x} \\left(\\frac {Q K _ {i}}{\\sqrt {d}} + \\mathcal {M}\\right) V _ {i} + \\operatorname {S o f t m a x} \\left(\\frac {Q K _ {t}}{\\sqrt {d}}\\right) V _ {t} \\tag {4}", + "image_path": "daf9760ff60a6cddc2c220aca9b2f5e5159c3b00f5d0240869a04fe247644626.jpg" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 55, + 533, + 295, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 533, + 295, + 714 + ], + "spans": [ + { + "bbox": [ + 55, + 533, + 295, + 714 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 55, + 533, + 295, + 714 + ], + "type": "inline_equation", + "content": "Q = W_{q}h" + }, + { + "bbox": [ + 55, + 533, + 295, + 714 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 55, + 533, + 295, + 714 + ], + "type": "inline_equation", + "content": "K_{i} = W_{k_{i}}c_{i}" + }, + { + "bbox": [ + 55, + 533, + 295, + 714 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 55, + 533, + 295, + 714 + ], + "type": "inline_equation", + "content": "V_{i} = W_{v_{i}}c_{i}" + }, + { + "bbox": [ + 55, + 533, + 295, + 714 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 55, + 533, + 295, + 714 + ], + "type": "inline_equation", + "content": "K_{t} = W_{k_{t}}c_{t}" + }, + { + "bbox": [ + 55, + 533, + 295, + 714 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 55, + 533, + 295, + 714 + ], + "type": "inline_equation", + "content": "V_{t} = W_{v_{t}}c_{t}" + }, + { + "bbox": [ + 55, + 533, + 295, + 714 + ], + "type": "text", + "content": ", and " + }, + { + "bbox": [ + 55, + 533, + 295, + 714 + ], + "type": "inline_equation", + "content": "\\gamma" + }, + { + "bbox": [ + 55, + 533, + 295, + 714 + ], + "type": "text", + "content": " is a hyperparameter for scaling control from " + }, + { + "bbox": [ + 55, + 533, + 295, + 714 + ], + "type": "inline_equation", + "content": "c_{i}" + }, + { + "bbox": [ + 55, + 533, + 295, + 714 + ], + "type": "text", + "content": ". For clarity, we omit the final linear layer " + }, + { + "bbox": [ + 55, + 533, + 295, + 714 + ], + "type": "inline_equation", + "content": "W_{out}" + }, + { + "bbox": [ + 55, + 533, + 295, + 714 + ], + "type": "text", + "content": ". The relation information mainly depends on text prompts. Therefore, we inject LoRA layers in all of the " + }, + { + "bbox": [ + 55, + 533, + 295, + 714 + ], + "type": "inline_equation", + "content": "W_{q}" + }, + { + "bbox": [ + 55, + 533, + 295, + 714 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 55, + 533, + 295, + 714 + ], + "type": "inline_equation", + "content": "W_{k_{t}}" + }, + { + "bbox": [ + 55, + 533, + 295, + 714 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 55, + 533, + 295, + 714 + ], + "type": "inline_equation", + "content": "W_{v_{t}}" + }, + { + "bbox": [ + 55, + 533, + 295, + 714 + ], + "type": "text", + "content": ", and " + }, + { + "bbox": [ + 55, + 533, + 295, + 714 + ], + "type": "inline_equation", + "content": "W_{out}" + }, + { + "bbox": [ + 55, + 533, + 295, + 714 + ], + "type": "text", + "content": " to encourage the model to pay more attention to the relationship in " + }, + { + "bbox": [ + 55, + 533, + 295, + 714 + ], + "type": "inline_equation", + "content": "c_{t}" + }, + { + "bbox": [ + 55, + 533, + 295, + 714 + ], + "type": "text", + "content": ". We freeze all other parameters during fine-tuning. As shown in Fig. 4, benefiting from the decoupled identity information in " + }, + { + "bbox": [ + 55, + 533, + 295, + 714 + ], + "type": "inline_equation", + "content": "c_{i}" + }, + { + "bbox": [ + 55, + 533, + 295, + 714 + ], + "type": "text", + "content": ", our method effectively captures the relationship in " + }, + { + "bbox": [ + 55, + 533, + 295, + 714 + ], + "type": "inline_equation", + "content": "x_{k}" + }, + { + "bbox": [ + 55, + 533, + 295, + 714 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 55, + 533, + 295, + 714 + ], + "type": "inline_equation", + "content": "c_{t}" + }, + { + "bbox": [ + 55, + 533, + 295, + 714 + ], + "type": "text", + "content": ", accurately generating the corresponding images. Notably, after fine-tuning, our method can also be directly integrated into SDXL [27], effectively enabling the model to address the relation-aware generation task. The extensive results are presented in the Supplementary Material (SM)." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 314, + 72, + 448, + 85 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 314, + 72, + 448, + 85 + ], + "spans": [ + { + "bbox": [ + 314, + 72, + 448, + 85 + ], + "type": "text", + "content": "3.4.2. Keypoint Matching Loss" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 313, + 87, + 555, + 182 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 87, + 555, + 182 + ], + "spans": [ + { + "bbox": [ + 313, + 87, + 555, + 182 + ], + "type": "text", + "content": "Most relationships have specific requirements for the pose, such as \"hugging\" requiring the arms to cross over the object's body, and \"riding a bicycle\" requiring the feet of the object to be on the pedal plate. Therefore, it's important to manipulate the pose of objects in the right status accurately. Intuitively, we introduce explicit supervision in the diffusion latent space during the tuning process, named Keypoint Matching Loss (KML)." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 313, + 183, + 555, + 290 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 183, + 555, + 290 + ], + "spans": [ + { + "bbox": [ + 313, + 183, + 555, + 290 + ], + "type": "text", + "content": "Specifically, considering our task only involves common objects instead of humans, we employ X-Pose [41] as a keypoint detector because it can detect any keypoints in complex real-world scenarios. We detect 17 keypoints for each object in " + }, + { + "bbox": [ + 313, + 183, + 555, + 290 + ], + "type": "inline_equation", + "content": "x_{k}" + }, + { + "bbox": [ + 313, + 183, + 555, + 290 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 313, + 183, + 555, + 290 + ], + "type": "inline_equation", + "content": "c_{i}" + }, + { + "bbox": [ + 313, + 183, + 555, + 290 + ], + "type": "text", + "content": ", and denote the keypoint coordinates as " + }, + { + "bbox": [ + 313, + 183, + 555, + 290 + ], + "type": "inline_equation", + "content": "c_{kp}^{x_k}, c_{kp}^{c_i} \\in \\mathcal{R}^{17 \\times 2}" + }, + { + "bbox": [ + 313, + 183, + 555, + 290 + ], + "type": "text", + "content": " respectively. To encourage the model to generate accurate poses in " + }, + { + "bbox": [ + 313, + 183, + 555, + 290 + ], + "type": "inline_equation", + "content": "\\mathcal{D}(\\hat{z}_0)" + }, + { + "bbox": [ + 313, + 183, + 555, + 290 + ], + "type": "text", + "content": ", where " + }, + { + "bbox": [ + 313, + 183, + 555, + 290 + ], + "type": "inline_equation", + "content": "\\mathcal{D}" + }, + { + "bbox": [ + 313, + 183, + 555, + 290 + ], + "type": "text", + "content": " is the VAE decoder. We use the U-Net's output " + }, + { + "bbox": [ + 313, + 183, + 555, + 290 + ], + "type": "inline_equation", + "content": "\\hat{\\epsilon}" + }, + { + "bbox": [ + 313, + 183, + 555, + 290 + ], + "type": "text", + "content": " to predict " + }, + { + "bbox": [ + 313, + 183, + 555, + 290 + ], + "type": "inline_equation", + "content": "z_0" + }, + { + "bbox": [ + 313, + 183, + 555, + 290 + ], + "type": "text", + "content": " during fine-tuning:" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 390, + 289, + 554, + 316 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 390, + 289, + 554, + 316 + ], + "spans": [ + { + "bbox": [ + 390, + 289, + 554, + 316 + ], + "type": "interline_equation", + "content": "\\hat {z} _ {0} = \\frac {z _ {t} - \\sqrt {1 - \\bar {\\alpha} _ {t}} \\hat {\\epsilon}}{\\sqrt {\\bar {\\alpha} _ {t}}} \\tag {5}", + "image_path": "9b150346557efef25e17f71bac2688f2fb4d53ce76558416b1a1372072b590e4.jpg" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 313, + 319, + 555, + 366 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 319, + 555, + 366 + ], + "spans": [ + { + "bbox": [ + 313, + 319, + 555, + 366 + ], + "type": "text", + "content": "We use the VAE encoder " + }, + { + "bbox": [ + 313, + 319, + 555, + 366 + ], + "type": "inline_equation", + "content": "\\mathcal{E}" + }, + { + "bbox": [ + 313, + 319, + 555, + 366 + ], + "type": "text", + "content": " to obtain the latent representation of image prompts " + }, + { + "bbox": [ + 313, + 319, + 555, + 366 + ], + "type": "inline_equation", + "content": "\\mathcal{E}(c_i)" + }, + { + "bbox": [ + 313, + 319, + 555, + 366 + ], + "type": "text", + "content": ". Then we calculate the MSE loss on the corresponding keypoint locations between " + }, + { + "bbox": [ + 313, + 319, + 555, + 366 + ], + "type": "inline_equation", + "content": "\\mathcal{E}(c_i)" + }, + { + "bbox": [ + 313, + 319, + 555, + 366 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 313, + 319, + 555, + 366 + ], + "type": "inline_equation", + "content": "\\hat{z}_0" + }, + { + "bbox": [ + 313, + 319, + 555, + 366 + ], + "type": "text", + "content": ":" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 341, + 373, + 555, + 399 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 341, + 373, + 555, + 399 + ], + "spans": [ + { + "bbox": [ + 341, + 373, + 555, + 399 + ], + "type": "interline_equation", + "content": "\\mathcal {L} _ {K M L} = \\frac {1}{n _ {k p}} \\mathbb {E} _ {\\boldsymbol {z} _ {t}, c _ {i}} \\left\\| \\mathcal {E} (c _ {i}) [ c _ {k p} ^ {c _ {i}} ] - \\hat {z} _ {0} [ c _ {k p} ^ {x _ {k}} ] \\right\\| _ {2} ^ {2} \\tag {6}", + "image_path": "cabe9fd7d11a79988de7f23c8bfb86aa694da4b4d973721864bc627c40b07f53.jpg" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 364, + 400, + 554, + 414 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 364, + 400, + 554, + 414 + ], + "spans": [ + { + "bbox": [ + 364, + 400, + 554, + 414 + ], + "type": "interline_equation", + "content": "\\mathcal {L} = \\mathcal {L} _ {\\text {d e n o i s e}} + \\lambda \\cdot \\mathcal {L} _ {K M L} \\tag {7}", + "image_path": "3a5a029c85c7c33a0b42faf5cfea8dc56ddc0a1a721d3a945a6102e2df17c33c.jpg" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 313, + 421, + 554, + 506 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 421, + 554, + 506 + ], + "spans": [ + { + "bbox": [ + 313, + 421, + 554, + 506 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 313, + 421, + 554, + 506 + ], + "type": "inline_equation", + "content": "n_{kp}" + }, + { + "bbox": [ + 313, + 421, + 554, + 506 + ], + "type": "text", + "content": " denote the number of keypoint, " + }, + { + "bbox": [ + 313, + 421, + 554, + 506 + ], + "type": "inline_equation", + "content": "c_{kp}^{c_i}" + }, + { + "bbox": [ + 313, + 421, + 554, + 506 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 313, + 421, + 554, + 506 + ], + "type": "inline_equation", + "content": "c_{kp}^{x_k}" + }, + { + "bbox": [ + 313, + 421, + 554, + 506 + ], + "type": "text", + "content": " are keypoint's coordinates in " + }, + { + "bbox": [ + 313, + 421, + 554, + 506 + ], + "type": "inline_equation", + "content": "c_i" + }, + { + "bbox": [ + 313, + 421, + 554, + 506 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 313, + 421, + 554, + 506 + ], + "type": "inline_equation", + "content": "x_k" + }, + { + "bbox": [ + 313, + 421, + 554, + 506 + ], + "type": "text", + "content": " respectively. We scale the coordinates to fit the size of " + }, + { + "bbox": [ + 313, + 421, + 554, + 506 + ], + "type": "inline_equation", + "content": "\\mathcal{E}(c_i)" + }, + { + "bbox": [ + 313, + 421, + 554, + 506 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 313, + 421, + 554, + 506 + ], + "type": "inline_equation", + "content": "\\hat{z}_0 \\lambda" + }, + { + "bbox": [ + 313, + 421, + 554, + 506 + ], + "type": "text", + "content": " controls for the relative weight of KML. The part inside the dotted box in Fig. 2 illustrates the model fine-tuning with the KML. We find KML is effective in encouraging the model to manipulate the pose for relation-aware generation." + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 314, + 510, + 436, + 522 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 314, + 510, + 436, + 522 + ], + "spans": [ + { + "bbox": [ + 314, + 510, + 436, + 522 + ], + "type": "text", + "content": "3.4.3. Local Token Injection" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 313, + 525, + 555, + 657 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 525, + 555, + 657 + ], + "spans": [ + { + "bbox": [ + 313, + 525, + 555, + 657 + ], + "type": "text", + "content": "When generating relationships, local and detailed features are essential—for instance, the feature of arms to accurately construct relationships like \"carrying\". However, the coarse image-level features extracted by the CLIP Image Encoder fail to provide such fine-grained details, leading to confusion between objects in Fig. 7. To address this limitation, we introduce dense features from the CLIP Image Encoder [44]. Specifically, we modify the last layer of the CLIP Image Encoder to obtain dense features: " + }, + { + "bbox": [ + 313, + 525, + 555, + 657 + ], + "type": "inline_equation", + "content": "h_{\\mathrm{dense}} =" + }, + { + "bbox": [ + 313, + 525, + 555, + 657 + ], + "type": "text", + "content": " ModifiedAttention(" + }, + { + "bbox": [ + 313, + 525, + 555, + 657 + ], + "type": "inline_equation", + "content": "h_{\\mathrm{clip}}'" + }, + { + "bbox": [ + 313, + 525, + 555, + 657 + ], + "type": "text", + "content": "), where " + }, + { + "bbox": [ + 313, + 525, + 555, + 657 + ], + "type": "inline_equation", + "content": "h_{\\mathrm{clip}}'" + }, + { + "bbox": [ + 313, + 525, + 555, + 657 + ], + "type": "text", + "content": " represents the hidden states of the second-to-last layer." + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 376, + 665, + 555, + 678 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 376, + 665, + 555, + 678 + ], + "spans": [ + { + "bbox": [ + 376, + 665, + 555, + 678 + ], + "type": "interline_equation", + "content": "h _ {\\mathrm {t m p}} = \\operatorname {P r o j} _ {v} \\left(\\operatorname {n o r m} \\left(h _ {\\mathrm {c l i p}} ^ {\\prime}\\right)\\right), \\tag {8}", + "image_path": "df19198b818a9c31f055f713297e337952f885919788e957184398c7de2e01a0.jpg" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 377, + 681, + 555, + 695 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 377, + 681, + 555, + 695 + ], + "spans": [ + { + "bbox": [ + 377, + 681, + 555, + 695 + ], + "type": "interline_equation", + "content": "h _ {\\mathrm {t m p}} = h _ {\\mathrm {c l i p}} ^ {\\prime} + \\operatorname {P r o j} _ {\\text {o u t}} \\left(h _ {\\mathrm {t m p}}\\right), \\tag {9}", + "image_path": "1eea3bc3dffb021ddd65ccf6d1a8a0c849acfdd54b3cbe8511085f55edbf8de6.jpg" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 372, + 697, + 554, + 711 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 372, + 697, + 554, + 711 + ], + "spans": [ + { + "bbox": [ + 372, + 697, + 554, + 711 + ], + "type": "interline_equation", + "content": "h _ {\\text {d e n s e}} = h _ {\\text {t m p}} + \\operatorname {F F N} \\left(h _ {\\text {t m p}}\\right) \\tag {10}", + "image_path": "1034e52318b18b6afe6758dc0b105bff42c9096f666a74a7a36e043e383c0e20.jpg" + } + ] + } + ], + "index": 19 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "text", + "content": "15727" + } + ] + } + ], + "index": 20 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 4 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 107, + 70, + 503, + 224 + ], + "blocks": [ + { + "bbox": [ + 107, + 70, + 503, + 224 + ], + "lines": [ + { + "bbox": [ + 107, + 70, + 503, + 224 + ], + "spans": [ + { + "bbox": [ + 107, + 70, + 503, + 224 + ], + "type": "image", + "image_path": "809ee864e0f3795e33b31f9468f23149cc2ed865d2a79d8a8d671053d613027d.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 55, + 229, + 555, + 251 + ], + "lines": [ + { + "bbox": [ + 55, + 229, + 555, + 251 + ], + "spans": [ + { + "bbox": [ + 55, + 229, + 555, + 251 + ], + "type": "text", + "content": "Figure 5. Additional results demonstrate the effectiveness of relation-aware generation. DreamRelation adapts the same pair of objects to different relationships in a natural and accurate manner." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "bbox": [ + 55, + 262, + 295, + 334 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 262, + 295, + 334 + ], + "spans": [ + { + "bbox": [ + 55, + 262, + 295, + 334 + ], + "type": "text", + "content": "To further enhance the compatibility between dense features and image-level features, we employ self-distillation on the CLIP Image Encoder [39] to derive our Local Image Encoder. Specifically, we align local regions of the dense features with their corresponding image-level features using cosine similarity." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 55, + 335, + 296, + 407 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 335, + 296, + 407 + ], + "spans": [ + { + "bbox": [ + 55, + 335, + 296, + 407 + ], + "type": "text", + "content": "During inference, we extract local tokens " + }, + { + "bbox": [ + 55, + 335, + 296, + 407 + ], + "type": "inline_equation", + "content": "tok_{\\mathrm{local}}" + }, + { + "bbox": [ + 55, + 335, + 296, + 407 + ], + "type": "text", + "content": " from the dense feature " + }, + { + "bbox": [ + 55, + 335, + 296, + 407 + ], + "type": "inline_equation", + "content": "h_{\\mathrm{dense}}" + }, + { + "bbox": [ + 55, + 335, + 296, + 407 + ], + "type": "text", + "content": " through partitioning and pooling. These local tokens are then concatenated with the image-level tokens " + }, + { + "bbox": [ + 55, + 335, + 296, + 407 + ], + "type": "inline_equation", + "content": "tok_{\\mathrm{image}}" + }, + { + "bbox": [ + 55, + 335, + 296, + 407 + ], + "type": "text", + "content": " and passed to the ID extractor. The ID extractor, which is a transformer-based architecture as illustrated in Fig. 2 (b), can be written as:" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 70, + 417, + 295, + 432 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 70, + 417, + 295, + 432 + ], + "spans": [ + { + "bbox": [ + 70, + 417, + 295, + 432 + ], + "type": "interline_equation", + "content": "q = q + \\text {A t t e n t i o n} (\\text {c o n c a t} [ q, \\text {t o k} _ {\\text {i m a g e}}, \\text {t o k} _ {\\text {l o c a l}} ]) \\tag {11}", + "image_path": "83c68338212db7845961b8fac41a44ba19744ff57ecdd8d2c29c72b8b72a607d.jpg" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 55, + 441, + 296, + 537 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 441, + 296, + 537 + ], + "spans": [ + { + "bbox": [ + 55, + 441, + 296, + 537 + ], + "type": "text", + "content": "Regarding the compatibility between CLIP dense features and the original image-level features, experimental results demonstrate that our Local Image Encoder outperforms other dense representations [3]. For a detailed discussion, we refer readers to the Supplementary Material. Additionally, the injection of local tokens enhances identity preservation, as evidenced by improvements in evaluation metrics and visual results, shown in Fig. 7 and Tab. 3." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 55, + 549, + 137, + 562 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 549, + 137, + 562 + ], + "spans": [ + { + "bbox": [ + 55, + 549, + 137, + 562 + ], + "type": "text", + "content": "4. Experiments" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 55, + 570, + 296, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 570, + 296, + 713 + ], + "spans": [ + { + "bbox": [ + 55, + 570, + 296, + 713 + ], + "type": "text", + "content": "Implementation Details. We generate a tuning set for relation learning with 4-6 samples per relationship. We incorporate LoRA layers into all text cross-attention layers of the U-Net. We set the LoRA rank to " + }, + { + "bbox": [ + 55, + 570, + 296, + 713 + ], + "type": "inline_equation", + "content": "r = 4" + }, + { + "bbox": [ + 55, + 570, + 296, + 713 + ], + "type": "text", + "content": ", the parallel cross-attention scaling factor to " + }, + { + "bbox": [ + 55, + 570, + 296, + 713 + ], + "type": "inline_equation", + "content": "\\gamma = 0.6" + }, + { + "bbox": [ + 55, + 570, + 296, + 713 + ], + "type": "text", + "content": ", and the keypoint matching loss weight to " + }, + { + "bbox": [ + 55, + 570, + 296, + 713 + ], + "type": "inline_equation", + "content": "\\lambda = 1e - 3" + }, + { + "bbox": [ + 55, + 570, + 296, + 713 + ], + "type": "text", + "content": ". The model is finetuned for 500 steps, using 2 A100 GPUs, with a total batch size of 8, completing the process in 10 minutes. We use the Adam optimizer [12] with a learning rate of " + }, + { + "bbox": [ + 55, + 570, + 296, + 713 + ], + "type": "inline_equation", + "content": "1e - 4" + }, + { + "bbox": [ + 55, + 570, + 296, + 713 + ], + "type": "text", + "content": " and no weight decay, resulting in a total of 3.1M trainable parameters. Note that our learnable parameters are only in the text cross-attention layers, making it compatible with" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 313, + 262, + 555, + 334 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 262, + 555, + 334 + ], + "spans": [ + { + "bbox": [ + 313, + 262, + 555, + 334 + ], + "type": "text", + "content": "any SDXL-based models. Considering the strong identity preservation capabilities of MS-Diffusion, we implement our DreamRelation on MS-Diffusion. During fine-tuning and inference, we concatenate local tokens with image tokens. We put more self-distillation details of the Local Image Encoder in the Supplementary Material." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 313, + 335, + 556, + 575 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 335, + 556, + 575 + ], + "spans": [ + { + "bbox": [ + 313, + 335, + 556, + 575 + ], + "type": "text", + "content": "Evaluation. To evaluate relation-aware customized image generation, we propose RelationBench, consisting of 44 objects from DreamBench [31] and CustomConcept101 [15], along with 25 relationships. The object categories include pets, plushies, toys, people, and cartoons. Using GPT-4, we generate 100 cases for single-object and multi-object evaluations. Following previous works [31, 35], we evaluate our method on three metrics: (1) Identity Preservation, which assesses the alignment between the generated images and the image prompts, using the CLIP image score and DINO score that calculate the cosine similarity between the class embeddings of images, referred to as CLIP-I and DINO, respectively; (2) Text Alignment, which evaluates how well the generated images align with the text prompts, using the CLIP image-text score, denoted as CLIP-T; and (3) Relation Alignment. We observed that nouns in the prompts can inflate CLIP-T scores, making them less accurate for evaluating relation generation. To address this, we extract the relation word from the text prompt using spaCy and calculate the CLIP image-text score, denoted as CLIP-R." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 313, + 586, + 400, + 597 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 586, + 400, + 597 + ], + "spans": [ + { + "bbox": [ + 313, + 586, + 400, + 597 + ], + "type": "text", + "content": "4.1. Main Results" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 313, + 604, + 555, + 700 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 604, + 555, + 700 + ], + "spans": [ + { + "bbox": [ + 313, + 604, + 555, + 700 + ], + "type": "text", + "content": "We evaluate our method through both quantitative and qualitative results in single- and multi-object cases. In single-object cases, we primarily focus on the alignment of the object's pose with the relationship in the text prompt and the model's ability to preserve identity. In multi-object cases, in addition to relation generation and identity preservation, we also need to ensure no confusion between the multiple objects." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 313, + 701, + 555, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 701, + 555, + 713 + ], + "spans": [ + { + "bbox": [ + 313, + 701, + 555, + 713 + ], + "type": "text", + "content": "Quantitative results. We compare our method with" + } + ] + } + ], + "index": 12 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "text", + "content": "15728" + } + ] + } + ], + "index": 13 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 5 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 59, + 83, + 88, + 139 + ], + "blocks": [ + { + "bbox": [ + 59, + 72, + 86, + 82 + ], + "lines": [ + { + "bbox": [ + 59, + 72, + 86, + 82 + ], + "spans": [ + { + "bbox": [ + 59, + 72, + 86, + 82 + ], + "type": "text", + "content": "Objects" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 59, + 83, + 88, + 139 + ], + "lines": [ + { + "bbox": [ + 59, + 83, + 88, + 139 + ], + "spans": [ + { + "bbox": [ + 59, + 83, + 88, + 139 + ], + "type": "image", + "image_path": "379cf19086e316ab79992f9b391916aedabc0f7ca89b83b404a29b1f4b25a375.jpg" + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 97, + 98, + 132, + 124 + ], + "lines": [ + { + "bbox": [ + 97, + 98, + 132, + 124 + ], + "spans": [ + { + "bbox": [ + 97, + 98, + 132, + 124 + ], + "type": "text", + "content": "A cat is kissing with a dog." + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 95, + 152, + 135, + 186 + ], + "lines": [ + { + "bbox": [ + 95, + 152, + 135, + 186 + ], + "spans": [ + { + "bbox": [ + 95, + 152, + 135, + 186 + ], + "type": "text", + "content": "A dog is shaking hands with a plushie bear." + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 95, + 209, + 134, + 243 + ], + "lines": [ + { + "bbox": [ + 95, + 209, + 134, + 243 + ], + "spans": [ + { + "bbox": [ + 95, + 209, + 134, + 243 + ], + "type": "text", + "content": "A dog is fencing with a plushie panda." + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "image_caption" + } + ], + "index": 1 + }, + { + "type": "image", + "bbox": [ + 141, + 83, + 197, + 138 + ], + "blocks": [ + { + "bbox": [ + 100, + 72, + 128, + 81 + ], + "lines": [ + { + "bbox": [ + 100, + 72, + 128, + 81 + ], + "spans": [ + { + "bbox": [ + 100, + 72, + 128, + 81 + ], + "type": "text", + "content": "Prompts" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 162, + 72, + 177, + 80 + ], + "lines": [ + { + "bbox": [ + 162, + 72, + 177, + 80 + ], + "spans": [ + { + "bbox": [ + 162, + 72, + 177, + 80 + ], + "type": "text", + "content": "Ours" + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 141, + 83, + 197, + 138 + ], + "lines": [ + { + "bbox": [ + 141, + 83, + 197, + 138 + ], + "spans": [ + { + "bbox": [ + 141, + 83, + 197, + 138 + ], + "type": "image", + "image_path": "e102369285e21b5cc8ae72b36faee5b2f7b6ac7514d0d985b4b00b054557487d.jpg" + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "image_body" + } + ], + "index": 7 + }, + { + "type": "image", + "bbox": [ + 200, + 83, + 255, + 137 + ], + "blocks": [ + { + "bbox": [ + 203, + 72, + 252, + 80 + ], + "lines": [ + { + "bbox": [ + 203, + 72, + 252, + 80 + ], + "spans": [ + { + "bbox": [ + 203, + 72, + 252, + 80 + ], + "type": "text", + "content": "ReVersion+MS" + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 200, + 83, + 255, + 137 + ], + "lines": [ + { + "bbox": [ + 200, + 83, + 255, + 137 + ], + "spans": [ + { + "bbox": [ + 200, + 83, + 255, + 137 + ], + "type": "image", + "image_path": "e10a75568002dc11826e587a9e8eac413e349b8c993468977332e3f7949ecef8.jpg" + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "image_body" + } + ], + "index": 9 + }, + { + "type": "image", + "bbox": [ + 258, + 83, + 312, + 137 + ], + "blocks": [ + { + "bbox": [ + 262, + 72, + 306, + 80 + ], + "lines": [ + { + "bbox": [ + 262, + 72, + 306, + 80 + ], + "spans": [ + { + "bbox": [ + 262, + 72, + 306, + 80 + ], + "type": "text", + "content": "MS-Diffusion" + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 258, + 83, + 312, + 137 + ], + "lines": [ + { + "bbox": [ + 258, + 83, + 312, + 137 + ], + "spans": [ + { + "bbox": [ + 258, + 83, + 312, + 137 + ], + "type": "image", + "image_path": "8d19a2a803bcd79ee9e96e9cb83dae13b0b3883e0cf39d8b9e7d01008fc1be9a.jpg" + } + ] + } + ], + "index": 11, + "angle": 0, + "type": "image_body" + } + ], + "index": 11 + }, + { + "type": "image", + "bbox": [ + 315, + 83, + 370, + 137 + ], + "blocks": [ + { + "bbox": [ + 321, + 72, + 364, + 80 + ], + "lines": [ + { + "bbox": [ + 321, + 72, + 364, + 80 + ], + "spans": [ + { + "bbox": [ + 321, + 72, + 364, + 80 + ], + "type": "text", + "content": "SSR-Encoder" + } + ] + } + ], + "index": 12, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 315, + 83, + 370, + 137 + ], + "lines": [ + { + "bbox": [ + 315, + 83, + 370, + 137 + ], + "spans": [ + { + "bbox": [ + 315, + 83, + 370, + 137 + ], + "type": "image", + "image_path": "b2589730f24b1c2f3f12031d4b1773e0ab117f2a7f7b744cc00913ab8f3ae562.jpg" + } + ] + } + ], + "index": 13, + "angle": 0, + "type": "image_body" + } + ], + "index": 13 + }, + { + "type": "image", + "bbox": [ + 373, + 83, + 428, + 137 + ], + "blocks": [ + { + "bbox": [ + 387, + 72, + 415, + 81 + ], + "lines": [ + { + "bbox": [ + 387, + 72, + 415, + 81 + ], + "spans": [ + { + "bbox": [ + 387, + 72, + 415, + 81 + ], + "type": "inline_equation", + "content": "\\lambda" + }, + { + "bbox": [ + 387, + 72, + 415, + 81 + ], + "type": "text", + "content": " -eclipse" + } + ] + } + ], + "index": 14, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 373, + 83, + 428, + 137 + ], + "lines": [ + { + "bbox": [ + 373, + 83, + 428, + 137 + ], + "spans": [ + { + "bbox": [ + 373, + 83, + 428, + 137 + ], + "type": "image", + "image_path": "8bf4caad1cc5881f70b03d4d991c0d7957631c5a62c953cf9b34fa1c2e3dc06d.jpg" + } + ] + } + ], + "index": 15, + "angle": 0, + "type": "image_body" + } + ], + "index": 15 + }, + { + "type": "image", + "bbox": [ + 373, + 140, + 428, + 195 + ], + "blocks": [ + { + "bbox": [ + 373, + 140, + 428, + 195 + ], + "lines": [ + { + "bbox": [ + 373, + 140, + 428, + 195 + ], + "spans": [ + { + "bbox": [ + 373, + 140, + 428, + 195 + ], + "type": "image", + "image_path": "988c9e1efedd07f5f5656173d3d0ebd0e3fb88b7835f4af9f27664ad6a1e4ff5.jpg" + } + ] + } + ], + "index": 17, + "angle": 0, + "type": "image_body" + } + ], + "index": 17 + }, + { + "type": "image", + "bbox": [ + 430, + 83, + 485, + 137 + ], + "blocks": [ + { + "bbox": [ + 429, + 72, + 552, + 80 + ], + "lines": [ + { + "bbox": [ + 429, + 72, + 552, + 80 + ], + "spans": [ + { + "bbox": [ + 429, + 72, + 552, + 80 + ], + "type": "text", + "content": "Custom Diffusion DreamBooth (SDXL)" + } + ] + } + ], + "index": 16, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 430, + 83, + 485, + 137 + ], + "lines": [ + { + "bbox": [ + 430, + 83, + 485, + 137 + ], + "spans": [ + { + "bbox": [ + 430, + 83, + 485, + 137 + ], + "type": "image", + "image_path": "ede574f721266b94a4be719cd335833cd807c661ecf5343115e7796ac2214c53.jpg" + } + ] + } + ], + "index": 18, + "angle": 0, + "type": "image_body" + } + ], + "index": 18 + }, + { + "type": "image", + "bbox": [ + 430, + 140, + 485, + 195 + ], + "blocks": [ + { + "bbox": [ + 430, + 140, + 485, + 195 + ], + "lines": [ + { + "bbox": [ + 430, + 140, + 485, + 195 + ], + "spans": [ + { + "bbox": [ + 430, + 140, + 485, + 195 + ], + "type": "image", + "image_path": "1d6107a00e1ba9db511ff8589de8ebd015dc5b79f7e44311a182427c896f2645.jpg" + } + ] + } + ], + "index": 19, + "angle": 0, + "type": "image_body" + } + ], + "index": 19 + }, + { + "type": "image", + "bbox": [ + 488, + 83, + 542, + 137 + ], + "blocks": [ + { + "bbox": [ + 488, + 83, + 542, + 137 + ], + "lines": [ + { + "bbox": [ + 488, + 83, + 542, + 137 + ], + "spans": [ + { + "bbox": [ + 488, + 83, + 542, + 137 + ], + "type": "image", + "image_path": "95eec3e4856dc4ea959accb9c0817c96907d14e9dd7b83c270aec1448d66b0f1.jpg" + } + ] + } + ], + "index": 20, + "angle": 0, + "type": "image_body" + } + ], + "index": 20 + }, + { + "type": "image", + "bbox": [ + 430, + 198, + 485, + 253 + ], + "blocks": [ + { + "bbox": [ + 430, + 198, + 485, + 253 + ], + "lines": [ + { + "bbox": [ + 430, + 198, + 485, + 253 + ], + "spans": [ + { + "bbox": [ + 430, + 198, + 485, + 253 + ], + "type": "image", + "image_path": "d2a93e7beb1b95cff6a234f46d3fc23c03ce3861b22648b1aa57bfa2a26af9b3.jpg" + } + ] + } + ], + "index": 21, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 55, + 261, + 555, + 294 + ], + "lines": [ + { + "bbox": [ + 55, + 261, + 555, + 294 + ], + "spans": [ + { + "bbox": [ + 55, + 261, + 555, + 294 + ], + "type": "text", + "content": "Figure 6. We compare multi-object performance with training-based and tuning-based methods. Additionally, we integrate ReVersion [11] into MS-Diffusion, where the learnable text token is injected into the text prompt during MS-Diffusion's inference, resulting in ReVersion+MS. Our method shows a clear advantage in relation-aware generation and avoiding object confusion in overlapping scenarios." + } + ] + } + ], + "index": 24, + "angle": 0, + "type": "image_caption" + } + ], + "index": 21 + }, + { + "type": "image", + "bbox": [ + 488, + 140, + 542, + 195 + ], + "blocks": [ + { + "bbox": [ + 488, + 140, + 542, + 195 + ], + "lines": [ + { + "bbox": [ + 488, + 140, + 542, + 195 + ], + "spans": [ + { + "bbox": [ + 488, + 140, + 542, + 195 + ], + "type": "image", + "image_path": "9e418c165bf878515dc4e89a4e5c8c317ba989b5ffebd56d776fa23f1ad0a435.jpg" + } + ] + } + ], + "index": 22, + "angle": 0, + "type": "image_body" + } + ], + "index": 22 + }, + { + "type": "image", + "bbox": [ + 488, + 198, + 542, + 253 + ], + "blocks": [ + { + "bbox": [ + 488, + 198, + 542, + 253 + ], + "lines": [ + { + "bbox": [ + 488, + 198, + 542, + 253 + ], + "spans": [ + { + "bbox": [ + 488, + 198, + 542, + 253 + ], + "type": "image", + "image_path": "82bc6c3f3d392bd90ffb7eaa46d951666f771eb021df5195fc4f4f89b0aa459f.jpg" + } + ] + } + ], + "index": 23, + "angle": 0, + "type": "image_body" + } + ], + "index": 23 + }, + { + "type": "table", + "bbox": [ + 79, + 300, + 526, + 453 + ], + "blocks": [ + { + "bbox": [ + 79, + 300, + 526, + 453 + ], + "lines": [ + { + "bbox": [ + 79, + 300, + 526, + 453 + ], + "spans": [ + { + "bbox": [ + 79, + 300, + 526, + 453 + ], + "type": "table", + "html": "
MethodSingle-objectMulti-object
CLIP-TCLIP-RCLIP-IDINOCLIP-TCLIP-RCLIP-IDINO
DreamBooth (SDXL) [31]27.818.274.262.624.316.267.857.2
Custom Diffusion [15]26.515.273.258.820.115.464.755.3
Cones-V2 [20]24.413.572.157.221.315.264.354.2
ELITE [36]25.714.975.461.5————
AnyDoor [5]24.514.777.462.221.614.969.759.8
BLIP-Diffusion [16]26.215.777.457.7————
SSR-Encoder [43]25.515.980.459.424.214.672.156.2
MS-Diffusion [35]26.518.878.764.526.918.973.858.8
Reversion+MS [11]27.819.377.863.127.219.274.259.7
Ours30.621.477.963.428.920.475.462.1
", + "image_path": "c4d76e37d26021ad0d3e4ea128069e1f75f10b1dd61b42c522c80369a3ffb3fc.jpg" + } + ] + } + ], + "index": 25, + "angle": 0, + "type": "table_body" + } + ], + "index": 25 + }, + { + "type": "table", + "bbox": [ + 57, + 477, + 290, + 541 + ], + "blocks": [ + { + "bbox": [ + 83, + 456, + 525, + 468 + ], + "lines": [ + { + "bbox": [ + 83, + 456, + 525, + 468 + ], + "spans": [ + { + "bbox": [ + 83, + 456, + 525, + 468 + ], + "type": "text", + "content": "Table 2. Quantitative comparison on RelationBench, Bold and underline represent the highest and second-highest metrics." + } + ] + } + ], + "index": 26, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 57, + 477, + 290, + 541 + ], + "lines": [ + { + "bbox": [ + 57, + 477, + 290, + 541 + ], + "spans": [ + { + "bbox": [ + 57, + 477, + 290, + 541 + ], + "type": "table", + "html": "
MethodMulti-object
CLIP-TCLIP-RCLIP-IDINO
w/o Relation-aware Data27.319.475.359.8
w/o Local Token Injection28.519.575.159.9
w/o Keypoint Matching Loss27.419.275.261.2
Full Model28.920.475.462.1
", + "image_path": "529124363985afe02154831bb0db72c73d48d300b31c18a41d74ba80825e04b6.jpg" + } + ] + } + ], + "index": 27, + "angle": 0, + "type": "table_body" + } + ], + "index": 27 + }, + { + "bbox": [ + 78, + 544, + 272, + 555 + ], + "lines": [ + { + "bbox": [ + 78, + 544, + 272, + 555 + ], + "spans": [ + { + "bbox": [ + 78, + 544, + 272, + 555 + ], + "type": "text", + "content": "Table 3. Ablation study on our proposed components" + } + ] + } + ], + "index": 28, + "angle": 0, + "type": "text" + }, + { + "bbox": [ + 55, + 570, + 296, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 570, + 296, + 713 + ], + "spans": [ + { + "bbox": [ + 55, + 570, + 296, + 713 + ], + "type": "text", + "content": "baseline models across three benchmarks: Relation-Bench, DreamBench, and multi-object cases in Custom-Concept101. As shown in Tab. 2, our method demonstrates a clear advantage on RelationBench for single-object cases, particularly in the CLIP-T and CLIP-R metrics. However, its performance is slightly lower in the CLIP-I metric, likely due to significant pose variations in the objects, which may decrease the CLIP-I score. For multi-object generation, our method outperforms other approaches across all evaluation metrics, which we attribute to its ability to effectively avoid object confusion and significantly enhance CLIP-I performance. Furthermore, as shown in Tab. 4, we evaluate" + } + ] + } + ], + "index": 29 + }, + { + "bbox": [ + 313, + 479, + 555, + 574 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 479, + 555, + 574 + ], + "spans": [ + { + "bbox": [ + 313, + 479, + 555, + 574 + ], + "type": "text", + "content": "single-object generation performance on DreamBench. Our method outperforms others in the CLIP-T metric while delivering competitive results in CLIP-I and DINO. We conduct experiments on multi-object cases from CustomConcept101, which we denote as M-CustomConcept101. As shown in Fig. 5, our model achieves the highest performance in the DINO score and ranks second in both the CLIP-T and CLIP-I metrics." + } + ] + } + ], + "index": 30 + }, + { + "bbox": [ + 313, + 576, + 555, + 659 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 576, + 555, + 659 + ], + "spans": [ + { + "bbox": [ + 313, + 576, + 555, + 659 + ], + "type": "text", + "content": "Qualitative Comparison. We conduct qualitative comparison experiments on RelationBench to evaluate relation-aware customized image generation. As shown in Fig. 6, our method outperforms others in generating multi-object relationships, effectively avoiding object confusion while accurately capturing the intended relationships. Additional results can be found in the SM." + } + ] + } + ], + "index": 31 + }, + { + "bbox": [ + 313, + 670, + 468, + 684 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 670, + 468, + 684 + ], + "spans": [ + { + "bbox": [ + 313, + 670, + 468, + 684 + ], + "type": "text", + "content": "4.2. Ablation study and Analysis" + } + ] + } + ], + "index": 32 + }, + { + "bbox": [ + 313, + 689, + 554, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 689, + 554, + 714 + ], + "spans": [ + { + "bbox": [ + 313, + 689, + 554, + 714 + ], + "type": "text", + "content": "Relation-Aware Data Engine. Instead of fine-tuning on our curated " + }, + { + "bbox": [ + 313, + 689, + 554, + 714 + ], + "type": "inline_equation", + "content": "(x_{k},c_{i},c_{t})" + }, + { + "bbox": [ + 313, + 689, + 554, + 714 + ], + "type": "text", + "content": " pairs. We train a text embedding " + }, + { + "bbox": [ + 313, + 689, + 554, + 714 + ], + "type": "inline_equation", + "content": "R^{*}" + } + ] + } + ], + "index": 33 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "text", + "content": "15729" + } + ] + } + ], + "index": 34 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 6 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 66, + 83, + 110, + 171 + ], + "blocks": [ + { + "bbox": [ + 62, + 70, + 102, + 81 + ], + "lines": [ + { + "bbox": [ + 62, + 70, + 102, + 81 + ], + "spans": [ + { + "bbox": [ + 62, + 70, + 102, + 81 + ], + "type": "text", + "content": "a) objects" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 66, + 83, + 110, + 171 + ], + "lines": [ + { + "bbox": [ + 66, + 83, + 110, + 171 + ], + "spans": [ + { + "bbox": [ + 66, + 83, + 110, + 171 + ], + "type": "image", + "image_path": "c32e4c51e7a31b10383fb3a9b685ecc9b41537c37ccecf9bf7fb6145c57a1a43.jpg" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_body" + } + ], + "index": 3 + }, + { + "type": "image", + "bbox": [ + 110, + 83, + 197, + 171 + ], + "blocks": [ + { + "bbox": [ + 138, + 70, + 169, + 79 + ], + "lines": [ + { + "bbox": [ + 138, + 70, + 169, + 79 + ], + "spans": [ + { + "bbox": [ + 138, + 70, + 169, + 79 + ], + "type": "text", + "content": "wKML" + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 110, + 83, + 197, + 171 + ], + "lines": [ + { + "bbox": [ + 110, + 83, + 197, + 171 + ], + "spans": [ + { + "bbox": [ + 110, + 83, + 197, + 171 + ], + "type": "image", + "image_path": "0cc68667bf281a2204618555ea89da1f93e41bba7b0e13375c76845b552ad4cc.jpg" + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 113, + 172, + 238, + 182 + ], + "lines": [ + { + "bbox": [ + 113, + 172, + 238, + 182 + ], + "spans": [ + { + "bbox": [ + 113, + 172, + 238, + 182 + ], + "type": "text", + "content": "A cat is shaking hands with a dog." + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "image_caption" + } + ], + "index": 4 + }, + { + "type": "image", + "bbox": [ + 197, + 83, + 285, + 171 + ], + "blocks": [ + { + "bbox": [ + 225, + 70, + 261, + 79 + ], + "lines": [ + { + "bbox": [ + 225, + 70, + 261, + 79 + ], + "spans": [ + { + "bbox": [ + 225, + 70, + 261, + 79 + ], + "type": "text", + "content": "w/o KML" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 197, + 83, + 285, + 171 + ], + "lines": [ + { + "bbox": [ + 197, + 83, + 285, + 171 + ], + "spans": [ + { + "bbox": [ + 197, + 83, + 285, + 171 + ], + "type": "image", + "image_path": "2cc0ebf8b7f7e3aea1fbb1cf771a434fd3853238e211b5d69d00071345cc47d3.jpg" + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "image_body" + } + ], + "index": 5 + }, + { + "type": "image", + "bbox": [ + 66, + 199, + 111, + 244 + ], + "blocks": [ + { + "bbox": [ + 62, + 183, + 102, + 196 + ], + "lines": [ + { + "bbox": [ + 62, + 183, + 102, + 196 + ], + "spans": [ + { + "bbox": [ + 62, + 183, + 102, + 196 + ], + "type": "text", + "content": "b) objects" + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 66, + 199, + 111, + 244 + ], + "lines": [ + { + "bbox": [ + 66, + 199, + 111, + 244 + ], + "spans": [ + { + "bbox": [ + 66, + 199, + 111, + 244 + ], + "type": "image", + "image_path": "8436a73e636a2148e8b7a4b22337942cefc829a9361b77e0a299c9032012c955.jpg" + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "image_body" + } + ], + "index": 10 + }, + { + "type": "image", + "bbox": [ + 111, + 199, + 198, + 287 + ], + "blocks": [ + { + "bbox": [ + 132, + 186, + 186, + 196 + ], + "lines": [ + { + "bbox": [ + 132, + 186, + 186, + 196 + ], + "spans": [ + { + "bbox": [ + 132, + 186, + 186, + 196 + ], + "type": "text", + "content": "w local tokens" + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 111, + 199, + 198, + 287 + ], + "lines": [ + { + "bbox": [ + 111, + 199, + 198, + 287 + ], + "spans": [ + { + "bbox": [ + 111, + 199, + 198, + 287 + ], + "type": "image", + "image_path": "585e297493747fd77bf7493be8bdea864588b9d43a1d8f9f364a84618a821200.jpg" + } + ] + } + ], + "index": 11, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 132, + 289, + 223, + 299 + ], + "lines": [ + { + "bbox": [ + 132, + 289, + 223, + 299 + ], + "spans": [ + { + "bbox": [ + 132, + 289, + 223, + 299 + ], + "type": "text", + "content": "A dog is carrying a sloth." + } + ] + } + ], + "index": 13, + "angle": 0, + "type": "image_caption" + } + ], + "index": 11 + }, + { + "type": "image", + "bbox": [ + 198, + 199, + 286, + 287 + ], + "blocks": [ + { + "bbox": [ + 212, + 186, + 272, + 195 + ], + "lines": [ + { + "bbox": [ + 212, + 186, + 272, + 195 + ], + "spans": [ + { + "bbox": [ + 212, + 186, + 272, + 195 + ], + "type": "text", + "content": "w/o local tokens" + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 198, + 199, + 286, + 287 + ], + "lines": [ + { + "bbox": [ + 198, + 199, + 286, + 287 + ], + "spans": [ + { + "bbox": [ + 198, + 199, + 286, + 287 + ], + "type": "image", + "image_path": "48d030056b3ff97702524a71b39ca2139dcc3c05dac93881cabd07bad6aaff44.jpg" + } + ] + } + ], + "index": 12, + "angle": 0, + "type": "image_body" + } + ], + "index": 12 + }, + { + "type": "image", + "bbox": [ + 63, + 338, + 80, + 392 + ], + "blocks": [ + { + "bbox": [ + 55, + 303, + 295, + 335 + ], + "lines": [ + { + "bbox": [ + 55, + 303, + 295, + 335 + ], + "spans": [ + { + "bbox": [ + 55, + 303, + 295, + 335 + ], + "type": "text", + "content": "Figure 7. Ablation study of keypoint matching loss (KML) and local tokens: The images are generated using the same random seed. We use green and red boxes to highlight the main differences." + } + ] + } + ], + "index": 14, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 63, + 338, + 80, + 392 + ], + "lines": [ + { + "bbox": [ + 63, + 338, + 80, + 392 + ], + "spans": [ + { + "bbox": [ + 63, + 338, + 80, + 392 + ], + "type": "image", + "image_path": "41e954c311fecc6c76611c1ed057e5c60c72fa2762f1539244342bd52da504b9.jpg" + } + ] + } + ], + "index": 15, + "angle": 0, + "type": "image_body" + } + ], + "index": 15 + }, + { + "type": "image", + "bbox": [ + 80, + 339, + 112, + 393 + ], + "blocks": [ + { + "bbox": [ + 80, + 339, + 112, + 393 + ], + "lines": [ + { + "bbox": [ + 80, + 339, + 112, + 393 + ], + "spans": [ + { + "bbox": [ + 80, + 339, + 112, + 393 + ], + "type": "image", + "image_path": "bbab11eb667e11037d0aaa1da0d501ce47968c0ae8263f930f8f5c0d24132ae1.jpg" + } + ] + } + ], + "index": 16, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 80, + 395, + 168, + 399 + ], + "lines": [ + { + "bbox": [ + 80, + 395, + 168, + 399 + ], + "spans": [ + { + "bbox": [ + 80, + 395, + 168, + 399 + ], + "type": "text", + "content": "A dog, a cat, a plushie bear hug with each other." + } + ] + } + ], + "index": 21, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 55, + 405, + 295, + 427 + ], + "lines": [ + { + "bbox": [ + 55, + 405, + 295, + 427 + ], + "spans": [ + { + "bbox": [ + 55, + 405, + 295, + 427 + ], + "type": "text", + "content": "Figure 8. DreamRelation for three objects. Compared to the baseline model, DreamRelation shows a clear advantage." + } + ] + } + ], + "index": 23, + "angle": 0, + "type": "image_caption" + } + ], + "index": 16 + }, + { + "type": "image", + "bbox": [ + 112, + 341, + 129, + 393 + ], + "blocks": [ + { + "bbox": [ + 112, + 341, + 129, + 393 + ], + "lines": [ + { + "bbox": [ + 112, + 341, + 129, + 393 + ], + "spans": [ + { + "bbox": [ + 112, + 341, + 129, + 393 + ], + "type": "image", + "image_path": "f413434c347e23d2334146db8ee38434698549a26d84273dae987ff814449779.jpg" + } + ] + } + ], + "index": 17, + "angle": 0, + "type": "image_body" + } + ], + "index": 17 + }, + { + "type": "image", + "bbox": [ + 129, + 341, + 181, + 393 + ], + "blocks": [ + { + "bbox": [ + 129, + 341, + 181, + 393 + ], + "lines": [ + { + "bbox": [ + 129, + 341, + 181, + 393 + ], + "spans": [ + { + "bbox": [ + 129, + 341, + 181, + 393 + ], + "type": "image", + "image_path": "bdeac0a670eabb1140cb1c1fb6c118c3778e4b7f077aba905985c69174d782f6.jpg" + } + ] + } + ], + "index": 18, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 174, + 395, + 287, + 400 + ], + "lines": [ + { + "bbox": [ + 174, + 395, + 287, + 400 + ], + "spans": [ + { + "bbox": [ + 174, + 395, + 287, + 400 + ], + "type": "text", + "content": "A dog, a cat, and a plushie bear are fighting with one another." + } + ] + } + ], + "index": 22, + "angle": 0, + "type": "image_caption" + } + ], + "index": 18 + }, + { + "type": "image", + "bbox": [ + 181, + 341, + 216, + 393 + ], + "blocks": [ + { + "bbox": [ + 181, + 341, + 216, + 393 + ], + "lines": [ + { + "bbox": [ + 181, + 341, + 216, + 393 + ], + "spans": [ + { + "bbox": [ + 181, + 341, + 216, + 393 + ], + "type": "image", + "image_path": "c6946fccc9d404af3675a38a9f8de4efe96dce4e18597be7b39406edfcc50e73.jpg" + } + ] + } + ], + "index": 19, + "angle": 0, + "type": "image_body" + } + ], + "index": 19 + }, + { + "type": "image", + "bbox": [ + 216, + 341, + 280, + 393 + ], + "blocks": [ + { + "bbox": [ + 216, + 341, + 280, + 393 + ], + "lines": [ + { + "bbox": [ + 216, + 341, + 280, + 393 + ], + "spans": [ + { + "bbox": [ + 216, + 341, + 280, + 393 + ], + "type": "image", + "image_path": "0a85a82b8ac6fbae2b0a50ec754485584c551cd7d3372904df6bc833592c4e57.jpg" + } + ] + } + ], + "index": 20, + "angle": 0, + "type": "image_body" + } + ], + "index": 20 + }, + { + "bbox": [ + 54, + 437, + 295, + 498 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 54, + 437, + 295, + 498 + ], + "spans": [ + { + "bbox": [ + 54, + 437, + 295, + 498 + ], + "type": "text", + "content": "using a set of relation-specific data " + }, + { + "bbox": [ + 54, + 437, + 295, + 498 + ], + "type": "inline_equation", + "content": "(x_{k},c_{i}^{\\prime},c_{t})" + }, + { + "bbox": [ + 54, + 437, + 295, + 498 + ], + "type": "text", + "content": ", where " + }, + { + "bbox": [ + 54, + 437, + 295, + 498 + ], + "type": "inline_equation", + "content": "c_{i}^{\\prime}" + }, + { + "bbox": [ + 54, + 437, + 295, + 498 + ], + "type": "text", + "content": " is cropped from " + }, + { + "bbox": [ + 54, + 437, + 295, + 498 + ], + "type": "inline_equation", + "content": "x_{k}" + }, + { + "bbox": [ + 54, + 437, + 295, + 498 + ], + "type": "text", + "content": ". The well-trained text embedding " + }, + { + "bbox": [ + 54, + 437, + 295, + 498 + ], + "type": "inline_equation", + "content": "R^{*}" + }, + { + "bbox": [ + 54, + 437, + 295, + 498 + ], + "type": "text", + "content": " is injected into the text embedding during inference. The quantitative results in Tab. 3 show that our relation-aware data engine outperforms in both CLIP-T and CLIP-R scores." + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 54, + 498, + 295, + 593 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 54, + 498, + 295, + 593 + ], + "spans": [ + { + "bbox": [ + 54, + 498, + 295, + 593 + ], + "type": "text", + "content": "Keypoint Matching Loss (KML). We omit KML during fine-tuning. As shown in Fig. 7 a), including the KML leads to more accurate and natural relation-aware customized images using the same random seed. Quantitative results, presented in Tab. 3, show a decline across all evaluation metrics when the KML is removed. Additionally, we experimented with different values for " + }, + { + "bbox": [ + 54, + 498, + 295, + 593 + ], + "type": "inline_equation", + "content": "\\lambda" + }, + { + "bbox": [ + 54, + 498, + 295, + 593 + ], + "type": "text", + "content": " and found that the best performance occurs at " + }, + { + "bbox": [ + 54, + 498, + 295, + 593 + ], + "type": "inline_equation", + "content": "\\lambda = 1e - 3" + }, + { + "bbox": [ + 54, + 498, + 295, + 593 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 25 + }, + { + "bbox": [ + 54, + 594, + 295, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 54, + 594, + 295, + 714 + ], + "spans": [ + { + "bbox": [ + 54, + 594, + 295, + 714 + ], + "type": "text", + "content": "Local Token Injection. For local token injection, the visual comparison in Fig. 7 b) shows severe object confusion when local tokens are omitted. The quantitative results in Tab. 3 demonstrate that local tokens improve relation-aware customization. We conducted an ablation study on the number of local tokens, as presented in Tab. 6, and determined the optimal number to be 16. Additional ablation studies are shown in the SM. Combining these methods ensures that DreamRelation achieves high-quality relation-aware customized generation." + } + ] + } + ], + "index": 26 + }, + { + "type": "table", + "bbox": [ + 336, + 70, + 529, + 173 + ], + "blocks": [ + { + "bbox": [ + 336, + 70, + 529, + 173 + ], + "lines": [ + { + "bbox": [ + 336, + 70, + 529, + 173 + ], + "spans": [ + { + "bbox": [ + 336, + 70, + 529, + 173 + ], + "type": "table", + "html": "
MethodSingle-object
CLIP-TCLIP-IDINO
DreamBooth (SDXL) [31]31.281.569.2
Custom Diffusion [15]28.477.266.8
Cones-V2 [20]31.076.567.2
ELITE [36]29.877.462.5
AnyDoor [5]25.582.167.8
SSR-Encoder [43]30.882.161.2
MS-Diffusion [35]31.579.368.2
Ours31.681.768.4
", + "image_path": "238eeb2740c11e0e48e04604c379bf23c9126ca30bf212cbe022135be0102ba7.jpg" + } + ] + } + ], + "index": 27, + "angle": 0, + "type": "table_body" + } + ], + "index": 27 + }, + { + "type": "table", + "bbox": [ + 336, + 190, + 529, + 294 + ], + "blocks": [ + { + "bbox": [ + 352, + 177, + 516, + 188 + ], + "lines": [ + { + "bbox": [ + 352, + 177, + 516, + 188 + ], + "spans": [ + { + "bbox": [ + 352, + 177, + 516, + 188 + ], + "type": "text", + "content": "Table 4. Evaluation Results on DreamBench." + } + ] + } + ], + "index": 28, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 336, + 190, + 529, + 294 + ], + "lines": [ + { + "bbox": [ + 336, + 190, + 529, + 294 + ], + "spans": [ + { + "bbox": [ + 336, + 190, + 529, + 294 + ], + "type": "table", + "html": "
MethodMulti-object
CLIP-TCLIP-IDINO
DreamBooth (SDXL) [31]29.567.449.2
Custom Diffusion [15]28.166.248.8
Cones-V2 [20]29.266.547.2
AnyDoor [5]20.272.151.2
SSR-Encoder [43]30.671.152.2
λ-eclipse [24]29.268.248.2
MS-Diffusion [35]28.070.251.2
Ours29.771.452.3
", + "image_path": "0337ebcb900b60ce15fee8b5234bcd7d2c6148199ab55049ad12caf1c9f6ac16.jpg" + } + ] + } + ], + "index": 29, + "angle": 0, + "type": "table_body" + } + ], + "index": 29 + }, + { + "type": "table", + "bbox": [ + 331, + 311, + 533, + 365 + ], + "blocks": [ + { + "bbox": [ + 334, + 297, + 533, + 308 + ], + "lines": [ + { + "bbox": [ + 334, + 297, + 533, + 308 + ], + "spans": [ + { + "bbox": [ + 334, + 297, + 533, + 308 + ], + "type": "text", + "content": "Table 5. Evaluation Results on M-CustomConcept101." + } + ] + } + ], + "index": 30, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 331, + 311, + 533, + 365 + ], + "lines": [ + { + "bbox": [ + 331, + 311, + 533, + 365 + ], + "spans": [ + { + "bbox": [ + 331, + 311, + 533, + 365 + ], + "type": "table", + "html": "
Num Local TokensMulti-object
CLIP-TCLIP-RCLIP-IDINO
N=2×228.319.475.360.1
N=4×428.920.475.462.1
N=8×827.918.174.159.4
", + "image_path": "a36c412af49e3504c03b18093270e3a83379801ffa29ebdb4acb8ce996950ed2.jpg" + } + ] + } + ], + "index": 31, + "angle": 0, + "type": "table_body" + } + ], + "index": 31 + }, + { + "type": "table", + "bbox": [ + 337, + 383, + 527, + 444 + ], + "blocks": [ + { + "bbox": [ + 335, + 369, + 533, + 380 + ], + "lines": [ + { + "bbox": [ + 335, + 369, + 533, + 380 + ], + "spans": [ + { + "bbox": [ + 335, + 369, + 533, + 380 + ], + "type": "text", + "content": "Table 6. Ablation study on the number of local tokens." + } + ] + } + ], + "index": 32, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 337, + 383, + 527, + 444 + ], + "lines": [ + { + "bbox": [ + 337, + 383, + 527, + 444 + ], + "spans": [ + { + "bbox": [ + 337, + 383, + 527, + 444 + ], + "type": "table", + "html": "
LambdaMulti-object
CLIP-TCLIP-RCLIP-IDINO
λ=1e-227.519.372.659.6
λ=1e-328.920.475.462.1
λ=1e-426.818.173.960.4
", + "image_path": "d4100e646cefbf17bc65414a9c0b9eb42b9b3abcb5a85c69476e5d178cfdfc26.jpg" + } + ] + } + ], + "index": 33, + "angle": 0, + "type": "table_body" + } + ], + "index": 33 + }, + { + "bbox": [ + 314, + 447, + 552, + 459 + ], + "lines": [ + { + "bbox": [ + 314, + 447, + 552, + 459 + ], + "spans": [ + { + "bbox": [ + 314, + 447, + 552, + 459 + ], + "type": "text", + "content": "Table 7. Ablation study on " + }, + { + "bbox": [ + 314, + 447, + 552, + 459 + ], + "type": "inline_equation", + "content": "\\lambda" + }, + { + "bbox": [ + 314, + 447, + 552, + 459 + ], + "type": "text", + "content": " ,controlling relative weight of KML." + } + ] + } + ], + "index": 34, + "angle": 0, + "type": "text" + }, + { + "bbox": [ + 314, + 469, + 388, + 481 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 314, + 469, + 388, + 481 + ], + "spans": [ + { + "bbox": [ + 314, + 469, + 388, + 481 + ], + "type": "text", + "content": "5. Conclusion" + } + ] + } + ], + "index": 35 + }, + { + "bbox": [ + 313, + 489, + 555, + 669 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 489, + 555, + 669 + ], + "spans": [ + { + "bbox": [ + 313, + 489, + 555, + 669 + ], + "type": "text", + "content": "In this work, we introduce a new challenging task: relation-aware customized image generation. This task aims to generate objects that adhere to the relationship specified in the text prompt and preserve the identities of the user-provided images. To support this, we propose a data engine for generating high-quality fine-tuning data. Our method, DreamRelation, incorporates Keypoint Matching Loss and Local Token Injection, effectively capturing relation information and generating natural relations between customized objects. We also present a fair comparison with other methods on a new benchmark, RelationBench. Extensive experiments demonstrate the effectiveness of our method, highlighting its potential for applications in interactive scenario generation, relation detection, and more. Our research can inspire the direction of relation-aware generation in the community." + } + ] + } + ], + "index": 36 + }, + { + "bbox": [ + 313, + 669, + 555, + 703 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 669, + 555, + 703 + ], + "spans": [ + { + "bbox": [ + 313, + 669, + 555, + 703 + ], + "type": "text", + "content": "Acknowledgement. This work is supported by the National Key Research and Development Program of China (No. 2023YFC3807600)." + } + ] + } + ], + "index": 37 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "type": "text", + "content": "15730" + } + ] + } + ], + "index": 38 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 7 + }, + { + "para_blocks": [ + { + "bbox": [ + 56, + 71, + 115, + 83 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 71, + 115, + 83 + ], + "spans": [ + { + "bbox": [ + 56, + 71, + 115, + 83 + ], + "type": "text", + "content": "References" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 57, + 91, + 296, + 715 + ], + "type": "list", + "angle": 0, + "index": 15, + "blocks": [ + { + "bbox": [ + 61, + 91, + 296, + 145 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 61, + 91, + 296, + 145 + ], + "spans": [ + { + "bbox": [ + 61, + 91, + 296, + 145 + ], + "type": "text", + "content": "[1] Jinbin Bai, Tian Ye, Wei Chow, Enxin Song, Qing-Guo Chen, Xiangtai Li, Zhen Dong, Lei Zhu, and Shuicheng Yan. Meissonic: Revitalizing masked generative transformers for efficient high-resolution text-to-image synthesis. arXiv preprint arXiv:2410.08261, 2024. 2" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 61, + 146, + 296, + 191 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 61, + 146, + 296, + 191 + ], + "spans": [ + { + "bbox": [ + 61, + 146, + 296, + 191 + ], + "type": "text", + "content": "[2] James Betker, Gabriel Goh, Li Jing, Tim Brooks, Jianfeng Wang, Linjie Li, Long Ouyang, Juntang Zhuang, Joyce Lee, Yufei Guo, et al. Improving image generation with better captions. Computer Science., 2:3, 2023. 2" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 62, + 191, + 295, + 235 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 191, + 295, + 235 + ], + "spans": [ + { + "bbox": [ + 62, + 191, + 295, + 235 + ], + "type": "text", + "content": "[3] Mathilde Caron, Hugo Touvron, Ishan Misra, Herve Jégou, Julien Mairal, Piotr Bojanowski, and Armand Joulin. Emerging properties in self-supervised vision transformers. In Int. Conf. Comput. Vis., 2021. 6" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 62, + 236, + 295, + 289 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 236, + 295, + 289 + ], + "spans": [ + { + "bbox": [ + 62, + 236, + 295, + 289 + ], + "type": "text", + "content": "[4] Junsong Chen, Jincheng Yu, Chongjian Ge, Lewei Yao, Enze Xie, Yue Wu, Zhongdao Wang, James Kwok, Ping Luo, Huchuan Lu, et al. Pixart-α: Fast training of diffusion transformer for photorealistic text-to-image synthesis. In ICLR, 2024. 2" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 62, + 291, + 295, + 324 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 291, + 295, + 324 + ], + "spans": [ + { + "bbox": [ + 62, + 291, + 295, + 324 + ], + "type": "text", + "content": "[5] Xi Chen, Lianghua Huang, Yu Liu, Yujun Shen, Deli Zhao, and Hengshuang Zhao. Anydoor: Zero-shot object-level image customization. In CVPR, 2024. 2, 3, 4, 5, 7, 8" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 62, + 325, + 295, + 378 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 325, + 295, + 378 + ], + "spans": [ + { + "bbox": [ + 62, + 325, + 295, + 378 + ], + "type": "text", + "content": "[6] Hyung Won Chung, Le Hou, Shayne Longpre, Barret Zoph, Yi Tay, William Fedus, Yunxuan Li, Xuezhi Wang, Mostafa Dehghani, Siddhartha Brahma, et al. Scaling instruction-finetuned language models. arXiv preprint arXiv:2210.11416, 2022. 2" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 62, + 380, + 295, + 423 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 380, + 295, + 423 + ], + "spans": [ + { + "bbox": [ + 62, + 380, + 295, + 423 + ], + "type": "text", + "content": "[7] Rinon Gal, Yuval Alaluf, Yuval Atzmon, Or Patashnik, Amit Haim Bermano, Gal Chechik, and Daniel Cohen-Or. An image is worth one word: Personalizing text-to-image generation using textual inversion. In ICLR, 2023. 3, 5" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 62, + 425, + 295, + 479 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 425, + 295, + 479 + ], + "spans": [ + { + "bbox": [ + 62, + 425, + 295, + 479 + ], + "type": "text", + "content": "[8] Yuchao Gu, Xintao Wang, Jay Zhangjie Wu, Yujun Shi, Yunpeng Chen, Zihan Fan, Wuyou Xiao, Rui Zhao, Shuning Chang, Weijia Wu, et al. Mix-of-show: Decentralized low-rank adaptation for multi-concept customization of diffusion models. NeurIPS, 2024. 2, 3, 5" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 62, + 480, + 295, + 523 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 480, + 295, + 523 + ], + "spans": [ + { + "bbox": [ + 62, + 480, + 295, + 523 + ], + "type": "text", + "content": "[9] Edward J Hu, Yelong Shen, Phillip Wallis abd Zeyuan Allen-Zhu, Yuanzhi Li, Shean Wang, Lu Wang, and Weizhu Chen. Lora: Low-rank adaptation of large language models. In ICLR, 2022. 2" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 57, + 525, + 295, + 567 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 525, + 295, + 567 + ], + "spans": [ + { + "bbox": [ + 57, + 525, + 295, + 567 + ], + "type": "text", + "content": "[10] Siteng Huang, Biao Gong, Yutong Feng, Xi Chen, Yuqian Fu, Yu Liu, and Donglin Wang. Learning disentangled identifiers for action-customized text-to-image generation. In CVPR, 2024. 5" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 57, + 569, + 295, + 612 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 569, + 295, + 612 + ], + "spans": [ + { + "bbox": [ + 57, + 569, + 295, + 612 + ], + "type": "text", + "content": "[11] Ziqi Huang, Tianxing Wu, Yuming Jiang, Kelvin CK Chan, and Ziwei Liu. Reversion: Diffusion-based relation inversion from images. arXiv preprint arXiv:2303.13495, 2023. 2, 3, 4, 5, 7" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 57, + 613, + 295, + 645 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 613, + 295, + 645 + ], + "spans": [ + { + "bbox": [ + 57, + 613, + 295, + 645 + ], + "type": "text", + "content": "[12] Diederik P Kingma and Jimmy Ba. Adam: A method for stochastic optimization. arXiv preprint arXiv:1412.6980, 2014. 6" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 57, + 647, + 295, + 690 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 647, + 295, + 690 + ], + "spans": [ + { + "bbox": [ + 57, + 647, + 295, + 690 + ], + "type": "text", + "content": "[13] Alexander Kirillov, Eric Mintun, Nikhila Ravi, Hanzi Mao, Chloe Rolland, Laura Gustafson, Tete Xiao, Spencer Whitehead, Alexander C. Berg, Wan-Yen Lo, Piotr Dollar, and Ross B. Girshick. Segment anything. In ICCV, 2023. 4" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 57, + 691, + 295, + 715 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 691, + 295, + 715 + ], + "spans": [ + { + "bbox": [ + 57, + 691, + 295, + 715 + ], + "type": "text", + "content": "[14] Lingjie Kong, Kai Wu, Xiaobin Hu, Wenhui Han, Jinlong Peng, Chengming Xu, Donghao Luo, Jiangning Zhang," + } + ] + } + ], + "index": 14 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 316, + 73, + 553, + 713 + ], + "type": "list", + "angle": 0, + "index": 31, + "blocks": [ + { + "bbox": [ + 333, + 73, + 553, + 106 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 333, + 73, + 553, + 106 + ], + "spans": [ + { + "bbox": [ + 333, + 73, + 553, + 106 + ], + "type": "text", + "content": "Chengjie Wang, and Yanwei Fu. Anymaker: Zero-shot general object customization via decoupled dual-level id injection. arXiv preprint arXiv:2406.11643, 2024. 2" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 316, + 107, + 553, + 140 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 107, + 553, + 140 + ], + "spans": [ + { + "bbox": [ + 316, + 107, + 553, + 140 + ], + "type": "text", + "content": "[15] Nupur Kumari, Bingliang Zhang, Richard Zhang, Eli Shechtman, and Jun-Yan Zhu. Multi-concept customization of text-to-image diffusion. In CVPR, 2023. 2, 3, 5, 6, 7, 8" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 316, + 141, + 553, + 184 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 141, + 553, + 184 + ], + "spans": [ + { + "bbox": [ + 316, + 141, + 553, + 184 + ], + "type": "text", + "content": "[16] Dongxu Li, Junnan Li, and Steven C. H. Hoi. Blip-diffusion: Pre-trained subject representation for controllable text-to-image generation and editing. In NeurIPS, 2023. 2, 3, 5, 7" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 316, + 186, + 553, + 230 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 186, + 553, + 230 + ], + "spans": [ + { + "bbox": [ + 316, + 186, + 553, + 230 + ], + "type": "text", + "content": "[17] Long Lian, Boyi Li, Adam Yala, and Trevor Darrell. Llm-grounded diffusion: Enhancing prompt understanding of text-to-image diffusion models with large language models. arXiv preprint arXiv:2305.13655, 2023. 2" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 316, + 232, + 553, + 284 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 232, + 553, + 284 + ], + "spans": [ + { + "bbox": [ + 316, + 232, + 553, + 284 + ], + "type": "text", + "content": "[18] Wang Lin, Jingyuan Chen, Jiaxin Shi, Yichen Zhu, Chen Liang, Junzhong Miao, Tao Jin, Zhou Zhao, Fei Wu, Shuicheng Yan, et al. Non-confusing generation of customized concepts in diffusion models. ICML, 2024. 2, 3, 5" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 316, + 286, + 553, + 308 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 286, + 553, + 308 + ], + "spans": [ + { + "bbox": [ + 316, + 286, + 553, + 308 + ], + "type": "text", + "content": "[19] Haotian Liu, Chunyuan Li, Qingyang Wu, and Yong Jae Lee. Visual instruction tuning. NeurIPS, 2024. 4" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 316, + 310, + 553, + 354 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 310, + 553, + 354 + ], + "spans": [ + { + "bbox": [ + 316, + 310, + 553, + 354 + ], + "type": "text", + "content": "[20] Zhiheng Liu, Yifei Zhang, Yujun Shen, Kecheng Zheng, Kai Zhu, Ruili Feng, Yu Liu, Deli Zhao, Jingren Zhou, and Yang Cao. Cones 2: Customizable image synthesis with multiple subjects. In NeurIPS, 2023. 2, 3, 7, 8" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 316, + 355, + 553, + 398 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 355, + 553, + 398 + ], + "spans": [ + { + "bbox": [ + 316, + 355, + 553, + 398 + ], + "type": "text", + "content": "[21] Joanna Materzynska, Josef Sivic, Eli Shechtman, Antonio Torralba, Richard Zhang, and Bryan Russell. Customizing motion in text-to-video diffusion models. arXiv preprint arXiv:2312.04966, 2023. 2" + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 316, + 399, + 553, + 453 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 399, + 553, + 453 + ], + "spans": [ + { + "bbox": [ + 316, + 399, + 553, + 453 + ], + "type": "text", + "content": "[22] mengmeng Ge, Xu Jia, Takashi Isobe, Xiaomin Li, Qinghe Wang, Jing Mu, Dong Zhou, liwang Amd, Huchuan Lu, Lu Tian, Ashish Sirasao, and Emad Barsoum. Customizing text-to-image generation with inverted interaction. In ACM MM, 2024. 2, 3" + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 316, + 456, + 553, + 499 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 456, + 553, + 499 + ], + "spans": [ + { + "bbox": [ + 316, + 456, + 553, + 499 + ], + "type": "text", + "content": "[23] Lianyu Pang, Jian Yin, Baoquan Zhao, Feize Wu, Fu Lee Wang, Qing Li, and Xudong Mao. Attndreambooth: Towards text-aligned personalized text-to-image generation. arXiv preprint arXiv:2406.05000, 2024. 2, 3" + } + ] + } + ], + "index": 25 + }, + { + "bbox": [ + 316, + 501, + 553, + 544 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 501, + 553, + 544 + ], + "spans": [ + { + "bbox": [ + 316, + 501, + 553, + 544 + ], + "type": "text", + "content": "[24] Maitreya Patel, Sangmin Jung, Chitta Baral, and Yezhou Yang. λ-eclipse: Multi-concept personalized text-to-image diffusion models by leveraging clip latent space. arXiv preprint arXiv:2402.05195, 2024. 2, 3, 5, 8" + } + ] + } + ], + "index": 26 + }, + { + "bbox": [ + 316, + 545, + 553, + 567 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 545, + 553, + 567 + ], + "spans": [ + { + "bbox": [ + 316, + 545, + 553, + 567 + ], + "type": "text", + "content": "[25] William Peebles and Saining Xie. Scalable diffusion models with transformers. In ICCV, 2023. 2" + } + ] + } + ], + "index": 27 + }, + { + "bbox": [ + 316, + 568, + 553, + 612 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 568, + 553, + 612 + ], + "spans": [ + { + "bbox": [ + 316, + 568, + 553, + 612 + ], + "type": "text", + "content": "[26] Pablo Pernias, Dominic Rampas, Mats Leon Richter, Christopher Pal, and Marc Aubreville. Würstchen: An efficient architecture for large-scale text-to-image diffusion models. In ICLR, 2023. 2" + } + ] + } + ], + "index": 28 + }, + { + "bbox": [ + 316, + 613, + 553, + 667 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 613, + 553, + 667 + ], + "spans": [ + { + "bbox": [ + 316, + 613, + 553, + 667 + ], + "type": "text", + "content": "[27] Dustin Podell, Zion English, Kyle Lacey, Andreas Blattmann, Tim Dockhorn, Jonas Müller, Joe Penna, and Robin Rombach. Sdxl: Improving latent diffusion models for high-resolution image synthesis. arXiv preprint arXiv:2307.01952, 2023. 2, 5" + } + ] + } + ], + "index": 29 + }, + { + "bbox": [ + 316, + 670, + 553, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 670, + 553, + 713 + ], + "spans": [ + { + "bbox": [ + 316, + 670, + 553, + 713 + ], + "type": "text", + "content": "[28] Alec Radford, Jong Wook Kim, Chris Hallacy, Aditya Ramesh, Gabriel Goh, Sandhini Agarwal, Girish Sastry, Amanda Askell, Pamela Mishkin, Jack Clark, Gretchen Krueger, and Ilya Sutskever. Learning transferable visual" + } + ] + } + ], + "index": 30 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "text", + "content": "15731" + } + ] + } + ], + "index": 32 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 8 + }, + { + "para_blocks": [ + { + "bbox": [ + 56, + 72, + 295, + 712 + ], + "type": "list", + "angle": 0, + "index": 15, + "blocks": [ + { + "bbox": [ + 76, + 72, + 294, + 94 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 76, + 72, + 294, + 94 + ], + "spans": [ + { + "bbox": [ + 76, + 72, + 294, + 94 + ], + "type": "text", + "content": "models from natural language supervision. In ICML, 2021. 2,3" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 56, + 96, + 295, + 140 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 96, + 295, + 140 + ], + "spans": [ + { + "bbox": [ + 56, + 96, + 295, + 140 + ], + "type": "text", + "content": "[29] Jingjing Ren, Wenbo Li, Haoyu Chen, Renjing Pei, Bin Shao, Yong Guo, Long Peng, Fenglong Song, and Lei Zhu. Ultrapixel: Advancing ultra-high-resolution image synthesis to new peaks. arXiv preprint arXiv:2407.02158, 2024. 2" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 56, + 143, + 294, + 175 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 143, + 294, + 175 + ], + "spans": [ + { + "bbox": [ + 56, + 143, + 294, + 175 + ], + "type": "text", + "content": "[30] Robin Rombach, Andreas Blattmann, Dominik Lorenz, Patrick Esser, and Björn Ommer. High-resolution image synthesis with latent diffusion models. In CVPR, 2022. 2" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 56, + 177, + 294, + 220 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 177, + 294, + 220 + ], + "spans": [ + { + "bbox": [ + 56, + 177, + 294, + 220 + ], + "type": "text", + "content": "[31] Nataniel Ruiz, Yuanzhen Li, Varun Jampani, Yael Pritch, Michael Rubinstein, and Kfir Aberman. Dreambooth: Fine tuning text-to-image diffusion models for subject-driven generation. In CVPR, 2023. 2, 3, 5, 6, 7, 8" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 56, + 223, + 294, + 289 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 223, + 294, + 289 + ], + "spans": [ + { + "bbox": [ + 56, + 223, + 294, + 289 + ], + "type": "text", + "content": "[32] Chitwan Sahara, William Chan, Saurabh Saxena, Lala Li, Jay Whang, Emily L. Denton, Seyed Kamyar Seyed Ghasemipour, Raphael Gontijo Lopes, Burcu Karagol Ayan, Tim Salimans, Jonathan Ho, David J. Fleet, and Mohammad Norouzi. Photorealistic text-to-image diffusion models with deep language understanding. In NeurIPS, 2022. 2" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 56, + 291, + 294, + 323 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 291, + 294, + 323 + ], + "spans": [ + { + "bbox": [ + 56, + 291, + 294, + 323 + ], + "type": "text", + "content": "[33] Ashish Vaswani, Noam Shazeer, Niki Parmar, Jakob Uszkoreit, Llion Jones, Aidan N Gomez, Lukasz Kaiser, and Illia Polosukhin. Attention is all you need. In NeurIPS, 2017. 2" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 56, + 326, + 294, + 367 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 326, + 294, + 367 + ], + "spans": [ + { + "bbox": [ + 56, + 326, + 294, + 367 + ], + "type": "text", + "content": "[34] Andrey Voynov, Qinghao Chu, Daniel Cohen-Or, and Kfir Aberman. p+: Extended textual conditioning in text-to-image generation. arXiv preprint arXiv:2303.09522, 2023. 3" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 56, + 371, + 294, + 415 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 371, + 294, + 415 + ], + "spans": [ + { + "bbox": [ + 56, + 371, + 294, + 415 + ], + "type": "text", + "content": "[35] X Wang, Siming Fu, Qihan Huang, Wanggui He, and Hao Jiang. Ms-diffusion: Multi-subject zero-shot image personalization with layout guidance. arXiv preprint arXiv:2406.07209, 2024. 2, 3, 4, 5, 6, 7, 8" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 56, + 417, + 294, + 460 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 417, + 294, + 460 + ], + "spans": [ + { + "bbox": [ + 56, + 417, + 294, + 460 + ], + "type": "text", + "content": "[36] Yuxiang Wei, Yabo Zhang, Zhilong Ji, Jinfeng Bai, Lei Zhang, and Wangmeng Zuo. ELITE: encoding visual concepts into textual embeddings for customized text-to-image generation. In ICCV, 2023. 2, 3, 5, 7, 8" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 56, + 463, + 294, + 506 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 463, + 294, + 506 + ], + "spans": [ + { + "bbox": [ + 56, + 463, + 294, + 506 + ], + "type": "text", + "content": "[37] Yujie Wei, Shiwei Zhang, Zhiwu Qing, Hangjie Yuan, Zhiheng Liu, Yu Liu, Yingya Zhang, Jingren Zhou, and Hongming Shan. Dreamvideo: Composing your dream videos with customized subject and motion. In CVPR, 2024. 3" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 56, + 509, + 294, + 552 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 509, + 294, + 552 + ], + "spans": [ + { + "bbox": [ + 56, + 509, + 294, + 552 + ], + "type": "text", + "content": "[38] Jianzong Wu, Xiangtai Li, Yanhong Zeng, Jiangning Zhang, Qianyu Zhou, Yining Li, Yunhai Tong, and Kai Chen. Motionbooth: Motion-aware customized text-to-video generation. arXiv preprint arXiv:2406.17758, 2024. 2" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 56, + 554, + 294, + 597 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 554, + 294, + 597 + ], + "spans": [ + { + "bbox": [ + 56, + 554, + 294, + 597 + ], + "type": "text", + "content": "[39] Size Wu, Wenwei Zhang, Lumin Xu, Sheng Jin, Xiangtai Li, Wentao Liu, and Chen Change Loy. Clipself: Vision transformer distills itself for open-vocabulary dense prediction. ICLR, 2024. 2, 6" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 56, + 600, + 294, + 643 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 600, + 294, + 643 + ], + "spans": [ + { + "bbox": [ + 56, + 600, + 294, + 643 + ], + "type": "text", + "content": "[40] Binxin Yang, Shuyang Gu, Bo Zhang, Ting Zhang, Xuejin Chen, Xiaoyan Sun, Dong Chen, and Fang Wen. Paint by example: Exemplar-based image editing with diffusion models. In CVPR, 2023. 2" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 56, + 646, + 294, + 667 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 646, + 294, + 667 + ], + "spans": [ + { + "bbox": [ + 56, + 646, + 294, + 667 + ], + "type": "text", + "content": "[41] Jie Yang, Ailing Zeng, Ruimao Zhang, and Lei Zhang. Unipose: Detecting any keypoints. ECCV, 2024. 4, 5" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 56, + 670, + 294, + 712 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 670, + 294, + 712 + ], + "spans": [ + { + "bbox": [ + 56, + 670, + 294, + 712 + ], + "type": "text", + "content": "[42] Hu Ye, Jun Zhang, Sibo Liu, Xiao Han, and Wei Yang. Ip-adapter: Text compatible image prompt adapter for text-to-image diffusion models. arXiv preprint arxiv:2308.06721, 2023. 2" + } + ] + } + ], + "index": 14 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 315, + 72, + 554, + 174 + ], + "type": "list", + "angle": 0, + "index": 19, + "blocks": [ + { + "bbox": [ + 315, + 72, + 554, + 116 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 315, + 72, + 554, + 116 + ], + "spans": [ + { + "bbox": [ + 315, + 72, + 554, + 116 + ], + "type": "text", + "content": "[43] Yuxuan Zhang, Yiren Song, Jiaming Liu, Rui Wang, Jinpeng Yu, Hao Tang, Huaxia Li, Xu Tang, Yao Hu, Han Pan, et al. Ssr-encoder: Encoding selective subject representation for subject-driven generation. In CVPR, 2024. 2, 3, 5, 7, 8" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 315, + 118, + 553, + 139 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 315, + 118, + 553, + 139 + ], + "spans": [ + { + "bbox": [ + 315, + 118, + 553, + 139 + ], + "type": "text", + "content": "[44] Chong Zhou, Chen Change Loy, and Bo Dai. Extract free dense labels from clip. In ECCV, 2022. 5" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 316, + 141, + 553, + 174 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 141, + 553, + 174 + ], + "spans": [ + { + "bbox": [ + 316, + 141, + 553, + 174 + ], + "type": "text", + "content": "[45] Chenyang Zhu, Kai Li, Yue Ma, Chunming He, and Li Xiu. Multibooth: Towards generating all your concepts in an image from text. arXiv preprint arXiv:2404.14239, 2024. 5" + } + ] + } + ], + "index": 18 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 749, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 749, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 749, + 317, + 757 + ], + "type": "text", + "content": "15732" + } + ] + } + ], + "index": 20 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 9 + } + ], + "_backend": "vlm", + "_version_name": "2.6.4" +} \ No newline at end of file diff --git a/2025/DreamText_ High Fidelity Scene Text Synthesis/6bc8ea68-7656-478a-8d11-90857d6ca90e_content_list.json b/2025/DreamText_ High Fidelity Scene Text Synthesis/6bc8ea68-7656-478a-8d11-90857d6ca90e_content_list.json new file mode 100644 index 0000000000000000000000000000000000000000..5aff7754a4dcb5ed6f4620c2f536806a2329309c --- /dev/null +++ b/2025/DreamText_ High Fidelity Scene Text Synthesis/6bc8ea68-7656-478a-8d11-90857d6ca90e_content_list.json @@ -0,0 +1,1469 @@ +[ + { + "type": "text", + "text": "DREAMTEXT: High Fidelity Scene Text Synthesis", + "text_level": 1, + "bbox": [ + 246, + 130, + 751, + 152 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Yibin Wang $^{1,2}$ , Weizhong Zhang $^{1,4}$ , Honghui Xu $^{5}$ , Cheng Jin $^{1,3\\dagger}$", + "bbox": [ + 236, + 178, + 766, + 198 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "$^{1}$ Fudan University $^{2}$ Shanghai Innovation Institute", + "bbox": [ + 303, + 198, + 692, + 215 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "$^{3}$ Innovation Center of Calligraphy and Painting Creation Technology, MCT", + "bbox": [ + 200, + 215, + 795, + 234 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "$^{4}$ Shanghai Key Laboratory of Intelligent Information Processing $^{5}$ Zhejiang University of Technology", + "bbox": [ + 102, + 233, + 895, + 252 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "yibinwang1121@163.com, weizhongzhang@fudan.edu.cn, xhh@zjut.edu.cn, jc@fudan.edu.cn", + "bbox": [ + 129, + 253, + 866, + 268 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/6ebdc6a9714dca6a4f1b52ca44fd15efe7bf40df27eece4554c3bf0a39fc7067.jpg", + "image_caption": [ + "Figure 1. Displayed are the results generated using our DREAMTEXT, showcasing its prowess across varied inputs." + ], + "image_footnote": [], + "bbox": [ + 109, + 305, + 883, + 508 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Abstract", + "text_level": 1, + "bbox": [ + 246, + 547, + 326, + 561 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Scene text synthesis involves rendering specified texts onto arbitrary images. Current methods typically formulate this task in an end-to-end manner but lack effective character-level guidance during training. Besides, their text encoders, pre-trained on a single font type, struggle to adapt to the diverse font styles encountered in practical applications. Consequently, these methods suffer from character distortion, repetition, and absence, particularly in polystylistic scenarios. To this end, this paper proposes DREAMTEXT for high-fidelity scene text synthesis. Our key idea is to reconstruct the diffusion training process, introducing more refined guidance tailored to this task, to expose and rectify the model's attention at the character level and strengthen its learning of text regions. This transformation poses a hybrid optimization challenge, involving both discrete and continuous variables. To effectively tackle this challenge, we employ a heuristic alternate optimization strategy. Meanwhile, we jointly train the text encoder and generator to comprehensively learn and utilize the diverse font present", + "bbox": [ + 88, + 580, + 485, + 869 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "in the training dataset. This joint training is seamlessly integrated into the alternate optimization process, fostering a synergistic relationship between learning character embedding and re-estimating character attention. Specifically, in each step, we first encode potential character-generated position information from cross-attention maps into latent character masks. These masks are then utilized to update the representation of specific characters in the current step, which, in turn, enables the generator to correct the character's attention in the subsequent steps. Both qualitative and quantitative results demonstrate the superiority of our method to the state of the art. Our project page is here.", + "bbox": [ + 511, + 549, + 908, + 731 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "1. Introduction", + "text_level": 1, + "bbox": [ + 513, + 766, + 645, + 782 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Scene text synthesis involves modifying or inserting specified text in arbitrary images while maintaining its natural and realistic appearance. Earlier GAN-based style transfer methods [13, 14, 16, 27] accomplish this by transferring the text style from a reference image to the rendered target text image. However, they are constrained in their capacity to generate text in arbitrary styles, such as font and color. These", + "bbox": [ + 511, + 794, + 908, + 902 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "CVF", + "bbox": [ + 106, + 2, + 181, + 42 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore.", + "bbox": [ + 236, + 0, + 810, + 46 + ], + "page_idx": 0 + }, + { + "type": "page_footnote", + "text": "† Corresponding author.", + "bbox": [ + 93, + 886, + 220, + 900 + ], + "page_idx": 0 + }, + { + "type": "page_number", + "text": "28555", + "bbox": [ + 478, + 944, + 519, + 957 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/208bd9f60383832fc13a1d2ca4503e20137e067a0add3a23ea91ab20c390de19.jpg", + "image_caption": [ + "Problem 1: Character repetition and absence", + "Problem 2: Character distortion" + ], + "image_footnote": [], + "bbox": [ + 135, + 107, + 864, + 224 + ], + "page_idx": 1 + }, + { + "type": "image", + "img_path": "images/35252f8ca281af7ac77b74acb845e6df83f2855272fe11c25ac8f977ebd04778.jpg", + "image_caption": [ + "Figure 2. Current methods encounter significant challenges, i.e., character repetition and absence (top) and character distortion (bottom)." + ], + "image_footnote": [], + "bbox": [ + 135, + 244, + 864, + 359 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "limitations are effectively addressed by utilizing the diffusion model [5, 8, 11, 12, 17, 21, 22], harnessing its inherent prior knowledge and robust capabilities. Nevertheless, these methods still encounter challenges in generating accurate text within polystylistic images due to the inadequate conditional guidance provided by their character-unaware text encoder. To this end, a recent study [29] proposes a character-level text encoder, providing more robust conditional guidance for the diffusion model.", + "bbox": [ + 89, + 398, + 483, + 532 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Despite their effectiveness, these methods are still difficult to accurately render text within complex scenes. Specifically, these methods such as UDiffText [29] and TextDiffuser [5] solely utilize a single font type to pre-train the text encoder, which is subsequently employed to fine-tune the generator. However, the diverse font styles encountered in practical applications present a significant challenge: the restricted representation domain hampers their ability to render text with unseen font styles, e.g., cases shown in Fig. 2 (bottom). Furthermore, these methods rely on character segmentation masks to supervise characters' attention for position control. We argue this approach has significant limitations: (1) unlike classification tasks with uniquely determined labels, the optimal generation positions for characters can be diverse. Therefore, using the specific masks to rigidly constrain the model may limit its flexibility in estimating optimal positions, hindering its ability to adapt to varied and complex scenarios. Besides, (2) these masks generated by a pre-trained segmentation model [5] are imprecise, tending to over-segment character regions. As a result, the model encounters issues of character repetition and absence as illustrated in Fig. 2 (top left) and (top right) respectively. We suspect these issues stem from characters' attention maps, as they reflect the underlying response to the character's", + "bbox": [ + 89, + 537, + 483, + 901 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "final generated position [23, 24]. Therefore, we visualize the attention maps of their problematic results by rendering attention directly onto the images. We observe certain characters' attention misalignment, resulting in character duplication (see Fig. 3 (a) and (c)). Besides, excessive dispersion of attention also leads to character absence (see Fig. 3 (b)). This observation underscores the insufficiency of these methods in guiding the model autonomously estimating optimal character-generated position. Further, we also assess specific methods [5, 29] quantitatively on two datasets to evaluate their proficiency in directing characters' attention toward their optimal generation regions throughout multiple global training steps. Specifically, for each test sample, we extract potential position masks for all characters based on their cross-attention maps from the inference process. Then, we compute the mean Intersection over Union (mIoU) between these masks and ground truth character segmentation masks. Their inoptimal results depicted in Fig. 4 further corroborate our hypothesis.", + "bbox": [ + 511, + 398, + 906, + 684 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "In summary, current methods encounter challenges in accurately rendering text within complex scenes due to their constrained representation domain and synthesizing characters in their ideal regions due to the lack of effective guidance for the model to estimate optimal character-generated position. To this end, this paper proposes DREAMTEXT for high-fidelity text synthesis. We reconstruct the diffusion training process, embedding more refined guidance: we expose and rectify the characters' attention to focus more precisely on their ideal generation areas, and enhance the learning of text regions and character representation by additional constraints. However, this reconstruction inevitably introduces a complex hybrid optimization problem encompassing both discrete and continuous variables. Therefore,", + "bbox": [ + 511, + 689, + 908, + 900 + ], + "page_idx": 1 + }, + { + "type": "page_number", + "text": "28556", + "bbox": [ + 478, + 945, + 517, + 955 + ], + "page_idx": 1 + }, + { + "type": "image", + "img_path": "images/b05c4cdd7650ce0b011e58a8ec5430088a0bad5fe880c76c564d249393632ff1.jpg", + "image_caption": [ + "Figure 3. The problematic results rendered by characters' attention maps (AMs)." + ], + "image_footnote": [], + "bbox": [ + 93, + 88, + 480, + 391 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "we design a heuristic alternate optimization strategy. Concurrently, we jointly train the text encoder and the generator, leveraging the diverse font styles within the training dataset to enrich character representation space. We demonstrate that this joint training can be seamlessly integrated into the heuristic alternate optimization process, facilitating a synergistic interplay between character representations learning and character attention re-estimation. Specifically, in each iteration, the pipeline begins by encoding the potential character-generated position information from cross-attention maps into latent character masks. These masks are subsequently employed to refine the representation of specific characters within the text encoder, facilitating the calibration of the characters' attention in the subsequent step. This iterative process contributes to both learning of better character representation and, as a by-product, explicit guidance for autonomously estimating character position.", + "bbox": [ + 88, + 446, + 482, + 702 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Notably, the generator may initially struggle to direct attention to the desired generation region for each character, resulting in suboptimal latent masks and impacting training progress. Unlike existing methods [5, 29] that rigidly constrain character attention, we employ a balanced supervision strategy. Initially, we assist in calibrating the attention using character segmentation masks for a warm-up in the earlier stage. Once the model has preliminarily acquired the ability to estimate ideal generation positions, we remove this guidance, allowing the model to autonomous learning iteratively. This strategy strikes a balance between constraining the model and unleashing its flexibility in estimating optimal generation positions. Fig. 4 illustrates the efficacy of our", + "bbox": [ + 89, + 704, + 485, + 901 + ], + "page_idx": 2 + }, + { + "type": "image", + "img_path": "images/41b6185fbff08675be8f802b118e82b3e53adf16bbdb725d7cb43bf017de6b4c.jpg", + "image_caption": [ + "Figure 4. The mIoU scores of UDiffText, TextDiffuser, and our method on LAION-OCR and SynthText over global training steps. Our method adopts a balanced supervision strategy: we initially use latent character masks to steer the character's attention for a warm-up in the earlier stage and stop guiding after 25,000 steps." + ], + "image_footnote": [], + "bbox": [ + 516, + 87, + 903, + 181 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "method in continuously improving the generator's ability to estimate the optimal characters' position.", + "bbox": [ + 511, + 271, + 906, + 301 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Our contributions are: (1) The proposed DREAMTEXT effectively alleviates the issues of character repetition, absence, and distortion encountered by existing methods. (2) Our heuristic alternate optimization strategy, integrating the joint learning of the text encoder and U-Net, orchestrates a symbiotic relationship between learning character representations and re-estimating character attention. (3) Our balanced supervision strategy strikes a balance between constraining the model and unleashing its flexibility in estimating optimal generation positions. (4) Both qualitative and quantitative results demonstrate the superiority of our method.", + "bbox": [ + 511, + 301, + 908, + 467 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "2. Related Work", + "text_level": 1, + "bbox": [ + 513, + 481, + 653, + 498 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Earlier GAN-based methods achieve scene text synthesis by transferring the text style in a reference image to the rendered target text image [10, 16, 18, 20, 27]. Specifically, STEFANN [16] employs a FANnet to edit individual characters and incorporates a placement algorithm to generate the desired word. SRNet [25] and MOSTEL [14], on the other hand, split the task into two main parts: background inpainting and text style transfer. This approach allows for the end-to-end synthesizing of entire words. Despite their simplicity and effectiveness, they are limited in their ability to generate text in arbitrary styles and locations, often producing less natural-looking images.", + "bbox": [ + 511, + 507, + 908, + 688 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "To this end, several diffusion-based approaches have emerged to address these challenges. These approaches harness the robust capabilities of diffusion models to edit or generate scene text, thereby enhancing the quality and diversity of the generated content. DiffSTE [8] proposes a dual encoder structure, comprising a character text encoder and an instruction text encoder. These components are employed for instruction tuning, providing improved control over the backbone network. DiffUTE [4] employs an OCR-based glyph encoder to extract glyph guidance from the rendered glyph image. TextDiffuser [5] concatenates the segmentation mask as conditional input and employs character-aware loss to control the generated characters precisely. However, they still face challenges in generating accurate text", + "bbox": [ + 511, + 689, + 910, + 900 + ], + "page_idx": 2 + }, + { + "type": "page_number", + "text": "28557", + "bbox": [ + 478, + 944, + 517, + 955 + ], + "page_idx": 2 + }, + { + "type": "image", + "img_path": "images/016c3053ed95d0bb11c7831077688a40c8caf28d5bbd776bb086f7c16f4c4442.jpg", + "image_caption": [ + "Figure 5. An overview of proposed heuristic alternate optimization strategy." + ], + "image_footnote": [], + "bbox": [ + 93, + 88, + 908, + 390 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "within polystylistic images, primarily due to the insufficient conditional guidance provided by their character-unaware text encoder. Motivated by this, UDiffText [29] replaces the original CLIP text encoder in Stable Diffusion with a character-level text encoder with a unitary font. It also employs local attention loss for position control, relying on ground truth character segmentation maps.", + "bbox": [ + 88, + 424, + 482, + 530 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "However, these methods still encounter challenges in accurately rendering text within complex scenes and synthesizing characters in their ideal regions, attributed to their constrained character representation domain and lack of effective guidance for the model to autonomously estimate character-generated position, respectively.", + "bbox": [ + 88, + 530, + 485, + 623 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "3. Method", + "text_level": 1, + "bbox": [ + 89, + 633, + 183, + 648 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Given an input image $z$ , a text region specified by a binary mask $B$ and a text condition $c$ , scene text synthesis is designed to render the given text onto the input image at the given text region. A natural idea for scene text synthesis is to formulate it into a latent diffusion process [15], which is equivalent to solving the following optimization problem:", + "bbox": [ + 89, + 659, + 483, + 751 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\min _ {(\\theta , \\vartheta)} \\mathcal {L} _ {L D M} \\triangleq \\mathbb {E} _ {\\boldsymbol {z}, c, \\epsilon \\sim N (0, 1), t} \\| \\boldsymbol {\\epsilon} - \\epsilon_ {\\theta} (\\boldsymbol {z} _ {t}, t, \\psi_ {\\vartheta} (c), \\boldsymbol {B}) \\| _ {2} ^ {2},\n$$\n", + "text_format": "latex", + "bbox": [ + 94, + 758, + 478, + 787 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "where $\\psi_{\\vartheta}$ is the learnable character-level text encoder initialized by UDifftext [29] and $\\theta$ is the model parameter vector. However, we argue that such a straightforward formulation could be problematic due to the following reasons:", + "bbox": [ + 89, + 794, + 485, + 854 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "- Without explicit guidance for rectifying characters' attention, character repetition and absence can inevitably occur in the rendered results as shown in Fig. 2.", + "bbox": [ + 89, + 854, + 485, + 901 + ], + "page_idx": 3 + }, + { + "type": "list", + "sub_type": "text", + "list_items": [ + "- This formulation uniformly measures the mean distance between all pixels, without emphasizing the text regions. Consequently, text-unrelated regions can adversely affect the learning of the text encoder during backpropagation. The experimental analysis is provided in Sec. 4.3.", + "- Due to the intricate interactions among the U-Net and text encoder, end-to-end training is particularly challenging as shown in Sec. 4.3. Therefore, a more refined guidance for the training process is necessary." + ], + "bbox": [ + 513, + 424, + 906, + 561 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "In this paper, we propose DREAMTEXT and develop a heuristic iterative optimization pipeline to enable efficient training. Our key idea is to explicitly model the potential character positions in attention maps by encoding them into latent character masks in each step. These masks are then used to enhance the learning of the text encoder through our carefully crafted losses, facilitating the calibration of character attention in the subsequent step. This heuristic iterative process allows our model to dynamically alternate between optimizing character embedding and re-estimating character generation positions. The details are presented in the following sections.", + "bbox": [ + 511, + 569, + 906, + 750 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "3.1. Latent Diffusion with Refined Guidance", + "text_level": 1, + "bbox": [ + 511, + 782, + 856, + 797 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "3.1.1. Latent Character Mask", + "text_level": 1, + "bbox": [ + 511, + 813, + 723, + 827 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "We utilize cross-attention maps to generate latent character masks, extracting characters' latent position information in the current step. Given latent image $z_{t}$ and text embedding $\\psi_{\\vartheta}(c)$ , the cross-attention map in layer $l$ is defined as follows:", + "bbox": [ + 511, + 840, + 906, + 902 + ], + "page_idx": 3 + }, + { + "type": "page_number", + "text": "28558", + "bbox": [ + 478, + 944, + 519, + 955 + ], + "page_idx": 3 + }, + { + "type": "image", + "img_path": "images/0c95447a95ca4921c541351b13ab80dc92346dc58fafc23aec3b6dea3c217086.jpg", + "image_caption": [ + "Figure 6. Qualitative comparative results against state-of-the-art methods." + ], + "image_footnote": [], + "bbox": [ + 133, + 88, + 864, + 329 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\boldsymbol {Q} _ {l} = \\boldsymbol {z} _ {t} \\boldsymbol {W} _ {l} ^ {q}, \\quad \\boldsymbol {K} _ {l} = \\psi_ {\\vartheta} (c) \\boldsymbol {W} _ {l} ^ {k}, \\tag {1}\n$$\n", + "text_format": "latex", + "bbox": [ + 181, + 388, + 483, + 407 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\boldsymbol {A} _ {l} = \\operatorname {s o f t m a x} \\left(\\frac {\\boldsymbol {Q} _ {l} \\boldsymbol {K} _ {l} ^ {T}}{\\sqrt {d}}\\right), \\tag {2}\n$$\n", + "text_format": "latex", + "bbox": [ + 212, + 410, + 483, + 446 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "where $\\pmb{W}_l^q$ and $\\pmb{W}_l^k$ are learnable parameters and $d$ is the embedding dimension. The attention maps $\\pmb{A}_l$ is reshaped to $N \\times H \\times W$ , where $N$ represents the number of text tokens such that each slice $\\pmb{A}_l^i \\in \\mathbb{R}^{H \\times W}$ represents the region attended by the $i$ -th token. These attention maps are then averaged across all layers of the U-Net to get the mean response $\\bar{\\pmb{A}} = \\frac{1}{L} \\sum_{l=1}^{L} \\pmb{A}_l$ for the regions attended by character tokens.", + "bbox": [ + 89, + 457, + 483, + 575 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "To obtain the latent character masks $M \\in \\mathbb{R}^{N \\times H \\times W}$ , we first apply Gaussian blur to the attention map to execute low-pass filtering $\\mathrm{blur}(\\bar{A})$ . This step helps mitigate excessive variance in the attended regions, ensuring a more uniform distribution of attention across relevant character regions. Subsequently, we employ a straightforward thresholding process to convert the blurred attention map into binary character masks. This process is encapsulated by the function $f(\\cdot)$ , which applies a threshold and assigns a value of 1 to pixel values exceeding the threshold, and 0 otherwise. The threshold is calculated as the mean value plus twice the variance of averaged attention maps in this work. That is:", + "bbox": [ + 89, + 577, + 483, + 760 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\nf (\\boldsymbol {X}) = \\left\\{ \\begin{array}{l} 1, \\text {i f} x _ {i, j} > \\operatorname {m e a n} (\\boldsymbol {X}) + 2 \\operatorname {s t d} (\\boldsymbol {X}) \\\\ 0, \\text {o t h e r w i s e} \\end{array} \\right. \\tag {3}\n$$\n", + "text_format": "latex", + "bbox": [ + 130, + 782, + 483, + 823 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Therefore, the equation for obtaining $M$ can be expressed as:", + "bbox": [ + 89, + 832, + 483, + 848 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\boldsymbol {M} = f (\\operatorname {b l u r} (\\bar {\\boldsymbol {A}})). \\tag {4}\n$$\n", + "text_format": "latex", + "bbox": [ + 223, + 862, + 483, + 880 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Based on the latent character masks $M$ , we employ several", + "bbox": [ + 109, + 885, + 482, + 901 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "loss functions to optimize both the text encoder and U-Net iteratively. We will delve into these in the following sections.", + "bbox": [ + 511, + 364, + 906, + 395 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "3.1.2. Masked Diffusion Loss", + "text_level": 1, + "bbox": [ + 511, + 401, + 718, + 415 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "We first extend the Eq. 1 to highlight all desired character tokens in the current step. Specifically, within a text $c$ containing $k$ tokens of interest, the diffusion loss of the corresponding pixels obtained from each token's latent character mask is applied with an additional weighting factor $\\gamma$ . Formally, considering $M_{k} = \\bigvee_{i=1}^{k} M_{i}$ to the union of the pixels of $k$ characters, the masked diffusion loss is formulated as,", + "bbox": [ + 511, + 420, + 906, + 527 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {L} _ {\\text {m a s k}} = \\mathbb {E} _ {z, c, \\epsilon \\sim N (0, 1), t} \\| (1 + \\gamma \\boldsymbol {M} _ {k}) (\\boldsymbol {\\epsilon} - \\epsilon_ {\\theta} (\\boldsymbol {z} _ {t}, t, \\psi_ {\\vartheta} (c))) \\| _ {2} ^ {2}.\n$$\n", + "text_format": "latex", + "bbox": [ + 511, + 537, + 919, + 556 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "We omit binary mask $B$ here for the simplicity.", + "bbox": [ + 511, + 565, + 823, + 580 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "3.1.3. Cross Attention Loss", + "text_level": 1, + "bbox": [ + 511, + 587, + 705, + 599 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "The masked diffusion loss is designed to focus on all desired concept tokens within a prompt. Additionally, to ensure that each token encodes information specific to its corresponding synthesis position, we incorporate the cross-attention loss [2]. This loss encourages tokens to attend exclusively to their corresponding target regions,", + "bbox": [ + 511, + 606, + 905, + 698 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {L} _ {a t t n} = \\mathbb {E} _ {\\boldsymbol {z}, c, t} \\| C _ {a t t n} (\\boldsymbol {z} _ {t}, \\psi_ {\\vartheta} (c) _ {i}) - \\boldsymbol {M} _ {i} \\| _ {2} ^ {2}. \\tag {5}\n$$\n", + "text_format": "latex", + "bbox": [ + 558, + 705, + 906, + 724 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Here, $C_{attn}(z_t, \\psi_\\vartheta(c)_i)$ represents the cross-attention map between the visual representation $z_t$ and the token $\\psi_\\vartheta(c)_i$ , while $M_i$ denotes the latent character mask of $\\psi_\\vartheta(c)_i$ .", + "bbox": [ + 511, + 734, + 906, + 779 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "However, the losses mentioned above alone are inadequate for achieving optimal character representation. This is primarily due to the inherent noise in the latent character masks, which frequently results in under- or oversegmentation of the ideal character regions during training. To mitigate the influence of noise and prevent skewed learning of character representation, we introduce additional losses [29] as shown below.", + "bbox": [ + 511, + 780, + 908, + 900 + ], + "page_idx": 4 + }, + { + "type": "page_number", + "text": "28559", + "bbox": [ + 478, + 944, + 517, + 955 + ], + "page_idx": 4 + }, + { + "type": "table", + "img_path": "images/1893a91a5b39adc9f2f30341d1e44749dcc6465f6009021479943b7320c6f303.jpg", + "table_caption": [ + "Table 1. Quantitative comparison against baselines." + ], + "table_footnote": [], + "table_body": "
MethodsSeqAcc-ReconSeqAcc-EditingFIDLPIPS
ICDAR13(8ch)ICDAR13TextSegLAION-OCRICDAR13(8ch)ICDAR13TextSegLAION-OCR
CVPR'22 SD-Inpainting [15]0.320.290.110.150.080.070.040.0526.780.0696
arXiv'23 DiffSTE [8]0.450.370.500.410.340.290.470.2751.670.1050
AAAI'23 MOSTEL [14]0.750.680.640.710.350.280.250.4425.090.0605
NIPS'23 TextDiffuser [5]0.870.810.680.800.820.750.660.6432.250.0834
ICLR'24 AnyText [19]0.890.870.810.860.810.790.800.7222.730.0651
ECCV'24 UDiffText [29]0.940.910.930.900.840.830.840.7815.790.0564
DreamText0.950.940.960.930.870.890.910.8812.130.0328
", + "bbox": [ + 93, + 109, + 903, + 212 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "3.1.4. Cross-modal Aligned Loss", + "text_level": 1, + "bbox": [ + 89, + 228, + 320, + 242 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "To obtain accurate and robust embeddings, we introduce an image encoder $\\xi$ , text head $H_{t}$ , and visual head $H_{v}$ to align cross-modal character features. Specifically, we compute the cosine similarity objective between visual and text representations to maximize the alignment between two modalities:", + "bbox": [ + 89, + 246, + 485, + 323 + ], + "page_idx": 5 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {L} _ {\\text {a l i g n}} = \\frac {\\left\\langle H _ {t} (\\mathbf {y}) , H _ {v} (\\xi (\\mathbf {I})) \\right\\rangle}{\\| H _ {t} (\\mathbf {y}) \\| _ {2} \\cdot \\| H _ {v} (\\xi (\\mathbf {I})) \\| _ {2}}. \\tag {6}\n$$\n", + "text_format": "latex", + "bbox": [ + 169, + 332, + 483, + 366 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "where $I$ represents the text image segmented via the bounding box from datasets and $\\langle ,\\rangle$ presents the inner product of two vectors. Note that to alleviate noise stemming from background and color variations, we preprocess the text image by converting it to grayscale.", + "bbox": [ + 89, + 368, + 483, + 443 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Through training, the text head learns to map the whole information of multiple characters into the text feature space. Similarly, the image head maps the overall visual text information into the visual feature space. This effectively ensures the overall textual and visual representations correspond, capturing high-level cross-modal consistency.", + "bbox": [ + 89, + 443, + 483, + 532 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "3.1.5. Character Id Loss", + "text_level": 1, + "bbox": [ + 89, + 539, + 264, + 551 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Additionally, we introduce character Id loss to ensure that the learned embeddings are highly distinguishable. To be precise, a multi-label classification head $H_{l}$ is introduced to predict character indices from text embeddings $y$ . Then, the cross-entropy objective is formulated,", + "bbox": [ + 89, + 556, + 483, + 633 + ], + "page_idx": 5 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {L} _ {i d} = - \\sum_ {i = 1} ^ {N} \\sum_ {j = 1} ^ {K} l _ {i, j} \\log \\left(H _ {l} (\\mathbf {y}) _ {j}\\right), \\tag {7}\n$$\n", + "text_format": "latex", + "bbox": [ + 174, + 638, + 483, + 681 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "where $N$ and $K$ represent the number of characters and the possible indices respectively while $l$ denotes the ground truth labels. This loss aggregates over all characters in the target text, ensuring the text encoder produces distinguishable embeddings for each character.", + "bbox": [ + 89, + 686, + 483, + 762 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "The complete objective of our training strategy can be expressed as", + "bbox": [ + 89, + 763, + 482, + 792 + ], + "page_idx": 5 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {L} = \\mathcal {L} _ {\\text {m a s k}} + \\alpha \\mathcal {L} _ {\\text {a t t n}} + \\beta (\\mathcal {L} _ {\\text {a l i g n}} + \\mathcal {L} _ {i d}). \\tag {8}\n$$\n", + "text_format": "latex", + "bbox": [ + 145, + 799, + 483, + 816 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "3.2. Optimization Strategy", + "text_level": 1, + "bbox": [ + 89, + 829, + 297, + 845 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "3.2.1. Heuristic Alternate Optimization", + "text_level": 1, + "bbox": [ + 89, + 851, + 367, + 866 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Our objective involves numerous discrete variables, rendering it non-differentiable and making vanilla SGD invalid.", + "bbox": [ + 89, + 869, + 485, + 900 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Thus, we employ a heuristic alternate optimization strategy, utilizing latent character masks that encapsulate potential synthesis position information, to facilitate a symbiotic relationship between text encoder and U-Net. Specifically, we perform an alternating update between the tokens given the latent character masks and the masks themselves. During optimization, the masks are computed by Eq. 4, while for other parameters, we fix the masks and calculate gradients accordingly. Using the losses above, we first optimize the representation of specific characters given masks in each iteration, which, in turn, enables the generator to rectify the character's attention in subsequent steps. This process allows model to dynamically alternate between optimizing character embeddings and re-estimating character masks.", + "bbox": [ + 511, + 228, + 906, + 439 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "3.2.2. Balanced Supervision on Character Attention", + "text_level": 1, + "bbox": [ + 511, + 449, + 877, + 465 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "The initial training phase may challenge the generator in directing attention to the intended generation region for each character. Unlike existing methods that excessively constrain character attention which inevitably limits the model's flexibility, we adopt a balanced supervision. Initially, we guide attention calibration by cross-entropy objective between latent character masks and character segmentation masks from datasets similar to Eq. 7 for a warm-up in the earlier stage. Once the model preliminarily acquires the capability to estimate ideal generation positions (approximately 25k steps as shown in Fig. 4), we cease guidance, enabling it to engage in autonomous iterative learning. This strategy balances constraining the model and unleashing its flexibility in estimating characters' optimal generation positions.", + "bbox": [ + 511, + 468, + 908, + 681 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "4. Experiments", + "text_level": 1, + "bbox": [ + 511, + 696, + 645, + 713 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "4.1. Implementation Details", + "text_level": 1, + "bbox": [ + 511, + 722, + 730, + 737 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "4.1.1. Datasets", + "text_level": 1, + "bbox": [ + 511, + 744, + 619, + 758 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Several datasets are used in our experiments: (1) SynthText [6] comprises 800,000 images containing around 8 million synthetic word instances. Each text instance is annotated with its corresponding text string, along with word-level and character-level bounding boxes. (2) LAION-OCR [5] contains 9,194,613 filtered high-quality text images including advertisements, notes, posters, covers, memes, logos, etc. (3) ICDAR13 [9] is widely recognized as the benchmark dataset for evaluating near-horizontal text detection, which consists", + "bbox": [ + 511, + 763, + 906, + 900 + ], + "page_idx": 5 + }, + { + "type": "page_number", + "text": "28560", + "bbox": [ + 478, + 944, + 519, + 955 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/fe9665a40efabaf5f6e9303953e02affd1798d3abb6d02842a27edd39ddb7132.jpg", + "image_caption": [ + "Figure 7. Qualitative comparative results against state-of-the-art methods." + ], + "image_footnote": [], + "bbox": [ + 133, + 89, + 864, + 333 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "of 233 test images used for evaluation purposes. (4) TextSeg [26] comprises 4,024 real-world text images sourced from various sources such as posters, greeting cards, book covers, logos, road signs, billboards, digital designs, handwritten notes, and more. Note that only SynthText and LAION-OCR datasets provide the character segmentation maps. We train our model using the training subsets of these datasets and randomly select 100 images from test subsets for testing.", + "bbox": [ + 89, + 369, + 483, + 489 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "4.1.2. Training Configurations", + "text_level": 1, + "bbox": [ + 89, + 500, + 305, + 516 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "We train our model based on the pre-trained checkpoint of SD-v2.0 inpainting version and text encoder in [29]. The pre-trained Vision Transformer [1] is used as our image encoder. The model is fine-tuned using 4 NVIDIA A100 GPUs on LAION-OCR for 200k steps, SynthText for 150k steps, TextSeg for 50k steps, and ICDAR13 for an additional 10k steps. We set $\\alpha$ to 0.01, and $\\beta$ to 0.001. Additionally, we utilize a batch size of 16 and a learning rate of $5 \\times 10^{-5}$ . The inference time of our DREAMTEXT is 8.5 seconds only.", + "bbox": [ + 89, + 520, + 483, + 657 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "4.1.3. Evaluation", + "text_level": 1, + "bbox": [ + 89, + 667, + 212, + 681 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Text accuracy: We assess all methods on two tasks: scene text reconstruction and scene text editing [29]. For the scene text reconstruction task, we use the models to reconstruct the text image using the provided ground truth text label and binary mask. In the scene text editing task, we replace the original text in each image with a random word and evaluate the models by generating images containing the edited text. We use an off-the-shelf scene text recognition (STR) model [3] to identify the rendered text and then evaluate word-level correctness using sequence accuracy (SeqAcc) by comparing the STR result with the ground truth.", + "bbox": [ + 89, + 686, + 482, + 853 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Image quality: we utilize Fréchet Inception Distance (FID) [7] to quantify the distance between the text images in the dataset and the generated images. Additionally, the Learned", + "bbox": [ + 89, + 854, + 483, + 900 + ], + "page_idx": 6 + }, + { + "type": "table", + "img_path": "images/bf3818e8e2f7df03057767a530d4af9d31480bace7282be36512207d36d8dffe.jpg", + "table_caption": [ + "Table 2. Accumulated results of losses." + ], + "table_footnote": [], + "table_body": "
SettingAverage SeqAccFIDLPIPS
ReconEditing
Base0.2180.06026.780.0696
+Mask0.4250.25923.210.0528
+Attn0.6980.53219.720.0483
+Align0.8840.80115.410.0392
+Id0.9400.88712.130.0328
", + "bbox": [ + 516, + 393, + 903, + 502 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Perceptual Image Patch Similarity (LPIPS) [28] is utilized as an additional metric.", + "bbox": [ + 511, + 517, + 906, + 547 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "4.2. Results", + "text_level": 1, + "bbox": [ + 511, + 559, + 607, + 574 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Quantitatively: In Tab. 6, we present a quantitative analysis comparing DREAMTEXT against baseline methods. Our method demonstrates a significant advantage across all quantitative metrics, highlighting the exceptional visual quality of the scene text images it generates. Notably, when compared to the previous state-of-the-art model UDiffText, DREAMTEXT outperforms it by 3.66 in terms of FID. While UDiffText and AnyText excel in image quality compared to other methods, except DREAMTEXT, they tend to perform less satisfactorily in terms of sequence accuracy, especially in text editing. This discrepancy may be attributed to its restricted representation domain and deflected attention, resulting in less accurate text rendering within complex scenes.", + "bbox": [ + 511, + 582, + 906, + 777 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Qualitatively: As depicted in Fig. 6, our method outperforms other baselines in synthesizing more coherent and accurate text within polystylistic scenes. For example, in Fig. 6 (3rd row), MOSTEL and Stable Diffusion fail to generate the text, while TextDiffuser and UDiffText encounter the issue of character absence. Similarly, in Fig. 6 (1st row), all baselines fail to generate the text on the provided poster. Moreover, we further compare our method with the latest ap", + "bbox": [ + 511, + 779, + 908, + 900 + ], + "page_idx": 6 + }, + { + "type": "page_number", + "text": "28561", + "bbox": [ + 478, + 944, + 517, + 955 + ], + "page_idx": 6 + }, + { + "type": "table", + "img_path": "images/78aef0913cf99b00fc5319c19312619df5671a706a1991fc3d74f43e673e9af8.jpg", + "table_caption": [ + "Table 3. Hyperparameter analysis." + ], + "table_footnote": [], + "table_body": "
α\n(β=0)Average SeqAccβ\n(α=0.01)Average SeqAcc
ReconEditingReconEditing
10.6720.55810.8640.766
0.10.7240.6150.10.8960.817
0.010.7480.6230.010.9210.864
0.0010.7530.6040.0010.9400.887
inputstep 0step 0.2Tstep 0.4Tstep 0.8Tstep T
SUNSHINE\nSTATE\nOF MINDSUNSHINE\nSTATE\nOF MINDSUNSHINE\nSTATE\nOF MINDSUNSHINE\nSTATE\nOF MINDSUNSHINE\nSTATE\nOF MINDSUNSHINE\nSTATE\nOF MIND
The Bride's CrewThe Bride's\nCrewThe Bride's\nCrewThe Bride's\nCrewThe Bride's\nCrewThe Bride's\nCrew
WILBUR\nSMITHWILBUR\nSMITHWILBUR\nSMITHWILBUR\nSMITHWILBUR\nSMITHWILBUR\nSMITH
BLUE\nHORIZONBLUE\nHORIZONBLUE\nHORIZONBLUE\nHORIZONBLUE\nHORIZONBLUE\nHORIZON
", + "bbox": [ + 91, + 116, + 478, + 369 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Figure 8. The visualized attention results of all characters across several steps during training.", + "bbox": [ + 89, + 378, + 482, + 407 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "proach, UDiffText, using the samples provided in its original paper for additional qualitative evaluation. The visualization results are presented in Fig. 7, offering additional evidence of our method's superiority in generating high-fidelity text.", + "bbox": [ + 89, + 425, + 483, + 488 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "4.3. Ablation Studies and Effectiveness Analysis", + "text_level": 1, + "bbox": [ + 89, + 498, + 460, + 513 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Ablation on Losses: We employ SD-v2.0 inpainting version as the base method. As shown in Tab. 2, we observe a significant drop in performance across all metrics when these losses are removed. Additionally, these results highlight that the masked diffusion loss and cross-attention loss alone are insufficient for achieving satisfactory character representation. This may be attributed to the inherent noise in the latent character masks. To sum up, these ablation results underscore the effectiveness of each objective within our heuristic alternate optimization strategy in improving various aspects of synthesized scene texts.", + "bbox": [ + 88, + 521, + 482, + 686 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Choice of $\\alpha$ and $\\beta$ : In our experiments, we vary the values of $\\alpha$ and $\\beta$ within the range [0.001, 0.01, 0.1, 1]. The results are presented in Tab. 3. Initially, we focus on exploring the impact of $\\alpha$ while setting $\\beta = 0$ . We find that the optimal performance in text reconstruction occurs when $\\alpha = 0.001$ . However, this choice yields poor results in editing tasks. Consequently, we opt for a more balanced approach, selecting $\\alpha = 0.01$ . Subsequently, with $\\alpha$ fixed at 0.01, we investigate the effect of $\\beta$ . We find that the optimal performance is achieved when $\\beta = 0.001$ .", + "bbox": [ + 88, + 688, + 482, + 838 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Effectiveness of Heuristic Alternate Optimization: We analyze the efficacy of our strategy by visualizing the evolution of character attention maps across several training steps. As illustrated in Fig. 8, the model exhibits deflected attention", + "bbox": [ + 89, + 839, + 483, + 900 + ], + "page_idx": 7 + }, + { + "type": "table", + "img_path": "images/945091872fa1deea466d061cc9f744752a461642f324cb20c7da21cc48629947.jpg", + "table_caption": [ + "Table 4. Ablation study results on warm-up steps." + ], + "table_footnote": [], + "table_body": "
Average SeqAcc
Recon ↑Editing ↑FID ↓mIoU ↑
15k0.8840.85213.820.681
20k0.9130.87313.380.692
25k0.9400.88712.130.722
30k0.9210.89113.240.703
", + "bbox": [ + 517, + 116, + 903, + 210 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "initially but progressively corrects it in subsequent training steps. Ultimately, attention becomes concentrated on the desired positions for all characters, indicating the effectiveness of our approach. Besides, Fig. 4 also illustrates the efficacy of this strategy in continuously improving the generator's ability to estimate the optimal characters' position by autonomously rectifying the attention during training.", + "bbox": [ + 511, + 222, + 906, + 327 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Choice of Warm-up Steps: We conduct an ablation experiment to analyze the impact of different warm-up steps on model performance. The results, presented in Tab. 4, demonstrate how varying the number of warm-up steps affects the performance. As the number of warm-up steps increases from 15k to 25k, we observe consistent improvements across all metrics. However, at 30k steps, while the editing accuracy slightly improves, the FID score worsens, suggesting a potential trade-off between image fidelity and editing capability. Overall, these findings highlight that 25k warm-up steps strike the best balance between reconstruction accuracy, editing performance, and image quality, making it the optimal choice for our method.", + "bbox": [ + 511, + 329, + 908, + 525 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "5. Conclusion", + "text_level": 1, + "bbox": [ + 513, + 539, + 633, + 556 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "This paper presents DREAMTEXT, a novel approach for high-fidelity scene text synthesis. We reconstruct the diffusion training process to expose and rectify the model's attention at the character level while strengthening its learning of text regions. This inevitably involves a hybrid optimization challenge, combining discrete and continuous variables, which we effectively address through our heuristic alternate optimization strategy. Meanwhile, we jointly train the text encoder and generator to comprehensively learn and utilize diverse fonts present in the training dataset, fostering a synergistic relationship between learning character embeddings and re-estimating character attention. Both qualitative and quantitative results highlight the superiority of our method.", + "bbox": [ + 511, + 566, + 908, + 763 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "6. Acknowledgments", + "text_level": 1, + "bbox": [ + 511, + 777, + 692, + 795 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "This work was supported by the National Nature Science Foundation of China (62472097), AI for Science Foundation of Fudan University (FudanX24AI028), and Shanghai Innovation Institute. The computations in this research were performed on the CFFF platform of Fudan University. Cheng Jin is the corresponding author of this work.", + "bbox": [ + 511, + 801, + 908, + 898 + ], + "page_idx": 7 + }, + { + "type": "page_number", + "text": "28562", + "bbox": [ + 478, + 944, + 519, + 957 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "References", + "text_level": 1, + "bbox": [ + 91, + 90, + 187, + 104 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[1] Rowel Atienza. Vision transformer for fast and efficient scene text recognition. In ICDAR, pages 319-334, 2021. 7", + "[2] Omri Avrahami, Kfir Aberman, Ohad Fried, Daniel Cohen-Or, and Dani Lischinski. Break-a-scene: Extracting multiple concepts from a single image. In SIGGRAPH Asia, pages 1–12, 2023. 5", + "[3] Darwin Bautista and Rowel Atienza. Scene text recognition with permuted autoregressive sequence models. In ECCV, pages 178-196, 2022. 7", + "[4] Haoxing Chen, Zhuoer Xu, Zhangxuan Gu, Yaohui Li, Changhua Meng, Huijia Zhu, Weiqiang Wang, et al. Diffuse: Universal text editing diffusion model. In NeurIPS, 2024. 3", + "[5] Jingye Chen, Yupan Huang, Tengchao Lv, Lei Cui, Qifeng Chen, and Furu Wei. Textdiffuser: Diffusion models as text painters. In NeurIPS, 2023. 2, 3, 6", + "[6] Ankush Gupta, Andrea Vedaldi, and Andrew Zisserman. Synthetic data for text localisation in natural images. In CVPR, pages 2315-2324, 2016. 6", + "[7] Martin Heusel, Hubert Ramsauer, Thomas Unterthiner, Bernhard Nessler, and Sepp Hochreiter. Gans trained by a two time-scale update rule converge to a local nash equilibrium. In NeurIPS, 2017. 7", + "[8] Jiabao Ji, Guanhua Zhang, Zhaowen Wang, Bairu Hou, Zhifei Zhang, Brian Price, and Shiyu Chang. Improving diffusion models for scene text editing with dual encoders. arXiv preprint arXiv:2304.05568, 2023. 2, 3, 6", + "[9] Dimosthenis Karatzas, Faisal Shafait, Seiichi Uchida, Masakazu Iwamura, Lluis Gomez i Bigorda, Sergi Robles Mestre, Joan Mas, David Fernandez Mota, Jon Almazan Almazan, and Lluis Pere De Las Heras. Icdar 2013 robust reading competition. In ICDAR, pages 1484-1493, 2013. 6", + "[10] Yuxin Kong, Canjie Luo, Weihong Ma, Qiyuan Zhu, Shenggao Zhu, Nicholas Yuan, and Lianwen Jin. Look closer to supervise better: One-shot font generation via component-based discriminator. In CVPR, pages 13482-13491, 2022. 3", + "[11] Sanyam Lakhanpal, Shivang Chopra, Vinija Jain, Aman Chadha, and Man Luo. Refining text-to-image generation: Towards accurate training-free glyph-enhanced image generation. arXiv preprint arXiv:2403.16422, 2024. 2", + "[12] Zeyu Liu, Weicong Liang, Zhanhao Liang, Chong Luo, Ji Li, Gao Huang, and Yuhui Yuan. Glyph-byt5: A customized text encoder for accurate visual text rendering. arXiv preprint arXiv:2403.09622, 2024. 2", + "[13] Song Park, Sanghyuk Chun, Junbum Cha, Bado Lee, and Hyunjung Shim. Multiple heads are better than one: Few-shot font generation with multiple localized experts. In ICCV, pages 13900-13909, 2021. 1", + "[14] Yadong Qu, Qingfeng Tan, Hongtao Xie, Jianjun Xu, Yuxin Wang, and Yongdong Zhang. Exploring stroke-level modifications for scene text editing. In AAAI, pages 2119-2127, 2023. 1, 3, 6", + "[15] Robin Rombach, Andreas Blattmann, Dominik Lorenz, Patrick Esser, and Björn Ommer. High-resolution image" + ], + "bbox": [ + 93, + 114, + 483, + 900 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "synthesis with latent diffusion models. In CVPR, pages 10684-10695, 2022. 4, 6", + "[16] Prasun Roy, Saumik Bhattacharya, Subhankar Ghosh, and Umapada Pal. Stefann: scene text editor using font adaptive neural network. In CVPR, pages 13228-13237, 2020. 1, 3", + "[17] Chitwan Sahara, William Chan, Saurabh Saxena, Lala Li, Jay Whang, Emily L Denton, Kamyar Ghasemipour, Raphael Gontijo Lopes, Burcu Karagol Ayan, Tim Salimans, et al. Photorealistic text-to-image diffusion models with deep language understanding. In NeurIPS, pages 36479-36494, 2022. 2", + "[18] Wataru Shimoda, Daichi Haraguchi, Seiichi Uchida, and Kota Yamaguchi. De-rendering stylized texts. In ICCV, pages 1076-1085, 2021. 3", + "[19] Yuxiang Tuo, Wangmeng Xiang, Jun-Yan He, Yifeng Geng, and Xuansong Xie. Anytext: Multilingual visual text generation and editing. ICLR, 2024. 6", + "[20] Yibin Wang, Yuchao Feng, Jie Wu, Honghui Xu, and Jianwei Zheng. Ca-gan: Object placement via coalescing attention based generative adversarial network. In ICME, pages 2375-2380. IEEE, 2023. 3", + "[21] Yibin Wang, Honghui Xu, Shanghai Zhou, Weizhong Zhang, and Cheng Jin. Enhancing object coherence in layout-to-image synthesis. arXiv preprint arXiv:2311.10522, 2023. 2", + "[22] Yibin Wang, Weizhong Zhang, and Cheng Jin. Magicface: Training-free universal-style human image customized synthesis. arXiv preprint arXiv:2408.07433, 2024. 2", + "[23] Yibin Wang, Weizhong Zhang, Jianwei Zheng, and Cheng Jin. High-fidelity person-centric subject-to-image synthesis. In CVPR, 2024. 2", + "[24] Yibin Wang, Weizhong Zhang, Jianwei Zheng, and Cheng Jin. Primecomposer: Faster progressively combined diffusion for image composition with attention steering. arXiv preprint arXiv:2403.05053, 2024. 2", + "[25] Liang Wu, Chengquan Zhang, Jiaming Liu, Junyu Han, Jingtuo Liu, Errui Ding, and Xiang Bai. Editing text in the wild. In ACM MM, pages 1500–1508, 2019. 3", + "[26] Xingqian Xu, Zhifei Zhang, Zhaowen Wang, Brian Price, Zhonghao Wang, and Humphrey Shi. Rethinking text segmentation: A novel dataset and a text-specific refinement approach. In CVPR, pages 12045-12055, 2021. 7", + "[27] Qiangpeng Yang, Jun Huang, and Wei Lin. Swaptext: Image based texts transfer in scenes. In CVPR, pages 14700-14709, 2020. 1, 3", + "[28] Richard Zhang, Phillip Isola, Alexei A Efros, Eli Shechtman, and Oliver Wang. The unreasonable effectiveness of deep features as a perceptual metric. In CVPR, pages 586-595, 2018. 7", + "[29] Yiming Zhao and Zhouhui Lian. Udifftext: A unified framework for high-quality text synthesis in arbitrary images via character-aware diffusion models. In ECCV, 2024. 2, 3, 4, 5, 6, 7" + ], + "bbox": [ + 516, + 92, + 906, + 840 + ], + "page_idx": 8 + }, + { + "type": "page_number", + "text": "28563", + "bbox": [ + 478, + 944, + 517, + 955 + ], + "page_idx": 8 + } +] \ No newline at end of file diff --git a/2025/DreamText_ High Fidelity Scene Text Synthesis/6bc8ea68-7656-478a-8d11-90857d6ca90e_model.json b/2025/DreamText_ High Fidelity Scene Text Synthesis/6bc8ea68-7656-478a-8d11-90857d6ca90e_model.json new file mode 100644 index 0000000000000000000000000000000000000000..bb68d06aac65c054b06b8d0186815643e3c0494d --- /dev/null +++ b/2025/DreamText_ High Fidelity Scene Text Synthesis/6bc8ea68-7656-478a-8d11-90857d6ca90e_model.json @@ -0,0 +1,1857 @@ +[ + [ + { + "type": "header", + "bbox": [ + 0.107, + 0.003, + 0.182, + 0.044 + ], + "angle": 0, + "content": "CVF" + }, + { + "type": "header", + "bbox": [ + 0.237, + 0.001, + 0.812, + 0.047 + ], + "angle": 0, + "content": "This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore." + }, + { + "type": "title", + "bbox": [ + 0.247, + 0.131, + 0.753, + 0.154 + ], + "angle": 0, + "content": "DREAMTEXT: High Fidelity Scene Text Synthesis" + }, + { + "type": "text", + "bbox": [ + 0.238, + 0.179, + 0.767, + 0.199 + ], + "angle": 0, + "content": "Yibin Wang\\(^{1,2}\\), Weizhong Zhang\\(^{1,4}\\), Honghui Xu\\(^{5}\\), Cheng Jin\\(^{1,3\\dagger}\\)" + }, + { + "type": "text", + "bbox": [ + 0.305, + 0.199, + 0.694, + 0.217 + ], + "angle": 0, + "content": "\\(^{1}\\)Fudan University \\(^{2}\\)Shanghai Innovation Institute" + }, + { + "type": "text", + "bbox": [ + 0.202, + 0.216, + 0.796, + 0.235 + ], + "angle": 0, + "content": "\\(^{3}\\)Innovation Center of Calligraphy and Painting Creation Technology, MCT" + }, + { + "type": "text", + "bbox": [ + 0.104, + 0.234, + 0.896, + 0.253 + ], + "angle": 0, + "content": "\\(^{4}\\)Shanghai Key Laboratory of Intelligent Information Processing \\(^{5}\\)Zhejiang University of Technology" + }, + { + "type": "text", + "bbox": [ + 0.13, + 0.255, + 0.867, + 0.27 + ], + "angle": 0, + "content": "yibinwang1121@163.com, weizhongzhang@fudan.edu.cn, xhh@zjut.edu.cn, jc@fudan.edu.cn" + }, + { + "type": "image", + "bbox": [ + 0.11, + 0.306, + 0.885, + 0.51 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.157, + 0.521, + 0.84, + 0.536 + ], + "angle": 0, + "content": "Figure 1. Displayed are the results generated using our DREAMTEXT, showcasing its prowess across varied inputs." + }, + { + "type": "title", + "bbox": [ + 0.248, + 0.548, + 0.327, + 0.563 + ], + "angle": 0, + "content": "Abstract" + }, + { + "type": "text", + "bbox": [ + 0.089, + 0.582, + 0.486, + 0.87 + ], + "angle": 0, + "content": "Scene text synthesis involves rendering specified texts onto arbitrary images. Current methods typically formulate this task in an end-to-end manner but lack effective character-level guidance during training. Besides, their text encoders, pre-trained on a single font type, struggle to adapt to the diverse font styles encountered in practical applications. Consequently, these methods suffer from character distortion, repetition, and absence, particularly in polystylistic scenarios. To this end, this paper proposes DREAMTEXT for high-fidelity scene text synthesis. Our key idea is to reconstruct the diffusion training process, introducing more refined guidance tailored to this task, to expose and rectify the model's attention at the character level and strengthen its learning of text regions. This transformation poses a hybrid optimization challenge, involving both discrete and continuous variables. To effectively tackle this challenge, we employ a heuristic alternate optimization strategy. Meanwhile, we jointly train the text encoder and generator to comprehensively learn and utilize the diverse font present" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.55, + 0.909, + 0.732 + ], + "angle": 0, + "content": "in the training dataset. This joint training is seamlessly integrated into the alternate optimization process, fostering a synergistic relationship between learning character embedding and re-estimating character attention. Specifically, in each step, we first encode potential character-generated position information from cross-attention maps into latent character masks. These masks are then utilized to update the representation of specific characters in the current step, which, in turn, enables the generator to correct the character's attention in the subsequent steps. Both qualitative and quantitative results demonstrate the superiority of our method to the state of the art. Our project page is here." + }, + { + "type": "title", + "bbox": [ + 0.514, + 0.767, + 0.646, + 0.783 + ], + "angle": 0, + "content": "1. Introduction" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.795, + 0.909, + 0.903 + ], + "angle": 0, + "content": "Scene text synthesis involves modifying or inserting specified text in arbitrary images while maintaining its natural and realistic appearance. Earlier GAN-based style transfer methods [13, 14, 16, 27] accomplish this by transferring the text style from a reference image to the rendered target text image. However, they are constrained in their capacity to generate text in arbitrary styles, such as font and color. These" + }, + { + "type": "page_footnote", + "bbox": [ + 0.094, + 0.887, + 0.221, + 0.901 + ], + "angle": 0, + "content": "† Corresponding author." + }, + { + "type": "page_number", + "bbox": [ + 0.479, + 0.945, + 0.52, + 0.958 + ], + "angle": 0, + "content": "28555" + } + ], + [ + { + "type": "image_caption", + "bbox": [ + 0.322, + 0.089, + 0.674, + 0.102 + ], + "angle": 0, + "content": "Problem 1: Character repetition and absence" + }, + { + "type": "image", + "bbox": [ + 0.136, + 0.108, + 0.865, + 0.225 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.372, + 0.226, + 0.626, + 0.239 + ], + "angle": 0, + "content": "Problem 2: Character distortion" + }, + { + "type": "image", + "bbox": [ + 0.136, + 0.245, + 0.865, + 0.361 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.094, + 0.372, + 0.9, + 0.386 + ], + "angle": 0, + "content": "Figure 2. Current methods encounter significant challenges, i.e., character repetition and absence (top) and character distortion (bottom)." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.399, + 0.484, + 0.534 + ], + "angle": 0, + "content": "limitations are effectively addressed by utilizing the diffusion model [5, 8, 11, 12, 17, 21, 22], harnessing its inherent prior knowledge and robust capabilities. Nevertheless, these methods still encounter challenges in generating accurate text within polystylistic images due to the inadequate conditional guidance provided by their character-unaware text encoder. To this end, a recent study [29] proposes a character-level text encoder, providing more robust conditional guidance for the diffusion model." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.539, + 0.484, + 0.902 + ], + "angle": 0, + "content": "Despite their effectiveness, these methods are still difficult to accurately render text within complex scenes. Specifically, these methods such as UDiffText [29] and TextDiffuser [5] solely utilize a single font type to pre-train the text encoder, which is subsequently employed to fine-tune the generator. However, the diverse font styles encountered in practical applications present a significant challenge: the restricted representation domain hampers their ability to render text with unseen font styles, e.g., cases shown in Fig. 2 (bottom). Furthermore, these methods rely on character segmentation masks to supervise characters' attention for position control. We argue this approach has significant limitations: (1) unlike classification tasks with uniquely determined labels, the optimal generation positions for characters can be diverse. Therefore, using the specific masks to rigidly constrain the model may limit its flexibility in estimating optimal positions, hindering its ability to adapt to varied and complex scenarios. Besides, (2) these masks generated by a pre-trained segmentation model [5] are imprecise, tending to over-segment character regions. As a result, the model encounters issues of character repetition and absence as illustrated in Fig. 2 (top left) and (top right) respectively. We suspect these issues stem from characters' attention maps, as they reflect the underlying response to the character's" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.399, + 0.908, + 0.685 + ], + "angle": 0, + "content": "final generated position [23, 24]. Therefore, we visualize the attention maps of their problematic results by rendering attention directly onto the images. We observe certain characters' attention misalignment, resulting in character duplication (see Fig. 3 (a) and (c)). Besides, excessive dispersion of attention also leads to character absence (see Fig. 3 (b)). This observation underscores the insufficiency of these methods in guiding the model autonomously estimating optimal character-generated position. Further, we also assess specific methods [5, 29] quantitatively on two datasets to evaluate their proficiency in directing characters' attention toward their optimal generation regions throughout multiple global training steps. Specifically, for each test sample, we extract potential position masks for all characters based on their cross-attention maps from the inference process. Then, we compute the mean Intersection over Union (mIoU) between these masks and ground truth character segmentation masks. Their inoptimal results depicted in Fig. 4 further corroborate our hypothesis." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.69, + 0.909, + 0.901 + ], + "angle": 0, + "content": "In summary, current methods encounter challenges in accurately rendering text within complex scenes due to their constrained representation domain and synthesizing characters in their ideal regions due to the lack of effective guidance for the model to estimate optimal character-generated position. To this end, this paper proposes DREAMTEXT for high-fidelity text synthesis. We reconstruct the diffusion training process, embedding more refined guidance: we expose and rectify the characters' attention to focus more precisely on their ideal generation areas, and enhance the learning of text regions and character representation by additional constraints. However, this reconstruction inevitably introduces a complex hybrid optimization problem encompassing both discrete and continuous variables. Therefore," + }, + { + "type": "page_number", + "bbox": [ + 0.48, + 0.946, + 0.519, + 0.957 + ], + "angle": 0, + "content": "28556" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.094, + 0.089, + 0.482, + 0.392 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.09, + 0.403, + 0.483, + 0.434 + ], + "angle": 0, + "content": "Figure 3. The problematic results rendered by characters' attention maps (AMs)." + }, + { + "type": "text", + "bbox": [ + 0.089, + 0.447, + 0.483, + 0.703 + ], + "angle": 0, + "content": "we design a heuristic alternate optimization strategy. Concurrently, we jointly train the text encoder and the generator, leveraging the diverse font styles within the training dataset to enrich character representation space. We demonstrate that this joint training can be seamlessly integrated into the heuristic alternate optimization process, facilitating a synergistic interplay between character representations learning and character attention re-estimation. Specifically, in each iteration, the pipeline begins by encoding the potential character-generated position information from cross-attention maps into latent character masks. These masks are subsequently employed to refine the representation of specific characters within the text encoder, facilitating the calibration of the characters' attention in the subsequent step. This iterative process contributes to both learning of better character representation and, as a by-product, explicit guidance for autonomously estimating character position." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.705, + 0.486, + 0.902 + ], + "angle": 0, + "content": "Notably, the generator may initially struggle to direct attention to the desired generation region for each character, resulting in suboptimal latent masks and impacting training progress. Unlike existing methods [5, 29] that rigidly constrain character attention, we employ a balanced supervision strategy. Initially, we assist in calibrating the attention using character segmentation masks for a warm-up in the earlier stage. Once the model has preliminarily acquired the ability to estimate ideal generation positions, we remove this guidance, allowing the model to autonomous learning iteratively. This strategy strikes a balance between constraining the model and unleashing its flexibility in estimating optimal generation positions. Fig. 4 illustrates the efficacy of our" + }, + { + "type": "image", + "bbox": [ + 0.517, + 0.088, + 0.905, + 0.183 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.513, + 0.192, + 0.908, + 0.263 + ], + "angle": 0, + "content": "Figure 4. The mIoU scores of UDiffText, TextDiffuser, and our method on LAION-OCR and SynthText over global training steps. Our method adopts a balanced supervision strategy: we initially use latent character masks to steer the character's attention for a warm-up in the earlier stage and stop guiding after 25,000 steps." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.272, + 0.907, + 0.302 + ], + "angle": 0, + "content": "method in continuously improving the generator's ability to estimate the optimal characters' position." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.303, + 0.909, + 0.468 + ], + "angle": 0, + "content": "Our contributions are: (1) The proposed DREAMTEXT effectively alleviates the issues of character repetition, absence, and distortion encountered by existing methods. (2) Our heuristic alternate optimization strategy, integrating the joint learning of the text encoder and U-Net, orchestrates a symbiotic relationship between learning character representations and re-estimating character attention. (3) Our balanced supervision strategy strikes a balance between constraining the model and unleashing its flexibility in estimating optimal generation positions. (4) Both qualitative and quantitative results demonstrate the superiority of our method." + }, + { + "type": "title", + "bbox": [ + 0.514, + 0.482, + 0.655, + 0.499 + ], + "angle": 0, + "content": "2. Related Work" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.508, + 0.909, + 0.689 + ], + "angle": 0, + "content": "Earlier GAN-based methods achieve scene text synthesis by transferring the text style in a reference image to the rendered target text image [10, 16, 18, 20, 27]. Specifically, STEFANN [16] employs a FANnet to edit individual characters and incorporates a placement algorithm to generate the desired word. SRNet [25] and MOSTEL [14], on the other hand, split the task into two main parts: background inpainting and text style transfer. This approach allows for the end-to-end synthesizing of entire words. Despite their simplicity and effectiveness, they are limited in their ability to generate text in arbitrary styles and locations, often producing less natural-looking images." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.69, + 0.911, + 0.901 + ], + "angle": 0, + "content": "To this end, several diffusion-based approaches have emerged to address these challenges. These approaches harness the robust capabilities of diffusion models to edit or generate scene text, thereby enhancing the quality and diversity of the generated content. DiffSTE [8] proposes a dual encoder structure, comprising a character text encoder and an instruction text encoder. These components are employed for instruction tuning, providing improved control over the backbone network. DiffUTE [4] employs an OCR-based glyph encoder to extract glyph guidance from the rendered glyph image. TextDiffuser [5] concatenates the segmentation mask as conditional input and employs character-aware loss to control the generated characters precisely. However, they still face challenges in generating accurate text" + }, + { + "type": "page_number", + "bbox": [ + 0.479, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "28557" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.094, + 0.089, + 0.91, + 0.391 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.273, + 0.399, + 0.725, + 0.415 + ], + "angle": 0, + "content": "Figure 5. An overview of proposed heuristic alternate optimization strategy." + }, + { + "type": "text", + "bbox": [ + 0.089, + 0.425, + 0.483, + 0.531 + ], + "angle": 0, + "content": "within polystylistic images, primarily due to the insufficient conditional guidance provided by their character-unaware text encoder. Motivated by this, UDiffText [29] replaces the original CLIP text encoder in Stable Diffusion with a character-level text encoder with a unitary font. It also employs local attention loss for position control, relying on ground truth character segmentation maps." + }, + { + "type": "text", + "bbox": [ + 0.089, + 0.531, + 0.486, + 0.624 + ], + "angle": 0, + "content": "However, these methods still encounter challenges in accurately rendering text within complex scenes and synthesizing characters in their ideal regions, attributed to their constrained character representation domain and lack of effective guidance for the model to autonomously estimate character-generated position, respectively." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.635, + 0.184, + 0.65 + ], + "angle": 0, + "content": "3. Method" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.66, + 0.484, + 0.752 + ], + "angle": 0, + "content": "Given an input image \\( z \\), a text region specified by a binary mask \\( B \\) and a text condition \\( c \\), scene text synthesis is designed to render the given text onto the input image at the given text region. A natural idea for scene text synthesis is to formulate it into a latent diffusion process [15], which is equivalent to solving the following optimization problem:" + }, + { + "type": "equation", + "bbox": [ + 0.095, + 0.76, + 0.48, + 0.788 + ], + "angle": 0, + "content": "\\[\n\\min _ {(\\theta , \\vartheta)} \\mathcal {L} _ {L D M} \\triangleq \\mathbb {E} _ {\\boldsymbol {z}, c, \\epsilon \\sim N (0, 1), t} \\| \\boldsymbol {\\epsilon} - \\epsilon_ {\\theta} (\\boldsymbol {z} _ {t}, t, \\psi_ {\\vartheta} (c), \\boldsymbol {B}) \\| _ {2} ^ {2},\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.795, + 0.486, + 0.856 + ], + "angle": 0, + "content": "where \\(\\psi_{\\vartheta}\\) is the learnable character-level text encoder initialized by UDifftext [29] and \\(\\theta\\) is the model parameter vector. However, we argue that such a straightforward formulation could be problematic due to the following reasons:" + }, + { + "type": "text", + "bbox": [ + 0.091, + 0.856, + 0.486, + 0.902 + ], + "angle": 0, + "content": "- Without explicit guidance for rectifying characters' attention, character repetition and absence can inevitably occur in the rendered results as shown in Fig. 2." + }, + { + "type": "text", + "bbox": [ + 0.514, + 0.425, + 0.907, + 0.501 + ], + "angle": 0, + "content": "- This formulation uniformly measures the mean distance between all pixels, without emphasizing the text regions. Consequently, text-unrelated regions can adversely affect the learning of the text encoder during backpropagation. The experimental analysis is provided in Sec. 4.3." + }, + { + "type": "text", + "bbox": [ + 0.514, + 0.501, + 0.906, + 0.562 + ], + "angle": 0, + "content": "- Due to the intricate interactions among the U-Net and text encoder, end-to-end training is particularly challenging as shown in Sec. 4.3. Therefore, a more refined guidance for the training process is necessary." + }, + { + "type": "list", + "bbox": [ + 0.514, + 0.425, + 0.907, + 0.562 + ], + "angle": 0, + "content": null + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.57, + 0.907, + 0.751 + ], + "angle": 0, + "content": "In this paper, we propose DREAMTEXT and develop a heuristic iterative optimization pipeline to enable efficient training. Our key idea is to explicitly model the potential character positions in attention maps by encoding them into latent character masks in each step. These masks are then used to enhance the learning of the text encoder through our carefully crafted losses, facilitating the calibration of character attention in the subsequent step. This heuristic iterative process allows our model to dynamically alternate between optimizing character embedding and re-estimating character generation positions. The details are presented in the following sections." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.784, + 0.857, + 0.799 + ], + "angle": 0, + "content": "3.1. Latent Diffusion with Refined Guidance" + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.814, + 0.725, + 0.828 + ], + "angle": 0, + "content": "3.1.1. Latent Character Mask" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.841, + 0.907, + 0.903 + ], + "angle": 0, + "content": "We utilize cross-attention maps to generate latent character masks, extracting characters' latent position information in the current step. Given latent image \\( z_{t} \\) and text embedding \\( \\psi_{\\vartheta}(c) \\), the cross-attention map in layer \\( l \\) is defined as follows:" + }, + { + "type": "page_number", + "bbox": [ + 0.479, + 0.945, + 0.521, + 0.957 + ], + "angle": 0, + "content": "28558" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.134, + 0.089, + 0.865, + 0.33 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.279, + 0.338, + 0.719, + 0.354 + ], + "angle": 0, + "content": "Figure 6. Qualitative comparative results against state-of-the-art methods." + }, + { + "type": "equation", + "bbox": [ + 0.182, + 0.39, + 0.484, + 0.409 + ], + "angle": 0, + "content": "\\[\n\\boldsymbol {Q} _ {l} = \\boldsymbol {z} _ {t} \\boldsymbol {W} _ {l} ^ {q}, \\quad \\boldsymbol {K} _ {l} = \\psi_ {\\vartheta} (c) \\boldsymbol {W} _ {l} ^ {k}, \\tag {1}\n\\]" + }, + { + "type": "equation", + "bbox": [ + 0.214, + 0.411, + 0.484, + 0.448 + ], + "angle": 0, + "content": "\\[\n\\boldsymbol {A} _ {l} = \\operatorname {s o f t m a x} \\left(\\frac {\\boldsymbol {Q} _ {l} \\boldsymbol {K} _ {l} ^ {T}}{\\sqrt {d}}\\right), \\tag {2}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.458, + 0.484, + 0.577 + ], + "angle": 0, + "content": "where \\( \\pmb{W}_l^q \\) and \\( \\pmb{W}_l^k \\) are learnable parameters and \\( d \\) is the embedding dimension. The attention maps \\( \\pmb{A}_l \\) is reshaped to \\( N \\times H \\times W \\), where \\( N \\) represents the number of text tokens such that each slice \\( \\pmb{A}_l^i \\in \\mathbb{R}^{H \\times W} \\) represents the region attended by the \\( i \\)-th token. These attention maps are then averaged across all layers of the U-Net to get the mean response \\( \\bar{\\pmb{A}} = \\frac{1}{L} \\sum_{l=1}^{L} \\pmb{A}_l \\) for the regions attended by character tokens." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.578, + 0.484, + 0.761 + ], + "angle": 0, + "content": "To obtain the latent character masks \\(M \\in \\mathbb{R}^{N \\times H \\times W}\\), we first apply Gaussian blur to the attention map to execute low-pass filtering \\(\\mathrm{blur}(\\bar{A})\\). This step helps mitigate excessive variance in the attended regions, ensuring a more uniform distribution of attention across relevant character regions. Subsequently, we employ a straightforward thresholding process to convert the blurred attention map into binary character masks. This process is encapsulated by the function \\(f(\\cdot)\\), which applies a threshold and assigns a value of 1 to pixel values exceeding the threshold, and 0 otherwise. The threshold is calculated as the mean value plus twice the variance of averaged attention maps in this work. That is:" + }, + { + "type": "equation", + "bbox": [ + 0.132, + 0.783, + 0.484, + 0.824 + ], + "angle": 0, + "content": "\\[\nf (\\boldsymbol {X}) = \\left\\{ \\begin{array}{l} 1, \\text {i f} x _ {i, j} > \\operatorname {m e a n} (\\boldsymbol {X}) + 2 \\operatorname {s t d} (\\boldsymbol {X}) \\\\ 0, \\text {o t h e r w i s e} \\end{array} \\right. \\tag {3}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.091, + 0.833, + 0.485, + 0.849 + ], + "angle": 0, + "content": "Therefore, the equation for obtaining \\(M\\) can be expressed as:" + }, + { + "type": "equation", + "bbox": [ + 0.225, + 0.863, + 0.484, + 0.881 + ], + "angle": 0, + "content": "\\[\n\\boldsymbol {M} = f (\\operatorname {b l u r} (\\bar {\\boldsymbol {A}})). \\tag {4}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.111, + 0.886, + 0.483, + 0.902 + ], + "angle": 0, + "content": "Based on the latent character masks \\(M\\), we employ several" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.365, + 0.907, + 0.396 + ], + "angle": 0, + "content": "loss functions to optimize both the text encoder and U-Net iteratively. We will delve into these in the following sections." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.402, + 0.72, + 0.416 + ], + "angle": 0, + "content": "3.1.2. Masked Diffusion Loss" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.421, + 0.907, + 0.528 + ], + "angle": 0, + "content": "We first extend the Eq. 1 to highlight all desired character tokens in the current step. Specifically, within a text \\(c\\) containing \\(k\\) tokens of interest, the diffusion loss of the corresponding pixels obtained from each token's latent character mask is applied with an additional weighting factor \\(\\gamma\\). Formally, considering \\(M_{k} = \\bigvee_{i=1}^{k} M_{i}\\) to the union of the pixels of \\(k\\) characters, the masked diffusion loss is formulated as," + }, + { + "type": "equation", + "bbox": [ + 0.513, + 0.538, + 0.921, + 0.557 + ], + "angle": 0, + "content": "\\[\n\\mathcal {L} _ {\\text {m a s k}} = \\mathbb {E} _ {z, c, \\epsilon \\sim N (0, 1), t} \\| (1 + \\gamma \\boldsymbol {M} _ {k}) (\\boldsymbol {\\epsilon} - \\epsilon_ {\\theta} (\\boldsymbol {z} _ {t}, t, \\psi_ {\\vartheta} (c))) \\| _ {2} ^ {2}.\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.566, + 0.824, + 0.581 + ], + "angle": 0, + "content": "We omit binary mask \\(B\\) here for the simplicity." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.588, + 0.706, + 0.601 + ], + "angle": 0, + "content": "3.1.3. Cross Attention Loss" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.607, + 0.906, + 0.699 + ], + "angle": 0, + "content": "The masked diffusion loss is designed to focus on all desired concept tokens within a prompt. Additionally, to ensure that each token encodes information specific to its corresponding synthesis position, we incorporate the cross-attention loss [2]. This loss encourages tokens to attend exclusively to their corresponding target regions," + }, + { + "type": "equation", + "bbox": [ + 0.56, + 0.707, + 0.907, + 0.726 + ], + "angle": 0, + "content": "\\[\n\\mathcal {L} _ {a t t n} = \\mathbb {E} _ {\\boldsymbol {z}, c, t} \\| C _ {a t t n} (\\boldsymbol {z} _ {t}, \\psi_ {\\vartheta} (c) _ {i}) - \\boldsymbol {M} _ {i} \\| _ {2} ^ {2}. \\tag {5}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.735, + 0.907, + 0.78 + ], + "angle": 0, + "content": "Here, \\( C_{attn}(z_t, \\psi_\\vartheta(c)_i) \\) represents the cross-attention map between the visual representation \\( z_t \\) and the token \\( \\psi_\\vartheta(c)_i \\), while \\( M_i \\) denotes the latent character mask of \\( \\psi_\\vartheta(c)_i \\)." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.781, + 0.909, + 0.901 + ], + "angle": 0, + "content": "However, the losses mentioned above alone are inadequate for achieving optimal character representation. This is primarily due to the inherent noise in the latent character masks, which frequently results in under- or oversegmentation of the ideal character regions during training. To mitigate the influence of noise and prevent skewed learning of character representation, we introduce additional losses [29] as shown below." + }, + { + "type": "page_number", + "bbox": [ + 0.479, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "28559" + } + ], + [ + { + "type": "table_caption", + "bbox": [ + 0.345, + 0.09, + 0.652, + 0.104 + ], + "angle": 0, + "content": "Table 1. Quantitative comparison against baselines." + }, + { + "type": "table", + "bbox": [ + 0.094, + 0.11, + 0.905, + 0.213 + ], + "angle": 0, + "content": "
MethodsSeqAcc-ReconSeqAcc-EditingFIDLPIPS
ICDAR13(8ch)ICDAR13TextSegLAION-OCRICDAR13(8ch)ICDAR13TextSegLAION-OCR
CVPR'22 SD-Inpainting [15]0.320.290.110.150.080.070.040.0526.780.0696
arXiv'23 DiffSTE [8]0.450.370.500.410.340.290.470.2751.670.1050
AAAI'23 MOSTEL [14]0.750.680.640.710.350.280.250.4425.090.0605
NIPS'23 TextDiffuser [5]0.870.810.680.800.820.750.660.6432.250.0834
ICLR'24 AnyText [19]0.890.870.810.860.810.790.800.7222.730.0651
ECCV'24 UDiffText [29]0.940.910.930.900.840.830.840.7815.790.0564
DreamText0.950.940.960.930.870.890.910.8812.130.0328
" + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.229, + 0.321, + 0.243 + ], + "angle": 0, + "content": "3.1.4. Cross-modal Aligned Loss" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.247, + 0.486, + 0.324 + ], + "angle": 0, + "content": "To obtain accurate and robust embeddings, we introduce an image encoder \\(\\xi\\), text head \\(H_{t}\\), and visual head \\(H_{v}\\) to align cross-modal character features. Specifically, we compute the cosine similarity objective between visual and text representations to maximize the alignment between two modalities:" + }, + { + "type": "equation", + "bbox": [ + 0.17, + 0.333, + 0.484, + 0.367 + ], + "angle": 0, + "content": "\\[\n\\mathcal {L} _ {\\text {a l i g n}} = \\frac {\\left\\langle H _ {t} (\\mathbf {y}) , H _ {v} (\\xi (\\mathbf {I})) \\right\\rangle}{\\| H _ {t} (\\mathbf {y}) \\| _ {2} \\cdot \\| H _ {v} (\\xi (\\mathbf {I})) \\| _ {2}}. \\tag {6}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.369, + 0.484, + 0.444 + ], + "angle": 0, + "content": "where \\( I \\) represents the text image segmented via the bounding box from datasets and \\( \\langle ,\\rangle \\) presents the inner product of two vectors. Note that to alleviate noise stemming from background and color variations, we preprocess the text image by converting it to grayscale." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.444, + 0.484, + 0.534 + ], + "angle": 0, + "content": "Through training, the text head learns to map the whole information of multiple characters into the text feature space. Similarly, the image head maps the overall visual text information into the visual feature space. This effectively ensures the overall textual and visual representations correspond, capturing high-level cross-modal consistency." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.54, + 0.265, + 0.553 + ], + "angle": 0, + "content": "3.1.5. Character Id Loss" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.558, + 0.484, + 0.635 + ], + "angle": 0, + "content": "Additionally, we introduce character Id loss to ensure that the learned embeddings are highly distinguishable. To be precise, a multi-label classification head \\( H_{l} \\) is introduced to predict character indices from text embeddings \\( y \\). Then, the cross-entropy objective is formulated," + }, + { + "type": "equation", + "bbox": [ + 0.175, + 0.64, + 0.484, + 0.682 + ], + "angle": 0, + "content": "\\[\n\\mathcal {L} _ {i d} = - \\sum_ {i = 1} ^ {N} \\sum_ {j = 1} ^ {K} l _ {i, j} \\log \\left(H _ {l} (\\mathbf {y}) _ {j}\\right), \\tag {7}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.688, + 0.484, + 0.763 + ], + "angle": 0, + "content": "where \\( N \\) and \\( K \\) represent the number of characters and the possible indices respectively while \\( l \\) denotes the ground truth labels. This loss aggregates over all characters in the target text, ensuring the text encoder produces distinguishable embeddings for each character." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.764, + 0.483, + 0.793 + ], + "angle": 0, + "content": "The complete objective of our training strategy can be expressed as" + }, + { + "type": "equation", + "bbox": [ + 0.146, + 0.8, + 0.484, + 0.817 + ], + "angle": 0, + "content": "\\[\n\\mathcal {L} = \\mathcal {L} _ {\\text {m a s k}} + \\alpha \\mathcal {L} _ {\\text {a t t n}} + \\beta (\\mathcal {L} _ {\\text {a l i g n}} + \\mathcal {L} _ {i d}). \\tag {8}\n\\]" + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.83, + 0.299, + 0.846 + ], + "angle": 0, + "content": "3.2. Optimization Strategy" + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.852, + 0.368, + 0.867 + ], + "angle": 0, + "content": "3.2.1. Heuristic Alternate Optimization" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.871, + 0.486, + 0.901 + ], + "angle": 0, + "content": "Our objective involves numerous discrete variables, rendering it non-differentiable and making vanilla SGD invalid." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.229, + 0.907, + 0.44 + ], + "angle": 0, + "content": "Thus, we employ a heuristic alternate optimization strategy, utilizing latent character masks that encapsulate potential synthesis position information, to facilitate a symbiotic relationship between text encoder and U-Net. Specifically, we perform an alternating update between the tokens given the latent character masks and the masks themselves. During optimization, the masks are computed by Eq. 4, while for other parameters, we fix the masks and calculate gradients accordingly. Using the losses above, we first optimize the representation of specific characters given masks in each iteration, which, in turn, enables the generator to rectify the character's attention in subsequent steps. This process allows model to dynamically alternate between optimizing character embeddings and re-estimating character masks." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.45, + 0.878, + 0.466 + ], + "angle": 0, + "content": "3.2.2. Balanced Supervision on Character Attention" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.469, + 0.909, + 0.682 + ], + "angle": 0, + "content": "The initial training phase may challenge the generator in directing attention to the intended generation region for each character. Unlike existing methods that excessively constrain character attention which inevitably limits the model's flexibility, we adopt a balanced supervision. Initially, we guide attention calibration by cross-entropy objective between latent character masks and character segmentation masks from datasets similar to Eq. 7 for a warm-up in the earlier stage. Once the model preliminarily acquires the capability to estimate ideal generation positions (approximately 25k steps as shown in Fig. 4), we cease guidance, enabling it to engage in autonomous iterative learning. This strategy balances constraining the model and unleashing its flexibility in estimating characters' optimal generation positions." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.697, + 0.646, + 0.714 + ], + "angle": 0, + "content": "4. Experiments" + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.723, + 0.731, + 0.738 + ], + "angle": 0, + "content": "4.1. Implementation Details" + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.746, + 0.62, + 0.759 + ], + "angle": 0, + "content": "4.1.1. Datasets" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.765, + 0.907, + 0.901 + ], + "angle": 0, + "content": "Several datasets are used in our experiments: (1) SynthText [6] comprises 800,000 images containing around 8 million synthetic word instances. Each text instance is annotated with its corresponding text string, along with word-level and character-level bounding boxes. (2) LAION-OCR [5] contains 9,194,613 filtered high-quality text images including advertisements, notes, posters, covers, memes, logos, etc. (3) ICDAR13 [9] is widely recognized as the benchmark dataset for evaluating near-horizontal text detection, which consists" + }, + { + "type": "page_number", + "bbox": [ + 0.479, + 0.945, + 0.521, + 0.957 + ], + "angle": 0, + "content": "28560" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.134, + 0.09, + 0.865, + 0.334 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.279, + 0.343, + 0.719, + 0.358 + ], + "angle": 0, + "content": "Figure 7. Qualitative comparative results against state-of-the-art methods." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.37, + 0.484, + 0.491 + ], + "angle": 0, + "content": "of 233 test images used for evaluation purposes. (4) TextSeg [26] comprises 4,024 real-world text images sourced from various sources such as posters, greeting cards, book covers, logos, road signs, billboards, digital designs, handwritten notes, and more. Note that only SynthText and LAION-OCR datasets provide the character segmentation maps. We train our model using the training subsets of these datasets and randomly select 100 images from test subsets for testing." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.501, + 0.307, + 0.517 + ], + "angle": 0, + "content": "4.1.2. Training Configurations" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.521, + 0.484, + 0.658 + ], + "angle": 0, + "content": "We train our model based on the pre-trained checkpoint of SD-v2.0 inpainting version and text encoder in [29]. The pre-trained Vision Transformer [1] is used as our image encoder. The model is fine-tuned using 4 NVIDIA A100 GPUs on LAION-OCR for 200k steps, SynthText for 150k steps, TextSeg for 50k steps, and ICDAR13 for an additional 10k steps. We set \\(\\alpha\\) to 0.01, and \\(\\beta\\) to 0.001. Additionally, we utilize a batch size of 16 and a learning rate of \\(5 \\times 10^{-5}\\). The inference time of our DREAMTEXT is 8.5 seconds only." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.669, + 0.214, + 0.683 + ], + "angle": 0, + "content": "4.1.3. Evaluation" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.688, + 0.483, + 0.854 + ], + "angle": 0, + "content": "Text accuracy: We assess all methods on two tasks: scene text reconstruction and scene text editing [29]. For the scene text reconstruction task, we use the models to reconstruct the text image using the provided ground truth text label and binary mask. In the scene text editing task, we replace the original text in each image with a random word and evaluate the models by generating images containing the edited text. We use an off-the-shelf scene text recognition (STR) model [3] to identify the rendered text and then evaluate word-level correctness using sequence accuracy (SeqAcc) by comparing the STR result with the ground truth." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.856, + 0.484, + 0.901 + ], + "angle": 0, + "content": "Image quality: we utilize Fréchet Inception Distance (FID) [7] to quantify the distance between the text images in the dataset and the generated images. Additionally, the Learned" + }, + { + "type": "table_caption", + "bbox": [ + 0.592, + 0.368, + 0.826, + 0.381 + ], + "angle": 0, + "content": "Table 2. Accumulated results of losses." + }, + { + "type": "table", + "bbox": [ + 0.517, + 0.395, + 0.905, + 0.503 + ], + "angle": 0, + "content": "
SettingAverage SeqAccFIDLPIPS
ReconEditing
Base0.2180.06026.780.0696
+Mask0.4250.25923.210.0528
+Attn0.6980.53219.720.0483
+Align0.8840.80115.410.0392
+Id0.9400.88712.130.0328
" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.518, + 0.907, + 0.548 + ], + "angle": 0, + "content": "Perceptual Image Patch Similarity (LPIPS) [28] is utilized as an additional metric." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.56, + 0.608, + 0.575 + ], + "angle": 0, + "content": "4.2. Results" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.583, + 0.908, + 0.779 + ], + "angle": 0, + "content": "Quantitatively: In Tab. 6, we present a quantitative analysis comparing DREAMTEXT against baseline methods. Our method demonstrates a significant advantage across all quantitative metrics, highlighting the exceptional visual quality of the scene text images it generates. Notably, when compared to the previous state-of-the-art model UDiffText, DREAMTEXT outperforms it by 3.66 in terms of FID. While UDiffText and AnyText excel in image quality compared to other methods, except DREAMTEXT, they tend to perform less satisfactorily in terms of sequence accuracy, especially in text editing. This discrepancy may be attributed to its restricted representation domain and deflected attention, resulting in less accurate text rendering within complex scenes." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.78, + 0.909, + 0.901 + ], + "angle": 0, + "content": "Qualitatively: As depicted in Fig. 6, our method outperforms other baselines in synthesizing more coherent and accurate text within polystylistic scenes. For example, in Fig. 6 (3rd row), MOSTEL and Stable Diffusion fail to generate the text, while TextDiffuser and UDiffText encounter the issue of character absence. Similarly, in Fig. 6 (1st row), all baselines fail to generate the text on the provided poster. Moreover, we further compare our method with the latest ap" + }, + { + "type": "page_number", + "bbox": [ + 0.479, + 0.945, + 0.518, + 0.957 + ], + "angle": 0, + "content": "28561" + } + ], + [ + { + "type": "table_caption", + "bbox": [ + 0.184, + 0.09, + 0.39, + 0.104 + ], + "angle": 0, + "content": "Table 3. Hyperparameter analysis." + }, + { + "type": "table", + "bbox": [ + 0.093, + 0.117, + 0.48, + 0.37 + ], + "angle": 0, + "content": "
α\n(β=0)Average SeqAccβ\n(α=0.01)Average SeqAcc
ReconEditingReconEditing
10.6720.55810.8640.766
0.10.7240.6150.10.8960.817
0.010.7480.6230.010.9210.864
0.0010.7530.6040.0010.9400.887
inputstep 0step 0.2Tstep 0.4Tstep 0.8Tstep T
SUNSHINE\nSTATE\nOF MINDSUNSHINE\nSTATE\nOF MINDSUNSHINE\nSTATE\nOF MINDSUNSHINE\nSTATE\nOF MINDSUNSHINE\nSTATE\nOF MINDSUNSHINE\nSTATE\nOF MIND
The Bride's CrewThe Bride's\nCrewThe Bride's\nCrewThe Bride's\nCrewThe Bride's\nCrewThe Bride's\nCrew
WILBUR\nSMITHWILBUR\nSMITHWILBUR\nSMITHWILBUR\nSMITHWILBUR\nSMITHWILBUR\nSMITH
BLUE\nHORIZONBLUE\nHORIZONBLUE\nHORIZONBLUE\nHORIZONBLUE\nHORIZONBLUE\nHORIZON
" + }, + { + "type": "image_caption", + "bbox": [ + 0.09, + 0.38, + 0.483, + 0.409 + ], + "angle": 0, + "content": "Figure 8. The visualized attention results of all characters across several steps during training." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.426, + 0.484, + 0.489 + ], + "angle": 0, + "content": "proach, UDiffText, using the samples provided in its original paper for additional qualitative evaluation. The visualization results are presented in Fig. 7, offering additional evidence of our method's superiority in generating high-fidelity text." + }, + { + "type": "title", + "bbox": [ + 0.09, + 0.499, + 0.462, + 0.515 + ], + "angle": 0, + "content": "4.3. Ablation Studies and Effectiveness Analysis" + }, + { + "type": "text", + "bbox": [ + 0.089, + 0.522, + 0.483, + 0.687 + ], + "angle": 0, + "content": "Ablation on Losses: We employ SD-v2.0 inpainting version as the base method. As shown in Tab. 2, we observe a significant drop in performance across all metrics when these losses are removed. Additionally, these results highlight that the masked diffusion loss and cross-attention loss alone are insufficient for achieving satisfactory character representation. This may be attributed to the inherent noise in the latent character masks. To sum up, these ablation results underscore the effectiveness of each objective within our heuristic alternate optimization strategy in improving various aspects of synthesized scene texts." + }, + { + "type": "text", + "bbox": [ + 0.089, + 0.689, + 0.483, + 0.839 + ], + "angle": 0, + "content": "Choice of \\(\\alpha\\) and \\(\\beta\\): In our experiments, we vary the values of \\(\\alpha\\) and \\(\\beta\\) within the range [0.001, 0.01, 0.1, 1]. The results are presented in Tab. 3. Initially, we focus on exploring the impact of \\(\\alpha\\) while setting \\(\\beta = 0\\). We find that the optimal performance in text reconstruction occurs when \\(\\alpha = 0.001\\). However, this choice yields poor results in editing tasks. Consequently, we opt for a more balanced approach, selecting \\(\\alpha = 0.01\\). Subsequently, with \\(\\alpha\\) fixed at 0.01, we investigate the effect of \\(\\beta\\). We find that the optimal performance is achieved when \\(\\beta = 0.001\\)." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.84, + 0.484, + 0.901 + ], + "angle": 0, + "content": "Effectiveness of Heuristic Alternate Optimization: We analyze the efficacy of our strategy by visualizing the evolution of character attention maps across several training steps. As illustrated in Fig. 8, the model exhibits deflected attention" + }, + { + "type": "table_caption", + "bbox": [ + 0.561, + 0.09, + 0.857, + 0.104 + ], + "angle": 0, + "content": "Table 4. Ablation study results on warm-up steps." + }, + { + "type": "table", + "bbox": [ + 0.518, + 0.117, + 0.905, + 0.212 + ], + "angle": 0, + "content": "
Average SeqAcc
Recon ↑Editing ↑FID ↓mIoU ↑
15k0.8840.85213.820.681
20k0.9130.87313.380.692
25k0.9400.88712.130.722
30k0.9210.89113.240.703
" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.223, + 0.907, + 0.328 + ], + "angle": 0, + "content": "initially but progressively corrects it in subsequent training steps. Ultimately, attention becomes concentrated on the desired positions for all characters, indicating the effectiveness of our approach. Besides, Fig. 4 also illustrates the efficacy of this strategy in continuously improving the generator's ability to estimate the optimal characters' position by autonomously rectifying the attention during training." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.33, + 0.909, + 0.526 + ], + "angle": 0, + "content": "Choice of Warm-up Steps: We conduct an ablation experiment to analyze the impact of different warm-up steps on model performance. The results, presented in Tab. 4, demonstrate how varying the number of warm-up steps affects the performance. As the number of warm-up steps increases from 15k to 25k, we observe consistent improvements across all metrics. However, at 30k steps, while the editing accuracy slightly improves, the FID score worsens, suggesting a potential trade-off between image fidelity and editing capability. Overall, these findings highlight that 25k warm-up steps strike the best balance between reconstruction accuracy, editing performance, and image quality, making it the optimal choice for our method." + }, + { + "type": "title", + "bbox": [ + 0.514, + 0.54, + 0.634, + 0.557 + ], + "angle": 0, + "content": "5. Conclusion" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.567, + 0.909, + 0.765 + ], + "angle": 0, + "content": "This paper presents DREAMTEXT, a novel approach for high-fidelity scene text synthesis. We reconstruct the diffusion training process to expose and rectify the model's attention at the character level while strengthening its learning of text regions. This inevitably involves a hybrid optimization challenge, combining discrete and continuous variables, which we effectively address through our heuristic alternate optimization strategy. Meanwhile, we jointly train the text encoder and generator to comprehensively learn and utilize diverse fonts present in the training dataset, fostering a synergistic relationship between learning character embeddings and re-estimating character attention. Both qualitative and quantitative results highlight the superiority of our method." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.779, + 0.693, + 0.796 + ], + "angle": 0, + "content": "6. Acknowledgments" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.803, + 0.909, + 0.9 + ], + "angle": 0, + "content": "This work was supported by the National Nature Science Foundation of China (62472097), AI for Science Foundation of Fudan University (FudanX24AI028), and Shanghai Innovation Institute. The computations in this research were performed on the CFFF platform of Fudan University. Cheng Jin is the corresponding author of this work." + }, + { + "type": "page_number", + "bbox": [ + 0.479, + 0.945, + 0.521, + 0.958 + ], + "angle": 0, + "content": "28562" + } + ], + [ + { + "type": "title", + "bbox": [ + 0.093, + 0.091, + 0.188, + 0.106 + ], + "angle": 0, + "content": "References" + }, + { + "type": "ref_text", + "bbox": [ + 0.101, + 0.116, + 0.482, + 0.143 + ], + "angle": 0, + "content": "[1] Rowel Atienza. Vision transformer for fast and efficient scene text recognition. In ICDAR, pages 319-334, 2021. 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.101, + 0.145, + 0.483, + 0.2 + ], + "angle": 0, + "content": "[2] Omri Avrahami, Kfir Aberman, Ohad Fried, Daniel Cohen-Or, and Dani Lischinski. Break-a-scene: Extracting multiple concepts from a single image. In SIGGRAPH Asia, pages 1–12, 2023. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.203, + 0.483, + 0.243 + ], + "angle": 0, + "content": "[3] Darwin Bautista and Rowel Atienza. Scene text recognition with permuted autoregressive sequence models. In ECCV, pages 178-196, 2022. 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.246, + 0.483, + 0.299 + ], + "angle": 0, + "content": "[4] Haoxing Chen, Zhuoer Xu, Zhangxuan Gu, Yaohui Li, Changhua Meng, Huijia Zhu, Weiqiang Wang, et al. Diffuse: Universal text editing diffusion model. In NeurIPS, 2024. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.303, + 0.483, + 0.343 + ], + "angle": 0, + "content": "[5] Jingye Chen, Yupan Huang, Tengchao Lv, Lei Cui, Qifeng Chen, and Furu Wei. Textdiffuser: Diffusion models as text painters. In NeurIPS, 2023. 2, 3, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.346, + 0.483, + 0.386 + ], + "angle": 0, + "content": "[6] Ankush Gupta, Andrea Vedaldi, and Andrew Zisserman. Synthetic data for text localisation in natural images. In CVPR, pages 2315-2324, 2016. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.389, + 0.483, + 0.443 + ], + "angle": 0, + "content": "[7] Martin Heusel, Hubert Ramsauer, Thomas Unterthiner, Bernhard Nessler, and Sepp Hochreiter. Gans trained by a two time-scale update rule converge to a local nash equilibrium. In NeurIPS, 2017. 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.446, + 0.483, + 0.5 + ], + "angle": 0, + "content": "[8] Jiabao Ji, Guanhua Zhang, Zhaowen Wang, Bairu Hou, Zhifei Zhang, Brian Price, and Shiyu Chang. Improving diffusion models for scene text editing with dual encoders. arXiv preprint arXiv:2304.05568, 2023. 2, 3, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.503, + 0.483, + 0.572 + ], + "angle": 0, + "content": "[9] Dimosthenis Karatzas, Faisal Shafait, Seiichi Uchida, Masakazu Iwamura, Lluis Gomez i Bigorda, Sergi Robles Mestre, Joan Mas, David Fernandez Mota, Jon Almazan Almazan, and Lluis Pere De Las Heras. Icdar 2013 robust reading competition. In ICDAR, pages 1484-1493, 2013. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.575, + 0.485, + 0.641 + ], + "angle": 0, + "content": "[10] Yuxin Kong, Canjie Luo, Weihong Ma, Qiyuan Zhu, Shenggao Zhu, Nicholas Yuan, and Lianwen Jin. Look closer to supervise better: One-shot font generation via component-based discriminator. In CVPR, pages 13482-13491, 2022. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.645, + 0.483, + 0.699 + ], + "angle": 0, + "content": "[11] Sanyam Lakhanpal, Shivang Chopra, Vinija Jain, Aman Chadha, and Man Luo. Refining text-to-image generation: Towards accurate training-free glyph-enhanced image generation. arXiv preprint arXiv:2403.16422, 2024. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.702, + 0.483, + 0.756 + ], + "angle": 0, + "content": "[12] Zeyu Liu, Weicong Liang, Zhanhao Liang, Chong Luo, Ji Li, Gao Huang, and Yuhui Yuan. Glyph-byt5: A customized text encoder for accurate visual text rendering. arXiv preprint arXiv:2403.09622, 2024. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.759, + 0.483, + 0.814 + ], + "angle": 0, + "content": "[13] Song Park, Sanghyuk Chun, Junbum Cha, Bado Lee, and Hyunjung Shim. Multiple heads are better than one: Few-shot font generation with multiple localized experts. In ICCV, pages 13900-13909, 2021. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.817, + 0.483, + 0.87 + ], + "angle": 0, + "content": "[14] Yadong Qu, Qingfeng Tan, Hongtao Xie, Jianjun Xu, Yuxin Wang, and Yongdong Zhang. Exploring stroke-level modifications for scene text editing. In AAAI, pages 2119-2127, 2023. 1, 3, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.873, + 0.483, + 0.901 + ], + "angle": 0, + "content": "[15] Robin Rombach, Andreas Blattmann, Dominik Lorenz, Patrick Esser, and Björn Ommer. High-resolution image" + }, + { + "type": "list", + "bbox": [ + 0.094, + 0.116, + 0.485, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.549, + 0.093, + 0.908, + 0.12 + ], + "angle": 0, + "content": "synthesis with latent diffusion models. In CVPR, pages 10684-10695, 2022. 4, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.122, + 0.906, + 0.163 + ], + "angle": 0, + "content": "[16] Prasun Roy, Saumik Bhattacharya, Subhankar Ghosh, and Umapada Pal. Stefann: scene text editor using font adaptive neural network. In CVPR, pages 13228-13237, 2020. 1, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.165, + 0.907, + 0.245 + ], + "angle": 0, + "content": "[17] Chitwan Sahara, William Chan, Saurabh Saxena, Lala Li, Jay Whang, Emily L Denton, Kamyar Ghasemipour, Raphael Gontijo Lopes, Burcu Karagol Ayan, Tim Salimans, et al. Photorealistic text-to-image diffusion models with deep language understanding. In NeurIPS, pages 36479-36494, 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.249, + 0.906, + 0.289 + ], + "angle": 0, + "content": "[18] Wataru Shimoda, Daichi Haraguchi, Seiichi Uchida, and Kota Yamaguchi. De-rendering stylized texts. In ICCV, pages 1076-1085, 2021. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.292, + 0.907, + 0.332 + ], + "angle": 0, + "content": "[19] Yuxiang Tuo, Wangmeng Xiang, Jun-Yan He, Yifeng Geng, and Xuansong Xie. Anytext: Multilingual visual text generation and editing. ICLR, 2024. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.334, + 0.907, + 0.388 + ], + "angle": 0, + "content": "[20] Yibin Wang, Yuchao Feng, Jie Wu, Honghui Xu, and Jianwei Zheng. Ca-gan: Object placement via coalescing attention based generative adversarial network. In ICME, pages 2375-2380. IEEE, 2023. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.391, + 0.907, + 0.445 + ], + "angle": 0, + "content": "[21] Yibin Wang, Honghui Xu, Shanghai Zhou, Weizhong Zhang, and Cheng Jin. Enhancing object coherence in layout-to-image synthesis. arXiv preprint arXiv:2311.10522, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.448, + 0.907, + 0.489 + ], + "angle": 0, + "content": "[22] Yibin Wang, Weizhong Zhang, and Cheng Jin. Magicface: Training-free universal-style human image customized synthesis. arXiv preprint arXiv:2408.07433, 2024. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.491, + 0.906, + 0.53 + ], + "angle": 0, + "content": "[23] Yibin Wang, Weizhong Zhang, Jianwei Zheng, and Cheng Jin. High-fidelity person-centric subject-to-image synthesis. In CVPR, 2024. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.533, + 0.906, + 0.587 + ], + "angle": 0, + "content": "[24] Yibin Wang, Weizhong Zhang, Jianwei Zheng, and Cheng Jin. Primecomposer: Faster progressively combined diffusion for image composition with attention steering. arXiv preprint arXiv:2403.05053, 2024. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.59, + 0.907, + 0.631 + ], + "angle": 0, + "content": "[25] Liang Wu, Chengquan Zhang, Jiaming Liu, Junyu Han, Jingtuo Liu, Errui Ding, and Xiang Bai. Editing text in the wild. In ACM MM, pages 1500–1508, 2019. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.633, + 0.907, + 0.687 + ], + "angle": 0, + "content": "[26] Xingqian Xu, Zhifei Zhang, Zhaowen Wang, Brian Price, Zhonghao Wang, and Humphrey Shi. Rethinking text segmentation: A novel dataset and a text-specific refinement approach. In CVPR, pages 12045-12055, 2021. 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.689, + 0.907, + 0.729 + ], + "angle": 0, + "content": "[27] Qiangpeng Yang, Jun Huang, and Wei Lin. Swaptext: Image based texts transfer in scenes. In CVPR, pages 14700-14709, 2020. 1, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.733, + 0.907, + 0.785 + ], + "angle": 0, + "content": "[28] Richard Zhang, Phillip Isola, Alexei A Efros, Eli Shechtman, and Oliver Wang. The unreasonable effectiveness of deep features as a perceptual metric. In CVPR, pages 586-595, 2018. 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.789, + 0.907, + 0.842 + ], + "angle": 0, + "content": "[29] Yiming Zhao and Zhouhui Lian. Udifftext: A unified framework for high-quality text synthesis in arbitrary images via character-aware diffusion models. In ECCV, 2024. 2, 3, 4, 5, 6, 7" + }, + { + "type": "list", + "bbox": [ + 0.517, + 0.093, + 0.908, + 0.842 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.48, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "28563" + } + ] +] \ No newline at end of file diff --git a/2025/DreamText_ High Fidelity Scene Text Synthesis/6bc8ea68-7656-478a-8d11-90857d6ca90e_origin.pdf b/2025/DreamText_ High Fidelity Scene Text Synthesis/6bc8ea68-7656-478a-8d11-90857d6ca90e_origin.pdf new file mode 100644 index 0000000000000000000000000000000000000000..d67d0d21153fbe48f878b6624f25bce0e12c50f6 --- /dev/null +++ b/2025/DreamText_ High Fidelity Scene Text Synthesis/6bc8ea68-7656-478a-8d11-90857d6ca90e_origin.pdf @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:7fc1891b13afcefb5f0ab4ed15f3e49db5ac8bae0782184d912bb7dde2c30f5c +size 8896456 diff --git a/2025/DreamText_ High Fidelity Scene Text Synthesis/full.md b/2025/DreamText_ High Fidelity Scene Text Synthesis/full.md new file mode 100644 index 0000000000000000000000000000000000000000..4989991f5bfdddc630ece32436b021e71da7025a --- /dev/null +++ b/2025/DreamText_ High Fidelity Scene Text Synthesis/full.md @@ -0,0 +1,316 @@ +# DREAMTEXT: High Fidelity Scene Text Synthesis + +Yibin Wang $^{1,2}$ , Weizhong Zhang $^{1,4}$ , Honghui Xu $^{5}$ , Cheng Jin $^{1,3\dagger}$ + +$^{1}$ Fudan University $^{2}$ Shanghai Innovation Institute + +$^{3}$ Innovation Center of Calligraphy and Painting Creation Technology, MCT + +$^{4}$ Shanghai Key Laboratory of Intelligent Information Processing $^{5}$ Zhejiang University of Technology + +yibinwang1121@163.com, weizhongzhang@fudan.edu.cn, xhh@zjut.edu.cn, jc@fudan.edu.cn + +![](images/6ebdc6a9714dca6a4f1b52ca44fd15efe7bf40df27eece4554c3bf0a39fc7067.jpg) +Figure 1. Displayed are the results generated using our DREAMTEXT, showcasing its prowess across varied inputs. + +# Abstract + +Scene text synthesis involves rendering specified texts onto arbitrary images. Current methods typically formulate this task in an end-to-end manner but lack effective character-level guidance during training. Besides, their text encoders, pre-trained on a single font type, struggle to adapt to the diverse font styles encountered in practical applications. Consequently, these methods suffer from character distortion, repetition, and absence, particularly in polystylistic scenarios. To this end, this paper proposes DREAMTEXT for high-fidelity scene text synthesis. Our key idea is to reconstruct the diffusion training process, introducing more refined guidance tailored to this task, to expose and rectify the model's attention at the character level and strengthen its learning of text regions. This transformation poses a hybrid optimization challenge, involving both discrete and continuous variables. To effectively tackle this challenge, we employ a heuristic alternate optimization strategy. Meanwhile, we jointly train the text encoder and generator to comprehensively learn and utilize the diverse font present + +in the training dataset. This joint training is seamlessly integrated into the alternate optimization process, fostering a synergistic relationship between learning character embedding and re-estimating character attention. Specifically, in each step, we first encode potential character-generated position information from cross-attention maps into latent character masks. These masks are then utilized to update the representation of specific characters in the current step, which, in turn, enables the generator to correct the character's attention in the subsequent steps. Both qualitative and quantitative results demonstrate the superiority of our method to the state of the art. Our project page is here. + +# 1. Introduction + +Scene text synthesis involves modifying or inserting specified text in arbitrary images while maintaining its natural and realistic appearance. Earlier GAN-based style transfer methods [13, 14, 16, 27] accomplish this by transferring the text style from a reference image to the rendered target text image. However, they are constrained in their capacity to generate text in arbitrary styles, such as font and color. These + +![](images/208bd9f60383832fc13a1d2ca4503e20137e067a0add3a23ea91ab20c390de19.jpg) +Problem 1: Character repetition and absence +Problem 2: Character distortion + +![](images/35252f8ca281af7ac77b74acb845e6df83f2855272fe11c25ac8f977ebd04778.jpg) +Figure 2. Current methods encounter significant challenges, i.e., character repetition and absence (top) and character distortion (bottom). + +limitations are effectively addressed by utilizing the diffusion model [5, 8, 11, 12, 17, 21, 22], harnessing its inherent prior knowledge and robust capabilities. Nevertheless, these methods still encounter challenges in generating accurate text within polystylistic images due to the inadequate conditional guidance provided by their character-unaware text encoder. To this end, a recent study [29] proposes a character-level text encoder, providing more robust conditional guidance for the diffusion model. + +Despite their effectiveness, these methods are still difficult to accurately render text within complex scenes. Specifically, these methods such as UDiffText [29] and TextDiffuser [5] solely utilize a single font type to pre-train the text encoder, which is subsequently employed to fine-tune the generator. However, the diverse font styles encountered in practical applications present a significant challenge: the restricted representation domain hampers their ability to render text with unseen font styles, e.g., cases shown in Fig. 2 (bottom). Furthermore, these methods rely on character segmentation masks to supervise characters' attention for position control. We argue this approach has significant limitations: (1) unlike classification tasks with uniquely determined labels, the optimal generation positions for characters can be diverse. Therefore, using the specific masks to rigidly constrain the model may limit its flexibility in estimating optimal positions, hindering its ability to adapt to varied and complex scenarios. Besides, (2) these masks generated by a pre-trained segmentation model [5] are imprecise, tending to over-segment character regions. As a result, the model encounters issues of character repetition and absence as illustrated in Fig. 2 (top left) and (top right) respectively. We suspect these issues stem from characters' attention maps, as they reflect the underlying response to the character's + +final generated position [23, 24]. Therefore, we visualize the attention maps of their problematic results by rendering attention directly onto the images. We observe certain characters' attention misalignment, resulting in character duplication (see Fig. 3 (a) and (c)). Besides, excessive dispersion of attention also leads to character absence (see Fig. 3 (b)). This observation underscores the insufficiency of these methods in guiding the model autonomously estimating optimal character-generated position. Further, we also assess specific methods [5, 29] quantitatively on two datasets to evaluate their proficiency in directing characters' attention toward their optimal generation regions throughout multiple global training steps. Specifically, for each test sample, we extract potential position masks for all characters based on their cross-attention maps from the inference process. Then, we compute the mean Intersection over Union (mIoU) between these masks and ground truth character segmentation masks. Their inoptimal results depicted in Fig. 4 further corroborate our hypothesis. + +In summary, current methods encounter challenges in accurately rendering text within complex scenes due to their constrained representation domain and synthesizing characters in their ideal regions due to the lack of effective guidance for the model to estimate optimal character-generated position. To this end, this paper proposes DREAMTEXT for high-fidelity text synthesis. We reconstruct the diffusion training process, embedding more refined guidance: we expose and rectify the characters' attention to focus more precisely on their ideal generation areas, and enhance the learning of text regions and character representation by additional constraints. However, this reconstruction inevitably introduces a complex hybrid optimization problem encompassing both discrete and continuous variables. Therefore, + +![](images/b05c4cdd7650ce0b011e58a8ec5430088a0bad5fe880c76c564d249393632ff1.jpg) +Figure 3. The problematic results rendered by characters' attention maps (AMs). + +we design a heuristic alternate optimization strategy. Concurrently, we jointly train the text encoder and the generator, leveraging the diverse font styles within the training dataset to enrich character representation space. We demonstrate that this joint training can be seamlessly integrated into the heuristic alternate optimization process, facilitating a synergistic interplay between character representations learning and character attention re-estimation. Specifically, in each iteration, the pipeline begins by encoding the potential character-generated position information from cross-attention maps into latent character masks. These masks are subsequently employed to refine the representation of specific characters within the text encoder, facilitating the calibration of the characters' attention in the subsequent step. This iterative process contributes to both learning of better character representation and, as a by-product, explicit guidance for autonomously estimating character position. + +Notably, the generator may initially struggle to direct attention to the desired generation region for each character, resulting in suboptimal latent masks and impacting training progress. Unlike existing methods [5, 29] that rigidly constrain character attention, we employ a balanced supervision strategy. Initially, we assist in calibrating the attention using character segmentation masks for a warm-up in the earlier stage. Once the model has preliminarily acquired the ability to estimate ideal generation positions, we remove this guidance, allowing the model to autonomous learning iteratively. This strategy strikes a balance between constraining the model and unleashing its flexibility in estimating optimal generation positions. Fig. 4 illustrates the efficacy of our + +![](images/41b6185fbff08675be8f802b118e82b3e53adf16bbdb725d7cb43bf017de6b4c.jpg) +Figure 4. The mIoU scores of UDiffText, TextDiffuser, and our method on LAION-OCR and SynthText over global training steps. Our method adopts a balanced supervision strategy: we initially use latent character masks to steer the character's attention for a warm-up in the earlier stage and stop guiding after 25,000 steps. + +method in continuously improving the generator's ability to estimate the optimal characters' position. + +Our contributions are: (1) The proposed DREAMTEXT effectively alleviates the issues of character repetition, absence, and distortion encountered by existing methods. (2) Our heuristic alternate optimization strategy, integrating the joint learning of the text encoder and U-Net, orchestrates a symbiotic relationship between learning character representations and re-estimating character attention. (3) Our balanced supervision strategy strikes a balance between constraining the model and unleashing its flexibility in estimating optimal generation positions. (4) Both qualitative and quantitative results demonstrate the superiority of our method. + +# 2. Related Work + +Earlier GAN-based methods achieve scene text synthesis by transferring the text style in a reference image to the rendered target text image [10, 16, 18, 20, 27]. Specifically, STEFANN [16] employs a FANnet to edit individual characters and incorporates a placement algorithm to generate the desired word. SRNet [25] and MOSTEL [14], on the other hand, split the task into two main parts: background inpainting and text style transfer. This approach allows for the end-to-end synthesizing of entire words. Despite their simplicity and effectiveness, they are limited in their ability to generate text in arbitrary styles and locations, often producing less natural-looking images. + +To this end, several diffusion-based approaches have emerged to address these challenges. These approaches harness the robust capabilities of diffusion models to edit or generate scene text, thereby enhancing the quality and diversity of the generated content. DiffSTE [8] proposes a dual encoder structure, comprising a character text encoder and an instruction text encoder. These components are employed for instruction tuning, providing improved control over the backbone network. DiffUTE [4] employs an OCR-based glyph encoder to extract glyph guidance from the rendered glyph image. TextDiffuser [5] concatenates the segmentation mask as conditional input and employs character-aware loss to control the generated characters precisely. However, they still face challenges in generating accurate text + +![](images/016c3053ed95d0bb11c7831077688a40c8caf28d5bbd776bb086f7c16f4c4442.jpg) +Figure 5. An overview of proposed heuristic alternate optimization strategy. + +within polystylistic images, primarily due to the insufficient conditional guidance provided by their character-unaware text encoder. Motivated by this, UDiffText [29] replaces the original CLIP text encoder in Stable Diffusion with a character-level text encoder with a unitary font. It also employs local attention loss for position control, relying on ground truth character segmentation maps. + +However, these methods still encounter challenges in accurately rendering text within complex scenes and synthesizing characters in their ideal regions, attributed to their constrained character representation domain and lack of effective guidance for the model to autonomously estimate character-generated position, respectively. + +# 3. Method + +Given an input image $z$ , a text region specified by a binary mask $B$ and a text condition $c$ , scene text synthesis is designed to render the given text onto the input image at the given text region. A natural idea for scene text synthesis is to formulate it into a latent diffusion process [15], which is equivalent to solving the following optimization problem: + +$$ +\min _ {(\theta , \vartheta)} \mathcal {L} _ {L D M} \triangleq \mathbb {E} _ {\boldsymbol {z}, c, \epsilon \sim N (0, 1), t} \| \boldsymbol {\epsilon} - \epsilon_ {\theta} (\boldsymbol {z} _ {t}, t, \psi_ {\vartheta} (c), \boldsymbol {B}) \| _ {2} ^ {2}, +$$ + +where $\psi_{\vartheta}$ is the learnable character-level text encoder initialized by UDifftext [29] and $\theta$ is the model parameter vector. However, we argue that such a straightforward formulation could be problematic due to the following reasons: + +- Without explicit guidance for rectifying characters' attention, character repetition and absence can inevitably occur in the rendered results as shown in Fig. 2. + +- This formulation uniformly measures the mean distance between all pixels, without emphasizing the text regions. Consequently, text-unrelated regions can adversely affect the learning of the text encoder during backpropagation. The experimental analysis is provided in Sec. 4.3. +- Due to the intricate interactions among the U-Net and text encoder, end-to-end training is particularly challenging as shown in Sec. 4.3. Therefore, a more refined guidance for the training process is necessary. + +In this paper, we propose DREAMTEXT and develop a heuristic iterative optimization pipeline to enable efficient training. Our key idea is to explicitly model the potential character positions in attention maps by encoding them into latent character masks in each step. These masks are then used to enhance the learning of the text encoder through our carefully crafted losses, facilitating the calibration of character attention in the subsequent step. This heuristic iterative process allows our model to dynamically alternate between optimizing character embedding and re-estimating character generation positions. The details are presented in the following sections. + +# 3.1. Latent Diffusion with Refined Guidance + +# 3.1.1. Latent Character Mask + +We utilize cross-attention maps to generate latent character masks, extracting characters' latent position information in the current step. Given latent image $z_{t}$ and text embedding $\psi_{\vartheta}(c)$ , the cross-attention map in layer $l$ is defined as follows: + +![](images/0c95447a95ca4921c541351b13ab80dc92346dc58fafc23aec3b6dea3c217086.jpg) +Figure 6. Qualitative comparative results against state-of-the-art methods. + +$$ +\boldsymbol {Q} _ {l} = \boldsymbol {z} _ {t} \boldsymbol {W} _ {l} ^ {q}, \quad \boldsymbol {K} _ {l} = \psi_ {\vartheta} (c) \boldsymbol {W} _ {l} ^ {k}, \tag {1} +$$ + +$$ +\boldsymbol {A} _ {l} = \operatorname {s o f t m a x} \left(\frac {\boldsymbol {Q} _ {l} \boldsymbol {K} _ {l} ^ {T}}{\sqrt {d}}\right), \tag {2} +$$ + +where $\pmb{W}_l^q$ and $\pmb{W}_l^k$ are learnable parameters and $d$ is the embedding dimension. The attention maps $\pmb{A}_l$ is reshaped to $N \times H \times W$ , where $N$ represents the number of text tokens such that each slice $\pmb{A}_l^i \in \mathbb{R}^{H \times W}$ represents the region attended by the $i$ -th token. These attention maps are then averaged across all layers of the U-Net to get the mean response $\bar{\pmb{A}} = \frac{1}{L} \sum_{l=1}^{L} \pmb{A}_l$ for the regions attended by character tokens. + +To obtain the latent character masks $M \in \mathbb{R}^{N \times H \times W}$ , we first apply Gaussian blur to the attention map to execute low-pass filtering $\mathrm{blur}(\bar{A})$ . This step helps mitigate excessive variance in the attended regions, ensuring a more uniform distribution of attention across relevant character regions. Subsequently, we employ a straightforward thresholding process to convert the blurred attention map into binary character masks. This process is encapsulated by the function $f(\cdot)$ , which applies a threshold and assigns a value of 1 to pixel values exceeding the threshold, and 0 otherwise. The threshold is calculated as the mean value plus twice the variance of averaged attention maps in this work. That is: + +$$ +f (\boldsymbol {X}) = \left\{ \begin{array}{l} 1, \text {i f} x _ {i, j} > \operatorname {m e a n} (\boldsymbol {X}) + 2 \operatorname {s t d} (\boldsymbol {X}) \\ 0, \text {o t h e r w i s e} \end{array} \right. \tag {3} +$$ + +Therefore, the equation for obtaining $M$ can be expressed as: + +$$ +\boldsymbol {M} = f (\operatorname {b l u r} (\bar {\boldsymbol {A}})). \tag {4} +$$ + +Based on the latent character masks $M$ , we employ several + +loss functions to optimize both the text encoder and U-Net iteratively. We will delve into these in the following sections. + +# 3.1.2. Masked Diffusion Loss + +We first extend the Eq. 1 to highlight all desired character tokens in the current step. Specifically, within a text $c$ containing $k$ tokens of interest, the diffusion loss of the corresponding pixels obtained from each token's latent character mask is applied with an additional weighting factor $\gamma$ . Formally, considering $M_{k} = \bigvee_{i=1}^{k} M_{i}$ to the union of the pixels of $k$ characters, the masked diffusion loss is formulated as, + +$$ +\mathcal {L} _ {\text {m a s k}} = \mathbb {E} _ {z, c, \epsilon \sim N (0, 1), t} \| (1 + \gamma \boldsymbol {M} _ {k}) (\boldsymbol {\epsilon} - \epsilon_ {\theta} (\boldsymbol {z} _ {t}, t, \psi_ {\vartheta} (c))) \| _ {2} ^ {2}. +$$ + +We omit binary mask $B$ here for the simplicity. + +# 3.1.3. Cross Attention Loss + +The masked diffusion loss is designed to focus on all desired concept tokens within a prompt. Additionally, to ensure that each token encodes information specific to its corresponding synthesis position, we incorporate the cross-attention loss [2]. This loss encourages tokens to attend exclusively to their corresponding target regions, + +$$ +\mathcal {L} _ {a t t n} = \mathbb {E} _ {\boldsymbol {z}, c, t} \| C _ {a t t n} (\boldsymbol {z} _ {t}, \psi_ {\vartheta} (c) _ {i}) - \boldsymbol {M} _ {i} \| _ {2} ^ {2}. \tag {5} +$$ + +Here, $C_{attn}(z_t, \psi_\vartheta(c)_i)$ represents the cross-attention map between the visual representation $z_t$ and the token $\psi_\vartheta(c)_i$ , while $M_i$ denotes the latent character mask of $\psi_\vartheta(c)_i$ . + +However, the losses mentioned above alone are inadequate for achieving optimal character representation. This is primarily due to the inherent noise in the latent character masks, which frequently results in under- or oversegmentation of the ideal character regions during training. To mitigate the influence of noise and prevent skewed learning of character representation, we introduce additional losses [29] as shown below. + +Table 1. Quantitative comparison against baselines. + +
MethodsSeqAcc-ReconSeqAcc-EditingFIDLPIPS
ICDAR13(8ch)ICDAR13TextSegLAION-OCRICDAR13(8ch)ICDAR13TextSegLAION-OCR
CVPR'22 SD-Inpainting [15]0.320.290.110.150.080.070.040.0526.780.0696
arXiv'23 DiffSTE [8]0.450.370.500.410.340.290.470.2751.670.1050
AAAI'23 MOSTEL [14]0.750.680.640.710.350.280.250.4425.090.0605
NIPS'23 TextDiffuser [5]0.870.810.680.800.820.750.660.6432.250.0834
ICLR'24 AnyText [19]0.890.870.810.860.810.790.800.7222.730.0651
ECCV'24 UDiffText [29]0.940.910.930.900.840.830.840.7815.790.0564
DreamText0.950.940.960.930.870.890.910.8812.130.0328
+ +# 3.1.4. Cross-modal Aligned Loss + +To obtain accurate and robust embeddings, we introduce an image encoder $\xi$ , text head $H_{t}$ , and visual head $H_{v}$ to align cross-modal character features. Specifically, we compute the cosine similarity objective between visual and text representations to maximize the alignment between two modalities: + +$$ +\mathcal {L} _ {\text {a l i g n}} = \frac {\left\langle H _ {t} (\mathbf {y}) , H _ {v} (\xi (\mathbf {I})) \right\rangle}{\| H _ {t} (\mathbf {y}) \| _ {2} \cdot \| H _ {v} (\xi (\mathbf {I})) \| _ {2}}. \tag {6} +$$ + +where $I$ represents the text image segmented via the bounding box from datasets and $\langle ,\rangle$ presents the inner product of two vectors. Note that to alleviate noise stemming from background and color variations, we preprocess the text image by converting it to grayscale. + +Through training, the text head learns to map the whole information of multiple characters into the text feature space. Similarly, the image head maps the overall visual text information into the visual feature space. This effectively ensures the overall textual and visual representations correspond, capturing high-level cross-modal consistency. + +# 3.1.5. Character Id Loss + +Additionally, we introduce character Id loss to ensure that the learned embeddings are highly distinguishable. To be precise, a multi-label classification head $H_{l}$ is introduced to predict character indices from text embeddings $y$ . Then, the cross-entropy objective is formulated, + +$$ +\mathcal {L} _ {i d} = - \sum_ {i = 1} ^ {N} \sum_ {j = 1} ^ {K} l _ {i, j} \log \left(H _ {l} (\mathbf {y}) _ {j}\right), \tag {7} +$$ + +where $N$ and $K$ represent the number of characters and the possible indices respectively while $l$ denotes the ground truth labels. This loss aggregates over all characters in the target text, ensuring the text encoder produces distinguishable embeddings for each character. + +The complete objective of our training strategy can be expressed as + +$$ +\mathcal {L} = \mathcal {L} _ {\text {m a s k}} + \alpha \mathcal {L} _ {\text {a t t n}} + \beta (\mathcal {L} _ {\text {a l i g n}} + \mathcal {L} _ {i d}). \tag {8} +$$ + +# 3.2. Optimization Strategy + +# 3.2.1. Heuristic Alternate Optimization + +Our objective involves numerous discrete variables, rendering it non-differentiable and making vanilla SGD invalid. + +Thus, we employ a heuristic alternate optimization strategy, utilizing latent character masks that encapsulate potential synthesis position information, to facilitate a symbiotic relationship between text encoder and U-Net. Specifically, we perform an alternating update between the tokens given the latent character masks and the masks themselves. During optimization, the masks are computed by Eq. 4, while for other parameters, we fix the masks and calculate gradients accordingly. Using the losses above, we first optimize the representation of specific characters given masks in each iteration, which, in turn, enables the generator to rectify the character's attention in subsequent steps. This process allows model to dynamically alternate between optimizing character embeddings and re-estimating character masks. + +# 3.2.2. Balanced Supervision on Character Attention + +The initial training phase may challenge the generator in directing attention to the intended generation region for each character. Unlike existing methods that excessively constrain character attention which inevitably limits the model's flexibility, we adopt a balanced supervision. Initially, we guide attention calibration by cross-entropy objective between latent character masks and character segmentation masks from datasets similar to Eq. 7 for a warm-up in the earlier stage. Once the model preliminarily acquires the capability to estimate ideal generation positions (approximately 25k steps as shown in Fig. 4), we cease guidance, enabling it to engage in autonomous iterative learning. This strategy balances constraining the model and unleashing its flexibility in estimating characters' optimal generation positions. + +# 4. Experiments + +# 4.1. Implementation Details + +# 4.1.1. Datasets + +Several datasets are used in our experiments: (1) SynthText [6] comprises 800,000 images containing around 8 million synthetic word instances. Each text instance is annotated with its corresponding text string, along with word-level and character-level bounding boxes. (2) LAION-OCR [5] contains 9,194,613 filtered high-quality text images including advertisements, notes, posters, covers, memes, logos, etc. (3) ICDAR13 [9] is widely recognized as the benchmark dataset for evaluating near-horizontal text detection, which consists + +![](images/fe9665a40efabaf5f6e9303953e02affd1798d3abb6d02842a27edd39ddb7132.jpg) +Figure 7. Qualitative comparative results against state-of-the-art methods. + +of 233 test images used for evaluation purposes. (4) TextSeg [26] comprises 4,024 real-world text images sourced from various sources such as posters, greeting cards, book covers, logos, road signs, billboards, digital designs, handwritten notes, and more. Note that only SynthText and LAION-OCR datasets provide the character segmentation maps. We train our model using the training subsets of these datasets and randomly select 100 images from test subsets for testing. + +# 4.1.2. Training Configurations + +We train our model based on the pre-trained checkpoint of SD-v2.0 inpainting version and text encoder in [29]. The pre-trained Vision Transformer [1] is used as our image encoder. The model is fine-tuned using 4 NVIDIA A100 GPUs on LAION-OCR for 200k steps, SynthText for 150k steps, TextSeg for 50k steps, and ICDAR13 for an additional 10k steps. We set $\alpha$ to 0.01, and $\beta$ to 0.001. Additionally, we utilize a batch size of 16 and a learning rate of $5 \times 10^{-5}$ . The inference time of our DREAMTEXT is 8.5 seconds only. + +# 4.1.3. Evaluation + +Text accuracy: We assess all methods on two tasks: scene text reconstruction and scene text editing [29]. For the scene text reconstruction task, we use the models to reconstruct the text image using the provided ground truth text label and binary mask. In the scene text editing task, we replace the original text in each image with a random word and evaluate the models by generating images containing the edited text. We use an off-the-shelf scene text recognition (STR) model [3] to identify the rendered text and then evaluate word-level correctness using sequence accuracy (SeqAcc) by comparing the STR result with the ground truth. + +Image quality: we utilize Fréchet Inception Distance (FID) [7] to quantify the distance between the text images in the dataset and the generated images. Additionally, the Learned + +Table 2. Accumulated results of losses. + +
SettingAverage SeqAccFIDLPIPS
ReconEditing
Base0.2180.06026.780.0696
+Mask0.4250.25923.210.0528
+Attn0.6980.53219.720.0483
+Align0.8840.80115.410.0392
+Id0.9400.88712.130.0328
+ +Perceptual Image Patch Similarity (LPIPS) [28] is utilized as an additional metric. + +# 4.2. Results + +Quantitatively: In Tab. 6, we present a quantitative analysis comparing DREAMTEXT against baseline methods. Our method demonstrates a significant advantage across all quantitative metrics, highlighting the exceptional visual quality of the scene text images it generates. Notably, when compared to the previous state-of-the-art model UDiffText, DREAMTEXT outperforms it by 3.66 in terms of FID. While UDiffText and AnyText excel in image quality compared to other methods, except DREAMTEXT, they tend to perform less satisfactorily in terms of sequence accuracy, especially in text editing. This discrepancy may be attributed to its restricted representation domain and deflected attention, resulting in less accurate text rendering within complex scenes. + +Qualitatively: As depicted in Fig. 6, our method outperforms other baselines in synthesizing more coherent and accurate text within polystylistic scenes. For example, in Fig. 6 (3rd row), MOSTEL and Stable Diffusion fail to generate the text, while TextDiffuser and UDiffText encounter the issue of character absence. Similarly, in Fig. 6 (1st row), all baselines fail to generate the text on the provided poster. Moreover, we further compare our method with the latest ap + +Table 3. Hyperparameter analysis. + +
α +(β=0)Average SeqAccβ +(α=0.01)Average SeqAcc
ReconEditingReconEditing
10.6720.55810.8640.766
0.10.7240.6150.10.8960.817
0.010.7480.6230.010.9210.864
0.0010.7530.6040.0010.9400.887
inputstep 0step 0.2Tstep 0.4Tstep 0.8Tstep T
SUNSHINE +STATE +OF MINDSUNSHINE +STATE +OF MINDSUNSHINE +STATE +OF MINDSUNSHINE +STATE +OF MINDSUNSHINE +STATE +OF MINDSUNSHINE +STATE +OF MIND
The Bride's CrewThe Bride's +CrewThe Bride's +CrewThe Bride's +CrewThe Bride's +CrewThe Bride's +Crew
WILBUR +SMITHWILBUR +SMITHWILBUR +SMITHWILBUR +SMITHWILBUR +SMITHWILBUR +SMITH
BLUE +HORIZONBLUE +HORIZONBLUE +HORIZONBLUE +HORIZONBLUE +HORIZONBLUE +HORIZON
+ +Figure 8. The visualized attention results of all characters across several steps during training. + +proach, UDiffText, using the samples provided in its original paper for additional qualitative evaluation. The visualization results are presented in Fig. 7, offering additional evidence of our method's superiority in generating high-fidelity text. + +# 4.3. Ablation Studies and Effectiveness Analysis + +Ablation on Losses: We employ SD-v2.0 inpainting version as the base method. As shown in Tab. 2, we observe a significant drop in performance across all metrics when these losses are removed. Additionally, these results highlight that the masked diffusion loss and cross-attention loss alone are insufficient for achieving satisfactory character representation. This may be attributed to the inherent noise in the latent character masks. To sum up, these ablation results underscore the effectiveness of each objective within our heuristic alternate optimization strategy in improving various aspects of synthesized scene texts. + +Choice of $\alpha$ and $\beta$ : In our experiments, we vary the values of $\alpha$ and $\beta$ within the range [0.001, 0.01, 0.1, 1]. The results are presented in Tab. 3. Initially, we focus on exploring the impact of $\alpha$ while setting $\beta = 0$ . We find that the optimal performance in text reconstruction occurs when $\alpha = 0.001$ . However, this choice yields poor results in editing tasks. Consequently, we opt for a more balanced approach, selecting $\alpha = 0.01$ . Subsequently, with $\alpha$ fixed at 0.01, we investigate the effect of $\beta$ . We find that the optimal performance is achieved when $\beta = 0.001$ . + +Effectiveness of Heuristic Alternate Optimization: We analyze the efficacy of our strategy by visualizing the evolution of character attention maps across several training steps. As illustrated in Fig. 8, the model exhibits deflected attention + +Table 4. Ablation study results on warm-up steps. + +
Average SeqAcc
Recon ↑Editing ↑FID ↓mIoU ↑
15k0.8840.85213.820.681
20k0.9130.87313.380.692
25k0.9400.88712.130.722
30k0.9210.89113.240.703
+ +initially but progressively corrects it in subsequent training steps. Ultimately, attention becomes concentrated on the desired positions for all characters, indicating the effectiveness of our approach. Besides, Fig. 4 also illustrates the efficacy of this strategy in continuously improving the generator's ability to estimate the optimal characters' position by autonomously rectifying the attention during training. + +Choice of Warm-up Steps: We conduct an ablation experiment to analyze the impact of different warm-up steps on model performance. The results, presented in Tab. 4, demonstrate how varying the number of warm-up steps affects the performance. As the number of warm-up steps increases from 15k to 25k, we observe consistent improvements across all metrics. However, at 30k steps, while the editing accuracy slightly improves, the FID score worsens, suggesting a potential trade-off between image fidelity and editing capability. Overall, these findings highlight that 25k warm-up steps strike the best balance between reconstruction accuracy, editing performance, and image quality, making it the optimal choice for our method. + +# 5. Conclusion + +This paper presents DREAMTEXT, a novel approach for high-fidelity scene text synthesis. We reconstruct the diffusion training process to expose and rectify the model's attention at the character level while strengthening its learning of text regions. This inevitably involves a hybrid optimization challenge, combining discrete and continuous variables, which we effectively address through our heuristic alternate optimization strategy. Meanwhile, we jointly train the text encoder and generator to comprehensively learn and utilize diverse fonts present in the training dataset, fostering a synergistic relationship between learning character embeddings and re-estimating character attention. Both qualitative and quantitative results highlight the superiority of our method. + +# 6. Acknowledgments + +This work was supported by the National Nature Science Foundation of China (62472097), AI for Science Foundation of Fudan University (FudanX24AI028), and Shanghai Innovation Institute. The computations in this research were performed on the CFFF platform of Fudan University. Cheng Jin is the corresponding author of this work. + +# References + +[1] Rowel Atienza. Vision transformer for fast and efficient scene text recognition. In ICDAR, pages 319-334, 2021. 7 +[2] Omri Avrahami, Kfir Aberman, Ohad Fried, Daniel Cohen-Or, and Dani Lischinski. Break-a-scene: Extracting multiple concepts from a single image. In SIGGRAPH Asia, pages 1–12, 2023. 5 +[3] Darwin Bautista and Rowel Atienza. Scene text recognition with permuted autoregressive sequence models. In ECCV, pages 178-196, 2022. 7 +[4] Haoxing Chen, Zhuoer Xu, Zhangxuan Gu, Yaohui Li, Changhua Meng, Huijia Zhu, Weiqiang Wang, et al. Diffuse: Universal text editing diffusion model. In NeurIPS, 2024. 3 +[5] Jingye Chen, Yupan Huang, Tengchao Lv, Lei Cui, Qifeng Chen, and Furu Wei. Textdiffuser: Diffusion models as text painters. In NeurIPS, 2023. 2, 3, 6 +[6] Ankush Gupta, Andrea Vedaldi, and Andrew Zisserman. Synthetic data for text localisation in natural images. In CVPR, pages 2315-2324, 2016. 6 +[7] Martin Heusel, Hubert Ramsauer, Thomas Unterthiner, Bernhard Nessler, and Sepp Hochreiter. Gans trained by a two time-scale update rule converge to a local nash equilibrium. In NeurIPS, 2017. 7 +[8] Jiabao Ji, Guanhua Zhang, Zhaowen Wang, Bairu Hou, Zhifei Zhang, Brian Price, and Shiyu Chang. Improving diffusion models for scene text editing with dual encoders. arXiv preprint arXiv:2304.05568, 2023. 2, 3, 6 +[9] Dimosthenis Karatzas, Faisal Shafait, Seiichi Uchida, Masakazu Iwamura, Lluis Gomez i Bigorda, Sergi Robles Mestre, Joan Mas, David Fernandez Mota, Jon Almazan Almazan, and Lluis Pere De Las Heras. Icdar 2013 robust reading competition. In ICDAR, pages 1484-1493, 2013. 6 +[10] Yuxin Kong, Canjie Luo, Weihong Ma, Qiyuan Zhu, Shenggao Zhu, Nicholas Yuan, and Lianwen Jin. Look closer to supervise better: One-shot font generation via component-based discriminator. In CVPR, pages 13482-13491, 2022. 3 +[11] Sanyam Lakhanpal, Shivang Chopra, Vinija Jain, Aman Chadha, and Man Luo. Refining text-to-image generation: Towards accurate training-free glyph-enhanced image generation. arXiv preprint arXiv:2403.16422, 2024. 2 +[12] Zeyu Liu, Weicong Liang, Zhanhao Liang, Chong Luo, Ji Li, Gao Huang, and Yuhui Yuan. Glyph-byt5: A customized text encoder for accurate visual text rendering. arXiv preprint arXiv:2403.09622, 2024. 2 +[13] Song Park, Sanghyuk Chun, Junbum Cha, Bado Lee, and Hyunjung Shim. Multiple heads are better than one: Few-shot font generation with multiple localized experts. In ICCV, pages 13900-13909, 2021. 1 +[14] Yadong Qu, Qingfeng Tan, Hongtao Xie, Jianjun Xu, Yuxin Wang, and Yongdong Zhang. Exploring stroke-level modifications for scene text editing. In AAAI, pages 2119-2127, 2023. 1, 3, 6 +[15] Robin Rombach, Andreas Blattmann, Dominik Lorenz, Patrick Esser, and Björn Ommer. High-resolution image + +synthesis with latent diffusion models. In CVPR, pages 10684-10695, 2022. 4, 6 +[16] Prasun Roy, Saumik Bhattacharya, Subhankar Ghosh, and Umapada Pal. Stefann: scene text editor using font adaptive neural network. In CVPR, pages 13228-13237, 2020. 1, 3 +[17] Chitwan Sahara, William Chan, Saurabh Saxena, Lala Li, Jay Whang, Emily L Denton, Kamyar Ghasemipour, Raphael Gontijo Lopes, Burcu Karagol Ayan, Tim Salimans, et al. Photorealistic text-to-image diffusion models with deep language understanding. In NeurIPS, pages 36479-36494, 2022. 2 +[18] Wataru Shimoda, Daichi Haraguchi, Seiichi Uchida, and Kota Yamaguchi. De-rendering stylized texts. In ICCV, pages 1076-1085, 2021. 3 +[19] Yuxiang Tuo, Wangmeng Xiang, Jun-Yan He, Yifeng Geng, and Xuansong Xie. Anytext: Multilingual visual text generation and editing. ICLR, 2024. 6 +[20] Yibin Wang, Yuchao Feng, Jie Wu, Honghui Xu, and Jianwei Zheng. Ca-gan: Object placement via coalescing attention based generative adversarial network. In ICME, pages 2375-2380. IEEE, 2023. 3 +[21] Yibin Wang, Honghui Xu, Shanghai Zhou, Weizhong Zhang, and Cheng Jin. Enhancing object coherence in layout-to-image synthesis. arXiv preprint arXiv:2311.10522, 2023. 2 +[22] Yibin Wang, Weizhong Zhang, and Cheng Jin. Magicface: Training-free universal-style human image customized synthesis. arXiv preprint arXiv:2408.07433, 2024. 2 +[23] Yibin Wang, Weizhong Zhang, Jianwei Zheng, and Cheng Jin. High-fidelity person-centric subject-to-image synthesis. In CVPR, 2024. 2 +[24] Yibin Wang, Weizhong Zhang, Jianwei Zheng, and Cheng Jin. Primecomposer: Faster progressively combined diffusion for image composition with attention steering. arXiv preprint arXiv:2403.05053, 2024. 2 +[25] Liang Wu, Chengquan Zhang, Jiaming Liu, Junyu Han, Jingtuo Liu, Errui Ding, and Xiang Bai. Editing text in the wild. In ACM MM, pages 1500–1508, 2019. 3 +[26] Xingqian Xu, Zhifei Zhang, Zhaowen Wang, Brian Price, Zhonghao Wang, and Humphrey Shi. Rethinking text segmentation: A novel dataset and a text-specific refinement approach. In CVPR, pages 12045-12055, 2021. 7 +[27] Qiangpeng Yang, Jun Huang, and Wei Lin. Swaptext: Image based texts transfer in scenes. In CVPR, pages 14700-14709, 2020. 1, 3 +[28] Richard Zhang, Phillip Isola, Alexei A Efros, Eli Shechtman, and Oliver Wang. The unreasonable effectiveness of deep features as a perceptual metric. In CVPR, pages 586-595, 2018. 7 +[29] Yiming Zhao and Zhouhui Lian. Udifftext: A unified framework for high-quality text synthesis in arbitrary images via character-aware diffusion models. In ECCV, 2024. 2, 3, 4, 5, 6, 7 \ No newline at end of file diff --git a/2025/DreamText_ High Fidelity Scene Text Synthesis/images.zip b/2025/DreamText_ High Fidelity Scene Text Synthesis/images.zip new file mode 100644 index 0000000000000000000000000000000000000000..d7f67423097f8e3053e2174b4630899f92177d6b --- /dev/null +++ b/2025/DreamText_ High Fidelity Scene Text Synthesis/images.zip @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:e30b9395881b2457b42d8943789c062519e85d2429e1606c82728cff009a9407 +size 918044 diff --git a/2025/DreamText_ High Fidelity Scene Text Synthesis/layout.json b/2025/DreamText_ High Fidelity Scene Text Synthesis/layout.json new file mode 100644 index 0000000000000000000000000000000000000000..5857257fa5994e44215081ba7385d09d5f947837 --- /dev/null +++ b/2025/DreamText_ High Fidelity Scene Text Synthesis/layout.json @@ -0,0 +1,6997 @@ +{ + "pdf_info": [ + { + "para_blocks": [ + { + "bbox": [ + 151, + 103, + 460, + 121 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 151, + 103, + 460, + 121 + ], + "spans": [ + { + "bbox": [ + 151, + 103, + 460, + 121 + ], + "type": "text", + "content": "DREAMTEXT: High Fidelity Scene Text Synthesis" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 145, + 141, + 469, + 157 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 145, + 141, + 469, + 157 + ], + "spans": [ + { + "bbox": [ + 145, + 141, + 469, + 157 + ], + "type": "text", + "content": "Yibin Wang" + }, + { + "bbox": [ + 145, + 141, + 469, + 157 + ], + "type": "inline_equation", + "content": "^{1,2}" + }, + { + "bbox": [ + 145, + 141, + 469, + 157 + ], + "type": "text", + "content": ", Weizhong Zhang" + }, + { + "bbox": [ + 145, + 141, + 469, + 157 + ], + "type": "inline_equation", + "content": "^{1,4}" + }, + { + "bbox": [ + 145, + 141, + 469, + 157 + ], + "type": "text", + "content": ", Honghui Xu" + }, + { + "bbox": [ + 145, + 141, + 469, + 157 + ], + "type": "inline_equation", + "content": "^{5}" + }, + { + "bbox": [ + 145, + 141, + 469, + 157 + ], + "type": "text", + "content": ", Cheng Jin" + }, + { + "bbox": [ + 145, + 141, + 469, + 157 + ], + "type": "inline_equation", + "content": "^{1,3\\dagger}" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 186, + 157, + 424, + 171 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 186, + 157, + 424, + 171 + ], + "spans": [ + { + "bbox": [ + 186, + 157, + 424, + 171 + ], + "type": "inline_equation", + "content": "^{1}" + }, + { + "bbox": [ + 186, + 157, + 424, + 171 + ], + "type": "text", + "content": "Fudan University " + }, + { + "bbox": [ + 186, + 157, + 424, + 171 + ], + "type": "inline_equation", + "content": "^{2}" + }, + { + "bbox": [ + 186, + 157, + 424, + 171 + ], + "type": "text", + "content": "Shanghai Innovation Institute" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 123, + 171, + 487, + 186 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 123, + 171, + 487, + 186 + ], + "spans": [ + { + "bbox": [ + 123, + 171, + 487, + 186 + ], + "type": "inline_equation", + "content": "^{3}" + }, + { + "bbox": [ + 123, + 171, + 487, + 186 + ], + "type": "text", + "content": "Innovation Center of Calligraphy and Painting Creation Technology, MCT" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 63, + 185, + 548, + 200 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 63, + 185, + 548, + 200 + ], + "spans": [ + { + "bbox": [ + 63, + 185, + 548, + 200 + ], + "type": "inline_equation", + "content": "^{4}" + }, + { + "bbox": [ + 63, + 185, + 548, + 200 + ], + "type": "text", + "content": "Shanghai Key Laboratory of Intelligent Information Processing " + }, + { + "bbox": [ + 63, + 185, + 548, + 200 + ], + "type": "inline_equation", + "content": "^{5}" + }, + { + "bbox": [ + 63, + 185, + 548, + 200 + ], + "type": "text", + "content": "Zhejiang University of Technology" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 79, + 201, + 530, + 213 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 79, + 201, + 530, + 213 + ], + "spans": [ + { + "bbox": [ + 79, + 201, + 530, + 213 + ], + "type": "text", + "content": "yibinwang1121@163.com, weizhongzhang@fudan.edu.cn, xhh@zjut.edu.cn, jc@fudan.edu.cn" + } + ] + } + ], + "index": 7 + }, + { + "type": "image", + "bbox": [ + 67, + 242, + 541, + 403 + ], + "blocks": [ + { + "bbox": [ + 67, + 242, + 541, + 403 + ], + "lines": [ + { + "bbox": [ + 67, + 242, + 541, + 403 + ], + "spans": [ + { + "bbox": [ + 67, + 242, + 541, + 403 + ], + "type": "image", + "image_path": "6ebdc6a9714dca6a4f1b52ca44fd15efe7bf40df27eece4554c3bf0a39fc7067.jpg" + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 96, + 412, + 514, + 424 + ], + "lines": [ + { + "bbox": [ + 96, + 412, + 514, + 424 + ], + "spans": [ + { + "bbox": [ + 96, + 412, + 514, + 424 + ], + "type": "text", + "content": "Figure 1. Displayed are the results generated using our DREAMTEXT, showcasing its prowess across varied inputs." + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "image_caption" + } + ], + "index": 8 + }, + { + "bbox": [ + 151, + 434, + 200, + 445 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 151, + 434, + 200, + 445 + ], + "spans": [ + { + "bbox": [ + 151, + 434, + 200, + 445 + ], + "type": "text", + "content": "Abstract" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 54, + 460, + 297, + 689 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 54, + 460, + 297, + 689 + ], + "spans": [ + { + "bbox": [ + 54, + 460, + 297, + 689 + ], + "type": "text", + "content": "Scene text synthesis involves rendering specified texts onto arbitrary images. Current methods typically formulate this task in an end-to-end manner but lack effective character-level guidance during training. Besides, their text encoders, pre-trained on a single font type, struggle to adapt to the diverse font styles encountered in practical applications. Consequently, these methods suffer from character distortion, repetition, and absence, particularly in polystylistic scenarios. To this end, this paper proposes DREAMTEXT for high-fidelity scene text synthesis. Our key idea is to reconstruct the diffusion training process, introducing more refined guidance tailored to this task, to expose and rectify the model's attention at the character level and strengthen its learning of text regions. This transformation poses a hybrid optimization challenge, involving both discrete and continuous variables. To effectively tackle this challenge, we employ a heuristic alternate optimization strategy. Meanwhile, we jointly train the text encoder and generator to comprehensively learn and utilize the diverse font present" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 313, + 435, + 556, + 579 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 435, + 556, + 579 + ], + "spans": [ + { + "bbox": [ + 313, + 435, + 556, + 579 + ], + "type": "text", + "content": "in the training dataset. This joint training is seamlessly integrated into the alternate optimization process, fostering a synergistic relationship between learning character embedding and re-estimating character attention. Specifically, in each step, we first encode potential character-generated position information from cross-attention maps into latent character masks. These masks are then utilized to update the representation of specific characters in the current step, which, in turn, enables the generator to correct the character's attention in the subsequent steps. Both qualitative and quantitative results demonstrate the superiority of our method to the state of the art. Our project page is here." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 314, + 607, + 395, + 620 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 314, + 607, + 395, + 620 + ], + "spans": [ + { + "bbox": [ + 314, + 607, + 395, + 620 + ], + "type": "text", + "content": "1. Introduction" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 313, + 629, + 556, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 629, + 556, + 715 + ], + "spans": [ + { + "bbox": [ + 313, + 629, + 556, + 715 + ], + "type": "text", + "content": "Scene text synthesis involves modifying or inserting specified text in arbitrary images while maintaining its natural and realistic appearance. Earlier GAN-based style transfer methods [13, 14, 16, 27] accomplish this by transferring the text style from a reference image to the rendered target text image. However, they are constrained in their capacity to generate text in arbitrary styles, such as font and color. These" + } + ] + } + ], + "index": 14 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "spans": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "type": "text", + "content": "CVF" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 145, + 0, + 496, + 37 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 145, + 0, + 496, + 37 + ], + "spans": [ + { + "bbox": [ + 145, + 0, + 496, + 37 + ], + "type": "text", + "content": "This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 57, + 702, + 135, + 713 + ], + "type": "page_footnote", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 702, + 135, + 713 + ], + "spans": [ + { + "bbox": [ + 57, + 702, + 135, + 713 + ], + "type": "text", + "content": "† Corresponding author." + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "type": "text", + "content": "28555" + } + ] + } + ], + "index": 16 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 0 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 83, + 85, + 529, + 178 + ], + "blocks": [ + { + "bbox": [ + 197, + 70, + 412, + 80 + ], + "lines": [ + { + "bbox": [ + 197, + 70, + 412, + 80 + ], + "spans": [ + { + "bbox": [ + 197, + 70, + 412, + 80 + ], + "type": "text", + "content": "Problem 1: Character repetition and absence" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 83, + 85, + 529, + 178 + ], + "lines": [ + { + "bbox": [ + 83, + 85, + 529, + 178 + ], + "spans": [ + { + "bbox": [ + 83, + 85, + 529, + 178 + ], + "type": "image", + "image_path": "208bd9f60383832fc13a1d2ca4503e20137e067a0add3a23ea91ab20c390de19.jpg" + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 227, + 178, + 383, + 189 + ], + "lines": [ + { + "bbox": [ + 227, + 178, + 383, + 189 + ], + "spans": [ + { + "bbox": [ + 227, + 178, + 383, + 189 + ], + "type": "text", + "content": "Problem 2: Character distortion" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_caption" + } + ], + "index": 1 + }, + { + "type": "image", + "bbox": [ + 83, + 194, + 529, + 285 + ], + "blocks": [ + { + "bbox": [ + 83, + 194, + 529, + 285 + ], + "lines": [ + { + "bbox": [ + 83, + 194, + 529, + 285 + ], + "spans": [ + { + "bbox": [ + 83, + 194, + 529, + 285 + ], + "type": "image", + "image_path": "35252f8ca281af7ac77b74acb845e6df83f2855272fe11c25ac8f977ebd04778.jpg" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 57, + 294, + 550, + 305 + ], + "lines": [ + { + "bbox": [ + 57, + 294, + 550, + 305 + ], + "spans": [ + { + "bbox": [ + 57, + 294, + 550, + 305 + ], + "type": "text", + "content": "Figure 2. Current methods encounter significant challenges, i.e., character repetition and absence (top) and character distortion (bottom)." + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_caption" + } + ], + "index": 3 + }, + { + "bbox": [ + 55, + 316, + 296, + 422 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 316, + 296, + 422 + ], + "spans": [ + { + "bbox": [ + 55, + 316, + 296, + 422 + ], + "type": "text", + "content": "limitations are effectively addressed by utilizing the diffusion model [5, 8, 11, 12, 17, 21, 22], harnessing its inherent prior knowledge and robust capabilities. Nevertheless, these methods still encounter challenges in generating accurate text within polystylistic images due to the inadequate conditional guidance provided by their character-unaware text encoder. To this end, a recent study [29] proposes a character-level text encoder, providing more robust conditional guidance for the diffusion model." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 55, + 426, + 296, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 426, + 296, + 714 + ], + "spans": [ + { + "bbox": [ + 55, + 426, + 296, + 714 + ], + "type": "text", + "content": "Despite their effectiveness, these methods are still difficult to accurately render text within complex scenes. Specifically, these methods such as UDiffText [29] and TextDiffuser [5] solely utilize a single font type to pre-train the text encoder, which is subsequently employed to fine-tune the generator. However, the diverse font styles encountered in practical applications present a significant challenge: the restricted representation domain hampers their ability to render text with unseen font styles, e.g., cases shown in Fig. 2 (bottom). Furthermore, these methods rely on character segmentation masks to supervise characters' attention for position control. We argue this approach has significant limitations: (1) unlike classification tasks with uniquely determined labels, the optimal generation positions for characters can be diverse. Therefore, using the specific masks to rigidly constrain the model may limit its flexibility in estimating optimal positions, hindering its ability to adapt to varied and complex scenarios. Besides, (2) these masks generated by a pre-trained segmentation model [5] are imprecise, tending to over-segment character regions. As a result, the model encounters issues of character repetition and absence as illustrated in Fig. 2 (top left) and (top right) respectively. We suspect these issues stem from characters' attention maps, as they reflect the underlying response to the character's" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 313, + 316, + 555, + 542 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 316, + 555, + 542 + ], + "spans": [ + { + "bbox": [ + 313, + 316, + 555, + 542 + ], + "type": "text", + "content": "final generated position [23, 24]. Therefore, we visualize the attention maps of their problematic results by rendering attention directly onto the images. We observe certain characters' attention misalignment, resulting in character duplication (see Fig. 3 (a) and (c)). Besides, excessive dispersion of attention also leads to character absence (see Fig. 3 (b)). This observation underscores the insufficiency of these methods in guiding the model autonomously estimating optimal character-generated position. Further, we also assess specific methods [5, 29] quantitatively on two datasets to evaluate their proficiency in directing characters' attention toward their optimal generation regions throughout multiple global training steps. Specifically, for each test sample, we extract potential position masks for all characters based on their cross-attention maps from the inference process. Then, we compute the mean Intersection over Union (mIoU) between these masks and ground truth character segmentation masks. Their inoptimal results depicted in Fig. 4 further corroborate our hypothesis." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 313, + 546, + 556, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 546, + 556, + 713 + ], + "spans": [ + { + "bbox": [ + 313, + 546, + 556, + 713 + ], + "type": "text", + "content": "In summary, current methods encounter challenges in accurately rendering text within complex scenes due to their constrained representation domain and synthesizing characters in their ideal regions due to the lack of effective guidance for the model to estimate optimal character-generated position. To this end, this paper proposes DREAMTEXT for high-fidelity text synthesis. We reconstruct the diffusion training process, embedding more refined guidance: we expose and rectify the characters' attention to focus more precisely on their ideal generation areas, and enhance the learning of text regions and character representation by additional constraints. However, this reconstruction inevitably introduces a complex hybrid optimization problem encompassing both discrete and continuous variables. Therefore," + } + ] + } + ], + "index": 8 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 749, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 749, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 749, + 317, + 757 + ], + "type": "text", + "content": "28556" + } + ] + } + ], + "index": 9 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 1 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 57, + 70, + 294, + 310 + ], + "blocks": [ + { + "bbox": [ + 57, + 70, + 294, + 310 + ], + "lines": [ + { + "bbox": [ + 57, + 70, + 294, + 310 + ], + "spans": [ + { + "bbox": [ + 57, + 70, + 294, + 310 + ], + "type": "image", + "image_path": "b05c4cdd7650ce0b011e58a8ec5430088a0bad5fe880c76c564d249393632ff1.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 55, + 319, + 295, + 343 + ], + "lines": [ + { + "bbox": [ + 55, + 319, + 295, + 343 + ], + "spans": [ + { + "bbox": [ + 55, + 319, + 295, + 343 + ], + "type": "text", + "content": "Figure 3. The problematic results rendered by characters' attention maps (AMs)." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "bbox": [ + 54, + 354, + 295, + 556 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 54, + 354, + 295, + 556 + ], + "spans": [ + { + "bbox": [ + 54, + 354, + 295, + 556 + ], + "type": "text", + "content": "we design a heuristic alternate optimization strategy. Concurrently, we jointly train the text encoder and the generator, leveraging the diverse font styles within the training dataset to enrich character representation space. We demonstrate that this joint training can be seamlessly integrated into the heuristic alternate optimization process, facilitating a synergistic interplay between character representations learning and character attention re-estimation. Specifically, in each iteration, the pipeline begins by encoding the potential character-generated position information from cross-attention maps into latent character masks. These masks are subsequently employed to refine the representation of specific characters within the text encoder, facilitating the calibration of the characters' attention in the subsequent step. This iterative process contributes to both learning of better character representation and, as a by-product, explicit guidance for autonomously estimating character position." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 55, + 558, + 297, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 558, + 297, + 714 + ], + "spans": [ + { + "bbox": [ + 55, + 558, + 297, + 714 + ], + "type": "text", + "content": "Notably, the generator may initially struggle to direct attention to the desired generation region for each character, resulting in suboptimal latent masks and impacting training progress. Unlike existing methods [5, 29] that rigidly constrain character attention, we employ a balanced supervision strategy. Initially, we assist in calibrating the attention using character segmentation masks for a warm-up in the earlier stage. Once the model has preliminarily acquired the ability to estimate ideal generation positions, we remove this guidance, allowing the model to autonomous learning iteratively. This strategy strikes a balance between constraining the model and unleashing its flexibility in estimating optimal generation positions. Fig. 4 illustrates the efficacy of our" + } + ] + } + ], + "index": 3 + }, + { + "type": "image", + "bbox": [ + 316, + 69, + 553, + 144 + ], + "blocks": [ + { + "bbox": [ + 316, + 69, + 553, + 144 + ], + "lines": [ + { + "bbox": [ + 316, + 69, + 553, + 144 + ], + "spans": [ + { + "bbox": [ + 316, + 69, + 553, + 144 + ], + "type": "image", + "image_path": "41b6185fbff08675be8f802b118e82b3e53adf16bbdb725d7cb43bf017de6b4c.jpg" + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 313, + 152, + 555, + 208 + ], + "lines": [ + { + "bbox": [ + 313, + 152, + 555, + 208 + ], + "spans": [ + { + "bbox": [ + 313, + 152, + 555, + 208 + ], + "type": "text", + "content": "Figure 4. The mIoU scores of UDiffText, TextDiffuser, and our method on LAION-OCR and SynthText over global training steps. Our method adopts a balanced supervision strategy: we initially use latent character masks to steer the character's attention for a warm-up in the earlier stage and stop guiding after 25,000 steps." + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "image_caption" + } + ], + "index": 4 + }, + { + "bbox": [ + 313, + 215, + 555, + 239 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 215, + 555, + 239 + ], + "spans": [ + { + "bbox": [ + 313, + 215, + 555, + 239 + ], + "type": "text", + "content": "method in continuously improving the generator's ability to estimate the optimal characters' position." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 313, + 239, + 556, + 370 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 239, + 556, + 370 + ], + "spans": [ + { + "bbox": [ + 313, + 239, + 556, + 370 + ], + "type": "text", + "content": "Our contributions are: (1) The proposed DREAMTEXT effectively alleviates the issues of character repetition, absence, and distortion encountered by existing methods. (2) Our heuristic alternate optimization strategy, integrating the joint learning of the text encoder and U-Net, orchestrates a symbiotic relationship between learning character representations and re-estimating character attention. (3) Our balanced supervision strategy strikes a balance between constraining the model and unleashing its flexibility in estimating optimal generation positions. (4) Both qualitative and quantitative results demonstrate the superiority of our method." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 314, + 381, + 400, + 395 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 314, + 381, + 400, + 395 + ], + "spans": [ + { + "bbox": [ + 314, + 381, + 400, + 395 + ], + "type": "text", + "content": "2. Related Work" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 313, + 402, + 556, + 545 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 402, + 556, + 545 + ], + "spans": [ + { + "bbox": [ + 313, + 402, + 556, + 545 + ], + "type": "text", + "content": "Earlier GAN-based methods achieve scene text synthesis by transferring the text style in a reference image to the rendered target text image [10, 16, 18, 20, 27]. Specifically, STEFANN [16] employs a FANnet to edit individual characters and incorporates a placement algorithm to generate the desired word. SRNet [25] and MOSTEL [14], on the other hand, split the task into two main parts: background inpainting and text style transfer. This approach allows for the end-to-end synthesizing of entire words. Despite their simplicity and effectiveness, they are limited in their ability to generate text in arbitrary styles and locations, often producing less natural-looking images." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 313, + 546, + 557, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 546, + 557, + 713 + ], + "spans": [ + { + "bbox": [ + 313, + 546, + 557, + 713 + ], + "type": "text", + "content": "To this end, several diffusion-based approaches have emerged to address these challenges. These approaches harness the robust capabilities of diffusion models to edit or generate scene text, thereby enhancing the quality and diversity of the generated content. DiffSTE [8] proposes a dual encoder structure, comprising a character text encoder and an instruction text encoder. These components are employed for instruction tuning, providing improved control over the backbone network. DiffUTE [4] employs an OCR-based glyph encoder to extract glyph guidance from the rendered glyph image. TextDiffuser [5] concatenates the segmentation mask as conditional input and employs character-aware loss to control the generated characters precisely. However, they still face challenges in generating accurate text" + } + ] + } + ], + "index": 10 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "text", + "content": "28557" + } + ] + } + ], + "index": 11 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 2 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 57, + 70, + 556, + 309 + ], + "blocks": [ + { + "bbox": [ + 57, + 70, + 556, + 309 + ], + "lines": [ + { + "bbox": [ + 57, + 70, + 556, + 309 + ], + "spans": [ + { + "bbox": [ + 57, + 70, + 556, + 309 + ], + "type": "image", + "image_path": "016c3053ed95d0bb11c7831077688a40c8caf28d5bbd776bb086f7c16f4c4442.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 167, + 316, + 443, + 328 + ], + "lines": [ + { + "bbox": [ + 167, + 316, + 443, + 328 + ], + "spans": [ + { + "bbox": [ + 167, + 316, + 443, + 328 + ], + "type": "text", + "content": "Figure 5. An overview of proposed heuristic alternate optimization strategy." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "bbox": [ + 54, + 336, + 295, + 420 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 54, + 336, + 295, + 420 + ], + "spans": [ + { + "bbox": [ + 54, + 336, + 295, + 420 + ], + "type": "text", + "content": "within polystylistic images, primarily due to the insufficient conditional guidance provided by their character-unaware text encoder. Motivated by this, UDiffText [29] replaces the original CLIP text encoder in Stable Diffusion with a character-level text encoder with a unitary font. It also employs local attention loss for position control, relying on ground truth character segmentation maps." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 54, + 420, + 297, + 494 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 54, + 420, + 297, + 494 + ], + "spans": [ + { + "bbox": [ + 54, + 420, + 297, + 494 + ], + "type": "text", + "content": "However, these methods still encounter challenges in accurately rendering text within complex scenes and synthesizing characters in their ideal regions, attributed to their constrained character representation domain and lack of effective guidance for the model to autonomously estimate character-generated position, respectively." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 55, + 502, + 112, + 514 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 502, + 112, + 514 + ], + "spans": [ + { + "bbox": [ + 55, + 502, + 112, + 514 + ], + "type": "text", + "content": "3. Method" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 55, + 522, + 296, + 595 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 522, + 296, + 595 + ], + "spans": [ + { + "bbox": [ + 55, + 522, + 296, + 595 + ], + "type": "text", + "content": "Given an input image " + }, + { + "bbox": [ + 55, + 522, + 296, + 595 + ], + "type": "inline_equation", + "content": "z" + }, + { + "bbox": [ + 55, + 522, + 296, + 595 + ], + "type": "text", + "content": ", a text region specified by a binary mask " + }, + { + "bbox": [ + 55, + 522, + 296, + 595 + ], + "type": "inline_equation", + "content": "B" + }, + { + "bbox": [ + 55, + 522, + 296, + 595 + ], + "type": "text", + "content": " and a text condition " + }, + { + "bbox": [ + 55, + 522, + 296, + 595 + ], + "type": "inline_equation", + "content": "c" + }, + { + "bbox": [ + 55, + 522, + 296, + 595 + ], + "type": "text", + "content": ", scene text synthesis is designed to render the given text onto the input image at the given text region. A natural idea for scene text synthesis is to formulate it into a latent diffusion process [15], which is equivalent to solving the following optimization problem:" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 58, + 601, + 293, + 624 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 58, + 601, + 293, + 624 + ], + "spans": [ + { + "bbox": [ + 58, + 601, + 293, + 624 + ], + "type": "interline_equation", + "content": "\\min _ {(\\theta , \\vartheta)} \\mathcal {L} _ {L D M} \\triangleq \\mathbb {E} _ {\\boldsymbol {z}, c, \\epsilon \\sim N (0, 1), t} \\| \\boldsymbol {\\epsilon} - \\epsilon_ {\\theta} (\\boldsymbol {z} _ {t}, t, \\psi_ {\\vartheta} (c), \\boldsymbol {B}) \\| _ {2} ^ {2},", + "image_path": "15da0ae7d8695947284ebc6a3a71def64679981e829783b49645770a48d3770d.jpg" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 55, + 629, + 297, + 677 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 629, + 297, + 677 + ], + "spans": [ + { + "bbox": [ + 55, + 629, + 297, + 677 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 55, + 629, + 297, + 677 + ], + "type": "inline_equation", + "content": "\\psi_{\\vartheta}" + }, + { + "bbox": [ + 55, + 629, + 297, + 677 + ], + "type": "text", + "content": " is the learnable character-level text encoder initialized by UDifftext [29] and " + }, + { + "bbox": [ + 55, + 629, + 297, + 677 + ], + "type": "inline_equation", + "content": "\\theta" + }, + { + "bbox": [ + 55, + 629, + 297, + 677 + ], + "type": "text", + "content": " is the model parameter vector. However, we argue that such a straightforward formulation could be problematic due to the following reasons:" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 55, + 677, + 297, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 677, + 297, + 714 + ], + "spans": [ + { + "bbox": [ + 55, + 677, + 297, + 714 + ], + "type": "text", + "content": "- Without explicit guidance for rectifying characters' attention, character repetition and absence can inevitably occur in the rendered results as shown in Fig. 2." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 314, + 336, + 555, + 445 + ], + "type": "list", + "angle": 0, + "index": 11, + "blocks": [ + { + "bbox": [ + 314, + 336, + 555, + 396 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 314, + 336, + 555, + 396 + ], + "spans": [ + { + "bbox": [ + 314, + 336, + 555, + 396 + ], + "type": "text", + "content": "- This formulation uniformly measures the mean distance between all pixels, without emphasizing the text regions. Consequently, text-unrelated regions can adversely affect the learning of the text encoder during backpropagation. The experimental analysis is provided in Sec. 4.3." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 314, + 396, + 554, + 445 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 314, + 396, + 554, + 445 + ], + "spans": [ + { + "bbox": [ + 314, + 396, + 554, + 445 + ], + "type": "text", + "content": "- Due to the intricate interactions among the U-Net and text encoder, end-to-end training is particularly challenging as shown in Sec. 4.3. Therefore, a more refined guidance for the training process is necessary." + } + ] + } + ], + "index": 10 + } + ], + "sub_type": "text" + }, + { + "bbox": [ + 313, + 451, + 555, + 594 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 451, + 555, + 594 + ], + "spans": [ + { + "bbox": [ + 313, + 451, + 555, + 594 + ], + "type": "text", + "content": "In this paper, we propose DREAMTEXT and develop a heuristic iterative optimization pipeline to enable efficient training. Our key idea is to explicitly model the potential character positions in attention maps by encoding them into latent character masks in each step. These masks are then used to enhance the learning of the text encoder through our carefully crafted losses, facilitating the calibration of character attention in the subsequent step. This heuristic iterative process allows our model to dynamically alternate between optimizing character embedding and re-estimating character generation positions. The details are presented in the following sections." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 313, + 620, + 524, + 632 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 620, + 524, + 632 + ], + "spans": [ + { + "bbox": [ + 313, + 620, + 524, + 632 + ], + "type": "text", + "content": "3.1. Latent Diffusion with Refined Guidance" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 313, + 644, + 443, + 655 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 644, + 443, + 655 + ], + "spans": [ + { + "bbox": [ + 313, + 644, + 443, + 655 + ], + "type": "text", + "content": "3.1.1. Latent Character Mask" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 313, + 666, + 555, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 666, + 555, + 715 + ], + "spans": [ + { + "bbox": [ + 313, + 666, + 555, + 715 + ], + "type": "text", + "content": "We utilize cross-attention maps to generate latent character masks, extracting characters' latent position information in the current step. Given latent image " + }, + { + "bbox": [ + 313, + 666, + 555, + 715 + ], + "type": "inline_equation", + "content": "z_{t}" + }, + { + "bbox": [ + 313, + 666, + 555, + 715 + ], + "type": "text", + "content": " and text embedding " + }, + { + "bbox": [ + 313, + 666, + 555, + 715 + ], + "type": "inline_equation", + "content": "\\psi_{\\vartheta}(c)" + }, + { + "bbox": [ + 313, + 666, + 555, + 715 + ], + "type": "text", + "content": ", the cross-attention map in layer " + }, + { + "bbox": [ + 313, + 666, + 555, + 715 + ], + "type": "inline_equation", + "content": "l" + }, + { + "bbox": [ + 313, + 666, + 555, + 715 + ], + "type": "text", + "content": " is defined as follows:" + } + ] + } + ], + "index": 15 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 318, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 318, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 318, + 757 + ], + "type": "text", + "content": "28558" + } + ] + } + ], + "index": 16 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 3 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 82, + 70, + 529, + 261 + ], + "blocks": [ + { + "bbox": [ + 82, + 70, + 529, + 261 + ], + "lines": [ + { + "bbox": [ + 82, + 70, + 529, + 261 + ], + "spans": [ + { + "bbox": [ + 82, + 70, + 529, + 261 + ], + "type": "image", + "image_path": "0c95447a95ca4921c541351b13ab80dc92346dc58fafc23aec3b6dea3c217086.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 170, + 267, + 440, + 280 + ], + "lines": [ + { + "bbox": [ + 170, + 267, + 440, + 280 + ], + "spans": [ + { + "bbox": [ + 170, + 267, + 440, + 280 + ], + "type": "text", + "content": "Figure 6. Qualitative comparative results against state-of-the-art methods." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "bbox": [ + 111, + 308, + 296, + 323 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 111, + 308, + 296, + 323 + ], + "spans": [ + { + "bbox": [ + 111, + 308, + 296, + 323 + ], + "type": "interline_equation", + "content": "\\boldsymbol {Q} _ {l} = \\boldsymbol {z} _ {t} \\boldsymbol {W} _ {l} ^ {q}, \\quad \\boldsymbol {K} _ {l} = \\psi_ {\\vartheta} (c) \\boldsymbol {W} _ {l} ^ {k}, \\tag {1}", + "image_path": "c126f71c494756b48eace21d8208681c83480f0d680e1b99e4962aa0cca762cd.jpg" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 130, + 325, + 296, + 354 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 325, + 296, + 354 + ], + "spans": [ + { + "bbox": [ + 130, + 325, + 296, + 354 + ], + "type": "interline_equation", + "content": "\\boldsymbol {A} _ {l} = \\operatorname {s o f t m a x} \\left(\\frac {\\boldsymbol {Q} _ {l} \\boldsymbol {K} _ {l} ^ {T}}{\\sqrt {d}}\\right), \\tag {2}", + "image_path": "90032e48080fc9b0478bdc5fe79d18211503a4b53479312e307ebbc1e388cb51.jpg" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 55, + 362, + 296, + 456 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 362, + 296, + 456 + ], + "spans": [ + { + "bbox": [ + 55, + 362, + 296, + 456 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 55, + 362, + 296, + 456 + ], + "type": "inline_equation", + "content": "\\pmb{W}_l^q" + }, + { + "bbox": [ + 55, + 362, + 296, + 456 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 55, + 362, + 296, + 456 + ], + "type": "inline_equation", + "content": "\\pmb{W}_l^k" + }, + { + "bbox": [ + 55, + 362, + 296, + 456 + ], + "type": "text", + "content": " are learnable parameters and " + }, + { + "bbox": [ + 55, + 362, + 296, + 456 + ], + "type": "inline_equation", + "content": "d" + }, + { + "bbox": [ + 55, + 362, + 296, + 456 + ], + "type": "text", + "content": " is the embedding dimension. The attention maps " + }, + { + "bbox": [ + 55, + 362, + 296, + 456 + ], + "type": "inline_equation", + "content": "\\pmb{A}_l" + }, + { + "bbox": [ + 55, + 362, + 296, + 456 + ], + "type": "text", + "content": " is reshaped to " + }, + { + "bbox": [ + 55, + 362, + 296, + 456 + ], + "type": "inline_equation", + "content": "N \\times H \\times W" + }, + { + "bbox": [ + 55, + 362, + 296, + 456 + ], + "type": "text", + "content": ", where " + }, + { + "bbox": [ + 55, + 362, + 296, + 456 + ], + "type": "inline_equation", + "content": "N" + }, + { + "bbox": [ + 55, + 362, + 296, + 456 + ], + "type": "text", + "content": " represents the number of text tokens such that each slice " + }, + { + "bbox": [ + 55, + 362, + 296, + 456 + ], + "type": "inline_equation", + "content": "\\pmb{A}_l^i \\in \\mathbb{R}^{H \\times W}" + }, + { + "bbox": [ + 55, + 362, + 296, + 456 + ], + "type": "text", + "content": " represents the region attended by the " + }, + { + "bbox": [ + 55, + 362, + 296, + 456 + ], + "type": "inline_equation", + "content": "i" + }, + { + "bbox": [ + 55, + 362, + 296, + 456 + ], + "type": "text", + "content": "-th token. These attention maps are then averaged across all layers of the U-Net to get the mean response " + }, + { + "bbox": [ + 55, + 362, + 296, + 456 + ], + "type": "inline_equation", + "content": "\\bar{\\pmb{A}} = \\frac{1}{L} \\sum_{l=1}^{L} \\pmb{A}_l" + }, + { + "bbox": [ + 55, + 362, + 296, + 456 + ], + "type": "text", + "content": " for the regions attended by character tokens." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 55, + 457, + 296, + 602 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 457, + 296, + 602 + ], + "spans": [ + { + "bbox": [ + 55, + 457, + 296, + 602 + ], + "type": "text", + "content": "To obtain the latent character masks " + }, + { + "bbox": [ + 55, + 457, + 296, + 602 + ], + "type": "inline_equation", + "content": "M \\in \\mathbb{R}^{N \\times H \\times W}" + }, + { + "bbox": [ + 55, + 457, + 296, + 602 + ], + "type": "text", + "content": ", we first apply Gaussian blur to the attention map to execute low-pass filtering " + }, + { + "bbox": [ + 55, + 457, + 296, + 602 + ], + "type": "inline_equation", + "content": "\\mathrm{blur}(\\bar{A})" + }, + { + "bbox": [ + 55, + 457, + 296, + 602 + ], + "type": "text", + "content": ". This step helps mitigate excessive variance in the attended regions, ensuring a more uniform distribution of attention across relevant character regions. Subsequently, we employ a straightforward thresholding process to convert the blurred attention map into binary character masks. This process is encapsulated by the function " + }, + { + "bbox": [ + 55, + 457, + 296, + 602 + ], + "type": "inline_equation", + "content": "f(\\cdot)" + }, + { + "bbox": [ + 55, + 457, + 296, + 602 + ], + "type": "text", + "content": ", which applies a threshold and assigns a value of 1 to pixel values exceeding the threshold, and 0 otherwise. The threshold is calculated as the mean value plus twice the variance of averaged attention maps in this work. That is:" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 80, + 620, + 296, + 652 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 80, + 620, + 296, + 652 + ], + "spans": [ + { + "bbox": [ + 80, + 620, + 296, + 652 + ], + "type": "interline_equation", + "content": "f (\\boldsymbol {X}) = \\left\\{ \\begin{array}{l} 1, \\text {i f} x _ {i, j} > \\operatorname {m e a n} (\\boldsymbol {X}) + 2 \\operatorname {s t d} (\\boldsymbol {X}) \\\\ 0, \\text {o t h e r w i s e} \\end{array} \\right. \\tag {3}", + "image_path": "49f42bc93ff7ccf21efec40226c4247061abbb5d1d7f7a6049d795ce910253c5.jpg" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 55, + 659, + 296, + 672 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 659, + 296, + 672 + ], + "spans": [ + { + "bbox": [ + 55, + 659, + 296, + 672 + ], + "type": "text", + "content": "Therefore, the equation for obtaining " + }, + { + "bbox": [ + 55, + 659, + 296, + 672 + ], + "type": "inline_equation", + "content": "M" + }, + { + "bbox": [ + 55, + 659, + 296, + 672 + ], + "type": "text", + "content": " can be expressed as:" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 137, + 683, + 296, + 697 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 137, + 683, + 296, + 697 + ], + "spans": [ + { + "bbox": [ + 137, + 683, + 296, + 697 + ], + "type": "interline_equation", + "content": "\\boldsymbol {M} = f (\\operatorname {b l u r} (\\bar {\\boldsymbol {A}})). \\tag {4}", + "image_path": "02c1722d7c80e199335c6b2751f484a58d1066d64aa1cd1ad541998c069151c5.jpg" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 67, + 701, + 295, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 67, + 701, + 295, + 714 + ], + "spans": [ + { + "bbox": [ + 67, + 701, + 295, + 714 + ], + "type": "text", + "content": "Based on the latent character masks " + }, + { + "bbox": [ + 67, + 701, + 295, + 714 + ], + "type": "inline_equation", + "content": "M" + }, + { + "bbox": [ + 67, + 701, + 295, + 714 + ], + "type": "text", + "content": ", we employ several" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 313, + 289, + 555, + 313 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 289, + 555, + 313 + ], + "spans": [ + { + "bbox": [ + 313, + 289, + 555, + 313 + ], + "type": "text", + "content": "loss functions to optimize both the text encoder and U-Net iteratively. We will delve into these in the following sections." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 313, + 318, + 440, + 329 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 318, + 440, + 329 + ], + "spans": [ + { + "bbox": [ + 313, + 318, + 440, + 329 + ], + "type": "text", + "content": "3.1.2. Masked Diffusion Loss" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 313, + 333, + 555, + 418 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 333, + 555, + 418 + ], + "spans": [ + { + "bbox": [ + 313, + 333, + 555, + 418 + ], + "type": "text", + "content": "We first extend the Eq. 1 to highlight all desired character tokens in the current step. Specifically, within a text " + }, + { + "bbox": [ + 313, + 333, + 555, + 418 + ], + "type": "inline_equation", + "content": "c" + }, + { + "bbox": [ + 313, + 333, + 555, + 418 + ], + "type": "text", + "content": " containing " + }, + { + "bbox": [ + 313, + 333, + 555, + 418 + ], + "type": "inline_equation", + "content": "k" + }, + { + "bbox": [ + 313, + 333, + 555, + 418 + ], + "type": "text", + "content": " tokens of interest, the diffusion loss of the corresponding pixels obtained from each token's latent character mask is applied with an additional weighting factor " + }, + { + "bbox": [ + 313, + 333, + 555, + 418 + ], + "type": "inline_equation", + "content": "\\gamma" + }, + { + "bbox": [ + 313, + 333, + 555, + 418 + ], + "type": "text", + "content": ". Formally, considering " + }, + { + "bbox": [ + 313, + 333, + 555, + 418 + ], + "type": "inline_equation", + "content": "M_{k} = \\bigvee_{i=1}^{k} M_{i}" + }, + { + "bbox": [ + 313, + 333, + 555, + 418 + ], + "type": "text", + "content": " to the union of the pixels of " + }, + { + "bbox": [ + 313, + 333, + 555, + 418 + ], + "type": "inline_equation", + "content": "k" + }, + { + "bbox": [ + 313, + 333, + 555, + 418 + ], + "type": "text", + "content": " characters, the masked diffusion loss is formulated as," + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 313, + 426, + 563, + 441 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 426, + 563, + 441 + ], + "spans": [ + { + "bbox": [ + 313, + 426, + 563, + 441 + ], + "type": "interline_equation", + "content": "\\mathcal {L} _ {\\text {m a s k}} = \\mathbb {E} _ {z, c, \\epsilon \\sim N (0, 1), t} \\| (1 + \\gamma \\boldsymbol {M} _ {k}) (\\boldsymbol {\\epsilon} - \\epsilon_ {\\theta} (\\boldsymbol {z} _ {t}, t, \\psi_ {\\vartheta} (c))) \\| _ {2} ^ {2}.", + "image_path": "fbd654a75acb4427ca1dbbfd509ab1ab019b6f903e3b9bad80d53d3926c5b942.jpg" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 313, + 448, + 504, + 460 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 448, + 504, + 460 + ], + "spans": [ + { + "bbox": [ + 313, + 448, + 504, + 460 + ], + "type": "text", + "content": "We omit binary mask " + }, + { + "bbox": [ + 313, + 448, + 504, + 460 + ], + "type": "inline_equation", + "content": "B" + }, + { + "bbox": [ + 313, + 448, + 504, + 460 + ], + "type": "text", + "content": " here for the simplicity." + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 313, + 465, + 432, + 475 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 465, + 432, + 475 + ], + "spans": [ + { + "bbox": [ + 313, + 465, + 432, + 475 + ], + "type": "text", + "content": "3.1.3. Cross Attention Loss" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 313, + 480, + 554, + 553 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 480, + 554, + 553 + ], + "spans": [ + { + "bbox": [ + 313, + 480, + 554, + 553 + ], + "type": "text", + "content": "The masked diffusion loss is designed to focus on all desired concept tokens within a prompt. Additionally, to ensure that each token encodes information specific to its corresponding synthesis position, we incorporate the cross-attention loss [2]. This loss encourages tokens to attend exclusively to their corresponding target regions," + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 342, + 559, + 555, + 574 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 342, + 559, + 555, + 574 + ], + "spans": [ + { + "bbox": [ + 342, + 559, + 555, + 574 + ], + "type": "interline_equation", + "content": "\\mathcal {L} _ {a t t n} = \\mathbb {E} _ {\\boldsymbol {z}, c, t} \\| C _ {a t t n} (\\boldsymbol {z} _ {t}, \\psi_ {\\vartheta} (c) _ {i}) - \\boldsymbol {M} _ {i} \\| _ {2} ^ {2}. \\tag {5}", + "image_path": "776c620b17d644ff4bdbc799a88d8f6b779d8484207c9bb14d8aff4fec13d860.jpg" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 313, + 582, + 555, + 617 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 582, + 555, + 617 + ], + "spans": [ + { + "bbox": [ + 313, + 582, + 555, + 617 + ], + "type": "text", + "content": "Here, " + }, + { + "bbox": [ + 313, + 582, + 555, + 617 + ], + "type": "inline_equation", + "content": "C_{attn}(z_t, \\psi_\\vartheta(c)_i)" + }, + { + "bbox": [ + 313, + 582, + 555, + 617 + ], + "type": "text", + "content": " represents the cross-attention map between the visual representation " + }, + { + "bbox": [ + 313, + 582, + 555, + 617 + ], + "type": "inline_equation", + "content": "z_t" + }, + { + "bbox": [ + 313, + 582, + 555, + 617 + ], + "type": "text", + "content": " and the token " + }, + { + "bbox": [ + 313, + 582, + 555, + 617 + ], + "type": "inline_equation", + "content": "\\psi_\\vartheta(c)_i" + }, + { + "bbox": [ + 313, + 582, + 555, + 617 + ], + "type": "text", + "content": ", while " + }, + { + "bbox": [ + 313, + 582, + 555, + 617 + ], + "type": "inline_equation", + "content": "M_i" + }, + { + "bbox": [ + 313, + 582, + 555, + 617 + ], + "type": "text", + "content": " denotes the latent character mask of " + }, + { + "bbox": [ + 313, + 582, + 555, + 617 + ], + "type": "inline_equation", + "content": "\\psi_\\vartheta(c)_i" + }, + { + "bbox": [ + 313, + 582, + 555, + 617 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 313, + 618, + 556, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 618, + 556, + 713 + ], + "spans": [ + { + "bbox": [ + 313, + 618, + 556, + 713 + ], + "type": "text", + "content": "However, the losses mentioned above alone are inadequate for achieving optimal character representation. This is primarily due to the inherent noise in the latent character masks, which frequently results in under- or oversegmentation of the ideal character regions during training. To mitigate the influence of noise and prevent skewed learning of character representation, we introduce additional losses [29] as shown below." + } + ] + } + ], + "index": 19 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "text", + "content": "28559" + } + ] + } + ], + "index": 20 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 4 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 57, + 87, + 553, + 168 + ], + "blocks": [ + { + "bbox": [ + 211, + 71, + 399, + 82 + ], + "lines": [ + { + "bbox": [ + 211, + 71, + 399, + 82 + ], + "spans": [ + { + "bbox": [ + 211, + 71, + 399, + 82 + ], + "type": "text", + "content": "Table 1. Quantitative comparison against baselines." + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 57, + 87, + 553, + 168 + ], + "lines": [ + { + "bbox": [ + 57, + 87, + 553, + 168 + ], + "spans": [ + { + "bbox": [ + 57, + 87, + 553, + 168 + ], + "type": "table", + "html": "
MethodsSeqAcc-ReconSeqAcc-EditingFIDLPIPS
ICDAR13(8ch)ICDAR13TextSegLAION-OCRICDAR13(8ch)ICDAR13TextSegLAION-OCR
CVPR'22 SD-Inpainting [15]0.320.290.110.150.080.070.040.0526.780.0696
arXiv'23 DiffSTE [8]0.450.370.500.410.340.290.470.2751.670.1050
AAAI'23 MOSTEL [14]0.750.680.640.710.350.280.250.4425.090.0605
NIPS'23 TextDiffuser [5]0.870.810.680.800.820.750.660.6432.250.0834
ICLR'24 AnyText [19]0.890.870.810.860.810.790.800.7222.730.0651
ECCV'24 UDiffText [29]0.940.910.930.900.840.830.840.7815.790.0564
DreamText0.950.940.960.930.870.890.910.8812.130.0328
", + "image_path": "1893a91a5b39adc9f2f30341d1e44749dcc6465f6009021479943b7320c6f303.jpg" + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "table_body" + } + ], + "index": 1 + }, + { + "bbox": [ + 55, + 181, + 196, + 192 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 181, + 196, + 192 + ], + "spans": [ + { + "bbox": [ + 55, + 181, + 196, + 192 + ], + "type": "text", + "content": "3.1.4. Cross-modal Aligned Loss" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 55, + 195, + 297, + 256 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 195, + 297, + 256 + ], + "spans": [ + { + "bbox": [ + 55, + 195, + 297, + 256 + ], + "type": "text", + "content": "To obtain accurate and robust embeddings, we introduce an image encoder " + }, + { + "bbox": [ + 55, + 195, + 297, + 256 + ], + "type": "inline_equation", + "content": "\\xi" + }, + { + "bbox": [ + 55, + 195, + 297, + 256 + ], + "type": "text", + "content": ", text head " + }, + { + "bbox": [ + 55, + 195, + 297, + 256 + ], + "type": "inline_equation", + "content": "H_{t}" + }, + { + "bbox": [ + 55, + 195, + 297, + 256 + ], + "type": "text", + "content": ", and visual head " + }, + { + "bbox": [ + 55, + 195, + 297, + 256 + ], + "type": "inline_equation", + "content": "H_{v}" + }, + { + "bbox": [ + 55, + 195, + 297, + 256 + ], + "type": "text", + "content": " to align cross-modal character features. Specifically, we compute the cosine similarity objective between visual and text representations to maximize the alignment between two modalities:" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 104, + 263, + 296, + 290 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 104, + 263, + 296, + 290 + ], + "spans": [ + { + "bbox": [ + 104, + 263, + 296, + 290 + ], + "type": "interline_equation", + "content": "\\mathcal {L} _ {\\text {a l i g n}} = \\frac {\\left\\langle H _ {t} (\\mathbf {y}) , H _ {v} (\\xi (\\mathbf {I})) \\right\\rangle}{\\| H _ {t} (\\mathbf {y}) \\| _ {2} \\cdot \\| H _ {v} (\\xi (\\mathbf {I})) \\| _ {2}}. \\tag {6}", + "image_path": "f2736a9507d3dcb8b19477a0991fb7fc4a14dcb65e9a4ef5d4c6a7a774bee884.jpg" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 55, + 292, + 296, + 351 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 292, + 296, + 351 + ], + "spans": [ + { + "bbox": [ + 55, + 292, + 296, + 351 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 55, + 292, + 296, + 351 + ], + "type": "inline_equation", + "content": "I" + }, + { + "bbox": [ + 55, + 292, + 296, + 351 + ], + "type": "text", + "content": " represents the text image segmented via the bounding box from datasets and " + }, + { + "bbox": [ + 55, + 292, + 296, + 351 + ], + "type": "inline_equation", + "content": "\\langle ,\\rangle" + }, + { + "bbox": [ + 55, + 292, + 296, + 351 + ], + "type": "text", + "content": " presents the inner product of two vectors. Note that to alleviate noise stemming from background and color variations, we preprocess the text image by converting it to grayscale." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 55, + 351, + 296, + 422 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 351, + 296, + 422 + ], + "spans": [ + { + "bbox": [ + 55, + 351, + 296, + 422 + ], + "type": "text", + "content": "Through training, the text head learns to map the whole information of multiple characters into the text feature space. Similarly, the image head maps the overall visual text information into the visual feature space. This effectively ensures the overall textual and visual representations correspond, capturing high-level cross-modal consistency." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 55, + 427, + 162, + 437 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 427, + 162, + 437 + ], + "spans": [ + { + "bbox": [ + 55, + 427, + 162, + 437 + ], + "type": "text", + "content": "3.1.5. Character Id Loss" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 55, + 441, + 296, + 502 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 441, + 296, + 502 + ], + "spans": [ + { + "bbox": [ + 55, + 441, + 296, + 502 + ], + "type": "text", + "content": "Additionally, we introduce character Id loss to ensure that the learned embeddings are highly distinguishable. To be precise, a multi-label classification head " + }, + { + "bbox": [ + 55, + 441, + 296, + 502 + ], + "type": "inline_equation", + "content": "H_{l}" + }, + { + "bbox": [ + 55, + 441, + 296, + 502 + ], + "type": "text", + "content": " is introduced to predict character indices from text embeddings " + }, + { + "bbox": [ + 55, + 441, + 296, + 502 + ], + "type": "inline_equation", + "content": "y" + }, + { + "bbox": [ + 55, + 441, + 296, + 502 + ], + "type": "text", + "content": ". Then, the cross-entropy objective is formulated," + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 107, + 506, + 296, + 540 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 107, + 506, + 296, + 540 + ], + "spans": [ + { + "bbox": [ + 107, + 506, + 296, + 540 + ], + "type": "interline_equation", + "content": "\\mathcal {L} _ {i d} = - \\sum_ {i = 1} ^ {N} \\sum_ {j = 1} ^ {K} l _ {i, j} \\log \\left(H _ {l} (\\mathbf {y}) _ {j}\\right), \\tag {7}", + "image_path": "2865fd4642be2694fab3c5a1113188e4a25d97419338f0356d67427f15cd533b.jpg" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 55, + 544, + 296, + 604 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 544, + 296, + 604 + ], + "spans": [ + { + "bbox": [ + 55, + 544, + 296, + 604 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 55, + 544, + 296, + 604 + ], + "type": "inline_equation", + "content": "N" + }, + { + "bbox": [ + 55, + 544, + 296, + 604 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 55, + 544, + 296, + 604 + ], + "type": "inline_equation", + "content": "K" + }, + { + "bbox": [ + 55, + 544, + 296, + 604 + ], + "type": "text", + "content": " represent the number of characters and the possible indices respectively while " + }, + { + "bbox": [ + 55, + 544, + 296, + 604 + ], + "type": "inline_equation", + "content": "l" + }, + { + "bbox": [ + 55, + 544, + 296, + 604 + ], + "type": "text", + "content": " denotes the ground truth labels. This loss aggregates over all characters in the target text, ensuring the text encoder produces distinguishable embeddings for each character." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 55, + 605, + 295, + 628 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 605, + 295, + 628 + ], + "spans": [ + { + "bbox": [ + 55, + 605, + 295, + 628 + ], + "type": "text", + "content": "The complete objective of our training strategy can be expressed as" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 89, + 633, + 296, + 647 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 89, + 633, + 296, + 647 + ], + "spans": [ + { + "bbox": [ + 89, + 633, + 296, + 647 + ], + "type": "interline_equation", + "content": "\\mathcal {L} = \\mathcal {L} _ {\\text {m a s k}} + \\alpha \\mathcal {L} _ {\\text {a t t n}} + \\beta (\\mathcal {L} _ {\\text {a l i g n}} + \\mathcal {L} _ {i d}). \\tag {8}", + "image_path": "0de0c7796d56faffe7fc63c458a6b1f27f4d0135f091c83540a71aa449f45b3e.jpg" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 55, + 657, + 182, + 670 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 657, + 182, + 670 + ], + "spans": [ + { + "bbox": [ + 55, + 657, + 182, + 670 + ], + "type": "text", + "content": "3.2. Optimization Strategy" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 55, + 674, + 225, + 686 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 674, + 225, + 686 + ], + "spans": [ + { + "bbox": [ + 55, + 674, + 225, + 686 + ], + "type": "text", + "content": "3.2.1. Heuristic Alternate Optimization" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 55, + 689, + 297, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 689, + 297, + 713 + ], + "spans": [ + { + "bbox": [ + 55, + 689, + 297, + 713 + ], + "type": "text", + "content": "Our objective involves numerous discrete variables, rendering it non-differentiable and making vanilla SGD invalid." + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 313, + 181, + 555, + 348 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 181, + 555, + 348 + ], + "spans": [ + { + "bbox": [ + 313, + 181, + 555, + 348 + ], + "type": "text", + "content": "Thus, we employ a heuristic alternate optimization strategy, utilizing latent character masks that encapsulate potential synthesis position information, to facilitate a symbiotic relationship between text encoder and U-Net. Specifically, we perform an alternating update between the tokens given the latent character masks and the masks themselves. During optimization, the masks are computed by Eq. 4, while for other parameters, we fix the masks and calculate gradients accordingly. Using the losses above, we first optimize the representation of specific characters given masks in each iteration, which, in turn, enables the generator to rectify the character's attention in subsequent steps. This process allows model to dynamically alternate between optimizing character embeddings and re-estimating character masks." + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 313, + 356, + 537, + 369 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 356, + 537, + 369 + ], + "spans": [ + { + "bbox": [ + 313, + 356, + 537, + 369 + ], + "type": "text", + "content": "3.2.2. Balanced Supervision on Character Attention" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 313, + 371, + 556, + 540 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 371, + 556, + 540 + ], + "spans": [ + { + "bbox": [ + 313, + 371, + 556, + 540 + ], + "type": "text", + "content": "The initial training phase may challenge the generator in directing attention to the intended generation region for each character. Unlike existing methods that excessively constrain character attention which inevitably limits the model's flexibility, we adopt a balanced supervision. Initially, we guide attention calibration by cross-entropy objective between latent character masks and character segmentation masks from datasets similar to Eq. 7 for a warm-up in the earlier stage. Once the model preliminarily acquires the capability to estimate ideal generation positions (approximately 25k steps as shown in Fig. 4), we cease guidance, enabling it to engage in autonomous iterative learning. This strategy balances constraining the model and unleashing its flexibility in estimating characters' optimal generation positions." + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 313, + 552, + 395, + 565 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 552, + 395, + 565 + ], + "spans": [ + { + "bbox": [ + 313, + 552, + 395, + 565 + ], + "type": "text", + "content": "4. Experiments" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 313, + 572, + 447, + 584 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 572, + 447, + 584 + ], + "spans": [ + { + "bbox": [ + 313, + 572, + 447, + 584 + ], + "type": "text", + "content": "4.1. Implementation Details" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 313, + 590, + 379, + 601 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 590, + 379, + 601 + ], + "spans": [ + { + "bbox": [ + 313, + 590, + 379, + 601 + ], + "type": "text", + "content": "4.1.1. Datasets" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 313, + 605, + 555, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 605, + 555, + 713 + ], + "spans": [ + { + "bbox": [ + 313, + 605, + 555, + 713 + ], + "type": "text", + "content": "Several datasets are used in our experiments: (1) SynthText [6] comprises 800,000 images containing around 8 million synthetic word instances. Each text instance is annotated with its corresponding text string, along with word-level and character-level bounding boxes. (2) LAION-OCR [5] contains 9,194,613 filtered high-quality text images including advertisements, notes, posters, covers, memes, logos, etc. (3) ICDAR13 [9] is widely recognized as the benchmark dataset for evaluating near-horizontal text detection, which consists" + } + ] + } + ], + "index": 22 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 318, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 318, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 318, + 757 + ], + "type": "text", + "content": "28560" + } + ] + } + ], + "index": 23 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 5 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 82, + 71, + 529, + 264 + ], + "blocks": [ + { + "bbox": [ + 82, + 71, + 529, + 264 + ], + "lines": [ + { + "bbox": [ + 82, + 71, + 529, + 264 + ], + "spans": [ + { + "bbox": [ + 82, + 71, + 529, + 264 + ], + "type": "image", + "image_path": "fe9665a40efabaf5f6e9303953e02affd1798d3abb6d02842a27edd39ddb7132.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 170, + 271, + 440, + 283 + ], + "lines": [ + { + "bbox": [ + 170, + 271, + 440, + 283 + ], + "spans": [ + { + "bbox": [ + 170, + 271, + 440, + 283 + ], + "type": "text", + "content": "Figure 7. Qualitative comparative results against state-of-the-art methods." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "bbox": [ + 55, + 293, + 296, + 388 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 293, + 296, + 388 + ], + "spans": [ + { + "bbox": [ + 55, + 293, + 296, + 388 + ], + "type": "text", + "content": "of 233 test images used for evaluation purposes. (4) TextSeg [26] comprises 4,024 real-world text images sourced from various sources such as posters, greeting cards, book covers, logos, road signs, billboards, digital designs, handwritten notes, and more. Note that only SynthText and LAION-OCR datasets provide the character segmentation maps. We train our model using the training subsets of these datasets and randomly select 100 images from test subsets for testing." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 55, + 396, + 187, + 409 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 396, + 187, + 409 + ], + "spans": [ + { + "bbox": [ + 55, + 396, + 187, + 409 + ], + "type": "text", + "content": "4.1.2. Training Configurations" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 55, + 412, + 296, + 521 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 412, + 296, + 521 + ], + "spans": [ + { + "bbox": [ + 55, + 412, + 296, + 521 + ], + "type": "text", + "content": "We train our model based on the pre-trained checkpoint of SD-v2.0 inpainting version and text encoder in [29]. The pre-trained Vision Transformer [1] is used as our image encoder. The model is fine-tuned using 4 NVIDIA A100 GPUs on LAION-OCR for 200k steps, SynthText for 150k steps, TextSeg for 50k steps, and ICDAR13 for an additional 10k steps. We set " + }, + { + "bbox": [ + 55, + 412, + 296, + 521 + ], + "type": "inline_equation", + "content": "\\alpha" + }, + { + "bbox": [ + 55, + 412, + 296, + 521 + ], + "type": "text", + "content": " to 0.01, and " + }, + { + "bbox": [ + 55, + 412, + 296, + 521 + ], + "type": "inline_equation", + "content": "\\beta" + }, + { + "bbox": [ + 55, + 412, + 296, + 521 + ], + "type": "text", + "content": " to 0.001. Additionally, we utilize a batch size of 16 and a learning rate of " + }, + { + "bbox": [ + 55, + 412, + 296, + 521 + ], + "type": "inline_equation", + "content": "5 \\times 10^{-5}" + }, + { + "bbox": [ + 55, + 412, + 296, + 521 + ], + "type": "text", + "content": ". The inference time of our DREAMTEXT is 8.5 seconds only." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 55, + 529, + 130, + 540 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 529, + 130, + 540 + ], + "spans": [ + { + "bbox": [ + 55, + 529, + 130, + 540 + ], + "type": "text", + "content": "4.1.3. Evaluation" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 55, + 544, + 295, + 676 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 544, + 295, + 676 + ], + "spans": [ + { + "bbox": [ + 55, + 544, + 295, + 676 + ], + "type": "text", + "content": "Text accuracy: We assess all methods on two tasks: scene text reconstruction and scene text editing [29]. For the scene text reconstruction task, we use the models to reconstruct the text image using the provided ground truth text label and binary mask. In the scene text editing task, we replace the original text in each image with a random word and evaluate the models by generating images containing the edited text. We use an off-the-shelf scene text recognition (STR) model [3] to identify the rendered text and then evaluate word-level correctness using sequence accuracy (SeqAcc) by comparing the STR result with the ground truth." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 55, + 677, + 296, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 677, + 296, + 713 + ], + "spans": [ + { + "bbox": [ + 55, + 677, + 296, + 713 + ], + "type": "text", + "content": "Image quality: we utilize Fréchet Inception Distance (FID) [7] to quantify the distance between the text images in the dataset and the generated images. Additionally, the Learned" + } + ] + } + ], + "index": 7 + }, + { + "type": "table", + "bbox": [ + 316, + 312, + 553, + 398 + ], + "blocks": [ + { + "bbox": [ + 362, + 291, + 505, + 301 + ], + "lines": [ + { + "bbox": [ + 362, + 291, + 505, + 301 + ], + "spans": [ + { + "bbox": [ + 362, + 291, + 505, + 301 + ], + "type": "text", + "content": "Table 2. Accumulated results of losses." + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 316, + 312, + 553, + 398 + ], + "lines": [ + { + "bbox": [ + 316, + 312, + 553, + 398 + ], + "spans": [ + { + "bbox": [ + 316, + 312, + 553, + 398 + ], + "type": "table", + "html": "
SettingAverage SeqAccFIDLPIPS
ReconEditing
Base0.2180.06026.780.0696
+Mask0.4250.25923.210.0528
+Attn0.6980.53219.720.0483
+Align0.8840.80115.410.0392
+Id0.9400.88712.130.0328
", + "image_path": "bf3818e8e2f7df03057767a530d4af9d31480bace7282be36512207d36d8dffe.jpg" + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "table_body" + } + ], + "index": 9 + }, + { + "bbox": [ + 313, + 410, + 555, + 434 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 410, + 555, + 434 + ], + "spans": [ + { + "bbox": [ + 313, + 410, + 555, + 434 + ], + "type": "text", + "content": "Perceptual Image Patch Similarity (LPIPS) [28] is utilized as an additional metric." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 313, + 443, + 372, + 455 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 443, + 372, + 455 + ], + "spans": [ + { + "bbox": [ + 313, + 443, + 372, + 455 + ], + "type": "text", + "content": "4.2. Results" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 313, + 461, + 555, + 616 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 461, + 555, + 616 + ], + "spans": [ + { + "bbox": [ + 313, + 461, + 555, + 616 + ], + "type": "text", + "content": "Quantitatively: In Tab. 6, we present a quantitative analysis comparing DREAMTEXT against baseline methods. Our method demonstrates a significant advantage across all quantitative metrics, highlighting the exceptional visual quality of the scene text images it generates. Notably, when compared to the previous state-of-the-art model UDiffText, DREAMTEXT outperforms it by 3.66 in terms of FID. While UDiffText and AnyText excel in image quality compared to other methods, except DREAMTEXT, they tend to perform less satisfactorily in terms of sequence accuracy, especially in text editing. This discrepancy may be attributed to its restricted representation domain and deflected attention, resulting in less accurate text rendering within complex scenes." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 313, + 617, + 556, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 617, + 556, + 713 + ], + "spans": [ + { + "bbox": [ + 313, + 617, + 556, + 713 + ], + "type": "text", + "content": "Qualitatively: As depicted in Fig. 6, our method outperforms other baselines in synthesizing more coherent and accurate text within polystylistic scenes. For example, in Fig. 6 (3rd row), MOSTEL and Stable Diffusion fail to generate the text, while TextDiffuser and UDiffText encounter the issue of character absence. Similarly, in Fig. 6 (1st row), all baselines fail to generate the text on the provided poster. Moreover, we further compare our method with the latest ap" + } + ] + } + ], + "index": 13 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "text", + "content": "28561" + } + ] + } + ], + "index": 14 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 6 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 56, + 92, + 293, + 293 + ], + "blocks": [ + { + "bbox": [ + 112, + 71, + 238, + 82 + ], + "lines": [ + { + "bbox": [ + 112, + 71, + 238, + 82 + ], + "spans": [ + { + "bbox": [ + 112, + 71, + 238, + 82 + ], + "type": "text", + "content": "Table 3. Hyperparameter analysis." + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 56, + 92, + 293, + 293 + ], + "lines": [ + { + "bbox": [ + 56, + 92, + 293, + 293 + ], + "spans": [ + { + "bbox": [ + 56, + 92, + 293, + 293 + ], + "type": "table", + "html": "
α\n(β=0)Average SeqAccβ\n(α=0.01)Average SeqAcc
ReconEditingReconEditing
10.6720.55810.8640.766
0.10.7240.6150.10.8960.817
0.010.7480.6230.010.9210.864
0.0010.7530.6040.0010.9400.887
inputstep 0step 0.2Tstep 0.4Tstep 0.8Tstep T
SUNSHINE\nSTATE\nOF MINDSUNSHINE\nSTATE\nOF MINDSUNSHINE\nSTATE\nOF MINDSUNSHINE\nSTATE\nOF MINDSUNSHINE\nSTATE\nOF MINDSUNSHINE\nSTATE\nOF MIND
The Bride's CrewThe Bride's\nCrewThe Bride's\nCrewThe Bride's\nCrewThe Bride's\nCrewThe Bride's\nCrew
WILBUR\nSMITHWILBUR\nSMITHWILBUR\nSMITHWILBUR\nSMITHWILBUR\nSMITHWILBUR\nSMITH
BLUE\nHORIZONBLUE\nHORIZONBLUE\nHORIZONBLUE\nHORIZONBLUE\nHORIZONBLUE\nHORIZON
", + "image_path": "78aef0913cf99b00fc5319c19312619df5671a706a1991fc3d74f43e673e9af8.jpg" + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "table_body" + } + ], + "index": 1 + }, + { + "bbox": [ + 55, + 300, + 295, + 323 + ], + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 300, + 295, + 323 + ], + "spans": [ + { + "bbox": [ + 55, + 300, + 295, + 323 + ], + "type": "text", + "content": "Figure 8. The visualized attention results of all characters across several steps during training." + } + ] + } + ], + "index": 2, + "type": "text" + }, + { + "bbox": [ + 55, + 337, + 296, + 387 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 337, + 296, + 387 + ], + "spans": [ + { + "bbox": [ + 55, + 337, + 296, + 387 + ], + "type": "text", + "content": "proach, UDiffText, using the samples provided in its original paper for additional qualitative evaluation. The visualization results are presented in Fig. 7, offering additional evidence of our method's superiority in generating high-fidelity text." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 55, + 395, + 282, + 407 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 395, + 282, + 407 + ], + "spans": [ + { + "bbox": [ + 55, + 395, + 282, + 407 + ], + "type": "text", + "content": "4.3. Ablation Studies and Effectiveness Analysis" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 54, + 413, + 295, + 544 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 54, + 413, + 295, + 544 + ], + "spans": [ + { + "bbox": [ + 54, + 413, + 295, + 544 + ], + "type": "text", + "content": "Ablation on Losses: We employ SD-v2.0 inpainting version as the base method. As shown in Tab. 2, we observe a significant drop in performance across all metrics when these losses are removed. Additionally, these results highlight that the masked diffusion loss and cross-attention loss alone are insufficient for achieving satisfactory character representation. This may be attributed to the inherent noise in the latent character masks. To sum up, these ablation results underscore the effectiveness of each objective within our heuristic alternate optimization strategy in improving various aspects of synthesized scene texts." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 54, + 545, + 295, + 664 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 54, + 545, + 295, + 664 + ], + "spans": [ + { + "bbox": [ + 54, + 545, + 295, + 664 + ], + "type": "text", + "content": "Choice of " + }, + { + "bbox": [ + 54, + 545, + 295, + 664 + ], + "type": "inline_equation", + "content": "\\alpha" + }, + { + "bbox": [ + 54, + 545, + 295, + 664 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 54, + 545, + 295, + 664 + ], + "type": "inline_equation", + "content": "\\beta" + }, + { + "bbox": [ + 54, + 545, + 295, + 664 + ], + "type": "text", + "content": ": In our experiments, we vary the values of " + }, + { + "bbox": [ + 54, + 545, + 295, + 664 + ], + "type": "inline_equation", + "content": "\\alpha" + }, + { + "bbox": [ + 54, + 545, + 295, + 664 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 54, + 545, + 295, + 664 + ], + "type": "inline_equation", + "content": "\\beta" + }, + { + "bbox": [ + 54, + 545, + 295, + 664 + ], + "type": "text", + "content": " within the range [0.001, 0.01, 0.1, 1]. The results are presented in Tab. 3. Initially, we focus on exploring the impact of " + }, + { + "bbox": [ + 54, + 545, + 295, + 664 + ], + "type": "inline_equation", + "content": "\\alpha" + }, + { + "bbox": [ + 54, + 545, + 295, + 664 + ], + "type": "text", + "content": " while setting " + }, + { + "bbox": [ + 54, + 545, + 295, + 664 + ], + "type": "inline_equation", + "content": "\\beta = 0" + }, + { + "bbox": [ + 54, + 545, + 295, + 664 + ], + "type": "text", + "content": ". We find that the optimal performance in text reconstruction occurs when " + }, + { + "bbox": [ + 54, + 545, + 295, + 664 + ], + "type": "inline_equation", + "content": "\\alpha = 0.001" + }, + { + "bbox": [ + 54, + 545, + 295, + 664 + ], + "type": "text", + "content": ". However, this choice yields poor results in editing tasks. Consequently, we opt for a more balanced approach, selecting " + }, + { + "bbox": [ + 54, + 545, + 295, + 664 + ], + "type": "inline_equation", + "content": "\\alpha = 0.01" + }, + { + "bbox": [ + 54, + 545, + 295, + 664 + ], + "type": "text", + "content": ". Subsequently, with " + }, + { + "bbox": [ + 54, + 545, + 295, + 664 + ], + "type": "inline_equation", + "content": "\\alpha" + }, + { + "bbox": [ + 54, + 545, + 295, + 664 + ], + "type": "text", + "content": " fixed at 0.01, we investigate the effect of " + }, + { + "bbox": [ + 54, + 545, + 295, + 664 + ], + "type": "inline_equation", + "content": "\\beta" + }, + { + "bbox": [ + 54, + 545, + 295, + 664 + ], + "type": "text", + "content": ". We find that the optimal performance is achieved when " + }, + { + "bbox": [ + 54, + 545, + 295, + 664 + ], + "type": "inline_equation", + "content": "\\beta = 0.001" + }, + { + "bbox": [ + 54, + 545, + 295, + 664 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 55, + 665, + 296, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 665, + 296, + 713 + ], + "spans": [ + { + "bbox": [ + 55, + 665, + 296, + 713 + ], + "type": "text", + "content": "Effectiveness of Heuristic Alternate Optimization: We analyze the efficacy of our strategy by visualizing the evolution of character attention maps across several training steps. As illustrated in Fig. 8, the model exhibits deflected attention" + } + ] + } + ], + "index": 7 + }, + { + "type": "table", + "bbox": [ + 317, + 92, + 553, + 167 + ], + "blocks": [ + { + "bbox": [ + 343, + 71, + 524, + 82 + ], + "lines": [ + { + "bbox": [ + 343, + 71, + 524, + 82 + ], + "spans": [ + { + "bbox": [ + 343, + 71, + 524, + 82 + ], + "type": "text", + "content": "Table 4. Ablation study results on warm-up steps." + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 317, + 92, + 553, + 167 + ], + "lines": [ + { + "bbox": [ + 317, + 92, + 553, + 167 + ], + "spans": [ + { + "bbox": [ + 317, + 92, + 553, + 167 + ], + "type": "table", + "html": "
Average SeqAcc
Recon ↑Editing ↑FID ↓mIoU ↑
15k0.8840.85213.820.681
20k0.9130.87313.380.692
25k0.9400.88712.130.722
30k0.9210.89113.240.703
", + "image_path": "945091872fa1deea466d061cc9f744752a461642f324cb20c7da21cc48629947.jpg" + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "table_body" + } + ], + "index": 9 + }, + { + "bbox": [ + 313, + 176, + 555, + 259 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 176, + 555, + 259 + ], + "spans": [ + { + "bbox": [ + 313, + 176, + 555, + 259 + ], + "type": "text", + "content": "initially but progressively corrects it in subsequent training steps. Ultimately, attention becomes concentrated on the desired positions for all characters, indicating the effectiveness of our approach. Besides, Fig. 4 also illustrates the efficacy of this strategy in continuously improving the generator's ability to estimate the optimal characters' position by autonomously rectifying the attention during training." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 313, + 261, + 556, + 416 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 261, + 556, + 416 + ], + "spans": [ + { + "bbox": [ + 313, + 261, + 556, + 416 + ], + "type": "text", + "content": "Choice of Warm-up Steps: We conduct an ablation experiment to analyze the impact of different warm-up steps on model performance. The results, presented in Tab. 4, demonstrate how varying the number of warm-up steps affects the performance. As the number of warm-up steps increases from 15k to 25k, we observe consistent improvements across all metrics. However, at 30k steps, while the editing accuracy slightly improves, the FID score worsens, suggesting a potential trade-off between image fidelity and editing capability. Overall, these findings highlight that 25k warm-up steps strike the best balance between reconstruction accuracy, editing performance, and image quality, making it the optimal choice for our method." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 314, + 427, + 388, + 441 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 314, + 427, + 388, + 441 + ], + "spans": [ + { + "bbox": [ + 314, + 427, + 388, + 441 + ], + "type": "text", + "content": "5. Conclusion" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 313, + 449, + 556, + 605 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 449, + 556, + 605 + ], + "spans": [ + { + "bbox": [ + 313, + 449, + 556, + 605 + ], + "type": "text", + "content": "This paper presents DREAMTEXT, a novel approach for high-fidelity scene text synthesis. We reconstruct the diffusion training process to expose and rectify the model's attention at the character level while strengthening its learning of text regions. This inevitably involves a hybrid optimization challenge, combining discrete and continuous variables, which we effectively address through our heuristic alternate optimization strategy. Meanwhile, we jointly train the text encoder and generator to comprehensively learn and utilize diverse fonts present in the training dataset, fostering a synergistic relationship between learning character embeddings and re-estimating character attention. Both qualitative and quantitative results highlight the superiority of our method." + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 313, + 616, + 424, + 630 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 616, + 424, + 630 + ], + "spans": [ + { + "bbox": [ + 313, + 616, + 424, + 630 + ], + "type": "text", + "content": "6. Acknowledgments" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 313, + 635, + 556, + 712 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 635, + 556, + 712 + ], + "spans": [ + { + "bbox": [ + 313, + 635, + 556, + 712 + ], + "type": "text", + "content": "This work was supported by the National Nature Science Foundation of China (62472097), AI for Science Foundation of Fudan University (FudanX24AI028), and Shanghai Innovation Institute. The computations in this research were performed on the CFFF platform of Fudan University. Cheng Jin is the corresponding author of this work." + } + ] + } + ], + "index": 15 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "type": "text", + "content": "28562" + } + ] + } + ], + "index": 16 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 7 + }, + { + "para_blocks": [ + { + "bbox": [ + 56, + 72, + 115, + 83 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 72, + 115, + 83 + ], + "spans": [ + { + "bbox": [ + 56, + 72, + 115, + 83 + ], + "type": "text", + "content": "References" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 57, + 91, + 296, + 713 + ], + "type": "list", + "angle": 0, + "index": 16, + "blocks": [ + { + "bbox": [ + 61, + 91, + 294, + 113 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 61, + 91, + 294, + 113 + ], + "spans": [ + { + "bbox": [ + 61, + 91, + 294, + 113 + ], + "type": "text", + "content": "[1] Rowel Atienza. Vision transformer for fast and efficient scene text recognition. In ICDAR, pages 319-334, 2021. 7" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 61, + 114, + 295, + 158 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 61, + 114, + 295, + 158 + ], + "spans": [ + { + "bbox": [ + 61, + 114, + 295, + 158 + ], + "type": "text", + "content": "[2] Omri Avrahami, Kfir Aberman, Ohad Fried, Daniel Cohen-Or, and Dani Lischinski. Break-a-scene: Extracting multiple concepts from a single image. In SIGGRAPH Asia, pages 1–12, 2023. 5" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 62, + 160, + 295, + 192 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 160, + 295, + 192 + ], + "spans": [ + { + "bbox": [ + 62, + 160, + 295, + 192 + ], + "type": "text", + "content": "[3] Darwin Bautista and Rowel Atienza. Scene text recognition with permuted autoregressive sequence models. In ECCV, pages 178-196, 2022. 7" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 62, + 194, + 295, + 236 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 194, + 295, + 236 + ], + "spans": [ + { + "bbox": [ + 62, + 194, + 295, + 236 + ], + "type": "text", + "content": "[4] Haoxing Chen, Zhuoer Xu, Zhangxuan Gu, Yaohui Li, Changhua Meng, Huijia Zhu, Weiqiang Wang, et al. Diffuse: Universal text editing diffusion model. In NeurIPS, 2024. 3" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 62, + 239, + 295, + 271 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 239, + 295, + 271 + ], + "spans": [ + { + "bbox": [ + 62, + 239, + 295, + 271 + ], + "type": "text", + "content": "[5] Jingye Chen, Yupan Huang, Tengchao Lv, Lei Cui, Qifeng Chen, and Furu Wei. Textdiffuser: Diffusion models as text painters. In NeurIPS, 2023. 2, 3, 6" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 62, + 274, + 295, + 305 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 274, + 295, + 305 + ], + "spans": [ + { + "bbox": [ + 62, + 274, + 295, + 305 + ], + "type": "text", + "content": "[6] Ankush Gupta, Andrea Vedaldi, and Andrew Zisserman. Synthetic data for text localisation in natural images. In CVPR, pages 2315-2324, 2016. 6" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 62, + 308, + 295, + 350 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 308, + 295, + 350 + ], + "spans": [ + { + "bbox": [ + 62, + 308, + 295, + 350 + ], + "type": "text", + "content": "[7] Martin Heusel, Hubert Ramsauer, Thomas Unterthiner, Bernhard Nessler, and Sepp Hochreiter. Gans trained by a two time-scale update rule converge to a local nash equilibrium. In NeurIPS, 2017. 7" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 62, + 353, + 295, + 396 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 353, + 295, + 396 + ], + "spans": [ + { + "bbox": [ + 62, + 353, + 295, + 396 + ], + "type": "text", + "content": "[8] Jiabao Ji, Guanhua Zhang, Zhaowen Wang, Bairu Hou, Zhifei Zhang, Brian Price, and Shiyu Chang. Improving diffusion models for scene text editing with dual encoders. arXiv preprint arXiv:2304.05568, 2023. 2, 3, 6" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 62, + 398, + 295, + 453 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 398, + 295, + 453 + ], + "spans": [ + { + "bbox": [ + 62, + 398, + 295, + 453 + ], + "type": "text", + "content": "[9] Dimosthenis Karatzas, Faisal Shafait, Seiichi Uchida, Masakazu Iwamura, Lluis Gomez i Bigorda, Sergi Robles Mestre, Joan Mas, David Fernandez Mota, Jon Almazan Almazan, and Lluis Pere De Las Heras. Icdar 2013 robust reading competition. In ICDAR, pages 1484-1493, 2013. 6" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 57, + 455, + 296, + 507 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 455, + 296, + 507 + ], + "spans": [ + { + "bbox": [ + 57, + 455, + 296, + 507 + ], + "type": "text", + "content": "[10] Yuxin Kong, Canjie Luo, Weihong Ma, Qiyuan Zhu, Shenggao Zhu, Nicholas Yuan, and Lianwen Jin. Look closer to supervise better: One-shot font generation via component-based discriminator. In CVPR, pages 13482-13491, 2022. 3" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 57, + 510, + 295, + 553 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 510, + 295, + 553 + ], + "spans": [ + { + "bbox": [ + 57, + 510, + 295, + 553 + ], + "type": "text", + "content": "[11] Sanyam Lakhanpal, Shivang Chopra, Vinija Jain, Aman Chadha, and Man Luo. Refining text-to-image generation: Towards accurate training-free glyph-enhanced image generation. arXiv preprint arXiv:2403.16422, 2024. 2" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 57, + 555, + 295, + 598 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 555, + 295, + 598 + ], + "spans": [ + { + "bbox": [ + 57, + 555, + 295, + 598 + ], + "type": "text", + "content": "[12] Zeyu Liu, Weicong Liang, Zhanhao Liang, Chong Luo, Ji Li, Gao Huang, and Yuhui Yuan. Glyph-byt5: A customized text encoder for accurate visual text rendering. arXiv preprint arXiv:2403.09622, 2024. 2" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 57, + 601, + 295, + 644 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 601, + 295, + 644 + ], + "spans": [ + { + "bbox": [ + 57, + 601, + 295, + 644 + ], + "type": "text", + "content": "[13] Song Park, Sanghyuk Chun, Junbum Cha, Bado Lee, and Hyunjung Shim. Multiple heads are better than one: Few-shot font generation with multiple localized experts. In ICCV, pages 13900-13909, 2021. 1" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 57, + 647, + 295, + 689 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 647, + 295, + 689 + ], + "spans": [ + { + "bbox": [ + 57, + 647, + 295, + 689 + ], + "type": "text", + "content": "[14] Yadong Qu, Qingfeng Tan, Hongtao Xie, Jianjun Xu, Yuxin Wang, and Yongdong Zhang. Exploring stroke-level modifications for scene text editing. In AAAI, pages 2119-2127, 2023. 1, 3, 6" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 57, + 691, + 295, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 691, + 295, + 713 + ], + "spans": [ + { + "bbox": [ + 57, + 691, + 295, + 713 + ], + "type": "text", + "content": "[15] Robin Rombach, Andreas Blattmann, Dominik Lorenz, Patrick Esser, and Björn Ommer. High-resolution image" + } + ] + } + ], + "index": 15 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 316, + 73, + 555, + 666 + ], + "type": "list", + "angle": 0, + "index": 32, + "blocks": [ + { + "bbox": [ + 335, + 73, + 555, + 95 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 335, + 73, + 555, + 95 + ], + "spans": [ + { + "bbox": [ + 335, + 73, + 555, + 95 + ], + "type": "text", + "content": "synthesis with latent diffusion models. In CVPR, pages 10684-10695, 2022. 4, 6" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 317, + 96, + 554, + 129 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 96, + 554, + 129 + ], + "spans": [ + { + "bbox": [ + 317, + 96, + 554, + 129 + ], + "type": "text", + "content": "[16] Prasun Roy, Saumik Bhattacharya, Subhankar Ghosh, and Umapada Pal. Stefann: scene text editor using font adaptive neural network. In CVPR, pages 13228-13237, 2020. 1, 3" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 317, + 130, + 555, + 194 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 130, + 555, + 194 + ], + "spans": [ + { + "bbox": [ + 317, + 130, + 555, + 194 + ], + "type": "text", + "content": "[17] Chitwan Sahara, William Chan, Saurabh Saxena, Lala Li, Jay Whang, Emily L Denton, Kamyar Ghasemipour, Raphael Gontijo Lopes, Burcu Karagol Ayan, Tim Salimans, et al. Photorealistic text-to-image diffusion models with deep language understanding. In NeurIPS, pages 36479-36494, 2022. 2" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 316, + 197, + 554, + 228 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 197, + 554, + 228 + ], + "spans": [ + { + "bbox": [ + 316, + 197, + 554, + 228 + ], + "type": "text", + "content": "[18] Wataru Shimoda, Daichi Haraguchi, Seiichi Uchida, and Kota Yamaguchi. De-rendering stylized texts. In ICCV, pages 1076-1085, 2021. 3" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 317, + 231, + 555, + 262 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 231, + 555, + 262 + ], + "spans": [ + { + "bbox": [ + 317, + 231, + 555, + 262 + ], + "type": "text", + "content": "[19] Yuxiang Tuo, Wangmeng Xiang, Jun-Yan He, Yifeng Geng, and Xuansong Xie. Anytext: Multilingual visual text generation and editing. ICLR, 2024. 6" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 316, + 264, + 555, + 307 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 264, + 555, + 307 + ], + "spans": [ + { + "bbox": [ + 316, + 264, + 555, + 307 + ], + "type": "text", + "content": "[20] Yibin Wang, Yuchao Feng, Jie Wu, Honghui Xu, and Jianwei Zheng. Ca-gan: Object placement via coalescing attention based generative adversarial network. In ICME, pages 2375-2380. IEEE, 2023. 3" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 316, + 309, + 555, + 352 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 309, + 555, + 352 + ], + "spans": [ + { + "bbox": [ + 316, + 309, + 555, + 352 + ], + "type": "text", + "content": "[21] Yibin Wang, Honghui Xu, Shanghai Zhou, Weizhong Zhang, and Cheng Jin. Enhancing object coherence in layout-to-image synthesis. arXiv preprint arXiv:2311.10522, 2023. 2" + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 316, + 354, + 555, + 387 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 354, + 555, + 387 + ], + "spans": [ + { + "bbox": [ + 316, + 354, + 555, + 387 + ], + "type": "text", + "content": "[22] Yibin Wang, Weizhong Zhang, and Cheng Jin. Magicface: Training-free universal-style human image customized synthesis. arXiv preprint arXiv:2408.07433, 2024. 2" + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 316, + 388, + 554, + 419 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 388, + 554, + 419 + ], + "spans": [ + { + "bbox": [ + 316, + 388, + 554, + 419 + ], + "type": "text", + "content": "[23] Yibin Wang, Weizhong Zhang, Jianwei Zheng, and Cheng Jin. High-fidelity person-centric subject-to-image synthesis. In CVPR, 2024. 2" + } + ] + } + ], + "index": 25 + }, + { + "bbox": [ + 317, + 422, + 554, + 464 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 422, + 554, + 464 + ], + "spans": [ + { + "bbox": [ + 317, + 422, + 554, + 464 + ], + "type": "text", + "content": "[24] Yibin Wang, Weizhong Zhang, Jianwei Zheng, and Cheng Jin. Primecomposer: Faster progressively combined diffusion for image composition with attention steering. arXiv preprint arXiv:2403.05053, 2024. 2" + } + ] + } + ], + "index": 26 + }, + { + "bbox": [ + 317, + 467, + 555, + 499 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 467, + 555, + 499 + ], + "spans": [ + { + "bbox": [ + 317, + 467, + 555, + 499 + ], + "type": "text", + "content": "[25] Liang Wu, Chengquan Zhang, Jiaming Liu, Junyu Han, Jingtuo Liu, Errui Ding, and Xiang Bai. Editing text in the wild. In ACM MM, pages 1500–1508, 2019. 3" + } + ] + } + ], + "index": 27 + }, + { + "bbox": [ + 317, + 501, + 555, + 544 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 501, + 555, + 544 + ], + "spans": [ + { + "bbox": [ + 317, + 501, + 555, + 544 + ], + "type": "text", + "content": "[26] Xingqian Xu, Zhifei Zhang, Zhaowen Wang, Brian Price, Zhonghao Wang, and Humphrey Shi. Rethinking text segmentation: A novel dataset and a text-specific refinement approach. In CVPR, pages 12045-12055, 2021. 7" + } + ] + } + ], + "index": 28 + }, + { + "bbox": [ + 317, + 545, + 555, + 577 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 545, + 555, + 577 + ], + "spans": [ + { + "bbox": [ + 317, + 545, + 555, + 577 + ], + "type": "text", + "content": "[27] Qiangpeng Yang, Jun Huang, and Wei Lin. Swaptext: Image based texts transfer in scenes. In CVPR, pages 14700-14709, 2020. 1, 3" + } + ] + } + ], + "index": 29 + }, + { + "bbox": [ + 317, + 580, + 555, + 621 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 580, + 555, + 621 + ], + "spans": [ + { + "bbox": [ + 317, + 580, + 555, + 621 + ], + "type": "text", + "content": "[28] Richard Zhang, Phillip Isola, Alexei A Efros, Eli Shechtman, and Oliver Wang. The unreasonable effectiveness of deep features as a perceptual metric. In CVPR, pages 586-595, 2018. 7" + } + ] + } + ], + "index": 30 + }, + { + "bbox": [ + 316, + 624, + 555, + 666 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 624, + 555, + 666 + ], + "spans": [ + { + "bbox": [ + 316, + 624, + 555, + 666 + ], + "type": "text", + "content": "[29] Yiming Zhao and Zhouhui Lian. Udifftext: A unified framework for high-quality text synthesis in arbitrary images via character-aware diffusion models. In ECCV, 2024. 2, 3, 4, 5, 6, 7" + } + ] + } + ], + "index": 31 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "text", + "content": "28563" + } + ] + } + ], + "index": 33 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 8 + } + ], + "_backend": "vlm", + "_version_name": "2.6.4" +} \ No newline at end of file diff --git a/2025/DreamTrack_ Dreaming the Future for Multimodal Visual Object Tracking/2e4cde12-368f-4feb-b1b3-a980bb1eb002_content_list.json b/2025/DreamTrack_ Dreaming the Future for Multimodal Visual Object Tracking/2e4cde12-368f-4feb-b1b3-a980bb1eb002_content_list.json new file mode 100644 index 0000000000000000000000000000000000000000..40299ab956f71165c776b4fd5826778701060f24 --- /dev/null +++ b/2025/DreamTrack_ Dreaming the Future for Multimodal Visual Object Tracking/2e4cde12-368f-4feb-b1b3-a980bb1eb002_content_list.json @@ -0,0 +1,1508 @@ +[ + { + "type": "text", + "text": "DreamTrack: Dreaming the Future for Multimodal Visual Object Tracking", + "text_level": 1, + "bbox": [ + 119, + 130, + 879, + 152 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Mingzhe Guo $^{1*}$ Weiping Tan $^{1}$ Wenyu Ran $^{1}$ Liping Jing $^{1\\dagger}$ Zhipeng Zhang $^{2*}$ Beijing Jiaotong University $^{1}$ , Shanghai Jiaotong University $^{2}$", + "bbox": [ + 156, + 179, + 844, + 219 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Abstract", + "text_level": 1, + "bbox": [ + 246, + 251, + 326, + 266 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Aiming to achieve class-agnostic perception in visual object tracking, current trackers commonly formulate tracking as a one-shot detection problem with the template-matching architecture. Despite the success, severe environmental variations in long-term tracking raise challenges to generalizing the tracker in novel situations. Temporal trackers try to fix it by preserving the time-validity of target information with historical predictions, e.g., updating the template. However, solely transmitting the previous observations instead of learning from them leads to an inferior capability of understanding the tracking scenario from past experience, which is critical for the generalization in new frames. To address this issue, we reformulate temporal learning in visual tracking as a History-to-Future process and propose a novel tracking framework DreamTrack. Our DreamTrack learns the temporal dynamics from past observations to dream the future variations of the environment, which boosts the generalization with the extended future information from history. Considering the uncertainty of future variation, multimodal prediction is designed to infer the target trajectory of each possible future situation. The experiments demonstrate that our DreamTrack achieves leading performance with real-time inference speed. In particular, DreamTrack obtains SUC scores of $76.6\\% /87.9\\%$ on LasOT/TrackingNet, surpassing all recent SOTA trackers.", + "bbox": [ + 89, + 282, + 483, + 660 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "1. Introduction", + "text_level": 1, + "bbox": [ + 91, + 686, + 220, + 702 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Visual Object Tracking (VOT) is one of the most fundamental computer vision problems with numerous applications [5, 28], which aims to estimate the position of an arbitrary target in a video sequence, given only its location in the initial frame. To achieve the class-agnostic perception, recent prevalent Siamese-based approaches [2, 6, 7, 9, 20, 31, 34, 47, 51, 52] typically formulate visual tracking as a one-shot detection problem. Specifically, these trackers perform general object tracking by matching with the initialized target template and then predicting the location in a single forward pass, as shown in Fig. 1(a).", + "bbox": [ + 89, + 712, + 483, + 878 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/85445e7bf89ad36005c9fa61614c3301311996a6f9e69ff695864fce586ee3ae.jpg", + "image_caption": [ + "Figure 1. Comparison of different temporal architectures. (a) The non-temporal architecture that regards visual tracking as a one-shot detection problem. (b) The temporal architecture with only historical information that limits the temporal messages in the transmission of previous observations. (c) Our proposed temporal architecture with both historical and future information, which learns environmental dynamics from past experiences and dreams the future states for better generalization in subsequent frames." + ], + "image_footnote": [], + "bbox": [ + 511, + 247, + 906, + 539 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Despite the demonstrated success, it's challenging for the non-temporal architecture to ensure the generalization in new frames with severe environmental variations compared with the initial frame, e.g., deformation of the target appearance and background in long-term tracking. The outdated template hardly provides effective target-aware knowledge but distracts the tracking process. Then many works [7, 10, 12, 25] intuitively try to fix it by keeping the time-validity of target information, e.g., updating the template with predictions of high confidence scores in previous frames [45, 50] or passing historical trajectory as prior knowledge [7, 47], as shown in Fig. 1(b). However, these efforts focus on only transmitting previous observations instead of learning from them, leading to an inferior capability of understanding the tracking scenario from the past expe", + "bbox": [ + 511, + 674, + 908, + 902 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "CVF", + "bbox": [ + 106, + 2, + 181, + 42 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore.", + "bbox": [ + 236, + 0, + 810, + 46 + ], + "page_idx": 0 + }, + { + "type": "page_footnote", + "text": "* Equal contribution. † Corresponding author.", + "bbox": [ + 107, + 886, + 352, + 900 + ], + "page_idx": 0 + }, + { + "type": "page_number", + "text": "7201", + "bbox": [ + 482, + 944, + 513, + 955 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "rience so that enables generalization to novel situations.", + "bbox": [ + 89, + 90, + 460, + 104 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "To address this issue, we reformulate the temporal learning in VOT as a History-to-Future process, as shown in Fig. 1(c). The pipeline dreams the future states of the tracking scenario based on learned temporal dynamics from previous observations, which empowers the tracker with superior generalization ability in complex environments of new frames. Our intuition stems from the human recognition process [21, 30]. The brain abstracts the world from past senses and imagines future variations, which aims to achieve goals in complex scenarios even though they are never encountered before. Such a temporal modeling manner is intuitively superior to the one limited in historical information, which further develops the generalization in new frames with the future information from dreamed states. To this end, we construct a future dreaming module, originally proposed in the world model for reinforcement learning tasks [21-23], to endow the tracker with the capability of predicting future variations of the target from the learned historical dynamics (see Fig. 1(c)). In contrast to non-temporal [6, 9, 31] or historical information-only trackers [7, 47, 50], our dreaming module extends the temporal messages from the history to its future, in which the dreamed future frames are self-supervised with ground-truth future observations. The future information undoubtedly facilitates better generalization for appearance deformation of the target and background variance.", + "bbox": [ + 89, + 108, + 485, + 500 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "More specifically, we propose a novel tracking framework dubbed DreamTrack, which has a dream-encoder-decoder architecture. The dreaming stage caches historical observations and predicts future states to represent the environmental variation in the next few frames. Then the features of the current and dreamed future frames are input to the encoder, which extracts target-aware features with the temporal messages from both history and future. Notably, after each feature extraction stage of the encoder, the temporal dynamics will be refined along the timeline, which aims to filter target-irrelevant interferences with the exposure of the target information. The final decoder predicts the trajectory of the target in the current and future frames. Considering the uncertainty of the future variation, we forecast the probabilities and target locations of different possible future scenarios, i.e., multimodal prediction.", + "bbox": [ + 89, + 501, + 483, + 743 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Extensive experiments on eight prevailing VOT benchmarks show that our proposed DreamTrack outperforms recent state-of-the-art trackers. For instance, under fair settings, DreamTrack- $\\mathrm{L}_{384}$ obtains a $76.6\\%$ SUC score on the LaSOT [15] dataset, surpassing ARTrackV2- $\\mathrm{L}_{384}$ [1] by $3.0\\%$ and LoRAT- $\\mathrm{L}_{378}$ [34] by $1.5\\%$ , respectively. In summary, the contributions of this work are as follows:", + "bbox": [ + 89, + 744, + 483, + 851 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "- We reformulate the tracking problem as a History-to-Future process and are the first to develop the temporal messages into the future frames for better generalization", + "bbox": [ + 89, + 854, + 483, + 902 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "of novel situations in visual object tracking.", + "bbox": [ + 526, + 90, + 816, + 106 + ], + "page_idx": 1 + }, + { + "type": "list", + "sub_type": "text", + "list_items": [ + "- We make the first attempt at predicting the future states of the target and performing multimodal prediction to include the uncertainty of different future scenarios, which empowers the tracker with superior generalization for appearance deformation and background variation.", + "- We prove the effectiveness of the proposed future dreaming to facilitate tracking with state-of-the-art performance on eight prevailing VOT benchmarks." + ], + "bbox": [ + 513, + 107, + 903, + 226 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "2. Related Work", + "text_level": 1, + "bbox": [ + 513, + 241, + 655, + 257 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Visual Object Tracking. Current prevailing trackers [2, 6, 9, 20, 31, 53] commonly employ a template-matching framework, which exploits the initialized template to match the target within the search region. Regarding tracking as a one-shot detection problem, these approaches develop a deep backbone to extract visual features [33, 51, 52], then predict the current target location following the detection paradigms [41, 55], such as object scale estimation [32, 44, 53] and center point localization [6, 20]. Observing the target information of the template may be weakened by the appearance deformation and background variation in long-term tracking, many works design modules to ensure the time-validity of target-relevant knowledge, e.g., updating the template with predictions of high confidence scores in previous frames [45, 50] or passing historical trajectory as prior knowledge [7, 47]. Despite the improved performance in long-term tracking, the limited understanding of the tracking scenario leads to suboptimal generalization in novel situations. Thus, we reformulate tracking as a History-to-Future process that learns the environmental dynamics from past experiences and dreams the future states for better adaptation and tracking capability in new frames.", + "bbox": [ + 511, + 267, + 906, + 599 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "World Model. The world model [30] aims to learn a general representation of the world and predicts future world states resulting from a sequence of actions, which has been widely studied in the game [21-23, 40] and lab environments [14, 18, 48]. Dreamer [22] learns a latent dynamics model from past experience to predict future state values and actions in a latent space. It's capable of handling challenging visual control tasks in the DeepMind Control Suite [43]. DreamerV2 [23] further achieves human-level performance on Atari games with discrete modeling. DreamerV3 [24] develops a general algorithm and allows solving challenging control problems such as obtaining diamonds in Minecraft from scratch given sparse rewards. Recently, learning world models in driving scenes has gained attention. MILE [26] employs a model-based imitation learning method to jointly learn a dynamics model and driving behavior. In this work, we introduce the mechanism of the world model to visual object tracking, which boosts environmental understanding with the dreamed future states.", + "bbox": [ + 511, + 613, + 908, + 900 + ], + "page_idx": 1 + }, + { + "type": "page_number", + "text": "7202", + "bbox": [ + 482, + 944, + 514, + 955 + ], + "page_idx": 1 + }, + { + "type": "image", + "img_path": "images/50d33d9bbd2e086ede54ab20a13743a7d7c9c12e918522a61f5a5823f0620ca4.jpg", + "image_caption": [ + "Figure 2. The dream-encoder-decoder framework of our DreamTrack. Initially, the dreaming stage takes the input image tokens to predict future states of the next two frames based on the cached temporal dynamics in the Future Dreaming module. Then features of the current frame and dreamed future are fed into the encoder for further modeling, with a Temporal Self-Attention after each encoder stage for target-aware refinement. In the end, the decoder performs multimodal prediction for the target, i.e., the probabilities and trajectories in the current and next two frames of three possible future scenarios. The prediction with the highest probability is selected as the final output." + ], + "image_footnote": [], + "bbox": [ + 91, + 88, + 906, + 316 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3. Method", + "text_level": 1, + "bbox": [ + 89, + 421, + 181, + 436 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3.1. Preliminaries", + "text_level": 1, + "bbox": [ + 89, + 446, + 230, + 460 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Formulation of VOT. Given the template image $\\mathbf{z}^0$ of a target initialized in the first frame, the objective of visual tracking is to localize the target $\\mathrm{b}^t$ from the continuous input image $\\mathbf{x}^t$ , where $\\mathbf{b}$ denotes the target state, i.e., a bounding box. The mainstream tracking paradigm aims to learn a deep tracking model $f$ that performs localization with the output $\\mathbf{y}^t$ following different detection paradigms, i.e., anchor-based [41] or anchor-free [55] architectures. The tracking process can be formulated as,", + "bbox": [ + 89, + 468, + 483, + 604 + ], + "page_idx": 2 + }, + { + "type": "equation", + "text": "\n$$\ny ^ {t} = f \\left(z ^ {0}, x ^ {t}\\right). \\tag {1}\n$$\n", + "text_format": "latex", + "bbox": [ + 233, + 613, + 480, + 631 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "To alleviate the influence of target's appearance deformation and background variation in long-term tracking, recent prevailing temporal trackers exploit previous outputs $\\mathrm{y}^{0:t - 1}$ as additional prior knowledge, e.g., updating the template [45, 50] or passing historical target location [7, 47]. The temporal tracking process is formulated as,", + "bbox": [ + 89, + 641, + 483, + 731 + ], + "page_idx": 2 + }, + { + "type": "equation", + "text": "\n$$\n\\mathrm {y} ^ {t} = f \\left(\\mathrm {z} ^ {0}, \\mathrm {x} ^ {t}, \\mathrm {y} ^ {0: t - 1}\\right). \\tag {2}\n$$\n", + "text_format": "latex", + "bbox": [ + 207, + 739, + 482, + 758 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "In contrast, we formulate temporal visual tracking as a History-to-Future process, which predicts the future trajectory of the target (including the current frame and the next two ones) based on the learned environmental dynamics from historical observations (cached in the target queries $\\mathbf{Q}_{\\mathrm{tgt}}^{t:t + 2}$ of $f$ , described in Sec. 3.2). This modeling pipeline can be expressed as,", + "bbox": [ + 89, + 768, + 483, + 875 + ], + "page_idx": 2 + }, + { + "type": "equation", + "text": "\n$$\ny ^ {t: t + 2} = f \\left[ Q _ {\\mathrm {t g t}} ^ {t: t + 2} \\right] \\left(z ^ {0}, x ^ {t}\\right). \\tag {3}\n$$\n", + "text_format": "latex", + "bbox": [ + 192, + 883, + 482, + 902 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "The parameterized modeling of environmental dynamics $(\\mathbf{Q}_{\\mathrm{tg}}^{t:t + 2})$ instead of explicit transmission of historical output $\\mathrm{y}^{0:t - 1}$ , not only enables better generalization in novel situations of future frames, but also reduces the risk of error accumulation from past experiences.", + "bbox": [ + 511, + 422, + 906, + 500 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Trajectory Coordinate Mapping. To reduce computation costs, most methods perform tracking on the cropped search region instead of the full-resolution frame, which outputs target's coordinates in the current frame, relative to the search region. Thus, it's necessary to map the predicted target location of the relative coordinate system to the original frame after inference, which we denote as the global coordinate system. Correspondingly, the ground-truth boxes for training are also transformed from the global to relative coordinate system. For our DreamTrack that learns dynamics and predicts target trajectory over a video sequence, we additionally cache the ground-truth boxes of the preceding two frames in the global coordinate system and map them to the relative coordinate system of the current frame after the search region is cropped. Notably, the search region patches of the two future frames are cropped by referring to the ground-truth box of the current frame, which represent the target's motion and serve for the supervision of future image reconstruction (described in Sec. 3.2 and Fig. 3).", + "bbox": [ + 511, + 508, + 906, + 795 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3.2. Network Architecture", + "text_level": 1, + "bbox": [ + 511, + 803, + 718, + 816 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "The overall framework of DreamTrack is depicted in Fig. 2. Given the target sequences constructed from object trajectories, we use a novel dream-encoder-decoder structure for learning and inference. The target state is described by a trajectory including the current and next two frames.", + "bbox": [ + 511, + 824, + 905, + 900 + ], + "page_idx": 2 + }, + { + "type": "page_number", + "text": "7203", + "bbox": [ + 482, + 944, + 514, + 955 + ], + "page_idx": 2 + }, + { + "type": "image", + "img_path": "images/f5bbecd65d43e3839113a97abc270b539916e8ca1521b7ce2122cfb022edb8fb.jpg", + "image_caption": [ + "Figure 3. Illustration of the proposed Future Dreaming module. The target queries auto-regressively update the temporal dynamics with the patch embeddings of the current frame through GRU, which are then exploited to predict the future states. For training, the distributions of the predicted future states and decoded future images are self-supervised by the ground-truth future frames." + ], + "image_footnote": [], + "bbox": [ + 96, + 90, + 478, + 277 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Dream. We propose the Future Dreaming module to learn temporal dynamics of the tracking scenario. As shown in Fig. 3, the module evolves the dynamics with current observation and dreams the future states, which endows the tracker with the capability of future prediction by supervising with ground-truth future frames. Specifically, three trainable target queries $\\mathbf{Q}_{\\mathrm{tgt}}^{t - 1:t + 1} \\in \\mathbb{R}^{3 \\times C}$ , which represent the transition dynamics of the current and next two frames, are initialized. With the template $z^0$ and search region $x^t$ of each new frame, the images are split and flattened to build current observation feature $\\mathbf{F}^{\\mathrm{t}} = \\{\\mathbf{F}_z^0 \\in \\mathbb{R}^{H_z W_z \\times C}, \\mathbf{F}_x^t \\in \\mathbb{R}^{H_x W_x \\times C}\\}$ with a linear projection. Then target queries auto-regressively update the cached temporal messages with $\\mathbf{F}^{\\mathrm{t}}$ through a gated recurrent unit (GRU) [8],", + "bbox": [ + 89, + 386, + 483, + 598 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\mathbf {Q} _ {\\mathrm {t g t}} ^ {t: t + 2} = \\operatorname {G R U} \\left(\\mathbf {Q} _ {\\mathrm {t g t}} ^ {t - 1: t + 1}, \\mathbf {F} ^ {t}\\right). \\tag {4}\n$$\n", + "text_format": "latex", + "bbox": [ + 179, + 611, + 482, + 630 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "As the aim of our DreamTrack is to enhance the generalization in novel situations of new frames, it's a prerequisite to evolve the future variations of the tracking environment. Thus, the updated target queries $\\mathbf{Q}_{\\mathrm{tgt}}^{t:t + 2}$ are exploited to backward imagine the future observation feature $\\hat{\\mathbf{F}}^{t:t + 2} = \\{\\hat{\\mathbf{F}}_{\\mathrm{z}}^{t:t + 2}\\in \\mathbb{R}^{3\\times H_{\\mathrm{z}}W_{\\mathrm{z}}\\times C},\\hat{\\mathbf{F}}_{\\mathrm{x}}^{t:t + 2}\\in \\mathbb{R}^{3\\times H_{\\mathrm{x}}W_{\\mathrm{x}}\\times C}\\}$ based on $\\mathbf{F}^{\\mathrm{t}}$ with a cross-attention,", + "bbox": [ + 89, + 633, + 483, + 739 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\hat {\\mathbf {F}} ^ {t: t + 2} = \\operatorname {C r o s s A t t e n t i o n} (q = \\mathbf {F} ^ {t}, k v = \\mathbf {Q} _ {\\mathrm {t g t}} ^ {t: t + 2}). \\quad (5)\n$$\n", + "text_format": "latex", + "bbox": [ + 109, + 746, + 482, + 767 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Notably, the interaction between $\\mathbf{F}^t$ and $\\mathbf{Q}_{\\mathrm{tgt}}^t$ to generate $\\hat{\\mathbf{F}}^t$ intends to enhance the feature quality of the current frame by introducing historical information, which originates from $\\mathbf{Q}_{\\mathrm{tgt}}^{t - 1}$ before the GRU update (Eq. 4). The output $\\hat{\\mathbf{F}}^{t:t + 2}$ of the module will be further modeled with the Temporal Self-Attention, which is detailed in the encoder.", + "bbox": [ + 89, + 773, + 482, + 869 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Following the loss design in world models [22-24], we respectively map the dreamed future observation fea", + "bbox": [ + 89, + 869, + 483, + 900 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "ture $\\hat{\\mathbf{F}}_{\\mathrm{x}}^{t:t + 2}$ and ground-truth observation feature $\\mathbf{F}_{\\mathrm{x}}^{t:t + 2}$ to distributions of $\\{\\hat{\\mu}_{\\mathrm{x}}^{t:t + 2},\\hat{\\sigma}_{\\mathrm{x}}^{t:t + 2}\\in \\mathbb{R}^{3\\times H_{\\mathrm{x}}W_{\\mathrm{x}}\\times C^{\\prime}}\\}$ and $\\{\\mu_{\\mathrm{x}}^{t:t + 2},\\sigma_{\\mathrm{x}}^{t:t + 2}\\in \\mathbb{R}^{3\\times H_{\\mathrm{x}}W_{\\mathrm{x}}\\times C^{\\prime}}\\}$ with MLP layers, and then minimize their KL divergence. Notably, we only consider supervising the dreaming process of the search region, as the input template always comes from the initial frame. The prior distribution from $\\hat{\\mathbf{F}}_{\\mathrm{x}}^{t:t + 2}$ is regarded as a prediction of the future dynamics. In contrast, the posterior distribution from $\\mathbf{F}_{\\mathrm{x}}^{t:t + 2}$ represents the future dynamics from the ground-truth observation $\\mathbf{x}^{t:t + 2}$ . The KL divergence measures the gap between the imagined future (prior) and the true future (posterior). We expect to enhance the capability of future prediction for long-term tracking, which is realized by optimizing the dreaming loss $\\mathcal{L}_{\\mathrm{drm}}$", + "bbox": [ + 511, + 89, + 903, + 303 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {L} _ {\\mathrm {d r m}} = \\mathrm {K L} \\left(\\left\\{\\hat {\\mu} _ {\\mathrm {x}} ^ {t: t + 2}, \\hat {\\sigma} _ {\\mathrm {x}} ^ {t: t + 2} \\right\\} \\mid \\mid \\left\\{\\mu_ {\\mathrm {x}} ^ {t: t + 2}, \\sigma_ {\\mathrm {x}} ^ {t: t + 2} \\right\\}\\right), \\tag {6}\n$$\n", + "text_format": "latex", + "bbox": [ + 539, + 309, + 903, + 329 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "where $\\mathrm{KL}(\\cdot ||\\cdot)$ denotes the KL divergence.", + "bbox": [ + 511, + 335, + 799, + 349 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Besides the supervision from the distribution difference, we further improve the accuracy of future prediction by reconstructing the future search regions based on $\\hat{\\mathbf{F}}_{\\mathrm{x}}^{t:t + 2}$ with a decoder of several CNN upsampling layers. The decoded future search regions $\\hat{\\mathbf{x}}^{t:t + 2}$ is supervised by the groundtruth search regions $\\mathbf{x}^{t:t + 2}$ by optimizing the reconstruction loss $\\mathcal{L}_{\\mathrm{RCT}}$", + "bbox": [ + 511, + 351, + 903, + 455 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {L} _ {\\mathrm {r c t}} = \\operatorname {M S E} \\left(\\hat {\\mathrm {x}} ^ {t: t + 2} | | \\mathrm {x} ^ {t: t + 2}\\right), \\tag {7}\n$$\n", + "text_format": "latex", + "bbox": [ + 612, + 455, + 903, + 472 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "where $\\mathrm{MSE}(\\cdot ||\\cdot)$ is the mean squared error.", + "bbox": [ + 511, + 477, + 800, + 493 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Encoder. Following previous works [7, 47, 51], we adopt a plain vision transformer (ViT) [13] as the encoder and model the search and template images jointly. The encoder extracts the features of the search and template images and learns feature-level correspondence to filter target-relevant distractors. The input dreamed observation features $\\{\\hat{\\mathbf{F}}_{\\mathrm{z}}^{t:t + 2},\\hat{\\mathbf{F}}_{\\mathrm{x}}^{t:t + 2}\\}$ are concatenated along the spatial dimension and fed into the encoder for feature extraction,", + "bbox": [ + 511, + 500, + 905, + 621 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\begin{array}{l} \\hat {\\mathbf {F}} _ {\\mathrm {z x}, i} ^ {t: t + 2} = \\operatorname {C o n c a t} \\left(\\hat {\\mathbf {F}} _ {\\mathrm {z}} ^ {t: t + 2}, \\hat {\\mathbf {F}} _ {\\mathrm {x}} ^ {t: t + 2}\\right), i = 0 \\\\ \\hat {\\mathbf {F}} _ {\\mathrm {z x}, i} ^ {t: t + 2} = \\operatorname {B l o c k} _ {i} \\left(\\hat {\\mathbf {F}} _ {\\mathrm {z x}, i - 1} ^ {t: t + 2}\\right), \\quad 1 \\leq i \\leq l _ {\\text {s p t}} \\tag {8} \\\\ \\end{array}\n$$\n", + "text_format": "latex", + "bbox": [ + 562, + 627, + 903, + 670 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "where $\\hat{\\mathbf{F}}_{\\mathrm{zx},i}^{t:t + 2}\\in \\mathbb{R}^{3\\times (H_{\\mathrm{z}}W_{\\mathrm{z}} + H_{\\mathrm{x}}W_{\\mathrm{x}})\\times C}$ , Block denotes the transformer block in the encoder for spatial modeling and $l_{\\mathrm{spt}}$ is the layer number.", + "bbox": [ + 511, + 676, + 905, + 724 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Besides the spatial modeling between the template and search region in Block, it's also important to exploit the extracted target-aware information to refine the temporal dynamics along the timeline. Thus, after each feature extraction stage of the encoder, we arrange a Temporal Self-Attention module to filter target-irrelevant interferences with the exposure of the target information from different levels. As shown in Fig. 4, the dreamed observation features $\\hat{\\mathbf{F}}_{\\mathrm{zx}}^{t:t + 2}$ are transposed and input to the attention for the interaction on the temporal dimension,", + "bbox": [ + 511, + 724, + 905, + 876 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\hat {\\mathbf {F}} _ {\\mathrm {z x}, i} ^ {t: t + 2} = \\operatorname {T e m A t t n} _ {\\text {s e l f}} ^ {i} \\left(\\hat {\\mathbf {F}} _ {\\mathrm {z x}, i - 1} ^ {t: t + 2}\\right), \\quad 1 \\leq i \\leq l _ {\\text {t e m}} \\tag {9}\n$$\n", + "text_format": "latex", + "bbox": [ + 539, + 882, + 903, + 902 + ], + "page_idx": 3 + }, + { + "type": "page_number", + "text": "7204", + "bbox": [ + 482, + 944, + 514, + 955 + ], + "page_idx": 3 + }, + { + "type": "image", + "img_path": "images/86554f4977ac27dede5172890e3b01c454a47ba1a6fee32f3591f2f46a97ddc2.jpg", + "image_caption": [ + "Figure 4. Illustration of the Temporal Self/Cross-Attention. For self-attention, patch embeddings of the template and search region are jointly modeled on the temporal dimension. In contrast, the cross-attention takes target queries as $\\mathbf{Q}$ and patch embeddings of search region as $\\mathbf{K}$ , $\\mathbf{V}$ to extract localization-relevant information." + ], + "image_footnote": [], + "bbox": [ + 93, + 88, + 482, + 255 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "where $\\mathrm{TemAttn}_{\\mathrm{self}}$ denotes the Temporal Self-Attention module and $l_{\\mathrm{tem}}$ is the module number corresponding to the four stages of the encoder (see Fig. 2).", + "bbox": [ + 89, + 347, + 483, + 393 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Decoder. The final decoder is responsible for predicting the target trajectory in the current and next two frames based on the updated target queries $\\mathbf{Q}_{\\mathrm{tgt}}^{t:t + 2}$ and target-aware observation feature of the search region $\\hat{\\mathbf{F}}_{\\mathrm{x}}^{t:t + 2}$ , which is split from the encoder output $\\hat{\\mathbf{F}}_{\\mathrm{zx}}^{t:t + 2}$ . Notably, the dreamed feature of the template $\\hat{\\mathbf{F}}_{\\mathrm{z}}^{t:t + 2}$ is not employed as the ultimate is to locate the target in the search region and the target information has been introduced into $\\hat{\\mathbf{F}}_{\\mathrm{x}}^{t:t + 2}$ after encoder modeling. The target queries $\\mathbf{Q}_{\\mathrm{tgt}}^{t:t + 2}$ first extract localization-relevant messages from $\\hat{\\mathbf{F}}_{\\mathrm{x}}^{t:t + 2}$ with the Temporal Cross-Attention module. As shown in Fig. 4, each target query interacts with the region feature of the same timestamp,", + "bbox": [ + 89, + 401, + 483, + 589 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\hat {\\mathbf {Q}} _ {\\mathrm {t g t}} ^ {t: t + 2} = \\operatorname {T e m A t t n} _ {\\text {c r o s s}} (q = \\mathbf {Q} _ {\\mathrm {t g t}} ^ {t: t + 2}, k v = \\hat {\\mathbf {F}} _ {\\mathrm {x}} ^ {t: t + 2}), \\tag {10}\n$$\n", + "text_format": "latex", + "bbox": [ + 99, + 593, + 480, + 614 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "where $\\mathrm{TemAttn}_{\\mathrm{cross}}$ is the Temporal Cross-Attention module. The output target queries $\\hat{\\mathbf{Q}}_{\\mathrm{tgt}}^{t:t + 2}$ are then used to predict the target trajectory of the current and next two frames.", + "bbox": [ + 89, + 619, + 482, + 665 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "The key challenge to future prediction is the high degree of uncertainty, in large part due to not knowing the intents and latent characteristics of the tracking target. For example, the target commonly has a multimodal distribution of futures: it could turn, go straight, slow down, speed up, etc. The high degree of multimodality also exists in the motion prediction of autonomous driving [17, 36], which correspondingly designs a multimodal prediction to generate multiple plausible trajectories of the target vehicle. We draw inspiration from this solution and design the tailored Multimodal Prediction to our DreamTrack. As shown in Fig. 5, each target query of one timestamp forecasts the probabilities $\\hat{\\mathbf{p}}^T\\in \\mathbb{R}^{3\\times 1}$ and target states $\\hat{\\mathbf{b}}^T\\in \\mathbb{R}^{3\\times 4}$ of three possible future situations with a stack of MLP layers,", + "bbox": [ + 89, + 666, + 483, + 876 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\hat {\\mathrm {p}} ^ {T}, \\hat {\\mathrm {b}} ^ {T} = \\operatorname {M L P} \\left(\\hat {\\mathbf {Q}} _ {\\mathrm {t g t}} ^ {T}\\right). T \\in \\{t, t + 1, t + 2 \\} \\tag {11}\n$$\n", + "text_format": "latex", + "bbox": [ + 120, + 882, + 482, + 902 + ], + "page_idx": 4 + }, + { + "type": "image", + "img_path": "images/4360af44e38dd4565c9f5f0537d5afcc39942769629616cd044483b759b6afee.jpg", + "image_caption": [ + "Figure 5. Illustration of the Multimodal Prediction. The target queries of different timestamps respectively predict the probabilities and locations of three possible future situations, in which the one of the highest probability score is the final output." + ], + "image_footnote": [], + "bbox": [ + 514, + 88, + 903, + 234 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "After the prediction, the target state with the highest probability of each timestamp constitutes the final target trajectory. For training, we first calculate GIoU [42] between the predicted target state of each modality $\\hat{\\mathbf{b}}_j^T\\in \\hat{\\mathbf{b}}^T (1\\leq j\\leq 3)$ and the ground-truth $\\mathrm{b}^T$ . The prediction with the largest GIoU is assigned with the probability label of 1 (the other two ones are 0), and L1 loss $\\mathcal{L}_{\\mathrm{L1}}$ and GIoU loss $\\mathcal{L}_{\\mathrm{GIoU}}$ are applied to the output target states. The probabilities are supervised by the 0-1 labels with a cross-entropy loss $\\mathcal{L}_{\\mathrm{ce}}$ .", + "bbox": [ + 511, + 305, + 906, + 443 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "3.3. Training and Inference", + "text_level": 1, + "bbox": [ + 511, + 450, + 725, + 465 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Training. Our DreamTrack performs video sequence-level training and facilitates environmental understanding in an end-to-end manner. We construct video clips in which each contains 4 frames for training (i.e., one frame as template and three ones as search sequence), following the protocol of coordinate mapping described in Sec. 3.1. For every new input of the video clip, the hidden state in GRU (see Sec. 3.2) will be initialized, which avoids being distracted by the last tracking video. Considering the consistent time step between training and inference, we sample consecutive three frames for the search sequence.", + "bbox": [ + 511, + 472, + 905, + 637 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "The overall optimization objective for DreamTrack includes dreaming loss $\\mathcal{L}_{\\mathrm{drm}}$ , reconstruction loss $\\mathcal{L}_{\\mathrm{rect}}$ from the Future Dreaming module, and probability loss $\\mathcal{L}_{\\mathrm{ce}}$ , L1 loss $\\mathcal{L}_{\\mathrm{L1}}$ and GIoU loss $\\mathcal{L}_{\\mathrm{GIoU}}$ from the Multimodal Prediction,", + "bbox": [ + 511, + 638, + 903, + 699 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {L} = \\lambda_ {1} \\mathcal {L} _ {\\mathrm {d r m}} + \\lambda_ {2} \\mathcal {L} _ {\\mathrm {r c t}} + \\lambda_ {3} \\mathcal {L} _ {\\mathrm {c e}} + \\lambda_ {4} \\mathcal {L} _ {\\mathrm {L 1}} + \\lambda_ {5} \\mathcal {L} _ {\\mathrm {G l o U}}, \\tag {12}\n$$\n", + "text_format": "latex", + "bbox": [ + 524, + 705, + 903, + 722 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "where $\\lambda$ values are the balancing weight coefficients.", + "bbox": [ + 511, + 727, + 862, + 742 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Inference. During inference, we initialize the hidden state of GRU for each new video sequence as previously described. Notably, it's no need to sample the video clip for inference. With the defined target queries that automatically cache the historical dynamics, we perform frame-level input for DreamTrack, which is consistent with previous nontemporal works [9, 51] and friendly for the deployment. Besides, the distribution mapping and image reconstruction in Fig. 3 are not required in the inference phase, which could be extended to other tasks, e.g., video generation.", + "bbox": [ + 511, + 750, + 906, + 900 + ], + "page_idx": 4 + }, + { + "type": "page_number", + "text": "7205", + "bbox": [ + 482, + 944, + 514, + 955 + ], + "page_idx": 4 + }, + { + "type": "table", + "img_path": "images/7f3a3000d974339bda5adb640005ef0aa20f29e8cd6d81fd374dd553487279c2.jpg", + "table_caption": [], + "table_footnote": [], + "table_body": "
MethodsLaSOT [15]LaSOText [16]GOT-10k* [27]TrackingNet [39]TNL2K [46]
SUC(%)PNorm(%)P(%)SUC(%)PNorm(%)P(%)AO(%)SR0.5(%)SR0.75(%)SUC(%)PNorm(%)P(%)SUC(%)P(%)
SiamFC255[2]33.642.033.923.031.126.934.835.39.857.166.353.329.528.6
ECO224[10]32.433.830.122.025.224.031.630.911.155.461.849.232.631.7
SiamRPN++255[32]49.656.949.134.041.639.651.761.632.573.380.069.441.341.2
DiMP288[3]56.965.056.739.247.645.161.171.749.274.080.168.744.743.4
SiamRCNN255[44]64.872.2----64.972.859.781.285.480.052.352.8
Ocean255[53]56.065.156.6---61.172.147.3---38.438.7
AutoMatch255[54]58.3-59.937.6-43.065.276.654.376.0-72.647.243.5
TrDiMP352[45]63.9-61.4---67.177.758.378.483.373.1--
TransT256[6]64.973.869.044.852.352.567.176.860.981.486.780.350.751.7
STARK320[50]67.177.0-47.856.255.168.878.164.182.086.9---
SwinTrack-B384[33]71.3-76.549.1-55.672.480.567.884.0-82.855.957.1
MixFormer-L320[9]70.179.976.3---72.682.268.883.988.983.1--
OSTrack384[51]71.181.177.650.561.357.673.783.270.883.988.583.255.956.7
GRM256[19]69.979.375.8---73.482.970.484.088.783.3--
MixViT288[9]69.679.975.9---72.582.469.983.588.382.0--
ROMTrack384[4]71.481.478.251.362.458.674.284.372.484.189.083.7--
MixViT-L384[9]72.482.280.1---75.785.375.185.490.285.7--
SeqTrack-L384[7]72.581.579.350.761.657.574.881.972.285.589.885.857.8-
ARTrack256[47]70.479.576.646.456.552.373.582.270.984.288.783.557.5-
ARTrack-L384[47]73.182.280.352.862.959.778.587.477.885.689.686.060.3-
ARTrackV2256[1]71.680.277.250.861.957.775.985.472.784.989.384.559.2-
ARTrackV2-L384[1]73.682.881.153.463.760.279.587.879.686.190.486.261.6-
LoRAT224[34]71.780.977.350.361.657.172.181.870.783.587.982.158.861.3
LoRAT-L378[34]75.184.182.056.669.065.177.586.278.185.689.785.462.367.0
DreamTrack25673.883.480.653.164.159.877.587.174.285.890.085.360.463.2
DreamTrack38475.084.281.754.565.361.178.387.976.686.590.685.961.264.8
DreamTrack-L38476.685.683.157.169.465.479.988.580.387.991.386.862.967.4
", + "bbox": [ + 112, + 88, + 883, + 419 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Table 1. State-of-the-art comparison on LaSOT [15], LaSOText [16], GOT-10k [27], TrackingNet [39] and TNL2K [46]. * indicates only trained on GOT-10k. The number in the subscript denotes the search region resolution.", + "bbox": [ + 89, + 422, + 906, + 454 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "4. Experiment", + "text_level": 1, + "bbox": [ + 89, + 464, + 217, + 481 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "4.1. Implementation Details", + "text_level": 1, + "bbox": [ + 89, + 489, + 307, + 505 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Model Variants. We develop three variants of DreamTrack with different configurations as follows:", + "bbox": [ + 89, + 511, + 482, + 541 + ], + "page_idx": 5 + }, + { + "type": "list", + "sub_type": "text", + "list_items": [ + "- DreamTrack256. Backbone: ViT-Base; Template size: $[128 \\times 128]$ ; Search region size: $[256 \\times 256]$ ;", + "- DreamTrack384. Backbone: ViT-Base; Template size: $[192 \\times 192]$ ; Search region size: $[384 \\times 384]$ ;", + "- DreamTrack-L384. Backbone: ViT-Large; Template size: $[192 \\times 192]$ ; Search region size: $[384 \\times 384]$ ." + ], + "bbox": [ + 89, + 541, + 482, + 631 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "The ViT backbone [13] is regarded as the encoder of our DreamTrack (see Fig. 2), which is divided into four stages for feature modeling with the proposed Temporal Self-Attention. The four stages contain 2,2,6,2 encoder layers for ViT-Base, and 4,4,12,4 ones for ViT-Large, respectively.", + "bbox": [ + 89, + 631, + 482, + 708 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Training Strategy. The training setting and evaluation protocol follow previous Transformer-based trackers (e.g., OSTrack [51]). The training datasets include COCO [35], LaSOT [15], TrackingNet [39] and GOT-10k [27] (with 1k sequences removed from GOT-10k train split, as per [50]). To ensure a fair evaluation of the GOT-10k test set, our models learn from the entire GOT-10k training split following its one-shot protocol. Our DreamTrack performs training with video clips, as described in Sec. 3.3. To align with the data form in the image dataset COCO that performs frame-level sampling, we repeat the sampled search frame with random shifts to constitute a pseudo video clip.", + "bbox": [ + 89, + 719, + 482, + 901 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "We train our model using the AdamW [37] optimizer with a weight decay of $1 \\times 10^{-4}$ . The learning rate for the encoder is set to $4 \\times 10^{-5}$ , while other parameters use a learning rate of $4 \\times 10^{-4}$ . The training process is conducted on 8 NVIDIA RTX3090 GPUs, comprising 300 epochs with 60k video sequences per epoch. The learning rate decreases by a factor of 10 after 240 epochs.", + "bbox": [ + 511, + 465, + 906, + 571 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "4.2. Main Results", + "text_level": 1, + "bbox": [ + 511, + 578, + 653, + 593 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "We evaluate our models on eight prevailing benchmark datasets, including LaSOT [15], LaSOText [16], GOT-10k [27], TrackingNet [39], TNL2K [46], OTB100 [49], NFS [29] and UAV123 [38].", + "bbox": [ + 511, + 599, + 905, + 661 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "LaSOT [15]. LaSOT is a benchmark designed for long-term tracking, comprising 280 videos in its test set. Tab. 1 demonstrates that our DreamTrack $_{256}$ already outperforms ARTrack- $L_{384}$ , which has a larger backbone. Furthermore, our DreamTrack- $L_{384}$ achieved the best performance with the SUC score of $76.6\\%$ , without the use of any online template update strategy or post-processing. This proves the superiority of our temporal design with future dreaming.", + "bbox": [ + 511, + 667, + 906, + 789 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "LaSOText [16]. LaSOText serves as an extension of LaSOT, which includes additional 150 video sequences. As shown in Tab. 1, our DreamTrack $_{256}$ achieves superior performance compared with other trackers with the same backbone, e.g., ARTrackV2 $_{256}$ and LoRAT $_{224}$ . With the larger backbone, the performance of our DreamTrack-L $_{384}$ further improves with the best SUC score of $57.1\\%$ .", + "bbox": [ + 511, + 794, + 908, + 902 + ], + "page_idx": 5 + }, + { + "type": "page_number", + "text": "7206", + "bbox": [ + 482, + 944, + 516, + 955 + ], + "page_idx": 5 + }, + { + "type": "table", + "img_path": "images/e0b73f55898f0d6051a23c65e2e478f14c10801d8816365c7c26ca7b581abaa9.jpg", + "table_caption": [], + "table_footnote": [], + "table_body": "
MethodOTB100 [49]NFS [29]UAV123 [38]
DreamTrack-L38472.467.473.1
DreamTrack38472.066.172.5
DreamTrack25671.865.472.1
LoRAT-L378 [34]72.066.772.5
LoRAT224 [34]71.464.072.6
ARTrackV2-L384 [1]-67.669.9
SeqTrack-L384 [7]-66.268.5
OSTrack384 [51]-66.570.7
TransT256 [6]69.565.769.1
DiMP288 [3]68.461.864.3
Ocean255 [53]68.449.457.4
ATOM288 [11]66.358.363.2
ECO224 [10]66.652.253.5
SiamFC255 [2]58.337.746.8
", + "bbox": [ + 133, + 88, + 439, + 268 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "GOT-10k [27]. GOT-10k provides 180 videos for testing with a strict protocol that only allows trackers to be trained exclusively on its training split. We adhere to this protocol and report the results in Tab. 1. Our DreamTrack- $\\mathrm{L}_{384}$ outperforms recent SOTA trackers across all metrics. Notably, our DreamTrack $_{256}$ surpasses other trackers with the same backbone, proving its effectiveness.", + "bbox": [ + 89, + 311, + 483, + 417 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "TrackingNet [39]. TrackingNet is an extensive tracking dataset that covers a wide range of real-world scenarios. We assess the performance of DreamTrack on its test set of 511 videos, as illustrated in Tab. 1. It shows that not only does our DreamTrack $_{384}$ outperform all other trackers in SUC, but our DreamTrack- $L_{384}$ also establishes a new state-of-the-art in three metrics on this large-scale benchmark.", + "bbox": [ + 89, + 425, + 483, + 532 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "TNL2K [46]. TNL2K is a recently proposed large-scale benchmark for tracking by natural language and bounding box initialization. We evaluate our tracker on its 700 testing videos and the results are listed in Tab. 1. Our DreamTrack-L384 achieves the best performance with the SUC and precision of $62.9\\%$ and $67.4\\%$ respectively, showing the effectiveness of our temporal learning.", + "bbox": [ + 89, + 537, + 483, + 643 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "OTB100 [49], NFS [29] and UAV123 [38]. We also evaluate our tracker on three additional benchmarks, i.e., OTB100, NFS and UAV123 including 100, 100 and 123 videos respectively. Tab. 2 shows that DreamTrack- $\\mathrm{L}_{384}$ achieves the best performance on all three benchmarks, demonstrating the strong generalizability of our framework.", + "bbox": [ + 89, + 650, + 482, + 743 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Efficiency Comparison. Tab. 3 compares the efficiency of our DreamTrack against other Transformer-based trackers in terms of inference speed, floating point operations, and number of parameters. Benefiting from the efficient one-stream architecture and sparse multimodal prediction, our DreamTrack shows impressive inference speed with small computation costs. Our most compact variant, DreamTrack $_{256}$ , achieves the state-of-the-art performance while operating at 139 frames per second (FPS), proving the effectiveness and efficiency of our method.", + "bbox": [ + 89, + 750, + 482, + 902 + ], + "page_idx": 6 + }, + { + "type": "table", + "img_path": "images/e543e4f843122d63975c181448d6bf1f63df8ac7bd2b0fb8b63d4bbc189f47a7.jpg", + "table_caption": [ + "Table 2. Performance comparison on OTB100 [49], NFS [29] and UAV123 [38] in terms of the SUC score $(\\%)$ ." + ], + "table_footnote": [], + "table_body": "
TrackerSpeed (FPS)FLOPs (G)Params (M)
SwinTrack-B384 [33]4569.791
OSTrack256 [51]105.421.594
OSTrack384 [51]58.148.394
SeqTrack-B256 [7]406689
SeqTrack-L384 [7]5524309
ARTrack256 [47]26395.6117
ROMTrack256 [4]6234.592
ROMTrack384 [4]2877.792
LoRAT224 [34]2093099
LoRAT-L378 [34]63325336
DreamTrack25613937.687
DreamTrack-L38432359.4317
", + "bbox": [ + 555, + 88, + 867, + 238 + ], + "page_idx": 6 + }, + { + "type": "table", + "img_path": "images/25e8051964c7dd2b8f88ed1e62c5d7bc714eed54fd705889b3b180c92ba4f146.jpg", + "table_caption": [ + "Table 3. Efficiency comparison. The inference speed is measured on a single NVIDIA RTX2080Ti GPU." + ], + "table_footnote": [], + "table_body": "
#Model VariantsLaSOT [15]TNL2K [46]FPS
SUC(%)PNorm(%)P(%)SUC(%)P(%)
①Baseline69.379.175.354.756.2162
②+ Future Dreaming Module71.781.377.756.959.5153
③+ Temporal Self-Attention72.682.078.559.160.4145
④+ Temporal Cross-Attention73.182.479.059.561.1143
⑤+ Multimodal Prediction73.883.480.660.463.2139
", + "bbox": [ + 517, + 281, + 903, + 367 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Table 4. Ablation on each proposed module in DreamTrack.", + "bbox": [ + 529, + 372, + 887, + 386 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "4.3. Component-wise Ablation", + "text_level": 1, + "bbox": [ + 511, + 402, + 751, + 419 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "We analyze the influence of each component in our DreamTrack $_{256}$ to show the effectiveness and rationality of the proposed modules, as shown in Tab. 4. When removing all designed modules and directly predicting the target location of the current frame based on the target query, the performance is still comparable with the fastest inference speed of 162 FPS. By applying the key Future Dreaming module on the baseline version, it obtains impressive SUC gains of $2.4\\%$ on LaSOT (2 v.s. 1). This verifies that deeper environmental understanding from past experience benefits the generalizability in novel situations. The Temporal Self/Cross-Attention further improves the performance for $0.9\\% / 0.5\\%$ SUC on LaSOT (4 v.s. 3 v.s. 2), proving the effectiveness of target-aware temporal refinement. The final Multimodal Prediction brings $0.7\\% / 0.9\\%$ SUC gains on LaSOT and TNL2K, respectively. It indicates that modeling the uncertainty of different future scenarios helps to forecast the target's motion and perform robust tracking.", + "bbox": [ + 511, + 425, + 906, + 699 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "4.4. Further Analysis", + "text_level": 1, + "bbox": [ + 511, + 710, + 681, + 727 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Self-supervision in Future Dreaming. As mentioned in Sec. 3.2, the capability of predicting the environmental future states relies on the self-supervision in the proposed Future Dreaming module. We explore the influence by ablating the loss functions based on $\\mathrm{DreamTrack}_{256}$ . As shown in Tab. 5, removing both the dreaming and reconstruction losses makes the dreamed future unreliable to reflect the real scenario in the next two frames, degrading the quality of learned temporal dynamics (①). Applying each of the two loss functions brings performance gains of $4.4\\% / 4.0\\%$ SUC on LaSOT respectively (②,③ v.s. ①), proving the ef", + "bbox": [ + 511, + 734, + 906, + 902 + ], + "page_idx": 6 + }, + { + "type": "page_number", + "text": "7207", + "bbox": [ + 482, + 944, + 514, + 955 + ], + "page_idx": 6 + }, + { + "type": "table", + "img_path": "images/2b7a13a23ddbdf4090df6471d23cf536367d954e8b0ea6a98b8cfbdea387c142.jpg", + "table_caption": [], + "table_footnote": [], + "table_body": "
#Dreaming LossReconstruction LossLaSOT [15]TNL2K [46]
SUC(%)PNorm(%)P(%)SUC(%)P(%)
①--68.377.473.853.753.2
②✓-72.782.679.859.161.7
③-✓72.382.179.258.560.3
④✓✓73.883.480.660.463.2
", + "bbox": [ + 112, + 88, + 459, + 161 + ], + "page_idx": 7 + }, + { + "type": "table", + "img_path": "images/64f99d191aab67809aea80d2ad9af5d5a2d1100286badfec1c09245431167a72.jpg", + "table_caption": [ + "Table 5. Ablation on loss functions in Future Dreaming module." + ], + "table_footnote": [], + "table_body": "
#Dreaming Step For EvaluationLaSOT [15]LaSOText [16]TNL2K [46]
SUC(%)P(%)SUC(%)P(%)SUC(%)P(%)
①073.880.653.159.860.463.2
②171.377.551.458.258.661.6
③268.574.748.955.055.357.8
", + "bbox": [ + 112, + 189, + 459, + 257 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Table 6. Performance comparison of different dreaming steps.", + "bbox": [ + 102, + 262, + 470, + 276 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "fectiveness. With the full supervision, the performance of DreamTrack $_{256}$ further improves with the SUC of $73.8\\%$ on LaSOT (4), which indicates the cooperation of our design.", + "bbox": [ + 89, + 286, + 483, + 333 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Performance of Future Prediction. With the defined target queries of different timestamps, our DreamTrack is capable of predicting the target trajectory including the current and next two frames. We then evaluate the performance of the future predictions by regarding them as the results, e.g., the future prediction for frame 10 from current frame 8 is used for the metric computation. As shown in Tab. 6, the performance of future prediction is inferior to current prediction (2,3 v.s. 1), and the tracking metrics degrade more with the prediction of further future. The underlying reason is that the future observation is unknowable for current prediction, in which the high degree of uncertainty raises challenges to future forecasting. In contrast, the previous trackers only predict the present target location with current observation. Despite these, the comparable performance demonstrates the potential of our design in the motion prediction of the target and benefits visual tracking.", + "bbox": [ + 89, + 344, + 483, + 602 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Activation Analysis and Result Visualization. Future dreaming endows our DreamTrack with superior generalization capability in novel situations. For deeper insights into the mechanism, we first visualize the activation maps $w$ and $w / o$ . The Future Dreaming module. As shown in Fig. 6, the attention areas without the future dreaming are easily distracted by the target's appearance deformation and background clutters, finally losing the target. This proves our claim that insufficient environmental understanding hardly handles the challenges in long-term tracking. In contrast, the version with future dreaming successfully locates the target area with the learned dynamics from past experience, showing the effectiveness of our design.", + "bbox": [ + 89, + 613, + 482, + 809 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "We also visualize the results of different trackers, as well as the dreamed future frames. As shown in Fig. 7, the proposed DreamTrack $_{256}$ delivers more robust tracking under deformation (the first row), occlusion and similar interferences (the last two rows). It demonstrates the superiority of our temporal learning in benefiting from both future and", + "bbox": [ + 89, + 810, + 483, + 900 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/44de60cf934d0d237447725bd03730110fddc3ca38ff9a3273de995db2a39023.jpg", + "image_caption": [ + "Figure 6. Attention visualization of our DreamTrack $w/$ and $w/o$ . the Future Dreaming module. The predictions and ground-truth boxes are marked in red and green, respectively." + ], + "image_footnote": [], + "bbox": [ + 514, + 89, + 903, + 284 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/d4712ebe160d156b9a462f42a1bb983b1b91303d6d6a8bff241e48e1845f00aa.jpg", + "image_caption": [ + "Figure 7. Results visualization of different trackers and dreamed future frames based on learned environmental dynamics." + ], + "image_footnote": [], + "bbox": [ + 516, + 339, + 903, + 503 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "historical information. Notably, the dreamed future frames correctly indicate the evolution of the complex environment, and the predicted trajectory successfully locates the tracking target. This indicates the potential of our method to be developed into video generation and annotation tasks.", + "bbox": [ + 511, + 544, + 906, + 619 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Due to limited space, we present more analyses in the supplementary material, including 1) more challenging benchmarks, 2) generalization in different paradigms, 3) position of future dreaming, 4) more visualizations, etc.", + "bbox": [ + 511, + 619, + 905, + 680 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "5. Conclusion", + "text_level": 1, + "bbox": [ + 511, + 694, + 633, + 709 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "We propose DreamTrack, a novel temporal framework with both historical and future information for visual object tracking. We reformulate the temporal learning in VOT as a History-to-Future process, which dreams the future states of the tracking scenario based on learned temporal dynamics from past experience, which empowers the tracker with the generalization in novel situations of new frames. The uncertainty of the future variation is modeled by forecasting the target trajectories of different possible future scenarios with multimodal prediction. Extensive experiments prove our tracker outperforms other mainstream trackers and achieves SOTA performance on prevailing benchmark datasets.", + "bbox": [ + 511, + 719, + 906, + 900 + ], + "page_idx": 7 + }, + { + "type": "page_number", + "text": "7208", + "bbox": [ + 482, + 944, + 514, + 955 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Acknowledgments", + "text_level": 1, + "bbox": [ + 91, + 90, + 250, + 107 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "This work was supported by the National Natural Science Foundation of China under Grant (62436001,62176020); the Joint Foundation of the Ministry of Education for Innovation Team (8091B042235); the Fundamental Research Funds for the Central Universities (2019JBZ110); and the State Key Laboratory of Rail Traffic Control and Safety (RCS2023K006), Beijing Jiaotong University.", + "bbox": [ + 89, + 114, + 483, + 220 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "References", + "text_level": 1, + "bbox": [ + 91, + 234, + 187, + 250 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[1] Yifan Bai, Zeyang Zhao, Yihong Gong, and Xing Wei. Artrackv2: Prompting autoregressive tracker where to look and how to describe. In CVPR, 2024. 2, 6, 7", + "[2] Luca Bertinetto, Jack Valmadre, João F Henriques, Andrea Vedaldi, and Philip H S Torr. Fully-convolutional siamese networks for object tracking. In ECCVW, 2016. 1, 2, 6, 7", + "[3] Goutam Bhat, Martin Danelljan, Luc Van Gool, and Radu Timofte. Learning discriminative model prediction for tracking. In ICCV, 2019. 6, 7", + "[4] Yidong Cai, Jie Liu, Jie Tang, and Gangshan Wu. Robust object modeling for visual tracking. In ICCV, 2023. 6, 7", + "[5] Fei Chen, Xiaodong Wang, Yunxiang Zhao, Shaohe Lv, and Xin Niu. Visual object tracking: A survey. Computer Vision and Image Understanding, 2022. 1", + "[6] Xin Chen, Bin Yan, Jiawen Zhu, Dong Wang, Xiaoyun Yang, and Huchuan Lu. Transformer tracking. In CVPR, 2021. 1, 2, 6, 7", + "[7] Xin Chen, Houwen Peng, Dong Wang, Hutchuan Lu, and Han Hu. Seqtrack: Sequence to sequence learning for visual object tracking. In CVPR, 2023. 1, 2, 3, 4, 6, 7", + "[8] Junyoung Chung, Caglar Gulcehre, KyungHyun Cho, and Yoshua Bengio. Empirical evaluation of gated recurrent neural networks on sequence modeling. arXiv preprint arXiv:1412.3555, 2014. 4", + "[9] Yutao Cui, Jiang Cheng, Limin Wang, and Gangshan Wu. Mixformer: End-to-end tracking with iterative mixed attention. In CVPR, 2022. 1, 2, 5, 6", + "[10] Martin Danelljan, Goutam Bhat, Fahad Shahbaz Khan, and Michael Felsberg. Eco: Efficient convolution operators for tracking. In CVPR, 2017. 1, 6, 7", + "[11] Martin Danelljan, Goutam Bhat, Fahad Shahbaz Khan, and Michael Felsberg. ATOM: Accurate tracking by overlap maximization. In CVPR, 2019. 7", + "[12] Martin Danelljan, Luc Van Gool, and Radu Timofte. Probabilistic regression for visual tracking. In CVPR, 2020. 1", + "[13] Alexey Dosovitskiy, Lucas Beyer, Alexander Kolesnikov, Dirk Weissenborn, Xiaohua Zhai, Thomas Unterthiner, Mostafa Dehghani, Matthias Minderer, Georg Heigold, Sylvain Gelly, et al. An image is worth 16x16 words: Transformers for image recognition at scale. In ICLR, 2021. 4, 6", + "[14] Frederik Ebert, Chelsea Finn, Sudeep Dasari, Annie Xie, Alex Lee, and Sergey Levine. Visual foresight: Model-based deep reinforcement learning for vision-based robotic control. arXiv preprint arXiv:1812.00568, 2018. 2" + ], + "bbox": [ + 93, + 258, + 483, + 900 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[15] Heng Fan, Liting Lin, Fan Yang, Peng Chu, Ge Deng, Sijia Yu, Hexin Bai, Yong Xu, Chunyuan Liao, and Haibin Ling. LaSOT: A high-quality benchmark for large-scale single object tracking. In CVPR, 2019. 2, 6, 7, 8", + "[16] Heng Fan, Hexin Bai, Liting Lin, Fan Yang, Peng Chu, Ge Deng, Sijia Yu, Mingzhen Huang, Juehuan Liu, Yong Xu, et al. Lasot: A high-quality large-scale single object tracking benchmark. IJCV, 2021. 6, 8", + "[17] Liangji Fang, Qinhong Jiang, Jianping Shi, and Bolei Zhou. Tpnet: Trajectory proposal network for motion prediction. In CVPR, 2020. 5", + "[18] Chelsea Finn and Sergey Levine. Deep visual foresight for planning robot motion. ICRA, 2016. 2", + "[19] Shenyuan Gao, Chunluan Zhou, and Jun Zhang. Generalized relation modeling for transformer tracking. In CVPR, 2023. 6", + "[20] Dongyan Guo, Jun Wang, Ying Cui, Zhenhua Wang, and Shengyong Chen. SiamCAR: Siamese fully convolutional classification and regression for visual tracking. In CVPR, 2020. 1, 2", + "[21] David Ha and Jürgen Schmidhuber. Recurrent world models facilitate policy evolution. NeurIPS, 2018. 2", + "[22] Danijar Hafner, Timothy Lillicrap, Jimmy Ba, and Mohammad Norouzi. Dream to control: Learning behaviors by latent imagination. *ICLR*, 2020. 2, 4", + "[23] Danijar Hafner, Timothy Lillicrap, Mohammad Norouzi, and Jimmy Ba. Mastering atari with discrete world models. *ICLR*, 2021. 2", + "[24] Danijar Hafner, Jurgis Pasukonis, Jimmy Ba, and Timothy Lillicrap. Mastering diverse domains through world models. arXiv preprint arXiv:2301.04104, 2023. 2, 4", + "[25] João F Henriques, Rui Caseiro, Pedro Martins, and Jorge Batista. High-speed tracking with kernelized correlation filters. IEEE T-PAMI, 2014. 1", + "[26] Anthony Hu, Gianluca Corrado, Nicolas Griffiths, Zachary Murez, Corina Gurau, Hudson Yeo, Alex Kendall, Roberto Cipolla, and Jamie Shotton. Model-based imitation learning for urban driving. NeurIPS, 2022. 2", + "[27] Lianghua Huang, Xin Zhao, and Kaiqi Huang. Got-10k: A large high-diversity benchmark for generic object tracking in the wild. IEEE T-PAMI, 2019. 6, 7", + "[28] Licheng Jiao, Dan Wang, Yidong Bai, Puhua Chen, and Fang Liu. Deep learning in visual tracking: A review. IEEE TNNLS, 2021. 1", + "[29] Hamed Kiani Galoogahi, Ashton Fagg, Chen Huang, Deva Ramanan, and Simon Lucey. Need for speed: A benchmark for higher frame rate object tracking. In ICCV, 2017. 6, 7", + "[30] Yann LeCun. A path towards autonomous machine intelligence version 0.9.2, 2022-06-27. Open Review, 2022. 2", + "[31] Bo Li, Junjie Yan, Wei Wu, Zheng Zhu, and Xiaolin Hu. High performance visual tracking with siamese region proposal network. In CVPR, 2018. 1, 2", + "[32] Bo Li, Wei Wu, Qiang Wang, Fangyi Zhang, Junliang Xing, and Junjie Yan. Siamrpn++: Evolution of siamese visual tracking with very deep networks. In CVPR, 2019. 2, 6", + "[33] Liting Lin, Heng Fan, Zhipeng Zhang, Yong Xu, and Haibin Ling. Swintrack: A simple and strong baseline for transformer tracking. NeurIPS, 2022. 2, 6, 7" + ], + "bbox": [ + 516, + 92, + 903, + 900 + ], + "page_idx": 8 + }, + { + "type": "page_number", + "text": "7209", + "bbox": [ + 482, + 944, + 514, + 955 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[34] Liting Lin, Heng Fan, Zhipeng Zhang, Yaowei Wang, Yong Xu, and Haibin Ling. Tracking meets lora: Faster training, larger model, stronger performance. ECCV, 2024. 1, 2, 6, 7", + "[35] Tsung-Yi Lin, Michael Maire, Serge J. Belongie, Lubomir D. Bourdev, Ross B. Girshick, James Hays, Pietro Perona, Deva Ramanan, Piotr Dollar, and C. Lawrence Zitnick. Microsoft COCO: Common objects in context. In ECCV, 2014. 6", + "[36] Yicheng Liu, Jinghuai Zhang, Liangji Fang, Qinhong Jiang, and Bolei Zhou. Multimodal motion prediction with stacked transformers. In CVPR, 2021. 5", + "[37] Ilya Loshchilov and Frank Hutter. Decoupled weight decay regularization. In ICLR, 2019. 6", + "[38] Matthias Mueller, Neil Smith, and Bernard Ghanem. A benchmark and simulator for uav tracking. In ECCV, 2016. 6, 7", + "[39] Matthias Muller, Adel Bibi, Silvio Giancola, Salman Alsubaihi, and Bernard Ghanem. TrackingNet: A large-scale dataset and benchmark for object tracking in the wild. In ECCV, 2018. 6, 7", + "[40] Scott Reed, Konrad Zolna, Emilio Parisotto, Sergio Gomez Colmenarejo, Alexander Novikov, Gabriel Barth-maron, Mai Giménez, Yury Sulsky, Jackie Kay, Jost Tobias Springenberg, et al. A generalist agent. TMLR, 2022. 2", + "[41] Shaoqing Ren, Kaiming He, Ross Girshick, and Jian Sun. Faster R-CNN: Towards real-time object detection with region proposal networks. In NIPS, 2015. 2, 3", + "[42] Hamid Rezatofighi, Nathan Tsoi, JunYoung Gwak, Amir Sadeghian, Ian Reid, and Silvio Savarese. Generalized intersection over union: A metric and a loss for bounding box regression. In CVPR, 2019. 5", + "[43] Yuval Tassa, Yotam Doron, Alistair Muldal, Tom Erez, Yazhe Li, Diego de Las Casas, David Budden, Abbas Abdelmaleki, Josh Merel, Andrew Lefrancq, et al. Deepmind control suite. arXiv preprint arXiv:1801.00690, 2018. 2", + "[44] Paul Voigtlaender, Jonathon Luiten, Philip HS Torr, and Bastian Leibe. Siam r-cnn: Visual tracking by re-detection. In CVPR, 2020. 2, 6", + "[45] Ning Wang, Wengang Zhou, Jie Wang, and Houqiang Li. Transformer meets tracker: Exploiting temporal context for robust visual tracking. In CVPR, 2021. 1, 2, 3, 6", + "[46] Xiao Wang, Xiujun Shu, Zhipeng Zhang, Bo Jiang, Yaowei Wang, Yonghong Tian, and Feng Wu. Towards more flexible and accurate object tracking with natural language: Algorithms and benchmark. In CVPR, 2021. 6, 7, 8", + "[47] Xing Wei, Yifan Bai, Yongchao Zheng, Dahu Shi, and Yihong Gong. Autoregressive visual tracking. In CVPR, 2023. 1, 2, 3, 4, 6, 7", + "[48] Philipp Wu, Alejandro Escontrela, Danijar Hafner, Pieter Abbeel, and Ken Goldberg. Daydreamer: World models for physical robot learning. In CoRL. PMLR, 2023. 2", + "[49] Yi Wu, Jongwoo Lim, and Ming Hsuan Yang. Object tracking benchmark. IEEE T-PAMI, 2015. 6, 7", + "[50] Bin Yan, Houwen Peng, Jianlong Fu, Dong Wang, and Huchuan Lu. Learning spatio-temporal transformer for visual tracking. In ICCV, 2021. 1, 2, 3, 6", + "[51] Botao Ye, Hong Chang, Bingpeng Ma, Shiguang Shan, and Xilin Chen. Joint feature learning and relation modeling for" + ], + "bbox": [ + 91, + 92, + 480, + 900 + ], + "page_idx": 9 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "tracking: A one-stream framework. In ECCV, 2022. 1, 2, 4, 5, 6, 7", + "[52] Zhipeng Zhang and Houwen Peng. Deeper and wider siamese networks for real-time visual tracking. In CVPR, 2019. 1, 2", + "[53] Zhipeng Zhang, Houwen Peng, Jianlong Fu, Bing Li, and Weiming Hu. Ocean: Object-aware anchor-free tracking. In ECCV, 2020. 2, 6, 7", + "[54] Zhipeng Zhang, Yihao Liu, Xiao Wang, Bing Li, and Weiming Hu. Learn to match: Automatic matching network design for visual tracking. ICCV, 2021. 6", + "[55] Xingyi Zhou, Dequan Wang, and Philipp Krahenbuhl. Objects as points. arXiv preprint arXiv:1904.07850, 2019. 2, 3" + ], + "bbox": [ + 516, + 92, + 903, + 287 + ], + "page_idx": 9 + }, + { + "type": "page_number", + "text": "7210", + "bbox": [ + 482, + 944, + 516, + 955 + ], + "page_idx": 9 + } +] \ No newline at end of file diff --git a/2025/DreamTrack_ Dreaming the Future for Multimodal Visual Object Tracking/2e4cde12-368f-4feb-b1b3-a980bb1eb002_model.json b/2025/DreamTrack_ Dreaming the Future for Multimodal Visual Object Tracking/2e4cde12-368f-4feb-b1b3-a980bb1eb002_model.json new file mode 100644 index 0000000000000000000000000000000000000000..69d18e337230b72d0fe15f3f7b95ece285f81814 --- /dev/null +++ b/2025/DreamTrack_ Dreaming the Future for Multimodal Visual Object Tracking/2e4cde12-368f-4feb-b1b3-a980bb1eb002_model.json @@ -0,0 +1,2156 @@ +[ + [ + { + "type": "header", + "bbox": [ + 0.107, + 0.003, + 0.182, + 0.044 + ], + "angle": 0, + "content": "CVF" + }, + { + "type": "header", + "bbox": [ + 0.237, + 0.001, + 0.812, + 0.047 + ], + "angle": 0, + "content": "This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore." + }, + { + "type": "title", + "bbox": [ + 0.12, + 0.131, + 0.88, + 0.154 + ], + "angle": 0, + "content": "DreamTrack: Dreaming the Future for Multimodal Visual Object Tracking" + }, + { + "type": "text", + "bbox": [ + 0.158, + 0.18, + 0.846, + 0.22 + ], + "angle": 0, + "content": "Mingzhe Guo \\(^{1*}\\) Weiping Tan \\(^{1}\\) Wenyu Ran \\(^{1}\\) Liping Jing \\(^{1\\dagger}\\) Zhipeng Zhang \\(^{2*}\\) Beijing Jiaotong University \\(^{1}\\), Shanghai Jiaotong University \\(^{2}\\)" + }, + { + "type": "title", + "bbox": [ + 0.248, + 0.252, + 0.327, + 0.267 + ], + "angle": 0, + "content": "Abstract" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.283, + 0.485, + 0.661 + ], + "angle": 0, + "content": "Aiming to achieve class-agnostic perception in visual object tracking, current trackers commonly formulate tracking as a one-shot detection problem with the template-matching architecture. Despite the success, severe environmental variations in long-term tracking raise challenges to generalizing the tracker in novel situations. Temporal trackers try to fix it by preserving the time-validity of target information with historical predictions, e.g., updating the template. However, solely transmitting the previous observations instead of learning from them leads to an inferior capability of understanding the tracking scenario from past experience, which is critical for the generalization in new frames. To address this issue, we reformulate temporal learning in visual tracking as a History-to-Future process and propose a novel tracking framework DreamTrack. Our DreamTrack learns the temporal dynamics from past observations to dream the future variations of the environment, which boosts the generalization with the extended future information from history. Considering the uncertainty of future variation, multimodal prediction is designed to infer the target trajectory of each possible future situation. The experiments demonstrate that our DreamTrack achieves leading performance with real-time inference speed. In particular, DreamTrack obtains SUC scores of \\(76.6\\% /87.9\\%\\) on LasOT/TrackingNet, surpassing all recent SOTA trackers." + }, + { + "type": "title", + "bbox": [ + 0.092, + 0.688, + 0.222, + 0.703 + ], + "angle": 0, + "content": "1. Introduction" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.713, + 0.484, + 0.88 + ], + "angle": 0, + "content": "Visual Object Tracking (VOT) is one of the most fundamental computer vision problems with numerous applications [5, 28], which aims to estimate the position of an arbitrary target in a video sequence, given only its location in the initial frame. To achieve the class-agnostic perception, recent prevalent Siamese-based approaches [2, 6, 7, 9, 20, 31, 34, 47, 51, 52] typically formulate visual tracking as a one-shot detection problem. Specifically, these trackers perform general object tracking by matching with the initialized target template and then predicting the location in a single forward pass, as shown in Fig. 1(a)." + }, + { + "type": "image", + "bbox": [ + 0.513, + 0.248, + 0.907, + 0.54 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.512, + 0.544, + 0.907, + 0.656 + ], + "angle": 0, + "content": "Figure 1. Comparison of different temporal architectures. (a) The non-temporal architecture that regards visual tracking as a one-shot detection problem. (b) The temporal architecture with only historical information that limits the temporal messages in the transmission of previous observations. (c) Our proposed temporal architecture with both historical and future information, which learns environmental dynamics from past experiences and dreams the future states for better generalization in subsequent frames." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.675, + 0.909, + 0.903 + ], + "angle": 0, + "content": "Despite the demonstrated success, it's challenging for the non-temporal architecture to ensure the generalization in new frames with severe environmental variations compared with the initial frame, e.g., deformation of the target appearance and background in long-term tracking. The outdated template hardly provides effective target-aware knowledge but distracts the tracking process. Then many works [7, 10, 12, 25] intuitively try to fix it by keeping the time-validity of target information, e.g., updating the template with predictions of high confidence scores in previous frames [45, 50] or passing historical trajectory as prior knowledge [7, 47], as shown in Fig. 1(b). However, these efforts focus on only transmitting previous observations instead of learning from them, leading to an inferior capability of understanding the tracking scenario from the past expe" + }, + { + "type": "page_footnote", + "bbox": [ + 0.109, + 0.887, + 0.353, + 0.901 + ], + "angle": 0, + "content": "* Equal contribution. † Corresponding author." + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.514, + 0.957 + ], + "angle": 0, + "content": "7201" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.091, + 0.092, + 0.461, + 0.106 + ], + "angle": 0, + "content": "rience so that enables generalization to novel situations." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.109, + 0.486, + 0.501 + ], + "angle": 0, + "content": "To address this issue, we reformulate the temporal learning in VOT as a History-to-Future process, as shown in Fig. 1(c). The pipeline dreams the future states of the tracking scenario based on learned temporal dynamics from previous observations, which empowers the tracker with superior generalization ability in complex environments of new frames. Our intuition stems from the human recognition process [21, 30]. The brain abstracts the world from past senses and imagines future variations, which aims to achieve goals in complex scenarios even though they are never encountered before. Such a temporal modeling manner is intuitively superior to the one limited in historical information, which further develops the generalization in new frames with the future information from dreamed states. To this end, we construct a future dreaming module, originally proposed in the world model for reinforcement learning tasks [21-23], to endow the tracker with the capability of predicting future variations of the target from the learned historical dynamics (see Fig. 1(c)). In contrast to non-temporal [6, 9, 31] or historical information-only trackers [7, 47, 50], our dreaming module extends the temporal messages from the history to its future, in which the dreamed future frames are self-supervised with ground-truth future observations. The future information undoubtedly facilitates better generalization for appearance deformation of the target and background variance." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.502, + 0.485, + 0.744 + ], + "angle": 0, + "content": "More specifically, we propose a novel tracking framework dubbed DreamTrack, which has a dream-encoder-decoder architecture. The dreaming stage caches historical observations and predicts future states to represent the environmental variation in the next few frames. Then the features of the current and dreamed future frames are input to the encoder, which extracts target-aware features with the temporal messages from both history and future. Notably, after each feature extraction stage of the encoder, the temporal dynamics will be refined along the timeline, which aims to filter target-irrelevant interferences with the exposure of the target information. The final decoder predicts the trajectory of the target in the current and future frames. Considering the uncertainty of the future variation, we forecast the probabilities and target locations of different possible future scenarios, i.e., multimodal prediction." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.745, + 0.485, + 0.852 + ], + "angle": 0, + "content": "Extensive experiments on eight prevailing VOT benchmarks show that our proposed DreamTrack outperforms recent state-of-the-art trackers. For instance, under fair settings, DreamTrack-\\(\\mathrm{L}_{384}\\) obtains a \\(76.6\\%\\) SUC score on the LaSOT [15] dataset, surpassing ARTrackV2-\\(\\mathrm{L}_{384}\\) [1] by \\(3.0\\%\\) and LoRAT-\\(\\mathrm{L}_{378}\\) [34] by \\(1.5\\%\\), respectively. In summary, the contributions of this work are as follows:" + }, + { + "type": "text", + "bbox": [ + 0.091, + 0.856, + 0.485, + 0.903 + ], + "angle": 0, + "content": "- We reformulate the tracking problem as a History-to-Future process and are the first to develop the temporal messages into the future frames for better generalization" + }, + { + "type": "text", + "bbox": [ + 0.527, + 0.092, + 0.818, + 0.107 + ], + "angle": 0, + "content": "of novel situations in visual object tracking." + }, + { + "type": "text", + "bbox": [ + 0.514, + 0.108, + 0.905, + 0.182 + ], + "angle": 0, + "content": "- We make the first attempt at predicting the future states of the target and performing multimodal prediction to include the uncertainty of different future scenarios, which empowers the tracker with superior generalization for appearance deformation and background variation." + }, + { + "type": "text", + "bbox": [ + 0.514, + 0.183, + 0.905, + 0.227 + ], + "angle": 0, + "content": "- We prove the effectiveness of the proposed future dreaming to facilitate tracking with state-of-the-art performance on eight prevailing VOT benchmarks." + }, + { + "type": "list", + "bbox": [ + 0.514, + 0.108, + 0.905, + 0.227 + ], + "angle": 0, + "content": null + }, + { + "type": "title", + "bbox": [ + 0.514, + 0.242, + 0.656, + 0.258 + ], + "angle": 0, + "content": "2. Related Work" + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.268, + 0.908, + 0.601 + ], + "angle": 0, + "content": "Visual Object Tracking. Current prevailing trackers [2, 6, 9, 20, 31, 53] commonly employ a template-matching framework, which exploits the initialized template to match the target within the search region. Regarding tracking as a one-shot detection problem, these approaches develop a deep backbone to extract visual features [33, 51, 52], then predict the current target location following the detection paradigms [41, 55], such as object scale estimation [32, 44, 53] and center point localization [6, 20]. Observing the target information of the template may be weakened by the appearance deformation and background variation in long-term tracking, many works design modules to ensure the time-validity of target-relevant knowledge, e.g., updating the template with predictions of high confidence scores in previous frames [45, 50] or passing historical trajectory as prior knowledge [7, 47]. Despite the improved performance in long-term tracking, the limited understanding of the tracking scenario leads to suboptimal generalization in novel situations. Thus, we reformulate tracking as a History-to-Future process that learns the environmental dynamics from past experiences and dreams the future states for better adaptation and tracking capability in new frames." + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.614, + 0.909, + 0.901 + ], + "angle": 0, + "content": "World Model. The world model [30] aims to learn a general representation of the world and predicts future world states resulting from a sequence of actions, which has been widely studied in the game [21-23, 40] and lab environments [14, 18, 48]. Dreamer [22] learns a latent dynamics model from past experience to predict future state values and actions in a latent space. It's capable of handling challenging visual control tasks in the DeepMind Control Suite [43]. DreamerV2 [23] further achieves human-level performance on Atari games with discrete modeling. DreamerV3 [24] develops a general algorithm and allows solving challenging control problems such as obtaining diamonds in Minecraft from scratch given sparse rewards. Recently, learning world models in driving scenes has gained attention. MILE [26] employs a model-based imitation learning method to jointly learn a dynamics model and driving behavior. In this work, we introduce the mechanism of the world model to visual object tracking, which boosts environmental understanding with the dreamed future states." + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.516, + 0.957 + ], + "angle": 0, + "content": "7202" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.093, + 0.089, + 0.907, + 0.317 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.09, + 0.328, + 0.907, + 0.398 + ], + "angle": 0, + "content": "Figure 2. The dream-encoder-decoder framework of our DreamTrack. Initially, the dreaming stage takes the input image tokens to predict future states of the next two frames based on the cached temporal dynamics in the Future Dreaming module. Then features of the current frame and dreamed future are fed into the encoder for further modeling, with a Temporal Self-Attention after each encoder stage for target-aware refinement. In the end, the decoder performs multimodal prediction for the target, i.e., the probabilities and trajectories in the current and next two frames of three possible future scenarios. The prediction with the highest probability is selected as the final output." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.422, + 0.182, + 0.437 + ], + "angle": 0, + "content": "3. Method" + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.447, + 0.232, + 0.461 + ], + "angle": 0, + "content": "3.1. Preliminaries" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.469, + 0.484, + 0.605 + ], + "angle": 0, + "content": "Formulation of VOT. Given the template image \\( \\mathbf{z}^0 \\) of a target initialized in the first frame, the objective of visual tracking is to localize the target \\( \\mathrm{b}^t \\) from the continuous input image \\( \\mathbf{x}^t \\), where \\( \\mathbf{b} \\) denotes the target state, i.e., a bounding box. The mainstream tracking paradigm aims to learn a deep tracking model \\( f \\) that performs localization with the output \\( \\mathbf{y}^t \\) following different detection paradigms, i.e., anchor-based [41] or anchor-free [55] architectures. The tracking process can be formulated as," + }, + { + "type": "equation", + "bbox": [ + 0.234, + 0.614, + 0.482, + 0.632 + ], + "angle": 0, + "content": "\\[\ny ^ {t} = f \\left(z ^ {0}, x ^ {t}\\right). \\tag {1}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.642, + 0.484, + 0.732 + ], + "angle": 0, + "content": "To alleviate the influence of target's appearance deformation and background variation in long-term tracking, recent prevailing temporal trackers exploit previous outputs \\( \\mathrm{y}^{0:t - 1} \\) as additional prior knowledge, e.g., updating the template [45, 50] or passing historical target location [7, 47]. The temporal tracking process is formulated as," + }, + { + "type": "equation", + "bbox": [ + 0.209, + 0.741, + 0.483, + 0.759 + ], + "angle": 0, + "content": "\\[\n\\mathrm {y} ^ {t} = f \\left(\\mathrm {z} ^ {0}, \\mathrm {x} ^ {t}, \\mathrm {y} ^ {0: t - 1}\\right). \\tag {2}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.77, + 0.484, + 0.875 + ], + "angle": 0, + "content": "In contrast, we formulate temporal visual tracking as a History-to-Future process, which predicts the future trajectory of the target (including the current frame and the next two ones) based on the learned environmental dynamics from historical observations (cached in the target queries \\(\\mathbf{Q}_{\\mathrm{tgt}}^{t:t + 2}\\) of \\(f\\), described in Sec. 3.2). This modeling pipeline can be expressed as," + }, + { + "type": "equation", + "bbox": [ + 0.194, + 0.884, + 0.483, + 0.903 + ], + "angle": 0, + "content": "\\[\ny ^ {t: t + 2} = f \\left[ Q _ {\\mathrm {t g t}} ^ {t: t + 2} \\right] \\left(z ^ {0}, x ^ {t}\\right). \\tag {3}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.424, + 0.907, + 0.5 + ], + "angle": 0, + "content": "The parameterized modeling of environmental dynamics \\((\\mathbf{Q}_{\\mathrm{tg}}^{t:t + 2})\\) instead of explicit transmission of historical output \\(\\mathrm{y}^{0:t - 1}\\), not only enables better generalization in novel situations of future frames, but also reduces the risk of error accumulation from past experiences." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.509, + 0.907, + 0.796 + ], + "angle": 0, + "content": "Trajectory Coordinate Mapping. To reduce computation costs, most methods perform tracking on the cropped search region instead of the full-resolution frame, which outputs target's coordinates in the current frame, relative to the search region. Thus, it's necessary to map the predicted target location of the relative coordinate system to the original frame after inference, which we denote as the global coordinate system. Correspondingly, the ground-truth boxes for training are also transformed from the global to relative coordinate system. For our DreamTrack that learns dynamics and predicts target trajectory over a video sequence, we additionally cache the ground-truth boxes of the preceding two frames in the global coordinate system and map them to the relative coordinate system of the current frame after the search region is cropped. Notably, the search region patches of the two future frames are cropped by referring to the ground-truth box of the current frame, which represent the target's motion and serve for the supervision of future image reconstruction (described in Sec. 3.2 and Fig. 3)." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.804, + 0.72, + 0.818 + ], + "angle": 0, + "content": "3.2. Network Architecture" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.825, + 0.906, + 0.901 + ], + "angle": 0, + "content": "The overall framework of DreamTrack is depicted in Fig. 2. Given the target sequences constructed from object trajectories, we use a novel dream-encoder-decoder structure for learning and inference. The target state is described by a trajectory including the current and next two frames." + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.516, + 0.957 + ], + "angle": 0, + "content": "7203" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.097, + 0.092, + 0.48, + 0.278 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.09, + 0.285, + 0.483, + 0.37 + ], + "angle": 0, + "content": "Figure 3. Illustration of the proposed Future Dreaming module. The target queries auto-regressively update the temporal dynamics with the patch embeddings of the current frame through GRU, which are then exploited to predict the future states. For training, the distributions of the predicted future states and decoded future images are self-supervised by the ground-truth future frames." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.387, + 0.484, + 0.599 + ], + "angle": 0, + "content": "Dream. We propose the Future Dreaming module to learn temporal dynamics of the tracking scenario. As shown in Fig. 3, the module evolves the dynamics with current observation and dreams the future states, which endows the tracker with the capability of future prediction by supervising with ground-truth future frames. Specifically, three trainable target queries \\(\\mathbf{Q}_{\\mathrm{tgt}}^{t - 1:t + 1} \\in \\mathbb{R}^{3 \\times C}\\), which represent the transition dynamics of the current and next two frames, are initialized. With the template \\(z^0\\) and search region \\(x^t\\) of each new frame, the images are split and flattened to build current observation feature \\(\\mathbf{F}^{\\mathrm{t}} = \\{\\mathbf{F}_z^0 \\in \\mathbb{R}^{H_z W_z \\times C}, \\mathbf{F}_x^t \\in \\mathbb{R}^{H_x W_x \\times C}\\}\\) with a linear projection. Then target queries auto-regressively update the cached temporal messages with \\(\\mathbf{F}^{\\mathrm{t}}\\) through a gated recurrent unit (GRU) [8]," + }, + { + "type": "equation", + "bbox": [ + 0.18, + 0.612, + 0.483, + 0.631 + ], + "angle": 0, + "content": "\\[\n\\mathbf {Q} _ {\\mathrm {t g t}} ^ {t: t + 2} = \\operatorname {G R U} \\left(\\mathbf {Q} _ {\\mathrm {t g t}} ^ {t - 1: t + 1}, \\mathbf {F} ^ {t}\\right). \\tag {4}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.635, + 0.484, + 0.74 + ], + "angle": 0, + "content": "As the aim of our DreamTrack is to enhance the generalization in novel situations of new frames, it's a prerequisite to evolve the future variations of the tracking environment. Thus, the updated target queries \\(\\mathbf{Q}_{\\mathrm{tgt}}^{t:t + 2}\\) are exploited to backward imagine the future observation feature \\(\\hat{\\mathbf{F}}^{t:t + 2} = \\{\\hat{\\mathbf{F}}_{\\mathrm{z}}^{t:t + 2}\\in \\mathbb{R}^{3\\times H_{\\mathrm{z}}W_{\\mathrm{z}}\\times C},\\hat{\\mathbf{F}}_{\\mathrm{x}}^{t:t + 2}\\in \\mathbb{R}^{3\\times H_{\\mathrm{x}}W_{\\mathrm{x}}\\times C}\\}\\) based on \\(\\mathbf{F}^{\\mathrm{t}}\\) with a cross-attention," + }, + { + "type": "equation", + "bbox": [ + 0.11, + 0.747, + 0.483, + 0.768 + ], + "angle": 0, + "content": "\\[\n\\hat {\\mathbf {F}} ^ {t: t + 2} = \\operatorname {C r o s s A t t e n t i o n} (q = \\mathbf {F} ^ {t}, k v = \\mathbf {Q} _ {\\mathrm {t g t}} ^ {t: t + 2}). \\quad (5)\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.774, + 0.483, + 0.87 + ], + "angle": 0, + "content": "Notably, the interaction between \\(\\mathbf{F}^t\\) and \\(\\mathbf{Q}_{\\mathrm{tgt}}^t\\) to generate \\(\\hat{\\mathbf{F}}^t\\) intends to enhance the feature quality of the current frame by introducing historical information, which originates from \\(\\mathbf{Q}_{\\mathrm{tgt}}^{t - 1}\\) before the GRU update (Eq. 4). The output \\(\\hat{\\mathbf{F}}^{t:t + 2}\\) of the module will be further modeled with the Temporal Self-Attention, which is detailed in the encoder." + }, + { + "type": "text", + "bbox": [ + 0.091, + 0.871, + 0.484, + 0.901 + ], + "angle": 0, + "content": "Following the loss design in world models [22-24], we respectively map the dreamed future observation fea" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.09, + 0.905, + 0.304 + ], + "angle": 0, + "content": "ture \\(\\hat{\\mathbf{F}}_{\\mathrm{x}}^{t:t + 2}\\) and ground-truth observation feature \\(\\mathbf{F}_{\\mathrm{x}}^{t:t + 2}\\) to distributions of \\(\\{\\hat{\\mu}_{\\mathrm{x}}^{t:t + 2},\\hat{\\sigma}_{\\mathrm{x}}^{t:t + 2}\\in \\mathbb{R}^{3\\times H_{\\mathrm{x}}W_{\\mathrm{x}}\\times C^{\\prime}}\\}\\) and \\(\\{\\mu_{\\mathrm{x}}^{t:t + 2},\\sigma_{\\mathrm{x}}^{t:t + 2}\\in \\mathbb{R}^{3\\times H_{\\mathrm{x}}W_{\\mathrm{x}}\\times C^{\\prime}}\\}\\) with MLP layers, and then minimize their KL divergence. Notably, we only consider supervising the dreaming process of the search region, as the input template always comes from the initial frame. The prior distribution from \\(\\hat{\\mathbf{F}}_{\\mathrm{x}}^{t:t + 2}\\) is regarded as a prediction of the future dynamics. In contrast, the posterior distribution from \\(\\mathbf{F}_{\\mathrm{x}}^{t:t + 2}\\) represents the future dynamics from the ground-truth observation \\(\\mathbf{x}^{t:t + 2}\\). The KL divergence measures the gap between the imagined future (prior) and the true future (posterior). We expect to enhance the capability of future prediction for long-term tracking, which is realized by optimizing the dreaming loss \\(\\mathcal{L}_{\\mathrm{drm}}\\)" + }, + { + "type": "equation", + "bbox": [ + 0.54, + 0.31, + 0.905, + 0.33 + ], + "angle": 0, + "content": "\\[\n\\mathcal {L} _ {\\mathrm {d r m}} = \\mathrm {K L} \\left(\\left\\{\\hat {\\mu} _ {\\mathrm {x}} ^ {t: t + 2}, \\hat {\\sigma} _ {\\mathrm {x}} ^ {t: t + 2} \\right\\} \\mid \\mid \\left\\{\\mu_ {\\mathrm {x}} ^ {t: t + 2}, \\sigma_ {\\mathrm {x}} ^ {t: t + 2} \\right\\}\\right), \\tag {6}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.336, + 0.8, + 0.351 + ], + "angle": 0, + "content": "where \\(\\mathrm{KL}(\\cdot ||\\cdot)\\) denotes the KL divergence." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.352, + 0.905, + 0.456 + ], + "angle": 0, + "content": "Besides the supervision from the distribution difference, we further improve the accuracy of future prediction by reconstructing the future search regions based on \\(\\hat{\\mathbf{F}}_{\\mathrm{x}}^{t:t + 2}\\) with a decoder of several CNN upsampling layers. The decoded future search regions \\(\\hat{\\mathbf{x}}^{t:t + 2}\\) is supervised by the groundtruth search regions \\(\\mathbf{x}^{t:t + 2}\\) by optimizing the reconstruction loss \\(\\mathcal{L}_{\\mathrm{RCT}}\\)" + }, + { + "type": "equation", + "bbox": [ + 0.613, + 0.456, + 0.905, + 0.473 + ], + "angle": 0, + "content": "\\[\n\\mathcal {L} _ {\\mathrm {r c t}} = \\operatorname {M S E} \\left(\\hat {\\mathrm {x}} ^ {t: t + 2} | | \\mathrm {x} ^ {t: t + 2}\\right), \\tag {7}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.478, + 0.801, + 0.494 + ], + "angle": 0, + "content": "where \\(\\mathrm{MSE}(\\cdot ||\\cdot)\\) is the mean squared error." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.501, + 0.906, + 0.622 + ], + "angle": 0, + "content": "Encoder. Following previous works [7, 47, 51], we adopt a plain vision transformer (ViT) [13] as the encoder and model the search and template images jointly. The encoder extracts the features of the search and template images and learns feature-level correspondence to filter target-relevant distractors. The input dreamed observation features \\(\\{\\hat{\\mathbf{F}}_{\\mathrm{z}}^{t:t + 2},\\hat{\\mathbf{F}}_{\\mathrm{x}}^{t:t + 2}\\}\\) are concatenated along the spatial dimension and fed into the encoder for feature extraction," + }, + { + "type": "equation", + "bbox": [ + 0.563, + 0.628, + 0.905, + 0.671 + ], + "angle": 0, + "content": "\\[\n\\begin{array}{l} \\hat {\\mathbf {F}} _ {\\mathrm {z x}, i} ^ {t: t + 2} = \\operatorname {C o n c a t} \\left(\\hat {\\mathbf {F}} _ {\\mathrm {z}} ^ {t: t + 2}, \\hat {\\mathbf {F}} _ {\\mathrm {x}} ^ {t: t + 2}\\right), i = 0 \\\\ \\hat {\\mathbf {F}} _ {\\mathrm {z x}, i} ^ {t: t + 2} = \\operatorname {B l o c k} _ {i} \\left(\\hat {\\mathbf {F}} _ {\\mathrm {z x}, i - 1} ^ {t: t + 2}\\right), \\quad 1 \\leq i \\leq l _ {\\text {s p t}} \\tag {8} \\\\ \\end{array}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.678, + 0.906, + 0.725 + ], + "angle": 0, + "content": "where \\(\\hat{\\mathbf{F}}_{\\mathrm{zx},i}^{t:t + 2}\\in \\mathbb{R}^{3\\times (H_{\\mathrm{z}}W_{\\mathrm{z}} + H_{\\mathrm{x}}W_{\\mathrm{x}})\\times C}\\), Block denotes the transformer block in the encoder for spatial modeling and \\(l_{\\mathrm{spt}}\\) is the layer number." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.726, + 0.906, + 0.877 + ], + "angle": 0, + "content": "Besides the spatial modeling between the template and search region in Block, it's also important to exploit the extracted target-aware information to refine the temporal dynamics along the timeline. Thus, after each feature extraction stage of the encoder, we arrange a Temporal Self-Attention module to filter target-irrelevant interferences with the exposure of the target information from different levels. As shown in Fig. 4, the dreamed observation features \\(\\hat{\\mathbf{F}}_{\\mathrm{zx}}^{t:t + 2}\\) are transposed and input to the attention for the interaction on the temporal dimension," + }, + { + "type": "equation", + "bbox": [ + 0.54, + 0.883, + 0.905, + 0.904 + ], + "angle": 0, + "content": "\\[\n\\hat {\\mathbf {F}} _ {\\mathrm {z x}, i} ^ {t: t + 2} = \\operatorname {T e m A t t n} _ {\\text {s e l f}} ^ {i} \\left(\\hat {\\mathbf {F}} _ {\\mathrm {z x}, i - 1} ^ {t: t + 2}\\right), \\quad 1 \\leq i \\leq l _ {\\text {t e m}} \\tag {9}\n\\]" + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.516, + 0.957 + ], + "angle": 0, + "content": "7204" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.094, + 0.089, + 0.483, + 0.256 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.09, + 0.262, + 0.484, + 0.332 + ], + "angle": 0, + "content": "Figure 4. Illustration of the Temporal Self/Cross-Attention. For self-attention, patch embeddings of the template and search region are jointly modeled on the temporal dimension. In contrast, the cross-attention takes target queries as \\(\\mathbf{Q}\\) and patch embeddings of search region as \\(\\mathbf{K}\\), \\(\\mathbf{V}\\) to extract localization-relevant information." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.348, + 0.484, + 0.395 + ], + "angle": 0, + "content": "where \\(\\mathrm{TemAttn}_{\\mathrm{self}}\\) denotes the Temporal Self-Attention module and \\(l_{\\mathrm{tem}}\\) is the module number corresponding to the four stages of the encoder (see Fig. 2)." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.402, + 0.484, + 0.59 + ], + "angle": 0, + "content": "Decoder. The final decoder is responsible for predicting the target trajectory in the current and next two frames based on the updated target queries \\(\\mathbf{Q}_{\\mathrm{tgt}}^{t:t + 2}\\) and target-aware observation feature of the search region \\(\\hat{\\mathbf{F}}_{\\mathrm{x}}^{t:t + 2}\\), which is split from the encoder output \\(\\hat{\\mathbf{F}}_{\\mathrm{zx}}^{t:t + 2}\\). Notably, the dreamed feature of the template \\(\\hat{\\mathbf{F}}_{\\mathrm{z}}^{t:t + 2}\\) is not employed as the ultimate is to locate the target in the search region and the target information has been introduced into \\(\\hat{\\mathbf{F}}_{\\mathrm{x}}^{t:t + 2}\\) after encoder modeling. The target queries \\(\\mathbf{Q}_{\\mathrm{tgt}}^{t:t + 2}\\) first extract localization-relevant messages from \\(\\hat{\\mathbf{F}}_{\\mathrm{x}}^{t:t + 2}\\) with the Temporal Cross-Attention module. As shown in Fig. 4, each target query interacts with the region feature of the same timestamp," + }, + { + "type": "equation", + "bbox": [ + 0.1, + 0.594, + 0.482, + 0.615 + ], + "angle": 0, + "content": "\\[\n\\hat {\\mathbf {Q}} _ {\\mathrm {t g t}} ^ {t: t + 2} = \\operatorname {T e m A t t n} _ {\\text {c r o s s}} (q = \\mathbf {Q} _ {\\mathrm {t g t}} ^ {t: t + 2}, k v = \\hat {\\mathbf {F}} _ {\\mathrm {x}} ^ {t: t + 2}), \\tag {10}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.621, + 0.483, + 0.666 + ], + "angle": 0, + "content": "where \\(\\mathrm{TemAttn}_{\\mathrm{cross}}\\) is the Temporal Cross-Attention module. The output target queries \\(\\hat{\\mathbf{Q}}_{\\mathrm{tgt}}^{t:t + 2}\\) are then used to predict the target trajectory of the current and next two frames." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.667, + 0.484, + 0.877 + ], + "angle": 0, + "content": "The key challenge to future prediction is the high degree of uncertainty, in large part due to not knowing the intents and latent characteristics of the tracking target. For example, the target commonly has a multimodal distribution of futures: it could turn, go straight, slow down, speed up, etc. The high degree of multimodality also exists in the motion prediction of autonomous driving [17, 36], which correspondingly designs a multimodal prediction to generate multiple plausible trajectories of the target vehicle. We draw inspiration from this solution and design the tailored Multimodal Prediction to our DreamTrack. As shown in Fig. 5, each target query of one timestamp forecasts the probabilities \\(\\hat{\\mathbf{p}}^T\\in \\mathbb{R}^{3\\times 1}\\) and target states \\(\\hat{\\mathbf{b}}^T\\in \\mathbb{R}^{3\\times 4}\\) of three possible future situations with a stack of MLP layers," + }, + { + "type": "equation", + "bbox": [ + 0.121, + 0.883, + 0.483, + 0.904 + ], + "angle": 0, + "content": "\\[\n\\hat {\\mathrm {p}} ^ {T}, \\hat {\\mathrm {b}} ^ {T} = \\operatorname {M L P} \\left(\\hat {\\mathbf {Q}} _ {\\mathrm {t g t}} ^ {T}\\right). T \\in \\{t, t + 1, t + 2 \\} \\tag {11}\n\\]" + }, + { + "type": "image", + "bbox": [ + 0.516, + 0.089, + 0.905, + 0.235 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.513, + 0.237, + 0.905, + 0.293 + ], + "angle": 0, + "content": "Figure 5. Illustration of the Multimodal Prediction. The target queries of different timestamps respectively predict the probabilities and locations of three possible future situations, in which the one of the highest probability score is the final output." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.306, + 0.907, + 0.444 + ], + "angle": 0, + "content": "After the prediction, the target state with the highest probability of each timestamp constitutes the final target trajectory. For training, we first calculate GIoU [42] between the predicted target state of each modality \\(\\hat{\\mathbf{b}}_j^T\\in \\hat{\\mathbf{b}}^T (1\\leq j\\leq 3)\\) and the ground-truth \\(\\mathrm{b}^T\\). The prediction with the largest GIoU is assigned with the probability label of 1 (the other two ones are 0), and L1 loss \\(\\mathcal{L}_{\\mathrm{L1}}\\) and GIoU loss \\(\\mathcal{L}_{\\mathrm{GIoU}}\\) are applied to the output target states. The probabilities are supervised by the 0-1 labels with a cross-entropy loss \\(\\mathcal{L}_{\\mathrm{ce}}\\)." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.451, + 0.727, + 0.466 + ], + "angle": 0, + "content": "3.3. Training and Inference" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.473, + 0.906, + 0.638 + ], + "angle": 0, + "content": "Training. Our DreamTrack performs video sequence-level training and facilitates environmental understanding in an end-to-end manner. We construct video clips in which each contains 4 frames for training (i.e., one frame as template and three ones as search sequence), following the protocol of coordinate mapping described in Sec. 3.1. For every new input of the video clip, the hidden state in GRU (see Sec. 3.2) will be initialized, which avoids being distracted by the last tracking video. Considering the consistent time step between training and inference, we sample consecutive three frames for the search sequence." + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.639, + 0.905, + 0.7 + ], + "angle": 0, + "content": "The overall optimization objective for DreamTrack includes dreaming loss \\(\\mathcal{L}_{\\mathrm{drm}}\\), reconstruction loss \\(\\mathcal{L}_{\\mathrm{rect}}\\) from the Future Dreaming module, and probability loss \\(\\mathcal{L}_{\\mathrm{ce}}\\), L1 loss \\(\\mathcal{L}_{\\mathrm{L1}}\\) and GIoU loss \\(\\mathcal{L}_{\\mathrm{GIoU}}\\) from the Multimodal Prediction," + }, + { + "type": "equation", + "bbox": [ + 0.525, + 0.707, + 0.905, + 0.723 + ], + "angle": 0, + "content": "\\[\n\\mathcal {L} = \\lambda_ {1} \\mathcal {L} _ {\\mathrm {d r m}} + \\lambda_ {2} \\mathcal {L} _ {\\mathrm {r c t}} + \\lambda_ {3} \\mathcal {L} _ {\\mathrm {c e}} + \\lambda_ {4} \\mathcal {L} _ {\\mathrm {L 1}} + \\lambda_ {5} \\mathcal {L} _ {\\mathrm {G l o U}}, \\tag {12}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.728, + 0.864, + 0.743 + ], + "angle": 0, + "content": "where \\(\\lambda\\) values are the balancing weight coefficients." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.75, + 0.907, + 0.901 + ], + "angle": 0, + "content": "Inference. During inference, we initialize the hidden state of GRU for each new video sequence as previously described. Notably, it's no need to sample the video clip for inference. With the defined target queries that automatically cache the historical dynamics, we perform frame-level input for DreamTrack, which is consistent with previous nontemporal works [9, 51] and friendly for the deployment. Besides, the distribution mapping and image reconstruction in Fig. 3 are not required in the inference phase, which could be extended to other tasks, e.g., video generation." + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.516, + 0.957 + ], + "angle": 0, + "content": "7205" + } + ], + [ + { + "type": "table", + "bbox": [ + 0.113, + 0.089, + 0.885, + 0.42 + ], + "angle": 0, + "content": "
MethodsLaSOT [15]LaSOText [16]GOT-10k* [27]TrackingNet [39]TNL2K [46]
SUC(%)PNorm(%)P(%)SUC(%)PNorm(%)P(%)AO(%)SR0.5(%)SR0.75(%)SUC(%)PNorm(%)P(%)SUC(%)P(%)
SiamFC255[2]33.642.033.923.031.126.934.835.39.857.166.353.329.528.6
ECO224[10]32.433.830.122.025.224.031.630.911.155.461.849.232.631.7
SiamRPN++255[32]49.656.949.134.041.639.651.761.632.573.380.069.441.341.2
DiMP288[3]56.965.056.739.247.645.161.171.749.274.080.168.744.743.4
SiamRCNN255[44]64.872.2----64.972.859.781.285.480.052.352.8
Ocean255[53]56.065.156.6---61.172.147.3---38.438.7
AutoMatch255[54]58.3-59.937.6-43.065.276.654.376.0-72.647.243.5
TrDiMP352[45]63.9-61.4---67.177.758.378.483.373.1--
TransT256[6]64.973.869.044.852.352.567.176.860.981.486.780.350.751.7
STARK320[50]67.177.0-47.856.255.168.878.164.182.086.9---
SwinTrack-B384[33]71.3-76.549.1-55.672.480.567.884.0-82.855.957.1
MixFormer-L320[9]70.179.976.3---72.682.268.883.988.983.1--
OSTrack384[51]71.181.177.650.561.357.673.783.270.883.988.583.255.956.7
GRM256[19]69.979.375.8---73.482.970.484.088.783.3--
MixViT288[9]69.679.975.9---72.582.469.983.588.382.0--
ROMTrack384[4]71.481.478.251.362.458.674.284.372.484.189.083.7--
MixViT-L384[9]72.482.280.1---75.785.375.185.490.285.7--
SeqTrack-L384[7]72.581.579.350.761.657.574.881.972.285.589.885.857.8-
ARTrack256[47]70.479.576.646.456.552.373.582.270.984.288.783.557.5-
ARTrack-L384[47]73.182.280.352.862.959.778.587.477.885.689.686.060.3-
ARTrackV2256[1]71.680.277.250.861.957.775.985.472.784.989.384.559.2-
ARTrackV2-L384[1]73.682.881.153.463.760.279.587.879.686.190.486.261.6-
LoRAT224[34]71.780.977.350.361.657.172.181.870.783.587.982.158.861.3
LoRAT-L378[34]75.184.182.056.669.065.177.586.278.185.689.785.462.367.0
DreamTrack25673.883.480.653.164.159.877.587.174.285.890.085.360.463.2
DreamTrack38475.084.281.754.565.361.178.387.976.686.590.685.961.264.8
DreamTrack-L38476.685.683.157.169.465.479.988.580.387.991.386.862.967.4
" + }, + { + "type": "table_caption", + "bbox": [ + 0.09, + 0.424, + 0.907, + 0.455 + ], + "angle": 0, + "content": "Table 1. State-of-the-art comparison on LaSOT [15], LaSOText [16], GOT-10k [27], TrackingNet [39] and TNL2K [46]. * indicates only trained on GOT-10k. The number in the subscript denotes the search region resolution." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.465, + 0.218, + 0.482 + ], + "angle": 0, + "content": "4. Experiment" + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.49, + 0.308, + 0.506 + ], + "angle": 0, + "content": "4.1. Implementation Details" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.512, + 0.483, + 0.542 + ], + "angle": 0, + "content": "Model Variants. We develop three variants of DreamTrack with different configurations as follows:" + }, + { + "type": "text", + "bbox": [ + 0.091, + 0.542, + 0.483, + 0.571 + ], + "angle": 0, + "content": "- DreamTrack256. Backbone: ViT-Base; Template size: \\([128 \\times 128]\\); Search region size: \\([256 \\times 256]\\);" + }, + { + "type": "text", + "bbox": [ + 0.092, + 0.572, + 0.483, + 0.602 + ], + "angle": 0, + "content": "- DreamTrack384. Backbone: ViT-Base; Template size: \\([192 \\times 192]\\); Search region size: \\([384 \\times 384]\\);" + }, + { + "type": "text", + "bbox": [ + 0.091, + 0.602, + 0.483, + 0.632 + ], + "angle": 0, + "content": "- DreamTrack-L384. Backbone: ViT-Large; Template size: \\([192 \\times 192]\\); Search region size: \\([384 \\times 384]\\)." + }, + { + "type": "list", + "bbox": [ + 0.091, + 0.542, + 0.483, + 0.632 + ], + "angle": 0, + "content": null + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.632, + 0.483, + 0.709 + ], + "angle": 0, + "content": "The ViT backbone [13] is regarded as the encoder of our DreamTrack (see Fig. 2), which is divided into four stages for feature modeling with the proposed Temporal Self-Attention. The four stages contain 2,2,6,2 encoder layers for ViT-Base, and 4,4,12,4 ones for ViT-Large, respectively." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.72, + 0.483, + 0.902 + ], + "angle": 0, + "content": "Training Strategy. The training setting and evaluation protocol follow previous Transformer-based trackers (e.g., OSTrack [51]). The training datasets include COCO [35], LaSOT [15], TrackingNet [39] and GOT-10k [27] (with 1k sequences removed from GOT-10k train split, as per [50]). To ensure a fair evaluation of the GOT-10k test set, our models learn from the entire GOT-10k training split following its one-shot protocol. Our DreamTrack performs training with video clips, as described in Sec. 3.3. To align with the data form in the image dataset COCO that performs frame-level sampling, we repeat the sampled search frame with random shifts to constitute a pseudo video clip." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.467, + 0.907, + 0.573 + ], + "angle": 0, + "content": "We train our model using the AdamW [37] optimizer with a weight decay of \\(1 \\times 10^{-4}\\). The learning rate for the encoder is set to \\(4 \\times 10^{-5}\\), while other parameters use a learning rate of \\(4 \\times 10^{-4}\\). The training process is conducted on 8 NVIDIA RTX3090 GPUs, comprising 300 epochs with 60k video sequences per epoch. The learning rate decreases by a factor of 10 after 240 epochs." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.579, + 0.655, + 0.594 + ], + "angle": 0, + "content": "4.2. Main Results" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.601, + 0.906, + 0.662 + ], + "angle": 0, + "content": "We evaluate our models on eight prevailing benchmark datasets, including LaSOT [15], LaSOText [16], GOT-10k [27], TrackingNet [39], TNL2K [46], OTB100 [49], NFS [29] and UAV123 [38]." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.668, + 0.907, + 0.79 + ], + "angle": 0, + "content": "LaSOT [15]. LaSOT is a benchmark designed for long-term tracking, comprising 280 videos in its test set. Tab. 1 demonstrates that our DreamTrack\\(_{256}\\) already outperforms ARTrack-\\(L_{384}\\), which has a larger backbone. Furthermore, our DreamTrack-\\(L_{384}\\) achieved the best performance with the SUC score of \\(76.6\\%\\), without the use of any online template update strategy or post-processing. This proves the superiority of our temporal design with future dreaming." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.795, + 0.909, + 0.903 + ], + "angle": 0, + "content": "LaSOText [16]. LaSOText serves as an extension of LaSOT, which includes additional 150 video sequences. As shown in Tab. 1, our DreamTrack\\(_{256}\\) achieves superior performance compared with other trackers with the same backbone, e.g., ARTrackV2\\(_{256}\\) and LoRAT\\(_{224}\\). With the larger backbone, the performance of our DreamTrack-L\\(_{384}\\) further improves with the best SUC score of \\(57.1\\%\\)." + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.517, + 0.957 + ], + "angle": 0, + "content": "7206" + } + ], + [ + { + "type": "table", + "bbox": [ + 0.134, + 0.089, + 0.44, + 0.269 + ], + "angle": 0, + "content": "
MethodOTB100 [49]NFS [29]UAV123 [38]
DreamTrack-L38472.467.473.1
DreamTrack38472.066.172.5
DreamTrack25671.865.472.1
LoRAT-L378 [34]72.066.772.5
LoRAT224 [34]71.464.072.6
ARTrackV2-L384 [1]-67.669.9
SeqTrack-L384 [7]-66.268.5
OSTrack384 [51]-66.570.7
TransT256 [6]69.565.769.1
DiMP288 [3]68.461.864.3
Ocean255 [53]68.449.457.4
ATOM288 [11]66.358.363.2
ECO224 [10]66.652.253.5
SiamFC255 [2]58.337.746.8
" + }, + { + "type": "table_caption", + "bbox": [ + 0.09, + 0.274, + 0.483, + 0.304 + ], + "angle": 0, + "content": "Table 2. Performance comparison on OTB100 [49], NFS [29] and UAV123 [38] in terms of the SUC score \\((\\%)\\)." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.312, + 0.484, + 0.419 + ], + "angle": 0, + "content": "GOT-10k [27]. GOT-10k provides 180 videos for testing with a strict protocol that only allows trackers to be trained exclusively on its training split. We adhere to this protocol and report the results in Tab. 1. Our DreamTrack-\\(\\mathrm{L}_{384}\\) outperforms recent SOTA trackers across all metrics. Notably, our DreamTrack\\(_{256}\\) surpasses other trackers with the same backbone, proving its effectiveness." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.426, + 0.484, + 0.533 + ], + "angle": 0, + "content": "TrackingNet [39]. TrackingNet is an extensive tracking dataset that covers a wide range of real-world scenarios. We assess the performance of DreamTrack on its test set of 511 videos, as illustrated in Tab. 1. It shows that not only does our DreamTrack\\(_{384}\\) outperform all other trackers in SUC, but our DreamTrack-\\(L_{384}\\) also establishes a new state-of-the-art in three metrics on this large-scale benchmark." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.538, + 0.484, + 0.645 + ], + "angle": 0, + "content": "TNL2K [46]. TNL2K is a recently proposed large-scale benchmark for tracking by natural language and bounding box initialization. We evaluate our tracker on its 700 testing videos and the results are listed in Tab. 1. Our DreamTrack-L384 achieves the best performance with the SUC and precision of \\(62.9\\%\\) and \\(67.4\\%\\) respectively, showing the effectiveness of our temporal learning." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.651, + 0.483, + 0.744 + ], + "angle": 0, + "content": "OTB100 [49], NFS [29] and UAV123 [38]. We also evaluate our tracker on three additional benchmarks, i.e., OTB100, NFS and UAV123 including 100, 100 and 123 videos respectively. Tab. 2 shows that DreamTrack-\\(\\mathrm{L}_{384}\\) achieves the best performance on all three benchmarks, demonstrating the strong generalizability of our framework." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.75, + 0.483, + 0.903 + ], + "angle": 0, + "content": "Efficiency Comparison. Tab. 3 compares the efficiency of our DreamTrack against other Transformer-based trackers in terms of inference speed, floating point operations, and number of parameters. Benefiting from the efficient one-stream architecture and sparse multimodal prediction, our DreamTrack shows impressive inference speed with small computation costs. Our most compact variant, DreamTrack\\(_{256}\\), achieves the state-of-the-art performance while operating at 139 frames per second (FPS), proving the effectiveness and efficiency of our method." + }, + { + "type": "table", + "bbox": [ + 0.556, + 0.089, + 0.868, + 0.239 + ], + "angle": 0, + "content": "
TrackerSpeed (FPS)FLOPs (G)Params (M)
SwinTrack-B384 [33]4569.791
OSTrack256 [51]105.421.594
OSTrack384 [51]58.148.394
SeqTrack-B256 [7]406689
SeqTrack-L384 [7]5524309
ARTrack256 [47]26395.6117
ROMTrack256 [4]6234.592
ROMTrack384 [4]2877.792
LoRAT224 [34]2093099
LoRAT-L378 [34]63325336
DreamTrack25613937.687
DreamTrack-L38432359.4317
" + }, + { + "type": "table_caption", + "bbox": [ + 0.513, + 0.244, + 0.907, + 0.272 + ], + "angle": 0, + "content": "Table 3. Efficiency comparison. The inference speed is measured on a single NVIDIA RTX2080Ti GPU." + }, + { + "type": "table", + "bbox": [ + 0.518, + 0.282, + 0.905, + 0.368 + ], + "angle": 0, + "content": "
#Model VariantsLaSOT [15]TNL2K [46]FPS
SUC(%)PNorm(%)P(%)SUC(%)P(%)
①Baseline69.379.175.354.756.2162
②+ Future Dreaming Module71.781.377.756.959.5153
③+ Temporal Self-Attention72.682.078.559.160.4145
④+ Temporal Cross-Attention73.182.479.059.561.1143
⑤+ Multimodal Prediction73.883.480.660.463.2139
" + }, + { + "type": "table_caption", + "bbox": [ + 0.53, + 0.373, + 0.888, + 0.387 + ], + "angle": 0, + "content": "Table 4. Ablation on each proposed module in DreamTrack." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.404, + 0.753, + 0.42 + ], + "angle": 0, + "content": "4.3. Component-wise Ablation" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.426, + 0.907, + 0.7 + ], + "angle": 0, + "content": "We analyze the influence of each component in our DreamTrack\\(_{256}\\) to show the effectiveness and rationality of the proposed modules, as shown in Tab. 4. When removing all designed modules and directly predicting the target location of the current frame based on the target query, the performance is still comparable with the fastest inference speed of 162 FPS. By applying the key Future Dreaming module on the baseline version, it obtains impressive SUC gains of \\(2.4\\%\\) on LaSOT (2 v.s. 1). This verifies that deeper environmental understanding from past experience benefits the generalizability in novel situations. The Temporal Self/Cross-Attention further improves the performance for \\(0.9\\% / 0.5\\%\\) SUC on LaSOT (4 v.s. 3 v.s. 2), proving the effectiveness of target-aware temporal refinement. The final Multimodal Prediction brings \\(0.7\\% / 0.9\\%\\) SUC gains on LaSOT and TNL2K, respectively. It indicates that modeling the uncertainty of different future scenarios helps to forecast the target's motion and perform robust tracking." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.712, + 0.682, + 0.728 + ], + "angle": 0, + "content": "4.4. Further Analysis" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.735, + 0.907, + 0.903 + ], + "angle": 0, + "content": "Self-supervision in Future Dreaming. As mentioned in Sec. 3.2, the capability of predicting the environmental future states relies on the self-supervision in the proposed Future Dreaming module. We explore the influence by ablating the loss functions based on \\(\\mathrm{DreamTrack}_{256}\\). As shown in Tab. 5, removing both the dreaming and reconstruction losses makes the dreamed future unreliable to reflect the real scenario in the next two frames, degrading the quality of learned temporal dynamics (①). Applying each of the two loss functions brings performance gains of \\(4.4\\% / 4.0\\%\\) SUC on LaSOT respectively (②,③ v.s. ①), proving the ef" + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.516, + 0.957 + ], + "angle": 0, + "content": "7207" + } + ], + [ + { + "type": "table", + "bbox": [ + 0.113, + 0.089, + 0.46, + 0.162 + ], + "angle": 0, + "content": "
#Dreaming LossReconstruction LossLaSOT [15]TNL2K [46]
SUC(%)PNorm(%)P(%)SUC(%)P(%)
①--68.377.473.853.753.2
②✓-72.782.679.859.161.7
③-✓72.382.179.258.560.3
④✓✓73.883.480.660.463.2
" + }, + { + "type": "table_caption", + "bbox": [ + 0.097, + 0.167, + 0.477, + 0.181 + ], + "angle": 0, + "content": "Table 5. Ablation on loss functions in Future Dreaming module." + }, + { + "type": "table", + "bbox": [ + 0.113, + 0.19, + 0.46, + 0.258 + ], + "angle": 0, + "content": "
#Dreaming Step For EvaluationLaSOT [15]LaSOText [16]TNL2K [46]
SUC(%)P(%)SUC(%)P(%)SUC(%)P(%)
①073.880.653.159.860.463.2
②171.377.551.458.258.661.6
③268.574.748.955.055.357.8
" + }, + { + "type": "text", + "bbox": [ + 0.103, + 0.263, + 0.471, + 0.277 + ], + "angle": 0, + "content": "Table 6. Performance comparison of different dreaming steps." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.287, + 0.484, + 0.334 + ], + "angle": 0, + "content": "fectiveness. With the full supervision, the performance of DreamTrack\\(_{256}\\) further improves with the SUC of \\(73.8\\%\\) on LaSOT (4), which indicates the cooperation of our design." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.345, + 0.484, + 0.603 + ], + "angle": 0, + "content": "Performance of Future Prediction. With the defined target queries of different timestamps, our DreamTrack is capable of predicting the target trajectory including the current and next two frames. We then evaluate the performance of the future predictions by regarding them as the results, e.g., the future prediction for frame 10 from current frame 8 is used for the metric computation. As shown in Tab. 6, the performance of future prediction is inferior to current prediction (2,3 v.s. 1), and the tracking metrics degrade more with the prediction of further future. The underlying reason is that the future observation is unknowable for current prediction, in which the high degree of uncertainty raises challenges to future forecasting. In contrast, the previous trackers only predict the present target location with current observation. Despite these, the comparable performance demonstrates the potential of our design in the motion prediction of the target and benefits visual tracking." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.614, + 0.483, + 0.81 + ], + "angle": 0, + "content": "Activation Analysis and Result Visualization. Future dreaming endows our DreamTrack with superior generalization capability in novel situations. For deeper insights into the mechanism, we first visualize the activation maps \\( w \\) and \\( w / o \\). The Future Dreaming module. As shown in Fig. 6, the attention areas without the future dreaming are easily distracted by the target's appearance deformation and background clutters, finally losing the target. This proves our claim that insufficient environmental understanding hardly handles the challenges in long-term tracking. In contrast, the version with future dreaming successfully locates the target area with the learned dynamics from past experience, showing the effectiveness of our design." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.811, + 0.484, + 0.901 + ], + "angle": 0, + "content": "We also visualize the results of different trackers, as well as the dreamed future frames. As shown in Fig. 7, the proposed DreamTrack\\(_{256}\\) delivers more robust tracking under deformation (the first row), occlusion and similar interferences (the last two rows). It demonstrates the superiority of our temporal learning in benefiting from both future and" + }, + { + "type": "image", + "bbox": [ + 0.516, + 0.09, + 0.905, + 0.285 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.513, + 0.29, + 0.907, + 0.332 + ], + "angle": 0, + "content": "Figure 6. Attention visualization of our DreamTrack \\( w/ \\) and \\( w/o \\). the Future Dreaming module. The predictions and ground-truth boxes are marked in red and green, respectively." + }, + { + "type": "image", + "bbox": [ + 0.517, + 0.34, + 0.905, + 0.504 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.513, + 0.508, + 0.907, + 0.537 + ], + "angle": 0, + "content": "Figure 7. Results visualization of different trackers and dreamed future frames based on learned environmental dynamics." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.545, + 0.907, + 0.62 + ], + "angle": 0, + "content": "historical information. Notably, the dreamed future frames correctly indicate the evolution of the complex environment, and the predicted trajectory successfully locates the tracking target. This indicates the potential of our method to be developed into video generation and annotation tasks." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.62, + 0.906, + 0.681 + ], + "angle": 0, + "content": "Due to limited space, we present more analyses in the supplementary material, including 1) more challenging benchmarks, 2) generalization in different paradigms, 3) position of future dreaming, 4) more visualizations, etc." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.695, + 0.634, + 0.71 + ], + "angle": 0, + "content": "5. Conclusion" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.72, + 0.907, + 0.901 + ], + "angle": 0, + "content": "We propose DreamTrack, a novel temporal framework with both historical and future information for visual object tracking. We reformulate the temporal learning in VOT as a History-to-Future process, which dreams the future states of the tracking scenario based on learned temporal dynamics from past experience, which empowers the tracker with the generalization in novel situations of new frames. The uncertainty of the future variation is modeled by forecasting the target trajectories of different possible future scenarios with multimodal prediction. Extensive experiments prove our tracker outperforms other mainstream trackers and achieves SOTA performance on prevailing benchmark datasets." + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.516, + 0.957 + ], + "angle": 0, + "content": "7208" + } + ], + [ + { + "type": "title", + "bbox": [ + 0.092, + 0.091, + 0.251, + 0.108 + ], + "angle": 0, + "content": "Acknowledgments" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.115, + 0.484, + 0.222 + ], + "angle": 0, + "content": "This work was supported by the National Natural Science Foundation of China under Grant (62436001,62176020); the Joint Foundation of the Ministry of Education for Innovation Team (8091B042235); the Fundamental Research Funds for the Central Universities (2019JBZ110); and the State Key Laboratory of Rail Traffic Control and Safety (RCS2023K006), Beijing Jiaotong University." + }, + { + "type": "title", + "bbox": [ + 0.093, + 0.235, + 0.188, + 0.251 + ], + "angle": 0, + "content": "References" + }, + { + "type": "ref_text", + "bbox": [ + 0.101, + 0.26, + 0.484, + 0.302 + ], + "angle": 0, + "content": "[1] Yifan Bai, Zeyang Zhao, Yihong Gong, and Xing Wei. Artrackv2: Prompting autoregressive tracker where to look and how to describe. In CVPR, 2024. 2, 6, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.1, + 0.304, + 0.484, + 0.345 + ], + "angle": 0, + "content": "[2] Luca Bertinetto, Jack Valmadre, João F Henriques, Andrea Vedaldi, and Philip H S Torr. Fully-convolutional siamese networks for object tracking. In ECCVW, 2016. 1, 2, 6, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.347, + 0.483, + 0.387 + ], + "angle": 0, + "content": "[3] Goutam Bhat, Martin Danelljan, Luc Van Gool, and Radu Timofte. Learning discriminative model prediction for tracking. In ICCV, 2019. 6, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.39, + 0.482, + 0.417 + ], + "angle": 0, + "content": "[4] Yidong Cai, Jie Liu, Jie Tang, and Gangshan Wu. Robust object modeling for visual tracking. In ICCV, 2023. 6, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.419, + 0.483, + 0.46 + ], + "angle": 0, + "content": "[5] Fei Chen, Xiaodong Wang, Yunxiang Zhao, Shaohe Lv, and Xin Niu. Visual object tracking: A survey. Computer Vision and Image Understanding, 2022. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.462, + 0.483, + 0.502 + ], + "angle": 0, + "content": "[6] Xin Chen, Bin Yan, Jiawen Zhu, Dong Wang, Xiaoyun Yang, and Huchuan Lu. Transformer tracking. In CVPR, 2021. 1, 2, 6, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.505, + 0.483, + 0.545 + ], + "angle": 0, + "content": "[7] Xin Chen, Houwen Peng, Dong Wang, Hutchuan Lu, and Han Hu. Seqtrack: Sequence to sequence learning for visual object tracking. In CVPR, 2023. 1, 2, 3, 4, 6, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.548, + 0.483, + 0.601 + ], + "angle": 0, + "content": "[8] Junyoung Chung, Caglar Gulcehre, KyungHyun Cho, and Yoshua Bengio. Empirical evaluation of gated recurrent neural networks on sequence modeling. arXiv preprint arXiv:1412.3555, 2014. 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.604, + 0.482, + 0.644 + ], + "angle": 0, + "content": "[9] Yutao Cui, Jiang Cheng, Limin Wang, and Gangshan Wu. Mixformer: End-to-end tracking with iterative mixed attention. In CVPR, 2022. 1, 2, 5, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.647, + 0.483, + 0.687 + ], + "angle": 0, + "content": "[10] Martin Danelljan, Goutam Bhat, Fahad Shahbaz Khan, and Michael Felsberg. Eco: Efficient convolution operators for tracking. In CVPR, 2017. 1, 6, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.69, + 0.482, + 0.731 + ], + "angle": 0, + "content": "[11] Martin Danelljan, Goutam Bhat, Fahad Shahbaz Khan, and Michael Felsberg. ATOM: Accurate tracking by overlap maximization. In CVPR, 2019. 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.733, + 0.482, + 0.76 + ], + "angle": 0, + "content": "[12] Martin Danelljan, Luc Van Gool, and Radu Timofte. Probabilistic regression for visual tracking. In CVPR, 2020. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.762, + 0.482, + 0.842 + ], + "angle": 0, + "content": "[13] Alexey Dosovitskiy, Lucas Beyer, Alexander Kolesnikov, Dirk Weissenborn, Xiaohua Zhai, Thomas Unterthiner, Mostafa Dehghani, Matthias Minderer, Georg Heigold, Sylvain Gelly, et al. An image is worth 16x16 words: Transformers for image recognition at scale. In ICLR, 2021. 4, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.846, + 0.482, + 0.901 + ], + "angle": 0, + "content": "[14] Frederik Ebert, Chelsea Finn, Sudeep Dasari, Annie Xie, Alex Lee, and Sergey Levine. Visual foresight: Model-based deep reinforcement learning for vision-based robotic control. arXiv preprint arXiv:1812.00568, 2018. 2" + }, + { + "type": "list", + "bbox": [ + 0.094, + 0.26, + 0.484, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.093, + 0.905, + 0.147 + ], + "angle": 0, + "content": "[15] Heng Fan, Liting Lin, Fan Yang, Peng Chu, Ge Deng, Sijia Yu, Hexin Bai, Yong Xu, Chunyuan Liao, and Haibin Ling. LaSOT: A high-quality benchmark for large-scale single object tracking. In CVPR, 2019. 2, 6, 7, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.149, + 0.905, + 0.203 + ], + "angle": 0, + "content": "[16] Heng Fan, Hexin Bai, Liting Lin, Fan Yang, Peng Chu, Ge Deng, Sijia Yu, Mingzhen Huang, Juehuan Liu, Yong Xu, et al. Lasot: A high-quality large-scale single object tracking benchmark. IJCV, 2021. 6, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.205, + 0.904, + 0.244 + ], + "angle": 0, + "content": "[17] Liangji Fang, Qinhong Jiang, Jianping Shi, and Bolei Zhou. Tpnet: Trajectory proposal network for motion prediction. In CVPR, 2020. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.247, + 0.905, + 0.273 + ], + "angle": 0, + "content": "[18] Chelsea Finn and Sergey Levine. Deep visual foresight for planning robot motion. ICRA, 2016. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.275, + 0.904, + 0.314 + ], + "angle": 0, + "content": "[19] Shenyuan Gao, Chunluan Zhou, and Jun Zhang. Generalized relation modeling for transformer tracking. In CVPR, 2023. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.317, + 0.905, + 0.369 + ], + "angle": 0, + "content": "[20] Dongyan Guo, Jun Wang, Ying Cui, Zhenhua Wang, and Shengyong Chen. SiamCAR: Siamese fully convolutional classification and regression for visual tracking. In CVPR, 2020. 1, 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.372, + 0.904, + 0.398 + ], + "angle": 0, + "content": "[21] David Ha and Jürgen Schmidhuber. Recurrent world models facilitate policy evolution. NeurIPS, 2018. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.401, + 0.904, + 0.44 + ], + "angle": 0, + "content": "[22] Danijar Hafner, Timothy Lillicrap, Jimmy Ba, and Mohammad Norouzi. Dream to control: Learning behaviors by latent imagination. *ICLR*, 2020. 2, 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.442, + 0.904, + 0.481 + ], + "angle": 0, + "content": "[23] Danijar Hafner, Timothy Lillicrap, Mohammad Norouzi, and Jimmy Ba. Mastering atari with discrete world models. *ICLR*, 2021. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.484, + 0.904, + 0.524 + ], + "angle": 0, + "content": "[24] Danijar Hafner, Jurgis Pasukonis, Jimmy Ba, and Timothy Lillicrap. Mastering diverse domains through world models. arXiv preprint arXiv:2301.04104, 2023. 2, 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.526, + 0.904, + 0.565 + ], + "angle": 0, + "content": "[25] João F Henriques, Rui Caseiro, Pedro Martins, and Jorge Batista. High-speed tracking with kernelized correlation filters. IEEE T-PAMI, 2014. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.568, + 0.905, + 0.621 + ], + "angle": 0, + "content": "[26] Anthony Hu, Gianluca Corrado, Nicolas Griffiths, Zachary Murez, Corina Gurau, Hudson Yeo, Alex Kendall, Roberto Cipolla, and Jamie Shotton. Model-based imitation learning for urban driving. NeurIPS, 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.624, + 0.905, + 0.663 + ], + "angle": 0, + "content": "[27] Lianghua Huang, Xin Zhao, and Kaiqi Huang. Got-10k: A large high-diversity benchmark for generic object tracking in the wild. IEEE T-PAMI, 2019. 6, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.665, + 0.905, + 0.704 + ], + "angle": 0, + "content": "[28] Licheng Jiao, Dan Wang, Yidong Bai, Puhua Chen, and Fang Liu. Deep learning in visual tracking: A review. IEEE TNNLS, 2021. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.707, + 0.905, + 0.747 + ], + "angle": 0, + "content": "[29] Hamed Kiani Galoogahi, Ashton Fagg, Chen Huang, Deva Ramanan, and Simon Lucey. Need for speed: A benchmark for higher frame rate object tracking. In ICCV, 2017. 6, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.749, + 0.904, + 0.775 + ], + "angle": 0, + "content": "[30] Yann LeCun. A path towards autonomous machine intelligence version 0.9.2, 2022-06-27. Open Review, 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.777, + 0.904, + 0.816 + ], + "angle": 0, + "content": "[31] Bo Li, Junjie Yan, Wei Wu, Zheng Zhu, and Xiaolin Hu. High performance visual tracking with siamese region proposal network. In CVPR, 2018. 1, 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.818, + 0.904, + 0.859 + ], + "angle": 0, + "content": "[32] Bo Li, Wei Wu, Qiang Wang, Fangyi Zhang, Junliang Xing, and Junjie Yan. Siamrpn++: Evolution of siamese visual tracking with very deep networks. In CVPR, 2019. 2, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.861, + 0.905, + 0.901 + ], + "angle": 0, + "content": "[33] Liting Lin, Heng Fan, Zhipeng Zhang, Yong Xu, and Haibin Ling. Swintrack: A simple and strong baseline for transformer tracking. NeurIPS, 2022. 2, 6, 7" + }, + { + "type": "list", + "bbox": [ + 0.517, + 0.093, + 0.905, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.516, + 0.957 + ], + "angle": 0, + "content": "7209" + } + ], + [ + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.093, + 0.482, + 0.135 + ], + "angle": 0, + "content": "[34] Liting Lin, Heng Fan, Zhipeng Zhang, Yaowei Wang, Yong Xu, and Haibin Ling. Tracking meets lora: Faster training, larger model, stronger performance. ECCV, 2024. 1, 2, 6, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.136, + 0.482, + 0.19 + ], + "angle": 0, + "content": "[35] Tsung-Yi Lin, Michael Maire, Serge J. Belongie, Lubomir D. Bourdev, Ross B. Girshick, James Hays, Pietro Perona, Deva Ramanan, Piotr Dollar, and C. Lawrence Zitnick. Microsoft COCO: Common objects in context. In ECCV, 2014. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.192, + 0.482, + 0.233 + ], + "angle": 0, + "content": "[36] Yicheng Liu, Jinghuai Zhang, Liangji Fang, Qinhong Jiang, and Bolei Zhou. Multimodal motion prediction with stacked transformers. In CVPR, 2021. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.235, + 0.482, + 0.262 + ], + "angle": 0, + "content": "[37] Ilya Loshchilov and Frank Hutter. Decoupled weight decay regularization. In ICLR, 2019. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.264, + 0.482, + 0.304 + ], + "angle": 0, + "content": "[38] Matthias Mueller, Neil Smith, and Bernard Ghanem. A benchmark and simulator for uav tracking. In ECCV, 2016. 6, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.307, + 0.482, + 0.36 + ], + "angle": 0, + "content": "[39] Matthias Muller, Adel Bibi, Silvio Giancola, Salman Alsubaihi, and Bernard Ghanem. TrackingNet: A large-scale dataset and benchmark for object tracking in the wild. In ECCV, 2018. 6, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.362, + 0.482, + 0.417 + ], + "angle": 0, + "content": "[40] Scott Reed, Konrad Zolna, Emilio Parisotto, Sergio Gomez Colmenarejo, Alexander Novikov, Gabriel Barth-maron, Mai Giménez, Yury Sulsky, Jackie Kay, Jost Tobias Springenberg, et al. A generalist agent. TMLR, 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.42, + 0.482, + 0.46 + ], + "angle": 0, + "content": "[41] Shaoqing Ren, Kaiming He, Ross Girshick, and Jian Sun. Faster R-CNN: Towards real-time object detection with region proposal networks. In NIPS, 2015. 2, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.462, + 0.482, + 0.517 + ], + "angle": 0, + "content": "[42] Hamid Rezatofighi, Nathan Tsoi, JunYoung Gwak, Amir Sadeghian, Ian Reid, and Silvio Savarese. Generalized intersection over union: A metric and a loss for bounding box regression. In CVPR, 2019. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.519, + 0.482, + 0.573 + ], + "angle": 0, + "content": "[43] Yuval Tassa, Yotam Doron, Alistair Muldal, Tom Erez, Yazhe Li, Diego de Las Casas, David Budden, Abbas Abdelmaleki, Josh Merel, Andrew Lefrancq, et al. Deepmind control suite. arXiv preprint arXiv:1801.00690, 2018. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.575, + 0.482, + 0.616 + ], + "angle": 0, + "content": "[44] Paul Voigtlaender, Jonathon Luiten, Philip HS Torr, and Bastian Leibe. Siam r-cnn: Visual tracking by re-detection. In CVPR, 2020. 2, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.618, + 0.482, + 0.657 + ], + "angle": 0, + "content": "[45] Ning Wang, Wengang Zhou, Jie Wang, and Houqiang Li. Transformer meets tracker: Exploiting temporal context for robust visual tracking. In CVPR, 2021. 1, 2, 3, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.66, + 0.482, + 0.715 + ], + "angle": 0, + "content": "[46] Xiao Wang, Xiujun Shu, Zhipeng Zhang, Bo Jiang, Yaowei Wang, Yonghong Tian, and Feng Wu. Towards more flexible and accurate object tracking with natural language: Algorithms and benchmark. In CVPR, 2021. 6, 7, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.717, + 0.482, + 0.757 + ], + "angle": 0, + "content": "[47] Xing Wei, Yifan Bai, Yongchao Zheng, Dahu Shi, and Yihong Gong. Autoregressive visual tracking. In CVPR, 2023. 1, 2, 3, 4, 6, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.76, + 0.482, + 0.801 + ], + "angle": 0, + "content": "[48] Philipp Wu, Alejandro Escontrela, Danijar Hafner, Pieter Abbeel, and Ken Goldberg. Daydreamer: World models for physical robot learning. In CoRL. PMLR, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.803, + 0.482, + 0.83 + ], + "angle": 0, + "content": "[49] Yi Wu, Jongwoo Lim, and Ming Hsuan Yang. Object tracking benchmark. IEEE T-PAMI, 2015. 6, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.832, + 0.482, + 0.872 + ], + "angle": 0, + "content": "[50] Bin Yan, Houwen Peng, Jianlong Fu, Dong Wang, and Huchuan Lu. Learning spatio-temporal transformer for visual tracking. In ICCV, 2021. 1, 2, 3, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.874, + 0.482, + 0.901 + ], + "angle": 0, + "content": "[51] Botao Ye, Hong Chang, Bingpeng Ma, Shiguang Shan, and Xilin Chen. Joint feature learning and relation modeling for" + }, + { + "type": "list", + "bbox": [ + 0.093, + 0.093, + 0.482, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.545, + 0.093, + 0.905, + 0.12 + ], + "angle": 0, + "content": "tracking: A one-stream framework. In ECCV, 2022. 1, 2, 4, 5, 6, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.122, + 0.905, + 0.162 + ], + "angle": 0, + "content": "[52] Zhipeng Zhang and Houwen Peng. Deeper and wider siamese networks for real-time visual tracking. In CVPR, 2019. 1, 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.165, + 0.905, + 0.205 + ], + "angle": 0, + "content": "[53] Zhipeng Zhang, Houwen Peng, Jianlong Fu, Bing Li, and Weiming Hu. Ocean: Object-aware anchor-free tracking. In ECCV, 2020. 2, 6, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.208, + 0.905, + 0.248 + ], + "angle": 0, + "content": "[54] Zhipeng Zhang, Yihao Liu, Xiao Wang, Bing Li, and Weiming Hu. Learn to match: Automatic matching network design for visual tracking. ICCV, 2021. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.251, + 0.905, + 0.289 + ], + "angle": 0, + "content": "[55] Xingyi Zhou, Dequan Wang, and Philipp Krahenbuhl. Objects as points. arXiv preprint arXiv:1904.07850, 2019. 2, 3" + }, + { + "type": "list", + "bbox": [ + 0.517, + 0.093, + 0.905, + 0.289 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.517, + 0.957 + ], + "angle": 0, + "content": "7210" + } + ] +] \ No newline at end of file diff --git a/2025/DreamTrack_ Dreaming the Future for Multimodal Visual Object Tracking/2e4cde12-368f-4feb-b1b3-a980bb1eb002_origin.pdf b/2025/DreamTrack_ Dreaming the Future for Multimodal Visual Object Tracking/2e4cde12-368f-4feb-b1b3-a980bb1eb002_origin.pdf new file mode 100644 index 0000000000000000000000000000000000000000..d504cce71b25018574e9d42cc107cffca203c868 --- /dev/null +++ b/2025/DreamTrack_ Dreaming the Future for Multimodal Visual Object Tracking/2e4cde12-368f-4feb-b1b3-a980bb1eb002_origin.pdf @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:1e9876fec9d9b42bf4b700cfe33d95f0840278d6d887a54ebe0a8690131cf40d +size 1748656 diff --git a/2025/DreamTrack_ Dreaming the Future for Multimodal Visual Object Tracking/full.md b/2025/DreamTrack_ Dreaming the Future for Multimodal Visual Object Tracking/full.md new file mode 100644 index 0000000000000000000000000000000000000000..7051374d9168884c8076e2c37162f2afe42967df --- /dev/null +++ b/2025/DreamTrack_ Dreaming the Future for Multimodal Visual Object Tracking/full.md @@ -0,0 +1,311 @@ +# DreamTrack: Dreaming the Future for Multimodal Visual Object Tracking + +Mingzhe Guo $^{1*}$ Weiping Tan $^{1}$ Wenyu Ran $^{1}$ Liping Jing $^{1\dagger}$ Zhipeng Zhang $^{2*}$ Beijing Jiaotong University $^{1}$ , Shanghai Jiaotong University $^{2}$ + +# Abstract + +Aiming to achieve class-agnostic perception in visual object tracking, current trackers commonly formulate tracking as a one-shot detection problem with the template-matching architecture. Despite the success, severe environmental variations in long-term tracking raise challenges to generalizing the tracker in novel situations. Temporal trackers try to fix it by preserving the time-validity of target information with historical predictions, e.g., updating the template. However, solely transmitting the previous observations instead of learning from them leads to an inferior capability of understanding the tracking scenario from past experience, which is critical for the generalization in new frames. To address this issue, we reformulate temporal learning in visual tracking as a History-to-Future process and propose a novel tracking framework DreamTrack. Our DreamTrack learns the temporal dynamics from past observations to dream the future variations of the environment, which boosts the generalization with the extended future information from history. Considering the uncertainty of future variation, multimodal prediction is designed to infer the target trajectory of each possible future situation. The experiments demonstrate that our DreamTrack achieves leading performance with real-time inference speed. In particular, DreamTrack obtains SUC scores of $76.6\% /87.9\%$ on LasOT/TrackingNet, surpassing all recent SOTA trackers. + +# 1. Introduction + +Visual Object Tracking (VOT) is one of the most fundamental computer vision problems with numerous applications [5, 28], which aims to estimate the position of an arbitrary target in a video sequence, given only its location in the initial frame. To achieve the class-agnostic perception, recent prevalent Siamese-based approaches [2, 6, 7, 9, 20, 31, 34, 47, 51, 52] typically formulate visual tracking as a one-shot detection problem. Specifically, these trackers perform general object tracking by matching with the initialized target template and then predicting the location in a single forward pass, as shown in Fig. 1(a). + +![](images/85445e7bf89ad36005c9fa61614c3301311996a6f9e69ff695864fce586ee3ae.jpg) +Figure 1. Comparison of different temporal architectures. (a) The non-temporal architecture that regards visual tracking as a one-shot detection problem. (b) The temporal architecture with only historical information that limits the temporal messages in the transmission of previous observations. (c) Our proposed temporal architecture with both historical and future information, which learns environmental dynamics from past experiences and dreams the future states for better generalization in subsequent frames. + +Despite the demonstrated success, it's challenging for the non-temporal architecture to ensure the generalization in new frames with severe environmental variations compared with the initial frame, e.g., deformation of the target appearance and background in long-term tracking. The outdated template hardly provides effective target-aware knowledge but distracts the tracking process. Then many works [7, 10, 12, 25] intuitively try to fix it by keeping the time-validity of target information, e.g., updating the template with predictions of high confidence scores in previous frames [45, 50] or passing historical trajectory as prior knowledge [7, 47], as shown in Fig. 1(b). However, these efforts focus on only transmitting previous observations instead of learning from them, leading to an inferior capability of understanding the tracking scenario from the past expe + +rience so that enables generalization to novel situations. + +To address this issue, we reformulate the temporal learning in VOT as a History-to-Future process, as shown in Fig. 1(c). The pipeline dreams the future states of the tracking scenario based on learned temporal dynamics from previous observations, which empowers the tracker with superior generalization ability in complex environments of new frames. Our intuition stems from the human recognition process [21, 30]. The brain abstracts the world from past senses and imagines future variations, which aims to achieve goals in complex scenarios even though they are never encountered before. Such a temporal modeling manner is intuitively superior to the one limited in historical information, which further develops the generalization in new frames with the future information from dreamed states. To this end, we construct a future dreaming module, originally proposed in the world model for reinforcement learning tasks [21-23], to endow the tracker with the capability of predicting future variations of the target from the learned historical dynamics (see Fig. 1(c)). In contrast to non-temporal [6, 9, 31] or historical information-only trackers [7, 47, 50], our dreaming module extends the temporal messages from the history to its future, in which the dreamed future frames are self-supervised with ground-truth future observations. The future information undoubtedly facilitates better generalization for appearance deformation of the target and background variance. + +More specifically, we propose a novel tracking framework dubbed DreamTrack, which has a dream-encoder-decoder architecture. The dreaming stage caches historical observations and predicts future states to represent the environmental variation in the next few frames. Then the features of the current and dreamed future frames are input to the encoder, which extracts target-aware features with the temporal messages from both history and future. Notably, after each feature extraction stage of the encoder, the temporal dynamics will be refined along the timeline, which aims to filter target-irrelevant interferences with the exposure of the target information. The final decoder predicts the trajectory of the target in the current and future frames. Considering the uncertainty of the future variation, we forecast the probabilities and target locations of different possible future scenarios, i.e., multimodal prediction. + +Extensive experiments on eight prevailing VOT benchmarks show that our proposed DreamTrack outperforms recent state-of-the-art trackers. For instance, under fair settings, DreamTrack- $\mathrm{L}_{384}$ obtains a $76.6\%$ SUC score on the LaSOT [15] dataset, surpassing ARTrackV2- $\mathrm{L}_{384}$ [1] by $3.0\%$ and LoRAT- $\mathrm{L}_{378}$ [34] by $1.5\%$ , respectively. In summary, the contributions of this work are as follows: + +- We reformulate the tracking problem as a History-to-Future process and are the first to develop the temporal messages into the future frames for better generalization + +of novel situations in visual object tracking. + +- We make the first attempt at predicting the future states of the target and performing multimodal prediction to include the uncertainty of different future scenarios, which empowers the tracker with superior generalization for appearance deformation and background variation. +- We prove the effectiveness of the proposed future dreaming to facilitate tracking with state-of-the-art performance on eight prevailing VOT benchmarks. + +# 2. Related Work + +Visual Object Tracking. Current prevailing trackers [2, 6, 9, 20, 31, 53] commonly employ a template-matching framework, which exploits the initialized template to match the target within the search region. Regarding tracking as a one-shot detection problem, these approaches develop a deep backbone to extract visual features [33, 51, 52], then predict the current target location following the detection paradigms [41, 55], such as object scale estimation [32, 44, 53] and center point localization [6, 20]. Observing the target information of the template may be weakened by the appearance deformation and background variation in long-term tracking, many works design modules to ensure the time-validity of target-relevant knowledge, e.g., updating the template with predictions of high confidence scores in previous frames [45, 50] or passing historical trajectory as prior knowledge [7, 47]. Despite the improved performance in long-term tracking, the limited understanding of the tracking scenario leads to suboptimal generalization in novel situations. Thus, we reformulate tracking as a History-to-Future process that learns the environmental dynamics from past experiences and dreams the future states for better adaptation and tracking capability in new frames. + +World Model. The world model [30] aims to learn a general representation of the world and predicts future world states resulting from a sequence of actions, which has been widely studied in the game [21-23, 40] and lab environments [14, 18, 48]. Dreamer [22] learns a latent dynamics model from past experience to predict future state values and actions in a latent space. It's capable of handling challenging visual control tasks in the DeepMind Control Suite [43]. DreamerV2 [23] further achieves human-level performance on Atari games with discrete modeling. DreamerV3 [24] develops a general algorithm and allows solving challenging control problems such as obtaining diamonds in Minecraft from scratch given sparse rewards. Recently, learning world models in driving scenes has gained attention. MILE [26] employs a model-based imitation learning method to jointly learn a dynamics model and driving behavior. In this work, we introduce the mechanism of the world model to visual object tracking, which boosts environmental understanding with the dreamed future states. + +![](images/50d33d9bbd2e086ede54ab20a13743a7d7c9c12e918522a61f5a5823f0620ca4.jpg) +Figure 2. The dream-encoder-decoder framework of our DreamTrack. Initially, the dreaming stage takes the input image tokens to predict future states of the next two frames based on the cached temporal dynamics in the Future Dreaming module. Then features of the current frame and dreamed future are fed into the encoder for further modeling, with a Temporal Self-Attention after each encoder stage for target-aware refinement. In the end, the decoder performs multimodal prediction for the target, i.e., the probabilities and trajectories in the current and next two frames of three possible future scenarios. The prediction with the highest probability is selected as the final output. + +# 3. Method + +# 3.1. Preliminaries + +Formulation of VOT. Given the template image $\mathbf{z}^0$ of a target initialized in the first frame, the objective of visual tracking is to localize the target $\mathrm{b}^t$ from the continuous input image $\mathbf{x}^t$ , where $\mathbf{b}$ denotes the target state, i.e., a bounding box. The mainstream tracking paradigm aims to learn a deep tracking model $f$ that performs localization with the output $\mathbf{y}^t$ following different detection paradigms, i.e., anchor-based [41] or anchor-free [55] architectures. The tracking process can be formulated as, + +$$ +y ^ {t} = f \left(z ^ {0}, x ^ {t}\right). \tag {1} +$$ + +To alleviate the influence of target's appearance deformation and background variation in long-term tracking, recent prevailing temporal trackers exploit previous outputs $\mathrm{y}^{0:t - 1}$ as additional prior knowledge, e.g., updating the template [45, 50] or passing historical target location [7, 47]. The temporal tracking process is formulated as, + +$$ +\mathrm {y} ^ {t} = f \left(\mathrm {z} ^ {0}, \mathrm {x} ^ {t}, \mathrm {y} ^ {0: t - 1}\right). \tag {2} +$$ + +In contrast, we formulate temporal visual tracking as a History-to-Future process, which predicts the future trajectory of the target (including the current frame and the next two ones) based on the learned environmental dynamics from historical observations (cached in the target queries $\mathbf{Q}_{\mathrm{tgt}}^{t:t + 2}$ of $f$ , described in Sec. 3.2). This modeling pipeline can be expressed as, + +$$ +y ^ {t: t + 2} = f \left[ Q _ {\mathrm {t g t}} ^ {t: t + 2} \right] \left(z ^ {0}, x ^ {t}\right). \tag {3} +$$ + +The parameterized modeling of environmental dynamics $(\mathbf{Q}_{\mathrm{tg}}^{t:t + 2})$ instead of explicit transmission of historical output $\mathrm{y}^{0:t - 1}$ , not only enables better generalization in novel situations of future frames, but also reduces the risk of error accumulation from past experiences. + +Trajectory Coordinate Mapping. To reduce computation costs, most methods perform tracking on the cropped search region instead of the full-resolution frame, which outputs target's coordinates in the current frame, relative to the search region. Thus, it's necessary to map the predicted target location of the relative coordinate system to the original frame after inference, which we denote as the global coordinate system. Correspondingly, the ground-truth boxes for training are also transformed from the global to relative coordinate system. For our DreamTrack that learns dynamics and predicts target trajectory over a video sequence, we additionally cache the ground-truth boxes of the preceding two frames in the global coordinate system and map them to the relative coordinate system of the current frame after the search region is cropped. Notably, the search region patches of the two future frames are cropped by referring to the ground-truth box of the current frame, which represent the target's motion and serve for the supervision of future image reconstruction (described in Sec. 3.2 and Fig. 3). + +# 3.2. Network Architecture + +The overall framework of DreamTrack is depicted in Fig. 2. Given the target sequences constructed from object trajectories, we use a novel dream-encoder-decoder structure for learning and inference. The target state is described by a trajectory including the current and next two frames. + +![](images/f5bbecd65d43e3839113a97abc270b539916e8ca1521b7ce2122cfb022edb8fb.jpg) +Figure 3. Illustration of the proposed Future Dreaming module. The target queries auto-regressively update the temporal dynamics with the patch embeddings of the current frame through GRU, which are then exploited to predict the future states. For training, the distributions of the predicted future states and decoded future images are self-supervised by the ground-truth future frames. + +Dream. We propose the Future Dreaming module to learn temporal dynamics of the tracking scenario. As shown in Fig. 3, the module evolves the dynamics with current observation and dreams the future states, which endows the tracker with the capability of future prediction by supervising with ground-truth future frames. Specifically, three trainable target queries $\mathbf{Q}_{\mathrm{tgt}}^{t - 1:t + 1} \in \mathbb{R}^{3 \times C}$ , which represent the transition dynamics of the current and next two frames, are initialized. With the template $z^0$ and search region $x^t$ of each new frame, the images are split and flattened to build current observation feature $\mathbf{F}^{\mathrm{t}} = \{\mathbf{F}_z^0 \in \mathbb{R}^{H_z W_z \times C}, \mathbf{F}_x^t \in \mathbb{R}^{H_x W_x \times C}\}$ with a linear projection. Then target queries auto-regressively update the cached temporal messages with $\mathbf{F}^{\mathrm{t}}$ through a gated recurrent unit (GRU) [8], + +$$ +\mathbf {Q} _ {\mathrm {t g t}} ^ {t: t + 2} = \operatorname {G R U} \left(\mathbf {Q} _ {\mathrm {t g t}} ^ {t - 1: t + 1}, \mathbf {F} ^ {t}\right). \tag {4} +$$ + +As the aim of our DreamTrack is to enhance the generalization in novel situations of new frames, it's a prerequisite to evolve the future variations of the tracking environment. Thus, the updated target queries $\mathbf{Q}_{\mathrm{tgt}}^{t:t + 2}$ are exploited to backward imagine the future observation feature $\hat{\mathbf{F}}^{t:t + 2} = \{\hat{\mathbf{F}}_{\mathrm{z}}^{t:t + 2}\in \mathbb{R}^{3\times H_{\mathrm{z}}W_{\mathrm{z}}\times C},\hat{\mathbf{F}}_{\mathrm{x}}^{t:t + 2}\in \mathbb{R}^{3\times H_{\mathrm{x}}W_{\mathrm{x}}\times C}\}$ based on $\mathbf{F}^{\mathrm{t}}$ with a cross-attention, + +$$ +\hat {\mathbf {F}} ^ {t: t + 2} = \operatorname {C r o s s A t t e n t i o n} (q = \mathbf {F} ^ {t}, k v = \mathbf {Q} _ {\mathrm {t g t}} ^ {t: t + 2}). \quad (5) +$$ + +Notably, the interaction between $\mathbf{F}^t$ and $\mathbf{Q}_{\mathrm{tgt}}^t$ to generate $\hat{\mathbf{F}}^t$ intends to enhance the feature quality of the current frame by introducing historical information, which originates from $\mathbf{Q}_{\mathrm{tgt}}^{t - 1}$ before the GRU update (Eq. 4). The output $\hat{\mathbf{F}}^{t:t + 2}$ of the module will be further modeled with the Temporal Self-Attention, which is detailed in the encoder. + +Following the loss design in world models [22-24], we respectively map the dreamed future observation fea + +ture $\hat{\mathbf{F}}_{\mathrm{x}}^{t:t + 2}$ and ground-truth observation feature $\mathbf{F}_{\mathrm{x}}^{t:t + 2}$ to distributions of $\{\hat{\mu}_{\mathrm{x}}^{t:t + 2},\hat{\sigma}_{\mathrm{x}}^{t:t + 2}\in \mathbb{R}^{3\times H_{\mathrm{x}}W_{\mathrm{x}}\times C^{\prime}}\}$ and $\{\mu_{\mathrm{x}}^{t:t + 2},\sigma_{\mathrm{x}}^{t:t + 2}\in \mathbb{R}^{3\times H_{\mathrm{x}}W_{\mathrm{x}}\times C^{\prime}}\}$ with MLP layers, and then minimize their KL divergence. Notably, we only consider supervising the dreaming process of the search region, as the input template always comes from the initial frame. The prior distribution from $\hat{\mathbf{F}}_{\mathrm{x}}^{t:t + 2}$ is regarded as a prediction of the future dynamics. In contrast, the posterior distribution from $\mathbf{F}_{\mathrm{x}}^{t:t + 2}$ represents the future dynamics from the ground-truth observation $\mathbf{x}^{t:t + 2}$ . The KL divergence measures the gap between the imagined future (prior) and the true future (posterior). We expect to enhance the capability of future prediction for long-term tracking, which is realized by optimizing the dreaming loss $\mathcal{L}_{\mathrm{drm}}$ + +$$ +\mathcal {L} _ {\mathrm {d r m}} = \mathrm {K L} \left(\left\{\hat {\mu} _ {\mathrm {x}} ^ {t: t + 2}, \hat {\sigma} _ {\mathrm {x}} ^ {t: t + 2} \right\} \mid \mid \left\{\mu_ {\mathrm {x}} ^ {t: t + 2}, \sigma_ {\mathrm {x}} ^ {t: t + 2} \right\}\right), \tag {6} +$$ + +where $\mathrm{KL}(\cdot ||\cdot)$ denotes the KL divergence. + +Besides the supervision from the distribution difference, we further improve the accuracy of future prediction by reconstructing the future search regions based on $\hat{\mathbf{F}}_{\mathrm{x}}^{t:t + 2}$ with a decoder of several CNN upsampling layers. The decoded future search regions $\hat{\mathbf{x}}^{t:t + 2}$ is supervised by the groundtruth search regions $\mathbf{x}^{t:t + 2}$ by optimizing the reconstruction loss $\mathcal{L}_{\mathrm{RCT}}$ + +$$ +\mathcal {L} _ {\mathrm {r c t}} = \operatorname {M S E} \left(\hat {\mathrm {x}} ^ {t: t + 2} | | \mathrm {x} ^ {t: t + 2}\right), \tag {7} +$$ + +where $\mathrm{MSE}(\cdot ||\cdot)$ is the mean squared error. + +Encoder. Following previous works [7, 47, 51], we adopt a plain vision transformer (ViT) [13] as the encoder and model the search and template images jointly. The encoder extracts the features of the search and template images and learns feature-level correspondence to filter target-relevant distractors. The input dreamed observation features $\{\hat{\mathbf{F}}_{\mathrm{z}}^{t:t + 2},\hat{\mathbf{F}}_{\mathrm{x}}^{t:t + 2}\}$ are concatenated along the spatial dimension and fed into the encoder for feature extraction, + +$$ +\begin{array}{l} \hat {\mathbf {F}} _ {\mathrm {z x}, i} ^ {t: t + 2} = \operatorname {C o n c a t} \left(\hat {\mathbf {F}} _ {\mathrm {z}} ^ {t: t + 2}, \hat {\mathbf {F}} _ {\mathrm {x}} ^ {t: t + 2}\right), i = 0 \\ \hat {\mathbf {F}} _ {\mathrm {z x}, i} ^ {t: t + 2} = \operatorname {B l o c k} _ {i} \left(\hat {\mathbf {F}} _ {\mathrm {z x}, i - 1} ^ {t: t + 2}\right), \quad 1 \leq i \leq l _ {\text {s p t}} \tag {8} \\ \end{array} +$$ + +where $\hat{\mathbf{F}}_{\mathrm{zx},i}^{t:t + 2}\in \mathbb{R}^{3\times (H_{\mathrm{z}}W_{\mathrm{z}} + H_{\mathrm{x}}W_{\mathrm{x}})\times C}$ , Block denotes the transformer block in the encoder for spatial modeling and $l_{\mathrm{spt}}$ is the layer number. + +Besides the spatial modeling between the template and search region in Block, it's also important to exploit the extracted target-aware information to refine the temporal dynamics along the timeline. Thus, after each feature extraction stage of the encoder, we arrange a Temporal Self-Attention module to filter target-irrelevant interferences with the exposure of the target information from different levels. As shown in Fig. 4, the dreamed observation features $\hat{\mathbf{F}}_{\mathrm{zx}}^{t:t + 2}$ are transposed and input to the attention for the interaction on the temporal dimension, + +$$ +\hat {\mathbf {F}} _ {\mathrm {z x}, i} ^ {t: t + 2} = \operatorname {T e m A t t n} _ {\text {s e l f}} ^ {i} \left(\hat {\mathbf {F}} _ {\mathrm {z x}, i - 1} ^ {t: t + 2}\right), \quad 1 \leq i \leq l _ {\text {t e m}} \tag {9} +$$ + +![](images/86554f4977ac27dede5172890e3b01c454a47ba1a6fee32f3591f2f46a97ddc2.jpg) +Figure 4. Illustration of the Temporal Self/Cross-Attention. For self-attention, patch embeddings of the template and search region are jointly modeled on the temporal dimension. In contrast, the cross-attention takes target queries as $\mathbf{Q}$ and patch embeddings of search region as $\mathbf{K}$ , $\mathbf{V}$ to extract localization-relevant information. + +where $\mathrm{TemAttn}_{\mathrm{self}}$ denotes the Temporal Self-Attention module and $l_{\mathrm{tem}}$ is the module number corresponding to the four stages of the encoder (see Fig. 2). + +Decoder. The final decoder is responsible for predicting the target trajectory in the current and next two frames based on the updated target queries $\mathbf{Q}_{\mathrm{tgt}}^{t:t + 2}$ and target-aware observation feature of the search region $\hat{\mathbf{F}}_{\mathrm{x}}^{t:t + 2}$ , which is split from the encoder output $\hat{\mathbf{F}}_{\mathrm{zx}}^{t:t + 2}$ . Notably, the dreamed feature of the template $\hat{\mathbf{F}}_{\mathrm{z}}^{t:t + 2}$ is not employed as the ultimate is to locate the target in the search region and the target information has been introduced into $\hat{\mathbf{F}}_{\mathrm{x}}^{t:t + 2}$ after encoder modeling. The target queries $\mathbf{Q}_{\mathrm{tgt}}^{t:t + 2}$ first extract localization-relevant messages from $\hat{\mathbf{F}}_{\mathrm{x}}^{t:t + 2}$ with the Temporal Cross-Attention module. As shown in Fig. 4, each target query interacts with the region feature of the same timestamp, + +$$ +\hat {\mathbf {Q}} _ {\mathrm {t g t}} ^ {t: t + 2} = \operatorname {T e m A t t n} _ {\text {c r o s s}} (q = \mathbf {Q} _ {\mathrm {t g t}} ^ {t: t + 2}, k v = \hat {\mathbf {F}} _ {\mathrm {x}} ^ {t: t + 2}), \tag {10} +$$ + +where $\mathrm{TemAttn}_{\mathrm{cross}}$ is the Temporal Cross-Attention module. The output target queries $\hat{\mathbf{Q}}_{\mathrm{tgt}}^{t:t + 2}$ are then used to predict the target trajectory of the current and next two frames. + +The key challenge to future prediction is the high degree of uncertainty, in large part due to not knowing the intents and latent characteristics of the tracking target. For example, the target commonly has a multimodal distribution of futures: it could turn, go straight, slow down, speed up, etc. The high degree of multimodality also exists in the motion prediction of autonomous driving [17, 36], which correspondingly designs a multimodal prediction to generate multiple plausible trajectories of the target vehicle. We draw inspiration from this solution and design the tailored Multimodal Prediction to our DreamTrack. As shown in Fig. 5, each target query of one timestamp forecasts the probabilities $\hat{\mathbf{p}}^T\in \mathbb{R}^{3\times 1}$ and target states $\hat{\mathbf{b}}^T\in \mathbb{R}^{3\times 4}$ of three possible future situations with a stack of MLP layers, + +$$ +\hat {\mathrm {p}} ^ {T}, \hat {\mathrm {b}} ^ {T} = \operatorname {M L P} \left(\hat {\mathbf {Q}} _ {\mathrm {t g t}} ^ {T}\right). T \in \{t, t + 1, t + 2 \} \tag {11} +$$ + +![](images/4360af44e38dd4565c9f5f0537d5afcc39942769629616cd044483b759b6afee.jpg) +Figure 5. Illustration of the Multimodal Prediction. The target queries of different timestamps respectively predict the probabilities and locations of three possible future situations, in which the one of the highest probability score is the final output. + +After the prediction, the target state with the highest probability of each timestamp constitutes the final target trajectory. For training, we first calculate GIoU [42] between the predicted target state of each modality $\hat{\mathbf{b}}_j^T\in \hat{\mathbf{b}}^T (1\leq j\leq 3)$ and the ground-truth $\mathrm{b}^T$ . The prediction with the largest GIoU is assigned with the probability label of 1 (the other two ones are 0), and L1 loss $\mathcal{L}_{\mathrm{L1}}$ and GIoU loss $\mathcal{L}_{\mathrm{GIoU}}$ are applied to the output target states. The probabilities are supervised by the 0-1 labels with a cross-entropy loss $\mathcal{L}_{\mathrm{ce}}$ . + +# 3.3. Training and Inference + +Training. Our DreamTrack performs video sequence-level training and facilitates environmental understanding in an end-to-end manner. We construct video clips in which each contains 4 frames for training (i.e., one frame as template and three ones as search sequence), following the protocol of coordinate mapping described in Sec. 3.1. For every new input of the video clip, the hidden state in GRU (see Sec. 3.2) will be initialized, which avoids being distracted by the last tracking video. Considering the consistent time step between training and inference, we sample consecutive three frames for the search sequence. + +The overall optimization objective for DreamTrack includes dreaming loss $\mathcal{L}_{\mathrm{drm}}$ , reconstruction loss $\mathcal{L}_{\mathrm{rect}}$ from the Future Dreaming module, and probability loss $\mathcal{L}_{\mathrm{ce}}$ , L1 loss $\mathcal{L}_{\mathrm{L1}}$ and GIoU loss $\mathcal{L}_{\mathrm{GIoU}}$ from the Multimodal Prediction, + +$$ +\mathcal {L} = \lambda_ {1} \mathcal {L} _ {\mathrm {d r m}} + \lambda_ {2} \mathcal {L} _ {\mathrm {r c t}} + \lambda_ {3} \mathcal {L} _ {\mathrm {c e}} + \lambda_ {4} \mathcal {L} _ {\mathrm {L 1}} + \lambda_ {5} \mathcal {L} _ {\mathrm {G l o U}}, \tag {12} +$$ + +where $\lambda$ values are the balancing weight coefficients. + +Inference. During inference, we initialize the hidden state of GRU for each new video sequence as previously described. Notably, it's no need to sample the video clip for inference. With the defined target queries that automatically cache the historical dynamics, we perform frame-level input for DreamTrack, which is consistent with previous nontemporal works [9, 51] and friendly for the deployment. Besides, the distribution mapping and image reconstruction in Fig. 3 are not required in the inference phase, which could be extended to other tasks, e.g., video generation. + +
MethodsLaSOT [15]LaSOText [16]GOT-10k* [27]TrackingNet [39]TNL2K [46]
SUC(%)PNorm(%)P(%)SUC(%)PNorm(%)P(%)AO(%)SR0.5(%)SR0.75(%)SUC(%)PNorm(%)P(%)SUC(%)P(%)
SiamFC255[2]33.642.033.923.031.126.934.835.39.857.166.353.329.528.6
ECO224[10]32.433.830.122.025.224.031.630.911.155.461.849.232.631.7
SiamRPN++255[32]49.656.949.134.041.639.651.761.632.573.380.069.441.341.2
DiMP288[3]56.965.056.739.247.645.161.171.749.274.080.168.744.743.4
SiamRCNN255[44]64.872.2----64.972.859.781.285.480.052.352.8
Ocean255[53]56.065.156.6---61.172.147.3---38.438.7
AutoMatch255[54]58.3-59.937.6-43.065.276.654.376.0-72.647.243.5
TrDiMP352[45]63.9-61.4---67.177.758.378.483.373.1--
TransT256[6]64.973.869.044.852.352.567.176.860.981.486.780.350.751.7
STARK320[50]67.177.0-47.856.255.168.878.164.182.086.9---
SwinTrack-B384[33]71.3-76.549.1-55.672.480.567.884.0-82.855.957.1
MixFormer-L320[9]70.179.976.3---72.682.268.883.988.983.1--
OSTrack384[51]71.181.177.650.561.357.673.783.270.883.988.583.255.956.7
GRM256[19]69.979.375.8---73.482.970.484.088.783.3--
MixViT288[9]69.679.975.9---72.582.469.983.588.382.0--
ROMTrack384[4]71.481.478.251.362.458.674.284.372.484.189.083.7--
MixViT-L384[9]72.482.280.1---75.785.375.185.490.285.7--
SeqTrack-L384[7]72.581.579.350.761.657.574.881.972.285.589.885.857.8-
ARTrack256[47]70.479.576.646.456.552.373.582.270.984.288.783.557.5-
ARTrack-L384[47]73.182.280.352.862.959.778.587.477.885.689.686.060.3-
ARTrackV2256[1]71.680.277.250.861.957.775.985.472.784.989.384.559.2-
ARTrackV2-L384[1]73.682.881.153.463.760.279.587.879.686.190.486.261.6-
LoRAT224[34]71.780.977.350.361.657.172.181.870.783.587.982.158.861.3
LoRAT-L378[34]75.184.182.056.669.065.177.586.278.185.689.785.462.367.0
DreamTrack25673.883.480.653.164.159.877.587.174.285.890.085.360.463.2
DreamTrack38475.084.281.754.565.361.178.387.976.686.590.685.961.264.8
DreamTrack-L38476.685.683.157.169.465.479.988.580.387.991.386.862.967.4
+ +Table 1. State-of-the-art comparison on LaSOT [15], LaSOText [16], GOT-10k [27], TrackingNet [39] and TNL2K [46]. * indicates only trained on GOT-10k. The number in the subscript denotes the search region resolution. + +# 4. Experiment + +# 4.1. Implementation Details + +Model Variants. We develop three variants of DreamTrack with different configurations as follows: + +- DreamTrack256. Backbone: ViT-Base; Template size: $[128 \times 128]$ ; Search region size: $[256 \times 256]$ ; +- DreamTrack384. Backbone: ViT-Base; Template size: $[192 \times 192]$ ; Search region size: $[384 \times 384]$ ; +- DreamTrack-L384. Backbone: ViT-Large; Template size: $[192 \times 192]$ ; Search region size: $[384 \times 384]$ . + +The ViT backbone [13] is regarded as the encoder of our DreamTrack (see Fig. 2), which is divided into four stages for feature modeling with the proposed Temporal Self-Attention. The four stages contain 2,2,6,2 encoder layers for ViT-Base, and 4,4,12,4 ones for ViT-Large, respectively. + +Training Strategy. The training setting and evaluation protocol follow previous Transformer-based trackers (e.g., OSTrack [51]). The training datasets include COCO [35], LaSOT [15], TrackingNet [39] and GOT-10k [27] (with 1k sequences removed from GOT-10k train split, as per [50]). To ensure a fair evaluation of the GOT-10k test set, our models learn from the entire GOT-10k training split following its one-shot protocol. Our DreamTrack performs training with video clips, as described in Sec. 3.3. To align with the data form in the image dataset COCO that performs frame-level sampling, we repeat the sampled search frame with random shifts to constitute a pseudo video clip. + +We train our model using the AdamW [37] optimizer with a weight decay of $1 \times 10^{-4}$ . The learning rate for the encoder is set to $4 \times 10^{-5}$ , while other parameters use a learning rate of $4 \times 10^{-4}$ . The training process is conducted on 8 NVIDIA RTX3090 GPUs, comprising 300 epochs with 60k video sequences per epoch. The learning rate decreases by a factor of 10 after 240 epochs. + +# 4.2. Main Results + +We evaluate our models on eight prevailing benchmark datasets, including LaSOT [15], LaSOText [16], GOT-10k [27], TrackingNet [39], TNL2K [46], OTB100 [49], NFS [29] and UAV123 [38]. + +LaSOT [15]. LaSOT is a benchmark designed for long-term tracking, comprising 280 videos in its test set. Tab. 1 demonstrates that our DreamTrack $_{256}$ already outperforms ARTrack- $L_{384}$ , which has a larger backbone. Furthermore, our DreamTrack- $L_{384}$ achieved the best performance with the SUC score of $76.6\%$ , without the use of any online template update strategy or post-processing. This proves the superiority of our temporal design with future dreaming. + +LaSOText [16]. LaSOText serves as an extension of LaSOT, which includes additional 150 video sequences. As shown in Tab. 1, our DreamTrack $_{256}$ achieves superior performance compared with other trackers with the same backbone, e.g., ARTrackV2 $_{256}$ and LoRAT $_{224}$ . With the larger backbone, the performance of our DreamTrack-L $_{384}$ further improves with the best SUC score of $57.1\%$ . + +
MethodOTB100 [49]NFS [29]UAV123 [38]
DreamTrack-L38472.467.473.1
DreamTrack38472.066.172.5
DreamTrack25671.865.472.1
LoRAT-L378 [34]72.066.772.5
LoRAT224 [34]71.464.072.6
ARTrackV2-L384 [1]-67.669.9
SeqTrack-L384 [7]-66.268.5
OSTrack384 [51]-66.570.7
TransT256 [6]69.565.769.1
DiMP288 [3]68.461.864.3
Ocean255 [53]68.449.457.4
ATOM288 [11]66.358.363.2
ECO224 [10]66.652.253.5
SiamFC255 [2]58.337.746.8
+ +GOT-10k [27]. GOT-10k provides 180 videos for testing with a strict protocol that only allows trackers to be trained exclusively on its training split. We adhere to this protocol and report the results in Tab. 1. Our DreamTrack- $\mathrm{L}_{384}$ outperforms recent SOTA trackers across all metrics. Notably, our DreamTrack $_{256}$ surpasses other trackers with the same backbone, proving its effectiveness. + +TrackingNet [39]. TrackingNet is an extensive tracking dataset that covers a wide range of real-world scenarios. We assess the performance of DreamTrack on its test set of 511 videos, as illustrated in Tab. 1. It shows that not only does our DreamTrack $_{384}$ outperform all other trackers in SUC, but our DreamTrack- $L_{384}$ also establishes a new state-of-the-art in three metrics on this large-scale benchmark. + +TNL2K [46]. TNL2K is a recently proposed large-scale benchmark for tracking by natural language and bounding box initialization. We evaluate our tracker on its 700 testing videos and the results are listed in Tab. 1. Our DreamTrack-L384 achieves the best performance with the SUC and precision of $62.9\%$ and $67.4\%$ respectively, showing the effectiveness of our temporal learning. + +OTB100 [49], NFS [29] and UAV123 [38]. We also evaluate our tracker on three additional benchmarks, i.e., OTB100, NFS and UAV123 including 100, 100 and 123 videos respectively. Tab. 2 shows that DreamTrack- $\mathrm{L}_{384}$ achieves the best performance on all three benchmarks, demonstrating the strong generalizability of our framework. + +Efficiency Comparison. Tab. 3 compares the efficiency of our DreamTrack against other Transformer-based trackers in terms of inference speed, floating point operations, and number of parameters. Benefiting from the efficient one-stream architecture and sparse multimodal prediction, our DreamTrack shows impressive inference speed with small computation costs. Our most compact variant, DreamTrack $_{256}$ , achieves the state-of-the-art performance while operating at 139 frames per second (FPS), proving the effectiveness and efficiency of our method. + +Table 2. Performance comparison on OTB100 [49], NFS [29] and UAV123 [38] in terms of the SUC score $(\%)$ . + +
TrackerSpeed (FPS)FLOPs (G)Params (M)
SwinTrack-B384 [33]4569.791
OSTrack256 [51]105.421.594
OSTrack384 [51]58.148.394
SeqTrack-B256 [7]406689
SeqTrack-L384 [7]5524309
ARTrack256 [47]26395.6117
ROMTrack256 [4]6234.592
ROMTrack384 [4]2877.792
LoRAT224 [34]2093099
LoRAT-L378 [34]63325336
DreamTrack25613937.687
DreamTrack-L38432359.4317
+ +Table 3. Efficiency comparison. The inference speed is measured on a single NVIDIA RTX2080Ti GPU. + +
#Model VariantsLaSOT [15]TNL2K [46]FPS
SUC(%)PNorm(%)P(%)SUC(%)P(%)
①Baseline69.379.175.354.756.2162
②+ Future Dreaming Module71.781.377.756.959.5153
③+ Temporal Self-Attention72.682.078.559.160.4145
④+ Temporal Cross-Attention73.182.479.059.561.1143
⑤+ Multimodal Prediction73.883.480.660.463.2139
+ +Table 4. Ablation on each proposed module in DreamTrack. + +# 4.3. Component-wise Ablation + +We analyze the influence of each component in our DreamTrack $_{256}$ to show the effectiveness and rationality of the proposed modules, as shown in Tab. 4. When removing all designed modules and directly predicting the target location of the current frame based on the target query, the performance is still comparable with the fastest inference speed of 162 FPS. By applying the key Future Dreaming module on the baseline version, it obtains impressive SUC gains of $2.4\%$ on LaSOT (2 v.s. 1). This verifies that deeper environmental understanding from past experience benefits the generalizability in novel situations. The Temporal Self/Cross-Attention further improves the performance for $0.9\% / 0.5\%$ SUC on LaSOT (4 v.s. 3 v.s. 2), proving the effectiveness of target-aware temporal refinement. The final Multimodal Prediction brings $0.7\% / 0.9\%$ SUC gains on LaSOT and TNL2K, respectively. It indicates that modeling the uncertainty of different future scenarios helps to forecast the target's motion and perform robust tracking. + +# 4.4. Further Analysis + +Self-supervision in Future Dreaming. As mentioned in Sec. 3.2, the capability of predicting the environmental future states relies on the self-supervision in the proposed Future Dreaming module. We explore the influence by ablating the loss functions based on $\mathrm{DreamTrack}_{256}$ . As shown in Tab. 5, removing both the dreaming and reconstruction losses makes the dreamed future unreliable to reflect the real scenario in the next two frames, degrading the quality of learned temporal dynamics (①). Applying each of the two loss functions brings performance gains of $4.4\% / 4.0\%$ SUC on LaSOT respectively (②,③ v.s. ①), proving the ef + +
#Dreaming LossReconstruction LossLaSOT [15]TNL2K [46]
SUC(%)PNorm(%)P(%)SUC(%)P(%)
①--68.377.473.853.753.2
②✓-72.782.679.859.161.7
③-✓72.382.179.258.560.3
④✓✓73.883.480.660.463.2
+ +Table 5. Ablation on loss functions in Future Dreaming module. + +
#Dreaming Step For EvaluationLaSOT [15]LaSOText [16]TNL2K [46]
SUC(%)P(%)SUC(%)P(%)SUC(%)P(%)
①073.880.653.159.860.463.2
②171.377.551.458.258.661.6
③268.574.748.955.055.357.8
+ +Table 6. Performance comparison of different dreaming steps. + +fectiveness. With the full supervision, the performance of DreamTrack $_{256}$ further improves with the SUC of $73.8\%$ on LaSOT (4), which indicates the cooperation of our design. + +Performance of Future Prediction. With the defined target queries of different timestamps, our DreamTrack is capable of predicting the target trajectory including the current and next two frames. We then evaluate the performance of the future predictions by regarding them as the results, e.g., the future prediction for frame 10 from current frame 8 is used for the metric computation. As shown in Tab. 6, the performance of future prediction is inferior to current prediction (2,3 v.s. 1), and the tracking metrics degrade more with the prediction of further future. The underlying reason is that the future observation is unknowable for current prediction, in which the high degree of uncertainty raises challenges to future forecasting. In contrast, the previous trackers only predict the present target location with current observation. Despite these, the comparable performance demonstrates the potential of our design in the motion prediction of the target and benefits visual tracking. + +Activation Analysis and Result Visualization. Future dreaming endows our DreamTrack with superior generalization capability in novel situations. For deeper insights into the mechanism, we first visualize the activation maps $w$ and $w / o$ . The Future Dreaming module. As shown in Fig. 6, the attention areas without the future dreaming are easily distracted by the target's appearance deformation and background clutters, finally losing the target. This proves our claim that insufficient environmental understanding hardly handles the challenges in long-term tracking. In contrast, the version with future dreaming successfully locates the target area with the learned dynamics from past experience, showing the effectiveness of our design. + +We also visualize the results of different trackers, as well as the dreamed future frames. As shown in Fig. 7, the proposed DreamTrack $_{256}$ delivers more robust tracking under deformation (the first row), occlusion and similar interferences (the last two rows). It demonstrates the superiority of our temporal learning in benefiting from both future and + +![](images/44de60cf934d0d237447725bd03730110fddc3ca38ff9a3273de995db2a39023.jpg) +Figure 6. Attention visualization of our DreamTrack $w/$ and $w/o$ . the Future Dreaming module. The predictions and ground-truth boxes are marked in red and green, respectively. + +![](images/d4712ebe160d156b9a462f42a1bb983b1b91303d6d6a8bff241e48e1845f00aa.jpg) +Figure 7. Results visualization of different trackers and dreamed future frames based on learned environmental dynamics. + +historical information. Notably, the dreamed future frames correctly indicate the evolution of the complex environment, and the predicted trajectory successfully locates the tracking target. This indicates the potential of our method to be developed into video generation and annotation tasks. + +Due to limited space, we present more analyses in the supplementary material, including 1) more challenging benchmarks, 2) generalization in different paradigms, 3) position of future dreaming, 4) more visualizations, etc. + +# 5. Conclusion + +We propose DreamTrack, a novel temporal framework with both historical and future information for visual object tracking. We reformulate the temporal learning in VOT as a History-to-Future process, which dreams the future states of the tracking scenario based on learned temporal dynamics from past experience, which empowers the tracker with the generalization in novel situations of new frames. The uncertainty of the future variation is modeled by forecasting the target trajectories of different possible future scenarios with multimodal prediction. Extensive experiments prove our tracker outperforms other mainstream trackers and achieves SOTA performance on prevailing benchmark datasets. + +# Acknowledgments + +This work was supported by the National Natural Science Foundation of China under Grant (62436001,62176020); the Joint Foundation of the Ministry of Education for Innovation Team (8091B042235); the Fundamental Research Funds for the Central Universities (2019JBZ110); and the State Key Laboratory of Rail Traffic Control and Safety (RCS2023K006), Beijing Jiaotong University. + +# References + +[1] Yifan Bai, Zeyang Zhao, Yihong Gong, and Xing Wei. Artrackv2: Prompting autoregressive tracker where to look and how to describe. In CVPR, 2024. 2, 6, 7 +[2] Luca Bertinetto, Jack Valmadre, João F Henriques, Andrea Vedaldi, and Philip H S Torr. Fully-convolutional siamese networks for object tracking. In ECCVW, 2016. 1, 2, 6, 7 +[3] Goutam Bhat, Martin Danelljan, Luc Van Gool, and Radu Timofte. Learning discriminative model prediction for tracking. In ICCV, 2019. 6, 7 +[4] Yidong Cai, Jie Liu, Jie Tang, and Gangshan Wu. Robust object modeling for visual tracking. In ICCV, 2023. 6, 7 +[5] Fei Chen, Xiaodong Wang, Yunxiang Zhao, Shaohe Lv, and Xin Niu. Visual object tracking: A survey. Computer Vision and Image Understanding, 2022. 1 +[6] Xin Chen, Bin Yan, Jiawen Zhu, Dong Wang, Xiaoyun Yang, and Huchuan Lu. Transformer tracking. In CVPR, 2021. 1, 2, 6, 7 +[7] Xin Chen, Houwen Peng, Dong Wang, Hutchuan Lu, and Han Hu. Seqtrack: Sequence to sequence learning for visual object tracking. In CVPR, 2023. 1, 2, 3, 4, 6, 7 +[8] Junyoung Chung, Caglar Gulcehre, KyungHyun Cho, and Yoshua Bengio. Empirical evaluation of gated recurrent neural networks on sequence modeling. arXiv preprint arXiv:1412.3555, 2014. 4 +[9] Yutao Cui, Jiang Cheng, Limin Wang, and Gangshan Wu. Mixformer: End-to-end tracking with iterative mixed attention. In CVPR, 2022. 1, 2, 5, 6 +[10] Martin Danelljan, Goutam Bhat, Fahad Shahbaz Khan, and Michael Felsberg. Eco: Efficient convolution operators for tracking. In CVPR, 2017. 1, 6, 7 +[11] Martin Danelljan, Goutam Bhat, Fahad Shahbaz Khan, and Michael Felsberg. ATOM: Accurate tracking by overlap maximization. In CVPR, 2019. 7 +[12] Martin Danelljan, Luc Van Gool, and Radu Timofte. Probabilistic regression for visual tracking. In CVPR, 2020. 1 +[13] Alexey Dosovitskiy, Lucas Beyer, Alexander Kolesnikov, Dirk Weissenborn, Xiaohua Zhai, Thomas Unterthiner, Mostafa Dehghani, Matthias Minderer, Georg Heigold, Sylvain Gelly, et al. An image is worth 16x16 words: Transformers for image recognition at scale. In ICLR, 2021. 4, 6 +[14] Frederik Ebert, Chelsea Finn, Sudeep Dasari, Annie Xie, Alex Lee, and Sergey Levine. Visual foresight: Model-based deep reinforcement learning for vision-based robotic control. arXiv preprint arXiv:1812.00568, 2018. 2 + +[15] Heng Fan, Liting Lin, Fan Yang, Peng Chu, Ge Deng, Sijia Yu, Hexin Bai, Yong Xu, Chunyuan Liao, and Haibin Ling. LaSOT: A high-quality benchmark for large-scale single object tracking. In CVPR, 2019. 2, 6, 7, 8 +[16] Heng Fan, Hexin Bai, Liting Lin, Fan Yang, Peng Chu, Ge Deng, Sijia Yu, Mingzhen Huang, Juehuan Liu, Yong Xu, et al. Lasot: A high-quality large-scale single object tracking benchmark. IJCV, 2021. 6, 8 +[17] Liangji Fang, Qinhong Jiang, Jianping Shi, and Bolei Zhou. Tpnet: Trajectory proposal network for motion prediction. In CVPR, 2020. 5 +[18] Chelsea Finn and Sergey Levine. Deep visual foresight for planning robot motion. ICRA, 2016. 2 +[19] Shenyuan Gao, Chunluan Zhou, and Jun Zhang. Generalized relation modeling for transformer tracking. In CVPR, 2023. 6 +[20] Dongyan Guo, Jun Wang, Ying Cui, Zhenhua Wang, and Shengyong Chen. SiamCAR: Siamese fully convolutional classification and regression for visual tracking. In CVPR, 2020. 1, 2 +[21] David Ha and Jürgen Schmidhuber. Recurrent world models facilitate policy evolution. NeurIPS, 2018. 2 +[22] Danijar Hafner, Timothy Lillicrap, Jimmy Ba, and Mohammad Norouzi. Dream to control: Learning behaviors by latent imagination. *ICLR*, 2020. 2, 4 +[23] Danijar Hafner, Timothy Lillicrap, Mohammad Norouzi, and Jimmy Ba. Mastering atari with discrete world models. *ICLR*, 2021. 2 +[24] Danijar Hafner, Jurgis Pasukonis, Jimmy Ba, and Timothy Lillicrap. Mastering diverse domains through world models. arXiv preprint arXiv:2301.04104, 2023. 2, 4 +[25] João F Henriques, Rui Caseiro, Pedro Martins, and Jorge Batista. High-speed tracking with kernelized correlation filters. IEEE T-PAMI, 2014. 1 +[26] Anthony Hu, Gianluca Corrado, Nicolas Griffiths, Zachary Murez, Corina Gurau, Hudson Yeo, Alex Kendall, Roberto Cipolla, and Jamie Shotton. Model-based imitation learning for urban driving. NeurIPS, 2022. 2 +[27] Lianghua Huang, Xin Zhao, and Kaiqi Huang. Got-10k: A large high-diversity benchmark for generic object tracking in the wild. IEEE T-PAMI, 2019. 6, 7 +[28] Licheng Jiao, Dan Wang, Yidong Bai, Puhua Chen, and Fang Liu. Deep learning in visual tracking: A review. IEEE TNNLS, 2021. 1 +[29] Hamed Kiani Galoogahi, Ashton Fagg, Chen Huang, Deva Ramanan, and Simon Lucey. Need for speed: A benchmark for higher frame rate object tracking. In ICCV, 2017. 6, 7 +[30] Yann LeCun. A path towards autonomous machine intelligence version 0.9.2, 2022-06-27. Open Review, 2022. 2 +[31] Bo Li, Junjie Yan, Wei Wu, Zheng Zhu, and Xiaolin Hu. High performance visual tracking with siamese region proposal network. In CVPR, 2018. 1, 2 +[32] Bo Li, Wei Wu, Qiang Wang, Fangyi Zhang, Junliang Xing, and Junjie Yan. Siamrpn++: Evolution of siamese visual tracking with very deep networks. In CVPR, 2019. 2, 6 +[33] Liting Lin, Heng Fan, Zhipeng Zhang, Yong Xu, and Haibin Ling. Swintrack: A simple and strong baseline for transformer tracking. NeurIPS, 2022. 2, 6, 7 + +[34] Liting Lin, Heng Fan, Zhipeng Zhang, Yaowei Wang, Yong Xu, and Haibin Ling. Tracking meets lora: Faster training, larger model, stronger performance. ECCV, 2024. 1, 2, 6, 7 +[35] Tsung-Yi Lin, Michael Maire, Serge J. Belongie, Lubomir D. Bourdev, Ross B. Girshick, James Hays, Pietro Perona, Deva Ramanan, Piotr Dollar, and C. Lawrence Zitnick. Microsoft COCO: Common objects in context. In ECCV, 2014. 6 +[36] Yicheng Liu, Jinghuai Zhang, Liangji Fang, Qinhong Jiang, and Bolei Zhou. Multimodal motion prediction with stacked transformers. In CVPR, 2021. 5 +[37] Ilya Loshchilov and Frank Hutter. Decoupled weight decay regularization. In ICLR, 2019. 6 +[38] Matthias Mueller, Neil Smith, and Bernard Ghanem. A benchmark and simulator for uav tracking. In ECCV, 2016. 6, 7 +[39] Matthias Muller, Adel Bibi, Silvio Giancola, Salman Alsubaihi, and Bernard Ghanem. TrackingNet: A large-scale dataset and benchmark for object tracking in the wild. In ECCV, 2018. 6, 7 +[40] Scott Reed, Konrad Zolna, Emilio Parisotto, Sergio Gomez Colmenarejo, Alexander Novikov, Gabriel Barth-maron, Mai Giménez, Yury Sulsky, Jackie Kay, Jost Tobias Springenberg, et al. A generalist agent. TMLR, 2022. 2 +[41] Shaoqing Ren, Kaiming He, Ross Girshick, and Jian Sun. Faster R-CNN: Towards real-time object detection with region proposal networks. In NIPS, 2015. 2, 3 +[42] Hamid Rezatofighi, Nathan Tsoi, JunYoung Gwak, Amir Sadeghian, Ian Reid, and Silvio Savarese. Generalized intersection over union: A metric and a loss for bounding box regression. In CVPR, 2019. 5 +[43] Yuval Tassa, Yotam Doron, Alistair Muldal, Tom Erez, Yazhe Li, Diego de Las Casas, David Budden, Abbas Abdelmaleki, Josh Merel, Andrew Lefrancq, et al. Deepmind control suite. arXiv preprint arXiv:1801.00690, 2018. 2 +[44] Paul Voigtlaender, Jonathon Luiten, Philip HS Torr, and Bastian Leibe. Siam r-cnn: Visual tracking by re-detection. In CVPR, 2020. 2, 6 +[45] Ning Wang, Wengang Zhou, Jie Wang, and Houqiang Li. Transformer meets tracker: Exploiting temporal context for robust visual tracking. In CVPR, 2021. 1, 2, 3, 6 +[46] Xiao Wang, Xiujun Shu, Zhipeng Zhang, Bo Jiang, Yaowei Wang, Yonghong Tian, and Feng Wu. Towards more flexible and accurate object tracking with natural language: Algorithms and benchmark. In CVPR, 2021. 6, 7, 8 +[47] Xing Wei, Yifan Bai, Yongchao Zheng, Dahu Shi, and Yihong Gong. Autoregressive visual tracking. In CVPR, 2023. 1, 2, 3, 4, 6, 7 +[48] Philipp Wu, Alejandro Escontrela, Danijar Hafner, Pieter Abbeel, and Ken Goldberg. Daydreamer: World models for physical robot learning. In CoRL. PMLR, 2023. 2 +[49] Yi Wu, Jongwoo Lim, and Ming Hsuan Yang. Object tracking benchmark. IEEE T-PAMI, 2015. 6, 7 +[50] Bin Yan, Houwen Peng, Jianlong Fu, Dong Wang, and Huchuan Lu. Learning spatio-temporal transformer for visual tracking. In ICCV, 2021. 1, 2, 3, 6 +[51] Botao Ye, Hong Chang, Bingpeng Ma, Shiguang Shan, and Xilin Chen. Joint feature learning and relation modeling for + +tracking: A one-stream framework. In ECCV, 2022. 1, 2, 4, 5, 6, 7 +[52] Zhipeng Zhang and Houwen Peng. Deeper and wider siamese networks for real-time visual tracking. In CVPR, 2019. 1, 2 +[53] Zhipeng Zhang, Houwen Peng, Jianlong Fu, Bing Li, and Weiming Hu. Ocean: Object-aware anchor-free tracking. In ECCV, 2020. 2, 6, 7 +[54] Zhipeng Zhang, Yihao Liu, Xiao Wang, Bing Li, and Weiming Hu. Learn to match: Automatic matching network design for visual tracking. ICCV, 2021. 6 +[55] Xingyi Zhou, Dequan Wang, and Philipp Krahenbuhl. Objects as points. arXiv preprint arXiv:1904.07850, 2019. 2, 3 \ No newline at end of file diff --git a/2025/DreamTrack_ Dreaming the Future for Multimodal Visual Object Tracking/images.zip b/2025/DreamTrack_ Dreaming the Future for Multimodal Visual Object Tracking/images.zip new file mode 100644 index 0000000000000000000000000000000000000000..3848b68ebb51c6dae00a0b32a6f5fb68a1da47e0 --- /dev/null +++ b/2025/DreamTrack_ Dreaming the Future for Multimodal Visual Object Tracking/images.zip @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:802860324eb246dde2fac389373b755817d745b86963807eed4fd0c6d98490f2 +size 737188 diff --git a/2025/DreamTrack_ Dreaming the Future for Multimodal Visual Object Tracking/layout.json b/2025/DreamTrack_ Dreaming the Future for Multimodal Visual Object Tracking/layout.json new file mode 100644 index 0000000000000000000000000000000000000000..bea89ddb1a49584af2f421add37c0197ecc105ae --- /dev/null +++ b/2025/DreamTrack_ Dreaming the Future for Multimodal Visual Object Tracking/layout.json @@ -0,0 +1,9008 @@ +{ + "pdf_info": [ + { + "para_blocks": [ + { + "bbox": [ + 73, + 103, + 538, + 121 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 73, + 103, + 538, + 121 + ], + "spans": [ + { + "bbox": [ + 73, + 103, + 538, + 121 + ], + "type": "text", + "content": "DreamTrack: Dreaming the Future for Multimodal Visual Object Tracking" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 96, + 142, + 517, + 174 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 96, + 142, + 517, + 174 + ], + "spans": [ + { + "bbox": [ + 96, + 142, + 517, + 174 + ], + "type": "text", + "content": "Mingzhe Guo " + }, + { + "bbox": [ + 96, + 142, + 517, + 174 + ], + "type": "inline_equation", + "content": "^{1*}" + }, + { + "bbox": [ + 96, + 142, + 517, + 174 + ], + "type": "text", + "content": " Weiping Tan " + }, + { + "bbox": [ + 96, + 142, + 517, + 174 + ], + "type": "inline_equation", + "content": "^{1}" + }, + { + "bbox": [ + 96, + 142, + 517, + 174 + ], + "type": "text", + "content": " Wenyu Ran " + }, + { + "bbox": [ + 96, + 142, + 517, + 174 + ], + "type": "inline_equation", + "content": "^{1}" + }, + { + "bbox": [ + 96, + 142, + 517, + 174 + ], + "type": "text", + "content": " Liping Jing " + }, + { + "bbox": [ + 96, + 142, + 517, + 174 + ], + "type": "inline_equation", + "content": "^{1\\dagger}" + }, + { + "bbox": [ + 96, + 142, + 517, + 174 + ], + "type": "text", + "content": " Zhipeng Zhang " + }, + { + "bbox": [ + 96, + 142, + 517, + 174 + ], + "type": "inline_equation", + "content": "^{2*}" + }, + { + "bbox": [ + 96, + 142, + 517, + 174 + ], + "type": "text", + "content": " Beijing Jiaotong University " + }, + { + "bbox": [ + 96, + 142, + 517, + 174 + ], + "type": "inline_equation", + "content": "^{1}" + }, + { + "bbox": [ + 96, + 142, + 517, + 174 + ], + "type": "text", + "content": ", Shanghai Jiaotong University " + }, + { + "bbox": [ + 96, + 142, + 517, + 174 + ], + "type": "inline_equation", + "content": "^{2}" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 151, + 199, + 200, + 211 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 151, + 199, + 200, + 211 + ], + "spans": [ + { + "bbox": [ + 151, + 199, + 200, + 211 + ], + "type": "text", + "content": "Abstract" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 55, + 224, + 296, + 523 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 224, + 296, + 523 + ], + "spans": [ + { + "bbox": [ + 55, + 224, + 296, + 523 + ], + "type": "text", + "content": "Aiming to achieve class-agnostic perception in visual object tracking, current trackers commonly formulate tracking as a one-shot detection problem with the template-matching architecture. Despite the success, severe environmental variations in long-term tracking raise challenges to generalizing the tracker in novel situations. Temporal trackers try to fix it by preserving the time-validity of target information with historical predictions, e.g., updating the template. However, solely transmitting the previous observations instead of learning from them leads to an inferior capability of understanding the tracking scenario from past experience, which is critical for the generalization in new frames. To address this issue, we reformulate temporal learning in visual tracking as a History-to-Future process and propose a novel tracking framework DreamTrack. Our DreamTrack learns the temporal dynamics from past observations to dream the future variations of the environment, which boosts the generalization with the extended future information from history. Considering the uncertainty of future variation, multimodal prediction is designed to infer the target trajectory of each possible future situation. The experiments demonstrate that our DreamTrack achieves leading performance with real-time inference speed. In particular, DreamTrack obtains SUC scores of " + }, + { + "bbox": [ + 55, + 224, + 296, + 523 + ], + "type": "inline_equation", + "content": "76.6\\% /87.9\\%" + }, + { + "bbox": [ + 55, + 224, + 296, + 523 + ], + "type": "text", + "content": " on LasOT/TrackingNet, surpassing all recent SOTA trackers." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 56, + 544, + 135, + 556 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 544, + 135, + 556 + ], + "spans": [ + { + "bbox": [ + 56, + 544, + 135, + 556 + ], + "type": "text", + "content": "1. Introduction" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 55, + 564, + 296, + 696 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 564, + 296, + 696 + ], + "spans": [ + { + "bbox": [ + 55, + 564, + 296, + 696 + ], + "type": "text", + "content": "Visual Object Tracking (VOT) is one of the most fundamental computer vision problems with numerous applications [5, 28], which aims to estimate the position of an arbitrary target in a video sequence, given only its location in the initial frame. To achieve the class-agnostic perception, recent prevalent Siamese-based approaches [2, 6, 7, 9, 20, 31, 34, 47, 51, 52] typically formulate visual tracking as a one-shot detection problem. Specifically, these trackers perform general object tracking by matching with the initialized target template and then predicting the location in a single forward pass, as shown in Fig. 1(a)." + } + ] + } + ], + "index": 7 + }, + { + "type": "image", + "bbox": [ + 313, + 196, + 555, + 427 + ], + "blocks": [ + { + "bbox": [ + 313, + 196, + 555, + 427 + ], + "lines": [ + { + "bbox": [ + 313, + 196, + 555, + 427 + ], + "spans": [ + { + "bbox": [ + 313, + 196, + 555, + 427 + ], + "type": "image", + "image_path": "85445e7bf89ad36005c9fa61614c3301311996a6f9e69ff695864fce586ee3ae.jpg" + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 313, + 430, + 555, + 519 + ], + "lines": [ + { + "bbox": [ + 313, + 430, + 555, + 519 + ], + "spans": [ + { + "bbox": [ + 313, + 430, + 555, + 519 + ], + "type": "text", + "content": "Figure 1. Comparison of different temporal architectures. (a) The non-temporal architecture that regards visual tracking as a one-shot detection problem. (b) The temporal architecture with only historical information that limits the temporal messages in the transmission of previous observations. (c) Our proposed temporal architecture with both historical and future information, which learns environmental dynamics from past experiences and dreams the future states for better generalization in subsequent frames." + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "image_caption" + } + ], + "index": 8 + }, + { + "bbox": [ + 313, + 534, + 556, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 534, + 556, + 715 + ], + "spans": [ + { + "bbox": [ + 313, + 534, + 556, + 715 + ], + "type": "text", + "content": "Despite the demonstrated success, it's challenging for the non-temporal architecture to ensure the generalization in new frames with severe environmental variations compared with the initial frame, e.g., deformation of the target appearance and background in long-term tracking. The outdated template hardly provides effective target-aware knowledge but distracts the tracking process. Then many works [7, 10, 12, 25] intuitively try to fix it by keeping the time-validity of target information, e.g., updating the template with predictions of high confidence scores in previous frames [45, 50] or passing historical trajectory as prior knowledge [7, 47], as shown in Fig. 1(b). However, these efforts focus on only transmitting previous observations instead of learning from them, leading to an inferior capability of understanding the tracking scenario from the past expe" + } + ] + } + ], + "index": 10 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "spans": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "type": "text", + "content": "CVF" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 145, + 0, + 496, + 37 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 145, + 0, + 496, + 37 + ], + "spans": [ + { + "bbox": [ + 145, + 0, + 496, + 37 + ], + "type": "text", + "content": "This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 66, + 702, + 216, + 713 + ], + "type": "page_footnote", + "angle": 0, + "lines": [ + { + "bbox": [ + 66, + 702, + 216, + 713 + ], + "spans": [ + { + "bbox": [ + 66, + 702, + 216, + 713 + ], + "type": "text", + "content": "* Equal contribution. † Corresponding author." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 295, + 748, + 314, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 314, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 314, + 757 + ], + "type": "text", + "content": "7201" + } + ] + } + ], + "index": 12 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 0 + }, + { + "para_blocks": [ + { + "bbox": [ + 55, + 72, + 282, + 83 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 72, + 282, + 83 + ], + "spans": [ + { + "bbox": [ + 55, + 72, + 282, + 83 + ], + "type": "text", + "content": "rience so that enables generalization to novel situations." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 55, + 86, + 297, + 396 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 86, + 297, + 396 + ], + "spans": [ + { + "bbox": [ + 55, + 86, + 297, + 396 + ], + "type": "text", + "content": "To address this issue, we reformulate the temporal learning in VOT as a History-to-Future process, as shown in Fig. 1(c). The pipeline dreams the future states of the tracking scenario based on learned temporal dynamics from previous observations, which empowers the tracker with superior generalization ability in complex environments of new frames. Our intuition stems from the human recognition process [21, 30]. The brain abstracts the world from past senses and imagines future variations, which aims to achieve goals in complex scenarios even though they are never encountered before. Such a temporal modeling manner is intuitively superior to the one limited in historical information, which further develops the generalization in new frames with the future information from dreamed states. To this end, we construct a future dreaming module, originally proposed in the world model for reinforcement learning tasks [21-23], to endow the tracker with the capability of predicting future variations of the target from the learned historical dynamics (see Fig. 1(c)). In contrast to non-temporal [6, 9, 31] or historical information-only trackers [7, 47, 50], our dreaming module extends the temporal messages from the history to its future, in which the dreamed future frames are self-supervised with ground-truth future observations. The future information undoubtedly facilitates better generalization for appearance deformation of the target and background variance." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 55, + 397, + 296, + 589 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 397, + 296, + 589 + ], + "spans": [ + { + "bbox": [ + 55, + 397, + 296, + 589 + ], + "type": "text", + "content": "More specifically, we propose a novel tracking framework dubbed DreamTrack, which has a dream-encoder-decoder architecture. The dreaming stage caches historical observations and predicts future states to represent the environmental variation in the next few frames. Then the features of the current and dreamed future frames are input to the encoder, which extracts target-aware features with the temporal messages from both history and future. Notably, after each feature extraction stage of the encoder, the temporal dynamics will be refined along the timeline, which aims to filter target-irrelevant interferences with the exposure of the target information. The final decoder predicts the trajectory of the target in the current and future frames. Considering the uncertainty of the future variation, we forecast the probabilities and target locations of different possible future scenarios, i.e., multimodal prediction." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 55, + 590, + 296, + 674 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 590, + 296, + 674 + ], + "spans": [ + { + "bbox": [ + 55, + 590, + 296, + 674 + ], + "type": "text", + "content": "Extensive experiments on eight prevailing VOT benchmarks show that our proposed DreamTrack outperforms recent state-of-the-art trackers. For instance, under fair settings, DreamTrack-" + }, + { + "bbox": [ + 55, + 590, + 296, + 674 + ], + "type": "inline_equation", + "content": "\\mathrm{L}_{384}" + }, + { + "bbox": [ + 55, + 590, + 296, + 674 + ], + "type": "text", + "content": " obtains a " + }, + { + "bbox": [ + 55, + 590, + 296, + 674 + ], + "type": "inline_equation", + "content": "76.6\\%" + }, + { + "bbox": [ + 55, + 590, + 296, + 674 + ], + "type": "text", + "content": " SUC score on the LaSOT [15] dataset, surpassing ARTrackV2-" + }, + { + "bbox": [ + 55, + 590, + 296, + 674 + ], + "type": "inline_equation", + "content": "\\mathrm{L}_{384}" + }, + { + "bbox": [ + 55, + 590, + 296, + 674 + ], + "type": "text", + "content": " [1] by " + }, + { + "bbox": [ + 55, + 590, + 296, + 674 + ], + "type": "inline_equation", + "content": "3.0\\%" + }, + { + "bbox": [ + 55, + 590, + 296, + 674 + ], + "type": "text", + "content": " and LoRAT-" + }, + { + "bbox": [ + 55, + 590, + 296, + 674 + ], + "type": "inline_equation", + "content": "\\mathrm{L}_{378}" + }, + { + "bbox": [ + 55, + 590, + 296, + 674 + ], + "type": "text", + "content": " [34] by " + }, + { + "bbox": [ + 55, + 590, + 296, + 674 + ], + "type": "inline_equation", + "content": "1.5\\%" + }, + { + "bbox": [ + 55, + 590, + 296, + 674 + ], + "type": "text", + "content": ", respectively. In summary, the contributions of this work are as follows:" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 55, + 677, + 296, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 677, + 296, + 715 + ], + "spans": [ + { + "bbox": [ + 55, + 677, + 296, + 715 + ], + "type": "text", + "content": "- We reformulate the tracking problem as a History-to-Future process and are the first to develop the temporal messages into the future frames for better generalization" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 322, + 72, + 500, + 84 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 322, + 72, + 500, + 84 + ], + "spans": [ + { + "bbox": [ + 322, + 72, + 500, + 84 + ], + "type": "text", + "content": "of novel situations in visual object tracking." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 314, + 85, + 553, + 179 + ], + "type": "list", + "angle": 0, + "index": 8, + "blocks": [ + { + "bbox": [ + 314, + 85, + 553, + 144 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 314, + 85, + 553, + 144 + ], + "spans": [ + { + "bbox": [ + 314, + 85, + 553, + 144 + ], + "type": "text", + "content": "- We make the first attempt at predicting the future states of the target and performing multimodal prediction to include the uncertainty of different future scenarios, which empowers the tracker with superior generalization for appearance deformation and background variation." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 314, + 144, + 553, + 179 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 314, + 144, + 553, + 179 + ], + "spans": [ + { + "bbox": [ + 314, + 144, + 553, + 179 + ], + "type": "text", + "content": "- We prove the effectiveness of the proposed future dreaming to facilitate tracking with state-of-the-art performance on eight prevailing VOT benchmarks." + } + ] + } + ], + "index": 7 + } + ], + "sub_type": "text" + }, + { + "bbox": [ + 314, + 191, + 401, + 204 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 314, + 191, + 401, + 204 + ], + "spans": [ + { + "bbox": [ + 314, + 191, + 401, + 204 + ], + "type": "text", + "content": "2. Related Work" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 313, + 212, + 555, + 475 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 212, + 555, + 475 + ], + "spans": [ + { + "bbox": [ + 313, + 212, + 555, + 475 + ], + "type": "text", + "content": "Visual Object Tracking. Current prevailing trackers [2, 6, 9, 20, 31, 53] commonly employ a template-matching framework, which exploits the initialized template to match the target within the search region. Regarding tracking as a one-shot detection problem, these approaches develop a deep backbone to extract visual features [33, 51, 52], then predict the current target location following the detection paradigms [41, 55], such as object scale estimation [32, 44, 53] and center point localization [6, 20]. Observing the target information of the template may be weakened by the appearance deformation and background variation in long-term tracking, many works design modules to ensure the time-validity of target-relevant knowledge, e.g., updating the template with predictions of high confidence scores in previous frames [45, 50] or passing historical trajectory as prior knowledge [7, 47]. Despite the improved performance in long-term tracking, the limited understanding of the tracking scenario leads to suboptimal generalization in novel situations. Thus, we reformulate tracking as a History-to-Future process that learns the environmental dynamics from past experiences and dreams the future states for better adaptation and tracking capability in new frames." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 313, + 486, + 556, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 486, + 556, + 713 + ], + "spans": [ + { + "bbox": [ + 313, + 486, + 556, + 713 + ], + "type": "text", + "content": "World Model. The world model [30] aims to learn a general representation of the world and predicts future world states resulting from a sequence of actions, which has been widely studied in the game [21-23, 40] and lab environments [14, 18, 48]. Dreamer [22] learns a latent dynamics model from past experience to predict future state values and actions in a latent space. It's capable of handling challenging visual control tasks in the DeepMind Control Suite [43]. DreamerV2 [23] further achieves human-level performance on Atari games with discrete modeling. DreamerV3 [24] develops a general algorithm and allows solving challenging control problems such as obtaining diamonds in Minecraft from scratch given sparse rewards. Recently, learning world models in driving scenes has gained attention. MILE [26] employs a model-based imitation learning method to jointly learn a dynamics model and driving behavior. In this work, we introduce the mechanism of the world model to visual object tracking, which boosts environmental understanding with the dreamed future states." + } + ] + } + ], + "index": 11 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "text", + "content": "7202" + } + ] + } + ], + "index": 12 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 1 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 56, + 70, + 555, + 251 + ], + "blocks": [ + { + "bbox": [ + 56, + 70, + 555, + 251 + ], + "lines": [ + { + "bbox": [ + 56, + 70, + 555, + 251 + ], + "spans": [ + { + "bbox": [ + 56, + 70, + 555, + 251 + ], + "type": "image", + "image_path": "50d33d9bbd2e086ede54ab20a13743a7d7c9c12e918522a61f5a5823f0620ca4.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 55, + 259, + 555, + 315 + ], + "lines": [ + { + "bbox": [ + 55, + 259, + 555, + 315 + ], + "spans": [ + { + "bbox": [ + 55, + 259, + 555, + 315 + ], + "type": "text", + "content": "Figure 2. The dream-encoder-decoder framework of our DreamTrack. Initially, the dreaming stage takes the input image tokens to predict future states of the next two frames based on the cached temporal dynamics in the Future Dreaming module. Then features of the current frame and dreamed future are fed into the encoder for further modeling, with a Temporal Self-Attention after each encoder stage for target-aware refinement. In the end, the decoder performs multimodal prediction for the target, i.e., the probabilities and trajectories in the current and next two frames of three possible future scenarios. The prediction with the highest probability is selected as the final output." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "bbox": [ + 55, + 334, + 111, + 346 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 334, + 111, + 346 + ], + "spans": [ + { + "bbox": [ + 55, + 334, + 111, + 346 + ], + "type": "text", + "content": "3. Method" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 55, + 354, + 141, + 365 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 354, + 141, + 365 + ], + "spans": [ + { + "bbox": [ + 55, + 354, + 141, + 365 + ], + "type": "text", + "content": "3.1. Preliminaries" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 55, + 371, + 296, + 479 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 371, + 296, + 479 + ], + "spans": [ + { + "bbox": [ + 55, + 371, + 296, + 479 + ], + "type": "text", + "content": "Formulation of VOT. Given the template image " + }, + { + "bbox": [ + 55, + 371, + 296, + 479 + ], + "type": "inline_equation", + "content": "\\mathbf{z}^0" + }, + { + "bbox": [ + 55, + 371, + 296, + 479 + ], + "type": "text", + "content": " of a target initialized in the first frame, the objective of visual tracking is to localize the target " + }, + { + "bbox": [ + 55, + 371, + 296, + 479 + ], + "type": "inline_equation", + "content": "\\mathrm{b}^t" + }, + { + "bbox": [ + 55, + 371, + 296, + 479 + ], + "type": "text", + "content": " from the continuous input image " + }, + { + "bbox": [ + 55, + 371, + 296, + 479 + ], + "type": "inline_equation", + "content": "\\mathbf{x}^t" + }, + { + "bbox": [ + 55, + 371, + 296, + 479 + ], + "type": "text", + "content": ", where " + }, + { + "bbox": [ + 55, + 371, + 296, + 479 + ], + "type": "inline_equation", + "content": "\\mathbf{b}" + }, + { + "bbox": [ + 55, + 371, + 296, + 479 + ], + "type": "text", + "content": " denotes the target state, i.e., a bounding box. The mainstream tracking paradigm aims to learn a deep tracking model " + }, + { + "bbox": [ + 55, + 371, + 296, + 479 + ], + "type": "inline_equation", + "content": "f" + }, + { + "bbox": [ + 55, + 371, + 296, + 479 + ], + "type": "text", + "content": " that performs localization with the output " + }, + { + "bbox": [ + 55, + 371, + 296, + 479 + ], + "type": "inline_equation", + "content": "\\mathbf{y}^t" + }, + { + "bbox": [ + 55, + 371, + 296, + 479 + ], + "type": "text", + "content": " following different detection paradigms, i.e., anchor-based [41] or anchor-free [55] architectures. The tracking process can be formulated as," + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 143, + 486, + 294, + 500 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 143, + 486, + 294, + 500 + ], + "spans": [ + { + "bbox": [ + 143, + 486, + 294, + 500 + ], + "type": "interline_equation", + "content": "y ^ {t} = f \\left(z ^ {0}, x ^ {t}\\right). \\tag {1}", + "image_path": "de17dbeb3b0b53ac9d485eed221b076302e3340ffcf899d0beda0957f8f34459.jpg" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 55, + 508, + 296, + 579 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 508, + 296, + 579 + ], + "spans": [ + { + "bbox": [ + 55, + 508, + 296, + 579 + ], + "type": "text", + "content": "To alleviate the influence of target's appearance deformation and background variation in long-term tracking, recent prevailing temporal trackers exploit previous outputs " + }, + { + "bbox": [ + 55, + 508, + 296, + 579 + ], + "type": "inline_equation", + "content": "\\mathrm{y}^{0:t - 1}" + }, + { + "bbox": [ + 55, + 508, + 296, + 579 + ], + "type": "text", + "content": " as additional prior knowledge, e.g., updating the template [45, 50] or passing historical target location [7, 47]. The temporal tracking process is formulated as," + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 127, + 586, + 295, + 601 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 127, + 586, + 295, + 601 + ], + "spans": [ + { + "bbox": [ + 127, + 586, + 295, + 601 + ], + "type": "interline_equation", + "content": "\\mathrm {y} ^ {t} = f \\left(\\mathrm {z} ^ {0}, \\mathrm {x} ^ {t}, \\mathrm {y} ^ {0: t - 1}\\right). \\tag {2}", + "image_path": "52a8b624da2a8c5d95c3141e5c67e8a09c83338e36b6587f7aca922fdc85af0d.jpg" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 55, + 609, + 296, + 693 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 609, + 296, + 693 + ], + "spans": [ + { + "bbox": [ + 55, + 609, + 296, + 693 + ], + "type": "text", + "content": "In contrast, we formulate temporal visual tracking as a History-to-Future process, which predicts the future trajectory of the target (including the current frame and the next two ones) based on the learned environmental dynamics from historical observations (cached in the target queries " + }, + { + "bbox": [ + 55, + 609, + 296, + 693 + ], + "type": "inline_equation", + "content": "\\mathbf{Q}_{\\mathrm{tgt}}^{t:t + 2}" + }, + { + "bbox": [ + 55, + 609, + 296, + 693 + ], + "type": "text", + "content": " of " + }, + { + "bbox": [ + 55, + 609, + 296, + 693 + ], + "type": "inline_equation", + "content": "f" + }, + { + "bbox": [ + 55, + 609, + 296, + 693 + ], + "type": "text", + "content": ", described in Sec. 3.2). This modeling pipeline can be expressed as," + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 118, + 700, + 295, + 715 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 118, + 700, + 295, + 715 + ], + "spans": [ + { + "bbox": [ + 118, + 700, + 295, + 715 + ], + "type": "interline_equation", + "content": "y ^ {t: t + 2} = f \\left[ Q _ {\\mathrm {t g t}} ^ {t: t + 2} \\right] \\left(z ^ {0}, x ^ {t}\\right). \\tag {3}", + "image_path": "cd29251a4f25c6d5b0048bcd9636b471ac4428d6a378d4f31d88d7dfe7fbee4a.jpg" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 313, + 335, + 555, + 396 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 335, + 555, + 396 + ], + "spans": [ + { + "bbox": [ + 313, + 335, + 555, + 396 + ], + "type": "text", + "content": "The parameterized modeling of environmental dynamics " + }, + { + "bbox": [ + 313, + 335, + 555, + 396 + ], + "type": "inline_equation", + "content": "(\\mathbf{Q}_{\\mathrm{tg}}^{t:t + 2})" + }, + { + "bbox": [ + 313, + 335, + 555, + 396 + ], + "type": "text", + "content": " instead of explicit transmission of historical output " + }, + { + "bbox": [ + 313, + 335, + 555, + 396 + ], + "type": "inline_equation", + "content": "\\mathrm{y}^{0:t - 1}" + }, + { + "bbox": [ + 313, + 335, + 555, + 396 + ], + "type": "text", + "content": ", not only enables better generalization in novel situations of future frames, but also reduces the risk of error accumulation from past experiences." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 313, + 403, + 555, + 630 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 403, + 555, + 630 + ], + "spans": [ + { + "bbox": [ + 313, + 403, + 555, + 630 + ], + "type": "text", + "content": "Trajectory Coordinate Mapping. To reduce computation costs, most methods perform tracking on the cropped search region instead of the full-resolution frame, which outputs target's coordinates in the current frame, relative to the search region. Thus, it's necessary to map the predicted target location of the relative coordinate system to the original frame after inference, which we denote as the global coordinate system. Correspondingly, the ground-truth boxes for training are also transformed from the global to relative coordinate system. For our DreamTrack that learns dynamics and predicts target trajectory over a video sequence, we additionally cache the ground-truth boxes of the preceding two frames in the global coordinate system and map them to the relative coordinate system of the current frame after the search region is cropped. Notably, the search region patches of the two future frames are cropped by referring to the ground-truth box of the current frame, which represent the target's motion and serve for the supervision of future image reconstruction (described in Sec. 3.2 and Fig. 3)." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 313, + 636, + 440, + 647 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 636, + 440, + 647 + ], + "spans": [ + { + "bbox": [ + 313, + 636, + 440, + 647 + ], + "type": "text", + "content": "3.2. Network Architecture" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 313, + 653, + 554, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 653, + 554, + 713 + ], + "spans": [ + { + "bbox": [ + 313, + 653, + 554, + 713 + ], + "type": "text", + "content": "The overall framework of DreamTrack is depicted in Fig. 2. Given the target sequences constructed from object trajectories, we use a novel dream-encoder-decoder structure for learning and inference. The target state is described by a trajectory including the current and next two frames." + } + ] + } + ], + "index": 13 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "text", + "content": "7203" + } + ] + } + ], + "index": 14 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 2 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 59, + 72, + 293, + 220 + ], + "blocks": [ + { + "bbox": [ + 59, + 72, + 293, + 220 + ], + "lines": [ + { + "bbox": [ + 59, + 72, + 293, + 220 + ], + "spans": [ + { + "bbox": [ + 59, + 72, + 293, + 220 + ], + "type": "image", + "image_path": "f5bbecd65d43e3839113a97abc270b539916e8ca1521b7ce2122cfb022edb8fb.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 55, + 225, + 295, + 293 + ], + "lines": [ + { + "bbox": [ + 55, + 225, + 295, + 293 + ], + "spans": [ + { + "bbox": [ + 55, + 225, + 295, + 293 + ], + "type": "text", + "content": "Figure 3. Illustration of the proposed Future Dreaming module. The target queries auto-regressively update the temporal dynamics with the patch embeddings of the current frame through GRU, which are then exploited to predict the future states. For training, the distributions of the predicted future states and decoded future images are self-supervised by the ground-truth future frames." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "bbox": [ + 55, + 306, + 296, + 474 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 306, + 296, + 474 + ], + "spans": [ + { + "bbox": [ + 55, + 306, + 296, + 474 + ], + "type": "text", + "content": "Dream. We propose the Future Dreaming module to learn temporal dynamics of the tracking scenario. As shown in Fig. 3, the module evolves the dynamics with current observation and dreams the future states, which endows the tracker with the capability of future prediction by supervising with ground-truth future frames. Specifically, three trainable target queries " + }, + { + "bbox": [ + 55, + 306, + 296, + 474 + ], + "type": "inline_equation", + "content": "\\mathbf{Q}_{\\mathrm{tgt}}^{t - 1:t + 1} \\in \\mathbb{R}^{3 \\times C}" + }, + { + "bbox": [ + 55, + 306, + 296, + 474 + ], + "type": "text", + "content": ", which represent the transition dynamics of the current and next two frames, are initialized. With the template " + }, + { + "bbox": [ + 55, + 306, + 296, + 474 + ], + "type": "inline_equation", + "content": "z^0" + }, + { + "bbox": [ + 55, + 306, + 296, + 474 + ], + "type": "text", + "content": " and search region " + }, + { + "bbox": [ + 55, + 306, + 296, + 474 + ], + "type": "inline_equation", + "content": "x^t" + }, + { + "bbox": [ + 55, + 306, + 296, + 474 + ], + "type": "text", + "content": " of each new frame, the images are split and flattened to build current observation feature " + }, + { + "bbox": [ + 55, + 306, + 296, + 474 + ], + "type": "inline_equation", + "content": "\\mathbf{F}^{\\mathrm{t}} = \\{\\mathbf{F}_z^0 \\in \\mathbb{R}^{H_z W_z \\times C}, \\mathbf{F}_x^t \\in \\mathbb{R}^{H_x W_x \\times C}\\}" + }, + { + "bbox": [ + 55, + 306, + 296, + 474 + ], + "type": "text", + "content": " with a linear projection. Then target queries auto-regressively update the cached temporal messages with " + }, + { + "bbox": [ + 55, + 306, + 296, + 474 + ], + "type": "inline_equation", + "content": "\\mathbf{F}^{\\mathrm{t}}" + }, + { + "bbox": [ + 55, + 306, + 296, + 474 + ], + "type": "text", + "content": " through a gated recurrent unit (GRU) [8]," + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 110, + 484, + 295, + 499 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 110, + 484, + 295, + 499 + ], + "spans": [ + { + "bbox": [ + 110, + 484, + 295, + 499 + ], + "type": "interline_equation", + "content": "\\mathbf {Q} _ {\\mathrm {t g t}} ^ {t: t + 2} = \\operatorname {G R U} \\left(\\mathbf {Q} _ {\\mathrm {t g t}} ^ {t - 1: t + 1}, \\mathbf {F} ^ {t}\\right). \\tag {4}", + "image_path": "b892d5fe145eb72c362482561506327524b9a17d9a2219232d911173e99f38af.jpg" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 55, + 502, + 296, + 586 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 502, + 296, + 586 + ], + "spans": [ + { + "bbox": [ + 55, + 502, + 296, + 586 + ], + "type": "text", + "content": "As the aim of our DreamTrack is to enhance the generalization in novel situations of new frames, it's a prerequisite to evolve the future variations of the tracking environment. Thus, the updated target queries " + }, + { + "bbox": [ + 55, + 502, + 296, + 586 + ], + "type": "inline_equation", + "content": "\\mathbf{Q}_{\\mathrm{tgt}}^{t:t + 2}" + }, + { + "bbox": [ + 55, + 502, + 296, + 586 + ], + "type": "text", + "content": " are exploited to backward imagine the future observation feature " + }, + { + "bbox": [ + 55, + 502, + 296, + 586 + ], + "type": "inline_equation", + "content": "\\hat{\\mathbf{F}}^{t:t + 2} = \\{\\hat{\\mathbf{F}}_{\\mathrm{z}}^{t:t + 2}\\in \\mathbb{R}^{3\\times H_{\\mathrm{z}}W_{\\mathrm{z}}\\times C},\\hat{\\mathbf{F}}_{\\mathrm{x}}^{t:t + 2}\\in \\mathbb{R}^{3\\times H_{\\mathrm{x}}W_{\\mathrm{x}}\\times C}\\}" + }, + { + "bbox": [ + 55, + 502, + 296, + 586 + ], + "type": "text", + "content": " based on " + }, + { + "bbox": [ + 55, + 502, + 296, + 586 + ], + "type": "inline_equation", + "content": "\\mathbf{F}^{\\mathrm{t}}" + }, + { + "bbox": [ + 55, + 502, + 296, + 586 + ], + "type": "text", + "content": " with a cross-attention," + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 67, + 591, + 295, + 608 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 67, + 591, + 295, + 608 + ], + "spans": [ + { + "bbox": [ + 67, + 591, + 295, + 608 + ], + "type": "interline_equation", + "content": "\\hat {\\mathbf {F}} ^ {t: t + 2} = \\operatorname {C r o s s A t t e n t i o n} (q = \\mathbf {F} ^ {t}, k v = \\mathbf {Q} _ {\\mathrm {t g t}} ^ {t: t + 2}). \\quad (5)", + "image_path": "faf91e8b85555989e626e941c3ea9352d7373b08cfe9be5ec68612a36b9d5d5a.jpg" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 55, + 613, + 295, + 689 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 613, + 295, + 689 + ], + "spans": [ + { + "bbox": [ + 55, + 613, + 295, + 689 + ], + "type": "text", + "content": "Notably, the interaction between " + }, + { + "bbox": [ + 55, + 613, + 295, + 689 + ], + "type": "inline_equation", + "content": "\\mathbf{F}^t" + }, + { + "bbox": [ + 55, + 613, + 295, + 689 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 55, + 613, + 295, + 689 + ], + "type": "inline_equation", + "content": "\\mathbf{Q}_{\\mathrm{tgt}}^t" + }, + { + "bbox": [ + 55, + 613, + 295, + 689 + ], + "type": "text", + "content": " to generate " + }, + { + "bbox": [ + 55, + 613, + 295, + 689 + ], + "type": "inline_equation", + "content": "\\hat{\\mathbf{F}}^t" + }, + { + "bbox": [ + 55, + 613, + 295, + 689 + ], + "type": "text", + "content": " intends to enhance the feature quality of the current frame by introducing historical information, which originates from " + }, + { + "bbox": [ + 55, + 613, + 295, + 689 + ], + "type": "inline_equation", + "content": "\\mathbf{Q}_{\\mathrm{tgt}}^{t - 1}" + }, + { + "bbox": [ + 55, + 613, + 295, + 689 + ], + "type": "text", + "content": " before the GRU update (Eq. 4). The output " + }, + { + "bbox": [ + 55, + 613, + 295, + 689 + ], + "type": "inline_equation", + "content": "\\hat{\\mathbf{F}}^{t:t + 2}" + }, + { + "bbox": [ + 55, + 613, + 295, + 689 + ], + "type": "text", + "content": " of the module will be further modeled with the Temporal Self-Attention, which is detailed in the encoder." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 55, + 689, + 296, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 689, + 296, + 713 + ], + "spans": [ + { + "bbox": [ + 55, + 689, + 296, + 713 + ], + "type": "text", + "content": "Following the loss design in world models [22-24], we respectively map the dreamed future observation fea" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 313, + 71, + 553, + 240 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 71, + 553, + 240 + ], + "spans": [ + { + "bbox": [ + 313, + 71, + 553, + 240 + ], + "type": "text", + "content": "ture " + }, + { + "bbox": [ + 313, + 71, + 553, + 240 + ], + "type": "inline_equation", + "content": "\\hat{\\mathbf{F}}_{\\mathrm{x}}^{t:t + 2}" + }, + { + "bbox": [ + 313, + 71, + 553, + 240 + ], + "type": "text", + "content": " and ground-truth observation feature " + }, + { + "bbox": [ + 313, + 71, + 553, + 240 + ], + "type": "inline_equation", + "content": "\\mathbf{F}_{\\mathrm{x}}^{t:t + 2}" + }, + { + "bbox": [ + 313, + 71, + 553, + 240 + ], + "type": "text", + "content": " to distributions of " + }, + { + "bbox": [ + 313, + 71, + 553, + 240 + ], + "type": "inline_equation", + "content": "\\{\\hat{\\mu}_{\\mathrm{x}}^{t:t + 2},\\hat{\\sigma}_{\\mathrm{x}}^{t:t + 2}\\in \\mathbb{R}^{3\\times H_{\\mathrm{x}}W_{\\mathrm{x}}\\times C^{\\prime}}\\}" + }, + { + "bbox": [ + 313, + 71, + 553, + 240 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 313, + 71, + 553, + 240 + ], + "type": "inline_equation", + "content": "\\{\\mu_{\\mathrm{x}}^{t:t + 2},\\sigma_{\\mathrm{x}}^{t:t + 2}\\in \\mathbb{R}^{3\\times H_{\\mathrm{x}}W_{\\mathrm{x}}\\times C^{\\prime}}\\}" + }, + { + "bbox": [ + 313, + 71, + 553, + 240 + ], + "type": "text", + "content": " with MLP layers, and then minimize their KL divergence. Notably, we only consider supervising the dreaming process of the search region, as the input template always comes from the initial frame. The prior distribution from " + }, + { + "bbox": [ + 313, + 71, + 553, + 240 + ], + "type": "inline_equation", + "content": "\\hat{\\mathbf{F}}_{\\mathrm{x}}^{t:t + 2}" + }, + { + "bbox": [ + 313, + 71, + 553, + 240 + ], + "type": "text", + "content": " is regarded as a prediction of the future dynamics. In contrast, the posterior distribution from " + }, + { + "bbox": [ + 313, + 71, + 553, + 240 + ], + "type": "inline_equation", + "content": "\\mathbf{F}_{\\mathrm{x}}^{t:t + 2}" + }, + { + "bbox": [ + 313, + 71, + 553, + 240 + ], + "type": "text", + "content": " represents the future dynamics from the ground-truth observation " + }, + { + "bbox": [ + 313, + 71, + 553, + 240 + ], + "type": "inline_equation", + "content": "\\mathbf{x}^{t:t + 2}" + }, + { + "bbox": [ + 313, + 71, + 553, + 240 + ], + "type": "text", + "content": ". The KL divergence measures the gap between the imagined future (prior) and the true future (posterior). We expect to enhance the capability of future prediction for long-term tracking, which is realized by optimizing the dreaming loss " + }, + { + "bbox": [ + 313, + 71, + 553, + 240 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{\\mathrm{drm}}" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 330, + 245, + 553, + 261 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 330, + 245, + 553, + 261 + ], + "spans": [ + { + "bbox": [ + 330, + 245, + 553, + 261 + ], + "type": "interline_equation", + "content": "\\mathcal {L} _ {\\mathrm {d r m}} = \\mathrm {K L} \\left(\\left\\{\\hat {\\mu} _ {\\mathrm {x}} ^ {t: t + 2}, \\hat {\\sigma} _ {\\mathrm {x}} ^ {t: t + 2} \\right\\} \\mid \\mid \\left\\{\\mu_ {\\mathrm {x}} ^ {t: t + 2}, \\sigma_ {\\mathrm {x}} ^ {t: t + 2} \\right\\}\\right), \\tag {6}", + "image_path": "293502095095a75976eb0fcf3f9caef69e7a7977ec0ca30a617dd9a1837a3eb2.jpg" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 313, + 266, + 489, + 277 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 266, + 489, + 277 + ], + "spans": [ + { + "bbox": [ + 313, + 266, + 489, + 277 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 313, + 266, + 489, + 277 + ], + "type": "inline_equation", + "content": "\\mathrm{KL}(\\cdot ||\\cdot)" + }, + { + "bbox": [ + 313, + 266, + 489, + 277 + ], + "type": "text", + "content": " denotes the KL divergence." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 313, + 278, + 553, + 361 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 278, + 553, + 361 + ], + "spans": [ + { + "bbox": [ + 313, + 278, + 553, + 361 + ], + "type": "text", + "content": "Besides the supervision from the distribution difference, we further improve the accuracy of future prediction by reconstructing the future search regions based on " + }, + { + "bbox": [ + 313, + 278, + 553, + 361 + ], + "type": "inline_equation", + "content": "\\hat{\\mathbf{F}}_{\\mathrm{x}}^{t:t + 2}" + }, + { + "bbox": [ + 313, + 278, + 553, + 361 + ], + "type": "text", + "content": " with a decoder of several CNN upsampling layers. The decoded future search regions " + }, + { + "bbox": [ + 313, + 278, + 553, + 361 + ], + "type": "inline_equation", + "content": "\\hat{\\mathbf{x}}^{t:t + 2}" + }, + { + "bbox": [ + 313, + 278, + 553, + 361 + ], + "type": "text", + "content": " is supervised by the groundtruth search regions " + }, + { + "bbox": [ + 313, + 278, + 553, + 361 + ], + "type": "inline_equation", + "content": "\\mathbf{x}^{t:t + 2}" + }, + { + "bbox": [ + 313, + 278, + 553, + 361 + ], + "type": "text", + "content": " by optimizing the reconstruction loss " + }, + { + "bbox": [ + 313, + 278, + 553, + 361 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{\\mathrm{RCT}}" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 375, + 361, + 553, + 374 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 375, + 361, + 553, + 374 + ], + "spans": [ + { + "bbox": [ + 375, + 361, + 553, + 374 + ], + "type": "interline_equation", + "content": "\\mathcal {L} _ {\\mathrm {r c t}} = \\operatorname {M S E} \\left(\\hat {\\mathrm {x}} ^ {t: t + 2} | | \\mathrm {x} ^ {t: t + 2}\\right), \\tag {7}", + "image_path": "502fd72cb7d1e2496a461da8a063abbf4a50e4d2569774438f1c96460f540597.jpg" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 313, + 378, + 490, + 391 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 378, + 490, + 391 + ], + "spans": [ + { + "bbox": [ + 313, + 378, + 490, + 391 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 313, + 378, + 490, + 391 + ], + "type": "inline_equation", + "content": "\\mathrm{MSE}(\\cdot ||\\cdot)" + }, + { + "bbox": [ + 313, + 378, + 490, + 391 + ], + "type": "text", + "content": " is the mean squared error." + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 313, + 396, + 554, + 492 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 396, + 554, + 492 + ], + "spans": [ + { + "bbox": [ + 313, + 396, + 554, + 492 + ], + "type": "text", + "content": "Encoder. Following previous works [7, 47, 51], we adopt a plain vision transformer (ViT) [13] as the encoder and model the search and template images jointly. The encoder extracts the features of the search and template images and learns feature-level correspondence to filter target-relevant distractors. The input dreamed observation features " + }, + { + "bbox": [ + 313, + 396, + 554, + 492 + ], + "type": "inline_equation", + "content": "\\{\\hat{\\mathbf{F}}_{\\mathrm{z}}^{t:t + 2},\\hat{\\mathbf{F}}_{\\mathrm{x}}^{t:t + 2}\\}" + }, + { + "bbox": [ + 313, + 396, + 554, + 492 + ], + "type": "text", + "content": " are concatenated along the spatial dimension and fed into the encoder for feature extraction," + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 344, + 497, + 553, + 531 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 344, + 497, + 553, + 531 + ], + "spans": [ + { + "bbox": [ + 344, + 497, + 553, + 531 + ], + "type": "interline_equation", + "content": "\\begin{array}{l} \\hat {\\mathbf {F}} _ {\\mathrm {z x}, i} ^ {t: t + 2} = \\operatorname {C o n c a t} \\left(\\hat {\\mathbf {F}} _ {\\mathrm {z}} ^ {t: t + 2}, \\hat {\\mathbf {F}} _ {\\mathrm {x}} ^ {t: t + 2}\\right), i = 0 \\\\ \\hat {\\mathbf {F}} _ {\\mathrm {z x}, i} ^ {t: t + 2} = \\operatorname {B l o c k} _ {i} \\left(\\hat {\\mathbf {F}} _ {\\mathrm {z x}, i - 1} ^ {t: t + 2}\\right), \\quad 1 \\leq i \\leq l _ {\\text {s p t}} \\tag {8} \\\\ \\end{array}", + "image_path": "6e12c4901b8031eeadb5b5f24b1e36b27bfc6c226cc6fdf5dc74d3bc61f547f0.jpg" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 313, + 536, + 554, + 574 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 536, + 554, + 574 + ], + "spans": [ + { + "bbox": [ + 313, + 536, + 554, + 574 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 313, + 536, + 554, + 574 + ], + "type": "inline_equation", + "content": "\\hat{\\mathbf{F}}_{\\mathrm{zx},i}^{t:t + 2}\\in \\mathbb{R}^{3\\times (H_{\\mathrm{z}}W_{\\mathrm{z}} + H_{\\mathrm{x}}W_{\\mathrm{x}})\\times C}" + }, + { + "bbox": [ + 313, + 536, + 554, + 574 + ], + "type": "text", + "content": ", Block denotes the transformer block in the encoder for spatial modeling and " + }, + { + "bbox": [ + 313, + 536, + 554, + 574 + ], + "type": "inline_equation", + "content": "l_{\\mathrm{spt}}" + }, + { + "bbox": [ + 313, + 536, + 554, + 574 + ], + "type": "text", + "content": " is the layer number." + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 313, + 574, + 554, + 694 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 574, + 554, + 694 + ], + "spans": [ + { + "bbox": [ + 313, + 574, + 554, + 694 + ], + "type": "text", + "content": "Besides the spatial modeling between the template and search region in Block, it's also important to exploit the extracted target-aware information to refine the temporal dynamics along the timeline. Thus, after each feature extraction stage of the encoder, we arrange a Temporal Self-Attention module to filter target-irrelevant interferences with the exposure of the target information from different levels. As shown in Fig. 4, the dreamed observation features " + }, + { + "bbox": [ + 313, + 574, + 554, + 694 + ], + "type": "inline_equation", + "content": "\\hat{\\mathbf{F}}_{\\mathrm{zx}}^{t:t + 2}" + }, + { + "bbox": [ + 313, + 574, + 554, + 694 + ], + "type": "text", + "content": " are transposed and input to the attention for the interaction on the temporal dimension," + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 330, + 699, + 553, + 715 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 330, + 699, + 553, + 715 + ], + "spans": [ + { + "bbox": [ + 330, + 699, + 553, + 715 + ], + "type": "interline_equation", + "content": "\\hat {\\mathbf {F}} _ {\\mathrm {z x}, i} ^ {t: t + 2} = \\operatorname {T e m A t t n} _ {\\text {s e l f}} ^ {i} \\left(\\hat {\\mathbf {F}} _ {\\mathrm {z x}, i - 1} ^ {t: t + 2}\\right), \\quad 1 \\leq i \\leq l _ {\\text {t e m}} \\tag {9}", + "image_path": "0d24709b7bdc657c03c2e6376aba69a4b2cc2ed27216f8ad9041e4abdd28d678.jpg" + } + ] + } + ], + "index": 18 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "text", + "content": "7204" + } + ] + } + ], + "index": 19 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 3 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 57, + 70, + 295, + 202 + ], + "blocks": [ + { + "bbox": [ + 57, + 70, + 295, + 202 + ], + "lines": [ + { + "bbox": [ + 57, + 70, + 295, + 202 + ], + "spans": [ + { + "bbox": [ + 57, + 70, + 295, + 202 + ], + "type": "image", + "image_path": "86554f4977ac27dede5172890e3b01c454a47ba1a6fee32f3591f2f46a97ddc2.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 55, + 207, + 296, + 262 + ], + "lines": [ + { + "bbox": [ + 55, + 207, + 296, + 262 + ], + "spans": [ + { + "bbox": [ + 55, + 207, + 296, + 262 + ], + "type": "text", + "content": "Figure 4. Illustration of the Temporal Self/Cross-Attention. For self-attention, patch embeddings of the template and search region are jointly modeled on the temporal dimension. In contrast, the cross-attention takes target queries as " + }, + { + "bbox": [ + 55, + 207, + 296, + 262 + ], + "type": "inline_equation", + "content": "\\mathbf{Q}" + }, + { + "bbox": [ + 55, + 207, + 296, + 262 + ], + "type": "text", + "content": " and patch embeddings of search region as " + }, + { + "bbox": [ + 55, + 207, + 296, + 262 + ], + "type": "inline_equation", + "content": "\\mathbf{K}" + }, + { + "bbox": [ + 55, + 207, + 296, + 262 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 55, + 207, + 296, + 262 + ], + "type": "inline_equation", + "content": "\\mathbf{V}" + }, + { + "bbox": [ + 55, + 207, + 296, + 262 + ], + "type": "text", + "content": " to extract localization-relevant information." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "bbox": [ + 55, + 275, + 296, + 312 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 275, + 296, + 312 + ], + "spans": [ + { + "bbox": [ + 55, + 275, + 296, + 312 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 55, + 275, + 296, + 312 + ], + "type": "inline_equation", + "content": "\\mathrm{TemAttn}_{\\mathrm{self}}" + }, + { + "bbox": [ + 55, + 275, + 296, + 312 + ], + "type": "text", + "content": " denotes the Temporal Self-Attention module and " + }, + { + "bbox": [ + 55, + 275, + 296, + 312 + ], + "type": "inline_equation", + "content": "l_{\\mathrm{tem}}" + }, + { + "bbox": [ + 55, + 275, + 296, + 312 + ], + "type": "text", + "content": " is the module number corresponding to the four stages of the encoder (see Fig. 2)." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 55, + 318, + 296, + 467 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 318, + 296, + 467 + ], + "spans": [ + { + "bbox": [ + 55, + 318, + 296, + 467 + ], + "type": "text", + "content": "Decoder. The final decoder is responsible for predicting the target trajectory in the current and next two frames based on the updated target queries " + }, + { + "bbox": [ + 55, + 318, + 296, + 467 + ], + "type": "inline_equation", + "content": "\\mathbf{Q}_{\\mathrm{tgt}}^{t:t + 2}" + }, + { + "bbox": [ + 55, + 318, + 296, + 467 + ], + "type": "text", + "content": " and target-aware observation feature of the search region " + }, + { + "bbox": [ + 55, + 318, + 296, + 467 + ], + "type": "inline_equation", + "content": "\\hat{\\mathbf{F}}_{\\mathrm{x}}^{t:t + 2}" + }, + { + "bbox": [ + 55, + 318, + 296, + 467 + ], + "type": "text", + "content": ", which is split from the encoder output " + }, + { + "bbox": [ + 55, + 318, + 296, + 467 + ], + "type": "inline_equation", + "content": "\\hat{\\mathbf{F}}_{\\mathrm{zx}}^{t:t + 2}" + }, + { + "bbox": [ + 55, + 318, + 296, + 467 + ], + "type": "text", + "content": ". Notably, the dreamed feature of the template " + }, + { + "bbox": [ + 55, + 318, + 296, + 467 + ], + "type": "inline_equation", + "content": "\\hat{\\mathbf{F}}_{\\mathrm{z}}^{t:t + 2}" + }, + { + "bbox": [ + 55, + 318, + 296, + 467 + ], + "type": "text", + "content": " is not employed as the ultimate is to locate the target in the search region and the target information has been introduced into " + }, + { + "bbox": [ + 55, + 318, + 296, + 467 + ], + "type": "inline_equation", + "content": "\\hat{\\mathbf{F}}_{\\mathrm{x}}^{t:t + 2}" + }, + { + "bbox": [ + 55, + 318, + 296, + 467 + ], + "type": "text", + "content": " after encoder modeling. The target queries " + }, + { + "bbox": [ + 55, + 318, + 296, + 467 + ], + "type": "inline_equation", + "content": "\\mathbf{Q}_{\\mathrm{tgt}}^{t:t + 2}" + }, + { + "bbox": [ + 55, + 318, + 296, + 467 + ], + "type": "text", + "content": " first extract localization-relevant messages from " + }, + { + "bbox": [ + 55, + 318, + 296, + 467 + ], + "type": "inline_equation", + "content": "\\hat{\\mathbf{F}}_{\\mathrm{x}}^{t:t + 2}" + }, + { + "bbox": [ + 55, + 318, + 296, + 467 + ], + "type": "text", + "content": " with the Temporal Cross-Attention module. As shown in Fig. 4, each target query interacts with the region feature of the same timestamp," + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 61, + 470, + 294, + 487 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 61, + 470, + 294, + 487 + ], + "spans": [ + { + "bbox": [ + 61, + 470, + 294, + 487 + ], + "type": "interline_equation", + "content": "\\hat {\\mathbf {Q}} _ {\\mathrm {t g t}} ^ {t: t + 2} = \\operatorname {T e m A t t n} _ {\\text {c r o s s}} (q = \\mathbf {Q} _ {\\mathrm {t g t}} ^ {t: t + 2}, k v = \\hat {\\mathbf {F}} _ {\\mathrm {x}} ^ {t: t + 2}), \\tag {10}", + "image_path": "20644ac1a72f49d3a5a501bd77d0ebae3d4ff2d339a47a226d9ccd81cd07b0e2.jpg" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 55, + 491, + 295, + 527 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 491, + 295, + 527 + ], + "spans": [ + { + "bbox": [ + 55, + 491, + 295, + 527 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 55, + 491, + 295, + 527 + ], + "type": "inline_equation", + "content": "\\mathrm{TemAttn}_{\\mathrm{cross}}" + }, + { + "bbox": [ + 55, + 491, + 295, + 527 + ], + "type": "text", + "content": " is the Temporal Cross-Attention module. The output target queries " + }, + { + "bbox": [ + 55, + 491, + 295, + 527 + ], + "type": "inline_equation", + "content": "\\hat{\\mathbf{Q}}_{\\mathrm{tgt}}^{t:t + 2}" + }, + { + "bbox": [ + 55, + 491, + 295, + 527 + ], + "type": "text", + "content": " are then used to predict the target trajectory of the current and next two frames." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 55, + 528, + 296, + 694 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 528, + 296, + 694 + ], + "spans": [ + { + "bbox": [ + 55, + 528, + 296, + 694 + ], + "type": "text", + "content": "The key challenge to future prediction is the high degree of uncertainty, in large part due to not knowing the intents and latent characteristics of the tracking target. For example, the target commonly has a multimodal distribution of futures: it could turn, go straight, slow down, speed up, etc. The high degree of multimodality also exists in the motion prediction of autonomous driving [17, 36], which correspondingly designs a multimodal prediction to generate multiple plausible trajectories of the target vehicle. We draw inspiration from this solution and design the tailored Multimodal Prediction to our DreamTrack. As shown in Fig. 5, each target query of one timestamp forecasts the probabilities " + }, + { + "bbox": [ + 55, + 528, + 296, + 694 + ], + "type": "inline_equation", + "content": "\\hat{\\mathbf{p}}^T\\in \\mathbb{R}^{3\\times 1}" + }, + { + "bbox": [ + 55, + 528, + 296, + 694 + ], + "type": "text", + "content": " and target states " + }, + { + "bbox": [ + 55, + 528, + 296, + 694 + ], + "type": "inline_equation", + "content": "\\hat{\\mathbf{b}}^T\\in \\mathbb{R}^{3\\times 4}" + }, + { + "bbox": [ + 55, + 528, + 296, + 694 + ], + "type": "text", + "content": " of three possible future situations with a stack of MLP layers," + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 74, + 699, + 295, + 715 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 74, + 699, + 295, + 715 + ], + "spans": [ + { + "bbox": [ + 74, + 699, + 295, + 715 + ], + "type": "interline_equation", + "content": "\\hat {\\mathrm {p}} ^ {T}, \\hat {\\mathrm {b}} ^ {T} = \\operatorname {M L P} \\left(\\hat {\\mathbf {Q}} _ {\\mathrm {t g t}} ^ {T}\\right). T \\in \\{t, t + 1, t + 2 \\} \\tag {11}", + "image_path": "02dc68d4d4177969406704c41ef7ec4d4eedc08c7b12ce48d0b1a588c4c27114.jpg" + } + ] + } + ], + "index": 7 + }, + { + "type": "image", + "bbox": [ + 315, + 70, + 553, + 186 + ], + "blocks": [ + { + "bbox": [ + 315, + 70, + 553, + 186 + ], + "lines": [ + { + "bbox": [ + 315, + 70, + 553, + 186 + ], + "spans": [ + { + "bbox": [ + 315, + 70, + 553, + 186 + ], + "type": "image", + "image_path": "4360af44e38dd4565c9f5f0537d5afcc39942769629616cd044483b759b6afee.jpg" + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 313, + 187, + 553, + 232 + ], + "lines": [ + { + "bbox": [ + 313, + 187, + 553, + 232 + ], + "spans": [ + { + "bbox": [ + 313, + 187, + 553, + 232 + ], + "type": "text", + "content": "Figure 5. Illustration of the Multimodal Prediction. The target queries of different timestamps respectively predict the probabilities and locations of three possible future situations, in which the one of the highest probability score is the final output." + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "image_caption" + } + ], + "index": 8 + }, + { + "bbox": [ + 313, + 242, + 555, + 351 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 242, + 555, + 351 + ], + "spans": [ + { + "bbox": [ + 313, + 242, + 555, + 351 + ], + "type": "text", + "content": "After the prediction, the target state with the highest probability of each timestamp constitutes the final target trajectory. For training, we first calculate GIoU [42] between the predicted target state of each modality " + }, + { + "bbox": [ + 313, + 242, + 555, + 351 + ], + "type": "inline_equation", + "content": "\\hat{\\mathbf{b}}_j^T\\in \\hat{\\mathbf{b}}^T (1\\leq j\\leq 3)" + }, + { + "bbox": [ + 313, + 242, + 555, + 351 + ], + "type": "text", + "content": " and the ground-truth " + }, + { + "bbox": [ + 313, + 242, + 555, + 351 + ], + "type": "inline_equation", + "content": "\\mathrm{b}^T" + }, + { + "bbox": [ + 313, + 242, + 555, + 351 + ], + "type": "text", + "content": ". The prediction with the largest GIoU is assigned with the probability label of 1 (the other two ones are 0), and L1 loss " + }, + { + "bbox": [ + 313, + 242, + 555, + 351 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{\\mathrm{L1}}" + }, + { + "bbox": [ + 313, + 242, + 555, + 351 + ], + "type": "text", + "content": " and GIoU loss " + }, + { + "bbox": [ + 313, + 242, + 555, + 351 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{\\mathrm{GIoU}}" + }, + { + "bbox": [ + 313, + 242, + 555, + 351 + ], + "type": "text", + "content": " are applied to the output target states. The probabilities are supervised by the 0-1 labels with a cross-entropy loss " + }, + { + "bbox": [ + 313, + 242, + 555, + 351 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{\\mathrm{ce}}" + }, + { + "bbox": [ + 313, + 242, + 555, + 351 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 313, + 357, + 444, + 369 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 357, + 444, + 369 + ], + "spans": [ + { + "bbox": [ + 313, + 357, + 444, + 369 + ], + "type": "text", + "content": "3.3. Training and Inference" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 313, + 374, + 554, + 505 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 374, + 554, + 505 + ], + "spans": [ + { + "bbox": [ + 313, + 374, + 554, + 505 + ], + "type": "text", + "content": "Training. Our DreamTrack performs video sequence-level training and facilitates environmental understanding in an end-to-end manner. We construct video clips in which each contains 4 frames for training (i.e., one frame as template and three ones as search sequence), following the protocol of coordinate mapping described in Sec. 3.1. For every new input of the video clip, the hidden state in GRU (see Sec. 3.2) will be initialized, which avoids being distracted by the last tracking video. Considering the consistent time step between training and inference, we sample consecutive three frames for the search sequence." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 313, + 506, + 553, + 554 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 506, + 553, + 554 + ], + "spans": [ + { + "bbox": [ + 313, + 506, + 553, + 554 + ], + "type": "text", + "content": "The overall optimization objective for DreamTrack includes dreaming loss " + }, + { + "bbox": [ + 313, + 506, + 553, + 554 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{\\mathrm{drm}}" + }, + { + "bbox": [ + 313, + 506, + 553, + 554 + ], + "type": "text", + "content": ", reconstruction loss " + }, + { + "bbox": [ + 313, + 506, + 553, + 554 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{\\mathrm{rect}}" + }, + { + "bbox": [ + 313, + 506, + 553, + 554 + ], + "type": "text", + "content": " from the Future Dreaming module, and probability loss " + }, + { + "bbox": [ + 313, + 506, + 553, + 554 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{\\mathrm{ce}}" + }, + { + "bbox": [ + 313, + 506, + 553, + 554 + ], + "type": "text", + "content": ", L1 loss " + }, + { + "bbox": [ + 313, + 506, + 553, + 554 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{\\mathrm{L1}}" + }, + { + "bbox": [ + 313, + 506, + 553, + 554 + ], + "type": "text", + "content": " and GIoU loss " + }, + { + "bbox": [ + 313, + 506, + 553, + 554 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{\\mathrm{GIoU}}" + }, + { + "bbox": [ + 313, + 506, + 553, + 554 + ], + "type": "text", + "content": " from the Multimodal Prediction," + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 321, + 559, + 553, + 572 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 321, + 559, + 553, + 572 + ], + "spans": [ + { + "bbox": [ + 321, + 559, + 553, + 572 + ], + "type": "interline_equation", + "content": "\\mathcal {L} = \\lambda_ {1} \\mathcal {L} _ {\\mathrm {d r m}} + \\lambda_ {2} \\mathcal {L} _ {\\mathrm {r c t}} + \\lambda_ {3} \\mathcal {L} _ {\\mathrm {c e}} + \\lambda_ {4} \\mathcal {L} _ {\\mathrm {L 1}} + \\lambda_ {5} \\mathcal {L} _ {\\mathrm {G l o U}}, \\tag {12}", + "image_path": "38e56cf13aec00ee8be672c1689897c717dd1d3d11b803ba25edbb1ec8f3ca38.jpg" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 313, + 576, + 528, + 588 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 576, + 528, + 588 + ], + "spans": [ + { + "bbox": [ + 313, + 576, + 528, + 588 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 313, + 576, + 528, + 588 + ], + "type": "inline_equation", + "content": "\\lambda" + }, + { + "bbox": [ + 313, + 576, + 528, + 588 + ], + "type": "text", + "content": " values are the balancing weight coefficients." + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 313, + 594, + 555, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 594, + 555, + 713 + ], + "spans": [ + { + "bbox": [ + 313, + 594, + 555, + 713 + ], + "type": "text", + "content": "Inference. During inference, we initialize the hidden state of GRU for each new video sequence as previously described. Notably, it's no need to sample the video clip for inference. With the defined target queries that automatically cache the historical dynamics, we perform frame-level input for DreamTrack, which is consistent with previous nontemporal works [9, 51] and friendly for the deployment. Besides, the distribution mapping and image reconstruction in Fig. 3 are not required in the inference phase, which could be extended to other tasks, e.g., video generation." + } + ] + } + ], + "index": 16 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "text", + "content": "7205" + } + ] + } + ], + "index": 17 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 4 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 69, + 70, + 541, + 332 + ], + "blocks": [ + { + "bbox": [ + 69, + 70, + 541, + 332 + ], + "lines": [ + { + "bbox": [ + 69, + 70, + 541, + 332 + ], + "spans": [ + { + "bbox": [ + 69, + 70, + 541, + 332 + ], + "type": "table", + "html": "
MethodsLaSOT [15]LaSOText [16]GOT-10k* [27]TrackingNet [39]TNL2K [46]
SUC(%)PNorm(%)P(%)SUC(%)PNorm(%)P(%)AO(%)SR0.5(%)SR0.75(%)SUC(%)PNorm(%)P(%)SUC(%)P(%)
SiamFC255[2]33.642.033.923.031.126.934.835.39.857.166.353.329.528.6
ECO224[10]32.433.830.122.025.224.031.630.911.155.461.849.232.631.7
SiamRPN++255[32]49.656.949.134.041.639.651.761.632.573.380.069.441.341.2
DiMP288[3]56.965.056.739.247.645.161.171.749.274.080.168.744.743.4
SiamRCNN255[44]64.872.2----64.972.859.781.285.480.052.352.8
Ocean255[53]56.065.156.6---61.172.147.3---38.438.7
AutoMatch255[54]58.3-59.937.6-43.065.276.654.376.0-72.647.243.5
TrDiMP352[45]63.9-61.4---67.177.758.378.483.373.1--
TransT256[6]64.973.869.044.852.352.567.176.860.981.486.780.350.751.7
STARK320[50]67.177.0-47.856.255.168.878.164.182.086.9---
SwinTrack-B384[33]71.3-76.549.1-55.672.480.567.884.0-82.855.957.1
MixFormer-L320[9]70.179.976.3---72.682.268.883.988.983.1--
OSTrack384[51]71.181.177.650.561.357.673.783.270.883.988.583.255.956.7
GRM256[19]69.979.375.8---73.482.970.484.088.783.3--
MixViT288[9]69.679.975.9---72.582.469.983.588.382.0--
ROMTrack384[4]71.481.478.251.362.458.674.284.372.484.189.083.7--
MixViT-L384[9]72.482.280.1---75.785.375.185.490.285.7--
SeqTrack-L384[7]72.581.579.350.761.657.574.881.972.285.589.885.857.8-
ARTrack256[47]70.479.576.646.456.552.373.582.270.984.288.783.557.5-
ARTrack-L384[47]73.182.280.352.862.959.778.587.477.885.689.686.060.3-
ARTrackV2256[1]71.680.277.250.861.957.775.985.472.784.989.384.559.2-
ARTrackV2-L384[1]73.682.881.153.463.760.279.587.879.686.190.486.261.6-
LoRAT224[34]71.780.977.350.361.657.172.181.870.783.587.982.158.861.3
LoRAT-L378[34]75.184.182.056.669.065.177.586.278.185.689.785.462.367.0
DreamTrack25673.883.480.653.164.159.877.587.174.285.890.085.360.463.2
DreamTrack38475.084.281.754.565.361.178.387.976.686.590.685.961.264.8
DreamTrack-L38476.685.683.157.169.465.479.988.580.387.991.386.862.967.4
", + "image_path": "7f3a3000d974339bda5adb640005ef0aa20f29e8cd6d81fd374dd553487279c2.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "table_body" + } + ], + "index": 0 + }, + { + "bbox": [ + 55, + 335, + 555, + 360 + ], + "lines": [ + { + "bbox": [ + 55, + 335, + 555, + 360 + ], + "spans": [ + { + "bbox": [ + 55, + 335, + 555, + 360 + ], + "type": "text", + "content": "Table 1. State-of-the-art comparison on LaSOT [15], LaSOText [16], GOT-10k [27], TrackingNet [39] and TNL2K [46]. * indicates only trained on GOT-10k. The number in the subscript denotes the search region resolution." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "text" + }, + { + "bbox": [ + 55, + 368, + 133, + 381 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 368, + 133, + 381 + ], + "spans": [ + { + "bbox": [ + 55, + 368, + 133, + 381 + ], + "type": "text", + "content": "4. Experiment" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 55, + 388, + 188, + 400 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 388, + 188, + 400 + ], + "spans": [ + { + "bbox": [ + 55, + 388, + 188, + 400 + ], + "type": "text", + "content": "4.1. Implementation Details" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 55, + 405, + 295, + 429 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 405, + 295, + 429 + ], + "spans": [ + { + "bbox": [ + 55, + 405, + 295, + 429 + ], + "type": "text", + "content": "Model Variants. We develop three variants of DreamTrack with different configurations as follows:" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 55, + 429, + 295, + 500 + ], + "type": "list", + "angle": 0, + "index": 8, + "blocks": [ + { + "bbox": [ + 55, + 429, + 295, + 452 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 429, + 295, + 452 + ], + "spans": [ + { + "bbox": [ + 55, + 429, + 295, + 452 + ], + "type": "text", + "content": "- DreamTrack256. Backbone: ViT-Base; Template size: " + }, + { + "bbox": [ + 55, + 429, + 295, + 452 + ], + "type": "inline_equation", + "content": "[128 \\times 128]" + }, + { + "bbox": [ + 55, + 429, + 295, + 452 + ], + "type": "text", + "content": "; Search region size: " + }, + { + "bbox": [ + 55, + 429, + 295, + 452 + ], + "type": "inline_equation", + "content": "[256 \\times 256]" + }, + { + "bbox": [ + 55, + 429, + 295, + 452 + ], + "type": "text", + "content": ";" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 56, + 453, + 295, + 476 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 453, + 295, + 476 + ], + "spans": [ + { + "bbox": [ + 56, + 453, + 295, + 476 + ], + "type": "text", + "content": "- DreamTrack384. Backbone: ViT-Base; Template size: " + }, + { + "bbox": [ + 56, + 453, + 295, + 476 + ], + "type": "inline_equation", + "content": "[192 \\times 192]" + }, + { + "bbox": [ + 56, + 453, + 295, + 476 + ], + "type": "text", + "content": "; Search region size: " + }, + { + "bbox": [ + 56, + 453, + 295, + 476 + ], + "type": "inline_equation", + "content": "[384 \\times 384]" + }, + { + "bbox": [ + 56, + 453, + 295, + 476 + ], + "type": "text", + "content": ";" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 55, + 476, + 295, + 500 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 476, + 295, + 500 + ], + "spans": [ + { + "bbox": [ + 55, + 476, + 295, + 500 + ], + "type": "text", + "content": "- DreamTrack-L384. Backbone: ViT-Large; Template size: " + }, + { + "bbox": [ + 55, + 476, + 295, + 500 + ], + "type": "inline_equation", + "content": "[192 \\times 192]" + }, + { + "bbox": [ + 55, + 476, + 295, + 500 + ], + "type": "text", + "content": "; Search region size: " + }, + { + "bbox": [ + 55, + 476, + 295, + 500 + ], + "type": "inline_equation", + "content": "[384 \\times 384]" + }, + { + "bbox": [ + 55, + 476, + 295, + 500 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 7 + } + ], + "sub_type": "text" + }, + { + "bbox": [ + 55, + 500, + 295, + 561 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 500, + 295, + 561 + ], + "spans": [ + { + "bbox": [ + 55, + 500, + 295, + 561 + ], + "type": "text", + "content": "The ViT backbone [13] is regarded as the encoder of our DreamTrack (see Fig. 2), which is divided into four stages for feature modeling with the proposed Temporal Self-Attention. The four stages contain 2,2,6,2 encoder layers for ViT-Base, and 4,4,12,4 ones for ViT-Large, respectively." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 55, + 570, + 295, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 570, + 295, + 714 + ], + "spans": [ + { + "bbox": [ + 55, + 570, + 295, + 714 + ], + "type": "text", + "content": "Training Strategy. The training setting and evaluation protocol follow previous Transformer-based trackers (e.g., OSTrack [51]). The training datasets include COCO [35], LaSOT [15], TrackingNet [39] and GOT-10k [27] (with 1k sequences removed from GOT-10k train split, as per [50]). To ensure a fair evaluation of the GOT-10k test set, our models learn from the entire GOT-10k training split following its one-shot protocol. Our DreamTrack performs training with video clips, as described in Sec. 3.3. To align with the data form in the image dataset COCO that performs frame-level sampling, we repeat the sampled search frame with random shifts to constitute a pseudo video clip." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 313, + 369, + 555, + 453 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 369, + 555, + 453 + ], + "spans": [ + { + "bbox": [ + 313, + 369, + 555, + 453 + ], + "type": "text", + "content": "We train our model using the AdamW [37] optimizer with a weight decay of " + }, + { + "bbox": [ + 313, + 369, + 555, + 453 + ], + "type": "inline_equation", + "content": "1 \\times 10^{-4}" + }, + { + "bbox": [ + 313, + 369, + 555, + 453 + ], + "type": "text", + "content": ". The learning rate for the encoder is set to " + }, + { + "bbox": [ + 313, + 369, + 555, + 453 + ], + "type": "inline_equation", + "content": "4 \\times 10^{-5}" + }, + { + "bbox": [ + 313, + 369, + 555, + 453 + ], + "type": "text", + "content": ", while other parameters use a learning rate of " + }, + { + "bbox": [ + 313, + 369, + 555, + 453 + ], + "type": "inline_equation", + "content": "4 \\times 10^{-4}" + }, + { + "bbox": [ + 313, + 369, + 555, + 453 + ], + "type": "text", + "content": ". The training process is conducted on 8 NVIDIA RTX3090 GPUs, comprising 300 epochs with 60k video sequences per epoch. The learning rate decreases by a factor of 10 after 240 epochs." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 313, + 458, + 400, + 470 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 458, + 400, + 470 + ], + "spans": [ + { + "bbox": [ + 313, + 458, + 400, + 470 + ], + "type": "text", + "content": "4.2. Main Results" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 313, + 475, + 554, + 524 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 475, + 554, + 524 + ], + "spans": [ + { + "bbox": [ + 313, + 475, + 554, + 524 + ], + "type": "text", + "content": "We evaluate our models on eight prevailing benchmark datasets, including LaSOT [15], LaSOText [16], GOT-10k [27], TrackingNet [39], TNL2K [46], OTB100 [49], NFS [29] and UAV123 [38]." + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 313, + 529, + 555, + 625 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 529, + 555, + 625 + ], + "spans": [ + { + "bbox": [ + 313, + 529, + 555, + 625 + ], + "type": "text", + "content": "LaSOT [15]. LaSOT is a benchmark designed for long-term tracking, comprising 280 videos in its test set. Tab. 1 demonstrates that our DreamTrack" + }, + { + "bbox": [ + 313, + 529, + 555, + 625 + ], + "type": "inline_equation", + "content": "_{256}" + }, + { + "bbox": [ + 313, + 529, + 555, + 625 + ], + "type": "text", + "content": " already outperforms ARTrack-" + }, + { + "bbox": [ + 313, + 529, + 555, + 625 + ], + "type": "inline_equation", + "content": "L_{384}" + }, + { + "bbox": [ + 313, + 529, + 555, + 625 + ], + "type": "text", + "content": ", which has a larger backbone. Furthermore, our DreamTrack-" + }, + { + "bbox": [ + 313, + 529, + 555, + 625 + ], + "type": "inline_equation", + "content": "L_{384}" + }, + { + "bbox": [ + 313, + 529, + 555, + 625 + ], + "type": "text", + "content": " achieved the best performance with the SUC score of " + }, + { + "bbox": [ + 313, + 529, + 555, + 625 + ], + "type": "inline_equation", + "content": "76.6\\%" + }, + { + "bbox": [ + 313, + 529, + 555, + 625 + ], + "type": "text", + "content": ", without the use of any online template update strategy or post-processing. This proves the superiority of our temporal design with future dreaming." + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 313, + 629, + 556, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 629, + 556, + 715 + ], + "spans": [ + { + "bbox": [ + 313, + 629, + 556, + 715 + ], + "type": "text", + "content": "LaSOText [16]. LaSOText serves as an extension of LaSOT, which includes additional 150 video sequences. As shown in Tab. 1, our DreamTrack" + }, + { + "bbox": [ + 313, + 629, + 556, + 715 + ], + "type": "inline_equation", + "content": "_{256}" + }, + { + "bbox": [ + 313, + 629, + 556, + 715 + ], + "type": "text", + "content": " achieves superior performance compared with other trackers with the same backbone, e.g., ARTrackV2" + }, + { + "bbox": [ + 313, + 629, + 556, + 715 + ], + "type": "inline_equation", + "content": "_{256}" + }, + { + "bbox": [ + 313, + 629, + 556, + 715 + ], + "type": "text", + "content": " and LoRAT" + }, + { + "bbox": [ + 313, + 629, + 556, + 715 + ], + "type": "inline_equation", + "content": "_{224}" + }, + { + "bbox": [ + 313, + 629, + 556, + 715 + ], + "type": "text", + "content": ". With the larger backbone, the performance of our DreamTrack-L" + }, + { + "bbox": [ + 313, + 629, + 556, + 715 + ], + "type": "inline_equation", + "content": "_{384}" + }, + { + "bbox": [ + 313, + 629, + 556, + 715 + ], + "type": "text", + "content": " further improves with the best SUC score of " + }, + { + "bbox": [ + 313, + 629, + 556, + 715 + ], + "type": "inline_equation", + "content": "57.1\\%" + }, + { + "bbox": [ + 313, + 629, + 556, + 715 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 15 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 748, + 316, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 316, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 316, + 757 + ], + "type": "text", + "content": "7206" + } + ] + } + ], + "index": 16 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 5 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 82, + 70, + 269, + 213 + ], + "blocks": [ + { + "bbox": [ + 82, + 70, + 269, + 213 + ], + "lines": [ + { + "bbox": [ + 82, + 70, + 269, + 213 + ], + "spans": [ + { + "bbox": [ + 82, + 70, + 269, + 213 + ], + "type": "table", + "html": "
MethodOTB100 [49]NFS [29]UAV123 [38]
DreamTrack-L38472.467.473.1
DreamTrack38472.066.172.5
DreamTrack25671.865.472.1
LoRAT-L378 [34]72.066.772.5
LoRAT224 [34]71.464.072.6
ARTrackV2-L384 [1]-67.669.9
SeqTrack-L384 [7]-66.268.5
OSTrack384 [51]-66.570.7
TransT256 [6]69.565.769.1
DiMP288 [3]68.461.864.3
Ocean255 [53]68.449.457.4
ATOM288 [11]66.358.363.2
ECO224 [10]66.652.253.5
SiamFC255 [2]58.337.746.8
", + "image_path": "e0b73f55898f0d6051a23c65e2e478f14c10801d8816365c7c26ca7b581abaa9.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "table_body" + } + ], + "index": 0 + }, + { + "bbox": [ + 55, + 247, + 296, + 331 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 247, + 296, + 331 + ], + "spans": [ + { + "bbox": [ + 55, + 247, + 296, + 331 + ], + "type": "text", + "content": "GOT-10k [27]. GOT-10k provides 180 videos for testing with a strict protocol that only allows trackers to be trained exclusively on its training split. We adhere to this protocol and report the results in Tab. 1. Our DreamTrack-" + }, + { + "bbox": [ + 55, + 247, + 296, + 331 + ], + "type": "inline_equation", + "content": "\\mathrm{L}_{384}" + }, + { + "bbox": [ + 55, + 247, + 296, + 331 + ], + "type": "text", + "content": " outperforms recent SOTA trackers across all metrics. Notably, our DreamTrack" + }, + { + "bbox": [ + 55, + 247, + 296, + 331 + ], + "type": "inline_equation", + "content": "_{256}" + }, + { + "bbox": [ + 55, + 247, + 296, + 331 + ], + "type": "text", + "content": " surpasses other trackers with the same backbone, proving its effectiveness." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 55, + 337, + 296, + 422 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 337, + 296, + 422 + ], + "spans": [ + { + "bbox": [ + 55, + 337, + 296, + 422 + ], + "type": "text", + "content": "TrackingNet [39]. TrackingNet is an extensive tracking dataset that covers a wide range of real-world scenarios. We assess the performance of DreamTrack on its test set of 511 videos, as illustrated in Tab. 1. It shows that not only does our DreamTrack" + }, + { + "bbox": [ + 55, + 337, + 296, + 422 + ], + "type": "inline_equation", + "content": "_{384}" + }, + { + "bbox": [ + 55, + 337, + 296, + 422 + ], + "type": "text", + "content": " outperform all other trackers in SUC, but our DreamTrack-" + }, + { + "bbox": [ + 55, + 337, + 296, + 422 + ], + "type": "inline_equation", + "content": "L_{384}" + }, + { + "bbox": [ + 55, + 337, + 296, + 422 + ], + "type": "text", + "content": " also establishes a new state-of-the-art in three metrics on this large-scale benchmark." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 55, + 426, + 296, + 510 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 426, + 296, + 510 + ], + "spans": [ + { + "bbox": [ + 55, + 426, + 296, + 510 + ], + "type": "text", + "content": "TNL2K [46]. TNL2K is a recently proposed large-scale benchmark for tracking by natural language and bounding box initialization. We evaluate our tracker on its 700 testing videos and the results are listed in Tab. 1. Our DreamTrack-L384 achieves the best performance with the SUC and precision of " + }, + { + "bbox": [ + 55, + 426, + 296, + 510 + ], + "type": "inline_equation", + "content": "62.9\\%" + }, + { + "bbox": [ + 55, + 426, + 296, + 510 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 55, + 426, + 296, + 510 + ], + "type": "inline_equation", + "content": "67.4\\%" + }, + { + "bbox": [ + 55, + 426, + 296, + 510 + ], + "type": "text", + "content": " respectively, showing the effectiveness of our temporal learning." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 55, + 515, + 295, + 589 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 515, + 295, + 589 + ], + "spans": [ + { + "bbox": [ + 55, + 515, + 295, + 589 + ], + "type": "text", + "content": "OTB100 [49], NFS [29] and UAV123 [38]. We also evaluate our tracker on three additional benchmarks, i.e., OTB100, NFS and UAV123 including 100, 100 and 123 videos respectively. Tab. 2 shows that DreamTrack-" + }, + { + "bbox": [ + 55, + 515, + 295, + 589 + ], + "type": "inline_equation", + "content": "\\mathrm{L}_{384}" + }, + { + "bbox": [ + 55, + 515, + 295, + 589 + ], + "type": "text", + "content": " achieves the best performance on all three benchmarks, demonstrating the strong generalizability of our framework." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 55, + 594, + 295, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 594, + 295, + 715 + ], + "spans": [ + { + "bbox": [ + 55, + 594, + 295, + 715 + ], + "type": "text", + "content": "Efficiency Comparison. Tab. 3 compares the efficiency of our DreamTrack against other Transformer-based trackers in terms of inference speed, floating point operations, and number of parameters. Benefiting from the efficient one-stream architecture and sparse multimodal prediction, our DreamTrack shows impressive inference speed with small computation costs. Our most compact variant, DreamTrack" + }, + { + "bbox": [ + 55, + 594, + 295, + 715 + ], + "type": "inline_equation", + "content": "_{256}" + }, + { + "bbox": [ + 55, + 594, + 295, + 715 + ], + "type": "text", + "content": ", achieves the state-of-the-art performance while operating at 139 frames per second (FPS), proving the effectiveness and efficiency of our method." + } + ] + } + ], + "index": 6 + }, + { + "type": "table", + "bbox": [ + 340, + 70, + 531, + 189 + ], + "blocks": [ + { + "bbox": [ + 55, + 217, + 295, + 240 + ], + "lines": [ + { + "bbox": [ + 55, + 217, + 295, + 240 + ], + "spans": [ + { + "bbox": [ + 55, + 217, + 295, + 240 + ], + "type": "text", + "content": "Table 2. Performance comparison on OTB100 [49], NFS [29] and UAV123 [38] in terms of the SUC score " + }, + { + "bbox": [ + 55, + 217, + 295, + 240 + ], + "type": "inline_equation", + "content": "(\\%)" + }, + { + "bbox": [ + 55, + 217, + 295, + 240 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 340, + 70, + 531, + 189 + ], + "lines": [ + { + "bbox": [ + 340, + 70, + 531, + 189 + ], + "spans": [ + { + "bbox": [ + 340, + 70, + 531, + 189 + ], + "type": "table", + "html": "
TrackerSpeed (FPS)FLOPs (G)Params (M)
SwinTrack-B384 [33]4569.791
OSTrack256 [51]105.421.594
OSTrack384 [51]58.148.394
SeqTrack-B256 [7]406689
SeqTrack-L384 [7]5524309
ARTrack256 [47]26395.6117
ROMTrack256 [4]6234.592
ROMTrack384 [4]2877.792
LoRAT224 [34]2093099
LoRAT-L378 [34]63325336
DreamTrack25613937.687
DreamTrack-L38432359.4317
", + "image_path": "e543e4f843122d63975c181448d6bf1f63df8ac7bd2b0fb8b63d4bbc189f47a7.jpg" + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "table_body" + } + ], + "index": 7 + }, + { + "type": "table", + "bbox": [ + 317, + 223, + 553, + 291 + ], + "blocks": [ + { + "bbox": [ + 313, + 193, + 555, + 215 + ], + "lines": [ + { + "bbox": [ + 313, + 193, + 555, + 215 + ], + "spans": [ + { + "bbox": [ + 313, + 193, + 555, + 215 + ], + "type": "text", + "content": "Table 3. Efficiency comparison. The inference speed is measured on a single NVIDIA RTX2080Ti GPU." + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 317, + 223, + 553, + 291 + ], + "lines": [ + { + "bbox": [ + 317, + 223, + 553, + 291 + ], + "spans": [ + { + "bbox": [ + 317, + 223, + 553, + 291 + ], + "type": "table", + "html": "
#Model VariantsLaSOT [15]TNL2K [46]FPS
SUC(%)PNorm(%)P(%)SUC(%)P(%)
①Baseline69.379.175.354.756.2162
②+ Future Dreaming Module71.781.377.756.959.5153
③+ Temporal Self-Attention72.682.078.559.160.4145
④+ Temporal Cross-Attention73.182.479.059.561.1143
⑤+ Multimodal Prediction73.883.480.660.463.2139
", + "image_path": "25e8051964c7dd2b8f88ed1e62c5d7bc714eed54fd705889b3b180c92ba4f146.jpg" + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "table_body" + } + ], + "index": 9 + }, + { + "bbox": [ + 324, + 295, + 543, + 306 + ], + "lines": [ + { + "bbox": [ + 324, + 295, + 543, + 306 + ], + "spans": [ + { + "bbox": [ + 324, + 295, + 543, + 306 + ], + "type": "text", + "content": "Table 4. Ablation on each proposed module in DreamTrack." + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "text" + }, + { + "bbox": [ + 313, + 319, + 460, + 332 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 319, + 460, + 332 + ], + "spans": [ + { + "bbox": [ + 313, + 319, + 460, + 332 + ], + "type": "text", + "content": "4.3. Component-wise Ablation" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 313, + 337, + 555, + 554 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 337, + 555, + 554 + ], + "spans": [ + { + "bbox": [ + 313, + 337, + 555, + 554 + ], + "type": "text", + "content": "We analyze the influence of each component in our DreamTrack" + }, + { + "bbox": [ + 313, + 337, + 555, + 554 + ], + "type": "inline_equation", + "content": "_{256}" + }, + { + "bbox": [ + 313, + 337, + 555, + 554 + ], + "type": "text", + "content": " to show the effectiveness and rationality of the proposed modules, as shown in Tab. 4. When removing all designed modules and directly predicting the target location of the current frame based on the target query, the performance is still comparable with the fastest inference speed of 162 FPS. By applying the key Future Dreaming module on the baseline version, it obtains impressive SUC gains of " + }, + { + "bbox": [ + 313, + 337, + 555, + 554 + ], + "type": "inline_equation", + "content": "2.4\\%" + }, + { + "bbox": [ + 313, + 337, + 555, + 554 + ], + "type": "text", + "content": " on LaSOT (2 v.s. 1). This verifies that deeper environmental understanding from past experience benefits the generalizability in novel situations. The Temporal Self/Cross-Attention further improves the performance for " + }, + { + "bbox": [ + 313, + 337, + 555, + 554 + ], + "type": "inline_equation", + "content": "0.9\\% / 0.5\\%" + }, + { + "bbox": [ + 313, + 337, + 555, + 554 + ], + "type": "text", + "content": " SUC on LaSOT (4 v.s. 3 v.s. 2), proving the effectiveness of target-aware temporal refinement. The final Multimodal Prediction brings " + }, + { + "bbox": [ + 313, + 337, + 555, + 554 + ], + "type": "inline_equation", + "content": "0.7\\% / 0.9\\%" + }, + { + "bbox": [ + 313, + 337, + 555, + 554 + ], + "type": "text", + "content": " SUC gains on LaSOT and TNL2K, respectively. It indicates that modeling the uncertainty of different future scenarios helps to forecast the target's motion and perform robust tracking." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 313, + 563, + 417, + 576 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 563, + 417, + 576 + ], + "spans": [ + { + "bbox": [ + 313, + 563, + 417, + 576 + ], + "type": "text", + "content": "4.4. Further Analysis" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 313, + 582, + 555, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 582, + 555, + 715 + ], + "spans": [ + { + "bbox": [ + 313, + 582, + 555, + 715 + ], + "type": "text", + "content": "Self-supervision in Future Dreaming. As mentioned in Sec. 3.2, the capability of predicting the environmental future states relies on the self-supervision in the proposed Future Dreaming module. We explore the influence by ablating the loss functions based on " + }, + { + "bbox": [ + 313, + 582, + 555, + 715 + ], + "type": "inline_equation", + "content": "\\mathrm{DreamTrack}_{256}" + }, + { + "bbox": [ + 313, + 582, + 555, + 715 + ], + "type": "text", + "content": ". As shown in Tab. 5, removing both the dreaming and reconstruction losses makes the dreamed future unreliable to reflect the real scenario in the next two frames, degrading the quality of learned temporal dynamics (①). Applying each of the two loss functions brings performance gains of " + }, + { + "bbox": [ + 313, + 582, + 555, + 715 + ], + "type": "inline_equation", + "content": "4.4\\% / 4.0\\%" + }, + { + "bbox": [ + 313, + 582, + 555, + 715 + ], + "type": "text", + "content": " SUC on LaSOT respectively (②,③ v.s. ①), proving the ef" + } + ] + } + ], + "index": 14 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "text", + "content": "7207" + } + ] + } + ], + "index": 15 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 6 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 69, + 70, + 281, + 128 + ], + "blocks": [ + { + "bbox": [ + 69, + 70, + 281, + 128 + ], + "lines": [ + { + "bbox": [ + 69, + 70, + 281, + 128 + ], + "spans": [ + { + "bbox": [ + 69, + 70, + 281, + 128 + ], + "type": "table", + "html": "
#Dreaming LossReconstruction LossLaSOT [15]TNL2K [46]
SUC(%)PNorm(%)P(%)SUC(%)P(%)
①--68.377.473.853.753.2
②✓-72.782.679.859.161.7
③-✓72.382.179.258.560.3
④✓✓73.883.480.660.463.2
", + "image_path": "2b7a13a23ddbdf4090df6471d23cf536367d954e8b0ea6a98b8cfbdea387c142.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "table_body" + } + ], + "index": 0 + }, + { + "type": "table", + "bbox": [ + 69, + 150, + 281, + 204 + ], + "blocks": [ + { + "bbox": [ + 59, + 132, + 291, + 143 + ], + "lines": [ + { + "bbox": [ + 59, + 132, + 291, + 143 + ], + "spans": [ + { + "bbox": [ + 59, + 132, + 291, + 143 + ], + "type": "text", + "content": "Table 5. Ablation on loss functions in Future Dreaming module." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 69, + 150, + 281, + 204 + ], + "lines": [ + { + "bbox": [ + 69, + 150, + 281, + 204 + ], + "spans": [ + { + "bbox": [ + 69, + 150, + 281, + 204 + ], + "type": "table", + "html": "
#Dreaming Step For EvaluationLaSOT [15]LaSOText [16]TNL2K [46]
SUC(%)P(%)SUC(%)P(%)SUC(%)P(%)
①073.880.653.159.860.463.2
②171.377.551.458.258.661.6
③268.574.748.955.055.357.8
", + "image_path": "64f99d191aab67809aea80d2ad9af5d5a2d1100286badfec1c09245431167a72.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "table_body" + } + ], + "index": 2 + }, + { + "bbox": [ + 63, + 208, + 288, + 219 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 63, + 208, + 288, + 219 + ], + "spans": [ + { + "bbox": [ + 63, + 208, + 288, + 219 + ], + "type": "text", + "content": "Table 6. Performance comparison of different dreaming steps." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 55, + 227, + 296, + 264 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 227, + 296, + 264 + ], + "spans": [ + { + "bbox": [ + 55, + 227, + 296, + 264 + ], + "type": "text", + "content": "fectiveness. With the full supervision, the performance of DreamTrack" + }, + { + "bbox": [ + 55, + 227, + 296, + 264 + ], + "type": "inline_equation", + "content": "_{256}" + }, + { + "bbox": [ + 55, + 227, + 296, + 264 + ], + "type": "text", + "content": " further improves with the SUC of " + }, + { + "bbox": [ + 55, + 227, + 296, + 264 + ], + "type": "inline_equation", + "content": "73.8\\%" + }, + { + "bbox": [ + 55, + 227, + 296, + 264 + ], + "type": "text", + "content": " on LaSOT (4), which indicates the cooperation of our design." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 55, + 273, + 296, + 477 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 273, + 296, + 477 + ], + "spans": [ + { + "bbox": [ + 55, + 273, + 296, + 477 + ], + "type": "text", + "content": "Performance of Future Prediction. With the defined target queries of different timestamps, our DreamTrack is capable of predicting the target trajectory including the current and next two frames. We then evaluate the performance of the future predictions by regarding them as the results, e.g., the future prediction for frame 10 from current frame 8 is used for the metric computation. As shown in Tab. 6, the performance of future prediction is inferior to current prediction (2,3 v.s. 1), and the tracking metrics degrade more with the prediction of further future. The underlying reason is that the future observation is unknowable for current prediction, in which the high degree of uncertainty raises challenges to future forecasting. In contrast, the previous trackers only predict the present target location with current observation. Despite these, the comparable performance demonstrates the potential of our design in the motion prediction of the target and benefits visual tracking." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 55, + 486, + 295, + 641 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 486, + 295, + 641 + ], + "spans": [ + { + "bbox": [ + 55, + 486, + 295, + 641 + ], + "type": "text", + "content": "Activation Analysis and Result Visualization. Future dreaming endows our DreamTrack with superior generalization capability in novel situations. For deeper insights into the mechanism, we first visualize the activation maps " + }, + { + "bbox": [ + 55, + 486, + 295, + 641 + ], + "type": "inline_equation", + "content": "w" + }, + { + "bbox": [ + 55, + 486, + 295, + 641 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 55, + 486, + 295, + 641 + ], + "type": "inline_equation", + "content": "w / o" + }, + { + "bbox": [ + 55, + 486, + 295, + 641 + ], + "type": "text", + "content": ". The Future Dreaming module. As shown in Fig. 6, the attention areas without the future dreaming are easily distracted by the target's appearance deformation and background clutters, finally losing the target. This proves our claim that insufficient environmental understanding hardly handles the challenges in long-term tracking. In contrast, the version with future dreaming successfully locates the target area with the learned dynamics from past experience, showing the effectiveness of our design." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 55, + 642, + 296, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 642, + 296, + 713 + ], + "spans": [ + { + "bbox": [ + 55, + 642, + 296, + 713 + ], + "type": "text", + "content": "We also visualize the results of different trackers, as well as the dreamed future frames. As shown in Fig. 7, the proposed DreamTrack" + }, + { + "bbox": [ + 55, + 642, + 296, + 713 + ], + "type": "inline_equation", + "content": "_{256}" + }, + { + "bbox": [ + 55, + 642, + 296, + 713 + ], + "type": "text", + "content": " delivers more robust tracking under deformation (the first row), occlusion and similar interferences (the last two rows). It demonstrates the superiority of our temporal learning in benefiting from both future and" + } + ] + } + ], + "index": 7 + }, + { + "type": "image", + "bbox": [ + 315, + 71, + 553, + 225 + ], + "blocks": [ + { + "bbox": [ + 315, + 71, + 553, + 225 + ], + "lines": [ + { + "bbox": [ + 315, + 71, + 553, + 225 + ], + "spans": [ + { + "bbox": [ + 315, + 71, + 553, + 225 + ], + "type": "image", + "image_path": "44de60cf934d0d237447725bd03730110fddc3ca38ff9a3273de995db2a39023.jpg" + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 313, + 229, + 555, + 262 + ], + "lines": [ + { + "bbox": [ + 313, + 229, + 555, + 262 + ], + "spans": [ + { + "bbox": [ + 313, + 229, + 555, + 262 + ], + "type": "text", + "content": "Figure 6. Attention visualization of our DreamTrack " + }, + { + "bbox": [ + 313, + 229, + 555, + 262 + ], + "type": "inline_equation", + "content": "w/" + }, + { + "bbox": [ + 313, + 229, + 555, + 262 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 313, + 229, + 555, + 262 + ], + "type": "inline_equation", + "content": "w/o" + }, + { + "bbox": [ + 313, + 229, + 555, + 262 + ], + "type": "text", + "content": ". the Future Dreaming module. The predictions and ground-truth boxes are marked in red and green, respectively." + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "image_caption" + } + ], + "index": 8 + }, + { + "type": "image", + "bbox": [ + 316, + 269, + 553, + 399 + ], + "blocks": [ + { + "bbox": [ + 316, + 269, + 553, + 399 + ], + "lines": [ + { + "bbox": [ + 316, + 269, + 553, + 399 + ], + "spans": [ + { + "bbox": [ + 316, + 269, + 553, + 399 + ], + "type": "image", + "image_path": "d4712ebe160d156b9a462f42a1bb983b1b91303d6d6a8bff241e48e1845f00aa.jpg" + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 313, + 402, + 555, + 425 + ], + "lines": [ + { + "bbox": [ + 313, + 402, + 555, + 425 + ], + "spans": [ + { + "bbox": [ + 313, + 402, + 555, + 425 + ], + "type": "text", + "content": "Figure 7. Results visualization of different trackers and dreamed future frames based on learned environmental dynamics." + } + ] + } + ], + "index": 11, + "angle": 0, + "type": "image_caption" + } + ], + "index": 10 + }, + { + "bbox": [ + 313, + 431, + 555, + 491 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 431, + 555, + 491 + ], + "spans": [ + { + "bbox": [ + 313, + 431, + 555, + 491 + ], + "type": "text", + "content": "historical information. Notably, the dreamed future frames correctly indicate the evolution of the complex environment, and the predicted trajectory successfully locates the tracking target. This indicates the potential of our method to be developed into video generation and annotation tasks." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 313, + 491, + 554, + 539 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 491, + 554, + 539 + ], + "spans": [ + { + "bbox": [ + 313, + 491, + 554, + 539 + ], + "type": "text", + "content": "Due to limited space, we present more analyses in the supplementary material, including 1) more challenging benchmarks, 2) generalization in different paradigms, 3) position of future dreaming, 4) more visualizations, etc." + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 313, + 550, + 388, + 562 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 550, + 388, + 562 + ], + "spans": [ + { + "bbox": [ + 313, + 550, + 388, + 562 + ], + "type": "text", + "content": "5. Conclusion" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 313, + 570, + 555, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 570, + 555, + 713 + ], + "spans": [ + { + "bbox": [ + 313, + 570, + 555, + 713 + ], + "type": "text", + "content": "We propose DreamTrack, a novel temporal framework with both historical and future information for visual object tracking. We reformulate the temporal learning in VOT as a History-to-Future process, which dreams the future states of the tracking scenario based on learned temporal dynamics from past experience, which empowers the tracker with the generalization in novel situations of new frames. The uncertainty of the future variation is modeled by forecasting the target trajectories of different possible future scenarios with multimodal prediction. Extensive experiments prove our tracker outperforms other mainstream trackers and achieves SOTA performance on prevailing benchmark datasets." + } + ] + } + ], + "index": 15 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "text", + "content": "7208" + } + ] + } + ], + "index": 16 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 7 + }, + { + "para_blocks": [ + { + "bbox": [ + 56, + 72, + 153, + 85 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 72, + 153, + 85 + ], + "spans": [ + { + "bbox": [ + 56, + 72, + 153, + 85 + ], + "type": "text", + "content": "Acknowledgments" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 55, + 91, + 296, + 175 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 91, + 296, + 175 + ], + "spans": [ + { + "bbox": [ + 55, + 91, + 296, + 175 + ], + "type": "text", + "content": "This work was supported by the National Natural Science Foundation of China under Grant (62436001,62176020); the Joint Foundation of the Ministry of Education for Innovation Team (8091B042235); the Fundamental Research Funds for the Central Universities (2019JBZ110); and the State Key Laboratory of Rail Traffic Control and Safety (RCS2023K006), Beijing Jiaotong University." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 56, + 186, + 115, + 198 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 186, + 115, + 198 + ], + "spans": [ + { + "bbox": [ + 56, + 186, + 115, + 198 + ], + "type": "text", + "content": "References" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 57, + 205, + 296, + 713 + ], + "type": "list", + "angle": 0, + "index": 17, + "blocks": [ + { + "bbox": [ + 61, + 205, + 296, + 239 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 61, + 205, + 296, + 239 + ], + "spans": [ + { + "bbox": [ + 61, + 205, + 296, + 239 + ], + "type": "text", + "content": "[1] Yifan Bai, Zeyang Zhao, Yihong Gong, and Xing Wei. Artrackv2: Prompting autoregressive tracker where to look and how to describe. In CVPR, 2024. 2, 6, 7" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 61, + 240, + 296, + 273 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 61, + 240, + 296, + 273 + ], + "spans": [ + { + "bbox": [ + 61, + 240, + 296, + 273 + ], + "type": "text", + "content": "[2] Luca Bertinetto, Jack Valmadre, João F Henriques, Andrea Vedaldi, and Philip H S Torr. Fully-convolutional siamese networks for object tracking. In ECCVW, 2016. 1, 2, 6, 7" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 62, + 274, + 295, + 306 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 274, + 295, + 306 + ], + "spans": [ + { + "bbox": [ + 62, + 274, + 295, + 306 + ], + "type": "text", + "content": "[3] Goutam Bhat, Martin Danelljan, Luc Van Gool, and Radu Timofte. Learning discriminative model prediction for tracking. In ICCV, 2019. 6, 7" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 62, + 308, + 294, + 330 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 308, + 294, + 330 + ], + "spans": [ + { + "bbox": [ + 62, + 308, + 294, + 330 + ], + "type": "text", + "content": "[4] Yidong Cai, Jie Liu, Jie Tang, and Gangshan Wu. Robust object modeling for visual tracking. In ICCV, 2023. 6, 7" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 62, + 331, + 295, + 364 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 331, + 295, + 364 + ], + "spans": [ + { + "bbox": [ + 62, + 331, + 295, + 364 + ], + "type": "text", + "content": "[5] Fei Chen, Xiaodong Wang, Yunxiang Zhao, Shaohe Lv, and Xin Niu. Visual object tracking: A survey. Computer Vision and Image Understanding, 2022. 1" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 62, + 365, + 295, + 397 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 365, + 295, + 397 + ], + "spans": [ + { + "bbox": [ + 62, + 365, + 295, + 397 + ], + "type": "text", + "content": "[6] Xin Chen, Bin Yan, Jiawen Zhu, Dong Wang, Xiaoyun Yang, and Huchuan Lu. Transformer tracking. In CVPR, 2021. 1, 2, 6, 7" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 62, + 399, + 295, + 431 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 399, + 295, + 431 + ], + "spans": [ + { + "bbox": [ + 62, + 399, + 295, + 431 + ], + "type": "text", + "content": "[7] Xin Chen, Houwen Peng, Dong Wang, Hutchuan Lu, and Han Hu. Seqtrack: Sequence to sequence learning for visual object tracking. In CVPR, 2023. 1, 2, 3, 4, 6, 7" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 62, + 434, + 295, + 475 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 434, + 295, + 475 + ], + "spans": [ + { + "bbox": [ + 62, + 434, + 295, + 475 + ], + "type": "text", + "content": "[8] Junyoung Chung, Caglar Gulcehre, KyungHyun Cho, and Yoshua Bengio. Empirical evaluation of gated recurrent neural networks on sequence modeling. arXiv preprint arXiv:1412.3555, 2014. 4" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 62, + 478, + 294, + 510 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 478, + 294, + 510 + ], + "spans": [ + { + "bbox": [ + 62, + 478, + 294, + 510 + ], + "type": "text", + "content": "[9] Yutao Cui, Jiang Cheng, Limin Wang, and Gangshan Wu. Mixformer: End-to-end tracking with iterative mixed attention. In CVPR, 2022. 1, 2, 5, 6" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 57, + 512, + 295, + 544 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 512, + 295, + 544 + ], + "spans": [ + { + "bbox": [ + 57, + 512, + 295, + 544 + ], + "type": "text", + "content": "[10] Martin Danelljan, Goutam Bhat, Fahad Shahbaz Khan, and Michael Felsberg. Eco: Efficient convolution operators for tracking. In CVPR, 2017. 1, 6, 7" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 57, + 546, + 294, + 578 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 546, + 294, + 578 + ], + "spans": [ + { + "bbox": [ + 57, + 546, + 294, + 578 + ], + "type": "text", + "content": "[11] Martin Danelljan, Goutam Bhat, Fahad Shahbaz Khan, and Michael Felsberg. ATOM: Accurate tracking by overlap maximization. In CVPR, 2019. 7" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 57, + 580, + 294, + 601 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 580, + 294, + 601 + ], + "spans": [ + { + "bbox": [ + 57, + 580, + 294, + 601 + ], + "type": "text", + "content": "[12] Martin Danelljan, Luc Van Gool, and Radu Timofte. Probabilistic regression for visual tracking. In CVPR, 2020. 1" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 57, + 603, + 294, + 666 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 603, + 294, + 666 + ], + "spans": [ + { + "bbox": [ + 57, + 603, + 294, + 666 + ], + "type": "text", + "content": "[13] Alexey Dosovitskiy, Lucas Beyer, Alexander Kolesnikov, Dirk Weissenborn, Xiaohua Zhai, Thomas Unterthiner, Mostafa Dehghani, Matthias Minderer, Georg Heigold, Sylvain Gelly, et al. An image is worth 16x16 words: Transformers for image recognition at scale. In ICLR, 2021. 4, 6" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 57, + 670, + 294, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 670, + 294, + 713 + ], + "spans": [ + { + "bbox": [ + 57, + 670, + 294, + 713 + ], + "type": "text", + "content": "[14] Frederik Ebert, Chelsea Finn, Sudeep Dasari, Annie Xie, Alex Lee, and Sergey Levine. Visual foresight: Model-based deep reinforcement learning for vision-based robotic control. arXiv preprint arXiv:1812.00568, 2018. 2" + } + ] + } + ], + "index": 16 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 316, + 73, + 553, + 713 + ], + "type": "list", + "angle": 0, + "index": 37, + "blocks": [ + { + "bbox": [ + 316, + 73, + 553, + 116 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 73, + 553, + 116 + ], + "spans": [ + { + "bbox": [ + 316, + 73, + 553, + 116 + ], + "type": "text", + "content": "[15] Heng Fan, Liting Lin, Fan Yang, Peng Chu, Ge Deng, Sijia Yu, Hexin Bai, Yong Xu, Chunyuan Liao, and Haibin Ling. LaSOT: A high-quality benchmark for large-scale single object tracking. In CVPR, 2019. 2, 6, 7, 8" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 317, + 118, + 553, + 160 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 118, + 553, + 160 + ], + "spans": [ + { + "bbox": [ + 317, + 118, + 553, + 160 + ], + "type": "text", + "content": "[16] Heng Fan, Hexin Bai, Liting Lin, Fan Yang, Peng Chu, Ge Deng, Sijia Yu, Mingzhen Huang, Juehuan Liu, Yong Xu, et al. Lasot: A high-quality large-scale single object tracking benchmark. IJCV, 2021. 6, 8" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 317, + 162, + 553, + 193 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 162, + 553, + 193 + ], + "spans": [ + { + "bbox": [ + 317, + 162, + 553, + 193 + ], + "type": "text", + "content": "[17] Liangji Fang, Qinhong Jiang, Jianping Shi, and Bolei Zhou. Tpnet: Trajectory proposal network for motion prediction. In CVPR, 2020. 5" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 316, + 195, + 553, + 216 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 195, + 553, + 216 + ], + "spans": [ + { + "bbox": [ + 316, + 195, + 553, + 216 + ], + "type": "text", + "content": "[18] Chelsea Finn and Sergey Levine. Deep visual foresight for planning robot motion. ICRA, 2016. 2" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 316, + 217, + 553, + 248 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 217, + 553, + 248 + ], + "spans": [ + { + "bbox": [ + 316, + 217, + 553, + 248 + ], + "type": "text", + "content": "[19] Shenyuan Gao, Chunluan Zhou, and Jun Zhang. Generalized relation modeling for transformer tracking. In CVPR, 2023. 6" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 316, + 251, + 553, + 292 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 251, + 553, + 292 + ], + "spans": [ + { + "bbox": [ + 316, + 251, + 553, + 292 + ], + "type": "text", + "content": "[20] Dongyan Guo, Jun Wang, Ying Cui, Zhenhua Wang, and Shengyong Chen. SiamCAR: Siamese fully convolutional classification and regression for visual tracking. In CVPR, 2020. 1, 2" + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 316, + 294, + 553, + 315 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 294, + 553, + 315 + ], + "spans": [ + { + "bbox": [ + 316, + 294, + 553, + 315 + ], + "type": "text", + "content": "[21] David Ha and Jürgen Schmidhuber. Recurrent world models facilitate policy evolution. NeurIPS, 2018. 2" + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 317, + 317, + 553, + 348 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 317, + 553, + 348 + ], + "spans": [ + { + "bbox": [ + 317, + 317, + 553, + 348 + ], + "type": "text", + "content": "[22] Danijar Hafner, Timothy Lillicrap, Jimmy Ba, and Mohammad Norouzi. Dream to control: Learning behaviors by latent imagination. *ICLR*, 2020. 2, 4" + } + ] + } + ], + "index": 25 + }, + { + "bbox": [ + 316, + 350, + 553, + 380 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 350, + 553, + 380 + ], + "spans": [ + { + "bbox": [ + 316, + 350, + 553, + 380 + ], + "type": "text", + "content": "[23] Danijar Hafner, Timothy Lillicrap, Mohammad Norouzi, and Jimmy Ba. Mastering atari with discrete world models. *ICLR*, 2021. 2" + } + ] + } + ], + "index": 26 + }, + { + "bbox": [ + 316, + 383, + 553, + 415 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 383, + 553, + 415 + ], + "spans": [ + { + "bbox": [ + 316, + 383, + 553, + 415 + ], + "type": "text", + "content": "[24] Danijar Hafner, Jurgis Pasukonis, Jimmy Ba, and Timothy Lillicrap. Mastering diverse domains through world models. arXiv preprint arXiv:2301.04104, 2023. 2, 4" + } + ] + } + ], + "index": 27 + }, + { + "bbox": [ + 316, + 416, + 553, + 447 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 416, + 553, + 447 + ], + "spans": [ + { + "bbox": [ + 316, + 416, + 553, + 447 + ], + "type": "text", + "content": "[25] João F Henriques, Rui Caseiro, Pedro Martins, and Jorge Batista. High-speed tracking with kernelized correlation filters. IEEE T-PAMI, 2014. 1" + } + ] + } + ], + "index": 28 + }, + { + "bbox": [ + 317, + 449, + 553, + 491 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 449, + 553, + 491 + ], + "spans": [ + { + "bbox": [ + 317, + 449, + 553, + 491 + ], + "type": "text", + "content": "[26] Anthony Hu, Gianluca Corrado, Nicolas Griffiths, Zachary Murez, Corina Gurau, Hudson Yeo, Alex Kendall, Roberto Cipolla, and Jamie Shotton. Model-based imitation learning for urban driving. NeurIPS, 2022. 2" + } + ] + } + ], + "index": 29 + }, + { + "bbox": [ + 317, + 494, + 553, + 525 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 494, + 553, + 525 + ], + "spans": [ + { + "bbox": [ + 317, + 494, + 553, + 525 + ], + "type": "text", + "content": "[27] Lianghua Huang, Xin Zhao, and Kaiqi Huang. Got-10k: A large high-diversity benchmark for generic object tracking in the wild. IEEE T-PAMI, 2019. 6, 7" + } + ] + } + ], + "index": 30 + }, + { + "bbox": [ + 317, + 526, + 553, + 557 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 526, + 553, + 557 + ], + "spans": [ + { + "bbox": [ + 317, + 526, + 553, + 557 + ], + "type": "text", + "content": "[28] Licheng Jiao, Dan Wang, Yidong Bai, Puhua Chen, and Fang Liu. Deep learning in visual tracking: A review. IEEE TNNLS, 2021. 1" + } + ] + } + ], + "index": 31 + }, + { + "bbox": [ + 317, + 559, + 553, + 591 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 559, + 553, + 591 + ], + "spans": [ + { + "bbox": [ + 317, + 559, + 553, + 591 + ], + "type": "text", + "content": "[29] Hamed Kiani Galoogahi, Ashton Fagg, Chen Huang, Deva Ramanan, and Simon Lucey. Need for speed: A benchmark for higher frame rate object tracking. In ICCV, 2017. 6, 7" + } + ] + } + ], + "index": 32 + }, + { + "bbox": [ + 316, + 593, + 553, + 613 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 593, + 553, + 613 + ], + "spans": [ + { + "bbox": [ + 316, + 593, + 553, + 613 + ], + "type": "text", + "content": "[30] Yann LeCun. A path towards autonomous machine intelligence version 0.9.2, 2022-06-27. Open Review, 2022. 2" + } + ] + } + ], + "index": 33 + }, + { + "bbox": [ + 317, + 615, + 553, + 646 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 615, + 553, + 646 + ], + "spans": [ + { + "bbox": [ + 317, + 615, + 553, + 646 + ], + "type": "text", + "content": "[31] Bo Li, Junjie Yan, Wei Wu, Zheng Zhu, and Xiaolin Hu. High performance visual tracking with siamese region proposal network. In CVPR, 2018. 1, 2" + } + ] + } + ], + "index": 34 + }, + { + "bbox": [ + 316, + 647, + 553, + 680 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 647, + 553, + 680 + ], + "spans": [ + { + "bbox": [ + 316, + 647, + 553, + 680 + ], + "type": "text", + "content": "[32] Bo Li, Wei Wu, Qiang Wang, Fangyi Zhang, Junliang Xing, and Junjie Yan. Siamrpn++: Evolution of siamese visual tracking with very deep networks. In CVPR, 2019. 2, 6" + } + ] + } + ], + "index": 35 + }, + { + "bbox": [ + 316, + 681, + 553, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 681, + 553, + 713 + ], + "spans": [ + { + "bbox": [ + 316, + 681, + 553, + 713 + ], + "type": "text", + "content": "[33] Liting Lin, Heng Fan, Zhipeng Zhang, Yong Xu, and Haibin Ling. Swintrack: A simple and strong baseline for transformer tracking. NeurIPS, 2022. 2, 6, 7" + } + ] + } + ], + "index": 36 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "text", + "content": "7209" + } + ] + } + ], + "index": 38 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 8 + }, + { + "para_blocks": [ + { + "bbox": [ + 56, + 73, + 294, + 713 + ], + "type": "list", + "angle": 0, + "index": 18, + "blocks": [ + { + "bbox": [ + 56, + 73, + 294, + 106 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 73, + 294, + 106 + ], + "spans": [ + { + "bbox": [ + 56, + 73, + 294, + 106 + ], + "type": "text", + "content": "[34] Liting Lin, Heng Fan, Zhipeng Zhang, Yaowei Wang, Yong Xu, and Haibin Ling. Tracking meets lora: Faster training, larger model, stronger performance. ECCV, 2024. 1, 2, 6, 7" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 56, + 107, + 294, + 150 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 107, + 294, + 150 + ], + "spans": [ + { + "bbox": [ + 56, + 107, + 294, + 150 + ], + "type": "text", + "content": "[35] Tsung-Yi Lin, Michael Maire, Serge J. Belongie, Lubomir D. Bourdev, Ross B. Girshick, James Hays, Pietro Perona, Deva Ramanan, Piotr Dollar, and C. Lawrence Zitnick. Microsoft COCO: Common objects in context. In ECCV, 2014. 6" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 56, + 152, + 294, + 184 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 152, + 294, + 184 + ], + "spans": [ + { + "bbox": [ + 56, + 152, + 294, + 184 + ], + "type": "text", + "content": "[36] Yicheng Liu, Jinghuai Zhang, Liangji Fang, Qinhong Jiang, and Bolei Zhou. Multimodal motion prediction with stacked transformers. In CVPR, 2021. 5" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 56, + 186, + 294, + 207 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 186, + 294, + 207 + ], + "spans": [ + { + "bbox": [ + 56, + 186, + 294, + 207 + ], + "type": "text", + "content": "[37] Ilya Loshchilov and Frank Hutter. Decoupled weight decay regularization. In ICLR, 2019. 6" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 56, + 209, + 294, + 240 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 209, + 294, + 240 + ], + "spans": [ + { + "bbox": [ + 56, + 209, + 294, + 240 + ], + "type": "text", + "content": "[38] Matthias Mueller, Neil Smith, and Bernard Ghanem. A benchmark and simulator for uav tracking. In ECCV, 2016. 6, 7" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 56, + 243, + 294, + 285 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 243, + 294, + 285 + ], + "spans": [ + { + "bbox": [ + 56, + 243, + 294, + 285 + ], + "type": "text", + "content": "[39] Matthias Muller, Adel Bibi, Silvio Giancola, Salman Alsubaihi, and Bernard Ghanem. TrackingNet: A large-scale dataset and benchmark for object tracking in the wild. In ECCV, 2018. 6, 7" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 56, + 286, + 294, + 330 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 286, + 294, + 330 + ], + "spans": [ + { + "bbox": [ + 56, + 286, + 294, + 330 + ], + "type": "text", + "content": "[40] Scott Reed, Konrad Zolna, Emilio Parisotto, Sergio Gomez Colmenarejo, Alexander Novikov, Gabriel Barth-maron, Mai Giménez, Yury Sulsky, Jackie Kay, Jost Tobias Springenberg, et al. A generalist agent. TMLR, 2022. 2" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 56, + 332, + 294, + 364 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 332, + 294, + 364 + ], + "spans": [ + { + "bbox": [ + 56, + 332, + 294, + 364 + ], + "type": "text", + "content": "[41] Shaoqing Ren, Kaiming He, Ross Girshick, and Jian Sun. Faster R-CNN: Towards real-time object detection with region proposal networks. In NIPS, 2015. 2, 3" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 56, + 365, + 294, + 409 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 365, + 294, + 409 + ], + "spans": [ + { + "bbox": [ + 56, + 365, + 294, + 409 + ], + "type": "text", + "content": "[42] Hamid Rezatofighi, Nathan Tsoi, JunYoung Gwak, Amir Sadeghian, Ian Reid, and Silvio Savarese. Generalized intersection over union: A metric and a loss for bounding box regression. In CVPR, 2019. 5" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 56, + 411, + 294, + 453 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 411, + 294, + 453 + ], + "spans": [ + { + "bbox": [ + 56, + 411, + 294, + 453 + ], + "type": "text", + "content": "[43] Yuval Tassa, Yotam Doron, Alistair Muldal, Tom Erez, Yazhe Li, Diego de Las Casas, David Budden, Abbas Abdelmaleki, Josh Merel, Andrew Lefrancq, et al. Deepmind control suite. arXiv preprint arXiv:1801.00690, 2018. 2" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 56, + 455, + 294, + 487 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 455, + 294, + 487 + ], + "spans": [ + { + "bbox": [ + 56, + 455, + 294, + 487 + ], + "type": "text", + "content": "[44] Paul Voigtlaender, Jonathon Luiten, Philip HS Torr, and Bastian Leibe. Siam r-cnn: Visual tracking by re-detection. In CVPR, 2020. 2, 6" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 56, + 489, + 294, + 520 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 489, + 294, + 520 + ], + "spans": [ + { + "bbox": [ + 56, + 489, + 294, + 520 + ], + "type": "text", + "content": "[45] Ning Wang, Wengang Zhou, Jie Wang, and Houqiang Li. Transformer meets tracker: Exploiting temporal context for robust visual tracking. In CVPR, 2021. 1, 2, 3, 6" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 56, + 522, + 294, + 566 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 522, + 294, + 566 + ], + "spans": [ + { + "bbox": [ + 56, + 522, + 294, + 566 + ], + "type": "text", + "content": "[46] Xiao Wang, Xiujun Shu, Zhipeng Zhang, Bo Jiang, Yaowei Wang, Yonghong Tian, and Feng Wu. Towards more flexible and accurate object tracking with natural language: Algorithms and benchmark. In CVPR, 2021. 6, 7, 8" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 56, + 567, + 294, + 599 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 567, + 294, + 599 + ], + "spans": [ + { + "bbox": [ + 56, + 567, + 294, + 599 + ], + "type": "text", + "content": "[47] Xing Wei, Yifan Bai, Yongchao Zheng, Dahu Shi, and Yihong Gong. Autoregressive visual tracking. In CVPR, 2023. 1, 2, 3, 4, 6, 7" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 56, + 601, + 294, + 634 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 601, + 294, + 634 + ], + "spans": [ + { + "bbox": [ + 56, + 601, + 294, + 634 + ], + "type": "text", + "content": "[48] Philipp Wu, Alejandro Escontrela, Danijar Hafner, Pieter Abbeel, and Ken Goldberg. Daydreamer: World models for physical robot learning. In CoRL. PMLR, 2023. 2" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 56, + 635, + 294, + 657 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 635, + 294, + 657 + ], + "spans": [ + { + "bbox": [ + 56, + 635, + 294, + 657 + ], + "type": "text", + "content": "[49] Yi Wu, Jongwoo Lim, and Ming Hsuan Yang. Object tracking benchmark. IEEE T-PAMI, 2015. 6, 7" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 56, + 658, + 294, + 690 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 658, + 294, + 690 + ], + "spans": [ + { + "bbox": [ + 56, + 658, + 294, + 690 + ], + "type": "text", + "content": "[50] Bin Yan, Houwen Peng, Jianlong Fu, Dong Wang, and Huchuan Lu. Learning spatio-temporal transformer for visual tracking. In ICCV, 2021. 1, 2, 3, 6" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 56, + 692, + 294, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 692, + 294, + 713 + ], + "spans": [ + { + "bbox": [ + 56, + 692, + 294, + 713 + ], + "type": "text", + "content": "[51] Botao Ye, Hong Chang, Bingpeng Ma, Shiguang Shan, and Xilin Chen. Joint feature learning and relation modeling for" + } + ] + } + ], + "index": 17 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 316, + 73, + 553, + 228 + ], + "type": "list", + "angle": 0, + "index": 24, + "blocks": [ + { + "bbox": [ + 333, + 73, + 553, + 95 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 333, + 73, + 553, + 95 + ], + "spans": [ + { + "bbox": [ + 333, + 73, + 553, + 95 + ], + "type": "text", + "content": "tracking: A one-stream framework. In ECCV, 2022. 1, 2, 4, 5, 6, 7" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 316, + 96, + 553, + 128 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 96, + 553, + 128 + ], + "spans": [ + { + "bbox": [ + 316, + 96, + 553, + 128 + ], + "type": "text", + "content": "[52] Zhipeng Zhang and Houwen Peng. Deeper and wider siamese networks for real-time visual tracking. In CVPR, 2019. 1, 2" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 316, + 130, + 553, + 162 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 130, + 553, + 162 + ], + "spans": [ + { + "bbox": [ + 316, + 130, + 553, + 162 + ], + "type": "text", + "content": "[53] Zhipeng Zhang, Houwen Peng, Jianlong Fu, Bing Li, and Weiming Hu. Ocean: Object-aware anchor-free tracking. In ECCV, 2020. 2, 6, 7" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 316, + 164, + 553, + 196 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 164, + 553, + 196 + ], + "spans": [ + { + "bbox": [ + 316, + 164, + 553, + 196 + ], + "type": "text", + "content": "[54] Zhipeng Zhang, Yihao Liu, Xiao Wang, Bing Li, and Weiming Hu. Learn to match: Automatic matching network design for visual tracking. ICCV, 2021. 6" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 316, + 198, + 553, + 228 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 198, + 553, + 228 + ], + "spans": [ + { + "bbox": [ + 316, + 198, + 553, + 228 + ], + "type": "text", + "content": "[55] Xingyi Zhou, Dequan Wang, and Philipp Krahenbuhl. Objects as points. arXiv preprint arXiv:1904.07850, 2019. 2, 3" + } + ] + } + ], + "index": 23 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 748, + 316, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 316, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 316, + 757 + ], + "type": "text", + "content": "7210" + } + ] + } + ], + "index": 25 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 9 + } + ], + "_backend": "vlm", + "_version_name": "2.6.4" +} \ No newline at end of file diff --git a/2025/DriveDreamer4D_ World Models Are Effective Data Machines for 4D Driving Scene Representation/aa9a31fd-776c-4a2a-a4f7-87c23be589e2_content_list.json b/2025/DriveDreamer4D_ World Models Are Effective Data Machines for 4D Driving Scene Representation/aa9a31fd-776c-4a2a-a4f7-87c23be589e2_content_list.json new file mode 100644 index 0000000000000000000000000000000000000000..4f2bc5c7cac292eedfc4fc538c0574a24b2fc5f9 --- /dev/null +++ b/2025/DriveDreamer4D_ World Models Are Effective Data Machines for 4D Driving Scene Representation/aa9a31fd-776c-4a2a-a4f7-87c23be589e2_content_list.json @@ -0,0 +1,1664 @@ +[ + { + "type": "text", + "text": "DriveDreamer4D: World Models Are Effective Data Machines for 4D Driving Scene Representation", + "text_level": 1, + "bbox": [ + 184, + 130, + 816, + 176 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Guosheng Zhao $^{1,2,3*}$ , Chaojun Ni $^{4*}$ , Xiaofeng Wang $^{1,2,3*}$ , Zheng Zhu $^{5*†}$ , Xueyang Zhang $^{6}$ , Yida Wang $^{6}$ , Guan Huang $^{5}$ , Xinze Chen $^{5}$ , Boyuan Wang $^{1,2,3}$ , Youyi Zhang $^{7}$ , Wenjun Mei $^{4}$ , Xingang Wang $^{2,3†}$", + "bbox": [ + 184, + 202, + 810, + 256 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "$^{1}$ Institute of Automation, Chinese Academy of Sciences, China", + "bbox": [ + 248, + 256, + 748, + 273 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "$^{2}$ School of Artificial Intelligence, University of Chinese Academy of Sciences, China", + "bbox": [ + 163, + 273, + 834, + 291 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "$^{3}$ Luoyang Institute for Robot and Intelligent Equipment, China $^{4}$ Peking University, China", + "bbox": [ + 142, + 291, + 854, + 309 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "$^{5}$ GigaAI, China $^{6}$ Li Auto Inc., China $^{7}$ Technical University of Munich, Germany", + "bbox": [ + 163, + 309, + 823, + 327 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/6ba133e2acea1579799ea73bcd998fdbe1c2c70ac15f65aa6c0bbf4c5a45d5ae.jpg", + "image_caption": [ + "Figure 1. Previous 4D Gaussian Splatting methods (e.g., PVG [9], $\\mathbf{S}^3$ Gaussian [31], Deformable-GS [80]) face challenges in rendering novel trajectories, such as lane change. DriveDreamer4D addresses this by enhancing 4D driving scene representation via integrating priors from world models, significantly improving rendering quality under complex scenarios and novel trajectory viewpoints." + ], + "image_footnote": [], + "bbox": [ + 101, + 363, + 401, + 542 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/8efd397d45a96f34d57b8bb9515a9a19b37fd1c7a3f3b5b3d99c2458b9a15631.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 405, + 363, + 568, + 449 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/072b40ab0261fb08a51bd24b96659ba57628b3314c9abcc27781e29a8099789d.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 570, + 363, + 733, + 449 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/4070e216b614821aff83b4a7c933acb565f4a799bb1d0c9791c887f96426c61c.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 735, + 363, + 900, + 449 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/454527b856f3366bf507c346142eaf349a1ce1a499fdd4bd8990f29746c8c851.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 403, + 455, + 568, + 542 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/ada2abcf22dd8589c7ee99ef00da41b9ebc46b78aea07b0c604ad4f4d8914c58.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 571, + 455, + 733, + 542 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/e315c8238599f15e77d9cfd012c952f9017e1f56d8b3a16e790c5bdeaa441aa4.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 736, + 455, + 900, + 542 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Abstract", + "text_level": 1, + "bbox": [ + 246, + 609, + 326, + 625 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Closed-loop simulation is essential for advancing end-to-end autonomous driving systems. Contemporary sensor simulation methods, such as NeRF and 3DGS, rely predominantly on conditions closely aligned with training data distributions, which are largely confined to forward-driving scenarios. Consequently, these methods face limitations when rendering complex maneuvers (e.g., lane change, acceleration, deceleration). Recent advancements in autonomous-driving world models have demonstrated the potential to generate diverse driving videos. However, these approaches remain constrained to 2D video generation, inherently lacking the spatiotemporal coherence required to capture intricacies of dynamic driving environments. In this paper, we introduce DriveDreamer4D, which", + "bbox": [ + 88, + 641, + 485, + 853 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "enhances 4D driving scene representation leveraging world model priors. Specifically, we utilize the world model as a data machine to synthesize novel trajectory videos, where structured conditions are explicitly leveraged to control the spatial-temporal consistency of traffic elements. Besides, the cousin data training strategy is proposed to facilitate merging real and synthetic data for optimizing 4DGS. To our knowledge, DriveDreamer4D is the first to utilize video generation models for improving 4D reconstruction in driving scenarios. Experimental results reveal that DriveDreamer4D significantly enhances generation quality under novel trajectory views, achieving a relative improvement in FID by $32.1\\%$ , $46.4\\%$ , and $16.3\\%$ compared to PVG, $S^3$ Gaussian, and Deformable-GS. Moreover, DriveDreamer4D markedly enhances the spatiotemporal coherence of driving agents, which is verified by a comprehensive user study and the relative increases of $22.6\\%$ , $43.5\\%$ , and $15.6\\%$ in the NTA-IoU metric.", + "bbox": [ + 511, + 611, + 908, + 882 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "CVF", + "bbox": [ + 106, + 2, + 181, + 42 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore.", + "bbox": [ + 236, + 0, + 810, + 46 + ], + "page_idx": 0 + }, + { + "type": "page_footnote", + "text": "*These authors contributed equally to this work. \n†Corresponding authors. zhengzhu@ieee.org, xingang.wang@ia.ac.cn. \n‡Project Page: https://drivedreamer4d.github.io", + "bbox": [ + 107, + 862, + 483, + 900 + ], + "page_idx": 0 + }, + { + "type": "page_number", + "text": "12015", + "bbox": [ + 480, + 944, + 519, + 957 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "1. Introduction", + "text_level": 1, + "bbox": [ + 91, + 89, + 222, + 106 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "End-to-end planning [29, 30, 34], which directly maps sensor inputs to control signals, is among the most critical and promising tasks in autonomous driving. However, current open-loop evaluations are inadequate for accurately assessing end-to-end planning algorithms, highlighting an urgent need for enhanced evaluation methods [41, 43, 86]. A compelling solution lies in closed-loop evaluations within real-world scenarios, which require retrieving sensor data from arbitrarily specified viewpoints. This necessitates constructing a 4D driving scene representation capable of reconstructing complex, dynamic driving environments.", + "bbox": [ + 89, + 114, + 480, + 281 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Closed-loop simulation in driving environments predominantly relies on scene reconstruction techniques such as Neural Radiance Fields (NeRF) [20, 49, 77, 79] and 3D Gaussian Splatting (3DGS) [11, 31, 36, 76], which are inherently limited by the density of input data. Specifically, these methods can render scenes effectively only under conditions closely aligned with their training data distributions—primarily forward-driving scenarios—and struggle to perform accurately during complex maneuvers (see Fig. 1). To mitigate these limitations, methods like SGD [84] and GGS [22] leverage generative models to extend the range of training viewpoints. However, these approaches primarily supplement sparse image data or static background elements, falling short of modeling the intricacies of dynamic, interactive driving scenes. Recently, advancements in autonomous driving world models [18, 28, 66, 68, 69, 87] have introduced the capability to generate diverse, command-aligned video viewpoints, offering renewed promise for closed-loop simulation in autonomous driving. Nonetheless, these models remain constrained to 2D videos, lacking the spatial-temporal coherence essential for accurately modeling complex driving scenarios.", + "bbox": [ + 89, + 281, + 482, + 613 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "In this paper, we introduce DriveDreamer4D, which improves 4D driving scene representation by integrating priors from autonomous driving world models. Our approach utilizes an autonomous driving world model [87] as a generative engine, synthesizing novel trajectory video data that densifies real-world driving datasets for enhanced training. Notably, we propose the Novel Trajectory Generation Module (NTGM) to generate diverse structured traffic conditions, and DriveDreamer4D applies these conditions to independently regulate the motion dynamics of foreground and background elements in complex driving environments. These conditions undergo view projection synchronized with vehicle maneuvers, ensuring that the synthesized data adheres to the spatiotemporal constraints. Subsequently, the Cousin Data Training Strategy (CDTS) is proposed to merge temporal-aligned real and synthetic data for training 4DGS. In CDTS, a regularization loss is further incorporated to ensure perceptual coherence. To the best of our knowledge, DriveDreamer4D is the first", + "bbox": [ + 89, + 613, + 482, + 900 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "framework to harness video generation models for elevating 4D scene reconstruction quality in autonomous driving, providing richly varied viewpoint data for scenarios including lane change, acceleration, and deceleration. As shown in Fig. 1, experiment results demonstrate that DriveDreamer4D significantly enhances generation fidelity for novel trajectory viewpoints, achieving a relative improvement in FID by $32.1\\%$ , $46.4\\%$ , and $16.3\\%$ compared to PVG [9], $\\mathrm{S}^3$ Gaussian [31], and Deformable-GS [80]. Besides, DriveDreamer4D fortifies the spatiotemporal coherence between foreground and background elements, with respective increases of $22.6\\%$ , $43.5\\%$ , and $15.6\\%$ in the NTA-IoU metric. Furthermore, a comprehensive user study confirms that the average win rate of DriveDreamer4D exceeds $80\\%$ , compared to three baselines.", + "bbox": [ + 511, + 90, + 903, + 316 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "The primary contributions of this work are as follows: (1) We present DriveDreamer4D, the first framework to leverage world model priors for advancing 4D scene reconstruction in autonomous driving. (2) The NTGM is proposed to automate the generation of structured conditions, allowing DriveDreamer4D to create novel trajectory videos with complex maneuvers while ensuring spatial-temporal consistency. Additionally, the CDTS is introduced to merge temporal-aligned real and synthetic data for training 4DGS, using a regularization loss to maintain perceptual coherence. (3) We perform comprehensive experiments to validate that DriveDreamer4D notably enhances generation quality across novel trajectory viewpoints, as well as the spatiotemporal coherence of driving scene elements.", + "bbox": [ + 511, + 318, + 903, + 531 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "2. Related Work", + "text_level": 1, + "bbox": [ + 511, + 547, + 653, + 564 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "2.1. Driving Scene Representation", + "text_level": 1, + "bbox": [ + 511, + 575, + 777, + 590 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "NeRF and 3DGS have emerged as leading approaches for 3D scene representation. NeRF models [2, 3, 49, 50] continuous volumetric scenes using multi-layer perceptron (MLP) networks, enabling highly detailed scene reconstructions with remarkable rendering quality. More recently, 3DGS [13, 36, 83] introduces an innovative method by defining a set of anisotropic Gaussians in 3D space, leveraging adaptive density control to achieve high-quality renderings from sparse point cloud inputs. Several works have extended NeRF [12, 20, 23, 33, 47, 56, 62, 63, 77, 79] or 3DGS [9, 11, 14, 31, 32, 44, 76, 84, 88, 89] to autonomous driving scenarios. Given the dynamic nature of driving environments, there has also been significant effort in modeling 4D driving scene representations. Some approaches encode time as an additional input to parameterize 4D scenes [1, 15, 31, 42, 46, 52, 60], while others represent scenes as a composition of moving object models alongside a static background model [39, 51, 64, 71, 74, 79]. Despite these advancements, methods based on NeRF and 3DGS face limitations tied to the density of input data. These tech-", + "bbox": [ + 511, + 598, + 903, + 900 + ], + "page_idx": 1 + }, + { + "type": "page_number", + "text": "12016", + "bbox": [ + 480, + 945, + 519, + 955 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "niques can only render scenes effectively when sensor data closely matches the training data distribution, which is typically confined to forward-driving scenarios.", + "bbox": [ + 89, + 90, + 482, + 137 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "2.2. World Models", + "text_level": 1, + "bbox": [ + 89, + 152, + 236, + 167 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "The world model module predicts possible future world states as a function of imagined action sequences proposed by the actor [40, 90]. Approaches such as [4, 5, 19, 21, 25-27, 38, 48, 67, 68, 72, 75, 81, 85] simulate environments through video generation controlled by free-text actions. At the forefront of this evolution is Sora [6], which leverages advanced generative techniques to produce intricate visual sequences that respect the fundamental laws of physics. This ability to deeply understand and simulate the environment not only improves video generation quality but also has substantial implications for real-world driving scenarios. Autonomous driving world models [18, 28, 66, 69, 78, 87] employ predictive methodologies to interpret driving environments, thereby generating realistic driving scenarios and learning key driving elements and policies from video data. Although these models successfully produce diverse driving video data conditioned on complex driving actions, they remain limited to 2D outputs and lack the spatial-temporal coherence needed to accurately capture the complexities of dynamic driving environments.", + "bbox": [ + 89, + 176, + 483, + 479 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "2.3. Diffusion Prior for 3D Representation", + "text_level": 1, + "bbox": [ + 89, + 494, + 418, + 512 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Constructing comprehensive 3D scenes from limited observations demands generative prior, particularly for unseen areas. Earlier studies distill the knowledge from text-to-image diffusion models [53, 55, 57, 58] into a 3D representation model. Specifically, the Score Distillation Sampling (SDS) [45, 54, 70] is adopted to synthesize a 3D object from the text prompt. Furthermore, to enhance 3D consistency, several approaches extend the multi-view diffusion models [17, 59] and video diffusion models [4, 10, 65] to 3D scene generation. To extend the diffusion prior to complex, dynamic, large-scale driving scenes for 3D reconstruction, methods such as SGD [84], GGS [22] and MagicDrive3D [16] employ generative models to broaden the range of training viewpoints. Nonetheless, these approaches mainly address sparse image data or static background elements, lacking the capacity to fully capture the complexities inherent in the 4D driving environments.", + "bbox": [ + 89, + 518, + 483, + 777 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3. Method", + "text_level": 1, + "bbox": [ + 89, + 797, + 181, + 811 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "In this section, we first elaborate on the preliminaries of 4D driving scene representation and world models for driving video generation. Then we present the details of DriveDreamer4D, which enhances 4D driving scene representation leveraging priors from driving world models.", + "bbox": [ + 89, + 825, + 483, + 901 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3.1. Preliminary", + "text_level": 1, + "bbox": [ + 513, + 90, + 643, + 107 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3.1.1. 4D Driving Scene Representation", + "text_level": 1, + "bbox": [ + 511, + 112, + 790, + 128 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "4DGS models the driving scene with a collection of 3DGS and a temporal field module. Each 3DGS [36] is parameterized by its center position $x$ , opacity $\\gamma$ , covariance $\\Sigma$ , and view-dependent RGB color $c$ , controlled via spherical harmonics. For stability, each covariance matrix $\\Sigma$ is decomposed by:", + "bbox": [ + 511, + 131, + 903, + 220 + ], + "page_idx": 2 + }, + { + "type": "equation", + "text": "\n$$\n\\boldsymbol {\\Sigma} = \\boldsymbol {R} \\boldsymbol {S} \\boldsymbol {S} ^ {T} \\boldsymbol {R} ^ {T}, \\tag {1}\n$$\n", + "text_format": "latex", + "bbox": [ + 650, + 220, + 903, + 238 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "where scaling matrix $S$ and a rotation matrix $\\pmb{R}$ are learnable parameters, represented by scaling $\\pmb{s}$ and quaternion $\\pmb{r}$ . All trainable parameters of a single 3D Gaussian are collectively denoted as $\\phi = \\{\\pmb{x},\\pmb{\\gamma},\\pmb{s},\\pmb{r},\\pmb{c}\\}$ . The temporal field $\\mathcal{F}$ takes $\\phi$ and a time step $t_{gs}$ as input, outputting the offset $\\delta \\phi = \\{\\delta x,\\delta \\gamma ,\\delta s,\\delta r,\\delta c\\}$ for each Gaussian relative to canonical space. The 4D Gaussian $\\phi^{\\prime} = \\{\\pmb{x}^{\\prime},\\pmb{\\gamma}^{\\prime},\\pmb{s}^{\\prime},\\pmb{r}^{\\prime},\\pmb{c}^{\\prime}\\}$ is then computed by:", + "bbox": [ + 511, + 244, + 906, + 367 + ], + "page_idx": 2 + }, + { + "type": "equation", + "text": "\n$$\n\\phi^ {\\prime} = \\phi + \\delta \\phi = \\phi + \\mathcal {F} (\\phi , \\boldsymbol {t} _ {g s}). \\tag {2}\n$$\n", + "text_format": "latex", + "bbox": [ + 599, + 380, + 903, + 398 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Following [82], a differentiable Gaussian Splating renderer is employed to project 4D Gaussians $\\phi$ into camera coordinates, yielding the covariance matrix $\\boldsymbol{\\Sigma}^{\\prime} = J\\boldsymbol{V}\\boldsymbol{\\Sigma}\\boldsymbol{V}^{T}\\boldsymbol{J}^{T}$ , where $\\boldsymbol{J}$ is the Jacobian matrix of the perspective projection, and $\\boldsymbol{V}$ is the transform matrix. The color of each pixel is calculated by $N$ ordered points using $\\alpha$ -blending:", + "bbox": [ + 511, + 404, + 905, + 496 + ], + "page_idx": 2 + }, + { + "type": "equation", + "text": "\n$$\nC = \\sum_ {i \\in N} T _ {i} \\boldsymbol {c} _ {i} ^ {\\prime} \\alpha_ {i}, \\tag {3}\n$$\n", + "text_format": "latex", + "bbox": [ + 650, + 506, + 903, + 539 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "where $T_{i}$ is the transmittance defined by $\\prod_{j = 1}^{i - 1}(1 - \\alpha_j)$ , $c_{i}^{\\prime}$ denotes the color of each point, $\\alpha_{i}$ is given by evaluating a 2D Gaussian with covariance $\\Sigma^{\\prime}$ multiplied with a learned per-point opacity $\\gamma_{i}^{\\prime}$ . The trainable parameters $\\phi^{\\prime}$ can be optimized by a combination of RGB loss, depth loss and SSIM loss:", + "bbox": [ + 511, + 551, + 905, + 643 + ], + "page_idx": 2 + }, + { + "type": "equation", + "text": "\n$$\n\\begin{array}{l} \\mathcal {L} _ {\\mathrm {o r i}} \\left(\\phi^ {\\prime}\\right) = \\lambda_ {1} \\| \\hat {I} _ {\\mathrm {o r i}} - I _ {\\mathrm {o r i}} \\| _ {1} + \\lambda_ {2} \\| \\hat {D} _ {\\mathrm {o r i}} - D _ {\\mathrm {o r i}} \\| _ {1} \\tag {4} \\\\ + \\lambda_ {3} \\operatorname {S S I M} \\left(\\bar {I} _ {\\text {o r i}}, I _ {\\text {o r i}}\\right), \\\\ \\end{array}\n$$\n", + "text_format": "latex", + "bbox": [ + 542, + 652, + 903, + 691 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "where $\\hat{I}_{\\mathrm{ori}}$ and $I_{\\mathrm{ori}}$ represent the rendered image and the ground truth image. $\\hat{D}_{\\mathrm{ori}}$ and $D_{\\mathrm{ori}}$ are the rendered depth and the ground truth LiDAR depth map. SSIM $(\\cdot)$ refers to the operation of the Structural Similarity Index Measure, and $\\lambda_1,\\lambda_2,\\lambda_3$ are the loss weights.", + "bbox": [ + 511, + 705, + 905, + 784 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3.1.2. World Models for Driving Video Generation", + "text_level": 1, + "bbox": [ + 511, + 790, + 867, + 806 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "The world model module predicts possible future world states based on imagined action sequences [40]. Autonomous-driving world models [18, 66, 69, 87], typically based on diffusion models, leverage structured driving information or action controls to guide future video prediction. During training, these models first encode", + "bbox": [ + 511, + 809, + 905, + 901 + ], + "page_idx": 2 + }, + { + "type": "page_number", + "text": "12017", + "bbox": [ + 480, + 944, + 517, + 955 + ], + "page_idx": 2 + }, + { + "type": "image", + "img_path": "images/f77c6426aff52ffd7e45ad74d6da743b6570d299e99351b50ce9bf561b2edaf9.jpg", + "image_caption": [ + "Figure 2. The overall framework of DriveDreamer4D. Initially, by altering the actions of the original trajectory (e.g., steering angle, speed), new trajectories can be obtained. Conditioned on the first frame and the structured information (3D bounding boxes, HDMap) from the new trajectory, the novel trajectory videos are generated. Subsequently, the temporal-aligned cousin pair (original and novel trajectory videos) are merged to optimize the 4D Gaussian Splatting model, where a regularization loss is calculated to ensure perceptual coherence." + ], + "image_footnote": [], + "bbox": [ + 98, + 92, + 898, + 470 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "videos $\\pmb{v}$ into a lower-dimensional latent space $z = \\mathcal{E}(\\pmb{v})$ using a variational encoder $\\mathcal{E}$ . After adding noise $\\epsilon_{t}$ to the latent, the diffusion model learns a denoising process. This diffusion process is optimized by:", + "bbox": [ + 89, + 568, + 482, + 628 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {L} _ {d i f f} = \\mathbb {E} _ {\\boldsymbol {z}, \\epsilon \\sim \\mathcal {N} (0, 1), t} \\left[ \\| \\epsilon_ {t} - \\epsilon_ {\\theta} (\\boldsymbol {z} _ {t}, t, \\boldsymbol {f}) \\| _ {2} ^ {2} \\right], \\tag {5}\n$$\n", + "text_format": "latex", + "bbox": [ + 130, + 638, + 482, + 664 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "where $\\epsilon_{\\theta}$ is a parameterized denoising network, $t$ denotes the time step, representing the level of noise added or removed at each stage. Additionally, to improve the controllability of the generated data, conditional features $f$ (e.g., reference images, speed, steering angle, scene layouts, camera poses and textual information) can be introduced into the reverse diffusion process, ensuring that the generated outputs adhere to the input control signals. During inference, the world models can be conditioned on a reference image to control the style of the output scene, while predicting the future world states contingent upon the other input actions.", + "bbox": [ + 89, + 672, + 482, + 839 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "3.2. DriveDreamer4D", + "text_level": 1, + "bbox": [ + 89, + 848, + 261, + 862 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "The overall pipeline of DriveDreamer4D is depicted in Fig. 2. In the upper part, the Novel Trajectory Generation", + "bbox": [ + 89, + 869, + 482, + 902 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Module (NTGM) is proposed to adjust driving actions (e.g., steering angle, speed) to generate new trajectories. These novel trajectories provide new perspectives for extracting structured information like 3D boxes and HDMap. Subsequently, a controllable video diffusion model synthesizes videos from these updated viewpoints, incorporating specific priors associated with the modified trajectories. In the lower part, the Cousin Data Training Strategy (CDTS) is introduced to combine the temporal-aligned original and generated data for optimizing the 4DGS model, where a regularization loss is calculated to impose perceptual coherence. In the following sections, we delve into the details of novel trajectories video generation and then introduce the CDTS for 4D reconstruction.", + "bbox": [ + 511, + 568, + 906, + 779 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "3.2.1. Novel Trajectory Video Generation", + "text_level": 1, + "bbox": [ + 511, + 789, + 805, + 806 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "As previously mentioned, traditional 4DGS methods are limited in rendering complex maneuvers, largely due to the training data being dominated by straightforward driving scenarios. To overcome this, DriveDreamer4D leverages world model priors to generate diverse viewpoint data, enhancing the 4D scene representation. To achieve this,", + "bbox": [ + 511, + 810, + 906, + 900 + ], + "page_idx": 3 + }, + { + "type": "page_number", + "text": "12018", + "bbox": [ + 480, + 944, + 517, + 955 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "we propose the NTGM, which is designed to create new trajectories that serve as input for the world model, enabling the automated generation of complex maneuver data. NTGM comprises two main components: (1) novel trajectory proposal, (2) trajectory safety assessment. In the novel trajectory proposal stage, text-to-trajectory [87] can be adopted to automatically generate diverse complex trajectories. Additionally, trajectories can be custom-designed to meet specific requirements, allowing for tailored data generation based on precise needs. The overview of the custom-designed trajectory proposal (e.g., lane change) and trajectory safety assessment is shown in the Algo. 1. In a specific driving scenario, the original trajectory in the world coordinate system can be readily acquired as $\\mathcal{T}_{\\mathrm{ori}}^{\\mathrm{world}} = \\{p_i^{\\mathrm{world}}\\}_{i=0}^{K}$ , where $K$ denotes the number of frames and $p_i^{\\mathrm{world}} \\in \\mathbb{R}^3$ refers to the position of the ego-vehicle at the $i$ -th frame. To propose novel trajectories, the original trajectory $\\mathcal{T}_{\\mathrm{ori}}^{\\mathrm{world}}$ is transformed into the ego-vehicle coordinate system of the first frame, denoted as $\\mathcal{T}_{\\mathrm{ori}}^{\\mathrm{EgoStart}}$ and computed as:", + "bbox": [ + 89, + 90, + 483, + 393 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\left[ p _ {i} ^ {\\text {E g o S t a r t}}, 1 \\right] ^ {T} = M _ {0} ^ {- 1} \\times \\left[ p _ {i} ^ {\\text {w o r l d}}, 1 \\right] ^ {T}, \\tag {6}\n$$\n", + "text_format": "latex", + "bbox": [ + 166, + 395, + 482, + 415 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "where $M_0 \\in \\mathbb{R}^{4 \\times 4}$ represents the transformation matrix from the ego-vehicle coordinate system of the first frame to the world coordinate system, $[\\cdot]$ denotes the operation of the concat. In the ego-vehicle coordinate system, the vehicle's heading is aligned with the positive $x$ -axis, the $y$ -axis points to the left side of the vehicle, and the $z$ -axis is oriented vertically upwards, perpendicular to the plane of the vehicle. Consequently, changes in the vehicle's velocity and direction can be respectively represented by adjusting the value along the $x$ -axis and $y$ -axis. A final safety assessment is conducted for the newly generated trajectory points, which includes verifying whether the vehicle trajectories $p$ remain within drivable areas $\\mathcal{B}_{\\mathrm{road}}$ and ensuring that no collisions occur with pedestrians or other vehicles $\\{o_j\\}_{j=1}^M$ .", + "bbox": [ + 89, + 422, + 483, + 638 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\np \\in \\mathcal {B} _ {\\text {r o a d}}, \\tag {7}\n$$\n", + "text_format": "latex", + "bbox": [ + 150, + 651, + 482, + 678 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\| p - o _ {j} \\| \\geq d _ {\\min }, \\quad \\forall j \\in \\{1, \\dots , M \\},\n$$\n", + "text_format": "latex", + "bbox": [ + 153, + 678, + 421, + 691 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "where $d_{\\mathrm{min}}$ is the minimal distance between different agents. Once a novel trajectory that complies with traffic regulations is generated, the road structure and 3D bounding boxes can be projected onto the camera view from the perspective of the new trajectory, thereby generating structured information relative to the updated trajectory. This structured information, along with the initial frame and text, is fed into a world model [87] to produce the videos that follow the novel trajectories.", + "bbox": [ + 89, + 704, + 483, + 840 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "3.2.2. Cousin Data Training Strategy", + "text_level": 1, + "bbox": [ + 89, + 849, + 349, + 866 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "To better integrate generated data for training 4DGS, we propose the CDTS. Specifically, we construct temporally", + "bbox": [ + 89, + 869, + 483, + 901 + ], + "page_idx": 4 + }, + { + "type": "code", + "sub_type": "algorithm", + "code_caption": [ + "Algorithm 1 Novel Trajectory Generation Module" + ], + "code_body": "Input: Trajectory $\\mathcal{T}_{\\mathrm{ori}}^{\\mathrm{world}}$ , Transformation matrix $M_0$ \nOutput: Novel trajectory $\\mathcal{T}_{\\mathrm{novel}}^{\\mathrm{ego}}$ $\\mathcal{T}_{\\mathrm{novel}}^{\\mathrm{ego}}\\gets [[0,0,0]]$ \nOffset $\\leftarrow 0$ \nfor each $p_{\\mathrm{ori}}^{\\mathrm{world}}$ in $\\mathcal{T}_{\\mathrm{ori}}^{\\mathrm{world}}$ [1:] do \n $p^{\\mathrm{EgoStart}}\\gets \\mathrm{RelativeCoord}(p_{\\mathrm{ori}}^{\\mathrm{world}},M_0)$ \nMaxOffset $\\leftarrow 0.1$ \nwhile True do \nNewOffset $\\leftarrow$ Offset + RandOffset(0, MaxOffset) \n $p^{\\mathrm{EgoStart}'}\\gets p^{\\mathrm{EgoStart}} + [0,$ NewOffset, 0] \nif SafeCheck $(p^{\\mathrm{EgoStart}'}$ ) then \nAddElement $(\\mathcal{T}_{\\mathrm{novel}}^{\\mathrm{ego}},p^{\\mathrm{EgoStart}'})$ \nOffset $\\leftarrow$ NewOffset \nbreak \nelse \nMaxOffset $\\leftarrow$ MaxOffset/2 \nend if \nend while \nend for", + "bbox": [ + 511, + 109, + 893, + 401 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "aligned cousin pair data as a minimal training batch:", + "bbox": [ + 511, + 426, + 859, + 441 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\operatorname {B a t c h S t a c k} \\left(\\left\\{\\hat {I} _ {\\text {o r i}, t} \\right\\} _ {t = 0} ^ {T}, \\left\\{\\hat {I} _ {\\text {n o v e l}, t} \\right\\} _ {t = 0} ^ {T}\\right), \\tag {8}\n$$\n", + "text_format": "latex", + "bbox": [ + 583, + 450, + 903, + 469 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "where $\\mathrm{BatchStack}(\\cdot)$ is the data processor to stack temporal-aligned $\\{\\hat{I}_{\\mathrm{ori},t}\\}_{t=0}^{T}$ and $\\{\\hat{I}_{\\mathrm{novel},t}\\}_{t=0}^{T}$ into a training batch. By leveraging real and synthetic data aligned at each timestep, CDTS mitigates the data gap in 4DGS training, enhancing the model's ability to learn consistent representations across real and synthetic data. To optimize 4DGS, the temporal-aligned cousin pair is input per step before gradient optimization. The loss function $\\mathcal{L}_{\\mathrm{ori}}$ for the original data is defined in Eq. 4. And the loss function $\\mathcal{L}_{\\mathrm{novel}}$ for the generated data is akin to [9, 31], defined as follows:", + "bbox": [ + 511, + 479, + 906, + 630 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\begin{array}{l} \\mathcal {L} _ {\\text {n o v e l}} \\left(\\phi^ {\\prime}\\right) = \\lambda_ {1} \\| \\hat {I} _ {\\text {n o v e l}} - I _ {\\text {n o v e l}} \\| _ {1} \\tag {9} \\\\ + \\lambda_ {3} \\operatorname {S S I M} \\left(\\hat {I} _ {\\text {n o v e l}}, I _ {\\text {n o v e l}}\\right), \\\\ \\end{array}\n$$\n", + "text_format": "latex", + "bbox": [ + 583, + 640, + 903, + 679 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "where $I_{\\mathrm{novel}}$ represents the generated images corresponding to the novel trajectories as described in Sec. 3.2.1, and $\\hat{I}_{\\mathrm{novel}}$ denotes the rendered images under the novel trajectories via differentiable splatting [82]. Notably, different from [9, 31], depth maps are not employed as constraints in the optimization of 4DGS when using the generated dataset $D_{\\mathrm{novel}}$ . The limitation arises from the fact that LiDAR point cloud data is exclusively collected for the original trajectory. When these LiDAR points are projected onto a new trajectory, it cannot produce a complete depth map for the new perspective, as something visible in the novel trajectory may have been occluded in the original view. Consequently, the incorporation of such depth maps does not facilitate the optimization of the 4DGS model. More details are described in", + "bbox": [ + 511, + 689, + 906, + 900 + ], + "page_idx": 4 + }, + { + "type": "page_number", + "text": "12019", + "bbox": [ + 480, + 944, + 517, + 955 + ], + "page_idx": 4 + }, + { + "type": "table", + "img_path": "images/a2a03a8b7a8df68773d7970544348c429163245e149a3bad1b92c4cc8b367fdc.jpg", + "table_caption": [], + "table_footnote": [], + "table_body": "
MethodLane ChangeAccelerationDecelerationAverage
NTA-IoU ↑NTL-IoU ↑NTA-IoU ↑NTL-IoU ↑NTA-IoU ↑NTL-IoU ↑NTA-IoU ↑NTL-IoU ↑
PVG [9]0.25650.700.39653.080.39453.650.34952.48
DriveDreamer4D with PVG0.43853.060.42153.350.42453.890.42853.43
S3Gaussian [31]0.17549.050.43451.930.38452.140.33151.04
DriveDreamer4D with S3Gaussian0.49553.420.48452.630.44552.690.47552.91
Deformable-GS [80]0.24051.620.34652.170.37753.210.32152.33
DriveDreamer4D with Deformable-GS0.33552.930.37152.770.40653.790.37153.16
", + "bbox": [ + 93, + 88, + 906, + 214 + ], + "page_idx": 5 + }, + { + "type": "table", + "img_path": "images/bdd48b393301fe1bd23c7a1d711768df02977270c589031f80659982d2f59cad.jpg", + "table_caption": [ + "Table 1. Comparison of NTA-IoU and NTL-IoU scores across different novel trajectory views (lane change, acceleration, deceleration)." + ], + "table_footnote": [], + "table_body": "
MethodFID ↓
PVG[9]105.29
DriveDreamer4D with PVG71.52
S3Gaussian[31]124.90
DriveDreamer4D with S3Gaussian66.93
Deformable-GS[80]92.34
DriveDreamer4D with Deformable-GS77.32
", + "bbox": [ + 91, + 262, + 478, + 390 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Table 2. Comparison of FID scores in novel trajectory view synthesis (lane change) on the Waymo dataset.", + "bbox": [ + 89, + 398, + 482, + 428 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Sec. 4.3. Additionally, we propose a regularization loss to enhance the perceptual coherence, defined as follows:", + "bbox": [ + 89, + 454, + 482, + 486 + ], + "page_idx": 5 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {L} _ {\\text {r e g}} \\left(\\phi^ {\\prime}\\right) = \\left\\| \\mathcal {F} _ {p} \\left(\\hat {I} _ {\\text {o r i}}\\right) - \\mathcal {F} _ {p} \\left(\\hat {I} _ {\\text {n o v e l}}\\right) \\right\\| _ {1}, \\tag {10}\n$$\n", + "text_format": "latex", + "bbox": [ + 161, + 496, + 482, + 516 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "where $\\mathcal{F}_p(\\cdot)$ denotes the perception feature extraction model [24]. The overall loss function for mixed training is defined as follows:", + "bbox": [ + 89, + 527, + 483, + 571 + ], + "page_idx": 5 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {L} \\left(\\phi^ {\\prime}\\right) = \\mathcal {L} _ {\\text {o r i}} + \\lambda_ {\\text {n o v e l}} \\mathcal {L} _ {\\text {n o v e l}} + \\lambda_ {\\text {r e g}} \\mathcal {L} _ {\\text {r e g}}. \\tag {11}\n$$\n", + "text_format": "latex", + "bbox": [ + 156, + 585, + 482, + 604 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "4. Experiments", + "text_level": 1, + "bbox": [ + 89, + 628, + 223, + 647 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "In this section, we first outline the experimental setup, including the dataset, implementation details, and evaluation metrics. Subsequently, quantitative results across different datasets are provided to demonstrate that the proposed DriveDreamer4D significantly enhances rendering quality for novel trajectory viewpoints and improves the spatiotemporal coherence of foreground and background components. Finally, we conduct an ablation study on the hyperparameter settings, as well as the effects of depth loss and the proposed CDTS, including temporal-aligned pairs, and regularization loss.", + "bbox": [ + 88, + 655, + 482, + 821 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "4.1. Experiment Setup", + "text_level": 1, + "bbox": [ + 89, + 832, + 266, + 849 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Dataset. We conduct experiments using the Waymo dataset [61], known for its comprehensive real-world driving logs. We specifically select eight scenes characterized by highly", + "bbox": [ + 89, + 854, + 483, + 902 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "dynamic interactions, featuring numerous vehicles with diverse relative positions and intricate driving trajectories. Each selected segment contains approximately 40 frames, with segment IDs detailed in the supplement. In addition, we have also conduct experiments on the nuscenes [7], and pandas [73] datasets. For nuScenes, we select eight scenes rich in dynamic objects and complex lane structures to test large maneuver capabilities. For PandaSet, we adopt the ten scenes used in Unisim [79] and NeuRAD [64] for fair comparison.", + "bbox": [ + 511, + 263, + 906, + 415 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Implementation Details. To demonstrate the versatility and robustness of DriveDreamer4D, we incorporate various 4DGS baselines into our pipeline, including Deformable-GS [80], S $^3$ Gaussian [31], and PVG [9]. For a fair comparison, LiDAR supervision is introduced to Deformable-GS. Scenes are segmented into 40-frame clips, matching the generative model's output length, using only forward-facing camera data at a resolution of $640 \\times 960$ . Models are trained with the Adam optimizer [37], following the learning rate schedule from 3D Gaussian Splatting [36]. Hyperparameters align with each baseline [9, 31, 80], with training strategies unchanged except for incorporating CDTS. Moreover, to demonstrate the robustness of our method across different datasets, we employ PVG and DriveDreamer4D with PVG on nuScenes [7], and PandaSet [73]. Since existing methods typically conduct experiments on one or two datasets with varying scene selections, we have chosen different state-of-the-art (SOTA) methods for comparison on each dataset. This approach ensures a comprehensive evaluation of our method's performance under diverse conditions and highlights its adaptability and effectiveness across a wide range of scenarios. By benchmarking against various SOTA methods tailored to specific datasets, we provide strong evidence of the generalizability and robustness of DriveDreamer4D in handling complex and varied driving environments.", + "bbox": [ + 511, + 416, + 908, + 792 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Metrics. Traditional 3D reconstruction tasks typically employ PSNR and SSIM metrics for evaluation, with validation sets that closely match the training data distribution (i.e., uniformly sampling frames from video sequences for validation, with the remainder used for training). However, in closed-loop driving simulation, the focus shifts to evaluating model rendering performance under novel trajectories,", + "bbox": [ + 511, + 795, + 908, + 902 + ], + "page_idx": 5 + }, + { + "type": "page_number", + "text": "12020", + "bbox": [ + 480, + 944, + 519, + 957 + ], + "page_idx": 5 + }, + { + "type": "table", + "img_path": "images/6504ce337ab222e93d4439f569585040829750c6eac2669b3db7fb966a5c6005.jpg", + "table_caption": [], + "table_footnote": [], + "table_body": "
Counterpart MethodDriveDreamer4D Win Rate
Lane ChangeAccelerationDecelerationAverage
PVG [9]100.0%90.5%89.1%93.2%
S3Gaussian [31]100.0%97.9%92.2%96.7%
Deformable-GS [80]95.8%83.5%72.9%84.1%
", + "bbox": [ + 94, + 88, + 482, + 156 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "where corresponding sensor data are unavailable, making metrics like PSNR and SSIM inapplicable for evaluation. Therefore, we propose Novel Trajectory Agent IoU (NTA-IoU) and Novel Trajectory Lane IoU (NTL-IoU), which assess the spatiotemporal coherence of foreground and background traffic components in novel trajectory viewpoints.", + "bbox": [ + 89, + 220, + 482, + 311 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "For NTA-IoU, we use YOLO11 [35] to identify vehicles in images rendered from novel trajectory views, yielding 2D bounding boxes. Simultaneously, geometric transformations are applied to the original 3D bounding boxes, projecting them onto the new viewpoints to generate corresponding 2D bounding boxes. For each projected 2D box, we then identify the closest detector-generated 2D box and compute their Intersection over Union (IoU). To ensure accurate matching, a distance threshold $d_{\\mathrm{thresh}}$ is introduced: when the center-to-center distance $\\| c(B^{\\mathrm{proj}}) - c(B^{\\mathrm{det}})\\|$ between the nearest detected box $B^{\\mathrm{det}}$ and the correctly projected box $B^{\\mathrm{proj}}$ surpasses this threshold, their NTA-IoU is assigned a value of zero:", + "bbox": [ + 88, + 311, + 482, + 508 + ], + "page_idx": 6 + }, + { + "type": "equation", + "text": "\n$$\n\\mathrm {N T A - I o U} = \\left\\{ \\begin{array}{l l} 0 & \\text {i f} \\| c (B ^ {\\operatorname {p r o j}}) - c (B ^ {\\det }) \\| \\geq d _ {\\text {t h r e s h}} \\\\ \\operatorname {I o U} \\left(B ^ {\\operatorname {p r o j}}, B ^ {\\det }\\right) & \\text {o t h e r w i s e .} \\end{array} \\right. \\tag {12}\n$$\n", + "text_format": "latex", + "bbox": [ + 94, + 520, + 482, + 561 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "For NTL-IoU, we employ TwinLiteNet [8] to extract 2D lanes from rendered images. Ground truth lanes are also projected onto the 2D image plane. We then compute the mean Intersection-over-Union (mIoU) between the rendered and ground truth lanes $L^{\\mathrm{det}}$ and $L^{\\mathrm{proj}}$ :", + "bbox": [ + 89, + 571, + 483, + 647 + ], + "page_idx": 6 + }, + { + "type": "equation", + "text": "\n$$\n\\mathrm {N T L - I o U} = \\mathrm {m I o U} \\left(L ^ {\\text {p r o j}}, L ^ {\\text {d e t}}\\right). \\tag {13}\n$$\n", + "text_format": "latex", + "bbox": [ + 181, + 657, + 482, + 676 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Additionally, in lane change scenarios, we observe inaccuracies in relative positioning, as well as frequent occurrences of artifacts such as flying points and ghosting, which notably degrade image quality. To assess this, we employ the FID metric [24], which quantifies differences in feature distribution between rendered novel trajectory images and original trajectory images. This metric effectively reflects visual quality and is particularly sensitive to artifacts like flying points and ghosting, providing a robust measure of image fidelity in these complex scenes. Finally, a user study is conducted to evaluate the quality of the renderings, where participants compare the rendering results of each baseline with its DriveDreamer4D enhanced version across three novel trajectories. The evaluation criteria", + "bbox": [ + 88, + 688, + 482, + 900 + ], + "page_idx": 6 + }, + { + "type": "table", + "img_path": "images/9a085df4e87fb89850ccf0036681d802c616a5ce355d63656ba4f38ee226027d.jpg", + "table_caption": [ + "Table 3. User study comparison of DriveDreamer4D win rates across various novel trajectory view synthesis." + ], + "table_footnote": [], + "table_body": "
MethodNTA-IoU ↑NTL-IoU ↑FID ↓
PVG[9]0.20448.75103.92
DriveDreamer4D with PVG0.33850.3688.60
", + "bbox": [ + 516, + 88, + 905, + 143 + ], + "page_idx": 6 + }, + { + "type": "table", + "img_path": "images/030faf9bf9ccbaa0555f5832bc00880ea18fe3023294eda3deff1f70ce52bae7.jpg", + "table_caption": [ + "Table 4. Comparison of NTA-IoU, NTL-IoU and FID scores across lane change on nuScenes." + ], + "table_footnote": [], + "table_body": "
MethodNo shift ↓Lane 2m ↓Lane 3m ↓Vert. 1m ↓
UniSim [79]-74.797.5-
NeuRAD [64]25.072.393.976.3
PVG [9]24.767.376.882.9
DriveDreamer4D with PVG23.562.474.669.8
", + "bbox": [ + 516, + 196, + 905, + 266 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Table 5. FID when shifting pose of ego vehicle on PandaSet.", + "bbox": [ + 527, + 275, + 888, + 289 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "focus on overall video quality, with particular attention to foreground objects like vehicles. For each comparison, participants were asked to select the option they found most favorable. Further details are provided in the supplement.", + "bbox": [ + 511, + 318, + 906, + 380 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "4.2. Main Results", + "text_level": 1, + "bbox": [ + 511, + 392, + 651, + 406 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "In this subsection, we first demonstrate the effectiveness of DriveDreamer4D with various baseline methods. Then, through experimental results obtained from three different datasets, combined with detailed comparisons against a range of baselines, we highlight the robustness and adaptability of DriveDreamer4D in addressing a broad spectrum of scenarios and challenges.", + "bbox": [ + 511, + 414, + 905, + 520 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Waymo. As demonstrated in Tab. 1, integrating DriveDreamer4D with various 4DGS algorithms consistently yields superior NTA-IoU and NTL-IoU scores across complex maneuvers (e.g., lane changes, acceleration, and deceleration), surpassing baseline methods. Specifically, DriveDreamer4D enhances average NTA-IoU by $22.6\\%$ , $43.5\\%$ , and $15.6\\%$ across three baselines (PVG [9], $\\mathrm{S}^3$ Gaussian [31], Deformable-GS [80]), underscoring its capability to improve the spatiotemporal coherence of foreground agents. It also boosts average NTL-IoU by $1.8\\%$ , $3.7\\%$ , and $1.6\\%$ , enhancing background lane consistency in 4D driving scene rendering.", + "bbox": [ + 511, + 521, + 905, + 703 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "In addition to verifying the spatiotemporal consistency of rendered novel trajectory views, we leverage the FID metric to assess rendering quality under novel trajectories. Given that acceleration and deceleration scenarios yield rendered views with distributional similarities to ground truth, limiting FID's discriminative capability across algorithms, our FID comparisons focus specifically on lane change scenarios. Experiment results, as presented in Tab. 2, indicate that our method substantially outperforms the baseline methods (PVG [9], $\\mathrm{S}^3$ Gaussian [31], Deformable-GS [80]), with FID relative improvements of $32.1\\%$ , $46.4\\%$ , and $16.3\\%$ . These results highlight DriveDreamer4D's capability to enhance generation quality for", + "bbox": [ + 511, + 704, + 908, + 902 + ], + "page_idx": 6 + }, + { + "type": "page_number", + "text": "12021", + "bbox": [ + 480, + 944, + 517, + 955 + ], + "page_idx": 6 + }, + { + "type": "table", + "img_path": "images/c2b8488fa3a72f6558adda5f489b58b680bf61254922673c9450097e4b7ed026.jpg", + "table_caption": [], + "table_footnote": [], + "table_body": "
λnovelNTA-IoU ↑FID ↓
00.349105.29
0.50.40582.84
10.42079.54
1.50.41782.10
", + "bbox": [ + 96, + 88, + 285, + 195 + ], + "page_idx": 7 + }, + { + "type": "table", + "img_path": "images/972f2da01b3ed0d96993bb4bef9cb9c903a7133bcf89824ccb6c1e5ea6053e50.jpg", + "table_caption": [ + "Table 6. Ablation study on the training loss weight $\\lambda_{\\mathrm{novel}}$ for novel trajectory data." + ], + "table_footnote": [], + "table_body": "
λregNTA-IoU ↑FID ↓
00.42079.54
1e-20.411119.39
1e-30.42871.52
1e-40.42275.31
", + "bbox": [ + 297, + 88, + 478, + 196 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "novel trajectory viewpoints.", + "bbox": [ + 89, + 277, + 276, + 292 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Finally, we conduct a user study to evaluate the rendering quality of different methods on novel trajectories, with a specific focus on foreground agents. For each method, we generate three novel trajectory views—lane change, acceleration, and deceleration—across eight scenes from the Waymo dataset [61]. Participants are then asked to select the renderings they found most visually favorable in each comparison. The DriveDreamer4D win rates from this study, shown in Tab. 3, reveal a significant user preference for our method's renderings.", + "bbox": [ + 88, + 294, + 482, + 445 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "nuScenes. Tab. 4 provides a comprehensive performance comparison between DriveDreamer4D and PVG [9] in lane change scenarios on the nuScenes dataset [7]. The experimental results offer compelling evidence that our method achieves a substantial improvement over the baseline, showcasing its superior capability in handling complex driving situations. Specifically, as detailed in the table, DriveDreamer4D consistently outperforms PVG across multiple evaluation metrics, including NTA-IoU, NTL-IoU, and FID. Notably, our method achieves a $65.7\\%$ increase in NTA-IoU, a $3.3\\%$ increase in NTL-IoU, and a $14.7\\%$ reduction in FID. These results underscore the robustness of DriveDreamer4D in addressing challenging driving scenarios, highlighting its exceptional performance and adaptability across diverse evaluation criteria. Such advancements reaffirm the effectiveness of our approach in delivering more accurate and reliable trajectory predictions for dynamic environments.", + "bbox": [ + 91, + 446, + 482, + 717 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "PandaSet. Following the experimental setup of NeuRAD [64] and UniSim [79], we conduct experiments on the PandaSet dataset [73]. As shown in Tab. 5, we report the FID comparison under the original trajectory, a lane shift of 2 meters, a lane shift of 3 meters, and a perpendicular shift of 1 meter. These results demonstrate that our DriveDreamer4D achieves the best performance across all settings. Notably, DriveDreamer4D achieves an FID of 74.6 under the challenging scenario of a 3-meter lane shift, representing a significant improvement of $20.6\\%$ over the previous SOTA method, NeuRAD. This highlights the strong performance and competitiveness of our approach com", + "bbox": [ + 89, + 719, + 482, + 901 + ], + "page_idx": 7 + }, + { + "type": "table", + "img_path": "images/53c1a2ffc6d1651174125868c09b3bda3e1ba91e8f1ae21b40fbdb96834bade7.jpg", + "table_caption": [ + "Table 7. Ablation study on the regularization loss weight ${\\lambda }_{\\text{reg }}$ ." + ], + "table_footnote": [], + "table_body": "
Depth LossTemporal-aligned Cousin PairRegularization LossNTA-IoU ↑FID ↓
✓××0.40182.63
×××0.42079.54
×✓×0.42376.20
×✓✓0.42871.52
", + "bbox": [ + 517, + 88, + 905, + 178 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Table 8. Ablation study on the depth loss and CDTS.", + "bbox": [ + 550, + 188, + 864, + 202 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "pared to existing techniques, reaffirming its effectiveness in handling complex driving scenarios with high precision and robustness.", + "bbox": [ + 511, + 234, + 905, + 279 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "4.3. Ablation Studies", + "text_level": 1, + "bbox": [ + 511, + 297, + 676, + 311 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "We conduct an ablation study based on PVG [9], analyzing the effects of hyperparameters settings for $\\lambda_{\\mathrm{novel}}$ and $\\lambda_{\\mathrm{reg}}$ , as well as the impact of depth loss, temporal-aligned cousin pairs, and regularization loss. Following the experiments shown in Tab. 6 and 7, we set $\\lambda_{\\mathrm{novel}}$ to 1 and $\\lambda_{\\mathrm{reg}}$ to $1\\times 10^{-3}$ which yield the best performance. As shown in Tab. 8, the experiment confirms that the depth loss should be excluded when optimizing novel trajectory views, since LiDAR depth maps are incomplete due to occlusions. Furthermore, employing temporal-aligned cousin pairs achieves an FID of 76.20 and an NTA-IoU of 0.423, with the perception regularization loss further improving the FID to 71.52 and NTA-IoU to 0.428. These results highlight the effectiveness of CDTS, in achieving $\\sim 10\\%$ improvement in FID, along with a $\\sim 2\\%$ increase in NTA-IoU.", + "bbox": [ + 511, + 321, + 906, + 547 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "5. Discussion and Conclusion", + "text_level": 1, + "bbox": [ + 511, + 570, + 761, + 585 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "In this paper, we present DriveDreamer4D, a novel framework designed to advance 4D driving scene representations by harnessing priors from world models. Addressing key limitations of current sensor simulation methods—namely, their dependence on forward-driving training data distributions and inability to model complex maneuvers—DriveDreamer4D leverages a world model to generate novel trajectory videos that complement real-world driving data. By explicitly employing structured conditions, our framework maintains spatial-temporal consistency across traffic elements, ensuring that generated data adheres closely to the dynamics of real-world traffic scenarios. Our experiments demonstrate that DriveDreamer4D achieves superior quality in generating diverse simulation viewpoints, with significant improvements in both the rendering fidelity and spatiotemporal coherence of scene components. Notably, these results highlight DriveDreamer4D's potential as a foundation for closed-loop simulations that require high-fidelity reconstructions of dynamic driving scenes.", + "bbox": [ + 511, + 598, + 906, + 900 + ], + "page_idx": 7 + }, + { + "type": "page_number", + "text": "12022", + "bbox": [ + 480, + 944, + 517, + 955 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "References", + "text_level": 1, + "bbox": [ + 91, + 89, + 187, + 104 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[1] Benjamin Attal, Jia-Bin Huang, Christian Richardt, Michael Zollhoefer, Johannes Kopf, Matthew O'Toole, and Changil Kim. Hyperreel: High-fidelity 6-dof video with ray-conditioned sampling. In CVPR, 2023. 2", + "[2] Jonathan T Barron, Ben Mildenhall, Dor Verbin, Pratul P Srinivasan, and Peter Hedman. Mip-nerf 360: Unbounded anti-aliased neural radiance fields. In CVPR, 2022. 2", + "[3] Jonathan T Barron, Ben Mildenhall, Dor Verbin, Pratul P Srinivasan, and Peter Hedman. Zip-nerf: Anti-aliased grid-based neural radiance fields. In ICCV, 2023. 2", + "[4] Andreas Blattmann, Tim Dockhorn, Sumith Kulal, Daniel Mendelevitch, Maciej Kilian, Dominik Lorenz, Yam Levi, Zion English, Vikram Voleti, Adam Letts, et al. Stable video diffusion: Scaling latent video diffusion models to large datasets. arXiv preprint arXiv:2311.15127, 2023. 3", + "[5] Andreas Blattmann, Robin Rombach, Huan Ling, Tim Dockhorn, Seung Wook Kim, Sanja Fidler, and Karsten Kreis. Align your latents: High-resolution video synthesis with latent diffusion models. In CVPR, 2023. 3", + "[6] Tim Brooks, Bill Peebles, Connor Holmes, Will DePue, Yufei Guo, Li Jing, David Schnurr, Joe Taylor, Troy Luhman, Eric Luhman, Clarence Ng, Ricky Wang, and Aditya Ramesh. Video generation models as world simulators. 2024. 3", + "[7] Holger Caesar, Varun Bankiti, Alex H Lang, Sourabh Vora, Venice Erin Liong, Qiang Xu, Anush Krishnan, Yu Pan, Giancarlo Baldan, and Oscar Beijbom. nuscenes: A multimodal dataset for autonomous driving. In CVPR, 2020. 6, 8", + "[8] Quang-Huy Che, Dinh-Phuc Nguyen, Minh-Quan Pham, and Duc-Khai Lam. Twinlitenet: An efficient and lightweight model for driveable area and lane segmentation in self-driving cars. In MAPR, 2023. 7", + "[9] Yurui Chen, Chun Gu, Junzhe Jiang, Xiatian Zhu, and Li Zhang. Periodic vibration gaussian: Dynamic urban scene reconstruction and real-time rendering. arXiv preprint arXiv:2311.18561, 2023. 1, 2, 5, 6, 7, 8", + "[10] Zilong Chen, Yikai Wang, Feng Wang, Zhengyi Wang, and Huaping Liu. V3d: Video diffusion models are effective 3d generators. arXiv preprint arXiv:2403.06738, 2024. 3", + "[11] Ziyu Chen, Jiawei Yang, Jiahui Huang, Riccardo de Lutio, Janick Martinez Esturo, Boris Ivanovic, Or Litany, Zan Gojcic, Sanja Fidler, Marco Pavone, Li Song, and Yue Wang. Omnire: Omni urban scene reconstruction. arXiv preprint arXiv:2408.16760, 2024. 2", + "[12] Kai Cheng, Xiaoxiao Long, Wei Yin, Jin Wang, Zhiqiang Wu, Yuexin Ma, Kaixuan Wang, Xiaozhi Chen, and Xuejin Chen. Uc-nerf: Neural radiance field for under-calibrated multi-view cameras in autonomous driving. arXiv preprint arXiv:2311.16945, 2023. 2", + "[13] Xiaobiao Du, Yida Wang, and Xin Yu. Mvgs: Multi-view-regulated gaussian splattingg for novel view synthesis. arXiv preprint arXiv:2410.02103, 2024. 2", + "[14] Tobias Fischer, Jonas Kulhanek, Samuel Rota Bulo, Lorenzo Porzi, Marc Pollefeys, and Peter Kontschieder. Dynamic 3d gaussian fields for urban areas. NeurIPS, 2024. 2" + ], + "bbox": [ + 93, + 114, + 483, + 900 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[15] Sara Fridovich-Keil, Giacomo Meanti, Frederik Rahbæk Warburg, Benjamin Recht, and Angjoo Kanazawa. K-planes: Explicit radiance fields in space, time, and appearance. In CVPR, 2023. 2", + "[16] Ruiyuan Gao, Kai Chen, Zhihao Li, Lanqing Hong, Zhenguo Li, and Qiang Xu. Magicdrive3d: Controllable 3d generation for any-view rendering in street scenes. arXiv preprint arXiv:2405.14475, 2024. 3", + "[17] Ruiqi Gao, Aleksander Holynski, Philipp Henzler, Arthur Brussee, Ricardo Martin-Brualla, Pratul Srinivasan, Jonathan T Barron, and Ben Poole. Cat3d: Create anything in 3d with multi-view diffusion models. arXiv preprint arXiv:2405.10314, 2024. 3", + "[18] Shenyuan Gao, Jiazhi Yang, Li Chen, Kashyap Chitta, Yihang Qiu, Andreas Geiger, Jun Zhang, and Hongyang Li. Vista: A generalizable driving world model with high fidelity and versatile controllability. arXiv preprint arXiv:2405.17398, 2024. 2, 3", + "[19] Rohit Girdhar, Mannat Singh, Andrew Brown, Quentin Duval, Samaneh Azadi, Sai Saketh Rambhatla, Akbar Shah, Xi Yin, Devi Parikh, and Ishan Misra. Emu video: Factorizing text-to-video generation by explicit image conditioning. arXiv preprint arXiv:2311.10709, 2023. 3", + "[20] Jianfei Guo, Nianchen Deng, Xinyang Li, Yeqi Bai, Botian Shi, Chiyu Wang, Chenjing Ding, Dongliang Wang, and Yikang Li. Streetsurf: Extending multi-view implicit surface reconstruction to street views. arXiv preprint arXiv:2306.04988, 2023. 2", + "[21] Agrim Gupta, Lijun Yu, Kihyuk Sohn, Xiuye Gu, Meera Hahn, Li Fei-Fei, Irfan Essa, Lu Jiang, and José Lezama. Photorealistic video generation with diffusion models. arXiv preprint arXiv:2312.06662, 2023. 3", + "[22] Huasong Han, Kaixuan Zhou, Xiaoxiao Long, Yusen Wang, and Chunxia Xiao. Ggs: Generalizable gaussian splatting for lane switching in autonomous driving. arXiv preprint arXiv:2409.02382, 2024. 2, 3", + "[23] Ayaan Haque, Matthew Tancik, Alexei A Efros, Aleksander Holynski, and Angjoo Kanazawa. Instruct-nerf2nerf: Editing 3d scenes with instructions. In CVPR, 2023. 2", + "[24] Martin Heusel, Hubert Ramsauer, Thomas Unterthiner, Bernhard Nessler, and Sepp Hochreiter. Gans trained by a two time-scale update rule converge to a local nash equilibrium. NeurIPS, 2017. 6, 7", + "[25] Jonathan Ho, William Chan, Chitwan Sahara, Jay Whang, Ruiqi Gao, Alexey Gritsenko, Diederik P Kingma, Ben Poole, Mohammad Norouzi, David J Fleet, et al. Imagen video: High definition video generation with diffusion models. arXiv preprint arXiv:2210.02303, 2022. 3", + "[26] Jonathan Ho, Tim Salimans, Alexey Gritsenko, William Chan, Mohammad Norouzi, and David J Fleet. Video diffusion models. NeurIPS, 2022.", + "[27] Wenyi Hong, Ming Ding, Wendi Zheng, Xinghan Liu, and Jie Tang. Cogvideo: Large-scale pretraining for text-to-video generation via transformers. arXiv preprint arXiv:2205.15868, 2022.3", + "[28] Anthony Hu, Lloyd Russell, Hudson Yeo, Zak Murez, George Fedoseev, Alex Kendall, Jamie Shotton, and Gian" + ], + "bbox": [ + 516, + 92, + 905, + 900 + ], + "page_idx": 8 + }, + { + "type": "page_number", + "text": "12023", + "bbox": [ + 480, + 944, + 517, + 955 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "luca Corrado. Gaia-1: A generative world model for autonomous driving. arXiv preprint arXiv:2309.17080, 2023. 2, 3", + "[29] Shengchao Hu, Li Chen, Penghao Wu, Hongyang Li, Junchi Yan, and Dacheng Tao. St-p3: End-to-end vision-based autonomous driving via spatial-temporal feature learning. In ECCV, 2022. 2", + "[30] Yihan Hu, Jiazhi Yang, Li Chen, Keyu Li, Chonghao Sima, Xizhou Zhu, Siqi Chai, Senyao Du, Tianwei Lin, Wenhai Wang, et al. Planning-oriented autonomous driving. In CVPR, 2023. 2", + "[31] Nan Huang, Xiaobao Wei, Wenzhao Zheng, Pengju An, Ming Lu, Wei Zhan, Masayoshi Tomizuka, Kurt Keutzer, and Shanghang Zhang. $s^3$ gaussian: Self-supervised street gaussians for autonomous driving. arXiv preprint arXiv:2405.20323, 2024. 1, 2, 5, 6, 7", + "[32] Sungwon Hwang, Min-Jung Kim, Taewoong Kang, Jayeon Kang, and Jaegul Choo. Vegs: View extrapolation of urban scenes in 3d gaussian splatting using learned priors. In ECCV, 2024. 2", + "[33] Muhammad Zubair Irshad, Sergey Zakharov, Katherine Liu, Vitor Guizilini, Thomas Kollar, Adrien Gaidon, Zsolt Kira, and Rares Ambrus. Neo 360: Neural fields for sparse view synthesis of outdoor scenes. In ICCV, 2023. 2", + "[34] Bo Jiang, Shaoyu Chen, Qing Xu, Bencheng Liao, Jiajie Chen, Helong Zhou, Qian Zhang, Wenyu Liu, Chang Huang, and Xinggang Wang. Vad: Vectorized scene representation for efficient autonomous driving. In ICCV, 2023. 2", + "[35] Glenn Jocher and Jing Qiu. Ultralytics yolo11, 2024. 7", + "[36] Bernhard Kerbl, Georgios Kopanas, Thomas Leimkuhler, and George Drettakis. 3d gaussian splatting for real-time radiance field rendering. ACM ToG, 2023. 2, 3, 6", + "[37] Diederik P Kingma and Jimmy Ba. Adam: A method for stochastic optimization. arXiv preprint arXiv:1412.6980, 2014. 6", + "[38] Dan Kondratyuk, Lijun Yu, Xiuye Gu, José Lezama, Jonathan Huang, Rachel Hornung, Hartwig Adam, Hassan Akbari, Yair Alon, Vighnesh Birodkar, et al. Videopoet: A large language model for zero-shot video generation. arXiv preprint arXiv:2312.14125, 2023. 3", + "[39] Abhijit Kundu, Kyle Genova, Xiaoqi Yin, Alireza Fathi, Caroline Pantofaru, Leonidas J Guibas, Andrea Tagliasacchi, Frank Dellaert, and Thomas Funkhouser. Panoptic neural fields: A semantic object-aware neural scene representation. In CVPR, 2022. 2", + "[40] Yann LeCun and Courant. A path towards autonomous machine intelligence version 0.9.2, 2022-06-27. 2022. 3", + "[41] Hao Li, Ming Yuan, Yan Zhang, Chenming Wu, Chen Zhao, Chunyu Song, Haocheng Feng, Errui Ding, Dingwen Zhang, and Jingdong Wang. Xld: A cross-lane dataset for benchmarking novel driving view synthesis. arXiv preprint arXiv:2406.18360, 2024. 2", + "[42] Zhengqi Li, Simon Niklaus, Noah Snavely, and Oliver Wang. Neural scene flow fields for space-time view synthesis of dynamic scenes. In CVPR, 2021. 2", + "[43] Zhiqi Li, Zhiding Yu, Shiyi Lan, Jiahan Li, Jan Kautz, Tong Lu, and Jose M Alvarez. Is ego status all you need for open-loop end-to-end autonomous driving? In CVPR, 2024. 2" + ], + "bbox": [ + 91, + 90, + 482, + 900 + ], + "page_idx": 9 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[44] Zhuopeng Li, Yilin Zhang, Chenming Wu, Jianke Zhu, and Liangjun Zhang. Ho-gaussian: Hybrid optimization of 3d gaussian splatting for urban scenes. arXiv preprint arXiv:2403.20032, 2024. 2", + "[45] Chen-Hsuan Lin, Jun Gao, Luming Tang, Towaki Takikawa, Xiaohui Zeng, Xun Huang, Karsten Kreis, Sanja Fidler, Ming-Yu Liu, and Tsung-Yi Lin. Magic3d: High-resolution text-to-3d content creation. In CVPR, 2023. 3", + "[46] Haotong Lin, Sida Peng, Zhen Xu, Yunzhi Yan, Qing Shuai, Hujun Bao, and Xiaowei Zhou. Efficient neural radiance fields for interactive free-viewpoint video. In SIGGRAPH Asia, 2022. 2", + "[47] Fan Lu, Yan Xu, Guang Chen, Hongsheng Li, Kwan-Yee Lin, and Changjun Jiang. Urban radiance field representation with deformable neural mesh primitives. In ICCV, 2023. 2", + "[48] Xin Ma, Yaohui Wang, Gengyun Jia, Xinyuan Chen, Zwei Liu, Yuan-Fang Li, Cunjian Chen, and Yu Qiao. Latte: Latent diffusion transformer for video generation. arXiv preprint arXiv:2401.03048, 2024. 3", + "[49] Ben Mildenhall, Pratul P Srinivasan, Matthew Tancik, Jonathan T Barron, Ravi Ramamoorthi, and Ren Ng. Nerf: Representing scenes as neural radiance fields for view synthesis. CACM, 2021. 2", + "[50] Thomas Müller, Alex Evans, Christoph Schied, and Alexander Keller. Instant neural graphics primitives with a multiresolution hash encoding. ACM ToG, 2022. 2", + "[51] Julian Ost, Fahim Mannan, Nils Thuerey, Julian Knodt, and Felix Heide. Neural scene graphs for dynamic scenes. In CVPR, 2021. 2", + "[52] Keunhong Park, Utkarsh Sinha, Peter Hedman, Jonathan T Barron, Sofien Bouaziz, Dan B Goldman, Ricardo MartinBrualla, and Steven M Seitz. Hypernerf: A higher-dimensional representation for topologically varying neural radiance fields. arXiv preprint arXiv:2106.13228, 2021. 2", + "[53] Dustin Podell, Zion English, Kyle Lacey, Andreas Blattmann, Tim Dockhorn, Jonas Müller, Joe Penna, and Robin Rombach. Sdxl: Improving latent diffusion models for high-resolution image synthesis. arXiv preprint arXiv:2307.01952, 2023. 3", + "[54] Ben Poole, Ajay Jain, Jonathan T Barron, and Ben Mildenhall. Dreamfusion: Text-to-3d using 2d diffusion. arXiv preprint arXiv:2209.14988, 2022. 3", + "[55] Aditya Ramesh, Prafulla Dhariwal, Alex Nichol, Casey Chu, and Mark Chen. Hierarchical text-conditional image generation with clip latents. arXiv preprint arXiv:2204.06125, 2022.3", + "[56] Konstantinos Rematas, Andrew Liu, Pratul P Srinivasan, Jonathan T Barron, Andrea Tagliasacchi, Thomas Funkhouser, and Vittorio Ferrari. Urban radiance fields. In CVPR, 2022. 2", + "[57] Robin Rombach, Andreas Blattmann, Dominik Lorenz, Patrick Esser, and Björn Ommher. High-resolution image synthesis with latent diffusion models. In CVPR, 2022. 3", + "[58] Chitwan Sahara, William Chan, Saurabh Saxena, Lala Li, Jay Whang, Emily L Denton, Kamyar Ghasemipour, Raphael Gontijo Lopes, Burcu Karagol Ayan, Tim Salimans, et al. Photorealistic text-to-image diffusion models with deep language understanding. NeurIPS, 2022. 3" + ], + "bbox": [ + 516, + 90, + 905, + 900 + ], + "page_idx": 9 + }, + { + "type": "page_number", + "text": "12024", + "bbox": [ + 480, + 944, + 519, + 955 + ], + "page_idx": 9 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[59] Kyle Sargent, Zizhang Li, Tanmay Shah, Charles Herrmann, Hong-Xing Yu, Yunzhi Zhang, Eric Ryan Chan, Dmitry Lagun, Li Fei-Fei, Deqing Sun, et al. Zeronvs: Zero-shot 360-degree view synthesis from a single real image. arXiv preprint arXiv:2310.17994, 2023. 3", + "[60] Liangchen Song, Anpei Chen, Zhong Li, Zhang Chen, Lele Chen, Junsong Yuan, Yi Xu, and Andreas Geiger. Nerfplayer: A streamable dynamic scene representation with decomposed neural radiance fields. IEEE TVCG, 2023. 2", + "[61] Pei Sun, Henrik Kretzschmar, Xerxes Dotiwalla, Aurelien Chouard, Vijaysai Patnaik, Paul Tsui, James Guo, Yin Zhou, Yuning Chai, Benjamin Caine, Vijay Vasudevan, Wei Han, Jiquan Ngiam, Hang Zhao, Aleksei Timofeev, Scott Ettinger, Maxim Krivokon, Amy Gao, Aditya Joshi, Yu Zhang, Jonathon Shlens, Zhifeng Chen, and Dragomir Anguelov. Scalability in perception for autonomous driving: Waymo open dataset. In CVPR, 2020. 6, 8", + "[62] Shanlin Sun, Bingbing Zhuang, Ziyu Jiang, Buyu Liu, Xiaohui Xie, and Manmohan Chandraker. Lidarf: Delving into lidar for neural radiance field on street scenes. In CVPR, 2024. 2", + "[63] Matthew Tancik, Vincent Casser, Xinchen Yan, Sabeek Pradhan, Ben Mildenhall, Pratul P Srinivasan, Jonathan T Barron, and Henrik Kretzschmar. Block-nerf: Scalable large scene neural view synthesis. In CVPR, 2022. 2", + "[64] Adam Tonderski, Carl Lindström, Georg Hess, William Ljungbergh, Lennart Svensson, and Christoffer Petersson. Neurad: Neural rendering for autonomous driving. In CVPR, 2024. 2, 6, 7, 8", + "[65] Vikram Voleti, Chun-Han Yao, Mark Boss, Adam Letts, David Pankratz, Dmitry Tochilkin, Christian Laforte, Robin Rombach, and Varun Jampani. Sv3d: Novel multi-view synthesis and 3d generation from a single image using latent video diffusion. arXiv preprint arXiv:2403.12008, 2024. 3", + "[66] Xiaofeng Wang, Zheng Zhu, Guan Huang, Xinze Chen, Jiagang Zhu, and Jiwen Lu. Drivedreamer: Towards real-world-driven world models for autonomous driving. arXiv preprint arXiv:2309.09777, 2023. 2, 3", + "[67] Xiaofeng Wang, Kang Zhao, Feng Liu, Jiayu Wang, Guosheng Zhao, Xiaoyi Bao, Zheng Zhu, Yingya Zhang, and Xingang Wang. Egovid-5m: A large-scale video-action dataset for egocentric video generation. arXiv preprint arXiv:2411.08380, 2024. 3", + "[68] Xiaofeng Wang, Zheng Zhu, Guan Huang, Boyuan Wang, Xinze Chen, and Jiwen Lu. Worlddreamer: Towards general world models for video generation via predicting masked tokens. arXiv preprint arXiv:2401.09985, 2024. 2, 3", + "[69] Yuqi Wang, Jiawei He, Lue Fan, Hongxin Li, Yuntao Chen, and Zhaoxiang Zhang. Driving into the future: Multiview visual forecasting and planning with world model for autonomous driving. In CVPR, 2024. 2, 3", + "[70] Rundi Wu, Ben Mildenhall, Philipp Henzler, Keunhong Park, Ruiqi Gao, Daniel Watson, Pratul P Srinivasan, Dor Verbin, Jonathan T Barron, Ben Poole, et al. Reconfusion: 3d reconstruction with diffusion priors. In CVPR, 2024. 3", + "[71] Zirui Wu, Tianyu Liu, Liyi Luo, Zhide Zhong, Jianteng Chen, Hongmin Xiao, Chao Hou, Haozhe Lou, Yuantao" + ], + "bbox": [ + 91, + 90, + 480, + 900 + ], + "page_idx": 10 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "Chen, Runyi Yang, et al. Mars: An instance-aware, modular and realistic simulator for autonomous driving. In ICAI, 2023. 2", + "[72] Jiannan Xiang, Guangyi Liu, Yi Gu, Qiyue Gao, Yuting Ning, Yuheng Zha, Zeyu Feng, Tianhua Tao, Shibo Hao, Yemin Shi, et al. Pandora: Towards general world model with natural language actions and video states. arXiv preprint arXiv:2406.09455, 2024. 3", + "[73] Pengchuan Xiao, Zhenlei Shao, Steven Hao, Zishuo Zhang, Xiaolin Chai, Judy Jiao, Zesong Li, Jian Wu, Kai Sun, Kun Jiang, et al. Pandaset: Advanced sensor suite dataset for autonomous driving. In ITSC, 2021. 6, 8", + "[74] Ziyang Xie, Junge Zhang, Wenye Li, Feihu Zhang, and Li Zhang. S-nerf: Neural radiance fields for street views. arXiv preprint arXiv:2303.00749, 2023. 2", + "[75] Wilson Yan, Yunzhi Zhang, Pieter Abbeel, and Aravind Srinivas. Videogpt: Video generation using vq-vae and transformers. arXiv preprint arXiv:2104.10157, 2021. 3", + "[76] Yunzhi Yan, Haotong Lin, Chenxu Zhou, Weijie Wang, Haiyang Sun, Kun Zhan, Xianpeng Lang, Xiaowei Zhou, and Sida Peng. Street gaussians for modeling dynamic urban scenes. arXiv preprint arXiv:2401.01339, 2024. 2", + "[77] Jiawei Yang, Boris Ivanovic, Or Litany, Xinshuo Weng, Seung Wook Kim, Boyi Li, Tong Che, Danfei Xu, Sanja Fidler, Marco Pavone, et al. Emernerf: Emergent spatial-temporal scene decomposition via self-supervision. arXiv preprint arXiv:2311.02077, 2023. 2", + "[78] Xuemeng Yang, Licheng Wen, Yukai Ma, Jianbiao Mei, Xin Li, Tiantian Wei, Wenjie Lei, Daocheng Fu, Pinlong Cai, Min Dou, Botian Shi, Liang He, Yong Liu, and Yu Qiao. Drivearena: A closed-loop generative simulation platform for autonomous driving. arXiv preprint arXiv:2408.00415, 2024.3", + "[79] Ze Yang, Yun Chen, Jingkang Wang, Sivabalan Manivasagam, Wei-Chiu Ma, Anqi Joyce Yang, and Raquel Urtasun. Unisim: A neural closed-loop sensor simulator. In CVPR, 2023. 2, 6, 7, 8", + "[80] Ziyi Yang, Xinyu Gao, Wen Zhou, Shaohui Jiao, Yuqing Zhang, and Xiaogang Jin. Deformable 3d gaussians for high-fidelity monocular dynamic scene reconstruction. In CVPR, 2024. 1, 2, 6, 7", + "[81] Zhuoyi Yang, Jiayan Teng, Wendi Zheng, Ming Ding, Shiyu Huang, Jiazheng Xu, Yuanming Yang, Wenyi Hong, Xiaohan Zhang, Guanyu Feng, et al. Cogvideox: Text-to-video diffusion models with an expert transformer. arXiv preprint arXiv:2408.06072, 2024. 3", + "[82] Wang Yifan, Felice Serena, Shihao Wu, Cengiz Öztireli, and Olga Sorkine-Hornung. Differentiable surface splatting for point-based geometry processing. ACM TOG, 2019. 3, 5", + "[83] Zehao Yu, Anpei Chen, Binbin Huang, Torsten Sattler, and Andreas Geiger. Mip-splatting: Alias-free 3d gaussian splattering. In CVPR, 2024. 2", + "[84] Zhongrui Yu, Haoran Wang, Jinze Yang, Hanzhang Wang, Zeke Xie, Yunfeng Cai, Jiale Cao, Zhong Ji, and Mingming Sun. Sgd: Street view synthesis with gaussian splatting and diffusion prior. arXiv preprint arXiv:2403.20079, 2024. 2, 3", + "[85] Yan Zeng, Guoqiang Wei, Jiani Zheng, Jiaxin Zou, Yang Wei, Yuchen Zhang, and Hang Li. Make pixels" + ], + "bbox": [ + 516, + 92, + 903, + 900 + ], + "page_idx": 10 + }, + { + "type": "page_number", + "text": "12025", + "bbox": [ + 480, + 944, + 517, + 955 + ], + "page_idx": 10 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "dance: High-dynamic video generation. arXiv preprint arXiv:2311.10982, 2023. 3", + "[86] Jiang-Tian Zhai, Ze Feng, Jinhao Du, Yongqiang Mao, Jiang-Jiang Liu, Zichang Tan, Yifu Zhang, Xiaoqing Ye, and Jingdong Wang. Rethinking the open-loop evaluation of end-to-end autonomous driving in nuscenes. arXiv preprint arXiv:2305.10430, 2023. 2", + "[87] Guosheng Zhao, Xiaofeng Wang, Zheng Zhu, Xinze Chen, Guan Huang, Xiaoyi Bao, and Xingang Wang. Drivedreamer-2: Llm-enhanced world models for diverse driving video generation. arXiv preprint arXiv:2403.06845, 2024. 2, 3, 5", + "[88] Hongyu Zhou, Jiahao Shao, Lu Xu, Dongfeng Bai, Weichao Qiu, Bingbing Liu, Yue Wang, Andreas Geiger, and Yiyi Liao. Hugs: Holistic urban 3d scene understanding via gaussian splatting. In CVPR, 2024. 2", + "[89] Xiaoyu Zhou, Zhiwei Lin, Xiaojun Shan, Yongtao Wang, Deqing Sun, and Ming-Hsuan Yang. Drivinggaussian: Composite gaussian splatting for surrounding dynamic autonomous driving scenes. In CVPR, pages 21634-21643, 2024. 2", + "[90] Zheng Zhu, Xiaofeng Wang, Wangbo Zhao, Chen Min, Ni-anchen Deng, Min Dou, Yuqi Wang, Botian Shi, Kai Wang, Chi Zhang, et al. Is sora a world simulator? a comprehensive survey on general world models and beyond. arXiv preprint arXiv:2405.03520, 2024. 3" + ], + "bbox": [ + 91, + 90, + 483, + 457 + ], + "page_idx": 11 + }, + { + "type": "page_number", + "text": "12026", + "bbox": [ + 480, + 945, + 519, + 955 + ], + "page_idx": 11 + } +] \ No newline at end of file diff --git a/2025/DriveDreamer4D_ World Models Are Effective Data Machines for 4D Driving Scene Representation/aa9a31fd-776c-4a2a-a4f7-87c23be589e2_model.json b/2025/DriveDreamer4D_ World Models Are Effective Data Machines for 4D Driving Scene Representation/aa9a31fd-776c-4a2a-a4f7-87c23be589e2_model.json new file mode 100644 index 0000000000000000000000000000000000000000..e97b2f52b0ab421332c6f4b0f40d185450d724dc --- /dev/null +++ b/2025/DriveDreamer4D_ World Models Are Effective Data Machines for 4D Driving Scene Representation/aa9a31fd-776c-4a2a-a4f7-87c23be589e2_model.json @@ -0,0 +1,2600 @@ +[ + [ + { + "type": "header", + "bbox": [ + 0.107, + 0.003, + 0.182, + 0.043 + ], + "angle": 0, + "content": "CVF" + }, + { + "type": "header", + "bbox": [ + 0.237, + 0.0, + 0.812, + 0.047 + ], + "angle": 0, + "content": "This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore." + }, + { + "type": "title", + "bbox": [ + 0.185, + 0.131, + 0.817, + 0.177 + ], + "angle": 0, + "content": "DriveDreamer4D: World Models Are Effective Data Machines for 4D Driving Scene Representation" + }, + { + "type": "text", + "bbox": [ + 0.186, + 0.203, + 0.812, + 0.257 + ], + "angle": 0, + "content": "Guosheng Zhao\\(^{1,2,3*}\\), Chaojun Ni\\(^{4*}\\), Xiaofeng Wang\\(^{1,2,3*}\\), Zheng Zhu\\(^{5*†}\\), Xueyang Zhang\\(^{6}\\), Yida Wang\\(^{6}\\), Guan Huang\\(^{5}\\), Xinze Chen\\(^{5}\\), Boyuan Wang\\(^{1,2,3}\\), Youyi Zhang\\(^{7}\\), Wenjun Mei\\(^{4}\\), Xingang Wang\\(^{2,3†}\\)" + }, + { + "type": "text", + "bbox": [ + 0.249, + 0.257, + 0.749, + 0.274 + ], + "angle": 0, + "content": "\\(^{1}\\)Institute of Automation, Chinese Academy of Sciences, China" + }, + { + "type": "text", + "bbox": [ + 0.164, + 0.275, + 0.835, + 0.292 + ], + "angle": 0, + "content": "\\(^{2}\\)School of Artificial Intelligence, University of Chinese Academy of Sciences, China" + }, + { + "type": "text", + "bbox": [ + 0.143, + 0.292, + 0.856, + 0.31 + ], + "angle": 0, + "content": "\\(^{3}\\)Luoyang Institute for Robot and Intelligent Equipment, China \\(^{4}\\)Peking University, China" + }, + { + "type": "text", + "bbox": [ + 0.164, + 0.31, + 0.825, + 0.328 + ], + "angle": 0, + "content": "\\(^{5}\\)GigaAI, China \\(^{6}\\)Li Auto Inc., China \\(^{7}\\)Technical University of Munich, Germany" + }, + { + "type": "image", + "bbox": [ + 0.102, + 0.364, + 0.403, + 0.543 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.406, + 0.364, + 0.57, + 0.45 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.571, + 0.364, + 0.734, + 0.45 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.736, + 0.364, + 0.901, + 0.45 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.405, + 0.456, + 0.569, + 0.543 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.572, + 0.456, + 0.735, + 0.543 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.737, + 0.456, + 0.901, + 0.543 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.089, + 0.555, + 0.908, + 0.598 + ], + "angle": 0, + "content": "Figure 1. Previous 4D Gaussian Splatting methods (e.g., PVG [9], \\(\\mathbf{S}^3\\) Gaussian [31], Deformable-GS [80]) face challenges in rendering novel trajectories, such as lane change. DriveDreamer4D addresses this by enhancing 4D driving scene representation via integrating priors from world models, significantly improving rendering quality under complex scenarios and novel trajectory viewpoints." + }, + { + "type": "title", + "bbox": [ + 0.248, + 0.61, + 0.327, + 0.626 + ], + "angle": 0, + "content": "Abstract" + }, + { + "type": "text", + "bbox": [ + 0.089, + 0.642, + 0.486, + 0.854 + ], + "angle": 0, + "content": "Closed-loop simulation is essential for advancing end-to-end autonomous driving systems. Contemporary sensor simulation methods, such as NeRF and 3DGS, rely predominantly on conditions closely aligned with training data distributions, which are largely confined to forward-driving scenarios. Consequently, these methods face limitations when rendering complex maneuvers (e.g., lane change, acceleration, deceleration). Recent advancements in autonomous-driving world models have demonstrated the potential to generate diverse driving videos. However, these approaches remain constrained to 2D video generation, inherently lacking the spatiotemporal coherence required to capture intricacies of dynamic driving environments. In this paper, we introduce DriveDreamer4D, which" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.612, + 0.909, + 0.883 + ], + "angle": 0, + "content": "enhances 4D driving scene representation leveraging world model priors. Specifically, we utilize the world model as a data machine to synthesize novel trajectory videos, where structured conditions are explicitly leveraged to control the spatial-temporal consistency of traffic elements. Besides, the cousin data training strategy is proposed to facilitate merging real and synthetic data for optimizing 4DGS. To our knowledge, DriveDreamer4D is the first to utilize video generation models for improving 4D reconstruction in driving scenarios. Experimental results reveal that DriveDreamer4D significantly enhances generation quality under novel trajectory views, achieving a relative improvement in FID by \\(32.1\\%\\), \\(46.4\\%\\), and \\(16.3\\%\\) compared to PVG, \\(S^3\\) Gaussian, and Deformable-GS. Moreover, DriveDreamer4D markedly enhances the spatiotemporal coherence of driving agents, which is verified by a comprehensive user study and the relative increases of \\(22.6\\%\\), \\(43.5\\%\\), and \\(15.6\\%\\) in the NTA-IoU metric." + }, + { + "type": "page_footnote", + "bbox": [ + 0.108, + 0.863, + 0.485, + 0.901 + ], + "angle": 0, + "content": "*These authors contributed equally to this work. \n†Corresponding authors. zhengzhu@ieee.org, xingang.wang@ia.ac.cn. \n‡Project Page: https://drivedreamer4d.github.io" + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.52, + 0.958 + ], + "angle": 0, + "content": "12015" + } + ], + [ + { + "type": "title", + "bbox": [ + 0.093, + 0.09, + 0.223, + 0.107 + ], + "angle": 0, + "content": "1. Introduction" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.115, + 0.482, + 0.282 + ], + "angle": 0, + "content": "End-to-end planning [29, 30, 34], which directly maps sensor inputs to control signals, is among the most critical and promising tasks in autonomous driving. However, current open-loop evaluations are inadequate for accurately assessing end-to-end planning algorithms, highlighting an urgent need for enhanced evaluation methods [41, 43, 86]. A compelling solution lies in closed-loop evaluations within real-world scenarios, which require retrieving sensor data from arbitrarily specified viewpoints. This necessitates constructing a 4D driving scene representation capable of reconstructing complex, dynamic driving environments." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.282, + 0.483, + 0.614 + ], + "angle": 0, + "content": "Closed-loop simulation in driving environments predominantly relies on scene reconstruction techniques such as Neural Radiance Fields (NeRF) [20, 49, 77, 79] and 3D Gaussian Splatting (3DGS) [11, 31, 36, 76], which are inherently limited by the density of input data. Specifically, these methods can render scenes effectively only under conditions closely aligned with their training data distributions—primarily forward-driving scenarios—and struggle to perform accurately during complex maneuvers (see Fig. 1). To mitigate these limitations, methods like SGD [84] and GGS [22] leverage generative models to extend the range of training viewpoints. However, these approaches primarily supplement sparse image data or static background elements, falling short of modeling the intricacies of dynamic, interactive driving scenes. Recently, advancements in autonomous driving world models [18, 28, 66, 68, 69, 87] have introduced the capability to generate diverse, command-aligned video viewpoints, offering renewed promise for closed-loop simulation in autonomous driving. Nonetheless, these models remain constrained to 2D videos, lacking the spatial-temporal coherence essential for accurately modeling complex driving scenarios." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.614, + 0.483, + 0.901 + ], + "angle": 0, + "content": "In this paper, we introduce DriveDreamer4D, which improves 4D driving scene representation by integrating priors from autonomous driving world models. Our approach utilizes an autonomous driving world model [87] as a generative engine, synthesizing novel trajectory video data that densifies real-world driving datasets for enhanced training. Notably, we propose the Novel Trajectory Generation Module (NTGM) to generate diverse structured traffic conditions, and DriveDreamer4D applies these conditions to independently regulate the motion dynamics of foreground and background elements in complex driving environments. These conditions undergo view projection synchronized with vehicle maneuvers, ensuring that the synthesized data adheres to the spatiotemporal constraints. Subsequently, the Cousin Data Training Strategy (CDTS) is proposed to merge temporal-aligned real and synthetic data for training 4DGS. In CDTS, a regularization loss is further incorporated to ensure perceptual coherence. To the best of our knowledge, DriveDreamer4D is the first" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.092, + 0.905, + 0.318 + ], + "angle": 0, + "content": "framework to harness video generation models for elevating 4D scene reconstruction quality in autonomous driving, providing richly varied viewpoint data for scenarios including lane change, acceleration, and deceleration. As shown in Fig. 1, experiment results demonstrate that DriveDreamer4D significantly enhances generation fidelity for novel trajectory viewpoints, achieving a relative improvement in FID by \\(32.1\\%\\), \\(46.4\\%\\), and \\(16.3\\%\\) compared to PVG [9], \\(\\mathrm{S}^3\\) Gaussian [31], and Deformable-GS [80]. Besides, DriveDreamer4D fortifies the spatiotemporal coherence between foreground and background elements, with respective increases of \\(22.6\\%\\), \\(43.5\\%\\), and \\(15.6\\%\\) in the NTA-IoU metric. Furthermore, a comprehensive user study confirms that the average win rate of DriveDreamer4D exceeds \\(80\\%\\), compared to three baselines." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.319, + 0.905, + 0.532 + ], + "angle": 0, + "content": "The primary contributions of this work are as follows: (1) We present DriveDreamer4D, the first framework to leverage world model priors for advancing 4D scene reconstruction in autonomous driving. (2) The NTGM is proposed to automate the generation of structured conditions, allowing DriveDreamer4D to create novel trajectory videos with complex maneuvers while ensuring spatial-temporal consistency. Additionally, the CDTS is introduced to merge temporal-aligned real and synthetic data for training 4DGS, using a regularization loss to maintain perceptual coherence. (3) We perform comprehensive experiments to validate that DriveDreamer4D notably enhances generation quality across novel trajectory viewpoints, as well as the spatiotemporal coherence of driving scene elements." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.549, + 0.655, + 0.565 + ], + "angle": 0, + "content": "2. Related Work" + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.576, + 0.779, + 0.592 + ], + "angle": 0, + "content": "2.1. Driving Scene Representation" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.599, + 0.905, + 0.901 + ], + "angle": 0, + "content": "NeRF and 3DGS have emerged as leading approaches for 3D scene representation. NeRF models [2, 3, 49, 50] continuous volumetric scenes using multi-layer perceptron (MLP) networks, enabling highly detailed scene reconstructions with remarkable rendering quality. More recently, 3DGS [13, 36, 83] introduces an innovative method by defining a set of anisotropic Gaussians in 3D space, leveraging adaptive density control to achieve high-quality renderings from sparse point cloud inputs. Several works have extended NeRF [12, 20, 23, 33, 47, 56, 62, 63, 77, 79] or 3DGS [9, 11, 14, 31, 32, 44, 76, 84, 88, 89] to autonomous driving scenarios. Given the dynamic nature of driving environments, there has also been significant effort in modeling 4D driving scene representations. Some approaches encode time as an additional input to parameterize 4D scenes [1, 15, 31, 42, 46, 52, 60], while others represent scenes as a composition of moving object models alongside a static background model [39, 51, 64, 71, 74, 79]. Despite these advancements, methods based on NeRF and 3DGS face limitations tied to the density of input data. These tech-" + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.946, + 0.52, + 0.957 + ], + "angle": 0, + "content": "12016" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.09, + 0.092, + 0.483, + 0.138 + ], + "angle": 0, + "content": "niques can only render scenes effectively when sensor data closely matches the training data distribution, which is typically confined to forward-driving scenarios." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.153, + 0.238, + 0.168 + ], + "angle": 0, + "content": "2.2. World Models" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.178, + 0.484, + 0.481 + ], + "angle": 0, + "content": "The world model module predicts possible future world states as a function of imagined action sequences proposed by the actor [40, 90]. Approaches such as [4, 5, 19, 21, 25-27, 38, 48, 67, 68, 72, 75, 81, 85] simulate environments through video generation controlled by free-text actions. At the forefront of this evolution is Sora [6], which leverages advanced generative techniques to produce intricate visual sequences that respect the fundamental laws of physics. This ability to deeply understand and simulate the environment not only improves video generation quality but also has substantial implications for real-world driving scenarios. Autonomous driving world models [18, 28, 66, 69, 78, 87] employ predictive methodologies to interpret driving environments, thereby generating realistic driving scenarios and learning key driving elements and policies from video data. Although these models successfully produce diverse driving video data conditioned on complex driving actions, they remain limited to 2D outputs and lack the spatial-temporal coherence needed to accurately capture the complexities of dynamic driving environments." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.496, + 0.419, + 0.513 + ], + "angle": 0, + "content": "2.3. Diffusion Prior for 3D Representation" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.52, + 0.484, + 0.779 + ], + "angle": 0, + "content": "Constructing comprehensive 3D scenes from limited observations demands generative prior, particularly for unseen areas. Earlier studies distill the knowledge from text-to-image diffusion models [53, 55, 57, 58] into a 3D representation model. Specifically, the Score Distillation Sampling (SDS) [45, 54, 70] is adopted to synthesize a 3D object from the text prompt. Furthermore, to enhance 3D consistency, several approaches extend the multi-view diffusion models [17, 59] and video diffusion models [4, 10, 65] to 3D scene generation. To extend the diffusion prior to complex, dynamic, large-scale driving scenes for 3D reconstruction, methods such as SGD [84], GGS [22] and MagicDrive3D [16] employ generative models to broaden the range of training viewpoints. Nonetheless, these approaches mainly address sparse image data or static background elements, lacking the capacity to fully capture the complexities inherent in the 4D driving environments." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.798, + 0.182, + 0.813 + ], + "angle": 0, + "content": "3. Method" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.826, + 0.484, + 0.902 + ], + "angle": 0, + "content": "In this section, we first elaborate on the preliminaries of 4D driving scene representation and world models for driving video generation. Then we present the details of DriveDreamer4D, which enhances 4D driving scene representation leveraging priors from driving world models." + }, + { + "type": "title", + "bbox": [ + 0.514, + 0.091, + 0.644, + 0.108 + ], + "angle": 0, + "content": "3.1. Preliminary" + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.113, + 0.791, + 0.129 + ], + "angle": 0, + "content": "3.1.1. 4D Driving Scene Representation" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.132, + 0.905, + 0.222 + ], + "angle": 0, + "content": "4DGS models the driving scene with a collection of 3DGS and a temporal field module. Each 3DGS [36] is parameterized by its center position \\( x \\), opacity \\( \\gamma \\), covariance \\( \\Sigma \\), and view-dependent RGB color \\( c \\), controlled via spherical harmonics. For stability, each covariance matrix \\( \\Sigma \\) is decomposed by:" + }, + { + "type": "equation", + "bbox": [ + 0.651, + 0.221, + 0.905, + 0.239 + ], + "angle": 0, + "content": "\\[\n\\boldsymbol {\\Sigma} = \\boldsymbol {R} \\boldsymbol {S} \\boldsymbol {S} ^ {T} \\boldsymbol {R} ^ {T}, \\tag {1}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.246, + 0.907, + 0.368 + ], + "angle": 0, + "content": "where scaling matrix \\(S\\) and a rotation matrix \\(\\pmb{R}\\) are learnable parameters, represented by scaling \\(\\pmb{s}\\) and quaternion \\(\\pmb{r}\\). All trainable parameters of a single 3D Gaussian are collectively denoted as \\(\\phi = \\{\\pmb{x},\\pmb{\\gamma},\\pmb{s},\\pmb{r},\\pmb{c}\\}\\). The temporal field \\(\\mathcal{F}\\) takes \\(\\phi\\) and a time step \\(t_{gs}\\) as input, outputting the offset \\(\\delta \\phi = \\{\\delta x,\\delta \\gamma ,\\delta s,\\delta r,\\delta c\\}\\) for each Gaussian relative to canonical space. The 4D Gaussian \\(\\phi^{\\prime} = \\{\\pmb{x}^{\\prime},\\pmb{\\gamma}^{\\prime},\\pmb{s}^{\\prime},\\pmb{r}^{\\prime},\\pmb{c}^{\\prime}\\}\\) is then computed by:" + }, + { + "type": "equation", + "bbox": [ + 0.601, + 0.381, + 0.905, + 0.4 + ], + "angle": 0, + "content": "\\[\n\\phi^ {\\prime} = \\phi + \\delta \\phi = \\phi + \\mathcal {F} (\\phi , \\boldsymbol {t} _ {g s}). \\tag {2}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.405, + 0.906, + 0.497 + ], + "angle": 0, + "content": "Following [82], a differentiable Gaussian Splating renderer is employed to project 4D Gaussians \\(\\phi\\) into camera coordinates, yielding the covariance matrix \\(\\boldsymbol{\\Sigma}^{\\prime} = J\\boldsymbol{V}\\boldsymbol{\\Sigma}\\boldsymbol{V}^{T}\\boldsymbol{J}^{T}\\), where \\(\\boldsymbol{J}\\) is the Jacobian matrix of the perspective projection, and \\(\\boldsymbol{V}\\) is the transform matrix. The color of each pixel is calculated by \\(N\\) ordered points using \\(\\alpha\\)-blending:" + }, + { + "type": "equation", + "bbox": [ + 0.651, + 0.507, + 0.905, + 0.54 + ], + "angle": 0, + "content": "\\[\nC = \\sum_ {i \\in N} T _ {i} \\boldsymbol {c} _ {i} ^ {\\prime} \\alpha_ {i}, \\tag {3}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.553, + 0.906, + 0.645 + ], + "angle": 0, + "content": "where \\(T_{i}\\) is the transmittance defined by \\(\\prod_{j = 1}^{i - 1}(1 - \\alpha_j)\\), \\(c_{i}^{\\prime}\\) denotes the color of each point, \\(\\alpha_{i}\\) is given by evaluating a 2D Gaussian with covariance \\(\\Sigma^{\\prime}\\) multiplied with a learned per-point opacity \\(\\gamma_{i}^{\\prime}\\). The trainable parameters \\(\\phi^{\\prime}\\) can be optimized by a combination of RGB loss, depth loss and SSIM loss:" + }, + { + "type": "equation", + "bbox": [ + 0.543, + 0.654, + 0.905, + 0.693 + ], + "angle": 0, + "content": "\\[\n\\begin{array}{l} \\mathcal {L} _ {\\mathrm {o r i}} \\left(\\phi^ {\\prime}\\right) = \\lambda_ {1} \\| \\hat {I} _ {\\mathrm {o r i}} - I _ {\\mathrm {o r i}} \\| _ {1} + \\lambda_ {2} \\| \\hat {D} _ {\\mathrm {o r i}} - D _ {\\mathrm {o r i}} \\| _ {1} \\tag {4} \\\\ + \\lambda_ {3} \\operatorname {S S I M} \\left(\\bar {I} _ {\\text {o r i}}, I _ {\\text {o r i}}\\right), \\\\ \\end{array}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.707, + 0.906, + 0.785 + ], + "angle": 0, + "content": "where \\(\\hat{I}_{\\mathrm{ori}}\\) and \\(I_{\\mathrm{ori}}\\) represent the rendered image and the ground truth image. \\(\\hat{D}_{\\mathrm{ori}}\\) and \\(D_{\\mathrm{ori}}\\) are the rendered depth and the ground truth LiDAR depth map. SSIM \\((\\cdot)\\) refers to the operation of the Structural Similarity Index Measure, and \\(\\lambda_1,\\lambda_2,\\lambda_3\\) are the loss weights." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.791, + 0.868, + 0.807 + ], + "angle": 0, + "content": "3.1.2. World Models for Driving Video Generation" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.81, + 0.906, + 0.902 + ], + "angle": 0, + "content": "The world model module predicts possible future world states based on imagined action sequences [40]. Autonomous-driving world models [18, 66, 69, 87], typically based on diffusion models, leverage structured driving information or action controls to guide future video prediction. During training, these models first encode" + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "12017" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.099, + 0.093, + 0.9, + 0.471 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.09, + 0.486, + 0.907, + 0.543 + ], + "angle": 0, + "content": "Figure 2. The overall framework of DriveDreamer4D. Initially, by altering the actions of the original trajectory (e.g., steering angle, speed), new trajectories can be obtained. Conditioned on the first frame and the structured information (3D bounding boxes, HDMap) from the new trajectory, the novel trajectory videos are generated. Subsequently, the temporal-aligned cousin pair (original and novel trajectory videos) are merged to optimize the 4D Gaussian Splatting model, where a regularization loss is calculated to ensure perceptual coherence." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.569, + 0.483, + 0.63 + ], + "angle": 0, + "content": "videos \\(\\pmb{v}\\) into a lower-dimensional latent space \\(z = \\mathcal{E}(\\pmb{v})\\) using a variational encoder \\(\\mathcal{E}\\). After adding noise \\(\\epsilon_{t}\\) to the latent, the diffusion model learns a denoising process. This diffusion process is optimized by:" + }, + { + "type": "equation", + "bbox": [ + 0.132, + 0.639, + 0.483, + 0.665 + ], + "angle": 0, + "content": "\\[\n\\mathcal {L} _ {d i f f} = \\mathbb {E} _ {\\boldsymbol {z}, \\epsilon \\sim \\mathcal {N} (0, 1), t} \\left[ \\| \\epsilon_ {t} - \\epsilon_ {\\theta} (\\boldsymbol {z} _ {t}, t, \\boldsymbol {f}) \\| _ {2} ^ {2} \\right], \\tag {5}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.674, + 0.483, + 0.84 + ], + "angle": 0, + "content": "where \\(\\epsilon_{\\theta}\\) is a parameterized denoising network, \\(t\\) denotes the time step, representing the level of noise added or removed at each stage. Additionally, to improve the controllability of the generated data, conditional features \\(f\\) (e.g., reference images, speed, steering angle, scene layouts, camera poses and textual information) can be introduced into the reverse diffusion process, ensuring that the generated outputs adhere to the input control signals. During inference, the world models can be conditioned on a reference image to control the style of the output scene, while predicting the future world states contingent upon the other input actions." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.849, + 0.262, + 0.863 + ], + "angle": 0, + "content": "3.2. DriveDreamer4D" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.871, + 0.483, + 0.903 + ], + "angle": 0, + "content": "The overall pipeline of DriveDreamer4D is depicted in Fig. 2. In the upper part, the Novel Trajectory Generation" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.569, + 0.907, + 0.78 + ], + "angle": 0, + "content": "Module (NTGM) is proposed to adjust driving actions (e.g., steering angle, speed) to generate new trajectories. These novel trajectories provide new perspectives for extracting structured information like 3D boxes and HDMap. Subsequently, a controllable video diffusion model synthesizes videos from these updated viewpoints, incorporating specific priors associated with the modified trajectories. In the lower part, the Cousin Data Training Strategy (CDTS) is introduced to combine the temporal-aligned original and generated data for optimizing the 4DGS model, where a regularization loss is calculated to impose perceptual coherence. In the following sections, we delve into the details of novel trajectories video generation and then introduce the CDTS for 4D reconstruction." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.79, + 0.806, + 0.807 + ], + "angle": 0, + "content": "3.2.1. Novel Trajectory Video Generation" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.811, + 0.907, + 0.901 + ], + "angle": 0, + "content": "As previously mentioned, traditional 4DGS methods are limited in rendering complex maneuvers, largely due to the training data being dominated by straightforward driving scenarios. To overcome this, DriveDreamer4D leverages world model priors to generate diverse viewpoint data, enhancing the 4D scene representation. To achieve this," + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "12018" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.09, + 0.092, + 0.485, + 0.394 + ], + "angle": 0, + "content": "we propose the NTGM, which is designed to create new trajectories that serve as input for the world model, enabling the automated generation of complex maneuver data. NTGM comprises two main components: (1) novel trajectory proposal, (2) trajectory safety assessment. In the novel trajectory proposal stage, text-to-trajectory [87] can be adopted to automatically generate diverse complex trajectories. Additionally, trajectories can be custom-designed to meet specific requirements, allowing for tailored data generation based on precise needs. The overview of the custom-designed trajectory proposal (e.g., lane change) and trajectory safety assessment is shown in the Algo. 1. In a specific driving scenario, the original trajectory in the world coordinate system can be readily acquired as \\( \\mathcal{T}_{\\mathrm{ori}}^{\\mathrm{world}} = \\{p_i^{\\mathrm{world}}\\}_{i=0}^{K} \\), where \\( K \\) denotes the number of frames and \\( p_i^{\\mathrm{world}} \\in \\mathbb{R}^3 \\) refers to the position of the ego-vehicle at the \\( i \\)-th frame. To propose novel trajectories, the original trajectory \\( \\mathcal{T}_{\\mathrm{ori}}^{\\mathrm{world}} \\) is transformed into the ego-vehicle coordinate system of the first frame, denoted as \\( \\mathcal{T}_{\\mathrm{ori}}^{\\mathrm{EgoStart}} \\) and computed as:" + }, + { + "type": "equation", + "bbox": [ + 0.167, + 0.396, + 0.483, + 0.416 + ], + "angle": 0, + "content": "\\[\n\\left[ p _ {i} ^ {\\text {E g o S t a r t}}, 1 \\right] ^ {T} = M _ {0} ^ {- 1} \\times \\left[ p _ {i} ^ {\\text {w o r l d}}, 1 \\right] ^ {T}, \\tag {6}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.424, + 0.484, + 0.639 + ], + "angle": 0, + "content": "where \\(M_0 \\in \\mathbb{R}^{4 \\times 4}\\) represents the transformation matrix from the ego-vehicle coordinate system of the first frame to the world coordinate system, \\([\\cdot]\\) denotes the operation of the concat. In the ego-vehicle coordinate system, the vehicle's heading is aligned with the positive \\(x\\)-axis, the \\(y\\)-axis points to the left side of the vehicle, and the \\(z\\)-axis is oriented vertically upwards, perpendicular to the plane of the vehicle. Consequently, changes in the vehicle's velocity and direction can be respectively represented by adjusting the value along the \\(x\\)-axis and \\(y\\)-axis. A final safety assessment is conducted for the newly generated trajectory points, which includes verifying whether the vehicle trajectories \\(p\\) remain within drivable areas \\(\\mathcal{B}_{\\mathrm{road}}\\) and ensuring that no collisions occur with pedestrians or other vehicles \\(\\{o_j\\}_{j=1}^M\\)." + }, + { + "type": "equation", + "bbox": [ + 0.151, + 0.652, + 0.483, + 0.679 + ], + "angle": 0, + "content": "\\[\np \\in \\mathcal {B} _ {\\text {r o a d}}, \\tag {7}\n\\]" + }, + { + "type": "equation", + "bbox": [ + 0.154, + 0.679, + 0.422, + 0.693 + ], + "angle": 0, + "content": "\\[\n\\| p - o _ {j} \\| \\geq d _ {\\min }, \\quad \\forall j \\in \\{1, \\dots , M \\},\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.705, + 0.484, + 0.842 + ], + "angle": 0, + "content": "where \\( d_{\\mathrm{min}} \\) is the minimal distance between different agents. Once a novel trajectory that complies with traffic regulations is generated, the road structure and 3D bounding boxes can be projected onto the camera view from the perspective of the new trajectory, thereby generating structured information relative to the updated trajectory. This structured information, along with the initial frame and text, is fed into a world model [87] to produce the videos that follow the novel trajectories." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.851, + 0.351, + 0.867 + ], + "angle": 0, + "content": "3.2.2. Cousin Data Training Strategy" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.871, + 0.484, + 0.902 + ], + "angle": 0, + "content": "To better integrate generated data for training 4DGS, we propose the CDTS. Specifically, we construct temporally" + }, + { + "type": "code_caption", + "bbox": [ + 0.516, + 0.092, + 0.85, + 0.107 + ], + "angle": 0, + "content": "Algorithm 1 Novel Trajectory Generation Module" + }, + { + "type": "algorithm", + "bbox": [ + 0.513, + 0.111, + 0.894, + 0.402 + ], + "angle": 0, + "content": "Input: Trajectory \\(\\mathcal{T}_{\\mathrm{ori}}^{\\mathrm{world}}\\) , Transformation matrix \\(M_0\\) \nOutput: Novel trajectory \\(\\mathcal{T}_{\\mathrm{novel}}^{\\mathrm{ego}}\\) \n\\(\\mathcal{T}_{\\mathrm{novel}}^{\\mathrm{ego}}\\gets [[0,0,0]]\\) \nOffset \\(\\leftarrow 0\\) \nfor each \\(p_{\\mathrm{ori}}^{\\mathrm{world}}\\) in \\(\\mathcal{T}_{\\mathrm{ori}}^{\\mathrm{world}}\\) [1:] do \n\\(p^{\\mathrm{EgoStart}}\\gets \\mathrm{RelativeCoord}(p_{\\mathrm{ori}}^{\\mathrm{world}},M_0)\\) \nMaxOffset \\(\\leftarrow 0.1\\) \nwhile True do \nNewOffset \\(\\leftarrow\\) Offset + RandOffset(0, MaxOffset) \n\\(p^{\\mathrm{EgoStart}'}\\gets p^{\\mathrm{EgoStart}} + [0,\\) NewOffset, 0] \nif SafeCheck \\((p^{\\mathrm{EgoStart}'}\\) ) then \nAddElement \\((\\mathcal{T}_{\\mathrm{novel}}^{\\mathrm{ego}},p^{\\mathrm{EgoStart}'})\\) \nOffset \\(\\leftarrow\\) NewOffset \nbreak \nelse \nMaxOffset \\(\\leftarrow\\) MaxOffset/2 \nend if \nend while \nend for" + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.427, + 0.86, + 0.442 + ], + "angle": 0, + "content": "aligned cousin pair data as a minimal training batch:" + }, + { + "type": "equation", + "bbox": [ + 0.584, + 0.451, + 0.905, + 0.47 + ], + "angle": 0, + "content": "\\[\n\\operatorname {B a t c h S t a c k} \\left(\\left\\{\\hat {I} _ {\\text {o r i}, t} \\right\\} _ {t = 0} ^ {T}, \\left\\{\\hat {I} _ {\\text {n o v e l}, t} \\right\\} _ {t = 0} ^ {T}\\right), \\tag {8}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.48, + 0.907, + 0.631 + ], + "angle": 0, + "content": "where \\(\\mathrm{BatchStack}(\\cdot)\\) is the data processor to stack temporal-aligned \\(\\{\\hat{I}_{\\mathrm{ori},t}\\}_{t=0}^{T}\\) and \\(\\{\\hat{I}_{\\mathrm{novel},t}\\}_{t=0}^{T}\\) into a training batch. By leveraging real and synthetic data aligned at each timestep, CDTS mitigates the data gap in 4DGS training, enhancing the model's ability to learn consistent representations across real and synthetic data. To optimize 4DGS, the temporal-aligned cousin pair is input per step before gradient optimization. The loss function \\(\\mathcal{L}_{\\mathrm{ori}}\\) for the original data is defined in Eq. 4. And the loss function \\(\\mathcal{L}_{\\mathrm{novel}}\\) for the generated data is akin to [9, 31], defined as follows:" + }, + { + "type": "equation", + "bbox": [ + 0.584, + 0.641, + 0.905, + 0.68 + ], + "angle": 0, + "content": "\\[\n\\begin{array}{l} \\mathcal {L} _ {\\text {n o v e l}} \\left(\\phi^ {\\prime}\\right) = \\lambda_ {1} \\| \\hat {I} _ {\\text {n o v e l}} - I _ {\\text {n o v e l}} \\| _ {1} \\tag {9} \\\\ + \\lambda_ {3} \\operatorname {S S I M} \\left(\\hat {I} _ {\\text {n o v e l}}, I _ {\\text {n o v e l}}\\right), \\\\ \\end{array}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.69, + 0.907, + 0.901 + ], + "angle": 0, + "content": "where \\( I_{\\mathrm{novel}} \\) represents the generated images corresponding to the novel trajectories as described in Sec. 3.2.1, and \\( \\hat{I}_{\\mathrm{novel}} \\) denotes the rendered images under the novel trajectories via differentiable splatting [82]. Notably, different from [9, 31], depth maps are not employed as constraints in the optimization of 4DGS when using the generated dataset \\( D_{\\mathrm{novel}} \\). The limitation arises from the fact that LiDAR point cloud data is exclusively collected for the original trajectory. When these LiDAR points are projected onto a new trajectory, it cannot produce a complete depth map for the new perspective, as something visible in the novel trajectory may have been occluded in the original view. Consequently, the incorporation of such depth maps does not facilitate the optimization of the 4DGS model. More details are described in" + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "12019" + } + ], + [ + { + "type": "table", + "bbox": [ + 0.094, + 0.089, + 0.907, + 0.215 + ], + "angle": 0, + "content": "
MethodLane ChangeAccelerationDecelerationAverage
NTA-IoU ↑NTL-IoU ↑NTA-IoU ↑NTL-IoU ↑NTA-IoU ↑NTL-IoU ↑NTA-IoU ↑NTL-IoU ↑
PVG [9]0.25650.700.39653.080.39453.650.34952.48
DriveDreamer4D with PVG0.43853.060.42153.350.42453.890.42853.43
S3Gaussian [31]0.17549.050.43451.930.38452.140.33151.04
DriveDreamer4D with S3Gaussian0.49553.420.48452.630.44552.690.47552.91
Deformable-GS [80]0.24051.620.34652.170.37753.210.32152.33
DriveDreamer4D with Deformable-GS0.33552.930.37152.770.40653.790.37153.16
" + }, + { + "type": "table_caption", + "bbox": [ + 0.098, + 0.224, + 0.899, + 0.24 + ], + "angle": 0, + "content": "Table 1. Comparison of NTA-IoU and NTL-IoU scores across different novel trajectory views (lane change, acceleration, deceleration)." + }, + { + "type": "table", + "bbox": [ + 0.092, + 0.263, + 0.479, + 0.391 + ], + "angle": 0, + "content": "
MethodFID ↓
PVG[9]105.29
DriveDreamer4D with PVG71.52
S3Gaussian[31]124.90
DriveDreamer4D with S3Gaussian66.93
Deformable-GS[80]92.34
DriveDreamer4D with Deformable-GS77.32
" + }, + { + "type": "table_caption", + "bbox": [ + 0.09, + 0.399, + 0.483, + 0.429 + ], + "angle": 0, + "content": "Table 2. Comparison of FID scores in novel trajectory view synthesis (lane change) on the Waymo dataset." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.455, + 0.483, + 0.487 + ], + "angle": 0, + "content": "Sec. 4.3. Additionally, we propose a regularization loss to enhance the perceptual coherence, defined as follows:" + }, + { + "type": "equation", + "bbox": [ + 0.163, + 0.497, + 0.483, + 0.517 + ], + "angle": 0, + "content": "\\[\n\\mathcal {L} _ {\\text {r e g}} \\left(\\phi^ {\\prime}\\right) = \\left\\| \\mathcal {F} _ {p} \\left(\\hat {I} _ {\\text {o r i}}\\right) - \\mathcal {F} _ {p} \\left(\\hat {I} _ {\\text {n o v e l}}\\right) \\right\\| _ {1}, \\tag {10}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.528, + 0.484, + 0.573 + ], + "angle": 0, + "content": "where \\(\\mathcal{F}_p(\\cdot)\\) denotes the perception feature extraction model [24]. The overall loss function for mixed training is defined as follows:" + }, + { + "type": "equation", + "bbox": [ + 0.158, + 0.586, + 0.483, + 0.605 + ], + "angle": 0, + "content": "\\[\n\\mathcal {L} \\left(\\phi^ {\\prime}\\right) = \\mathcal {L} _ {\\text {o r i}} + \\lambda_ {\\text {n o v e l}} \\mathcal {L} _ {\\text {n o v e l}} + \\lambda_ {\\text {r e g}} \\mathcal {L} _ {\\text {r e g}}. \\tag {11}\n\\]" + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.63, + 0.224, + 0.648 + ], + "angle": 0, + "content": "4. Experiments" + }, + { + "type": "text", + "bbox": [ + 0.089, + 0.656, + 0.483, + 0.822 + ], + "angle": 0, + "content": "In this section, we first outline the experimental setup, including the dataset, implementation details, and evaluation metrics. Subsequently, quantitative results across different datasets are provided to demonstrate that the proposed DriveDreamer4D significantly enhances rendering quality for novel trajectory viewpoints and improves the spatiotemporal coherence of foreground and background components. Finally, we conduct an ablation study on the hyperparameter settings, as well as the effects of depth loss and the proposed CDTS, including temporal-aligned pairs, and regularization loss." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.833, + 0.267, + 0.85 + ], + "angle": 0, + "content": "4.1. Experiment Setup" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.856, + 0.484, + 0.903 + ], + "angle": 0, + "content": "Dataset. We conduct experiments using the Waymo dataset [61], known for its comprehensive real-world driving logs. We specifically select eight scenes characterized by highly" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.265, + 0.907, + 0.416 + ], + "angle": 0, + "content": "dynamic interactions, featuring numerous vehicles with diverse relative positions and intricate driving trajectories. Each selected segment contains approximately 40 frames, with segment IDs detailed in the supplement. In addition, we have also conduct experiments on the nuscenes [7], and pandas [73] datasets. For nuScenes, we select eight scenes rich in dynamic objects and complex lane structures to test large maneuver capabilities. For PandaSet, we adopt the ten scenes used in Unisim [79] and NeuRAD [64] for fair comparison." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.417, + 0.909, + 0.794 + ], + "angle": 0, + "content": "Implementation Details. To demonstrate the versatility and robustness of DriveDreamer4D, we incorporate various 4DGS baselines into our pipeline, including Deformable-GS [80], S\\(^3\\)Gaussian [31], and PVG [9]. For a fair comparison, LiDAR supervision is introduced to Deformable-GS. Scenes are segmented into 40-frame clips, matching the generative model's output length, using only forward-facing camera data at a resolution of \\(640 \\times 960\\). Models are trained with the Adam optimizer [37], following the learning rate schedule from 3D Gaussian Splatting [36]. Hyperparameters align with each baseline [9, 31, 80], with training strategies unchanged except for incorporating CDTS. Moreover, to demonstrate the robustness of our method across different datasets, we employ PVG and DriveDreamer4D with PVG on nuScenes [7], and PandaSet [73]. Since existing methods typically conduct experiments on one or two datasets with varying scene selections, we have chosen different state-of-the-art (SOTA) methods for comparison on each dataset. This approach ensures a comprehensive evaluation of our method's performance under diverse conditions and highlights its adaptability and effectiveness across a wide range of scenarios. By benchmarking against various SOTA methods tailored to specific datasets, we provide strong evidence of the generalizability and robustness of DriveDreamer4D in handling complex and varied driving environments." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.796, + 0.909, + 0.903 + ], + "angle": 0, + "content": "Metrics. Traditional 3D reconstruction tasks typically employ PSNR and SSIM metrics for evaluation, with validation sets that closely match the training data distribution (i.e., uniformly sampling frames from video sequences for validation, with the remainder used for training). However, in closed-loop driving simulation, the focus shifts to evaluating model rendering performance under novel trajectories," + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.52, + 0.958 + ], + "angle": 0, + "content": "12020" + } + ], + [ + { + "type": "table", + "bbox": [ + 0.095, + 0.089, + 0.483, + 0.157 + ], + "angle": 0, + "content": "
Counterpart MethodDriveDreamer4D Win Rate
Lane ChangeAccelerationDecelerationAverage
PVG [9]100.0%90.5%89.1%93.2%
S3Gaussian [31]100.0%97.9%92.2%96.7%
Deformable-GS [80]95.8%83.5%72.9%84.1%
" + }, + { + "type": "table_caption", + "bbox": [ + 0.09, + 0.166, + 0.483, + 0.194 + ], + "angle": 0, + "content": "Table 3. User study comparison of DriveDreamer4D win rates across various novel trajectory view synthesis." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.222, + 0.483, + 0.312 + ], + "angle": 0, + "content": "where corresponding sensor data are unavailable, making metrics like PSNR and SSIM inapplicable for evaluation. Therefore, we propose Novel Trajectory Agent IoU (NTA-IoU) and Novel Trajectory Lane IoU (NTL-IoU), which assess the spatiotemporal coherence of foreground and background traffic components in novel trajectory viewpoints." + }, + { + "type": "text", + "bbox": [ + 0.089, + 0.313, + 0.483, + 0.509 + ], + "angle": 0, + "content": "For NTA-IoU, we use YOLO11 [35] to identify vehicles in images rendered from novel trajectory views, yielding 2D bounding boxes. Simultaneously, geometric transformations are applied to the original 3D bounding boxes, projecting them onto the new viewpoints to generate corresponding 2D bounding boxes. For each projected 2D box, we then identify the closest detector-generated 2D box and compute their Intersection over Union (IoU). To ensure accurate matching, a distance threshold \\( d_{\\mathrm{thresh}} \\) is introduced: when the center-to-center distance \\( \\| c(B^{\\mathrm{proj}}) - c(B^{\\mathrm{det}})\\| \\) between the nearest detected box \\( B^{\\mathrm{det}} \\) and the correctly projected box \\( B^{\\mathrm{proj}} \\) surpasses this threshold, their NTA-IoU is assigned a value of zero:" + }, + { + "type": "equation", + "bbox": [ + 0.096, + 0.521, + 0.483, + 0.562 + ], + "angle": 0, + "content": "\\[\n\\mathrm {N T A - I o U} = \\left\\{ \\begin{array}{l l} 0 & \\text {i f} \\| c (B ^ {\\operatorname {p r o j}}) - c (B ^ {\\det }) \\| \\geq d _ {\\text {t h r e s h}} \\\\ \\operatorname {I o U} \\left(B ^ {\\operatorname {p r o j}}, B ^ {\\det }\\right) & \\text {o t h e r w i s e .} \\end{array} \\right. \\tag {12}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.572, + 0.484, + 0.648 + ], + "angle": 0, + "content": "For NTL-IoU, we employ TwinLiteNet [8] to extract 2D lanes from rendered images. Ground truth lanes are also projected onto the 2D image plane. We then compute the mean Intersection-over-Union (mIoU) between the rendered and ground truth lanes \\( L^{\\mathrm{det}} \\) and \\( L^{\\mathrm{proj}} \\):" + }, + { + "type": "equation", + "bbox": [ + 0.182, + 0.659, + 0.483, + 0.677 + ], + "angle": 0, + "content": "\\[\n\\mathrm {N T L - I o U} = \\mathrm {m I o U} \\left(L ^ {\\text {p r o j}}, L ^ {\\text {d e t}}\\right). \\tag {13}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.089, + 0.689, + 0.483, + 0.901 + ], + "angle": 0, + "content": "Additionally, in lane change scenarios, we observe inaccuracies in relative positioning, as well as frequent occurrences of artifacts such as flying points and ghosting, which notably degrade image quality. To assess this, we employ the FID metric [24], which quantifies differences in feature distribution between rendered novel trajectory images and original trajectory images. This metric effectively reflects visual quality and is particularly sensitive to artifacts like flying points and ghosting, providing a robust measure of image fidelity in these complex scenes. Finally, a user study is conducted to evaluate the quality of the renderings, where participants compare the rendering results of each baseline with its DriveDreamer4D enhanced version across three novel trajectories. The evaluation criteria" + }, + { + "type": "table", + "bbox": [ + 0.517, + 0.089, + 0.906, + 0.145 + ], + "angle": 0, + "content": "
MethodNTA-IoU ↑NTL-IoU ↑FID ↓
PVG[9]0.20448.75103.92
DriveDreamer4D with PVG0.33850.3688.60
" + }, + { + "type": "table_caption", + "bbox": [ + 0.513, + 0.154, + 0.907, + 0.183 + ], + "angle": 0, + "content": "Table 4. Comparison of NTA-IoU, NTL-IoU and FID scores across lane change on nuScenes." + }, + { + "type": "table", + "bbox": [ + 0.517, + 0.198, + 0.906, + 0.267 + ], + "angle": 0, + "content": "
MethodNo shift ↓Lane 2m ↓Lane 3m ↓Vert. 1m ↓
UniSim [79]-74.797.5-
NeuRAD [64]25.072.393.976.3
PVG [9]24.767.376.882.9
DriveDreamer4D with PVG23.562.474.669.8
" + }, + { + "type": "table_caption", + "bbox": [ + 0.529, + 0.276, + 0.89, + 0.29 + ], + "angle": 0, + "content": "Table 5. FID when shifting pose of ego vehicle on PandaSet." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.319, + 0.907, + 0.381 + ], + "angle": 0, + "content": "focus on overall video quality, with particular attention to foreground objects like vehicles. For each comparison, participants were asked to select the option they found most favorable. Further details are provided in the supplement." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.393, + 0.653, + 0.407 + ], + "angle": 0, + "content": "4.2. Main Results" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.415, + 0.906, + 0.521 + ], + "angle": 0, + "content": "In this subsection, we first demonstrate the effectiveness of DriveDreamer4D with various baseline methods. Then, through experimental results obtained from three different datasets, combined with detailed comparisons against a range of baselines, we highlight the robustness and adaptability of DriveDreamer4D in addressing a broad spectrum of scenarios and challenges." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.522, + 0.906, + 0.704 + ], + "angle": 0, + "content": "Waymo. As demonstrated in Tab. 1, integrating DriveDreamer4D with various 4DGS algorithms consistently yields superior NTA-IoU and NTL-IoU scores across complex maneuvers (e.g., lane changes, acceleration, and deceleration), surpassing baseline methods. Specifically, DriveDreamer4D enhances average NTA-IoU by \\(22.6\\%\\), \\(43.5\\%\\), and \\(15.6\\%\\) across three baselines (PVG [9], \\(\\mathrm{S}^3\\)Gaussian [31], Deformable-GS [80]), underscoring its capability to improve the spatiotemporal coherence of foreground agents. It also boosts average NTL-IoU by \\(1.8\\%\\), \\(3.7\\%\\), and \\(1.6\\%\\), enhancing background lane consistency in 4D driving scene rendering." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.705, + 0.909, + 0.903 + ], + "angle": 0, + "content": "In addition to verifying the spatiotemporal consistency of rendered novel trajectory views, we leverage the FID metric to assess rendering quality under novel trajectories. Given that acceleration and deceleration scenarios yield rendered views with distributional similarities to ground truth, limiting FID's discriminative capability across algorithms, our FID comparisons focus specifically on lane change scenarios. Experiment results, as presented in Tab. 2, indicate that our method substantially outperforms the baseline methods (PVG [9], \\(\\mathrm{S}^3\\) Gaussian [31], Deformable-GS [80]), with FID relative improvements of \\(32.1\\%\\), \\(46.4\\%\\), and \\(16.3\\%\\). These results highlight DriveDreamer4D's capability to enhance generation quality for" + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.518, + 0.957 + ], + "angle": 0, + "content": "12021" + } + ], + [ + { + "type": "table", + "bbox": [ + 0.098, + 0.089, + 0.287, + 0.196 + ], + "angle": 0, + "content": "
λnovelNTA-IoU ↑FID ↓
00.349105.29
0.50.40582.84
10.42079.54
1.50.41782.10
" + }, + { + "type": "table_caption", + "bbox": [ + 0.094, + 0.207, + 0.286, + 0.249 + ], + "angle": 0, + "content": "Table 6. Ablation study on the training loss weight \\( \\lambda_{\\mathrm{novel}} \\) for novel trajectory data." + }, + { + "type": "table", + "bbox": [ + 0.299, + 0.089, + 0.48, + 0.197 + ], + "angle": 0, + "content": "
λregNTA-IoU ↑FID ↓
00.42079.54
1e-20.411119.39
1e-30.42871.52
1e-40.42275.31
" + }, + { + "type": "table_caption", + "bbox": [ + 0.298, + 0.207, + 0.48, + 0.235 + ], + "angle": 0, + "content": "Table 7. Ablation study on the regularization loss weight \\( {\\lambda }_{\\text{reg }} \\) ." + }, + { + "type": "text", + "bbox": [ + 0.091, + 0.279, + 0.277, + 0.293 + ], + "angle": 0, + "content": "novel trajectory viewpoints." + }, + { + "type": "text", + "bbox": [ + 0.089, + 0.295, + 0.483, + 0.446 + ], + "angle": 0, + "content": "Finally, we conduct a user study to evaluate the rendering quality of different methods on novel trajectories, with a specific focus on foreground agents. For each method, we generate three novel trajectory views—lane change, acceleration, and deceleration—across eight scenes from the Waymo dataset [61]. Participants are then asked to select the renderings they found most visually favorable in each comparison. The DriveDreamer4D win rates from this study, shown in Tab. 3, reveal a significant user preference for our method's renderings." + }, + { + "type": "text", + "bbox": [ + 0.093, + 0.447, + 0.483, + 0.718 + ], + "angle": 0, + "content": "nuScenes. Tab. 4 provides a comprehensive performance comparison between DriveDreamer4D and PVG [9] in lane change scenarios on the nuScenes dataset [7]. The experimental results offer compelling evidence that our method achieves a substantial improvement over the baseline, showcasing its superior capability in handling complex driving situations. Specifically, as detailed in the table, DriveDreamer4D consistently outperforms PVG across multiple evaluation metrics, including NTA-IoU, NTL-IoU, and FID. Notably, our method achieves a \\(65.7\\%\\) increase in NTA-IoU, a \\(3.3\\%\\) increase in NTL-IoU, and a \\(14.7\\%\\) reduction in FID. These results underscore the robustness of DriveDreamer4D in addressing challenging driving scenarios, highlighting its exceptional performance and adaptability across diverse evaluation criteria. Such advancements reaffirm the effectiveness of our approach in delivering more accurate and reliable trajectory predictions for dynamic environments." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.72, + 0.483, + 0.902 + ], + "angle": 0, + "content": "PandaSet. Following the experimental setup of NeuRAD [64] and UniSim [79], we conduct experiments on the PandaSet dataset [73]. As shown in Tab. 5, we report the FID comparison under the original trajectory, a lane shift of 2 meters, a lane shift of 3 meters, and a perpendicular shift of 1 meter. These results demonstrate that our DriveDreamer4D achieves the best performance across all settings. Notably, DriveDreamer4D achieves an FID of 74.6 under the challenging scenario of a 3-meter lane shift, representing a significant improvement of \\(20.6\\%\\) over the previous SOTA method, NeuRAD. This highlights the strong performance and competitiveness of our approach com" + }, + { + "type": "table", + "bbox": [ + 0.518, + 0.089, + 0.906, + 0.179 + ], + "angle": 0, + "content": "
Depth LossTemporal-aligned Cousin PairRegularization LossNTA-IoU ↑FID ↓
✓××0.40182.63
×××0.42079.54
×✓×0.42376.20
×✓✓0.42871.52
" + }, + { + "type": "table_caption", + "bbox": [ + 0.551, + 0.189, + 0.866, + 0.203 + ], + "angle": 0, + "content": "Table 8. Ablation study on the depth loss and CDTS." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.235, + 0.906, + 0.28 + ], + "angle": 0, + "content": "pared to existing techniques, reaffirming its effectiveness in handling complex driving scenarios with high precision and robustness." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.298, + 0.678, + 0.312 + ], + "angle": 0, + "content": "4.3. Ablation Studies" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.323, + 0.907, + 0.549 + ], + "angle": 0, + "content": "We conduct an ablation study based on PVG [9], analyzing the effects of hyperparameters settings for \\(\\lambda_{\\mathrm{novel}}\\) and \\(\\lambda_{\\mathrm{reg}}\\), as well as the impact of depth loss, temporal-aligned cousin pairs, and regularization loss. Following the experiments shown in Tab. 6 and 7, we set \\(\\lambda_{\\mathrm{novel}}\\) to 1 and \\(\\lambda_{\\mathrm{reg}}\\) to \\(1\\times 10^{-3}\\) which yield the best performance. As shown in Tab. 8, the experiment confirms that the depth loss should be excluded when optimizing novel trajectory views, since LiDAR depth maps are incomplete due to occlusions. Furthermore, employing temporal-aligned cousin pairs achieves an FID of 76.20 and an NTA-IoU of 0.423, with the perception regularization loss further improving the FID to 71.52 and NTA-IoU to 0.428. These results highlight the effectiveness of CDTS, in achieving \\(\\sim 10\\%\\) improvement in FID, along with a \\(\\sim 2\\%\\) increase in NTA-IoU." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.571, + 0.762, + 0.587 + ], + "angle": 0, + "content": "5. Discussion and Conclusion" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.599, + 0.907, + 0.901 + ], + "angle": 0, + "content": "In this paper, we present DriveDreamer4D, a novel framework designed to advance 4D driving scene representations by harnessing priors from world models. Addressing key limitations of current sensor simulation methods—namely, their dependence on forward-driving training data distributions and inability to model complex maneuvers—DriveDreamer4D leverages a world model to generate novel trajectory videos that complement real-world driving data. By explicitly employing structured conditions, our framework maintains spatial-temporal consistency across traffic elements, ensuring that generated data adheres closely to the dynamics of real-world traffic scenarios. Our experiments demonstrate that DriveDreamer4D achieves superior quality in generating diverse simulation viewpoints, with significant improvements in both the rendering fidelity and spatiotemporal coherence of scene components. Notably, these results highlight DriveDreamer4D's potential as a foundation for closed-loop simulations that require high-fidelity reconstructions of dynamic driving scenes." + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "12022" + } + ], + [ + { + "type": "title", + "bbox": [ + 0.093, + 0.09, + 0.188, + 0.106 + ], + "angle": 0, + "content": "References" + }, + { + "type": "ref_text", + "bbox": [ + 0.101, + 0.115, + 0.484, + 0.171 + ], + "angle": 0, + "content": "[1] Benjamin Attal, Jia-Bin Huang, Christian Richardt, Michael Zollhoefer, Johannes Kopf, Matthew O'Toole, and Changil Kim. Hyperreel: High-fidelity 6-dof video with ray-conditioned sampling. In CVPR, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.101, + 0.172, + 0.484, + 0.212 + ], + "angle": 0, + "content": "[2] Jonathan T Barron, Ben Mildenhall, Dor Verbin, Pratul P Srinivasan, and Peter Hedman. Mip-nerf 360: Unbounded anti-aliased neural radiance fields. In CVPR, 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.213, + 0.483, + 0.255 + ], + "angle": 0, + "content": "[3] Jonathan T Barron, Ben Mildenhall, Dor Verbin, Pratul P Srinivasan, and Peter Hedman. Zip-nerf: Anti-aliased grid-based neural radiance fields. In ICCV, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.256, + 0.483, + 0.325 + ], + "angle": 0, + "content": "[4] Andreas Blattmann, Tim Dockhorn, Sumith Kulal, Daniel Mendelevitch, Maciej Kilian, Dominik Lorenz, Yam Levi, Zion English, Vikram Voleti, Adam Letts, et al. Stable video diffusion: Scaling latent video diffusion models to large datasets. arXiv preprint arXiv:2311.15127, 2023. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.327, + 0.483, + 0.381 + ], + "angle": 0, + "content": "[5] Andreas Blattmann, Robin Rombach, Huan Ling, Tim Dockhorn, Seung Wook Kim, Sanja Fidler, and Karsten Kreis. Align your latents: High-resolution video synthesis with latent diffusion models. In CVPR, 2023. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.382, + 0.483, + 0.45 + ], + "angle": 0, + "content": "[6] Tim Brooks, Bill Peebles, Connor Holmes, Will DePue, Yufei Guo, Li Jing, David Schnurr, Joe Taylor, Troy Luhman, Eric Luhman, Clarence Ng, Ricky Wang, and Aditya Ramesh. Video generation models as world simulators. 2024. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.452, + 0.483, + 0.52 + ], + "angle": 0, + "content": "[7] Holger Caesar, Varun Bankiti, Alex H Lang, Sourabh Vora, Venice Erin Liong, Qiang Xu, Anush Krishnan, Yu Pan, Giancarlo Baldan, and Oscar Beijbom. nuscenes: A multimodal dataset for autonomous driving. In CVPR, 2020. 6, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.522, + 0.483, + 0.577 + ], + "angle": 0, + "content": "[8] Quang-Huy Che, Dinh-Phuc Nguyen, Minh-Quan Pham, and Duc-Khai Lam. Twinlitenet: An efficient and lightweight model for driveable area and lane segmentation in self-driving cars. In MAPR, 2023. 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.579, + 0.483, + 0.633 + ], + "angle": 0, + "content": "[9] Yurui Chen, Chun Gu, Junzhe Jiang, Xiatian Zhu, and Li Zhang. Periodic vibration gaussian: Dynamic urban scene reconstruction and real-time rendering. arXiv preprint arXiv:2311.18561, 2023. 1, 2, 5, 6, 7, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.635, + 0.483, + 0.676 + ], + "angle": 0, + "content": "[10] Zilong Chen, Yikai Wang, Feng Wang, Zhengyi Wang, and Huaping Liu. V3d: Video diffusion models are effective 3d generators. arXiv preprint arXiv:2403.06738, 2024. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.677, + 0.483, + 0.746 + ], + "angle": 0, + "content": "[11] Ziyu Chen, Jiawei Yang, Jiahui Huang, Riccardo de Lutio, Janick Martinez Esturo, Boris Ivanovic, Or Litany, Zan Gojcic, Sanja Fidler, Marco Pavone, Li Song, and Yue Wang. Omnire: Omni urban scene reconstruction. arXiv preprint arXiv:2408.16760, 2024. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.747, + 0.483, + 0.815 + ], + "angle": 0, + "content": "[12] Kai Cheng, Xiaoxiao Long, Wei Yin, Jin Wang, Zhiqiang Wu, Yuexin Ma, Kaixuan Wang, Xiaozhi Chen, and Xuejin Chen. Uc-nerf: Neural radiance field for under-calibrated multi-view cameras in autonomous driving. arXiv preprint arXiv:2311.16945, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.817, + 0.483, + 0.859 + ], + "angle": 0, + "content": "[13] Xiaobiao Du, Yida Wang, and Xin Yu. Mvgs: Multi-view-regulated gaussian splattingg for novel view synthesis. arXiv preprint arXiv:2410.02103, 2024. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.86, + 0.483, + 0.901 + ], + "angle": 0, + "content": "[14] Tobias Fischer, Jonas Kulhanek, Samuel Rota Bulo, Lorenzo Porzi, Marc Pollefeys, and Peter Kontschieder. Dynamic 3d gaussian fields for urban areas. NeurIPS, 2024. 2" + }, + { + "type": "list", + "bbox": [ + 0.094, + 0.115, + 0.484, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.093, + 0.906, + 0.147 + ], + "angle": 0, + "content": "[15] Sara Fridovich-Keil, Giacomo Meanti, Frederik Rahbæk Warburg, Benjamin Recht, and Angjoo Kanazawa. K-planes: Explicit radiance fields in space, time, and appearance. In CVPR, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.15, + 0.906, + 0.204 + ], + "angle": 0, + "content": "[16] Ruiyuan Gao, Kai Chen, Zhihao Li, Lanqing Hong, Zhenguo Li, and Qiang Xu. Magicdrive3d: Controllable 3d generation for any-view rendering in street scenes. arXiv preprint arXiv:2405.14475, 2024. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.207, + 0.906, + 0.275 + ], + "angle": 0, + "content": "[17] Ruiqi Gao, Aleksander Holynski, Philipp Henzler, Arthur Brussee, Ricardo Martin-Brualla, Pratul Srinivasan, Jonathan T Barron, and Ben Poole. Cat3d: Create anything in 3d with multi-view diffusion models. arXiv preprint arXiv:2405.10314, 2024. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.277, + 0.906, + 0.346 + ], + "angle": 0, + "content": "[18] Shenyuan Gao, Jiazhi Yang, Li Chen, Kashyap Chitta, Yihang Qiu, Andreas Geiger, Jun Zhang, and Hongyang Li. Vista: A generalizable driving world model with high fidelity and versatile controllability. arXiv preprint arXiv:2405.17398, 2024. 2, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.348, + 0.906, + 0.416 + ], + "angle": 0, + "content": "[19] Rohit Girdhar, Mannat Singh, Andrew Brown, Quentin Duval, Samaneh Azadi, Sai Saketh Rambhatla, Akbar Shah, Xi Yin, Devi Parikh, and Ishan Misra. Emu video: Factorizing text-to-video generation by explicit image conditioning. arXiv preprint arXiv:2311.10709, 2023. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.418, + 0.906, + 0.487 + ], + "angle": 0, + "content": "[20] Jianfei Guo, Nianchen Deng, Xinyang Li, Yeqi Bai, Botian Shi, Chiyu Wang, Chenjing Ding, Dongliang Wang, and Yikang Li. Streetsurf: Extending multi-view implicit surface reconstruction to street views. arXiv preprint arXiv:2306.04988, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.489, + 0.906, + 0.544 + ], + "angle": 0, + "content": "[21] Agrim Gupta, Lijun Yu, Kihyuk Sohn, Xiuye Gu, Meera Hahn, Li Fei-Fei, Irfan Essa, Lu Jiang, and José Lezama. Photorealistic video generation with diffusion models. arXiv preprint arXiv:2312.06662, 2023. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.546, + 0.906, + 0.6 + ], + "angle": 0, + "content": "[22] Huasong Han, Kaixuan Zhou, Xiaoxiao Long, Yusen Wang, and Chunxia Xiao. Ggs: Generalizable gaussian splatting for lane switching in autonomous driving. arXiv preprint arXiv:2409.02382, 2024. 2, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.602, + 0.906, + 0.644 + ], + "angle": 0, + "content": "[23] Ayaan Haque, Matthew Tancik, Alexei A Efros, Aleksander Holynski, and Angjoo Kanazawa. Instruct-nerf2nerf: Editing 3d scenes with instructions. In CVPR, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.646, + 0.906, + 0.7 + ], + "angle": 0, + "content": "[24] Martin Heusel, Hubert Ramsauer, Thomas Unterthiner, Bernhard Nessler, and Sepp Hochreiter. Gans trained by a two time-scale update rule converge to a local nash equilibrium. NeurIPS, 2017. 6, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.702, + 0.906, + 0.771 + ], + "angle": 0, + "content": "[25] Jonathan Ho, William Chan, Chitwan Sahara, Jay Whang, Ruiqi Gao, Alexey Gritsenko, Diederik P Kingma, Ben Poole, Mohammad Norouzi, David J Fleet, et al. Imagen video: High definition video generation with diffusion models. arXiv preprint arXiv:2210.02303, 2022. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.773, + 0.906, + 0.814 + ], + "angle": 0, + "content": "[26] Jonathan Ho, Tim Salimans, Alexey Gritsenko, William Chan, Mohammad Norouzi, and David J Fleet. Video diffusion models. NeurIPS, 2022." + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.817, + 0.906, + 0.87 + ], + "angle": 0, + "content": "[27] Wenyi Hong, Ming Ding, Wendi Zheng, Xinghan Liu, and Jie Tang. Cogvideo: Large-scale pretraining for text-to-video generation via transformers. arXiv preprint arXiv:2205.15868, 2022.3" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.873, + 0.906, + 0.901 + ], + "angle": 0, + "content": "[28] Anthony Hu, Lloyd Russell, Hudson Yeo, Zak Murez, George Fedoseev, Alex Kendall, Jamie Shotton, and Gian" + }, + { + "type": "list", + "bbox": [ + 0.517, + 0.093, + 0.906, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "12023" + } + ], + [ + { + "type": "ref_text", + "bbox": [ + 0.125, + 0.092, + 0.482, + 0.133 + ], + "angle": 0, + "content": "luca Corrado. Gaia-1: A generative world model for autonomous driving. arXiv preprint arXiv:2309.17080, 2023. 2, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.134, + 0.483, + 0.189 + ], + "angle": 0, + "content": "[29] Shengchao Hu, Li Chen, Penghao Wu, Hongyang Li, Junchi Yan, and Dacheng Tao. St-p3: End-to-end vision-based autonomous driving via spatial-temporal feature learning. In ECCV, 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.19, + 0.482, + 0.245 + ], + "angle": 0, + "content": "[30] Yihan Hu, Jiazhi Yang, Li Chen, Keyu Li, Chonghao Sima, Xizhou Zhu, Siqi Chai, Senyao Du, Tianwei Lin, Wenhai Wang, et al. Planning-oriented autonomous driving. In CVPR, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.246, + 0.482, + 0.315 + ], + "angle": 0, + "content": "[31] Nan Huang, Xiaobao Wei, Wenzhao Zheng, Pengju An, Ming Lu, Wei Zhan, Masayoshi Tomizuka, Kurt Keutzer, and Shanghang Zhang. \\(s^3\\) gaussian: Self-supervised street gaussians for autonomous driving. arXiv preprint arXiv:2405.20323, 2024. 1, 2, 5, 6, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.316, + 0.482, + 0.37 + ], + "angle": 0, + "content": "[32] Sungwon Hwang, Min-Jung Kim, Taewoong Kang, Jayeon Kang, and Jaegul Choo. Vegs: View extrapolation of urban scenes in 3d gaussian splatting using learned priors. In ECCV, 2024. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.371, + 0.482, + 0.426 + ], + "angle": 0, + "content": "[33] Muhammad Zubair Irshad, Sergey Zakharov, Katherine Liu, Vitor Guizilini, Thomas Kollar, Adrien Gaidon, Zsolt Kira, and Rares Ambrus. Neo 360: Neural fields for sparse view synthesis of outdoor scenes. In ICCV, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.427, + 0.482, + 0.482 + ], + "angle": 0, + "content": "[34] Bo Jiang, Shaoyu Chen, Qing Xu, Bencheng Liao, Jiajie Chen, Helong Zhou, Qian Zhang, Wenyu Liu, Chang Huang, and Xinggang Wang. Vad: Vectorized scene representation for efficient autonomous driving. In ICCV, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.483, + 0.454, + 0.497 + ], + "angle": 0, + "content": "[35] Glenn Jocher and Jing Qiu. Ultralytics yolo11, 2024. 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.497, + 0.482, + 0.538 + ], + "angle": 0, + "content": "[36] Bernhard Kerbl, Georgios Kopanas, Thomas Leimkuhler, and George Drettakis. 3d gaussian splatting for real-time radiance field rendering. ACM ToG, 2023. 2, 3, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.539, + 0.482, + 0.579 + ], + "angle": 0, + "content": "[37] Diederik P Kingma and Jimmy Ba. Adam: A method for stochastic optimization. arXiv preprint arXiv:1412.6980, 2014. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.58, + 0.482, + 0.65 + ], + "angle": 0, + "content": "[38] Dan Kondratyuk, Lijun Yu, Xiuye Gu, José Lezama, Jonathan Huang, Rachel Hornung, Hartwig Adam, Hassan Akbari, Yair Alon, Vighnesh Birodkar, et al. Videopoet: A large language model for zero-shot video generation. arXiv preprint arXiv:2312.14125, 2023. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.65, + 0.482, + 0.718 + ], + "angle": 0, + "content": "[39] Abhijit Kundu, Kyle Genova, Xiaoqi Yin, Alireza Fathi, Caroline Pantofaru, Leonidas J Guibas, Andrea Tagliasacchi, Frank Dellaert, and Thomas Funkhouser. Panoptic neural fields: A semantic object-aware neural scene representation. In CVPR, 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.72, + 0.482, + 0.747 + ], + "angle": 0, + "content": "[40] Yann LeCun and Courant. A path towards autonomous machine intelligence version 0.9.2, 2022-06-27. 2022. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.748, + 0.482, + 0.816 + ], + "angle": 0, + "content": "[41] Hao Li, Ming Yuan, Yan Zhang, Chenming Wu, Chen Zhao, Chunyu Song, Haocheng Feng, Errui Ding, Dingwen Zhang, and Jingdong Wang. Xld: A cross-lane dataset for benchmarking novel driving view synthesis. arXiv preprint arXiv:2406.18360, 2024. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.817, + 0.482, + 0.858 + ], + "angle": 0, + "content": "[42] Zhengqi Li, Simon Niklaus, Noah Snavely, and Oliver Wang. Neural scene flow fields for space-time view synthesis of dynamic scenes. In CVPR, 2021. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.859, + 0.482, + 0.901 + ], + "angle": 0, + "content": "[43] Zhiqi Li, Zhiding Yu, Shiyi Lan, Jiahan Li, Jan Kautz, Tong Lu, and Jose M Alvarez. Is ego status all you need for open-loop end-to-end autonomous driving? In CVPR, 2024. 2" + }, + { + "type": "list", + "bbox": [ + 0.093, + 0.092, + 0.483, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.092, + 0.906, + 0.147 + ], + "angle": 0, + "content": "[44] Zhuopeng Li, Yilin Zhang, Chenming Wu, Jianke Zhu, and Liangjun Zhang. Ho-gaussian: Hybrid optimization of 3d gaussian splatting for urban scenes. arXiv preprint arXiv:2403.20032, 2024. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.149, + 0.905, + 0.203 + ], + "angle": 0, + "content": "[45] Chen-Hsuan Lin, Jun Gao, Luming Tang, Towaki Takikawa, Xiaohui Zeng, Xun Huang, Karsten Kreis, Sanja Fidler, Ming-Yu Liu, and Tsung-Yi Lin. Magic3d: High-resolution text-to-3d content creation. In CVPR, 2023. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.204, + 0.905, + 0.258 + ], + "angle": 0, + "content": "[46] Haotong Lin, Sida Peng, Zhen Xu, Yunzhi Yan, Qing Shuai, Hujun Bao, and Xiaowei Zhou. Efficient neural radiance fields for interactive free-viewpoint video. In SIGGRAPH Asia, 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.26, + 0.905, + 0.301 + ], + "angle": 0, + "content": "[47] Fan Lu, Yan Xu, Guang Chen, Hongsheng Li, Kwan-Yee Lin, and Changjun Jiang. Urban radiance field representation with deformable neural mesh primitives. In ICCV, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.302, + 0.905, + 0.357 + ], + "angle": 0, + "content": "[48] Xin Ma, Yaohui Wang, Gengyun Jia, Xinyuan Chen, Zwei Liu, Yuan-Fang Li, Cunjian Chen, and Yu Qiao. Latte: Latent diffusion transformer for video generation. arXiv preprint arXiv:2401.03048, 2024. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.358, + 0.905, + 0.412 + ], + "angle": 0, + "content": "[49] Ben Mildenhall, Pratul P Srinivasan, Matthew Tancik, Jonathan T Barron, Ravi Ramamoorthi, and Ren Ng. Nerf: Representing scenes as neural radiance fields for view synthesis. CACM, 2021. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.413, + 0.905, + 0.455 + ], + "angle": 0, + "content": "[50] Thomas Müller, Alex Evans, Christoph Schied, and Alexander Keller. Instant neural graphics primitives with a multiresolution hash encoding. ACM ToG, 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.455, + 0.905, + 0.496 + ], + "angle": 0, + "content": "[51] Julian Ost, Fahim Mannan, Nils Thuerey, Julian Knodt, and Felix Heide. Neural scene graphs for dynamic scenes. In CVPR, 2021. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.497, + 0.905, + 0.566 + ], + "angle": 0, + "content": "[52] Keunhong Park, Utkarsh Sinha, Peter Hedman, Jonathan T Barron, Sofien Bouaziz, Dan B Goldman, Ricardo MartinBrualla, and Steven M Seitz. Hypernerf: A higher-dimensional representation for topologically varying neural radiance fields. arXiv preprint arXiv:2106.13228, 2021. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.567, + 0.905, + 0.635 + ], + "angle": 0, + "content": "[53] Dustin Podell, Zion English, Kyle Lacey, Andreas Blattmann, Tim Dockhorn, Jonas Müller, Joe Penna, and Robin Rombach. Sdxl: Improving latent diffusion models for high-resolution image synthesis. arXiv preprint arXiv:2307.01952, 2023. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.636, + 0.905, + 0.678 + ], + "angle": 0, + "content": "[54] Ben Poole, Ajay Jain, Jonathan T Barron, and Ben Mildenhall. Dreamfusion: Text-to-3d using 2d diffusion. arXiv preprint arXiv:2209.14988, 2022. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.679, + 0.905, + 0.733 + ], + "angle": 0, + "content": "[55] Aditya Ramesh, Prafulla Dhariwal, Alex Nichol, Casey Chu, and Mark Chen. Hierarchical text-conditional image generation with clip latents. arXiv preprint arXiv:2204.06125, 2022.3" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.734, + 0.905, + 0.788 + ], + "angle": 0, + "content": "[56] Konstantinos Rematas, Andrew Liu, Pratul P Srinivasan, Jonathan T Barron, Andrea Tagliasacchi, Thomas Funkhouser, and Vittorio Ferrari. Urban radiance fields. In CVPR, 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.789, + 0.905, + 0.83 + ], + "angle": 0, + "content": "[57] Robin Rombach, Andreas Blattmann, Dominik Lorenz, Patrick Esser, and Björn Ommher. High-resolution image synthesis with latent diffusion models. In CVPR, 2022. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.831, + 0.905, + 0.901 + ], + "angle": 0, + "content": "[58] Chitwan Sahara, William Chan, Saurabh Saxena, Lala Li, Jay Whang, Emily L Denton, Kamyar Ghasemipour, Raphael Gontijo Lopes, Burcu Karagol Ayan, Tim Salimans, et al. Photorealistic text-to-image diffusion models with deep language understanding. NeurIPS, 2022. 3" + }, + { + "type": "list", + "bbox": [ + 0.517, + 0.092, + 0.906, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.52, + 0.957 + ], + "angle": 0, + "content": "12024" + } + ], + [ + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.092, + 0.482, + 0.161 + ], + "angle": 0, + "content": "[59] Kyle Sargent, Zizhang Li, Tanmay Shah, Charles Herrmann, Hong-Xing Yu, Yunzhi Zhang, Eric Ryan Chan, Dmitry Lagun, Li Fei-Fei, Deqing Sun, et al. Zeronvs: Zero-shot 360-degree view synthesis from a single real image. arXiv preprint arXiv:2310.17994, 2023. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.164, + 0.482, + 0.219 + ], + "angle": 0, + "content": "[60] Liangchen Song, Anpei Chen, Zhong Li, Zhang Chen, Lele Chen, Junsong Yuan, Yi Xu, and Andreas Geiger. Nerfplayer: A streamable dynamic scene representation with decomposed neural radiance fields. IEEE TVCG, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.221, + 0.482, + 0.331 + ], + "angle": 0, + "content": "[61] Pei Sun, Henrik Kretzschmar, Xerxes Dotiwalla, Aurelien Chouard, Vijaysai Patnaik, Paul Tsui, James Guo, Yin Zhou, Yuning Chai, Benjamin Caine, Vijay Vasudevan, Wei Han, Jiquan Ngiam, Hang Zhao, Aleksei Timofeev, Scott Ettinger, Maxim Krivokon, Amy Gao, Aditya Joshi, Yu Zhang, Jonathon Shlens, Zhifeng Chen, and Dragomir Anguelov. Scalability in perception for autonomous driving: Waymo open dataset. In CVPR, 2020. 6, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.333, + 0.482, + 0.387 + ], + "angle": 0, + "content": "[62] Shanlin Sun, Bingbing Zhuang, Ziyu Jiang, Buyu Liu, Xiaohui Xie, and Manmohan Chandraker. Lidarf: Delving into lidar for neural radiance field on street scenes. In CVPR, 2024. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.39, + 0.482, + 0.445 + ], + "angle": 0, + "content": "[63] Matthew Tancik, Vincent Casser, Xinchen Yan, Sabeek Pradhan, Ben Mildenhall, Pratul P Srinivasan, Jonathan T Barron, and Henrik Kretzschmar. Block-nerf: Scalable large scene neural view synthesis. In CVPR, 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.447, + 0.482, + 0.502 + ], + "angle": 0, + "content": "[64] Adam Tonderski, Carl Lindström, Georg Hess, William Ljungbergh, Lennart Svensson, and Christoffer Petersson. Neurad: Neural rendering for autonomous driving. In CVPR, 2024. 2, 6, 7, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.504, + 0.482, + 0.573 + ], + "angle": 0, + "content": "[65] Vikram Voleti, Chun-Han Yao, Mark Boss, Adam Letts, David Pankratz, Dmitry Tochilkin, Christian Laforte, Robin Rombach, and Varun Jampani. Sv3d: Novel multi-view synthesis and 3d generation from a single image using latent video diffusion. arXiv preprint arXiv:2403.12008, 2024. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.575, + 0.482, + 0.63 + ], + "angle": 0, + "content": "[66] Xiaofeng Wang, Zheng Zhu, Guan Huang, Xinze Chen, Jiagang Zhu, and Jiwen Lu. Drivedreamer: Towards real-world-driven world models for autonomous driving. arXiv preprint arXiv:2309.09777, 2023. 2, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.632, + 0.482, + 0.7 + ], + "angle": 0, + "content": "[67] Xiaofeng Wang, Kang Zhao, Feng Liu, Jiayu Wang, Guosheng Zhao, Xiaoyi Bao, Zheng Zhu, Yingya Zhang, and Xingang Wang. Egovid-5m: A large-scale video-action dataset for egocentric video generation. arXiv preprint arXiv:2411.08380, 2024. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.702, + 0.482, + 0.758 + ], + "angle": 0, + "content": "[68] Xiaofeng Wang, Zheng Zhu, Guan Huang, Boyuan Wang, Xinze Chen, and Jiwen Lu. Worlddreamer: Towards general world models for video generation via predicting masked tokens. arXiv preprint arXiv:2401.09985, 2024. 2, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.76, + 0.482, + 0.814 + ], + "angle": 0, + "content": "[69] Yuqi Wang, Jiawei He, Lue Fan, Hongxin Li, Yuntao Chen, and Zhaoxiang Zhang. Driving into the future: Multiview visual forecasting and planning with world model for autonomous driving. In CVPR, 2024. 2, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.817, + 0.482, + 0.871 + ], + "angle": 0, + "content": "[70] Rundi Wu, Ben Mildenhall, Philipp Henzler, Keunhong Park, Ruiqi Gao, Daniel Watson, Pratul P Srinivasan, Dor Verbin, Jonathan T Barron, Ben Poole, et al. Reconfusion: 3d reconstruction with diffusion priors. In CVPR, 2024. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.873, + 0.482, + 0.901 + ], + "angle": 0, + "content": "[71] Zirui Wu, Tianyu Liu, Liyi Luo, Zhide Zhong, Jianteng Chen, Hongmin Xiao, Chao Hou, Haozhe Lou, Yuantao" + }, + { + "type": "list", + "bbox": [ + 0.093, + 0.092, + 0.482, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.545, + 0.093, + 0.905, + 0.133 + ], + "angle": 0, + "content": "Chen, Runyi Yang, et al. Mars: An instance-aware, modular and realistic simulator for autonomous driving. In ICAI, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.135, + 0.905, + 0.204 + ], + "angle": 0, + "content": "[72] Jiannan Xiang, Guangyi Liu, Yi Gu, Qiyue Gao, Yuting Ning, Yuheng Zha, Zeyu Feng, Tianhua Tao, Shibo Hao, Yemin Shi, et al. Pandora: Towards general world model with natural language actions and video states. arXiv preprint arXiv:2406.09455, 2024. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.205, + 0.905, + 0.259 + ], + "angle": 0, + "content": "[73] Pengchuan Xiao, Zhenlei Shao, Steven Hao, Zishuo Zhang, Xiaolin Chai, Judy Jiao, Zesong Li, Jian Wu, Kai Sun, Kun Jiang, et al. Pandaset: Advanced sensor suite dataset for autonomous driving. In ITSC, 2021. 6, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.261, + 0.905, + 0.3 + ], + "angle": 0, + "content": "[74] Ziyang Xie, Junge Zhang, Wenye Li, Feihu Zhang, and Li Zhang. S-nerf: Neural radiance fields for street views. arXiv preprint arXiv:2303.00749, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.302, + 0.905, + 0.343 + ], + "angle": 0, + "content": "[75] Wilson Yan, Yunzhi Zhang, Pieter Abbeel, and Aravind Srinivas. Videogpt: Video generation using vq-vae and transformers. arXiv preprint arXiv:2104.10157, 2021. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.345, + 0.905, + 0.399 + ], + "angle": 0, + "content": "[76] Yunzhi Yan, Haotong Lin, Chenxu Zhou, Weijie Wang, Haiyang Sun, Kun Zhan, Xianpeng Lang, Xiaowei Zhou, and Sida Peng. Street gaussians for modeling dynamic urban scenes. arXiv preprint arXiv:2401.01339, 2024. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.401, + 0.905, + 0.467 + ], + "angle": 0, + "content": "[77] Jiawei Yang, Boris Ivanovic, Or Litany, Xinshuo Weng, Seung Wook Kim, Boyi Li, Tong Che, Danfei Xu, Sanja Fidler, Marco Pavone, et al. Emernerf: Emergent spatial-temporal scene decomposition via self-supervision. arXiv preprint arXiv:2311.02077, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.47, + 0.905, + 0.55 + ], + "angle": 0, + "content": "[78] Xuemeng Yang, Licheng Wen, Yukai Ma, Jianbiao Mei, Xin Li, Tiantian Wei, Wenjie Lei, Daocheng Fu, Pinlong Cai, Min Dou, Botian Shi, Liang He, Yong Liu, and Yu Qiao. Drivearena: A closed-loop generative simulation platform for autonomous driving. arXiv preprint arXiv:2408.00415, 2024.3" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.553, + 0.905, + 0.607 + ], + "angle": 0, + "content": "[79] Ze Yang, Yun Chen, Jingkang Wang, Sivabalan Manivasagam, Wei-Chiu Ma, Anqi Joyce Yang, and Raquel Urtasun. Unisim: A neural closed-loop sensor simulator. In CVPR, 2023. 2, 6, 7, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.609, + 0.905, + 0.663 + ], + "angle": 0, + "content": "[80] Ziyi Yang, Xinyu Gao, Wen Zhou, Shaohui Jiao, Yuqing Zhang, and Xiaogang Jin. Deformable 3d gaussians for high-fidelity monocular dynamic scene reconstruction. In CVPR, 2024. 1, 2, 6, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.665, + 0.905, + 0.732 + ], + "angle": 0, + "content": "[81] Zhuoyi Yang, Jiayan Teng, Wendi Zheng, Ming Ding, Shiyu Huang, Jiazheng Xu, Yuanming Yang, Wenyi Hong, Xiaohan Zhang, Guanyu Feng, et al. Cogvideox: Text-to-video diffusion models with an expert transformer. arXiv preprint arXiv:2408.06072, 2024. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.734, + 0.905, + 0.775 + ], + "angle": 0, + "content": "[82] Wang Yifan, Felice Serena, Shihao Wu, Cengiz Öztireli, and Olga Sorkine-Hornung. Differentiable surface splatting for point-based geometry processing. ACM TOG, 2019. 3, 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.777, + 0.905, + 0.816 + ], + "angle": 0, + "content": "[83] Zehao Yu, Anpei Chen, Binbin Huang, Torsten Sattler, and Andreas Geiger. Mip-splatting: Alias-free 3d gaussian splattering. In CVPR, 2024. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.818, + 0.905, + 0.872 + ], + "angle": 0, + "content": "[84] Zhongrui Yu, Haoran Wang, Jinze Yang, Hanzhang Wang, Zeke Xie, Yunfeng Cai, Jiale Cao, Zhong Ji, and Mingming Sun. Sgd: Street view synthesis with gaussian splatting and diffusion prior. arXiv preprint arXiv:2403.20079, 2024. 2, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.874, + 0.905, + 0.901 + ], + "angle": 0, + "content": "[85] Yan Zeng, Guoqiang Wei, Jiani Zheng, Jiaxin Zou, Yang Wei, Yuchen Zhang, and Hang Li. Make pixels" + }, + { + "type": "list", + "bbox": [ + 0.517, + 0.093, + 0.905, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "12025" + } + ], + [ + { + "type": "ref_text", + "bbox": [ + 0.125, + 0.092, + 0.484, + 0.12 + ], + "angle": 0, + "content": "dance: High-dynamic video generation. arXiv preprint arXiv:2311.10982, 2023. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.122, + 0.484, + 0.19 + ], + "angle": 0, + "content": "[86] Jiang-Tian Zhai, Ze Feng, Jinhao Du, Yongqiang Mao, Jiang-Jiang Liu, Zichang Tan, Yifu Zhang, Xiaoqing Ye, and Jingdong Wang. Rethinking the open-loop evaluation of end-to-end autonomous driving in nuscenes. arXiv preprint arXiv:2305.10430, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.192, + 0.483, + 0.261 + ], + "angle": 0, + "content": "[87] Guosheng Zhao, Xiaofeng Wang, Zheng Zhu, Xinze Chen, Guan Huang, Xiaoyi Bao, and Xingang Wang. Drivedreamer-2: Llm-enhanced world models for diverse driving video generation. arXiv preprint arXiv:2403.06845, 2024. 2, 3, 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.263, + 0.483, + 0.319 + ], + "angle": 0, + "content": "[88] Hongyu Zhou, Jiahao Shao, Lu Xu, Dongfeng Bai, Weichao Qiu, Bingbing Liu, Yue Wang, Andreas Geiger, and Yiyi Liao. Hugs: Holistic urban 3d scene understanding via gaussian splatting. In CVPR, 2024. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.32, + 0.483, + 0.387 + ], + "angle": 0, + "content": "[89] Xiaoyu Zhou, Zhiwei Lin, Xiaojun Shan, Yongtao Wang, Deqing Sun, and Ming-Hsuan Yang. Drivinggaussian: Composite gaussian splatting for surrounding dynamic autonomous driving scenes. In CVPR, pages 21634-21643, 2024. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.39, + 0.483, + 0.458 + ], + "angle": 0, + "content": "[90] Zheng Zhu, Xiaofeng Wang, Wangbo Zhao, Chen Min, Ni-anchen Deng, Min Dou, Yuqi Wang, Botian Shi, Kai Wang, Chi Zhang, et al. Is sora a world simulator? a comprehensive survey on general world models and beyond. arXiv preprint arXiv:2405.03520, 2024. 3" + }, + { + "type": "list", + "bbox": [ + 0.093, + 0.092, + 0.484, + 0.458 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.946, + 0.52, + 0.957 + ], + "angle": 0, + "content": "12026" + } + ] +] \ No newline at end of file diff --git a/2025/DriveDreamer4D_ World Models Are Effective Data Machines for 4D Driving Scene Representation/aa9a31fd-776c-4a2a-a4f7-87c23be589e2_origin.pdf b/2025/DriveDreamer4D_ World Models Are Effective Data Machines for 4D Driving Scene Representation/aa9a31fd-776c-4a2a-a4f7-87c23be589e2_origin.pdf new file mode 100644 index 0000000000000000000000000000000000000000..24c74a9b3e8d1f7a256320be8284e31a00295af4 --- /dev/null +++ b/2025/DriveDreamer4D_ World Models Are Effective Data Machines for 4D Driving Scene Representation/aa9a31fd-776c-4a2a-a4f7-87c23be589e2_origin.pdf @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:da3af97993247509987e1360f87de4b36abfebd82d8c6746629b04455cc93d41 +size 2783126 diff --git a/2025/DriveDreamer4D_ World Models Are Effective Data Machines for 4D Driving Scene Representation/full.md b/2025/DriveDreamer4D_ World Models Are Effective Data Machines for 4D Driving Scene Representation/full.md new file mode 100644 index 0000000000000000000000000000000000000000..623665c42a2fbe076db9f77e9e3fcf5c68748902 --- /dev/null +++ b/2025/DriveDreamer4D_ World Models Are Effective Data Machines for 4D Driving Scene Representation/full.md @@ -0,0 +1,379 @@ +# DriveDreamer4D: World Models Are Effective Data Machines for 4D Driving Scene Representation + +Guosheng Zhao $^{1,2,3*}$ , Chaojun Ni $^{4*}$ , Xiaofeng Wang $^{1,2,3*}$ , Zheng Zhu $^{5*†}$ , Xueyang Zhang $^{6}$ , Yida Wang $^{6}$ , Guan Huang $^{5}$ , Xinze Chen $^{5}$ , Boyuan Wang $^{1,2,3}$ , Youyi Zhang $^{7}$ , Wenjun Mei $^{4}$ , Xingang Wang $^{2,3†}$ + +$^{1}$ Institute of Automation, Chinese Academy of Sciences, China + +$^{2}$ School of Artificial Intelligence, University of Chinese Academy of Sciences, China + +$^{3}$ Luoyang Institute for Robot and Intelligent Equipment, China $^{4}$ Peking University, China + +$^{5}$ GigaAI, China $^{6}$ Li Auto Inc., China $^{7}$ Technical University of Munich, Germany + +![](images/6ba133e2acea1579799ea73bcd998fdbe1c2c70ac15f65aa6c0bbf4c5a45d5ae.jpg) +Figure 1. Previous 4D Gaussian Splatting methods (e.g., PVG [9], $\mathbf{S}^3$ Gaussian [31], Deformable-GS [80]) face challenges in rendering novel trajectories, such as lane change. DriveDreamer4D addresses this by enhancing 4D driving scene representation via integrating priors from world models, significantly improving rendering quality under complex scenarios and novel trajectory viewpoints. + +![](images/8efd397d45a96f34d57b8bb9515a9a19b37fd1c7a3f3b5b3d99c2458b9a15631.jpg) + +![](images/072b40ab0261fb08a51bd24b96659ba57628b3314c9abcc27781e29a8099789d.jpg) + +![](images/4070e216b614821aff83b4a7c933acb565f4a799bb1d0c9791c887f96426c61c.jpg) + +![](images/454527b856f3366bf507c346142eaf349a1ce1a499fdd4bd8990f29746c8c851.jpg) + +![](images/ada2abcf22dd8589c7ee99ef00da41b9ebc46b78aea07b0c604ad4f4d8914c58.jpg) + +![](images/e315c8238599f15e77d9cfd012c952f9017e1f56d8b3a16e790c5bdeaa441aa4.jpg) + +# Abstract + +Closed-loop simulation is essential for advancing end-to-end autonomous driving systems. Contemporary sensor simulation methods, such as NeRF and 3DGS, rely predominantly on conditions closely aligned with training data distributions, which are largely confined to forward-driving scenarios. Consequently, these methods face limitations when rendering complex maneuvers (e.g., lane change, acceleration, deceleration). Recent advancements in autonomous-driving world models have demonstrated the potential to generate diverse driving videos. However, these approaches remain constrained to 2D video generation, inherently lacking the spatiotemporal coherence required to capture intricacies of dynamic driving environments. In this paper, we introduce DriveDreamer4D, which + +enhances 4D driving scene representation leveraging world model priors. Specifically, we utilize the world model as a data machine to synthesize novel trajectory videos, where structured conditions are explicitly leveraged to control the spatial-temporal consistency of traffic elements. Besides, the cousin data training strategy is proposed to facilitate merging real and synthetic data for optimizing 4DGS. To our knowledge, DriveDreamer4D is the first to utilize video generation models for improving 4D reconstruction in driving scenarios. Experimental results reveal that DriveDreamer4D significantly enhances generation quality under novel trajectory views, achieving a relative improvement in FID by $32.1\%$ , $46.4\%$ , and $16.3\%$ compared to PVG, $S^3$ Gaussian, and Deformable-GS. Moreover, DriveDreamer4D markedly enhances the spatiotemporal coherence of driving agents, which is verified by a comprehensive user study and the relative increases of $22.6\%$ , $43.5\%$ , and $15.6\%$ in the NTA-IoU metric. + +# 1. Introduction + +End-to-end planning [29, 30, 34], which directly maps sensor inputs to control signals, is among the most critical and promising tasks in autonomous driving. However, current open-loop evaluations are inadequate for accurately assessing end-to-end planning algorithms, highlighting an urgent need for enhanced evaluation methods [41, 43, 86]. A compelling solution lies in closed-loop evaluations within real-world scenarios, which require retrieving sensor data from arbitrarily specified viewpoints. This necessitates constructing a 4D driving scene representation capable of reconstructing complex, dynamic driving environments. + +Closed-loop simulation in driving environments predominantly relies on scene reconstruction techniques such as Neural Radiance Fields (NeRF) [20, 49, 77, 79] and 3D Gaussian Splatting (3DGS) [11, 31, 36, 76], which are inherently limited by the density of input data. Specifically, these methods can render scenes effectively only under conditions closely aligned with their training data distributions—primarily forward-driving scenarios—and struggle to perform accurately during complex maneuvers (see Fig. 1). To mitigate these limitations, methods like SGD [84] and GGS [22] leverage generative models to extend the range of training viewpoints. However, these approaches primarily supplement sparse image data or static background elements, falling short of modeling the intricacies of dynamic, interactive driving scenes. Recently, advancements in autonomous driving world models [18, 28, 66, 68, 69, 87] have introduced the capability to generate diverse, command-aligned video viewpoints, offering renewed promise for closed-loop simulation in autonomous driving. Nonetheless, these models remain constrained to 2D videos, lacking the spatial-temporal coherence essential for accurately modeling complex driving scenarios. + +In this paper, we introduce DriveDreamer4D, which improves 4D driving scene representation by integrating priors from autonomous driving world models. Our approach utilizes an autonomous driving world model [87] as a generative engine, synthesizing novel trajectory video data that densifies real-world driving datasets for enhanced training. Notably, we propose the Novel Trajectory Generation Module (NTGM) to generate diverse structured traffic conditions, and DriveDreamer4D applies these conditions to independently regulate the motion dynamics of foreground and background elements in complex driving environments. These conditions undergo view projection synchronized with vehicle maneuvers, ensuring that the synthesized data adheres to the spatiotemporal constraints. Subsequently, the Cousin Data Training Strategy (CDTS) is proposed to merge temporal-aligned real and synthetic data for training 4DGS. In CDTS, a regularization loss is further incorporated to ensure perceptual coherence. To the best of our knowledge, DriveDreamer4D is the first + +framework to harness video generation models for elevating 4D scene reconstruction quality in autonomous driving, providing richly varied viewpoint data for scenarios including lane change, acceleration, and deceleration. As shown in Fig. 1, experiment results demonstrate that DriveDreamer4D significantly enhances generation fidelity for novel trajectory viewpoints, achieving a relative improvement in FID by $32.1\%$ , $46.4\%$ , and $16.3\%$ compared to PVG [9], $\mathrm{S}^3$ Gaussian [31], and Deformable-GS [80]. Besides, DriveDreamer4D fortifies the spatiotemporal coherence between foreground and background elements, with respective increases of $22.6\%$ , $43.5\%$ , and $15.6\%$ in the NTA-IoU metric. Furthermore, a comprehensive user study confirms that the average win rate of DriveDreamer4D exceeds $80\%$ , compared to three baselines. + +The primary contributions of this work are as follows: (1) We present DriveDreamer4D, the first framework to leverage world model priors for advancing 4D scene reconstruction in autonomous driving. (2) The NTGM is proposed to automate the generation of structured conditions, allowing DriveDreamer4D to create novel trajectory videos with complex maneuvers while ensuring spatial-temporal consistency. Additionally, the CDTS is introduced to merge temporal-aligned real and synthetic data for training 4DGS, using a regularization loss to maintain perceptual coherence. (3) We perform comprehensive experiments to validate that DriveDreamer4D notably enhances generation quality across novel trajectory viewpoints, as well as the spatiotemporal coherence of driving scene elements. + +# 2. Related Work + +# 2.1. Driving Scene Representation + +NeRF and 3DGS have emerged as leading approaches for 3D scene representation. NeRF models [2, 3, 49, 50] continuous volumetric scenes using multi-layer perceptron (MLP) networks, enabling highly detailed scene reconstructions with remarkable rendering quality. More recently, 3DGS [13, 36, 83] introduces an innovative method by defining a set of anisotropic Gaussians in 3D space, leveraging adaptive density control to achieve high-quality renderings from sparse point cloud inputs. Several works have extended NeRF [12, 20, 23, 33, 47, 56, 62, 63, 77, 79] or 3DGS [9, 11, 14, 31, 32, 44, 76, 84, 88, 89] to autonomous driving scenarios. Given the dynamic nature of driving environments, there has also been significant effort in modeling 4D driving scene representations. Some approaches encode time as an additional input to parameterize 4D scenes [1, 15, 31, 42, 46, 52, 60], while others represent scenes as a composition of moving object models alongside a static background model [39, 51, 64, 71, 74, 79]. Despite these advancements, methods based on NeRF and 3DGS face limitations tied to the density of input data. These tech- + +niques can only render scenes effectively when sensor data closely matches the training data distribution, which is typically confined to forward-driving scenarios. + +# 2.2. World Models + +The world model module predicts possible future world states as a function of imagined action sequences proposed by the actor [40, 90]. Approaches such as [4, 5, 19, 21, 25-27, 38, 48, 67, 68, 72, 75, 81, 85] simulate environments through video generation controlled by free-text actions. At the forefront of this evolution is Sora [6], which leverages advanced generative techniques to produce intricate visual sequences that respect the fundamental laws of physics. This ability to deeply understand and simulate the environment not only improves video generation quality but also has substantial implications for real-world driving scenarios. Autonomous driving world models [18, 28, 66, 69, 78, 87] employ predictive methodologies to interpret driving environments, thereby generating realistic driving scenarios and learning key driving elements and policies from video data. Although these models successfully produce diverse driving video data conditioned on complex driving actions, they remain limited to 2D outputs and lack the spatial-temporal coherence needed to accurately capture the complexities of dynamic driving environments. + +# 2.3. Diffusion Prior for 3D Representation + +Constructing comprehensive 3D scenes from limited observations demands generative prior, particularly for unseen areas. Earlier studies distill the knowledge from text-to-image diffusion models [53, 55, 57, 58] into a 3D representation model. Specifically, the Score Distillation Sampling (SDS) [45, 54, 70] is adopted to synthesize a 3D object from the text prompt. Furthermore, to enhance 3D consistency, several approaches extend the multi-view diffusion models [17, 59] and video diffusion models [4, 10, 65] to 3D scene generation. To extend the diffusion prior to complex, dynamic, large-scale driving scenes for 3D reconstruction, methods such as SGD [84], GGS [22] and MagicDrive3D [16] employ generative models to broaden the range of training viewpoints. Nonetheless, these approaches mainly address sparse image data or static background elements, lacking the capacity to fully capture the complexities inherent in the 4D driving environments. + +# 3. Method + +In this section, we first elaborate on the preliminaries of 4D driving scene representation and world models for driving video generation. Then we present the details of DriveDreamer4D, which enhances 4D driving scene representation leveraging priors from driving world models. + +# 3.1. Preliminary + +# 3.1.1. 4D Driving Scene Representation + +4DGS models the driving scene with a collection of 3DGS and a temporal field module. Each 3DGS [36] is parameterized by its center position $x$ , opacity $\gamma$ , covariance $\Sigma$ , and view-dependent RGB color $c$ , controlled via spherical harmonics. For stability, each covariance matrix $\Sigma$ is decomposed by: + +$$ +\boldsymbol {\Sigma} = \boldsymbol {R} \boldsymbol {S} \boldsymbol {S} ^ {T} \boldsymbol {R} ^ {T}, \tag {1} +$$ + +where scaling matrix $S$ and a rotation matrix $\pmb{R}$ are learnable parameters, represented by scaling $\pmb{s}$ and quaternion $\pmb{r}$ . All trainable parameters of a single 3D Gaussian are collectively denoted as $\phi = \{\pmb{x},\pmb{\gamma},\pmb{s},\pmb{r},\pmb{c}\}$ . The temporal field $\mathcal{F}$ takes $\phi$ and a time step $t_{gs}$ as input, outputting the offset $\delta \phi = \{\delta x,\delta \gamma ,\delta s,\delta r,\delta c\}$ for each Gaussian relative to canonical space. The 4D Gaussian $\phi^{\prime} = \{\pmb{x}^{\prime},\pmb{\gamma}^{\prime},\pmb{s}^{\prime},\pmb{r}^{\prime},\pmb{c}^{\prime}\}$ is then computed by: + +$$ +\phi^ {\prime} = \phi + \delta \phi = \phi + \mathcal {F} (\phi , \boldsymbol {t} _ {g s}). \tag {2} +$$ + +Following [82], a differentiable Gaussian Splating renderer is employed to project 4D Gaussians $\phi$ into camera coordinates, yielding the covariance matrix $\boldsymbol{\Sigma}^{\prime} = J\boldsymbol{V}\boldsymbol{\Sigma}\boldsymbol{V}^{T}\boldsymbol{J}^{T}$ , where $\boldsymbol{J}$ is the Jacobian matrix of the perspective projection, and $\boldsymbol{V}$ is the transform matrix. The color of each pixel is calculated by $N$ ordered points using $\alpha$ -blending: + +$$ +C = \sum_ {i \in N} T _ {i} \boldsymbol {c} _ {i} ^ {\prime} \alpha_ {i}, \tag {3} +$$ + +where $T_{i}$ is the transmittance defined by $\prod_{j = 1}^{i - 1}(1 - \alpha_j)$ , $c_{i}^{\prime}$ denotes the color of each point, $\alpha_{i}$ is given by evaluating a 2D Gaussian with covariance $\Sigma^{\prime}$ multiplied with a learned per-point opacity $\gamma_{i}^{\prime}$ . The trainable parameters $\phi^{\prime}$ can be optimized by a combination of RGB loss, depth loss and SSIM loss: + +$$ +\begin{array}{l} \mathcal {L} _ {\mathrm {o r i}} \left(\phi^ {\prime}\right) = \lambda_ {1} \| \hat {I} _ {\mathrm {o r i}} - I _ {\mathrm {o r i}} \| _ {1} + \lambda_ {2} \| \hat {D} _ {\mathrm {o r i}} - D _ {\mathrm {o r i}} \| _ {1} \tag {4} \\ + \lambda_ {3} \operatorname {S S I M} \left(\bar {I} _ {\text {o r i}}, I _ {\text {o r i}}\right), \\ \end{array} +$$ + +where $\hat{I}_{\mathrm{ori}}$ and $I_{\mathrm{ori}}$ represent the rendered image and the ground truth image. $\hat{D}_{\mathrm{ori}}$ and $D_{\mathrm{ori}}$ are the rendered depth and the ground truth LiDAR depth map. SSIM $(\cdot)$ refers to the operation of the Structural Similarity Index Measure, and $\lambda_1,\lambda_2,\lambda_3$ are the loss weights. + +# 3.1.2. World Models for Driving Video Generation + +The world model module predicts possible future world states based on imagined action sequences [40]. Autonomous-driving world models [18, 66, 69, 87], typically based on diffusion models, leverage structured driving information or action controls to guide future video prediction. During training, these models first encode + +![](images/f77c6426aff52ffd7e45ad74d6da743b6570d299e99351b50ce9bf561b2edaf9.jpg) +Figure 2. The overall framework of DriveDreamer4D. Initially, by altering the actions of the original trajectory (e.g., steering angle, speed), new trajectories can be obtained. Conditioned on the first frame and the structured information (3D bounding boxes, HDMap) from the new trajectory, the novel trajectory videos are generated. Subsequently, the temporal-aligned cousin pair (original and novel trajectory videos) are merged to optimize the 4D Gaussian Splatting model, where a regularization loss is calculated to ensure perceptual coherence. + +videos $\pmb{v}$ into a lower-dimensional latent space $z = \mathcal{E}(\pmb{v})$ using a variational encoder $\mathcal{E}$ . After adding noise $\epsilon_{t}$ to the latent, the diffusion model learns a denoising process. This diffusion process is optimized by: + +$$ +\mathcal {L} _ {d i f f} = \mathbb {E} _ {\boldsymbol {z}, \epsilon \sim \mathcal {N} (0, 1), t} \left[ \| \epsilon_ {t} - \epsilon_ {\theta} (\boldsymbol {z} _ {t}, t, \boldsymbol {f}) \| _ {2} ^ {2} \right], \tag {5} +$$ + +where $\epsilon_{\theta}$ is a parameterized denoising network, $t$ denotes the time step, representing the level of noise added or removed at each stage. Additionally, to improve the controllability of the generated data, conditional features $f$ (e.g., reference images, speed, steering angle, scene layouts, camera poses and textual information) can be introduced into the reverse diffusion process, ensuring that the generated outputs adhere to the input control signals. During inference, the world models can be conditioned on a reference image to control the style of the output scene, while predicting the future world states contingent upon the other input actions. + +# 3.2. DriveDreamer4D + +The overall pipeline of DriveDreamer4D is depicted in Fig. 2. In the upper part, the Novel Trajectory Generation + +Module (NTGM) is proposed to adjust driving actions (e.g., steering angle, speed) to generate new trajectories. These novel trajectories provide new perspectives for extracting structured information like 3D boxes and HDMap. Subsequently, a controllable video diffusion model synthesizes videos from these updated viewpoints, incorporating specific priors associated with the modified trajectories. In the lower part, the Cousin Data Training Strategy (CDTS) is introduced to combine the temporal-aligned original and generated data for optimizing the 4DGS model, where a regularization loss is calculated to impose perceptual coherence. In the following sections, we delve into the details of novel trajectories video generation and then introduce the CDTS for 4D reconstruction. + +# 3.2.1. Novel Trajectory Video Generation + +As previously mentioned, traditional 4DGS methods are limited in rendering complex maneuvers, largely due to the training data being dominated by straightforward driving scenarios. To overcome this, DriveDreamer4D leverages world model priors to generate diverse viewpoint data, enhancing the 4D scene representation. To achieve this, + +we propose the NTGM, which is designed to create new trajectories that serve as input for the world model, enabling the automated generation of complex maneuver data. NTGM comprises two main components: (1) novel trajectory proposal, (2) trajectory safety assessment. In the novel trajectory proposal stage, text-to-trajectory [87] can be adopted to automatically generate diverse complex trajectories. Additionally, trajectories can be custom-designed to meet specific requirements, allowing for tailored data generation based on precise needs. The overview of the custom-designed trajectory proposal (e.g., lane change) and trajectory safety assessment is shown in the Algo. 1. In a specific driving scenario, the original trajectory in the world coordinate system can be readily acquired as $\mathcal{T}_{\mathrm{ori}}^{\mathrm{world}} = \{p_i^{\mathrm{world}}\}_{i=0}^{K}$ , where $K$ denotes the number of frames and $p_i^{\mathrm{world}} \in \mathbb{R}^3$ refers to the position of the ego-vehicle at the $i$ -th frame. To propose novel trajectories, the original trajectory $\mathcal{T}_{\mathrm{ori}}^{\mathrm{world}}$ is transformed into the ego-vehicle coordinate system of the first frame, denoted as $\mathcal{T}_{\mathrm{ori}}^{\mathrm{EgoStart}}$ and computed as: + +$$ +\left[ p _ {i} ^ {\text {E g o S t a r t}}, 1 \right] ^ {T} = M _ {0} ^ {- 1} \times \left[ p _ {i} ^ {\text {w o r l d}}, 1 \right] ^ {T}, \tag {6} +$$ + +where $M_0 \in \mathbb{R}^{4 \times 4}$ represents the transformation matrix from the ego-vehicle coordinate system of the first frame to the world coordinate system, $[\cdot]$ denotes the operation of the concat. In the ego-vehicle coordinate system, the vehicle's heading is aligned with the positive $x$ -axis, the $y$ -axis points to the left side of the vehicle, and the $z$ -axis is oriented vertically upwards, perpendicular to the plane of the vehicle. Consequently, changes in the vehicle's velocity and direction can be respectively represented by adjusting the value along the $x$ -axis and $y$ -axis. A final safety assessment is conducted for the newly generated trajectory points, which includes verifying whether the vehicle trajectories $p$ remain within drivable areas $\mathcal{B}_{\mathrm{road}}$ and ensuring that no collisions occur with pedestrians or other vehicles $\{o_j\}_{j=1}^M$ . + +$$ +p \in \mathcal {B} _ {\text {r o a d}}, \tag {7} +$$ + +$$ +\| p - o _ {j} \| \geq d _ {\min }, \quad \forall j \in \{1, \dots , M \}, +$$ + +where $d_{\mathrm{min}}$ is the minimal distance between different agents. Once a novel trajectory that complies with traffic regulations is generated, the road structure and 3D bounding boxes can be projected onto the camera view from the perspective of the new trajectory, thereby generating structured information relative to the updated trajectory. This structured information, along with the initial frame and text, is fed into a world model [87] to produce the videos that follow the novel trajectories. + +# 3.2.2. Cousin Data Training Strategy + +To better integrate generated data for training 4DGS, we propose the CDTS. Specifically, we construct temporally + +Algorithm 1 Novel Trajectory Generation Module +Input: Trajectory $\mathcal{T}_{\mathrm{ori}}^{\mathrm{world}}$ , Transformation matrix $M_0$ +Output: Novel trajectory $\mathcal{T}_{\mathrm{novel}}^{\mathrm{ego}}$ $\mathcal{T}_{\mathrm{novel}}^{\mathrm{ego}}\gets [[0,0,0]]$ +Offset $\leftarrow 0$ +for each $p_{\mathrm{ori}}^{\mathrm{world}}$ in $\mathcal{T}_{\mathrm{ori}}^{\mathrm{world}}$ [1:] do + $p^{\mathrm{EgoStart}}\gets \mathrm{RelativeCoord}(p_{\mathrm{ori}}^{\mathrm{world}},M_0)$ +MaxOffset $\leftarrow 0.1$ +while True do +NewOffset $\leftarrow$ Offset + RandOffset(0, MaxOffset) + $p^{\mathrm{EgoStart}'}\gets p^{\mathrm{EgoStart}} + [0,$ NewOffset, 0] +if SafeCheck $(p^{\mathrm{EgoStart}'}$ ) then +AddElement $(\mathcal{T}_{\mathrm{novel}}^{\mathrm{ego}},p^{\mathrm{EgoStart}'})$ +Offset $\leftarrow$ NewOffset +break +else +MaxOffset $\leftarrow$ MaxOffset/2 +end if +end while +end for + +aligned cousin pair data as a minimal training batch: + +$$ +\operatorname {B a t c h S t a c k} \left(\left\{\hat {I} _ {\text {o r i}, t} \right\} _ {t = 0} ^ {T}, \left\{\hat {I} _ {\text {n o v e l}, t} \right\} _ {t = 0} ^ {T}\right), \tag {8} +$$ + +where $\mathrm{BatchStack}(\cdot)$ is the data processor to stack temporal-aligned $\{\hat{I}_{\mathrm{ori},t}\}_{t=0}^{T}$ and $\{\hat{I}_{\mathrm{novel},t}\}_{t=0}^{T}$ into a training batch. By leveraging real and synthetic data aligned at each timestep, CDTS mitigates the data gap in 4DGS training, enhancing the model's ability to learn consistent representations across real and synthetic data. To optimize 4DGS, the temporal-aligned cousin pair is input per step before gradient optimization. The loss function $\mathcal{L}_{\mathrm{ori}}$ for the original data is defined in Eq. 4. And the loss function $\mathcal{L}_{\mathrm{novel}}$ for the generated data is akin to [9, 31], defined as follows: + +$$ +\begin{array}{l} \mathcal {L} _ {\text {n o v e l}} \left(\phi^ {\prime}\right) = \lambda_ {1} \| \hat {I} _ {\text {n o v e l}} - I _ {\text {n o v e l}} \| _ {1} \tag {9} \\ + \lambda_ {3} \operatorname {S S I M} \left(\hat {I} _ {\text {n o v e l}}, I _ {\text {n o v e l}}\right), \\ \end{array} +$$ + +where $I_{\mathrm{novel}}$ represents the generated images corresponding to the novel trajectories as described in Sec. 3.2.1, and $\hat{I}_{\mathrm{novel}}$ denotes the rendered images under the novel trajectories via differentiable splatting [82]. Notably, different from [9, 31], depth maps are not employed as constraints in the optimization of 4DGS when using the generated dataset $D_{\mathrm{novel}}$ . The limitation arises from the fact that LiDAR point cloud data is exclusively collected for the original trajectory. When these LiDAR points are projected onto a new trajectory, it cannot produce a complete depth map for the new perspective, as something visible in the novel trajectory may have been occluded in the original view. Consequently, the incorporation of such depth maps does not facilitate the optimization of the 4DGS model. More details are described in + +
MethodLane ChangeAccelerationDecelerationAverage
NTA-IoU ↑NTL-IoU ↑NTA-IoU ↑NTL-IoU ↑NTA-IoU ↑NTL-IoU ↑NTA-IoU ↑NTL-IoU ↑
PVG [9]0.25650.700.39653.080.39453.650.34952.48
DriveDreamer4D with PVG0.43853.060.42153.350.42453.890.42853.43
S3Gaussian [31]0.17549.050.43451.930.38452.140.33151.04
DriveDreamer4D with S3Gaussian0.49553.420.48452.630.44552.690.47552.91
Deformable-GS [80]0.24051.620.34652.170.37753.210.32152.33
DriveDreamer4D with Deformable-GS0.33552.930.37152.770.40653.790.37153.16
+ +Table 1. Comparison of NTA-IoU and NTL-IoU scores across different novel trajectory views (lane change, acceleration, deceleration). + +
MethodFID ↓
PVG[9]105.29
DriveDreamer4D with PVG71.52
S3Gaussian[31]124.90
DriveDreamer4D with S3Gaussian66.93
Deformable-GS[80]92.34
DriveDreamer4D with Deformable-GS77.32
+ +Table 2. Comparison of FID scores in novel trajectory view synthesis (lane change) on the Waymo dataset. + +Sec. 4.3. Additionally, we propose a regularization loss to enhance the perceptual coherence, defined as follows: + +$$ +\mathcal {L} _ {\text {r e g}} \left(\phi^ {\prime}\right) = \left\| \mathcal {F} _ {p} \left(\hat {I} _ {\text {o r i}}\right) - \mathcal {F} _ {p} \left(\hat {I} _ {\text {n o v e l}}\right) \right\| _ {1}, \tag {10} +$$ + +where $\mathcal{F}_p(\cdot)$ denotes the perception feature extraction model [24]. The overall loss function for mixed training is defined as follows: + +$$ +\mathcal {L} \left(\phi^ {\prime}\right) = \mathcal {L} _ {\text {o r i}} + \lambda_ {\text {n o v e l}} \mathcal {L} _ {\text {n o v e l}} + \lambda_ {\text {r e g}} \mathcal {L} _ {\text {r e g}}. \tag {11} +$$ + +# 4. Experiments + +In this section, we first outline the experimental setup, including the dataset, implementation details, and evaluation metrics. Subsequently, quantitative results across different datasets are provided to demonstrate that the proposed DriveDreamer4D significantly enhances rendering quality for novel trajectory viewpoints and improves the spatiotemporal coherence of foreground and background components. Finally, we conduct an ablation study on the hyperparameter settings, as well as the effects of depth loss and the proposed CDTS, including temporal-aligned pairs, and regularization loss. + +# 4.1. Experiment Setup + +Dataset. We conduct experiments using the Waymo dataset [61], known for its comprehensive real-world driving logs. We specifically select eight scenes characterized by highly + +dynamic interactions, featuring numerous vehicles with diverse relative positions and intricate driving trajectories. Each selected segment contains approximately 40 frames, with segment IDs detailed in the supplement. In addition, we have also conduct experiments on the nuscenes [7], and pandas [73] datasets. For nuScenes, we select eight scenes rich in dynamic objects and complex lane structures to test large maneuver capabilities. For PandaSet, we adopt the ten scenes used in Unisim [79] and NeuRAD [64] for fair comparison. + +Implementation Details. To demonstrate the versatility and robustness of DriveDreamer4D, we incorporate various 4DGS baselines into our pipeline, including Deformable-GS [80], S $^3$ Gaussian [31], and PVG [9]. For a fair comparison, LiDAR supervision is introduced to Deformable-GS. Scenes are segmented into 40-frame clips, matching the generative model's output length, using only forward-facing camera data at a resolution of $640 \times 960$ . Models are trained with the Adam optimizer [37], following the learning rate schedule from 3D Gaussian Splatting [36]. Hyperparameters align with each baseline [9, 31, 80], with training strategies unchanged except for incorporating CDTS. Moreover, to demonstrate the robustness of our method across different datasets, we employ PVG and DriveDreamer4D with PVG on nuScenes [7], and PandaSet [73]. Since existing methods typically conduct experiments on one or two datasets with varying scene selections, we have chosen different state-of-the-art (SOTA) methods for comparison on each dataset. This approach ensures a comprehensive evaluation of our method's performance under diverse conditions and highlights its adaptability and effectiveness across a wide range of scenarios. By benchmarking against various SOTA methods tailored to specific datasets, we provide strong evidence of the generalizability and robustness of DriveDreamer4D in handling complex and varied driving environments. + +Metrics. Traditional 3D reconstruction tasks typically employ PSNR and SSIM metrics for evaluation, with validation sets that closely match the training data distribution (i.e., uniformly sampling frames from video sequences for validation, with the remainder used for training). However, in closed-loop driving simulation, the focus shifts to evaluating model rendering performance under novel trajectories, + +
Counterpart MethodDriveDreamer4D Win Rate
Lane ChangeAccelerationDecelerationAverage
PVG [9]100.0%90.5%89.1%93.2%
S3Gaussian [31]100.0%97.9%92.2%96.7%
Deformable-GS [80]95.8%83.5%72.9%84.1%
+ +where corresponding sensor data are unavailable, making metrics like PSNR and SSIM inapplicable for evaluation. Therefore, we propose Novel Trajectory Agent IoU (NTA-IoU) and Novel Trajectory Lane IoU (NTL-IoU), which assess the spatiotemporal coherence of foreground and background traffic components in novel trajectory viewpoints. + +For NTA-IoU, we use YOLO11 [35] to identify vehicles in images rendered from novel trajectory views, yielding 2D bounding boxes. Simultaneously, geometric transformations are applied to the original 3D bounding boxes, projecting them onto the new viewpoints to generate corresponding 2D bounding boxes. For each projected 2D box, we then identify the closest detector-generated 2D box and compute their Intersection over Union (IoU). To ensure accurate matching, a distance threshold $d_{\mathrm{thresh}}$ is introduced: when the center-to-center distance $\| c(B^{\mathrm{proj}}) - c(B^{\mathrm{det}})\|$ between the nearest detected box $B^{\mathrm{det}}$ and the correctly projected box $B^{\mathrm{proj}}$ surpasses this threshold, their NTA-IoU is assigned a value of zero: + +$$ +\mathrm {N T A - I o U} = \left\{ \begin{array}{l l} 0 & \text {i f} \| c (B ^ {\operatorname {p r o j}}) - c (B ^ {\det }) \| \geq d _ {\text {t h r e s h}} \\ \operatorname {I o U} \left(B ^ {\operatorname {p r o j}}, B ^ {\det }\right) & \text {o t h e r w i s e .} \end{array} \right. \tag {12} +$$ + +For NTL-IoU, we employ TwinLiteNet [8] to extract 2D lanes from rendered images. Ground truth lanes are also projected onto the 2D image plane. We then compute the mean Intersection-over-Union (mIoU) between the rendered and ground truth lanes $L^{\mathrm{det}}$ and $L^{\mathrm{proj}}$ : + +$$ +\mathrm {N T L - I o U} = \mathrm {m I o U} \left(L ^ {\text {p r o j}}, L ^ {\text {d e t}}\right). \tag {13} +$$ + +Additionally, in lane change scenarios, we observe inaccuracies in relative positioning, as well as frequent occurrences of artifacts such as flying points and ghosting, which notably degrade image quality. To assess this, we employ the FID metric [24], which quantifies differences in feature distribution between rendered novel trajectory images and original trajectory images. This metric effectively reflects visual quality and is particularly sensitive to artifacts like flying points and ghosting, providing a robust measure of image fidelity in these complex scenes. Finally, a user study is conducted to evaluate the quality of the renderings, where participants compare the rendering results of each baseline with its DriveDreamer4D enhanced version across three novel trajectories. The evaluation criteria + +Table 3. User study comparison of DriveDreamer4D win rates across various novel trajectory view synthesis. + +
MethodNTA-IoU ↑NTL-IoU ↑FID ↓
PVG[9]0.20448.75103.92
DriveDreamer4D with PVG0.33850.3688.60
+ +Table 4. Comparison of NTA-IoU, NTL-IoU and FID scores across lane change on nuScenes. + +
MethodNo shift ↓Lane 2m ↓Lane 3m ↓Vert. 1m ↓
UniSim [79]-74.797.5-
NeuRAD [64]25.072.393.976.3
PVG [9]24.767.376.882.9
DriveDreamer4D with PVG23.562.474.669.8
+ +Table 5. FID when shifting pose of ego vehicle on PandaSet. + +focus on overall video quality, with particular attention to foreground objects like vehicles. For each comparison, participants were asked to select the option they found most favorable. Further details are provided in the supplement. + +# 4.2. Main Results + +In this subsection, we first demonstrate the effectiveness of DriveDreamer4D with various baseline methods. Then, through experimental results obtained from three different datasets, combined with detailed comparisons against a range of baselines, we highlight the robustness and adaptability of DriveDreamer4D in addressing a broad spectrum of scenarios and challenges. + +Waymo. As demonstrated in Tab. 1, integrating DriveDreamer4D with various 4DGS algorithms consistently yields superior NTA-IoU and NTL-IoU scores across complex maneuvers (e.g., lane changes, acceleration, and deceleration), surpassing baseline methods. Specifically, DriveDreamer4D enhances average NTA-IoU by $22.6\%$ , $43.5\%$ , and $15.6\%$ across three baselines (PVG [9], $\mathrm{S}^3$ Gaussian [31], Deformable-GS [80]), underscoring its capability to improve the spatiotemporal coherence of foreground agents. It also boosts average NTL-IoU by $1.8\%$ , $3.7\%$ , and $1.6\%$ , enhancing background lane consistency in 4D driving scene rendering. + +In addition to verifying the spatiotemporal consistency of rendered novel trajectory views, we leverage the FID metric to assess rendering quality under novel trajectories. Given that acceleration and deceleration scenarios yield rendered views with distributional similarities to ground truth, limiting FID's discriminative capability across algorithms, our FID comparisons focus specifically on lane change scenarios. Experiment results, as presented in Tab. 2, indicate that our method substantially outperforms the baseline methods (PVG [9], $\mathrm{S}^3$ Gaussian [31], Deformable-GS [80]), with FID relative improvements of $32.1\%$ , $46.4\%$ , and $16.3\%$ . These results highlight DriveDreamer4D's capability to enhance generation quality for + +
λnovelNTA-IoU ↑FID ↓
00.349105.29
0.50.40582.84
10.42079.54
1.50.41782.10
+ +Table 6. Ablation study on the training loss weight $\lambda_{\mathrm{novel}}$ for novel trajectory data. + +
λregNTA-IoU ↑FID ↓
00.42079.54
1e-20.411119.39
1e-30.42871.52
1e-40.42275.31
+ +novel trajectory viewpoints. + +Finally, we conduct a user study to evaluate the rendering quality of different methods on novel trajectories, with a specific focus on foreground agents. For each method, we generate three novel trajectory views—lane change, acceleration, and deceleration—across eight scenes from the Waymo dataset [61]. Participants are then asked to select the renderings they found most visually favorable in each comparison. The DriveDreamer4D win rates from this study, shown in Tab. 3, reveal a significant user preference for our method's renderings. + +nuScenes. Tab. 4 provides a comprehensive performance comparison between DriveDreamer4D and PVG [9] in lane change scenarios on the nuScenes dataset [7]. The experimental results offer compelling evidence that our method achieves a substantial improvement over the baseline, showcasing its superior capability in handling complex driving situations. Specifically, as detailed in the table, DriveDreamer4D consistently outperforms PVG across multiple evaluation metrics, including NTA-IoU, NTL-IoU, and FID. Notably, our method achieves a $65.7\%$ increase in NTA-IoU, a $3.3\%$ increase in NTL-IoU, and a $14.7\%$ reduction in FID. These results underscore the robustness of DriveDreamer4D in addressing challenging driving scenarios, highlighting its exceptional performance and adaptability across diverse evaluation criteria. Such advancements reaffirm the effectiveness of our approach in delivering more accurate and reliable trajectory predictions for dynamic environments. + +PandaSet. Following the experimental setup of NeuRAD [64] and UniSim [79], we conduct experiments on the PandaSet dataset [73]. As shown in Tab. 5, we report the FID comparison under the original trajectory, a lane shift of 2 meters, a lane shift of 3 meters, and a perpendicular shift of 1 meter. These results demonstrate that our DriveDreamer4D achieves the best performance across all settings. Notably, DriveDreamer4D achieves an FID of 74.6 under the challenging scenario of a 3-meter lane shift, representing a significant improvement of $20.6\%$ over the previous SOTA method, NeuRAD. This highlights the strong performance and competitiveness of our approach com + +Table 7. Ablation study on the regularization loss weight ${\lambda }_{\text{reg }}$ . + +
Depth LossTemporal-aligned Cousin PairRegularization LossNTA-IoU ↑FID ↓
✓××0.40182.63
×××0.42079.54
×✓×0.42376.20
×✓✓0.42871.52
+ +Table 8. Ablation study on the depth loss and CDTS. + +pared to existing techniques, reaffirming its effectiveness in handling complex driving scenarios with high precision and robustness. + +# 4.3. Ablation Studies + +We conduct an ablation study based on PVG [9], analyzing the effects of hyperparameters settings for $\lambda_{\mathrm{novel}}$ and $\lambda_{\mathrm{reg}}$ , as well as the impact of depth loss, temporal-aligned cousin pairs, and regularization loss. Following the experiments shown in Tab. 6 and 7, we set $\lambda_{\mathrm{novel}}$ to 1 and $\lambda_{\mathrm{reg}}$ to $1\times 10^{-3}$ which yield the best performance. As shown in Tab. 8, the experiment confirms that the depth loss should be excluded when optimizing novel trajectory views, since LiDAR depth maps are incomplete due to occlusions. Furthermore, employing temporal-aligned cousin pairs achieves an FID of 76.20 and an NTA-IoU of 0.423, with the perception regularization loss further improving the FID to 71.52 and NTA-IoU to 0.428. These results highlight the effectiveness of CDTS, in achieving $\sim 10\%$ improvement in FID, along with a $\sim 2\%$ increase in NTA-IoU. + +# 5. Discussion and Conclusion + +In this paper, we present DriveDreamer4D, a novel framework designed to advance 4D driving scene representations by harnessing priors from world models. Addressing key limitations of current sensor simulation methods—namely, their dependence on forward-driving training data distributions and inability to model complex maneuvers—DriveDreamer4D leverages a world model to generate novel trajectory videos that complement real-world driving data. By explicitly employing structured conditions, our framework maintains spatial-temporal consistency across traffic elements, ensuring that generated data adheres closely to the dynamics of real-world traffic scenarios. Our experiments demonstrate that DriveDreamer4D achieves superior quality in generating diverse simulation viewpoints, with significant improvements in both the rendering fidelity and spatiotemporal coherence of scene components. Notably, these results highlight DriveDreamer4D's potential as a foundation for closed-loop simulations that require high-fidelity reconstructions of dynamic driving scenes. + +# References + +[1] Benjamin Attal, Jia-Bin Huang, Christian Richardt, Michael Zollhoefer, Johannes Kopf, Matthew O'Toole, and Changil Kim. Hyperreel: High-fidelity 6-dof video with ray-conditioned sampling. In CVPR, 2023. 2 +[2] Jonathan T Barron, Ben Mildenhall, Dor Verbin, Pratul P Srinivasan, and Peter Hedman. Mip-nerf 360: Unbounded anti-aliased neural radiance fields. In CVPR, 2022. 2 +[3] Jonathan T Barron, Ben Mildenhall, Dor Verbin, Pratul P Srinivasan, and Peter Hedman. Zip-nerf: Anti-aliased grid-based neural radiance fields. In ICCV, 2023. 2 +[4] Andreas Blattmann, Tim Dockhorn, Sumith Kulal, Daniel Mendelevitch, Maciej Kilian, Dominik Lorenz, Yam Levi, Zion English, Vikram Voleti, Adam Letts, et al. Stable video diffusion: Scaling latent video diffusion models to large datasets. arXiv preprint arXiv:2311.15127, 2023. 3 +[5] Andreas Blattmann, Robin Rombach, Huan Ling, Tim Dockhorn, Seung Wook Kim, Sanja Fidler, and Karsten Kreis. Align your latents: High-resolution video synthesis with latent diffusion models. In CVPR, 2023. 3 +[6] Tim Brooks, Bill Peebles, Connor Holmes, Will DePue, Yufei Guo, Li Jing, David Schnurr, Joe Taylor, Troy Luhman, Eric Luhman, Clarence Ng, Ricky Wang, and Aditya Ramesh. Video generation models as world simulators. 2024. 3 +[7] Holger Caesar, Varun Bankiti, Alex H Lang, Sourabh Vora, Venice Erin Liong, Qiang Xu, Anush Krishnan, Yu Pan, Giancarlo Baldan, and Oscar Beijbom. nuscenes: A multimodal dataset for autonomous driving. In CVPR, 2020. 6, 8 +[8] Quang-Huy Che, Dinh-Phuc Nguyen, Minh-Quan Pham, and Duc-Khai Lam. Twinlitenet: An efficient and lightweight model for driveable area and lane segmentation in self-driving cars. In MAPR, 2023. 7 +[9] Yurui Chen, Chun Gu, Junzhe Jiang, Xiatian Zhu, and Li Zhang. Periodic vibration gaussian: Dynamic urban scene reconstruction and real-time rendering. arXiv preprint arXiv:2311.18561, 2023. 1, 2, 5, 6, 7, 8 +[10] Zilong Chen, Yikai Wang, Feng Wang, Zhengyi Wang, and Huaping Liu. V3d: Video diffusion models are effective 3d generators. arXiv preprint arXiv:2403.06738, 2024. 3 +[11] Ziyu Chen, Jiawei Yang, Jiahui Huang, Riccardo de Lutio, Janick Martinez Esturo, Boris Ivanovic, Or Litany, Zan Gojcic, Sanja Fidler, Marco Pavone, Li Song, and Yue Wang. Omnire: Omni urban scene reconstruction. arXiv preprint arXiv:2408.16760, 2024. 2 +[12] Kai Cheng, Xiaoxiao Long, Wei Yin, Jin Wang, Zhiqiang Wu, Yuexin Ma, Kaixuan Wang, Xiaozhi Chen, and Xuejin Chen. Uc-nerf: Neural radiance field for under-calibrated multi-view cameras in autonomous driving. arXiv preprint arXiv:2311.16945, 2023. 2 +[13] Xiaobiao Du, Yida Wang, and Xin Yu. Mvgs: Multi-view-regulated gaussian splattingg for novel view synthesis. arXiv preprint arXiv:2410.02103, 2024. 2 +[14] Tobias Fischer, Jonas Kulhanek, Samuel Rota Bulo, Lorenzo Porzi, Marc Pollefeys, and Peter Kontschieder. Dynamic 3d gaussian fields for urban areas. NeurIPS, 2024. 2 + +[15] Sara Fridovich-Keil, Giacomo Meanti, Frederik Rahbæk Warburg, Benjamin Recht, and Angjoo Kanazawa. K-planes: Explicit radiance fields in space, time, and appearance. In CVPR, 2023. 2 +[16] Ruiyuan Gao, Kai Chen, Zhihao Li, Lanqing Hong, Zhenguo Li, and Qiang Xu. Magicdrive3d: Controllable 3d generation for any-view rendering in street scenes. arXiv preprint arXiv:2405.14475, 2024. 3 +[17] Ruiqi Gao, Aleksander Holynski, Philipp Henzler, Arthur Brussee, Ricardo Martin-Brualla, Pratul Srinivasan, Jonathan T Barron, and Ben Poole. Cat3d: Create anything in 3d with multi-view diffusion models. arXiv preprint arXiv:2405.10314, 2024. 3 +[18] Shenyuan Gao, Jiazhi Yang, Li Chen, Kashyap Chitta, Yihang Qiu, Andreas Geiger, Jun Zhang, and Hongyang Li. Vista: A generalizable driving world model with high fidelity and versatile controllability. arXiv preprint arXiv:2405.17398, 2024. 2, 3 +[19] Rohit Girdhar, Mannat Singh, Andrew Brown, Quentin Duval, Samaneh Azadi, Sai Saketh Rambhatla, Akbar Shah, Xi Yin, Devi Parikh, and Ishan Misra. Emu video: Factorizing text-to-video generation by explicit image conditioning. arXiv preprint arXiv:2311.10709, 2023. 3 +[20] Jianfei Guo, Nianchen Deng, Xinyang Li, Yeqi Bai, Botian Shi, Chiyu Wang, Chenjing Ding, Dongliang Wang, and Yikang Li. Streetsurf: Extending multi-view implicit surface reconstruction to street views. arXiv preprint arXiv:2306.04988, 2023. 2 +[21] Agrim Gupta, Lijun Yu, Kihyuk Sohn, Xiuye Gu, Meera Hahn, Li Fei-Fei, Irfan Essa, Lu Jiang, and José Lezama. Photorealistic video generation with diffusion models. arXiv preprint arXiv:2312.06662, 2023. 3 +[22] Huasong Han, Kaixuan Zhou, Xiaoxiao Long, Yusen Wang, and Chunxia Xiao. Ggs: Generalizable gaussian splatting for lane switching in autonomous driving. arXiv preprint arXiv:2409.02382, 2024. 2, 3 +[23] Ayaan Haque, Matthew Tancik, Alexei A Efros, Aleksander Holynski, and Angjoo Kanazawa. Instruct-nerf2nerf: Editing 3d scenes with instructions. In CVPR, 2023. 2 +[24] Martin Heusel, Hubert Ramsauer, Thomas Unterthiner, Bernhard Nessler, and Sepp Hochreiter. Gans trained by a two time-scale update rule converge to a local nash equilibrium. NeurIPS, 2017. 6, 7 +[25] Jonathan Ho, William Chan, Chitwan Sahara, Jay Whang, Ruiqi Gao, Alexey Gritsenko, Diederik P Kingma, Ben Poole, Mohammad Norouzi, David J Fleet, et al. Imagen video: High definition video generation with diffusion models. arXiv preprint arXiv:2210.02303, 2022. 3 +[26] Jonathan Ho, Tim Salimans, Alexey Gritsenko, William Chan, Mohammad Norouzi, and David J Fleet. Video diffusion models. NeurIPS, 2022. +[27] Wenyi Hong, Ming Ding, Wendi Zheng, Xinghan Liu, and Jie Tang. Cogvideo: Large-scale pretraining for text-to-video generation via transformers. arXiv preprint arXiv:2205.15868, 2022.3 +[28] Anthony Hu, Lloyd Russell, Hudson Yeo, Zak Murez, George Fedoseev, Alex Kendall, Jamie Shotton, and Gian + +luca Corrado. Gaia-1: A generative world model for autonomous driving. arXiv preprint arXiv:2309.17080, 2023. 2, 3 +[29] Shengchao Hu, Li Chen, Penghao Wu, Hongyang Li, Junchi Yan, and Dacheng Tao. St-p3: End-to-end vision-based autonomous driving via spatial-temporal feature learning. In ECCV, 2022. 2 +[30] Yihan Hu, Jiazhi Yang, Li Chen, Keyu Li, Chonghao Sima, Xizhou Zhu, Siqi Chai, Senyao Du, Tianwei Lin, Wenhai Wang, et al. Planning-oriented autonomous driving. In CVPR, 2023. 2 +[31] Nan Huang, Xiaobao Wei, Wenzhao Zheng, Pengju An, Ming Lu, Wei Zhan, Masayoshi Tomizuka, Kurt Keutzer, and Shanghang Zhang. $s^3$ gaussian: Self-supervised street gaussians for autonomous driving. arXiv preprint arXiv:2405.20323, 2024. 1, 2, 5, 6, 7 +[32] Sungwon Hwang, Min-Jung Kim, Taewoong Kang, Jayeon Kang, and Jaegul Choo. Vegs: View extrapolation of urban scenes in 3d gaussian splatting using learned priors. In ECCV, 2024. 2 +[33] Muhammad Zubair Irshad, Sergey Zakharov, Katherine Liu, Vitor Guizilini, Thomas Kollar, Adrien Gaidon, Zsolt Kira, and Rares Ambrus. Neo 360: Neural fields for sparse view synthesis of outdoor scenes. In ICCV, 2023. 2 +[34] Bo Jiang, Shaoyu Chen, Qing Xu, Bencheng Liao, Jiajie Chen, Helong Zhou, Qian Zhang, Wenyu Liu, Chang Huang, and Xinggang Wang. Vad: Vectorized scene representation for efficient autonomous driving. In ICCV, 2023. 2 +[35] Glenn Jocher and Jing Qiu. Ultralytics yolo11, 2024. 7 +[36] Bernhard Kerbl, Georgios Kopanas, Thomas Leimkuhler, and George Drettakis. 3d gaussian splatting for real-time radiance field rendering. ACM ToG, 2023. 2, 3, 6 +[37] Diederik P Kingma and Jimmy Ba. Adam: A method for stochastic optimization. arXiv preprint arXiv:1412.6980, 2014. 6 +[38] Dan Kondratyuk, Lijun Yu, Xiuye Gu, José Lezama, Jonathan Huang, Rachel Hornung, Hartwig Adam, Hassan Akbari, Yair Alon, Vighnesh Birodkar, et al. Videopoet: A large language model for zero-shot video generation. arXiv preprint arXiv:2312.14125, 2023. 3 +[39] Abhijit Kundu, Kyle Genova, Xiaoqi Yin, Alireza Fathi, Caroline Pantofaru, Leonidas J Guibas, Andrea Tagliasacchi, Frank Dellaert, and Thomas Funkhouser. Panoptic neural fields: A semantic object-aware neural scene representation. In CVPR, 2022. 2 +[40] Yann LeCun and Courant. A path towards autonomous machine intelligence version 0.9.2, 2022-06-27. 2022. 3 +[41] Hao Li, Ming Yuan, Yan Zhang, Chenming Wu, Chen Zhao, Chunyu Song, Haocheng Feng, Errui Ding, Dingwen Zhang, and Jingdong Wang. Xld: A cross-lane dataset for benchmarking novel driving view synthesis. arXiv preprint arXiv:2406.18360, 2024. 2 +[42] Zhengqi Li, Simon Niklaus, Noah Snavely, and Oliver Wang. Neural scene flow fields for space-time view synthesis of dynamic scenes. In CVPR, 2021. 2 +[43] Zhiqi Li, Zhiding Yu, Shiyi Lan, Jiahan Li, Jan Kautz, Tong Lu, and Jose M Alvarez. Is ego status all you need for open-loop end-to-end autonomous driving? In CVPR, 2024. 2 + +[44] Zhuopeng Li, Yilin Zhang, Chenming Wu, Jianke Zhu, and Liangjun Zhang. Ho-gaussian: Hybrid optimization of 3d gaussian splatting for urban scenes. arXiv preprint arXiv:2403.20032, 2024. 2 +[45] Chen-Hsuan Lin, Jun Gao, Luming Tang, Towaki Takikawa, Xiaohui Zeng, Xun Huang, Karsten Kreis, Sanja Fidler, Ming-Yu Liu, and Tsung-Yi Lin. Magic3d: High-resolution text-to-3d content creation. In CVPR, 2023. 3 +[46] Haotong Lin, Sida Peng, Zhen Xu, Yunzhi Yan, Qing Shuai, Hujun Bao, and Xiaowei Zhou. Efficient neural radiance fields for interactive free-viewpoint video. In SIGGRAPH Asia, 2022. 2 +[47] Fan Lu, Yan Xu, Guang Chen, Hongsheng Li, Kwan-Yee Lin, and Changjun Jiang. Urban radiance field representation with deformable neural mesh primitives. In ICCV, 2023. 2 +[48] Xin Ma, Yaohui Wang, Gengyun Jia, Xinyuan Chen, Zwei Liu, Yuan-Fang Li, Cunjian Chen, and Yu Qiao. Latte: Latent diffusion transformer for video generation. arXiv preprint arXiv:2401.03048, 2024. 3 +[49] Ben Mildenhall, Pratul P Srinivasan, Matthew Tancik, Jonathan T Barron, Ravi Ramamoorthi, and Ren Ng. Nerf: Representing scenes as neural radiance fields for view synthesis. CACM, 2021. 2 +[50] Thomas Müller, Alex Evans, Christoph Schied, and Alexander Keller. Instant neural graphics primitives with a multiresolution hash encoding. ACM ToG, 2022. 2 +[51] Julian Ost, Fahim Mannan, Nils Thuerey, Julian Knodt, and Felix Heide. Neural scene graphs for dynamic scenes. In CVPR, 2021. 2 +[52] Keunhong Park, Utkarsh Sinha, Peter Hedman, Jonathan T Barron, Sofien Bouaziz, Dan B Goldman, Ricardo MartinBrualla, and Steven M Seitz. Hypernerf: A higher-dimensional representation for topologically varying neural radiance fields. arXiv preprint arXiv:2106.13228, 2021. 2 +[53] Dustin Podell, Zion English, Kyle Lacey, Andreas Blattmann, Tim Dockhorn, Jonas Müller, Joe Penna, and Robin Rombach. Sdxl: Improving latent diffusion models for high-resolution image synthesis. arXiv preprint arXiv:2307.01952, 2023. 3 +[54] Ben Poole, Ajay Jain, Jonathan T Barron, and Ben Mildenhall. Dreamfusion: Text-to-3d using 2d diffusion. arXiv preprint arXiv:2209.14988, 2022. 3 +[55] Aditya Ramesh, Prafulla Dhariwal, Alex Nichol, Casey Chu, and Mark Chen. Hierarchical text-conditional image generation with clip latents. arXiv preprint arXiv:2204.06125, 2022.3 +[56] Konstantinos Rematas, Andrew Liu, Pratul P Srinivasan, Jonathan T Barron, Andrea Tagliasacchi, Thomas Funkhouser, and Vittorio Ferrari. Urban radiance fields. In CVPR, 2022. 2 +[57] Robin Rombach, Andreas Blattmann, Dominik Lorenz, Patrick Esser, and Björn Ommher. High-resolution image synthesis with latent diffusion models. In CVPR, 2022. 3 +[58] Chitwan Sahara, William Chan, Saurabh Saxena, Lala Li, Jay Whang, Emily L Denton, Kamyar Ghasemipour, Raphael Gontijo Lopes, Burcu Karagol Ayan, Tim Salimans, et al. Photorealistic text-to-image diffusion models with deep language understanding. NeurIPS, 2022. 3 + +[59] Kyle Sargent, Zizhang Li, Tanmay Shah, Charles Herrmann, Hong-Xing Yu, Yunzhi Zhang, Eric Ryan Chan, Dmitry Lagun, Li Fei-Fei, Deqing Sun, et al. Zeronvs: Zero-shot 360-degree view synthesis from a single real image. arXiv preprint arXiv:2310.17994, 2023. 3 +[60] Liangchen Song, Anpei Chen, Zhong Li, Zhang Chen, Lele Chen, Junsong Yuan, Yi Xu, and Andreas Geiger. Nerfplayer: A streamable dynamic scene representation with decomposed neural radiance fields. IEEE TVCG, 2023. 2 +[61] Pei Sun, Henrik Kretzschmar, Xerxes Dotiwalla, Aurelien Chouard, Vijaysai Patnaik, Paul Tsui, James Guo, Yin Zhou, Yuning Chai, Benjamin Caine, Vijay Vasudevan, Wei Han, Jiquan Ngiam, Hang Zhao, Aleksei Timofeev, Scott Ettinger, Maxim Krivokon, Amy Gao, Aditya Joshi, Yu Zhang, Jonathon Shlens, Zhifeng Chen, and Dragomir Anguelov. Scalability in perception for autonomous driving: Waymo open dataset. In CVPR, 2020. 6, 8 +[62] Shanlin Sun, Bingbing Zhuang, Ziyu Jiang, Buyu Liu, Xiaohui Xie, and Manmohan Chandraker. Lidarf: Delving into lidar for neural radiance field on street scenes. In CVPR, 2024. 2 +[63] Matthew Tancik, Vincent Casser, Xinchen Yan, Sabeek Pradhan, Ben Mildenhall, Pratul P Srinivasan, Jonathan T Barron, and Henrik Kretzschmar. Block-nerf: Scalable large scene neural view synthesis. In CVPR, 2022. 2 +[64] Adam Tonderski, Carl Lindström, Georg Hess, William Ljungbergh, Lennart Svensson, and Christoffer Petersson. Neurad: Neural rendering for autonomous driving. In CVPR, 2024. 2, 6, 7, 8 +[65] Vikram Voleti, Chun-Han Yao, Mark Boss, Adam Letts, David Pankratz, Dmitry Tochilkin, Christian Laforte, Robin Rombach, and Varun Jampani. Sv3d: Novel multi-view synthesis and 3d generation from a single image using latent video diffusion. arXiv preprint arXiv:2403.12008, 2024. 3 +[66] Xiaofeng Wang, Zheng Zhu, Guan Huang, Xinze Chen, Jiagang Zhu, and Jiwen Lu. Drivedreamer: Towards real-world-driven world models for autonomous driving. arXiv preprint arXiv:2309.09777, 2023. 2, 3 +[67] Xiaofeng Wang, Kang Zhao, Feng Liu, Jiayu Wang, Guosheng Zhao, Xiaoyi Bao, Zheng Zhu, Yingya Zhang, and Xingang Wang. Egovid-5m: A large-scale video-action dataset for egocentric video generation. arXiv preprint arXiv:2411.08380, 2024. 3 +[68] Xiaofeng Wang, Zheng Zhu, Guan Huang, Boyuan Wang, Xinze Chen, and Jiwen Lu. Worlddreamer: Towards general world models for video generation via predicting masked tokens. arXiv preprint arXiv:2401.09985, 2024. 2, 3 +[69] Yuqi Wang, Jiawei He, Lue Fan, Hongxin Li, Yuntao Chen, and Zhaoxiang Zhang. Driving into the future: Multiview visual forecasting and planning with world model for autonomous driving. In CVPR, 2024. 2, 3 +[70] Rundi Wu, Ben Mildenhall, Philipp Henzler, Keunhong Park, Ruiqi Gao, Daniel Watson, Pratul P Srinivasan, Dor Verbin, Jonathan T Barron, Ben Poole, et al. Reconfusion: 3d reconstruction with diffusion priors. In CVPR, 2024. 3 +[71] Zirui Wu, Tianyu Liu, Liyi Luo, Zhide Zhong, Jianteng Chen, Hongmin Xiao, Chao Hou, Haozhe Lou, Yuantao + +Chen, Runyi Yang, et al. Mars: An instance-aware, modular and realistic simulator for autonomous driving. In ICAI, 2023. 2 +[72] Jiannan Xiang, Guangyi Liu, Yi Gu, Qiyue Gao, Yuting Ning, Yuheng Zha, Zeyu Feng, Tianhua Tao, Shibo Hao, Yemin Shi, et al. Pandora: Towards general world model with natural language actions and video states. arXiv preprint arXiv:2406.09455, 2024. 3 +[73] Pengchuan Xiao, Zhenlei Shao, Steven Hao, Zishuo Zhang, Xiaolin Chai, Judy Jiao, Zesong Li, Jian Wu, Kai Sun, Kun Jiang, et al. Pandaset: Advanced sensor suite dataset for autonomous driving. In ITSC, 2021. 6, 8 +[74] Ziyang Xie, Junge Zhang, Wenye Li, Feihu Zhang, and Li Zhang. S-nerf: Neural radiance fields for street views. arXiv preprint arXiv:2303.00749, 2023. 2 +[75] Wilson Yan, Yunzhi Zhang, Pieter Abbeel, and Aravind Srinivas. Videogpt: Video generation using vq-vae and transformers. arXiv preprint arXiv:2104.10157, 2021. 3 +[76] Yunzhi Yan, Haotong Lin, Chenxu Zhou, Weijie Wang, Haiyang Sun, Kun Zhan, Xianpeng Lang, Xiaowei Zhou, and Sida Peng. Street gaussians for modeling dynamic urban scenes. arXiv preprint arXiv:2401.01339, 2024. 2 +[77] Jiawei Yang, Boris Ivanovic, Or Litany, Xinshuo Weng, Seung Wook Kim, Boyi Li, Tong Che, Danfei Xu, Sanja Fidler, Marco Pavone, et al. Emernerf: Emergent spatial-temporal scene decomposition via self-supervision. arXiv preprint arXiv:2311.02077, 2023. 2 +[78] Xuemeng Yang, Licheng Wen, Yukai Ma, Jianbiao Mei, Xin Li, Tiantian Wei, Wenjie Lei, Daocheng Fu, Pinlong Cai, Min Dou, Botian Shi, Liang He, Yong Liu, and Yu Qiao. Drivearena: A closed-loop generative simulation platform for autonomous driving. arXiv preprint arXiv:2408.00415, 2024.3 +[79] Ze Yang, Yun Chen, Jingkang Wang, Sivabalan Manivasagam, Wei-Chiu Ma, Anqi Joyce Yang, and Raquel Urtasun. Unisim: A neural closed-loop sensor simulator. In CVPR, 2023. 2, 6, 7, 8 +[80] Ziyi Yang, Xinyu Gao, Wen Zhou, Shaohui Jiao, Yuqing Zhang, and Xiaogang Jin. Deformable 3d gaussians for high-fidelity monocular dynamic scene reconstruction. In CVPR, 2024. 1, 2, 6, 7 +[81] Zhuoyi Yang, Jiayan Teng, Wendi Zheng, Ming Ding, Shiyu Huang, Jiazheng Xu, Yuanming Yang, Wenyi Hong, Xiaohan Zhang, Guanyu Feng, et al. Cogvideox: Text-to-video diffusion models with an expert transformer. arXiv preprint arXiv:2408.06072, 2024. 3 +[82] Wang Yifan, Felice Serena, Shihao Wu, Cengiz Öztireli, and Olga Sorkine-Hornung. Differentiable surface splatting for point-based geometry processing. ACM TOG, 2019. 3, 5 +[83] Zehao Yu, Anpei Chen, Binbin Huang, Torsten Sattler, and Andreas Geiger. Mip-splatting: Alias-free 3d gaussian splattering. In CVPR, 2024. 2 +[84] Zhongrui Yu, Haoran Wang, Jinze Yang, Hanzhang Wang, Zeke Xie, Yunfeng Cai, Jiale Cao, Zhong Ji, and Mingming Sun. Sgd: Street view synthesis with gaussian splatting and diffusion prior. arXiv preprint arXiv:2403.20079, 2024. 2, 3 +[85] Yan Zeng, Guoqiang Wei, Jiani Zheng, Jiaxin Zou, Yang Wei, Yuchen Zhang, and Hang Li. Make pixels + +dance: High-dynamic video generation. arXiv preprint arXiv:2311.10982, 2023. 3 +[86] Jiang-Tian Zhai, Ze Feng, Jinhao Du, Yongqiang Mao, Jiang-Jiang Liu, Zichang Tan, Yifu Zhang, Xiaoqing Ye, and Jingdong Wang. Rethinking the open-loop evaluation of end-to-end autonomous driving in nuscenes. arXiv preprint arXiv:2305.10430, 2023. 2 +[87] Guosheng Zhao, Xiaofeng Wang, Zheng Zhu, Xinze Chen, Guan Huang, Xiaoyi Bao, and Xingang Wang. Drivedreamer-2: Llm-enhanced world models for diverse driving video generation. arXiv preprint arXiv:2403.06845, 2024. 2, 3, 5 +[88] Hongyu Zhou, Jiahao Shao, Lu Xu, Dongfeng Bai, Weichao Qiu, Bingbing Liu, Yue Wang, Andreas Geiger, and Yiyi Liao. Hugs: Holistic urban 3d scene understanding via gaussian splatting. In CVPR, 2024. 2 +[89] Xiaoyu Zhou, Zhiwei Lin, Xiaojun Shan, Yongtao Wang, Deqing Sun, and Ming-Hsuan Yang. Drivinggaussian: Composite gaussian splatting for surrounding dynamic autonomous driving scenes. In CVPR, pages 21634-21643, 2024. 2 +[90] Zheng Zhu, Xiaofeng Wang, Wangbo Zhao, Chen Min, Ni-anchen Deng, Min Dou, Yuqi Wang, Botian Shi, Kai Wang, Chi Zhang, et al. Is sora a world simulator? a comprehensive survey on general world models and beyond. arXiv preprint arXiv:2405.03520, 2024. 3 \ No newline at end of file diff --git a/2025/DriveDreamer4D_ World Models Are Effective Data Machines for 4D Driving Scene Representation/images.zip b/2025/DriveDreamer4D_ World Models Are Effective Data Machines for 4D Driving Scene Representation/images.zip new file mode 100644 index 0000000000000000000000000000000000000000..e553064ad41745cf3423ada0d1b75f1bee9cba2c --- /dev/null +++ b/2025/DriveDreamer4D_ World Models Are Effective Data Machines for 4D Driving Scene Representation/images.zip @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:f830affd0171602653f7b87479a25cc387d8e1bcda4d68d4c4ff6ba5d6e38041 +size 447370 diff --git a/2025/DriveDreamer4D_ World Models Are Effective Data Machines for 4D Driving Scene Representation/layout.json b/2025/DriveDreamer4D_ World Models Are Effective Data Machines for 4D Driving Scene Representation/layout.json new file mode 100644 index 0000000000000000000000000000000000000000..924b9b83b0d67d5b48fda668ce54d75aa1cccedc --- /dev/null +++ b/2025/DriveDreamer4D_ World Models Are Effective Data Machines for 4D Driving Scene Representation/layout.json @@ -0,0 +1,10894 @@ +{ + "pdf_info": [ + { + "para_blocks": [ + { + "bbox": [ + 113, + 103, + 500, + 140 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 113, + 103, + 500, + 140 + ], + "spans": [ + { + "bbox": [ + 113, + 103, + 500, + 140 + ], + "type": "text", + "content": "DriveDreamer4D: World Models Are Effective Data Machines for 4D Driving Scene Representation" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 113, + 160, + 496, + 203 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 113, + 160, + 496, + 203 + ], + "spans": [ + { + "bbox": [ + 113, + 160, + 496, + 203 + ], + "type": "text", + "content": "Guosheng Zhao" + }, + { + "bbox": [ + 113, + 160, + 496, + 203 + ], + "type": "inline_equation", + "content": "^{1,2,3*}" + }, + { + "bbox": [ + 113, + 160, + 496, + 203 + ], + "type": "text", + "content": ", Chaojun Ni" + }, + { + "bbox": [ + 113, + 160, + 496, + 203 + ], + "type": "inline_equation", + "content": "^{4*}" + }, + { + "bbox": [ + 113, + 160, + 496, + 203 + ], + "type": "text", + "content": ", Xiaofeng Wang" + }, + { + "bbox": [ + 113, + 160, + 496, + 203 + ], + "type": "inline_equation", + "content": "^{1,2,3*}" + }, + { + "bbox": [ + 113, + 160, + 496, + 203 + ], + "type": "text", + "content": ", Zheng Zhu" + }, + { + "bbox": [ + 113, + 160, + 496, + 203 + ], + "type": "inline_equation", + "content": "^{5*†}" + }, + { + "bbox": [ + 113, + 160, + 496, + 203 + ], + "type": "text", + "content": ", Xueyang Zhang" + }, + { + "bbox": [ + 113, + 160, + 496, + 203 + ], + "type": "inline_equation", + "content": "^{6}" + }, + { + "bbox": [ + 113, + 160, + 496, + 203 + ], + "type": "text", + "content": ", Yida Wang" + }, + { + "bbox": [ + 113, + 160, + 496, + 203 + ], + "type": "inline_equation", + "content": "^{6}" + }, + { + "bbox": [ + 113, + 160, + 496, + 203 + ], + "type": "text", + "content": ", Guan Huang" + }, + { + "bbox": [ + 113, + 160, + 496, + 203 + ], + "type": "inline_equation", + "content": "^{5}" + }, + { + "bbox": [ + 113, + 160, + 496, + 203 + ], + "type": "text", + "content": ", Xinze Chen" + }, + { + "bbox": [ + 113, + 160, + 496, + 203 + ], + "type": "inline_equation", + "content": "^{5}" + }, + { + "bbox": [ + 113, + 160, + 496, + 203 + ], + "type": "text", + "content": ", Boyuan Wang" + }, + { + "bbox": [ + 113, + 160, + 496, + 203 + ], + "type": "inline_equation", + "content": "^{1,2,3}" + }, + { + "bbox": [ + 113, + 160, + 496, + 203 + ], + "type": "text", + "content": ", Youyi Zhang" + }, + { + "bbox": [ + 113, + 160, + 496, + 203 + ], + "type": "inline_equation", + "content": "^{7}" + }, + { + "bbox": [ + 113, + 160, + 496, + 203 + ], + "type": "text", + "content": ", Wenjun Mei" + }, + { + "bbox": [ + 113, + 160, + 496, + 203 + ], + "type": "inline_equation", + "content": "^{4}" + }, + { + "bbox": [ + 113, + 160, + 496, + 203 + ], + "type": "text", + "content": ", Xingang Wang" + }, + { + "bbox": [ + 113, + 160, + 496, + 203 + ], + "type": "inline_equation", + "content": "^{2,3†}" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 152, + 203, + 458, + 217 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 152, + 203, + 458, + 217 + ], + "spans": [ + { + "bbox": [ + 152, + 203, + 458, + 217 + ], + "type": "inline_equation", + "content": "^{1}" + }, + { + "bbox": [ + 152, + 203, + 458, + 217 + ], + "type": "text", + "content": "Institute of Automation, Chinese Academy of Sciences, China" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 100, + 217, + 511, + 231 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 100, + 217, + 511, + 231 + ], + "spans": [ + { + "bbox": [ + 100, + 217, + 511, + 231 + ], + "type": "inline_equation", + "content": "^{2}" + }, + { + "bbox": [ + 100, + 217, + 511, + 231 + ], + "type": "text", + "content": "School of Artificial Intelligence, University of Chinese Academy of Sciences, China" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 87, + 231, + 523, + 245 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 87, + 231, + 523, + 245 + ], + "spans": [ + { + "bbox": [ + 87, + 231, + 523, + 245 + ], + "type": "inline_equation", + "content": "^{3}" + }, + { + "bbox": [ + 87, + 231, + 523, + 245 + ], + "type": "text", + "content": "Luoyang Institute for Robot and Intelligent Equipment, China " + }, + { + "bbox": [ + 87, + 231, + 523, + 245 + ], + "type": "inline_equation", + "content": "^{4}" + }, + { + "bbox": [ + 87, + 231, + 523, + 245 + ], + "type": "text", + "content": "Peking University, China" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 100, + 245, + 504, + 259 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 100, + 245, + 504, + 259 + ], + "spans": [ + { + "bbox": [ + 100, + 245, + 504, + 259 + ], + "type": "inline_equation", + "content": "^{5}" + }, + { + "bbox": [ + 100, + 245, + 504, + 259 + ], + "type": "text", + "content": "GigaAI, China " + }, + { + "bbox": [ + 100, + 245, + 504, + 259 + ], + "type": "inline_equation", + "content": "^{6}" + }, + { + "bbox": [ + 100, + 245, + 504, + 259 + ], + "type": "text", + "content": "Li Auto Inc., China " + }, + { + "bbox": [ + 100, + 245, + 504, + 259 + ], + "type": "inline_equation", + "content": "^{7}" + }, + { + "bbox": [ + 100, + 245, + 504, + 259 + ], + "type": "text", + "content": "Technical University of Munich, Germany" + } + ] + } + ], + "index": 7 + }, + { + "type": "image", + "bbox": [ + 62, + 288, + 246, + 430 + ], + "blocks": [ + { + "bbox": [ + 62, + 288, + 246, + 430 + ], + "lines": [ + { + "bbox": [ + 62, + 288, + 246, + 430 + ], + "spans": [ + { + "bbox": [ + 62, + 288, + 246, + 430 + ], + "type": "image", + "image_path": "6ba133e2acea1579799ea73bcd998fdbe1c2c70ac15f65aa6c0bbf4c5a45d5ae.jpg" + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 54, + 439, + 555, + 473 + ], + "lines": [ + { + "bbox": [ + 54, + 439, + 555, + 473 + ], + "spans": [ + { + "bbox": [ + 54, + 439, + 555, + 473 + ], + "type": "text", + "content": "Figure 1. Previous 4D Gaussian Splatting methods (e.g., PVG [9], " + }, + { + "bbox": [ + 54, + 439, + 555, + 473 + ], + "type": "inline_equation", + "content": "\\mathbf{S}^3" + }, + { + "bbox": [ + 54, + 439, + 555, + 473 + ], + "type": "text", + "content": " Gaussian [31], Deformable-GS [80]) face challenges in rendering novel trajectories, such as lane change. DriveDreamer4D addresses this by enhancing 4D driving scene representation via integrating priors from world models, significantly improving rendering quality under complex scenarios and novel trajectory viewpoints." + } + ] + } + ], + "index": 15, + "angle": 0, + "type": "image_caption" + } + ], + "index": 8 + }, + { + "type": "image", + "bbox": [ + 248, + 288, + 348, + 356 + ], + "blocks": [ + { + "bbox": [ + 248, + 288, + 348, + 356 + ], + "lines": [ + { + "bbox": [ + 248, + 288, + 348, + 356 + ], + "spans": [ + { + "bbox": [ + 248, + 288, + 348, + 356 + ], + "type": "image", + "image_path": "8efd397d45a96f34d57b8bb9515a9a19b37fd1c7a3f3b5b3d99c2458b9a15631.jpg" + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "image_body" + } + ], + "index": 9 + }, + { + "type": "image", + "bbox": [ + 349, + 288, + 449, + 356 + ], + "blocks": [ + { + "bbox": [ + 349, + 288, + 449, + 356 + ], + "lines": [ + { + "bbox": [ + 349, + 288, + 449, + 356 + ], + "spans": [ + { + "bbox": [ + 349, + 288, + 449, + 356 + ], + "type": "image", + "image_path": "072b40ab0261fb08a51bd24b96659ba57628b3314c9abcc27781e29a8099789d.jpg" + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "image_body" + } + ], + "index": 10 + }, + { + "type": "image", + "bbox": [ + 450, + 288, + 551, + 356 + ], + "blocks": [ + { + "bbox": [ + 450, + 288, + 551, + 356 + ], + "lines": [ + { + "bbox": [ + 450, + 288, + 551, + 356 + ], + "spans": [ + { + "bbox": [ + 450, + 288, + 551, + 356 + ], + "type": "image", + "image_path": "4070e216b614821aff83b4a7c933acb565f4a799bb1d0c9791c887f96426c61c.jpg" + } + ] + } + ], + "index": 11, + "angle": 0, + "type": "image_body" + } + ], + "index": 11 + }, + { + "type": "image", + "bbox": [ + 247, + 361, + 348, + 430 + ], + "blocks": [ + { + "bbox": [ + 247, + 361, + 348, + 430 + ], + "lines": [ + { + "bbox": [ + 247, + 361, + 348, + 430 + ], + "spans": [ + { + "bbox": [ + 247, + 361, + 348, + 430 + ], + "type": "image", + "image_path": "454527b856f3366bf507c346142eaf349a1ce1a499fdd4bd8990f29746c8c851.jpg" + } + ] + } + ], + "index": 12, + "angle": 0, + "type": "image_body" + } + ], + "index": 12 + }, + { + "type": "image", + "bbox": [ + 350, + 361, + 449, + 430 + ], + "blocks": [ + { + "bbox": [ + 350, + 361, + 449, + 430 + ], + "lines": [ + { + "bbox": [ + 350, + 361, + 449, + 430 + ], + "spans": [ + { + "bbox": [ + 350, + 361, + 449, + 430 + ], + "type": "image", + "image_path": "ada2abcf22dd8589c7ee99ef00da41b9ebc46b78aea07b0c604ad4f4d8914c58.jpg" + } + ] + } + ], + "index": 13, + "angle": 0, + "type": "image_body" + } + ], + "index": 13 + }, + { + "type": "image", + "bbox": [ + 451, + 361, + 551, + 430 + ], + "blocks": [ + { + "bbox": [ + 451, + 361, + 551, + 430 + ], + "lines": [ + { + "bbox": [ + 451, + 361, + 551, + 430 + ], + "spans": [ + { + "bbox": [ + 451, + 361, + 551, + 430 + ], + "type": "image", + "image_path": "e315c8238599f15e77d9cfd012c952f9017e1f56d8b3a16e790c5bdeaa441aa4.jpg" + } + ] + } + ], + "index": 14, + "angle": 0, + "type": "image_body" + } + ], + "index": 14 + }, + { + "bbox": [ + 151, + 483, + 200, + 495 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 151, + 483, + 200, + 495 + ], + "spans": [ + { + "bbox": [ + 151, + 483, + 200, + 495 + ], + "type": "text", + "content": "Abstract" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 54, + 508, + 297, + 676 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 54, + 508, + 297, + 676 + ], + "spans": [ + { + "bbox": [ + 54, + 508, + 297, + 676 + ], + "type": "text", + "content": "Closed-loop simulation is essential for advancing end-to-end autonomous driving systems. Contemporary sensor simulation methods, such as NeRF and 3DGS, rely predominantly on conditions closely aligned with training data distributions, which are largely confined to forward-driving scenarios. Consequently, these methods face limitations when rendering complex maneuvers (e.g., lane change, acceleration, deceleration). Recent advancements in autonomous-driving world models have demonstrated the potential to generate diverse driving videos. However, these approaches remain constrained to 2D video generation, inherently lacking the spatiotemporal coherence required to capture intricacies of dynamic driving environments. In this paper, we introduce DriveDreamer4D, which" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 313, + 484, + 556, + 699 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 484, + 556, + 699 + ], + "spans": [ + { + "bbox": [ + 313, + 484, + 556, + 699 + ], + "type": "text", + "content": "enhances 4D driving scene representation leveraging world model priors. Specifically, we utilize the world model as a data machine to synthesize novel trajectory videos, where structured conditions are explicitly leveraged to control the spatial-temporal consistency of traffic elements. Besides, the cousin data training strategy is proposed to facilitate merging real and synthetic data for optimizing 4DGS. To our knowledge, DriveDreamer4D is the first to utilize video generation models for improving 4D reconstruction in driving scenarios. Experimental results reveal that DriveDreamer4D significantly enhances generation quality under novel trajectory views, achieving a relative improvement in FID by " + }, + { + "bbox": [ + 313, + 484, + 556, + 699 + ], + "type": "inline_equation", + "content": "32.1\\%" + }, + { + "bbox": [ + 313, + 484, + 556, + 699 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 313, + 484, + 556, + 699 + ], + "type": "inline_equation", + "content": "46.4\\%" + }, + { + "bbox": [ + 313, + 484, + 556, + 699 + ], + "type": "text", + "content": ", and " + }, + { + "bbox": [ + 313, + 484, + 556, + 699 + ], + "type": "inline_equation", + "content": "16.3\\%" + }, + { + "bbox": [ + 313, + 484, + 556, + 699 + ], + "type": "text", + "content": " compared to PVG, " + }, + { + "bbox": [ + 313, + 484, + 556, + 699 + ], + "type": "inline_equation", + "content": "S^3" + }, + { + "bbox": [ + 313, + 484, + 556, + 699 + ], + "type": "text", + "content": " Gaussian, and Deformable-GS. Moreover, DriveDreamer4D markedly enhances the spatiotemporal coherence of driving agents, which is verified by a comprehensive user study and the relative increases of " + }, + { + "bbox": [ + 313, + 484, + 556, + 699 + ], + "type": "inline_equation", + "content": "22.6\\%" + }, + { + "bbox": [ + 313, + 484, + 556, + 699 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 313, + 484, + 556, + 699 + ], + "type": "inline_equation", + "content": "43.5\\%" + }, + { + "bbox": [ + 313, + 484, + 556, + 699 + ], + "type": "text", + "content": ", and " + }, + { + "bbox": [ + 313, + 484, + 556, + 699 + ], + "type": "inline_equation", + "content": "15.6\\%" + }, + { + "bbox": [ + 313, + 484, + 556, + 699 + ], + "type": "text", + "content": " in the NTA-IoU metric." + } + ] + } + ], + "index": 18 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "spans": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "type": "text", + "content": "CVF" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 145, + 0, + 496, + 37 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 145, + 0, + 496, + 37 + ], + "spans": [ + { + "bbox": [ + 145, + 0, + 496, + 37 + ], + "type": "text", + "content": "This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 66, + 683, + 296, + 713 + ], + "type": "page_footnote", + "angle": 0, + "lines": [ + { + "bbox": [ + 66, + 683, + 296, + 713 + ], + "spans": [ + { + "bbox": [ + 66, + 683, + 296, + 713 + ], + "type": "text", + "content": "*These authors contributed equally to this work. \n†Corresponding authors. zhengzhu@ieee.org, xingang.wang@ia.ac.cn. \n‡Project Page: https://drivedreamer4d.github.io" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 294, + 748, + 318, + 758 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 318, + 758 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 318, + 758 + ], + "type": "text", + "content": "12015" + } + ] + } + ], + "index": 20 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 0 + }, + { + "para_blocks": [ + { + "bbox": [ + 56, + 71, + 136, + 84 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 71, + 136, + 84 + ], + "spans": [ + { + "bbox": [ + 56, + 71, + 136, + 84 + ], + "type": "text", + "content": "1. Introduction" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 55, + 91, + 294, + 223 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 91, + 294, + 223 + ], + "spans": [ + { + "bbox": [ + 55, + 91, + 294, + 223 + ], + "type": "text", + "content": "End-to-end planning [29, 30, 34], which directly maps sensor inputs to control signals, is among the most critical and promising tasks in autonomous driving. However, current open-loop evaluations are inadequate for accurately assessing end-to-end planning algorithms, highlighting an urgent need for enhanced evaluation methods [41, 43, 86]. A compelling solution lies in closed-loop evaluations within real-world scenarios, which require retrieving sensor data from arbitrarily specified viewpoints. This necessitates constructing a 4D driving scene representation capable of reconstructing complex, dynamic driving environments." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 55, + 223, + 295, + 486 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 223, + 295, + 486 + ], + "spans": [ + { + "bbox": [ + 55, + 223, + 295, + 486 + ], + "type": "text", + "content": "Closed-loop simulation in driving environments predominantly relies on scene reconstruction techniques such as Neural Radiance Fields (NeRF) [20, 49, 77, 79] and 3D Gaussian Splatting (3DGS) [11, 31, 36, 76], which are inherently limited by the density of input data. Specifically, these methods can render scenes effectively only under conditions closely aligned with their training data distributions—primarily forward-driving scenarios—and struggle to perform accurately during complex maneuvers (see Fig. 1). To mitigate these limitations, methods like SGD [84] and GGS [22] leverage generative models to extend the range of training viewpoints. However, these approaches primarily supplement sparse image data or static background elements, falling short of modeling the intricacies of dynamic, interactive driving scenes. Recently, advancements in autonomous driving world models [18, 28, 66, 68, 69, 87] have introduced the capability to generate diverse, command-aligned video viewpoints, offering renewed promise for closed-loop simulation in autonomous driving. Nonetheless, these models remain constrained to 2D videos, lacking the spatial-temporal coherence essential for accurately modeling complex driving scenarios." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 55, + 486, + 295, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 486, + 295, + 713 + ], + "spans": [ + { + "bbox": [ + 55, + 486, + 295, + 713 + ], + "type": "text", + "content": "In this paper, we introduce DriveDreamer4D, which improves 4D driving scene representation by integrating priors from autonomous driving world models. Our approach utilizes an autonomous driving world model [87] as a generative engine, synthesizing novel trajectory video data that densifies real-world driving datasets for enhanced training. Notably, we propose the Novel Trajectory Generation Module (NTGM) to generate diverse structured traffic conditions, and DriveDreamer4D applies these conditions to independently regulate the motion dynamics of foreground and background elements in complex driving environments. These conditions undergo view projection synchronized with vehicle maneuvers, ensuring that the synthesized data adheres to the spatiotemporal constraints. Subsequently, the Cousin Data Training Strategy (CDTS) is proposed to merge temporal-aligned real and synthetic data for training 4DGS. In CDTS, a regularization loss is further incorporated to ensure perceptual coherence. To the best of our knowledge, DriveDreamer4D is the first" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 313, + 72, + 553, + 251 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 72, + 553, + 251 + ], + "spans": [ + { + "bbox": [ + 313, + 72, + 553, + 251 + ], + "type": "text", + "content": "framework to harness video generation models for elevating 4D scene reconstruction quality in autonomous driving, providing richly varied viewpoint data for scenarios including lane change, acceleration, and deceleration. As shown in Fig. 1, experiment results demonstrate that DriveDreamer4D significantly enhances generation fidelity for novel trajectory viewpoints, achieving a relative improvement in FID by " + }, + { + "bbox": [ + 313, + 72, + 553, + 251 + ], + "type": "inline_equation", + "content": "32.1\\%" + }, + { + "bbox": [ + 313, + 72, + 553, + 251 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 313, + 72, + 553, + 251 + ], + "type": "inline_equation", + "content": "46.4\\%" + }, + { + "bbox": [ + 313, + 72, + 553, + 251 + ], + "type": "text", + "content": ", and " + }, + { + "bbox": [ + 313, + 72, + 553, + 251 + ], + "type": "inline_equation", + "content": "16.3\\%" + }, + { + "bbox": [ + 313, + 72, + 553, + 251 + ], + "type": "text", + "content": " compared to PVG [9], " + }, + { + "bbox": [ + 313, + 72, + 553, + 251 + ], + "type": "inline_equation", + "content": "\\mathrm{S}^3" + }, + { + "bbox": [ + 313, + 72, + 553, + 251 + ], + "type": "text", + "content": " Gaussian [31], and Deformable-GS [80]. Besides, DriveDreamer4D fortifies the spatiotemporal coherence between foreground and background elements, with respective increases of " + }, + { + "bbox": [ + 313, + 72, + 553, + 251 + ], + "type": "inline_equation", + "content": "22.6\\%" + }, + { + "bbox": [ + 313, + 72, + 553, + 251 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 313, + 72, + 553, + 251 + ], + "type": "inline_equation", + "content": "43.5\\%" + }, + { + "bbox": [ + 313, + 72, + 553, + 251 + ], + "type": "text", + "content": ", and " + }, + { + "bbox": [ + 313, + 72, + 553, + 251 + ], + "type": "inline_equation", + "content": "15.6\\%" + }, + { + "bbox": [ + 313, + 72, + 553, + 251 + ], + "type": "text", + "content": " in the NTA-IoU metric. Furthermore, a comprehensive user study confirms that the average win rate of DriveDreamer4D exceeds " + }, + { + "bbox": [ + 313, + 72, + 553, + 251 + ], + "type": "inline_equation", + "content": "80\\%" + }, + { + "bbox": [ + 313, + 72, + 553, + 251 + ], + "type": "text", + "content": ", compared to three baselines." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 313, + 252, + 553, + 421 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 252, + 553, + 421 + ], + "spans": [ + { + "bbox": [ + 313, + 252, + 553, + 421 + ], + "type": "text", + "content": "The primary contributions of this work are as follows: (1) We present DriveDreamer4D, the first framework to leverage world model priors for advancing 4D scene reconstruction in autonomous driving. (2) The NTGM is proposed to automate the generation of structured conditions, allowing DriveDreamer4D to create novel trajectory videos with complex maneuvers while ensuring spatial-temporal consistency. Additionally, the CDTS is introduced to merge temporal-aligned real and synthetic data for training 4DGS, using a regularization loss to maintain perceptual coherence. (3) We perform comprehensive experiments to validate that DriveDreamer4D notably enhances generation quality across novel trajectory viewpoints, as well as the spatiotemporal coherence of driving scene elements." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 313, + 434, + 400, + 447 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 434, + 400, + 447 + ], + "spans": [ + { + "bbox": [ + 313, + 434, + 400, + 447 + ], + "type": "text", + "content": "2. Related Work" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 313, + 456, + 476, + 468 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 456, + 476, + 468 + ], + "spans": [ + { + "bbox": [ + 313, + 456, + 476, + 468 + ], + "type": "text", + "content": "2.1. Driving Scene Representation" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 313, + 474, + 553, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 474, + 553, + 713 + ], + "spans": [ + { + "bbox": [ + 313, + 474, + 553, + 713 + ], + "type": "text", + "content": "NeRF and 3DGS have emerged as leading approaches for 3D scene representation. NeRF models [2, 3, 49, 50] continuous volumetric scenes using multi-layer perceptron (MLP) networks, enabling highly detailed scene reconstructions with remarkable rendering quality. More recently, 3DGS [13, 36, 83] introduces an innovative method by defining a set of anisotropic Gaussians in 3D space, leveraging adaptive density control to achieve high-quality renderings from sparse point cloud inputs. Several works have extended NeRF [12, 20, 23, 33, 47, 56, 62, 63, 77, 79] or 3DGS [9, 11, 14, 31, 32, 44, 76, 84, 88, 89] to autonomous driving scenarios. Given the dynamic nature of driving environments, there has also been significant effort in modeling 4D driving scene representations. Some approaches encode time as an additional input to parameterize 4D scenes [1, 15, 31, 42, 46, 52, 60], while others represent scenes as a composition of moving object models alongside a static background model [39, 51, 64, 71, 74, 79]. Despite these advancements, methods based on NeRF and 3DGS face limitations tied to the density of input data. These tech-" + } + ] + } + ], + "index": 8 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 749, + 318, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 749, + 318, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 749, + 318, + 757 + ], + "type": "text", + "content": "12016" + } + ] + } + ], + "index": 9 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 1 + }, + { + "para_blocks": [ + { + "bbox": [ + 55, + 72, + 295, + 109 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 72, + 295, + 109 + ], + "spans": [ + { + "bbox": [ + 55, + 72, + 295, + 109 + ], + "type": "text", + "content": "niques can only render scenes effectively when sensor data closely matches the training data distribution, which is typically confined to forward-driving scenarios." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 55, + 121, + 145, + 133 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 121, + 145, + 133 + ], + "spans": [ + { + "bbox": [ + 55, + 121, + 145, + 133 + ], + "type": "text", + "content": "2.2. World Models" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 55, + 140, + 296, + 380 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 140, + 296, + 380 + ], + "spans": [ + { + "bbox": [ + 55, + 140, + 296, + 380 + ], + "type": "text", + "content": "The world model module predicts possible future world states as a function of imagined action sequences proposed by the actor [40, 90]. Approaches such as [4, 5, 19, 21, 25-27, 38, 48, 67, 68, 72, 75, 81, 85] simulate environments through video generation controlled by free-text actions. At the forefront of this evolution is Sora [6], which leverages advanced generative techniques to produce intricate visual sequences that respect the fundamental laws of physics. This ability to deeply understand and simulate the environment not only improves video generation quality but also has substantial implications for real-world driving scenarios. Autonomous driving world models [18, 28, 66, 69, 78, 87] employ predictive methodologies to interpret driving environments, thereby generating realistic driving scenarios and learning key driving elements and policies from video data. Although these models successfully produce diverse driving video data conditioned on complex driving actions, they remain limited to 2D outputs and lack the spatial-temporal coherence needed to accurately capture the complexities of dynamic driving environments." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 55, + 392, + 256, + 406 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 392, + 256, + 406 + ], + "spans": [ + { + "bbox": [ + 55, + 392, + 256, + 406 + ], + "type": "text", + "content": "2.3. Diffusion Prior for 3D Representation" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 55, + 411, + 296, + 616 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 411, + 296, + 616 + ], + "spans": [ + { + "bbox": [ + 55, + 411, + 296, + 616 + ], + "type": "text", + "content": "Constructing comprehensive 3D scenes from limited observations demands generative prior, particularly for unseen areas. Earlier studies distill the knowledge from text-to-image diffusion models [53, 55, 57, 58] into a 3D representation model. Specifically, the Score Distillation Sampling (SDS) [45, 54, 70] is adopted to synthesize a 3D object from the text prompt. Furthermore, to enhance 3D consistency, several approaches extend the multi-view diffusion models [17, 59] and video diffusion models [4, 10, 65] to 3D scene generation. To extend the diffusion prior to complex, dynamic, large-scale driving scenes for 3D reconstruction, methods such as SGD [84], GGS [22] and MagicDrive3D [16] employ generative models to broaden the range of training viewpoints. Nonetheless, these approaches mainly address sparse image data or static background elements, lacking the capacity to fully capture the complexities inherent in the 4D driving environments." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 55, + 632, + 111, + 643 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 632, + 111, + 643 + ], + "spans": [ + { + "bbox": [ + 55, + 632, + 111, + 643 + ], + "type": "text", + "content": "3. Method" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 55, + 654, + 296, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 654, + 296, + 714 + ], + "spans": [ + { + "bbox": [ + 55, + 654, + 296, + 714 + ], + "type": "text", + "content": "In this section, we first elaborate on the preliminaries of 4D driving scene representation and world models for driving video generation. Then we present the details of DriveDreamer4D, which enhances 4D driving scene representation leveraging priors from driving world models." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 314, + 72, + 394, + 85 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 314, + 72, + 394, + 85 + ], + "spans": [ + { + "bbox": [ + 314, + 72, + 394, + 85 + ], + "type": "text", + "content": "3.1. Preliminary" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 313, + 89, + 484, + 102 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 89, + 484, + 102 + ], + "spans": [ + { + "bbox": [ + 313, + 89, + 484, + 102 + ], + "type": "text", + "content": "3.1.1. 4D Driving Scene Representation" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 313, + 104, + 553, + 175 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 104, + 553, + 175 + ], + "spans": [ + { + "bbox": [ + 313, + 104, + 553, + 175 + ], + "type": "text", + "content": "4DGS models the driving scene with a collection of 3DGS and a temporal field module. Each 3DGS [36] is parameterized by its center position " + }, + { + "bbox": [ + 313, + 104, + 553, + 175 + ], + "type": "inline_equation", + "content": "x" + }, + { + "bbox": [ + 313, + 104, + 553, + 175 + ], + "type": "text", + "content": ", opacity " + }, + { + "bbox": [ + 313, + 104, + 553, + 175 + ], + "type": "inline_equation", + "content": "\\gamma" + }, + { + "bbox": [ + 313, + 104, + 553, + 175 + ], + "type": "text", + "content": ", covariance " + }, + { + "bbox": [ + 313, + 104, + 553, + 175 + ], + "type": "inline_equation", + "content": "\\Sigma" + }, + { + "bbox": [ + 313, + 104, + 553, + 175 + ], + "type": "text", + "content": ", and view-dependent RGB color " + }, + { + "bbox": [ + 313, + 104, + 553, + 175 + ], + "type": "inline_equation", + "content": "c" + }, + { + "bbox": [ + 313, + 104, + 553, + 175 + ], + "type": "text", + "content": ", controlled via spherical harmonics. For stability, each covariance matrix " + }, + { + "bbox": [ + 313, + 104, + 553, + 175 + ], + "type": "inline_equation", + "content": "\\Sigma" + }, + { + "bbox": [ + 313, + 104, + 553, + 175 + ], + "type": "text", + "content": " is decomposed by:" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 398, + 175, + 553, + 189 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 398, + 175, + 553, + 189 + ], + "spans": [ + { + "bbox": [ + 398, + 175, + 553, + 189 + ], + "type": "interline_equation", + "content": "\\boldsymbol {\\Sigma} = \\boldsymbol {R} \\boldsymbol {S} \\boldsymbol {S} ^ {T} \\boldsymbol {R} ^ {T}, \\tag {1}", + "image_path": "9d8e9c0c14b2e75b319886aaa0ea1543d04a0bdaa31f4bfb07c449c4dcb2dcdc.jpg" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 313, + 194, + 555, + 291 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 194, + 555, + 291 + ], + "spans": [ + { + "bbox": [ + 313, + 194, + 555, + 291 + ], + "type": "text", + "content": "where scaling matrix " + }, + { + "bbox": [ + 313, + 194, + 555, + 291 + ], + "type": "inline_equation", + "content": "S" + }, + { + "bbox": [ + 313, + 194, + 555, + 291 + ], + "type": "text", + "content": " and a rotation matrix " + }, + { + "bbox": [ + 313, + 194, + 555, + 291 + ], + "type": "inline_equation", + "content": "\\pmb{R}" + }, + { + "bbox": [ + 313, + 194, + 555, + 291 + ], + "type": "text", + "content": " are learnable parameters, represented by scaling " + }, + { + "bbox": [ + 313, + 194, + 555, + 291 + ], + "type": "inline_equation", + "content": "\\pmb{s}" + }, + { + "bbox": [ + 313, + 194, + 555, + 291 + ], + "type": "text", + "content": " and quaternion " + }, + { + "bbox": [ + 313, + 194, + 555, + 291 + ], + "type": "inline_equation", + "content": "\\pmb{r}" + }, + { + "bbox": [ + 313, + 194, + 555, + 291 + ], + "type": "text", + "content": ". All trainable parameters of a single 3D Gaussian are collectively denoted as " + }, + { + "bbox": [ + 313, + 194, + 555, + 291 + ], + "type": "inline_equation", + "content": "\\phi = \\{\\pmb{x},\\pmb{\\gamma},\\pmb{s},\\pmb{r},\\pmb{c}\\}" + }, + { + "bbox": [ + 313, + 194, + 555, + 291 + ], + "type": "text", + "content": ". The temporal field " + }, + { + "bbox": [ + 313, + 194, + 555, + 291 + ], + "type": "inline_equation", + "content": "\\mathcal{F}" + }, + { + "bbox": [ + 313, + 194, + 555, + 291 + ], + "type": "text", + "content": " takes " + }, + { + "bbox": [ + 313, + 194, + 555, + 291 + ], + "type": "inline_equation", + "content": "\\phi" + }, + { + "bbox": [ + 313, + 194, + 555, + 291 + ], + "type": "text", + "content": " and a time step " + }, + { + "bbox": [ + 313, + 194, + 555, + 291 + ], + "type": "inline_equation", + "content": "t_{gs}" + }, + { + "bbox": [ + 313, + 194, + 555, + 291 + ], + "type": "text", + "content": " as input, outputting the offset " + }, + { + "bbox": [ + 313, + 194, + 555, + 291 + ], + "type": "inline_equation", + "content": "\\delta \\phi = \\{\\delta x,\\delta \\gamma ,\\delta s,\\delta r,\\delta c\\}" + }, + { + "bbox": [ + 313, + 194, + 555, + 291 + ], + "type": "text", + "content": " for each Gaussian relative to canonical space. The 4D Gaussian " + }, + { + "bbox": [ + 313, + 194, + 555, + 291 + ], + "type": "inline_equation", + "content": "\\phi^{\\prime} = \\{\\pmb{x}^{\\prime},\\pmb{\\gamma}^{\\prime},\\pmb{s}^{\\prime},\\pmb{r}^{\\prime},\\pmb{c}^{\\prime}\\}" + }, + { + "bbox": [ + 313, + 194, + 555, + 291 + ], + "type": "text", + "content": " is then computed by:" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 367, + 301, + 553, + 316 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 367, + 301, + 553, + 316 + ], + "spans": [ + { + "bbox": [ + 367, + 301, + 553, + 316 + ], + "type": "interline_equation", + "content": "\\phi^ {\\prime} = \\phi + \\delta \\phi = \\phi + \\mathcal {F} (\\phi , \\boldsymbol {t} _ {g s}). \\tag {2}", + "image_path": "24e295c634b412f9a5a3311c2ae582caac38d030e4815974caeee293956707b5.jpg" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 313, + 320, + 554, + 393 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 320, + 554, + 393 + ], + "spans": [ + { + "bbox": [ + 313, + 320, + 554, + 393 + ], + "type": "text", + "content": "Following [82], a differentiable Gaussian Splating renderer is employed to project 4D Gaussians " + }, + { + "bbox": [ + 313, + 320, + 554, + 393 + ], + "type": "inline_equation", + "content": "\\phi" + }, + { + "bbox": [ + 313, + 320, + 554, + 393 + ], + "type": "text", + "content": " into camera coordinates, yielding the covariance matrix " + }, + { + "bbox": [ + 313, + 320, + 554, + 393 + ], + "type": "inline_equation", + "content": "\\boldsymbol{\\Sigma}^{\\prime} = J\\boldsymbol{V}\\boldsymbol{\\Sigma}\\boldsymbol{V}^{T}\\boldsymbol{J}^{T}" + }, + { + "bbox": [ + 313, + 320, + 554, + 393 + ], + "type": "text", + "content": ", where " + }, + { + "bbox": [ + 313, + 320, + 554, + 393 + ], + "type": "inline_equation", + "content": "\\boldsymbol{J}" + }, + { + "bbox": [ + 313, + 320, + 554, + 393 + ], + "type": "text", + "content": " is the Jacobian matrix of the perspective projection, and " + }, + { + "bbox": [ + 313, + 320, + 554, + 393 + ], + "type": "inline_equation", + "content": "\\boldsymbol{V}" + }, + { + "bbox": [ + 313, + 320, + 554, + 393 + ], + "type": "text", + "content": " is the transform matrix. The color of each pixel is calculated by " + }, + { + "bbox": [ + 313, + 320, + 554, + 393 + ], + "type": "inline_equation", + "content": "N" + }, + { + "bbox": [ + 313, + 320, + 554, + 393 + ], + "type": "text", + "content": " ordered points using " + }, + { + "bbox": [ + 313, + 320, + 554, + 393 + ], + "type": "inline_equation", + "content": "\\alpha" + }, + { + "bbox": [ + 313, + 320, + 554, + 393 + ], + "type": "text", + "content": "-blending:" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 398, + 401, + 553, + 427 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 398, + 401, + 553, + 427 + ], + "spans": [ + { + "bbox": [ + 398, + 401, + 553, + 427 + ], + "type": "interline_equation", + "content": "C = \\sum_ {i \\in N} T _ {i} \\boldsymbol {c} _ {i} ^ {\\prime} \\alpha_ {i}, \\tag {3}", + "image_path": "77a02622da855c2264866ad2966c9471009bdf2f4e691c4161d59a71c8e0e27f.jpg" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 313, + 437, + 554, + 510 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 437, + 554, + 510 + ], + "spans": [ + { + "bbox": [ + 313, + 437, + 554, + 510 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 313, + 437, + 554, + 510 + ], + "type": "inline_equation", + "content": "T_{i}" + }, + { + "bbox": [ + 313, + 437, + 554, + 510 + ], + "type": "text", + "content": " is the transmittance defined by " + }, + { + "bbox": [ + 313, + 437, + 554, + 510 + ], + "type": "inline_equation", + "content": "\\prod_{j = 1}^{i - 1}(1 - \\alpha_j)" + }, + { + "bbox": [ + 313, + 437, + 554, + 510 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 313, + 437, + 554, + 510 + ], + "type": "inline_equation", + "content": "c_{i}^{\\prime}" + }, + { + "bbox": [ + 313, + 437, + 554, + 510 + ], + "type": "text", + "content": " denotes the color of each point, " + }, + { + "bbox": [ + 313, + 437, + 554, + 510 + ], + "type": "inline_equation", + "content": "\\alpha_{i}" + }, + { + "bbox": [ + 313, + 437, + 554, + 510 + ], + "type": "text", + "content": " is given by evaluating a 2D Gaussian with covariance " + }, + { + "bbox": [ + 313, + 437, + 554, + 510 + ], + "type": "inline_equation", + "content": "\\Sigma^{\\prime}" + }, + { + "bbox": [ + 313, + 437, + 554, + 510 + ], + "type": "text", + "content": " multiplied with a learned per-point opacity " + }, + { + "bbox": [ + 313, + 437, + 554, + 510 + ], + "type": "inline_equation", + "content": "\\gamma_{i}^{\\prime}" + }, + { + "bbox": [ + 313, + 437, + 554, + 510 + ], + "type": "text", + "content": ". The trainable parameters " + }, + { + "bbox": [ + 313, + 437, + 554, + 510 + ], + "type": "inline_equation", + "content": "\\phi^{\\prime}" + }, + { + "bbox": [ + 313, + 437, + 554, + 510 + ], + "type": "text", + "content": " can be optimized by a combination of RGB loss, depth loss and SSIM loss:" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 332, + 517, + 553, + 548 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 332, + 517, + 553, + 548 + ], + "spans": [ + { + "bbox": [ + 332, + 517, + 553, + 548 + ], + "type": "interline_equation", + "content": "\\begin{array}{l} \\mathcal {L} _ {\\mathrm {o r i}} \\left(\\phi^ {\\prime}\\right) = \\lambda_ {1} \\| \\hat {I} _ {\\mathrm {o r i}} - I _ {\\mathrm {o r i}} \\| _ {1} + \\lambda_ {2} \\| \\hat {D} _ {\\mathrm {o r i}} - D _ {\\mathrm {o r i}} \\| _ {1} \\tag {4} \\\\ + \\lambda_ {3} \\operatorname {S S I M} \\left(\\bar {I} _ {\\text {o r i}}, I _ {\\text {o r i}}\\right), \\\\ \\end{array}", + "image_path": "b7932feb5278b242d1b789f041a17cbaea1a9a364dc086993b8c96cefcaa55b1.jpg" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 313, + 559, + 554, + 621 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 559, + 554, + 621 + ], + "spans": [ + { + "bbox": [ + 313, + 559, + 554, + 621 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 313, + 559, + 554, + 621 + ], + "type": "inline_equation", + "content": "\\hat{I}_{\\mathrm{ori}}" + }, + { + "bbox": [ + 313, + 559, + 554, + 621 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 313, + 559, + 554, + 621 + ], + "type": "inline_equation", + "content": "I_{\\mathrm{ori}}" + }, + { + "bbox": [ + 313, + 559, + 554, + 621 + ], + "type": "text", + "content": " represent the rendered image and the ground truth image. " + }, + { + "bbox": [ + 313, + 559, + 554, + 621 + ], + "type": "inline_equation", + "content": "\\hat{D}_{\\mathrm{ori}}" + }, + { + "bbox": [ + 313, + 559, + 554, + 621 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 313, + 559, + 554, + 621 + ], + "type": "inline_equation", + "content": "D_{\\mathrm{ori}}" + }, + { + "bbox": [ + 313, + 559, + 554, + 621 + ], + "type": "text", + "content": " are the rendered depth and the ground truth LiDAR depth map. SSIM " + }, + { + "bbox": [ + 313, + 559, + 554, + 621 + ], + "type": "inline_equation", + "content": "(\\cdot)" + }, + { + "bbox": [ + 313, + 559, + 554, + 621 + ], + "type": "text", + "content": " refers to the operation of the Structural Similarity Index Measure, and " + }, + { + "bbox": [ + 313, + 559, + 554, + 621 + ], + "type": "inline_equation", + "content": "\\lambda_1,\\lambda_2,\\lambda_3" + }, + { + "bbox": [ + 313, + 559, + 554, + 621 + ], + "type": "text", + "content": " are the loss weights." + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 313, + 626, + 531, + 639 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 626, + 531, + 639 + ], + "spans": [ + { + "bbox": [ + 313, + 626, + 531, + 639 + ], + "type": "text", + "content": "3.1.2. World Models for Driving Video Generation" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 313, + 641, + 554, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 641, + 554, + 714 + ], + "spans": [ + { + "bbox": [ + 313, + 641, + 554, + 714 + ], + "type": "text", + "content": "The world model module predicts possible future world states based on imagined action sequences [40]. Autonomous-driving world models [18, 66, 69, 87], typically based on diffusion models, leverage structured driving information or action controls to guide future video prediction. During training, these models first encode" + } + ] + } + ], + "index": 19 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "text", + "content": "12017" + } + ] + } + ], + "index": 20 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 2 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 60, + 73, + 550, + 373 + ], + "blocks": [ + { + "bbox": [ + 60, + 73, + 550, + 373 + ], + "lines": [ + { + "bbox": [ + 60, + 73, + 550, + 373 + ], + "spans": [ + { + "bbox": [ + 60, + 73, + 550, + 373 + ], + "type": "image", + "image_path": "f77c6426aff52ffd7e45ad74d6da743b6570d299e99351b50ce9bf561b2edaf9.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 55, + 384, + 555, + 430 + ], + "lines": [ + { + "bbox": [ + 55, + 384, + 555, + 430 + ], + "spans": [ + { + "bbox": [ + 55, + 384, + 555, + 430 + ], + "type": "text", + "content": "Figure 2. The overall framework of DriveDreamer4D. Initially, by altering the actions of the original trajectory (e.g., steering angle, speed), new trajectories can be obtained. Conditioned on the first frame and the structured information (3D bounding boxes, HDMap) from the new trajectory, the novel trajectory videos are generated. Subsequently, the temporal-aligned cousin pair (original and novel trajectory videos) are merged to optimize the 4D Gaussian Splatting model, where a regularization loss is calculated to ensure perceptual coherence." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "bbox": [ + 55, + 450, + 295, + 498 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 450, + 295, + 498 + ], + "spans": [ + { + "bbox": [ + 55, + 450, + 295, + 498 + ], + "type": "text", + "content": "videos " + }, + { + "bbox": [ + 55, + 450, + 295, + 498 + ], + "type": "inline_equation", + "content": "\\pmb{v}" + }, + { + "bbox": [ + 55, + 450, + 295, + 498 + ], + "type": "text", + "content": " into a lower-dimensional latent space " + }, + { + "bbox": [ + 55, + 450, + 295, + 498 + ], + "type": "inline_equation", + "content": "z = \\mathcal{E}(\\pmb{v})" + }, + { + "bbox": [ + 55, + 450, + 295, + 498 + ], + "type": "text", + "content": " using a variational encoder " + }, + { + "bbox": [ + 55, + 450, + 295, + 498 + ], + "type": "inline_equation", + "content": "\\mathcal{E}" + }, + { + "bbox": [ + 55, + 450, + 295, + 498 + ], + "type": "text", + "content": ". After adding noise " + }, + { + "bbox": [ + 55, + 450, + 295, + 498 + ], + "type": "inline_equation", + "content": "\\epsilon_{t}" + }, + { + "bbox": [ + 55, + 450, + 295, + 498 + ], + "type": "text", + "content": " to the latent, the diffusion model learns a denoising process. This diffusion process is optimized by:" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 80, + 506, + 295, + 526 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 80, + 506, + 295, + 526 + ], + "spans": [ + { + "bbox": [ + 80, + 506, + 295, + 526 + ], + "type": "interline_equation", + "content": "\\mathcal {L} _ {d i f f} = \\mathbb {E} _ {\\boldsymbol {z}, \\epsilon \\sim \\mathcal {N} (0, 1), t} \\left[ \\| \\epsilon_ {t} - \\epsilon_ {\\theta} (\\boldsymbol {z} _ {t}, t, \\boldsymbol {f}) \\| _ {2} ^ {2} \\right], \\tag {5}", + "image_path": "bc227c87d01c33b46e72763f91fa75ccf4dedd10c9b6814dc006547bde72ec8f.jpg" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 55, + 533, + 295, + 665 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 533, + 295, + 665 + ], + "spans": [ + { + "bbox": [ + 55, + 533, + 295, + 665 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 55, + 533, + 295, + 665 + ], + "type": "inline_equation", + "content": "\\epsilon_{\\theta}" + }, + { + "bbox": [ + 55, + 533, + 295, + 665 + ], + "type": "text", + "content": " is a parameterized denoising network, " + }, + { + "bbox": [ + 55, + 533, + 295, + 665 + ], + "type": "inline_equation", + "content": "t" + }, + { + "bbox": [ + 55, + 533, + 295, + 665 + ], + "type": "text", + "content": " denotes the time step, representing the level of noise added or removed at each stage. Additionally, to improve the controllability of the generated data, conditional features " + }, + { + "bbox": [ + 55, + 533, + 295, + 665 + ], + "type": "inline_equation", + "content": "f" + }, + { + "bbox": [ + 55, + 533, + 295, + 665 + ], + "type": "text", + "content": " (e.g., reference images, speed, steering angle, scene layouts, camera poses and textual information) can be introduced into the reverse diffusion process, ensuring that the generated outputs adhere to the input control signals. During inference, the world models can be conditioned on a reference image to control the style of the output scene, while predicting the future world states contingent upon the other input actions." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 55, + 672, + 160, + 683 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 672, + 160, + 683 + ], + "spans": [ + { + "bbox": [ + 55, + 672, + 160, + 683 + ], + "type": "text", + "content": "3.2. DriveDreamer4D" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 55, + 689, + 295, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 689, + 295, + 715 + ], + "spans": [ + { + "bbox": [ + 55, + 689, + 295, + 715 + ], + "type": "text", + "content": "The overall pipeline of DriveDreamer4D is depicted in Fig. 2. In the upper part, the Novel Trajectory Generation" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 313, + 450, + 555, + 617 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 450, + 555, + 617 + ], + "spans": [ + { + "bbox": [ + 313, + 450, + 555, + 617 + ], + "type": "text", + "content": "Module (NTGM) is proposed to adjust driving actions (e.g., steering angle, speed) to generate new trajectories. These novel trajectories provide new perspectives for extracting structured information like 3D boxes and HDMap. Subsequently, a controllable video diffusion model synthesizes videos from these updated viewpoints, incorporating specific priors associated with the modified trajectories. In the lower part, the Cousin Data Training Strategy (CDTS) is introduced to combine the temporal-aligned original and generated data for optimizing the 4DGS model, where a regularization loss is calculated to impose perceptual coherence. In the following sections, we delve into the details of novel trajectories video generation and then introduce the CDTS for 4D reconstruction." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 313, + 625, + 493, + 639 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 625, + 493, + 639 + ], + "spans": [ + { + "bbox": [ + 313, + 625, + 493, + 639 + ], + "type": "text", + "content": "3.2.1. Novel Trajectory Video Generation" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 313, + 642, + 555, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 642, + 555, + 713 + ], + "spans": [ + { + "bbox": [ + 313, + 642, + 555, + 713 + ], + "type": "text", + "content": "As previously mentioned, traditional 4DGS methods are limited in rendering complex maneuvers, largely due to the training data being dominated by straightforward driving scenarios. To overcome this, DriveDreamer4D leverages world model priors to generate diverse viewpoint data, enhancing the 4D scene representation. To achieve this," + } + ] + } + ], + "index": 9 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "text", + "content": "12018" + } + ] + } + ], + "index": 10 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 3 + }, + { + "para_blocks": [ + { + "bbox": [ + 55, + 72, + 296, + 312 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 72, + 296, + 312 + ], + "spans": [ + { + "bbox": [ + 55, + 72, + 296, + 312 + ], + "type": "text", + "content": "we propose the NTGM, which is designed to create new trajectories that serve as input for the world model, enabling the automated generation of complex maneuver data. NTGM comprises two main components: (1) novel trajectory proposal, (2) trajectory safety assessment. In the novel trajectory proposal stage, text-to-trajectory [87] can be adopted to automatically generate diverse complex trajectories. Additionally, trajectories can be custom-designed to meet specific requirements, allowing for tailored data generation based on precise needs. The overview of the custom-designed trajectory proposal (e.g., lane change) and trajectory safety assessment is shown in the Algo. 1. In a specific driving scenario, the original trajectory in the world coordinate system can be readily acquired as " + }, + { + "bbox": [ + 55, + 72, + 296, + 312 + ], + "type": "inline_equation", + "content": "\\mathcal{T}_{\\mathrm{ori}}^{\\mathrm{world}} = \\{p_i^{\\mathrm{world}}\\}_{i=0}^{K}" + }, + { + "bbox": [ + 55, + 72, + 296, + 312 + ], + "type": "text", + "content": ", where " + }, + { + "bbox": [ + 55, + 72, + 296, + 312 + ], + "type": "inline_equation", + "content": "K" + }, + { + "bbox": [ + 55, + 72, + 296, + 312 + ], + "type": "text", + "content": " denotes the number of frames and " + }, + { + "bbox": [ + 55, + 72, + 296, + 312 + ], + "type": "inline_equation", + "content": "p_i^{\\mathrm{world}} \\in \\mathbb{R}^3" + }, + { + "bbox": [ + 55, + 72, + 296, + 312 + ], + "type": "text", + "content": " refers to the position of the ego-vehicle at the " + }, + { + "bbox": [ + 55, + 72, + 296, + 312 + ], + "type": "inline_equation", + "content": "i" + }, + { + "bbox": [ + 55, + 72, + 296, + 312 + ], + "type": "text", + "content": "-th frame. To propose novel trajectories, the original trajectory " + }, + { + "bbox": [ + 55, + 72, + 296, + 312 + ], + "type": "inline_equation", + "content": "\\mathcal{T}_{\\mathrm{ori}}^{\\mathrm{world}}" + }, + { + "bbox": [ + 55, + 72, + 296, + 312 + ], + "type": "text", + "content": " is transformed into the ego-vehicle coordinate system of the first frame, denoted as " + }, + { + "bbox": [ + 55, + 72, + 296, + 312 + ], + "type": "inline_equation", + "content": "\\mathcal{T}_{\\mathrm{ori}}^{\\mathrm{EgoStart}}" + }, + { + "bbox": [ + 55, + 72, + 296, + 312 + ], + "type": "text", + "content": " and computed as:" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 102, + 313, + 295, + 329 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 102, + 313, + 295, + 329 + ], + "spans": [ + { + "bbox": [ + 102, + 313, + 295, + 329 + ], + "type": "interline_equation", + "content": "\\left[ p _ {i} ^ {\\text {E g o S t a r t}}, 1 \\right] ^ {T} = M _ {0} ^ {- 1} \\times \\left[ p _ {i} ^ {\\text {w o r l d}}, 1 \\right] ^ {T}, \\tag {6}", + "image_path": "754d251bb05e998add66e3ac94c019b87a57fc76f525dddef045e097f21ac59f.jpg" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 55, + 335, + 296, + 506 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 335, + 296, + 506 + ], + "spans": [ + { + "bbox": [ + 55, + 335, + 296, + 506 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 55, + 335, + 296, + 506 + ], + "type": "inline_equation", + "content": "M_0 \\in \\mathbb{R}^{4 \\times 4}" + }, + { + "bbox": [ + 55, + 335, + 296, + 506 + ], + "type": "text", + "content": " represents the transformation matrix from the ego-vehicle coordinate system of the first frame to the world coordinate system, " + }, + { + "bbox": [ + 55, + 335, + 296, + 506 + ], + "type": "inline_equation", + "content": "[\\cdot]" + }, + { + "bbox": [ + 55, + 335, + 296, + 506 + ], + "type": "text", + "content": " denotes the operation of the concat. In the ego-vehicle coordinate system, the vehicle's heading is aligned with the positive " + }, + { + "bbox": [ + 55, + 335, + 296, + 506 + ], + "type": "inline_equation", + "content": "x" + }, + { + "bbox": [ + 55, + 335, + 296, + 506 + ], + "type": "text", + "content": "-axis, the " + }, + { + "bbox": [ + 55, + 335, + 296, + 506 + ], + "type": "inline_equation", + "content": "y" + }, + { + "bbox": [ + 55, + 335, + 296, + 506 + ], + "type": "text", + "content": "-axis points to the left side of the vehicle, and the " + }, + { + "bbox": [ + 55, + 335, + 296, + 506 + ], + "type": "inline_equation", + "content": "z" + }, + { + "bbox": [ + 55, + 335, + 296, + 506 + ], + "type": "text", + "content": "-axis is oriented vertically upwards, perpendicular to the plane of the vehicle. Consequently, changes in the vehicle's velocity and direction can be respectively represented by adjusting the value along the " + }, + { + "bbox": [ + 55, + 335, + 296, + 506 + ], + "type": "inline_equation", + "content": "x" + }, + { + "bbox": [ + 55, + 335, + 296, + 506 + ], + "type": "text", + "content": "-axis and " + }, + { + "bbox": [ + 55, + 335, + 296, + 506 + ], + "type": "inline_equation", + "content": "y" + }, + { + "bbox": [ + 55, + 335, + 296, + 506 + ], + "type": "text", + "content": "-axis. A final safety assessment is conducted for the newly generated trajectory points, which includes verifying whether the vehicle trajectories " + }, + { + "bbox": [ + 55, + 335, + 296, + 506 + ], + "type": "inline_equation", + "content": "p" + }, + { + "bbox": [ + 55, + 335, + 296, + 506 + ], + "type": "text", + "content": " remain within drivable areas " + }, + { + "bbox": [ + 55, + 335, + 296, + 506 + ], + "type": "inline_equation", + "content": "\\mathcal{B}_{\\mathrm{road}}" + }, + { + "bbox": [ + 55, + 335, + 296, + 506 + ], + "type": "text", + "content": " and ensuring that no collisions occur with pedestrians or other vehicles " + }, + { + "bbox": [ + 55, + 335, + 296, + 506 + ], + "type": "inline_equation", + "content": "\\{o_j\\}_{j=1}^M" + }, + { + "bbox": [ + 55, + 335, + 296, + 506 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 92, + 516, + 295, + 537 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 92, + 516, + 295, + 537 + ], + "spans": [ + { + "bbox": [ + 92, + 516, + 295, + 537 + ], + "type": "interline_equation", + "content": "p \\in \\mathcal {B} _ {\\text {r o a d}}, \\tag {7}", + "image_path": "13e79debc675d177eb822c3dccd19eee194c4caed2630d14f5037ac7f21c28ef.jpg" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 94, + 537, + 258, + 548 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 94, + 537, + 258, + 548 + ], + "spans": [ + { + "bbox": [ + 94, + 537, + 258, + 548 + ], + "type": "interline_equation", + "content": "\\| p - o _ {j} \\| \\geq d _ {\\min }, \\quad \\forall j \\in \\{1, \\dots , M \\},", + "image_path": "7836c07fcc2a76578c219ccc3405ede261b19dd068aa4c102b5c3d1abbbb2539.jpg" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 55, + 558, + 296, + 666 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 558, + 296, + 666 + ], + "spans": [ + { + "bbox": [ + 55, + 558, + 296, + 666 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 55, + 558, + 296, + 666 + ], + "type": "inline_equation", + "content": "d_{\\mathrm{min}}" + }, + { + "bbox": [ + 55, + 558, + 296, + 666 + ], + "type": "text", + "content": " is the minimal distance between different agents. Once a novel trajectory that complies with traffic regulations is generated, the road structure and 3D bounding boxes can be projected onto the camera view from the perspective of the new trajectory, thereby generating structured information relative to the updated trajectory. This structured information, along with the initial frame and text, is fed into a world model [87] to produce the videos that follow the novel trajectories." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 55, + 673, + 214, + 686 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 673, + 214, + 686 + ], + "spans": [ + { + "bbox": [ + 55, + 673, + 214, + 686 + ], + "type": "text", + "content": "3.2.2. Cousin Data Training Strategy" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 55, + 689, + 296, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 689, + 296, + 714 + ], + "spans": [ + { + "bbox": [ + 55, + 689, + 296, + 714 + ], + "type": "text", + "content": "To better integrate generated data for training 4DGS, we propose the CDTS. Specifically, we construct temporally" + } + ] + } + ], + "index": 7 + }, + { + "type": "code", + "bbox": [ + 313, + 87, + 547, + 318 + ], + "blocks": [ + { + "bbox": [ + 315, + 72, + 520, + 84 + ], + "lines": [ + { + "bbox": [ + 315, + 72, + 520, + 84 + ], + "spans": [ + { + "bbox": [ + 315, + 72, + 520, + 84 + ], + "type": "text", + "content": "Algorithm 1 Novel Trajectory Generation Module" + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "code_caption" + }, + { + "bbox": [ + 313, + 87, + 547, + 318 + ], + "lines": [ + { + "bbox": [ + 313, + 87, + 547, + 318 + ], + "spans": [ + { + "bbox": [ + 313, + 87, + 547, + 318 + ], + "type": "text", + "content": "Input: Trajectory " + }, + { + "bbox": [ + 313, + 87, + 547, + 318 + ], + "type": "inline_equation", + "content": "\\mathcal{T}_{\\mathrm{ori}}^{\\mathrm{world}}" + }, + { + "bbox": [ + 313, + 87, + 547, + 318 + ], + "type": "text", + "content": " , Transformation matrix " + }, + { + "bbox": [ + 313, + 87, + 547, + 318 + ], + "type": "inline_equation", + "content": "M_0" + }, + { + "bbox": [ + 313, + 87, + 547, + 318 + ], + "type": "text", + "content": " \nOutput: Novel trajectory " + }, + { + "bbox": [ + 313, + 87, + 547, + 318 + ], + "type": "inline_equation", + "content": "\\mathcal{T}_{\\mathrm{novel}}^{\\mathrm{ego}}" + }, + { + "bbox": [ + 313, + 87, + 547, + 318 + ], + "type": "inline_equation", + "content": "\\mathcal{T}_{\\mathrm{novel}}^{\\mathrm{ego}}\\gets [[0,0,0]]" + }, + { + "bbox": [ + 313, + 87, + 547, + 318 + ], + "type": "text", + "content": " \nOffset " + }, + { + "bbox": [ + 313, + 87, + 547, + 318 + ], + "type": "inline_equation", + "content": "\\leftarrow 0" + }, + { + "bbox": [ + 313, + 87, + 547, + 318 + ], + "type": "text", + "content": " \nfor each " + }, + { + "bbox": [ + 313, + 87, + 547, + 318 + ], + "type": "inline_equation", + "content": "p_{\\mathrm{ori}}^{\\mathrm{world}}" + }, + { + "bbox": [ + 313, + 87, + 547, + 318 + ], + "type": "text", + "content": " in " + }, + { + "bbox": [ + 313, + 87, + 547, + 318 + ], + "type": "inline_equation", + "content": "\\mathcal{T}_{\\mathrm{ori}}^{\\mathrm{world}}" + }, + { + "bbox": [ + 313, + 87, + 547, + 318 + ], + "type": "text", + "content": " [1:] do \n" + }, + { + "bbox": [ + 313, + 87, + 547, + 318 + ], + "type": "inline_equation", + "content": "p^{\\mathrm{EgoStart}}\\gets \\mathrm{RelativeCoord}(p_{\\mathrm{ori}}^{\\mathrm{world}},M_0)" + }, + { + "bbox": [ + 313, + 87, + 547, + 318 + ], + "type": "text", + "content": " \nMaxOffset " + }, + { + "bbox": [ + 313, + 87, + 547, + 318 + ], + "type": "inline_equation", + "content": "\\leftarrow 0.1" + }, + { + "bbox": [ + 313, + 87, + 547, + 318 + ], + "type": "text", + "content": " \nwhile True do \nNewOffset " + }, + { + "bbox": [ + 313, + 87, + 547, + 318 + ], + "type": "inline_equation", + "content": "\\leftarrow" + }, + { + "bbox": [ + 313, + 87, + 547, + 318 + ], + "type": "text", + "content": " Offset + RandOffset(0, MaxOffset) \n" + }, + { + "bbox": [ + 313, + 87, + 547, + 318 + ], + "type": "inline_equation", + "content": "p^{\\mathrm{EgoStart}'}\\gets p^{\\mathrm{EgoStart}} + [0," + }, + { + "bbox": [ + 313, + 87, + 547, + 318 + ], + "type": "text", + "content": " NewOffset, 0] \nif SafeCheck " + }, + { + "bbox": [ + 313, + 87, + 547, + 318 + ], + "type": "inline_equation", + "content": "(p^{\\mathrm{EgoStart}'}" + }, + { + "bbox": [ + 313, + 87, + 547, + 318 + ], + "type": "text", + "content": " ) then \nAddElement " + }, + { + "bbox": [ + 313, + 87, + 547, + 318 + ], + "type": "inline_equation", + "content": "(\\mathcal{T}_{\\mathrm{novel}}^{\\mathrm{ego}},p^{\\mathrm{EgoStart}'})" + }, + { + "bbox": [ + 313, + 87, + 547, + 318 + ], + "type": "text", + "content": " \nOffset " + }, + { + "bbox": [ + 313, + 87, + 547, + 318 + ], + "type": "inline_equation", + "content": "\\leftarrow" + }, + { + "bbox": [ + 313, + 87, + 547, + 318 + ], + "type": "text", + "content": " NewOffset \nbreak \nelse \nMaxOffset " + }, + { + "bbox": [ + 313, + 87, + 547, + 318 + ], + "type": "inline_equation", + "content": "\\leftarrow" + }, + { + "bbox": [ + 313, + 87, + 547, + 318 + ], + "type": "text", + "content": " MaxOffset/2 \nend if \nend while \nend for" + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "code_body" + } + ], + "index": 9, + "sub_type": "algorithm" + }, + { + "bbox": [ + 313, + 338, + 526, + 350 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 338, + 526, + 350 + ], + "spans": [ + { + "bbox": [ + 313, + 338, + 526, + 350 + ], + "type": "text", + "content": "aligned cousin pair data as a minimal training batch:" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 357, + 357, + 553, + 372 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 357, + 357, + 553, + 372 + ], + "spans": [ + { + "bbox": [ + 357, + 357, + 553, + 372 + ], + "type": "interline_equation", + "content": "\\operatorname {B a t c h S t a c k} \\left(\\left\\{\\hat {I} _ {\\text {o r i}, t} \\right\\} _ {t = 0} ^ {T}, \\left\\{\\hat {I} _ {\\text {n o v e l}, t} \\right\\} _ {t = 0} ^ {T}\\right), \\tag {8}", + "image_path": "ddee6c471ba5798765aa01d7d082a4f2c1301e4185ca79583ea8d97699420581.jpg" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 313, + 380, + 555, + 499 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 380, + 555, + 499 + ], + "spans": [ + { + "bbox": [ + 313, + 380, + 555, + 499 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 313, + 380, + 555, + 499 + ], + "type": "inline_equation", + "content": "\\mathrm{BatchStack}(\\cdot)" + }, + { + "bbox": [ + 313, + 380, + 555, + 499 + ], + "type": "text", + "content": " is the data processor to stack temporal-aligned " + }, + { + "bbox": [ + 313, + 380, + 555, + 499 + ], + "type": "inline_equation", + "content": "\\{\\hat{I}_{\\mathrm{ori},t}\\}_{t=0}^{T}" + }, + { + "bbox": [ + 313, + 380, + 555, + 499 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 313, + 380, + 555, + 499 + ], + "type": "inline_equation", + "content": "\\{\\hat{I}_{\\mathrm{novel},t}\\}_{t=0}^{T}" + }, + { + "bbox": [ + 313, + 380, + 555, + 499 + ], + "type": "text", + "content": " into a training batch. By leveraging real and synthetic data aligned at each timestep, CDTS mitigates the data gap in 4DGS training, enhancing the model's ability to learn consistent representations across real and synthetic data. To optimize 4DGS, the temporal-aligned cousin pair is input per step before gradient optimization. The loss function " + }, + { + "bbox": [ + 313, + 380, + 555, + 499 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{\\mathrm{ori}}" + }, + { + "bbox": [ + 313, + 380, + 555, + 499 + ], + "type": "text", + "content": " for the original data is defined in Eq. 4. And the loss function " + }, + { + "bbox": [ + 313, + 380, + 555, + 499 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{\\mathrm{novel}}" + }, + { + "bbox": [ + 313, + 380, + 555, + 499 + ], + "type": "text", + "content": " for the generated data is akin to [9, 31], defined as follows:" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 357, + 507, + 553, + 538 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 357, + 507, + 553, + 538 + ], + "spans": [ + { + "bbox": [ + 357, + 507, + 553, + 538 + ], + "type": "interline_equation", + "content": "\\begin{array}{l} \\mathcal {L} _ {\\text {n o v e l}} \\left(\\phi^ {\\prime}\\right) = \\lambda_ {1} \\| \\hat {I} _ {\\text {n o v e l}} - I _ {\\text {n o v e l}} \\| _ {1} \\tag {9} \\\\ + \\lambda_ {3} \\operatorname {S S I M} \\left(\\hat {I} _ {\\text {n o v e l}}, I _ {\\text {n o v e l}}\\right), \\\\ \\end{array}", + "image_path": "2c396f242f5963586752ab4a146f949c6d5004015701abf7eaa142944e3f5f40.jpg" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 313, + 546, + 555, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 546, + 555, + 713 + ], + "spans": [ + { + "bbox": [ + 313, + 546, + 555, + 713 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 313, + 546, + 555, + 713 + ], + "type": "inline_equation", + "content": "I_{\\mathrm{novel}}" + }, + { + "bbox": [ + 313, + 546, + 555, + 713 + ], + "type": "text", + "content": " represents the generated images corresponding to the novel trajectories as described in Sec. 3.2.1, and " + }, + { + "bbox": [ + 313, + 546, + 555, + 713 + ], + "type": "inline_equation", + "content": "\\hat{I}_{\\mathrm{novel}}" + }, + { + "bbox": [ + 313, + 546, + 555, + 713 + ], + "type": "text", + "content": " denotes the rendered images under the novel trajectories via differentiable splatting [82]. Notably, different from [9, 31], depth maps are not employed as constraints in the optimization of 4DGS when using the generated dataset " + }, + { + "bbox": [ + 313, + 546, + 555, + 713 + ], + "type": "inline_equation", + "content": "D_{\\mathrm{novel}}" + }, + { + "bbox": [ + 313, + 546, + 555, + 713 + ], + "type": "text", + "content": ". The limitation arises from the fact that LiDAR point cloud data is exclusively collected for the original trajectory. When these LiDAR points are projected onto a new trajectory, it cannot produce a complete depth map for the new perspective, as something visible in the novel trajectory may have been occluded in the original view. Consequently, the incorporation of such depth maps does not facilitate the optimization of the 4DGS model. More details are described in" + } + ] + } + ], + "index": 14 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "text", + "content": "12019" + } + ] + } + ], + "index": 15 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 4 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 57, + 70, + 555, + 170 + ], + "blocks": [ + { + "bbox": [ + 57, + 70, + 555, + 170 + ], + "lines": [ + { + "bbox": [ + 57, + 70, + 555, + 170 + ], + "spans": [ + { + "bbox": [ + 57, + 70, + 555, + 170 + ], + "type": "table", + "html": "
MethodLane ChangeAccelerationDecelerationAverage
NTA-IoU ↑NTL-IoU ↑NTA-IoU ↑NTL-IoU ↑NTA-IoU ↑NTL-IoU ↑NTA-IoU ↑NTL-IoU ↑
PVG [9]0.25650.700.39653.080.39453.650.34952.48
DriveDreamer4D with PVG0.43853.060.42153.350.42453.890.42853.43
S3Gaussian [31]0.17549.050.43451.930.38452.140.33151.04
DriveDreamer4D with S3Gaussian0.49553.420.48452.630.44552.690.47552.91
Deformable-GS [80]0.24051.620.34652.170.37753.210.32152.33
DriveDreamer4D with Deformable-GS0.33552.930.37152.770.40653.790.37153.16
", + "image_path": "a2a03a8b7a8df68773d7970544348c429163245e149a3bad1b92c4cc8b367fdc.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "table_body" + } + ], + "index": 0 + }, + { + "type": "table", + "bbox": [ + 56, + 208, + 293, + 309 + ], + "blocks": [ + { + "bbox": [ + 59, + 177, + 550, + 190 + ], + "lines": [ + { + "bbox": [ + 59, + 177, + 550, + 190 + ], + "spans": [ + { + "bbox": [ + 59, + 177, + 550, + 190 + ], + "type": "text", + "content": "Table 1. Comparison of NTA-IoU and NTL-IoU scores across different novel trajectory views (lane change, acceleration, deceleration)." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 56, + 208, + 293, + 309 + ], + "lines": [ + { + "bbox": [ + 56, + 208, + 293, + 309 + ], + "spans": [ + { + "bbox": [ + 56, + 208, + 293, + 309 + ], + "type": "table", + "html": "
MethodFID ↓
PVG[9]105.29
DriveDreamer4D with PVG71.52
S3Gaussian[31]124.90
DriveDreamer4D with S3Gaussian66.93
Deformable-GS[80]92.34
DriveDreamer4D with Deformable-GS77.32
", + "image_path": "bdd48b393301fe1bd23c7a1d711768df02977270c589031f80659982d2f59cad.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "table_body" + } + ], + "index": 2 + }, + { + "bbox": [ + 55, + 316, + 295, + 339 + ], + "lines": [ + { + "bbox": [ + 55, + 316, + 295, + 339 + ], + "spans": [ + { + "bbox": [ + 55, + 316, + 295, + 339 + ], + "type": "text", + "content": "Table 2. Comparison of FID scores in novel trajectory view synthesis (lane change) on the Waymo dataset." + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "text" + }, + { + "bbox": [ + 55, + 360, + 295, + 385 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 360, + 295, + 385 + ], + "spans": [ + { + "bbox": [ + 55, + 360, + 295, + 385 + ], + "type": "text", + "content": "Sec. 4.3. Additionally, we propose a regularization loss to enhance the perceptual coherence, defined as follows:" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 99, + 393, + 295, + 409 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 99, + 393, + 295, + 409 + ], + "spans": [ + { + "bbox": [ + 99, + 393, + 295, + 409 + ], + "type": "interline_equation", + "content": "\\mathcal {L} _ {\\text {r e g}} \\left(\\phi^ {\\prime}\\right) = \\left\\| \\mathcal {F} _ {p} \\left(\\hat {I} _ {\\text {o r i}}\\right) - \\mathcal {F} _ {p} \\left(\\hat {I} _ {\\text {n o v e l}}\\right) \\right\\| _ {1}, \\tag {10}", + "image_path": "564b59bef166396e6b7cb00367b16519c728f1eed7e04e207b10571f2783badb.jpg" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 55, + 418, + 296, + 453 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 418, + 296, + 453 + ], + "spans": [ + { + "bbox": [ + 55, + 418, + 296, + 453 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 55, + 418, + 296, + 453 + ], + "type": "inline_equation", + "content": "\\mathcal{F}_p(\\cdot)" + }, + { + "bbox": [ + 55, + 418, + 296, + 453 + ], + "type": "text", + "content": " denotes the perception feature extraction model [24]. The overall loss function for mixed training is defined as follows:" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 96, + 464, + 295, + 479 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 96, + 464, + 295, + 479 + ], + "spans": [ + { + "bbox": [ + 96, + 464, + 295, + 479 + ], + "type": "interline_equation", + "content": "\\mathcal {L} \\left(\\phi^ {\\prime}\\right) = \\mathcal {L} _ {\\text {o r i}} + \\lambda_ {\\text {n o v e l}} \\mathcal {L} _ {\\text {n o v e l}} + \\lambda_ {\\text {r e g}} \\mathcal {L} _ {\\text {r e g}}. \\tag {11}", + "image_path": "bfd45c532c570018b62ad46960ccae56d93d8ef1676b57896228c1ea83d9477d.jpg" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 55, + 498, + 137, + 513 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 498, + 137, + 513 + ], + "spans": [ + { + "bbox": [ + 55, + 498, + 137, + 513 + ], + "type": "text", + "content": "4. Experiments" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 54, + 519, + 295, + 651 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 54, + 519, + 295, + 651 + ], + "spans": [ + { + "bbox": [ + 54, + 519, + 295, + 651 + ], + "type": "text", + "content": "In this section, we first outline the experimental setup, including the dataset, implementation details, and evaluation metrics. Subsequently, quantitative results across different datasets are provided to demonstrate that the proposed DriveDreamer4D significantly enhances rendering quality for novel trajectory viewpoints and improves the spatiotemporal coherence of foreground and background components. Finally, we conduct an ablation study on the hyperparameter settings, as well as the effects of depth loss and the proposed CDTS, including temporal-aligned pairs, and regularization loss." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 55, + 659, + 163, + 673 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 659, + 163, + 673 + ], + "spans": [ + { + "bbox": [ + 55, + 659, + 163, + 673 + ], + "type": "text", + "content": "4.1. Experiment Setup" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 55, + 677, + 296, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 677, + 296, + 715 + ], + "spans": [ + { + "bbox": [ + 55, + 677, + 296, + 715 + ], + "type": "text", + "content": "Dataset. We conduct experiments using the Waymo dataset [61], known for its comprehensive real-world driving logs. We specifically select eight scenes characterized by highly" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 313, + 209, + 555, + 329 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 209, + 555, + 329 + ], + "spans": [ + { + "bbox": [ + 313, + 209, + 555, + 329 + ], + "type": "text", + "content": "dynamic interactions, featuring numerous vehicles with diverse relative positions and intricate driving trajectories. Each selected segment contains approximately 40 frames, with segment IDs detailed in the supplement. In addition, we have also conduct experiments on the nuscenes [7], and pandas [73] datasets. For nuScenes, we select eight scenes rich in dynamic objects and complex lane structures to test large maneuver capabilities. For PandaSet, we adopt the ten scenes used in Unisim [79] and NeuRAD [64] for fair comparison." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 313, + 330, + 556, + 628 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 330, + 556, + 628 + ], + "spans": [ + { + "bbox": [ + 313, + 330, + 556, + 628 + ], + "type": "text", + "content": "Implementation Details. To demonstrate the versatility and robustness of DriveDreamer4D, we incorporate various 4DGS baselines into our pipeline, including Deformable-GS [80], S" + }, + { + "bbox": [ + 313, + 330, + 556, + 628 + ], + "type": "inline_equation", + "content": "^3" + }, + { + "bbox": [ + 313, + 330, + 556, + 628 + ], + "type": "text", + "content": "Gaussian [31], and PVG [9]. For a fair comparison, LiDAR supervision is introduced to Deformable-GS. Scenes are segmented into 40-frame clips, matching the generative model's output length, using only forward-facing camera data at a resolution of " + }, + { + "bbox": [ + 313, + 330, + 556, + 628 + ], + "type": "inline_equation", + "content": "640 \\times 960" + }, + { + "bbox": [ + 313, + 330, + 556, + 628 + ], + "type": "text", + "content": ". Models are trained with the Adam optimizer [37], following the learning rate schedule from 3D Gaussian Splatting [36]. Hyperparameters align with each baseline [9, 31, 80], with training strategies unchanged except for incorporating CDTS. Moreover, to demonstrate the robustness of our method across different datasets, we employ PVG and DriveDreamer4D with PVG on nuScenes [7], and PandaSet [73]. Since existing methods typically conduct experiments on one or two datasets with varying scene selections, we have chosen different state-of-the-art (SOTA) methods for comparison on each dataset. This approach ensures a comprehensive evaluation of our method's performance under diverse conditions and highlights its adaptability and effectiveness across a wide range of scenarios. By benchmarking against various SOTA methods tailored to specific datasets, we provide strong evidence of the generalizability and robustness of DriveDreamer4D in handling complex and varied driving environments." + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 313, + 630, + 556, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 630, + 556, + 715 + ], + "spans": [ + { + "bbox": [ + 313, + 630, + 556, + 715 + ], + "type": "text", + "content": "Metrics. Traditional 3D reconstruction tasks typically employ PSNR and SSIM metrics for evaluation, with validation sets that closely match the training data distribution (i.e., uniformly sampling frames from video sequences for validation, with the remainder used for training). However, in closed-loop driving simulation, the focus shifts to evaluating model rendering performance under novel trajectories," + } + ] + } + ], + "index": 14 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 318, + 758 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 318, + 758 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 318, + 758 + ], + "type": "text", + "content": "12020" + } + ] + } + ], + "index": 15 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 5 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 58, + 70, + 295, + 124 + ], + "blocks": [ + { + "bbox": [ + 58, + 70, + 295, + 124 + ], + "lines": [ + { + "bbox": [ + 58, + 70, + 295, + 124 + ], + "spans": [ + { + "bbox": [ + 58, + 70, + 295, + 124 + ], + "type": "table", + "html": "
Counterpart MethodDriveDreamer4D Win Rate
Lane ChangeAccelerationDecelerationAverage
PVG [9]100.0%90.5%89.1%93.2%
S3Gaussian [31]100.0%97.9%92.2%96.7%
Deformable-GS [80]95.8%83.5%72.9%84.1%
", + "image_path": "6504ce337ab222e93d4439f569585040829750c6eac2669b3db7fb966a5c6005.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "table_body" + } + ], + "index": 0 + }, + { + "bbox": [ + 55, + 175, + 295, + 247 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 175, + 295, + 247 + ], + "spans": [ + { + "bbox": [ + 55, + 175, + 295, + 247 + ], + "type": "text", + "content": "where corresponding sensor data are unavailable, making metrics like PSNR and SSIM inapplicable for evaluation. Therefore, we propose Novel Trajectory Agent IoU (NTA-IoU) and Novel Trajectory Lane IoU (NTL-IoU), which assess the spatiotemporal coherence of foreground and background traffic components in novel trajectory viewpoints." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 54, + 247, + 295, + 403 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 54, + 247, + 295, + 403 + ], + "spans": [ + { + "bbox": [ + 54, + 247, + 295, + 403 + ], + "type": "text", + "content": "For NTA-IoU, we use YOLO11 [35] to identify vehicles in images rendered from novel trajectory views, yielding 2D bounding boxes. Simultaneously, geometric transformations are applied to the original 3D bounding boxes, projecting them onto the new viewpoints to generate corresponding 2D bounding boxes. For each projected 2D box, we then identify the closest detector-generated 2D box and compute their Intersection over Union (IoU). To ensure accurate matching, a distance threshold " + }, + { + "bbox": [ + 54, + 247, + 295, + 403 + ], + "type": "inline_equation", + "content": "d_{\\mathrm{thresh}}" + }, + { + "bbox": [ + 54, + 247, + 295, + 403 + ], + "type": "text", + "content": " is introduced: when the center-to-center distance " + }, + { + "bbox": [ + 54, + 247, + 295, + 403 + ], + "type": "inline_equation", + "content": "\\| c(B^{\\mathrm{proj}}) - c(B^{\\mathrm{det}})\\|" + }, + { + "bbox": [ + 54, + 247, + 295, + 403 + ], + "type": "text", + "content": " between the nearest detected box " + }, + { + "bbox": [ + 54, + 247, + 295, + 403 + ], + "type": "inline_equation", + "content": "B^{\\mathrm{det}}" + }, + { + "bbox": [ + 54, + 247, + 295, + 403 + ], + "type": "text", + "content": " and the correctly projected box " + }, + { + "bbox": [ + 54, + 247, + 295, + 403 + ], + "type": "inline_equation", + "content": "B^{\\mathrm{proj}}" + }, + { + "bbox": [ + 54, + 247, + 295, + 403 + ], + "type": "text", + "content": " surpasses this threshold, their NTA-IoU is assigned a value of zero:" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 58, + 412, + 295, + 445 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 58, + 412, + 295, + 445 + ], + "spans": [ + { + "bbox": [ + 58, + 412, + 295, + 445 + ], + "type": "interline_equation", + "content": "\\mathrm {N T A - I o U} = \\left\\{ \\begin{array}{l l} 0 & \\text {i f} \\| c (B ^ {\\operatorname {p r o j}}) - c (B ^ {\\det }) \\| \\geq d _ {\\text {t h r e s h}} \\\\ \\operatorname {I o U} \\left(B ^ {\\operatorname {p r o j}}, B ^ {\\det }\\right) & \\text {o t h e r w i s e .} \\end{array} \\right. \\tag {12}", + "image_path": "2d9681ca1a1110f736828a4aa9f9f8a177074e3ec6e69197b8cbc53c31410ccd.jpg" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 55, + 453, + 296, + 513 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 453, + 296, + 513 + ], + "spans": [ + { + "bbox": [ + 55, + 453, + 296, + 513 + ], + "type": "text", + "content": "For NTL-IoU, we employ TwinLiteNet [8] to extract 2D lanes from rendered images. Ground truth lanes are also projected onto the 2D image plane. We then compute the mean Intersection-over-Union (mIoU) between the rendered and ground truth lanes " + }, + { + "bbox": [ + 55, + 453, + 296, + 513 + ], + "type": "inline_equation", + "content": "L^{\\mathrm{det}}" + }, + { + "bbox": [ + 55, + 453, + 296, + 513 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 55, + 453, + 296, + 513 + ], + "type": "inline_equation", + "content": "L^{\\mathrm{proj}}" + }, + { + "bbox": [ + 55, + 453, + 296, + 513 + ], + "type": "text", + "content": ":" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 111, + 521, + 295, + 536 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 111, + 521, + 295, + 536 + ], + "spans": [ + { + "bbox": [ + 111, + 521, + 295, + 536 + ], + "type": "interline_equation", + "content": "\\mathrm {N T L - I o U} = \\mathrm {m I o U} \\left(L ^ {\\text {p r o j}}, L ^ {\\text {d e t}}\\right). \\tag {13}", + "image_path": "3b00bfa92583e774f7e5968c474bee3b44c227f79b045e3bc0750cdc1207dcf2.jpg" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 54, + 545, + 295, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 54, + 545, + 295, + 713 + ], + "spans": [ + { + "bbox": [ + 54, + 545, + 295, + 713 + ], + "type": "text", + "content": "Additionally, in lane change scenarios, we observe inaccuracies in relative positioning, as well as frequent occurrences of artifacts such as flying points and ghosting, which notably degrade image quality. To assess this, we employ the FID metric [24], which quantifies differences in feature distribution between rendered novel trajectory images and original trajectory images. This metric effectively reflects visual quality and is particularly sensitive to artifacts like flying points and ghosting, providing a robust measure of image fidelity in these complex scenes. Finally, a user study is conducted to evaluate the quality of the renderings, where participants compare the rendering results of each baseline with its DriveDreamer4D enhanced version across three novel trajectories. The evaluation criteria" + } + ] + } + ], + "index": 7 + }, + { + "type": "table", + "bbox": [ + 316, + 70, + 554, + 114 + ], + "blocks": [ + { + "bbox": [ + 55, + 131, + 295, + 153 + ], + "lines": [ + { + "bbox": [ + 55, + 131, + 295, + 153 + ], + "spans": [ + { + "bbox": [ + 55, + 131, + 295, + 153 + ], + "type": "text", + "content": "Table 3. User study comparison of DriveDreamer4D win rates across various novel trajectory view synthesis." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 316, + 70, + 554, + 114 + ], + "lines": [ + { + "bbox": [ + 316, + 70, + 554, + 114 + ], + "spans": [ + { + "bbox": [ + 316, + 70, + 554, + 114 + ], + "type": "table", + "html": "
MethodNTA-IoU ↑NTL-IoU ↑FID ↓
PVG[9]0.20448.75103.92
DriveDreamer4D with PVG0.33850.3688.60
", + "image_path": "9a085df4e87fb89850ccf0036681d802c616a5ce355d63656ba4f38ee226027d.jpg" + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "table_body" + } + ], + "index": 8 + }, + { + "type": "table", + "bbox": [ + 316, + 156, + 554, + 211 + ], + "blocks": [ + { + "bbox": [ + 313, + 121, + 555, + 144 + ], + "lines": [ + { + "bbox": [ + 313, + 121, + 555, + 144 + ], + "spans": [ + { + "bbox": [ + 313, + 121, + 555, + 144 + ], + "type": "text", + "content": "Table 4. Comparison of NTA-IoU, NTL-IoU and FID scores across lane change on nuScenes." + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 316, + 156, + 554, + 211 + ], + "lines": [ + { + "bbox": [ + 316, + 156, + 554, + 211 + ], + "spans": [ + { + "bbox": [ + 316, + 156, + 554, + 211 + ], + "type": "table", + "html": "
MethodNo shift ↓Lane 2m ↓Lane 3m ↓Vert. 1m ↓
UniSim [79]-74.797.5-
NeuRAD [64]25.072.393.976.3
PVG [9]24.767.376.882.9
DriveDreamer4D with PVG23.562.474.669.8
", + "image_path": "030faf9bf9ccbaa0555f5832bc00880ea18fe3023294eda3deff1f70ce52bae7.jpg" + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "table_body" + } + ], + "index": 10 + }, + { + "bbox": [ + 323, + 218, + 544, + 229 + ], + "lines": [ + { + "bbox": [ + 323, + 218, + 544, + 229 + ], + "spans": [ + { + "bbox": [ + 323, + 218, + 544, + 229 + ], + "type": "text", + "content": "Table 5. FID when shifting pose of ego vehicle on PandaSet." + } + ] + } + ], + "index": 11, + "angle": 0, + "type": "text" + }, + { + "bbox": [ + 313, + 252, + 555, + 301 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 252, + 555, + 301 + ], + "spans": [ + { + "bbox": [ + 313, + 252, + 555, + 301 + ], + "type": "text", + "content": "focus on overall video quality, with particular attention to foreground objects like vehicles. For each comparison, participants were asked to select the option they found most favorable. Further details are provided in the supplement." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 313, + 311, + 399, + 322 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 311, + 399, + 322 + ], + "spans": [ + { + "bbox": [ + 313, + 311, + 399, + 322 + ], + "type": "text", + "content": "4.2. Main Results" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 313, + 328, + 554, + 412 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 328, + 554, + 412 + ], + "spans": [ + { + "bbox": [ + 313, + 328, + 554, + 412 + ], + "type": "text", + "content": "In this subsection, we first demonstrate the effectiveness of DriveDreamer4D with various baseline methods. Then, through experimental results obtained from three different datasets, combined with detailed comparisons against a range of baselines, we highlight the robustness and adaptability of DriveDreamer4D in addressing a broad spectrum of scenarios and challenges." + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 313, + 413, + 554, + 557 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 413, + 554, + 557 + ], + "spans": [ + { + "bbox": [ + 313, + 413, + 554, + 557 + ], + "type": "text", + "content": "Waymo. As demonstrated in Tab. 1, integrating DriveDreamer4D with various 4DGS algorithms consistently yields superior NTA-IoU and NTL-IoU scores across complex maneuvers (e.g., lane changes, acceleration, and deceleration), surpassing baseline methods. Specifically, DriveDreamer4D enhances average NTA-IoU by " + }, + { + "bbox": [ + 313, + 413, + 554, + 557 + ], + "type": "inline_equation", + "content": "22.6\\%" + }, + { + "bbox": [ + 313, + 413, + 554, + 557 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 313, + 413, + 554, + 557 + ], + "type": "inline_equation", + "content": "43.5\\%" + }, + { + "bbox": [ + 313, + 413, + 554, + 557 + ], + "type": "text", + "content": ", and " + }, + { + "bbox": [ + 313, + 413, + 554, + 557 + ], + "type": "inline_equation", + "content": "15.6\\%" + }, + { + "bbox": [ + 313, + 413, + 554, + 557 + ], + "type": "text", + "content": " across three baselines (PVG [9], " + }, + { + "bbox": [ + 313, + 413, + 554, + 557 + ], + "type": "inline_equation", + "content": "\\mathrm{S}^3" + }, + { + "bbox": [ + 313, + 413, + 554, + 557 + ], + "type": "text", + "content": "Gaussian [31], Deformable-GS [80]), underscoring its capability to improve the spatiotemporal coherence of foreground agents. It also boosts average NTL-IoU by " + }, + { + "bbox": [ + 313, + 413, + 554, + 557 + ], + "type": "inline_equation", + "content": "1.8\\%" + }, + { + "bbox": [ + 313, + 413, + 554, + 557 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 313, + 413, + 554, + 557 + ], + "type": "inline_equation", + "content": "3.7\\%" + }, + { + "bbox": [ + 313, + 413, + 554, + 557 + ], + "type": "text", + "content": ", and " + }, + { + "bbox": [ + 313, + 413, + 554, + 557 + ], + "type": "inline_equation", + "content": "1.6\\%" + }, + { + "bbox": [ + 313, + 413, + 554, + 557 + ], + "type": "text", + "content": ", enhancing background lane consistency in 4D driving scene rendering." + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 313, + 558, + 556, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 558, + 556, + 715 + ], + "spans": [ + { + "bbox": [ + 313, + 558, + 556, + 715 + ], + "type": "text", + "content": "In addition to verifying the spatiotemporal consistency of rendered novel trajectory views, we leverage the FID metric to assess rendering quality under novel trajectories. Given that acceleration and deceleration scenarios yield rendered views with distributional similarities to ground truth, limiting FID's discriminative capability across algorithms, our FID comparisons focus specifically on lane change scenarios. Experiment results, as presented in Tab. 2, indicate that our method substantially outperforms the baseline methods (PVG [9], " + }, + { + "bbox": [ + 313, + 558, + 556, + 715 + ], + "type": "inline_equation", + "content": "\\mathrm{S}^3" + }, + { + "bbox": [ + 313, + 558, + 556, + 715 + ], + "type": "text", + "content": " Gaussian [31], Deformable-GS [80]), with FID relative improvements of " + }, + { + "bbox": [ + 313, + 558, + 556, + 715 + ], + "type": "inline_equation", + "content": "32.1\\%" + }, + { + "bbox": [ + 313, + 558, + 556, + 715 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 313, + 558, + 556, + 715 + ], + "type": "inline_equation", + "content": "46.4\\%" + }, + { + "bbox": [ + 313, + 558, + 556, + 715 + ], + "type": "text", + "content": ", and " + }, + { + "bbox": [ + 313, + 558, + 556, + 715 + ], + "type": "inline_equation", + "content": "16.3\\%" + }, + { + "bbox": [ + 313, + 558, + 556, + 715 + ], + "type": "text", + "content": ". These results highlight DriveDreamer4D's capability to enhance generation quality for" + } + ] + } + ], + "index": 16 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "text", + "content": "12021" + } + ] + } + ], + "index": 17 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 6 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 59, + 70, + 175, + 155 + ], + "blocks": [ + { + "bbox": [ + 59, + 70, + 175, + 155 + ], + "lines": [ + { + "bbox": [ + 59, + 70, + 175, + 155 + ], + "spans": [ + { + "bbox": [ + 59, + 70, + 175, + 155 + ], + "type": "table", + "html": "
λnovelNTA-IoU ↑FID ↓
00.349105.29
0.50.40582.84
10.42079.54
1.50.41782.10
", + "image_path": "c2b8488fa3a72f6558adda5f489b58b680bf61254922673c9450097e4b7ed026.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "table_body" + } + ], + "index": 0 + }, + { + "type": "table", + "bbox": [ + 182, + 70, + 293, + 156 + ], + "blocks": [ + { + "bbox": [ + 57, + 163, + 175, + 197 + ], + "lines": [ + { + "bbox": [ + 57, + 163, + 175, + 197 + ], + "spans": [ + { + "bbox": [ + 57, + 163, + 175, + 197 + ], + "type": "text", + "content": "Table 6. Ablation study on the training loss weight " + }, + { + "bbox": [ + 57, + 163, + 175, + 197 + ], + "type": "inline_equation", + "content": "\\lambda_{\\mathrm{novel}}" + }, + { + "bbox": [ + 57, + 163, + 175, + 197 + ], + "type": "text", + "content": " for novel trajectory data." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 182, + 70, + 293, + 156 + ], + "lines": [ + { + "bbox": [ + 182, + 70, + 293, + 156 + ], + "spans": [ + { + "bbox": [ + 182, + 70, + 293, + 156 + ], + "type": "table", + "html": "
λregNTA-IoU ↑FID ↓
00.42079.54
1e-20.411119.39
1e-30.42871.52
1e-40.42275.31
", + "image_path": "972f2da01b3ed0d96993bb4bef9cb9c903a7133bcf89824ccb6c1e5ea6053e50.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "table_body" + } + ], + "index": 2 + }, + { + "bbox": [ + 55, + 220, + 169, + 232 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 220, + 169, + 232 + ], + "spans": [ + { + "bbox": [ + 55, + 220, + 169, + 232 + ], + "type": "text", + "content": "novel trajectory viewpoints." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 54, + 233, + 295, + 353 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 54, + 233, + 295, + 353 + ], + "spans": [ + { + "bbox": [ + 54, + 233, + 295, + 353 + ], + "type": "text", + "content": "Finally, we conduct a user study to evaluate the rendering quality of different methods on novel trajectories, with a specific focus on foreground agents. For each method, we generate three novel trajectory views—lane change, acceleration, and deceleration—across eight scenes from the Waymo dataset [61]. Participants are then asked to select the renderings they found most visually favorable in each comparison. The DriveDreamer4D win rates from this study, shown in Tab. 3, reveal a significant user preference for our method's renderings." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 56, + 354, + 295, + 568 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 354, + 295, + 568 + ], + "spans": [ + { + "bbox": [ + 56, + 354, + 295, + 568 + ], + "type": "text", + "content": "nuScenes. Tab. 4 provides a comprehensive performance comparison between DriveDreamer4D and PVG [9] in lane change scenarios on the nuScenes dataset [7]. The experimental results offer compelling evidence that our method achieves a substantial improvement over the baseline, showcasing its superior capability in handling complex driving situations. Specifically, as detailed in the table, DriveDreamer4D consistently outperforms PVG across multiple evaluation metrics, including NTA-IoU, NTL-IoU, and FID. Notably, our method achieves a " + }, + { + "bbox": [ + 56, + 354, + 295, + 568 + ], + "type": "inline_equation", + "content": "65.7\\%" + }, + { + "bbox": [ + 56, + 354, + 295, + 568 + ], + "type": "text", + "content": " increase in NTA-IoU, a " + }, + { + "bbox": [ + 56, + 354, + 295, + 568 + ], + "type": "inline_equation", + "content": "3.3\\%" + }, + { + "bbox": [ + 56, + 354, + 295, + 568 + ], + "type": "text", + "content": " increase in NTL-IoU, and a " + }, + { + "bbox": [ + 56, + 354, + 295, + 568 + ], + "type": "inline_equation", + "content": "14.7\\%" + }, + { + "bbox": [ + 56, + 354, + 295, + 568 + ], + "type": "text", + "content": " reduction in FID. These results underscore the robustness of DriveDreamer4D in addressing challenging driving scenarios, highlighting its exceptional performance and adaptability across diverse evaluation criteria. Such advancements reaffirm the effectiveness of our approach in delivering more accurate and reliable trajectory predictions for dynamic environments." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 55, + 570, + 295, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 570, + 295, + 714 + ], + "spans": [ + { + "bbox": [ + 55, + 570, + 295, + 714 + ], + "type": "text", + "content": "PandaSet. Following the experimental setup of NeuRAD [64] and UniSim [79], we conduct experiments on the PandaSet dataset [73]. As shown in Tab. 5, we report the FID comparison under the original trajectory, a lane shift of 2 meters, a lane shift of 3 meters, and a perpendicular shift of 1 meter. These results demonstrate that our DriveDreamer4D achieves the best performance across all settings. Notably, DriveDreamer4D achieves an FID of 74.6 under the challenging scenario of a 3-meter lane shift, representing a significant improvement of " + }, + { + "bbox": [ + 55, + 570, + 295, + 714 + ], + "type": "inline_equation", + "content": "20.6\\%" + }, + { + "bbox": [ + 55, + 570, + 295, + 714 + ], + "type": "text", + "content": " over the previous SOTA method, NeuRAD. This highlights the strong performance and competitiveness of our approach com" + } + ] + } + ], + "index": 7 + }, + { + "type": "table", + "bbox": [ + 317, + 70, + 554, + 141 + ], + "blocks": [ + { + "bbox": [ + 182, + 163, + 293, + 186 + ], + "lines": [ + { + "bbox": [ + 182, + 163, + 293, + 186 + ], + "spans": [ + { + "bbox": [ + 182, + 163, + 293, + 186 + ], + "type": "text", + "content": "Table 7. Ablation study on the regularization loss weight " + }, + { + "bbox": [ + 182, + 163, + 293, + 186 + ], + "type": "inline_equation", + "content": "{\\lambda }_{\\text{reg }}" + }, + { + "bbox": [ + 182, + 163, + 293, + 186 + ], + "type": "text", + "content": " ." + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 317, + 70, + 554, + 141 + ], + "lines": [ + { + "bbox": [ + 317, + 70, + 554, + 141 + ], + "spans": [ + { + "bbox": [ + 317, + 70, + 554, + 141 + ], + "type": "table", + "html": "
Depth LossTemporal-aligned Cousin PairRegularization LossNTA-IoU ↑FID ↓
✓××0.40182.63
×××0.42079.54
×✓×0.42376.20
×✓✓0.42871.52
", + "image_path": "53c1a2ffc6d1651174125868c09b3bda3e1ba91e8f1ae21b40fbdb96834bade7.jpg" + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "table_body" + } + ], + "index": 8 + }, + { + "bbox": [ + 337, + 149, + 529, + 160 + ], + "lines": [ + { + "bbox": [ + 337, + 149, + 529, + 160 + ], + "spans": [ + { + "bbox": [ + 337, + 149, + 529, + 160 + ], + "type": "text", + "content": "Table 8. Ablation study on the depth loss and CDTS." + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "text" + }, + { + "bbox": [ + 313, + 186, + 554, + 221 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 186, + 554, + 221 + ], + "spans": [ + { + "bbox": [ + 313, + 186, + 554, + 221 + ], + "type": "text", + "content": "pared to existing techniques, reaffirming its effectiveness in handling complex driving scenarios with high precision and robustness." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 313, + 236, + 414, + 247 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 236, + 414, + 247 + ], + "spans": [ + { + "bbox": [ + 313, + 236, + 414, + 247 + ], + "type": "text", + "content": "4.3. Ablation Studies" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 313, + 255, + 555, + 434 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 255, + 555, + 434 + ], + "spans": [ + { + "bbox": [ + 313, + 255, + 555, + 434 + ], + "type": "text", + "content": "We conduct an ablation study based on PVG [9], analyzing the effects of hyperparameters settings for " + }, + { + "bbox": [ + 313, + 255, + 555, + 434 + ], + "type": "inline_equation", + "content": "\\lambda_{\\mathrm{novel}}" + }, + { + "bbox": [ + 313, + 255, + 555, + 434 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 313, + 255, + 555, + 434 + ], + "type": "inline_equation", + "content": "\\lambda_{\\mathrm{reg}}" + }, + { + "bbox": [ + 313, + 255, + 555, + 434 + ], + "type": "text", + "content": ", as well as the impact of depth loss, temporal-aligned cousin pairs, and regularization loss. Following the experiments shown in Tab. 6 and 7, we set " + }, + { + "bbox": [ + 313, + 255, + 555, + 434 + ], + "type": "inline_equation", + "content": "\\lambda_{\\mathrm{novel}}" + }, + { + "bbox": [ + 313, + 255, + 555, + 434 + ], + "type": "text", + "content": " to 1 and " + }, + { + "bbox": [ + 313, + 255, + 555, + 434 + ], + "type": "inline_equation", + "content": "\\lambda_{\\mathrm{reg}}" + }, + { + "bbox": [ + 313, + 255, + 555, + 434 + ], + "type": "text", + "content": " to " + }, + { + "bbox": [ + 313, + 255, + 555, + 434 + ], + "type": "inline_equation", + "content": "1\\times 10^{-3}" + }, + { + "bbox": [ + 313, + 255, + 555, + 434 + ], + "type": "text", + "content": " which yield the best performance. As shown in Tab. 8, the experiment confirms that the depth loss should be excluded when optimizing novel trajectory views, since LiDAR depth maps are incomplete due to occlusions. Furthermore, employing temporal-aligned cousin pairs achieves an FID of 76.20 and an NTA-IoU of 0.423, with the perception regularization loss further improving the FID to 71.52 and NTA-IoU to 0.428. These results highlight the effectiveness of CDTS, in achieving " + }, + { + "bbox": [ + 313, + 255, + 555, + 434 + ], + "type": "inline_equation", + "content": "\\sim 10\\%" + }, + { + "bbox": [ + 313, + 255, + 555, + 434 + ], + "type": "text", + "content": " improvement in FID, along with a " + }, + { + "bbox": [ + 313, + 255, + 555, + 434 + ], + "type": "inline_equation", + "content": "\\sim 2\\%" + }, + { + "bbox": [ + 313, + 255, + 555, + 434 + ], + "type": "text", + "content": " increase in NTA-IoU." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 313, + 452, + 466, + 464 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 452, + 466, + 464 + ], + "spans": [ + { + "bbox": [ + 313, + 452, + 466, + 464 + ], + "type": "text", + "content": "5. Discussion and Conclusion" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 313, + 474, + 555, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 474, + 555, + 713 + ], + "spans": [ + { + "bbox": [ + 313, + 474, + 555, + 713 + ], + "type": "text", + "content": "In this paper, we present DriveDreamer4D, a novel framework designed to advance 4D driving scene representations by harnessing priors from world models. Addressing key limitations of current sensor simulation methods—namely, their dependence on forward-driving training data distributions and inability to model complex maneuvers—DriveDreamer4D leverages a world model to generate novel trajectory videos that complement real-world driving data. By explicitly employing structured conditions, our framework maintains spatial-temporal consistency across traffic elements, ensuring that generated data adheres closely to the dynamics of real-world traffic scenarios. Our experiments demonstrate that DriveDreamer4D achieves superior quality in generating diverse simulation viewpoints, with significant improvements in both the rendering fidelity and spatiotemporal coherence of scene components. Notably, these results highlight DriveDreamer4D's potential as a foundation for closed-loop simulations that require high-fidelity reconstructions of dynamic driving scenes." + } + ] + } + ], + "index": 14 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "text", + "content": "12022" + } + ] + } + ], + "index": 15 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 7 + }, + { + "para_blocks": [ + { + "bbox": [ + 56, + 71, + 115, + 83 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 71, + 115, + 83 + ], + "spans": [ + { + "bbox": [ + 56, + 71, + 115, + 83 + ], + "type": "text", + "content": "References" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 57, + 91, + 296, + 713 + ], + "type": "list", + "angle": 0, + "index": 15, + "blocks": [ + { + "bbox": [ + 61, + 91, + 296, + 135 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 61, + 91, + 296, + 135 + ], + "spans": [ + { + "bbox": [ + 61, + 91, + 296, + 135 + ], + "type": "text", + "content": "[1] Benjamin Attal, Jia-Bin Huang, Christian Richardt, Michael Zollhoefer, Johannes Kopf, Matthew O'Toole, and Changil Kim. Hyperreel: High-fidelity 6-dof video with ray-conditioned sampling. In CVPR, 2023. 2" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 61, + 136, + 296, + 167 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 61, + 136, + 296, + 167 + ], + "spans": [ + { + "bbox": [ + 61, + 136, + 296, + 167 + ], + "type": "text", + "content": "[2] Jonathan T Barron, Ben Mildenhall, Dor Verbin, Pratul P Srinivasan, and Peter Hedman. Mip-nerf 360: Unbounded anti-aliased neural radiance fields. In CVPR, 2022. 2" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 62, + 168, + 295, + 201 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 168, + 295, + 201 + ], + "spans": [ + { + "bbox": [ + 62, + 168, + 295, + 201 + ], + "type": "text", + "content": "[3] Jonathan T Barron, Ben Mildenhall, Dor Verbin, Pratul P Srinivasan, and Peter Hedman. Zip-nerf: Anti-aliased grid-based neural radiance fields. In ICCV, 2023. 2" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 62, + 202, + 295, + 257 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 202, + 295, + 257 + ], + "spans": [ + { + "bbox": [ + 62, + 202, + 295, + 257 + ], + "type": "text", + "content": "[4] Andreas Blattmann, Tim Dockhorn, Sumith Kulal, Daniel Mendelevitch, Maciej Kilian, Dominik Lorenz, Yam Levi, Zion English, Vikram Voleti, Adam Letts, et al. Stable video diffusion: Scaling latent video diffusion models to large datasets. arXiv preprint arXiv:2311.15127, 2023. 3" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 62, + 258, + 295, + 301 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 258, + 295, + 301 + ], + "spans": [ + { + "bbox": [ + 62, + 258, + 295, + 301 + ], + "type": "text", + "content": "[5] Andreas Blattmann, Robin Rombach, Huan Ling, Tim Dockhorn, Seung Wook Kim, Sanja Fidler, and Karsten Kreis. Align your latents: High-resolution video synthesis with latent diffusion models. In CVPR, 2023. 3" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 62, + 302, + 295, + 356 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 302, + 295, + 356 + ], + "spans": [ + { + "bbox": [ + 62, + 302, + 295, + 356 + ], + "type": "text", + "content": "[6] Tim Brooks, Bill Peebles, Connor Holmes, Will DePue, Yufei Guo, Li Jing, David Schnurr, Joe Taylor, Troy Luhman, Eric Luhman, Clarence Ng, Ricky Wang, and Aditya Ramesh. Video generation models as world simulators. 2024. 3" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 62, + 357, + 295, + 411 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 357, + 295, + 411 + ], + "spans": [ + { + "bbox": [ + 62, + 357, + 295, + 411 + ], + "type": "text", + "content": "[7] Holger Caesar, Varun Bankiti, Alex H Lang, Sourabh Vora, Venice Erin Liong, Qiang Xu, Anush Krishnan, Yu Pan, Giancarlo Baldan, and Oscar Beijbom. nuscenes: A multimodal dataset for autonomous driving. In CVPR, 2020. 6, 8" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 62, + 413, + 295, + 456 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 413, + 295, + 456 + ], + "spans": [ + { + "bbox": [ + 62, + 413, + 295, + 456 + ], + "type": "text", + "content": "[8] Quang-Huy Che, Dinh-Phuc Nguyen, Minh-Quan Pham, and Duc-Khai Lam. Twinlitenet: An efficient and lightweight model for driveable area and lane segmentation in self-driving cars. In MAPR, 2023. 7" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 62, + 458, + 295, + 501 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 458, + 295, + 501 + ], + "spans": [ + { + "bbox": [ + 62, + 458, + 295, + 501 + ], + "type": "text", + "content": "[9] Yurui Chen, Chun Gu, Junzhe Jiang, Xiatian Zhu, and Li Zhang. Periodic vibration gaussian: Dynamic urban scene reconstruction and real-time rendering. arXiv preprint arXiv:2311.18561, 2023. 1, 2, 5, 6, 7, 8" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 57, + 502, + 295, + 535 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 502, + 295, + 535 + ], + "spans": [ + { + "bbox": [ + 57, + 502, + 295, + 535 + ], + "type": "text", + "content": "[10] Zilong Chen, Yikai Wang, Feng Wang, Zhengyi Wang, and Huaping Liu. V3d: Video diffusion models are effective 3d generators. arXiv preprint arXiv:2403.06738, 2024. 3" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 57, + 536, + 295, + 590 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 536, + 295, + 590 + ], + "spans": [ + { + "bbox": [ + 57, + 536, + 295, + 590 + ], + "type": "text", + "content": "[11] Ziyu Chen, Jiawei Yang, Jiahui Huang, Riccardo de Lutio, Janick Martinez Esturo, Boris Ivanovic, Or Litany, Zan Gojcic, Sanja Fidler, Marco Pavone, Li Song, and Yue Wang. Omnire: Omni urban scene reconstruction. arXiv preprint arXiv:2408.16760, 2024. 2" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 57, + 591, + 295, + 645 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 591, + 295, + 645 + ], + "spans": [ + { + "bbox": [ + 57, + 591, + 295, + 645 + ], + "type": "text", + "content": "[12] Kai Cheng, Xiaoxiao Long, Wei Yin, Jin Wang, Zhiqiang Wu, Yuexin Ma, Kaixuan Wang, Xiaozhi Chen, and Xuejin Chen. Uc-nerf: Neural radiance field for under-calibrated multi-view cameras in autonomous driving. arXiv preprint arXiv:2311.16945, 2023. 2" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 57, + 647, + 295, + 680 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 647, + 295, + 680 + ], + "spans": [ + { + "bbox": [ + 57, + 647, + 295, + 680 + ], + "type": "text", + "content": "[13] Xiaobiao Du, Yida Wang, and Xin Yu. Mvgs: Multi-view-regulated gaussian splattingg for novel view synthesis. arXiv preprint arXiv:2410.02103, 2024. 2" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 57, + 681, + 295, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 681, + 295, + 713 + ], + "spans": [ + { + "bbox": [ + 57, + 681, + 295, + 713 + ], + "type": "text", + "content": "[14] Tobias Fischer, Jonas Kulhanek, Samuel Rota Bulo, Lorenzo Porzi, Marc Pollefeys, and Peter Kontschieder. Dynamic 3d gaussian fields for urban areas. NeurIPS, 2024. 2" + } + ] + } + ], + "index": 14 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 316, + 73, + 554, + 713 + ], + "type": "list", + "angle": 0, + "index": 30, + "blocks": [ + { + "bbox": [ + 316, + 73, + 554, + 116 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 73, + 554, + 116 + ], + "spans": [ + { + "bbox": [ + 316, + 73, + 554, + 116 + ], + "type": "text", + "content": "[15] Sara Fridovich-Keil, Giacomo Meanti, Frederik Rahbæk Warburg, Benjamin Recht, and Angjoo Kanazawa. K-planes: Explicit radiance fields in space, time, and appearance. In CVPR, 2023. 2" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 317, + 118, + 554, + 161 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 118, + 554, + 161 + ], + "spans": [ + { + "bbox": [ + 317, + 118, + 554, + 161 + ], + "type": "text", + "content": "[16] Ruiyuan Gao, Kai Chen, Zhihao Li, Lanqing Hong, Zhenguo Li, and Qiang Xu. Magicdrive3d: Controllable 3d generation for any-view rendering in street scenes. arXiv preprint arXiv:2405.14475, 2024. 3" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 317, + 163, + 554, + 217 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 163, + 554, + 217 + ], + "spans": [ + { + "bbox": [ + 317, + 163, + 554, + 217 + ], + "type": "text", + "content": "[17] Ruiqi Gao, Aleksander Holynski, Philipp Henzler, Arthur Brussee, Ricardo Martin-Brualla, Pratul Srinivasan, Jonathan T Barron, and Ben Poole. Cat3d: Create anything in 3d with multi-view diffusion models. arXiv preprint arXiv:2405.10314, 2024. 3" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 317, + 219, + 554, + 274 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 219, + 554, + 274 + ], + "spans": [ + { + "bbox": [ + 317, + 219, + 554, + 274 + ], + "type": "text", + "content": "[18] Shenyuan Gao, Jiazhi Yang, Li Chen, Kashyap Chitta, Yihang Qiu, Andreas Geiger, Jun Zhang, and Hongyang Li. Vista: A generalizable driving world model with high fidelity and versatile controllability. arXiv preprint arXiv:2405.17398, 2024. 2, 3" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 317, + 275, + 554, + 329 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 275, + 554, + 329 + ], + "spans": [ + { + "bbox": [ + 317, + 275, + 554, + 329 + ], + "type": "text", + "content": "[19] Rohit Girdhar, Mannat Singh, Andrew Brown, Quentin Duval, Samaneh Azadi, Sai Saketh Rambhatla, Akbar Shah, Xi Yin, Devi Parikh, and Ishan Misra. Emu video: Factorizing text-to-video generation by explicit image conditioning. arXiv preprint arXiv:2311.10709, 2023. 3" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 317, + 331, + 554, + 385 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 331, + 554, + 385 + ], + "spans": [ + { + "bbox": [ + 317, + 331, + 554, + 385 + ], + "type": "text", + "content": "[20] Jianfei Guo, Nianchen Deng, Xinyang Li, Yeqi Bai, Botian Shi, Chiyu Wang, Chenjing Ding, Dongliang Wang, and Yikang Li. Streetsurf: Extending multi-view implicit surface reconstruction to street views. arXiv preprint arXiv:2306.04988, 2023. 2" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 317, + 387, + 554, + 430 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 387, + 554, + 430 + ], + "spans": [ + { + "bbox": [ + 317, + 387, + 554, + 430 + ], + "type": "text", + "content": "[21] Agrim Gupta, Lijun Yu, Kihyuk Sohn, Xiuye Gu, Meera Hahn, Li Fei-Fei, Irfan Essa, Lu Jiang, and José Lezama. Photorealistic video generation with diffusion models. arXiv preprint arXiv:2312.06662, 2023. 3" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 317, + 432, + 554, + 475 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 432, + 554, + 475 + ], + "spans": [ + { + "bbox": [ + 317, + 432, + 554, + 475 + ], + "type": "text", + "content": "[22] Huasong Han, Kaixuan Zhou, Xiaoxiao Long, Yusen Wang, and Chunxia Xiao. Ggs: Generalizable gaussian splatting for lane switching in autonomous driving. arXiv preprint arXiv:2409.02382, 2024. 2, 3" + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 317, + 476, + 554, + 510 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 476, + 554, + 510 + ], + "spans": [ + { + "bbox": [ + 317, + 476, + 554, + 510 + ], + "type": "text", + "content": "[23] Ayaan Haque, Matthew Tancik, Alexei A Efros, Aleksander Holynski, and Angjoo Kanazawa. Instruct-nerf2nerf: Editing 3d scenes with instructions. In CVPR, 2023. 2" + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 317, + 511, + 554, + 554 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 511, + 554, + 554 + ], + "spans": [ + { + "bbox": [ + 317, + 511, + 554, + 554 + ], + "type": "text", + "content": "[24] Martin Heusel, Hubert Ramsauer, Thomas Unterthiner, Bernhard Nessler, and Sepp Hochreiter. Gans trained by a two time-scale update rule converge to a local nash equilibrium. NeurIPS, 2017. 6, 7" + } + ] + } + ], + "index": 25 + }, + { + "bbox": [ + 317, + 555, + 554, + 610 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 555, + 554, + 610 + ], + "spans": [ + { + "bbox": [ + 317, + 555, + 554, + 610 + ], + "type": "text", + "content": "[25] Jonathan Ho, William Chan, Chitwan Sahara, Jay Whang, Ruiqi Gao, Alexey Gritsenko, Diederik P Kingma, Ben Poole, Mohammad Norouzi, David J Fleet, et al. Imagen video: High definition video generation with diffusion models. arXiv preprint arXiv:2210.02303, 2022. 3" + } + ] + } + ], + "index": 26 + }, + { + "bbox": [ + 317, + 612, + 554, + 644 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 612, + 554, + 644 + ], + "spans": [ + { + "bbox": [ + 317, + 612, + 554, + 644 + ], + "type": "text", + "content": "[26] Jonathan Ho, Tim Salimans, Alexey Gritsenko, William Chan, Mohammad Norouzi, and David J Fleet. Video diffusion models. NeurIPS, 2022." + } + ] + } + ], + "index": 27 + }, + { + "bbox": [ + 317, + 647, + 554, + 689 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 647, + 554, + 689 + ], + "spans": [ + { + "bbox": [ + 317, + 647, + 554, + 689 + ], + "type": "text", + "content": "[27] Wenyi Hong, Ming Ding, Wendi Zheng, Xinghan Liu, and Jie Tang. Cogvideo: Large-scale pretraining for text-to-video generation via transformers. arXiv preprint arXiv:2205.15868, 2022.3" + } + ] + } + ], + "index": 28 + }, + { + "bbox": [ + 316, + 691, + 554, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 691, + 554, + 713 + ], + "spans": [ + { + "bbox": [ + 316, + 691, + 554, + 713 + ], + "type": "text", + "content": "[28] Anthony Hu, Lloyd Russell, Hudson Yeo, Zak Murez, George Fedoseev, Alex Kendall, Jamie Shotton, and Gian" + } + ] + } + ], + "index": 29 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "text", + "content": "12023" + } + ] + } + ], + "index": 31 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 8 + }, + { + "para_blocks": [ + { + "bbox": [ + 56, + 72, + 295, + 713 + ], + "type": "list", + "angle": 0, + "index": 16, + "blocks": [ + { + "bbox": [ + 76, + 72, + 294, + 105 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 76, + 72, + 294, + 105 + ], + "spans": [ + { + "bbox": [ + 76, + 72, + 294, + 105 + ], + "type": "text", + "content": "luca Corrado. Gaia-1: A generative world model for autonomous driving. arXiv preprint arXiv:2309.17080, 2023. 2, 3" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 56, + 106, + 295, + 149 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 106, + 295, + 149 + ], + "spans": [ + { + "bbox": [ + 56, + 106, + 295, + 149 + ], + "type": "text", + "content": "[29] Shengchao Hu, Li Chen, Penghao Wu, Hongyang Li, Junchi Yan, and Dacheng Tao. St-p3: End-to-end vision-based autonomous driving via spatial-temporal feature learning. In ECCV, 2022. 2" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 56, + 150, + 294, + 194 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 150, + 294, + 194 + ], + "spans": [ + { + "bbox": [ + 56, + 150, + 294, + 194 + ], + "type": "text", + "content": "[30] Yihan Hu, Jiazhi Yang, Li Chen, Keyu Li, Chonghao Sima, Xizhou Zhu, Siqi Chai, Senyao Du, Tianwei Lin, Wenhai Wang, et al. Planning-oriented autonomous driving. In CVPR, 2023. 2" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 56, + 194, + 294, + 249 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 194, + 294, + 249 + ], + "spans": [ + { + "bbox": [ + 56, + 194, + 294, + 249 + ], + "type": "text", + "content": "[31] Nan Huang, Xiaobao Wei, Wenzhao Zheng, Pengju An, Ming Lu, Wei Zhan, Masayoshi Tomizuka, Kurt Keutzer, and Shanghang Zhang. " + }, + { + "bbox": [ + 56, + 194, + 294, + 249 + ], + "type": "inline_equation", + "content": "s^3" + }, + { + "bbox": [ + 56, + 194, + 294, + 249 + ], + "type": "text", + "content": " gaussian: Self-supervised street gaussians for autonomous driving. arXiv preprint arXiv:2405.20323, 2024. 1, 2, 5, 6, 7" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 56, + 250, + 294, + 293 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 250, + 294, + 293 + ], + "spans": [ + { + "bbox": [ + 56, + 250, + 294, + 293 + ], + "type": "text", + "content": "[32] Sungwon Hwang, Min-Jung Kim, Taewoong Kang, Jayeon Kang, and Jaegul Choo. Vegs: View extrapolation of urban scenes in 3d gaussian splatting using learned priors. In ECCV, 2024. 2" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 56, + 293, + 294, + 337 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 293, + 294, + 337 + ], + "spans": [ + { + "bbox": [ + 56, + 293, + 294, + 337 + ], + "type": "text", + "content": "[33] Muhammad Zubair Irshad, Sergey Zakharov, Katherine Liu, Vitor Guizilini, Thomas Kollar, Adrien Gaidon, Zsolt Kira, and Rares Ambrus. Neo 360: Neural fields for sparse view synthesis of outdoor scenes. In ICCV, 2023. 2" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 56, + 338, + 294, + 381 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 338, + 294, + 381 + ], + "spans": [ + { + "bbox": [ + 56, + 338, + 294, + 381 + ], + "type": "text", + "content": "[34] Bo Jiang, Shaoyu Chen, Qing Xu, Bencheng Liao, Jiajie Chen, Helong Zhou, Qian Zhang, Wenyu Liu, Chang Huang, and Xinggang Wang. Vad: Vectorized scene representation for efficient autonomous driving. In ICCV, 2023. 2" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 56, + 382, + 277, + 393 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 382, + 277, + 393 + ], + "spans": [ + { + "bbox": [ + 56, + 382, + 277, + 393 + ], + "type": "text", + "content": "[35] Glenn Jocher and Jing Qiu. Ultralytics yolo11, 2024. 7" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 56, + 393, + 294, + 426 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 393, + 294, + 426 + ], + "spans": [ + { + "bbox": [ + 56, + 393, + 294, + 426 + ], + "type": "text", + "content": "[36] Bernhard Kerbl, Georgios Kopanas, Thomas Leimkuhler, and George Drettakis. 3d gaussian splatting for real-time radiance field rendering. ACM ToG, 2023. 2, 3, 6" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 56, + 426, + 294, + 458 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 426, + 294, + 458 + ], + "spans": [ + { + "bbox": [ + 56, + 426, + 294, + 458 + ], + "type": "text", + "content": "[37] Diederik P Kingma and Jimmy Ba. Adam: A method for stochastic optimization. arXiv preprint arXiv:1412.6980, 2014. 6" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 56, + 459, + 294, + 514 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 459, + 294, + 514 + ], + "spans": [ + { + "bbox": [ + 56, + 459, + 294, + 514 + ], + "type": "text", + "content": "[38] Dan Kondratyuk, Lijun Yu, Xiuye Gu, José Lezama, Jonathan Huang, Rachel Hornung, Hartwig Adam, Hassan Akbari, Yair Alon, Vighnesh Birodkar, et al. Videopoet: A large language model for zero-shot video generation. arXiv preprint arXiv:2312.14125, 2023. 3" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 56, + 514, + 294, + 568 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 514, + 294, + 568 + ], + "spans": [ + { + "bbox": [ + 56, + 514, + 294, + 568 + ], + "type": "text", + "content": "[39] Abhijit Kundu, Kyle Genova, Xiaoqi Yin, Alireza Fathi, Caroline Pantofaru, Leonidas J Guibas, Andrea Tagliasacchi, Frank Dellaert, and Thomas Funkhouser. Panoptic neural fields: A semantic object-aware neural scene representation. In CVPR, 2022. 2" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 56, + 570, + 294, + 591 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 570, + 294, + 591 + ], + "spans": [ + { + "bbox": [ + 56, + 570, + 294, + 591 + ], + "type": "text", + "content": "[40] Yann LeCun and Courant. A path towards autonomous machine intelligence version 0.9.2, 2022-06-27. 2022. 3" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 56, + 592, + 294, + 646 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 592, + 294, + 646 + ], + "spans": [ + { + "bbox": [ + 56, + 592, + 294, + 646 + ], + "type": "text", + "content": "[41] Hao Li, Ming Yuan, Yan Zhang, Chenming Wu, Chen Zhao, Chunyu Song, Haocheng Feng, Errui Ding, Dingwen Zhang, and Jingdong Wang. Xld: A cross-lane dataset for benchmarking novel driving view synthesis. arXiv preprint arXiv:2406.18360, 2024. 2" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 56, + 647, + 294, + 679 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 647, + 294, + 679 + ], + "spans": [ + { + "bbox": [ + 56, + 647, + 294, + 679 + ], + "type": "text", + "content": "[42] Zhengqi Li, Simon Niklaus, Noah Snavely, and Oliver Wang. Neural scene flow fields for space-time view synthesis of dynamic scenes. In CVPR, 2021. 2" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 56, + 680, + 294, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 680, + 294, + 713 + ], + "spans": [ + { + "bbox": [ + 56, + 680, + 294, + 713 + ], + "type": "text", + "content": "[43] Zhiqi Li, Zhiding Yu, Shiyi Lan, Jiahan Li, Jan Kautz, Tong Lu, and Jose M Alvarez. Is ego status all you need for open-loop end-to-end autonomous driving? In CVPR, 2024. 2" + } + ] + } + ], + "index": 15 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 316, + 72, + 554, + 713 + ], + "type": "list", + "angle": 0, + "index": 32, + "blocks": [ + { + "bbox": [ + 316, + 72, + 554, + 116 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 72, + 554, + 116 + ], + "spans": [ + { + "bbox": [ + 316, + 72, + 554, + 116 + ], + "type": "text", + "content": "[44] Zhuopeng Li, Yilin Zhang, Chenming Wu, Jianke Zhu, and Liangjun Zhang. Ho-gaussian: Hybrid optimization of 3d gaussian splatting for urban scenes. arXiv preprint arXiv:2403.20032, 2024. 2" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 316, + 118, + 553, + 160 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 118, + 553, + 160 + ], + "spans": [ + { + "bbox": [ + 316, + 118, + 553, + 160 + ], + "type": "text", + "content": "[45] Chen-Hsuan Lin, Jun Gao, Luming Tang, Towaki Takikawa, Xiaohui Zeng, Xun Huang, Karsten Kreis, Sanja Fidler, Ming-Yu Liu, and Tsung-Yi Lin. Magic3d: High-resolution text-to-3d content creation. In CVPR, 2023. 3" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 316, + 161, + 553, + 204 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 161, + 553, + 204 + ], + "spans": [ + { + "bbox": [ + 316, + 161, + 553, + 204 + ], + "type": "text", + "content": "[46] Haotong Lin, Sida Peng, Zhen Xu, Yunzhi Yan, Qing Shuai, Hujun Bao, and Xiaowei Zhou. Efficient neural radiance fields for interactive free-viewpoint video. In SIGGRAPH Asia, 2022. 2" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 316, + 205, + 553, + 238 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 205, + 553, + 238 + ], + "spans": [ + { + "bbox": [ + 316, + 205, + 553, + 238 + ], + "type": "text", + "content": "[47] Fan Lu, Yan Xu, Guang Chen, Hongsheng Li, Kwan-Yee Lin, and Changjun Jiang. Urban radiance field representation with deformable neural mesh primitives. In ICCV, 2023. 2" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 316, + 239, + 553, + 282 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 239, + 553, + 282 + ], + "spans": [ + { + "bbox": [ + 316, + 239, + 553, + 282 + ], + "type": "text", + "content": "[48] Xin Ma, Yaohui Wang, Gengyun Jia, Xinyuan Chen, Zwei Liu, Yuan-Fang Li, Cunjian Chen, and Yu Qiao. Latte: Latent diffusion transformer for video generation. arXiv preprint arXiv:2401.03048, 2024. 3" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 316, + 283, + 553, + 326 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 283, + 553, + 326 + ], + "spans": [ + { + "bbox": [ + 316, + 283, + 553, + 326 + ], + "type": "text", + "content": "[49] Ben Mildenhall, Pratul P Srinivasan, Matthew Tancik, Jonathan T Barron, Ravi Ramamoorthi, and Ren Ng. Nerf: Representing scenes as neural radiance fields for view synthesis. CACM, 2021. 2" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 316, + 327, + 553, + 360 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 327, + 553, + 360 + ], + "spans": [ + { + "bbox": [ + 316, + 327, + 553, + 360 + ], + "type": "text", + "content": "[50] Thomas Müller, Alex Evans, Christoph Schied, and Alexander Keller. Instant neural graphics primitives with a multiresolution hash encoding. ACM ToG, 2022. 2" + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 316, + 360, + 553, + 392 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 360, + 553, + 392 + ], + "spans": [ + { + "bbox": [ + 316, + 360, + 553, + 392 + ], + "type": "text", + "content": "[51] Julian Ost, Fahim Mannan, Nils Thuerey, Julian Knodt, and Felix Heide. Neural scene graphs for dynamic scenes. In CVPR, 2021. 2" + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 316, + 393, + 553, + 448 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 393, + 553, + 448 + ], + "spans": [ + { + "bbox": [ + 316, + 393, + 553, + 448 + ], + "type": "text", + "content": "[52] Keunhong Park, Utkarsh Sinha, Peter Hedman, Jonathan T Barron, Sofien Bouaziz, Dan B Goldman, Ricardo MartinBrualla, and Steven M Seitz. Hypernerf: A higher-dimensional representation for topologically varying neural radiance fields. arXiv preprint arXiv:2106.13228, 2021. 2" + } + ] + } + ], + "index": 25 + }, + { + "bbox": [ + 316, + 449, + 553, + 502 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 449, + 553, + 502 + ], + "spans": [ + { + "bbox": [ + 316, + 449, + 553, + 502 + ], + "type": "text", + "content": "[53] Dustin Podell, Zion English, Kyle Lacey, Andreas Blattmann, Tim Dockhorn, Jonas Müller, Joe Penna, and Robin Rombach. Sdxl: Improving latent diffusion models for high-resolution image synthesis. arXiv preprint arXiv:2307.01952, 2023. 3" + } + ] + } + ], + "index": 26 + }, + { + "bbox": [ + 316, + 503, + 553, + 536 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 503, + 553, + 536 + ], + "spans": [ + { + "bbox": [ + 316, + 503, + 553, + 536 + ], + "type": "text", + "content": "[54] Ben Poole, Ajay Jain, Jonathan T Barron, and Ben Mildenhall. Dreamfusion: Text-to-3d using 2d diffusion. arXiv preprint arXiv:2209.14988, 2022. 3" + } + ] + } + ], + "index": 27 + }, + { + "bbox": [ + 316, + 537, + 553, + 580 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 537, + 553, + 580 + ], + "spans": [ + { + "bbox": [ + 316, + 537, + 553, + 580 + ], + "type": "text", + "content": "[55] Aditya Ramesh, Prafulla Dhariwal, Alex Nichol, Casey Chu, and Mark Chen. Hierarchical text-conditional image generation with clip latents. arXiv preprint arXiv:2204.06125, 2022.3" + } + ] + } + ], + "index": 28 + }, + { + "bbox": [ + 316, + 581, + 553, + 624 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 581, + 553, + 624 + ], + "spans": [ + { + "bbox": [ + 316, + 581, + 553, + 624 + ], + "type": "text", + "content": "[56] Konstantinos Rematas, Andrew Liu, Pratul P Srinivasan, Jonathan T Barron, Andrea Tagliasacchi, Thomas Funkhouser, and Vittorio Ferrari. Urban radiance fields. In CVPR, 2022. 2" + } + ] + } + ], + "index": 29 + }, + { + "bbox": [ + 316, + 624, + 553, + 657 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 624, + 553, + 657 + ], + "spans": [ + { + "bbox": [ + 316, + 624, + 553, + 657 + ], + "type": "text", + "content": "[57] Robin Rombach, Andreas Blattmann, Dominik Lorenz, Patrick Esser, and Björn Ommher. High-resolution image synthesis with latent diffusion models. In CVPR, 2022. 3" + } + ] + } + ], + "index": 30 + }, + { + "bbox": [ + 316, + 658, + 553, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 658, + 553, + 713 + ], + "spans": [ + { + "bbox": [ + 316, + 658, + 553, + 713 + ], + "type": "text", + "content": "[58] Chitwan Sahara, William Chan, Saurabh Saxena, Lala Li, Jay Whang, Emily L Denton, Kamyar Ghasemipour, Raphael Gontijo Lopes, Burcu Karagol Ayan, Tim Salimans, et al. Photorealistic text-to-image diffusion models with deep language understanding. NeurIPS, 2022. 3" + } + ] + } + ], + "index": 31 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "type": "text", + "content": "12024" + } + ] + } + ], + "index": 33 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 9 + }, + { + "para_blocks": [ + { + "bbox": [ + 56, + 72, + 294, + 713 + ], + "type": "list", + "angle": 0, + "index": 13, + "blocks": [ + { + "bbox": [ + 56, + 72, + 294, + 127 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 72, + 294, + 127 + ], + "spans": [ + { + "bbox": [ + 56, + 72, + 294, + 127 + ], + "type": "text", + "content": "[59] Kyle Sargent, Zizhang Li, Tanmay Shah, Charles Herrmann, Hong-Xing Yu, Yunzhi Zhang, Eric Ryan Chan, Dmitry Lagun, Li Fei-Fei, Deqing Sun, et al. Zeronvs: Zero-shot 360-degree view synthesis from a single real image. arXiv preprint arXiv:2310.17994, 2023. 3" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 56, + 129, + 294, + 173 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 129, + 294, + 173 + ], + "spans": [ + { + "bbox": [ + 56, + 129, + 294, + 173 + ], + "type": "text", + "content": "[60] Liangchen Song, Anpei Chen, Zhong Li, Zhang Chen, Lele Chen, Junsong Yuan, Yi Xu, and Andreas Geiger. Nerfplayer: A streamable dynamic scene representation with decomposed neural radiance fields. IEEE TVCG, 2023. 2" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 56, + 175, + 294, + 262 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 175, + 294, + 262 + ], + "spans": [ + { + "bbox": [ + 56, + 175, + 294, + 262 + ], + "type": "text", + "content": "[61] Pei Sun, Henrik Kretzschmar, Xerxes Dotiwalla, Aurelien Chouard, Vijaysai Patnaik, Paul Tsui, James Guo, Yin Zhou, Yuning Chai, Benjamin Caine, Vijay Vasudevan, Wei Han, Jiquan Ngiam, Hang Zhao, Aleksei Timofeev, Scott Ettinger, Maxim Krivokon, Amy Gao, Aditya Joshi, Yu Zhang, Jonathon Shlens, Zhifeng Chen, and Dragomir Anguelov. Scalability in perception for autonomous driving: Waymo open dataset. In CVPR, 2020. 6, 8" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 56, + 263, + 294, + 306 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 263, + 294, + 306 + ], + "spans": [ + { + "bbox": [ + 56, + 263, + 294, + 306 + ], + "type": "text", + "content": "[62] Shanlin Sun, Bingbing Zhuang, Ziyu Jiang, Buyu Liu, Xiaohui Xie, and Manmohan Chandraker. Lidarf: Delving into lidar for neural radiance field on street scenes. In CVPR, 2024. 2" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 56, + 308, + 294, + 352 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 308, + 294, + 352 + ], + "spans": [ + { + "bbox": [ + 56, + 308, + 294, + 352 + ], + "type": "text", + "content": "[63] Matthew Tancik, Vincent Casser, Xinchen Yan, Sabeek Pradhan, Ben Mildenhall, Pratul P Srinivasan, Jonathan T Barron, and Henrik Kretzschmar. Block-nerf: Scalable large scene neural view synthesis. In CVPR, 2022. 2" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 56, + 354, + 294, + 397 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 354, + 294, + 397 + ], + "spans": [ + { + "bbox": [ + 56, + 354, + 294, + 397 + ], + "type": "text", + "content": "[64] Adam Tonderski, Carl Lindström, Georg Hess, William Ljungbergh, Lennart Svensson, and Christoffer Petersson. Neurad: Neural rendering for autonomous driving. In CVPR, 2024. 2, 6, 7, 8" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 56, + 399, + 294, + 453 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 399, + 294, + 453 + ], + "spans": [ + { + "bbox": [ + 56, + 399, + 294, + 453 + ], + "type": "text", + "content": "[65] Vikram Voleti, Chun-Han Yao, Mark Boss, Adam Letts, David Pankratz, Dmitry Tochilkin, Christian Laforte, Robin Rombach, and Varun Jampani. Sv3d: Novel multi-view synthesis and 3d generation from a single image using latent video diffusion. arXiv preprint arXiv:2403.12008, 2024. 3" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 56, + 455, + 294, + 498 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 455, + 294, + 498 + ], + "spans": [ + { + "bbox": [ + 56, + 455, + 294, + 498 + ], + "type": "text", + "content": "[66] Xiaofeng Wang, Zheng Zhu, Guan Huang, Xinze Chen, Jiagang Zhu, and Jiwen Lu. Drivedreamer: Towards real-world-driven world models for autonomous driving. arXiv preprint arXiv:2309.09777, 2023. 2, 3" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 56, + 500, + 294, + 554 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 500, + 294, + 554 + ], + "spans": [ + { + "bbox": [ + 56, + 500, + 294, + 554 + ], + "type": "text", + "content": "[67] Xiaofeng Wang, Kang Zhao, Feng Liu, Jiayu Wang, Guosheng Zhao, Xiaoyi Bao, Zheng Zhu, Yingya Zhang, and Xingang Wang. Egovid-5m: A large-scale video-action dataset for egocentric video generation. arXiv preprint arXiv:2411.08380, 2024. 3" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 56, + 555, + 294, + 600 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 555, + 294, + 600 + ], + "spans": [ + { + "bbox": [ + 56, + 555, + 294, + 600 + ], + "type": "text", + "content": "[68] Xiaofeng Wang, Zheng Zhu, Guan Huang, Boyuan Wang, Xinze Chen, and Jiwen Lu. Worlddreamer: Towards general world models for video generation via predicting masked tokens. arXiv preprint arXiv:2401.09985, 2024. 2, 3" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 56, + 601, + 294, + 644 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 601, + 294, + 644 + ], + "spans": [ + { + "bbox": [ + 56, + 601, + 294, + 644 + ], + "type": "text", + "content": "[69] Yuqi Wang, Jiawei He, Lue Fan, Hongxin Li, Yuntao Chen, and Zhaoxiang Zhang. Driving into the future: Multiview visual forecasting and planning with world model for autonomous driving. In CVPR, 2024. 2, 3" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 56, + 647, + 294, + 689 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 647, + 294, + 689 + ], + "spans": [ + { + "bbox": [ + 56, + 647, + 294, + 689 + ], + "type": "text", + "content": "[70] Rundi Wu, Ben Mildenhall, Philipp Henzler, Keunhong Park, Ruiqi Gao, Daniel Watson, Pratul P Srinivasan, Dor Verbin, Jonathan T Barron, Ben Poole, et al. Reconfusion: 3d reconstruction with diffusion priors. In CVPR, 2024. 3" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 56, + 691, + 294, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 691, + 294, + 713 + ], + "spans": [ + { + "bbox": [ + 56, + 691, + 294, + 713 + ], + "type": "text", + "content": "[71] Zirui Wu, Tianyu Liu, Liyi Luo, Zhide Zhong, Jianteng Chen, Hongmin Xiao, Chao Hou, Haozhe Lou, Yuantao" + } + ] + } + ], + "index": 12 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 316, + 73, + 553, + 713 + ], + "type": "list", + "angle": 0, + "index": 29, + "blocks": [ + { + "bbox": [ + 333, + 73, + 553, + 105 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 333, + 73, + 553, + 105 + ], + "spans": [ + { + "bbox": [ + 333, + 73, + 553, + 105 + ], + "type": "text", + "content": "Chen, Runyi Yang, et al. Mars: An instance-aware, modular and realistic simulator for autonomous driving. In ICAI, 2023. 2" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 316, + 106, + 553, + 161 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 106, + 553, + 161 + ], + "spans": [ + { + "bbox": [ + 316, + 106, + 553, + 161 + ], + "type": "text", + "content": "[72] Jiannan Xiang, Guangyi Liu, Yi Gu, Qiyue Gao, Yuting Ning, Yuheng Zha, Zeyu Feng, Tianhua Tao, Shibo Hao, Yemin Shi, et al. Pandora: Towards general world model with natural language actions and video states. arXiv preprint arXiv:2406.09455, 2024. 3" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 316, + 162, + 553, + 205 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 162, + 553, + 205 + ], + "spans": [ + { + "bbox": [ + 316, + 162, + 553, + 205 + ], + "type": "text", + "content": "[73] Pengchuan Xiao, Zhenlei Shao, Steven Hao, Zishuo Zhang, Xiaolin Chai, Judy Jiao, Zesong Li, Jian Wu, Kai Sun, Kun Jiang, et al. Pandaset: Advanced sensor suite dataset for autonomous driving. In ITSC, 2021. 6, 8" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 316, + 206, + 553, + 237 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 206, + 553, + 237 + ], + "spans": [ + { + "bbox": [ + 316, + 206, + 553, + 237 + ], + "type": "text", + "content": "[74] Ziyang Xie, Junge Zhang, Wenye Li, Feihu Zhang, and Li Zhang. S-nerf: Neural radiance fields for street views. arXiv preprint arXiv:2303.00749, 2023. 2" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 316, + 239, + 553, + 271 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 239, + 553, + 271 + ], + "spans": [ + { + "bbox": [ + 316, + 239, + 553, + 271 + ], + "type": "text", + "content": "[75] Wilson Yan, Yunzhi Zhang, Pieter Abbeel, and Aravind Srinivas. Videogpt: Video generation using vq-vae and transformers. arXiv preprint arXiv:2104.10157, 2021. 3" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 316, + 273, + 553, + 316 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 273, + 553, + 316 + ], + "spans": [ + { + "bbox": [ + 316, + 273, + 553, + 316 + ], + "type": "text", + "content": "[76] Yunzhi Yan, Haotong Lin, Chenxu Zhou, Weijie Wang, Haiyang Sun, Kun Zhan, Xianpeng Lang, Xiaowei Zhou, and Sida Peng. Street gaussians for modeling dynamic urban scenes. arXiv preprint arXiv:2401.01339, 2024. 2" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 316, + 317, + 553, + 369 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 317, + 553, + 369 + ], + "spans": [ + { + "bbox": [ + 316, + 317, + 553, + 369 + ], + "type": "text", + "content": "[77] Jiawei Yang, Boris Ivanovic, Or Litany, Xinshuo Weng, Seung Wook Kim, Boyi Li, Tong Che, Danfei Xu, Sanja Fidler, Marco Pavone, et al. Emernerf: Emergent spatial-temporal scene decomposition via self-supervision. arXiv preprint arXiv:2311.02077, 2023. 2" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 316, + 372, + 553, + 435 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 372, + 553, + 435 + ], + "spans": [ + { + "bbox": [ + 316, + 372, + 553, + 435 + ], + "type": "text", + "content": "[78] Xuemeng Yang, Licheng Wen, Yukai Ma, Jianbiao Mei, Xin Li, Tiantian Wei, Wenjie Lei, Daocheng Fu, Pinlong Cai, Min Dou, Botian Shi, Liang He, Yong Liu, and Yu Qiao. Drivearena: A closed-loop generative simulation platform for autonomous driving. arXiv preprint arXiv:2408.00415, 2024.3" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 316, + 437, + 553, + 480 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 437, + 553, + 480 + ], + "spans": [ + { + "bbox": [ + 316, + 437, + 553, + 480 + ], + "type": "text", + "content": "[79] Ze Yang, Yun Chen, Jingkang Wang, Sivabalan Manivasagam, Wei-Chiu Ma, Anqi Joyce Yang, and Raquel Urtasun. Unisim: A neural closed-loop sensor simulator. In CVPR, 2023. 2, 6, 7, 8" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 316, + 482, + 553, + 525 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 482, + 553, + 525 + ], + "spans": [ + { + "bbox": [ + 316, + 482, + 553, + 525 + ], + "type": "text", + "content": "[80] Ziyi Yang, Xinyu Gao, Wen Zhou, Shaohui Jiao, Yuqing Zhang, and Xiaogang Jin. Deformable 3d gaussians for high-fidelity monocular dynamic scene reconstruction. In CVPR, 2024. 1, 2, 6, 7" + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 316, + 526, + 553, + 579 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 526, + 553, + 579 + ], + "spans": [ + { + "bbox": [ + 316, + 526, + 553, + 579 + ], + "type": "text", + "content": "[81] Zhuoyi Yang, Jiayan Teng, Wendi Zheng, Ming Ding, Shiyu Huang, Jiazheng Xu, Yuanming Yang, Wenyi Hong, Xiaohan Zhang, Guanyu Feng, et al. Cogvideox: Text-to-video diffusion models with an expert transformer. arXiv preprint arXiv:2408.06072, 2024. 3" + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 316, + 581, + 553, + 613 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 581, + 553, + 613 + ], + "spans": [ + { + "bbox": [ + 316, + 581, + 553, + 613 + ], + "type": "text", + "content": "[82] Wang Yifan, Felice Serena, Shihao Wu, Cengiz Öztireli, and Olga Sorkine-Hornung. Differentiable surface splatting for point-based geometry processing. ACM TOG, 2019. 3, 5" + } + ] + } + ], + "index": 25 + }, + { + "bbox": [ + 316, + 615, + 553, + 646 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 615, + 553, + 646 + ], + "spans": [ + { + "bbox": [ + 316, + 615, + 553, + 646 + ], + "type": "text", + "content": "[83] Zehao Yu, Anpei Chen, Binbin Huang, Torsten Sattler, and Andreas Geiger. Mip-splatting: Alias-free 3d gaussian splattering. In CVPR, 2024. 2" + } + ] + } + ], + "index": 26 + }, + { + "bbox": [ + 316, + 647, + 553, + 690 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 647, + 553, + 690 + ], + "spans": [ + { + "bbox": [ + 316, + 647, + 553, + 690 + ], + "type": "text", + "content": "[84] Zhongrui Yu, Haoran Wang, Jinze Yang, Hanzhang Wang, Zeke Xie, Yunfeng Cai, Jiale Cao, Zhong Ji, and Mingming Sun. Sgd: Street view synthesis with gaussian splatting and diffusion prior. arXiv preprint arXiv:2403.20079, 2024. 2, 3" + } + ] + } + ], + "index": 27 + }, + { + "bbox": [ + 316, + 692, + 553, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 692, + 553, + 713 + ], + "spans": [ + { + "bbox": [ + 316, + 692, + 553, + 713 + ], + "type": "text", + "content": "[85] Yan Zeng, Guoqiang Wei, Jiani Zheng, Jiaxin Zou, Yang Wei, Yuchen Zhang, and Hang Li. Make pixels" + } + ] + } + ], + "index": 28 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "text", + "content": "12025" + } + ] + } + ], + "index": 30 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 10 + }, + { + "para_blocks": [ + { + "bbox": [ + 56, + 72, + 296, + 362 + ], + "type": "list", + "angle": 0, + "index": 6, + "blocks": [ + { + "bbox": [ + 76, + 72, + 296, + 95 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 76, + 72, + 296, + 95 + ], + "spans": [ + { + "bbox": [ + 76, + 72, + 296, + 95 + ], + "type": "text", + "content": "dance: High-dynamic video generation. arXiv preprint arXiv:2311.10982, 2023. 3" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 56, + 96, + 296, + 150 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 96, + 296, + 150 + ], + "spans": [ + { + "bbox": [ + 56, + 96, + 296, + 150 + ], + "type": "text", + "content": "[86] Jiang-Tian Zhai, Ze Feng, Jinhao Du, Yongqiang Mao, Jiang-Jiang Liu, Zichang Tan, Yifu Zhang, Xiaoqing Ye, and Jingdong Wang. Rethinking the open-loop evaluation of end-to-end autonomous driving in nuscenes. arXiv preprint arXiv:2305.10430, 2023. 2" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 56, + 152, + 295, + 206 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 152, + 295, + 206 + ], + "spans": [ + { + "bbox": [ + 56, + 152, + 295, + 206 + ], + "type": "text", + "content": "[87] Guosheng Zhao, Xiaofeng Wang, Zheng Zhu, Xinze Chen, Guan Huang, Xiaoyi Bao, and Xingang Wang. Drivedreamer-2: Llm-enhanced world models for diverse driving video generation. arXiv preprint arXiv:2403.06845, 2024. 2, 3, 5" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 56, + 208, + 295, + 252 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 208, + 295, + 252 + ], + "spans": [ + { + "bbox": [ + 56, + 208, + 295, + 252 + ], + "type": "text", + "content": "[88] Hongyu Zhou, Jiahao Shao, Lu Xu, Dongfeng Bai, Weichao Qiu, Bingbing Liu, Yue Wang, Andreas Geiger, and Yiyi Liao. Hugs: Holistic urban 3d scene understanding via gaussian splatting. In CVPR, 2024. 2" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 56, + 253, + 295, + 306 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 253, + 295, + 306 + ], + "spans": [ + { + "bbox": [ + 56, + 253, + 295, + 306 + ], + "type": "text", + "content": "[89] Xiaoyu Zhou, Zhiwei Lin, Xiaojun Shan, Yongtao Wang, Deqing Sun, and Ming-Hsuan Yang. Drivinggaussian: Composite gaussian splatting for surrounding dynamic autonomous driving scenes. In CVPR, pages 21634-21643, 2024. 2" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 56, + 308, + 295, + 362 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 308, + 295, + 362 + ], + "spans": [ + { + "bbox": [ + 56, + 308, + 295, + 362 + ], + "type": "text", + "content": "[90] Zheng Zhu, Xiaofeng Wang, Wangbo Zhao, Chen Min, Ni-anchen Deng, Min Dou, Yuqi Wang, Botian Shi, Kai Wang, Chi Zhang, et al. Is sora a world simulator? a comprehensive survey on general world models and beyond. arXiv preprint arXiv:2405.03520, 2024. 3" + } + ] + } + ], + "index": 5 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 749, + 318, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 749, + 318, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 749, + 318, + 757 + ], + "type": "text", + "content": "12026" + } + ] + } + ], + "index": 7 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 11 + } + ], + "_backend": "vlm", + "_version_name": "2.6.4" +} \ No newline at end of file diff --git a/2025/DriveGEN_ Generalized and Robust 3D Detection in Driving via Controllable Text-to-Image Diffusion Generation/d5fb6699-8753-41df-8e0e-5b82ff7bf5a3_content_list.json b/2025/DriveGEN_ Generalized and Robust 3D Detection in Driving via Controllable Text-to-Image Diffusion Generation/d5fb6699-8753-41df-8e0e-5b82ff7bf5a3_content_list.json new file mode 100644 index 0000000000000000000000000000000000000000..4544be940cce2f23e0e2c98e00214d32cb538c1b --- /dev/null +++ b/2025/DriveGEN_ Generalized and Robust 3D Detection in Driving via Controllable Text-to-Image Diffusion Generation/d5fb6699-8753-41df-8e0e-5b82ff7bf5a3_content_list.json @@ -0,0 +1,1364 @@ +[ + { + "type": "text", + "text": "DriveGEN: Generalized and Robust 3D Detection in Driving via Controllable Text-to-Image Diffusion Generation", + "text_level": 1, + "bbox": [ + 194, + 130, + 805, + 176 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Hongbin Lin $^{1,2}$ Zilu Guo $^{1,2}$ Yifan Zhang $^{3}$ Shuaicheng Niu $^{4}$ Yafeng Li $^{5}$ Ruimao Zhang $^{6}$ Shuguang Cui $^{2,1}$ Zhen Li $^{2,1*}$", + "bbox": [ + 232, + 202, + 759, + 239 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "$^{1}$ FNii-Shenzhen $^{2}$ SSE, CUHK-Shenzhen $^{3}$ National University of Singapore $^{4}$ Nanyang Technological University $^{5}$ Baoji University of Arts and Sciences $^{6}$ Sun Yat-sen University", + "bbox": [ + 101, + 239, + 893, + 275 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/f4971ec62903b3190e1607d19ff2aa63e8d36e158741595d7c392d6280f7eebb.jpg", + "image_caption": [ + "Figure 1. An illustration of DriveGEN for enhancing vision-centric 3D detectors. DriveGEN extends original training images to diverse Out-of-Distribution (OOD) scenarios without additional diffusion model training, preserving all annotated objects. Even with a single augmentation, i.e., 'Snow', the augmented detector [63] is improved not only on the original KITTI [10] val set (Ideal) and already known 'Snow' scenario, but also in unseen 'Fog' and 'Defocus' scenarios, demonstrating a comprehensive model generalizability improvement." + ], + "image_footnote": [], + "bbox": [ + 138, + 286, + 859, + 583 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Abstract", + "text_level": 1, + "bbox": [ + 248, + 686, + 326, + 702 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "In autonomous driving, vision-centric 3D detection aims to identify 3D objects from images. However, high data collection costs and diverse real-world scenarios limit the scale of training data. Once distribution shifts occur between training and test data, existing methods often suffer from performance degradation, known as Out-of-Distribution (OOD) problems. To address this, controllable Text-to-Image (T2I) diffusion offers a potential solution for training data enhancement, which is required to generate diverse OOD scenarios with precise 3D object geometry. Nevertheless, exist", + "bbox": [ + 89, + 720, + 485, + 872 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "ing controllable T2I approaches are restricted by the limited scale of training data or struggle to preserve all annotated 3D objects. In this paper, we present DriveGEN, a method designed to improve the robustness of 3D detectors in Driving via Training-Free Controllable Text-to-Image Diffusion Generation. Without extra diffusion model training, DriveGEN consistently preserves objects with precise 3D geometry across diverse OOD generations, consisting of 2 stages: 1) Self-Prototype Extraction: We empirically find that self-attention features are semantic-aware but require accurate region selection for 3D objects. Thus, we extract precise object features via layouts to capture 3D object geometry, termed self-prototypes. 2) Prototype-Guided Diffusion: To preserve objects across various OOD scenarios, we per", + "bbox": [ + 511, + 688, + 908, + 902 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "CVF", + "bbox": [ + 106, + 2, + 181, + 42 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore.", + "bbox": [ + 236, + 0, + 810, + 46 + ], + "page_idx": 0 + }, + { + "type": "page_footnote", + "text": "*Corresponding author.", + "bbox": [ + 109, + 887, + 236, + 900 + ], + "page_idx": 0 + }, + { + "type": "page_number", + "text": "27497", + "bbox": [ + 478, + 944, + 519, + 957 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "form semantic-aware feature alignment and shallow feature alignment during denoising. Extensive experiments demonstrate our effectiveness in improving 3D detection. The code is available at github.com/Hongbin98/DriveGEN.", + "bbox": [ + 88, + 90, + 480, + 151 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "1. Introduction", + "text_level": 1, + "bbox": [ + 89, + 180, + 222, + 196 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Three-dimensional (3D) object detection intends to identify objects and assess their spatial and dimensional attributes via various sensor inputs, which is widely explored in Autonomous Driving (AD) and robotics [3-5, 19, 49, 50, 58]. To save sensor costs, there is a growing trend towards implementing perception systems by vision-centric 3D detection which relies solely on single or multi-view RGB images along with camera calibration information. Despite challenges in this field, existing methods [53, 56, 60] have demonstrated promising results across various benchmarks.", + "bbox": [ + 91, + 205, + 482, + 356 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Behind the success, one prerequisite is that test images share the same distribution as training images. However, this assumption is often not held in diverse real-world scenarios [21, 32, 62, 64], especially for perception systems that are required for long durations, like autonomous vehicles navigating through different weather conditions. Once the environment changes, well-trained detectors may suffer performance degradation due to the existence of data distribution shifts between training images and test images. As shown in Figure 1, the original detector performs well in the ideal (sunny) scenario while failing to maintain stable performance in the Snow and Fog scenario due to the out-of-distribution (OOD) issue. Considering the widespread application of 3D detection, significant performance degradation on OOD test data may result in traffic accidents and serious safety risks. Thus, it is essential to address the OOD generalization problem of vision-centric 3D detection.", + "bbox": [ + 91, + 357, + 482, + 613 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Recently, two attempts sought to solve the distribution shifts in AD by test-time model adaptation [22] or restoring OOD scenarios to ideal conditions with a weather-adaptive diffusion model [28]. However, they require additional consumption during the test phase, leading to additional energy costs while the driving range is crucial in AD. To save test-time computation costs, we explore controllable Text-to-Image (T2I) diffusion generation for training data augmentation, thereby improving the robustness of 3D detectors. Prior training-based controllable T2I approaches such as ControlNet [59] and T2i-adapter [27] offer users fine-grained spatial control but rely on substantial training data to train auxiliary modules, making them impractical for handling OOD scenarios in AD. As for training-free methods, they [25, 46] leverage low-resolution self-attention features to capture image structures for diffusion model guidance. However, these features are relatively coarse for preserving object geometric characteristics of different classes in 3D object detection. As shown in Figure 2, the object", + "bbox": [ + 91, + 614, + 482, + 900 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "geometry may be lost after multiple rounds of downsampling $(32\\mathrm{x})$ , leading to severe orientation error, position misalignment and omissions, particularly for tiny objects. Therefore, an eligible training-free controllable T2I diffusion method for training data enhancement in 3D detection should preserve all objects with precise 3D geometry.", + "bbox": [ + 511, + 90, + 903, + 181 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "To address these challenges, we propose a method namely DriveGEN for Generalized and Robust 3D Detection in Driving via Training-Free Controllable Text-to-Image Diffusion GENeration. DriveGEN enhances the model robustness within various OOD scenarios generation, consisting of two stages: 1) Self-Prototype Extraction. To capture accurate object geometric characteristics, we first extract principal components of self-attention features [25, 46] to obtain coarse semantic-aware features. Then, we adopt layouts (i.e., bounding boxes) to re-weight features in object regions via a peak function, termed self-prototypes. 2) Prototype-Guided Diffusion. To preserve objects with precise 3D geometry, we perform semantic-aware feature alignment with self-prototypes during the denoising phase. However, these semantic-aware features are relatively coarse to represent tiny objects as shown in Figure 2. Therefore, we devise a shallow feature alignment strategy to constrain the diffusion process and retrain tiny objects.", + "bbox": [ + 511, + 181, + 903, + 453 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Contributions: 1) To the best of our knowledge, we are the first to explore training-free controllable T2I diffusion generation to enhance the model robustness of vision-centric 3D detection. Even with a single augmentation, DriveGEN improves the 3D detector [63] by an average of $7.6\\mathrm{mAP}$ across 13 OOD scenarios. 2) With self-prototype extraction and prototype-guided diffusion stages, DriveGEN enables OOD scenario augmentation while preserving precise 3D geometry for all objects, enhancing the applicability of diffusion models in autonomous driving. 3) Extensive experiments show DriveGEN significantly improves existing monocular 3D detectors across 13 OOD corruption on KITTI and multi-view 3D detectors on the real scenarios (Night, Rainy) of nuScenes, demonstrating the effectiveness of DriveGEN in boosting model generalizability.", + "bbox": [ + 511, + 454, + 903, + 679 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "2. Related Work", + "text_level": 1, + "bbox": [ + 513, + 694, + 653, + 709 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "We first review the literature on explorations of model robustness in autonomous driving, and then discuss controllable T2I diffusion generation methods. More discussions on vision-centric 3D object detection are in Appendix A.", + "bbox": [ + 511, + 719, + 903, + 779 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Model Robustness in Autonomous Driving. The reliability of vision-centric autonomous driving systems depends on the robustness of perception models, especially under challenging scenarios like adverse weather, varied lighting, and complex urban environments [52]. To improve robustness, existing methods employ diffusion models to generate synthetic driving scenes for training data augmentation. Specifically, BEVGen [44] and BEVControl [57] employ", + "bbox": [ + 511, + 780, + 903, + 900 + ], + "page_idx": 1 + }, + { + "type": "page_number", + "text": "27498", + "bbox": [ + 478, + 944, + 517, + 955 + ], + "page_idx": 1 + }, + { + "type": "image", + "img_path": "images/2b113af7cb6544eac7c504aaefffdea8f96a57ef6a32ee4179d0ab7bbd2d0212.jpg", + "image_caption": [ + "Figure 2. Illustration of the importance of object preservation in 3d detection based on KITTI. ControlNet [59] suffers object orientation errors and omissions even if accurate segmentation masks [34] and rich prompts [6] are provided, showing that training-based methods may struggle with spatial control with limited training data. Additionally, Freecontrol [25] relies on coarse low-resolution features to capture semantic structures that may encounter potential object geometry loss, resulting in object position misalignment and omission issues." + ], + "image_footnote": [], + "bbox": [ + 94, + 89, + 903, + 180 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Layouts to create flexible scene configurations. Innovative frameworks like Panacea [43] and MagicDrive [14] extend this controllability to multi-view scenes, while models such as DriveDreamer [66] and DrivingDiffusion [16] focus on generating realistic videos for model validation. Recently, RoboFusion [42] has sought to tackle OOD noise scenarios by leveraging visual foundation models like SAM.", + "bbox": [ + 88, + 273, + 480, + 380 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Controllable T2I Diffusion. Controllable text-to-image (T2I) diffusion models [35, 61, 69] are increasingly applied in real-world scenarios, where text prompts or scene descriptions guide the creation of synthetic data. By leveraging pre-trained models like Stable Diffusion [35] or other large-scale T2I models [33, 37], existing methods can generate high-fidelity and diverse images based on textual input, allowing for controlled variations [2, 65]. Techniques such as Uni-ControlNet [67], UniControl [30] and Layout-diffusion [68] further enhance control by integrating multimodal inputs, creating contextually accurate and condition-specific scenes. Recent works also explore the utility of cross-attention mechanisms and latent embeddings to align generated images closely with text descriptions, supporting the generation for perception tasks like semantic segmentation and depth estimation [15, 17]. However, these methods require a large amount of training data to train auxiliary modules, making them impractical for handling OOD scenarios in AD for the limited data scale.", + "bbox": [ + 91, + 380, + 483, + 666 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Another alternative choice is training-free controllable T2I diffusion, PnP [46] directly injects the self-attention features of condition images to guide the diffusion process while Freecontrol [25] utilizes PCA to extract principle components of self-attention features for semantic structures extraction. However, existing approaches solely rely on the relatively coarse self-attention features, which often results in object region omissions and misalignment as we mentioned before. To this end, we propose our DriveGEN to preserve all objects with precise 3D geometry.", + "bbox": [ + 89, + 667, + 482, + 819 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3. Preliminary", + "text_level": 1, + "bbox": [ + 89, + 829, + 217, + 847 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Pre-trained diffusion models [7, 13, 35] are probabilistic generative models, which consists of two complementary stochastic processes: 1) In the forward process, Gaussian", + "bbox": [ + 89, + 854, + 482, + 902 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "noise is incrementally added to a clean image $\\mathbf{x}_0$ . 2) As for the backward process, it conducts the iterative process of denoising the initial Gaussian noise image $\\mathbf{z}_{N_t}$ , where a cleaner image is given at each step. To be specific, this process leverages a denoising network $\\epsilon_{\\theta}(\\mathbf{z}_t,t)$ to estimate the added noise in the forward process. Subsequently, for $\\mathbf{z}_t$ at each step, the network removes the estimated noise perturbation to achieve a cleaner $\\mathbf{z}_{t - 1}$ .", + "bbox": [ + 511, + 273, + 906, + 393 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "For T2I diffusion models, a previous method [41] has shown that $\\epsilon_{\\theta}$ can approximate the score function for the marginal distributions $p_t$ , i.e., $-\\sigma_t\\nabla_{\\mathbf{z}_t}\\log p_t(\\mathbf{z}_t|\\mathbf{c})$ where $\\sigma_t$ denotes a noise schedule and $\\mathbf{c}$ is a text prompt. To incorporate auxiliary information $y$ into the diffusion sampling process, existing methods [7, 8, 25] guide the diffusion process by adding a time-dependent energy function termed $g(\\mathbf{z}_t;t,y)$ with a strength coefficient $s$ :", + "bbox": [ + 511, + 395, + 908, + 517 + ], + "page_idx": 2 + }, + { + "type": "equation", + "text": "\n$$\n\\hat {\\epsilon} _ {\\boldsymbol {\\theta}} (\\mathbf {z} _ {t}; t, c) = \\epsilon_ {\\boldsymbol {\\theta}} (\\mathbf {z} _ {t}; t, c) - s \\cdot g (\\mathbf {z} _ {t}; t, y),\n$$\n", + "text_format": "latex", + "bbox": [ + 576, + 529, + 841, + 545 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "where $g$ can be defined via bounding boxes [51]. In this paper, we adopt a standard choice $\\epsilon_{\\theta}$ following existing methods [25, 46], i.e., U-Net architecture [36] with multi-level self-attention and cross-attention layers [47].", + "bbox": [ + 511, + 556, + 906, + 617 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "4. Training-Free Controllable T2I Diffusion", + "text_level": 1, + "bbox": [ + 511, + 631, + 882, + 648 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "4.1. Problem Statement", + "text_level": 1, + "bbox": [ + 511, + 666, + 697, + 681 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Without loss of generality, we denote the well-trained 3D detector as $f_{\\Theta_d}(\\cdot)$ which is obtained by training on labeled training images $\\mathcal{D} = \\{(\\mathbf{x}_i,\\mathbf{y}_i)\\}_{i = 1}^N$ . The training images follow the original training distribution $P(\\mathbf{x})$ (i.e., $\\mathbf{x} \\sim P(\\mathbf{x})$ ). Here, $\\Theta_d$ represents the parameters of the detector and $N$ is the total number of training images. After training, the model is applied to unlabeled test images $\\mathcal{D}_t = \\{\\mathbf{x}^t\\}$ . Unfortunately, the well-trained model often encounters Out-of-distribution (OOD) test scenarios in autonomous driving (AD) due to prevalent natural corruptions, namely distribution shifts, i.e., $\\mathbf{x}^t \\sim Q(\\mathbf{x})$ and $P(\\mathbf{x}) \\neq Q(\\mathbf{x})$ .", + "bbox": [ + 511, + 688, + 906, + 854 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "As previously discussed, existing methods [22, 28] are relatively time-consuming due to their additional computation costs at test time. Hence, we investigate controllable", + "bbox": [ + 511, + 854, + 906, + 900 + ], + "page_idx": 2 + }, + { + "type": "page_number", + "text": "27499", + "bbox": [ + 478, + 944, + 519, + 957 + ], + "page_idx": 2 + }, + { + "type": "image", + "img_path": "images/fe48bb7292bb6f94362cb3bfbe03c238620ceaea624228e71750f9429cd5d363.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 140, + 95, + 861, + 290 + ], + "page_idx": 3 + }, + { + "type": "image", + "img_path": "images/a2ac8b0a0517294c12ee272b6275b087c7c24d8403dc581423b81f02c7cdc9e0.jpg", + "image_caption": [ + "Figure 3. An overview of our DriveGEN, consists of two stages: 1) The Self-Prototype Extraction stage is devised to extract accurate semantic structures of multiple objects. To capture precise locations, we achieve fine-grained self-prototypes via leveraging layouts and the peak function to re-weight object regions rather than directly using coarse self-attention features. 2) The Prototype-Guided Diffusion stage conducts semantic-aware feature alignment for semantic structure matching and shallow feature alignment for tiny object preservation." + ], + "image_footnote": [], + "bbox": [ + 140, + 295, + 859, + 463 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Text-to-Image (T2I) diffusion generation for training data augmentation, thereby improving the robustness of 3D detectors. Particularly, it is crucial to preserve object information because the annotations would be reused during detector training. However, training-based approaches [27, 59] require substantial training data to train auxiliary modules for spatial control, which is often unsatisfying in AD. Besides, training-free methods [25, 46] mainly consider the relatively coarse self-attention features, making it challenging to preserve relatively tiny objects.", + "bbox": [ + 88, + 560, + 482, + 712 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "4.2. Overall Scheme", + "text_level": 1, + "bbox": [ + 89, + 724, + 248, + 739 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "After examining the characteristics and challenges of robust 3D object detection, we introduce DriveGEN, a Training-Free Controllable T2I Diffusion Generation method devised to improve the robustness of vision-centric 3D detectors. As illustrated in Figure 3, DriveGEN consists of two stages: 1) Self-Prototype Extraction and 2) Prototype-Guided Diffusion. We briefly introduce each stage below.", + "bbox": [ + 89, + 748, + 482, + 853 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "First, we develop the self-prototype extraction stage (c.f. Section 4.3) to capture the geometric characteristics of objects accurately. Existing methods apply principal compo", + "bbox": [ + 89, + 854, + 483, + 902 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "nent analysis (PCA) on the self-attention features of the first decoder layer of the diffusion model $\\epsilon_{\\theta}$ to capture semantic structures of condition images [25, 46]. However, we empirically find that solely relying on cross-attention relevance between self-attention features and object prompts for region selection may result in object preservation errors (c.f. Figure 2). To address this, DriveGEN leverages layouts to re-weight object regions in the principal components of self-attention features via a peak function and then achieves the self-prototypes which help the diffusion model focus on the accurate object regions during generation.", + "bbox": [ + 511, + 560, + 906, + 728 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Second, we introduce the prototype-guided diffusion stage (c.f. Section 4.4) to preserve all objects during the generation process. On the one hand, we conduct semantic-aware feature alignment between self-attention features and self-prototypes to capture accurate object semantic structures. On the other hand, DriveGEN performs shallow feature alignment to further preserve object information by constraining the noisy latent at the shallow level, compensating for the gaps in semantic-aware feature alignment regarding tiny object preservation. The Pseudo-code of DriveGEN is summarized in Algorithm 1.", + "bbox": [ + 511, + 734, + 908, + 902 + ], + "page_idx": 3 + }, + { + "type": "page_number", + "text": "27500", + "bbox": [ + 478, + 944, + 519, + 957 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "4.3. Self-Prototype Extraction", + "text_level": 1, + "bbox": [ + 89, + 90, + 323, + 107 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "To accurately capture the geometric characteristics of multiple objects, we propose a self-prototype extraction strategy that consists of two components: 1) Image structure extraction and 2) Object region re-weighting.", + "bbox": [ + 89, + 112, + 482, + 172 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Image Structure Extraction. To achieve zero-shot spatial control, capturing the semantic structures of input images is essential. Prior works [25, 46] demonstrate that self-attention features (e.g., keys) of the T2I diffusion model $\\epsilon_{\\theta}$ are semantic-aware [35]. Thus, given the input image $\\mathbf{x}_i$ , we first generate the text prompt $\\mathbf{c}_i$ via a simple template 'A photo of $\\{\\text{object}\\} ,\\ldots,$ and $\\{\\text{object}\\}'$ . Then, we leverage DDIM Inversion [39] to achieve a series of time-dependent diffusion features $\\{\\mathbf{z}_t^I\\in \\mathbb{R}^{N_c\\times H\\times W}\\}_{t = 1}^{N_t}$ where $N_{t}$ denotes the number of DDIM inversion and sampling steps, $N_{c}$ represents the number of feature channels, and $H,W$ indicate the height and width, respectively. During sampling, we apply PCA on self-attention features to extract principal components $\\{\\mathbf{P}_t\\}_{t = 1}^{N_p}$ , each $\\mathbf{P}_t\\in \\mathbb{R}^{N_b\\times H\\times W}$ and $\\mathbf{P}_t = \\{\\mathbf{p}_t^1,\\dots,\\mathbf{p}_t^{N_b}\\}$ . Here $N_{b}$ denotes the number of principal components and we store $\\mathbf{P}_t$ of the initial $N_{p}$ steps.", + "bbox": [ + 89, + 172, + 482, + 416 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Object Region Re-weighting. Existing methods [9, 25] select foreground regions by leveraging cross-attention maps between specific words (i.e., object names) and image features to derive a concept region mask $\\mathbf{M}$ . However, relying solely on such attention maps may result in omissions and misalignment of object regions, as illustrated in Figure 2. To address this, we introduce a peak function $m(\\cdot)$ to reweight $\\mathbf{M}$ for each object region based on layouts $\\mathbf{L}$ after the coordinate transformation through downsampling. To be specific, given a point $(p,q)$ from $\\mathbf{L}_i$ of $\\mathbf{x}_i$ , the corresponding object center is defined as $(\\hat{p},\\hat{q})$ and then we can calculate $m(p,q)$ by:", + "bbox": [ + 89, + 416, + 482, + 598 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\nm (p, q) = \\exp \\left(- \\frac {(p - \\hat {p}) ^ {2} + (q - \\hat {q}) ^ {2}}{2 \\sigma^ {2}}\\right) \\tag {1}\n$$\n", + "text_format": "latex", + "bbox": [ + 145, + 604, + 482, + 641 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "where $\\sigma \\in (0,1)$ controls the range of the influence area. Then, we update the concept token mask $\\mathbf{M}$ by:", + "bbox": [ + 89, + 647, + 482, + 678 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\tilde {\\mathbf {M}} (p, q) = \\mathbf {M} (p, q) + m (p, q) \\cdot \\mathbf {L} _ {i} (p, q). \\tag {2}\n$$\n", + "text_format": "latex", + "bbox": [ + 148, + 686, + 482, + 705 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "The object regions in $\\mathbf{L}_i$ are marked as 1 while all other regions as 0. Eventually, we obtain the self-prototypes by $\\{\\hat{\\mathbf{P}}_t\\}_{t=1}^{N_p} = \\{\\mathbf{P}_t\\}_{t=1}^{N_p}\\big|_{\\hat{\\mathbf{M}}}$ which guide the diffusion model to localize precise locations of objects and combine $\\hat{\\mathbf{M}}$ for accurate object region selection.", + "bbox": [ + 89, + 712, + 482, + 792 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "4.4. Prototype-Guided Diffusion", + "text_level": 1, + "bbox": [ + 89, + 801, + 341, + 816 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "To preserve all objects with precise 3D geometry during diffusion, we leverage the self-prototypes $\\{\\hat{\\mathbf{P}}_t\\}_{t=1}^{N_p}$ and time-dependent diffusion features $\\{\\mathbf{z}_t^I\\}_{t=1}^{N_t}$ for feature alignment, including: 1) semantic-aware feature alignment and 2) shallow feature alignment.", + "bbox": [ + 89, + 823, + 482, + 901 + ], + "page_idx": 4 + }, + { + "type": "code", + "sub_type": "algorithm", + "code_caption": [ + "Algorithm 1 The pipeline of the proposed DriveGEN" + ], + "code_body": "Require: Training data $\\{(\\mathbf{x}_i^s,\\mathbf{y}_i^s)\\}_{i = 1}^N$ Pretrained T2I diffusion model $\\epsilon_{\\theta}$ Hyper-parameters $s,\\sigma ,\\tau$ The OOD scenario. \n#Stage 1: Self-Prototype Extraction \n1: for each training image $x_{i}$ do \n2: Get the text prompt $c_{i}$ and layout $\\mathbf{L}_i$ based on $\\mathbf{y}_i$ \n3: Undergo DDIM Inversion to get $\\mathbf{P}_t$ \n4: Get updated M via Eqn. (1), Eqn. (2) for self-prototypes $\\hat{\\mathbf{P}}_t$ \n5: end for \n#Stage 2: Prototype-Guided Diffusion \n6: for each training image $\\mathbf{x}_i$ do \n7: Update $\\mathbf{c}_i$ with specific OOD type; \n8: for diffusion step $t = 1\\rightarrow N_t$ do \n9: if $t\\leq N_p$ then \n10: Calculate $g_{sa},g_{sl}$ based on Eqn. (3), Eqn. (4); \n11: Get the gradient based on Eqn.(5); \n12: end if \n13: Update the noisy latent $\\mathbf{z}_t\\to \\mathbf{z}_{t - 1}$ \n14: end for \n15: end for \n16: return Output images for all $\\mathbf{x}_i$ in the OOD scenario.", + "bbox": [ + 516, + 109, + 903, + 353 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Semantic-aware Feature Alignment. As mentioned before, the self-attention features of the first decoder layer can represent the semantic structures of input images. Thus, we first initialize the noisy latent $\\mathbf{z}_{N_t}$ via the $\\mathbf{z}_{N_t}^I$ . For the step $t \\leq N_p$ , we perform PCA to obtain the principal components $\\mathbf{P}_t$ of the self-attention features. Then, we adopt a pre-defined threshold $\\tau$ following [9] to select informative regions by $\\hat{\\mathbf{M}} = \\mathbb{I}(\\hat{\\mathbf{M}} > \\tau)$ . Specifically, we calculate the semantic-aware feature alignment loss with $\\mathbf{P}_t$ and $\\hat{\\mathbf{P}}_t$ :", + "bbox": [ + 511, + 378, + 906, + 516 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\ng _ {s a} = \\| \\hat {\\mathbf {M}} \\odot (\\mathbf {P} _ {t} - \\hat {\\mathbf {P}} _ {t}) \\| _ {2} ^ {2}, \\tag {3}\n$$\n", + "text_format": "latex", + "bbox": [ + 616, + 522, + 903, + 542 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "where $\\odot$ represents element-wise multiplication, applying the mask $\\hat{\\mathbf{M}}$ to the difference between $\\mathbf{P}_t$ and $\\hat{\\mathbf{P}}_t$ .", + "bbox": [ + 511, + 547, + 903, + 578 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Shallow Feature Alignment. As shown in Figure 2, we empirically find that principal components lack sufficient details to represent objects, especially tiny objects. For instance, given a bounding box of a cyclist with a height of 20 pixels and a width of 5 pixels, this object is unable to occupy even one element within principal components after multiple rounds of downsampling (e.g., $32\\mathrm{x}$ ).", + "bbox": [ + 511, + 578, + 905, + 684 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Therefore, DriveGEN introduces shallow feature alignment before the forward process of diffusion, which aims to preserve fine-grained object details by constraining the noisy latent. Specifically, based on the layout $\\mathbf{L}_i$ with $N_{i}$ objects, we calculate the shallow feature alignment loss by:", + "bbox": [ + 511, + 684, + 905, + 760 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\ng _ {s l} = \\frac {1}{N _ {i}} \\sum_ {p, q} \\| \\mathbf {L} _ {i} (p, q) \\odot \\left(\\mathbf {z} _ {t} ^ {I} (p, q) - \\mathbf {z} _ {t} (p, q)\\right) \\| _ {2} ^ {2}, \\tag {4}\n$$\n", + "text_format": "latex", + "bbox": [ + 539, + 779, + 903, + 815 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Overall, the total scheme of DriveGEN is as follows:", + "bbox": [ + 531, + 821, + 880, + 837 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\hat {\\epsilon} _ {t} = (1 + s) \\epsilon_ {\\theta} (\\mathbf {z} _ {t}; t, c) - s \\epsilon_ {\\theta} (\\mathbf {z} _ {t}; t, \\emptyset) + g _ {s a} + g _ {s l}. \\tag {5}\n$$\n", + "text_format": "latex", + "bbox": [ + 529, + 845, + 903, + 864 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Note that $\\varnothing$ denotes null and DriveGEN is a classifier-free guidance [12] process with hyper-parameters $s$ , $\\sigma$ and $\\tau$ .", + "bbox": [ + 511, + 869, + 903, + 901 + ], + "page_idx": 4 + }, + { + "type": "page_number", + "text": "27501", + "bbox": [ + 478, + 944, + 517, + 955 + ], + "page_idx": 4 + }, + { + "type": "table", + "img_path": "images/1482d99c0dbd94b4c53cc3b2d346753d6c4dbfb68004c74c71a327b13c275189.jpg", + "table_caption": [ + "Table 1. Comparison on the KITTI-C dataset, severity level 1 regarding Mean $AP_{3D|R_{40}}$ . The bold number indicates the best result." + ], + "table_footnote": [], + "table_body": "
Car, IoU @ 0.7, 0.5, 0.5
MethodTraining-free diffusionNoiseBlurWeatherDigitalAvg.
Gauss.ShotImpul.Defoc.GlassMotionSnowFrostFogBrit.Contr.PixelSat.
Monoflex [63]13.0620.9114.0920.1728.5930.3433.6430.3119.5845.2220.0129.0738.8526.45
• Color Jitter (Traditional aug.)9.5515.8111.9022.6725.3830.1234.0830.0019.2942.1019.9317.1736.4824.19
• Brightness (Traditional aug.)11.4419.4212.7311.1818.9522.0326.6421.7013.0439.6113.0821.1129.7320.05
• ControlNet (Only Snow aug.)X0.321.181.284.6511.3217.0422.8419.729.5334.798.731.2516.8811.50
• ControlNet (3 scenarios aug.)X0.701.140.290.350.370.834.073.051.209.430.940.483.202.00
• ControlNet (6 scenarios aug.)X0.000.000.000.000.001.940.000.000.000.000.000.000.000.15
• Freecontrol (Only Snow aug.)✓20.3928.0522.2712.5221.6721.0927.9717.9110.0735.999.9028.5635.2722.44
• Freecontrol (3 scenarios aug.)✓15.4321.2514.0415.3319.2216.5322.0117.4115.1824.6616.6625.3931.4419.58
• Freecontrol (6 scenarios aug.)✓12.8518.7116.3112.0616.3112.8617.6416.1814.2324.5215.6422.5624.2817.24
• DriveGEN (Only Snow aug.)✓16.4826.7224.9831.1735.5538.1341.3938.6127.8549.7629.2838.8844.1234.07
• DriveGEN (3 scenarios aug.)✓25.6337.0429.1334.1339.1536.8138.5837.9833.9345.3934.6639.3643.8136.58
• DriveGEN (6 scenarios aug.)✓24.7733.7928.2736.9240.3340.4540.6040.5638.1044.8339.2841.8144.0537.98
MonoGround [31]13.0521.7718.8720.7930.7432.0234.4327.0214.1546.2114.6333.4135.6026.36
• Color Jitter (Traditional aug.)12.8824.3118.9523.0730.4431.4235.9430.4319.8944.6620.6129.7536.6526.36
• Brightness (Traditional aug.)14.0223.5220.1423.9531.7828.7935.0831.8718.8742.9417.7525.5537.1827.03
• ControlNet (Only Snow aug.)X1.763.234.635.2012.9514.1117.7011.583.0435.212.987.2913.9810.28
• ControlNet (3 scenarios aug.)X0.000.000.241.421.681.614.905.400.5717.901.127.796.823.80
• ControlNet (6 scenarios aug.)X0.000.000.001.681.260.351.130.520.444.080.382.221.771.06
• Freecontrol (Only Snow aug.)✓11.7521.8915.7617.7021.4521.6932.0820.6013.5736.0514.0326.7538.3522.43
• Freecontrol (3 scenarios aug.)✓16.3120.7517.6113.1016.8414.8217.7215.8911.8725.2013.5522.5724.4917.75
• Freecontrol (6 scenarios aug.)✓15.2022.5915.3522.0021.1818.9517.6914.8514.8224.0216.9722.9926.1219.44
• DriveGEN (Only Snow aug.)✓17.0726.7823.7832.8937.5239.0640.6134.9125.2946.2127.1238.2544.4533.38
• DriveGEN (3 scenarios aug.)✓19.7931.4427.6336.8440.1039.3539.1736.4228.7945.4229.2942.6044.9935.53
• DriveGEN (6 scenarios aug.)✓23.8432.5930.3438.5741.2040.1938.1638.4032.5343.9534.8044.1045.1337.21
", + "bbox": [ + 133, + 114, + 862, + 398 + ], + "page_idx": 5 + }, + { + "type": "table", + "img_path": "images/61aec76d56408943344bd2f8f355ae78621d9479e165cc697d3f4c1b1437051b.jpg", + "table_caption": [], + "table_footnote": [], + "table_body": "
Pedestrian, IoU @ 0.5, 0.25, 0.25
Monoflex [63]1.163.540.788.0617.7015.2612.729.255.6119.875.351.498.658.42
• Color Jitter (Traditional aug.)0.993.531.4110.0514.6312.0014.7312.137.7219.029.231.1111.299.07
• Brightness (Traditional aug.)0.631.851.205.0113.7212.157.935.841.8716.892.300.483.775.67
• ControlNet (Only Snow aug.)X0.000.000.001.788.325.844.033.751.3111.001.230.000.932.94
• ControlNet (3 scenarios aug.)X0.000.000.001.153.231.141.641.350.834.062.010.000.001.19
• ControlNet (6 scenarios aug.)X0.000.000.000.000.360.000.000.000.002.500.000.000.000.22
• Freecontrol (Only Snow aug.)✓3.625.173.213.235.165.944.683.524.287.704.228.517.705.15
• Freecontrol (3 scenarios aug.)✓2.263.543.043.274.974.257.424.417.2710.916.186.837.435.52
• Freecontrol (6 scenarios aug.)✓5.999.109.084.807.365.847.227.587.1010.819.409.9010.248.03
• DriveGEN (Only Snow aug.)✓1.253.224.5616.2119.8820.6119.8014.468.8024.629.119.7718.0413.10
• DriveGEN (3 scenarios aug.)✓6.099.859.6014.9618.5916.4715.4516.3915.8821.9816.8514.5518.1314.98
• DriveGEN (6 scenarios aug.)✓6.349.227.7117.6019.2120.2016.8816.1915.7223.5716.7514.3717.3115.47
MonoGround [31]2.673.255.7617.5718.9117.7112.969.354.3724.155.893.277.1610.23
• Color Jitter (Traditional aug.)2.443.244.1115.3718.4616.5015.4512.389.1424.719.822.077.8110.89
• Brightness (Traditional aug.)2.794.107.6114.5114.1314.5212.1212.665.5520.805.362.3610.319.75
• ControlNet (Only Snow aug.)X1.851.030.817.189.978.321.322.921.3712.991.400.280.903.87
• ControlNet (3 scenarios aug.)X0.000.000.003.603.891.961.141.031.504.211.250.260.771.51
• ControlNet (6 scenarios aug.)X0.000.000.000.000.001.670.000.000.000.000.000.000.000.13
• Freecontrol (Only Snow aug.)✓10.0413.0911.7014.9112.6113.0315.1412.358.1315.6912.3416.3214.5713.07
• Freecontrol (3 scenarios aug.)✓0.771.811.982.193.083.963.290.522.553.763.104.105.372.81
• Freecontrol (6 scenarios aug.)✓6.817.547.675.545.455.102.191.313.638.595.358.866.265.72
• DriveGEN (Only Snow aug.)✓6.226.899.2715.0317.2418.8116.4013.389.8123.2811.5012.0413.9213.37
• DriveGEN (3 scenarios aug.)✓6.539.2611.1716.1519.6021.0218.6716.1713.7224.8014.5219.9819.3616.23
• DriveGEN (6 scenarios aug.)✓9.7013.6813.5217.0017.7420.3917.2318.7915.5023.0915.8020.9918.3417.06
", + "bbox": [ + 133, + 400, + 862, + 657 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "5. Experiments", + "text_level": 1, + "bbox": [ + 89, + 695, + 223, + 713 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "We conduct experiments to validate DriveGEN on monocular and multi-view 3D object detection, including four different training settings with various augmented OOD scenarios: 1) Augmented with traditional techniques (Traditional aug.); 2) Only augmented with Snow (Only Snow aug.); 3) Augmented with Snow, Rain and Fog (3 scenarios aug.); 4) Augmented with Snow, Rain, Fog, Night, Defocus and Sandstorm (6 scenarios aug.).", + "bbox": [ + 89, + 727, + 482, + 848 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Datasets. For monocular 3d object detection, we split the images of KITTI [10] into a training set (3712 images) and a validation set (3769 images) following the protocol from", + "bbox": [ + 89, + 854, + 483, + 902 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Monoflex [63], encompassing three classes: Car, Pedestrian, and Cyclist. We adopt the KITTI-C dataset [22] for validation, which includes 13 types of data corruptions on the validation set across four categories: Noise, Blur, Weather, and Digital [11].", + "bbox": [ + 511, + 696, + 906, + 773 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "As for multi-view 3D object detection, we first split real-world Night and Rainy validation scenarios based on descriptions of nuScenes [1] following [24]. Next, we augment 500 daytime training scenes under the Snow condition. The augmented detectors are then evaluated on real-world Night and Rainy scenarios, as well as on the widely used Robo3D benchmark [52]. More details of datasets are provided in Appendix B.", + "bbox": [ + 511, + 779, + 908, + 902 + ], + "page_idx": 5 + }, + { + "type": "page_number", + "text": "27502", + "bbox": [ + 478, + 944, + 519, + 957 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/409d28a1f3ac3a5e981ced131940fcf6503911c35a3773f84cb90ffafc2b14da.jpg", + "image_caption": [ + "Figure 4. Based on MonoCD [54], we provide more comparisons with KITTI-C, regarding Mean $AP_{3D|R_{40}}$ ." + ], + "image_footnote": [], + "bbox": [ + 91, + 88, + 906, + 224 + ], + "page_idx": 6 + }, + { + "type": "table", + "img_path": "images/78457a8cb79eb82a9e327e5765cc0b6bf4c0af6fd5c938196aa65b617ed52f03.jpg", + "table_caption": [ + "Table 2. Detection results on nuScenes-C and real-world Night and Rainy scenarios of nuScenes, regarding mAP and NDS." + ], + "table_footnote": [], + "table_body": "
MetricMethodnuScenes-CReal-world Scenarios
BrightnessCameraCrashColorQuantFogFrameLostLowLightMotionBlurSnowAvg.nuScenes-NightnuScenes-RainyAvg.
mAPBEVFormer-tiny24.2622.0123.9121.9820.4316.1120.7810.8320.049.3126.7618.04
• DriveGEN (3k Snow)26.0424.1325.7824.1321.1317.2022.3911.7221.57(↑1.53)12.1929.1120.65 (↑2.61)
NDSBEVFormer-tiny34.9333.0634.5633.0431.4228.4231.8123.2731.3117.9237.3527.64
• DriveGEN (3k Snow)37.1435.3436.9335.3533.3230.4234.2324.0933.35(↑2.04)19.3740.9530.16 (↑2.52)
", + "bbox": [ + 107, + 286, + 888, + 357 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Implementation Details. We implement our method and other baselines in PyTorch [29]. Following Freecontrol [25], we adopt keys from the self-attention of the first U-Net decoder layer as the features. In the self-prototype extraction stage, we run DDIM inversion with 200 steps. Subsequently, we also run 200 steps of DDIM sampling in the prototype-guided diffusion stage. We set $s$ and $\\tau$ to 7.5 and 0.3 following FreeControl [25], while $\\sigma$ is set to 0.1. All 3D detectors are trained on a mix of original and augmented data. All results in the manuscript are based on Stable Diffusion (SD) 1.5 [35]. More details and the results of SD 2.1 and XL 1.0 are put in Appendix C.", + "bbox": [ + 88, + 393, + 480, + 574 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Compared Methods. Based on typical or state-of-the-art (SOTA) 3D detectors [20, 31, 54, 63], we fully compare DriveGEN with following methods: 1) source-only, i.e., directly apply the well-trained model to corrupted test data; 2) Traditional data augmentation, i.e., Color Jitter and Brightness; 3) training-based controllable T2I diffusion: ControlNet [59] trains auxiliary modules for spatial control; 4) training-free controllable T2I diffusion: Freecontrol [25] enables zero-shot control of pretrained diffusion models.", + "bbox": [ + 88, + 577, + 482, + 713 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Evaluation Protocols. We report the experimental results in the Average Precision (AP) for 3D bounding boxes, denoted as $AP_{3D|R_{40}}$ . On the KITTI-C dataset, the results present the mean values of three difficulty levels and Intersection over Union (IoU) thresholds are set to 0.7, 0.5, 0.5 for Cars and 0.5, 0.25, 0.25 for Pedestrians and Cyclists.", + "bbox": [ + 89, + 715, + 482, + 808 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "As for nuScenes, the mean average precision (mAP) and nuScenes detection score (NDS) are calculated following BEVFormer [20]. The metrics also include five true positive metrics, including ATE, ASE, AOE, AVE, and AAE for measuring errors in translation, scale, orientation, velocity, and attributes.", + "bbox": [ + 89, + 810, + 483, + 898 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "5.1. Comparisons with Previous Methods", + "text_level": 1, + "bbox": [ + 511, + 392, + 831, + 407 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "For monocular 3d object detection, Table 1 gives the following observations: 1) Applying the well-trained detector to corrupted scenarios suffers significant performance degradation due to the data distribution shifts and traditional data augmentation offers limited gains. 2) ControlNet (pretrained on ade20k [70]) fails to enhance 3d detectors with limited training data even if it is provided with accurate segmentation masks [34] and rich prompts [6] for fine-tuning. 3) Freecontrol achieves minor gains in a few cases but generally degrades, indicating that training-free T2I diffusion is a feasible solution but requires careful object preservation. 4) DriveGEN consistently outperforms all compared methods over all categories within various base models for both known and unseen scenarios, like improving MonoFlex by an average of $7.6\\mathrm{mAP}$ across 13 various OOD scenarios with the single Snow scenario augmentation.", + "bbox": [ + 509, + 414, + 906, + 655 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "As for multi-view 3D object detection, we randomly select 3k daytime training images and apply Snow augmentation (3k Snow) to enhance the detector [20]. Table 2 shows DriveGEN consistently enhances BEVFormer-tiny across 8 OOD scenarios with an average of $1.53\\mathrm{mAP}$ and 2.04 NDS, further confirming our effectiveness. Moreover, we validate the detector on real Night and Rainy scenarios of nuScenes. With the help of DriveGEN, it also achieves notable gains in real-world tasks with an average improvement of $2.61\\mathrm{mAP}$ and 2.52 NDS, further demonstrating our superiority. More detailed results are put in the Appendix D.", + "bbox": [ + 511, + 656, + 908, + 824 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "5.2. More Augmented Scenarios for Training", + "text_level": 1, + "bbox": [ + 511, + 832, + 861, + 849 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "An intuitive concern is whether the robustness of detectors progressively improves with the increasing of additional augmented scenarios (i.e., Only Snow aug. $\\rightarrow 3$ scenarios", + "bbox": [ + 511, + 854, + 906, + 902 + ], + "page_idx": 6 + }, + { + "type": "page_number", + "text": "27503", + "bbox": [ + 478, + 944, + 517, + 955 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/e9f4557a8a6b85aaaccd5c3dc8551acd280a52881c407ab09a85a1631be846da.jpg", + "image_caption": [ + "Figure 5. Ablation studies on semantic feature alignment loss $g_{sa}$ , self-prototypes $\\hat{\\mathbf{P}}_t$ and shallow feature alignment loss $g_{sl}$ ." + ], + "image_footnote": [], + "bbox": [ + 107, + 88, + 890, + 224 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/e9e976d8ab81f03eca7bcbdf2bdbca6ee1396e6a104efa635b40e427d48b212c.jpg", + "image_caption": [ + "Figure 6. Qualitative results of DriveGEN for multi-view 3D object detection on nuScenes. DriveGEN supports vision-centric 3D detection tasks since our method only requires input images and corresponding object annotations without any additional diffusion model training." + ], + "image_footnote": [], + "bbox": [ + 109, + 268, + 890, + 441 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "aug. $\\rightarrow 6$ scenarios aug.). Based on Table 1 and Figure 4, we are given additional observations: 1) For ControlNet, the model performance drop for OOD scenarios grows as the number of augmentation scenarios increases, demonstrating that severe object omission leads to noisy training data. 2) For Freecontrol, the model performance across various OOD scenarios remains stable or tends to decline, indicating its object preservation may be unstable. 3) Owing to the precise object preservation, DriveGEN obtains continuous overall performance (i.e., Avg.) improvement across all categories with the incorporation of OOD scenarios.", + "bbox": [ + 89, + 510, + 483, + 676 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "5.3. Ablation Studies", + "text_level": 1, + "bbox": [ + 89, + 705, + 254, + 720 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "To examine DriveGEN, we present qualitative results guided by different loss settings. As shown in Figure 5, compared with only generated via text prompts, introducing semantic-aware feature alignment $g_{sa}$ helps preserve more image structures. Next, by guiding the diffusion process with our self-prototypes $\\hat{\\mathbf{P}}_t$ , we achieve more accurate object geometry, alleviating potential object geometry loss. Eventually, introducing our shallow feature alignment $g_{sl}$ further improves object preservation with precise 3D geometry, especially for small objects (highlighted by red circles). More ablation studies are provided in Appendix D.", + "bbox": [ + 89, + 734, + 483, + 902 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "5.4. Quantitative Results", + "text_level": 1, + "bbox": [ + 513, + 508, + 705, + 525 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "We provide visualizations for multi-view 3D object detection on nuScenes in Snow scenarios as shown in Figure 6. It is evident that DriveGEN supports existing vision-centric 3D detection tasks since it requires only input images and corresponding object annotations without any additional diffusion model training, demonstrating that DriveGEN achieves superior results even in challenging multiview tasks. More visualizations are put in Appendix E.", + "bbox": [ + 511, + 531, + 906, + 652 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "6. Conclusion", + "text_level": 1, + "bbox": [ + 513, + 664, + 633, + 679 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "In this paper, we propose a method for robust 3D detection in driving via training-free controllable text-to-image diffusion generation, namely DriveGEN. Specifically, DriveGEN consists of: 1) Self-Prototype Extraction: To improve the guidance of self-attention features, we extract the self-prototypes by layouts to capture accurate object geometry, leveraging more precise features to guide diffusion. 2) Prototype-Guided Diffusion: To further preserve 3D object geometry, we conduct semantic-aware feature and shallow feature alignment during diffusion, alleviating the object misalignment and omission issues compared with previous methods. Experiments on KITTI-C and nuScenes demonstrate the effectiveness of DriveGEN in improving model robustness for vision-centric 3D detection.", + "bbox": [ + 511, + 689, + 906, + 900 + ], + "page_idx": 7 + }, + { + "type": "page_number", + "text": "27504", + "bbox": [ + 478, + 944, + 519, + 957 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Acknowledgements", + "text_level": 1, + "bbox": [ + 91, + 90, + 259, + 107 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "This work was supported by NSFC with Grant No. 62293482, by the Basic Research Project No. HZQB-KCZYZ-2021067 of Hetao Shenzhen HK S&T Cooperation Zone, by Shenzhen General Program No. JCYJ20220530143600001, by Shenzhen-Hong Kong Joint Funding No. SGDX20211123112401002, by the Shenzhen Outstanding Talents Training Fund 202002, by Guangdong Research Project No. 2017ZT07X152 and No. 2019CX01X104, by the Guangdong Provincial Key Laboratory of Future Networks of Intelligence (Grant No. 2022B1212010001), by the Guangdong Provincial Key Laboratory of Big Data Computing, CHUK-Shenzhen, by the NSFC 61931024&12326610, by the Key Area R&D Program of Guangdong Province with grant No. 2018B030338001, by the Shenzhen Key Laboratory of Big Data and Artificial Intelligence (Grant No. ZDSYS201707251409055), by Shaanxi Mathematical Basic Science Research Project(No.23JSY047), and by Tencent & Huawei Open Fund, by China Association for Science and Technology Youth Care Program.", + "bbox": [ + 89, + 114, + 485, + 419 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "References", + "text_level": 1, + "bbox": [ + 91, + 431, + 187, + 448 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[1] Holger Caesar, Varun Bankiti, Alex H Lang, and et.al. nuscene: A multimodal dataset for autonomous driving. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 11621-11631, 2020. 6", + "[2] Changjian Chen, Fei Lv, Yalong Guan, Pengcheng Wang, Shengjie Yu, Yifan Zhang, and Zhuo Tang. Human-guided image generation for expanding small-scale training image datasets. arXiv preprint arXiv:2412.16839, 2024. 3", + "[3] Xiaozhi Chen, Kaustav Kundu, Yukun Zhu, and et. al. 3d object proposals for accurate object class detection. Advances in neural information processing systems, 28, 2015. 2", + "[4] Xiaozhi Chen, Huimin Ma, Ji Wan, Bo Li, and Tian Xia. Multi-view 3d object detection network for autonomous driving. In Proceedings of the IEEE conference on Computer Vision and Pattern Recognition, pages 1907-1915, 2017.", + "[5] Yukang Chen, Jianhui Liu, Xiangyu Zhang, Xiaojuan Qi, and Jiaya Jia. Voxelnext: Fully sparse voxelnet for 3d object detection and tracking. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 21674-21683, 2023. 2", + "[6] Zhe Chen, Jiannan Wu, Wenhai Wang, Weijie Su, Guo Chen, Sen Xing, Muyan Zhong, Qinglong Zhang, Xizhou Zhu, Lewei Lu, et al. Internvl: Scaling up vision foundation models and aligning for generic visual-linguistic tasks. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 24185-24198, 2024. 3, 7", + "[7] Prafulla Dhariwal and Alexander Nichol. Diffusion models beat gans on image synthesis. Advances in neural information processing systems, 34:8780-8794, 2021. 3, 1", + "[8] Dave Epstein, Allan Jabri, Ben Poole, Alexei Efros, and Aleksander Holynski. Diffusion self-guidance for control" + ], + "bbox": [ + 99, + 457, + 482, + 900 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "able image generation. Advances in Neural Information Processing Systems, 36:16222-16239, 2023. 3", + "[9] Songwei Ge, Taesung Park, Jun-Yan Zhu, and Jia-Bin Huang. Expressive text-to-image generation with rich text. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 7545-7556, 2023. 5", + "[10] Andreas Geiger, Philip Lenz, and Raquel Urtasun. Are we ready for autonomous driving? the kitti vision benchmark suite. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 3354-3361. IEEE, 2012. 1, 6", + "[11] Dan Hendrycks and Thomas Dietterich. Benchmarking neural network robustness to common corruptions and perturbations. In International Conference on Learning Representations, 2018. 6", + "[12] Jonathan Ho and Tim Salimans. Classifier-free diffusion guidance. arXiv preprint arXiv:2207.12598, 2022. 5", + "[13] Jonathan Ho, Ajay Jain, and Pieter Abbeel. Denoising diffusion probabilistic models. Advances in neural information processing systems, 33:6840-6851, 2020. 3, 1", + "[14] Yuming Hong, Weiyao Xu, Yi Yang, Hao Zhou, and Xiaoxue Wang. Magicdrive: 3d controllable driving video synthesis using layout and 3d representations. Proceedings of the IEEE/CVF International Conference on Computer Vision (ICCV), 2021. 3", + "[15] Xinyu Huang, Mingyu Chen, Yunqi Li, Hao Wang, and Yuxing Zhao. Diffumask: Synthetic image and semantic mask annotation using cross-attention in diffusion models. arXiv preprint arXiv:2211.09123, 2022. 3", + "[16] Junsoo Kim, Yeongjae Ye, Dongheon Cho, Jaechan Seo, and Jiseob Jeong. Drivingdiffusion: Realistic multi-view video generation with diffusion models. Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), 2021. 3", + "[17] Feng Li, Yuan Zhang, Xinyang Wang, and Meng Liu. Vpd: Vision-based perception diffusion models for human pose estimation and dense visual annotations. arXiv preprint arXiv:2109.05270, 2021. 3", + "[18] Hongyu Li, Pengfei Zhao, Wei Jiang, and Zihan Chen. Sensor fusion with diffusion models for enhanced 3d object detection. arXiv preprint arXiv:2205.02574, 2022. 1", + "[19] Peiliang Li, Xiaozhi Chen, and Shaojie Shen. Stereo r-cnn based 3d object detection for autonomous driving. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 7644-7652, 2019. 2", + "[20] Zhiqi Li, Wenhai Wang, Hongyang Li, Enze Xie, Chonghao Sima, Tong Lu, Yu Qiao, and Jifeng Dai. Bevformer: Learning bird's-eye-view representation from multi-camera images via spatiotemporal transformers. In European conference on computer vision, pages 1-18. Springer, 2022. 7, 2, 4", + "[21] Hongbin Lin, Yifan Zhang, Zhen Qiu, Shuaicheng Niu, Chuang Gan, Yanxia Liu, and Mingkui Tan. Prototype-guided continual adaptation for class-incremental unsupervised domain adaptation. In European Conference on Computer Vision, pages 351-368. Springer, 2022. 2" + ], + "bbox": [ + 516, + 92, + 906, + 900 + ], + "page_idx": 8 + }, + { + "type": "page_number", + "text": "27505", + "bbox": [ + 478, + 944, + 517, + 955 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[22] Hongbin Lin, Yifan Zhang, Shuaicheng Niu, Shuguang Cui, and Zhen Li. Monotta: Fully test-time adaptation for monocular 3d object detection. In European Conference on Computer Vision, pages 96-114. Springer, 2025. 2, 3, 6, 1, 4", + "[23] Xiaoyang Liu, Yikang Wang, Xinyu Zhang, Wei Ma, Chao Wang, and Yun Chen. Multi-modal fusion for 3d object detection in autonomous driving. arXiv preprint arXiv:2204.01234, 2022. 1", + "[24] Zhijian Liu, Haotian Tang, Alexander Amini, Xinyu Yang, Huizi Mao, Daniela L Rus, and Song Han. Bevfusion: Multi-task multi-sensor fusion with unified bird's-eye view representation. In 2023 IEEE international conference on robotics and automation (ICRA), pages 2774-2781. IEEE, 2023. 6, 1", + "[25] Sicheng Mo, Fangzhou Mu, Kuan Heng Lin, Yanli Liu, Bochen Guan, Yin Li, and Bolei Zhou. Freecontrol: Training-free spatial control of any text-to-image diffusion model with any condition. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 7465-7475, 2024. 2, 3, 4, 5, 7", + "[26] Seokhwa Moon, Sanghyuk Hwang, Hyunsu Lee, and Minje Kim. Temporal fusion with attention mechanisms for driving scenarios. Proceedings of the IEEE/CVF International Conference on Computer Vision (ICCV), 2022. 1", + "[27] Chong Mou, Xintao Wang, Liangbin Xie, Yanze Wu, Jian Zhang, Zhongang Qi, and Ying Shan. T2i-adapter: Learning adapters to dig out more controllable ability for text-to-image diffusion models. In Proceedings of the AAAI Conference on Artificial Intelligence, pages 4296–4304, 2024. 2, 4", + "[28] Youngmin Oh, Hyung-II Kim, Seong Tae Kim, and Jung Uk Kim. Monowad: Weather-adaptive diffusion model for robust monocular 3d object detection. In European Conference on Computer Vision, 2025. 2, 3", + "[29] Adam Paszke, Sam Gross, Francisco Massa, Adam Lerer, James Bradbury, Gregory Chanan, Trevor Killeen, Zeming Lin, Natalia Gimelshein, Luca Antiga, et al. Pytorch: An imperative style, high-performance deep learning library. In Advances in neural information processing systems, 2019. 7, 2", + "[30] Yuheng Qin, Zhaoyang Zhang, Han Zhang, Dong Li, Baining Chen, Lu Zhang, Jiuming Gu, and Dongdong Zhang. Uncontrol: A unified diffusion model for controllable visual generation in the wild. arXiv preprint arXiv:2305.11147, 2023. 3", + "[31] Zequn Qin and Xi Li. Monoground: Detecting monocular 3d objects from the ground. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 3793-3802, 2022. 6, 7, 4", + "[32] Zhen Qiu, Yifan Zhang, Hongbin Lin, Shuaicheng Niu, Yanxia Liu, Qing Du, and Mingkui Tan. Source-free domain adaptation via avatar prototype generation and adaptation. arXiv preprint arXiv:2106.15326, 2021. 2", + "[33] Aditya Ramesh, Prafulla Dhariwal, Alex Nichol, Casey Chu, and Mark Chen. Hierarchical text-conditional image generation with clip latents. arXiv preprint arXiv:2204.06125, 2022. 3", + "[34] Nikhila Ravi, Valentin Gabeur, Yuan-Ting Hu, Ronghang Hu, Chaitanya Ryali, Tengyu Ma, Haitham Khedr, Roman" + ], + "bbox": [ + 91, + 90, + 482, + 900 + ], + "page_idx": 9 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "Rädle, Chloe Rolland, Laura Gustafson, et al. Sam 2: Segment anything in images and videos. arXiv preprint arXiv:2408.00714, 2024. 3, 7", + "[35] Robin Rombach, Andreas Blattmann, Dominik Lorenz, Patrick Esser, and Björn Ommer. High-resolution image synthesis with latent diffusion models. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 10684-10695, 2022. 3, 5, 7, 1", + "[36] Olaf Ronneberger, Philipp Fischer, and Thomas Brox. U-net: Convolutional networks for biomedical image segmentation. In Medical image computing and computer-assisted intervention-MICCAI 2015: 18th international conference, Munich, Germany, October 5-9, 2015, proceedings, part III 18, pages 234-241. Springer, 2015. 3, 2", + "[37] Chitwan Sahara, William Chan, Saurabh Saxena, Lala Li, Jay Whang, Emily Denton, Seyed Kamyar Seyed Ghasemipour, Burcu Karagol Ayan, Tim Salimans, Jonathan Ho, et al. Imagen: Photorealistic text-to-image diffusion models. arXiv preprint arXiv:2205.11487, 2022. 3", + "[38] Jascha Sohl-Dickstein, Eric A Weiss, Niru Maheswaranathan, and Surya Ganguli. Deep unsupervised learning using nonequilibrium thermodynamics. Proceedings of the 32nd International Conference on Machine Learning, 2015. 1", + "[39] Jiaming Song, Chenlin Meng, and Stefano Ermon. Denoising diffusion implicit models. arXiv preprint arXiv:2010.02502, 2020. 5", + "[40] Yang Song, Jascha Sohl-Dickstein, Diederik P Kingma, Abhishek Kumar, Stefano Ermon, and Ben Poole. Score-based generative modeling through stochastic differential equations. arXiv preprint arXiv:2011.13456, 2021. 1", + "[41] Yang Song, Jascha Sohl-Dickstein, Diederik P Kingma, Abhishek Kumar, Stefano Ermon, and Ben Poole. Score-based generative modeling through stochastic differential equations. In International Conference on Learning Representations, 2021. 3", + "[42] Ziying Song, Guoxing Zhang, Lin Liu, Lei Yang, Shaoqing Xu, Caiyan Jia, Feiyang Jia, and Li Wang. Robofusion: towards robust multi-modal 3d object detection via sam. In Proceedings of the Thirty-Third International Joint Conference on Artificial Intelligence, pages 1272–1280, 2024. 3", + "[43] Lin Sun, Liangliang Cao, Yang Zheng, Bo Wang, Zhiming Li, Wenzhe Xie, and Xu Sun. Panacea: A framework for diverse, controllable generation of urban scenes. arXiv preprint arXiv:2207.10701, 2022. 3", + "[44] Alexander Swerdlow, Runsheng Xu, and Bolei Zhou. Streetview image generation from a bird's-eye view layout. IEEE Robotics and Automation Letters, 2024. 2", + "[45] Sergey Tulyakov, Ming-Yu Liu, Xiaodong Yang, and Jan Kautz. Mocogan: Decomposing motion and content for video generation. Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition (CVPR), 2018. 1", + "[46] Narek Tumanyan, Michal Geyer, Shai Bagon, and Tali Dekel. Plug-and-play diffusion features for text-driven image-to-image translation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 1921-1930, 2023. 2, 3, 4, 5" + ], + "bbox": [ + 516, + 92, + 906, + 900 + ], + "page_idx": 9 + }, + { + "type": "page_number", + "text": "27506", + "bbox": [ + 478, + 944, + 519, + 955 + ], + "page_idx": 9 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[47] A Vaswani. Attention is all you need. Advances in Neural Information Processing Systems, 2017. 3", + "[48] Pascal Vincent, Hugo Larochelle, Yoshua Bengio, and Pierre-Antoine Manzagol. Extracting and composing robust features with denoising autoencoders. Proceedings of the 25th International Conference on Machine Learning, 2008. 1", + "[49] Yan Wang, Wei-Lun Chao, Divyansh Garg, and et. al. Pseudo-lidar from visual depth estimation: Bridging the gap in 3d object detection for autonomous driving. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 8445-8453, 2019. 2", + "[50] Hai Wu, Chenglu Wen, Shaoshuai Shi, Xin Li, and Cheng Wang. Virtual sparse convolution for multimodal 3d object detection. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 21653-21662, 2023. 2", + "[51] Jinheng Xie, Yuexiang Li, Yawen Huang, Haozhe Liu, Wentian Zhang, Yefeng Zheng, and Mike Zheng Shou. Boxdiff: Text-to-image synthesis with training-free box-constrained diffusion. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 7452-7461, 2023. 3", + "[52] Shaoyuan Xie, Lingdong Kong, Wenwei Zhang, Jiawei Ren, Liang Pan, Kai Chen, and Ziwei Liu. Benchmarking and improving bird's eye view perception robustness in autonomous driving. IEEE Transactions on Pattern Analysis and Machine Intelligence, 2025. 2, 6", + "[53] Junkai Xu, Liang Peng, Haoran Cheng, Hao Li, Wei Qian, Ke Li, Wenxiao Wang, and Deng Cai. Mononerd: Nerflike representations for monocular 3d object detection. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 6814-6824, 2023. 2", + "[54] Longfei Yan, Pei Yan, Shengzhou Xiong, Xuanyu Xiang, and Yihua Tan. Monocd: Monocular 3d object detection with complementary depths. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 10248-10257, 2024. 7", + "[55] Xu Yan, Weiming Zeng, Jiaming Zhu, Hailin Chen, Yaowei Li, and Deng Cai. Videogpt: Video generation using vq-vae and transformers. arXiv preprint arXiv:2104.10157, 2021. 1", + "[56] Chenyu Yang, Yuntao Chen, Hao Tian, Chenxin Tao, Xizhou Zhu, Zhaoxiang Zhang, Gao Huang, Hongyang Li, Yu Qiao, Lewei Lu, et al. Bevformer v2: Adapting modern image backbones to bird's-eye-view recognition via perspective supervision. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 17830-17839, 2023. 2", + "[57] Kairui Yang, Enhui Ma, Jibin Peng, Qing Guo, Di Lin, and Kaicheng Yu. Bevcontrol: Accurately controlling street-view elements with multi-perspective consistency via bev sketch layout. arXiv preprint arXiv:2308.01661, 2023. 2", + "[58] Tianwei Yin, Xingyi Zhou, and Philipp Krahenbuhl. Center-based 3d object detection and tracking. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 11784-11793, 2021. 2", + "[59] Lvmin Zhang, Anyi Rao, and Maneesh Agrawala. Adding conditional control to text-to-image diffusion models. In" + ], + "bbox": [ + 89, + 92, + 483, + 900 + ], + "page_idx": 10 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 3836-3847, 2023. 2, 3, 4, 7", + "[60] Renrui Zhang, Han Qiu, Tai Wang, Ziyu Guo, Ziteng Cui, Yu Qiao, Hongsheng Li, and Peng Gao. Monodetr: Depth-guided transformer for monocular 3d object detection. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 9155-9166, 2023. 2", + "[61] Yifan Zhang and Bryan Hooi. Hipa: enabling one-step text-to-image diffusion models via high-frequency-promoting adaptation. arXiv preprint arXiv:2311.18158, 2023. 3", + "[62] Yifan Zhang, Ying Wei, Qingyao Wu, Peilin Zhao, Shuaicheng Niu, Junzhou Huang, and Mingkui Tan. Collaborative unsupervised domain adaptation for medical image diagnosis. IEEE Transactions on Image Processing, 29: 7834-7844, 2020. 2", + "[63] Yunpeng Zhang, Jiwen Lu, and Jie Zhou. Objects are different: Flexible monocular 3d object detection. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 3289-3298, 2021. 1, 2, 6, 7, 3, 4", + "[64] Yifan Zhang, Bingyi Kang, Bryan Hooi, Shuicheng Yan, and Jiashi Feng. Deep long-tailed learning: A survey. IEEE transactions on pattern analysis and machine intelligence, 45(9):10795-10816, 2023. 2", + "[65] Yifan Zhang, Daquan Zhou, Bryan Hooi, Kai Wang, and Ji-ashi Feng. Expanding small-scale datasets with guided imagination. In Advances in neural information processing systems, pages 76558–76618, 2023. 3", + "[66] Bowen Zhao, Feng Lu, Cheng Zhang, Jiaqi Hu, and Lin Song. Drivedreamer: Realistic multi-view video generation of driving scenes with future prediction. Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition (CVPR), 2022. 3", + "[67] Shihao Zhao, Dongdong Chen, Yen-Chun Chen, Jianmin Bao, Dongdong Zhang, Lu Yuan, Han Zhang, Dong Li, Baining Chen, and Lu Zhang. Uni-controlnet: All-in-one control to text-to-image diffusion models. arXiv preprint arXiv:2305.16322, 2023. 3", + "[68] Guangcong Zheng, Xianpan Zhou, Xuewei Li, Zhongang Qi, Ying Shan, and Xi Li. Layoutdiffusion: Controllable diffusion model for layout-to-image generation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 22490-22499, 2023. 3", + "[69] Longtao Zheng, Yifan Zhang, Hanzhong Guo, Jiachun Pan, Zhenxiong Tan, Jiahao Lu, Chuanxin Tang, Bo An, and Shuicheng Yan. Memo: Memory-guided diffusion for expressive talking video generation. arXiv preprint arXiv:2412.04448, 2024. 3", + "[70] Bolei Zhou, Hang Zhao, Xavier Puig, Sanja Fidler, Adela Barriuso, and Antonio Torralba. Scene parsing through ade20k dataset. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 633-641, 2017. 7" + ], + "bbox": [ + 516, + 92, + 903, + 824 + ], + "page_idx": 10 + }, + { + "type": "page_number", + "text": "27507", + "bbox": [ + 478, + 944, + 517, + 955 + ], + "page_idx": 10 + } +] \ No newline at end of file diff --git a/2025/DriveGEN_ Generalized and Robust 3D Detection in Driving via Controllable Text-to-Image Diffusion Generation/d5fb6699-8753-41df-8e0e-5b82ff7bf5a3_model.json b/2025/DriveGEN_ Generalized and Robust 3D Detection in Driving via Controllable Text-to-Image Diffusion Generation/d5fb6699-8753-41df-8e0e-5b82ff7bf5a3_model.json new file mode 100644 index 0000000000000000000000000000000000000000..9626f5cf5e57bb40bde6b130b8aa29765377b1c6 --- /dev/null +++ b/2025/DriveGEN_ Generalized and Robust 3D Detection in Driving via Controllable Text-to-Image Diffusion Generation/d5fb6699-8753-41df-8e0e-5b82ff7bf5a3_model.json @@ -0,0 +1,2136 @@ +[ + [ + { + "type": "header", + "bbox": [ + 0.107, + 0.003, + 0.182, + 0.044 + ], + "angle": 0, + "content": "CVF" + }, + { + "type": "header", + "bbox": [ + 0.237, + 0.001, + 0.812, + 0.047 + ], + "angle": 0, + "content": "This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore." + }, + { + "type": "title", + "bbox": [ + 0.195, + 0.131, + 0.806, + 0.177 + ], + "angle": 0, + "content": "DriveGEN: Generalized and Robust 3D Detection in Driving via Controllable Text-to-Image Diffusion Generation" + }, + { + "type": "text", + "bbox": [ + 0.233, + 0.203, + 0.76, + 0.24 + ], + "angle": 0, + "content": "Hongbin Lin\\(^{1,2}\\) Zilu Guo\\(^{1,2}\\) Yifan Zhang\\(^{3}\\) Shuaicheng Niu\\(^{4}\\) Yafeng Li\\(^{5}\\) Ruimao Zhang\\(^{6}\\) Shuguang Cui\\(^{2,1}\\) Zhen Li\\(^{2,1*}\\)" + }, + { + "type": "text", + "bbox": [ + 0.102, + 0.24, + 0.895, + 0.276 + ], + "angle": 0, + "content": "\\(^{1}\\)FNii-Shenzhen \\(^{2}\\) SSE, CUHK-Shenzhen \\(^{3}\\) National University of Singapore \\(^{4}\\) Nanyang Technological University \\(^{5}\\)Baoji University of Arts and Sciences \\(^{6}\\)Sun Yat-sen University" + }, + { + "type": "image", + "bbox": [ + 0.139, + 0.287, + 0.86, + 0.584 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.09, + 0.598, + 0.908, + 0.655 + ], + "angle": 0, + "content": "Figure 1. An illustration of DriveGEN for enhancing vision-centric 3D detectors. DriveGEN extends original training images to diverse Out-of-Distribution (OOD) scenarios without additional diffusion model training, preserving all annotated objects. Even with a single augmentation, i.e., 'Snow', the augmented detector [63] is improved not only on the original KITTI [10] val set (Ideal) and already known 'Snow' scenario, but also in unseen 'Fog' and 'Defocus' scenarios, demonstrating a comprehensive model generalizability improvement." + }, + { + "type": "title", + "bbox": [ + 0.249, + 0.688, + 0.327, + 0.703 + ], + "angle": 0, + "content": "Abstract" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.721, + 0.486, + 0.873 + ], + "angle": 0, + "content": "In autonomous driving, vision-centric 3D detection aims to identify 3D objects from images. However, high data collection costs and diverse real-world scenarios limit the scale of training data. Once distribution shifts occur between training and test data, existing methods often suffer from performance degradation, known as Out-of-Distribution (OOD) problems. To address this, controllable Text-to-Image (T2I) diffusion offers a potential solution for training data enhancement, which is required to generate diverse OOD scenarios with precise 3D object geometry. Nevertheless, exist" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.689, + 0.909, + 0.903 + ], + "angle": 0, + "content": "ing controllable T2I approaches are restricted by the limited scale of training data or struggle to preserve all annotated 3D objects. In this paper, we present DriveGEN, a method designed to improve the robustness of 3D detectors in Driving via Training-Free Controllable Text-to-Image Diffusion Generation. Without extra diffusion model training, DriveGEN consistently preserves objects with precise 3D geometry across diverse OOD generations, consisting of 2 stages: 1) Self-Prototype Extraction: We empirically find that self-attention features are semantic-aware but require accurate region selection for 3D objects. Thus, we extract precise object features via layouts to capture 3D object geometry, termed self-prototypes. 2) Prototype-Guided Diffusion: To preserve objects across various OOD scenarios, we per" + }, + { + "type": "page_footnote", + "bbox": [ + 0.11, + 0.888, + 0.237, + 0.901 + ], + "angle": 0, + "content": "*Corresponding author." + }, + { + "type": "page_number", + "bbox": [ + 0.479, + 0.945, + 0.52, + 0.958 + ], + "angle": 0, + "content": "27497" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.089, + 0.092, + 0.482, + 0.152 + ], + "angle": 0, + "content": "form semantic-aware feature alignment and shallow feature alignment during denoising. Extensive experiments demonstrate our effectiveness in improving 3D detection. The code is available at github.com/Hongbin98/DriveGEN." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.181, + 0.223, + 0.197 + ], + "angle": 0, + "content": "1. Introduction" + }, + { + "type": "text", + "bbox": [ + 0.093, + 0.206, + 0.483, + 0.357 + ], + "angle": 0, + "content": "Three-dimensional (3D) object detection intends to identify objects and assess their spatial and dimensional attributes via various sensor inputs, which is widely explored in Autonomous Driving (AD) and robotics [3-5, 19, 49, 50, 58]. To save sensor costs, there is a growing trend towards implementing perception systems by vision-centric 3D detection which relies solely on single or multi-view RGB images along with camera calibration information. Despite challenges in this field, existing methods [53, 56, 60] have demonstrated promising results across various benchmarks." + }, + { + "type": "text", + "bbox": [ + 0.093, + 0.358, + 0.483, + 0.614 + ], + "angle": 0, + "content": "Behind the success, one prerequisite is that test images share the same distribution as training images. However, this assumption is often not held in diverse real-world scenarios [21, 32, 62, 64], especially for perception systems that are required for long durations, like autonomous vehicles navigating through different weather conditions. Once the environment changes, well-trained detectors may suffer performance degradation due to the existence of data distribution shifts between training images and test images. As shown in Figure 1, the original detector performs well in the ideal (sunny) scenario while failing to maintain stable performance in the Snow and Fog scenario due to the out-of-distribution (OOD) issue. Considering the widespread application of 3D detection, significant performance degradation on OOD test data may result in traffic accidents and serious safety risks. Thus, it is essential to address the OOD generalization problem of vision-centric 3D detection." + }, + { + "type": "text", + "bbox": [ + 0.093, + 0.615, + 0.483, + 0.901 + ], + "angle": 0, + "content": "Recently, two attempts sought to solve the distribution shifts in AD by test-time model adaptation [22] or restoring OOD scenarios to ideal conditions with a weather-adaptive diffusion model [28]. However, they require additional consumption during the test phase, leading to additional energy costs while the driving range is crucial in AD. To save test-time computation costs, we explore controllable Text-to-Image (T2I) diffusion generation for training data augmentation, thereby improving the robustness of 3D detectors. Prior training-based controllable T2I approaches such as ControlNet [59] and T2i-adapter [27] offer users fine-grained spatial control but rely on substantial training data to train auxiliary modules, making them impractical for handling OOD scenarios in AD. As for training-free methods, they [25, 46] leverage low-resolution self-attention features to capture image structures for diffusion model guidance. However, these features are relatively coarse for preserving object geometric characteristics of different classes in 3D object detection. As shown in Figure 2, the object" + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.092, + 0.905, + 0.182 + ], + "angle": 0, + "content": "geometry may be lost after multiple rounds of downsampling \\((32\\mathrm{x})\\), leading to severe orientation error, position misalignment and omissions, particularly for tiny objects. Therefore, an eligible training-free controllable T2I diffusion method for training data enhancement in 3D detection should preserve all objects with precise 3D geometry." + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.183, + 0.905, + 0.454 + ], + "angle": 0, + "content": "To address these challenges, we propose a method namely DriveGEN for Generalized and Robust 3D Detection in Driving via Training-Free Controllable Text-to-Image Diffusion GENeration. DriveGEN enhances the model robustness within various OOD scenarios generation, consisting of two stages: 1) Self-Prototype Extraction. To capture accurate object geometric characteristics, we first extract principal components of self-attention features [25, 46] to obtain coarse semantic-aware features. Then, we adopt layouts (i.e., bounding boxes) to re-weight features in object regions via a peak function, termed self-prototypes. 2) Prototype-Guided Diffusion. To preserve objects with precise 3D geometry, we perform semantic-aware feature alignment with self-prototypes during the denoising phase. However, these semantic-aware features are relatively coarse to represent tiny objects as shown in Figure 2. Therefore, we devise a shallow feature alignment strategy to constrain the diffusion process and retrain tiny objects." + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.455, + 0.905, + 0.68 + ], + "angle": 0, + "content": "Contributions: 1) To the best of our knowledge, we are the first to explore training-free controllable T2I diffusion generation to enhance the model robustness of vision-centric 3D detection. Even with a single augmentation, DriveGEN improves the 3D detector [63] by an average of \\(7.6\\mathrm{mAP}\\) across 13 OOD scenarios. 2) With self-prototype extraction and prototype-guided diffusion stages, DriveGEN enables OOD scenario augmentation while preserving precise 3D geometry for all objects, enhancing the applicability of diffusion models in autonomous driving. 3) Extensive experiments show DriveGEN significantly improves existing monocular 3D detectors across 13 OOD corruption on KITTI and multi-view 3D detectors on the real scenarios (Night, Rainy) of nuScenes, demonstrating the effectiveness of DriveGEN in boosting model generalizability." + }, + { + "type": "title", + "bbox": [ + 0.514, + 0.695, + 0.655, + 0.71 + ], + "angle": 0, + "content": "2. Related Work" + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.72, + 0.905, + 0.78 + ], + "angle": 0, + "content": "We first review the literature on explorations of model robustness in autonomous driving, and then discuss controllable T2I diffusion generation methods. More discussions on vision-centric 3D object detection are in Appendix A." + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.781, + 0.905, + 0.901 + ], + "angle": 0, + "content": "Model Robustness in Autonomous Driving. The reliability of vision-centric autonomous driving systems depends on the robustness of perception models, especially under challenging scenarios like adverse weather, varied lighting, and complex urban environments [52]. To improve robustness, existing methods employ diffusion models to generate synthetic driving scenes for training data augmentation. Specifically, BEVGen [44] and BEVControl [57] employ" + }, + { + "type": "page_number", + "bbox": [ + 0.48, + 0.945, + 0.518, + 0.956 + ], + "angle": 0, + "content": "27498" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.095, + 0.09, + 0.904, + 0.181 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.09, + 0.192, + 0.908, + 0.25 + ], + "angle": 0, + "content": "Figure 2. Illustration of the importance of object preservation in 3d detection based on KITTI. ControlNet [59] suffers object orientation errors and omissions even if accurate segmentation masks [34] and rich prompts [6] are provided, showing that training-based methods may struggle with spatial control with limited training data. Additionally, Freecontrol [25] relies on coarse low-resolution features to capture semantic structures that may encounter potential object geometry loss, resulting in object position misalignment and omission issues." + }, + { + "type": "text", + "bbox": [ + 0.089, + 0.275, + 0.482, + 0.381 + ], + "angle": 0, + "content": "Layouts to create flexible scene configurations. Innovative frameworks like Panacea [43] and MagicDrive [14] extend this controllability to multi-view scenes, while models such as DriveDreamer [66] and DrivingDiffusion [16] focus on generating realistic videos for model validation. Recently, RoboFusion [42] has sought to tackle OOD noise scenarios by leveraging visual foundation models like SAM." + }, + { + "type": "text", + "bbox": [ + 0.093, + 0.381, + 0.485, + 0.667 + ], + "angle": 0, + "content": "Controllable T2I Diffusion. Controllable text-to-image (T2I) diffusion models [35, 61, 69] are increasingly applied in real-world scenarios, where text prompts or scene descriptions guide the creation of synthetic data. By leveraging pre-trained models like Stable Diffusion [35] or other large-scale T2I models [33, 37], existing methods can generate high-fidelity and diverse images based on textual input, allowing for controlled variations [2, 65]. Techniques such as Uni-ControlNet [67], UniControl [30] and Layout-diffusion [68] further enhance control by integrating multimodal inputs, creating contextually accurate and condition-specific scenes. Recent works also explore the utility of cross-attention mechanisms and latent embeddings to align generated images closely with text descriptions, supporting the generation for perception tasks like semantic segmentation and depth estimation [15, 17]. However, these methods require a large amount of training data to train auxiliary modules, making them impractical for handling OOD scenarios in AD for the limited data scale." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.668, + 0.483, + 0.82 + ], + "angle": 0, + "content": "Another alternative choice is training-free controllable T2I diffusion, PnP [46] directly injects the self-attention features of condition images to guide the diffusion process while Freecontrol [25] utilizes PCA to extract principle components of self-attention features for semantic structures extraction. However, existing approaches solely rely on the relatively coarse self-attention features, which often results in object region omissions and misalignment as we mentioned before. To this end, we propose our DriveGEN to preserve all objects with precise 3D geometry." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.83, + 0.218, + 0.848 + ], + "angle": 0, + "content": "3. Preliminary" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.855, + 0.483, + 0.903 + ], + "angle": 0, + "content": "Pre-trained diffusion models [7, 13, 35] are probabilistic generative models, which consists of two complementary stochastic processes: 1) In the forward process, Gaussian" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.275, + 0.907, + 0.395 + ], + "angle": 0, + "content": "noise is incrementally added to a clean image \\(\\mathbf{x}_0\\). 2) As for the backward process, it conducts the iterative process of denoising the initial Gaussian noise image \\(\\mathbf{z}_{N_t}\\), where a cleaner image is given at each step. To be specific, this process leverages a denoising network \\(\\epsilon_{\\theta}(\\mathbf{z}_t,t)\\) to estimate the added noise in the forward process. Subsequently, for \\(\\mathbf{z}_t\\) at each step, the network removes the estimated noise perturbation to achieve a cleaner \\(\\mathbf{z}_{t - 1}\\)." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.396, + 0.909, + 0.518 + ], + "angle": 0, + "content": "For T2I diffusion models, a previous method [41] has shown that \\(\\epsilon_{\\theta}\\) can approximate the score function for the marginal distributions \\(p_t\\), i.e., \\(-\\sigma_t\\nabla_{\\mathbf{z}_t}\\log p_t(\\mathbf{z}_t|\\mathbf{c})\\) where \\(\\sigma_t\\) denotes a noise schedule and \\(\\mathbf{c}\\) is a text prompt. To incorporate auxiliary information \\(y\\) into the diffusion sampling process, existing methods [7, 8, 25] guide the diffusion process by adding a time-dependent energy function termed \\(g(\\mathbf{z}_t;t,y)\\) with a strength coefficient \\(s\\):" + }, + { + "type": "equation", + "bbox": [ + 0.578, + 0.53, + 0.842, + 0.546 + ], + "angle": 0, + "content": "\\[\n\\hat {\\epsilon} _ {\\boldsymbol {\\theta}} (\\mathbf {z} _ {t}; t, c) = \\epsilon_ {\\boldsymbol {\\theta}} (\\mathbf {z} _ {t}; t, c) - s \\cdot g (\\mathbf {z} _ {t}; t, y),\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.557, + 0.907, + 0.618 + ], + "angle": 0, + "content": "where \\( g \\) can be defined via bounding boxes [51]. In this paper, we adopt a standard choice \\( \\epsilon_{\\theta} \\) following existing methods [25, 46], i.e., U-Net architecture [36] with multi-level self-attention and cross-attention layers [47]." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.632, + 0.883, + 0.649 + ], + "angle": 0, + "content": "4. Training-Free Controllable T2I Diffusion" + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.667, + 0.698, + 0.683 + ], + "angle": 0, + "content": "4.1. Problem Statement" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.689, + 0.907, + 0.856 + ], + "angle": 0, + "content": "Without loss of generality, we denote the well-trained 3D detector as \\( f_{\\Theta_d}(\\cdot) \\) which is obtained by training on labeled training images \\( \\mathcal{D} = \\{(\\mathbf{x}_i,\\mathbf{y}_i)\\}_{i = 1}^N \\). The training images follow the original training distribution \\( P(\\mathbf{x}) \\) (i.e., \\( \\mathbf{x} \\sim P(\\mathbf{x}) \\)). Here, \\( \\Theta_d \\) represents the parameters of the detector and \\( N \\) is the total number of training images. After training, the model is applied to unlabeled test images \\( \\mathcal{D}_t = \\{\\mathbf{x}^t\\} \\). Unfortunately, the well-trained model often encounters Out-of-distribution (OOD) test scenarios in autonomous driving (AD) due to prevalent natural corruptions, namely distribution shifts, i.e., \\( \\mathbf{x}^t \\sim Q(\\mathbf{x}) \\) and \\( P(\\mathbf{x}) \\neq Q(\\mathbf{x}) \\)." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.856, + 0.907, + 0.901 + ], + "angle": 0, + "content": "As previously discussed, existing methods [22, 28] are relatively time-consuming due to their additional computation costs at test time. Hence, we investigate controllable" + }, + { + "type": "page_number", + "bbox": [ + 0.479, + 0.945, + 0.521, + 0.958 + ], + "angle": 0, + "content": "27499" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.141, + 0.097, + 0.862, + 0.291 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.141, + 0.296, + 0.861, + 0.464 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.089, + 0.479, + 0.908, + 0.536 + ], + "angle": 0, + "content": "Figure 3. An overview of our DriveGEN, consists of two stages: 1) The Self-Prototype Extraction stage is devised to extract accurate semantic structures of multiple objects. To capture precise locations, we achieve fine-grained self-prototypes via leveraging layouts and the peak function to re-weight object regions rather than directly using coarse self-attention features. 2) The Prototype-Guided Diffusion stage conducts semantic-aware feature alignment for semantic structure matching and shallow feature alignment for tiny object preservation." + }, + { + "type": "text", + "bbox": [ + 0.089, + 0.561, + 0.483, + 0.713 + ], + "angle": 0, + "content": "Text-to-Image (T2I) diffusion generation for training data augmentation, thereby improving the robustness of 3D detectors. Particularly, it is crucial to preserve object information because the annotations would be reused during detector training. However, training-based approaches [27, 59] require substantial training data to train auxiliary modules for spatial control, which is often unsatisfying in AD. Besides, training-free methods [25, 46] mainly consider the relatively coarse self-attention features, making it challenging to preserve relatively tiny objects." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.726, + 0.249, + 0.74 + ], + "angle": 0, + "content": "4.2. Overall Scheme" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.749, + 0.483, + 0.854 + ], + "angle": 0, + "content": "After examining the characteristics and challenges of robust 3D object detection, we introduce DriveGEN, a Training-Free Controllable T2I Diffusion Generation method devised to improve the robustness of vision-centric 3D detectors. As illustrated in Figure 3, DriveGEN consists of two stages: 1) Self-Prototype Extraction and 2) Prototype-Guided Diffusion. We briefly introduce each stage below." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.856, + 0.484, + 0.903 + ], + "angle": 0, + "content": "First, we develop the self-prototype extraction stage (c.f. Section 4.3) to capture the geometric characteristics of objects accurately. Existing methods apply principal compo" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.561, + 0.907, + 0.729 + ], + "angle": 0, + "content": "nent analysis (PCA) on the self-attention features of the first decoder layer of the diffusion model \\(\\epsilon_{\\theta}\\) to capture semantic structures of condition images [25, 46]. However, we empirically find that solely relying on cross-attention relevance between self-attention features and object prompts for region selection may result in object preservation errors (c.f. Figure 2). To address this, DriveGEN leverages layouts to re-weight object regions in the principal components of self-attention features via a peak function and then achieves the self-prototypes which help the diffusion model focus on the accurate object regions during generation." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.735, + 0.909, + 0.903 + ], + "angle": 0, + "content": "Second, we introduce the prototype-guided diffusion stage (c.f. Section 4.4) to preserve all objects during the generation process. On the one hand, we conduct semantic-aware feature alignment between self-attention features and self-prototypes to capture accurate object semantic structures. On the other hand, DriveGEN performs shallow feature alignment to further preserve object information by constraining the noisy latent at the shallow level, compensating for the gaps in semantic-aware feature alignment regarding tiny object preservation. The Pseudo-code of DriveGEN is summarized in Algorithm 1." + }, + { + "type": "page_number", + "bbox": [ + 0.479, + 0.945, + 0.521, + 0.958 + ], + "angle": 0, + "content": "27500" + } + ], + [ + { + "type": "title", + "bbox": [ + 0.091, + 0.091, + 0.325, + 0.108 + ], + "angle": 0, + "content": "4.3. Self-Prototype Extraction" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.113, + 0.483, + 0.174 + ], + "angle": 0, + "content": "To accurately capture the geometric characteristics of multiple objects, we propose a self-prototype extraction strategy that consists of two components: 1) Image structure extraction and 2) Object region re-weighting." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.174, + 0.483, + 0.417 + ], + "angle": 0, + "content": "Image Structure Extraction. To achieve zero-shot spatial control, capturing the semantic structures of input images is essential. Prior works [25, 46] demonstrate that self-attention features (e.g., keys) of the T2I diffusion model \\(\\epsilon_{\\theta}\\) are semantic-aware [35]. Thus, given the input image \\(\\mathbf{x}_i\\), we first generate the text prompt \\(\\mathbf{c}_i\\) via a simple template 'A photo of \\(\\{\\text{object}\\} ,\\ldots,\\) and \\(\\{\\text{object}\\}'\\). Then, we leverage DDIM Inversion [39] to achieve a series of time-dependent diffusion features \\(\\{\\mathbf{z}_t^I\\in \\mathbb{R}^{N_c\\times H\\times W}\\}_{t = 1}^{N_t}\\) where \\(N_{t}\\) denotes the number of DDIM inversion and sampling steps, \\(N_{c}\\) represents the number of feature channels, and \\(H,W\\) indicate the height and width, respectively. During sampling, we apply PCA on self-attention features to extract principal components \\(\\{\\mathbf{P}_t\\}_{t = 1}^{N_p}\\), each \\(\\mathbf{P}_t\\in \\mathbb{R}^{N_b\\times H\\times W}\\) and \\(\\mathbf{P}_t = \\{\\mathbf{p}_t^1,\\dots,\\mathbf{p}_t^{N_b}\\}\\). Here \\(N_{b}\\) denotes the number of principal components and we store \\(\\mathbf{P}_t\\) of the initial \\(N_{p}\\) steps." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.417, + 0.483, + 0.599 + ], + "angle": 0, + "content": "Object Region Re-weighting. Existing methods [9, 25] select foreground regions by leveraging cross-attention maps between specific words (i.e., object names) and image features to derive a concept region mask \\(\\mathbf{M}\\). However, relying solely on such attention maps may result in omissions and misalignment of object regions, as illustrated in Figure 2. To address this, we introduce a peak function \\(m(\\cdot)\\) to reweight \\(\\mathbf{M}\\) for each object region based on layouts \\(\\mathbf{L}\\) after the coordinate transformation through downsampling. To be specific, given a point \\((p,q)\\) from \\(\\mathbf{L}_i\\) of \\(\\mathbf{x}_i\\), the corresponding object center is defined as \\((\\hat{p},\\hat{q})\\) and then we can calculate \\(m(p,q)\\) by:" + }, + { + "type": "equation", + "bbox": [ + 0.147, + 0.606, + 0.483, + 0.642 + ], + "angle": 0, + "content": "\\[\nm (p, q) = \\exp \\left(- \\frac {(p - \\hat {p}) ^ {2} + (q - \\hat {q}) ^ {2}}{2 \\sigma^ {2}}\\right) \\tag {1}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.648, + 0.483, + 0.679 + ], + "angle": 0, + "content": "where \\(\\sigma \\in (0,1)\\) controls the range of the influence area. Then, we update the concept token mask \\(\\mathbf{M}\\) by:" + }, + { + "type": "equation", + "bbox": [ + 0.15, + 0.688, + 0.483, + 0.706 + ], + "angle": 0, + "content": "\\[\n\\tilde {\\mathbf {M}} (p, q) = \\mathbf {M} (p, q) + m (p, q) \\cdot \\mathbf {L} _ {i} (p, q). \\tag {2}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.713, + 0.483, + 0.794 + ], + "angle": 0, + "content": "The object regions in \\(\\mathbf{L}_i\\) are marked as 1 while all other regions as 0. Eventually, we obtain the self-prototypes by \\(\\{\\hat{\\mathbf{P}}_t\\}_{t=1}^{N_p} = \\{\\mathbf{P}_t\\}_{t=1}^{N_p}\\big|_{\\hat{\\mathbf{M}}}\\) which guide the diffusion model to localize precise locations of objects and combine \\(\\hat{\\mathbf{M}}\\) for accurate object region selection." + }, + { + "type": "title", + "bbox": [ + 0.09, + 0.802, + 0.342, + 0.818 + ], + "angle": 0, + "content": "4.4. Prototype-Guided Diffusion" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.824, + 0.483, + 0.902 + ], + "angle": 0, + "content": "To preserve all objects with precise 3D geometry during diffusion, we leverage the self-prototypes \\(\\{\\hat{\\mathbf{P}}_t\\}_{t=1}^{N_p}\\) and time-dependent diffusion features \\(\\{\\mathbf{z}_t^I\\}_{t=1}^{N_t}\\) for feature alignment, including: 1) semantic-aware feature alignment and 2) shallow feature alignment." + }, + { + "type": "code_caption", + "bbox": [ + 0.517, + 0.092, + 0.869, + 0.107 + ], + "angle": 0, + "content": "Algorithm 1 The pipeline of the proposed DriveGEN" + }, + { + "type": "algorithm", + "bbox": [ + 0.517, + 0.111, + 0.905, + 0.354 + ], + "angle": 0, + "content": "Require: Training data \\(\\{(\\mathbf{x}_i^s,\\mathbf{y}_i^s)\\}_{i = 1}^N\\) Pretrained T2I diffusion model \\(\\epsilon_{\\theta}\\) Hyper-parameters \\(s,\\sigma ,\\tau\\) The OOD scenario. \n#Stage 1: Self-Prototype Extraction \n1: for each training image \\(x_{i}\\) do \n2: Get the text prompt \\(c_{i}\\) and layout \\(\\mathbf{L}_i\\) based on \\(\\mathbf{y}_i\\) \n3: Undergo DDIM Inversion to get \\(\\mathbf{P}_t\\) \n4: Get updated M via Eqn. (1), Eqn. (2) for self-prototypes \\(\\hat{\\mathbf{P}}_t\\) \n5: end for \n#Stage 2: Prototype-Guided Diffusion \n6: for each training image \\(\\mathbf{x}_i\\) do \n7: Update \\(\\mathbf{c}_i\\) with specific OOD type; \n8: for diffusion step \\(t = 1\\rightarrow N_t\\) do \n9: if \\(t\\leq N_p\\) then \n10: Calculate \\(g_{sa},g_{sl}\\) based on Eqn. (3), Eqn. (4); \n11: Get the gradient based on Eqn.(5); \n12: end if \n13: Update the noisy latent \\(\\mathbf{z}_t\\to \\mathbf{z}_{t - 1}\\) \n14: end for \n15: end for \n16: return Output images for all \\(\\mathbf{x}_i\\) in the OOD scenario." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.379, + 0.907, + 0.517 + ], + "angle": 0, + "content": "Semantic-aware Feature Alignment. As mentioned before, the self-attention features of the first decoder layer can represent the semantic structures of input images. Thus, we first initialize the noisy latent \\(\\mathbf{z}_{N_t}\\) via the \\(\\mathbf{z}_{N_t}^I\\). For the step \\(t \\leq N_p\\), we perform PCA to obtain the principal components \\(\\mathbf{P}_t\\) of the self-attention features. Then, we adopt a pre-defined threshold \\(\\tau\\) following [9] to select informative regions by \\(\\hat{\\mathbf{M}} = \\mathbb{I}(\\hat{\\mathbf{M}} > \\tau)\\). Specifically, we calculate the semantic-aware feature alignment loss with \\(\\mathbf{P}_t\\) and \\(\\hat{\\mathbf{P}}_t\\):" + }, + { + "type": "equation", + "bbox": [ + 0.617, + 0.523, + 0.905, + 0.543 + ], + "angle": 0, + "content": "\\[\ng _ {s a} = \\| \\hat {\\mathbf {M}} \\odot (\\mathbf {P} _ {t} - \\hat {\\mathbf {P}} _ {t}) \\| _ {2} ^ {2}, \\tag {3}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.549, + 0.905, + 0.579 + ], + "angle": 0, + "content": "where \\(\\odot\\) represents element-wise multiplication, applying the mask \\(\\hat{\\mathbf{M}}\\) to the difference between \\(\\mathbf{P}_t\\) and \\(\\hat{\\mathbf{P}}_t\\)." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.579, + 0.906, + 0.685 + ], + "angle": 0, + "content": "Shallow Feature Alignment. As shown in Figure 2, we empirically find that principal components lack sufficient details to represent objects, especially tiny objects. For instance, given a bounding box of a cyclist with a height of 20 pixels and a width of 5 pixels, this object is unable to occupy even one element within principal components after multiple rounds of downsampling (e.g., \\(32\\mathrm{x}\\))." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.685, + 0.906, + 0.761 + ], + "angle": 0, + "content": "Therefore, DriveGEN introduces shallow feature alignment before the forward process of diffusion, which aims to preserve fine-grained object details by constraining the noisy latent. Specifically, based on the layout \\(\\mathbf{L}_i\\) with \\(N_{i}\\) objects, we calculate the shallow feature alignment loss by:" + }, + { + "type": "equation", + "bbox": [ + 0.54, + 0.78, + 0.905, + 0.816 + ], + "angle": 0, + "content": "\\[\ng _ {s l} = \\frac {1}{N _ {i}} \\sum_ {p, q} \\| \\mathbf {L} _ {i} (p, q) \\odot \\left(\\mathbf {z} _ {t} ^ {I} (p, q) - \\mathbf {z} _ {t} (p, q)\\right) \\| _ {2} ^ {2}, \\tag {4}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.532, + 0.823, + 0.882, + 0.838 + ], + "angle": 0, + "content": "Overall, the total scheme of DriveGEN is as follows:" + }, + { + "type": "equation", + "bbox": [ + 0.531, + 0.846, + 0.905, + 0.865 + ], + "angle": 0, + "content": "\\[\n\\hat {\\epsilon} _ {t} = (1 + s) \\epsilon_ {\\theta} (\\mathbf {z} _ {t}; t, c) - s \\epsilon_ {\\theta} (\\mathbf {z} _ {t}; t, \\emptyset) + g _ {s a} + g _ {s l}. \\tag {5}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.871, + 0.905, + 0.902 + ], + "angle": 0, + "content": "Note that \\(\\varnothing\\) denotes null and DriveGEN is a classifier-free guidance [12] process with hyper-parameters \\(s\\), \\(\\sigma\\) and \\(\\tau\\)." + }, + { + "type": "page_number", + "bbox": [ + 0.479, + 0.945, + 0.518, + 0.957 + ], + "angle": 0, + "content": "27501" + } + ], + [ + { + "type": "table_caption", + "bbox": [ + 0.105, + 0.089, + 0.891, + 0.105 + ], + "angle": 0, + "content": "Table 1. Comparison on the KITTI-C dataset, severity level 1 regarding Mean \\(AP_{3D|R_{40}}\\). The bold number indicates the best result." + }, + { + "type": "table", + "bbox": [ + 0.134, + 0.116, + 0.864, + 0.399 + ], + "angle": 0, + "content": "
Car, IoU @ 0.7, 0.5, 0.5
MethodTraining-free diffusionNoiseBlurWeatherDigitalAvg.
Gauss.ShotImpul.Defoc.GlassMotionSnowFrostFogBrit.Contr.PixelSat.
Monoflex [63]13.0620.9114.0920.1728.5930.3433.6430.3119.5845.2220.0129.0738.8526.45
• Color Jitter (Traditional aug.)9.5515.8111.9022.6725.3830.1234.0830.0019.2942.1019.9317.1736.4824.19
• Brightness (Traditional aug.)11.4419.4212.7311.1818.9522.0326.6421.7013.0439.6113.0821.1129.7320.05
• ControlNet (Only Snow aug.)X0.321.181.284.6511.3217.0422.8419.729.5334.798.731.2516.8811.50
• ControlNet (3 scenarios aug.)X0.701.140.290.350.370.834.073.051.209.430.940.483.202.00
• ControlNet (6 scenarios aug.)X0.000.000.000.000.001.940.000.000.000.000.000.000.000.15
• Freecontrol (Only Snow aug.)✓20.3928.0522.2712.5221.6721.0927.9717.9110.0735.999.9028.5635.2722.44
• Freecontrol (3 scenarios aug.)✓15.4321.2514.0415.3319.2216.5322.0117.4115.1824.6616.6625.3931.4419.58
• Freecontrol (6 scenarios aug.)✓12.8518.7116.3112.0616.3112.8617.6416.1814.2324.5215.6422.5624.2817.24
• DriveGEN (Only Snow aug.)✓16.4826.7224.9831.1735.5538.1341.3938.6127.8549.7629.2838.8844.1234.07
• DriveGEN (3 scenarios aug.)✓25.6337.0429.1334.1339.1536.8138.5837.9833.9345.3934.6639.3643.8136.58
• DriveGEN (6 scenarios aug.)✓24.7733.7928.2736.9240.3340.4540.6040.5638.1044.8339.2841.8144.0537.98
MonoGround [31]13.0521.7718.8720.7930.7432.0234.4327.0214.1546.2114.6333.4135.6026.36
• Color Jitter (Traditional aug.)12.8824.3118.9523.0730.4431.4235.9430.4319.8944.6620.6129.7536.6526.36
• Brightness (Traditional aug.)14.0223.5220.1423.9531.7828.7935.0831.8718.8742.9417.7525.5537.1827.03
• ControlNet (Only Snow aug.)X1.763.234.635.2012.9514.1117.7011.583.0435.212.987.2913.9810.28
• ControlNet (3 scenarios aug.)X0.000.000.241.421.681.614.905.400.5717.901.127.796.823.80
• ControlNet (6 scenarios aug.)X0.000.000.001.681.260.351.130.520.444.080.382.221.771.06
• Freecontrol (Only Snow aug.)✓11.7521.8915.7617.7021.4521.6932.0820.6013.5736.0514.0326.7538.3522.43
• Freecontrol (3 scenarios aug.)✓16.3120.7517.6113.1016.8414.8217.7215.8911.8725.2013.5522.5724.4917.75
• Freecontrol (6 scenarios aug.)✓15.2022.5915.3522.0021.1818.9517.6914.8514.8224.0216.9722.9926.1219.44
• DriveGEN (Only Snow aug.)✓17.0726.7823.7832.8937.5239.0640.6134.9125.2946.2127.1238.2544.4533.38
• DriveGEN (3 scenarios aug.)✓19.7931.4427.6336.8440.1039.3539.1736.4228.7945.4229.2942.6044.9935.53
• DriveGEN (6 scenarios aug.)✓23.8432.5930.3438.5741.2040.1938.1638.4032.5343.9534.8044.1045.1337.21
" + }, + { + "type": "table", + "bbox": [ + 0.134, + 0.401, + 0.864, + 0.658 + ], + "angle": 0, + "content": "
Pedestrian, IoU @ 0.5, 0.25, 0.25
Monoflex [63]1.163.540.788.0617.7015.2612.729.255.6119.875.351.498.658.42
• Color Jitter (Traditional aug.)0.993.531.4110.0514.6312.0014.7312.137.7219.029.231.1111.299.07
• Brightness (Traditional aug.)0.631.851.205.0113.7212.157.935.841.8716.892.300.483.775.67
• ControlNet (Only Snow aug.)X0.000.000.001.788.325.844.033.751.3111.001.230.000.932.94
• ControlNet (3 scenarios aug.)X0.000.000.001.153.231.141.641.350.834.062.010.000.001.19
• ControlNet (6 scenarios aug.)X0.000.000.000.000.360.000.000.000.002.500.000.000.000.22
• Freecontrol (Only Snow aug.)✓3.625.173.213.235.165.944.683.524.287.704.228.517.705.15
• Freecontrol (3 scenarios aug.)✓2.263.543.043.274.974.257.424.417.2710.916.186.837.435.52
• Freecontrol (6 scenarios aug.)✓5.999.109.084.807.365.847.227.587.1010.819.409.9010.248.03
• DriveGEN (Only Snow aug.)✓1.253.224.5616.2119.8820.6119.8014.468.8024.629.119.7718.0413.10
• DriveGEN (3 scenarios aug.)✓6.099.859.6014.9618.5916.4715.4516.3915.8821.9816.8514.5518.1314.98
• DriveGEN (6 scenarios aug.)✓6.349.227.7117.6019.2120.2016.8816.1915.7223.5716.7514.3717.3115.47
MonoGround [31]2.673.255.7617.5718.9117.7112.969.354.3724.155.893.277.1610.23
• Color Jitter (Traditional aug.)2.443.244.1115.3718.4616.5015.4512.389.1424.719.822.077.8110.89
• Brightness (Traditional aug.)2.794.107.6114.5114.1314.5212.1212.665.5520.805.362.3610.319.75
• ControlNet (Only Snow aug.)X1.851.030.817.189.978.321.322.921.3712.991.400.280.903.87
• ControlNet (3 scenarios aug.)X0.000.000.003.603.891.961.141.031.504.211.250.260.771.51
• ControlNet (6 scenarios aug.)X0.000.000.000.000.001.670.000.000.000.000.000.000.000.13
• Freecontrol (Only Snow aug.)✓10.0413.0911.7014.9112.6113.0315.1412.358.1315.6912.3416.3214.5713.07
• Freecontrol (3 scenarios aug.)✓0.771.811.982.193.083.963.290.522.553.763.104.105.372.81
• Freecontrol (6 scenarios aug.)✓6.817.547.675.545.455.102.191.313.638.595.358.866.265.72
• DriveGEN (Only Snow aug.)✓6.226.899.2715.0317.2418.8116.4013.389.8123.2811.5012.0413.9213.37
• DriveGEN (3 scenarios aug.)✓6.539.2611.1716.1519.6021.0218.6716.1713.7224.8014.5219.9819.3616.23
• DriveGEN (6 scenarios aug.)✓9.7013.6813.5217.0017.7420.3917.2318.7915.5023.0915.8020.9918.3417.06
" + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.696, + 0.224, + 0.714 + ], + "angle": 0, + "content": "5. Experiments" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.728, + 0.483, + 0.849 + ], + "angle": 0, + "content": "We conduct experiments to validate DriveGEN on monocular and multi-view 3D object detection, including four different training settings with various augmented OOD scenarios: 1) Augmented with traditional techniques (Traditional aug.); 2) Only augmented with Snow (Only Snow aug.); 3) Augmented with Snow, Rain and Fog (3 scenarios aug.); 4) Augmented with Snow, Rain, Fog, Night, Defocus and Sandstorm (6 scenarios aug.)." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.856, + 0.484, + 0.903 + ], + "angle": 0, + "content": "Datasets. For monocular 3d object detection, we split the images of KITTI [10] into a training set (3712 images) and a validation set (3769 images) following the protocol from" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.697, + 0.907, + 0.775 + ], + "angle": 0, + "content": "Monoflex [63], encompassing three classes: Car, Pedestrian, and Cyclist. We adopt the KITTI-C dataset [22] for validation, which includes 13 types of data corruptions on the validation set across four categories: Noise, Blur, Weather, and Digital [11]." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.78, + 0.909, + 0.903 + ], + "angle": 0, + "content": "As for multi-view 3D object detection, we first split real-world Night and Rainy validation scenarios based on descriptions of nuScenes [1] following [24]. Next, we augment 500 daytime training scenes under the Snow condition. The augmented detectors are then evaluated on real-world Night and Rainy scenarios, as well as on the widely used Robo3D benchmark [52]. More details of datasets are provided in Appendix B." + }, + { + "type": "page_number", + "bbox": [ + 0.479, + 0.945, + 0.521, + 0.958 + ], + "angle": 0, + "content": "27502" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.093, + 0.089, + 0.907, + 0.226 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.137, + 0.233, + 0.856, + 0.25 + ], + "angle": 0, + "content": "Figure 4. Based on MonoCD [54], we provide more comparisons with KITTI-C, regarding Mean \\(AP_{3D|R_{40}}\\)." + }, + { + "type": "table_caption", + "bbox": [ + 0.132, + 0.262, + 0.863, + 0.277 + ], + "angle": 0, + "content": "Table 2. Detection results on nuScenes-C and real-world Night and Rainy scenarios of nuScenes, regarding mAP and NDS." + }, + { + "type": "table", + "bbox": [ + 0.109, + 0.287, + 0.89, + 0.358 + ], + "angle": 0, + "content": "
MetricMethodnuScenes-CReal-world Scenarios
BrightnessCameraCrashColorQuantFogFrameLostLowLightMotionBlurSnowAvg.nuScenes-NightnuScenes-RainyAvg.
mAPBEVFormer-tiny24.2622.0123.9121.9820.4316.1120.7810.8320.049.3126.7618.04
• DriveGEN (3k Snow)26.0424.1325.7824.1321.1317.2022.3911.7221.57(↑1.53)12.1929.1120.65 (↑2.61)
NDSBEVFormer-tiny34.9333.0634.5633.0431.4228.4231.8123.2731.3117.9237.3527.64
• DriveGEN (3k Snow)37.1435.3436.9335.3533.3230.4234.2324.0933.35(↑2.04)19.3740.9530.16 (↑2.52)
" + }, + { + "type": "text", + "bbox": [ + 0.089, + 0.394, + 0.482, + 0.575 + ], + "angle": 0, + "content": "Implementation Details. We implement our method and other baselines in PyTorch [29]. Following Freecontrol [25], we adopt keys from the self-attention of the first U-Net decoder layer as the features. In the self-prototype extraction stage, we run DDIM inversion with 200 steps. Subsequently, we also run 200 steps of DDIM sampling in the prototype-guided diffusion stage. We set \\( s \\) and \\( \\tau \\) to 7.5 and 0.3 following FreeControl [25], while \\( \\sigma \\) is set to 0.1. All 3D detectors are trained on a mix of original and augmented data. All results in the manuscript are based on Stable Diffusion (SD) 1.5 [35]. More details and the results of SD 2.1 and XL 1.0 are put in Appendix C." + }, + { + "type": "text", + "bbox": [ + 0.089, + 0.578, + 0.483, + 0.714 + ], + "angle": 0, + "content": "Compared Methods. Based on typical or state-of-the-art (SOTA) 3D detectors [20, 31, 54, 63], we fully compare DriveGEN with following methods: 1) source-only, i.e., directly apply the well-trained model to corrupted test data; 2) Traditional data augmentation, i.e., Color Jitter and Brightness; 3) training-based controllable T2I diffusion: ControlNet [59] trains auxiliary modules for spatial control; 4) training-free controllable T2I diffusion: Freecontrol [25] enables zero-shot control of pretrained diffusion models." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.717, + 0.483, + 0.809 + ], + "angle": 0, + "content": "Evaluation Protocols. We report the experimental results in the Average Precision (AP) for 3D bounding boxes, denoted as \\( AP_{3D|R_{40}} \\). On the KITTI-C dataset, the results present the mean values of three difficulty levels and Intersection over Union (IoU) thresholds are set to 0.7, 0.5, 0.5 for Cars and 0.5, 0.25, 0.25 for Pedestrians and Cyclists." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.811, + 0.484, + 0.9 + ], + "angle": 0, + "content": "As for nuScenes, the mean average precision (mAP) and nuScenes detection score (NDS) are calculated following BEVFormer [20]. The metrics also include five true positive metrics, including ATE, ASE, AOE, AVE, and AAE for measuring errors in translation, scale, orientation, velocity, and attributes." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.393, + 0.833, + 0.409 + ], + "angle": 0, + "content": "5.1. Comparisons with Previous Methods" + }, + { + "type": "text", + "bbox": [ + 0.511, + 0.415, + 0.907, + 0.656 + ], + "angle": 0, + "content": "For monocular 3d object detection, Table 1 gives the following observations: 1) Applying the well-trained detector to corrupted scenarios suffers significant performance degradation due to the data distribution shifts and traditional data augmentation offers limited gains. 2) ControlNet (pretrained on ade20k [70]) fails to enhance 3d detectors with limited training data even if it is provided with accurate segmentation masks [34] and rich prompts [6] for fine-tuning. 3) Freecontrol achieves minor gains in a few cases but generally degrades, indicating that training-free T2I diffusion is a feasible solution but requires careful object preservation. 4) DriveGEN consistently outperforms all compared methods over all categories within various base models for both known and unseen scenarios, like improving MonoFlex by an average of \\(7.6\\mathrm{mAP}\\) across 13 various OOD scenarios with the single Snow scenario augmentation." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.657, + 0.909, + 0.825 + ], + "angle": 0, + "content": "As for multi-view 3D object detection, we randomly select 3k daytime training images and apply Snow augmentation (3k Snow) to enhance the detector [20]. Table 2 shows DriveGEN consistently enhances BEVFormer-tiny across 8 OOD scenarios with an average of \\(1.53\\mathrm{mAP}\\) and 2.04 NDS, further confirming our effectiveness. Moreover, we validate the detector on real Night and Rainy scenarios of nuScenes. With the help of DriveGEN, it also achieves notable gains in real-world tasks with an average improvement of \\(2.61\\mathrm{mAP}\\) and 2.52 NDS, further demonstrating our superiority. More detailed results are put in the Appendix D." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.833, + 0.862, + 0.85 + ], + "angle": 0, + "content": "5.2. More Augmented Scenarios for Training" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.856, + 0.908, + 0.903 + ], + "angle": 0, + "content": "An intuitive concern is whether the robustness of detectors progressively improves with the increasing of additional augmented scenarios (i.e., Only Snow aug. \\(\\rightarrow 3\\) scenarios" + }, + { + "type": "page_number", + "bbox": [ + 0.479, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "27503" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.109, + 0.089, + 0.891, + 0.226 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.129, + 0.239, + 0.867, + 0.255 + ], + "angle": 0, + "content": "Figure 5. Ablation studies on semantic feature alignment loss \\( g_{sa} \\), self-prototypes \\( \\hat{\\mathbf{P}}_t \\) and shallow feature alignment loss \\( g_{sl} \\)." + }, + { + "type": "image", + "bbox": [ + 0.111, + 0.27, + 0.891, + 0.442 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.09, + 0.455, + 0.908, + 0.486 + ], + "angle": 0, + "content": "Figure 6. Qualitative results of DriveGEN for multi-view 3D object detection on nuScenes. DriveGEN supports vision-centric 3D detection tasks since our method only requires input images and corresponding object annotations without any additional diffusion model training." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.511, + 0.484, + 0.678 + ], + "angle": 0, + "content": "aug. \\(\\rightarrow 6\\) scenarios aug.). Based on Table 1 and Figure 4, we are given additional observations: 1) For ControlNet, the model performance drop for OOD scenarios grows as the number of augmentation scenarios increases, demonstrating that severe object omission leads to noisy training data. 2) For Freecontrol, the model performance across various OOD scenarios remains stable or tends to decline, indicating its object preservation may be unstable. 3) Owing to the precise object preservation, DriveGEN obtains continuous overall performance (i.e., Avg.) improvement across all categories with the incorporation of OOD scenarios." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.706, + 0.255, + 0.721 + ], + "angle": 0, + "content": "5.3. Ablation Studies" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.735, + 0.484, + 0.903 + ], + "angle": 0, + "content": "To examine DriveGEN, we present qualitative results guided by different loss settings. As shown in Figure 5, compared with only generated via text prompts, introducing semantic-aware feature alignment \\( g_{sa} \\) helps preserve more image structures. Next, by guiding the diffusion process with our self-prototypes \\( \\hat{\\mathbf{P}}_t \\), we achieve more accurate object geometry, alleviating potential object geometry loss. Eventually, introducing our shallow feature alignment \\( g_{sl} \\) further improves object preservation with precise 3D geometry, especially for small objects (highlighted by red circles). More ablation studies are provided in Appendix D." + }, + { + "type": "title", + "bbox": [ + 0.514, + 0.51, + 0.707, + 0.526 + ], + "angle": 0, + "content": "5.4. Quantitative Results" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.532, + 0.907, + 0.654 + ], + "angle": 0, + "content": "We provide visualizations for multi-view 3D object detection on nuScenes in Snow scenarios as shown in Figure 6. It is evident that DriveGEN supports existing vision-centric 3D detection tasks since it requires only input images and corresponding object annotations without any additional diffusion model training, demonstrating that DriveGEN achieves superior results even in challenging multiview tasks. More visualizations are put in Appendix E." + }, + { + "type": "title", + "bbox": [ + 0.514, + 0.665, + 0.634, + 0.68 + ], + "angle": 0, + "content": "6. Conclusion" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.69, + 0.907, + 0.901 + ], + "angle": 0, + "content": "In this paper, we propose a method for robust 3D detection in driving via training-free controllable text-to-image diffusion generation, namely DriveGEN. Specifically, DriveGEN consists of: 1) Self-Prototype Extraction: To improve the guidance of self-attention features, we extract the self-prototypes by layouts to capture accurate object geometry, leveraging more precise features to guide diffusion. 2) Prototype-Guided Diffusion: To further preserve 3D object geometry, we conduct semantic-aware feature and shallow feature alignment during diffusion, alleviating the object misalignment and omission issues compared with previous methods. Experiments on KITTI-C and nuScenes demonstrate the effectiveness of DriveGEN in improving model robustness for vision-centric 3D detection." + }, + { + "type": "page_number", + "bbox": [ + 0.479, + 0.945, + 0.521, + 0.958 + ], + "angle": 0, + "content": "27504" + } + ], + [ + { + "type": "title", + "bbox": [ + 0.092, + 0.091, + 0.26, + 0.108 + ], + "angle": 0, + "content": "Acknowledgements" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.116, + 0.486, + 0.42 + ], + "angle": 0, + "content": "This work was supported by NSFC with Grant No. 62293482, by the Basic Research Project No. HZQB-KCZYZ-2021067 of Hetao Shenzhen HK S&T Cooperation Zone, by Shenzhen General Program No. JCYJ20220530143600001, by Shenzhen-Hong Kong Joint Funding No. SGDX20211123112401002, by the Shenzhen Outstanding Talents Training Fund 202002, by Guangdong Research Project No. 2017ZT07X152 and No. 2019CX01X104, by the Guangdong Provincial Key Laboratory of Future Networks of Intelligence (Grant No. 2022B1212010001), by the Guangdong Provincial Key Laboratory of Big Data Computing, CHUK-Shenzhen, by the NSFC 61931024&12326610, by the Key Area R&D Program of Guangdong Province with grant No. 2018B030338001, by the Shenzhen Key Laboratory of Big Data and Artificial Intelligence (Grant No. ZDSYS201707251409055), by Shaanxi Mathematical Basic Science Research Project(No.23JSY047), and by Tencent & Huawei Open Fund, by China Association for Science and Technology Youth Care Program." + }, + { + "type": "title", + "bbox": [ + 0.093, + 0.433, + 0.188, + 0.449 + ], + "angle": 0, + "content": "References" + }, + { + "type": "ref_text", + "bbox": [ + 0.1, + 0.458, + 0.483, + 0.514 + ], + "angle": 0, + "content": "[1] Holger Caesar, Varun Bankiti, Alex H Lang, and et.al. nuscene: A multimodal dataset for autonomous driving. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 11621-11631, 2020. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.101, + 0.516, + 0.483, + 0.571 + ], + "angle": 0, + "content": "[2] Changjian Chen, Fei Lv, Yalong Guan, Pengcheng Wang, Shengjie Yu, Yifan Zhang, and Zhuo Tang. Human-guided image generation for expanding small-scale training image datasets. arXiv preprint arXiv:2412.16839, 2024. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.574, + 0.483, + 0.615 + ], + "angle": 0, + "content": "[3] Xiaozhi Chen, Kaustav Kundu, Yukun Zhu, and et. al. 3d object proposals for accurate object class detection. Advances in neural information processing systems, 28, 2015. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.617, + 0.483, + 0.672 + ], + "angle": 0, + "content": "[4] Xiaozhi Chen, Huimin Ma, Ji Wan, Bo Li, and Tian Xia. Multi-view 3d object detection network for autonomous driving. In Proceedings of the IEEE conference on Computer Vision and Pattern Recognition, pages 1907-1915, 2017." + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.674, + 0.483, + 0.741 + ], + "angle": 0, + "content": "[5] Yukang Chen, Jianhui Liu, Xiangyu Zhang, Xiaojuan Qi, and Jiaya Jia. Voxelnext: Fully sparse voxelnet for 3d object detection and tracking. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 21674-21683, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.745, + 0.483, + 0.828 + ], + "angle": 0, + "content": "[6] Zhe Chen, Jiannan Wu, Wenhai Wang, Weijie Su, Guo Chen, Sen Xing, Muyan Zhong, Qinglong Zhang, Xizhou Zhu, Lewei Lu, et al. Internvl: Scaling up vision foundation models and aligning for generic visual-linguistic tasks. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 24185-24198, 2024. 3, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.83, + 0.483, + 0.871 + ], + "angle": 0, + "content": "[7] Prafulla Dhariwal and Alexander Nichol. Diffusion models beat gans on image synthesis. Advances in neural information processing systems, 34:8780-8794, 2021. 3, 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.874, + 0.483, + 0.901 + ], + "angle": 0, + "content": "[8] Dave Epstein, Allan Jabri, Ben Poole, Alexei Efros, and Aleksander Holynski. Diffusion self-guidance for control" + }, + { + "type": "list", + "bbox": [ + 0.1, + 0.458, + 0.483, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.545, + 0.093, + 0.905, + 0.12 + ], + "angle": 0, + "content": "able image generation. Advances in Neural Information Processing Systems, 36:16222-16239, 2023. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.524, + 0.124, + 0.905, + 0.178 + ], + "angle": 0, + "content": "[9] Songwei Ge, Taesung Park, Jun-Yan Zhu, and Jia-Bin Huang. Expressive text-to-image generation with rich text. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 7545-7556, 2023. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.181, + 0.905, + 0.248 + ], + "angle": 0, + "content": "[10] Andreas Geiger, Philip Lenz, and Raquel Urtasun. Are we ready for autonomous driving? the kitti vision benchmark suite. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 3354-3361. IEEE, 2012. 1, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.253, + 0.905, + 0.307 + ], + "angle": 0, + "content": "[11] Dan Hendrycks and Thomas Dietterich. Benchmarking neural network robustness to common corruptions and perturbations. In International Conference on Learning Representations, 2018. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.311, + 0.905, + 0.338 + ], + "angle": 0, + "content": "[12] Jonathan Ho and Tim Salimans. Classifier-free diffusion guidance. arXiv preprint arXiv:2207.12598, 2022. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.341, + 0.905, + 0.382 + ], + "angle": 0, + "content": "[13] Jonathan Ho, Ajay Jain, and Pieter Abbeel. Denoising diffusion probabilistic models. Advances in neural information processing systems, 33:6840-6851, 2020. 3, 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.385, + 0.907, + 0.453 + ], + "angle": 0, + "content": "[14] Yuming Hong, Weiyao Xu, Yi Yang, Hao Zhou, and Xiaoxue Wang. Magicdrive: 3d controllable driving video synthesis using layout and 3d representations. Proceedings of the IEEE/CVF International Conference on Computer Vision (ICCV), 2021. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.457, + 0.905, + 0.512 + ], + "angle": 0, + "content": "[15] Xinyu Huang, Mingyu Chen, Yunqi Li, Hao Wang, and Yuxing Zhao. Diffumask: Synthetic image and semantic mask annotation using cross-attention in diffusion models. arXiv preprint arXiv:2211.09123, 2022. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.515, + 0.907, + 0.583 + ], + "angle": 0, + "content": "[16] Junsoo Kim, Yeongjae Ye, Dongheon Cho, Jaechan Seo, and Jiseob Jeong. Drivingdiffusion: Realistic multi-view video generation with diffusion models. Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), 2021. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.586, + 0.905, + 0.641 + ], + "angle": 0, + "content": "[17] Feng Li, Yuan Zhang, Xinyang Wang, and Meng Liu. Vpd: Vision-based perception diffusion models for human pose estimation and dense visual annotations. arXiv preprint arXiv:2109.05270, 2021. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.644, + 0.905, + 0.685 + ], + "angle": 0, + "content": "[18] Hongyu Li, Pengfei Zhao, Wei Jiang, and Zihan Chen. Sensor fusion with diffusion models for enhanced 3d object detection. arXiv preprint arXiv:2205.02574, 2022. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.688, + 0.905, + 0.743 + ], + "angle": 0, + "content": "[19] Peiliang Li, Xiaozhi Chen, and Shaojie Shen. Stereo r-cnn based 3d object detection for autonomous driving. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 7644-7652, 2019. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.746, + 0.905, + 0.828 + ], + "angle": 0, + "content": "[20] Zhiqi Li, Wenhai Wang, Hongyang Li, Enze Xie, Chonghao Sima, Tong Lu, Yu Qiao, and Jifeng Dai. Bevformer: Learning bird's-eye-view representation from multi-camera images via spatiotemporal transformers. In European conference on computer vision, pages 1-18. Springer, 2022. 7, 2, 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.832, + 0.905, + 0.901 + ], + "angle": 0, + "content": "[21] Hongbin Lin, Yifan Zhang, Zhen Qiu, Shuaicheng Niu, Chuang Gan, Yanxia Liu, and Mingkui Tan. Prototype-guided continual adaptation for class-incremental unsupervised domain adaptation. In European Conference on Computer Vision, pages 351-368. Springer, 2022. 2" + }, + { + "type": "list", + "bbox": [ + 0.517, + 0.093, + 0.907, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.48, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "27505" + } + ], + [ + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.092, + 0.482, + 0.148 + ], + "angle": 0, + "content": "[22] Hongbin Lin, Yifan Zhang, Shuaicheng Niu, Shuguang Cui, and Zhen Li. Monotta: Fully test-time adaptation for monocular 3d object detection. In European Conference on Computer Vision, pages 96-114. Springer, 2025. 2, 3, 6, 1, 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.149, + 0.483, + 0.205 + ], + "angle": 0, + "content": "[23] Xiaoyang Liu, Yikang Wang, Xinyu Zhang, Wei Ma, Chao Wang, and Yun Chen. Multi-modal fusion for 3d object detection in autonomous driving. arXiv preprint arXiv:2204.01234, 2022. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.207, + 0.483, + 0.276 + ], + "angle": 0, + "content": "[24] Zhijian Liu, Haotian Tang, Alexander Amini, Xinyu Yang, Huizi Mao, Daniela L Rus, and Song Han. Bevfusion: Multi-task multi-sensor fusion with unified bird's-eye view representation. In 2023 IEEE international conference on robotics and automation (ICRA), pages 2774-2781. IEEE, 2023. 6, 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.278, + 0.483, + 0.36 + ], + "angle": 0, + "content": "[25] Sicheng Mo, Fangzhou Mu, Kuan Heng Lin, Yanli Liu, Bochen Guan, Yin Li, and Bolei Zhou. Freecontrol: Training-free spatial control of any text-to-image diffusion model with any condition. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 7465-7475, 2024. 2, 3, 4, 5, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.362, + 0.483, + 0.417 + ], + "angle": 0, + "content": "[26] Seokhwa Moon, Sanghyuk Hwang, Hyunsu Lee, and Minje Kim. Temporal fusion with attention mechanisms for driving scenarios. Proceedings of the IEEE/CVF International Conference on Computer Vision (ICCV), 2022. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.419, + 0.483, + 0.488 + ], + "angle": 0, + "content": "[27] Chong Mou, Xintao Wang, Liangbin Xie, Yanze Wu, Jian Zhang, Zhongang Qi, and Ying Shan. T2i-adapter: Learning adapters to dig out more controllable ability for text-to-image diffusion models. In Proceedings of the AAAI Conference on Artificial Intelligence, pages 4296–4304, 2024. 2, 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.49, + 0.483, + 0.545 + ], + "angle": 0, + "content": "[28] Youngmin Oh, Hyung-II Kim, Seong Tae Kim, and Jung Uk Kim. Monowad: Weather-adaptive diffusion model for robust monocular 3d object detection. In European Conference on Computer Vision, 2025. 2, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.547, + 0.483, + 0.629 + ], + "angle": 0, + "content": "[29] Adam Paszke, Sam Gross, Francisco Massa, Adam Lerer, James Bradbury, Gregory Chanan, Trevor Killeen, Zeming Lin, Natalia Gimelshein, Luca Antiga, et al. Pytorch: An imperative style, high-performance deep learning library. In Advances in neural information processing systems, 2019. 7, 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.632, + 0.483, + 0.699 + ], + "angle": 0, + "content": "[30] Yuheng Qin, Zhaoyang Zhang, Han Zhang, Dong Li, Baining Chen, Lu Zhang, Jiuming Gu, and Dongdong Zhang. Uncontrol: A unified diffusion model for controllable visual generation in the wild. arXiv preprint arXiv:2305.11147, 2023. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.703, + 0.483, + 0.758 + ], + "angle": 0, + "content": "[31] Zequn Qin and Xi Li. Monoground: Detecting monocular 3d objects from the ground. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 3793-3802, 2022. 6, 7, 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.76, + 0.483, + 0.814 + ], + "angle": 0, + "content": "[32] Zhen Qiu, Yifan Zhang, Hongbin Lin, Shuaicheng Niu, Yanxia Liu, Qing Du, and Mingkui Tan. Source-free domain adaptation via avatar prototype generation and adaptation. arXiv preprint arXiv:2106.15326, 2021. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.817, + 0.483, + 0.87 + ], + "angle": 0, + "content": "[33] Aditya Ramesh, Prafulla Dhariwal, Alex Nichol, Casey Chu, and Mark Chen. Hierarchical text-conditional image generation with clip latents. arXiv preprint arXiv:2204.06125, 2022. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.873, + 0.483, + 0.901 + ], + "angle": 0, + "content": "[34] Nikhila Ravi, Valentin Gabeur, Yuan-Ting Hu, Ronghang Hu, Chaitanya Ryali, Tengyu Ma, Haitham Khedr, Roman" + }, + { + "type": "list", + "bbox": [ + 0.093, + 0.092, + 0.483, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.545, + 0.093, + 0.905, + 0.134 + ], + "angle": 0, + "content": "Rädle, Chloe Rolland, Laura Gustafson, et al. Sam 2: Segment anything in images and videos. arXiv preprint arXiv:2408.00714, 2024. 3, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.137, + 0.907, + 0.205 + ], + "angle": 0, + "content": "[35] Robin Rombach, Andreas Blattmann, Dominik Lorenz, Patrick Esser, and Björn Ommer. High-resolution image synthesis with latent diffusion models. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 10684-10695, 2022. 3, 5, 7, 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.207, + 0.907, + 0.29 + ], + "angle": 0, + "content": "[36] Olaf Ronneberger, Philipp Fischer, and Thomas Brox. U-net: Convolutional networks for biomedical image segmentation. In Medical image computing and computer-assisted intervention-MICCAI 2015: 18th international conference, Munich, Germany, October 5-9, 2015, proceedings, part III 18, pages 234-241. Springer, 2015. 3, 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.292, + 0.906, + 0.36 + ], + "angle": 0, + "content": "[37] Chitwan Sahara, William Chan, Saurabh Saxena, Lala Li, Jay Whang, Emily Denton, Seyed Kamyar Seyed Ghasemipour, Burcu Karagol Ayan, Tim Salimans, Jonathan Ho, et al. Imagen: Photorealistic text-to-image diffusion models. arXiv preprint arXiv:2205.11487, 2022. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.362, + 0.906, + 0.431 + ], + "angle": 0, + "content": "[38] Jascha Sohl-Dickstein, Eric A Weiss, Niru Maheswaranathan, and Surya Ganguli. Deep unsupervised learning using nonequilibrium thermodynamics. Proceedings of the 32nd International Conference on Machine Learning, 2015. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.433, + 0.906, + 0.473 + ], + "angle": 0, + "content": "[39] Jiaming Song, Chenlin Meng, and Stefano Ermon. Denoising diffusion implicit models. arXiv preprint arXiv:2010.02502, 2020. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.476, + 0.906, + 0.532 + ], + "angle": 0, + "content": "[40] Yang Song, Jascha Sohl-Dickstein, Diederik P Kingma, Abhishek Kumar, Stefano Ermon, and Ben Poole. Score-based generative modeling through stochastic differential equations. arXiv preprint arXiv:2011.13456, 2021. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.533, + 0.906, + 0.601 + ], + "angle": 0, + "content": "[41] Yang Song, Jascha Sohl-Dickstein, Diederik P Kingma, Abhishek Kumar, Stefano Ermon, and Ben Poole. Score-based generative modeling through stochastic differential equations. In International Conference on Learning Representations, 2021. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.604, + 0.906, + 0.673 + ], + "angle": 0, + "content": "[42] Ziying Song, Guoxing Zhang, Lin Liu, Lei Yang, Shaoqing Xu, Caiyan Jia, Feiyang Jia, and Li Wang. Robofusion: towards robust multi-modal 3d object detection via sam. In Proceedings of the Thirty-Third International Joint Conference on Artificial Intelligence, pages 1272–1280, 2024. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.675, + 0.906, + 0.731 + ], + "angle": 0, + "content": "[43] Lin Sun, Liangliang Cao, Yang Zheng, Bo Wang, Zhiming Li, Wenzhe Xie, and Xu Sun. Panacea: A framework for diverse, controllable generation of urban scenes. arXiv preprint arXiv:2207.10701, 2022. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.732, + 0.906, + 0.772 + ], + "angle": 0, + "content": "[44] Alexander Swerdlow, Runsheng Xu, and Bolei Zhou. Streetview image generation from a bird's-eye view layout. IEEE Robotics and Automation Letters, 2024. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.775, + 0.906, + 0.83 + ], + "angle": 0, + "content": "[45] Sergey Tulyakov, Ming-Yu Liu, Xiaodong Yang, and Jan Kautz. Mocogan: Decomposing motion and content for video generation. Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition (CVPR), 2018. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.832, + 0.906, + 0.901 + ], + "angle": 0, + "content": "[46] Narek Tumanyan, Michal Geyer, Shai Bagon, and Tali Dekel. Plug-and-play diffusion features for text-driven image-to-image translation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 1921-1930, 2023. 2, 3, 4, 5" + }, + { + "type": "list", + "bbox": [ + 0.517, + 0.093, + 0.907, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.48, + 0.945, + 0.52, + 0.957 + ], + "angle": 0, + "content": "27506" + } + ], + [ + { + "type": "ref_text", + "bbox": [ + 0.092, + 0.093, + 0.484, + 0.12 + ], + "angle": 0, + "content": "[47] A Vaswani. Attention is all you need. Advances in Neural Information Processing Systems, 2017. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.091, + 0.122, + 0.484, + 0.191 + ], + "angle": 0, + "content": "[48] Pascal Vincent, Hugo Larochelle, Yoshua Bengio, and Pierre-Antoine Manzagol. Extracting and composing robust features with denoising autoencoders. Proceedings of the 25th International Conference on Machine Learning, 2008. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.193, + 0.484, + 0.262 + ], + "angle": 0, + "content": "[49] Yan Wang, Wei-Lun Chao, Divyansh Garg, and et. al. Pseudo-lidar from visual depth estimation: Bridging the gap in 3d object detection for autonomous driving. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 8445-8453, 2019. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.264, + 0.483, + 0.332 + ], + "angle": 0, + "content": "[50] Hai Wu, Chenglu Wen, Shaoshuai Shi, Xin Li, and Cheng Wang. Virtual sparse convolution for multimodal 3d object detection. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 21653-21662, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.334, + 0.484, + 0.404 + ], + "angle": 0, + "content": "[51] Jinheng Xie, Yuexiang Li, Yawen Huang, Haozhe Liu, Wentian Zhang, Yefeng Zheng, and Mike Zheng Shou. Boxdiff: Text-to-image synthesis with training-free box-constrained diffusion. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 7452-7461, 2023. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.406, + 0.484, + 0.475 + ], + "angle": 0, + "content": "[52] Shaoyuan Xie, Lingdong Kong, Wenwei Zhang, Jiawei Ren, Liang Pan, Kai Chen, and Ziwei Liu. Benchmarking and improving bird's eye view perception robustness in autonomous driving. IEEE Transactions on Pattern Analysis and Machine Intelligence, 2025. 2, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.477, + 0.484, + 0.546 + ], + "angle": 0, + "content": "[53] Junkai Xu, Liang Peng, Haoran Cheng, Hao Li, Wei Qian, Ke Li, Wenxiao Wang, and Deng Cai. Mononerd: Nerflike representations for monocular 3d object detection. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 6814-6824, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.548, + 0.484, + 0.616 + ], + "angle": 0, + "content": "[54] Longfei Yan, Pei Yan, Shengzhou Xiong, Xuanyu Xiang, and Yihua Tan. Monocd: Monocular 3d object detection with complementary depths. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 10248-10257, 2024. 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.618, + 0.484, + 0.659 + ], + "angle": 0, + "content": "[55] Xu Yan, Weiming Zeng, Jiaming Zhu, Hailin Chen, Yaowei Li, and Deng Cai. Videogpt: Video generation using vq-vae and transformers. arXiv preprint arXiv:2104.10157, 2021. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.661, + 0.484, + 0.757 + ], + "angle": 0, + "content": "[56] Chenyu Yang, Yuntao Chen, Hao Tian, Chenxin Tao, Xizhou Zhu, Zhaoxiang Zhang, Gao Huang, Hongyang Li, Yu Qiao, Lewei Lu, et al. Bevformer v2: Adapting modern image backbones to bird's-eye-view recognition via perspective supervision. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 17830-17839, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.759, + 0.484, + 0.815 + ], + "angle": 0, + "content": "[57] Kairui Yang, Enhui Ma, Jibin Peng, Qing Guo, Di Lin, and Kaicheng Yu. Bevcontrol: Accurately controlling street-view elements with multi-perspective consistency via bev sketch layout. arXiv preprint arXiv:2308.01661, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.817, + 0.484, + 0.872 + ], + "angle": 0, + "content": "[58] Tianwei Yin, Xingyi Zhou, and Philipp Krahenbuhl. Center-based 3d object detection and tracking. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 11784-11793, 2021. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.874, + 0.484, + 0.901 + ], + "angle": 0, + "content": "[59] Lvmin Zhang, Anyi Rao, and Maneesh Agrawala. Adding conditional control to text-to-image diffusion models. In" + }, + { + "type": "list", + "bbox": [ + 0.091, + 0.093, + 0.484, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.545, + 0.093, + 0.905, + 0.12 + ], + "angle": 0, + "content": "Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 3836-3847, 2023. 2, 3, 4, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.122, + 0.905, + 0.19 + ], + "angle": 0, + "content": "[60] Renrui Zhang, Han Qiu, Tai Wang, Ziyu Guo, Ziteng Cui, Yu Qiao, Hongsheng Li, and Peng Gao. Monodetr: Depth-guided transformer for monocular 3d object detection. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 9155-9166, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.193, + 0.905, + 0.234 + ], + "angle": 0, + "content": "[61] Yifan Zhang and Bryan Hooi. Hipa: enabling one-step text-to-image diffusion models via high-frequency-promoting adaptation. arXiv preprint arXiv:2311.18158, 2023. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.236, + 0.905, + 0.303 + ], + "angle": 0, + "content": "[62] Yifan Zhang, Ying Wei, Qingyao Wu, Peilin Zhao, Shuaicheng Niu, Junzhou Huang, and Mingkui Tan. Collaborative unsupervised domain adaptation for medical image diagnosis. IEEE Transactions on Image Processing, 29: 7834-7844, 2020. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.305, + 0.905, + 0.361 + ], + "angle": 0, + "content": "[63] Yunpeng Zhang, Jiwen Lu, and Jie Zhou. Objects are different: Flexible monocular 3d object detection. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 3289-3298, 2021. 1, 2, 6, 7, 3, 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.363, + 0.905, + 0.417 + ], + "angle": 0, + "content": "[64] Yifan Zhang, Bingyi Kang, Bryan Hooi, Shuicheng Yan, and Jiashi Feng. Deep long-tailed learning: A survey. IEEE transactions on pattern analysis and machine intelligence, 45(9):10795-10816, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.419, + 0.905, + 0.473 + ], + "angle": 0, + "content": "[65] Yifan Zhang, Daquan Zhou, Bryan Hooi, Kai Wang, and Ji-ashi Feng. Expanding small-scale datasets with guided imagination. In Advances in neural information processing systems, pages 76558–76618, 2023. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.476, + 0.905, + 0.544 + ], + "angle": 0, + "content": "[66] Bowen Zhao, Feng Lu, Cheng Zhang, Jiaqi Hu, and Lin Song. Drivedreamer: Realistic multi-view video generation of driving scenes with future prediction. Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition (CVPR), 2022. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.546, + 0.905, + 0.613 + ], + "angle": 0, + "content": "[67] Shihao Zhao, Dongdong Chen, Yen-Chun Chen, Jianmin Bao, Dongdong Zhang, Lu Yuan, Han Zhang, Dong Li, Baining Chen, and Lu Zhang. Uni-controlnet: All-in-one control to text-to-image diffusion models. arXiv preprint arXiv:2305.16322, 2023. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.616, + 0.905, + 0.685 + ], + "angle": 0, + "content": "[68] Guangcong Zheng, Xianpan Zhou, Xuewei Li, Zhongang Qi, Ying Shan, and Xi Li. Layoutdiffusion: Controllable diffusion model for layout-to-image generation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 22490-22499, 2023. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.687, + 0.905, + 0.755 + ], + "angle": 0, + "content": "[69] Longtao Zheng, Yifan Zhang, Hanzhong Guo, Jiachun Pan, Zhenxiong Tan, Jiahao Lu, Chuanxin Tang, Bo An, and Shuicheng Yan. Memo: Memory-guided diffusion for expressive talking video generation. arXiv preprint arXiv:2412.04448, 2024. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.757, + 0.905, + 0.825 + ], + "angle": 0, + "content": "[70] Bolei Zhou, Hang Zhao, Xavier Puig, Sanja Fidler, Adela Barriuso, and Antonio Torralba. Scene parsing through ade20k dataset. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 633-641, 2017. 7" + }, + { + "type": "list", + "bbox": [ + 0.517, + 0.093, + 0.905, + 0.825 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.48, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "27507" + } + ] +] \ No newline at end of file diff --git a/2025/DriveGEN_ Generalized and Robust 3D Detection in Driving via Controllable Text-to-Image Diffusion Generation/d5fb6699-8753-41df-8e0e-5b82ff7bf5a3_origin.pdf b/2025/DriveGEN_ Generalized and Robust 3D Detection in Driving via Controllable Text-to-Image Diffusion Generation/d5fb6699-8753-41df-8e0e-5b82ff7bf5a3_origin.pdf new file mode 100644 index 0000000000000000000000000000000000000000..c2e8b4072b1613b58b4b5057ddf4247ce99fef11 --- /dev/null +++ b/2025/DriveGEN_ Generalized and Robust 3D Detection in Driving via Controllable Text-to-Image Diffusion Generation/d5fb6699-8753-41df-8e0e-5b82ff7bf5a3_origin.pdf @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:8db5773ec8702edd168ef59636c27297a8a286ce4d47eb7eba66e72cfb6af9bf +size 5356380 diff --git a/2025/DriveGEN_ Generalized and Robust 3D Detection in Driving via Controllable Text-to-Image Diffusion Generation/full.md b/2025/DriveGEN_ Generalized and Robust 3D Detection in Driving via Controllable Text-to-Image Diffusion Generation/full.md new file mode 100644 index 0000000000000000000000000000000000000000..cbe8bb608fc36eb1d20695aeb7e2e6f46d571717 --- /dev/null +++ b/2025/DriveGEN_ Generalized and Robust 3D Detection in Driving via Controllable Text-to-Image Diffusion Generation/full.md @@ -0,0 +1,299 @@ +# DriveGEN: Generalized and Robust 3D Detection in Driving via Controllable Text-to-Image Diffusion Generation + +Hongbin Lin $^{1,2}$ Zilu Guo $^{1,2}$ Yifan Zhang $^{3}$ Shuaicheng Niu $^{4}$ Yafeng Li $^{5}$ Ruimao Zhang $^{6}$ Shuguang Cui $^{2,1}$ Zhen Li $^{2,1*}$ + +$^{1}$ FNii-Shenzhen $^{2}$ SSE, CUHK-Shenzhen $^{3}$ National University of Singapore $^{4}$ Nanyang Technological University $^{5}$ Baoji University of Arts and Sciences $^{6}$ Sun Yat-sen University + +![](images/f4971ec62903b3190e1607d19ff2aa63e8d36e158741595d7c392d6280f7eebb.jpg) +Figure 1. An illustration of DriveGEN for enhancing vision-centric 3D detectors. DriveGEN extends original training images to diverse Out-of-Distribution (OOD) scenarios without additional diffusion model training, preserving all annotated objects. Even with a single augmentation, i.e., 'Snow', the augmented detector [63] is improved not only on the original KITTI [10] val set (Ideal) and already known 'Snow' scenario, but also in unseen 'Fog' and 'Defocus' scenarios, demonstrating a comprehensive model generalizability improvement. + +# Abstract + +In autonomous driving, vision-centric 3D detection aims to identify 3D objects from images. However, high data collection costs and diverse real-world scenarios limit the scale of training data. Once distribution shifts occur between training and test data, existing methods often suffer from performance degradation, known as Out-of-Distribution (OOD) problems. To address this, controllable Text-to-Image (T2I) diffusion offers a potential solution for training data enhancement, which is required to generate diverse OOD scenarios with precise 3D object geometry. Nevertheless, exist + +ing controllable T2I approaches are restricted by the limited scale of training data or struggle to preserve all annotated 3D objects. In this paper, we present DriveGEN, a method designed to improve the robustness of 3D detectors in Driving via Training-Free Controllable Text-to-Image Diffusion Generation. Without extra diffusion model training, DriveGEN consistently preserves objects with precise 3D geometry across diverse OOD generations, consisting of 2 stages: 1) Self-Prototype Extraction: We empirically find that self-attention features are semantic-aware but require accurate region selection for 3D objects. Thus, we extract precise object features via layouts to capture 3D object geometry, termed self-prototypes. 2) Prototype-Guided Diffusion: To preserve objects across various OOD scenarios, we per + +form semantic-aware feature alignment and shallow feature alignment during denoising. Extensive experiments demonstrate our effectiveness in improving 3D detection. The code is available at github.com/Hongbin98/DriveGEN. + +# 1. Introduction + +Three-dimensional (3D) object detection intends to identify objects and assess their spatial and dimensional attributes via various sensor inputs, which is widely explored in Autonomous Driving (AD) and robotics [3-5, 19, 49, 50, 58]. To save sensor costs, there is a growing trend towards implementing perception systems by vision-centric 3D detection which relies solely on single or multi-view RGB images along with camera calibration information. Despite challenges in this field, existing methods [53, 56, 60] have demonstrated promising results across various benchmarks. + +Behind the success, one prerequisite is that test images share the same distribution as training images. However, this assumption is often not held in diverse real-world scenarios [21, 32, 62, 64], especially for perception systems that are required for long durations, like autonomous vehicles navigating through different weather conditions. Once the environment changes, well-trained detectors may suffer performance degradation due to the existence of data distribution shifts between training images and test images. As shown in Figure 1, the original detector performs well in the ideal (sunny) scenario while failing to maintain stable performance in the Snow and Fog scenario due to the out-of-distribution (OOD) issue. Considering the widespread application of 3D detection, significant performance degradation on OOD test data may result in traffic accidents and serious safety risks. Thus, it is essential to address the OOD generalization problem of vision-centric 3D detection. + +Recently, two attempts sought to solve the distribution shifts in AD by test-time model adaptation [22] or restoring OOD scenarios to ideal conditions with a weather-adaptive diffusion model [28]. However, they require additional consumption during the test phase, leading to additional energy costs while the driving range is crucial in AD. To save test-time computation costs, we explore controllable Text-to-Image (T2I) diffusion generation for training data augmentation, thereby improving the robustness of 3D detectors. Prior training-based controllable T2I approaches such as ControlNet [59] and T2i-adapter [27] offer users fine-grained spatial control but rely on substantial training data to train auxiliary modules, making them impractical for handling OOD scenarios in AD. As for training-free methods, they [25, 46] leverage low-resolution self-attention features to capture image structures for diffusion model guidance. However, these features are relatively coarse for preserving object geometric characteristics of different classes in 3D object detection. As shown in Figure 2, the object + +geometry may be lost after multiple rounds of downsampling $(32\mathrm{x})$ , leading to severe orientation error, position misalignment and omissions, particularly for tiny objects. Therefore, an eligible training-free controllable T2I diffusion method for training data enhancement in 3D detection should preserve all objects with precise 3D geometry. + +To address these challenges, we propose a method namely DriveGEN for Generalized and Robust 3D Detection in Driving via Training-Free Controllable Text-to-Image Diffusion GENeration. DriveGEN enhances the model robustness within various OOD scenarios generation, consisting of two stages: 1) Self-Prototype Extraction. To capture accurate object geometric characteristics, we first extract principal components of self-attention features [25, 46] to obtain coarse semantic-aware features. Then, we adopt layouts (i.e., bounding boxes) to re-weight features in object regions via a peak function, termed self-prototypes. 2) Prototype-Guided Diffusion. To preserve objects with precise 3D geometry, we perform semantic-aware feature alignment with self-prototypes during the denoising phase. However, these semantic-aware features are relatively coarse to represent tiny objects as shown in Figure 2. Therefore, we devise a shallow feature alignment strategy to constrain the diffusion process and retrain tiny objects. + +Contributions: 1) To the best of our knowledge, we are the first to explore training-free controllable T2I diffusion generation to enhance the model robustness of vision-centric 3D detection. Even with a single augmentation, DriveGEN improves the 3D detector [63] by an average of $7.6\mathrm{mAP}$ across 13 OOD scenarios. 2) With self-prototype extraction and prototype-guided diffusion stages, DriveGEN enables OOD scenario augmentation while preserving precise 3D geometry for all objects, enhancing the applicability of diffusion models in autonomous driving. 3) Extensive experiments show DriveGEN significantly improves existing monocular 3D detectors across 13 OOD corruption on KITTI and multi-view 3D detectors on the real scenarios (Night, Rainy) of nuScenes, demonstrating the effectiveness of DriveGEN in boosting model generalizability. + +# 2. Related Work + +We first review the literature on explorations of model robustness in autonomous driving, and then discuss controllable T2I diffusion generation methods. More discussions on vision-centric 3D object detection are in Appendix A. + +Model Robustness in Autonomous Driving. The reliability of vision-centric autonomous driving systems depends on the robustness of perception models, especially under challenging scenarios like adverse weather, varied lighting, and complex urban environments [52]. To improve robustness, existing methods employ diffusion models to generate synthetic driving scenes for training data augmentation. Specifically, BEVGen [44] and BEVControl [57] employ + +![](images/2b113af7cb6544eac7c504aaefffdea8f96a57ef6a32ee4179d0ab7bbd2d0212.jpg) +Figure 2. Illustration of the importance of object preservation in 3d detection based on KITTI. ControlNet [59] suffers object orientation errors and omissions even if accurate segmentation masks [34] and rich prompts [6] are provided, showing that training-based methods may struggle with spatial control with limited training data. Additionally, Freecontrol [25] relies on coarse low-resolution features to capture semantic structures that may encounter potential object geometry loss, resulting in object position misalignment and omission issues. + +Layouts to create flexible scene configurations. Innovative frameworks like Panacea [43] and MagicDrive [14] extend this controllability to multi-view scenes, while models such as DriveDreamer [66] and DrivingDiffusion [16] focus on generating realistic videos for model validation. Recently, RoboFusion [42] has sought to tackle OOD noise scenarios by leveraging visual foundation models like SAM. + +Controllable T2I Diffusion. Controllable text-to-image (T2I) diffusion models [35, 61, 69] are increasingly applied in real-world scenarios, where text prompts or scene descriptions guide the creation of synthetic data. By leveraging pre-trained models like Stable Diffusion [35] or other large-scale T2I models [33, 37], existing methods can generate high-fidelity and diverse images based on textual input, allowing for controlled variations [2, 65]. Techniques such as Uni-ControlNet [67], UniControl [30] and Layout-diffusion [68] further enhance control by integrating multimodal inputs, creating contextually accurate and condition-specific scenes. Recent works also explore the utility of cross-attention mechanisms and latent embeddings to align generated images closely with text descriptions, supporting the generation for perception tasks like semantic segmentation and depth estimation [15, 17]. However, these methods require a large amount of training data to train auxiliary modules, making them impractical for handling OOD scenarios in AD for the limited data scale. + +Another alternative choice is training-free controllable T2I diffusion, PnP [46] directly injects the self-attention features of condition images to guide the diffusion process while Freecontrol [25] utilizes PCA to extract principle components of self-attention features for semantic structures extraction. However, existing approaches solely rely on the relatively coarse self-attention features, which often results in object region omissions and misalignment as we mentioned before. To this end, we propose our DriveGEN to preserve all objects with precise 3D geometry. + +# 3. Preliminary + +Pre-trained diffusion models [7, 13, 35] are probabilistic generative models, which consists of two complementary stochastic processes: 1) In the forward process, Gaussian + +noise is incrementally added to a clean image $\mathbf{x}_0$ . 2) As for the backward process, it conducts the iterative process of denoising the initial Gaussian noise image $\mathbf{z}_{N_t}$ , where a cleaner image is given at each step. To be specific, this process leverages a denoising network $\epsilon_{\theta}(\mathbf{z}_t,t)$ to estimate the added noise in the forward process. Subsequently, for $\mathbf{z}_t$ at each step, the network removes the estimated noise perturbation to achieve a cleaner $\mathbf{z}_{t - 1}$ . + +For T2I diffusion models, a previous method [41] has shown that $\epsilon_{\theta}$ can approximate the score function for the marginal distributions $p_t$ , i.e., $-\sigma_t\nabla_{\mathbf{z}_t}\log p_t(\mathbf{z}_t|\mathbf{c})$ where $\sigma_t$ denotes a noise schedule and $\mathbf{c}$ is a text prompt. To incorporate auxiliary information $y$ into the diffusion sampling process, existing methods [7, 8, 25] guide the diffusion process by adding a time-dependent energy function termed $g(\mathbf{z}_t;t,y)$ with a strength coefficient $s$ : + +$$ +\hat {\epsilon} _ {\boldsymbol {\theta}} (\mathbf {z} _ {t}; t, c) = \epsilon_ {\boldsymbol {\theta}} (\mathbf {z} _ {t}; t, c) - s \cdot g (\mathbf {z} _ {t}; t, y), +$$ + +where $g$ can be defined via bounding boxes [51]. In this paper, we adopt a standard choice $\epsilon_{\theta}$ following existing methods [25, 46], i.e., U-Net architecture [36] with multi-level self-attention and cross-attention layers [47]. + +# 4. Training-Free Controllable T2I Diffusion + +# 4.1. Problem Statement + +Without loss of generality, we denote the well-trained 3D detector as $f_{\Theta_d}(\cdot)$ which is obtained by training on labeled training images $\mathcal{D} = \{(\mathbf{x}_i,\mathbf{y}_i)\}_{i = 1}^N$ . The training images follow the original training distribution $P(\mathbf{x})$ (i.e., $\mathbf{x} \sim P(\mathbf{x})$ ). Here, $\Theta_d$ represents the parameters of the detector and $N$ is the total number of training images. After training, the model is applied to unlabeled test images $\mathcal{D}_t = \{\mathbf{x}^t\}$ . Unfortunately, the well-trained model often encounters Out-of-distribution (OOD) test scenarios in autonomous driving (AD) due to prevalent natural corruptions, namely distribution shifts, i.e., $\mathbf{x}^t \sim Q(\mathbf{x})$ and $P(\mathbf{x}) \neq Q(\mathbf{x})$ . + +As previously discussed, existing methods [22, 28] are relatively time-consuming due to their additional computation costs at test time. Hence, we investigate controllable + +![](images/fe48bb7292bb6f94362cb3bfbe03c238620ceaea624228e71750f9429cd5d363.jpg) + +![](images/a2ac8b0a0517294c12ee272b6275b087c7c24d8403dc581423b81f02c7cdc9e0.jpg) +Figure 3. An overview of our DriveGEN, consists of two stages: 1) The Self-Prototype Extraction stage is devised to extract accurate semantic structures of multiple objects. To capture precise locations, we achieve fine-grained self-prototypes via leveraging layouts and the peak function to re-weight object regions rather than directly using coarse self-attention features. 2) The Prototype-Guided Diffusion stage conducts semantic-aware feature alignment for semantic structure matching and shallow feature alignment for tiny object preservation. + +Text-to-Image (T2I) diffusion generation for training data augmentation, thereby improving the robustness of 3D detectors. Particularly, it is crucial to preserve object information because the annotations would be reused during detector training. However, training-based approaches [27, 59] require substantial training data to train auxiliary modules for spatial control, which is often unsatisfying in AD. Besides, training-free methods [25, 46] mainly consider the relatively coarse self-attention features, making it challenging to preserve relatively tiny objects. + +# 4.2. Overall Scheme + +After examining the characteristics and challenges of robust 3D object detection, we introduce DriveGEN, a Training-Free Controllable T2I Diffusion Generation method devised to improve the robustness of vision-centric 3D detectors. As illustrated in Figure 3, DriveGEN consists of two stages: 1) Self-Prototype Extraction and 2) Prototype-Guided Diffusion. We briefly introduce each stage below. + +First, we develop the self-prototype extraction stage (c.f. Section 4.3) to capture the geometric characteristics of objects accurately. Existing methods apply principal compo + +nent analysis (PCA) on the self-attention features of the first decoder layer of the diffusion model $\epsilon_{\theta}$ to capture semantic structures of condition images [25, 46]. However, we empirically find that solely relying on cross-attention relevance between self-attention features and object prompts for region selection may result in object preservation errors (c.f. Figure 2). To address this, DriveGEN leverages layouts to re-weight object regions in the principal components of self-attention features via a peak function and then achieves the self-prototypes which help the diffusion model focus on the accurate object regions during generation. + +Second, we introduce the prototype-guided diffusion stage (c.f. Section 4.4) to preserve all objects during the generation process. On the one hand, we conduct semantic-aware feature alignment between self-attention features and self-prototypes to capture accurate object semantic structures. On the other hand, DriveGEN performs shallow feature alignment to further preserve object information by constraining the noisy latent at the shallow level, compensating for the gaps in semantic-aware feature alignment regarding tiny object preservation. The Pseudo-code of DriveGEN is summarized in Algorithm 1. + +# 4.3. Self-Prototype Extraction + +To accurately capture the geometric characteristics of multiple objects, we propose a self-prototype extraction strategy that consists of two components: 1) Image structure extraction and 2) Object region re-weighting. + +Image Structure Extraction. To achieve zero-shot spatial control, capturing the semantic structures of input images is essential. Prior works [25, 46] demonstrate that self-attention features (e.g., keys) of the T2I diffusion model $\epsilon_{\theta}$ are semantic-aware [35]. Thus, given the input image $\mathbf{x}_i$ , we first generate the text prompt $\mathbf{c}_i$ via a simple template 'A photo of $\{\text{object}\} ,\ldots,$ and $\{\text{object}\}'$ . Then, we leverage DDIM Inversion [39] to achieve a series of time-dependent diffusion features $\{\mathbf{z}_t^I\in \mathbb{R}^{N_c\times H\times W}\}_{t = 1}^{N_t}$ where $N_{t}$ denotes the number of DDIM inversion and sampling steps, $N_{c}$ represents the number of feature channels, and $H,W$ indicate the height and width, respectively. During sampling, we apply PCA on self-attention features to extract principal components $\{\mathbf{P}_t\}_{t = 1}^{N_p}$ , each $\mathbf{P}_t\in \mathbb{R}^{N_b\times H\times W}$ and $\mathbf{P}_t = \{\mathbf{p}_t^1,\dots,\mathbf{p}_t^{N_b}\}$ . Here $N_{b}$ denotes the number of principal components and we store $\mathbf{P}_t$ of the initial $N_{p}$ steps. + +Object Region Re-weighting. Existing methods [9, 25] select foreground regions by leveraging cross-attention maps between specific words (i.e., object names) and image features to derive a concept region mask $\mathbf{M}$ . However, relying solely on such attention maps may result in omissions and misalignment of object regions, as illustrated in Figure 2. To address this, we introduce a peak function $m(\cdot)$ to reweight $\mathbf{M}$ for each object region based on layouts $\mathbf{L}$ after the coordinate transformation through downsampling. To be specific, given a point $(p,q)$ from $\mathbf{L}_i$ of $\mathbf{x}_i$ , the corresponding object center is defined as $(\hat{p},\hat{q})$ and then we can calculate $m(p,q)$ by: + +$$ +m (p, q) = \exp \left(- \frac {(p - \hat {p}) ^ {2} + (q - \hat {q}) ^ {2}}{2 \sigma^ {2}}\right) \tag {1} +$$ + +where $\sigma \in (0,1)$ controls the range of the influence area. Then, we update the concept token mask $\mathbf{M}$ by: + +$$ +\tilde {\mathbf {M}} (p, q) = \mathbf {M} (p, q) + m (p, q) \cdot \mathbf {L} _ {i} (p, q). \tag {2} +$$ + +The object regions in $\mathbf{L}_i$ are marked as 1 while all other regions as 0. Eventually, we obtain the self-prototypes by $\{\hat{\mathbf{P}}_t\}_{t=1}^{N_p} = \{\mathbf{P}_t\}_{t=1}^{N_p}\big|_{\hat{\mathbf{M}}}$ which guide the diffusion model to localize precise locations of objects and combine $\hat{\mathbf{M}}$ for accurate object region selection. + +# 4.4. Prototype-Guided Diffusion + +To preserve all objects with precise 3D geometry during diffusion, we leverage the self-prototypes $\{\hat{\mathbf{P}}_t\}_{t=1}^{N_p}$ and time-dependent diffusion features $\{\mathbf{z}_t^I\}_{t=1}^{N_t}$ for feature alignment, including: 1) semantic-aware feature alignment and 2) shallow feature alignment. + +Algorithm 1 The pipeline of the proposed DriveGEN +Require: Training data $\{(\mathbf{x}_i^s,\mathbf{y}_i^s)\}_{i = 1}^N$ Pretrained T2I diffusion model $\epsilon_{\theta}$ Hyper-parameters $s,\sigma ,\tau$ The OOD scenario. +#Stage 1: Self-Prototype Extraction +1: for each training image $x_{i}$ do +2: Get the text prompt $c_{i}$ and layout $\mathbf{L}_i$ based on $\mathbf{y}_i$ +3: Undergo DDIM Inversion to get $\mathbf{P}_t$ +4: Get updated M via Eqn. (1), Eqn. (2) for self-prototypes $\hat{\mathbf{P}}_t$ +5: end for +#Stage 2: Prototype-Guided Diffusion +6: for each training image $\mathbf{x}_i$ do +7: Update $\mathbf{c}_i$ with specific OOD type; +8: for diffusion step $t = 1\rightarrow N_t$ do +9: if $t\leq N_p$ then +10: Calculate $g_{sa},g_{sl}$ based on Eqn. (3), Eqn. (4); +11: Get the gradient based on Eqn.(5); +12: end if +13: Update the noisy latent $\mathbf{z}_t\to \mathbf{z}_{t - 1}$ +14: end for +15: end for +16: return Output images for all $\mathbf{x}_i$ in the OOD scenario. + +Semantic-aware Feature Alignment. As mentioned before, the self-attention features of the first decoder layer can represent the semantic structures of input images. Thus, we first initialize the noisy latent $\mathbf{z}_{N_t}$ via the $\mathbf{z}_{N_t}^I$ . For the step $t \leq N_p$ , we perform PCA to obtain the principal components $\mathbf{P}_t$ of the self-attention features. Then, we adopt a pre-defined threshold $\tau$ following [9] to select informative regions by $\hat{\mathbf{M}} = \mathbb{I}(\hat{\mathbf{M}} > \tau)$ . Specifically, we calculate the semantic-aware feature alignment loss with $\mathbf{P}_t$ and $\hat{\mathbf{P}}_t$ : + +$$ +g _ {s a} = \| \hat {\mathbf {M}} \odot (\mathbf {P} _ {t} - \hat {\mathbf {P}} _ {t}) \| _ {2} ^ {2}, \tag {3} +$$ + +where $\odot$ represents element-wise multiplication, applying the mask $\hat{\mathbf{M}}$ to the difference between $\mathbf{P}_t$ and $\hat{\mathbf{P}}_t$ . + +Shallow Feature Alignment. As shown in Figure 2, we empirically find that principal components lack sufficient details to represent objects, especially tiny objects. For instance, given a bounding box of a cyclist with a height of 20 pixels and a width of 5 pixels, this object is unable to occupy even one element within principal components after multiple rounds of downsampling (e.g., $32\mathrm{x}$ ). + +Therefore, DriveGEN introduces shallow feature alignment before the forward process of diffusion, which aims to preserve fine-grained object details by constraining the noisy latent. Specifically, based on the layout $\mathbf{L}_i$ with $N_{i}$ objects, we calculate the shallow feature alignment loss by: + +$$ +g _ {s l} = \frac {1}{N _ {i}} \sum_ {p, q} \| \mathbf {L} _ {i} (p, q) \odot \left(\mathbf {z} _ {t} ^ {I} (p, q) - \mathbf {z} _ {t} (p, q)\right) \| _ {2} ^ {2}, \tag {4} +$$ + +Overall, the total scheme of DriveGEN is as follows: + +$$ +\hat {\epsilon} _ {t} = (1 + s) \epsilon_ {\theta} (\mathbf {z} _ {t}; t, c) - s \epsilon_ {\theta} (\mathbf {z} _ {t}; t, \emptyset) + g _ {s a} + g _ {s l}. \tag {5} +$$ + +Note that $\varnothing$ denotes null and DriveGEN is a classifier-free guidance [12] process with hyper-parameters $s$ , $\sigma$ and $\tau$ . + +Table 1. Comparison on the KITTI-C dataset, severity level 1 regarding Mean $AP_{3D|R_{40}}$ . The bold number indicates the best result. + +
Car, IoU @ 0.7, 0.5, 0.5
MethodTraining-free diffusionNoiseBlurWeatherDigitalAvg.
Gauss.ShotImpul.Defoc.GlassMotionSnowFrostFogBrit.Contr.PixelSat.
Monoflex [63]13.0620.9114.0920.1728.5930.3433.6430.3119.5845.2220.0129.0738.8526.45
• Color Jitter (Traditional aug.)9.5515.8111.9022.6725.3830.1234.0830.0019.2942.1019.9317.1736.4824.19
• Brightness (Traditional aug.)11.4419.4212.7311.1818.9522.0326.6421.7013.0439.6113.0821.1129.7320.05
• ControlNet (Only Snow aug.)X0.321.181.284.6511.3217.0422.8419.729.5334.798.731.2516.8811.50
• ControlNet (3 scenarios aug.)X0.701.140.290.350.370.834.073.051.209.430.940.483.202.00
• ControlNet (6 scenarios aug.)X0.000.000.000.000.001.940.000.000.000.000.000.000.000.15
• Freecontrol (Only Snow aug.)✓20.3928.0522.2712.5221.6721.0927.9717.9110.0735.999.9028.5635.2722.44
• Freecontrol (3 scenarios aug.)✓15.4321.2514.0415.3319.2216.5322.0117.4115.1824.6616.6625.3931.4419.58
• Freecontrol (6 scenarios aug.)✓12.8518.7116.3112.0616.3112.8617.6416.1814.2324.5215.6422.5624.2817.24
• DriveGEN (Only Snow aug.)✓16.4826.7224.9831.1735.5538.1341.3938.6127.8549.7629.2838.8844.1234.07
• DriveGEN (3 scenarios aug.)✓25.6337.0429.1334.1339.1536.8138.5837.9833.9345.3934.6639.3643.8136.58
• DriveGEN (6 scenarios aug.)✓24.7733.7928.2736.9240.3340.4540.6040.5638.1044.8339.2841.8144.0537.98
MonoGround [31]13.0521.7718.8720.7930.7432.0234.4327.0214.1546.2114.6333.4135.6026.36
• Color Jitter (Traditional aug.)12.8824.3118.9523.0730.4431.4235.9430.4319.8944.6620.6129.7536.6526.36
• Brightness (Traditional aug.)14.0223.5220.1423.9531.7828.7935.0831.8718.8742.9417.7525.5537.1827.03
• ControlNet (Only Snow aug.)X1.763.234.635.2012.9514.1117.7011.583.0435.212.987.2913.9810.28
• ControlNet (3 scenarios aug.)X0.000.000.241.421.681.614.905.400.5717.901.127.796.823.80
• ControlNet (6 scenarios aug.)X0.000.000.001.681.260.351.130.520.444.080.382.221.771.06
• Freecontrol (Only Snow aug.)✓11.7521.8915.7617.7021.4521.6932.0820.6013.5736.0514.0326.7538.3522.43
• Freecontrol (3 scenarios aug.)✓16.3120.7517.6113.1016.8414.8217.7215.8911.8725.2013.5522.5724.4917.75
• Freecontrol (6 scenarios aug.)✓15.2022.5915.3522.0021.1818.9517.6914.8514.8224.0216.9722.9926.1219.44
• DriveGEN (Only Snow aug.)✓17.0726.7823.7832.8937.5239.0640.6134.9125.2946.2127.1238.2544.4533.38
• DriveGEN (3 scenarios aug.)✓19.7931.4427.6336.8440.1039.3539.1736.4228.7945.4229.2942.6044.9935.53
• DriveGEN (6 scenarios aug.)✓23.8432.5930.3438.5741.2040.1938.1638.4032.5343.9534.8044.1045.1337.21
+ +
Pedestrian, IoU @ 0.5, 0.25, 0.25
Monoflex [63]1.163.540.788.0617.7015.2612.729.255.6119.875.351.498.658.42
• Color Jitter (Traditional aug.)0.993.531.4110.0514.6312.0014.7312.137.7219.029.231.1111.299.07
• Brightness (Traditional aug.)0.631.851.205.0113.7212.157.935.841.8716.892.300.483.775.67
• ControlNet (Only Snow aug.)X0.000.000.001.788.325.844.033.751.3111.001.230.000.932.94
• ControlNet (3 scenarios aug.)X0.000.000.001.153.231.141.641.350.834.062.010.000.001.19
• ControlNet (6 scenarios aug.)X0.000.000.000.000.360.000.000.000.002.500.000.000.000.22
• Freecontrol (Only Snow aug.)✓3.625.173.213.235.165.944.683.524.287.704.228.517.705.15
• Freecontrol (3 scenarios aug.)✓2.263.543.043.274.974.257.424.417.2710.916.186.837.435.52
• Freecontrol (6 scenarios aug.)✓5.999.109.084.807.365.847.227.587.1010.819.409.9010.248.03
• DriveGEN (Only Snow aug.)✓1.253.224.5616.2119.8820.6119.8014.468.8024.629.119.7718.0413.10
• DriveGEN (3 scenarios aug.)✓6.099.859.6014.9618.5916.4715.4516.3915.8821.9816.8514.5518.1314.98
• DriveGEN (6 scenarios aug.)✓6.349.227.7117.6019.2120.2016.8816.1915.7223.5716.7514.3717.3115.47
MonoGround [31]2.673.255.7617.5718.9117.7112.969.354.3724.155.893.277.1610.23
• Color Jitter (Traditional aug.)2.443.244.1115.3718.4616.5015.4512.389.1424.719.822.077.8110.89
• Brightness (Traditional aug.)2.794.107.6114.5114.1314.5212.1212.665.5520.805.362.3610.319.75
• ControlNet (Only Snow aug.)X1.851.030.817.189.978.321.322.921.3712.991.400.280.903.87
• ControlNet (3 scenarios aug.)X0.000.000.003.603.891.961.141.031.504.211.250.260.771.51
• ControlNet (6 scenarios aug.)X0.000.000.000.000.001.670.000.000.000.000.000.000.000.13
• Freecontrol (Only Snow aug.)✓10.0413.0911.7014.9112.6113.0315.1412.358.1315.6912.3416.3214.5713.07
• Freecontrol (3 scenarios aug.)✓0.771.811.982.193.083.963.290.522.553.763.104.105.372.81
• Freecontrol (6 scenarios aug.)✓6.817.547.675.545.455.102.191.313.638.595.358.866.265.72
• DriveGEN (Only Snow aug.)✓6.226.899.2715.0317.2418.8116.4013.389.8123.2811.5012.0413.9213.37
• DriveGEN (3 scenarios aug.)✓6.539.2611.1716.1519.6021.0218.6716.1713.7224.8014.5219.9819.3616.23
• DriveGEN (6 scenarios aug.)✓9.7013.6813.5217.0017.7420.3917.2318.7915.5023.0915.8020.9918.3417.06
+ +# 5. Experiments + +We conduct experiments to validate DriveGEN on monocular and multi-view 3D object detection, including four different training settings with various augmented OOD scenarios: 1) Augmented with traditional techniques (Traditional aug.); 2) Only augmented with Snow (Only Snow aug.); 3) Augmented with Snow, Rain and Fog (3 scenarios aug.); 4) Augmented with Snow, Rain, Fog, Night, Defocus and Sandstorm (6 scenarios aug.). + +Datasets. For monocular 3d object detection, we split the images of KITTI [10] into a training set (3712 images) and a validation set (3769 images) following the protocol from + +Monoflex [63], encompassing three classes: Car, Pedestrian, and Cyclist. We adopt the KITTI-C dataset [22] for validation, which includes 13 types of data corruptions on the validation set across four categories: Noise, Blur, Weather, and Digital [11]. + +As for multi-view 3D object detection, we first split real-world Night and Rainy validation scenarios based on descriptions of nuScenes [1] following [24]. Next, we augment 500 daytime training scenes under the Snow condition. The augmented detectors are then evaluated on real-world Night and Rainy scenarios, as well as on the widely used Robo3D benchmark [52]. More details of datasets are provided in Appendix B. + +![](images/409d28a1f3ac3a5e981ced131940fcf6503911c35a3773f84cb90ffafc2b14da.jpg) +Figure 4. Based on MonoCD [54], we provide more comparisons with KITTI-C, regarding Mean $AP_{3D|R_{40}}$ . + +Table 2. Detection results on nuScenes-C and real-world Night and Rainy scenarios of nuScenes, regarding mAP and NDS. + +
MetricMethodnuScenes-CReal-world Scenarios
BrightnessCameraCrashColorQuantFogFrameLostLowLightMotionBlurSnowAvg.nuScenes-NightnuScenes-RainyAvg.
mAPBEVFormer-tiny24.2622.0123.9121.9820.4316.1120.7810.8320.049.3126.7618.04
• DriveGEN (3k Snow)26.0424.1325.7824.1321.1317.2022.3911.7221.57(↑1.53)12.1929.1120.65 (↑2.61)
NDSBEVFormer-tiny34.9333.0634.5633.0431.4228.4231.8123.2731.3117.9237.3527.64
• DriveGEN (3k Snow)37.1435.3436.9335.3533.3230.4234.2324.0933.35(↑2.04)19.3740.9530.16 (↑2.52)
+ +Implementation Details. We implement our method and other baselines in PyTorch [29]. Following Freecontrol [25], we adopt keys from the self-attention of the first U-Net decoder layer as the features. In the self-prototype extraction stage, we run DDIM inversion with 200 steps. Subsequently, we also run 200 steps of DDIM sampling in the prototype-guided diffusion stage. We set $s$ and $\tau$ to 7.5 and 0.3 following FreeControl [25], while $\sigma$ is set to 0.1. All 3D detectors are trained on a mix of original and augmented data. All results in the manuscript are based on Stable Diffusion (SD) 1.5 [35]. More details and the results of SD 2.1 and XL 1.0 are put in Appendix C. + +Compared Methods. Based on typical or state-of-the-art (SOTA) 3D detectors [20, 31, 54, 63], we fully compare DriveGEN with following methods: 1) source-only, i.e., directly apply the well-trained model to corrupted test data; 2) Traditional data augmentation, i.e., Color Jitter and Brightness; 3) training-based controllable T2I diffusion: ControlNet [59] trains auxiliary modules for spatial control; 4) training-free controllable T2I diffusion: Freecontrol [25] enables zero-shot control of pretrained diffusion models. + +Evaluation Protocols. We report the experimental results in the Average Precision (AP) for 3D bounding boxes, denoted as $AP_{3D|R_{40}}$ . On the KITTI-C dataset, the results present the mean values of three difficulty levels and Intersection over Union (IoU) thresholds are set to 0.7, 0.5, 0.5 for Cars and 0.5, 0.25, 0.25 for Pedestrians and Cyclists. + +As for nuScenes, the mean average precision (mAP) and nuScenes detection score (NDS) are calculated following BEVFormer [20]. The metrics also include five true positive metrics, including ATE, ASE, AOE, AVE, and AAE for measuring errors in translation, scale, orientation, velocity, and attributes. + +# 5.1. Comparisons with Previous Methods + +For monocular 3d object detection, Table 1 gives the following observations: 1) Applying the well-trained detector to corrupted scenarios suffers significant performance degradation due to the data distribution shifts and traditional data augmentation offers limited gains. 2) ControlNet (pretrained on ade20k [70]) fails to enhance 3d detectors with limited training data even if it is provided with accurate segmentation masks [34] and rich prompts [6] for fine-tuning. 3) Freecontrol achieves minor gains in a few cases but generally degrades, indicating that training-free T2I diffusion is a feasible solution but requires careful object preservation. 4) DriveGEN consistently outperforms all compared methods over all categories within various base models for both known and unseen scenarios, like improving MonoFlex by an average of $7.6\mathrm{mAP}$ across 13 various OOD scenarios with the single Snow scenario augmentation. + +As for multi-view 3D object detection, we randomly select 3k daytime training images and apply Snow augmentation (3k Snow) to enhance the detector [20]. Table 2 shows DriveGEN consistently enhances BEVFormer-tiny across 8 OOD scenarios with an average of $1.53\mathrm{mAP}$ and 2.04 NDS, further confirming our effectiveness. Moreover, we validate the detector on real Night and Rainy scenarios of nuScenes. With the help of DriveGEN, it also achieves notable gains in real-world tasks with an average improvement of $2.61\mathrm{mAP}$ and 2.52 NDS, further demonstrating our superiority. More detailed results are put in the Appendix D. + +# 5.2. More Augmented Scenarios for Training + +An intuitive concern is whether the robustness of detectors progressively improves with the increasing of additional augmented scenarios (i.e., Only Snow aug. $\rightarrow 3$ scenarios + +![](images/e9f4557a8a6b85aaaccd5c3dc8551acd280a52881c407ab09a85a1631be846da.jpg) +Figure 5. Ablation studies on semantic feature alignment loss $g_{sa}$ , self-prototypes $\hat{\mathbf{P}}_t$ and shallow feature alignment loss $g_{sl}$ . + +![](images/e9e976d8ab81f03eca7bcbdf2bdbca6ee1396e6a104efa635b40e427d48b212c.jpg) +Figure 6. Qualitative results of DriveGEN for multi-view 3D object detection on nuScenes. DriveGEN supports vision-centric 3D detection tasks since our method only requires input images and corresponding object annotations without any additional diffusion model training. + +aug. $\rightarrow 6$ scenarios aug.). Based on Table 1 and Figure 4, we are given additional observations: 1) For ControlNet, the model performance drop for OOD scenarios grows as the number of augmentation scenarios increases, demonstrating that severe object omission leads to noisy training data. 2) For Freecontrol, the model performance across various OOD scenarios remains stable or tends to decline, indicating its object preservation may be unstable. 3) Owing to the precise object preservation, DriveGEN obtains continuous overall performance (i.e., Avg.) improvement across all categories with the incorporation of OOD scenarios. + +# 5.3. Ablation Studies + +To examine DriveGEN, we present qualitative results guided by different loss settings. As shown in Figure 5, compared with only generated via text prompts, introducing semantic-aware feature alignment $g_{sa}$ helps preserve more image structures. Next, by guiding the diffusion process with our self-prototypes $\hat{\mathbf{P}}_t$ , we achieve more accurate object geometry, alleviating potential object geometry loss. Eventually, introducing our shallow feature alignment $g_{sl}$ further improves object preservation with precise 3D geometry, especially for small objects (highlighted by red circles). More ablation studies are provided in Appendix D. + +# 5.4. Quantitative Results + +We provide visualizations for multi-view 3D object detection on nuScenes in Snow scenarios as shown in Figure 6. It is evident that DriveGEN supports existing vision-centric 3D detection tasks since it requires only input images and corresponding object annotations without any additional diffusion model training, demonstrating that DriveGEN achieves superior results even in challenging multiview tasks. More visualizations are put in Appendix E. + +# 6. Conclusion + +In this paper, we propose a method for robust 3D detection in driving via training-free controllable text-to-image diffusion generation, namely DriveGEN. Specifically, DriveGEN consists of: 1) Self-Prototype Extraction: To improve the guidance of self-attention features, we extract the self-prototypes by layouts to capture accurate object geometry, leveraging more precise features to guide diffusion. 2) Prototype-Guided Diffusion: To further preserve 3D object geometry, we conduct semantic-aware feature and shallow feature alignment during diffusion, alleviating the object misalignment and omission issues compared with previous methods. Experiments on KITTI-C and nuScenes demonstrate the effectiveness of DriveGEN in improving model robustness for vision-centric 3D detection. + +# Acknowledgements + +This work was supported by NSFC with Grant No. 62293482, by the Basic Research Project No. HZQB-KCZYZ-2021067 of Hetao Shenzhen HK S&T Cooperation Zone, by Shenzhen General Program No. JCYJ20220530143600001, by Shenzhen-Hong Kong Joint Funding No. SGDX20211123112401002, by the Shenzhen Outstanding Talents Training Fund 202002, by Guangdong Research Project No. 2017ZT07X152 and No. 2019CX01X104, by the Guangdong Provincial Key Laboratory of Future Networks of Intelligence (Grant No. 2022B1212010001), by the Guangdong Provincial Key Laboratory of Big Data Computing, CHUK-Shenzhen, by the NSFC 61931024&12326610, by the Key Area R&D Program of Guangdong Province with grant No. 2018B030338001, by the Shenzhen Key Laboratory of Big Data and Artificial Intelligence (Grant No. ZDSYS201707251409055), by Shaanxi Mathematical Basic Science Research Project(No.23JSY047), and by Tencent & Huawei Open Fund, by China Association for Science and Technology Youth Care Program. + +# References + +[1] Holger Caesar, Varun Bankiti, Alex H Lang, and et.al. nuscene: A multimodal dataset for autonomous driving. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 11621-11631, 2020. 6 +[2] Changjian Chen, Fei Lv, Yalong Guan, Pengcheng Wang, Shengjie Yu, Yifan Zhang, and Zhuo Tang. Human-guided image generation for expanding small-scale training image datasets. arXiv preprint arXiv:2412.16839, 2024. 3 +[3] Xiaozhi Chen, Kaustav Kundu, Yukun Zhu, and et. al. 3d object proposals for accurate object class detection. Advances in neural information processing systems, 28, 2015. 2 +[4] Xiaozhi Chen, Huimin Ma, Ji Wan, Bo Li, and Tian Xia. Multi-view 3d object detection network for autonomous driving. In Proceedings of the IEEE conference on Computer Vision and Pattern Recognition, pages 1907-1915, 2017. +[5] Yukang Chen, Jianhui Liu, Xiangyu Zhang, Xiaojuan Qi, and Jiaya Jia. Voxelnext: Fully sparse voxelnet for 3d object detection and tracking. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 21674-21683, 2023. 2 +[6] Zhe Chen, Jiannan Wu, Wenhai Wang, Weijie Su, Guo Chen, Sen Xing, Muyan Zhong, Qinglong Zhang, Xizhou Zhu, Lewei Lu, et al. Internvl: Scaling up vision foundation models and aligning for generic visual-linguistic tasks. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 24185-24198, 2024. 3, 7 +[7] Prafulla Dhariwal and Alexander Nichol. Diffusion models beat gans on image synthesis. Advances in neural information processing systems, 34:8780-8794, 2021. 3, 1 +[8] Dave Epstein, Allan Jabri, Ben Poole, Alexei Efros, and Aleksander Holynski. Diffusion self-guidance for control + +able image generation. Advances in Neural Information Processing Systems, 36:16222-16239, 2023. 3 +[9] Songwei Ge, Taesung Park, Jun-Yan Zhu, and Jia-Bin Huang. Expressive text-to-image generation with rich text. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 7545-7556, 2023. 5 +[10] Andreas Geiger, Philip Lenz, and Raquel Urtasun. Are we ready for autonomous driving? the kitti vision benchmark suite. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 3354-3361. IEEE, 2012. 1, 6 +[11] Dan Hendrycks and Thomas Dietterich. Benchmarking neural network robustness to common corruptions and perturbations. In International Conference on Learning Representations, 2018. 6 +[12] Jonathan Ho and Tim Salimans. Classifier-free diffusion guidance. arXiv preprint arXiv:2207.12598, 2022. 5 +[13] Jonathan Ho, Ajay Jain, and Pieter Abbeel. Denoising diffusion probabilistic models. Advances in neural information processing systems, 33:6840-6851, 2020. 3, 1 +[14] Yuming Hong, Weiyao Xu, Yi Yang, Hao Zhou, and Xiaoxue Wang. Magicdrive: 3d controllable driving video synthesis using layout and 3d representations. Proceedings of the IEEE/CVF International Conference on Computer Vision (ICCV), 2021. 3 +[15] Xinyu Huang, Mingyu Chen, Yunqi Li, Hao Wang, and Yuxing Zhao. Diffumask: Synthetic image and semantic mask annotation using cross-attention in diffusion models. arXiv preprint arXiv:2211.09123, 2022. 3 +[16] Junsoo Kim, Yeongjae Ye, Dongheon Cho, Jaechan Seo, and Jiseob Jeong. Drivingdiffusion: Realistic multi-view video generation with diffusion models. Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), 2021. 3 +[17] Feng Li, Yuan Zhang, Xinyang Wang, and Meng Liu. Vpd: Vision-based perception diffusion models for human pose estimation and dense visual annotations. arXiv preprint arXiv:2109.05270, 2021. 3 +[18] Hongyu Li, Pengfei Zhao, Wei Jiang, and Zihan Chen. Sensor fusion with diffusion models for enhanced 3d object detection. arXiv preprint arXiv:2205.02574, 2022. 1 +[19] Peiliang Li, Xiaozhi Chen, and Shaojie Shen. Stereo r-cnn based 3d object detection for autonomous driving. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 7644-7652, 2019. 2 +[20] Zhiqi Li, Wenhai Wang, Hongyang Li, Enze Xie, Chonghao Sima, Tong Lu, Yu Qiao, and Jifeng Dai. Bevformer: Learning bird's-eye-view representation from multi-camera images via spatiotemporal transformers. In European conference on computer vision, pages 1-18. Springer, 2022. 7, 2, 4 +[21] Hongbin Lin, Yifan Zhang, Zhen Qiu, Shuaicheng Niu, Chuang Gan, Yanxia Liu, and Mingkui Tan. Prototype-guided continual adaptation for class-incremental unsupervised domain adaptation. In European Conference on Computer Vision, pages 351-368. Springer, 2022. 2 + +[22] Hongbin Lin, Yifan Zhang, Shuaicheng Niu, Shuguang Cui, and Zhen Li. Monotta: Fully test-time adaptation for monocular 3d object detection. In European Conference on Computer Vision, pages 96-114. Springer, 2025. 2, 3, 6, 1, 4 +[23] Xiaoyang Liu, Yikang Wang, Xinyu Zhang, Wei Ma, Chao Wang, and Yun Chen. Multi-modal fusion for 3d object detection in autonomous driving. arXiv preprint arXiv:2204.01234, 2022. 1 +[24] Zhijian Liu, Haotian Tang, Alexander Amini, Xinyu Yang, Huizi Mao, Daniela L Rus, and Song Han. Bevfusion: Multi-task multi-sensor fusion with unified bird's-eye view representation. In 2023 IEEE international conference on robotics and automation (ICRA), pages 2774-2781. IEEE, 2023. 6, 1 +[25] Sicheng Mo, Fangzhou Mu, Kuan Heng Lin, Yanli Liu, Bochen Guan, Yin Li, and Bolei Zhou. Freecontrol: Training-free spatial control of any text-to-image diffusion model with any condition. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 7465-7475, 2024. 2, 3, 4, 5, 7 +[26] Seokhwa Moon, Sanghyuk Hwang, Hyunsu Lee, and Minje Kim. Temporal fusion with attention mechanisms for driving scenarios. Proceedings of the IEEE/CVF International Conference on Computer Vision (ICCV), 2022. 1 +[27] Chong Mou, Xintao Wang, Liangbin Xie, Yanze Wu, Jian Zhang, Zhongang Qi, and Ying Shan. T2i-adapter: Learning adapters to dig out more controllable ability for text-to-image diffusion models. In Proceedings of the AAAI Conference on Artificial Intelligence, pages 4296–4304, 2024. 2, 4 +[28] Youngmin Oh, Hyung-II Kim, Seong Tae Kim, and Jung Uk Kim. Monowad: Weather-adaptive diffusion model for robust monocular 3d object detection. In European Conference on Computer Vision, 2025. 2, 3 +[29] Adam Paszke, Sam Gross, Francisco Massa, Adam Lerer, James Bradbury, Gregory Chanan, Trevor Killeen, Zeming Lin, Natalia Gimelshein, Luca Antiga, et al. Pytorch: An imperative style, high-performance deep learning library. In Advances in neural information processing systems, 2019. 7, 2 +[30] Yuheng Qin, Zhaoyang Zhang, Han Zhang, Dong Li, Baining Chen, Lu Zhang, Jiuming Gu, and Dongdong Zhang. Uncontrol: A unified diffusion model for controllable visual generation in the wild. arXiv preprint arXiv:2305.11147, 2023. 3 +[31] Zequn Qin and Xi Li. Monoground: Detecting monocular 3d objects from the ground. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 3793-3802, 2022. 6, 7, 4 +[32] Zhen Qiu, Yifan Zhang, Hongbin Lin, Shuaicheng Niu, Yanxia Liu, Qing Du, and Mingkui Tan. Source-free domain adaptation via avatar prototype generation and adaptation. arXiv preprint arXiv:2106.15326, 2021. 2 +[33] Aditya Ramesh, Prafulla Dhariwal, Alex Nichol, Casey Chu, and Mark Chen. Hierarchical text-conditional image generation with clip latents. arXiv preprint arXiv:2204.06125, 2022. 3 +[34] Nikhila Ravi, Valentin Gabeur, Yuan-Ting Hu, Ronghang Hu, Chaitanya Ryali, Tengyu Ma, Haitham Khedr, Roman + +Rädle, Chloe Rolland, Laura Gustafson, et al. Sam 2: Segment anything in images and videos. arXiv preprint arXiv:2408.00714, 2024. 3, 7 +[35] Robin Rombach, Andreas Blattmann, Dominik Lorenz, Patrick Esser, and Björn Ommer. High-resolution image synthesis with latent diffusion models. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 10684-10695, 2022. 3, 5, 7, 1 +[36] Olaf Ronneberger, Philipp Fischer, and Thomas Brox. U-net: Convolutional networks for biomedical image segmentation. In Medical image computing and computer-assisted intervention-MICCAI 2015: 18th international conference, Munich, Germany, October 5-9, 2015, proceedings, part III 18, pages 234-241. Springer, 2015. 3, 2 +[37] Chitwan Sahara, William Chan, Saurabh Saxena, Lala Li, Jay Whang, Emily Denton, Seyed Kamyar Seyed Ghasemipour, Burcu Karagol Ayan, Tim Salimans, Jonathan Ho, et al. Imagen: Photorealistic text-to-image diffusion models. arXiv preprint arXiv:2205.11487, 2022. 3 +[38] Jascha Sohl-Dickstein, Eric A Weiss, Niru Maheswaranathan, and Surya Ganguli. Deep unsupervised learning using nonequilibrium thermodynamics. Proceedings of the 32nd International Conference on Machine Learning, 2015. 1 +[39] Jiaming Song, Chenlin Meng, and Stefano Ermon. Denoising diffusion implicit models. arXiv preprint arXiv:2010.02502, 2020. 5 +[40] Yang Song, Jascha Sohl-Dickstein, Diederik P Kingma, Abhishek Kumar, Stefano Ermon, and Ben Poole. Score-based generative modeling through stochastic differential equations. arXiv preprint arXiv:2011.13456, 2021. 1 +[41] Yang Song, Jascha Sohl-Dickstein, Diederik P Kingma, Abhishek Kumar, Stefano Ermon, and Ben Poole. Score-based generative modeling through stochastic differential equations. In International Conference on Learning Representations, 2021. 3 +[42] Ziying Song, Guoxing Zhang, Lin Liu, Lei Yang, Shaoqing Xu, Caiyan Jia, Feiyang Jia, and Li Wang. Robofusion: towards robust multi-modal 3d object detection via sam. In Proceedings of the Thirty-Third International Joint Conference on Artificial Intelligence, pages 1272–1280, 2024. 3 +[43] Lin Sun, Liangliang Cao, Yang Zheng, Bo Wang, Zhiming Li, Wenzhe Xie, and Xu Sun. Panacea: A framework for diverse, controllable generation of urban scenes. arXiv preprint arXiv:2207.10701, 2022. 3 +[44] Alexander Swerdlow, Runsheng Xu, and Bolei Zhou. Streetview image generation from a bird's-eye view layout. IEEE Robotics and Automation Letters, 2024. 2 +[45] Sergey Tulyakov, Ming-Yu Liu, Xiaodong Yang, and Jan Kautz. Mocogan: Decomposing motion and content for video generation. Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition (CVPR), 2018. 1 +[46] Narek Tumanyan, Michal Geyer, Shai Bagon, and Tali Dekel. Plug-and-play diffusion features for text-driven image-to-image translation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 1921-1930, 2023. 2, 3, 4, 5 + +[47] A Vaswani. Attention is all you need. Advances in Neural Information Processing Systems, 2017. 3 +[48] Pascal Vincent, Hugo Larochelle, Yoshua Bengio, and Pierre-Antoine Manzagol. Extracting and composing robust features with denoising autoencoders. Proceedings of the 25th International Conference on Machine Learning, 2008. 1 +[49] Yan Wang, Wei-Lun Chao, Divyansh Garg, and et. al. Pseudo-lidar from visual depth estimation: Bridging the gap in 3d object detection for autonomous driving. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 8445-8453, 2019. 2 +[50] Hai Wu, Chenglu Wen, Shaoshuai Shi, Xin Li, and Cheng Wang. Virtual sparse convolution for multimodal 3d object detection. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 21653-21662, 2023. 2 +[51] Jinheng Xie, Yuexiang Li, Yawen Huang, Haozhe Liu, Wentian Zhang, Yefeng Zheng, and Mike Zheng Shou. Boxdiff: Text-to-image synthesis with training-free box-constrained diffusion. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 7452-7461, 2023. 3 +[52] Shaoyuan Xie, Lingdong Kong, Wenwei Zhang, Jiawei Ren, Liang Pan, Kai Chen, and Ziwei Liu. Benchmarking and improving bird's eye view perception robustness in autonomous driving. IEEE Transactions on Pattern Analysis and Machine Intelligence, 2025. 2, 6 +[53] Junkai Xu, Liang Peng, Haoran Cheng, Hao Li, Wei Qian, Ke Li, Wenxiao Wang, and Deng Cai. Mononerd: Nerflike representations for monocular 3d object detection. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 6814-6824, 2023. 2 +[54] Longfei Yan, Pei Yan, Shengzhou Xiong, Xuanyu Xiang, and Yihua Tan. Monocd: Monocular 3d object detection with complementary depths. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 10248-10257, 2024. 7 +[55] Xu Yan, Weiming Zeng, Jiaming Zhu, Hailin Chen, Yaowei Li, and Deng Cai. Videogpt: Video generation using vq-vae and transformers. arXiv preprint arXiv:2104.10157, 2021. 1 +[56] Chenyu Yang, Yuntao Chen, Hao Tian, Chenxin Tao, Xizhou Zhu, Zhaoxiang Zhang, Gao Huang, Hongyang Li, Yu Qiao, Lewei Lu, et al. Bevformer v2: Adapting modern image backbones to bird's-eye-view recognition via perspective supervision. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 17830-17839, 2023. 2 +[57] Kairui Yang, Enhui Ma, Jibin Peng, Qing Guo, Di Lin, and Kaicheng Yu. Bevcontrol: Accurately controlling street-view elements with multi-perspective consistency via bev sketch layout. arXiv preprint arXiv:2308.01661, 2023. 2 +[58] Tianwei Yin, Xingyi Zhou, and Philipp Krahenbuhl. Center-based 3d object detection and tracking. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 11784-11793, 2021. 2 +[59] Lvmin Zhang, Anyi Rao, and Maneesh Agrawala. Adding conditional control to text-to-image diffusion models. In + +Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 3836-3847, 2023. 2, 3, 4, 7 +[60] Renrui Zhang, Han Qiu, Tai Wang, Ziyu Guo, Ziteng Cui, Yu Qiao, Hongsheng Li, and Peng Gao. Monodetr: Depth-guided transformer for monocular 3d object detection. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 9155-9166, 2023. 2 +[61] Yifan Zhang and Bryan Hooi. Hipa: enabling one-step text-to-image diffusion models via high-frequency-promoting adaptation. arXiv preprint arXiv:2311.18158, 2023. 3 +[62] Yifan Zhang, Ying Wei, Qingyao Wu, Peilin Zhao, Shuaicheng Niu, Junzhou Huang, and Mingkui Tan. Collaborative unsupervised domain adaptation for medical image diagnosis. IEEE Transactions on Image Processing, 29: 7834-7844, 2020. 2 +[63] Yunpeng Zhang, Jiwen Lu, and Jie Zhou. Objects are different: Flexible monocular 3d object detection. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 3289-3298, 2021. 1, 2, 6, 7, 3, 4 +[64] Yifan Zhang, Bingyi Kang, Bryan Hooi, Shuicheng Yan, and Jiashi Feng. Deep long-tailed learning: A survey. IEEE transactions on pattern analysis and machine intelligence, 45(9):10795-10816, 2023. 2 +[65] Yifan Zhang, Daquan Zhou, Bryan Hooi, Kai Wang, and Ji-ashi Feng. Expanding small-scale datasets with guided imagination. In Advances in neural information processing systems, pages 76558–76618, 2023. 3 +[66] Bowen Zhao, Feng Lu, Cheng Zhang, Jiaqi Hu, and Lin Song. Drivedreamer: Realistic multi-view video generation of driving scenes with future prediction. Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition (CVPR), 2022. 3 +[67] Shihao Zhao, Dongdong Chen, Yen-Chun Chen, Jianmin Bao, Dongdong Zhang, Lu Yuan, Han Zhang, Dong Li, Baining Chen, and Lu Zhang. Uni-controlnet: All-in-one control to text-to-image diffusion models. arXiv preprint arXiv:2305.16322, 2023. 3 +[68] Guangcong Zheng, Xianpan Zhou, Xuewei Li, Zhongang Qi, Ying Shan, and Xi Li. Layoutdiffusion: Controllable diffusion model for layout-to-image generation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 22490-22499, 2023. 3 +[69] Longtao Zheng, Yifan Zhang, Hanzhong Guo, Jiachun Pan, Zhenxiong Tan, Jiahao Lu, Chuanxin Tang, Bo An, and Shuicheng Yan. Memo: Memory-guided diffusion for expressive talking video generation. arXiv preprint arXiv:2412.04448, 2024. 3 +[70] Bolei Zhou, Hang Zhao, Xavier Puig, Sanja Fidler, Adela Barriuso, and Antonio Torralba. Scene parsing through ade20k dataset. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 633-641, 2017. 7 \ No newline at end of file diff --git a/2025/DriveGEN_ Generalized and Robust 3D Detection in Driving via Controllable Text-to-Image Diffusion Generation/images.zip b/2025/DriveGEN_ Generalized and Robust 3D Detection in Driving via Controllable Text-to-Image Diffusion Generation/images.zip new file mode 100644 index 0000000000000000000000000000000000000000..88b90187c8ecbc8de34f32b902bc9f72e386db5d --- /dev/null +++ b/2025/DriveGEN_ Generalized and Robust 3D Detection in Driving via Controllable Text-to-Image Diffusion Generation/images.zip @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:4cdfb65c8b75dc32ca9affb7dde485dc3cd44c9007cb4346898bbfa7e3eab190 +size 1014219 diff --git a/2025/DriveGEN_ Generalized and Robust 3D Detection in Driving via Controllable Text-to-Image Diffusion Generation/layout.json b/2025/DriveGEN_ Generalized and Robust 3D Detection in Driving via Controllable Text-to-Image Diffusion Generation/layout.json new file mode 100644 index 0000000000000000000000000000000000000000..20060d5be71e30332e90aca51c232156b82446f2 --- /dev/null +++ b/2025/DriveGEN_ Generalized and Robust 3D Detection in Driving via Controllable Text-to-Image Diffusion Generation/layout.json @@ -0,0 +1,8951 @@ +{ + "pdf_info": [ + { + "para_blocks": [ + { + "bbox": [ + 119, + 103, + 493, + 140 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 119, + 103, + 493, + 140 + ], + "spans": [ + { + "bbox": [ + 119, + 103, + 493, + 140 + ], + "type": "text", + "content": "DriveGEN: Generalized and Robust 3D Detection in Driving via Controllable Text-to-Image Diffusion Generation" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 142, + 160, + 465, + 190 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 142, + 160, + 465, + 190 + ], + "spans": [ + { + "bbox": [ + 142, + 160, + 465, + 190 + ], + "type": "text", + "content": "Hongbin Lin" + }, + { + "bbox": [ + 142, + 160, + 465, + 190 + ], + "type": "inline_equation", + "content": "^{1,2}" + }, + { + "bbox": [ + 142, + 160, + 465, + 190 + ], + "type": "text", + "content": " Zilu Guo" + }, + { + "bbox": [ + 142, + 160, + 465, + 190 + ], + "type": "inline_equation", + "content": "^{1,2}" + }, + { + "bbox": [ + 142, + 160, + 465, + 190 + ], + "type": "text", + "content": " Yifan Zhang" + }, + { + "bbox": [ + 142, + 160, + 465, + 190 + ], + "type": "inline_equation", + "content": "^{3}" + }, + { + "bbox": [ + 142, + 160, + 465, + 190 + ], + "type": "text", + "content": " Shuaicheng Niu" + }, + { + "bbox": [ + 142, + 160, + 465, + 190 + ], + "type": "inline_equation", + "content": "^{4}" + }, + { + "bbox": [ + 142, + 160, + 465, + 190 + ], + "type": "text", + "content": " Yafeng Li" + }, + { + "bbox": [ + 142, + 160, + 465, + 190 + ], + "type": "inline_equation", + "content": "^{5}" + }, + { + "bbox": [ + 142, + 160, + 465, + 190 + ], + "type": "text", + "content": " Ruimao Zhang" + }, + { + "bbox": [ + 142, + 160, + 465, + 190 + ], + "type": "inline_equation", + "content": "^{6}" + }, + { + "bbox": [ + 142, + 160, + 465, + 190 + ], + "type": "text", + "content": " Shuguang Cui" + }, + { + "bbox": [ + 142, + 160, + 465, + 190 + ], + "type": "inline_equation", + "content": "^{2,1}" + }, + { + "bbox": [ + 142, + 160, + 465, + 190 + ], + "type": "text", + "content": " Zhen Li" + }, + { + "bbox": [ + 142, + 160, + 465, + 190 + ], + "type": "inline_equation", + "content": "^{2,1*}" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 62, + 190, + 547, + 218 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 190, + 547, + 218 + ], + "spans": [ + { + "bbox": [ + 62, + 190, + 547, + 218 + ], + "type": "inline_equation", + "content": "^{1}" + }, + { + "bbox": [ + 62, + 190, + 547, + 218 + ], + "type": "text", + "content": "FNii-Shenzhen " + }, + { + "bbox": [ + 62, + 190, + 547, + 218 + ], + "type": "inline_equation", + "content": "^{2}" + }, + { + "bbox": [ + 62, + 190, + 547, + 218 + ], + "type": "text", + "content": " SSE, CUHK-Shenzhen " + }, + { + "bbox": [ + 62, + 190, + 547, + 218 + ], + "type": "inline_equation", + "content": "^{3}" + }, + { + "bbox": [ + 62, + 190, + 547, + 218 + ], + "type": "text", + "content": " National University of Singapore " + }, + { + "bbox": [ + 62, + 190, + 547, + 218 + ], + "type": "inline_equation", + "content": "^{4}" + }, + { + "bbox": [ + 62, + 190, + 547, + 218 + ], + "type": "text", + "content": " Nanyang Technological University " + }, + { + "bbox": [ + 62, + 190, + 547, + 218 + ], + "type": "inline_equation", + "content": "^{5}" + }, + { + "bbox": [ + 62, + 190, + 547, + 218 + ], + "type": "text", + "content": "Baoji University of Arts and Sciences " + }, + { + "bbox": [ + 62, + 190, + 547, + 218 + ], + "type": "inline_equation", + "content": "^{6}" + }, + { + "bbox": [ + 62, + 190, + 547, + 218 + ], + "type": "text", + "content": "Sun Yat-sen University" + } + ] + } + ], + "index": 4 + }, + { + "type": "image", + "bbox": [ + 85, + 227, + 526, + 462 + ], + "blocks": [ + { + "bbox": [ + 85, + 227, + 526, + 462 + ], + "lines": [ + { + "bbox": [ + 85, + 227, + 526, + 462 + ], + "spans": [ + { + "bbox": [ + 85, + 227, + 526, + 462 + ], + "type": "image", + "image_path": "f4971ec62903b3190e1607d19ff2aa63e8d36e158741595d7c392d6280f7eebb.jpg" + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 55, + 473, + 555, + 518 + ], + "lines": [ + { + "bbox": [ + 55, + 473, + 555, + 518 + ], + "spans": [ + { + "bbox": [ + 55, + 473, + 555, + 518 + ], + "type": "text", + "content": "Figure 1. An illustration of DriveGEN for enhancing vision-centric 3D detectors. DriveGEN extends original training images to diverse Out-of-Distribution (OOD) scenarios without additional diffusion model training, preserving all annotated objects. Even with a single augmentation, i.e., 'Snow', the augmented detector [63] is improved not only on the original KITTI [10] val set (Ideal) and already known 'Snow' scenario, but also in unseen 'Fog' and 'Defocus' scenarios, demonstrating a comprehensive model generalizability improvement." + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "image_caption" + } + ], + "index": 5 + }, + { + "bbox": [ + 152, + 544, + 200, + 556 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 152, + 544, + 200, + 556 + ], + "spans": [ + { + "bbox": [ + 152, + 544, + 200, + 556 + ], + "type": "text", + "content": "Abstract" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 55, + 571, + 297, + 691 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 571, + 297, + 691 + ], + "spans": [ + { + "bbox": [ + 55, + 571, + 297, + 691 + ], + "type": "text", + "content": "In autonomous driving, vision-centric 3D detection aims to identify 3D objects from images. However, high data collection costs and diverse real-world scenarios limit the scale of training data. Once distribution shifts occur between training and test data, existing methods often suffer from performance degradation, known as Out-of-Distribution (OOD) problems. To address this, controllable Text-to-Image (T2I) diffusion offers a potential solution for training data enhancement, which is required to generate diverse OOD scenarios with precise 3D object geometry. Nevertheless, exist" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 313, + 545, + 556, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 545, + 556, + 715 + ], + "spans": [ + { + "bbox": [ + 313, + 545, + 556, + 715 + ], + "type": "text", + "content": "ing controllable T2I approaches are restricted by the limited scale of training data or struggle to preserve all annotated 3D objects. In this paper, we present DriveGEN, a method designed to improve the robustness of 3D detectors in Driving via Training-Free Controllable Text-to-Image Diffusion Generation. Without extra diffusion model training, DriveGEN consistently preserves objects with precise 3D geometry across diverse OOD generations, consisting of 2 stages: 1) Self-Prototype Extraction: We empirically find that self-attention features are semantic-aware but require accurate region selection for 3D objects. Thus, we extract precise object features via layouts to capture 3D object geometry, termed self-prototypes. 2) Prototype-Guided Diffusion: To preserve objects across various OOD scenarios, we per" + } + ] + } + ], + "index": 9 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "spans": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "type": "text", + "content": "CVF" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 145, + 0, + 496, + 37 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 145, + 0, + 496, + 37 + ], + "spans": [ + { + "bbox": [ + 145, + 0, + 496, + 37 + ], + "type": "text", + "content": "This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 67, + 703, + 145, + 713 + ], + "type": "page_footnote", + "angle": 0, + "lines": [ + { + "bbox": [ + 67, + 703, + 145, + 713 + ], + "spans": [ + { + "bbox": [ + 67, + 703, + 145, + 713 + ], + "type": "text", + "content": "*Corresponding author." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "type": "text", + "content": "27497" + } + ] + } + ], + "index": 11 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 0 + }, + { + "para_blocks": [ + { + "bbox": [ + 54, + 72, + 294, + 120 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 54, + 72, + 294, + 120 + ], + "spans": [ + { + "bbox": [ + 54, + 72, + 294, + 120 + ], + "type": "text", + "content": "form semantic-aware feature alignment and shallow feature alignment during denoising. Extensive experiments demonstrate our effectiveness in improving 3D detection. The code is available at github.com/Hongbin98/DriveGEN." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 55, + 143, + 136, + 156 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 143, + 136, + 156 + ], + "spans": [ + { + "bbox": [ + 55, + 143, + 136, + 156 + ], + "type": "text", + "content": "1. Introduction" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 56, + 163, + 295, + 282 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 163, + 295, + 282 + ], + "spans": [ + { + "bbox": [ + 56, + 163, + 295, + 282 + ], + "type": "text", + "content": "Three-dimensional (3D) object detection intends to identify objects and assess their spatial and dimensional attributes via various sensor inputs, which is widely explored in Autonomous Driving (AD) and robotics [3-5, 19, 49, 50, 58]. To save sensor costs, there is a growing trend towards implementing perception systems by vision-centric 3D detection which relies solely on single or multi-view RGB images along with camera calibration information. Despite challenges in this field, existing methods [53, 56, 60] have demonstrated promising results across various benchmarks." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 56, + 283, + 295, + 486 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 283, + 295, + 486 + ], + "spans": [ + { + "bbox": [ + 56, + 283, + 295, + 486 + ], + "type": "text", + "content": "Behind the success, one prerequisite is that test images share the same distribution as training images. However, this assumption is often not held in diverse real-world scenarios [21, 32, 62, 64], especially for perception systems that are required for long durations, like autonomous vehicles navigating through different weather conditions. Once the environment changes, well-trained detectors may suffer performance degradation due to the existence of data distribution shifts between training images and test images. As shown in Figure 1, the original detector performs well in the ideal (sunny) scenario while failing to maintain stable performance in the Snow and Fog scenario due to the out-of-distribution (OOD) issue. Considering the widespread application of 3D detection, significant performance degradation on OOD test data may result in traffic accidents and serious safety risks. Thus, it is essential to address the OOD generalization problem of vision-centric 3D detection." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 56, + 487, + 295, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 487, + 295, + 713 + ], + "spans": [ + { + "bbox": [ + 56, + 487, + 295, + 713 + ], + "type": "text", + "content": "Recently, two attempts sought to solve the distribution shifts in AD by test-time model adaptation [22] or restoring OOD scenarios to ideal conditions with a weather-adaptive diffusion model [28]. However, they require additional consumption during the test phase, leading to additional energy costs while the driving range is crucial in AD. To save test-time computation costs, we explore controllable Text-to-Image (T2I) diffusion generation for training data augmentation, thereby improving the robustness of 3D detectors. Prior training-based controllable T2I approaches such as ControlNet [59] and T2i-adapter [27] offer users fine-grained spatial control but rely on substantial training data to train auxiliary modules, making them impractical for handling OOD scenarios in AD. As for training-free methods, they [25, 46] leverage low-resolution self-attention features to capture image structures for diffusion model guidance. However, these features are relatively coarse for preserving object geometric characteristics of different classes in 3D object detection. As shown in Figure 2, the object" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 313, + 72, + 553, + 144 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 72, + 553, + 144 + ], + "spans": [ + { + "bbox": [ + 313, + 72, + 553, + 144 + ], + "type": "text", + "content": "geometry may be lost after multiple rounds of downsampling " + }, + { + "bbox": [ + 313, + 72, + 553, + 144 + ], + "type": "inline_equation", + "content": "(32\\mathrm{x})" + }, + { + "bbox": [ + 313, + 72, + 553, + 144 + ], + "type": "text", + "content": ", leading to severe orientation error, position misalignment and omissions, particularly for tiny objects. Therefore, an eligible training-free controllable T2I diffusion method for training data enhancement in 3D detection should preserve all objects with precise 3D geometry." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 313, + 144, + 553, + 359 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 144, + 553, + 359 + ], + "spans": [ + { + "bbox": [ + 313, + 144, + 553, + 359 + ], + "type": "text", + "content": "To address these challenges, we propose a method namely DriveGEN for Generalized and Robust 3D Detection in Driving via Training-Free Controllable Text-to-Image Diffusion GENeration. DriveGEN enhances the model robustness within various OOD scenarios generation, consisting of two stages: 1) Self-Prototype Extraction. To capture accurate object geometric characteristics, we first extract principal components of self-attention features [25, 46] to obtain coarse semantic-aware features. Then, we adopt layouts (i.e., bounding boxes) to re-weight features in object regions via a peak function, termed self-prototypes. 2) Prototype-Guided Diffusion. To preserve objects with precise 3D geometry, we perform semantic-aware feature alignment with self-prototypes during the denoising phase. However, these semantic-aware features are relatively coarse to represent tiny objects as shown in Figure 2. Therefore, we devise a shallow feature alignment strategy to constrain the diffusion process and retrain tiny objects." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 313, + 360, + 553, + 538 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 360, + 553, + 538 + ], + "spans": [ + { + "bbox": [ + 313, + 360, + 553, + 538 + ], + "type": "text", + "content": "Contributions: 1) To the best of our knowledge, we are the first to explore training-free controllable T2I diffusion generation to enhance the model robustness of vision-centric 3D detection. Even with a single augmentation, DriveGEN improves the 3D detector [63] by an average of " + }, + { + "bbox": [ + 313, + 360, + 553, + 538 + ], + "type": "inline_equation", + "content": "7.6\\mathrm{mAP}" + }, + { + "bbox": [ + 313, + 360, + 553, + 538 + ], + "type": "text", + "content": " across 13 OOD scenarios. 2) With self-prototype extraction and prototype-guided diffusion stages, DriveGEN enables OOD scenario augmentation while preserving precise 3D geometry for all objects, enhancing the applicability of diffusion models in autonomous driving. 3) Extensive experiments show DriveGEN significantly improves existing monocular 3D detectors across 13 OOD corruption on KITTI and multi-view 3D detectors on the real scenarios (Night, Rainy) of nuScenes, demonstrating the effectiveness of DriveGEN in boosting model generalizability." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 314, + 550, + 400, + 562 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 314, + 550, + 400, + 562 + ], + "spans": [ + { + "bbox": [ + 314, + 550, + 400, + 562 + ], + "type": "text", + "content": "2. Related Work" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 313, + 570, + 553, + 617 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 570, + 553, + 617 + ], + "spans": [ + { + "bbox": [ + 313, + 570, + 553, + 617 + ], + "type": "text", + "content": "We first review the literature on explorations of model robustness in autonomous driving, and then discuss controllable T2I diffusion generation methods. More discussions on vision-centric 3D object detection are in Appendix A." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 313, + 618, + 553, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 618, + 553, + 713 + ], + "spans": [ + { + "bbox": [ + 313, + 618, + 553, + 713 + ], + "type": "text", + "content": "Model Robustness in Autonomous Driving. The reliability of vision-centric autonomous driving systems depends on the robustness of perception models, especially under challenging scenarios like adverse weather, varied lighting, and complex urban environments [52]. To improve robustness, existing methods employ diffusion models to generate synthetic driving scenes for training data augmentation. Specifically, BEVGen [44] and BEVControl [57] employ" + } + ] + } + ], + "index": 10 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "text", + "content": "27498" + } + ] + } + ], + "index": 11 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 1 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 58, + 71, + 553, + 143 + ], + "blocks": [ + { + "bbox": [ + 58, + 71, + 553, + 143 + ], + "lines": [ + { + "bbox": [ + 58, + 71, + 553, + 143 + ], + "spans": [ + { + "bbox": [ + 58, + 71, + 553, + 143 + ], + "type": "image", + "image_path": "2b113af7cb6544eac7c504aaefffdea8f96a57ef6a32ee4179d0ab7bbd2d0212.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 55, + 152, + 555, + 198 + ], + "lines": [ + { + "bbox": [ + 55, + 152, + 555, + 198 + ], + "spans": [ + { + "bbox": [ + 55, + 152, + 555, + 198 + ], + "type": "text", + "content": "Figure 2. Illustration of the importance of object preservation in 3d detection based on KITTI. ControlNet [59] suffers object orientation errors and omissions even if accurate segmentation masks [34] and rich prompts [6] are provided, showing that training-based methods may struggle with spatial control with limited training data. Additionally, Freecontrol [25] relies on coarse low-resolution features to capture semantic structures that may encounter potential object geometry loss, resulting in object position misalignment and omission issues." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "bbox": [ + 54, + 217, + 294, + 301 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 54, + 217, + 294, + 301 + ], + "spans": [ + { + "bbox": [ + 54, + 217, + 294, + 301 + ], + "type": "text", + "content": "Layouts to create flexible scene configurations. Innovative frameworks like Panacea [43] and MagicDrive [14] extend this controllability to multi-view scenes, while models such as DriveDreamer [66] and DrivingDiffusion [16] focus on generating realistic videos for model validation. Recently, RoboFusion [42] has sought to tackle OOD noise scenarios by leveraging visual foundation models like SAM." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 56, + 301, + 296, + 528 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 301, + 296, + 528 + ], + "spans": [ + { + "bbox": [ + 56, + 301, + 296, + 528 + ], + "type": "text", + "content": "Controllable T2I Diffusion. Controllable text-to-image (T2I) diffusion models [35, 61, 69] are increasingly applied in real-world scenarios, where text prompts or scene descriptions guide the creation of synthetic data. By leveraging pre-trained models like Stable Diffusion [35] or other large-scale T2I models [33, 37], existing methods can generate high-fidelity and diverse images based on textual input, allowing for controlled variations [2, 65]. Techniques such as Uni-ControlNet [67], UniControl [30] and Layout-diffusion [68] further enhance control by integrating multimodal inputs, creating contextually accurate and condition-specific scenes. Recent works also explore the utility of cross-attention mechanisms and latent embeddings to align generated images closely with text descriptions, supporting the generation for perception tasks like semantic segmentation and depth estimation [15, 17]. However, these methods require a large amount of training data to train auxiliary modules, making them impractical for handling OOD scenarios in AD for the limited data scale." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 55, + 529, + 295, + 649 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 529, + 295, + 649 + ], + "spans": [ + { + "bbox": [ + 55, + 529, + 295, + 649 + ], + "type": "text", + "content": "Another alternative choice is training-free controllable T2I diffusion, PnP [46] directly injects the self-attention features of condition images to guide the diffusion process while Freecontrol [25] utilizes PCA to extract principle components of self-attention features for semantic structures extraction. However, existing approaches solely rely on the relatively coarse self-attention features, which often results in object region omissions and misalignment as we mentioned before. To this end, we propose our DriveGEN to preserve all objects with precise 3D geometry." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 55, + 657, + 133, + 671 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 657, + 133, + 671 + ], + "spans": [ + { + "bbox": [ + 55, + 657, + 133, + 671 + ], + "type": "text", + "content": "3. Preliminary" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 55, + 677, + 295, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 677, + 295, + 715 + ], + "spans": [ + { + "bbox": [ + 55, + 677, + 295, + 715 + ], + "type": "text", + "content": "Pre-trained diffusion models [7, 13, 35] are probabilistic generative models, which consists of two complementary stochastic processes: 1) In the forward process, Gaussian" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 313, + 217, + 555, + 312 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 217, + 555, + 312 + ], + "spans": [ + { + "bbox": [ + 313, + 217, + 555, + 312 + ], + "type": "text", + "content": "noise is incrementally added to a clean image " + }, + { + "bbox": [ + 313, + 217, + 555, + 312 + ], + "type": "inline_equation", + "content": "\\mathbf{x}_0" + }, + { + "bbox": [ + 313, + 217, + 555, + 312 + ], + "type": "text", + "content": ". 2) As for the backward process, it conducts the iterative process of denoising the initial Gaussian noise image " + }, + { + "bbox": [ + 313, + 217, + 555, + 312 + ], + "type": "inline_equation", + "content": "\\mathbf{z}_{N_t}" + }, + { + "bbox": [ + 313, + 217, + 555, + 312 + ], + "type": "text", + "content": ", where a cleaner image is given at each step. To be specific, this process leverages a denoising network " + }, + { + "bbox": [ + 313, + 217, + 555, + 312 + ], + "type": "inline_equation", + "content": "\\epsilon_{\\theta}(\\mathbf{z}_t,t)" + }, + { + "bbox": [ + 313, + 217, + 555, + 312 + ], + "type": "text", + "content": " to estimate the added noise in the forward process. Subsequently, for " + }, + { + "bbox": [ + 313, + 217, + 555, + 312 + ], + "type": "inline_equation", + "content": "\\mathbf{z}_t" + }, + { + "bbox": [ + 313, + 217, + 555, + 312 + ], + "type": "text", + "content": " at each step, the network removes the estimated noise perturbation to achieve a cleaner " + }, + { + "bbox": [ + 313, + 217, + 555, + 312 + ], + "type": "inline_equation", + "content": "\\mathbf{z}_{t - 1}" + }, + { + "bbox": [ + 313, + 217, + 555, + 312 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 313, + 313, + 556, + 410 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 313, + 556, + 410 + ], + "spans": [ + { + "bbox": [ + 313, + 313, + 556, + 410 + ], + "type": "text", + "content": "For T2I diffusion models, a previous method [41] has shown that " + }, + { + "bbox": [ + 313, + 313, + 556, + 410 + ], + "type": "inline_equation", + "content": "\\epsilon_{\\theta}" + }, + { + "bbox": [ + 313, + 313, + 556, + 410 + ], + "type": "text", + "content": " can approximate the score function for the marginal distributions " + }, + { + "bbox": [ + 313, + 313, + 556, + 410 + ], + "type": "inline_equation", + "content": "p_t" + }, + { + "bbox": [ + 313, + 313, + 556, + 410 + ], + "type": "text", + "content": ", i.e., " + }, + { + "bbox": [ + 313, + 313, + 556, + 410 + ], + "type": "inline_equation", + "content": "-\\sigma_t\\nabla_{\\mathbf{z}_t}\\log p_t(\\mathbf{z}_t|\\mathbf{c})" + }, + { + "bbox": [ + 313, + 313, + 556, + 410 + ], + "type": "text", + "content": " where " + }, + { + "bbox": [ + 313, + 313, + 556, + 410 + ], + "type": "inline_equation", + "content": "\\sigma_t" + }, + { + "bbox": [ + 313, + 313, + 556, + 410 + ], + "type": "text", + "content": " denotes a noise schedule and " + }, + { + "bbox": [ + 313, + 313, + 556, + 410 + ], + "type": "inline_equation", + "content": "\\mathbf{c}" + }, + { + "bbox": [ + 313, + 313, + 556, + 410 + ], + "type": "text", + "content": " is a text prompt. To incorporate auxiliary information " + }, + { + "bbox": [ + 313, + 313, + 556, + 410 + ], + "type": "inline_equation", + "content": "y" + }, + { + "bbox": [ + 313, + 313, + 556, + 410 + ], + "type": "text", + "content": " into the diffusion sampling process, existing methods [7, 8, 25] guide the diffusion process by adding a time-dependent energy function termed " + }, + { + "bbox": [ + 313, + 313, + 556, + 410 + ], + "type": "inline_equation", + "content": "g(\\mathbf{z}_t;t,y)" + }, + { + "bbox": [ + 313, + 313, + 556, + 410 + ], + "type": "text", + "content": " with a strength coefficient " + }, + { + "bbox": [ + 313, + 313, + 556, + 410 + ], + "type": "inline_equation", + "content": "s" + }, + { + "bbox": [ + 313, + 313, + 556, + 410 + ], + "type": "text", + "content": ":" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 353, + 419, + 515, + 432 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 353, + 419, + 515, + 432 + ], + "spans": [ + { + "bbox": [ + 353, + 419, + 515, + 432 + ], + "type": "interline_equation", + "content": "\\hat {\\epsilon} _ {\\boldsymbol {\\theta}} (\\mathbf {z} _ {t}; t, c) = \\epsilon_ {\\boldsymbol {\\theta}} (\\mathbf {z} _ {t}; t, c) - s \\cdot g (\\mathbf {z} _ {t}; t, y),", + "image_path": "ce2862b8a3609787e3112ecf52ec0f0ff9c05cba62cdb04273338dd0c096f591.jpg" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 313, + 441, + 555, + 489 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 441, + 555, + 489 + ], + "spans": [ + { + "bbox": [ + 313, + 441, + 555, + 489 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 313, + 441, + 555, + 489 + ], + "type": "inline_equation", + "content": "g" + }, + { + "bbox": [ + 313, + 441, + 555, + 489 + ], + "type": "text", + "content": " can be defined via bounding boxes [51]. In this paper, we adopt a standard choice " + }, + { + "bbox": [ + 313, + 441, + 555, + 489 + ], + "type": "inline_equation", + "content": "\\epsilon_{\\theta}" + }, + { + "bbox": [ + 313, + 441, + 555, + 489 + ], + "type": "text", + "content": " following existing methods [25, 46], i.e., U-Net architecture [36] with multi-level self-attention and cross-attention layers [47]." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 313, + 500, + 540, + 514 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 500, + 540, + 514 + ], + "spans": [ + { + "bbox": [ + 313, + 500, + 540, + 514 + ], + "type": "text", + "content": "4. Training-Free Controllable T2I Diffusion" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 313, + 528, + 427, + 540 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 528, + 427, + 540 + ], + "spans": [ + { + "bbox": [ + 313, + 528, + 427, + 540 + ], + "type": "text", + "content": "4.1. Problem Statement" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 313, + 545, + 555, + 677 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 545, + 555, + 677 + ], + "spans": [ + { + "bbox": [ + 313, + 545, + 555, + 677 + ], + "type": "text", + "content": "Without loss of generality, we denote the well-trained 3D detector as " + }, + { + "bbox": [ + 313, + 545, + 555, + 677 + ], + "type": "inline_equation", + "content": "f_{\\Theta_d}(\\cdot)" + }, + { + "bbox": [ + 313, + 545, + 555, + 677 + ], + "type": "text", + "content": " which is obtained by training on labeled training images " + }, + { + "bbox": [ + 313, + 545, + 555, + 677 + ], + "type": "inline_equation", + "content": "\\mathcal{D} = \\{(\\mathbf{x}_i,\\mathbf{y}_i)\\}_{i = 1}^N" + }, + { + "bbox": [ + 313, + 545, + 555, + 677 + ], + "type": "text", + "content": ". The training images follow the original training distribution " + }, + { + "bbox": [ + 313, + 545, + 555, + 677 + ], + "type": "inline_equation", + "content": "P(\\mathbf{x})" + }, + { + "bbox": [ + 313, + 545, + 555, + 677 + ], + "type": "text", + "content": " (i.e., " + }, + { + "bbox": [ + 313, + 545, + 555, + 677 + ], + "type": "inline_equation", + "content": "\\mathbf{x} \\sim P(\\mathbf{x})" + }, + { + "bbox": [ + 313, + 545, + 555, + 677 + ], + "type": "text", + "content": "). Here, " + }, + { + "bbox": [ + 313, + 545, + 555, + 677 + ], + "type": "inline_equation", + "content": "\\Theta_d" + }, + { + "bbox": [ + 313, + 545, + 555, + 677 + ], + "type": "text", + "content": " represents the parameters of the detector and " + }, + { + "bbox": [ + 313, + 545, + 555, + 677 + ], + "type": "inline_equation", + "content": "N" + }, + { + "bbox": [ + 313, + 545, + 555, + 677 + ], + "type": "text", + "content": " is the total number of training images. After training, the model is applied to unlabeled test images " + }, + { + "bbox": [ + 313, + 545, + 555, + 677 + ], + "type": "inline_equation", + "content": "\\mathcal{D}_t = \\{\\mathbf{x}^t\\}" + }, + { + "bbox": [ + 313, + 545, + 555, + 677 + ], + "type": "text", + "content": ". Unfortunately, the well-trained model often encounters Out-of-distribution (OOD) test scenarios in autonomous driving (AD) due to prevalent natural corruptions, namely distribution shifts, i.e., " + }, + { + "bbox": [ + 313, + 545, + 555, + 677 + ], + "type": "inline_equation", + "content": "\\mathbf{x}^t \\sim Q(\\mathbf{x})" + }, + { + "bbox": [ + 313, + 545, + 555, + 677 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 313, + 545, + 555, + 677 + ], + "type": "inline_equation", + "content": "P(\\mathbf{x}) \\neq Q(\\mathbf{x})" + }, + { + "bbox": [ + 313, + 545, + 555, + 677 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 313, + 677, + 555, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 677, + 555, + 713 + ], + "spans": [ + { + "bbox": [ + 313, + 677, + 555, + 713 + ], + "type": "text", + "content": "As previously discussed, existing methods [22, 28] are relatively time-consuming due to their additional computation costs at test time. Hence, we investigate controllable" + } + ] + } + ], + "index": 14 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "type": "text", + "content": "27499" + } + ] + } + ], + "index": 15 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 2 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 86, + 76, + 527, + 230 + ], + "blocks": [ + { + "bbox": [ + 86, + 76, + 527, + 230 + ], + "lines": [ + { + "bbox": [ + 86, + 76, + 527, + 230 + ], + "spans": [ + { + "bbox": [ + 86, + 76, + 527, + 230 + ], + "type": "image", + "image_path": "fe48bb7292bb6f94362cb3bfbe03c238620ceaea624228e71750f9429cd5d363.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + } + ], + "index": 0 + }, + { + "type": "image", + "bbox": [ + 86, + 234, + 526, + 367 + ], + "blocks": [ + { + "bbox": [ + 86, + 234, + 526, + 367 + ], + "lines": [ + { + "bbox": [ + 86, + 234, + 526, + 367 + ], + "spans": [ + { + "bbox": [ + 86, + 234, + 526, + 367 + ], + "type": "image", + "image_path": "a2ac8b0a0517294c12ee272b6275b087c7c24d8403dc581423b81f02c7cdc9e0.jpg" + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 54, + 379, + 555, + 424 + ], + "lines": [ + { + "bbox": [ + 54, + 379, + 555, + 424 + ], + "spans": [ + { + "bbox": [ + 54, + 379, + 555, + 424 + ], + "type": "text", + "content": "Figure 3. An overview of our DriveGEN, consists of two stages: 1) The Self-Prototype Extraction stage is devised to extract accurate semantic structures of multiple objects. To capture precise locations, we achieve fine-grained self-prototypes via leveraging layouts and the peak function to re-weight object regions rather than directly using coarse self-attention features. 2) The Prototype-Guided Diffusion stage conducts semantic-aware feature alignment for semantic structure matching and shallow feature alignment for tiny object preservation." + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_caption" + } + ], + "index": 1 + }, + { + "bbox": [ + 54, + 444, + 295, + 564 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 54, + 444, + 295, + 564 + ], + "spans": [ + { + "bbox": [ + 54, + 444, + 295, + 564 + ], + "type": "text", + "content": "Text-to-Image (T2I) diffusion generation for training data augmentation, thereby improving the robustness of 3D detectors. Particularly, it is crucial to preserve object information because the annotations would be reused during detector training. However, training-based approaches [27, 59] require substantial training data to train auxiliary modules for spatial control, which is often unsatisfying in AD. Besides, training-free methods [25, 46] mainly consider the relatively coarse self-attention features, making it challenging to preserve relatively tiny objects." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 55, + 574, + 152, + 586 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 574, + 152, + 586 + ], + "spans": [ + { + "bbox": [ + 55, + 574, + 152, + 586 + ], + "type": "text", + "content": "4.2. Overall Scheme" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 55, + 593, + 295, + 676 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 593, + 295, + 676 + ], + "spans": [ + { + "bbox": [ + 55, + 593, + 295, + 676 + ], + "type": "text", + "content": "After examining the characteristics and challenges of robust 3D object detection, we introduce DriveGEN, a Training-Free Controllable T2I Diffusion Generation method devised to improve the robustness of vision-centric 3D detectors. As illustrated in Figure 3, DriveGEN consists of two stages: 1) Self-Prototype Extraction and 2) Prototype-Guided Diffusion. We briefly introduce each stage below." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 55, + 677, + 296, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 677, + 296, + 715 + ], + "spans": [ + { + "bbox": [ + 55, + 677, + 296, + 715 + ], + "type": "text", + "content": "First, we develop the self-prototype extraction stage (c.f. Section 4.3) to capture the geometric characteristics of objects accurately. Existing methods apply principal compo" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 313, + 444, + 555, + 577 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 444, + 555, + 577 + ], + "spans": [ + { + "bbox": [ + 313, + 444, + 555, + 577 + ], + "type": "text", + "content": "nent analysis (PCA) on the self-attention features of the first decoder layer of the diffusion model " + }, + { + "bbox": [ + 313, + 444, + 555, + 577 + ], + "type": "inline_equation", + "content": "\\epsilon_{\\theta}" + }, + { + "bbox": [ + 313, + 444, + 555, + 577 + ], + "type": "text", + "content": " to capture semantic structures of condition images [25, 46]. However, we empirically find that solely relying on cross-attention relevance between self-attention features and object prompts for region selection may result in object preservation errors (c.f. Figure 2). To address this, DriveGEN leverages layouts to re-weight object regions in the principal components of self-attention features via a peak function and then achieves the self-prototypes which help the diffusion model focus on the accurate object regions during generation." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 313, + 582, + 556, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 582, + 556, + 715 + ], + "spans": [ + { + "bbox": [ + 313, + 582, + 556, + 715 + ], + "type": "text", + "content": "Second, we introduce the prototype-guided diffusion stage (c.f. Section 4.4) to preserve all objects during the generation process. On the one hand, we conduct semantic-aware feature alignment between self-attention features and self-prototypes to capture accurate object semantic structures. On the other hand, DriveGEN performs shallow feature alignment to further preserve object information by constraining the noisy latent at the shallow level, compensating for the gaps in semantic-aware feature alignment regarding tiny object preservation. The Pseudo-code of DriveGEN is summarized in Algorithm 1." + } + ] + } + ], + "index": 8 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "type": "text", + "content": "27500" + } + ] + } + ], + "index": 9 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 3 + }, + { + "para_blocks": [ + { + "bbox": [ + 55, + 72, + 198, + 85 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 72, + 198, + 85 + ], + "spans": [ + { + "bbox": [ + 55, + 72, + 198, + 85 + ], + "type": "text", + "content": "4.3. Self-Prototype Extraction" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 55, + 89, + 295, + 137 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 89, + 295, + 137 + ], + "spans": [ + { + "bbox": [ + 55, + 89, + 295, + 137 + ], + "type": "text", + "content": "To accurately capture the geometric characteristics of multiple objects, we propose a self-prototype extraction strategy that consists of two components: 1) Image structure extraction and 2) Object region re-weighting." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 55, + 137, + 295, + 330 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 137, + 295, + 330 + ], + "spans": [ + { + "bbox": [ + 55, + 137, + 295, + 330 + ], + "type": "text", + "content": "Image Structure Extraction. To achieve zero-shot spatial control, capturing the semantic structures of input images is essential. Prior works [25, 46] demonstrate that self-attention features (e.g., keys) of the T2I diffusion model " + }, + { + "bbox": [ + 55, + 137, + 295, + 330 + ], + "type": "inline_equation", + "content": "\\epsilon_{\\theta}" + }, + { + "bbox": [ + 55, + 137, + 295, + 330 + ], + "type": "text", + "content": " are semantic-aware [35]. Thus, given the input image " + }, + { + "bbox": [ + 55, + 137, + 295, + 330 + ], + "type": "inline_equation", + "content": "\\mathbf{x}_i" + }, + { + "bbox": [ + 55, + 137, + 295, + 330 + ], + "type": "text", + "content": ", we first generate the text prompt " + }, + { + "bbox": [ + 55, + 137, + 295, + 330 + ], + "type": "inline_equation", + "content": "\\mathbf{c}_i" + }, + { + "bbox": [ + 55, + 137, + 295, + 330 + ], + "type": "text", + "content": " via a simple template 'A photo of " + }, + { + "bbox": [ + 55, + 137, + 295, + 330 + ], + "type": "inline_equation", + "content": "\\{\\text{object}\\} ,\\ldots," + }, + { + "bbox": [ + 55, + 137, + 295, + 330 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 55, + 137, + 295, + 330 + ], + "type": "inline_equation", + "content": "\\{\\text{object}\\}'" + }, + { + "bbox": [ + 55, + 137, + 295, + 330 + ], + "type": "text", + "content": ". Then, we leverage DDIM Inversion [39] to achieve a series of time-dependent diffusion features " + }, + { + "bbox": [ + 55, + 137, + 295, + 330 + ], + "type": "inline_equation", + "content": "\\{\\mathbf{z}_t^I\\in \\mathbb{R}^{N_c\\times H\\times W}\\}_{t = 1}^{N_t}" + }, + { + "bbox": [ + 55, + 137, + 295, + 330 + ], + "type": "text", + "content": " where " + }, + { + "bbox": [ + 55, + 137, + 295, + 330 + ], + "type": "inline_equation", + "content": "N_{t}" + }, + { + "bbox": [ + 55, + 137, + 295, + 330 + ], + "type": "text", + "content": " denotes the number of DDIM inversion and sampling steps, " + }, + { + "bbox": [ + 55, + 137, + 295, + 330 + ], + "type": "inline_equation", + "content": "N_{c}" + }, + { + "bbox": [ + 55, + 137, + 295, + 330 + ], + "type": "text", + "content": " represents the number of feature channels, and " + }, + { + "bbox": [ + 55, + 137, + 295, + 330 + ], + "type": "inline_equation", + "content": "H,W" + }, + { + "bbox": [ + 55, + 137, + 295, + 330 + ], + "type": "text", + "content": " indicate the height and width, respectively. During sampling, we apply PCA on self-attention features to extract principal components " + }, + { + "bbox": [ + 55, + 137, + 295, + 330 + ], + "type": "inline_equation", + "content": "\\{\\mathbf{P}_t\\}_{t = 1}^{N_p}" + }, + { + "bbox": [ + 55, + 137, + 295, + 330 + ], + "type": "text", + "content": ", each " + }, + { + "bbox": [ + 55, + 137, + 295, + 330 + ], + "type": "inline_equation", + "content": "\\mathbf{P}_t\\in \\mathbb{R}^{N_b\\times H\\times W}" + }, + { + "bbox": [ + 55, + 137, + 295, + 330 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 55, + 137, + 295, + 330 + ], + "type": "inline_equation", + "content": "\\mathbf{P}_t = \\{\\mathbf{p}_t^1,\\dots,\\mathbf{p}_t^{N_b}\\}" + }, + { + "bbox": [ + 55, + 137, + 295, + 330 + ], + "type": "text", + "content": ". Here " + }, + { + "bbox": [ + 55, + 137, + 295, + 330 + ], + "type": "inline_equation", + "content": "N_{b}" + }, + { + "bbox": [ + 55, + 137, + 295, + 330 + ], + "type": "text", + "content": " denotes the number of principal components and we store " + }, + { + "bbox": [ + 55, + 137, + 295, + 330 + ], + "type": "inline_equation", + "content": "\\mathbf{P}_t" + }, + { + "bbox": [ + 55, + 137, + 295, + 330 + ], + "type": "text", + "content": " of the initial " + }, + { + "bbox": [ + 55, + 137, + 295, + 330 + ], + "type": "inline_equation", + "content": "N_{p}" + }, + { + "bbox": [ + 55, + 137, + 295, + 330 + ], + "type": "text", + "content": " steps." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 55, + 330, + 295, + 474 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 330, + 295, + 474 + ], + "spans": [ + { + "bbox": [ + 55, + 330, + 295, + 474 + ], + "type": "text", + "content": "Object Region Re-weighting. Existing methods [9, 25] select foreground regions by leveraging cross-attention maps between specific words (i.e., object names) and image features to derive a concept region mask " + }, + { + "bbox": [ + 55, + 330, + 295, + 474 + ], + "type": "inline_equation", + "content": "\\mathbf{M}" + }, + { + "bbox": [ + 55, + 330, + 295, + 474 + ], + "type": "text", + "content": ". However, relying solely on such attention maps may result in omissions and misalignment of object regions, as illustrated in Figure 2. To address this, we introduce a peak function " + }, + { + "bbox": [ + 55, + 330, + 295, + 474 + ], + "type": "inline_equation", + "content": "m(\\cdot)" + }, + { + "bbox": [ + 55, + 330, + 295, + 474 + ], + "type": "text", + "content": " to reweight " + }, + { + "bbox": [ + 55, + 330, + 295, + 474 + ], + "type": "inline_equation", + "content": "\\mathbf{M}" + }, + { + "bbox": [ + 55, + 330, + 295, + 474 + ], + "type": "text", + "content": " for each object region based on layouts " + }, + { + "bbox": [ + 55, + 330, + 295, + 474 + ], + "type": "inline_equation", + "content": "\\mathbf{L}" + }, + { + "bbox": [ + 55, + 330, + 295, + 474 + ], + "type": "text", + "content": " after the coordinate transformation through downsampling. To be specific, given a point " + }, + { + "bbox": [ + 55, + 330, + 295, + 474 + ], + "type": "inline_equation", + "content": "(p,q)" + }, + { + "bbox": [ + 55, + 330, + 295, + 474 + ], + "type": "text", + "content": " from " + }, + { + "bbox": [ + 55, + 330, + 295, + 474 + ], + "type": "inline_equation", + "content": "\\mathbf{L}_i" + }, + { + "bbox": [ + 55, + 330, + 295, + 474 + ], + "type": "text", + "content": " of " + }, + { + "bbox": [ + 55, + 330, + 295, + 474 + ], + "type": "inline_equation", + "content": "\\mathbf{x}_i" + }, + { + "bbox": [ + 55, + 330, + 295, + 474 + ], + "type": "text", + "content": ", the corresponding object center is defined as " + }, + { + "bbox": [ + 55, + 330, + 295, + 474 + ], + "type": "inline_equation", + "content": "(\\hat{p},\\hat{q})" + }, + { + "bbox": [ + 55, + 330, + 295, + 474 + ], + "type": "text", + "content": " and then we can calculate " + }, + { + "bbox": [ + 55, + 330, + 295, + 474 + ], + "type": "inline_equation", + "content": "m(p,q)" + }, + { + "bbox": [ + 55, + 330, + 295, + 474 + ], + "type": "text", + "content": " by:" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 89, + 479, + 295, + 508 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 89, + 479, + 295, + 508 + ], + "spans": [ + { + "bbox": [ + 89, + 479, + 295, + 508 + ], + "type": "interline_equation", + "content": "m (p, q) = \\exp \\left(- \\frac {(p - \\hat {p}) ^ {2} + (q - \\hat {q}) ^ {2}}{2 \\sigma^ {2}}\\right) \\tag {1}", + "image_path": "e8abe5e5b38c2712d58ec8eae8639bf654d7c6d8a0b98ff0c23edc0210ebd8a0.jpg" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 55, + 513, + 295, + 537 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 513, + 295, + 537 + ], + "spans": [ + { + "bbox": [ + 55, + 513, + 295, + 537 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 55, + 513, + 295, + 537 + ], + "type": "inline_equation", + "content": "\\sigma \\in (0,1)" + }, + { + "bbox": [ + 55, + 513, + 295, + 537 + ], + "type": "text", + "content": " controls the range of the influence area. Then, we update the concept token mask " + }, + { + "bbox": [ + 55, + 513, + 295, + 537 + ], + "type": "inline_equation", + "content": "\\mathbf{M}" + }, + { + "bbox": [ + 55, + 513, + 295, + 537 + ], + "type": "text", + "content": " by:" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 91, + 544, + 295, + 559 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 91, + 544, + 295, + 559 + ], + "spans": [ + { + "bbox": [ + 91, + 544, + 295, + 559 + ], + "type": "interline_equation", + "content": "\\tilde {\\mathbf {M}} (p, q) = \\mathbf {M} (p, q) + m (p, q) \\cdot \\mathbf {L} _ {i} (p, q). \\tag {2}", + "image_path": "bf96b284c7d43f7971732b08da4044dbcccaa372331d887751acf9e606056621.jpg" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 55, + 564, + 295, + 628 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 564, + 295, + 628 + ], + "spans": [ + { + "bbox": [ + 55, + 564, + 295, + 628 + ], + "type": "text", + "content": "The object regions in " + }, + { + "bbox": [ + 55, + 564, + 295, + 628 + ], + "type": "inline_equation", + "content": "\\mathbf{L}_i" + }, + { + "bbox": [ + 55, + 564, + 295, + 628 + ], + "type": "text", + "content": " are marked as 1 while all other regions as 0. Eventually, we obtain the self-prototypes by " + }, + { + "bbox": [ + 55, + 564, + 295, + 628 + ], + "type": "inline_equation", + "content": "\\{\\hat{\\mathbf{P}}_t\\}_{t=1}^{N_p} = \\{\\mathbf{P}_t\\}_{t=1}^{N_p}\\big|_{\\hat{\\mathbf{M}}}" + }, + { + "bbox": [ + 55, + 564, + 295, + 628 + ], + "type": "text", + "content": " which guide the diffusion model to localize precise locations of objects and combine " + }, + { + "bbox": [ + 55, + 564, + 295, + 628 + ], + "type": "inline_equation", + "content": "\\hat{\\mathbf{M}}" + }, + { + "bbox": [ + 55, + 564, + 295, + 628 + ], + "type": "text", + "content": " for accurate object region selection." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 55, + 635, + 209, + 647 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 635, + 209, + 647 + ], + "spans": [ + { + "bbox": [ + 55, + 635, + 209, + 647 + ], + "type": "text", + "content": "4.4. Prototype-Guided Diffusion" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 55, + 652, + 295, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 652, + 295, + 714 + ], + "spans": [ + { + "bbox": [ + 55, + 652, + 295, + 714 + ], + "type": "text", + "content": "To preserve all objects with precise 3D geometry during diffusion, we leverage the self-prototypes " + }, + { + "bbox": [ + 55, + 652, + 295, + 714 + ], + "type": "inline_equation", + "content": "\\{\\hat{\\mathbf{P}}_t\\}_{t=1}^{N_p}" + }, + { + "bbox": [ + 55, + 652, + 295, + 714 + ], + "type": "text", + "content": " and time-dependent diffusion features " + }, + { + "bbox": [ + 55, + 652, + 295, + 714 + ], + "type": "inline_equation", + "content": "\\{\\mathbf{z}_t^I\\}_{t=1}^{N_t}" + }, + { + "bbox": [ + 55, + 652, + 295, + 714 + ], + "type": "text", + "content": " for feature alignment, including: 1) semantic-aware feature alignment and 2) shallow feature alignment." + } + ] + } + ], + "index": 9 + }, + { + "type": "code", + "bbox": [ + 316, + 87, + 553, + 280 + ], + "blocks": [ + { + "bbox": [ + 316, + 72, + 531, + 84 + ], + "lines": [ + { + "bbox": [ + 316, + 72, + 531, + 84 + ], + "spans": [ + { + "bbox": [ + 316, + 72, + 531, + 84 + ], + "type": "text", + "content": "Algorithm 1 The pipeline of the proposed DriveGEN" + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "code_caption" + }, + { + "bbox": [ + 316, + 87, + 553, + 280 + ], + "lines": [ + { + "bbox": [ + 316, + 87, + 553, + 280 + ], + "spans": [ + { + "bbox": [ + 316, + 87, + 553, + 280 + ], + "type": "text", + "content": "Require: Training data " + }, + { + "bbox": [ + 316, + 87, + 553, + 280 + ], + "type": "inline_equation", + "content": "\\{(\\mathbf{x}_i^s,\\mathbf{y}_i^s)\\}_{i = 1}^N" + }, + { + "bbox": [ + 316, + 87, + 553, + 280 + ], + "type": "text", + "content": " Pretrained T2I diffusion model " + }, + { + "bbox": [ + 316, + 87, + 553, + 280 + ], + "type": "inline_equation", + "content": "\\epsilon_{\\theta}" + }, + { + "bbox": [ + 316, + 87, + 553, + 280 + ], + "type": "text", + "content": " Hyper-parameters " + }, + { + "bbox": [ + 316, + 87, + 553, + 280 + ], + "type": "inline_equation", + "content": "s,\\sigma ,\\tau" + }, + { + "bbox": [ + 316, + 87, + 553, + 280 + ], + "type": "text", + "content": " The OOD scenario. \n#Stage 1: Self-Prototype Extraction \n1: for each training image " + }, + { + "bbox": [ + 316, + 87, + 553, + 280 + ], + "type": "inline_equation", + "content": "x_{i}" + }, + { + "bbox": [ + 316, + 87, + 553, + 280 + ], + "type": "text", + "content": " do \n2: Get the text prompt " + }, + { + "bbox": [ + 316, + 87, + 553, + 280 + ], + "type": "inline_equation", + "content": "c_{i}" + }, + { + "bbox": [ + 316, + 87, + 553, + 280 + ], + "type": "text", + "content": " and layout " + }, + { + "bbox": [ + 316, + 87, + 553, + 280 + ], + "type": "inline_equation", + "content": "\\mathbf{L}_i" + }, + { + "bbox": [ + 316, + 87, + 553, + 280 + ], + "type": "text", + "content": " based on " + }, + { + "bbox": [ + 316, + 87, + 553, + 280 + ], + "type": "inline_equation", + "content": "\\mathbf{y}_i" + }, + { + "bbox": [ + 316, + 87, + 553, + 280 + ], + "type": "text", + "content": " \n3: Undergo DDIM Inversion to get " + }, + { + "bbox": [ + 316, + 87, + 553, + 280 + ], + "type": "inline_equation", + "content": "\\mathbf{P}_t" + }, + { + "bbox": [ + 316, + 87, + 553, + 280 + ], + "type": "text", + "content": " \n4: Get updated M via Eqn. (1), Eqn. (2) for self-prototypes " + }, + { + "bbox": [ + 316, + 87, + 553, + 280 + ], + "type": "inline_equation", + "content": "\\hat{\\mathbf{P}}_t" + }, + { + "bbox": [ + 316, + 87, + 553, + 280 + ], + "type": "text", + "content": " \n5: end for \n#Stage 2: Prototype-Guided Diffusion \n6: for each training image " + }, + { + "bbox": [ + 316, + 87, + 553, + 280 + ], + "type": "inline_equation", + "content": "\\mathbf{x}_i" + }, + { + "bbox": [ + 316, + 87, + 553, + 280 + ], + "type": "text", + "content": " do \n7: Update " + }, + { + "bbox": [ + 316, + 87, + 553, + 280 + ], + "type": "inline_equation", + "content": "\\mathbf{c}_i" + }, + { + "bbox": [ + 316, + 87, + 553, + 280 + ], + "type": "text", + "content": " with specific OOD type; \n8: for diffusion step " + }, + { + "bbox": [ + 316, + 87, + 553, + 280 + ], + "type": "inline_equation", + "content": "t = 1\\rightarrow N_t" + }, + { + "bbox": [ + 316, + 87, + 553, + 280 + ], + "type": "text", + "content": " do \n9: if " + }, + { + "bbox": [ + 316, + 87, + 553, + 280 + ], + "type": "inline_equation", + "content": "t\\leq N_p" + }, + { + "bbox": [ + 316, + 87, + 553, + 280 + ], + "type": "text", + "content": " then \n10: Calculate " + }, + { + "bbox": [ + 316, + 87, + 553, + 280 + ], + "type": "inline_equation", + "content": "g_{sa},g_{sl}" + }, + { + "bbox": [ + 316, + 87, + 553, + 280 + ], + "type": "text", + "content": " based on Eqn. (3), Eqn. (4); \n11: Get the gradient based on Eqn.(5); \n12: end if \n13: Update the noisy latent " + }, + { + "bbox": [ + 316, + 87, + 553, + 280 + ], + "type": "inline_equation", + "content": "\\mathbf{z}_t\\to \\mathbf{z}_{t - 1}" + }, + { + "bbox": [ + 316, + 87, + 553, + 280 + ], + "type": "text", + "content": " \n14: end for \n15: end for \n16: return Output images for all " + }, + { + "bbox": [ + 316, + 87, + 553, + 280 + ], + "type": "inline_equation", + "content": "\\mathbf{x}_i" + }, + { + "bbox": [ + 316, + 87, + 553, + 280 + ], + "type": "text", + "content": " in the OOD scenario." + } + ] + } + ], + "index": 11, + "angle": 0, + "type": "code_body" + } + ], + "index": 11, + "sub_type": "algorithm" + }, + { + "bbox": [ + 313, + 300, + 555, + 409 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 300, + 555, + 409 + ], + "spans": [ + { + "bbox": [ + 313, + 300, + 555, + 409 + ], + "type": "text", + "content": "Semantic-aware Feature Alignment. As mentioned before, the self-attention features of the first decoder layer can represent the semantic structures of input images. Thus, we first initialize the noisy latent " + }, + { + "bbox": [ + 313, + 300, + 555, + 409 + ], + "type": "inline_equation", + "content": "\\mathbf{z}_{N_t}" + }, + { + "bbox": [ + 313, + 300, + 555, + 409 + ], + "type": "text", + "content": " via the " + }, + { + "bbox": [ + 313, + 300, + 555, + 409 + ], + "type": "inline_equation", + "content": "\\mathbf{z}_{N_t}^I" + }, + { + "bbox": [ + 313, + 300, + 555, + 409 + ], + "type": "text", + "content": ". For the step " + }, + { + "bbox": [ + 313, + 300, + 555, + 409 + ], + "type": "inline_equation", + "content": "t \\leq N_p" + }, + { + "bbox": [ + 313, + 300, + 555, + 409 + ], + "type": "text", + "content": ", we perform PCA to obtain the principal components " + }, + { + "bbox": [ + 313, + 300, + 555, + 409 + ], + "type": "inline_equation", + "content": "\\mathbf{P}_t" + }, + { + "bbox": [ + 313, + 300, + 555, + 409 + ], + "type": "text", + "content": " of the self-attention features. Then, we adopt a pre-defined threshold " + }, + { + "bbox": [ + 313, + 300, + 555, + 409 + ], + "type": "inline_equation", + "content": "\\tau" + }, + { + "bbox": [ + 313, + 300, + 555, + 409 + ], + "type": "text", + "content": " following [9] to select informative regions by " + }, + { + "bbox": [ + 313, + 300, + 555, + 409 + ], + "type": "inline_equation", + "content": "\\hat{\\mathbf{M}} = \\mathbb{I}(\\hat{\\mathbf{M}} > \\tau)" + }, + { + "bbox": [ + 313, + 300, + 555, + 409 + ], + "type": "text", + "content": ". Specifically, we calculate the semantic-aware feature alignment loss with " + }, + { + "bbox": [ + 313, + 300, + 555, + 409 + ], + "type": "inline_equation", + "content": "\\mathbf{P}_t" + }, + { + "bbox": [ + 313, + 300, + 555, + 409 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 313, + 300, + 555, + 409 + ], + "type": "inline_equation", + "content": "\\hat{\\mathbf{P}}_t" + }, + { + "bbox": [ + 313, + 300, + 555, + 409 + ], + "type": "text", + "content": ":" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 377, + 414, + 553, + 430 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 377, + 414, + 553, + 430 + ], + "spans": [ + { + "bbox": [ + 377, + 414, + 553, + 430 + ], + "type": "interline_equation", + "content": "g _ {s a} = \\| \\hat {\\mathbf {M}} \\odot (\\mathbf {P} _ {t} - \\hat {\\mathbf {P}} _ {t}) \\| _ {2} ^ {2}, \\tag {3}", + "image_path": "4f44ee7d0c4d161e19b1700ef49b0f68f5bc40ce9de9f19ed5afe4ff14e2a699.jpg" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 313, + 434, + 553, + 458 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 434, + 553, + 458 + ], + "spans": [ + { + "bbox": [ + 313, + 434, + 553, + 458 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 313, + 434, + 553, + 458 + ], + "type": "inline_equation", + "content": "\\odot" + }, + { + "bbox": [ + 313, + 434, + 553, + 458 + ], + "type": "text", + "content": " represents element-wise multiplication, applying the mask " + }, + { + "bbox": [ + 313, + 434, + 553, + 458 + ], + "type": "inline_equation", + "content": "\\hat{\\mathbf{M}}" + }, + { + "bbox": [ + 313, + 434, + 553, + 458 + ], + "type": "text", + "content": " to the difference between " + }, + { + "bbox": [ + 313, + 434, + 553, + 458 + ], + "type": "inline_equation", + "content": "\\mathbf{P}_t" + }, + { + "bbox": [ + 313, + 434, + 553, + 458 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 313, + 434, + 553, + 458 + ], + "type": "inline_equation", + "content": "\\hat{\\mathbf{P}}_t" + }, + { + "bbox": [ + 313, + 434, + 553, + 458 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 313, + 458, + 554, + 542 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 458, + 554, + 542 + ], + "spans": [ + { + "bbox": [ + 313, + 458, + 554, + 542 + ], + "type": "text", + "content": "Shallow Feature Alignment. As shown in Figure 2, we empirically find that principal components lack sufficient details to represent objects, especially tiny objects. For instance, given a bounding box of a cyclist with a height of 20 pixels and a width of 5 pixels, this object is unable to occupy even one element within principal components after multiple rounds of downsampling (e.g., " + }, + { + "bbox": [ + 313, + 458, + 554, + 542 + ], + "type": "inline_equation", + "content": "32\\mathrm{x}" + }, + { + "bbox": [ + 313, + 458, + 554, + 542 + ], + "type": "text", + "content": ")." + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 313, + 542, + 554, + 602 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 542, + 554, + 602 + ], + "spans": [ + { + "bbox": [ + 313, + 542, + 554, + 602 + ], + "type": "text", + "content": "Therefore, DriveGEN introduces shallow feature alignment before the forward process of diffusion, which aims to preserve fine-grained object details by constraining the noisy latent. Specifically, based on the layout " + }, + { + "bbox": [ + 313, + 542, + 554, + 602 + ], + "type": "inline_equation", + "content": "\\mathbf{L}_i" + }, + { + "bbox": [ + 313, + 542, + 554, + 602 + ], + "type": "text", + "content": " with " + }, + { + "bbox": [ + 313, + 542, + 554, + 602 + ], + "type": "inline_equation", + "content": "N_{i}" + }, + { + "bbox": [ + 313, + 542, + 554, + 602 + ], + "type": "text", + "content": " objects, we calculate the shallow feature alignment loss by:" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 330, + 617, + 553, + 646 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 330, + 617, + 553, + 646 + ], + "spans": [ + { + "bbox": [ + 330, + 617, + 553, + 646 + ], + "type": "interline_equation", + "content": "g _ {s l} = \\frac {1}{N _ {i}} \\sum_ {p, q} \\| \\mathbf {L} _ {i} (p, q) \\odot \\left(\\mathbf {z} _ {t} ^ {I} (p, q) - \\mathbf {z} _ {t} (p, q)\\right) \\| _ {2} ^ {2}, \\tag {4}", + "image_path": "510ef379b489638714bc64a7f7678d45f139b2d9d04a80072d39d5a8ff43da27.jpg" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 325, + 651, + 539, + 663 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 325, + 651, + 539, + 663 + ], + "spans": [ + { + "bbox": [ + 325, + 651, + 539, + 663 + ], + "type": "text", + "content": "Overall, the total scheme of DriveGEN is as follows:" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 324, + 670, + 553, + 685 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 324, + 670, + 553, + 685 + ], + "spans": [ + { + "bbox": [ + 324, + 670, + 553, + 685 + ], + "type": "interline_equation", + "content": "\\hat {\\epsilon} _ {t} = (1 + s) \\epsilon_ {\\theta} (\\mathbf {z} _ {t}; t, c) - s \\epsilon_ {\\theta} (\\mathbf {z} _ {t}; t, \\emptyset) + g _ {s a} + g _ {s l}. \\tag {5}", + "image_path": "f42250441b8dad9dee76eae2c241ef5912076f2fbb66ee79d6749efb0e37d31d.jpg" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 313, + 689, + 553, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 689, + 553, + 714 + ], + "spans": [ + { + "bbox": [ + 313, + 689, + 553, + 714 + ], + "type": "text", + "content": "Note that " + }, + { + "bbox": [ + 313, + 689, + 553, + 714 + ], + "type": "inline_equation", + "content": "\\varnothing" + }, + { + "bbox": [ + 313, + 689, + 553, + 714 + ], + "type": "text", + "content": " denotes null and DriveGEN is a classifier-free guidance [12] process with hyper-parameters " + }, + { + "bbox": [ + 313, + 689, + 553, + 714 + ], + "type": "inline_equation", + "content": "s" + }, + { + "bbox": [ + 313, + 689, + 553, + 714 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 313, + 689, + 553, + 714 + ], + "type": "inline_equation", + "content": "\\sigma" + }, + { + "bbox": [ + 313, + 689, + 553, + 714 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 313, + 689, + 553, + 714 + ], + "type": "inline_equation", + "content": "\\tau" + }, + { + "bbox": [ + 313, + 689, + 553, + 714 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 20 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "text", + "content": "27501" + } + ] + } + ], + "index": 21 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 4 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 82, + 91, + 528, + 316 + ], + "blocks": [ + { + "bbox": [ + 64, + 70, + 545, + 83 + ], + "lines": [ + { + "bbox": [ + 64, + 70, + 545, + 83 + ], + "spans": [ + { + "bbox": [ + 64, + 70, + 545, + 83 + ], + "type": "text", + "content": "Table 1. Comparison on the KITTI-C dataset, severity level 1 regarding Mean " + }, + { + "bbox": [ + 64, + 70, + 545, + 83 + ], + "type": "inline_equation", + "content": "AP_{3D|R_{40}}" + }, + { + "bbox": [ + 64, + 70, + 545, + 83 + ], + "type": "text", + "content": ". The bold number indicates the best result." + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 82, + 91, + 528, + 316 + ], + "lines": [ + { + "bbox": [ + 82, + 91, + 528, + 316 + ], + "spans": [ + { + "bbox": [ + 82, + 91, + 528, + 316 + ], + "type": "table", + "html": "
Car, IoU @ 0.7, 0.5, 0.5
MethodTraining-free diffusionNoiseBlurWeatherDigitalAvg.
Gauss.ShotImpul.Defoc.GlassMotionSnowFrostFogBrit.Contr.PixelSat.
Monoflex [63]13.0620.9114.0920.1728.5930.3433.6430.3119.5845.2220.0129.0738.8526.45
• Color Jitter (Traditional aug.)9.5515.8111.9022.6725.3830.1234.0830.0019.2942.1019.9317.1736.4824.19
• Brightness (Traditional aug.)11.4419.4212.7311.1818.9522.0326.6421.7013.0439.6113.0821.1129.7320.05
• ControlNet (Only Snow aug.)X0.321.181.284.6511.3217.0422.8419.729.5334.798.731.2516.8811.50
• ControlNet (3 scenarios aug.)X0.701.140.290.350.370.834.073.051.209.430.940.483.202.00
• ControlNet (6 scenarios aug.)X0.000.000.000.000.001.940.000.000.000.000.000.000.000.15
• Freecontrol (Only Snow aug.)✓20.3928.0522.2712.5221.6721.0927.9717.9110.0735.999.9028.5635.2722.44
• Freecontrol (3 scenarios aug.)✓15.4321.2514.0415.3319.2216.5322.0117.4115.1824.6616.6625.3931.4419.58
• Freecontrol (6 scenarios aug.)✓12.8518.7116.3112.0616.3112.8617.6416.1814.2324.5215.6422.5624.2817.24
• DriveGEN (Only Snow aug.)✓16.4826.7224.9831.1735.5538.1341.3938.6127.8549.7629.2838.8844.1234.07
• DriveGEN (3 scenarios aug.)✓25.6337.0429.1334.1339.1536.8138.5837.9833.9345.3934.6639.3643.8136.58
• DriveGEN (6 scenarios aug.)✓24.7733.7928.2736.9240.3340.4540.6040.5638.1044.8339.2841.8144.0537.98
MonoGround [31]13.0521.7718.8720.7930.7432.0234.4327.0214.1546.2114.6333.4135.6026.36
• Color Jitter (Traditional aug.)12.8824.3118.9523.0730.4431.4235.9430.4319.8944.6620.6129.7536.6526.36
• Brightness (Traditional aug.)14.0223.5220.1423.9531.7828.7935.0831.8718.8742.9417.7525.5537.1827.03
• ControlNet (Only Snow aug.)X1.763.234.635.2012.9514.1117.7011.583.0435.212.987.2913.9810.28
• ControlNet (3 scenarios aug.)X0.000.000.241.421.681.614.905.400.5717.901.127.796.823.80
• ControlNet (6 scenarios aug.)X0.000.000.001.681.260.351.130.520.444.080.382.221.771.06
• Freecontrol (Only Snow aug.)✓11.7521.8915.7617.7021.4521.6932.0820.6013.5736.0514.0326.7538.3522.43
• Freecontrol (3 scenarios aug.)✓16.3120.7517.6113.1016.8414.8217.7215.8911.8725.2013.5522.5724.4917.75
• Freecontrol (6 scenarios aug.)✓15.2022.5915.3522.0021.1818.9517.6914.8514.8224.0216.9722.9926.1219.44
• DriveGEN (Only Snow aug.)✓17.0726.7823.7832.8937.5239.0640.6134.9125.2946.2127.1238.2544.4533.38
• DriveGEN (3 scenarios aug.)✓19.7931.4427.6336.8440.1039.3539.1736.4228.7945.4229.2942.6044.9935.53
• DriveGEN (6 scenarios aug.)✓23.8432.5930.3438.5741.2040.1938.1638.4032.5343.9534.8044.1045.1337.21
", + "image_path": "1482d99c0dbd94b4c53cc3b2d346753d6c4dbfb68004c74c71a327b13c275189.jpg" + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "table_body" + } + ], + "index": 1 + }, + { + "type": "table", + "bbox": [ + 82, + 317, + 528, + 521 + ], + "blocks": [ + { + "bbox": [ + 82, + 317, + 528, + 521 + ], + "lines": [ + { + "bbox": [ + 82, + 317, + 528, + 521 + ], + "spans": [ + { + "bbox": [ + 82, + 317, + 528, + 521 + ], + "type": "table", + "html": "
Pedestrian, IoU @ 0.5, 0.25, 0.25
Monoflex [63]1.163.540.788.0617.7015.2612.729.255.6119.875.351.498.658.42
• Color Jitter (Traditional aug.)0.993.531.4110.0514.6312.0014.7312.137.7219.029.231.1111.299.07
• Brightness (Traditional aug.)0.631.851.205.0113.7212.157.935.841.8716.892.300.483.775.67
• ControlNet (Only Snow aug.)X0.000.000.001.788.325.844.033.751.3111.001.230.000.932.94
• ControlNet (3 scenarios aug.)X0.000.000.001.153.231.141.641.350.834.062.010.000.001.19
• ControlNet (6 scenarios aug.)X0.000.000.000.000.360.000.000.000.002.500.000.000.000.22
• Freecontrol (Only Snow aug.)✓3.625.173.213.235.165.944.683.524.287.704.228.517.705.15
• Freecontrol (3 scenarios aug.)✓2.263.543.043.274.974.257.424.417.2710.916.186.837.435.52
• Freecontrol (6 scenarios aug.)✓5.999.109.084.807.365.847.227.587.1010.819.409.9010.248.03
• DriveGEN (Only Snow aug.)✓1.253.224.5616.2119.8820.6119.8014.468.8024.629.119.7718.0413.10
• DriveGEN (3 scenarios aug.)✓6.099.859.6014.9618.5916.4715.4516.3915.8821.9816.8514.5518.1314.98
• DriveGEN (6 scenarios aug.)✓6.349.227.7117.6019.2120.2016.8816.1915.7223.5716.7514.3717.3115.47
MonoGround [31]2.673.255.7617.5718.9117.7112.969.354.3724.155.893.277.1610.23
• Color Jitter (Traditional aug.)2.443.244.1115.3718.4616.5015.4512.389.1424.719.822.077.8110.89
• Brightness (Traditional aug.)2.794.107.6114.5114.1314.5212.1212.665.5520.805.362.3610.319.75
• ControlNet (Only Snow aug.)X1.851.030.817.189.978.321.322.921.3712.991.400.280.903.87
• ControlNet (3 scenarios aug.)X0.000.000.003.603.891.961.141.031.504.211.250.260.771.51
• ControlNet (6 scenarios aug.)X0.000.000.000.000.001.670.000.000.000.000.000.000.000.13
• Freecontrol (Only Snow aug.)✓10.0413.0911.7014.9112.6113.0315.1412.358.1315.6912.3416.3214.5713.07
• Freecontrol (3 scenarios aug.)✓0.771.811.982.193.083.963.290.522.553.763.104.105.372.81
• Freecontrol (6 scenarios aug.)✓6.817.547.675.545.455.102.191.313.638.595.358.866.265.72
• DriveGEN (Only Snow aug.)✓6.226.899.2715.0317.2418.8116.4013.389.8123.2811.5012.0413.9213.37
• DriveGEN (3 scenarios aug.)✓6.539.2611.1716.1519.6021.0218.6716.1713.7224.8014.5219.9819.3616.23
• DriveGEN (6 scenarios aug.)✓9.7013.6813.5217.0017.7420.3917.2318.7915.5023.0915.8020.9918.3417.06
", + "image_path": "61aec76d56408943344bd2f8f355ae78621d9479e165cc697d3f4c1b1437051b.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "table_body" + } + ], + "index": 2 + }, + { + "bbox": [ + 55, + 551, + 137, + 565 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 551, + 137, + 565 + ], + "spans": [ + { + "bbox": [ + 55, + 551, + 137, + 565 + ], + "type": "text", + "content": "5. Experiments" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 55, + 576, + 295, + 672 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 576, + 295, + 672 + ], + "spans": [ + { + "bbox": [ + 55, + 576, + 295, + 672 + ], + "type": "text", + "content": "We conduct experiments to validate DriveGEN on monocular and multi-view 3D object detection, including four different training settings with various augmented OOD scenarios: 1) Augmented with traditional techniques (Traditional aug.); 2) Only augmented with Snow (Only Snow aug.); 3) Augmented with Snow, Rain and Fog (3 scenarios aug.); 4) Augmented with Snow, Rain, Fog, Night, Defocus and Sandstorm (6 scenarios aug.)." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 55, + 677, + 296, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 677, + 296, + 715 + ], + "spans": [ + { + "bbox": [ + 55, + 677, + 296, + 715 + ], + "type": "text", + "content": "Datasets. For monocular 3d object detection, we split the images of KITTI [10] into a training set (3712 images) and a validation set (3769 images) following the protocol from" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 313, + 552, + 555, + 613 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 552, + 555, + 613 + ], + "spans": [ + { + "bbox": [ + 313, + 552, + 555, + 613 + ], + "type": "text", + "content": "Monoflex [63], encompassing three classes: Car, Pedestrian, and Cyclist. We adopt the KITTI-C dataset [22] for validation, which includes 13 types of data corruptions on the validation set across four categories: Noise, Blur, Weather, and Digital [11]." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 313, + 617, + 556, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 617, + 556, + 715 + ], + "spans": [ + { + "bbox": [ + 313, + 617, + 556, + 715 + ], + "type": "text", + "content": "As for multi-view 3D object detection, we first split real-world Night and Rainy validation scenarios based on descriptions of nuScenes [1] following [24]. Next, we augment 500 daytime training scenes under the Snow condition. The augmented detectors are then evaluated on real-world Night and Rainy scenarios, as well as on the widely used Robo3D benchmark [52]. More details of datasets are provided in Appendix B." + } + ] + } + ], + "index": 7 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "type": "text", + "content": "27502" + } + ] + } + ], + "index": 8 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 5 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 56, + 70, + 555, + 178 + ], + "blocks": [ + { + "bbox": [ + 56, + 70, + 555, + 178 + ], + "lines": [ + { + "bbox": [ + 56, + 70, + 555, + 178 + ], + "spans": [ + { + "bbox": [ + 56, + 70, + 555, + 178 + ], + "type": "image", + "image_path": "409d28a1f3ac3a5e981ced131940fcf6503911c35a3773f84cb90ffafc2b14da.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 83, + 184, + 523, + 198 + ], + "lines": [ + { + "bbox": [ + 83, + 184, + 523, + 198 + ], + "spans": [ + { + "bbox": [ + 83, + 184, + 523, + 198 + ], + "type": "text", + "content": "Figure 4. Based on MonoCD [54], we provide more comparisons with KITTI-C, regarding Mean " + }, + { + "bbox": [ + 83, + 184, + 523, + 198 + ], + "type": "inline_equation", + "content": "AP_{3D|R_{40}}" + }, + { + "bbox": [ + 83, + 184, + 523, + 198 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "type": "table", + "bbox": [ + 66, + 227, + 544, + 283 + ], + "blocks": [ + { + "bbox": [ + 80, + 207, + 528, + 219 + ], + "lines": [ + { + "bbox": [ + 80, + 207, + 528, + 219 + ], + "spans": [ + { + "bbox": [ + 80, + 207, + 528, + 219 + ], + "type": "text", + "content": "Table 2. Detection results on nuScenes-C and real-world Night and Rainy scenarios of nuScenes, regarding mAP and NDS." + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 66, + 227, + 544, + 283 + ], + "lines": [ + { + "bbox": [ + 66, + 227, + 544, + 283 + ], + "spans": [ + { + "bbox": [ + 66, + 227, + 544, + 283 + ], + "type": "table", + "html": "
MetricMethodnuScenes-CReal-world Scenarios
BrightnessCameraCrashColorQuantFogFrameLostLowLightMotionBlurSnowAvg.nuScenes-NightnuScenes-RainyAvg.
mAPBEVFormer-tiny24.2622.0123.9121.9820.4316.1120.7810.8320.049.3126.7618.04
• DriveGEN (3k Snow)26.0424.1325.7824.1321.1317.2022.3911.7221.57(↑1.53)12.1929.1120.65 (↑2.61)
NDSBEVFormer-tiny34.9333.0634.5633.0431.4228.4231.8123.2731.3117.9237.3527.64
• DriveGEN (3k Snow)37.1435.3436.9335.3533.3230.4234.2324.0933.35(↑2.04)19.3740.9530.16 (↑2.52)
", + "image_path": "78457a8cb79eb82a9e327e5765cc0b6bf4c0af6fd5c938196aa65b617ed52f03.jpg" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "table_body" + } + ], + "index": 3 + }, + { + "bbox": [ + 54, + 312, + 294, + 455 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 54, + 312, + 294, + 455 + ], + "spans": [ + { + "bbox": [ + 54, + 312, + 294, + 455 + ], + "type": "text", + "content": "Implementation Details. We implement our method and other baselines in PyTorch [29]. Following Freecontrol [25], we adopt keys from the self-attention of the first U-Net decoder layer as the features. In the self-prototype extraction stage, we run DDIM inversion with 200 steps. Subsequently, we also run 200 steps of DDIM sampling in the prototype-guided diffusion stage. We set " + }, + { + "bbox": [ + 54, + 312, + 294, + 455 + ], + "type": "inline_equation", + "content": "s" + }, + { + "bbox": [ + 54, + 312, + 294, + 455 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 54, + 312, + 294, + 455 + ], + "type": "inline_equation", + "content": "\\tau" + }, + { + "bbox": [ + 54, + 312, + 294, + 455 + ], + "type": "text", + "content": " to 7.5 and 0.3 following FreeControl [25], while " + }, + { + "bbox": [ + 54, + 312, + 294, + 455 + ], + "type": "inline_equation", + "content": "\\sigma" + }, + { + "bbox": [ + 54, + 312, + 294, + 455 + ], + "type": "text", + "content": " is set to 0.1. All 3D detectors are trained on a mix of original and augmented data. All results in the manuscript are based on Stable Diffusion (SD) 1.5 [35]. More details and the results of SD 2.1 and XL 1.0 are put in Appendix C." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 54, + 457, + 295, + 565 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 54, + 457, + 295, + 565 + ], + "spans": [ + { + "bbox": [ + 54, + 457, + 295, + 565 + ], + "type": "text", + "content": "Compared Methods. Based on typical or state-of-the-art (SOTA) 3D detectors [20, 31, 54, 63], we fully compare DriveGEN with following methods: 1) source-only, i.e., directly apply the well-trained model to corrupted test data; 2) Traditional data augmentation, i.e., Color Jitter and Brightness; 3) training-based controllable T2I diffusion: ControlNet [59] trains auxiliary modules for spatial control; 4) training-free controllable T2I diffusion: Freecontrol [25] enables zero-shot control of pretrained diffusion models." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 55, + 567, + 295, + 640 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 567, + 295, + 640 + ], + "spans": [ + { + "bbox": [ + 55, + 567, + 295, + 640 + ], + "type": "text", + "content": "Evaluation Protocols. We report the experimental results in the Average Precision (AP) for 3D bounding boxes, denoted as " + }, + { + "bbox": [ + 55, + 567, + 295, + 640 + ], + "type": "inline_equation", + "content": "AP_{3D|R_{40}}" + }, + { + "bbox": [ + 55, + 567, + 295, + 640 + ], + "type": "text", + "content": ". On the KITTI-C dataset, the results present the mean values of three difficulty levels and Intersection over Union (IoU) thresholds are set to 0.7, 0.5, 0.5 for Cars and 0.5, 0.25, 0.25 for Pedestrians and Cyclists." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 55, + 642, + 296, + 712 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 642, + 296, + 712 + ], + "spans": [ + { + "bbox": [ + 55, + 642, + 296, + 712 + ], + "type": "text", + "content": "As for nuScenes, the mean average precision (mAP) and nuScenes detection score (NDS) are calculated following BEVFormer [20]. The metrics also include five true positive metrics, including ATE, ASE, AOE, AVE, and AAE for measuring errors in translation, scale, orientation, velocity, and attributes." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 313, + 311, + 509, + 323 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 311, + 509, + 323 + ], + "spans": [ + { + "bbox": [ + 313, + 311, + 509, + 323 + ], + "type": "text", + "content": "5.1. Comparisons with Previous Methods" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 312, + 328, + 555, + 519 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 312, + 328, + 555, + 519 + ], + "spans": [ + { + "bbox": [ + 312, + 328, + 555, + 519 + ], + "type": "text", + "content": "For monocular 3d object detection, Table 1 gives the following observations: 1) Applying the well-trained detector to corrupted scenarios suffers significant performance degradation due to the data distribution shifts and traditional data augmentation offers limited gains. 2) ControlNet (pretrained on ade20k [70]) fails to enhance 3d detectors with limited training data even if it is provided with accurate segmentation masks [34] and rich prompts [6] for fine-tuning. 3) Freecontrol achieves minor gains in a few cases but generally degrades, indicating that training-free T2I diffusion is a feasible solution but requires careful object preservation. 4) DriveGEN consistently outperforms all compared methods over all categories within various base models for both known and unseen scenarios, like improving MonoFlex by an average of " + }, + { + "bbox": [ + 312, + 328, + 555, + 519 + ], + "type": "inline_equation", + "content": "7.6\\mathrm{mAP}" + }, + { + "bbox": [ + 312, + 328, + 555, + 519 + ], + "type": "text", + "content": " across 13 various OOD scenarios with the single Snow scenario augmentation." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 313, + 520, + 556, + 653 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 520, + 556, + 653 + ], + "spans": [ + { + "bbox": [ + 313, + 520, + 556, + 653 + ], + "type": "text", + "content": "As for multi-view 3D object detection, we randomly select 3k daytime training images and apply Snow augmentation (3k Snow) to enhance the detector [20]. Table 2 shows DriveGEN consistently enhances BEVFormer-tiny across 8 OOD scenarios with an average of " + }, + { + "bbox": [ + 313, + 520, + 556, + 653 + ], + "type": "inline_equation", + "content": "1.53\\mathrm{mAP}" + }, + { + "bbox": [ + 313, + 520, + 556, + 653 + ], + "type": "text", + "content": " and 2.04 NDS, further confirming our effectiveness. Moreover, we validate the detector on real Night and Rainy scenarios of nuScenes. With the help of DriveGEN, it also achieves notable gains in real-world tasks with an average improvement of " + }, + { + "bbox": [ + 313, + 520, + 556, + 653 + ], + "type": "inline_equation", + "content": "2.61\\mathrm{mAP}" + }, + { + "bbox": [ + 313, + 520, + 556, + 653 + ], + "type": "text", + "content": " and 2.52 NDS, further demonstrating our superiority. More detailed results are put in the Appendix D." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 313, + 659, + 527, + 673 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 659, + 527, + 673 + ], + "spans": [ + { + "bbox": [ + 313, + 659, + 527, + 673 + ], + "type": "text", + "content": "5.2. More Augmented Scenarios for Training" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 313, + 677, + 555, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 677, + 555, + 715 + ], + "spans": [ + { + "bbox": [ + 313, + 677, + 555, + 715 + ], + "type": "text", + "content": "An intuitive concern is whether the robustness of detectors progressively improves with the increasing of additional augmented scenarios (i.e., Only Snow aug. " + }, + { + "bbox": [ + 313, + 677, + 555, + 715 + ], + "type": "inline_equation", + "content": "\\rightarrow 3" + }, + { + "bbox": [ + 313, + 677, + 555, + 715 + ], + "type": "text", + "content": " scenarios" + } + ] + } + ], + "index": 12 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "text", + "content": "27503" + } + ] + } + ], + "index": 13 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 6 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 66, + 70, + 545, + 178 + ], + "blocks": [ + { + "bbox": [ + 66, + 70, + 545, + 178 + ], + "lines": [ + { + "bbox": [ + 66, + 70, + 545, + 178 + ], + "spans": [ + { + "bbox": [ + 66, + 70, + 545, + 178 + ], + "type": "image", + "image_path": "e9f4557a8a6b85aaaccd5c3dc8551acd280a52881c407ab09a85a1631be846da.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 78, + 189, + 530, + 201 + ], + "lines": [ + { + "bbox": [ + 78, + 189, + 530, + 201 + ], + "spans": [ + { + "bbox": [ + 78, + 189, + 530, + 201 + ], + "type": "text", + "content": "Figure 5. Ablation studies on semantic feature alignment loss " + }, + { + "bbox": [ + 78, + 189, + 530, + 201 + ], + "type": "inline_equation", + "content": "g_{sa}" + }, + { + "bbox": [ + 78, + 189, + 530, + 201 + ], + "type": "text", + "content": ", self-prototypes " + }, + { + "bbox": [ + 78, + 189, + 530, + 201 + ], + "type": "inline_equation", + "content": "\\hat{\\mathbf{P}}_t" + }, + { + "bbox": [ + 78, + 189, + 530, + 201 + ], + "type": "text", + "content": " and shallow feature alignment loss " + }, + { + "bbox": [ + 78, + 189, + 530, + 201 + ], + "type": "inline_equation", + "content": "g_{sl}" + }, + { + "bbox": [ + 78, + 189, + 530, + 201 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "type": "image", + "bbox": [ + 67, + 213, + 545, + 350 + ], + "blocks": [ + { + "bbox": [ + 67, + 213, + 545, + 350 + ], + "lines": [ + { + "bbox": [ + 67, + 213, + 545, + 350 + ], + "spans": [ + { + "bbox": [ + 67, + 213, + 545, + 350 + ], + "type": "image", + "image_path": "e9e976d8ab81f03eca7bcbdf2bdbca6ee1396e6a104efa635b40e427d48b212c.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 55, + 360, + 555, + 384 + ], + "lines": [ + { + "bbox": [ + 55, + 360, + 555, + 384 + ], + "spans": [ + { + "bbox": [ + 55, + 360, + 555, + 384 + ], + "type": "text", + "content": "Figure 6. Qualitative results of DriveGEN for multi-view 3D object detection on nuScenes. DriveGEN supports vision-centric 3D detection tasks since our method only requires input images and corresponding object annotations without any additional diffusion model training." + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_caption" + } + ], + "index": 2 + }, + { + "bbox": [ + 55, + 404, + 296, + 536 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 404, + 296, + 536 + ], + "spans": [ + { + "bbox": [ + 55, + 404, + 296, + 536 + ], + "type": "text", + "content": "aug. " + }, + { + "bbox": [ + 55, + 404, + 296, + 536 + ], + "type": "inline_equation", + "content": "\\rightarrow 6" + }, + { + "bbox": [ + 55, + 404, + 296, + 536 + ], + "type": "text", + "content": " scenarios aug.). Based on Table 1 and Figure 4, we are given additional observations: 1) For ControlNet, the model performance drop for OOD scenarios grows as the number of augmentation scenarios increases, demonstrating that severe object omission leads to noisy training data. 2) For Freecontrol, the model performance across various OOD scenarios remains stable or tends to decline, indicating its object preservation may be unstable. 3) Owing to the precise object preservation, DriveGEN obtains continuous overall performance (i.e., Avg.) improvement across all categories with the incorporation of OOD scenarios." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 55, + 559, + 156, + 571 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 559, + 156, + 571 + ], + "spans": [ + { + "bbox": [ + 55, + 559, + 156, + 571 + ], + "type": "text", + "content": "5.3. Ablation Studies" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 55, + 582, + 296, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 582, + 296, + 715 + ], + "spans": [ + { + "bbox": [ + 55, + 582, + 296, + 715 + ], + "type": "text", + "content": "To examine DriveGEN, we present qualitative results guided by different loss settings. As shown in Figure 5, compared with only generated via text prompts, introducing semantic-aware feature alignment " + }, + { + "bbox": [ + 55, + 582, + 296, + 715 + ], + "type": "inline_equation", + "content": "g_{sa}" + }, + { + "bbox": [ + 55, + 582, + 296, + 715 + ], + "type": "text", + "content": " helps preserve more image structures. Next, by guiding the diffusion process with our self-prototypes " + }, + { + "bbox": [ + 55, + 582, + 296, + 715 + ], + "type": "inline_equation", + "content": "\\hat{\\mathbf{P}}_t" + }, + { + "bbox": [ + 55, + 582, + 296, + 715 + ], + "type": "text", + "content": ", we achieve more accurate object geometry, alleviating potential object geometry loss. Eventually, introducing our shallow feature alignment " + }, + { + "bbox": [ + 55, + 582, + 296, + 715 + ], + "type": "inline_equation", + "content": "g_{sl}" + }, + { + "bbox": [ + 55, + 582, + 296, + 715 + ], + "type": "text", + "content": " further improves object preservation with precise 3D geometry, especially for small objects (highlighted by red circles). More ablation studies are provided in Appendix D." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 314, + 403, + 432, + 416 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 314, + 403, + 432, + 416 + ], + "spans": [ + { + "bbox": [ + 314, + 403, + 432, + 416 + ], + "type": "text", + "content": "5.4. Quantitative Results" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 313, + 421, + 555, + 517 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 421, + 555, + 517 + ], + "spans": [ + { + "bbox": [ + 313, + 421, + 555, + 517 + ], + "type": "text", + "content": "We provide visualizations for multi-view 3D object detection on nuScenes in Snow scenarios as shown in Figure 6. It is evident that DriveGEN supports existing vision-centric 3D detection tasks since it requires only input images and corresponding object annotations without any additional diffusion model training, demonstrating that DriveGEN achieves superior results even in challenging multiview tasks. More visualizations are put in Appendix E." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 314, + 526, + 388, + 538 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 314, + 526, + 388, + 538 + ], + "spans": [ + { + "bbox": [ + 314, + 526, + 388, + 538 + ], + "type": "text", + "content": "6. Conclusion" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 313, + 546, + 555, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 546, + 555, + 713 + ], + "spans": [ + { + "bbox": [ + 313, + 546, + 555, + 713 + ], + "type": "text", + "content": "In this paper, we propose a method for robust 3D detection in driving via training-free controllable text-to-image diffusion generation, namely DriveGEN. Specifically, DriveGEN consists of: 1) Self-Prototype Extraction: To improve the guidance of self-attention features, we extract the self-prototypes by layouts to capture accurate object geometry, leveraging more precise features to guide diffusion. 2) Prototype-Guided Diffusion: To further preserve 3D object geometry, we conduct semantic-aware feature and shallow feature alignment during diffusion, alleviating the object misalignment and omission issues compared with previous methods. Experiments on KITTI-C and nuScenes demonstrate the effectiveness of DriveGEN in improving model robustness for vision-centric 3D detection." + } + ] + } + ], + "index": 10 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "type": "text", + "content": "27504" + } + ] + } + ], + "index": 11 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 7 + }, + { + "para_blocks": [ + { + "bbox": [ + 56, + 72, + 159, + 85 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 72, + 159, + 85 + ], + "spans": [ + { + "bbox": [ + 56, + 72, + 159, + 85 + ], + "type": "text", + "content": "Acknowledgements" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 55, + 91, + 297, + 332 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 91, + 297, + 332 + ], + "spans": [ + { + "bbox": [ + 55, + 91, + 297, + 332 + ], + "type": "text", + "content": "This work was supported by NSFC with Grant No. 62293482, by the Basic Research Project No. HZQB-KCZYZ-2021067 of Hetao Shenzhen HK S&T Cooperation Zone, by Shenzhen General Program No. JCYJ20220530143600001, by Shenzhen-Hong Kong Joint Funding No. SGDX20211123112401002, by the Shenzhen Outstanding Talents Training Fund 202002, by Guangdong Research Project No. 2017ZT07X152 and No. 2019CX01X104, by the Guangdong Provincial Key Laboratory of Future Networks of Intelligence (Grant No. 2022B1212010001), by the Guangdong Provincial Key Laboratory of Big Data Computing, CHUK-Shenzhen, by the NSFC 61931024&12326610, by the Key Area R&D Program of Guangdong Province with grant No. 2018B030338001, by the Shenzhen Key Laboratory of Big Data and Artificial Intelligence (Grant No. ZDSYS201707251409055), by Shaanxi Mathematical Basic Science Research Project(No.23JSY047), and by Tencent & Huawei Open Fund, by China Association for Science and Technology Youth Care Program." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 56, + 342, + 115, + 355 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 342, + 115, + 355 + ], + "spans": [ + { + "bbox": [ + 56, + 342, + 115, + 355 + ], + "type": "text", + "content": "References" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 61, + 362, + 295, + 713 + ], + "type": "list", + "angle": 0, + "index": 11, + "blocks": [ + { + "bbox": [ + 61, + 362, + 295, + 407 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 61, + 362, + 295, + 407 + ], + "spans": [ + { + "bbox": [ + 61, + 362, + 295, + 407 + ], + "type": "text", + "content": "[1] Holger Caesar, Varun Bankiti, Alex H Lang, and et.al. nuscene: A multimodal dataset for autonomous driving. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 11621-11631, 2020. 6" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 61, + 408, + 295, + 452 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 61, + 408, + 295, + 452 + ], + "spans": [ + { + "bbox": [ + 61, + 408, + 295, + 452 + ], + "type": "text", + "content": "[2] Changjian Chen, Fei Lv, Yalong Guan, Pengcheng Wang, Shengjie Yu, Yifan Zhang, and Zhuo Tang. Human-guided image generation for expanding small-scale training image datasets. arXiv preprint arXiv:2412.16839, 2024. 3" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 62, + 454, + 295, + 487 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 454, + 295, + 487 + ], + "spans": [ + { + "bbox": [ + 62, + 454, + 295, + 487 + ], + "type": "text", + "content": "[3] Xiaozhi Chen, Kaustav Kundu, Yukun Zhu, and et. al. 3d object proposals for accurate object class detection. Advances in neural information processing systems, 28, 2015. 2" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 62, + 488, + 295, + 532 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 488, + 295, + 532 + ], + "spans": [ + { + "bbox": [ + 62, + 488, + 295, + 532 + ], + "type": "text", + "content": "[4] Xiaozhi Chen, Huimin Ma, Ji Wan, Bo Li, and Tian Xia. Multi-view 3d object detection network for autonomous driving. In Proceedings of the IEEE conference on Computer Vision and Pattern Recognition, pages 1907-1915, 2017." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 62, + 533, + 295, + 586 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 533, + 295, + 586 + ], + "spans": [ + { + "bbox": [ + 62, + 533, + 295, + 586 + ], + "type": "text", + "content": "[5] Yukang Chen, Jianhui Liu, Xiangyu Zhang, Xiaojuan Qi, and Jiaya Jia. Voxelnext: Fully sparse voxelnet for 3d object detection and tracking. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 21674-21683, 2023. 2" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 62, + 590, + 295, + 655 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 590, + 295, + 655 + ], + "spans": [ + { + "bbox": [ + 62, + 590, + 295, + 655 + ], + "type": "text", + "content": "[6] Zhe Chen, Jiannan Wu, Wenhai Wang, Weijie Su, Guo Chen, Sen Xing, Muyan Zhong, Qinglong Zhang, Xizhou Zhu, Lewei Lu, et al. Internvl: Scaling up vision foundation models and aligning for generic visual-linguistic tasks. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 24185-24198, 2024. 3, 7" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 62, + 657, + 295, + 689 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 657, + 295, + 689 + ], + "spans": [ + { + "bbox": [ + 62, + 657, + 295, + 689 + ], + "type": "text", + "content": "[7] Prafulla Dhariwal and Alexander Nichol. Diffusion models beat gans on image synthesis. Advances in neural information processing systems, 34:8780-8794, 2021. 3, 1" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 62, + 692, + 295, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 692, + 295, + 713 + ], + "spans": [ + { + "bbox": [ + 62, + 692, + 295, + 713 + ], + "type": "text", + "content": "[8] Dave Epstein, Allan Jabri, Ben Poole, Alexei Efros, and Aleksander Holynski. Diffusion self-guidance for control" + } + ] + } + ], + "index": 10 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 316, + 73, + 555, + 713 + ], + "type": "list", + "angle": 0, + "index": 26, + "blocks": [ + { + "bbox": [ + 333, + 73, + 553, + 95 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 333, + 73, + 553, + 95 + ], + "spans": [ + { + "bbox": [ + 333, + 73, + 553, + 95 + ], + "type": "text", + "content": "able image generation. Advances in Neural Information Processing Systems, 36:16222-16239, 2023. 3" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 320, + 98, + 553, + 140 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 320, + 98, + 553, + 140 + ], + "spans": [ + { + "bbox": [ + 320, + 98, + 553, + 140 + ], + "type": "text", + "content": "[9] Songwei Ge, Taesung Park, Jun-Yan Zhu, and Jia-Bin Huang. Expressive text-to-image generation with rich text. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 7545-7556, 2023. 5" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 316, + 143, + 553, + 196 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 143, + 553, + 196 + ], + "spans": [ + { + "bbox": [ + 316, + 143, + 553, + 196 + ], + "type": "text", + "content": "[10] Andreas Geiger, Philip Lenz, and Raquel Urtasun. Are we ready for autonomous driving? the kitti vision benchmark suite. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 3354-3361. IEEE, 2012. 1, 6" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 316, + 200, + 553, + 243 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 200, + 553, + 243 + ], + "spans": [ + { + "bbox": [ + 316, + 200, + 553, + 243 + ], + "type": "text", + "content": "[11] Dan Hendrycks and Thomas Dietterich. Benchmarking neural network robustness to common corruptions and perturbations. In International Conference on Learning Representations, 2018. 6" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 316, + 246, + 553, + 267 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 246, + 553, + 267 + ], + "spans": [ + { + "bbox": [ + 316, + 246, + 553, + 267 + ], + "type": "text", + "content": "[12] Jonathan Ho and Tim Salimans. Classifier-free diffusion guidance. arXiv preprint arXiv:2207.12598, 2022. 5" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 316, + 270, + 553, + 302 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 270, + 553, + 302 + ], + "spans": [ + { + "bbox": [ + 316, + 270, + 553, + 302 + ], + "type": "text", + "content": "[13] Jonathan Ho, Ajay Jain, and Pieter Abbeel. Denoising diffusion probabilistic models. Advances in neural information processing systems, 33:6840-6851, 2020. 3, 1" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 316, + 304, + 555, + 358 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 304, + 555, + 358 + ], + "spans": [ + { + "bbox": [ + 316, + 304, + 555, + 358 + ], + "type": "text", + "content": "[14] Yuming Hong, Weiyao Xu, Yi Yang, Hao Zhou, and Xiaoxue Wang. Magicdrive: 3d controllable driving video synthesis using layout and 3d representations. Proceedings of the IEEE/CVF International Conference on Computer Vision (ICCV), 2021. 3" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 316, + 361, + 553, + 405 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 361, + 553, + 405 + ], + "spans": [ + { + "bbox": [ + 316, + 361, + 553, + 405 + ], + "type": "text", + "content": "[15] Xinyu Huang, Mingyu Chen, Yunqi Li, Hao Wang, and Yuxing Zhao. Diffumask: Synthetic image and semantic mask annotation using cross-attention in diffusion models. arXiv preprint arXiv:2211.09123, 2022. 3" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 317, + 407, + 555, + 461 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 407, + 555, + 461 + ], + "spans": [ + { + "bbox": [ + 317, + 407, + 555, + 461 + ], + "type": "text", + "content": "[16] Junsoo Kim, Yeongjae Ye, Dongheon Cho, Jaechan Seo, and Jiseob Jeong. Drivingdiffusion: Realistic multi-view video generation with diffusion models. Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), 2021. 3" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 317, + 464, + 553, + 507 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 464, + 553, + 507 + ], + "spans": [ + { + "bbox": [ + 317, + 464, + 553, + 507 + ], + "type": "text", + "content": "[17] Feng Li, Yuan Zhang, Xinyang Wang, and Meng Liu. Vpd: Vision-based perception diffusion models for human pose estimation and dense visual annotations. arXiv preprint arXiv:2109.05270, 2021. 3" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 317, + 510, + 553, + 542 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 510, + 553, + 542 + ], + "spans": [ + { + "bbox": [ + 317, + 510, + 553, + 542 + ], + "type": "text", + "content": "[18] Hongyu Li, Pengfei Zhao, Wei Jiang, and Zihan Chen. Sensor fusion with diffusion models for enhanced 3d object detection. arXiv preprint arXiv:2205.02574, 2022. 1" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 317, + 544, + 553, + 588 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 544, + 553, + 588 + ], + "spans": [ + { + "bbox": [ + 317, + 544, + 553, + 588 + ], + "type": "text", + "content": "[19] Peiliang Li, Xiaozhi Chen, and Shaojie Shen. Stereo r-cnn based 3d object detection for autonomous driving. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 7644-7652, 2019. 2" + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 317, + 590, + 553, + 655 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 590, + 553, + 655 + ], + "spans": [ + { + "bbox": [ + 317, + 590, + 553, + 655 + ], + "type": "text", + "content": "[20] Zhiqi Li, Wenhai Wang, Hongyang Li, Enze Xie, Chonghao Sima, Tong Lu, Yu Qiao, and Jifeng Dai. Bevformer: Learning bird's-eye-view representation from multi-camera images via spatiotemporal transformers. In European conference on computer vision, pages 1-18. Springer, 2022. 7, 2, 4" + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 316, + 658, + 553, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 658, + 553, + 713 + ], + "spans": [ + { + "bbox": [ + 316, + 658, + 553, + 713 + ], + "type": "text", + "content": "[21] Hongbin Lin, Yifan Zhang, Zhen Qiu, Shuaicheng Niu, Chuang Gan, Yanxia Liu, and Mingkui Tan. Prototype-guided continual adaptation for class-incremental unsupervised domain adaptation. In European Conference on Computer Vision, pages 351-368. Springer, 2022. 2" + } + ] + } + ], + "index": 25 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "text", + "content": "27505" + } + ] + } + ], + "index": 27 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 8 + }, + { + "para_blocks": [ + { + "bbox": [ + 56, + 72, + 295, + 713 + ], + "type": "list", + "angle": 0, + "index": 13, + "blocks": [ + { + "bbox": [ + 56, + 72, + 294, + 117 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 72, + 294, + 117 + ], + "spans": [ + { + "bbox": [ + 56, + 72, + 294, + 117 + ], + "type": "text", + "content": "[22] Hongbin Lin, Yifan Zhang, Shuaicheng Niu, Shuguang Cui, and Zhen Li. Monotta: Fully test-time adaptation for monocular 3d object detection. In European Conference on Computer Vision, pages 96-114. Springer, 2025. 2, 3, 6, 1, 4" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 56, + 118, + 295, + 162 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 118, + 295, + 162 + ], + "spans": [ + { + "bbox": [ + 56, + 118, + 295, + 162 + ], + "type": "text", + "content": "[23] Xiaoyang Liu, Yikang Wang, Xinyu Zhang, Wei Ma, Chao Wang, and Yun Chen. Multi-modal fusion for 3d object detection in autonomous driving. arXiv preprint arXiv:2204.01234, 2022. 1" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 56, + 163, + 295, + 218 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 163, + 295, + 218 + ], + "spans": [ + { + "bbox": [ + 56, + 163, + 295, + 218 + ], + "type": "text", + "content": "[24] Zhijian Liu, Haotian Tang, Alexander Amini, Xinyu Yang, Huizi Mao, Daniela L Rus, and Song Han. Bevfusion: Multi-task multi-sensor fusion with unified bird's-eye view representation. In 2023 IEEE international conference on robotics and automation (ICRA), pages 2774-2781. IEEE, 2023. 6, 1" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 56, + 220, + 295, + 285 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 220, + 295, + 285 + ], + "spans": [ + { + "bbox": [ + 56, + 220, + 295, + 285 + ], + "type": "text", + "content": "[25] Sicheng Mo, Fangzhou Mu, Kuan Heng Lin, Yanli Liu, Bochen Guan, Yin Li, and Bolei Zhou. Freecontrol: Training-free spatial control of any text-to-image diffusion model with any condition. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 7465-7475, 2024. 2, 3, 4, 5, 7" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 56, + 286, + 295, + 330 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 286, + 295, + 330 + ], + "spans": [ + { + "bbox": [ + 56, + 286, + 295, + 330 + ], + "type": "text", + "content": "[26] Seokhwa Moon, Sanghyuk Hwang, Hyunsu Lee, and Minje Kim. Temporal fusion with attention mechanisms for driving scenarios. Proceedings of the IEEE/CVF International Conference on Computer Vision (ICCV), 2022. 1" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 56, + 331, + 295, + 386 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 331, + 295, + 386 + ], + "spans": [ + { + "bbox": [ + 56, + 331, + 295, + 386 + ], + "type": "text", + "content": "[27] Chong Mou, Xintao Wang, Liangbin Xie, Yanze Wu, Jian Zhang, Zhongang Qi, and Ying Shan. T2i-adapter: Learning adapters to dig out more controllable ability for text-to-image diffusion models. In Proceedings of the AAAI Conference on Artificial Intelligence, pages 4296–4304, 2024. 2, 4" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 56, + 388, + 295, + 431 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 388, + 295, + 431 + ], + "spans": [ + { + "bbox": [ + 56, + 388, + 295, + 431 + ], + "type": "text", + "content": "[28] Youngmin Oh, Hyung-II Kim, Seong Tae Kim, and Jung Uk Kim. Monowad: Weather-adaptive diffusion model for robust monocular 3d object detection. In European Conference on Computer Vision, 2025. 2, 3" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 56, + 433, + 295, + 498 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 433, + 295, + 498 + ], + "spans": [ + { + "bbox": [ + 56, + 433, + 295, + 498 + ], + "type": "text", + "content": "[29] Adam Paszke, Sam Gross, Francisco Massa, Adam Lerer, James Bradbury, Gregory Chanan, Trevor Killeen, Zeming Lin, Natalia Gimelshein, Luca Antiga, et al. Pytorch: An imperative style, high-performance deep learning library. In Advances in neural information processing systems, 2019. 7, 2" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 56, + 500, + 295, + 553 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 500, + 295, + 553 + ], + "spans": [ + { + "bbox": [ + 56, + 500, + 295, + 553 + ], + "type": "text", + "content": "[30] Yuheng Qin, Zhaoyang Zhang, Han Zhang, Dong Li, Baining Chen, Lu Zhang, Jiuming Gu, and Dongdong Zhang. Uncontrol: A unified diffusion model for controllable visual generation in the wild. arXiv preprint arXiv:2305.11147, 2023. 3" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 56, + 556, + 295, + 600 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 556, + 295, + 600 + ], + "spans": [ + { + "bbox": [ + 56, + 556, + 295, + 600 + ], + "type": "text", + "content": "[31] Zequn Qin and Xi Li. Monoground: Detecting monocular 3d objects from the ground. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 3793-3802, 2022. 6, 7, 4" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 56, + 601, + 295, + 644 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 601, + 295, + 644 + ], + "spans": [ + { + "bbox": [ + 56, + 601, + 295, + 644 + ], + "type": "text", + "content": "[32] Zhen Qiu, Yifan Zhang, Hongbin Lin, Shuaicheng Niu, Yanxia Liu, Qing Du, and Mingkui Tan. Source-free domain adaptation via avatar prototype generation and adaptation. arXiv preprint arXiv:2106.15326, 2021. 2" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 56, + 647, + 295, + 689 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 647, + 295, + 689 + ], + "spans": [ + { + "bbox": [ + 56, + 647, + 295, + 689 + ], + "type": "text", + "content": "[33] Aditya Ramesh, Prafulla Dhariwal, Alex Nichol, Casey Chu, and Mark Chen. Hierarchical text-conditional image generation with clip latents. arXiv preprint arXiv:2204.06125, 2022. 3" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 56, + 691, + 295, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 691, + 295, + 713 + ], + "spans": [ + { + "bbox": [ + 56, + 691, + 295, + 713 + ], + "type": "text", + "content": "[34] Nikhila Ravi, Valentin Gabeur, Yuan-Ting Hu, Ronghang Hu, Chaitanya Ryali, Tengyu Ma, Haitham Khedr, Roman" + } + ] + } + ], + "index": 12 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 316, + 73, + 555, + 713 + ], + "type": "list", + "angle": 0, + "index": 27, + "blocks": [ + { + "bbox": [ + 333, + 73, + 553, + 106 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 333, + 73, + 553, + 106 + ], + "spans": [ + { + "bbox": [ + 333, + 73, + 553, + 106 + ], + "type": "text", + "content": "Rädle, Chloe Rolland, Laura Gustafson, et al. Sam 2: Segment anything in images and videos. arXiv preprint arXiv:2408.00714, 2024. 3, 7" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 316, + 108, + 555, + 162 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 108, + 555, + 162 + ], + "spans": [ + { + "bbox": [ + 316, + 108, + 555, + 162 + ], + "type": "text", + "content": "[35] Robin Rombach, Andreas Blattmann, Dominik Lorenz, Patrick Esser, and Björn Ommer. High-resolution image synthesis with latent diffusion models. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 10684-10695, 2022. 3, 5, 7, 1" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 316, + 163, + 555, + 229 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 163, + 555, + 229 + ], + "spans": [ + { + "bbox": [ + 316, + 163, + 555, + 229 + ], + "type": "text", + "content": "[36] Olaf Ronneberger, Philipp Fischer, and Thomas Brox. U-net: Convolutional networks for biomedical image segmentation. In Medical image computing and computer-assisted intervention-MICCAI 2015: 18th international conference, Munich, Germany, October 5-9, 2015, proceedings, part III 18, pages 234-241. Springer, 2015. 3, 2" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 316, + 231, + 554, + 285 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 231, + 554, + 285 + ], + "spans": [ + { + "bbox": [ + 316, + 231, + 554, + 285 + ], + "type": "text", + "content": "[37] Chitwan Sahara, William Chan, Saurabh Saxena, Lala Li, Jay Whang, Emily Denton, Seyed Kamyar Seyed Ghasemipour, Burcu Karagol Ayan, Tim Salimans, Jonathan Ho, et al. Imagen: Photorealistic text-to-image diffusion models. arXiv preprint arXiv:2205.11487, 2022. 3" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 316, + 286, + 554, + 341 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 286, + 554, + 341 + ], + "spans": [ + { + "bbox": [ + 316, + 286, + 554, + 341 + ], + "type": "text", + "content": "[38] Jascha Sohl-Dickstein, Eric A Weiss, Niru Maheswaranathan, and Surya Ganguli. Deep unsupervised learning using nonequilibrium thermodynamics. Proceedings of the 32nd International Conference on Machine Learning, 2015. 1" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 316, + 342, + 554, + 374 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 342, + 554, + 374 + ], + "spans": [ + { + "bbox": [ + 316, + 342, + 554, + 374 + ], + "type": "text", + "content": "[39] Jiaming Song, Chenlin Meng, and Stefano Ermon. Denoising diffusion implicit models. arXiv preprint arXiv:2010.02502, 2020. 5" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 316, + 376, + 554, + 421 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 376, + 554, + 421 + ], + "spans": [ + { + "bbox": [ + 316, + 376, + 554, + 421 + ], + "type": "text", + "content": "[40] Yang Song, Jascha Sohl-Dickstein, Diederik P Kingma, Abhishek Kumar, Stefano Ermon, and Ben Poole. Score-based generative modeling through stochastic differential equations. arXiv preprint arXiv:2011.13456, 2021. 1" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 316, + 422, + 554, + 475 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 422, + 554, + 475 + ], + "spans": [ + { + "bbox": [ + 316, + 422, + 554, + 475 + ], + "type": "text", + "content": "[41] Yang Song, Jascha Sohl-Dickstein, Diederik P Kingma, Abhishek Kumar, Stefano Ermon, and Ben Poole. Score-based generative modeling through stochastic differential equations. In International Conference on Learning Representations, 2021. 3" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 316, + 478, + 554, + 533 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 478, + 554, + 533 + ], + "spans": [ + { + "bbox": [ + 316, + 478, + 554, + 533 + ], + "type": "text", + "content": "[42] Ziying Song, Guoxing Zhang, Lin Liu, Lei Yang, Shaoqing Xu, Caiyan Jia, Feiyang Jia, and Li Wang. Robofusion: towards robust multi-modal 3d object detection via sam. In Proceedings of the Thirty-Third International Joint Conference on Artificial Intelligence, pages 1272–1280, 2024. 3" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 316, + 534, + 554, + 578 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 534, + 554, + 578 + ], + "spans": [ + { + "bbox": [ + 316, + 534, + 554, + 578 + ], + "type": "text", + "content": "[43] Lin Sun, Liangliang Cao, Yang Zheng, Bo Wang, Zhiming Li, Wenzhe Xie, and Xu Sun. Panacea: A framework for diverse, controllable generation of urban scenes. arXiv preprint arXiv:2207.10701, 2022. 3" + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 316, + 579, + 554, + 611 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 579, + 554, + 611 + ], + "spans": [ + { + "bbox": [ + 316, + 579, + 554, + 611 + ], + "type": "text", + "content": "[44] Alexander Swerdlow, Runsheng Xu, and Bolei Zhou. Streetview image generation from a bird's-eye view layout. IEEE Robotics and Automation Letters, 2024. 2" + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 316, + 613, + 554, + 657 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 613, + 554, + 657 + ], + "spans": [ + { + "bbox": [ + 316, + 613, + 554, + 657 + ], + "type": "text", + "content": "[45] Sergey Tulyakov, Ming-Yu Liu, Xiaodong Yang, and Jan Kautz. Mocogan: Decomposing motion and content for video generation. Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition (CVPR), 2018. 1" + } + ] + } + ], + "index": 25 + }, + { + "bbox": [ + 316, + 658, + 554, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 658, + 554, + 713 + ], + "spans": [ + { + "bbox": [ + 316, + 658, + 554, + 713 + ], + "type": "text", + "content": "[46] Narek Tumanyan, Michal Geyer, Shai Bagon, and Tali Dekel. Plug-and-play diffusion features for text-driven image-to-image translation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 1921-1930, 2023. 2, 3, 4, 5" + } + ] + } + ], + "index": 26 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 318, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 318, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 318, + 757 + ], + "type": "text", + "content": "27506" + } + ] + } + ], + "index": 28 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 9 + }, + { + "para_blocks": [ + { + "bbox": [ + 55, + 73, + 296, + 713 + ], + "type": "list", + "angle": 0, + "index": 13, + "blocks": [ + { + "bbox": [ + 56, + 73, + 296, + 95 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 73, + 296, + 95 + ], + "spans": [ + { + "bbox": [ + 56, + 73, + 296, + 95 + ], + "type": "text", + "content": "[47] A Vaswani. Attention is all you need. Advances in Neural Information Processing Systems, 2017. 3" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 55, + 96, + 296, + 151 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 96, + 296, + 151 + ], + "spans": [ + { + "bbox": [ + 55, + 96, + 296, + 151 + ], + "type": "text", + "content": "[48] Pascal Vincent, Hugo Larochelle, Yoshua Bengio, and Pierre-Antoine Manzagol. Extracting and composing robust features with denoising autoencoders. Proceedings of the 25th International Conference on Machine Learning, 2008. 1" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 56, + 152, + 296, + 207 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 152, + 296, + 207 + ], + "spans": [ + { + "bbox": [ + 56, + 152, + 296, + 207 + ], + "type": "text", + "content": "[49] Yan Wang, Wei-Lun Chao, Divyansh Garg, and et. al. Pseudo-lidar from visual depth estimation: Bridging the gap in 3d object detection for autonomous driving. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 8445-8453, 2019. 2" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 56, + 209, + 295, + 262 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 209, + 295, + 262 + ], + "spans": [ + { + "bbox": [ + 56, + 209, + 295, + 262 + ], + "type": "text", + "content": "[50] Hai Wu, Chenglu Wen, Shaoshuai Shi, Xin Li, and Cheng Wang. Virtual sparse convolution for multimodal 3d object detection. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 21653-21662, 2023. 2" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 56, + 264, + 296, + 319 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 264, + 296, + 319 + ], + "spans": [ + { + "bbox": [ + 56, + 264, + 296, + 319 + ], + "type": "text", + "content": "[51] Jinheng Xie, Yuexiang Li, Yawen Huang, Haozhe Liu, Wentian Zhang, Yefeng Zheng, and Mike Zheng Shou. Boxdiff: Text-to-image synthesis with training-free box-constrained diffusion. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 7452-7461, 2023. 3" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 56, + 321, + 296, + 376 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 321, + 296, + 376 + ], + "spans": [ + { + "bbox": [ + 56, + 321, + 296, + 376 + ], + "type": "text", + "content": "[52] Shaoyuan Xie, Lingdong Kong, Wenwei Zhang, Jiawei Ren, Liang Pan, Kai Chen, and Ziwei Liu. Benchmarking and improving bird's eye view perception robustness in autonomous driving. IEEE Transactions on Pattern Analysis and Machine Intelligence, 2025. 2, 6" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 56, + 377, + 296, + 432 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 377, + 296, + 432 + ], + "spans": [ + { + "bbox": [ + 56, + 377, + 296, + 432 + ], + "type": "text", + "content": "[53] Junkai Xu, Liang Peng, Haoran Cheng, Hao Li, Wei Qian, Ke Li, Wenxiao Wang, and Deng Cai. Mononerd: Nerflike representations for monocular 3d object detection. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 6814-6824, 2023. 2" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 56, + 434, + 296, + 487 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 434, + 296, + 487 + ], + "spans": [ + { + "bbox": [ + 56, + 434, + 296, + 487 + ], + "type": "text", + "content": "[54] Longfei Yan, Pei Yan, Shengzhou Xiong, Xuanyu Xiang, and Yihua Tan. Monocd: Monocular 3d object detection with complementary depths. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 10248-10257, 2024. 7" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 56, + 489, + 296, + 521 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 489, + 296, + 521 + ], + "spans": [ + { + "bbox": [ + 56, + 489, + 296, + 521 + ], + "type": "text", + "content": "[55] Xu Yan, Weiming Zeng, Jiaming Zhu, Hailin Chen, Yaowei Li, and Deng Cai. Videogpt: Video generation using vq-vae and transformers. arXiv preprint arXiv:2104.10157, 2021. 1" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 56, + 523, + 296, + 599 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 523, + 296, + 599 + ], + "spans": [ + { + "bbox": [ + 56, + 523, + 296, + 599 + ], + "type": "text", + "content": "[56] Chenyu Yang, Yuntao Chen, Hao Tian, Chenxin Tao, Xizhou Zhu, Zhaoxiang Zhang, Gao Huang, Hongyang Li, Yu Qiao, Lewei Lu, et al. Bevformer v2: Adapting modern image backbones to bird's-eye-view recognition via perspective supervision. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 17830-17839, 2023. 2" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 56, + 601, + 296, + 645 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 601, + 296, + 645 + ], + "spans": [ + { + "bbox": [ + 56, + 601, + 296, + 645 + ], + "type": "text", + "content": "[57] Kairui Yang, Enhui Ma, Jibin Peng, Qing Guo, Di Lin, and Kaicheng Yu. Bevcontrol: Accurately controlling street-view elements with multi-perspective consistency via bev sketch layout. arXiv preprint arXiv:2308.01661, 2023. 2" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 56, + 647, + 296, + 690 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 647, + 296, + 690 + ], + "spans": [ + { + "bbox": [ + 56, + 647, + 296, + 690 + ], + "type": "text", + "content": "[58] Tianwei Yin, Xingyi Zhou, and Philipp Krahenbuhl. Center-based 3d object detection and tracking. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 11784-11793, 2021. 2" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 56, + 692, + 296, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 692, + 296, + 713 + ], + "spans": [ + { + "bbox": [ + 56, + 692, + 296, + 713 + ], + "type": "text", + "content": "[59] Lvmin Zhang, Anyi Rao, and Maneesh Agrawala. Adding conditional control to text-to-image diffusion models. In" + } + ] + } + ], + "index": 12 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 316, + 73, + 553, + 653 + ], + "type": "list", + "angle": 0, + "index": 26, + "blocks": [ + { + "bbox": [ + 333, + 73, + 553, + 95 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 333, + 73, + 553, + 95 + ], + "spans": [ + { + "bbox": [ + 333, + 73, + 553, + 95 + ], + "type": "text", + "content": "Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 3836-3847, 2023. 2, 3, 4, 7" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 316, + 96, + 553, + 150 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 96, + 553, + 150 + ], + "spans": [ + { + "bbox": [ + 316, + 96, + 553, + 150 + ], + "type": "text", + "content": "[60] Renrui Zhang, Han Qiu, Tai Wang, Ziyu Guo, Ziteng Cui, Yu Qiao, Hongsheng Li, and Peng Gao. Monodetr: Depth-guided transformer for monocular 3d object detection. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 9155-9166, 2023. 2" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 316, + 152, + 553, + 185 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 152, + 553, + 185 + ], + "spans": [ + { + "bbox": [ + 316, + 152, + 553, + 185 + ], + "type": "text", + "content": "[61] Yifan Zhang and Bryan Hooi. Hipa: enabling one-step text-to-image diffusion models via high-frequency-promoting adaptation. arXiv preprint arXiv:2311.18158, 2023. 3" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 317, + 186, + 553, + 239 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 186, + 553, + 239 + ], + "spans": [ + { + "bbox": [ + 317, + 186, + 553, + 239 + ], + "type": "text", + "content": "[62] Yifan Zhang, Ying Wei, Qingyao Wu, Peilin Zhao, Shuaicheng Niu, Junzhou Huang, and Mingkui Tan. Collaborative unsupervised domain adaptation for medical image diagnosis. IEEE Transactions on Image Processing, 29: 7834-7844, 2020. 2" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 316, + 241, + 553, + 285 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 241, + 553, + 285 + ], + "spans": [ + { + "bbox": [ + 316, + 241, + 553, + 285 + ], + "type": "text", + "content": "[63] Yunpeng Zhang, Jiwen Lu, and Jie Zhou. Objects are different: Flexible monocular 3d object detection. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 3289-3298, 2021. 1, 2, 6, 7, 3, 4" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 316, + 287, + 553, + 330 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 287, + 553, + 330 + ], + "spans": [ + { + "bbox": [ + 316, + 287, + 553, + 330 + ], + "type": "text", + "content": "[64] Yifan Zhang, Bingyi Kang, Bryan Hooi, Shuicheng Yan, and Jiashi Feng. Deep long-tailed learning: A survey. IEEE transactions on pattern analysis and machine intelligence, 45(9):10795-10816, 2023. 2" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 316, + 331, + 553, + 374 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 331, + 553, + 374 + ], + "spans": [ + { + "bbox": [ + 316, + 331, + 553, + 374 + ], + "type": "text", + "content": "[65] Yifan Zhang, Daquan Zhou, Bryan Hooi, Kai Wang, and Ji-ashi Feng. Expanding small-scale datasets with guided imagination. In Advances in neural information processing systems, pages 76558–76618, 2023. 3" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 317, + 376, + 553, + 430 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 376, + 553, + 430 + ], + "spans": [ + { + "bbox": [ + 317, + 376, + 553, + 430 + ], + "type": "text", + "content": "[66] Bowen Zhao, Feng Lu, Cheng Zhang, Jiaqi Hu, and Lin Song. Drivedreamer: Realistic multi-view video generation of driving scenes with future prediction. Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition (CVPR), 2022. 3" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 316, + 432, + 553, + 485 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 432, + 553, + 485 + ], + "spans": [ + { + "bbox": [ + 316, + 432, + 553, + 485 + ], + "type": "text", + "content": "[67] Shihao Zhao, Dongdong Chen, Yen-Chun Chen, Jianmin Bao, Dongdong Zhang, Lu Yuan, Han Zhang, Dong Li, Baining Chen, and Lu Zhang. Uni-controlnet: All-in-one control to text-to-image diffusion models. arXiv preprint arXiv:2305.16322, 2023. 3" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 317, + 487, + 553, + 542 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 487, + 553, + 542 + ], + "spans": [ + { + "bbox": [ + 317, + 487, + 553, + 542 + ], + "type": "text", + "content": "[68] Guangcong Zheng, Xianpan Zhou, Xuewei Li, Zhongang Qi, Ying Shan, and Xi Li. Layoutdiffusion: Controllable diffusion model for layout-to-image generation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 22490-22499, 2023. 3" + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 317, + 544, + 553, + 597 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 544, + 553, + 597 + ], + "spans": [ + { + "bbox": [ + 317, + 544, + 553, + 597 + ], + "type": "text", + "content": "[69] Longtao Zheng, Yifan Zhang, Hanzhong Guo, Jiachun Pan, Zhenxiong Tan, Jiahao Lu, Chuanxin Tang, Bo An, and Shuicheng Yan. Memo: Memory-guided diffusion for expressive talking video generation. arXiv preprint arXiv:2412.04448, 2024. 3" + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 316, + 599, + 553, + 653 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 599, + 553, + 653 + ], + "spans": [ + { + "bbox": [ + 316, + 599, + 553, + 653 + ], + "type": "text", + "content": "[70] Bolei Zhou, Hang Zhao, Xavier Puig, Sanja Fidler, Adela Barriuso, and Antonio Torralba. Scene parsing through ade20k dataset. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 633-641, 2017. 7" + } + ] + } + ], + "index": 25 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "text", + "content": "27507" + } + ] + } + ], + "index": 27 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 10 + } + ], + "_backend": "vlm", + "_version_name": "2.6.4" +} \ No newline at end of file diff --git a/2025/DriveGPT4-V2_ Harnessing Large Language Model Capabilities for Enhanced Closed-Loop Autonomous Driving/31ff1a5c-7171-4c25-a1d0-ee2bacb8057b_content_list.json b/2025/DriveGPT4-V2_ Harnessing Large Language Model Capabilities for Enhanced Closed-Loop Autonomous Driving/31ff1a5c-7171-4c25-a1d0-ee2bacb8057b_content_list.json new file mode 100644 index 0000000000000000000000000000000000000000..3718c16c50c347ac5250caa987cba2b262d74db2 --- /dev/null +++ b/2025/DriveGPT4-V2_ Harnessing Large Language Model Capabilities for Enhanced Closed-Loop Autonomous Driving/31ff1a5c-7171-4c25-a1d0-ee2bacb8057b_content_list.json @@ -0,0 +1,1406 @@ +[ + { + "type": "text", + "text": "DriveGPT4-V2: Harnessing Large Language Model Capabilities for Enhanced Closed-Loop Autonomous Driving", + "text_level": 1, + "bbox": [ + 155, + 128, + 843, + 176 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Zhenhua Xu $^{1,2}$ Yan Bai $^{3}$ Yujia Zhang $^{1}$ Zhuoling Li $^{1}$ Fei Xia $^{3}$ Kwan-Yee K. Wong $^{1}$ Jianqiang Wang $^{2}$ Hengshuang Zhao $^{1*}$ The University of Hong Kong Tsinghua University Meituan", + "bbox": [ + 217, + 202, + 779, + 257 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Abstract", + "text_level": 1, + "bbox": [ + 246, + 291, + 326, + 306 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Multimodal large language models (MLLMs) possess the ability to comprehend visual images or videos, and show impressive reasoning ability thanks to the vast amounts of pretrained knowledge, making them highly suitable for autonomous driving applications. Unlike the previous work, DriveGPT4-V1, which focused on open-loop tasks, this study explores the capabilities of LLMs in enhancing closed-loop autonomous driving. DriveGPT4-V2 processes camera images and vehicle states as input to generate low-level control signals for end-to-end vehicle operation. A multi-view visual tokenizer (MV-VT) is employed enabling DriveGPT4-V2 to perceive the environment with an extensive range while maintaining critical details. The model architecture has been refined to improve decision prediction and inference speed. To further enhance the performance, an additional expert LLM is trained for online imitation learning. The expert LLM, sharing a similar structure with DriveGPT4-V2, can access privileged information about surrounding objects for more robust and reliable predictions. Experimental results show that DriveGPT4-V2 outperforms all baselines on the challenging CARLA Longest6 benchmark. The code and data of DriveGPT4-V2 will be publicly available.", + "bbox": [ + 89, + 323, + 483, + 672 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "1. Introduction", + "text_level": 1, + "bbox": [ + 89, + 700, + 220, + 715 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Autonomous driving has seen rapid advancements over the past decade in both academia and industry [31, 56]. Today, autonomous vehicles are deployed commercially in diverse applications such as robotaxi services [62], goods delivery [33], and security patrols [45]. Additionally, assisted driving systems are increasingly integrated into consumer vehicles, enhancing safety and convenience. However, the majority of current autonomous driving systems still rely heavily on rule-based decision-making modules, where planning and control are governed by predefined algorithms and", + "bbox": [ + 89, + 724, + 483, + 878 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/e2c2dfe1095507cf1537101d0eb554fb6c817cee0ee39103ff335f5c6fae8a11.jpg", + "image_caption": [ + "Figure 1. DriveGPT4-V2 for closed-loop autonomous driving. Taken as input multi-view camera images and vehicle state information, DriveGPT4-V2 predicts high-level vehicle decisions and converts them to low-level vehicle control signals in an end-to-end manner. DriveGPT4-V2 presents outstanding effectiveness and efficiency, serving as a reliable baseline method for future research on autonomous driving with LLMs." + ], + "image_footnote": [], + "bbox": [ + 516, + 290, + 905, + 506 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "rules. This reliance restricts the ability of these systems to operate effectively in varied, real-world scenarios, limiting the broader advancement of autonomous driving.", + "bbox": [ + 511, + 642, + 906, + 686 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "In response, recent research has shifted toward end-to-end autonomous driving systems based on learning-based methods [8, 9, 12, 17, 21, 23, 45]. These methods enable more streamlined and intelligent system designs by directly learning control from input sensor data, without the need for complicated intermediate modules. From a learning theory perspective, end-to-end autonomous driving can optimize the entire system on the final outputs, rather than through the isolated optimization of individual modules, which potentially improves overall performance.", + "bbox": [ + 511, + 688, + 908, + 839 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "The rise of large language models (LLMs) has demonstrated their impressive ability to perform natural language processing (NLP) tasks [1, 2, 13, 34, 49]. Commercial LLMs, such as ChatGPT [34] and Claude3 [1], are already", + "bbox": [ + 511, + 839, + 908, + 902 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "CVF", + "bbox": [ + 106, + 2, + 181, + 42 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore.", + "bbox": [ + 236, + 0, + 810, + 46 + ], + "page_idx": 0 + }, + { + "type": "page_footnote", + "text": "*Corresponding author.", + "bbox": [ + 109, + 887, + 235, + 898 + ], + "page_idx": 0 + }, + { + "type": "page_number", + "text": "17261", + "bbox": [ + 480, + 944, + 517, + 955 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "widely used in daily life. Motivated by their success in NLP tasks, LLMs are being extended to multimodal domains, enabling more generalized tasks in fields like image analysis [28-30], video comprehension [18, 26], and embodied AI [24, 25, 46, 55]. Given their versatility, multimodal LLMs have been applied to autonomous driving for tasks such as interpretability [15, 36, 44, 54], vehicle control [10, 27, 27], and risk assessment [3, 60]. While multimodal LLMs have demonstrated promising vehicle control performance due to their extensive pretrained knowledge, previous research has primarily concentrated on open-loop settings, which are not implemented for real-time vehicle control. Open-loop systems lack feedback mechanisms, making them less adaptable to changes and disturbances, limiting their applicability in real-world applications. Thus, multimodal LLMs need to be designed and evaluated specifically for closed-loop autonomous driving scenarios [27, 39, 43, 51].", + "bbox": [ + 89, + 90, + 480, + 347 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Behavior cloning [35, 48], the most commonly employed imitation learning algorithm, is used for most end-to-end autonomous driving works [44, 54]. However, models trained only by behavior cloning can encounter drifting issues in closed-loop tasks, where errors accumulate, potentially leading the vehicle into unsafe situations, such as deviating from the path, colliding, or getting stuck.", + "bbox": [ + 89, + 349, + 480, + 455 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "To address the aforementioned issue, we incorporate online (or on-policy) imitation learning [5, 40, 41] into closed-loop autonomous driving. Leveraging the reasoning capabilities and pretrained knowledge of LLMs, we propose using a multimodal LLM as a fully automatic expert for online imitation learning without human efforts. Let $M_E$ and $M_A$ denote the model of expert and agent, respectively. $M_E$ can access privileged information about nearby objects while $M_A$ cannot, thus presenting a more powerful and reliable performance. After being trained by behavior cloning, $M_A$ is deployed on the training environment for data aggregation. $M_A$ might make incorrect predictions that lead the vehicle into exception states. $M_E$ can provide $M_A$ with on-policy supervision and take over the vehicle if the error made by $M_A$ is large enough. When $M_E$ takes over the vehicle, it collects one data sample at the moment and aggregates it into the dataset. The aggregated dataset is leveraged to further train the agent model to handle errors.", + "bbox": [ + 89, + 459, + 482, + 731 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "In this paper, we introduce DriveGPT4-V2, an end-to-end autonomous driving system. DriveGPT4-V2 extends DriveGPT4-V1 [54] to closed-loop tasks and integrates online imitation learning for enhanced robustness and intelligence. The \"4\" in the system's name signifies its multimodal capacity, inspired by the MiniGPT4 model [65]. DriveGPT4-V2 is powered by multimodal LLMs, enabling it to directly generate low-level vehicle control signals (i.e., throttle, brake and steer) based on multimodal input data (i.e., vehicle states and camera images). The structure of DriveGPT4-V2 is refined for our task, including a", + "bbox": [ + 89, + 734, + 482, + 900 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "multi-view visual tokenizer (MV-VT) that possesses a large enough perception range without losing details, as well as output decision heads (DeciHeads) specially designed for numerical vehicle decision prediction. An additional multimodal LLM with access to privileged information (e.g., ground truth of nearby objects in text) serves as an expert for online imitation learning, which is inspired by the LBC framework [5]. DriveGPT4-V2 achieves state-of-the-art performance on the challenging CARLA Longest6 benchmark [12], which evaluates systems over 36 long routes demanding high levels of intelligence and reliability. The overview of DriveGPT4-V2 is visualized in Fig. 1. The contributions of this paper are:", + "bbox": [ + 511, + 90, + 903, + 287 + ], + "page_idx": 1 + }, + { + "type": "list", + "sub_type": "text", + "list_items": [ + "- We extend DriveGPT4-V1, an open-loop autonomous driving system with multimodal LLMs, and propose DriveGPT4-V2 for end-to-end closed-loop autonomous driving. DriveGPT4-V2 can better harness the capability of LLMs and be deployed for continuous vehicle control.", + "- We introduce the use of LLMs with privileged information as an expert for online (on-policy) imitation learning, facilitating the generation of labeled data for handling errors, which significantly enhances the performance and robustness of closed-loop autonomous driving.", + "- We optimize the model structure of MLLMs, including a multi-view visual tokenizer (MV-VT) that possesses a large perception range and preserves details, as well as output decision heads (DeciHeads) specially designed for numerical vehicle decision prediction.", + "- DriveGPT4-V2 demonstrates superior performance on the CARLA Longest6 benchmark, substantially outperforming existing methods." + ], + "bbox": [ + 513, + 287, + 903, + 559 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "2. Related Works", + "text_level": 1, + "bbox": [ + 513, + 571, + 661, + 588 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Large language models for autonomous driving. Owing to the powerful reasoning capabilities and extensive pretrained knowledge, large language models (LLMs) have found wide application in autonomous driving tasks, such as vehicle decision-making [7, 10, 15, 27, 32, 39, 43, 54, 59], scene understanding [20, 36, 44, 54], risk analysis [3, 60], and simulation [47, 50]. GPT-Driver [32] represents the entire surrounding environment textually and employs a fine-tuned ChatGPT model to predict vehicle trajectories. DriveGPT4 [54] leverages multimodal LLMs for end-to-end vehicle action prediction and scene understanding. It processes 8-frame video clips and human user commands to predict vehicle control signals and action explanations. However, most of these approaches are designed for open-loop scenarios and present unsatisfactory performance in closed-loop tasks, limiting their applicability to real-world driving deployment. Additionally, many existing methods face efficiency challenges due to heavy model architectures.", + "bbox": [ + 511, + 598, + 906, + 869 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Closed-loop end-to-end autonomous driving. Closed-loop end-to-end autonomous driving has recently emerged", + "bbox": [ + 511, + 869, + 903, + 900 + ], + "page_idx": 1 + }, + { + "type": "page_number", + "text": "17262", + "bbox": [ + 480, + 944, + 517, + 955 + ], + "page_idx": 1 + }, + { + "type": "image", + "img_path": "images/955fc14f1df6defcabc113baf58f6b04fddc39a90f3b17b5024c9d5e203d8c8c.jpg", + "image_caption": [ + "Figure 2. DriveGPT4-V2 diagram. DriveGPT4-V2 takes multimodal input data to generate numerical control signals for end-to-end vehicle driving. The input includes multi-view images and vehicle state information. The images are transformed into the text domain via a multi-view vision tokenizer (MV-VT). The current speed of the vehicle and target point are tokenized by the LLM tokenizer. The LLM then outputs four tokens. Each token is used to predict one vehicle decision by an MLP decision head (DeciHead). These predicted decisions are subsequently converted to low-level commands via PID controllers to operate the vehicle. The LLM expert model, which shares a similar structure to DriveGPT4-V2, has access to privileged information about surroundings (shown in the purple module). The expert provides on-policy supervision to DriveGPT4-V2 to enhance closed-loop performance. This figure is best viewed in color." + ], + "image_footnote": [], + "bbox": [ + 98, + 92, + 898, + 435 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "as a prominent research focus, addressing the challenge of bridging the gap between academic research and real-world applications. Various approaches improved transformer architectures and training methods to enhance driving performance [4, 5, 12, 16, 21, 22, 42, 58, 63]. For instance, Transfuser++ [21] integrates multimodal sensor inputs, including camera imagery, LiDAR point clouds and vehicle dynamics, to predict speed and trajectory, achieving state-of-the-art performance across multiple benchmarks. Even though most early LLM-based methods have primarily targeted open-loop tasks, recent studies [27, 39, 43, 51] have attempted closed-loop approaches. But most of them are deployed in simple environments like Highway-env [51], instead of real-world scenarios. LMDrive [43], pioneering work on end-to-end closed-loop driving with multimodal LLMs, demonstrated the potential of LLMs in this domain. It is deployed in the CARLA simulator [14] and controls the vehicle like human drivers. However, it relies solely on behavior cloning and presents unsatisfactory driving performance. Online imitation learning has been highlighted as a potential way to enhance model robustness.", + "bbox": [ + 89, + 575, + 485, + 893 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Imitation learning in autonomous driving. Imitation learning [19] aims to train an agent to mimic the expert behavior. Behavior cloning is the most widely used imitation learning algorithm in autonomous driving [4, 12, 21, 39, 43, 44, 52-54]. In this method, data is off-policy sampled from expert trajectories for model training to replicate expert actions. Although behavior cloning can yield reasonable performance, it suffers from error accumulation during closed-loop operation, leading the model to encounter unseen situations that may compromise safety. To mitigate this issue, online imitation learning [40, 41] has been employed in various works [4, 5]. The LBC framework [5] uses the DAgger algorithm [41], and trains an agent alongside an expert network with privileged information that achieves superior performance. The expert provides on-policy supervision to the agent and collects additional data samples for data aggregation, which can be used to further train the agent model to handle errors. Inspired by LBC, DriveGPT4-V2 harnesses the capabilities of LLMs and leverages a privileged LLM expert to provide on-policy supervision, facilitating improved driving performance and robustness.", + "bbox": [ + 511, + 575, + 908, + 893 + ], + "page_idx": 2 + }, + { + "type": "page_number", + "text": "17263", + "bbox": [ + 480, + 944, + 519, + 957 + ], + "page_idx": 2 + }, + { + "type": "image", + "img_path": "images/995738dfecd7412e6c9b2dddf492c5aeb9f0b263cf8538985befe2873bc33c90.jpg", + "image_caption": [ + "Figure 3. Multi-view visual tokenizer (MV-VT) structure. The input images consist of three front views. Each patch is processed through a visual encoder to extract features. Finally, a trained projection layer maps the downsampled feature into the text domain for further processing.", + "Figure 4. Visualization of waypoints and route points. The ego vehicle is represented by the green rectangle, and the red point denotes the target point. The grey line is the route that the vehicle should follow. (a) Waypoints (blue points) represent vehicle positions in a 4-second future. (b) Route points are evenly sampled from the global planned route in front. These two kinds of points can better supervise the training of DriveGPT4-V2." + ], + "image_footnote": [], + "bbox": [ + 99, + 90, + 473, + 306 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "3. DriveGPT4-V2", + "text_level": 1, + "bbox": [ + 89, + 417, + 243, + 431 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Building on DriveGPT4-V1 [54], a flexible multimodal LLM designed for open-loop autonomous driving, DriveGPT4-V2 has been specifically developed for closed-loop driving tasks. This paper employs LLMs both as planners and as expert models for online imitation learning. A multi-view visual tokenizer (MV-VT) is leveraged for effective feature extraction. Unlike standard LLMs used for NLP, the architecture of DriveGPT4-V2 is tailored for vehicle planning, with additional output decision heads (DeciHeads) to directly optimize numerical decision-making predictions. This design enhances model efficiency by reducing the token count required for inference. The expert model, which shares a similar architecture to DriveGPT4-V2, has access to privileged ground-truth information about the surroundings and is used solely for on-policy supervision during training, not for inference.", + "bbox": [ + 88, + 443, + 482, + 685 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "3.1. Model Design", + "text_level": 1, + "bbox": [ + 89, + 696, + 233, + 712 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "DriveGPT4-V2 takes multimodal input data and predicts high-level vehicle decisions for end-to-end autonomous driving. These predictions are subsequently converted into low-level control signals through PID controllers. The model structure is shown in Fig. 2.", + "bbox": [ + 89, + 718, + 482, + 794 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Multimodal input. The input consists of two modalities: image and text. The image input is captured by three front-view cameras at a resolution of $384 \\times 384$ , while the text input includes the vehicle's current speed and the target point. These three front-view images ensure coverage of both sides of the vehicle, reducing the risk of missing adjacent vehicles or potential collisions. The image data is", + "bbox": [ + 89, + 795, + 483, + 901 + ], + "page_idx": 3 + }, + { + "type": "image", + "img_path": "images/c33d38c0cd5c96737f087a11329e34e686c48f64f9cfa5cb70a639771da57302.jpg", + "image_caption": [ + "(a) Waypoints.", + "(b) Route points." + ], + "image_footnote": [], + "bbox": [ + 514, + 95, + 906, + 104 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "mapped into the text domain via a visual tokenizer, as implemented in LLaVA [30], while text inputs are tokenized with standard LLM tokenizers.", + "bbox": [ + 511, + 263, + 906, + 306 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Multi-view visual tokenizer (MV-VT). The visual tokenizer is visualized in Fig. 3. It first employs pretrained visual encoders, such as CLIP [37] or SigLIP [57], to extract features from the input images. The size of each image is specifically optimized for compatibility with the SigLIP encoder (i.e., $384 \\times 384$ ). The three images are acquired from front-left, front, and front-right cameras respectively, ensuring comprehensive coverage of the vehicle's frontal and lateral views. In CARLA benchmarks, back-view images exhibit limited utilization frequency and demonstrate no measurable improvement in planning performance. To maintain computational efficiency, MV-VT deliberately excludes back-view inputs.", + "bbox": [ + 511, + 309, + 905, + 503 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Output representations. DriveGPT4-V2 generates three low-level control signals to drive the vehicle, i.e., throttle, brake, and steering. However, direct prediction of these signals requires temporal consistency; otherwise, the vehicle may display unstable zigzag trajectories. To ensure stable control, DriveGPT4-V2 instead predicts high-level decision variables, which are then converted to low-level commands by PID controllers. The predictions of the model include:", + "bbox": [ + 511, + 506, + 905, + 626 + ], + "page_idx": 3 + }, + { + "type": "list", + "sub_type": "text", + "list_items": [ + "- Target Speed: The desired speed for the vehicle. A speed PID controller adjusts longitudinal control based on the predicted target speed and current speed.", + "- Target Angle: The direction in which the vehicle should steer. A steering PID controller uses it for lateral control.", + "- Waypoints: Predicted x-y positions of the vehicle in the future. Waypoints consist of eight points.", + "- Route Points: Points evenly sampled along the global route in front, enhancing the vehicle's perception ability." + ], + "bbox": [ + 511, + 627, + 903, + 762 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Waypoints and route points are visualized in Fig. 4.", + "bbox": [ + 511, + 763, + 854, + 777 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Decision heads (DeciHeads). For closed-loop autonomous driving, DriveGPT4-V2 employs dedicated output heads (MLPs) to predict numerical vehicle decisions. The original LLM's vocabulary output head is replaced with four decision heads, each predicting a decision variable (i.e., target speed, target angle, waypoints, and route points) based on one token. This setup allows DriveGPT4-V2 to produce only four tokens per prediction cycle, achieving a huge", + "bbox": [ + 511, + 779, + 906, + 900 + ], + "page_idx": 3 + }, + { + "type": "page_number", + "text": "17264", + "bbox": [ + 480, + 944, + 519, + 955 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "speed increase over traditional text-based outputs. For example, a float number with 3 digits (e.g., 1.538) needs 5 tokens for text output representation but only requires a single token for our design. DeciHead also enables direct optimization of decision variables using regression loss.", + "bbox": [ + 89, + 90, + 480, + 166 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Expert model and online imitation learning. The expert model has a similar architecture to DriveGPT4-V2 but has access to privileged information. The privileged information consists of ground truth (GT) of nearby objects and hazard information provided by the CARLA simulator (e.g., traffic light and stop sign violation; vehicle or pedestrian collision). This privileged data allows the expert to better perceive the environment, so that it can focus on vehicle decisions and planning. The privileged information greatly boosts the model capacity, resulting in enhanced closed-loop performance. The expert model provides on-policy supervision to DriveGPT4-V2 during training.", + "bbox": [ + 89, + 167, + 482, + 347 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "When DriveGPT4-V2 runs on training scenarios and routes, the expert model receives identical input and generates concurrent predictions. If the discrepancy between DriveGPT4-V2's prediction and the expert's exceeds a threshold, the expert's prediction is used instead to control the vehicle. Such a situation is marked as an exception case. A data sample of the current moment is added to the training set, where the expert's output is used as the label. DriveGPT4-V2 is fine-tuned on this aggregated dataset under the expert's on-policy supervision for enhanced closed-loop autonomous driving performance.", + "bbox": [ + 89, + 347, + 482, + 513 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "The expert LLM is used exclusively for on-policy supervision during training and is not accessed during inference.", + "bbox": [ + 89, + 513, + 482, + 545 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "3.2. Training", + "text_level": 1, + "bbox": [ + 89, + 553, + 194, + 569 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "The training process of DriveGPT4-V2 and the expert LLM includes two stages. In the first stage, both models are trained on data collected by a rule-based autopilot using behavior cloning. In the second stage, DriveGPT4-V2 runs on the training scenarios and routes under the expert's on-policy supervision, and generates additional data for error correction. The training process is visualized in Fig. 5.", + "bbox": [ + 89, + 575, + 482, + 681 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "3.2.1. Stage 1: behavior cloning", + "text_level": 1, + "bbox": [ + 89, + 689, + 316, + 704 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "In the first stage, a rule-based autopilot [21] is deployed in CARLA to collect data. The autopilot agent records data at fixed time intervals, including camera images, current vehicle speed, privileged information on surrounding objects (e.g., vehicles and traffic lights), and a target point. Labels for the vehicle decision variables are calculated using rule-based algorithms. After data collection, both DriveGPT4-V2 and the expert LLM are trained on the dataset, with the expert having access to privileged information.", + "bbox": [ + 89, + 708, + 482, + 843 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "3.2.2. Stage 2: DAgger with on-policy supervision", + "text_level": 1, + "bbox": [ + 89, + 851, + 439, + 867 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "The expert, with access to privileged information, achieves notably higher performance by leveraging reliable priors on", + "bbox": [ + 89, + 869, + 482, + 901 + ], + "page_idx": 4 + }, + { + "type": "image", + "img_path": "images/501a1f5a0dfb3947b86c41c55fe6c3d1a51e80bb06853d866e54b38248154e56.jpg", + "image_caption": [ + "(a) Stage 1: Behavior cloning." + ], + "image_footnote": [], + "bbox": [ + 540, + 89, + 898, + 186 + ], + "page_idx": 4 + }, + { + "type": "image", + "img_path": "images/d7e4f976a0eeeff15f2c7c35142437c32eb23e636d668a8e8889e42777e2c61c.jpg", + "image_caption": [ + "(b) Stage 2: DAgger online imitation learning.", + "Figure 5. Diagram of the two-stage training process. (a) In the first stage, both DriveGPT4-V2 and the expert LLM are trained on data collected by a rule-based autopilot. (b) In the second stage, DriveGPT4-V2 runs on the training scenarios and routes. When the discrepancy between DriveGPT4-V2's predictions and those of the expert exceeds a predefined threshold, the expert's predictions are used to control the vehicle. Data from these cases is then added to the dataset for data aggregation." + ], + "image_footnote": [], + "bbox": [ + 517, + 205, + 919, + 339 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "surrounding objects, allowing it to handle challenging scenarios effectively. DriveGPT4-V2, trained only by behavior cloning, may struggle with error accumulation and exceptional situations (e.g., vehicle collisions or getting stuck). Inspired by LBC [5], the expert LLM is used as a teacher for on-policy supervision. On-policy supervision is a training paradigm in imitation learning, where supervision or feedback is provided to an agent while it is acting according to its current policy. This approach contrasts with off-policy methods, where supervision comes from data collected by a different policy or behavior. During the operation of DriveGPT4-V2, the expert and DriveGPT4-V2 receive identical input, and their outputs are compared. If the discrepancy in predicted decisions exceeds a threshold, the expert takes control to avoid exceptions. One data sample is collected at that moment and aggregated into the dataset. The output of the expert is used as the training label. After traversing the training routes under expert supervision, DriveGPT4-V2 is fine-tuned on the aggregated dataset.", + "bbox": [ + 511, + 507, + 906, + 794 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "3.3. Dataset", + "text_level": 1, + "bbox": [ + 513, + 803, + 609, + 816 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Following Transfuser++ [21], in stage 1, the rule-based autopilot agent is used to collect data samples in various CARLA simulator towns under varying weather, lighting, and daytime conditions. Vehicle speed and position are estimated by Kalman filters based on IMU data. Each sam", + "bbox": [ + 511, + 824, + 905, + 900 + ], + "page_idx": 4 + }, + { + "type": "page_number", + "text": "17265", + "bbox": [ + 480, + 944, + 517, + 955 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "ple records objects within a fixed radius, including vehicles, pedestrians, traffic lights, and stop signs. Target speeds and angles are calculated by the autopilot's PID controller. The ground-truth waypoints contain 8 points and record the vehicle position in a 4-second future. Route points have 10 points sampled along the planned route in front.", + "bbox": [ + 88, + 90, + 480, + 181 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "In stage 2, DriveGPT4-V2 runs on selected training routes for data aggregation. Simple routes are not considered in this stage for data collection efficiency. If the difference between the predicted decisions of DriveGPT4-V2 and the expert exceeds a threshold, the expert takes control and records the data sample. DriveGPT4-V2 is retrained on the aggregated dataset for better performance. Under common circumstances, DriveGPT4-V2 achieves satisfactory performance after one round of DAgger.", + "bbox": [ + 89, + 181, + 482, + 316 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "3.4. Loss Function", + "text_level": 1, + "bbox": [ + 89, + 324, + 235, + 339 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "During training, the L1 loss function supervises all predicted decision variables, including target speed $(TS)$ , target angle $(TA)$ , waypoints $(WP)$ , and route points $(RP)$ . The overall loss is calculated by:", + "bbox": [ + 89, + 347, + 482, + 406 + ], + "page_idx": 5 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {L} = \\mathcal {L} _ {T S} + \\mathcal {L} _ {T A} + \\mathcal {L} _ {W P} + \\mathcal {L} _ {R P} \\tag {1}\n$$\n", + "text_format": "latex", + "bbox": [ + 171, + 412, + 480, + 429 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "This approach allows the model to optimize final numerical vehicle decision variables directly, instead of relying on vocabulary-based classification as in NLP tasks.", + "bbox": [ + 89, + 434, + 482, + 479 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "4. Experiments", + "text_level": 1, + "bbox": [ + 89, + 491, + 220, + 508 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "4.1. Dataset and Benchmark", + "text_level": 1, + "bbox": [ + 89, + 516, + 312, + 530 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "All experiments in this study were conducted in the CARLA simulator [14]. We followed Transfuser++ for training data preparation. In stage 1, a rule-based autopilot was employed to collect 350K behavior-cloning samples. The autopilot traversed multiple routes in Towns 1, 2, 3, 4, 5, 6, 7, and 10 of CARLA, encountering various scenarios such as different weather and lighting conditions. Data were collected at a frequency of $2\\mathrm{Hz}$ and subsequently filtered to remove incorrect or redundant samples, such as those captured when the autopilot was stuck. This process resulted in a dataset of 300K samples.", + "bbox": [ + 89, + 537, + 482, + 703 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "In stage 2, DriveGPT4-V2 trained by behavior cloning runs on selected training scenarios and routes, utilizing the expert for on-policy supervision. The expert can access privileged information about objects within a 30-meter distance. Approximately 150K samples were gathered in this stage. These aggregated data were then combined with the original dataset to further fine-tune DriveGPT4-V2, enhancing its performance.", + "bbox": [ + 89, + 704, + 482, + 824 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "All methods were evaluated on the challenging CARLA Longest6 benchmark [12], which consists of 36 extended routes encompassing various scenarios. The length and complexity of routes make them suitable for a comprehensive evaluation of closed-loop autonomous driving.", + "bbox": [ + 89, + 825, + 482, + 900 + ], + "page_idx": 5 + }, + { + "type": "table", + "img_path": "images/1bcd09b5ebe2bcdb690dda945a2fb256bb939c7af0d5d8b7a80ee1b7dcb764da.jpg", + "table_caption": [ + "Table 1. Infraction penalty factors." + ], + "table_footnote": [], + "table_body": "
ScenePenaltyScenePenalty
Pedestrian collision0.5Vehicle collision0.6
Static object collision0.65Traffic light violation0.7
Stop sign violation0.7
", + "bbox": [ + 514, + 114, + 903, + 186 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "4.2. DriveGPT4-V2 Configuration", + "text_level": 1, + "bbox": [ + 511, + 210, + 777, + 227 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "LLM. For efficient data collection and inference, we used tiny-scale LLMs in DriveGPT4-V2, such as Qwen-0.5B [2] and TinyLLaMA [61]. These models were fine-tuned for multimodal understanding, following methods outlined in LLaVA [29, 30] and TinyLLaVA [64]. Large-scale LLMs, like LLaMA3-8B [49] and vicuna-7B [11], are not considered at the moment due to their high computational demands and reduced inference speed.", + "bbox": [ + 511, + 233, + 903, + 353 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Training. DriveGPT4-V2 was fully fine-tuned in an end-to-end manner, except for the SigLIP $384 \\times 384$ visual encoder [57]. The model was trained with a learning rate of $2e^{-5}$ over 60 epochs by employing a cosine annealing schedule for better learning rate control. The training takes about 75 hours on 16 A800 GPUs.", + "bbox": [ + 511, + 354, + 903, + 443 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Inference. Two PID controllers were designed to convert DriveGPT4-V2's predicted decision variables into low-level control signals (i.e., throttle, steer, and brake). These control signals were used to directly drive the ego vehicle in the CARLA simulator. In the experiments, only the predicted target speed and target angle were used by the PID controllers, while waypoints and route points served as additional supervision for improved training results. The model runs inference on a single A800 GPU with FP16 precision.", + "bbox": [ + 511, + 445, + 905, + 580 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "4.3. Evaluation Criteria", + "text_level": 1, + "bbox": [ + 511, + 590, + 699, + 604 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Following the CARLA leaderboard, the methods were evaluated mainly based on three metrics: Driving Score (DS), Route Completion (RC), and Infraction Score (IS). RC indicates the percentage of routes completed, while IS penalizes infractions such as traffic violations and collisions during inference. Each infraction multiplicatively reduces IS by a penalty factor. The penalty factors are listed in Tab. 1. DS, as the product of RC and IS, provides a comprehensive evaluation score. In the evaluation, we also report ratios of commonly seen exception events, including pedestrian collision (Ped), vehicle collision (Veh), static object collision (Stat), red light violation (Red), route deviations (Dev), scenario time out (TO) and agent block (Block).", + "bbox": [ + 511, + 612, + 905, + 808 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "4.4. Comparison Experiments", + "text_level": 1, + "bbox": [ + 511, + 818, + 746, + 834 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Baselines. We selected several representative prior works as baselines for comparison:", + "bbox": [ + 511, + 840, + 903, + 869 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "- WOR [6] (ICCV 2021): WOR uses commands instead of target points for global planning instructions, providing", + "bbox": [ + 513, + 869, + 903, + 900 + ], + "page_idx": 5 + }, + { + "type": "page_number", + "text": "17266", + "bbox": [ + 480, + 944, + 517, + 955 + ], + "page_idx": 5 + }, + { + "type": "table", + "img_path": "images/2ffcb13e728c53ea93460d7aa573e4cec76d457684b9e19ab53cd0ebad510c73.jpg", + "table_caption": [ + "Table 2. Closed-loop experiments performance on CARLA Longest6. \"Visual\" indicates visual input modalities, while \"C\" and \"L\" represent camera and LiDAR, respectively. Bold numbers highlight SOTA metric scores of all models; while underlined numbers represent the best metric scores of baseline methods. * denotes models implemented by ourselves based on official open-sourced code. † represents the model without data augmentation." + ], + "table_footnote": [], + "table_body": "
MethodVisualDS↑RC↑IS↑Ped↓Veh↓Stat↓Red↓Dev↓TO↓Block↓
WOR [6]C21480.560.181.050.371.280.880.080.20
LAV v1 [4]C&L33700.510.160.830.150.960.060.120.45
Interfuser [42]C47740.630.061.140.110.240.000.520.06
TransFuser [12]C&L47930.500.032.450.070.160.000.060.10
LAV v2 [4]C&L58830.680.000.690.150.230.080.320.11
Perception PlanT [38]C&L58880.650.070.970.110.090.000.130.13
Transfuser++* [21]C&L65900.720.000.990.010.070.000.100.12
Transfuser++*† [21]C&L58890.650.011.150.010.100.000.140.13
LMDrive* [43]C&L36690.520.071.030.181.010.090.110.22
DriveGPT4-V2C70910.770.000.800.010.040.000.070.09
", + "bbox": [ + 91, + 155, + 906, + 339 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "enriched labels and supervision for possible actions.", + "bbox": [ + 102, + 363, + 447, + 378 + ], + "page_idx": 6 + }, + { + "type": "list", + "sub_type": "text", + "list_items": [ + "- LAV v1 & v2 [4] (CVPR 2022): LAV predicts waypoints by integrating information from nearby vehicles, which is further refined through a GRU.", + "- Transfuser [12] (TPAMI 2022): Transfuser is a widely used baseline model for closed-loop driving that has inspired subsequent works.", + "- Perception PlanT [38] (CoRL 2022): PlanT is based on imitation learning with object-level input representation. Such representation is elegant and effective.", + "- InterFuser [42] (CoRL 2023): InterFuser predicts additional density maps and traffic rule flags to enhance waypoint prediction.", + "- Transfuser++ [21] (ICCV 2023): Transfuser++ predicts target speeds and routes instead of waypoints, integrating point cloud and camera images. It is currently the state-of-the-art (SOTA) method for CARLA Longest6. Transfuser++ utilizes triple training data by repeatedly collecting expert trajectories for augmentation.", + "- LMDrive [43] (CVPR 2024): LMDrive is a pioneering work for closed-loop driving with multimodal LLMs, training a vicuna-7B LLM for end-to-end vehicle control." + ], + "bbox": [ + 89, + 378, + 480, + 695 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Comparison results. The comparison results on the CARLA Longest6 benchmark are shown in Tab. 2. It is observed that DriveGPT4-V2 outperforms all baseline models by a large margin. Thanks to the huge amount pretrained knowledge of LLMs and visual encoders, DriveGPT4-V2 can better handle complex urban scenes under various conditions, such as nighttime, rainy weather, etc. For most baselines, we list the results reported on the CARLA Longest6 leaderboard. However, some baselines cannot achieve reported performance by our implementation and test, thus we report our implemented results separately. LMDrive [43] is the only open-sourced LLM method that can be tested in a closed-loop environment. But it fails", + "bbox": [ + 89, + 704, + 482, + 900 + ], + "page_idx": 6 + }, + { + "type": "table", + "img_path": "images/d1dbaecddd538c546c66bec96d5ac8fe3860c661fead26e6998d4ec67bf28290.jpg", + "table_caption": [ + "Table 3. Efficiency analysis." + ], + "table_footnote": [], + "table_body": "
LLMDSTrainFPS
LLaVA-LLaMA3.1-8B6511.2h/epoch0.4
TinyLLaVA-LLaMA-1.5B633.0h/epoch2.9
LLaVA-Qwen-0.5B631.3h/epoch8.1
", + "bbox": [ + 514, + 386, + 908, + 460 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "to obtain satisfactory performance. It first converts multimodal data to BEV and projects BEV to the text domain. However, BEV does not have good-quality pretrained multimodal LLMs. Training from scratch severely degrades its final performance, which is also discussed by DriveGPT4-V1 [54]. Transfuser++ [21] is the SOTA method by fusing point clouds and camera images for closed-loop control. Transfuser++ uses triple training data and achieves 65 DS, but only has 58 DS with the same amount of training data (denoted by $\\dagger$ ) as our method. Thanks to the pretrained knowledge and reasoning ability of LLMs, DriveGPT4-V2 can outperform all baselines with camera images as input.", + "bbox": [ + 511, + 486, + 906, + 667 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "4.5. Model Efficiency", + "text_level": 1, + "bbox": [ + 511, + 681, + 679, + 696 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "In our experiments, DriveGPT4-V2 and the expert are implemented by tiny LLMs. LLMs with 7B or even more parameters show powerful reasoning and generalization ability, but they severely enlarge the time consumption and demand much more training resources, which is unacceptable in real-world applications. In this section, we report efficiency results of DriveGPT4-V2 with LLaVA-0.5B, TinyLLaVA-1.5B and LLaVA-7B LLMs. The results are shown in Tab. 3. It is found that merely scaling up the LLM model does not gain corresponding performance enhancement, but slows down the whole system more than 10 times. Therefore, DriveGPT4-V2 mainly considers 0.5B LLMs as the planner.", + "bbox": [ + 511, + 704, + 906, + 900 + ], + "page_idx": 6 + }, + { + "type": "page_number", + "text": "17267", + "bbox": [ + 480, + 944, + 517, + 955 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "4.6. Ablation Studies and Discussion", + "text_level": 1, + "bbox": [ + 89, + 90, + 372, + 104 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "We conducted multiple ablation studies to justify the design of DriveGPT4-V2. The results are presented in Tab. 4, 5 and 6. Due to huge time consumption, DriveGPT4-V2 in ablation studies is not trained with DAgger data by default.", + "bbox": [ + 89, + 112, + 480, + 172 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "LLM visual pretraining. DriveGPT4-V2 relies on LLaVA [30, 64] as the planner, which is pretrained on a huge amount of visual understanding data. The visual pretraining benefits DriveGPT4-V2 on the autonomous driving task with multimodal input. Directly training DriveGPT4-V2 on scratch (i.e., LLM for NLP tasks without multimodal ability) slows down convergence and degrades the final results.", + "bbox": [ + 89, + 172, + 480, + 279 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Visual tokenizer. The input camera images are processed by the proposed visual tokenizer. Inspired by LLaVA 1.5 [30], the features of images are extracted separately by the visual encoder and then concatenated as the input. This design ensures a sufficient perception range without compromising image detail. If the input is a single image with a large scope, it might lead to information loss during preprocessing steps (e.g., resize or crop); if the input is a small square image, the vehicle cannot have a sufficient perception range for effective planning.", + "bbox": [ + 89, + 279, + 480, + 431 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Waypoints and route points. Although DriveGPT4-V2 does not rely on waypoints and route points for generating low-level control signals, they provide valuable supervision during training. Predicting waypoints and route points helps the model perceive the target route and anticipate future actions. Removing them negatively impacts performance.", + "bbox": [ + 89, + 431, + 480, + 521 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Expert on-policy supervision. The expert model, benefiting from privileged information, provides robust on-policy supervision to DriveGPT4-V2. Without expert guidance, DriveGPT4-V2 relies solely on behavior cloning instead of DAGger imitation learning, which can lead to error accumulation and degraded performance.", + "bbox": [ + 89, + 521, + 480, + 612 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "PID controllers. Previous works often generated low-level control signals based on waypoints or route points [12, 21] by calculating target speed and angle from these predictions. However, even minor prediction errors in waypoints or route points can lead to substantial errors in speed and angle. As shown in Tab. 5, using waypoints or route points for PID control can make control signals noisy and unstable, potentially driving the vehicle into exceptional states. Therefore it presents inferior results.", + "bbox": [ + 89, + 613, + 480, + 748 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Decision heads. During inference, DriveGPT4-V2 generates four tokens in an auto-regressive manner, mapped to vehicle decision variables by four decision heads. Predictions of target speed and angle involve one-to-one mappings (i.e., one token for one number), while waypoint and route point predictions involve one-to-many mappings (e.g., waypoints contain 8 x-y points, thus each token is mapped to 16 numbers). Another design would involve predicting each waypoint and route point one at a time, requiring 8 tokens for waypoints and 10 tokens for route points. Such design", + "bbox": [ + 89, + 750, + 482, + 901 + ], + "page_idx": 7 + }, + { + "type": "table", + "img_path": "images/582556cd97569854a55b8c0bef212c1740407aaacc1c0df524794c42e9a52a3f.jpg", + "table_caption": [ + "Table 4. Ablation studies of DriveGPT4-V2. \"WP\" and \"RP\" represent waypoints and route points, respectively." + ], + "table_footnote": [], + "table_body": "
DSRCIS
Baseline47780.60
+ LLM Visual Pretraining56870.64
+ Visual Tokenizer60880.68
+ WP&RP63900.70
+ Expert Supervision70910.77
", + "bbox": [ + 516, + 122, + 900, + 222 + ], + "page_idx": 7 + }, + { + "type": "table", + "img_path": "images/a1112eaa6830ab7fb1654f13e6b848c0d3b07bc7a7d964a6d0d2b7fe56fe7b29.jpg", + "table_caption": [ + "Table 5. Ablation studies on PID controllers. \"WP\" indicates utilizing predicted waypoints for PID control; while \"TS&RP\" means PID control by predicted target speed and route points." + ], + "table_footnote": [], + "table_body": "
PID ControllerDSRCIS
WP53850.62
TS & RP59880.67
DriveGPT4-V263900.70
", + "bbox": [ + 517, + 275, + 898, + 353 + ], + "page_idx": 7 + }, + { + "type": "table", + "img_path": "images/8668107942c6bcfbdf41c9c51c64a5833ea2912a0c375fffa81dc5c05f6ca9b4.jpg", + "table_caption": [ + "Table 6. Ablation studies on decision heads. \"Additional tokens\" indicates using more output tokens for prediction." + ], + "table_footnote": [], + "table_body": "
DSRCISFPS
Additional tokens64910.701.4
DriveGPT4-V263900.708.1
", + "bbox": [ + 516, + 393, + 900, + 458 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "does not significantly improve performance and severely affects efficiency. Using text to represent output numbers would require approximately 160 tokens, making it impractical for real-time autonomous driving. The evaluation results are shown in Tab. 6", + "bbox": [ + 511, + 472, + 903, + 547 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "5. Conclusion and Future Work", + "text_level": 1, + "bbox": [ + 511, + 561, + 782, + 577 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "In this paper, we presented DriveGPT4-V2, a novel LLM-based framework for closed-loop, end-to-end autonomous driving. DriveGPT4-V2 processes camera images and vehicle states as input to generate low-level control signals for direct vehicle operation. Leveraging the extensive pretrained knowledge of MLLMs, DriveGPT4-V2 demonstrates the ability to navigate complex urban scenarios under diverse and challenging conditions. The model architecture is specifically optimized for precise numerical vehicle decision prediction. An additional expert LLM, which shares a structure similar to DriveGPT4-V2, has been trained to provide on-policy supervision. Experimental results highlight that DriveGPT4-V2 achieves state-of-the-art performance on the challenging CARLA Longest6 benchmark, outperforming all baselines. In the future, we aim to extend the capabilities of DriveGPT4-V2 for broader applications in closed-loop autonomous driving tasks, including video games and real-world deployment.", + "bbox": [ + 511, + 585, + 903, + 858 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Acknowledgment. This work is partially supported by the National Natural Science Foundation of China (No. 62422606, 62201484, 52221005).", + "bbox": [ + 511, + 858, + 903, + 900 + ], + "page_idx": 7 + }, + { + "type": "page_number", + "text": "17268", + "bbox": [ + 480, + 944, + 517, + 955 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "References", + "text_level": 1, + "bbox": [ + 91, + 90, + 187, + 104 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[1] Anthropic. Claude. https://www.anthropic.com, 2024.", + "[2] Jinze Bai, Shuai Bai, Yunfei Chu, Zeyu Cui, Kai Dang, Xiaodong Deng, Yang Fan, Wenbin Ge, Yu Han, Fei Huang, et al. Qwen technical report. arXiv:2309.16609, 2023.", + "[3] Xu Cao, Tong Zhou, Yunsheng Ma, Wenqian Ye, Can Cui, Kun Tang, Zhipeng Cao, Kaizhao Liang, Ziran Wang, James M Rehg, et al. Maplm: A real-world large-scale vision-language benchmark for map and traffic scene understanding. In CVPR, 2024.", + "[4] Dian Chen and Philipp Krahenbuhl. Learning from all vehicles. In CVPR, 2022.", + "[5] Dian Chen, Brady Zhou, Vladlen Koltun, and Philipp Krahenbuhl. Learning by cheating. In CoRL, 2020.", + "[6] Dian Chen, Vladlen Koltun, and Philipp Krahenbuhl. Learning to drive from a world on rails. In ICCV, 2021.", + "[7] Long Chen, Oleg Sinavski, Jan Hünermann, Alice Karnsund, Andrew James Willmott, Danny Birch, Daniel Maund, and Jamie Shotton. Driving with llms: Fusing object-level vector modality for explainable autonomous driving. In ICRA, 2024.", + "[8] Li Chen, Penghao Wu, Kashyap Chitta, Bernhard Jaeger, Andreas Geiger, and Hongyang Li. End-to-end autonomous driving: Challenges and frontiers. TPAMI, 2024.", + "[9] Shaoyu Chen, Bo Jiang, Hao Gao, Bencheng Liao, Qing Xu, Qian Zhang, Chang Huang, Wenyu Liu, and Xinggang Wang. Vadv2: End-to-end vectorized autonomous driving via probabilistic planning. arXiv:2402.13243, 2024.", + "[10] Yuan Chen, Zi-han Ding, Ziqin Wang, Yan Wang, Lijun Zhang, and Si Liu. Asynchronous large language model enhanced planner for autonomous driving. In ECCV, 2024.", + "[11] Wei-Lin Chiang, Zhuohan Li, Zi Lin, Ying Sheng, Zhang-hao Wu, Hao Zhang, Lianmin Zheng, Siyuan Zhuang, Yong-hao Zhuang, Joseph E Gonzalez, et al. Vicuna: An opensource chatbot impressing gpt-4 with $90\\%$ * chatgpt quality. See https://vicuna.lmsys.org, 2023.", + "[12] Kashyap Chitta, Aditya Prakash, Bernhard Jaeger, Zehao Yu, Katrin Renz, and Andreas Geiger. Transfuser: Imitation with transformer-based sensor fusion for autonomous driving. TPAMI, 2022.", + "[13] Google DeepMind. Gemini ai. https://deepmind.com, 2024.", + "[14] Alexey Dosovitskiy, German Ros, Felipe Codevilla, Antonio Lopez, and Vladlen Koltun. Carla: An open urban driving simulator. In CoRL, 2017.", + "[15] Yuchao Feng, Zhen Feng, Wei Hua, and Yuxiang Sun. Multimodal-xad: Explainable autonomous driving based on multimodal environment descriptions. TITS, 2024.", + "[16] Jiawei Fu, Yanqing Shen, Zhiqiang Jian, Shitao Chen, Jing-min Xin, and Nanning Zheng. Interactionnet: Joint planning and prediction for autonomous driving with transformers. In IROS, 2023.", + "[17] Yihan Hu, Jiazhi Yang, Li Chen, Keyu Li, Chonghao Sima, Xizhou Zhu, Siqi Chai, Senyao Du, Tianwei Lin, Wenhai Wang, et al. Planning-oriented autonomous driving. In CVPR, 2023." + ], + "bbox": [ + 93, + 114, + 480, + 898 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[18] Bin Huang, Xin Wang, Hong Chen, Zihan Song, and Wenwu Zhu. Vtimellm: Empower llm to grasp video moments. In CVPR, 2024.", + "[19] Ahmed Hussein, Mohamed Medhat Gaber, Eyad Elyan, and Chrisina Jayne. Imitation learning: A survey of learning methods. ACM Computing Surveys, 2017.", + "[20] Yuichi Inoue, Yuki Yada, Kotaro Tanahashi, and Yu Yamaguchi. Nuscenes-mqa: Integrated evaluation of captions and qa for autonomous driving datasets using markup annotations. In WACV, 2024.", + "[21] Bernhard Jaeger, Kashyap Chitta, and Andreas Geiger. Hidden biases of end-to-end driving models. In ICCV, 2023.", + "[22] Xiaosong Jia, Penghao Wu, Li Chen, Jiangwei Xie, Conghui He, Junchi Yan, and Hongyang Li. Think twice before driving: Towards scalable decoders for end-to-end autonomous driving. In CVPR, 2023.", + "[23] Bo Jiang, Shaoyu Chen, Qing Xu, Benchcheng Liao, Jiajie Chen, Helong Zhou, Qian Zhang, Wenyu Liu, Chang Huang, and Xinggang Wang. Vad: Vectorized scene representation for efficient autonomous driving. In ICCV, 2023.", + "[24] Zhuoling Li, Liangliang Ren, Jinrong Yang, Yong Zhao, Xiaoyang Wu, Zhenhua Xu, Xiang Bai, and Hengshuang Zhao. Virt: Vision instructed transformer for robotic manipulation. arXiv:2410.07169, 2024.", + "[25] Zhuoling Li, Xiaogang Xu, Zhenhua Xu, SerNam Lim, and Hengshuang Zhao. Larm: Large auto-regressive model for long-horizon embodied intelligence. arXiv:2405.17424, 2024.", + "[26] Bin Lin, Yang Ye, Bin Zhu, Jiaxi Cui, Munan Ning, Peng Jin, and Li Yuan. Video-llava: Learning united visual representation by alignment before projection. arXiv:2311.10122, 2023.", + "[27] Yuanfei Lin, Chenran Li, Mingyu Ding, Masayoshi Tomizuka, Wei Zhan, and Matthias Althoff. Drplanner: Diagnosis and repair of motion planners for automated vehicles using large language models. RA-L, 2024.", + "[28] Haotian Liu, Chunyuan Li, Yuheng Li, and Yong Jae Lee. Improved baselines with visual instruction tuning. In arXiv:2310.03744, 2023.", + "[29] Haotian Liu, Chunyuan Li, Qingyang Wu, and Yong Jae Lee. Visual instruction tuning. In NeurIPS, 2023.", + "[30] Haotian Liu, Chunyuan Li, Yuheng Li, Bo Li, Yuanhan Zhang, Sheng Shen, and Yong Jae Lee. Llava-last: Improved reasoning,OCR, and world knowledge. https://llava-v1.github.io/blog/2024-01-30-llava-last, 2024.", + "[31] Tianyu Liu, Qing hai Liao, Lu Gan, Fulong Ma, Jie Cheng, Xupeng Xie, Zhe Wang, Yingbing Chen, Yilong Zhu, Shuyang Zhang, et al. The role of the hercules autonomous vehicle during the COVID-19 pandemic: An autonomous logistic vehicle for contactless goods transportation. IEEE Robotics & Automation Magazine, 2021.", + "[32] Jiageng Mao, Yuxi Qian, Junjie Ye, Hang Zhao, and Yue Wang. Gpt-driver: Learning to drive with gpt. In NeurIPS Workshop, 2023.", + "[33] Maxim Mnyakin. Challenges and opportunities of integrating autonomous vehicles into urban retail delivery services. Reviews of Contemporary Business Analytics, 2023." + ], + "bbox": [ + 516, + 92, + 903, + 900 + ], + "page_idx": 8 + }, + { + "type": "page_number", + "text": "17269", + "bbox": [ + 480, + 944, + 517, + 955 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[34] OpenAI. Chatgpt. https://chat.openai.com, 2024.", + "[35] Takayuki Osa, Joni Pajarinen, Gerhard Neumann, J Andrew Bagnell, Pieter Abbeel, Jan Peters, et al. An algorithmic perspective on imitation learning. Foundations and Trends® in Robotics, 2018.", + "[36] SungYeon Park, MinJae Lee, JiHyuk Kang, Hahyeon Choi, Yoonah Park, Juhwan Cho, Adam Lee, and DongKyu Kim. Vlaad: Vision and language assistant for autonomous driving. In WACV, 2024.", + "[37] Alec Radford, Jong Wook Kim, Chris Hallacy, Aditya Ramesh, Gabriel Goh, Sandhini Agarwal, Girish Sastry, Amanda Askell, Pamela Mishkin, Jack Clark, et al. Learning transferable visual models from natural language supervision. In ICML, 2021.", + "[38] Katrin Renz, Kashyap Chitta, Otniel-Bogdan Mercea, A Koepke, Zeynep Akata, and Andreas Geiger. Plant: Explainable planning transformers via object-level representations. In CoRL, 2022.", + "[39] Katrin Renz, Long Chen, Ana-Maria Marcu, Jan Hunermann, Benoit Hanotte, Alice Karnsund, Jamie Shotton, Elahe Arani, and Oleg Sinavski. Carllava: Vision language models for camera-only closed-loop driving. arXiv:2406.10165, 2024.", + "[40] Stephane Ross and J. Andrew Bagnell. Reinforcement and imitation learning via interactive no-regret learning. In arXiv:1406.5979, 2014.", + "[41] Stéphane Ross, Geoffrey Gordon, and Drew Bagnell. A reduction of imitation learning and structured prediction to no-regret online learning. In AISTATS, 2011.", + "[42] Hao Shao, Letian Wang, Ruobing Chen, Hongsheng Li, and Yu Liu. Safety-enhanced autonomous driving using interpretable sensor fusion transformer. In CoRL, 2023.", + "[43] Hao Shao, Yuxuan Hu, Letian Wang, Guanglu Song, Steven L Waslander, Yu Liu, and Hongsheng Li. Lmdrive: Closed-loop end-to-end driving with large language models. In CVPR, 2024.", + "[44] Chonghao Sima, Katrin Renz, Kashyap Chitta, Li Chen, Hanxue Zhang, Chengen Xie, Jens Beiwenger, Ping Luo, Andreas Geiger, and Hongyang Li. Drivelm: Driving with graph visual question answering. In ECCV, 2024.", + "[45] Robert Soler, Alae Moudni, Gabriel Roskowski, Xinrui Yu, Mikhail Gormov, and Jafar Saniee. Autonomous patrol and threat detection through integrated mapping and computer vision. In IEEE International Conference on Electro Information Technology, 2024.", + "[46] Chan Hee Song, Jiaman Wu, Clayton Washington, Brian M Sadler, Wei-Lun Chao, and Yu Su. Llm-planner: Few-shot grounded planning for embodied agents with large language models. In ICCV, 2023.", + "[47] Yafu Tian, Alexander Carballo, Ruifeng Li, Simon Thompson, and Kazuya Takeda. Rsg-search plus: An advanced traffic scene retrieval methods based on road scene graph. In IV, 2024.", + "[48] Faraz Torabi, Garrett Warnell, and Peter Stone. Behavioral cloning from observation. In *IJCAI*, 2018.", + "[49] Hugo Touvron, Thibaut Lavril, Gautier Izacard, Xavier Martinet, Marie-Anne Lachaux, Timothée Lacroix, Baptiste" + ], + "bbox": [ + 91, + 92, + 482, + 900 + ], + "page_idx": 9 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "Rozière, Naman Goyal, Eric Hambro, Faisal Azhar, et al. Llama: Open and efficient foundation language models. arXiv:2302.13971, 2023.", + "[50] Yuxi Wei, Zi Wang, Yifan Lu, Chenxin Xu, Changxing Liu, Hao Zhao, Siheng Chen, and Yanfeng Wang. Editable scene simulation for autonomous driving via collaborative llm-agents. In CVPR, 2024.", + "[51] Licheng Wen, Daocheng Fu, Xin Li, Xinyu Cai, Tao Ma, Pinlong Cai, Min Dou, Botian Shi, Liang He, and Yu Qiao. Dilu: A knowledge-driven approach to autonomous driving with large language models. In ICLR, 2024.", + "[52] Zhenhua Xu, Yuxuan Liu, Lu Gan, Yuxiang Sun, Ming Liu, and Lujia Wang. Ringdet: Road network graph detection by transformer in aerial images. TGRS, 2022.", + "[53] Zhenhua Xu, Yuxuan Liu, Yuxiang Sun, Ming Liu, and Lujia Wang. Rngdet++: Road network graph detection by transformer with instance segmentation and multi-scale features enhancement. RA-L, 2023.", + "[54] Zhenhua Xu, Yujia Zhang, Enze Xie, Zhen Zhao, Yong Guo, Kwan-Yee K Wong, Zhenguo Li, and Hengshuang Zhao. Drivegpt4: Interpretable end-to-end autonomous driving via large language model. RA-L, 2024.", + "[55] Yijun Yang, Tianyi Zhou, Kanxue Li, Dapeng Tao, Lusong Li, Li Shen, Xiaodong He, Jing Jiang, and Yuhui Shi. Embodied multi-modal agent trained by an llm from a parallel textworld. In CVPR, 2024.", + "[56] Ekim Yurtsever, Jacob Lambert, Alexander Carballo, and Kazuya Takeda. A survey of autonomous driving: Common practices and emerging technologies. IEEE Access.", + "[57] Xiaohua Zhai, Basil Mustafa, Alexander Kolesnikov, and Lucas Beyer. Sigmoid loss for language image pre-training. In ICCV, 2023.", + "[58] Jimuyang Zhang, Zanming Huang, and Eshed Ohn-Bar. Coaching a teachable student. In CVPR, 2023.", + "[59] Jimuyang Zhang, Zanming Huang, Arjit Ray, and Eshed Ohn-Bar. Feedback-guided autonomous driving. In CVPR, 2024.", + "[60] Jiawei Zhang, Chejian Xu, and Bo Li. Chatscene: Knowledge-enabled safety-critical scenario generation for autonomous vehicles. In CVPR, 2024.", + "[61] Peiyuan Zhang, Guangtao Zeng, Tianduo Wang, and Wei Lu. Tinyllama: An open-source small language model. arXiv:2401.02385, 2024.", + "[62] Xiang Zhang, Haojie Sun, Xiaoyang Pei, Linghui Guan, and Zihao Wang. Evolution of technology investment and development of robotaxi services. Transportation Research Part E: Logistics and Transportation Review, 2024.", + "[63] Zhejun Zhang, Alexander Liniger, Dengxin Dai, Fisher Yu, and Luc Van Gool. End-to-end urban driving by imitating a reinforcement learning coach. In ICCV, 2021.", + "[64] Baichuan Zhou, Ying Hu, Xi Weng, Junlong Jia, Jie Luo, Xien Liu, Ji Wu, and Lei Huang. Tinyllava: A framework of small-scale large multimodal models. arXiv:2402.14289, 2024.", + "[65] Deyao Zhu, Jun Chen, Xiaoqian Shen, Xiang Li, and Mohamed Elhoseiny. Minigpt-4: Enhancing vision-language understanding with advanced large language models. In ICLR, 2024." + ], + "bbox": [ + 516, + 92, + 903, + 898 + ], + "page_idx": 9 + }, + { + "type": "page_number", + "text": "17270", + "bbox": [ + 480, + 944, + 519, + 955 + ], + "page_idx": 9 + } +] \ No newline at end of file diff --git a/2025/DriveGPT4-V2_ Harnessing Large Language Model Capabilities for Enhanced Closed-Loop Autonomous Driving/31ff1a5c-7171-4c25-a1d0-ee2bacb8057b_model.json b/2025/DriveGPT4-V2_ Harnessing Large Language Model Capabilities for Enhanced Closed-Loop Autonomous Driving/31ff1a5c-7171-4c25-a1d0-ee2bacb8057b_model.json new file mode 100644 index 0000000000000000000000000000000000000000..f7b782503a35180fe5112482b9ec50af00a3af19 --- /dev/null +++ b/2025/DriveGPT4-V2_ Harnessing Large Language Model Capabilities for Enhanced Closed-Loop Autonomous Driving/31ff1a5c-7171-4c25-a1d0-ee2bacb8057b_model.json @@ -0,0 +1,2299 @@ +[ + [ + { + "type": "header", + "bbox": [ + 0.107, + 0.003, + 0.182, + 0.043 + ], + "angle": 0, + "content": "CVF" + }, + { + "type": "header", + "bbox": [ + 0.237, + 0.001, + 0.812, + 0.047 + ], + "angle": 0, + "content": "This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore." + }, + { + "type": "title", + "bbox": [ + 0.156, + 0.13, + 0.844, + 0.177 + ], + "angle": 0, + "content": "DriveGPT4-V2: Harnessing Large Language Model Capabilities for Enhanced Closed-Loop Autonomous Driving" + }, + { + "type": "text", + "bbox": [ + 0.218, + 0.203, + 0.78, + 0.258 + ], + "angle": 0, + "content": "Zhenhua Xu\\(^{1,2}\\) Yan Bai\\(^{3}\\) Yujia Zhang\\(^{1}\\) Zhuoling Li\\(^{1}\\) Fei Xia\\(^{3}\\) Kwan-Yee K. Wong\\(^{1}\\) Jianqiang Wang\\(^{2}\\) Hengshuang Zhao\\(^{1*}\\) The University of Hong Kong Tsinghua University Meituan" + }, + { + "type": "title", + "bbox": [ + 0.248, + 0.292, + 0.327, + 0.308 + ], + "angle": 0, + "content": "Abstract" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.324, + 0.485, + 0.673 + ], + "angle": 0, + "content": "Multimodal large language models (MLLMs) possess the ability to comprehend visual images or videos, and show impressive reasoning ability thanks to the vast amounts of pretrained knowledge, making them highly suitable for autonomous driving applications. Unlike the previous work, DriveGPT4-V1, which focused on open-loop tasks, this study explores the capabilities of LLMs in enhancing closed-loop autonomous driving. DriveGPT4-V2 processes camera images and vehicle states as input to generate low-level control signals for end-to-end vehicle operation. A multi-view visual tokenizer (MV-VT) is employed enabling DriveGPT4-V2 to perceive the environment with an extensive range while maintaining critical details. The model architecture has been refined to improve decision prediction and inference speed. To further enhance the performance, an additional expert LLM is trained for online imitation learning. The expert LLM, sharing a similar structure with DriveGPT4-V2, can access privileged information about surrounding objects for more robust and reliable predictions. Experimental results show that DriveGPT4-V2 outperforms all baselines on the challenging CARLA Longest6 benchmark. The code and data of DriveGPT4-V2 will be publicly available." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.701, + 0.222, + 0.716 + ], + "angle": 0, + "content": "1. Introduction" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.726, + 0.484, + 0.879 + ], + "angle": 0, + "content": "Autonomous driving has seen rapid advancements over the past decade in both academia and industry [31, 56]. Today, autonomous vehicles are deployed commercially in diverse applications such as robotaxi services [62], goods delivery [33], and security patrols [45]. Additionally, assisted driving systems are increasingly integrated into consumer vehicles, enhancing safety and convenience. However, the majority of current autonomous driving systems still rely heavily on rule-based decision-making modules, where planning and control are governed by predefined algorithms and" + }, + { + "type": "image", + "bbox": [ + 0.517, + 0.291, + 0.906, + 0.507 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.512, + 0.517, + 0.907, + 0.617 + ], + "angle": 0, + "content": "Figure 1. DriveGPT4-V2 for closed-loop autonomous driving. Taken as input multi-view camera images and vehicle state information, DriveGPT4-V2 predicts high-level vehicle decisions and converts them to low-level vehicle control signals in an end-to-end manner. DriveGPT4-V2 presents outstanding effectiveness and efficiency, serving as a reliable baseline method for future research on autonomous driving with LLMs." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.643, + 0.907, + 0.688 + ], + "angle": 0, + "content": "rules. This reliance restricts the ability of these systems to operate effectively in varied, real-world scenarios, limiting the broader advancement of autonomous driving." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.689, + 0.909, + 0.84 + ], + "angle": 0, + "content": "In response, recent research has shifted toward end-to-end autonomous driving systems based on learning-based methods [8, 9, 12, 17, 21, 23, 45]. These methods enable more streamlined and intelligent system designs by directly learning control from input sensor data, without the need for complicated intermediate modules. From a learning theory perspective, end-to-end autonomous driving can optimize the entire system on the final outputs, rather than through the isolated optimization of individual modules, which potentially improves overall performance." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.84, + 0.909, + 0.903 + ], + "angle": 0, + "content": "The rise of large language models (LLMs) has demonstrated their impressive ability to perform natural language processing (NLP) tasks [1, 2, 13, 34, 49]. Commercial LLMs, such as ChatGPT [34] and Claude3 [1], are already" + }, + { + "type": "page_footnote", + "bbox": [ + 0.11, + 0.888, + 0.236, + 0.9 + ], + "angle": 0, + "content": "*Corresponding author." + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.518, + 0.957 + ], + "angle": 0, + "content": "17261" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.09, + 0.091, + 0.482, + 0.348 + ], + "angle": 0, + "content": "widely used in daily life. Motivated by their success in NLP tasks, LLMs are being extended to multimodal domains, enabling more generalized tasks in fields like image analysis [28-30], video comprehension [18, 26], and embodied AI [24, 25, 46, 55]. Given their versatility, multimodal LLMs have been applied to autonomous driving for tasks such as interpretability [15, 36, 44, 54], vehicle control [10, 27, 27], and risk assessment [3, 60]. While multimodal LLMs have demonstrated promising vehicle control performance due to their extensive pretrained knowledge, previous research has primarily concentrated on open-loop settings, which are not implemented for real-time vehicle control. Open-loop systems lack feedback mechanisms, making them less adaptable to changes and disturbances, limiting their applicability in real-world applications. Thus, multimodal LLMs need to be designed and evaluated specifically for closed-loop autonomous driving scenarios [27, 39, 43, 51]." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.351, + 0.482, + 0.457 + ], + "angle": 0, + "content": "Behavior cloning [35, 48], the most commonly employed imitation learning algorithm, is used for most end-to-end autonomous driving works [44, 54]. However, models trained only by behavior cloning can encounter drifting issues in closed-loop tasks, where errors accumulate, potentially leading the vehicle into unsafe situations, such as deviating from the path, colliding, or getting stuck." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.46, + 0.483, + 0.732 + ], + "angle": 0, + "content": "To address the aforementioned issue, we incorporate online (or on-policy) imitation learning [5, 40, 41] into closed-loop autonomous driving. Leveraging the reasoning capabilities and pretrained knowledge of LLMs, we propose using a multimodal LLM as a fully automatic expert for online imitation learning without human efforts. Let \\( M_E \\) and \\( M_A \\) denote the model of expert and agent, respectively. \\( M_E \\) can access privileged information about nearby objects while \\( M_A \\) cannot, thus presenting a more powerful and reliable performance. After being trained by behavior cloning, \\( M_A \\) is deployed on the training environment for data aggregation. \\( M_A \\) might make incorrect predictions that lead the vehicle into exception states. \\( M_E \\) can provide \\( M_A \\) with on-policy supervision and take over the vehicle if the error made by \\( M_A \\) is large enough. When \\( M_E \\) takes over the vehicle, it collects one data sample at the moment and aggregates it into the dataset. The aggregated dataset is leveraged to further train the agent model to handle errors." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.735, + 0.483, + 0.901 + ], + "angle": 0, + "content": "In this paper, we introduce DriveGPT4-V2, an end-to-end autonomous driving system. DriveGPT4-V2 extends DriveGPT4-V1 [54] to closed-loop tasks and integrates online imitation learning for enhanced robustness and intelligence. The \"4\" in the system's name signifies its multimodal capacity, inspired by the MiniGPT4 model [65]. DriveGPT4-V2 is powered by multimodal LLMs, enabling it to directly generate low-level vehicle control signals (i.e., throttle, brake and steer) based on multimodal input data (i.e., vehicle states and camera images). The structure of DriveGPT4-V2 is refined for our task, including a" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.092, + 0.905, + 0.288 + ], + "angle": 0, + "content": "multi-view visual tokenizer (MV-VT) that possesses a large enough perception range without losing details, as well as output decision heads (DeciHeads) specially designed for numerical vehicle decision prediction. An additional multimodal LLM with access to privileged information (e.g., ground truth of nearby objects in text) serves as an expert for online imitation learning, which is inspired by the LBC framework [5]. DriveGPT4-V2 achieves state-of-the-art performance on the challenging CARLA Longest6 benchmark [12], which evaluates systems over 36 long routes demanding high levels of intelligence and reliability. The overview of DriveGPT4-V2 is visualized in Fig. 1. The contributions of this paper are:" + }, + { + "type": "text", + "bbox": [ + 0.514, + 0.289, + 0.905, + 0.363 + ], + "angle": 0, + "content": "- We extend DriveGPT4-V1, an open-loop autonomous driving system with multimodal LLMs, and propose DriveGPT4-V2 for end-to-end closed-loop autonomous driving. DriveGPT4-V2 can better harness the capability of LLMs and be deployed for continuous vehicle control." + }, + { + "type": "text", + "bbox": [ + 0.514, + 0.364, + 0.905, + 0.438 + ], + "angle": 0, + "content": "- We introduce the use of LLMs with privileged information as an expert for online (on-policy) imitation learning, facilitating the generation of labeled data for handling errors, which significantly enhances the performance and robustness of closed-loop autonomous driving." + }, + { + "type": "text", + "bbox": [ + 0.514, + 0.44, + 0.905, + 0.514 + ], + "angle": 0, + "content": "- We optimize the model structure of MLLMs, including a multi-view visual tokenizer (MV-VT) that possesses a large perception range and preserves details, as well as output decision heads (DeciHeads) specially designed for numerical vehicle decision prediction." + }, + { + "type": "text", + "bbox": [ + 0.514, + 0.515, + 0.905, + 0.56 + ], + "angle": 0, + "content": "- DriveGPT4-V2 demonstrates superior performance on the CARLA Longest6 benchmark, substantially outperforming existing methods." + }, + { + "type": "list", + "bbox": [ + 0.514, + 0.289, + 0.905, + 0.56 + ], + "angle": 0, + "content": null + }, + { + "type": "title", + "bbox": [ + 0.514, + 0.573, + 0.663, + 0.589 + ], + "angle": 0, + "content": "2. Related Works" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.599, + 0.907, + 0.871 + ], + "angle": 0, + "content": "Large language models for autonomous driving. Owing to the powerful reasoning capabilities and extensive pretrained knowledge, large language models (LLMs) have found wide application in autonomous driving tasks, such as vehicle decision-making [7, 10, 15, 27, 32, 39, 43, 54, 59], scene understanding [20, 36, 44, 54], risk analysis [3, 60], and simulation [47, 50]. GPT-Driver [32] represents the entire surrounding environment textually and employs a fine-tuned ChatGPT model to predict vehicle trajectories. DriveGPT4 [54] leverages multimodal LLMs for end-to-end vehicle action prediction and scene understanding. It processes 8-frame video clips and human user commands to predict vehicle control signals and action explanations. However, most of these approaches are designed for open-loop scenarios and present unsatisfactory performance in closed-loop tasks, limiting their applicability to real-world driving deployment. Additionally, many existing methods face efficiency challenges due to heavy model architectures." + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.871, + 0.905, + 0.901 + ], + "angle": 0, + "content": "Closed-loop end-to-end autonomous driving. Closed-loop end-to-end autonomous driving has recently emerged" + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "17262" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.099, + 0.093, + 0.9, + 0.436 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.089, + 0.452, + 0.907, + 0.551 + ], + "angle": 0, + "content": "Figure 2. DriveGPT4-V2 diagram. DriveGPT4-V2 takes multimodal input data to generate numerical control signals for end-to-end vehicle driving. The input includes multi-view images and vehicle state information. The images are transformed into the text domain via a multi-view vision tokenizer (MV-VT). The current speed of the vehicle and target point are tokenized by the LLM tokenizer. The LLM then outputs four tokens. Each token is used to predict one vehicle decision by an MLP decision head (DeciHead). These predicted decisions are subsequently converted to low-level commands via PID controllers to operate the vehicle. The LLM expert model, which shares a similar structure to DriveGPT4-V2, has access to privileged information about surroundings (shown in the purple module). The expert provides on-policy supervision to DriveGPT4-V2 to enhance closed-loop performance. This figure is best viewed in color." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.576, + 0.486, + 0.895 + ], + "angle": 0, + "content": "as a prominent research focus, addressing the challenge of bridging the gap between academic research and real-world applications. Various approaches improved transformer architectures and training methods to enhance driving performance [4, 5, 12, 16, 21, 22, 42, 58, 63]. For instance, Transfuser++ [21] integrates multimodal sensor inputs, including camera imagery, LiDAR point clouds and vehicle dynamics, to predict speed and trajectory, achieving state-of-the-art performance across multiple benchmarks. Even though most early LLM-based methods have primarily targeted open-loop tasks, recent studies [27, 39, 43, 51] have attempted closed-loop approaches. But most of them are deployed in simple environments like Highway-env [51], instead of real-world scenarios. LMDrive [43], pioneering work on end-to-end closed-loop driving with multimodal LLMs, demonstrated the potential of LLMs in this domain. It is deployed in the CARLA simulator [14] and controls the vehicle like human drivers. However, it relies solely on behavior cloning and presents unsatisfactory driving performance. Online imitation learning has been highlighted as a potential way to enhance model robustness." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.576, + 0.909, + 0.895 + ], + "angle": 0, + "content": "Imitation learning in autonomous driving. Imitation learning [19] aims to train an agent to mimic the expert behavior. Behavior cloning is the most widely used imitation learning algorithm in autonomous driving [4, 12, 21, 39, 43, 44, 52-54]. In this method, data is off-policy sampled from expert trajectories for model training to replicate expert actions. Although behavior cloning can yield reasonable performance, it suffers from error accumulation during closed-loop operation, leading the model to encounter unseen situations that may compromise safety. To mitigate this issue, online imitation learning [40, 41] has been employed in various works [4, 5]. The LBC framework [5] uses the DAgger algorithm [41], and trains an agent alongside an expert network with privileged information that achieves superior performance. The expert provides on-policy supervision to the agent and collects additional data samples for data aggregation, which can be used to further train the agent model to handle errors. Inspired by LBC, DriveGPT4-V2 harnesses the capabilities of LLMs and leverages a privileged LLM expert to provide on-policy supervision, facilitating improved driving performance and robustness." + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.52, + 0.958 + ], + "angle": 0, + "content": "17263" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.1, + 0.092, + 0.475, + 0.307 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.09, + 0.322, + 0.484, + 0.392 + ], + "angle": 0, + "content": "Figure 3. Multi-view visual tokenizer (MV-VT) structure. The input images consist of three front views. Each patch is processed through a visual encoder to extract features. Finally, a trained projection layer maps the downsampled feature into the text domain for further processing." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.418, + 0.245, + 0.433 + ], + "angle": 0, + "content": "3. DriveGPT4-V2" + }, + { + "type": "text", + "bbox": [ + 0.089, + 0.444, + 0.483, + 0.686 + ], + "angle": 0, + "content": "Building on DriveGPT4-V1 [54], a flexible multimodal LLM designed for open-loop autonomous driving, DriveGPT4-V2 has been specifically developed for closed-loop driving tasks. This paper employs LLMs both as planners and as expert models for online imitation learning. A multi-view visual tokenizer (MV-VT) is leveraged for effective feature extraction. Unlike standard LLMs used for NLP, the architecture of DriveGPT4-V2 is tailored for vehicle planning, with additional output decision heads (DeciHeads) to directly optimize numerical decision-making predictions. This design enhances model efficiency by reducing the token count required for inference. The expert model, which shares a similar architecture to DriveGPT4-V2, has access to privileged ground-truth information about the surroundings and is used solely for on-policy supervision during training, not for inference." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.697, + 0.234, + 0.713 + ], + "angle": 0, + "content": "3.1. Model Design" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.719, + 0.483, + 0.795 + ], + "angle": 0, + "content": "DriveGPT4-V2 takes multimodal input data and predicts high-level vehicle decisions for end-to-end autonomous driving. These predictions are subsequently converted into low-level control signals through PID controllers. The model structure is shown in Fig. 2." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.796, + 0.484, + 0.902 + ], + "angle": 0, + "content": "Multimodal input. The input consists of two modalities: image and text. The image input is captured by three front-view cameras at a resolution of \\(384 \\times 384\\), while the text input includes the vehicle's current speed and the target point. These three front-view images ensure coverage of both sides of the vehicle, reducing the risk of missing adjacent vehicles or potential collisions. The image data is" + }, + { + "type": "image", + "bbox": [ + 0.515, + 0.096, + 0.907, + 0.106 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.571, + 0.114, + 0.65, + 0.127 + ], + "angle": 0, + "content": "(a) Waypoints." + }, + { + "type": "image_caption", + "bbox": [ + 0.765, + 0.115, + 0.856, + 0.127 + ], + "angle": 0, + "content": "(b) Route points." + }, + { + "type": "image_caption", + "bbox": [ + 0.512, + 0.139, + 0.907, + 0.237 + ], + "angle": 0, + "content": "Figure 4. Visualization of waypoints and route points. The ego vehicle is represented by the green rectangle, and the red point denotes the target point. The grey line is the route that the vehicle should follow. (a) Waypoints (blue points) represent vehicle positions in a 4-second future. (b) Route points are evenly sampled from the global planned route in front. These two kinds of points can better supervise the training of DriveGPT4-V2." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.265, + 0.907, + 0.308 + ], + "angle": 0, + "content": "mapped into the text domain via a visual tokenizer, as implemented in LLaVA [30], while text inputs are tokenized with standard LLM tokenizers." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.31, + 0.906, + 0.505 + ], + "angle": 0, + "content": "Multi-view visual tokenizer (MV-VT). The visual tokenizer is visualized in Fig. 3. It first employs pretrained visual encoders, such as CLIP [37] or SigLIP [57], to extract features from the input images. The size of each image is specifically optimized for compatibility with the SigLIP encoder (i.e., \\(384 \\times 384\\)). The three images are acquired from front-left, front, and front-right cameras respectively, ensuring comprehensive coverage of the vehicle's frontal and lateral views. In CARLA benchmarks, back-view images exhibit limited utilization frequency and demonstrate no measurable improvement in planning performance. To maintain computational efficiency, MV-VT deliberately excludes back-view inputs." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.507, + 0.906, + 0.627 + ], + "angle": 0, + "content": "Output representations. DriveGPT4-V2 generates three low-level control signals to drive the vehicle, i.e., throttle, brake, and steering. However, direct prediction of these signals requires temporal consistency; otherwise, the vehicle may display unstable zigzag trajectories. To ensure stable control, DriveGPT4-V2 instead predicts high-level decision variables, which are then converted to low-level commands by PID controllers. The predictions of the model include:" + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.628, + 0.905, + 0.673 + ], + "angle": 0, + "content": "- Target Speed: The desired speed for the vehicle. A speed PID controller adjusts longitudinal control based on the predicted target speed and current speed." + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.674, + 0.905, + 0.703 + ], + "angle": 0, + "content": "- Target Angle: The direction in which the vehicle should steer. A steering PID controller uses it for lateral control." + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.704, + 0.905, + 0.734 + ], + "angle": 0, + "content": "- Waypoints: Predicted x-y positions of the vehicle in the future. Waypoints consist of eight points." + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.734, + 0.905, + 0.763 + ], + "angle": 0, + "content": "- Route Points: Points evenly sampled along the global route in front, enhancing the vehicle's perception ability." + }, + { + "type": "list", + "bbox": [ + 0.513, + 0.628, + 0.905, + 0.763 + ], + "angle": 0, + "content": null + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.765, + 0.855, + 0.779 + ], + "angle": 0, + "content": "Waypoints and route points are visualized in Fig. 4." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.78, + 0.907, + 0.901 + ], + "angle": 0, + "content": "Decision heads (DeciHeads). For closed-loop autonomous driving, DriveGPT4-V2 employs dedicated output heads (MLPs) to predict numerical vehicle decisions. The original LLM's vocabulary output head is replaced with four decision heads, each predicting a decision variable (i.e., target speed, target angle, waypoints, and route points) based on one token. This setup allows DriveGPT4-V2 to produce only four tokens per prediction cycle, achieving a huge" + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.52, + 0.957 + ], + "angle": 0, + "content": "17264" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.09, + 0.092, + 0.482, + 0.167 + ], + "angle": 0, + "content": "speed increase over traditional text-based outputs. For example, a float number with 3 digits (e.g., 1.538) needs 5 tokens for text output representation but only requires a single token for our design. DeciHead also enables direct optimization of decision variables using regression loss." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.168, + 0.483, + 0.348 + ], + "angle": 0, + "content": "Expert model and online imitation learning. The expert model has a similar architecture to DriveGPT4-V2 but has access to privileged information. The privileged information consists of ground truth (GT) of nearby objects and hazard information provided by the CARLA simulator (e.g., traffic light and stop sign violation; vehicle or pedestrian collision). This privileged data allows the expert to better perceive the environment, so that it can focus on vehicle decisions and planning. The privileged information greatly boosts the model capacity, resulting in enhanced closed-loop performance. The expert model provides on-policy supervision to DriveGPT4-V2 during training." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.348, + 0.483, + 0.514 + ], + "angle": 0, + "content": "When DriveGPT4-V2 runs on training scenarios and routes, the expert model receives identical input and generates concurrent predictions. If the discrepancy between DriveGPT4-V2's prediction and the expert's exceeds a threshold, the expert's prediction is used instead to control the vehicle. Such a situation is marked as an exception case. A data sample of the current moment is added to the training set, where the expert's output is used as the label. DriveGPT4-V2 is fine-tuned on this aggregated dataset under the expert's on-policy supervision for enhanced closed-loop autonomous driving performance." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.515, + 0.483, + 0.546 + ], + "angle": 0, + "content": "The expert LLM is used exclusively for on-policy supervision during training and is not accessed during inference." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.554, + 0.196, + 0.57 + ], + "angle": 0, + "content": "3.2. Training" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.576, + 0.483, + 0.682 + ], + "angle": 0, + "content": "The training process of DriveGPT4-V2 and the expert LLM includes two stages. In the first stage, both models are trained on data collected by a rule-based autopilot using behavior cloning. In the second stage, DriveGPT4-V2 runs on the training scenarios and routes under the expert's on-policy supervision, and generates additional data for error correction. The training process is visualized in Fig. 5." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.69, + 0.317, + 0.705 + ], + "angle": 0, + "content": "3.2.1. Stage 1: behavior cloning" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.709, + 0.483, + 0.844 + ], + "angle": 0, + "content": "In the first stage, a rule-based autopilot [21] is deployed in CARLA to collect data. The autopilot agent records data at fixed time intervals, including camera images, current vehicle speed, privileged information on surrounding objects (e.g., vehicles and traffic lights), and a target point. Labels for the vehicle decision variables are calculated using rule-based algorithms. After data collection, both DriveGPT4-V2 and the expert LLM are trained on the dataset, with the expert having access to privileged information." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.852, + 0.44, + 0.868 + ], + "angle": 0, + "content": "3.2.2. Stage 2: DAgger with on-policy supervision" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.871, + 0.483, + 0.902 + ], + "angle": 0, + "content": "The expert, with access to privileged information, achieves notably higher performance by leveraging reliable priors on" + }, + { + "type": "image", + "bbox": [ + 0.541, + 0.09, + 0.9, + 0.188 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.639, + 0.192, + 0.8, + 0.204 + ], + "angle": 0, + "content": "(a) Stage 1: Behavior cloning." + }, + { + "type": "image", + "bbox": [ + 0.519, + 0.206, + 0.92, + 0.34 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.596, + 0.346, + 0.842, + 0.359 + ], + "angle": 0, + "content": "(b) Stage 2: DAgger online imitation learning." + }, + { + "type": "image_caption", + "bbox": [ + 0.512, + 0.371, + 0.907, + 0.483 + ], + "angle": 0, + "content": "Figure 5. Diagram of the two-stage training process. (a) In the first stage, both DriveGPT4-V2 and the expert LLM are trained on data collected by a rule-based autopilot. (b) In the second stage, DriveGPT4-V2 runs on the training scenarios and routes. When the discrepancy between DriveGPT4-V2's predictions and those of the expert exceeds a predefined threshold, the expert's predictions are used to control the vehicle. Data from these cases is then added to the dataset for data aggregation." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.508, + 0.907, + 0.795 + ], + "angle": 0, + "content": "surrounding objects, allowing it to handle challenging scenarios effectively. DriveGPT4-V2, trained only by behavior cloning, may struggle with error accumulation and exceptional situations (e.g., vehicle collisions or getting stuck). Inspired by LBC [5], the expert LLM is used as a teacher for on-policy supervision. On-policy supervision is a training paradigm in imitation learning, where supervision or feedback is provided to an agent while it is acting according to its current policy. This approach contrasts with off-policy methods, where supervision comes from data collected by a different policy or behavior. During the operation of DriveGPT4-V2, the expert and DriveGPT4-V2 receive identical input, and their outputs are compared. If the discrepancy in predicted decisions exceeds a threshold, the expert takes control to avoid exceptions. One data sample is collected at that moment and aggregated into the dataset. The output of the expert is used as the training label. After traversing the training routes under expert supervision, DriveGPT4-V2 is fine-tuned on the aggregated dataset." + }, + { + "type": "title", + "bbox": [ + 0.514, + 0.804, + 0.61, + 0.818 + ], + "angle": 0, + "content": "3.3. Dataset" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.825, + 0.906, + 0.901 + ], + "angle": 0, + "content": "Following Transfuser++ [21], in stage 1, the rule-based autopilot agent is used to collect data samples in various CARLA simulator towns under varying weather, lighting, and daytime conditions. Vehicle speed and position are estimated by Kalman filters based on IMU data. Each sam" + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "17265" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.089, + 0.092, + 0.482, + 0.182 + ], + "angle": 0, + "content": "ple records objects within a fixed radius, including vehicles, pedestrians, traffic lights, and stop signs. Target speeds and angles are calculated by the autopilot's PID controller. The ground-truth waypoints contain 8 points and record the vehicle position in a 4-second future. Route points have 10 points sampled along the planned route in front." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.183, + 0.483, + 0.318 + ], + "angle": 0, + "content": "In stage 2, DriveGPT4-V2 runs on selected training routes for data aggregation. Simple routes are not considered in this stage for data collection efficiency. If the difference between the predicted decisions of DriveGPT4-V2 and the expert exceeds a threshold, the expert takes control and records the data sample. DriveGPT4-V2 is retrained on the aggregated dataset for better performance. Under common circumstances, DriveGPT4-V2 achieves satisfactory performance after one round of DAgger." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.325, + 0.236, + 0.34 + ], + "angle": 0, + "content": "3.4. Loss Function" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.348, + 0.483, + 0.407 + ], + "angle": 0, + "content": "During training, the L1 loss function supervises all predicted decision variables, including target speed \\((TS)\\), target angle \\((TA)\\), waypoints \\((WP)\\), and route points \\((RP)\\). The overall loss is calculated by:" + }, + { + "type": "equation", + "bbox": [ + 0.173, + 0.414, + 0.482, + 0.43 + ], + "angle": 0, + "content": "\\[\n\\mathcal {L} = \\mathcal {L} _ {T S} + \\mathcal {L} _ {T A} + \\mathcal {L} _ {W P} + \\mathcal {L} _ {R P} \\tag {1}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.435, + 0.483, + 0.481 + ], + "angle": 0, + "content": "This approach allows the model to optimize final numerical vehicle decision variables directly, instead of relying on vocabulary-based classification as in NLP tasks." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.492, + 0.222, + 0.509 + ], + "angle": 0, + "content": "4. Experiments" + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.517, + 0.313, + 0.531 + ], + "angle": 0, + "content": "4.1. Dataset and Benchmark" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.539, + 0.483, + 0.704 + ], + "angle": 0, + "content": "All experiments in this study were conducted in the CARLA simulator [14]. We followed Transfuser++ for training data preparation. In stage 1, a rule-based autopilot was employed to collect 350K behavior-cloning samples. The autopilot traversed multiple routes in Towns 1, 2, 3, 4, 5, 6, 7, and 10 of CARLA, encountering various scenarios such as different weather and lighting conditions. Data were collected at a frequency of \\(2\\mathrm{Hz}\\) and subsequently filtered to remove incorrect or redundant samples, such as those captured when the autopilot was stuck. This process resulted in a dataset of 300K samples." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.705, + 0.483, + 0.825 + ], + "angle": 0, + "content": "In stage 2, DriveGPT4-V2 trained by behavior cloning runs on selected training scenarios and routes, utilizing the expert for on-policy supervision. The expert can access privileged information about objects within a 30-meter distance. Approximately 150K samples were gathered in this stage. These aggregated data were then combined with the original dataset to further fine-tune DriveGPT4-V2, enhancing its performance." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.826, + 0.483, + 0.901 + ], + "angle": 0, + "content": "All methods were evaluated on the challenging CARLA Longest6 benchmark [12], which consists of 36 extended routes encompassing various scenarios. The length and complexity of routes make them suitable for a comprehensive evaluation of closed-loop autonomous driving." + }, + { + "type": "table_caption", + "bbox": [ + 0.606, + 0.09, + 0.813, + 0.104 + ], + "angle": 0, + "content": "Table 1. Infraction penalty factors." + }, + { + "type": "table", + "bbox": [ + 0.515, + 0.115, + 0.905, + 0.187 + ], + "angle": 0, + "content": "
ScenePenaltyScenePenalty
Pedestrian collision0.5Vehicle collision0.6
Static object collision0.65Traffic light violation0.7
Stop sign violation0.7
" + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.212, + 0.779, + 0.228 + ], + "angle": 0, + "content": "4.2. DriveGPT4-V2 Configuration" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.234, + 0.905, + 0.354 + ], + "angle": 0, + "content": "LLM. For efficient data collection and inference, we used tiny-scale LLMs in DriveGPT4-V2, such as Qwen-0.5B [2] and TinyLLaMA [61]. These models were fine-tuned for multimodal understanding, following methods outlined in LLaVA [29, 30] and TinyLLaVA [64]. Large-scale LLMs, like LLaMA3-8B [49] and vicuna-7B [11], are not considered at the moment due to their high computational demands and reduced inference speed." + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.355, + 0.905, + 0.444 + ], + "angle": 0, + "content": "Training. DriveGPT4-V2 was fully fine-tuned in an end-to-end manner, except for the SigLIP \\(384 \\times 384\\) visual encoder [57]. The model was trained with a learning rate of \\(2e^{-5}\\) over 60 epochs by employing a cosine annealing schedule for better learning rate control. The training takes about 75 hours on 16 A800 GPUs." + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.446, + 0.906, + 0.581 + ], + "angle": 0, + "content": "Inference. Two PID controllers were designed to convert DriveGPT4-V2's predicted decision variables into low-level control signals (i.e., throttle, steer, and brake). These control signals were used to directly drive the ego vehicle in the CARLA simulator. In the experiments, only the predicted target speed and target angle were used by the PID controllers, while waypoints and route points served as additional supervision for improved training results. The model runs inference on a single A800 GPU with FP16 precision." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.591, + 0.7, + 0.605 + ], + "angle": 0, + "content": "4.3. Evaluation Criteria" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.613, + 0.906, + 0.809 + ], + "angle": 0, + "content": "Following the CARLA leaderboard, the methods were evaluated mainly based on three metrics: Driving Score (DS), Route Completion (RC), and Infraction Score (IS). RC indicates the percentage of routes completed, while IS penalizes infractions such as traffic violations and collisions during inference. Each infraction multiplicatively reduces IS by a penalty factor. The penalty factors are listed in Tab. 1. DS, as the product of RC and IS, provides a comprehensive evaluation score. In the evaluation, we also report ratios of commonly seen exception events, including pedestrian collision (Ped), vehicle collision (Veh), static object collision (Stat), red light violation (Red), route deviations (Dev), scenario time out (TO) and agent block (Block)." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.819, + 0.747, + 0.835 + ], + "angle": 0, + "content": "4.4. Comparison Experiments" + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.841, + 0.905, + 0.871 + ], + "angle": 0, + "content": "Baselines. We selected several representative prior works as baselines for comparison:" + }, + { + "type": "text", + "bbox": [ + 0.514, + 0.871, + 0.905, + 0.901 + ], + "angle": 0, + "content": "- WOR [6] (ICCV 2021): WOR uses commands instead of target points for global planning instructions, providing" + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "17266" + } + ], + [ + { + "type": "table_caption", + "bbox": [ + 0.09, + 0.089, + 0.907, + 0.146 + ], + "angle": 0, + "content": "Table 2. Closed-loop experiments performance on CARLA Longest6. \"Visual\" indicates visual input modalities, while \"C\" and \"L\" represent camera and LiDAR, respectively. Bold numbers highlight SOTA metric scores of all models; while underlined numbers represent the best metric scores of baseline methods. * denotes models implemented by ourselves based on official open-sourced code. † represents the model without data augmentation." + }, + { + "type": "table", + "bbox": [ + 0.092, + 0.156, + 0.908, + 0.34 + ], + "angle": 0, + "content": "
MethodVisualDS↑RC↑IS↑Ped↓Veh↓Stat↓Red↓Dev↓TO↓Block↓
WOR [6]C21480.560.181.050.371.280.880.080.20
LAV v1 [4]C&L33700.510.160.830.150.960.060.120.45
Interfuser [42]C47740.630.061.140.110.240.000.520.06
TransFuser [12]C&L47930.500.032.450.070.160.000.060.10
LAV v2 [4]C&L58830.680.000.690.150.230.080.320.11
Perception PlanT [38]C&L58880.650.070.970.110.090.000.130.13
Transfuser++* [21]C&L65900.720.000.990.010.070.000.100.12
Transfuser++*† [21]C&L58890.650.011.150.010.100.000.140.13
LMDrive* [43]C&L36690.520.071.030.181.010.090.110.22
DriveGPT4-V2C70910.770.000.800.010.040.000.070.09
" + }, + { + "type": "text", + "bbox": [ + 0.104, + 0.364, + 0.449, + 0.379 + ], + "angle": 0, + "content": "enriched labels and supervision for possible actions." + }, + { + "type": "text", + "bbox": [ + 0.091, + 0.38, + 0.482, + 0.424 + ], + "angle": 0, + "content": "- LAV v1 & v2 [4] (CVPR 2022): LAV predicts waypoints by integrating information from nearby vehicles, which is further refined through a GRU." + }, + { + "type": "text", + "bbox": [ + 0.091, + 0.425, + 0.482, + 0.469 + ], + "angle": 0, + "content": "- Transfuser [12] (TPAMI 2022): Transfuser is a widely used baseline model for closed-loop driving that has inspired subsequent works." + }, + { + "type": "text", + "bbox": [ + 0.091, + 0.47, + 0.482, + 0.514 + ], + "angle": 0, + "content": "- Perception PlanT [38] (CoRL 2022): PlanT is based on imitation learning with object-level input representation. Such representation is elegant and effective." + }, + { + "type": "text", + "bbox": [ + 0.091, + 0.515, + 0.482, + 0.56 + ], + "angle": 0, + "content": "- InterFuser [42] (CoRL 2023): InterFuser predicts additional density maps and traffic rule flags to enhance waypoint prediction." + }, + { + "type": "text", + "bbox": [ + 0.091, + 0.56, + 0.482, + 0.65 + ], + "angle": 0, + "content": "- Transfuser++ [21] (ICCV 2023): Transfuser++ predicts target speeds and routes instead of waypoints, integrating point cloud and camera images. It is currently the state-of-the-art (SOTA) method for CARLA Longest6. Transfuser++ utilizes triple training data by repeatedly collecting expert trajectories for augmentation." + }, + { + "type": "text", + "bbox": [ + 0.091, + 0.651, + 0.482, + 0.696 + ], + "angle": 0, + "content": "- LMDrive [43] (CVPR 2024): LMDrive is a pioneering work for closed-loop driving with multimodal LLMs, training a vicuna-7B LLM for end-to-end vehicle control." + }, + { + "type": "list", + "bbox": [ + 0.091, + 0.38, + 0.482, + 0.696 + ], + "angle": 0, + "content": null + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.705, + 0.483, + 0.901 + ], + "angle": 0, + "content": "Comparison results. The comparison results on the CARLA Longest6 benchmark are shown in Tab. 2. It is observed that DriveGPT4-V2 outperforms all baseline models by a large margin. Thanks to the huge amount pretrained knowledge of LLMs and visual encoders, DriveGPT4-V2 can better handle complex urban scenes under various conditions, such as nighttime, rainy weather, etc. For most baselines, we list the results reported on the CARLA Longest6 leaderboard. However, some baselines cannot achieve reported performance by our implementation and test, thus we report our implemented results separately. LMDrive [43] is the only open-sourced LLM method that can be tested in a closed-loop environment. But it fails" + }, + { + "type": "table_caption", + "bbox": [ + 0.626, + 0.362, + 0.794, + 0.377 + ], + "angle": 0, + "content": "Table 3. Efficiency analysis." + }, + { + "type": "table", + "bbox": [ + 0.515, + 0.387, + 0.909, + 0.461 + ], + "angle": 0, + "content": "
LLMDSTrainFPS
LLaVA-LLaMA3.1-8B6511.2h/epoch0.4
TinyLLaVA-LLaMA-1.5B633.0h/epoch2.9
LLaVA-Qwen-0.5B631.3h/epoch8.1
" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.487, + 0.907, + 0.669 + ], + "angle": 0, + "content": "to obtain satisfactory performance. It first converts multimodal data to BEV and projects BEV to the text domain. However, BEV does not have good-quality pretrained multimodal LLMs. Training from scratch severely degrades its final performance, which is also discussed by DriveGPT4-V1 [54]. Transfuser++ [21] is the SOTA method by fusing point clouds and camera images for closed-loop control. Transfuser++ uses triple training data and achieves 65 DS, but only has 58 DS with the same amount of training data (denoted by \\(\\dagger\\)) as our method. Thanks to the pretrained knowledge and reasoning ability of LLMs, DriveGPT4-V2 can outperform all baselines with camera images as input." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.682, + 0.681, + 0.698 + ], + "angle": 0, + "content": "4.5. Model Efficiency" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.705, + 0.907, + 0.901 + ], + "angle": 0, + "content": "In our experiments, DriveGPT4-V2 and the expert are implemented by tiny LLMs. LLMs with 7B or even more parameters show powerful reasoning and generalization ability, but they severely enlarge the time consumption and demand much more training resources, which is unacceptable in real-world applications. In this section, we report efficiency results of DriveGPT4-V2 with LLaVA-0.5B, TinyLLaVA-1.5B and LLaVA-7B LLMs. The results are shown in Tab. 3. It is found that merely scaling up the LLM model does not gain corresponding performance enhancement, but slows down the whole system more than 10 times. Therefore, DriveGPT4-V2 mainly considers 0.5B LLMs as the planner." + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "17267" + } + ], + [ + { + "type": "title", + "bbox": [ + 0.091, + 0.091, + 0.373, + 0.106 + ], + "angle": 0, + "content": "4.6. Ablation Studies and Discussion" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.113, + 0.482, + 0.174 + ], + "angle": 0, + "content": "We conducted multiple ablation studies to justify the design of DriveGPT4-V2. The results are presented in Tab. 4, 5 and 6. Due to huge time consumption, DriveGPT4-V2 in ablation studies is not trained with DAgger data by default." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.174, + 0.482, + 0.28 + ], + "angle": 0, + "content": "LLM visual pretraining. DriveGPT4-V2 relies on LLaVA [30, 64] as the planner, which is pretrained on a huge amount of visual understanding data. The visual pretraining benefits DriveGPT4-V2 on the autonomous driving task with multimodal input. Directly training DriveGPT4-V2 on scratch (i.e., LLM for NLP tasks without multimodal ability) slows down convergence and degrades the final results." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.28, + 0.482, + 0.432 + ], + "angle": 0, + "content": "Visual tokenizer. The input camera images are processed by the proposed visual tokenizer. Inspired by LLaVA 1.5 [30], the features of images are extracted separately by the visual encoder and then concatenated as the input. This design ensures a sufficient perception range without compromising image detail. If the input is a single image with a large scope, it might lead to information loss during preprocessing steps (e.g., resize or crop); if the input is a small square image, the vehicle cannot have a sufficient perception range for effective planning." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.432, + 0.482, + 0.522 + ], + "angle": 0, + "content": "Waypoints and route points. Although DriveGPT4-V2 does not rely on waypoints and route points for generating low-level control signals, they provide valuable supervision during training. Predicting waypoints and route points helps the model perceive the target route and anticipate future actions. Removing them negatively impacts performance." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.522, + 0.482, + 0.613 + ], + "angle": 0, + "content": "Expert on-policy supervision. The expert model, benefiting from privileged information, provides robust on-policy supervision to DriveGPT4-V2. Without expert guidance, DriveGPT4-V2 relies solely on behavior cloning instead of DAGger imitation learning, which can lead to error accumulation and degraded performance." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.614, + 0.482, + 0.749 + ], + "angle": 0, + "content": "PID controllers. Previous works often generated low-level control signals based on waypoints or route points [12, 21] by calculating target speed and angle from these predictions. However, even minor prediction errors in waypoints or route points can lead to substantial errors in speed and angle. As shown in Tab. 5, using waypoints or route points for PID control can make control signals noisy and unstable, potentially driving the vehicle into exceptional states. Therefore it presents inferior results." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.75, + 0.483, + 0.902 + ], + "angle": 0, + "content": "Decision heads. During inference, DriveGPT4-V2 generates four tokens in an auto-regressive manner, mapped to vehicle decision variables by four decision heads. Predictions of target speed and angle involve one-to-one mappings (i.e., one token for one number), while waypoint and route point predictions involve one-to-many mappings (e.g., waypoints contain 8 x-y points, thus each token is mapped to 16 numbers). Another design would involve predicting each waypoint and route point one at a time, requiring 8 tokens for waypoints and 10 tokens for route points. Such design" + }, + { + "type": "table_caption", + "bbox": [ + 0.513, + 0.089, + 0.905, + 0.118 + ], + "angle": 0, + "content": "Table 4. Ablation studies of DriveGPT4-V2. \"WP\" and \"RP\" represent waypoints and route points, respectively." + }, + { + "type": "table", + "bbox": [ + 0.517, + 0.123, + 0.901, + 0.223 + ], + "angle": 0, + "content": "
DSRCIS
Baseline47780.60
+ LLM Visual Pretraining56870.64
+ Visual Tokenizer60880.68
+ WP&RP63900.70
+ Expert Supervision70910.77
" + }, + { + "type": "table_caption", + "bbox": [ + 0.513, + 0.229, + 0.905, + 0.272 + ], + "angle": 0, + "content": "Table 5. Ablation studies on PID controllers. \"WP\" indicates utilizing predicted waypoints for PID control; while \"TS&RP\" means PID control by predicted target speed and route points." + }, + { + "type": "table", + "bbox": [ + 0.518, + 0.276, + 0.9, + 0.354 + ], + "angle": 0, + "content": "
PID ControllerDSRCIS
WP53850.62
TS & RP59880.67
DriveGPT4-V263900.70
" + }, + { + "type": "table_caption", + "bbox": [ + 0.513, + 0.361, + 0.905, + 0.39 + ], + "angle": 0, + "content": "Table 6. Ablation studies on decision heads. \"Additional tokens\" indicates using more output tokens for prediction." + }, + { + "type": "table", + "bbox": [ + 0.517, + 0.395, + 0.901, + 0.459 + ], + "angle": 0, + "content": "
DSRCISFPS
Additional tokens64910.701.4
DriveGPT4-V263900.708.1
" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.473, + 0.905, + 0.548 + ], + "angle": 0, + "content": "does not significantly improve performance and severely affects efficiency. Using text to represent output numbers would require approximately 160 tokens, making it impractical for real-time autonomous driving. The evaluation results are shown in Tab. 6" + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.563, + 0.783, + 0.578 + ], + "angle": 0, + "content": "5. Conclusion and Future Work" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.587, + 0.905, + 0.859 + ], + "angle": 0, + "content": "In this paper, we presented DriveGPT4-V2, a novel LLM-based framework for closed-loop, end-to-end autonomous driving. DriveGPT4-V2 processes camera images and vehicle states as input to generate low-level control signals for direct vehicle operation. Leveraging the extensive pretrained knowledge of MLLMs, DriveGPT4-V2 demonstrates the ability to navigate complex urban scenarios under diverse and challenging conditions. The model architecture is specifically optimized for precise numerical vehicle decision prediction. An additional expert LLM, which shares a structure similar to DriveGPT4-V2, has been trained to provide on-policy supervision. Experimental results highlight that DriveGPT4-V2 achieves state-of-the-art performance on the challenging CARLA Longest6 benchmark, outperforming all baselines. In the future, we aim to extend the capabilities of DriveGPT4-V2 for broader applications in closed-loop autonomous driving tasks, including video games and real-world deployment." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.859, + 0.905, + 0.901 + ], + "angle": 0, + "content": "Acknowledgment. This work is partially supported by the National Natural Science Foundation of China (No. 62422606, 62201484, 52221005)." + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "17268" + } + ], + [ + { + "type": "title", + "bbox": [ + 0.093, + 0.091, + 0.188, + 0.106 + ], + "angle": 0, + "content": "References" + }, + { + "type": "ref_text", + "bbox": [ + 0.101, + 0.116, + 0.482, + 0.141 + ], + "angle": 0, + "content": "[1] Anthropic. Claude. https://www.anthropic.com, 2024." + }, + { + "type": "ref_text", + "bbox": [ + 0.101, + 0.144, + 0.482, + 0.185 + ], + "angle": 0, + "content": "[2] Jinze Bai, Shuai Bai, Yunfei Chu, Zeyu Cui, Kai Dang, Xiaodong Deng, Yang Fan, Wenbin Ge, Yu Han, Fei Huang, et al. Qwen technical report. arXiv:2309.16609, 2023." + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.187, + 0.482, + 0.255 + ], + "angle": 0, + "content": "[3] Xu Cao, Tong Zhou, Yunsheng Ma, Wenqian Ye, Can Cui, Kun Tang, Zhipeng Cao, Kaizhao Liang, Ziran Wang, James M Rehg, et al. Maplm: A real-world large-scale vision-language benchmark for map and traffic scene understanding. In CVPR, 2024." + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.256, + 0.482, + 0.282 + ], + "angle": 0, + "content": "[4] Dian Chen and Philipp Krahenbuhl. Learning from all vehicles. In CVPR, 2022." + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.284, + 0.482, + 0.311 + ], + "angle": 0, + "content": "[5] Dian Chen, Brady Zhou, Vladlen Koltun, and Philipp Krahenbuhl. Learning by cheating. In CoRL, 2020." + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.313, + 0.482, + 0.339 + ], + "angle": 0, + "content": "[6] Dian Chen, Vladlen Koltun, and Philipp Krahenbuhl. Learning to drive from a world on rails. In ICCV, 2021." + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.341, + 0.482, + 0.408 + ], + "angle": 0, + "content": "[7] Long Chen, Oleg Sinavski, Jan Hünermann, Alice Karnsund, Andrew James Willmott, Danny Birch, Daniel Maund, and Jamie Shotton. Driving with llms: Fusing object-level vector modality for explainable autonomous driving. In ICRA, 2024." + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.411, + 0.482, + 0.451 + ], + "angle": 0, + "content": "[8] Li Chen, Penghao Wu, Kashyap Chitta, Bernhard Jaeger, Andreas Geiger, and Hongyang Li. End-to-end autonomous driving: Challenges and frontiers. TPAMI, 2024." + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.453, + 0.482, + 0.508 + ], + "angle": 0, + "content": "[9] Shaoyu Chen, Bo Jiang, Hao Gao, Bencheng Liao, Qing Xu, Qian Zhang, Chang Huang, Wenyu Liu, and Xinggang Wang. Vadv2: End-to-end vectorized autonomous driving via probabilistic planning. arXiv:2402.13243, 2024." + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.51, + 0.482, + 0.55 + ], + "angle": 0, + "content": "[10] Yuan Chen, Zi-han Ding, Ziqin Wang, Yan Wang, Lijun Zhang, and Si Liu. Asynchronous large language model enhanced planner for autonomous driving. In ECCV, 2024." + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.551, + 0.482, + 0.62 + ], + "angle": 0, + "content": "[11] Wei-Lin Chiang, Zhuohan Li, Zi Lin, Ying Sheng, Zhang-hao Wu, Hao Zhang, Lianmin Zheng, Siyuan Zhuang, Yong-hao Zhuang, Joseph E Gonzalez, et al. Vicuna: An opensource chatbot impressing gpt-4 with \\(90\\%\\) * chatgpt quality. See https://vicuna.lmsys.org, 2023." + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.621, + 0.482, + 0.676 + ], + "angle": 0, + "content": "[12] Kashyap Chitta, Aditya Prakash, Bernhard Jaeger, Zehao Yu, Katrin Renz, and Andreas Geiger. Transfuser: Imitation with transformer-based sensor fusion for autonomous driving. TPAMI, 2022." + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.677, + 0.481, + 0.703 + ], + "angle": 0, + "content": "[13] Google DeepMind. Gemini ai. https://deepmind.com, 2024." + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.705, + 0.482, + 0.746 + ], + "angle": 0, + "content": "[14] Alexey Dosovitskiy, German Ros, Felipe Codevilla, Antonio Lopez, and Vladlen Koltun. Carla: An open urban driving simulator. In CoRL, 2017." + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.748, + 0.482, + 0.788 + ], + "angle": 0, + "content": "[15] Yuchao Feng, Zhen Feng, Wei Hua, and Yuxiang Sun. Multimodal-xad: Explainable autonomous driving based on multimodal environment descriptions. TITS, 2024." + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.79, + 0.482, + 0.843 + ], + "angle": 0, + "content": "[16] Jiawei Fu, Yanqing Shen, Zhiqiang Jian, Shitao Chen, Jing-min Xin, and Nanning Zheng. Interactionnet: Joint planning and prediction for autonomous driving with transformers. In IROS, 2023." + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.845, + 0.482, + 0.899 + ], + "angle": 0, + "content": "[17] Yihan Hu, Jiazhi Yang, Li Chen, Keyu Li, Chonghao Sima, Xizhou Zhu, Siqi Chai, Senyao Du, Tianwei Lin, Wenhai Wang, et al. Planning-oriented autonomous driving. In CVPR, 2023." + }, + { + "type": "list", + "bbox": [ + 0.094, + 0.116, + 0.482, + 0.899 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.093, + 0.905, + 0.133 + ], + "angle": 0, + "content": "[18] Bin Huang, Xin Wang, Hong Chen, Zihan Song, and Wenwu Zhu. Vtimellm: Empower llm to grasp video moments. In CVPR, 2024." + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.136, + 0.905, + 0.175 + ], + "angle": 0, + "content": "[19] Ahmed Hussein, Mohamed Medhat Gaber, Eyad Elyan, and Chrisina Jayne. Imitation learning: A survey of learning methods. ACM Computing Surveys, 2017." + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.178, + 0.905, + 0.23 + ], + "angle": 0, + "content": "[20] Yuichi Inoue, Yuki Yada, Kotaro Tanahashi, and Yu Yamaguchi. Nuscenes-mqa: Integrated evaluation of captions and qa for autonomous driving datasets using markup annotations. In WACV, 2024." + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.233, + 0.905, + 0.259 + ], + "angle": 0, + "content": "[21] Bernhard Jaeger, Kashyap Chitta, and Andreas Geiger. Hidden biases of end-to-end driving models. In ICCV, 2023." + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.261, + 0.905, + 0.315 + ], + "angle": 0, + "content": "[22] Xiaosong Jia, Penghao Wu, Li Chen, Jiangwei Xie, Conghui He, Junchi Yan, and Hongyang Li. Think twice before driving: Towards scalable decoders for end-to-end autonomous driving. In CVPR, 2023." + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.317, + 0.905, + 0.37 + ], + "angle": 0, + "content": "[23] Bo Jiang, Shaoyu Chen, Qing Xu, Benchcheng Liao, Jiajie Chen, Helong Zhou, Qian Zhang, Wenyu Liu, Chang Huang, and Xinggang Wang. Vad: Vectorized scene representation for efficient autonomous driving. In ICCV, 2023." + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.372, + 0.905, + 0.426 + ], + "angle": 0, + "content": "[24] Zhuoling Li, Liangliang Ren, Jinrong Yang, Yong Zhao, Xiaoyang Wu, Zhenhua Xu, Xiang Bai, and Hengshuang Zhao. Virt: Vision instructed transformer for robotic manipulation. arXiv:2410.07169, 2024." + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.428, + 0.905, + 0.481 + ], + "angle": 0, + "content": "[25] Zhuoling Li, Xiaogang Xu, Zhenhua Xu, SerNam Lim, and Hengshuang Zhao. Larm: Large auto-regressive model for long-horizon embodied intelligence. arXiv:2405.17424, 2024." + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.484, + 0.905, + 0.537 + ], + "angle": 0, + "content": "[26] Bin Lin, Yang Ye, Bin Zhu, Jiaxi Cui, Munan Ning, Peng Jin, and Li Yuan. Video-llava: Learning united visual representation by alignment before projection. arXiv:2311.10122, 2023." + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.539, + 0.905, + 0.594 + ], + "angle": 0, + "content": "[27] Yuanfei Lin, Chenran Li, Mingyu Ding, Masayoshi Tomizuka, Wei Zhan, and Matthias Althoff. Drplanner: Diagnosis and repair of motion planners for automated vehicles using large language models. RA-L, 2024." + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.596, + 0.905, + 0.636 + ], + "angle": 0, + "content": "[28] Haotian Liu, Chunyuan Li, Yuheng Li, and Yong Jae Lee. Improved baselines with visual instruction tuning. In arXiv:2310.03744, 2023." + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.638, + 0.904, + 0.664 + ], + "angle": 0, + "content": "[29] Haotian Liu, Chunyuan Li, Qingyang Wu, and Yong Jae Lee. Visual instruction tuning. In NeurIPS, 2023." + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.665, + 0.905, + 0.733 + ], + "angle": 0, + "content": "[30] Haotian Liu, Chunyuan Li, Yuheng Li, Bo Li, Yuanhan Zhang, Sheng Shen, and Yong Jae Lee. Llava-last: Improved reasoning,OCR, and world knowledge. https://llava-v1.github.io/blog/2024-01-30-llava-last, 2024." + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.735, + 0.905, + 0.816 + ], + "angle": 0, + "content": "[31] Tianyu Liu, Qing hai Liao, Lu Gan, Fulong Ma, Jie Cheng, Xupeng Xie, Zhe Wang, Yingbing Chen, Yilong Zhu, Shuyang Zhang, et al. The role of the hercules autonomous vehicle during the COVID-19 pandemic: An autonomous logistic vehicle for contactless goods transportation. IEEE Robotics & Automation Magazine, 2021." + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.818, + 0.905, + 0.859 + ], + "angle": 0, + "content": "[32] Jiageng Mao, Yuxi Qian, Junjie Ye, Hang Zhao, and Yue Wang. Gpt-driver: Learning to drive with gpt. In NeurIPS Workshop, 2023." + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.86, + 0.905, + 0.901 + ], + "angle": 0, + "content": "[33] Maxim Mnyakin. Challenges and opportunities of integrating autonomous vehicles into urban retail delivery services. Reviews of Contemporary Business Analytics, 2023." + }, + { + "type": "list", + "bbox": [ + 0.517, + 0.093, + 0.905, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "17269" + } + ], + [ + { + "type": "ref_text", + "bbox": [ + 0.092, + 0.093, + 0.482, + 0.107 + ], + "angle": 0, + "content": "[34] OpenAI. Chatgpt. https://chat.openai.com, 2024." + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.108, + 0.483, + 0.163 + ], + "angle": 0, + "content": "[35] Takayuki Osa, Joni Pajarinen, Gerhard Neumann, J Andrew Bagnell, Pieter Abbeel, Jan Peters, et al. An algorithmic perspective on imitation learning. Foundations and Trends® in Robotics, 2018." + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.165, + 0.482, + 0.22 + ], + "angle": 0, + "content": "[36] SungYeon Park, MinJae Lee, JiHyuk Kang, Hahyeon Choi, Yoonah Park, Juhwan Cho, Adam Lee, and DongKyu Kim. Vlaad: Vision and language assistant for autonomous driving. In WACV, 2024." + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.222, + 0.483, + 0.29 + ], + "angle": 0, + "content": "[37] Alec Radford, Jong Wook Kim, Chris Hallacy, Aditya Ramesh, Gabriel Goh, Sandhini Agarwal, Girish Sastry, Amanda Askell, Pamela Mishkin, Jack Clark, et al. Learning transferable visual models from natural language supervision. In ICML, 2021." + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.292, + 0.483, + 0.347 + ], + "angle": 0, + "content": "[38] Katrin Renz, Kashyap Chitta, Otniel-Bogdan Mercea, A Koepke, Zeynep Akata, and Andreas Geiger. Plant: Explainable planning transformers via object-level representations. In CoRL, 2022." + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.349, + 0.483, + 0.417 + ], + "angle": 0, + "content": "[39] Katrin Renz, Long Chen, Ana-Maria Marcu, Jan Hunermann, Benoit Hanotte, Alice Karnsund, Jamie Shotton, Elahe Arani, and Oleg Sinavski. Carllava: Vision language models for camera-only closed-loop driving. arXiv:2406.10165, 2024." + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.42, + 0.482, + 0.46 + ], + "angle": 0, + "content": "[40] Stephane Ross and J. Andrew Bagnell. Reinforcement and imitation learning via interactive no-regret learning. In arXiv:1406.5979, 2014." + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.462, + 0.482, + 0.503 + ], + "angle": 0, + "content": "[41] Stéphane Ross, Geoffrey Gordon, and Drew Bagnell. A reduction of imitation learning and structured prediction to no-regret online learning. In AISTATS, 2011." + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.505, + 0.482, + 0.546 + ], + "angle": 0, + "content": "[42] Hao Shao, Letian Wang, Ruobing Chen, Hongsheng Li, and Yu Liu. Safety-enhanced autonomous driving using interpretable sensor fusion transformer. In CoRL, 2023." + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.548, + 0.482, + 0.602 + ], + "angle": 0, + "content": "[43] Hao Shao, Yuxuan Hu, Letian Wang, Guanglu Song, Steven L Waslander, Yu Liu, and Hongsheng Li. Lmdrive: Closed-loop end-to-end driving with large language models. In CVPR, 2024." + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.604, + 0.482, + 0.659 + ], + "angle": 0, + "content": "[44] Chonghao Sima, Katrin Renz, Kashyap Chitta, Li Chen, Hanxue Zhang, Chengen Xie, Jens Beiwenger, Ping Luo, Andreas Geiger, and Hongyang Li. Drivelm: Driving with graph visual question answering. In ECCV, 2024." + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.661, + 0.482, + 0.729 + ], + "angle": 0, + "content": "[45] Robert Soler, Alae Moudni, Gabriel Roskowski, Xinrui Yu, Mikhail Gormov, and Jafar Saniee. Autonomous patrol and threat detection through integrated mapping and computer vision. In IEEE International Conference on Electro Information Technology, 2024." + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.732, + 0.482, + 0.785 + ], + "angle": 0, + "content": "[46] Chan Hee Song, Jiaman Wu, Clayton Washington, Brian M Sadler, Wei-Lun Chao, and Yu Su. Llm-planner: Few-shot grounded planning for embodied agents with large language models. In ICCV, 2023." + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.788, + 0.482, + 0.842 + ], + "angle": 0, + "content": "[47] Yafu Tian, Alexander Carballo, Ruifeng Li, Simon Thompson, and Kazuya Takeda. Rsg-search plus: An advanced traffic scene retrieval methods based on road scene graph. In IV, 2024." + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.845, + 0.482, + 0.872 + ], + "angle": 0, + "content": "[48] Faraz Torabi, Garrett Warnell, and Peter Stone. Behavioral cloning from observation. In *IJCAI*, 2018." + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.874, + 0.482, + 0.901 + ], + "angle": 0, + "content": "[49] Hugo Touvron, Thibaut Lavril, Gautier Izacard, Xavier Martinet, Marie-Anne Lachaux, Timothée Lacroix, Baptiste" + }, + { + "type": "list", + "bbox": [ + 0.092, + 0.093, + 0.483, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.545, + 0.093, + 0.905, + 0.134 + ], + "angle": 0, + "content": "Rozière, Naman Goyal, Eric Hambro, Faisal Azhar, et al. Llama: Open and efficient foundation language models. arXiv:2302.13971, 2023." + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.136, + 0.905, + 0.189 + ], + "angle": 0, + "content": "[50] Yuxi Wei, Zi Wang, Yifan Lu, Chenxin Xu, Changxing Liu, Hao Zhao, Siheng Chen, and Yanfeng Wang. Editable scene simulation for autonomous driving via collaborative llm-agents. In CVPR, 2024." + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.191, + 0.905, + 0.245 + ], + "angle": 0, + "content": "[51] Licheng Wen, Daocheng Fu, Xin Li, Xinyu Cai, Tao Ma, Pinlong Cai, Min Dou, Botian Shi, Liang He, and Yu Qiao. Dilu: A knowledge-driven approach to autonomous driving with large language models. In ICLR, 2024." + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.247, + 0.905, + 0.287 + ], + "angle": 0, + "content": "[52] Zhenhua Xu, Yuxuan Liu, Lu Gan, Yuxiang Sun, Ming Liu, and Lujia Wang. Ringdet: Road network graph detection by transformer in aerial images. TGRS, 2022." + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.289, + 0.905, + 0.342 + ], + "angle": 0, + "content": "[53] Zhenhua Xu, Yuxuan Liu, Yuxiang Sun, Ming Liu, and Lujia Wang. Rngdet++: Road network graph detection by transformer with instance segmentation and multi-scale features enhancement. RA-L, 2023." + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.344, + 0.905, + 0.399 + ], + "angle": 0, + "content": "[54] Zhenhua Xu, Yujia Zhang, Enze Xie, Zhen Zhao, Yong Guo, Kwan-Yee K Wong, Zhenguo Li, and Hengshuang Zhao. Drivegpt4: Interpretable end-to-end autonomous driving via large language model. RA-L, 2024." + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.401, + 0.905, + 0.453 + ], + "angle": 0, + "content": "[55] Yijun Yang, Tianyi Zhou, Kanxue Li, Dapeng Tao, Lusong Li, Li Shen, Xiaodong He, Jing Jiang, and Yuhui Shi. Embodied multi-modal agent trained by an llm from a parallel textworld. In CVPR, 2024." + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.455, + 0.905, + 0.496 + ], + "angle": 0, + "content": "[56] Ekim Yurtsever, Jacob Lambert, Alexander Carballo, and Kazuya Takeda. A survey of autonomous driving: Common practices and emerging technologies. IEEE Access." + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.498, + 0.905, + 0.537 + ], + "angle": 0, + "content": "[57] Xiaohua Zhai, Basil Mustafa, Alexander Kolesnikov, and Lucas Beyer. Sigmoid loss for language image pre-training. In ICCV, 2023." + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.539, + 0.905, + 0.566 + ], + "angle": 0, + "content": "[58] Jimuyang Zhang, Zanming Huang, and Eshed Ohn-Bar. Coaching a teachable student. In CVPR, 2023." + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.568, + 0.905, + 0.607 + ], + "angle": 0, + "content": "[59] Jimuyang Zhang, Zanming Huang, Arjit Ray, and Eshed Ohn-Bar. Feedback-guided autonomous driving. In CVPR, 2024." + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.609, + 0.905, + 0.649 + ], + "angle": 0, + "content": "[60] Jiawei Zhang, Chejian Xu, and Bo Li. Chatscene: Knowledge-enabled safety-critical scenario generation for autonomous vehicles. In CVPR, 2024." + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.651, + 0.905, + 0.691 + ], + "angle": 0, + "content": "[61] Peiyuan Zhang, Guangtao Zeng, Tianduo Wang, and Wei Lu. Tinyllama: An open-source small language model. arXiv:2401.02385, 2024." + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.693, + 0.905, + 0.748 + ], + "angle": 0, + "content": "[62] Xiang Zhang, Haojie Sun, Xiaoyang Pei, Linghui Guan, and Zihao Wang. Evolution of technology investment and development of robotaxi services. Transportation Research Part E: Logistics and Transportation Review, 2024." + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.749, + 0.905, + 0.789 + ], + "angle": 0, + "content": "[63] Zhejun Zhang, Alexander Liniger, Dengxin Dai, Fisher Yu, and Luc Van Gool. End-to-end urban driving by imitating a reinforcement learning coach. In ICCV, 2021." + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.791, + 0.905, + 0.843 + ], + "angle": 0, + "content": "[64] Baichuan Zhou, Ying Hu, Xi Weng, Junlong Jia, Jie Luo, Xien Liu, Ji Wu, and Lei Huang. Tinyllava: A framework of small-scale large multimodal models. arXiv:2402.14289, 2024." + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.846, + 0.905, + 0.9 + ], + "angle": 0, + "content": "[65] Deyao Zhu, Jun Chen, Xiaoqian Shen, Xiang Li, and Mohamed Elhoseiny. Minigpt-4: Enhancing vision-language understanding with advanced large language models. In ICLR, 2024." + }, + { + "type": "list", + "bbox": [ + 0.517, + 0.093, + 0.905, + 0.9 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.52, + 0.957 + ], + "angle": 0, + "content": "17270" + } + ] +] \ No newline at end of file diff --git a/2025/DriveGPT4-V2_ Harnessing Large Language Model Capabilities for Enhanced Closed-Loop Autonomous Driving/31ff1a5c-7171-4c25-a1d0-ee2bacb8057b_origin.pdf b/2025/DriveGPT4-V2_ Harnessing Large Language Model Capabilities for Enhanced Closed-Loop Autonomous Driving/31ff1a5c-7171-4c25-a1d0-ee2bacb8057b_origin.pdf new file mode 100644 index 0000000000000000000000000000000000000000..21d5d5f9c85c275c753b3a009af0b8211f65bb6e --- /dev/null +++ b/2025/DriveGPT4-V2_ Harnessing Large Language Model Capabilities for Enhanced Closed-Loop Autonomous Driving/31ff1a5c-7171-4c25-a1d0-ee2bacb8057b_origin.pdf @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:0fbf529c958322c4fe21e2d93ef07fd967b80d4f549541821487f51772092909 +size 959639 diff --git a/2025/DriveGPT4-V2_ Harnessing Large Language Model Capabilities for Enhanced Closed-Loop Autonomous Driving/full.md b/2025/DriveGPT4-V2_ Harnessing Large Language Model Capabilities for Enhanced Closed-Loop Autonomous Driving/full.md new file mode 100644 index 0000000000000000000000000000000000000000..24237a3cb71bf93eeab884da33982c837b524692 --- /dev/null +++ b/2025/DriveGPT4-V2_ Harnessing Large Language Model Capabilities for Enhanced Closed-Loop Autonomous Driving/full.md @@ -0,0 +1,293 @@ +# DriveGPT4-V2: Harnessing Large Language Model Capabilities for Enhanced Closed-Loop Autonomous Driving + +Zhenhua Xu $^{1,2}$ Yan Bai $^{3}$ Yujia Zhang $^{1}$ Zhuoling Li $^{1}$ Fei Xia $^{3}$ Kwan-Yee K. Wong $^{1}$ Jianqiang Wang $^{2}$ Hengshuang Zhao $^{1*}$ The University of Hong Kong Tsinghua University Meituan + +# Abstract + +Multimodal large language models (MLLMs) possess the ability to comprehend visual images or videos, and show impressive reasoning ability thanks to the vast amounts of pretrained knowledge, making them highly suitable for autonomous driving applications. Unlike the previous work, DriveGPT4-V1, which focused on open-loop tasks, this study explores the capabilities of LLMs in enhancing closed-loop autonomous driving. DriveGPT4-V2 processes camera images and vehicle states as input to generate low-level control signals for end-to-end vehicle operation. A multi-view visual tokenizer (MV-VT) is employed enabling DriveGPT4-V2 to perceive the environment with an extensive range while maintaining critical details. The model architecture has been refined to improve decision prediction and inference speed. To further enhance the performance, an additional expert LLM is trained for online imitation learning. The expert LLM, sharing a similar structure with DriveGPT4-V2, can access privileged information about surrounding objects for more robust and reliable predictions. Experimental results show that DriveGPT4-V2 outperforms all baselines on the challenging CARLA Longest6 benchmark. The code and data of DriveGPT4-V2 will be publicly available. + +# 1. Introduction + +Autonomous driving has seen rapid advancements over the past decade in both academia and industry [31, 56]. Today, autonomous vehicles are deployed commercially in diverse applications such as robotaxi services [62], goods delivery [33], and security patrols [45]. Additionally, assisted driving systems are increasingly integrated into consumer vehicles, enhancing safety and convenience. However, the majority of current autonomous driving systems still rely heavily on rule-based decision-making modules, where planning and control are governed by predefined algorithms and + +![](images/e2c2dfe1095507cf1537101d0eb554fb6c817cee0ee39103ff335f5c6fae8a11.jpg) +Figure 1. DriveGPT4-V2 for closed-loop autonomous driving. Taken as input multi-view camera images and vehicle state information, DriveGPT4-V2 predicts high-level vehicle decisions and converts them to low-level vehicle control signals in an end-to-end manner. DriveGPT4-V2 presents outstanding effectiveness and efficiency, serving as a reliable baseline method for future research on autonomous driving with LLMs. + +rules. This reliance restricts the ability of these systems to operate effectively in varied, real-world scenarios, limiting the broader advancement of autonomous driving. + +In response, recent research has shifted toward end-to-end autonomous driving systems based on learning-based methods [8, 9, 12, 17, 21, 23, 45]. These methods enable more streamlined and intelligent system designs by directly learning control from input sensor data, without the need for complicated intermediate modules. From a learning theory perspective, end-to-end autonomous driving can optimize the entire system on the final outputs, rather than through the isolated optimization of individual modules, which potentially improves overall performance. + +The rise of large language models (LLMs) has demonstrated their impressive ability to perform natural language processing (NLP) tasks [1, 2, 13, 34, 49]. Commercial LLMs, such as ChatGPT [34] and Claude3 [1], are already + +widely used in daily life. Motivated by their success in NLP tasks, LLMs are being extended to multimodal domains, enabling more generalized tasks in fields like image analysis [28-30], video comprehension [18, 26], and embodied AI [24, 25, 46, 55]. Given their versatility, multimodal LLMs have been applied to autonomous driving for tasks such as interpretability [15, 36, 44, 54], vehicle control [10, 27, 27], and risk assessment [3, 60]. While multimodal LLMs have demonstrated promising vehicle control performance due to their extensive pretrained knowledge, previous research has primarily concentrated on open-loop settings, which are not implemented for real-time vehicle control. Open-loop systems lack feedback mechanisms, making them less adaptable to changes and disturbances, limiting their applicability in real-world applications. Thus, multimodal LLMs need to be designed and evaluated specifically for closed-loop autonomous driving scenarios [27, 39, 43, 51]. + +Behavior cloning [35, 48], the most commonly employed imitation learning algorithm, is used for most end-to-end autonomous driving works [44, 54]. However, models trained only by behavior cloning can encounter drifting issues in closed-loop tasks, where errors accumulate, potentially leading the vehicle into unsafe situations, such as deviating from the path, colliding, or getting stuck. + +To address the aforementioned issue, we incorporate online (or on-policy) imitation learning [5, 40, 41] into closed-loop autonomous driving. Leveraging the reasoning capabilities and pretrained knowledge of LLMs, we propose using a multimodal LLM as a fully automatic expert for online imitation learning without human efforts. Let $M_E$ and $M_A$ denote the model of expert and agent, respectively. $M_E$ can access privileged information about nearby objects while $M_A$ cannot, thus presenting a more powerful and reliable performance. After being trained by behavior cloning, $M_A$ is deployed on the training environment for data aggregation. $M_A$ might make incorrect predictions that lead the vehicle into exception states. $M_E$ can provide $M_A$ with on-policy supervision and take over the vehicle if the error made by $M_A$ is large enough. When $M_E$ takes over the vehicle, it collects one data sample at the moment and aggregates it into the dataset. The aggregated dataset is leveraged to further train the agent model to handle errors. + +In this paper, we introduce DriveGPT4-V2, an end-to-end autonomous driving system. DriveGPT4-V2 extends DriveGPT4-V1 [54] to closed-loop tasks and integrates online imitation learning for enhanced robustness and intelligence. The "4" in the system's name signifies its multimodal capacity, inspired by the MiniGPT4 model [65]. DriveGPT4-V2 is powered by multimodal LLMs, enabling it to directly generate low-level vehicle control signals (i.e., throttle, brake and steer) based on multimodal input data (i.e., vehicle states and camera images). The structure of DriveGPT4-V2 is refined for our task, including a + +multi-view visual tokenizer (MV-VT) that possesses a large enough perception range without losing details, as well as output decision heads (DeciHeads) specially designed for numerical vehicle decision prediction. An additional multimodal LLM with access to privileged information (e.g., ground truth of nearby objects in text) serves as an expert for online imitation learning, which is inspired by the LBC framework [5]. DriveGPT4-V2 achieves state-of-the-art performance on the challenging CARLA Longest6 benchmark [12], which evaluates systems over 36 long routes demanding high levels of intelligence and reliability. The overview of DriveGPT4-V2 is visualized in Fig. 1. The contributions of this paper are: + +- We extend DriveGPT4-V1, an open-loop autonomous driving system with multimodal LLMs, and propose DriveGPT4-V2 for end-to-end closed-loop autonomous driving. DriveGPT4-V2 can better harness the capability of LLMs and be deployed for continuous vehicle control. +- We introduce the use of LLMs with privileged information as an expert for online (on-policy) imitation learning, facilitating the generation of labeled data for handling errors, which significantly enhances the performance and robustness of closed-loop autonomous driving. +- We optimize the model structure of MLLMs, including a multi-view visual tokenizer (MV-VT) that possesses a large perception range and preserves details, as well as output decision heads (DeciHeads) specially designed for numerical vehicle decision prediction. +- DriveGPT4-V2 demonstrates superior performance on the CARLA Longest6 benchmark, substantially outperforming existing methods. + +# 2. Related Works + +Large language models for autonomous driving. Owing to the powerful reasoning capabilities and extensive pretrained knowledge, large language models (LLMs) have found wide application in autonomous driving tasks, such as vehicle decision-making [7, 10, 15, 27, 32, 39, 43, 54, 59], scene understanding [20, 36, 44, 54], risk analysis [3, 60], and simulation [47, 50]. GPT-Driver [32] represents the entire surrounding environment textually and employs a fine-tuned ChatGPT model to predict vehicle trajectories. DriveGPT4 [54] leverages multimodal LLMs for end-to-end vehicle action prediction and scene understanding. It processes 8-frame video clips and human user commands to predict vehicle control signals and action explanations. However, most of these approaches are designed for open-loop scenarios and present unsatisfactory performance in closed-loop tasks, limiting their applicability to real-world driving deployment. Additionally, many existing methods face efficiency challenges due to heavy model architectures. + +Closed-loop end-to-end autonomous driving. Closed-loop end-to-end autonomous driving has recently emerged + +![](images/955fc14f1df6defcabc113baf58f6b04fddc39a90f3b17b5024c9d5e203d8c8c.jpg) +Figure 2. DriveGPT4-V2 diagram. DriveGPT4-V2 takes multimodal input data to generate numerical control signals for end-to-end vehicle driving. The input includes multi-view images and vehicle state information. The images are transformed into the text domain via a multi-view vision tokenizer (MV-VT). The current speed of the vehicle and target point are tokenized by the LLM tokenizer. The LLM then outputs four tokens. Each token is used to predict one vehicle decision by an MLP decision head (DeciHead). These predicted decisions are subsequently converted to low-level commands via PID controllers to operate the vehicle. The LLM expert model, which shares a similar structure to DriveGPT4-V2, has access to privileged information about surroundings (shown in the purple module). The expert provides on-policy supervision to DriveGPT4-V2 to enhance closed-loop performance. This figure is best viewed in color. + +as a prominent research focus, addressing the challenge of bridging the gap between academic research and real-world applications. Various approaches improved transformer architectures and training methods to enhance driving performance [4, 5, 12, 16, 21, 22, 42, 58, 63]. For instance, Transfuser++ [21] integrates multimodal sensor inputs, including camera imagery, LiDAR point clouds and vehicle dynamics, to predict speed and trajectory, achieving state-of-the-art performance across multiple benchmarks. Even though most early LLM-based methods have primarily targeted open-loop tasks, recent studies [27, 39, 43, 51] have attempted closed-loop approaches. But most of them are deployed in simple environments like Highway-env [51], instead of real-world scenarios. LMDrive [43], pioneering work on end-to-end closed-loop driving with multimodal LLMs, demonstrated the potential of LLMs in this domain. It is deployed in the CARLA simulator [14] and controls the vehicle like human drivers. However, it relies solely on behavior cloning and presents unsatisfactory driving performance. Online imitation learning has been highlighted as a potential way to enhance model robustness. + +Imitation learning in autonomous driving. Imitation learning [19] aims to train an agent to mimic the expert behavior. Behavior cloning is the most widely used imitation learning algorithm in autonomous driving [4, 12, 21, 39, 43, 44, 52-54]. In this method, data is off-policy sampled from expert trajectories for model training to replicate expert actions. Although behavior cloning can yield reasonable performance, it suffers from error accumulation during closed-loop operation, leading the model to encounter unseen situations that may compromise safety. To mitigate this issue, online imitation learning [40, 41] has been employed in various works [4, 5]. The LBC framework [5] uses the DAgger algorithm [41], and trains an agent alongside an expert network with privileged information that achieves superior performance. The expert provides on-policy supervision to the agent and collects additional data samples for data aggregation, which can be used to further train the agent model to handle errors. Inspired by LBC, DriveGPT4-V2 harnesses the capabilities of LLMs and leverages a privileged LLM expert to provide on-policy supervision, facilitating improved driving performance and robustness. + +![](images/995738dfecd7412e6c9b2dddf492c5aeb9f0b263cf8538985befe2873bc33c90.jpg) +Figure 3. Multi-view visual tokenizer (MV-VT) structure. The input images consist of three front views. Each patch is processed through a visual encoder to extract features. Finally, a trained projection layer maps the downsampled feature into the text domain for further processing. +Figure 4. Visualization of waypoints and route points. The ego vehicle is represented by the green rectangle, and the red point denotes the target point. The grey line is the route that the vehicle should follow. (a) Waypoints (blue points) represent vehicle positions in a 4-second future. (b) Route points are evenly sampled from the global planned route in front. These two kinds of points can better supervise the training of DriveGPT4-V2. + +# 3. DriveGPT4-V2 + +Building on DriveGPT4-V1 [54], a flexible multimodal LLM designed for open-loop autonomous driving, DriveGPT4-V2 has been specifically developed for closed-loop driving tasks. This paper employs LLMs both as planners and as expert models for online imitation learning. A multi-view visual tokenizer (MV-VT) is leveraged for effective feature extraction. Unlike standard LLMs used for NLP, the architecture of DriveGPT4-V2 is tailored for vehicle planning, with additional output decision heads (DeciHeads) to directly optimize numerical decision-making predictions. This design enhances model efficiency by reducing the token count required for inference. The expert model, which shares a similar architecture to DriveGPT4-V2, has access to privileged ground-truth information about the surroundings and is used solely for on-policy supervision during training, not for inference. + +# 3.1. Model Design + +DriveGPT4-V2 takes multimodal input data and predicts high-level vehicle decisions for end-to-end autonomous driving. These predictions are subsequently converted into low-level control signals through PID controllers. The model structure is shown in Fig. 2. + +Multimodal input. The input consists of two modalities: image and text. The image input is captured by three front-view cameras at a resolution of $384 \times 384$ , while the text input includes the vehicle's current speed and the target point. These three front-view images ensure coverage of both sides of the vehicle, reducing the risk of missing adjacent vehicles or potential collisions. The image data is + +![](images/c33d38c0cd5c96737f087a11329e34e686c48f64f9cfa5cb70a639771da57302.jpg) +(a) Waypoints. +(b) Route points. + +mapped into the text domain via a visual tokenizer, as implemented in LLaVA [30], while text inputs are tokenized with standard LLM tokenizers. + +Multi-view visual tokenizer (MV-VT). The visual tokenizer is visualized in Fig. 3. It first employs pretrained visual encoders, such as CLIP [37] or SigLIP [57], to extract features from the input images. The size of each image is specifically optimized for compatibility with the SigLIP encoder (i.e., $384 \times 384$ ). The three images are acquired from front-left, front, and front-right cameras respectively, ensuring comprehensive coverage of the vehicle's frontal and lateral views. In CARLA benchmarks, back-view images exhibit limited utilization frequency and demonstrate no measurable improvement in planning performance. To maintain computational efficiency, MV-VT deliberately excludes back-view inputs. + +Output representations. DriveGPT4-V2 generates three low-level control signals to drive the vehicle, i.e., throttle, brake, and steering. However, direct prediction of these signals requires temporal consistency; otherwise, the vehicle may display unstable zigzag trajectories. To ensure stable control, DriveGPT4-V2 instead predicts high-level decision variables, which are then converted to low-level commands by PID controllers. The predictions of the model include: + +- Target Speed: The desired speed for the vehicle. A speed PID controller adjusts longitudinal control based on the predicted target speed and current speed. +- Target Angle: The direction in which the vehicle should steer. A steering PID controller uses it for lateral control. +- Waypoints: Predicted x-y positions of the vehicle in the future. Waypoints consist of eight points. +- Route Points: Points evenly sampled along the global route in front, enhancing the vehicle's perception ability. + +Waypoints and route points are visualized in Fig. 4. + +Decision heads (DeciHeads). For closed-loop autonomous driving, DriveGPT4-V2 employs dedicated output heads (MLPs) to predict numerical vehicle decisions. The original LLM's vocabulary output head is replaced with four decision heads, each predicting a decision variable (i.e., target speed, target angle, waypoints, and route points) based on one token. This setup allows DriveGPT4-V2 to produce only four tokens per prediction cycle, achieving a huge + +speed increase over traditional text-based outputs. For example, a float number with 3 digits (e.g., 1.538) needs 5 tokens for text output representation but only requires a single token for our design. DeciHead also enables direct optimization of decision variables using regression loss. + +Expert model and online imitation learning. The expert model has a similar architecture to DriveGPT4-V2 but has access to privileged information. The privileged information consists of ground truth (GT) of nearby objects and hazard information provided by the CARLA simulator (e.g., traffic light and stop sign violation; vehicle or pedestrian collision). This privileged data allows the expert to better perceive the environment, so that it can focus on vehicle decisions and planning. The privileged information greatly boosts the model capacity, resulting in enhanced closed-loop performance. The expert model provides on-policy supervision to DriveGPT4-V2 during training. + +When DriveGPT4-V2 runs on training scenarios and routes, the expert model receives identical input and generates concurrent predictions. If the discrepancy between DriveGPT4-V2's prediction and the expert's exceeds a threshold, the expert's prediction is used instead to control the vehicle. Such a situation is marked as an exception case. A data sample of the current moment is added to the training set, where the expert's output is used as the label. DriveGPT4-V2 is fine-tuned on this aggregated dataset under the expert's on-policy supervision for enhanced closed-loop autonomous driving performance. + +The expert LLM is used exclusively for on-policy supervision during training and is not accessed during inference. + +# 3.2. Training + +The training process of DriveGPT4-V2 and the expert LLM includes two stages. In the first stage, both models are trained on data collected by a rule-based autopilot using behavior cloning. In the second stage, DriveGPT4-V2 runs on the training scenarios and routes under the expert's on-policy supervision, and generates additional data for error correction. The training process is visualized in Fig. 5. + +# 3.2.1. Stage 1: behavior cloning + +In the first stage, a rule-based autopilot [21] is deployed in CARLA to collect data. The autopilot agent records data at fixed time intervals, including camera images, current vehicle speed, privileged information on surrounding objects (e.g., vehicles and traffic lights), and a target point. Labels for the vehicle decision variables are calculated using rule-based algorithms. After data collection, both DriveGPT4-V2 and the expert LLM are trained on the dataset, with the expert having access to privileged information. + +# 3.2.2. Stage 2: DAgger with on-policy supervision + +The expert, with access to privileged information, achieves notably higher performance by leveraging reliable priors on + +![](images/501a1f5a0dfb3947b86c41c55fe6c3d1a51e80bb06853d866e54b38248154e56.jpg) +(a) Stage 1: Behavior cloning. + +![](images/d7e4f976a0eeeff15f2c7c35142437c32eb23e636d668a8e8889e42777e2c61c.jpg) +(b) Stage 2: DAgger online imitation learning. +Figure 5. Diagram of the two-stage training process. (a) In the first stage, both DriveGPT4-V2 and the expert LLM are trained on data collected by a rule-based autopilot. (b) In the second stage, DriveGPT4-V2 runs on the training scenarios and routes. When the discrepancy between DriveGPT4-V2's predictions and those of the expert exceeds a predefined threshold, the expert's predictions are used to control the vehicle. Data from these cases is then added to the dataset for data aggregation. + +surrounding objects, allowing it to handle challenging scenarios effectively. DriveGPT4-V2, trained only by behavior cloning, may struggle with error accumulation and exceptional situations (e.g., vehicle collisions or getting stuck). Inspired by LBC [5], the expert LLM is used as a teacher for on-policy supervision. On-policy supervision is a training paradigm in imitation learning, where supervision or feedback is provided to an agent while it is acting according to its current policy. This approach contrasts with off-policy methods, where supervision comes from data collected by a different policy or behavior. During the operation of DriveGPT4-V2, the expert and DriveGPT4-V2 receive identical input, and their outputs are compared. If the discrepancy in predicted decisions exceeds a threshold, the expert takes control to avoid exceptions. One data sample is collected at that moment and aggregated into the dataset. The output of the expert is used as the training label. After traversing the training routes under expert supervision, DriveGPT4-V2 is fine-tuned on the aggregated dataset. + +# 3.3. Dataset + +Following Transfuser++ [21], in stage 1, the rule-based autopilot agent is used to collect data samples in various CARLA simulator towns under varying weather, lighting, and daytime conditions. Vehicle speed and position are estimated by Kalman filters based on IMU data. Each sam + +ple records objects within a fixed radius, including vehicles, pedestrians, traffic lights, and stop signs. Target speeds and angles are calculated by the autopilot's PID controller. The ground-truth waypoints contain 8 points and record the vehicle position in a 4-second future. Route points have 10 points sampled along the planned route in front. + +In stage 2, DriveGPT4-V2 runs on selected training routes for data aggregation. Simple routes are not considered in this stage for data collection efficiency. If the difference between the predicted decisions of DriveGPT4-V2 and the expert exceeds a threshold, the expert takes control and records the data sample. DriveGPT4-V2 is retrained on the aggregated dataset for better performance. Under common circumstances, DriveGPT4-V2 achieves satisfactory performance after one round of DAgger. + +# 3.4. Loss Function + +During training, the L1 loss function supervises all predicted decision variables, including target speed $(TS)$ , target angle $(TA)$ , waypoints $(WP)$ , and route points $(RP)$ . The overall loss is calculated by: + +$$ +\mathcal {L} = \mathcal {L} _ {T S} + \mathcal {L} _ {T A} + \mathcal {L} _ {W P} + \mathcal {L} _ {R P} \tag {1} +$$ + +This approach allows the model to optimize final numerical vehicle decision variables directly, instead of relying on vocabulary-based classification as in NLP tasks. + +# 4. Experiments + +# 4.1. Dataset and Benchmark + +All experiments in this study were conducted in the CARLA simulator [14]. We followed Transfuser++ for training data preparation. In stage 1, a rule-based autopilot was employed to collect 350K behavior-cloning samples. The autopilot traversed multiple routes in Towns 1, 2, 3, 4, 5, 6, 7, and 10 of CARLA, encountering various scenarios such as different weather and lighting conditions. Data were collected at a frequency of $2\mathrm{Hz}$ and subsequently filtered to remove incorrect or redundant samples, such as those captured when the autopilot was stuck. This process resulted in a dataset of 300K samples. + +In stage 2, DriveGPT4-V2 trained by behavior cloning runs on selected training scenarios and routes, utilizing the expert for on-policy supervision. The expert can access privileged information about objects within a 30-meter distance. Approximately 150K samples were gathered in this stage. These aggregated data were then combined with the original dataset to further fine-tune DriveGPT4-V2, enhancing its performance. + +All methods were evaluated on the challenging CARLA Longest6 benchmark [12], which consists of 36 extended routes encompassing various scenarios. The length and complexity of routes make them suitable for a comprehensive evaluation of closed-loop autonomous driving. + +Table 1. Infraction penalty factors. + +
ScenePenaltyScenePenalty
Pedestrian collision0.5Vehicle collision0.6
Static object collision0.65Traffic light violation0.7
Stop sign violation0.7
+ +# 4.2. DriveGPT4-V2 Configuration + +LLM. For efficient data collection and inference, we used tiny-scale LLMs in DriveGPT4-V2, such as Qwen-0.5B [2] and TinyLLaMA [61]. These models were fine-tuned for multimodal understanding, following methods outlined in LLaVA [29, 30] and TinyLLaVA [64]. Large-scale LLMs, like LLaMA3-8B [49] and vicuna-7B [11], are not considered at the moment due to their high computational demands and reduced inference speed. + +Training. DriveGPT4-V2 was fully fine-tuned in an end-to-end manner, except for the SigLIP $384 \times 384$ visual encoder [57]. The model was trained with a learning rate of $2e^{-5}$ over 60 epochs by employing a cosine annealing schedule for better learning rate control. The training takes about 75 hours on 16 A800 GPUs. + +Inference. Two PID controllers were designed to convert DriveGPT4-V2's predicted decision variables into low-level control signals (i.e., throttle, steer, and brake). These control signals were used to directly drive the ego vehicle in the CARLA simulator. In the experiments, only the predicted target speed and target angle were used by the PID controllers, while waypoints and route points served as additional supervision for improved training results. The model runs inference on a single A800 GPU with FP16 precision. + +# 4.3. Evaluation Criteria + +Following the CARLA leaderboard, the methods were evaluated mainly based on three metrics: Driving Score (DS), Route Completion (RC), and Infraction Score (IS). RC indicates the percentage of routes completed, while IS penalizes infractions such as traffic violations and collisions during inference. Each infraction multiplicatively reduces IS by a penalty factor. The penalty factors are listed in Tab. 1. DS, as the product of RC and IS, provides a comprehensive evaluation score. In the evaluation, we also report ratios of commonly seen exception events, including pedestrian collision (Ped), vehicle collision (Veh), static object collision (Stat), red light violation (Red), route deviations (Dev), scenario time out (TO) and agent block (Block). + +# 4.4. Comparison Experiments + +Baselines. We selected several representative prior works as baselines for comparison: + +- WOR [6] (ICCV 2021): WOR uses commands instead of target points for global planning instructions, providing + +Table 2. Closed-loop experiments performance on CARLA Longest6. "Visual" indicates visual input modalities, while "C" and "L" represent camera and LiDAR, respectively. Bold numbers highlight SOTA metric scores of all models; while underlined numbers represent the best metric scores of baseline methods. * denotes models implemented by ourselves based on official open-sourced code. † represents the model without data augmentation. + +
MethodVisualDS↑RC↑IS↑Ped↓Veh↓Stat↓Red↓Dev↓TO↓Block↓
WOR [6]C21480.560.181.050.371.280.880.080.20
LAV v1 [4]C&L33700.510.160.830.150.960.060.120.45
Interfuser [42]C47740.630.061.140.110.240.000.520.06
TransFuser [12]C&L47930.500.032.450.070.160.000.060.10
LAV v2 [4]C&L58830.680.000.690.150.230.080.320.11
Perception PlanT [38]C&L58880.650.070.970.110.090.000.130.13
Transfuser++* [21]C&L65900.720.000.990.010.070.000.100.12
Transfuser++*† [21]C&L58890.650.011.150.010.100.000.140.13
LMDrive* [43]C&L36690.520.071.030.181.010.090.110.22
DriveGPT4-V2C70910.770.000.800.010.040.000.070.09
+ +enriched labels and supervision for possible actions. + +- LAV v1 & v2 [4] (CVPR 2022): LAV predicts waypoints by integrating information from nearby vehicles, which is further refined through a GRU. +- Transfuser [12] (TPAMI 2022): Transfuser is a widely used baseline model for closed-loop driving that has inspired subsequent works. +- Perception PlanT [38] (CoRL 2022): PlanT is based on imitation learning with object-level input representation. Such representation is elegant and effective. +- InterFuser [42] (CoRL 2023): InterFuser predicts additional density maps and traffic rule flags to enhance waypoint prediction. +- Transfuser++ [21] (ICCV 2023): Transfuser++ predicts target speeds and routes instead of waypoints, integrating point cloud and camera images. It is currently the state-of-the-art (SOTA) method for CARLA Longest6. Transfuser++ utilizes triple training data by repeatedly collecting expert trajectories for augmentation. +- LMDrive [43] (CVPR 2024): LMDrive is a pioneering work for closed-loop driving with multimodal LLMs, training a vicuna-7B LLM for end-to-end vehicle control. + +Comparison results. The comparison results on the CARLA Longest6 benchmark are shown in Tab. 2. It is observed that DriveGPT4-V2 outperforms all baseline models by a large margin. Thanks to the huge amount pretrained knowledge of LLMs and visual encoders, DriveGPT4-V2 can better handle complex urban scenes under various conditions, such as nighttime, rainy weather, etc. For most baselines, we list the results reported on the CARLA Longest6 leaderboard. However, some baselines cannot achieve reported performance by our implementation and test, thus we report our implemented results separately. LMDrive [43] is the only open-sourced LLM method that can be tested in a closed-loop environment. But it fails + +Table 3. Efficiency analysis. + +
LLMDSTrainFPS
LLaVA-LLaMA3.1-8B6511.2h/epoch0.4
TinyLLaVA-LLaMA-1.5B633.0h/epoch2.9
LLaVA-Qwen-0.5B631.3h/epoch8.1
+ +to obtain satisfactory performance. It first converts multimodal data to BEV and projects BEV to the text domain. However, BEV does not have good-quality pretrained multimodal LLMs. Training from scratch severely degrades its final performance, which is also discussed by DriveGPT4-V1 [54]. Transfuser++ [21] is the SOTA method by fusing point clouds and camera images for closed-loop control. Transfuser++ uses triple training data and achieves 65 DS, but only has 58 DS with the same amount of training data (denoted by $\dagger$ ) as our method. Thanks to the pretrained knowledge and reasoning ability of LLMs, DriveGPT4-V2 can outperform all baselines with camera images as input. + +# 4.5. Model Efficiency + +In our experiments, DriveGPT4-V2 and the expert are implemented by tiny LLMs. LLMs with 7B or even more parameters show powerful reasoning and generalization ability, but they severely enlarge the time consumption and demand much more training resources, which is unacceptable in real-world applications. In this section, we report efficiency results of DriveGPT4-V2 with LLaVA-0.5B, TinyLLaVA-1.5B and LLaVA-7B LLMs. The results are shown in Tab. 3. It is found that merely scaling up the LLM model does not gain corresponding performance enhancement, but slows down the whole system more than 10 times. Therefore, DriveGPT4-V2 mainly considers 0.5B LLMs as the planner. + +# 4.6. Ablation Studies and Discussion + +We conducted multiple ablation studies to justify the design of DriveGPT4-V2. The results are presented in Tab. 4, 5 and 6. Due to huge time consumption, DriveGPT4-V2 in ablation studies is not trained with DAgger data by default. + +LLM visual pretraining. DriveGPT4-V2 relies on LLaVA [30, 64] as the planner, which is pretrained on a huge amount of visual understanding data. The visual pretraining benefits DriveGPT4-V2 on the autonomous driving task with multimodal input. Directly training DriveGPT4-V2 on scratch (i.e., LLM for NLP tasks without multimodal ability) slows down convergence and degrades the final results. + +Visual tokenizer. The input camera images are processed by the proposed visual tokenizer. Inspired by LLaVA 1.5 [30], the features of images are extracted separately by the visual encoder and then concatenated as the input. This design ensures a sufficient perception range without compromising image detail. If the input is a single image with a large scope, it might lead to information loss during preprocessing steps (e.g., resize or crop); if the input is a small square image, the vehicle cannot have a sufficient perception range for effective planning. + +Waypoints and route points. Although DriveGPT4-V2 does not rely on waypoints and route points for generating low-level control signals, they provide valuable supervision during training. Predicting waypoints and route points helps the model perceive the target route and anticipate future actions. Removing them negatively impacts performance. + +Expert on-policy supervision. The expert model, benefiting from privileged information, provides robust on-policy supervision to DriveGPT4-V2. Without expert guidance, DriveGPT4-V2 relies solely on behavior cloning instead of DAGger imitation learning, which can lead to error accumulation and degraded performance. + +PID controllers. Previous works often generated low-level control signals based on waypoints or route points [12, 21] by calculating target speed and angle from these predictions. However, even minor prediction errors in waypoints or route points can lead to substantial errors in speed and angle. As shown in Tab. 5, using waypoints or route points for PID control can make control signals noisy and unstable, potentially driving the vehicle into exceptional states. Therefore it presents inferior results. + +Decision heads. During inference, DriveGPT4-V2 generates four tokens in an auto-regressive manner, mapped to vehicle decision variables by four decision heads. Predictions of target speed and angle involve one-to-one mappings (i.e., one token for one number), while waypoint and route point predictions involve one-to-many mappings (e.g., waypoints contain 8 x-y points, thus each token is mapped to 16 numbers). Another design would involve predicting each waypoint and route point one at a time, requiring 8 tokens for waypoints and 10 tokens for route points. Such design + +Table 4. Ablation studies of DriveGPT4-V2. "WP" and "RP" represent waypoints and route points, respectively. + +
DSRCIS
Baseline47780.60
+ LLM Visual Pretraining56870.64
+ Visual Tokenizer60880.68
+ WP&RP63900.70
+ Expert Supervision70910.77
+ +Table 5. Ablation studies on PID controllers. "WP" indicates utilizing predicted waypoints for PID control; while "TS&RP" means PID control by predicted target speed and route points. + +
PID ControllerDSRCIS
WP53850.62
TS & RP59880.67
DriveGPT4-V263900.70
+ +Table 6. Ablation studies on decision heads. "Additional tokens" indicates using more output tokens for prediction. + +
DSRCISFPS
Additional tokens64910.701.4
DriveGPT4-V263900.708.1
+ +does not significantly improve performance and severely affects efficiency. Using text to represent output numbers would require approximately 160 tokens, making it impractical for real-time autonomous driving. The evaluation results are shown in Tab. 6 + +# 5. Conclusion and Future Work + +In this paper, we presented DriveGPT4-V2, a novel LLM-based framework for closed-loop, end-to-end autonomous driving. DriveGPT4-V2 processes camera images and vehicle states as input to generate low-level control signals for direct vehicle operation. Leveraging the extensive pretrained knowledge of MLLMs, DriveGPT4-V2 demonstrates the ability to navigate complex urban scenarios under diverse and challenging conditions. The model architecture is specifically optimized for precise numerical vehicle decision prediction. An additional expert LLM, which shares a structure similar to DriveGPT4-V2, has been trained to provide on-policy supervision. Experimental results highlight that DriveGPT4-V2 achieves state-of-the-art performance on the challenging CARLA Longest6 benchmark, outperforming all baselines. In the future, we aim to extend the capabilities of DriveGPT4-V2 for broader applications in closed-loop autonomous driving tasks, including video games and real-world deployment. + +Acknowledgment. This work is partially supported by the National Natural Science Foundation of China (No. 62422606, 62201484, 52221005). + +# References + +[1] Anthropic. Claude. https://www.anthropic.com, 2024. +[2] Jinze Bai, Shuai Bai, Yunfei Chu, Zeyu Cui, Kai Dang, Xiaodong Deng, Yang Fan, Wenbin Ge, Yu Han, Fei Huang, et al. Qwen technical report. arXiv:2309.16609, 2023. +[3] Xu Cao, Tong Zhou, Yunsheng Ma, Wenqian Ye, Can Cui, Kun Tang, Zhipeng Cao, Kaizhao Liang, Ziran Wang, James M Rehg, et al. Maplm: A real-world large-scale vision-language benchmark for map and traffic scene understanding. In CVPR, 2024. +[4] Dian Chen and Philipp Krahenbuhl. Learning from all vehicles. In CVPR, 2022. +[5] Dian Chen, Brady Zhou, Vladlen Koltun, and Philipp Krahenbuhl. Learning by cheating. In CoRL, 2020. +[6] Dian Chen, Vladlen Koltun, and Philipp Krahenbuhl. Learning to drive from a world on rails. In ICCV, 2021. +[7] Long Chen, Oleg Sinavski, Jan Hünermann, Alice Karnsund, Andrew James Willmott, Danny Birch, Daniel Maund, and Jamie Shotton. Driving with llms: Fusing object-level vector modality for explainable autonomous driving. In ICRA, 2024. +[8] Li Chen, Penghao Wu, Kashyap Chitta, Bernhard Jaeger, Andreas Geiger, and Hongyang Li. End-to-end autonomous driving: Challenges and frontiers. TPAMI, 2024. +[9] Shaoyu Chen, Bo Jiang, Hao Gao, Bencheng Liao, Qing Xu, Qian Zhang, Chang Huang, Wenyu Liu, and Xinggang Wang. Vadv2: End-to-end vectorized autonomous driving via probabilistic planning. arXiv:2402.13243, 2024. +[10] Yuan Chen, Zi-han Ding, Ziqin Wang, Yan Wang, Lijun Zhang, and Si Liu. Asynchronous large language model enhanced planner for autonomous driving. In ECCV, 2024. +[11] Wei-Lin Chiang, Zhuohan Li, Zi Lin, Ying Sheng, Zhang-hao Wu, Hao Zhang, Lianmin Zheng, Siyuan Zhuang, Yong-hao Zhuang, Joseph E Gonzalez, et al. Vicuna: An opensource chatbot impressing gpt-4 with $90\%$ * chatgpt quality. See https://vicuna.lmsys.org, 2023. +[12] Kashyap Chitta, Aditya Prakash, Bernhard Jaeger, Zehao Yu, Katrin Renz, and Andreas Geiger. Transfuser: Imitation with transformer-based sensor fusion for autonomous driving. TPAMI, 2022. +[13] Google DeepMind. Gemini ai. https://deepmind.com, 2024. +[14] Alexey Dosovitskiy, German Ros, Felipe Codevilla, Antonio Lopez, and Vladlen Koltun. Carla: An open urban driving simulator. In CoRL, 2017. +[15] Yuchao Feng, Zhen Feng, Wei Hua, and Yuxiang Sun. Multimodal-xad: Explainable autonomous driving based on multimodal environment descriptions. TITS, 2024. +[16] Jiawei Fu, Yanqing Shen, Zhiqiang Jian, Shitao Chen, Jing-min Xin, and Nanning Zheng. Interactionnet: Joint planning and prediction for autonomous driving with transformers. In IROS, 2023. +[17] Yihan Hu, Jiazhi Yang, Li Chen, Keyu Li, Chonghao Sima, Xizhou Zhu, Siqi Chai, Senyao Du, Tianwei Lin, Wenhai Wang, et al. Planning-oriented autonomous driving. In CVPR, 2023. + +[18] Bin Huang, Xin Wang, Hong Chen, Zihan Song, and Wenwu Zhu. Vtimellm: Empower llm to grasp video moments. In CVPR, 2024. +[19] Ahmed Hussein, Mohamed Medhat Gaber, Eyad Elyan, and Chrisina Jayne. Imitation learning: A survey of learning methods. ACM Computing Surveys, 2017. +[20] Yuichi Inoue, Yuki Yada, Kotaro Tanahashi, and Yu Yamaguchi. Nuscenes-mqa: Integrated evaluation of captions and qa for autonomous driving datasets using markup annotations. In WACV, 2024. +[21] Bernhard Jaeger, Kashyap Chitta, and Andreas Geiger. Hidden biases of end-to-end driving models. In ICCV, 2023. +[22] Xiaosong Jia, Penghao Wu, Li Chen, Jiangwei Xie, Conghui He, Junchi Yan, and Hongyang Li. Think twice before driving: Towards scalable decoders for end-to-end autonomous driving. In CVPR, 2023. +[23] Bo Jiang, Shaoyu Chen, Qing Xu, Benchcheng Liao, Jiajie Chen, Helong Zhou, Qian Zhang, Wenyu Liu, Chang Huang, and Xinggang Wang. Vad: Vectorized scene representation for efficient autonomous driving. In ICCV, 2023. +[24] Zhuoling Li, Liangliang Ren, Jinrong Yang, Yong Zhao, Xiaoyang Wu, Zhenhua Xu, Xiang Bai, and Hengshuang Zhao. Virt: Vision instructed transformer for robotic manipulation. arXiv:2410.07169, 2024. +[25] Zhuoling Li, Xiaogang Xu, Zhenhua Xu, SerNam Lim, and Hengshuang Zhao. Larm: Large auto-regressive model for long-horizon embodied intelligence. arXiv:2405.17424, 2024. +[26] Bin Lin, Yang Ye, Bin Zhu, Jiaxi Cui, Munan Ning, Peng Jin, and Li Yuan. Video-llava: Learning united visual representation by alignment before projection. arXiv:2311.10122, 2023. +[27] Yuanfei Lin, Chenran Li, Mingyu Ding, Masayoshi Tomizuka, Wei Zhan, and Matthias Althoff. Drplanner: Diagnosis and repair of motion planners for automated vehicles using large language models. RA-L, 2024. +[28] Haotian Liu, Chunyuan Li, Yuheng Li, and Yong Jae Lee. Improved baselines with visual instruction tuning. In arXiv:2310.03744, 2023. +[29] Haotian Liu, Chunyuan Li, Qingyang Wu, and Yong Jae Lee. Visual instruction tuning. In NeurIPS, 2023. +[30] Haotian Liu, Chunyuan Li, Yuheng Li, Bo Li, Yuanhan Zhang, Sheng Shen, and Yong Jae Lee. Llava-last: Improved reasoning,OCR, and world knowledge. https://llava-v1.github.io/blog/2024-01-30-llava-last, 2024. +[31] Tianyu Liu, Qing hai Liao, Lu Gan, Fulong Ma, Jie Cheng, Xupeng Xie, Zhe Wang, Yingbing Chen, Yilong Zhu, Shuyang Zhang, et al. The role of the hercules autonomous vehicle during the COVID-19 pandemic: An autonomous logistic vehicle for contactless goods transportation. IEEE Robotics & Automation Magazine, 2021. +[32] Jiageng Mao, Yuxi Qian, Junjie Ye, Hang Zhao, and Yue Wang. Gpt-driver: Learning to drive with gpt. In NeurIPS Workshop, 2023. +[33] Maxim Mnyakin. Challenges and opportunities of integrating autonomous vehicles into urban retail delivery services. Reviews of Contemporary Business Analytics, 2023. + +[34] OpenAI. Chatgpt. https://chat.openai.com, 2024. +[35] Takayuki Osa, Joni Pajarinen, Gerhard Neumann, J Andrew Bagnell, Pieter Abbeel, Jan Peters, et al. An algorithmic perspective on imitation learning. Foundations and Trends® in Robotics, 2018. +[36] SungYeon Park, MinJae Lee, JiHyuk Kang, Hahyeon Choi, Yoonah Park, Juhwan Cho, Adam Lee, and DongKyu Kim. Vlaad: Vision and language assistant for autonomous driving. In WACV, 2024. +[37] Alec Radford, Jong Wook Kim, Chris Hallacy, Aditya Ramesh, Gabriel Goh, Sandhini Agarwal, Girish Sastry, Amanda Askell, Pamela Mishkin, Jack Clark, et al. Learning transferable visual models from natural language supervision. In ICML, 2021. +[38] Katrin Renz, Kashyap Chitta, Otniel-Bogdan Mercea, A Koepke, Zeynep Akata, and Andreas Geiger. Plant: Explainable planning transformers via object-level representations. In CoRL, 2022. +[39] Katrin Renz, Long Chen, Ana-Maria Marcu, Jan Hunermann, Benoit Hanotte, Alice Karnsund, Jamie Shotton, Elahe Arani, and Oleg Sinavski. Carllava: Vision language models for camera-only closed-loop driving. arXiv:2406.10165, 2024. +[40] Stephane Ross and J. Andrew Bagnell. Reinforcement and imitation learning via interactive no-regret learning. In arXiv:1406.5979, 2014. +[41] Stéphane Ross, Geoffrey Gordon, and Drew Bagnell. A reduction of imitation learning and structured prediction to no-regret online learning. In AISTATS, 2011. +[42] Hao Shao, Letian Wang, Ruobing Chen, Hongsheng Li, and Yu Liu. Safety-enhanced autonomous driving using interpretable sensor fusion transformer. In CoRL, 2023. +[43] Hao Shao, Yuxuan Hu, Letian Wang, Guanglu Song, Steven L Waslander, Yu Liu, and Hongsheng Li. Lmdrive: Closed-loop end-to-end driving with large language models. In CVPR, 2024. +[44] Chonghao Sima, Katrin Renz, Kashyap Chitta, Li Chen, Hanxue Zhang, Chengen Xie, Jens Beiwenger, Ping Luo, Andreas Geiger, and Hongyang Li. Drivelm: Driving with graph visual question answering. In ECCV, 2024. +[45] Robert Soler, Alae Moudni, Gabriel Roskowski, Xinrui Yu, Mikhail Gormov, and Jafar Saniee. Autonomous patrol and threat detection through integrated mapping and computer vision. In IEEE International Conference on Electro Information Technology, 2024. +[46] Chan Hee Song, Jiaman Wu, Clayton Washington, Brian M Sadler, Wei-Lun Chao, and Yu Su. Llm-planner: Few-shot grounded planning for embodied agents with large language models. In ICCV, 2023. +[47] Yafu Tian, Alexander Carballo, Ruifeng Li, Simon Thompson, and Kazuya Takeda. Rsg-search plus: An advanced traffic scene retrieval methods based on road scene graph. In IV, 2024. +[48] Faraz Torabi, Garrett Warnell, and Peter Stone. Behavioral cloning from observation. In *IJCAI*, 2018. +[49] Hugo Touvron, Thibaut Lavril, Gautier Izacard, Xavier Martinet, Marie-Anne Lachaux, Timothée Lacroix, Baptiste + +Rozière, Naman Goyal, Eric Hambro, Faisal Azhar, et al. Llama: Open and efficient foundation language models. arXiv:2302.13971, 2023. +[50] Yuxi Wei, Zi Wang, Yifan Lu, Chenxin Xu, Changxing Liu, Hao Zhao, Siheng Chen, and Yanfeng Wang. Editable scene simulation for autonomous driving via collaborative llm-agents. In CVPR, 2024. +[51] Licheng Wen, Daocheng Fu, Xin Li, Xinyu Cai, Tao Ma, Pinlong Cai, Min Dou, Botian Shi, Liang He, and Yu Qiao. Dilu: A knowledge-driven approach to autonomous driving with large language models. In ICLR, 2024. +[52] Zhenhua Xu, Yuxuan Liu, Lu Gan, Yuxiang Sun, Ming Liu, and Lujia Wang. Ringdet: Road network graph detection by transformer in aerial images. TGRS, 2022. +[53] Zhenhua Xu, Yuxuan Liu, Yuxiang Sun, Ming Liu, and Lujia Wang. Rngdet++: Road network graph detection by transformer with instance segmentation and multi-scale features enhancement. RA-L, 2023. +[54] Zhenhua Xu, Yujia Zhang, Enze Xie, Zhen Zhao, Yong Guo, Kwan-Yee K Wong, Zhenguo Li, and Hengshuang Zhao. Drivegpt4: Interpretable end-to-end autonomous driving via large language model. RA-L, 2024. +[55] Yijun Yang, Tianyi Zhou, Kanxue Li, Dapeng Tao, Lusong Li, Li Shen, Xiaodong He, Jing Jiang, and Yuhui Shi. Embodied multi-modal agent trained by an llm from a parallel textworld. In CVPR, 2024. +[56] Ekim Yurtsever, Jacob Lambert, Alexander Carballo, and Kazuya Takeda. A survey of autonomous driving: Common practices and emerging technologies. IEEE Access. +[57] Xiaohua Zhai, Basil Mustafa, Alexander Kolesnikov, and Lucas Beyer. Sigmoid loss for language image pre-training. In ICCV, 2023. +[58] Jimuyang Zhang, Zanming Huang, and Eshed Ohn-Bar. Coaching a teachable student. In CVPR, 2023. +[59] Jimuyang Zhang, Zanming Huang, Arjit Ray, and Eshed Ohn-Bar. Feedback-guided autonomous driving. In CVPR, 2024. +[60] Jiawei Zhang, Chejian Xu, and Bo Li. Chatscene: Knowledge-enabled safety-critical scenario generation for autonomous vehicles. In CVPR, 2024. +[61] Peiyuan Zhang, Guangtao Zeng, Tianduo Wang, and Wei Lu. Tinyllama: An open-source small language model. arXiv:2401.02385, 2024. +[62] Xiang Zhang, Haojie Sun, Xiaoyang Pei, Linghui Guan, and Zihao Wang. Evolution of technology investment and development of robotaxi services. Transportation Research Part E: Logistics and Transportation Review, 2024. +[63] Zhejun Zhang, Alexander Liniger, Dengxin Dai, Fisher Yu, and Luc Van Gool. End-to-end urban driving by imitating a reinforcement learning coach. In ICCV, 2021. +[64] Baichuan Zhou, Ying Hu, Xi Weng, Junlong Jia, Jie Luo, Xien Liu, Ji Wu, and Lei Huang. Tinyllava: A framework of small-scale large multimodal models. arXiv:2402.14289, 2024. +[65] Deyao Zhu, Jun Chen, Xiaoqian Shen, Xiang Li, and Mohamed Elhoseiny. Minigpt-4: Enhancing vision-language understanding with advanced large language models. In ICLR, 2024. \ No newline at end of file diff --git a/2025/DriveGPT4-V2_ Harnessing Large Language Model Capabilities for Enhanced Closed-Loop Autonomous Driving/images.zip b/2025/DriveGPT4-V2_ Harnessing Large Language Model Capabilities for Enhanced Closed-Loop Autonomous Driving/images.zip new file mode 100644 index 0000000000000000000000000000000000000000..be65d622593109a5f39a5251d538657b6b7206ec --- /dev/null +++ b/2025/DriveGPT4-V2_ Harnessing Large Language Model Capabilities for Enhanced Closed-Loop Autonomous Driving/images.zip @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:8cc4b29a7e11b37be678545168abccac4256ae49823be4596ddf72ede37e168d +size 342585 diff --git a/2025/DriveGPT4-V2_ Harnessing Large Language Model Capabilities for Enhanced Closed-Loop Autonomous Driving/layout.json b/2025/DriveGPT4-V2_ Harnessing Large Language Model Capabilities for Enhanced Closed-Loop Autonomous Driving/layout.json new file mode 100644 index 0000000000000000000000000000000000000000..53652039efcdaa622c913aa4f6d69468c613db09 --- /dev/null +++ b/2025/DriveGPT4-V2_ Harnessing Large Language Model Capabilities for Enhanced Closed-Loop Autonomous Driving/layout.json @@ -0,0 +1,7545 @@ +{ + "pdf_info": [ + { + "para_blocks": [ + { + "bbox": [ + 95, + 102, + 516, + 140 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 95, + 102, + 516, + 140 + ], + "spans": [ + { + "bbox": [ + 95, + 102, + 516, + 140 + ], + "type": "text", + "content": "DriveGPT4-V2: Harnessing Large Language Model Capabilities for Enhanced Closed-Loop Autonomous Driving" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 133, + 160, + 477, + 204 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 160, + 477, + 204 + ], + "spans": [ + { + "bbox": [ + 133, + 160, + 477, + 204 + ], + "type": "text", + "content": "Zhenhua Xu" + }, + { + "bbox": [ + 133, + 160, + 477, + 204 + ], + "type": "inline_equation", + "content": "^{1,2}" + }, + { + "bbox": [ + 133, + 160, + 477, + 204 + ], + "type": "text", + "content": " Yan Bai" + }, + { + "bbox": [ + 133, + 160, + 477, + 204 + ], + "type": "inline_equation", + "content": "^{3}" + }, + { + "bbox": [ + 133, + 160, + 477, + 204 + ], + "type": "text", + "content": " Yujia Zhang" + }, + { + "bbox": [ + 133, + 160, + 477, + 204 + ], + "type": "inline_equation", + "content": "^{1}" + }, + { + "bbox": [ + 133, + 160, + 477, + 204 + ], + "type": "text", + "content": " Zhuoling Li" + }, + { + "bbox": [ + 133, + 160, + 477, + 204 + ], + "type": "inline_equation", + "content": "^{1}" + }, + { + "bbox": [ + 133, + 160, + 477, + 204 + ], + "type": "text", + "content": " Fei Xia" + }, + { + "bbox": [ + 133, + 160, + 477, + 204 + ], + "type": "inline_equation", + "content": "^{3}" + }, + { + "bbox": [ + 133, + 160, + 477, + 204 + ], + "type": "text", + "content": " Kwan-Yee K. Wong" + }, + { + "bbox": [ + 133, + 160, + 477, + 204 + ], + "type": "inline_equation", + "content": "^{1}" + }, + { + "bbox": [ + 133, + 160, + 477, + 204 + ], + "type": "text", + "content": " Jianqiang Wang" + }, + { + "bbox": [ + 133, + 160, + 477, + 204 + ], + "type": "inline_equation", + "content": "^{2}" + }, + { + "bbox": [ + 133, + 160, + 477, + 204 + ], + "type": "text", + "content": " Hengshuang Zhao" + }, + { + "bbox": [ + 133, + 160, + 477, + 204 + ], + "type": "inline_equation", + "content": "^{1*}" + }, + { + "bbox": [ + 133, + 160, + 477, + 204 + ], + "type": "text", + "content": " The University of Hong Kong Tsinghua University Meituan" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 151, + 231, + 200, + 243 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 151, + 231, + 200, + 243 + ], + "spans": [ + { + "bbox": [ + 151, + 231, + 200, + 243 + ], + "type": "text", + "content": "Abstract" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 55, + 256, + 296, + 533 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 256, + 296, + 533 + ], + "spans": [ + { + "bbox": [ + 55, + 256, + 296, + 533 + ], + "type": "text", + "content": "Multimodal large language models (MLLMs) possess the ability to comprehend visual images or videos, and show impressive reasoning ability thanks to the vast amounts of pretrained knowledge, making them highly suitable for autonomous driving applications. Unlike the previous work, DriveGPT4-V1, which focused on open-loop tasks, this study explores the capabilities of LLMs in enhancing closed-loop autonomous driving. DriveGPT4-V2 processes camera images and vehicle states as input to generate low-level control signals for end-to-end vehicle operation. A multi-view visual tokenizer (MV-VT) is employed enabling DriveGPT4-V2 to perceive the environment with an extensive range while maintaining critical details. The model architecture has been refined to improve decision prediction and inference speed. To further enhance the performance, an additional expert LLM is trained for online imitation learning. The expert LLM, sharing a similar structure with DriveGPT4-V2, can access privileged information about surrounding objects for more robust and reliable predictions. Experimental results show that DriveGPT4-V2 outperforms all baselines on the challenging CARLA Longest6 benchmark. The code and data of DriveGPT4-V2 will be publicly available." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 55, + 555, + 135, + 567 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 555, + 135, + 567 + ], + "spans": [ + { + "bbox": [ + 55, + 555, + 135, + 567 + ], + "type": "text", + "content": "1. Introduction" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 55, + 574, + 296, + 696 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 574, + 296, + 696 + ], + "spans": [ + { + "bbox": [ + 55, + 574, + 296, + 696 + ], + "type": "text", + "content": "Autonomous driving has seen rapid advancements over the past decade in both academia and industry [31, 56]. Today, autonomous vehicles are deployed commercially in diverse applications such as robotaxi services [62], goods delivery [33], and security patrols [45]. Additionally, assisted driving systems are increasingly integrated into consumer vehicles, enhancing safety and convenience. However, the majority of current autonomous driving systems still rely heavily on rule-based decision-making modules, where planning and control are governed by predefined algorithms and" + } + ] + } + ], + "index": 7 + }, + { + "type": "image", + "bbox": [ + 316, + 230, + 554, + 401 + ], + "blocks": [ + { + "bbox": [ + 316, + 230, + 554, + 401 + ], + "lines": [ + { + "bbox": [ + 316, + 230, + 554, + 401 + ], + "spans": [ + { + "bbox": [ + 316, + 230, + 554, + 401 + ], + "type": "image", + "image_path": "e2c2dfe1095507cf1537101d0eb554fb6c817cee0ee39103ff335f5c6fae8a11.jpg" + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 313, + 409, + 555, + 488 + ], + "lines": [ + { + "bbox": [ + 313, + 409, + 555, + 488 + ], + "spans": [ + { + "bbox": [ + 313, + 409, + 555, + 488 + ], + "type": "text", + "content": "Figure 1. DriveGPT4-V2 for closed-loop autonomous driving. Taken as input multi-view camera images and vehicle state information, DriveGPT4-V2 predicts high-level vehicle decisions and converts them to low-level vehicle control signals in an end-to-end manner. DriveGPT4-V2 presents outstanding effectiveness and efficiency, serving as a reliable baseline method for future research on autonomous driving with LLMs." + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "image_caption" + } + ], + "index": 8 + }, + { + "bbox": [ + 313, + 509, + 555, + 544 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 509, + 555, + 544 + ], + "spans": [ + { + "bbox": [ + 313, + 509, + 555, + 544 + ], + "type": "text", + "content": "rules. This reliance restricts the ability of these systems to operate effectively in varied, real-world scenarios, limiting the broader advancement of autonomous driving." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 313, + 545, + 556, + 665 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 545, + 556, + 665 + ], + "spans": [ + { + "bbox": [ + 313, + 545, + 556, + 665 + ], + "type": "text", + "content": "In response, recent research has shifted toward end-to-end autonomous driving systems based on learning-based methods [8, 9, 12, 17, 21, 23, 45]. These methods enable more streamlined and intelligent system designs by directly learning control from input sensor data, without the need for complicated intermediate modules. From a learning theory perspective, end-to-end autonomous driving can optimize the entire system on the final outputs, rather than through the isolated optimization of individual modules, which potentially improves overall performance." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 313, + 665, + 556, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 665, + 556, + 715 + ], + "spans": [ + { + "bbox": [ + 313, + 665, + 556, + 715 + ], + "type": "text", + "content": "The rise of large language models (LLMs) has demonstrated their impressive ability to perform natural language processing (NLP) tasks [1, 2, 13, 34, 49]. Commercial LLMs, such as ChatGPT [34] and Claude3 [1], are already" + } + ] + } + ], + "index": 12 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "spans": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "type": "text", + "content": "CVF" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 145, + 0, + 496, + 37 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 145, + 0, + 496, + 37 + ], + "spans": [ + { + "bbox": [ + 145, + 0, + 496, + 37 + ], + "type": "text", + "content": "This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 67, + 703, + 144, + 712 + ], + "type": "page_footnote", + "angle": 0, + "lines": [ + { + "bbox": [ + 67, + 703, + 144, + 712 + ], + "spans": [ + { + "bbox": [ + 67, + 703, + 144, + 712 + ], + "type": "text", + "content": "*Corresponding author." + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "text", + "content": "17261" + } + ] + } + ], + "index": 14 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 0 + }, + { + "para_blocks": [ + { + "bbox": [ + 55, + 72, + 294, + 275 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 72, + 294, + 275 + ], + "spans": [ + { + "bbox": [ + 55, + 72, + 294, + 275 + ], + "type": "text", + "content": "widely used in daily life. Motivated by their success in NLP tasks, LLMs are being extended to multimodal domains, enabling more generalized tasks in fields like image analysis [28-30], video comprehension [18, 26], and embodied AI [24, 25, 46, 55]. Given their versatility, multimodal LLMs have been applied to autonomous driving for tasks such as interpretability [15, 36, 44, 54], vehicle control [10, 27, 27], and risk assessment [3, 60]. While multimodal LLMs have demonstrated promising vehicle control performance due to their extensive pretrained knowledge, previous research has primarily concentrated on open-loop settings, which are not implemented for real-time vehicle control. Open-loop systems lack feedback mechanisms, making them less adaptable to changes and disturbances, limiting their applicability in real-world applications. Thus, multimodal LLMs need to be designed and evaluated specifically for closed-loop autonomous driving scenarios [27, 39, 43, 51]." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 55, + 277, + 294, + 361 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 277, + 294, + 361 + ], + "spans": [ + { + "bbox": [ + 55, + 277, + 294, + 361 + ], + "type": "text", + "content": "Behavior cloning [35, 48], the most commonly employed imitation learning algorithm, is used for most end-to-end autonomous driving works [44, 54]. However, models trained only by behavior cloning can encounter drifting issues in closed-loop tasks, where errors accumulate, potentially leading the vehicle into unsafe situations, such as deviating from the path, colliding, or getting stuck." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 55, + 364, + 295, + 579 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 364, + 295, + 579 + ], + "spans": [ + { + "bbox": [ + 55, + 364, + 295, + 579 + ], + "type": "text", + "content": "To address the aforementioned issue, we incorporate online (or on-policy) imitation learning [5, 40, 41] into closed-loop autonomous driving. Leveraging the reasoning capabilities and pretrained knowledge of LLMs, we propose using a multimodal LLM as a fully automatic expert for online imitation learning without human efforts. Let " + }, + { + "bbox": [ + 55, + 364, + 295, + 579 + ], + "type": "inline_equation", + "content": "M_E" + }, + { + "bbox": [ + 55, + 364, + 295, + 579 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 55, + 364, + 295, + 579 + ], + "type": "inline_equation", + "content": "M_A" + }, + { + "bbox": [ + 55, + 364, + 295, + 579 + ], + "type": "text", + "content": " denote the model of expert and agent, respectively. " + }, + { + "bbox": [ + 55, + 364, + 295, + 579 + ], + "type": "inline_equation", + "content": "M_E" + }, + { + "bbox": [ + 55, + 364, + 295, + 579 + ], + "type": "text", + "content": " can access privileged information about nearby objects while " + }, + { + "bbox": [ + 55, + 364, + 295, + 579 + ], + "type": "inline_equation", + "content": "M_A" + }, + { + "bbox": [ + 55, + 364, + 295, + 579 + ], + "type": "text", + "content": " cannot, thus presenting a more powerful and reliable performance. After being trained by behavior cloning, " + }, + { + "bbox": [ + 55, + 364, + 295, + 579 + ], + "type": "inline_equation", + "content": "M_A" + }, + { + "bbox": [ + 55, + 364, + 295, + 579 + ], + "type": "text", + "content": " is deployed on the training environment for data aggregation. " + }, + { + "bbox": [ + 55, + 364, + 295, + 579 + ], + "type": "inline_equation", + "content": "M_A" + }, + { + "bbox": [ + 55, + 364, + 295, + 579 + ], + "type": "text", + "content": " might make incorrect predictions that lead the vehicle into exception states. " + }, + { + "bbox": [ + 55, + 364, + 295, + 579 + ], + "type": "inline_equation", + "content": "M_E" + }, + { + "bbox": [ + 55, + 364, + 295, + 579 + ], + "type": "text", + "content": " can provide " + }, + { + "bbox": [ + 55, + 364, + 295, + 579 + ], + "type": "inline_equation", + "content": "M_A" + }, + { + "bbox": [ + 55, + 364, + 295, + 579 + ], + "type": "text", + "content": " with on-policy supervision and take over the vehicle if the error made by " + }, + { + "bbox": [ + 55, + 364, + 295, + 579 + ], + "type": "inline_equation", + "content": "M_A" + }, + { + "bbox": [ + 55, + 364, + 295, + 579 + ], + "type": "text", + "content": " is large enough. When " + }, + { + "bbox": [ + 55, + 364, + 295, + 579 + ], + "type": "inline_equation", + "content": "M_E" + }, + { + "bbox": [ + 55, + 364, + 295, + 579 + ], + "type": "text", + "content": " takes over the vehicle, it collects one data sample at the moment and aggregates it into the dataset. The aggregated dataset is leveraged to further train the agent model to handle errors." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 55, + 582, + 295, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 582, + 295, + 713 + ], + "spans": [ + { + "bbox": [ + 55, + 582, + 295, + 713 + ], + "type": "text", + "content": "In this paper, we introduce DriveGPT4-V2, an end-to-end autonomous driving system. DriveGPT4-V2 extends DriveGPT4-V1 [54] to closed-loop tasks and integrates online imitation learning for enhanced robustness and intelligence. The \"4\" in the system's name signifies its multimodal capacity, inspired by the MiniGPT4 model [65]. DriveGPT4-V2 is powered by multimodal LLMs, enabling it to directly generate low-level vehicle control signals (i.e., throttle, brake and steer) based on multimodal input data (i.e., vehicle states and camera images). The structure of DriveGPT4-V2 is refined for our task, including a" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 313, + 72, + 553, + 228 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 72, + 553, + 228 + ], + "spans": [ + { + "bbox": [ + 313, + 72, + 553, + 228 + ], + "type": "text", + "content": "multi-view visual tokenizer (MV-VT) that possesses a large enough perception range without losing details, as well as output decision heads (DeciHeads) specially designed for numerical vehicle decision prediction. An additional multimodal LLM with access to privileged information (e.g., ground truth of nearby objects in text) serves as an expert for online imitation learning, which is inspired by the LBC framework [5]. DriveGPT4-V2 achieves state-of-the-art performance on the challenging CARLA Longest6 benchmark [12], which evaluates systems over 36 long routes demanding high levels of intelligence and reliability. The overview of DriveGPT4-V2 is visualized in Fig. 1. The contributions of this paper are:" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 314, + 228, + 553, + 443 + ], + "type": "list", + "angle": 0, + "index": 9, + "blocks": [ + { + "bbox": [ + 314, + 228, + 553, + 287 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 314, + 228, + 553, + 287 + ], + "spans": [ + { + "bbox": [ + 314, + 228, + 553, + 287 + ], + "type": "text", + "content": "- We extend DriveGPT4-V1, an open-loop autonomous driving system with multimodal LLMs, and propose DriveGPT4-V2 for end-to-end closed-loop autonomous driving. DriveGPT4-V2 can better harness the capability of LLMs and be deployed for continuous vehicle control." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 314, + 288, + 553, + 346 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 314, + 288, + 553, + 346 + ], + "spans": [ + { + "bbox": [ + 314, + 288, + 553, + 346 + ], + "type": "text", + "content": "- We introduce the use of LLMs with privileged information as an expert for online (on-policy) imitation learning, facilitating the generation of labeled data for handling errors, which significantly enhances the performance and robustness of closed-loop autonomous driving." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 314, + 348, + 553, + 407 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 314, + 348, + 553, + 407 + ], + "spans": [ + { + "bbox": [ + 314, + 348, + 553, + 407 + ], + "type": "text", + "content": "- We optimize the model structure of MLLMs, including a multi-view visual tokenizer (MV-VT) that possesses a large perception range and preserves details, as well as output decision heads (DeciHeads) specially designed for numerical vehicle decision prediction." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 314, + 407, + 553, + 443 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 314, + 407, + 553, + 443 + ], + "spans": [ + { + "bbox": [ + 314, + 407, + 553, + 443 + ], + "type": "text", + "content": "- DriveGPT4-V2 demonstrates superior performance on the CARLA Longest6 benchmark, substantially outperforming existing methods." + } + ] + } + ], + "index": 8 + } + ], + "sub_type": "text" + }, + { + "bbox": [ + 314, + 453, + 405, + 466 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 314, + 453, + 405, + 466 + ], + "spans": [ + { + "bbox": [ + 314, + 453, + 405, + 466 + ], + "type": "text", + "content": "2. Related Works" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 313, + 474, + 555, + 689 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 474, + 555, + 689 + ], + "spans": [ + { + "bbox": [ + 313, + 474, + 555, + 689 + ], + "type": "text", + "content": "Large language models for autonomous driving. Owing to the powerful reasoning capabilities and extensive pretrained knowledge, large language models (LLMs) have found wide application in autonomous driving tasks, such as vehicle decision-making [7, 10, 15, 27, 32, 39, 43, 54, 59], scene understanding [20, 36, 44, 54], risk analysis [3, 60], and simulation [47, 50]. GPT-Driver [32] represents the entire surrounding environment textually and employs a fine-tuned ChatGPT model to predict vehicle trajectories. DriveGPT4 [54] leverages multimodal LLMs for end-to-end vehicle action prediction and scene understanding. It processes 8-frame video clips and human user commands to predict vehicle control signals and action explanations. However, most of these approaches are designed for open-loop scenarios and present unsatisfactory performance in closed-loop tasks, limiting their applicability to real-world driving deployment. Additionally, many existing methods face efficiency challenges due to heavy model architectures." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 313, + 689, + 553, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 689, + 553, + 713 + ], + "spans": [ + { + "bbox": [ + 313, + 689, + 553, + 713 + ], + "type": "text", + "content": "Closed-loop end-to-end autonomous driving. Closed-loop end-to-end autonomous driving has recently emerged" + } + ] + } + ], + "index": 12 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "text", + "content": "17262" + } + ] + } + ], + "index": 13 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 1 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 60, + 73, + 550, + 345 + ], + "blocks": [ + { + "bbox": [ + 60, + 73, + 550, + 345 + ], + "lines": [ + { + "bbox": [ + 60, + 73, + 550, + 345 + ], + "spans": [ + { + "bbox": [ + 60, + 73, + 550, + 345 + ], + "type": "image", + "image_path": "955fc14f1df6defcabc113baf58f6b04fddc39a90f3b17b5024c9d5e203d8c8c.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 54, + 357, + 555, + 436 + ], + "lines": [ + { + "bbox": [ + 54, + 357, + 555, + 436 + ], + "spans": [ + { + "bbox": [ + 54, + 357, + 555, + 436 + ], + "type": "text", + "content": "Figure 2. DriveGPT4-V2 diagram. DriveGPT4-V2 takes multimodal input data to generate numerical control signals for end-to-end vehicle driving. The input includes multi-view images and vehicle state information. The images are transformed into the text domain via a multi-view vision tokenizer (MV-VT). The current speed of the vehicle and target point are tokenized by the LLM tokenizer. The LLM then outputs four tokens. Each token is used to predict one vehicle decision by an MLP decision head (DeciHead). These predicted decisions are subsequently converted to low-level commands via PID controllers to operate the vehicle. The LLM expert model, which shares a similar structure to DriveGPT4-V2, has access to privileged information about surroundings (shown in the purple module). The expert provides on-policy supervision to DriveGPT4-V2 to enhance closed-loop performance. This figure is best viewed in color." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "bbox": [ + 55, + 456, + 297, + 708 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 456, + 297, + 708 + ], + "spans": [ + { + "bbox": [ + 55, + 456, + 297, + 708 + ], + "type": "text", + "content": "as a prominent research focus, addressing the challenge of bridging the gap between academic research and real-world applications. Various approaches improved transformer architectures and training methods to enhance driving performance [4, 5, 12, 16, 21, 22, 42, 58, 63]. For instance, Transfuser++ [21] integrates multimodal sensor inputs, including camera imagery, LiDAR point clouds and vehicle dynamics, to predict speed and trajectory, achieving state-of-the-art performance across multiple benchmarks. Even though most early LLM-based methods have primarily targeted open-loop tasks, recent studies [27, 39, 43, 51] have attempted closed-loop approaches. But most of them are deployed in simple environments like Highway-env [51], instead of real-world scenarios. LMDrive [43], pioneering work on end-to-end closed-loop driving with multimodal LLMs, demonstrated the potential of LLMs in this domain. It is deployed in the CARLA simulator [14] and controls the vehicle like human drivers. However, it relies solely on behavior cloning and presents unsatisfactory driving performance. Online imitation learning has been highlighted as a potential way to enhance model robustness." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 313, + 456, + 556, + 708 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 456, + 556, + 708 + ], + "spans": [ + { + "bbox": [ + 313, + 456, + 556, + 708 + ], + "type": "text", + "content": "Imitation learning in autonomous driving. Imitation learning [19] aims to train an agent to mimic the expert behavior. Behavior cloning is the most widely used imitation learning algorithm in autonomous driving [4, 12, 21, 39, 43, 44, 52-54]. In this method, data is off-policy sampled from expert trajectories for model training to replicate expert actions. Although behavior cloning can yield reasonable performance, it suffers from error accumulation during closed-loop operation, leading the model to encounter unseen situations that may compromise safety. To mitigate this issue, online imitation learning [40, 41] has been employed in various works [4, 5]. The LBC framework [5] uses the DAgger algorithm [41], and trains an agent alongside an expert network with privileged information that achieves superior performance. The expert provides on-policy supervision to the agent and collects additional data samples for data aggregation, which can be used to further train the agent model to handle errors. Inspired by LBC, DriveGPT4-V2 harnesses the capabilities of LLMs and leverages a privileged LLM expert to provide on-policy supervision, facilitating improved driving performance and robustness." + } + ] + } + ], + "index": 3 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 318, + 758 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 318, + 758 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 318, + 758 + ], + "type": "text", + "content": "17263" + } + ] + } + ], + "index": 4 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 2 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 61, + 72, + 290, + 243 + ], + "blocks": [ + { + "bbox": [ + 61, + 72, + 290, + 243 + ], + "lines": [ + { + "bbox": [ + 61, + 72, + 290, + 243 + ], + "spans": [ + { + "bbox": [ + 61, + 72, + 290, + 243 + ], + "type": "image", + "image_path": "995738dfecd7412e6c9b2dddf492c5aeb9f0b263cf8538985befe2873bc33c90.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 55, + 255, + 296, + 310 + ], + "lines": [ + { + "bbox": [ + 55, + 255, + 296, + 310 + ], + "spans": [ + { + "bbox": [ + 55, + 255, + 296, + 310 + ], + "type": "text", + "content": "Figure 3. Multi-view visual tokenizer (MV-VT) structure. The input images consist of three front views. Each patch is processed through a visual encoder to extract features. Finally, a trained projection layer maps the downsampled feature into the text domain for further processing." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 313, + 110, + 555, + 187 + ], + "lines": [ + { + "bbox": [ + 313, + 110, + 555, + 187 + ], + "spans": [ + { + "bbox": [ + 313, + 110, + 555, + 187 + ], + "type": "text", + "content": "Figure 4. Visualization of waypoints and route points. The ego vehicle is represented by the green rectangle, and the red point denotes the target point. The grey line is the route that the vehicle should follow. (a) Waypoints (blue points) represent vehicle positions in a 4-second future. (b) Route points are evenly sampled from the global planned route in front. These two kinds of points can better supervise the training of DriveGPT4-V2." + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "bbox": [ + 55, + 331, + 149, + 342 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 331, + 149, + 342 + ], + "spans": [ + { + "bbox": [ + 55, + 331, + 149, + 342 + ], + "type": "text", + "content": "3. DriveGPT4-V2" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 54, + 351, + 295, + 543 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 54, + 351, + 295, + 543 + ], + "spans": [ + { + "bbox": [ + 54, + 351, + 295, + 543 + ], + "type": "text", + "content": "Building on DriveGPT4-V1 [54], a flexible multimodal LLM designed for open-loop autonomous driving, DriveGPT4-V2 has been specifically developed for closed-loop driving tasks. This paper employs LLMs both as planners and as expert models for online imitation learning. A multi-view visual tokenizer (MV-VT) is leveraged for effective feature extraction. Unlike standard LLMs used for NLP, the architecture of DriveGPT4-V2 is tailored for vehicle planning, with additional output decision heads (DeciHeads) to directly optimize numerical decision-making predictions. This design enhances model efficiency by reducing the token count required for inference. The expert model, which shares a similar architecture to DriveGPT4-V2, has access to privileged ground-truth information about the surroundings and is used solely for on-policy supervision during training, not for inference." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 55, + 552, + 143, + 564 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 552, + 143, + 564 + ], + "spans": [ + { + "bbox": [ + 55, + 552, + 143, + 564 + ], + "type": "text", + "content": "3.1. Model Design" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 55, + 569, + 295, + 629 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 569, + 295, + 629 + ], + "spans": [ + { + "bbox": [ + 55, + 569, + 295, + 629 + ], + "type": "text", + "content": "DriveGPT4-V2 takes multimodal input data and predicts high-level vehicle decisions for end-to-end autonomous driving. These predictions are subsequently converted into low-level control signals through PID controllers. The model structure is shown in Fig. 2." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 55, + 630, + 296, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 630, + 296, + 714 + ], + "spans": [ + { + "bbox": [ + 55, + 630, + 296, + 714 + ], + "type": "text", + "content": "Multimodal input. The input consists of two modalities: image and text. The image input is captured by three front-view cameras at a resolution of " + }, + { + "bbox": [ + 55, + 630, + 296, + 714 + ], + "type": "inline_equation", + "content": "384 \\times 384" + }, + { + "bbox": [ + 55, + 630, + 296, + 714 + ], + "type": "text", + "content": ", while the text input includes the vehicle's current speed and the target point. These three front-view images ensure coverage of both sides of the vehicle, reducing the risk of missing adjacent vehicles or potential collisions. The image data is" + } + ] + } + ], + "index": 6 + }, + { + "type": "image", + "bbox": [ + 315, + 76, + 555, + 83 + ], + "blocks": [ + { + "bbox": [ + 315, + 76, + 555, + 83 + ], + "lines": [ + { + "bbox": [ + 315, + 76, + 555, + 83 + ], + "spans": [ + { + "bbox": [ + 315, + 76, + 555, + 83 + ], + "type": "image", + "image_path": "c33d38c0cd5c96737f087a11329e34e686c48f64f9cfa5cb70a639771da57302.jpg" + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 349, + 90, + 397, + 100 + ], + "lines": [ + { + "bbox": [ + 349, + 90, + 397, + 100 + ], + "spans": [ + { + "bbox": [ + 349, + 90, + 397, + 100 + ], + "type": "text", + "content": "(a) Waypoints." + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 468, + 91, + 523, + 100 + ], + "lines": [ + { + "bbox": [ + 468, + 91, + 523, + 100 + ], + "spans": [ + { + "bbox": [ + 468, + 91, + 523, + 100 + ], + "type": "text", + "content": "(b) Route points." + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "image_caption" + } + ], + "index": 7 + }, + { + "bbox": [ + 313, + 209, + 555, + 243 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 209, + 555, + 243 + ], + "spans": [ + { + "bbox": [ + 313, + 209, + 555, + 243 + ], + "type": "text", + "content": "mapped into the text domain via a visual tokenizer, as implemented in LLaVA [30], while text inputs are tokenized with standard LLM tokenizers." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 313, + 245, + 554, + 399 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 245, + 554, + 399 + ], + "spans": [ + { + "bbox": [ + 313, + 245, + 554, + 399 + ], + "type": "text", + "content": "Multi-view visual tokenizer (MV-VT). The visual tokenizer is visualized in Fig. 3. It first employs pretrained visual encoders, such as CLIP [37] or SigLIP [57], to extract features from the input images. The size of each image is specifically optimized for compatibility with the SigLIP encoder (i.e., " + }, + { + "bbox": [ + 313, + 245, + 554, + 399 + ], + "type": "inline_equation", + "content": "384 \\times 384" + }, + { + "bbox": [ + 313, + 245, + 554, + 399 + ], + "type": "text", + "content": "). The three images are acquired from front-left, front, and front-right cameras respectively, ensuring comprehensive coverage of the vehicle's frontal and lateral views. In CARLA benchmarks, back-view images exhibit limited utilization frequency and demonstrate no measurable improvement in planning performance. To maintain computational efficiency, MV-VT deliberately excludes back-view inputs." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 313, + 401, + 554, + 496 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 401, + 554, + 496 + ], + "spans": [ + { + "bbox": [ + 313, + 401, + 554, + 496 + ], + "type": "text", + "content": "Output representations. DriveGPT4-V2 generates three low-level control signals to drive the vehicle, i.e., throttle, brake, and steering. However, direct prediction of these signals requires temporal consistency; otherwise, the vehicle may display unstable zigzag trajectories. To ensure stable control, DriveGPT4-V2 instead predicts high-level decision variables, which are then converted to low-level commands by PID controllers. The predictions of the model include:" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 313, + 497, + 553, + 604 + ], + "type": "list", + "angle": 0, + "index": 18, + "blocks": [ + { + "bbox": [ + 313, + 497, + 553, + 533 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 497, + 553, + 533 + ], + "spans": [ + { + "bbox": [ + 313, + 497, + 553, + 533 + ], + "type": "text", + "content": "- Target Speed: The desired speed for the vehicle. A speed PID controller adjusts longitudinal control based on the predicted target speed and current speed." + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 313, + 533, + 553, + 556 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 533, + 553, + 556 + ], + "spans": [ + { + "bbox": [ + 313, + 533, + 553, + 556 + ], + "type": "text", + "content": "- Target Angle: The direction in which the vehicle should steer. A steering PID controller uses it for lateral control." + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 313, + 557, + 553, + 581 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 557, + 553, + 581 + ], + "spans": [ + { + "bbox": [ + 313, + 557, + 553, + 581 + ], + "type": "text", + "content": "- Waypoints: Predicted x-y positions of the vehicle in the future. Waypoints consist of eight points." + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 313, + 581, + 553, + 604 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 581, + 553, + 604 + ], + "spans": [ + { + "bbox": [ + 313, + 581, + 553, + 604 + ], + "type": "text", + "content": "- Route Points: Points evenly sampled along the global route in front, enhancing the vehicle's perception ability." + } + ] + } + ], + "index": 17 + } + ], + "sub_type": "text" + }, + { + "bbox": [ + 313, + 605, + 523, + 616 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 605, + 523, + 616 + ], + "spans": [ + { + "bbox": [ + 313, + 605, + 523, + 616 + ], + "type": "text", + "content": "Waypoints and route points are visualized in Fig. 4." + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 313, + 617, + 555, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 617, + 555, + 713 + ], + "spans": [ + { + "bbox": [ + 313, + 617, + 555, + 713 + ], + "type": "text", + "content": "Decision heads (DeciHeads). For closed-loop autonomous driving, DriveGPT4-V2 employs dedicated output heads (MLPs) to predict numerical vehicle decisions. The original LLM's vocabulary output head is replaced with four decision heads, each predicting a decision variable (i.e., target speed, target angle, waypoints, and route points) based on one token. This setup allows DriveGPT4-V2 to produce only four tokens per prediction cycle, achieving a huge" + } + ] + } + ], + "index": 20 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "type": "text", + "content": "17264" + } + ] + } + ], + "index": 21 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 3 + }, + { + "para_blocks": [ + { + "bbox": [ + 55, + 72, + 294, + 132 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 72, + 294, + 132 + ], + "spans": [ + { + "bbox": [ + 55, + 72, + 294, + 132 + ], + "type": "text", + "content": "speed increase over traditional text-based outputs. For example, a float number with 3 digits (e.g., 1.538) needs 5 tokens for text output representation but only requires a single token for our design. DeciHead also enables direct optimization of decision variables using regression loss." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 55, + 133, + 295, + 275 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 133, + 295, + 275 + ], + "spans": [ + { + "bbox": [ + 55, + 133, + 295, + 275 + ], + "type": "text", + "content": "Expert model and online imitation learning. The expert model has a similar architecture to DriveGPT4-V2 but has access to privileged information. The privileged information consists of ground truth (GT) of nearby objects and hazard information provided by the CARLA simulator (e.g., traffic light and stop sign violation; vehicle or pedestrian collision). This privileged data allows the expert to better perceive the environment, so that it can focus on vehicle decisions and planning. The privileged information greatly boosts the model capacity, resulting in enhanced closed-loop performance. The expert model provides on-policy supervision to DriveGPT4-V2 during training." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 55, + 275, + 295, + 407 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 275, + 295, + 407 + ], + "spans": [ + { + "bbox": [ + 55, + 275, + 295, + 407 + ], + "type": "text", + "content": "When DriveGPT4-V2 runs on training scenarios and routes, the expert model receives identical input and generates concurrent predictions. If the discrepancy between DriveGPT4-V2's prediction and the expert's exceeds a threshold, the expert's prediction is used instead to control the vehicle. Such a situation is marked as an exception case. A data sample of the current moment is added to the training set, where the expert's output is used as the label. DriveGPT4-V2 is fine-tuned on this aggregated dataset under the expert's on-policy supervision for enhanced closed-loop autonomous driving performance." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 55, + 407, + 295, + 432 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 407, + 295, + 432 + ], + "spans": [ + { + "bbox": [ + 55, + 407, + 295, + 432 + ], + "type": "text", + "content": "The expert LLM is used exclusively for on-policy supervision during training and is not accessed during inference." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 55, + 438, + 119, + 451 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 438, + 119, + 451 + ], + "spans": [ + { + "bbox": [ + 55, + 438, + 119, + 451 + ], + "type": "text", + "content": "3.2. Training" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 55, + 456, + 295, + 540 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 456, + 295, + 540 + ], + "spans": [ + { + "bbox": [ + 55, + 456, + 295, + 540 + ], + "type": "text", + "content": "The training process of DriveGPT4-V2 and the expert LLM includes two stages. In the first stage, both models are trained on data collected by a rule-based autopilot using behavior cloning. In the second stage, DriveGPT4-V2 runs on the training scenarios and routes under the expert's on-policy supervision, and generates additional data for error correction. The training process is visualized in Fig. 5." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 55, + 546, + 194, + 558 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 546, + 194, + 558 + ], + "spans": [ + { + "bbox": [ + 55, + 546, + 194, + 558 + ], + "type": "text", + "content": "3.2.1. Stage 1: behavior cloning" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 55, + 561, + 295, + 668 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 561, + 295, + 668 + ], + "spans": [ + { + "bbox": [ + 55, + 561, + 295, + 668 + ], + "type": "text", + "content": "In the first stage, a rule-based autopilot [21] is deployed in CARLA to collect data. The autopilot agent records data at fixed time intervals, including camera images, current vehicle speed, privileged information on surrounding objects (e.g., vehicles and traffic lights), and a target point. Labels for the vehicle decision variables are calculated using rule-based algorithms. After data collection, both DriveGPT4-V2 and the expert LLM are trained on the dataset, with the expert having access to privileged information." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 55, + 674, + 269, + 687 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 674, + 269, + 687 + ], + "spans": [ + { + "bbox": [ + 55, + 674, + 269, + 687 + ], + "type": "text", + "content": "3.2.2. Stage 2: DAgger with on-policy supervision" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 55, + 689, + 295, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 689, + 295, + 714 + ], + "spans": [ + { + "bbox": [ + 55, + 689, + 295, + 714 + ], + "type": "text", + "content": "The expert, with access to privileged information, achieves notably higher performance by leveraging reliable priors on" + } + ] + } + ], + "index": 9 + }, + { + "type": "image", + "bbox": [ + 331, + 71, + 550, + 148 + ], + "blocks": [ + { + "bbox": [ + 331, + 71, + 550, + 148 + ], + "lines": [ + { + "bbox": [ + 331, + 71, + 550, + 148 + ], + "spans": [ + { + "bbox": [ + 331, + 71, + 550, + 148 + ], + "type": "image", + "image_path": "501a1f5a0dfb3947b86c41c55fe6c3d1a51e80bb06853d866e54b38248154e56.jpg" + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 391, + 152, + 489, + 161 + ], + "lines": [ + { + "bbox": [ + 391, + 152, + 489, + 161 + ], + "spans": [ + { + "bbox": [ + 391, + 152, + 489, + 161 + ], + "type": "text", + "content": "(a) Stage 1: Behavior cloning." + } + ] + } + ], + "index": 11, + "angle": 0, + "type": "image_caption" + } + ], + "index": 10 + }, + { + "type": "image", + "bbox": [ + 317, + 163, + 563, + 269 + ], + "blocks": [ + { + "bbox": [ + 317, + 163, + 563, + 269 + ], + "lines": [ + { + "bbox": [ + 317, + 163, + 563, + 269 + ], + "spans": [ + { + "bbox": [ + 317, + 163, + 563, + 269 + ], + "type": "image", + "image_path": "d7e4f976a0eeeff15f2c7c35142437c32eb23e636d668a8e8889e42777e2c61c.jpg" + } + ] + } + ], + "index": 12, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 364, + 274, + 515, + 284 + ], + "lines": [ + { + "bbox": [ + 364, + 274, + 515, + 284 + ], + "spans": [ + { + "bbox": [ + 364, + 274, + 515, + 284 + ], + "type": "text", + "content": "(b) Stage 2: DAgger online imitation learning." + } + ] + } + ], + "index": 13, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 313, + 293, + 555, + 382 + ], + "lines": [ + { + "bbox": [ + 313, + 293, + 555, + 382 + ], + "spans": [ + { + "bbox": [ + 313, + 293, + 555, + 382 + ], + "type": "text", + "content": "Figure 5. Diagram of the two-stage training process. (a) In the first stage, both DriveGPT4-V2 and the expert LLM are trained on data collected by a rule-based autopilot. (b) In the second stage, DriveGPT4-V2 runs on the training scenarios and routes. When the discrepancy between DriveGPT4-V2's predictions and those of the expert exceeds a predefined threshold, the expert's predictions are used to control the vehicle. Data from these cases is then added to the dataset for data aggregation." + } + ] + } + ], + "index": 14, + "angle": 0, + "type": "image_caption" + } + ], + "index": 12 + }, + { + "bbox": [ + 313, + 402, + 555, + 629 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 402, + 555, + 629 + ], + "spans": [ + { + "bbox": [ + 313, + 402, + 555, + 629 + ], + "type": "text", + "content": "surrounding objects, allowing it to handle challenging scenarios effectively. DriveGPT4-V2, trained only by behavior cloning, may struggle with error accumulation and exceptional situations (e.g., vehicle collisions or getting stuck). Inspired by LBC [5], the expert LLM is used as a teacher for on-policy supervision. On-policy supervision is a training paradigm in imitation learning, where supervision or feedback is provided to an agent while it is acting according to its current policy. This approach contrasts with off-policy methods, where supervision comes from data collected by a different policy or behavior. During the operation of DriveGPT4-V2, the expert and DriveGPT4-V2 receive identical input, and their outputs are compared. If the discrepancy in predicted decisions exceeds a threshold, the expert takes control to avoid exceptions. One data sample is collected at that moment and aggregated into the dataset. The output of the expert is used as the training label. After traversing the training routes under expert supervision, DriveGPT4-V2 is fine-tuned on the aggregated dataset." + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 314, + 636, + 373, + 647 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 314, + 636, + 373, + 647 + ], + "spans": [ + { + "bbox": [ + 314, + 636, + 373, + 647 + ], + "type": "text", + "content": "3.3. Dataset" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 313, + 653, + 554, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 653, + 554, + 713 + ], + "spans": [ + { + "bbox": [ + 313, + 653, + 554, + 713 + ], + "type": "text", + "content": "Following Transfuser++ [21], in stage 1, the rule-based autopilot agent is used to collect data samples in various CARLA simulator towns under varying weather, lighting, and daytime conditions. Vehicle speed and position are estimated by Kalman filters based on IMU data. Each sam" + } + ] + } + ], + "index": 17 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "text", + "content": "17265" + } + ] + } + ], + "index": 18 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 4 + }, + { + "para_blocks": [ + { + "bbox": [ + 54, + 72, + 294, + 144 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 54, + 72, + 294, + 144 + ], + "spans": [ + { + "bbox": [ + 54, + 72, + 294, + 144 + ], + "type": "text", + "content": "ple records objects within a fixed radius, including vehicles, pedestrians, traffic lights, and stop signs. Target speeds and angles are calculated by the autopilot's PID controller. The ground-truth waypoints contain 8 points and record the vehicle position in a 4-second future. Route points have 10 points sampled along the planned route in front." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 55, + 144, + 295, + 251 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 144, + 295, + 251 + ], + "spans": [ + { + "bbox": [ + 55, + 144, + 295, + 251 + ], + "type": "text", + "content": "In stage 2, DriveGPT4-V2 runs on selected training routes for data aggregation. Simple routes are not considered in this stage for data collection efficiency. If the difference between the predicted decisions of DriveGPT4-V2 and the expert exceeds a threshold, the expert takes control and records the data sample. DriveGPT4-V2 is retrained on the aggregated dataset for better performance. Under common circumstances, DriveGPT4-V2 achieves satisfactory performance after one round of DAgger." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 55, + 257, + 144, + 269 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 257, + 144, + 269 + ], + "spans": [ + { + "bbox": [ + 55, + 257, + 144, + 269 + ], + "type": "text", + "content": "3.4. Loss Function" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 55, + 275, + 295, + 322 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 275, + 295, + 322 + ], + "spans": [ + { + "bbox": [ + 55, + 275, + 295, + 322 + ], + "type": "text", + "content": "During training, the L1 loss function supervises all predicted decision variables, including target speed " + }, + { + "bbox": [ + 55, + 275, + 295, + 322 + ], + "type": "inline_equation", + "content": "(TS)" + }, + { + "bbox": [ + 55, + 275, + 295, + 322 + ], + "type": "text", + "content": ", target angle " + }, + { + "bbox": [ + 55, + 275, + 295, + 322 + ], + "type": "inline_equation", + "content": "(TA)" + }, + { + "bbox": [ + 55, + 275, + 295, + 322 + ], + "type": "text", + "content": ", waypoints " + }, + { + "bbox": [ + 55, + 275, + 295, + 322 + ], + "type": "inline_equation", + "content": "(WP)" + }, + { + "bbox": [ + 55, + 275, + 295, + 322 + ], + "type": "text", + "content": ", and route points " + }, + { + "bbox": [ + 55, + 275, + 295, + 322 + ], + "type": "inline_equation", + "content": "(RP)" + }, + { + "bbox": [ + 55, + 275, + 295, + 322 + ], + "type": "text", + "content": ". The overall loss is calculated by:" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 105, + 327, + 294, + 340 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 105, + 327, + 294, + 340 + ], + "spans": [ + { + "bbox": [ + 105, + 327, + 294, + 340 + ], + "type": "interline_equation", + "content": "\\mathcal {L} = \\mathcal {L} _ {T S} + \\mathcal {L} _ {T A} + \\mathcal {L} _ {W P} + \\mathcal {L} _ {R P} \\tag {1}", + "image_path": "fd7a8d0a03b1e9ee253ff3afbe997bdb9b1ee7ad76e73b9135289615bc0b2f41.jpg" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 55, + 344, + 295, + 380 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 344, + 295, + 380 + ], + "spans": [ + { + "bbox": [ + 55, + 344, + 295, + 380 + ], + "type": "text", + "content": "This approach allows the model to optimize final numerical vehicle decision variables directly, instead of relying on vocabulary-based classification as in NLP tasks." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 55, + 389, + 135, + 403 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 389, + 135, + 403 + ], + "spans": [ + { + "bbox": [ + 55, + 389, + 135, + 403 + ], + "type": "text", + "content": "4. Experiments" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 55, + 409, + 191, + 420 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 409, + 191, + 420 + ], + "spans": [ + { + "bbox": [ + 55, + 409, + 191, + 420 + ], + "type": "text", + "content": "4.1. Dataset and Benchmark" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 55, + 426, + 295, + 557 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 426, + 295, + 557 + ], + "spans": [ + { + "bbox": [ + 55, + 426, + 295, + 557 + ], + "type": "text", + "content": "All experiments in this study were conducted in the CARLA simulator [14]. We followed Transfuser++ for training data preparation. In stage 1, a rule-based autopilot was employed to collect 350K behavior-cloning samples. The autopilot traversed multiple routes in Towns 1, 2, 3, 4, 5, 6, 7, and 10 of CARLA, encountering various scenarios such as different weather and lighting conditions. Data were collected at a frequency of " + }, + { + "bbox": [ + 55, + 426, + 295, + 557 + ], + "type": "inline_equation", + "content": "2\\mathrm{Hz}" + }, + { + "bbox": [ + 55, + 426, + 295, + 557 + ], + "type": "text", + "content": " and subsequently filtered to remove incorrect or redundant samples, such as those captured when the autopilot was stuck. This process resulted in a dataset of 300K samples." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 55, + 558, + 295, + 653 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 558, + 295, + 653 + ], + "spans": [ + { + "bbox": [ + 55, + 558, + 295, + 653 + ], + "type": "text", + "content": "In stage 2, DriveGPT4-V2 trained by behavior cloning runs on selected training scenarios and routes, utilizing the expert for on-policy supervision. The expert can access privileged information about objects within a 30-meter distance. Approximately 150K samples were gathered in this stage. These aggregated data were then combined with the original dataset to further fine-tune DriveGPT4-V2, enhancing its performance." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 55, + 654, + 295, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 654, + 295, + 713 + ], + "spans": [ + { + "bbox": [ + 55, + 654, + 295, + 713 + ], + "type": "text", + "content": "All methods were evaluated on the challenging CARLA Longest6 benchmark [12], which consists of 36 extended routes encompassing various scenarios. The length and complexity of routes make them suitable for a comprehensive evaluation of closed-loop autonomous driving." + } + ] + } + ], + "index": 10 + }, + { + "type": "table", + "bbox": [ + 315, + 91, + 553, + 148 + ], + "blocks": [ + { + "bbox": [ + 370, + 71, + 497, + 82 + ], + "lines": [ + { + "bbox": [ + 370, + 71, + 497, + 82 + ], + "spans": [ + { + "bbox": [ + 370, + 71, + 497, + 82 + ], + "type": "text", + "content": "Table 1. Infraction penalty factors." + } + ] + } + ], + "index": 11, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 315, + 91, + 553, + 148 + ], + "lines": [ + { + "bbox": [ + 315, + 91, + 553, + 148 + ], + "spans": [ + { + "bbox": [ + 315, + 91, + 553, + 148 + ], + "type": "table", + "html": "
ScenePenaltyScenePenalty
Pedestrian collision0.5Vehicle collision0.6
Static object collision0.65Traffic light violation0.7
Stop sign violation0.7
", + "image_path": "1bcd09b5ebe2bcdb690dda945a2fb256bb939c7af0d5d8b7a80ee1b7dcb764da.jpg" + } + ] + } + ], + "index": 12, + "angle": 0, + "type": "table_body" + } + ], + "index": 12 + }, + { + "bbox": [ + 313, + 167, + 476, + 180 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 167, + 476, + 180 + ], + "spans": [ + { + "bbox": [ + 313, + 167, + 476, + 180 + ], + "type": "text", + "content": "4.2. DriveGPT4-V2 Configuration" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 313, + 185, + 553, + 280 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 185, + 553, + 280 + ], + "spans": [ + { + "bbox": [ + 313, + 185, + 553, + 280 + ], + "type": "text", + "content": "LLM. For efficient data collection and inference, we used tiny-scale LLMs in DriveGPT4-V2, such as Qwen-0.5B [2] and TinyLLaMA [61]. These models were fine-tuned for multimodal understanding, following methods outlined in LLaVA [29, 30] and TinyLLaVA [64]. Large-scale LLMs, like LLaMA3-8B [49] and vicuna-7B [11], are not considered at the moment due to their high computational demands and reduced inference speed." + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 313, + 281, + 553, + 351 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 281, + 553, + 351 + ], + "spans": [ + { + "bbox": [ + 313, + 281, + 553, + 351 + ], + "type": "text", + "content": "Training. DriveGPT4-V2 was fully fine-tuned in an end-to-end manner, except for the SigLIP " + }, + { + "bbox": [ + 313, + 281, + 553, + 351 + ], + "type": "inline_equation", + "content": "384 \\times 384" + }, + { + "bbox": [ + 313, + 281, + 553, + 351 + ], + "type": "text", + "content": " visual encoder [57]. The model was trained with a learning rate of " + }, + { + "bbox": [ + 313, + 281, + 553, + 351 + ], + "type": "inline_equation", + "content": "2e^{-5}" + }, + { + "bbox": [ + 313, + 281, + 553, + 351 + ], + "type": "text", + "content": " over 60 epochs by employing a cosine annealing schedule for better learning rate control. The training takes about 75 hours on 16 A800 GPUs." + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 313, + 353, + 554, + 460 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 353, + 554, + 460 + ], + "spans": [ + { + "bbox": [ + 313, + 353, + 554, + 460 + ], + "type": "text", + "content": "Inference. Two PID controllers were designed to convert DriveGPT4-V2's predicted decision variables into low-level control signals (i.e., throttle, steer, and brake). These control signals were used to directly drive the ego vehicle in the CARLA simulator. In the experiments, only the predicted target speed and target angle were used by the PID controllers, while waypoints and route points served as additional supervision for improved training results. The model runs inference on a single A800 GPU with FP16 precision." + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 313, + 468, + 428, + 479 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 468, + 428, + 479 + ], + "spans": [ + { + "bbox": [ + 313, + 468, + 428, + 479 + ], + "type": "text", + "content": "4.3. Evaluation Criteria" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 313, + 485, + 554, + 640 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 485, + 554, + 640 + ], + "spans": [ + { + "bbox": [ + 313, + 485, + 554, + 640 + ], + "type": "text", + "content": "Following the CARLA leaderboard, the methods were evaluated mainly based on three metrics: Driving Score (DS), Route Completion (RC), and Infraction Score (IS). RC indicates the percentage of routes completed, while IS penalizes infractions such as traffic violations and collisions during inference. Each infraction multiplicatively reduces IS by a penalty factor. The penalty factors are listed in Tab. 1. DS, as the product of RC and IS, provides a comprehensive evaluation score. In the evaluation, we also report ratios of commonly seen exception events, including pedestrian collision (Ped), vehicle collision (Veh), static object collision (Stat), red light violation (Red), route deviations (Dev), scenario time out (TO) and agent block (Block)." + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 313, + 648, + 457, + 661 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 648, + 457, + 661 + ], + "spans": [ + { + "bbox": [ + 313, + 648, + 457, + 661 + ], + "type": "text", + "content": "4.4. Comparison Experiments" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 313, + 666, + 553, + 689 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 666, + 553, + 689 + ], + "spans": [ + { + "bbox": [ + 313, + 666, + 553, + 689 + ], + "type": "text", + "content": "Baselines. We selected several representative prior works as baselines for comparison:" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 314, + 689, + 553, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 314, + 689, + 553, + 713 + ], + "spans": [ + { + "bbox": [ + 314, + 689, + 553, + 713 + ], + "type": "text", + "content": "- WOR [6] (ICCV 2021): WOR uses commands instead of target points for global planning instructions, providing" + } + ] + } + ], + "index": 21 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "text", + "content": "17266" + } + ] + } + ], + "index": 22 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 5 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 56, + 123, + 555, + 269 + ], + "blocks": [ + { + "bbox": [ + 55, + 70, + 555, + 115 + ], + "lines": [ + { + "bbox": [ + 55, + 70, + 555, + 115 + ], + "spans": [ + { + "bbox": [ + 55, + 70, + 555, + 115 + ], + "type": "text", + "content": "Table 2. Closed-loop experiments performance on CARLA Longest6. \"Visual\" indicates visual input modalities, while \"C\" and \"L\" represent camera and LiDAR, respectively. Bold numbers highlight SOTA metric scores of all models; while underlined numbers represent the best metric scores of baseline methods. * denotes models implemented by ourselves based on official open-sourced code. † represents the model without data augmentation." + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 56, + 123, + 555, + 269 + ], + "lines": [ + { + "bbox": [ + 56, + 123, + 555, + 269 + ], + "spans": [ + { + "bbox": [ + 56, + 123, + 555, + 269 + ], + "type": "table", + "html": "
MethodVisualDS↑RC↑IS↑Ped↓Veh↓Stat↓Red↓Dev↓TO↓Block↓
WOR [6]C21480.560.181.050.371.280.880.080.20
LAV v1 [4]C&L33700.510.160.830.150.960.060.120.45
Interfuser [42]C47740.630.061.140.110.240.000.520.06
TransFuser [12]C&L47930.500.032.450.070.160.000.060.10
LAV v2 [4]C&L58830.680.000.690.150.230.080.320.11
Perception PlanT [38]C&L58880.650.070.970.110.090.000.130.13
Transfuser++* [21]C&L65900.720.000.990.010.070.000.100.12
Transfuser++*† [21]C&L58890.650.011.150.010.100.000.140.13
LMDrive* [43]C&L36690.520.071.030.181.010.090.110.22
DriveGPT4-V2C70910.770.000.800.010.040.000.070.09
", + "image_path": "2ffcb13e728c53ea93460d7aa573e4cec76d457684b9e19ab53cd0ebad510c73.jpg" + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "table_body" + } + ], + "index": 1 + }, + { + "bbox": [ + 63, + 288, + 274, + 300 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 63, + 288, + 274, + 300 + ], + "spans": [ + { + "bbox": [ + 63, + 288, + 274, + 300 + ], + "type": "text", + "content": "enriched labels and supervision for possible actions." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 55, + 300, + 294, + 551 + ], + "type": "list", + "angle": 0, + "index": 9, + "blocks": [ + { + "bbox": [ + 55, + 300, + 294, + 335 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 300, + 294, + 335 + ], + "spans": [ + { + "bbox": [ + 55, + 300, + 294, + 335 + ], + "type": "text", + "content": "- LAV v1 & v2 [4] (CVPR 2022): LAV predicts waypoints by integrating information from nearby vehicles, which is further refined through a GRU." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 55, + 336, + 294, + 371 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 336, + 294, + 371 + ], + "spans": [ + { + "bbox": [ + 55, + 336, + 294, + 371 + ], + "type": "text", + "content": "- Transfuser [12] (TPAMI 2022): Transfuser is a widely used baseline model for closed-loop driving that has inspired subsequent works." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 55, + 372, + 294, + 407 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 372, + 294, + 407 + ], + "spans": [ + { + "bbox": [ + 55, + 372, + 294, + 407 + ], + "type": "text", + "content": "- Perception PlanT [38] (CoRL 2022): PlanT is based on imitation learning with object-level input representation. Such representation is elegant and effective." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 55, + 407, + 294, + 443 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 407, + 294, + 443 + ], + "spans": [ + { + "bbox": [ + 55, + 407, + 294, + 443 + ], + "type": "text", + "content": "- InterFuser [42] (CoRL 2023): InterFuser predicts additional density maps and traffic rule flags to enhance waypoint prediction." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 55, + 443, + 294, + 514 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 443, + 294, + 514 + ], + "spans": [ + { + "bbox": [ + 55, + 443, + 294, + 514 + ], + "type": "text", + "content": "- Transfuser++ [21] (ICCV 2023): Transfuser++ predicts target speeds and routes instead of waypoints, integrating point cloud and camera images. It is currently the state-of-the-art (SOTA) method for CARLA Longest6. Transfuser++ utilizes triple training data by repeatedly collecting expert trajectories for augmentation." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 55, + 515, + 294, + 551 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 515, + 294, + 551 + ], + "spans": [ + { + "bbox": [ + 55, + 515, + 294, + 551 + ], + "type": "text", + "content": "- LMDrive [43] (CVPR 2024): LMDrive is a pioneering work for closed-loop driving with multimodal LLMs, training a vicuna-7B LLM for end-to-end vehicle control." + } + ] + } + ], + "index": 8 + } + ], + "sub_type": "text" + }, + { + "bbox": [ + 55, + 558, + 295, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 558, + 295, + 713 + ], + "spans": [ + { + "bbox": [ + 55, + 558, + 295, + 713 + ], + "type": "text", + "content": "Comparison results. The comparison results on the CARLA Longest6 benchmark are shown in Tab. 2. It is observed that DriveGPT4-V2 outperforms all baseline models by a large margin. Thanks to the huge amount pretrained knowledge of LLMs and visual encoders, DriveGPT4-V2 can better handle complex urban scenes under various conditions, such as nighttime, rainy weather, etc. For most baselines, we list the results reported on the CARLA Longest6 leaderboard. However, some baselines cannot achieve reported performance by our implementation and test, thus we report our implemented results separately. LMDrive [43] is the only open-sourced LLM method that can be tested in a closed-loop environment. But it fails" + } + ] + } + ], + "index": 10 + }, + { + "type": "table", + "bbox": [ + 315, + 306, + 556, + 365 + ], + "blocks": [ + { + "bbox": [ + 383, + 286, + 485, + 298 + ], + "lines": [ + { + "bbox": [ + 383, + 286, + 485, + 298 + ], + "spans": [ + { + "bbox": [ + 383, + 286, + 485, + 298 + ], + "type": "text", + "content": "Table 3. Efficiency analysis." + } + ] + } + ], + "index": 11, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 315, + 306, + 556, + 365 + ], + "lines": [ + { + "bbox": [ + 315, + 306, + 556, + 365 + ], + "spans": [ + { + "bbox": [ + 315, + 306, + 556, + 365 + ], + "type": "table", + "html": "
LLMDSTrainFPS
LLaVA-LLaMA3.1-8B6511.2h/epoch0.4
TinyLLaVA-LLaMA-1.5B633.0h/epoch2.9
LLaVA-Qwen-0.5B631.3h/epoch8.1
", + "image_path": "d1dbaecddd538c546c66bec96d5ac8fe3860c661fead26e6998d4ec67bf28290.jpg" + } + ] + } + ], + "index": 12, + "angle": 0, + "type": "table_body" + } + ], + "index": 12 + }, + { + "bbox": [ + 313, + 385, + 555, + 529 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 385, + 555, + 529 + ], + "spans": [ + { + "bbox": [ + 313, + 385, + 555, + 529 + ], + "type": "text", + "content": "to obtain satisfactory performance. It first converts multimodal data to BEV and projects BEV to the text domain. However, BEV does not have good-quality pretrained multimodal LLMs. Training from scratch severely degrades its final performance, which is also discussed by DriveGPT4-V1 [54]. Transfuser++ [21] is the SOTA method by fusing point clouds and camera images for closed-loop control. Transfuser++ uses triple training data and achieves 65 DS, but only has 58 DS with the same amount of training data (denoted by " + }, + { + "bbox": [ + 313, + 385, + 555, + 529 + ], + "type": "inline_equation", + "content": "\\dagger" + }, + { + "bbox": [ + 313, + 385, + 555, + 529 + ], + "type": "text", + "content": ") as our method. Thanks to the pretrained knowledge and reasoning ability of LLMs, DriveGPT4-V2 can outperform all baselines with camera images as input." + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 313, + 540, + 416, + 552 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 540, + 416, + 552 + ], + "spans": [ + { + "bbox": [ + 313, + 540, + 416, + 552 + ], + "type": "text", + "content": "4.5. Model Efficiency" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 313, + 558, + 555, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 558, + 555, + 713 + ], + "spans": [ + { + "bbox": [ + 313, + 558, + 555, + 713 + ], + "type": "text", + "content": "In our experiments, DriveGPT4-V2 and the expert are implemented by tiny LLMs. LLMs with 7B or even more parameters show powerful reasoning and generalization ability, but they severely enlarge the time consumption and demand much more training resources, which is unacceptable in real-world applications. In this section, we report efficiency results of DriveGPT4-V2 with LLaVA-0.5B, TinyLLaVA-1.5B and LLaVA-7B LLMs. The results are shown in Tab. 3. It is found that merely scaling up the LLM model does not gain corresponding performance enhancement, but slows down the whole system more than 10 times. Therefore, DriveGPT4-V2 mainly considers 0.5B LLMs as the planner." + } + ] + } + ], + "index": 15 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "text", + "content": "17267" + } + ] + } + ], + "index": 16 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 6 + }, + { + "para_blocks": [ + { + "bbox": [ + 55, + 72, + 228, + 83 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 72, + 228, + 83 + ], + "spans": [ + { + "bbox": [ + 55, + 72, + 228, + 83 + ], + "type": "text", + "content": "4.6. Ablation Studies and Discussion" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 55, + 89, + 294, + 137 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 89, + 294, + 137 + ], + "spans": [ + { + "bbox": [ + 55, + 89, + 294, + 137 + ], + "type": "text", + "content": "We conducted multiple ablation studies to justify the design of DriveGPT4-V2. The results are presented in Tab. 4, 5 and 6. Due to huge time consumption, DriveGPT4-V2 in ablation studies is not trained with DAgger data by default." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 55, + 137, + 294, + 221 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 137, + 294, + 221 + ], + "spans": [ + { + "bbox": [ + 55, + 137, + 294, + 221 + ], + "type": "text", + "content": "LLM visual pretraining. DriveGPT4-V2 relies on LLaVA [30, 64] as the planner, which is pretrained on a huge amount of visual understanding data. The visual pretraining benefits DriveGPT4-V2 on the autonomous driving task with multimodal input. Directly training DriveGPT4-V2 on scratch (i.e., LLM for NLP tasks without multimodal ability) slows down convergence and degrades the final results." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 55, + 221, + 294, + 342 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 221, + 294, + 342 + ], + "spans": [ + { + "bbox": [ + 55, + 221, + 294, + 342 + ], + "type": "text", + "content": "Visual tokenizer. The input camera images are processed by the proposed visual tokenizer. Inspired by LLaVA 1.5 [30], the features of images are extracted separately by the visual encoder and then concatenated as the input. This design ensures a sufficient perception range without compromising image detail. If the input is a single image with a large scope, it might lead to information loss during preprocessing steps (e.g., resize or crop); if the input is a small square image, the vehicle cannot have a sufficient perception range for effective planning." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 55, + 342, + 294, + 413 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 342, + 294, + 413 + ], + "spans": [ + { + "bbox": [ + 55, + 342, + 294, + 413 + ], + "type": "text", + "content": "Waypoints and route points. Although DriveGPT4-V2 does not rely on waypoints and route points for generating low-level control signals, they provide valuable supervision during training. Predicting waypoints and route points helps the model perceive the target route and anticipate future actions. Removing them negatively impacts performance." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 55, + 413, + 294, + 485 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 413, + 294, + 485 + ], + "spans": [ + { + "bbox": [ + 55, + 413, + 294, + 485 + ], + "type": "text", + "content": "Expert on-policy supervision. The expert model, benefiting from privileged information, provides robust on-policy supervision to DriveGPT4-V2. Without expert guidance, DriveGPT4-V2 relies solely on behavior cloning instead of DAGger imitation learning, which can lead to error accumulation and degraded performance." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 55, + 486, + 294, + 593 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 486, + 294, + 593 + ], + "spans": [ + { + "bbox": [ + 55, + 486, + 294, + 593 + ], + "type": "text", + "content": "PID controllers. Previous works often generated low-level control signals based on waypoints or route points [12, 21] by calculating target speed and angle from these predictions. However, even minor prediction errors in waypoints or route points can lead to substantial errors in speed and angle. As shown in Tab. 5, using waypoints or route points for PID control can make control signals noisy and unstable, potentially driving the vehicle into exceptional states. Therefore it presents inferior results." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 55, + 594, + 295, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 594, + 295, + 714 + ], + "spans": [ + { + "bbox": [ + 55, + 594, + 295, + 714 + ], + "type": "text", + "content": "Decision heads. During inference, DriveGPT4-V2 generates four tokens in an auto-regressive manner, mapped to vehicle decision variables by four decision heads. Predictions of target speed and angle involve one-to-one mappings (i.e., one token for one number), while waypoint and route point predictions involve one-to-many mappings (e.g., waypoints contain 8 x-y points, thus each token is mapped to 16 numbers). Another design would involve predicting each waypoint and route point one at a time, requiring 8 tokens for waypoints and 10 tokens for route points. Such design" + } + ] + } + ], + "index": 7 + }, + { + "type": "table", + "bbox": [ + 316, + 97, + 551, + 176 + ], + "blocks": [ + { + "bbox": [ + 313, + 70, + 553, + 93 + ], + "lines": [ + { + "bbox": [ + 313, + 70, + 553, + 93 + ], + "spans": [ + { + "bbox": [ + 313, + 70, + 553, + 93 + ], + "type": "text", + "content": "Table 4. Ablation studies of DriveGPT4-V2. \"WP\" and \"RP\" represent waypoints and route points, respectively." + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 316, + 97, + 551, + 176 + ], + "lines": [ + { + "bbox": [ + 316, + 97, + 551, + 176 + ], + "spans": [ + { + "bbox": [ + 316, + 97, + 551, + 176 + ], + "type": "table", + "html": "
DSRCIS
Baseline47780.60
+ LLM Visual Pretraining56870.64
+ Visual Tokenizer60880.68
+ WP&RP63900.70
+ Expert Supervision70910.77
", + "image_path": "582556cd97569854a55b8c0bef212c1740407aaacc1c0df524794c42e9a52a3f.jpg" + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "table_body" + } + ], + "index": 9 + }, + { + "type": "table", + "bbox": [ + 317, + 218, + 550, + 280 + ], + "blocks": [ + { + "bbox": [ + 313, + 181, + 553, + 215 + ], + "lines": [ + { + "bbox": [ + 313, + 181, + 553, + 215 + ], + "spans": [ + { + "bbox": [ + 313, + 181, + 553, + 215 + ], + "type": "text", + "content": "Table 5. Ablation studies on PID controllers. \"WP\" indicates utilizing predicted waypoints for PID control; while \"TS&RP\" means PID control by predicted target speed and route points." + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 317, + 218, + 550, + 280 + ], + "lines": [ + { + "bbox": [ + 317, + 218, + 550, + 280 + ], + "spans": [ + { + "bbox": [ + 317, + 218, + 550, + 280 + ], + "type": "table", + "html": "
PID ControllerDSRCIS
WP53850.62
TS & RP59880.67
DriveGPT4-V263900.70
", + "image_path": "a1112eaa6830ab7fb1654f13e6b848c0d3b07bc7a7d964a6d0d2b7fe56fe7b29.jpg" + } + ] + } + ], + "index": 11, + "angle": 0, + "type": "table_body" + } + ], + "index": 11 + }, + { + "type": "table", + "bbox": [ + 316, + 312, + 551, + 363 + ], + "blocks": [ + { + "bbox": [ + 313, + 285, + 553, + 308 + ], + "lines": [ + { + "bbox": [ + 313, + 285, + 553, + 308 + ], + "spans": [ + { + "bbox": [ + 313, + 285, + 553, + 308 + ], + "type": "text", + "content": "Table 6. Ablation studies on decision heads. \"Additional tokens\" indicates using more output tokens for prediction." + } + ] + } + ], + "index": 12, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 316, + 312, + 551, + 363 + ], + "lines": [ + { + "bbox": [ + 316, + 312, + 551, + 363 + ], + "spans": [ + { + "bbox": [ + 316, + 312, + 551, + 363 + ], + "type": "table", + "html": "
DSRCISFPS
Additional tokens64910.701.4
DriveGPT4-V263900.708.1
", + "image_path": "8668107942c6bcfbdf41c9c51c64a5833ea2912a0c375fffa81dc5c05f6ca9b4.jpg" + } + ] + } + ], + "index": 13, + "angle": 0, + "type": "table_body" + } + ], + "index": 13 + }, + { + "bbox": [ + 313, + 374, + 553, + 434 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 374, + 553, + 434 + ], + "spans": [ + { + "bbox": [ + 313, + 374, + 553, + 434 + ], + "type": "text", + "content": "does not significantly improve performance and severely affects efficiency. Using text to represent output numbers would require approximately 160 tokens, making it impractical for real-time autonomous driving. The evaluation results are shown in Tab. 6" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 313, + 445, + 479, + 457 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 445, + 479, + 457 + ], + "spans": [ + { + "bbox": [ + 313, + 445, + 479, + 457 + ], + "type": "text", + "content": "5. Conclusion and Future Work" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 313, + 464, + 553, + 680 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 464, + 553, + 680 + ], + "spans": [ + { + "bbox": [ + 313, + 464, + 553, + 680 + ], + "type": "text", + "content": "In this paper, we presented DriveGPT4-V2, a novel LLM-based framework for closed-loop, end-to-end autonomous driving. DriveGPT4-V2 processes camera images and vehicle states as input to generate low-level control signals for direct vehicle operation. Leveraging the extensive pretrained knowledge of MLLMs, DriveGPT4-V2 demonstrates the ability to navigate complex urban scenarios under diverse and challenging conditions. The model architecture is specifically optimized for precise numerical vehicle decision prediction. An additional expert LLM, which shares a structure similar to DriveGPT4-V2, has been trained to provide on-policy supervision. Experimental results highlight that DriveGPT4-V2 achieves state-of-the-art performance on the challenging CARLA Longest6 benchmark, outperforming all baselines. In the future, we aim to extend the capabilities of DriveGPT4-V2 for broader applications in closed-loop autonomous driving tasks, including video games and real-world deployment." + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 313, + 680, + 553, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 680, + 553, + 713 + ], + "spans": [ + { + "bbox": [ + 313, + 680, + 553, + 713 + ], + "type": "text", + "content": "Acknowledgment. This work is partially supported by the National Natural Science Foundation of China (No. 62422606, 62201484, 52221005)." + } + ] + } + ], + "index": 17 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "text", + "content": "17268" + } + ] + } + ], + "index": 18 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 7 + }, + { + "para_blocks": [ + { + "bbox": [ + 56, + 72, + 115, + 83 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 72, + 115, + 83 + ], + "spans": [ + { + "bbox": [ + 56, + 72, + 115, + 83 + ], + "type": "text", + "content": "References" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 57, + 91, + 294, + 712 + ], + "type": "list", + "angle": 0, + "index": 18, + "blocks": [ + { + "bbox": [ + 61, + 91, + 294, + 111 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 61, + 91, + 294, + 111 + ], + "spans": [ + { + "bbox": [ + 61, + 91, + 294, + 111 + ], + "type": "text", + "content": "[1] Anthropic. Claude. https://www.anthropic.com, 2024." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 61, + 114, + 294, + 146 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 61, + 114, + 294, + 146 + ], + "spans": [ + { + "bbox": [ + 61, + 114, + 294, + 146 + ], + "type": "text", + "content": "[2] Jinze Bai, Shuai Bai, Yunfei Chu, Zeyu Cui, Kai Dang, Xiaodong Deng, Yang Fan, Wenbin Ge, Yu Han, Fei Huang, et al. Qwen technical report. arXiv:2309.16609, 2023." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 62, + 148, + 294, + 201 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 148, + 294, + 201 + ], + "spans": [ + { + "bbox": [ + 62, + 148, + 294, + 201 + ], + "type": "text", + "content": "[3] Xu Cao, Tong Zhou, Yunsheng Ma, Wenqian Ye, Can Cui, Kun Tang, Zhipeng Cao, Kaizhao Liang, Ziran Wang, James M Rehg, et al. Maplm: A real-world large-scale vision-language benchmark for map and traffic scene understanding. In CVPR, 2024." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 62, + 202, + 294, + 223 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 202, + 294, + 223 + ], + "spans": [ + { + "bbox": [ + 62, + 202, + 294, + 223 + ], + "type": "text", + "content": "[4] Dian Chen and Philipp Krahenbuhl. Learning from all vehicles. In CVPR, 2022." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 62, + 224, + 294, + 246 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 224, + 294, + 246 + ], + "spans": [ + { + "bbox": [ + 62, + 224, + 294, + 246 + ], + "type": "text", + "content": "[5] Dian Chen, Brady Zhou, Vladlen Koltun, and Philipp Krahenbuhl. Learning by cheating. In CoRL, 2020." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 62, + 247, + 294, + 268 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 247, + 294, + 268 + ], + "spans": [ + { + "bbox": [ + 62, + 247, + 294, + 268 + ], + "type": "text", + "content": "[6] Dian Chen, Vladlen Koltun, and Philipp Krahenbuhl. Learning to drive from a world on rails. In ICCV, 2021." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 62, + 270, + 294, + 323 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 270, + 294, + 323 + ], + "spans": [ + { + "bbox": [ + 62, + 270, + 294, + 323 + ], + "type": "text", + "content": "[7] Long Chen, Oleg Sinavski, Jan Hünermann, Alice Karnsund, Andrew James Willmott, Danny Birch, Daniel Maund, and Jamie Shotton. Driving with llms: Fusing object-level vector modality for explainable autonomous driving. In ICRA, 2024." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 62, + 325, + 294, + 357 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 325, + 294, + 357 + ], + "spans": [ + { + "bbox": [ + 62, + 325, + 294, + 357 + ], + "type": "text", + "content": "[8] Li Chen, Penghao Wu, Kashyap Chitta, Bernhard Jaeger, Andreas Geiger, and Hongyang Li. End-to-end autonomous driving: Challenges and frontiers. TPAMI, 2024." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 62, + 358, + 294, + 402 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 358, + 294, + 402 + ], + "spans": [ + { + "bbox": [ + 62, + 358, + 294, + 402 + ], + "type": "text", + "content": "[9] Shaoyu Chen, Bo Jiang, Hao Gao, Bencheng Liao, Qing Xu, Qian Zhang, Chang Huang, Wenyu Liu, and Xinggang Wang. Vadv2: End-to-end vectorized autonomous driving via probabilistic planning. arXiv:2402.13243, 2024." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 57, + 403, + 294, + 435 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 403, + 294, + 435 + ], + "spans": [ + { + "bbox": [ + 57, + 403, + 294, + 435 + ], + "type": "text", + "content": "[10] Yuan Chen, Zi-han Ding, Ziqin Wang, Yan Wang, Lijun Zhang, and Si Liu. Asynchronous large language model enhanced planner for autonomous driving. In ECCV, 2024." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 57, + 436, + 294, + 491 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 436, + 294, + 491 + ], + "spans": [ + { + "bbox": [ + 57, + 436, + 294, + 491 + ], + "type": "text", + "content": "[11] Wei-Lin Chiang, Zhuohan Li, Zi Lin, Ying Sheng, Zhang-hao Wu, Hao Zhang, Lianmin Zheng, Siyuan Zhuang, Yong-hao Zhuang, Joseph E Gonzalez, et al. Vicuna: An opensource chatbot impressing gpt-4 with " + }, + { + "bbox": [ + 57, + 436, + 294, + 491 + ], + "type": "inline_equation", + "content": "90\\%" + }, + { + "bbox": [ + 57, + 436, + 294, + 491 + ], + "type": "text", + "content": " * chatgpt quality. See https://vicuna.lmsys.org, 2023." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 57, + 491, + 294, + 535 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 491, + 294, + 535 + ], + "spans": [ + { + "bbox": [ + 57, + 491, + 294, + 535 + ], + "type": "text", + "content": "[12] Kashyap Chitta, Aditya Prakash, Bernhard Jaeger, Zehao Yu, Katrin Renz, and Andreas Geiger. Transfuser: Imitation with transformer-based sensor fusion for autonomous driving. TPAMI, 2022." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 57, + 536, + 294, + 556 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 536, + 294, + 556 + ], + "spans": [ + { + "bbox": [ + 57, + 536, + 294, + 556 + ], + "type": "text", + "content": "[13] Google DeepMind. Gemini ai. https://deepmind.com, 2024." + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 57, + 558, + 294, + 590 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 558, + 294, + 590 + ], + "spans": [ + { + "bbox": [ + 57, + 558, + 294, + 590 + ], + "type": "text", + "content": "[14] Alexey Dosovitskiy, German Ros, Felipe Codevilla, Antonio Lopez, and Vladlen Koltun. Carla: An open urban driving simulator. In CoRL, 2017." + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 57, + 592, + 294, + 624 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 592, + 294, + 624 + ], + "spans": [ + { + "bbox": [ + 57, + 592, + 294, + 624 + ], + "type": "text", + "content": "[15] Yuchao Feng, Zhen Feng, Wei Hua, and Yuxiang Sun. Multimodal-xad: Explainable autonomous driving based on multimodal environment descriptions. TITS, 2024." + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 57, + 625, + 294, + 667 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 625, + 294, + 667 + ], + "spans": [ + { + "bbox": [ + 57, + 625, + 294, + 667 + ], + "type": "text", + "content": "[16] Jiawei Fu, Yanqing Shen, Zhiqiang Jian, Shitao Chen, Jing-min Xin, and Nanning Zheng. Interactionnet: Joint planning and prediction for autonomous driving with transformers. In IROS, 2023." + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 57, + 669, + 294, + 712 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 669, + 294, + 712 + ], + "spans": [ + { + "bbox": [ + 57, + 669, + 294, + 712 + ], + "type": "text", + "content": "[17] Yihan Hu, Jiazhi Yang, Li Chen, Keyu Li, Chonghao Sima, Xizhou Zhu, Siqi Chai, Senyao Du, Tianwei Lin, Wenhai Wang, et al. Planning-oriented autonomous driving. In CVPR, 2023." + } + ] + } + ], + "index": 17 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 316, + 73, + 553, + 713 + ], + "type": "list", + "angle": 0, + "index": 35, + "blocks": [ + { + "bbox": [ + 316, + 73, + 553, + 105 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 73, + 553, + 105 + ], + "spans": [ + { + "bbox": [ + 316, + 73, + 553, + 105 + ], + "type": "text", + "content": "[18] Bin Huang, Xin Wang, Hong Chen, Zihan Song, and Wenwu Zhu. Vtimellm: Empower llm to grasp video moments. In CVPR, 2024." + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 317, + 107, + 553, + 138 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 107, + 553, + 138 + ], + "spans": [ + { + "bbox": [ + 317, + 107, + 553, + 138 + ], + "type": "text", + "content": "[19] Ahmed Hussein, Mohamed Medhat Gaber, Eyad Elyan, and Chrisina Jayne. Imitation learning: A survey of learning methods. ACM Computing Surveys, 2017." + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 317, + 140, + 553, + 182 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 140, + 553, + 182 + ], + "spans": [ + { + "bbox": [ + 317, + 140, + 553, + 182 + ], + "type": "text", + "content": "[20] Yuichi Inoue, Yuki Yada, Kotaro Tanahashi, and Yu Yamaguchi. Nuscenes-mqa: Integrated evaluation of captions and qa for autonomous driving datasets using markup annotations. In WACV, 2024." + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 316, + 184, + 553, + 205 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 184, + 553, + 205 + ], + "spans": [ + { + "bbox": [ + 316, + 184, + 553, + 205 + ], + "type": "text", + "content": "[21] Bernhard Jaeger, Kashyap Chitta, and Andreas Geiger. Hidden biases of end-to-end driving models. In ICCV, 2023." + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 317, + 206, + 553, + 249 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 206, + 553, + 249 + ], + "spans": [ + { + "bbox": [ + 317, + 206, + 553, + 249 + ], + "type": "text", + "content": "[22] Xiaosong Jia, Penghao Wu, Li Chen, Jiangwei Xie, Conghui He, Junchi Yan, and Hongyang Li. Think twice before driving: Towards scalable decoders for end-to-end autonomous driving. In CVPR, 2023." + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 317, + 251, + 553, + 293 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 251, + 553, + 293 + ], + "spans": [ + { + "bbox": [ + 317, + 251, + 553, + 293 + ], + "type": "text", + "content": "[23] Bo Jiang, Shaoyu Chen, Qing Xu, Benchcheng Liao, Jiajie Chen, Helong Zhou, Qian Zhang, Wenyu Liu, Chang Huang, and Xinggang Wang. Vad: Vectorized scene representation for efficient autonomous driving. In ICCV, 2023." + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 317, + 294, + 553, + 337 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 294, + 553, + 337 + ], + "spans": [ + { + "bbox": [ + 317, + 294, + 553, + 337 + ], + "type": "text", + "content": "[24] Zhuoling Li, Liangliang Ren, Jinrong Yang, Yong Zhao, Xiaoyang Wu, Zhenhua Xu, Xiang Bai, and Hengshuang Zhao. Virt: Vision instructed transformer for robotic manipulation. arXiv:2410.07169, 2024." + } + ] + } + ], + "index": 25 + }, + { + "bbox": [ + 317, + 338, + 553, + 380 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 338, + 553, + 380 + ], + "spans": [ + { + "bbox": [ + 317, + 338, + 553, + 380 + ], + "type": "text", + "content": "[25] Zhuoling Li, Xiaogang Xu, Zhenhua Xu, SerNam Lim, and Hengshuang Zhao. Larm: Large auto-regressive model for long-horizon embodied intelligence. arXiv:2405.17424, 2024." + } + ] + } + ], + "index": 26 + }, + { + "bbox": [ + 317, + 383, + 553, + 425 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 383, + 553, + 425 + ], + "spans": [ + { + "bbox": [ + 317, + 383, + 553, + 425 + ], + "type": "text", + "content": "[26] Bin Lin, Yang Ye, Bin Zhu, Jiaxi Cui, Munan Ning, Peng Jin, and Li Yuan. Video-llava: Learning united visual representation by alignment before projection. arXiv:2311.10122, 2023." + } + ] + } + ], + "index": 27 + }, + { + "bbox": [ + 317, + 426, + 553, + 470 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 426, + 553, + 470 + ], + "spans": [ + { + "bbox": [ + 317, + 426, + 553, + 470 + ], + "type": "text", + "content": "[27] Yuanfei Lin, Chenran Li, Mingyu Ding, Masayoshi Tomizuka, Wei Zhan, and Matthias Althoff. Drplanner: Diagnosis and repair of motion planners for automated vehicles using large language models. RA-L, 2024." + } + ] + } + ], + "index": 28 + }, + { + "bbox": [ + 317, + 472, + 553, + 503 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 472, + 553, + 503 + ], + "spans": [ + { + "bbox": [ + 317, + 472, + 553, + 503 + ], + "type": "text", + "content": "[28] Haotian Liu, Chunyuan Li, Yuheng Li, and Yong Jae Lee. Improved baselines with visual instruction tuning. In arXiv:2310.03744, 2023." + } + ] + } + ], + "index": 29 + }, + { + "bbox": [ + 317, + 505, + 553, + 525 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 505, + 553, + 525 + ], + "spans": [ + { + "bbox": [ + 317, + 505, + 553, + 525 + ], + "type": "text", + "content": "[29] Haotian Liu, Chunyuan Li, Qingyang Wu, and Yong Jae Lee. Visual instruction tuning. In NeurIPS, 2023." + } + ] + } + ], + "index": 30 + }, + { + "bbox": [ + 317, + 526, + 553, + 580 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 526, + 553, + 580 + ], + "spans": [ + { + "bbox": [ + 317, + 526, + 553, + 580 + ], + "type": "text", + "content": "[30] Haotian Liu, Chunyuan Li, Yuheng Li, Bo Li, Yuanhan Zhang, Sheng Shen, and Yong Jae Lee. Llava-last: Improved reasoning,OCR, and world knowledge. https://llava-v1.github.io/blog/2024-01-30-llava-last, 2024." + } + ] + } + ], + "index": 31 + }, + { + "bbox": [ + 317, + 582, + 553, + 646 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 582, + 553, + 646 + ], + "spans": [ + { + "bbox": [ + 317, + 582, + 553, + 646 + ], + "type": "text", + "content": "[31] Tianyu Liu, Qing hai Liao, Lu Gan, Fulong Ma, Jie Cheng, Xupeng Xie, Zhe Wang, Yingbing Chen, Yilong Zhu, Shuyang Zhang, et al. The role of the hercules autonomous vehicle during the COVID-19 pandemic: An autonomous logistic vehicle for contactless goods transportation. IEEE Robotics & Automation Magazine, 2021." + } + ] + } + ], + "index": 32 + }, + { + "bbox": [ + 317, + 647, + 553, + 680 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 647, + 553, + 680 + ], + "spans": [ + { + "bbox": [ + 317, + 647, + 553, + 680 + ], + "type": "text", + "content": "[32] Jiageng Mao, Yuxi Qian, Junjie Ye, Hang Zhao, and Yue Wang. Gpt-driver: Learning to drive with gpt. In NeurIPS Workshop, 2023." + } + ] + } + ], + "index": 33 + }, + { + "bbox": [ + 317, + 681, + 553, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 681, + 553, + 713 + ], + "spans": [ + { + "bbox": [ + 317, + 681, + 553, + 713 + ], + "type": "text", + "content": "[33] Maxim Mnyakin. Challenges and opportunities of integrating autonomous vehicles into urban retail delivery services. Reviews of Contemporary Business Analytics, 2023." + } + ] + } + ], + "index": 34 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "text", + "content": "17269" + } + ] + } + ], + "index": 36 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 8 + }, + { + "para_blocks": [ + { + "bbox": [ + 56, + 73, + 295, + 713 + ], + "type": "list", + "angle": 0, + "index": 16, + "blocks": [ + { + "bbox": [ + 56, + 73, + 294, + 84 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 73, + 294, + 84 + ], + "spans": [ + { + "bbox": [ + 56, + 73, + 294, + 84 + ], + "type": "text", + "content": "[34] OpenAI. Chatgpt. https://chat.openai.com, 2024." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 57, + 85, + 295, + 129 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 85, + 295, + 129 + ], + "spans": [ + { + "bbox": [ + 57, + 85, + 295, + 129 + ], + "type": "text", + "content": "[35] Takayuki Osa, Joni Pajarinen, Gerhard Neumann, J Andrew Bagnell, Pieter Abbeel, Jan Peters, et al. An algorithmic perspective on imitation learning. Foundations and Trends® in Robotics, 2018." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 57, + 130, + 294, + 174 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 130, + 294, + 174 + ], + "spans": [ + { + "bbox": [ + 57, + 130, + 294, + 174 + ], + "type": "text", + "content": "[36] SungYeon Park, MinJae Lee, JiHyuk Kang, Hahyeon Choi, Yoonah Park, Juhwan Cho, Adam Lee, and DongKyu Kim. Vlaad: Vision and language assistant for autonomous driving. In WACV, 2024." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 57, + 175, + 295, + 229 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 175, + 295, + 229 + ], + "spans": [ + { + "bbox": [ + 57, + 175, + 295, + 229 + ], + "type": "text", + "content": "[37] Alec Radford, Jong Wook Kim, Chris Hallacy, Aditya Ramesh, Gabriel Goh, Sandhini Agarwal, Girish Sastry, Amanda Askell, Pamela Mishkin, Jack Clark, et al. Learning transferable visual models from natural language supervision. In ICML, 2021." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 57, + 231, + 295, + 274 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 231, + 295, + 274 + ], + "spans": [ + { + "bbox": [ + 57, + 231, + 295, + 274 + ], + "type": "text", + "content": "[38] Katrin Renz, Kashyap Chitta, Otniel-Bogdan Mercea, A Koepke, Zeynep Akata, and Andreas Geiger. Plant: Explainable planning transformers via object-level representations. In CoRL, 2022." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 56, + 276, + 295, + 330 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 276, + 295, + 330 + ], + "spans": [ + { + "bbox": [ + 56, + 276, + 295, + 330 + ], + "type": "text", + "content": "[39] Katrin Renz, Long Chen, Ana-Maria Marcu, Jan Hunermann, Benoit Hanotte, Alice Karnsund, Jamie Shotton, Elahe Arani, and Oleg Sinavski. Carllava: Vision language models for camera-only closed-loop driving. arXiv:2406.10165, 2024." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 57, + 332, + 294, + 364 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 332, + 294, + 364 + ], + "spans": [ + { + "bbox": [ + 57, + 332, + 294, + 364 + ], + "type": "text", + "content": "[40] Stephane Ross and J. Andrew Bagnell. Reinforcement and imitation learning via interactive no-regret learning. In arXiv:1406.5979, 2014." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 57, + 365, + 294, + 398 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 365, + 294, + 398 + ], + "spans": [ + { + "bbox": [ + 57, + 365, + 294, + 398 + ], + "type": "text", + "content": "[41] Stéphane Ross, Geoffrey Gordon, and Drew Bagnell. A reduction of imitation learning and structured prediction to no-regret online learning. In AISTATS, 2011." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 57, + 399, + 294, + 432 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 399, + 294, + 432 + ], + "spans": [ + { + "bbox": [ + 57, + 399, + 294, + 432 + ], + "type": "text", + "content": "[42] Hao Shao, Letian Wang, Ruobing Chen, Hongsheng Li, and Yu Liu. Safety-enhanced autonomous driving using interpretable sensor fusion transformer. In CoRL, 2023." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 57, + 434, + 294, + 476 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 434, + 294, + 476 + ], + "spans": [ + { + "bbox": [ + 57, + 434, + 294, + 476 + ], + "type": "text", + "content": "[43] Hao Shao, Yuxuan Hu, Letian Wang, Guanglu Song, Steven L Waslander, Yu Liu, and Hongsheng Li. Lmdrive: Closed-loop end-to-end driving with large language models. In CVPR, 2024." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 57, + 478, + 294, + 521 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 478, + 294, + 521 + ], + "spans": [ + { + "bbox": [ + 57, + 478, + 294, + 521 + ], + "type": "text", + "content": "[44] Chonghao Sima, Katrin Renz, Kashyap Chitta, Li Chen, Hanxue Zhang, Chengen Xie, Jens Beiwenger, Ping Luo, Andreas Geiger, and Hongyang Li. Drivelm: Driving with graph visual question answering. In ECCV, 2024." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 57, + 523, + 294, + 577 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 523, + 294, + 577 + ], + "spans": [ + { + "bbox": [ + 57, + 523, + 294, + 577 + ], + "type": "text", + "content": "[45] Robert Soler, Alae Moudni, Gabriel Roskowski, Xinrui Yu, Mikhail Gormov, and Jafar Saniee. Autonomous patrol and threat detection through integrated mapping and computer vision. In IEEE International Conference on Electro Information Technology, 2024." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 57, + 579, + 294, + 621 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 579, + 294, + 621 + ], + "spans": [ + { + "bbox": [ + 57, + 579, + 294, + 621 + ], + "type": "text", + "content": "[46] Chan Hee Song, Jiaman Wu, Clayton Washington, Brian M Sadler, Wei-Lun Chao, and Yu Su. Llm-planner: Few-shot grounded planning for embodied agents with large language models. In ICCV, 2023." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 57, + 624, + 294, + 666 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 624, + 294, + 666 + ], + "spans": [ + { + "bbox": [ + 57, + 624, + 294, + 666 + ], + "type": "text", + "content": "[47] Yafu Tian, Alexander Carballo, Ruifeng Li, Simon Thompson, and Kazuya Takeda. Rsg-search plus: An advanced traffic scene retrieval methods based on road scene graph. In IV, 2024." + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 57, + 669, + 294, + 690 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 669, + 294, + 690 + ], + "spans": [ + { + "bbox": [ + 57, + 669, + 294, + 690 + ], + "type": "text", + "content": "[48] Faraz Torabi, Garrett Warnell, and Peter Stone. Behavioral cloning from observation. In *IJCAI*, 2018." + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 57, + 692, + 294, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 692, + 294, + 713 + ], + "spans": [ + { + "bbox": [ + 57, + 692, + 294, + 713 + ], + "type": "text", + "content": "[49] Hugo Touvron, Thibaut Lavril, Gautier Izacard, Xavier Martinet, Marie-Anne Lachaux, Timothée Lacroix, Baptiste" + } + ] + } + ], + "index": 15 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 316, + 73, + 553, + 712 + ], + "type": "list", + "angle": 0, + "index": 34, + "blocks": [ + { + "bbox": [ + 333, + 73, + 553, + 106 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 333, + 73, + 553, + 106 + ], + "spans": [ + { + "bbox": [ + 333, + 73, + 553, + 106 + ], + "type": "text", + "content": "Rozière, Naman Goyal, Eric Hambro, Faisal Azhar, et al. Llama: Open and efficient foundation language models. arXiv:2302.13971, 2023." + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 316, + 107, + 553, + 149 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 107, + 553, + 149 + ], + "spans": [ + { + "bbox": [ + 316, + 107, + 553, + 149 + ], + "type": "text", + "content": "[50] Yuxi Wei, Zi Wang, Yifan Lu, Chenxin Xu, Changxing Liu, Hao Zhao, Siheng Chen, and Yanfeng Wang. Editable scene simulation for autonomous driving via collaborative llm-agents. In CVPR, 2024." + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 316, + 151, + 553, + 194 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 151, + 553, + 194 + ], + "spans": [ + { + "bbox": [ + 316, + 151, + 553, + 194 + ], + "type": "text", + "content": "[51] Licheng Wen, Daocheng Fu, Xin Li, Xinyu Cai, Tao Ma, Pinlong Cai, Min Dou, Botian Shi, Liang He, and Yu Qiao. Dilu: A knowledge-driven approach to autonomous driving with large language models. In ICLR, 2024." + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 316, + 195, + 553, + 227 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 195, + 553, + 227 + ], + "spans": [ + { + "bbox": [ + 316, + 195, + 553, + 227 + ], + "type": "text", + "content": "[52] Zhenhua Xu, Yuxuan Liu, Lu Gan, Yuxiang Sun, Ming Liu, and Lujia Wang. Ringdet: Road network graph detection by transformer in aerial images. TGRS, 2022." + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 316, + 228, + 553, + 270 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 228, + 553, + 270 + ], + "spans": [ + { + "bbox": [ + 316, + 228, + 553, + 270 + ], + "type": "text", + "content": "[53] Zhenhua Xu, Yuxuan Liu, Yuxiang Sun, Ming Liu, and Lujia Wang. Rngdet++: Road network graph detection by transformer with instance segmentation and multi-scale features enhancement. RA-L, 2023." + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 316, + 272, + 553, + 316 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 272, + 553, + 316 + ], + "spans": [ + { + "bbox": [ + 316, + 272, + 553, + 316 + ], + "type": "text", + "content": "[54] Zhenhua Xu, Yujia Zhang, Enze Xie, Zhen Zhao, Yong Guo, Kwan-Yee K Wong, Zhenguo Li, and Hengshuang Zhao. Drivegpt4: Interpretable end-to-end autonomous driving via large language model. RA-L, 2024." + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 316, + 317, + 553, + 358 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 317, + 553, + 358 + ], + "spans": [ + { + "bbox": [ + 316, + 317, + 553, + 358 + ], + "type": "text", + "content": "[55] Yijun Yang, Tianyi Zhou, Kanxue Li, Dapeng Tao, Lusong Li, Li Shen, Xiaodong He, Jing Jiang, and Yuhui Shi. Embodied multi-modal agent trained by an llm from a parallel textworld. In CVPR, 2024." + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 316, + 360, + 553, + 392 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 360, + 553, + 392 + ], + "spans": [ + { + "bbox": [ + 316, + 360, + 553, + 392 + ], + "type": "text", + "content": "[56] Ekim Yurtsever, Jacob Lambert, Alexander Carballo, and Kazuya Takeda. A survey of autonomous driving: Common practices and emerging technologies. IEEE Access." + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 316, + 394, + 553, + 425 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 394, + 553, + 425 + ], + "spans": [ + { + "bbox": [ + 316, + 394, + 553, + 425 + ], + "type": "text", + "content": "[57] Xiaohua Zhai, Basil Mustafa, Alexander Kolesnikov, and Lucas Beyer. Sigmoid loss for language image pre-training. In ICCV, 2023." + } + ] + } + ], + "index": 25 + }, + { + "bbox": [ + 316, + 426, + 553, + 448 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 426, + 553, + 448 + ], + "spans": [ + { + "bbox": [ + 316, + 426, + 553, + 448 + ], + "type": "text", + "content": "[58] Jimuyang Zhang, Zanming Huang, and Eshed Ohn-Bar. Coaching a teachable student. In CVPR, 2023." + } + ] + } + ], + "index": 26 + }, + { + "bbox": [ + 316, + 449, + 553, + 480 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 449, + 553, + 480 + ], + "spans": [ + { + "bbox": [ + 316, + 449, + 553, + 480 + ], + "type": "text", + "content": "[59] Jimuyang Zhang, Zanming Huang, Arjit Ray, and Eshed Ohn-Bar. Feedback-guided autonomous driving. In CVPR, 2024." + } + ] + } + ], + "index": 27 + }, + { + "bbox": [ + 316, + 482, + 553, + 514 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 482, + 553, + 514 + ], + "spans": [ + { + "bbox": [ + 316, + 482, + 553, + 514 + ], + "type": "text", + "content": "[60] Jiawei Zhang, Chejian Xu, and Bo Li. Chatscene: Knowledge-enabled safety-critical scenario generation for autonomous vehicles. In CVPR, 2024." + } + ] + } + ], + "index": 28 + }, + { + "bbox": [ + 316, + 515, + 553, + 547 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 515, + 553, + 547 + ], + "spans": [ + { + "bbox": [ + 316, + 515, + 553, + 547 + ], + "type": "text", + "content": "[61] Peiyuan Zhang, Guangtao Zeng, Tianduo Wang, and Wei Lu. Tinyllama: An open-source small language model. arXiv:2401.02385, 2024." + } + ] + } + ], + "index": 29 + }, + { + "bbox": [ + 316, + 548, + 553, + 592 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 548, + 553, + 592 + ], + "spans": [ + { + "bbox": [ + 316, + 548, + 553, + 592 + ], + "type": "text", + "content": "[62] Xiang Zhang, Haojie Sun, Xiaoyang Pei, Linghui Guan, and Zihao Wang. Evolution of technology investment and development of robotaxi services. Transportation Research Part E: Logistics and Transportation Review, 2024." + } + ] + } + ], + "index": 30 + }, + { + "bbox": [ + 316, + 593, + 553, + 624 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 593, + 553, + 624 + ], + "spans": [ + { + "bbox": [ + 316, + 593, + 553, + 624 + ], + "type": "text", + "content": "[63] Zhejun Zhang, Alexander Liniger, Dengxin Dai, Fisher Yu, and Luc Van Gool. End-to-end urban driving by imitating a reinforcement learning coach. In ICCV, 2021." + } + ] + } + ], + "index": 31 + }, + { + "bbox": [ + 316, + 626, + 553, + 667 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 626, + 553, + 667 + ], + "spans": [ + { + "bbox": [ + 316, + 626, + 553, + 667 + ], + "type": "text", + "content": "[64] Baichuan Zhou, Ying Hu, Xi Weng, Junlong Jia, Jie Luo, Xien Liu, Ji Wu, and Lei Huang. Tinyllava: A framework of small-scale large multimodal models. arXiv:2402.14289, 2024." + } + ] + } + ], + "index": 32 + }, + { + "bbox": [ + 316, + 670, + 553, + 712 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 670, + 553, + 712 + ], + "spans": [ + { + "bbox": [ + 316, + 670, + 553, + 712 + ], + "type": "text", + "content": "[65] Deyao Zhu, Jun Chen, Xiaoqian Shen, Xiang Li, and Mohamed Elhoseiny. Minigpt-4: Enhancing vision-language understanding with advanced large language models. In ICLR, 2024." + } + ] + } + ], + "index": 33 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "type": "text", + "content": "17270" + } + ] + } + ], + "index": 35 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 9 + } + ], + "_backend": "vlm", + "_version_name": "2.6.4" +} \ No newline at end of file diff --git a/2025/DriveScape_ High-Resolution Driving Video Generation by Multi-View Feature Fusion/6698f995-6b34-4c34-b673-bcdca2c64dfc_content_list.json b/2025/DriveScape_ High-Resolution Driving Video Generation by Multi-View Feature Fusion/6698f995-6b34-4c34-b673-bcdca2c64dfc_content_list.json new file mode 100644 index 0000000000000000000000000000000000000000..6cd1094ca828703c86cbf9a0c8195fba6a25b550 --- /dev/null +++ b/2025/DriveScape_ High-Resolution Driving Video Generation by Multi-View Feature Fusion/6698f995-6b34-4c34-b673-bcdca2c64dfc_content_list.json @@ -0,0 +1,1302 @@ +[ + { + "type": "text", + "text": "DriveScape: High-Resolution Driving Video Generation by Multi-View Feature Fusion", + "text_level": 1, + "bbox": [ + 202, + 130, + 795, + 174 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Wei Wu $^{1,2}$ , Xi Guo $^{2}$ , Weixuan Tang $^{2}$ , Tingxuan Huang $^{3}$ , Chiyu Wang $^{2}$ , Chenjing Ding $^{2*}$ $^{1}$ Tsinghua University $^{2}$ Sensetime Research $^{3}$ Northeastern University {wuwei, quoxi, dingchenjing} @ sensetime.com, {tangweixuan, wangchiyu} @ senseauto.com, {huangtingxuan} @stu.neu.edu.cn", + "bbox": [ + 217, + 202, + 779, + 310 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Abstract", + "text_level": 1, + "bbox": [ + 246, + 344, + 326, + 359 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Recent advancements in generative models offer promising solutions for synthesizing realistic driving videos, aiding in training autonomous driving perception models. However, existing methods often struggle with high-resolution multi-view generation, mainly due to the significant memory and computational overhead caused by simultaneously inputting multi-view videos into denoising diffusion models. In this paper, we propose a driving video generation framework based on multi-view feature fusion named DriveScape for multi-view 3D condition-guided video generation. We introduce a Bi-Directional Modulated Transformer (BiMoT) module to encode, fuse and inject multiview features along with various 3D road structures and objects, which enables high-resolution multi-view generation. Consequently, our approach allows precise control over video generation, greatly enhancing realism and providing a robust solution for creating high-quality, multi-view driving videos. Our framework achieves state-of-the-art results on the nuScenes dataset, demonstrating impressive generative quality metrics with an FID score of 8.34 and an FVD score of 76.39, as well as superior performance across various perception tasks. This lays the foundation for more accurate environment simulation in autonomous driving. Our project homepage.", + "bbox": [ + 88, + 376, + 485, + 739 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "1. Introduction", + "text_level": 1, + "bbox": [ + 91, + 771, + 223, + 787 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Autonomous driving has garnered significant attention, emphasizing the need for precise environmental understanding to ensure safe driving decisions [13, 38]. Bird's-eye view (BEV) maps, generated from multi-view images, serve as a crucial structured representation for various tasks, includ", + "bbox": [ + 89, + 797, + 483, + 875 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "ing 3D object detection, segmentation, tracking, depth estimation, and trajectory prediction [14, 15, 17, 18, 20, 34, 36, 39]. However, acquiring high-quality multi-view video data poses significant challenges due to the high labeling costs, increasing the demand for generating temporally consistent video data aligned with real-world distributions and 3D labels. Recent generative models have demonstrated the potential of synthetic data to enhance visual tasks [7, 25, 35], while diffusion models have advanced the generation of diverse and realistic driving videos, which are crucial for developing robust autonomous driving systems.", + "bbox": [ + 511, + 345, + 906, + 512 + ], + "page_idx": 0 + }, + { + "type": "table", + "img_path": "images/a08c3e9b9b415b39b3ccabdaacf4a8a3edfe270de75fbdceb21cbbea1b1677ff.jpg", + "table_caption": [], + "table_footnote": [], + "table_body": "
ModelSpatial Res.Temporal Res. (Hz)sparse condition
Drive-WM192 × 3842×
WoVoGen256 × 4482×
Delphi512 × 5122×
GenAD256 × 4482×
MagicDrive272 × 736-×
DriveDreamer 1&2448 × 2562×
DriveDiffusion512 × 5122×
Panacea256 × 5122×
Ours (DriveScape)576 × 10242 - 10✓
", + "bbox": [ + 535, + 526, + 883, + 631 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Table 1. Real-world multi-view driving controllable world models. DriveScape (ours) operates at a high spatiotemporal resolution, offering support for sparse layout controllability, as well as unified control for both mono-condition and multi-condition scenarios.", + "bbox": [ + 511, + 641, + 906, + 696 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "The most relevant approach to our work is the generation of multi-view driving videos controlled by 3D conditions. In terms of spatial resolution, general video generation tasks, such as those addressed by SVD, have achieved high-resolution video generation at $1024 \\times 576$ . However, in driving scenarios, multi-camera panoramic generation is often required. Previous methods typically need to denoise multiple views simultaneously, leading to memory constraints during both training and inference. For temporal resolution, previous methods rely on frame-by-frame 3D layout control. Due to high annotation costs, public datasets like NuScenes provide annotations at a low frame rate (2Hz), restricting models from generating videos at the", + "bbox": [ + 511, + 704, + 908, + 902 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "CVF", + "bbox": [ + 106, + 2, + 181, + 42 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore.", + "bbox": [ + 236, + 0, + 810, + 46 + ], + "page_idx": 0 + }, + { + "type": "page_footnote", + "text": "*Corresponding author", + "bbox": [ + 109, + 887, + 235, + 900 + ], + "page_idx": 0 + }, + { + "type": "page_number", + "text": "17187", + "bbox": [ + 480, + 944, + 519, + 955 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "rate (2Hz), restricting models from generating videos at the same frame rate. Additionally, the cost of obtaining different types of annotations varies in practice. For instance, annotating 3D lane lines is typically more expensive than annotating bounding boxes, often resulting in lane lines being labeled at a lower frame rate. This necessitates generative models that can handle sparse conditional control and remain compatible with both sparse and non-sparse conditions simultaneously.", + "bbox": [ + 89, + 90, + 480, + 227 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "To address these challenges, we propose DriveScape, a solution for multi-view 3D condition-guided video generation with high spatio-temporal resolution. Unlike existing methods, DriveScape integrates adjacent views and keyframes through multi-view fusion, avoiding the computational burden of denoising multiple views simultaneously. We introduce a streamlined training and inference pipeline, free from post-processing, supporting high spatial and temporal resolution for sparse condition control. To align 3D road structural information, we propose a Bi-Directional Condition Alignment module, integrating multiple conditions into an aligned latent space for fine-grained layout controllability. By leveraging temporal and spatial information, DriveScape establishes a unified model for consistent multi-view video generation, significantly enhancing realism and applicability.", + "bbox": [ + 91, + 229, + 482, + 469 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "In summary, our key contributions are as follows:", + "bbox": [ + 109, + 470, + 437, + 484 + ], + "page_idx": 1 + }, + { + "type": "list", + "sub_type": "text", + "list_items": [ + "- We propose a unified multi-view feature fusion framework for driving video generation without any post-processing. To the best of our knowledge, our model is the first to achieve high-resolution, high-frame-rate, sparse condition control for multi-view driving video generation.", + "- We introduce a Bi-Directional Condition Alignment module to facilitate diverse modalities of condition guidance and sparse condition control in video generation tasks. Experiments demonstrate that our model achieves decoupled, sparse, and precise control of dynamic foreground and static background.", + "- We achieve state-of-the-art video synthesis performance on the nuScenes dataset, excelling not only in the generation quality of multi-view images and videos but also across various perception applications." + ], + "bbox": [ + 91, + 487, + 482, + 728 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "2. Related Works", + "text_level": 1, + "bbox": [ + 89, + 744, + 238, + 762 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "2.1. Controllable generation", + "text_level": 1, + "bbox": [ + 89, + 771, + 308, + 787 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "With the advent of diffusion models, significant progress has been made in text-to-video generation [1, 2, 5, 8, 10, 12, 24, 30, 40]. Video LDM [4] employs a latent diffusion pipeline where denoising occurs in the image latent space, greatly accelerating the process. However, text alone cannot precisely control video generation. Later methods introduced image blocks alongside text as prompts for the", + "bbox": [ + 89, + 794, + 482, + 900 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "denoising network [37]. Our goal is to generate highly realistic street-view videos, a complex scene featuring numerous elements and interactions (e.g., intricate street layouts and moving vehicles). To achieve fine-grained control, we incorporate road maps, 3D bounding boxes, and BEV keyframes, going beyond text and images to ensure detailed and precise video generation.", + "bbox": [ + 511, + 90, + 903, + 196 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "2.2. Multi-view video generation", + "text_level": 1, + "bbox": [ + 511, + 207, + 764, + 223 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Multi-view consistency and temporal consistency are two critical challenges in multi-view video generation. To address multi-view consistency, MVDiffusion [26] introduced a correspondence-aware attention module to align information across views. Tseng et al. [27] utilized epipolar geometry to regularize consistency between different views. MagicDrive [7] leveraged camera poses, bounding boxes, and road maps as priors, incorporating an additional cross-view attention block to enhance consistency. However, these methods are limited to generating multi-view images rather than videos.", + "bbox": [ + 511, + 229, + 906, + 395 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "2.3. Street view generation", + "text_level": 1, + "bbox": [ + 511, + 407, + 720, + 422 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Most street-view generation models condition on 2D layouts, such as BEV maps, 2D bounding boxes, and semantic segmentation. BEVGen [25] encapsulates all semantics within BEV for street-view generation. BEVControl [35] employs a two-stage method to generate multi-view urban scene images from BEV layouts, where a controller generates foreground and background objects, and a coordinator combines them to maintain visual consistency across views. However, projecting 3D information to 2D loses geometric details, making direct extension to video generation prone to inconsistencies across frames. To address this, we incorporate 3D bounding boxes as a condition to guide generation.DrivingDiffusion [16] adopts a multi-stage pipeline with two models and two stages of post-processing to generate frames and extend them into videos. In contrast, our method avoids the complexity of multi-stage pipelines by using an efficient, end-to-end approach.", + "bbox": [ + 511, + 429, + 906, + 686 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "3. Method", + "text_level": 1, + "bbox": [ + 511, + 700, + 604, + 715 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "3.1. Overview", + "text_level": 1, + "bbox": [ + 511, + 726, + 624, + 741 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "The overview of DriveScape is shown in Fig. 1. Based on the LDM pipeline [22], DriveScape generates street-view videos conditioned on scene annotations, such as BEV maps, 3D bounding boxes, ego states, and text for each view. The unified model, described in Section 3.2, enables high-resolution multi-view video generation without post-refinement.", + "bbox": [ + 511, + 750, + 903, + 852 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "The Bi-Directional Condition Alignment module, detailed in Section 3.5, ensures effective alignment between 3D road structures and temporal sparse conditions.", + "bbox": [ + 511, + 854, + 903, + 900 + ], + "page_idx": 1 + }, + { + "type": "page_number", + "text": "17188", + "bbox": [ + 480, + 944, + 517, + 955 + ], + "page_idx": 1 + }, + { + "type": "image", + "img_path": "images/49b1dfe8d80e4b6efba66d7778586e6a150f77efc9405d2da381e86e00bfe69f.jpg", + "image_caption": [ + "Figure 1. End-to-End Multi-view Video Generation pipeline. We use learnable embedding vectors to represent different cameras and categorize camera views as key views and neighbour views. Our training scheme guides the generation of key view videos through their neighboring frames. Additionally, we introduce the key-frames condition along with a training and inference scheme to ensure multi-view consistency simultaneously. Furthermore, our model does not require any post-refine process as DriveDiffusion ([16]). It is capable of learning multi-view and temporal consistency simultaneously, resulting in high-fidelity street-view synthesis." + ], + "image_footnote": [], + "bbox": [ + 133, + 88, + 867, + 375 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3.2. Unified Model", + "text_level": 1, + "bbox": [ + 89, + 481, + 236, + 494 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Our model leverages SVD [3] as the backbone, with each denoising step processing a video from a single viewpoint. Control conditions are injected through a unified control module, including frame-wise images from adjacent viewpoints, road maps, 3D bounding boxes, and the first frame from each viewpoint. Starting with initial multi-view frames $x_{0:N-1}^{0}(N$ is the number of views), DriveScape predicts subsequent frames $x_{0:N-1}^{1:T}(T$ is the number of frames) under conditions such as BEV maps $m_{0:N-1}^{0:T}$ , 3D bounding boxes $b_{0:N-1}^{0:T}$ , BEV key frames, and neighboring camera videos. BEV maps and 3D bounding boxes are projected into the camera's First-Person View (FPV) and encoded with convolutional blocks.", + "bbox": [ + 88, + 503, + 482, + 700 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "As shown in Fig. 1, we use learnable embedding vectors $\\{e_i\\}$ to represent different cameras. Similar to the processing of canbus data (e.g., velocity and direction angle), these embeddings are passed through a multilayer perceptron (MLP) followed by a Sigmoid Linear Unit (SiLU) activation. This enables the embeddings to perform crossattention with the image latent space, allowing the unified model to predict videos from multiple perspectives.", + "bbox": [ + 88, + 702, + 482, + 824 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Embedding camera information alone cannot ensure spatial and temporal consistency, as it only captures global camera characteristics. Consistency is achieved through model design and training. For temporal consistency, we integrate temporal attention into the latent diffusion UNet [23]", + "bbox": [ + 88, + 825, + 482, + 901 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "and control alignment module (Section 3.5). For multi-view consistency, we introduce keyframe conditions and a carefully designed training and inference scheme.", + "bbox": [ + 511, + 481, + 906, + 527 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3.3. Training Stage", + "text_level": 1, + "bbox": [ + 511, + 542, + 663, + 558 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Neighbor frames feature fusion. DriveScape categorizes camera views into key and neighbor views (Fig. 2). Frames of the same type have minimal overlap, ensuring distinct regions. The training scheme guides key view generation using neighbor frames. During training, neighbor and key views are selected sequentially. For neighbor views, video conditions from adjacent key views are input into the network. Key views are generated sequentially, stored, and incorporated into the cross-attention mechanism with the image latent space, ensuring effective guidance for neighbor view generation.", + "bbox": [ + 511, + 566, + 906, + 733 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Multi-view key-frame feature fusion. A challenge arises from the lack of specific multi-view constraints during key-view generation, leading to degraded performance. While this is generally acceptable due to minimal overlapping regions, long-term generation introduces interdependencies among views. For example, during overtaking, a vehicle initially captured by the rear camera later appears in the front camera. To address this, we propose key-frame conditions, which are the starting frames of all views, to establish correlations within the image latent space. Combined with temporal attention, the model learns the relationships across", + "bbox": [ + 511, + 734, + 908, + 902 + ], + "page_idx": 2 + }, + { + "type": "page_number", + "text": "17189", + "bbox": [ + 480, + 944, + 519, + 955 + ], + "page_idx": 2 + }, + { + "type": "image", + "img_path": "images/6bc43f600d0da8e506cfdfe42a49ad5713f793b7e007bf4805b60fbd4652ffc8.jpg", + "image_caption": [ + "Figure 2. The DriveScape pipeline, built on the LDM framework, generates street-view videos conditioned on scene annotations, BEV maps, and 3D bounding boxes for each view. Keyframe conditions are introduced, in collaboration with temporal and spatial self-attention modules, to ensure consistency across spatial and temporal dimensions. Our unified model enables multi-view video generation without the need for complex post-processing or refinement procedures. Additionally, the Bi-Directional Modulated Transformer (BiMoT) module, consisting of two cross-attention layers with opposite directions and a temporal self-attention layer, ensures effective alignment and integration of various 3D road structural information, enabling precise control over video generation." + ], + "image_footnote": [], + "bbox": [ + 133, + 88, + 867, + 375 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "all views at the initial frame and the preceding frames of neighboring cameras.", + "bbox": [ + 88, + 494, + 482, + 526 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Sparse multi-modal feature fusion. Our model supports training with videos at various frame rates, ranging from 2Hz to 10Hz, while maintaining sample conditions at 2Hz. A lightweight encoder is used to encode the 2Hz map and bounding boxes (BBoxes), with condition-missing frames padded with zeros. Additionally, an extra mask channel is introduced to indicate the presence of conditions at each frame, where mask values are 0 or 1. These inputs are processed through the BiMoT module. Embeddings for unconditioned frames are learned and integrated into the conditions to align with the video frame rate. This diverse frame-rate training scheme, combined with the BiMoT module, enhances sparse condition control capability.", + "bbox": [ + 88, + 527, + 482, + 724 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Conditions dropout. To enhance robustness, key views are trained in random order. Neighbor camera videos and conditions are dropped with probabilities of $50\\%$ and $20\\%$ , respectively. Unlike [16], our model requires no postrefinement, simultaneously learning multi-view and temporal consistency to improve quality and realism.", + "bbox": [ + 88, + 726, + 482, + 816 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "3.4. Inference Stage", + "text_level": 1, + "bbox": [ + 89, + 830, + 246, + 848 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "During inference, our pipeline enables efficient and consistent generation of multi-view videos. Similar to the training stage, the process consists of two patterns: key-view infer", + "bbox": [ + 89, + 854, + 482, + 901 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "ence and neighbor-view inference, as shown in Fig. 2. Keyview camera videos are generated simultaneously without cross-view condition control, as the key-view cameras are intentionally placed non-adjacently. For example, one camera is positioned at the front, while two are at the rear—one on the left and the other on the right. This strategic arrangement ensures comprehensive coverage from multiple perspectives with minimal overlap.", + "bbox": [ + 511, + 494, + 903, + 618 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "In this configuration, neighbor-view cameras—two at the front and one at the rear—naturally receive cross-view conditional information from their adjacent key-view cameras. This design allows neighbor-view inference to maintain consistency with conditions from neighboring cameras, ensuring cross-view coherence.", + "bbox": [ + 511, + 619, + 905, + 708 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Leveraging this setup, our inference pipeline excels in parallel processing, requiring only two consecutive iterations to generate multi-view videos. This approach not only accelerates video production but also ensures cross-view consistency and realism throughout.", + "bbox": [ + 511, + 710, + 905, + 787 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "3.5. Bi-Directional Modulated Transformer", + "text_level": 1, + "bbox": [ + 511, + 801, + 848, + 816 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "To integrate multi-conditions from different modalities into the UNet model and ensure effective articulation of each condition, we propose the Bi-Directional Modulated Transformer (BiMoT) module. This module aligns features from various modalities within the latent space, reinforces tem", + "bbox": [ + 511, + 825, + 903, + 900 + ], + "page_idx": 3 + }, + { + "type": "page_number", + "text": "17190", + "bbox": [ + 480, + 944, + 517, + 955 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "poral coherence among the merged conditions, and injects the aligned conditions into the network. As illustrated in Fig. 1, the module consists of two cross-attention layers with opposite directions and a temporal self-attention layer.", + "bbox": [ + 89, + 90, + 480, + 151 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Latent to Condition Attention To address diverse conditions with different modalities, we employ a latent-to-condition cross-attention mechanism to standardize feature representations. Pixel-level features are used as signals to merge various condition features. Specifically, let $F_{cond}$ represent condition features and $F_{latent}$ represent video latent features. The output is computed as:", + "bbox": [ + 89, + 152, + 480, + 257 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\nF _ {c o n d} ^ {\\prime} = \\operatorname {c r o s s a t t n} \\left(F _ {c o n d}, F _ {l a t e n t}\\right), \\tag {1}\n$$\n", + "text_format": "latex", + "bbox": [ + 169, + 268, + 482, + 290 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "where the query is $F_{cond}$ , and the key and value are $F_{latent}$ . This cross-attention operation facilitates the flow of video latent information into the condition space, providing a unified representation for integrating all modalities.", + "bbox": [ + 89, + 301, + 480, + 361 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Temporal Attention. We apply temporal self-attention to ensure that the merged features propagate effectively along the temporal dimension. Sparse condition information is also incorporated into the image latent space, naturally unifying all modalities into a single representation. This enables temporal information from latent features to interact with sparse conditions over time, guiding the generation process with sparse control. This mechanism ensures consistency and precise control over the temporal dimensions.", + "bbox": [ + 89, + 362, + 480, + 497 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Condition to Video Attention. Similar to the latent-to-condition cross-attention, a condition-to-video cross-attention mechanism is applied:", + "bbox": [ + 89, + 500, + 480, + 544 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\nF _ {l a t e n t} ^ {\\prime} = \\operatorname {c r o s s a t t n} \\left(F _ {l a t e n t}, F _ {c o n d}\\right). \\tag {2}\n$$\n", + "text_format": "latex", + "bbox": [ + 166, + 556, + 482, + 575 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "This operation facilitates the flow of merged condition feature information into the image latents, ultimately enabling unified control over video generation.", + "bbox": [ + 89, + 587, + 480, + 632 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Coarse to Fine Condition Mixture. BiMoT is flexible in accommodating various types of conditions, enabling easy extension to additional inputs. We begin with maps and 3D layouts, then introduce BEV key-frames, and finally integrate neighbor camera videos. This progression allows generation control to transition from a coarse level, driven by maps and 3D layouts, to a finer level, guided by pixel-level details. This ensures both structural coherence and detailed fidelity. The flexibility of BiMoT enables fine-grained layout controllability with unified mono-condition and multi-condition control.", + "bbox": [ + 89, + 633, + 482, + 797 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "4. Experiments", + "text_level": 1, + "bbox": [ + 89, + 814, + 220, + 830 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "In this section, we evaluate our model through key metrics, outlining the experimental setup, dataset, and evaluation criteria. We then analyze qualitative results and conduct ablation studies to validate our approach.", + "bbox": [ + 89, + 839, + 480, + 900 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "4.1. Experiment Details", + "text_level": 1, + "bbox": [ + 513, + 90, + 697, + 106 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Dataset. The training data is sourced from the real-world driving dataset nuScenes [6], a widely used benchmark for BEV segmentation and 3D object detection in driving scenarios. Following the official configuration, we use 700 street-view scenes for training and 150 for validation, targeting 8 object classes and 8 road classes.", + "bbox": [ + 511, + 112, + 903, + 202 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Evaluation Metrics. To evaluate the effectiveness of our approach in delivering realism, continuity, and precise control over multiple conditions, we employ four key metrics for comparison against existing multi-view image and video generation methods. For realism, we use the Fréchet Inception Distance (FID) [11] to measure image quality. Temporal coherence is assessed using the Fréchet Video Distance (FVD) [28], ensuring consistent and fluid motion in the generated videos. Controllability is evaluated through two perception-based methods: CVT [39] for BEV segmentation and BEVFusion [19] for 3D object detection. Both methods are trained on the nuScenes dataset, providing a reliable benchmark for performance evaluation. Additionally, since our model generates videos, we follow Panacea [33] and use the video-based perception model StreamPETR [29] to evaluate the 3D object detection task.", + "bbox": [ + 511, + 203, + 903, + 445 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "4.2. Experiments Results", + "text_level": 1, + "bbox": [ + 513, + 455, + 707, + 470 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Qualitative Results. We showcase the quality and consistency of our video generation results in Fig. 3. Our model achieves notable consistency across both frames and viewpoints. Specifically, vehicles are accurately positioned according to the provided 3D layouts, and street delineations closely match the supplied projected maps. Remarkably, even when bounding boxes extend beyond the visible area, our model adheres to the specified conditions and generates the remaining objects with high fidelity.", + "bbox": [ + 511, + 477, + 903, + 613 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Sparse Condition Control As shown in Fig. 5, our model generates consistent videos in both time and space under sparse conditions, achieving a resolution of $576 \\times 1024$ and a frame rate of 10 fps. To compare with related works on sparse control [9], we implemented a sparse control model with multi-condition fusion, as existing methods lack support for multi-condition control. A visualization comparing the reproduced SparseCtrl [9] and our method is presented in Fig. 4. Our method clearly demonstrates superior spatial consistency. Moreover, the FID and FVD scores for the reproduced SparseCtrl are 11.15 and 116.00, respectively, significantly worse than those achieved by our model (8.34 and 76.39). For additional visual comparisons and details on the multi-condition fusion SparseCtrl, please refer to the supplementary materials and the accompanying web page.", + "bbox": [ + 511, + 613, + 903, + 840 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Quantitative Results. We report quantitative metrics on the nuScenes validation set in Table 2. Using FID and FVD, we evaluate image and video quality, respectively. Following MagicDrive [7] and DrivingDiffusion [16], we", + "bbox": [ + 511, + 840, + 903, + 900 + ], + "page_idx": 4 + }, + { + "type": "page_number", + "text": "17191", + "bbox": [ + 480, + 944, + 517, + 955 + ], + "page_idx": 4 + }, + { + "type": "table", + "img_path": "images/f7b62ef89dfcd47602eae26a73d1501b9a138f0d8cdf26b3ae3047913cd962a3.jpg", + "table_caption": [], + "table_footnote": [], + "table_body": "
MethodFID↓FVD↓Road mIoU ↑Vehicle mIoU ↑Object NDS ↑
Oracle--73.6834.8441.21
BEVGen[25]25.54-50.20 (-23.48)5.89 (-28.95)-
BEVControl [35]24.85-60.80 (-12.88)26.80 (-8.04)-
MagicDrive [7]16.20-61.05 (-12.63)27.01 (-7.83)23.32 (-17.89)
DriveDreamer [31]52.60452.0---
DrivingDiffusion [16]15.83332.063.20 (-10.48)31.60 (-3.24)33.1 (-8.11)
Drive-WM [32]15.80122.7065.07 (-8.61)27.19 (-7.65)-
Delphi [21]15.08113.5--36.58 (-4.63)
SparseCtrl[9]11.15116.00---
Ours8.3476.3968.21 (-5.47)30.27 (-4.57)36.5 (-4.71)
", + "bbox": [ + 179, + 87, + 823, + 260 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Table 2. The comparative analysis of our method against others on the nuScenes validation dataset is detailed for each task. Our end-to-end approach outperformed most baseline models without requiring any post-refinement process.", + "bbox": [ + 89, + 270, + 906, + 299 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/1e538604e8c82383bd5b8a362be456dd51e85689478209710856b9e55e99afd1.jpg", + "image_caption": [ + "Figure 3. Showcases for multi-view video generation by DriveScape demonstrate frame-to-frame and viewpoint consistency. Vehicles generated by our model are accurately positioned according to the provided 3D layouts, while the street depictions closely align with the projected maps." + ], + "image_footnote": [], + "bbox": [ + 117, + 311, + 880, + 604 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/6f481d41397efc002e0ab76879aaeaff7071e2d6825cfcd60703a6388f91a836.jpg", + "image_caption": [ + "Figure 4. Results in each row are Ours, SparseCtrl, input condition and GT respectively, which illustrate that our method has better spatial consistency than SparseCtrl." + ], + "image_footnote": [], + "bbox": [ + 91, + 686, + 486, + 803 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "generate data from validation set labels and assess it with a pre-trained perception model. The decline in Road mIoU", + "bbox": [ + 89, + 869, + 483, + 902 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "and Vehicle mIoU, compared to original data, measures the effectiveness of our framework in producing perception-friendly synthetic data.", + "bbox": [ + 511, + 689, + 906, + 733 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "To align with previous studies, we resized generated images to $224 \\times 400$ . We compared our FID results with three image generation models: BEVGen [25], BEVControl [35], and MagicDrive [7]. Our method significantly outperformed these models. Notably, video generation models typically yield higher FID scores than image models. For instance, Drive-WM, using only an image model, achieved better FID than the full model, highlighting the challenges of video generation.", + "bbox": [ + 511, + 734, + 908, + 869 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "For video quality, assessed via the FVD metric, our approach showed substantial improvements over recent works", + "bbox": [ + 511, + 869, + 906, + 902 + ], + "page_idx": 5 + }, + { + "type": "page_number", + "text": "17192", + "bbox": [ + 480, + 944, + 519, + 957 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/fc1328671988eaa6e2b3787d4b8518b4976c53b1944803ee87c2d7fe01020ea6.jpg", + "image_caption": [ + "Figure 5. Showcase for sparse Condition Control." + ], + "image_footnote": [], + "bbox": [ + 98, + 89, + 478, + 316 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "such as DriveDreamer [31], DrivingDiffusion [16], and Drive-WM [32]. In BEV segmentation metrics, our method achieved superior perceptual performance, except for a slightly lower Road mIoU compared to DrivingDiffusion, which utilized a post-processing network for refinement. To maintain simplicity, our end-to-end system applies no post-processing beyond image resizing for perceptual evaluation. The perception models we used operate on single-image inference, not leveraging the temporal consistency of our generated videos. Given our significant lead in FVD, our method holds strong potential for training video BEV perception models.", + "bbox": [ + 89, + 357, + 483, + 540 + ], + "page_idx": 6 + }, + { + "type": "table", + "img_path": "images/c6b4307be11c23f07a633bfc78fa69b423663bd96d9411df9aeb7da1f18e1d5a.jpg", + "table_caption": [], + "table_footnote": [], + "table_body": "
dataNDS ↑FID ↓FVD ↓
Oracle46.9--
Panacea32.1 (-14.8)16.96139
Oracle53.7--
Ours44.5 (-9.2)8.3476.39
", + "bbox": [ + 138, + 551, + 434, + 633 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "4.3. Ablation Studies", + "text_level": 1, + "bbox": [ + 89, + 681, + 254, + 696 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "To refine and evaluate the efficacy of our proposed framework, we conducted ablation studies, as detailed in Table 4. Various strategies for implementing conditional guidance were tested. Removing the entire BiMoT module resulted in the FID increasing from 8.34 to 20.98 and the FVD rising from 76.39 to 142.81, underscoring the module's importance for multi-condition encoding. Similarly, removing the temporal attention layer within BiMoT caused a significant performance drop, with FID and FVD increasing to 20.67 and 123.03, respectively, highlighting the role of temporal modelling in BiMoT. Additionally, during testing, we separately removed the key-frame condition and the neighbor condition. Both led to a performance decline, indicating the", + "bbox": [ + 89, + 704, + 483, + 900 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "importance of these conditions for facilitating information exchange between views.", + "bbox": [ + 511, + 90, + 906, + 121 + ], + "page_idx": 6 + }, + { + "type": "table", + "img_path": "images/d2e976422d91156f1cb968661638bbaf32f000f6f21cf73380cd2a4732b6852e.jpg", + "table_caption": [ + "Table 3. Compared with Panacea, we evaluate results on the nuScenes validation set using the pre-trained StreamPETR model." + ], + "table_footnote": [], + "table_body": "
TeAttnBiMOTKFNeFIDFVD
✓✓✓20.98142.81
✓✓20.67123.03
✓✓✓11.60108.80
✓✓✓12.18111.36
✓✓✓✓8.3476.39
", + "bbox": [ + 540, + 133, + 880, + 229 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Table 4. Ablation Study. TeAttn is temporal attention module in BiMOT, KF is key-frame condition and Ne is neighbor condition", + "bbox": [ + 511, + 239, + 906, + 268 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "4.4. Visualization of Different Settings", + "text_level": 1, + "bbox": [ + 511, + 294, + 808, + 311 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "In our visual showcases, we demonstrate the model's proficiency in manipulating video predictions for a single scene under various control conditions. The control conditions, sourced from frames of identical perspectives but randomly selected from other scenes, test the decoupling of conditions and the robustness of our method across diverse scenarios.", + "bbox": [ + 511, + 316, + 903, + 406 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "As shown in Fig. 7, the experiments explore the influence of different 3D layout controls. Using a single reference frame from a highway segment, we first depict vehicles driving normally, then transition to vehicles traversing grassy terrain. Remarkably, our model maintains temporal consistency even in this unconventional context. Additional experiments involve removing or adding large vehicles, with results aligning closely with the specified conditions, showcasing precise control.", + "bbox": [ + 511, + 407, + 905, + 542 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Furthermore, as illustrated in Fig. 8, we investigate the impact of different maps on video generation. The generated scenes align impressively with the shapes of the provided map regions, maintaining layout feature consistency despite map changes. These experiments provide strong evidence of our model's capability to render specified control conditions into precise and coherent visual narratives within the generated video content.", + "bbox": [ + 511, + 544, + 906, + 664 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "5. Conclusions", + "text_level": 1, + "bbox": [ + 511, + 679, + 638, + 694 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "The paper introduces DriveScape, the first end-to-end multi-view driving video generative model. It enables consistent multi-view video generation without complex postprocessing or refinement. This unified framework streamlines multi-view condition-guided video generation. We also propose a Bi-Directional Modulated Transformer module for diverse condition guidance. Our approach achieves decoupled and precise control of dynamic foreground and static background, delivering state-of-the-art video synthesis on the nuScenes dataset. It excels in multi-view image and video generation quality and perception applications. In summary, our method outperforms others in generation quality and ease of application.", + "bbox": [ + 511, + 704, + 906, + 900 + ], + "page_idx": 6 + }, + { + "type": "page_number", + "text": "17193", + "bbox": [ + 480, + 944, + 517, + 955 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/b565b73afdef9a50ea2277898823f855298278724094fe607063f07c0966b9f0.jpg", + "image_caption": [ + "Figure 6. Showcase for maps and 3D layouts control under different conditions. (a): control by the reference maps and 3D Layouts; (b) change 3D Layouts with more vehicles; (c) change the maps with curved pedestrian crosswalk; (d) remove all the 3D Layouts." + ], + "image_footnote": [], + "bbox": [ + 117, + 103, + 883, + 287 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/8c640171de84fd850e911dc18e017793ae07b0d09d31c897555c874686e42fc4.jpg", + "image_caption": [ + "Figure 7. Showcases for 3D layouts control under different conditions. DriveScape can precisely control dynamic objects. (a) reference layouts; (b) move car to the grass; (c) remove all cars; (d) add trucks to the scene." + ], + "image_footnote": [], + "bbox": [ + 114, + 359, + 880, + 559 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/0812ffb9511188af0cf2e05b163d70fb142cf525153a32c4b5a3203d8538d3ba.jpg", + "image_caption": [ + "Figure 8. Showcases for maps control under. DriveScape can precisely control static background. (a) replace the friable area with green belts; (b) replace the friable area with green belts; (c) replace the friable area with sidewalks; (d) add island on the road." + ], + "image_footnote": [], + "bbox": [ + 114, + 641, + 883, + 840 + ], + "page_idx": 7 + }, + { + "type": "page_number", + "text": "17194", + "bbox": [ + 480, + 944, + 519, + 955 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "References", + "text_level": 1, + "bbox": [ + 91, + 90, + 187, + 104 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[1] Jie An, Songyang Zhang, Harry Yang, Sonal Gupta, Jia-Bin Huang, Jiebo Luo, and Xi Yin. Latent-shift: Latent diffusion with temporal shift for efficient text-to-video generation. arXiv preprint arXiv:2304.08477, 2023. 2", + "[2] Andreas Blattmann, Tim Dockhorn, Sumith Kulal, Daniel Mendelevitch, Maciej Kilian, Dominik Lorenz, Yam Levi, Zion English, Vikram Voleti, Adam Letts, Varun Jampani, and Robin Rombach. Stable video diffusion: Scaling latent video diffusion models to large datasets, 2023. 2", + "[3] Andreas Blattmann, Tim Dockhorn, Sumith Kulal, Daniel Mendelevitch, Maciej Kilian, Dominik Lorenz, Yam Levi, Zion English, Vikram Voleti, Adam Letts, et al. Stable video diffusion: Scaling latent video diffusion models to large datasets. arXiv preprint arXiv:2311.15127, 2023. 2", + "[4] Andreas Blattmann, Robin Rombach, Huan Ling, Tim Dockhorn, Seung Wook Kim, Sanja Fidler, and Karsten Kreis. Align your latents: High-resolution video synthesis with latent diffusion models. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 22563-22575, 2023. 2", + "[5] Andreas Blattmann, Robin Rombach, Huan Ling, Tim Dockhorn, Seung Wook Kim, Sanja Fidler, and Karsten Kreis. Align your latents: High-resolution video synthesis with latent diffusion models. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 22563-22575, 2023. 2", + "[6] Holger Caesar, Varun Bankiti, Alex H Lang, Sourabh Vora, Venice Erin Liong, Qiang Xu, Anush Krishnan, Yu Pan, Giancarlo Baldan, and Oscar Beijbom. nuscenes: A multimodal dataset for autonomous driving. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 11621-11631, 2020. 5", + "[7] Ruiyuan Gao, Kai Chen, Enze Xie, Lanqing Hong, Zhenguo Li, Dit-Yan Yeung, and Qiang Xu. Magicdrive: Street view generation with diverse 3d geometry control. arXiv preprint arXiv:2310.02601, 2023. 1, 2, 5, 6", + "[8] Yuwei Guo, Ceyuan Yang, Anyi Rao, Yaohui Wang, Yu Qiao, Dahua Lin, and Bo Dai. Animatediff:Animate your personalized text-to-image diffusion models without specific tuning. arXiv preprint arXiv:2307.04725, 2023. 2", + "[9] Yuwei Guo, Ceyuan Yang, Anyi Rao, Maneesh Agrawala, Dahua Lin, and Bo Dai. Sparsectrl: Adding sparse controls to text-to-video diffusion models. In European Conference on Computer Vision, pages 330-348. Springer, 2025. 5, 6", + "[10] Yingqing He, Tianyu Yang, Yong Zhang, Ying Shan, and Qifeng Chen. Latent video diffusion models for high-fidelity video generation with arbitrary lengths. arXiv preprint arXiv:2211.13221, 2022. 2", + "[11] Martin Heusel, Hubert Ramsauer, Thomas Unterthiner, Bernhard Nessler, and Sepp Hochreiter. Gans trained by a two time-scale update rule converge to a local nash equilibrium. Advances in neural information processing systems, 30, 2017. 5", + "[12] Jonathan Ho, William Chan, Chitwan Sahara, Jay Whang, Ruiqi Gao, Alexey Gritsenko, Diederik P Kingma, Ben Poole, Mohammad Norouzi, David J Fleet, et al. Imagen" + ], + "bbox": [ + 93, + 114, + 483, + 900 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "video: High definition video generation with diffusion models. arXiv preprint arXiv:2210.02303, 2022. 2", + "[13] Yihan Hu, Jiazhi Yang, Li Chen, Keyu Li, Chonghao Sima, Xizhou Zhu, Siqi Chai, Senyao Du, Tianwei Lin, Wenhai Wang, et al. Goal-oriented autonomous driving. arXiv preprint arXiv:2212.10156, 2022. 1", + "[14] Junjie Huang and Guan Huang. Bevdet4d: Exploit temporal cues in multi-camera 3d object detection. arXiv preprint arXiv:2203.17054, 2022. 1", + "[15] Junjie Huang, Guan Huang, Zheng Zhu, Yun Ye, and Dalong Du. Bevdet: High-performance multi-camera 3d object detection in bird-eye-view. arXiv preprint arXiv:2112.11790, 2021. 1", + "[16] Xiaofan Li, Yifu Zhang, and Xiaoqing Ye. Drivingdiffusion: Layout-guided multi-view driving scene video generation with latent diffusion model. arXiv preprint arXiv:2310.07771, 2023. 2, 3, 4, 5, 6", + "[17] Yinhao Li, Zheng Ge, Guanyi Yu, Jinrong Yang, Zengran Wang, Yukang Shi, Jianjian Sun, and Zeming Li. Bevdepth: Acquisition of reliable depth for multi-view 3d object detection. In Proceedings of the AAAI Conference on Artificial Intelligence, pages 1477-1485, 2023. 1", + "[18] Zhiqi Li, Wenhai Wang, Hongyang Li, Enze Xie, Chonghao Sima, Tong Lu, Yu Qiao, and Jifeng Dai. Bevformer: Learning bird's-eye-view representation from multi-camera images via spatiotemporal transformers. In European conference on computer vision, pages 1-18. Springer, 2022. 1", + "[19] Tingting Liang, Hongwei Xie, Kaicheng Yu, Zhongyu Xia, Zhiwei Lin, Yongtao Wang, Tao Tang, Bing Wang, and Zhi Tang. Bevfusion: A simple and robust lidar-camera fusion framework. Advances in Neural Information Processing Systems, 35:10421-10434, 2022. 5", + "[20] Haicheng Liao, Yongkang Li, Zhenning Li, Chengyue Wang, Zhiyong Cui, Shengbo Eben Li, and Chengzhong Xu. A cognitive-based trajectory prediction approach for autonomous driving. arXiv preprint arXiv:2402.19251, 2024. 1", + "[21] Enhui Ma, Lijun Zhou, Tao Tang, Zhan Zhang, Dong Han, Junpeng Jiang, Kun Zhan, Peng Jia, Xianpeng Lang, Haiyang Sun, et al. Unleashing generalization of end-to-end autonomous driving with controllable long video generation. arXiv preprint arXiv:2406.01349, 2024. 6", + "[22] Robin Rombach, Andreas Blattmann, Dominik Lorenz, Patrick Esser, and Björn Ommer. High-resolution image synthesis with latent diffusion models. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 10684-10695, 2022. 2", + "[23] Olaf Ronneberger, Philipp Fischer, and Thomas Brox. U-net: Convolutional networks for biomedical image segmentation. In Medical Image Computing and Computer-Assisted Intervention-MICCAI 2015: 18th International Conference, Munich, Germany, October 5-9, 2015, Proceedings, Part III 18, pages 234-241. Springer, 2015. 3", + "[24] Uriel Singer, Adam Polyak, Thomas Hayes, Xi Yin, Jie An, Songyang Zhang, Qiyuan Hu, Harry Yang, Oron Ashual, Oran Gafni, Devi Parikh, Sonal Gupta, and Yaniv Taigman. Make-a-video: Text-to-video generation without text-video data, 2022. 2" + ], + "bbox": [ + 516, + 92, + 906, + 898 + ], + "page_idx": 8 + }, + { + "type": "page_number", + "text": "17195", + "bbox": [ + 480, + 944, + 517, + 955 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[25] Alexander Swerdlow, Runsheng Xu, and Bolei Zhou. Street-view image generation from a bird's-eye view layout. IEEE Robotics and Automation Letters, 2024. 1, 2, 6", + "[26] Shitao Tang, Fuyang Zhang, Jiacheng Chen, Peng Wang, and Yasutaka Furukawa. Mvdiffusion: Enabling holistic multiview image generation with correspondence-aware diffusion, 2023. 2", + "[27] Hung-Yu Tseng, Qinbo Li, Changil Kim, Suhib Alsisan, Jia-Bin Huang, and Johannes Kopf. Consistent view synthesis with pose-guided diffusion models. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 16773-16783, 2023. 2", + "[28] Thomas Unterthiner, Sjoerd Van Steenkiste, Karol Kurach, Raphael Marinier, Marcin Michalski, and Sylvain Gelly. Towards accurate generative models of video: A new metric & challenges. arXiv preprint arXiv:1812.01717, 2018. 5", + "[29] Shihao Wang, Yingfei Liu, Tiancai Wang, Ying Li, and Xi-angyu Zhang. Exploring object-centric temporal modeling for efficient multi-view 3d object detection. In CVPR, pages 3621–3631, 2023. 5", + "[30] Wenjing Wang, Huan Yang, Zixi Tuo, Huiguo He, Junchen Zhu, Jianlong Fu, and Jiaying Liu. Videofactory: Swap attention in spatiotemporal diffusions for text-to-video generation, 2023. 2", + "[31] Xiaofeng Wang, Zheng Zhu, Guan Huang, Xinze Chen, and Jiwen Lu. Drivedreamer: Towards real-world-driven world models for autonomous driving. arXiv preprint arXiv:2309.09777, 2023. 6", + "[32] Yuqi Wang, Jiawei He, Lue Fan, Hongxin Li, Yuntao Chen, and Zhaoxiang Zhang. Driving into the future: Multiview visual forecasting and planning with world model for autonomous driving. arXiv preprint arXiv:2311.17918, 2023. 6", + "[33] Yuqing Wen, Yucheng Zhao, Yingfei Liu, Fan Jia, Yanhui Wang, Chong Luo, Chi Zhang, Tiancai Wang, Xiaoyan Sun, and Xiangyu Zhang. Panacea: Panoramic and controllable video generation for autonomous driving. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 6902-6912, 2024. 5", + "[34] Chenyu Yang, Yuntao Chen, Hao Tian, Chenxin Tao, Xizhou Zhu, Zhaoxiang Zhang, Gao Huang, Hongyang Li, Yu Qiao, Lewei Lu, et al. Bevformer v2: Adapting modern image backbones to bird's-eye-view recognition via perspective supervision. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 17830-17839, 2023. 1", + "[35] Kairui Yang, Enhui Ma, Jibin Peng, Qing Guo, Di Lin, and Kaicheng Yu. Bevcontrol: Accurately controlling street-view elements with multi-perspective consistency via bev sketch layout. arXiv preprint arXiv:2308.01661, 2023. 1, 2, 6", + "[36] Yuxiang Yang, Yingqi Deng, Jing Zhang, Jiahao Nie, and Zheng-Jun Zha. Bevtrack: A simple and strong baseline for 3d single object tracking in bird's-eye view. arXiv e-prints, pages arXiv-2309, 2023. 1", + "[37] David Junhao Zhang, Dongxu Li, Hung Le, Mike Zheng Shou, Caiming Xiong, and Doyen Sahoo. Moonshot: To" + ], + "bbox": [ + 91, + 92, + 480, + 900 + ], + "page_idx": 9 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "wards controllable video generation and editing with multimodal conditions, 2024. 2", + "[38] Wenzhao Zheng, Ruiqi Song, Xianda Guo, and Long Chen. Genad: Generative end-to-end autonomous driving. arXiv preprint arXiv:2402.11502, 2024. 1", + "[39] Brady Zhou and Philipp Krahenbuhl. Cross-view transformers for real-time map-view semantic segmentation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 13760-13769, 2022. 1, 5", + "[40] Daquan Zhou, Weimin Wang, Hanshu Yan, Weiwei Lv, Yizhe Zhu, and Jiashi Feng. Magicvideo: Efficient video generation with latent diffusion models, 2023. 2" + ], + "bbox": [ + 516, + 92, + 903, + 261 + ], + "page_idx": 9 + }, + { + "type": "page_number", + "text": "17196", + "bbox": [ + 480, + 945, + 519, + 955 + ], + "page_idx": 9 + } +] \ No newline at end of file diff --git a/2025/DriveScape_ High-Resolution Driving Video Generation by Multi-View Feature Fusion/6698f995-6b34-4c34-b673-bcdca2c64dfc_model.json b/2025/DriveScape_ High-Resolution Driving Video Generation by Multi-View Feature Fusion/6698f995-6b34-4c34-b673-bcdca2c64dfc_model.json new file mode 100644 index 0000000000000000000000000000000000000000..de00996c3b0b960205bb6bef68415270fe3a1647 --- /dev/null +++ b/2025/DriveScape_ High-Resolution Driving Video Generation by Multi-View Feature Fusion/6698f995-6b34-4c34-b673-bcdca2c64dfc_model.json @@ -0,0 +1,1793 @@ +[ + [ + { + "type": "header", + "bbox": [ + 0.107, + 0.003, + 0.182, + 0.043 + ], + "angle": 0, + "content": "CVF" + }, + { + "type": "header", + "bbox": [ + 0.237, + 0.001, + 0.812, + 0.047 + ], + "angle": 0, + "content": "This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore." + }, + { + "type": "title", + "bbox": [ + 0.203, + 0.131, + 0.797, + 0.175 + ], + "angle": 0, + "content": "DriveScape: High-Resolution Driving Video Generation by Multi-View Feature Fusion" + }, + { + "type": "text", + "bbox": [ + 0.218, + 0.203, + 0.78, + 0.311 + ], + "angle": 0, + "content": "Wei Wu\\(^{1,2}\\), Xi Guo\\(^{2}\\), Weixuan Tang\\(^{2}\\), Tingxuan Huang\\(^{3}\\), Chiyu Wang\\(^{2}\\), Chenjing Ding\\(^{2*}\\) \n\\(^{1}\\)Tsinghua University \\(^{2}\\)Sensetime Research \\(^{3}\\)Northeastern University {wuwei, quoxi, dingchenjing} @ sensetime.com, {tangweixuan, wangchiyu} @ senseauto.com, {huangtingxuan} @stu.neu.edu.cn" + }, + { + "type": "title", + "bbox": [ + 0.248, + 0.345, + 0.327, + 0.361 + ], + "angle": 0, + "content": "Abstract" + }, + { + "type": "text", + "bbox": [ + 0.089, + 0.377, + 0.486, + 0.741 + ], + "angle": 0, + "content": "Recent advancements in generative models offer promising solutions for synthesizing realistic driving videos, aiding in training autonomous driving perception models. However, existing methods often struggle with high-resolution multi-view generation, mainly due to the significant memory and computational overhead caused by simultaneously inputting multi-view videos into denoising diffusion models. In this paper, we propose a driving video generation framework based on multi-view feature fusion named DriveScape for multi-view 3D condition-guided video generation. We introduce a Bi-Directional Modulated Transformer (BiMoT) module to encode, fuse and inject multiview features along with various 3D road structures and objects, which enables high-resolution multi-view generation. Consequently, our approach allows precise control over video generation, greatly enhancing realism and providing a robust solution for creating high-quality, multi-view driving videos. Our framework achieves state-of-the-art results on the nuScenes dataset, demonstrating impressive generative quality metrics with an FID score of 8.34 and an FVD score of 76.39, as well as superior performance across various perception tasks. This lays the foundation for more accurate environment simulation in autonomous driving. Our project homepage." + }, + { + "type": "title", + "bbox": [ + 0.092, + 0.772, + 0.224, + 0.789 + ], + "angle": 0, + "content": "1. Introduction" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.799, + 0.485, + 0.875 + ], + "angle": 0, + "content": "Autonomous driving has garnered significant attention, emphasizing the need for precise environmental understanding to ensure safe driving decisions [13, 38]. Bird's-eye view (BEV) maps, generated from multi-view images, serve as a crucial structured representation for various tasks, includ" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.346, + 0.908, + 0.513 + ], + "angle": 0, + "content": "ing 3D object detection, segmentation, tracking, depth estimation, and trajectory prediction [14, 15, 17, 18, 20, 34, 36, 39]. However, acquiring high-quality multi-view video data poses significant challenges due to the high labeling costs, increasing the demand for generating temporally consistent video data aligned with real-world distributions and 3D labels. Recent generative models have demonstrated the potential of synthetic data to enhance visual tasks [7, 25, 35], while diffusion models have advanced the generation of diverse and realistic driving videos, which are crucial for developing robust autonomous driving systems." + }, + { + "type": "table", + "bbox": [ + 0.537, + 0.527, + 0.885, + 0.632 + ], + "angle": 0, + "content": "
ModelSpatial Res.Temporal Res. (Hz)sparse condition
Drive-WM192 × 3842×
WoVoGen256 × 4482×
Delphi512 × 5122×
GenAD256 × 4482×
MagicDrive272 × 736-×
DriveDreamer 1&2448 × 2562×
DriveDiffusion512 × 5122×
Panacea256 × 5122×
Ours (DriveScape)576 × 10242 - 10✓
" + }, + { + "type": "table_caption", + "bbox": [ + 0.513, + 0.642, + 0.907, + 0.698 + ], + "angle": 0, + "content": "Table 1. Real-world multi-view driving controllable world models. DriveScape (ours) operates at a high spatiotemporal resolution, offering support for sparse layout controllability, as well as unified control for both mono-condition and multi-condition scenarios." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.705, + 0.909, + 0.903 + ], + "angle": 0, + "content": "The most relevant approach to our work is the generation of multi-view driving videos controlled by 3D conditions. In terms of spatial resolution, general video generation tasks, such as those addressed by SVD, have achieved high-resolution video generation at \\(1024 \\times 576\\). However, in driving scenarios, multi-camera panoramic generation is often required. Previous methods typically need to denoise multiple views simultaneously, leading to memory constraints during both training and inference. For temporal resolution, previous methods rely on frame-by-frame 3D layout control. Due to high annotation costs, public datasets like NuScenes provide annotations at a low frame rate (2Hz), restricting models from generating videos at the" + }, + { + "type": "page_footnote", + "bbox": [ + 0.11, + 0.888, + 0.236, + 0.901 + ], + "angle": 0, + "content": "*Corresponding author" + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.52, + 0.957 + ], + "angle": 0, + "content": "17187" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.09, + 0.092, + 0.482, + 0.228 + ], + "angle": 0, + "content": "rate (2Hz), restricting models from generating videos at the same frame rate. Additionally, the cost of obtaining different types of annotations varies in practice. For instance, annotating 3D lane lines is typically more expensive than annotating bounding boxes, often resulting in lane lines being labeled at a lower frame rate. This necessitates generative models that can handle sparse conditional control and remain compatible with both sparse and non-sparse conditions simultaneously." + }, + { + "type": "text", + "bbox": [ + 0.093, + 0.23, + 0.483, + 0.47 + ], + "angle": 0, + "content": "To address these challenges, we propose DriveScape, a solution for multi-view 3D condition-guided video generation with high spatio-temporal resolution. Unlike existing methods, DriveScape integrates adjacent views and keyframes through multi-view fusion, avoiding the computational burden of denoising multiple views simultaneously. We introduce a streamlined training and inference pipeline, free from post-processing, supporting high spatial and temporal resolution for sparse condition control. To align 3D road structural information, we propose a Bi-Directional Condition Alignment module, integrating multiple conditions into an aligned latent space for fine-grained layout controllability. By leveraging temporal and spatial information, DriveScape establishes a unified model for consistent multi-view video generation, significantly enhancing realism and applicability." + }, + { + "type": "text", + "bbox": [ + 0.11, + 0.472, + 0.438, + 0.486 + ], + "angle": 0, + "content": "In summary, our key contributions are as follows:" + }, + { + "type": "text", + "bbox": [ + 0.092, + 0.488, + 0.483, + 0.577 + ], + "angle": 0, + "content": "- We propose a unified multi-view feature fusion framework for driving video generation without any post-processing. To the best of our knowledge, our model is the first to achieve high-resolution, high-frame-rate, sparse condition control for multi-view driving video generation." + }, + { + "type": "text", + "bbox": [ + 0.092, + 0.578, + 0.483, + 0.668 + ], + "angle": 0, + "content": "- We introduce a Bi-Directional Condition Alignment module to facilitate diverse modalities of condition guidance and sparse condition control in video generation tasks. Experiments demonstrate that our model achieves decoupled, sparse, and precise control of dynamic foreground and static background." + }, + { + "type": "text", + "bbox": [ + 0.092, + 0.669, + 0.483, + 0.729 + ], + "angle": 0, + "content": "- We achieve state-of-the-art video synthesis performance on the nuScenes dataset, excelling not only in the generation quality of multi-view images and videos but also across various perception applications." + }, + { + "type": "list", + "bbox": [ + 0.092, + 0.488, + 0.483, + 0.729 + ], + "angle": 0, + "content": null + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.746, + 0.24, + 0.763 + ], + "angle": 0, + "content": "2. Related Works" + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.772, + 0.31, + 0.789 + ], + "angle": 0, + "content": "2.1. Controllable generation" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.795, + 0.483, + 0.901 + ], + "angle": 0, + "content": "With the advent of diffusion models, significant progress has been made in text-to-video generation [1, 2, 5, 8, 10, 12, 24, 30, 40]. Video LDM [4] employs a latent diffusion pipeline where denoising occurs in the image latent space, greatly accelerating the process. However, text alone cannot precisely control video generation. Later methods introduced image blocks alongside text as prompts for the" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.092, + 0.905, + 0.198 + ], + "angle": 0, + "content": "denoising network [37]. Our goal is to generate highly realistic street-view videos, a complex scene featuring numerous elements and interactions (e.g., intricate street layouts and moving vehicles). To achieve fine-grained control, we incorporate road maps, 3D bounding boxes, and BEV keyframes, going beyond text and images to ensure detailed and precise video generation." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.208, + 0.766, + 0.224 + ], + "angle": 0, + "content": "2.2. Multi-view video generation" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.231, + 0.907, + 0.396 + ], + "angle": 0, + "content": "Multi-view consistency and temporal consistency are two critical challenges in multi-view video generation. To address multi-view consistency, MVDiffusion [26] introduced a correspondence-aware attention module to align information across views. Tseng et al. [27] utilized epipolar geometry to regularize consistency between different views. MagicDrive [7] leveraged camera poses, bounding boxes, and road maps as priors, incorporating an additional cross-view attention block to enhance consistency. However, these methods are limited to generating multi-view images rather than videos." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.408, + 0.722, + 0.423 + ], + "angle": 0, + "content": "2.3. Street view generation" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.43, + 0.907, + 0.688 + ], + "angle": 0, + "content": "Most street-view generation models condition on 2D layouts, such as BEV maps, 2D bounding boxes, and semantic segmentation. BEVGen [25] encapsulates all semantics within BEV for street-view generation. BEVControl [35] employs a two-stage method to generate multi-view urban scene images from BEV layouts, where a controller generates foreground and background objects, and a coordinator combines them to maintain visual consistency across views. However, projecting 3D information to 2D loses geometric details, making direct extension to video generation prone to inconsistencies across frames. To address this, we incorporate 3D bounding boxes as a condition to guide generation.DrivingDiffusion [16] adopts a multi-stage pipeline with two models and two stages of post-processing to generate frames and extend them into videos. In contrast, our method avoids the complexity of multi-stage pipelines by using an efficient, end-to-end approach." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.702, + 0.605, + 0.717 + ], + "angle": 0, + "content": "3. Method" + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.727, + 0.625, + 0.742 + ], + "angle": 0, + "content": "3.1. Overview" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.75, + 0.905, + 0.853 + ], + "angle": 0, + "content": "The overview of DriveScape is shown in Fig. 1. Based on the LDM pipeline [22], DriveScape generates street-view videos conditioned on scene annotations, such as BEV maps, 3D bounding boxes, ego states, and text for each view. The unified model, described in Section 3.2, enables high-resolution multi-view video generation without post-refinement." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.856, + 0.905, + 0.901 + ], + "angle": 0, + "content": "The Bi-Directional Condition Alignment module, detailed in Section 3.5, ensures effective alignment between 3D road structures and temporal sparse conditions." + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "17188" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.134, + 0.089, + 0.868, + 0.375 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.089, + 0.386, + 0.908, + 0.459 + ], + "angle": 0, + "content": "Figure 1. End-to-End Multi-view Video Generation pipeline. We use learnable embedding vectors to represent different cameras and categorize camera views as key views and neighbour views. Our training scheme guides the generation of key view videos through their neighboring frames. Additionally, we introduce the key-frames condition along with a training and inference scheme to ensure multi-view consistency simultaneously. Furthermore, our model does not require any post-refine process as DriveDiffusion ([16]). It is capable of learning multi-view and temporal consistency simultaneously, resulting in high-fidelity street-view synthesis." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.482, + 0.238, + 0.496 + ], + "angle": 0, + "content": "3.2. Unified Model" + }, + { + "type": "text", + "bbox": [ + 0.089, + 0.505, + 0.483, + 0.701 + ], + "angle": 0, + "content": "Our model leverages SVD [3] as the backbone, with each denoising step processing a video from a single viewpoint. Control conditions are injected through a unified control module, including frame-wise images from adjacent viewpoints, road maps, 3D bounding boxes, and the first frame from each viewpoint. Starting with initial multi-view frames \\( x_{0:N-1}^{0}(N \\) is the number of views), DriveScape predicts subsequent frames \\( x_{0:N-1}^{1:T}(T \\) is the number of frames) under conditions such as BEV maps \\( m_{0:N-1}^{0:T} \\), 3D bounding boxes \\( b_{0:N-1}^{0:T} \\), BEV key frames, and neighboring camera videos. BEV maps and 3D bounding boxes are projected into the camera's First-Person View (FPV) and encoded with convolutional blocks." + }, + { + "type": "text", + "bbox": [ + 0.089, + 0.703, + 0.483, + 0.825 + ], + "angle": 0, + "content": "As shown in Fig. 1, we use learnable embedding vectors \\(\\{e_i\\}\\) to represent different cameras. Similar to the processing of canbus data (e.g., velocity and direction angle), these embeddings are passed through a multilayer perceptron (MLP) followed by a Sigmoid Linear Unit (SiLU) activation. This enables the embeddings to perform crossattention with the image latent space, allowing the unified model to predict videos from multiple perspectives." + }, + { + "type": "text", + "bbox": [ + 0.089, + 0.826, + 0.483, + 0.902 + ], + "angle": 0, + "content": "Embedding camera information alone cannot ensure spatial and temporal consistency, as it only captures global camera characteristics. Consistency is achieved through model design and training. For temporal consistency, we integrate temporal attention into the latent diffusion UNet [23]" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.482, + 0.907, + 0.528 + ], + "angle": 0, + "content": "and control alignment module (Section 3.5). For multi-view consistency, we introduce keyframe conditions and a carefully designed training and inference scheme." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.543, + 0.665, + 0.559 + ], + "angle": 0, + "content": "3.3. Training Stage" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.567, + 0.907, + 0.734 + ], + "angle": 0, + "content": "Neighbor frames feature fusion. DriveScape categorizes camera views into key and neighbor views (Fig. 2). Frames of the same type have minimal overlap, ensuring distinct regions. The training scheme guides key view generation using neighbor frames. During training, neighbor and key views are selected sequentially. For neighbor views, video conditions from adjacent key views are input into the network. Key views are generated sequentially, stored, and incorporated into the cross-attention mechanism with the image latent space, ensuring effective guidance for neighbor view generation." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.735, + 0.909, + 0.903 + ], + "angle": 0, + "content": "Multi-view key-frame feature fusion. A challenge arises from the lack of specific multi-view constraints during key-view generation, leading to degraded performance. While this is generally acceptable due to minimal overlapping regions, long-term generation introduces interdependencies among views. For example, during overtaking, a vehicle initially captured by the rear camera later appears in the front camera. To address this, we propose key-frame conditions, which are the starting frames of all views, to establish correlations within the image latent space. Combined with temporal attention, the model learns the relationships across" + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.52, + 0.957 + ], + "angle": 0, + "content": "17189" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.134, + 0.089, + 0.868, + 0.376 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.089, + 0.386, + 0.908, + 0.471 + ], + "angle": 0, + "content": "Figure 2. The DriveScape pipeline, built on the LDM framework, generates street-view videos conditioned on scene annotations, BEV maps, and 3D bounding boxes for each view. Keyframe conditions are introduced, in collaboration with temporal and spatial self-attention modules, to ensure consistency across spatial and temporal dimensions. Our unified model enables multi-view video generation without the need for complex post-processing or refinement procedures. Additionally, the Bi-Directional Modulated Transformer (BiMoT) module, consisting of two cross-attention layers with opposite directions and a temporal self-attention layer, ensures effective alignment and integration of various 3D road structural information, enabling precise control over video generation." + }, + { + "type": "text", + "bbox": [ + 0.089, + 0.496, + 0.483, + 0.527 + ], + "angle": 0, + "content": "all views at the initial frame and the preceding frames of neighboring cameras." + }, + { + "type": "text", + "bbox": [ + 0.089, + 0.529, + 0.483, + 0.725 + ], + "angle": 0, + "content": "Sparse multi-modal feature fusion. Our model supports training with videos at various frame rates, ranging from 2Hz to 10Hz, while maintaining sample conditions at 2Hz. A lightweight encoder is used to encode the 2Hz map and bounding boxes (BBoxes), with condition-missing frames padded with zeros. Additionally, an extra mask channel is introduced to indicate the presence of conditions at each frame, where mask values are 0 or 1. These inputs are processed through the BiMoT module. Embeddings for unconditioned frames are learned and integrated into the conditions to align with the video frame rate. This diverse frame-rate training scheme, combined with the BiMoT module, enhances sparse condition control capability." + }, + { + "type": "text", + "bbox": [ + 0.089, + 0.727, + 0.483, + 0.817 + ], + "angle": 0, + "content": "Conditions dropout. To enhance robustness, key views are trained in random order. Neighbor camera videos and conditions are dropped with probabilities of \\(50\\%\\) and \\(20\\%\\), respectively. Unlike [16], our model requires no postrefinement, simultaneously learning multi-view and temporal consistency to improve quality and realism." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.832, + 0.248, + 0.849 + ], + "angle": 0, + "content": "3.4. Inference Stage" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.856, + 0.483, + 0.902 + ], + "angle": 0, + "content": "During inference, our pipeline enables efficient and consistent generation of multi-view videos. Similar to the training stage, the process consists of two patterns: key-view infer" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.496, + 0.905, + 0.619 + ], + "angle": 0, + "content": "ence and neighbor-view inference, as shown in Fig. 2. Keyview camera videos are generated simultaneously without cross-view condition control, as the key-view cameras are intentionally placed non-adjacently. For example, one camera is positioned at the front, while two are at the rear—one on the left and the other on the right. This strategic arrangement ensures comprehensive coverage from multiple perspectives with minimal overlap." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.62, + 0.906, + 0.709 + ], + "angle": 0, + "content": "In this configuration, neighbor-view cameras—two at the front and one at the rear—naturally receive cross-view conditional information from their adjacent key-view cameras. This design allows neighbor-view inference to maintain consistency with conditions from neighboring cameras, ensuring cross-view coherence." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.711, + 0.906, + 0.788 + ], + "angle": 0, + "content": "Leveraging this setup, our inference pipeline excels in parallel processing, requiring only two consecutive iterations to generate multi-view videos. This approach not only accelerates video production but also ensures cross-view consistency and realism throughout." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.802, + 0.849, + 0.817 + ], + "angle": 0, + "content": "3.5. Bi-Directional Modulated Transformer" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.826, + 0.905, + 0.901 + ], + "angle": 0, + "content": "To integrate multi-conditions from different modalities into the UNet model and ensure effective articulation of each condition, we propose the Bi-Directional Modulated Transformer (BiMoT) module. This module aligns features from various modalities within the latent space, reinforces tem" + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "17190" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.09, + 0.092, + 0.482, + 0.152 + ], + "angle": 0, + "content": "poral coherence among the merged conditions, and injects the aligned conditions into the network. As illustrated in Fig. 1, the module consists of two cross-attention layers with opposite directions and a temporal self-attention layer." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.153, + 0.482, + 0.258 + ], + "angle": 0, + "content": "Latent to Condition Attention To address diverse conditions with different modalities, we employ a latent-to-condition cross-attention mechanism to standardize feature representations. Pixel-level features are used as signals to merge various condition features. Specifically, let \\( F_{cond} \\) represent condition features and \\( F_{latent} \\) represent video latent features. The output is computed as:" + }, + { + "type": "equation", + "bbox": [ + 0.17, + 0.27, + 0.483, + 0.291 + ], + "angle": 0, + "content": "\\[\nF _ {c o n d} ^ {\\prime} = \\operatorname {c r o s s a t t n} \\left(F _ {c o n d}, F _ {l a t e n t}\\right), \\tag {1}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.302, + 0.482, + 0.362 + ], + "angle": 0, + "content": "where the query is \\( F_{cond} \\), and the key and value are \\( F_{latent} \\). This cross-attention operation facilitates the flow of video latent information into the condition space, providing a unified representation for integrating all modalities." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.363, + 0.482, + 0.498 + ], + "angle": 0, + "content": "Temporal Attention. We apply temporal self-attention to ensure that the merged features propagate effectively along the temporal dimension. Sparse condition information is also incorporated into the image latent space, naturally unifying all modalities into a single representation. This enables temporal information from latent features to interact with sparse conditions over time, guiding the generation process with sparse control. This mechanism ensures consistency and precise control over the temporal dimensions." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.5, + 0.482, + 0.545 + ], + "angle": 0, + "content": "Condition to Video Attention. Similar to the latent-to-condition cross-attention, a condition-to-video cross-attention mechanism is applied:" + }, + { + "type": "equation", + "bbox": [ + 0.167, + 0.557, + 0.483, + 0.577 + ], + "angle": 0, + "content": "\\[\nF _ {l a t e n t} ^ {\\prime} = \\operatorname {c r o s s a t t n} \\left(F _ {l a t e n t}, F _ {c o n d}\\right). \\tag {2}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.588, + 0.482, + 0.633 + ], + "angle": 0, + "content": "This operation facilitates the flow of merged condition feature information into the image latents, ultimately enabling unified control over video generation." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.634, + 0.483, + 0.799 + ], + "angle": 0, + "content": "Coarse to Fine Condition Mixture. BiMoT is flexible in accommodating various types of conditions, enabling easy extension to additional inputs. We begin with maps and 3D layouts, then introduce BEV key-frames, and finally integrate neighbor camera videos. This progression allows generation control to transition from a coarse level, driven by maps and 3D layouts, to a finer level, guided by pixel-level details. This ensures both structural coherence and detailed fidelity. The flexibility of BiMoT enables fine-grained layout controllability with unified mono-condition and multi-condition control." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.815, + 0.222, + 0.832 + ], + "angle": 0, + "content": "4. Experiments" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.84, + 0.482, + 0.901 + ], + "angle": 0, + "content": "In this section, we evaluate our model through key metrics, outlining the experimental setup, dataset, and evaluation criteria. We then analyze qualitative results and conduct ablation studies to validate our approach." + }, + { + "type": "title", + "bbox": [ + 0.514, + 0.092, + 0.698, + 0.107 + ], + "angle": 0, + "content": "4.1. Experiment Details" + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.113, + 0.905, + 0.203 + ], + "angle": 0, + "content": "Dataset. The training data is sourced from the real-world driving dataset nuScenes [6], a widely used benchmark for BEV segmentation and 3D object detection in driving scenarios. Following the official configuration, we use 700 street-view scenes for training and 150 for validation, targeting 8 object classes and 8 road classes." + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.204, + 0.905, + 0.446 + ], + "angle": 0, + "content": "Evaluation Metrics. To evaluate the effectiveness of our approach in delivering realism, continuity, and precise control over multiple conditions, we employ four key metrics for comparison against existing multi-view image and video generation methods. For realism, we use the Fréchet Inception Distance (FID) [11] to measure image quality. Temporal coherence is assessed using the Fréchet Video Distance (FVD) [28], ensuring consistent and fluid motion in the generated videos. Controllability is evaluated through two perception-based methods: CVT [39] for BEV segmentation and BEVFusion [19] for 3D object detection. Both methods are trained on the nuScenes dataset, providing a reliable benchmark for performance evaluation. Additionally, since our model generates videos, we follow Panacea [33] and use the video-based perception model StreamPETR [29] to evaluate the 3D object detection task." + }, + { + "type": "title", + "bbox": [ + 0.514, + 0.456, + 0.708, + 0.471 + ], + "angle": 0, + "content": "4.2. Experiments Results" + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.478, + 0.905, + 0.614 + ], + "angle": 0, + "content": "Qualitative Results. We showcase the quality and consistency of our video generation results in Fig. 3. Our model achieves notable consistency across both frames and viewpoints. Specifically, vehicles are accurately positioned according to the provided 3D layouts, and street delineations closely match the supplied projected maps. Remarkably, even when bounding boxes extend beyond the visible area, our model adheres to the specified conditions and generates the remaining objects with high fidelity." + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.614, + 0.905, + 0.841 + ], + "angle": 0, + "content": "Sparse Condition Control As shown in Fig. 5, our model generates consistent videos in both time and space under sparse conditions, achieving a resolution of \\(576 \\times 1024\\) and a frame rate of 10 fps. To compare with related works on sparse control [9], we implemented a sparse control model with multi-condition fusion, as existing methods lack support for multi-condition control. A visualization comparing the reproduced SparseCtrl [9] and our method is presented in Fig. 4. Our method clearly demonstrates superior spatial consistency. Moreover, the FID and FVD scores for the reproduced SparseCtrl are 11.15 and 116.00, respectively, significantly worse than those achieved by our model (8.34 and 76.39). For additional visual comparisons and details on the multi-condition fusion SparseCtrl, please refer to the supplementary materials and the accompanying web page." + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.841, + 0.905, + 0.901 + ], + "angle": 0, + "content": "Quantitative Results. We report quantitative metrics on the nuScenes validation set in Table 2. Using FID and FVD, we evaluate image and video quality, respectively. Following MagicDrive [7] and DrivingDiffusion [16], we" + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.518, + 0.957 + ], + "angle": 0, + "content": "17191" + } + ], + [ + { + "type": "table", + "bbox": [ + 0.18, + 0.088, + 0.825, + 0.261 + ], + "angle": 0, + "content": "
MethodFID↓FVD↓Road mIoU ↑Vehicle mIoU ↑Object NDS ↑
Oracle--73.6834.8441.21
BEVGen[25]25.54-50.20 (-23.48)5.89 (-28.95)-
BEVControl [35]24.85-60.80 (-12.88)26.80 (-8.04)-
MagicDrive [7]16.20-61.05 (-12.63)27.01 (-7.83)23.32 (-17.89)
DriveDreamer [31]52.60452.0---
DrivingDiffusion [16]15.83332.063.20 (-10.48)31.60 (-3.24)33.1 (-8.11)
Drive-WM [32]15.80122.7065.07 (-8.61)27.19 (-7.65)-
Delphi [21]15.08113.5--36.58 (-4.63)
SparseCtrl[9]11.15116.00---
Ours8.3476.3968.21 (-5.47)30.27 (-4.57)36.5 (-4.71)
" + }, + { + "type": "table_caption", + "bbox": [ + 0.09, + 0.271, + 0.907, + 0.3 + ], + "angle": 0, + "content": "Table 2. The comparative analysis of our method against others on the nuScenes validation dataset is detailed for each task. Our end-to-end approach outperformed most baseline models without requiring any post-refinement process." + }, + { + "type": "image", + "bbox": [ + 0.118, + 0.313, + 0.882, + 0.606 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.09, + 0.621, + 0.908, + 0.665 + ], + "angle": 0, + "content": "Figure 3. Showcases for multi-view video generation by DriveScape demonstrate frame-to-frame and viewpoint consistency. Vehicles generated by our model are accurately positioned according to the provided 3D layouts, while the street depictions closely align with the projected maps." + }, + { + "type": "image", + "bbox": [ + 0.093, + 0.687, + 0.487, + 0.804 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.09, + 0.812, + 0.484, + 0.856 + ], + "angle": 0, + "content": "Figure 4. Results in each row are Ours, SparseCtrl, input condition and GT respectively, which illustrate that our method has better spatial consistency than SparseCtrl." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.871, + 0.484, + 0.903 + ], + "angle": 0, + "content": "generate data from validation set labels and assess it with a pre-trained perception model. The decline in Road mIoU" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.69, + 0.907, + 0.734 + ], + "angle": 0, + "content": "and Vehicle mIoU, compared to original data, measures the effectiveness of our framework in producing perception-friendly synthetic data." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.735, + 0.909, + 0.87 + ], + "angle": 0, + "content": "To align with previous studies, we resized generated images to \\(224 \\times 400\\). We compared our FID results with three image generation models: BEVGen [25], BEVControl [35], and MagicDrive [7]. Our method significantly outperformed these models. Notably, video generation models typically yield higher FID scores than image models. For instance, Drive-WM, using only an image model, achieved better FID than the full model, highlighting the challenges of video generation." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.871, + 0.908, + 0.903 + ], + "angle": 0, + "content": "For video quality, assessed via the FVD metric, our approach showed substantial improvements over recent works" + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.52, + 0.958 + ], + "angle": 0, + "content": "17192" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.099, + 0.09, + 0.48, + 0.317 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.139, + 0.327, + 0.434, + 0.341 + ], + "angle": 0, + "content": "Figure 5. Showcase for sparse Condition Control." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.358, + 0.484, + 0.541 + ], + "angle": 0, + "content": "such as DriveDreamer [31], DrivingDiffusion [16], and Drive-WM [32]. In BEV segmentation metrics, our method achieved superior perceptual performance, except for a slightly lower Road mIoU compared to DrivingDiffusion, which utilized a post-processing network for refinement. To maintain simplicity, our end-to-end system applies no post-processing beyond image resizing for perceptual evaluation. The perception models we used operate on single-image inference, not leveraging the temporal consistency of our generated videos. Given our significant lead in FVD, our method holds strong potential for training video BEV perception models." + }, + { + "type": "table", + "bbox": [ + 0.14, + 0.552, + 0.436, + 0.634 + ], + "angle": 0, + "content": "
dataNDS ↑FID ↓FVD ↓
Oracle46.9--
Panacea32.1 (-14.8)16.96139
Oracle53.7--
Ours44.5 (-9.2)8.3476.39
" + }, + { + "type": "table_caption", + "bbox": [ + 0.09, + 0.644, + 0.483, + 0.672 + ], + "angle": 0, + "content": "Table 3. Compared with Panacea, we evaluate results on the nuScenes validation set using the pre-trained StreamPETR model." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.682, + 0.255, + 0.697 + ], + "angle": 0, + "content": "4.3. Ablation Studies" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.705, + 0.484, + 0.901 + ], + "angle": 0, + "content": "To refine and evaluate the efficacy of our proposed framework, we conducted ablation studies, as detailed in Table 4. Various strategies for implementing conditional guidance were tested. Removing the entire BiMoT module resulted in the FID increasing from 8.34 to 20.98 and the FVD rising from 76.39 to 142.81, underscoring the module's importance for multi-condition encoding. Similarly, removing the temporal attention layer within BiMoT caused a significant performance drop, with FID and FVD increasing to 20.67 and 123.03, respectively, highlighting the role of temporal modelling in BiMoT. Additionally, during testing, we separately removed the key-frame condition and the neighbor condition. Both led to a performance decline, indicating the" + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.092, + 0.907, + 0.122 + ], + "angle": 0, + "content": "importance of these conditions for facilitating information exchange between views." + }, + { + "type": "table", + "bbox": [ + 0.541, + 0.134, + 0.882, + 0.23 + ], + "angle": 0, + "content": "
TeAttnBiMOTKFNeFIDFVD
✓✓✓20.98142.81
✓✓20.67123.03
✓✓✓11.60108.80
✓✓✓12.18111.36
✓✓✓✓8.3476.39
" + }, + { + "type": "table_caption", + "bbox": [ + 0.513, + 0.241, + 0.907, + 0.269 + ], + "angle": 0, + "content": "Table 4. Ablation Study. TeAttn is temporal attention module in BiMOT, KF is key-frame condition and Ne is neighbor condition" + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.295, + 0.81, + 0.312 + ], + "angle": 0, + "content": "4.4. Visualization of Different Settings" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.318, + 0.905, + 0.407 + ], + "angle": 0, + "content": "In our visual showcases, we demonstrate the model's proficiency in manipulating video predictions for a single scene under various control conditions. The control conditions, sourced from frames of identical perspectives but randomly selected from other scenes, test the decoupling of conditions and the robustness of our method across diverse scenarios." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.408, + 0.906, + 0.544 + ], + "angle": 0, + "content": "As shown in Fig. 7, the experiments explore the influence of different 3D layout controls. Using a single reference frame from a highway segment, we first depict vehicles driving normally, then transition to vehicles traversing grassy terrain. Remarkably, our model maintains temporal consistency even in this unconventional context. Additional experiments involve removing or adding large vehicles, with results aligning closely with the specified conditions, showcasing precise control." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.545, + 0.907, + 0.665 + ], + "angle": 0, + "content": "Furthermore, as illustrated in Fig. 8, we investigate the impact of different maps on video generation. The generated scenes align impressively with the shapes of the provided map regions, maintaining layout feature consistency despite map changes. These experiments provide strong evidence of our model's capability to render specified control conditions into precise and coherent visual narratives within the generated video content." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.68, + 0.64, + 0.695 + ], + "angle": 0, + "content": "5. Conclusions" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.705, + 0.907, + 0.901 + ], + "angle": 0, + "content": "The paper introduces DriveScape, the first end-to-end multi-view driving video generative model. It enables consistent multi-view video generation without complex postprocessing or refinement. This unified framework streamlines multi-view condition-guided video generation. We also propose a Bi-Directional Modulated Transformer module for diverse condition guidance. Our approach achieves decoupled and precise control of dynamic foreground and static background, delivering state-of-the-art video synthesis on the nuScenes dataset. It excels in multi-view image and video generation quality and perception applications. In summary, our method outperforms others in generation quality and ease of application." + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "17193" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.118, + 0.104, + 0.885, + 0.289 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.09, + 0.303, + 0.908, + 0.331 + ], + "angle": 0, + "content": "Figure 6. Showcase for maps and 3D layouts control under different conditions. (a): control by the reference maps and 3D Layouts; (b) change 3D Layouts with more vehicles; (c) change the maps with curved pedestrian crosswalk; (d) remove all the 3D Layouts." + }, + { + "type": "image", + "bbox": [ + 0.115, + 0.361, + 0.882, + 0.56 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.09, + 0.571, + 0.908, + 0.598 + ], + "angle": 0, + "content": "Figure 7. Showcases for 3D layouts control under different conditions. DriveScape can precisely control dynamic objects. (a) reference layouts; (b) move car to the grass; (c) remove all cars; (d) add trucks to the scene." + }, + { + "type": "image", + "bbox": [ + 0.115, + 0.642, + 0.885, + 0.841 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.09, + 0.852, + 0.908, + 0.88 + ], + "angle": 0, + "content": "Figure 8. Showcases for maps control under. DriveScape can precisely control static background. (a) replace the friable area with green belts; (b) replace the friable area with green belts; (c) replace the friable area with sidewalks; (d) add island on the road." + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.52, + 0.957 + ], + "angle": 0, + "content": "17194" + } + ], + [ + { + "type": "title", + "bbox": [ + 0.093, + 0.091, + 0.188, + 0.106 + ], + "angle": 0, + "content": "References" + }, + { + "type": "ref_text", + "bbox": [ + 0.101, + 0.115, + 0.484, + 0.17 + ], + "angle": 0, + "content": "[1] Jie An, Songyang Zhang, Harry Yang, Sonal Gupta, Jia-Bin Huang, Jiebo Luo, and Xi Yin. Latent-shift: Latent diffusion with temporal shift for efficient text-to-video generation. arXiv preprint arXiv:2304.08477, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.101, + 0.171, + 0.484, + 0.241 + ], + "angle": 0, + "content": "[2] Andreas Blattmann, Tim Dockhorn, Sumith Kulal, Daniel Mendelevitch, Maciej Kilian, Dominik Lorenz, Yam Levi, Zion English, Vikram Voleti, Adam Letts, Varun Jampani, and Robin Rombach. Stable video diffusion: Scaling latent video diffusion models to large datasets, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.242, + 0.483, + 0.311 + ], + "angle": 0, + "content": "[3] Andreas Blattmann, Tim Dockhorn, Sumith Kulal, Daniel Mendelevitch, Maciej Kilian, Dominik Lorenz, Yam Levi, Zion English, Vikram Voleti, Adam Letts, et al. Stable video diffusion: Scaling latent video diffusion models to large datasets. arXiv preprint arXiv:2311.15127, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.312, + 0.483, + 0.394 + ], + "angle": 0, + "content": "[4] Andreas Blattmann, Robin Rombach, Huan Ling, Tim Dockhorn, Seung Wook Kim, Sanja Fidler, and Karsten Kreis. Align your latents: High-resolution video synthesis with latent diffusion models. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 22563-22575, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.396, + 0.484, + 0.478 + ], + "angle": 0, + "content": "[5] Andreas Blattmann, Robin Rombach, Huan Ling, Tim Dockhorn, Seung Wook Kim, Sanja Fidler, and Karsten Kreis. Align your latents: High-resolution video synthesis with latent diffusion models. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 22563-22575, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.48, + 0.485, + 0.563 + ], + "angle": 0, + "content": "[6] Holger Caesar, Varun Bankiti, Alex H Lang, Sourabh Vora, Venice Erin Liong, Qiang Xu, Anush Krishnan, Yu Pan, Giancarlo Baldan, and Oscar Beijbom. nuscenes: A multimodal dataset for autonomous driving. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 11621-11631, 2020. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.565, + 0.483, + 0.619 + ], + "angle": 0, + "content": "[7] Ruiyuan Gao, Kai Chen, Enze Xie, Lanqing Hong, Zhenguo Li, Dit-Yan Yeung, and Qiang Xu. Magicdrive: Street view generation with diverse 3d geometry control. arXiv preprint arXiv:2310.02601, 2023. 1, 2, 5, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.621, + 0.483, + 0.675 + ], + "angle": 0, + "content": "[8] Yuwei Guo, Ceyuan Yang, Anyi Rao, Yaohui Wang, Yu Qiao, Dahua Lin, and Bo Dai. Animatediff:Animate your personalized text-to-image diffusion models without specific tuning. arXiv preprint arXiv:2307.04725, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.677, + 0.483, + 0.732 + ], + "angle": 0, + "content": "[9] Yuwei Guo, Ceyuan Yang, Anyi Rao, Maneesh Agrawala, Dahua Lin, and Bo Dai. Sparsectrl: Adding sparse controls to text-to-video diffusion models. In European Conference on Computer Vision, pages 330-348. Springer, 2025. 5, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.734, + 0.483, + 0.787 + ], + "angle": 0, + "content": "[10] Yingqing He, Tianyu Yang, Yong Zhang, Ying Shan, and Qifeng Chen. Latent video diffusion models for high-fidelity video generation with arbitrary lengths. arXiv preprint arXiv:2211.13221, 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.789, + 0.483, + 0.857 + ], + "angle": 0, + "content": "[11] Martin Heusel, Hubert Ramsauer, Thomas Unterthiner, Bernhard Nessler, and Sepp Hochreiter. Gans trained by a two time-scale update rule converge to a local nash equilibrium. Advances in neural information processing systems, 30, 2017. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.859, + 0.483, + 0.901 + ], + "angle": 0, + "content": "[12] Jonathan Ho, William Chan, Chitwan Sahara, Jay Whang, Ruiqi Gao, Alexey Gritsenko, Diederik P Kingma, Ben Poole, Mohammad Norouzi, David J Fleet, et al. Imagen" + }, + { + "type": "list", + "bbox": [ + 0.094, + 0.115, + 0.485, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.548, + 0.093, + 0.905, + 0.12 + ], + "angle": 0, + "content": "video: High definition video generation with diffusion models. arXiv preprint arXiv:2210.02303, 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.122, + 0.905, + 0.176 + ], + "angle": 0, + "content": "[13] Yihan Hu, Jiazhi Yang, Li Chen, Keyu Li, Chonghao Sima, Xizhou Zhu, Siqi Chai, Senyao Du, Tianwei Lin, Wenhai Wang, et al. Goal-oriented autonomous driving. arXiv preprint arXiv:2212.10156, 2022. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.178, + 0.905, + 0.217 + ], + "angle": 0, + "content": "[14] Junjie Huang and Guan Huang. Bevdet4d: Exploit temporal cues in multi-camera 3d object detection. arXiv preprint arXiv:2203.17054, 2022. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.219, + 0.905, + 0.272 + ], + "angle": 0, + "content": "[15] Junjie Huang, Guan Huang, Zheng Zhu, Yun Ye, and Dalong Du. Bevdet: High-performance multi-camera 3d object detection in bird-eye-view. arXiv preprint arXiv:2112.11790, 2021. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.274, + 0.905, + 0.328 + ], + "angle": 0, + "content": "[16] Xiaofan Li, Yifu Zhang, and Xiaoqing Ye. Drivingdiffusion: Layout-guided multi-view driving scene video generation with latent diffusion model. arXiv preprint arXiv:2310.07771, 2023. 2, 3, 4, 5, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.33, + 0.905, + 0.399 + ], + "angle": 0, + "content": "[17] Yinhao Li, Zheng Ge, Guanyi Yu, Jinrong Yang, Zengran Wang, Yukang Shi, Jianjian Sun, and Zeming Li. Bevdepth: Acquisition of reliable depth for multi-view 3d object detection. In Proceedings of the AAAI Conference on Artificial Intelligence, pages 1477-1485, 2023. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.401, + 0.905, + 0.469 + ], + "angle": 0, + "content": "[18] Zhiqi Li, Wenhai Wang, Hongyang Li, Enze Xie, Chonghao Sima, Tong Lu, Yu Qiao, and Jifeng Dai. Bevformer: Learning bird's-eye-view representation from multi-camera images via spatiotemporal transformers. In European conference on computer vision, pages 1-18. Springer, 2022. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.47, + 0.905, + 0.537 + ], + "angle": 0, + "content": "[19] Tingting Liang, Hongwei Xie, Kaicheng Yu, Zhongyu Xia, Zhiwei Lin, Yongtao Wang, Tao Tang, Bing Wang, and Zhi Tang. Bevfusion: A simple and robust lidar-camera fusion framework. Advances in Neural Information Processing Systems, 35:10421-10434, 2022. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.54, + 0.905, + 0.607 + ], + "angle": 0, + "content": "[20] Haicheng Liao, Yongkang Li, Zhenning Li, Chengyue Wang, Zhiyong Cui, Shengbo Eben Li, and Chengzhong Xu. A cognitive-based trajectory prediction approach for autonomous driving. arXiv preprint arXiv:2402.19251, 2024. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.609, + 0.905, + 0.678 + ], + "angle": 0, + "content": "[21] Enhui Ma, Lijun Zhou, Tao Tang, Zhan Zhang, Dong Han, Junpeng Jiang, Kun Zhan, Peng Jia, Xianpeng Lang, Haiyang Sun, et al. Unleashing generalization of end-to-end autonomous driving with controllable long video generation. arXiv preprint arXiv:2406.01349, 2024. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.679, + 0.907, + 0.748 + ], + "angle": 0, + "content": "[22] Robin Rombach, Andreas Blattmann, Dominik Lorenz, Patrick Esser, and Björn Ommer. High-resolution image synthesis with latent diffusion models. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 10684-10695, 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.749, + 0.907, + 0.831 + ], + "angle": 0, + "content": "[23] Olaf Ronneberger, Philipp Fischer, and Thomas Brox. U-net: Convolutional networks for biomedical image segmentation. In Medical Image Computing and Computer-Assisted Intervention-MICCAI 2015: 18th International Conference, Munich, Germany, October 5-9, 2015, Proceedings, Part III 18, pages 234-241. Springer, 2015. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.832, + 0.905, + 0.9 + ], + "angle": 0, + "content": "[24] Uriel Singer, Adam Polyak, Thomas Hayes, Xi Yin, Jie An, Songyang Zhang, Qiyuan Hu, Harry Yang, Oron Ashual, Oran Gafni, Devi Parikh, Sonal Gupta, and Yaniv Taigman. Make-a-video: Text-to-video generation without text-video data, 2022. 2" + }, + { + "type": "list", + "bbox": [ + 0.517, + 0.093, + 0.907, + 0.9 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "17195" + } + ], + [ + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.093, + 0.482, + 0.134 + ], + "angle": 0, + "content": "[25] Alexander Swerdlow, Runsheng Xu, and Bolei Zhou. Street-view image generation from a bird's-eye view layout. IEEE Robotics and Automation Letters, 2024. 1, 2, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.136, + 0.482, + 0.19 + ], + "angle": 0, + "content": "[26] Shitao Tang, Fuyang Zhang, Jiacheng Chen, Peng Wang, and Yasutaka Furukawa. Mvdiffusion: Enabling holistic multiview image generation with correspondence-aware diffusion, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.193, + 0.482, + 0.262 + ], + "angle": 0, + "content": "[27] Hung-Yu Tseng, Qinbo Li, Changil Kim, Suhib Alsisan, Jia-Bin Huang, and Johannes Kopf. Consistent view synthesis with pose-guided diffusion models. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 16773-16783, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.264, + 0.482, + 0.319 + ], + "angle": 0, + "content": "[28] Thomas Unterthiner, Sjoerd Van Steenkiste, Karol Kurach, Raphael Marinier, Marcin Michalski, and Sylvain Gelly. Towards accurate generative models of video: A new metric & challenges. arXiv preprint arXiv:1812.01717, 2018. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.321, + 0.482, + 0.375 + ], + "angle": 0, + "content": "[29] Shihao Wang, Yingfei Liu, Tiancai Wang, Ying Li, and Xi-angyu Zhang. Exploring object-centric temporal modeling for efficient multi-view 3d object detection. In CVPR, pages 3621–3631, 2023. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.378, + 0.482, + 0.433 + ], + "angle": 0, + "content": "[30] Wenjing Wang, Huan Yang, Zixi Tuo, Huiguo He, Junchen Zhu, Jianlong Fu, and Jiaying Liu. Videofactory: Swap attention in spatiotemporal diffusions for text-to-video generation, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.435, + 0.482, + 0.489 + ], + "angle": 0, + "content": "[31] Xiaofeng Wang, Zheng Zhu, Guan Huang, Xinze Chen, and Jiwen Lu. Drivedreamer: Towards real-world-driven world models for autonomous driving. arXiv preprint arXiv:2309.09777, 2023. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.491, + 0.482, + 0.559 + ], + "angle": 0, + "content": "[32] Yuqi Wang, Jiawei He, Lue Fan, Hongxin Li, Yuntao Chen, and Zhaoxiang Zhang. Driving into the future: Multiview visual forecasting and planning with world model for autonomous driving. arXiv preprint arXiv:2311.17918, 2023. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.563, + 0.482, + 0.645 + ], + "angle": 0, + "content": "[33] Yuqing Wen, Yucheng Zhao, Yingfei Liu, Fan Jia, Yanhui Wang, Chong Luo, Chi Zhang, Tiancai Wang, Xiaoyan Sun, and Xiangyu Zhang. Panacea: Panoramic and controllable video generation for autonomous driving. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 6902-6912, 2024. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.647, + 0.482, + 0.743 + ], + "angle": 0, + "content": "[34] Chenyu Yang, Yuntao Chen, Hao Tian, Chenxin Tao, Xizhou Zhu, Zhaoxiang Zhang, Gao Huang, Hongyang Li, Yu Qiao, Lewei Lu, et al. Bevformer v2: Adapting modern image backbones to bird's-eye-view recognition via perspective supervision. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 17830-17839, 2023. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.746, + 0.482, + 0.814 + ], + "angle": 0, + "content": "[35] Kairui Yang, Enhui Ma, Jibin Peng, Qing Guo, Di Lin, and Kaicheng Yu. Bevcontrol: Accurately controlling street-view elements with multi-perspective consistency via bev sketch layout. arXiv preprint arXiv:2308.01661, 2023. 1, 2, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.817, + 0.482, + 0.871 + ], + "angle": 0, + "content": "[36] Yuxiang Yang, Yingqi Deng, Jing Zhang, Jiahao Nie, and Zheng-Jun Zha. Bevtrack: A simple and strong baseline for 3d single object tracking in bird's-eye view. arXiv e-prints, pages arXiv-2309, 2023. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.873, + 0.482, + 0.901 + ], + "angle": 0, + "content": "[37] David Junhao Zhang, Dongxu Li, Hung Le, Mike Zheng Shou, Caiming Xiong, and Doyen Sahoo. Moonshot: To" + }, + { + "type": "list", + "bbox": [ + 0.093, + 0.093, + 0.482, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.549, + 0.093, + 0.905, + 0.12 + ], + "angle": 0, + "content": "wards controllable video generation and editing with multimodal conditions, 2024. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.122, + 0.905, + 0.163 + ], + "angle": 0, + "content": "[38] Wenzhao Zheng, Ruiqi Song, Xianda Guo, and Long Chen. Genad: Generative end-to-end autonomous driving. arXiv preprint arXiv:2402.11502, 2024. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.165, + 0.905, + 0.22 + ], + "angle": 0, + "content": "[39] Brady Zhou and Philipp Krahenbuhl. Cross-view transformers for real-time map-view semantic segmentation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 13760-13769, 2022. 1, 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.222, + 0.905, + 0.262 + ], + "angle": 0, + "content": "[40] Daquan Zhou, Weimin Wang, Hanshu Yan, Weiwei Lv, Yizhe Zhu, and Jiashi Feng. Magicvideo: Efficient video generation with latent diffusion models, 2023. 2" + }, + { + "type": "list", + "bbox": [ + 0.517, + 0.093, + 0.905, + 0.262 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.946, + 0.52, + 0.957 + ], + "angle": 0, + "content": "17196" + } + ] +] \ No newline at end of file diff --git a/2025/DriveScape_ High-Resolution Driving Video Generation by Multi-View Feature Fusion/6698f995-6b34-4c34-b673-bcdca2c64dfc_origin.pdf b/2025/DriveScape_ High-Resolution Driving Video Generation by Multi-View Feature Fusion/6698f995-6b34-4c34-b673-bcdca2c64dfc_origin.pdf new file mode 100644 index 0000000000000000000000000000000000000000..bed9deffb618a57174047d98c9020df171a1f411 --- /dev/null +++ b/2025/DriveScape_ High-Resolution Driving Video Generation by Multi-View Feature Fusion/6698f995-6b34-4c34-b673-bcdca2c64dfc_origin.pdf @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:afbb26e4115b71f9e868eaf19ba7ca5d5be7544a05e524d605207600d1288c28 +size 9685655 diff --git a/2025/DriveScape_ High-Resolution Driving Video Generation by Multi-View Feature Fusion/full.md b/2025/DriveScape_ High-Resolution Driving Video Generation by Multi-View Feature Fusion/full.md new file mode 100644 index 0000000000000000000000000000000000000000..654848ac55db4be660801e66d4995a8bf062e32e --- /dev/null +++ b/2025/DriveScape_ High-Resolution Driving Video Generation by Multi-View Feature Fusion/full.md @@ -0,0 +1,241 @@ +# DriveScape: High-Resolution Driving Video Generation by Multi-View Feature Fusion + +Wei Wu $^{1,2}$ , Xi Guo $^{2}$ , Weixuan Tang $^{2}$ , Tingxuan Huang $^{3}$ , Chiyu Wang $^{2}$ , Chenjing Ding $^{2*}$ $^{1}$ Tsinghua University $^{2}$ Sensetime Research $^{3}$ Northeastern University {wuwei, quoxi, dingchenjing} @ sensetime.com, {tangweixuan, wangchiyu} @ senseauto.com, {huangtingxuan} @stu.neu.edu.cn + +# Abstract + +Recent advancements in generative models offer promising solutions for synthesizing realistic driving videos, aiding in training autonomous driving perception models. However, existing methods often struggle with high-resolution multi-view generation, mainly due to the significant memory and computational overhead caused by simultaneously inputting multi-view videos into denoising diffusion models. In this paper, we propose a driving video generation framework based on multi-view feature fusion named DriveScape for multi-view 3D condition-guided video generation. We introduce a Bi-Directional Modulated Transformer (BiMoT) module to encode, fuse and inject multiview features along with various 3D road structures and objects, which enables high-resolution multi-view generation. Consequently, our approach allows precise control over video generation, greatly enhancing realism and providing a robust solution for creating high-quality, multi-view driving videos. Our framework achieves state-of-the-art results on the nuScenes dataset, demonstrating impressive generative quality metrics with an FID score of 8.34 and an FVD score of 76.39, as well as superior performance across various perception tasks. This lays the foundation for more accurate environment simulation in autonomous driving. Our project homepage. + +# 1. Introduction + +Autonomous driving has garnered significant attention, emphasizing the need for precise environmental understanding to ensure safe driving decisions [13, 38]. Bird's-eye view (BEV) maps, generated from multi-view images, serve as a crucial structured representation for various tasks, includ + +ing 3D object detection, segmentation, tracking, depth estimation, and trajectory prediction [14, 15, 17, 18, 20, 34, 36, 39]. However, acquiring high-quality multi-view video data poses significant challenges due to the high labeling costs, increasing the demand for generating temporally consistent video data aligned with real-world distributions and 3D labels. Recent generative models have demonstrated the potential of synthetic data to enhance visual tasks [7, 25, 35], while diffusion models have advanced the generation of diverse and realistic driving videos, which are crucial for developing robust autonomous driving systems. + +
ModelSpatial Res.Temporal Res. (Hz)sparse condition
Drive-WM192 × 3842×
WoVoGen256 × 4482×
Delphi512 × 5122×
GenAD256 × 4482×
MagicDrive272 × 736-×
DriveDreamer 1&2448 × 2562×
DriveDiffusion512 × 5122×
Panacea256 × 5122×
Ours (DriveScape)576 × 10242 - 10✓
+ +Table 1. Real-world multi-view driving controllable world models. DriveScape (ours) operates at a high spatiotemporal resolution, offering support for sparse layout controllability, as well as unified control for both mono-condition and multi-condition scenarios. + +The most relevant approach to our work is the generation of multi-view driving videos controlled by 3D conditions. In terms of spatial resolution, general video generation tasks, such as those addressed by SVD, have achieved high-resolution video generation at $1024 \times 576$ . However, in driving scenarios, multi-camera panoramic generation is often required. Previous methods typically need to denoise multiple views simultaneously, leading to memory constraints during both training and inference. For temporal resolution, previous methods rely on frame-by-frame 3D layout control. Due to high annotation costs, public datasets like NuScenes provide annotations at a low frame rate (2Hz), restricting models from generating videos at the + +rate (2Hz), restricting models from generating videos at the same frame rate. Additionally, the cost of obtaining different types of annotations varies in practice. For instance, annotating 3D lane lines is typically more expensive than annotating bounding boxes, often resulting in lane lines being labeled at a lower frame rate. This necessitates generative models that can handle sparse conditional control and remain compatible with both sparse and non-sparse conditions simultaneously. + +To address these challenges, we propose DriveScape, a solution for multi-view 3D condition-guided video generation with high spatio-temporal resolution. Unlike existing methods, DriveScape integrates adjacent views and keyframes through multi-view fusion, avoiding the computational burden of denoising multiple views simultaneously. We introduce a streamlined training and inference pipeline, free from post-processing, supporting high spatial and temporal resolution for sparse condition control. To align 3D road structural information, we propose a Bi-Directional Condition Alignment module, integrating multiple conditions into an aligned latent space for fine-grained layout controllability. By leveraging temporal and spatial information, DriveScape establishes a unified model for consistent multi-view video generation, significantly enhancing realism and applicability. + +In summary, our key contributions are as follows: + +- We propose a unified multi-view feature fusion framework for driving video generation without any post-processing. To the best of our knowledge, our model is the first to achieve high-resolution, high-frame-rate, sparse condition control for multi-view driving video generation. +- We introduce a Bi-Directional Condition Alignment module to facilitate diverse modalities of condition guidance and sparse condition control in video generation tasks. Experiments demonstrate that our model achieves decoupled, sparse, and precise control of dynamic foreground and static background. +- We achieve state-of-the-art video synthesis performance on the nuScenes dataset, excelling not only in the generation quality of multi-view images and videos but also across various perception applications. + +# 2. Related Works + +# 2.1. Controllable generation + +With the advent of diffusion models, significant progress has been made in text-to-video generation [1, 2, 5, 8, 10, 12, 24, 30, 40]. Video LDM [4] employs a latent diffusion pipeline where denoising occurs in the image latent space, greatly accelerating the process. However, text alone cannot precisely control video generation. Later methods introduced image blocks alongside text as prompts for the + +denoising network [37]. Our goal is to generate highly realistic street-view videos, a complex scene featuring numerous elements and interactions (e.g., intricate street layouts and moving vehicles). To achieve fine-grained control, we incorporate road maps, 3D bounding boxes, and BEV keyframes, going beyond text and images to ensure detailed and precise video generation. + +# 2.2. Multi-view video generation + +Multi-view consistency and temporal consistency are two critical challenges in multi-view video generation. To address multi-view consistency, MVDiffusion [26] introduced a correspondence-aware attention module to align information across views. Tseng et al. [27] utilized epipolar geometry to regularize consistency between different views. MagicDrive [7] leveraged camera poses, bounding boxes, and road maps as priors, incorporating an additional cross-view attention block to enhance consistency. However, these methods are limited to generating multi-view images rather than videos. + +# 2.3. Street view generation + +Most street-view generation models condition on 2D layouts, such as BEV maps, 2D bounding boxes, and semantic segmentation. BEVGen [25] encapsulates all semantics within BEV for street-view generation. BEVControl [35] employs a two-stage method to generate multi-view urban scene images from BEV layouts, where a controller generates foreground and background objects, and a coordinator combines them to maintain visual consistency across views. However, projecting 3D information to 2D loses geometric details, making direct extension to video generation prone to inconsistencies across frames. To address this, we incorporate 3D bounding boxes as a condition to guide generation.DrivingDiffusion [16] adopts a multi-stage pipeline with two models and two stages of post-processing to generate frames and extend them into videos. In contrast, our method avoids the complexity of multi-stage pipelines by using an efficient, end-to-end approach. + +# 3. Method + +# 3.1. Overview + +The overview of DriveScape is shown in Fig. 1. Based on the LDM pipeline [22], DriveScape generates street-view videos conditioned on scene annotations, such as BEV maps, 3D bounding boxes, ego states, and text for each view. The unified model, described in Section 3.2, enables high-resolution multi-view video generation without post-refinement. + +The Bi-Directional Condition Alignment module, detailed in Section 3.5, ensures effective alignment between 3D road structures and temporal sparse conditions. + +![](images/49b1dfe8d80e4b6efba66d7778586e6a150f77efc9405d2da381e86e00bfe69f.jpg) +Figure 1. End-to-End Multi-view Video Generation pipeline. We use learnable embedding vectors to represent different cameras and categorize camera views as key views and neighbour views. Our training scheme guides the generation of key view videos through their neighboring frames. Additionally, we introduce the key-frames condition along with a training and inference scheme to ensure multi-view consistency simultaneously. Furthermore, our model does not require any post-refine process as DriveDiffusion ([16]). It is capable of learning multi-view and temporal consistency simultaneously, resulting in high-fidelity street-view synthesis. + +# 3.2. Unified Model + +Our model leverages SVD [3] as the backbone, with each denoising step processing a video from a single viewpoint. Control conditions are injected through a unified control module, including frame-wise images from adjacent viewpoints, road maps, 3D bounding boxes, and the first frame from each viewpoint. Starting with initial multi-view frames $x_{0:N-1}^{0}(N$ is the number of views), DriveScape predicts subsequent frames $x_{0:N-1}^{1:T}(T$ is the number of frames) under conditions such as BEV maps $m_{0:N-1}^{0:T}$ , 3D bounding boxes $b_{0:N-1}^{0:T}$ , BEV key frames, and neighboring camera videos. BEV maps and 3D bounding boxes are projected into the camera's First-Person View (FPV) and encoded with convolutional blocks. + +As shown in Fig. 1, we use learnable embedding vectors $\{e_i\}$ to represent different cameras. Similar to the processing of canbus data (e.g., velocity and direction angle), these embeddings are passed through a multilayer perceptron (MLP) followed by a Sigmoid Linear Unit (SiLU) activation. This enables the embeddings to perform crossattention with the image latent space, allowing the unified model to predict videos from multiple perspectives. + +Embedding camera information alone cannot ensure spatial and temporal consistency, as it only captures global camera characteristics. Consistency is achieved through model design and training. For temporal consistency, we integrate temporal attention into the latent diffusion UNet [23] + +and control alignment module (Section 3.5). For multi-view consistency, we introduce keyframe conditions and a carefully designed training and inference scheme. + +# 3.3. Training Stage + +Neighbor frames feature fusion. DriveScape categorizes camera views into key and neighbor views (Fig. 2). Frames of the same type have minimal overlap, ensuring distinct regions. The training scheme guides key view generation using neighbor frames. During training, neighbor and key views are selected sequentially. For neighbor views, video conditions from adjacent key views are input into the network. Key views are generated sequentially, stored, and incorporated into the cross-attention mechanism with the image latent space, ensuring effective guidance for neighbor view generation. + +Multi-view key-frame feature fusion. A challenge arises from the lack of specific multi-view constraints during key-view generation, leading to degraded performance. While this is generally acceptable due to minimal overlapping regions, long-term generation introduces interdependencies among views. For example, during overtaking, a vehicle initially captured by the rear camera later appears in the front camera. To address this, we propose key-frame conditions, which are the starting frames of all views, to establish correlations within the image latent space. Combined with temporal attention, the model learns the relationships across + +![](images/6bc43f600d0da8e506cfdfe42a49ad5713f793b7e007bf4805b60fbd4652ffc8.jpg) +Figure 2. The DriveScape pipeline, built on the LDM framework, generates street-view videos conditioned on scene annotations, BEV maps, and 3D bounding boxes for each view. Keyframe conditions are introduced, in collaboration with temporal and spatial self-attention modules, to ensure consistency across spatial and temporal dimensions. Our unified model enables multi-view video generation without the need for complex post-processing or refinement procedures. Additionally, the Bi-Directional Modulated Transformer (BiMoT) module, consisting of two cross-attention layers with opposite directions and a temporal self-attention layer, ensures effective alignment and integration of various 3D road structural information, enabling precise control over video generation. + +all views at the initial frame and the preceding frames of neighboring cameras. + +Sparse multi-modal feature fusion. Our model supports training with videos at various frame rates, ranging from 2Hz to 10Hz, while maintaining sample conditions at 2Hz. A lightweight encoder is used to encode the 2Hz map and bounding boxes (BBoxes), with condition-missing frames padded with zeros. Additionally, an extra mask channel is introduced to indicate the presence of conditions at each frame, where mask values are 0 or 1. These inputs are processed through the BiMoT module. Embeddings for unconditioned frames are learned and integrated into the conditions to align with the video frame rate. This diverse frame-rate training scheme, combined with the BiMoT module, enhances sparse condition control capability. + +Conditions dropout. To enhance robustness, key views are trained in random order. Neighbor camera videos and conditions are dropped with probabilities of $50\%$ and $20\%$ , respectively. Unlike [16], our model requires no postrefinement, simultaneously learning multi-view and temporal consistency to improve quality and realism. + +# 3.4. Inference Stage + +During inference, our pipeline enables efficient and consistent generation of multi-view videos. Similar to the training stage, the process consists of two patterns: key-view infer + +ence and neighbor-view inference, as shown in Fig. 2. Keyview camera videos are generated simultaneously without cross-view condition control, as the key-view cameras are intentionally placed non-adjacently. For example, one camera is positioned at the front, while two are at the rear—one on the left and the other on the right. This strategic arrangement ensures comprehensive coverage from multiple perspectives with minimal overlap. + +In this configuration, neighbor-view cameras—two at the front and one at the rear—naturally receive cross-view conditional information from their adjacent key-view cameras. This design allows neighbor-view inference to maintain consistency with conditions from neighboring cameras, ensuring cross-view coherence. + +Leveraging this setup, our inference pipeline excels in parallel processing, requiring only two consecutive iterations to generate multi-view videos. This approach not only accelerates video production but also ensures cross-view consistency and realism throughout. + +# 3.5. Bi-Directional Modulated Transformer + +To integrate multi-conditions from different modalities into the UNet model and ensure effective articulation of each condition, we propose the Bi-Directional Modulated Transformer (BiMoT) module. This module aligns features from various modalities within the latent space, reinforces tem + +poral coherence among the merged conditions, and injects the aligned conditions into the network. As illustrated in Fig. 1, the module consists of two cross-attention layers with opposite directions and a temporal self-attention layer. + +Latent to Condition Attention To address diverse conditions with different modalities, we employ a latent-to-condition cross-attention mechanism to standardize feature representations. Pixel-level features are used as signals to merge various condition features. Specifically, let $F_{cond}$ represent condition features and $F_{latent}$ represent video latent features. The output is computed as: + +$$ +F _ {c o n d} ^ {\prime} = \operatorname {c r o s s a t t n} \left(F _ {c o n d}, F _ {l a t e n t}\right), \tag {1} +$$ + +where the query is $F_{cond}$ , and the key and value are $F_{latent}$ . This cross-attention operation facilitates the flow of video latent information into the condition space, providing a unified representation for integrating all modalities. + +Temporal Attention. We apply temporal self-attention to ensure that the merged features propagate effectively along the temporal dimension. Sparse condition information is also incorporated into the image latent space, naturally unifying all modalities into a single representation. This enables temporal information from latent features to interact with sparse conditions over time, guiding the generation process with sparse control. This mechanism ensures consistency and precise control over the temporal dimensions. + +Condition to Video Attention. Similar to the latent-to-condition cross-attention, a condition-to-video cross-attention mechanism is applied: + +$$ +F _ {l a t e n t} ^ {\prime} = \operatorname {c r o s s a t t n} \left(F _ {l a t e n t}, F _ {c o n d}\right). \tag {2} +$$ + +This operation facilitates the flow of merged condition feature information into the image latents, ultimately enabling unified control over video generation. + +Coarse to Fine Condition Mixture. BiMoT is flexible in accommodating various types of conditions, enabling easy extension to additional inputs. We begin with maps and 3D layouts, then introduce BEV key-frames, and finally integrate neighbor camera videos. This progression allows generation control to transition from a coarse level, driven by maps and 3D layouts, to a finer level, guided by pixel-level details. This ensures both structural coherence and detailed fidelity. The flexibility of BiMoT enables fine-grained layout controllability with unified mono-condition and multi-condition control. + +# 4. Experiments + +In this section, we evaluate our model through key metrics, outlining the experimental setup, dataset, and evaluation criteria. We then analyze qualitative results and conduct ablation studies to validate our approach. + +# 4.1. Experiment Details + +Dataset. The training data is sourced from the real-world driving dataset nuScenes [6], a widely used benchmark for BEV segmentation and 3D object detection in driving scenarios. Following the official configuration, we use 700 street-view scenes for training and 150 for validation, targeting 8 object classes and 8 road classes. + +Evaluation Metrics. To evaluate the effectiveness of our approach in delivering realism, continuity, and precise control over multiple conditions, we employ four key metrics for comparison against existing multi-view image and video generation methods. For realism, we use the Fréchet Inception Distance (FID) [11] to measure image quality. Temporal coherence is assessed using the Fréchet Video Distance (FVD) [28], ensuring consistent and fluid motion in the generated videos. Controllability is evaluated through two perception-based methods: CVT [39] for BEV segmentation and BEVFusion [19] for 3D object detection. Both methods are trained on the nuScenes dataset, providing a reliable benchmark for performance evaluation. Additionally, since our model generates videos, we follow Panacea [33] and use the video-based perception model StreamPETR [29] to evaluate the 3D object detection task. + +# 4.2. Experiments Results + +Qualitative Results. We showcase the quality and consistency of our video generation results in Fig. 3. Our model achieves notable consistency across both frames and viewpoints. Specifically, vehicles are accurately positioned according to the provided 3D layouts, and street delineations closely match the supplied projected maps. Remarkably, even when bounding boxes extend beyond the visible area, our model adheres to the specified conditions and generates the remaining objects with high fidelity. + +Sparse Condition Control As shown in Fig. 5, our model generates consistent videos in both time and space under sparse conditions, achieving a resolution of $576 \times 1024$ and a frame rate of 10 fps. To compare with related works on sparse control [9], we implemented a sparse control model with multi-condition fusion, as existing methods lack support for multi-condition control. A visualization comparing the reproduced SparseCtrl [9] and our method is presented in Fig. 4. Our method clearly demonstrates superior spatial consistency. Moreover, the FID and FVD scores for the reproduced SparseCtrl are 11.15 and 116.00, respectively, significantly worse than those achieved by our model (8.34 and 76.39). For additional visual comparisons and details on the multi-condition fusion SparseCtrl, please refer to the supplementary materials and the accompanying web page. + +Quantitative Results. We report quantitative metrics on the nuScenes validation set in Table 2. Using FID and FVD, we evaluate image and video quality, respectively. Following MagicDrive [7] and DrivingDiffusion [16], we + +
MethodFID↓FVD↓Road mIoU ↑Vehicle mIoU ↑Object NDS ↑
Oracle--73.6834.8441.21
BEVGen[25]25.54-50.20 (-23.48)5.89 (-28.95)-
BEVControl [35]24.85-60.80 (-12.88)26.80 (-8.04)-
MagicDrive [7]16.20-61.05 (-12.63)27.01 (-7.83)23.32 (-17.89)
DriveDreamer [31]52.60452.0---
DrivingDiffusion [16]15.83332.063.20 (-10.48)31.60 (-3.24)33.1 (-8.11)
Drive-WM [32]15.80122.7065.07 (-8.61)27.19 (-7.65)-
Delphi [21]15.08113.5--36.58 (-4.63)
SparseCtrl[9]11.15116.00---
Ours8.3476.3968.21 (-5.47)30.27 (-4.57)36.5 (-4.71)
+ +Table 2. The comparative analysis of our method against others on the nuScenes validation dataset is detailed for each task. Our end-to-end approach outperformed most baseline models without requiring any post-refinement process. + +![](images/1e538604e8c82383bd5b8a362be456dd51e85689478209710856b9e55e99afd1.jpg) +Figure 3. Showcases for multi-view video generation by DriveScape demonstrate frame-to-frame and viewpoint consistency. Vehicles generated by our model are accurately positioned according to the provided 3D layouts, while the street depictions closely align with the projected maps. + +![](images/6f481d41397efc002e0ab76879aaeaff7071e2d6825cfcd60703a6388f91a836.jpg) +Figure 4. Results in each row are Ours, SparseCtrl, input condition and GT respectively, which illustrate that our method has better spatial consistency than SparseCtrl. + +generate data from validation set labels and assess it with a pre-trained perception model. The decline in Road mIoU + +and Vehicle mIoU, compared to original data, measures the effectiveness of our framework in producing perception-friendly synthetic data. + +To align with previous studies, we resized generated images to $224 \times 400$ . We compared our FID results with three image generation models: BEVGen [25], BEVControl [35], and MagicDrive [7]. Our method significantly outperformed these models. Notably, video generation models typically yield higher FID scores than image models. For instance, Drive-WM, using only an image model, achieved better FID than the full model, highlighting the challenges of video generation. + +For video quality, assessed via the FVD metric, our approach showed substantial improvements over recent works + +![](images/fc1328671988eaa6e2b3787d4b8518b4976c53b1944803ee87c2d7fe01020ea6.jpg) +Figure 5. Showcase for sparse Condition Control. + +such as DriveDreamer [31], DrivingDiffusion [16], and Drive-WM [32]. In BEV segmentation metrics, our method achieved superior perceptual performance, except for a slightly lower Road mIoU compared to DrivingDiffusion, which utilized a post-processing network for refinement. To maintain simplicity, our end-to-end system applies no post-processing beyond image resizing for perceptual evaluation. The perception models we used operate on single-image inference, not leveraging the temporal consistency of our generated videos. Given our significant lead in FVD, our method holds strong potential for training video BEV perception models. + +
dataNDS ↑FID ↓FVD ↓
Oracle46.9--
Panacea32.1 (-14.8)16.96139
Oracle53.7--
Ours44.5 (-9.2)8.3476.39
+ +# 4.3. Ablation Studies + +To refine and evaluate the efficacy of our proposed framework, we conducted ablation studies, as detailed in Table 4. Various strategies for implementing conditional guidance were tested. Removing the entire BiMoT module resulted in the FID increasing from 8.34 to 20.98 and the FVD rising from 76.39 to 142.81, underscoring the module's importance for multi-condition encoding. Similarly, removing the temporal attention layer within BiMoT caused a significant performance drop, with FID and FVD increasing to 20.67 and 123.03, respectively, highlighting the role of temporal modelling in BiMoT. Additionally, during testing, we separately removed the key-frame condition and the neighbor condition. Both led to a performance decline, indicating the + +importance of these conditions for facilitating information exchange between views. + +Table 3. Compared with Panacea, we evaluate results on the nuScenes validation set using the pre-trained StreamPETR model. + +
TeAttnBiMOTKFNeFIDFVD
✓✓✓20.98142.81
✓✓20.67123.03
✓✓✓11.60108.80
✓✓✓12.18111.36
✓✓✓✓8.3476.39
+ +Table 4. Ablation Study. TeAttn is temporal attention module in BiMOT, KF is key-frame condition and Ne is neighbor condition + +# 4.4. Visualization of Different Settings + +In our visual showcases, we demonstrate the model's proficiency in manipulating video predictions for a single scene under various control conditions. The control conditions, sourced from frames of identical perspectives but randomly selected from other scenes, test the decoupling of conditions and the robustness of our method across diverse scenarios. + +As shown in Fig. 7, the experiments explore the influence of different 3D layout controls. Using a single reference frame from a highway segment, we first depict vehicles driving normally, then transition to vehicles traversing grassy terrain. Remarkably, our model maintains temporal consistency even in this unconventional context. Additional experiments involve removing or adding large vehicles, with results aligning closely with the specified conditions, showcasing precise control. + +Furthermore, as illustrated in Fig. 8, we investigate the impact of different maps on video generation. The generated scenes align impressively with the shapes of the provided map regions, maintaining layout feature consistency despite map changes. These experiments provide strong evidence of our model's capability to render specified control conditions into precise and coherent visual narratives within the generated video content. + +# 5. Conclusions + +The paper introduces DriveScape, the first end-to-end multi-view driving video generative model. It enables consistent multi-view video generation without complex postprocessing or refinement. This unified framework streamlines multi-view condition-guided video generation. We also propose a Bi-Directional Modulated Transformer module for diverse condition guidance. Our approach achieves decoupled and precise control of dynamic foreground and static background, delivering state-of-the-art video synthesis on the nuScenes dataset. It excels in multi-view image and video generation quality and perception applications. In summary, our method outperforms others in generation quality and ease of application. + +![](images/b565b73afdef9a50ea2277898823f855298278724094fe607063f07c0966b9f0.jpg) +Figure 6. Showcase for maps and 3D layouts control under different conditions. (a): control by the reference maps and 3D Layouts; (b) change 3D Layouts with more vehicles; (c) change the maps with curved pedestrian crosswalk; (d) remove all the 3D Layouts. + +![](images/8c640171de84fd850e911dc18e017793ae07b0d09d31c897555c874686e42fc4.jpg) +Figure 7. Showcases for 3D layouts control under different conditions. DriveScape can precisely control dynamic objects. (a) reference layouts; (b) move car to the grass; (c) remove all cars; (d) add trucks to the scene. + +![](images/0812ffb9511188af0cf2e05b163d70fb142cf525153a32c4b5a3203d8538d3ba.jpg) +Figure 8. Showcases for maps control under. DriveScape can precisely control static background. (a) replace the friable area with green belts; (b) replace the friable area with green belts; (c) replace the friable area with sidewalks; (d) add island on the road. + +# References + +[1] Jie An, Songyang Zhang, Harry Yang, Sonal Gupta, Jia-Bin Huang, Jiebo Luo, and Xi Yin. Latent-shift: Latent diffusion with temporal shift for efficient text-to-video generation. arXiv preprint arXiv:2304.08477, 2023. 2 +[2] Andreas Blattmann, Tim Dockhorn, Sumith Kulal, Daniel Mendelevitch, Maciej Kilian, Dominik Lorenz, Yam Levi, Zion English, Vikram Voleti, Adam Letts, Varun Jampani, and Robin Rombach. Stable video diffusion: Scaling latent video diffusion models to large datasets, 2023. 2 +[3] Andreas Blattmann, Tim Dockhorn, Sumith Kulal, Daniel Mendelevitch, Maciej Kilian, Dominik Lorenz, Yam Levi, Zion English, Vikram Voleti, Adam Letts, et al. Stable video diffusion: Scaling latent video diffusion models to large datasets. arXiv preprint arXiv:2311.15127, 2023. 2 +[4] Andreas Blattmann, Robin Rombach, Huan Ling, Tim Dockhorn, Seung Wook Kim, Sanja Fidler, and Karsten Kreis. Align your latents: High-resolution video synthesis with latent diffusion models. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 22563-22575, 2023. 2 +[5] Andreas Blattmann, Robin Rombach, Huan Ling, Tim Dockhorn, Seung Wook Kim, Sanja Fidler, and Karsten Kreis. Align your latents: High-resolution video synthesis with latent diffusion models. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 22563-22575, 2023. 2 +[6] Holger Caesar, Varun Bankiti, Alex H Lang, Sourabh Vora, Venice Erin Liong, Qiang Xu, Anush Krishnan, Yu Pan, Giancarlo Baldan, and Oscar Beijbom. nuscenes: A multimodal dataset for autonomous driving. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 11621-11631, 2020. 5 +[7] Ruiyuan Gao, Kai Chen, Enze Xie, Lanqing Hong, Zhenguo Li, Dit-Yan Yeung, and Qiang Xu. Magicdrive: Street view generation with diverse 3d geometry control. arXiv preprint arXiv:2310.02601, 2023. 1, 2, 5, 6 +[8] Yuwei Guo, Ceyuan Yang, Anyi Rao, Yaohui Wang, Yu Qiao, Dahua Lin, and Bo Dai. Animatediff:Animate your personalized text-to-image diffusion models without specific tuning. arXiv preprint arXiv:2307.04725, 2023. 2 +[9] Yuwei Guo, Ceyuan Yang, Anyi Rao, Maneesh Agrawala, Dahua Lin, and Bo Dai. Sparsectrl: Adding sparse controls to text-to-video diffusion models. In European Conference on Computer Vision, pages 330-348. Springer, 2025. 5, 6 +[10] Yingqing He, Tianyu Yang, Yong Zhang, Ying Shan, and Qifeng Chen. Latent video diffusion models for high-fidelity video generation with arbitrary lengths. arXiv preprint arXiv:2211.13221, 2022. 2 +[11] Martin Heusel, Hubert Ramsauer, Thomas Unterthiner, Bernhard Nessler, and Sepp Hochreiter. Gans trained by a two time-scale update rule converge to a local nash equilibrium. Advances in neural information processing systems, 30, 2017. 5 +[12] Jonathan Ho, William Chan, Chitwan Sahara, Jay Whang, Ruiqi Gao, Alexey Gritsenko, Diederik P Kingma, Ben Poole, Mohammad Norouzi, David J Fleet, et al. Imagen + +video: High definition video generation with diffusion models. arXiv preprint arXiv:2210.02303, 2022. 2 +[13] Yihan Hu, Jiazhi Yang, Li Chen, Keyu Li, Chonghao Sima, Xizhou Zhu, Siqi Chai, Senyao Du, Tianwei Lin, Wenhai Wang, et al. Goal-oriented autonomous driving. arXiv preprint arXiv:2212.10156, 2022. 1 +[14] Junjie Huang and Guan Huang. Bevdet4d: Exploit temporal cues in multi-camera 3d object detection. arXiv preprint arXiv:2203.17054, 2022. 1 +[15] Junjie Huang, Guan Huang, Zheng Zhu, Yun Ye, and Dalong Du. Bevdet: High-performance multi-camera 3d object detection in bird-eye-view. arXiv preprint arXiv:2112.11790, 2021. 1 +[16] Xiaofan Li, Yifu Zhang, and Xiaoqing Ye. Drivingdiffusion: Layout-guided multi-view driving scene video generation with latent diffusion model. arXiv preprint arXiv:2310.07771, 2023. 2, 3, 4, 5, 6 +[17] Yinhao Li, Zheng Ge, Guanyi Yu, Jinrong Yang, Zengran Wang, Yukang Shi, Jianjian Sun, and Zeming Li. Bevdepth: Acquisition of reliable depth for multi-view 3d object detection. In Proceedings of the AAAI Conference on Artificial Intelligence, pages 1477-1485, 2023. 1 +[18] Zhiqi Li, Wenhai Wang, Hongyang Li, Enze Xie, Chonghao Sima, Tong Lu, Yu Qiao, and Jifeng Dai. Bevformer: Learning bird's-eye-view representation from multi-camera images via spatiotemporal transformers. In European conference on computer vision, pages 1-18. Springer, 2022. 1 +[19] Tingting Liang, Hongwei Xie, Kaicheng Yu, Zhongyu Xia, Zhiwei Lin, Yongtao Wang, Tao Tang, Bing Wang, and Zhi Tang. Bevfusion: A simple and robust lidar-camera fusion framework. Advances in Neural Information Processing Systems, 35:10421-10434, 2022. 5 +[20] Haicheng Liao, Yongkang Li, Zhenning Li, Chengyue Wang, Zhiyong Cui, Shengbo Eben Li, and Chengzhong Xu. A cognitive-based trajectory prediction approach for autonomous driving. arXiv preprint arXiv:2402.19251, 2024. 1 +[21] Enhui Ma, Lijun Zhou, Tao Tang, Zhan Zhang, Dong Han, Junpeng Jiang, Kun Zhan, Peng Jia, Xianpeng Lang, Haiyang Sun, et al. Unleashing generalization of end-to-end autonomous driving with controllable long video generation. arXiv preprint arXiv:2406.01349, 2024. 6 +[22] Robin Rombach, Andreas Blattmann, Dominik Lorenz, Patrick Esser, and Björn Ommer. High-resolution image synthesis with latent diffusion models. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 10684-10695, 2022. 2 +[23] Olaf Ronneberger, Philipp Fischer, and Thomas Brox. U-net: Convolutional networks for biomedical image segmentation. In Medical Image Computing and Computer-Assisted Intervention-MICCAI 2015: 18th International Conference, Munich, Germany, October 5-9, 2015, Proceedings, Part III 18, pages 234-241. Springer, 2015. 3 +[24] Uriel Singer, Adam Polyak, Thomas Hayes, Xi Yin, Jie An, Songyang Zhang, Qiyuan Hu, Harry Yang, Oron Ashual, Oran Gafni, Devi Parikh, Sonal Gupta, and Yaniv Taigman. Make-a-video: Text-to-video generation without text-video data, 2022. 2 + +[25] Alexander Swerdlow, Runsheng Xu, and Bolei Zhou. Street-view image generation from a bird's-eye view layout. IEEE Robotics and Automation Letters, 2024. 1, 2, 6 +[26] Shitao Tang, Fuyang Zhang, Jiacheng Chen, Peng Wang, and Yasutaka Furukawa. Mvdiffusion: Enabling holistic multiview image generation with correspondence-aware diffusion, 2023. 2 +[27] Hung-Yu Tseng, Qinbo Li, Changil Kim, Suhib Alsisan, Jia-Bin Huang, and Johannes Kopf. Consistent view synthesis with pose-guided diffusion models. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 16773-16783, 2023. 2 +[28] Thomas Unterthiner, Sjoerd Van Steenkiste, Karol Kurach, Raphael Marinier, Marcin Michalski, and Sylvain Gelly. Towards accurate generative models of video: A new metric & challenges. arXiv preprint arXiv:1812.01717, 2018. 5 +[29] Shihao Wang, Yingfei Liu, Tiancai Wang, Ying Li, and Xi-angyu Zhang. Exploring object-centric temporal modeling for efficient multi-view 3d object detection. In CVPR, pages 3621–3631, 2023. 5 +[30] Wenjing Wang, Huan Yang, Zixi Tuo, Huiguo He, Junchen Zhu, Jianlong Fu, and Jiaying Liu. Videofactory: Swap attention in spatiotemporal diffusions for text-to-video generation, 2023. 2 +[31] Xiaofeng Wang, Zheng Zhu, Guan Huang, Xinze Chen, and Jiwen Lu. Drivedreamer: Towards real-world-driven world models for autonomous driving. arXiv preprint arXiv:2309.09777, 2023. 6 +[32] Yuqi Wang, Jiawei He, Lue Fan, Hongxin Li, Yuntao Chen, and Zhaoxiang Zhang. Driving into the future: Multiview visual forecasting and planning with world model for autonomous driving. arXiv preprint arXiv:2311.17918, 2023. 6 +[33] Yuqing Wen, Yucheng Zhao, Yingfei Liu, Fan Jia, Yanhui Wang, Chong Luo, Chi Zhang, Tiancai Wang, Xiaoyan Sun, and Xiangyu Zhang. Panacea: Panoramic and controllable video generation for autonomous driving. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 6902-6912, 2024. 5 +[34] Chenyu Yang, Yuntao Chen, Hao Tian, Chenxin Tao, Xizhou Zhu, Zhaoxiang Zhang, Gao Huang, Hongyang Li, Yu Qiao, Lewei Lu, et al. Bevformer v2: Adapting modern image backbones to bird's-eye-view recognition via perspective supervision. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 17830-17839, 2023. 1 +[35] Kairui Yang, Enhui Ma, Jibin Peng, Qing Guo, Di Lin, and Kaicheng Yu. Bevcontrol: Accurately controlling street-view elements with multi-perspective consistency via bev sketch layout. arXiv preprint arXiv:2308.01661, 2023. 1, 2, 6 +[36] Yuxiang Yang, Yingqi Deng, Jing Zhang, Jiahao Nie, and Zheng-Jun Zha. Bevtrack: A simple and strong baseline for 3d single object tracking in bird's-eye view. arXiv e-prints, pages arXiv-2309, 2023. 1 +[37] David Junhao Zhang, Dongxu Li, Hung Le, Mike Zheng Shou, Caiming Xiong, and Doyen Sahoo. Moonshot: To + +wards controllable video generation and editing with multimodal conditions, 2024. 2 +[38] Wenzhao Zheng, Ruiqi Song, Xianda Guo, and Long Chen. Genad: Generative end-to-end autonomous driving. arXiv preprint arXiv:2402.11502, 2024. 1 +[39] Brady Zhou and Philipp Krahenbuhl. Cross-view transformers for real-time map-view semantic segmentation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 13760-13769, 2022. 1, 5 +[40] Daquan Zhou, Weimin Wang, Hanshu Yan, Weiwei Lv, Yizhe Zhu, and Jiashi Feng. Magicvideo: Efficient video generation with latent diffusion models, 2023. 2 \ No newline at end of file diff --git a/2025/DriveScape_ High-Resolution Driving Video Generation by Multi-View Feature Fusion/images.zip b/2025/DriveScape_ High-Resolution Driving Video Generation by Multi-View Feature Fusion/images.zip new file mode 100644 index 0000000000000000000000000000000000000000..a1406379f35f959658a3d12b922986977e708826 --- /dev/null +++ b/2025/DriveScape_ High-Resolution Driving Video Generation by Multi-View Feature Fusion/images.zip @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:567b5d7d434e7f29c030a11db8b4fcad6d9fa7c30d140c0b9365813030d7c2f0 +size 765957 diff --git a/2025/DriveScape_ High-Resolution Driving Video Generation by Multi-View Feature Fusion/layout.json b/2025/DriveScape_ High-Resolution Driving Video Generation by Multi-View Feature Fusion/layout.json new file mode 100644 index 0000000000000000000000000000000000000000..3c3e7ba73cd82ac96a2ae075f80a728f985b5eb4 --- /dev/null +++ b/2025/DriveScape_ High-Resolution Driving Video Generation by Multi-View Feature Fusion/layout.json @@ -0,0 +1,5934 @@ +{ + "pdf_info": [ + { + "para_blocks": [ + { + "bbox": [ + 124, + 103, + 487, + 138 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 124, + 103, + 487, + 138 + ], + "spans": [ + { + "bbox": [ + 124, + 103, + 487, + 138 + ], + "type": "text", + "content": "DriveScape: High-Resolution Driving Video Generation by Multi-View Feature Fusion" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 133, + 160, + 477, + 246 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 160, + 477, + 246 + ], + "spans": [ + { + "bbox": [ + 133, + 160, + 477, + 246 + ], + "type": "text", + "content": "Wei Wu" + }, + { + "bbox": [ + 133, + 160, + 477, + 246 + ], + "type": "inline_equation", + "content": "^{1,2}" + }, + { + "bbox": [ + 133, + 160, + 477, + 246 + ], + "type": "text", + "content": ", Xi Guo" + }, + { + "bbox": [ + 133, + 160, + 477, + 246 + ], + "type": "inline_equation", + "content": "^{2}" + }, + { + "bbox": [ + 133, + 160, + 477, + 246 + ], + "type": "text", + "content": ", Weixuan Tang" + }, + { + "bbox": [ + 133, + 160, + 477, + 246 + ], + "type": "inline_equation", + "content": "^{2}" + }, + { + "bbox": [ + 133, + 160, + 477, + 246 + ], + "type": "text", + "content": ", Tingxuan Huang" + }, + { + "bbox": [ + 133, + 160, + 477, + 246 + ], + "type": "inline_equation", + "content": "^{3}" + }, + { + "bbox": [ + 133, + 160, + 477, + 246 + ], + "type": "text", + "content": ", Chiyu Wang" + }, + { + "bbox": [ + 133, + 160, + 477, + 246 + ], + "type": "inline_equation", + "content": "^{2}" + }, + { + "bbox": [ + 133, + 160, + 477, + 246 + ], + "type": "text", + "content": ", Chenjing Ding" + }, + { + "bbox": [ + 133, + 160, + 477, + 246 + ], + "type": "inline_equation", + "content": "^{2*}" + }, + { + "bbox": [ + 133, + 160, + 477, + 246 + ], + "type": "inline_equation", + "content": "^{1}" + }, + { + "bbox": [ + 133, + 160, + 477, + 246 + ], + "type": "text", + "content": "Tsinghua University " + }, + { + "bbox": [ + 133, + 160, + 477, + 246 + ], + "type": "inline_equation", + "content": "^{2}" + }, + { + "bbox": [ + 133, + 160, + 477, + 246 + ], + "type": "text", + "content": "Sensetime Research " + }, + { + "bbox": [ + 133, + 160, + 477, + 246 + ], + "type": "inline_equation", + "content": "^{3}" + }, + { + "bbox": [ + 133, + 160, + 477, + 246 + ], + "type": "text", + "content": "Northeastern University {wuwei, quoxi, dingchenjing} @ sensetime.com, {tangweixuan, wangchiyu} @ senseauto.com, {huangtingxuan} @stu.neu.edu.cn" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 151, + 273, + 200, + 285 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 151, + 273, + 200, + 285 + ], + "spans": [ + { + "bbox": [ + 151, + 273, + 200, + 285 + ], + "type": "text", + "content": "Abstract" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 54, + 298, + 297, + 586 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 54, + 298, + 297, + 586 + ], + "spans": [ + { + "bbox": [ + 54, + 298, + 297, + 586 + ], + "type": "text", + "content": "Recent advancements in generative models offer promising solutions for synthesizing realistic driving videos, aiding in training autonomous driving perception models. However, existing methods often struggle with high-resolution multi-view generation, mainly due to the significant memory and computational overhead caused by simultaneously inputting multi-view videos into denoising diffusion models. In this paper, we propose a driving video generation framework based on multi-view feature fusion named DriveScape for multi-view 3D condition-guided video generation. We introduce a Bi-Directional Modulated Transformer (BiMoT) module to encode, fuse and inject multiview features along with various 3D road structures and objects, which enables high-resolution multi-view generation. Consequently, our approach allows precise control over video generation, greatly enhancing realism and providing a robust solution for creating high-quality, multi-view driving videos. Our framework achieves state-of-the-art results on the nuScenes dataset, demonstrating impressive generative quality metrics with an FID score of 8.34 and an FVD score of 76.39, as well as superior performance across various perception tasks. This lays the foundation for more accurate environment simulation in autonomous driving. Our project homepage." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 56, + 611, + 137, + 624 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 611, + 137, + 624 + ], + "spans": [ + { + "bbox": [ + 56, + 611, + 137, + 624 + ], + "type": "text", + "content": "1. Introduction" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 55, + 632, + 296, + 693 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 632, + 296, + 693 + ], + "spans": [ + { + "bbox": [ + 55, + 632, + 296, + 693 + ], + "type": "text", + "content": "Autonomous driving has garnered significant attention, emphasizing the need for precise environmental understanding to ensure safe driving decisions [13, 38]. Bird's-eye view (BEV) maps, generated from multi-view images, serve as a crucial structured representation for various tasks, includ" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 313, + 274, + 555, + 406 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 274, + 555, + 406 + ], + "spans": [ + { + "bbox": [ + 313, + 274, + 555, + 406 + ], + "type": "text", + "content": "ing 3D object detection, segmentation, tracking, depth estimation, and trajectory prediction [14, 15, 17, 18, 20, 34, 36, 39]. However, acquiring high-quality multi-view video data poses significant challenges due to the high labeling costs, increasing the demand for generating temporally consistent video data aligned with real-world distributions and 3D labels. Recent generative models have demonstrated the potential of synthetic data to enhance visual tasks [7, 25, 35], while diffusion models have advanced the generation of diverse and realistic driving videos, which are crucial for developing robust autonomous driving systems." + } + ] + } + ], + "index": 8 + }, + { + "type": "table", + "bbox": [ + 328, + 417, + 541, + 500 + ], + "blocks": [ + { + "bbox": [ + 328, + 417, + 541, + 500 + ], + "lines": [ + { + "bbox": [ + 328, + 417, + 541, + 500 + ], + "spans": [ + { + "bbox": [ + 328, + 417, + 541, + 500 + ], + "type": "table", + "html": "
ModelSpatial Res.Temporal Res. (Hz)sparse condition
Drive-WM192 × 3842×
WoVoGen256 × 4482×
Delphi512 × 5122×
GenAD256 × 4482×
MagicDrive272 × 736-×
DriveDreamer 1&2448 × 2562×
DriveDiffusion512 × 5122×
Panacea256 × 5122×
Ours (DriveScape)576 × 10242 - 10✓
", + "image_path": "a08c3e9b9b415b39b3ccabdaacf4a8a3edfe270de75fbdceb21cbbea1b1677ff.jpg" + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "table_body" + } + ], + "index": 9 + }, + { + "bbox": [ + 313, + 508, + 555, + 552 + ], + "lines": [ + { + "bbox": [ + 313, + 508, + 555, + 552 + ], + "spans": [ + { + "bbox": [ + 313, + 508, + 555, + 552 + ], + "type": "text", + "content": "Table 1. Real-world multi-view driving controllable world models. DriveScape (ours) operates at a high spatiotemporal resolution, offering support for sparse layout controllability, as well as unified control for both mono-condition and multi-condition scenarios." + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "text" + }, + { + "bbox": [ + 313, + 558, + 556, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 558, + 556, + 715 + ], + "spans": [ + { + "bbox": [ + 313, + 558, + 556, + 715 + ], + "type": "text", + "content": "The most relevant approach to our work is the generation of multi-view driving videos controlled by 3D conditions. In terms of spatial resolution, general video generation tasks, such as those addressed by SVD, have achieved high-resolution video generation at " + }, + { + "bbox": [ + 313, + 558, + 556, + 715 + ], + "type": "inline_equation", + "content": "1024 \\times 576" + }, + { + "bbox": [ + 313, + 558, + 556, + 715 + ], + "type": "text", + "content": ". However, in driving scenarios, multi-camera panoramic generation is often required. Previous methods typically need to denoise multiple views simultaneously, leading to memory constraints during both training and inference. For temporal resolution, previous methods rely on frame-by-frame 3D layout control. Due to high annotation costs, public datasets like NuScenes provide annotations at a low frame rate (2Hz), restricting models from generating videos at the" + } + ] + } + ], + "index": 11 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "spans": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "type": "text", + "content": "CVF" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 145, + 0, + 496, + 37 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 145, + 0, + 496, + 37 + ], + "spans": [ + { + "bbox": [ + 145, + 0, + 496, + 37 + ], + "type": "text", + "content": "This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 67, + 703, + 144, + 713 + ], + "type": "page_footnote", + "angle": 0, + "lines": [ + { + "bbox": [ + 67, + 703, + 144, + 713 + ], + "spans": [ + { + "bbox": [ + 67, + 703, + 144, + 713 + ], + "type": "text", + "content": "*Corresponding author" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "type": "text", + "content": "17187" + } + ] + } + ], + "index": 13 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 0 + }, + { + "para_blocks": [ + { + "bbox": [ + 55, + 72, + 294, + 180 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 72, + 294, + 180 + ], + "spans": [ + { + "bbox": [ + 55, + 72, + 294, + 180 + ], + "type": "text", + "content": "rate (2Hz), restricting models from generating videos at the same frame rate. Additionally, the cost of obtaining different types of annotations varies in practice. For instance, annotating 3D lane lines is typically more expensive than annotating bounding boxes, often resulting in lane lines being labeled at a lower frame rate. This necessitates generative models that can handle sparse conditional control and remain compatible with both sparse and non-sparse conditions simultaneously." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 56, + 182, + 295, + 372 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 182, + 295, + 372 + ], + "spans": [ + { + "bbox": [ + 56, + 182, + 295, + 372 + ], + "type": "text", + "content": "To address these challenges, we propose DriveScape, a solution for multi-view 3D condition-guided video generation with high spatio-temporal resolution. Unlike existing methods, DriveScape integrates adjacent views and keyframes through multi-view fusion, avoiding the computational burden of denoising multiple views simultaneously. We introduce a streamlined training and inference pipeline, free from post-processing, supporting high spatial and temporal resolution for sparse condition control. To align 3D road structural information, we propose a Bi-Directional Condition Alignment module, integrating multiple conditions into an aligned latent space for fine-grained layout controllability. By leveraging temporal and spatial information, DriveScape establishes a unified model for consistent multi-view video generation, significantly enhancing realism and applicability." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 67, + 373, + 268, + 384 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 67, + 373, + 268, + 384 + ], + "spans": [ + { + "bbox": [ + 67, + 373, + 268, + 384 + ], + "type": "text", + "content": "In summary, our key contributions are as follows:" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 56, + 386, + 295, + 577 + ], + "type": "list", + "angle": 0, + "index": 6, + "blocks": [ + { + "bbox": [ + 56, + 386, + 295, + 456 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 386, + 295, + 456 + ], + "spans": [ + { + "bbox": [ + 56, + 386, + 295, + 456 + ], + "type": "text", + "content": "- We propose a unified multi-view feature fusion framework for driving video generation without any post-processing. To the best of our knowledge, our model is the first to achieve high-resolution, high-frame-rate, sparse condition control for multi-view driving video generation." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 56, + 457, + 295, + 529 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 457, + 295, + 529 + ], + "spans": [ + { + "bbox": [ + 56, + 457, + 295, + 529 + ], + "type": "text", + "content": "- We introduce a Bi-Directional Condition Alignment module to facilitate diverse modalities of condition guidance and sparse condition control in video generation tasks. Experiments demonstrate that our model achieves decoupled, sparse, and precise control of dynamic foreground and static background." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 56, + 529, + 295, + 577 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 529, + 295, + 577 + ], + "spans": [ + { + "bbox": [ + 56, + 529, + 295, + 577 + ], + "type": "text", + "content": "- We achieve state-of-the-art video synthesis performance on the nuScenes dataset, excelling not only in the generation quality of multi-view images and videos but also across various perception applications." + } + ] + } + ], + "index": 5 + } + ], + "sub_type": "text" + }, + { + "bbox": [ + 55, + 590, + 146, + 604 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 590, + 146, + 604 + ], + "spans": [ + { + "bbox": [ + 55, + 590, + 146, + 604 + ], + "type": "text", + "content": "2. Related Works" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 55, + 611, + 189, + 624 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 611, + 189, + 624 + ], + "spans": [ + { + "bbox": [ + 55, + 611, + 189, + 624 + ], + "type": "text", + "content": "2.1. Controllable generation" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 55, + 629, + 295, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 629, + 295, + 713 + ], + "spans": [ + { + "bbox": [ + 55, + 629, + 295, + 713 + ], + "type": "text", + "content": "With the advent of diffusion models, significant progress has been made in text-to-video generation [1, 2, 5, 8, 10, 12, 24, 30, 40]. Video LDM [4] employs a latent diffusion pipeline where denoising occurs in the image latent space, greatly accelerating the process. However, text alone cannot precisely control video generation. Later methods introduced image blocks alongside text as prompts for the" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 313, + 72, + 553, + 156 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 72, + 553, + 156 + ], + "spans": [ + { + "bbox": [ + 313, + 72, + 553, + 156 + ], + "type": "text", + "content": "denoising network [37]. Our goal is to generate highly realistic street-view videos, a complex scene featuring numerous elements and interactions (e.g., intricate street layouts and moving vehicles). To achieve fine-grained control, we incorporate road maps, 3D bounding boxes, and BEV keyframes, going beyond text and images to ensure detailed and precise video generation." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 313, + 164, + 468, + 177 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 164, + 468, + 177 + ], + "spans": [ + { + "bbox": [ + 313, + 164, + 468, + 177 + ], + "type": "text", + "content": "2.2. Multi-view video generation" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 313, + 182, + 555, + 313 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 182, + 555, + 313 + ], + "spans": [ + { + "bbox": [ + 313, + 182, + 555, + 313 + ], + "type": "text", + "content": "Multi-view consistency and temporal consistency are two critical challenges in multi-view video generation. To address multi-view consistency, MVDiffusion [26] introduced a correspondence-aware attention module to align information across views. Tseng et al. [27] utilized epipolar geometry to regularize consistency between different views. MagicDrive [7] leveraged camera poses, bounding boxes, and road maps as priors, incorporating an additional cross-view attention block to enhance consistency. However, these methods are limited to generating multi-view images rather than videos." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 313, + 323, + 441, + 335 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 323, + 441, + 335 + ], + "spans": [ + { + "bbox": [ + 313, + 323, + 441, + 335 + ], + "type": "text", + "content": "2.3. Street view generation" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 313, + 340, + 555, + 544 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 340, + 555, + 544 + ], + "spans": [ + { + "bbox": [ + 313, + 340, + 555, + 544 + ], + "type": "text", + "content": "Most street-view generation models condition on 2D layouts, such as BEV maps, 2D bounding boxes, and semantic segmentation. BEVGen [25] encapsulates all semantics within BEV for street-view generation. BEVControl [35] employs a two-stage method to generate multi-view urban scene images from BEV layouts, where a controller generates foreground and background objects, and a coordinator combines them to maintain visual consistency across views. However, projecting 3D information to 2D loses geometric details, making direct extension to video generation prone to inconsistencies across frames. To address this, we incorporate 3D bounding boxes as a condition to guide generation.DrivingDiffusion [16] adopts a multi-stage pipeline with two models and two stages of post-processing to generate frames and extend them into videos. In contrast, our method avoids the complexity of multi-stage pipelines by using an efficient, end-to-end approach." + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 313, + 555, + 370, + 567 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 555, + 370, + 567 + ], + "spans": [ + { + "bbox": [ + 313, + 555, + 370, + 567 + ], + "type": "text", + "content": "3. Method" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 313, + 575, + 382, + 587 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 575, + 382, + 587 + ], + "spans": [ + { + "bbox": [ + 313, + 575, + 382, + 587 + ], + "type": "text", + "content": "3.1. Overview" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 313, + 594, + 553, + 675 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 594, + 553, + 675 + ], + "spans": [ + { + "bbox": [ + 313, + 594, + 553, + 675 + ], + "type": "text", + "content": "The overview of DriveScape is shown in Fig. 1. Based on the LDM pipeline [22], DriveScape generates street-view videos conditioned on scene annotations, such as BEV maps, 3D bounding boxes, ego states, and text for each view. The unified model, described in Section 3.2, enables high-resolution multi-view video generation without post-refinement." + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 313, + 677, + 553, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 677, + 553, + 713 + ], + "spans": [ + { + "bbox": [ + 313, + 677, + 553, + 713 + ], + "type": "text", + "content": "The Bi-Directional Condition Alignment module, detailed in Section 3.5, ensures effective alignment between 3D road structures and temporal sparse conditions." + } + ] + } + ], + "index": 18 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "text", + "content": "17188" + } + ] + } + ], + "index": 19 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 1 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 82, + 70, + 531, + 297 + ], + "blocks": [ + { + "bbox": [ + 82, + 70, + 531, + 297 + ], + "lines": [ + { + "bbox": [ + 82, + 70, + 531, + 297 + ], + "spans": [ + { + "bbox": [ + 82, + 70, + 531, + 297 + ], + "type": "image", + "image_path": "49b1dfe8d80e4b6efba66d7778586e6a150f77efc9405d2da381e86e00bfe69f.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 54, + 305, + 555, + 363 + ], + "lines": [ + { + "bbox": [ + 54, + 305, + 555, + 363 + ], + "spans": [ + { + "bbox": [ + 54, + 305, + 555, + 363 + ], + "type": "text", + "content": "Figure 1. End-to-End Multi-view Video Generation pipeline. We use learnable embedding vectors to represent different cameras and categorize camera views as key views and neighbour views. Our training scheme guides the generation of key view videos through their neighboring frames. Additionally, we introduce the key-frames condition along with a training and inference scheme to ensure multi-view consistency simultaneously. Furthermore, our model does not require any post-refine process as DriveDiffusion ([16]). It is capable of learning multi-view and temporal consistency simultaneously, resulting in high-fidelity street-view synthesis." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "bbox": [ + 55, + 381, + 145, + 392 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 381, + 145, + 392 + ], + "spans": [ + { + "bbox": [ + 55, + 381, + 145, + 392 + ], + "type": "text", + "content": "3.2. Unified Model" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 54, + 399, + 295, + 555 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 54, + 399, + 295, + 555 + ], + "spans": [ + { + "bbox": [ + 54, + 399, + 295, + 555 + ], + "type": "text", + "content": "Our model leverages SVD [3] as the backbone, with each denoising step processing a video from a single viewpoint. Control conditions are injected through a unified control module, including frame-wise images from adjacent viewpoints, road maps, 3D bounding boxes, and the first frame from each viewpoint. Starting with initial multi-view frames " + }, + { + "bbox": [ + 54, + 399, + 295, + 555 + ], + "type": "inline_equation", + "content": "x_{0:N-1}^{0}(N" + }, + { + "bbox": [ + 54, + 399, + 295, + 555 + ], + "type": "text", + "content": " is the number of views), DriveScape predicts subsequent frames " + }, + { + "bbox": [ + 54, + 399, + 295, + 555 + ], + "type": "inline_equation", + "content": "x_{0:N-1}^{1:T}(T" + }, + { + "bbox": [ + 54, + 399, + 295, + 555 + ], + "type": "text", + "content": " is the number of frames) under conditions such as BEV maps " + }, + { + "bbox": [ + 54, + 399, + 295, + 555 + ], + "type": "inline_equation", + "content": "m_{0:N-1}^{0:T}" + }, + { + "bbox": [ + 54, + 399, + 295, + 555 + ], + "type": "text", + "content": ", 3D bounding boxes " + }, + { + "bbox": [ + 54, + 399, + 295, + 555 + ], + "type": "inline_equation", + "content": "b_{0:N-1}^{0:T}" + }, + { + "bbox": [ + 54, + 399, + 295, + 555 + ], + "type": "text", + "content": ", BEV key frames, and neighboring camera videos. BEV maps and 3D bounding boxes are projected into the camera's First-Person View (FPV) and encoded with convolutional blocks." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 54, + 556, + 295, + 653 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 54, + 556, + 295, + 653 + ], + "spans": [ + { + "bbox": [ + 54, + 556, + 295, + 653 + ], + "type": "text", + "content": "As shown in Fig. 1, we use learnable embedding vectors " + }, + { + "bbox": [ + 54, + 556, + 295, + 653 + ], + "type": "inline_equation", + "content": "\\{e_i\\}" + }, + { + "bbox": [ + 54, + 556, + 295, + 653 + ], + "type": "text", + "content": " to represent different cameras. Similar to the processing of canbus data (e.g., velocity and direction angle), these embeddings are passed through a multilayer perceptron (MLP) followed by a Sigmoid Linear Unit (SiLU) activation. This enables the embeddings to perform crossattention with the image latent space, allowing the unified model to predict videos from multiple perspectives." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 54, + 654, + 295, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 54, + 654, + 295, + 714 + ], + "spans": [ + { + "bbox": [ + 54, + 654, + 295, + 714 + ], + "type": "text", + "content": "Embedding camera information alone cannot ensure spatial and temporal consistency, as it only captures global camera characteristics. Consistency is achieved through model design and training. For temporal consistency, we integrate temporal attention into the latent diffusion UNet [23]" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 313, + 381, + 555, + 418 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 381, + 555, + 418 + ], + "spans": [ + { + "bbox": [ + 313, + 381, + 555, + 418 + ], + "type": "text", + "content": "and control alignment module (Section 3.5). For multi-view consistency, we introduce keyframe conditions and a carefully designed training and inference scheme." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 313, + 430, + 406, + 442 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 430, + 406, + 442 + ], + "spans": [ + { + "bbox": [ + 313, + 430, + 406, + 442 + ], + "type": "text", + "content": "3.3. Training Stage" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 313, + 449, + 555, + 581 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 449, + 555, + 581 + ], + "spans": [ + { + "bbox": [ + 313, + 449, + 555, + 581 + ], + "type": "text", + "content": "Neighbor frames feature fusion. DriveScape categorizes camera views into key and neighbor views (Fig. 2). Frames of the same type have minimal overlap, ensuring distinct regions. The training scheme guides key view generation using neighbor frames. During training, neighbor and key views are selected sequentially. For neighbor views, video conditions from adjacent key views are input into the network. Key views are generated sequentially, stored, and incorporated into the cross-attention mechanism with the image latent space, ensuring effective guidance for neighbor view generation." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 313, + 582, + 556, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 582, + 556, + 715 + ], + "spans": [ + { + "bbox": [ + 313, + 582, + 556, + 715 + ], + "type": "text", + "content": "Multi-view key-frame feature fusion. A challenge arises from the lack of specific multi-view constraints during key-view generation, leading to degraded performance. While this is generally acceptable due to minimal overlapping regions, long-term generation introduces interdependencies among views. For example, during overtaking, a vehicle initially captured by the rear camera later appears in the front camera. To address this, we propose key-frame conditions, which are the starting frames of all views, to establish correlations within the image latent space. Combined with temporal attention, the model learns the relationships across" + } + ] + } + ], + "index": 9 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "type": "text", + "content": "17189" + } + ] + } + ], + "index": 10 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 2 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 82, + 70, + 531, + 297 + ], + "blocks": [ + { + "bbox": [ + 82, + 70, + 531, + 297 + ], + "lines": [ + { + "bbox": [ + 82, + 70, + 531, + 297 + ], + "spans": [ + { + "bbox": [ + 82, + 70, + 531, + 297 + ], + "type": "image", + "image_path": "6bc43f600d0da8e506cfdfe42a49ad5713f793b7e007bf4805b60fbd4652ffc8.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 54, + 305, + 555, + 373 + ], + "lines": [ + { + "bbox": [ + 54, + 305, + 555, + 373 + ], + "spans": [ + { + "bbox": [ + 54, + 305, + 555, + 373 + ], + "type": "text", + "content": "Figure 2. The DriveScape pipeline, built on the LDM framework, generates street-view videos conditioned on scene annotations, BEV maps, and 3D bounding boxes for each view. Keyframe conditions are introduced, in collaboration with temporal and spatial self-attention modules, to ensure consistency across spatial and temporal dimensions. Our unified model enables multi-view video generation without the need for complex post-processing or refinement procedures. Additionally, the Bi-Directional Modulated Transformer (BiMoT) module, consisting of two cross-attention layers with opposite directions and a temporal self-attention layer, ensures effective alignment and integration of various 3D road structural information, enabling precise control over video generation." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "bbox": [ + 54, + 392, + 295, + 417 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 54, + 392, + 295, + 417 + ], + "spans": [ + { + "bbox": [ + 54, + 392, + 295, + 417 + ], + "type": "text", + "content": "all views at the initial frame and the preceding frames of neighboring cameras." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 54, + 418, + 295, + 574 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 54, + 418, + 295, + 574 + ], + "spans": [ + { + "bbox": [ + 54, + 418, + 295, + 574 + ], + "type": "text", + "content": "Sparse multi-modal feature fusion. Our model supports training with videos at various frame rates, ranging from 2Hz to 10Hz, while maintaining sample conditions at 2Hz. A lightweight encoder is used to encode the 2Hz map and bounding boxes (BBoxes), with condition-missing frames padded with zeros. Additionally, an extra mask channel is introduced to indicate the presence of conditions at each frame, where mask values are 0 or 1. These inputs are processed through the BiMoT module. Embeddings for unconditioned frames are learned and integrated into the conditions to align with the video frame rate. This diverse frame-rate training scheme, combined with the BiMoT module, enhances sparse condition control capability." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 54, + 575, + 295, + 647 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 54, + 575, + 295, + 647 + ], + "spans": [ + { + "bbox": [ + 54, + 575, + 295, + 647 + ], + "type": "text", + "content": "Conditions dropout. To enhance robustness, key views are trained in random order. Neighbor camera videos and conditions are dropped with probabilities of " + }, + { + "bbox": [ + 54, + 575, + 295, + 647 + ], + "type": "inline_equation", + "content": "50\\%" + }, + { + "bbox": [ + 54, + 575, + 295, + 647 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 54, + 575, + 295, + 647 + ], + "type": "inline_equation", + "content": "20\\%" + }, + { + "bbox": [ + 54, + 575, + 295, + 647 + ], + "type": "text", + "content": ", respectively. Unlike [16], our model requires no postrefinement, simultaneously learning multi-view and temporal consistency to improve quality and realism." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 55, + 658, + 151, + 672 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 658, + 151, + 672 + ], + "spans": [ + { + "bbox": [ + 55, + 658, + 151, + 672 + ], + "type": "text", + "content": "3.4. Inference Stage" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 55, + 677, + 295, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 677, + 295, + 714 + ], + "spans": [ + { + "bbox": [ + 55, + 677, + 295, + 714 + ], + "type": "text", + "content": "During inference, our pipeline enables efficient and consistent generation of multi-view videos. Similar to the training stage, the process consists of two patterns: key-view infer" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 313, + 392, + 553, + 490 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 392, + 553, + 490 + ], + "spans": [ + { + "bbox": [ + 313, + 392, + 553, + 490 + ], + "type": "text", + "content": "ence and neighbor-view inference, as shown in Fig. 2. Keyview camera videos are generated simultaneously without cross-view condition control, as the key-view cameras are intentionally placed non-adjacently. For example, one camera is positioned at the front, while two are at the rear—one on the left and the other on the right. This strategic arrangement ensures comprehensive coverage from multiple perspectives with minimal overlap." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 313, + 491, + 554, + 561 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 491, + 554, + 561 + ], + "spans": [ + { + "bbox": [ + 313, + 491, + 554, + 561 + ], + "type": "text", + "content": "In this configuration, neighbor-view cameras—two at the front and one at the rear—naturally receive cross-view conditional information from their adjacent key-view cameras. This design allows neighbor-view inference to maintain consistency with conditions from neighboring cameras, ensuring cross-view coherence." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 313, + 563, + 554, + 624 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 563, + 554, + 624 + ], + "spans": [ + { + "bbox": [ + 313, + 563, + 554, + 624 + ], + "type": "text", + "content": "Leveraging this setup, our inference pipeline excels in parallel processing, requiring only two consecutive iterations to generate multi-view videos. This approach not only accelerates video production but also ensures cross-view consistency and realism throughout." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 313, + 635, + 519, + 647 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 635, + 519, + 647 + ], + "spans": [ + { + "bbox": [ + 313, + 635, + 519, + 647 + ], + "type": "text", + "content": "3.5. Bi-Directional Modulated Transformer" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 313, + 654, + 553, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 654, + 553, + 713 + ], + "spans": [ + { + "bbox": [ + 313, + 654, + 553, + 713 + ], + "type": "text", + "content": "To integrate multi-conditions from different modalities into the UNet model and ensure effective articulation of each condition, we propose the Bi-Directional Modulated Transformer (BiMoT) module. This module aligns features from various modalities within the latent space, reinforces tem" + } + ] + } + ], + "index": 11 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "text", + "content": "17190" + } + ] + } + ], + "index": 12 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 3 + }, + { + "para_blocks": [ + { + "bbox": [ + 55, + 72, + 294, + 120 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 72, + 294, + 120 + ], + "spans": [ + { + "bbox": [ + 55, + 72, + 294, + 120 + ], + "type": "text", + "content": "poral coherence among the merged conditions, and injects the aligned conditions into the network. As illustrated in Fig. 1, the module consists of two cross-attention layers with opposite directions and a temporal self-attention layer." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 55, + 121, + 294, + 204 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 121, + 294, + 204 + ], + "spans": [ + { + "bbox": [ + 55, + 121, + 294, + 204 + ], + "type": "text", + "content": "Latent to Condition Attention To address diverse conditions with different modalities, we employ a latent-to-condition cross-attention mechanism to standardize feature representations. Pixel-level features are used as signals to merge various condition features. Specifically, let " + }, + { + "bbox": [ + 55, + 121, + 294, + 204 + ], + "type": "inline_equation", + "content": "F_{cond}" + }, + { + "bbox": [ + 55, + 121, + 294, + 204 + ], + "type": "text", + "content": " represent condition features and " + }, + { + "bbox": [ + 55, + 121, + 294, + 204 + ], + "type": "inline_equation", + "content": "F_{latent}" + }, + { + "bbox": [ + 55, + 121, + 294, + 204 + ], + "type": "text", + "content": " represent video latent features. The output is computed as:" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 104, + 213, + 295, + 230 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 104, + 213, + 295, + 230 + ], + "spans": [ + { + "bbox": [ + 104, + 213, + 295, + 230 + ], + "type": "interline_equation", + "content": "F _ {c o n d} ^ {\\prime} = \\operatorname {c r o s s a t t n} \\left(F _ {c o n d}, F _ {l a t e n t}\\right), \\tag {1}", + "image_path": "3e00fc6202f8f982223ea85deddc9bbc93e91eef097d64ed1e95dbfb62e8eced.jpg" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 55, + 239, + 294, + 286 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 239, + 294, + 286 + ], + "spans": [ + { + "bbox": [ + 55, + 239, + 294, + 286 + ], + "type": "text", + "content": "where the query is " + }, + { + "bbox": [ + 55, + 239, + 294, + 286 + ], + "type": "inline_equation", + "content": "F_{cond}" + }, + { + "bbox": [ + 55, + 239, + 294, + 286 + ], + "type": "text", + "content": ", and the key and value are " + }, + { + "bbox": [ + 55, + 239, + 294, + 286 + ], + "type": "inline_equation", + "content": "F_{latent}" + }, + { + "bbox": [ + 55, + 239, + 294, + 286 + ], + "type": "text", + "content": ". This cross-attention operation facilitates the flow of video latent information into the condition space, providing a unified representation for integrating all modalities." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 55, + 287, + 294, + 394 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 287, + 294, + 394 + ], + "spans": [ + { + "bbox": [ + 55, + 287, + 294, + 394 + ], + "type": "text", + "content": "Temporal Attention. We apply temporal self-attention to ensure that the merged features propagate effectively along the temporal dimension. Sparse condition information is also incorporated into the image latent space, naturally unifying all modalities into a single representation. This enables temporal information from latent features to interact with sparse conditions over time, guiding the generation process with sparse control. This mechanism ensures consistency and precise control over the temporal dimensions." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 55, + 396, + 294, + 431 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 396, + 294, + 431 + ], + "spans": [ + { + "bbox": [ + 55, + 396, + 294, + 431 + ], + "type": "text", + "content": "Condition to Video Attention. Similar to the latent-to-condition cross-attention, a condition-to-video cross-attention mechanism is applied:" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 102, + 441, + 295, + 456 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 102, + 441, + 295, + 456 + ], + "spans": [ + { + "bbox": [ + 102, + 441, + 295, + 456 + ], + "type": "interline_equation", + "content": "F _ {l a t e n t} ^ {\\prime} = \\operatorname {c r o s s a t t n} \\left(F _ {l a t e n t}, F _ {c o n d}\\right). \\tag {2}", + "image_path": "17d5282aa255a5dccd06facecc36f928e75be6947d852b27e7eaee9348ee675a.jpg" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 55, + 465, + 294, + 501 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 465, + 294, + 501 + ], + "spans": [ + { + "bbox": [ + 55, + 465, + 294, + 501 + ], + "type": "text", + "content": "This operation facilitates the flow of merged condition feature information into the image latents, ultimately enabling unified control over video generation." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 55, + 502, + 295, + 632 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 502, + 295, + 632 + ], + "spans": [ + { + "bbox": [ + 55, + 502, + 295, + 632 + ], + "type": "text", + "content": "Coarse to Fine Condition Mixture. BiMoT is flexible in accommodating various types of conditions, enabling easy extension to additional inputs. We begin with maps and 3D layouts, then introduce BEV key-frames, and finally integrate neighbor camera videos. This progression allows generation control to transition from a coarse level, driven by maps and 3D layouts, to a finer level, guided by pixel-level details. This ensures both structural coherence and detailed fidelity. The flexibility of BiMoT enables fine-grained layout controllability with unified mono-condition and multi-condition control." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 55, + 645, + 135, + 658 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 645, + 135, + 658 + ], + "spans": [ + { + "bbox": [ + 55, + 645, + 135, + 658 + ], + "type": "text", + "content": "4. Experiments" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 55, + 665, + 294, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 665, + 294, + 713 + ], + "spans": [ + { + "bbox": [ + 55, + 665, + 294, + 713 + ], + "type": "text", + "content": "In this section, we evaluate our model through key metrics, outlining the experimental setup, dataset, and evaluation criteria. We then analyze qualitative results and conduct ablation studies to validate our approach." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 314, + 72, + 427, + 84 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 314, + 72, + 427, + 84 + ], + "spans": [ + { + "bbox": [ + 314, + 72, + 427, + 84 + ], + "type": "text", + "content": "4.1. Experiment Details" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 313, + 89, + 553, + 160 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 89, + 553, + 160 + ], + "spans": [ + { + "bbox": [ + 313, + 89, + 553, + 160 + ], + "type": "text", + "content": "Dataset. The training data is sourced from the real-world driving dataset nuScenes [6], a widely used benchmark for BEV segmentation and 3D object detection in driving scenarios. Following the official configuration, we use 700 street-view scenes for training and 150 for validation, targeting 8 object classes and 8 road classes." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 313, + 161, + 553, + 353 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 161, + 553, + 353 + ], + "spans": [ + { + "bbox": [ + 313, + 161, + 553, + 353 + ], + "type": "text", + "content": "Evaluation Metrics. To evaluate the effectiveness of our approach in delivering realism, continuity, and precise control over multiple conditions, we employ four key metrics for comparison against existing multi-view image and video generation methods. For realism, we use the Fréchet Inception Distance (FID) [11] to measure image quality. Temporal coherence is assessed using the Fréchet Video Distance (FVD) [28], ensuring consistent and fluid motion in the generated videos. Controllability is evaluated through two perception-based methods: CVT [39] for BEV segmentation and BEVFusion [19] for 3D object detection. Both methods are trained on the nuScenes dataset, providing a reliable benchmark for performance evaluation. Additionally, since our model generates videos, we follow Panacea [33] and use the video-based perception model StreamPETR [29] to evaluate the 3D object detection task." + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 314, + 361, + 433, + 373 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 314, + 361, + 433, + 373 + ], + "spans": [ + { + "bbox": [ + 314, + 361, + 433, + 373 + ], + "type": "text", + "content": "4.2. Experiments Results" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 313, + 378, + 553, + 486 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 378, + 553, + 486 + ], + "spans": [ + { + "bbox": [ + 313, + 378, + 553, + 486 + ], + "type": "text", + "content": "Qualitative Results. We showcase the quality and consistency of our video generation results in Fig. 3. Our model achieves notable consistency across both frames and viewpoints. Specifically, vehicles are accurately positioned according to the provided 3D layouts, and street delineations closely match the supplied projected maps. Remarkably, even when bounding boxes extend beyond the visible area, our model adheres to the specified conditions and generates the remaining objects with high fidelity." + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 313, + 486, + 553, + 666 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 486, + 553, + 666 + ], + "spans": [ + { + "bbox": [ + 313, + 486, + 553, + 666 + ], + "type": "text", + "content": "Sparse Condition Control As shown in Fig. 5, our model generates consistent videos in both time and space under sparse conditions, achieving a resolution of " + }, + { + "bbox": [ + 313, + 486, + 553, + 666 + ], + "type": "inline_equation", + "content": "576 \\times 1024" + }, + { + "bbox": [ + 313, + 486, + 553, + 666 + ], + "type": "text", + "content": " and a frame rate of 10 fps. To compare with related works on sparse control [9], we implemented a sparse control model with multi-condition fusion, as existing methods lack support for multi-condition control. A visualization comparing the reproduced SparseCtrl [9] and our method is presented in Fig. 4. Our method clearly demonstrates superior spatial consistency. Moreover, the FID and FVD scores for the reproduced SparseCtrl are 11.15 and 116.00, respectively, significantly worse than those achieved by our model (8.34 and 76.39). For additional visual comparisons and details on the multi-condition fusion SparseCtrl, please refer to the supplementary materials and the accompanying web page." + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 313, + 666, + 553, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 666, + 553, + 713 + ], + "spans": [ + { + "bbox": [ + 313, + 666, + 553, + 713 + ], + "type": "text", + "content": "Quantitative Results. We report quantitative metrics on the nuScenes validation set in Table 2. Using FID and FVD, we evaluate image and video quality, respectively. Following MagicDrive [7] and DrivingDiffusion [16], we" + } + ] + } + ], + "index": 17 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "text", + "content": "17191" + } + ] + } + ], + "index": 18 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 4 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 110, + 69, + 504, + 206 + ], + "blocks": [ + { + "bbox": [ + 110, + 69, + 504, + 206 + ], + "lines": [ + { + "bbox": [ + 110, + 69, + 504, + 206 + ], + "spans": [ + { + "bbox": [ + 110, + 69, + 504, + 206 + ], + "type": "table", + "html": "
MethodFID↓FVD↓Road mIoU ↑Vehicle mIoU ↑Object NDS ↑
Oracle--73.6834.8441.21
BEVGen[25]25.54-50.20 (-23.48)5.89 (-28.95)-
BEVControl [35]24.85-60.80 (-12.88)26.80 (-8.04)-
MagicDrive [7]16.20-61.05 (-12.63)27.01 (-7.83)23.32 (-17.89)
DriveDreamer [31]52.60452.0---
DrivingDiffusion [16]15.83332.063.20 (-10.48)31.60 (-3.24)33.1 (-8.11)
Drive-WM [32]15.80122.7065.07 (-8.61)27.19 (-7.65)-
Delphi [21]15.08113.5--36.58 (-4.63)
SparseCtrl[9]11.15116.00---
Ours8.3476.3968.21 (-5.47)30.27 (-4.57)36.5 (-4.71)
", + "image_path": "f7b62ef89dfcd47602eae26a73d1501b9a138f0d8cdf26b3ae3047913cd962a3.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "table_body" + } + ], + "index": 0 + }, + { + "bbox": [ + 55, + 214, + 555, + 237 + ], + "lines": [ + { + "bbox": [ + 55, + 214, + 555, + 237 + ], + "spans": [ + { + "bbox": [ + 55, + 214, + 555, + 237 + ], + "type": "text", + "content": "Table 2. The comparative analysis of our method against others on the nuScenes validation dataset is detailed for each task. Our end-to-end approach outperformed most baseline models without requiring any post-refinement process." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "text" + }, + { + "type": "image", + "bbox": [ + 72, + 247, + 539, + 479 + ], + "blocks": [ + { + "bbox": [ + 72, + 247, + 539, + 479 + ], + "lines": [ + { + "bbox": [ + 72, + 247, + 539, + 479 + ], + "spans": [ + { + "bbox": [ + 72, + 247, + 539, + 479 + ], + "type": "image", + "image_path": "1e538604e8c82383bd5b8a362be456dd51e85689478209710856b9e55e99afd1.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 55, + 491, + 555, + 526 + ], + "lines": [ + { + "bbox": [ + 55, + 491, + 555, + 526 + ], + "spans": [ + { + "bbox": [ + 55, + 491, + 555, + 526 + ], + "type": "text", + "content": "Figure 3. Showcases for multi-view video generation by DriveScape demonstrate frame-to-frame and viewpoint consistency. Vehicles generated by our model are accurately positioned according to the provided 3D layouts, while the street depictions closely align with the projected maps." + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_caption" + } + ], + "index": 2 + }, + { + "type": "image", + "bbox": [ + 56, + 544, + 298, + 636 + ], + "blocks": [ + { + "bbox": [ + 56, + 544, + 298, + 636 + ], + "lines": [ + { + "bbox": [ + 56, + 544, + 298, + 636 + ], + "spans": [ + { + "bbox": [ + 56, + 544, + 298, + 636 + ], + "type": "image", + "image_path": "6f481d41397efc002e0ab76879aaeaff7071e2d6825cfcd60703a6388f91a836.jpg" + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 55, + 643, + 296, + 677 + ], + "lines": [ + { + "bbox": [ + 55, + 643, + 296, + 677 + ], + "spans": [ + { + "bbox": [ + 55, + 643, + 296, + 677 + ], + "type": "text", + "content": "Figure 4. Results in each row are Ours, SparseCtrl, input condition and GT respectively, which illustrate that our method has better spatial consistency than SparseCtrl." + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "image_caption" + } + ], + "index": 4 + }, + { + "bbox": [ + 55, + 689, + 296, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 689, + 296, + 715 + ], + "spans": [ + { + "bbox": [ + 55, + 689, + 296, + 715 + ], + "type": "text", + "content": "generate data from validation set labels and assess it with a pre-trained perception model. The decline in Road mIoU" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 313, + 546, + 555, + 581 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 546, + 555, + 581 + ], + "spans": [ + { + "bbox": [ + 313, + 546, + 555, + 581 + ], + "type": "text", + "content": "and Vehicle mIoU, compared to original data, measures the effectiveness of our framework in producing perception-friendly synthetic data." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 313, + 582, + 556, + 689 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 582, + 556, + 689 + ], + "spans": [ + { + "bbox": [ + 313, + 582, + 556, + 689 + ], + "type": "text", + "content": "To align with previous studies, we resized generated images to " + }, + { + "bbox": [ + 313, + 582, + 556, + 689 + ], + "type": "inline_equation", + "content": "224 \\times 400" + }, + { + "bbox": [ + 313, + 582, + 556, + 689 + ], + "type": "text", + "content": ". We compared our FID results with three image generation models: BEVGen [25], BEVControl [35], and MagicDrive [7]. Our method significantly outperformed these models. Notably, video generation models typically yield higher FID scores than image models. For instance, Drive-WM, using only an image model, achieved better FID than the full model, highlighting the challenges of video generation." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 313, + 689, + 555, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 689, + 555, + 715 + ], + "spans": [ + { + "bbox": [ + 313, + 689, + 555, + 715 + ], + "type": "text", + "content": "For video quality, assessed via the FVD metric, our approach showed substantial improvements over recent works" + } + ] + } + ], + "index": 9 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 318, + 758 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 318, + 758 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 318, + 758 + ], + "type": "text", + "content": "17192" + } + ] + } + ], + "index": 10 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 5 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 60, + 71, + 293, + 251 + ], + "blocks": [ + { + "bbox": [ + 60, + 71, + 293, + 251 + ], + "lines": [ + { + "bbox": [ + 60, + 71, + 293, + 251 + ], + "spans": [ + { + "bbox": [ + 60, + 71, + 293, + 251 + ], + "type": "image", + "image_path": "fc1328671988eaa6e2b3787d4b8518b4976c53b1944803ee87c2d7fe01020ea6.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 85, + 258, + 265, + 270 + ], + "lines": [ + { + "bbox": [ + 85, + 258, + 265, + 270 + ], + "spans": [ + { + "bbox": [ + 85, + 258, + 265, + 270 + ], + "type": "text", + "content": "Figure 5. Showcase for sparse Condition Control." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "bbox": [ + 55, + 283, + 296, + 428 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 283, + 296, + 428 + ], + "spans": [ + { + "bbox": [ + 55, + 283, + 296, + 428 + ], + "type": "text", + "content": "such as DriveDreamer [31], DrivingDiffusion [16], and Drive-WM [32]. In BEV segmentation metrics, our method achieved superior perceptual performance, except for a slightly lower Road mIoU compared to DrivingDiffusion, which utilized a post-processing network for refinement. To maintain simplicity, our end-to-end system applies no post-processing beyond image resizing for perceptual evaluation. The perception models we used operate on single-image inference, not leveraging the temporal consistency of our generated videos. Given our significant lead in FVD, our method holds strong potential for training video BEV perception models." + } + ] + } + ], + "index": 2 + }, + { + "type": "table", + "bbox": [ + 85, + 437, + 266, + 502 + ], + "blocks": [ + { + "bbox": [ + 85, + 437, + 266, + 502 + ], + "lines": [ + { + "bbox": [ + 85, + 437, + 266, + 502 + ], + "spans": [ + { + "bbox": [ + 85, + 437, + 266, + 502 + ], + "type": "table", + "html": "
dataNDS ↑FID ↓FVD ↓
Oracle46.9--
Panacea32.1 (-14.8)16.96139
Oracle53.7--
Ours44.5 (-9.2)8.3476.39
", + "image_path": "c6b4307be11c23f07a633bfc78fa69b423663bd96d9411df9aeb7da1f18e1d5a.jpg" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "table_body" + } + ], + "index": 3 + }, + { + "bbox": [ + 55, + 540, + 156, + 552 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 540, + 156, + 552 + ], + "spans": [ + { + "bbox": [ + 55, + 540, + 156, + 552 + ], + "type": "text", + "content": "4.3. Ablation Studies" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 55, + 558, + 296, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 558, + 296, + 713 + ], + "spans": [ + { + "bbox": [ + 55, + 558, + 296, + 713 + ], + "type": "text", + "content": "To refine and evaluate the efficacy of our proposed framework, we conducted ablation studies, as detailed in Table 4. Various strategies for implementing conditional guidance were tested. Removing the entire BiMoT module resulted in the FID increasing from 8.34 to 20.98 and the FVD rising from 76.39 to 142.81, underscoring the module's importance for multi-condition encoding. Similarly, removing the temporal attention layer within BiMoT caused a significant performance drop, with FID and FVD increasing to 20.67 and 123.03, respectively, highlighting the role of temporal modelling in BiMoT. Additionally, during testing, we separately removed the key-frame condition and the neighbor condition. Both led to a performance decline, indicating the" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 313, + 72, + 555, + 96 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 72, + 555, + 96 + ], + "spans": [ + { + "bbox": [ + 313, + 72, + 555, + 96 + ], + "type": "text", + "content": "importance of these conditions for facilitating information exchange between views." + } + ] + } + ], + "index": 7 + }, + { + "type": "table", + "bbox": [ + 331, + 106, + 539, + 182 + ], + "blocks": [ + { + "bbox": [ + 55, + 510, + 295, + 532 + ], + "lines": [ + { + "bbox": [ + 55, + 510, + 295, + 532 + ], + "spans": [ + { + "bbox": [ + 55, + 510, + 295, + 532 + ], + "type": "text", + "content": "Table 3. Compared with Panacea, we evaluate results on the nuScenes validation set using the pre-trained StreamPETR model." + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 331, + 106, + 539, + 182 + ], + "lines": [ + { + "bbox": [ + 331, + 106, + 539, + 182 + ], + "spans": [ + { + "bbox": [ + 331, + 106, + 539, + 182 + ], + "type": "table", + "html": "
TeAttnBiMOTKFNeFIDFVD
✓✓✓20.98142.81
✓✓20.67123.03
✓✓✓11.60108.80
✓✓✓12.18111.36
✓✓✓✓8.3476.39
", + "image_path": "d2e976422d91156f1cb968661638bbaf32f000f6f21cf73380cd2a4732b6852e.jpg" + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "table_body" + } + ], + "index": 8 + }, + { + "bbox": [ + 313, + 190, + 555, + 213 + ], + "lines": [ + { + "bbox": [ + 313, + 190, + 555, + 213 + ], + "spans": [ + { + "bbox": [ + 313, + 190, + 555, + 213 + ], + "type": "text", + "content": "Table 4. Ablation Study. TeAttn is temporal attention module in BiMOT, KF is key-frame condition and Ne is neighbor condition" + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "text" + }, + { + "bbox": [ + 313, + 233, + 495, + 247 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 233, + 495, + 247 + ], + "spans": [ + { + "bbox": [ + 313, + 233, + 495, + 247 + ], + "type": "text", + "content": "4.4. Visualization of Different Settings" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 313, + 251, + 553, + 322 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 251, + 553, + 322 + ], + "spans": [ + { + "bbox": [ + 313, + 251, + 553, + 322 + ], + "type": "text", + "content": "In our visual showcases, we demonstrate the model's proficiency in manipulating video predictions for a single scene under various control conditions. The control conditions, sourced from frames of identical perspectives but randomly selected from other scenes, test the decoupling of conditions and the robustness of our method across diverse scenarios." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 313, + 323, + 554, + 430 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 323, + 554, + 430 + ], + "spans": [ + { + "bbox": [ + 313, + 323, + 554, + 430 + ], + "type": "text", + "content": "As shown in Fig. 7, the experiments explore the influence of different 3D layout controls. Using a single reference frame from a highway segment, we first depict vehicles driving normally, then transition to vehicles traversing grassy terrain. Remarkably, our model maintains temporal consistency even in this unconventional context. Additional experiments involve removing or adding large vehicles, with results aligning closely with the specified conditions, showcasing precise control." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 313, + 431, + 555, + 526 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 431, + 555, + 526 + ], + "spans": [ + { + "bbox": [ + 313, + 431, + 555, + 526 + ], + "type": "text", + "content": "Furthermore, as illustrated in Fig. 8, we investigate the impact of different maps on video generation. The generated scenes align impressively with the shapes of the provided map regions, maintaining layout feature consistency despite map changes. These experiments provide strong evidence of our model's capability to render specified control conditions into precise and coherent visual narratives within the generated video content." + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 313, + 538, + 391, + 550 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 538, + 391, + 550 + ], + "spans": [ + { + "bbox": [ + 313, + 538, + 391, + 550 + ], + "type": "text", + "content": "5. Conclusions" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 313, + 558, + 555, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 558, + 555, + 713 + ], + "spans": [ + { + "bbox": [ + 313, + 558, + 555, + 713 + ], + "type": "text", + "content": "The paper introduces DriveScape, the first end-to-end multi-view driving video generative model. It enables consistent multi-view video generation without complex postprocessing or refinement. This unified framework streamlines multi-view condition-guided video generation. We also propose a Bi-Directional Modulated Transformer module for diverse condition guidance. Our approach achieves decoupled and precise control of dynamic foreground and static background, delivering state-of-the-art video synthesis on the nuScenes dataset. It excels in multi-view image and video generation quality and perception applications. In summary, our method outperforms others in generation quality and ease of application." + } + ] + } + ], + "index": 15 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "text", + "content": "17193" + } + ] + } + ], + "index": 16 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 6 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 72, + 82, + 541, + 228 + ], + "blocks": [ + { + "bbox": [ + 72, + 82, + 541, + 228 + ], + "lines": [ + { + "bbox": [ + 72, + 82, + 541, + 228 + ], + "spans": [ + { + "bbox": [ + 72, + 82, + 541, + 228 + ], + "type": "image", + "image_path": "b565b73afdef9a50ea2277898823f855298278724094fe607063f07c0966b9f0.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 55, + 239, + 555, + 262 + ], + "lines": [ + { + "bbox": [ + 55, + 239, + 555, + 262 + ], + "spans": [ + { + "bbox": [ + 55, + 239, + 555, + 262 + ], + "type": "text", + "content": "Figure 6. Showcase for maps and 3D layouts control under different conditions. (a): control by the reference maps and 3D Layouts; (b) change 3D Layouts with more vehicles; (c) change the maps with curved pedestrian crosswalk; (d) remove all the 3D Layouts." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "type": "image", + "bbox": [ + 70, + 285, + 539, + 443 + ], + "blocks": [ + { + "bbox": [ + 70, + 285, + 539, + 443 + ], + "lines": [ + { + "bbox": [ + 70, + 285, + 539, + 443 + ], + "spans": [ + { + "bbox": [ + 70, + 285, + 539, + 443 + ], + "type": "image", + "image_path": "8c640171de84fd850e911dc18e017793ae07b0d09d31c897555c874686e42fc4.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 55, + 452, + 555, + 473 + ], + "lines": [ + { + "bbox": [ + 55, + 452, + 555, + 473 + ], + "spans": [ + { + "bbox": [ + 55, + 452, + 555, + 473 + ], + "type": "text", + "content": "Figure 7. Showcases for 3D layouts control under different conditions. DriveScape can precisely control dynamic objects. (a) reference layouts; (b) move car to the grass; (c) remove all cars; (d) add trucks to the scene." + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_caption" + } + ], + "index": 2 + }, + { + "type": "image", + "bbox": [ + 70, + 508, + 541, + 666 + ], + "blocks": [ + { + "bbox": [ + 70, + 508, + 541, + 666 + ], + "lines": [ + { + "bbox": [ + 70, + 508, + 541, + 666 + ], + "spans": [ + { + "bbox": [ + 70, + 508, + 541, + 666 + ], + "type": "image", + "image_path": "0812ffb9511188af0cf2e05b163d70fb142cf525153a32c4b5a3203d8538d3ba.jpg" + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 55, + 674, + 555, + 696 + ], + "lines": [ + { + "bbox": [ + 55, + 674, + 555, + 696 + ], + "spans": [ + { + "bbox": [ + 55, + 674, + 555, + 696 + ], + "type": "text", + "content": "Figure 8. Showcases for maps control under. DriveScape can precisely control static background. (a) replace the friable area with green belts; (b) replace the friable area with green belts; (c) replace the friable area with sidewalks; (d) add island on the road." + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "image_caption" + } + ], + "index": 4 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "type": "text", + "content": "17194" + } + ] + } + ], + "index": 6 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 7 + }, + { + "para_blocks": [ + { + "bbox": [ + 56, + 72, + 115, + 83 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 72, + 115, + 83 + ], + "spans": [ + { + "bbox": [ + 56, + 72, + 115, + 83 + ], + "type": "text", + "content": "References" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 57, + 91, + 296, + 713 + ], + "type": "list", + "angle": 0, + "index": 13, + "blocks": [ + { + "bbox": [ + 61, + 91, + 296, + 134 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 61, + 91, + 296, + 134 + ], + "spans": [ + { + "bbox": [ + 61, + 91, + 296, + 134 + ], + "type": "text", + "content": "[1] Jie An, Songyang Zhang, Harry Yang, Sonal Gupta, Jia-Bin Huang, Jiebo Luo, and Xi Yin. Latent-shift: Latent diffusion with temporal shift for efficient text-to-video generation. arXiv preprint arXiv:2304.08477, 2023. 2" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 61, + 135, + 296, + 190 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 61, + 135, + 296, + 190 + ], + "spans": [ + { + "bbox": [ + 61, + 135, + 296, + 190 + ], + "type": "text", + "content": "[2] Andreas Blattmann, Tim Dockhorn, Sumith Kulal, Daniel Mendelevitch, Maciej Kilian, Dominik Lorenz, Yam Levi, Zion English, Vikram Voleti, Adam Letts, Varun Jampani, and Robin Rombach. Stable video diffusion: Scaling latent video diffusion models to large datasets, 2023. 2" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 62, + 191, + 295, + 246 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 191, + 295, + 246 + ], + "spans": [ + { + "bbox": [ + 62, + 191, + 295, + 246 + ], + "type": "text", + "content": "[3] Andreas Blattmann, Tim Dockhorn, Sumith Kulal, Daniel Mendelevitch, Maciej Kilian, Dominik Lorenz, Yam Levi, Zion English, Vikram Voleti, Adam Letts, et al. Stable video diffusion: Scaling latent video diffusion models to large datasets. arXiv preprint arXiv:2311.15127, 2023. 2" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 62, + 247, + 295, + 312 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 247, + 295, + 312 + ], + "spans": [ + { + "bbox": [ + 62, + 247, + 295, + 312 + ], + "type": "text", + "content": "[4] Andreas Blattmann, Robin Rombach, Huan Ling, Tim Dockhorn, Seung Wook Kim, Sanja Fidler, and Karsten Kreis. Align your latents: High-resolution video synthesis with latent diffusion models. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 22563-22575, 2023. 2" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 62, + 313, + 296, + 378 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 313, + 296, + 378 + ], + "spans": [ + { + "bbox": [ + 62, + 313, + 296, + 378 + ], + "type": "text", + "content": "[5] Andreas Blattmann, Robin Rombach, Huan Ling, Tim Dockhorn, Seung Wook Kim, Sanja Fidler, and Karsten Kreis. Align your latents: High-resolution video synthesis with latent diffusion models. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 22563-22575, 2023. 2" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 62, + 380, + 296, + 445 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 380, + 296, + 445 + ], + "spans": [ + { + "bbox": [ + 62, + 380, + 296, + 445 + ], + "type": "text", + "content": "[6] Holger Caesar, Varun Bankiti, Alex H Lang, Sourabh Vora, Venice Erin Liong, Qiang Xu, Anush Krishnan, Yu Pan, Giancarlo Baldan, and Oscar Beijbom. nuscenes: A multimodal dataset for autonomous driving. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 11621-11631, 2020. 5" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 62, + 447, + 295, + 490 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 447, + 295, + 490 + ], + "spans": [ + { + "bbox": [ + 62, + 447, + 295, + 490 + ], + "type": "text", + "content": "[7] Ruiyuan Gao, Kai Chen, Enze Xie, Lanqing Hong, Zhenguo Li, Dit-Yan Yeung, and Qiang Xu. Magicdrive: Street view generation with diverse 3d geometry control. arXiv preprint arXiv:2310.02601, 2023. 1, 2, 5, 6" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 62, + 491, + 295, + 534 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 491, + 295, + 534 + ], + "spans": [ + { + "bbox": [ + 62, + 491, + 295, + 534 + ], + "type": "text", + "content": "[8] Yuwei Guo, Ceyuan Yang, Anyi Rao, Yaohui Wang, Yu Qiao, Dahua Lin, and Bo Dai. Animatediff:Animate your personalized text-to-image diffusion models without specific tuning. arXiv preprint arXiv:2307.04725, 2023. 2" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 62, + 536, + 295, + 579 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 536, + 295, + 579 + ], + "spans": [ + { + "bbox": [ + 62, + 536, + 295, + 579 + ], + "type": "text", + "content": "[9] Yuwei Guo, Ceyuan Yang, Anyi Rao, Maneesh Agrawala, Dahua Lin, and Bo Dai. Sparsectrl: Adding sparse controls to text-to-video diffusion models. In European Conference on Computer Vision, pages 330-348. Springer, 2025. 5, 6" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 57, + 581, + 295, + 623 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 581, + 295, + 623 + ], + "spans": [ + { + "bbox": [ + 57, + 581, + 295, + 623 + ], + "type": "text", + "content": "[10] Yingqing He, Tianyu Yang, Yong Zhang, Ying Shan, and Qifeng Chen. Latent video diffusion models for high-fidelity video generation with arbitrary lengths. arXiv preprint arXiv:2211.13221, 2022. 2" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 57, + 624, + 295, + 678 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 624, + 295, + 678 + ], + "spans": [ + { + "bbox": [ + 57, + 624, + 295, + 678 + ], + "type": "text", + "content": "[11] Martin Heusel, Hubert Ramsauer, Thomas Unterthiner, Bernhard Nessler, and Sepp Hochreiter. Gans trained by a two time-scale update rule converge to a local nash equilibrium. Advances in neural information processing systems, 30, 2017. 5" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 57, + 680, + 295, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 680, + 295, + 713 + ], + "spans": [ + { + "bbox": [ + 57, + 680, + 295, + 713 + ], + "type": "text", + "content": "[12] Jonathan Ho, William Chan, Chitwan Sahara, Jay Whang, Ruiqi Gao, Alexey Gritsenko, Diederik P Kingma, Ben Poole, Mohammad Norouzi, David J Fleet, et al. Imagen" + } + ] + } + ], + "index": 12 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 316, + 73, + 555, + 712 + ], + "type": "list", + "angle": 0, + "index": 27, + "blocks": [ + { + "bbox": [ + 335, + 73, + 553, + 95 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 335, + 73, + 553, + 95 + ], + "spans": [ + { + "bbox": [ + 335, + 73, + 553, + 95 + ], + "type": "text", + "content": "video: High definition video generation with diffusion models. arXiv preprint arXiv:2210.02303, 2022. 2" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 316, + 96, + 553, + 139 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 96, + 553, + 139 + ], + "spans": [ + { + "bbox": [ + 316, + 96, + 553, + 139 + ], + "type": "text", + "content": "[13] Yihan Hu, Jiazhi Yang, Li Chen, Keyu Li, Chonghao Sima, Xizhou Zhu, Siqi Chai, Senyao Du, Tianwei Lin, Wenhai Wang, et al. Goal-oriented autonomous driving. arXiv preprint arXiv:2212.10156, 2022. 1" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 317, + 140, + 553, + 171 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 140, + 553, + 171 + ], + "spans": [ + { + "bbox": [ + 317, + 140, + 553, + 171 + ], + "type": "text", + "content": "[14] Junjie Huang and Guan Huang. Bevdet4d: Exploit temporal cues in multi-camera 3d object detection. arXiv preprint arXiv:2203.17054, 2022. 1" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 317, + 173, + 553, + 215 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 173, + 553, + 215 + ], + "spans": [ + { + "bbox": [ + 317, + 173, + 553, + 215 + ], + "type": "text", + "content": "[15] Junjie Huang, Guan Huang, Zheng Zhu, Yun Ye, and Dalong Du. Bevdet: High-performance multi-camera 3d object detection in bird-eye-view. arXiv preprint arXiv:2112.11790, 2021. 1" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 316, + 217, + 553, + 259 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 217, + 553, + 259 + ], + "spans": [ + { + "bbox": [ + 316, + 217, + 553, + 259 + ], + "type": "text", + "content": "[16] Xiaofan Li, Yifu Zhang, and Xiaoqing Ye. Drivingdiffusion: Layout-guided multi-view driving scene video generation with latent diffusion model. arXiv preprint arXiv:2310.07771, 2023. 2, 3, 4, 5, 6" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 316, + 261, + 553, + 316 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 261, + 553, + 316 + ], + "spans": [ + { + "bbox": [ + 316, + 261, + 553, + 316 + ], + "type": "text", + "content": "[17] Yinhao Li, Zheng Ge, Guanyi Yu, Jinrong Yang, Zengran Wang, Yukang Shi, Jianjian Sun, and Zeming Li. Bevdepth: Acquisition of reliable depth for multi-view 3d object detection. In Proceedings of the AAAI Conference on Artificial Intelligence, pages 1477-1485, 2023. 1" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 317, + 317, + 553, + 371 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 317, + 553, + 371 + ], + "spans": [ + { + "bbox": [ + 317, + 317, + 553, + 371 + ], + "type": "text", + "content": "[18] Zhiqi Li, Wenhai Wang, Hongyang Li, Enze Xie, Chonghao Sima, Tong Lu, Yu Qiao, and Jifeng Dai. Bevformer: Learning bird's-eye-view representation from multi-camera images via spatiotemporal transformers. In European conference on computer vision, pages 1-18. Springer, 2022. 1" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 317, + 372, + 553, + 425 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 372, + 553, + 425 + ], + "spans": [ + { + "bbox": [ + 317, + 372, + 553, + 425 + ], + "type": "text", + "content": "[19] Tingting Liang, Hongwei Xie, Kaicheng Yu, Zhongyu Xia, Zhiwei Lin, Yongtao Wang, Tao Tang, Bing Wang, and Zhi Tang. Bevfusion: A simple and robust lidar-camera fusion framework. Advances in Neural Information Processing Systems, 35:10421-10434, 2022. 5" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 317, + 427, + 553, + 480 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 427, + 553, + 480 + ], + "spans": [ + { + "bbox": [ + 317, + 427, + 553, + 480 + ], + "type": "text", + "content": "[20] Haicheng Liao, Yongkang Li, Zhenning Li, Chengyue Wang, Zhiyong Cui, Shengbo Eben Li, and Chengzhong Xu. A cognitive-based trajectory prediction approach for autonomous driving. arXiv preprint arXiv:2402.19251, 2024. 1" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 317, + 482, + 553, + 536 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 482, + 553, + 536 + ], + "spans": [ + { + "bbox": [ + 317, + 482, + 553, + 536 + ], + "type": "text", + "content": "[21] Enhui Ma, Lijun Zhou, Tao Tang, Zhan Zhang, Dong Han, Junpeng Jiang, Kun Zhan, Peng Jia, Xianpeng Lang, Haiyang Sun, et al. Unleashing generalization of end-to-end autonomous driving with controllable long video generation. arXiv preprint arXiv:2406.01349, 2024. 6" + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 317, + 537, + 555, + 592 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 537, + 555, + 592 + ], + "spans": [ + { + "bbox": [ + 317, + 537, + 555, + 592 + ], + "type": "text", + "content": "[22] Robin Rombach, Andreas Blattmann, Dominik Lorenz, Patrick Esser, and Björn Ommer. High-resolution image synthesis with latent diffusion models. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 10684-10695, 2022. 2" + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 317, + 593, + 555, + 658 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 593, + 555, + 658 + ], + "spans": [ + { + "bbox": [ + 317, + 593, + 555, + 658 + ], + "type": "text", + "content": "[23] Olaf Ronneberger, Philipp Fischer, and Thomas Brox. U-net: Convolutional networks for biomedical image segmentation. In Medical Image Computing and Computer-Assisted Intervention-MICCAI 2015: 18th International Conference, Munich, Germany, October 5-9, 2015, Proceedings, Part III 18, pages 234-241. Springer, 2015. 3" + } + ] + } + ], + "index": 25 + }, + { + "bbox": [ + 317, + 658, + 553, + 712 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 658, + 553, + 712 + ], + "spans": [ + { + "bbox": [ + 317, + 658, + 553, + 712 + ], + "type": "text", + "content": "[24] Uriel Singer, Adam Polyak, Thomas Hayes, Xi Yin, Jie An, Songyang Zhang, Qiyuan Hu, Harry Yang, Oron Ashual, Oran Gafni, Devi Parikh, Sonal Gupta, and Yaniv Taigman. Make-a-video: Text-to-video generation without text-video data, 2022. 2" + } + ] + } + ], + "index": 26 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "text", + "content": "17195" + } + ] + } + ], + "index": 28 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 8 + }, + { + "para_blocks": [ + { + "bbox": [ + 56, + 73, + 294, + 713 + ], + "type": "list", + "angle": 0, + "index": 13, + "blocks": [ + { + "bbox": [ + 56, + 73, + 294, + 106 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 73, + 294, + 106 + ], + "spans": [ + { + "bbox": [ + 56, + 73, + 294, + 106 + ], + "type": "text", + "content": "[25] Alexander Swerdlow, Runsheng Xu, and Bolei Zhou. Street-view image generation from a bird's-eye view layout. IEEE Robotics and Automation Letters, 2024. 1, 2, 6" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 56, + 107, + 294, + 150 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 107, + 294, + 150 + ], + "spans": [ + { + "bbox": [ + 56, + 107, + 294, + 150 + ], + "type": "text", + "content": "[26] Shitao Tang, Fuyang Zhang, Jiacheng Chen, Peng Wang, and Yasutaka Furukawa. Mvdiffusion: Enabling holistic multiview image generation with correspondence-aware diffusion, 2023. 2" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 56, + 152, + 294, + 207 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 152, + 294, + 207 + ], + "spans": [ + { + "bbox": [ + 56, + 152, + 294, + 207 + ], + "type": "text", + "content": "[27] Hung-Yu Tseng, Qinbo Li, Changil Kim, Suhib Alsisan, Jia-Bin Huang, and Johannes Kopf. Consistent view synthesis with pose-guided diffusion models. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 16773-16783, 2023. 2" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 56, + 209, + 294, + 252 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 209, + 294, + 252 + ], + "spans": [ + { + "bbox": [ + 56, + 209, + 294, + 252 + ], + "type": "text", + "content": "[28] Thomas Unterthiner, Sjoerd Van Steenkiste, Karol Kurach, Raphael Marinier, Marcin Michalski, and Sylvain Gelly. Towards accurate generative models of video: A new metric & challenges. arXiv preprint arXiv:1812.01717, 2018. 5" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 56, + 254, + 294, + 297 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 254, + 294, + 297 + ], + "spans": [ + { + "bbox": [ + 56, + 254, + 294, + 297 + ], + "type": "text", + "content": "[29] Shihao Wang, Yingfei Liu, Tiancai Wang, Ying Li, and Xi-angyu Zhang. Exploring object-centric temporal modeling for efficient multi-view 3d object detection. In CVPR, pages 3621–3631, 2023. 5" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 56, + 299, + 294, + 342 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 299, + 294, + 342 + ], + "spans": [ + { + "bbox": [ + 56, + 299, + 294, + 342 + ], + "type": "text", + "content": "[30] Wenjing Wang, Huan Yang, Zixi Tuo, Huiguo He, Junchen Zhu, Jianlong Fu, and Jiaying Liu. Videofactory: Swap attention in spatiotemporal diffusions for text-to-video generation, 2023. 2" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 56, + 344, + 294, + 387 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 344, + 294, + 387 + ], + "spans": [ + { + "bbox": [ + 56, + 344, + 294, + 387 + ], + "type": "text", + "content": "[31] Xiaofeng Wang, Zheng Zhu, Guan Huang, Xinze Chen, and Jiwen Lu. Drivedreamer: Towards real-world-driven world models for autonomous driving. arXiv preprint arXiv:2309.09777, 2023. 6" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 56, + 388, + 294, + 442 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 388, + 294, + 442 + ], + "spans": [ + { + "bbox": [ + 56, + 388, + 294, + 442 + ], + "type": "text", + "content": "[32] Yuqi Wang, Jiawei He, Lue Fan, Hongxin Li, Yuntao Chen, and Zhaoxiang Zhang. Driving into the future: Multiview visual forecasting and planning with world model for autonomous driving. arXiv preprint arXiv:2311.17918, 2023. 6" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 56, + 445, + 294, + 510 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 445, + 294, + 510 + ], + "spans": [ + { + "bbox": [ + 56, + 445, + 294, + 510 + ], + "type": "text", + "content": "[33] Yuqing Wen, Yucheng Zhao, Yingfei Liu, Fan Jia, Yanhui Wang, Chong Luo, Chi Zhang, Tiancai Wang, Xiaoyan Sun, and Xiangyu Zhang. Panacea: Panoramic and controllable video generation for autonomous driving. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 6902-6912, 2024. 5" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 56, + 512, + 294, + 588 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 512, + 294, + 588 + ], + "spans": [ + { + "bbox": [ + 56, + 512, + 294, + 588 + ], + "type": "text", + "content": "[34] Chenyu Yang, Yuntao Chen, Hao Tian, Chenxin Tao, Xizhou Zhu, Zhaoxiang Zhang, Gao Huang, Hongyang Li, Yu Qiao, Lewei Lu, et al. Bevformer v2: Adapting modern image backbones to bird's-eye-view recognition via perspective supervision. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 17830-17839, 2023. 1" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 56, + 590, + 294, + 644 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 590, + 294, + 644 + ], + "spans": [ + { + "bbox": [ + 56, + 590, + 294, + 644 + ], + "type": "text", + "content": "[35] Kairui Yang, Enhui Ma, Jibin Peng, Qing Guo, Di Lin, and Kaicheng Yu. Bevcontrol: Accurately controlling street-view elements with multi-perspective consistency via bev sketch layout. arXiv preprint arXiv:2308.01661, 2023. 1, 2, 6" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 56, + 647, + 294, + 689 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 647, + 294, + 689 + ], + "spans": [ + { + "bbox": [ + 56, + 647, + 294, + 689 + ], + "type": "text", + "content": "[36] Yuxiang Yang, Yingqi Deng, Jing Zhang, Jiahao Nie, and Zheng-Jun Zha. Bevtrack: A simple and strong baseline for 3d single object tracking in bird's-eye view. arXiv e-prints, pages arXiv-2309, 2023. 1" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 56, + 691, + 294, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 691, + 294, + 713 + ], + "spans": [ + { + "bbox": [ + 56, + 691, + 294, + 713 + ], + "type": "text", + "content": "[37] David Junhao Zhang, Dongxu Li, Hung Le, Mike Zheng Shou, Caiming Xiong, and Doyen Sahoo. Moonshot: To" + } + ] + } + ], + "index": 12 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 316, + 73, + 553, + 207 + ], + "type": "list", + "angle": 0, + "index": 18, + "blocks": [ + { + "bbox": [ + 335, + 73, + 553, + 95 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 335, + 73, + 553, + 95 + ], + "spans": [ + { + "bbox": [ + 335, + 73, + 553, + 95 + ], + "type": "text", + "content": "wards controllable video generation and editing with multimodal conditions, 2024. 2" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 316, + 96, + 553, + 129 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 96, + 553, + 129 + ], + "spans": [ + { + "bbox": [ + 316, + 96, + 553, + 129 + ], + "type": "text", + "content": "[38] Wenzhao Zheng, Ruiqi Song, Xianda Guo, and Long Chen. Genad: Generative end-to-end autonomous driving. arXiv preprint arXiv:2402.11502, 2024. 1" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 316, + 130, + 553, + 174 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 130, + 553, + 174 + ], + "spans": [ + { + "bbox": [ + 316, + 130, + 553, + 174 + ], + "type": "text", + "content": "[39] Brady Zhou and Philipp Krahenbuhl. Cross-view transformers for real-time map-view semantic segmentation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 13760-13769, 2022. 1, 5" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 316, + 175, + 553, + 207 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 175, + 553, + 207 + ], + "spans": [ + { + "bbox": [ + 316, + 175, + 553, + 207 + ], + "type": "text", + "content": "[40] Daquan Zhou, Weimin Wang, Hanshu Yan, Weiwei Lv, Yizhe Zhu, and Jiashi Feng. Magicvideo: Efficient video generation with latent diffusion models, 2023. 2" + } + ] + } + ], + "index": 17 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 749, + 318, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 749, + 318, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 749, + 318, + 757 + ], + "type": "text", + "content": "17196" + } + ] + } + ], + "index": 19 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 9 + } + ], + "_backend": "vlm", + "_version_name": "2.6.4" +} \ No newline at end of file diff --git a/2025/Driving by the Rules_ A Benchmark for Integrating Traffic Sign Regulations into Vectorized HD Map/cbe7a2b6-7e9c-4185-bab6-ca66fed01b85_content_list.json b/2025/Driving by the Rules_ A Benchmark for Integrating Traffic Sign Regulations into Vectorized HD Map/cbe7a2b6-7e9c-4185-bab6-ca66fed01b85_content_list.json new file mode 100644 index 0000000000000000000000000000000000000000..eb3ab4a83925061ac1a59facff1b143e4c3e5680 --- /dev/null +++ b/2025/Driving by the Rules_ A Benchmark for Integrating Traffic Sign Regulations into Vectorized HD Map/cbe7a2b6-7e9c-4185-bab6-ca66fed01b85_content_list.json @@ -0,0 +1,1420 @@ +[ + { + "type": "text", + "text": "Driving by the Rules: A Benchmark for Integrating Traffic Sign Regulations into Vectorized HD Map", + "text_level": 1, + "bbox": [ + 91, + 128, + 906, + 176 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Xinyuan Chang $^{1*}$ Maixuan Xue $^{2*†}$ Xinran Liu $^{1}$ Zheng Pan $^{1}$ Xing Wei $^{2‡}$ $^{1}$ Amap, Alibaba Group $^{2}$ Xi’an Jiaotong University \n{changxinyuan.cxy, tom.lxr, panzheng.pan}@alibaba-inc.com \nxmx0809@stu.xjtu.edu.cn, weixing@mail.xjtu.edu.cn", + "bbox": [ + 133, + 202, + 870, + 275 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/fa98e39510a06b77ea1def57287ac98a0b8b5c58fe12f39b937e421ac5f465cb.jpg", + "image_caption": [ + "Figure 1. MapDR Overview and Motivation. The image on the left depicts the comprehensive results of HD map construction for an intersection scene, whereas the image on the right illustrates the outcomes after partitioning it into three layers. Existing online mapping methods primarily emphasize the construction of the geometric and connectivity layers, neglecting the traffic regulation layer. However, precise comprehension of traffic signs and their correlation with lanes is vital for ensuring the safety of autonomous driving." + ], + "image_footnote": [], + "bbox": [ + 140, + 330, + 686, + 542 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/df3966c6489e0b02b577968def196aff7d0c31af7f128e2738124bf701658d61.jpg", + "image_caption": [ + "Geometric layer" + ], + "image_footnote": [], + "bbox": [ + 700, + 332, + 859, + 393 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/f005aea12fc3690efb0eaff4f4543bcdbb88f79353f3d6c94bdf59d688586855.jpg", + "image_caption": [ + "Connectivity layer" + ], + "image_footnote": [], + "bbox": [ + 702, + 407, + 857, + 467 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/218888cb25494023aaae06c9ff7b5f0bb0ce06a5e6579921a59ae6257f422520.jpg", + "image_caption": [ + "Traffic regulation layer" + ], + "image_footnote": [], + "bbox": [ + 702, + 479, + 857, + 540 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Abstract", + "text_level": 1, + "bbox": [ + 246, + 643, + 326, + 660 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Ensuring adherence to traffic sign regulations is essential for both human and autonomous vehicle navigation. While current online mapping solutions often prioritize the construction of the geometric and connectivity layers of HD maps, overlooking the construction of the traffic regulation layer within HD maps. Addressing this gap, we introduce MapDR, a novel dataset designed for the extraction of Driving Rules from traffic signs and their association with vectorized, locally perceived HD Maps. MapDR features over 10,000 annotated video clips that capture the intricate correlation between traffic sign regulations and", + "bbox": [ + 89, + 680, + 485, + 847 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "lanes. Built upon this benchmark and the newly defined task of integrating traffic regulations into online HD maps, we provide modular and end-to-end solutions: VLE-MEE and RuleVLM, offering a strong baseline for advancing autonomous driving technology. It fills a critical gap in the integration of traffic sign rules, contributing to the development of reliable autonomous driving systems. Code is available at https://github.com/MIV-XJTU/MapDR.", + "bbox": [ + 511, + 645, + 908, + 767 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "1. Introduction", + "text_level": 1, + "bbox": [ + 513, + 799, + 645, + 814 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "The rise of autonomous vehicles and intelligent transportation systems has emphasized the importance of accurate, reliable navigational data. High-Definition (HD) maps, with their detailed representations of road elements, have become essential to supporting these systems. An HD map typically", + "bbox": [ + 511, + 825, + 908, + 902 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "CVF", + "bbox": [ + 106, + 2, + 181, + 42 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore.", + "bbox": [ + 236, + 0, + 810, + 46 + ], + "page_idx": 0 + }, + { + "type": "page_footnote", + "text": "*Equal contribution.", + "bbox": [ + 107, + 864, + 220, + 876 + ], + "page_idx": 0 + }, + { + "type": "page_footnote", + "text": "$^{\\dagger}$ Work done during the internship at Amap, Alibaba Group.", + "bbox": [ + 109, + 877, + 421, + 888 + ], + "page_idx": 0 + }, + { + "type": "page_footnote", + "text": "$\\ddagger$ Corresponding author.", + "bbox": [ + 109, + 888, + 235, + 901 + ], + "page_idx": 0 + }, + { + "type": "page_number", + "text": "6823", + "bbox": [ + 482, + 944, + 514, + 955 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "consists of three primary layers: the geometric layer, connectivity layer, and traffic regulation layer [16, 35], as shown in Fig. 1. The geometric layer provides vector information, such as dividers and lane centerlines; the connectivity layer defines lane relationships to assist in path planning; and the traffic regulation layer encodes rule information (e.g., HOV lanes, bus lanes, speed-limit zones) associated with lanes, ensuring compliant driving.", + "bbox": [ + 89, + 90, + 482, + 210 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "A limitation of conventional HD maps is their inability to support rapid, real-time updates, making the construction of online HD maps an emerging trend. Existing methods, such as MapTR [30] and TopoMLP [54], address only the geometric and connectivity layers, overlooking the traffic regulation layer. As a result, autonomous driving systems currently still rely on offline maps to obtain traffic regulations, an approach that runs counter to the trend of online HD map construction.", + "bbox": [ + 89, + 213, + 483, + 349 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Traffic signs serve as a \"visual language\" on the road, playing a critical role in defining traffic regulations. Inspired by human driving processes, the creation of traffic regulations can be divided into two stages: 1) understanding the rules conveyed by traffic signs, and 2) determining the lanes to which these rules apply. Although OpenLane-V2 [52] has made strides in associating signs with lanes in online mapping, it has two primary limitations. First, it considers only directional signs, neglecting other types of regulations. Second, its label annotations are limited to sign categories, falling short of the detailed rule descriptions required for autonomous driving. A more structured description, in alignment with HD map standards, is essential for supporting comprehensive autonomous driving applications [16, 35].", + "bbox": [ + 89, + 352, + 483, + 564 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "To bridge this gap, we introduce MapDR, the first dataset specifically designed for the challenging task of integrating traffic regulations into existing online HD maps. MapDR focuses on complex traffic scenes with diverse signage, collecting over 10,000 video clips across three representative cities in China, with each clip containing at least one traffic sign. In addition to vectorized representations of dividers, boundaries, and centerlines, MapDR provides structural annotations of traffic regulations and their associations with lanes, offering a comprehensive foundation for advancing online HD maps. More details of MapDR can be found in Sec. 4.", + "bbox": [ + 89, + 566, + 483, + 744 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Based on MapDR, in addition to the end-to-end task definition, we also divide this task into two innovative subtasks aimed at bolstering research in this domain: 1) rule extraction from traffic signs, and 2) rule-lane correspondence reasoning. We introduce Vision-Language Encoder (VLE) and Map Element Encoder (MEE) to handle the interaction of multimodal data, encompassing images, texts, and vectors. Through the concatenation of these two models, integrated traffic regulations can be obtained. Additionally, inspired by recent advancements in Multimodal Large Lan", + "bbox": [ + 89, + 750, + 483, + 900 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Table 1. Comparison of the existing datasets. \"Sign\" and \"Lane\" denote whether the dataset focuses on traffic signs and lanes. Only those annotated with formatted (\"Fmt.\") rules and the correspondence (\"Corr.\") between rules and lanes can form driving rules. \"Clip\" represents whether the data is organized in the form of video clips. \"* denotes that these samples are not newly collected and are built upon the previous dataset.", + "bbox": [ + 511, + 88, + 906, + 186 + ], + "page_idx": 1 + }, + { + "type": "table", + "img_path": "images/b25a9b0b908a767c82761e6ed9b6d10682d6595a78579c0a40c19fa8f92f6111.jpg", + "table_caption": [], + "table_footnote": [], + "table_body": "
DatasetSignLaneDriving RulesNumber of SamplesYear
Fmt.Corr.ImageClipRegion
nuScenes [9]✓1400K1KWorldwide2019
Argoverse2 [53]✓2100K1KUSA2021
CTSU [19]✓5K/China2021
OpenLane [11]✓200K*1K*Worldwide2022
RS10K [20]✓✓10K/China2023
OpenLane-V2 [52]✓✓✓466K*2K*Worldwide2023
Waymo [47]✓✓390K2KUSA2024
MapDR(ours)✓✓✓✓400K10KChina2024
", + "bbox": [ + 516, + 196, + 903, + 303 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "guage Models (MLLMs), we also explore an end-to-end MLLM solution called RuleVLM to solve this task holistically. These approaches establish robust baselines and aim to inspire future research in this area. For detailed descriptions of the proposed tasks and metrics, please refer to Sec. 3. For approaches, please refer to Secs. 5 and 6.", + "bbox": [ + 511, + 330, + 906, + 421 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "To sum up, our contributions are as follows:", + "bbox": [ + 532, + 422, + 823, + 436 + ], + "page_idx": 1 + }, + { + "type": "list", + "sub_type": "text", + "list_items": [ + "- For the first time, We introduce the novel task of extracting lane-level rules from traffic signs and integrating them into vectorized HD maps, alongside the MapDR dataset and evaluation metrics for benchmarking.", + "- MapDR comprises over 10,000 video clips and 400,000 front-view images captured across multiple cities in China, covering diverse traffic conditions and including more than 18,000 annotated lane-level rules. All data are newly collected and meticulously validated.", + "- We propose two modeling approaches, modular and end-to-end, to address the integration of lane-level rules into vectorized HD maps, providing effective baselines for future research." + ], + "bbox": [ + 545, + 441, + 906, + 667 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "2. Related Work", + "text_level": 1, + "bbox": [ + 513, + 686, + 653, + 700 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "2.1. HD Map Construction", + "text_level": 1, + "bbox": [ + 511, + 710, + 722, + 727 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "HD maps construction have seen significant advancements, with a focus on traffic element perception, including lane detection and traffic sign recognition [8, 9, 17, 18, 22, 46, 53, 58, 60]. The shift towards BEV perception and vectorization for end-to-end HD maps construction has gained traction [9, 11, 53]. Notable works include HDMapNet, which aggregates semantic segmentation results [29], LSS [40] estimates depth to transfer image features to BEV features, while VectorMapNet [36] is the first end-to-end framework for sequential vector point prediction to generate HD maps without post-processing. MapTR [30] and its enhanced", + "bbox": [ + 511, + 734, + 906, + 900 + ], + "page_idx": 1 + }, + { + "type": "page_number", + "text": "6824", + "bbox": [ + 482, + 944, + 514, + 955 + ], + "page_idx": 1 + }, + { + "type": "image", + "img_path": "images/307ba3b46ca57ad5adb2d0d4c501f5f7f069580f9a1c1f82b612dfb815c84d96.jpg", + "image_caption": [ + "Figure 2. Overview of the sub-tasks. Step 1 ~ Step 4 shows a case of driving by the rules. Step 2 and Step 3 demonstrates the specific role of two sub-tasks, respectively." + ], + "image_footnote": [], + "bbox": [ + 133, + 88, + 859, + 229 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "version, MapTRv2 [31], introduced a unified permutation-equivalent modeling approach and extended it to a general framework supporting centerline learning and 3D map construction. However, these efforts have largely overlooked the integration of traffic sign rules into HD maps.", + "bbox": [ + 88, + 296, + 485, + 372 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "2.2. Traffic Element Association", + "text_level": 1, + "bbox": [ + 89, + 380, + 341, + 395 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Traffic element association aims to link elements like traffic signs with lanes. As demonstrated in Tab. 1, CTSU has initiated internal elements association to describe traffic sign in $\\{key: value\\}$ form, however lacking both generalization of driving rules from description and lane association [19]. VTKGG [20] propose to utilize a graph model for connectivity but also lacks structured expression of driving rules for motion planning and requires complex integration into HD maps, which is typically expressed in the BEV space. OpenLane-V2 [52] advances traffic sign and lane association in BEV space, but is constrained by single-label classification for traffic sign rather than structured descriptions for fine-grained driving rules. This makes it insufficient for signs with multiple rules, which are complex but common in real scenarios. Recent MLLM-based benchmarks [10, 38, 41, 43, 44] for autonomous driving, such as MAPLM [10], prioritize end-to-end motion planning over precise rule extraction from traffic sign, lacking evaluation for rule reasoning. MapDR addresses this gap by focusing on traffic sign rule extraction and lane association.", + "bbox": [ + 91, + 401, + 483, + 704 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "2.3. Vision-Language Models", + "text_level": 1, + "bbox": [ + 89, + 712, + 318, + 728 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Vision-Language Models (VLMs) facilitates multimodal applications by learning joint representations of vision and language data. Visual Question Answer (VQA) tasks provide answers to image-related questions [4], while Visual Information Extraction (VIE) tasks extract structured information from visual and textual data [4, 23, 55, 56]. In Autonomous Driving (AD), VLMs are increasingly used for comprehensive traffic scene understanding and decision-making. The field has seen various approaches, including using transformers [51] for joint encoding [23, 25], excelling at multimodal information interaction, and independent en", + "bbox": [ + 89, + 734, + 485, + 900 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "coders for different modalities [24, 42] that are proficient in multimodal retrieval. Cross-modal representation methods [26, 59] combine these advantages, and the latest LLM-based research [28, 32-34] has achieved state-of-the-art results in various multimodal tasks. Nowadays, an increasing number of methods are leveraging LLMs to achieve impressive results [12, 37, 45, 57], with works like DriveLLM [13] showing significant potential in AD. However, addressing hallucination [6] remains the most crucial aspect for LLM-based approaches.", + "bbox": [ + 511, + 296, + 908, + 448 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3. Task Definition : Integrating Traffic Sign Regulations into HD Maps", + "text_level": 1, + "bbox": [ + 511, + 467, + 906, + 503 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "The ability to discern rules from traffic signs and to associate them with specific lanes is pivotal for autonomous navigation. Our proposed task focuses on extracting lane-level driving rules denoted as $R = \\{r_i\\}_{i=1}^m$ and reasoning the correspondence between these rules and the centerlines $L = \\{l_i\\}_{i=1}^k$ in the local vectorized HD map, where $m$ is the number of rules and $k$ is the number of centerlines. Each rule $r_i$ is a set of pre-defined properties in $\\{key : value\\}$ pairs, as shown in Fig. 3. The final correspondence forms a bipartite graph $G = (R \\cup L, E)$ , where $E \\subseteq \\{0,1\\}^{m \\times k}$ and the element $E_{ij}$ in the $i$ -th row and $j$ -th column of matrix $E$ represents the corresponding status between $r_i$ and $l_j$ . In this case, we assume the entire approach as $\\mathcal{F}$ and the overall process can be formulated as $G = \\mathcal{F}(X, L)$ , where $X = \\{x_i\\}_{i=1}^n$ represents a series of image sequences and $n$ is the number of frames. To independently investigate the capabilities of rule extraction and correspondence reasoning within the overall task, we further divided the overall task into two sub-tasks, defined as follows.", + "bbox": [ + 511, + 512, + 908, + 800 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3.1. Rule Extraction from Traffic Sign", + "text_level": 1, + "bbox": [ + 511, + 815, + 808, + 832 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "As shown in Step 2 of Fig. 2, this sub-task involves extracting multiple rules $R = \\{r_i\\}_{i=1}^m$ from a series of image sequences $X = \\{x_i\\}_{i=1}^n$ . The rule extraction model, denoted as $\\mathcal{M}$ , can be expressed as $R = \\mathcal{M}(X)$ .", + "bbox": [ + 511, + 839, + 908, + 902 + ], + "page_idx": 2 + }, + { + "type": "page_number", + "text": "6825", + "bbox": [ + 482, + 944, + 514, + 955 + ], + "page_idx": 2 + }, + { + "type": "image", + "img_path": "images/732294b9cf93258cb44928eb52493a5934660a5921e35324248e2eb038033e09.jpg", + "image_caption": [ + "Figure 3. Visualization of dataset demo. Multiple lane-level rules of a single traffic sign are annotated in $\\{key: value\\}$ format. Directed lines indicate the correspondence between rules and particular centerlines." + ], + "image_footnote": [], + "bbox": [ + 112, + 88, + 459, + 195 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "3.2. Rule-Lane Correspondence Reasoning", + "text_level": 1, + "bbox": [ + 89, + 290, + 421, + 306 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "As shown in Step 3 of Fig. 2, the reasoning process establishes the correspondence between centerlines $L = \\{l_i\\}_{i=1}^k$ and all rules $R$ . We denote the correspondence reasoning model as $\\mathcal{T}$ , and this process can be described as $E = \\mathcal{T}(R, L)$ . The final reasoning result forms a bipartite graph $G = (R \\cup L, E)$ , which means corresponding relationships only exist between rules and centerlines.", + "bbox": [ + 89, + 313, + 483, + 419 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "4. The MapDR Dataset & Benchmark", + "text_level": 1, + "bbox": [ + 89, + 433, + 413, + 450 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "We introduce the MapDR dataset, meticulously annotated with traffic sign regulations and their correspondences to lanes, as shown in Fig. 3. The dataset encompasses a diverse range of scenarios, weather conditions, and traffic situations, with over 10,000 traffic scene segments, 18,000 driving rules, and 400,000 images. Traffic signs typically have varying textual descriptions, text layouts, and positions on the road, which add complexity to the task.", + "bbox": [ + 89, + 459, + 483, + 579 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "The majority of the data originates from Beijing and Shanghai, with additional scenes from Guangzhou. Fig. 4 illustrates the geographic spread and variety of traffic signs. The dataset reflects a natural long-tail distribution, with a prevalence of bus and direction lanes and a scarcity of tidalflow lanes. We primarily focus on traffic signs that indicate lane-level rules, collected from cities with the most complex and diverse traffic scenarios in China, ensuring realistic and representative data. All images have undergone privacy and safety processing to obscure license plates and faces. More comprehensive statistic of dataset and case demonstrations can be found in the supplement.", + "bbox": [ + 89, + 580, + 483, + 762 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "4.1. Raw Data & Annotation", + "text_level": 1, + "bbox": [ + 89, + 772, + 313, + 787 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Raw Data. MapDR is collected from real-world traffic scenes, each scene segment (video clip) captures front-view images within a $100m \\times 100m$ area centered on the traffic sign, with a consistent resolution of $1920 \\times 1240$ . Each clip contains 30 to 60 frames, captured at 1 frame every 2 meters, ensuring consistent spatial intervals. Each video clip focuses on a single traffic sign and provides its position in 3D space.", + "bbox": [ + 89, + 794, + 483, + 900 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Camera intrinsics and poses are provided for each frame, and coordinates for each clip are transformed to distinct ENU systems.", + "bbox": [ + 511, + 90, + 903, + 135 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "All vectors of local map in the target area are provided as 3D point lists, generated using our algorithm similar to MapTRv2 [31]. Each lane vector has a type, such as divider, centerline, crosswalk, or boundary. For example, the centerline is defined as $L = \\{l_i\\}_{i=1}^k$ , where each vector $l_i$ is composed of multiple 3D points $l_i = [p_1, \\ldots, p_n]$ , and $p_j = (x_j, y_j, z_j)$ represents the coordinates of the current point. The pipeline of dataset production is illustrated in Fig. 5, and detailed data acquisition and annotation procedures can be found in the supplement material.", + "bbox": [ + 511, + 137, + 906, + 289 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "**Formatted Rules.** Each video clip may contain multiple lane-level rules, denoted as $R$ . Each rule is expressed by symbols and text on the sign, requiring interpretation. We reference existing HD map design and data specifications [1, 16, 35] to define lane-level driving rules in the JSON format. As shown in Fig. 3, each rule $r_i$ comprises 8 predefined properties in the form of {key : value} pairs. This definition schema encompasses most scenarios and supports integration into existing autonomous driving systems. We enclose the symbols and texts denoting each distinct rule on traffic signs with polygons and project them into 3D space as $P_i = [p_1, \\ldots, p_n]$ , where $n$ varies. Researchers can optionally use this information to facilitate rule extraction.", + "bbox": [ + 511, + 313, + 908, + 508 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Correspondence between Rules & Lanes. Based on formatted rules $R$ and centerlines $L$ , corresponding centerlines of each rule are annotated as shown in Fig. 3. Therefore correspondence between rules and centerlines can be formed as a bipartite graph $G = (R \\cup L, E)$ , where $E \\subseteq \\{0,1\\}^{|R| \\times |L|}$ and the positive edges only exist between $R$ and $L$ as demonstrated in Sec. 3.2. Specifically, $E_{ij} = 1$ represents that", + "bbox": [ + 511, + 532, + 908, + 642 + ], + "page_idx": 3 + }, + { + "type": "image", + "img_path": "images/8d41f316abb264102e743f75839b8ebe308aa55d8e30be049b22a61d8fd1b753.jpg", + "image_caption": [ + "Figure 4. Geographic location distribution of the collected traffic signs and proportions of various lane types represented in all signs. The geographic distribution is visualized based on OpenStreetMap [2]." + ], + "image_footnote": [], + "bbox": [ + 535, + 666, + 885, + 830 + ], + "page_idx": 3 + }, + { + "type": "page_number", + "text": "6826", + "bbox": [ + 482, + 944, + 514, + 955 + ], + "page_idx": 3 + }, + { + "type": "image", + "img_path": "images/6ac67f847f7504ee637375f74cc16c051b9fc9e4b23c72a868597759f49e1979.jpg", + "image_caption": [ + "Figure 5. Pipeline of dataset production. The location of traffic signs are sampled from existing database then front-view images of each sign are newly collected. Vectorized map is processed in cloud server. Finally formatted rules and correspondence between rules and centerlines are annotated and organized as MapDR." + ], + "image_footnote": [], + "bbox": [ + 109, + 87, + 464, + 154 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "vehicle driving on the lane with centerline $l_{j}$ should follow the driving rule $r_{i}$ .", + "bbox": [ + 89, + 262, + 483, + 292 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "4.2. Evaluation Metrics", + "text_level": 1, + "bbox": [ + 89, + 303, + 274, + 316 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Based on the task defined in Sec. 3, we propose evaluation metrics for entire task as well as for two sub-tasks separately. The $F_{1}$ score is adopted as the benchmark for methods rank.", + "bbox": [ + 89, + 325, + 485, + 371 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Rule Extraction (R.E.). Given the ground truth $R$ and predicted rules $\\hat{R}$ , we propose to calculate the Precision $(P_{R,E.})$ and Recall $(R_{R,E.})$ to evaluate the capability of rules extraction as defined in Eq. (1), where $\\hat{r_i} = r_i$ represents all the properties are predicted correctly.", + "bbox": [ + 89, + 390, + 483, + 467 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\nP _ {R. E.} = \\frac {\\left| \\hat {R} \\cap R \\right|}{\\left| \\hat {R} \\right|} \\quad R _ {R. E.} = \\frac {\\left| \\hat {R} \\cap R \\right|}{\\left| R \\right|} \\tag {1}\n$$\n", + "text_format": "latex", + "bbox": [ + 148, + 478, + 483, + 513 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Correspondence Reasoning (C.R.). Given the ground truth of correspondence bipartite graph $G = (R \\cup L, E)$ and predicted graph $\\hat{G} = (R \\cup L, \\hat{E})$ , we propose to calculate Precision ( $P_{C.R.}$ ) and Recall ( $R_{C.R.}$ ) of edge set $E$ to evaluate the capability of correspondence reasoning individually. Metrics are defined as Eq. (2).", + "bbox": [ + 89, + 532, + 483, + 625 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\nP _ {C. R.} = \\frac {\\left| \\hat {E} \\cap E \\right|}{\\left| \\hat {E} \\right|} \\quad R _ {C. R.} = \\frac {\\left| \\hat {E} \\cap E \\right|}{\\left| E \\right|} \\tag {2}\n$$\n", + "text_format": "latex", + "bbox": [ + 148, + 633, + 482, + 671 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Overall. As defined in Sec. 3, the final predicted results can be obtained through an end-to-end prediction or a combination of two sequential sub-tasks. Given the predicted rules, correspondence should be reasoned between $\\hat{R}$ and $L$ which means the prediction of entire task is $\\hat{G} = (\\hat{R} \\cup L, \\hat{E})$ and the ground truth is consistent $G = (R \\cup L, E)$ . We evaluate Precision $(P_{all})$ and Recall $(R_{all})$ using the subgraph $G^{s}$ , where $G^{s} = \\{g_{ij}^{s}\\}_{i=1,j=1}^{m,k}$ , $g_{ij}^{s} = (\\{r_{i}, l_{j}\\}, e_{ij})$ . In set of sub-graph $G^{s}$ , $m$ is the number of rules, and $k$ is the number of centerlines. Furthermore, we propose the $F_{1}$ for the final benchmark ranking. Metrics are defined in Eq. (3) and Eq. (4). The $F_{1}$ is derived from the calculations of $P_{all}$ and $R_{all}$ . We provide an example of calculating the Overall metrics in the supplement material.", + "bbox": [ + 89, + 688, + 485, + 900 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\nP _ {a l l} = \\frac {\\left| \\hat {G} ^ {s} \\cap G ^ {s} \\right|}{\\left| \\hat {G} ^ {s} \\right|} \\quad R _ {a l l} = \\frac {\\left| \\hat {G} ^ {s} \\cap G ^ {s} \\right|}{\\left| G ^ {s} \\right|} \\tag {3}\n$$\n", + "text_format": "latex", + "bbox": [ + 553, + 114, + 906, + 152 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\nF _ {1} = 2 \\times \\frac {P _ {\\text {a l l}} \\times R _ {\\text {a l l}}}{P _ {\\text {a l l}} + R _ {\\text {a l l}}} \\tag {4}\n$$\n", + "text_format": "latex", + "bbox": [ + 633, + 162, + 906, + 196 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "5. Modular Approach", + "text_level": 1, + "bbox": [ + 511, + 205, + 702, + 223 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "In order to conduct a comprehensive analysis of the entire task, a modular approach is implemented in this section to address the two sub-tasks. It is demonstrated that these interconnected modules can effectively achieve the ultimate objective of integrating driving rules into the HD map. To address the multimodal information interaction encompassing images, texts, and vectors, a Vision-Language Encoder (VLE) and a Map Element Encoder (MEE) are developed. The following sections expound upon their structures, applications, and the experimental results on MapDR.", + "bbox": [ + 511, + 229, + 906, + 381 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "5.1. Architecture", + "text_level": 1, + "bbox": [ + 511, + 391, + 648, + 405 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Vision-Language Encoder. Inspired by vision-language frameworks [7, 25-27, 42], we designed a vision-language fusion model named VLE, following [26]. As shown in Fig. 6, VLE uses ViT-b16 [15] as the vision encoder, with the text encoder and multimodal fusion encoder each consisting of $L$ transformer layers [51]. Each layer of the fusion encoder includes a cross-attention module for fusion [26]. In practice, distinct rules are represented by varying numbers of symbols and texts, as shown in the OCR results in Fig. 6. To address the challenge of representing variable-length input as fixed-length features, we introduce a [CLS] token for an entire rule and several [STC] tokens for sentence-level representation. The specific usage of these tokens is detailed in Sec. 5.2. Furthermore, we incorporate inter-instance and intra-instance attention mechanisms [31] to enhance model performance by capturing interactions and independence between and within sentences. In addition to content, layout captures the relative positions of symbols and texts, offering important semantic meaning. To leverage this, we encode the layout using the method from [48] and the relative positions of characters as position embedding following [14]. As shown in VLE in Fig. 6, text embedding, layout embedding, and position embedding together form the input of the text encoder.", + "bbox": [ + 511, + 412, + 908, + 775 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Map Element Encoder. Vectors can be represented as sequences of points, similar to words in sentences. Inspired by this, we designed MEE akin to language models [14]. The MEE employs $M$ transformer layers for vector encoding and $N$ cross-attention layers for multimodal fusion. Utilizing the method from [48], points of each vector are embedded as point embedding. To achieve a fixed-length representation,", + "bbox": [ + 511, + 794, + 908, + 902 + ], + "page_idx": 4 + }, + { + "type": "page_number", + "text": "6827", + "bbox": [ + 482, + 944, + 514, + 955 + ], + "page_idx": 4 + }, + { + "type": "image", + "img_path": "images/20dd1837f8f401df33c212c7e9b6b9ac2bae69f347a90298fbacad52d2affe46.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 138, + 92, + 857, + 277 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/7eac7cfdb22f5c436f60b18c8f4eedb176cd67f2f171a678d7f6ba2ad7dbae6e.jpg", + "image_caption": [ + "Figure 6. Overview of the modular approach. Entire approach can be divided into two main parts: Rule Extraction from Traffic Sign (top) and Rule-Lane Correspondence Reasoning (bottom). Rule Extraction model consists of two sequential stages with the same structure VLE but unshared parameters, and the training procedure is independent." + ], + "image_footnote": [], + "bbox": [ + 140, + 279, + 857, + 409 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "we add [VEC] tokens as the first token of each vector, similar to [STC] tokens in the VLE. We also introduce learnable type embedding for vector types, learnable instance embedding to distinguish vector instances, and position embedding from [14] to encode the relative positions of multiple points within a vector. These embedding are aggregated as the input of vector encoder, as shown in Fig. 7. In addition, we employ inter-instance and intra-instance attention mechanisms [31] to prioritize interactions within vectors over interactions between vectors, as depicted in the dashed box on the right side of Fig. 7. The [VEC] token in output serves as fused feature of rules and vectors, enabling the final prediction of their relationships through association head.", + "bbox": [ + 89, + 491, + 483, + 686 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "5.2. Implementation", + "text_level": 1, + "bbox": [ + 89, + 694, + 250, + 709 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "We utilize VLE and MEE as backbones to integrate multiple modalities and address these two sub-tasks. The specific procedures are detailed as follows:", + "bbox": [ + 89, + 715, + 482, + 762 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Rule Extraction from Traffic sign. To clarify the objectives of model, we first cluster symbols and texts into groups. As shown in the upper part of Fig. 6, the VLE is used to encode OCR results and images. By calculating the cosine similarity between [STC] tokens, different symbols and texts are clustered into groups. This process is supervised by contrastive loss during training. Next, using grouped OCR results as text input and maintaining the VLE structure, we", + "bbox": [ + 89, + 779, + 483, + 900 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "extract lane-level rules. We employ multiple linear layers as understanding head for the [CLS] token to predict the corresponding value for each property of the rules. This facilitates to express all rules as $\\{key:value\\}$ pairs.", + "bbox": [ + 511, + 491, + 906, + 551 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Rule-Lane Correspondence Reasoning. MEE is designed for vector encoding and interaction with rules. Each formatted rule is mapped to an embedding through MLP and", + "bbox": [ + 511, + 574, + 908, + 621 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/27b818cb5a8c6c71eff3683acaf306cecfb29692e6fad6b0e40f1e43d71736d1.jpg", + "image_caption": [ + "Figure 7. Structure of MEE. MEE serves as correspondence reasoning model. Learnable embeddings are introduced within input to enhance the representing capacity of vector types. inter & intra-instance attention mechanisms facilitate to capture the relationships and independence of individual vectors." + ], + "image_footnote": [], + "bbox": [ + 545, + 646, + 885, + 816 + ], + "page_idx": 5 + }, + { + "type": "page_number", + "text": "6828", + "bbox": [ + 482, + 944, + 514, + 955 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/81fe351a1fe39d9a99bba6400a5b27163cf6ccb63532f0941af583291dd262e1.jpg", + "image_caption": [ + "Figure 8. Overview of end-to-end approaches. Based on different vectorized HD maps encoding methods, approaches can be categorized into three types." + ], + "image_footnote": [], + "bbox": [ + 133, + 85, + 867, + 305 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "fused with vector features in the fusion encoder, as shown in the lower part of Fig. 6. We add a binary classification head after each [VEC] token to determine the relationship between the current centerline and rule. Details can be found in supplement.", + "bbox": [ + 89, + 372, + 483, + 448 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "5.3. Experiment", + "text_level": 1, + "bbox": [ + 89, + 460, + 218, + 477 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Setup. The dataset is split into train and test sets in the ratio of $9:1$ . $L = 6$ in VLE and $M = 2$ , $N = 2$ in MEE. Input images are resized to $256 \\times 256$ and the feature dimension is 768 with consistent 12 attention heads. We initialize VLE with pre-trained weights of DeiT [50] and BERT [14] while MEE is trained from scratch. The training procedure runs 50 and 120 epochs for VLE and MEE, respectively. More details can be found in the supplement.", + "bbox": [ + 89, + 483, + 485, + 604 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Results. We use a heuristic method based on OCR character matching and nearest lane matching as the modular baseline. Meanwhile we make modifications to ALBEF [26] and BERT [14] to adapt them to our task. As shown in Tab. 2, the heuristic method performs poorly in both R.E. and C.R. tasks, underscoring the challenging nature of the MapDR dataset. ALBEF-BERT method significantly enhances performance in the R.E. sub-task however lacking adequate adaptation to vector information hampers its performance on the C.R.. Tab. 3 indicates the attention mechanisms significantly improve $R_{R,E}$ , while layout of text brings marginal improvement. For C.R., the attention mechanisms is the critical part that promotes the effectiveness of MEE. Instance embedding slightly improves $P_{C,R}$ and $R_{C,R}$ , while type embedding significantly enhances both, indicating that vector types help the model establish rule-lane correspondence. The separate evaluation results of all lane types and the strategies of heuristic method can be found in the supplement.", + "bbox": [ + 89, + 628, + 485, + 901 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "6. End-to-End Approach", + "text_level": 1, + "bbox": [ + 511, + 371, + 725, + 388 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Recently, there has been notable progress in the development of Multimodal Large Language Models (MLLMs) [3, 5, 39, 49], demonstrating significant capabilities in processing multimodal data and acquiring implicit knowledge of modality alignment. Their large-scale pre-training endows them with exceptional understanding and generative abilities, thereby enabling the comprehensive execution of the task proposed by MapDR in an end-to-end manner. This section aims to address the overall task using Qwen-VL [5] and to investigate three distinct vector encoding methods. Moreover, we introduce an innovative RuleVLM model based on MEE, designed to process inputs from three modalities (images, texts, and vectors) and to produce driving rules, which yield favorable results on MapDR. Further elaboration on these concepts is presented in subsequent sections.", + "bbox": [ + 511, + 400, + 908, + 627 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "6.1. Architecture", + "text_level": 1, + "bbox": [ + 513, + 648, + 648, + 662 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "As shown in Fig. 8, we classify MLLM methods into three types: Qwen-VL(TextPrompt), Qwen-VL(VisualPrompt), and RuleVLM, based on different vector encoding methods. These methods utilize the framework of Qwen-VL [5], with ViT [15] serving as the image encoder. Qwen-VL(TextPrompt) encodes centerline coordinates in the PV image as text, inputs them into QwenLM; Qwen-VL(VisualPrompt) visualizes the centerline and its index in the PV image and inputs it as Visual Prompt into QwenLM, while RuleVLM uses MEE with the cross-attention layer removed to encode vectorized centerline results and aligns it with LLM through an adapter. The goal of all these methods is to output a serialized set of driving rules in text form, decode it through a json decoder to restore formatted data, and then integrate it into HD maps for further applications.", + "bbox": [ + 511, + 674, + 908, + 900 + ], + "page_idx": 6 + }, + { + "type": "page_number", + "text": "6829", + "bbox": [ + 482, + 945, + 514, + 955 + ], + "page_idx": 6 + }, + { + "type": "table", + "img_path": "images/67b8ace9a40906a3302c9d5f37377a47290759d42fca61d9445c7cacc5497b81.jpg", + "table_caption": [ + "Table 2. Evaluation of the overall task. The heuristic method and the Qwen-VL(TextPrompt) method serve as the baselines for the modular and end-to-end approach, respectively. \" - \" denotes end-to-end approach is not suitable to independent evaluations of C.R. because these approaches do not utilize ground truth of rules for correspondence reasoning independently." + ], + "table_footnote": [], + "table_body": "
ModelTypeR.E.C.R.Overall
PR.E.(%)RR.E.(%)PC.R.(%)RC.R.(%)Pall(%)Rall(%)F1 Score
Heuristic18.0111.5133.0517.995.012.730.035
ALBEF-BERTModular75.7857.564.1417.250.240.780.003
VLE-MEE76.6774.5478.0582.1663.3567.370.653
Qwen-VL(TextPrompt)42.2141.09--8.398.170.083
Qwen-VL(VisualPrompt)End-to-End89.2989.50--39.1439.230.392
RuleVLM89.2889.44--64.1664.280.642
", + "bbox": [ + 91, + 142, + 903, + 275 + ], + "page_idx": 7 + }, + { + "type": "table", + "img_path": "images/86aee286f340f50ca0c7b7b5bc47de37d5f25d5de6d12cb383f029a27437521d.jpg", + "table_caption": [ + "Table 3. Evaluation of sub-tasks. \"Attn.\" indicates intra & inter-instance attention mechanisms. \"Layout\" refers to the text layout applied in VLE. \"In.E.\" and \"Ty.E.\" denotes instance and type embedding in MEE, respectively." + ], + "table_footnote": [], + "table_body": "
VLEPR.E.(%)RR.E.(%)
Attn.Layout
XX75.7857.56
✓X76.8671.75
✓✓76.6774.54
MEEPC.R.(%)RC.R.(%)
Attn.In.E.Ty.E.
XXX4.1417.25
✓XX68.9171.39
✓✓X69.6872.76
✓✓✓78.0582.16
", + "bbox": [ + 116, + 364, + 444, + 559 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "6.2. Implementation", + "text_level": 1, + "bbox": [ + 89, + 585, + 250, + 602 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "The JSON-formatted driving rules in MapDR are serialized into text and used as the answers in the QA format for training and evaluation. For all methods based on QwenVL, the pre-trained weights of Qwen-VL-Chat (9.6B) are used for initialization. For method using MEE, the trained weights of MEE from the modular approach are utilized for its initialization. During the supervised fine-tuning, we perform LoRA [21] fine-tuning of the QwenLM, while keeping the parameters of ViT frozen. For RuleVLM, we employ a linear layer as the adapter between QwenLM and MEE, while the parameters of MEE and the adapter remains trainable. We use special tokens to separate the centerline tokens extracted by MEE from the visual and text tokens. Additionally, to prevent overfitting, we shuffle the order of centerlines in the local vectorized map during training.", + "bbox": [ + 88, + 608, + 485, + 835 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "6.3. Experiment.", + "text_level": 1, + "bbox": [ + 89, + 847, + 222, + 863 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Setup. All experiments are conducted on QA formatted MapDR dataset with 20 epochs of training on NVIDIA RTX", + "bbox": [ + 89, + 869, + 483, + 902 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "A6000. The dimension of hidden states remains 4096 while $r$ and alpha of LoRA configuration are set 64 and 16 respectively. Meanwhile training batch size is 6 and the max training token length remains 2048.", + "bbox": [ + 511, + 300, + 906, + 362 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Results. We conducted zero-shot evaluations on a subset of MapDR using existing MLLMs [3, 5, 39, 49] as qualitative investigation, but we encounter challenges in achieving stable structured outputs, further details are provided in the supplement. Additionally, as shown in the Tab. 2, Qwen-VL(TextPrompt) method serves as a end-to-end baseline do not perform better on both the R.E. and overall evaluation. Textual representations of centerline coordinates are challenging for MLLMs to perform spatial reasoning and also resulted in excessive length of sequence, which affected $P_{R.E.}$ and $R_{R.E.}$ . Qwen-VL(VisualPrompt) visualizes the centerline within the image, resulting in reduced text sequence length, and demonstrates notable enhancements in the R.E. task compared to the baseline. However, the overall metrics are suboptimal, suggesting that image-based spatial reasoning remains an exceedingly challenging endeavor. The RuleVLM method that we propose introduces MEE for the independent encoding of vectorized HD maps information. This facilitates a more direct extraction of centerline features, resulting in significant enhancements in the overall metrics.", + "bbox": [ + 511, + 378, + 908, + 681 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "7. Conclusion", + "text_level": 1, + "bbox": [ + 511, + 694, + 633, + 709 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "In this work, we defined the task of integrating traffic regulations into vectorized HD maps to facilitate the construction of the traffic regulation layer within online HD maps. We constructed the MapDR dataset, which includes over 10,000 video clips, more than 400,000 images, and at least 18,000 driving rules, covering a total mileage of approximately $800Km$ . Additionally, we proposed the modular approach MEE-VLE and the end-to-end approach RuleVLM, establishing effective baselines for further research. We hope that MapDR can serve as a starting point, and that more researchers can contribute to expanding this dataset or proposing even better solutions.", + "bbox": [ + 511, + 719, + 908, + 900 + ], + "page_idx": 7 + }, + { + "type": "page_number", + "text": "6830", + "bbox": [ + 482, + 944, + 514, + 955 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "References", + "text_level": 1, + "bbox": [ + 91, + 90, + 187, + 104 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[1] Navigation data standard. https://nds-association.org/.4", + "[2] Openstreetmap. https://github.com/openmaptiles/openmaptiles.4", + "[3] Anthropic. Claude-3. https://www.anthropic.com/news/claudi-3-family, 2024.7,8", + "[4] Stanislaw Antol, Aishwarya Agrawal, Jiasen Lu, Margaret Mitchell, Dhruv Batra, C. Lawrence Zitnick, and Devi Parikh. VQA: visual question answering. In ICCV, 2015. 3", + "[5] Jinze Bai, Shuai Bai, Shusheng Yang, Shijie Wang, Sinan Tan, Peng Wang, Junyang Lin, Chang Zhou, and Jingren Zhou. Qwen-vl: A frontier large vision-language model with versatile abilities. arXiv preprint arXiv:2308.12966, 2023.7, 8", + "[6] Zechen Bai, Pichao Wang, Tianjun Xiao, Tong He, Zongbo Han, Zheng Zhang, and Mike Zheng Shou. Hallucination of multimodal large language models: A survey. arXiv preprint arXiv:2404.18930, 2024. 3", + "[7] Hangbo Bao, Wenhui Wang, Li Dong, Qiang Liu, Owais Khan Mohammed, Kriti Aggarwal, Subhojit Som, Songhao Piao, and Furu Wei. Vlmo: Unified vision-language pre-training with mixture-of-modality-experts. In NeurIPS, 2022. 5", + "[8] Karsten Behrendt, Libor Novak, and Rami Botros. A deep learning approach to traffic lights: Detection, tracking, and classification. In ICRA, 2017. 2", + "[9] Holger Caesar, Varun Bankiti, Alex H. Lang, Sourabh Vora, Venice Erin Liong, Qiang Xu, Anush Krishnan, Yu Pan, Giancarlo Baldan, and Oscar Beijbom. nuscenes: A multimodal dataset for autonomous driving. In CVPR, 2020. 2", + "[10] Xu Cao, Tong Zhou, Yunsheng Ma, Wenqian Ye, Can Cui, Kun Tang, Zhipeng Cao, Kaizhao Liang, Ziran Wang, James Rehg, and Chao Zheng. Maplm: A real-world large-scale vision-language dataset for map and traffic scene understanding. In CVPR, 2024. 3", + "[11] Li Chen, Chonghao Sima, Yang Li, Zehan Zheng, Jiajie Xu, Xiangwei Geng, Hongyang Li, Conghui He, Jianping Shi, Yu Qiao, and Junchi Yan. Persformer: 3d lane detection via perspective transformer and the openlane benchmark. In ECCV, 2022. 2", + "[12] Long Chen, Oleg Sinavski, Jan Hunermann, Alice Karnsund, Andrew James Willmott, Danny Birch, Daniel Maund, and Jamie Shotton. Driving with llms: Fusing object-level vector modality for explainable autonomous driving. In ICRA, 2024. 3", + "[13] Yaodong Cui, Shucheng Huang, Jiaming Zhong, Zhenan Liu, Yutong Wang, Chen Sun, Bai Li, Xiao Wang, and Amir Khajepour. Drivellm: Charting the path toward full autonomous driving with large language models. IEEE TIV, 2024. 3", + "[14] Jacob Devlin, Ming-Wei Chang, Kenton Lee, and Kristina Toutanova. BERT: pre-training of deep bidirectional transformers for language understanding. In NAACL, 2019. 5, 6, 7", + "[15] Alexey Dosovitskiy, Lucas Beyer, Alexander Kolesnikov, Dirk Weissenborn, Xiaohua Zhai, Thomas Unterthiner," + ], + "bbox": [ + 93, + 114, + 485, + 898 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "Mostafa Dehghani, Matthias Minderer, Georg Heigold, Sylvain Gelly, Jakob Uszkoreit, and Neil Houlsby. An image is worth 16x16 words: Transformers for image recognition at scale. In ICLR, 2021. 5, 7", + "[16] Gamal Elghazaly, Raphaël Frank, Scott Harvey, and Stefan Safko. High-definition maps: Comprehensive survey, challenges, and future perspectives. IEEE Open Journal of Intelligent Transportation Systems, 2023. 2, 4", + "[17] Andreas Fregin, Julian Müller, Ulrich Krebel, and Klaus Dietmayer. The driveu traffic light dataset: Introduction and comparison with existing datasets. In ICRA, 2018. 2", + "[18] Shuo Gu, Yigong Zhang, Jinhui Tang, Jian Yang, and Hui Kong. Road detection through CRF based lidar-camera fusion. In ICRA, 2019. 2", + "[19] Yunfei Guo, Wei Feng, Fei Yin, Tao Xue, Shuqi Mei, and Cheng-Lin Liu. Learning to understand traffic signs. In ACMMM, 2021. 2, 3", + "[20] Yunfei Guo, Fei Yin, Xiao-Hui Li, Xudong Yan, Tao Xue, Shuqi Mei, and Cheng-Lin Liu. Visual traffic knowledge graph generation from scene images. In ICCV, 2023. 2, 3", + "[21] Edward J. Hu, Yelong Shen, Phillip Wallis, Zeyuan Allen-Zhu, Yuanzhi Li, Shean Wang, Lu Wang, and Weizhu Chen. Lora: Low-rank adaptation of large language models. In The Tenth International Conference on Learning Representations, ICLR 2022, Virtual Event, April 25-29, 2022, 2022. 8", + "[22] Xinyu Huang, Peng Wang, Xinjing Cheng, Dingfu Zhou, Qichuan Geng, and Ruigang Yang. The apolloscope open dataset for autonomous driving and its application. IEEE TPAMI, 2020. 2", + "[23] Yupan Huang, Tengchao Lv, Lei Cui, Yutong Lu, and Furu Wei. Layoutmv3: Pre-training for document AI with unified text and image masking. In ACMMM, 2022. 3", + "[24] Chao Jia, Yinfei Yang, Ye Xia, Yi-Ting Chen, Zarana Parekh, Hieu Pham, Quoc V. Le, Yun-Hsuan Sung, Zhen Li, and Tom Duerig. Scaling up visual and vision-language representation learning with noisy text supervision. In ICML, 2021. 3", + "[25] Wonjae Kim, Bokyung Son, and Ildoo Kim. Vilt: Vision-and-language transformer without convolution or region supervision. In ICML, 2021. 3, 5", + "[26] Junnan Li, Ramprasaath R. Selvaraju, Akhilesh Gotmare, Shafiq R. Joty, Caiming Xiong, and Steven Chu-Hong Hoi. Align before fuse: Vision and language representation learning with momentum distillation. In NeurIPS, 2021. 3, 5, 7", + "[27] Junnan Li, Dongxu Li, Caiming Xiong, and Steven C. H. Hoi. BLIP: bootstrapping language-image pre-training for unified vision-language understanding and generation. In ICML, 2022. 5", + "[28] Junnan Li, Dongxu Li, Silvio Savarese, and Steven C. H. Hoi. BLIP-2: bootstrapping language-image pre-training with frozen image encoders and large language models. In ICML, 2023. 3", + "[29] Qi Li, Yue Wang, Yilun Wang, and Hang Zhao. Hdmapnet: An online HD map construction and evaluation framework. In ICRA, 2022. 2", + "[30] Benchcheng Liao, Shaoyu Chen, Xinggang Wang, Tianheng Cheng, Qian Zhang, Wenyu Liu, and Chang Huang. Maptr:" + ], + "bbox": [ + 516, + 92, + 906, + 901 + ], + "page_idx": 8 + }, + { + "type": "page_number", + "text": "6831", + "bbox": [ + 482, + 945, + 513, + 955 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "Structured modeling and learning for online vectorized HD map construction. In ICLR, 2023. 2", + "[31] Benchcheng Liao, Shaoyu Chen, Yunchi Zhang, Bo Jiang, Qian Zhang, Wenyu Liu, Chang Huang, and Xinggang Wang. Maptrv2: An end-to-end framework for online vectorized hd map construction. arXiv preprint arXiv:2308.05736, 2023. 3, 4, 5, 6", + "[32] Haotian Liu, Chunyuan Li, Yuheng Li, and Yong Jae Lee. Improved baselines with visual instruction tuning. arXiv preprint arXiv:2310.03744, 2023. 3", + "[33] Haotian Liu, Chunyuan Li, Qingyang Wu, and Yong Jae Lee. Visual instruction tuning. In NeurIPS, 2023.", + "[34] Haotian Liu, Chunyuan Li, Yuheng Li, Bo Li, Yuanhan Zhang, Sheng Shen, and Yong Jae Lee. Llava-last: Improved reasoning,OCR,and world knowledge.https://llava-v1.github.io/blog/2024-01-30-llava-last, 2024.3", + "[35] Rong Liu, Jinling Wang, and Bingqi Zhang. High definition map for automated driving: Overview and analysis. Journal of Navigation, 2020. 2, 4", + "[36] Yicheng Liu, Tianyuan Yuan, Yue Wang, Yilun Wang, and Hang Zhao. Vectormapnet: End-to-end vectorized HD map learning. In ICML, 2023. 2", + "[37] Jiageng Mao, Yuxi Qian, Hang Zhao, and Yue Wang. Gptdriver: Learning to drive with GPT. In NeurIPS Workshop, 2023. 3", + "[38] Ana-Maria Marcu, Long Chen, Jan Hunermann, Alice Karnsund, Benoit Hanotte, Prajwal Chidananda, Saurabh Nair, Vijay Badrinarayanan, Alex Kendall, Jamie Shotton, and Oleg Sinavski. Lingoqa: Video question answering for autonomous driving. arXiv preprint arXiv:2312.14115, 2023. 3", + "[39] OpenAI. GPT-4 technical report. arXiv preprint arXiv:2303.08774, 2024. 7, 8", + "[40] Jonah Philion and Sanja Fidler. Lift, splat, shoot: Encoding images from arbitrary camera rigs by implicitly unprojecting to 3d. In ECCV, 2020. 2", + "[41] Tianwen Qian, Jingjing Chen, Linhai Zhuo, Yang Jiao, and Yu-Gang Jiang. Nuscenes-qa: A multi-modal visual question answering benchmark for autonomous driving scenario. In AAAI, 2024. 3", + "[42] Alec Radford, Jong Wook Kim, Chris Hallacy, Aditya Ramesh, Gabriel Goh, Sandhini Agarwal, Girish Sastry, Amanda Askell, Pamela Mishkin, Jack Clark, Gretchen Krueger, and Ilya Sutskever. Learning transferable visual models from natural language supervision. In ICML, 2021. 3, 5", + "[43] Enna Sachdeva, Nakul Agarwal, Suhas Chundi, Sean Roelofs, Jiachen Li, Mykel J. Kochenderfer, Chiho Choi, and Behzad Dariush. Rank2tell: A multimodal driving dataset for joint importance ranking and reasoning. In WACV, 2024. 3", + "[44] Chonghao Sima, Katrin Renz, Kashyap Chitta, Li Chen, Hanxue Zhang, Chengen Xie, Ping Luo, Andreas Geiger, and Hongyang Li. Drivelm: Driving with graph visual question answering. arXiv preprint arXiv:2312.14150, 2023. 3", + "[45] Chonghao Sima, Katrin Renz, Kashyap Chitta, Li Chen, Hanxue Zhang, Chengen Xie, Jens Beiwenger, Ping Luo, Andreas Geiger, and Hongyang Li. Drivelm: Driving with graph visual question answering. In ECCV, 2024. 3" + ], + "bbox": [ + 91, + 90, + 485, + 900 + ], + "page_idx": 9 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[46] Johannes Stallkamp, Marc Schlipsing, Jan Salmen, and Christian Igel. Man vs. computer: Benchmarking machine learning algorithms for traffic sign recognition. *Neural Networks*, 2012. 2", + "[47] Pei Sun, Henrik Kretzschmar, Xerxes Dotiwalla, Aurelien Chouard, Vijaysai Patnaik, Paul Tsui, James Guo, Yin Zhou, Yuning Chai, Benjamin Caine, Vijay Vasudevan, Wei Han, Jiquan Ngiam, Hang Zhao, Aleksei Timofeev, Scott Ettinger, Maxim Krivokon, Amy Gao, Aditya Joshi, Yu Zhang, Jonathon Shlens, Zhifeng Chen, and Dragomir Anguelov. Scalability in perception for autonomous driving: Waymo open dataset. In CVPR, 2020. 2", + "[48] Matthew Tancik, Pratul P. Srinivasan, Ben Mildenhall, Sara Fridovich-Keil, Nithin Raghavan, Utkarsh Singhal, Ravi Ramamoorthi, Jonathan T. Barron, and Ren Ng. Fourier features let networks learn high frequency functions in low dimensional domains. In NeurIPS, 2020. 5", + "[49] Gemini Team, Rohan Anil, Sebastian Borgeaud, Yonghui Wu, Jean-Baptiste Alayrac, Jiahui Yu, Radu Soricut, Johan Schalkwyk, Andrew M. Dai, Anja Hauth, and et al. Gemini: A family of highly capable multimodal models. arXiv preprint arXiv:2312.11805, 2024. 7, 8", + "[50] Hugo Touvron, Matthieu Cord, Matthijs Douze, Francisco Massa, Alexandre Sablayrolles, and Hervé Jégou. Training data-efficient image transformers & distillation through attention. In ICML, 2021. 7", + "[51] Ashish Vaswani, Noam Shazeer, Niki Parmar, Jakob Uszkoreit, Llion Jones, Aidan N. Gomez, Lukasz Kaiser, and Illia Polosukhin. Attention is all you need. In NeurIPS, 2017. 3, 5", + "[52] Huijie Wang, Tianyu Li, Yang Li, Li Chen, Chonghao Sima, Zhenbo Liu, Bangjun Wang, Peijin Jia, Yuting Wang, Shengyin Jiang, Feng Wen, Hang Xu, Ping Luo, Junchi Yan, Wei Zhang, and Hongyang Li. Openlane-v2: A topology reasoning benchmark for unified 3d HD mapping. In NeurIPS, 2023. 2, 3", + "[53] Benjamin Wilson, William Qi, Tanmay Agarwal, John Lambert, Jagjeet Singh, Siddhesh Khandelwal, Bowen Pan, Ratnesh Kumar, Andrew Hartnett, Jhony Kaesemodel Pontes, Deva Ramanan, Peter Carr, and James Hays. Argoverse 2: Next generation datasets for self-driving perception and forecasting. In NeurIPS, 2021. 2", + "[54] Dongming Wu, Jiahao Chang, Fan Jia, Yingfei Liu, Tiancai Wang, and Jianbing Shen. Topomlp: A simple yet strong pipeline for driving topology reasoning. In ICLR, 2024. 2", + "[55] Yiheng Xu, Minghao Li, Lei Cui, Shaohan Huang, Furu Wei, and Ming Zhou. Layoutm: Pre-training of text and layout for document image understanding. In KDD, 2020. 3", + "[56] Yang Xu, Yiheng Xu, Tengchao Lv, Lei Cui, Furu Wei, Guoxin Wang, Yijuan Lu, Dinei A. F. Floréncio, Cha Zhang, Wanxiang Che, Min Zhang, and Lidong Zhou. Layoutmv2: Multi-modal pre-training for visually-rich document understanding. In ACL, 2021. 3", + "[57] Zhenhua Xu, Yujia Zhang, Enze Xie, Zhen Zhao, Yong Guo, Kwan-Yee K. Wong, Zhenguo Li, and Hengshuang Zhao. Drivegpt4: Interpretable end-to-end autonomous driving via large language model. IEEE Robotics Autom. Lett., 2024. 3", + "[58] Fisher Yu, Haofeng Chen, Xin Wang, Wenqi Xian, Yingying Chen, Fangchen Liu, Vashisht Madhavan, and Trevor Dar" + ], + "bbox": [ + 516, + 92, + 906, + 898 + ], + "page_idx": 9 + }, + { + "type": "page_number", + "text": "6832", + "bbox": [ + 482, + 944, + 514, + 955 + ], + "page_idx": 9 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "rell. BDD100K: A diverse driving dataset for heterogeneous multitask learning. In CVPR, 2020. 2", + "[59] Jiahui Yu, Zirui Wang, Vijay Vasudevan, Legg Yeung, Mojtaba Seyedhosseini, and Yonghui Wu. Coca: Contrastive captioners are image-text foundation models. TMLR, 2022. 3", + "[60] Zhe Zhu, Dun Liang, Song-Hai Zhang, Xiaolei Huang, Baoli Li, and Shi-Min Hu. Traffic-sign detection and classification in the wild. In CVPR, 2016. 2" + ], + "bbox": [ + 91, + 90, + 483, + 204 + ], + "page_idx": 10 + }, + { + "type": "page_number", + "text": "6833", + "bbox": [ + 482, + 945, + 514, + 955 + ], + "page_idx": 10 + } +] \ No newline at end of file diff --git a/2025/Driving by the Rules_ A Benchmark for Integrating Traffic Sign Regulations into Vectorized HD Map/cbe7a2b6-7e9c-4185-bab6-ca66fed01b85_model.json b/2025/Driving by the Rules_ A Benchmark for Integrating Traffic Sign Regulations into Vectorized HD Map/cbe7a2b6-7e9c-4185-bab6-ca66fed01b85_model.json new file mode 100644 index 0000000000000000000000000000000000000000..c0d66d56f05d97972251070f931aad932a1121ad --- /dev/null +++ b/2025/Driving by the Rules_ A Benchmark for Integrating Traffic Sign Regulations into Vectorized HD Map/cbe7a2b6-7e9c-4185-bab6-ca66fed01b85_model.json @@ -0,0 +1,2158 @@ +[ + [ + { + "type": "header", + "bbox": [ + 0.107, + 0.003, + 0.182, + 0.043 + ], + "angle": 0, + "content": "CVF" + }, + { + "type": "header", + "bbox": [ + 0.237, + 0.001, + 0.812, + 0.047 + ], + "angle": 0, + "content": "This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore." + }, + { + "type": "title", + "bbox": [ + 0.093, + 0.13, + 0.908, + 0.178 + ], + "angle": 0, + "content": "Driving by the Rules: A Benchmark for Integrating Traffic Sign Regulations into Vectorized HD Map" + }, + { + "type": "text", + "bbox": [ + 0.135, + 0.203, + 0.871, + 0.276 + ], + "angle": 0, + "content": "Xinyuan Chang \\(^{1*}\\) Maixuan Xue \\(^{2*†}\\) Xinran Liu \\(^{1}\\) Zheng Pan \\(^{1}\\) Xing Wei \\(^{2‡}\\) \n\\(^{1}\\)Amap, Alibaba Group \\(^{2}\\)Xi’an Jiaotong University \n{changxinyuan.cxy, tom.lxr, panzheng.pan}@alibaba-inc.com \nxmx0809@stu.xjtu.edu.cn, weixing@mail.xjtu.edu.cn" + }, + { + "type": "image", + "bbox": [ + 0.141, + 0.332, + 0.687, + 0.543 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.702, + 0.333, + 0.86, + 0.394 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.736, + 0.395, + 0.829, + 0.406 + ], + "angle": 0, + "content": "Geometric layer" + }, + { + "type": "image", + "bbox": [ + 0.703, + 0.408, + 0.859, + 0.468 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.731, + 0.468, + 0.837, + 0.479 + ], + "angle": 0, + "content": "Connectivity layer" + }, + { + "type": "image", + "bbox": [ + 0.703, + 0.48, + 0.859, + 0.541 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.719, + 0.542, + 0.851, + 0.554 + ], + "angle": 0, + "content": "Traffic regulation layer" + }, + { + "type": "image_caption", + "bbox": [ + 0.089, + 0.568, + 0.908, + 0.626 + ], + "angle": 0, + "content": "Figure 1. MapDR Overview and Motivation. The image on the left depicts the comprehensive results of HD map construction for an intersection scene, whereas the image on the right illustrates the outcomes after partitioning it into three layers. Existing online mapping methods primarily emphasize the construction of the geometric and connectivity layers, neglecting the traffic regulation layer. However, precise comprehension of traffic signs and their correlation with lanes is vital for ensuring the safety of autonomous driving." + }, + { + "type": "title", + "bbox": [ + 0.248, + 0.645, + 0.327, + 0.661 + ], + "angle": 0, + "content": "Abstract" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.681, + 0.486, + 0.848 + ], + "angle": 0, + "content": "Ensuring adherence to traffic sign regulations is essential for both human and autonomous vehicle navigation. While current online mapping solutions often prioritize the construction of the geometric and connectivity layers of HD maps, overlooking the construction of the traffic regulation layer within HD maps. Addressing this gap, we introduce MapDR, a novel dataset designed for the extraction of Driving Rules from traffic signs and their association with vectorized, locally perceived HD Maps. MapDR features over 10,000 annotated video clips that capture the intricate correlation between traffic sign regulations and" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.646, + 0.909, + 0.768 + ], + "angle": 0, + "content": "lanes. Built upon this benchmark and the newly defined task of integrating traffic regulations into online HD maps, we provide modular and end-to-end solutions: VLE-MEE and RuleVLM, offering a strong baseline for advancing autonomous driving technology. It fills a critical gap in the integration of traffic sign rules, contributing to the development of reliable autonomous driving systems. Code is available at https://github.com/MIV-XJTU/MapDR." + }, + { + "type": "title", + "bbox": [ + 0.514, + 0.8, + 0.646, + 0.815 + ], + "angle": 0, + "content": "1. Introduction" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.827, + 0.909, + 0.904 + ], + "angle": 0, + "content": "The rise of autonomous vehicles and intelligent transportation systems has emphasized the importance of accurate, reliable navigational data. High-Definition (HD) maps, with their detailed representations of road elements, have become essential to supporting these systems. An HD map typically" + }, + { + "type": "page_footnote", + "bbox": [ + 0.109, + 0.865, + 0.221, + 0.877 + ], + "angle": 0, + "content": "*Equal contribution." + }, + { + "type": "page_footnote", + "bbox": [ + 0.111, + 0.878, + 0.423, + 0.89 + ], + "angle": 0, + "content": "\\(^{\\dagger}\\)Work done during the internship at Amap, Alibaba Group." + }, + { + "type": "page_footnote", + "bbox": [ + 0.111, + 0.89, + 0.236, + 0.902 + ], + "angle": 0, + "content": "\\(\\ddagger\\) Corresponding author." + }, + { + "type": "list", + "bbox": [ + 0.109, + 0.865, + 0.423, + 0.902 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.516, + 0.957 + ], + "angle": 0, + "content": "6823" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.09, + 0.092, + 0.483, + 0.212 + ], + "angle": 0, + "content": "consists of three primary layers: the geometric layer, connectivity layer, and traffic regulation layer [16, 35], as shown in Fig. 1. The geometric layer provides vector information, such as dividers and lane centerlines; the connectivity layer defines lane relationships to assist in path planning; and the traffic regulation layer encodes rule information (e.g., HOV lanes, bus lanes, speed-limit zones) associated with lanes, ensuring compliant driving." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.214, + 0.484, + 0.35 + ], + "angle": 0, + "content": "A limitation of conventional HD maps is their inability to support rapid, real-time updates, making the construction of online HD maps an emerging trend. Existing methods, such as MapTR [30] and TopoMLP [54], address only the geometric and connectivity layers, overlooking the traffic regulation layer. As a result, autonomous driving systems currently still rely on offline maps to obtain traffic regulations, an approach that runs counter to the trend of online HD map construction." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.353, + 0.484, + 0.565 + ], + "angle": 0, + "content": "Traffic signs serve as a \"visual language\" on the road, playing a critical role in defining traffic regulations. Inspired by human driving processes, the creation of traffic regulations can be divided into two stages: 1) understanding the rules conveyed by traffic signs, and 2) determining the lanes to which these rules apply. Although OpenLane-V2 [52] has made strides in associating signs with lanes in online mapping, it has two primary limitations. First, it considers only directional signs, neglecting other types of regulations. Second, its label annotations are limited to sign categories, falling short of the detailed rule descriptions required for autonomous driving. A more structured description, in alignment with HD map standards, is essential for supporting comprehensive autonomous driving applications [16, 35]." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.567, + 0.484, + 0.746 + ], + "angle": 0, + "content": "To bridge this gap, we introduce MapDR, the first dataset specifically designed for the challenging task of integrating traffic regulations into existing online HD maps. MapDR focuses on complex traffic scenes with diverse signage, collecting over 10,000 video clips across three representative cities in China, with each clip containing at least one traffic sign. In addition to vectorized representations of dividers, boundaries, and centerlines, MapDR provides structural annotations of traffic regulations and their associations with lanes, offering a comprehensive foundation for advancing online HD maps. More details of MapDR can be found in Sec. 4." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.75, + 0.484, + 0.901 + ], + "angle": 0, + "content": "Based on MapDR, in addition to the end-to-end task definition, we also divide this task into two innovative subtasks aimed at bolstering research in this domain: 1) rule extraction from traffic signs, and 2) rule-lane correspondence reasoning. We introduce Vision-Language Encoder (VLE) and Map Element Encoder (MEE) to handle the interaction of multimodal data, encompassing images, texts, and vectors. Through the concatenation of these two models, integrated traffic regulations can be obtained. Additionally, inspired by recent advancements in Multimodal Large Lan" + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.089, + 0.907, + 0.187 + ], + "angle": 0, + "content": "Table 1. Comparison of the existing datasets. \"Sign\" and \"Lane\" denote whether the dataset focuses on traffic signs and lanes. Only those annotated with formatted (\"Fmt.\") rules and the correspondence (\"Corr.\") between rules and lanes can form driving rules. \"Clip\" represents whether the data is organized in the form of video clips. \"* denotes that these samples are not newly collected and are built upon the previous dataset." + }, + { + "type": "table", + "bbox": [ + 0.517, + 0.198, + 0.905, + 0.304 + ], + "angle": 0, + "content": "
DatasetSignLaneDriving RulesNumber of SamplesYear
Fmt.Corr.ImageClipRegion
nuScenes [9]✓1400K1KWorldwide2019
Argoverse2 [53]✓2100K1KUSA2021
CTSU [19]✓5K/China2021
OpenLane [11]✓200K*1K*Worldwide2022
RS10K [20]✓✓10K/China2023
OpenLane-V2 [52]✓✓✓466K*2K*Worldwide2023
Waymo [47]✓✓390K2KUSA2024
MapDR(ours)✓✓✓✓400K10KChina2024
" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.332, + 0.907, + 0.422 + ], + "angle": 0, + "content": "guage Models (MLLMs), we also explore an end-to-end MLLM solution called RuleVLM to solve this task holistically. These approaches establish robust baselines and aim to inspire future research in this area. For detailed descriptions of the proposed tasks and metrics, please refer to Sec. 3. For approaches, please refer to Secs. 5 and 6." + }, + { + "type": "text", + "bbox": [ + 0.534, + 0.424, + 0.825, + 0.437 + ], + "angle": 0, + "content": "To sum up, our contributions are as follows:" + }, + { + "type": "text", + "bbox": [ + 0.547, + 0.442, + 0.907, + 0.517 + ], + "angle": 0, + "content": "- For the first time, We introduce the novel task of extracting lane-level rules from traffic signs and integrating them into vectorized HD maps, alongside the MapDR dataset and evaluation metrics for benchmarking." + }, + { + "type": "text", + "bbox": [ + 0.547, + 0.518, + 0.907, + 0.606 + ], + "angle": 0, + "content": "- MapDR comprises over 10,000 video clips and 400,000 front-view images captured across multiple cities in China, covering diverse traffic conditions and including more than 18,000 annotated lane-level rules. All data are newly collected and meticulously validated." + }, + { + "type": "text", + "bbox": [ + 0.547, + 0.608, + 0.905, + 0.668 + ], + "angle": 0, + "content": "- We propose two modeling approaches, modular and end-to-end, to address the integration of lane-level rules into vectorized HD maps, providing effective baselines for future research." + }, + { + "type": "list", + "bbox": [ + 0.547, + 0.442, + 0.907, + 0.668 + ], + "angle": 0, + "content": null + }, + { + "type": "title", + "bbox": [ + 0.514, + 0.687, + 0.655, + 0.702 + ], + "angle": 0, + "content": "2. Related Work" + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.712, + 0.723, + 0.728 + ], + "angle": 0, + "content": "2.1. HD Map Construction" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.735, + 0.907, + 0.901 + ], + "angle": 0, + "content": "HD maps construction have seen significant advancements, with a focus on traffic element perception, including lane detection and traffic sign recognition [8, 9, 17, 18, 22, 46, 53, 58, 60]. The shift towards BEV perception and vectorization for end-to-end HD maps construction has gained traction [9, 11, 53]. Notable works include HDMapNet, which aggregates semantic segmentation results [29], LSS [40] estimates depth to transfer image features to BEV features, while VectorMapNet [36] is the first end-to-end framework for sequential vector point prediction to generate HD maps without post-processing. MapTR [30] and its enhanced" + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.516, + 0.956 + ], + "angle": 0, + "content": "6824" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.134, + 0.089, + 0.861, + 0.23 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.089, + 0.242, + 0.908, + 0.272 + ], + "angle": 0, + "content": "Figure 2. Overview of the sub-tasks. Step 1 ~ Step 4 shows a case of driving by the rules. Step 2 and Step 3 demonstrates the specific role of two sub-tasks, respectively." + }, + { + "type": "text", + "bbox": [ + 0.089, + 0.297, + 0.486, + 0.373 + ], + "angle": 0, + "content": "version, MapTRv2 [31], introduced a unified permutation-equivalent modeling approach and extended it to a general framework supporting centerline learning and 3D map construction. However, these efforts have largely overlooked the integration of traffic sign rules into HD maps." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.381, + 0.342, + 0.396 + ], + "angle": 0, + "content": "2.2. Traffic Element Association" + }, + { + "type": "text", + "bbox": [ + 0.093, + 0.402, + 0.485, + 0.705 + ], + "angle": 0, + "content": "Traffic element association aims to link elements like traffic signs with lanes. As demonstrated in Tab. 1, CTSU has initiated internal elements association to describe traffic sign in \\(\\{key: value\\}\\) form, however lacking both generalization of driving rules from description and lane association [19]. VTKGG [20] propose to utilize a graph model for connectivity but also lacks structured expression of driving rules for motion planning and requires complex integration into HD maps, which is typically expressed in the BEV space. OpenLane-V2 [52] advances traffic sign and lane association in BEV space, but is constrained by single-label classification for traffic sign rather than structured descriptions for fine-grained driving rules. This makes it insufficient for signs with multiple rules, which are complex but common in real scenarios. Recent MLLM-based benchmarks [10, 38, 41, 43, 44] for autonomous driving, such as MAPLM [10], prioritize end-to-end motion planning over precise rule extraction from traffic sign, lacking evaluation for rule reasoning. MapDR addresses this gap by focusing on traffic sign rule extraction and lane association." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.713, + 0.32, + 0.729 + ], + "angle": 0, + "content": "2.3. Vision-Language Models" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.735, + 0.486, + 0.901 + ], + "angle": 0, + "content": "Vision-Language Models (VLMs) facilitates multimodal applications by learning joint representations of vision and language data. Visual Question Answer (VQA) tasks provide answers to image-related questions [4], while Visual Information Extraction (VIE) tasks extract structured information from visual and textual data [4, 23, 55, 56]. In Autonomous Driving (AD), VLMs are increasingly used for comprehensive traffic scene understanding and decision-making. The field has seen various approaches, including using transformers [51] for joint encoding [23, 25], excelling at multimodal information interaction, and independent en" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.297, + 0.909, + 0.449 + ], + "angle": 0, + "content": "coders for different modalities [24, 42] that are proficient in multimodal retrieval. Cross-modal representation methods [26, 59] combine these advantages, and the latest LLM-based research [28, 32-34] has achieved state-of-the-art results in various multimodal tasks. Nowadays, an increasing number of methods are leveraging LLMs to achieve impressive results [12, 37, 45, 57], with works like DriveLLM [13] showing significant potential in AD. However, addressing hallucination [6] remains the most crucial aspect for LLM-based approaches." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.468, + 0.907, + 0.504 + ], + "angle": 0, + "content": "3. Task Definition : Integrating Traffic Sign Regulations into HD Maps" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.513, + 0.909, + 0.801 + ], + "angle": 0, + "content": "The ability to discern rules from traffic signs and to associate them with specific lanes is pivotal for autonomous navigation. Our proposed task focuses on extracting lane-level driving rules denoted as \\( R = \\{r_i\\}_{i=1}^m \\) and reasoning the correspondence between these rules and the centerlines \\( L = \\{l_i\\}_{i=1}^k \\) in the local vectorized HD map, where \\( m \\) is the number of rules and \\( k \\) is the number of centerlines. Each rule \\( r_i \\) is a set of pre-defined properties in \\( \\{key : value\\} \\) pairs, as shown in Fig. 3. The final correspondence forms a bipartite graph \\( G = (R \\cup L, E) \\), where \\( E \\subseteq \\{0,1\\}^{m \\times k} \\) and the element \\( E_{ij} \\) in the \\( i \\)-th row and \\( j \\)-th column of matrix \\( E \\) represents the corresponding status between \\( r_i \\) and \\( l_j \\). In this case, we assume the entire approach as \\( \\mathcal{F} \\) and the overall process can be formulated as \\( G = \\mathcal{F}(X, L) \\), where \\( X = \\{x_i\\}_{i=1}^n \\) represents a series of image sequences and \\( n \\) is the number of frames. To independently investigate the capabilities of rule extraction and correspondence reasoning within the overall task, we further divided the overall task into two sub-tasks, defined as follows." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.816, + 0.81, + 0.833 + ], + "angle": 0, + "content": "3.1. Rule Extraction from Traffic Sign" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.84, + 0.91, + 0.903 + ], + "angle": 0, + "content": "As shown in Step 2 of Fig. 2, this sub-task involves extracting multiple rules \\( R = \\{r_i\\}_{i=1}^m \\) from a series of image sequences \\( X = \\{x_i\\}_{i=1}^n \\). The rule extraction model, denoted as \\( \\mathcal{M} \\), can be expressed as \\( R = \\mathcal{M}(X) \\)." + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.516, + 0.957 + ], + "angle": 0, + "content": "6825" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.114, + 0.089, + 0.46, + 0.196 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.09, + 0.208, + 0.486, + 0.264 + ], + "angle": 0, + "content": "Figure 3. Visualization of dataset demo. Multiple lane-level rules of a single traffic sign are annotated in \\(\\{key: value\\}\\) format. Directed lines indicate the correspondence between rules and particular centerlines." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.291, + 0.423, + 0.308 + ], + "angle": 0, + "content": "3.2. Rule-Lane Correspondence Reasoning" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.314, + 0.484, + 0.42 + ], + "angle": 0, + "content": "As shown in Step 3 of Fig. 2, the reasoning process establishes the correspondence between centerlines \\( L = \\{l_i\\}_{i=1}^k \\) and all rules \\( R \\). We denote the correspondence reasoning model as \\( \\mathcal{T} \\), and this process can be described as \\( E = \\mathcal{T}(R, L) \\). The final reasoning result forms a bipartite graph \\( G = (R \\cup L, E) \\), which means corresponding relationships only exist between rules and centerlines." + }, + { + "type": "title", + "bbox": [ + 0.09, + 0.434, + 0.414, + 0.451 + ], + "angle": 0, + "content": "4. The MapDR Dataset & Benchmark" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.46, + 0.484, + 0.58 + ], + "angle": 0, + "content": "We introduce the MapDR dataset, meticulously annotated with traffic sign regulations and their correspondences to lanes, as shown in Fig. 3. The dataset encompasses a diverse range of scenarios, weather conditions, and traffic situations, with over 10,000 traffic scene segments, 18,000 driving rules, and 400,000 images. Traffic signs typically have varying textual descriptions, text layouts, and positions on the road, which add complexity to the task." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.582, + 0.484, + 0.763 + ], + "angle": 0, + "content": "The majority of the data originates from Beijing and Shanghai, with additional scenes from Guangzhou. Fig. 4 illustrates the geographic spread and variety of traffic signs. The dataset reflects a natural long-tail distribution, with a prevalence of bus and direction lanes and a scarcity of tidalflow lanes. We primarily focus on traffic signs that indicate lane-level rules, collected from cities with the most complex and diverse traffic scenarios in China, ensuring realistic and representative data. All images have undergone privacy and safety processing to obscure license plates and faces. More comprehensive statistic of dataset and case demonstrations can be found in the supplement." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.773, + 0.315, + 0.789 + ], + "angle": 0, + "content": "4.1. Raw Data & Annotation" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.795, + 0.484, + 0.901 + ], + "angle": 0, + "content": "Raw Data. MapDR is collected from real-world traffic scenes, each scene segment (video clip) captures front-view images within a \\(100m \\times 100m\\) area centered on the traffic sign, with a consistent resolution of \\(1920 \\times 1240\\). Each clip contains 30 to 60 frames, captured at 1 frame every 2 meters, ensuring consistent spatial intervals. Each video clip focuses on a single traffic sign and provides its position in 3D space." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.092, + 0.905, + 0.136 + ], + "angle": 0, + "content": "Camera intrinsics and poses are provided for each frame, and coordinates for each clip are transformed to distinct ENU systems." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.138, + 0.907, + 0.29 + ], + "angle": 0, + "content": "All vectors of local map in the target area are provided as 3D point lists, generated using our algorithm similar to MapTRv2 [31]. Each lane vector has a type, such as divider, centerline, crosswalk, or boundary. For example, the centerline is defined as \\( L = \\{l_i\\}_{i=1}^k \\), where each vector \\( l_i \\) is composed of multiple 3D points \\( l_i = [p_1, \\ldots, p_n] \\), and \\( p_j = (x_j, y_j, z_j) \\) represents the coordinates of the current point. The pipeline of dataset production is illustrated in Fig. 5, and detailed data acquisition and annotation procedures can be found in the supplement material." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.314, + 0.909, + 0.51 + ], + "angle": 0, + "content": "**Formatted Rules.** Each video clip may contain multiple lane-level rules, denoted as \\( R \\). Each rule is expressed by symbols and text on the sign, requiring interpretation. We reference existing HD map design and data specifications [1, 16, 35] to define lane-level driving rules in the JSON format. As shown in Fig. 3, each rule \\( r_i \\) comprises 8 predefined properties in the form of {key : value} pairs. This definition schema encompasses most scenarios and supports integration into existing autonomous driving systems. We enclose the symbols and texts denoting each distinct rule on traffic signs with polygons and project them into 3D space as \\( P_i = [p_1, \\ldots, p_n] \\), where \\( n \\) varies. Researchers can optionally use this information to facilitate rule extraction." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.534, + 0.909, + 0.643 + ], + "angle": 0, + "content": "Correspondence between Rules & Lanes. Based on formatted rules \\( R \\) and centerlines \\( L \\), corresponding centerlines of each rule are annotated as shown in Fig. 3. Therefore correspondence between rules and centerlines can be formed as a bipartite graph \\( G = (R \\cup L, E) \\), where \\( E \\subseteq \\{0,1\\}^{|R| \\times |L|} \\) and the positive edges only exist between \\( R \\) and \\( L \\) as demonstrated in Sec. 3.2. Specifically, \\( E_{ij} = 1 \\) represents that" + }, + { + "type": "image", + "bbox": [ + 0.536, + 0.667, + 0.887, + 0.832 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.512, + 0.842, + 0.909, + 0.898 + ], + "angle": 0, + "content": "Figure 4. Geographic location distribution of the collected traffic signs and proportions of various lane types represented in all signs. The geographic distribution is visualized based on OpenStreetMap [2]." + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.516, + 0.956 + ], + "angle": 0, + "content": "6826" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.111, + 0.088, + 0.465, + 0.155 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.09, + 0.166, + 0.483, + 0.237 + ], + "angle": 0, + "content": "Figure 5. Pipeline of dataset production. The location of traffic signs are sampled from existing database then front-view images of each sign are newly collected. Vectorized map is processed in cloud server. Finally formatted rules and correspondence between rules and centerlines are annotated and organized as MapDR." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.263, + 0.484, + 0.294 + ], + "angle": 0, + "content": "vehicle driving on the lane with centerline \\( l_{j} \\) should follow the driving rule \\( r_{i} \\)." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.304, + 0.276, + 0.318 + ], + "angle": 0, + "content": "4.2. Evaluation Metrics" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.326, + 0.486, + 0.372 + ], + "angle": 0, + "content": "Based on the task defined in Sec. 3, we propose evaluation metrics for entire task as well as for two sub-tasks separately. The \\(F_{1}\\) score is adopted as the benchmark for methods rank." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.391, + 0.484, + 0.468 + ], + "angle": 0, + "content": "Rule Extraction (R.E.). Given the ground truth \\( R \\) and predicted rules \\( \\hat{R} \\), we propose to calculate the Precision \\( (P_{R,E.}) \\) and Recall \\( (R_{R,E.}) \\) to evaluate the capability of rules extraction as defined in Eq. (1), where \\( \\hat{r_i} = r_i \\) represents all the properties are predicted correctly." + }, + { + "type": "equation", + "bbox": [ + 0.15, + 0.479, + 0.484, + 0.515 + ], + "angle": 0, + "content": "\\[\nP _ {R. E.} = \\frac {\\left| \\hat {R} \\cap R \\right|}{\\left| \\hat {R} \\right|} \\quad R _ {R. E.} = \\frac {\\left| \\hat {R} \\cap R \\right|}{\\left| R \\right|} \\tag {1}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.533, + 0.485, + 0.625 + ], + "angle": 0, + "content": "Correspondence Reasoning (C.R.). Given the ground truth of correspondence bipartite graph \\( G = (R \\cup L, E) \\) and predicted graph \\( \\hat{G} = (R \\cup L, \\hat{E}) \\), we propose to calculate Precision (\\( P_{C.R.} \\)) and Recall (\\( R_{C.R.} \\)) of edge set \\( E \\) to evaluate the capability of correspondence reasoning individually. Metrics are defined as Eq. (2)." + }, + { + "type": "equation", + "bbox": [ + 0.15, + 0.635, + 0.483, + 0.672 + ], + "angle": 0, + "content": "\\[\nP _ {C. R.} = \\frac {\\left| \\hat {E} \\cap E \\right|}{\\left| \\hat {E} \\right|} \\quad R _ {C. R.} = \\frac {\\left| \\hat {E} \\cap E \\right|}{\\left| E \\right|} \\tag {2}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.689, + 0.486, + 0.901 + ], + "angle": 0, + "content": "Overall. As defined in Sec. 3, the final predicted results can be obtained through an end-to-end prediction or a combination of two sequential sub-tasks. Given the predicted rules, correspondence should be reasoned between \\(\\hat{R}\\) and \\(L\\) which means the prediction of entire task is \\(\\hat{G} = (\\hat{R} \\cup L, \\hat{E})\\) and the ground truth is consistent \\(G = (R \\cup L, E)\\). We evaluate Precision \\((P_{all})\\) and Recall \\((R_{all})\\) using the subgraph \\(G^{s}\\), where \\(G^{s} = \\{g_{ij}^{s}\\}_{i=1,j=1}^{m,k}\\), \\(g_{ij}^{s} = (\\{r_{i}, l_{j}\\}, e_{ij})\\). In set of sub-graph \\(G^{s}\\), \\(m\\) is the number of rules, and \\(k\\) is the number of centerlines. Furthermore, we propose the \\(F_{1}\\) for the final benchmark ranking. Metrics are defined in Eq. (3) and Eq. (4). The \\(F_{1}\\) is derived from the calculations of \\(P_{all}\\) and \\(R_{all}\\). We provide an example of calculating the Overall metrics in the supplement material." + }, + { + "type": "equation", + "bbox": [ + 0.555, + 0.115, + 0.907, + 0.153 + ], + "angle": 0, + "content": "\\[\nP _ {a l l} = \\frac {\\left| \\hat {G} ^ {s} \\cap G ^ {s} \\right|}{\\left| \\hat {G} ^ {s} \\right|} \\quad R _ {a l l} = \\frac {\\left| \\hat {G} ^ {s} \\cap G ^ {s} \\right|}{\\left| G ^ {s} \\right|} \\tag {3}\n\\]" + }, + { + "type": "equation", + "bbox": [ + 0.634, + 0.164, + 0.907, + 0.197 + ], + "angle": 0, + "content": "\\[\nF _ {1} = 2 \\times \\frac {P _ {\\text {a l l}} \\times R _ {\\text {a l l}}}{P _ {\\text {a l l}} + R _ {\\text {a l l}}} \\tag {4}\n\\]" + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.207, + 0.704, + 0.224 + ], + "angle": 0, + "content": "5. Modular Approach" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.231, + 0.908, + 0.382 + ], + "angle": 0, + "content": "In order to conduct a comprehensive analysis of the entire task, a modular approach is implemented in this section to address the two sub-tasks. It is demonstrated that these interconnected modules can effectively achieve the ultimate objective of integrating driving rules into the HD map. To address the multimodal information interaction encompassing images, texts, and vectors, a Vision-Language Encoder (VLE) and a Map Element Encoder (MEE) are developed. The following sections expound upon their structures, applications, and the experimental results on MapDR." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.392, + 0.649, + 0.406 + ], + "angle": 0, + "content": "5.1. Architecture" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.414, + 0.909, + 0.776 + ], + "angle": 0, + "content": "Vision-Language Encoder. Inspired by vision-language frameworks [7, 25-27, 42], we designed a vision-language fusion model named VLE, following [26]. As shown in Fig. 6, VLE uses ViT-b16 [15] as the vision encoder, with the text encoder and multimodal fusion encoder each consisting of \\( L \\) transformer layers [51]. Each layer of the fusion encoder includes a cross-attention module for fusion [26]. In practice, distinct rules are represented by varying numbers of symbols and texts, as shown in the OCR results in Fig. 6. To address the challenge of representing variable-length input as fixed-length features, we introduce a [CLS] token for an entire rule and several [STC] tokens for sentence-level representation. The specific usage of these tokens is detailed in Sec. 5.2. Furthermore, we incorporate inter-instance and intra-instance attention mechanisms [31] to enhance model performance by capturing interactions and independence between and within sentences. In addition to content, layout captures the relative positions of symbols and texts, offering important semantic meaning. To leverage this, we encode the layout using the method from [48] and the relative positions of characters as position embedding following [14]. As shown in VLE in Fig. 6, text embedding, layout embedding, and position embedding together form the input of the text encoder." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.795, + 0.909, + 0.903 + ], + "angle": 0, + "content": "Map Element Encoder. Vectors can be represented as sequences of points, similar to words in sentences. Inspired by this, we designed MEE akin to language models [14]. The MEE employs \\(M\\) transformer layers for vector encoding and \\(N\\) cross-attention layers for multimodal fusion. Utilizing the method from [48], points of each vector are embedded as point embedding. To achieve a fixed-length representation," + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.516, + 0.957 + ], + "angle": 0, + "content": "6827" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.139, + 0.093, + 0.858, + 0.279 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.141, + 0.28, + 0.859, + 0.41 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.09, + 0.423, + 0.907, + 0.466 + ], + "angle": 0, + "content": "Figure 6. Overview of the modular approach. Entire approach can be divided into two main parts: Rule Extraction from Traffic Sign (top) and Rule-Lane Correspondence Reasoning (bottom). Rule Extraction model consists of two sequential stages with the same structure VLE but unshared parameters, and the training procedure is independent." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.492, + 0.484, + 0.688 + ], + "angle": 0, + "content": "we add [VEC] tokens as the first token of each vector, similar to [STC] tokens in the VLE. We also introduce learnable type embedding for vector types, learnable instance embedding to distinguish vector instances, and position embedding from [14] to encode the relative positions of multiple points within a vector. These embedding are aggregated as the input of vector encoder, as shown in Fig. 7. In addition, we employ inter-instance and intra-instance attention mechanisms [31] to prioritize interactions within vectors over interactions between vectors, as depicted in the dashed box on the right side of Fig. 7. The [VEC] token in output serves as fused feature of rules and vectors, enabling the final prediction of their relationships through association head." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.695, + 0.251, + 0.71 + ], + "angle": 0, + "content": "5.2. Implementation" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.717, + 0.483, + 0.763 + ], + "angle": 0, + "content": "We utilize VLE and MEE as backbones to integrate multiple modalities and address these two sub-tasks. The specific procedures are detailed as follows:" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.78, + 0.484, + 0.901 + ], + "angle": 0, + "content": "Rule Extraction from Traffic sign. To clarify the objectives of model, we first cluster symbols and texts into groups. As shown in the upper part of Fig. 6, the VLE is used to encode OCR results and images. By calculating the cosine similarity between [STC] tokens, different symbols and texts are clustered into groups. This process is supervised by contrastive loss during training. Next, using grouped OCR results as text input and maintaining the VLE structure, we" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.492, + 0.907, + 0.553 + ], + "angle": 0, + "content": "extract lane-level rules. We employ multiple linear layers as understanding head for the [CLS] token to predict the corresponding value for each property of the rules. This facilitates to express all rules as \\(\\{key:value\\}\\) pairs." + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.575, + 0.909, + 0.622 + ], + "angle": 0, + "content": "Rule-Lane Correspondence Reasoning. MEE is designed for vector encoding and interaction with rules. Each formatted rule is mapped to an embedding through MLP and" + }, + { + "type": "image", + "bbox": [ + 0.546, + 0.647, + 0.887, + 0.817 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.513, + 0.827, + 0.909, + 0.898 + ], + "angle": 0, + "content": "Figure 7. Structure of MEE. MEE serves as correspondence reasoning model. Learnable embeddings are introduced within input to enhance the representing capacity of vector types. inter & intra-instance attention mechanisms facilitate to capture the relationships and independence of individual vectors." + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.516, + 0.957 + ], + "angle": 0, + "content": "6828" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.134, + 0.087, + 0.868, + 0.306 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.09, + 0.318, + 0.908, + 0.349 + ], + "angle": 0, + "content": "Figure 8. Overview of end-to-end approaches. Based on different vectorized HD maps encoding methods, approaches can be categorized into three types." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.373, + 0.484, + 0.449 + ], + "angle": 0, + "content": "fused with vector features in the fusion encoder, as shown in the lower part of Fig. 6. We add a binary classification head after each [VEC] token to determine the relationship between the current centerline and rule. Details can be found in supplement." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.461, + 0.22, + 0.478 + ], + "angle": 0, + "content": "5.3. Experiment" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.484, + 0.486, + 0.606 + ], + "angle": 0, + "content": "Setup. The dataset is split into train and test sets in the ratio of \\(9:1\\). \\(L = 6\\) in VLE and \\(M = 2\\), \\(N = 2\\) in MEE. Input images are resized to \\(256 \\times 256\\) and the feature dimension is 768 with consistent 12 attention heads. We initialize VLE with pre-trained weights of DeiT [50] and BERT [14] while MEE is trained from scratch. The training procedure runs 50 and 120 epochs for VLE and MEE, respectively. More details can be found in the supplement." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.629, + 0.486, + 0.902 + ], + "angle": 0, + "content": "Results. We use a heuristic method based on OCR character matching and nearest lane matching as the modular baseline. Meanwhile we make modifications to ALBEF [26] and BERT [14] to adapt them to our task. As shown in Tab. 2, the heuristic method performs poorly in both R.E. and C.R. tasks, underscoring the challenging nature of the MapDR dataset. ALBEF-BERT method significantly enhances performance in the R.E. sub-task however lacking adequate adaptation to vector information hampers its performance on the C.R.. Tab. 3 indicates the attention mechanisms significantly improve \\(R_{R,E}\\), while layout of text brings marginal improvement. For C.R., the attention mechanisms is the critical part that promotes the effectiveness of MEE. Instance embedding slightly improves \\(P_{C,R}\\) and \\(R_{C,R}\\), while type embedding significantly enhances both, indicating that vector types help the model establish rule-lane correspondence. The separate evaluation results of all lane types and the strategies of heuristic method can be found in the supplement." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.372, + 0.727, + 0.39 + ], + "angle": 0, + "content": "6. End-to-End Approach" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.401, + 0.909, + 0.628 + ], + "angle": 0, + "content": "Recently, there has been notable progress in the development of Multimodal Large Language Models (MLLMs) [3, 5, 39, 49], demonstrating significant capabilities in processing multimodal data and acquiring implicit knowledge of modality alignment. Their large-scale pre-training endows them with exceptional understanding and generative abilities, thereby enabling the comprehensive execution of the task proposed by MapDR in an end-to-end manner. This section aims to address the overall task using Qwen-VL [5] and to investigate three distinct vector encoding methods. Moreover, we introduce an innovative RuleVLM model based on MEE, designed to process inputs from three modalities (images, texts, and vectors) and to produce driving rules, which yield favorable results on MapDR. Further elaboration on these concepts is presented in subsequent sections." + }, + { + "type": "title", + "bbox": [ + 0.514, + 0.649, + 0.649, + 0.663 + ], + "angle": 0, + "content": "6.1. Architecture" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.675, + 0.909, + 0.901 + ], + "angle": 0, + "content": "As shown in Fig. 8, we classify MLLM methods into three types: Qwen-VL(TextPrompt), Qwen-VL(VisualPrompt), and RuleVLM, based on different vector encoding methods. These methods utilize the framework of Qwen-VL [5], with ViT [15] serving as the image encoder. Qwen-VL(TextPrompt) encodes centerline coordinates in the PV image as text, inputs them into QwenLM; Qwen-VL(VisualPrompt) visualizes the centerline and its index in the PV image and inputs it as Visual Prompt into QwenLM, while RuleVLM uses MEE with the cross-attention layer removed to encode vectorized centerline results and aligns it with LLM through an adapter. The goal of all these methods is to output a serialized set of driving rules in text form, decode it through a json decoder to restore formatted data, and then integrate it into HD maps for further applications." + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.946, + 0.516, + 0.957 + ], + "angle": 0, + "content": "6829" + } + ], + [ + { + "type": "table_caption", + "bbox": [ + 0.09, + 0.089, + 0.907, + 0.133 + ], + "angle": 0, + "content": "Table 2. Evaluation of the overall task. The heuristic method and the Qwen-VL(TextPrompt) method serve as the baselines for the modular and end-to-end approach, respectively. \" - \" denotes end-to-end approach is not suitable to independent evaluations of C.R. because these approaches do not utilize ground truth of rules for correspondence reasoning independently." + }, + { + "type": "table", + "bbox": [ + 0.092, + 0.143, + 0.905, + 0.276 + ], + "angle": 0, + "content": "
ModelTypeR.E.C.R.Overall
PR.E.(%)RR.E.(%)PC.R.(%)RC.R.(%)Pall(%)Rall(%)F1 Score
Heuristic18.0111.5133.0517.995.012.730.035
ALBEF-BERTModular75.7857.564.1417.250.240.780.003
VLE-MEE76.6774.5478.0582.1663.3567.370.653
Qwen-VL(TextPrompt)42.2141.09--8.398.170.083
Qwen-VL(VisualPrompt)End-to-End89.2989.50--39.1439.230.392
RuleVLM89.2889.44--64.1664.280.642
" + }, + { + "type": "table_caption", + "bbox": [ + 0.09, + 0.299, + 0.486, + 0.356 + ], + "angle": 0, + "content": "Table 3. Evaluation of sub-tasks. \"Attn.\" indicates intra & inter-instance attention mechanisms. \"Layout\" refers to the text layout applied in VLE. \"In.E.\" and \"Ty.E.\" denotes instance and type embedding in MEE, respectively." + }, + { + "type": "table", + "bbox": [ + 0.117, + 0.366, + 0.446, + 0.56 + ], + "angle": 0, + "content": "
VLEPR.E.(%)RR.E.(%)
Attn.Layout
XX75.7857.56
✓X76.8671.75
✓✓76.6774.54
MEEPC.R.(%)RC.R.(%)
Attn.In.E.Ty.E.
XXX4.1417.25
✓XX68.9171.39
✓✓X69.6872.76
✓✓✓78.0582.16
" + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.586, + 0.251, + 0.603 + ], + "angle": 0, + "content": "6.2. Implementation" + }, + { + "type": "text", + "bbox": [ + 0.089, + 0.609, + 0.486, + 0.837 + ], + "angle": 0, + "content": "The JSON-formatted driving rules in MapDR are serialized into text and used as the answers in the QA format for training and evaluation. For all methods based on QwenVL, the pre-trained weights of Qwen-VL-Chat (9.6B) are used for initialization. For method using MEE, the trained weights of MEE from the modular approach are utilized for its initialization. During the supervised fine-tuning, we perform LoRA [21] fine-tuning of the QwenLM, while keeping the parameters of ViT frozen. For RuleVLM, we employ a linear layer as the adapter between QwenLM and MEE, while the parameters of MEE and the adapter remains trainable. We use special tokens to separate the centerline tokens extracted by MEE from the visual and text tokens. Additionally, to prevent overfitting, we shuffle the order of centerlines in the local vectorized map during training." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.848, + 0.223, + 0.864 + ], + "angle": 0, + "content": "6.3. Experiment." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.871, + 0.485, + 0.903 + ], + "angle": 0, + "content": "Setup. All experiments are conducted on QA formatted MapDR dataset with 20 epochs of training on NVIDIA RTX" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.301, + 0.907, + 0.363 + ], + "angle": 0, + "content": "A6000. The dimension of hidden states remains 4096 while \\( r \\) and alpha of LoRA configuration are set 64 and 16 respectively. Meanwhile training batch size is 6 and the max training token length remains 2048." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.38, + 0.909, + 0.683 + ], + "angle": 0, + "content": "Results. We conducted zero-shot evaluations on a subset of MapDR using existing MLLMs [3, 5, 39, 49] as qualitative investigation, but we encounter challenges in achieving stable structured outputs, further details are provided in the supplement. Additionally, as shown in the Tab. 2, Qwen-VL(TextPrompt) method serves as a end-to-end baseline do not perform better on both the R.E. and overall evaluation. Textual representations of centerline coordinates are challenging for MLLMs to perform spatial reasoning and also resulted in excessive length of sequence, which affected \\( P_{R.E.} \\) and \\( R_{R.E.} \\). Qwen-VL(VisualPrompt) visualizes the centerline within the image, resulting in reduced text sequence length, and demonstrates notable enhancements in the R.E. task compared to the baseline. However, the overall metrics are suboptimal, suggesting that image-based spatial reasoning remains an exceedingly challenging endeavor. The RuleVLM method that we propose introduces MEE for the independent encoding of vectorized HD maps information. This facilitates a more direct extraction of centerline features, resulting in significant enhancements in the overall metrics." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.695, + 0.634, + 0.71 + ], + "angle": 0, + "content": "7. Conclusion" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.72, + 0.909, + 0.901 + ], + "angle": 0, + "content": "In this work, we defined the task of integrating traffic regulations into vectorized HD maps to facilitate the construction of the traffic regulation layer within online HD maps. We constructed the MapDR dataset, which includes over 10,000 video clips, more than 400,000 images, and at least 18,000 driving rules, covering a total mileage of approximately \\(800Km\\). Additionally, we proposed the modular approach MEE-VLE and the end-to-end approach RuleVLM, establishing effective baselines for further research. We hope that MapDR can serve as a starting point, and that more researchers can contribute to expanding this dataset or proposing even better solutions." + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.516, + 0.956 + ], + "angle": 0, + "content": "6830" + } + ], + [ + { + "type": "title", + "bbox": [ + 0.093, + 0.091, + 0.188, + 0.106 + ], + "angle": 0, + "content": "References" + }, + { + "type": "ref_text", + "bbox": [ + 0.101, + 0.116, + 0.486, + 0.143 + ], + "angle": 0, + "content": "[1] Navigation data standard. https://nds-association.org/.4" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.146, + 0.485, + 0.172 + ], + "angle": 0, + "content": "[2] Openstreetmap. https://github.com/openmaptiles/openmaptiles.4" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.175, + 0.483, + 0.202 + ], + "angle": 0, + "content": "[3] Anthropic. Claude-3. https://www.anthropic.com/news/claudi-3-family, 2024.7,8" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.204, + 0.483, + 0.245 + ], + "angle": 0, + "content": "[4] Stanislaw Antol, Aishwarya Agrawal, Jiasen Lu, Margaret Mitchell, Dhruv Batra, C. Lawrence Zitnick, and Devi Parikh. VQA: visual question answering. In ICCV, 2015. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.248, + 0.483, + 0.315 + ], + "angle": 0, + "content": "[5] Jinze Bai, Shuai Bai, Shusheng Yang, Shijie Wang, Sinan Tan, Peng Wang, Junyang Lin, Chang Zhou, and Jingren Zhou. Qwen-vl: A frontier large vision-language model with versatile abilities. arXiv preprint arXiv:2308.12966, 2023.7, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.318, + 0.483, + 0.372 + ], + "angle": 0, + "content": "[6] Zechen Bai, Pichao Wang, Tianjun Xiao, Tong He, Zongbo Han, Zheng Zhang, and Mike Zheng Shou. Hallucination of multimodal large language models: A survey. arXiv preprint arXiv:2404.18930, 2024. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.375, + 0.483, + 0.442 + ], + "angle": 0, + "content": "[7] Hangbo Bao, Wenhui Wang, Li Dong, Qiang Liu, Owais Khan Mohammed, Kriti Aggarwal, Subhojit Som, Songhao Piao, and Furu Wei. Vlmo: Unified vision-language pre-training with mixture-of-modality-experts. In NeurIPS, 2022. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.446, + 0.483, + 0.487 + ], + "angle": 0, + "content": "[8] Karsten Behrendt, Libor Novak, and Rami Botros. A deep learning approach to traffic lights: Detection, tracking, and classification. In ICRA, 2017. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.49, + 0.483, + 0.544 + ], + "angle": 0, + "content": "[9] Holger Caesar, Varun Bankiti, Alex H. Lang, Sourabh Vora, Venice Erin Liong, Qiang Xu, Anush Krishnan, Yu Pan, Giancarlo Baldan, and Oscar Beijbom. nuscenes: A multimodal dataset for autonomous driving. In CVPR, 2020. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.547, + 0.483, + 0.615 + ], + "angle": 0, + "content": "[10] Xu Cao, Tong Zhou, Yunsheng Ma, Wenqian Ye, Can Cui, Kun Tang, Zhipeng Cao, Kaizhao Liang, Ziran Wang, James Rehg, and Chao Zheng. Maplm: A real-world large-scale vision-language dataset for map and traffic scene understanding. In CVPR, 2024. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.617, + 0.483, + 0.685 + ], + "angle": 0, + "content": "[11] Li Chen, Chonghao Sima, Yang Li, Zehan Zheng, Jiajie Xu, Xiangwei Geng, Hongyang Li, Conghui He, Jianping Shi, Yu Qiao, and Junchi Yan. Persformer: 3d lane detection via perspective transformer and the openlane benchmark. In ECCV, 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.688, + 0.483, + 0.755 + ], + "angle": 0, + "content": "[12] Long Chen, Oleg Sinavski, Jan Hunermann, Alice Karnsund, Andrew James Willmott, Danny Birch, Daniel Maund, and Jamie Shotton. Driving with llms: Fusing object-level vector modality for explainable autonomous driving. In ICRA, 2024. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.759, + 0.483, + 0.815 + ], + "angle": 0, + "content": "[13] Yaodong Cui, Shucheng Huang, Jiaming Zhong, Zhenan Liu, Yutong Wang, Chen Sun, Bai Li, Xiao Wang, and Amir Khajepour. Drivellm: Charting the path toward full autonomous driving with large language models. IEEE TIV, 2024. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.817, + 0.483, + 0.869 + ], + "angle": 0, + "content": "[14] Jacob Devlin, Ming-Wei Chang, Kenton Lee, and Kristina Toutanova. BERT: pre-training of deep bidirectional transformers for language understanding. In NAACL, 2019. 5, 6, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.873, + 0.483, + 0.9 + ], + "angle": 0, + "content": "[15] Alexey Dosovitskiy, Lucas Beyer, Alexander Kolesnikov, Dirk Weissenborn, Xiaohua Zhai, Thomas Unterthiner," + }, + { + "type": "list", + "bbox": [ + 0.094, + 0.116, + 0.486, + 0.9 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.548, + 0.093, + 0.907, + 0.147 + ], + "angle": 0, + "content": "Mostafa Dehghani, Matthias Minderer, Georg Heigold, Sylvain Gelly, Jakob Uszkoreit, and Neil Houlsby. An image is worth 16x16 words: Transformers for image recognition at scale. In ICLR, 2021. 5, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.15, + 0.907, + 0.205 + ], + "angle": 0, + "content": "[16] Gamal Elghazaly, Raphaël Frank, Scott Harvey, and Stefan Safko. High-definition maps: Comprehensive survey, challenges, and future perspectives. IEEE Open Journal of Intelligent Transportation Systems, 2023. 2, 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.207, + 0.906, + 0.247 + ], + "angle": 0, + "content": "[17] Andreas Fregin, Julian Müller, Ulrich Krebel, and Klaus Dietmayer. The driveu traffic light dataset: Introduction and comparison with existing datasets. In ICRA, 2018. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.249, + 0.907, + 0.289 + ], + "angle": 0, + "content": "[18] Shuo Gu, Yigong Zhang, Jinhui Tang, Jian Yang, and Hui Kong. Road detection through CRF based lidar-camera fusion. In ICRA, 2019. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.292, + 0.906, + 0.332 + ], + "angle": 0, + "content": "[19] Yunfei Guo, Wei Feng, Fei Yin, Tao Xue, Shuqi Mei, and Cheng-Lin Liu. Learning to understand traffic signs. In ACMMM, 2021. 2, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.334, + 0.906, + 0.376 + ], + "angle": 0, + "content": "[20] Yunfei Guo, Fei Yin, Xiao-Hui Li, Xudong Yan, Tao Xue, Shuqi Mei, and Cheng-Lin Liu. Visual traffic knowledge graph generation from scene images. In ICCV, 2023. 2, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.378, + 0.907, + 0.446 + ], + "angle": 0, + "content": "[21] Edward J. Hu, Yelong Shen, Phillip Wallis, Zeyuan Allen-Zhu, Yuanzhi Li, Shean Wang, Lu Wang, and Weizhu Chen. Lora: Low-rank adaptation of large language models. In The Tenth International Conference on Learning Representations, ICLR 2022, Virtual Event, April 25-29, 2022, 2022. 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.448, + 0.906, + 0.502 + ], + "angle": 0, + "content": "[22] Xinyu Huang, Peng Wang, Xinjing Cheng, Dingfu Zhou, Qichuan Geng, and Ruigang Yang. The apolloscope open dataset for autonomous driving and its application. IEEE TPAMI, 2020. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.504, + 0.906, + 0.546 + ], + "angle": 0, + "content": "[23] Yupan Huang, Tengchao Lv, Lei Cui, Yutong Lu, and Furu Wei. Layoutmv3: Pre-training for document AI with unified text and image masking. In ACMMM, 2022. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.548, + 0.906, + 0.603 + ], + "angle": 0, + "content": "[24] Chao Jia, Yinfei Yang, Ye Xia, Yi-Ting Chen, Zarana Parekh, Hieu Pham, Quoc V. Le, Yun-Hsuan Sung, Zhen Li, and Tom Duerig. Scaling up visual and vision-language representation learning with noisy text supervision. In ICML, 2021. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.604, + 0.907, + 0.644 + ], + "angle": 0, + "content": "[25] Wonjae Kim, Bokyung Son, and Ildoo Kim. Vilt: Vision-and-language transformer without convolution or region supervision. In ICML, 2021. 3, 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.647, + 0.907, + 0.713 + ], + "angle": 0, + "content": "[26] Junnan Li, Ramprasaath R. Selvaraju, Akhilesh Gotmare, Shafiq R. Joty, Caiming Xiong, and Steven Chu-Hong Hoi. Align before fuse: Vision and language representation learning with momentum distillation. In NeurIPS, 2021. 3, 5, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.717, + 0.907, + 0.772 + ], + "angle": 0, + "content": "[27] Junnan Li, Dongxu Li, Caiming Xiong, and Steven C. H. Hoi. BLIP: bootstrapping language-image pre-training for unified vision-language understanding and generation. In ICML, 2022. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.774, + 0.906, + 0.828 + ], + "angle": 0, + "content": "[28] Junnan Li, Dongxu Li, Silvio Savarese, and Steven C. H. Hoi. BLIP-2: bootstrapping language-image pre-training with frozen image encoders and large language models. In ICML, 2023. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.831, + 0.906, + 0.871 + ], + "angle": 0, + "content": "[29] Qi Li, Yue Wang, Yilun Wang, and Hang Zhao. Hdmapnet: An online HD map construction and evaluation framework. In ICRA, 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.873, + 0.906, + 0.902 + ], + "angle": 0, + "content": "[30] Benchcheng Liao, Shaoyu Chen, Xinggang Wang, Tianheng Cheng, Qian Zhang, Wenyu Liu, and Chang Huang. Maptr:" + }, + { + "type": "list", + "bbox": [ + 0.517, + 0.093, + 0.907, + 0.902 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.946, + 0.514, + 0.957 + ], + "angle": 0, + "content": "6831" + } + ], + [ + { + "type": "ref_text", + "bbox": [ + 0.126, + 0.092, + 0.482, + 0.12 + ], + "angle": 0, + "content": "Structured modeling and learning for online vectorized HD map construction. In ICLR, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.121, + 0.486, + 0.188 + ], + "angle": 0, + "content": "[31] Benchcheng Liao, Shaoyu Chen, Yunchi Zhang, Bo Jiang, Qian Zhang, Wenyu Liu, Chang Huang, and Xinggang Wang. Maptrv2: An end-to-end framework for online vectorized hd map construction. arXiv preprint arXiv:2308.05736, 2023. 3, 4, 5, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.19, + 0.485, + 0.232 + ], + "angle": 0, + "content": "[32] Haotian Liu, Chunyuan Li, Yuheng Li, and Yong Jae Lee. Improved baselines with visual instruction tuning. arXiv preprint arXiv:2310.03744, 2023. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.233, + 0.484, + 0.259 + ], + "angle": 0, + "content": "[33] Haotian Liu, Chunyuan Li, Qingyang Wu, and Yong Jae Lee. Visual instruction tuning. In NeurIPS, 2023." + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.261, + 0.484, + 0.327 + ], + "angle": 0, + "content": "[34] Haotian Liu, Chunyuan Li, Yuheng Li, Bo Li, Yuanhan Zhang, Sheng Shen, and Yong Jae Lee. Llava-last: Improved reasoning,OCR,and world knowledge.https://llava-v1.github.io/blog/2024-01-30-llava-last, 2024.3" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.33, + 0.484, + 0.371 + ], + "angle": 0, + "content": "[35] Rong Liu, Jinling Wang, and Bingqi Zhang. High definition map for automated driving: Overview and analysis. Journal of Navigation, 2020. 2, 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.372, + 0.484, + 0.412 + ], + "angle": 0, + "content": "[36] Yicheng Liu, Tianyuan Yuan, Yue Wang, Yilun Wang, and Hang Zhao. Vectormapnet: End-to-end vectorized HD map learning. In ICML, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.414, + 0.484, + 0.453 + ], + "angle": 0, + "content": "[37] Jiageng Mao, Yuxi Qian, Hang Zhao, and Yue Wang. Gptdriver: Learning to drive with GPT. In NeurIPS Workshop, 2023. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.456, + 0.484, + 0.524 + ], + "angle": 0, + "content": "[38] Ana-Maria Marcu, Long Chen, Jan Hunermann, Alice Karnsund, Benoit Hanotte, Prajwal Chidananda, Saurabh Nair, Vijay Badrinarayanan, Alex Kendall, Jamie Shotton, and Oleg Sinavski. Lingoqa: Video question answering for autonomous driving. arXiv preprint arXiv:2312.14115, 2023. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.526, + 0.484, + 0.551 + ], + "angle": 0, + "content": "[39] OpenAI. GPT-4 technical report. arXiv preprint arXiv:2303.08774, 2024. 7, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.553, + 0.484, + 0.593 + ], + "angle": 0, + "content": "[40] Jonah Philion and Sanja Fidler. Lift, splat, shoot: Encoding images from arbitrary camera rigs by implicitly unprojecting to 3d. In ECCV, 2020. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.595, + 0.484, + 0.649 + ], + "angle": 0, + "content": "[41] Tianwen Qian, Jingjing Chen, Linhai Zhuo, Yang Jiao, and Yu-Gang Jiang. Nuscenes-qa: A multi-modal visual question answering benchmark for autonomous driving scenario. In AAAI, 2024. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.651, + 0.484, + 0.732 + ], + "angle": 0, + "content": "[42] Alec Radford, Jong Wook Kim, Chris Hallacy, Aditya Ramesh, Gabriel Goh, Sandhini Agarwal, Girish Sastry, Amanda Askell, Pamela Mishkin, Jack Clark, Gretchen Krueger, and Ilya Sutskever. Learning transferable visual models from natural language supervision. In ICML, 2021. 3, 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.735, + 0.484, + 0.789 + ], + "angle": 0, + "content": "[43] Enna Sachdeva, Nakul Agarwal, Suhas Chundi, Sean Roelofs, Jiachen Li, Mykel J. Kochenderfer, Chiho Choi, and Behzad Dariush. Rank2tell: A multimodal driving dataset for joint importance ranking and reasoning. In WACV, 2024. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.79, + 0.484, + 0.844 + ], + "angle": 0, + "content": "[44] Chonghao Sima, Katrin Renz, Kashyap Chitta, Li Chen, Hanxue Zhang, Chengen Xie, Ping Luo, Andreas Geiger, and Hongyang Li. Drivelm: Driving with graph visual question answering. arXiv preprint arXiv:2312.14150, 2023. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.846, + 0.484, + 0.901 + ], + "angle": 0, + "content": "[45] Chonghao Sima, Katrin Renz, Kashyap Chitta, Li Chen, Hanxue Zhang, Chengen Xie, Jens Beiwenger, Ping Luo, Andreas Geiger, and Hongyang Li. Drivelm: Driving with graph visual question answering. In ECCV, 2024. 3" + }, + { + "type": "list", + "bbox": [ + 0.093, + 0.092, + 0.486, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.093, + 0.908, + 0.146 + ], + "angle": 0, + "content": "[46] Johannes Stallkamp, Marc Schlipsing, Jan Salmen, and Christian Igel. Man vs. computer: Benchmarking machine learning algorithms for traffic sign recognition. *Neural Networks*, 2012. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.149, + 0.908, + 0.259 + ], + "angle": 0, + "content": "[47] Pei Sun, Henrik Kretzschmar, Xerxes Dotiwalla, Aurelien Chouard, Vijaysai Patnaik, Paul Tsui, James Guo, Yin Zhou, Yuning Chai, Benjamin Caine, Vijay Vasudevan, Wei Han, Jiquan Ngiam, Hang Zhao, Aleksei Timofeev, Scott Ettinger, Maxim Krivokon, Amy Gao, Aditya Joshi, Yu Zhang, Jonathon Shlens, Zhifeng Chen, and Dragomir Anguelov. Scalability in perception for autonomous driving: Waymo open dataset. In CVPR, 2020. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.261, + 0.908, + 0.327 + ], + "angle": 0, + "content": "[48] Matthew Tancik, Pratul P. Srinivasan, Ben Mildenhall, Sara Fridovich-Keil, Nithin Raghavan, Utkarsh Singhal, Ravi Ramamoorthi, Jonathan T. Barron, and Ren Ng. Fourier features let networks learn high frequency functions in low dimensional domains. In NeurIPS, 2020. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.33, + 0.908, + 0.397 + ], + "angle": 0, + "content": "[49] Gemini Team, Rohan Anil, Sebastian Borgeaud, Yonghui Wu, Jean-Baptiste Alayrac, Jiahui Yu, Radu Soricut, Johan Schalkwyk, Andrew M. Dai, Anja Hauth, and et al. Gemini: A family of highly capable multimodal models. arXiv preprint arXiv:2312.11805, 2024. 7, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.4, + 0.908, + 0.453 + ], + "angle": 0, + "content": "[50] Hugo Touvron, Matthieu Cord, Matthijs Douze, Francisco Massa, Alexandre Sablayrolles, and Hervé Jégou. Training data-efficient image transformers & distillation through attention. In ICML, 2021. 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.456, + 0.908, + 0.495 + ], + "angle": 0, + "content": "[51] Ashish Vaswani, Noam Shazeer, Niki Parmar, Jakob Uszkoreit, Llion Jones, Aidan N. Gomez, Lukasz Kaiser, and Illia Polosukhin. Attention is all you need. In NeurIPS, 2017. 3, 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.498, + 0.908, + 0.578 + ], + "angle": 0, + "content": "[52] Huijie Wang, Tianyu Li, Yang Li, Li Chen, Chonghao Sima, Zhenbo Liu, Bangjun Wang, Peijin Jia, Yuting Wang, Shengyin Jiang, Feng Wen, Hang Xu, Ping Luo, Junchi Yan, Wei Zhang, and Hongyang Li. Openlane-v2: A topology reasoning benchmark for unified 3d HD mapping. In NeurIPS, 2023. 2, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.581, + 0.908, + 0.662 + ], + "angle": 0, + "content": "[53] Benjamin Wilson, William Qi, Tanmay Agarwal, John Lambert, Jagjeet Singh, Siddhesh Khandelwal, Bowen Pan, Ratnesh Kumar, Andrew Hartnett, Jhony Kaesemodel Pontes, Deva Ramanan, Peter Carr, and James Hays. Argoverse 2: Next generation datasets for self-driving perception and forecasting. In NeurIPS, 2021. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.665, + 0.905, + 0.705 + ], + "angle": 0, + "content": "[54] Dongming Wu, Jiahao Chang, Fan Jia, Yingfei Liu, Tiancai Wang, and Jianbing Shen. Topomlp: A simple yet strong pipeline for driving topology reasoning. In ICLR, 2024. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.707, + 0.905, + 0.747 + ], + "angle": 0, + "content": "[55] Yiheng Xu, Minghao Li, Lei Cui, Shaohan Huang, Furu Wei, and Ming Zhou. Layoutm: Pre-training of text and layout for document image understanding. In KDD, 2020. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.749, + 0.908, + 0.816 + ], + "angle": 0, + "content": "[56] Yang Xu, Yiheng Xu, Tengchao Lv, Lei Cui, Furu Wei, Guoxin Wang, Yijuan Lu, Dinei A. F. Floréncio, Cha Zhang, Wanxiang Che, Min Zhang, and Lidong Zhou. Layoutmv2: Multi-modal pre-training for visually-rich document understanding. In ACL, 2021. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.818, + 0.908, + 0.872 + ], + "angle": 0, + "content": "[57] Zhenhua Xu, Yujia Zhang, Enze Xie, Zhen Zhao, Yong Guo, Kwan-Yee K. Wong, Zhenguo Li, and Hengshuang Zhao. Drivegpt4: Interpretable end-to-end autonomous driving via large language model. IEEE Robotics Autom. Lett., 2024. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.874, + 0.908, + 0.9 + ], + "angle": 0, + "content": "[58] Fisher Yu, Haofeng Chen, Xin Wang, Wenqi Xian, Yingying Chen, Fangchen Liu, Vashisht Madhavan, and Trevor Dar" + }, + { + "type": "list", + "bbox": [ + 0.517, + 0.093, + 0.908, + 0.9 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.516, + 0.956 + ], + "angle": 0, + "content": "6832" + } + ], + [ + { + "type": "ref_text", + "bbox": [ + 0.125, + 0.092, + 0.482, + 0.12 + ], + "angle": 0, + "content": "rell. BDD100K: A diverse driving dataset for heterogeneous multitask learning. In CVPR, 2020. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.122, + 0.484, + 0.163 + ], + "angle": 0, + "content": "[59] Jiahui Yu, Zirui Wang, Vijay Vasudevan, Legg Yeung, Mojtaba Seyedhosseini, and Yonghui Wu. Coca: Contrastive captioners are image-text foundation models. TMLR, 2022. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.165, + 0.482, + 0.205 + ], + "angle": 0, + "content": "[60] Zhe Zhu, Dun Liang, Song-Hai Zhang, Xiaolei Huang, Baoli Li, and Shi-Min Hu. Traffic-sign detection and classification in the wild. In CVPR, 2016. 2" + }, + { + "type": "list", + "bbox": [ + 0.093, + 0.092, + 0.484, + 0.205 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.946, + 0.515, + 0.956 + ], + "angle": 0, + "content": "6833" + } + ] +] \ No newline at end of file diff --git a/2025/Driving by the Rules_ A Benchmark for Integrating Traffic Sign Regulations into Vectorized HD Map/cbe7a2b6-7e9c-4185-bab6-ca66fed01b85_origin.pdf b/2025/Driving by the Rules_ A Benchmark for Integrating Traffic Sign Regulations into Vectorized HD Map/cbe7a2b6-7e9c-4185-bab6-ca66fed01b85_origin.pdf new file mode 100644 index 0000000000000000000000000000000000000000..70cea9282158bc516c62defa762db4224a0a0358 --- /dev/null +++ b/2025/Driving by the Rules_ A Benchmark for Integrating Traffic Sign Regulations into Vectorized HD Map/cbe7a2b6-7e9c-4185-bab6-ca66fed01b85_origin.pdf @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:ef52e445117ae25b59b894bfcb3e4bdb25f160eb656549a07a21768bd8eacdc0 +size 7490316 diff --git a/2025/Driving by the Rules_ A Benchmark for Integrating Traffic Sign Regulations into Vectorized HD Map/full.md b/2025/Driving by the Rules_ A Benchmark for Integrating Traffic Sign Regulations into Vectorized HD Map/full.md new file mode 100644 index 0000000000000000000000000000000000000000..9976e3839cf57a8810373f1805760a9a3f40fecf --- /dev/null +++ b/2025/Driving by the Rules_ A Benchmark for Integrating Traffic Sign Regulations into Vectorized HD Map/full.md @@ -0,0 +1,280 @@ +# Driving by the Rules: A Benchmark for Integrating Traffic Sign Regulations into Vectorized HD Map + +Xinyuan Chang $^{1*}$ Maixuan Xue $^{2*†}$ Xinran Liu $^{1}$ Zheng Pan $^{1}$ Xing Wei $^{2‡}$ $^{1}$ Amap, Alibaba Group $^{2}$ Xi’an Jiaotong University +{changxinyuan.cxy, tom.lxr, panzheng.pan}@alibaba-inc.com +xmx0809@stu.xjtu.edu.cn, weixing@mail.xjtu.edu.cn + +![](images/fa98e39510a06b77ea1def57287ac98a0b8b5c58fe12f39b937e421ac5f465cb.jpg) +Figure 1. MapDR Overview and Motivation. The image on the left depicts the comprehensive results of HD map construction for an intersection scene, whereas the image on the right illustrates the outcomes after partitioning it into three layers. Existing online mapping methods primarily emphasize the construction of the geometric and connectivity layers, neglecting the traffic regulation layer. However, precise comprehension of traffic signs and their correlation with lanes is vital for ensuring the safety of autonomous driving. + +![](images/df3966c6489e0b02b577968def196aff7d0c31af7f128e2738124bf701658d61.jpg) +Geometric layer + +![](images/f005aea12fc3690efb0eaff4f4543bcdbb88f79353f3d6c94bdf59d688586855.jpg) +Connectivity layer + +![](images/218888cb25494023aaae06c9ff7b5f0bb0ce06a5e6579921a59ae6257f422520.jpg) +Traffic regulation layer + +# Abstract + +Ensuring adherence to traffic sign regulations is essential for both human and autonomous vehicle navigation. While current online mapping solutions often prioritize the construction of the geometric and connectivity layers of HD maps, overlooking the construction of the traffic regulation layer within HD maps. Addressing this gap, we introduce MapDR, a novel dataset designed for the extraction of Driving Rules from traffic signs and their association with vectorized, locally perceived HD Maps. MapDR features over 10,000 annotated video clips that capture the intricate correlation between traffic sign regulations and + +lanes. Built upon this benchmark and the newly defined task of integrating traffic regulations into online HD maps, we provide modular and end-to-end solutions: VLE-MEE and RuleVLM, offering a strong baseline for advancing autonomous driving technology. It fills a critical gap in the integration of traffic sign rules, contributing to the development of reliable autonomous driving systems. Code is available at https://github.com/MIV-XJTU/MapDR. + +# 1. Introduction + +The rise of autonomous vehicles and intelligent transportation systems has emphasized the importance of accurate, reliable navigational data. High-Definition (HD) maps, with their detailed representations of road elements, have become essential to supporting these systems. An HD map typically + +consists of three primary layers: the geometric layer, connectivity layer, and traffic regulation layer [16, 35], as shown in Fig. 1. The geometric layer provides vector information, such as dividers and lane centerlines; the connectivity layer defines lane relationships to assist in path planning; and the traffic regulation layer encodes rule information (e.g., HOV lanes, bus lanes, speed-limit zones) associated with lanes, ensuring compliant driving. + +A limitation of conventional HD maps is their inability to support rapid, real-time updates, making the construction of online HD maps an emerging trend. Existing methods, such as MapTR [30] and TopoMLP [54], address only the geometric and connectivity layers, overlooking the traffic regulation layer. As a result, autonomous driving systems currently still rely on offline maps to obtain traffic regulations, an approach that runs counter to the trend of online HD map construction. + +Traffic signs serve as a "visual language" on the road, playing a critical role in defining traffic regulations. Inspired by human driving processes, the creation of traffic regulations can be divided into two stages: 1) understanding the rules conveyed by traffic signs, and 2) determining the lanes to which these rules apply. Although OpenLane-V2 [52] has made strides in associating signs with lanes in online mapping, it has two primary limitations. First, it considers only directional signs, neglecting other types of regulations. Second, its label annotations are limited to sign categories, falling short of the detailed rule descriptions required for autonomous driving. A more structured description, in alignment with HD map standards, is essential for supporting comprehensive autonomous driving applications [16, 35]. + +To bridge this gap, we introduce MapDR, the first dataset specifically designed for the challenging task of integrating traffic regulations into existing online HD maps. MapDR focuses on complex traffic scenes with diverse signage, collecting over 10,000 video clips across three representative cities in China, with each clip containing at least one traffic sign. In addition to vectorized representations of dividers, boundaries, and centerlines, MapDR provides structural annotations of traffic regulations and their associations with lanes, offering a comprehensive foundation for advancing online HD maps. More details of MapDR can be found in Sec. 4. + +Based on MapDR, in addition to the end-to-end task definition, we also divide this task into two innovative subtasks aimed at bolstering research in this domain: 1) rule extraction from traffic signs, and 2) rule-lane correspondence reasoning. We introduce Vision-Language Encoder (VLE) and Map Element Encoder (MEE) to handle the interaction of multimodal data, encompassing images, texts, and vectors. Through the concatenation of these two models, integrated traffic regulations can be obtained. Additionally, inspired by recent advancements in Multimodal Large Lan + +Table 1. Comparison of the existing datasets. "Sign" and "Lane" denote whether the dataset focuses on traffic signs and lanes. Only those annotated with formatted ("Fmt.") rules and the correspondence ("Corr.") between rules and lanes can form driving rules. "Clip" represents whether the data is organized in the form of video clips. "* denotes that these samples are not newly collected and are built upon the previous dataset. + +
DatasetSignLaneDriving RulesNumber of SamplesYear
Fmt.Corr.ImageClipRegion
nuScenes [9]✓1400K1KWorldwide2019
Argoverse2 [53]✓2100K1KUSA2021
CTSU [19]✓5K/China2021
OpenLane [11]✓200K*1K*Worldwide2022
RS10K [20]✓✓10K/China2023
OpenLane-V2 [52]✓✓✓466K*2K*Worldwide2023
Waymo [47]✓✓390K2KUSA2024
MapDR(ours)✓✓✓✓400K10KChina2024
+ +guage Models (MLLMs), we also explore an end-to-end MLLM solution called RuleVLM to solve this task holistically. These approaches establish robust baselines and aim to inspire future research in this area. For detailed descriptions of the proposed tasks and metrics, please refer to Sec. 3. For approaches, please refer to Secs. 5 and 6. + +To sum up, our contributions are as follows: + +- For the first time, We introduce the novel task of extracting lane-level rules from traffic signs and integrating them into vectorized HD maps, alongside the MapDR dataset and evaluation metrics for benchmarking. +- MapDR comprises over 10,000 video clips and 400,000 front-view images captured across multiple cities in China, covering diverse traffic conditions and including more than 18,000 annotated lane-level rules. All data are newly collected and meticulously validated. +- We propose two modeling approaches, modular and end-to-end, to address the integration of lane-level rules into vectorized HD maps, providing effective baselines for future research. + +# 2. Related Work + +# 2.1. HD Map Construction + +HD maps construction have seen significant advancements, with a focus on traffic element perception, including lane detection and traffic sign recognition [8, 9, 17, 18, 22, 46, 53, 58, 60]. The shift towards BEV perception and vectorization for end-to-end HD maps construction has gained traction [9, 11, 53]. Notable works include HDMapNet, which aggregates semantic segmentation results [29], LSS [40] estimates depth to transfer image features to BEV features, while VectorMapNet [36] is the first end-to-end framework for sequential vector point prediction to generate HD maps without post-processing. MapTR [30] and its enhanced + +![](images/307ba3b46ca57ad5adb2d0d4c501f5f7f069580f9a1c1f82b612dfb815c84d96.jpg) +Figure 2. Overview of the sub-tasks. Step 1 ~ Step 4 shows a case of driving by the rules. Step 2 and Step 3 demonstrates the specific role of two sub-tasks, respectively. + +version, MapTRv2 [31], introduced a unified permutation-equivalent modeling approach and extended it to a general framework supporting centerline learning and 3D map construction. However, these efforts have largely overlooked the integration of traffic sign rules into HD maps. + +# 2.2. Traffic Element Association + +Traffic element association aims to link elements like traffic signs with lanes. As demonstrated in Tab. 1, CTSU has initiated internal elements association to describe traffic sign in $\{key: value\}$ form, however lacking both generalization of driving rules from description and lane association [19]. VTKGG [20] propose to utilize a graph model for connectivity but also lacks structured expression of driving rules for motion planning and requires complex integration into HD maps, which is typically expressed in the BEV space. OpenLane-V2 [52] advances traffic sign and lane association in BEV space, but is constrained by single-label classification for traffic sign rather than structured descriptions for fine-grained driving rules. This makes it insufficient for signs with multiple rules, which are complex but common in real scenarios. Recent MLLM-based benchmarks [10, 38, 41, 43, 44] for autonomous driving, such as MAPLM [10], prioritize end-to-end motion planning over precise rule extraction from traffic sign, lacking evaluation for rule reasoning. MapDR addresses this gap by focusing on traffic sign rule extraction and lane association. + +# 2.3. Vision-Language Models + +Vision-Language Models (VLMs) facilitates multimodal applications by learning joint representations of vision and language data. Visual Question Answer (VQA) tasks provide answers to image-related questions [4], while Visual Information Extraction (VIE) tasks extract structured information from visual and textual data [4, 23, 55, 56]. In Autonomous Driving (AD), VLMs are increasingly used for comprehensive traffic scene understanding and decision-making. The field has seen various approaches, including using transformers [51] for joint encoding [23, 25], excelling at multimodal information interaction, and independent en + +coders for different modalities [24, 42] that are proficient in multimodal retrieval. Cross-modal representation methods [26, 59] combine these advantages, and the latest LLM-based research [28, 32-34] has achieved state-of-the-art results in various multimodal tasks. Nowadays, an increasing number of methods are leveraging LLMs to achieve impressive results [12, 37, 45, 57], with works like DriveLLM [13] showing significant potential in AD. However, addressing hallucination [6] remains the most crucial aspect for LLM-based approaches. + +# 3. Task Definition : Integrating Traffic Sign Regulations into HD Maps + +The ability to discern rules from traffic signs and to associate them with specific lanes is pivotal for autonomous navigation. Our proposed task focuses on extracting lane-level driving rules denoted as $R = \{r_i\}_{i=1}^m$ and reasoning the correspondence between these rules and the centerlines $L = \{l_i\}_{i=1}^k$ in the local vectorized HD map, where $m$ is the number of rules and $k$ is the number of centerlines. Each rule $r_i$ is a set of pre-defined properties in $\{key : value\}$ pairs, as shown in Fig. 3. The final correspondence forms a bipartite graph $G = (R \cup L, E)$ , where $E \subseteq \{0,1\}^{m \times k}$ and the element $E_{ij}$ in the $i$ -th row and $j$ -th column of matrix $E$ represents the corresponding status between $r_i$ and $l_j$ . In this case, we assume the entire approach as $\mathcal{F}$ and the overall process can be formulated as $G = \mathcal{F}(X, L)$ , where $X = \{x_i\}_{i=1}^n$ represents a series of image sequences and $n$ is the number of frames. To independently investigate the capabilities of rule extraction and correspondence reasoning within the overall task, we further divided the overall task into two sub-tasks, defined as follows. + +# 3.1. Rule Extraction from Traffic Sign + +As shown in Step 2 of Fig. 2, this sub-task involves extracting multiple rules $R = \{r_i\}_{i=1}^m$ from a series of image sequences $X = \{x_i\}_{i=1}^n$ . The rule extraction model, denoted as $\mathcal{M}$ , can be expressed as $R = \mathcal{M}(X)$ . + +![](images/732294b9cf93258cb44928eb52493a5934660a5921e35324248e2eb038033e09.jpg) +Figure 3. Visualization of dataset demo. Multiple lane-level rules of a single traffic sign are annotated in $\{key: value\}$ format. Directed lines indicate the correspondence between rules and particular centerlines. + +# 3.2. Rule-Lane Correspondence Reasoning + +As shown in Step 3 of Fig. 2, the reasoning process establishes the correspondence between centerlines $L = \{l_i\}_{i=1}^k$ and all rules $R$ . We denote the correspondence reasoning model as $\mathcal{T}$ , and this process can be described as $E = \mathcal{T}(R, L)$ . The final reasoning result forms a bipartite graph $G = (R \cup L, E)$ , which means corresponding relationships only exist between rules and centerlines. + +# 4. The MapDR Dataset & Benchmark + +We introduce the MapDR dataset, meticulously annotated with traffic sign regulations and their correspondences to lanes, as shown in Fig. 3. The dataset encompasses a diverse range of scenarios, weather conditions, and traffic situations, with over 10,000 traffic scene segments, 18,000 driving rules, and 400,000 images. Traffic signs typically have varying textual descriptions, text layouts, and positions on the road, which add complexity to the task. + +The majority of the data originates from Beijing and Shanghai, with additional scenes from Guangzhou. Fig. 4 illustrates the geographic spread and variety of traffic signs. The dataset reflects a natural long-tail distribution, with a prevalence of bus and direction lanes and a scarcity of tidalflow lanes. We primarily focus on traffic signs that indicate lane-level rules, collected from cities with the most complex and diverse traffic scenarios in China, ensuring realistic and representative data. All images have undergone privacy and safety processing to obscure license plates and faces. More comprehensive statistic of dataset and case demonstrations can be found in the supplement. + +# 4.1. Raw Data & Annotation + +Raw Data. MapDR is collected from real-world traffic scenes, each scene segment (video clip) captures front-view images within a $100m \times 100m$ area centered on the traffic sign, with a consistent resolution of $1920 \times 1240$ . Each clip contains 30 to 60 frames, captured at 1 frame every 2 meters, ensuring consistent spatial intervals. Each video clip focuses on a single traffic sign and provides its position in 3D space. + +Camera intrinsics and poses are provided for each frame, and coordinates for each clip are transformed to distinct ENU systems. + +All vectors of local map in the target area are provided as 3D point lists, generated using our algorithm similar to MapTRv2 [31]. Each lane vector has a type, such as divider, centerline, crosswalk, or boundary. For example, the centerline is defined as $L = \{l_i\}_{i=1}^k$ , where each vector $l_i$ is composed of multiple 3D points $l_i = [p_1, \ldots, p_n]$ , and $p_j = (x_j, y_j, z_j)$ represents the coordinates of the current point. The pipeline of dataset production is illustrated in Fig. 5, and detailed data acquisition and annotation procedures can be found in the supplement material. + +**Formatted Rules.** Each video clip may contain multiple lane-level rules, denoted as $R$ . Each rule is expressed by symbols and text on the sign, requiring interpretation. We reference existing HD map design and data specifications [1, 16, 35] to define lane-level driving rules in the JSON format. As shown in Fig. 3, each rule $r_i$ comprises 8 predefined properties in the form of {key : value} pairs. This definition schema encompasses most scenarios and supports integration into existing autonomous driving systems. We enclose the symbols and texts denoting each distinct rule on traffic signs with polygons and project them into 3D space as $P_i = [p_1, \ldots, p_n]$ , where $n$ varies. Researchers can optionally use this information to facilitate rule extraction. + +Correspondence between Rules & Lanes. Based on formatted rules $R$ and centerlines $L$ , corresponding centerlines of each rule are annotated as shown in Fig. 3. Therefore correspondence between rules and centerlines can be formed as a bipartite graph $G = (R \cup L, E)$ , where $E \subseteq \{0,1\}^{|R| \times |L|}$ and the positive edges only exist between $R$ and $L$ as demonstrated in Sec. 3.2. Specifically, $E_{ij} = 1$ represents that + +![](images/8d41f316abb264102e743f75839b8ebe308aa55d8e30be049b22a61d8fd1b753.jpg) +Figure 4. Geographic location distribution of the collected traffic signs and proportions of various lane types represented in all signs. The geographic distribution is visualized based on OpenStreetMap [2]. + +![](images/6ac67f847f7504ee637375f74cc16c051b9fc9e4b23c72a868597759f49e1979.jpg) +Figure 5. Pipeline of dataset production. The location of traffic signs are sampled from existing database then front-view images of each sign are newly collected. Vectorized map is processed in cloud server. Finally formatted rules and correspondence between rules and centerlines are annotated and organized as MapDR. + +vehicle driving on the lane with centerline $l_{j}$ should follow the driving rule $r_{i}$ . + +# 4.2. Evaluation Metrics + +Based on the task defined in Sec. 3, we propose evaluation metrics for entire task as well as for two sub-tasks separately. The $F_{1}$ score is adopted as the benchmark for methods rank. + +Rule Extraction (R.E.). Given the ground truth $R$ and predicted rules $\hat{R}$ , we propose to calculate the Precision $(P_{R,E.})$ and Recall $(R_{R,E.})$ to evaluate the capability of rules extraction as defined in Eq. (1), where $\hat{r_i} = r_i$ represents all the properties are predicted correctly. + +$$ +P _ {R. E.} = \frac {\left| \hat {R} \cap R \right|}{\left| \hat {R} \right|} \quad R _ {R. E.} = \frac {\left| \hat {R} \cap R \right|}{\left| R \right|} \tag {1} +$$ + +Correspondence Reasoning (C.R.). Given the ground truth of correspondence bipartite graph $G = (R \cup L, E)$ and predicted graph $\hat{G} = (R \cup L, \hat{E})$ , we propose to calculate Precision ( $P_{C.R.}$ ) and Recall ( $R_{C.R.}$ ) of edge set $E$ to evaluate the capability of correspondence reasoning individually. Metrics are defined as Eq. (2). + +$$ +P _ {C. R.} = \frac {\left| \hat {E} \cap E \right|}{\left| \hat {E} \right|} \quad R _ {C. R.} = \frac {\left| \hat {E} \cap E \right|}{\left| E \right|} \tag {2} +$$ + +Overall. As defined in Sec. 3, the final predicted results can be obtained through an end-to-end prediction or a combination of two sequential sub-tasks. Given the predicted rules, correspondence should be reasoned between $\hat{R}$ and $L$ which means the prediction of entire task is $\hat{G} = (\hat{R} \cup L, \hat{E})$ and the ground truth is consistent $G = (R \cup L, E)$ . We evaluate Precision $(P_{all})$ and Recall $(R_{all})$ using the subgraph $G^{s}$ , where $G^{s} = \{g_{ij}^{s}\}_{i=1,j=1}^{m,k}$ , $g_{ij}^{s} = (\{r_{i}, l_{j}\}, e_{ij})$ . In set of sub-graph $G^{s}$ , $m$ is the number of rules, and $k$ is the number of centerlines. Furthermore, we propose the $F_{1}$ for the final benchmark ranking. Metrics are defined in Eq. (3) and Eq. (4). The $F_{1}$ is derived from the calculations of $P_{all}$ and $R_{all}$ . We provide an example of calculating the Overall metrics in the supplement material. + +$$ +P _ {a l l} = \frac {\left| \hat {G} ^ {s} \cap G ^ {s} \right|}{\left| \hat {G} ^ {s} \right|} \quad R _ {a l l} = \frac {\left| \hat {G} ^ {s} \cap G ^ {s} \right|}{\left| G ^ {s} \right|} \tag {3} +$$ + +$$ +F _ {1} = 2 \times \frac {P _ {\text {a l l}} \times R _ {\text {a l l}}}{P _ {\text {a l l}} + R _ {\text {a l l}}} \tag {4} +$$ + +# 5. Modular Approach + +In order to conduct a comprehensive analysis of the entire task, a modular approach is implemented in this section to address the two sub-tasks. It is demonstrated that these interconnected modules can effectively achieve the ultimate objective of integrating driving rules into the HD map. To address the multimodal information interaction encompassing images, texts, and vectors, a Vision-Language Encoder (VLE) and a Map Element Encoder (MEE) are developed. The following sections expound upon their structures, applications, and the experimental results on MapDR. + +# 5.1. Architecture + +Vision-Language Encoder. Inspired by vision-language frameworks [7, 25-27, 42], we designed a vision-language fusion model named VLE, following [26]. As shown in Fig. 6, VLE uses ViT-b16 [15] as the vision encoder, with the text encoder and multimodal fusion encoder each consisting of $L$ transformer layers [51]. Each layer of the fusion encoder includes a cross-attention module for fusion [26]. In practice, distinct rules are represented by varying numbers of symbols and texts, as shown in the OCR results in Fig. 6. To address the challenge of representing variable-length input as fixed-length features, we introduce a [CLS] token for an entire rule and several [STC] tokens for sentence-level representation. The specific usage of these tokens is detailed in Sec. 5.2. Furthermore, we incorporate inter-instance and intra-instance attention mechanisms [31] to enhance model performance by capturing interactions and independence between and within sentences. In addition to content, layout captures the relative positions of symbols and texts, offering important semantic meaning. To leverage this, we encode the layout using the method from [48] and the relative positions of characters as position embedding following [14]. As shown in VLE in Fig. 6, text embedding, layout embedding, and position embedding together form the input of the text encoder. + +Map Element Encoder. Vectors can be represented as sequences of points, similar to words in sentences. Inspired by this, we designed MEE akin to language models [14]. The MEE employs $M$ transformer layers for vector encoding and $N$ cross-attention layers for multimodal fusion. Utilizing the method from [48], points of each vector are embedded as point embedding. To achieve a fixed-length representation, + +![](images/20dd1837f8f401df33c212c7e9b6b9ac2bae69f347a90298fbacad52d2affe46.jpg) + +![](images/7eac7cfdb22f5c436f60b18c8f4eedb176cd67f2f171a678d7f6ba2ad7dbae6e.jpg) +Figure 6. Overview of the modular approach. Entire approach can be divided into two main parts: Rule Extraction from Traffic Sign (top) and Rule-Lane Correspondence Reasoning (bottom). Rule Extraction model consists of two sequential stages with the same structure VLE but unshared parameters, and the training procedure is independent. + +we add [VEC] tokens as the first token of each vector, similar to [STC] tokens in the VLE. We also introduce learnable type embedding for vector types, learnable instance embedding to distinguish vector instances, and position embedding from [14] to encode the relative positions of multiple points within a vector. These embedding are aggregated as the input of vector encoder, as shown in Fig. 7. In addition, we employ inter-instance and intra-instance attention mechanisms [31] to prioritize interactions within vectors over interactions between vectors, as depicted in the dashed box on the right side of Fig. 7. The [VEC] token in output serves as fused feature of rules and vectors, enabling the final prediction of their relationships through association head. + +# 5.2. Implementation + +We utilize VLE and MEE as backbones to integrate multiple modalities and address these two sub-tasks. The specific procedures are detailed as follows: + +Rule Extraction from Traffic sign. To clarify the objectives of model, we first cluster symbols and texts into groups. As shown in the upper part of Fig. 6, the VLE is used to encode OCR results and images. By calculating the cosine similarity between [STC] tokens, different symbols and texts are clustered into groups. This process is supervised by contrastive loss during training. Next, using grouped OCR results as text input and maintaining the VLE structure, we + +extract lane-level rules. We employ multiple linear layers as understanding head for the [CLS] token to predict the corresponding value for each property of the rules. This facilitates to express all rules as $\{key:value\}$ pairs. + +Rule-Lane Correspondence Reasoning. MEE is designed for vector encoding and interaction with rules. Each formatted rule is mapped to an embedding through MLP and + +![](images/27b818cb5a8c6c71eff3683acaf306cecfb29692e6fad6b0e40f1e43d71736d1.jpg) +Figure 7. Structure of MEE. MEE serves as correspondence reasoning model. Learnable embeddings are introduced within input to enhance the representing capacity of vector types. inter & intra-instance attention mechanisms facilitate to capture the relationships and independence of individual vectors. + +![](images/81fe351a1fe39d9a99bba6400a5b27163cf6ccb63532f0941af583291dd262e1.jpg) +Figure 8. Overview of end-to-end approaches. Based on different vectorized HD maps encoding methods, approaches can be categorized into three types. + +fused with vector features in the fusion encoder, as shown in the lower part of Fig. 6. We add a binary classification head after each [VEC] token to determine the relationship between the current centerline and rule. Details can be found in supplement. + +# 5.3. Experiment + +Setup. The dataset is split into train and test sets in the ratio of $9:1$ . $L = 6$ in VLE and $M = 2$ , $N = 2$ in MEE. Input images are resized to $256 \times 256$ and the feature dimension is 768 with consistent 12 attention heads. We initialize VLE with pre-trained weights of DeiT [50] and BERT [14] while MEE is trained from scratch. The training procedure runs 50 and 120 epochs for VLE and MEE, respectively. More details can be found in the supplement. + +Results. We use a heuristic method based on OCR character matching and nearest lane matching as the modular baseline. Meanwhile we make modifications to ALBEF [26] and BERT [14] to adapt them to our task. As shown in Tab. 2, the heuristic method performs poorly in both R.E. and C.R. tasks, underscoring the challenging nature of the MapDR dataset. ALBEF-BERT method significantly enhances performance in the R.E. sub-task however lacking adequate adaptation to vector information hampers its performance on the C.R.. Tab. 3 indicates the attention mechanisms significantly improve $R_{R,E}$ , while layout of text brings marginal improvement. For C.R., the attention mechanisms is the critical part that promotes the effectiveness of MEE. Instance embedding slightly improves $P_{C,R}$ and $R_{C,R}$ , while type embedding significantly enhances both, indicating that vector types help the model establish rule-lane correspondence. The separate evaluation results of all lane types and the strategies of heuristic method can be found in the supplement. + +# 6. End-to-End Approach + +Recently, there has been notable progress in the development of Multimodal Large Language Models (MLLMs) [3, 5, 39, 49], demonstrating significant capabilities in processing multimodal data and acquiring implicit knowledge of modality alignment. Their large-scale pre-training endows them with exceptional understanding and generative abilities, thereby enabling the comprehensive execution of the task proposed by MapDR in an end-to-end manner. This section aims to address the overall task using Qwen-VL [5] and to investigate three distinct vector encoding methods. Moreover, we introduce an innovative RuleVLM model based on MEE, designed to process inputs from three modalities (images, texts, and vectors) and to produce driving rules, which yield favorable results on MapDR. Further elaboration on these concepts is presented in subsequent sections. + +# 6.1. Architecture + +As shown in Fig. 8, we classify MLLM methods into three types: Qwen-VL(TextPrompt), Qwen-VL(VisualPrompt), and RuleVLM, based on different vector encoding methods. These methods utilize the framework of Qwen-VL [5], with ViT [15] serving as the image encoder. Qwen-VL(TextPrompt) encodes centerline coordinates in the PV image as text, inputs them into QwenLM; Qwen-VL(VisualPrompt) visualizes the centerline and its index in the PV image and inputs it as Visual Prompt into QwenLM, while RuleVLM uses MEE with the cross-attention layer removed to encode vectorized centerline results and aligns it with LLM through an adapter. The goal of all these methods is to output a serialized set of driving rules in text form, decode it through a json decoder to restore formatted data, and then integrate it into HD maps for further applications. + +Table 2. Evaluation of the overall task. The heuristic method and the Qwen-VL(TextPrompt) method serve as the baselines for the modular and end-to-end approach, respectively. " - " denotes end-to-end approach is not suitable to independent evaluations of C.R. because these approaches do not utilize ground truth of rules for correspondence reasoning independently. + +
ModelTypeR.E.C.R.Overall
PR.E.(%)RR.E.(%)PC.R.(%)RC.R.(%)Pall(%)Rall(%)F1 Score
Heuristic18.0111.5133.0517.995.012.730.035
ALBEF-BERTModular75.7857.564.1417.250.240.780.003
VLE-MEE76.6774.5478.0582.1663.3567.370.653
Qwen-VL(TextPrompt)42.2141.09--8.398.170.083
Qwen-VL(VisualPrompt)End-to-End89.2989.50--39.1439.230.392
RuleVLM89.2889.44--64.1664.280.642
+ +Table 3. Evaluation of sub-tasks. "Attn." indicates intra & inter-instance attention mechanisms. "Layout" refers to the text layout applied in VLE. "In.E." and "Ty.E." denotes instance and type embedding in MEE, respectively. + +
VLEPR.E.(%)RR.E.(%)
Attn.Layout
XX75.7857.56
✓X76.8671.75
✓✓76.6774.54
MEEPC.R.(%)RC.R.(%)
Attn.In.E.Ty.E.
XXX4.1417.25
✓XX68.9171.39
✓✓X69.6872.76
✓✓✓78.0582.16
+ +# 6.2. Implementation + +The JSON-formatted driving rules in MapDR are serialized into text and used as the answers in the QA format for training and evaluation. For all methods based on QwenVL, the pre-trained weights of Qwen-VL-Chat (9.6B) are used for initialization. For method using MEE, the trained weights of MEE from the modular approach are utilized for its initialization. During the supervised fine-tuning, we perform LoRA [21] fine-tuning of the QwenLM, while keeping the parameters of ViT frozen. For RuleVLM, we employ a linear layer as the adapter between QwenLM and MEE, while the parameters of MEE and the adapter remains trainable. We use special tokens to separate the centerline tokens extracted by MEE from the visual and text tokens. Additionally, to prevent overfitting, we shuffle the order of centerlines in the local vectorized map during training. + +# 6.3. Experiment. + +Setup. All experiments are conducted on QA formatted MapDR dataset with 20 epochs of training on NVIDIA RTX + +A6000. The dimension of hidden states remains 4096 while $r$ and alpha of LoRA configuration are set 64 and 16 respectively. Meanwhile training batch size is 6 and the max training token length remains 2048. + +Results. We conducted zero-shot evaluations on a subset of MapDR using existing MLLMs [3, 5, 39, 49] as qualitative investigation, but we encounter challenges in achieving stable structured outputs, further details are provided in the supplement. Additionally, as shown in the Tab. 2, Qwen-VL(TextPrompt) method serves as a end-to-end baseline do not perform better on both the R.E. and overall evaluation. Textual representations of centerline coordinates are challenging for MLLMs to perform spatial reasoning and also resulted in excessive length of sequence, which affected $P_{R.E.}$ and $R_{R.E.}$ . Qwen-VL(VisualPrompt) visualizes the centerline within the image, resulting in reduced text sequence length, and demonstrates notable enhancements in the R.E. task compared to the baseline. However, the overall metrics are suboptimal, suggesting that image-based spatial reasoning remains an exceedingly challenging endeavor. The RuleVLM method that we propose introduces MEE for the independent encoding of vectorized HD maps information. This facilitates a more direct extraction of centerline features, resulting in significant enhancements in the overall metrics. + +# 7. Conclusion + +In this work, we defined the task of integrating traffic regulations into vectorized HD maps to facilitate the construction of the traffic regulation layer within online HD maps. We constructed the MapDR dataset, which includes over 10,000 video clips, more than 400,000 images, and at least 18,000 driving rules, covering a total mileage of approximately $800Km$ . Additionally, we proposed the modular approach MEE-VLE and the end-to-end approach RuleVLM, establishing effective baselines for further research. We hope that MapDR can serve as a starting point, and that more researchers can contribute to expanding this dataset or proposing even better solutions. + +# References + +[1] Navigation data standard. https://nds-association.org/.4 +[2] Openstreetmap. https://github.com/openmaptiles/openmaptiles.4 +[3] Anthropic. Claude-3. https://www.anthropic.com/news/claudi-3-family, 2024.7,8 +[4] Stanislaw Antol, Aishwarya Agrawal, Jiasen Lu, Margaret Mitchell, Dhruv Batra, C. Lawrence Zitnick, and Devi Parikh. VQA: visual question answering. In ICCV, 2015. 3 +[5] Jinze Bai, Shuai Bai, Shusheng Yang, Shijie Wang, Sinan Tan, Peng Wang, Junyang Lin, Chang Zhou, and Jingren Zhou. Qwen-vl: A frontier large vision-language model with versatile abilities. arXiv preprint arXiv:2308.12966, 2023.7, 8 +[6] Zechen Bai, Pichao Wang, Tianjun Xiao, Tong He, Zongbo Han, Zheng Zhang, and Mike Zheng Shou. Hallucination of multimodal large language models: A survey. arXiv preprint arXiv:2404.18930, 2024. 3 +[7] Hangbo Bao, Wenhui Wang, Li Dong, Qiang Liu, Owais Khan Mohammed, Kriti Aggarwal, Subhojit Som, Songhao Piao, and Furu Wei. Vlmo: Unified vision-language pre-training with mixture-of-modality-experts. In NeurIPS, 2022. 5 +[8] Karsten Behrendt, Libor Novak, and Rami Botros. A deep learning approach to traffic lights: Detection, tracking, and classification. In ICRA, 2017. 2 +[9] Holger Caesar, Varun Bankiti, Alex H. Lang, Sourabh Vora, Venice Erin Liong, Qiang Xu, Anush Krishnan, Yu Pan, Giancarlo Baldan, and Oscar Beijbom. nuscenes: A multimodal dataset for autonomous driving. In CVPR, 2020. 2 +[10] Xu Cao, Tong Zhou, Yunsheng Ma, Wenqian Ye, Can Cui, Kun Tang, Zhipeng Cao, Kaizhao Liang, Ziran Wang, James Rehg, and Chao Zheng. Maplm: A real-world large-scale vision-language dataset for map and traffic scene understanding. In CVPR, 2024. 3 +[11] Li Chen, Chonghao Sima, Yang Li, Zehan Zheng, Jiajie Xu, Xiangwei Geng, Hongyang Li, Conghui He, Jianping Shi, Yu Qiao, and Junchi Yan. Persformer: 3d lane detection via perspective transformer and the openlane benchmark. In ECCV, 2022. 2 +[12] Long Chen, Oleg Sinavski, Jan Hunermann, Alice Karnsund, Andrew James Willmott, Danny Birch, Daniel Maund, and Jamie Shotton. Driving with llms: Fusing object-level vector modality for explainable autonomous driving. In ICRA, 2024. 3 +[13] Yaodong Cui, Shucheng Huang, Jiaming Zhong, Zhenan Liu, Yutong Wang, Chen Sun, Bai Li, Xiao Wang, and Amir Khajepour. Drivellm: Charting the path toward full autonomous driving with large language models. IEEE TIV, 2024. 3 +[14] Jacob Devlin, Ming-Wei Chang, Kenton Lee, and Kristina Toutanova. BERT: pre-training of deep bidirectional transformers for language understanding. In NAACL, 2019. 5, 6, 7 +[15] Alexey Dosovitskiy, Lucas Beyer, Alexander Kolesnikov, Dirk Weissenborn, Xiaohua Zhai, Thomas Unterthiner, + +Mostafa Dehghani, Matthias Minderer, Georg Heigold, Sylvain Gelly, Jakob Uszkoreit, and Neil Houlsby. An image is worth 16x16 words: Transformers for image recognition at scale. In ICLR, 2021. 5, 7 +[16] Gamal Elghazaly, Raphaël Frank, Scott Harvey, and Stefan Safko. High-definition maps: Comprehensive survey, challenges, and future perspectives. IEEE Open Journal of Intelligent Transportation Systems, 2023. 2, 4 +[17] Andreas Fregin, Julian Müller, Ulrich Krebel, and Klaus Dietmayer. The driveu traffic light dataset: Introduction and comparison with existing datasets. In ICRA, 2018. 2 +[18] Shuo Gu, Yigong Zhang, Jinhui Tang, Jian Yang, and Hui Kong. Road detection through CRF based lidar-camera fusion. In ICRA, 2019. 2 +[19] Yunfei Guo, Wei Feng, Fei Yin, Tao Xue, Shuqi Mei, and Cheng-Lin Liu. Learning to understand traffic signs. In ACMMM, 2021. 2, 3 +[20] Yunfei Guo, Fei Yin, Xiao-Hui Li, Xudong Yan, Tao Xue, Shuqi Mei, and Cheng-Lin Liu. Visual traffic knowledge graph generation from scene images. In ICCV, 2023. 2, 3 +[21] Edward J. Hu, Yelong Shen, Phillip Wallis, Zeyuan Allen-Zhu, Yuanzhi Li, Shean Wang, Lu Wang, and Weizhu Chen. Lora: Low-rank adaptation of large language models. In The Tenth International Conference on Learning Representations, ICLR 2022, Virtual Event, April 25-29, 2022, 2022. 8 +[22] Xinyu Huang, Peng Wang, Xinjing Cheng, Dingfu Zhou, Qichuan Geng, and Ruigang Yang. The apolloscope open dataset for autonomous driving and its application. IEEE TPAMI, 2020. 2 +[23] Yupan Huang, Tengchao Lv, Lei Cui, Yutong Lu, and Furu Wei. Layoutmv3: Pre-training for document AI with unified text and image masking. In ACMMM, 2022. 3 +[24] Chao Jia, Yinfei Yang, Ye Xia, Yi-Ting Chen, Zarana Parekh, Hieu Pham, Quoc V. Le, Yun-Hsuan Sung, Zhen Li, and Tom Duerig. Scaling up visual and vision-language representation learning with noisy text supervision. In ICML, 2021. 3 +[25] Wonjae Kim, Bokyung Son, and Ildoo Kim. Vilt: Vision-and-language transformer without convolution or region supervision. In ICML, 2021. 3, 5 +[26] Junnan Li, Ramprasaath R. Selvaraju, Akhilesh Gotmare, Shafiq R. Joty, Caiming Xiong, and Steven Chu-Hong Hoi. Align before fuse: Vision and language representation learning with momentum distillation. In NeurIPS, 2021. 3, 5, 7 +[27] Junnan Li, Dongxu Li, Caiming Xiong, and Steven C. H. Hoi. BLIP: bootstrapping language-image pre-training for unified vision-language understanding and generation. In ICML, 2022. 5 +[28] Junnan Li, Dongxu Li, Silvio Savarese, and Steven C. H. Hoi. BLIP-2: bootstrapping language-image pre-training with frozen image encoders and large language models. In ICML, 2023. 3 +[29] Qi Li, Yue Wang, Yilun Wang, and Hang Zhao. Hdmapnet: An online HD map construction and evaluation framework. In ICRA, 2022. 2 +[30] Benchcheng Liao, Shaoyu Chen, Xinggang Wang, Tianheng Cheng, Qian Zhang, Wenyu Liu, and Chang Huang. Maptr: + +Structured modeling and learning for online vectorized HD map construction. In ICLR, 2023. 2 +[31] Benchcheng Liao, Shaoyu Chen, Yunchi Zhang, Bo Jiang, Qian Zhang, Wenyu Liu, Chang Huang, and Xinggang Wang. Maptrv2: An end-to-end framework for online vectorized hd map construction. arXiv preprint arXiv:2308.05736, 2023. 3, 4, 5, 6 +[32] Haotian Liu, Chunyuan Li, Yuheng Li, and Yong Jae Lee. Improved baselines with visual instruction tuning. arXiv preprint arXiv:2310.03744, 2023. 3 +[33] Haotian Liu, Chunyuan Li, Qingyang Wu, and Yong Jae Lee. Visual instruction tuning. In NeurIPS, 2023. +[34] Haotian Liu, Chunyuan Li, Yuheng Li, Bo Li, Yuanhan Zhang, Sheng Shen, and Yong Jae Lee. Llava-last: Improved reasoning,OCR,and world knowledge.https://llava-v1.github.io/blog/2024-01-30-llava-last, 2024.3 +[35] Rong Liu, Jinling Wang, and Bingqi Zhang. High definition map for automated driving: Overview and analysis. Journal of Navigation, 2020. 2, 4 +[36] Yicheng Liu, Tianyuan Yuan, Yue Wang, Yilun Wang, and Hang Zhao. Vectormapnet: End-to-end vectorized HD map learning. In ICML, 2023. 2 +[37] Jiageng Mao, Yuxi Qian, Hang Zhao, and Yue Wang. Gptdriver: Learning to drive with GPT. In NeurIPS Workshop, 2023. 3 +[38] Ana-Maria Marcu, Long Chen, Jan Hunermann, Alice Karnsund, Benoit Hanotte, Prajwal Chidananda, Saurabh Nair, Vijay Badrinarayanan, Alex Kendall, Jamie Shotton, and Oleg Sinavski. Lingoqa: Video question answering for autonomous driving. arXiv preprint arXiv:2312.14115, 2023. 3 +[39] OpenAI. GPT-4 technical report. arXiv preprint arXiv:2303.08774, 2024. 7, 8 +[40] Jonah Philion and Sanja Fidler. Lift, splat, shoot: Encoding images from arbitrary camera rigs by implicitly unprojecting to 3d. In ECCV, 2020. 2 +[41] Tianwen Qian, Jingjing Chen, Linhai Zhuo, Yang Jiao, and Yu-Gang Jiang. Nuscenes-qa: A multi-modal visual question answering benchmark for autonomous driving scenario. In AAAI, 2024. 3 +[42] Alec Radford, Jong Wook Kim, Chris Hallacy, Aditya Ramesh, Gabriel Goh, Sandhini Agarwal, Girish Sastry, Amanda Askell, Pamela Mishkin, Jack Clark, Gretchen Krueger, and Ilya Sutskever. Learning transferable visual models from natural language supervision. In ICML, 2021. 3, 5 +[43] Enna Sachdeva, Nakul Agarwal, Suhas Chundi, Sean Roelofs, Jiachen Li, Mykel J. Kochenderfer, Chiho Choi, and Behzad Dariush. Rank2tell: A multimodal driving dataset for joint importance ranking and reasoning. In WACV, 2024. 3 +[44] Chonghao Sima, Katrin Renz, Kashyap Chitta, Li Chen, Hanxue Zhang, Chengen Xie, Ping Luo, Andreas Geiger, and Hongyang Li. Drivelm: Driving with graph visual question answering. arXiv preprint arXiv:2312.14150, 2023. 3 +[45] Chonghao Sima, Katrin Renz, Kashyap Chitta, Li Chen, Hanxue Zhang, Chengen Xie, Jens Beiwenger, Ping Luo, Andreas Geiger, and Hongyang Li. Drivelm: Driving with graph visual question answering. In ECCV, 2024. 3 + +[46] Johannes Stallkamp, Marc Schlipsing, Jan Salmen, and Christian Igel. Man vs. computer: Benchmarking machine learning algorithms for traffic sign recognition. *Neural Networks*, 2012. 2 +[47] Pei Sun, Henrik Kretzschmar, Xerxes Dotiwalla, Aurelien Chouard, Vijaysai Patnaik, Paul Tsui, James Guo, Yin Zhou, Yuning Chai, Benjamin Caine, Vijay Vasudevan, Wei Han, Jiquan Ngiam, Hang Zhao, Aleksei Timofeev, Scott Ettinger, Maxim Krivokon, Amy Gao, Aditya Joshi, Yu Zhang, Jonathon Shlens, Zhifeng Chen, and Dragomir Anguelov. Scalability in perception for autonomous driving: Waymo open dataset. In CVPR, 2020. 2 +[48] Matthew Tancik, Pratul P. Srinivasan, Ben Mildenhall, Sara Fridovich-Keil, Nithin Raghavan, Utkarsh Singhal, Ravi Ramamoorthi, Jonathan T. Barron, and Ren Ng. Fourier features let networks learn high frequency functions in low dimensional domains. In NeurIPS, 2020. 5 +[49] Gemini Team, Rohan Anil, Sebastian Borgeaud, Yonghui Wu, Jean-Baptiste Alayrac, Jiahui Yu, Radu Soricut, Johan Schalkwyk, Andrew M. Dai, Anja Hauth, and et al. Gemini: A family of highly capable multimodal models. arXiv preprint arXiv:2312.11805, 2024. 7, 8 +[50] Hugo Touvron, Matthieu Cord, Matthijs Douze, Francisco Massa, Alexandre Sablayrolles, and Hervé Jégou. Training data-efficient image transformers & distillation through attention. In ICML, 2021. 7 +[51] Ashish Vaswani, Noam Shazeer, Niki Parmar, Jakob Uszkoreit, Llion Jones, Aidan N. Gomez, Lukasz Kaiser, and Illia Polosukhin. Attention is all you need. In NeurIPS, 2017. 3, 5 +[52] Huijie Wang, Tianyu Li, Yang Li, Li Chen, Chonghao Sima, Zhenbo Liu, Bangjun Wang, Peijin Jia, Yuting Wang, Shengyin Jiang, Feng Wen, Hang Xu, Ping Luo, Junchi Yan, Wei Zhang, and Hongyang Li. Openlane-v2: A topology reasoning benchmark for unified 3d HD mapping. In NeurIPS, 2023. 2, 3 +[53] Benjamin Wilson, William Qi, Tanmay Agarwal, John Lambert, Jagjeet Singh, Siddhesh Khandelwal, Bowen Pan, Ratnesh Kumar, Andrew Hartnett, Jhony Kaesemodel Pontes, Deva Ramanan, Peter Carr, and James Hays. Argoverse 2: Next generation datasets for self-driving perception and forecasting. In NeurIPS, 2021. 2 +[54] Dongming Wu, Jiahao Chang, Fan Jia, Yingfei Liu, Tiancai Wang, and Jianbing Shen. Topomlp: A simple yet strong pipeline for driving topology reasoning. In ICLR, 2024. 2 +[55] Yiheng Xu, Minghao Li, Lei Cui, Shaohan Huang, Furu Wei, and Ming Zhou. Layoutm: Pre-training of text and layout for document image understanding. In KDD, 2020. 3 +[56] Yang Xu, Yiheng Xu, Tengchao Lv, Lei Cui, Furu Wei, Guoxin Wang, Yijuan Lu, Dinei A. F. Floréncio, Cha Zhang, Wanxiang Che, Min Zhang, and Lidong Zhou. Layoutmv2: Multi-modal pre-training for visually-rich document understanding. In ACL, 2021. 3 +[57] Zhenhua Xu, Yujia Zhang, Enze Xie, Zhen Zhao, Yong Guo, Kwan-Yee K. Wong, Zhenguo Li, and Hengshuang Zhao. Drivegpt4: Interpretable end-to-end autonomous driving via large language model. IEEE Robotics Autom. Lett., 2024. 3 +[58] Fisher Yu, Haofeng Chen, Xin Wang, Wenqi Xian, Yingying Chen, Fangchen Liu, Vashisht Madhavan, and Trevor Dar + +rell. BDD100K: A diverse driving dataset for heterogeneous multitask learning. In CVPR, 2020. 2 +[59] Jiahui Yu, Zirui Wang, Vijay Vasudevan, Legg Yeung, Mojtaba Seyedhosseini, and Yonghui Wu. Coca: Contrastive captioners are image-text foundation models. TMLR, 2022. 3 +[60] Zhe Zhu, Dun Liang, Song-Hai Zhang, Xiaolei Huang, Baoli Li, and Shi-Min Hu. Traffic-sign detection and classification in the wild. In CVPR, 2016. 2 \ No newline at end of file diff --git a/2025/Driving by the Rules_ A Benchmark for Integrating Traffic Sign Regulations into Vectorized HD Map/images.zip b/2025/Driving by the Rules_ A Benchmark for Integrating Traffic Sign Regulations into Vectorized HD Map/images.zip new file mode 100644 index 0000000000000000000000000000000000000000..31c72c0c18a9d97ca742b22ec23136213cad9dc1 --- /dev/null +++ b/2025/Driving by the Rules_ A Benchmark for Integrating Traffic Sign Regulations into Vectorized HD Map/images.zip @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:c013c5c92708d84d843a8abe390d78cdae0cabb53d10d9cdcc2b3886e1a47fc7 +size 590678 diff --git a/2025/Driving by the Rules_ A Benchmark for Integrating Traffic Sign Regulations into Vectorized HD Map/layout.json b/2025/Driving by the Rules_ A Benchmark for Integrating Traffic Sign Regulations into Vectorized HD Map/layout.json new file mode 100644 index 0000000000000000000000000000000000000000..8060f4e87d2c73e32d2d63fc1715af69a2f83b9c --- /dev/null +++ b/2025/Driving by the Rules_ A Benchmark for Integrating Traffic Sign Regulations into Vectorized HD Map/layout.json @@ -0,0 +1,8539 @@ +{ + "pdf_info": [ + { + "para_blocks": [ + { + "bbox": [ + 56, + 102, + 555, + 140 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 102, + 555, + 140 + ], + "spans": [ + { + "bbox": [ + 56, + 102, + 555, + 140 + ], + "type": "text", + "content": "Driving by the Rules: A Benchmark for Integrating Traffic Sign Regulations into Vectorized HD Map" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 82, + 160, + 533, + 218 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 82, + 160, + 533, + 218 + ], + "spans": [ + { + "bbox": [ + 82, + 160, + 533, + 218 + ], + "type": "text", + "content": "Xinyuan Chang " + }, + { + "bbox": [ + 82, + 160, + 533, + 218 + ], + "type": "inline_equation", + "content": "^{1*}" + }, + { + "bbox": [ + 82, + 160, + 533, + 218 + ], + "type": "text", + "content": " Maixuan Xue " + }, + { + "bbox": [ + 82, + 160, + 533, + 218 + ], + "type": "inline_equation", + "content": "^{2*†}" + }, + { + "bbox": [ + 82, + 160, + 533, + 218 + ], + "type": "text", + "content": " Xinran Liu " + }, + { + "bbox": [ + 82, + 160, + 533, + 218 + ], + "type": "inline_equation", + "content": "^{1}" + }, + { + "bbox": [ + 82, + 160, + 533, + 218 + ], + "type": "text", + "content": " Zheng Pan " + }, + { + "bbox": [ + 82, + 160, + 533, + 218 + ], + "type": "inline_equation", + "content": "^{1}" + }, + { + "bbox": [ + 82, + 160, + 533, + 218 + ], + "type": "text", + "content": " Xing Wei " + }, + { + "bbox": [ + 82, + 160, + 533, + 218 + ], + "type": "inline_equation", + "content": "^{2‡}" + }, + { + "bbox": [ + 82, + 160, + 533, + 218 + ], + "type": "inline_equation", + "content": "^{1}" + }, + { + "bbox": [ + 82, + 160, + 533, + 218 + ], + "type": "text", + "content": "Amap, Alibaba Group " + }, + { + "bbox": [ + 82, + 160, + 533, + 218 + ], + "type": "inline_equation", + "content": "^{2}" + }, + { + "bbox": [ + 82, + 160, + 533, + 218 + ], + "type": "text", + "content": "Xi’an Jiaotong University \n{changxinyuan.cxy, tom.lxr, panzheng.pan}@alibaba-inc.com \nxmx0809@stu.xjtu.edu.cn, weixing@mail.xjtu.edu.cn" + } + ] + } + ], + "index": 3 + }, + { + "type": "image", + "bbox": [ + 86, + 262, + 420, + 430 + ], + "blocks": [ + { + "bbox": [ + 86, + 262, + 420, + 430 + ], + "lines": [ + { + "bbox": [ + 86, + 262, + 420, + 430 + ], + "spans": [ + { + "bbox": [ + 86, + 262, + 420, + 430 + ], + "type": "image", + "image_path": "fa98e39510a06b77ea1def57287ac98a0b8b5c58fe12f39b937e421ac5f465cb.jpg" + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 54, + 449, + 555, + 495 + ], + "lines": [ + { + "bbox": [ + 54, + 449, + 555, + 495 + ], + "spans": [ + { + "bbox": [ + 54, + 449, + 555, + 495 + ], + "type": "text", + "content": "Figure 1. MapDR Overview and Motivation. The image on the left depicts the comprehensive results of HD map construction for an intersection scene, whereas the image on the right illustrates the outcomes after partitioning it into three layers. Existing online mapping methods primarily emphasize the construction of the geometric and connectivity layers, neglecting the traffic regulation layer. However, precise comprehension of traffic signs and their correlation with lanes is vital for ensuring the safety of autonomous driving." + } + ] + } + ], + "index": 11, + "angle": 0, + "type": "image_caption" + } + ], + "index": 4 + }, + { + "type": "image", + "bbox": [ + 429, + 263, + 526, + 312 + ], + "blocks": [ + { + "bbox": [ + 429, + 263, + 526, + 312 + ], + "lines": [ + { + "bbox": [ + 429, + 263, + 526, + 312 + ], + "spans": [ + { + "bbox": [ + 429, + 263, + 526, + 312 + ], + "type": "image", + "image_path": "df3966c6489e0b02b577968def196aff7d0c31af7f128e2738124bf701658d61.jpg" + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 450, + 312, + 507, + 321 + ], + "lines": [ + { + "bbox": [ + 450, + 312, + 507, + 321 + ], + "spans": [ + { + "bbox": [ + 450, + 312, + 507, + 321 + ], + "type": "text", + "content": "Geometric layer" + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "image_caption" + } + ], + "index": 5 + }, + { + "type": "image", + "bbox": [ + 430, + 323, + 525, + 370 + ], + "blocks": [ + { + "bbox": [ + 430, + 323, + 525, + 370 + ], + "lines": [ + { + "bbox": [ + 430, + 323, + 525, + 370 + ], + "spans": [ + { + "bbox": [ + 430, + 323, + 525, + 370 + ], + "type": "image", + "image_path": "f005aea12fc3690efb0eaff4f4543bcdbb88f79353f3d6c94bdf59d688586855.jpg" + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 447, + 370, + 512, + 379 + ], + "lines": [ + { + "bbox": [ + 447, + 370, + 512, + 379 + ], + "spans": [ + { + "bbox": [ + 447, + 370, + 512, + 379 + ], + "type": "text", + "content": "Connectivity layer" + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "image_caption" + } + ], + "index": 7 + }, + { + "type": "image", + "bbox": [ + 430, + 380, + 525, + 428 + ], + "blocks": [ + { + "bbox": [ + 430, + 380, + 525, + 428 + ], + "lines": [ + { + "bbox": [ + 430, + 380, + 525, + 428 + ], + "spans": [ + { + "bbox": [ + 430, + 380, + 525, + 428 + ], + "type": "image", + "image_path": "218888cb25494023aaae06c9ff7b5f0bb0ce06a5e6579921a59ae6257f422520.jpg" + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 440, + 429, + 520, + 438 + ], + "lines": [ + { + "bbox": [ + 440, + 429, + 520, + 438 + ], + "spans": [ + { + "bbox": [ + 440, + 429, + 520, + 438 + ], + "type": "text", + "content": "Traffic regulation layer" + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "image_caption" + } + ], + "index": 9 + }, + { + "bbox": [ + 151, + 510, + 200, + 523 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 151, + 510, + 200, + 523 + ], + "spans": [ + { + "bbox": [ + 151, + 510, + 200, + 523 + ], + "type": "text", + "content": "Abstract" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 55, + 539, + 297, + 671 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 539, + 297, + 671 + ], + "spans": [ + { + "bbox": [ + 55, + 539, + 297, + 671 + ], + "type": "text", + "content": "Ensuring adherence to traffic sign regulations is essential for both human and autonomous vehicle navigation. While current online mapping solutions often prioritize the construction of the geometric and connectivity layers of HD maps, overlooking the construction of the traffic regulation layer within HD maps. Addressing this gap, we introduce MapDR, a novel dataset designed for the extraction of Driving Rules from traffic signs and their association with vectorized, locally perceived HD Maps. MapDR features over 10,000 annotated video clips that capture the intricate correlation between traffic sign regulations and" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 313, + 511, + 556, + 608 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 511, + 556, + 608 + ], + "spans": [ + { + "bbox": [ + 313, + 511, + 556, + 608 + ], + "type": "text", + "content": "lanes. Built upon this benchmark and the newly defined task of integrating traffic regulations into online HD maps, we provide modular and end-to-end solutions: VLE-MEE and RuleVLM, offering a strong baseline for advancing autonomous driving technology. It fills a critical gap in the integration of traffic sign rules, contributing to the development of reliable autonomous driving systems. Code is available at https://github.com/MIV-XJTU/MapDR." + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 314, + 633, + 395, + 645 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 314, + 633, + 395, + 645 + ], + "spans": [ + { + "bbox": [ + 314, + 633, + 395, + 645 + ], + "type": "text", + "content": "1. Introduction" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 313, + 654, + 556, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 654, + 556, + 715 + ], + "spans": [ + { + "bbox": [ + 313, + 654, + 556, + 715 + ], + "type": "text", + "content": "The rise of autonomous vehicles and intelligent transportation systems has emphasized the importance of accurate, reliable navigational data. High-Definition (HD) maps, with their detailed representations of road elements, have become essential to supporting these systems. An HD map typically" + } + ] + } + ], + "index": 16 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "spans": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "type": "text", + "content": "CVF" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 145, + 0, + 496, + 37 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 145, + 0, + 496, + 37 + ], + "spans": [ + { + "bbox": [ + 145, + 0, + 496, + 37 + ], + "type": "text", + "content": "This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 66, + 685, + 135, + 694 + ], + "type": "page_footnote", + "angle": 0, + "lines": [ + { + "bbox": [ + 66, + 685, + 135, + 694 + ], + "spans": [ + { + "bbox": [ + 66, + 685, + 135, + 694 + ], + "type": "text", + "content": "*Equal contribution." + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 67, + 695, + 258, + 704 + ], + "type": "page_footnote", + "angle": 0, + "lines": [ + { + "bbox": [ + 67, + 695, + 258, + 704 + ], + "spans": [ + { + "bbox": [ + 67, + 695, + 258, + 704 + ], + "type": "inline_equation", + "content": "^{\\dagger}" + }, + { + "bbox": [ + 67, + 695, + 258, + 704 + ], + "type": "text", + "content": "Work done during the internship at Amap, Alibaba Group." + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 67, + 704, + 144, + 714 + ], + "type": "page_footnote", + "angle": 0, + "lines": [ + { + "bbox": [ + 67, + 704, + 144, + 714 + ], + "spans": [ + { + "bbox": [ + 67, + 704, + 144, + 714 + ], + "type": "inline_equation", + "content": "\\ddagger" + }, + { + "bbox": [ + 67, + 704, + 144, + 714 + ], + "type": "text", + "content": " Corresponding author." + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "text", + "content": "6823" + } + ] + } + ], + "index": 21 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 0 + }, + { + "para_blocks": [ + { + "bbox": [ + 55, + 72, + 295, + 167 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 72, + 295, + 167 + ], + "spans": [ + { + "bbox": [ + 55, + 72, + 295, + 167 + ], + "type": "text", + "content": "consists of three primary layers: the geometric layer, connectivity layer, and traffic regulation layer [16, 35], as shown in Fig. 1. The geometric layer provides vector information, such as dividers and lane centerlines; the connectivity layer defines lane relationships to assist in path planning; and the traffic regulation layer encodes rule information (e.g., HOV lanes, bus lanes, speed-limit zones) associated with lanes, ensuring compliant driving." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 55, + 169, + 296, + 277 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 169, + 296, + 277 + ], + "spans": [ + { + "bbox": [ + 55, + 169, + 296, + 277 + ], + "type": "text", + "content": "A limitation of conventional HD maps is their inability to support rapid, real-time updates, making the construction of online HD maps an emerging trend. Existing methods, such as MapTR [30] and TopoMLP [54], address only the geometric and connectivity layers, overlooking the traffic regulation layer. As a result, autonomous driving systems currently still rely on offline maps to obtain traffic regulations, an approach that runs counter to the trend of online HD map construction." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 55, + 279, + 296, + 447 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 279, + 296, + 447 + ], + "spans": [ + { + "bbox": [ + 55, + 279, + 296, + 447 + ], + "type": "text", + "content": "Traffic signs serve as a \"visual language\" on the road, playing a critical role in defining traffic regulations. Inspired by human driving processes, the creation of traffic regulations can be divided into two stages: 1) understanding the rules conveyed by traffic signs, and 2) determining the lanes to which these rules apply. Although OpenLane-V2 [52] has made strides in associating signs with lanes in online mapping, it has two primary limitations. First, it considers only directional signs, neglecting other types of regulations. Second, its label annotations are limited to sign categories, falling short of the detailed rule descriptions required for autonomous driving. A more structured description, in alignment with HD map standards, is essential for supporting comprehensive autonomous driving applications [16, 35]." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 55, + 449, + 296, + 590 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 449, + 296, + 590 + ], + "spans": [ + { + "bbox": [ + 55, + 449, + 296, + 590 + ], + "type": "text", + "content": "To bridge this gap, we introduce MapDR, the first dataset specifically designed for the challenging task of integrating traffic regulations into existing online HD maps. MapDR focuses on complex traffic scenes with diverse signage, collecting over 10,000 video clips across three representative cities in China, with each clip containing at least one traffic sign. In addition to vectorized representations of dividers, boundaries, and centerlines, MapDR provides structural annotations of traffic regulations and their associations with lanes, offering a comprehensive foundation for advancing online HD maps. More details of MapDR can be found in Sec. 4." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 55, + 594, + 296, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 594, + 296, + 713 + ], + "spans": [ + { + "bbox": [ + 55, + 594, + 296, + 713 + ], + "type": "text", + "content": "Based on MapDR, in addition to the end-to-end task definition, we also divide this task into two innovative subtasks aimed at bolstering research in this domain: 1) rule extraction from traffic signs, and 2) rule-lane correspondence reasoning. We introduce Vision-Language Encoder (VLE) and Map Element Encoder (MEE) to handle the interaction of multimodal data, encompassing images, texts, and vectors. Through the concatenation of these two models, integrated traffic regulations can be obtained. Additionally, inspired by recent advancements in Multimodal Large Lan" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 313, + 70, + 555, + 148 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 70, + 555, + 148 + ], + "spans": [ + { + "bbox": [ + 313, + 70, + 555, + 148 + ], + "type": "text", + "content": "Table 1. Comparison of the existing datasets. \"Sign\" and \"Lane\" denote whether the dataset focuses on traffic signs and lanes. Only those annotated with formatted (\"Fmt.\") rules and the correspondence (\"Corr.\") between rules and lanes can form driving rules. \"Clip\" represents whether the data is organized in the form of video clips. \"* denotes that these samples are not newly collected and are built upon the previous dataset." + } + ] + } + ], + "index": 5 + }, + { + "type": "table", + "bbox": [ + 316, + 156, + 553, + 240 + ], + "blocks": [ + { + "bbox": [ + 316, + 156, + 553, + 240 + ], + "lines": [ + { + "bbox": [ + 316, + 156, + 553, + 240 + ], + "spans": [ + { + "bbox": [ + 316, + 156, + 553, + 240 + ], + "type": "table", + "html": "
DatasetSignLaneDriving RulesNumber of SamplesYear
Fmt.Corr.ImageClipRegion
nuScenes [9]✓1400K1KWorldwide2019
Argoverse2 [53]✓2100K1KUSA2021
CTSU [19]✓5K/China2021
OpenLane [11]✓200K*1K*Worldwide2022
RS10K [20]✓✓10K/China2023
OpenLane-V2 [52]✓✓✓466K*2K*Worldwide2023
Waymo [47]✓✓390K2KUSA2024
MapDR(ours)✓✓✓✓400K10KChina2024
", + "image_path": "b25a9b0b908a767c82761e6ed9b6d10682d6595a78579c0a40c19fa8f92f6111.jpg" + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "table_body" + } + ], + "index": 6 + }, + { + "bbox": [ + 313, + 262, + 555, + 334 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 262, + 555, + 334 + ], + "spans": [ + { + "bbox": [ + 313, + 262, + 555, + 334 + ], + "type": "text", + "content": "guage Models (MLLMs), we also explore an end-to-end MLLM solution called RuleVLM to solve this task holistically. These approaches establish robust baselines and aim to inspire future research in this area. For detailed descriptions of the proposed tasks and metrics, please refer to Sec. 3. For approaches, please refer to Secs. 5 and 6." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 326, + 335, + 504, + 346 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 326, + 335, + 504, + 346 + ], + "spans": [ + { + "bbox": [ + 326, + 335, + 504, + 346 + ], + "type": "text", + "content": "To sum up, our contributions are as follows:" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 334, + 350, + 555, + 529 + ], + "type": "list", + "angle": 0, + "index": 12, + "blocks": [ + { + "bbox": [ + 334, + 350, + 555, + 409 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 334, + 350, + 555, + 409 + ], + "spans": [ + { + "bbox": [ + 334, + 350, + 555, + 409 + ], + "type": "text", + "content": "- For the first time, We introduce the novel task of extracting lane-level rules from traffic signs and integrating them into vectorized HD maps, alongside the MapDR dataset and evaluation metrics for benchmarking." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 334, + 410, + 555, + 479 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 334, + 410, + 555, + 479 + ], + "spans": [ + { + "bbox": [ + 334, + 410, + 555, + 479 + ], + "type": "text", + "content": "- MapDR comprises over 10,000 video clips and 400,000 front-view images captured across multiple cities in China, covering diverse traffic conditions and including more than 18,000 annotated lane-level rules. All data are newly collected and meticulously validated." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 334, + 481, + 553, + 529 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 334, + 481, + 553, + 529 + ], + "spans": [ + { + "bbox": [ + 334, + 481, + 553, + 529 + ], + "type": "text", + "content": "- We propose two modeling approaches, modular and end-to-end, to address the integration of lane-level rules into vectorized HD maps, providing effective baselines for future research." + } + ] + } + ], + "index": 11 + } + ], + "sub_type": "text" + }, + { + "bbox": [ + 314, + 544, + 400, + 555 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 314, + 544, + 400, + 555 + ], + "spans": [ + { + "bbox": [ + 314, + 544, + 400, + 555 + ], + "type": "text", + "content": "2. Related Work" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 313, + 563, + 442, + 576 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 563, + 442, + 576 + ], + "spans": [ + { + "bbox": [ + 313, + 563, + 442, + 576 + ], + "type": "text", + "content": "2.1. HD Map Construction" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 313, + 582, + 555, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 582, + 555, + 713 + ], + "spans": [ + { + "bbox": [ + 313, + 582, + 555, + 713 + ], + "type": "text", + "content": "HD maps construction have seen significant advancements, with a focus on traffic element perception, including lane detection and traffic sign recognition [8, 9, 17, 18, 22, 46, 53, 58, 60]. The shift towards BEV perception and vectorization for end-to-end HD maps construction has gained traction [9, 11, 53]. Notable works include HDMapNet, which aggregates semantic segmentation results [29], LSS [40] estimates depth to transfer image features to BEV features, while VectorMapNet [36] is the first end-to-end framework for sequential vector point prediction to generate HD maps without post-processing. MapTR [30] and its enhanced" + } + ] + } + ], + "index": 15 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "text", + "content": "6824" + } + ] + } + ], + "index": 16 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 1 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 82, + 70, + 526, + 182 + ], + "blocks": [ + { + "bbox": [ + 82, + 70, + 526, + 182 + ], + "lines": [ + { + "bbox": [ + 82, + 70, + 526, + 182 + ], + "spans": [ + { + "bbox": [ + 82, + 70, + 526, + 182 + ], + "type": "image", + "image_path": "307ba3b46ca57ad5adb2d0d4c501f5f7f069580f9a1c1f82b612dfb815c84d96.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 54, + 191, + 555, + 215 + ], + "lines": [ + { + "bbox": [ + 54, + 191, + 555, + 215 + ], + "spans": [ + { + "bbox": [ + 54, + 191, + 555, + 215 + ], + "type": "text", + "content": "Figure 2. Overview of the sub-tasks. Step 1 ~ Step 4 shows a case of driving by the rules. Step 2 and Step 3 demonstrates the specific role of two sub-tasks, respectively." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "bbox": [ + 54, + 235, + 297, + 295 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 54, + 235, + 297, + 295 + ], + "spans": [ + { + "bbox": [ + 54, + 235, + 297, + 295 + ], + "type": "text", + "content": "version, MapTRv2 [31], introduced a unified permutation-equivalent modeling approach and extended it to a general framework supporting centerline learning and 3D map construction. However, these efforts have largely overlooked the integration of traffic sign rules into HD maps." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 55, + 301, + 209, + 313 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 301, + 209, + 313 + ], + "spans": [ + { + "bbox": [ + 55, + 301, + 209, + 313 + ], + "type": "text", + "content": "2.2. Traffic Element Association" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 56, + 318, + 296, + 558 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 318, + 296, + 558 + ], + "spans": [ + { + "bbox": [ + 56, + 318, + 296, + 558 + ], + "type": "text", + "content": "Traffic element association aims to link elements like traffic signs with lanes. As demonstrated in Tab. 1, CTSU has initiated internal elements association to describe traffic sign in " + }, + { + "bbox": [ + 56, + 318, + 296, + 558 + ], + "type": "inline_equation", + "content": "\\{key: value\\}" + }, + { + "bbox": [ + 56, + 318, + 296, + 558 + ], + "type": "text", + "content": " form, however lacking both generalization of driving rules from description and lane association [19]. VTKGG [20] propose to utilize a graph model for connectivity but also lacks structured expression of driving rules for motion planning and requires complex integration into HD maps, which is typically expressed in the BEV space. OpenLane-V2 [52] advances traffic sign and lane association in BEV space, but is constrained by single-label classification for traffic sign rather than structured descriptions for fine-grained driving rules. This makes it insufficient for signs with multiple rules, which are complex but common in real scenarios. Recent MLLM-based benchmarks [10, 38, 41, 43, 44] for autonomous driving, such as MAPLM [10], prioritize end-to-end motion planning over precise rule extraction from traffic sign, lacking evaluation for rule reasoning. MapDR addresses this gap by focusing on traffic sign rule extraction and lane association." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 55, + 564, + 195, + 577 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 564, + 195, + 577 + ], + "spans": [ + { + "bbox": [ + 55, + 564, + 195, + 577 + ], + "type": "text", + "content": "2.3. Vision-Language Models" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 55, + 582, + 297, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 582, + 297, + 713 + ], + "spans": [ + { + "bbox": [ + 55, + 582, + 297, + 713 + ], + "type": "text", + "content": "Vision-Language Models (VLMs) facilitates multimodal applications by learning joint representations of vision and language data. Visual Question Answer (VQA) tasks provide answers to image-related questions [4], while Visual Information Extraction (VIE) tasks extract structured information from visual and textual data [4, 23, 55, 56]. In Autonomous Driving (AD), VLMs are increasingly used for comprehensive traffic scene understanding and decision-making. The field has seen various approaches, including using transformers [51] for joint encoding [23, 25], excelling at multimodal information interaction, and independent en" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 313, + 235, + 556, + 355 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 235, + 556, + 355 + ], + "spans": [ + { + "bbox": [ + 313, + 235, + 556, + 355 + ], + "type": "text", + "content": "coders for different modalities [24, 42] that are proficient in multimodal retrieval. Cross-modal representation methods [26, 59] combine these advantages, and the latest LLM-based research [28, 32-34] has achieved state-of-the-art results in various multimodal tasks. Nowadays, an increasing number of methods are leveraging LLMs to achieve impressive results [12, 37, 45, 57], with works like DriveLLM [13] showing significant potential in AD. However, addressing hallucination [6] remains the most crucial aspect for LLM-based approaches." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 313, + 370, + 555, + 399 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 370, + 555, + 399 + ], + "spans": [ + { + "bbox": [ + 313, + 370, + 555, + 399 + ], + "type": "text", + "content": "3. Task Definition : Integrating Traffic Sign Regulations into HD Maps" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 313, + 406, + 556, + 634 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 406, + 556, + 634 + ], + "spans": [ + { + "bbox": [ + 313, + 406, + 556, + 634 + ], + "type": "text", + "content": "The ability to discern rules from traffic signs and to associate them with specific lanes is pivotal for autonomous navigation. Our proposed task focuses on extracting lane-level driving rules denoted as " + }, + { + "bbox": [ + 313, + 406, + 556, + 634 + ], + "type": "inline_equation", + "content": "R = \\{r_i\\}_{i=1}^m" + }, + { + "bbox": [ + 313, + 406, + 556, + 634 + ], + "type": "text", + "content": " and reasoning the correspondence between these rules and the centerlines " + }, + { + "bbox": [ + 313, + 406, + 556, + 634 + ], + "type": "inline_equation", + "content": "L = \\{l_i\\}_{i=1}^k" + }, + { + "bbox": [ + 313, + 406, + 556, + 634 + ], + "type": "text", + "content": " in the local vectorized HD map, where " + }, + { + "bbox": [ + 313, + 406, + 556, + 634 + ], + "type": "inline_equation", + "content": "m" + }, + { + "bbox": [ + 313, + 406, + 556, + 634 + ], + "type": "text", + "content": " is the number of rules and " + }, + { + "bbox": [ + 313, + 406, + 556, + 634 + ], + "type": "inline_equation", + "content": "k" + }, + { + "bbox": [ + 313, + 406, + 556, + 634 + ], + "type": "text", + "content": " is the number of centerlines. Each rule " + }, + { + "bbox": [ + 313, + 406, + 556, + 634 + ], + "type": "inline_equation", + "content": "r_i" + }, + { + "bbox": [ + 313, + 406, + 556, + 634 + ], + "type": "text", + "content": " is a set of pre-defined properties in " + }, + { + "bbox": [ + 313, + 406, + 556, + 634 + ], + "type": "inline_equation", + "content": "\\{key : value\\}" + }, + { + "bbox": [ + 313, + 406, + 556, + 634 + ], + "type": "text", + "content": " pairs, as shown in Fig. 3. The final correspondence forms a bipartite graph " + }, + { + "bbox": [ + 313, + 406, + 556, + 634 + ], + "type": "inline_equation", + "content": "G = (R \\cup L, E)" + }, + { + "bbox": [ + 313, + 406, + 556, + 634 + ], + "type": "text", + "content": ", where " + }, + { + "bbox": [ + 313, + 406, + 556, + 634 + ], + "type": "inline_equation", + "content": "E \\subseteq \\{0,1\\}^{m \\times k}" + }, + { + "bbox": [ + 313, + 406, + 556, + 634 + ], + "type": "text", + "content": " and the element " + }, + { + "bbox": [ + 313, + 406, + 556, + 634 + ], + "type": "inline_equation", + "content": "E_{ij}" + }, + { + "bbox": [ + 313, + 406, + 556, + 634 + ], + "type": "text", + "content": " in the " + }, + { + "bbox": [ + 313, + 406, + 556, + 634 + ], + "type": "inline_equation", + "content": "i" + }, + { + "bbox": [ + 313, + 406, + 556, + 634 + ], + "type": "text", + "content": "-th row and " + }, + { + "bbox": [ + 313, + 406, + 556, + 634 + ], + "type": "inline_equation", + "content": "j" + }, + { + "bbox": [ + 313, + 406, + 556, + 634 + ], + "type": "text", + "content": "-th column of matrix " + }, + { + "bbox": [ + 313, + 406, + 556, + 634 + ], + "type": "inline_equation", + "content": "E" + }, + { + "bbox": [ + 313, + 406, + 556, + 634 + ], + "type": "text", + "content": " represents the corresponding status between " + }, + { + "bbox": [ + 313, + 406, + 556, + 634 + ], + "type": "inline_equation", + "content": "r_i" + }, + { + "bbox": [ + 313, + 406, + 556, + 634 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 313, + 406, + 556, + 634 + ], + "type": "inline_equation", + "content": "l_j" + }, + { + "bbox": [ + 313, + 406, + 556, + 634 + ], + "type": "text", + "content": ". In this case, we assume the entire approach as " + }, + { + "bbox": [ + 313, + 406, + 556, + 634 + ], + "type": "inline_equation", + "content": "\\mathcal{F}" + }, + { + "bbox": [ + 313, + 406, + 556, + 634 + ], + "type": "text", + "content": " and the overall process can be formulated as " + }, + { + "bbox": [ + 313, + 406, + 556, + 634 + ], + "type": "inline_equation", + "content": "G = \\mathcal{F}(X, L)" + }, + { + "bbox": [ + 313, + 406, + 556, + 634 + ], + "type": "text", + "content": ", where " + }, + { + "bbox": [ + 313, + 406, + 556, + 634 + ], + "type": "inline_equation", + "content": "X = \\{x_i\\}_{i=1}^n" + }, + { + "bbox": [ + 313, + 406, + 556, + 634 + ], + "type": "text", + "content": " represents a series of image sequences and " + }, + { + "bbox": [ + 313, + 406, + 556, + 634 + ], + "type": "inline_equation", + "content": "n" + }, + { + "bbox": [ + 313, + 406, + 556, + 634 + ], + "type": "text", + "content": " is the number of frames. To independently investigate the capabilities of rule extraction and correspondence reasoning within the overall task, we further divided the overall task into two sub-tasks, defined as follows." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 313, + 646, + 495, + 659 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 646, + 495, + 659 + ], + "spans": [ + { + "bbox": [ + 313, + 646, + 495, + 659 + ], + "type": "text", + "content": "3.1. Rule Extraction from Traffic Sign" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 313, + 665, + 556, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 665, + 556, + 715 + ], + "spans": [ + { + "bbox": [ + 313, + 665, + 556, + 715 + ], + "type": "text", + "content": "As shown in Step 2 of Fig. 2, this sub-task involves extracting multiple rules " + }, + { + "bbox": [ + 313, + 665, + 556, + 715 + ], + "type": "inline_equation", + "content": "R = \\{r_i\\}_{i=1}^m" + }, + { + "bbox": [ + 313, + 665, + 556, + 715 + ], + "type": "text", + "content": " from a series of image sequences " + }, + { + "bbox": [ + 313, + 665, + 556, + 715 + ], + "type": "inline_equation", + "content": "X = \\{x_i\\}_{i=1}^n" + }, + { + "bbox": [ + 313, + 665, + 556, + 715 + ], + "type": "text", + "content": ". The rule extraction model, denoted as " + }, + { + "bbox": [ + 313, + 665, + 556, + 715 + ], + "type": "inline_equation", + "content": "\\mathcal{M}" + }, + { + "bbox": [ + 313, + 665, + 556, + 715 + ], + "type": "text", + "content": ", can be expressed as " + }, + { + "bbox": [ + 313, + 665, + 556, + 715 + ], + "type": "inline_equation", + "content": "R = \\mathcal{M}(X)" + }, + { + "bbox": [ + 313, + 665, + 556, + 715 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 11 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "text", + "content": "6825" + } + ] + } + ], + "index": 12 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 2 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 69, + 70, + 281, + 155 + ], + "blocks": [ + { + "bbox": [ + 69, + 70, + 281, + 155 + ], + "lines": [ + { + "bbox": [ + 69, + 70, + 281, + 155 + ], + "spans": [ + { + "bbox": [ + 69, + 70, + 281, + 155 + ], + "type": "image", + "image_path": "732294b9cf93258cb44928eb52493a5934660a5921e35324248e2eb038033e09.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 55, + 164, + 297, + 209 + ], + "lines": [ + { + "bbox": [ + 55, + 164, + 297, + 209 + ], + "spans": [ + { + "bbox": [ + 55, + 164, + 297, + 209 + ], + "type": "text", + "content": "Figure 3. Visualization of dataset demo. Multiple lane-level rules of a single traffic sign are annotated in " + }, + { + "bbox": [ + 55, + 164, + 297, + 209 + ], + "type": "inline_equation", + "content": "\\{key: value\\}" + }, + { + "bbox": [ + 55, + 164, + 297, + 209 + ], + "type": "text", + "content": " format. Directed lines indicate the correspondence between rules and particular centerlines." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "bbox": [ + 55, + 230, + 258, + 243 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 230, + 258, + 243 + ], + "spans": [ + { + "bbox": [ + 55, + 230, + 258, + 243 + ], + "type": "text", + "content": "3.2. Rule-Lane Correspondence Reasoning" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 55, + 248, + 296, + 332 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 248, + 296, + 332 + ], + "spans": [ + { + "bbox": [ + 55, + 248, + 296, + 332 + ], + "type": "text", + "content": "As shown in Step 3 of Fig. 2, the reasoning process establishes the correspondence between centerlines " + }, + { + "bbox": [ + 55, + 248, + 296, + 332 + ], + "type": "inline_equation", + "content": "L = \\{l_i\\}_{i=1}^k" + }, + { + "bbox": [ + 55, + 248, + 296, + 332 + ], + "type": "text", + "content": " and all rules " + }, + { + "bbox": [ + 55, + 248, + 296, + 332 + ], + "type": "inline_equation", + "content": "R" + }, + { + "bbox": [ + 55, + 248, + 296, + 332 + ], + "type": "text", + "content": ". We denote the correspondence reasoning model as " + }, + { + "bbox": [ + 55, + 248, + 296, + 332 + ], + "type": "inline_equation", + "content": "\\mathcal{T}" + }, + { + "bbox": [ + 55, + 248, + 296, + 332 + ], + "type": "text", + "content": ", and this process can be described as " + }, + { + "bbox": [ + 55, + 248, + 296, + 332 + ], + "type": "inline_equation", + "content": "E = \\mathcal{T}(R, L)" + }, + { + "bbox": [ + 55, + 248, + 296, + 332 + ], + "type": "text", + "content": ". The final reasoning result forms a bipartite graph " + }, + { + "bbox": [ + 55, + 248, + 296, + 332 + ], + "type": "inline_equation", + "content": "G = (R \\cup L, E)" + }, + { + "bbox": [ + 55, + 248, + 296, + 332 + ], + "type": "text", + "content": ", which means corresponding relationships only exist between rules and centerlines." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 55, + 343, + 253, + 357 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 343, + 253, + 357 + ], + "spans": [ + { + "bbox": [ + 55, + 343, + 253, + 357 + ], + "type": "text", + "content": "4. The MapDR Dataset & Benchmark" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 55, + 364, + 296, + 459 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 364, + 296, + 459 + ], + "spans": [ + { + "bbox": [ + 55, + 364, + 296, + 459 + ], + "type": "text", + "content": "We introduce the MapDR dataset, meticulously annotated with traffic sign regulations and their correspondences to lanes, as shown in Fig. 3. The dataset encompasses a diverse range of scenarios, weather conditions, and traffic situations, with over 10,000 traffic scene segments, 18,000 driving rules, and 400,000 images. Traffic signs typically have varying textual descriptions, text layouts, and positions on the road, which add complexity to the task." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 55, + 460, + 296, + 604 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 460, + 296, + 604 + ], + "spans": [ + { + "bbox": [ + 55, + 460, + 296, + 604 + ], + "type": "text", + "content": "The majority of the data originates from Beijing and Shanghai, with additional scenes from Guangzhou. Fig. 4 illustrates the geographic spread and variety of traffic signs. The dataset reflects a natural long-tail distribution, with a prevalence of bus and direction lanes and a scarcity of tidalflow lanes. We primarily focus on traffic signs that indicate lane-level rules, collected from cities with the most complex and diverse traffic scenarios in China, ensuring realistic and representative data. All images have undergone privacy and safety processing to obscure license plates and faces. More comprehensive statistic of dataset and case demonstrations can be found in the supplement." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 55, + 612, + 192, + 624 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 612, + 192, + 624 + ], + "spans": [ + { + "bbox": [ + 55, + 612, + 192, + 624 + ], + "type": "text", + "content": "4.1. Raw Data & Annotation" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 55, + 629, + 296, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 629, + 296, + 713 + ], + "spans": [ + { + "bbox": [ + 55, + 629, + 296, + 713 + ], + "type": "text", + "content": "Raw Data. MapDR is collected from real-world traffic scenes, each scene segment (video clip) captures front-view images within a " + }, + { + "bbox": [ + 55, + 629, + 296, + 713 + ], + "type": "inline_equation", + "content": "100m \\times 100m" + }, + { + "bbox": [ + 55, + 629, + 296, + 713 + ], + "type": "text", + "content": " area centered on the traffic sign, with a consistent resolution of " + }, + { + "bbox": [ + 55, + 629, + 296, + 713 + ], + "type": "inline_equation", + "content": "1920 \\times 1240" + }, + { + "bbox": [ + 55, + 629, + 296, + 713 + ], + "type": "text", + "content": ". Each clip contains 30 to 60 frames, captured at 1 frame every 2 meters, ensuring consistent spatial intervals. Each video clip focuses on a single traffic sign and provides its position in 3D space." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 313, + 72, + 553, + 107 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 72, + 553, + 107 + ], + "spans": [ + { + "bbox": [ + 313, + 72, + 553, + 107 + ], + "type": "text", + "content": "Camera intrinsics and poses are provided for each frame, and coordinates for each clip are transformed to distinct ENU systems." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 313, + 109, + 555, + 229 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 109, + 555, + 229 + ], + "spans": [ + { + "bbox": [ + 313, + 109, + 555, + 229 + ], + "type": "text", + "content": "All vectors of local map in the target area are provided as 3D point lists, generated using our algorithm similar to MapTRv2 [31]. Each lane vector has a type, such as divider, centerline, crosswalk, or boundary. For example, the centerline is defined as " + }, + { + "bbox": [ + 313, + 109, + 555, + 229 + ], + "type": "inline_equation", + "content": "L = \\{l_i\\}_{i=1}^k" + }, + { + "bbox": [ + 313, + 109, + 555, + 229 + ], + "type": "text", + "content": ", where each vector " + }, + { + "bbox": [ + 313, + 109, + 555, + 229 + ], + "type": "inline_equation", + "content": "l_i" + }, + { + "bbox": [ + 313, + 109, + 555, + 229 + ], + "type": "text", + "content": " is composed of multiple 3D points " + }, + { + "bbox": [ + 313, + 109, + 555, + 229 + ], + "type": "inline_equation", + "content": "l_i = [p_1, \\ldots, p_n]" + }, + { + "bbox": [ + 313, + 109, + 555, + 229 + ], + "type": "text", + "content": ", and " + }, + { + "bbox": [ + 313, + 109, + 555, + 229 + ], + "type": "inline_equation", + "content": "p_j = (x_j, y_j, z_j)" + }, + { + "bbox": [ + 313, + 109, + 555, + 229 + ], + "type": "text", + "content": " represents the coordinates of the current point. The pipeline of dataset production is illustrated in Fig. 5, and detailed data acquisition and annotation procedures can be found in the supplement material." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 313, + 248, + 556, + 403 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 248, + 556, + 403 + ], + "spans": [ + { + "bbox": [ + 313, + 248, + 556, + 403 + ], + "type": "text", + "content": "**Formatted Rules.** Each video clip may contain multiple lane-level rules, denoted as " + }, + { + "bbox": [ + 313, + 248, + 556, + 403 + ], + "type": "inline_equation", + "content": "R" + }, + { + "bbox": [ + 313, + 248, + 556, + 403 + ], + "type": "text", + "content": ". Each rule is expressed by symbols and text on the sign, requiring interpretation. We reference existing HD map design and data specifications [1, 16, 35] to define lane-level driving rules in the JSON format. As shown in Fig. 3, each rule " + }, + { + "bbox": [ + 313, + 248, + 556, + 403 + ], + "type": "inline_equation", + "content": "r_i" + }, + { + "bbox": [ + 313, + 248, + 556, + 403 + ], + "type": "text", + "content": " comprises 8 predefined properties in the form of {key : value} pairs. This definition schema encompasses most scenarios and supports integration into existing autonomous driving systems. We enclose the symbols and texts denoting each distinct rule on traffic signs with polygons and project them into 3D space as " + }, + { + "bbox": [ + 313, + 248, + 556, + 403 + ], + "type": "inline_equation", + "content": "P_i = [p_1, \\ldots, p_n]" + }, + { + "bbox": [ + 313, + 248, + 556, + 403 + ], + "type": "text", + "content": ", where " + }, + { + "bbox": [ + 313, + 248, + 556, + 403 + ], + "type": "inline_equation", + "content": "n" + }, + { + "bbox": [ + 313, + 248, + 556, + 403 + ], + "type": "text", + "content": " varies. Researchers can optionally use this information to facilitate rule extraction." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 313, + 422, + 556, + 509 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 422, + 556, + 509 + ], + "spans": [ + { + "bbox": [ + 313, + 422, + 556, + 509 + ], + "type": "text", + "content": "Correspondence between Rules & Lanes. Based on formatted rules " + }, + { + "bbox": [ + 313, + 422, + 556, + 509 + ], + "type": "inline_equation", + "content": "R" + }, + { + "bbox": [ + 313, + 422, + 556, + 509 + ], + "type": "text", + "content": " and centerlines " + }, + { + "bbox": [ + 313, + 422, + 556, + 509 + ], + "type": "inline_equation", + "content": "L" + }, + { + "bbox": [ + 313, + 422, + 556, + 509 + ], + "type": "text", + "content": ", corresponding centerlines of each rule are annotated as shown in Fig. 3. Therefore correspondence between rules and centerlines can be formed as a bipartite graph " + }, + { + "bbox": [ + 313, + 422, + 556, + 509 + ], + "type": "inline_equation", + "content": "G = (R \\cup L, E)" + }, + { + "bbox": [ + 313, + 422, + 556, + 509 + ], + "type": "text", + "content": ", where " + }, + { + "bbox": [ + 313, + 422, + 556, + 509 + ], + "type": "inline_equation", + "content": "E \\subseteq \\{0,1\\}^{|R| \\times |L|}" + }, + { + "bbox": [ + 313, + 422, + 556, + 509 + ], + "type": "text", + "content": " and the positive edges only exist between " + }, + { + "bbox": [ + 313, + 422, + 556, + 509 + ], + "type": "inline_equation", + "content": "R" + }, + { + "bbox": [ + 313, + 422, + 556, + 509 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 313, + 422, + 556, + 509 + ], + "type": "inline_equation", + "content": "L" + }, + { + "bbox": [ + 313, + 422, + 556, + 509 + ], + "type": "text", + "content": " as demonstrated in Sec. 3.2. Specifically, " + }, + { + "bbox": [ + 313, + 422, + 556, + 509 + ], + "type": "inline_equation", + "content": "E_{ij} = 1" + }, + { + "bbox": [ + 313, + 422, + 556, + 509 + ], + "type": "text", + "content": " represents that" + } + ] + } + ], + "index": 12 + }, + { + "type": "image", + "bbox": [ + 328, + 528, + 542, + 658 + ], + "blocks": [ + { + "bbox": [ + 328, + 528, + 542, + 658 + ], + "lines": [ + { + "bbox": [ + 328, + 528, + 542, + 658 + ], + "spans": [ + { + "bbox": [ + 328, + 528, + 542, + 658 + ], + "type": "image", + "image_path": "8d41f316abb264102e743f75839b8ebe308aa55d8e30be049b22a61d8fd1b753.jpg" + } + ] + } + ], + "index": 13, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 313, + 666, + 556, + 711 + ], + "lines": [ + { + "bbox": [ + 313, + 666, + 556, + 711 + ], + "spans": [ + { + "bbox": [ + 313, + 666, + 556, + 711 + ], + "type": "text", + "content": "Figure 4. Geographic location distribution of the collected traffic signs and proportions of various lane types represented in all signs. The geographic distribution is visualized based on OpenStreetMap [2]." + } + ] + } + ], + "index": 14, + "angle": 0, + "type": "image_caption" + } + ], + "index": 13 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "text", + "content": "6826" + } + ] + } + ], + "index": 15 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 3 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 67, + 69, + 284, + 122 + ], + "blocks": [ + { + "bbox": [ + 67, + 69, + 284, + 122 + ], + "lines": [ + { + "bbox": [ + 67, + 69, + 284, + 122 + ], + "spans": [ + { + "bbox": [ + 67, + 69, + 284, + 122 + ], + "type": "image", + "image_path": "6ac67f847f7504ee637375f74cc16c051b9fc9e4b23c72a868597759f49e1979.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 55, + 131, + 295, + 187 + ], + "lines": [ + { + "bbox": [ + 55, + 131, + 295, + 187 + ], + "spans": [ + { + "bbox": [ + 55, + 131, + 295, + 187 + ], + "type": "text", + "content": "Figure 5. Pipeline of dataset production. The location of traffic signs are sampled from existing database then front-view images of each sign are newly collected. Vectorized map is processed in cloud server. Finally formatted rules and correspondence between rules and centerlines are annotated and organized as MapDR." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "bbox": [ + 55, + 208, + 296, + 232 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 208, + 296, + 232 + ], + "spans": [ + { + "bbox": [ + 55, + 208, + 296, + 232 + ], + "type": "text", + "content": "vehicle driving on the lane with centerline " + }, + { + "bbox": [ + 55, + 208, + 296, + 232 + ], + "type": "inline_equation", + "content": "l_{j}" + }, + { + "bbox": [ + 55, + 208, + 296, + 232 + ], + "type": "text", + "content": " should follow the driving rule " + }, + { + "bbox": [ + 55, + 208, + 296, + 232 + ], + "type": "inline_equation", + "content": "r_{i}" + }, + { + "bbox": [ + 55, + 208, + 296, + 232 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 55, + 240, + 168, + 251 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 240, + 168, + 251 + ], + "spans": [ + { + "bbox": [ + 55, + 240, + 168, + 251 + ], + "type": "text", + "content": "4.2. Evaluation Metrics" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 55, + 258, + 297, + 294 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 258, + 297, + 294 + ], + "spans": [ + { + "bbox": [ + 55, + 258, + 297, + 294 + ], + "type": "text", + "content": "Based on the task defined in Sec. 3, we propose evaluation metrics for entire task as well as for two sub-tasks separately. The " + }, + { + "bbox": [ + 55, + 258, + 297, + 294 + ], + "type": "inline_equation", + "content": "F_{1}" + }, + { + "bbox": [ + 55, + 258, + 297, + 294 + ], + "type": "text", + "content": " score is adopted as the benchmark for methods rank." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 55, + 309, + 296, + 370 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 309, + 296, + 370 + ], + "spans": [ + { + "bbox": [ + 55, + 309, + 296, + 370 + ], + "type": "text", + "content": "Rule Extraction (R.E.). Given the ground truth " + }, + { + "bbox": [ + 55, + 309, + 296, + 370 + ], + "type": "inline_equation", + "content": "R" + }, + { + "bbox": [ + 55, + 309, + 296, + 370 + ], + "type": "text", + "content": " and predicted rules " + }, + { + "bbox": [ + 55, + 309, + 296, + 370 + ], + "type": "inline_equation", + "content": "\\hat{R}" + }, + { + "bbox": [ + 55, + 309, + 296, + 370 + ], + "type": "text", + "content": ", we propose to calculate the Precision " + }, + { + "bbox": [ + 55, + 309, + 296, + 370 + ], + "type": "inline_equation", + "content": "(P_{R,E.})" + }, + { + "bbox": [ + 55, + 309, + 296, + 370 + ], + "type": "text", + "content": " and Recall " + }, + { + "bbox": [ + 55, + 309, + 296, + 370 + ], + "type": "inline_equation", + "content": "(R_{R,E.})" + }, + { + "bbox": [ + 55, + 309, + 296, + 370 + ], + "type": "text", + "content": " to evaluate the capability of rules extraction as defined in Eq. (1), where " + }, + { + "bbox": [ + 55, + 309, + 296, + 370 + ], + "type": "inline_equation", + "content": "\\hat{r_i} = r_i" + }, + { + "bbox": [ + 55, + 309, + 296, + 370 + ], + "type": "text", + "content": " represents all the properties are predicted correctly." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 91, + 379, + 296, + 407 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 91, + 379, + 296, + 407 + ], + "spans": [ + { + "bbox": [ + 91, + 379, + 296, + 407 + ], + "type": "interline_equation", + "content": "P _ {R. E.} = \\frac {\\left| \\hat {R} \\cap R \\right|}{\\left| \\hat {R} \\right|} \\quad R _ {R. E.} = \\frac {\\left| \\hat {R} \\cap R \\right|}{\\left| R \\right|} \\tag {1}", + "image_path": "60a454c8b6648e7bead52cd102b0d63a19398d7f78b2b1451e40275e6d46ba59.jpg" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 55, + 422, + 296, + 495 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 422, + 296, + 495 + ], + "spans": [ + { + "bbox": [ + 55, + 422, + 296, + 495 + ], + "type": "text", + "content": "Correspondence Reasoning (C.R.). Given the ground truth of correspondence bipartite graph " + }, + { + "bbox": [ + 55, + 422, + 296, + 495 + ], + "type": "inline_equation", + "content": "G = (R \\cup L, E)" + }, + { + "bbox": [ + 55, + 422, + 296, + 495 + ], + "type": "text", + "content": " and predicted graph " + }, + { + "bbox": [ + 55, + 422, + 296, + 495 + ], + "type": "inline_equation", + "content": "\\hat{G} = (R \\cup L, \\hat{E})" + }, + { + "bbox": [ + 55, + 422, + 296, + 495 + ], + "type": "text", + "content": ", we propose to calculate Precision (" + }, + { + "bbox": [ + 55, + 422, + 296, + 495 + ], + "type": "inline_equation", + "content": "P_{C.R.}" + }, + { + "bbox": [ + 55, + 422, + 296, + 495 + ], + "type": "text", + "content": ") and Recall (" + }, + { + "bbox": [ + 55, + 422, + 296, + 495 + ], + "type": "inline_equation", + "content": "R_{C.R.}" + }, + { + "bbox": [ + 55, + 422, + 296, + 495 + ], + "type": "text", + "content": ") of edge set " + }, + { + "bbox": [ + 55, + 422, + 296, + 495 + ], + "type": "inline_equation", + "content": "E" + }, + { + "bbox": [ + 55, + 422, + 296, + 495 + ], + "type": "text", + "content": " to evaluate the capability of correspondence reasoning individually. Metrics are defined as Eq. (2)." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 91, + 502, + 295, + 532 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 91, + 502, + 295, + 532 + ], + "spans": [ + { + "bbox": [ + 91, + 502, + 295, + 532 + ], + "type": "interline_equation", + "content": "P _ {C. R.} = \\frac {\\left| \\hat {E} \\cap E \\right|}{\\left| \\hat {E} \\right|} \\quad R _ {C. R.} = \\frac {\\left| \\hat {E} \\cap E \\right|}{\\left| E \\right|} \\tag {2}", + "image_path": "5e95810d568876b0892c2296e49f4cde3e0f26c7b6e19ecc0c8353582ecd37d0.jpg" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 55, + 545, + 297, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 545, + 297, + 713 + ], + "spans": [ + { + "bbox": [ + 55, + 545, + 297, + 713 + ], + "type": "text", + "content": "Overall. As defined in Sec. 3, the final predicted results can be obtained through an end-to-end prediction or a combination of two sequential sub-tasks. Given the predicted rules, correspondence should be reasoned between " + }, + { + "bbox": [ + 55, + 545, + 297, + 713 + ], + "type": "inline_equation", + "content": "\\hat{R}" + }, + { + "bbox": [ + 55, + 545, + 297, + 713 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 55, + 545, + 297, + 713 + ], + "type": "inline_equation", + "content": "L" + }, + { + "bbox": [ + 55, + 545, + 297, + 713 + ], + "type": "text", + "content": " which means the prediction of entire task is " + }, + { + "bbox": [ + 55, + 545, + 297, + 713 + ], + "type": "inline_equation", + "content": "\\hat{G} = (\\hat{R} \\cup L, \\hat{E})" + }, + { + "bbox": [ + 55, + 545, + 297, + 713 + ], + "type": "text", + "content": " and the ground truth is consistent " + }, + { + "bbox": [ + 55, + 545, + 297, + 713 + ], + "type": "inline_equation", + "content": "G = (R \\cup L, E)" + }, + { + "bbox": [ + 55, + 545, + 297, + 713 + ], + "type": "text", + "content": ". We evaluate Precision " + }, + { + "bbox": [ + 55, + 545, + 297, + 713 + ], + "type": "inline_equation", + "content": "(P_{all})" + }, + { + "bbox": [ + 55, + 545, + 297, + 713 + ], + "type": "text", + "content": " and Recall " + }, + { + "bbox": [ + 55, + 545, + 297, + 713 + ], + "type": "inline_equation", + "content": "(R_{all})" + }, + { + "bbox": [ + 55, + 545, + 297, + 713 + ], + "type": "text", + "content": " using the subgraph " + }, + { + "bbox": [ + 55, + 545, + 297, + 713 + ], + "type": "inline_equation", + "content": "G^{s}" + }, + { + "bbox": [ + 55, + 545, + 297, + 713 + ], + "type": "text", + "content": ", where " + }, + { + "bbox": [ + 55, + 545, + 297, + 713 + ], + "type": "inline_equation", + "content": "G^{s} = \\{g_{ij}^{s}\\}_{i=1,j=1}^{m,k}" + }, + { + "bbox": [ + 55, + 545, + 297, + 713 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 55, + 545, + 297, + 713 + ], + "type": "inline_equation", + "content": "g_{ij}^{s} = (\\{r_{i}, l_{j}\\}, e_{ij})" + }, + { + "bbox": [ + 55, + 545, + 297, + 713 + ], + "type": "text", + "content": ". In set of sub-graph " + }, + { + "bbox": [ + 55, + 545, + 297, + 713 + ], + "type": "inline_equation", + "content": "G^{s}" + }, + { + "bbox": [ + 55, + 545, + 297, + 713 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 55, + 545, + 297, + 713 + ], + "type": "inline_equation", + "content": "m" + }, + { + "bbox": [ + 55, + 545, + 297, + 713 + ], + "type": "text", + "content": " is the number of rules, and " + }, + { + "bbox": [ + 55, + 545, + 297, + 713 + ], + "type": "inline_equation", + "content": "k" + }, + { + "bbox": [ + 55, + 545, + 297, + 713 + ], + "type": "text", + "content": " is the number of centerlines. Furthermore, we propose the " + }, + { + "bbox": [ + 55, + 545, + 297, + 713 + ], + "type": "inline_equation", + "content": "F_{1}" + }, + { + "bbox": [ + 55, + 545, + 297, + 713 + ], + "type": "text", + "content": " for the final benchmark ranking. Metrics are defined in Eq. (3) and Eq. (4). The " + }, + { + "bbox": [ + 55, + 545, + 297, + 713 + ], + "type": "inline_equation", + "content": "F_{1}" + }, + { + "bbox": [ + 55, + 545, + 297, + 713 + ], + "type": "text", + "content": " is derived from the calculations of " + }, + { + "bbox": [ + 55, + 545, + 297, + 713 + ], + "type": "inline_equation", + "content": "P_{all}" + }, + { + "bbox": [ + 55, + 545, + 297, + 713 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 55, + 545, + 297, + 713 + ], + "type": "inline_equation", + "content": "R_{all}" + }, + { + "bbox": [ + 55, + 545, + 297, + 713 + ], + "type": "text", + "content": ". We provide an example of calculating the Overall metrics in the supplement material." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 339, + 91, + 555, + 121 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 339, + 91, + 555, + 121 + ], + "spans": [ + { + "bbox": [ + 339, + 91, + 555, + 121 + ], + "type": "interline_equation", + "content": "P _ {a l l} = \\frac {\\left| \\hat {G} ^ {s} \\cap G ^ {s} \\right|}{\\left| \\hat {G} ^ {s} \\right|} \\quad R _ {a l l} = \\frac {\\left| \\hat {G} ^ {s} \\cap G ^ {s} \\right|}{\\left| G ^ {s} \\right|} \\tag {3}", + "image_path": "0d718320be024a43a8db5cb629a3e1c115592f256719546393a75e35842b4ce6.jpg" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 388, + 129, + 555, + 156 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 388, + 129, + 555, + 156 + ], + "spans": [ + { + "bbox": [ + 388, + 129, + 555, + 156 + ], + "type": "interline_equation", + "content": "F _ {1} = 2 \\times \\frac {P _ {\\text {a l l}} \\times R _ {\\text {a l l}}}{P _ {\\text {a l l}} + R _ {\\text {a l l}}} \\tag {4}", + "image_path": "eaf71cdf887f47311763c112e83043e5437d3b03474373429ff8aef422903c31.jpg" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 313, + 163, + 430, + 177 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 163, + 430, + 177 + ], + "spans": [ + { + "bbox": [ + 313, + 163, + 430, + 177 + ], + "type": "text", + "content": "5. Modular Approach" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 313, + 182, + 555, + 302 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 182, + 555, + 302 + ], + "spans": [ + { + "bbox": [ + 313, + 182, + 555, + 302 + ], + "type": "text", + "content": "In order to conduct a comprehensive analysis of the entire task, a modular approach is implemented in this section to address the two sub-tasks. It is demonstrated that these interconnected modules can effectively achieve the ultimate objective of integrating driving rules into the HD map. To address the multimodal information interaction encompassing images, texts, and vectors, a Vision-Language Encoder (VLE) and a Map Element Encoder (MEE) are developed. The following sections expound upon their structures, applications, and the experimental results on MapDR." + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 313, + 310, + 397, + 321 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 310, + 397, + 321 + ], + "spans": [ + { + "bbox": [ + 313, + 310, + 397, + 321 + ], + "type": "text", + "content": "5.1. Architecture" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 313, + 327, + 556, + 614 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 327, + 556, + 614 + ], + "spans": [ + { + "bbox": [ + 313, + 327, + 556, + 614 + ], + "type": "text", + "content": "Vision-Language Encoder. Inspired by vision-language frameworks [7, 25-27, 42], we designed a vision-language fusion model named VLE, following [26]. As shown in Fig. 6, VLE uses ViT-b16 [15] as the vision encoder, with the text encoder and multimodal fusion encoder each consisting of " + }, + { + "bbox": [ + 313, + 327, + 556, + 614 + ], + "type": "inline_equation", + "content": "L" + }, + { + "bbox": [ + 313, + 327, + 556, + 614 + ], + "type": "text", + "content": " transformer layers [51]. Each layer of the fusion encoder includes a cross-attention module for fusion [26]. In practice, distinct rules are represented by varying numbers of symbols and texts, as shown in the OCR results in Fig. 6. To address the challenge of representing variable-length input as fixed-length features, we introduce a [CLS] token for an entire rule and several [STC] tokens for sentence-level representation. The specific usage of these tokens is detailed in Sec. 5.2. Furthermore, we incorporate inter-instance and intra-instance attention mechanisms [31] to enhance model performance by capturing interactions and independence between and within sentences. In addition to content, layout captures the relative positions of symbols and texts, offering important semantic meaning. To leverage this, we encode the layout using the method from [48] and the relative positions of characters as position embedding following [14]. As shown in VLE in Fig. 6, text embedding, layout embedding, and position embedding together form the input of the text encoder." + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 313, + 629, + 556, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 629, + 556, + 715 + ], + "spans": [ + { + "bbox": [ + 313, + 629, + 556, + 715 + ], + "type": "text", + "content": "Map Element Encoder. Vectors can be represented as sequences of points, similar to words in sentences. Inspired by this, we designed MEE akin to language models [14]. The MEE employs " + }, + { + "bbox": [ + 313, + 629, + 556, + 715 + ], + "type": "inline_equation", + "content": "M" + }, + { + "bbox": [ + 313, + 629, + 556, + 715 + ], + "type": "text", + "content": " transformer layers for vector encoding and " + }, + { + "bbox": [ + 313, + 629, + 556, + 715 + ], + "type": "inline_equation", + "content": "N" + }, + { + "bbox": [ + 313, + 629, + 556, + 715 + ], + "type": "text", + "content": " cross-attention layers for multimodal fusion. Utilizing the method from [48], points of each vector are embedded as point embedding. To achieve a fixed-length representation," + } + ] + } + ], + "index": 16 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "text", + "content": "6827" + } + ] + } + ], + "index": 17 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 4 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 85, + 73, + 525, + 220 + ], + "blocks": [ + { + "bbox": [ + 85, + 73, + 525, + 220 + ], + "lines": [ + { + "bbox": [ + 85, + 73, + 525, + 220 + ], + "spans": [ + { + "bbox": [ + 85, + 73, + 525, + 220 + ], + "type": "image", + "image_path": "20dd1837f8f401df33c212c7e9b6b9ac2bae69f347a90298fbacad52d2affe46.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + } + ], + "index": 0 + }, + { + "type": "image", + "bbox": [ + 86, + 221, + 525, + 324 + ], + "blocks": [ + { + "bbox": [ + 86, + 221, + 525, + 324 + ], + "lines": [ + { + "bbox": [ + 86, + 221, + 525, + 324 + ], + "spans": [ + { + "bbox": [ + 86, + 221, + 525, + 324 + ], + "type": "image", + "image_path": "7eac7cfdb22f5c436f60b18c8f4eedb176cd67f2f171a678d7f6ba2ad7dbae6e.jpg" + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 55, + 335, + 555, + 369 + ], + "lines": [ + { + "bbox": [ + 55, + 335, + 555, + 369 + ], + "spans": [ + { + "bbox": [ + 55, + 335, + 555, + 369 + ], + "type": "text", + "content": "Figure 6. Overview of the modular approach. Entire approach can be divided into two main parts: Rule Extraction from Traffic Sign (top) and Rule-Lane Correspondence Reasoning (bottom). Rule Extraction model consists of two sequential stages with the same structure VLE but unshared parameters, and the training procedure is independent." + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_caption" + } + ], + "index": 1 + }, + { + "bbox": [ + 55, + 389, + 296, + 544 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 389, + 296, + 544 + ], + "spans": [ + { + "bbox": [ + 55, + 389, + 296, + 544 + ], + "type": "text", + "content": "we add [VEC] tokens as the first token of each vector, similar to [STC] tokens in the VLE. We also introduce learnable type embedding for vector types, learnable instance embedding to distinguish vector instances, and position embedding from [14] to encode the relative positions of multiple points within a vector. These embedding are aggregated as the input of vector encoder, as shown in Fig. 7. In addition, we employ inter-instance and intra-instance attention mechanisms [31] to prioritize interactions within vectors over interactions between vectors, as depicted in the dashed box on the right side of Fig. 7. The [VEC] token in output serves as fused feature of rules and vectors, enabling the final prediction of their relationships through association head." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 55, + 550, + 153, + 562 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 550, + 153, + 562 + ], + "spans": [ + { + "bbox": [ + 55, + 550, + 153, + 562 + ], + "type": "text", + "content": "5.2. Implementation" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 55, + 567, + 295, + 604 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 567, + 295, + 604 + ], + "spans": [ + { + "bbox": [ + 55, + 567, + 295, + 604 + ], + "type": "text", + "content": "We utilize VLE and MEE as backbones to integrate multiple modalities and address these two sub-tasks. The specific procedures are detailed as follows:" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 55, + 617, + 296, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 617, + 296, + 713 + ], + "spans": [ + { + "bbox": [ + 55, + 617, + 296, + 713 + ], + "type": "text", + "content": "Rule Extraction from Traffic sign. To clarify the objectives of model, we first cluster symbols and texts into groups. As shown in the upper part of Fig. 6, the VLE is used to encode OCR results and images. By calculating the cosine similarity between [STC] tokens, different symbols and texts are clustered into groups. This process is supervised by contrastive loss during training. Next, using grouped OCR results as text input and maintaining the VLE structure, we" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 313, + 389, + 555, + 437 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 389, + 555, + 437 + ], + "spans": [ + { + "bbox": [ + 313, + 389, + 555, + 437 + ], + "type": "text", + "content": "extract lane-level rules. We employ multiple linear layers as understanding head for the [CLS] token to predict the corresponding value for each property of the rules. This facilitates to express all rules as " + }, + { + "bbox": [ + 313, + 389, + 555, + 437 + ], + "type": "inline_equation", + "content": "\\{key:value\\}" + }, + { + "bbox": [ + 313, + 389, + 555, + 437 + ], + "type": "text", + "content": " pairs." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 313, + 455, + 556, + 492 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 455, + 556, + 492 + ], + "spans": [ + { + "bbox": [ + 313, + 455, + 556, + 492 + ], + "type": "text", + "content": "Rule-Lane Correspondence Reasoning. MEE is designed for vector encoding and interaction with rules. Each formatted rule is mapped to an embedding through MLP and" + } + ] + } + ], + "index": 8 + }, + { + "type": "image", + "bbox": [ + 334, + 512, + 542, + 647 + ], + "blocks": [ + { + "bbox": [ + 334, + 512, + 542, + 647 + ], + "lines": [ + { + "bbox": [ + 334, + 512, + 542, + 647 + ], + "spans": [ + { + "bbox": [ + 334, + 512, + 542, + 647 + ], + "type": "image", + "image_path": "27b818cb5a8c6c71eff3683acaf306cecfb29692e6fad6b0e40f1e43d71736d1.jpg" + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 313, + 654, + 556, + 711 + ], + "lines": [ + { + "bbox": [ + 313, + 654, + 556, + 711 + ], + "spans": [ + { + "bbox": [ + 313, + 654, + 556, + 711 + ], + "type": "text", + "content": "Figure 7. Structure of MEE. MEE serves as correspondence reasoning model. Learnable embeddings are introduced within input to enhance the representing capacity of vector types. inter & intra-instance attention mechanisms facilitate to capture the relationships and independence of individual vectors." + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "image_caption" + } + ], + "index": 9 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "text", + "content": "6828" + } + ] + } + ], + "index": 11 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 5 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 82, + 68, + 531, + 242 + ], + "blocks": [ + { + "bbox": [ + 82, + 68, + 531, + 242 + ], + "lines": [ + { + "bbox": [ + 82, + 68, + 531, + 242 + ], + "spans": [ + { + "bbox": [ + 82, + 68, + 531, + 242 + ], + "type": "image", + "image_path": "81fe351a1fe39d9a99bba6400a5b27163cf6ccb63532f0941af583291dd262e1.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 55, + 251, + 555, + 276 + ], + "lines": [ + { + "bbox": [ + 55, + 251, + 555, + 276 + ], + "spans": [ + { + "bbox": [ + 55, + 251, + 555, + 276 + ], + "type": "text", + "content": "Figure 8. Overview of end-to-end approaches. Based on different vectorized HD maps encoding methods, approaches can be categorized into three types." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "bbox": [ + 55, + 295, + 296, + 355 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 295, + 296, + 355 + ], + "spans": [ + { + "bbox": [ + 55, + 295, + 296, + 355 + ], + "type": "text", + "content": "fused with vector features in the fusion encoder, as shown in the lower part of Fig. 6. We add a binary classification head after each [VEC] token to determine the relationship between the current centerline and rule. Details can be found in supplement." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 55, + 365, + 134, + 378 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 365, + 134, + 378 + ], + "spans": [ + { + "bbox": [ + 55, + 365, + 134, + 378 + ], + "type": "text", + "content": "5.3. Experiment" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 55, + 383, + 297, + 479 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 383, + 297, + 479 + ], + "spans": [ + { + "bbox": [ + 55, + 383, + 297, + 479 + ], + "type": "text", + "content": "Setup. The dataset is split into train and test sets in the ratio of " + }, + { + "bbox": [ + 55, + 383, + 297, + 479 + ], + "type": "inline_equation", + "content": "9:1" + }, + { + "bbox": [ + 55, + 383, + 297, + 479 + ], + "type": "text", + "content": ". " + }, + { + "bbox": [ + 55, + 383, + 297, + 479 + ], + "type": "inline_equation", + "content": "L = 6" + }, + { + "bbox": [ + 55, + 383, + 297, + 479 + ], + "type": "text", + "content": " in VLE and " + }, + { + "bbox": [ + 55, + 383, + 297, + 479 + ], + "type": "inline_equation", + "content": "M = 2" + }, + { + "bbox": [ + 55, + 383, + 297, + 479 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 55, + 383, + 297, + 479 + ], + "type": "inline_equation", + "content": "N = 2" + }, + { + "bbox": [ + 55, + 383, + 297, + 479 + ], + "type": "text", + "content": " in MEE. Input images are resized to " + }, + { + "bbox": [ + 55, + 383, + 297, + 479 + ], + "type": "inline_equation", + "content": "256 \\times 256" + }, + { + "bbox": [ + 55, + 383, + 297, + 479 + ], + "type": "text", + "content": " and the feature dimension is 768 with consistent 12 attention heads. We initialize VLE with pre-trained weights of DeiT [50] and BERT [14] while MEE is trained from scratch. The training procedure runs 50 and 120 epochs for VLE and MEE, respectively. More details can be found in the supplement." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 55, + 498, + 297, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 498, + 297, + 714 + ], + "spans": [ + { + "bbox": [ + 55, + 498, + 297, + 714 + ], + "type": "text", + "content": "Results. We use a heuristic method based on OCR character matching and nearest lane matching as the modular baseline. Meanwhile we make modifications to ALBEF [26] and BERT [14] to adapt them to our task. As shown in Tab. 2, the heuristic method performs poorly in both R.E. and C.R. tasks, underscoring the challenging nature of the MapDR dataset. ALBEF-BERT method significantly enhances performance in the R.E. sub-task however lacking adequate adaptation to vector information hampers its performance on the C.R.. Tab. 3 indicates the attention mechanisms significantly improve " + }, + { + "bbox": [ + 55, + 498, + 297, + 714 + ], + "type": "inline_equation", + "content": "R_{R,E}" + }, + { + "bbox": [ + 55, + 498, + 297, + 714 + ], + "type": "text", + "content": ", while layout of text brings marginal improvement. For C.R., the attention mechanisms is the critical part that promotes the effectiveness of MEE. Instance embedding slightly improves " + }, + { + "bbox": [ + 55, + 498, + 297, + 714 + ], + "type": "inline_equation", + "content": "P_{C,R}" + }, + { + "bbox": [ + 55, + 498, + 297, + 714 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 55, + 498, + 297, + 714 + ], + "type": "inline_equation", + "content": "R_{C,R}" + }, + { + "bbox": [ + 55, + 498, + 297, + 714 + ], + "type": "text", + "content": ", while type embedding significantly enhances both, indicating that vector types help the model establish rule-lane correspondence. The separate evaluation results of all lane types and the strategies of heuristic method can be found in the supplement." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 313, + 294, + 444, + 308 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 294, + 444, + 308 + ], + "spans": [ + { + "bbox": [ + 313, + 294, + 444, + 308 + ], + "type": "text", + "content": "6. End-to-End Approach" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 313, + 317, + 556, + 497 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 317, + 556, + 497 + ], + "spans": [ + { + "bbox": [ + 313, + 317, + 556, + 497 + ], + "type": "text", + "content": "Recently, there has been notable progress in the development of Multimodal Large Language Models (MLLMs) [3, 5, 39, 49], demonstrating significant capabilities in processing multimodal data and acquiring implicit knowledge of modality alignment. Their large-scale pre-training endows them with exceptional understanding and generative abilities, thereby enabling the comprehensive execution of the task proposed by MapDR in an end-to-end manner. This section aims to address the overall task using Qwen-VL [5] and to investigate three distinct vector encoding methods. Moreover, we introduce an innovative RuleVLM model based on MEE, designed to process inputs from three modalities (images, texts, and vectors) and to produce driving rules, which yield favorable results on MapDR. Further elaboration on these concepts is presented in subsequent sections." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 314, + 514, + 397, + 525 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 314, + 514, + 397, + 525 + ], + "spans": [ + { + "bbox": [ + 314, + 514, + 397, + 525 + ], + "type": "text", + "content": "6.1. Architecture" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 313, + 534, + 556, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 534, + 556, + 713 + ], + "spans": [ + { + "bbox": [ + 313, + 534, + 556, + 713 + ], + "type": "text", + "content": "As shown in Fig. 8, we classify MLLM methods into three types: Qwen-VL(TextPrompt), Qwen-VL(VisualPrompt), and RuleVLM, based on different vector encoding methods. These methods utilize the framework of Qwen-VL [5], with ViT [15] serving as the image encoder. Qwen-VL(TextPrompt) encodes centerline coordinates in the PV image as text, inputs them into QwenLM; Qwen-VL(VisualPrompt) visualizes the centerline and its index in the PV image and inputs it as Visual Prompt into QwenLM, while RuleVLM uses MEE with the cross-attention layer removed to encode vectorized centerline results and aligns it with LLM through an adapter. The goal of all these methods is to output a serialized set of driving rules in text form, decode it through a json decoder to restore formatted data, and then integrate it into HD maps for further applications." + } + ] + } + ], + "index": 9 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 749, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 749, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 749, + 315, + 757 + ], + "type": "text", + "content": "6829" + } + ] + } + ], + "index": 10 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 6 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 56, + 113, + 553, + 218 + ], + "blocks": [ + { + "bbox": [ + 55, + 70, + 555, + 105 + ], + "lines": [ + { + "bbox": [ + 55, + 70, + 555, + 105 + ], + "spans": [ + { + "bbox": [ + 55, + 70, + 555, + 105 + ], + "type": "text", + "content": "Table 2. Evaluation of the overall task. The heuristic method and the Qwen-VL(TextPrompt) method serve as the baselines for the modular and end-to-end approach, respectively. \" - \" denotes end-to-end approach is not suitable to independent evaluations of C.R. because these approaches do not utilize ground truth of rules for correspondence reasoning independently." + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 56, + 113, + 553, + 218 + ], + "lines": [ + { + "bbox": [ + 56, + 113, + 553, + 218 + ], + "spans": [ + { + "bbox": [ + 56, + 113, + 553, + 218 + ], + "type": "table", + "html": "
ModelTypeR.E.C.R.Overall
PR.E.(%)RR.E.(%)PC.R.(%)RC.R.(%)Pall(%)Rall(%)F1 Score
Heuristic18.0111.5133.0517.995.012.730.035
ALBEF-BERTModular75.7857.564.1417.250.240.780.003
VLE-MEE76.6774.5478.0582.1663.3567.370.653
Qwen-VL(TextPrompt)42.2141.09--8.398.170.083
Qwen-VL(VisualPrompt)End-to-End89.2989.50--39.1439.230.392
RuleVLM89.2889.44--64.1664.280.642
", + "image_path": "67b8ace9a40906a3302c9d5f37377a47290759d42fca61d9445c7cacc5497b81.jpg" + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "table_body" + } + ], + "index": 1 + }, + { + "type": "table", + "bbox": [ + 71, + 289, + 272, + 443 + ], + "blocks": [ + { + "bbox": [ + 55, + 236, + 297, + 281 + ], + "lines": [ + { + "bbox": [ + 55, + 236, + 297, + 281 + ], + "spans": [ + { + "bbox": [ + 55, + 236, + 297, + 281 + ], + "type": "text", + "content": "Table 3. Evaluation of sub-tasks. \"Attn.\" indicates intra & inter-instance attention mechanisms. \"Layout\" refers to the text layout applied in VLE. \"In.E.\" and \"Ty.E.\" denotes instance and type embedding in MEE, respectively." + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 71, + 289, + 272, + 443 + ], + "lines": [ + { + "bbox": [ + 71, + 289, + 272, + 443 + ], + "spans": [ + { + "bbox": [ + 71, + 289, + 272, + 443 + ], + "type": "table", + "html": "
VLEPR.E.(%)RR.E.(%)
Attn.Layout
XX75.7857.56
✓X76.8671.75
✓✓76.6774.54
MEEPC.R.(%)RC.R.(%)
Attn.In.E.Ty.E.
XXX4.1417.25
✓XX68.9171.39
✓✓X69.6872.76
✓✓✓78.0582.16
", + "image_path": "86aee286f340f50ca0c7b7b5bc47de37d5f25d5de6d12cb383f029a27437521d.jpg" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "table_body" + } + ], + "index": 3 + }, + { + "bbox": [ + 55, + 464, + 153, + 477 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 464, + 153, + 477 + ], + "spans": [ + { + "bbox": [ + 55, + 464, + 153, + 477 + ], + "type": "text", + "content": "6.2. Implementation" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 54, + 482, + 297, + 662 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 54, + 482, + 297, + 662 + ], + "spans": [ + { + "bbox": [ + 54, + 482, + 297, + 662 + ], + "type": "text", + "content": "The JSON-formatted driving rules in MapDR are serialized into text and used as the answers in the QA format for training and evaluation. For all methods based on QwenVL, the pre-trained weights of Qwen-VL-Chat (9.6B) are used for initialization. For method using MEE, the trained weights of MEE from the modular approach are utilized for its initialization. During the supervised fine-tuning, we perform LoRA [21] fine-tuning of the QwenLM, while keeping the parameters of ViT frozen. For RuleVLM, we employ a linear layer as the adapter between QwenLM and MEE, while the parameters of MEE and the adapter remains trainable. We use special tokens to separate the centerline tokens extracted by MEE from the visual and text tokens. Additionally, to prevent overfitting, we shuffle the order of centerlines in the local vectorized map during training." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 55, + 671, + 136, + 684 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 671, + 136, + 684 + ], + "spans": [ + { + "bbox": [ + 55, + 671, + 136, + 684 + ], + "type": "text", + "content": "6.3. Experiment." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 55, + 689, + 296, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 689, + 296, + 715 + ], + "spans": [ + { + "bbox": [ + 55, + 689, + 296, + 715 + ], + "type": "text", + "content": "Setup. All experiments are conducted on QA formatted MapDR dataset with 20 epochs of training on NVIDIA RTX" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 313, + 238, + 555, + 287 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 238, + 555, + 287 + ], + "spans": [ + { + "bbox": [ + 313, + 238, + 555, + 287 + ], + "type": "text", + "content": "A6000. The dimension of hidden states remains 4096 while " + }, + { + "bbox": [ + 313, + 238, + 555, + 287 + ], + "type": "inline_equation", + "content": "r" + }, + { + "bbox": [ + 313, + 238, + 555, + 287 + ], + "type": "text", + "content": " and alpha of LoRA configuration are set 64 and 16 respectively. Meanwhile training batch size is 6 and the max training token length remains 2048." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 313, + 300, + 556, + 540 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 300, + 556, + 540 + ], + "spans": [ + { + "bbox": [ + 313, + 300, + 556, + 540 + ], + "type": "text", + "content": "Results. We conducted zero-shot evaluations on a subset of MapDR using existing MLLMs [3, 5, 39, 49] as qualitative investigation, but we encounter challenges in achieving stable structured outputs, further details are provided in the supplement. Additionally, as shown in the Tab. 2, Qwen-VL(TextPrompt) method serves as a end-to-end baseline do not perform better on both the R.E. and overall evaluation. Textual representations of centerline coordinates are challenging for MLLMs to perform spatial reasoning and also resulted in excessive length of sequence, which affected " + }, + { + "bbox": [ + 313, + 300, + 556, + 540 + ], + "type": "inline_equation", + "content": "P_{R.E.}" + }, + { + "bbox": [ + 313, + 300, + 556, + 540 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 313, + 300, + 556, + 540 + ], + "type": "inline_equation", + "content": "R_{R.E.}" + }, + { + "bbox": [ + 313, + 300, + 556, + 540 + ], + "type": "text", + "content": ". Qwen-VL(VisualPrompt) visualizes the centerline within the image, resulting in reduced text sequence length, and demonstrates notable enhancements in the R.E. task compared to the baseline. However, the overall metrics are suboptimal, suggesting that image-based spatial reasoning remains an exceedingly challenging endeavor. The RuleVLM method that we propose introduces MEE for the independent encoding of vectorized HD maps information. This facilitates a more direct extraction of centerline features, resulting in significant enhancements in the overall metrics." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 313, + 550, + 388, + 562 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 550, + 388, + 562 + ], + "spans": [ + { + "bbox": [ + 313, + 550, + 388, + 562 + ], + "type": "text", + "content": "7. Conclusion" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 313, + 570, + 556, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 570, + 556, + 713 + ], + "spans": [ + { + "bbox": [ + 313, + 570, + 556, + 713 + ], + "type": "text", + "content": "In this work, we defined the task of integrating traffic regulations into vectorized HD maps to facilitate the construction of the traffic regulation layer within online HD maps. We constructed the MapDR dataset, which includes over 10,000 video clips, more than 400,000 images, and at least 18,000 driving rules, covering a total mileage of approximately " + }, + { + "bbox": [ + 313, + 570, + 556, + 713 + ], + "type": "inline_equation", + "content": "800Km" + }, + { + "bbox": [ + 313, + 570, + 556, + 713 + ], + "type": "text", + "content": ". Additionally, we proposed the modular approach MEE-VLE and the end-to-end approach RuleVLM, establishing effective baselines for further research. We hope that MapDR can serve as a starting point, and that more researchers can contribute to expanding this dataset or proposing even better solutions." + } + ] + } + ], + "index": 11 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "text", + "content": "6830" + } + ] + } + ], + "index": 12 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 7 + }, + { + "para_blocks": [ + { + "bbox": [ + 56, + 72, + 115, + 83 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 72, + 115, + 83 + ], + "spans": [ + { + "bbox": [ + 56, + 72, + 115, + 83 + ], + "type": "text", + "content": "References" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 57, + 91, + 297, + 712 + ], + "type": "list", + "angle": 0, + "index": 16, + "blocks": [ + { + "bbox": [ + 61, + 91, + 297, + 113 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 61, + 91, + 297, + 113 + ], + "spans": [ + { + "bbox": [ + 61, + 91, + 297, + 113 + ], + "type": "text", + "content": "[1] Navigation data standard. https://nds-association.org/.4" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 62, + 115, + 296, + 136 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 115, + 296, + 136 + ], + "spans": [ + { + "bbox": [ + 62, + 115, + 296, + 136 + ], + "type": "text", + "content": "[2] Openstreetmap. https://github.com/openmaptiles/openmaptiles.4" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 62, + 138, + 295, + 159 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 138, + 295, + 159 + ], + "spans": [ + { + "bbox": [ + 62, + 138, + 295, + 159 + ], + "type": "text", + "content": "[3] Anthropic. Claude-3. https://www.anthropic.com/news/claudi-3-family, 2024.7,8" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 62, + 161, + 295, + 194 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 161, + 295, + 194 + ], + "spans": [ + { + "bbox": [ + 62, + 161, + 295, + 194 + ], + "type": "text", + "content": "[4] Stanislaw Antol, Aishwarya Agrawal, Jiasen Lu, Margaret Mitchell, Dhruv Batra, C. Lawrence Zitnick, and Devi Parikh. VQA: visual question answering. In ICCV, 2015. 3" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 62, + 196, + 295, + 249 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 196, + 295, + 249 + ], + "spans": [ + { + "bbox": [ + 62, + 196, + 295, + 249 + ], + "type": "text", + "content": "[5] Jinze Bai, Shuai Bai, Shusheng Yang, Shijie Wang, Sinan Tan, Peng Wang, Junyang Lin, Chang Zhou, and Jingren Zhou. Qwen-vl: A frontier large vision-language model with versatile abilities. arXiv preprint arXiv:2308.12966, 2023.7, 8" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 62, + 251, + 295, + 294 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 251, + 295, + 294 + ], + "spans": [ + { + "bbox": [ + 62, + 251, + 295, + 294 + ], + "type": "text", + "content": "[6] Zechen Bai, Pichao Wang, Tianjun Xiao, Tong He, Zongbo Han, Zheng Zhang, and Mike Zheng Shou. Hallucination of multimodal large language models: A survey. arXiv preprint arXiv:2404.18930, 2024. 3" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 62, + 297, + 295, + 350 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 297, + 295, + 350 + ], + "spans": [ + { + "bbox": [ + 62, + 297, + 295, + 350 + ], + "type": "text", + "content": "[7] Hangbo Bao, Wenhui Wang, Li Dong, Qiang Liu, Owais Khan Mohammed, Kriti Aggarwal, Subhojit Som, Songhao Piao, and Furu Wei. Vlmo: Unified vision-language pre-training with mixture-of-modality-experts. In NeurIPS, 2022. 5" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 62, + 353, + 295, + 385 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 353, + 295, + 385 + ], + "spans": [ + { + "bbox": [ + 62, + 353, + 295, + 385 + ], + "type": "text", + "content": "[8] Karsten Behrendt, Libor Novak, and Rami Botros. A deep learning approach to traffic lights: Detection, tracking, and classification. In ICRA, 2017. 2" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 62, + 388, + 295, + 430 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 388, + 295, + 430 + ], + "spans": [ + { + "bbox": [ + 62, + 388, + 295, + 430 + ], + "type": "text", + "content": "[9] Holger Caesar, Varun Bankiti, Alex H. Lang, Sourabh Vora, Venice Erin Liong, Qiang Xu, Anush Krishnan, Yu Pan, Giancarlo Baldan, and Oscar Beijbom. nuscenes: A multimodal dataset for autonomous driving. In CVPR, 2020. 2" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 57, + 433, + 295, + 487 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 433, + 295, + 487 + ], + "spans": [ + { + "bbox": [ + 57, + 433, + 295, + 487 + ], + "type": "text", + "content": "[10] Xu Cao, Tong Zhou, Yunsheng Ma, Wenqian Ye, Can Cui, Kun Tang, Zhipeng Cao, Kaizhao Liang, Ziran Wang, James Rehg, and Chao Zheng. Maplm: A real-world large-scale vision-language dataset for map and traffic scene understanding. In CVPR, 2024. 3" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 57, + 488, + 295, + 542 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 488, + 295, + 542 + ], + "spans": [ + { + "bbox": [ + 57, + 488, + 295, + 542 + ], + "type": "text", + "content": "[11] Li Chen, Chonghao Sima, Yang Li, Zehan Zheng, Jiajie Xu, Xiangwei Geng, Hongyang Li, Conghui He, Jianping Shi, Yu Qiao, and Junchi Yan. Persformer: 3d lane detection via perspective transformer and the openlane benchmark. In ECCV, 2022. 2" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 57, + 544, + 295, + 597 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 544, + 295, + 597 + ], + "spans": [ + { + "bbox": [ + 57, + 544, + 295, + 597 + ], + "type": "text", + "content": "[12] Long Chen, Oleg Sinavski, Jan Hunermann, Alice Karnsund, Andrew James Willmott, Danny Birch, Daniel Maund, and Jamie Shotton. Driving with llms: Fusing object-level vector modality for explainable autonomous driving. In ICRA, 2024. 3" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 57, + 601, + 295, + 645 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 601, + 295, + 645 + ], + "spans": [ + { + "bbox": [ + 57, + 601, + 295, + 645 + ], + "type": "text", + "content": "[13] Yaodong Cui, Shucheng Huang, Jiaming Zhong, Zhenan Liu, Yutong Wang, Chen Sun, Bai Li, Xiao Wang, and Amir Khajepour. Drivellm: Charting the path toward full autonomous driving with large language models. IEEE TIV, 2024. 3" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 57, + 647, + 295, + 688 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 647, + 295, + 688 + ], + "spans": [ + { + "bbox": [ + 57, + 647, + 295, + 688 + ], + "type": "text", + "content": "[14] Jacob Devlin, Ming-Wei Chang, Kenton Lee, and Kristina Toutanova. BERT: pre-training of deep bidirectional transformers for language understanding. In NAACL, 2019. 5, 6, 7" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 57, + 691, + 295, + 712 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 691, + 295, + 712 + ], + "spans": [ + { + "bbox": [ + 57, + 691, + 295, + 712 + ], + "type": "text", + "content": "[15] Alexey Dosovitskiy, Lucas Beyer, Alexander Kolesnikov, Dirk Weissenborn, Xiaohua Zhai, Thomas Unterthiner," + } + ] + } + ], + "index": 15 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 316, + 73, + 555, + 714 + ], + "type": "list", + "angle": 0, + "index": 33, + "blocks": [ + { + "bbox": [ + 335, + 73, + 555, + 116 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 335, + 73, + 555, + 116 + ], + "spans": [ + { + "bbox": [ + 335, + 73, + 555, + 116 + ], + "type": "text", + "content": "Mostafa Dehghani, Matthias Minderer, Georg Heigold, Sylvain Gelly, Jakob Uszkoreit, and Neil Houlsby. An image is worth 16x16 words: Transformers for image recognition at scale. In ICLR, 2021. 5, 7" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 317, + 118, + 555, + 162 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 118, + 555, + 162 + ], + "spans": [ + { + "bbox": [ + 317, + 118, + 555, + 162 + ], + "type": "text", + "content": "[16] Gamal Elghazaly, Raphaël Frank, Scott Harvey, and Stefan Safko. High-definition maps: Comprehensive survey, challenges, and future perspectives. IEEE Open Journal of Intelligent Transportation Systems, 2023. 2, 4" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 317, + 163, + 554, + 195 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 163, + 554, + 195 + ], + "spans": [ + { + "bbox": [ + 317, + 163, + 554, + 195 + ], + "type": "text", + "content": "[17] Andreas Fregin, Julian Müller, Ulrich Krebel, and Klaus Dietmayer. The driveu traffic light dataset: Introduction and comparison with existing datasets. In ICRA, 2018. 2" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 317, + 197, + 555, + 228 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 197, + 555, + 228 + ], + "spans": [ + { + "bbox": [ + 317, + 197, + 555, + 228 + ], + "type": "text", + "content": "[18] Shuo Gu, Yigong Zhang, Jinhui Tang, Jian Yang, and Hui Kong. Road detection through CRF based lidar-camera fusion. In ICRA, 2019. 2" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 317, + 231, + 554, + 262 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 231, + 554, + 262 + ], + "spans": [ + { + "bbox": [ + 317, + 231, + 554, + 262 + ], + "type": "text", + "content": "[19] Yunfei Guo, Wei Feng, Fei Yin, Tao Xue, Shuqi Mei, and Cheng-Lin Liu. Learning to understand traffic signs. In ACMMM, 2021. 2, 3" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 317, + 264, + 554, + 297 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 264, + 554, + 297 + ], + "spans": [ + { + "bbox": [ + 317, + 264, + 554, + 297 + ], + "type": "text", + "content": "[20] Yunfei Guo, Fei Yin, Xiao-Hui Li, Xudong Yan, Tao Xue, Shuqi Mei, and Cheng-Lin Liu. Visual traffic knowledge graph generation from scene images. In ICCV, 2023. 2, 3" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 317, + 299, + 555, + 353 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 299, + 555, + 353 + ], + "spans": [ + { + "bbox": [ + 317, + 299, + 555, + 353 + ], + "type": "text", + "content": "[21] Edward J. Hu, Yelong Shen, Phillip Wallis, Zeyuan Allen-Zhu, Yuanzhi Li, Shean Wang, Lu Wang, and Weizhu Chen. Lora: Low-rank adaptation of large language models. In The Tenth International Conference on Learning Representations, ICLR 2022, Virtual Event, April 25-29, 2022, 2022. 8" + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 317, + 354, + 554, + 397 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 354, + 554, + 397 + ], + "spans": [ + { + "bbox": [ + 317, + 354, + 554, + 397 + ], + "type": "text", + "content": "[22] Xinyu Huang, Peng Wang, Xinjing Cheng, Dingfu Zhou, Qichuan Geng, and Ruigang Yang. The apolloscope open dataset for autonomous driving and its application. IEEE TPAMI, 2020. 2" + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 317, + 399, + 554, + 432 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 399, + 554, + 432 + ], + "spans": [ + { + "bbox": [ + 317, + 399, + 554, + 432 + ], + "type": "text", + "content": "[23] Yupan Huang, Tengchao Lv, Lei Cui, Yutong Lu, and Furu Wei. Layoutmv3: Pre-training for document AI with unified text and image masking. In ACMMM, 2022. 3" + } + ] + } + ], + "index": 25 + }, + { + "bbox": [ + 317, + 434, + 554, + 477 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 434, + 554, + 477 + ], + "spans": [ + { + "bbox": [ + 317, + 434, + 554, + 477 + ], + "type": "text", + "content": "[24] Chao Jia, Yinfei Yang, Ye Xia, Yi-Ting Chen, Zarana Parekh, Hieu Pham, Quoc V. Le, Yun-Hsuan Sung, Zhen Li, and Tom Duerig. Scaling up visual and vision-language representation learning with noisy text supervision. In ICML, 2021. 3" + } + ] + } + ], + "index": 26 + }, + { + "bbox": [ + 317, + 478, + 555, + 510 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 478, + 555, + 510 + ], + "spans": [ + { + "bbox": [ + 317, + 478, + 555, + 510 + ], + "type": "text", + "content": "[25] Wonjae Kim, Bokyung Son, and Ildoo Kim. Vilt: Vision-and-language transformer without convolution or region supervision. In ICML, 2021. 3, 5" + } + ] + } + ], + "index": 27 + }, + { + "bbox": [ + 317, + 512, + 555, + 564 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 512, + 555, + 564 + ], + "spans": [ + { + "bbox": [ + 317, + 512, + 555, + 564 + ], + "type": "text", + "content": "[26] Junnan Li, Ramprasaath R. Selvaraju, Akhilesh Gotmare, Shafiq R. Joty, Caiming Xiong, and Steven Chu-Hong Hoi. Align before fuse: Vision and language representation learning with momentum distillation. In NeurIPS, 2021. 3, 5, 7" + } + ] + } + ], + "index": 28 + }, + { + "bbox": [ + 317, + 567, + 555, + 611 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 567, + 555, + 611 + ], + "spans": [ + { + "bbox": [ + 317, + 567, + 555, + 611 + ], + "type": "text", + "content": "[27] Junnan Li, Dongxu Li, Caiming Xiong, and Steven C. H. Hoi. BLIP: bootstrapping language-image pre-training for unified vision-language understanding and generation. In ICML, 2022. 5" + } + ] + } + ], + "index": 29 + }, + { + "bbox": [ + 317, + 613, + 554, + 655 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 613, + 554, + 655 + ], + "spans": [ + { + "bbox": [ + 317, + 613, + 554, + 655 + ], + "type": "text", + "content": "[28] Junnan Li, Dongxu Li, Silvio Savarese, and Steven C. H. Hoi. BLIP-2: bootstrapping language-image pre-training with frozen image encoders and large language models. In ICML, 2023. 3" + } + ] + } + ], + "index": 30 + }, + { + "bbox": [ + 317, + 658, + 554, + 689 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 658, + 554, + 689 + ], + "spans": [ + { + "bbox": [ + 317, + 658, + 554, + 689 + ], + "type": "text", + "content": "[29] Qi Li, Yue Wang, Yilun Wang, and Hang Zhao. Hdmapnet: An online HD map construction and evaluation framework. In ICRA, 2022. 2" + } + ] + } + ], + "index": 31 + }, + { + "bbox": [ + 316, + 691, + 554, + 714 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 691, + 554, + 714 + ], + "spans": [ + { + "bbox": [ + 316, + 691, + 554, + 714 + ], + "type": "text", + "content": "[30] Benchcheng Liao, Shaoyu Chen, Xinggang Wang, Tianheng Cheng, Qian Zhang, Wenyu Liu, and Chang Huang. Maptr:" + } + ] + } + ], + "index": 32 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 749, + 314, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 749, + 314, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 749, + 314, + 757 + ], + "type": "text", + "content": "6831" + } + ] + } + ], + "index": 34 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 8 + }, + { + "para_blocks": [ + { + "bbox": [ + 56, + 72, + 297, + 713 + ], + "type": "list", + "angle": 0, + "index": 16, + "blocks": [ + { + "bbox": [ + 77, + 72, + 294, + 95 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 77, + 72, + 294, + 95 + ], + "spans": [ + { + "bbox": [ + 77, + 72, + 294, + 95 + ], + "type": "text", + "content": "Structured modeling and learning for online vectorized HD map construction. In ICLR, 2023. 2" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 56, + 95, + 297, + 148 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 95, + 297, + 148 + ], + "spans": [ + { + "bbox": [ + 56, + 95, + 297, + 148 + ], + "type": "text", + "content": "[31] Benchcheng Liao, Shaoyu Chen, Yunchi Zhang, Bo Jiang, Qian Zhang, Wenyu Liu, Chang Huang, and Xinggang Wang. Maptrv2: An end-to-end framework for online vectorized hd map construction. arXiv preprint arXiv:2308.05736, 2023. 3, 4, 5, 6" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 56, + 150, + 296, + 183 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 150, + 296, + 183 + ], + "spans": [ + { + "bbox": [ + 56, + 150, + 296, + 183 + ], + "type": "text", + "content": "[32] Haotian Liu, Chunyuan Li, Yuheng Li, and Yong Jae Lee. Improved baselines with visual instruction tuning. arXiv preprint arXiv:2310.03744, 2023. 3" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 56, + 184, + 296, + 205 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 184, + 296, + 205 + ], + "spans": [ + { + "bbox": [ + 56, + 184, + 296, + 205 + ], + "type": "text", + "content": "[33] Haotian Liu, Chunyuan Li, Qingyang Wu, and Yong Jae Lee. Visual instruction tuning. In NeurIPS, 2023." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 56, + 206, + 296, + 258 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 206, + 296, + 258 + ], + "spans": [ + { + "bbox": [ + 56, + 206, + 296, + 258 + ], + "type": "text", + "content": "[34] Haotian Liu, Chunyuan Li, Yuheng Li, Bo Li, Yuanhan Zhang, Sheng Shen, and Yong Jae Lee. Llava-last: Improved reasoning,OCR,and world knowledge.https://llava-v1.github.io/blog/2024-01-30-llava-last, 2024.3" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 56, + 261, + 296, + 293 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 261, + 296, + 293 + ], + "spans": [ + { + "bbox": [ + 56, + 261, + 296, + 293 + ], + "type": "text", + "content": "[35] Rong Liu, Jinling Wang, and Bingqi Zhang. High definition map for automated driving: Overview and analysis. Journal of Navigation, 2020. 2, 4" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 56, + 294, + 296, + 326 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 294, + 296, + 326 + ], + "spans": [ + { + "bbox": [ + 56, + 294, + 296, + 326 + ], + "type": "text", + "content": "[36] Yicheng Liu, Tianyuan Yuan, Yue Wang, Yilun Wang, and Hang Zhao. Vectormapnet: End-to-end vectorized HD map learning. In ICML, 2023. 2" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 56, + 327, + 296, + 358 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 327, + 296, + 358 + ], + "spans": [ + { + "bbox": [ + 56, + 327, + 296, + 358 + ], + "type": "text", + "content": "[37] Jiageng Mao, Yuxi Qian, Hang Zhao, and Yue Wang. Gptdriver: Learning to drive with GPT. In NeurIPS Workshop, 2023. 3" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 56, + 361, + 296, + 415 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 361, + 296, + 415 + ], + "spans": [ + { + "bbox": [ + 56, + 361, + 296, + 415 + ], + "type": "text", + "content": "[38] Ana-Maria Marcu, Long Chen, Jan Hunermann, Alice Karnsund, Benoit Hanotte, Prajwal Chidananda, Saurabh Nair, Vijay Badrinarayanan, Alex Kendall, Jamie Shotton, and Oleg Sinavski. Lingoqa: Video question answering for autonomous driving. arXiv preprint arXiv:2312.14115, 2023. 3" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 56, + 416, + 296, + 436 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 416, + 296, + 436 + ], + "spans": [ + { + "bbox": [ + 56, + 416, + 296, + 436 + ], + "type": "text", + "content": "[39] OpenAI. GPT-4 technical report. arXiv preprint arXiv:2303.08774, 2024. 7, 8" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 56, + 437, + 296, + 469 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 437, + 296, + 469 + ], + "spans": [ + { + "bbox": [ + 56, + 437, + 296, + 469 + ], + "type": "text", + "content": "[40] Jonah Philion and Sanja Fidler. Lift, splat, shoot: Encoding images from arbitrary camera rigs by implicitly unprojecting to 3d. In ECCV, 2020. 2" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 56, + 471, + 296, + 514 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 471, + 296, + 514 + ], + "spans": [ + { + "bbox": [ + 56, + 471, + 296, + 514 + ], + "type": "text", + "content": "[41] Tianwen Qian, Jingjing Chen, Linhai Zhuo, Yang Jiao, and Yu-Gang Jiang. Nuscenes-qa: A multi-modal visual question answering benchmark for autonomous driving scenario. In AAAI, 2024. 3" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 56, + 515, + 296, + 579 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 515, + 296, + 579 + ], + "spans": [ + { + "bbox": [ + 56, + 515, + 296, + 579 + ], + "type": "text", + "content": "[42] Alec Radford, Jong Wook Kim, Chris Hallacy, Aditya Ramesh, Gabriel Goh, Sandhini Agarwal, Girish Sastry, Amanda Askell, Pamela Mishkin, Jack Clark, Gretchen Krueger, and Ilya Sutskever. Learning transferable visual models from natural language supervision. In ICML, 2021. 3, 5" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 56, + 582, + 296, + 624 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 582, + 296, + 624 + ], + "spans": [ + { + "bbox": [ + 56, + 582, + 296, + 624 + ], + "type": "text", + "content": "[43] Enna Sachdeva, Nakul Agarwal, Suhas Chundi, Sean Roelofs, Jiachen Li, Mykel J. Kochenderfer, Chiho Choi, and Behzad Dariush. Rank2tell: A multimodal driving dataset for joint importance ranking and reasoning. In WACV, 2024. 3" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 56, + 625, + 296, + 668 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 625, + 296, + 668 + ], + "spans": [ + { + "bbox": [ + 56, + 625, + 296, + 668 + ], + "type": "text", + "content": "[44] Chonghao Sima, Katrin Renz, Kashyap Chitta, Li Chen, Hanxue Zhang, Chengen Xie, Ping Luo, Andreas Geiger, and Hongyang Li. Drivelm: Driving with graph visual question answering. arXiv preprint arXiv:2312.14150, 2023. 3" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 56, + 670, + 296, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 670, + 296, + 713 + ], + "spans": [ + { + "bbox": [ + 56, + 670, + 296, + 713 + ], + "type": "text", + "content": "[45] Chonghao Sima, Katrin Renz, Kashyap Chitta, Li Chen, Hanxue Zhang, Chengen Xie, Jens Beiwenger, Ping Luo, Andreas Geiger, and Hongyang Li. Drivelm: Driving with graph visual question answering. In ECCV, 2024. 3" + } + ] + } + ], + "index": 15 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 316, + 73, + 555, + 712 + ], + "type": "list", + "angle": 0, + "index": 30, + "blocks": [ + { + "bbox": [ + 316, + 73, + 555, + 115 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 73, + 555, + 115 + ], + "spans": [ + { + "bbox": [ + 316, + 73, + 555, + 115 + ], + "type": "text", + "content": "[46] Johannes Stallkamp, Marc Schlipsing, Jan Salmen, and Christian Igel. Man vs. computer: Benchmarking machine learning algorithms for traffic sign recognition. *Neural Networks*, 2012. 2" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 316, + 118, + 555, + 205 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 118, + 555, + 205 + ], + "spans": [ + { + "bbox": [ + 316, + 118, + 555, + 205 + ], + "type": "text", + "content": "[47] Pei Sun, Henrik Kretzschmar, Xerxes Dotiwalla, Aurelien Chouard, Vijaysai Patnaik, Paul Tsui, James Guo, Yin Zhou, Yuning Chai, Benjamin Caine, Vijay Vasudevan, Wei Han, Jiquan Ngiam, Hang Zhao, Aleksei Timofeev, Scott Ettinger, Maxim Krivokon, Amy Gao, Aditya Joshi, Yu Zhang, Jonathon Shlens, Zhifeng Chen, and Dragomir Anguelov. Scalability in perception for autonomous driving: Waymo open dataset. In CVPR, 2020. 2" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 316, + 206, + 555, + 258 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 206, + 555, + 258 + ], + "spans": [ + { + "bbox": [ + 316, + 206, + 555, + 258 + ], + "type": "text", + "content": "[48] Matthew Tancik, Pratul P. Srinivasan, Ben Mildenhall, Sara Fridovich-Keil, Nithin Raghavan, Utkarsh Singhal, Ravi Ramamoorthi, Jonathan T. Barron, and Ren Ng. Fourier features let networks learn high frequency functions in low dimensional domains. In NeurIPS, 2020. 5" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 316, + 261, + 555, + 314 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 261, + 555, + 314 + ], + "spans": [ + { + "bbox": [ + 316, + 261, + 555, + 314 + ], + "type": "text", + "content": "[49] Gemini Team, Rohan Anil, Sebastian Borgeaud, Yonghui Wu, Jean-Baptiste Alayrac, Jiahui Yu, Radu Soricut, Johan Schalkwyk, Andrew M. Dai, Anja Hauth, and et al. Gemini: A family of highly capable multimodal models. arXiv preprint arXiv:2312.11805, 2024. 7, 8" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 316, + 316, + 555, + 358 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 316, + 555, + 358 + ], + "spans": [ + { + "bbox": [ + 316, + 316, + 555, + 358 + ], + "type": "text", + "content": "[50] Hugo Touvron, Matthieu Cord, Matthijs Douze, Francisco Massa, Alexandre Sablayrolles, and Hervé Jégou. Training data-efficient image transformers & distillation through attention. In ICML, 2021. 7" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 316, + 361, + 555, + 392 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 361, + 555, + 392 + ], + "spans": [ + { + "bbox": [ + 316, + 361, + 555, + 392 + ], + "type": "text", + "content": "[51] Ashish Vaswani, Noam Shazeer, Niki Parmar, Jakob Uszkoreit, Llion Jones, Aidan N. Gomez, Lukasz Kaiser, and Illia Polosukhin. Attention is all you need. In NeurIPS, 2017. 3, 5" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 316, + 394, + 555, + 457 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 394, + 555, + 457 + ], + "spans": [ + { + "bbox": [ + 316, + 394, + 555, + 457 + ], + "type": "text", + "content": "[52] Huijie Wang, Tianyu Li, Yang Li, Li Chen, Chonghao Sima, Zhenbo Liu, Bangjun Wang, Peijin Jia, Yuting Wang, Shengyin Jiang, Feng Wen, Hang Xu, Ping Luo, Junchi Yan, Wei Zhang, and Hongyang Li. Openlane-v2: A topology reasoning benchmark for unified 3d HD mapping. In NeurIPS, 2023. 2, 3" + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 316, + 460, + 555, + 524 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 460, + 555, + 524 + ], + "spans": [ + { + "bbox": [ + 316, + 460, + 555, + 524 + ], + "type": "text", + "content": "[53] Benjamin Wilson, William Qi, Tanmay Agarwal, John Lambert, Jagjeet Singh, Siddhesh Khandelwal, Bowen Pan, Ratnesh Kumar, Andrew Hartnett, Jhony Kaesemodel Pontes, Deva Ramanan, Peter Carr, and James Hays. Argoverse 2: Next generation datasets for self-driving perception and forecasting. In NeurIPS, 2021. 2" + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 316, + 526, + 553, + 558 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 526, + 553, + 558 + ], + "spans": [ + { + "bbox": [ + 316, + 526, + 553, + 558 + ], + "type": "text", + "content": "[54] Dongming Wu, Jiahao Chang, Fan Jia, Yingfei Liu, Tiancai Wang, and Jianbing Shen. Topomlp: A simple yet strong pipeline for driving topology reasoning. In ICLR, 2024. 2" + } + ] + } + ], + "index": 25 + }, + { + "bbox": [ + 316, + 559, + 553, + 591 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 559, + 553, + 591 + ], + "spans": [ + { + "bbox": [ + 316, + 559, + 553, + 591 + ], + "type": "text", + "content": "[55] Yiheng Xu, Minghao Li, Lei Cui, Shaohan Huang, Furu Wei, and Ming Zhou. Layoutm: Pre-training of text and layout for document image understanding. In KDD, 2020. 3" + } + ] + } + ], + "index": 26 + }, + { + "bbox": [ + 316, + 593, + 555, + 646 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 593, + 555, + 646 + ], + "spans": [ + { + "bbox": [ + 316, + 593, + 555, + 646 + ], + "type": "text", + "content": "[56] Yang Xu, Yiheng Xu, Tengchao Lv, Lei Cui, Furu Wei, Guoxin Wang, Yijuan Lu, Dinei A. F. Floréncio, Cha Zhang, Wanxiang Che, Min Zhang, and Lidong Zhou. Layoutmv2: Multi-modal pre-training for visually-rich document understanding. In ACL, 2021. 3" + } + ] + } + ], + "index": 27 + }, + { + "bbox": [ + 316, + 647, + 555, + 690 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 647, + 555, + 690 + ], + "spans": [ + { + "bbox": [ + 316, + 647, + 555, + 690 + ], + "type": "text", + "content": "[57] Zhenhua Xu, Yujia Zhang, Enze Xie, Zhen Zhao, Yong Guo, Kwan-Yee K. Wong, Zhenguo Li, and Hengshuang Zhao. Drivegpt4: Interpretable end-to-end autonomous driving via large language model. IEEE Robotics Autom. Lett., 2024. 3" + } + ] + } + ], + "index": 28 + }, + { + "bbox": [ + 316, + 692, + 555, + 712 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 692, + 555, + 712 + ], + "spans": [ + { + "bbox": [ + 316, + 692, + 555, + 712 + ], + "type": "text", + "content": "[58] Fisher Yu, Haofeng Chen, Xin Wang, Wenqi Xian, Yingying Chen, Fangchen Liu, Vashisht Madhavan, and Trevor Dar" + } + ] + } + ], + "index": 29 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "text", + "content": "6832" + } + ] + } + ], + "index": 31 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 9 + }, + { + "para_blocks": [ + { + "bbox": [ + 56, + 72, + 296, + 162 + ], + "type": "list", + "angle": 0, + "index": 3, + "blocks": [ + { + "bbox": [ + 76, + 72, + 294, + 95 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 76, + 72, + 294, + 95 + ], + "spans": [ + { + "bbox": [ + 76, + 72, + 294, + 95 + ], + "type": "text", + "content": "rell. BDD100K: A diverse driving dataset for heterogeneous multitask learning. In CVPR, 2020. 2" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 56, + 96, + 296, + 129 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 96, + 296, + 129 + ], + "spans": [ + { + "bbox": [ + 56, + 96, + 296, + 129 + ], + "type": "text", + "content": "[59] Jiahui Yu, Zirui Wang, Vijay Vasudevan, Legg Yeung, Mojtaba Seyedhosseini, and Yonghui Wu. Coca: Contrastive captioners are image-text foundation models. TMLR, 2022. 3" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 56, + 130, + 294, + 162 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 130, + 294, + 162 + ], + "spans": [ + { + "bbox": [ + 56, + 130, + 294, + 162 + ], + "type": "text", + "content": "[60] Zhe Zhu, Dun Liang, Song-Hai Zhang, Xiaolei Huang, Baoli Li, and Shi-Min Hu. Traffic-sign detection and classification in the wild. In CVPR, 2016. 2" + } + ] + } + ], + "index": 2 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 749, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 749, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 749, + 315, + 757 + ], + "type": "text", + "content": "6833" + } + ] + } + ], + "index": 4 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 10 + } + ], + "_backend": "vlm", + "_version_name": "2.6.4" +} \ No newline at end of file diff --git a/2025/DrivingSphere_ Building a High-fidelity 4D World for Closed-loop Simulation/e8e4b9c2-6cbc-49b7-a622-765862a5895e_content_list.json b/2025/DrivingSphere_ Building a High-fidelity 4D World for Closed-loop Simulation/e8e4b9c2-6cbc-49b7-a622-765862a5895e_content_list.json new file mode 100644 index 0000000000000000000000000000000000000000..df56f049dd98b6262961a817c9eaf8b217c2ff44 --- /dev/null +++ b/2025/DrivingSphere_ Building a High-fidelity 4D World for Closed-loop Simulation/e8e4b9c2-6cbc-49b7-a622-765862a5895e_content_list.json @@ -0,0 +1,1615 @@ +[ + { + "type": "text", + "text": "DrivingSphere: Building a High-fidelity 4D World for Closed-loop Simulation", + "text_level": 1, + "bbox": [ + 94, + 130, + 872, + 152 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Tianyi Yan $^{1}$ , Dongming Wu $^{3}$ , Wencheng Han $^{1}$ , Junpeng Jiang $^{2}$ , Xia Zhou $^{2}$ , Kun Zhan $^{2}$ , Cheng-zhong Xu $^{1}$ , Jianbing Shen $^{1*}$", + "bbox": [ + 232, + 179, + 733, + 215 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "$^{1}$ SKL-IOTSC, Computer and Information Science, University of Macau, $^{2}$ Li Auto Inc., $^{3}$ School of Computer Science, Beijing Institute of Technology", + "bbox": [ + 143, + 215, + 826, + 252 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Abstract", + "text_level": 1, + "bbox": [ + 233, + 286, + 313, + 301 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Autonomous driving evaluation requires simulation environments that closely replicate actual road conditions, including real-world sensory data and responsive feedback loops. However, many existing simulations need to predict waypoints along fixed routes on public datasets or synthetic photorealistic data, i.e., open-loop simulation usually lacks the ability to assess dynamic decision-making. While the recent efforts of closed-loop simulation offer feedback-driven environments, they cannot process visual sensor inputs or produce outputs that differ from real-world data. To address these challenges, we propose DrivingSphere, a realistic and closed-loop simulation framework. Its core idea is to build 4D world representation and generate real-life and controllable driving scenarios. In specific, our framework includes a Dynamic Environment Composition module that constructs a detailed 4D driving world with a format of occupancy equipping with static backgrounds and dynamic objects, and a Visual Scene Synthesis module that transforms this data into high-fidelity, multi-view video outputs, ensuring spatial and temporal consistency. By providing a dynamic and realistic simulation environment, DrivingSphere enables comprehensive testing and validation of autonomous driving algorithms, ultimately advancing the development of more reliable autonomous cars. The benchmark will be publicly released.", + "bbox": [ + 76, + 319, + 472, + 696 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "1. Introduction", + "text_level": 1, + "bbox": [ + 76, + 728, + 209, + 742 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "End-to-end autonomous driving algorithms [5-7, 15, 19, 20, 25, 36, 42, 48, 53, 54, 59] have made significant progress in recent years, making the accurate evaluation of these models a very urgent task. For safe and responsible evaluation,", + "bbox": [ + 75, + 753, + 468, + 815 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "it is essential to have a precise simulation environment that accurately reflects real-world driving conditions. This requirement usually includes two aspects: firstly, the generation of high-fidelity sensory data, and secondly, the implementation of closed-loop feedback mechanisms.", + "bbox": [ + 500, + 287, + 890, + 363 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "The most common evaluation methods involve the forecasting of waypoints along fixed routes within collected public datasets [3, 37], that is, open-loop simulation approaches. While these benchmarks offer realistic driving data, they have relatively fixed distributions and lack diversity [24], limiting their ability to assess the generalization capabilities of autonomous driving algorithms. Even though several recent works [13, 43, 57] excel at producing diverse, photo-realistic data, continue to rely on the prediction of waypoints that are pre-established. In summary, despite high-fidelity sensory data, these open-loop evaluation solutions cannot provide the dynamic feedback necessary for assessing how autonomous systems respond to dynamic changes and decision-making.", + "bbox": [ + 496, + 366, + 892, + 578 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "In contrast, closed-loop simulations [10, 23, 46, 56] offer feedback-driven systems where an agent's actions influence and are influenced by other agents and the environment. For example, traffic flow-based simulation methods [12, 46, 56] successfully enable multi-agent simulations. However, they lack the ability to process visual sensor inputs, limiting their interplay with vision-based end-to-end models. Game engine-based simulators [10, 23] create scalable and physically realistic environments, but their outputs often diverge from real-world sensor data, restricting their utility in validating algorithms trained on actual inputs.", + "bbox": [ + 496, + 580, + 893, + 747 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "To address these challenges, we propose DrivingSphere, a novel generative closed-loop simulation framework that leverages geometric-wise prior information to generate realistic and controllable driving scenarios. In comparison with existing simulations, DrivingSphere has three distinguished features: ① Rich simulation granularity. Unlike past methods that model only roads and cars [13, 51], our approach allows for including previously unmodeled elements, such as buildings, vegetation, and other environmental structures. Although these non-traffic elements are not", + "bbox": [ + 496, + 750, + 893, + 901 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "CVF", + "bbox": [ + 106, + 2, + 181, + 42 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore.", + "bbox": [ + 236, + 0, + 810, + 44 + ], + "page_idx": 0 + }, + { + "type": "page_footnote", + "text": "*Corresponding author: Jianbing Shen. This work was supported in part by the Science and Technology Development Fund of Macau SAR (FDCT) under grants 0102/2023/RIA2 and 0154/2022/A3 and 001/2024/SKL, the Jiangyin Hi-tech Industrial Development Zone under the Taihu Innovation Scheme (EF2025-00003-SKL-IOTSC), the University of Macau SRG2022-00023-IOTSC grant.", + "bbox": [ + 75, + 827, + 468, + 900 + ], + "page_idx": 0 + }, + { + "type": "page_number", + "text": "27531", + "bbox": [ + 478, + 944, + 517, + 955 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/25b82b33c5d51a3320144bd16ea17e52fe7cd31faa60d464b746dcd10e3109a1.jpg", + "image_caption": [ + "Figure 1. Comparison of frameworks for evaluating end-to-end Autonomous Driving (AD) algorithms. (a) Open-loop evaluation uses waypoint predictions along fixed routes in pre-collected datasets [3]. Generative models [13, 43, 57] create diverse, realistic data but lack dynamic feedback to assess AD responses to dynamic changes. (b) Simulated closed-loop evaluation [10, 23] offers feedback-driven, scalable environments where agent actions impact simulation dynamics; however, sensory outputs often differ from real-world data, limiting effectiveness for algorithms trained on real data. (c) Our generative closed-loop simulation framework, *DrivingSphere*, addresses these limitations by delivering realistic visual inputs and continuous, responsive feedback between the AD agent and environment." + ], + "image_footnote": [], + "bbox": [ + 83, + 89, + 344, + 255 + ], + "page_idx": 1 + }, + { + "type": "image", + "img_path": "images/b28a90788599772a538a3d8c7ecc3fbd7075c7e529e415d2d9894d3d5d01cd3b.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 346, + 90, + 580, + 255 + ], + "page_idx": 1 + }, + { + "type": "image", + "img_path": "images/35657bbd26f8ca421f8cbbdc3ecd32048e8d534e2b1bd8693faf728c134b1a61.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 586, + 90, + 883, + 255 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "directly involved in traffic flow, their presence can significantly impact the input to driving models, thereby influencing decision-making processes in complex driving scenarios. 2 Physical and spatial realism. Due to our model explicitly representing the scene and traffic actors in 4D space, it is capable of precisely depicting physical interactions and occlusion relationships across different traffic elements. This ensures that each viewpoint and position naturally adheres to physical principles like depth and occlusion, thus achieving structured coordination of global road layout, traffic participants, and their behaviors. 3 High visual consistency and fidelity. Our model pays more effort to associate each traffic actor's appearance and unique ID within the scene, providing stable and high-fidelity temporal and spatial coherence across frames and views.", + "bbox": [ + 75, + 357, + 472, + 583 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "The proposed DrivingSphere contains two key modules: Dynamic Environment Composition and Visual Scene Synthesis. First, the Dynamic Environment Composition module aims to produce a coarse-grained 4D world with an occupancy grid format, which integrates large static backgrounds with dynamic traffic actors. For this, we introduce a BEV-conditioned occupancy diffusion model, named OccDreamer, to generate static backgrounds. It provides an infinite city-scale scene via progressive region extension. Besides, we build an actor bank to manage key actors or objects, and update actors' spatial-temporal position.", + "bbox": [ + 75, + 590, + 468, + 758 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Once the scene is generated, the second module, Visual Scene Synthesis, converts it into high-fidelity, multi-view video outputs for AD system perception. To fully capture the complexity of the 4D environment, we employ a dual-path condition encoding strategy. Specifically, the global branch, using a pretrained 4D encoder [41], extracts geometric information and spatial-temporal relationships directly from the occupancy data, while the local projection branch produces view-specific semantic maps that align on", + "bbox": [ + 75, + 763, + 470, + 902 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "a pixel-to-pixel basis, accurately capturing occlusions and depth variations. To further ensure visual consistency, we design an ID-aware actor encoding mechanism that binds the appearance and identity information of traffic actors to their scene locations, which preserves spatial correlation across views and temporal consistency across frames. With the custom-designed dual-path condition encoding and ID-aware actor encoding, the Visual Scene Synthesis generates spatially and temporally coherent video sequences. This results in a highly realistic simulation environment that is ideal for testing and validating AD systems.", + "bbox": [ + 496, + 357, + 893, + 523 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "By responsive interplay between AD algorithms and the simulated environment, DrivingSphere supports a closed-loop feedback loop, where actions from AD systems [6, 19, 20, 36] influence surrounding actors and scenes. This simulation platform closely mirrors real-world conditions, providing a reliable foundation for comprehensive AD validation. Our extensive experiments show that DrivingSphere achieves superior performance in visual fidelity and temporal consistency, significantly reducing the domain gap between simulated and real-world environments. By excelling in both open-loop and closed-loop evaluations, DrivingSphere proves to be a reliable platform for dynamic agent-environment interplay in rigorous AD testing. Ultimately, DrivingSphere will drive progress toward the development of safer and more reliable autonomous vehicles. Overall, our contributions are four-fold:", + "bbox": [ + 496, + 530, + 893, + 772 + ], + "page_idx": 1 + }, + { + "type": "list", + "sub_type": "text", + "list_items": [ + "- We introduce DrivingSphere, a novel geometry-aware closed-loop simulation framework that captures 2D visual and 3D geometric properties while seamlessly integrating with vision-based end-to-end driving agents.", + "- We propose the Dynamic Environment Composition that provides a comprehensive 4D driving world by modeling both static and dynamic elements, enabling intricate simulations across diverse environments." + ], + "bbox": [ + 500, + 780, + 893, + 898 + ], + "page_idx": 1 + }, + { + "type": "page_number", + "text": "27532", + "bbox": [ + 478, + 944, + 519, + 957 + ], + "page_idx": 1 + }, + { + "type": "list", + "sub_type": "text", + "list_items": [ + "- We design the Visual Scene Synthesis that enables long-term consistency and high fidelity, addressing challenges in generating continuous and multi-view video sequences for AD simulations.", + "- We introduce the Agent Coordination module for scalable multi-agent systems, enabling the integration of multiple vehicle agents and establishing a crucial feedback loop for validating autonomous driving algorithms." + ], + "bbox": [ + 76, + 90, + 467, + 212 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "2. Related Work", + "text_level": 1, + "bbox": [ + 76, + 224, + 217, + 239 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Generative Models in Driving. Denoising Diffusion Probabilistic Models (DDPMs) [17] have transformed image generation by iteratively refining Gaussian noise into high-quality images. Based on this technique, generative models in autonomous driving have greatly advanced. For example, BEVGen [38] and BEVControl [50] use diffusion models to produce controllable street-view images from BEV layouts. More recently, diffusion-based video generation models [2, 16] enable methods like [13, 14, 43-45, 57] to generate multi-view driving videos, effectively supporting perception tasks. Inspired by these 2D methods, advances in 3D scene generation have also emerged. Integrating diffusion processes into LiDAR generation improves the realism of LiDAR scenes [18, 30, 34, 49, 61, 62]. Occupancy-based models have further progressed, with occupancy grids now central to 3D scene generation. SemCity [22] enhances outdoor 3D occupancy generation, while models like OccWorld [58] and OccSora [41] use occupancy data for vehicle motion prediction and trajectory-conditioned 4D occupancy generation, respectively. Unlike these, our OccDreamer combines text and BEV conditions to generate 3D occupancy data, dynamically enhanced with foreground motion and ego pose to form a complete 4D driving world. By conditioning on occupancy data, our VideoDreamer allows precise control over non-traffic elements and ensures spatial and temporal coherence, capturing essential physical relationships such as depth and occlusion.", + "bbox": [ + 76, + 250, + 468, + 657 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Autonomous Driving Simulations. Autonomous driving simulation methods are developed from basic behavioral models to complex, interactive environments. Flow-based simulations [9, 11, 12, 39, 46, 56] enable multi-agent interactions through long-term traffic planning. However, these methods typically operate in 2D space and lack real-life visual outputs. Platforms like CARLA [10], SUMO [27], and MetaDrive [23] address these limitations by modeling traffic dynamics and supporting multi-agent interactions with visual contexts, but they have visual fidelity differences from real-world data, limiting their direct use with algorithms trained on such data. While reconstruction-based methods [21, 26, 29] and frameworks such as UniSim [52], and MARS [47] capture high-fidelity scenes, they are limited by the scope of the data used, restricting their ability to simulate varied and long-tail scenarios. DriveArena [51]", + "bbox": [ + 76, + 659, + 467, + 900 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "recently advanced generative closed-loop simulation by using 2D traffic sketches to control visual outputs, integrating well with vision-based driving agents. However, reliance on 2D inputs constrains its modeling granularity and visual realism. Our approach overcomes these challenges by leveraging 4D occupancy data, generating visually coherent and geometrically accurate scenes that rigorously test driving agents in closed-loop simulations, equipping them to handle complex real-world conditions effectively.", + "bbox": [ + 496, + 90, + 890, + 227 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3. Methodology", + "text_level": 1, + "bbox": [ + 500, + 239, + 633, + 257 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "DrivingSphere acts as a generative closed-loop simulation framework that combines occupancy-based 4D world modeling with advanced video rendering techniques, providing high-fidelity visual outputs that enhance both simulation realism and agent-environment interplay in autonomous driving scenarios. Specifically, DrivingSphere begins with Dynamic Environment Composition (Sec. 3.1), which generates static background from a map sketch, selects traffic participants from the actor bank, and updates the actor location, to formulate a 4D driving environment with a format of occupancy. Next, Visual Scene Synthesis (Sec. 3.2) conditions the occupancy data surrounding the ego vehicle, accurately capturing occlusion relationships and fine-grained semantic information to generate high-fidelity multi-view videos. Finally, the closed-loop feedback mechanism (Sec. 3.3) enables dynamic, responsive adjustments, where autonomous agents continuously receive updated visual data and generate control signals that modify the simulation environment, providing a comprehensive platform for algorithm testing and refinement. The overall framework is shown in Fig. 2.", + "bbox": [ + 496, + 263, + 890, + 566 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3.1. Dynamic Environment Composition", + "text_level": 1, + "bbox": [ + 500, + 575, + 812, + 590 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Previous driving simulation methods have often overlooked static and varied elements like buildings, obstacles, and vegetation [10, 51]. Although these elements are not direct traffic participants, they are part of the perception input of AD systems, influencing final driving decision-making. For instance, static objects like buildings may block sensors' line-of-sight, causing occlusions of other vehicles or pedestrians. Obstacles or vegetation can introduce sensor artifacts or false positives for AD systems. Both situations may result in incorrect decisions like unnecessary braking or lane changes. To address this issue, we propose a Dynamic Environment Composition to construct a comprehensive 4D Driving World equipping complex digital assets.", + "bbox": [ + 496, + 598, + 890, + 794 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "4D Driving World Representation. Our 4D driving world representation includes three critical digital assets: static background scenes, dynamic foreground actors, and spatial position of these actors. Formally, we define this representation as $\\pmb{W} = \\{\\pmb{S}_{\\mathrm{city}},(\\pmb{A}_1,\\pmb{P}_1),\\dots ,(\\pmb{A}_N,\\pmb{P}_N)\\}_{n = 1}^N$ . Here, $S_{\\mathrm{city}}$ is a composite of multiple regional static scene $\\{S_1,\\ldots ,S_K\\}$ . Each $S_{k}\\in \\mathbb{R}^{H\\times W\\times D}$ is the $k^{th}$ static back", + "bbox": [ + 496, + 795, + 890, + 900 + ], + "page_idx": 2 + }, + { + "type": "page_number", + "text": "27533", + "bbox": [ + 478, + 944, + 517, + 955 + ], + "page_idx": 2 + }, + { + "type": "image", + "img_path": "images/c0a17a0c91e0ce25c99b5a8dd7ade49231f8cd6eae4e025749f89be402c165a2.jpg", + "image_caption": [ + "Figure 2. Overview of DrivingSphere framework. (a) The Dynamic Environment Composition module builds a 4D driving world, simulating real driving scenarios with backgrounds generated by OccDreamer, dynamic actors from an actor bank, and trajectories guided by a transportation simulator. (b) The Visual Scene Synthesis produces high-fidelity, photo-realistic video frames conditioned on global semantics, view-specific details, and scene prompts, supporting close-loop evaluation. Control signals enable adaptive feedback for driving agents, facilitating continuous testing and evaluation of driving algorithms in the simulated environment." + ], + "image_footnote": [], + "bbox": [ + 86, + 95, + 890, + 330 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "ground scene that captures the spatial layout and static elements within the region. $A_{n}\\in \\mathbb{R}^{4}$ is the $n^{th}$ actor of $N$ actors, such as vehicles and pedestrians, defined by a 3D coordinate and a semantic label. It is worth noticing that both the static scene $\\pmb{S}_{k}$ and dynamic actor $A_{n}$ are represented as occupancy grids. $P_{n} = \\left\\{P_{n}^{0},\\dots ,P_{n}^{T}\\right\\}_{t = 0}^{T}$ describe actor position across $T$ frames, where $P_{n}^{t}\\in \\mathbb{R}^{4}$ consists of a 3D relative position of the $n$ -th actor at time $t$ with respect to the ego vehicle and a yaw rotation $\\theta$ . In the following, we will detail how to generate three assets.", + "bbox": [ + 75, + 406, + 470, + 561 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Static Scene Generation. A straightforward approach generating static 3D scenes $S_{\\mathrm{city}}$ is directly using ground truth occupancy data from existing datasets, like the Boston area in nuScenes [3]. However, this method is limited to specific regions captured during data collection, restricting its applicability to other urban areas. To address this limitation, we propose an occupancy diffusion model, OccDreamer, which is conditioned on BEV maps and textual descriptions, capable of generating static scenes for any desired urban region. The framework of OccDreamer, which integrates the following components, is illustrated in Fig. 3.", + "bbox": [ + 75, + 561, + 468, + 727 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Firstly, to efficiently train the diffusion model while addressing the computational complexity of processing 3D data, we adopt VQVAE [17, 35, 40] as an occupancy tokenizer, mapping $S_{k}$ to a latent feature $Z^{S}$ : $Z^{S} = \\mathcal{F}_{\\mathrm{VAE,Enc}}^{occ}(S_{k})$ . The reconstructed scene is defined as $S_{k}^{\\prime} = \\mathcal{F}_{\\mathrm{VAE,Dec}}^{occ}(Z^{S_{k}})$ . $\\mathcal{F}_{\\mathrm{VAE}}^{occ}$ is trained with a combined loss $\\mathcal{L}_{occ}$ :", + "bbox": [ + 76, + 727, + 468, + 821 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {L} _ {o c c} = \\mathcal {L} _ {C E} \\left(\\boldsymbol {S} _ {k}, \\boldsymbol {S} _ {k} ^ {\\prime}\\right) + \\alpha \\mathcal {L} _ {L o v} \\left(\\boldsymbol {S} _ {k}, \\boldsymbol {S} _ {k} ^ {\\prime}\\right) + \\beta \\mathcal {L} _ {e m b}, \\tag {1}\n$$\n", + "text_format": "latex", + "bbox": [ + 91, + 830, + 468, + 848 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "where $\\mathcal{L}_{CE}$ is the CE loss and $\\mathcal{L}_{Lov}$ is the Lovász loss [1].", + "bbox": [ + 76, + 854, + 464, + 869 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Secondly, considering varying road structures and complex regional topographies, we propose a controllable re", + "bbox": [ + 76, + 869, + 468, + 901 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "gion occupancy generation module. It accepts a regional BEV map capturing the road structure and a language prompt describing abstract features (like suburban areas with rich vegetation or commercial districts with buildings lining both sides) as input, and outputs a region occupancy $S_{k}$ . Following the principles of diffusion models [17, 55], a CLIP encoder [31] converts the text prompt into an embedding $F_{\\mathrm{region}}$ . Then, $F_{\\mathrm{region}}$ is injected into the denoiser $\\epsilon_{\\theta}^{s}$ via cross-attention. Simultaneously, the BEV map is processed by a pre-trained image VAE [17] to extract the corresponding road embedding $F_{M}$ , serving as input for the ControlNet branch $\\epsilon_{\\phi}^{s}$ . This enables precise control over the diffusion learning process in the latent space:", + "bbox": [ + 496, + 406, + 890, + 603 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {L} _ {\\text {o c c}} = \\mathbb {E} _ {\\boldsymbol {F} _ {M}, \\boldsymbol {F} _ {\\text {r e g i o n}}, \\boldsymbol {Z}} s _ {k, \\epsilon^ {s}, \\tau} \\left[ \\left\\| \\epsilon^ {s} - \\epsilon_ {\\theta} ^ {s} \\left(\\boldsymbol {Z} _ {\\tau} ^ {\\boldsymbol {S} _ {k}}, \\boldsymbol {F} _ {\\text {r e g i o n}}, \\tau , \\epsilon_ {\\phi} ^ {s} (\\boldsymbol {F} _ {M})\\right) \\right\\| ^ {2} \\right], \\tag {2}\n$$\n", + "text_format": "latex", + "bbox": [ + 500, + 608, + 906, + 643 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "where $Z_{\\tau}^{S_k}$ is the noisy version of $Z^{S_k}$ by adding $\\tau$ steps noise $\\epsilon^s$ . After training, we use OccDreamer to create $Z_{k}^{S}$ and decode it with $\\mathcal{F}_{\\mathrm{VAE,Decode}}^{\\mathrm{occ}}$ to obtain $S_{k}$ .", + "bbox": [ + 496, + 643, + 890, + 691 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Thirdly, for spatial consistency across the entire 3D scene, we propose a Scene Expansion Mechanism for building a coherent and city-level static scene $S_{\\mathrm{city}}$ . Its core idea is to expand an initial region $S_{k}$ to a neighboring region $S_{k+1}$ by using the overlap between these regions as a conditioning constraint. Formally, let $O$ represent a binary mask that defines the overlapping area between two adjacent regions $S_{k}$ and $S_{k+1}$ . This overlap serves as the conditional context for generating the next region, enforcing spatial coherence at the boundary. To generate $S_{k+1}$ , we first create a partially masked scene $S_{k}^{\\mathrm{partial}}$ by applying $O$ to $S_{k}$ :", + "bbox": [ + 496, + 690, + 890, + 859 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\boldsymbol {S} _ {k} ^ {\\text {p a r t i a l}} = \\boldsymbol {S} _ {k} \\odot \\boldsymbol {O}, \\tag {3}\n$$\n", + "text_format": "latex", + "bbox": [ + 635, + 864, + 890, + 882 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "where $\\odot$ denotes element-wise multiplication, extracting", + "bbox": [ + 500, + 886, + 890, + 901 + ], + "page_idx": 3 + }, + { + "type": "page_number", + "text": "27534", + "bbox": [ + 478, + 944, + 517, + 955 + ], + "page_idx": 3 + }, + { + "type": "image", + "img_path": "images/b69dc1588143795e3d83e491b935aa255de9e42f0321e925119a0a0e0aa2d57d.jpg", + "image_caption": [ + "Figure 3. The framework of OccDreamer, which includes Occupancy Tokenizer, Region Occupancy Generation, and Scene Extensio to generate a city-level background for 4D driving world." + ], + "image_footnote": [], + "bbox": [ + 83, + 88, + 467, + 333 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "the overlapping section of $S_{k}$ . We then employ a diffusion process to generate the neighboring region $S_{k + 1}$ , conditioning on both the BEV map $M_{k + 1}$ (capturing the road structure of the target area) and the partially masked scene $S_{k}^{\\mathrm{partial}}$ . This diffusion process is represented as:", + "bbox": [ + 75, + 390, + 468, + 467 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\boldsymbol {Z} ^ {\\boldsymbol {S} _ {k + 1}} = \\mathcal {F} _ {\\text {o u t p a i n t i n g}} \\left(\\epsilon^ {s}, \\boldsymbol {M} _ {k + 1}, \\boldsymbol {S} _ {k} ^ {\\text {p a r t i a l}}\\right), \\tag {4}\n$$\n", + "text_format": "latex", + "bbox": [ + 145, + 476, + 468, + 496 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "where $\\epsilon^s$ is the input noise, and $\\mathcal{F}_{\\mathrm{outpainting}}$ denotes the diffusion model conditioned on the masked scene and BEV map. After obtaining the latent representation $Z^{S_{k + 1}}$ , we decode it using the occupancy VAE decoder, resulting in the expanded region $S_{k + 1}$ . The two regions are then combined into a larger scene $S_{\\mathrm{merged}}$ through a merging operation:", + "bbox": [ + 75, + 506, + 468, + 597 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\boldsymbol {S} _ {\\text {m e r g e d}} = \\operatorname {M e r g e} \\left(\\boldsymbol {S} _ {k} \\odot (1 - \\boldsymbol {O}), \\boldsymbol {S} _ {k + 1}\\right), \\tag {5}\n$$\n", + "text_format": "latex", + "bbox": [ + 145, + 607, + 468, + 625 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "where Merge ensures spatial continuity and logical consistency between adjacent regions. The spatially continuous and logically consistent city scene $S_{\\mathrm{city}}$ serves as an excellent digital background for the 4D driving world.", + "bbox": [ + 75, + 628, + 468, + 688 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Dynamic Actor Selection. To complement static scenes, we populate the 4D driving world with dynamic actors, creating realistic traffic flows. An actor bank $\\mathcal{B}$ is constructed, where each actor $\\mathbf{A}_n \\in \\mathcal{B}$ consists of instance-level voxel data segmented from real or synthetic occupancy scenes, with unique geometric, semantic, and behavioral attributes, as well as a descriptive caption $L_{\\mathbf{A}_n}$ and an ID $I_{\\mathbf{A}_n}$ to ensure consistency across frames. Actors are selected based on semantic similarity to user-provided descriptions via CLIP, or randomly sampled from relevant categories if unspecified, ensuring contextual diversity. This flexible selection process allows for dynamically relevant and diverse actor integration within the 4D driving world, supporting realistic and adaptable traffic simulation.", + "bbox": [ + 75, + 689, + 470, + 900 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Spatial Position of Actors. The generation of actor positions $\\pmb{P}$ is driven by the control signals output from autonomous agents (Sec. 3.3). Each actor is assigned an control signal set $C_n = \\{c_n^0, \\dots, c_n^T\\}$ , where $c_n^t$ represents the received control signal of actor $n$ at time $t$ . These signals iteratively update the dynamic state of the world. For $n$ -th actor, we have the position:", + "bbox": [ + 496, + 90, + 890, + 196 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\boldsymbol {P} _ {n} ^ {t + 1} = \\boldsymbol {P} _ {n} ^ {t} + \\triangle \\boldsymbol {P} _ {n} ^ {t} \\left(\\boldsymbol {c} _ {n} ^ {t}\\right), \\tag {6}\n$$\n", + "text_format": "latex", + "bbox": [ + 614, + 202, + 890, + 219 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "where $\\triangle P_n^t (c_N^t)$ is the positional change resulting from the control signal $c_{N}^{t}$ at time $t$", + "bbox": [ + 496, + 220, + 890, + 252 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "4D World Composition. With the static background, dynamic foreground actors and their positions calculated, we integrate them into the comprehensive 4D driving world. The world state at any time $t$ is represented by:", + "bbox": [ + 496, + 252, + 890, + 313 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\boldsymbol {W} ^ {t} = \\pi (\\boldsymbol {S}, \\left(\\boldsymbol {A} _ {n}, \\boldsymbol {c} _ {n} ^ {t}\\right) _ {n = 1} ^ {N}), \\tag {7}\n$$\n", + "text_format": "latex", + "bbox": [ + 609, + 319, + 890, + 339 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "where $\\pi$ is the composition operator that aligns each actor based on their current pose $P_{n}^{t}$ , represents the set of poses for all actors, forming the complete 4D driving world at time $t$ . By iterating this process over time, we construct the dynamic 4D world that accounts for actor interplay, occlusions, and the evolving scene geometry.", + "bbox": [ + 496, + 340, + 890, + 431 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "3.2. Visual Scene Synthesis", + "text_level": 1, + "bbox": [ + 498, + 439, + 709, + 455 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Previous generative models [13, 28, 43, 57] tend to employ 2D visual conditions, which fail to accurately capture the geometric and semantic complexities inherent in real-world driving scenes. In response, our Visual Scene Synthesis employs VideoDreamer, a custom-designed video diffusion model, to transform the occupancy-driven 4D driving world constructed in the previous section into high-fidelity visual results. The overall framework is shown in Fig. 4. Specifically, we introduce a Dual-path Condition Encoding strategy that focuses on encoding occupancy data as its primary condition. We further enhance the consistency of appearances across views and frames by developing an ID-aware Actor Encoding method. Finally, we integrate the Spatial-Temporal Diffusion Transformer (ST-DiT) from OpenSora [60] as the foundational technology to ensure visual consistency and generate artifact-free frames.", + "bbox": [ + 496, + 462, + 890, + 703 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Dual-path Condition Encoding. We design a dual-path condition encoding strategy aimed at effectively capturing occupancy data. Given a set of driving world data $\\mathbf{W} = \\{\\mathbf{W}^0, \\mathbf{W}^1, \\dots, \\mathbf{W}^T\\}$ spanning frames from 0 up to $T$ , we first encode them into a global feature $\\mathbf{F}_{\\mathrm{global}} = \\mathcal{F}_{\\mathrm{VAE}}^{\\mathrm{4Docc}}(\\mathbf{W})$ , capturing comprehensive geometric information and the spatial-temporal relationships within the scene. Here, $\\mathcal{F}_{\\mathrm{VAE}}^{\\mathrm{4Docc}}$ is the well-trained 4D encoder from occupancy VAE [41], responsible for capturing global geometry of the scene. This ensures that the overall structure and spatial layout of the scene are accurately represented.", + "bbox": [ + 496, + 704, + 890, + 869 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Meanwhile, for $t^{th}$ frame (where $t = 0,\\dots,T$ ), 3D occupancy data $\\mathbf{W}^t$ is projected into 2D semantic maps $\\mathcal{M}_v^t$ for", + "bbox": [ + 496, + 869, + 890, + 901 + ], + "page_idx": 4 + }, + { + "type": "page_number", + "text": "27535", + "bbox": [ + 478, + 944, + 517, + 955 + ], + "page_idx": 4 + }, + { + "type": "image", + "img_path": "images/e45d3593a899356336edcaa9f22d52370655dd9052d2eebc971e830c92a63463.jpg", + "image_caption": [ + "Figure 4. Overview of the VideoDreamer. The model conditions on 4D driving world and enriched actor embeddings (e.g., actor ID, position and caption). This ensures high visual fidelity and geometric consistency in the generated driving simulations. Each view $v$ by camera extrinsics $\\mathbf{T}^t$ and intrinsics $\\mathbf{K}_v$ :" + ], + "image_footnote": [], + "bbox": [ + 83, + 90, + 467, + 311 + ], + "page_idx": 5 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {M} _ {v} ^ {t} = \\mathcal {Q} \\left(\\boldsymbol {W} ^ {t}, \\mathbf {K} _ {v}, \\mathbf {T} ^ {t}\\right), \\tag {8}\n$$\n", + "text_format": "latex", + "bbox": [ + 194, + 393, + 468, + 411 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "where $\\mathcal{Q}$ is the render and projection function [33]. This process yields a sequence of semantic maps, $\\mathcal{M} \\in \\mathbb{R}^{T \\times 6 \\times H \\times W}$ , with precise pixel-to-pixel alignment across views. Such alignment is essential for accurately capturing view-specific occlusions and depth variations. $\\mathcal{M}$ is then encoded into view-specific features $F_{\\mathrm{view}} = \\mathcal{F}_{\\mathrm{VAE}}^{\\mathrm{Image}}(\\mathcal{M})$ . $F_{\\mathrm{view}}$ and $F_{\\mathrm{global}}$ are then integrated through cross-attention to obtain the unified representation $F_{\\mathrm{occ}}$ .", + "bbox": [ + 75, + 415, + 468, + 536 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "ID-aware Actor Encoding. To ensure coherence of each actor's appearance and identity information across the scene, we extract a fused sequence embedding $F_{\\mathrm{fuse}}$ :", + "bbox": [ + 75, + 536, + 468, + 582 + ], + "page_idx": 5 + }, + { + "type": "equation", + "text": "\n$$\n\\boldsymbol {F} _ {\\text {f u s e}} = \\operatorname {C O N C A T} \\left[ \\boldsymbol {F} _ {\\boldsymbol {W}}, \\boldsymbol {F} _ {\\boldsymbol {A} _ {1}}, \\dots , \\boldsymbol {F} _ {\\boldsymbol {A} _ {N}} \\right], \\tag {9}\n$$\n", + "text_format": "latex", + "bbox": [ + 150, + 590, + 468, + 606 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "where $F_W$ is the embedding of the scene caption $\\mathcal{L}_{\\mathrm{scene}}$ , derived from T5 [32]. Each actor embedding $F_{A_n}$ is:", + "bbox": [ + 75, + 611, + 468, + 642 + ], + "page_idx": 5 + }, + { + "type": "equation", + "text": "\n$$\n\\boldsymbol {F} _ {\\boldsymbol {A} _ {n}} = \\operatorname {C O N C A T} \\left[ \\mathcal {F} _ {\\text {F o u r i e r}} \\left(\\boldsymbol {P} _ {n}\\right), \\mathcal {F} _ {\\text {F o u r i e r}} \\left(\\boldsymbol {I} _ {n}\\right), \\mathcal {F} _ {T 5} \\left(\\boldsymbol {L} _ {\\boldsymbol {A} _ {i}}\\right) \\right]. \\tag {10}\n$$\n", + "text_format": "latex", + "bbox": [ + 94, + 648, + 468, + 676 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Here, $P_{n}$ represents the position of the $n$ -th actor and $I_{n}$ is its unique ID. Both are encoded using Fourier encoding [29]. The actor caption $L_{A_i}$ is encoded using the T5 [32]. Overall, by integrating geometry positions, textual descriptions, and identity information, we ensure coherent behavior and long-term visual consistency across frames.", + "bbox": [ + 75, + 674, + 468, + 763 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "ControlNet-DiT. To enhance visual fidelity and temporal consistency in generated videos, we integrate the ST-DiT [60] as our denoiser $\\epsilon_{\\theta}^{v}$ , utilizing ST-DiT blocks stacked with View-aware Spatial Self-Attention (VSSA), Temporal Self-Attention, Cross-Attention, and FFN.", + "bbox": [ + 75, + 763, + 468, + 838 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "In the video generation process, the input video sequence $\\mathbf{V} \\in \\mathbb{R}^{v \\times T \\times H \\times W}$ is first encoded into latent features $Z^{\\mathbf{V}} \\in \\mathbb{R}^{v \\times T \\times H' \\times W' \\times C}$ . To ensure multi-view consistency while managing computational complexity, we mod-", + "bbox": [ + 76, + 839, + 468, + 900 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/0045c73d182897ed8245df6fd9a28d3bf755bf90f20c45357484a7fd1c0013fc.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 498, + 89, + 696, + 181 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/77719c565930512c098d4fa8afe8ad5394909ae1e69dc07f626950a3e246d90a.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 702, + 90, + 890, + 181 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/857accd1b061e6fc3e1ffbc5c3c56509b22384a59361029f5893695d78415640.jpg", + "image_caption": [ + "Figure 5. Qualitative results of generated 3D scene. OccDreamer uses text prompt and bev map as guidance to generate 3D scenes with controllable regional content and road structure." + ], + "image_footnote": [], + "bbox": [ + 506, + 184, + 880, + 253 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "if the original Spatial self-attention [60] as VSSA. The latent feature $Z^{\\mathbf{V}}$ is reshaped from $v\\times T\\times H^{\\prime}\\times W^{\\prime}\\times C$ to $T\\times (vH^{\\prime}W^{\\prime})\\times C$ , treating the combined view, height, and width dimensions $(vH^{\\prime}W^{\\prime})$ as the sequence length. This view-flatted approach reduces the parameter overhead associated with cross-view attention [13, 44] while maintaining spatial coherence across different views. For external conditions, we apply a cross-attention layer that injects the fused sequence embedding $F_{\\mathrm{fuse}}$ (derived from scene context and actor identities) as: $Z^{V} = \\mathrm{CrossAttention}(Z^{V},F_{\\mathrm{fuse}})$ . Last but not least, we introduce ControlNet branch $\\epsilon_{\\phi}^{v}$ to ensure the occupancy control, inspired by PixArt-δ [4]. This branch duplicates blocks from $\\epsilon_{\\theta}^{v}$ and takes $F_{\\mathrm{occ}}$ as the input. The final diffusion process is:", + "bbox": [ + 496, + 301, + 892, + 513 + ], + "page_idx": 5 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {L} _ {\\text {v i d e o}} = \\mathbb {E} _ {\\boldsymbol {F} _ {\\text {o c c}}, \\boldsymbol {F} _ {\\text {f u s e}}, \\boldsymbol {Z} V, \\epsilon^ {v}, \\tau} \\left[ \\left\\| \\epsilon^ {v} - \\epsilon_ {\\theta} ^ {v} \\left(\\boldsymbol {Z} _ {\\tau} ^ {\\boldsymbol {V}}, \\boldsymbol {F} _ {\\text {f u s e}}, \\tau , \\epsilon_ {\\phi} ^ {v} \\left(\\boldsymbol {F} _ {\\text {o c c}}\\right)\\right) \\right\\| ^ {2} \\right]. \\tag {11}\n$$\n", + "text_format": "latex", + "bbox": [ + 498, + 521, + 893, + 555 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "This approach ensures fine-grained control over spatial and temporal coherence, producing artifact-free frames that meet the high fidelity requirements for AD simulation.", + "bbox": [ + 496, + 553, + 890, + 597 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Auto-regressive Generation. To generate temporally consistent and content-coherent video frames, we employ an auto-regressive generation strategy [60] based on first- $f$ frame mask. This approach models the dependency of each video frame on the previous frames, ensuring smooth transitions and maintaining visual continuity over time:", + "bbox": [ + 496, + 598, + 890, + 688 + ], + "page_idx": 5 + }, + { + "type": "equation", + "text": "\n$$\n\\boldsymbol {V} ^ {t} = \\mathcal {F} _ {\\text {v i d e o}} (\\boldsymbol {V} ^ {t - 1}, \\boldsymbol {Z} ^ {V}, t),\n$$\n", + "text_format": "latex", + "bbox": [ + 602, + 696, + 787, + 713 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "where $V^t$ is the video frame at time $t$ , and $\\mathcal{F}_{\\mathrm{video}}$ is the video generation function that conditions on the latent representation $Z^{V}$ and the previous frame $V^{t-1}$ . This ensures temporal coherence and infinite video sequences.", + "bbox": [ + 496, + 717, + 890, + 777 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "3.3. Agent Interplay and Closed-Loop Simulation", + "text_level": 1, + "bbox": [ + 498, + 786, + 880, + 803 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "We enable seamless coordination of autonomous driving agents within the DrivingSphere simulation environment, categorizing agents into two primary types: the Ego Agent and Environment Agents.", + "bbox": [ + 496, + 809, + 890, + 869 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Ego Agent: The Ego Agent represents the autonomous driving system under evaluation. Driven by end-to-end", + "bbox": [ + 496, + 869, + 890, + 900 + ], + "page_idx": 5 + }, + { + "type": "page_number", + "text": "27536", + "bbox": [ + 478, + 944, + 519, + 955 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "models [6, 19, 20], it receives visual input frames $V^{t}$ and outputs the predicted control signal $c_{\\mathrm{ego}}^{t + 1}$ at each time $t$ :", + "bbox": [ + 76, + 90, + 468, + 125 + ], + "page_idx": 6 + }, + { + "type": "equation", + "text": "\n$$\n\\boldsymbol {c} _ {\\text {e g o}} ^ {t + 1} = \\mathcal {F} _ {\\text {d r i v i n g}} \\left(\\boldsymbol {V} ^ {t}\\right), \\tag {12}\n$$\n", + "text_format": "latex", + "bbox": [ + 205, + 133, + 468, + 152 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "where $\\mathcal{F}_{\\mathrm{driving}}$ denotes the driving algorithm [6, 19, 20].", + "bbox": [ + 76, + 161, + 442, + 176 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Environment Agents: Environment Agents are responsible for controlling the behavior and actions of all other actors within the simulation world $W$ . To achieve realistic info exchanges, we utilize the traffic flow engine [46], which supports multi-agent simulation. Environment Agents receive inputs from the simulation's state and output control signal $c_{\\mathrm{n}}^{t + 1}$ that dictate the movements and interplay of actors. Formally, this process can be described as:", + "bbox": [ + 75, + 176, + 468, + 296 + ], + "page_idx": 6 + }, + { + "type": "equation", + "text": "\n$$\n\\boldsymbol {c} _ {\\mathrm {n}} ^ {t + 1} = \\mathcal {F} _ {\\text {s y s t e m}} (\\boldsymbol {W} ^ {t}), \\tag {13}\n$$\n", + "text_format": "latex", + "bbox": [ + 202, + 306, + 468, + 324 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "where $\\mathcal{F}_{\\mathrm{system}}$ denotes the transportation simulator.", + "bbox": [ + 76, + 334, + 410, + 349 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "These agents take as input the visual frames $V^t$ or previous state $\\pmb{W}^{t}$ , and output actors' respective control signals $\\pmb{c}_{\\mathrm{ego}}^{t + 1}$ and $\\pmb{c}_{\\mathrm{n}}^{t + 1}$ . These signals influence the actor positions using Eq. (6) and further dynamically update the 4D driving world $\\pmb{W}^{t + 1}$ , maintaining the closed-loop simulation where the environment continuously adapts to the agents' actions.", + "bbox": [ + 76, + 349, + 468, + 441 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "4. Experiments", + "text_level": 1, + "bbox": [ + 76, + 454, + 207, + 470 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "4.1. Settings", + "text_level": 1, + "bbox": [ + 76, + 479, + 174, + 494 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Dataset and Evaluation Metrics. We utilize real-world driving datasets nuScenes [3]. These datasets provide rich sensor data, including LiDAR point clouds, camera images, and high-definition maps, which are essential for generating accurate occupancy grids and semantic information. We employ the following metrics to evaluate the performance of our simulation framework: FVD and UniAD-based metrics [19] for visual quality and fidelity. As for the evaluation of OccDreamer, we use FID and MMD [22]. In open-loop evaluation, we borrow metrics from NAVSIM [8]. RC and ADS for the closed-loop following [51].", + "bbox": [ + 75, + 503, + 468, + 667 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Implement Details are included in the supplementary.", + "bbox": [ + 76, + 670, + 436, + 685 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "4.2. Visual Fidelity and Realism", + "text_level": 1, + "bbox": [ + 76, + 695, + 323, + 710 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "We first evaluate the fidelity of the voxel-based scenes from OccDreamer and video sequences from VideoDreamer against real nuScenes data to assess the domain gap.", + "bbox": [ + 76, + 719, + 468, + 763 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Scene Generation. We conduct quantitative and qualitative analyses to assess the fidelity of OccDreamer's generated occupancy data. Our method outperforms existing approaches quantitatively, as shown in Tab. 2. Specifically, SemCity employs a VAE trained on three planes to generate occupancy data while OccDreamer achieves pixel-to-pixel alignment based on BEV maps, resulting in more robust and accurate occupancy generation. Qualitatively, Fig. 5 presents the results of OccDreamer with both caption and", + "bbox": [ + 75, + 765, + 468, + 900 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "BEV map. To the best of our knowledge, our approach is the first to achieve text-guided occupancy generation. As demonstrated, abstract textual descriptions effectively guide the model to generate diverse scenes, including outdoor environments and urban landscapes. Under BEV map control, the generated samples accurately reflect road structures and lane configurations. These capabilities establish a solid foundation for the comprehensive modeling of the 4D driving world, ensuring that the generated scenes are both spatially continuous and logically consistent. More visualizations are provided in the supplementary materials.", + "bbox": [ + 496, + 90, + 890, + 257 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Video Generation. We further assess the performance of our video generation model, VideoDreamer, in generating coherent video. We conduct a comparative analysis against leading models such as MagicDrive [13], DriveArena [51], and Panacea [45]. Using UniAD [19] as the evaluation framework, we assess the performance across crucial metrics including 3D object detection, BEV map segmentation, and trajectory planning. Our experimental results, summarized in Tab. 1, indicate that VideoDreamer significantly outperforms the baseline methods in these key metrics. This performance underscores the advantages of occupancy data over traditional 2D conditions, affirming VideoDreamer's effectiveness in creating realistic, dynamic scenes that accurately simulate real-world driving complexities.", + "bbox": [ + 496, + 257, + 892, + 469 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Qualitatively, as shown in Fig. 6, our approach demonstrates markedly superior performance in maintaining coherence and fidelity over time, significantly outperforming others. This success is primarily due to the geometry-aware occupancy encoding and instance encoding that ensures appearance consistency, highlighting VideoDreamer's robustness in creating visually consistent, detailed driving environments that capture real-world scene dynamics. More visualizations are included in the supplementary materials.", + "bbox": [ + 496, + 469, + 890, + 604 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "4.3. Open-loop Evaluation", + "text_level": 1, + "bbox": [ + 500, + 616, + 705, + 632 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "In the open-loop setting, where AD agents passively receive environment input without influencing simulation dynamics, we assess the performance of UniAD [19] on DrivingSphere and alternative simulators. Using 100 predefined trajectories, Tab. 3 demonstrates better PDMS scores achieved by AD agents in DrivingSphere, suggesting that its higher visual fidelity reduces perceptual inaccuracies that can misguide agent decisions. This improvement highlights DrivingSphere's suitability for testing AD systems in environments that closely emulate real-world conditions.", + "bbox": [ + 496, + 638, + 890, + 790 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "4.4. Closed-loop Evaluation", + "text_level": 1, + "bbox": [ + 500, + 801, + 714, + 816 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "In the closed-loop evaluation, AD agents receive visual input and output control signals, thus interactively shaping the simulation. This evaluation setup involves 100 pre-defined trajectories to test across controlled yet varied scenarios. As shown in Tab. 4, UniAD [19] exhibits relatively low Route", + "bbox": [ + 496, + 825, + 890, + 900 + ], + "page_idx": 6 + }, + { + "type": "page_number", + "text": "27537", + "bbox": [ + 478, + 944, + 517, + 955 + ], + "page_idx": 6 + }, + { + "type": "table", + "img_path": "images/202e29fdce86c8ef0906abf1ae09262c9a0a4b2ea6f1396fa126ca8938b29d37.jpg", + "table_caption": [], + "table_footnote": [], + "table_body": "
Source DataQuality3DODSegmentationL2(↓)
FVD(↓)mAP(↑)NDS(↑)Lanes(↑)Drivable(↑)Divider(↑)1.0s2.0s3.0s
ori nuScenes[3]—37.9849.8531.3169.1425.930.510.981.65
MagicDrive [13]218.1212.9228.3621.9551.4617.100.571.141.95
Panacea [45]13913.7227.7318.2352.3717.210.581.141.97
DriveArena [51]185.3216.0630.0326.1459.3720.790.561.101.89
DrivingSphere103.4221.4534.1627.9962.8722.290.541.101.76
", + "bbox": [ + 81, + 87, + 893, + 198 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/b7b56650c7a3f8988505033fe7f558153600651b2314a532f9834720382bb0c1.jpg", + "image_caption": [ + "Figure 6. Comparison of generated driving simulations from (a) Panacea [45], (b) Our method, and (c) Real Scene [3] across multiple views and time steps. Red boxes highlight instance-level appearance consistency, with our method maintaining more coherent actor appearances across frames. Yellow boxes illustrate temporal consistency, where our approach achieves smoother transitions over time. Green boxes and background elements demonstrate our model's capability to capture and accurately model non-direct traffic participants, such as buildings and other environmental features, enhancing realism in the generated scene." + ], + "image_footnote": [], + "bbox": [ + 81, + 244, + 890, + 518 + ], + "page_idx": 7 + }, + { + "type": "table", + "img_path": "images/1e09fc163594c60288afdecd05eae41b240c5f278a979d68077fec3fbdf5b09c.jpg", + "table_caption": [ + "Table 1. Quantitative comparison of video generation fidelity for autonomous driving simulations on the nuScenes validation set. Using UniAD [19] for evaluation on generated data, DrivingSphere achieves high fidelity in video generation, closely matching real-world data across key perception and planning metrics." + ], + "table_footnote": [], + "table_body": "
MethodsFID(↓)MMD(↓)
Semicity [22]6340.251
Ours2740.082
", + "bbox": [ + 81, + 599, + 468, + 643 + ], + "page_idx": 7 + }, + { + "type": "table", + "img_path": "images/b586617ac8e94759fb2b61b792e5a2d6faf500a614d67eeaf165693268613904.jpg", + "table_caption": [ + "Table 2. Quantitative comparison of scene generation performance, following the setup in SemCity. The scene data generated by our method is rendered into 2D images." + ], + "table_footnote": [], + "table_body": "
Data SourceNC(↑)DAC(↑)TTC(↑)PDMS(↑)
nuScenes [3]0.9930.9950.9470.910
DriveArena [51]0.7920.9420.7710.636
DriveArena* [51]0.8290.9640.8120.698
DrivingSphere0.8520.9680.8530.742
", + "bbox": [ + 81, + 685, + 468, + 758 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Completion (RC) scores, with completing an average of only $11.7\\%$ of each route. Compared to DriveArena [51], our simulation yields consistently better performance metrics, and, along with the superior visual fidelity demonstrated in Tab. 1, affirms DrivingSphere's effectiveness as a high-fidelity platform for AD algorithm validation.", + "bbox": [ + 75, + 809, + 470, + 902 + ], + "page_idx": 7 + }, + { + "type": "table", + "img_path": "images/fbeaff437d5147da414590b97a395889f4a29c4b098e75a77f880aad9d70406d.jpg", + "table_caption": [ + "Table 3. Open-loop evaluation of UniAD [19] on original nuScenes sequences, DriveArena's reported results, and our predefined 100-trajectory test for DriveArena* and DrivingSphere." + ], + "table_footnote": [], + "table_body": "
RoutePDMS(↑)RC(↑)ADS(↑)
DriveArena [51]0.69010.06410.0508
DrivingSphere0.72810.11700.0851
", + "bbox": [ + 503, + 599, + 890, + 645 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Table 4. Closed-loop evaluation of UniAD's [19] performance on 100 pre-defined routes.", + "bbox": [ + 498, + 646, + 890, + 675 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "5. Conclusion", + "text_level": 1, + "bbox": [ + 500, + 688, + 619, + 704 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "We presented *DrivingSphere*, a generative closed-loop simulation framework that bridges the gap between traditional closed-loop simulations and open-loop generative models. Through advanced occupancy-based modeling and controllable generation mechanisms, *DrivingSphere* creates realistic, high-fidelity simulation for autonomous driving. Our experiments demonstrate superior visual quality, temporal consistency, and the ability to effectively test AD algorithms in dynamic environments. In future work, we aim to enhance dynamic agent behaviors and incorporate more complex real-world scenarios to further improve the robustness of our simulations.", + "bbox": [ + 496, + 719, + 893, + 900 + ], + "page_idx": 7 + }, + { + "type": "page_number", + "text": "27538", + "bbox": [ + 478, + 944, + 519, + 955 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "References", + "text_level": 1, + "bbox": [ + 78, + 89, + 174, + 106 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[1] Maxim Berman, Amal Rannen Triki, and Matthew B Blaschko. The lovasz-softmax loss: A tractable surrogate for the optimization of the intersection-over-union measure in neural networks. In CVPR, 2018. 4", + "[2] Andreas Blattmann, Tim Dockhorn, Sumith Kulal, Daniel Mendelevitch, Maciej Kilian, Dominik Lorenz, Yam Levi, Zion English, Vikram Voleti, Adam Letts, et al. Stable video diffusion: Scaling latent video diffusion models to large datasets. arXiv preprint arXiv:2311.15127, 2023. 3", + "[3] Holger Caesar, Varun Bankiti, Alex H. Lang, Sourabh Vora, Venice Erin Liong, Qiang Xu, Anush Krishnan, Yu Pan, Giancarlo Baldan, and Oscar Beijbom. nuscenes: A multimodal dataset for autonomous driving. In CVPR, 2020. 1, 2, 4, 7, 8", + "[4] Junsong Chen, Yue Wu, Simian Luo, Enze Xie, Sayak Paul, Ping Luo, Hang Zhao, and Zhenguo Li. Pixart-{\\delta} : Fast and controllable image generation with latent consistency models. arXiv preprint arXiv:2401.05252, 2024. 6", + "[5] Li Chen, Penghao Wu, Kashyap Chitta, Bernhard Jaeger, Andreas Geiger, and Hongyang Li. End-to-end autonomous driving: Challenges and frontiers. T-PAMI, 2024. 1", + "[6] Shaoyu Chen, Bo Jiang, Hao Gao, Bencheng Liao, Qing Xu, Qian Zhang, Chang Huang, Wenyu Liu, and Xinggang Wang. Vadv2: End-to-end vectorized autonomous driving via probabilistic planning. arXiv preprint arXiv:2402.13243, 2024. 2, 7", + "[7] Yuan Chen, Zi-han Ding, Ziqin Wang, Yan Wang, Lijun Zhang, and Si Liu. Asynchronous large language model enhanced planner for autonomous driving. In ECCV, 2025. 1", + "[8] Daniel Dauner, Marcel Hallgarten, Tianyu Li, Xinshuo Weng, Zhiyu Huang, Zetong Yang, Hongyang Li, Igor Gilitschenski, Boris Ivanovic, Marco Pavone, Andreas Geiger, and Kashyap Chitta. Navsim: Data-driven non-reactive autonomous vehicle simulation and benchmarking. In NeurIPS, 2024. 7, 1", + "[9] Wenhao Ding, Yulong Cao, Ding Zhao, Chaowei Xiao, and Marco Pavone. Realgen: Retrieval augmented generation for controllable traffic scenarios. arXiv preprint arXiv:2312.13303, 2023. 3", + "[10] Alexey Dosovitskiy, German Ros, Felipe Codevilla, Antonio Lopez, and Vladlen Koltun. Carla: An open urban driving simulator. In CoRL, 2017. 1, 2, 3", + "[11] Lan Feng, Quanyi Li, Zhenghao Peng, Shuhan Tan, and Bolei Zhou. Trafficgen: Learning to generate diverse and realistic traffic scenarios. In ICRA, 2023. 3", + "[12] Daocheng Fu, Wenjie Lei, Licheng Wen, Pinlong Cai, Song Mao, Min Dou, Botian Shi, and Yu Qiao. Limsim++: A closed-loop platform for deploying multimodal llms in autonomous driving. arXiv preprint arXiv:2402.01246, 2024. 1, 3", + "[13] Ruiyuan Gao, Kai Chen, Enze Xie, Lanqing Hong, Zhenguo Li, Dit-Yan Yeung, and Qiang Xu. Magicdrive: Street view generation with diverse 3d geometry control. arXiv preprint arXiv:2310.02601, 2023. 1, 2, 3, 5, 6, 7, 8", + "[14] Shenyuan Gao, Jiazhi Yang, Li Chen, Kashyap Chitta, Yihang Qiu, Andreas Geiger, Jun Zhang, and Hongyang Li." + ], + "bbox": [ + 78, + 114, + 470, + 902 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "Vista: A generalizable driving world model with high fidelity and versatile controllability. In NeurIPS, 2024. 3", + "[15] Wencheng Han, Dongqian Guo, Cheng-Zhong Xu, and Jianbing Shen. Dme-driver: Integrating human decision logic and 3d scene perception in autonomous driving. arXiv preprint arXiv:2401.03641, 2024. 1", + "[16] Yingqing He, Tianyu Yang, Yong Zhang, Ying Shan, and Qifeng Chen. Latent video diffusion models for high-fidelity long video generation. arXiv preprint arXiv:2211.13221, 2022. 3, 1", + "[17] Jonathan Ho, Ajay Jain, and Pieter Abbeel. Denoising diffusion probabilistic models. NeurIPS, 2020. 3, 4, 1", + "[18] Qianjiang Hu, Zhimin Zhang, and Wei Hu. Rangeldm: Fast realistic lidar point cloud generation. In ECCV, 2025. 3", + "[19] Yihan Hu, Jiazhi Yang, Li Chen, Keyu Li, Chonghao Sima, Xizhou Zhu, Siqi Chai, Senyao Du, Tianwei Lin, Wenhai Wang, et al. Planning-oriented autonomous driving. In CVPR, 2023. 1, 2, 7, 8", + "[20] Bo Jiang, Shaoyu Chen, Qing Xu, Bencheng Liao, Jiajie Chen, Helong Zhou, Qian Zhang, Wenyu Liu, Chang Huang, and Xinggang Wang. Vad: Vectorized scene representation for efficient autonomous driving. In ICCV, 2023. 1, 2, 7", + "[21] Bernhard Kerbl, Georgios Kopanas, Thomas Leimkuhler, and George Drettakis. 3d gaussian splatting for real-time radiance field rendering. ACM Trans. Graph., 2023. 3", + "[22] Jumin Lee, Sebin Lee, Changho Jo, Woobin Im, Juhyeong Seon, and Sung-Eui Yoon. Semcity: Semantic scene generation with triplane diffusion. In CVPR, 2024. 3, 7, 8, 2", + "[23] Quanyi Li, Zhenghao Peng, Lan Feng, Qihang Zhang, Zhenghai Xue, and Bolei Zhou. Metadrive: Composing diverse driving scenarios for generalizable reinforcement learning. T-PAMI, 2022. 1, 2, 3", + "[24] Zhiqi Li, Zhiding Yu, Shiyi Lan, Jiahan Li, Jan Kautz, Tong Lu, and Jose M Alvarez. Is ego status all you need for open-loop end-to-end autonomous driving? In CVPR, 2024. 1", + "[25] Zhiqi Li, Zhiding Yu, Shiyi Lan, Jiahan Li, Jan Kautz, Tong Lu, and Jose M Alvarez. Is ego status all you need for open-loop end-to-end autonomous driving? In CVPR, 2024. 1", + "[26] William Ljungbergh, Adam Tonderski, Joakim Johnander, Holger Caesar, Kalle Åström, Michael Felsberg, and Christoffer Petersson. Neuroncap: Photorealistic closed-loop safety testing for autonomous driving. In ECCV, 2025. 3", + "[27] Pablo Alvarez Lopez, Michael Behrisch, Laura Bieker-Walz, Jakob Erdmann, Yun-Pang Flotteröd, Robert Hilbrich, Leonhard Lücken, Johannes Rummel, Peter Wagner, and Evamarie Wießner. Microscopic traffic simulation using sumo. In IEEE ITSC, 2018. 3", + "[28] Enhui Ma, Lijun Zhou, Tao Tang, Zhan Zhang, Dong Han, Junpeng Jiang, Kun Zhan, Peng Jia, Xianpeng Lang, Haiyang Sun, et al. Unleashing generalization of end-to-end autonomous driving with controllable long video generation. arXiv preprint arXiv:2406.01349, 2024. 5, 1", + "[29] Ben Mildenhall, Pratul P Srinivasan, Matthew Tancik, Jonathan T Barron, Ravi Ramamoorthi, and Ren Ng. Nerf: Representing scenes as neural radiance fields for view synthesis. Communications of the ACM, 2021. 3, 6" + ], + "bbox": [ + 501, + 92, + 893, + 900 + ], + "page_idx": 8 + }, + { + "type": "page_number", + "text": "27539", + "bbox": [ + 478, + 944, + 519, + 957 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[30] Kazuto Nakashima and Ryo Kurazume. Lidar data synthesis with denoising diffusion probabilistic models. In ICRA, 2024. 3", + "[31] Alec Radford, Jong Wook Kim, Chris Hallacy, Aditya Ramesh, Gabriel Goh, Sandhini Agarwal, Girish Sastry, Amanda Askell, Pamela Mishkin, Jack Clark, et al. Learning transferable visual models from natural language supervision. In ICML, 2021. 4", + "[32] Colin Raffel, Noam Shazeer, Adam Roberts, Katherine Lee, Sharan Narang, Michael Matena, Yanqi Zhou, Wei Li, and Peter J. Liu. Exploring the limits of transfer learning with a unified text-to-text transformer. Journal of Machine Learning Research, 2020. 6", + "[33] Prabhu Ramachandran and Gaël Varoquaux. Mayavi: 3d visualization of scientific data. Computing in Science & Engineering, 2011. 6", + "[34] Haoxi Ran, Vitor Guizilini, and Yue Wang. Towards realistic scene generation with lidar diffusion models. In ECCV, 2024. 3", + "[35] Robin Rombach, Andreas Blattmann, Dominik Lorenz, Patrick Esser, and Björn Ommer. High-resolution image synthesis with latent diffusion models. In CVPR, 2022. 4", + "[36] Hao Shao, Yuxuan Hu, Letian Wang, Guanglu Song, Steven L Waslander, Yu Liu, and Hongsheng Li. Lmdrive: Closed-loop end-to-end driving with large language models. In CVPR, 2024. 1, 2", + "[37] Pei Sun, Henrik Kretzschmar, Xerxes Dotiwalla, Aurelien Chouard, Vijaysai Patnaik, Paul Tsui, James Guo, Yin Zhou, Yuning Chai, Benjamin Caine, et al. Scalability in perception for autonomous driving: Waymo open dataset. In CVPR, 2020. 1", + "[38] Alexander Swerdlow, Runsheng Xu, and Bolei Zhou. Streetview image generation from a bird's-eye view layout. IEEE RA-L, 2024. 3", + "[39] Shuhan Tan, Boris Ivanovic, Xinshuo Weng, Marco Pavone, and Philipp Kraehenbuehl. Language conditioned traffic generation. arXiv preprint arXiv:2307.07947, 2023. 3", + "[40] Aaron Van Den Oord, Oriol Vinyals, et al. Neural discrete representation learning. NeurIPS, 2017. 4", + "[41] Lening Wang, Wenzhao Zheng, Yilong Ren, Han Jiang, Zhiyong Cui, Haiyang Yu, and Jiwen Lu. Occsora: 4d occupancy generation models as world simulators for autonomous driving. arXiv preprint arXiv:2405.20337, 2024. 2, 3, 5, 1", + "[42] Wenhai Wang, Jiangwei Xie, ChuanYang Hu, Haoming Zou, Jianan Fan, Wenwen Tong, Yang Wen, Silei Wu, Hanming Deng, Zhiqi Li, et al. Drivemlm: Aligning multi-modal large language models with behavioral planning states for autonomous driving. arXiv preprint arXiv:2312.09245, 2023. 1", + "[43] Xiaofeng Wang, Zheng Zhu, Guan Huang, Xinze Chen, Jia-gang Zhu, and Jiwen Lu. Drivedreamer: Towards real-world-driven world models for autonomous driving. arXiv preprint arXiv:2309.09777, 2023. 1, 2, 3, 5", + "[44] Yuqi Wang, Jiawei He, Lue Fan, Hongxin Li, Yuntao Chen, and Zhaoxiang Zhang. Driving into the future: Multiview visual forecasting and planning with world model for autonomous driving. In CVPR, 2024. 6, 1, 2" + ], + "bbox": [ + 78, + 90, + 470, + 900 + ], + "page_idx": 9 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[45] Yuqing Wen, Yucheng Zhao, Yingfei Liu, Fan Jia, Yanhui Wang, Chong Luo, Chi Zhang, Tiancai Wang, Xiaoyan Sun, and Xiangyu Zhang. Panacea: Panoramic and controllable video generation for autonomous driving. In CVPR, 2024. 3, 7, 8, 1, 2", + "[46] Licheng Wenl, Daocheng Fu, Song Mao, Pinlong Cai, Min Dou, Yikang Li, and Yu Qiao. Limsim: A long-term interactive multi-scenario traffic simulator. In IEEE ITSC, 2023. 1, 3, 7", + "[47] Zirui Wu, Tianyu Liu, Liyi Luo, Zhide Zhong, Jianteng Chen, Hongmin Xiao, Chao Hou, Haozhe Lou, Yuantao Chen, Runyi Yang, et al. Mars: An instance-aware, modular and realistic simulator for autonomous driving. In CAAI International Conference on Artificial Intelligence, 2023. 3", + "[48] Zhenhua Xu, Yujia Zhang, Enze Xie, Zhen Zhao, Yong Guo, Kwan-Yee K Wong, Zhenguo Li, and Hengshuang Zhao. Drivegpt4: Interpretable end-to-end autonomous driving via large language model. IEEE RA-L, 2024. 1", + "[49] Tianyi Yan, Junbo Yin, Xianpeng Lang, Ruigang Yang, Cheng-Zhong Xu, and Jianbing Shen. Olidm: Object-aware lidar diffusion models for autonomous driving. In AAAI, 2025. 3", + "[50] Kairui Yang, Enhui Ma, Jibin Peng, Qing Guo, Di Lin, and Kaicheng Yu. Bevcontrol: Accurately controlling street-view elements with multi-perspective consistency via bev sketch layout. arXiv preprint arXiv:2308.01661, 2023. 3", + "[51] Xuemeng Yang, Licheng Wen, Yukai Ma, Jianbiao Mei, Xin Li, Tiantian Wei, Wenjie Lei, Daocheng Fu, Pinlong Cai, Min Dou, Botian Shi, Liang He, Yong Liu, and Yu Qiao. Drivearena: A closed-loop generative simulation platform for autonomous driving. arXiv preprint arXiv:2408.00415, 2024. 1, 3, 7, 8, 2, 5", + "[52] Ze Yang, Yun Chen, Jingkang Wang, Sivabalan Manivasagam, Wei-Chiu Ma, Anqi Joyce Yang, and Raquel Urtasun. Unisim: A neural closed-loop sensor simulator. In CVPR, 2023. 3", + "[53] Jimuyang Zhang, Zanming Huang, Arijit Ray, and Eshed Ohn-Bar. Feedback-guided autonomous driving. In CVPR, 2024. 1", + "[54] Linrui Zhang, Zhenghao Peng, Quanyi Li, and Bolei Zhou. Cat: Closed-loop adversarial training for safe end-to-end driving. In 7th Annual Conference on Robot Learning, 2023. 1", + "[55] Lvmin Zhang, Anyi Rao, and Maneesh Agrawala. Adding conditional control to text-to-image diffusion models. In ICCV, 2023. 4", + "[56] Yuheng Zhang, Tianjian Ouyang, Fudan Yu, Cong Ma, Lei Qiao, Wei Wu, Jian Yuan, and Yong Li. Lcsim: A large-scale controllable traffic simulator. arXiv preprint arXiv:2406.19781, 2024. 1, 3", + "[57] Guosheng Zhao, Xiaofeng Wang, Zheng Zhu, Xinze Chen, Guan Huang, Xiaoyi Bao, and Xingang Wang. Drivedreamer-2: Llm-enhanced world models for diverse driving video generation. arXiv preprint arXiv:2403.06845, 2024. 1, 2, 3, 5", + "[58] Wenzhao Zheng, Weiliang Chen, Yuanhui Huang, Borui Zhang, Yueqi Duan, and Jiwen Lu. Occworld: Learning a 3d" + ], + "bbox": [ + 501, + 90, + 892, + 902 + ], + "page_idx": 9 + }, + { + "type": "page_number", + "text": "27540", + "bbox": [ + 478, + 944, + 519, + 955 + ], + "page_idx": 9 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "occupancy world model for autonomous driving. In ECCV, 2024.3,1,2", + "[59] Wenzhao Zheng, Ruiqi Song, Xianda Guo, Chenming Zhang, and Long Chen. Genad: Generative end-to-end autonomous driving. arXiv preprint arXiv:2402.11502, 2024. 1", + "[60] Zangwei Zheng, Xiangyu Peng, Tianji Yang, Chenhui Shen, Shenggui Li, Hongxin Liu, Yukun Zhou, Tianyi Li, and Yang You. Open-sora: Democratizing efficient video production for all, 2024. 5, 6, 1", + "[61] Vlas Zyrianov, Xiyue Zhu, and Shenlong Wang. Learning to generate realistic lidar point cloud. In ECCV, 2022. 3", + "[62] Vlas Zyrianov, Henry Che, Zhijian Liu, and Shenlong Wang. Lidardm: Generative lidar simulation in a generated world. arXiv preprint arXiv:2404.02903, 2024. 3" + ], + "bbox": [ + 78, + 90, + 467, + 304 + ], + "page_idx": 10 + }, + { + "type": "page_number", + "text": "27541", + "bbox": [ + 478, + 944, + 517, + 955 + ], + "page_idx": 10 + } +] \ No newline at end of file diff --git a/2025/DrivingSphere_ Building a High-fidelity 4D World for Closed-loop Simulation/e8e4b9c2-6cbc-49b7-a622-765862a5895e_model.json b/2025/DrivingSphere_ Building a High-fidelity 4D World for Closed-loop Simulation/e8e4b9c2-6cbc-49b7-a622-765862a5895e_model.json new file mode 100644 index 0000000000000000000000000000000000000000..6296238233842ba4e3f113685e64a36a30c353d0 --- /dev/null +++ b/2025/DrivingSphere_ Building a High-fidelity 4D World for Closed-loop Simulation/e8e4b9c2-6cbc-49b7-a622-765862a5895e_model.json @@ -0,0 +1,2323 @@ +[ + [ + { + "type": "header", + "bbox": [ + 0.107, + 0.003, + 0.182, + 0.043 + ], + "angle": 0, + "content": "CVF" + }, + { + "type": "header", + "bbox": [ + 0.237, + 0.0, + 0.812, + 0.045 + ], + "angle": 0, + "content": "This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore." + }, + { + "type": "title", + "bbox": [ + 0.096, + 0.131, + 0.873, + 0.153 + ], + "angle": 0, + "content": "DrivingSphere: Building a High-fidelity 4D World for Closed-loop Simulation" + }, + { + "type": "text", + "bbox": [ + 0.233, + 0.18, + 0.735, + 0.217 + ], + "angle": 0, + "content": "Tianyi Yan\\(^{1}\\), Dongming Wu\\(^{3}\\), Wencheng Han\\(^{1}\\), Junpeng Jiang\\(^{2}\\), Xia Zhou\\(^{2}\\), Kun Zhan\\(^{2}\\), Cheng-zhong Xu\\(^{1}\\), Jianbing Shen\\(^{1*}\\)" + }, + { + "type": "text", + "bbox": [ + 0.144, + 0.217, + 0.827, + 0.253 + ], + "angle": 0, + "content": "\\(^{1}\\)SKL-IOTSC, Computer and Information Science, University of Macau, \\(^{2}\\)Li Auto Inc., \\(^{3}\\)School of Computer Science, Beijing Institute of Technology" + }, + { + "type": "title", + "bbox": [ + 0.235, + 0.287, + 0.314, + 0.303 + ], + "angle": 0, + "content": "Abstract" + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.32, + 0.473, + 0.698 + ], + "angle": 0, + "content": "Autonomous driving evaluation requires simulation environments that closely replicate actual road conditions, including real-world sensory data and responsive feedback loops. However, many existing simulations need to predict waypoints along fixed routes on public datasets or synthetic photorealistic data, i.e., open-loop simulation usually lacks the ability to assess dynamic decision-making. While the recent efforts of closed-loop simulation offer feedback-driven environments, they cannot process visual sensor inputs or produce outputs that differ from real-world data. To address these challenges, we propose DrivingSphere, a realistic and closed-loop simulation framework. Its core idea is to build 4D world representation and generate real-life and controllable driving scenarios. In specific, our framework includes a Dynamic Environment Composition module that constructs a detailed 4D driving world with a format of occupancy equipping with static backgrounds and dynamic objects, and a Visual Scene Synthesis module that transforms this data into high-fidelity, multi-view video outputs, ensuring spatial and temporal consistency. By providing a dynamic and realistic simulation environment, DrivingSphere enables comprehensive testing and validation of autonomous driving algorithms, ultimately advancing the development of more reliable autonomous cars. The benchmark will be publicly released." + }, + { + "type": "title", + "bbox": [ + 0.078, + 0.729, + 0.21, + 0.743 + ], + "angle": 0, + "content": "1. Introduction" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.754, + 0.47, + 0.816 + ], + "angle": 0, + "content": "End-to-end autonomous driving algorithms [5-7, 15, 19, 20, 25, 36, 42, 48, 53, 54, 59] have made significant progress in recent years, making the accurate evaluation of these models a very urgent task. For safe and responsible evaluation," + }, + { + "type": "text", + "bbox": [ + 0.5, + 0.288, + 0.892, + 0.364 + ], + "angle": 0, + "content": "it is essential to have a precise simulation environment that accurately reflects real-world driving conditions. This requirement usually includes two aspects: firstly, the generation of high-fidelity sensory data, and secondly, the implementation of closed-loop feedback mechanisms." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.367, + 0.893, + 0.579 + ], + "angle": 0, + "content": "The most common evaluation methods involve the forecasting of waypoints along fixed routes within collected public datasets [3, 37], that is, open-loop simulation approaches. While these benchmarks offer realistic driving data, they have relatively fixed distributions and lack diversity [24], limiting their ability to assess the generalization capabilities of autonomous driving algorithms. Even though several recent works [13, 43, 57] excel at producing diverse, photo-realistic data, continue to rely on the prediction of waypoints that are pre-established. In summary, despite high-fidelity sensory data, these open-loop evaluation solutions cannot provide the dynamic feedback necessary for assessing how autonomous systems respond to dynamic changes and decision-making." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.581, + 0.895, + 0.748 + ], + "angle": 0, + "content": "In contrast, closed-loop simulations [10, 23, 46, 56] offer feedback-driven systems where an agent's actions influence and are influenced by other agents and the environment. For example, traffic flow-based simulation methods [12, 46, 56] successfully enable multi-agent simulations. However, they lack the ability to process visual sensor inputs, limiting their interplay with vision-based end-to-end models. Game engine-based simulators [10, 23] create scalable and physically realistic environments, but their outputs often diverge from real-world sensor data, restricting their utility in validating algorithms trained on actual inputs." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.75, + 0.895, + 0.902 + ], + "angle": 0, + "content": "To address these challenges, we propose DrivingSphere, a novel generative closed-loop simulation framework that leverages geometric-wise prior information to generate realistic and controllable driving scenarios. In comparison with existing simulations, DrivingSphere has three distinguished features: ① Rich simulation granularity. Unlike past methods that model only roads and cars [13, 51], our approach allows for including previously unmodeled elements, such as buildings, vegetation, and other environmental structures. Although these non-traffic elements are not" + }, + { + "type": "page_footnote", + "bbox": [ + 0.076, + 0.828, + 0.47, + 0.901 + ], + "angle": 0, + "content": "*Corresponding author: Jianbing Shen. This work was supported in part by the Science and Technology Development Fund of Macau SAR (FDCT) under grants 0102/2023/RIA2 and 0154/2022/A3 and 001/2024/SKL, the Jiangyin Hi-tech Industrial Development Zone under the Taihu Innovation Scheme (EF2025-00003-SKL-IOTSC), the University of Macau SRG2022-00023-IOTSC grant." + }, + { + "type": "page_number", + "bbox": [ + 0.479, + 0.945, + 0.518, + 0.957 + ], + "angle": 0, + "content": "27531" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.084, + 0.09, + 0.345, + 0.256 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.347, + 0.091, + 0.581, + 0.256 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.587, + 0.092, + 0.885, + 0.256 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.076, + 0.258, + 0.895, + 0.343 + ], + "angle": 0, + "content": "Figure 1. Comparison of frameworks for evaluating end-to-end Autonomous Driving (AD) algorithms. (a) Open-loop evaluation uses waypoint predictions along fixed routes in pre-collected datasets [3]. Generative models [13, 43, 57] create diverse, realistic data but lack dynamic feedback to assess AD responses to dynamic changes. (b) Simulated closed-loop evaluation [10, 23] offers feedback-driven, scalable environments where agent actions impact simulation dynamics; however, sensory outputs often differ from real-world data, limiting effectiveness for algorithms trained on real data. (c) Our generative closed-loop simulation framework, *DrivingSphere*, addresses these limitations by delivering realistic visual inputs and continuous, responsive feedback between the AD agent and environment." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.358, + 0.473, + 0.584 + ], + "angle": 0, + "content": "directly involved in traffic flow, their presence can significantly impact the input to driving models, thereby influencing decision-making processes in complex driving scenarios. 2 Physical and spatial realism. Due to our model explicitly representing the scene and traffic actors in 4D space, it is capable of precisely depicting physical interactions and occlusion relationships across different traffic elements. This ensures that each viewpoint and position naturally adheres to physical principles like depth and occlusion, thus achieving structured coordination of global road layout, traffic participants, and their behaviors. 3 High visual consistency and fidelity. Our model pays more effort to associate each traffic actor's appearance and unique ID within the scene, providing stable and high-fidelity temporal and spatial coherence across frames and views." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.592, + 0.47, + 0.759 + ], + "angle": 0, + "content": "The proposed DrivingSphere contains two key modules: Dynamic Environment Composition and Visual Scene Synthesis. First, the Dynamic Environment Composition module aims to produce a coarse-grained 4D world with an occupancy grid format, which integrates large static backgrounds with dynamic traffic actors. For this, we introduce a BEV-conditioned occupancy diffusion model, named OccDreamer, to generate static backgrounds. It provides an infinite city-scale scene via progressive region extension. Besides, we build an actor bank to manage key actors or objects, and update actors' spatial-temporal position." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.765, + 0.471, + 0.903 + ], + "angle": 0, + "content": "Once the scene is generated, the second module, Visual Scene Synthesis, converts it into high-fidelity, multi-view video outputs for AD system perception. To fully capture the complexity of the 4D environment, we employ a dual-path condition encoding strategy. Specifically, the global branch, using a pretrained 4D encoder [41], extracts geometric information and spatial-temporal relationships directly from the occupancy data, while the local projection branch produces view-specific semantic maps that align on" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.358, + 0.895, + 0.525 + ], + "angle": 0, + "content": "a pixel-to-pixel basis, accurately capturing occlusions and depth variations. To further ensure visual consistency, we design an ID-aware actor encoding mechanism that binds the appearance and identity information of traffic actors to their scene locations, which preserves spatial correlation across views and temporal consistency across frames. With the custom-designed dual-path condition encoding and ID-aware actor encoding, the Visual Scene Synthesis generates spatially and temporally coherent video sequences. This results in a highly realistic simulation environment that is ideal for testing and validating AD systems." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.531, + 0.895, + 0.773 + ], + "angle": 0, + "content": "By responsive interplay between AD algorithms and the simulated environment, DrivingSphere supports a closed-loop feedback loop, where actions from AD systems [6, 19, 20, 36] influence surrounding actors and scenes. This simulation platform closely mirrors real-world conditions, providing a reliable foundation for comprehensive AD validation. Our extensive experiments show that DrivingSphere achieves superior performance in visual fidelity and temporal consistency, significantly reducing the domain gap between simulated and real-world environments. By excelling in both open-loop and closed-loop evaluations, DrivingSphere proves to be a reliable platform for dynamic agent-environment interplay in rigorous AD testing. Ultimately, DrivingSphere will drive progress toward the development of safer and more reliable autonomous vehicles. Overall, our contributions are four-fold:" + }, + { + "type": "text", + "bbox": [ + 0.5, + 0.781, + 0.892, + 0.841 + ], + "angle": 0, + "content": "- We introduce DrivingSphere, a novel geometry-aware closed-loop simulation framework that captures 2D visual and 3D geometric properties while seamlessly integrating with vision-based end-to-end driving agents." + }, + { + "type": "text", + "bbox": [ + 0.5, + 0.841, + 0.894, + 0.9 + ], + "angle": 0, + "content": "- We propose the Dynamic Environment Composition that provides a comprehensive 4D driving world by modeling both static and dynamic elements, enabling intricate simulations across diverse environments." + }, + { + "type": "list", + "bbox": [ + 0.5, + 0.781, + 0.894, + 0.9 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.479, + 0.945, + 0.521, + 0.958 + ], + "angle": 0, + "content": "27532" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.077, + 0.092, + 0.468, + 0.15 + ], + "angle": 0, + "content": "- We design the Visual Scene Synthesis that enables long-term consistency and high fidelity, addressing challenges in generating continuous and multi-view video sequences for AD simulations." + }, + { + "type": "text", + "bbox": [ + 0.078, + 0.152, + 0.468, + 0.213 + ], + "angle": 0, + "content": "- We introduce the Agent Coordination module for scalable multi-agent systems, enabling the integration of multiple vehicle agents and establishing a crucial feedback loop for validating autonomous driving algorithms." + }, + { + "type": "list", + "bbox": [ + 0.077, + 0.092, + 0.468, + 0.213 + ], + "angle": 0, + "content": null + }, + { + "type": "title", + "bbox": [ + 0.078, + 0.226, + 0.218, + 0.241 + ], + "angle": 0, + "content": "2. Related Work" + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.251, + 0.47, + 0.659 + ], + "angle": 0, + "content": "Generative Models in Driving. Denoising Diffusion Probabilistic Models (DDPMs) [17] have transformed image generation by iteratively refining Gaussian noise into high-quality images. Based on this technique, generative models in autonomous driving have greatly advanced. For example, BEVGen [38] and BEVControl [50] use diffusion models to produce controllable street-view images from BEV layouts. More recently, diffusion-based video generation models [2, 16] enable methods like [13, 14, 43-45, 57] to generate multi-view driving videos, effectively supporting perception tasks. Inspired by these 2D methods, advances in 3D scene generation have also emerged. Integrating diffusion processes into LiDAR generation improves the realism of LiDAR scenes [18, 30, 34, 49, 61, 62]. Occupancy-based models have further progressed, with occupancy grids now central to 3D scene generation. SemCity [22] enhances outdoor 3D occupancy generation, while models like OccWorld [58] and OccSora [41] use occupancy data for vehicle motion prediction and trajectory-conditioned 4D occupancy generation, respectively. Unlike these, our OccDreamer combines text and BEV conditions to generate 3D occupancy data, dynamically enhanced with foreground motion and ego pose to form a complete 4D driving world. By conditioning on occupancy data, our VideoDreamer allows precise control over non-traffic elements and ensures spatial and temporal coherence, capturing essential physical relationships such as depth and occlusion." + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.66, + 0.468, + 0.901 + ], + "angle": 0, + "content": "Autonomous Driving Simulations. Autonomous driving simulation methods are developed from basic behavioral models to complex, interactive environments. Flow-based simulations [9, 11, 12, 39, 46, 56] enable multi-agent interactions through long-term traffic planning. However, these methods typically operate in 2D space and lack real-life visual outputs. Platforms like CARLA [10], SUMO [27], and MetaDrive [23] address these limitations by modeling traffic dynamics and supporting multi-agent interactions with visual contexts, but they have visual fidelity differences from real-world data, limiting their direct use with algorithms trained on such data. While reconstruction-based methods [21, 26, 29] and frameworks such as UniSim [52], and MARS [47] capture high-fidelity scenes, they are limited by the scope of the data used, restricting their ability to simulate varied and long-tail scenarios. DriveArena [51]" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.092, + 0.892, + 0.228 + ], + "angle": 0, + "content": "recently advanced generative closed-loop simulation by using 2D traffic sketches to control visual outputs, integrating well with vision-based driving agents. However, reliance on 2D inputs constrains its modeling granularity and visual realism. Our approach overcomes these challenges by leveraging 4D occupancy data, generating visually coherent and geometrically accurate scenes that rigorously test driving agents in closed-loop simulations, equipping them to handle complex real-world conditions effectively." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.24, + 0.634, + 0.258 + ], + "angle": 0, + "content": "3. Methodology" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.265, + 0.892, + 0.568 + ], + "angle": 0, + "content": "DrivingSphere acts as a generative closed-loop simulation framework that combines occupancy-based 4D world modeling with advanced video rendering techniques, providing high-fidelity visual outputs that enhance both simulation realism and agent-environment interplay in autonomous driving scenarios. Specifically, DrivingSphere begins with Dynamic Environment Composition (Sec. 3.1), which generates static background from a map sketch, selects traffic participants from the actor bank, and updates the actor location, to formulate a 4D driving environment with a format of occupancy. Next, Visual Scene Synthesis (Sec. 3.2) conditions the occupancy data surrounding the ego vehicle, accurately capturing occlusion relationships and fine-grained semantic information to generate high-fidelity multi-view videos. Finally, the closed-loop feedback mechanism (Sec. 3.3) enables dynamic, responsive adjustments, where autonomous agents continuously receive updated visual data and generate control signals that modify the simulation environment, providing a comprehensive platform for algorithm testing and refinement. The overall framework is shown in Fig. 2." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.576, + 0.813, + 0.592 + ], + "angle": 0, + "content": "3.1. Dynamic Environment Composition" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.599, + 0.892, + 0.795 + ], + "angle": 0, + "content": "Previous driving simulation methods have often overlooked static and varied elements like buildings, obstacles, and vegetation [10, 51]. Although these elements are not direct traffic participants, they are part of the perception input of AD systems, influencing final driving decision-making. For instance, static objects like buildings may block sensors' line-of-sight, causing occlusions of other vehicles or pedestrians. Obstacles or vegetation can introduce sensor artifacts or false positives for AD systems. Both situations may result in incorrect decisions like unnecessary braking or lane changes. To address this issue, we propose a Dynamic Environment Composition to construct a comprehensive 4D Driving World equipping complex digital assets." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.796, + 0.892, + 0.901 + ], + "angle": 0, + "content": "4D Driving World Representation. Our 4D driving world representation includes three critical digital assets: static background scenes, dynamic foreground actors, and spatial position of these actors. Formally, we define this representation as \\( \\pmb{W} = \\{\\pmb{S}_{\\mathrm{city}},(\\pmb{A}_1,\\pmb{P}_1),\\dots ,(\\pmb{A}_N,\\pmb{P}_N)\\}_{n = 1}^N \\). Here, \\( S_{\\mathrm{city}} \\) is a composite of multiple regional static scene \\( \\{S_1,\\ldots ,S_K\\} \\). Each \\( S_{k}\\in \\mathbb{R}^{H\\times W\\times D} \\) is the \\( k^{th} \\) static back" + }, + { + "type": "page_number", + "bbox": [ + 0.479, + 0.945, + 0.518, + 0.957 + ], + "angle": 0, + "content": "27533" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.087, + 0.096, + 0.891, + 0.331 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.076, + 0.333, + 0.894, + 0.402 + ], + "angle": 0, + "content": "Figure 2. Overview of DrivingSphere framework. (a) The Dynamic Environment Composition module builds a 4D driving world, simulating real driving scenarios with backgrounds generated by OccDreamer, dynamic actors from an actor bank, and trajectories guided by a transportation simulator. (b) The Visual Scene Synthesis produces high-fidelity, photo-realistic video frames conditioned on global semantics, view-specific details, and scene prompts, supporting close-loop evaluation. Control signals enable adaptive feedback for driving agents, facilitating continuous testing and evaluation of driving algorithms in the simulated environment." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.407, + 0.471, + 0.562 + ], + "angle": 0, + "content": "ground scene that captures the spatial layout and static elements within the region. \\(A_{n}\\in \\mathbb{R}^{4}\\) is the \\(n^{th}\\) actor of \\(N\\) actors, such as vehicles and pedestrians, defined by a 3D coordinate and a semantic label. It is worth noticing that both the static scene \\(\\pmb{S}_{k}\\) and dynamic actor \\(A_{n}\\) are represented as occupancy grids. \\(P_{n} = \\left\\{P_{n}^{0},\\dots ,P_{n}^{T}\\right\\}_{t = 0}^{T}\\) describe actor position across \\(T\\) frames, where \\(P_{n}^{t}\\in \\mathbb{R}^{4}\\) consists of a 3D relative position of the \\(n\\)-th actor at time \\(t\\) with respect to the ego vehicle and a yaw rotation \\(\\theta\\). In the following, we will detail how to generate three assets." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.562, + 0.47, + 0.728 + ], + "angle": 0, + "content": "Static Scene Generation. A straightforward approach generating static 3D scenes \\( S_{\\mathrm{city}} \\) is directly using ground truth occupancy data from existing datasets, like the Boston area in nuScenes [3]. However, this method is limited to specific regions captured during data collection, restricting its applicability to other urban areas. To address this limitation, we propose an occupancy diffusion model, OccDreamer, which is conditioned on BEV maps and textual descriptions, capable of generating static scenes for any desired urban region. The framework of OccDreamer, which integrates the following components, is illustrated in Fig. 3." + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.728, + 0.47, + 0.822 + ], + "angle": 0, + "content": "Firstly, to efficiently train the diffusion model while addressing the computational complexity of processing 3D data, we adopt VQVAE [17, 35, 40] as an occupancy tokenizer, mapping \\( S_{k} \\) to a latent feature \\( Z^{S} \\): \\( Z^{S} = \\mathcal{F}_{\\mathrm{VAE,Enc}}^{occ}(S_{k}) \\). The reconstructed scene is defined as \\( S_{k}^{\\prime} = \\mathcal{F}_{\\mathrm{VAE,Dec}}^{occ}(Z^{S_{k}}) \\). \\( \\mathcal{F}_{\\mathrm{VAE}}^{occ} \\) is trained with a combined loss \\( \\mathcal{L}_{occ} \\):" + }, + { + "type": "equation", + "bbox": [ + 0.093, + 0.832, + 0.469, + 0.849 + ], + "angle": 0, + "content": "\\[\n\\mathcal {L} _ {o c c} = \\mathcal {L} _ {C E} \\left(\\boldsymbol {S} _ {k}, \\boldsymbol {S} _ {k} ^ {\\prime}\\right) + \\alpha \\mathcal {L} _ {L o v} \\left(\\boldsymbol {S} _ {k}, \\boldsymbol {S} _ {k} ^ {\\prime}\\right) + \\beta \\mathcal {L} _ {e m b}, \\tag {1}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.855, + 0.465, + 0.871 + ], + "angle": 0, + "content": "where \\(\\mathcal{L}_{CE}\\) is the CE loss and \\(\\mathcal{L}_{Lov}\\) is the Lovász loss [1]." + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.871, + 0.469, + 0.902 + ], + "angle": 0, + "content": "Secondly, considering varying road structures and complex regional topographies, we propose a controllable re" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.407, + 0.892, + 0.604 + ], + "angle": 0, + "content": "gion occupancy generation module. It accepts a regional BEV map capturing the road structure and a language prompt describing abstract features (like suburban areas with rich vegetation or commercial districts with buildings lining both sides) as input, and outputs a region occupancy \\( S_{k} \\). Following the principles of diffusion models [17, 55], a CLIP encoder [31] converts the text prompt into an embedding \\( F_{\\mathrm{region}} \\). Then, \\( F_{\\mathrm{region}} \\) is injected into the denoiser \\( \\epsilon_{\\theta}^{s} \\) via cross-attention. Simultaneously, the BEV map is processed by a pre-trained image VAE [17] to extract the corresponding road embedding \\( F_{M} \\), serving as input for the ControlNet branch \\( \\epsilon_{\\phi}^{s} \\). This enables precise control over the diffusion learning process in the latent space:" + }, + { + "type": "equation", + "bbox": [ + 0.5, + 0.609, + 0.908, + 0.645 + ], + "angle": 0, + "content": "\\[\n\\mathcal {L} _ {\\text {o c c}} = \\mathbb {E} _ {\\boldsymbol {F} _ {M}, \\boldsymbol {F} _ {\\text {r e g i o n}}, \\boldsymbol {Z}} s _ {k, \\epsilon^ {s}, \\tau} \\left[ \\left\\| \\epsilon^ {s} - \\epsilon_ {\\theta} ^ {s} \\left(\\boldsymbol {Z} _ {\\tau} ^ {\\boldsymbol {S} _ {k}}, \\boldsymbol {F} _ {\\text {r e g i o n}}, \\tau , \\epsilon_ {\\phi} ^ {s} (\\boldsymbol {F} _ {M})\\right) \\right\\| ^ {2} \\right], \\tag {2}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.645, + 0.892, + 0.692 + ], + "angle": 0, + "content": "where \\( Z_{\\tau}^{S_k} \\) is the noisy version of \\( Z^{S_k} \\) by adding \\( \\tau \\) steps noise \\( \\epsilon^s \\). After training, we use OccDreamer to create \\( Z_{k}^{S} \\) and decode it with \\( \\mathcal{F}_{\\mathrm{VAE,Decode}}^{\\mathrm{occ}} \\) to obtain \\( S_{k} \\)." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.691, + 0.892, + 0.86 + ], + "angle": 0, + "content": "Thirdly, for spatial consistency across the entire 3D scene, we propose a Scene Expansion Mechanism for building a coherent and city-level static scene \\( S_{\\mathrm{city}} \\). Its core idea is to expand an initial region \\( S_{k} \\) to a neighboring region \\( S_{k+1} \\) by using the overlap between these regions as a conditioning constraint. Formally, let \\( O \\) represent a binary mask that defines the overlapping area between two adjacent regions \\( S_{k} \\) and \\( S_{k+1} \\). This overlap serves as the conditional context for generating the next region, enforcing spatial coherence at the boundary. To generate \\( S_{k+1} \\), we first create a partially masked scene \\( S_{k}^{\\mathrm{partial}} \\) by applying \\( O \\) to \\( S_{k} \\):" + }, + { + "type": "equation", + "bbox": [ + 0.636, + 0.865, + 0.891, + 0.883 + ], + "angle": 0, + "content": "\\[\n\\boldsymbol {S} _ {k} ^ {\\text {p a r t i a l}} = \\boldsymbol {S} _ {k} \\odot \\boldsymbol {O}, \\tag {3}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.5, + 0.887, + 0.892, + 0.902 + ], + "angle": 0, + "content": "where \\(\\odot\\) denotes element-wise multiplication, extracting" + }, + { + "type": "page_number", + "bbox": [ + 0.479, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "27534" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.084, + 0.089, + 0.468, + 0.334 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.076, + 0.334, + 0.47, + 0.375 + ], + "angle": 0, + "content": "Figure 3. The framework of OccDreamer, which includes Occupancy Tokenizer, Region Occupancy Generation, and Scene Extensio to generate a city-level background for 4D driving world." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.391, + 0.47, + 0.468 + ], + "angle": 0, + "content": "the overlapping section of \\( S_{k} \\). We then employ a diffusion process to generate the neighboring region \\( S_{k + 1} \\), conditioning on both the BEV map \\( M_{k + 1} \\) (capturing the road structure of the target area) and the partially masked scene \\( S_{k}^{\\mathrm{partial}} \\). This diffusion process is represented as:" + }, + { + "type": "equation", + "bbox": [ + 0.146, + 0.477, + 0.47, + 0.497 + ], + "angle": 0, + "content": "\\[\n\\boldsymbol {Z} ^ {\\boldsymbol {S} _ {k + 1}} = \\mathcal {F} _ {\\text {o u t p a i n t i n g}} \\left(\\epsilon^ {s}, \\boldsymbol {M} _ {k + 1}, \\boldsymbol {S} _ {k} ^ {\\text {p a r t i a l}}\\right), \\tag {4}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.507, + 0.47, + 0.598 + ], + "angle": 0, + "content": "where \\(\\epsilon^s\\) is the input noise, and \\(\\mathcal{F}_{\\mathrm{outpainting}}\\) denotes the diffusion model conditioned on the masked scene and BEV map. After obtaining the latent representation \\(Z^{S_{k + 1}}\\), we decode it using the occupancy VAE decoder, resulting in the expanded region \\(S_{k + 1}\\). The two regions are then combined into a larger scene \\(S_{\\mathrm{merged}}\\) through a merging operation:" + }, + { + "type": "equation", + "bbox": [ + 0.147, + 0.608, + 0.469, + 0.625 + ], + "angle": 0, + "content": "\\[\n\\boldsymbol {S} _ {\\text {m e r g e d}} = \\operatorname {M e r g e} \\left(\\boldsymbol {S} _ {k} \\odot (1 - \\boldsymbol {O}), \\boldsymbol {S} _ {k + 1}\\right), \\tag {5}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.629, + 0.469, + 0.689 + ], + "angle": 0, + "content": "where Merge ensures spatial continuity and logical consistency between adjacent regions. The spatially continuous and logically consistent city scene \\( S_{\\mathrm{city}} \\) serves as an excellent digital background for the 4D driving world." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.69, + 0.471, + 0.901 + ], + "angle": 0, + "content": "Dynamic Actor Selection. To complement static scenes, we populate the 4D driving world with dynamic actors, creating realistic traffic flows. An actor bank \\(\\mathcal{B}\\) is constructed, where each actor \\(\\mathbf{A}_n \\in \\mathcal{B}\\) consists of instance-level voxel data segmented from real or synthetic occupancy scenes, with unique geometric, semantic, and behavioral attributes, as well as a descriptive caption \\(L_{\\mathbf{A}_n}\\) and an ID \\(I_{\\mathbf{A}_n}\\) to ensure consistency across frames. Actors are selected based on semantic similarity to user-provided descriptions via CLIP, or randomly sampled from relevant categories if unspecified, ensuring contextual diversity. This flexible selection process allows for dynamically relevant and diverse actor integration within the 4D driving world, supporting realistic and adaptable traffic simulation." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.092, + 0.892, + 0.198 + ], + "angle": 0, + "content": "Spatial Position of Actors. The generation of actor positions \\( \\pmb{P} \\) is driven by the control signals output from autonomous agents (Sec. 3.3). Each actor is assigned an control signal set \\( C_n = \\{c_n^0, \\dots, c_n^T\\} \\), where \\( c_n^t \\) represents the received control signal of actor \\( n \\) at time \\( t \\). These signals iteratively update the dynamic state of the world. For \\( n \\)-th actor, we have the position:" + }, + { + "type": "equation", + "bbox": [ + 0.616, + 0.203, + 0.892, + 0.22 + ], + "angle": 0, + "content": "\\[\n\\boldsymbol {P} _ {n} ^ {t + 1} = \\boldsymbol {P} _ {n} ^ {t} + \\triangle \\boldsymbol {P} _ {n} ^ {t} \\left(\\boldsymbol {c} _ {n} ^ {t}\\right), \\tag {6}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.222, + 0.892, + 0.253 + ], + "angle": 0, + "content": "where \\(\\triangle P_n^t (c_N^t)\\) is the positional change resulting from the control signal \\(c_{N}^{t}\\) at time \\(t\\)" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.253, + 0.892, + 0.314 + ], + "angle": 0, + "content": "4D World Composition. With the static background, dynamic foreground actors and their positions calculated, we integrate them into the comprehensive 4D driving world. The world state at any time \\( t \\) is represented by:" + }, + { + "type": "equation", + "bbox": [ + 0.61, + 0.32, + 0.892, + 0.34 + ], + "angle": 0, + "content": "\\[\n\\boldsymbol {W} ^ {t} = \\pi (\\boldsymbol {S}, \\left(\\boldsymbol {A} _ {n}, \\boldsymbol {c} _ {n} ^ {t}\\right) _ {n = 1} ^ {N}), \\tag {7}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.342, + 0.892, + 0.432 + ], + "angle": 0, + "content": "where \\(\\pi\\) is the composition operator that aligns each actor based on their current pose \\(P_{n}^{t}\\), represents the set of poses for all actors, forming the complete 4D driving world at time \\(t\\). By iterating this process over time, we construct the dynamic 4D world that accounts for actor interplay, occlusions, and the evolving scene geometry." + }, + { + "type": "title", + "bbox": [ + 0.499, + 0.44, + 0.71, + 0.456 + ], + "angle": 0, + "content": "3.2. Visual Scene Synthesis" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.463, + 0.892, + 0.704 + ], + "angle": 0, + "content": "Previous generative models [13, 28, 43, 57] tend to employ 2D visual conditions, which fail to accurately capture the geometric and semantic complexities inherent in real-world driving scenes. In response, our Visual Scene Synthesis employs VideoDreamer, a custom-designed video diffusion model, to transform the occupancy-driven 4D driving world constructed in the previous section into high-fidelity visual results. The overall framework is shown in Fig. 4. Specifically, we introduce a Dual-path Condition Encoding strategy that focuses on encoding occupancy data as its primary condition. We further enhance the consistency of appearances across views and frames by developing an ID-aware Actor Encoding method. Finally, we integrate the Spatial-Temporal Diffusion Transformer (ST-DiT) from OpenSora [60] as the foundational technology to ensure visual consistency and generate artifact-free frames." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.705, + 0.892, + 0.87 + ], + "angle": 0, + "content": "Dual-path Condition Encoding. We design a dual-path condition encoding strategy aimed at effectively capturing occupancy data. Given a set of driving world data \\(\\mathbf{W} = \\{\\mathbf{W}^0, \\mathbf{W}^1, \\dots, \\mathbf{W}^T\\}\\) spanning frames from 0 up to \\(T\\), we first encode them into a global feature \\(\\mathbf{F}_{\\mathrm{global}} = \\mathcal{F}_{\\mathrm{VAE}}^{\\mathrm{4Docc}}(\\mathbf{W})\\), capturing comprehensive geometric information and the spatial-temporal relationships within the scene. Here, \\(\\mathcal{F}_{\\mathrm{VAE}}^{\\mathrm{4Docc}}\\) is the well-trained 4D encoder from occupancy VAE [41], responsible for capturing global geometry of the scene. This ensures that the overall structure and spatial layout of the scene are accurately represented." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.871, + 0.892, + 0.902 + ], + "angle": 0, + "content": "Meanwhile, for \\(t^{th}\\) frame (where \\(t = 0,\\dots,T\\)), 3D occupancy data \\(\\mathbf{W}^t\\) is projected into 2D semantic maps \\(\\mathcal{M}_v^t\\) for" + }, + { + "type": "page_number", + "bbox": [ + 0.479, + 0.945, + 0.518, + 0.957 + ], + "angle": 0, + "content": "27535" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.084, + 0.091, + 0.468, + 0.313 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.076, + 0.314, + 0.47, + 0.388 + ], + "angle": 0, + "content": "Figure 4. Overview of the VideoDreamer. The model conditions on 4D driving world and enriched actor embeddings (e.g., actor ID, position and caption). This ensures high visual fidelity and geometric consistency in the generated driving simulations. Each view \\( v \\) by camera extrinsics \\( \\mathbf{T}^t \\) and intrinsics \\( \\mathbf{K}_v \\):" + }, + { + "type": "equation", + "bbox": [ + 0.195, + 0.395, + 0.47, + 0.412 + ], + "angle": 0, + "content": "\\[\n\\mathcal {M} _ {v} ^ {t} = \\mathcal {Q} \\left(\\boldsymbol {W} ^ {t}, \\mathbf {K} _ {v}, \\mathbf {T} ^ {t}\\right), \\tag {8}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.416, + 0.47, + 0.537 + ], + "angle": 0, + "content": "where \\(\\mathcal{Q}\\) is the render and projection function [33]. This process yields a sequence of semantic maps, \\(\\mathcal{M} \\in \\mathbb{R}^{T \\times 6 \\times H \\times W}\\), with precise pixel-to-pixel alignment across views. Such alignment is essential for accurately capturing view-specific occlusions and depth variations. \\(\\mathcal{M}\\) is then encoded into view-specific features \\(F_{\\mathrm{view}} = \\mathcal{F}_{\\mathrm{VAE}}^{\\mathrm{Image}}(\\mathcal{M})\\). \\(F_{\\mathrm{view}}\\) and \\(F_{\\mathrm{global}}\\) are then integrated through cross-attention to obtain the unified representation \\(F_{\\mathrm{occ}}\\)." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.537, + 0.47, + 0.583 + ], + "angle": 0, + "content": "ID-aware Actor Encoding. To ensure coherence of each actor's appearance and identity information across the scene, we extract a fused sequence embedding \\( F_{\\mathrm{fuse}} \\):" + }, + { + "type": "equation", + "bbox": [ + 0.151, + 0.592, + 0.469, + 0.607 + ], + "angle": 0, + "content": "\\[\n\\boldsymbol {F} _ {\\text {f u s e}} = \\operatorname {C O N C A T} \\left[ \\boldsymbol {F} _ {\\boldsymbol {W}}, \\boldsymbol {F} _ {\\boldsymbol {A} _ {1}}, \\dots , \\boldsymbol {F} _ {\\boldsymbol {A} _ {N}} \\right], \\tag {9}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.612, + 0.469, + 0.643 + ], + "angle": 0, + "content": "where \\( F_W \\) is the embedding of the scene caption \\( \\mathcal{L}_{\\mathrm{scene}} \\), derived from T5 [32]. Each actor embedding \\( F_{A_n} \\) is:" + }, + { + "type": "equation", + "bbox": [ + 0.096, + 0.65, + 0.47, + 0.677 + ], + "angle": 0, + "content": "\\[\n\\boldsymbol {F} _ {\\boldsymbol {A} _ {n}} = \\operatorname {C O N C A T} \\left[ \\mathcal {F} _ {\\text {F o u r i e r}} \\left(\\boldsymbol {P} _ {n}\\right), \\mathcal {F} _ {\\text {F o u r i e r}} \\left(\\boldsymbol {I} _ {n}\\right), \\mathcal {F} _ {T 5} \\left(\\boldsymbol {L} _ {\\boldsymbol {A} _ {i}}\\right) \\right]. \\tag {10}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.675, + 0.469, + 0.764 + ], + "angle": 0, + "content": "Here, \\(P_{n}\\) represents the position of the \\(n\\)-th actor and \\(I_{n}\\) is its unique ID. Both are encoded using Fourier encoding [29]. The actor caption \\(L_{A_i}\\) is encoded using the T5 [32]. Overall, by integrating geometry positions, textual descriptions, and identity information, we ensure coherent behavior and long-term visual consistency across frames." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.765, + 0.469, + 0.839 + ], + "angle": 0, + "content": "ControlNet-DiT. To enhance visual fidelity and temporal consistency in generated videos, we integrate the ST-DiT [60] as our denoiser \\(\\epsilon_{\\theta}^{v}\\), utilizing ST-DiT blocks stacked with View-aware Spatial Self-Attention (VSSA), Temporal Self-Attention, Cross-Attention, and FFN." + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.84, + 0.47, + 0.901 + ], + "angle": 0, + "content": "In the video generation process, the input video sequence \\( \\mathbf{V} \\in \\mathbb{R}^{v \\times T \\times H \\times W} \\) is first encoded into latent features \\( Z^{\\mathbf{V}} \\in \\mathbb{R}^{v \\times T \\times H' \\times W' \\times C} \\). To ensure multi-view consistency while managing computational complexity, we mod-" + }, + { + "type": "image", + "bbox": [ + 0.499, + 0.09, + 0.697, + 0.183 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.703, + 0.091, + 0.892, + 0.182 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.507, + 0.185, + 0.882, + 0.254 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.498, + 0.256, + 0.892, + 0.298 + ], + "angle": 0, + "content": "Figure 5. Qualitative results of generated 3D scene. OccDreamer uses text prompt and bev map as guidance to generate 3D scenes with controllable regional content and road structure." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.302, + 0.893, + 0.514 + ], + "angle": 0, + "content": "if the original Spatial self-attention [60] as VSSA. The latent feature \\( Z^{\\mathbf{V}} \\) is reshaped from \\( v\\times T\\times H^{\\prime}\\times W^{\\prime}\\times C \\) to \\( T\\times (vH^{\\prime}W^{\\prime})\\times C \\), treating the combined view, height, and width dimensions \\( (vH^{\\prime}W^{\\prime}) \\) as the sequence length. This view-flatted approach reduces the parameter overhead associated with cross-view attention [13, 44] while maintaining spatial coherence across different views. For external conditions, we apply a cross-attention layer that injects the fused sequence embedding \\( F_{\\mathrm{fuse}} \\) (derived from scene context and actor identities) as: \\( Z^{V} = \\mathrm{CrossAttention}(Z^{V},F_{\\mathrm{fuse}}) \\). Last but not least, we introduce ControlNet branch \\( \\epsilon_{\\phi}^{v} \\) to ensure the occupancy control, inspired by PixArt-δ [4]. This branch duplicates blocks from \\( \\epsilon_{\\theta}^{v} \\) and takes \\( F_{\\mathrm{occ}} \\) as the input. The final diffusion process is:" + }, + { + "type": "equation", + "bbox": [ + 0.499, + 0.522, + 0.895, + 0.556 + ], + "angle": 0, + "content": "\\[\n\\mathcal {L} _ {\\text {v i d e o}} = \\mathbb {E} _ {\\boldsymbol {F} _ {\\text {o c c}}, \\boldsymbol {F} _ {\\text {f u s e}}, \\boldsymbol {Z} V, \\epsilon^ {v}, \\tau} \\left[ \\left\\| \\epsilon^ {v} - \\epsilon_ {\\theta} ^ {v} \\left(\\boldsymbol {Z} _ {\\tau} ^ {\\boldsymbol {V}}, \\boldsymbol {F} _ {\\text {f u s e}}, \\tau , \\epsilon_ {\\phi} ^ {v} \\left(\\boldsymbol {F} _ {\\text {o c c}}\\right)\\right) \\right\\| ^ {2} \\right]. \\tag {11}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.554, + 0.892, + 0.598 + ], + "angle": 0, + "content": "This approach ensures fine-grained control over spatial and temporal coherence, producing artifact-free frames that meet the high fidelity requirements for AD simulation." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.599, + 0.892, + 0.689 + ], + "angle": 0, + "content": "Auto-regressive Generation. To generate temporally consistent and content-coherent video frames, we employ an auto-regressive generation strategy [60] based on first-\\(f\\) frame mask. This approach models the dependency of each video frame on the previous frames, ensuring smooth transitions and maintaining visual continuity over time:" + }, + { + "type": "equation", + "bbox": [ + 0.603, + 0.697, + 0.788, + 0.714 + ], + "angle": 0, + "content": "\\[\n\\boldsymbol {V} ^ {t} = \\mathcal {F} _ {\\text {v i d e o}} (\\boldsymbol {V} ^ {t - 1}, \\boldsymbol {Z} ^ {V}, t),\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.718, + 0.892, + 0.779 + ], + "angle": 0, + "content": "where \\( V^t \\) is the video frame at time \\( t \\), and \\( \\mathcal{F}_{\\mathrm{video}} \\) is the video generation function that conditions on the latent representation \\( Z^{V} \\) and the previous frame \\( V^{t-1} \\). This ensures temporal coherence and infinite video sequences." + }, + { + "type": "title", + "bbox": [ + 0.499, + 0.787, + 0.882, + 0.804 + ], + "angle": 0, + "content": "3.3. Agent Interplay and Closed-Loop Simulation" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.81, + 0.892, + 0.871 + ], + "angle": 0, + "content": "We enable seamless coordination of autonomous driving agents within the DrivingSphere simulation environment, categorizing agents into two primary types: the Ego Agent and Environment Agents." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.871, + 0.892, + 0.901 + ], + "angle": 0, + "content": "Ego Agent: The Ego Agent represents the autonomous driving system under evaluation. Driven by end-to-end" + }, + { + "type": "page_number", + "bbox": [ + 0.479, + 0.945, + 0.52, + 0.957 + ], + "angle": 0, + "content": "27536" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.077, + 0.091, + 0.47, + 0.125 + ], + "angle": 0, + "content": "models [6, 19, 20], it receives visual input frames \\( V^{t} \\) and outputs the predicted control signal \\( c_{\\mathrm{ego}}^{t + 1} \\) at each time \\( t \\):" + }, + { + "type": "equation", + "bbox": [ + 0.206, + 0.134, + 0.469, + 0.153 + ], + "angle": 0, + "content": "\\[\n\\boldsymbol {c} _ {\\text {e g o}} ^ {t + 1} = \\mathcal {F} _ {\\text {d r i v i n g}} \\left(\\boldsymbol {V} ^ {t}\\right), \\tag {12}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.162, + 0.444, + 0.177 + ], + "angle": 0, + "content": "where \\(\\mathcal{F}_{\\mathrm{driving}}\\) denotes the driving algorithm [6, 19, 20]." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.178, + 0.469, + 0.297 + ], + "angle": 0, + "content": "Environment Agents: Environment Agents are responsible for controlling the behavior and actions of all other actors within the simulation world \\( W \\). To achieve realistic info exchanges, we utilize the traffic flow engine [46], which supports multi-agent simulation. Environment Agents receive inputs from the simulation's state and output control signal \\( c_{\\mathrm{n}}^{t + 1} \\) that dictate the movements and interplay of actors. Formally, this process can be described as:" + }, + { + "type": "equation", + "bbox": [ + 0.204, + 0.307, + 0.469, + 0.325 + ], + "angle": 0, + "content": "\\[\n\\boldsymbol {c} _ {\\mathrm {n}} ^ {t + 1} = \\mathcal {F} _ {\\text {s y s t e m}} (\\boldsymbol {W} ^ {t}), \\tag {13}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.335, + 0.411, + 0.35 + ], + "angle": 0, + "content": "where \\(\\mathcal{F}_{\\mathrm{system}}\\) denotes the transportation simulator." + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.351, + 0.469, + 0.442 + ], + "angle": 0, + "content": "These agents take as input the visual frames \\(V^t\\) or previous state \\(\\pmb{W}^{t}\\), and output actors' respective control signals \\(\\pmb{c}_{\\mathrm{ego}}^{t + 1}\\) and \\(\\pmb{c}_{\\mathrm{n}}^{t + 1}\\). These signals influence the actor positions using Eq. (6) and further dynamically update the 4D driving world \\(\\pmb{W}^{t + 1}\\), maintaining the closed-loop simulation where the environment continuously adapts to the agents' actions." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.455, + 0.208, + 0.472 + ], + "angle": 0, + "content": "4. Experiments" + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.481, + 0.175, + 0.496 + ], + "angle": 0, + "content": "4.1. Settings" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.504, + 0.469, + 0.669 + ], + "angle": 0, + "content": "Dataset and Evaluation Metrics. We utilize real-world driving datasets nuScenes [3]. These datasets provide rich sensor data, including LiDAR point clouds, camera images, and high-definition maps, which are essential for generating accurate occupancy grids and semantic information. We employ the following metrics to evaluate the performance of our simulation framework: FVD and UniAD-based metrics [19] for visual quality and fidelity. As for the evaluation of OccDreamer, we use FID and MMD [22]. In open-loop evaluation, we borrow metrics from NAVSIM [8]. RC and ADS for the closed-loop following [51]." + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.671, + 0.437, + 0.686 + ], + "angle": 0, + "content": "Implement Details are included in the supplementary." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.696, + 0.325, + 0.712 + ], + "angle": 0, + "content": "4.2. Visual Fidelity and Realism" + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.72, + 0.469, + 0.765 + ], + "angle": 0, + "content": "We first evaluate the fidelity of the voxel-based scenes from OccDreamer and video sequences from VideoDreamer against real nuScenes data to assess the domain gap." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.766, + 0.469, + 0.901 + ], + "angle": 0, + "content": "Scene Generation. We conduct quantitative and qualitative analyses to assess the fidelity of OccDreamer's generated occupancy data. Our method outperforms existing approaches quantitatively, as shown in Tab. 2. Specifically, SemCity employs a VAE trained on three planes to generate occupancy data while OccDreamer achieves pixel-to-pixel alignment based on BEV maps, resulting in more robust and accurate occupancy generation. Qualitatively, Fig. 5 presents the results of OccDreamer with both caption and" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.092, + 0.892, + 0.258 + ], + "angle": 0, + "content": "BEV map. To the best of our knowledge, our approach is the first to achieve text-guided occupancy generation. As demonstrated, abstract textual descriptions effectively guide the model to generate diverse scenes, including outdoor environments and urban landscapes. Under BEV map control, the generated samples accurately reflect road structures and lane configurations. These capabilities establish a solid foundation for the comprehensive modeling of the 4D driving world, ensuring that the generated scenes are both spatially continuous and logically consistent. More visualizations are provided in the supplementary materials." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.258, + 0.893, + 0.47 + ], + "angle": 0, + "content": "Video Generation. We further assess the performance of our video generation model, VideoDreamer, in generating coherent video. We conduct a comparative analysis against leading models such as MagicDrive [13], DriveArena [51], and Panacea [45]. Using UniAD [19] as the evaluation framework, we assess the performance across crucial metrics including 3D object detection, BEV map segmentation, and trajectory planning. Our experimental results, summarized in Tab. 1, indicate that VideoDreamer significantly outperforms the baseline methods in these key metrics. This performance underscores the advantages of occupancy data over traditional 2D conditions, affirming VideoDreamer's effectiveness in creating realistic, dynamic scenes that accurately simulate real-world driving complexities." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.47, + 0.892, + 0.606 + ], + "angle": 0, + "content": "Qualitatively, as shown in Fig. 6, our approach demonstrates markedly superior performance in maintaining coherence and fidelity over time, significantly outperforming others. This success is primarily due to the geometry-aware occupancy encoding and instance encoding that ensures appearance consistency, highlighting VideoDreamer's robustness in creating visually consistent, detailed driving environments that capture real-world scene dynamics. More visualizations are included in the supplementary materials." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.617, + 0.706, + 0.633 + ], + "angle": 0, + "content": "4.3. Open-loop Evaluation" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.64, + 0.892, + 0.791 + ], + "angle": 0, + "content": "In the open-loop setting, where AD agents passively receive environment input without influencing simulation dynamics, we assess the performance of UniAD [19] on DrivingSphere and alternative simulators. Using 100 predefined trajectories, Tab. 3 demonstrates better PDMS scores achieved by AD agents in DrivingSphere, suggesting that its higher visual fidelity reduces perceptual inaccuracies that can misguide agent decisions. This improvement highlights DrivingSphere's suitability for testing AD systems in environments that closely emulate real-world conditions." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.802, + 0.715, + 0.817 + ], + "angle": 0, + "content": "4.4. Closed-loop Evaluation" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.826, + 0.892, + 0.901 + ], + "angle": 0, + "content": "In the closed-loop evaluation, AD agents receive visual input and output control signals, thus interactively shaping the simulation. This evaluation setup involves 100 pre-defined trajectories to test across controlled yet varied scenarios. As shown in Tab. 4, UniAD [19] exhibits relatively low Route" + }, + { + "type": "page_number", + "bbox": [ + 0.479, + 0.945, + 0.518, + 0.957 + ], + "angle": 0, + "content": "27537" + } + ], + [ + { + "type": "table", + "bbox": [ + 0.082, + 0.088, + 0.894, + 0.199 + ], + "angle": 0, + "content": "
Source DataQuality3DODSegmentationL2(↓)
FVD(↓)mAP(↑)NDS(↑)Lanes(↑)Drivable(↑)Divider(↑)1.0s2.0s3.0s
ori nuScenes[3]—37.9849.8531.3169.1425.930.510.981.65
MagicDrive [13]218.1212.9228.3621.9551.4617.100.571.141.95
Panacea [45]13913.7227.7318.2352.3717.210.581.141.97
DriveArena [51]185.3216.0630.0326.1459.3720.790.561.101.89
DrivingSphere103.4221.4534.1627.9962.8722.290.541.101.76
" + }, + { + "type": "table_caption", + "bbox": [ + 0.076, + 0.201, + 0.895, + 0.244 + ], + "angle": 0, + "content": "Table 1. Quantitative comparison of video generation fidelity for autonomous driving simulations on the nuScenes validation set. Using UniAD [19] for evaluation on generated data, DrivingSphere achieves high fidelity in video generation, closely matching real-world data across key perception and planning metrics." + }, + { + "type": "image", + "bbox": [ + 0.082, + 0.246, + 0.891, + 0.519 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.076, + 0.52, + 0.895, + 0.59 + ], + "angle": 0, + "content": "Figure 6. Comparison of generated driving simulations from (a) Panacea [45], (b) Our method, and (c) Real Scene [3] across multiple views and time steps. Red boxes highlight instance-level appearance consistency, with our method maintaining more coherent actor appearances across frames. Yellow boxes illustrate temporal consistency, where our approach achieves smoother transitions over time. Green boxes and background elements demonstrate our model's capability to capture and accurately model non-direct traffic participants, such as buildings and other environmental features, enhancing realism in the generated scene." + }, + { + "type": "table", + "bbox": [ + 0.082, + 0.6, + 0.47, + 0.644 + ], + "angle": 0, + "content": "
MethodsFID(↓)MMD(↓)
Semicity [22]6340.251
Ours2740.082
" + }, + { + "type": "table_caption", + "bbox": [ + 0.077, + 0.645, + 0.47, + 0.685 + ], + "angle": 0, + "content": "Table 2. Quantitative comparison of scene generation performance, following the setup in SemCity. The scene data generated by our method is rendered into 2D images." + }, + { + "type": "table", + "bbox": [ + 0.082, + 0.686, + 0.469, + 0.759 + ], + "angle": 0, + "content": "
Data SourceNC(↑)DAC(↑)TTC(↑)PDMS(↑)
nuScenes [3]0.9930.9950.9470.910
DriveArena [51]0.7920.9420.7710.636
DriveArena* [51]0.8290.9640.8120.698
DrivingSphere0.8520.9680.8530.742
" + }, + { + "type": "table_caption", + "bbox": [ + 0.076, + 0.76, + 0.47, + 0.802 + ], + "angle": 0, + "content": "Table 3. Open-loop evaluation of UniAD [19] on original nuScenes sequences, DriveArena's reported results, and our predefined 100-trajectory test for DriveArena* and DrivingSphere." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.81, + 0.471, + 0.903 + ], + "angle": 0, + "content": "Completion (RC) scores, with completing an average of only \\(11.7\\%\\) of each route. Compared to DriveArena [51], our simulation yields consistently better performance metrics, and, along with the superior visual fidelity demonstrated in Tab. 1, affirms DrivingSphere's effectiveness as a high-fidelity platform for AD algorithm validation." + }, + { + "type": "table", + "bbox": [ + 0.504, + 0.6, + 0.892, + 0.646 + ], + "angle": 0, + "content": "
RoutePDMS(↑)RC(↑)ADS(↑)
DriveArena [51]0.69010.06410.0508
DrivingSphere0.72810.11700.0851
" + }, + { + "type": "table_caption", + "bbox": [ + 0.499, + 0.647, + 0.892, + 0.676 + ], + "angle": 0, + "content": "Table 4. Closed-loop evaluation of UniAD's [19] performance on 100 pre-defined routes." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.689, + 0.62, + 0.705 + ], + "angle": 0, + "content": "5. Conclusion" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.72, + 0.894, + 0.901 + ], + "angle": 0, + "content": "We presented *DrivingSphere*, a generative closed-loop simulation framework that bridges the gap between traditional closed-loop simulations and open-loop generative models. Through advanced occupancy-based modeling and controllable generation mechanisms, *DrivingSphere* creates realistic, high-fidelity simulation for autonomous driving. Our experiments demonstrate superior visual quality, temporal consistency, and the ability to effectively test AD algorithms in dynamic environments. In future work, we aim to enhance dynamic agent behaviors and incorporate more complex real-world scenarios to further improve the robustness of our simulations." + }, + { + "type": "page_number", + "bbox": [ + 0.479, + 0.945, + 0.52, + 0.957 + ], + "angle": 0, + "content": "27538" + } + ], + [ + { + "type": "title", + "bbox": [ + 0.079, + 0.09, + 0.176, + 0.107 + ], + "angle": 0, + "content": "References" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.115, + 0.47, + 0.17 + ], + "angle": 0, + "content": "[1] Maxim Berman, Amal Rannen Triki, and Matthew B Blaschko. The lovasz-softmax loss: A tractable surrogate for the optimization of the intersection-over-union measure in neural networks. In CVPR, 2018. 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.171, + 0.472, + 0.242 + ], + "angle": 0, + "content": "[2] Andreas Blattmann, Tim Dockhorn, Sumith Kulal, Daniel Mendelevitch, Maciej Kilian, Dominik Lorenz, Yam Levi, Zion English, Vikram Voleti, Adam Letts, et al. Stable video diffusion: Scaling latent video diffusion models to large datasets. arXiv preprint arXiv:2311.15127, 2023. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.242, + 0.471, + 0.31 + ], + "angle": 0, + "content": "[3] Holger Caesar, Varun Bankiti, Alex H. Lang, Sourabh Vora, Venice Erin Liong, Qiang Xu, Anush Krishnan, Yu Pan, Giancarlo Baldan, and Oscar Beijbom. nuscenes: A multimodal dataset for autonomous driving. In CVPR, 2020. 1, 2, 4, 7, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.311, + 0.471, + 0.368 + ], + "angle": 0, + "content": "[4] Junsong Chen, Yue Wu, Simian Luo, Enze Xie, Sayak Paul, Ping Luo, Hang Zhao, and Zhenguo Li. Pixart-{\\delta} : Fast and controllable image generation with latent consistency models. arXiv preprint arXiv:2401.05252, 2024. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.368, + 0.471, + 0.41 + ], + "angle": 0, + "content": "[5] Li Chen, Penghao Wu, Kashyap Chitta, Bernhard Jaeger, Andreas Geiger, and Hongyang Li. End-to-end autonomous driving: Challenges and frontiers. T-PAMI, 2024. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.41, + 0.471, + 0.478 + ], + "angle": 0, + "content": "[6] Shaoyu Chen, Bo Jiang, Hao Gao, Bencheng Liao, Qing Xu, Qian Zhang, Chang Huang, Wenyu Liu, and Xinggang Wang. Vadv2: End-to-end vectorized autonomous driving via probabilistic planning. arXiv preprint arXiv:2402.13243, 2024. 2, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.48, + 0.471, + 0.522 + ], + "angle": 0, + "content": "[7] Yuan Chen, Zi-han Ding, Ziqin Wang, Yan Wang, Lijun Zhang, and Si Liu. Asynchronous large language model enhanced planner for autonomous driving. In ECCV, 2025. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.522, + 0.471, + 0.605 + ], + "angle": 0, + "content": "[8] Daniel Dauner, Marcel Hallgarten, Tianyu Li, Xinshuo Weng, Zhiyu Huang, Zetong Yang, Hongyang Li, Igor Gilitschenski, Boris Ivanovic, Marco Pavone, Andreas Geiger, and Kashyap Chitta. Navsim: Data-driven non-reactive autonomous vehicle simulation and benchmarking. In NeurIPS, 2024. 7, 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.606, + 0.471, + 0.661 + ], + "angle": 0, + "content": "[9] Wenhao Ding, Yulong Cao, Ding Zhao, Chaowei Xiao, and Marco Pavone. Realgen: Retrieval augmented generation for controllable traffic scenarios. arXiv preprint arXiv:2312.13303, 2023. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.079, + 0.662, + 0.471, + 0.704 + ], + "angle": 0, + "content": "[10] Alexey Dosovitskiy, German Ros, Felipe Codevilla, Antonio Lopez, and Vladlen Koltun. Carla: An open urban driving simulator. In CoRL, 2017. 1, 2, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.079, + 0.705, + 0.471, + 0.746 + ], + "angle": 0, + "content": "[11] Lan Feng, Quanyi Li, Zhenghao Peng, Shuhan Tan, and Bolei Zhou. Trafficgen: Learning to generate diverse and realistic traffic scenarios. In ICRA, 2023. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.079, + 0.747, + 0.471, + 0.815 + ], + "angle": 0, + "content": "[12] Daocheng Fu, Wenjie Lei, Licheng Wen, Pinlong Cai, Song Mao, Min Dou, Botian Shi, and Yu Qiao. Limsim++: A closed-loop platform for deploying multimodal llms in autonomous driving. arXiv preprint arXiv:2402.01246, 2024. 1, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.079, + 0.817, + 0.471, + 0.872 + ], + "angle": 0, + "content": "[13] Ruiyuan Gao, Kai Chen, Enze Xie, Lanqing Hong, Zhenguo Li, Dit-Yan Yeung, and Qiang Xu. Magicdrive: Street view generation with diverse 3d geometry control. arXiv preprint arXiv:2310.02601, 2023. 1, 2, 3, 5, 6, 7, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.079, + 0.873, + 0.471, + 0.903 + ], + "angle": 0, + "content": "[14] Shenyuan Gao, Jiazhi Yang, Li Chen, Kashyap Chitta, Yihang Qiu, Andreas Geiger, Jun Zhang, and Hongyang Li." + }, + { + "type": "list", + "bbox": [ + 0.079, + 0.115, + 0.472, + 0.903 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.533, + 0.093, + 0.892, + 0.12 + ], + "angle": 0, + "content": "Vista: A generalizable driving world model with high fidelity and versatile controllability. In NeurIPS, 2024. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.121, + 0.894, + 0.178 + ], + "angle": 0, + "content": "[15] Wencheng Han, Dongqian Guo, Cheng-Zhong Xu, and Jianbing Shen. Dme-driver: Integrating human decision logic and 3d scene perception in autonomous driving. arXiv preprint arXiv:2401.03641, 2024. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.179, + 0.894, + 0.233 + ], + "angle": 0, + "content": "[16] Yingqing He, Tianyu Yang, Yong Zhang, Ying Shan, and Qifeng Chen. Latent video diffusion models for high-fidelity long video generation. arXiv preprint arXiv:2211.13221, 2022. 3, 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.234, + 0.892, + 0.262 + ], + "angle": 0, + "content": "[17] Jonathan Ho, Ajay Jain, and Pieter Abbeel. Denoising diffusion probabilistic models. NeurIPS, 2020. 3, 4, 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.263, + 0.892, + 0.292 + ], + "angle": 0, + "content": "[18] Qianjiang Hu, Zhimin Zhang, and Wei Hu. Rangeldm: Fast realistic lidar point cloud generation. In ECCV, 2025. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.293, + 0.892, + 0.348 + ], + "angle": 0, + "content": "[19] Yihan Hu, Jiazhi Yang, Li Chen, Keyu Li, Chonghao Sima, Xizhou Zhu, Siqi Chai, Senyao Du, Tianwei Lin, Wenhai Wang, et al. Planning-oriented autonomous driving. In CVPR, 2023. 1, 2, 7, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.349, + 0.892, + 0.405 + ], + "angle": 0, + "content": "[20] Bo Jiang, Shaoyu Chen, Qing Xu, Bencheng Liao, Jiajie Chen, Helong Zhou, Qian Zhang, Wenyu Liu, Chang Huang, and Xinggang Wang. Vad: Vectorized scene representation for efficient autonomous driving. In ICCV, 2023. 1, 2, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.406, + 0.892, + 0.449 + ], + "angle": 0, + "content": "[21] Bernhard Kerbl, Georgios Kopanas, Thomas Leimkuhler, and George Drettakis. 3d gaussian splatting for real-time radiance field rendering. ACM Trans. Graph., 2023. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.449, + 0.892, + 0.491 + ], + "angle": 0, + "content": "[22] Jumin Lee, Sebin Lee, Changho Jo, Woobin Im, Juhyeong Seon, and Sung-Eui Yoon. Semcity: Semantic scene generation with triplane diffusion. In CVPR, 2024. 3, 7, 8, 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.491, + 0.892, + 0.547 + ], + "angle": 0, + "content": "[23] Quanyi Li, Zhenghao Peng, Lan Feng, Qihang Zhang, Zhenghai Xue, and Bolei Zhou. Metadrive: Composing diverse driving scenarios for generalizable reinforcement learning. T-PAMI, 2022. 1, 2, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.548, + 0.892, + 0.591 + ], + "angle": 0, + "content": "[24] Zhiqi Li, Zhiding Yu, Shiyi Lan, Jiahan Li, Jan Kautz, Tong Lu, and Jose M Alvarez. Is ego status all you need for open-loop end-to-end autonomous driving? In CVPR, 2024. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.591, + 0.892, + 0.633 + ], + "angle": 0, + "content": "[25] Zhiqi Li, Zhiding Yu, Shiyi Lan, Jiahan Li, Jan Kautz, Tong Lu, and Jose M Alvarez. Is ego status all you need for open-loop end-to-end autonomous driving? In CVPR, 2024. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.634, + 0.892, + 0.702 + ], + "angle": 0, + "content": "[26] William Ljungbergh, Adam Tonderski, Joakim Johnander, Holger Caesar, Kalle Åström, Michael Felsberg, and Christoffer Petersson. Neuroncap: Photorealistic closed-loop safety testing for autonomous driving. In ECCV, 2025. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.704, + 0.892, + 0.773 + ], + "angle": 0, + "content": "[27] Pablo Alvarez Lopez, Michael Behrisch, Laura Bieker-Walz, Jakob Erdmann, Yun-Pang Flotteröd, Robert Hilbrich, Leonhard Lücken, Johannes Rummel, Peter Wagner, and Evamarie Wießner. Microscopic traffic simulation using sumo. In IEEE ITSC, 2018. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.774, + 0.892, + 0.845 + ], + "angle": 0, + "content": "[28] Enhui Ma, Lijun Zhou, Tao Tang, Zhan Zhang, Dong Han, Junpeng Jiang, Kun Zhan, Peng Jia, Xianpeng Lang, Haiyang Sun, et al. Unleashing generalization of end-to-end autonomous driving with controllable long video generation. arXiv preprint arXiv:2406.01349, 2024. 5, 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.845, + 0.892, + 0.901 + ], + "angle": 0, + "content": "[29] Ben Mildenhall, Pratul P Srinivasan, Matthew Tancik, Jonathan T Barron, Ravi Ramamoorthi, and Ren Ng. Nerf: Representing scenes as neural radiance fields for view synthesis. Communications of the ACM, 2021. 3, 6" + }, + { + "type": "list", + "bbox": [ + 0.503, + 0.093, + 0.894, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.479, + 0.945, + 0.521, + 0.958 + ], + "angle": 0, + "content": "27539" + } + ], + [ + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.092, + 0.47, + 0.133 + ], + "angle": 0, + "content": "[30] Kazuto Nakashima and Ryo Kurazume. Lidar data synthesis with denoising diffusion probabilistic models. In ICRA, 2024. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.134, + 0.47, + 0.203 + ], + "angle": 0, + "content": "[31] Alec Radford, Jong Wook Kim, Chris Hallacy, Aditya Ramesh, Gabriel Goh, Sandhini Agarwal, Girish Sastry, Amanda Askell, Pamela Mishkin, Jack Clark, et al. Learning transferable visual models from natural language supervision. In ICML, 2021. 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.204, + 0.471, + 0.273 + ], + "angle": 0, + "content": "[32] Colin Raffel, Noam Shazeer, Adam Roberts, Katherine Lee, Sharan Narang, Michael Matena, Yanqi Zhou, Wei Li, and Peter J. Liu. Exploring the limits of transfer learning with a unified text-to-text transformer. Journal of Machine Learning Research, 2020. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.274, + 0.47, + 0.315 + ], + "angle": 0, + "content": "[33] Prabhu Ramachandran and Gaël Varoquaux. Mayavi: 3d visualization of scientific data. Computing in Science & Engineering, 2011. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.316, + 0.47, + 0.357 + ], + "angle": 0, + "content": "[34] Haoxi Ran, Vitor Guizilini, and Yue Wang. Towards realistic scene generation with lidar diffusion models. In ECCV, 2024. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.358, + 0.47, + 0.399 + ], + "angle": 0, + "content": "[35] Robin Rombach, Andreas Blattmann, Dominik Lorenz, Patrick Esser, and Björn Ommer. High-resolution image synthesis with latent diffusion models. In CVPR, 2022. 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.4, + 0.47, + 0.455 + ], + "angle": 0, + "content": "[36] Hao Shao, Yuxuan Hu, Letian Wang, Guanglu Song, Steven L Waslander, Yu Liu, and Hongsheng Li. Lmdrive: Closed-loop end-to-end driving with large language models. In CVPR, 2024. 1, 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.456, + 0.47, + 0.523 + ], + "angle": 0, + "content": "[37] Pei Sun, Henrik Kretzschmar, Xerxes Dotiwalla, Aurelien Chouard, Vijaysai Patnaik, Paul Tsui, James Guo, Yin Zhou, Yuning Chai, Benjamin Caine, et al. Scalability in perception for autonomous driving: Waymo open dataset. In CVPR, 2020. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.525, + 0.47, + 0.566 + ], + "angle": 0, + "content": "[38] Alexander Swerdlow, Runsheng Xu, and Bolei Zhou. Streetview image generation from a bird's-eye view layout. IEEE RA-L, 2024. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.567, + 0.47, + 0.609 + ], + "angle": 0, + "content": "[39] Shuhan Tan, Boris Ivanovic, Xinshuo Weng, Marco Pavone, and Philipp Kraehenbuehl. Language conditioned traffic generation. arXiv preprint arXiv:2307.07947, 2023. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.609, + 0.47, + 0.637 + ], + "angle": 0, + "content": "[40] Aaron Van Den Oord, Oriol Vinyals, et al. Neural discrete representation learning. NeurIPS, 2017. 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.637, + 0.47, + 0.705 + ], + "angle": 0, + "content": "[41] Lening Wang, Wenzhao Zheng, Yilong Ren, Han Jiang, Zhiyong Cui, Haiyang Yu, and Jiwen Lu. Occsora: 4d occupancy generation models as world simulators for autonomous driving. arXiv preprint arXiv:2405.20337, 2024. 2, 3, 5, 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.706, + 0.47, + 0.789 + ], + "angle": 0, + "content": "[42] Wenhai Wang, Jiangwei Xie, ChuanYang Hu, Haoming Zou, Jianan Fan, Wenwen Tong, Yang Wen, Silei Wu, Hanming Deng, Zhiqi Li, et al. Drivemlm: Aligning multi-modal large language models with behavioral planning states for autonomous driving. arXiv preprint arXiv:2312.09245, 2023. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.79, + 0.47, + 0.845 + ], + "angle": 0, + "content": "[43] Xiaofeng Wang, Zheng Zhu, Guan Huang, Xinze Chen, Jia-gang Zhu, and Jiwen Lu. Drivedreamer: Towards real-world-driven world models for autonomous driving. arXiv preprint arXiv:2309.09777, 2023. 1, 2, 3, 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.846, + 0.47, + 0.901 + ], + "angle": 0, + "content": "[44] Yuqi Wang, Jiawei He, Lue Fan, Hongxin Li, Yuntao Chen, and Zhaoxiang Zhang. Driving into the future: Multiview visual forecasting and planning with world model for autonomous driving. In CVPR, 2024. 6, 1, 2" + }, + { + "type": "list", + "bbox": [ + 0.08, + 0.092, + 0.471, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.092, + 0.893, + 0.161 + ], + "angle": 0, + "content": "[45] Yuqing Wen, Yucheng Zhao, Yingfei Liu, Fan Jia, Yanhui Wang, Chong Luo, Chi Zhang, Tiancai Wang, Xiaoyan Sun, and Xiangyu Zhang. Panacea: Panoramic and controllable video generation for autonomous driving. In CVPR, 2024. 3, 7, 8, 1, 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.163, + 0.893, + 0.218 + ], + "angle": 0, + "content": "[46] Licheng Wenl, Daocheng Fu, Song Mao, Pinlong Cai, Min Dou, Yikang Li, and Yu Qiao. Limsim: A long-term interactive multi-scenario traffic simulator. In IEEE ITSC, 2023. 1, 3, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.22, + 0.893, + 0.29 + ], + "angle": 0, + "content": "[47] Zirui Wu, Tianyu Liu, Liyi Luo, Zhide Zhong, Jianteng Chen, Hongmin Xiao, Chao Hou, Haozhe Lou, Yuantao Chen, Runyi Yang, et al. Mars: An instance-aware, modular and realistic simulator for autonomous driving. In CAAI International Conference on Artificial Intelligence, 2023. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.291, + 0.893, + 0.347 + ], + "angle": 0, + "content": "[48] Zhenhua Xu, Yujia Zhang, Enze Xie, Zhen Zhao, Yong Guo, Kwan-Yee K Wong, Zhenguo Li, and Hengshuang Zhao. Drivegpt4: Interpretable end-to-end autonomous driving via large language model. IEEE RA-L, 2024. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.348, + 0.893, + 0.403 + ], + "angle": 0, + "content": "[49] Tianyi Yan, Junbo Yin, Xianpeng Lang, Ruigang Yang, Cheng-Zhong Xu, and Jianbing Shen. Olidm: Object-aware lidar diffusion models for autonomous driving. In AAAI, 2025. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.404, + 0.893, + 0.461 + ], + "angle": 0, + "content": "[50] Kairui Yang, Enhui Ma, Jibin Peng, Qing Guo, Di Lin, and Kaicheng Yu. Bevcontrol: Accurately controlling street-view elements with multi-perspective consistency via bev sketch layout. arXiv preprint arXiv:2308.01661, 2023. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.461, + 0.893, + 0.545 + ], + "angle": 0, + "content": "[51] Xuemeng Yang, Licheng Wen, Yukai Ma, Jianbiao Mei, Xin Li, Tiantian Wei, Wenjie Lei, Daocheng Fu, Pinlong Cai, Min Dou, Botian Shi, Liang He, Yong Liu, and Yu Qiao. Drivearena: A closed-loop generative simulation platform for autonomous driving. arXiv preprint arXiv:2408.00415, 2024. 1, 3, 7, 8, 2, 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.546, + 0.893, + 0.601 + ], + "angle": 0, + "content": "[52] Ze Yang, Yun Chen, Jingkang Wang, Sivabalan Manivasagam, Wei-Chiu Ma, Anqi Joyce Yang, and Raquel Urtasun. Unisim: A neural closed-loop sensor simulator. In CVPR, 2023. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.602, + 0.893, + 0.643 + ], + "angle": 0, + "content": "[53] Jimuyang Zhang, Zanming Huang, Arijit Ray, and Eshed Ohn-Bar. Feedback-guided autonomous driving. In CVPR, 2024. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.645, + 0.893, + 0.7 + ], + "angle": 0, + "content": "[54] Linrui Zhang, Zhenghao Peng, Quanyi Li, and Bolei Zhou. Cat: Closed-loop adversarial training for safe end-to-end driving. In 7th Annual Conference on Robot Learning, 2023. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.702, + 0.893, + 0.744 + ], + "angle": 0, + "content": "[55] Lvmin Zhang, Anyi Rao, and Maneesh Agrawala. Adding conditional control to text-to-image diffusion models. In ICCV, 2023. 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.745, + 0.893, + 0.801 + ], + "angle": 0, + "content": "[56] Yuheng Zhang, Tianjian Ouyang, Fudan Yu, Cong Ma, Lei Qiao, Wei Wu, Jian Yuan, and Yong Li. Lcsim: A large-scale controllable traffic simulator. arXiv preprint arXiv:2406.19781, 2024. 1, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.802, + 0.893, + 0.871 + ], + "angle": 0, + "content": "[57] Guosheng Zhao, Xiaofeng Wang, Zheng Zhu, Xinze Chen, Guan Huang, Xiaoyi Bao, and Xingang Wang. Drivedreamer-2: Llm-enhanced world models for diverse driving video generation. arXiv preprint arXiv:2403.06845, 2024. 1, 2, 3, 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.872, + 0.893, + 0.903 + ], + "angle": 0, + "content": "[58] Wenzhao Zheng, Weiliang Chen, Yuanhui Huang, Borui Zhang, Yueqi Duan, and Jiwen Lu. Occworld: Learning a 3d" + }, + { + "type": "list", + "bbox": [ + 0.503, + 0.092, + 0.893, + 0.903 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.48, + 0.945, + 0.52, + 0.957 + ], + "angle": 0, + "content": "27540" + } + ], + [ + { + "type": "ref_text", + "bbox": [ + 0.108, + 0.092, + 0.468, + 0.119 + ], + "angle": 0, + "content": "occupancy world model for autonomous driving. In ECCV, 2024.3,1,2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.122, + 0.468, + 0.175 + ], + "angle": 0, + "content": "[59] Wenzhao Zheng, Ruiqi Song, Xianda Guo, Chenming Zhang, and Long Chen. Genad: Generative end-to-end autonomous driving. arXiv preprint arXiv:2402.11502, 2024. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.178, + 0.468, + 0.233 + ], + "angle": 0, + "content": "[60] Zangwei Zheng, Xiangyu Peng, Tianji Yang, Chenhui Shen, Shenggui Li, Hongxin Liu, Yukun Zhou, Tianyi Li, and Yang You. Open-sora: Democratizing efficient video production for all, 2024. 5, 6, 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.235, + 0.468, + 0.262 + ], + "angle": 0, + "content": "[61] Vlas Zyrianov, Xiyue Zhu, and Shenlong Wang. Learning to generate realistic lidar point cloud. In ECCV, 2022. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.264, + 0.468, + 0.305 + ], + "angle": 0, + "content": "[62] Vlas Zyrianov, Henry Che, Zhijian Liu, and Shenlong Wang. Lidardm: Generative lidar simulation in a generated world. arXiv preprint arXiv:2404.02903, 2024. 3" + }, + { + "type": "list", + "bbox": [ + 0.08, + 0.092, + 0.468, + 0.305 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.48, + 0.945, + 0.518, + 0.956 + ], + "angle": 0, + "content": "27541" + } + ] +] \ No newline at end of file diff --git a/2025/DrivingSphere_ Building a High-fidelity 4D World for Closed-loop Simulation/e8e4b9c2-6cbc-49b7-a622-765862a5895e_origin.pdf b/2025/DrivingSphere_ Building a High-fidelity 4D World for Closed-loop Simulation/e8e4b9c2-6cbc-49b7-a622-765862a5895e_origin.pdf new file mode 100644 index 0000000000000000000000000000000000000000..8d23beb3856b661f0f2eda930e8f353441d54f96 --- /dev/null +++ b/2025/DrivingSphere_ Building a High-fidelity 4D World for Closed-loop Simulation/e8e4b9c2-6cbc-49b7-a622-765862a5895e_origin.pdf @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:fd520cf309c9dbd750f6224b9edd23b3af18ceff131a2f4a9ce6ced45c62e418 +size 5645137 diff --git a/2025/DrivingSphere_ Building a High-fidelity 4D World for Closed-loop Simulation/full.md b/2025/DrivingSphere_ Building a High-fidelity 4D World for Closed-loop Simulation/full.md new file mode 100644 index 0000000000000000000000000000000000000000..9a8d539798c6a964f07bac45745a1c0510e45dad --- /dev/null +++ b/2025/DrivingSphere_ Building a High-fidelity 4D World for Closed-loop Simulation/full.md @@ -0,0 +1,336 @@ +# DrivingSphere: Building a High-fidelity 4D World for Closed-loop Simulation + +Tianyi Yan $^{1}$ , Dongming Wu $^{3}$ , Wencheng Han $^{1}$ , Junpeng Jiang $^{2}$ , Xia Zhou $^{2}$ , Kun Zhan $^{2}$ , Cheng-zhong Xu $^{1}$ , Jianbing Shen $^{1*}$ + +$^{1}$ SKL-IOTSC, Computer and Information Science, University of Macau, $^{2}$ Li Auto Inc., $^{3}$ School of Computer Science, Beijing Institute of Technology + +# Abstract + +Autonomous driving evaluation requires simulation environments that closely replicate actual road conditions, including real-world sensory data and responsive feedback loops. However, many existing simulations need to predict waypoints along fixed routes on public datasets or synthetic photorealistic data, i.e., open-loop simulation usually lacks the ability to assess dynamic decision-making. While the recent efforts of closed-loop simulation offer feedback-driven environments, they cannot process visual sensor inputs or produce outputs that differ from real-world data. To address these challenges, we propose DrivingSphere, a realistic and closed-loop simulation framework. Its core idea is to build 4D world representation and generate real-life and controllable driving scenarios. In specific, our framework includes a Dynamic Environment Composition module that constructs a detailed 4D driving world with a format of occupancy equipping with static backgrounds and dynamic objects, and a Visual Scene Synthesis module that transforms this data into high-fidelity, multi-view video outputs, ensuring spatial and temporal consistency. By providing a dynamic and realistic simulation environment, DrivingSphere enables comprehensive testing and validation of autonomous driving algorithms, ultimately advancing the development of more reliable autonomous cars. The benchmark will be publicly released. + +# 1. Introduction + +End-to-end autonomous driving algorithms [5-7, 15, 19, 20, 25, 36, 42, 48, 53, 54, 59] have made significant progress in recent years, making the accurate evaluation of these models a very urgent task. For safe and responsible evaluation, + +it is essential to have a precise simulation environment that accurately reflects real-world driving conditions. This requirement usually includes two aspects: firstly, the generation of high-fidelity sensory data, and secondly, the implementation of closed-loop feedback mechanisms. + +The most common evaluation methods involve the forecasting of waypoints along fixed routes within collected public datasets [3, 37], that is, open-loop simulation approaches. While these benchmarks offer realistic driving data, they have relatively fixed distributions and lack diversity [24], limiting their ability to assess the generalization capabilities of autonomous driving algorithms. Even though several recent works [13, 43, 57] excel at producing diverse, photo-realistic data, continue to rely on the prediction of waypoints that are pre-established. In summary, despite high-fidelity sensory data, these open-loop evaluation solutions cannot provide the dynamic feedback necessary for assessing how autonomous systems respond to dynamic changes and decision-making. + +In contrast, closed-loop simulations [10, 23, 46, 56] offer feedback-driven systems where an agent's actions influence and are influenced by other agents and the environment. For example, traffic flow-based simulation methods [12, 46, 56] successfully enable multi-agent simulations. However, they lack the ability to process visual sensor inputs, limiting their interplay with vision-based end-to-end models. Game engine-based simulators [10, 23] create scalable and physically realistic environments, but their outputs often diverge from real-world sensor data, restricting their utility in validating algorithms trained on actual inputs. + +To address these challenges, we propose DrivingSphere, a novel generative closed-loop simulation framework that leverages geometric-wise prior information to generate realistic and controllable driving scenarios. In comparison with existing simulations, DrivingSphere has three distinguished features: ① Rich simulation granularity. Unlike past methods that model only roads and cars [13, 51], our approach allows for including previously unmodeled elements, such as buildings, vegetation, and other environmental structures. Although these non-traffic elements are not + +![](images/25b82b33c5d51a3320144bd16ea17e52fe7cd31faa60d464b746dcd10e3109a1.jpg) +Figure 1. Comparison of frameworks for evaluating end-to-end Autonomous Driving (AD) algorithms. (a) Open-loop evaluation uses waypoint predictions along fixed routes in pre-collected datasets [3]. Generative models [13, 43, 57] create diverse, realistic data but lack dynamic feedback to assess AD responses to dynamic changes. (b) Simulated closed-loop evaluation [10, 23] offers feedback-driven, scalable environments where agent actions impact simulation dynamics; however, sensory outputs often differ from real-world data, limiting effectiveness for algorithms trained on real data. (c) Our generative closed-loop simulation framework, *DrivingSphere*, addresses these limitations by delivering realistic visual inputs and continuous, responsive feedback between the AD agent and environment. + +![](images/b28a90788599772a538a3d8c7ecc3fbd7075c7e529e415d2d9894d3d5d01cd3b.jpg) + +![](images/35657bbd26f8ca421f8cbbdc3ecd32048e8d534e2b1bd8693faf728c134b1a61.jpg) + +directly involved in traffic flow, their presence can significantly impact the input to driving models, thereby influencing decision-making processes in complex driving scenarios. 2 Physical and spatial realism. Due to our model explicitly representing the scene and traffic actors in 4D space, it is capable of precisely depicting physical interactions and occlusion relationships across different traffic elements. This ensures that each viewpoint and position naturally adheres to physical principles like depth and occlusion, thus achieving structured coordination of global road layout, traffic participants, and their behaviors. 3 High visual consistency and fidelity. Our model pays more effort to associate each traffic actor's appearance and unique ID within the scene, providing stable and high-fidelity temporal and spatial coherence across frames and views. + +The proposed DrivingSphere contains two key modules: Dynamic Environment Composition and Visual Scene Synthesis. First, the Dynamic Environment Composition module aims to produce a coarse-grained 4D world with an occupancy grid format, which integrates large static backgrounds with dynamic traffic actors. For this, we introduce a BEV-conditioned occupancy diffusion model, named OccDreamer, to generate static backgrounds. It provides an infinite city-scale scene via progressive region extension. Besides, we build an actor bank to manage key actors or objects, and update actors' spatial-temporal position. + +Once the scene is generated, the second module, Visual Scene Synthesis, converts it into high-fidelity, multi-view video outputs for AD system perception. To fully capture the complexity of the 4D environment, we employ a dual-path condition encoding strategy. Specifically, the global branch, using a pretrained 4D encoder [41], extracts geometric information and spatial-temporal relationships directly from the occupancy data, while the local projection branch produces view-specific semantic maps that align on + +a pixel-to-pixel basis, accurately capturing occlusions and depth variations. To further ensure visual consistency, we design an ID-aware actor encoding mechanism that binds the appearance and identity information of traffic actors to their scene locations, which preserves spatial correlation across views and temporal consistency across frames. With the custom-designed dual-path condition encoding and ID-aware actor encoding, the Visual Scene Synthesis generates spatially and temporally coherent video sequences. This results in a highly realistic simulation environment that is ideal for testing and validating AD systems. + +By responsive interplay between AD algorithms and the simulated environment, DrivingSphere supports a closed-loop feedback loop, where actions from AD systems [6, 19, 20, 36] influence surrounding actors and scenes. This simulation platform closely mirrors real-world conditions, providing a reliable foundation for comprehensive AD validation. Our extensive experiments show that DrivingSphere achieves superior performance in visual fidelity and temporal consistency, significantly reducing the domain gap between simulated and real-world environments. By excelling in both open-loop and closed-loop evaluations, DrivingSphere proves to be a reliable platform for dynamic agent-environment interplay in rigorous AD testing. Ultimately, DrivingSphere will drive progress toward the development of safer and more reliable autonomous vehicles. Overall, our contributions are four-fold: + +- We introduce DrivingSphere, a novel geometry-aware closed-loop simulation framework that captures 2D visual and 3D geometric properties while seamlessly integrating with vision-based end-to-end driving agents. +- We propose the Dynamic Environment Composition that provides a comprehensive 4D driving world by modeling both static and dynamic elements, enabling intricate simulations across diverse environments. + +- We design the Visual Scene Synthesis that enables long-term consistency and high fidelity, addressing challenges in generating continuous and multi-view video sequences for AD simulations. +- We introduce the Agent Coordination module for scalable multi-agent systems, enabling the integration of multiple vehicle agents and establishing a crucial feedback loop for validating autonomous driving algorithms. + +# 2. Related Work + +Generative Models in Driving. Denoising Diffusion Probabilistic Models (DDPMs) [17] have transformed image generation by iteratively refining Gaussian noise into high-quality images. Based on this technique, generative models in autonomous driving have greatly advanced. For example, BEVGen [38] and BEVControl [50] use diffusion models to produce controllable street-view images from BEV layouts. More recently, diffusion-based video generation models [2, 16] enable methods like [13, 14, 43-45, 57] to generate multi-view driving videos, effectively supporting perception tasks. Inspired by these 2D methods, advances in 3D scene generation have also emerged. Integrating diffusion processes into LiDAR generation improves the realism of LiDAR scenes [18, 30, 34, 49, 61, 62]. Occupancy-based models have further progressed, with occupancy grids now central to 3D scene generation. SemCity [22] enhances outdoor 3D occupancy generation, while models like OccWorld [58] and OccSora [41] use occupancy data for vehicle motion prediction and trajectory-conditioned 4D occupancy generation, respectively. Unlike these, our OccDreamer combines text and BEV conditions to generate 3D occupancy data, dynamically enhanced with foreground motion and ego pose to form a complete 4D driving world. By conditioning on occupancy data, our VideoDreamer allows precise control over non-traffic elements and ensures spatial and temporal coherence, capturing essential physical relationships such as depth and occlusion. + +Autonomous Driving Simulations. Autonomous driving simulation methods are developed from basic behavioral models to complex, interactive environments. Flow-based simulations [9, 11, 12, 39, 46, 56] enable multi-agent interactions through long-term traffic planning. However, these methods typically operate in 2D space and lack real-life visual outputs. Platforms like CARLA [10], SUMO [27], and MetaDrive [23] address these limitations by modeling traffic dynamics and supporting multi-agent interactions with visual contexts, but they have visual fidelity differences from real-world data, limiting their direct use with algorithms trained on such data. While reconstruction-based methods [21, 26, 29] and frameworks such as UniSim [52], and MARS [47] capture high-fidelity scenes, they are limited by the scope of the data used, restricting their ability to simulate varied and long-tail scenarios. DriveArena [51] + +recently advanced generative closed-loop simulation by using 2D traffic sketches to control visual outputs, integrating well with vision-based driving agents. However, reliance on 2D inputs constrains its modeling granularity and visual realism. Our approach overcomes these challenges by leveraging 4D occupancy data, generating visually coherent and geometrically accurate scenes that rigorously test driving agents in closed-loop simulations, equipping them to handle complex real-world conditions effectively. + +# 3. Methodology + +DrivingSphere acts as a generative closed-loop simulation framework that combines occupancy-based 4D world modeling with advanced video rendering techniques, providing high-fidelity visual outputs that enhance both simulation realism and agent-environment interplay in autonomous driving scenarios. Specifically, DrivingSphere begins with Dynamic Environment Composition (Sec. 3.1), which generates static background from a map sketch, selects traffic participants from the actor bank, and updates the actor location, to formulate a 4D driving environment with a format of occupancy. Next, Visual Scene Synthesis (Sec. 3.2) conditions the occupancy data surrounding the ego vehicle, accurately capturing occlusion relationships and fine-grained semantic information to generate high-fidelity multi-view videos. Finally, the closed-loop feedback mechanism (Sec. 3.3) enables dynamic, responsive adjustments, where autonomous agents continuously receive updated visual data and generate control signals that modify the simulation environment, providing a comprehensive platform for algorithm testing and refinement. The overall framework is shown in Fig. 2. + +# 3.1. Dynamic Environment Composition + +Previous driving simulation methods have often overlooked static and varied elements like buildings, obstacles, and vegetation [10, 51]. Although these elements are not direct traffic participants, they are part of the perception input of AD systems, influencing final driving decision-making. For instance, static objects like buildings may block sensors' line-of-sight, causing occlusions of other vehicles or pedestrians. Obstacles or vegetation can introduce sensor artifacts or false positives for AD systems. Both situations may result in incorrect decisions like unnecessary braking or lane changes. To address this issue, we propose a Dynamic Environment Composition to construct a comprehensive 4D Driving World equipping complex digital assets. + +4D Driving World Representation. Our 4D driving world representation includes three critical digital assets: static background scenes, dynamic foreground actors, and spatial position of these actors. Formally, we define this representation as $\pmb{W} = \{\pmb{S}_{\mathrm{city}},(\pmb{A}_1,\pmb{P}_1),\dots ,(\pmb{A}_N,\pmb{P}_N)\}_{n = 1}^N$ . Here, $S_{\mathrm{city}}$ is a composite of multiple regional static scene $\{S_1,\ldots ,S_K\}$ . Each $S_{k}\in \mathbb{R}^{H\times W\times D}$ is the $k^{th}$ static back + +![](images/c0a17a0c91e0ce25c99b5a8dd7ade49231f8cd6eae4e025749f89be402c165a2.jpg) +Figure 2. Overview of DrivingSphere framework. (a) The Dynamic Environment Composition module builds a 4D driving world, simulating real driving scenarios with backgrounds generated by OccDreamer, dynamic actors from an actor bank, and trajectories guided by a transportation simulator. (b) The Visual Scene Synthesis produces high-fidelity, photo-realistic video frames conditioned on global semantics, view-specific details, and scene prompts, supporting close-loop evaluation. Control signals enable adaptive feedback for driving agents, facilitating continuous testing and evaluation of driving algorithms in the simulated environment. + +ground scene that captures the spatial layout and static elements within the region. $A_{n}\in \mathbb{R}^{4}$ is the $n^{th}$ actor of $N$ actors, such as vehicles and pedestrians, defined by a 3D coordinate and a semantic label. It is worth noticing that both the static scene $\pmb{S}_{k}$ and dynamic actor $A_{n}$ are represented as occupancy grids. $P_{n} = \left\{P_{n}^{0},\dots ,P_{n}^{T}\right\}_{t = 0}^{T}$ describe actor position across $T$ frames, where $P_{n}^{t}\in \mathbb{R}^{4}$ consists of a 3D relative position of the $n$ -th actor at time $t$ with respect to the ego vehicle and a yaw rotation $\theta$ . In the following, we will detail how to generate three assets. + +Static Scene Generation. A straightforward approach generating static 3D scenes $S_{\mathrm{city}}$ is directly using ground truth occupancy data from existing datasets, like the Boston area in nuScenes [3]. However, this method is limited to specific regions captured during data collection, restricting its applicability to other urban areas. To address this limitation, we propose an occupancy diffusion model, OccDreamer, which is conditioned on BEV maps and textual descriptions, capable of generating static scenes for any desired urban region. The framework of OccDreamer, which integrates the following components, is illustrated in Fig. 3. + +Firstly, to efficiently train the diffusion model while addressing the computational complexity of processing 3D data, we adopt VQVAE [17, 35, 40] as an occupancy tokenizer, mapping $S_{k}$ to a latent feature $Z^{S}$ : $Z^{S} = \mathcal{F}_{\mathrm{VAE,Enc}}^{occ}(S_{k})$ . The reconstructed scene is defined as $S_{k}^{\prime} = \mathcal{F}_{\mathrm{VAE,Dec}}^{occ}(Z^{S_{k}})$ . $\mathcal{F}_{\mathrm{VAE}}^{occ}$ is trained with a combined loss $\mathcal{L}_{occ}$ : + +$$ +\mathcal {L} _ {o c c} = \mathcal {L} _ {C E} \left(\boldsymbol {S} _ {k}, \boldsymbol {S} _ {k} ^ {\prime}\right) + \alpha \mathcal {L} _ {L o v} \left(\boldsymbol {S} _ {k}, \boldsymbol {S} _ {k} ^ {\prime}\right) + \beta \mathcal {L} _ {e m b}, \tag {1} +$$ + +where $\mathcal{L}_{CE}$ is the CE loss and $\mathcal{L}_{Lov}$ is the Lovász loss [1]. + +Secondly, considering varying road structures and complex regional topographies, we propose a controllable re + +gion occupancy generation module. It accepts a regional BEV map capturing the road structure and a language prompt describing abstract features (like suburban areas with rich vegetation or commercial districts with buildings lining both sides) as input, and outputs a region occupancy $S_{k}$ . Following the principles of diffusion models [17, 55], a CLIP encoder [31] converts the text prompt into an embedding $F_{\mathrm{region}}$ . Then, $F_{\mathrm{region}}$ is injected into the denoiser $\epsilon_{\theta}^{s}$ via cross-attention. Simultaneously, the BEV map is processed by a pre-trained image VAE [17] to extract the corresponding road embedding $F_{M}$ , serving as input for the ControlNet branch $\epsilon_{\phi}^{s}$ . This enables precise control over the diffusion learning process in the latent space: + +$$ +\mathcal {L} _ {\text {o c c}} = \mathbb {E} _ {\boldsymbol {F} _ {M}, \boldsymbol {F} _ {\text {r e g i o n}}, \boldsymbol {Z}} s _ {k, \epsilon^ {s}, \tau} \left[ \left\| \epsilon^ {s} - \epsilon_ {\theta} ^ {s} \left(\boldsymbol {Z} _ {\tau} ^ {\boldsymbol {S} _ {k}}, \boldsymbol {F} _ {\text {r e g i o n}}, \tau , \epsilon_ {\phi} ^ {s} (\boldsymbol {F} _ {M})\right) \right\| ^ {2} \right], \tag {2} +$$ + +where $Z_{\tau}^{S_k}$ is the noisy version of $Z^{S_k}$ by adding $\tau$ steps noise $\epsilon^s$ . After training, we use OccDreamer to create $Z_{k}^{S}$ and decode it with $\mathcal{F}_{\mathrm{VAE,Decode}}^{\mathrm{occ}}$ to obtain $S_{k}$ . + +Thirdly, for spatial consistency across the entire 3D scene, we propose a Scene Expansion Mechanism for building a coherent and city-level static scene $S_{\mathrm{city}}$ . Its core idea is to expand an initial region $S_{k}$ to a neighboring region $S_{k+1}$ by using the overlap between these regions as a conditioning constraint. Formally, let $O$ represent a binary mask that defines the overlapping area between two adjacent regions $S_{k}$ and $S_{k+1}$ . This overlap serves as the conditional context for generating the next region, enforcing spatial coherence at the boundary. To generate $S_{k+1}$ , we first create a partially masked scene $S_{k}^{\mathrm{partial}}$ by applying $O$ to $S_{k}$ : + +$$ +\boldsymbol {S} _ {k} ^ {\text {p a r t i a l}} = \boldsymbol {S} _ {k} \odot \boldsymbol {O}, \tag {3} +$$ + +where $\odot$ denotes element-wise multiplication, extracting + +![](images/b69dc1588143795e3d83e491b935aa255de9e42f0321e925119a0a0e0aa2d57d.jpg) +Figure 3. The framework of OccDreamer, which includes Occupancy Tokenizer, Region Occupancy Generation, and Scene Extensio to generate a city-level background for 4D driving world. + +the overlapping section of $S_{k}$ . We then employ a diffusion process to generate the neighboring region $S_{k + 1}$ , conditioning on both the BEV map $M_{k + 1}$ (capturing the road structure of the target area) and the partially masked scene $S_{k}^{\mathrm{partial}}$ . This diffusion process is represented as: + +$$ +\boldsymbol {Z} ^ {\boldsymbol {S} _ {k + 1}} = \mathcal {F} _ {\text {o u t p a i n t i n g}} \left(\epsilon^ {s}, \boldsymbol {M} _ {k + 1}, \boldsymbol {S} _ {k} ^ {\text {p a r t i a l}}\right), \tag {4} +$$ + +where $\epsilon^s$ is the input noise, and $\mathcal{F}_{\mathrm{outpainting}}$ denotes the diffusion model conditioned on the masked scene and BEV map. After obtaining the latent representation $Z^{S_{k + 1}}$ , we decode it using the occupancy VAE decoder, resulting in the expanded region $S_{k + 1}$ . The two regions are then combined into a larger scene $S_{\mathrm{merged}}$ through a merging operation: + +$$ +\boldsymbol {S} _ {\text {m e r g e d}} = \operatorname {M e r g e} \left(\boldsymbol {S} _ {k} \odot (1 - \boldsymbol {O}), \boldsymbol {S} _ {k + 1}\right), \tag {5} +$$ + +where Merge ensures spatial continuity and logical consistency between adjacent regions. The spatially continuous and logically consistent city scene $S_{\mathrm{city}}$ serves as an excellent digital background for the 4D driving world. + +Dynamic Actor Selection. To complement static scenes, we populate the 4D driving world with dynamic actors, creating realistic traffic flows. An actor bank $\mathcal{B}$ is constructed, where each actor $\mathbf{A}_n \in \mathcal{B}$ consists of instance-level voxel data segmented from real or synthetic occupancy scenes, with unique geometric, semantic, and behavioral attributes, as well as a descriptive caption $L_{\mathbf{A}_n}$ and an ID $I_{\mathbf{A}_n}$ to ensure consistency across frames. Actors are selected based on semantic similarity to user-provided descriptions via CLIP, or randomly sampled from relevant categories if unspecified, ensuring contextual diversity. This flexible selection process allows for dynamically relevant and diverse actor integration within the 4D driving world, supporting realistic and adaptable traffic simulation. + +Spatial Position of Actors. The generation of actor positions $\pmb{P}$ is driven by the control signals output from autonomous agents (Sec. 3.3). Each actor is assigned an control signal set $C_n = \{c_n^0, \dots, c_n^T\}$ , where $c_n^t$ represents the received control signal of actor $n$ at time $t$ . These signals iteratively update the dynamic state of the world. For $n$ -th actor, we have the position: + +$$ +\boldsymbol {P} _ {n} ^ {t + 1} = \boldsymbol {P} _ {n} ^ {t} + \triangle \boldsymbol {P} _ {n} ^ {t} \left(\boldsymbol {c} _ {n} ^ {t}\right), \tag {6} +$$ + +where $\triangle P_n^t (c_N^t)$ is the positional change resulting from the control signal $c_{N}^{t}$ at time $t$ + +4D World Composition. With the static background, dynamic foreground actors and their positions calculated, we integrate them into the comprehensive 4D driving world. The world state at any time $t$ is represented by: + +$$ +\boldsymbol {W} ^ {t} = \pi (\boldsymbol {S}, \left(\boldsymbol {A} _ {n}, \boldsymbol {c} _ {n} ^ {t}\right) _ {n = 1} ^ {N}), \tag {7} +$$ + +where $\pi$ is the composition operator that aligns each actor based on their current pose $P_{n}^{t}$ , represents the set of poses for all actors, forming the complete 4D driving world at time $t$ . By iterating this process over time, we construct the dynamic 4D world that accounts for actor interplay, occlusions, and the evolving scene geometry. + +# 3.2. Visual Scene Synthesis + +Previous generative models [13, 28, 43, 57] tend to employ 2D visual conditions, which fail to accurately capture the geometric and semantic complexities inherent in real-world driving scenes. In response, our Visual Scene Synthesis employs VideoDreamer, a custom-designed video diffusion model, to transform the occupancy-driven 4D driving world constructed in the previous section into high-fidelity visual results. The overall framework is shown in Fig. 4. Specifically, we introduce a Dual-path Condition Encoding strategy that focuses on encoding occupancy data as its primary condition. We further enhance the consistency of appearances across views and frames by developing an ID-aware Actor Encoding method. Finally, we integrate the Spatial-Temporal Diffusion Transformer (ST-DiT) from OpenSora [60] as the foundational technology to ensure visual consistency and generate artifact-free frames. + +Dual-path Condition Encoding. We design a dual-path condition encoding strategy aimed at effectively capturing occupancy data. Given a set of driving world data $\mathbf{W} = \{\mathbf{W}^0, \mathbf{W}^1, \dots, \mathbf{W}^T\}$ spanning frames from 0 up to $T$ , we first encode them into a global feature $\mathbf{F}_{\mathrm{global}} = \mathcal{F}_{\mathrm{VAE}}^{\mathrm{4Docc}}(\mathbf{W})$ , capturing comprehensive geometric information and the spatial-temporal relationships within the scene. Here, $\mathcal{F}_{\mathrm{VAE}}^{\mathrm{4Docc}}$ is the well-trained 4D encoder from occupancy VAE [41], responsible for capturing global geometry of the scene. This ensures that the overall structure and spatial layout of the scene are accurately represented. + +Meanwhile, for $t^{th}$ frame (where $t = 0,\dots,T$ ), 3D occupancy data $\mathbf{W}^t$ is projected into 2D semantic maps $\mathcal{M}_v^t$ for + +![](images/e45d3593a899356336edcaa9f22d52370655dd9052d2eebc971e830c92a63463.jpg) +Figure 4. Overview of the VideoDreamer. The model conditions on 4D driving world and enriched actor embeddings (e.g., actor ID, position and caption). This ensures high visual fidelity and geometric consistency in the generated driving simulations. Each view $v$ by camera extrinsics $\mathbf{T}^t$ and intrinsics $\mathbf{K}_v$ : + +$$ +\mathcal {M} _ {v} ^ {t} = \mathcal {Q} \left(\boldsymbol {W} ^ {t}, \mathbf {K} _ {v}, \mathbf {T} ^ {t}\right), \tag {8} +$$ + +where $\mathcal{Q}$ is the render and projection function [33]. This process yields a sequence of semantic maps, $\mathcal{M} \in \mathbb{R}^{T \times 6 \times H \times W}$ , with precise pixel-to-pixel alignment across views. Such alignment is essential for accurately capturing view-specific occlusions and depth variations. $\mathcal{M}$ is then encoded into view-specific features $F_{\mathrm{view}} = \mathcal{F}_{\mathrm{VAE}}^{\mathrm{Image}}(\mathcal{M})$ . $F_{\mathrm{view}}$ and $F_{\mathrm{global}}$ are then integrated through cross-attention to obtain the unified representation $F_{\mathrm{occ}}$ . + +ID-aware Actor Encoding. To ensure coherence of each actor's appearance and identity information across the scene, we extract a fused sequence embedding $F_{\mathrm{fuse}}$ : + +$$ +\boldsymbol {F} _ {\text {f u s e}} = \operatorname {C O N C A T} \left[ \boldsymbol {F} _ {\boldsymbol {W}}, \boldsymbol {F} _ {\boldsymbol {A} _ {1}}, \dots , \boldsymbol {F} _ {\boldsymbol {A} _ {N}} \right], \tag {9} +$$ + +where $F_W$ is the embedding of the scene caption $\mathcal{L}_{\mathrm{scene}}$ , derived from T5 [32]. Each actor embedding $F_{A_n}$ is: + +$$ +\boldsymbol {F} _ {\boldsymbol {A} _ {n}} = \operatorname {C O N C A T} \left[ \mathcal {F} _ {\text {F o u r i e r}} \left(\boldsymbol {P} _ {n}\right), \mathcal {F} _ {\text {F o u r i e r}} \left(\boldsymbol {I} _ {n}\right), \mathcal {F} _ {T 5} \left(\boldsymbol {L} _ {\boldsymbol {A} _ {i}}\right) \right]. \tag {10} +$$ + +Here, $P_{n}$ represents the position of the $n$ -th actor and $I_{n}$ is its unique ID. Both are encoded using Fourier encoding [29]. The actor caption $L_{A_i}$ is encoded using the T5 [32]. Overall, by integrating geometry positions, textual descriptions, and identity information, we ensure coherent behavior and long-term visual consistency across frames. + +ControlNet-DiT. To enhance visual fidelity and temporal consistency in generated videos, we integrate the ST-DiT [60] as our denoiser $\epsilon_{\theta}^{v}$ , utilizing ST-DiT blocks stacked with View-aware Spatial Self-Attention (VSSA), Temporal Self-Attention, Cross-Attention, and FFN. + +In the video generation process, the input video sequence $\mathbf{V} \in \mathbb{R}^{v \times T \times H \times W}$ is first encoded into latent features $Z^{\mathbf{V}} \in \mathbb{R}^{v \times T \times H' \times W' \times C}$ . To ensure multi-view consistency while managing computational complexity, we mod- + +![](images/0045c73d182897ed8245df6fd9a28d3bf755bf90f20c45357484a7fd1c0013fc.jpg) + +![](images/77719c565930512c098d4fa8afe8ad5394909ae1e69dc07f626950a3e246d90a.jpg) + +![](images/857accd1b061e6fc3e1ffbc5c3c56509b22384a59361029f5893695d78415640.jpg) +Figure 5. Qualitative results of generated 3D scene. OccDreamer uses text prompt and bev map as guidance to generate 3D scenes with controllable regional content and road structure. + +if the original Spatial self-attention [60] as VSSA. The latent feature $Z^{\mathbf{V}}$ is reshaped from $v\times T\times H^{\prime}\times W^{\prime}\times C$ to $T\times (vH^{\prime}W^{\prime})\times C$ , treating the combined view, height, and width dimensions $(vH^{\prime}W^{\prime})$ as the sequence length. This view-flatted approach reduces the parameter overhead associated with cross-view attention [13, 44] while maintaining spatial coherence across different views. For external conditions, we apply a cross-attention layer that injects the fused sequence embedding $F_{\mathrm{fuse}}$ (derived from scene context and actor identities) as: $Z^{V} = \mathrm{CrossAttention}(Z^{V},F_{\mathrm{fuse}})$ . Last but not least, we introduce ControlNet branch $\epsilon_{\phi}^{v}$ to ensure the occupancy control, inspired by PixArt-δ [4]. This branch duplicates blocks from $\epsilon_{\theta}^{v}$ and takes $F_{\mathrm{occ}}$ as the input. The final diffusion process is: + +$$ +\mathcal {L} _ {\text {v i d e o}} = \mathbb {E} _ {\boldsymbol {F} _ {\text {o c c}}, \boldsymbol {F} _ {\text {f u s e}}, \boldsymbol {Z} V, \epsilon^ {v}, \tau} \left[ \left\| \epsilon^ {v} - \epsilon_ {\theta} ^ {v} \left(\boldsymbol {Z} _ {\tau} ^ {\boldsymbol {V}}, \boldsymbol {F} _ {\text {f u s e}}, \tau , \epsilon_ {\phi} ^ {v} \left(\boldsymbol {F} _ {\text {o c c}}\right)\right) \right\| ^ {2} \right]. \tag {11} +$$ + +This approach ensures fine-grained control over spatial and temporal coherence, producing artifact-free frames that meet the high fidelity requirements for AD simulation. + +Auto-regressive Generation. To generate temporally consistent and content-coherent video frames, we employ an auto-regressive generation strategy [60] based on first- $f$ frame mask. This approach models the dependency of each video frame on the previous frames, ensuring smooth transitions and maintaining visual continuity over time: + +$$ +\boldsymbol {V} ^ {t} = \mathcal {F} _ {\text {v i d e o}} (\boldsymbol {V} ^ {t - 1}, \boldsymbol {Z} ^ {V}, t), +$$ + +where $V^t$ is the video frame at time $t$ , and $\mathcal{F}_{\mathrm{video}}$ is the video generation function that conditions on the latent representation $Z^{V}$ and the previous frame $V^{t-1}$ . This ensures temporal coherence and infinite video sequences. + +# 3.3. Agent Interplay and Closed-Loop Simulation + +We enable seamless coordination of autonomous driving agents within the DrivingSphere simulation environment, categorizing agents into two primary types: the Ego Agent and Environment Agents. + +Ego Agent: The Ego Agent represents the autonomous driving system under evaluation. Driven by end-to-end + +models [6, 19, 20], it receives visual input frames $V^{t}$ and outputs the predicted control signal $c_{\mathrm{ego}}^{t + 1}$ at each time $t$ : + +$$ +\boldsymbol {c} _ {\text {e g o}} ^ {t + 1} = \mathcal {F} _ {\text {d r i v i n g}} \left(\boldsymbol {V} ^ {t}\right), \tag {12} +$$ + +where $\mathcal{F}_{\mathrm{driving}}$ denotes the driving algorithm [6, 19, 20]. + +Environment Agents: Environment Agents are responsible for controlling the behavior and actions of all other actors within the simulation world $W$ . To achieve realistic info exchanges, we utilize the traffic flow engine [46], which supports multi-agent simulation. Environment Agents receive inputs from the simulation's state and output control signal $c_{\mathrm{n}}^{t + 1}$ that dictate the movements and interplay of actors. Formally, this process can be described as: + +$$ +\boldsymbol {c} _ {\mathrm {n}} ^ {t + 1} = \mathcal {F} _ {\text {s y s t e m}} (\boldsymbol {W} ^ {t}), \tag {13} +$$ + +where $\mathcal{F}_{\mathrm{system}}$ denotes the transportation simulator. + +These agents take as input the visual frames $V^t$ or previous state $\pmb{W}^{t}$ , and output actors' respective control signals $\pmb{c}_{\mathrm{ego}}^{t + 1}$ and $\pmb{c}_{\mathrm{n}}^{t + 1}$ . These signals influence the actor positions using Eq. (6) and further dynamically update the 4D driving world $\pmb{W}^{t + 1}$ , maintaining the closed-loop simulation where the environment continuously adapts to the agents' actions. + +# 4. Experiments + +# 4.1. Settings + +Dataset and Evaluation Metrics. We utilize real-world driving datasets nuScenes [3]. These datasets provide rich sensor data, including LiDAR point clouds, camera images, and high-definition maps, which are essential for generating accurate occupancy grids and semantic information. We employ the following metrics to evaluate the performance of our simulation framework: FVD and UniAD-based metrics [19] for visual quality and fidelity. As for the evaluation of OccDreamer, we use FID and MMD [22]. In open-loop evaluation, we borrow metrics from NAVSIM [8]. RC and ADS for the closed-loop following [51]. + +Implement Details are included in the supplementary. + +# 4.2. Visual Fidelity and Realism + +We first evaluate the fidelity of the voxel-based scenes from OccDreamer and video sequences from VideoDreamer against real nuScenes data to assess the domain gap. + +Scene Generation. We conduct quantitative and qualitative analyses to assess the fidelity of OccDreamer's generated occupancy data. Our method outperforms existing approaches quantitatively, as shown in Tab. 2. Specifically, SemCity employs a VAE trained on three planes to generate occupancy data while OccDreamer achieves pixel-to-pixel alignment based on BEV maps, resulting in more robust and accurate occupancy generation. Qualitatively, Fig. 5 presents the results of OccDreamer with both caption and + +BEV map. To the best of our knowledge, our approach is the first to achieve text-guided occupancy generation. As demonstrated, abstract textual descriptions effectively guide the model to generate diverse scenes, including outdoor environments and urban landscapes. Under BEV map control, the generated samples accurately reflect road structures and lane configurations. These capabilities establish a solid foundation for the comprehensive modeling of the 4D driving world, ensuring that the generated scenes are both spatially continuous and logically consistent. More visualizations are provided in the supplementary materials. + +Video Generation. We further assess the performance of our video generation model, VideoDreamer, in generating coherent video. We conduct a comparative analysis against leading models such as MagicDrive [13], DriveArena [51], and Panacea [45]. Using UniAD [19] as the evaluation framework, we assess the performance across crucial metrics including 3D object detection, BEV map segmentation, and trajectory planning. Our experimental results, summarized in Tab. 1, indicate that VideoDreamer significantly outperforms the baseline methods in these key metrics. This performance underscores the advantages of occupancy data over traditional 2D conditions, affirming VideoDreamer's effectiveness in creating realistic, dynamic scenes that accurately simulate real-world driving complexities. + +Qualitatively, as shown in Fig. 6, our approach demonstrates markedly superior performance in maintaining coherence and fidelity over time, significantly outperforming others. This success is primarily due to the geometry-aware occupancy encoding and instance encoding that ensures appearance consistency, highlighting VideoDreamer's robustness in creating visually consistent, detailed driving environments that capture real-world scene dynamics. More visualizations are included in the supplementary materials. + +# 4.3. Open-loop Evaluation + +In the open-loop setting, where AD agents passively receive environment input without influencing simulation dynamics, we assess the performance of UniAD [19] on DrivingSphere and alternative simulators. Using 100 predefined trajectories, Tab. 3 demonstrates better PDMS scores achieved by AD agents in DrivingSphere, suggesting that its higher visual fidelity reduces perceptual inaccuracies that can misguide agent decisions. This improvement highlights DrivingSphere's suitability for testing AD systems in environments that closely emulate real-world conditions. + +# 4.4. Closed-loop Evaluation + +In the closed-loop evaluation, AD agents receive visual input and output control signals, thus interactively shaping the simulation. This evaluation setup involves 100 pre-defined trajectories to test across controlled yet varied scenarios. As shown in Tab. 4, UniAD [19] exhibits relatively low Route + +
Source DataQuality3DODSegmentationL2(↓)
FVD(↓)mAP(↑)NDS(↑)Lanes(↑)Drivable(↑)Divider(↑)1.0s2.0s3.0s
ori nuScenes[3]—37.9849.8531.3169.1425.930.510.981.65
MagicDrive [13]218.1212.9228.3621.9551.4617.100.571.141.95
Panacea [45]13913.7227.7318.2352.3717.210.581.141.97
DriveArena [51]185.3216.0630.0326.1459.3720.790.561.101.89
DrivingSphere103.4221.4534.1627.9962.8722.290.541.101.76
+ +![](images/b7b56650c7a3f8988505033fe7f558153600651b2314a532f9834720382bb0c1.jpg) +Figure 6. Comparison of generated driving simulations from (a) Panacea [45], (b) Our method, and (c) Real Scene [3] across multiple views and time steps. Red boxes highlight instance-level appearance consistency, with our method maintaining more coherent actor appearances across frames. Yellow boxes illustrate temporal consistency, where our approach achieves smoother transitions over time. Green boxes and background elements demonstrate our model's capability to capture and accurately model non-direct traffic participants, such as buildings and other environmental features, enhancing realism in the generated scene. + +Table 1. Quantitative comparison of video generation fidelity for autonomous driving simulations on the nuScenes validation set. Using UniAD [19] for evaluation on generated data, DrivingSphere achieves high fidelity in video generation, closely matching real-world data across key perception and planning metrics. + +
MethodsFID(↓)MMD(↓)
Semicity [22]6340.251
Ours2740.082
+ +Table 2. Quantitative comparison of scene generation performance, following the setup in SemCity. The scene data generated by our method is rendered into 2D images. + +
Data SourceNC(↑)DAC(↑)TTC(↑)PDMS(↑)
nuScenes [3]0.9930.9950.9470.910
DriveArena [51]0.7920.9420.7710.636
DriveArena* [51]0.8290.9640.8120.698
DrivingSphere0.8520.9680.8530.742
+ +Completion (RC) scores, with completing an average of only $11.7\%$ of each route. Compared to DriveArena [51], our simulation yields consistently better performance metrics, and, along with the superior visual fidelity demonstrated in Tab. 1, affirms DrivingSphere's effectiveness as a high-fidelity platform for AD algorithm validation. + +Table 3. Open-loop evaluation of UniAD [19] on original nuScenes sequences, DriveArena's reported results, and our predefined 100-trajectory test for DriveArena* and DrivingSphere. + +
RoutePDMS(↑)RC(↑)ADS(↑)
DriveArena [51]0.69010.06410.0508
DrivingSphere0.72810.11700.0851
+ +Table 4. Closed-loop evaluation of UniAD's [19] performance on 100 pre-defined routes. + +# 5. Conclusion + +We presented *DrivingSphere*, a generative closed-loop simulation framework that bridges the gap between traditional closed-loop simulations and open-loop generative models. Through advanced occupancy-based modeling and controllable generation mechanisms, *DrivingSphere* creates realistic, high-fidelity simulation for autonomous driving. Our experiments demonstrate superior visual quality, temporal consistency, and the ability to effectively test AD algorithms in dynamic environments. In future work, we aim to enhance dynamic agent behaviors and incorporate more complex real-world scenarios to further improve the robustness of our simulations. + +# References + +[1] Maxim Berman, Amal Rannen Triki, and Matthew B Blaschko. The lovasz-softmax loss: A tractable surrogate for the optimization of the intersection-over-union measure in neural networks. In CVPR, 2018. 4 +[2] Andreas Blattmann, Tim Dockhorn, Sumith Kulal, Daniel Mendelevitch, Maciej Kilian, Dominik Lorenz, Yam Levi, Zion English, Vikram Voleti, Adam Letts, et al. Stable video diffusion: Scaling latent video diffusion models to large datasets. arXiv preprint arXiv:2311.15127, 2023. 3 +[3] Holger Caesar, Varun Bankiti, Alex H. Lang, Sourabh Vora, Venice Erin Liong, Qiang Xu, Anush Krishnan, Yu Pan, Giancarlo Baldan, and Oscar Beijbom. nuscenes: A multimodal dataset for autonomous driving. In CVPR, 2020. 1, 2, 4, 7, 8 +[4] Junsong Chen, Yue Wu, Simian Luo, Enze Xie, Sayak Paul, Ping Luo, Hang Zhao, and Zhenguo Li. Pixart-{\delta} : Fast and controllable image generation with latent consistency models. arXiv preprint arXiv:2401.05252, 2024. 6 +[5] Li Chen, Penghao Wu, Kashyap Chitta, Bernhard Jaeger, Andreas Geiger, and Hongyang Li. End-to-end autonomous driving: Challenges and frontiers. T-PAMI, 2024. 1 +[6] Shaoyu Chen, Bo Jiang, Hao Gao, Bencheng Liao, Qing Xu, Qian Zhang, Chang Huang, Wenyu Liu, and Xinggang Wang. Vadv2: End-to-end vectorized autonomous driving via probabilistic planning. arXiv preprint arXiv:2402.13243, 2024. 2, 7 +[7] Yuan Chen, Zi-han Ding, Ziqin Wang, Yan Wang, Lijun Zhang, and Si Liu. Asynchronous large language model enhanced planner for autonomous driving. In ECCV, 2025. 1 +[8] Daniel Dauner, Marcel Hallgarten, Tianyu Li, Xinshuo Weng, Zhiyu Huang, Zetong Yang, Hongyang Li, Igor Gilitschenski, Boris Ivanovic, Marco Pavone, Andreas Geiger, and Kashyap Chitta. Navsim: Data-driven non-reactive autonomous vehicle simulation and benchmarking. In NeurIPS, 2024. 7, 1 +[9] Wenhao Ding, Yulong Cao, Ding Zhao, Chaowei Xiao, and Marco Pavone. Realgen: Retrieval augmented generation for controllable traffic scenarios. arXiv preprint arXiv:2312.13303, 2023. 3 +[10] Alexey Dosovitskiy, German Ros, Felipe Codevilla, Antonio Lopez, and Vladlen Koltun. Carla: An open urban driving simulator. In CoRL, 2017. 1, 2, 3 +[11] Lan Feng, Quanyi Li, Zhenghao Peng, Shuhan Tan, and Bolei Zhou. Trafficgen: Learning to generate diverse and realistic traffic scenarios. In ICRA, 2023. 3 +[12] Daocheng Fu, Wenjie Lei, Licheng Wen, Pinlong Cai, Song Mao, Min Dou, Botian Shi, and Yu Qiao. Limsim++: A closed-loop platform for deploying multimodal llms in autonomous driving. arXiv preprint arXiv:2402.01246, 2024. 1, 3 +[13] Ruiyuan Gao, Kai Chen, Enze Xie, Lanqing Hong, Zhenguo Li, Dit-Yan Yeung, and Qiang Xu. Magicdrive: Street view generation with diverse 3d geometry control. arXiv preprint arXiv:2310.02601, 2023. 1, 2, 3, 5, 6, 7, 8 +[14] Shenyuan Gao, Jiazhi Yang, Li Chen, Kashyap Chitta, Yihang Qiu, Andreas Geiger, Jun Zhang, and Hongyang Li. + +Vista: A generalizable driving world model with high fidelity and versatile controllability. In NeurIPS, 2024. 3 +[15] Wencheng Han, Dongqian Guo, Cheng-Zhong Xu, and Jianbing Shen. Dme-driver: Integrating human decision logic and 3d scene perception in autonomous driving. arXiv preprint arXiv:2401.03641, 2024. 1 +[16] Yingqing He, Tianyu Yang, Yong Zhang, Ying Shan, and Qifeng Chen. Latent video diffusion models for high-fidelity long video generation. arXiv preprint arXiv:2211.13221, 2022. 3, 1 +[17] Jonathan Ho, Ajay Jain, and Pieter Abbeel. Denoising diffusion probabilistic models. NeurIPS, 2020. 3, 4, 1 +[18] Qianjiang Hu, Zhimin Zhang, and Wei Hu. Rangeldm: Fast realistic lidar point cloud generation. In ECCV, 2025. 3 +[19] Yihan Hu, Jiazhi Yang, Li Chen, Keyu Li, Chonghao Sima, Xizhou Zhu, Siqi Chai, Senyao Du, Tianwei Lin, Wenhai Wang, et al. Planning-oriented autonomous driving. In CVPR, 2023. 1, 2, 7, 8 +[20] Bo Jiang, Shaoyu Chen, Qing Xu, Bencheng Liao, Jiajie Chen, Helong Zhou, Qian Zhang, Wenyu Liu, Chang Huang, and Xinggang Wang. Vad: Vectorized scene representation for efficient autonomous driving. In ICCV, 2023. 1, 2, 7 +[21] Bernhard Kerbl, Georgios Kopanas, Thomas Leimkuhler, and George Drettakis. 3d gaussian splatting for real-time radiance field rendering. ACM Trans. Graph., 2023. 3 +[22] Jumin Lee, Sebin Lee, Changho Jo, Woobin Im, Juhyeong Seon, and Sung-Eui Yoon. Semcity: Semantic scene generation with triplane diffusion. In CVPR, 2024. 3, 7, 8, 2 +[23] Quanyi Li, Zhenghao Peng, Lan Feng, Qihang Zhang, Zhenghai Xue, and Bolei Zhou. Metadrive: Composing diverse driving scenarios for generalizable reinforcement learning. T-PAMI, 2022. 1, 2, 3 +[24] Zhiqi Li, Zhiding Yu, Shiyi Lan, Jiahan Li, Jan Kautz, Tong Lu, and Jose M Alvarez. Is ego status all you need for open-loop end-to-end autonomous driving? In CVPR, 2024. 1 +[25] Zhiqi Li, Zhiding Yu, Shiyi Lan, Jiahan Li, Jan Kautz, Tong Lu, and Jose M Alvarez. Is ego status all you need for open-loop end-to-end autonomous driving? In CVPR, 2024. 1 +[26] William Ljungbergh, Adam Tonderski, Joakim Johnander, Holger Caesar, Kalle Åström, Michael Felsberg, and Christoffer Petersson. Neuroncap: Photorealistic closed-loop safety testing for autonomous driving. In ECCV, 2025. 3 +[27] Pablo Alvarez Lopez, Michael Behrisch, Laura Bieker-Walz, Jakob Erdmann, Yun-Pang Flotteröd, Robert Hilbrich, Leonhard Lücken, Johannes Rummel, Peter Wagner, and Evamarie Wießner. Microscopic traffic simulation using sumo. In IEEE ITSC, 2018. 3 +[28] Enhui Ma, Lijun Zhou, Tao Tang, Zhan Zhang, Dong Han, Junpeng Jiang, Kun Zhan, Peng Jia, Xianpeng Lang, Haiyang Sun, et al. Unleashing generalization of end-to-end autonomous driving with controllable long video generation. arXiv preprint arXiv:2406.01349, 2024. 5, 1 +[29] Ben Mildenhall, Pratul P Srinivasan, Matthew Tancik, Jonathan T Barron, Ravi Ramamoorthi, and Ren Ng. Nerf: Representing scenes as neural radiance fields for view synthesis. Communications of the ACM, 2021. 3, 6 + +[30] Kazuto Nakashima and Ryo Kurazume. Lidar data synthesis with denoising diffusion probabilistic models. In ICRA, 2024. 3 +[31] Alec Radford, Jong Wook Kim, Chris Hallacy, Aditya Ramesh, Gabriel Goh, Sandhini Agarwal, Girish Sastry, Amanda Askell, Pamela Mishkin, Jack Clark, et al. Learning transferable visual models from natural language supervision. In ICML, 2021. 4 +[32] Colin Raffel, Noam Shazeer, Adam Roberts, Katherine Lee, Sharan Narang, Michael Matena, Yanqi Zhou, Wei Li, and Peter J. Liu. Exploring the limits of transfer learning with a unified text-to-text transformer. Journal of Machine Learning Research, 2020. 6 +[33] Prabhu Ramachandran and Gaël Varoquaux. Mayavi: 3d visualization of scientific data. Computing in Science & Engineering, 2011. 6 +[34] Haoxi Ran, Vitor Guizilini, and Yue Wang. Towards realistic scene generation with lidar diffusion models. In ECCV, 2024. 3 +[35] Robin Rombach, Andreas Blattmann, Dominik Lorenz, Patrick Esser, and Björn Ommer. High-resolution image synthesis with latent diffusion models. In CVPR, 2022. 4 +[36] Hao Shao, Yuxuan Hu, Letian Wang, Guanglu Song, Steven L Waslander, Yu Liu, and Hongsheng Li. Lmdrive: Closed-loop end-to-end driving with large language models. In CVPR, 2024. 1, 2 +[37] Pei Sun, Henrik Kretzschmar, Xerxes Dotiwalla, Aurelien Chouard, Vijaysai Patnaik, Paul Tsui, James Guo, Yin Zhou, Yuning Chai, Benjamin Caine, et al. Scalability in perception for autonomous driving: Waymo open dataset. In CVPR, 2020. 1 +[38] Alexander Swerdlow, Runsheng Xu, and Bolei Zhou. Streetview image generation from a bird's-eye view layout. IEEE RA-L, 2024. 3 +[39] Shuhan Tan, Boris Ivanovic, Xinshuo Weng, Marco Pavone, and Philipp Kraehenbuehl. Language conditioned traffic generation. arXiv preprint arXiv:2307.07947, 2023. 3 +[40] Aaron Van Den Oord, Oriol Vinyals, et al. Neural discrete representation learning. NeurIPS, 2017. 4 +[41] Lening Wang, Wenzhao Zheng, Yilong Ren, Han Jiang, Zhiyong Cui, Haiyang Yu, and Jiwen Lu. Occsora: 4d occupancy generation models as world simulators for autonomous driving. arXiv preprint arXiv:2405.20337, 2024. 2, 3, 5, 1 +[42] Wenhai Wang, Jiangwei Xie, ChuanYang Hu, Haoming Zou, Jianan Fan, Wenwen Tong, Yang Wen, Silei Wu, Hanming Deng, Zhiqi Li, et al. Drivemlm: Aligning multi-modal large language models with behavioral planning states for autonomous driving. arXiv preprint arXiv:2312.09245, 2023. 1 +[43] Xiaofeng Wang, Zheng Zhu, Guan Huang, Xinze Chen, Jia-gang Zhu, and Jiwen Lu. Drivedreamer: Towards real-world-driven world models for autonomous driving. arXiv preprint arXiv:2309.09777, 2023. 1, 2, 3, 5 +[44] Yuqi Wang, Jiawei He, Lue Fan, Hongxin Li, Yuntao Chen, and Zhaoxiang Zhang. Driving into the future: Multiview visual forecasting and planning with world model for autonomous driving. In CVPR, 2024. 6, 1, 2 + +[45] Yuqing Wen, Yucheng Zhao, Yingfei Liu, Fan Jia, Yanhui Wang, Chong Luo, Chi Zhang, Tiancai Wang, Xiaoyan Sun, and Xiangyu Zhang. Panacea: Panoramic and controllable video generation for autonomous driving. In CVPR, 2024. 3, 7, 8, 1, 2 +[46] Licheng Wenl, Daocheng Fu, Song Mao, Pinlong Cai, Min Dou, Yikang Li, and Yu Qiao. Limsim: A long-term interactive multi-scenario traffic simulator. In IEEE ITSC, 2023. 1, 3, 7 +[47] Zirui Wu, Tianyu Liu, Liyi Luo, Zhide Zhong, Jianteng Chen, Hongmin Xiao, Chao Hou, Haozhe Lou, Yuantao Chen, Runyi Yang, et al. Mars: An instance-aware, modular and realistic simulator for autonomous driving. In CAAI International Conference on Artificial Intelligence, 2023. 3 +[48] Zhenhua Xu, Yujia Zhang, Enze Xie, Zhen Zhao, Yong Guo, Kwan-Yee K Wong, Zhenguo Li, and Hengshuang Zhao. Drivegpt4: Interpretable end-to-end autonomous driving via large language model. IEEE RA-L, 2024. 1 +[49] Tianyi Yan, Junbo Yin, Xianpeng Lang, Ruigang Yang, Cheng-Zhong Xu, and Jianbing Shen. Olidm: Object-aware lidar diffusion models for autonomous driving. In AAAI, 2025. 3 +[50] Kairui Yang, Enhui Ma, Jibin Peng, Qing Guo, Di Lin, and Kaicheng Yu. Bevcontrol: Accurately controlling street-view elements with multi-perspective consistency via bev sketch layout. arXiv preprint arXiv:2308.01661, 2023. 3 +[51] Xuemeng Yang, Licheng Wen, Yukai Ma, Jianbiao Mei, Xin Li, Tiantian Wei, Wenjie Lei, Daocheng Fu, Pinlong Cai, Min Dou, Botian Shi, Liang He, Yong Liu, and Yu Qiao. Drivearena: A closed-loop generative simulation platform for autonomous driving. arXiv preprint arXiv:2408.00415, 2024. 1, 3, 7, 8, 2, 5 +[52] Ze Yang, Yun Chen, Jingkang Wang, Sivabalan Manivasagam, Wei-Chiu Ma, Anqi Joyce Yang, and Raquel Urtasun. Unisim: A neural closed-loop sensor simulator. In CVPR, 2023. 3 +[53] Jimuyang Zhang, Zanming Huang, Arijit Ray, and Eshed Ohn-Bar. Feedback-guided autonomous driving. In CVPR, 2024. 1 +[54] Linrui Zhang, Zhenghao Peng, Quanyi Li, and Bolei Zhou. Cat: Closed-loop adversarial training for safe end-to-end driving. In 7th Annual Conference on Robot Learning, 2023. 1 +[55] Lvmin Zhang, Anyi Rao, and Maneesh Agrawala. Adding conditional control to text-to-image diffusion models. In ICCV, 2023. 4 +[56] Yuheng Zhang, Tianjian Ouyang, Fudan Yu, Cong Ma, Lei Qiao, Wei Wu, Jian Yuan, and Yong Li. Lcsim: A large-scale controllable traffic simulator. arXiv preprint arXiv:2406.19781, 2024. 1, 3 +[57] Guosheng Zhao, Xiaofeng Wang, Zheng Zhu, Xinze Chen, Guan Huang, Xiaoyi Bao, and Xingang Wang. Drivedreamer-2: Llm-enhanced world models for diverse driving video generation. arXiv preprint arXiv:2403.06845, 2024. 1, 2, 3, 5 +[58] Wenzhao Zheng, Weiliang Chen, Yuanhui Huang, Borui Zhang, Yueqi Duan, and Jiwen Lu. Occworld: Learning a 3d + +occupancy world model for autonomous driving. In ECCV, 2024.3,1,2 +[59] Wenzhao Zheng, Ruiqi Song, Xianda Guo, Chenming Zhang, and Long Chen. Genad: Generative end-to-end autonomous driving. arXiv preprint arXiv:2402.11502, 2024. 1 +[60] Zangwei Zheng, Xiangyu Peng, Tianji Yang, Chenhui Shen, Shenggui Li, Hongxin Liu, Yukun Zhou, Tianyi Li, and Yang You. Open-sora: Democratizing efficient video production for all, 2024. 5, 6, 1 +[61] Vlas Zyrianov, Xiyue Zhu, and Shenlong Wang. Learning to generate realistic lidar point cloud. In ECCV, 2022. 3 +[62] Vlas Zyrianov, Henry Che, Zhijian Liu, and Shenlong Wang. Lidardm: Generative lidar simulation in a generated world. arXiv preprint arXiv:2404.02903, 2024. 3 \ No newline at end of file diff --git a/2025/DrivingSphere_ Building a High-fidelity 4D World for Closed-loop Simulation/images.zip b/2025/DrivingSphere_ Building a High-fidelity 4D World for Closed-loop Simulation/images.zip new file mode 100644 index 0000000000000000000000000000000000000000..fa55caff9146f15a63de75246bb3d34cf4a5d9e1 --- /dev/null +++ b/2025/DrivingSphere_ Building a High-fidelity 4D World for Closed-loop Simulation/images.zip @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:3a657afd37831b0799eed468d6fd74c81c3b09cf999cb6cfb77fd31c0d5069f8 +size 686327 diff --git a/2025/DrivingSphere_ Building a High-fidelity 4D World for Closed-loop Simulation/layout.json b/2025/DrivingSphere_ Building a High-fidelity 4D World for Closed-loop Simulation/layout.json new file mode 100644 index 0000000000000000000000000000000000000000..19e32121cfd72e2db4bea779b054af128fdebef1 --- /dev/null +++ b/2025/DrivingSphere_ Building a High-fidelity 4D World for Closed-loop Simulation/layout.json @@ -0,0 +1,9927 @@ +{ + "pdf_info": [ + { + "para_blocks": [ + { + "bbox": [ + 58, + 103, + 534, + 121 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 58, + 103, + 534, + 121 + ], + "spans": [ + { + "bbox": [ + 58, + 103, + 534, + 121 + ], + "type": "text", + "content": "DrivingSphere: Building a High-fidelity 4D World for Closed-loop Simulation" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 142, + 142, + 449, + 171 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 142, + 142, + 449, + 171 + ], + "spans": [ + { + "bbox": [ + 142, + 142, + 449, + 171 + ], + "type": "text", + "content": "Tianyi Yan" + }, + { + "bbox": [ + 142, + 142, + 449, + 171 + ], + "type": "inline_equation", + "content": "^{1}" + }, + { + "bbox": [ + 142, + 142, + 449, + 171 + ], + "type": "text", + "content": ", Dongming Wu" + }, + { + "bbox": [ + 142, + 142, + 449, + 171 + ], + "type": "inline_equation", + "content": "^{3}" + }, + { + "bbox": [ + 142, + 142, + 449, + 171 + ], + "type": "text", + "content": ", Wencheng Han" + }, + { + "bbox": [ + 142, + 142, + 449, + 171 + ], + "type": "inline_equation", + "content": "^{1}" + }, + { + "bbox": [ + 142, + 142, + 449, + 171 + ], + "type": "text", + "content": ", Junpeng Jiang" + }, + { + "bbox": [ + 142, + 142, + 449, + 171 + ], + "type": "inline_equation", + "content": "^{2}" + }, + { + "bbox": [ + 142, + 142, + 449, + 171 + ], + "type": "text", + "content": ", Xia Zhou" + }, + { + "bbox": [ + 142, + 142, + 449, + 171 + ], + "type": "inline_equation", + "content": "^{2}" + }, + { + "bbox": [ + 142, + 142, + 449, + 171 + ], + "type": "text", + "content": ", Kun Zhan" + }, + { + "bbox": [ + 142, + 142, + 449, + 171 + ], + "type": "inline_equation", + "content": "^{2}" + }, + { + "bbox": [ + 142, + 142, + 449, + 171 + ], + "type": "text", + "content": ", Cheng-zhong Xu" + }, + { + "bbox": [ + 142, + 142, + 449, + 171 + ], + "type": "inline_equation", + "content": "^{1}" + }, + { + "bbox": [ + 142, + 142, + 449, + 171 + ], + "type": "text", + "content": ", Jianbing Shen" + }, + { + "bbox": [ + 142, + 142, + 449, + 171 + ], + "type": "inline_equation", + "content": "^{1*}" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 88, + 171, + 506, + 200 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 88, + 171, + 506, + 200 + ], + "spans": [ + { + "bbox": [ + 88, + 171, + 506, + 200 + ], + "type": "inline_equation", + "content": "^{1}" + }, + { + "bbox": [ + 88, + 171, + 506, + 200 + ], + "type": "text", + "content": "SKL-IOTSC, Computer and Information Science, University of Macau, " + }, + { + "bbox": [ + 88, + 171, + 506, + 200 + ], + "type": "inline_equation", + "content": "^{2}" + }, + { + "bbox": [ + 88, + 171, + 506, + 200 + ], + "type": "text", + "content": "Li Auto Inc., " + }, + { + "bbox": [ + 88, + 171, + 506, + 200 + ], + "type": "inline_equation", + "content": "^{3}" + }, + { + "bbox": [ + 88, + 171, + 506, + 200 + ], + "type": "text", + "content": "School of Computer Science, Beijing Institute of Technology" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 143, + 227, + 192, + 239 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 143, + 227, + 192, + 239 + ], + "spans": [ + { + "bbox": [ + 143, + 227, + 192, + 239 + ], + "type": "text", + "content": "Abstract" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 47, + 253, + 289, + 552 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 253, + 289, + 552 + ], + "spans": [ + { + "bbox": [ + 47, + 253, + 289, + 552 + ], + "type": "text", + "content": "Autonomous driving evaluation requires simulation environments that closely replicate actual road conditions, including real-world sensory data and responsive feedback loops. However, many existing simulations need to predict waypoints along fixed routes on public datasets or synthetic photorealistic data, i.e., open-loop simulation usually lacks the ability to assess dynamic decision-making. While the recent efforts of closed-loop simulation offer feedback-driven environments, they cannot process visual sensor inputs or produce outputs that differ from real-world data. To address these challenges, we propose DrivingSphere, a realistic and closed-loop simulation framework. Its core idea is to build 4D world representation and generate real-life and controllable driving scenarios. In specific, our framework includes a Dynamic Environment Composition module that constructs a detailed 4D driving world with a format of occupancy equipping with static backgrounds and dynamic objects, and a Visual Scene Synthesis module that transforms this data into high-fidelity, multi-view video outputs, ensuring spatial and temporal consistency. By providing a dynamic and realistic simulation environment, DrivingSphere enables comprehensive testing and validation of autonomous driving algorithms, ultimately advancing the development of more reliable autonomous cars. The benchmark will be publicly released." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 47, + 577, + 128, + 588 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 577, + 128, + 588 + ], + "spans": [ + { + "bbox": [ + 47, + 577, + 128, + 588 + ], + "type": "text", + "content": "1. Introduction" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 46, + 597, + 287, + 646 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 597, + 287, + 646 + ], + "spans": [ + { + "bbox": [ + 46, + 597, + 287, + 646 + ], + "type": "text", + "content": "End-to-end autonomous driving algorithms [5-7, 15, 19, 20, 25, 36, 42, 48, 53, 54, 59] have made significant progress in recent years, making the accurate evaluation of these models a very urgent task. For safe and responsible evaluation," + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 306, + 228, + 545, + 288 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 228, + 545, + 288 + ], + "spans": [ + { + "bbox": [ + 306, + 228, + 545, + 288 + ], + "type": "text", + "content": "it is essential to have a precise simulation environment that accurately reflects real-world driving conditions. This requirement usually includes two aspects: firstly, the generation of high-fidelity sensory data, and secondly, the implementation of closed-loop feedback mechanisms." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 304, + 290, + 546, + 458 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 290, + 546, + 458 + ], + "spans": [ + { + "bbox": [ + 304, + 290, + 546, + 458 + ], + "type": "text", + "content": "The most common evaluation methods involve the forecasting of waypoints along fixed routes within collected public datasets [3, 37], that is, open-loop simulation approaches. While these benchmarks offer realistic driving data, they have relatively fixed distributions and lack diversity [24], limiting their ability to assess the generalization capabilities of autonomous driving algorithms. Even though several recent works [13, 43, 57] excel at producing diverse, photo-realistic data, continue to rely on the prediction of waypoints that are pre-established. In summary, despite high-fidelity sensory data, these open-loop evaluation solutions cannot provide the dynamic feedback necessary for assessing how autonomous systems respond to dynamic changes and decision-making." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 304, + 460, + 547, + 592 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 460, + 547, + 592 + ], + "spans": [ + { + "bbox": [ + 304, + 460, + 547, + 592 + ], + "type": "text", + "content": "In contrast, closed-loop simulations [10, 23, 46, 56] offer feedback-driven systems where an agent's actions influence and are influenced by other agents and the environment. For example, traffic flow-based simulation methods [12, 46, 56] successfully enable multi-agent simulations. However, they lack the ability to process visual sensor inputs, limiting their interplay with vision-based end-to-end models. Game engine-based simulators [10, 23] create scalable and physically realistic environments, but their outputs often diverge from real-world sensor data, restricting their utility in validating algorithms trained on actual inputs." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 304, + 594, + 547, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 594, + 547, + 714 + ], + "spans": [ + { + "bbox": [ + 304, + 594, + 547, + 714 + ], + "type": "text", + "content": "To address these challenges, we propose DrivingSphere, a novel generative closed-loop simulation framework that leverages geometric-wise prior information to generate realistic and controllable driving scenarios. In comparison with existing simulations, DrivingSphere has three distinguished features: ① Rich simulation granularity. Unlike past methods that model only roads and cars [13, 51], our approach allows for including previously unmodeled elements, such as buildings, vegetation, and other environmental structures. Although these non-traffic elements are not" + } + ] + } + ], + "index": 12 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "spans": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "type": "text", + "content": "CVF" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 145, + 0, + 496, + 35 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 145, + 0, + 496, + 35 + ], + "spans": [ + { + "bbox": [ + 145, + 0, + 496, + 35 + ], + "type": "text", + "content": "This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 46, + 655, + 287, + 713 + ], + "type": "page_footnote", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 655, + 287, + 713 + ], + "spans": [ + { + "bbox": [ + 46, + 655, + 287, + 713 + ], + "type": "text", + "content": "*Corresponding author: Jianbing Shen. This work was supported in part by the Science and Technology Development Fund of Macau SAR (FDCT) under grants 0102/2023/RIA2 and 0154/2022/A3 and 001/2024/SKL, the Jiangyin Hi-tech Industrial Development Zone under the Taihu Innovation Scheme (EF2025-00003-SKL-IOTSC), the University of Macau SRG2022-00023-IOTSC grant." + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "text", + "content": "27531" + } + ] + } + ], + "index": 14 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 0 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 51, + 71, + 211, + 202 + ], + "blocks": [ + { + "bbox": [ + 51, + 71, + 211, + 202 + ], + "lines": [ + { + "bbox": [ + 51, + 71, + 211, + 202 + ], + "spans": [ + { + "bbox": [ + 51, + 71, + 211, + 202 + ], + "type": "image", + "image_path": "25b82b33c5d51a3320144bd16ea17e52fe7cd31faa60d464b746dcd10e3109a1.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 46, + 204, + 547, + 271 + ], + "lines": [ + { + "bbox": [ + 46, + 204, + 547, + 271 + ], + "spans": [ + { + "bbox": [ + 46, + 204, + 547, + 271 + ], + "type": "text", + "content": "Figure 1. Comparison of frameworks for evaluating end-to-end Autonomous Driving (AD) algorithms. (a) Open-loop evaluation uses waypoint predictions along fixed routes in pre-collected datasets [3]. Generative models [13, 43, 57] create diverse, realistic data but lack dynamic feedback to assess AD responses to dynamic changes. (b) Simulated closed-loop evaluation [10, 23] offers feedback-driven, scalable environments where agent actions impact simulation dynamics; however, sensory outputs often differ from real-world data, limiting effectiveness for algorithms trained on real data. (c) Our generative closed-loop simulation framework, *DrivingSphere*, addresses these limitations by delivering realistic visual inputs and continuous, responsive feedback between the AD agent and environment." + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "type": "image", + "bbox": [ + 212, + 72, + 355, + 202 + ], + "blocks": [ + { + "bbox": [ + 212, + 72, + 355, + 202 + ], + "lines": [ + { + "bbox": [ + 212, + 72, + 355, + 202 + ], + "spans": [ + { + "bbox": [ + 212, + 72, + 355, + 202 + ], + "type": "image", + "image_path": "b28a90788599772a538a3d8c7ecc3fbd7075c7e529e415d2d9894d3d5d01cd3b.jpg" + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_body" + } + ], + "index": 1 + }, + { + "type": "image", + "bbox": [ + 359, + 72, + 541, + 202 + ], + "blocks": [ + { + "bbox": [ + 359, + 72, + 541, + 202 + ], + "lines": [ + { + "bbox": [ + 359, + 72, + 541, + 202 + ], + "spans": [ + { + "bbox": [ + 359, + 72, + 541, + 202 + ], + "type": "image", + "image_path": "35657bbd26f8ca421f8cbbdc3ecd32048e8d534e2b1bd8693faf728c134b1a61.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + } + ], + "index": 2 + }, + { + "bbox": [ + 46, + 283, + 289, + 462 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 283, + 289, + 462 + ], + "spans": [ + { + "bbox": [ + 46, + 283, + 289, + 462 + ], + "type": "text", + "content": "directly involved in traffic flow, their presence can significantly impact the input to driving models, thereby influencing decision-making processes in complex driving scenarios. 2 Physical and spatial realism. Due to our model explicitly representing the scene and traffic actors in 4D space, it is capable of precisely depicting physical interactions and occlusion relationships across different traffic elements. This ensures that each viewpoint and position naturally adheres to physical principles like depth and occlusion, thus achieving structured coordination of global road layout, traffic participants, and their behaviors. 3 High visual consistency and fidelity. Our model pays more effort to associate each traffic actor's appearance and unique ID within the scene, providing stable and high-fidelity temporal and spatial coherence across frames and views." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 46, + 468, + 287, + 601 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 468, + 287, + 601 + ], + "spans": [ + { + "bbox": [ + 46, + 468, + 287, + 601 + ], + "type": "text", + "content": "The proposed DrivingSphere contains two key modules: Dynamic Environment Composition and Visual Scene Synthesis. First, the Dynamic Environment Composition module aims to produce a coarse-grained 4D world with an occupancy grid format, which integrates large static backgrounds with dynamic traffic actors. For this, we introduce a BEV-conditioned occupancy diffusion model, named OccDreamer, to generate static backgrounds. It provides an infinite city-scale scene via progressive region extension. Besides, we build an actor bank to manage key actors or objects, and update actors' spatial-temporal position." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 46, + 605, + 288, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 605, + 288, + 715 + ], + "spans": [ + { + "bbox": [ + 46, + 605, + 288, + 715 + ], + "type": "text", + "content": "Once the scene is generated, the second module, Visual Scene Synthesis, converts it into high-fidelity, multi-view video outputs for AD system perception. To fully capture the complexity of the 4D environment, we employ a dual-path condition encoding strategy. Specifically, the global branch, using a pretrained 4D encoder [41], extracts geometric information and spatial-temporal relationships directly from the occupancy data, while the local projection branch produces view-specific semantic maps that align on" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 304, + 283, + 547, + 415 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 283, + 547, + 415 + ], + "spans": [ + { + "bbox": [ + 304, + 283, + 547, + 415 + ], + "type": "text", + "content": "a pixel-to-pixel basis, accurately capturing occlusions and depth variations. To further ensure visual consistency, we design an ID-aware actor encoding mechanism that binds the appearance and identity information of traffic actors to their scene locations, which preserves spatial correlation across views and temporal consistency across frames. With the custom-designed dual-path condition encoding and ID-aware actor encoding, the Visual Scene Synthesis generates spatially and temporally coherent video sequences. This results in a highly realistic simulation environment that is ideal for testing and validating AD systems." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 304, + 420, + 547, + 612 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 420, + 547, + 612 + ], + "spans": [ + { + "bbox": [ + 304, + 420, + 547, + 612 + ], + "type": "text", + "content": "By responsive interplay between AD algorithms and the simulated environment, DrivingSphere supports a closed-loop feedback loop, where actions from AD systems [6, 19, 20, 36] influence surrounding actors and scenes. This simulation platform closely mirrors real-world conditions, providing a reliable foundation for comprehensive AD validation. Our extensive experiments show that DrivingSphere achieves superior performance in visual fidelity and temporal consistency, significantly reducing the domain gap between simulated and real-world environments. By excelling in both open-loop and closed-loop evaluations, DrivingSphere proves to be a reliable platform for dynamic agent-environment interplay in rigorous AD testing. Ultimately, DrivingSphere will drive progress toward the development of safer and more reliable autonomous vehicles. Overall, our contributions are four-fold:" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 306, + 618, + 547, + 712 + ], + "type": "list", + "angle": 0, + "index": 11, + "blocks": [ + { + "bbox": [ + 306, + 618, + 545, + 666 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 618, + 545, + 666 + ], + "spans": [ + { + "bbox": [ + 306, + 618, + 545, + 666 + ], + "type": "text", + "content": "- We introduce DrivingSphere, a novel geometry-aware closed-loop simulation framework that captures 2D visual and 3D geometric properties while seamlessly integrating with vision-based end-to-end driving agents." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 306, + 666, + 547, + 712 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 666, + 547, + 712 + ], + "spans": [ + { + "bbox": [ + 306, + 666, + 547, + 712 + ], + "type": "text", + "content": "- We propose the Dynamic Environment Composition that provides a comprehensive 4D driving world by modeling both static and dynamic elements, enabling intricate simulations across diverse environments." + } + ] + } + ], + "index": 10 + } + ], + "sub_type": "text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "type": "text", + "content": "27532" + } + ] + } + ], + "index": 12 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 1 + }, + { + "para_blocks": [ + { + "bbox": [ + 47, + 72, + 286, + 168 + ], + "type": "list", + "angle": 0, + "index": 2, + "blocks": [ + { + "bbox": [ + 47, + 72, + 286, + 118 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 72, + 286, + 118 + ], + "spans": [ + { + "bbox": [ + 47, + 72, + 286, + 118 + ], + "type": "text", + "content": "- We design the Visual Scene Synthesis that enables long-term consistency and high fidelity, addressing challenges in generating continuous and multi-view video sequences for AD simulations." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 47, + 120, + 286, + 168 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 120, + 286, + 168 + ], + "spans": [ + { + "bbox": [ + 47, + 120, + 286, + 168 + ], + "type": "text", + "content": "- We introduce the Agent Coordination module for scalable multi-agent systems, enabling the integration of multiple vehicle agents and establishing a crucial feedback loop for validating autonomous driving algorithms." + } + ] + } + ], + "index": 1 + } + ], + "sub_type": "text" + }, + { + "bbox": [ + 47, + 178, + 133, + 190 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 178, + 133, + 190 + ], + "spans": [ + { + "bbox": [ + 47, + 178, + 133, + 190 + ], + "type": "text", + "content": "2. Related Work" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 47, + 198, + 287, + 521 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 198, + 287, + 521 + ], + "spans": [ + { + "bbox": [ + 47, + 198, + 287, + 521 + ], + "type": "text", + "content": "Generative Models in Driving. Denoising Diffusion Probabilistic Models (DDPMs) [17] have transformed image generation by iteratively refining Gaussian noise into high-quality images. Based on this technique, generative models in autonomous driving have greatly advanced. For example, BEVGen [38] and BEVControl [50] use diffusion models to produce controllable street-view images from BEV layouts. More recently, diffusion-based video generation models [2, 16] enable methods like [13, 14, 43-45, 57] to generate multi-view driving videos, effectively supporting perception tasks. Inspired by these 2D methods, advances in 3D scene generation have also emerged. Integrating diffusion processes into LiDAR generation improves the realism of LiDAR scenes [18, 30, 34, 49, 61, 62]. Occupancy-based models have further progressed, with occupancy grids now central to 3D scene generation. SemCity [22] enhances outdoor 3D occupancy generation, while models like OccWorld [58] and OccSora [41] use occupancy data for vehicle motion prediction and trajectory-conditioned 4D occupancy generation, respectively. Unlike these, our OccDreamer combines text and BEV conditions to generate 3D occupancy data, dynamically enhanced with foreground motion and ego pose to form a complete 4D driving world. By conditioning on occupancy data, our VideoDreamer allows precise control over non-traffic elements and ensures spatial and temporal coherence, capturing essential physical relationships such as depth and occlusion." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 47, + 522, + 286, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 522, + 286, + 713 + ], + "spans": [ + { + "bbox": [ + 47, + 522, + 286, + 713 + ], + "type": "text", + "content": "Autonomous Driving Simulations. Autonomous driving simulation methods are developed from basic behavioral models to complex, interactive environments. Flow-based simulations [9, 11, 12, 39, 46, 56] enable multi-agent interactions through long-term traffic planning. However, these methods typically operate in 2D space and lack real-life visual outputs. Platforms like CARLA [10], SUMO [27], and MetaDrive [23] address these limitations by modeling traffic dynamics and supporting multi-agent interactions with visual contexts, but they have visual fidelity differences from real-world data, limiting their direct use with algorithms trained on such data. While reconstruction-based methods [21, 26, 29] and frameworks such as UniSim [52], and MARS [47] capture high-fidelity scenes, they are limited by the scope of the data used, restricting their ability to simulate varied and long-tail scenarios. DriveArena [51]" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 304, + 72, + 545, + 180 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 72, + 545, + 180 + ], + "spans": [ + { + "bbox": [ + 304, + 72, + 545, + 180 + ], + "type": "text", + "content": "recently advanced generative closed-loop simulation by using 2D traffic sketches to control visual outputs, integrating well with vision-based driving agents. However, reliance on 2D inputs constrains its modeling granularity and visual realism. Our approach overcomes these challenges by leveraging 4D occupancy data, generating visually coherent and geometrically accurate scenes that rigorously test driving agents in closed-loop simulations, equipping them to handle complex real-world conditions effectively." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 306, + 190, + 388, + 204 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 190, + 388, + 204 + ], + "spans": [ + { + "bbox": [ + 306, + 190, + 388, + 204 + ], + "type": "text", + "content": "3. Methodology" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 304, + 209, + 545, + 449 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 209, + 545, + 449 + ], + "spans": [ + { + "bbox": [ + 304, + 209, + 545, + 449 + ], + "type": "text", + "content": "DrivingSphere acts as a generative closed-loop simulation framework that combines occupancy-based 4D world modeling with advanced video rendering techniques, providing high-fidelity visual outputs that enhance both simulation realism and agent-environment interplay in autonomous driving scenarios. Specifically, DrivingSphere begins with Dynamic Environment Composition (Sec. 3.1), which generates static background from a map sketch, selects traffic participants from the actor bank, and updates the actor location, to formulate a 4D driving environment with a format of occupancy. Next, Visual Scene Synthesis (Sec. 3.2) conditions the occupancy data surrounding the ego vehicle, accurately capturing occlusion relationships and fine-grained semantic information to generate high-fidelity multi-view videos. Finally, the closed-loop feedback mechanism (Sec. 3.3) enables dynamic, responsive adjustments, where autonomous agents continuously receive updated visual data and generate control signals that modify the simulation environment, providing a comprehensive platform for algorithm testing and refinement. The overall framework is shown in Fig. 2." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 306, + 456, + 497, + 468 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 456, + 497, + 468 + ], + "spans": [ + { + "bbox": [ + 306, + 456, + 497, + 468 + ], + "type": "text", + "content": "3.1. Dynamic Environment Composition" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 304, + 474, + 545, + 629 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 474, + 545, + 629 + ], + "spans": [ + { + "bbox": [ + 304, + 474, + 545, + 629 + ], + "type": "text", + "content": "Previous driving simulation methods have often overlooked static and varied elements like buildings, obstacles, and vegetation [10, 51]. Although these elements are not direct traffic participants, they are part of the perception input of AD systems, influencing final driving decision-making. For instance, static objects like buildings may block sensors' line-of-sight, causing occlusions of other vehicles or pedestrians. Obstacles or vegetation can introduce sensor artifacts or false positives for AD systems. Both situations may result in incorrect decisions like unnecessary braking or lane changes. To address this issue, we propose a Dynamic Environment Composition to construct a comprehensive 4D Driving World equipping complex digital assets." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 304, + 630, + 545, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 630, + 545, + 713 + ], + "spans": [ + { + "bbox": [ + 304, + 630, + 545, + 713 + ], + "type": "text", + "content": "4D Driving World Representation. Our 4D driving world representation includes three critical digital assets: static background scenes, dynamic foreground actors, and spatial position of these actors. Formally, we define this representation as " + }, + { + "bbox": [ + 304, + 630, + 545, + 713 + ], + "type": "inline_equation", + "content": "\\pmb{W} = \\{\\pmb{S}_{\\mathrm{city}},(\\pmb{A}_1,\\pmb{P}_1),\\dots ,(\\pmb{A}_N,\\pmb{P}_N)\\}_{n = 1}^N" + }, + { + "bbox": [ + 304, + 630, + 545, + 713 + ], + "type": "text", + "content": ". Here, " + }, + { + "bbox": [ + 304, + 630, + 545, + 713 + ], + "type": "inline_equation", + "content": "S_{\\mathrm{city}}" + }, + { + "bbox": [ + 304, + 630, + 545, + 713 + ], + "type": "text", + "content": " is a composite of multiple regional static scene " + }, + { + "bbox": [ + 304, + 630, + 545, + 713 + ], + "type": "inline_equation", + "content": "\\{S_1,\\ldots ,S_K\\}" + }, + { + "bbox": [ + 304, + 630, + 545, + 713 + ], + "type": "text", + "content": ". Each " + }, + { + "bbox": [ + 304, + 630, + 545, + 713 + ], + "type": "inline_equation", + "content": "S_{k}\\in \\mathbb{R}^{H\\times W\\times D}" + }, + { + "bbox": [ + 304, + 630, + 545, + 713 + ], + "type": "text", + "content": " is the " + }, + { + "bbox": [ + 304, + 630, + 545, + 713 + ], + "type": "inline_equation", + "content": "k^{th}" + }, + { + "bbox": [ + 304, + 630, + 545, + 713 + ], + "type": "text", + "content": " static back" + } + ] + } + ], + "index": 11 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "text", + "content": "27533" + } + ] + } + ], + "index": 12 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 2 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 53, + 76, + 545, + 262 + ], + "blocks": [ + { + "bbox": [ + 53, + 76, + 545, + 262 + ], + "lines": [ + { + "bbox": [ + 53, + 76, + 545, + 262 + ], + "spans": [ + { + "bbox": [ + 53, + 76, + 545, + 262 + ], + "type": "image", + "image_path": "c0a17a0c91e0ce25c99b5a8dd7ade49231f8cd6eae4e025749f89be402c165a2.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 46, + 263, + 547, + 318 + ], + "lines": [ + { + "bbox": [ + 46, + 263, + 547, + 318 + ], + "spans": [ + { + "bbox": [ + 46, + 263, + 547, + 318 + ], + "type": "text", + "content": "Figure 2. Overview of DrivingSphere framework. (a) The Dynamic Environment Composition module builds a 4D driving world, simulating real driving scenarios with backgrounds generated by OccDreamer, dynamic actors from an actor bank, and trajectories guided by a transportation simulator. (b) The Visual Scene Synthesis produces high-fidelity, photo-realistic video frames conditioned on global semantics, view-specific details, and scene prompts, supporting close-loop evaluation. Control signals enable adaptive feedback for driving agents, facilitating continuous testing and evaluation of driving algorithms in the simulated environment." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "bbox": [ + 46, + 322, + 288, + 445 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 322, + 288, + 445 + ], + "spans": [ + { + "bbox": [ + 46, + 322, + 288, + 445 + ], + "type": "text", + "content": "ground scene that captures the spatial layout and static elements within the region. " + }, + { + "bbox": [ + 46, + 322, + 288, + 445 + ], + "type": "inline_equation", + "content": "A_{n}\\in \\mathbb{R}^{4}" + }, + { + "bbox": [ + 46, + 322, + 288, + 445 + ], + "type": "text", + "content": " is the " + }, + { + "bbox": [ + 46, + 322, + 288, + 445 + ], + "type": "inline_equation", + "content": "n^{th}" + }, + { + "bbox": [ + 46, + 322, + 288, + 445 + ], + "type": "text", + "content": " actor of " + }, + { + "bbox": [ + 46, + 322, + 288, + 445 + ], + "type": "inline_equation", + "content": "N" + }, + { + "bbox": [ + 46, + 322, + 288, + 445 + ], + "type": "text", + "content": " actors, such as vehicles and pedestrians, defined by a 3D coordinate and a semantic label. It is worth noticing that both the static scene " + }, + { + "bbox": [ + 46, + 322, + 288, + 445 + ], + "type": "inline_equation", + "content": "\\pmb{S}_{k}" + }, + { + "bbox": [ + 46, + 322, + 288, + 445 + ], + "type": "text", + "content": " and dynamic actor " + }, + { + "bbox": [ + 46, + 322, + 288, + 445 + ], + "type": "inline_equation", + "content": "A_{n}" + }, + { + "bbox": [ + 46, + 322, + 288, + 445 + ], + "type": "text", + "content": " are represented as occupancy grids. " + }, + { + "bbox": [ + 46, + 322, + 288, + 445 + ], + "type": "inline_equation", + "content": "P_{n} = \\left\\{P_{n}^{0},\\dots ,P_{n}^{T}\\right\\}_{t = 0}^{T}" + }, + { + "bbox": [ + 46, + 322, + 288, + 445 + ], + "type": "text", + "content": " describe actor position across " + }, + { + "bbox": [ + 46, + 322, + 288, + 445 + ], + "type": "inline_equation", + "content": "T" + }, + { + "bbox": [ + 46, + 322, + 288, + 445 + ], + "type": "text", + "content": " frames, where " + }, + { + "bbox": [ + 46, + 322, + 288, + 445 + ], + "type": "inline_equation", + "content": "P_{n}^{t}\\in \\mathbb{R}^{4}" + }, + { + "bbox": [ + 46, + 322, + 288, + 445 + ], + "type": "text", + "content": " consists of a 3D relative position of the " + }, + { + "bbox": [ + 46, + 322, + 288, + 445 + ], + "type": "inline_equation", + "content": "n" + }, + { + "bbox": [ + 46, + 322, + 288, + 445 + ], + "type": "text", + "content": "-th actor at time " + }, + { + "bbox": [ + 46, + 322, + 288, + 445 + ], + "type": "inline_equation", + "content": "t" + }, + { + "bbox": [ + 46, + 322, + 288, + 445 + ], + "type": "text", + "content": " with respect to the ego vehicle and a yaw rotation " + }, + { + "bbox": [ + 46, + 322, + 288, + 445 + ], + "type": "inline_equation", + "content": "\\theta" + }, + { + "bbox": [ + 46, + 322, + 288, + 445 + ], + "type": "text", + "content": ". In the following, we will detail how to generate three assets." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 46, + 445, + 287, + 576 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 445, + 287, + 576 + ], + "spans": [ + { + "bbox": [ + 46, + 445, + 287, + 576 + ], + "type": "text", + "content": "Static Scene Generation. A straightforward approach generating static 3D scenes " + }, + { + "bbox": [ + 46, + 445, + 287, + 576 + ], + "type": "inline_equation", + "content": "S_{\\mathrm{city}}" + }, + { + "bbox": [ + 46, + 445, + 287, + 576 + ], + "type": "text", + "content": " is directly using ground truth occupancy data from existing datasets, like the Boston area in nuScenes [3]. However, this method is limited to specific regions captured during data collection, restricting its applicability to other urban areas. To address this limitation, we propose an occupancy diffusion model, OccDreamer, which is conditioned on BEV maps and textual descriptions, capable of generating static scenes for any desired urban region. The framework of OccDreamer, which integrates the following components, is illustrated in Fig. 3." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 47, + 576, + 287, + 651 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 576, + 287, + 651 + ], + "spans": [ + { + "bbox": [ + 47, + 576, + 287, + 651 + ], + "type": "text", + "content": "Firstly, to efficiently train the diffusion model while addressing the computational complexity of processing 3D data, we adopt VQVAE [17, 35, 40] as an occupancy tokenizer, mapping " + }, + { + "bbox": [ + 47, + 576, + 287, + 651 + ], + "type": "inline_equation", + "content": "S_{k}" + }, + { + "bbox": [ + 47, + 576, + 287, + 651 + ], + "type": "text", + "content": " to a latent feature " + }, + { + "bbox": [ + 47, + 576, + 287, + 651 + ], + "type": "inline_equation", + "content": "Z^{S}" + }, + { + "bbox": [ + 47, + 576, + 287, + 651 + ], + "type": "text", + "content": ": " + }, + { + "bbox": [ + 47, + 576, + 287, + 651 + ], + "type": "inline_equation", + "content": "Z^{S} = \\mathcal{F}_{\\mathrm{VAE,Enc}}^{occ}(S_{k})" + }, + { + "bbox": [ + 47, + 576, + 287, + 651 + ], + "type": "text", + "content": ". The reconstructed scene is defined as " + }, + { + "bbox": [ + 47, + 576, + 287, + 651 + ], + "type": "inline_equation", + "content": "S_{k}^{\\prime} = \\mathcal{F}_{\\mathrm{VAE,Dec}}^{occ}(Z^{S_{k}})" + }, + { + "bbox": [ + 47, + 576, + 287, + 651 + ], + "type": "text", + "content": ". " + }, + { + "bbox": [ + 47, + 576, + 287, + 651 + ], + "type": "inline_equation", + "content": "\\mathcal{F}_{\\mathrm{VAE}}^{occ}" + }, + { + "bbox": [ + 47, + 576, + 287, + 651 + ], + "type": "text", + "content": " is trained with a combined loss " + }, + { + "bbox": [ + 47, + 576, + 287, + 651 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{occ}" + }, + { + "bbox": [ + 47, + 576, + 287, + 651 + ], + "type": "text", + "content": ":" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 56, + 658, + 287, + 672 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 658, + 287, + 672 + ], + "spans": [ + { + "bbox": [ + 56, + 658, + 287, + 672 + ], + "type": "interline_equation", + "content": "\\mathcal {L} _ {o c c} = \\mathcal {L} _ {C E} \\left(\\boldsymbol {S} _ {k}, \\boldsymbol {S} _ {k} ^ {\\prime}\\right) + \\alpha \\mathcal {L} _ {L o v} \\left(\\boldsymbol {S} _ {k}, \\boldsymbol {S} _ {k} ^ {\\prime}\\right) + \\beta \\mathcal {L} _ {e m b}, \\tag {1}", + "image_path": "0849a734dca6eef9dc056f71a2944056d7817f4c68fc158a03e5b4dde2d2c51e.jpg" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 47, + 677, + 284, + 689 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 677, + 284, + 689 + ], + "spans": [ + { + "bbox": [ + 47, + 677, + 284, + 689 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 47, + 677, + 284, + 689 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{CE}" + }, + { + "bbox": [ + 47, + 677, + 284, + 689 + ], + "type": "text", + "content": " is the CE loss and " + }, + { + "bbox": [ + 47, + 677, + 284, + 689 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{Lov}" + }, + { + "bbox": [ + 47, + 677, + 284, + 689 + ], + "type": "text", + "content": " is the Lovász loss [1]." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 47, + 689, + 287, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 689, + 287, + 714 + ], + "spans": [ + { + "bbox": [ + 47, + 689, + 287, + 714 + ], + "type": "text", + "content": "Secondly, considering varying road structures and complex regional topographies, we propose a controllable re" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 304, + 322, + 545, + 478 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 322, + 545, + 478 + ], + "spans": [ + { + "bbox": [ + 304, + 322, + 545, + 478 + ], + "type": "text", + "content": "gion occupancy generation module. It accepts a regional BEV map capturing the road structure and a language prompt describing abstract features (like suburban areas with rich vegetation or commercial districts with buildings lining both sides) as input, and outputs a region occupancy " + }, + { + "bbox": [ + 304, + 322, + 545, + 478 + ], + "type": "inline_equation", + "content": "S_{k}" + }, + { + "bbox": [ + 304, + 322, + 545, + 478 + ], + "type": "text", + "content": ". Following the principles of diffusion models [17, 55], a CLIP encoder [31] converts the text prompt into an embedding " + }, + { + "bbox": [ + 304, + 322, + 545, + 478 + ], + "type": "inline_equation", + "content": "F_{\\mathrm{region}}" + }, + { + "bbox": [ + 304, + 322, + 545, + 478 + ], + "type": "text", + "content": ". Then, " + }, + { + "bbox": [ + 304, + 322, + 545, + 478 + ], + "type": "inline_equation", + "content": "F_{\\mathrm{region}}" + }, + { + "bbox": [ + 304, + 322, + 545, + 478 + ], + "type": "text", + "content": " is injected into the denoiser " + }, + { + "bbox": [ + 304, + 322, + 545, + 478 + ], + "type": "inline_equation", + "content": "\\epsilon_{\\theta}^{s}" + }, + { + "bbox": [ + 304, + 322, + 545, + 478 + ], + "type": "text", + "content": " via cross-attention. Simultaneously, the BEV map is processed by a pre-trained image VAE [17] to extract the corresponding road embedding " + }, + { + "bbox": [ + 304, + 322, + 545, + 478 + ], + "type": "inline_equation", + "content": "F_{M}" + }, + { + "bbox": [ + 304, + 322, + 545, + 478 + ], + "type": "text", + "content": ", serving as input for the ControlNet branch " + }, + { + "bbox": [ + 304, + 322, + 545, + 478 + ], + "type": "inline_equation", + "content": "\\epsilon_{\\phi}^{s}" + }, + { + "bbox": [ + 304, + 322, + 545, + 478 + ], + "type": "text", + "content": ". This enables precise control over the diffusion learning process in the latent space:" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 306, + 482, + 555, + 510 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 482, + 555, + 510 + ], + "spans": [ + { + "bbox": [ + 306, + 482, + 555, + 510 + ], + "type": "interline_equation", + "content": "\\mathcal {L} _ {\\text {o c c}} = \\mathbb {E} _ {\\boldsymbol {F} _ {M}, \\boldsymbol {F} _ {\\text {r e g i o n}}, \\boldsymbol {Z}} s _ {k, \\epsilon^ {s}, \\tau} \\left[ \\left\\| \\epsilon^ {s} - \\epsilon_ {\\theta} ^ {s} \\left(\\boldsymbol {Z} _ {\\tau} ^ {\\boldsymbol {S} _ {k}}, \\boldsymbol {F} _ {\\text {r e g i o n}}, \\tau , \\epsilon_ {\\phi} ^ {s} (\\boldsymbol {F} _ {M})\\right) \\right\\| ^ {2} \\right], \\tag {2}", + "image_path": "8d718b02f0b4804a93481f94abe3ba429575f86f889b47f8e7f415009a910192.jpg" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 304, + 510, + 545, + 548 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 510, + 545, + 548 + ], + "spans": [ + { + "bbox": [ + 304, + 510, + 545, + 548 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 304, + 510, + 545, + 548 + ], + "type": "inline_equation", + "content": "Z_{\\tau}^{S_k}" + }, + { + "bbox": [ + 304, + 510, + 545, + 548 + ], + "type": "text", + "content": " is the noisy version of " + }, + { + "bbox": [ + 304, + 510, + 545, + 548 + ], + "type": "inline_equation", + "content": "Z^{S_k}" + }, + { + "bbox": [ + 304, + 510, + 545, + 548 + ], + "type": "text", + "content": " by adding " + }, + { + "bbox": [ + 304, + 510, + 545, + 548 + ], + "type": "inline_equation", + "content": "\\tau" + }, + { + "bbox": [ + 304, + 510, + 545, + 548 + ], + "type": "text", + "content": " steps noise " + }, + { + "bbox": [ + 304, + 510, + 545, + 548 + ], + "type": "inline_equation", + "content": "\\epsilon^s" + }, + { + "bbox": [ + 304, + 510, + 545, + 548 + ], + "type": "text", + "content": ". After training, we use OccDreamer to create " + }, + { + "bbox": [ + 304, + 510, + 545, + 548 + ], + "type": "inline_equation", + "content": "Z_{k}^{S}" + }, + { + "bbox": [ + 304, + 510, + 545, + 548 + ], + "type": "text", + "content": " and decode it with " + }, + { + "bbox": [ + 304, + 510, + 545, + 548 + ], + "type": "inline_equation", + "content": "\\mathcal{F}_{\\mathrm{VAE,Decode}}^{\\mathrm{occ}}" + }, + { + "bbox": [ + 304, + 510, + 545, + 548 + ], + "type": "text", + "content": " to obtain " + }, + { + "bbox": [ + 304, + 510, + 545, + 548 + ], + "type": "inline_equation", + "content": "S_{k}" + }, + { + "bbox": [ + 304, + 510, + 545, + 548 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 304, + 547, + 545, + 681 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 547, + 545, + 681 + ], + "spans": [ + { + "bbox": [ + 304, + 547, + 545, + 681 + ], + "type": "text", + "content": "Thirdly, for spatial consistency across the entire 3D scene, we propose a Scene Expansion Mechanism for building a coherent and city-level static scene " + }, + { + "bbox": [ + 304, + 547, + 545, + 681 + ], + "type": "inline_equation", + "content": "S_{\\mathrm{city}}" + }, + { + "bbox": [ + 304, + 547, + 545, + 681 + ], + "type": "text", + "content": ". Its core idea is to expand an initial region " + }, + { + "bbox": [ + 304, + 547, + 545, + 681 + ], + "type": "inline_equation", + "content": "S_{k}" + }, + { + "bbox": [ + 304, + 547, + 545, + 681 + ], + "type": "text", + "content": " to a neighboring region " + }, + { + "bbox": [ + 304, + 547, + 545, + 681 + ], + "type": "inline_equation", + "content": "S_{k+1}" + }, + { + "bbox": [ + 304, + 547, + 545, + 681 + ], + "type": "text", + "content": " by using the overlap between these regions as a conditioning constraint. Formally, let " + }, + { + "bbox": [ + 304, + 547, + 545, + 681 + ], + "type": "inline_equation", + "content": "O" + }, + { + "bbox": [ + 304, + 547, + 545, + 681 + ], + "type": "text", + "content": " represent a binary mask that defines the overlapping area between two adjacent regions " + }, + { + "bbox": [ + 304, + 547, + 545, + 681 + ], + "type": "inline_equation", + "content": "S_{k}" + }, + { + "bbox": [ + 304, + 547, + 545, + 681 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 304, + 547, + 545, + 681 + ], + "type": "inline_equation", + "content": "S_{k+1}" + }, + { + "bbox": [ + 304, + 547, + 545, + 681 + ], + "type": "text", + "content": ". This overlap serves as the conditional context for generating the next region, enforcing spatial coherence at the boundary. To generate " + }, + { + "bbox": [ + 304, + 547, + 545, + 681 + ], + "type": "inline_equation", + "content": "S_{k+1}" + }, + { + "bbox": [ + 304, + 547, + 545, + 681 + ], + "type": "text", + "content": ", we first create a partially masked scene " + }, + { + "bbox": [ + 304, + 547, + 545, + 681 + ], + "type": "inline_equation", + "content": "S_{k}^{\\mathrm{partial}}" + }, + { + "bbox": [ + 304, + 547, + 545, + 681 + ], + "type": "text", + "content": " by applying " + }, + { + "bbox": [ + 304, + 547, + 545, + 681 + ], + "type": "inline_equation", + "content": "O" + }, + { + "bbox": [ + 304, + 547, + 545, + 681 + ], + "type": "text", + "content": " to " + }, + { + "bbox": [ + 304, + 547, + 545, + 681 + ], + "type": "inline_equation", + "content": "S_{k}" + }, + { + "bbox": [ + 304, + 547, + 545, + 681 + ], + "type": "text", + "content": ":" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 389, + 685, + 545, + 699 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 389, + 685, + 545, + 699 + ], + "spans": [ + { + "bbox": [ + 389, + 685, + 545, + 699 + ], + "type": "interline_equation", + "content": "\\boldsymbol {S} _ {k} ^ {\\text {p a r t i a l}} = \\boldsymbol {S} _ {k} \\odot \\boldsymbol {O}, \\tag {3}", + "image_path": "fba77297340df8242a3c683fcf076df13cede55d4314c000dbcc2374723330fe.jpg" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 306, + 702, + 545, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 702, + 545, + 714 + ], + "spans": [ + { + "bbox": [ + 306, + 702, + 545, + 714 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 306, + 702, + 545, + 714 + ], + "type": "inline_equation", + "content": "\\odot" + }, + { + "bbox": [ + 306, + 702, + 545, + 714 + ], + "type": "text", + "content": " denotes element-wise multiplication, extracting" + } + ] + } + ], + "index": 13 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "text", + "content": "27534" + } + ] + } + ], + "index": 14 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 3 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 51, + 70, + 286, + 264 + ], + "blocks": [ + { + "bbox": [ + 51, + 70, + 286, + 264 + ], + "lines": [ + { + "bbox": [ + 51, + 70, + 286, + 264 + ], + "spans": [ + { + "bbox": [ + 51, + 70, + 286, + 264 + ], + "type": "image", + "image_path": "b69dc1588143795e3d83e491b935aa255de9e42f0321e925119a0a0e0aa2d57d.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 46, + 264, + 287, + 297 + ], + "lines": [ + { + "bbox": [ + 46, + 264, + 287, + 297 + ], + "spans": [ + { + "bbox": [ + 46, + 264, + 287, + 297 + ], + "type": "text", + "content": "Figure 3. The framework of OccDreamer, which includes Occupancy Tokenizer, Region Occupancy Generation, and Scene Extensio to generate a city-level background for 4D driving world." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "bbox": [ + 46, + 309, + 287, + 370 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 309, + 287, + 370 + ], + "spans": [ + { + "bbox": [ + 46, + 309, + 287, + 370 + ], + "type": "text", + "content": "the overlapping section of " + }, + { + "bbox": [ + 46, + 309, + 287, + 370 + ], + "type": "inline_equation", + "content": "S_{k}" + }, + { + "bbox": [ + 46, + 309, + 287, + 370 + ], + "type": "text", + "content": ". We then employ a diffusion process to generate the neighboring region " + }, + { + "bbox": [ + 46, + 309, + 287, + 370 + ], + "type": "inline_equation", + "content": "S_{k + 1}" + }, + { + "bbox": [ + 46, + 309, + 287, + 370 + ], + "type": "text", + "content": ", conditioning on both the BEV map " + }, + { + "bbox": [ + 46, + 309, + 287, + 370 + ], + "type": "inline_equation", + "content": "M_{k + 1}" + }, + { + "bbox": [ + 46, + 309, + 287, + 370 + ], + "type": "text", + "content": " (capturing the road structure of the target area) and the partially masked scene " + }, + { + "bbox": [ + 46, + 309, + 287, + 370 + ], + "type": "inline_equation", + "content": "S_{k}^{\\mathrm{partial}}" + }, + { + "bbox": [ + 46, + 309, + 287, + 370 + ], + "type": "text", + "content": ". This diffusion process is represented as:" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 89, + 377, + 287, + 393 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 89, + 377, + 287, + 393 + ], + "spans": [ + { + "bbox": [ + 89, + 377, + 287, + 393 + ], + "type": "interline_equation", + "content": "\\boldsymbol {Z} ^ {\\boldsymbol {S} _ {k + 1}} = \\mathcal {F} _ {\\text {o u t p a i n t i n g}} \\left(\\epsilon^ {s}, \\boldsymbol {M} _ {k + 1}, \\boldsymbol {S} _ {k} ^ {\\text {p a r t i a l}}\\right), \\tag {4}", + "image_path": "860e1731e6f2ab8639cb8fa4283785ac94c6f2117caa35174a88a1fceacbcea5.jpg" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 46, + 401, + 287, + 473 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 401, + 287, + 473 + ], + "spans": [ + { + "bbox": [ + 46, + 401, + 287, + 473 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 46, + 401, + 287, + 473 + ], + "type": "inline_equation", + "content": "\\epsilon^s" + }, + { + "bbox": [ + 46, + 401, + 287, + 473 + ], + "type": "text", + "content": " is the input noise, and " + }, + { + "bbox": [ + 46, + 401, + 287, + 473 + ], + "type": "inline_equation", + "content": "\\mathcal{F}_{\\mathrm{outpainting}}" + }, + { + "bbox": [ + 46, + 401, + 287, + 473 + ], + "type": "text", + "content": " denotes the diffusion model conditioned on the masked scene and BEV map. After obtaining the latent representation " + }, + { + "bbox": [ + 46, + 401, + 287, + 473 + ], + "type": "inline_equation", + "content": "Z^{S_{k + 1}}" + }, + { + "bbox": [ + 46, + 401, + 287, + 473 + ], + "type": "text", + "content": ", we decode it using the occupancy VAE decoder, resulting in the expanded region " + }, + { + "bbox": [ + 46, + 401, + 287, + 473 + ], + "type": "inline_equation", + "content": "S_{k + 1}" + }, + { + "bbox": [ + 46, + 401, + 287, + 473 + ], + "type": "text", + "content": ". The two regions are then combined into a larger scene " + }, + { + "bbox": [ + 46, + 401, + 287, + 473 + ], + "type": "inline_equation", + "content": "S_{\\mathrm{merged}}" + }, + { + "bbox": [ + 46, + 401, + 287, + 473 + ], + "type": "text", + "content": " through a merging operation:" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 89, + 481, + 287, + 495 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 89, + 481, + 287, + 495 + ], + "spans": [ + { + "bbox": [ + 89, + 481, + 287, + 495 + ], + "type": "interline_equation", + "content": "\\boldsymbol {S} _ {\\text {m e r g e d}} = \\operatorname {M e r g e} \\left(\\boldsymbol {S} _ {k} \\odot (1 - \\boldsymbol {O}), \\boldsymbol {S} _ {k + 1}\\right), \\tag {5}", + "image_path": "eaeb1bbfe31ee0bb597b916e891e048ea1e063221fc82c6e32363d936c4bc1fe.jpg" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 46, + 498, + 287, + 545 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 498, + 287, + 545 + ], + "spans": [ + { + "bbox": [ + 46, + 498, + 287, + 545 + ], + "type": "text", + "content": "where Merge ensures spatial continuity and logical consistency between adjacent regions. The spatially continuous and logically consistent city scene " + }, + { + "bbox": [ + 46, + 498, + 287, + 545 + ], + "type": "inline_equation", + "content": "S_{\\mathrm{city}}" + }, + { + "bbox": [ + 46, + 498, + 287, + 545 + ], + "type": "text", + "content": " serves as an excellent digital background for the 4D driving world." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 46, + 546, + 288, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 546, + 288, + 713 + ], + "spans": [ + { + "bbox": [ + 46, + 546, + 288, + 713 + ], + "type": "text", + "content": "Dynamic Actor Selection. To complement static scenes, we populate the 4D driving world with dynamic actors, creating realistic traffic flows. An actor bank " + }, + { + "bbox": [ + 46, + 546, + 288, + 713 + ], + "type": "inline_equation", + "content": "\\mathcal{B}" + }, + { + "bbox": [ + 46, + 546, + 288, + 713 + ], + "type": "text", + "content": " is constructed, where each actor " + }, + { + "bbox": [ + 46, + 546, + 288, + 713 + ], + "type": "inline_equation", + "content": "\\mathbf{A}_n \\in \\mathcal{B}" + }, + { + "bbox": [ + 46, + 546, + 288, + 713 + ], + "type": "text", + "content": " consists of instance-level voxel data segmented from real or synthetic occupancy scenes, with unique geometric, semantic, and behavioral attributes, as well as a descriptive caption " + }, + { + "bbox": [ + 46, + 546, + 288, + 713 + ], + "type": "inline_equation", + "content": "L_{\\mathbf{A}_n}" + }, + { + "bbox": [ + 46, + 546, + 288, + 713 + ], + "type": "text", + "content": " and an ID " + }, + { + "bbox": [ + 46, + 546, + 288, + 713 + ], + "type": "inline_equation", + "content": "I_{\\mathbf{A}_n}" + }, + { + "bbox": [ + 46, + 546, + 288, + 713 + ], + "type": "text", + "content": " to ensure consistency across frames. Actors are selected based on semantic similarity to user-provided descriptions via CLIP, or randomly sampled from relevant categories if unspecified, ensuring contextual diversity. This flexible selection process allows for dynamically relevant and diverse actor integration within the 4D driving world, supporting realistic and adaptable traffic simulation." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 304, + 72, + 545, + 156 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 72, + 545, + 156 + ], + "spans": [ + { + "bbox": [ + 304, + 72, + 545, + 156 + ], + "type": "text", + "content": "Spatial Position of Actors. The generation of actor positions " + }, + { + "bbox": [ + 304, + 72, + 545, + 156 + ], + "type": "inline_equation", + "content": "\\pmb{P}" + }, + { + "bbox": [ + 304, + 72, + 545, + 156 + ], + "type": "text", + "content": " is driven by the control signals output from autonomous agents (Sec. 3.3). Each actor is assigned an control signal set " + }, + { + "bbox": [ + 304, + 72, + 545, + 156 + ], + "type": "inline_equation", + "content": "C_n = \\{c_n^0, \\dots, c_n^T\\}" + }, + { + "bbox": [ + 304, + 72, + 545, + 156 + ], + "type": "text", + "content": ", where " + }, + { + "bbox": [ + 304, + 72, + 545, + 156 + ], + "type": "inline_equation", + "content": "c_n^t" + }, + { + "bbox": [ + 304, + 72, + 545, + 156 + ], + "type": "text", + "content": " represents the received control signal of actor " + }, + { + "bbox": [ + 304, + 72, + 545, + 156 + ], + "type": "inline_equation", + "content": "n" + }, + { + "bbox": [ + 304, + 72, + 545, + 156 + ], + "type": "text", + "content": " at time " + }, + { + "bbox": [ + 304, + 72, + 545, + 156 + ], + "type": "inline_equation", + "content": "t" + }, + { + "bbox": [ + 304, + 72, + 545, + 156 + ], + "type": "text", + "content": ". These signals iteratively update the dynamic state of the world. For " + }, + { + "bbox": [ + 304, + 72, + 545, + 156 + ], + "type": "inline_equation", + "content": "n" + }, + { + "bbox": [ + 304, + 72, + 545, + 156 + ], + "type": "text", + "content": "-th actor, we have the position:" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 376, + 160, + 545, + 174 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 376, + 160, + 545, + 174 + ], + "spans": [ + { + "bbox": [ + 376, + 160, + 545, + 174 + ], + "type": "interline_equation", + "content": "\\boldsymbol {P} _ {n} ^ {t + 1} = \\boldsymbol {P} _ {n} ^ {t} + \\triangle \\boldsymbol {P} _ {n} ^ {t} \\left(\\boldsymbol {c} _ {n} ^ {t}\\right), \\tag {6}", + "image_path": "6220aeb02e10d011255c4efa61df569cccd857b204ecf5664a2084f38a19ad27.jpg" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 304, + 175, + 545, + 200 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 175, + 545, + 200 + ], + "spans": [ + { + "bbox": [ + 304, + 175, + 545, + 200 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 304, + 175, + 545, + 200 + ], + "type": "inline_equation", + "content": "\\triangle P_n^t (c_N^t)" + }, + { + "bbox": [ + 304, + 175, + 545, + 200 + ], + "type": "text", + "content": " is the positional change resulting from the control signal " + }, + { + "bbox": [ + 304, + 175, + 545, + 200 + ], + "type": "inline_equation", + "content": "c_{N}^{t}" + }, + { + "bbox": [ + 304, + 175, + 545, + 200 + ], + "type": "text", + "content": " at time " + }, + { + "bbox": [ + 304, + 175, + 545, + 200 + ], + "type": "inline_equation", + "content": "t" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 304, + 200, + 545, + 248 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 200, + 545, + 248 + ], + "spans": [ + { + "bbox": [ + 304, + 200, + 545, + 248 + ], + "type": "text", + "content": "4D World Composition. With the static background, dynamic foreground actors and their positions calculated, we integrate them into the comprehensive 4D driving world. The world state at any time " + }, + { + "bbox": [ + 304, + 200, + 545, + 248 + ], + "type": "inline_equation", + "content": "t" + }, + { + "bbox": [ + 304, + 200, + 545, + 248 + ], + "type": "text", + "content": " is represented by:" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 373, + 253, + 545, + 269 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 373, + 253, + 545, + 269 + ], + "spans": [ + { + "bbox": [ + 373, + 253, + 545, + 269 + ], + "type": "interline_equation", + "content": "\\boldsymbol {W} ^ {t} = \\pi (\\boldsymbol {S}, \\left(\\boldsymbol {A} _ {n}, \\boldsymbol {c} _ {n} ^ {t}\\right) _ {n = 1} ^ {N}), \\tag {7}", + "image_path": "4f4cb36d7598f08382a5fdec402bdad0485fa064b6129d5a8df35050a3fdc697.jpg" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 304, + 270, + 545, + 342 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 270, + 545, + 342 + ], + "spans": [ + { + "bbox": [ + 304, + 270, + 545, + 342 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 304, + 270, + 545, + 342 + ], + "type": "inline_equation", + "content": "\\pi" + }, + { + "bbox": [ + 304, + 270, + 545, + 342 + ], + "type": "text", + "content": " is the composition operator that aligns each actor based on their current pose " + }, + { + "bbox": [ + 304, + 270, + 545, + 342 + ], + "type": "inline_equation", + "content": "P_{n}^{t}" + }, + { + "bbox": [ + 304, + 270, + 545, + 342 + ], + "type": "text", + "content": ", represents the set of poses for all actors, forming the complete 4D driving world at time " + }, + { + "bbox": [ + 304, + 270, + 545, + 342 + ], + "type": "inline_equation", + "content": "t" + }, + { + "bbox": [ + 304, + 270, + 545, + 342 + ], + "type": "text", + "content": ". By iterating this process over time, we construct the dynamic 4D world that accounts for actor interplay, occlusions, and the evolving scene geometry." + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 305, + 348, + 434, + 361 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 348, + 434, + 361 + ], + "spans": [ + { + "bbox": [ + 305, + 348, + 434, + 361 + ], + "type": "text", + "content": "3.2. Visual Scene Synthesis" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 304, + 366, + 545, + 557 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 366, + 545, + 557 + ], + "spans": [ + { + "bbox": [ + 304, + 366, + 545, + 557 + ], + "type": "text", + "content": "Previous generative models [13, 28, 43, 57] tend to employ 2D visual conditions, which fail to accurately capture the geometric and semantic complexities inherent in real-world driving scenes. In response, our Visual Scene Synthesis employs VideoDreamer, a custom-designed video diffusion model, to transform the occupancy-driven 4D driving world constructed in the previous section into high-fidelity visual results. The overall framework is shown in Fig. 4. Specifically, we introduce a Dual-path Condition Encoding strategy that focuses on encoding occupancy data as its primary condition. We further enhance the consistency of appearances across views and frames by developing an ID-aware Actor Encoding method. Finally, we integrate the Spatial-Temporal Diffusion Transformer (ST-DiT) from OpenSora [60] as the foundational technology to ensure visual consistency and generate artifact-free frames." + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 304, + 558, + 545, + 689 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 558, + 545, + 689 + ], + "spans": [ + { + "bbox": [ + 304, + 558, + 545, + 689 + ], + "type": "text", + "content": "Dual-path Condition Encoding. We design a dual-path condition encoding strategy aimed at effectively capturing occupancy data. Given a set of driving world data " + }, + { + "bbox": [ + 304, + 558, + 545, + 689 + ], + "type": "inline_equation", + "content": "\\mathbf{W} = \\{\\mathbf{W}^0, \\mathbf{W}^1, \\dots, \\mathbf{W}^T\\}" + }, + { + "bbox": [ + 304, + 558, + 545, + 689 + ], + "type": "text", + "content": " spanning frames from 0 up to " + }, + { + "bbox": [ + 304, + 558, + 545, + 689 + ], + "type": "inline_equation", + "content": "T" + }, + { + "bbox": [ + 304, + 558, + 545, + 689 + ], + "type": "text", + "content": ", we first encode them into a global feature " + }, + { + "bbox": [ + 304, + 558, + 545, + 689 + ], + "type": "inline_equation", + "content": "\\mathbf{F}_{\\mathrm{global}} = \\mathcal{F}_{\\mathrm{VAE}}^{\\mathrm{4Docc}}(\\mathbf{W})" + }, + { + "bbox": [ + 304, + 558, + 545, + 689 + ], + "type": "text", + "content": ", capturing comprehensive geometric information and the spatial-temporal relationships within the scene. Here, " + }, + { + "bbox": [ + 304, + 558, + 545, + 689 + ], + "type": "inline_equation", + "content": "\\mathcal{F}_{\\mathrm{VAE}}^{\\mathrm{4Docc}}" + }, + { + "bbox": [ + 304, + 558, + 545, + 689 + ], + "type": "text", + "content": " is the well-trained 4D encoder from occupancy VAE [41], responsible for capturing global geometry of the scene. This ensures that the overall structure and spatial layout of the scene are accurately represented." + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 304, + 689, + 545, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 689, + 545, + 714 + ], + "spans": [ + { + "bbox": [ + 304, + 689, + 545, + 714 + ], + "type": "text", + "content": "Meanwhile, for " + }, + { + "bbox": [ + 304, + 689, + 545, + 714 + ], + "type": "inline_equation", + "content": "t^{th}" + }, + { + "bbox": [ + 304, + 689, + 545, + 714 + ], + "type": "text", + "content": " frame (where " + }, + { + "bbox": [ + 304, + 689, + 545, + 714 + ], + "type": "inline_equation", + "content": "t = 0,\\dots,T" + }, + { + "bbox": [ + 304, + 689, + 545, + 714 + ], + "type": "text", + "content": "), 3D occupancy data " + }, + { + "bbox": [ + 304, + 689, + 545, + 714 + ], + "type": "inline_equation", + "content": "\\mathbf{W}^t" + }, + { + "bbox": [ + 304, + 689, + 545, + 714 + ], + "type": "text", + "content": " is projected into 2D semantic maps " + }, + { + "bbox": [ + 304, + 689, + 545, + 714 + ], + "type": "inline_equation", + "content": "\\mathcal{M}_v^t" + }, + { + "bbox": [ + 304, + 689, + 545, + 714 + ], + "type": "text", + "content": " for" + } + ] + } + ], + "index": 17 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "text", + "content": "27535" + } + ] + } + ], + "index": 18 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 4 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 51, + 72, + 286, + 247 + ], + "blocks": [ + { + "bbox": [ + 51, + 72, + 286, + 247 + ], + "lines": [ + { + "bbox": [ + 51, + 72, + 286, + 247 + ], + "spans": [ + { + "bbox": [ + 51, + 72, + 286, + 247 + ], + "type": "image", + "image_path": "e45d3593a899356336edcaa9f22d52370655dd9052d2eebc971e830c92a63463.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 46, + 248, + 287, + 307 + ], + "lines": [ + { + "bbox": [ + 46, + 248, + 287, + 307 + ], + "spans": [ + { + "bbox": [ + 46, + 248, + 287, + 307 + ], + "type": "text", + "content": "Figure 4. Overview of the VideoDreamer. The model conditions on 4D driving world and enriched actor embeddings (e.g., actor ID, position and caption). This ensures high visual fidelity and geometric consistency in the generated driving simulations. Each view " + }, + { + "bbox": [ + 46, + 248, + 287, + 307 + ], + "type": "inline_equation", + "content": "v" + }, + { + "bbox": [ + 46, + 248, + 287, + 307 + ], + "type": "text", + "content": " by camera extrinsics " + }, + { + "bbox": [ + 46, + 248, + 287, + 307 + ], + "type": "inline_equation", + "content": "\\mathbf{T}^t" + }, + { + "bbox": [ + 46, + 248, + 287, + 307 + ], + "type": "text", + "content": " and intrinsics " + }, + { + "bbox": [ + 46, + 248, + 287, + 307 + ], + "type": "inline_equation", + "content": "\\mathbf{K}_v" + }, + { + "bbox": [ + 46, + 248, + 287, + 307 + ], + "type": "text", + "content": ":" + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "bbox": [ + 119, + 312, + 287, + 326 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 119, + 312, + 287, + 326 + ], + "spans": [ + { + "bbox": [ + 119, + 312, + 287, + 326 + ], + "type": "interline_equation", + "content": "\\mathcal {M} _ {v} ^ {t} = \\mathcal {Q} \\left(\\boldsymbol {W} ^ {t}, \\mathbf {K} _ {v}, \\mathbf {T} ^ {t}\\right), \\tag {8}", + "image_path": "5e8057d0c9493566a65042c2a2cd3d30e9a5eec194a066fdec9b28cdb738bd58.jpg" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 46, + 329, + 287, + 425 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 329, + 287, + 425 + ], + "spans": [ + { + "bbox": [ + 46, + 329, + 287, + 425 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 46, + 329, + 287, + 425 + ], + "type": "inline_equation", + "content": "\\mathcal{Q}" + }, + { + "bbox": [ + 46, + 329, + 287, + 425 + ], + "type": "text", + "content": " is the render and projection function [33]. This process yields a sequence of semantic maps, " + }, + { + "bbox": [ + 46, + 329, + 287, + 425 + ], + "type": "inline_equation", + "content": "\\mathcal{M} \\in \\mathbb{R}^{T \\times 6 \\times H \\times W}" + }, + { + "bbox": [ + 46, + 329, + 287, + 425 + ], + "type": "text", + "content": ", with precise pixel-to-pixel alignment across views. Such alignment is essential for accurately capturing view-specific occlusions and depth variations. " + }, + { + "bbox": [ + 46, + 329, + 287, + 425 + ], + "type": "inline_equation", + "content": "\\mathcal{M}" + }, + { + "bbox": [ + 46, + 329, + 287, + 425 + ], + "type": "text", + "content": " is then encoded into view-specific features " + }, + { + "bbox": [ + 46, + 329, + 287, + 425 + ], + "type": "inline_equation", + "content": "F_{\\mathrm{view}} = \\mathcal{F}_{\\mathrm{VAE}}^{\\mathrm{Image}}(\\mathcal{M})" + }, + { + "bbox": [ + 46, + 329, + 287, + 425 + ], + "type": "text", + "content": ". " + }, + { + "bbox": [ + 46, + 329, + 287, + 425 + ], + "type": "inline_equation", + "content": "F_{\\mathrm{view}}" + }, + { + "bbox": [ + 46, + 329, + 287, + 425 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 46, + 329, + 287, + 425 + ], + "type": "inline_equation", + "content": "F_{\\mathrm{global}}" + }, + { + "bbox": [ + 46, + 329, + 287, + 425 + ], + "type": "text", + "content": " are then integrated through cross-attention to obtain the unified representation " + }, + { + "bbox": [ + 46, + 329, + 287, + 425 + ], + "type": "inline_equation", + "content": "F_{\\mathrm{occ}}" + }, + { + "bbox": [ + 46, + 329, + 287, + 425 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 46, + 425, + 287, + 461 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 425, + 287, + 461 + ], + "spans": [ + { + "bbox": [ + 46, + 425, + 287, + 461 + ], + "type": "text", + "content": "ID-aware Actor Encoding. To ensure coherence of each actor's appearance and identity information across the scene, we extract a fused sequence embedding " + }, + { + "bbox": [ + 46, + 425, + 287, + 461 + ], + "type": "inline_equation", + "content": "F_{\\mathrm{fuse}}" + }, + { + "bbox": [ + 46, + 425, + 287, + 461 + ], + "type": "text", + "content": ":" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 92, + 468, + 287, + 480 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 92, + 468, + 287, + 480 + ], + "spans": [ + { + "bbox": [ + 92, + 468, + 287, + 480 + ], + "type": "interline_equation", + "content": "\\boldsymbol {F} _ {\\text {f u s e}} = \\operatorname {C O N C A T} \\left[ \\boldsymbol {F} _ {\\boldsymbol {W}}, \\boldsymbol {F} _ {\\boldsymbol {A} _ {1}}, \\dots , \\boldsymbol {F} _ {\\boldsymbol {A} _ {N}} \\right], \\tag {9}", + "image_path": "1fd2e6f98b269569878947e214cf1a6436ce5a0a9d8fb3706436fdd5a3ba8328.jpg" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 46, + 484, + 287, + 509 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 484, + 287, + 509 + ], + "spans": [ + { + "bbox": [ + 46, + 484, + 287, + 509 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 46, + 484, + 287, + 509 + ], + "type": "inline_equation", + "content": "F_W" + }, + { + "bbox": [ + 46, + 484, + 287, + 509 + ], + "type": "text", + "content": " is the embedding of the scene caption " + }, + { + "bbox": [ + 46, + 484, + 287, + 509 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{\\mathrm{scene}}" + }, + { + "bbox": [ + 46, + 484, + 287, + 509 + ], + "type": "text", + "content": ", derived from T5 [32]. Each actor embedding " + }, + { + "bbox": [ + 46, + 484, + 287, + 509 + ], + "type": "inline_equation", + "content": "F_{A_n}" + }, + { + "bbox": [ + 46, + 484, + 287, + 509 + ], + "type": "text", + "content": " is:" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 58, + 514, + 287, + 536 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 58, + 514, + 287, + 536 + ], + "spans": [ + { + "bbox": [ + 58, + 514, + 287, + 536 + ], + "type": "interline_equation", + "content": "\\boldsymbol {F} _ {\\boldsymbol {A} _ {n}} = \\operatorname {C O N C A T} \\left[ \\mathcal {F} _ {\\text {F o u r i e r}} \\left(\\boldsymbol {P} _ {n}\\right), \\mathcal {F} _ {\\text {F o u r i e r}} \\left(\\boldsymbol {I} _ {n}\\right), \\mathcal {F} _ {T 5} \\left(\\boldsymbol {L} _ {\\boldsymbol {A} _ {i}}\\right) \\right]. \\tag {10}", + "image_path": "71fe1039d64746c1cf4280f95dfa0e33ac8e96a88d3b831ee8d75c5aaf3adc5a.jpg" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 46, + 534, + 287, + 605 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 534, + 287, + 605 + ], + "spans": [ + { + "bbox": [ + 46, + 534, + 287, + 605 + ], + "type": "text", + "content": "Here, " + }, + { + "bbox": [ + 46, + 534, + 287, + 605 + ], + "type": "inline_equation", + "content": "P_{n}" + }, + { + "bbox": [ + 46, + 534, + 287, + 605 + ], + "type": "text", + "content": " represents the position of the " + }, + { + "bbox": [ + 46, + 534, + 287, + 605 + ], + "type": "inline_equation", + "content": "n" + }, + { + "bbox": [ + 46, + 534, + 287, + 605 + ], + "type": "text", + "content": "-th actor and " + }, + { + "bbox": [ + 46, + 534, + 287, + 605 + ], + "type": "inline_equation", + "content": "I_{n}" + }, + { + "bbox": [ + 46, + 534, + 287, + 605 + ], + "type": "text", + "content": " is its unique ID. Both are encoded using Fourier encoding [29]. The actor caption " + }, + { + "bbox": [ + 46, + 534, + 287, + 605 + ], + "type": "inline_equation", + "content": "L_{A_i}" + }, + { + "bbox": [ + 46, + 534, + 287, + 605 + ], + "type": "text", + "content": " is encoded using the T5 [32]. Overall, by integrating geometry positions, textual descriptions, and identity information, we ensure coherent behavior and long-term visual consistency across frames." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 46, + 605, + 287, + 664 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 605, + 287, + 664 + ], + "spans": [ + { + "bbox": [ + 46, + 605, + 287, + 664 + ], + "type": "text", + "content": "ControlNet-DiT. To enhance visual fidelity and temporal consistency in generated videos, we integrate the ST-DiT [60] as our denoiser " + }, + { + "bbox": [ + 46, + 605, + 287, + 664 + ], + "type": "inline_equation", + "content": "\\epsilon_{\\theta}^{v}" + }, + { + "bbox": [ + 46, + 605, + 287, + 664 + ], + "type": "text", + "content": ", utilizing ST-DiT blocks stacked with View-aware Spatial Self-Attention (VSSA), Temporal Self-Attention, Cross-Attention, and FFN." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 47, + 665, + 287, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 665, + 287, + 713 + ], + "spans": [ + { + "bbox": [ + 47, + 665, + 287, + 713 + ], + "type": "text", + "content": "In the video generation process, the input video sequence " + }, + { + "bbox": [ + 47, + 665, + 287, + 713 + ], + "type": "inline_equation", + "content": "\\mathbf{V} \\in \\mathbb{R}^{v \\times T \\times H \\times W}" + }, + { + "bbox": [ + 47, + 665, + 287, + 713 + ], + "type": "text", + "content": " is first encoded into latent features " + }, + { + "bbox": [ + 47, + 665, + 287, + 713 + ], + "type": "inline_equation", + "content": "Z^{\\mathbf{V}} \\in \\mathbb{R}^{v \\times T \\times H' \\times W' \\times C}" + }, + { + "bbox": [ + 47, + 665, + 287, + 713 + ], + "type": "text", + "content": ". To ensure multi-view consistency while managing computational complexity, we mod-" + } + ] + } + ], + "index": 10 + }, + { + "type": "image", + "bbox": [ + 305, + 71, + 426, + 144 + ], + "blocks": [ + { + "bbox": [ + 305, + 71, + 426, + 144 + ], + "lines": [ + { + "bbox": [ + 305, + 71, + 426, + 144 + ], + "spans": [ + { + "bbox": [ + 305, + 71, + 426, + 144 + ], + "type": "image", + "image_path": "0045c73d182897ed8245df6fd9a28d3bf755bf90f20c45357484a7fd1c0013fc.jpg" + } + ] + } + ], + "index": 11, + "angle": 0, + "type": "image_body" + } + ], + "index": 11 + }, + { + "type": "image", + "bbox": [ + 430, + 72, + 545, + 144 + ], + "blocks": [ + { + "bbox": [ + 430, + 72, + 545, + 144 + ], + "lines": [ + { + "bbox": [ + 430, + 72, + 545, + 144 + ], + "spans": [ + { + "bbox": [ + 430, + 72, + 545, + 144 + ], + "type": "image", + "image_path": "77719c565930512c098d4fa8afe8ad5394909ae1e69dc07f626950a3e246d90a.jpg" + } + ] + } + ], + "index": 12, + "angle": 0, + "type": "image_body" + } + ], + "index": 12 + }, + { + "type": "image", + "bbox": [ + 310, + 146, + 539, + 201 + ], + "blocks": [ + { + "bbox": [ + 310, + 146, + 539, + 201 + ], + "lines": [ + { + "bbox": [ + 310, + 146, + 539, + 201 + ], + "spans": [ + { + "bbox": [ + 310, + 146, + 539, + 201 + ], + "type": "image", + "image_path": "857accd1b061e6fc3e1ffbc5c3c56509b22384a59361029f5893695d78415640.jpg" + } + ] + } + ], + "index": 13, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 304, + 202, + 545, + 236 + ], + "lines": [ + { + "bbox": [ + 304, + 202, + 545, + 236 + ], + "spans": [ + { + "bbox": [ + 304, + 202, + 545, + 236 + ], + "type": "text", + "content": "Figure 5. Qualitative results of generated 3D scene. OccDreamer uses text prompt and bev map as guidance to generate 3D scenes with controllable regional content and road structure." + } + ] + } + ], + "index": 14, + "angle": 0, + "type": "image_caption" + } + ], + "index": 13 + }, + { + "bbox": [ + 304, + 239, + 546, + 407 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 239, + 546, + 407 + ], + "spans": [ + { + "bbox": [ + 304, + 239, + 546, + 407 + ], + "type": "text", + "content": "if the original Spatial self-attention [60] as VSSA. The latent feature " + }, + { + "bbox": [ + 304, + 239, + 546, + 407 + ], + "type": "inline_equation", + "content": "Z^{\\mathbf{V}}" + }, + { + "bbox": [ + 304, + 239, + 546, + 407 + ], + "type": "text", + "content": " is reshaped from " + }, + { + "bbox": [ + 304, + 239, + 546, + 407 + ], + "type": "inline_equation", + "content": "v\\times T\\times H^{\\prime}\\times W^{\\prime}\\times C" + }, + { + "bbox": [ + 304, + 239, + 546, + 407 + ], + "type": "text", + "content": " to " + }, + { + "bbox": [ + 304, + 239, + 546, + 407 + ], + "type": "inline_equation", + "content": "T\\times (vH^{\\prime}W^{\\prime})\\times C" + }, + { + "bbox": [ + 304, + 239, + 546, + 407 + ], + "type": "text", + "content": ", treating the combined view, height, and width dimensions " + }, + { + "bbox": [ + 304, + 239, + 546, + 407 + ], + "type": "inline_equation", + "content": "(vH^{\\prime}W^{\\prime})" + }, + { + "bbox": [ + 304, + 239, + 546, + 407 + ], + "type": "text", + "content": " as the sequence length. This view-flatted approach reduces the parameter overhead associated with cross-view attention [13, 44] while maintaining spatial coherence across different views. For external conditions, we apply a cross-attention layer that injects the fused sequence embedding " + }, + { + "bbox": [ + 304, + 239, + 546, + 407 + ], + "type": "inline_equation", + "content": "F_{\\mathrm{fuse}}" + }, + { + "bbox": [ + 304, + 239, + 546, + 407 + ], + "type": "text", + "content": " (derived from scene context and actor identities) as: " + }, + { + "bbox": [ + 304, + 239, + 546, + 407 + ], + "type": "inline_equation", + "content": "Z^{V} = \\mathrm{CrossAttention}(Z^{V},F_{\\mathrm{fuse}})" + }, + { + "bbox": [ + 304, + 239, + 546, + 407 + ], + "type": "text", + "content": ". Last but not least, we introduce ControlNet branch " + }, + { + "bbox": [ + 304, + 239, + 546, + 407 + ], + "type": "inline_equation", + "content": "\\epsilon_{\\phi}^{v}" + }, + { + "bbox": [ + 304, + 239, + 546, + 407 + ], + "type": "text", + "content": " to ensure the occupancy control, inspired by PixArt-δ [4]. This branch duplicates blocks from " + }, + { + "bbox": [ + 304, + 239, + 546, + 407 + ], + "type": "inline_equation", + "content": "\\epsilon_{\\theta}^{v}" + }, + { + "bbox": [ + 304, + 239, + 546, + 407 + ], + "type": "text", + "content": " and takes " + }, + { + "bbox": [ + 304, + 239, + 546, + 407 + ], + "type": "inline_equation", + "content": "F_{\\mathrm{occ}}" + }, + { + "bbox": [ + 304, + 239, + 546, + 407 + ], + "type": "text", + "content": " as the input. The final diffusion process is:" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 305, + 413, + 547, + 440 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 413, + 547, + 440 + ], + "spans": [ + { + "bbox": [ + 305, + 413, + 547, + 440 + ], + "type": "interline_equation", + "content": "\\mathcal {L} _ {\\text {v i d e o}} = \\mathbb {E} _ {\\boldsymbol {F} _ {\\text {o c c}}, \\boldsymbol {F} _ {\\text {f u s e}}, \\boldsymbol {Z} V, \\epsilon^ {v}, \\tau} \\left[ \\left\\| \\epsilon^ {v} - \\epsilon_ {\\theta} ^ {v} \\left(\\boldsymbol {Z} _ {\\tau} ^ {\\boldsymbol {V}}, \\boldsymbol {F} _ {\\text {f u s e}}, \\tau , \\epsilon_ {\\phi} ^ {v} \\left(\\boldsymbol {F} _ {\\text {o c c}}\\right)\\right) \\right\\| ^ {2} \\right]. \\tag {11}", + "image_path": "3bc0210cb0f699e0c8de791d5ba0ece9bc2d17a2cab6e8b86975940fc1f6f6d1.jpg" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 304, + 438, + 545, + 473 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 438, + 545, + 473 + ], + "spans": [ + { + "bbox": [ + 304, + 438, + 545, + 473 + ], + "type": "text", + "content": "This approach ensures fine-grained control over spatial and temporal coherence, producing artifact-free frames that meet the high fidelity requirements for AD simulation." + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 304, + 474, + 545, + 545 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 474, + 545, + 545 + ], + "spans": [ + { + "bbox": [ + 304, + 474, + 545, + 545 + ], + "type": "text", + "content": "Auto-regressive Generation. To generate temporally consistent and content-coherent video frames, we employ an auto-regressive generation strategy [60] based on first-" + }, + { + "bbox": [ + 304, + 474, + 545, + 545 + ], + "type": "inline_equation", + "content": "f" + }, + { + "bbox": [ + 304, + 474, + 545, + 545 + ], + "type": "text", + "content": " frame mask. This approach models the dependency of each video frame on the previous frames, ensuring smooth transitions and maintaining visual continuity over time:" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 369, + 552, + 482, + 565 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 369, + 552, + 482, + 565 + ], + "spans": [ + { + "bbox": [ + 369, + 552, + 482, + 565 + ], + "type": "interline_equation", + "content": "\\boldsymbol {V} ^ {t} = \\mathcal {F} _ {\\text {v i d e o}} (\\boldsymbol {V} ^ {t - 1}, \\boldsymbol {Z} ^ {V}, t),", + "image_path": "a930a52d0b85120e2a6f54416a7b225cb557c26a513d0435ea5f3279ab757a8f.jpg" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 304, + 568, + 545, + 616 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 568, + 545, + 616 + ], + "spans": [ + { + "bbox": [ + 304, + 568, + 545, + 616 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 304, + 568, + 545, + 616 + ], + "type": "inline_equation", + "content": "V^t" + }, + { + "bbox": [ + 304, + 568, + 545, + 616 + ], + "type": "text", + "content": " is the video frame at time " + }, + { + "bbox": [ + 304, + 568, + 545, + 616 + ], + "type": "inline_equation", + "content": "t" + }, + { + "bbox": [ + 304, + 568, + 545, + 616 + ], + "type": "text", + "content": ", and " + }, + { + "bbox": [ + 304, + 568, + 545, + 616 + ], + "type": "inline_equation", + "content": "\\mathcal{F}_{\\mathrm{video}}" + }, + { + "bbox": [ + 304, + 568, + 545, + 616 + ], + "type": "text", + "content": " is the video generation function that conditions on the latent representation " + }, + { + "bbox": [ + 304, + 568, + 545, + 616 + ], + "type": "inline_equation", + "content": "Z^{V}" + }, + { + "bbox": [ + 304, + 568, + 545, + 616 + ], + "type": "text", + "content": " and the previous frame " + }, + { + "bbox": [ + 304, + 568, + 545, + 616 + ], + "type": "inline_equation", + "content": "V^{t-1}" + }, + { + "bbox": [ + 304, + 568, + 545, + 616 + ], + "type": "text", + "content": ". This ensures temporal coherence and infinite video sequences." + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 305, + 623, + 539, + 636 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 623, + 539, + 636 + ], + "spans": [ + { + "bbox": [ + 305, + 623, + 539, + 636 + ], + "type": "text", + "content": "3.3. Agent Interplay and Closed-Loop Simulation" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 304, + 641, + 545, + 689 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 641, + 545, + 689 + ], + "spans": [ + { + "bbox": [ + 304, + 641, + 545, + 689 + ], + "type": "text", + "content": "We enable seamless coordination of autonomous driving agents within the DrivingSphere simulation environment, categorizing agents into two primary types: the Ego Agent and Environment Agents." + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 304, + 689, + 545, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 689, + 545, + 713 + ], + "spans": [ + { + "bbox": [ + 304, + 689, + 545, + 713 + ], + "type": "text", + "content": "Ego Agent: The Ego Agent represents the autonomous driving system under evaluation. Driven by end-to-end" + } + ] + } + ], + "index": 23 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 318, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 318, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 318, + 757 + ], + "type": "text", + "content": "27536" + } + ] + } + ], + "index": 24 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 5 + }, + { + "para_blocks": [ + { + "bbox": [ + 47, + 72, + 287, + 99 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 72, + 287, + 99 + ], + "spans": [ + { + "bbox": [ + 47, + 72, + 287, + 99 + ], + "type": "text", + "content": "models [6, 19, 20], it receives visual input frames " + }, + { + "bbox": [ + 47, + 72, + 287, + 99 + ], + "type": "inline_equation", + "content": "V^{t}" + }, + { + "bbox": [ + 47, + 72, + 287, + 99 + ], + "type": "text", + "content": " and outputs the predicted control signal " + }, + { + "bbox": [ + 47, + 72, + 287, + 99 + ], + "type": "inline_equation", + "content": "c_{\\mathrm{ego}}^{t + 1}" + }, + { + "bbox": [ + 47, + 72, + 287, + 99 + ], + "type": "text", + "content": " at each time " + }, + { + "bbox": [ + 47, + 72, + 287, + 99 + ], + "type": "inline_equation", + "content": "t" + }, + { + "bbox": [ + 47, + 72, + 287, + 99 + ], + "type": "text", + "content": ":" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 126, + 106, + 287, + 121 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 126, + 106, + 287, + 121 + ], + "spans": [ + { + "bbox": [ + 126, + 106, + 287, + 121 + ], + "type": "interline_equation", + "content": "\\boldsymbol {c} _ {\\text {e g o}} ^ {t + 1} = \\mathcal {F} _ {\\text {d r i v i n g}} \\left(\\boldsymbol {V} ^ {t}\\right), \\tag {12}", + "image_path": "fe4c75c5849b70a86140269fdb0f507be2e48e6ad7aa6335f2ace5f4eed5f22f.jpg" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 47, + 128, + 271, + 140 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 128, + 271, + 140 + ], + "spans": [ + { + "bbox": [ + 47, + 128, + 271, + 140 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 47, + 128, + 271, + 140 + ], + "type": "inline_equation", + "content": "\\mathcal{F}_{\\mathrm{driving}}" + }, + { + "bbox": [ + 47, + 128, + 271, + 140 + ], + "type": "text", + "content": " denotes the driving algorithm [6, 19, 20]." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 46, + 140, + 287, + 235 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 140, + 287, + 235 + ], + "spans": [ + { + "bbox": [ + 46, + 140, + 287, + 235 + ], + "type": "text", + "content": "Environment Agents: Environment Agents are responsible for controlling the behavior and actions of all other actors within the simulation world " + }, + { + "bbox": [ + 46, + 140, + 287, + 235 + ], + "type": "inline_equation", + "content": "W" + }, + { + "bbox": [ + 46, + 140, + 287, + 235 + ], + "type": "text", + "content": ". To achieve realistic info exchanges, we utilize the traffic flow engine [46], which supports multi-agent simulation. Environment Agents receive inputs from the simulation's state and output control signal " + }, + { + "bbox": [ + 46, + 140, + 287, + 235 + ], + "type": "inline_equation", + "content": "c_{\\mathrm{n}}^{t + 1}" + }, + { + "bbox": [ + 46, + 140, + 287, + 235 + ], + "type": "text", + "content": " that dictate the movements and interplay of actors. Formally, this process can be described as:" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 124, + 243, + 287, + 257 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 124, + 243, + 287, + 257 + ], + "spans": [ + { + "bbox": [ + 124, + 243, + 287, + 257 + ], + "type": "interline_equation", + "content": "\\boldsymbol {c} _ {\\mathrm {n}} ^ {t + 1} = \\mathcal {F} _ {\\text {s y s t e m}} (\\boldsymbol {W} ^ {t}), \\tag {13}", + "image_path": "a2072c0decd96fc2680f0aabf8811e9d47e350f8c2255fb474528335d291ca11.jpg" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 47, + 265, + 251, + 277 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 265, + 251, + 277 + ], + "spans": [ + { + "bbox": [ + 47, + 265, + 251, + 277 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 47, + 265, + 251, + 277 + ], + "type": "inline_equation", + "content": "\\mathcal{F}_{\\mathrm{system}}" + }, + { + "bbox": [ + 47, + 265, + 251, + 277 + ], + "type": "text", + "content": " denotes the transportation simulator." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 47, + 277, + 287, + 350 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 277, + 287, + 350 + ], + "spans": [ + { + "bbox": [ + 47, + 277, + 287, + 350 + ], + "type": "text", + "content": "These agents take as input the visual frames " + }, + { + "bbox": [ + 47, + 277, + 287, + 350 + ], + "type": "inline_equation", + "content": "V^t" + }, + { + "bbox": [ + 47, + 277, + 287, + 350 + ], + "type": "text", + "content": " or previous state " + }, + { + "bbox": [ + 47, + 277, + 287, + 350 + ], + "type": "inline_equation", + "content": "\\pmb{W}^{t}" + }, + { + "bbox": [ + 47, + 277, + 287, + 350 + ], + "type": "text", + "content": ", and output actors' respective control signals " + }, + { + "bbox": [ + 47, + 277, + 287, + 350 + ], + "type": "inline_equation", + "content": "\\pmb{c}_{\\mathrm{ego}}^{t + 1}" + }, + { + "bbox": [ + 47, + 277, + 287, + 350 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 47, + 277, + 287, + 350 + ], + "type": "inline_equation", + "content": "\\pmb{c}_{\\mathrm{n}}^{t + 1}" + }, + { + "bbox": [ + 47, + 277, + 287, + 350 + ], + "type": "text", + "content": ". These signals influence the actor positions using Eq. (6) and further dynamically update the 4D driving world " + }, + { + "bbox": [ + 47, + 277, + 287, + 350 + ], + "type": "inline_equation", + "content": "\\pmb{W}^{t + 1}" + }, + { + "bbox": [ + 47, + 277, + 287, + 350 + ], + "type": "text", + "content": ", maintaining the closed-loop simulation where the environment continuously adapts to the agents' actions." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 47, + 360, + 127, + 373 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 360, + 127, + 373 + ], + "spans": [ + { + "bbox": [ + 47, + 360, + 127, + 373 + ], + "type": "text", + "content": "4. Experiments" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 47, + 380, + 107, + 392 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 380, + 107, + 392 + ], + "spans": [ + { + "bbox": [ + 47, + 380, + 107, + 392 + ], + "type": "text", + "content": "4.1. Settings" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 46, + 399, + 287, + 529 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 399, + 287, + 529 + ], + "spans": [ + { + "bbox": [ + 46, + 399, + 287, + 529 + ], + "type": "text", + "content": "Dataset and Evaluation Metrics. We utilize real-world driving datasets nuScenes [3]. These datasets provide rich sensor data, including LiDAR point clouds, camera images, and high-definition maps, which are essential for generating accurate occupancy grids and semantic information. We employ the following metrics to evaluate the performance of our simulation framework: FVD and UniAD-based metrics [19] for visual quality and fidelity. As for the evaluation of OccDreamer, we use FID and MMD [22]. In open-loop evaluation, we borrow metrics from NAVSIM [8]. RC and ADS for the closed-loop following [51]." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 47, + 531, + 267, + 543 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 531, + 267, + 543 + ], + "spans": [ + { + "bbox": [ + 47, + 531, + 267, + 543 + ], + "type": "text", + "content": "Implement Details are included in the supplementary." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 47, + 551, + 198, + 563 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 551, + 198, + 563 + ], + "spans": [ + { + "bbox": [ + 47, + 551, + 198, + 563 + ], + "type": "text", + "content": "4.2. Visual Fidelity and Realism" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 47, + 570, + 287, + 605 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 570, + 287, + 605 + ], + "spans": [ + { + "bbox": [ + 47, + 570, + 287, + 605 + ], + "type": "text", + "content": "We first evaluate the fidelity of the voxel-based scenes from OccDreamer and video sequences from VideoDreamer against real nuScenes data to assess the domain gap." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 46, + 606, + 287, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 606, + 287, + 713 + ], + "spans": [ + { + "bbox": [ + 46, + 606, + 287, + 713 + ], + "type": "text", + "content": "Scene Generation. We conduct quantitative and qualitative analyses to assess the fidelity of OccDreamer's generated occupancy data. Our method outperforms existing approaches quantitatively, as shown in Tab. 2. Specifically, SemCity employs a VAE trained on three planes to generate occupancy data while OccDreamer achieves pixel-to-pixel alignment based on BEV maps, resulting in more robust and accurate occupancy generation. Qualitatively, Fig. 5 presents the results of OccDreamer with both caption and" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 304, + 72, + 545, + 204 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 72, + 545, + 204 + ], + "spans": [ + { + "bbox": [ + 304, + 72, + 545, + 204 + ], + "type": "text", + "content": "BEV map. To the best of our knowledge, our approach is the first to achieve text-guided occupancy generation. As demonstrated, abstract textual descriptions effectively guide the model to generate diverse scenes, including outdoor environments and urban landscapes. Under BEV map control, the generated samples accurately reflect road structures and lane configurations. These capabilities establish a solid foundation for the comprehensive modeling of the 4D driving world, ensuring that the generated scenes are both spatially continuous and logically consistent. More visualizations are provided in the supplementary materials." + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 304, + 204, + 546, + 372 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 204, + 546, + 372 + ], + "spans": [ + { + "bbox": [ + 304, + 204, + 546, + 372 + ], + "type": "text", + "content": "Video Generation. We further assess the performance of our video generation model, VideoDreamer, in generating coherent video. We conduct a comparative analysis against leading models such as MagicDrive [13], DriveArena [51], and Panacea [45]. Using UniAD [19] as the evaluation framework, we assess the performance across crucial metrics including 3D object detection, BEV map segmentation, and trajectory planning. Our experimental results, summarized in Tab. 1, indicate that VideoDreamer significantly outperforms the baseline methods in these key metrics. This performance underscores the advantages of occupancy data over traditional 2D conditions, affirming VideoDreamer's effectiveness in creating realistic, dynamic scenes that accurately simulate real-world driving complexities." + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 304, + 372, + 545, + 479 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 372, + 545, + 479 + ], + "spans": [ + { + "bbox": [ + 304, + 372, + 545, + 479 + ], + "type": "text", + "content": "Qualitatively, as shown in Fig. 6, our approach demonstrates markedly superior performance in maintaining coherence and fidelity over time, significantly outperforming others. This success is primarily due to the geometry-aware occupancy encoding and instance encoding that ensures appearance consistency, highlighting VideoDreamer's robustness in creating visually consistent, detailed driving environments that capture real-world scene dynamics. More visualizations are included in the supplementary materials." + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 306, + 488, + 432, + 501 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 488, + 432, + 501 + ], + "spans": [ + { + "bbox": [ + 306, + 488, + 432, + 501 + ], + "type": "text", + "content": "4.3. Open-loop Evaluation" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 304, + 506, + 545, + 626 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 506, + 545, + 626 + ], + "spans": [ + { + "bbox": [ + 304, + 506, + 545, + 626 + ], + "type": "text", + "content": "In the open-loop setting, where AD agents passively receive environment input without influencing simulation dynamics, we assess the performance of UniAD [19] on DrivingSphere and alternative simulators. Using 100 predefined trajectories, Tab. 3 demonstrates better PDMS scores achieved by AD agents in DrivingSphere, suggesting that its higher visual fidelity reduces perceptual inaccuracies that can misguide agent decisions. This improvement highlights DrivingSphere's suitability for testing AD systems in environments that closely emulate real-world conditions." + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 306, + 635, + 437, + 647 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 635, + 437, + 647 + ], + "spans": [ + { + "bbox": [ + 306, + 635, + 437, + 647 + ], + "type": "text", + "content": "4.4. Closed-loop Evaluation" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 304, + 654, + 545, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 654, + 545, + 713 + ], + "spans": [ + { + "bbox": [ + 304, + 654, + 545, + 713 + ], + "type": "text", + "content": "In the closed-loop evaluation, AD agents receive visual input and output control signals, thus interactively shaping the simulation. This evaluation setup involves 100 pre-defined trajectories to test across controlled yet varied scenarios. As shown in Tab. 4, UniAD [19] exhibits relatively low Route" + } + ] + } + ], + "index": 20 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "text", + "content": "27537" + } + ] + } + ], + "index": 21 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 6 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 50, + 69, + 547, + 157 + ], + "blocks": [ + { + "bbox": [ + 50, + 69, + 547, + 157 + ], + "lines": [ + { + "bbox": [ + 50, + 69, + 547, + 157 + ], + "spans": [ + { + "bbox": [ + 50, + 69, + 547, + 157 + ], + "type": "table", + "html": "
Source DataQuality3DODSegmentationL2(↓)
FVD(↓)mAP(↑)NDS(↑)Lanes(↑)Drivable(↑)Divider(↑)1.0s2.0s3.0s
ori nuScenes[3]—37.9849.8531.3169.1425.930.510.981.65
MagicDrive [13]218.1212.9228.3621.9551.4617.100.571.141.95
Panacea [45]13913.7227.7318.2352.3717.210.581.141.97
DriveArena [51]185.3216.0630.0326.1459.3720.790.561.101.89
DrivingSphere103.4221.4534.1627.9962.8722.290.541.101.76
", + "image_path": "202e29fdce86c8ef0906abf1ae09262c9a0a4b2ea6f1396fa126ca8938b29d37.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "table_body" + } + ], + "index": 0 + }, + { + "type": "image", + "bbox": [ + 50, + 194, + 545, + 411 + ], + "blocks": [ + { + "bbox": [ + 50, + 194, + 545, + 411 + ], + "lines": [ + { + "bbox": [ + 50, + 194, + 545, + 411 + ], + "spans": [ + { + "bbox": [ + 50, + 194, + 545, + 411 + ], + "type": "image", + "image_path": "b7b56650c7a3f8988505033fe7f558153600651b2314a532f9834720382bb0c1.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 46, + 411, + 547, + 467 + ], + "lines": [ + { + "bbox": [ + 46, + 411, + 547, + 467 + ], + "spans": [ + { + "bbox": [ + 46, + 411, + 547, + 467 + ], + "type": "text", + "content": "Figure 6. Comparison of generated driving simulations from (a) Panacea [45], (b) Our method, and (c) Real Scene [3] across multiple views and time steps. Red boxes highlight instance-level appearance consistency, with our method maintaining more coherent actor appearances across frames. Yellow boxes illustrate temporal consistency, where our approach achieves smoother transitions over time. Green boxes and background elements demonstrate our model's capability to capture and accurately model non-direct traffic participants, such as buildings and other environmental features, enhancing realism in the generated scene." + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_caption" + } + ], + "index": 2 + }, + { + "type": "table", + "bbox": [ + 50, + 475, + 287, + 510 + ], + "blocks": [ + { + "bbox": [ + 46, + 159, + 547, + 193 + ], + "lines": [ + { + "bbox": [ + 46, + 159, + 547, + 193 + ], + "spans": [ + { + "bbox": [ + 46, + 159, + 547, + 193 + ], + "type": "text", + "content": "Table 1. Quantitative comparison of video generation fidelity for autonomous driving simulations on the nuScenes validation set. Using UniAD [19] for evaluation on generated data, DrivingSphere achieves high fidelity in video generation, closely matching real-world data across key perception and planning metrics." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 50, + 475, + 287, + 510 + ], + "lines": [ + { + "bbox": [ + 50, + 475, + 287, + 510 + ], + "spans": [ + { + "bbox": [ + 50, + 475, + 287, + 510 + ], + "type": "table", + "html": "
MethodsFID(↓)MMD(↓)
Semicity [22]6340.251
Ours2740.082
", + "image_path": "1e09fc163594c60288afdecd05eae41b240c5f278a979d68077fec3fbdf5b09c.jpg" + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "table_body" + } + ], + "index": 4 + }, + { + "type": "table", + "bbox": [ + 50, + 543, + 287, + 601 + ], + "blocks": [ + { + "bbox": [ + 47, + 510, + 287, + 542 + ], + "lines": [ + { + "bbox": [ + 47, + 510, + 287, + 542 + ], + "spans": [ + { + "bbox": [ + 47, + 510, + 287, + 542 + ], + "type": "text", + "content": "Table 2. Quantitative comparison of scene generation performance, following the setup in SemCity. The scene data generated by our method is rendered into 2D images." + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 50, + 543, + 287, + 601 + ], + "lines": [ + { + "bbox": [ + 50, + 543, + 287, + 601 + ], + "spans": [ + { + "bbox": [ + 50, + 543, + 287, + 601 + ], + "type": "table", + "html": "
Data SourceNC(↑)DAC(↑)TTC(↑)PDMS(↑)
nuScenes [3]0.9930.9950.9470.910
DriveArena [51]0.7920.9420.7710.636
DriveArena* [51]0.8290.9640.8120.698
DrivingSphere0.8520.9680.8530.742
", + "image_path": "b586617ac8e94759fb2b61b792e5a2d6faf500a614d67eeaf165693268613904.jpg" + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "table_body" + } + ], + "index": 6 + }, + { + "bbox": [ + 46, + 641, + 288, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 641, + 288, + 715 + ], + "spans": [ + { + "bbox": [ + 46, + 641, + 288, + 715 + ], + "type": "text", + "content": "Completion (RC) scores, with completing an average of only " + }, + { + "bbox": [ + 46, + 641, + 288, + 715 + ], + "type": "inline_equation", + "content": "11.7\\%" + }, + { + "bbox": [ + 46, + 641, + 288, + 715 + ], + "type": "text", + "content": " of each route. Compared to DriveArena [51], our simulation yields consistently better performance metrics, and, along with the superior visual fidelity demonstrated in Tab. 1, affirms DrivingSphere's effectiveness as a high-fidelity platform for AD algorithm validation." + } + ] + } + ], + "index": 8 + }, + { + "type": "table", + "bbox": [ + 308, + 475, + 545, + 511 + ], + "blocks": [ + { + "bbox": [ + 46, + 601, + 287, + 635 + ], + "lines": [ + { + "bbox": [ + 46, + 601, + 287, + 635 + ], + "spans": [ + { + "bbox": [ + 46, + 601, + 287, + 635 + ], + "type": "text", + "content": "Table 3. Open-loop evaluation of UniAD [19] on original nuScenes sequences, DriveArena's reported results, and our predefined 100-trajectory test for DriveArena* and DrivingSphere." + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 308, + 475, + 545, + 511 + ], + "lines": [ + { + "bbox": [ + 308, + 475, + 545, + 511 + ], + "spans": [ + { + "bbox": [ + 308, + 475, + 545, + 511 + ], + "type": "table", + "html": "
RoutePDMS(↑)RC(↑)ADS(↑)
DriveArena [51]0.69010.06410.0508
DrivingSphere0.72810.11700.0851
", + "image_path": "fbeaff437d5147da414590b97a395889f4a29c4b098e75a77f880aad9d70406d.jpg" + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "table_body" + } + ], + "index": 9 + }, + { + "bbox": [ + 305, + 512, + 545, + 535 + ], + "lines": [ + { + "bbox": [ + 305, + 512, + 545, + 535 + ], + "spans": [ + { + "bbox": [ + 305, + 512, + 545, + 535 + ], + "type": "text", + "content": "Table 4. Closed-loop evaluation of UniAD's [19] performance on 100 pre-defined routes." + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "text" + }, + { + "bbox": [ + 306, + 545, + 379, + 558 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 545, + 379, + 558 + ], + "spans": [ + { + "bbox": [ + 306, + 545, + 379, + 558 + ], + "type": "text", + "content": "5. Conclusion" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 304, + 570, + 547, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 570, + 547, + 713 + ], + "spans": [ + { + "bbox": [ + 304, + 570, + 547, + 713 + ], + "type": "text", + "content": "We presented *DrivingSphere*, a generative closed-loop simulation framework that bridges the gap between traditional closed-loop simulations and open-loop generative models. Through advanced occupancy-based modeling and controllable generation mechanisms, *DrivingSphere* creates realistic, high-fidelity simulation for autonomous driving. Our experiments demonstrate superior visual quality, temporal consistency, and the ability to effectively test AD algorithms in dynamic environments. In future work, we aim to enhance dynamic agent behaviors and incorporate more complex real-world scenarios to further improve the robustness of our simulations." + } + ] + } + ], + "index": 12 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 318, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 318, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 318, + 757 + ], + "type": "text", + "content": "27538" + } + ] + } + ], + "index": 13 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 7 + }, + { + "para_blocks": [ + { + "bbox": [ + 48, + 71, + 107, + 84 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 71, + 107, + 84 + ], + "spans": [ + { + "bbox": [ + 48, + 71, + 107, + 84 + ], + "type": "text", + "content": "References" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 48, + 91, + 288, + 715 + ], + "type": "list", + "angle": 0, + "index": 15, + "blocks": [ + { + "bbox": [ + 53, + 91, + 287, + 134 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 91, + 287, + 134 + ], + "spans": [ + { + "bbox": [ + 53, + 91, + 287, + 134 + ], + "type": "text", + "content": "[1] Maxim Berman, Amal Rannen Triki, and Matthew B Blaschko. The lovasz-softmax loss: A tractable surrogate for the optimization of the intersection-over-union measure in neural networks. In CVPR, 2018. 4" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 53, + 135, + 288, + 191 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 135, + 288, + 191 + ], + "spans": [ + { + "bbox": [ + 53, + 135, + 288, + 191 + ], + "type": "text", + "content": "[2] Andreas Blattmann, Tim Dockhorn, Sumith Kulal, Daniel Mendelevitch, Maciej Kilian, Dominik Lorenz, Yam Levi, Zion English, Vikram Voleti, Adam Letts, et al. Stable video diffusion: Scaling latent video diffusion models to large datasets. arXiv preprint arXiv:2311.15127, 2023. 3" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 53, + 191, + 288, + 245 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 191, + 288, + 245 + ], + "spans": [ + { + "bbox": [ + 53, + 191, + 288, + 245 + ], + "type": "text", + "content": "[3] Holger Caesar, Varun Bankiti, Alex H. Lang, Sourabh Vora, Venice Erin Liong, Qiang Xu, Anush Krishnan, Yu Pan, Giancarlo Baldan, and Oscar Beijbom. nuscenes: A multimodal dataset for autonomous driving. In CVPR, 2020. 1, 2, 4, 7, 8" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 53, + 246, + 288, + 291 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 246, + 288, + 291 + ], + "spans": [ + { + "bbox": [ + 53, + 246, + 288, + 291 + ], + "type": "text", + "content": "[4] Junsong Chen, Yue Wu, Simian Luo, Enze Xie, Sayak Paul, Ping Luo, Hang Zhao, and Zhenguo Li. Pixart-{\\delta} : Fast and controllable image generation with latent consistency models. arXiv preprint arXiv:2401.05252, 2024. 6" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 53, + 291, + 288, + 324 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 291, + 288, + 324 + ], + "spans": [ + { + "bbox": [ + 53, + 291, + 288, + 324 + ], + "type": "text", + "content": "[5] Li Chen, Penghao Wu, Kashyap Chitta, Bernhard Jaeger, Andreas Geiger, and Hongyang Li. End-to-end autonomous driving: Challenges and frontiers. T-PAMI, 2024. 1" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 53, + 324, + 288, + 378 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 324, + 288, + 378 + ], + "spans": [ + { + "bbox": [ + 53, + 324, + 288, + 378 + ], + "type": "text", + "content": "[6] Shaoyu Chen, Bo Jiang, Hao Gao, Bencheng Liao, Qing Xu, Qian Zhang, Chang Huang, Wenyu Liu, and Xinggang Wang. Vadv2: End-to-end vectorized autonomous driving via probabilistic planning. arXiv preprint arXiv:2402.13243, 2024. 2, 7" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 53, + 380, + 288, + 413 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 380, + 288, + 413 + ], + "spans": [ + { + "bbox": [ + 53, + 380, + 288, + 413 + ], + "type": "text", + "content": "[7] Yuan Chen, Zi-han Ding, Ziqin Wang, Yan Wang, Lijun Zhang, and Si Liu. Asynchronous large language model enhanced planner for autonomous driving. In ECCV, 2025. 1" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 53, + 413, + 288, + 479 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 413, + 288, + 479 + ], + "spans": [ + { + "bbox": [ + 53, + 413, + 288, + 479 + ], + "type": "text", + "content": "[8] Daniel Dauner, Marcel Hallgarten, Tianyu Li, Xinshuo Weng, Zhiyu Huang, Zetong Yang, Hongyang Li, Igor Gilitschenski, Boris Ivanovic, Marco Pavone, Andreas Geiger, and Kashyap Chitta. Navsim: Data-driven non-reactive autonomous vehicle simulation and benchmarking. In NeurIPS, 2024. 7, 1" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 53, + 479, + 288, + 523 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 479, + 288, + 523 + ], + "spans": [ + { + "bbox": [ + 53, + 479, + 288, + 523 + ], + "type": "text", + "content": "[9] Wenhao Ding, Yulong Cao, Ding Zhao, Chaowei Xiao, and Marco Pavone. Realgen: Retrieval augmented generation for controllable traffic scenarios. arXiv preprint arXiv:2312.13303, 2023. 3" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 48, + 524, + 288, + 557 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 524, + 288, + 557 + ], + "spans": [ + { + "bbox": [ + 48, + 524, + 288, + 557 + ], + "type": "text", + "content": "[10] Alexey Dosovitskiy, German Ros, Felipe Codevilla, Antonio Lopez, and Vladlen Koltun. Carla: An open urban driving simulator. In CoRL, 2017. 1, 2, 3" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 48, + 558, + 288, + 590 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 558, + 288, + 590 + ], + "spans": [ + { + "bbox": [ + 48, + 558, + 288, + 590 + ], + "type": "text", + "content": "[11] Lan Feng, Quanyi Li, Zhenghao Peng, Shuhan Tan, and Bolei Zhou. Trafficgen: Learning to generate diverse and realistic traffic scenarios. In ICRA, 2023. 3" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 48, + 591, + 288, + 645 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 591, + 288, + 645 + ], + "spans": [ + { + "bbox": [ + 48, + 591, + 288, + 645 + ], + "type": "text", + "content": "[12] Daocheng Fu, Wenjie Lei, Licheng Wen, Pinlong Cai, Song Mao, Min Dou, Botian Shi, and Yu Qiao. Limsim++: A closed-loop platform for deploying multimodal llms in autonomous driving. arXiv preprint arXiv:2402.01246, 2024. 1, 3" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 48, + 647, + 288, + 690 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 647, + 288, + 690 + ], + "spans": [ + { + "bbox": [ + 48, + 647, + 288, + 690 + ], + "type": "text", + "content": "[13] Ruiyuan Gao, Kai Chen, Enze Xie, Lanqing Hong, Zhenguo Li, Dit-Yan Yeung, and Qiang Xu. Magicdrive: Street view generation with diverse 3d geometry control. arXiv preprint arXiv:2310.02601, 2023. 1, 2, 3, 5, 6, 7, 8" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 48, + 691, + 288, + 715 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 691, + 288, + 715 + ], + "spans": [ + { + "bbox": [ + 48, + 691, + 288, + 715 + ], + "type": "text", + "content": "[14] Shenyuan Gao, Jiazhi Yang, Li Chen, Kashyap Chitta, Yihang Qiu, Andreas Geiger, Jun Zhang, and Hongyang Li." + } + ] + } + ], + "index": 14 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 307, + 73, + 547, + 713 + ], + "type": "list", + "angle": 0, + "index": 32, + "blocks": [ + { + "bbox": [ + 326, + 73, + 545, + 95 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 326, + 73, + 545, + 95 + ], + "spans": [ + { + "bbox": [ + 326, + 73, + 545, + 95 + ], + "type": "text", + "content": "Vista: A generalizable driving world model with high fidelity and versatile controllability. In NeurIPS, 2024. 3" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 307, + 95, + 547, + 140 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 95, + 547, + 140 + ], + "spans": [ + { + "bbox": [ + 307, + 95, + 547, + 140 + ], + "type": "text", + "content": "[15] Wencheng Han, Dongqian Guo, Cheng-Zhong Xu, and Jianbing Shen. Dme-driver: Integrating human decision logic and 3d scene perception in autonomous driving. arXiv preprint arXiv:2401.03641, 2024. 1" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 307, + 141, + 547, + 184 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 141, + 547, + 184 + ], + "spans": [ + { + "bbox": [ + 307, + 141, + 547, + 184 + ], + "type": "text", + "content": "[16] Yingqing He, Tianyu Yang, Yong Zhang, Ying Shan, and Qifeng Chen. Latent video diffusion models for high-fidelity long video generation. arXiv preprint arXiv:2211.13221, 2022. 3, 1" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 307, + 185, + 545, + 207 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 185, + 545, + 207 + ], + "spans": [ + { + "bbox": [ + 307, + 185, + 545, + 207 + ], + "type": "text", + "content": "[17] Jonathan Ho, Ajay Jain, and Pieter Abbeel. Denoising diffusion probabilistic models. NeurIPS, 2020. 3, 4, 1" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 307, + 208, + 545, + 231 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 208, + 545, + 231 + ], + "spans": [ + { + "bbox": [ + 307, + 208, + 545, + 231 + ], + "type": "text", + "content": "[18] Qianjiang Hu, Zhimin Zhang, and Wei Hu. Rangeldm: Fast realistic lidar point cloud generation. In ECCV, 2025. 3" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 307, + 232, + 545, + 275 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 232, + 545, + 275 + ], + "spans": [ + { + "bbox": [ + 307, + 232, + 545, + 275 + ], + "type": "text", + "content": "[19] Yihan Hu, Jiazhi Yang, Li Chen, Keyu Li, Chonghao Sima, Xizhou Zhu, Siqi Chai, Senyao Du, Tianwei Lin, Wenhai Wang, et al. Planning-oriented autonomous driving. In CVPR, 2023. 1, 2, 7, 8" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 307, + 276, + 545, + 320 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 276, + 545, + 320 + ], + "spans": [ + { + "bbox": [ + 307, + 276, + 545, + 320 + ], + "type": "text", + "content": "[20] Bo Jiang, Shaoyu Chen, Qing Xu, Bencheng Liao, Jiajie Chen, Helong Zhou, Qian Zhang, Wenyu Liu, Chang Huang, and Xinggang Wang. Vad: Vectorized scene representation for efficient autonomous driving. In ICCV, 2023. 1, 2, 7" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 307, + 321, + 545, + 355 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 321, + 545, + 355 + ], + "spans": [ + { + "bbox": [ + 307, + 321, + 545, + 355 + ], + "type": "text", + "content": "[21] Bernhard Kerbl, Georgios Kopanas, Thomas Leimkuhler, and George Drettakis. 3d gaussian splatting for real-time radiance field rendering. ACM Trans. Graph., 2023. 3" + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 307, + 355, + 545, + 388 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 355, + 545, + 388 + ], + "spans": [ + { + "bbox": [ + 307, + 355, + 545, + 388 + ], + "type": "text", + "content": "[22] Jumin Lee, Sebin Lee, Changho Jo, Woobin Im, Juhyeong Seon, and Sung-Eui Yoon. Semcity: Semantic scene generation with triplane diffusion. In CVPR, 2024. 3, 7, 8, 2" + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 307, + 388, + 545, + 433 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 388, + 545, + 433 + ], + "spans": [ + { + "bbox": [ + 307, + 388, + 545, + 433 + ], + "type": "text", + "content": "[23] Quanyi Li, Zhenghao Peng, Lan Feng, Qihang Zhang, Zhenghai Xue, and Bolei Zhou. Metadrive: Composing diverse driving scenarios for generalizable reinforcement learning. T-PAMI, 2022. 1, 2, 3" + } + ] + } + ], + "index": 25 + }, + { + "bbox": [ + 307, + 434, + 545, + 468 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 434, + 545, + 468 + ], + "spans": [ + { + "bbox": [ + 307, + 434, + 545, + 468 + ], + "type": "text", + "content": "[24] Zhiqi Li, Zhiding Yu, Shiyi Lan, Jiahan Li, Jan Kautz, Tong Lu, and Jose M Alvarez. Is ego status all you need for open-loop end-to-end autonomous driving? In CVPR, 2024. 1" + } + ] + } + ], + "index": 26 + }, + { + "bbox": [ + 307, + 468, + 545, + 501 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 468, + 545, + 501 + ], + "spans": [ + { + "bbox": [ + 307, + 468, + 545, + 501 + ], + "type": "text", + "content": "[25] Zhiqi Li, Zhiding Yu, Shiyi Lan, Jiahan Li, Jan Kautz, Tong Lu, and Jose M Alvarez. Is ego status all you need for open-loop end-to-end autonomous driving? In CVPR, 2024. 1" + } + ] + } + ], + "index": 27 + }, + { + "bbox": [ + 307, + 502, + 545, + 555 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 502, + 545, + 555 + ], + "spans": [ + { + "bbox": [ + 307, + 502, + 545, + 555 + ], + "type": "text", + "content": "[26] William Ljungbergh, Adam Tonderski, Joakim Johnander, Holger Caesar, Kalle Åström, Michael Felsberg, and Christoffer Petersson. Neuroncap: Photorealistic closed-loop safety testing for autonomous driving. In ECCV, 2025. 3" + } + ] + } + ], + "index": 28 + }, + { + "bbox": [ + 307, + 557, + 545, + 612 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 557, + 545, + 612 + ], + "spans": [ + { + "bbox": [ + 307, + 557, + 545, + 612 + ], + "type": "text", + "content": "[27] Pablo Alvarez Lopez, Michael Behrisch, Laura Bieker-Walz, Jakob Erdmann, Yun-Pang Flotteröd, Robert Hilbrich, Leonhard Lücken, Johannes Rummel, Peter Wagner, and Evamarie Wießner. Microscopic traffic simulation using sumo. In IEEE ITSC, 2018. 3" + } + ] + } + ], + "index": 29 + }, + { + "bbox": [ + 307, + 613, + 545, + 669 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 613, + 545, + 669 + ], + "spans": [ + { + "bbox": [ + 307, + 613, + 545, + 669 + ], + "type": "text", + "content": "[28] Enhui Ma, Lijun Zhou, Tao Tang, Zhan Zhang, Dong Han, Junpeng Jiang, Kun Zhan, Peng Jia, Xianpeng Lang, Haiyang Sun, et al. Unleashing generalization of end-to-end autonomous driving with controllable long video generation. arXiv preprint arXiv:2406.01349, 2024. 5, 1" + } + ] + } + ], + "index": 30 + }, + { + "bbox": [ + 307, + 669, + 545, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 669, + 545, + 713 + ], + "spans": [ + { + "bbox": [ + 307, + 669, + 545, + 713 + ], + "type": "text", + "content": "[29] Ben Mildenhall, Pratul P Srinivasan, Matthew Tancik, Jonathan T Barron, Ravi Ramamoorthi, and Ren Ng. Nerf: Representing scenes as neural radiance fields for view synthesis. Communications of the ACM, 2021. 3, 6" + } + ] + } + ], + "index": 31 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "type": "text", + "content": "27539" + } + ] + } + ], + "index": 33 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 8 + }, + { + "para_blocks": [ + { + "bbox": [ + 48, + 72, + 288, + 713 + ], + "type": "list", + "angle": 0, + "index": 15, + "blocks": [ + { + "bbox": [ + 48, + 72, + 287, + 105 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 72, + 287, + 105 + ], + "spans": [ + { + "bbox": [ + 48, + 72, + 287, + 105 + ], + "type": "text", + "content": "[30] Kazuto Nakashima and Ryo Kurazume. Lidar data synthesis with denoising diffusion probabilistic models. In ICRA, 2024. 3" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 48, + 106, + 287, + 160 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 106, + 287, + 160 + ], + "spans": [ + { + "bbox": [ + 48, + 106, + 287, + 160 + ], + "type": "text", + "content": "[31] Alec Radford, Jong Wook Kim, Chris Hallacy, Aditya Ramesh, Gabriel Goh, Sandhini Agarwal, Girish Sastry, Amanda Askell, Pamela Mishkin, Jack Clark, et al. Learning transferable visual models from natural language supervision. In ICML, 2021. 4" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 48, + 161, + 288, + 216 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 161, + 288, + 216 + ], + "spans": [ + { + "bbox": [ + 48, + 161, + 288, + 216 + ], + "type": "text", + "content": "[32] Colin Raffel, Noam Shazeer, Adam Roberts, Katherine Lee, Sharan Narang, Michael Matena, Yanqi Zhou, Wei Li, and Peter J. Liu. Exploring the limits of transfer learning with a unified text-to-text transformer. Journal of Machine Learning Research, 2020. 6" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 48, + 217, + 287, + 249 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 217, + 287, + 249 + ], + "spans": [ + { + "bbox": [ + 48, + 217, + 287, + 249 + ], + "type": "text", + "content": "[33] Prabhu Ramachandran and Gaël Varoquaux. Mayavi: 3d visualization of scientific data. Computing in Science & Engineering, 2011. 6" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 48, + 250, + 287, + 282 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 250, + 287, + 282 + ], + "spans": [ + { + "bbox": [ + 48, + 250, + 287, + 282 + ], + "type": "text", + "content": "[34] Haoxi Ran, Vitor Guizilini, and Yue Wang. Towards realistic scene generation with lidar diffusion models. In ECCV, 2024. 3" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 48, + 283, + 287, + 316 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 283, + 287, + 316 + ], + "spans": [ + { + "bbox": [ + 48, + 283, + 287, + 316 + ], + "type": "text", + "content": "[35] Robin Rombach, Andreas Blattmann, Dominik Lorenz, Patrick Esser, and Björn Ommer. High-resolution image synthesis with latent diffusion models. In CVPR, 2022. 4" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 48, + 316, + 287, + 360 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 316, + 287, + 360 + ], + "spans": [ + { + "bbox": [ + 48, + 316, + 287, + 360 + ], + "type": "text", + "content": "[36] Hao Shao, Yuxuan Hu, Letian Wang, Guanglu Song, Steven L Waslander, Yu Liu, and Hongsheng Li. Lmdrive: Closed-loop end-to-end driving with large language models. In CVPR, 2024. 1, 2" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 48, + 361, + 287, + 414 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 361, + 287, + 414 + ], + "spans": [ + { + "bbox": [ + 48, + 361, + 287, + 414 + ], + "type": "text", + "content": "[37] Pei Sun, Henrik Kretzschmar, Xerxes Dotiwalla, Aurelien Chouard, Vijaysai Patnaik, Paul Tsui, James Guo, Yin Zhou, Yuning Chai, Benjamin Caine, et al. Scalability in perception for autonomous driving: Waymo open dataset. In CVPR, 2020. 1" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 48, + 415, + 287, + 448 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 415, + 287, + 448 + ], + "spans": [ + { + "bbox": [ + 48, + 415, + 287, + 448 + ], + "type": "text", + "content": "[38] Alexander Swerdlow, Runsheng Xu, and Bolei Zhou. Streetview image generation from a bird's-eye view layout. IEEE RA-L, 2024. 3" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 48, + 449, + 287, + 482 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 449, + 287, + 482 + ], + "spans": [ + { + "bbox": [ + 48, + 449, + 287, + 482 + ], + "type": "text", + "content": "[39] Shuhan Tan, Boris Ivanovic, Xinshuo Weng, Marco Pavone, and Philipp Kraehenbuehl. Language conditioned traffic generation. arXiv preprint arXiv:2307.07947, 2023. 3" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 48, + 482, + 287, + 504 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 482, + 287, + 504 + ], + "spans": [ + { + "bbox": [ + 48, + 482, + 287, + 504 + ], + "type": "text", + "content": "[40] Aaron Van Den Oord, Oriol Vinyals, et al. Neural discrete representation learning. NeurIPS, 2017. 4" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 48, + 504, + 287, + 558 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 504, + 287, + 558 + ], + "spans": [ + { + "bbox": [ + 48, + 504, + 287, + 558 + ], + "type": "text", + "content": "[41] Lening Wang, Wenzhao Zheng, Yilong Ren, Han Jiang, Zhiyong Cui, Haiyang Yu, and Jiwen Lu. Occsora: 4d occupancy generation models as world simulators for autonomous driving. arXiv preprint arXiv:2405.20337, 2024. 2, 3, 5, 1" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 48, + 559, + 287, + 624 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 559, + 287, + 624 + ], + "spans": [ + { + "bbox": [ + 48, + 559, + 287, + 624 + ], + "type": "text", + "content": "[42] Wenhai Wang, Jiangwei Xie, ChuanYang Hu, Haoming Zou, Jianan Fan, Wenwen Tong, Yang Wen, Silei Wu, Hanming Deng, Zhiqi Li, et al. Drivemlm: Aligning multi-modal large language models with behavioral planning states for autonomous driving. arXiv preprint arXiv:2312.09245, 2023. 1" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 48, + 625, + 287, + 669 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 625, + 287, + 669 + ], + "spans": [ + { + "bbox": [ + 48, + 625, + 287, + 669 + ], + "type": "text", + "content": "[43] Xiaofeng Wang, Zheng Zhu, Guan Huang, Xinze Chen, Jia-gang Zhu, and Jiwen Lu. Drivedreamer: Towards real-world-driven world models for autonomous driving. arXiv preprint arXiv:2309.09777, 2023. 1, 2, 3, 5" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 48, + 670, + 287, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 670, + 287, + 713 + ], + "spans": [ + { + "bbox": [ + 48, + 670, + 287, + 713 + ], + "type": "text", + "content": "[44] Yuqi Wang, Jiawei He, Lue Fan, Hongxin Li, Yuntao Chen, and Zhaoxiang Zhang. Driving into the future: Multiview visual forecasting and planning with world model for autonomous driving. In CVPR, 2024. 6, 1, 2" + } + ] + } + ], + "index": 14 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 307, + 72, + 546, + 715 + ], + "type": "list", + "angle": 0, + "index": 30, + "blocks": [ + { + "bbox": [ + 307, + 72, + 546, + 127 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 72, + 546, + 127 + ], + "spans": [ + { + "bbox": [ + 307, + 72, + 546, + 127 + ], + "type": "text", + "content": "[45] Yuqing Wen, Yucheng Zhao, Yingfei Liu, Fan Jia, Yanhui Wang, Chong Luo, Chi Zhang, Tiancai Wang, Xiaoyan Sun, and Xiangyu Zhang. Panacea: Panoramic and controllable video generation for autonomous driving. In CVPR, 2024. 3, 7, 8, 1, 2" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 307, + 129, + 546, + 172 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 129, + 546, + 172 + ], + "spans": [ + { + "bbox": [ + 307, + 129, + 546, + 172 + ], + "type": "text", + "content": "[46] Licheng Wenl, Daocheng Fu, Song Mao, Pinlong Cai, Min Dou, Yikang Li, and Yu Qiao. Limsim: A long-term interactive multi-scenario traffic simulator. In IEEE ITSC, 2023. 1, 3, 7" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 307, + 174, + 546, + 229 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 174, + 546, + 229 + ], + "spans": [ + { + "bbox": [ + 307, + 174, + 546, + 229 + ], + "type": "text", + "content": "[47] Zirui Wu, Tianyu Liu, Liyi Luo, Zhide Zhong, Jianteng Chen, Hongmin Xiao, Chao Hou, Haozhe Lou, Yuantao Chen, Runyi Yang, et al. Mars: An instance-aware, modular and realistic simulator for autonomous driving. In CAAI International Conference on Artificial Intelligence, 2023. 3" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 307, + 230, + 546, + 274 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 230, + 546, + 274 + ], + "spans": [ + { + "bbox": [ + 307, + 230, + 546, + 274 + ], + "type": "text", + "content": "[48] Zhenhua Xu, Yujia Zhang, Enze Xie, Zhen Zhao, Yong Guo, Kwan-Yee K Wong, Zhenguo Li, and Hengshuang Zhao. Drivegpt4: Interpretable end-to-end autonomous driving via large language model. IEEE RA-L, 2024. 1" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 307, + 275, + 546, + 319 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 275, + 546, + 319 + ], + "spans": [ + { + "bbox": [ + 307, + 275, + 546, + 319 + ], + "type": "text", + "content": "[49] Tianyi Yan, Junbo Yin, Xianpeng Lang, Ruigang Yang, Cheng-Zhong Xu, and Jianbing Shen. Olidm: Object-aware lidar diffusion models for autonomous driving. In AAAI, 2025. 3" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 307, + 319, + 546, + 365 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 319, + 546, + 365 + ], + "spans": [ + { + "bbox": [ + 307, + 319, + 546, + 365 + ], + "type": "text", + "content": "[50] Kairui Yang, Enhui Ma, Jibin Peng, Qing Guo, Di Lin, and Kaicheng Yu. Bevcontrol: Accurately controlling street-view elements with multi-perspective consistency via bev sketch layout. arXiv preprint arXiv:2308.01661, 2023. 3" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 307, + 365, + 546, + 431 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 365, + 546, + 431 + ], + "spans": [ + { + "bbox": [ + 307, + 365, + 546, + 431 + ], + "type": "text", + "content": "[51] Xuemeng Yang, Licheng Wen, Yukai Ma, Jianbiao Mei, Xin Li, Tiantian Wei, Wenjie Lei, Daocheng Fu, Pinlong Cai, Min Dou, Botian Shi, Liang He, Yong Liu, and Yu Qiao. Drivearena: A closed-loop generative simulation platform for autonomous driving. arXiv preprint arXiv:2408.00415, 2024. 1, 3, 7, 8, 2, 5" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 307, + 432, + 546, + 475 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 432, + 546, + 475 + ], + "spans": [ + { + "bbox": [ + 307, + 432, + 546, + 475 + ], + "type": "text", + "content": "[52] Ze Yang, Yun Chen, Jingkang Wang, Sivabalan Manivasagam, Wei-Chiu Ma, Anqi Joyce Yang, and Raquel Urtasun. Unisim: A neural closed-loop sensor simulator. In CVPR, 2023. 3" + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 307, + 476, + 546, + 509 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 476, + 546, + 509 + ], + "spans": [ + { + "bbox": [ + 307, + 476, + 546, + 509 + ], + "type": "text", + "content": "[53] Jimuyang Zhang, Zanming Huang, Arijit Ray, and Eshed Ohn-Bar. Feedback-guided autonomous driving. In CVPR, 2024. 1" + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 307, + 510, + 546, + 554 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 510, + 546, + 554 + ], + "spans": [ + { + "bbox": [ + 307, + 510, + 546, + 554 + ], + "type": "text", + "content": "[54] Linrui Zhang, Zhenghao Peng, Quanyi Li, and Bolei Zhou. Cat: Closed-loop adversarial training for safe end-to-end driving. In 7th Annual Conference on Robot Learning, 2023. 1" + } + ] + } + ], + "index": 25 + }, + { + "bbox": [ + 307, + 555, + 546, + 589 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 555, + 546, + 589 + ], + "spans": [ + { + "bbox": [ + 307, + 555, + 546, + 589 + ], + "type": "text", + "content": "[55] Lvmin Zhang, Anyi Rao, and Maneesh Agrawala. Adding conditional control to text-to-image diffusion models. In ICCV, 2023. 4" + } + ] + } + ], + "index": 26 + }, + { + "bbox": [ + 307, + 590, + 546, + 634 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 590, + 546, + 634 + ], + "spans": [ + { + "bbox": [ + 307, + 590, + 546, + 634 + ], + "type": "text", + "content": "[56] Yuheng Zhang, Tianjian Ouyang, Fudan Yu, Cong Ma, Lei Qiao, Wei Wu, Jian Yuan, and Yong Li. Lcsim: A large-scale controllable traffic simulator. arXiv preprint arXiv:2406.19781, 2024. 1, 3" + } + ] + } + ], + "index": 27 + }, + { + "bbox": [ + 307, + 635, + 546, + 689 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 635, + 546, + 689 + ], + "spans": [ + { + "bbox": [ + 307, + 635, + 546, + 689 + ], + "type": "text", + "content": "[57] Guosheng Zhao, Xiaofeng Wang, Zheng Zhu, Xinze Chen, Guan Huang, Xiaoyi Bao, and Xingang Wang. Drivedreamer-2: Llm-enhanced world models for diverse driving video generation. arXiv preprint arXiv:2403.06845, 2024. 1, 2, 3, 5" + } + ] + } + ], + "index": 28 + }, + { + "bbox": [ + 307, + 690, + 546, + 715 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 690, + 546, + 715 + ], + "spans": [ + { + "bbox": [ + 307, + 690, + 546, + 715 + ], + "type": "text", + "content": "[58] Wenzhao Zheng, Weiliang Chen, Yuanhui Huang, Borui Zhang, Yueqi Duan, and Jiwen Lu. Occworld: Learning a 3d" + } + ] + } + ], + "index": 29 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 318, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 318, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 318, + 757 + ], + "type": "text", + "content": "27540" + } + ] + } + ], + "index": 31 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 9 + }, + { + "para_blocks": [ + { + "bbox": [ + 48, + 72, + 286, + 241 + ], + "type": "list", + "angle": 0, + "index": 5, + "blocks": [ + { + "bbox": [ + 66, + 72, + 286, + 94 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 66, + 72, + 286, + 94 + ], + "spans": [ + { + "bbox": [ + 66, + 72, + 286, + 94 + ], + "type": "text", + "content": "occupancy world model for autonomous driving. In ECCV, 2024.3,1,2" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 48, + 96, + 286, + 138 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 96, + 286, + 138 + ], + "spans": [ + { + "bbox": [ + 48, + 96, + 286, + 138 + ], + "type": "text", + "content": "[59] Wenzhao Zheng, Ruiqi Song, Xianda Guo, Chenming Zhang, and Long Chen. Genad: Generative end-to-end autonomous driving. arXiv preprint arXiv:2402.11502, 2024. 1" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 48, + 140, + 286, + 184 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 140, + 286, + 184 + ], + "spans": [ + { + "bbox": [ + 48, + 140, + 286, + 184 + ], + "type": "text", + "content": "[60] Zangwei Zheng, Xiangyu Peng, Tianji Yang, Chenhui Shen, Shenggui Li, Hongxin Liu, Yukun Zhou, Tianyi Li, and Yang You. Open-sora: Democratizing efficient video production for all, 2024. 5, 6, 1" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 48, + 186, + 286, + 207 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 186, + 286, + 207 + ], + "spans": [ + { + "bbox": [ + 48, + 186, + 286, + 207 + ], + "type": "text", + "content": "[61] Vlas Zyrianov, Xiyue Zhu, and Shenlong Wang. Learning to generate realistic lidar point cloud. In ECCV, 2022. 3" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 48, + 209, + 286, + 241 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 209, + 286, + 241 + ], + "spans": [ + { + "bbox": [ + 48, + 209, + 286, + 241 + ], + "type": "text", + "content": "[62] Vlas Zyrianov, Henry Che, Zhijian Liu, and Shenlong Wang. Lidardm: Generative lidar simulation in a generated world. arXiv preprint arXiv:2404.02903, 2024. 3" + } + ] + } + ], + "index": 4 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "text", + "content": "27541" + } + ] + } + ], + "index": 6 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 10 + } + ], + "_backend": "vlm", + "_version_name": "2.6.4" +} \ No newline at end of file